diff --git a/v2.1/english/cond_prefill.mlmodelc/analytics/coremldata.bin b/v2.1/english/cond_prefill.mlmodelc/analytics/coremldata.bin new file mode 100644 index 0000000000000000000000000000000000000000..3b3bebe8e3864d029ad3c8e470015b188bd5f908 --- /dev/null +++ b/v2.1/english/cond_prefill.mlmodelc/analytics/coremldata.bin @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:324406d83f9685577c6665f231fdfa02518b8addceb54a780feaa968a04c33df +size 243 diff --git a/v2.1/english/cond_prefill.mlmodelc/coremldata.bin b/v2.1/english/cond_prefill.mlmodelc/coremldata.bin new file mode 100644 index 0000000000000000000000000000000000000000..a224d3b18292c6c40ebfacce1cb01029c1bdf650 --- /dev/null +++ b/v2.1/english/cond_prefill.mlmodelc/coremldata.bin @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:12714da8d2ebae282e353e6785d4472ff5b87f656661d60eff769c584969f415 +size 1151 diff --git a/v2.1/english/cond_prefill.mlmodelc/model.mil b/v2.1/english/cond_prefill.mlmodelc/model.mil new file mode 100644 index 0000000000000000000000000000000000000000..203b091a2c5043d4b8c07f4402d5a5f0319b3c3f --- /dev/null +++ b/v2.1/english/cond_prefill.mlmodelc/model.mil @@ -0,0 +1,1309 @@ +program(1.0) +[buildInfo = dict, tensor>({{"coremlc-component-MIL", "3520.4.1"}, {"coremlc-version", "3520.5.1"}, {"coremltools-component-torch", "2.12.0"}, {"coremltools-source-dialect", "TorchScript"}, {"coremltools-version", "9.0"}})] +{ + func main(tensor cache0, tensor cache1, tensor cache2, tensor cache3, tensor cache4, tensor cache5, tensor conditioning, tensor position0, tensor position1, tensor position2, tensor position3, tensor position4, tensor position5, tensor valid_len) { + tensor input_3_axes_0 = const()[name = tensor("input_3_axes_0"), val = tensor([-1])]; + tensor conditioning_to_fp16_dtype_0 = const()[name = tensor("conditioning_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor norm0_1_weight_to_fp16 = const()[name = tensor("norm0_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(64)))]; + tensor norm0_1_bias_to_fp16 = const()[name = tensor("norm0_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(2176)))]; + tensor var_48_to_fp16 = const()[name = tensor("op_48_to_fp16"), val = tensor(0x1.5p-17)]; + tensor conditioning_to_fp16 = cast(dtype = conditioning_to_fp16_dtype_0, x = conditioning)[name = tensor("cast_159")]; + tensor input_3_cast_fp16 = layer_norm(axes = input_3_axes_0, beta = norm0_1_bias_to_fp16, epsilon = var_48_to_fp16, gamma = norm0_1_weight_to_fp16, x = conditioning_to_fp16)[name = tensor("input_3_cast_fp16")]; + tensor attn0_in_proj_weight_to_fp16 = const()[name = tensor("attn0_in_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(4288)))]; + tensor linear_0_bias_0_to_fp16 = const()[name = tensor("linear_0_bias_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(6295808)))]; + tensor linear_0_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = attn0_in_proj_weight_to_fp16, x = input_3_cast_fp16)[name = tensor("linear_0_cast_fp16")]; + tensor var_62 = const()[name = tensor("op_62"), val = tensor([1, 256, 3, 16, 64])]; + tensor qkv_1_cast_fp16 = reshape(shape = var_62, x = linear_0_cast_fp16)[name = tensor("qkv_1_cast_fp16")]; + tensor q_1_begin_0 = const()[name = tensor("q_1_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor q_1_end_0 = const()[name = tensor("q_1_end_0"), val = tensor([1, 256, 1, 16, 64])]; + tensor q_1_end_mask_0 = const()[name = tensor("q_1_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor q_1_squeeze_mask_0 = const()[name = tensor("q_1_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor q_1_cast_fp16 = slice_by_index(begin = q_1_begin_0, end = q_1_end_0, end_mask = q_1_end_mask_0, squeeze_mask = q_1_squeeze_mask_0, x = qkv_1_cast_fp16)[name = tensor("q_1_cast_fp16")]; + tensor k_1_begin_0 = const()[name = tensor("k_1_begin_0"), val = tensor([0, 0, 1, 0, 0])]; + tensor k_1_end_0 = const()[name = tensor("k_1_end_0"), val = tensor([1, 256, 2, 16, 64])]; + tensor k_1_end_mask_0 = const()[name = tensor("k_1_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor k_1_squeeze_mask_0 = const()[name = tensor("k_1_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor k_1_cast_fp16 = slice_by_index(begin = k_1_begin_0, end = k_1_end_0, end_mask = k_1_end_mask_0, squeeze_mask = k_1_squeeze_mask_0, x = qkv_1_cast_fp16)[name = tensor("k_1_cast_fp16")]; + tensor v_1_begin_0 = const()[name = tensor("v_1_begin_0"), val = tensor([0, 0, 2, 0, 0])]; + tensor v_1_end_0 = const()[name = tensor("v_1_end_0"), val = tensor([1, 256, 3, 16, 64])]; + tensor v_1_end_mask_0 = const()[name = tensor("v_1_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor v_1_squeeze_mask_0 = const()[name = tensor("v_1_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor v_1_cast_fp16 = slice_by_index(begin = v_1_begin_0, end = v_1_end_0, end_mask = v_1_end_mask_0, squeeze_mask = v_1_squeeze_mask_0, x = qkv_1_cast_fp16)[name = tensor("v_1_cast_fp16")]; + tensor var_120 = const()[name = tensor("op_120"), val = tensor([1, 1])]; + tensor position0_to_fp16_dtype_0 = const()[name = tensor("position0_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor position0_to_fp16 = cast(dtype = position0_to_fp16_dtype_0, x = position0)[name = tensor("cast_158")]; + tensor var_121_cast_fp16 = reshape(shape = var_120, x = position0_to_fp16)[name = tensor("op_121_cast_fp16")]; + tensor ts_1_promoted_to_fp16 = const()[name = tensor("ts_1_promoted_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(6302016)))]; + tensor ts_3_cast_fp16 = add(x = ts_1_promoted_to_fp16, y = var_121_cast_fp16)[name = tensor("ts_3_cast_fp16")]; + tensor var_128 = const()[name = tensor("op_128"), val = tensor([1, 256, 1, 1])]; + tensor ts_5_cast_fp16 = reshape(shape = var_128, x = ts_3_cast_fp16)[name = tensor("ts_5_cast_fp16")]; + tensor var_135 = const()[name = tensor("op_135"), val = tensor([1, 256, 16, 32, 2])]; + tensor q_complex_1_cast_fp16 = reshape(shape = var_135, x = q_1_cast_fp16)[name = tensor("q_complex_1_cast_fp16")]; + tensor var_142 = const()[name = tensor("op_142"), val = tensor([1, 256, 16, 32, 2])]; + tensor k_complex_1_cast_fp16 = reshape(shape = var_142, x = k_1_cast_fp16)[name = tensor("k_complex_1_cast_fp16")]; + tensor var_146_begin_0 = const()[name = tensor("op_146_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_146_end_0 = const()[name = tensor("op_146_end_0"), val = tensor([1, 256, 16, 32, 1])]; + tensor var_146_end_mask_0 = const()[name = tensor("op_146_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_146_squeeze_mask_0 = const()[name = tensor("op_146_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_146_cast_fp16 = slice_by_index(begin = var_146_begin_0, end = var_146_end_0, end_mask = var_146_end_mask_0, squeeze_mask = var_146_squeeze_mask_0, x = q_complex_1_cast_fp16)[name = tensor("op_146_cast_fp16")]; + tensor var_154_begin_0 = const()[name = tensor("op_154_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_154_end_0 = const()[name = tensor("op_154_end_0"), val = tensor([1, 256, 16, 32, 2])]; + tensor var_154_end_mask_0 = const()[name = tensor("op_154_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_154_squeeze_mask_0 = const()[name = tensor("op_154_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_154_cast_fp16 = slice_by_index(begin = var_154_begin_0, end = var_154_end_0, end_mask = var_154_end_mask_0, squeeze_mask = var_154_squeeze_mask_0, x = q_complex_1_cast_fp16)[name = tensor("op_154_cast_fp16")]; + tensor var_162_begin_0 = const()[name = tensor("op_162_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_162_end_0 = const()[name = tensor("op_162_end_0"), val = tensor([1, 256, 16, 32, 1])]; + tensor var_162_end_mask_0 = const()[name = tensor("op_162_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_162_squeeze_mask_0 = const()[name = tensor("op_162_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_162_cast_fp16 = slice_by_index(begin = var_162_begin_0, end = var_162_end_0, end_mask = var_162_end_mask_0, squeeze_mask = var_162_squeeze_mask_0, x = k_complex_1_cast_fp16)[name = tensor("op_162_cast_fp16")]; + tensor var_170_begin_0 = const()[name = tensor("op_170_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_170_end_0 = const()[name = tensor("op_170_end_0"), val = tensor([1, 256, 16, 32, 2])]; + tensor var_170_end_mask_0 = const()[name = tensor("op_170_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_170_squeeze_mask_0 = const()[name = tensor("op_170_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_170_cast_fp16 = slice_by_index(begin = var_170_begin_0, end = var_170_end_0, end_mask = var_170_end_mask_0, squeeze_mask = var_170_squeeze_mask_0, x = k_complex_1_cast_fp16)[name = tensor("op_170_cast_fp16")]; + tensor freqs_1_to_fp16 = const()[name = tensor("freqs_1_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(6302592)))]; + tensor var_176_cast_fp16 = mul(x = freqs_1_to_fp16, y = ts_5_cast_fp16)[name = tensor("op_176_cast_fp16")]; + tensor rotr_1_cast_fp16 = cos(x = var_176_cast_fp16)[name = tensor("rotr_1_cast_fp16")]; + tensor roti_1_cast_fp16 = sin(x = var_176_cast_fp16)[name = tensor("roti_1_cast_fp16")]; + tensor var_180_cast_fp16 = mul(x = var_146_cast_fp16, y = rotr_1_cast_fp16)[name = tensor("op_180_cast_fp16")]; + tensor var_181_cast_fp16 = mul(x = var_154_cast_fp16, y = roti_1_cast_fp16)[name = tensor("op_181_cast_fp16")]; + tensor qor_1_cast_fp16 = sub(x = var_180_cast_fp16, y = var_181_cast_fp16)[name = tensor("qor_1_cast_fp16")]; + tensor var_184_cast_fp16 = mul(x = var_146_cast_fp16, y = roti_1_cast_fp16)[name = tensor("op_184_cast_fp16")]; + tensor var_185_cast_fp16 = mul(x = var_154_cast_fp16, y = rotr_1_cast_fp16)[name = tensor("op_185_cast_fp16")]; + tensor qoi_1_cast_fp16 = add(x = var_184_cast_fp16, y = var_185_cast_fp16)[name = tensor("qoi_1_cast_fp16")]; + tensor var_188_cast_fp16 = mul(x = var_162_cast_fp16, y = rotr_1_cast_fp16)[name = tensor("op_188_cast_fp16")]; + tensor var_189_cast_fp16 = mul(x = var_170_cast_fp16, y = roti_1_cast_fp16)[name = tensor("op_189_cast_fp16")]; + tensor kor_1_cast_fp16 = sub(x = var_188_cast_fp16, y = var_189_cast_fp16)[name = tensor("kor_1_cast_fp16")]; + tensor var_192_cast_fp16 = mul(x = var_162_cast_fp16, y = roti_1_cast_fp16)[name = tensor("op_192_cast_fp16")]; + tensor var_193_cast_fp16 = mul(x = var_170_cast_fp16, y = rotr_1_cast_fp16)[name = tensor("op_193_cast_fp16")]; + tensor koi_1_cast_fp16 = add(x = var_192_cast_fp16, y = var_193_cast_fp16)[name = tensor("koi_1_cast_fp16")]; + tensor qo_1_axis_0 = const()[name = tensor("qo_1_axis_0"), val = tensor(-1)]; + tensor qo_1_cast_fp16 = stack(axis = qo_1_axis_0, values = (qor_1_cast_fp16, qoi_1_cast_fp16))[name = tensor("qo_1_cast_fp16")]; + tensor ko_1_axis_0 = const()[name = tensor("ko_1_axis_0"), val = tensor(-1)]; + tensor ko_1_cast_fp16 = stack(axis = ko_1_axis_0, values = (kor_1_cast_fp16, koi_1_cast_fp16))[name = tensor("ko_1_cast_fp16")]; + tensor var_226 = const()[name = tensor("op_226"), val = tensor([1, 256, 16, 64])]; + tensor q_3_cast_fp16 = reshape(shape = var_226, x = qo_1_cast_fp16)[name = tensor("q_3_cast_fp16")]; + tensor var_232 = const()[name = tensor("op_232"), val = tensor([1, 256, 16, 64])]; + tensor k_3_cast_fp16 = reshape(shape = var_232, x = ko_1_cast_fp16)[name = tensor("k_3_cast_fp16")]; + tensor var_249_promoted_to_fp16 = const()[name = tensor("op_249_promoted_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(6302720)))]; + tensor write_indices_float_1_cast_fp16 = add(x = var_121_cast_fp16, y = var_249_promoted_to_fp16)[name = tensor("write_indices_float_1_cast_fp16")]; + tensor _inversed_253_y_0_to_fp16 = const()[name = tensor("_inversed_253_y_0_to_fp16"), val = tensor(0x1p-9)]; + tensor _inversed_253_cast_fp16 = mul(x = write_indices_float_1_cast_fp16, y = _inversed_253_y_0_to_fp16)[name = tensor("_inversed_253_cast_fp16")]; + tensor var_254_cast_fp16 = floor(x = _inversed_253_cast_fp16)[name = tensor("op_254_cast_fp16")]; + tensor var_255_to_fp16 = const()[name = tensor("op_255_to_fp16"), val = tensor(0x1p+9)]; + tensor var_256_cast_fp16 = mul(x = var_254_cast_fp16, y = var_255_to_fp16)[name = tensor("op_256_cast_fp16")]; + tensor write_indices_float_3_cast_fp16 = sub(x = write_indices_float_1_cast_fp16, y = var_256_cast_fp16)[name = tensor("write_indices_float_3_cast_fp16")]; + tensor var_265 = const()[name = tensor("op_265"), val = tensor([1, 1])]; + tensor valid_len_to_fp16_dtype_0 = const()[name = tensor("valid_len_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor valid_len_to_fp16 = cast(dtype = valid_len_to_fp16_dtype_0, x = valid_len)[name = tensor("cast_157")]; + tensor var_266_cast_fp16 = reshape(shape = var_265, x = valid_len_to_fp16)[name = tensor("op_266_cast_fp16")]; + tensor token_valid_1_cast_fp16 = less(x = var_249_promoted_to_fp16, y = var_266_cast_fp16)[name = tensor("token_valid_1_cast_fp16")]; + tensor var_274_value_0_to_fp16 = const()[name = tensor("op_274_value_0_to_fp16"), val = tensor(0x1.ffp+8)]; + tensor var_274_cast_fp16 = fill_like(ref_tensor = write_indices_float_3_cast_fp16, value = var_274_value_0_to_fp16)[name = tensor("op_274_cast_fp16")]; + tensor write_indices_float_5_cast_fp16 = select(a = write_indices_float_3_cast_fp16, b = var_274_cast_fp16, cond = token_valid_1_cast_fp16)[name = tensor("write_indices_float_5_cast_fp16")]; + tensor var_280_dtype_0 = const()[name = tensor("op_280_dtype_0"), val = tensor("int32")]; + tensor var_285 = const()[name = tensor("op_285"), val = tensor([1, 256, 1, 1])]; + tensor write_indices_float_5_cast_fp16_to_int32 = cast(dtype = var_280_dtype_0, x = write_indices_float_5_cast_fp16)[name = tensor("cast_156")]; + tensor var_286 = reshape(shape = var_285, x = write_indices_float_5_cast_fp16_to_int32)[name = tensor("op_286")]; + tensor write_indices_1_reps_0 = const()[name = tensor("write_indices_1_reps_0"), val = tensor([1, 1, 16, 64])]; + tensor write_indices_1 = tile(reps = write_indices_1_reps_0, x = var_286)[name = tensor("write_indices_1")]; + tensor var_296_begin_0 = const()[name = tensor("op_296_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_296_end_0 = const()[name = tensor("op_296_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor var_296_end_mask_0 = const()[name = tensor("op_296_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_296_squeeze_mask_0 = const()[name = tensor("op_296_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor cache0_to_fp16_dtype_0 = const()[name = tensor("cache0_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor cache0_to_fp16 = cast(dtype = cache0_to_fp16_dtype_0, x = cache0)[name = tensor("cast_155")]; + tensor var_296_cast_fp16 = slice_by_index(begin = var_296_begin_0, end = var_296_end_0, end_mask = var_296_end_mask_0, squeeze_mask = var_296_squeeze_mask_0, x = cache0_to_fp16)[name = tensor("op_296_cast_fp16")]; + tensor var_298_axis_0 = const()[name = tensor("op_298_axis_0"), val = tensor(1)]; + tensor var_298_mode_0 = const()[name = tensor("op_298_mode_0"), val = tensor("update")]; + tensor var_298_validate_indices_0 = const()[name = tensor("op_298_validate_indices_0"), val = tensor(false)]; + tensor var_298_cast_fp16 = scatter_along_axis(axis = var_298_axis_0, data = var_296_cast_fp16, indices = write_indices_1, mode = var_298_mode_0, updates = k_3_cast_fp16, validate_indices = var_298_validate_indices_0)[name = tensor("op_298_cast_fp16")]; + tensor concat_2 = const()[name = tensor("concat_2"), val = tensor([0, 0, 0, 0, 0])]; + tensor concat_3 = const()[name = tensor("concat_3"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_1_internal_tensor_assign_1_stride_0 = const()[name = tensor("new_cache_1_internal_tensor_assign_1_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_1_internal_tensor_assign_1_begin_mask_0 = const()[name = tensor("new_cache_1_internal_tensor_assign_1_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_1_internal_tensor_assign_1_end_mask_0 = const()[name = tensor("new_cache_1_internal_tensor_assign_1_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_1_internal_tensor_assign_1_squeeze_mask_0 = const()[name = tensor("new_cache_1_internal_tensor_assign_1_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_10 = const()[name = tensor("shape_10"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_0 = const()[name = tensor("reduce_prod_0"), val = tensor(1048576)]; + tensor range_1d_0_start_0 = const()[name = tensor("range_1d_0_start_0"), val = tensor(0)]; + tensor range_1d_0_step_0 = const()[name = tensor("range_1d_0_step_0"), val = tensor(1)]; + tensor range_1d_0 = range_1d(end = reduce_prod_0, start = range_1d_0_start_0, step = range_1d_0_step_0)[name = tensor("range_1d_0")]; + tensor reshape_0 = reshape(shape = shape_10, x = range_1d_0)[name = tensor("reshape_0")]; + tensor slice_by_index_0 = slice_by_index(begin = concat_2, begin_mask = new_cache_1_internal_tensor_assign_1_begin_mask_0, end = concat_3, end_mask = new_cache_1_internal_tensor_assign_1_end_mask_0, squeeze_mask = new_cache_1_internal_tensor_assign_1_squeeze_mask_0, stride = new_cache_1_internal_tensor_assign_1_stride_0, x = reshape_0)[name = tensor("slice_by_index_0")]; + tensor reshape_1_shape_0 = const()[name = tensor("reshape_1_shape_0"), val = tensor([-1])]; + tensor reshape_1 = reshape(shape = reshape_1_shape_0, x = slice_by_index_0)[name = tensor("reshape_1")]; + tensor reshape_2_shape_0 = const()[name = tensor("reshape_2_shape_0"), val = tensor([-1])]; + tensor reshape_2_cast_fp16 = reshape(shape = reshape_2_shape_0, x = var_298_cast_fp16)[name = tensor("reshape_2_cast_fp16")]; + tensor reshape_3_shape_0 = const()[name = tensor("reshape_3_shape_0"), val = tensor([-1])]; + tensor reshape_3_cast_fp16 = reshape(shape = reshape_3_shape_0, x = cache0_to_fp16)[name = tensor("reshape_3_cast_fp16")]; + tensor scatter_0_mode_0 = const()[name = tensor("scatter_0_mode_0"), val = tensor("update")]; + tensor scatter_0_axis_0 = const()[name = tensor("scatter_0_axis_0"), val = tensor(0)]; + tensor scatter_0_validate_indices_0 = const()[name = tensor("scatter_0_validate_indices_0"), val = tensor(false)]; + tensor scatter_0_cast_fp16 = scatter(axis = scatter_0_axis_0, data = reshape_3_cast_fp16, indices = reshape_1, mode = scatter_0_mode_0, updates = reshape_2_cast_fp16, validate_indices = scatter_0_validate_indices_0)[name = tensor("scatter_0_cast_fp16")]; + tensor reshape_4_cast_fp16 = reshape(shape = shape_10, x = scatter_0_cast_fp16)[name = tensor("reshape_4_cast_fp16")]; + tensor var_306_begin_0 = const()[name = tensor("op_306_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor var_306_end_0 = const()[name = tensor("op_306_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor var_306_end_mask_0 = const()[name = tensor("op_306_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_306_squeeze_mask_0 = const()[name = tensor("op_306_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor var_306_cast_fp16 = slice_by_index(begin = var_306_begin_0, end = var_306_end_0, end_mask = var_306_end_mask_0, squeeze_mask = var_306_squeeze_mask_0, x = reshape_4_cast_fp16)[name = tensor("op_306_cast_fp16")]; + tensor var_308_axis_0 = const()[name = tensor("op_308_axis_0"), val = tensor(1)]; + tensor var_308_mode_0 = const()[name = tensor("op_308_mode_0"), val = tensor("update")]; + tensor var_308_validate_indices_0 = const()[name = tensor("op_308_validate_indices_0"), val = tensor(false)]; + tensor var_308_cast_fp16 = scatter_along_axis(axis = var_308_axis_0, data = var_306_cast_fp16, indices = write_indices_1, mode = var_308_mode_0, updates = v_1_cast_fp16, validate_indices = var_308_validate_indices_0)[name = tensor("op_308_cast_fp16")]; + tensor concat_4 = const()[name = tensor("concat_4"), val = tensor([1, 0, 0, 0, 0])]; + tensor concat_5 = const()[name = tensor("concat_5"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_1_internal_tensor_assign_2_stride_0 = const()[name = tensor("new_cache_1_internal_tensor_assign_2_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_1_internal_tensor_assign_2_begin_mask_0 = const()[name = tensor("new_cache_1_internal_tensor_assign_2_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_1_internal_tensor_assign_2_end_mask_0 = const()[name = tensor("new_cache_1_internal_tensor_assign_2_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_1_internal_tensor_assign_2_squeeze_mask_0 = const()[name = tensor("new_cache_1_internal_tensor_assign_2_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_11 = const()[name = tensor("shape_11"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_1 = const()[name = tensor("reduce_prod_1"), val = tensor(1048576)]; + tensor range_1d_1_start_0 = const()[name = tensor("range_1d_1_start_0"), val = tensor(0)]; + tensor range_1d_1_step_0 = const()[name = tensor("range_1d_1_step_0"), val = tensor(1)]; + tensor range_1d_1 = range_1d(end = reduce_prod_1, start = range_1d_1_start_0, step = range_1d_1_step_0)[name = tensor("range_1d_1")]; + tensor reshape_5 = reshape(shape = shape_11, x = range_1d_1)[name = tensor("reshape_5")]; + tensor slice_by_index_1 = slice_by_index(begin = concat_4, begin_mask = new_cache_1_internal_tensor_assign_2_begin_mask_0, end = concat_5, end_mask = new_cache_1_internal_tensor_assign_2_end_mask_0, squeeze_mask = new_cache_1_internal_tensor_assign_2_squeeze_mask_0, stride = new_cache_1_internal_tensor_assign_2_stride_0, x = reshape_5)[name = tensor("slice_by_index_1")]; + tensor reshape_6_shape_0 = const()[name = tensor("reshape_6_shape_0"), val = tensor([-1])]; + tensor reshape_6 = reshape(shape = reshape_6_shape_0, x = slice_by_index_1)[name = tensor("reshape_6")]; + tensor reshape_7_shape_0 = const()[name = tensor("reshape_7_shape_0"), val = tensor([-1])]; + tensor reshape_7_cast_fp16 = reshape(shape = reshape_7_shape_0, x = var_308_cast_fp16)[name = tensor("reshape_7_cast_fp16")]; + tensor reshape_8_shape_0 = const()[name = tensor("reshape_8_shape_0"), val = tensor([-1])]; + tensor reshape_8_cast_fp16 = reshape(shape = reshape_8_shape_0, x = reshape_4_cast_fp16)[name = tensor("reshape_8_cast_fp16")]; + tensor scatter_1_mode_0 = const()[name = tensor("scatter_1_mode_0"), val = tensor("update")]; + tensor scatter_1_axis_0 = const()[name = tensor("scatter_1_axis_0"), val = tensor(0)]; + tensor scatter_1_validate_indices_0 = const()[name = tensor("scatter_1_validate_indices_0"), val = tensor(false)]; + tensor scatter_1_cast_fp16 = scatter(axis = scatter_1_axis_0, data = reshape_8_cast_fp16, indices = reshape_6, mode = scatter_1_mode_0, updates = reshape_7_cast_fp16, validate_indices = scatter_1_validate_indices_0)[name = tensor("scatter_1_cast_fp16")]; + tensor reshape_9_cast_fp16 = reshape(shape = shape_11, x = scatter_1_cast_fp16)[name = tensor("reshape_9_cast_fp16")]; + tensor reshape_9_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("reshape_9_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor keys_1_begin_0 = const()[name = tensor("keys_1_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor keys_1_end_0 = const()[name = tensor("keys_1_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor keys_1_end_mask_0 = const()[name = tensor("keys_1_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor keys_1_squeeze_mask_0 = const()[name = tensor("keys_1_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor keys_1_cast_fp16 = slice_by_index(begin = keys_1_begin_0, end = keys_1_end_0, end_mask = keys_1_end_mask_0, squeeze_mask = keys_1_squeeze_mask_0, x = reshape_9_cast_fp16)[name = tensor("keys_1_cast_fp16")]; + tensor values_1_begin_0 = const()[name = tensor("values_1_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor values_1_end_0 = const()[name = tensor("values_1_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor values_1_end_mask_0 = const()[name = tensor("values_1_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor values_1_squeeze_mask_0 = const()[name = tensor("values_1_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor values_1_cast_fp16 = slice_by_index(begin = values_1_begin_0, end = values_1_end_0, end_mask = values_1_end_mask_0, squeeze_mask = values_1_squeeze_mask_0, x = reshape_9_cast_fp16)[name = tensor("values_1_cast_fp16")]; + tensor var_320_cast_fp16 = not_equal(x = keys_1_cast_fp16, y = keys_1_cast_fp16)[name = tensor("op_320_cast_fp16")]; + tensor var_326_to_fp16 = const()[name = tensor("op_326_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(6303296)))]; + tensor keys_3_cast_fp16 = select(a = var_326_to_fp16, b = keys_1_cast_fp16, cond = var_320_cast_fp16)[name = tensor("keys_3_cast_fp16")]; + tensor var_328_cast_fp16 = not_equal(x = values_1_cast_fp16, y = values_1_cast_fp16)[name = tensor("op_328_cast_fp16")]; + tensor values_3_cast_fp16 = select(a = var_326_to_fp16, b = values_1_cast_fp16, cond = var_328_cast_fp16)[name = tensor("values_3_cast_fp16")]; + tensor var_352 = const()[name = tensor("op_352"), val = tensor([0, 2, 1, 3])]; + tensor var_368 = const()[name = tensor("op_368"), val = tensor([1, 1, 1])]; + tensor var_369_cast_fp16 = reshape(shape = var_368, x = position0_to_fp16)[name = tensor("op_369_cast_fp16")]; + tensor q_offsets_1_promoted_to_fp16 = const()[name = tensor("q_offsets_1_promoted_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(7351936)))]; + tensor q_positions_1_cast_fp16 = add(x = var_369_cast_fp16, y = q_offsets_1_promoted_to_fp16)[name = tensor("q_positions_1_cast_fp16")]; + tensor var_391 = const()[name = tensor("op_391"), val = tensor([1, 1, 1])]; + tensor var_392_cast_fp16 = reshape(shape = var_391, x = valid_len_to_fp16)[name = tensor("op_392_cast_fp16")]; + tensor valid_end_1_cast_fp16 = add(x = var_369_cast_fp16, y = var_392_cast_fp16)[name = tensor("valid_end_1_cast_fp16")]; + tensor k_positions_1_promoted_to_fp16 = const()[name = tensor("k_positions_1_promoted_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(7352512)))]; + tensor valid_mask_1_cast_fp16 = less(x = k_positions_1_promoted_to_fp16, y = valid_end_1_cast_fp16)[name = tensor("valid_mask_1_cast_fp16")]; + tensor causal_mask_1_cast_fp16 = less_equal(x = k_positions_1_promoted_to_fp16, y = q_positions_1_cast_fp16)[name = tensor("causal_mask_1_cast_fp16")]; + tensor var_397 = logical_and(x = valid_mask_1_cast_fp16, y = causal_mask_1_cast_fp16)[name = tensor("op_397")]; + tensor attn_mask_1_axes_0 = const()[name = tensor("attn_mask_1_axes_0"), val = tensor([1])]; + tensor attn_mask_1 = expand_dims(axes = attn_mask_1_axes_0, x = var_397)[name = tensor("attn_mask_1")]; + tensor var_411_transpose_x_0 = const()[name = tensor("op_411_transpose_x_0"), val = tensor(false)]; + tensor var_411_transpose_y_0 = const()[name = tensor("op_411_transpose_y_0"), val = tensor(false)]; + tensor transpose_15_perm_0 = const()[name = tensor("transpose_15_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_16_perm_0 = const()[name = tensor("transpose_16_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_16 = transpose(perm = transpose_16_perm_0, x = keys_3_cast_fp16)[name = tensor("transpose_42")]; + tensor transpose_15 = transpose(perm = transpose_15_perm_0, x = q_3_cast_fp16)[name = tensor("transpose_43")]; + tensor var_411_cast_fp16 = matmul(transpose_x = var_411_transpose_x_0, transpose_y = var_411_transpose_y_0, x = transpose_15, y = transpose_16)[name = tensor("op_411_cast_fp16")]; + tensor var_405_to_fp16 = const()[name = tensor("op_405_to_fp16"), val = tensor([0x1p-3])]; + tensor attn_weights_1_cast_fp16 = mul(x = var_411_cast_fp16, y = var_405_to_fp16)[name = tensor("attn_weights_1_cast_fp16")]; + tensor var_413 = logical_not(x = attn_mask_1)[name = tensor("op_413")]; + tensor var_414_to_fp16 = const()[name = tensor("op_414_to_fp16"), val = tensor(-inf)]; + tensor attn_weights_3_cast_fp16 = select(a = var_414_to_fp16, b = attn_weights_1_cast_fp16, cond = var_413)[name = tensor("attn_weights_3_cast_fp16")]; + tensor var_416 = const()[name = tensor("op_416"), val = tensor(-1)]; + tensor attn_weights_5_cast_fp16 = softmax(axis = var_416, x = attn_weights_3_cast_fp16)[name = tensor("attn_weights_5_cast_fp16")]; + tensor attn_output_1_transpose_x_0 = const()[name = tensor("attn_output_1_transpose_x_0"), val = tensor(false)]; + tensor attn_output_1_transpose_y_0 = const()[name = tensor("attn_output_1_transpose_y_0"), val = tensor(false)]; + tensor values_5_cast_fp16 = transpose(perm = var_352, x = values_3_cast_fp16)[name = tensor("transpose_44")]; + tensor attn_output_1_cast_fp16 = matmul(transpose_x = attn_output_1_transpose_x_0, transpose_y = attn_output_1_transpose_y_0, x = attn_weights_5_cast_fp16, y = values_5_cast_fp16)[name = tensor("attn_output_1_cast_fp16")]; + tensor var_424 = const()[name = tensor("op_424"), val = tensor([0, 2, 1, 3])]; + tensor var_429 = const()[name = tensor("op_429"), val = tensor([1, 256, 1024])]; + tensor var_425_cast_fp16 = transpose(perm = var_424, x = attn_output_1_cast_fp16)[name = tensor("transpose_41")]; + tensor input_5_cast_fp16 = reshape(shape = var_429, x = var_425_cast_fp16)[name = tensor("input_5_cast_fp16")]; + tensor attn0_out_proj_weight_to_fp16 = const()[name = tensor("attn0_out_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(7353600)))]; + tensor linear_1_bias_0_to_fp16 = const()[name = tensor("linear_1_bias_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9450816)))]; + tensor linear_1_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = attn0_out_proj_weight_to_fp16, x = input_5_cast_fp16)[name = tensor("linear_1_cast_fp16")]; + tensor var_438_cast_fp16 = add(x = position0_to_fp16, y = valid_len_to_fp16)[name = tensor("op_438_cast_fp16")]; + tensor var_438_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_438_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor input_7_cast_fp16 = add(x = conditioning_to_fp16, y = linear_1_cast_fp16)[name = tensor("input_7_cast_fp16")]; + tensor input_9_axes_0 = const()[name = tensor("input_9_axes_0"), val = tensor([-1])]; + tensor norm0_2_weight_to_fp16 = const()[name = tensor("norm0_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9452928)))]; + tensor norm0_2_bias_to_fp16 = const()[name = tensor("norm0_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9455040)))]; + tensor var_442_to_fp16 = const()[name = tensor("op_442_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_9_cast_fp16 = layer_norm(axes = input_9_axes_0, beta = norm0_2_bias_to_fp16, epsilon = var_442_to_fp16, gamma = norm0_2_weight_to_fp16, x = input_7_cast_fp16)[name = tensor("input_9_cast_fp16")]; + tensor linear0_1_weight_to_fp16 = const()[name = tensor("linear0_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9457152)))]; + tensor linear_2_bias_0_to_fp16 = const()[name = tensor("linear_2_bias_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(17845824)))]; + tensor linear_2_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = linear0_1_weight_to_fp16, x = input_9_cast_fp16)[name = tensor("linear_2_cast_fp16")]; + tensor input_11_mode_0 = const()[name = tensor("input_11_mode_0"), val = tensor("EXACT")]; + tensor input_11_cast_fp16 = gelu(mode = input_11_mode_0, x = linear_2_cast_fp16)[name = tensor("input_11_cast_fp16")]; + tensor linear0_2_weight_to_fp16 = const()[name = tensor("linear0_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(17854080)))]; + tensor linear_3_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = linear0_2_weight_to_fp16, x = input_11_cast_fp16)[name = tensor("linear_3_cast_fp16")]; + tensor input_13_cast_fp16 = add(x = input_7_cast_fp16, y = linear_3_cast_fp16)[name = tensor("input_13_cast_fp16")]; + tensor input_15_axes_0 = const()[name = tensor("input_15_axes_0"), val = tensor([-1])]; + tensor norm1_1_weight_to_fp16 = const()[name = tensor("norm1_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(26242752)))]; + tensor norm1_1_bias_to_fp16 = const()[name = tensor("norm1_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(26244864)))]; + tensor var_459_to_fp16 = const()[name = tensor("op_459_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_15_cast_fp16 = layer_norm(axes = input_15_axes_0, beta = norm1_1_bias_to_fp16, epsilon = var_459_to_fp16, gamma = norm1_1_weight_to_fp16, x = input_13_cast_fp16)[name = tensor("input_15_cast_fp16")]; + tensor attn1_in_proj_weight_to_fp16 = const()[name = tensor("attn1_in_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(26246976)))]; + tensor linear_4_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = attn1_in_proj_weight_to_fp16, x = input_15_cast_fp16)[name = tensor("linear_4_cast_fp16")]; + tensor var_473 = const()[name = tensor("op_473"), val = tensor([1, 256, 3, 16, 64])]; + tensor qkv_3_cast_fp16 = reshape(shape = var_473, x = linear_4_cast_fp16)[name = tensor("qkv_3_cast_fp16")]; + tensor q_7_begin_0 = const()[name = tensor("q_7_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor q_7_end_0 = const()[name = tensor("q_7_end_0"), val = tensor([1, 256, 1, 16, 64])]; + tensor q_7_end_mask_0 = const()[name = tensor("q_7_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor q_7_squeeze_mask_0 = const()[name = tensor("q_7_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor q_7_cast_fp16 = slice_by_index(begin = q_7_begin_0, end = q_7_end_0, end_mask = q_7_end_mask_0, squeeze_mask = q_7_squeeze_mask_0, x = qkv_3_cast_fp16)[name = tensor("q_7_cast_fp16")]; + tensor k_5_begin_0 = const()[name = tensor("k_5_begin_0"), val = tensor([0, 0, 1, 0, 0])]; + tensor k_5_end_0 = const()[name = tensor("k_5_end_0"), val = tensor([1, 256, 2, 16, 64])]; + tensor k_5_end_mask_0 = const()[name = tensor("k_5_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor k_5_squeeze_mask_0 = const()[name = tensor("k_5_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor k_5_cast_fp16 = slice_by_index(begin = k_5_begin_0, end = k_5_end_0, end_mask = k_5_end_mask_0, squeeze_mask = k_5_squeeze_mask_0, x = qkv_3_cast_fp16)[name = tensor("k_5_cast_fp16")]; + tensor v_3_begin_0 = const()[name = tensor("v_3_begin_0"), val = tensor([0, 0, 2, 0, 0])]; + tensor v_3_end_0 = const()[name = tensor("v_3_end_0"), val = tensor([1, 256, 3, 16, 64])]; + tensor v_3_end_mask_0 = const()[name = tensor("v_3_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor v_3_squeeze_mask_0 = const()[name = tensor("v_3_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor v_3_cast_fp16 = slice_by_index(begin = v_3_begin_0, end = v_3_end_0, end_mask = v_3_end_mask_0, squeeze_mask = v_3_squeeze_mask_0, x = qkv_3_cast_fp16)[name = tensor("v_3_cast_fp16")]; + tensor var_531 = const()[name = tensor("op_531"), val = tensor([1, 1])]; + tensor position1_to_fp16_dtype_0 = const()[name = tensor("position1_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor position1_to_fp16 = cast(dtype = position1_to_fp16_dtype_0, x = position1)[name = tensor("cast_152")]; + tensor var_532_cast_fp16 = reshape(shape = var_531, x = position1_to_fp16)[name = tensor("op_532_cast_fp16")]; + tensor ts_9_cast_fp16 = add(x = ts_1_promoted_to_fp16, y = var_532_cast_fp16)[name = tensor("ts_9_cast_fp16")]; + tensor var_539 = const()[name = tensor("op_539"), val = tensor([1, 256, 1, 1])]; + tensor ts_11_cast_fp16 = reshape(shape = var_539, x = ts_9_cast_fp16)[name = tensor("ts_11_cast_fp16")]; + tensor var_546 = const()[name = tensor("op_546"), val = tensor([1, 256, 16, 32, 2])]; + tensor q_complex_3_cast_fp16 = reshape(shape = var_546, x = q_7_cast_fp16)[name = tensor("q_complex_3_cast_fp16")]; + tensor var_553 = const()[name = tensor("op_553"), val = tensor([1, 256, 16, 32, 2])]; + tensor k_complex_3_cast_fp16 = reshape(shape = var_553, x = k_5_cast_fp16)[name = tensor("k_complex_3_cast_fp16")]; + tensor var_557_begin_0 = const()[name = tensor("op_557_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_557_end_0 = const()[name = tensor("op_557_end_0"), val = tensor([1, 256, 16, 32, 1])]; + tensor var_557_end_mask_0 = const()[name = tensor("op_557_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_557_squeeze_mask_0 = const()[name = tensor("op_557_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_557_cast_fp16 = slice_by_index(begin = var_557_begin_0, end = var_557_end_0, end_mask = var_557_end_mask_0, squeeze_mask = var_557_squeeze_mask_0, x = q_complex_3_cast_fp16)[name = tensor("op_557_cast_fp16")]; + tensor var_565_begin_0 = const()[name = tensor("op_565_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_565_end_0 = const()[name = tensor("op_565_end_0"), val = tensor([1, 256, 16, 32, 2])]; + tensor var_565_end_mask_0 = const()[name = tensor("op_565_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_565_squeeze_mask_0 = const()[name = tensor("op_565_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_565_cast_fp16 = slice_by_index(begin = var_565_begin_0, end = var_565_end_0, end_mask = var_565_end_mask_0, squeeze_mask = var_565_squeeze_mask_0, x = q_complex_3_cast_fp16)[name = tensor("op_565_cast_fp16")]; + tensor var_573_begin_0 = const()[name = tensor("op_573_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_573_end_0 = const()[name = tensor("op_573_end_0"), val = tensor([1, 256, 16, 32, 1])]; + tensor var_573_end_mask_0 = const()[name = tensor("op_573_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_573_squeeze_mask_0 = const()[name = tensor("op_573_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_573_cast_fp16 = slice_by_index(begin = var_573_begin_0, end = var_573_end_0, end_mask = var_573_end_mask_0, squeeze_mask = var_573_squeeze_mask_0, x = k_complex_3_cast_fp16)[name = tensor("op_573_cast_fp16")]; + tensor var_581_begin_0 = const()[name = tensor("op_581_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_581_end_0 = const()[name = tensor("op_581_end_0"), val = tensor([1, 256, 16, 32, 2])]; + tensor var_581_end_mask_0 = const()[name = tensor("op_581_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_581_squeeze_mask_0 = const()[name = tensor("op_581_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_581_cast_fp16 = slice_by_index(begin = var_581_begin_0, end = var_581_end_0, end_mask = var_581_end_mask_0, squeeze_mask = var_581_squeeze_mask_0, x = k_complex_3_cast_fp16)[name = tensor("op_581_cast_fp16")]; + tensor freqs_3_to_fp16 = const()[name = tensor("freqs_3_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(32538496)))]; + tensor var_587_cast_fp16 = mul(x = freqs_3_to_fp16, y = ts_11_cast_fp16)[name = tensor("op_587_cast_fp16")]; + tensor rotr_3_cast_fp16 = cos(x = var_587_cast_fp16)[name = tensor("rotr_3_cast_fp16")]; + tensor roti_3_cast_fp16 = sin(x = var_587_cast_fp16)[name = tensor("roti_3_cast_fp16")]; + tensor var_591_cast_fp16 = mul(x = var_557_cast_fp16, y = rotr_3_cast_fp16)[name = tensor("op_591_cast_fp16")]; + tensor var_592_cast_fp16 = mul(x = var_565_cast_fp16, y = roti_3_cast_fp16)[name = tensor("op_592_cast_fp16")]; + tensor qor_5_cast_fp16 = sub(x = var_591_cast_fp16, y = var_592_cast_fp16)[name = tensor("qor_5_cast_fp16")]; + tensor var_595_cast_fp16 = mul(x = var_557_cast_fp16, y = roti_3_cast_fp16)[name = tensor("op_595_cast_fp16")]; + tensor var_596_cast_fp16 = mul(x = var_565_cast_fp16, y = rotr_3_cast_fp16)[name = tensor("op_596_cast_fp16")]; + tensor qoi_5_cast_fp16 = add(x = var_595_cast_fp16, y = var_596_cast_fp16)[name = tensor("qoi_5_cast_fp16")]; + tensor var_599_cast_fp16 = mul(x = var_573_cast_fp16, y = rotr_3_cast_fp16)[name = tensor("op_599_cast_fp16")]; + tensor var_600_cast_fp16 = mul(x = var_581_cast_fp16, y = roti_3_cast_fp16)[name = tensor("op_600_cast_fp16")]; + tensor kor_5_cast_fp16 = sub(x = var_599_cast_fp16, y = var_600_cast_fp16)[name = tensor("kor_5_cast_fp16")]; + tensor var_603_cast_fp16 = mul(x = var_573_cast_fp16, y = roti_3_cast_fp16)[name = tensor("op_603_cast_fp16")]; + tensor var_604_cast_fp16 = mul(x = var_581_cast_fp16, y = rotr_3_cast_fp16)[name = tensor("op_604_cast_fp16")]; + tensor koi_5_cast_fp16 = add(x = var_603_cast_fp16, y = var_604_cast_fp16)[name = tensor("koi_5_cast_fp16")]; + tensor qo_3_axis_0 = const()[name = tensor("qo_3_axis_0"), val = tensor(-1)]; + tensor qo_3_cast_fp16 = stack(axis = qo_3_axis_0, values = (qor_5_cast_fp16, qoi_5_cast_fp16))[name = tensor("qo_3_cast_fp16")]; + tensor ko_3_axis_0 = const()[name = tensor("ko_3_axis_0"), val = tensor(-1)]; + tensor ko_3_cast_fp16 = stack(axis = ko_3_axis_0, values = (kor_5_cast_fp16, koi_5_cast_fp16))[name = tensor("ko_3_cast_fp16")]; + tensor var_637 = const()[name = tensor("op_637"), val = tensor([1, 256, 16, 64])]; + tensor q_9_cast_fp16 = reshape(shape = var_637, x = qo_3_cast_fp16)[name = tensor("q_9_cast_fp16")]; + tensor var_643 = const()[name = tensor("op_643"), val = tensor([1, 256, 16, 64])]; + tensor k_7_cast_fp16 = reshape(shape = var_643, x = ko_3_cast_fp16)[name = tensor("k_7_cast_fp16")]; + tensor write_indices_float_7_cast_fp16 = add(x = var_532_cast_fp16, y = var_249_promoted_to_fp16)[name = tensor("write_indices_float_7_cast_fp16")]; + tensor _inversed_664_y_0_to_fp16 = const()[name = tensor("_inversed_664_y_0_to_fp16"), val = tensor(0x1p-9)]; + tensor _inversed_664_cast_fp16 = mul(x = write_indices_float_7_cast_fp16, y = _inversed_664_y_0_to_fp16)[name = tensor("_inversed_664_cast_fp16")]; + tensor var_665_cast_fp16 = floor(x = _inversed_664_cast_fp16)[name = tensor("op_665_cast_fp16")]; + tensor var_666_to_fp16 = const()[name = tensor("op_666_to_fp16"), val = tensor(0x1p+9)]; + tensor var_667_cast_fp16 = mul(x = var_665_cast_fp16, y = var_666_to_fp16)[name = tensor("op_667_cast_fp16")]; + tensor write_indices_float_9_cast_fp16 = sub(x = write_indices_float_7_cast_fp16, y = var_667_cast_fp16)[name = tensor("write_indices_float_9_cast_fp16")]; + tensor var_685_value_0_to_fp16 = const()[name = tensor("op_685_value_0_to_fp16"), val = tensor(0x1.ffp+8)]; + tensor var_685_cast_fp16 = fill_like(ref_tensor = write_indices_float_9_cast_fp16, value = var_685_value_0_to_fp16)[name = tensor("op_685_cast_fp16")]; + tensor write_indices_float_11_cast_fp16 = select(a = write_indices_float_9_cast_fp16, b = var_685_cast_fp16, cond = token_valid_1_cast_fp16)[name = tensor("write_indices_float_11_cast_fp16")]; + tensor var_691_dtype_0 = const()[name = tensor("op_691_dtype_0"), val = tensor("int32")]; + tensor var_696 = const()[name = tensor("op_696"), val = tensor([1, 256, 1, 1])]; + tensor write_indices_float_11_cast_fp16_to_int32 = cast(dtype = var_691_dtype_0, x = write_indices_float_11_cast_fp16)[name = tensor("cast_151")]; + tensor var_697 = reshape(shape = var_696, x = write_indices_float_11_cast_fp16_to_int32)[name = tensor("op_697")]; + tensor write_indices_3_reps_0 = const()[name = tensor("write_indices_3_reps_0"), val = tensor([1, 1, 16, 64])]; + tensor write_indices_3 = tile(reps = write_indices_3_reps_0, x = var_697)[name = tensor("write_indices_3")]; + tensor var_707_begin_0 = const()[name = tensor("op_707_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_707_end_0 = const()[name = tensor("op_707_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor var_707_end_mask_0 = const()[name = tensor("op_707_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_707_squeeze_mask_0 = const()[name = tensor("op_707_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor cache1_to_fp16_dtype_0 = const()[name = tensor("cache1_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor cache1_to_fp16 = cast(dtype = cache1_to_fp16_dtype_0, x = cache1)[name = tensor("cast_150")]; + tensor var_707_cast_fp16 = slice_by_index(begin = var_707_begin_0, end = var_707_end_0, end_mask = var_707_end_mask_0, squeeze_mask = var_707_squeeze_mask_0, x = cache1_to_fp16)[name = tensor("op_707_cast_fp16")]; + tensor var_709_axis_0 = const()[name = tensor("op_709_axis_0"), val = tensor(1)]; + tensor var_709_mode_0 = const()[name = tensor("op_709_mode_0"), val = tensor("update")]; + tensor var_709_validate_indices_0 = const()[name = tensor("op_709_validate_indices_0"), val = tensor(false)]; + tensor var_709_cast_fp16 = scatter_along_axis(axis = var_709_axis_0, data = var_707_cast_fp16, indices = write_indices_3, mode = var_709_mode_0, updates = k_7_cast_fp16, validate_indices = var_709_validate_indices_0)[name = tensor("op_709_cast_fp16")]; + tensor concat_10 = const()[name = tensor("concat_10"), val = tensor([0, 0, 0, 0, 0])]; + tensor concat_11 = const()[name = tensor("concat_11"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_3_internal_tensor_assign_1_stride_0 = const()[name = tensor("new_cache_3_internal_tensor_assign_1_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_3_internal_tensor_assign_1_begin_mask_0 = const()[name = tensor("new_cache_3_internal_tensor_assign_1_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_3_internal_tensor_assign_1_end_mask_0 = const()[name = tensor("new_cache_3_internal_tensor_assign_1_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_3_internal_tensor_assign_1_squeeze_mask_0 = const()[name = tensor("new_cache_3_internal_tensor_assign_1_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_12 = const()[name = tensor("shape_12"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_2 = const()[name = tensor("reduce_prod_2"), val = tensor(1048576)]; + tensor range_1d_2_start_0 = const()[name = tensor("range_1d_2_start_0"), val = tensor(0)]; + tensor range_1d_2_step_0 = const()[name = tensor("range_1d_2_step_0"), val = tensor(1)]; + tensor range_1d_2 = range_1d(end = reduce_prod_2, start = range_1d_2_start_0, step = range_1d_2_step_0)[name = tensor("range_1d_2")]; + tensor reshape_10 = reshape(shape = shape_12, x = range_1d_2)[name = tensor("reshape_10")]; + tensor slice_by_index_2 = slice_by_index(begin = concat_10, begin_mask = new_cache_3_internal_tensor_assign_1_begin_mask_0, end = concat_11, end_mask = new_cache_3_internal_tensor_assign_1_end_mask_0, squeeze_mask = new_cache_3_internal_tensor_assign_1_squeeze_mask_0, stride = new_cache_3_internal_tensor_assign_1_stride_0, x = reshape_10)[name = tensor("slice_by_index_2")]; + tensor reshape_11_shape_0 = const()[name = tensor("reshape_11_shape_0"), val = tensor([-1])]; + tensor reshape_11 = reshape(shape = reshape_11_shape_0, x = slice_by_index_2)[name = tensor("reshape_11")]; + tensor reshape_12_shape_0 = const()[name = tensor("reshape_12_shape_0"), val = tensor([-1])]; + tensor reshape_12_cast_fp16 = reshape(shape = reshape_12_shape_0, x = var_709_cast_fp16)[name = tensor("reshape_12_cast_fp16")]; + tensor reshape_13_shape_0 = const()[name = tensor("reshape_13_shape_0"), val = tensor([-1])]; + tensor reshape_13_cast_fp16 = reshape(shape = reshape_13_shape_0, x = cache1_to_fp16)[name = tensor("reshape_13_cast_fp16")]; + tensor scatter_2_mode_0 = const()[name = tensor("scatter_2_mode_0"), val = tensor("update")]; + tensor scatter_2_axis_0 = const()[name = tensor("scatter_2_axis_0"), val = tensor(0)]; + tensor scatter_2_validate_indices_0 = const()[name = tensor("scatter_2_validate_indices_0"), val = tensor(false)]; + tensor scatter_2_cast_fp16 = scatter(axis = scatter_2_axis_0, data = reshape_13_cast_fp16, indices = reshape_11, mode = scatter_2_mode_0, updates = reshape_12_cast_fp16, validate_indices = scatter_2_validate_indices_0)[name = tensor("scatter_2_cast_fp16")]; + tensor reshape_14_cast_fp16 = reshape(shape = shape_12, x = scatter_2_cast_fp16)[name = tensor("reshape_14_cast_fp16")]; + tensor var_717_begin_0 = const()[name = tensor("op_717_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor var_717_end_0 = const()[name = tensor("op_717_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor var_717_end_mask_0 = const()[name = tensor("op_717_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_717_squeeze_mask_0 = const()[name = tensor("op_717_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor var_717_cast_fp16 = slice_by_index(begin = var_717_begin_0, end = var_717_end_0, end_mask = var_717_end_mask_0, squeeze_mask = var_717_squeeze_mask_0, x = reshape_14_cast_fp16)[name = tensor("op_717_cast_fp16")]; + tensor var_719_axis_0 = const()[name = tensor("op_719_axis_0"), val = tensor(1)]; + tensor var_719_mode_0 = const()[name = tensor("op_719_mode_0"), val = tensor("update")]; + tensor var_719_validate_indices_0 = const()[name = tensor("op_719_validate_indices_0"), val = tensor(false)]; + tensor var_719_cast_fp16 = scatter_along_axis(axis = var_719_axis_0, data = var_717_cast_fp16, indices = write_indices_3, mode = var_719_mode_0, updates = v_3_cast_fp16, validate_indices = var_719_validate_indices_0)[name = tensor("op_719_cast_fp16")]; + tensor concat_12 = const()[name = tensor("concat_12"), val = tensor([1, 0, 0, 0, 0])]; + tensor concat_13 = const()[name = tensor("concat_13"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_3_internal_tensor_assign_2_stride_0 = const()[name = tensor("new_cache_3_internal_tensor_assign_2_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_3_internal_tensor_assign_2_begin_mask_0 = const()[name = tensor("new_cache_3_internal_tensor_assign_2_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_3_internal_tensor_assign_2_end_mask_0 = const()[name = tensor("new_cache_3_internal_tensor_assign_2_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_3_internal_tensor_assign_2_squeeze_mask_0 = const()[name = tensor("new_cache_3_internal_tensor_assign_2_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_13 = const()[name = tensor("shape_13"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_3 = const()[name = tensor("reduce_prod_3"), val = tensor(1048576)]; + tensor range_1d_3_start_0 = const()[name = tensor("range_1d_3_start_0"), val = tensor(0)]; + tensor range_1d_3_step_0 = const()[name = tensor("range_1d_3_step_0"), val = tensor(1)]; + tensor range_1d_3 = range_1d(end = reduce_prod_3, start = range_1d_3_start_0, step = range_1d_3_step_0)[name = tensor("range_1d_3")]; + tensor reshape_15 = reshape(shape = shape_13, x = range_1d_3)[name = tensor("reshape_15")]; + tensor slice_by_index_3 = slice_by_index(begin = concat_12, begin_mask = new_cache_3_internal_tensor_assign_2_begin_mask_0, end = concat_13, end_mask = new_cache_3_internal_tensor_assign_2_end_mask_0, squeeze_mask = new_cache_3_internal_tensor_assign_2_squeeze_mask_0, stride = new_cache_3_internal_tensor_assign_2_stride_0, x = reshape_15)[name = tensor("slice_by_index_3")]; + tensor reshape_16_shape_0 = const()[name = tensor("reshape_16_shape_0"), val = tensor([-1])]; + tensor reshape_16 = reshape(shape = reshape_16_shape_0, x = slice_by_index_3)[name = tensor("reshape_16")]; + tensor reshape_17_shape_0 = const()[name = tensor("reshape_17_shape_0"), val = tensor([-1])]; + tensor reshape_17_cast_fp16 = reshape(shape = reshape_17_shape_0, x = var_719_cast_fp16)[name = tensor("reshape_17_cast_fp16")]; + tensor reshape_18_shape_0 = const()[name = tensor("reshape_18_shape_0"), val = tensor([-1])]; + tensor reshape_18_cast_fp16 = reshape(shape = reshape_18_shape_0, x = reshape_14_cast_fp16)[name = tensor("reshape_18_cast_fp16")]; + tensor scatter_3_mode_0 = const()[name = tensor("scatter_3_mode_0"), val = tensor("update")]; + tensor scatter_3_axis_0 = const()[name = tensor("scatter_3_axis_0"), val = tensor(0)]; + tensor scatter_3_validate_indices_0 = const()[name = tensor("scatter_3_validate_indices_0"), val = tensor(false)]; + tensor scatter_3_cast_fp16 = scatter(axis = scatter_3_axis_0, data = reshape_18_cast_fp16, indices = reshape_16, mode = scatter_3_mode_0, updates = reshape_17_cast_fp16, validate_indices = scatter_3_validate_indices_0)[name = tensor("scatter_3_cast_fp16")]; + tensor reshape_19_cast_fp16 = reshape(shape = shape_13, x = scatter_3_cast_fp16)[name = tensor("reshape_19_cast_fp16")]; + tensor reshape_19_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("reshape_19_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor keys_7_begin_0 = const()[name = tensor("keys_7_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor keys_7_end_0 = const()[name = tensor("keys_7_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor keys_7_end_mask_0 = const()[name = tensor("keys_7_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor keys_7_squeeze_mask_0 = const()[name = tensor("keys_7_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor keys_7_cast_fp16 = slice_by_index(begin = keys_7_begin_0, end = keys_7_end_0, end_mask = keys_7_end_mask_0, squeeze_mask = keys_7_squeeze_mask_0, x = reshape_19_cast_fp16)[name = tensor("keys_7_cast_fp16")]; + tensor values_7_begin_0 = const()[name = tensor("values_7_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor values_7_end_0 = const()[name = tensor("values_7_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor values_7_end_mask_0 = const()[name = tensor("values_7_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor values_7_squeeze_mask_0 = const()[name = tensor("values_7_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor values_7_cast_fp16 = slice_by_index(begin = values_7_begin_0, end = values_7_end_0, end_mask = values_7_end_mask_0, squeeze_mask = values_7_squeeze_mask_0, x = reshape_19_cast_fp16)[name = tensor("values_7_cast_fp16")]; + tensor var_731_cast_fp16 = not_equal(x = keys_7_cast_fp16, y = keys_7_cast_fp16)[name = tensor("op_731_cast_fp16")]; + tensor keys_9_cast_fp16 = select(a = var_326_to_fp16, b = keys_7_cast_fp16, cond = var_731_cast_fp16)[name = tensor("keys_9_cast_fp16")]; + tensor var_739_cast_fp16 = not_equal(x = values_7_cast_fp16, y = values_7_cast_fp16)[name = tensor("op_739_cast_fp16")]; + tensor values_9_cast_fp16 = select(a = var_326_to_fp16, b = values_7_cast_fp16, cond = var_739_cast_fp16)[name = tensor("values_9_cast_fp16")]; + tensor var_763 = const()[name = tensor("op_763"), val = tensor([0, 2, 1, 3])]; + tensor var_779 = const()[name = tensor("op_779"), val = tensor([1, 1, 1])]; + tensor var_780_cast_fp16 = reshape(shape = var_779, x = position1_to_fp16)[name = tensor("op_780_cast_fp16")]; + tensor q_positions_3_cast_fp16 = add(x = var_780_cast_fp16, y = q_offsets_1_promoted_to_fp16)[name = tensor("q_positions_3_cast_fp16")]; + tensor valid_end_3_cast_fp16 = add(x = var_780_cast_fp16, y = var_392_cast_fp16)[name = tensor("valid_end_3_cast_fp16")]; + tensor valid_mask_3_cast_fp16 = less(x = k_positions_1_promoted_to_fp16, y = valid_end_3_cast_fp16)[name = tensor("valid_mask_3_cast_fp16")]; + tensor causal_mask_3_cast_fp16 = less_equal(x = k_positions_1_promoted_to_fp16, y = q_positions_3_cast_fp16)[name = tensor("causal_mask_3_cast_fp16")]; + tensor var_808 = logical_and(x = valid_mask_3_cast_fp16, y = causal_mask_3_cast_fp16)[name = tensor("op_808")]; + tensor attn_mask_3_axes_0 = const()[name = tensor("attn_mask_3_axes_0"), val = tensor([1])]; + tensor attn_mask_3 = expand_dims(axes = attn_mask_3_axes_0, x = var_808)[name = tensor("attn_mask_3")]; + tensor var_822_transpose_x_0 = const()[name = tensor("op_822_transpose_x_0"), val = tensor(false)]; + tensor var_822_transpose_y_0 = const()[name = tensor("op_822_transpose_y_0"), val = tensor(false)]; + tensor transpose_17_perm_0 = const()[name = tensor("transpose_17_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_18_perm_0 = const()[name = tensor("transpose_18_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_18 = transpose(perm = transpose_18_perm_0, x = keys_9_cast_fp16)[name = tensor("transpose_38")]; + tensor transpose_17 = transpose(perm = transpose_17_perm_0, x = q_9_cast_fp16)[name = tensor("transpose_39")]; + tensor var_822_cast_fp16 = matmul(transpose_x = var_822_transpose_x_0, transpose_y = var_822_transpose_y_0, x = transpose_17, y = transpose_18)[name = tensor("op_822_cast_fp16")]; + tensor var_816_to_fp16 = const()[name = tensor("op_816_to_fp16"), val = tensor([0x1p-3])]; + tensor attn_weights_7_cast_fp16 = mul(x = var_822_cast_fp16, y = var_816_to_fp16)[name = tensor("attn_weights_7_cast_fp16")]; + tensor var_824 = logical_not(x = attn_mask_3)[name = tensor("op_824")]; + tensor var_825_to_fp16 = const()[name = tensor("op_825_to_fp16"), val = tensor(-inf)]; + tensor attn_weights_9_cast_fp16 = select(a = var_825_to_fp16, b = attn_weights_7_cast_fp16, cond = var_824)[name = tensor("attn_weights_9_cast_fp16")]; + tensor var_827 = const()[name = tensor("op_827"), val = tensor(-1)]; + tensor attn_weights_11_cast_fp16 = softmax(axis = var_827, x = attn_weights_9_cast_fp16)[name = tensor("attn_weights_11_cast_fp16")]; + tensor attn_output_3_transpose_x_0 = const()[name = tensor("attn_output_3_transpose_x_0"), val = tensor(false)]; + tensor attn_output_3_transpose_y_0 = const()[name = tensor("attn_output_3_transpose_y_0"), val = tensor(false)]; + tensor values_11_cast_fp16 = transpose(perm = var_763, x = values_9_cast_fp16)[name = tensor("transpose_40")]; + tensor attn_output_3_cast_fp16 = matmul(transpose_x = attn_output_3_transpose_x_0, transpose_y = attn_output_3_transpose_y_0, x = attn_weights_11_cast_fp16, y = values_11_cast_fp16)[name = tensor("attn_output_3_cast_fp16")]; + tensor var_835 = const()[name = tensor("op_835"), val = tensor([0, 2, 1, 3])]; + tensor var_840 = const()[name = tensor("op_840"), val = tensor([1, 256, 1024])]; + tensor var_836_cast_fp16 = transpose(perm = var_835, x = attn_output_3_cast_fp16)[name = tensor("transpose_37")]; + tensor input_17_cast_fp16 = reshape(shape = var_840, x = var_836_cast_fp16)[name = tensor("input_17_cast_fp16")]; + tensor attn1_out_proj_weight_to_fp16 = const()[name = tensor("attn1_out_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(32538624)))]; + tensor linear_5_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = attn1_out_proj_weight_to_fp16, x = input_17_cast_fp16)[name = tensor("linear_5_cast_fp16")]; + tensor var_849_cast_fp16 = add(x = position1_to_fp16, y = valid_len_to_fp16)[name = tensor("op_849_cast_fp16")]; + tensor var_849_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_849_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor input_19_cast_fp16 = add(x = input_13_cast_fp16, y = linear_5_cast_fp16)[name = tensor("input_19_cast_fp16")]; + tensor input_21_axes_0 = const()[name = tensor("input_21_axes_0"), val = tensor([-1])]; + tensor norm1_2_weight_to_fp16 = const()[name = tensor("norm1_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(34635840)))]; + tensor norm1_2_bias_to_fp16 = const()[name = tensor("norm1_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(34637952)))]; + tensor var_853_to_fp16 = const()[name = tensor("op_853_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_21_cast_fp16 = layer_norm(axes = input_21_axes_0, beta = norm1_2_bias_to_fp16, epsilon = var_853_to_fp16, gamma = norm1_2_weight_to_fp16, x = input_19_cast_fp16)[name = tensor("input_21_cast_fp16")]; + tensor linear1_1_weight_to_fp16 = const()[name = tensor("linear1_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(34640064)))]; + tensor linear_6_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = linear1_1_weight_to_fp16, x = input_21_cast_fp16)[name = tensor("linear_6_cast_fp16")]; + tensor input_23_mode_0 = const()[name = tensor("input_23_mode_0"), val = tensor("EXACT")]; + tensor input_23_cast_fp16 = gelu(mode = input_23_mode_0, x = linear_6_cast_fp16)[name = tensor("input_23_cast_fp16")]; + tensor linear1_2_weight_to_fp16 = const()[name = tensor("linear1_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(43028736)))]; + tensor linear_7_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = linear1_2_weight_to_fp16, x = input_23_cast_fp16)[name = tensor("linear_7_cast_fp16")]; + tensor input_25_cast_fp16 = add(x = input_19_cast_fp16, y = linear_7_cast_fp16)[name = tensor("input_25_cast_fp16")]; + tensor input_27_axes_0 = const()[name = tensor("input_27_axes_0"), val = tensor([-1])]; + tensor norm2_1_weight_to_fp16 = const()[name = tensor("norm2_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(51417408)))]; + tensor norm2_1_bias_to_fp16 = const()[name = tensor("norm2_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(51419520)))]; + tensor var_870_to_fp16 = const()[name = tensor("op_870_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_27_cast_fp16 = layer_norm(axes = input_27_axes_0, beta = norm2_1_bias_to_fp16, epsilon = var_870_to_fp16, gamma = norm2_1_weight_to_fp16, x = input_25_cast_fp16)[name = tensor("input_27_cast_fp16")]; + tensor attn2_in_proj_weight_to_fp16 = const()[name = tensor("attn2_in_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(51421632)))]; + tensor linear_8_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = attn2_in_proj_weight_to_fp16, x = input_27_cast_fp16)[name = tensor("linear_8_cast_fp16")]; + tensor var_884 = const()[name = tensor("op_884"), val = tensor([1, 256, 3, 16, 64])]; + tensor qkv_5_cast_fp16 = reshape(shape = var_884, x = linear_8_cast_fp16)[name = tensor("qkv_5_cast_fp16")]; + tensor q_13_begin_0 = const()[name = tensor("q_13_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor q_13_end_0 = const()[name = tensor("q_13_end_0"), val = tensor([1, 256, 1, 16, 64])]; + tensor q_13_end_mask_0 = const()[name = tensor("q_13_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor q_13_squeeze_mask_0 = const()[name = tensor("q_13_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor q_13_cast_fp16 = slice_by_index(begin = q_13_begin_0, end = q_13_end_0, end_mask = q_13_end_mask_0, squeeze_mask = q_13_squeeze_mask_0, x = qkv_5_cast_fp16)[name = tensor("q_13_cast_fp16")]; + tensor k_9_begin_0 = const()[name = tensor("k_9_begin_0"), val = tensor([0, 0, 1, 0, 0])]; + tensor k_9_end_0 = const()[name = tensor("k_9_end_0"), val = tensor([1, 256, 2, 16, 64])]; + tensor k_9_end_mask_0 = const()[name = tensor("k_9_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor k_9_squeeze_mask_0 = const()[name = tensor("k_9_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor k_9_cast_fp16 = slice_by_index(begin = k_9_begin_0, end = k_9_end_0, end_mask = k_9_end_mask_0, squeeze_mask = k_9_squeeze_mask_0, x = qkv_5_cast_fp16)[name = tensor("k_9_cast_fp16")]; + tensor v_5_begin_0 = const()[name = tensor("v_5_begin_0"), val = tensor([0, 0, 2, 0, 0])]; + tensor v_5_end_0 = const()[name = tensor("v_5_end_0"), val = tensor([1, 256, 3, 16, 64])]; + tensor v_5_end_mask_0 = const()[name = tensor("v_5_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor v_5_squeeze_mask_0 = const()[name = tensor("v_5_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor v_5_cast_fp16 = slice_by_index(begin = v_5_begin_0, end = v_5_end_0, end_mask = v_5_end_mask_0, squeeze_mask = v_5_squeeze_mask_0, x = qkv_5_cast_fp16)[name = tensor("v_5_cast_fp16")]; + tensor var_942 = const()[name = tensor("op_942"), val = tensor([1, 1])]; + tensor position2_to_fp16_dtype_0 = const()[name = tensor("position2_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor position2_to_fp16 = cast(dtype = position2_to_fp16_dtype_0, x = position2)[name = tensor("cast_147")]; + tensor var_943_cast_fp16 = reshape(shape = var_942, x = position2_to_fp16)[name = tensor("op_943_cast_fp16")]; + tensor ts_15_cast_fp16 = add(x = ts_1_promoted_to_fp16, y = var_943_cast_fp16)[name = tensor("ts_15_cast_fp16")]; + tensor var_950 = const()[name = tensor("op_950"), val = tensor([1, 256, 1, 1])]; + tensor ts_17_cast_fp16 = reshape(shape = var_950, x = ts_15_cast_fp16)[name = tensor("ts_17_cast_fp16")]; + tensor var_957 = const()[name = tensor("op_957"), val = tensor([1, 256, 16, 32, 2])]; + tensor q_complex_5_cast_fp16 = reshape(shape = var_957, x = q_13_cast_fp16)[name = tensor("q_complex_5_cast_fp16")]; + tensor var_964 = const()[name = tensor("op_964"), val = tensor([1, 256, 16, 32, 2])]; + tensor k_complex_5_cast_fp16 = reshape(shape = var_964, x = k_9_cast_fp16)[name = tensor("k_complex_5_cast_fp16")]; + tensor var_968_begin_0 = const()[name = tensor("op_968_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_968_end_0 = const()[name = tensor("op_968_end_0"), val = tensor([1, 256, 16, 32, 1])]; + tensor var_968_end_mask_0 = const()[name = tensor("op_968_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_968_squeeze_mask_0 = const()[name = tensor("op_968_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_968_cast_fp16 = slice_by_index(begin = var_968_begin_0, end = var_968_end_0, end_mask = var_968_end_mask_0, squeeze_mask = var_968_squeeze_mask_0, x = q_complex_5_cast_fp16)[name = tensor("op_968_cast_fp16")]; + tensor var_976_begin_0 = const()[name = tensor("op_976_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_976_end_0 = const()[name = tensor("op_976_end_0"), val = tensor([1, 256, 16, 32, 2])]; + tensor var_976_end_mask_0 = const()[name = tensor("op_976_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_976_squeeze_mask_0 = const()[name = tensor("op_976_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_976_cast_fp16 = slice_by_index(begin = var_976_begin_0, end = var_976_end_0, end_mask = var_976_end_mask_0, squeeze_mask = var_976_squeeze_mask_0, x = q_complex_5_cast_fp16)[name = tensor("op_976_cast_fp16")]; + tensor var_984_begin_0 = const()[name = tensor("op_984_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_984_end_0 = const()[name = tensor("op_984_end_0"), val = tensor([1, 256, 16, 32, 1])]; + tensor var_984_end_mask_0 = const()[name = tensor("op_984_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_984_squeeze_mask_0 = const()[name = tensor("op_984_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_984_cast_fp16 = slice_by_index(begin = var_984_begin_0, end = var_984_end_0, end_mask = var_984_end_mask_0, squeeze_mask = var_984_squeeze_mask_0, x = k_complex_5_cast_fp16)[name = tensor("op_984_cast_fp16")]; + tensor var_992_begin_0 = const()[name = tensor("op_992_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_992_end_0 = const()[name = tensor("op_992_end_0"), val = tensor([1, 256, 16, 32, 2])]; + tensor var_992_end_mask_0 = const()[name = tensor("op_992_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_992_squeeze_mask_0 = const()[name = tensor("op_992_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_992_cast_fp16 = slice_by_index(begin = var_992_begin_0, end = var_992_end_0, end_mask = var_992_end_mask_0, squeeze_mask = var_992_squeeze_mask_0, x = k_complex_5_cast_fp16)[name = tensor("op_992_cast_fp16")]; + tensor freqs_5_to_fp16 = const()[name = tensor("freqs_5_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(57713152)))]; + tensor var_998_cast_fp16 = mul(x = freqs_5_to_fp16, y = ts_17_cast_fp16)[name = tensor("op_998_cast_fp16")]; + tensor rotr_5_cast_fp16 = cos(x = var_998_cast_fp16)[name = tensor("rotr_5_cast_fp16")]; + tensor roti_5_cast_fp16 = sin(x = var_998_cast_fp16)[name = tensor("roti_5_cast_fp16")]; + tensor var_1002_cast_fp16 = mul(x = var_968_cast_fp16, y = rotr_5_cast_fp16)[name = tensor("op_1002_cast_fp16")]; + tensor var_1003_cast_fp16 = mul(x = var_976_cast_fp16, y = roti_5_cast_fp16)[name = tensor("op_1003_cast_fp16")]; + tensor qor_9_cast_fp16 = sub(x = var_1002_cast_fp16, y = var_1003_cast_fp16)[name = tensor("qor_9_cast_fp16")]; + tensor var_1006_cast_fp16 = mul(x = var_968_cast_fp16, y = roti_5_cast_fp16)[name = tensor("op_1006_cast_fp16")]; + tensor var_1007_cast_fp16 = mul(x = var_976_cast_fp16, y = rotr_5_cast_fp16)[name = tensor("op_1007_cast_fp16")]; + tensor qoi_9_cast_fp16 = add(x = var_1006_cast_fp16, y = var_1007_cast_fp16)[name = tensor("qoi_9_cast_fp16")]; + tensor var_1010_cast_fp16 = mul(x = var_984_cast_fp16, y = rotr_5_cast_fp16)[name = tensor("op_1010_cast_fp16")]; + tensor var_1011_cast_fp16 = mul(x = var_992_cast_fp16, y = roti_5_cast_fp16)[name = tensor("op_1011_cast_fp16")]; + tensor kor_9_cast_fp16 = sub(x = var_1010_cast_fp16, y = var_1011_cast_fp16)[name = tensor("kor_9_cast_fp16")]; + tensor var_1014_cast_fp16 = mul(x = var_984_cast_fp16, y = roti_5_cast_fp16)[name = tensor("op_1014_cast_fp16")]; + tensor var_1015_cast_fp16 = mul(x = var_992_cast_fp16, y = rotr_5_cast_fp16)[name = tensor("op_1015_cast_fp16")]; + tensor koi_9_cast_fp16 = add(x = var_1014_cast_fp16, y = var_1015_cast_fp16)[name = tensor("koi_9_cast_fp16")]; + tensor qo_5_axis_0 = const()[name = tensor("qo_5_axis_0"), val = tensor(-1)]; + tensor qo_5_cast_fp16 = stack(axis = qo_5_axis_0, values = (qor_9_cast_fp16, qoi_9_cast_fp16))[name = tensor("qo_5_cast_fp16")]; + tensor ko_5_axis_0 = const()[name = tensor("ko_5_axis_0"), val = tensor(-1)]; + tensor ko_5_cast_fp16 = stack(axis = ko_5_axis_0, values = (kor_9_cast_fp16, koi_9_cast_fp16))[name = tensor("ko_5_cast_fp16")]; + tensor var_1048 = const()[name = tensor("op_1048"), val = tensor([1, 256, 16, 64])]; + tensor q_15_cast_fp16 = reshape(shape = var_1048, x = qo_5_cast_fp16)[name = tensor("q_15_cast_fp16")]; + tensor var_1054 = const()[name = tensor("op_1054"), val = tensor([1, 256, 16, 64])]; + tensor k_11_cast_fp16 = reshape(shape = var_1054, x = ko_5_cast_fp16)[name = tensor("k_11_cast_fp16")]; + tensor write_indices_float_13_cast_fp16 = add(x = var_943_cast_fp16, y = var_249_promoted_to_fp16)[name = tensor("write_indices_float_13_cast_fp16")]; + tensor _inversed_1075_y_0_to_fp16 = const()[name = tensor("_inversed_1075_y_0_to_fp16"), val = tensor(0x1p-9)]; + tensor _inversed_1075_cast_fp16 = mul(x = write_indices_float_13_cast_fp16, y = _inversed_1075_y_0_to_fp16)[name = tensor("_inversed_1075_cast_fp16")]; + tensor var_1076_cast_fp16 = floor(x = _inversed_1075_cast_fp16)[name = tensor("op_1076_cast_fp16")]; + tensor var_1077_to_fp16 = const()[name = tensor("op_1077_to_fp16"), val = tensor(0x1p+9)]; + tensor var_1078_cast_fp16 = mul(x = var_1076_cast_fp16, y = var_1077_to_fp16)[name = tensor("op_1078_cast_fp16")]; + tensor write_indices_float_15_cast_fp16 = sub(x = write_indices_float_13_cast_fp16, y = var_1078_cast_fp16)[name = tensor("write_indices_float_15_cast_fp16")]; + tensor var_1096_value_0_to_fp16 = const()[name = tensor("op_1096_value_0_to_fp16"), val = tensor(0x1.ffp+8)]; + tensor var_1096_cast_fp16 = fill_like(ref_tensor = write_indices_float_15_cast_fp16, value = var_1096_value_0_to_fp16)[name = tensor("op_1096_cast_fp16")]; + tensor write_indices_float_17_cast_fp16 = select(a = write_indices_float_15_cast_fp16, b = var_1096_cast_fp16, cond = token_valid_1_cast_fp16)[name = tensor("write_indices_float_17_cast_fp16")]; + tensor var_1102_dtype_0 = const()[name = tensor("op_1102_dtype_0"), val = tensor("int32")]; + tensor var_1107 = const()[name = tensor("op_1107"), val = tensor([1, 256, 1, 1])]; + tensor write_indices_float_17_cast_fp16_to_int32 = cast(dtype = var_1102_dtype_0, x = write_indices_float_17_cast_fp16)[name = tensor("cast_146")]; + tensor var_1108 = reshape(shape = var_1107, x = write_indices_float_17_cast_fp16_to_int32)[name = tensor("op_1108")]; + tensor write_indices_5_reps_0 = const()[name = tensor("write_indices_5_reps_0"), val = tensor([1, 1, 16, 64])]; + tensor write_indices_5 = tile(reps = write_indices_5_reps_0, x = var_1108)[name = tensor("write_indices_5")]; + tensor var_1118_begin_0 = const()[name = tensor("op_1118_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_1118_end_0 = const()[name = tensor("op_1118_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor var_1118_end_mask_0 = const()[name = tensor("op_1118_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_1118_squeeze_mask_0 = const()[name = tensor("op_1118_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor cache2_to_fp16_dtype_0 = const()[name = tensor("cache2_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor cache2_to_fp16 = cast(dtype = cache2_to_fp16_dtype_0, x = cache2)[name = tensor("cast_145")]; + tensor var_1118_cast_fp16 = slice_by_index(begin = var_1118_begin_0, end = var_1118_end_0, end_mask = var_1118_end_mask_0, squeeze_mask = var_1118_squeeze_mask_0, x = cache2_to_fp16)[name = tensor("op_1118_cast_fp16")]; + tensor var_1120_axis_0 = const()[name = tensor("op_1120_axis_0"), val = tensor(1)]; + tensor var_1120_mode_0 = const()[name = tensor("op_1120_mode_0"), val = tensor("update")]; + tensor var_1120_validate_indices_0 = const()[name = tensor("op_1120_validate_indices_0"), val = tensor(false)]; + tensor var_1120_cast_fp16 = scatter_along_axis(axis = var_1120_axis_0, data = var_1118_cast_fp16, indices = write_indices_5, mode = var_1120_mode_0, updates = k_11_cast_fp16, validate_indices = var_1120_validate_indices_0)[name = tensor("op_1120_cast_fp16")]; + tensor concat_18 = const()[name = tensor("concat_18"), val = tensor([0, 0, 0, 0, 0])]; + tensor concat_19 = const()[name = tensor("concat_19"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_5_internal_tensor_assign_1_stride_0 = const()[name = tensor("new_cache_5_internal_tensor_assign_1_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_5_internal_tensor_assign_1_begin_mask_0 = const()[name = tensor("new_cache_5_internal_tensor_assign_1_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_5_internal_tensor_assign_1_end_mask_0 = const()[name = tensor("new_cache_5_internal_tensor_assign_1_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_5_internal_tensor_assign_1_squeeze_mask_0 = const()[name = tensor("new_cache_5_internal_tensor_assign_1_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_14 = const()[name = tensor("shape_14"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_4 = const()[name = tensor("reduce_prod_4"), val = tensor(1048576)]; + tensor range_1d_4_start_0 = const()[name = tensor("range_1d_4_start_0"), val = tensor(0)]; + tensor range_1d_4_step_0 = const()[name = tensor("range_1d_4_step_0"), val = tensor(1)]; + tensor range_1d_4 = range_1d(end = reduce_prod_4, start = range_1d_4_start_0, step = range_1d_4_step_0)[name = tensor("range_1d_4")]; + tensor reshape_20 = reshape(shape = shape_14, x = range_1d_4)[name = tensor("reshape_20")]; + tensor slice_by_index_4 = slice_by_index(begin = concat_18, begin_mask = new_cache_5_internal_tensor_assign_1_begin_mask_0, end = concat_19, end_mask = new_cache_5_internal_tensor_assign_1_end_mask_0, squeeze_mask = new_cache_5_internal_tensor_assign_1_squeeze_mask_0, stride = new_cache_5_internal_tensor_assign_1_stride_0, x = reshape_20)[name = tensor("slice_by_index_4")]; + tensor reshape_21_shape_0 = const()[name = tensor("reshape_21_shape_0"), val = tensor([-1])]; + tensor reshape_21 = reshape(shape = reshape_21_shape_0, x = slice_by_index_4)[name = tensor("reshape_21")]; + tensor reshape_22_shape_0 = const()[name = tensor("reshape_22_shape_0"), val = tensor([-1])]; + tensor reshape_22_cast_fp16 = reshape(shape = reshape_22_shape_0, x = var_1120_cast_fp16)[name = tensor("reshape_22_cast_fp16")]; + tensor reshape_23_shape_0 = const()[name = tensor("reshape_23_shape_0"), val = tensor([-1])]; + tensor reshape_23_cast_fp16 = reshape(shape = reshape_23_shape_0, x = cache2_to_fp16)[name = tensor("reshape_23_cast_fp16")]; + tensor scatter_4_mode_0 = const()[name = tensor("scatter_4_mode_0"), val = tensor("update")]; + tensor scatter_4_axis_0 = const()[name = tensor("scatter_4_axis_0"), val = tensor(0)]; + tensor scatter_4_validate_indices_0 = const()[name = tensor("scatter_4_validate_indices_0"), val = tensor(false)]; + tensor scatter_4_cast_fp16 = scatter(axis = scatter_4_axis_0, data = reshape_23_cast_fp16, indices = reshape_21, mode = scatter_4_mode_0, updates = reshape_22_cast_fp16, validate_indices = scatter_4_validate_indices_0)[name = tensor("scatter_4_cast_fp16")]; + tensor reshape_24_cast_fp16 = reshape(shape = shape_14, x = scatter_4_cast_fp16)[name = tensor("reshape_24_cast_fp16")]; + tensor var_1128_begin_0 = const()[name = tensor("op_1128_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor var_1128_end_0 = const()[name = tensor("op_1128_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor var_1128_end_mask_0 = const()[name = tensor("op_1128_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_1128_squeeze_mask_0 = const()[name = tensor("op_1128_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor var_1128_cast_fp16 = slice_by_index(begin = var_1128_begin_0, end = var_1128_end_0, end_mask = var_1128_end_mask_0, squeeze_mask = var_1128_squeeze_mask_0, x = reshape_24_cast_fp16)[name = tensor("op_1128_cast_fp16")]; + tensor var_1130_axis_0 = const()[name = tensor("op_1130_axis_0"), val = tensor(1)]; + tensor var_1130_mode_0 = const()[name = tensor("op_1130_mode_0"), val = tensor("update")]; + tensor var_1130_validate_indices_0 = const()[name = tensor("op_1130_validate_indices_0"), val = tensor(false)]; + tensor var_1130_cast_fp16 = scatter_along_axis(axis = var_1130_axis_0, data = var_1128_cast_fp16, indices = write_indices_5, mode = var_1130_mode_0, updates = v_5_cast_fp16, validate_indices = var_1130_validate_indices_0)[name = tensor("op_1130_cast_fp16")]; + tensor concat_20 = const()[name = tensor("concat_20"), val = tensor([1, 0, 0, 0, 0])]; + tensor concat_21 = const()[name = tensor("concat_21"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_5_internal_tensor_assign_2_stride_0 = const()[name = tensor("new_cache_5_internal_tensor_assign_2_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_5_internal_tensor_assign_2_begin_mask_0 = const()[name = tensor("new_cache_5_internal_tensor_assign_2_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_5_internal_tensor_assign_2_end_mask_0 = const()[name = tensor("new_cache_5_internal_tensor_assign_2_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_5_internal_tensor_assign_2_squeeze_mask_0 = const()[name = tensor("new_cache_5_internal_tensor_assign_2_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_15 = const()[name = tensor("shape_15"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_5 = const()[name = tensor("reduce_prod_5"), val = tensor(1048576)]; + tensor range_1d_5_start_0 = const()[name = tensor("range_1d_5_start_0"), val = tensor(0)]; + tensor range_1d_5_step_0 = const()[name = tensor("range_1d_5_step_0"), val = tensor(1)]; + tensor range_1d_5 = range_1d(end = reduce_prod_5, start = range_1d_5_start_0, step = range_1d_5_step_0)[name = tensor("range_1d_5")]; + tensor reshape_25 = reshape(shape = shape_15, x = range_1d_5)[name = tensor("reshape_25")]; + tensor slice_by_index_5 = slice_by_index(begin = concat_20, begin_mask = new_cache_5_internal_tensor_assign_2_begin_mask_0, end = concat_21, end_mask = new_cache_5_internal_tensor_assign_2_end_mask_0, squeeze_mask = new_cache_5_internal_tensor_assign_2_squeeze_mask_0, stride = new_cache_5_internal_tensor_assign_2_stride_0, x = reshape_25)[name = tensor("slice_by_index_5")]; + tensor reshape_26_shape_0 = const()[name = tensor("reshape_26_shape_0"), val = tensor([-1])]; + tensor reshape_26 = reshape(shape = reshape_26_shape_0, x = slice_by_index_5)[name = tensor("reshape_26")]; + tensor reshape_27_shape_0 = const()[name = tensor("reshape_27_shape_0"), val = tensor([-1])]; + tensor reshape_27_cast_fp16 = reshape(shape = reshape_27_shape_0, x = var_1130_cast_fp16)[name = tensor("reshape_27_cast_fp16")]; + tensor reshape_28_shape_0 = const()[name = tensor("reshape_28_shape_0"), val = tensor([-1])]; + tensor reshape_28_cast_fp16 = reshape(shape = reshape_28_shape_0, x = reshape_24_cast_fp16)[name = tensor("reshape_28_cast_fp16")]; + tensor scatter_5_mode_0 = const()[name = tensor("scatter_5_mode_0"), val = tensor("update")]; + tensor scatter_5_axis_0 = const()[name = tensor("scatter_5_axis_0"), val = tensor(0)]; + tensor scatter_5_validate_indices_0 = const()[name = tensor("scatter_5_validate_indices_0"), val = tensor(false)]; + tensor scatter_5_cast_fp16 = scatter(axis = scatter_5_axis_0, data = reshape_28_cast_fp16, indices = reshape_26, mode = scatter_5_mode_0, updates = reshape_27_cast_fp16, validate_indices = scatter_5_validate_indices_0)[name = tensor("scatter_5_cast_fp16")]; + tensor reshape_29_cast_fp16 = reshape(shape = shape_15, x = scatter_5_cast_fp16)[name = tensor("reshape_29_cast_fp16")]; + tensor reshape_29_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("reshape_29_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor keys_13_begin_0 = const()[name = tensor("keys_13_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor keys_13_end_0 = const()[name = tensor("keys_13_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor keys_13_end_mask_0 = const()[name = tensor("keys_13_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor keys_13_squeeze_mask_0 = const()[name = tensor("keys_13_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor keys_13_cast_fp16 = slice_by_index(begin = keys_13_begin_0, end = keys_13_end_0, end_mask = keys_13_end_mask_0, squeeze_mask = keys_13_squeeze_mask_0, x = reshape_29_cast_fp16)[name = tensor("keys_13_cast_fp16")]; + tensor values_13_begin_0 = const()[name = tensor("values_13_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor values_13_end_0 = const()[name = tensor("values_13_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor values_13_end_mask_0 = const()[name = tensor("values_13_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor values_13_squeeze_mask_0 = const()[name = tensor("values_13_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor values_13_cast_fp16 = slice_by_index(begin = values_13_begin_0, end = values_13_end_0, end_mask = values_13_end_mask_0, squeeze_mask = values_13_squeeze_mask_0, x = reshape_29_cast_fp16)[name = tensor("values_13_cast_fp16")]; + tensor var_1142_cast_fp16 = not_equal(x = keys_13_cast_fp16, y = keys_13_cast_fp16)[name = tensor("op_1142_cast_fp16")]; + tensor keys_15_cast_fp16 = select(a = var_326_to_fp16, b = keys_13_cast_fp16, cond = var_1142_cast_fp16)[name = tensor("keys_15_cast_fp16")]; + tensor var_1150_cast_fp16 = not_equal(x = values_13_cast_fp16, y = values_13_cast_fp16)[name = tensor("op_1150_cast_fp16")]; + tensor values_15_cast_fp16 = select(a = var_326_to_fp16, b = values_13_cast_fp16, cond = var_1150_cast_fp16)[name = tensor("values_15_cast_fp16")]; + tensor var_1174 = const()[name = tensor("op_1174"), val = tensor([0, 2, 1, 3])]; + tensor var_1190 = const()[name = tensor("op_1190"), val = tensor([1, 1, 1])]; + tensor var_1191_cast_fp16 = reshape(shape = var_1190, x = position2_to_fp16)[name = tensor("op_1191_cast_fp16")]; + tensor q_positions_5_cast_fp16 = add(x = var_1191_cast_fp16, y = q_offsets_1_promoted_to_fp16)[name = tensor("q_positions_5_cast_fp16")]; + tensor valid_end_5_cast_fp16 = add(x = var_1191_cast_fp16, y = var_392_cast_fp16)[name = tensor("valid_end_5_cast_fp16")]; + tensor valid_mask_5_cast_fp16 = less(x = k_positions_1_promoted_to_fp16, y = valid_end_5_cast_fp16)[name = tensor("valid_mask_5_cast_fp16")]; + tensor causal_mask_5_cast_fp16 = less_equal(x = k_positions_1_promoted_to_fp16, y = q_positions_5_cast_fp16)[name = tensor("causal_mask_5_cast_fp16")]; + tensor var_1219 = logical_and(x = valid_mask_5_cast_fp16, y = causal_mask_5_cast_fp16)[name = tensor("op_1219")]; + tensor attn_mask_5_axes_0 = const()[name = tensor("attn_mask_5_axes_0"), val = tensor([1])]; + tensor attn_mask_5 = expand_dims(axes = attn_mask_5_axes_0, x = var_1219)[name = tensor("attn_mask_5")]; + tensor var_1233_transpose_x_0 = const()[name = tensor("op_1233_transpose_x_0"), val = tensor(false)]; + tensor var_1233_transpose_y_0 = const()[name = tensor("op_1233_transpose_y_0"), val = tensor(false)]; + tensor transpose_19_perm_0 = const()[name = tensor("transpose_19_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_20_perm_0 = const()[name = tensor("transpose_20_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_20 = transpose(perm = transpose_20_perm_0, x = keys_15_cast_fp16)[name = tensor("transpose_34")]; + tensor transpose_19 = transpose(perm = transpose_19_perm_0, x = q_15_cast_fp16)[name = tensor("transpose_35")]; + tensor var_1233_cast_fp16 = matmul(transpose_x = var_1233_transpose_x_0, transpose_y = var_1233_transpose_y_0, x = transpose_19, y = transpose_20)[name = tensor("op_1233_cast_fp16")]; + tensor var_1227_to_fp16 = const()[name = tensor("op_1227_to_fp16"), val = tensor([0x1p-3])]; + tensor attn_weights_13_cast_fp16 = mul(x = var_1233_cast_fp16, y = var_1227_to_fp16)[name = tensor("attn_weights_13_cast_fp16")]; + tensor var_1235 = logical_not(x = attn_mask_5)[name = tensor("op_1235")]; + tensor var_1236_to_fp16 = const()[name = tensor("op_1236_to_fp16"), val = tensor(-inf)]; + tensor attn_weights_15_cast_fp16 = select(a = var_1236_to_fp16, b = attn_weights_13_cast_fp16, cond = var_1235)[name = tensor("attn_weights_15_cast_fp16")]; + tensor var_1238 = const()[name = tensor("op_1238"), val = tensor(-1)]; + tensor attn_weights_17_cast_fp16 = softmax(axis = var_1238, x = attn_weights_15_cast_fp16)[name = tensor("attn_weights_17_cast_fp16")]; + tensor attn_output_5_transpose_x_0 = const()[name = tensor("attn_output_5_transpose_x_0"), val = tensor(false)]; + tensor attn_output_5_transpose_y_0 = const()[name = tensor("attn_output_5_transpose_y_0"), val = tensor(false)]; + tensor values_17_cast_fp16 = transpose(perm = var_1174, x = values_15_cast_fp16)[name = tensor("transpose_36")]; + tensor attn_output_5_cast_fp16 = matmul(transpose_x = attn_output_5_transpose_x_0, transpose_y = attn_output_5_transpose_y_0, x = attn_weights_17_cast_fp16, y = values_17_cast_fp16)[name = tensor("attn_output_5_cast_fp16")]; + tensor var_1246 = const()[name = tensor("op_1246"), val = tensor([0, 2, 1, 3])]; + tensor var_1251 = const()[name = tensor("op_1251"), val = tensor([1, 256, 1024])]; + tensor var_1247_cast_fp16 = transpose(perm = var_1246, x = attn_output_5_cast_fp16)[name = tensor("transpose_33")]; + tensor input_29_cast_fp16 = reshape(shape = var_1251, x = var_1247_cast_fp16)[name = tensor("input_29_cast_fp16")]; + tensor attn2_out_proj_weight_to_fp16 = const()[name = tensor("attn2_out_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(57713280)))]; + tensor linear_9_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = attn2_out_proj_weight_to_fp16, x = input_29_cast_fp16)[name = tensor("linear_9_cast_fp16")]; + tensor var_1260_cast_fp16 = add(x = position2_to_fp16, y = valid_len_to_fp16)[name = tensor("op_1260_cast_fp16")]; + tensor var_1260_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_1260_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor input_31_cast_fp16 = add(x = input_25_cast_fp16, y = linear_9_cast_fp16)[name = tensor("input_31_cast_fp16")]; + tensor input_33_axes_0 = const()[name = tensor("input_33_axes_0"), val = tensor([-1])]; + tensor norm2_2_weight_to_fp16 = const()[name = tensor("norm2_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(59810496)))]; + tensor norm2_2_bias_to_fp16 = const()[name = tensor("norm2_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(59812608)))]; + tensor var_1264_to_fp16 = const()[name = tensor("op_1264_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_33_cast_fp16 = layer_norm(axes = input_33_axes_0, beta = norm2_2_bias_to_fp16, epsilon = var_1264_to_fp16, gamma = norm2_2_weight_to_fp16, x = input_31_cast_fp16)[name = tensor("input_33_cast_fp16")]; + tensor linear2_1_weight_to_fp16 = const()[name = tensor("linear2_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(59814720)))]; + tensor linear_10_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = linear2_1_weight_to_fp16, x = input_33_cast_fp16)[name = tensor("linear_10_cast_fp16")]; + tensor input_35_mode_0 = const()[name = tensor("input_35_mode_0"), val = tensor("EXACT")]; + tensor input_35_cast_fp16 = gelu(mode = input_35_mode_0, x = linear_10_cast_fp16)[name = tensor("input_35_cast_fp16")]; + tensor linear2_2_weight_to_fp16 = const()[name = tensor("linear2_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(68203392)))]; + tensor linear_11_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = linear2_2_weight_to_fp16, x = input_35_cast_fp16)[name = tensor("linear_11_cast_fp16")]; + tensor input_37_cast_fp16 = add(x = input_31_cast_fp16, y = linear_11_cast_fp16)[name = tensor("input_37_cast_fp16")]; + tensor input_39_axes_0 = const()[name = tensor("input_39_axes_0"), val = tensor([-1])]; + tensor norm3_1_weight_to_fp16 = const()[name = tensor("norm3_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(76592064)))]; + tensor norm3_1_bias_to_fp16 = const()[name = tensor("norm3_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(76594176)))]; + tensor var_1281_to_fp16 = const()[name = tensor("op_1281_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_39_cast_fp16 = layer_norm(axes = input_39_axes_0, beta = norm3_1_bias_to_fp16, epsilon = var_1281_to_fp16, gamma = norm3_1_weight_to_fp16, x = input_37_cast_fp16)[name = tensor("input_39_cast_fp16")]; + tensor attn3_in_proj_weight_to_fp16 = const()[name = tensor("attn3_in_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(76596288)))]; + tensor linear_12_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = attn3_in_proj_weight_to_fp16, x = input_39_cast_fp16)[name = tensor("linear_12_cast_fp16")]; + tensor var_1295 = const()[name = tensor("op_1295"), val = tensor([1, 256, 3, 16, 64])]; + tensor qkv_7_cast_fp16 = reshape(shape = var_1295, x = linear_12_cast_fp16)[name = tensor("qkv_7_cast_fp16")]; + tensor q_19_begin_0 = const()[name = tensor("q_19_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor q_19_end_0 = const()[name = tensor("q_19_end_0"), val = tensor([1, 256, 1, 16, 64])]; + tensor q_19_end_mask_0 = const()[name = tensor("q_19_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor q_19_squeeze_mask_0 = const()[name = tensor("q_19_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor q_19_cast_fp16 = slice_by_index(begin = q_19_begin_0, end = q_19_end_0, end_mask = q_19_end_mask_0, squeeze_mask = q_19_squeeze_mask_0, x = qkv_7_cast_fp16)[name = tensor("q_19_cast_fp16")]; + tensor k_13_begin_0 = const()[name = tensor("k_13_begin_0"), val = tensor([0, 0, 1, 0, 0])]; + tensor k_13_end_0 = const()[name = tensor("k_13_end_0"), val = tensor([1, 256, 2, 16, 64])]; + tensor k_13_end_mask_0 = const()[name = tensor("k_13_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor k_13_squeeze_mask_0 = const()[name = tensor("k_13_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor k_13_cast_fp16 = slice_by_index(begin = k_13_begin_0, end = k_13_end_0, end_mask = k_13_end_mask_0, squeeze_mask = k_13_squeeze_mask_0, x = qkv_7_cast_fp16)[name = tensor("k_13_cast_fp16")]; + tensor v_7_begin_0 = const()[name = tensor("v_7_begin_0"), val = tensor([0, 0, 2, 0, 0])]; + tensor v_7_end_0 = const()[name = tensor("v_7_end_0"), val = tensor([1, 256, 3, 16, 64])]; + tensor v_7_end_mask_0 = const()[name = tensor("v_7_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor v_7_squeeze_mask_0 = const()[name = tensor("v_7_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor v_7_cast_fp16 = slice_by_index(begin = v_7_begin_0, end = v_7_end_0, end_mask = v_7_end_mask_0, squeeze_mask = v_7_squeeze_mask_0, x = qkv_7_cast_fp16)[name = tensor("v_7_cast_fp16")]; + tensor var_1353 = const()[name = tensor("op_1353"), val = tensor([1, 1])]; + tensor position3_to_fp16_dtype_0 = const()[name = tensor("position3_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor position3_to_fp16 = cast(dtype = position3_to_fp16_dtype_0, x = position3)[name = tensor("cast_142")]; + tensor var_1354_cast_fp16 = reshape(shape = var_1353, x = position3_to_fp16)[name = tensor("op_1354_cast_fp16")]; + tensor ts_21_cast_fp16 = add(x = ts_1_promoted_to_fp16, y = var_1354_cast_fp16)[name = tensor("ts_21_cast_fp16")]; + tensor var_1361 = const()[name = tensor("op_1361"), val = tensor([1, 256, 1, 1])]; + tensor ts_23_cast_fp16 = reshape(shape = var_1361, x = ts_21_cast_fp16)[name = tensor("ts_23_cast_fp16")]; + tensor var_1368 = const()[name = tensor("op_1368"), val = tensor([1, 256, 16, 32, 2])]; + tensor q_complex_7_cast_fp16 = reshape(shape = var_1368, x = q_19_cast_fp16)[name = tensor("q_complex_7_cast_fp16")]; + tensor var_1375 = const()[name = tensor("op_1375"), val = tensor([1, 256, 16, 32, 2])]; + tensor k_complex_7_cast_fp16 = reshape(shape = var_1375, x = k_13_cast_fp16)[name = tensor("k_complex_7_cast_fp16")]; + tensor var_1379_begin_0 = const()[name = tensor("op_1379_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_1379_end_0 = const()[name = tensor("op_1379_end_0"), val = tensor([1, 256, 16, 32, 1])]; + tensor var_1379_end_mask_0 = const()[name = tensor("op_1379_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_1379_squeeze_mask_0 = const()[name = tensor("op_1379_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_1379_cast_fp16 = slice_by_index(begin = var_1379_begin_0, end = var_1379_end_0, end_mask = var_1379_end_mask_0, squeeze_mask = var_1379_squeeze_mask_0, x = q_complex_7_cast_fp16)[name = tensor("op_1379_cast_fp16")]; + tensor var_1387_begin_0 = const()[name = tensor("op_1387_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_1387_end_0 = const()[name = tensor("op_1387_end_0"), val = tensor([1, 256, 16, 32, 2])]; + tensor var_1387_end_mask_0 = const()[name = tensor("op_1387_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_1387_squeeze_mask_0 = const()[name = tensor("op_1387_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_1387_cast_fp16 = slice_by_index(begin = var_1387_begin_0, end = var_1387_end_0, end_mask = var_1387_end_mask_0, squeeze_mask = var_1387_squeeze_mask_0, x = q_complex_7_cast_fp16)[name = tensor("op_1387_cast_fp16")]; + tensor var_1395_begin_0 = const()[name = tensor("op_1395_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_1395_end_0 = const()[name = tensor("op_1395_end_0"), val = tensor([1, 256, 16, 32, 1])]; + tensor var_1395_end_mask_0 = const()[name = tensor("op_1395_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_1395_squeeze_mask_0 = const()[name = tensor("op_1395_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_1395_cast_fp16 = slice_by_index(begin = var_1395_begin_0, end = var_1395_end_0, end_mask = var_1395_end_mask_0, squeeze_mask = var_1395_squeeze_mask_0, x = k_complex_7_cast_fp16)[name = tensor("op_1395_cast_fp16")]; + tensor var_1403_begin_0 = const()[name = tensor("op_1403_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_1403_end_0 = const()[name = tensor("op_1403_end_0"), val = tensor([1, 256, 16, 32, 2])]; + tensor var_1403_end_mask_0 = const()[name = tensor("op_1403_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_1403_squeeze_mask_0 = const()[name = tensor("op_1403_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_1403_cast_fp16 = slice_by_index(begin = var_1403_begin_0, end = var_1403_end_0, end_mask = var_1403_end_mask_0, squeeze_mask = var_1403_squeeze_mask_0, x = k_complex_7_cast_fp16)[name = tensor("op_1403_cast_fp16")]; + tensor freqs_7_to_fp16 = const()[name = tensor("freqs_7_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(82887808)))]; + tensor var_1409_cast_fp16 = mul(x = freqs_7_to_fp16, y = ts_23_cast_fp16)[name = tensor("op_1409_cast_fp16")]; + tensor rotr_7_cast_fp16 = cos(x = var_1409_cast_fp16)[name = tensor("rotr_7_cast_fp16")]; + tensor roti_7_cast_fp16 = sin(x = var_1409_cast_fp16)[name = tensor("roti_7_cast_fp16")]; + tensor var_1413_cast_fp16 = mul(x = var_1379_cast_fp16, y = rotr_7_cast_fp16)[name = tensor("op_1413_cast_fp16")]; + tensor var_1414_cast_fp16 = mul(x = var_1387_cast_fp16, y = roti_7_cast_fp16)[name = tensor("op_1414_cast_fp16")]; + tensor qor_13_cast_fp16 = sub(x = var_1413_cast_fp16, y = var_1414_cast_fp16)[name = tensor("qor_13_cast_fp16")]; + tensor var_1417_cast_fp16 = mul(x = var_1379_cast_fp16, y = roti_7_cast_fp16)[name = tensor("op_1417_cast_fp16")]; + tensor var_1418_cast_fp16 = mul(x = var_1387_cast_fp16, y = rotr_7_cast_fp16)[name = tensor("op_1418_cast_fp16")]; + tensor qoi_13_cast_fp16 = add(x = var_1417_cast_fp16, y = var_1418_cast_fp16)[name = tensor("qoi_13_cast_fp16")]; + tensor var_1421_cast_fp16 = mul(x = var_1395_cast_fp16, y = rotr_7_cast_fp16)[name = tensor("op_1421_cast_fp16")]; + tensor var_1422_cast_fp16 = mul(x = var_1403_cast_fp16, y = roti_7_cast_fp16)[name = tensor("op_1422_cast_fp16")]; + tensor kor_13_cast_fp16 = sub(x = var_1421_cast_fp16, y = var_1422_cast_fp16)[name = tensor("kor_13_cast_fp16")]; + tensor var_1425_cast_fp16 = mul(x = var_1395_cast_fp16, y = roti_7_cast_fp16)[name = tensor("op_1425_cast_fp16")]; + tensor var_1426_cast_fp16 = mul(x = var_1403_cast_fp16, y = rotr_7_cast_fp16)[name = tensor("op_1426_cast_fp16")]; + tensor koi_13_cast_fp16 = add(x = var_1425_cast_fp16, y = var_1426_cast_fp16)[name = tensor("koi_13_cast_fp16")]; + tensor qo_7_axis_0 = const()[name = tensor("qo_7_axis_0"), val = tensor(-1)]; + tensor qo_7_cast_fp16 = stack(axis = qo_7_axis_0, values = (qor_13_cast_fp16, qoi_13_cast_fp16))[name = tensor("qo_7_cast_fp16")]; + tensor ko_7_axis_0 = const()[name = tensor("ko_7_axis_0"), val = tensor(-1)]; + tensor ko_7_cast_fp16 = stack(axis = ko_7_axis_0, values = (kor_13_cast_fp16, koi_13_cast_fp16))[name = tensor("ko_7_cast_fp16")]; + tensor var_1459 = const()[name = tensor("op_1459"), val = tensor([1, 256, 16, 64])]; + tensor q_21_cast_fp16 = reshape(shape = var_1459, x = qo_7_cast_fp16)[name = tensor("q_21_cast_fp16")]; + tensor var_1465 = const()[name = tensor("op_1465"), val = tensor([1, 256, 16, 64])]; + tensor k_15_cast_fp16 = reshape(shape = var_1465, x = ko_7_cast_fp16)[name = tensor("k_15_cast_fp16")]; + tensor write_indices_float_19_cast_fp16 = add(x = var_1354_cast_fp16, y = var_249_promoted_to_fp16)[name = tensor("write_indices_float_19_cast_fp16")]; + tensor _inversed_1486_y_0_to_fp16 = const()[name = tensor("_inversed_1486_y_0_to_fp16"), val = tensor(0x1p-9)]; + tensor _inversed_1486_cast_fp16 = mul(x = write_indices_float_19_cast_fp16, y = _inversed_1486_y_0_to_fp16)[name = tensor("_inversed_1486_cast_fp16")]; + tensor var_1487_cast_fp16 = floor(x = _inversed_1486_cast_fp16)[name = tensor("op_1487_cast_fp16")]; + tensor var_1488_to_fp16 = const()[name = tensor("op_1488_to_fp16"), val = tensor(0x1p+9)]; + tensor var_1489_cast_fp16 = mul(x = var_1487_cast_fp16, y = var_1488_to_fp16)[name = tensor("op_1489_cast_fp16")]; + tensor write_indices_float_21_cast_fp16 = sub(x = write_indices_float_19_cast_fp16, y = var_1489_cast_fp16)[name = tensor("write_indices_float_21_cast_fp16")]; + tensor var_1507_value_0_to_fp16 = const()[name = tensor("op_1507_value_0_to_fp16"), val = tensor(0x1.ffp+8)]; + tensor var_1507_cast_fp16 = fill_like(ref_tensor = write_indices_float_21_cast_fp16, value = var_1507_value_0_to_fp16)[name = tensor("op_1507_cast_fp16")]; + tensor write_indices_float_23_cast_fp16 = select(a = write_indices_float_21_cast_fp16, b = var_1507_cast_fp16, cond = token_valid_1_cast_fp16)[name = tensor("write_indices_float_23_cast_fp16")]; + tensor var_1513_dtype_0 = const()[name = tensor("op_1513_dtype_0"), val = tensor("int32")]; + tensor var_1518 = const()[name = tensor("op_1518"), val = tensor([1, 256, 1, 1])]; + tensor write_indices_float_23_cast_fp16_to_int32 = cast(dtype = var_1513_dtype_0, x = write_indices_float_23_cast_fp16)[name = tensor("cast_141")]; + tensor var_1519 = reshape(shape = var_1518, x = write_indices_float_23_cast_fp16_to_int32)[name = tensor("op_1519")]; + tensor write_indices_7_reps_0 = const()[name = tensor("write_indices_7_reps_0"), val = tensor([1, 1, 16, 64])]; + tensor write_indices_7 = tile(reps = write_indices_7_reps_0, x = var_1519)[name = tensor("write_indices_7")]; + tensor var_1529_begin_0 = const()[name = tensor("op_1529_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_1529_end_0 = const()[name = tensor("op_1529_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor var_1529_end_mask_0 = const()[name = tensor("op_1529_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_1529_squeeze_mask_0 = const()[name = tensor("op_1529_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor cache3_to_fp16_dtype_0 = const()[name = tensor("cache3_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor cache3_to_fp16 = cast(dtype = cache3_to_fp16_dtype_0, x = cache3)[name = tensor("cast_140")]; + tensor var_1529_cast_fp16 = slice_by_index(begin = var_1529_begin_0, end = var_1529_end_0, end_mask = var_1529_end_mask_0, squeeze_mask = var_1529_squeeze_mask_0, x = cache3_to_fp16)[name = tensor("op_1529_cast_fp16")]; + tensor var_1531_axis_0 = const()[name = tensor("op_1531_axis_0"), val = tensor(1)]; + tensor var_1531_mode_0 = const()[name = tensor("op_1531_mode_0"), val = tensor("update")]; + tensor var_1531_validate_indices_0 = const()[name = tensor("op_1531_validate_indices_0"), val = tensor(false)]; + tensor var_1531_cast_fp16 = scatter_along_axis(axis = var_1531_axis_0, data = var_1529_cast_fp16, indices = write_indices_7, mode = var_1531_mode_0, updates = k_15_cast_fp16, validate_indices = var_1531_validate_indices_0)[name = tensor("op_1531_cast_fp16")]; + tensor concat_26 = const()[name = tensor("concat_26"), val = tensor([0, 0, 0, 0, 0])]; + tensor concat_27 = const()[name = tensor("concat_27"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_7_internal_tensor_assign_1_stride_0 = const()[name = tensor("new_cache_7_internal_tensor_assign_1_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_7_internal_tensor_assign_1_begin_mask_0 = const()[name = tensor("new_cache_7_internal_tensor_assign_1_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_7_internal_tensor_assign_1_end_mask_0 = const()[name = tensor("new_cache_7_internal_tensor_assign_1_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_7_internal_tensor_assign_1_squeeze_mask_0 = const()[name = tensor("new_cache_7_internal_tensor_assign_1_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_16 = const()[name = tensor("shape_16"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_6 = const()[name = tensor("reduce_prod_6"), val = tensor(1048576)]; + tensor range_1d_6_start_0 = const()[name = tensor("range_1d_6_start_0"), val = tensor(0)]; + tensor range_1d_6_step_0 = const()[name = tensor("range_1d_6_step_0"), val = tensor(1)]; + tensor range_1d_6 = range_1d(end = reduce_prod_6, start = range_1d_6_start_0, step = range_1d_6_step_0)[name = tensor("range_1d_6")]; + tensor reshape_30 = reshape(shape = shape_16, x = range_1d_6)[name = tensor("reshape_30")]; + tensor slice_by_index_6 = slice_by_index(begin = concat_26, begin_mask = new_cache_7_internal_tensor_assign_1_begin_mask_0, end = concat_27, end_mask = new_cache_7_internal_tensor_assign_1_end_mask_0, squeeze_mask = new_cache_7_internal_tensor_assign_1_squeeze_mask_0, stride = new_cache_7_internal_tensor_assign_1_stride_0, x = reshape_30)[name = tensor("slice_by_index_6")]; + tensor reshape_31_shape_0 = const()[name = tensor("reshape_31_shape_0"), val = tensor([-1])]; + tensor reshape_31 = reshape(shape = reshape_31_shape_0, x = slice_by_index_6)[name = tensor("reshape_31")]; + tensor reshape_32_shape_0 = const()[name = tensor("reshape_32_shape_0"), val = tensor([-1])]; + tensor reshape_32_cast_fp16 = reshape(shape = reshape_32_shape_0, x = var_1531_cast_fp16)[name = tensor("reshape_32_cast_fp16")]; + tensor reshape_33_shape_0 = const()[name = tensor("reshape_33_shape_0"), val = tensor([-1])]; + tensor reshape_33_cast_fp16 = reshape(shape = reshape_33_shape_0, x = cache3_to_fp16)[name = tensor("reshape_33_cast_fp16")]; + tensor scatter_6_mode_0 = const()[name = tensor("scatter_6_mode_0"), val = tensor("update")]; + tensor scatter_6_axis_0 = const()[name = tensor("scatter_6_axis_0"), val = tensor(0)]; + tensor scatter_6_validate_indices_0 = const()[name = tensor("scatter_6_validate_indices_0"), val = tensor(false)]; + tensor scatter_6_cast_fp16 = scatter(axis = scatter_6_axis_0, data = reshape_33_cast_fp16, indices = reshape_31, mode = scatter_6_mode_0, updates = reshape_32_cast_fp16, validate_indices = scatter_6_validate_indices_0)[name = tensor("scatter_6_cast_fp16")]; + tensor reshape_34_cast_fp16 = reshape(shape = shape_16, x = scatter_6_cast_fp16)[name = tensor("reshape_34_cast_fp16")]; + tensor var_1539_begin_0 = const()[name = tensor("op_1539_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor var_1539_end_0 = const()[name = tensor("op_1539_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor var_1539_end_mask_0 = const()[name = tensor("op_1539_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_1539_squeeze_mask_0 = const()[name = tensor("op_1539_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor var_1539_cast_fp16 = slice_by_index(begin = var_1539_begin_0, end = var_1539_end_0, end_mask = var_1539_end_mask_0, squeeze_mask = var_1539_squeeze_mask_0, x = reshape_34_cast_fp16)[name = tensor("op_1539_cast_fp16")]; + tensor var_1541_axis_0 = const()[name = tensor("op_1541_axis_0"), val = tensor(1)]; + tensor var_1541_mode_0 = const()[name = tensor("op_1541_mode_0"), val = tensor("update")]; + tensor var_1541_validate_indices_0 = const()[name = tensor("op_1541_validate_indices_0"), val = tensor(false)]; + tensor var_1541_cast_fp16 = scatter_along_axis(axis = var_1541_axis_0, data = var_1539_cast_fp16, indices = write_indices_7, mode = var_1541_mode_0, updates = v_7_cast_fp16, validate_indices = var_1541_validate_indices_0)[name = tensor("op_1541_cast_fp16")]; + tensor concat_28 = const()[name = tensor("concat_28"), val = tensor([1, 0, 0, 0, 0])]; + tensor concat_29 = const()[name = tensor("concat_29"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_7_internal_tensor_assign_2_stride_0 = const()[name = tensor("new_cache_7_internal_tensor_assign_2_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_7_internal_tensor_assign_2_begin_mask_0 = const()[name = tensor("new_cache_7_internal_tensor_assign_2_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_7_internal_tensor_assign_2_end_mask_0 = const()[name = tensor("new_cache_7_internal_tensor_assign_2_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_7_internal_tensor_assign_2_squeeze_mask_0 = const()[name = tensor("new_cache_7_internal_tensor_assign_2_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_17 = const()[name = tensor("shape_17"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_7 = const()[name = tensor("reduce_prod_7"), val = tensor(1048576)]; + tensor range_1d_7_start_0 = const()[name = tensor("range_1d_7_start_0"), val = tensor(0)]; + tensor range_1d_7_step_0 = const()[name = tensor("range_1d_7_step_0"), val = tensor(1)]; + tensor range_1d_7 = range_1d(end = reduce_prod_7, start = range_1d_7_start_0, step = range_1d_7_step_0)[name = tensor("range_1d_7")]; + tensor reshape_35 = reshape(shape = shape_17, x = range_1d_7)[name = tensor("reshape_35")]; + tensor slice_by_index_7 = slice_by_index(begin = concat_28, begin_mask = new_cache_7_internal_tensor_assign_2_begin_mask_0, end = concat_29, end_mask = new_cache_7_internal_tensor_assign_2_end_mask_0, squeeze_mask = new_cache_7_internal_tensor_assign_2_squeeze_mask_0, stride = new_cache_7_internal_tensor_assign_2_stride_0, x = reshape_35)[name = tensor("slice_by_index_7")]; + tensor reshape_36_shape_0 = const()[name = tensor("reshape_36_shape_0"), val = tensor([-1])]; + tensor reshape_36 = reshape(shape = reshape_36_shape_0, x = slice_by_index_7)[name = tensor("reshape_36")]; + tensor reshape_37_shape_0 = const()[name = tensor("reshape_37_shape_0"), val = tensor([-1])]; + tensor reshape_37_cast_fp16 = reshape(shape = reshape_37_shape_0, x = var_1541_cast_fp16)[name = tensor("reshape_37_cast_fp16")]; + tensor reshape_38_shape_0 = const()[name = tensor("reshape_38_shape_0"), val = tensor([-1])]; + tensor reshape_38_cast_fp16 = reshape(shape = reshape_38_shape_0, x = reshape_34_cast_fp16)[name = tensor("reshape_38_cast_fp16")]; + tensor scatter_7_mode_0 = const()[name = tensor("scatter_7_mode_0"), val = tensor("update")]; + tensor scatter_7_axis_0 = const()[name = tensor("scatter_7_axis_0"), val = tensor(0)]; + tensor scatter_7_validate_indices_0 = const()[name = tensor("scatter_7_validate_indices_0"), val = tensor(false)]; + tensor scatter_7_cast_fp16 = scatter(axis = scatter_7_axis_0, data = reshape_38_cast_fp16, indices = reshape_36, mode = scatter_7_mode_0, updates = reshape_37_cast_fp16, validate_indices = scatter_7_validate_indices_0)[name = tensor("scatter_7_cast_fp16")]; + tensor reshape_39_cast_fp16 = reshape(shape = shape_17, x = scatter_7_cast_fp16)[name = tensor("reshape_39_cast_fp16")]; + tensor reshape_39_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("reshape_39_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor keys_19_begin_0 = const()[name = tensor("keys_19_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor keys_19_end_0 = const()[name = tensor("keys_19_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor keys_19_end_mask_0 = const()[name = tensor("keys_19_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor keys_19_squeeze_mask_0 = const()[name = tensor("keys_19_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor keys_19_cast_fp16 = slice_by_index(begin = keys_19_begin_0, end = keys_19_end_0, end_mask = keys_19_end_mask_0, squeeze_mask = keys_19_squeeze_mask_0, x = reshape_39_cast_fp16)[name = tensor("keys_19_cast_fp16")]; + tensor values_19_begin_0 = const()[name = tensor("values_19_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor values_19_end_0 = const()[name = tensor("values_19_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor values_19_end_mask_0 = const()[name = tensor("values_19_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor values_19_squeeze_mask_0 = const()[name = tensor("values_19_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor values_19_cast_fp16 = slice_by_index(begin = values_19_begin_0, end = values_19_end_0, end_mask = values_19_end_mask_0, squeeze_mask = values_19_squeeze_mask_0, x = reshape_39_cast_fp16)[name = tensor("values_19_cast_fp16")]; + tensor var_1553_cast_fp16 = not_equal(x = keys_19_cast_fp16, y = keys_19_cast_fp16)[name = tensor("op_1553_cast_fp16")]; + tensor keys_21_cast_fp16 = select(a = var_326_to_fp16, b = keys_19_cast_fp16, cond = var_1553_cast_fp16)[name = tensor("keys_21_cast_fp16")]; + tensor var_1561_cast_fp16 = not_equal(x = values_19_cast_fp16, y = values_19_cast_fp16)[name = tensor("op_1561_cast_fp16")]; + tensor values_21_cast_fp16 = select(a = var_326_to_fp16, b = values_19_cast_fp16, cond = var_1561_cast_fp16)[name = tensor("values_21_cast_fp16")]; + tensor var_1585 = const()[name = tensor("op_1585"), val = tensor([0, 2, 1, 3])]; + tensor var_1601 = const()[name = tensor("op_1601"), val = tensor([1, 1, 1])]; + tensor var_1602_cast_fp16 = reshape(shape = var_1601, x = position3_to_fp16)[name = tensor("op_1602_cast_fp16")]; + tensor q_positions_7_cast_fp16 = add(x = var_1602_cast_fp16, y = q_offsets_1_promoted_to_fp16)[name = tensor("q_positions_7_cast_fp16")]; + tensor valid_end_7_cast_fp16 = add(x = var_1602_cast_fp16, y = var_392_cast_fp16)[name = tensor("valid_end_7_cast_fp16")]; + tensor valid_mask_7_cast_fp16 = less(x = k_positions_1_promoted_to_fp16, y = valid_end_7_cast_fp16)[name = tensor("valid_mask_7_cast_fp16")]; + tensor causal_mask_7_cast_fp16 = less_equal(x = k_positions_1_promoted_to_fp16, y = q_positions_7_cast_fp16)[name = tensor("causal_mask_7_cast_fp16")]; + tensor var_1630 = logical_and(x = valid_mask_7_cast_fp16, y = causal_mask_7_cast_fp16)[name = tensor("op_1630")]; + tensor attn_mask_7_axes_0 = const()[name = tensor("attn_mask_7_axes_0"), val = tensor([1])]; + tensor attn_mask_7 = expand_dims(axes = attn_mask_7_axes_0, x = var_1630)[name = tensor("attn_mask_7")]; + tensor var_1644_transpose_x_0 = const()[name = tensor("op_1644_transpose_x_0"), val = tensor(false)]; + tensor var_1644_transpose_y_0 = const()[name = tensor("op_1644_transpose_y_0"), val = tensor(false)]; + tensor transpose_21_perm_0 = const()[name = tensor("transpose_21_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_22_perm_0 = const()[name = tensor("transpose_22_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_22 = transpose(perm = transpose_22_perm_0, x = keys_21_cast_fp16)[name = tensor("transpose_30")]; + tensor transpose_21 = transpose(perm = transpose_21_perm_0, x = q_21_cast_fp16)[name = tensor("transpose_31")]; + tensor var_1644_cast_fp16 = matmul(transpose_x = var_1644_transpose_x_0, transpose_y = var_1644_transpose_y_0, x = transpose_21, y = transpose_22)[name = tensor("op_1644_cast_fp16")]; + tensor var_1638_to_fp16 = const()[name = tensor("op_1638_to_fp16"), val = tensor([0x1p-3])]; + tensor attn_weights_19_cast_fp16 = mul(x = var_1644_cast_fp16, y = var_1638_to_fp16)[name = tensor("attn_weights_19_cast_fp16")]; + tensor var_1646 = logical_not(x = attn_mask_7)[name = tensor("op_1646")]; + tensor var_1647_to_fp16 = const()[name = tensor("op_1647_to_fp16"), val = tensor(-inf)]; + tensor attn_weights_21_cast_fp16 = select(a = var_1647_to_fp16, b = attn_weights_19_cast_fp16, cond = var_1646)[name = tensor("attn_weights_21_cast_fp16")]; + tensor var_1649 = const()[name = tensor("op_1649"), val = tensor(-1)]; + tensor attn_weights_23_cast_fp16 = softmax(axis = var_1649, x = attn_weights_21_cast_fp16)[name = tensor("attn_weights_23_cast_fp16")]; + tensor attn_output_7_transpose_x_0 = const()[name = tensor("attn_output_7_transpose_x_0"), val = tensor(false)]; + tensor attn_output_7_transpose_y_0 = const()[name = tensor("attn_output_7_transpose_y_0"), val = tensor(false)]; + tensor values_23_cast_fp16 = transpose(perm = var_1585, x = values_21_cast_fp16)[name = tensor("transpose_32")]; + tensor attn_output_7_cast_fp16 = matmul(transpose_x = attn_output_7_transpose_x_0, transpose_y = attn_output_7_transpose_y_0, x = attn_weights_23_cast_fp16, y = values_23_cast_fp16)[name = tensor("attn_output_7_cast_fp16")]; + tensor var_1657 = const()[name = tensor("op_1657"), val = tensor([0, 2, 1, 3])]; + tensor var_1662 = const()[name = tensor("op_1662"), val = tensor([1, 256, 1024])]; + tensor var_1658_cast_fp16 = transpose(perm = var_1657, x = attn_output_7_cast_fp16)[name = tensor("transpose_29")]; + tensor input_41_cast_fp16 = reshape(shape = var_1662, x = var_1658_cast_fp16)[name = tensor("input_41_cast_fp16")]; + tensor attn3_out_proj_weight_to_fp16 = const()[name = tensor("attn3_out_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(82887936)))]; + tensor linear_13_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = attn3_out_proj_weight_to_fp16, x = input_41_cast_fp16)[name = tensor("linear_13_cast_fp16")]; + tensor var_1671_cast_fp16 = add(x = position3_to_fp16, y = valid_len_to_fp16)[name = tensor("op_1671_cast_fp16")]; + tensor var_1671_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_1671_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor input_43_cast_fp16 = add(x = input_37_cast_fp16, y = linear_13_cast_fp16)[name = tensor("input_43_cast_fp16")]; + tensor input_45_axes_0 = const()[name = tensor("input_45_axes_0"), val = tensor([-1])]; + tensor norm3_2_weight_to_fp16 = const()[name = tensor("norm3_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(84985152)))]; + tensor norm3_2_bias_to_fp16 = const()[name = tensor("norm3_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(84987264)))]; + tensor var_1675_to_fp16 = const()[name = tensor("op_1675_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_45_cast_fp16 = layer_norm(axes = input_45_axes_0, beta = norm3_2_bias_to_fp16, epsilon = var_1675_to_fp16, gamma = norm3_2_weight_to_fp16, x = input_43_cast_fp16)[name = tensor("input_45_cast_fp16")]; + tensor linear3_1_weight_to_fp16 = const()[name = tensor("linear3_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(84989376)))]; + tensor linear_14_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = linear3_1_weight_to_fp16, x = input_45_cast_fp16)[name = tensor("linear_14_cast_fp16")]; + tensor input_47_mode_0 = const()[name = tensor("input_47_mode_0"), val = tensor("EXACT")]; + tensor input_47_cast_fp16 = gelu(mode = input_47_mode_0, x = linear_14_cast_fp16)[name = tensor("input_47_cast_fp16")]; + tensor linear3_2_weight_to_fp16 = const()[name = tensor("linear3_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(93378048)))]; + tensor linear_15_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = linear3_2_weight_to_fp16, x = input_47_cast_fp16)[name = tensor("linear_15_cast_fp16")]; + tensor input_49_cast_fp16 = add(x = input_43_cast_fp16, y = linear_15_cast_fp16)[name = tensor("input_49_cast_fp16")]; + tensor input_51_axes_0 = const()[name = tensor("input_51_axes_0"), val = tensor([-1])]; + tensor norm4_1_weight_to_fp16 = const()[name = tensor("norm4_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(101766720)))]; + tensor norm4_1_bias_to_fp16 = const()[name = tensor("norm4_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(101768832)))]; + tensor var_1692_to_fp16 = const()[name = tensor("op_1692_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_51_cast_fp16 = layer_norm(axes = input_51_axes_0, beta = norm4_1_bias_to_fp16, epsilon = var_1692_to_fp16, gamma = norm4_1_weight_to_fp16, x = input_49_cast_fp16)[name = tensor("input_51_cast_fp16")]; + tensor attn4_in_proj_weight_to_fp16 = const()[name = tensor("attn4_in_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(101770944)))]; + tensor linear_16_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = attn4_in_proj_weight_to_fp16, x = input_51_cast_fp16)[name = tensor("linear_16_cast_fp16")]; + tensor var_1706 = const()[name = tensor("op_1706"), val = tensor([1, 256, 3, 16, 64])]; + tensor qkv_9_cast_fp16 = reshape(shape = var_1706, x = linear_16_cast_fp16)[name = tensor("qkv_9_cast_fp16")]; + tensor q_25_begin_0 = const()[name = tensor("q_25_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor q_25_end_0 = const()[name = tensor("q_25_end_0"), val = tensor([1, 256, 1, 16, 64])]; + tensor q_25_end_mask_0 = const()[name = tensor("q_25_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor q_25_squeeze_mask_0 = const()[name = tensor("q_25_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor q_25_cast_fp16 = slice_by_index(begin = q_25_begin_0, end = q_25_end_0, end_mask = q_25_end_mask_0, squeeze_mask = q_25_squeeze_mask_0, x = qkv_9_cast_fp16)[name = tensor("q_25_cast_fp16")]; + tensor k_17_begin_0 = const()[name = tensor("k_17_begin_0"), val = tensor([0, 0, 1, 0, 0])]; + tensor k_17_end_0 = const()[name = tensor("k_17_end_0"), val = tensor([1, 256, 2, 16, 64])]; + tensor k_17_end_mask_0 = const()[name = tensor("k_17_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor k_17_squeeze_mask_0 = const()[name = tensor("k_17_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor k_17_cast_fp16 = slice_by_index(begin = k_17_begin_0, end = k_17_end_0, end_mask = k_17_end_mask_0, squeeze_mask = k_17_squeeze_mask_0, x = qkv_9_cast_fp16)[name = tensor("k_17_cast_fp16")]; + tensor v_9_begin_0 = const()[name = tensor("v_9_begin_0"), val = tensor([0, 0, 2, 0, 0])]; + tensor v_9_end_0 = const()[name = tensor("v_9_end_0"), val = tensor([1, 256, 3, 16, 64])]; + tensor v_9_end_mask_0 = const()[name = tensor("v_9_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor v_9_squeeze_mask_0 = const()[name = tensor("v_9_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor v_9_cast_fp16 = slice_by_index(begin = v_9_begin_0, end = v_9_end_0, end_mask = v_9_end_mask_0, squeeze_mask = v_9_squeeze_mask_0, x = qkv_9_cast_fp16)[name = tensor("v_9_cast_fp16")]; + tensor var_1764 = const()[name = tensor("op_1764"), val = tensor([1, 1])]; + tensor position4_to_fp16_dtype_0 = const()[name = tensor("position4_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor position4_to_fp16 = cast(dtype = position4_to_fp16_dtype_0, x = position4)[name = tensor("cast_137")]; + tensor var_1765_cast_fp16 = reshape(shape = var_1764, x = position4_to_fp16)[name = tensor("op_1765_cast_fp16")]; + tensor ts_27_cast_fp16 = add(x = ts_1_promoted_to_fp16, y = var_1765_cast_fp16)[name = tensor("ts_27_cast_fp16")]; + tensor var_1772 = const()[name = tensor("op_1772"), val = tensor([1, 256, 1, 1])]; + tensor ts_29_cast_fp16 = reshape(shape = var_1772, x = ts_27_cast_fp16)[name = tensor("ts_29_cast_fp16")]; + tensor var_1779 = const()[name = tensor("op_1779"), val = tensor([1, 256, 16, 32, 2])]; + tensor q_complex_9_cast_fp16 = reshape(shape = var_1779, x = q_25_cast_fp16)[name = tensor("q_complex_9_cast_fp16")]; + tensor var_1786 = const()[name = tensor("op_1786"), val = tensor([1, 256, 16, 32, 2])]; + tensor k_complex_9_cast_fp16 = reshape(shape = var_1786, x = k_17_cast_fp16)[name = tensor("k_complex_9_cast_fp16")]; + tensor var_1790_begin_0 = const()[name = tensor("op_1790_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_1790_end_0 = const()[name = tensor("op_1790_end_0"), val = tensor([1, 256, 16, 32, 1])]; + tensor var_1790_end_mask_0 = const()[name = tensor("op_1790_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_1790_squeeze_mask_0 = const()[name = tensor("op_1790_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_1790_cast_fp16 = slice_by_index(begin = var_1790_begin_0, end = var_1790_end_0, end_mask = var_1790_end_mask_0, squeeze_mask = var_1790_squeeze_mask_0, x = q_complex_9_cast_fp16)[name = tensor("op_1790_cast_fp16")]; + tensor var_1798_begin_0 = const()[name = tensor("op_1798_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_1798_end_0 = const()[name = tensor("op_1798_end_0"), val = tensor([1, 256, 16, 32, 2])]; + tensor var_1798_end_mask_0 = const()[name = tensor("op_1798_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_1798_squeeze_mask_0 = const()[name = tensor("op_1798_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_1798_cast_fp16 = slice_by_index(begin = var_1798_begin_0, end = var_1798_end_0, end_mask = var_1798_end_mask_0, squeeze_mask = var_1798_squeeze_mask_0, x = q_complex_9_cast_fp16)[name = tensor("op_1798_cast_fp16")]; + tensor var_1806_begin_0 = const()[name = tensor("op_1806_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_1806_end_0 = const()[name = tensor("op_1806_end_0"), val = tensor([1, 256, 16, 32, 1])]; + tensor var_1806_end_mask_0 = const()[name = tensor("op_1806_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_1806_squeeze_mask_0 = const()[name = tensor("op_1806_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_1806_cast_fp16 = slice_by_index(begin = var_1806_begin_0, end = var_1806_end_0, end_mask = var_1806_end_mask_0, squeeze_mask = var_1806_squeeze_mask_0, x = k_complex_9_cast_fp16)[name = tensor("op_1806_cast_fp16")]; + tensor var_1814_begin_0 = const()[name = tensor("op_1814_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_1814_end_0 = const()[name = tensor("op_1814_end_0"), val = tensor([1, 256, 16, 32, 2])]; + tensor var_1814_end_mask_0 = const()[name = tensor("op_1814_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_1814_squeeze_mask_0 = const()[name = tensor("op_1814_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_1814_cast_fp16 = slice_by_index(begin = var_1814_begin_0, end = var_1814_end_0, end_mask = var_1814_end_mask_0, squeeze_mask = var_1814_squeeze_mask_0, x = k_complex_9_cast_fp16)[name = tensor("op_1814_cast_fp16")]; + tensor freqs_9_to_fp16 = const()[name = tensor("freqs_9_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(108062464)))]; + tensor var_1820_cast_fp16 = mul(x = freqs_9_to_fp16, y = ts_29_cast_fp16)[name = tensor("op_1820_cast_fp16")]; + tensor rotr_9_cast_fp16 = cos(x = var_1820_cast_fp16)[name = tensor("rotr_9_cast_fp16")]; + tensor roti_9_cast_fp16 = sin(x = var_1820_cast_fp16)[name = tensor("roti_9_cast_fp16")]; + tensor var_1824_cast_fp16 = mul(x = var_1790_cast_fp16, y = rotr_9_cast_fp16)[name = tensor("op_1824_cast_fp16")]; + tensor var_1825_cast_fp16 = mul(x = var_1798_cast_fp16, y = roti_9_cast_fp16)[name = tensor("op_1825_cast_fp16")]; + tensor qor_17_cast_fp16 = sub(x = var_1824_cast_fp16, y = var_1825_cast_fp16)[name = tensor("qor_17_cast_fp16")]; + tensor var_1828_cast_fp16 = mul(x = var_1790_cast_fp16, y = roti_9_cast_fp16)[name = tensor("op_1828_cast_fp16")]; + tensor var_1829_cast_fp16 = mul(x = var_1798_cast_fp16, y = rotr_9_cast_fp16)[name = tensor("op_1829_cast_fp16")]; + tensor qoi_17_cast_fp16 = add(x = var_1828_cast_fp16, y = var_1829_cast_fp16)[name = tensor("qoi_17_cast_fp16")]; + tensor var_1832_cast_fp16 = mul(x = var_1806_cast_fp16, y = rotr_9_cast_fp16)[name = tensor("op_1832_cast_fp16")]; + tensor var_1833_cast_fp16 = mul(x = var_1814_cast_fp16, y = roti_9_cast_fp16)[name = tensor("op_1833_cast_fp16")]; + tensor kor_17_cast_fp16 = sub(x = var_1832_cast_fp16, y = var_1833_cast_fp16)[name = tensor("kor_17_cast_fp16")]; + tensor var_1836_cast_fp16 = mul(x = var_1806_cast_fp16, y = roti_9_cast_fp16)[name = tensor("op_1836_cast_fp16")]; + tensor var_1837_cast_fp16 = mul(x = var_1814_cast_fp16, y = rotr_9_cast_fp16)[name = tensor("op_1837_cast_fp16")]; + tensor koi_17_cast_fp16 = add(x = var_1836_cast_fp16, y = var_1837_cast_fp16)[name = tensor("koi_17_cast_fp16")]; + tensor qo_9_axis_0 = const()[name = tensor("qo_9_axis_0"), val = tensor(-1)]; + tensor qo_9_cast_fp16 = stack(axis = qo_9_axis_0, values = (qor_17_cast_fp16, qoi_17_cast_fp16))[name = tensor("qo_9_cast_fp16")]; + tensor ko_9_axis_0 = const()[name = tensor("ko_9_axis_0"), val = tensor(-1)]; + tensor ko_9_cast_fp16 = stack(axis = ko_9_axis_0, values = (kor_17_cast_fp16, koi_17_cast_fp16))[name = tensor("ko_9_cast_fp16")]; + tensor var_1870 = const()[name = tensor("op_1870"), val = tensor([1, 256, 16, 64])]; + tensor q_27_cast_fp16 = reshape(shape = var_1870, x = qo_9_cast_fp16)[name = tensor("q_27_cast_fp16")]; + tensor var_1876 = const()[name = tensor("op_1876"), val = tensor([1, 256, 16, 64])]; + tensor k_19_cast_fp16 = reshape(shape = var_1876, x = ko_9_cast_fp16)[name = tensor("k_19_cast_fp16")]; + tensor write_indices_float_25_cast_fp16 = add(x = var_1765_cast_fp16, y = var_249_promoted_to_fp16)[name = tensor("write_indices_float_25_cast_fp16")]; + tensor _inversed_1897_y_0_to_fp16 = const()[name = tensor("_inversed_1897_y_0_to_fp16"), val = tensor(0x1p-9)]; + tensor _inversed_1897_cast_fp16 = mul(x = write_indices_float_25_cast_fp16, y = _inversed_1897_y_0_to_fp16)[name = tensor("_inversed_1897_cast_fp16")]; + tensor var_1898_cast_fp16 = floor(x = _inversed_1897_cast_fp16)[name = tensor("op_1898_cast_fp16")]; + tensor var_1899_to_fp16 = const()[name = tensor("op_1899_to_fp16"), val = tensor(0x1p+9)]; + tensor var_1900_cast_fp16 = mul(x = var_1898_cast_fp16, y = var_1899_to_fp16)[name = tensor("op_1900_cast_fp16")]; + tensor write_indices_float_27_cast_fp16 = sub(x = write_indices_float_25_cast_fp16, y = var_1900_cast_fp16)[name = tensor("write_indices_float_27_cast_fp16")]; + tensor var_1918_value_0_to_fp16 = const()[name = tensor("op_1918_value_0_to_fp16"), val = tensor(0x1.ffp+8)]; + tensor var_1918_cast_fp16 = fill_like(ref_tensor = write_indices_float_27_cast_fp16, value = var_1918_value_0_to_fp16)[name = tensor("op_1918_cast_fp16")]; + tensor write_indices_float_29_cast_fp16 = select(a = write_indices_float_27_cast_fp16, b = var_1918_cast_fp16, cond = token_valid_1_cast_fp16)[name = tensor("write_indices_float_29_cast_fp16")]; + tensor var_1924_dtype_0 = const()[name = tensor("op_1924_dtype_0"), val = tensor("int32")]; + tensor var_1929 = const()[name = tensor("op_1929"), val = tensor([1, 256, 1, 1])]; + tensor write_indices_float_29_cast_fp16_to_int32 = cast(dtype = var_1924_dtype_0, x = write_indices_float_29_cast_fp16)[name = tensor("cast_136")]; + tensor var_1930 = reshape(shape = var_1929, x = write_indices_float_29_cast_fp16_to_int32)[name = tensor("op_1930")]; + tensor write_indices_9_reps_0 = const()[name = tensor("write_indices_9_reps_0"), val = tensor([1, 1, 16, 64])]; + tensor write_indices_9 = tile(reps = write_indices_9_reps_0, x = var_1930)[name = tensor("write_indices_9")]; + tensor var_1940_begin_0 = const()[name = tensor("op_1940_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_1940_end_0 = const()[name = tensor("op_1940_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor var_1940_end_mask_0 = const()[name = tensor("op_1940_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_1940_squeeze_mask_0 = const()[name = tensor("op_1940_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor cache4_to_fp16_dtype_0 = const()[name = tensor("cache4_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor cache4_to_fp16 = cast(dtype = cache4_to_fp16_dtype_0, x = cache4)[name = tensor("cast_135")]; + tensor var_1940_cast_fp16 = slice_by_index(begin = var_1940_begin_0, end = var_1940_end_0, end_mask = var_1940_end_mask_0, squeeze_mask = var_1940_squeeze_mask_0, x = cache4_to_fp16)[name = tensor("op_1940_cast_fp16")]; + tensor var_1942_axis_0 = const()[name = tensor("op_1942_axis_0"), val = tensor(1)]; + tensor var_1942_mode_0 = const()[name = tensor("op_1942_mode_0"), val = tensor("update")]; + tensor var_1942_validate_indices_0 = const()[name = tensor("op_1942_validate_indices_0"), val = tensor(false)]; + tensor var_1942_cast_fp16 = scatter_along_axis(axis = var_1942_axis_0, data = var_1940_cast_fp16, indices = write_indices_9, mode = var_1942_mode_0, updates = k_19_cast_fp16, validate_indices = var_1942_validate_indices_0)[name = tensor("op_1942_cast_fp16")]; + tensor concat_34 = const()[name = tensor("concat_34"), val = tensor([0, 0, 0, 0, 0])]; + tensor concat_35 = const()[name = tensor("concat_35"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_9_internal_tensor_assign_1_stride_0 = const()[name = tensor("new_cache_9_internal_tensor_assign_1_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_9_internal_tensor_assign_1_begin_mask_0 = const()[name = tensor("new_cache_9_internal_tensor_assign_1_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_9_internal_tensor_assign_1_end_mask_0 = const()[name = tensor("new_cache_9_internal_tensor_assign_1_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_9_internal_tensor_assign_1_squeeze_mask_0 = const()[name = tensor("new_cache_9_internal_tensor_assign_1_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_18 = const()[name = tensor("shape_18"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_8 = const()[name = tensor("reduce_prod_8"), val = tensor(1048576)]; + tensor range_1d_8_start_0 = const()[name = tensor("range_1d_8_start_0"), val = tensor(0)]; + tensor range_1d_8_step_0 = const()[name = tensor("range_1d_8_step_0"), val = tensor(1)]; + tensor range_1d_8 = range_1d(end = reduce_prod_8, start = range_1d_8_start_0, step = range_1d_8_step_0)[name = tensor("range_1d_8")]; + tensor reshape_40 = reshape(shape = shape_18, x = range_1d_8)[name = tensor("reshape_40")]; + tensor slice_by_index_8 = slice_by_index(begin = concat_34, begin_mask = new_cache_9_internal_tensor_assign_1_begin_mask_0, end = concat_35, end_mask = new_cache_9_internal_tensor_assign_1_end_mask_0, squeeze_mask = new_cache_9_internal_tensor_assign_1_squeeze_mask_0, stride = new_cache_9_internal_tensor_assign_1_stride_0, x = reshape_40)[name = tensor("slice_by_index_8")]; + tensor reshape_41_shape_0 = const()[name = tensor("reshape_41_shape_0"), val = tensor([-1])]; + tensor reshape_41 = reshape(shape = reshape_41_shape_0, x = slice_by_index_8)[name = tensor("reshape_41")]; + tensor reshape_42_shape_0 = const()[name = tensor("reshape_42_shape_0"), val = tensor([-1])]; + tensor reshape_42_cast_fp16 = reshape(shape = reshape_42_shape_0, x = var_1942_cast_fp16)[name = tensor("reshape_42_cast_fp16")]; + tensor reshape_43_shape_0 = const()[name = tensor("reshape_43_shape_0"), val = tensor([-1])]; + tensor reshape_43_cast_fp16 = reshape(shape = reshape_43_shape_0, x = cache4_to_fp16)[name = tensor("reshape_43_cast_fp16")]; + tensor scatter_8_mode_0 = const()[name = tensor("scatter_8_mode_0"), val = tensor("update")]; + tensor scatter_8_axis_0 = const()[name = tensor("scatter_8_axis_0"), val = tensor(0)]; + tensor scatter_8_validate_indices_0 = const()[name = tensor("scatter_8_validate_indices_0"), val = tensor(false)]; + tensor scatter_8_cast_fp16 = scatter(axis = scatter_8_axis_0, data = reshape_43_cast_fp16, indices = reshape_41, mode = scatter_8_mode_0, updates = reshape_42_cast_fp16, validate_indices = scatter_8_validate_indices_0)[name = tensor("scatter_8_cast_fp16")]; + tensor reshape_44_cast_fp16 = reshape(shape = shape_18, x = scatter_8_cast_fp16)[name = tensor("reshape_44_cast_fp16")]; + tensor var_1950_begin_0 = const()[name = tensor("op_1950_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor var_1950_end_0 = const()[name = tensor("op_1950_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor var_1950_end_mask_0 = const()[name = tensor("op_1950_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_1950_squeeze_mask_0 = const()[name = tensor("op_1950_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor var_1950_cast_fp16 = slice_by_index(begin = var_1950_begin_0, end = var_1950_end_0, end_mask = var_1950_end_mask_0, squeeze_mask = var_1950_squeeze_mask_0, x = reshape_44_cast_fp16)[name = tensor("op_1950_cast_fp16")]; + tensor var_1952_axis_0 = const()[name = tensor("op_1952_axis_0"), val = tensor(1)]; + tensor var_1952_mode_0 = const()[name = tensor("op_1952_mode_0"), val = tensor("update")]; + tensor var_1952_validate_indices_0 = const()[name = tensor("op_1952_validate_indices_0"), val = tensor(false)]; + tensor var_1952_cast_fp16 = scatter_along_axis(axis = var_1952_axis_0, data = var_1950_cast_fp16, indices = write_indices_9, mode = var_1952_mode_0, updates = v_9_cast_fp16, validate_indices = var_1952_validate_indices_0)[name = tensor("op_1952_cast_fp16")]; + tensor concat_36 = const()[name = tensor("concat_36"), val = tensor([1, 0, 0, 0, 0])]; + tensor concat_37 = const()[name = tensor("concat_37"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_9_internal_tensor_assign_2_stride_0 = const()[name = tensor("new_cache_9_internal_tensor_assign_2_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_9_internal_tensor_assign_2_begin_mask_0 = const()[name = tensor("new_cache_9_internal_tensor_assign_2_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_9_internal_tensor_assign_2_end_mask_0 = const()[name = tensor("new_cache_9_internal_tensor_assign_2_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_9_internal_tensor_assign_2_squeeze_mask_0 = const()[name = tensor("new_cache_9_internal_tensor_assign_2_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_19 = const()[name = tensor("shape_19"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_9 = const()[name = tensor("reduce_prod_9"), val = tensor(1048576)]; + tensor range_1d_9_start_0 = const()[name = tensor("range_1d_9_start_0"), val = tensor(0)]; + tensor range_1d_9_step_0 = const()[name = tensor("range_1d_9_step_0"), val = tensor(1)]; + tensor range_1d_9 = range_1d(end = reduce_prod_9, start = range_1d_9_start_0, step = range_1d_9_step_0)[name = tensor("range_1d_9")]; + tensor reshape_45 = reshape(shape = shape_19, x = range_1d_9)[name = tensor("reshape_45")]; + tensor slice_by_index_9 = slice_by_index(begin = concat_36, begin_mask = new_cache_9_internal_tensor_assign_2_begin_mask_0, end = concat_37, end_mask = new_cache_9_internal_tensor_assign_2_end_mask_0, squeeze_mask = new_cache_9_internal_tensor_assign_2_squeeze_mask_0, stride = new_cache_9_internal_tensor_assign_2_stride_0, x = reshape_45)[name = tensor("slice_by_index_9")]; + tensor reshape_46_shape_0 = const()[name = tensor("reshape_46_shape_0"), val = tensor([-1])]; + tensor reshape_46 = reshape(shape = reshape_46_shape_0, x = slice_by_index_9)[name = tensor("reshape_46")]; + tensor reshape_47_shape_0 = const()[name = tensor("reshape_47_shape_0"), val = tensor([-1])]; + tensor reshape_47_cast_fp16 = reshape(shape = reshape_47_shape_0, x = var_1952_cast_fp16)[name = tensor("reshape_47_cast_fp16")]; + tensor reshape_48_shape_0 = const()[name = tensor("reshape_48_shape_0"), val = tensor([-1])]; + tensor reshape_48_cast_fp16 = reshape(shape = reshape_48_shape_0, x = reshape_44_cast_fp16)[name = tensor("reshape_48_cast_fp16")]; + tensor scatter_9_mode_0 = const()[name = tensor("scatter_9_mode_0"), val = tensor("update")]; + tensor scatter_9_axis_0 = const()[name = tensor("scatter_9_axis_0"), val = tensor(0)]; + tensor scatter_9_validate_indices_0 = const()[name = tensor("scatter_9_validate_indices_0"), val = tensor(false)]; + tensor scatter_9_cast_fp16 = scatter(axis = scatter_9_axis_0, data = reshape_48_cast_fp16, indices = reshape_46, mode = scatter_9_mode_0, updates = reshape_47_cast_fp16, validate_indices = scatter_9_validate_indices_0)[name = tensor("scatter_9_cast_fp16")]; + tensor reshape_49_cast_fp16 = reshape(shape = shape_19, x = scatter_9_cast_fp16)[name = tensor("reshape_49_cast_fp16")]; + tensor reshape_49_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("reshape_49_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor keys_25_begin_0 = const()[name = tensor("keys_25_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor keys_25_end_0 = const()[name = tensor("keys_25_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor keys_25_end_mask_0 = const()[name = tensor("keys_25_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor keys_25_squeeze_mask_0 = const()[name = tensor("keys_25_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor keys_25_cast_fp16 = slice_by_index(begin = keys_25_begin_0, end = keys_25_end_0, end_mask = keys_25_end_mask_0, squeeze_mask = keys_25_squeeze_mask_0, x = reshape_49_cast_fp16)[name = tensor("keys_25_cast_fp16")]; + tensor values_25_begin_0 = const()[name = tensor("values_25_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor values_25_end_0 = const()[name = tensor("values_25_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor values_25_end_mask_0 = const()[name = tensor("values_25_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor values_25_squeeze_mask_0 = const()[name = tensor("values_25_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor values_25_cast_fp16 = slice_by_index(begin = values_25_begin_0, end = values_25_end_0, end_mask = values_25_end_mask_0, squeeze_mask = values_25_squeeze_mask_0, x = reshape_49_cast_fp16)[name = tensor("values_25_cast_fp16")]; + tensor var_1964_cast_fp16 = not_equal(x = keys_25_cast_fp16, y = keys_25_cast_fp16)[name = tensor("op_1964_cast_fp16")]; + tensor keys_27_cast_fp16 = select(a = var_326_to_fp16, b = keys_25_cast_fp16, cond = var_1964_cast_fp16)[name = tensor("keys_27_cast_fp16")]; + tensor var_1972_cast_fp16 = not_equal(x = values_25_cast_fp16, y = values_25_cast_fp16)[name = tensor("op_1972_cast_fp16")]; + tensor values_27_cast_fp16 = select(a = var_326_to_fp16, b = values_25_cast_fp16, cond = var_1972_cast_fp16)[name = tensor("values_27_cast_fp16")]; + tensor var_1996 = const()[name = tensor("op_1996"), val = tensor([0, 2, 1, 3])]; + tensor var_2012 = const()[name = tensor("op_2012"), val = tensor([1, 1, 1])]; + tensor var_2013_cast_fp16 = reshape(shape = var_2012, x = position4_to_fp16)[name = tensor("op_2013_cast_fp16")]; + tensor q_positions_9_cast_fp16 = add(x = var_2013_cast_fp16, y = q_offsets_1_promoted_to_fp16)[name = tensor("q_positions_9_cast_fp16")]; + tensor valid_end_9_cast_fp16 = add(x = var_2013_cast_fp16, y = var_392_cast_fp16)[name = tensor("valid_end_9_cast_fp16")]; + tensor valid_mask_9_cast_fp16 = less(x = k_positions_1_promoted_to_fp16, y = valid_end_9_cast_fp16)[name = tensor("valid_mask_9_cast_fp16")]; + tensor causal_mask_9_cast_fp16 = less_equal(x = k_positions_1_promoted_to_fp16, y = q_positions_9_cast_fp16)[name = tensor("causal_mask_9_cast_fp16")]; + tensor var_2041 = logical_and(x = valid_mask_9_cast_fp16, y = causal_mask_9_cast_fp16)[name = tensor("op_2041")]; + tensor attn_mask_9_axes_0 = const()[name = tensor("attn_mask_9_axes_0"), val = tensor([1])]; + tensor attn_mask_9 = expand_dims(axes = attn_mask_9_axes_0, x = var_2041)[name = tensor("attn_mask_9")]; + tensor var_2055_transpose_x_0 = const()[name = tensor("op_2055_transpose_x_0"), val = tensor(false)]; + tensor var_2055_transpose_y_0 = const()[name = tensor("op_2055_transpose_y_0"), val = tensor(false)]; + tensor transpose_23_perm_0 = const()[name = tensor("transpose_23_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_24_perm_0 = const()[name = tensor("transpose_24_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_24 = transpose(perm = transpose_24_perm_0, x = keys_27_cast_fp16)[name = tensor("transpose_26")]; + tensor transpose_23 = transpose(perm = transpose_23_perm_0, x = q_27_cast_fp16)[name = tensor("transpose_27")]; + tensor var_2055_cast_fp16 = matmul(transpose_x = var_2055_transpose_x_0, transpose_y = var_2055_transpose_y_0, x = transpose_23, y = transpose_24)[name = tensor("op_2055_cast_fp16")]; + tensor var_2049_to_fp16 = const()[name = tensor("op_2049_to_fp16"), val = tensor([0x1p-3])]; + tensor attn_weights_25_cast_fp16 = mul(x = var_2055_cast_fp16, y = var_2049_to_fp16)[name = tensor("attn_weights_25_cast_fp16")]; + tensor var_2057 = logical_not(x = attn_mask_9)[name = tensor("op_2057")]; + tensor var_2058_to_fp16 = const()[name = tensor("op_2058_to_fp16"), val = tensor(-inf)]; + tensor attn_weights_27_cast_fp16 = select(a = var_2058_to_fp16, b = attn_weights_25_cast_fp16, cond = var_2057)[name = tensor("attn_weights_27_cast_fp16")]; + tensor var_2060 = const()[name = tensor("op_2060"), val = tensor(-1)]; + tensor attn_weights_29_cast_fp16 = softmax(axis = var_2060, x = attn_weights_27_cast_fp16)[name = tensor("attn_weights_29_cast_fp16")]; + tensor attn_output_9_transpose_x_0 = const()[name = tensor("attn_output_9_transpose_x_0"), val = tensor(false)]; + tensor attn_output_9_transpose_y_0 = const()[name = tensor("attn_output_9_transpose_y_0"), val = tensor(false)]; + tensor values_29_cast_fp16 = transpose(perm = var_1996, x = values_27_cast_fp16)[name = tensor("transpose_28")]; + tensor attn_output_9_cast_fp16 = matmul(transpose_x = attn_output_9_transpose_x_0, transpose_y = attn_output_9_transpose_y_0, x = attn_weights_29_cast_fp16, y = values_29_cast_fp16)[name = tensor("attn_output_9_cast_fp16")]; + tensor var_2068 = const()[name = tensor("op_2068"), val = tensor([0, 2, 1, 3])]; + tensor var_2073 = const()[name = tensor("op_2073"), val = tensor([1, 256, 1024])]; + tensor var_2069_cast_fp16 = transpose(perm = var_2068, x = attn_output_9_cast_fp16)[name = tensor("transpose_25")]; + tensor input_53_cast_fp16 = reshape(shape = var_2073, x = var_2069_cast_fp16)[name = tensor("input_53_cast_fp16")]; + tensor attn4_out_proj_weight_to_fp16 = const()[name = tensor("attn4_out_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(108062592)))]; + tensor linear_17_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = attn4_out_proj_weight_to_fp16, x = input_53_cast_fp16)[name = tensor("linear_17_cast_fp16")]; + tensor var_2082_cast_fp16 = add(x = position4_to_fp16, y = valid_len_to_fp16)[name = tensor("op_2082_cast_fp16")]; + tensor var_2082_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_2082_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor input_55_cast_fp16 = add(x = input_49_cast_fp16, y = linear_17_cast_fp16)[name = tensor("input_55_cast_fp16")]; + tensor input_57_axes_0 = const()[name = tensor("input_57_axes_0"), val = tensor([-1])]; + tensor norm4_2_weight_to_fp16 = const()[name = tensor("norm4_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(110159808)))]; + tensor norm4_2_bias_to_fp16 = const()[name = tensor("norm4_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(110161920)))]; + tensor var_2086_to_fp16 = const()[name = tensor("op_2086_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_57_cast_fp16 = layer_norm(axes = input_57_axes_0, beta = norm4_2_bias_to_fp16, epsilon = var_2086_to_fp16, gamma = norm4_2_weight_to_fp16, x = input_55_cast_fp16)[name = tensor("input_57_cast_fp16")]; + tensor linear4_1_weight_to_fp16 = const()[name = tensor("linear4_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(110164032)))]; + tensor linear_18_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = linear4_1_weight_to_fp16, x = input_57_cast_fp16)[name = tensor("linear_18_cast_fp16")]; + tensor input_59_mode_0 = const()[name = tensor("input_59_mode_0"), val = tensor("EXACT")]; + tensor input_59_cast_fp16 = gelu(mode = input_59_mode_0, x = linear_18_cast_fp16)[name = tensor("input_59_cast_fp16")]; + tensor linear4_2_weight_to_fp16 = const()[name = tensor("linear4_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(118552704)))]; + tensor linear_19_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = linear4_2_weight_to_fp16, x = input_59_cast_fp16)[name = tensor("linear_19_cast_fp16")]; + tensor input_61_cast_fp16 = add(x = input_55_cast_fp16, y = linear_19_cast_fp16)[name = tensor("input_61_cast_fp16")]; + tensor input_63_axes_0 = const()[name = tensor("input_63_axes_0"), val = tensor([-1])]; + tensor norm5_1_weight_to_fp16 = const()[name = tensor("norm5_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(126941376)))]; + tensor norm5_1_bias_to_fp16 = const()[name = tensor("norm5_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(126943488)))]; + tensor var_2103_to_fp16 = const()[name = tensor("op_2103_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_63_cast_fp16 = layer_norm(axes = input_63_axes_0, beta = norm5_1_bias_to_fp16, epsilon = var_2103_to_fp16, gamma = norm5_1_weight_to_fp16, x = input_61_cast_fp16)[name = tensor("input_63_cast_fp16")]; + tensor attn5_in_proj_weight_to_fp16 = const()[name = tensor("attn5_in_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(126945600)))]; + tensor linear_20_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = attn5_in_proj_weight_to_fp16, x = input_63_cast_fp16)[name = tensor("linear_20_cast_fp16")]; + tensor var_2117 = const()[name = tensor("op_2117"), val = tensor([1, 256, 3, 16, 64])]; + tensor qkv_cast_fp16 = reshape(shape = var_2117, x = linear_20_cast_fp16)[name = tensor("qkv_cast_fp16")]; + tensor k_21_begin_0 = const()[name = tensor("k_21_begin_0"), val = tensor([0, 0, 1, 0, 0])]; + tensor k_21_end_0 = const()[name = tensor("k_21_end_0"), val = tensor([1, 256, 2, 16, 64])]; + tensor k_21_end_mask_0 = const()[name = tensor("k_21_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor k_21_squeeze_mask_0 = const()[name = tensor("k_21_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor k_21_cast_fp16 = slice_by_index(begin = k_21_begin_0, end = k_21_end_0, end_mask = k_21_end_mask_0, squeeze_mask = k_21_squeeze_mask_0, x = qkv_cast_fp16)[name = tensor("k_21_cast_fp16")]; + tensor v_begin_0 = const()[name = tensor("v_begin_0"), val = tensor([0, 0, 2, 0, 0])]; + tensor v_end_0 = const()[name = tensor("v_end_0"), val = tensor([1, 256, 3, 16, 64])]; + tensor v_end_mask_0 = const()[name = tensor("v_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor v_squeeze_mask_0 = const()[name = tensor("v_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor v_cast_fp16 = slice_by_index(begin = v_begin_0, end = v_end_0, end_mask = v_end_mask_0, squeeze_mask = v_squeeze_mask_0, x = qkv_cast_fp16)[name = tensor("v_cast_fp16")]; + tensor var_2162 = const()[name = tensor("op_2162"), val = tensor([1, 1])]; + tensor position5_to_fp16_dtype_0 = const()[name = tensor("position5_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor position5_to_fp16 = cast(dtype = position5_to_fp16_dtype_0, x = position5)[name = tensor("cast_132")]; + tensor var_2163_cast_fp16 = reshape(shape = var_2162, x = position5_to_fp16)[name = tensor("op_2163_cast_fp16")]; + tensor ts_33_cast_fp16 = add(x = ts_1_promoted_to_fp16, y = var_2163_cast_fp16)[name = tensor("ts_33_cast_fp16")]; + tensor var_2170 = const()[name = tensor("op_2170"), val = tensor([1, 256, 1, 1])]; + tensor ts_cast_fp16 = reshape(shape = var_2170, x = ts_33_cast_fp16)[name = tensor("ts_cast_fp16")]; + tensor var_2177 = const()[name = tensor("op_2177"), val = tensor([1, 256, 16, 32, 2])]; + tensor k_complex_cast_fp16 = reshape(shape = var_2177, x = k_21_cast_fp16)[name = tensor("k_complex_cast_fp16")]; + tensor var_2181_begin_0 = const()[name = tensor("op_2181_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_2181_end_0 = const()[name = tensor("op_2181_end_0"), val = tensor([1, 256, 16, 32, 1])]; + tensor var_2181_end_mask_0 = const()[name = tensor("op_2181_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_2181_squeeze_mask_0 = const()[name = tensor("op_2181_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_2181_cast_fp16 = slice_by_index(begin = var_2181_begin_0, end = var_2181_end_0, end_mask = var_2181_end_mask_0, squeeze_mask = var_2181_squeeze_mask_0, x = k_complex_cast_fp16)[name = tensor("op_2181_cast_fp16")]; + tensor var_2189_begin_0 = const()[name = tensor("op_2189_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_2189_end_0 = const()[name = tensor("op_2189_end_0"), val = tensor([1, 256, 16, 32, 2])]; + tensor var_2189_end_mask_0 = const()[name = tensor("op_2189_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_2189_squeeze_mask_0 = const()[name = tensor("op_2189_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_2189_cast_fp16 = slice_by_index(begin = var_2189_begin_0, end = var_2189_end_0, end_mask = var_2189_end_mask_0, squeeze_mask = var_2189_squeeze_mask_0, x = k_complex_cast_fp16)[name = tensor("op_2189_cast_fp16")]; + tensor freqs_to_fp16 = const()[name = tensor("freqs_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(133237120)))]; + tensor var_2195_cast_fp16 = mul(x = freqs_to_fp16, y = ts_cast_fp16)[name = tensor("op_2195_cast_fp16")]; + tensor rotr_cast_fp16 = cos(x = var_2195_cast_fp16)[name = tensor("rotr_cast_fp16")]; + tensor roti_cast_fp16 = sin(x = var_2195_cast_fp16)[name = tensor("roti_cast_fp16")]; + tensor var_2199_cast_fp16 = mul(x = var_2181_cast_fp16, y = rotr_cast_fp16)[name = tensor("op_2199_cast_fp16")]; + tensor var_2200_cast_fp16 = mul(x = var_2189_cast_fp16, y = roti_cast_fp16)[name = tensor("op_2200_cast_fp16")]; + tensor kor_21_cast_fp16 = sub(x = var_2199_cast_fp16, y = var_2200_cast_fp16)[name = tensor("kor_21_cast_fp16")]; + tensor var_2203_cast_fp16 = mul(x = var_2181_cast_fp16, y = roti_cast_fp16)[name = tensor("op_2203_cast_fp16")]; + tensor var_2204_cast_fp16 = mul(x = var_2189_cast_fp16, y = rotr_cast_fp16)[name = tensor("op_2204_cast_fp16")]; + tensor koi_21_cast_fp16 = add(x = var_2203_cast_fp16, y = var_2204_cast_fp16)[name = tensor("koi_21_cast_fp16")]; + tensor ko_axis_0 = const()[name = tensor("ko_axis_0"), val = tensor(-1)]; + tensor ko_cast_fp16 = stack(axis = ko_axis_0, values = (kor_21_cast_fp16, koi_21_cast_fp16))[name = tensor("ko_cast_fp16")]; + tensor var_2224 = const()[name = tensor("op_2224"), val = tensor([1, 256, 16, 64])]; + tensor k_cast_fp16 = reshape(shape = var_2224, x = ko_cast_fp16)[name = tensor("k_cast_fp16")]; + tensor write_indices_float_31_cast_fp16 = add(x = var_2163_cast_fp16, y = var_249_promoted_to_fp16)[name = tensor("write_indices_float_31_cast_fp16")]; + tensor _inversed_2245_y_0_to_fp16 = const()[name = tensor("_inversed_2245_y_0_to_fp16"), val = tensor(0x1p-9)]; + tensor _inversed_2245_cast_fp16 = mul(x = write_indices_float_31_cast_fp16, y = _inversed_2245_y_0_to_fp16)[name = tensor("_inversed_2245_cast_fp16")]; + tensor var_2246_cast_fp16 = floor(x = _inversed_2245_cast_fp16)[name = tensor("op_2246_cast_fp16")]; + tensor var_2247_to_fp16 = const()[name = tensor("op_2247_to_fp16"), val = tensor(0x1p+9)]; + tensor var_2248_cast_fp16 = mul(x = var_2246_cast_fp16, y = var_2247_to_fp16)[name = tensor("op_2248_cast_fp16")]; + tensor write_indices_float_33_cast_fp16 = sub(x = write_indices_float_31_cast_fp16, y = var_2248_cast_fp16)[name = tensor("write_indices_float_33_cast_fp16")]; + tensor var_2266_value_0_to_fp16 = const()[name = tensor("op_2266_value_0_to_fp16"), val = tensor(0x1.ffp+8)]; + tensor var_2266_cast_fp16 = fill_like(ref_tensor = write_indices_float_33_cast_fp16, value = var_2266_value_0_to_fp16)[name = tensor("op_2266_cast_fp16")]; + tensor write_indices_float_cast_fp16 = select(a = write_indices_float_33_cast_fp16, b = var_2266_cast_fp16, cond = token_valid_1_cast_fp16)[name = tensor("write_indices_float_cast_fp16")]; + tensor var_2272_dtype_0 = const()[name = tensor("op_2272_dtype_0"), val = tensor("int32")]; + tensor var_2277 = const()[name = tensor("op_2277"), val = tensor([1, 256, 1, 1])]; + tensor write_indices_float_cast_fp16_to_int32 = cast(dtype = var_2272_dtype_0, x = write_indices_float_cast_fp16)[name = tensor("cast_131")]; + tensor var_2278 = reshape(shape = var_2277, x = write_indices_float_cast_fp16_to_int32)[name = tensor("op_2278")]; + tensor write_indices_reps_0 = const()[name = tensor("write_indices_reps_0"), val = tensor([1, 1, 16, 64])]; + tensor write_indices = tile(reps = write_indices_reps_0, x = var_2278)[name = tensor("write_indices")]; + tensor var_2288_begin_0 = const()[name = tensor("op_2288_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_2288_end_0 = const()[name = tensor("op_2288_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor var_2288_end_mask_0 = const()[name = tensor("op_2288_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_2288_squeeze_mask_0 = const()[name = tensor("op_2288_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor cache5_to_fp16_dtype_0 = const()[name = tensor("cache5_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor cache5_to_fp16 = cast(dtype = cache5_to_fp16_dtype_0, x = cache5)[name = tensor("cast_130")]; + tensor var_2288_cast_fp16 = slice_by_index(begin = var_2288_begin_0, end = var_2288_end_0, end_mask = var_2288_end_mask_0, squeeze_mask = var_2288_squeeze_mask_0, x = cache5_to_fp16)[name = tensor("op_2288_cast_fp16")]; + tensor var_2290_axis_0 = const()[name = tensor("op_2290_axis_0"), val = tensor(1)]; + tensor var_2290_mode_0 = const()[name = tensor("op_2290_mode_0"), val = tensor("update")]; + tensor var_2290_validate_indices_0 = const()[name = tensor("op_2290_validate_indices_0"), val = tensor(false)]; + tensor var_2290_cast_fp16 = scatter_along_axis(axis = var_2290_axis_0, data = var_2288_cast_fp16, indices = write_indices, mode = var_2290_mode_0, updates = k_cast_fp16, validate_indices = var_2290_validate_indices_0)[name = tensor("op_2290_cast_fp16")]; + tensor concat_42 = const()[name = tensor("concat_42"), val = tensor([0, 0, 0, 0, 0])]; + tensor concat_43 = const()[name = tensor("concat_43"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_internal_tensor_assign_1_stride_0 = const()[name = tensor("new_cache_internal_tensor_assign_1_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_internal_tensor_assign_1_begin_mask_0 = const()[name = tensor("new_cache_internal_tensor_assign_1_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_internal_tensor_assign_1_end_mask_0 = const()[name = tensor("new_cache_internal_tensor_assign_1_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_internal_tensor_assign_1_squeeze_mask_0 = const()[name = tensor("new_cache_internal_tensor_assign_1_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_20 = const()[name = tensor("shape_20"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_10 = const()[name = tensor("reduce_prod_10"), val = tensor(1048576)]; + tensor range_1d_10_start_0 = const()[name = tensor("range_1d_10_start_0"), val = tensor(0)]; + tensor range_1d_10_step_0 = const()[name = tensor("range_1d_10_step_0"), val = tensor(1)]; + tensor range_1d_10 = range_1d(end = reduce_prod_10, start = range_1d_10_start_0, step = range_1d_10_step_0)[name = tensor("range_1d_10")]; + tensor reshape_50 = reshape(shape = shape_20, x = range_1d_10)[name = tensor("reshape_50")]; + tensor slice_by_index_10 = slice_by_index(begin = concat_42, begin_mask = new_cache_internal_tensor_assign_1_begin_mask_0, end = concat_43, end_mask = new_cache_internal_tensor_assign_1_end_mask_0, squeeze_mask = new_cache_internal_tensor_assign_1_squeeze_mask_0, stride = new_cache_internal_tensor_assign_1_stride_0, x = reshape_50)[name = tensor("slice_by_index_10")]; + tensor reshape_51_shape_0 = const()[name = tensor("reshape_51_shape_0"), val = tensor([-1])]; + tensor reshape_51 = reshape(shape = reshape_51_shape_0, x = slice_by_index_10)[name = tensor("reshape_51")]; + tensor reshape_52_shape_0 = const()[name = tensor("reshape_52_shape_0"), val = tensor([-1])]; + tensor reshape_52_cast_fp16 = reshape(shape = reshape_52_shape_0, x = var_2290_cast_fp16)[name = tensor("reshape_52_cast_fp16")]; + tensor reshape_53_shape_0 = const()[name = tensor("reshape_53_shape_0"), val = tensor([-1])]; + tensor reshape_53_cast_fp16 = reshape(shape = reshape_53_shape_0, x = cache5_to_fp16)[name = tensor("reshape_53_cast_fp16")]; + tensor scatter_10_mode_0 = const()[name = tensor("scatter_10_mode_0"), val = tensor("update")]; + tensor scatter_10_axis_0 = const()[name = tensor("scatter_10_axis_0"), val = tensor(0)]; + tensor scatter_10_validate_indices_0 = const()[name = tensor("scatter_10_validate_indices_0"), val = tensor(false)]; + tensor scatter_10_cast_fp16 = scatter(axis = scatter_10_axis_0, data = reshape_53_cast_fp16, indices = reshape_51, mode = scatter_10_mode_0, updates = reshape_52_cast_fp16, validate_indices = scatter_10_validate_indices_0)[name = tensor("scatter_10_cast_fp16")]; + tensor reshape_54_cast_fp16 = reshape(shape = shape_20, x = scatter_10_cast_fp16)[name = tensor("reshape_54_cast_fp16")]; + tensor var_2298_begin_0 = const()[name = tensor("op_2298_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor var_2298_end_0 = const()[name = tensor("op_2298_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor var_2298_end_mask_0 = const()[name = tensor("op_2298_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_2298_squeeze_mask_0 = const()[name = tensor("op_2298_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor var_2298_cast_fp16 = slice_by_index(begin = var_2298_begin_0, end = var_2298_end_0, end_mask = var_2298_end_mask_0, squeeze_mask = var_2298_squeeze_mask_0, x = reshape_54_cast_fp16)[name = tensor("op_2298_cast_fp16")]; + tensor var_2300_axis_0 = const()[name = tensor("op_2300_axis_0"), val = tensor(1)]; + tensor var_2300_mode_0 = const()[name = tensor("op_2300_mode_0"), val = tensor("update")]; + tensor var_2300_validate_indices_0 = const()[name = tensor("op_2300_validate_indices_0"), val = tensor(false)]; + tensor var_2300_cast_fp16 = scatter_along_axis(axis = var_2300_axis_0, data = var_2298_cast_fp16, indices = write_indices, mode = var_2300_mode_0, updates = v_cast_fp16, validate_indices = var_2300_validate_indices_0)[name = tensor("op_2300_cast_fp16")]; + tensor concat_44 = const()[name = tensor("concat_44"), val = tensor([1, 0, 0, 0, 0])]; + tensor concat_45 = const()[name = tensor("concat_45"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_internal_tensor_assign_2_stride_0 = const()[name = tensor("new_cache_internal_tensor_assign_2_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_internal_tensor_assign_2_begin_mask_0 = const()[name = tensor("new_cache_internal_tensor_assign_2_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_internal_tensor_assign_2_end_mask_0 = const()[name = tensor("new_cache_internal_tensor_assign_2_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_internal_tensor_assign_2_squeeze_mask_0 = const()[name = tensor("new_cache_internal_tensor_assign_2_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_21 = const()[name = tensor("shape_21"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_11 = const()[name = tensor("reduce_prod_11"), val = tensor(1048576)]; + tensor range_1d_11_start_0 = const()[name = tensor("range_1d_11_start_0"), val = tensor(0)]; + tensor range_1d_11_step_0 = const()[name = tensor("range_1d_11_step_0"), val = tensor(1)]; + tensor range_1d_11 = range_1d(end = reduce_prod_11, start = range_1d_11_start_0, step = range_1d_11_step_0)[name = tensor("range_1d_11")]; + tensor reshape_55 = reshape(shape = shape_21, x = range_1d_11)[name = tensor("reshape_55")]; + tensor slice_by_index_11 = slice_by_index(begin = concat_44, begin_mask = new_cache_internal_tensor_assign_2_begin_mask_0, end = concat_45, end_mask = new_cache_internal_tensor_assign_2_end_mask_0, squeeze_mask = new_cache_internal_tensor_assign_2_squeeze_mask_0, stride = new_cache_internal_tensor_assign_2_stride_0, x = reshape_55)[name = tensor("slice_by_index_11")]; + tensor reshape_56_shape_0 = const()[name = tensor("reshape_56_shape_0"), val = tensor([-1])]; + tensor reshape_56 = reshape(shape = reshape_56_shape_0, x = slice_by_index_11)[name = tensor("reshape_56")]; + tensor reshape_57_shape_0 = const()[name = tensor("reshape_57_shape_0"), val = tensor([-1])]; + tensor reshape_57_cast_fp16 = reshape(shape = reshape_57_shape_0, x = var_2300_cast_fp16)[name = tensor("reshape_57_cast_fp16")]; + tensor reshape_58_shape_0 = const()[name = tensor("reshape_58_shape_0"), val = tensor([-1])]; + tensor reshape_58_cast_fp16 = reshape(shape = reshape_58_shape_0, x = reshape_54_cast_fp16)[name = tensor("reshape_58_cast_fp16")]; + tensor scatter_11_mode_0 = const()[name = tensor("scatter_11_mode_0"), val = tensor("update")]; + tensor scatter_11_axis_0 = const()[name = tensor("scatter_11_axis_0"), val = tensor(0)]; + tensor scatter_11_validate_indices_0 = const()[name = tensor("scatter_11_validate_indices_0"), val = tensor(false)]; + tensor scatter_11_cast_fp16 = scatter(axis = scatter_11_axis_0, data = reshape_58_cast_fp16, indices = reshape_56, mode = scatter_11_mode_0, updates = reshape_57_cast_fp16, validate_indices = scatter_11_validate_indices_0)[name = tensor("scatter_11_cast_fp16")]; + tensor reshape_59_cast_fp16 = reshape(shape = shape_21, x = scatter_11_cast_fp16)[name = tensor("reshape_59_cast_fp16")]; + tensor reshape_59_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("reshape_59_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor var_2310_cast_fp16 = add(x = position5_to_fp16, y = valid_len_to_fp16)[name = tensor("op_2310_cast_fp16")]; + tensor var_2310_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_2310_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor var_2310 = cast(dtype = var_2310_cast_fp16_to_fp32_dtype_0, x = var_2310_cast_fp16)[name = tensor("cast_128")]; + tensor new_cache_internal_tensor_assign_2 = cast(dtype = reshape_59_cast_fp16_to_fp32_dtype_0, x = reshape_59_cast_fp16)[name = tensor("cast_129")]; + tensor var_2082 = cast(dtype = var_2082_cast_fp16_to_fp32_dtype_0, x = var_2082_cast_fp16)[name = tensor("cast_133")]; + tensor new_cache_9_internal_tensor_assign_2 = cast(dtype = reshape_49_cast_fp16_to_fp32_dtype_0, x = reshape_49_cast_fp16)[name = tensor("cast_134")]; + tensor var_1671 = cast(dtype = var_1671_cast_fp16_to_fp32_dtype_0, x = var_1671_cast_fp16)[name = tensor("cast_138")]; + tensor new_cache_7_internal_tensor_assign_2 = cast(dtype = reshape_39_cast_fp16_to_fp32_dtype_0, x = reshape_39_cast_fp16)[name = tensor("cast_139")]; + tensor var_1260 = cast(dtype = var_1260_cast_fp16_to_fp32_dtype_0, x = var_1260_cast_fp16)[name = tensor("cast_143")]; + tensor new_cache_5_internal_tensor_assign_2 = cast(dtype = reshape_29_cast_fp16_to_fp32_dtype_0, x = reshape_29_cast_fp16)[name = tensor("cast_144")]; + tensor var_849 = cast(dtype = var_849_cast_fp16_to_fp32_dtype_0, x = var_849_cast_fp16)[name = tensor("cast_148")]; + tensor new_cache_3_internal_tensor_assign_2 = cast(dtype = reshape_19_cast_fp16_to_fp32_dtype_0, x = reshape_19_cast_fp16)[name = tensor("cast_149")]; + tensor var_438 = cast(dtype = var_438_cast_fp16_to_fp32_dtype_0, x = var_438_cast_fp16)[name = tensor("cast_153")]; + tensor new_cache_1_internal_tensor_assign_2 = cast(dtype = reshape_9_cast_fp16_to_fp32_dtype_0, x = reshape_9_cast_fp16)[name = tensor("cast_154")]; + } -> (new_cache_1_internal_tensor_assign_2, var_438, new_cache_3_internal_tensor_assign_2, var_849, new_cache_5_internal_tensor_assign_2, var_1260, new_cache_7_internal_tensor_assign_2, var_1671, new_cache_9_internal_tensor_assign_2, var_2082, new_cache_internal_tensor_assign_2, var_2310); +} \ No newline at end of file diff --git a/v2.1/english/cond_prefill.mlmodelc/weights/weight.bin b/v2.1/english/cond_prefill.mlmodelc/weights/weight.bin new file mode 100644 index 0000000000000000000000000000000000000000..f5168cafdc2860e491e20d31fcc03f11476449e5 --- /dev/null +++ b/v2.1/english/cond_prefill.mlmodelc/weights/weight.bin @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:3a40a5ef3158ef91bfbf932be1c6ee44daca10a856fbb30b9a6720ad9424df8c +size 133237248 diff --git a/v2.1/english/cond_prefill.mlpackage/Data/com.apple.CoreML/model.mlmodel b/v2.1/english/cond_prefill.mlpackage/Data/com.apple.CoreML/model.mlmodel new file mode 100644 index 0000000000000000000000000000000000000000..d16bf7d79ce462f9612ef9e169f473b0c14898e9 --- /dev/null +++ b/v2.1/english/cond_prefill.mlpackage/Data/com.apple.CoreML/model.mlmodel @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:7f897b23ae1602a1db7284d7ce9eb31b6f615e12bf76eda3e9650a356ed89a8a +size 200480 diff --git a/v2.1/english/cond_prefill.mlpackage/Data/com.apple.CoreML/weights/weight.bin b/v2.1/english/cond_prefill.mlpackage/Data/com.apple.CoreML/weights/weight.bin new file mode 100644 index 0000000000000000000000000000000000000000..f5168cafdc2860e491e20d31fcc03f11476449e5 --- /dev/null +++ b/v2.1/english/cond_prefill.mlpackage/Data/com.apple.CoreML/weights/weight.bin @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:3a40a5ef3158ef91bfbf932be1c6ee44daca10a856fbb30b9a6720ad9424df8c +size 133237248 diff --git a/v2.1/english/cond_prefill.mlpackage/Manifest.json b/v2.1/english/cond_prefill.mlpackage/Manifest.json new file mode 100644 index 0000000000000000000000000000000000000000..8f207f224ee128ef98405532d9cccff0d004dc79 --- /dev/null +++ b/v2.1/english/cond_prefill.mlpackage/Manifest.json @@ -0,0 +1,18 @@ +{ + "fileFormatVersion": "1.0.0", + "itemInfoEntries": { + "01FF0335-23EA-4E69-BE20-620770FD067C": { + "author": "com.apple.CoreML", + "description": "CoreML Model Weights", + "name": "weights", + "path": "com.apple.CoreML/weights" + }, + "071D4BA1-DDC5-4C24-801B-076F42B76ACC": { + "author": "com.apple.CoreML", + "description": "CoreML Model Specification", + "name": "model.mlmodel", + "path": "com.apple.CoreML/model.mlmodel" + } + }, + "rootModelIdentifier": "071D4BA1-DDC5-4C24-801B-076F42B76ACC" +} diff --git a/v2.1/english/constants/bos_emb.npy b/v2.1/english/constants/bos_emb.npy new file mode 100644 index 0000000000000000000000000000000000000000..9b7f9d49f38d50b43db1e5fa4c5ab1f906a3ecd0 --- /dev/null +++ b/v2.1/english/constants/bos_emb.npy @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:4ae6121755d51695873cda7822b0959cba18b28543439a478b39d5c51b91e749 +size 256 diff --git a/v2.1/english/constants/emb_mean.npy b/v2.1/english/constants/emb_mean.npy new file mode 100644 index 0000000000000000000000000000000000000000..51004022389fb33665cf968b8b9315d43c02879c --- /dev/null +++ b/v2.1/english/constants/emb_mean.npy @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:600795312716d8776c2a3f438d83c458ab5f4ac64128cd0eb568079d8c9b8e9b +size 256 diff --git a/v2.1/english/constants/emb_std.npy b/v2.1/english/constants/emb_std.npy new file mode 100644 index 0000000000000000000000000000000000000000..47b4c72f446ab075015c8d038c0f24d98828b284 --- /dev/null +++ b/v2.1/english/constants/emb_std.npy @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:638c529cedbeef6c87739f56c5de4d7635a24bc43f13f6fec13dab804d3a208d +size 256 diff --git a/v2.1/english/constants/mimi_init_state.npz b/v2.1/english/constants/mimi_init_state.npz new file mode 100644 index 0000000000000000000000000000000000000000..3b52f383a86ef0439f9993506e4529c4c1234828 --- /dev/null +++ b/v2.1/english/constants/mimi_init_state.npz @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:515588e10e9c14a6b009f1cfda9fcd592542b1640ad856d5fc69e50024702a39 +size 1083108 diff --git a/v2.1/english/constants/quantizer_weight.npy b/v2.1/english/constants/quantizer_weight.npy new file mode 100644 index 0000000000000000000000000000000000000000..8c5776124753f1b4f6145f91832597b4381f6cd9 --- /dev/null +++ b/v2.1/english/constants/quantizer_weight.npy @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:3cb7b1a032c562af193adf9ffdd3366aca8f275da1f48ce0456a122a8c688f03 +size 65664 diff --git a/v2.1/english/constants/text_embed_table.npy b/v2.1/english/constants/text_embed_table.npy new file mode 100644 index 0000000000000000000000000000000000000000..5d3f32d66ad0a7f22655af42bee0df9602857ca2 --- /dev/null +++ b/v2.1/english/constants/text_embed_table.npy @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:0747d8e12ab8798d09be0ee390827b895b6d1babb82037b28aaca03aed8b2203 +size 16388224 diff --git a/v2.1/english/constants_bin/alba.safetensors b/v2.1/english/constants_bin/alba.safetensors new file mode 100644 index 0000000000000000000000000000000000000000..817ef268df79f674a4e3c12027e79b70de46f7fa --- /dev/null +++ b/v2.1/english/constants_bin/alba.safetensors @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:69c32db63ca56843d994f81f343f62e0bf2d73f7e4c9bc73e44bb1110b1d8845 +size 6194424 diff --git a/v2.1/english/constants_bin/anna.safetensors b/v2.1/english/constants_bin/anna.safetensors new file mode 100644 index 0000000000000000000000000000000000000000..abf48da23307ad5e0e5da9cdda647d8ebefbb9a6 --- /dev/null +++ b/v2.1/english/constants_bin/anna.safetensors @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:5ea82f78db006c9fd34e32ddd5aae82674b5b32646097977436458d00af80dfa +size 7816440 diff --git a/v2.1/english/constants_bin/azelma.safetensors b/v2.1/english/constants_bin/azelma.safetensors new file mode 100644 index 0000000000000000000000000000000000000000..1251c8a110324a1b69e6d40edb6baeca6e4d0600 --- /dev/null +++ b/v2.1/english/constants_bin/azelma.safetensors @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:9f3e69f29075f991fd47774566865ef0e0e637cb5a35992c9919761b5b84b1de +size 7963896 diff --git a/v2.1/english/constants_bin/bill_boerst.safetensors b/v2.1/english/constants_bin/bill_boerst.safetensors new file mode 100644 index 0000000000000000000000000000000000000000..a154072f3365807cf340aa6d024970a920a492be --- /dev/null +++ b/v2.1/english/constants_bin/bill_boerst.safetensors @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:75610127d44e0b05b442154f80f89f993df235aecc6cad7070f11000d006c188 +size 6735096 diff --git a/v2.1/english/constants_bin/bos_before_voice.bin b/v2.1/english/constants_bin/bos_before_voice.bin new file mode 100644 index 0000000000000000000000000000000000000000..510dc02b5324d05f4d8c6dcf676ca9241a6a30c8 --- /dev/null +++ b/v2.1/english/constants_bin/bos_before_voice.bin @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:6ac530a104e3f3d2b2ff15f3fad63540f2619b31db78b159c0595a108f142caf +size 4096 diff --git a/v2.1/english/constants_bin/bos_emb.bin b/v2.1/english/constants_bin/bos_emb.bin new file mode 100644 index 0000000000000000000000000000000000000000..89a20dad35b263becd3d5d7b7dd97372ad5ea290 --- /dev/null +++ b/v2.1/english/constants_bin/bos_emb.bin @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:52f1a157fa35fc100213e6e0fafa74703be99db8faa42de3e99feda607ac7b63 +size 128 diff --git a/v2.1/english/constants_bin/caro_davy.safetensors b/v2.1/english/constants_bin/caro_davy.safetensors new file mode 100644 index 0000000000000000000000000000000000000000..edde630b624c3e0d59e2f18081261412fdb3caf1 --- /dev/null +++ b/v2.1/english/constants_bin/caro_davy.safetensors @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:a5961b63a2e7a5cfd7edc383aa9042fb70fd14a9dee6310cdc633881a7f2449a +size 5260536 diff --git a/v2.1/english/constants_bin/charles.safetensors b/v2.1/english/constants_bin/charles.safetensors new file mode 100644 index 0000000000000000000000000000000000000000..9a9c871dc6b956bfed6123cbc09164d136aacdb5 --- /dev/null +++ b/v2.1/english/constants_bin/charles.safetensors @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:299edc20182eeccfbf94e308626f259da4fbf339daa8d5905218f2b1774639b8 +size 6194424 diff --git a/v2.1/english/constants_bin/cosette.safetensors b/v2.1/english/constants_bin/cosette.safetensors new file mode 100644 index 0000000000000000000000000000000000000000..ac272fb887b320313471ce0e80b116109f514b06 --- /dev/null +++ b/v2.1/english/constants_bin/cosette.safetensors @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:c4fdc15f5a3a20c44dd0064a37e87d15d25562936e8dbad7e07b9832015a545d +size 6194424 diff --git a/v2.1/english/constants_bin/eponine.safetensors b/v2.1/english/constants_bin/eponine.safetensors new file mode 100644 index 0000000000000000000000000000000000000000..551899137a06107bf652a6ed4beb155606a8d625 --- /dev/null +++ b/v2.1/english/constants_bin/eponine.safetensors @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:bda3b76a384ff355fe0350736387765946304ae8ca16e59f60ea3296a1c99cc6 +size 6931704 diff --git a/v2.1/english/constants_bin/estelle.safetensors b/v2.1/english/constants_bin/estelle.safetensors new file mode 100644 index 0000000000000000000000000000000000000000..aa90862f9b045a758ce4fa055d9513f7cabe0622 --- /dev/null +++ b/v2.1/english/constants_bin/estelle.safetensors @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:ccebef7f51762c7fc08870f5ecf268e8713551ae2c9f7984ddaec0c1e1c77153 +size 8258808 diff --git a/v2.1/english/constants_bin/eve.safetensors b/v2.1/english/constants_bin/eve.safetensors new file mode 100644 index 0000000000000000000000000000000000000000..d154c65d57fa9ce34a24799d15f2ba5f5777fe18 --- /dev/null +++ b/v2.1/english/constants_bin/eve.safetensors @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:ea9c2faf862a6c9d2cb61910fdf02842ae56940382cc8c1000fdb1b43269692b +size 6538488 diff --git a/v2.1/english/constants_bin/fantine.safetensors b/v2.1/english/constants_bin/fantine.safetensors new file mode 100644 index 0000000000000000000000000000000000000000..ce4051ddb4c6e815ea64380f04fb880c8fe59adc --- /dev/null +++ b/v2.1/english/constants_bin/fantine.safetensors @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:51a8a4355d7f912d4959e4b1918314fda85ad47eba0a33a1d78a4a505d3465f5 +size 6538488 diff --git a/v2.1/english/constants_bin/george.safetensors b/v2.1/english/constants_bin/george.safetensors new file mode 100644 index 0000000000000000000000000000000000000000..15d85dcb153b7767ca5a68e6b318b0f4431c1e3d --- /dev/null +++ b/v2.1/english/constants_bin/george.safetensors @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:0c1c6c57c55a98d81254b33728150c7776f40647fe95258d1a6c1a02780b5d02 +size 6243576 diff --git a/v2.1/english/constants_bin/giovanni.safetensors b/v2.1/english/constants_bin/giovanni.safetensors new file mode 100644 index 0000000000000000000000000000000000000000..10929aed38e3f15eb5d2b3445725b51cbfe05b0d --- /dev/null +++ b/v2.1/english/constants_bin/giovanni.safetensors @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:a5ee718157ec1c6fd9c1e66a7733c7e6337d474a74a7da756455d27717885c59 +size 4621552 diff --git a/v2.1/english/constants_bin/jane.safetensors b/v2.1/english/constants_bin/jane.safetensors new file mode 100644 index 0000000000000000000000000000000000000000..6dbc7613d1ba76ce86aaf07018e6e692e21c98f9 --- /dev/null +++ b/v2.1/english/constants_bin/jane.safetensors @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:37386227ca8ec5bf1b8e516c13d132ce5ff5437a304fe90129a1c62f41d9a008 +size 7374072 diff --git a/v2.1/english/constants_bin/javert.safetensors b/v2.1/english/constants_bin/javert.safetensors new file mode 100644 index 0000000000000000000000000000000000000000..f909190273dd2439ad163e9258fc3fd4cf4370ea --- /dev/null +++ b/v2.1/english/constants_bin/javert.safetensors @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:0ae88e03ca4e76a0e16cbf321a807428febda9d9e9bc0358c02e7f9c9e2c263b +size 6194424 diff --git a/v2.1/english/constants_bin/jean.safetensors b/v2.1/english/constants_bin/jean.safetensors new file mode 100644 index 0000000000000000000000000000000000000000..4876681da3b69083f95f33450ff7204393ac0234 --- /dev/null +++ b/v2.1/english/constants_bin/jean.safetensors @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:90be4b8f50bb4d2dbe27e3fb4e31417cf6a57928931f0a60426a1748821a3d12 +size 6194424 diff --git a/v2.1/english/constants_bin/juergen.safetensors b/v2.1/english/constants_bin/juergen.safetensors new file mode 100644 index 0000000000000000000000000000000000000000..d723c61e32a033a1cfdae60cdf4edaa4ba5dbee3 --- /dev/null +++ b/v2.1/english/constants_bin/juergen.safetensors @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:e74d67b38339fc01118e3bbe2c90d6d40e601b161dbf7be906421956ba80a532 +size 6243576 diff --git a/v2.1/english/constants_bin/lola.safetensors b/v2.1/english/constants_bin/lola.safetensors new file mode 100644 index 0000000000000000000000000000000000000000..bb3983222be592368c8fb4b30e1e20e95cf42c9c --- /dev/null +++ b/v2.1/english/constants_bin/lola.safetensors @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:34972e86b07b17272a8061460054a08119399cace9454979052b9e2d96664e8c +size 5948664 diff --git a/v2.1/english/constants_bin/marius.safetensors b/v2.1/english/constants_bin/marius.safetensors new file mode 100644 index 0000000000000000000000000000000000000000..f9dc2a51b003fd67097d850e4751ea9b1d8ec703 --- /dev/null +++ b/v2.1/english/constants_bin/marius.safetensors @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:04f84efcb77a0547ba582c058db496f7ff4920891d49d37b9950d128422582a8 +size 6194424 diff --git a/v2.1/english/constants_bin/mary.safetensors b/v2.1/english/constants_bin/mary.safetensors new file mode 100644 index 0000000000000000000000000000000000000000..8a39b4a2b80d6ee370b23de9788dac13d8e8e574 --- /dev/null +++ b/v2.1/english/constants_bin/mary.safetensors @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:a8f2adf260cab966fe0a113d6b549d6efdeaa79de544ae0ff34b5b6a41445a59 +size 6194424 diff --git a/v2.1/english/constants_bin/michael.safetensors b/v2.1/english/constants_bin/michael.safetensors new file mode 100644 index 0000000000000000000000000000000000000000..2d9027fc0c4aae18b3a4011987e142bff53b5014 --- /dev/null +++ b/v2.1/english/constants_bin/michael.safetensors @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:8937f724ac4719b9aa51ea0ba1f18f9de0af7a663ad6263558266c1a53c9722d +size 7275768 diff --git a/v2.1/english/constants_bin/paul.safetensors b/v2.1/english/constants_bin/paul.safetensors new file mode 100644 index 0000000000000000000000000000000000000000..c746dd2f3dcd784fb9ad4adae26b8ec3915fc10a --- /dev/null +++ b/v2.1/english/constants_bin/paul.safetensors @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:ed7a019168f94dfe77009f1b0de59387abc6fbb0db954d38ce722ecb77da61aa +size 6980856 diff --git a/v2.1/english/constants_bin/peter_yearsley.safetensors b/v2.1/english/constants_bin/peter_yearsley.safetensors new file mode 100644 index 0000000000000000000000000000000000000000..a2053b8f5811dd02dfa13527041193bdd995b0a1 --- /dev/null +++ b/v2.1/english/constants_bin/peter_yearsley.safetensors @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:dd977a6e15591e347c9a23fa7cc09e35a65b462917f5eeb162baff6dc9e3f685 +size 3736816 diff --git a/v2.1/english/constants_bin/rafael.safetensors b/v2.1/english/constants_bin/rafael.safetensors new file mode 100644 index 0000000000000000000000000000000000000000..1cce35bdee666ce2d1a5d60af81e618bec316299 --- /dev/null +++ b/v2.1/english/constants_bin/rafael.safetensors @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:ac19f099f6cd839875a629c3e2e91e0dfc2c197acf2875db168d0ae244fb58bd +size 6194424 diff --git a/v2.1/english/constants_bin/stuart_bell.safetensors b/v2.1/english/constants_bin/stuart_bell.safetensors new file mode 100644 index 0000000000000000000000000000000000000000..5c3dcf1453a254845fe3e55d8d5ee585fbf86bd2 --- /dev/null +++ b/v2.1/english/constants_bin/stuart_bell.safetensors @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:5a49da7ca5df05d02587ec4a0981c0d318e045f68e24423c4203ce474d9b33dc +size 5260536 diff --git a/v2.1/english/constants_bin/text_embed_table.bin b/v2.1/english/constants_bin/text_embed_table.bin new file mode 100644 index 0000000000000000000000000000000000000000..8a4c039a01c18b2009e5e3580cabde28be7fda38 --- /dev/null +++ b/v2.1/english/constants_bin/text_embed_table.bin @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:2512f9edd82ab0c68bb42a5fcd329a371eaf63ee6fb5275dfb1bc9536592bb0e +size 16388096 diff --git a/v2.1/english/constants_bin/tokenizer.model b/v2.1/english/constants_bin/tokenizer.model new file mode 100644 index 0000000000000000000000000000000000000000..1820a7cbb15efc6a33dd365113c07e3df9d28d80 --- /dev/null +++ b/v2.1/english/constants_bin/tokenizer.model @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:d461765ae179566678c93091c5fa6f2984c31bbe990bf1aa62d92c64d91bc3f6 +size 59339 diff --git a/v2.1/english/constants_bin/vera.safetensors b/v2.1/english/constants_bin/vera.safetensors new file mode 100644 index 0000000000000000000000000000000000000000..ff2699eb04d9a494be126faf11c4211c809fec46 --- /dev/null +++ b/v2.1/english/constants_bin/vera.safetensors @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:4bf50ddd957b5d218b264fdcf18efbbc7384d12da3eca98ca19b9e8dd6976acc +size 6735096 diff --git a/v2.1/english/flow_decoder_fused.mlmodelc/analytics/coremldata.bin b/v2.1/english/flow_decoder_fused.mlmodelc/analytics/coremldata.bin new file mode 100644 index 0000000000000000000000000000000000000000..de5da7392913995a2b334958380c1492877742af --- /dev/null +++ b/v2.1/english/flow_decoder_fused.mlmodelc/analytics/coremldata.bin @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:17cf344fd2be66e934982944a7a252da17087efbc9daac0d3dfb4e8e95797264 +size 243 diff --git a/v2.1/english/flow_decoder_fused.mlmodelc/coremldata.bin b/v2.1/english/flow_decoder_fused.mlmodelc/coremldata.bin new file mode 100644 index 0000000000000000000000000000000000000000..d9abf6be77d7d70d9fee2309055a8e92868c5d79 --- /dev/null +++ b/v2.1/english/flow_decoder_fused.mlmodelc/coremldata.bin @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:69ff9affa07e38b1a511ac8d71c594eeeb236389ddbe74f4fabdfa45db4edf7a +size 413 diff --git a/v2.1/english/flow_decoder_fused.mlmodelc/model.mil b/v2.1/english/flow_decoder_fused.mlmodelc/model.mil new file mode 100644 index 0000000000000000000000000000000000000000..c64a1f34555e7bbf9c8d4502d29f9ef52dd6b0ec --- /dev/null +++ b/v2.1/english/flow_decoder_fused.mlmodelc/model.mil @@ -0,0 +1,1915 @@ +program(1.0) +[buildInfo = dict, tensor>({{"coremlc-component-MIL", "3520.4.1"}, {"coremlc-version", "3520.5.1"}, {"coremltools-component-torch", "2.12.0"}, {"coremltools-source-dialect", "TorchScript"}, {"coremltools-version", "9.0"}})] +{ + func main(tensor latent_init, tensor transformer_out) { + tensor latent_init_to_fp16_dtype_0 = const()[name = tensor("latent_init_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor flow_net_input_proj_weight_to_fp16 = const()[name = tensor("flow_net_input_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(64)))]; + tensor flow_net_input_proj_bias_to_fp16 = const()[name = tensor("flow_net_input_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(32896)))]; + tensor latent_init_to_fp16 = cast(dtype = latent_init_to_fp16_dtype_0, x = latent_init)[name = tensor("cast_146")]; + tensor linear_0_cast_fp16 = linear(bias = flow_net_input_proj_bias_to_fp16, weight = flow_net_input_proj_weight_to_fp16, x = latent_init_to_fp16)[name = tensor("linear_0_cast_fp16")]; + tensor input_3_to_fp16 = const()[name = tensor("input_3_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(33984)))]; + tensor input_5_cast_fp16 = silu(x = input_3_to_fp16)[name = tensor("input_5_cast_fp16")]; + tensor flow_net_time_embed_0_mlp_2_weight_to_fp16 = const()[name = tensor("flow_net_time_embed_0_mlp_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(35072)))]; + tensor flow_net_time_embed_0_mlp_2_bias_to_fp16 = const()[name = tensor("flow_net_time_embed_0_mlp_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(559424)))]; + tensor linear_2_cast_fp16 = linear(bias = flow_net_time_embed_0_mlp_2_bias_to_fp16, weight = flow_net_time_embed_0_mlp_2_weight_to_fp16, x = input_5_cast_fp16)[name = tensor("linear_2_cast_fp16")]; + tensor reduce_mean_0_axes_0 = const()[name = tensor("reduce_mean_0_axes_0"), val = tensor([-1])]; + tensor reduce_mean_0_keep_dims_0 = const()[name = tensor("reduce_mean_0_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_0_cast_fp16 = reduce_mean(axes = reduce_mean_0_axes_0, keep_dims = reduce_mean_0_keep_dims_0, x = linear_2_cast_fp16)[name = tensor("reduce_mean_0_cast_fp16")]; + tensor sub_0_cast_fp16 = sub(x = linear_2_cast_fp16, y = reduce_mean_0_cast_fp16)[name = tensor("sub_0_cast_fp16")]; + tensor square_0_cast_fp16 = square(x = sub_0_cast_fp16)[name = tensor("square_0_cast_fp16")]; + tensor reduce_mean_1_axes_0 = const()[name = tensor("reduce_mean_1_axes_0"), val = tensor([-1])]; + tensor reduce_mean_1_keep_dims_0 = const()[name = tensor("reduce_mean_1_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_1_cast_fp16 = reduce_mean(axes = reduce_mean_1_axes_0, keep_dims = reduce_mean_1_keep_dims_0, x = square_0_cast_fp16)[name = tensor("reduce_mean_1_cast_fp16")]; + tensor real_div_0_to_fp16 = const()[name = tensor("real_div_0_to_fp16"), val = tensor(0x1.008p+0)]; + tensor mul_0_cast_fp16 = mul(x = reduce_mean_1_cast_fp16, y = real_div_0_to_fp16)[name = tensor("mul_0_cast_fp16")]; + tensor var_70_to_fp16 = const()[name = tensor("op_70_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_1_cast_fp16 = add(x = mul_0_cast_fp16, y = var_70_to_fp16)[name = tensor("var_1_cast_fp16")]; + tensor var_73_epsilon_0 = const()[name = tensor("op_73_epsilon_0"), val = tensor(0x1.197998p-40)]; + tensor var_73_cast_fp16 = rsqrt(epsilon = var_73_epsilon_0, x = var_1_cast_fp16)[name = tensor("op_73_cast_fp16")]; + tensor const_3_to_fp16 = const()[name = tensor("const_3_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(560512)))]; + tensor var_74_cast_fp16 = mul(x = const_3_to_fp16, y = var_73_cast_fp16)[name = tensor("op_74_cast_fp16")]; + tensor var_75_cast_fp16 = mul(x = linear_2_cast_fp16, y = var_74_cast_fp16)[name = tensor("op_75_cast_fp16")]; + tensor input_9_to_fp16 = const()[name = tensor("input_9_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(561600)))]; + tensor input_11_cast_fp16 = silu(x = input_9_to_fp16)[name = tensor("input_11_cast_fp16")]; + tensor flow_net_time_embed_1_mlp_2_weight_to_fp16 = const()[name = tensor("flow_net_time_embed_1_mlp_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(562688)))]; + tensor flow_net_time_embed_1_mlp_2_bias_to_fp16 = const()[name = tensor("flow_net_time_embed_1_mlp_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1087040)))]; + tensor linear_4_cast_fp16 = linear(bias = flow_net_time_embed_1_mlp_2_bias_to_fp16, weight = flow_net_time_embed_1_mlp_2_weight_to_fp16, x = input_11_cast_fp16)[name = tensor("linear_4_cast_fp16")]; + tensor reduce_mean_2_axes_0 = const()[name = tensor("reduce_mean_2_axes_0"), val = tensor([-1])]; + tensor reduce_mean_2_keep_dims_0 = const()[name = tensor("reduce_mean_2_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_2_cast_fp16 = reduce_mean(axes = reduce_mean_2_axes_0, keep_dims = reduce_mean_2_keep_dims_0, x = linear_4_cast_fp16)[name = tensor("reduce_mean_2_cast_fp16")]; + tensor sub_2_cast_fp16 = sub(x = linear_4_cast_fp16, y = reduce_mean_2_cast_fp16)[name = tensor("sub_2_cast_fp16")]; + tensor square_1_cast_fp16 = square(x = sub_2_cast_fp16)[name = tensor("square_1_cast_fp16")]; + tensor reduce_mean_3_axes_0 = const()[name = tensor("reduce_mean_3_axes_0"), val = tensor([-1])]; + tensor reduce_mean_3_keep_dims_0 = const()[name = tensor("reduce_mean_3_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_3_cast_fp16 = reduce_mean(axes = reduce_mean_3_axes_0, keep_dims = reduce_mean_3_keep_dims_0, x = square_1_cast_fp16)[name = tensor("reduce_mean_3_cast_fp16")]; + tensor real_div_1_to_fp16 = const()[name = tensor("real_div_1_to_fp16"), val = tensor(0x1.008p+0)]; + tensor mul_1_cast_fp16 = mul(x = reduce_mean_3_cast_fp16, y = real_div_1_to_fp16)[name = tensor("mul_1_cast_fp16")]; + tensor var_110_to_fp16 = const()[name = tensor("op_110_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_3_cast_fp16 = add(x = mul_1_cast_fp16, y = var_110_to_fp16)[name = tensor("var_3_cast_fp16")]; + tensor var_113_epsilon_0 = const()[name = tensor("op_113_epsilon_0"), val = tensor(0x1.197998p-40)]; + tensor var_113_cast_fp16 = rsqrt(epsilon = var_113_epsilon_0, x = var_3_cast_fp16)[name = tensor("op_113_cast_fp16")]; + tensor const_5_to_fp16 = const()[name = tensor("const_5_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1088128)))]; + tensor var_114_cast_fp16 = mul(x = const_5_to_fp16, y = var_113_cast_fp16)[name = tensor("op_114_cast_fp16")]; + tensor var_115_cast_fp16 = mul(x = linear_4_cast_fp16, y = var_114_cast_fp16)[name = tensor("op_115_cast_fp16")]; + tensor var_127_cast_fp16 = add(x = var_75_cast_fp16, y = var_115_cast_fp16)[name = tensor("op_127_cast_fp16")]; + tensor _inversed_t_combined_1_y_0_to_fp16 = const()[name = tensor("_inversed_t_combined_1_y_0_to_fp16"), val = tensor(0x1p-1)]; + tensor _inversed_t_combined_1_cast_fp16 = mul(x = var_127_cast_fp16, y = _inversed_t_combined_1_y_0_to_fp16)[name = tensor("_inversed_t_combined_1_cast_fp16")]; + tensor transformer_out_to_fp16_dtype_0 = const()[name = tensor("transformer_out_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor flow_net_cond_embed_weight_to_fp16 = const()[name = tensor("flow_net_cond_embed_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1089216)))]; + tensor flow_net_cond_embed_bias_to_fp16 = const()[name = tensor("flow_net_cond_embed_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(2137856)))]; + tensor transformer_out_to_fp16 = cast(dtype = transformer_out_to_fp16_dtype_0, x = transformer_out)[name = tensor("cast_145")]; + tensor linear_5_cast_fp16 = linear(bias = flow_net_cond_embed_bias_to_fp16, weight = flow_net_cond_embed_weight_to_fp16, x = transformer_out_to_fp16)[name = tensor("linear_5_cast_fp16")]; + tensor input_13_cast_fp16 = add(x = _inversed_t_combined_1_cast_fp16, y = linear_5_cast_fp16)[name = tensor("input_13_cast_fp16")]; + tensor input_15_cast_fp16 = silu(x = input_13_cast_fp16)[name = tensor("input_15_cast_fp16")]; + tensor flow_net_res_blocks_0_adaLN_modulation_1_weight_to_fp16 = const()[name = tensor("flow_net_res_blocks_0_adaLN_modulation_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(2138944)))]; + tensor flow_net_res_blocks_0_adaLN_modulation_1_bias_to_fp16 = const()[name = tensor("flow_net_res_blocks_0_adaLN_modulation_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(3711872)))]; + tensor linear_6_cast_fp16 = linear(bias = flow_net_res_blocks_0_adaLN_modulation_1_bias_to_fp16, weight = flow_net_res_blocks_0_adaLN_modulation_1_weight_to_fp16, x = input_15_cast_fp16)[name = tensor("linear_6_cast_fp16")]; + tensor var_142_split_sizes_0 = const()[name = tensor("op_142_split_sizes_0"), val = tensor([512, 512, 512])]; + tensor var_142_axis_0 = const()[name = tensor("op_142_axis_0"), val = tensor(-1)]; + tensor var_142_cast_fp16_0, tensor var_142_cast_fp16_1, tensor var_142_cast_fp16_2 = split(axis = var_142_axis_0, split_sizes = var_142_split_sizes_0, x = linear_6_cast_fp16)[name = tensor("op_142_cast_fp16")]; + tensor mean_1_axes_0 = const()[name = tensor("mean_1_axes_0"), val = tensor([-1])]; + tensor mean_1_keep_dims_0 = const()[name = tensor("mean_1_keep_dims_0"), val = tensor(true)]; + tensor mean_1_cast_fp16 = reduce_mean(axes = mean_1_axes_0, keep_dims = mean_1_keep_dims_0, x = linear_0_cast_fp16)[name = tensor("mean_1_cast_fp16")]; + tensor sub_4_cast_fp16 = sub(x = linear_0_cast_fp16, y = mean_1_cast_fp16)[name = tensor("sub_4_cast_fp16")]; + tensor square_2_cast_fp16 = square(x = sub_4_cast_fp16)[name = tensor("square_2_cast_fp16")]; + tensor reduce_mean_5_axes_0 = const()[name = tensor("reduce_mean_5_axes_0"), val = tensor([-1])]; + tensor reduce_mean_5_keep_dims_0 = const()[name = tensor("reduce_mean_5_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_5_cast_fp16 = reduce_mean(axes = reduce_mean_5_axes_0, keep_dims = reduce_mean_5_keep_dims_0, x = square_2_cast_fp16)[name = tensor("reduce_mean_5_cast_fp16")]; + tensor var_152_to_fp16 = const()[name = tensor("op_152_to_fp16"), val = tensor(0x1.1p-20)]; + tensor var_153_cast_fp16 = add(x = reduce_mean_5_cast_fp16, y = var_152_to_fp16)[name = tensor("op_153_cast_fp16")]; + tensor var_154_cast_fp16 = sqrt(x = var_153_cast_fp16)[name = tensor("op_154_cast_fp16")]; + tensor x_7_cast_fp16 = real_div(x = sub_4_cast_fp16, y = var_154_cast_fp16)[name = tensor("x_7_cast_fp16")]; + tensor flow_net_res_blocks_0_in_ln_weight_to_fp16 = const()[name = tensor("flow_net_res_blocks_0_in_ln_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(3715008)))]; + tensor var_156_cast_fp16 = mul(x = x_7_cast_fp16, y = flow_net_res_blocks_0_in_ln_weight_to_fp16)[name = tensor("op_156_cast_fp16")]; + tensor flow_net_res_blocks_0_in_ln_bias_to_fp16 = const()[name = tensor("flow_net_res_blocks_0_in_ln_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(3716096)))]; + tensor x_9_cast_fp16 = add(x = var_156_cast_fp16, y = flow_net_res_blocks_0_in_ln_bias_to_fp16)[name = tensor("x_9_cast_fp16")]; + tensor var_158_promoted_to_fp16 = const()[name = tensor("op_158_promoted_to_fp16"), val = tensor(0x1p+0)]; + tensor var_159_cast_fp16 = add(x = var_142_cast_fp16_1, y = var_158_promoted_to_fp16)[name = tensor("op_159_cast_fp16")]; + tensor var_160_cast_fp16 = mul(x = x_9_cast_fp16, y = var_159_cast_fp16)[name = tensor("op_160_cast_fp16")]; + tensor input_17_cast_fp16 = add(x = var_160_cast_fp16, y = var_142_cast_fp16_0)[name = tensor("input_17_cast_fp16")]; + tensor flow_net_res_blocks_0_mlp_0_weight_to_fp16 = const()[name = tensor("flow_net_res_blocks_0_mlp_0_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(3717184)))]; + tensor flow_net_res_blocks_0_mlp_0_bias_to_fp16 = const()[name = tensor("flow_net_res_blocks_0_mlp_0_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(4241536)))]; + tensor linear_7_cast_fp16 = linear(bias = flow_net_res_blocks_0_mlp_0_bias_to_fp16, weight = flow_net_res_blocks_0_mlp_0_weight_to_fp16, x = input_17_cast_fp16)[name = tensor("linear_7_cast_fp16")]; + tensor input_21_cast_fp16 = silu(x = linear_7_cast_fp16)[name = tensor("input_21_cast_fp16")]; + tensor flow_net_res_blocks_0_mlp_2_weight_to_fp16 = const()[name = tensor("flow_net_res_blocks_0_mlp_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(4242624)))]; + tensor flow_net_res_blocks_0_mlp_2_bias_to_fp16 = const()[name = tensor("flow_net_res_blocks_0_mlp_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(4766976)))]; + tensor linear_8_cast_fp16 = linear(bias = flow_net_res_blocks_0_mlp_2_bias_to_fp16, weight = flow_net_res_blocks_0_mlp_2_weight_to_fp16, x = input_21_cast_fp16)[name = tensor("linear_8_cast_fp16")]; + tensor var_171_cast_fp16 = mul(x = var_142_cast_fp16_2, y = linear_8_cast_fp16)[name = tensor("op_171_cast_fp16")]; + tensor x_11_cast_fp16 = add(x = linear_0_cast_fp16, y = var_171_cast_fp16)[name = tensor("x_11_cast_fp16")]; + tensor flow_net_res_blocks_1_adaLN_modulation_1_weight_to_fp16 = const()[name = tensor("flow_net_res_blocks_1_adaLN_modulation_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(4768064)))]; + tensor flow_net_res_blocks_1_adaLN_modulation_1_bias_to_fp16 = const()[name = tensor("flow_net_res_blocks_1_adaLN_modulation_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(6340992)))]; + tensor linear_9_cast_fp16 = linear(bias = flow_net_res_blocks_1_adaLN_modulation_1_bias_to_fp16, weight = flow_net_res_blocks_1_adaLN_modulation_1_weight_to_fp16, x = input_15_cast_fp16)[name = tensor("linear_9_cast_fp16")]; + tensor var_181_split_sizes_0 = const()[name = tensor("op_181_split_sizes_0"), val = tensor([512, 512, 512])]; + tensor var_181_axis_0 = const()[name = tensor("op_181_axis_0"), val = tensor(-1)]; + tensor var_181_cast_fp16_0, tensor var_181_cast_fp16_1, tensor var_181_cast_fp16_2 = split(axis = var_181_axis_0, split_sizes = var_181_split_sizes_0, x = linear_9_cast_fp16)[name = tensor("op_181_cast_fp16")]; + tensor mean_3_axes_0 = const()[name = tensor("mean_3_axes_0"), val = tensor([-1])]; + tensor mean_3_keep_dims_0 = const()[name = tensor("mean_3_keep_dims_0"), val = tensor(true)]; + tensor mean_3_cast_fp16 = reduce_mean(axes = mean_3_axes_0, keep_dims = mean_3_keep_dims_0, x = x_11_cast_fp16)[name = tensor("mean_3_cast_fp16")]; + tensor sub_5_cast_fp16 = sub(x = x_11_cast_fp16, y = mean_3_cast_fp16)[name = tensor("sub_5_cast_fp16")]; + tensor square_3_cast_fp16 = square(x = sub_5_cast_fp16)[name = tensor("square_3_cast_fp16")]; + tensor reduce_mean_7_axes_0 = const()[name = tensor("reduce_mean_7_axes_0"), val = tensor([-1])]; + tensor reduce_mean_7_keep_dims_0 = const()[name = tensor("reduce_mean_7_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_7_cast_fp16 = reduce_mean(axes = reduce_mean_7_axes_0, keep_dims = reduce_mean_7_keep_dims_0, x = square_3_cast_fp16)[name = tensor("reduce_mean_7_cast_fp16")]; + tensor var_191_to_fp16 = const()[name = tensor("op_191_to_fp16"), val = tensor(0x1.1p-20)]; + tensor var_192_cast_fp16 = add(x = reduce_mean_7_cast_fp16, y = var_191_to_fp16)[name = tensor("op_192_cast_fp16")]; + tensor var_193_cast_fp16 = sqrt(x = var_192_cast_fp16)[name = tensor("op_193_cast_fp16")]; + tensor x_13_cast_fp16 = real_div(x = sub_5_cast_fp16, y = var_193_cast_fp16)[name = tensor("x_13_cast_fp16")]; + tensor flow_net_res_blocks_1_in_ln_weight_to_fp16 = const()[name = tensor("flow_net_res_blocks_1_in_ln_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(6344128)))]; + tensor var_195_cast_fp16 = mul(x = x_13_cast_fp16, y = flow_net_res_blocks_1_in_ln_weight_to_fp16)[name = tensor("op_195_cast_fp16")]; + tensor flow_net_res_blocks_1_in_ln_bias_to_fp16 = const()[name = tensor("flow_net_res_blocks_1_in_ln_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(6345216)))]; + tensor x_15_cast_fp16 = add(x = var_195_cast_fp16, y = flow_net_res_blocks_1_in_ln_bias_to_fp16)[name = tensor("x_15_cast_fp16")]; + tensor var_197_promoted_to_fp16 = const()[name = tensor("op_197_promoted_to_fp16"), val = tensor(0x1p+0)]; + tensor var_198_cast_fp16 = add(x = var_181_cast_fp16_1, y = var_197_promoted_to_fp16)[name = tensor("op_198_cast_fp16")]; + tensor var_199_cast_fp16 = mul(x = x_15_cast_fp16, y = var_198_cast_fp16)[name = tensor("op_199_cast_fp16")]; + tensor input_25_cast_fp16 = add(x = var_199_cast_fp16, y = var_181_cast_fp16_0)[name = tensor("input_25_cast_fp16")]; + tensor flow_net_res_blocks_1_mlp_0_weight_to_fp16 = const()[name = tensor("flow_net_res_blocks_1_mlp_0_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(6346304)))]; + tensor flow_net_res_blocks_1_mlp_0_bias_to_fp16 = const()[name = tensor("flow_net_res_blocks_1_mlp_0_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(6870656)))]; + tensor linear_10_cast_fp16 = linear(bias = flow_net_res_blocks_1_mlp_0_bias_to_fp16, weight = flow_net_res_blocks_1_mlp_0_weight_to_fp16, x = input_25_cast_fp16)[name = tensor("linear_10_cast_fp16")]; + tensor input_29_cast_fp16 = silu(x = linear_10_cast_fp16)[name = tensor("input_29_cast_fp16")]; + tensor flow_net_res_blocks_1_mlp_2_weight_to_fp16 = const()[name = tensor("flow_net_res_blocks_1_mlp_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(6871744)))]; + tensor flow_net_res_blocks_1_mlp_2_bias_to_fp16 = const()[name = tensor("flow_net_res_blocks_1_mlp_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(7396096)))]; + tensor linear_11_cast_fp16 = linear(bias = flow_net_res_blocks_1_mlp_2_bias_to_fp16, weight = flow_net_res_blocks_1_mlp_2_weight_to_fp16, x = input_29_cast_fp16)[name = tensor("linear_11_cast_fp16")]; + tensor var_210_cast_fp16 = mul(x = var_181_cast_fp16_2, y = linear_11_cast_fp16)[name = tensor("op_210_cast_fp16")]; + tensor x_17_cast_fp16 = add(x = x_11_cast_fp16, y = var_210_cast_fp16)[name = tensor("x_17_cast_fp16")]; + tensor flow_net_res_blocks_2_adaLN_modulation_1_weight_to_fp16 = const()[name = tensor("flow_net_res_blocks_2_adaLN_modulation_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(7397184)))]; + tensor flow_net_res_blocks_2_adaLN_modulation_1_bias_to_fp16 = const()[name = tensor("flow_net_res_blocks_2_adaLN_modulation_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8970112)))]; + tensor linear_12_cast_fp16 = linear(bias = flow_net_res_blocks_2_adaLN_modulation_1_bias_to_fp16, weight = flow_net_res_blocks_2_adaLN_modulation_1_weight_to_fp16, x = input_15_cast_fp16)[name = tensor("linear_12_cast_fp16")]; + tensor var_220_split_sizes_0 = const()[name = tensor("op_220_split_sizes_0"), val = tensor([512, 512, 512])]; + tensor var_220_axis_0 = const()[name = tensor("op_220_axis_0"), val = tensor(-1)]; + tensor var_220_cast_fp16_0, tensor var_220_cast_fp16_1, tensor var_220_cast_fp16_2 = split(axis = var_220_axis_0, split_sizes = var_220_split_sizes_0, x = linear_12_cast_fp16)[name = tensor("op_220_cast_fp16")]; + tensor mean_5_axes_0 = const()[name = tensor("mean_5_axes_0"), val = tensor([-1])]; + tensor mean_5_keep_dims_0 = const()[name = tensor("mean_5_keep_dims_0"), val = tensor(true)]; + tensor mean_5_cast_fp16 = reduce_mean(axes = mean_5_axes_0, keep_dims = mean_5_keep_dims_0, x = x_17_cast_fp16)[name = tensor("mean_5_cast_fp16")]; + tensor sub_6_cast_fp16 = sub(x = x_17_cast_fp16, y = mean_5_cast_fp16)[name = tensor("sub_6_cast_fp16")]; + tensor square_4_cast_fp16 = square(x = sub_6_cast_fp16)[name = tensor("square_4_cast_fp16")]; + tensor reduce_mean_9_axes_0 = const()[name = tensor("reduce_mean_9_axes_0"), val = tensor([-1])]; + tensor reduce_mean_9_keep_dims_0 = const()[name = tensor("reduce_mean_9_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_9_cast_fp16 = reduce_mean(axes = reduce_mean_9_axes_0, keep_dims = reduce_mean_9_keep_dims_0, x = square_4_cast_fp16)[name = tensor("reduce_mean_9_cast_fp16")]; + tensor var_230_to_fp16 = const()[name = tensor("op_230_to_fp16"), val = tensor(0x1.1p-20)]; + tensor var_231_cast_fp16 = add(x = reduce_mean_9_cast_fp16, y = var_230_to_fp16)[name = tensor("op_231_cast_fp16")]; + tensor var_232_cast_fp16 = sqrt(x = var_231_cast_fp16)[name = tensor("op_232_cast_fp16")]; + tensor x_19_cast_fp16 = real_div(x = sub_6_cast_fp16, y = var_232_cast_fp16)[name = tensor("x_19_cast_fp16")]; + tensor flow_net_res_blocks_2_in_ln_weight_to_fp16 = const()[name = tensor("flow_net_res_blocks_2_in_ln_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8973248)))]; + tensor var_234_cast_fp16 = mul(x = x_19_cast_fp16, y = flow_net_res_blocks_2_in_ln_weight_to_fp16)[name = tensor("op_234_cast_fp16")]; + tensor flow_net_res_blocks_2_in_ln_bias_to_fp16 = const()[name = tensor("flow_net_res_blocks_2_in_ln_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8974336)))]; + tensor x_21_cast_fp16 = add(x = var_234_cast_fp16, y = flow_net_res_blocks_2_in_ln_bias_to_fp16)[name = tensor("x_21_cast_fp16")]; + tensor var_236_promoted_to_fp16 = const()[name = tensor("op_236_promoted_to_fp16"), val = tensor(0x1p+0)]; + tensor var_237_cast_fp16 = add(x = var_220_cast_fp16_1, y = var_236_promoted_to_fp16)[name = tensor("op_237_cast_fp16")]; + tensor var_238_cast_fp16 = mul(x = x_21_cast_fp16, y = var_237_cast_fp16)[name = tensor("op_238_cast_fp16")]; + tensor input_33_cast_fp16 = add(x = var_238_cast_fp16, y = var_220_cast_fp16_0)[name = tensor("input_33_cast_fp16")]; + tensor flow_net_res_blocks_2_mlp_0_weight_to_fp16 = const()[name = tensor("flow_net_res_blocks_2_mlp_0_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8975424)))]; + tensor flow_net_res_blocks_2_mlp_0_bias_to_fp16 = const()[name = tensor("flow_net_res_blocks_2_mlp_0_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9499776)))]; + tensor linear_13_cast_fp16 = linear(bias = flow_net_res_blocks_2_mlp_0_bias_to_fp16, weight = flow_net_res_blocks_2_mlp_0_weight_to_fp16, x = input_33_cast_fp16)[name = tensor("linear_13_cast_fp16")]; + tensor input_37_cast_fp16 = silu(x = linear_13_cast_fp16)[name = tensor("input_37_cast_fp16")]; + tensor flow_net_res_blocks_2_mlp_2_weight_to_fp16 = const()[name = tensor("flow_net_res_blocks_2_mlp_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9500864)))]; + tensor flow_net_res_blocks_2_mlp_2_bias_to_fp16 = const()[name = tensor("flow_net_res_blocks_2_mlp_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(10025216)))]; + tensor linear_14_cast_fp16 = linear(bias = flow_net_res_blocks_2_mlp_2_bias_to_fp16, weight = flow_net_res_blocks_2_mlp_2_weight_to_fp16, x = input_37_cast_fp16)[name = tensor("linear_14_cast_fp16")]; + tensor var_249_cast_fp16 = mul(x = var_220_cast_fp16_2, y = linear_14_cast_fp16)[name = tensor("op_249_cast_fp16")]; + tensor x_23_cast_fp16 = add(x = x_17_cast_fp16, y = var_249_cast_fp16)[name = tensor("x_23_cast_fp16")]; + tensor flow_net_res_blocks_3_adaLN_modulation_1_weight_to_fp16 = const()[name = tensor("flow_net_res_blocks_3_adaLN_modulation_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(10026304)))]; + tensor flow_net_res_blocks_3_adaLN_modulation_1_bias_to_fp16 = const()[name = tensor("flow_net_res_blocks_3_adaLN_modulation_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(11599232)))]; + tensor linear_15_cast_fp16 = linear(bias = flow_net_res_blocks_3_adaLN_modulation_1_bias_to_fp16, weight = flow_net_res_blocks_3_adaLN_modulation_1_weight_to_fp16, x = input_15_cast_fp16)[name = tensor("linear_15_cast_fp16")]; + tensor var_259_split_sizes_0 = const()[name = tensor("op_259_split_sizes_0"), val = tensor([512, 512, 512])]; + tensor var_259_axis_0 = const()[name = tensor("op_259_axis_0"), val = tensor(-1)]; + tensor var_259_cast_fp16_0, tensor var_259_cast_fp16_1, tensor var_259_cast_fp16_2 = split(axis = var_259_axis_0, split_sizes = var_259_split_sizes_0, x = linear_15_cast_fp16)[name = tensor("op_259_cast_fp16")]; + tensor mean_7_axes_0 = const()[name = tensor("mean_7_axes_0"), val = tensor([-1])]; + tensor mean_7_keep_dims_0 = const()[name = tensor("mean_7_keep_dims_0"), val = tensor(true)]; + tensor mean_7_cast_fp16 = reduce_mean(axes = mean_7_axes_0, keep_dims = mean_7_keep_dims_0, x = x_23_cast_fp16)[name = tensor("mean_7_cast_fp16")]; + tensor sub_7_cast_fp16 = sub(x = x_23_cast_fp16, y = mean_7_cast_fp16)[name = tensor("sub_7_cast_fp16")]; + tensor square_5_cast_fp16 = square(x = sub_7_cast_fp16)[name = tensor("square_5_cast_fp16")]; + tensor reduce_mean_11_axes_0 = const()[name = tensor("reduce_mean_11_axes_0"), val = tensor([-1])]; + tensor reduce_mean_11_keep_dims_0 = const()[name = tensor("reduce_mean_11_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_11_cast_fp16 = reduce_mean(axes = reduce_mean_11_axes_0, keep_dims = reduce_mean_11_keep_dims_0, x = square_5_cast_fp16)[name = tensor("reduce_mean_11_cast_fp16")]; + tensor var_269_to_fp16 = const()[name = tensor("op_269_to_fp16"), val = tensor(0x1.1p-20)]; + tensor var_270_cast_fp16 = add(x = reduce_mean_11_cast_fp16, y = var_269_to_fp16)[name = tensor("op_270_cast_fp16")]; + tensor var_271_cast_fp16 = sqrt(x = var_270_cast_fp16)[name = tensor("op_271_cast_fp16")]; + tensor x_25_cast_fp16 = real_div(x = sub_7_cast_fp16, y = var_271_cast_fp16)[name = tensor("x_25_cast_fp16")]; + tensor flow_net_res_blocks_3_in_ln_weight_to_fp16 = const()[name = tensor("flow_net_res_blocks_3_in_ln_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(11602368)))]; + tensor var_273_cast_fp16 = mul(x = x_25_cast_fp16, y = flow_net_res_blocks_3_in_ln_weight_to_fp16)[name = tensor("op_273_cast_fp16")]; + tensor flow_net_res_blocks_3_in_ln_bias_to_fp16 = const()[name = tensor("flow_net_res_blocks_3_in_ln_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(11603456)))]; + tensor x_27_cast_fp16 = add(x = var_273_cast_fp16, y = flow_net_res_blocks_3_in_ln_bias_to_fp16)[name = tensor("x_27_cast_fp16")]; + tensor var_275_promoted_to_fp16 = const()[name = tensor("op_275_promoted_to_fp16"), val = tensor(0x1p+0)]; + tensor var_276_cast_fp16 = add(x = var_259_cast_fp16_1, y = var_275_promoted_to_fp16)[name = tensor("op_276_cast_fp16")]; + tensor var_277_cast_fp16 = mul(x = x_27_cast_fp16, y = var_276_cast_fp16)[name = tensor("op_277_cast_fp16")]; + tensor input_41_cast_fp16 = add(x = var_277_cast_fp16, y = var_259_cast_fp16_0)[name = tensor("input_41_cast_fp16")]; + tensor flow_net_res_blocks_3_mlp_0_weight_to_fp16 = const()[name = tensor("flow_net_res_blocks_3_mlp_0_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(11604544)))]; + tensor flow_net_res_blocks_3_mlp_0_bias_to_fp16 = const()[name = tensor("flow_net_res_blocks_3_mlp_0_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(12128896)))]; + tensor linear_16_cast_fp16 = linear(bias = flow_net_res_blocks_3_mlp_0_bias_to_fp16, weight = flow_net_res_blocks_3_mlp_0_weight_to_fp16, x = input_41_cast_fp16)[name = tensor("linear_16_cast_fp16")]; + tensor input_45_cast_fp16 = silu(x = linear_16_cast_fp16)[name = tensor("input_45_cast_fp16")]; + tensor flow_net_res_blocks_3_mlp_2_weight_to_fp16 = const()[name = tensor("flow_net_res_blocks_3_mlp_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(12129984)))]; + tensor flow_net_res_blocks_3_mlp_2_bias_to_fp16 = const()[name = tensor("flow_net_res_blocks_3_mlp_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(12654336)))]; + tensor linear_17_cast_fp16 = linear(bias = flow_net_res_blocks_3_mlp_2_bias_to_fp16, weight = flow_net_res_blocks_3_mlp_2_weight_to_fp16, x = input_45_cast_fp16)[name = tensor("linear_17_cast_fp16")]; + tensor var_288_cast_fp16 = mul(x = var_259_cast_fp16_2, y = linear_17_cast_fp16)[name = tensor("op_288_cast_fp16")]; + tensor x_29_cast_fp16 = add(x = x_23_cast_fp16, y = var_288_cast_fp16)[name = tensor("x_29_cast_fp16")]; + tensor flow_net_res_blocks_4_adaLN_modulation_1_weight_to_fp16 = const()[name = tensor("flow_net_res_blocks_4_adaLN_modulation_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(12655424)))]; + tensor flow_net_res_blocks_4_adaLN_modulation_1_bias_to_fp16 = const()[name = tensor("flow_net_res_blocks_4_adaLN_modulation_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(14228352)))]; + tensor linear_18_cast_fp16 = linear(bias = flow_net_res_blocks_4_adaLN_modulation_1_bias_to_fp16, weight = flow_net_res_blocks_4_adaLN_modulation_1_weight_to_fp16, x = input_15_cast_fp16)[name = tensor("linear_18_cast_fp16")]; + tensor var_298_split_sizes_0 = const()[name = tensor("op_298_split_sizes_0"), val = tensor([512, 512, 512])]; + tensor var_298_axis_0 = const()[name = tensor("op_298_axis_0"), val = tensor(-1)]; + tensor var_298_cast_fp16_0, tensor var_298_cast_fp16_1, tensor var_298_cast_fp16_2 = split(axis = var_298_axis_0, split_sizes = var_298_split_sizes_0, x = linear_18_cast_fp16)[name = tensor("op_298_cast_fp16")]; + tensor mean_9_axes_0 = const()[name = tensor("mean_9_axes_0"), val = tensor([-1])]; + tensor mean_9_keep_dims_0 = const()[name = tensor("mean_9_keep_dims_0"), val = tensor(true)]; + tensor mean_9_cast_fp16 = reduce_mean(axes = mean_9_axes_0, keep_dims = mean_9_keep_dims_0, x = x_29_cast_fp16)[name = tensor("mean_9_cast_fp16")]; + tensor sub_8_cast_fp16 = sub(x = x_29_cast_fp16, y = mean_9_cast_fp16)[name = tensor("sub_8_cast_fp16")]; + tensor square_6_cast_fp16 = square(x = sub_8_cast_fp16)[name = tensor("square_6_cast_fp16")]; + tensor reduce_mean_13_axes_0 = const()[name = tensor("reduce_mean_13_axes_0"), val = tensor([-1])]; + tensor reduce_mean_13_keep_dims_0 = const()[name = tensor("reduce_mean_13_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_13_cast_fp16 = reduce_mean(axes = reduce_mean_13_axes_0, keep_dims = reduce_mean_13_keep_dims_0, x = square_6_cast_fp16)[name = tensor("reduce_mean_13_cast_fp16")]; + tensor var_308_to_fp16 = const()[name = tensor("op_308_to_fp16"), val = tensor(0x1.1p-20)]; + tensor var_309_cast_fp16 = add(x = reduce_mean_13_cast_fp16, y = var_308_to_fp16)[name = tensor("op_309_cast_fp16")]; + tensor var_310_cast_fp16 = sqrt(x = var_309_cast_fp16)[name = tensor("op_310_cast_fp16")]; + tensor x_31_cast_fp16 = real_div(x = sub_8_cast_fp16, y = var_310_cast_fp16)[name = tensor("x_31_cast_fp16")]; + tensor flow_net_res_blocks_4_in_ln_weight_to_fp16 = const()[name = tensor("flow_net_res_blocks_4_in_ln_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(14231488)))]; + tensor var_312_cast_fp16 = mul(x = x_31_cast_fp16, y = flow_net_res_blocks_4_in_ln_weight_to_fp16)[name = tensor("op_312_cast_fp16")]; + tensor flow_net_res_blocks_4_in_ln_bias_to_fp16 = const()[name = tensor("flow_net_res_blocks_4_in_ln_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(14232576)))]; + tensor x_33_cast_fp16 = add(x = var_312_cast_fp16, y = flow_net_res_blocks_4_in_ln_bias_to_fp16)[name = tensor("x_33_cast_fp16")]; + tensor var_314_promoted_to_fp16 = const()[name = tensor("op_314_promoted_to_fp16"), val = tensor(0x1p+0)]; + tensor var_315_cast_fp16 = add(x = var_298_cast_fp16_1, y = var_314_promoted_to_fp16)[name = tensor("op_315_cast_fp16")]; + tensor var_316_cast_fp16 = mul(x = x_33_cast_fp16, y = var_315_cast_fp16)[name = tensor("op_316_cast_fp16")]; + tensor input_49_cast_fp16 = add(x = var_316_cast_fp16, y = var_298_cast_fp16_0)[name = tensor("input_49_cast_fp16")]; + tensor flow_net_res_blocks_4_mlp_0_weight_to_fp16 = const()[name = tensor("flow_net_res_blocks_4_mlp_0_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(14233664)))]; + tensor flow_net_res_blocks_4_mlp_0_bias_to_fp16 = const()[name = tensor("flow_net_res_blocks_4_mlp_0_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(14758016)))]; + tensor linear_19_cast_fp16 = linear(bias = flow_net_res_blocks_4_mlp_0_bias_to_fp16, weight = flow_net_res_blocks_4_mlp_0_weight_to_fp16, x = input_49_cast_fp16)[name = tensor("linear_19_cast_fp16")]; + tensor input_53_cast_fp16 = silu(x = linear_19_cast_fp16)[name = tensor("input_53_cast_fp16")]; + tensor flow_net_res_blocks_4_mlp_2_weight_to_fp16 = const()[name = tensor("flow_net_res_blocks_4_mlp_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(14759104)))]; + tensor flow_net_res_blocks_4_mlp_2_bias_to_fp16 = const()[name = tensor("flow_net_res_blocks_4_mlp_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15283456)))]; + tensor linear_20_cast_fp16 = linear(bias = flow_net_res_blocks_4_mlp_2_bias_to_fp16, weight = flow_net_res_blocks_4_mlp_2_weight_to_fp16, x = input_53_cast_fp16)[name = tensor("linear_20_cast_fp16")]; + tensor var_327_cast_fp16 = mul(x = var_298_cast_fp16_2, y = linear_20_cast_fp16)[name = tensor("op_327_cast_fp16")]; + tensor x_35_cast_fp16 = add(x = x_29_cast_fp16, y = var_327_cast_fp16)[name = tensor("x_35_cast_fp16")]; + tensor flow_net_res_blocks_5_adaLN_modulation_1_weight_to_fp16 = const()[name = tensor("flow_net_res_blocks_5_adaLN_modulation_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15284544)))]; + tensor flow_net_res_blocks_5_adaLN_modulation_1_bias_to_fp16 = const()[name = tensor("flow_net_res_blocks_5_adaLN_modulation_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(16857472)))]; + tensor linear_21_cast_fp16 = linear(bias = flow_net_res_blocks_5_adaLN_modulation_1_bias_to_fp16, weight = flow_net_res_blocks_5_adaLN_modulation_1_weight_to_fp16, x = input_15_cast_fp16)[name = tensor("linear_21_cast_fp16")]; + tensor var_337_split_sizes_0 = const()[name = tensor("op_337_split_sizes_0"), val = tensor([512, 512, 512])]; + tensor var_337_axis_0 = const()[name = tensor("op_337_axis_0"), val = tensor(-1)]; + tensor var_337_cast_fp16_0, tensor var_337_cast_fp16_1, tensor var_337_cast_fp16_2 = split(axis = var_337_axis_0, split_sizes = var_337_split_sizes_0, x = linear_21_cast_fp16)[name = tensor("op_337_cast_fp16")]; + tensor mean_11_axes_0 = const()[name = tensor("mean_11_axes_0"), val = tensor([-1])]; + tensor mean_11_keep_dims_0 = const()[name = tensor("mean_11_keep_dims_0"), val = tensor(true)]; + tensor mean_11_cast_fp16 = reduce_mean(axes = mean_11_axes_0, keep_dims = mean_11_keep_dims_0, x = x_35_cast_fp16)[name = tensor("mean_11_cast_fp16")]; + tensor sub_9_cast_fp16 = sub(x = x_35_cast_fp16, y = mean_11_cast_fp16)[name = tensor("sub_9_cast_fp16")]; + tensor square_7_cast_fp16 = square(x = sub_9_cast_fp16)[name = tensor("square_7_cast_fp16")]; + tensor reduce_mean_15_axes_0 = const()[name = tensor("reduce_mean_15_axes_0"), val = tensor([-1])]; + tensor reduce_mean_15_keep_dims_0 = const()[name = tensor("reduce_mean_15_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_15_cast_fp16 = reduce_mean(axes = reduce_mean_15_axes_0, keep_dims = reduce_mean_15_keep_dims_0, x = square_7_cast_fp16)[name = tensor("reduce_mean_15_cast_fp16")]; + tensor var_347_to_fp16 = const()[name = tensor("op_347_to_fp16"), val = tensor(0x1.1p-20)]; + tensor var_348_cast_fp16 = add(x = reduce_mean_15_cast_fp16, y = var_347_to_fp16)[name = tensor("op_348_cast_fp16")]; + tensor var_349_cast_fp16 = sqrt(x = var_348_cast_fp16)[name = tensor("op_349_cast_fp16")]; + tensor x_37_cast_fp16 = real_div(x = sub_9_cast_fp16, y = var_349_cast_fp16)[name = tensor("x_37_cast_fp16")]; + tensor flow_net_res_blocks_5_in_ln_weight_to_fp16 = const()[name = tensor("flow_net_res_blocks_5_in_ln_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(16860608)))]; + tensor var_351_cast_fp16 = mul(x = x_37_cast_fp16, y = flow_net_res_blocks_5_in_ln_weight_to_fp16)[name = tensor("op_351_cast_fp16")]; + tensor flow_net_res_blocks_5_in_ln_bias_to_fp16 = const()[name = tensor("flow_net_res_blocks_5_in_ln_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(16861696)))]; + tensor x_39_cast_fp16 = add(x = var_351_cast_fp16, y = flow_net_res_blocks_5_in_ln_bias_to_fp16)[name = tensor("x_39_cast_fp16")]; + tensor var_353_promoted_to_fp16 = const()[name = tensor("op_353_promoted_to_fp16"), val = tensor(0x1p+0)]; + tensor var_354_cast_fp16 = add(x = var_337_cast_fp16_1, y = var_353_promoted_to_fp16)[name = tensor("op_354_cast_fp16")]; + tensor var_355_cast_fp16 = mul(x = x_39_cast_fp16, y = var_354_cast_fp16)[name = tensor("op_355_cast_fp16")]; + tensor input_57_cast_fp16 = add(x = var_355_cast_fp16, y = var_337_cast_fp16_0)[name = tensor("input_57_cast_fp16")]; + tensor flow_net_res_blocks_5_mlp_0_weight_to_fp16 = const()[name = tensor("flow_net_res_blocks_5_mlp_0_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(16862784)))]; + tensor flow_net_res_blocks_5_mlp_0_bias_to_fp16 = const()[name = tensor("flow_net_res_blocks_5_mlp_0_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(17387136)))]; + tensor linear_22_cast_fp16 = linear(bias = flow_net_res_blocks_5_mlp_0_bias_to_fp16, weight = flow_net_res_blocks_5_mlp_0_weight_to_fp16, x = input_57_cast_fp16)[name = tensor("linear_22_cast_fp16")]; + tensor input_61_cast_fp16 = silu(x = linear_22_cast_fp16)[name = tensor("input_61_cast_fp16")]; + tensor flow_net_res_blocks_5_mlp_2_weight_to_fp16 = const()[name = tensor("flow_net_res_blocks_5_mlp_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(17388224)))]; + tensor flow_net_res_blocks_5_mlp_2_bias_to_fp16 = const()[name = tensor("flow_net_res_blocks_5_mlp_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(17912576)))]; + tensor linear_23_cast_fp16 = linear(bias = flow_net_res_blocks_5_mlp_2_bias_to_fp16, weight = flow_net_res_blocks_5_mlp_2_weight_to_fp16, x = input_61_cast_fp16)[name = tensor("linear_23_cast_fp16")]; + tensor var_366_cast_fp16 = mul(x = var_337_cast_fp16_2, y = linear_23_cast_fp16)[name = tensor("op_366_cast_fp16")]; + tensor x_41_cast_fp16 = add(x = x_35_cast_fp16, y = var_366_cast_fp16)[name = tensor("x_41_cast_fp16")]; + tensor flow_net_final_layer_adaLN_modulation_1_weight_to_fp16 = const()[name = tensor("flow_net_final_layer_adaLN_modulation_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(17913664)))]; + tensor flow_net_final_layer_adaLN_modulation_1_bias_to_fp16 = const()[name = tensor("flow_net_final_layer_adaLN_modulation_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(18962304)))]; + tensor linear_24_cast_fp16 = linear(bias = flow_net_final_layer_adaLN_modulation_1_bias_to_fp16, weight = flow_net_final_layer_adaLN_modulation_1_weight_to_fp16, x = input_15_cast_fp16)[name = tensor("linear_24_cast_fp16")]; + tensor var_375_split_sizes_0 = const()[name = tensor("op_375_split_sizes_0"), val = tensor([512, 512])]; + tensor var_375_axis_0 = const()[name = tensor("op_375_axis_0"), val = tensor(-1)]; + tensor var_375_cast_fp16_0, tensor var_375_cast_fp16_1 = split(axis = var_375_axis_0, split_sizes = var_375_split_sizes_0, x = linear_24_cast_fp16)[name = tensor("op_375_cast_fp16")]; + tensor mean_13_axes_0 = const()[name = tensor("mean_13_axes_0"), val = tensor([-1])]; + tensor mean_13_keep_dims_0 = const()[name = tensor("mean_13_keep_dims_0"), val = tensor(true)]; + tensor mean_13_cast_fp16 = reduce_mean(axes = mean_13_axes_0, keep_dims = mean_13_keep_dims_0, x = x_41_cast_fp16)[name = tensor("mean_13_cast_fp16")]; + tensor sub_10_cast_fp16 = sub(x = x_41_cast_fp16, y = mean_13_cast_fp16)[name = tensor("sub_10_cast_fp16")]; + tensor square_8_cast_fp16 = square(x = sub_10_cast_fp16)[name = tensor("square_8_cast_fp16")]; + tensor reduce_mean_17_axes_0 = const()[name = tensor("reduce_mean_17_axes_0"), val = tensor([-1])]; + tensor reduce_mean_17_keep_dims_0 = const()[name = tensor("reduce_mean_17_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_17_cast_fp16 = reduce_mean(axes = reduce_mean_17_axes_0, keep_dims = reduce_mean_17_keep_dims_0, x = square_8_cast_fp16)[name = tensor("reduce_mean_17_cast_fp16")]; + tensor var_382_to_fp16 = const()[name = tensor("op_382_to_fp16"), val = tensor(0x1.1p-20)]; + tensor var_383_cast_fp16 = add(x = reduce_mean_17_cast_fp16, y = var_382_to_fp16)[name = tensor("op_383_cast_fp16")]; + tensor var_384_cast_fp16 = sqrt(x = var_383_cast_fp16)[name = tensor("op_384_cast_fp16")]; + tensor x_43_cast_fp16 = real_div(x = sub_10_cast_fp16, y = var_384_cast_fp16)[name = tensor("x_43_cast_fp16")]; + tensor var_386_promoted_to_fp16 = const()[name = tensor("op_386_promoted_to_fp16"), val = tensor(0x1p+0)]; + tensor var_387_cast_fp16 = add(x = var_375_cast_fp16_1, y = var_386_promoted_to_fp16)[name = tensor("op_387_cast_fp16")]; + tensor var_388_cast_fp16 = mul(x = x_43_cast_fp16, y = var_387_cast_fp16)[name = tensor("op_388_cast_fp16")]; + tensor input_65_cast_fp16 = add(x = var_388_cast_fp16, y = var_375_cast_fp16_0)[name = tensor("input_65_cast_fp16")]; + tensor flow_net_final_layer_linear_weight_to_fp16 = const()[name = tensor("flow_net_final_layer_linear_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(18964416)))]; + tensor flow_net_final_layer_linear_bias_to_fp16 = const()[name = tensor("flow_net_final_layer_linear_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(18997248)))]; + tensor linear_25_cast_fp16 = linear(bias = flow_net_final_layer_linear_bias_to_fp16, weight = flow_net_final_layer_linear_weight_to_fp16, x = input_65_cast_fp16)[name = tensor("linear_25_cast_fp16")]; + tensor var_399_to_fp16 = const()[name = tensor("op_399_to_fp16"), val = tensor(0x1p-3)]; + tensor var_400_cast_fp16 = mul(x = linear_25_cast_fp16, y = var_399_to_fp16)[name = tensor("op_400_cast_fp16")]; + tensor input_67_cast_fp16 = add(x = latent_init_to_fp16, y = var_400_cast_fp16)[name = tensor("input_67_cast_fp16")]; + tensor linear_26_cast_fp16 = linear(bias = flow_net_input_proj_bias_to_fp16, weight = flow_net_input_proj_weight_to_fp16, x = input_67_cast_fp16)[name = tensor("linear_26_cast_fp16")]; + tensor input_71_to_fp16 = const()[name = tensor("input_71_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(18997376)))]; + tensor input_73_cast_fp16 = silu(x = input_71_to_fp16)[name = tensor("input_73_cast_fp16")]; + tensor linear_28_cast_fp16 = linear(bias = flow_net_time_embed_0_mlp_2_bias_to_fp16, weight = flow_net_time_embed_0_mlp_2_weight_to_fp16, x = input_73_cast_fp16)[name = tensor("linear_28_cast_fp16")]; + tensor reduce_mean_18_axes_0 = const()[name = tensor("reduce_mean_18_axes_0"), val = tensor([-1])]; + tensor reduce_mean_18_keep_dims_0 = const()[name = tensor("reduce_mean_18_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_18_cast_fp16 = reduce_mean(axes = reduce_mean_18_axes_0, keep_dims = reduce_mean_18_keep_dims_0, x = linear_28_cast_fp16)[name = tensor("reduce_mean_18_cast_fp16")]; + tensor sub_11_cast_fp16 = sub(x = linear_28_cast_fp16, y = reduce_mean_18_cast_fp16)[name = tensor("sub_11_cast_fp16")]; + tensor square_9_cast_fp16 = square(x = sub_11_cast_fp16)[name = tensor("square_9_cast_fp16")]; + tensor reduce_mean_19_axes_0 = const()[name = tensor("reduce_mean_19_axes_0"), val = tensor([-1])]; + tensor reduce_mean_19_keep_dims_0 = const()[name = tensor("reduce_mean_19_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_19_cast_fp16 = reduce_mean(axes = reduce_mean_19_axes_0, keep_dims = reduce_mean_19_keep_dims_0, x = square_9_cast_fp16)[name = tensor("reduce_mean_19_cast_fp16")]; + tensor real_div_2_to_fp16 = const()[name = tensor("real_div_2_to_fp16"), val = tensor(0x1.008p+0)]; + tensor mul_2_cast_fp16 = mul(x = reduce_mean_19_cast_fp16, y = real_div_2_to_fp16)[name = tensor("mul_2_cast_fp16")]; + tensor var_466_to_fp16 = const()[name = tensor("op_466_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_19_cast_fp16 = add(x = mul_2_cast_fp16, y = var_466_to_fp16)[name = tensor("var_19_cast_fp16")]; + tensor var_469_epsilon_0 = const()[name = tensor("op_469_epsilon_0"), val = tensor(0x1.197998p-40)]; + tensor var_469_cast_fp16 = rsqrt(epsilon = var_469_epsilon_0, x = var_19_cast_fp16)[name = tensor("op_469_cast_fp16")]; + tensor var_470_cast_fp16 = mul(x = const_3_to_fp16, y = var_469_cast_fp16)[name = tensor("op_470_cast_fp16")]; + tensor var_471_cast_fp16 = mul(x = linear_28_cast_fp16, y = var_470_cast_fp16)[name = tensor("op_471_cast_fp16")]; + tensor input_77_to_fp16 = const()[name = tensor("input_77_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(18998464)))]; + tensor input_79_cast_fp16 = silu(x = input_77_to_fp16)[name = tensor("input_79_cast_fp16")]; + tensor linear_30_cast_fp16 = linear(bias = flow_net_time_embed_1_mlp_2_bias_to_fp16, weight = flow_net_time_embed_1_mlp_2_weight_to_fp16, x = input_79_cast_fp16)[name = tensor("linear_30_cast_fp16")]; + tensor reduce_mean_20_axes_0 = const()[name = tensor("reduce_mean_20_axes_0"), val = tensor([-1])]; + tensor reduce_mean_20_keep_dims_0 = const()[name = tensor("reduce_mean_20_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_20_cast_fp16 = reduce_mean(axes = reduce_mean_20_axes_0, keep_dims = reduce_mean_20_keep_dims_0, x = linear_30_cast_fp16)[name = tensor("reduce_mean_20_cast_fp16")]; + tensor sub_13_cast_fp16 = sub(x = linear_30_cast_fp16, y = reduce_mean_20_cast_fp16)[name = tensor("sub_13_cast_fp16")]; + tensor square_10_cast_fp16 = square(x = sub_13_cast_fp16)[name = tensor("square_10_cast_fp16")]; + tensor reduce_mean_21_axes_0 = const()[name = tensor("reduce_mean_21_axes_0"), val = tensor([-1])]; + tensor reduce_mean_21_keep_dims_0 = const()[name = tensor("reduce_mean_21_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_21_cast_fp16 = reduce_mean(axes = reduce_mean_21_axes_0, keep_dims = reduce_mean_21_keep_dims_0, x = square_10_cast_fp16)[name = tensor("reduce_mean_21_cast_fp16")]; + tensor real_div_3_to_fp16 = const()[name = tensor("real_div_3_to_fp16"), val = tensor(0x1.008p+0)]; + tensor mul_3_cast_fp16 = mul(x = reduce_mean_21_cast_fp16, y = real_div_3_to_fp16)[name = tensor("mul_3_cast_fp16")]; + tensor var_503_to_fp16 = const()[name = tensor("op_503_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_21_cast_fp16 = add(x = mul_3_cast_fp16, y = var_503_to_fp16)[name = tensor("var_21_cast_fp16")]; + tensor var_506_epsilon_0 = const()[name = tensor("op_506_epsilon_0"), val = tensor(0x1.197998p-40)]; + tensor var_506_cast_fp16 = rsqrt(epsilon = var_506_epsilon_0, x = var_21_cast_fp16)[name = tensor("op_506_cast_fp16")]; + tensor var_507_cast_fp16 = mul(x = const_5_to_fp16, y = var_506_cast_fp16)[name = tensor("op_507_cast_fp16")]; + tensor var_508_cast_fp16 = mul(x = linear_30_cast_fp16, y = var_507_cast_fp16)[name = tensor("op_508_cast_fp16")]; + tensor var_520_cast_fp16 = add(x = var_471_cast_fp16, y = var_508_cast_fp16)[name = tensor("op_520_cast_fp16")]; + tensor _inversed_t_combined_3_y_0_to_fp16 = const()[name = tensor("_inversed_t_combined_3_y_0_to_fp16"), val = tensor(0x1p-1)]; + tensor _inversed_t_combined_3_cast_fp16 = mul(x = var_520_cast_fp16, y = _inversed_t_combined_3_y_0_to_fp16)[name = tensor("_inversed_t_combined_3_cast_fp16")]; + tensor input_81_cast_fp16 = add(x = _inversed_t_combined_3_cast_fp16, y = linear_5_cast_fp16)[name = tensor("input_81_cast_fp16")]; + tensor input_83_cast_fp16 = silu(x = input_81_cast_fp16)[name = tensor("input_83_cast_fp16")]; + tensor linear_32_cast_fp16 = linear(bias = flow_net_res_blocks_0_adaLN_modulation_1_bias_to_fp16, weight = flow_net_res_blocks_0_adaLN_modulation_1_weight_to_fp16, x = input_83_cast_fp16)[name = tensor("linear_32_cast_fp16")]; + tensor var_535_split_sizes_0 = const()[name = tensor("op_535_split_sizes_0"), val = tensor([512, 512, 512])]; + tensor var_535_axis_0 = const()[name = tensor("op_535_axis_0"), val = tensor(-1)]; + tensor var_535_cast_fp16_0, tensor var_535_cast_fp16_1, tensor var_535_cast_fp16_2 = split(axis = var_535_axis_0, split_sizes = var_535_split_sizes_0, x = linear_32_cast_fp16)[name = tensor("op_535_cast_fp16")]; + tensor mean_15_axes_0 = const()[name = tensor("mean_15_axes_0"), val = tensor([-1])]; + tensor mean_15_keep_dims_0 = const()[name = tensor("mean_15_keep_dims_0"), val = tensor(true)]; + tensor mean_15_cast_fp16 = reduce_mean(axes = mean_15_axes_0, keep_dims = mean_15_keep_dims_0, x = linear_26_cast_fp16)[name = tensor("mean_15_cast_fp16")]; + tensor sub_15_cast_fp16 = sub(x = linear_26_cast_fp16, y = mean_15_cast_fp16)[name = tensor("sub_15_cast_fp16")]; + tensor square_11_cast_fp16 = square(x = sub_15_cast_fp16)[name = tensor("square_11_cast_fp16")]; + tensor reduce_mean_23_axes_0 = const()[name = tensor("reduce_mean_23_axes_0"), val = tensor([-1])]; + tensor reduce_mean_23_keep_dims_0 = const()[name = tensor("reduce_mean_23_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_23_cast_fp16 = reduce_mean(axes = reduce_mean_23_axes_0, keep_dims = reduce_mean_23_keep_dims_0, x = square_11_cast_fp16)[name = tensor("reduce_mean_23_cast_fp16")]; + tensor var_545_to_fp16 = const()[name = tensor("op_545_to_fp16"), val = tensor(0x1.1p-20)]; + tensor var_546_cast_fp16 = add(x = reduce_mean_23_cast_fp16, y = var_545_to_fp16)[name = tensor("op_546_cast_fp16")]; + tensor var_547_cast_fp16 = sqrt(x = var_546_cast_fp16)[name = tensor("op_547_cast_fp16")]; + tensor x_51_cast_fp16 = real_div(x = sub_15_cast_fp16, y = var_547_cast_fp16)[name = tensor("x_51_cast_fp16")]; + tensor var_549_cast_fp16 = mul(x = x_51_cast_fp16, y = flow_net_res_blocks_0_in_ln_weight_to_fp16)[name = tensor("op_549_cast_fp16")]; + tensor x_53_cast_fp16 = add(x = var_549_cast_fp16, y = flow_net_res_blocks_0_in_ln_bias_to_fp16)[name = tensor("x_53_cast_fp16")]; + tensor var_551_promoted_to_fp16 = const()[name = tensor("op_551_promoted_to_fp16"), val = tensor(0x1p+0)]; + tensor var_552_cast_fp16 = add(x = var_535_cast_fp16_1, y = var_551_promoted_to_fp16)[name = tensor("op_552_cast_fp16")]; + tensor var_553_cast_fp16 = mul(x = x_53_cast_fp16, y = var_552_cast_fp16)[name = tensor("op_553_cast_fp16")]; + tensor input_85_cast_fp16 = add(x = var_553_cast_fp16, y = var_535_cast_fp16_0)[name = tensor("input_85_cast_fp16")]; + tensor linear_33_cast_fp16 = linear(bias = flow_net_res_blocks_0_mlp_0_bias_to_fp16, weight = flow_net_res_blocks_0_mlp_0_weight_to_fp16, x = input_85_cast_fp16)[name = tensor("linear_33_cast_fp16")]; + tensor input_89_cast_fp16 = silu(x = linear_33_cast_fp16)[name = tensor("input_89_cast_fp16")]; + tensor linear_34_cast_fp16 = linear(bias = flow_net_res_blocks_0_mlp_2_bias_to_fp16, weight = flow_net_res_blocks_0_mlp_2_weight_to_fp16, x = input_89_cast_fp16)[name = tensor("linear_34_cast_fp16")]; + tensor var_564_cast_fp16 = mul(x = var_535_cast_fp16_2, y = linear_34_cast_fp16)[name = tensor("op_564_cast_fp16")]; + tensor x_55_cast_fp16 = add(x = linear_26_cast_fp16, y = var_564_cast_fp16)[name = tensor("x_55_cast_fp16")]; + tensor linear_35_cast_fp16 = linear(bias = flow_net_res_blocks_1_adaLN_modulation_1_bias_to_fp16, weight = flow_net_res_blocks_1_adaLN_modulation_1_weight_to_fp16, x = input_83_cast_fp16)[name = tensor("linear_35_cast_fp16")]; + tensor var_574_split_sizes_0 = const()[name = tensor("op_574_split_sizes_0"), val = tensor([512, 512, 512])]; + tensor var_574_axis_0 = const()[name = tensor("op_574_axis_0"), val = tensor(-1)]; + tensor var_574_cast_fp16_0, tensor var_574_cast_fp16_1, tensor var_574_cast_fp16_2 = split(axis = var_574_axis_0, split_sizes = var_574_split_sizes_0, x = linear_35_cast_fp16)[name = tensor("op_574_cast_fp16")]; + tensor mean_17_axes_0 = const()[name = tensor("mean_17_axes_0"), val = tensor([-1])]; + tensor mean_17_keep_dims_0 = const()[name = tensor("mean_17_keep_dims_0"), val = tensor(true)]; + tensor mean_17_cast_fp16 = reduce_mean(axes = mean_17_axes_0, keep_dims = mean_17_keep_dims_0, x = x_55_cast_fp16)[name = tensor("mean_17_cast_fp16")]; + tensor sub_16_cast_fp16 = sub(x = x_55_cast_fp16, y = mean_17_cast_fp16)[name = tensor("sub_16_cast_fp16")]; + tensor square_12_cast_fp16 = square(x = sub_16_cast_fp16)[name = tensor("square_12_cast_fp16")]; + tensor reduce_mean_25_axes_0 = const()[name = tensor("reduce_mean_25_axes_0"), val = tensor([-1])]; + tensor reduce_mean_25_keep_dims_0 = const()[name = tensor("reduce_mean_25_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_25_cast_fp16 = reduce_mean(axes = reduce_mean_25_axes_0, keep_dims = reduce_mean_25_keep_dims_0, x = square_12_cast_fp16)[name = tensor("reduce_mean_25_cast_fp16")]; + tensor var_584_to_fp16 = const()[name = tensor("op_584_to_fp16"), val = tensor(0x1.1p-20)]; + tensor var_585_cast_fp16 = add(x = reduce_mean_25_cast_fp16, y = var_584_to_fp16)[name = tensor("op_585_cast_fp16")]; + tensor var_586_cast_fp16 = sqrt(x = var_585_cast_fp16)[name = tensor("op_586_cast_fp16")]; + tensor x_57_cast_fp16 = real_div(x = sub_16_cast_fp16, y = var_586_cast_fp16)[name = tensor("x_57_cast_fp16")]; + tensor var_588_cast_fp16 = mul(x = x_57_cast_fp16, y = flow_net_res_blocks_1_in_ln_weight_to_fp16)[name = tensor("op_588_cast_fp16")]; + tensor x_59_cast_fp16 = add(x = var_588_cast_fp16, y = flow_net_res_blocks_1_in_ln_bias_to_fp16)[name = tensor("x_59_cast_fp16")]; + tensor var_590_promoted_to_fp16 = const()[name = tensor("op_590_promoted_to_fp16"), val = tensor(0x1p+0)]; + tensor var_591_cast_fp16 = add(x = var_574_cast_fp16_1, y = var_590_promoted_to_fp16)[name = tensor("op_591_cast_fp16")]; + tensor var_592_cast_fp16 = mul(x = x_59_cast_fp16, y = var_591_cast_fp16)[name = tensor("op_592_cast_fp16")]; + tensor input_93_cast_fp16 = add(x = var_592_cast_fp16, y = var_574_cast_fp16_0)[name = tensor("input_93_cast_fp16")]; + tensor linear_36_cast_fp16 = linear(bias = flow_net_res_blocks_1_mlp_0_bias_to_fp16, weight = flow_net_res_blocks_1_mlp_0_weight_to_fp16, x = input_93_cast_fp16)[name = tensor("linear_36_cast_fp16")]; + tensor input_97_cast_fp16 = silu(x = linear_36_cast_fp16)[name = tensor("input_97_cast_fp16")]; + tensor linear_37_cast_fp16 = linear(bias = flow_net_res_blocks_1_mlp_2_bias_to_fp16, weight = flow_net_res_blocks_1_mlp_2_weight_to_fp16, x = input_97_cast_fp16)[name = tensor("linear_37_cast_fp16")]; + tensor var_603_cast_fp16 = mul(x = var_574_cast_fp16_2, y = linear_37_cast_fp16)[name = tensor("op_603_cast_fp16")]; + tensor x_61_cast_fp16 = add(x = x_55_cast_fp16, y = var_603_cast_fp16)[name = tensor("x_61_cast_fp16")]; + tensor linear_38_cast_fp16 = linear(bias = flow_net_res_blocks_2_adaLN_modulation_1_bias_to_fp16, weight = flow_net_res_blocks_2_adaLN_modulation_1_weight_to_fp16, x = input_83_cast_fp16)[name = tensor("linear_38_cast_fp16")]; + tensor var_613_split_sizes_0 = const()[name = tensor("op_613_split_sizes_0"), val = tensor([512, 512, 512])]; + tensor var_613_axis_0 = const()[name = tensor("op_613_axis_0"), val = tensor(-1)]; + tensor var_613_cast_fp16_0, tensor var_613_cast_fp16_1, tensor var_613_cast_fp16_2 = split(axis = var_613_axis_0, split_sizes = var_613_split_sizes_0, x = linear_38_cast_fp16)[name = tensor("op_613_cast_fp16")]; + tensor mean_19_axes_0 = const()[name = tensor("mean_19_axes_0"), val = tensor([-1])]; + tensor mean_19_keep_dims_0 = const()[name = tensor("mean_19_keep_dims_0"), val = tensor(true)]; + tensor mean_19_cast_fp16 = reduce_mean(axes = mean_19_axes_0, keep_dims = mean_19_keep_dims_0, x = x_61_cast_fp16)[name = tensor("mean_19_cast_fp16")]; + tensor sub_17_cast_fp16 = sub(x = x_61_cast_fp16, y = mean_19_cast_fp16)[name = tensor("sub_17_cast_fp16")]; + tensor square_13_cast_fp16 = square(x = sub_17_cast_fp16)[name = tensor("square_13_cast_fp16")]; + tensor reduce_mean_27_axes_0 = const()[name = tensor("reduce_mean_27_axes_0"), val = tensor([-1])]; + tensor reduce_mean_27_keep_dims_0 = const()[name = tensor("reduce_mean_27_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_27_cast_fp16 = reduce_mean(axes = reduce_mean_27_axes_0, keep_dims = reduce_mean_27_keep_dims_0, x = square_13_cast_fp16)[name = tensor("reduce_mean_27_cast_fp16")]; + tensor var_623_to_fp16 = const()[name = tensor("op_623_to_fp16"), val = tensor(0x1.1p-20)]; + tensor var_624_cast_fp16 = add(x = reduce_mean_27_cast_fp16, y = var_623_to_fp16)[name = tensor("op_624_cast_fp16")]; + tensor var_625_cast_fp16 = sqrt(x = var_624_cast_fp16)[name = tensor("op_625_cast_fp16")]; + tensor x_63_cast_fp16 = real_div(x = sub_17_cast_fp16, y = var_625_cast_fp16)[name = tensor("x_63_cast_fp16")]; + tensor var_627_cast_fp16 = mul(x = x_63_cast_fp16, y = flow_net_res_blocks_2_in_ln_weight_to_fp16)[name = tensor("op_627_cast_fp16")]; + tensor x_65_cast_fp16 = add(x = var_627_cast_fp16, y = flow_net_res_blocks_2_in_ln_bias_to_fp16)[name = tensor("x_65_cast_fp16")]; + tensor var_629_promoted_to_fp16 = const()[name = tensor("op_629_promoted_to_fp16"), val = tensor(0x1p+0)]; + tensor var_630_cast_fp16 = add(x = var_613_cast_fp16_1, y = var_629_promoted_to_fp16)[name = tensor("op_630_cast_fp16")]; + tensor var_631_cast_fp16 = mul(x = x_65_cast_fp16, y = var_630_cast_fp16)[name = tensor("op_631_cast_fp16")]; + tensor input_101_cast_fp16 = add(x = var_631_cast_fp16, y = var_613_cast_fp16_0)[name = tensor("input_101_cast_fp16")]; + tensor linear_39_cast_fp16 = linear(bias = flow_net_res_blocks_2_mlp_0_bias_to_fp16, weight = flow_net_res_blocks_2_mlp_0_weight_to_fp16, x = input_101_cast_fp16)[name = tensor("linear_39_cast_fp16")]; + tensor input_105_cast_fp16 = silu(x = linear_39_cast_fp16)[name = tensor("input_105_cast_fp16")]; + tensor linear_40_cast_fp16 = linear(bias = flow_net_res_blocks_2_mlp_2_bias_to_fp16, weight = flow_net_res_blocks_2_mlp_2_weight_to_fp16, x = input_105_cast_fp16)[name = tensor("linear_40_cast_fp16")]; + tensor var_642_cast_fp16 = mul(x = var_613_cast_fp16_2, y = linear_40_cast_fp16)[name = tensor("op_642_cast_fp16")]; + tensor x_67_cast_fp16 = add(x = x_61_cast_fp16, y = var_642_cast_fp16)[name = tensor("x_67_cast_fp16")]; + tensor linear_41_cast_fp16 = linear(bias = flow_net_res_blocks_3_adaLN_modulation_1_bias_to_fp16, weight = flow_net_res_blocks_3_adaLN_modulation_1_weight_to_fp16, x = input_83_cast_fp16)[name = tensor("linear_41_cast_fp16")]; + tensor var_652_split_sizes_0 = const()[name = tensor("op_652_split_sizes_0"), val = tensor([512, 512, 512])]; + tensor var_652_axis_0 = const()[name = tensor("op_652_axis_0"), val = tensor(-1)]; + tensor var_652_cast_fp16_0, tensor var_652_cast_fp16_1, tensor var_652_cast_fp16_2 = split(axis = var_652_axis_0, split_sizes = var_652_split_sizes_0, x = linear_41_cast_fp16)[name = tensor("op_652_cast_fp16")]; + tensor mean_21_axes_0 = const()[name = tensor("mean_21_axes_0"), val = tensor([-1])]; + tensor mean_21_keep_dims_0 = const()[name = tensor("mean_21_keep_dims_0"), val = tensor(true)]; + tensor mean_21_cast_fp16 = reduce_mean(axes = mean_21_axes_0, keep_dims = mean_21_keep_dims_0, x = x_67_cast_fp16)[name = tensor("mean_21_cast_fp16")]; + tensor sub_18_cast_fp16 = sub(x = x_67_cast_fp16, y = mean_21_cast_fp16)[name = tensor("sub_18_cast_fp16")]; + tensor square_14_cast_fp16 = square(x = sub_18_cast_fp16)[name = tensor("square_14_cast_fp16")]; + tensor reduce_mean_29_axes_0 = const()[name = tensor("reduce_mean_29_axes_0"), val = tensor([-1])]; + tensor reduce_mean_29_keep_dims_0 = const()[name = tensor("reduce_mean_29_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_29_cast_fp16 = reduce_mean(axes = reduce_mean_29_axes_0, keep_dims = reduce_mean_29_keep_dims_0, x = square_14_cast_fp16)[name = tensor("reduce_mean_29_cast_fp16")]; + tensor var_662_to_fp16 = const()[name = tensor("op_662_to_fp16"), val = tensor(0x1.1p-20)]; + tensor var_663_cast_fp16 = add(x = reduce_mean_29_cast_fp16, y = var_662_to_fp16)[name = tensor("op_663_cast_fp16")]; + tensor var_664_cast_fp16 = sqrt(x = var_663_cast_fp16)[name = tensor("op_664_cast_fp16")]; + tensor x_69_cast_fp16 = real_div(x = sub_18_cast_fp16, y = var_664_cast_fp16)[name = tensor("x_69_cast_fp16")]; + tensor var_666_cast_fp16 = mul(x = x_69_cast_fp16, y = flow_net_res_blocks_3_in_ln_weight_to_fp16)[name = tensor("op_666_cast_fp16")]; + tensor x_71_cast_fp16 = add(x = var_666_cast_fp16, y = flow_net_res_blocks_3_in_ln_bias_to_fp16)[name = tensor("x_71_cast_fp16")]; + tensor var_668_promoted_to_fp16 = const()[name = tensor("op_668_promoted_to_fp16"), val = tensor(0x1p+0)]; + tensor var_669_cast_fp16 = add(x = var_652_cast_fp16_1, y = var_668_promoted_to_fp16)[name = tensor("op_669_cast_fp16")]; + tensor var_670_cast_fp16 = mul(x = x_71_cast_fp16, y = var_669_cast_fp16)[name = tensor("op_670_cast_fp16")]; + tensor input_109_cast_fp16 = add(x = var_670_cast_fp16, y = var_652_cast_fp16_0)[name = tensor("input_109_cast_fp16")]; + tensor linear_42_cast_fp16 = linear(bias = flow_net_res_blocks_3_mlp_0_bias_to_fp16, weight = flow_net_res_blocks_3_mlp_0_weight_to_fp16, x = input_109_cast_fp16)[name = tensor("linear_42_cast_fp16")]; + tensor input_113_cast_fp16 = silu(x = linear_42_cast_fp16)[name = tensor("input_113_cast_fp16")]; + tensor linear_43_cast_fp16 = linear(bias = flow_net_res_blocks_3_mlp_2_bias_to_fp16, weight = flow_net_res_blocks_3_mlp_2_weight_to_fp16, x = input_113_cast_fp16)[name = tensor("linear_43_cast_fp16")]; + tensor var_681_cast_fp16 = mul(x = var_652_cast_fp16_2, y = linear_43_cast_fp16)[name = tensor("op_681_cast_fp16")]; + tensor x_73_cast_fp16 = add(x = x_67_cast_fp16, y = var_681_cast_fp16)[name = tensor("x_73_cast_fp16")]; + tensor linear_44_cast_fp16 = linear(bias = flow_net_res_blocks_4_adaLN_modulation_1_bias_to_fp16, weight = flow_net_res_blocks_4_adaLN_modulation_1_weight_to_fp16, x = input_83_cast_fp16)[name = tensor("linear_44_cast_fp16")]; + tensor var_691_split_sizes_0 = const()[name = tensor("op_691_split_sizes_0"), val = tensor([512, 512, 512])]; + tensor var_691_axis_0 = const()[name = tensor("op_691_axis_0"), val = tensor(-1)]; + tensor var_691_cast_fp16_0, tensor var_691_cast_fp16_1, tensor var_691_cast_fp16_2 = split(axis = var_691_axis_0, split_sizes = var_691_split_sizes_0, x = linear_44_cast_fp16)[name = tensor("op_691_cast_fp16")]; + tensor mean_23_axes_0 = const()[name = tensor("mean_23_axes_0"), val = tensor([-1])]; + tensor mean_23_keep_dims_0 = const()[name = tensor("mean_23_keep_dims_0"), val = tensor(true)]; + tensor mean_23_cast_fp16 = reduce_mean(axes = mean_23_axes_0, keep_dims = mean_23_keep_dims_0, x = x_73_cast_fp16)[name = tensor("mean_23_cast_fp16")]; + tensor sub_19_cast_fp16 = sub(x = x_73_cast_fp16, y = mean_23_cast_fp16)[name = tensor("sub_19_cast_fp16")]; + tensor square_15_cast_fp16 = square(x = sub_19_cast_fp16)[name = tensor("square_15_cast_fp16")]; + tensor reduce_mean_31_axes_0 = const()[name = tensor("reduce_mean_31_axes_0"), val = tensor([-1])]; + tensor reduce_mean_31_keep_dims_0 = const()[name = tensor("reduce_mean_31_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_31_cast_fp16 = reduce_mean(axes = reduce_mean_31_axes_0, keep_dims = reduce_mean_31_keep_dims_0, x = square_15_cast_fp16)[name = tensor("reduce_mean_31_cast_fp16")]; + tensor var_701_to_fp16 = const()[name = tensor("op_701_to_fp16"), val = tensor(0x1.1p-20)]; + tensor var_702_cast_fp16 = add(x = reduce_mean_31_cast_fp16, y = var_701_to_fp16)[name = tensor("op_702_cast_fp16")]; + tensor var_703_cast_fp16 = sqrt(x = var_702_cast_fp16)[name = tensor("op_703_cast_fp16")]; + tensor x_75_cast_fp16 = real_div(x = sub_19_cast_fp16, y = var_703_cast_fp16)[name = tensor("x_75_cast_fp16")]; + tensor var_705_cast_fp16 = mul(x = x_75_cast_fp16, y = flow_net_res_blocks_4_in_ln_weight_to_fp16)[name = tensor("op_705_cast_fp16")]; + tensor x_77_cast_fp16 = add(x = var_705_cast_fp16, y = flow_net_res_blocks_4_in_ln_bias_to_fp16)[name = tensor("x_77_cast_fp16")]; + tensor var_707_promoted_to_fp16 = const()[name = tensor("op_707_promoted_to_fp16"), val = tensor(0x1p+0)]; + tensor var_708_cast_fp16 = add(x = var_691_cast_fp16_1, y = var_707_promoted_to_fp16)[name = tensor("op_708_cast_fp16")]; + tensor var_709_cast_fp16 = mul(x = x_77_cast_fp16, y = var_708_cast_fp16)[name = tensor("op_709_cast_fp16")]; + tensor input_117_cast_fp16 = add(x = var_709_cast_fp16, y = var_691_cast_fp16_0)[name = tensor("input_117_cast_fp16")]; + tensor linear_45_cast_fp16 = linear(bias = flow_net_res_blocks_4_mlp_0_bias_to_fp16, weight = flow_net_res_blocks_4_mlp_0_weight_to_fp16, x = input_117_cast_fp16)[name = tensor("linear_45_cast_fp16")]; + tensor input_121_cast_fp16 = silu(x = linear_45_cast_fp16)[name = tensor("input_121_cast_fp16")]; + tensor linear_46_cast_fp16 = linear(bias = flow_net_res_blocks_4_mlp_2_bias_to_fp16, weight = flow_net_res_blocks_4_mlp_2_weight_to_fp16, x = input_121_cast_fp16)[name = tensor("linear_46_cast_fp16")]; + tensor var_720_cast_fp16 = mul(x = var_691_cast_fp16_2, y = linear_46_cast_fp16)[name = tensor("op_720_cast_fp16")]; + tensor x_79_cast_fp16 = add(x = x_73_cast_fp16, y = var_720_cast_fp16)[name = tensor("x_79_cast_fp16")]; + tensor linear_47_cast_fp16 = linear(bias = flow_net_res_blocks_5_adaLN_modulation_1_bias_to_fp16, weight = flow_net_res_blocks_5_adaLN_modulation_1_weight_to_fp16, x = input_83_cast_fp16)[name = tensor("linear_47_cast_fp16")]; + tensor var_730_split_sizes_0 = const()[name = tensor("op_730_split_sizes_0"), val = tensor([512, 512, 512])]; + tensor var_730_axis_0 = const()[name = tensor("op_730_axis_0"), val = tensor(-1)]; + tensor var_730_cast_fp16_0, tensor var_730_cast_fp16_1, tensor var_730_cast_fp16_2 = split(axis = var_730_axis_0, split_sizes = var_730_split_sizes_0, x = linear_47_cast_fp16)[name = tensor("op_730_cast_fp16")]; + tensor mean_25_axes_0 = const()[name = tensor("mean_25_axes_0"), val = tensor([-1])]; + tensor mean_25_keep_dims_0 = const()[name = tensor("mean_25_keep_dims_0"), val = tensor(true)]; + tensor mean_25_cast_fp16 = reduce_mean(axes = mean_25_axes_0, keep_dims = mean_25_keep_dims_0, x = x_79_cast_fp16)[name = tensor("mean_25_cast_fp16")]; + tensor sub_20_cast_fp16 = sub(x = x_79_cast_fp16, y = mean_25_cast_fp16)[name = tensor("sub_20_cast_fp16")]; + tensor square_16_cast_fp16 = square(x = sub_20_cast_fp16)[name = tensor("square_16_cast_fp16")]; + tensor reduce_mean_33_axes_0 = const()[name = tensor("reduce_mean_33_axes_0"), val = tensor([-1])]; + tensor reduce_mean_33_keep_dims_0 = const()[name = tensor("reduce_mean_33_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_33_cast_fp16 = reduce_mean(axes = reduce_mean_33_axes_0, keep_dims = reduce_mean_33_keep_dims_0, x = square_16_cast_fp16)[name = tensor("reduce_mean_33_cast_fp16")]; + tensor var_740_to_fp16 = const()[name = tensor("op_740_to_fp16"), val = tensor(0x1.1p-20)]; + tensor var_741_cast_fp16 = add(x = reduce_mean_33_cast_fp16, y = var_740_to_fp16)[name = tensor("op_741_cast_fp16")]; + tensor var_742_cast_fp16 = sqrt(x = var_741_cast_fp16)[name = tensor("op_742_cast_fp16")]; + tensor x_81_cast_fp16 = real_div(x = sub_20_cast_fp16, y = var_742_cast_fp16)[name = tensor("x_81_cast_fp16")]; + tensor var_744_cast_fp16 = mul(x = x_81_cast_fp16, y = flow_net_res_blocks_5_in_ln_weight_to_fp16)[name = tensor("op_744_cast_fp16")]; + tensor x_83_cast_fp16 = add(x = var_744_cast_fp16, y = flow_net_res_blocks_5_in_ln_bias_to_fp16)[name = tensor("x_83_cast_fp16")]; + tensor var_746_promoted_to_fp16 = const()[name = tensor("op_746_promoted_to_fp16"), val = tensor(0x1p+0)]; + tensor var_747_cast_fp16 = add(x = var_730_cast_fp16_1, y = var_746_promoted_to_fp16)[name = tensor("op_747_cast_fp16")]; + tensor var_748_cast_fp16 = mul(x = x_83_cast_fp16, y = var_747_cast_fp16)[name = tensor("op_748_cast_fp16")]; + tensor input_125_cast_fp16 = add(x = var_748_cast_fp16, y = var_730_cast_fp16_0)[name = tensor("input_125_cast_fp16")]; + tensor linear_48_cast_fp16 = linear(bias = flow_net_res_blocks_5_mlp_0_bias_to_fp16, weight = flow_net_res_blocks_5_mlp_0_weight_to_fp16, x = input_125_cast_fp16)[name = tensor("linear_48_cast_fp16")]; + tensor input_129_cast_fp16 = silu(x = linear_48_cast_fp16)[name = tensor("input_129_cast_fp16")]; + tensor linear_49_cast_fp16 = linear(bias = flow_net_res_blocks_5_mlp_2_bias_to_fp16, weight = flow_net_res_blocks_5_mlp_2_weight_to_fp16, x = input_129_cast_fp16)[name = tensor("linear_49_cast_fp16")]; + tensor var_759_cast_fp16 = mul(x = var_730_cast_fp16_2, y = linear_49_cast_fp16)[name = tensor("op_759_cast_fp16")]; + tensor x_85_cast_fp16 = add(x = x_79_cast_fp16, y = var_759_cast_fp16)[name = tensor("x_85_cast_fp16")]; + tensor linear_50_cast_fp16 = linear(bias = flow_net_final_layer_adaLN_modulation_1_bias_to_fp16, weight = flow_net_final_layer_adaLN_modulation_1_weight_to_fp16, x = input_83_cast_fp16)[name = tensor("linear_50_cast_fp16")]; + tensor var_768_split_sizes_0 = const()[name = tensor("op_768_split_sizes_0"), val = tensor([512, 512])]; + tensor var_768_axis_0 = const()[name = tensor("op_768_axis_0"), val = tensor(-1)]; + tensor var_768_cast_fp16_0, tensor var_768_cast_fp16_1 = split(axis = var_768_axis_0, split_sizes = var_768_split_sizes_0, x = linear_50_cast_fp16)[name = tensor("op_768_cast_fp16")]; + tensor mean_27_axes_0 = const()[name = tensor("mean_27_axes_0"), val = tensor([-1])]; + tensor mean_27_keep_dims_0 = const()[name = tensor("mean_27_keep_dims_0"), val = tensor(true)]; + tensor mean_27_cast_fp16 = reduce_mean(axes = mean_27_axes_0, keep_dims = mean_27_keep_dims_0, x = x_85_cast_fp16)[name = tensor("mean_27_cast_fp16")]; + tensor sub_21_cast_fp16 = sub(x = x_85_cast_fp16, y = mean_27_cast_fp16)[name = tensor("sub_21_cast_fp16")]; + tensor square_17_cast_fp16 = square(x = sub_21_cast_fp16)[name = tensor("square_17_cast_fp16")]; + tensor reduce_mean_35_axes_0 = const()[name = tensor("reduce_mean_35_axes_0"), val = tensor([-1])]; + tensor reduce_mean_35_keep_dims_0 = const()[name = tensor("reduce_mean_35_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_35_cast_fp16 = reduce_mean(axes = reduce_mean_35_axes_0, keep_dims = reduce_mean_35_keep_dims_0, x = square_17_cast_fp16)[name = tensor("reduce_mean_35_cast_fp16")]; + tensor var_775_to_fp16 = const()[name = tensor("op_775_to_fp16"), val = tensor(0x1.1p-20)]; + tensor var_776_cast_fp16 = add(x = reduce_mean_35_cast_fp16, y = var_775_to_fp16)[name = tensor("op_776_cast_fp16")]; + tensor var_777_cast_fp16 = sqrt(x = var_776_cast_fp16)[name = tensor("op_777_cast_fp16")]; + tensor x_87_cast_fp16 = real_div(x = sub_21_cast_fp16, y = var_777_cast_fp16)[name = tensor("x_87_cast_fp16")]; + tensor var_779_promoted_to_fp16 = const()[name = tensor("op_779_promoted_to_fp16"), val = tensor(0x1p+0)]; + tensor var_780_cast_fp16 = add(x = var_768_cast_fp16_1, y = var_779_promoted_to_fp16)[name = tensor("op_780_cast_fp16")]; + tensor var_781_cast_fp16 = mul(x = x_87_cast_fp16, y = var_780_cast_fp16)[name = tensor("op_781_cast_fp16")]; + tensor input_133_cast_fp16 = add(x = var_781_cast_fp16, y = var_768_cast_fp16_0)[name = tensor("input_133_cast_fp16")]; + tensor linear_51_cast_fp16 = linear(bias = flow_net_final_layer_linear_bias_to_fp16, weight = flow_net_final_layer_linear_weight_to_fp16, x = input_133_cast_fp16)[name = tensor("linear_51_cast_fp16")]; + tensor var_792_to_fp16 = const()[name = tensor("op_792_to_fp16"), val = tensor(0x1p-3)]; + tensor var_793_cast_fp16 = mul(x = linear_51_cast_fp16, y = var_792_to_fp16)[name = tensor("op_793_cast_fp16")]; + tensor input_135_cast_fp16 = add(x = input_67_cast_fp16, y = var_793_cast_fp16)[name = tensor("input_135_cast_fp16")]; + tensor linear_52_cast_fp16 = linear(bias = flow_net_input_proj_bias_to_fp16, weight = flow_net_input_proj_weight_to_fp16, x = input_135_cast_fp16)[name = tensor("linear_52_cast_fp16")]; + tensor input_139_to_fp16 = const()[name = tensor("input_139_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(18999552)))]; + tensor input_141_cast_fp16 = silu(x = input_139_to_fp16)[name = tensor("input_141_cast_fp16")]; + tensor linear_54_cast_fp16 = linear(bias = flow_net_time_embed_0_mlp_2_bias_to_fp16, weight = flow_net_time_embed_0_mlp_2_weight_to_fp16, x = input_141_cast_fp16)[name = tensor("linear_54_cast_fp16")]; + tensor reduce_mean_36_axes_0 = const()[name = tensor("reduce_mean_36_axes_0"), val = tensor([-1])]; + tensor reduce_mean_36_keep_dims_0 = const()[name = tensor("reduce_mean_36_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_36_cast_fp16 = reduce_mean(axes = reduce_mean_36_axes_0, keep_dims = reduce_mean_36_keep_dims_0, x = linear_54_cast_fp16)[name = tensor("reduce_mean_36_cast_fp16")]; + tensor sub_22_cast_fp16 = sub(x = linear_54_cast_fp16, y = reduce_mean_36_cast_fp16)[name = tensor("sub_22_cast_fp16")]; + tensor square_18_cast_fp16 = square(x = sub_22_cast_fp16)[name = tensor("square_18_cast_fp16")]; + tensor reduce_mean_37_axes_0 = const()[name = tensor("reduce_mean_37_axes_0"), val = tensor([-1])]; + tensor reduce_mean_37_keep_dims_0 = const()[name = tensor("reduce_mean_37_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_37_cast_fp16 = reduce_mean(axes = reduce_mean_37_axes_0, keep_dims = reduce_mean_37_keep_dims_0, x = square_18_cast_fp16)[name = tensor("reduce_mean_37_cast_fp16")]; + tensor real_div_4_to_fp16 = const()[name = tensor("real_div_4_to_fp16"), val = tensor(0x1.008p+0)]; + tensor mul_4_cast_fp16 = mul(x = reduce_mean_37_cast_fp16, y = real_div_4_to_fp16)[name = tensor("mul_4_cast_fp16")]; + tensor var_859_to_fp16 = const()[name = tensor("op_859_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_37_cast_fp16 = add(x = mul_4_cast_fp16, y = var_859_to_fp16)[name = tensor("var_37_cast_fp16")]; + tensor var_862_epsilon_0 = const()[name = tensor("op_862_epsilon_0"), val = tensor(0x1.197998p-40)]; + tensor var_862_cast_fp16 = rsqrt(epsilon = var_862_epsilon_0, x = var_37_cast_fp16)[name = tensor("op_862_cast_fp16")]; + tensor var_863_cast_fp16 = mul(x = const_3_to_fp16, y = var_862_cast_fp16)[name = tensor("op_863_cast_fp16")]; + tensor var_864_cast_fp16 = mul(x = linear_54_cast_fp16, y = var_863_cast_fp16)[name = tensor("op_864_cast_fp16")]; + tensor input_145_to_fp16 = const()[name = tensor("input_145_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(19000640)))]; + tensor input_147_cast_fp16 = silu(x = input_145_to_fp16)[name = tensor("input_147_cast_fp16")]; + tensor linear_56_cast_fp16 = linear(bias = flow_net_time_embed_1_mlp_2_bias_to_fp16, weight = flow_net_time_embed_1_mlp_2_weight_to_fp16, x = input_147_cast_fp16)[name = tensor("linear_56_cast_fp16")]; + tensor reduce_mean_38_axes_0 = const()[name = tensor("reduce_mean_38_axes_0"), val = tensor([-1])]; + tensor reduce_mean_38_keep_dims_0 = const()[name = tensor("reduce_mean_38_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_38_cast_fp16 = reduce_mean(axes = reduce_mean_38_axes_0, keep_dims = reduce_mean_38_keep_dims_0, x = linear_56_cast_fp16)[name = tensor("reduce_mean_38_cast_fp16")]; + tensor sub_24_cast_fp16 = sub(x = linear_56_cast_fp16, y = reduce_mean_38_cast_fp16)[name = tensor("sub_24_cast_fp16")]; + tensor square_19_cast_fp16 = square(x = sub_24_cast_fp16)[name = tensor("square_19_cast_fp16")]; + tensor reduce_mean_39_axes_0 = const()[name = tensor("reduce_mean_39_axes_0"), val = tensor([-1])]; + tensor reduce_mean_39_keep_dims_0 = const()[name = tensor("reduce_mean_39_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_39_cast_fp16 = reduce_mean(axes = reduce_mean_39_axes_0, keep_dims = reduce_mean_39_keep_dims_0, x = square_19_cast_fp16)[name = tensor("reduce_mean_39_cast_fp16")]; + tensor real_div_5_to_fp16 = const()[name = tensor("real_div_5_to_fp16"), val = tensor(0x1.008p+0)]; + tensor mul_5_cast_fp16 = mul(x = reduce_mean_39_cast_fp16, y = real_div_5_to_fp16)[name = tensor("mul_5_cast_fp16")]; + tensor var_896_to_fp16 = const()[name = tensor("op_896_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_39_cast_fp16 = add(x = mul_5_cast_fp16, y = var_896_to_fp16)[name = tensor("var_39_cast_fp16")]; + tensor var_899_epsilon_0 = const()[name = tensor("op_899_epsilon_0"), val = tensor(0x1.197998p-40)]; + tensor var_899_cast_fp16 = rsqrt(epsilon = var_899_epsilon_0, x = var_39_cast_fp16)[name = tensor("op_899_cast_fp16")]; + tensor var_900_cast_fp16 = mul(x = const_5_to_fp16, y = var_899_cast_fp16)[name = tensor("op_900_cast_fp16")]; + tensor var_901_cast_fp16 = mul(x = linear_56_cast_fp16, y = var_900_cast_fp16)[name = tensor("op_901_cast_fp16")]; + tensor var_913_cast_fp16 = add(x = var_864_cast_fp16, y = var_901_cast_fp16)[name = tensor("op_913_cast_fp16")]; + tensor _inversed_t_combined_5_y_0_to_fp16 = const()[name = tensor("_inversed_t_combined_5_y_0_to_fp16"), val = tensor(0x1p-1)]; + tensor _inversed_t_combined_5_cast_fp16 = mul(x = var_913_cast_fp16, y = _inversed_t_combined_5_y_0_to_fp16)[name = tensor("_inversed_t_combined_5_cast_fp16")]; + tensor input_149_cast_fp16 = add(x = _inversed_t_combined_5_cast_fp16, y = linear_5_cast_fp16)[name = tensor("input_149_cast_fp16")]; + tensor input_151_cast_fp16 = silu(x = input_149_cast_fp16)[name = tensor("input_151_cast_fp16")]; + tensor linear_58_cast_fp16 = linear(bias = flow_net_res_blocks_0_adaLN_modulation_1_bias_to_fp16, weight = flow_net_res_blocks_0_adaLN_modulation_1_weight_to_fp16, x = input_151_cast_fp16)[name = tensor("linear_58_cast_fp16")]; + tensor var_928_split_sizes_0 = const()[name = tensor("op_928_split_sizes_0"), val = tensor([512, 512, 512])]; + tensor var_928_axis_0 = const()[name = tensor("op_928_axis_0"), val = tensor(-1)]; + tensor var_928_cast_fp16_0, tensor var_928_cast_fp16_1, tensor var_928_cast_fp16_2 = split(axis = var_928_axis_0, split_sizes = var_928_split_sizes_0, x = linear_58_cast_fp16)[name = tensor("op_928_cast_fp16")]; + tensor mean_29_axes_0 = const()[name = tensor("mean_29_axes_0"), val = tensor([-1])]; + tensor mean_29_keep_dims_0 = const()[name = tensor("mean_29_keep_dims_0"), val = tensor(true)]; + tensor mean_29_cast_fp16 = reduce_mean(axes = mean_29_axes_0, keep_dims = mean_29_keep_dims_0, x = linear_52_cast_fp16)[name = tensor("mean_29_cast_fp16")]; + tensor sub_26_cast_fp16 = sub(x = linear_52_cast_fp16, y = mean_29_cast_fp16)[name = tensor("sub_26_cast_fp16")]; + tensor square_20_cast_fp16 = square(x = sub_26_cast_fp16)[name = tensor("square_20_cast_fp16")]; + tensor reduce_mean_41_axes_0 = const()[name = tensor("reduce_mean_41_axes_0"), val = tensor([-1])]; + tensor reduce_mean_41_keep_dims_0 = const()[name = tensor("reduce_mean_41_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_41_cast_fp16 = reduce_mean(axes = reduce_mean_41_axes_0, keep_dims = reduce_mean_41_keep_dims_0, x = square_20_cast_fp16)[name = tensor("reduce_mean_41_cast_fp16")]; + tensor var_938_to_fp16 = const()[name = tensor("op_938_to_fp16"), val = tensor(0x1.1p-20)]; + tensor var_939_cast_fp16 = add(x = reduce_mean_41_cast_fp16, y = var_938_to_fp16)[name = tensor("op_939_cast_fp16")]; + tensor var_940_cast_fp16 = sqrt(x = var_939_cast_fp16)[name = tensor("op_940_cast_fp16")]; + tensor x_95_cast_fp16 = real_div(x = sub_26_cast_fp16, y = var_940_cast_fp16)[name = tensor("x_95_cast_fp16")]; + tensor var_942_cast_fp16 = mul(x = x_95_cast_fp16, y = flow_net_res_blocks_0_in_ln_weight_to_fp16)[name = tensor("op_942_cast_fp16")]; + tensor x_97_cast_fp16 = add(x = var_942_cast_fp16, y = flow_net_res_blocks_0_in_ln_bias_to_fp16)[name = tensor("x_97_cast_fp16")]; + tensor var_944_promoted_to_fp16 = const()[name = tensor("op_944_promoted_to_fp16"), val = tensor(0x1p+0)]; + tensor var_945_cast_fp16 = add(x = var_928_cast_fp16_1, y = var_944_promoted_to_fp16)[name = tensor("op_945_cast_fp16")]; + tensor var_946_cast_fp16 = mul(x = x_97_cast_fp16, y = var_945_cast_fp16)[name = tensor("op_946_cast_fp16")]; + tensor input_153_cast_fp16 = add(x = var_946_cast_fp16, y = var_928_cast_fp16_0)[name = tensor("input_153_cast_fp16")]; + tensor linear_59_cast_fp16 = linear(bias = flow_net_res_blocks_0_mlp_0_bias_to_fp16, weight = flow_net_res_blocks_0_mlp_0_weight_to_fp16, x = input_153_cast_fp16)[name = tensor("linear_59_cast_fp16")]; + tensor input_157_cast_fp16 = silu(x = linear_59_cast_fp16)[name = tensor("input_157_cast_fp16")]; + tensor linear_60_cast_fp16 = linear(bias = flow_net_res_blocks_0_mlp_2_bias_to_fp16, weight = flow_net_res_blocks_0_mlp_2_weight_to_fp16, x = input_157_cast_fp16)[name = tensor("linear_60_cast_fp16")]; + tensor var_957_cast_fp16 = mul(x = var_928_cast_fp16_2, y = linear_60_cast_fp16)[name = tensor("op_957_cast_fp16")]; + tensor x_99_cast_fp16 = add(x = linear_52_cast_fp16, y = var_957_cast_fp16)[name = tensor("x_99_cast_fp16")]; + tensor linear_61_cast_fp16 = linear(bias = flow_net_res_blocks_1_adaLN_modulation_1_bias_to_fp16, weight = flow_net_res_blocks_1_adaLN_modulation_1_weight_to_fp16, x = input_151_cast_fp16)[name = tensor("linear_61_cast_fp16")]; + tensor var_967_split_sizes_0 = const()[name = tensor("op_967_split_sizes_0"), val = tensor([512, 512, 512])]; + tensor var_967_axis_0 = const()[name = tensor("op_967_axis_0"), val = tensor(-1)]; + tensor var_967_cast_fp16_0, tensor var_967_cast_fp16_1, tensor var_967_cast_fp16_2 = split(axis = var_967_axis_0, split_sizes = var_967_split_sizes_0, x = linear_61_cast_fp16)[name = tensor("op_967_cast_fp16")]; + tensor mean_31_axes_0 = const()[name = tensor("mean_31_axes_0"), val = tensor([-1])]; + tensor mean_31_keep_dims_0 = const()[name = tensor("mean_31_keep_dims_0"), val = tensor(true)]; + tensor mean_31_cast_fp16 = reduce_mean(axes = mean_31_axes_0, keep_dims = mean_31_keep_dims_0, x = x_99_cast_fp16)[name = tensor("mean_31_cast_fp16")]; + tensor sub_27_cast_fp16 = sub(x = x_99_cast_fp16, y = mean_31_cast_fp16)[name = tensor("sub_27_cast_fp16")]; + tensor square_21_cast_fp16 = square(x = sub_27_cast_fp16)[name = tensor("square_21_cast_fp16")]; + tensor reduce_mean_43_axes_0 = const()[name = tensor("reduce_mean_43_axes_0"), val = tensor([-1])]; + tensor reduce_mean_43_keep_dims_0 = const()[name = tensor("reduce_mean_43_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_43_cast_fp16 = reduce_mean(axes = reduce_mean_43_axes_0, keep_dims = reduce_mean_43_keep_dims_0, x = square_21_cast_fp16)[name = tensor("reduce_mean_43_cast_fp16")]; + tensor var_977_to_fp16 = const()[name = tensor("op_977_to_fp16"), val = tensor(0x1.1p-20)]; + tensor var_978_cast_fp16 = add(x = reduce_mean_43_cast_fp16, y = var_977_to_fp16)[name = tensor("op_978_cast_fp16")]; + tensor var_979_cast_fp16 = sqrt(x = var_978_cast_fp16)[name = tensor("op_979_cast_fp16")]; + tensor x_101_cast_fp16 = real_div(x = sub_27_cast_fp16, y = var_979_cast_fp16)[name = tensor("x_101_cast_fp16")]; + tensor var_981_cast_fp16 = mul(x = x_101_cast_fp16, y = flow_net_res_blocks_1_in_ln_weight_to_fp16)[name = tensor("op_981_cast_fp16")]; + tensor x_103_cast_fp16 = add(x = var_981_cast_fp16, y = flow_net_res_blocks_1_in_ln_bias_to_fp16)[name = tensor("x_103_cast_fp16")]; + tensor var_983_promoted_to_fp16 = const()[name = tensor("op_983_promoted_to_fp16"), val = tensor(0x1p+0)]; + tensor var_984_cast_fp16 = add(x = var_967_cast_fp16_1, y = var_983_promoted_to_fp16)[name = tensor("op_984_cast_fp16")]; + tensor var_985_cast_fp16 = mul(x = x_103_cast_fp16, y = var_984_cast_fp16)[name = tensor("op_985_cast_fp16")]; + tensor input_161_cast_fp16 = add(x = var_985_cast_fp16, y = var_967_cast_fp16_0)[name = tensor("input_161_cast_fp16")]; + tensor linear_62_cast_fp16 = linear(bias = flow_net_res_blocks_1_mlp_0_bias_to_fp16, weight = flow_net_res_blocks_1_mlp_0_weight_to_fp16, x = input_161_cast_fp16)[name = tensor("linear_62_cast_fp16")]; + tensor input_165_cast_fp16 = silu(x = linear_62_cast_fp16)[name = tensor("input_165_cast_fp16")]; + tensor linear_63_cast_fp16 = linear(bias = flow_net_res_blocks_1_mlp_2_bias_to_fp16, weight = flow_net_res_blocks_1_mlp_2_weight_to_fp16, x = input_165_cast_fp16)[name = tensor("linear_63_cast_fp16")]; + tensor var_996_cast_fp16 = mul(x = var_967_cast_fp16_2, y = linear_63_cast_fp16)[name = tensor("op_996_cast_fp16")]; + tensor x_105_cast_fp16 = add(x = x_99_cast_fp16, y = var_996_cast_fp16)[name = tensor("x_105_cast_fp16")]; + tensor linear_64_cast_fp16 = linear(bias = flow_net_res_blocks_2_adaLN_modulation_1_bias_to_fp16, weight = flow_net_res_blocks_2_adaLN_modulation_1_weight_to_fp16, x = input_151_cast_fp16)[name = tensor("linear_64_cast_fp16")]; + tensor var_1006_split_sizes_0 = const()[name = tensor("op_1006_split_sizes_0"), val = tensor([512, 512, 512])]; + tensor var_1006_axis_0 = const()[name = tensor("op_1006_axis_0"), val = tensor(-1)]; + tensor var_1006_cast_fp16_0, tensor var_1006_cast_fp16_1, tensor var_1006_cast_fp16_2 = split(axis = var_1006_axis_0, split_sizes = var_1006_split_sizes_0, x = linear_64_cast_fp16)[name = tensor("op_1006_cast_fp16")]; + tensor mean_33_axes_0 = const()[name = tensor("mean_33_axes_0"), val = tensor([-1])]; + tensor mean_33_keep_dims_0 = const()[name = tensor("mean_33_keep_dims_0"), val = tensor(true)]; + tensor mean_33_cast_fp16 = reduce_mean(axes = mean_33_axes_0, keep_dims = mean_33_keep_dims_0, x = x_105_cast_fp16)[name = tensor("mean_33_cast_fp16")]; + tensor sub_28_cast_fp16 = sub(x = x_105_cast_fp16, y = mean_33_cast_fp16)[name = tensor("sub_28_cast_fp16")]; + tensor square_22_cast_fp16 = square(x = sub_28_cast_fp16)[name = tensor("square_22_cast_fp16")]; + tensor reduce_mean_45_axes_0 = const()[name = tensor("reduce_mean_45_axes_0"), val = tensor([-1])]; + tensor reduce_mean_45_keep_dims_0 = const()[name = tensor("reduce_mean_45_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_45_cast_fp16 = reduce_mean(axes = reduce_mean_45_axes_0, keep_dims = reduce_mean_45_keep_dims_0, x = square_22_cast_fp16)[name = tensor("reduce_mean_45_cast_fp16")]; + tensor var_1016_to_fp16 = const()[name = tensor("op_1016_to_fp16"), val = tensor(0x1.1p-20)]; + tensor var_1017_cast_fp16 = add(x = reduce_mean_45_cast_fp16, y = var_1016_to_fp16)[name = tensor("op_1017_cast_fp16")]; + tensor var_1018_cast_fp16 = sqrt(x = var_1017_cast_fp16)[name = tensor("op_1018_cast_fp16")]; + tensor x_107_cast_fp16 = real_div(x = sub_28_cast_fp16, y = var_1018_cast_fp16)[name = tensor("x_107_cast_fp16")]; + tensor var_1020_cast_fp16 = mul(x = x_107_cast_fp16, y = flow_net_res_blocks_2_in_ln_weight_to_fp16)[name = tensor("op_1020_cast_fp16")]; + tensor x_109_cast_fp16 = add(x = var_1020_cast_fp16, y = flow_net_res_blocks_2_in_ln_bias_to_fp16)[name = tensor("x_109_cast_fp16")]; + tensor var_1022_promoted_to_fp16 = const()[name = tensor("op_1022_promoted_to_fp16"), val = tensor(0x1p+0)]; + tensor var_1023_cast_fp16 = add(x = var_1006_cast_fp16_1, y = var_1022_promoted_to_fp16)[name = tensor("op_1023_cast_fp16")]; + tensor var_1024_cast_fp16 = mul(x = x_109_cast_fp16, y = var_1023_cast_fp16)[name = tensor("op_1024_cast_fp16")]; + tensor input_169_cast_fp16 = add(x = var_1024_cast_fp16, y = var_1006_cast_fp16_0)[name = tensor("input_169_cast_fp16")]; + tensor linear_65_cast_fp16 = linear(bias = flow_net_res_blocks_2_mlp_0_bias_to_fp16, weight = flow_net_res_blocks_2_mlp_0_weight_to_fp16, x = input_169_cast_fp16)[name = tensor("linear_65_cast_fp16")]; + tensor input_173_cast_fp16 = silu(x = linear_65_cast_fp16)[name = tensor("input_173_cast_fp16")]; + tensor linear_66_cast_fp16 = linear(bias = flow_net_res_blocks_2_mlp_2_bias_to_fp16, weight = flow_net_res_blocks_2_mlp_2_weight_to_fp16, x = input_173_cast_fp16)[name = tensor("linear_66_cast_fp16")]; + tensor var_1035_cast_fp16 = mul(x = var_1006_cast_fp16_2, y = linear_66_cast_fp16)[name = tensor("op_1035_cast_fp16")]; + tensor x_111_cast_fp16 = add(x = x_105_cast_fp16, y = var_1035_cast_fp16)[name = tensor("x_111_cast_fp16")]; + tensor linear_67_cast_fp16 = linear(bias = flow_net_res_blocks_3_adaLN_modulation_1_bias_to_fp16, weight = flow_net_res_blocks_3_adaLN_modulation_1_weight_to_fp16, x = input_151_cast_fp16)[name = tensor("linear_67_cast_fp16")]; + tensor var_1045_split_sizes_0 = const()[name = tensor("op_1045_split_sizes_0"), val = tensor([512, 512, 512])]; + tensor var_1045_axis_0 = const()[name = tensor("op_1045_axis_0"), val = tensor(-1)]; + tensor var_1045_cast_fp16_0, tensor var_1045_cast_fp16_1, tensor var_1045_cast_fp16_2 = split(axis = var_1045_axis_0, split_sizes = var_1045_split_sizes_0, x = linear_67_cast_fp16)[name = tensor("op_1045_cast_fp16")]; + tensor mean_35_axes_0 = const()[name = tensor("mean_35_axes_0"), val = tensor([-1])]; + tensor mean_35_keep_dims_0 = const()[name = tensor("mean_35_keep_dims_0"), val = tensor(true)]; + tensor mean_35_cast_fp16 = reduce_mean(axes = mean_35_axes_0, keep_dims = mean_35_keep_dims_0, x = x_111_cast_fp16)[name = tensor("mean_35_cast_fp16")]; + tensor sub_29_cast_fp16 = sub(x = x_111_cast_fp16, y = mean_35_cast_fp16)[name = tensor("sub_29_cast_fp16")]; + tensor square_23_cast_fp16 = square(x = sub_29_cast_fp16)[name = tensor("square_23_cast_fp16")]; + tensor reduce_mean_47_axes_0 = const()[name = tensor("reduce_mean_47_axes_0"), val = tensor([-1])]; + tensor reduce_mean_47_keep_dims_0 = const()[name = tensor("reduce_mean_47_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_47_cast_fp16 = reduce_mean(axes = reduce_mean_47_axes_0, keep_dims = reduce_mean_47_keep_dims_0, x = square_23_cast_fp16)[name = tensor("reduce_mean_47_cast_fp16")]; + tensor var_1055_to_fp16 = const()[name = tensor("op_1055_to_fp16"), val = tensor(0x1.1p-20)]; + tensor var_1056_cast_fp16 = add(x = reduce_mean_47_cast_fp16, y = var_1055_to_fp16)[name = tensor("op_1056_cast_fp16")]; + tensor var_1057_cast_fp16 = sqrt(x = var_1056_cast_fp16)[name = tensor("op_1057_cast_fp16")]; + tensor x_113_cast_fp16 = real_div(x = sub_29_cast_fp16, y = var_1057_cast_fp16)[name = tensor("x_113_cast_fp16")]; + tensor var_1059_cast_fp16 = mul(x = x_113_cast_fp16, y = flow_net_res_blocks_3_in_ln_weight_to_fp16)[name = tensor("op_1059_cast_fp16")]; + tensor x_115_cast_fp16 = add(x = var_1059_cast_fp16, y = flow_net_res_blocks_3_in_ln_bias_to_fp16)[name = tensor("x_115_cast_fp16")]; + tensor var_1061_promoted_to_fp16 = const()[name = tensor("op_1061_promoted_to_fp16"), val = tensor(0x1p+0)]; + tensor var_1062_cast_fp16 = add(x = var_1045_cast_fp16_1, y = var_1061_promoted_to_fp16)[name = tensor("op_1062_cast_fp16")]; + tensor var_1063_cast_fp16 = mul(x = x_115_cast_fp16, y = var_1062_cast_fp16)[name = tensor("op_1063_cast_fp16")]; + tensor input_177_cast_fp16 = add(x = var_1063_cast_fp16, y = var_1045_cast_fp16_0)[name = tensor("input_177_cast_fp16")]; + tensor linear_68_cast_fp16 = linear(bias = flow_net_res_blocks_3_mlp_0_bias_to_fp16, weight = flow_net_res_blocks_3_mlp_0_weight_to_fp16, x = input_177_cast_fp16)[name = tensor("linear_68_cast_fp16")]; + tensor input_181_cast_fp16 = silu(x = linear_68_cast_fp16)[name = tensor("input_181_cast_fp16")]; + tensor linear_69_cast_fp16 = linear(bias = flow_net_res_blocks_3_mlp_2_bias_to_fp16, weight = flow_net_res_blocks_3_mlp_2_weight_to_fp16, x = input_181_cast_fp16)[name = tensor("linear_69_cast_fp16")]; + tensor var_1074_cast_fp16 = mul(x = var_1045_cast_fp16_2, y = linear_69_cast_fp16)[name = tensor("op_1074_cast_fp16")]; + tensor x_117_cast_fp16 = add(x = x_111_cast_fp16, y = var_1074_cast_fp16)[name = tensor("x_117_cast_fp16")]; + tensor linear_70_cast_fp16 = linear(bias = flow_net_res_blocks_4_adaLN_modulation_1_bias_to_fp16, weight = flow_net_res_blocks_4_adaLN_modulation_1_weight_to_fp16, x = input_151_cast_fp16)[name = tensor("linear_70_cast_fp16")]; + tensor var_1084_split_sizes_0 = const()[name = tensor("op_1084_split_sizes_0"), val = tensor([512, 512, 512])]; + tensor var_1084_axis_0 = const()[name = tensor("op_1084_axis_0"), val = tensor(-1)]; + tensor var_1084_cast_fp16_0, tensor var_1084_cast_fp16_1, tensor var_1084_cast_fp16_2 = split(axis = var_1084_axis_0, split_sizes = var_1084_split_sizes_0, x = linear_70_cast_fp16)[name = tensor("op_1084_cast_fp16")]; + tensor mean_37_axes_0 = const()[name = tensor("mean_37_axes_0"), val = tensor([-1])]; + tensor mean_37_keep_dims_0 = const()[name = tensor("mean_37_keep_dims_0"), val = tensor(true)]; + tensor mean_37_cast_fp16 = reduce_mean(axes = mean_37_axes_0, keep_dims = mean_37_keep_dims_0, x = x_117_cast_fp16)[name = tensor("mean_37_cast_fp16")]; + tensor sub_30_cast_fp16 = sub(x = x_117_cast_fp16, y = mean_37_cast_fp16)[name = tensor("sub_30_cast_fp16")]; + tensor square_24_cast_fp16 = square(x = sub_30_cast_fp16)[name = tensor("square_24_cast_fp16")]; + tensor reduce_mean_49_axes_0 = const()[name = tensor("reduce_mean_49_axes_0"), val = tensor([-1])]; + tensor reduce_mean_49_keep_dims_0 = const()[name = tensor("reduce_mean_49_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_49_cast_fp16 = reduce_mean(axes = reduce_mean_49_axes_0, keep_dims = reduce_mean_49_keep_dims_0, x = square_24_cast_fp16)[name = tensor("reduce_mean_49_cast_fp16")]; + tensor var_1094_to_fp16 = const()[name = tensor("op_1094_to_fp16"), val = tensor(0x1.1p-20)]; + tensor var_1095_cast_fp16 = add(x = reduce_mean_49_cast_fp16, y = var_1094_to_fp16)[name = tensor("op_1095_cast_fp16")]; + tensor var_1096_cast_fp16 = sqrt(x = var_1095_cast_fp16)[name = tensor("op_1096_cast_fp16")]; + tensor x_119_cast_fp16 = real_div(x = sub_30_cast_fp16, y = var_1096_cast_fp16)[name = tensor("x_119_cast_fp16")]; + tensor var_1098_cast_fp16 = mul(x = x_119_cast_fp16, y = flow_net_res_blocks_4_in_ln_weight_to_fp16)[name = tensor("op_1098_cast_fp16")]; + tensor x_121_cast_fp16 = add(x = var_1098_cast_fp16, y = flow_net_res_blocks_4_in_ln_bias_to_fp16)[name = tensor("x_121_cast_fp16")]; + tensor var_1100_promoted_to_fp16 = const()[name = tensor("op_1100_promoted_to_fp16"), val = tensor(0x1p+0)]; + tensor var_1101_cast_fp16 = add(x = var_1084_cast_fp16_1, y = var_1100_promoted_to_fp16)[name = tensor("op_1101_cast_fp16")]; + tensor var_1102_cast_fp16 = mul(x = x_121_cast_fp16, y = var_1101_cast_fp16)[name = tensor("op_1102_cast_fp16")]; + tensor input_185_cast_fp16 = add(x = var_1102_cast_fp16, y = var_1084_cast_fp16_0)[name = tensor("input_185_cast_fp16")]; + tensor linear_71_cast_fp16 = linear(bias = flow_net_res_blocks_4_mlp_0_bias_to_fp16, weight = flow_net_res_blocks_4_mlp_0_weight_to_fp16, x = input_185_cast_fp16)[name = tensor("linear_71_cast_fp16")]; + tensor input_189_cast_fp16 = silu(x = linear_71_cast_fp16)[name = tensor("input_189_cast_fp16")]; + tensor linear_72_cast_fp16 = linear(bias = flow_net_res_blocks_4_mlp_2_bias_to_fp16, weight = flow_net_res_blocks_4_mlp_2_weight_to_fp16, x = input_189_cast_fp16)[name = tensor("linear_72_cast_fp16")]; + tensor var_1113_cast_fp16 = mul(x = var_1084_cast_fp16_2, y = linear_72_cast_fp16)[name = tensor("op_1113_cast_fp16")]; + tensor x_123_cast_fp16 = add(x = x_117_cast_fp16, y = var_1113_cast_fp16)[name = tensor("x_123_cast_fp16")]; + tensor linear_73_cast_fp16 = linear(bias = flow_net_res_blocks_5_adaLN_modulation_1_bias_to_fp16, weight = flow_net_res_blocks_5_adaLN_modulation_1_weight_to_fp16, x = input_151_cast_fp16)[name = tensor("linear_73_cast_fp16")]; + tensor var_1123_split_sizes_0 = const()[name = tensor("op_1123_split_sizes_0"), val = tensor([512, 512, 512])]; + tensor var_1123_axis_0 = const()[name = tensor("op_1123_axis_0"), val = tensor(-1)]; + tensor var_1123_cast_fp16_0, tensor var_1123_cast_fp16_1, tensor var_1123_cast_fp16_2 = split(axis = var_1123_axis_0, split_sizes = var_1123_split_sizes_0, x = linear_73_cast_fp16)[name = tensor("op_1123_cast_fp16")]; + tensor mean_39_axes_0 = const()[name = tensor("mean_39_axes_0"), val = tensor([-1])]; + tensor mean_39_keep_dims_0 = const()[name = tensor("mean_39_keep_dims_0"), val = tensor(true)]; + tensor mean_39_cast_fp16 = reduce_mean(axes = mean_39_axes_0, keep_dims = mean_39_keep_dims_0, x = x_123_cast_fp16)[name = tensor("mean_39_cast_fp16")]; + tensor sub_31_cast_fp16 = sub(x = x_123_cast_fp16, y = mean_39_cast_fp16)[name = tensor("sub_31_cast_fp16")]; + tensor square_25_cast_fp16 = square(x = sub_31_cast_fp16)[name = tensor("square_25_cast_fp16")]; + tensor reduce_mean_51_axes_0 = const()[name = tensor("reduce_mean_51_axes_0"), val = tensor([-1])]; + tensor reduce_mean_51_keep_dims_0 = const()[name = tensor("reduce_mean_51_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_51_cast_fp16 = reduce_mean(axes = reduce_mean_51_axes_0, keep_dims = reduce_mean_51_keep_dims_0, x = square_25_cast_fp16)[name = tensor("reduce_mean_51_cast_fp16")]; + tensor var_1133_to_fp16 = const()[name = tensor("op_1133_to_fp16"), val = tensor(0x1.1p-20)]; + tensor var_1134_cast_fp16 = add(x = reduce_mean_51_cast_fp16, y = var_1133_to_fp16)[name = tensor("op_1134_cast_fp16")]; + tensor var_1135_cast_fp16 = sqrt(x = var_1134_cast_fp16)[name = tensor("op_1135_cast_fp16")]; + tensor x_125_cast_fp16 = real_div(x = sub_31_cast_fp16, y = var_1135_cast_fp16)[name = tensor("x_125_cast_fp16")]; + tensor var_1137_cast_fp16 = mul(x = x_125_cast_fp16, y = flow_net_res_blocks_5_in_ln_weight_to_fp16)[name = tensor("op_1137_cast_fp16")]; + tensor x_127_cast_fp16 = add(x = var_1137_cast_fp16, y = flow_net_res_blocks_5_in_ln_bias_to_fp16)[name = tensor("x_127_cast_fp16")]; + tensor var_1139_promoted_to_fp16 = const()[name = tensor("op_1139_promoted_to_fp16"), val = tensor(0x1p+0)]; + tensor var_1140_cast_fp16 = add(x = var_1123_cast_fp16_1, y = var_1139_promoted_to_fp16)[name = tensor("op_1140_cast_fp16")]; + tensor var_1141_cast_fp16 = mul(x = x_127_cast_fp16, y = var_1140_cast_fp16)[name = tensor("op_1141_cast_fp16")]; + tensor input_193_cast_fp16 = add(x = var_1141_cast_fp16, y = var_1123_cast_fp16_0)[name = tensor("input_193_cast_fp16")]; + tensor linear_74_cast_fp16 = linear(bias = flow_net_res_blocks_5_mlp_0_bias_to_fp16, weight = flow_net_res_blocks_5_mlp_0_weight_to_fp16, x = input_193_cast_fp16)[name = tensor("linear_74_cast_fp16")]; + tensor input_197_cast_fp16 = silu(x = linear_74_cast_fp16)[name = tensor("input_197_cast_fp16")]; + tensor linear_75_cast_fp16 = linear(bias = flow_net_res_blocks_5_mlp_2_bias_to_fp16, weight = flow_net_res_blocks_5_mlp_2_weight_to_fp16, x = input_197_cast_fp16)[name = tensor("linear_75_cast_fp16")]; + tensor var_1152_cast_fp16 = mul(x = var_1123_cast_fp16_2, y = linear_75_cast_fp16)[name = tensor("op_1152_cast_fp16")]; + tensor x_129_cast_fp16 = add(x = x_123_cast_fp16, y = var_1152_cast_fp16)[name = tensor("x_129_cast_fp16")]; + tensor linear_76_cast_fp16 = linear(bias = flow_net_final_layer_adaLN_modulation_1_bias_to_fp16, weight = flow_net_final_layer_adaLN_modulation_1_weight_to_fp16, x = input_151_cast_fp16)[name = tensor("linear_76_cast_fp16")]; + tensor var_1161_split_sizes_0 = const()[name = tensor("op_1161_split_sizes_0"), val = tensor([512, 512])]; + tensor var_1161_axis_0 = const()[name = tensor("op_1161_axis_0"), val = tensor(-1)]; + tensor var_1161_cast_fp16_0, tensor var_1161_cast_fp16_1 = split(axis = var_1161_axis_0, split_sizes = var_1161_split_sizes_0, x = linear_76_cast_fp16)[name = tensor("op_1161_cast_fp16")]; + tensor mean_41_axes_0 = const()[name = tensor("mean_41_axes_0"), val = tensor([-1])]; + tensor mean_41_keep_dims_0 = const()[name = tensor("mean_41_keep_dims_0"), val = tensor(true)]; + tensor mean_41_cast_fp16 = reduce_mean(axes = mean_41_axes_0, keep_dims = mean_41_keep_dims_0, x = x_129_cast_fp16)[name = tensor("mean_41_cast_fp16")]; + tensor sub_32_cast_fp16 = sub(x = x_129_cast_fp16, y = mean_41_cast_fp16)[name = tensor("sub_32_cast_fp16")]; + tensor square_26_cast_fp16 = square(x = sub_32_cast_fp16)[name = tensor("square_26_cast_fp16")]; + tensor reduce_mean_53_axes_0 = const()[name = tensor("reduce_mean_53_axes_0"), val = tensor([-1])]; + tensor reduce_mean_53_keep_dims_0 = const()[name = tensor("reduce_mean_53_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_53_cast_fp16 = reduce_mean(axes = reduce_mean_53_axes_0, keep_dims = reduce_mean_53_keep_dims_0, x = square_26_cast_fp16)[name = tensor("reduce_mean_53_cast_fp16")]; + tensor var_1168_to_fp16 = const()[name = tensor("op_1168_to_fp16"), val = tensor(0x1.1p-20)]; + tensor var_1169_cast_fp16 = add(x = reduce_mean_53_cast_fp16, y = var_1168_to_fp16)[name = tensor("op_1169_cast_fp16")]; + tensor var_1170_cast_fp16 = sqrt(x = var_1169_cast_fp16)[name = tensor("op_1170_cast_fp16")]; + tensor x_131_cast_fp16 = real_div(x = sub_32_cast_fp16, y = var_1170_cast_fp16)[name = tensor("x_131_cast_fp16")]; + tensor var_1172_promoted_to_fp16 = const()[name = tensor("op_1172_promoted_to_fp16"), val = tensor(0x1p+0)]; + tensor var_1173_cast_fp16 = add(x = var_1161_cast_fp16_1, y = var_1172_promoted_to_fp16)[name = tensor("op_1173_cast_fp16")]; + tensor var_1174_cast_fp16 = mul(x = x_131_cast_fp16, y = var_1173_cast_fp16)[name = tensor("op_1174_cast_fp16")]; + tensor input_201_cast_fp16 = add(x = var_1174_cast_fp16, y = var_1161_cast_fp16_0)[name = tensor("input_201_cast_fp16")]; + tensor linear_77_cast_fp16 = linear(bias = flow_net_final_layer_linear_bias_to_fp16, weight = flow_net_final_layer_linear_weight_to_fp16, x = input_201_cast_fp16)[name = tensor("linear_77_cast_fp16")]; + tensor var_1185_to_fp16 = const()[name = tensor("op_1185_to_fp16"), val = tensor(0x1p-3)]; + tensor var_1186_cast_fp16 = mul(x = linear_77_cast_fp16, y = var_1185_to_fp16)[name = tensor("op_1186_cast_fp16")]; + tensor input_203_cast_fp16 = add(x = input_135_cast_fp16, y = var_1186_cast_fp16)[name = tensor("input_203_cast_fp16")]; + tensor linear_78_cast_fp16 = linear(bias = flow_net_input_proj_bias_to_fp16, weight = flow_net_input_proj_weight_to_fp16, x = input_203_cast_fp16)[name = tensor("linear_78_cast_fp16")]; + tensor input_207_to_fp16 = const()[name = tensor("input_207_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(19001728)))]; + tensor input_209_cast_fp16 = silu(x = input_207_to_fp16)[name = tensor("input_209_cast_fp16")]; + tensor linear_80_cast_fp16 = linear(bias = flow_net_time_embed_0_mlp_2_bias_to_fp16, weight = flow_net_time_embed_0_mlp_2_weight_to_fp16, x = input_209_cast_fp16)[name = tensor("linear_80_cast_fp16")]; + tensor reduce_mean_54_axes_0 = const()[name = tensor("reduce_mean_54_axes_0"), val = tensor([-1])]; + tensor reduce_mean_54_keep_dims_0 = const()[name = tensor("reduce_mean_54_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_54_cast_fp16 = reduce_mean(axes = reduce_mean_54_axes_0, keep_dims = reduce_mean_54_keep_dims_0, x = linear_80_cast_fp16)[name = tensor("reduce_mean_54_cast_fp16")]; + tensor sub_33_cast_fp16 = sub(x = linear_80_cast_fp16, y = reduce_mean_54_cast_fp16)[name = tensor("sub_33_cast_fp16")]; + tensor square_27_cast_fp16 = square(x = sub_33_cast_fp16)[name = tensor("square_27_cast_fp16")]; + tensor reduce_mean_55_axes_0 = const()[name = tensor("reduce_mean_55_axes_0"), val = tensor([-1])]; + tensor reduce_mean_55_keep_dims_0 = const()[name = tensor("reduce_mean_55_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_55_cast_fp16 = reduce_mean(axes = reduce_mean_55_axes_0, keep_dims = reduce_mean_55_keep_dims_0, x = square_27_cast_fp16)[name = tensor("reduce_mean_55_cast_fp16")]; + tensor real_div_6_to_fp16 = const()[name = tensor("real_div_6_to_fp16"), val = tensor(0x1.008p+0)]; + tensor mul_6_cast_fp16 = mul(x = reduce_mean_55_cast_fp16, y = real_div_6_to_fp16)[name = tensor("mul_6_cast_fp16")]; + tensor var_1252_to_fp16 = const()[name = tensor("op_1252_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_55_cast_fp16 = add(x = mul_6_cast_fp16, y = var_1252_to_fp16)[name = tensor("var_55_cast_fp16")]; + tensor var_1255_epsilon_0 = const()[name = tensor("op_1255_epsilon_0"), val = tensor(0x1.197998p-40)]; + tensor var_1255_cast_fp16 = rsqrt(epsilon = var_1255_epsilon_0, x = var_55_cast_fp16)[name = tensor("op_1255_cast_fp16")]; + tensor var_1256_cast_fp16 = mul(x = const_3_to_fp16, y = var_1255_cast_fp16)[name = tensor("op_1256_cast_fp16")]; + tensor var_1257_cast_fp16 = mul(x = linear_80_cast_fp16, y = var_1256_cast_fp16)[name = tensor("op_1257_cast_fp16")]; + tensor input_213_to_fp16 = const()[name = tensor("input_213_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(19002816)))]; + tensor input_215_cast_fp16 = silu(x = input_213_to_fp16)[name = tensor("input_215_cast_fp16")]; + tensor linear_82_cast_fp16 = linear(bias = flow_net_time_embed_1_mlp_2_bias_to_fp16, weight = flow_net_time_embed_1_mlp_2_weight_to_fp16, x = input_215_cast_fp16)[name = tensor("linear_82_cast_fp16")]; + tensor reduce_mean_56_axes_0 = const()[name = tensor("reduce_mean_56_axes_0"), val = tensor([-1])]; + tensor reduce_mean_56_keep_dims_0 = const()[name = tensor("reduce_mean_56_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_56_cast_fp16 = reduce_mean(axes = reduce_mean_56_axes_0, keep_dims = reduce_mean_56_keep_dims_0, x = linear_82_cast_fp16)[name = tensor("reduce_mean_56_cast_fp16")]; + tensor sub_35_cast_fp16 = sub(x = linear_82_cast_fp16, y = reduce_mean_56_cast_fp16)[name = tensor("sub_35_cast_fp16")]; + tensor square_28_cast_fp16 = square(x = sub_35_cast_fp16)[name = tensor("square_28_cast_fp16")]; + tensor reduce_mean_57_axes_0 = const()[name = tensor("reduce_mean_57_axes_0"), val = tensor([-1])]; + tensor reduce_mean_57_keep_dims_0 = const()[name = tensor("reduce_mean_57_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_57_cast_fp16 = reduce_mean(axes = reduce_mean_57_axes_0, keep_dims = reduce_mean_57_keep_dims_0, x = square_28_cast_fp16)[name = tensor("reduce_mean_57_cast_fp16")]; + tensor real_div_7_to_fp16 = const()[name = tensor("real_div_7_to_fp16"), val = tensor(0x1.008p+0)]; + tensor mul_7_cast_fp16 = mul(x = reduce_mean_57_cast_fp16, y = real_div_7_to_fp16)[name = tensor("mul_7_cast_fp16")]; + tensor var_1289_to_fp16 = const()[name = tensor("op_1289_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_57_cast_fp16 = add(x = mul_7_cast_fp16, y = var_1289_to_fp16)[name = tensor("var_57_cast_fp16")]; + tensor var_1292_epsilon_0 = const()[name = tensor("op_1292_epsilon_0"), val = tensor(0x1.197998p-40)]; + tensor var_1292_cast_fp16 = rsqrt(epsilon = var_1292_epsilon_0, x = var_57_cast_fp16)[name = tensor("op_1292_cast_fp16")]; + tensor var_1293_cast_fp16 = mul(x = const_5_to_fp16, y = var_1292_cast_fp16)[name = tensor("op_1293_cast_fp16")]; + tensor var_1294_cast_fp16 = mul(x = linear_82_cast_fp16, y = var_1293_cast_fp16)[name = tensor("op_1294_cast_fp16")]; + tensor var_1306_cast_fp16 = add(x = var_1257_cast_fp16, y = var_1294_cast_fp16)[name = tensor("op_1306_cast_fp16")]; + tensor _inversed_t_combined_7_y_0_to_fp16 = const()[name = tensor("_inversed_t_combined_7_y_0_to_fp16"), val = tensor(0x1p-1)]; + tensor _inversed_t_combined_7_cast_fp16 = mul(x = var_1306_cast_fp16, y = _inversed_t_combined_7_y_0_to_fp16)[name = tensor("_inversed_t_combined_7_cast_fp16")]; + tensor input_217_cast_fp16 = add(x = _inversed_t_combined_7_cast_fp16, y = linear_5_cast_fp16)[name = tensor("input_217_cast_fp16")]; + tensor input_219_cast_fp16 = silu(x = input_217_cast_fp16)[name = tensor("input_219_cast_fp16")]; + tensor linear_84_cast_fp16 = linear(bias = flow_net_res_blocks_0_adaLN_modulation_1_bias_to_fp16, weight = flow_net_res_blocks_0_adaLN_modulation_1_weight_to_fp16, x = input_219_cast_fp16)[name = tensor("linear_84_cast_fp16")]; + tensor var_1321_split_sizes_0 = const()[name = tensor("op_1321_split_sizes_0"), val = tensor([512, 512, 512])]; + tensor var_1321_axis_0 = const()[name = tensor("op_1321_axis_0"), val = tensor(-1)]; + tensor var_1321_cast_fp16_0, tensor var_1321_cast_fp16_1, tensor var_1321_cast_fp16_2 = split(axis = var_1321_axis_0, split_sizes = var_1321_split_sizes_0, x = linear_84_cast_fp16)[name = tensor("op_1321_cast_fp16")]; + tensor mean_43_axes_0 = const()[name = tensor("mean_43_axes_0"), val = tensor([-1])]; + tensor mean_43_keep_dims_0 = const()[name = tensor("mean_43_keep_dims_0"), val = tensor(true)]; + tensor mean_43_cast_fp16 = reduce_mean(axes = mean_43_axes_0, keep_dims = mean_43_keep_dims_0, x = linear_78_cast_fp16)[name = tensor("mean_43_cast_fp16")]; + tensor sub_37_cast_fp16 = sub(x = linear_78_cast_fp16, y = mean_43_cast_fp16)[name = tensor("sub_37_cast_fp16")]; + tensor square_29_cast_fp16 = square(x = sub_37_cast_fp16)[name = tensor("square_29_cast_fp16")]; + tensor reduce_mean_59_axes_0 = const()[name = tensor("reduce_mean_59_axes_0"), val = tensor([-1])]; + tensor reduce_mean_59_keep_dims_0 = const()[name = tensor("reduce_mean_59_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_59_cast_fp16 = reduce_mean(axes = reduce_mean_59_axes_0, keep_dims = reduce_mean_59_keep_dims_0, x = square_29_cast_fp16)[name = tensor("reduce_mean_59_cast_fp16")]; + tensor var_1331_to_fp16 = const()[name = tensor("op_1331_to_fp16"), val = tensor(0x1.1p-20)]; + tensor var_1332_cast_fp16 = add(x = reduce_mean_59_cast_fp16, y = var_1331_to_fp16)[name = tensor("op_1332_cast_fp16")]; + tensor var_1333_cast_fp16 = sqrt(x = var_1332_cast_fp16)[name = tensor("op_1333_cast_fp16")]; + tensor x_139_cast_fp16 = real_div(x = sub_37_cast_fp16, y = var_1333_cast_fp16)[name = tensor("x_139_cast_fp16")]; + tensor var_1335_cast_fp16 = mul(x = x_139_cast_fp16, y = flow_net_res_blocks_0_in_ln_weight_to_fp16)[name = tensor("op_1335_cast_fp16")]; + tensor x_141_cast_fp16 = add(x = var_1335_cast_fp16, y = flow_net_res_blocks_0_in_ln_bias_to_fp16)[name = tensor("x_141_cast_fp16")]; + tensor var_1337_promoted_to_fp16 = const()[name = tensor("op_1337_promoted_to_fp16"), val = tensor(0x1p+0)]; + tensor var_1338_cast_fp16 = add(x = var_1321_cast_fp16_1, y = var_1337_promoted_to_fp16)[name = tensor("op_1338_cast_fp16")]; + tensor var_1339_cast_fp16 = mul(x = x_141_cast_fp16, y = var_1338_cast_fp16)[name = tensor("op_1339_cast_fp16")]; + tensor input_221_cast_fp16 = add(x = var_1339_cast_fp16, y = var_1321_cast_fp16_0)[name = tensor("input_221_cast_fp16")]; + tensor linear_85_cast_fp16 = linear(bias = flow_net_res_blocks_0_mlp_0_bias_to_fp16, weight = flow_net_res_blocks_0_mlp_0_weight_to_fp16, x = input_221_cast_fp16)[name = tensor("linear_85_cast_fp16")]; + tensor input_225_cast_fp16 = silu(x = linear_85_cast_fp16)[name = tensor("input_225_cast_fp16")]; + tensor linear_86_cast_fp16 = linear(bias = flow_net_res_blocks_0_mlp_2_bias_to_fp16, weight = flow_net_res_blocks_0_mlp_2_weight_to_fp16, x = input_225_cast_fp16)[name = tensor("linear_86_cast_fp16")]; + tensor var_1350_cast_fp16 = mul(x = var_1321_cast_fp16_2, y = linear_86_cast_fp16)[name = tensor("op_1350_cast_fp16")]; + tensor x_143_cast_fp16 = add(x = linear_78_cast_fp16, y = var_1350_cast_fp16)[name = tensor("x_143_cast_fp16")]; + tensor linear_87_cast_fp16 = linear(bias = flow_net_res_blocks_1_adaLN_modulation_1_bias_to_fp16, weight = flow_net_res_blocks_1_adaLN_modulation_1_weight_to_fp16, x = input_219_cast_fp16)[name = tensor("linear_87_cast_fp16")]; + tensor var_1360_split_sizes_0 = const()[name = tensor("op_1360_split_sizes_0"), val = tensor([512, 512, 512])]; + tensor var_1360_axis_0 = const()[name = tensor("op_1360_axis_0"), val = tensor(-1)]; + tensor var_1360_cast_fp16_0, tensor var_1360_cast_fp16_1, tensor var_1360_cast_fp16_2 = split(axis = var_1360_axis_0, split_sizes = var_1360_split_sizes_0, x = linear_87_cast_fp16)[name = tensor("op_1360_cast_fp16")]; + tensor mean_45_axes_0 = const()[name = tensor("mean_45_axes_0"), val = tensor([-1])]; + tensor mean_45_keep_dims_0 = const()[name = tensor("mean_45_keep_dims_0"), val = tensor(true)]; + tensor mean_45_cast_fp16 = reduce_mean(axes = mean_45_axes_0, keep_dims = mean_45_keep_dims_0, x = x_143_cast_fp16)[name = tensor("mean_45_cast_fp16")]; + tensor sub_38_cast_fp16 = sub(x = x_143_cast_fp16, y = mean_45_cast_fp16)[name = tensor("sub_38_cast_fp16")]; + tensor square_30_cast_fp16 = square(x = sub_38_cast_fp16)[name = tensor("square_30_cast_fp16")]; + tensor reduce_mean_61_axes_0 = const()[name = tensor("reduce_mean_61_axes_0"), val = tensor([-1])]; + tensor reduce_mean_61_keep_dims_0 = const()[name = tensor("reduce_mean_61_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_61_cast_fp16 = reduce_mean(axes = reduce_mean_61_axes_0, keep_dims = reduce_mean_61_keep_dims_0, x = square_30_cast_fp16)[name = tensor("reduce_mean_61_cast_fp16")]; + tensor var_1370_to_fp16 = const()[name = tensor("op_1370_to_fp16"), val = tensor(0x1.1p-20)]; + tensor var_1371_cast_fp16 = add(x = reduce_mean_61_cast_fp16, y = var_1370_to_fp16)[name = tensor("op_1371_cast_fp16")]; + tensor var_1372_cast_fp16 = sqrt(x = var_1371_cast_fp16)[name = tensor("op_1372_cast_fp16")]; + tensor x_145_cast_fp16 = real_div(x = sub_38_cast_fp16, y = var_1372_cast_fp16)[name = tensor("x_145_cast_fp16")]; + tensor var_1374_cast_fp16 = mul(x = x_145_cast_fp16, y = flow_net_res_blocks_1_in_ln_weight_to_fp16)[name = tensor("op_1374_cast_fp16")]; + tensor x_147_cast_fp16 = add(x = var_1374_cast_fp16, y = flow_net_res_blocks_1_in_ln_bias_to_fp16)[name = tensor("x_147_cast_fp16")]; + tensor var_1376_promoted_to_fp16 = const()[name = tensor("op_1376_promoted_to_fp16"), val = tensor(0x1p+0)]; + tensor var_1377_cast_fp16 = add(x = var_1360_cast_fp16_1, y = var_1376_promoted_to_fp16)[name = tensor("op_1377_cast_fp16")]; + tensor var_1378_cast_fp16 = mul(x = x_147_cast_fp16, y = var_1377_cast_fp16)[name = tensor("op_1378_cast_fp16")]; + tensor input_229_cast_fp16 = add(x = var_1378_cast_fp16, y = var_1360_cast_fp16_0)[name = tensor("input_229_cast_fp16")]; + tensor linear_88_cast_fp16 = linear(bias = flow_net_res_blocks_1_mlp_0_bias_to_fp16, weight = flow_net_res_blocks_1_mlp_0_weight_to_fp16, x = input_229_cast_fp16)[name = tensor("linear_88_cast_fp16")]; + tensor input_233_cast_fp16 = silu(x = linear_88_cast_fp16)[name = tensor("input_233_cast_fp16")]; + tensor linear_89_cast_fp16 = linear(bias = flow_net_res_blocks_1_mlp_2_bias_to_fp16, weight = flow_net_res_blocks_1_mlp_2_weight_to_fp16, x = input_233_cast_fp16)[name = tensor("linear_89_cast_fp16")]; + tensor var_1389_cast_fp16 = mul(x = var_1360_cast_fp16_2, y = linear_89_cast_fp16)[name = tensor("op_1389_cast_fp16")]; + tensor x_149_cast_fp16 = add(x = x_143_cast_fp16, y = var_1389_cast_fp16)[name = tensor("x_149_cast_fp16")]; + tensor linear_90_cast_fp16 = linear(bias = flow_net_res_blocks_2_adaLN_modulation_1_bias_to_fp16, weight = flow_net_res_blocks_2_adaLN_modulation_1_weight_to_fp16, x = input_219_cast_fp16)[name = tensor("linear_90_cast_fp16")]; + tensor var_1399_split_sizes_0 = const()[name = tensor("op_1399_split_sizes_0"), val = tensor([512, 512, 512])]; + tensor var_1399_axis_0 = const()[name = tensor("op_1399_axis_0"), val = tensor(-1)]; + tensor var_1399_cast_fp16_0, tensor var_1399_cast_fp16_1, tensor var_1399_cast_fp16_2 = split(axis = var_1399_axis_0, split_sizes = var_1399_split_sizes_0, x = linear_90_cast_fp16)[name = tensor("op_1399_cast_fp16")]; + tensor mean_47_axes_0 = const()[name = tensor("mean_47_axes_0"), val = tensor([-1])]; + tensor mean_47_keep_dims_0 = const()[name = tensor("mean_47_keep_dims_0"), val = tensor(true)]; + tensor mean_47_cast_fp16 = reduce_mean(axes = mean_47_axes_0, keep_dims = mean_47_keep_dims_0, x = x_149_cast_fp16)[name = tensor("mean_47_cast_fp16")]; + tensor sub_39_cast_fp16 = sub(x = x_149_cast_fp16, y = mean_47_cast_fp16)[name = tensor("sub_39_cast_fp16")]; + tensor square_31_cast_fp16 = square(x = sub_39_cast_fp16)[name = tensor("square_31_cast_fp16")]; + tensor reduce_mean_63_axes_0 = const()[name = tensor("reduce_mean_63_axes_0"), val = tensor([-1])]; + tensor reduce_mean_63_keep_dims_0 = const()[name = tensor("reduce_mean_63_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_63_cast_fp16 = reduce_mean(axes = reduce_mean_63_axes_0, keep_dims = reduce_mean_63_keep_dims_0, x = square_31_cast_fp16)[name = tensor("reduce_mean_63_cast_fp16")]; + tensor var_1409_to_fp16 = const()[name = tensor("op_1409_to_fp16"), val = tensor(0x1.1p-20)]; + tensor var_1410_cast_fp16 = add(x = reduce_mean_63_cast_fp16, y = var_1409_to_fp16)[name = tensor("op_1410_cast_fp16")]; + tensor var_1411_cast_fp16 = sqrt(x = var_1410_cast_fp16)[name = tensor("op_1411_cast_fp16")]; + tensor x_151_cast_fp16 = real_div(x = sub_39_cast_fp16, y = var_1411_cast_fp16)[name = tensor("x_151_cast_fp16")]; + tensor var_1413_cast_fp16 = mul(x = x_151_cast_fp16, y = flow_net_res_blocks_2_in_ln_weight_to_fp16)[name = tensor("op_1413_cast_fp16")]; + tensor x_153_cast_fp16 = add(x = var_1413_cast_fp16, y = flow_net_res_blocks_2_in_ln_bias_to_fp16)[name = tensor("x_153_cast_fp16")]; + tensor var_1415_promoted_to_fp16 = const()[name = tensor("op_1415_promoted_to_fp16"), val = tensor(0x1p+0)]; + tensor var_1416_cast_fp16 = add(x = var_1399_cast_fp16_1, y = var_1415_promoted_to_fp16)[name = tensor("op_1416_cast_fp16")]; + tensor var_1417_cast_fp16 = mul(x = x_153_cast_fp16, y = var_1416_cast_fp16)[name = tensor("op_1417_cast_fp16")]; + tensor input_237_cast_fp16 = add(x = var_1417_cast_fp16, y = var_1399_cast_fp16_0)[name = tensor("input_237_cast_fp16")]; + tensor linear_91_cast_fp16 = linear(bias = flow_net_res_blocks_2_mlp_0_bias_to_fp16, weight = flow_net_res_blocks_2_mlp_0_weight_to_fp16, x = input_237_cast_fp16)[name = tensor("linear_91_cast_fp16")]; + tensor input_241_cast_fp16 = silu(x = linear_91_cast_fp16)[name = tensor("input_241_cast_fp16")]; + tensor linear_92_cast_fp16 = linear(bias = flow_net_res_blocks_2_mlp_2_bias_to_fp16, weight = flow_net_res_blocks_2_mlp_2_weight_to_fp16, x = input_241_cast_fp16)[name = tensor("linear_92_cast_fp16")]; + tensor var_1428_cast_fp16 = mul(x = var_1399_cast_fp16_2, y = linear_92_cast_fp16)[name = tensor("op_1428_cast_fp16")]; + tensor x_155_cast_fp16 = add(x = x_149_cast_fp16, y = var_1428_cast_fp16)[name = tensor("x_155_cast_fp16")]; + tensor linear_93_cast_fp16 = linear(bias = flow_net_res_blocks_3_adaLN_modulation_1_bias_to_fp16, weight = flow_net_res_blocks_3_adaLN_modulation_1_weight_to_fp16, x = input_219_cast_fp16)[name = tensor("linear_93_cast_fp16")]; + tensor var_1438_split_sizes_0 = const()[name = tensor("op_1438_split_sizes_0"), val = tensor([512, 512, 512])]; + tensor var_1438_axis_0 = const()[name = tensor("op_1438_axis_0"), val = tensor(-1)]; + tensor var_1438_cast_fp16_0, tensor var_1438_cast_fp16_1, tensor var_1438_cast_fp16_2 = split(axis = var_1438_axis_0, split_sizes = var_1438_split_sizes_0, x = linear_93_cast_fp16)[name = tensor("op_1438_cast_fp16")]; + tensor mean_49_axes_0 = const()[name = tensor("mean_49_axes_0"), val = tensor([-1])]; + tensor mean_49_keep_dims_0 = const()[name = tensor("mean_49_keep_dims_0"), val = tensor(true)]; + tensor mean_49_cast_fp16 = reduce_mean(axes = mean_49_axes_0, keep_dims = mean_49_keep_dims_0, x = x_155_cast_fp16)[name = tensor("mean_49_cast_fp16")]; + tensor sub_40_cast_fp16 = sub(x = x_155_cast_fp16, y = mean_49_cast_fp16)[name = tensor("sub_40_cast_fp16")]; + tensor square_32_cast_fp16 = square(x = sub_40_cast_fp16)[name = tensor("square_32_cast_fp16")]; + tensor reduce_mean_65_axes_0 = const()[name = tensor("reduce_mean_65_axes_0"), val = tensor([-1])]; + tensor reduce_mean_65_keep_dims_0 = const()[name = tensor("reduce_mean_65_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_65_cast_fp16 = reduce_mean(axes = reduce_mean_65_axes_0, keep_dims = reduce_mean_65_keep_dims_0, x = square_32_cast_fp16)[name = tensor("reduce_mean_65_cast_fp16")]; + tensor var_1448_to_fp16 = const()[name = tensor("op_1448_to_fp16"), val = tensor(0x1.1p-20)]; + tensor var_1449_cast_fp16 = add(x = reduce_mean_65_cast_fp16, y = var_1448_to_fp16)[name = tensor("op_1449_cast_fp16")]; + tensor var_1450_cast_fp16 = sqrt(x = var_1449_cast_fp16)[name = tensor("op_1450_cast_fp16")]; + tensor x_157_cast_fp16 = real_div(x = sub_40_cast_fp16, y = var_1450_cast_fp16)[name = tensor("x_157_cast_fp16")]; + tensor var_1452_cast_fp16 = mul(x = x_157_cast_fp16, y = flow_net_res_blocks_3_in_ln_weight_to_fp16)[name = tensor("op_1452_cast_fp16")]; + tensor x_159_cast_fp16 = add(x = var_1452_cast_fp16, y = flow_net_res_blocks_3_in_ln_bias_to_fp16)[name = tensor("x_159_cast_fp16")]; + tensor var_1454_promoted_to_fp16 = const()[name = tensor("op_1454_promoted_to_fp16"), val = tensor(0x1p+0)]; + tensor var_1455_cast_fp16 = add(x = var_1438_cast_fp16_1, y = var_1454_promoted_to_fp16)[name = tensor("op_1455_cast_fp16")]; + tensor var_1456_cast_fp16 = mul(x = x_159_cast_fp16, y = var_1455_cast_fp16)[name = tensor("op_1456_cast_fp16")]; + tensor input_245_cast_fp16 = add(x = var_1456_cast_fp16, y = var_1438_cast_fp16_0)[name = tensor("input_245_cast_fp16")]; + tensor linear_94_cast_fp16 = linear(bias = flow_net_res_blocks_3_mlp_0_bias_to_fp16, weight = flow_net_res_blocks_3_mlp_0_weight_to_fp16, x = input_245_cast_fp16)[name = tensor("linear_94_cast_fp16")]; + tensor input_249_cast_fp16 = silu(x = linear_94_cast_fp16)[name = tensor("input_249_cast_fp16")]; + tensor linear_95_cast_fp16 = linear(bias = flow_net_res_blocks_3_mlp_2_bias_to_fp16, weight = flow_net_res_blocks_3_mlp_2_weight_to_fp16, x = input_249_cast_fp16)[name = tensor("linear_95_cast_fp16")]; + tensor var_1467_cast_fp16 = mul(x = var_1438_cast_fp16_2, y = linear_95_cast_fp16)[name = tensor("op_1467_cast_fp16")]; + tensor x_161_cast_fp16 = add(x = x_155_cast_fp16, y = var_1467_cast_fp16)[name = tensor("x_161_cast_fp16")]; + tensor linear_96_cast_fp16 = linear(bias = flow_net_res_blocks_4_adaLN_modulation_1_bias_to_fp16, weight = flow_net_res_blocks_4_adaLN_modulation_1_weight_to_fp16, x = input_219_cast_fp16)[name = tensor("linear_96_cast_fp16")]; + tensor var_1477_split_sizes_0 = const()[name = tensor("op_1477_split_sizes_0"), val = tensor([512, 512, 512])]; + tensor var_1477_axis_0 = const()[name = tensor("op_1477_axis_0"), val = tensor(-1)]; + tensor var_1477_cast_fp16_0, tensor var_1477_cast_fp16_1, tensor var_1477_cast_fp16_2 = split(axis = var_1477_axis_0, split_sizes = var_1477_split_sizes_0, x = linear_96_cast_fp16)[name = tensor("op_1477_cast_fp16")]; + tensor mean_51_axes_0 = const()[name = tensor("mean_51_axes_0"), val = tensor([-1])]; + tensor mean_51_keep_dims_0 = const()[name = tensor("mean_51_keep_dims_0"), val = tensor(true)]; + tensor mean_51_cast_fp16 = reduce_mean(axes = mean_51_axes_0, keep_dims = mean_51_keep_dims_0, x = x_161_cast_fp16)[name = tensor("mean_51_cast_fp16")]; + tensor sub_41_cast_fp16 = sub(x = x_161_cast_fp16, y = mean_51_cast_fp16)[name = tensor("sub_41_cast_fp16")]; + tensor square_33_cast_fp16 = square(x = sub_41_cast_fp16)[name = tensor("square_33_cast_fp16")]; + tensor reduce_mean_67_axes_0 = const()[name = tensor("reduce_mean_67_axes_0"), val = tensor([-1])]; + tensor reduce_mean_67_keep_dims_0 = const()[name = tensor("reduce_mean_67_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_67_cast_fp16 = reduce_mean(axes = reduce_mean_67_axes_0, keep_dims = reduce_mean_67_keep_dims_0, x = square_33_cast_fp16)[name = tensor("reduce_mean_67_cast_fp16")]; + tensor var_1487_to_fp16 = const()[name = tensor("op_1487_to_fp16"), val = tensor(0x1.1p-20)]; + tensor var_1488_cast_fp16 = add(x = reduce_mean_67_cast_fp16, y = var_1487_to_fp16)[name = tensor("op_1488_cast_fp16")]; + tensor var_1489_cast_fp16 = sqrt(x = var_1488_cast_fp16)[name = tensor("op_1489_cast_fp16")]; + tensor x_163_cast_fp16 = real_div(x = sub_41_cast_fp16, y = var_1489_cast_fp16)[name = tensor("x_163_cast_fp16")]; + tensor var_1491_cast_fp16 = mul(x = x_163_cast_fp16, y = flow_net_res_blocks_4_in_ln_weight_to_fp16)[name = tensor("op_1491_cast_fp16")]; + tensor x_165_cast_fp16 = add(x = var_1491_cast_fp16, y = flow_net_res_blocks_4_in_ln_bias_to_fp16)[name = tensor("x_165_cast_fp16")]; + tensor var_1493_promoted_to_fp16 = const()[name = tensor("op_1493_promoted_to_fp16"), val = tensor(0x1p+0)]; + tensor var_1494_cast_fp16 = add(x = var_1477_cast_fp16_1, y = var_1493_promoted_to_fp16)[name = tensor("op_1494_cast_fp16")]; + tensor var_1495_cast_fp16 = mul(x = x_165_cast_fp16, y = var_1494_cast_fp16)[name = tensor("op_1495_cast_fp16")]; + tensor input_253_cast_fp16 = add(x = var_1495_cast_fp16, y = var_1477_cast_fp16_0)[name = tensor("input_253_cast_fp16")]; + tensor linear_97_cast_fp16 = linear(bias = flow_net_res_blocks_4_mlp_0_bias_to_fp16, weight = flow_net_res_blocks_4_mlp_0_weight_to_fp16, x = input_253_cast_fp16)[name = tensor("linear_97_cast_fp16")]; + tensor input_257_cast_fp16 = silu(x = linear_97_cast_fp16)[name = tensor("input_257_cast_fp16")]; + tensor linear_98_cast_fp16 = linear(bias = flow_net_res_blocks_4_mlp_2_bias_to_fp16, weight = flow_net_res_blocks_4_mlp_2_weight_to_fp16, x = input_257_cast_fp16)[name = tensor("linear_98_cast_fp16")]; + tensor var_1506_cast_fp16 = mul(x = var_1477_cast_fp16_2, y = linear_98_cast_fp16)[name = tensor("op_1506_cast_fp16")]; + tensor x_167_cast_fp16 = add(x = x_161_cast_fp16, y = var_1506_cast_fp16)[name = tensor("x_167_cast_fp16")]; + tensor linear_99_cast_fp16 = linear(bias = flow_net_res_blocks_5_adaLN_modulation_1_bias_to_fp16, weight = flow_net_res_blocks_5_adaLN_modulation_1_weight_to_fp16, x = input_219_cast_fp16)[name = tensor("linear_99_cast_fp16")]; + tensor var_1516_split_sizes_0 = const()[name = tensor("op_1516_split_sizes_0"), val = tensor([512, 512, 512])]; + tensor var_1516_axis_0 = const()[name = tensor("op_1516_axis_0"), val = tensor(-1)]; + tensor var_1516_cast_fp16_0, tensor var_1516_cast_fp16_1, tensor var_1516_cast_fp16_2 = split(axis = var_1516_axis_0, split_sizes = var_1516_split_sizes_0, x = linear_99_cast_fp16)[name = tensor("op_1516_cast_fp16")]; + tensor mean_53_axes_0 = const()[name = tensor("mean_53_axes_0"), val = tensor([-1])]; + tensor mean_53_keep_dims_0 = const()[name = tensor("mean_53_keep_dims_0"), val = tensor(true)]; + tensor mean_53_cast_fp16 = reduce_mean(axes = mean_53_axes_0, keep_dims = mean_53_keep_dims_0, x = x_167_cast_fp16)[name = tensor("mean_53_cast_fp16")]; + tensor sub_42_cast_fp16 = sub(x = x_167_cast_fp16, y = mean_53_cast_fp16)[name = tensor("sub_42_cast_fp16")]; + tensor square_34_cast_fp16 = square(x = sub_42_cast_fp16)[name = tensor("square_34_cast_fp16")]; + tensor reduce_mean_69_axes_0 = const()[name = tensor("reduce_mean_69_axes_0"), val = tensor([-1])]; + tensor reduce_mean_69_keep_dims_0 = const()[name = tensor("reduce_mean_69_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_69_cast_fp16 = reduce_mean(axes = reduce_mean_69_axes_0, keep_dims = reduce_mean_69_keep_dims_0, x = square_34_cast_fp16)[name = tensor("reduce_mean_69_cast_fp16")]; + tensor var_1526_to_fp16 = const()[name = tensor("op_1526_to_fp16"), val = tensor(0x1.1p-20)]; + tensor var_1527_cast_fp16 = add(x = reduce_mean_69_cast_fp16, y = var_1526_to_fp16)[name = tensor("op_1527_cast_fp16")]; + tensor var_1528_cast_fp16 = sqrt(x = var_1527_cast_fp16)[name = tensor("op_1528_cast_fp16")]; + tensor x_169_cast_fp16 = real_div(x = sub_42_cast_fp16, y = var_1528_cast_fp16)[name = tensor("x_169_cast_fp16")]; + tensor var_1530_cast_fp16 = mul(x = x_169_cast_fp16, y = flow_net_res_blocks_5_in_ln_weight_to_fp16)[name = tensor("op_1530_cast_fp16")]; + tensor x_171_cast_fp16 = add(x = var_1530_cast_fp16, y = flow_net_res_blocks_5_in_ln_bias_to_fp16)[name = tensor("x_171_cast_fp16")]; + tensor var_1532_promoted_to_fp16 = const()[name = tensor("op_1532_promoted_to_fp16"), val = tensor(0x1p+0)]; + tensor var_1533_cast_fp16 = add(x = var_1516_cast_fp16_1, y = var_1532_promoted_to_fp16)[name = tensor("op_1533_cast_fp16")]; + tensor var_1534_cast_fp16 = mul(x = x_171_cast_fp16, y = var_1533_cast_fp16)[name = tensor("op_1534_cast_fp16")]; + tensor input_261_cast_fp16 = add(x = var_1534_cast_fp16, y = var_1516_cast_fp16_0)[name = tensor("input_261_cast_fp16")]; + tensor linear_100_cast_fp16 = linear(bias = flow_net_res_blocks_5_mlp_0_bias_to_fp16, weight = flow_net_res_blocks_5_mlp_0_weight_to_fp16, x = input_261_cast_fp16)[name = tensor("linear_100_cast_fp16")]; + tensor input_265_cast_fp16 = silu(x = linear_100_cast_fp16)[name = tensor("input_265_cast_fp16")]; + tensor linear_101_cast_fp16 = linear(bias = flow_net_res_blocks_5_mlp_2_bias_to_fp16, weight = flow_net_res_blocks_5_mlp_2_weight_to_fp16, x = input_265_cast_fp16)[name = tensor("linear_101_cast_fp16")]; + tensor var_1545_cast_fp16 = mul(x = var_1516_cast_fp16_2, y = linear_101_cast_fp16)[name = tensor("op_1545_cast_fp16")]; + tensor x_173_cast_fp16 = add(x = x_167_cast_fp16, y = var_1545_cast_fp16)[name = tensor("x_173_cast_fp16")]; + tensor linear_102_cast_fp16 = linear(bias = flow_net_final_layer_adaLN_modulation_1_bias_to_fp16, weight = flow_net_final_layer_adaLN_modulation_1_weight_to_fp16, x = input_219_cast_fp16)[name = tensor("linear_102_cast_fp16")]; + tensor var_1554_split_sizes_0 = const()[name = tensor("op_1554_split_sizes_0"), val = tensor([512, 512])]; + tensor var_1554_axis_0 = const()[name = tensor("op_1554_axis_0"), val = tensor(-1)]; + tensor var_1554_cast_fp16_0, tensor var_1554_cast_fp16_1 = split(axis = var_1554_axis_0, split_sizes = var_1554_split_sizes_0, x = linear_102_cast_fp16)[name = tensor("op_1554_cast_fp16")]; + tensor mean_55_axes_0 = const()[name = tensor("mean_55_axes_0"), val = tensor([-1])]; + tensor mean_55_keep_dims_0 = const()[name = tensor("mean_55_keep_dims_0"), val = tensor(true)]; + tensor mean_55_cast_fp16 = reduce_mean(axes = mean_55_axes_0, keep_dims = mean_55_keep_dims_0, x = x_173_cast_fp16)[name = tensor("mean_55_cast_fp16")]; + tensor sub_43_cast_fp16 = sub(x = x_173_cast_fp16, y = mean_55_cast_fp16)[name = tensor("sub_43_cast_fp16")]; + tensor square_35_cast_fp16 = square(x = sub_43_cast_fp16)[name = tensor("square_35_cast_fp16")]; + tensor reduce_mean_71_axes_0 = const()[name = tensor("reduce_mean_71_axes_0"), val = tensor([-1])]; + tensor reduce_mean_71_keep_dims_0 = const()[name = tensor("reduce_mean_71_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_71_cast_fp16 = reduce_mean(axes = reduce_mean_71_axes_0, keep_dims = reduce_mean_71_keep_dims_0, x = square_35_cast_fp16)[name = tensor("reduce_mean_71_cast_fp16")]; + tensor var_1561_to_fp16 = const()[name = tensor("op_1561_to_fp16"), val = tensor(0x1.1p-20)]; + tensor var_1562_cast_fp16 = add(x = reduce_mean_71_cast_fp16, y = var_1561_to_fp16)[name = tensor("op_1562_cast_fp16")]; + tensor var_1563_cast_fp16 = sqrt(x = var_1562_cast_fp16)[name = tensor("op_1563_cast_fp16")]; + tensor x_175_cast_fp16 = real_div(x = sub_43_cast_fp16, y = var_1563_cast_fp16)[name = tensor("x_175_cast_fp16")]; + tensor var_1565_promoted_to_fp16 = const()[name = tensor("op_1565_promoted_to_fp16"), val = tensor(0x1p+0)]; + tensor var_1566_cast_fp16 = add(x = var_1554_cast_fp16_1, y = var_1565_promoted_to_fp16)[name = tensor("op_1566_cast_fp16")]; + tensor var_1567_cast_fp16 = mul(x = x_175_cast_fp16, y = var_1566_cast_fp16)[name = tensor("op_1567_cast_fp16")]; + tensor input_269_cast_fp16 = add(x = var_1567_cast_fp16, y = var_1554_cast_fp16_0)[name = tensor("input_269_cast_fp16")]; + tensor linear_103_cast_fp16 = linear(bias = flow_net_final_layer_linear_bias_to_fp16, weight = flow_net_final_layer_linear_weight_to_fp16, x = input_269_cast_fp16)[name = tensor("linear_103_cast_fp16")]; + tensor var_1578_to_fp16 = const()[name = tensor("op_1578_to_fp16"), val = tensor(0x1p-3)]; + tensor var_1579_cast_fp16 = mul(x = linear_103_cast_fp16, y = var_1578_to_fp16)[name = tensor("op_1579_cast_fp16")]; + tensor input_271_cast_fp16 = add(x = input_203_cast_fp16, y = var_1579_cast_fp16)[name = tensor("input_271_cast_fp16")]; + tensor linear_104_cast_fp16 = linear(bias = flow_net_input_proj_bias_to_fp16, weight = flow_net_input_proj_weight_to_fp16, x = input_271_cast_fp16)[name = tensor("linear_104_cast_fp16")]; + tensor input_275_to_fp16 = const()[name = tensor("input_275_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(19003904)))]; + tensor input_277_cast_fp16 = silu(x = input_275_to_fp16)[name = tensor("input_277_cast_fp16")]; + tensor linear_106_cast_fp16 = linear(bias = flow_net_time_embed_0_mlp_2_bias_to_fp16, weight = flow_net_time_embed_0_mlp_2_weight_to_fp16, x = input_277_cast_fp16)[name = tensor("linear_106_cast_fp16")]; + tensor reduce_mean_72_axes_0 = const()[name = tensor("reduce_mean_72_axes_0"), val = tensor([-1])]; + tensor reduce_mean_72_keep_dims_0 = const()[name = tensor("reduce_mean_72_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_72_cast_fp16 = reduce_mean(axes = reduce_mean_72_axes_0, keep_dims = reduce_mean_72_keep_dims_0, x = linear_106_cast_fp16)[name = tensor("reduce_mean_72_cast_fp16")]; + tensor sub_44_cast_fp16 = sub(x = linear_106_cast_fp16, y = reduce_mean_72_cast_fp16)[name = tensor("sub_44_cast_fp16")]; + tensor square_36_cast_fp16 = square(x = sub_44_cast_fp16)[name = tensor("square_36_cast_fp16")]; + tensor reduce_mean_73_axes_0 = const()[name = tensor("reduce_mean_73_axes_0"), val = tensor([-1])]; + tensor reduce_mean_73_keep_dims_0 = const()[name = tensor("reduce_mean_73_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_73_cast_fp16 = reduce_mean(axes = reduce_mean_73_axes_0, keep_dims = reduce_mean_73_keep_dims_0, x = square_36_cast_fp16)[name = tensor("reduce_mean_73_cast_fp16")]; + tensor real_div_8_to_fp16 = const()[name = tensor("real_div_8_to_fp16"), val = tensor(0x1.008p+0)]; + tensor mul_8_cast_fp16 = mul(x = reduce_mean_73_cast_fp16, y = real_div_8_to_fp16)[name = tensor("mul_8_cast_fp16")]; + tensor var_1645_to_fp16 = const()[name = tensor("op_1645_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_73_cast_fp16_0 = add(x = mul_8_cast_fp16, y = var_1645_to_fp16)[name = tensor("var_73_cast_fp16")]; + tensor var_1648_epsilon_0 = const()[name = tensor("op_1648_epsilon_0"), val = tensor(0x1.197998p-40)]; + tensor var_1648_cast_fp16 = rsqrt(epsilon = var_1648_epsilon_0, x = var_73_cast_fp16_0)[name = tensor("op_1648_cast_fp16")]; + tensor var_1649_cast_fp16 = mul(x = const_3_to_fp16, y = var_1648_cast_fp16)[name = tensor("op_1649_cast_fp16")]; + tensor var_1650_cast_fp16 = mul(x = linear_106_cast_fp16, y = var_1649_cast_fp16)[name = tensor("op_1650_cast_fp16")]; + tensor input_281_to_fp16 = const()[name = tensor("input_281_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(19004992)))]; + tensor input_283_cast_fp16 = silu(x = input_281_to_fp16)[name = tensor("input_283_cast_fp16")]; + tensor linear_108_cast_fp16 = linear(bias = flow_net_time_embed_1_mlp_2_bias_to_fp16, weight = flow_net_time_embed_1_mlp_2_weight_to_fp16, x = input_283_cast_fp16)[name = tensor("linear_108_cast_fp16")]; + tensor reduce_mean_74_axes_0 = const()[name = tensor("reduce_mean_74_axes_0"), val = tensor([-1])]; + tensor reduce_mean_74_keep_dims_0 = const()[name = tensor("reduce_mean_74_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_74_cast_fp16 = reduce_mean(axes = reduce_mean_74_axes_0, keep_dims = reduce_mean_74_keep_dims_0, x = linear_108_cast_fp16)[name = tensor("reduce_mean_74_cast_fp16")]; + tensor sub_46_cast_fp16 = sub(x = linear_108_cast_fp16, y = reduce_mean_74_cast_fp16)[name = tensor("sub_46_cast_fp16")]; + tensor square_37_cast_fp16 = square(x = sub_46_cast_fp16)[name = tensor("square_37_cast_fp16")]; + tensor reduce_mean_75_axes_0 = const()[name = tensor("reduce_mean_75_axes_0"), val = tensor([-1])]; + tensor reduce_mean_75_keep_dims_0 = const()[name = tensor("reduce_mean_75_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_75_cast_fp16 = reduce_mean(axes = reduce_mean_75_axes_0, keep_dims = reduce_mean_75_keep_dims_0, x = square_37_cast_fp16)[name = tensor("reduce_mean_75_cast_fp16")]; + tensor real_div_9_to_fp16 = const()[name = tensor("real_div_9_to_fp16"), val = tensor(0x1.008p+0)]; + tensor mul_9_cast_fp16 = mul(x = reduce_mean_75_cast_fp16, y = real_div_9_to_fp16)[name = tensor("mul_9_cast_fp16")]; + tensor var_1682_to_fp16 = const()[name = tensor("op_1682_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_75_cast_fp16_0 = add(x = mul_9_cast_fp16, y = var_1682_to_fp16)[name = tensor("var_75_cast_fp16")]; + tensor var_1685_epsilon_0 = const()[name = tensor("op_1685_epsilon_0"), val = tensor(0x1.197998p-40)]; + tensor var_1685_cast_fp16 = rsqrt(epsilon = var_1685_epsilon_0, x = var_75_cast_fp16_0)[name = tensor("op_1685_cast_fp16")]; + tensor var_1686_cast_fp16 = mul(x = const_5_to_fp16, y = var_1685_cast_fp16)[name = tensor("op_1686_cast_fp16")]; + tensor var_1687_cast_fp16 = mul(x = linear_108_cast_fp16, y = var_1686_cast_fp16)[name = tensor("op_1687_cast_fp16")]; + tensor var_1699_cast_fp16 = add(x = var_1650_cast_fp16, y = var_1687_cast_fp16)[name = tensor("op_1699_cast_fp16")]; + tensor _inversed_t_combined_9_y_0_to_fp16 = const()[name = tensor("_inversed_t_combined_9_y_0_to_fp16"), val = tensor(0x1p-1)]; + tensor _inversed_t_combined_9_cast_fp16 = mul(x = var_1699_cast_fp16, y = _inversed_t_combined_9_y_0_to_fp16)[name = tensor("_inversed_t_combined_9_cast_fp16")]; + tensor input_285_cast_fp16 = add(x = _inversed_t_combined_9_cast_fp16, y = linear_5_cast_fp16)[name = tensor("input_285_cast_fp16")]; + tensor input_287_cast_fp16 = silu(x = input_285_cast_fp16)[name = tensor("input_287_cast_fp16")]; + tensor linear_110_cast_fp16 = linear(bias = flow_net_res_blocks_0_adaLN_modulation_1_bias_to_fp16, weight = flow_net_res_blocks_0_adaLN_modulation_1_weight_to_fp16, x = input_287_cast_fp16)[name = tensor("linear_110_cast_fp16")]; + tensor var_1714_split_sizes_0 = const()[name = tensor("op_1714_split_sizes_0"), val = tensor([512, 512, 512])]; + tensor var_1714_axis_0 = const()[name = tensor("op_1714_axis_0"), val = tensor(-1)]; + tensor var_1714_cast_fp16_0, tensor var_1714_cast_fp16_1, tensor var_1714_cast_fp16_2 = split(axis = var_1714_axis_0, split_sizes = var_1714_split_sizes_0, x = linear_110_cast_fp16)[name = tensor("op_1714_cast_fp16")]; + tensor mean_57_axes_0 = const()[name = tensor("mean_57_axes_0"), val = tensor([-1])]; + tensor mean_57_keep_dims_0 = const()[name = tensor("mean_57_keep_dims_0"), val = tensor(true)]; + tensor mean_57_cast_fp16 = reduce_mean(axes = mean_57_axes_0, keep_dims = mean_57_keep_dims_0, x = linear_104_cast_fp16)[name = tensor("mean_57_cast_fp16")]; + tensor sub_48_cast_fp16 = sub(x = linear_104_cast_fp16, y = mean_57_cast_fp16)[name = tensor("sub_48_cast_fp16")]; + tensor square_38_cast_fp16 = square(x = sub_48_cast_fp16)[name = tensor("square_38_cast_fp16")]; + tensor reduce_mean_77_axes_0 = const()[name = tensor("reduce_mean_77_axes_0"), val = tensor([-1])]; + tensor reduce_mean_77_keep_dims_0 = const()[name = tensor("reduce_mean_77_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_77_cast_fp16 = reduce_mean(axes = reduce_mean_77_axes_0, keep_dims = reduce_mean_77_keep_dims_0, x = square_38_cast_fp16)[name = tensor("reduce_mean_77_cast_fp16")]; + tensor var_1724_to_fp16 = const()[name = tensor("op_1724_to_fp16"), val = tensor(0x1.1p-20)]; + tensor var_1725_cast_fp16 = add(x = reduce_mean_77_cast_fp16, y = var_1724_to_fp16)[name = tensor("op_1725_cast_fp16")]; + tensor var_1726_cast_fp16 = sqrt(x = var_1725_cast_fp16)[name = tensor("op_1726_cast_fp16")]; + tensor x_183_cast_fp16 = real_div(x = sub_48_cast_fp16, y = var_1726_cast_fp16)[name = tensor("x_183_cast_fp16")]; + tensor var_1728_cast_fp16 = mul(x = x_183_cast_fp16, y = flow_net_res_blocks_0_in_ln_weight_to_fp16)[name = tensor("op_1728_cast_fp16")]; + tensor x_185_cast_fp16 = add(x = var_1728_cast_fp16, y = flow_net_res_blocks_0_in_ln_bias_to_fp16)[name = tensor("x_185_cast_fp16")]; + tensor var_1730_promoted_to_fp16 = const()[name = tensor("op_1730_promoted_to_fp16"), val = tensor(0x1p+0)]; + tensor var_1731_cast_fp16 = add(x = var_1714_cast_fp16_1, y = var_1730_promoted_to_fp16)[name = tensor("op_1731_cast_fp16")]; + tensor var_1732_cast_fp16 = mul(x = x_185_cast_fp16, y = var_1731_cast_fp16)[name = tensor("op_1732_cast_fp16")]; + tensor input_289_cast_fp16 = add(x = var_1732_cast_fp16, y = var_1714_cast_fp16_0)[name = tensor("input_289_cast_fp16")]; + tensor linear_111_cast_fp16 = linear(bias = flow_net_res_blocks_0_mlp_0_bias_to_fp16, weight = flow_net_res_blocks_0_mlp_0_weight_to_fp16, x = input_289_cast_fp16)[name = tensor("linear_111_cast_fp16")]; + tensor input_293_cast_fp16 = silu(x = linear_111_cast_fp16)[name = tensor("input_293_cast_fp16")]; + tensor linear_112_cast_fp16 = linear(bias = flow_net_res_blocks_0_mlp_2_bias_to_fp16, weight = flow_net_res_blocks_0_mlp_2_weight_to_fp16, x = input_293_cast_fp16)[name = tensor("linear_112_cast_fp16")]; + tensor var_1743_cast_fp16 = mul(x = var_1714_cast_fp16_2, y = linear_112_cast_fp16)[name = tensor("op_1743_cast_fp16")]; + tensor x_187_cast_fp16 = add(x = linear_104_cast_fp16, y = var_1743_cast_fp16)[name = tensor("x_187_cast_fp16")]; + tensor linear_113_cast_fp16 = linear(bias = flow_net_res_blocks_1_adaLN_modulation_1_bias_to_fp16, weight = flow_net_res_blocks_1_adaLN_modulation_1_weight_to_fp16, x = input_287_cast_fp16)[name = tensor("linear_113_cast_fp16")]; + tensor var_1753_split_sizes_0 = const()[name = tensor("op_1753_split_sizes_0"), val = tensor([512, 512, 512])]; + tensor var_1753_axis_0 = const()[name = tensor("op_1753_axis_0"), val = tensor(-1)]; + tensor var_1753_cast_fp16_0, tensor var_1753_cast_fp16_1, tensor var_1753_cast_fp16_2 = split(axis = var_1753_axis_0, split_sizes = var_1753_split_sizes_0, x = linear_113_cast_fp16)[name = tensor("op_1753_cast_fp16")]; + tensor mean_59_axes_0 = const()[name = tensor("mean_59_axes_0"), val = tensor([-1])]; + tensor mean_59_keep_dims_0 = const()[name = tensor("mean_59_keep_dims_0"), val = tensor(true)]; + tensor mean_59_cast_fp16 = reduce_mean(axes = mean_59_axes_0, keep_dims = mean_59_keep_dims_0, x = x_187_cast_fp16)[name = tensor("mean_59_cast_fp16")]; + tensor sub_49_cast_fp16 = sub(x = x_187_cast_fp16, y = mean_59_cast_fp16)[name = tensor("sub_49_cast_fp16")]; + tensor square_39_cast_fp16 = square(x = sub_49_cast_fp16)[name = tensor("square_39_cast_fp16")]; + tensor reduce_mean_79_axes_0 = const()[name = tensor("reduce_mean_79_axes_0"), val = tensor([-1])]; + tensor reduce_mean_79_keep_dims_0 = const()[name = tensor("reduce_mean_79_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_79_cast_fp16 = reduce_mean(axes = reduce_mean_79_axes_0, keep_dims = reduce_mean_79_keep_dims_0, x = square_39_cast_fp16)[name = tensor("reduce_mean_79_cast_fp16")]; + tensor var_1763_to_fp16 = const()[name = tensor("op_1763_to_fp16"), val = tensor(0x1.1p-20)]; + tensor var_1764_cast_fp16 = add(x = reduce_mean_79_cast_fp16, y = var_1763_to_fp16)[name = tensor("op_1764_cast_fp16")]; + tensor var_1765_cast_fp16 = sqrt(x = var_1764_cast_fp16)[name = tensor("op_1765_cast_fp16")]; + tensor x_189_cast_fp16 = real_div(x = sub_49_cast_fp16, y = var_1765_cast_fp16)[name = tensor("x_189_cast_fp16")]; + tensor var_1767_cast_fp16 = mul(x = x_189_cast_fp16, y = flow_net_res_blocks_1_in_ln_weight_to_fp16)[name = tensor("op_1767_cast_fp16")]; + tensor x_191_cast_fp16 = add(x = var_1767_cast_fp16, y = flow_net_res_blocks_1_in_ln_bias_to_fp16)[name = tensor("x_191_cast_fp16")]; + tensor var_1769_promoted_to_fp16 = const()[name = tensor("op_1769_promoted_to_fp16"), val = tensor(0x1p+0)]; + tensor var_1770_cast_fp16 = add(x = var_1753_cast_fp16_1, y = var_1769_promoted_to_fp16)[name = tensor("op_1770_cast_fp16")]; + tensor var_1771_cast_fp16 = mul(x = x_191_cast_fp16, y = var_1770_cast_fp16)[name = tensor("op_1771_cast_fp16")]; + tensor input_297_cast_fp16 = add(x = var_1771_cast_fp16, y = var_1753_cast_fp16_0)[name = tensor("input_297_cast_fp16")]; + tensor linear_114_cast_fp16 = linear(bias = flow_net_res_blocks_1_mlp_0_bias_to_fp16, weight = flow_net_res_blocks_1_mlp_0_weight_to_fp16, x = input_297_cast_fp16)[name = tensor("linear_114_cast_fp16")]; + tensor input_301_cast_fp16 = silu(x = linear_114_cast_fp16)[name = tensor("input_301_cast_fp16")]; + tensor linear_115_cast_fp16 = linear(bias = flow_net_res_blocks_1_mlp_2_bias_to_fp16, weight = flow_net_res_blocks_1_mlp_2_weight_to_fp16, x = input_301_cast_fp16)[name = tensor("linear_115_cast_fp16")]; + tensor var_1782_cast_fp16 = mul(x = var_1753_cast_fp16_2, y = linear_115_cast_fp16)[name = tensor("op_1782_cast_fp16")]; + tensor x_193_cast_fp16 = add(x = x_187_cast_fp16, y = var_1782_cast_fp16)[name = tensor("x_193_cast_fp16")]; + tensor linear_116_cast_fp16 = linear(bias = flow_net_res_blocks_2_adaLN_modulation_1_bias_to_fp16, weight = flow_net_res_blocks_2_adaLN_modulation_1_weight_to_fp16, x = input_287_cast_fp16)[name = tensor("linear_116_cast_fp16")]; + tensor var_1792_split_sizes_0 = const()[name = tensor("op_1792_split_sizes_0"), val = tensor([512, 512, 512])]; + tensor var_1792_axis_0 = const()[name = tensor("op_1792_axis_0"), val = tensor(-1)]; + tensor var_1792_cast_fp16_0, tensor var_1792_cast_fp16_1, tensor var_1792_cast_fp16_2 = split(axis = var_1792_axis_0, split_sizes = var_1792_split_sizes_0, x = linear_116_cast_fp16)[name = tensor("op_1792_cast_fp16")]; + tensor mean_61_axes_0 = const()[name = tensor("mean_61_axes_0"), val = tensor([-1])]; + tensor mean_61_keep_dims_0 = const()[name = tensor("mean_61_keep_dims_0"), val = tensor(true)]; + tensor mean_61_cast_fp16 = reduce_mean(axes = mean_61_axes_0, keep_dims = mean_61_keep_dims_0, x = x_193_cast_fp16)[name = tensor("mean_61_cast_fp16")]; + tensor sub_50_cast_fp16 = sub(x = x_193_cast_fp16, y = mean_61_cast_fp16)[name = tensor("sub_50_cast_fp16")]; + tensor square_40_cast_fp16 = square(x = sub_50_cast_fp16)[name = tensor("square_40_cast_fp16")]; + tensor reduce_mean_81_axes_0 = const()[name = tensor("reduce_mean_81_axes_0"), val = tensor([-1])]; + tensor reduce_mean_81_keep_dims_0 = const()[name = tensor("reduce_mean_81_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_81_cast_fp16 = reduce_mean(axes = reduce_mean_81_axes_0, keep_dims = reduce_mean_81_keep_dims_0, x = square_40_cast_fp16)[name = tensor("reduce_mean_81_cast_fp16")]; + tensor var_1802_to_fp16 = const()[name = tensor("op_1802_to_fp16"), val = tensor(0x1.1p-20)]; + tensor var_1803_cast_fp16 = add(x = reduce_mean_81_cast_fp16, y = var_1802_to_fp16)[name = tensor("op_1803_cast_fp16")]; + tensor var_1804_cast_fp16 = sqrt(x = var_1803_cast_fp16)[name = tensor("op_1804_cast_fp16")]; + tensor x_195_cast_fp16 = real_div(x = sub_50_cast_fp16, y = var_1804_cast_fp16)[name = tensor("x_195_cast_fp16")]; + tensor var_1806_cast_fp16 = mul(x = x_195_cast_fp16, y = flow_net_res_blocks_2_in_ln_weight_to_fp16)[name = tensor("op_1806_cast_fp16")]; + tensor x_197_cast_fp16 = add(x = var_1806_cast_fp16, y = flow_net_res_blocks_2_in_ln_bias_to_fp16)[name = tensor("x_197_cast_fp16")]; + tensor var_1808_promoted_to_fp16 = const()[name = tensor("op_1808_promoted_to_fp16"), val = tensor(0x1p+0)]; + tensor var_1809_cast_fp16 = add(x = var_1792_cast_fp16_1, y = var_1808_promoted_to_fp16)[name = tensor("op_1809_cast_fp16")]; + tensor var_1810_cast_fp16 = mul(x = x_197_cast_fp16, y = var_1809_cast_fp16)[name = tensor("op_1810_cast_fp16")]; + tensor input_305_cast_fp16 = add(x = var_1810_cast_fp16, y = var_1792_cast_fp16_0)[name = tensor("input_305_cast_fp16")]; + tensor linear_117_cast_fp16 = linear(bias = flow_net_res_blocks_2_mlp_0_bias_to_fp16, weight = flow_net_res_blocks_2_mlp_0_weight_to_fp16, x = input_305_cast_fp16)[name = tensor("linear_117_cast_fp16")]; + tensor input_309_cast_fp16 = silu(x = linear_117_cast_fp16)[name = tensor("input_309_cast_fp16")]; + tensor linear_118_cast_fp16 = linear(bias = flow_net_res_blocks_2_mlp_2_bias_to_fp16, weight = flow_net_res_blocks_2_mlp_2_weight_to_fp16, x = input_309_cast_fp16)[name = tensor("linear_118_cast_fp16")]; + tensor var_1821_cast_fp16 = mul(x = var_1792_cast_fp16_2, y = linear_118_cast_fp16)[name = tensor("op_1821_cast_fp16")]; + tensor x_199_cast_fp16 = add(x = x_193_cast_fp16, y = var_1821_cast_fp16)[name = tensor("x_199_cast_fp16")]; + tensor linear_119_cast_fp16 = linear(bias = flow_net_res_blocks_3_adaLN_modulation_1_bias_to_fp16, weight = flow_net_res_blocks_3_adaLN_modulation_1_weight_to_fp16, x = input_287_cast_fp16)[name = tensor("linear_119_cast_fp16")]; + tensor var_1831_split_sizes_0 = const()[name = tensor("op_1831_split_sizes_0"), val = tensor([512, 512, 512])]; + tensor var_1831_axis_0 = const()[name = tensor("op_1831_axis_0"), val = tensor(-1)]; + tensor var_1831_cast_fp16_0, tensor var_1831_cast_fp16_1, tensor var_1831_cast_fp16_2 = split(axis = var_1831_axis_0, split_sizes = var_1831_split_sizes_0, x = linear_119_cast_fp16)[name = tensor("op_1831_cast_fp16")]; + tensor mean_63_axes_0 = const()[name = tensor("mean_63_axes_0"), val = tensor([-1])]; + tensor mean_63_keep_dims_0 = const()[name = tensor("mean_63_keep_dims_0"), val = tensor(true)]; + tensor mean_63_cast_fp16 = reduce_mean(axes = mean_63_axes_0, keep_dims = mean_63_keep_dims_0, x = x_199_cast_fp16)[name = tensor("mean_63_cast_fp16")]; + tensor sub_51_cast_fp16 = sub(x = x_199_cast_fp16, y = mean_63_cast_fp16)[name = tensor("sub_51_cast_fp16")]; + tensor square_41_cast_fp16 = square(x = sub_51_cast_fp16)[name = tensor("square_41_cast_fp16")]; + tensor reduce_mean_83_axes_0 = const()[name = tensor("reduce_mean_83_axes_0"), val = tensor([-1])]; + tensor reduce_mean_83_keep_dims_0 = const()[name = tensor("reduce_mean_83_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_83_cast_fp16 = reduce_mean(axes = reduce_mean_83_axes_0, keep_dims = reduce_mean_83_keep_dims_0, x = square_41_cast_fp16)[name = tensor("reduce_mean_83_cast_fp16")]; + tensor var_1841_to_fp16 = const()[name = tensor("op_1841_to_fp16"), val = tensor(0x1.1p-20)]; + tensor var_1842_cast_fp16 = add(x = reduce_mean_83_cast_fp16, y = var_1841_to_fp16)[name = tensor("op_1842_cast_fp16")]; + tensor var_1843_cast_fp16 = sqrt(x = var_1842_cast_fp16)[name = tensor("op_1843_cast_fp16")]; + tensor x_201_cast_fp16 = real_div(x = sub_51_cast_fp16, y = var_1843_cast_fp16)[name = tensor("x_201_cast_fp16")]; + tensor var_1845_cast_fp16 = mul(x = x_201_cast_fp16, y = flow_net_res_blocks_3_in_ln_weight_to_fp16)[name = tensor("op_1845_cast_fp16")]; + tensor x_203_cast_fp16 = add(x = var_1845_cast_fp16, y = flow_net_res_blocks_3_in_ln_bias_to_fp16)[name = tensor("x_203_cast_fp16")]; + tensor var_1847_promoted_to_fp16 = const()[name = tensor("op_1847_promoted_to_fp16"), val = tensor(0x1p+0)]; + tensor var_1848_cast_fp16 = add(x = var_1831_cast_fp16_1, y = var_1847_promoted_to_fp16)[name = tensor("op_1848_cast_fp16")]; + tensor var_1849_cast_fp16 = mul(x = x_203_cast_fp16, y = var_1848_cast_fp16)[name = tensor("op_1849_cast_fp16")]; + tensor input_313_cast_fp16 = add(x = var_1849_cast_fp16, y = var_1831_cast_fp16_0)[name = tensor("input_313_cast_fp16")]; + tensor linear_120_cast_fp16 = linear(bias = flow_net_res_blocks_3_mlp_0_bias_to_fp16, weight = flow_net_res_blocks_3_mlp_0_weight_to_fp16, x = input_313_cast_fp16)[name = tensor("linear_120_cast_fp16")]; + tensor input_317_cast_fp16 = silu(x = linear_120_cast_fp16)[name = tensor("input_317_cast_fp16")]; + tensor linear_121_cast_fp16 = linear(bias = flow_net_res_blocks_3_mlp_2_bias_to_fp16, weight = flow_net_res_blocks_3_mlp_2_weight_to_fp16, x = input_317_cast_fp16)[name = tensor("linear_121_cast_fp16")]; + tensor var_1860_cast_fp16 = mul(x = var_1831_cast_fp16_2, y = linear_121_cast_fp16)[name = tensor("op_1860_cast_fp16")]; + tensor x_205_cast_fp16 = add(x = x_199_cast_fp16, y = var_1860_cast_fp16)[name = tensor("x_205_cast_fp16")]; + tensor linear_122_cast_fp16 = linear(bias = flow_net_res_blocks_4_adaLN_modulation_1_bias_to_fp16, weight = flow_net_res_blocks_4_adaLN_modulation_1_weight_to_fp16, x = input_287_cast_fp16)[name = tensor("linear_122_cast_fp16")]; + tensor var_1870_split_sizes_0 = const()[name = tensor("op_1870_split_sizes_0"), val = tensor([512, 512, 512])]; + tensor var_1870_axis_0 = const()[name = tensor("op_1870_axis_0"), val = tensor(-1)]; + tensor var_1870_cast_fp16_0, tensor var_1870_cast_fp16_1, tensor var_1870_cast_fp16_2 = split(axis = var_1870_axis_0, split_sizes = var_1870_split_sizes_0, x = linear_122_cast_fp16)[name = tensor("op_1870_cast_fp16")]; + tensor mean_65_axes_0 = const()[name = tensor("mean_65_axes_0"), val = tensor([-1])]; + tensor mean_65_keep_dims_0 = const()[name = tensor("mean_65_keep_dims_0"), val = tensor(true)]; + tensor mean_65_cast_fp16 = reduce_mean(axes = mean_65_axes_0, keep_dims = mean_65_keep_dims_0, x = x_205_cast_fp16)[name = tensor("mean_65_cast_fp16")]; + tensor sub_52_cast_fp16 = sub(x = x_205_cast_fp16, y = mean_65_cast_fp16)[name = tensor("sub_52_cast_fp16")]; + tensor square_42_cast_fp16 = square(x = sub_52_cast_fp16)[name = tensor("square_42_cast_fp16")]; + tensor reduce_mean_85_axes_0 = const()[name = tensor("reduce_mean_85_axes_0"), val = tensor([-1])]; + tensor reduce_mean_85_keep_dims_0 = const()[name = tensor("reduce_mean_85_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_85_cast_fp16 = reduce_mean(axes = reduce_mean_85_axes_0, keep_dims = reduce_mean_85_keep_dims_0, x = square_42_cast_fp16)[name = tensor("reduce_mean_85_cast_fp16")]; + tensor var_1880_to_fp16 = const()[name = tensor("op_1880_to_fp16"), val = tensor(0x1.1p-20)]; + tensor var_1881_cast_fp16 = add(x = reduce_mean_85_cast_fp16, y = var_1880_to_fp16)[name = tensor("op_1881_cast_fp16")]; + tensor var_1882_cast_fp16 = sqrt(x = var_1881_cast_fp16)[name = tensor("op_1882_cast_fp16")]; + tensor x_207_cast_fp16 = real_div(x = sub_52_cast_fp16, y = var_1882_cast_fp16)[name = tensor("x_207_cast_fp16")]; + tensor var_1884_cast_fp16 = mul(x = x_207_cast_fp16, y = flow_net_res_blocks_4_in_ln_weight_to_fp16)[name = tensor("op_1884_cast_fp16")]; + tensor x_209_cast_fp16 = add(x = var_1884_cast_fp16, y = flow_net_res_blocks_4_in_ln_bias_to_fp16)[name = tensor("x_209_cast_fp16")]; + tensor var_1886_promoted_to_fp16 = const()[name = tensor("op_1886_promoted_to_fp16"), val = tensor(0x1p+0)]; + tensor var_1887_cast_fp16 = add(x = var_1870_cast_fp16_1, y = var_1886_promoted_to_fp16)[name = tensor("op_1887_cast_fp16")]; + tensor var_1888_cast_fp16 = mul(x = x_209_cast_fp16, y = var_1887_cast_fp16)[name = tensor("op_1888_cast_fp16")]; + tensor input_321_cast_fp16 = add(x = var_1888_cast_fp16, y = var_1870_cast_fp16_0)[name = tensor("input_321_cast_fp16")]; + tensor linear_123_cast_fp16 = linear(bias = flow_net_res_blocks_4_mlp_0_bias_to_fp16, weight = flow_net_res_blocks_4_mlp_0_weight_to_fp16, x = input_321_cast_fp16)[name = tensor("linear_123_cast_fp16")]; + tensor input_325_cast_fp16 = silu(x = linear_123_cast_fp16)[name = tensor("input_325_cast_fp16")]; + tensor linear_124_cast_fp16 = linear(bias = flow_net_res_blocks_4_mlp_2_bias_to_fp16, weight = flow_net_res_blocks_4_mlp_2_weight_to_fp16, x = input_325_cast_fp16)[name = tensor("linear_124_cast_fp16")]; + tensor var_1899_cast_fp16 = mul(x = var_1870_cast_fp16_2, y = linear_124_cast_fp16)[name = tensor("op_1899_cast_fp16")]; + tensor x_211_cast_fp16 = add(x = x_205_cast_fp16, y = var_1899_cast_fp16)[name = tensor("x_211_cast_fp16")]; + tensor linear_125_cast_fp16 = linear(bias = flow_net_res_blocks_5_adaLN_modulation_1_bias_to_fp16, weight = flow_net_res_blocks_5_adaLN_modulation_1_weight_to_fp16, x = input_287_cast_fp16)[name = tensor("linear_125_cast_fp16")]; + tensor var_1909_split_sizes_0 = const()[name = tensor("op_1909_split_sizes_0"), val = tensor([512, 512, 512])]; + tensor var_1909_axis_0 = const()[name = tensor("op_1909_axis_0"), val = tensor(-1)]; + tensor var_1909_cast_fp16_0, tensor var_1909_cast_fp16_1, tensor var_1909_cast_fp16_2 = split(axis = var_1909_axis_0, split_sizes = var_1909_split_sizes_0, x = linear_125_cast_fp16)[name = tensor("op_1909_cast_fp16")]; + tensor mean_67_axes_0 = const()[name = tensor("mean_67_axes_0"), val = tensor([-1])]; + tensor mean_67_keep_dims_0 = const()[name = tensor("mean_67_keep_dims_0"), val = tensor(true)]; + tensor mean_67_cast_fp16 = reduce_mean(axes = mean_67_axes_0, keep_dims = mean_67_keep_dims_0, x = x_211_cast_fp16)[name = tensor("mean_67_cast_fp16")]; + tensor sub_53_cast_fp16 = sub(x = x_211_cast_fp16, y = mean_67_cast_fp16)[name = tensor("sub_53_cast_fp16")]; + tensor square_43_cast_fp16 = square(x = sub_53_cast_fp16)[name = tensor("square_43_cast_fp16")]; + tensor reduce_mean_87_axes_0 = const()[name = tensor("reduce_mean_87_axes_0"), val = tensor([-1])]; + tensor reduce_mean_87_keep_dims_0 = const()[name = tensor("reduce_mean_87_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_87_cast_fp16 = reduce_mean(axes = reduce_mean_87_axes_0, keep_dims = reduce_mean_87_keep_dims_0, x = square_43_cast_fp16)[name = tensor("reduce_mean_87_cast_fp16")]; + tensor var_1919_to_fp16 = const()[name = tensor("op_1919_to_fp16"), val = tensor(0x1.1p-20)]; + tensor var_1920_cast_fp16 = add(x = reduce_mean_87_cast_fp16, y = var_1919_to_fp16)[name = tensor("op_1920_cast_fp16")]; + tensor var_1921_cast_fp16 = sqrt(x = var_1920_cast_fp16)[name = tensor("op_1921_cast_fp16")]; + tensor x_213_cast_fp16 = real_div(x = sub_53_cast_fp16, y = var_1921_cast_fp16)[name = tensor("x_213_cast_fp16")]; + tensor var_1923_cast_fp16 = mul(x = x_213_cast_fp16, y = flow_net_res_blocks_5_in_ln_weight_to_fp16)[name = tensor("op_1923_cast_fp16")]; + tensor x_215_cast_fp16 = add(x = var_1923_cast_fp16, y = flow_net_res_blocks_5_in_ln_bias_to_fp16)[name = tensor("x_215_cast_fp16")]; + tensor var_1925_promoted_to_fp16 = const()[name = tensor("op_1925_promoted_to_fp16"), val = tensor(0x1p+0)]; + tensor var_1926_cast_fp16 = add(x = var_1909_cast_fp16_1, y = var_1925_promoted_to_fp16)[name = tensor("op_1926_cast_fp16")]; + tensor var_1927_cast_fp16 = mul(x = x_215_cast_fp16, y = var_1926_cast_fp16)[name = tensor("op_1927_cast_fp16")]; + tensor input_329_cast_fp16 = add(x = var_1927_cast_fp16, y = var_1909_cast_fp16_0)[name = tensor("input_329_cast_fp16")]; + tensor linear_126_cast_fp16 = linear(bias = flow_net_res_blocks_5_mlp_0_bias_to_fp16, weight = flow_net_res_blocks_5_mlp_0_weight_to_fp16, x = input_329_cast_fp16)[name = tensor("linear_126_cast_fp16")]; + tensor input_333_cast_fp16 = silu(x = linear_126_cast_fp16)[name = tensor("input_333_cast_fp16")]; + tensor linear_127_cast_fp16 = linear(bias = flow_net_res_blocks_5_mlp_2_bias_to_fp16, weight = flow_net_res_blocks_5_mlp_2_weight_to_fp16, x = input_333_cast_fp16)[name = tensor("linear_127_cast_fp16")]; + tensor var_1938_cast_fp16 = mul(x = var_1909_cast_fp16_2, y = linear_127_cast_fp16)[name = tensor("op_1938_cast_fp16")]; + tensor x_217_cast_fp16 = add(x = x_211_cast_fp16, y = var_1938_cast_fp16)[name = tensor("x_217_cast_fp16")]; + tensor linear_128_cast_fp16 = linear(bias = flow_net_final_layer_adaLN_modulation_1_bias_to_fp16, weight = flow_net_final_layer_adaLN_modulation_1_weight_to_fp16, x = input_287_cast_fp16)[name = tensor("linear_128_cast_fp16")]; + tensor var_1947_split_sizes_0 = const()[name = tensor("op_1947_split_sizes_0"), val = tensor([512, 512])]; + tensor var_1947_axis_0 = const()[name = tensor("op_1947_axis_0"), val = tensor(-1)]; + tensor var_1947_cast_fp16_0, tensor var_1947_cast_fp16_1 = split(axis = var_1947_axis_0, split_sizes = var_1947_split_sizes_0, x = linear_128_cast_fp16)[name = tensor("op_1947_cast_fp16")]; + tensor mean_69_axes_0 = const()[name = tensor("mean_69_axes_0"), val = tensor([-1])]; + tensor mean_69_keep_dims_0 = const()[name = tensor("mean_69_keep_dims_0"), val = tensor(true)]; + tensor mean_69_cast_fp16 = reduce_mean(axes = mean_69_axes_0, keep_dims = mean_69_keep_dims_0, x = x_217_cast_fp16)[name = tensor("mean_69_cast_fp16")]; + tensor sub_54_cast_fp16 = sub(x = x_217_cast_fp16, y = mean_69_cast_fp16)[name = tensor("sub_54_cast_fp16")]; + tensor square_44_cast_fp16 = square(x = sub_54_cast_fp16)[name = tensor("square_44_cast_fp16")]; + tensor reduce_mean_89_axes_0 = const()[name = tensor("reduce_mean_89_axes_0"), val = tensor([-1])]; + tensor reduce_mean_89_keep_dims_0 = const()[name = tensor("reduce_mean_89_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_89_cast_fp16 = reduce_mean(axes = reduce_mean_89_axes_0, keep_dims = reduce_mean_89_keep_dims_0, x = square_44_cast_fp16)[name = tensor("reduce_mean_89_cast_fp16")]; + tensor var_1954_to_fp16 = const()[name = tensor("op_1954_to_fp16"), val = tensor(0x1.1p-20)]; + tensor var_1955_cast_fp16 = add(x = reduce_mean_89_cast_fp16, y = var_1954_to_fp16)[name = tensor("op_1955_cast_fp16")]; + tensor var_1956_cast_fp16 = sqrt(x = var_1955_cast_fp16)[name = tensor("op_1956_cast_fp16")]; + tensor x_219_cast_fp16 = real_div(x = sub_54_cast_fp16, y = var_1956_cast_fp16)[name = tensor("x_219_cast_fp16")]; + tensor var_1958_promoted_to_fp16 = const()[name = tensor("op_1958_promoted_to_fp16"), val = tensor(0x1p+0)]; + tensor var_1959_cast_fp16 = add(x = var_1947_cast_fp16_1, y = var_1958_promoted_to_fp16)[name = tensor("op_1959_cast_fp16")]; + tensor var_1960_cast_fp16 = mul(x = x_219_cast_fp16, y = var_1959_cast_fp16)[name = tensor("op_1960_cast_fp16")]; + tensor input_337_cast_fp16 = add(x = var_1960_cast_fp16, y = var_1947_cast_fp16_0)[name = tensor("input_337_cast_fp16")]; + tensor linear_129_cast_fp16 = linear(bias = flow_net_final_layer_linear_bias_to_fp16, weight = flow_net_final_layer_linear_weight_to_fp16, x = input_337_cast_fp16)[name = tensor("linear_129_cast_fp16")]; + tensor var_1971_to_fp16 = const()[name = tensor("op_1971_to_fp16"), val = tensor(0x1p-3)]; + tensor var_1972_cast_fp16 = mul(x = linear_129_cast_fp16, y = var_1971_to_fp16)[name = tensor("op_1972_cast_fp16")]; + tensor input_339_cast_fp16 = add(x = input_271_cast_fp16, y = var_1972_cast_fp16)[name = tensor("input_339_cast_fp16")]; + tensor linear_130_cast_fp16 = linear(bias = flow_net_input_proj_bias_to_fp16, weight = flow_net_input_proj_weight_to_fp16, x = input_339_cast_fp16)[name = tensor("linear_130_cast_fp16")]; + tensor input_343_to_fp16 = const()[name = tensor("input_343_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(19006080)))]; + tensor input_345_cast_fp16 = silu(x = input_343_to_fp16)[name = tensor("input_345_cast_fp16")]; + tensor linear_132_cast_fp16 = linear(bias = flow_net_time_embed_0_mlp_2_bias_to_fp16, weight = flow_net_time_embed_0_mlp_2_weight_to_fp16, x = input_345_cast_fp16)[name = tensor("linear_132_cast_fp16")]; + tensor reduce_mean_90_axes_0 = const()[name = tensor("reduce_mean_90_axes_0"), val = tensor([-1])]; + tensor reduce_mean_90_keep_dims_0 = const()[name = tensor("reduce_mean_90_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_90_cast_fp16 = reduce_mean(axes = reduce_mean_90_axes_0, keep_dims = reduce_mean_90_keep_dims_0, x = linear_132_cast_fp16)[name = tensor("reduce_mean_90_cast_fp16")]; + tensor sub_55_cast_fp16 = sub(x = linear_132_cast_fp16, y = reduce_mean_90_cast_fp16)[name = tensor("sub_55_cast_fp16")]; + tensor square_45_cast_fp16 = square(x = sub_55_cast_fp16)[name = tensor("square_45_cast_fp16")]; + tensor reduce_mean_91_axes_0 = const()[name = tensor("reduce_mean_91_axes_0"), val = tensor([-1])]; + tensor reduce_mean_91_keep_dims_0 = const()[name = tensor("reduce_mean_91_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_91_cast_fp16 = reduce_mean(axes = reduce_mean_91_axes_0, keep_dims = reduce_mean_91_keep_dims_0, x = square_45_cast_fp16)[name = tensor("reduce_mean_91_cast_fp16")]; + tensor real_div_10_to_fp16 = const()[name = tensor("real_div_10_to_fp16"), val = tensor(0x1.008p+0)]; + tensor mul_10_cast_fp16 = mul(x = reduce_mean_91_cast_fp16, y = real_div_10_to_fp16)[name = tensor("mul_10_cast_fp16")]; + tensor var_2038_to_fp16 = const()[name = tensor("op_2038_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_91_cast_fp16 = add(x = mul_10_cast_fp16, y = var_2038_to_fp16)[name = tensor("var_91_cast_fp16")]; + tensor var_2041_epsilon_0 = const()[name = tensor("op_2041_epsilon_0"), val = tensor(0x1.197998p-40)]; + tensor var_2041_cast_fp16 = rsqrt(epsilon = var_2041_epsilon_0, x = var_91_cast_fp16)[name = tensor("op_2041_cast_fp16")]; + tensor var_2042_cast_fp16 = mul(x = const_3_to_fp16, y = var_2041_cast_fp16)[name = tensor("op_2042_cast_fp16")]; + tensor var_2043_cast_fp16 = mul(x = linear_132_cast_fp16, y = var_2042_cast_fp16)[name = tensor("op_2043_cast_fp16")]; + tensor input_349_to_fp16 = const()[name = tensor("input_349_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(19007168)))]; + tensor input_351_cast_fp16 = silu(x = input_349_to_fp16)[name = tensor("input_351_cast_fp16")]; + tensor linear_134_cast_fp16 = linear(bias = flow_net_time_embed_1_mlp_2_bias_to_fp16, weight = flow_net_time_embed_1_mlp_2_weight_to_fp16, x = input_351_cast_fp16)[name = tensor("linear_134_cast_fp16")]; + tensor reduce_mean_92_axes_0 = const()[name = tensor("reduce_mean_92_axes_0"), val = tensor([-1])]; + tensor reduce_mean_92_keep_dims_0 = const()[name = tensor("reduce_mean_92_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_92_cast_fp16 = reduce_mean(axes = reduce_mean_92_axes_0, keep_dims = reduce_mean_92_keep_dims_0, x = linear_134_cast_fp16)[name = tensor("reduce_mean_92_cast_fp16")]; + tensor sub_57_cast_fp16 = sub(x = linear_134_cast_fp16, y = reduce_mean_92_cast_fp16)[name = tensor("sub_57_cast_fp16")]; + tensor square_46_cast_fp16 = square(x = sub_57_cast_fp16)[name = tensor("square_46_cast_fp16")]; + tensor reduce_mean_93_axes_0 = const()[name = tensor("reduce_mean_93_axes_0"), val = tensor([-1])]; + tensor reduce_mean_93_keep_dims_0 = const()[name = tensor("reduce_mean_93_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_93_cast_fp16 = reduce_mean(axes = reduce_mean_93_axes_0, keep_dims = reduce_mean_93_keep_dims_0, x = square_46_cast_fp16)[name = tensor("reduce_mean_93_cast_fp16")]; + tensor real_div_11_to_fp16 = const()[name = tensor("real_div_11_to_fp16"), val = tensor(0x1.008p+0)]; + tensor mul_11_cast_fp16 = mul(x = reduce_mean_93_cast_fp16, y = real_div_11_to_fp16)[name = tensor("mul_11_cast_fp16")]; + tensor var_2075_to_fp16 = const()[name = tensor("op_2075_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_93_cast_fp16 = add(x = mul_11_cast_fp16, y = var_2075_to_fp16)[name = tensor("var_93_cast_fp16")]; + tensor var_2078_epsilon_0 = const()[name = tensor("op_2078_epsilon_0"), val = tensor(0x1.197998p-40)]; + tensor var_2078_cast_fp16 = rsqrt(epsilon = var_2078_epsilon_0, x = var_93_cast_fp16)[name = tensor("op_2078_cast_fp16")]; + tensor var_2079_cast_fp16 = mul(x = const_5_to_fp16, y = var_2078_cast_fp16)[name = tensor("op_2079_cast_fp16")]; + tensor var_2080_cast_fp16 = mul(x = linear_134_cast_fp16, y = var_2079_cast_fp16)[name = tensor("op_2080_cast_fp16")]; + tensor var_2092_cast_fp16 = add(x = var_2043_cast_fp16, y = var_2080_cast_fp16)[name = tensor("op_2092_cast_fp16")]; + tensor _inversed_t_combined_11_y_0_to_fp16 = const()[name = tensor("_inversed_t_combined_11_y_0_to_fp16"), val = tensor(0x1p-1)]; + tensor _inversed_t_combined_11_cast_fp16 = mul(x = var_2092_cast_fp16, y = _inversed_t_combined_11_y_0_to_fp16)[name = tensor("_inversed_t_combined_11_cast_fp16")]; + tensor input_353_cast_fp16 = add(x = _inversed_t_combined_11_cast_fp16, y = linear_5_cast_fp16)[name = tensor("input_353_cast_fp16")]; + tensor input_355_cast_fp16 = silu(x = input_353_cast_fp16)[name = tensor("input_355_cast_fp16")]; + tensor linear_136_cast_fp16 = linear(bias = flow_net_res_blocks_0_adaLN_modulation_1_bias_to_fp16, weight = flow_net_res_blocks_0_adaLN_modulation_1_weight_to_fp16, x = input_355_cast_fp16)[name = tensor("linear_136_cast_fp16")]; + tensor var_2107_split_sizes_0 = const()[name = tensor("op_2107_split_sizes_0"), val = tensor([512, 512, 512])]; + tensor var_2107_axis_0 = const()[name = tensor("op_2107_axis_0"), val = tensor(-1)]; + tensor var_2107_cast_fp16_0, tensor var_2107_cast_fp16_1, tensor var_2107_cast_fp16_2 = split(axis = var_2107_axis_0, split_sizes = var_2107_split_sizes_0, x = linear_136_cast_fp16)[name = tensor("op_2107_cast_fp16")]; + tensor mean_71_axes_0 = const()[name = tensor("mean_71_axes_0"), val = tensor([-1])]; + tensor mean_71_keep_dims_0 = const()[name = tensor("mean_71_keep_dims_0"), val = tensor(true)]; + tensor mean_71_cast_fp16 = reduce_mean(axes = mean_71_axes_0, keep_dims = mean_71_keep_dims_0, x = linear_130_cast_fp16)[name = tensor("mean_71_cast_fp16")]; + tensor sub_59_cast_fp16 = sub(x = linear_130_cast_fp16, y = mean_71_cast_fp16)[name = tensor("sub_59_cast_fp16")]; + tensor square_47_cast_fp16 = square(x = sub_59_cast_fp16)[name = tensor("square_47_cast_fp16")]; + tensor reduce_mean_95_axes_0 = const()[name = tensor("reduce_mean_95_axes_0"), val = tensor([-1])]; + tensor reduce_mean_95_keep_dims_0 = const()[name = tensor("reduce_mean_95_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_95_cast_fp16 = reduce_mean(axes = reduce_mean_95_axes_0, keep_dims = reduce_mean_95_keep_dims_0, x = square_47_cast_fp16)[name = tensor("reduce_mean_95_cast_fp16")]; + tensor var_2117_to_fp16 = const()[name = tensor("op_2117_to_fp16"), val = tensor(0x1.1p-20)]; + tensor var_2118_cast_fp16 = add(x = reduce_mean_95_cast_fp16, y = var_2117_to_fp16)[name = tensor("op_2118_cast_fp16")]; + tensor var_2119_cast_fp16 = sqrt(x = var_2118_cast_fp16)[name = tensor("op_2119_cast_fp16")]; + tensor x_227_cast_fp16 = real_div(x = sub_59_cast_fp16, y = var_2119_cast_fp16)[name = tensor("x_227_cast_fp16")]; + tensor var_2121_cast_fp16 = mul(x = x_227_cast_fp16, y = flow_net_res_blocks_0_in_ln_weight_to_fp16)[name = tensor("op_2121_cast_fp16")]; + tensor x_229_cast_fp16 = add(x = var_2121_cast_fp16, y = flow_net_res_blocks_0_in_ln_bias_to_fp16)[name = tensor("x_229_cast_fp16")]; + tensor var_2123_promoted_to_fp16 = const()[name = tensor("op_2123_promoted_to_fp16"), val = tensor(0x1p+0)]; + tensor var_2124_cast_fp16 = add(x = var_2107_cast_fp16_1, y = var_2123_promoted_to_fp16)[name = tensor("op_2124_cast_fp16")]; + tensor var_2125_cast_fp16 = mul(x = x_229_cast_fp16, y = var_2124_cast_fp16)[name = tensor("op_2125_cast_fp16")]; + tensor input_357_cast_fp16 = add(x = var_2125_cast_fp16, y = var_2107_cast_fp16_0)[name = tensor("input_357_cast_fp16")]; + tensor linear_137_cast_fp16 = linear(bias = flow_net_res_blocks_0_mlp_0_bias_to_fp16, weight = flow_net_res_blocks_0_mlp_0_weight_to_fp16, x = input_357_cast_fp16)[name = tensor("linear_137_cast_fp16")]; + tensor input_361_cast_fp16 = silu(x = linear_137_cast_fp16)[name = tensor("input_361_cast_fp16")]; + tensor linear_138_cast_fp16 = linear(bias = flow_net_res_blocks_0_mlp_2_bias_to_fp16, weight = flow_net_res_blocks_0_mlp_2_weight_to_fp16, x = input_361_cast_fp16)[name = tensor("linear_138_cast_fp16")]; + tensor var_2136_cast_fp16 = mul(x = var_2107_cast_fp16_2, y = linear_138_cast_fp16)[name = tensor("op_2136_cast_fp16")]; + tensor x_231_cast_fp16 = add(x = linear_130_cast_fp16, y = var_2136_cast_fp16)[name = tensor("x_231_cast_fp16")]; + tensor linear_139_cast_fp16 = linear(bias = flow_net_res_blocks_1_adaLN_modulation_1_bias_to_fp16, weight = flow_net_res_blocks_1_adaLN_modulation_1_weight_to_fp16, x = input_355_cast_fp16)[name = tensor("linear_139_cast_fp16")]; + tensor var_2146_split_sizes_0 = const()[name = tensor("op_2146_split_sizes_0"), val = tensor([512, 512, 512])]; + tensor var_2146_axis_0 = const()[name = tensor("op_2146_axis_0"), val = tensor(-1)]; + tensor var_2146_cast_fp16_0, tensor var_2146_cast_fp16_1, tensor var_2146_cast_fp16_2 = split(axis = var_2146_axis_0, split_sizes = var_2146_split_sizes_0, x = linear_139_cast_fp16)[name = tensor("op_2146_cast_fp16")]; + tensor mean_73_axes_0 = const()[name = tensor("mean_73_axes_0"), val = tensor([-1])]; + tensor mean_73_keep_dims_0 = const()[name = tensor("mean_73_keep_dims_0"), val = tensor(true)]; + tensor mean_73_cast_fp16 = reduce_mean(axes = mean_73_axes_0, keep_dims = mean_73_keep_dims_0, x = x_231_cast_fp16)[name = tensor("mean_73_cast_fp16")]; + tensor sub_60_cast_fp16 = sub(x = x_231_cast_fp16, y = mean_73_cast_fp16)[name = tensor("sub_60_cast_fp16")]; + tensor square_48_cast_fp16 = square(x = sub_60_cast_fp16)[name = tensor("square_48_cast_fp16")]; + tensor reduce_mean_97_axes_0 = const()[name = tensor("reduce_mean_97_axes_0"), val = tensor([-1])]; + tensor reduce_mean_97_keep_dims_0 = const()[name = tensor("reduce_mean_97_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_97_cast_fp16 = reduce_mean(axes = reduce_mean_97_axes_0, keep_dims = reduce_mean_97_keep_dims_0, x = square_48_cast_fp16)[name = tensor("reduce_mean_97_cast_fp16")]; + tensor var_2156_to_fp16 = const()[name = tensor("op_2156_to_fp16"), val = tensor(0x1.1p-20)]; + tensor var_2157_cast_fp16 = add(x = reduce_mean_97_cast_fp16, y = var_2156_to_fp16)[name = tensor("op_2157_cast_fp16")]; + tensor var_2158_cast_fp16 = sqrt(x = var_2157_cast_fp16)[name = tensor("op_2158_cast_fp16")]; + tensor x_233_cast_fp16 = real_div(x = sub_60_cast_fp16, y = var_2158_cast_fp16)[name = tensor("x_233_cast_fp16")]; + tensor var_2160_cast_fp16 = mul(x = x_233_cast_fp16, y = flow_net_res_blocks_1_in_ln_weight_to_fp16)[name = tensor("op_2160_cast_fp16")]; + tensor x_235_cast_fp16 = add(x = var_2160_cast_fp16, y = flow_net_res_blocks_1_in_ln_bias_to_fp16)[name = tensor("x_235_cast_fp16")]; + tensor var_2162_promoted_to_fp16 = const()[name = tensor("op_2162_promoted_to_fp16"), val = tensor(0x1p+0)]; + tensor var_2163_cast_fp16 = add(x = var_2146_cast_fp16_1, y = var_2162_promoted_to_fp16)[name = tensor("op_2163_cast_fp16")]; + tensor var_2164_cast_fp16 = mul(x = x_235_cast_fp16, y = var_2163_cast_fp16)[name = tensor("op_2164_cast_fp16")]; + tensor input_365_cast_fp16 = add(x = var_2164_cast_fp16, y = var_2146_cast_fp16_0)[name = tensor("input_365_cast_fp16")]; + tensor linear_140_cast_fp16 = linear(bias = flow_net_res_blocks_1_mlp_0_bias_to_fp16, weight = flow_net_res_blocks_1_mlp_0_weight_to_fp16, x = input_365_cast_fp16)[name = tensor("linear_140_cast_fp16")]; + tensor input_369_cast_fp16 = silu(x = linear_140_cast_fp16)[name = tensor("input_369_cast_fp16")]; + tensor linear_141_cast_fp16 = linear(bias = flow_net_res_blocks_1_mlp_2_bias_to_fp16, weight = flow_net_res_blocks_1_mlp_2_weight_to_fp16, x = input_369_cast_fp16)[name = tensor("linear_141_cast_fp16")]; + tensor var_2175_cast_fp16 = mul(x = var_2146_cast_fp16_2, y = linear_141_cast_fp16)[name = tensor("op_2175_cast_fp16")]; + tensor x_237_cast_fp16 = add(x = x_231_cast_fp16, y = var_2175_cast_fp16)[name = tensor("x_237_cast_fp16")]; + tensor linear_142_cast_fp16 = linear(bias = flow_net_res_blocks_2_adaLN_modulation_1_bias_to_fp16, weight = flow_net_res_blocks_2_adaLN_modulation_1_weight_to_fp16, x = input_355_cast_fp16)[name = tensor("linear_142_cast_fp16")]; + tensor var_2185_split_sizes_0 = const()[name = tensor("op_2185_split_sizes_0"), val = tensor([512, 512, 512])]; + tensor var_2185_axis_0 = const()[name = tensor("op_2185_axis_0"), val = tensor(-1)]; + tensor var_2185_cast_fp16_0, tensor var_2185_cast_fp16_1, tensor var_2185_cast_fp16_2 = split(axis = var_2185_axis_0, split_sizes = var_2185_split_sizes_0, x = linear_142_cast_fp16)[name = tensor("op_2185_cast_fp16")]; + tensor mean_75_axes_0 = const()[name = tensor("mean_75_axes_0"), val = tensor([-1])]; + tensor mean_75_keep_dims_0 = const()[name = tensor("mean_75_keep_dims_0"), val = tensor(true)]; + tensor mean_75_cast_fp16 = reduce_mean(axes = mean_75_axes_0, keep_dims = mean_75_keep_dims_0, x = x_237_cast_fp16)[name = tensor("mean_75_cast_fp16")]; + tensor sub_61_cast_fp16 = sub(x = x_237_cast_fp16, y = mean_75_cast_fp16)[name = tensor("sub_61_cast_fp16")]; + tensor square_49_cast_fp16 = square(x = sub_61_cast_fp16)[name = tensor("square_49_cast_fp16")]; + tensor reduce_mean_99_axes_0 = const()[name = tensor("reduce_mean_99_axes_0"), val = tensor([-1])]; + tensor reduce_mean_99_keep_dims_0 = const()[name = tensor("reduce_mean_99_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_99_cast_fp16 = reduce_mean(axes = reduce_mean_99_axes_0, keep_dims = reduce_mean_99_keep_dims_0, x = square_49_cast_fp16)[name = tensor("reduce_mean_99_cast_fp16")]; + tensor var_2195_to_fp16 = const()[name = tensor("op_2195_to_fp16"), val = tensor(0x1.1p-20)]; + tensor var_2196_cast_fp16 = add(x = reduce_mean_99_cast_fp16, y = var_2195_to_fp16)[name = tensor("op_2196_cast_fp16")]; + tensor var_2197_cast_fp16 = sqrt(x = var_2196_cast_fp16)[name = tensor("op_2197_cast_fp16")]; + tensor x_239_cast_fp16 = real_div(x = sub_61_cast_fp16, y = var_2197_cast_fp16)[name = tensor("x_239_cast_fp16")]; + tensor var_2199_cast_fp16 = mul(x = x_239_cast_fp16, y = flow_net_res_blocks_2_in_ln_weight_to_fp16)[name = tensor("op_2199_cast_fp16")]; + tensor x_241_cast_fp16 = add(x = var_2199_cast_fp16, y = flow_net_res_blocks_2_in_ln_bias_to_fp16)[name = tensor("x_241_cast_fp16")]; + tensor var_2201_promoted_to_fp16 = const()[name = tensor("op_2201_promoted_to_fp16"), val = tensor(0x1p+0)]; + tensor var_2202_cast_fp16 = add(x = var_2185_cast_fp16_1, y = var_2201_promoted_to_fp16)[name = tensor("op_2202_cast_fp16")]; + tensor var_2203_cast_fp16 = mul(x = x_241_cast_fp16, y = var_2202_cast_fp16)[name = tensor("op_2203_cast_fp16")]; + tensor input_373_cast_fp16 = add(x = var_2203_cast_fp16, y = var_2185_cast_fp16_0)[name = tensor("input_373_cast_fp16")]; + tensor linear_143_cast_fp16 = linear(bias = flow_net_res_blocks_2_mlp_0_bias_to_fp16, weight = flow_net_res_blocks_2_mlp_0_weight_to_fp16, x = input_373_cast_fp16)[name = tensor("linear_143_cast_fp16")]; + tensor input_377_cast_fp16 = silu(x = linear_143_cast_fp16)[name = tensor("input_377_cast_fp16")]; + tensor linear_144_cast_fp16 = linear(bias = flow_net_res_blocks_2_mlp_2_bias_to_fp16, weight = flow_net_res_blocks_2_mlp_2_weight_to_fp16, x = input_377_cast_fp16)[name = tensor("linear_144_cast_fp16")]; + tensor var_2214_cast_fp16 = mul(x = var_2185_cast_fp16_2, y = linear_144_cast_fp16)[name = tensor("op_2214_cast_fp16")]; + tensor x_243_cast_fp16 = add(x = x_237_cast_fp16, y = var_2214_cast_fp16)[name = tensor("x_243_cast_fp16")]; + tensor linear_145_cast_fp16 = linear(bias = flow_net_res_blocks_3_adaLN_modulation_1_bias_to_fp16, weight = flow_net_res_blocks_3_adaLN_modulation_1_weight_to_fp16, x = input_355_cast_fp16)[name = tensor("linear_145_cast_fp16")]; + tensor var_2224_split_sizes_0 = const()[name = tensor("op_2224_split_sizes_0"), val = tensor([512, 512, 512])]; + tensor var_2224_axis_0 = const()[name = tensor("op_2224_axis_0"), val = tensor(-1)]; + tensor var_2224_cast_fp16_0, tensor var_2224_cast_fp16_1, tensor var_2224_cast_fp16_2 = split(axis = var_2224_axis_0, split_sizes = var_2224_split_sizes_0, x = linear_145_cast_fp16)[name = tensor("op_2224_cast_fp16")]; + tensor mean_77_axes_0 = const()[name = tensor("mean_77_axes_0"), val = tensor([-1])]; + tensor mean_77_keep_dims_0 = const()[name = tensor("mean_77_keep_dims_0"), val = tensor(true)]; + tensor mean_77_cast_fp16 = reduce_mean(axes = mean_77_axes_0, keep_dims = mean_77_keep_dims_0, x = x_243_cast_fp16)[name = tensor("mean_77_cast_fp16")]; + tensor sub_62_cast_fp16 = sub(x = x_243_cast_fp16, y = mean_77_cast_fp16)[name = tensor("sub_62_cast_fp16")]; + tensor square_50_cast_fp16 = square(x = sub_62_cast_fp16)[name = tensor("square_50_cast_fp16")]; + tensor reduce_mean_101_axes_0 = const()[name = tensor("reduce_mean_101_axes_0"), val = tensor([-1])]; + tensor reduce_mean_101_keep_dims_0 = const()[name = tensor("reduce_mean_101_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_101_cast_fp16 = reduce_mean(axes = reduce_mean_101_axes_0, keep_dims = reduce_mean_101_keep_dims_0, x = square_50_cast_fp16)[name = tensor("reduce_mean_101_cast_fp16")]; + tensor var_2234_to_fp16 = const()[name = tensor("op_2234_to_fp16"), val = tensor(0x1.1p-20)]; + tensor var_2235_cast_fp16 = add(x = reduce_mean_101_cast_fp16, y = var_2234_to_fp16)[name = tensor("op_2235_cast_fp16")]; + tensor var_2236_cast_fp16 = sqrt(x = var_2235_cast_fp16)[name = tensor("op_2236_cast_fp16")]; + tensor x_245_cast_fp16 = real_div(x = sub_62_cast_fp16, y = var_2236_cast_fp16)[name = tensor("x_245_cast_fp16")]; + tensor var_2238_cast_fp16 = mul(x = x_245_cast_fp16, y = flow_net_res_blocks_3_in_ln_weight_to_fp16)[name = tensor("op_2238_cast_fp16")]; + tensor x_247_cast_fp16 = add(x = var_2238_cast_fp16, y = flow_net_res_blocks_3_in_ln_bias_to_fp16)[name = tensor("x_247_cast_fp16")]; + tensor var_2240_promoted_to_fp16 = const()[name = tensor("op_2240_promoted_to_fp16"), val = tensor(0x1p+0)]; + tensor var_2241_cast_fp16 = add(x = var_2224_cast_fp16_1, y = var_2240_promoted_to_fp16)[name = tensor("op_2241_cast_fp16")]; + tensor var_2242_cast_fp16 = mul(x = x_247_cast_fp16, y = var_2241_cast_fp16)[name = tensor("op_2242_cast_fp16")]; + tensor input_381_cast_fp16 = add(x = var_2242_cast_fp16, y = var_2224_cast_fp16_0)[name = tensor("input_381_cast_fp16")]; + tensor linear_146_cast_fp16 = linear(bias = flow_net_res_blocks_3_mlp_0_bias_to_fp16, weight = flow_net_res_blocks_3_mlp_0_weight_to_fp16, x = input_381_cast_fp16)[name = tensor("linear_146_cast_fp16")]; + tensor input_385_cast_fp16 = silu(x = linear_146_cast_fp16)[name = tensor("input_385_cast_fp16")]; + tensor linear_147_cast_fp16 = linear(bias = flow_net_res_blocks_3_mlp_2_bias_to_fp16, weight = flow_net_res_blocks_3_mlp_2_weight_to_fp16, x = input_385_cast_fp16)[name = tensor("linear_147_cast_fp16")]; + tensor var_2253_cast_fp16 = mul(x = var_2224_cast_fp16_2, y = linear_147_cast_fp16)[name = tensor("op_2253_cast_fp16")]; + tensor x_249_cast_fp16 = add(x = x_243_cast_fp16, y = var_2253_cast_fp16)[name = tensor("x_249_cast_fp16")]; + tensor linear_148_cast_fp16 = linear(bias = flow_net_res_blocks_4_adaLN_modulation_1_bias_to_fp16, weight = flow_net_res_blocks_4_adaLN_modulation_1_weight_to_fp16, x = input_355_cast_fp16)[name = tensor("linear_148_cast_fp16")]; + tensor var_2263_split_sizes_0 = const()[name = tensor("op_2263_split_sizes_0"), val = tensor([512, 512, 512])]; + tensor var_2263_axis_0 = const()[name = tensor("op_2263_axis_0"), val = tensor(-1)]; + tensor var_2263_cast_fp16_0, tensor var_2263_cast_fp16_1, tensor var_2263_cast_fp16_2 = split(axis = var_2263_axis_0, split_sizes = var_2263_split_sizes_0, x = linear_148_cast_fp16)[name = tensor("op_2263_cast_fp16")]; + tensor mean_79_axes_0 = const()[name = tensor("mean_79_axes_0"), val = tensor([-1])]; + tensor mean_79_keep_dims_0 = const()[name = tensor("mean_79_keep_dims_0"), val = tensor(true)]; + tensor mean_79_cast_fp16 = reduce_mean(axes = mean_79_axes_0, keep_dims = mean_79_keep_dims_0, x = x_249_cast_fp16)[name = tensor("mean_79_cast_fp16")]; + tensor sub_63_cast_fp16 = sub(x = x_249_cast_fp16, y = mean_79_cast_fp16)[name = tensor("sub_63_cast_fp16")]; + tensor square_51_cast_fp16 = square(x = sub_63_cast_fp16)[name = tensor("square_51_cast_fp16")]; + tensor reduce_mean_103_axes_0 = const()[name = tensor("reduce_mean_103_axes_0"), val = tensor([-1])]; + tensor reduce_mean_103_keep_dims_0 = const()[name = tensor("reduce_mean_103_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_103_cast_fp16 = reduce_mean(axes = reduce_mean_103_axes_0, keep_dims = reduce_mean_103_keep_dims_0, x = square_51_cast_fp16)[name = tensor("reduce_mean_103_cast_fp16")]; + tensor var_2273_to_fp16 = const()[name = tensor("op_2273_to_fp16"), val = tensor(0x1.1p-20)]; + tensor var_2274_cast_fp16 = add(x = reduce_mean_103_cast_fp16, y = var_2273_to_fp16)[name = tensor("op_2274_cast_fp16")]; + tensor var_2275_cast_fp16 = sqrt(x = var_2274_cast_fp16)[name = tensor("op_2275_cast_fp16")]; + tensor x_251_cast_fp16 = real_div(x = sub_63_cast_fp16, y = var_2275_cast_fp16)[name = tensor("x_251_cast_fp16")]; + tensor var_2277_cast_fp16 = mul(x = x_251_cast_fp16, y = flow_net_res_blocks_4_in_ln_weight_to_fp16)[name = tensor("op_2277_cast_fp16")]; + tensor x_253_cast_fp16 = add(x = var_2277_cast_fp16, y = flow_net_res_blocks_4_in_ln_bias_to_fp16)[name = tensor("x_253_cast_fp16")]; + tensor var_2279_promoted_to_fp16 = const()[name = tensor("op_2279_promoted_to_fp16"), val = tensor(0x1p+0)]; + tensor var_2280_cast_fp16 = add(x = var_2263_cast_fp16_1, y = var_2279_promoted_to_fp16)[name = tensor("op_2280_cast_fp16")]; + tensor var_2281_cast_fp16 = mul(x = x_253_cast_fp16, y = var_2280_cast_fp16)[name = tensor("op_2281_cast_fp16")]; + tensor input_389_cast_fp16 = add(x = var_2281_cast_fp16, y = var_2263_cast_fp16_0)[name = tensor("input_389_cast_fp16")]; + tensor linear_149_cast_fp16 = linear(bias = flow_net_res_blocks_4_mlp_0_bias_to_fp16, weight = flow_net_res_blocks_4_mlp_0_weight_to_fp16, x = input_389_cast_fp16)[name = tensor("linear_149_cast_fp16")]; + tensor input_393_cast_fp16 = silu(x = linear_149_cast_fp16)[name = tensor("input_393_cast_fp16")]; + tensor linear_150_cast_fp16 = linear(bias = flow_net_res_blocks_4_mlp_2_bias_to_fp16, weight = flow_net_res_blocks_4_mlp_2_weight_to_fp16, x = input_393_cast_fp16)[name = tensor("linear_150_cast_fp16")]; + tensor var_2292_cast_fp16 = mul(x = var_2263_cast_fp16_2, y = linear_150_cast_fp16)[name = tensor("op_2292_cast_fp16")]; + tensor x_255_cast_fp16 = add(x = x_249_cast_fp16, y = var_2292_cast_fp16)[name = tensor("x_255_cast_fp16")]; + tensor linear_151_cast_fp16 = linear(bias = flow_net_res_blocks_5_adaLN_modulation_1_bias_to_fp16, weight = flow_net_res_blocks_5_adaLN_modulation_1_weight_to_fp16, x = input_355_cast_fp16)[name = tensor("linear_151_cast_fp16")]; + tensor var_2302_split_sizes_0 = const()[name = tensor("op_2302_split_sizes_0"), val = tensor([512, 512, 512])]; + tensor var_2302_axis_0 = const()[name = tensor("op_2302_axis_0"), val = tensor(-1)]; + tensor var_2302_cast_fp16_0, tensor var_2302_cast_fp16_1, tensor var_2302_cast_fp16_2 = split(axis = var_2302_axis_0, split_sizes = var_2302_split_sizes_0, x = linear_151_cast_fp16)[name = tensor("op_2302_cast_fp16")]; + tensor mean_81_axes_0 = const()[name = tensor("mean_81_axes_0"), val = tensor([-1])]; + tensor mean_81_keep_dims_0 = const()[name = tensor("mean_81_keep_dims_0"), val = tensor(true)]; + tensor mean_81_cast_fp16 = reduce_mean(axes = mean_81_axes_0, keep_dims = mean_81_keep_dims_0, x = x_255_cast_fp16)[name = tensor("mean_81_cast_fp16")]; + tensor sub_64_cast_fp16 = sub(x = x_255_cast_fp16, y = mean_81_cast_fp16)[name = tensor("sub_64_cast_fp16")]; + tensor square_52_cast_fp16 = square(x = sub_64_cast_fp16)[name = tensor("square_52_cast_fp16")]; + tensor reduce_mean_105_axes_0 = const()[name = tensor("reduce_mean_105_axes_0"), val = tensor([-1])]; + tensor reduce_mean_105_keep_dims_0 = const()[name = tensor("reduce_mean_105_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_105_cast_fp16 = reduce_mean(axes = reduce_mean_105_axes_0, keep_dims = reduce_mean_105_keep_dims_0, x = square_52_cast_fp16)[name = tensor("reduce_mean_105_cast_fp16")]; + tensor var_2312_to_fp16 = const()[name = tensor("op_2312_to_fp16"), val = tensor(0x1.1p-20)]; + tensor var_2313_cast_fp16 = add(x = reduce_mean_105_cast_fp16, y = var_2312_to_fp16)[name = tensor("op_2313_cast_fp16")]; + tensor var_2314_cast_fp16 = sqrt(x = var_2313_cast_fp16)[name = tensor("op_2314_cast_fp16")]; + tensor x_257_cast_fp16 = real_div(x = sub_64_cast_fp16, y = var_2314_cast_fp16)[name = tensor("x_257_cast_fp16")]; + tensor var_2316_cast_fp16 = mul(x = x_257_cast_fp16, y = flow_net_res_blocks_5_in_ln_weight_to_fp16)[name = tensor("op_2316_cast_fp16")]; + tensor x_259_cast_fp16 = add(x = var_2316_cast_fp16, y = flow_net_res_blocks_5_in_ln_bias_to_fp16)[name = tensor("x_259_cast_fp16")]; + tensor var_2318_promoted_to_fp16 = const()[name = tensor("op_2318_promoted_to_fp16"), val = tensor(0x1p+0)]; + tensor var_2319_cast_fp16 = add(x = var_2302_cast_fp16_1, y = var_2318_promoted_to_fp16)[name = tensor("op_2319_cast_fp16")]; + tensor var_2320_cast_fp16 = mul(x = x_259_cast_fp16, y = var_2319_cast_fp16)[name = tensor("op_2320_cast_fp16")]; + tensor input_397_cast_fp16 = add(x = var_2320_cast_fp16, y = var_2302_cast_fp16_0)[name = tensor("input_397_cast_fp16")]; + tensor linear_152_cast_fp16 = linear(bias = flow_net_res_blocks_5_mlp_0_bias_to_fp16, weight = flow_net_res_blocks_5_mlp_0_weight_to_fp16, x = input_397_cast_fp16)[name = tensor("linear_152_cast_fp16")]; + tensor input_401_cast_fp16 = silu(x = linear_152_cast_fp16)[name = tensor("input_401_cast_fp16")]; + tensor linear_153_cast_fp16 = linear(bias = flow_net_res_blocks_5_mlp_2_bias_to_fp16, weight = flow_net_res_blocks_5_mlp_2_weight_to_fp16, x = input_401_cast_fp16)[name = tensor("linear_153_cast_fp16")]; + tensor var_2331_cast_fp16 = mul(x = var_2302_cast_fp16_2, y = linear_153_cast_fp16)[name = tensor("op_2331_cast_fp16")]; + tensor x_261_cast_fp16 = add(x = x_255_cast_fp16, y = var_2331_cast_fp16)[name = tensor("x_261_cast_fp16")]; + tensor linear_154_cast_fp16 = linear(bias = flow_net_final_layer_adaLN_modulation_1_bias_to_fp16, weight = flow_net_final_layer_adaLN_modulation_1_weight_to_fp16, x = input_355_cast_fp16)[name = tensor("linear_154_cast_fp16")]; + tensor var_2340_split_sizes_0 = const()[name = tensor("op_2340_split_sizes_0"), val = tensor([512, 512])]; + tensor var_2340_axis_0 = const()[name = tensor("op_2340_axis_0"), val = tensor(-1)]; + tensor var_2340_cast_fp16_0, tensor var_2340_cast_fp16_1 = split(axis = var_2340_axis_0, split_sizes = var_2340_split_sizes_0, x = linear_154_cast_fp16)[name = tensor("op_2340_cast_fp16")]; + tensor mean_83_axes_0 = const()[name = tensor("mean_83_axes_0"), val = tensor([-1])]; + tensor mean_83_keep_dims_0 = const()[name = tensor("mean_83_keep_dims_0"), val = tensor(true)]; + tensor mean_83_cast_fp16 = reduce_mean(axes = mean_83_axes_0, keep_dims = mean_83_keep_dims_0, x = x_261_cast_fp16)[name = tensor("mean_83_cast_fp16")]; + tensor sub_65_cast_fp16 = sub(x = x_261_cast_fp16, y = mean_83_cast_fp16)[name = tensor("sub_65_cast_fp16")]; + tensor square_53_cast_fp16 = square(x = sub_65_cast_fp16)[name = tensor("square_53_cast_fp16")]; + tensor reduce_mean_107_axes_0 = const()[name = tensor("reduce_mean_107_axes_0"), val = tensor([-1])]; + tensor reduce_mean_107_keep_dims_0 = const()[name = tensor("reduce_mean_107_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_107_cast_fp16 = reduce_mean(axes = reduce_mean_107_axes_0, keep_dims = reduce_mean_107_keep_dims_0, x = square_53_cast_fp16)[name = tensor("reduce_mean_107_cast_fp16")]; + tensor var_2347_to_fp16 = const()[name = tensor("op_2347_to_fp16"), val = tensor(0x1.1p-20)]; + tensor var_2348_cast_fp16 = add(x = reduce_mean_107_cast_fp16, y = var_2347_to_fp16)[name = tensor("op_2348_cast_fp16")]; + tensor var_2349_cast_fp16 = sqrt(x = var_2348_cast_fp16)[name = tensor("op_2349_cast_fp16")]; + tensor x_263_cast_fp16 = real_div(x = sub_65_cast_fp16, y = var_2349_cast_fp16)[name = tensor("x_263_cast_fp16")]; + tensor var_2351_promoted_to_fp16 = const()[name = tensor("op_2351_promoted_to_fp16"), val = tensor(0x1p+0)]; + tensor var_2352_cast_fp16 = add(x = var_2340_cast_fp16_1, y = var_2351_promoted_to_fp16)[name = tensor("op_2352_cast_fp16")]; + tensor var_2353_cast_fp16 = mul(x = x_263_cast_fp16, y = var_2352_cast_fp16)[name = tensor("op_2353_cast_fp16")]; + tensor input_405_cast_fp16 = add(x = var_2353_cast_fp16, y = var_2340_cast_fp16_0)[name = tensor("input_405_cast_fp16")]; + tensor linear_155_cast_fp16 = linear(bias = flow_net_final_layer_linear_bias_to_fp16, weight = flow_net_final_layer_linear_weight_to_fp16, x = input_405_cast_fp16)[name = tensor("linear_155_cast_fp16")]; + tensor var_2364_to_fp16 = const()[name = tensor("op_2364_to_fp16"), val = tensor(0x1p-3)]; + tensor var_2365_cast_fp16 = mul(x = linear_155_cast_fp16, y = var_2364_to_fp16)[name = tensor("op_2365_cast_fp16")]; + tensor input_407_cast_fp16 = add(x = input_339_cast_fp16, y = var_2365_cast_fp16)[name = tensor("input_407_cast_fp16")]; + tensor linear_156_cast_fp16 = linear(bias = flow_net_input_proj_bias_to_fp16, weight = flow_net_input_proj_weight_to_fp16, x = input_407_cast_fp16)[name = tensor("linear_156_cast_fp16")]; + tensor input_411_to_fp16 = const()[name = tensor("input_411_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(19008256)))]; + tensor input_413_cast_fp16 = silu(x = input_411_to_fp16)[name = tensor("input_413_cast_fp16")]; + tensor linear_158_cast_fp16 = linear(bias = flow_net_time_embed_0_mlp_2_bias_to_fp16, weight = flow_net_time_embed_0_mlp_2_weight_to_fp16, x = input_413_cast_fp16)[name = tensor("linear_158_cast_fp16")]; + tensor reduce_mean_108_axes_0 = const()[name = tensor("reduce_mean_108_axes_0"), val = tensor([-1])]; + tensor reduce_mean_108_keep_dims_0 = const()[name = tensor("reduce_mean_108_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_108_cast_fp16 = reduce_mean(axes = reduce_mean_108_axes_0, keep_dims = reduce_mean_108_keep_dims_0, x = linear_158_cast_fp16)[name = tensor("reduce_mean_108_cast_fp16")]; + tensor sub_66_cast_fp16 = sub(x = linear_158_cast_fp16, y = reduce_mean_108_cast_fp16)[name = tensor("sub_66_cast_fp16")]; + tensor square_54_cast_fp16 = square(x = sub_66_cast_fp16)[name = tensor("square_54_cast_fp16")]; + tensor reduce_mean_109_axes_0 = const()[name = tensor("reduce_mean_109_axes_0"), val = tensor([-1])]; + tensor reduce_mean_109_keep_dims_0 = const()[name = tensor("reduce_mean_109_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_109_cast_fp16 = reduce_mean(axes = reduce_mean_109_axes_0, keep_dims = reduce_mean_109_keep_dims_0, x = square_54_cast_fp16)[name = tensor("reduce_mean_109_cast_fp16")]; + tensor real_div_12_to_fp16 = const()[name = tensor("real_div_12_to_fp16"), val = tensor(0x1.008p+0)]; + tensor mul_12_cast_fp16 = mul(x = reduce_mean_109_cast_fp16, y = real_div_12_to_fp16)[name = tensor("mul_12_cast_fp16")]; + tensor var_2431_to_fp16 = const()[name = tensor("op_2431_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_109_cast_fp16 = add(x = mul_12_cast_fp16, y = var_2431_to_fp16)[name = tensor("var_109_cast_fp16")]; + tensor var_2434_epsilon_0 = const()[name = tensor("op_2434_epsilon_0"), val = tensor(0x1.197998p-40)]; + tensor var_2434_cast_fp16 = rsqrt(epsilon = var_2434_epsilon_0, x = var_109_cast_fp16)[name = tensor("op_2434_cast_fp16")]; + tensor var_2435_cast_fp16 = mul(x = const_3_to_fp16, y = var_2434_cast_fp16)[name = tensor("op_2435_cast_fp16")]; + tensor var_2436_cast_fp16 = mul(x = linear_158_cast_fp16, y = var_2435_cast_fp16)[name = tensor("op_2436_cast_fp16")]; + tensor input_417_to_fp16 = const()[name = tensor("input_417_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(19009344)))]; + tensor input_419_cast_fp16 = silu(x = input_417_to_fp16)[name = tensor("input_419_cast_fp16")]; + tensor linear_160_cast_fp16 = linear(bias = flow_net_time_embed_1_mlp_2_bias_to_fp16, weight = flow_net_time_embed_1_mlp_2_weight_to_fp16, x = input_419_cast_fp16)[name = tensor("linear_160_cast_fp16")]; + tensor reduce_mean_110_axes_0 = const()[name = tensor("reduce_mean_110_axes_0"), val = tensor([-1])]; + tensor reduce_mean_110_keep_dims_0 = const()[name = tensor("reduce_mean_110_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_110_cast_fp16 = reduce_mean(axes = reduce_mean_110_axes_0, keep_dims = reduce_mean_110_keep_dims_0, x = linear_160_cast_fp16)[name = tensor("reduce_mean_110_cast_fp16")]; + tensor sub_68_cast_fp16 = sub(x = linear_160_cast_fp16, y = reduce_mean_110_cast_fp16)[name = tensor("sub_68_cast_fp16")]; + tensor square_55_cast_fp16 = square(x = sub_68_cast_fp16)[name = tensor("square_55_cast_fp16")]; + tensor reduce_mean_111_axes_0 = const()[name = tensor("reduce_mean_111_axes_0"), val = tensor([-1])]; + tensor reduce_mean_111_keep_dims_0 = const()[name = tensor("reduce_mean_111_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_111_cast_fp16 = reduce_mean(axes = reduce_mean_111_axes_0, keep_dims = reduce_mean_111_keep_dims_0, x = square_55_cast_fp16)[name = tensor("reduce_mean_111_cast_fp16")]; + tensor real_div_13_to_fp16 = const()[name = tensor("real_div_13_to_fp16"), val = tensor(0x1.008p+0)]; + tensor mul_13_cast_fp16 = mul(x = reduce_mean_111_cast_fp16, y = real_div_13_to_fp16)[name = tensor("mul_13_cast_fp16")]; + tensor var_2468_to_fp16 = const()[name = tensor("op_2468_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_111_cast_fp16 = add(x = mul_13_cast_fp16, y = var_2468_to_fp16)[name = tensor("var_111_cast_fp16")]; + tensor var_2471_epsilon_0 = const()[name = tensor("op_2471_epsilon_0"), val = tensor(0x1.197998p-40)]; + tensor var_2471_cast_fp16 = rsqrt(epsilon = var_2471_epsilon_0, x = var_111_cast_fp16)[name = tensor("op_2471_cast_fp16")]; + tensor var_2472_cast_fp16 = mul(x = const_5_to_fp16, y = var_2471_cast_fp16)[name = tensor("op_2472_cast_fp16")]; + tensor var_2473_cast_fp16 = mul(x = linear_160_cast_fp16, y = var_2472_cast_fp16)[name = tensor("op_2473_cast_fp16")]; + tensor var_2485_cast_fp16 = add(x = var_2436_cast_fp16, y = var_2473_cast_fp16)[name = tensor("op_2485_cast_fp16")]; + tensor _inversed_t_combined_13_y_0_to_fp16 = const()[name = tensor("_inversed_t_combined_13_y_0_to_fp16"), val = tensor(0x1p-1)]; + tensor _inversed_t_combined_13_cast_fp16 = mul(x = var_2485_cast_fp16, y = _inversed_t_combined_13_y_0_to_fp16)[name = tensor("_inversed_t_combined_13_cast_fp16")]; + tensor input_421_cast_fp16 = add(x = _inversed_t_combined_13_cast_fp16, y = linear_5_cast_fp16)[name = tensor("input_421_cast_fp16")]; + tensor input_423_cast_fp16 = silu(x = input_421_cast_fp16)[name = tensor("input_423_cast_fp16")]; + tensor linear_162_cast_fp16 = linear(bias = flow_net_res_blocks_0_adaLN_modulation_1_bias_to_fp16, weight = flow_net_res_blocks_0_adaLN_modulation_1_weight_to_fp16, x = input_423_cast_fp16)[name = tensor("linear_162_cast_fp16")]; + tensor var_2500_split_sizes_0 = const()[name = tensor("op_2500_split_sizes_0"), val = tensor([512, 512, 512])]; + tensor var_2500_axis_0 = const()[name = tensor("op_2500_axis_0"), val = tensor(-1)]; + tensor var_2500_cast_fp16_0, tensor var_2500_cast_fp16_1, tensor var_2500_cast_fp16_2 = split(axis = var_2500_axis_0, split_sizes = var_2500_split_sizes_0, x = linear_162_cast_fp16)[name = tensor("op_2500_cast_fp16")]; + tensor mean_85_axes_0 = const()[name = tensor("mean_85_axes_0"), val = tensor([-1])]; + tensor mean_85_keep_dims_0 = const()[name = tensor("mean_85_keep_dims_0"), val = tensor(true)]; + tensor mean_85_cast_fp16 = reduce_mean(axes = mean_85_axes_0, keep_dims = mean_85_keep_dims_0, x = linear_156_cast_fp16)[name = tensor("mean_85_cast_fp16")]; + tensor sub_70_cast_fp16 = sub(x = linear_156_cast_fp16, y = mean_85_cast_fp16)[name = tensor("sub_70_cast_fp16")]; + tensor square_56_cast_fp16 = square(x = sub_70_cast_fp16)[name = tensor("square_56_cast_fp16")]; + tensor reduce_mean_113_axes_0 = const()[name = tensor("reduce_mean_113_axes_0"), val = tensor([-1])]; + tensor reduce_mean_113_keep_dims_0 = const()[name = tensor("reduce_mean_113_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_113_cast_fp16 = reduce_mean(axes = reduce_mean_113_axes_0, keep_dims = reduce_mean_113_keep_dims_0, x = square_56_cast_fp16)[name = tensor("reduce_mean_113_cast_fp16")]; + tensor var_2510_to_fp16 = const()[name = tensor("op_2510_to_fp16"), val = tensor(0x1.1p-20)]; + tensor var_2511_cast_fp16 = add(x = reduce_mean_113_cast_fp16, y = var_2510_to_fp16)[name = tensor("op_2511_cast_fp16")]; + tensor var_2512_cast_fp16 = sqrt(x = var_2511_cast_fp16)[name = tensor("op_2512_cast_fp16")]; + tensor x_271_cast_fp16 = real_div(x = sub_70_cast_fp16, y = var_2512_cast_fp16)[name = tensor("x_271_cast_fp16")]; + tensor var_2514_cast_fp16 = mul(x = x_271_cast_fp16, y = flow_net_res_blocks_0_in_ln_weight_to_fp16)[name = tensor("op_2514_cast_fp16")]; + tensor x_273_cast_fp16 = add(x = var_2514_cast_fp16, y = flow_net_res_blocks_0_in_ln_bias_to_fp16)[name = tensor("x_273_cast_fp16")]; + tensor var_2516_promoted_to_fp16 = const()[name = tensor("op_2516_promoted_to_fp16"), val = tensor(0x1p+0)]; + tensor var_2517_cast_fp16 = add(x = var_2500_cast_fp16_1, y = var_2516_promoted_to_fp16)[name = tensor("op_2517_cast_fp16")]; + tensor var_2518_cast_fp16 = mul(x = x_273_cast_fp16, y = var_2517_cast_fp16)[name = tensor("op_2518_cast_fp16")]; + tensor input_425_cast_fp16 = add(x = var_2518_cast_fp16, y = var_2500_cast_fp16_0)[name = tensor("input_425_cast_fp16")]; + tensor linear_163_cast_fp16 = linear(bias = flow_net_res_blocks_0_mlp_0_bias_to_fp16, weight = flow_net_res_blocks_0_mlp_0_weight_to_fp16, x = input_425_cast_fp16)[name = tensor("linear_163_cast_fp16")]; + tensor input_429_cast_fp16 = silu(x = linear_163_cast_fp16)[name = tensor("input_429_cast_fp16")]; + tensor linear_164_cast_fp16 = linear(bias = flow_net_res_blocks_0_mlp_2_bias_to_fp16, weight = flow_net_res_blocks_0_mlp_2_weight_to_fp16, x = input_429_cast_fp16)[name = tensor("linear_164_cast_fp16")]; + tensor var_2529_cast_fp16 = mul(x = var_2500_cast_fp16_2, y = linear_164_cast_fp16)[name = tensor("op_2529_cast_fp16")]; + tensor x_275_cast_fp16 = add(x = linear_156_cast_fp16, y = var_2529_cast_fp16)[name = tensor("x_275_cast_fp16")]; + tensor linear_165_cast_fp16 = linear(bias = flow_net_res_blocks_1_adaLN_modulation_1_bias_to_fp16, weight = flow_net_res_blocks_1_adaLN_modulation_1_weight_to_fp16, x = input_423_cast_fp16)[name = tensor("linear_165_cast_fp16")]; + tensor var_2539_split_sizes_0 = const()[name = tensor("op_2539_split_sizes_0"), val = tensor([512, 512, 512])]; + tensor var_2539_axis_0 = const()[name = tensor("op_2539_axis_0"), val = tensor(-1)]; + tensor var_2539_cast_fp16_0, tensor var_2539_cast_fp16_1, tensor var_2539_cast_fp16_2 = split(axis = var_2539_axis_0, split_sizes = var_2539_split_sizes_0, x = linear_165_cast_fp16)[name = tensor("op_2539_cast_fp16")]; + tensor mean_87_axes_0 = const()[name = tensor("mean_87_axes_0"), val = tensor([-1])]; + tensor mean_87_keep_dims_0 = const()[name = tensor("mean_87_keep_dims_0"), val = tensor(true)]; + tensor mean_87_cast_fp16 = reduce_mean(axes = mean_87_axes_0, keep_dims = mean_87_keep_dims_0, x = x_275_cast_fp16)[name = tensor("mean_87_cast_fp16")]; + tensor sub_71_cast_fp16 = sub(x = x_275_cast_fp16, y = mean_87_cast_fp16)[name = tensor("sub_71_cast_fp16")]; + tensor square_57_cast_fp16 = square(x = sub_71_cast_fp16)[name = tensor("square_57_cast_fp16")]; + tensor reduce_mean_115_axes_0 = const()[name = tensor("reduce_mean_115_axes_0"), val = tensor([-1])]; + tensor reduce_mean_115_keep_dims_0 = const()[name = tensor("reduce_mean_115_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_115_cast_fp16 = reduce_mean(axes = reduce_mean_115_axes_0, keep_dims = reduce_mean_115_keep_dims_0, x = square_57_cast_fp16)[name = tensor("reduce_mean_115_cast_fp16")]; + tensor var_2549_to_fp16 = const()[name = tensor("op_2549_to_fp16"), val = tensor(0x1.1p-20)]; + tensor var_2550_cast_fp16 = add(x = reduce_mean_115_cast_fp16, y = var_2549_to_fp16)[name = tensor("op_2550_cast_fp16")]; + tensor var_2551_cast_fp16 = sqrt(x = var_2550_cast_fp16)[name = tensor("op_2551_cast_fp16")]; + tensor x_277_cast_fp16 = real_div(x = sub_71_cast_fp16, y = var_2551_cast_fp16)[name = tensor("x_277_cast_fp16")]; + tensor var_2553_cast_fp16 = mul(x = x_277_cast_fp16, y = flow_net_res_blocks_1_in_ln_weight_to_fp16)[name = tensor("op_2553_cast_fp16")]; + tensor x_279_cast_fp16 = add(x = var_2553_cast_fp16, y = flow_net_res_blocks_1_in_ln_bias_to_fp16)[name = tensor("x_279_cast_fp16")]; + tensor var_2555_promoted_to_fp16 = const()[name = tensor("op_2555_promoted_to_fp16"), val = tensor(0x1p+0)]; + tensor var_2556_cast_fp16 = add(x = var_2539_cast_fp16_1, y = var_2555_promoted_to_fp16)[name = tensor("op_2556_cast_fp16")]; + tensor var_2557_cast_fp16 = mul(x = x_279_cast_fp16, y = var_2556_cast_fp16)[name = tensor("op_2557_cast_fp16")]; + tensor input_433_cast_fp16 = add(x = var_2557_cast_fp16, y = var_2539_cast_fp16_0)[name = tensor("input_433_cast_fp16")]; + tensor linear_166_cast_fp16 = linear(bias = flow_net_res_blocks_1_mlp_0_bias_to_fp16, weight = flow_net_res_blocks_1_mlp_0_weight_to_fp16, x = input_433_cast_fp16)[name = tensor("linear_166_cast_fp16")]; + tensor input_437_cast_fp16 = silu(x = linear_166_cast_fp16)[name = tensor("input_437_cast_fp16")]; + tensor linear_167_cast_fp16 = linear(bias = flow_net_res_blocks_1_mlp_2_bias_to_fp16, weight = flow_net_res_blocks_1_mlp_2_weight_to_fp16, x = input_437_cast_fp16)[name = tensor("linear_167_cast_fp16")]; + tensor var_2568_cast_fp16 = mul(x = var_2539_cast_fp16_2, y = linear_167_cast_fp16)[name = tensor("op_2568_cast_fp16")]; + tensor x_281_cast_fp16 = add(x = x_275_cast_fp16, y = var_2568_cast_fp16)[name = tensor("x_281_cast_fp16")]; + tensor linear_168_cast_fp16 = linear(bias = flow_net_res_blocks_2_adaLN_modulation_1_bias_to_fp16, weight = flow_net_res_blocks_2_adaLN_modulation_1_weight_to_fp16, x = input_423_cast_fp16)[name = tensor("linear_168_cast_fp16")]; + tensor var_2578_split_sizes_0 = const()[name = tensor("op_2578_split_sizes_0"), val = tensor([512, 512, 512])]; + tensor var_2578_axis_0 = const()[name = tensor("op_2578_axis_0"), val = tensor(-1)]; + tensor var_2578_cast_fp16_0, tensor var_2578_cast_fp16_1, tensor var_2578_cast_fp16_2 = split(axis = var_2578_axis_0, split_sizes = var_2578_split_sizes_0, x = linear_168_cast_fp16)[name = tensor("op_2578_cast_fp16")]; + tensor mean_89_axes_0 = const()[name = tensor("mean_89_axes_0"), val = tensor([-1])]; + tensor mean_89_keep_dims_0 = const()[name = tensor("mean_89_keep_dims_0"), val = tensor(true)]; + tensor mean_89_cast_fp16 = reduce_mean(axes = mean_89_axes_0, keep_dims = mean_89_keep_dims_0, x = x_281_cast_fp16)[name = tensor("mean_89_cast_fp16")]; + tensor sub_72_cast_fp16 = sub(x = x_281_cast_fp16, y = mean_89_cast_fp16)[name = tensor("sub_72_cast_fp16")]; + tensor square_58_cast_fp16 = square(x = sub_72_cast_fp16)[name = tensor("square_58_cast_fp16")]; + tensor reduce_mean_117_axes_0 = const()[name = tensor("reduce_mean_117_axes_0"), val = tensor([-1])]; + tensor reduce_mean_117_keep_dims_0 = const()[name = tensor("reduce_mean_117_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_117_cast_fp16 = reduce_mean(axes = reduce_mean_117_axes_0, keep_dims = reduce_mean_117_keep_dims_0, x = square_58_cast_fp16)[name = tensor("reduce_mean_117_cast_fp16")]; + tensor var_2588_to_fp16 = const()[name = tensor("op_2588_to_fp16"), val = tensor(0x1.1p-20)]; + tensor var_2589_cast_fp16 = add(x = reduce_mean_117_cast_fp16, y = var_2588_to_fp16)[name = tensor("op_2589_cast_fp16")]; + tensor var_2590_cast_fp16 = sqrt(x = var_2589_cast_fp16)[name = tensor("op_2590_cast_fp16")]; + tensor x_283_cast_fp16 = real_div(x = sub_72_cast_fp16, y = var_2590_cast_fp16)[name = tensor("x_283_cast_fp16")]; + tensor var_2592_cast_fp16 = mul(x = x_283_cast_fp16, y = flow_net_res_blocks_2_in_ln_weight_to_fp16)[name = tensor("op_2592_cast_fp16")]; + tensor x_285_cast_fp16 = add(x = var_2592_cast_fp16, y = flow_net_res_blocks_2_in_ln_bias_to_fp16)[name = tensor("x_285_cast_fp16")]; + tensor var_2594_promoted_to_fp16 = const()[name = tensor("op_2594_promoted_to_fp16"), val = tensor(0x1p+0)]; + tensor var_2595_cast_fp16 = add(x = var_2578_cast_fp16_1, y = var_2594_promoted_to_fp16)[name = tensor("op_2595_cast_fp16")]; + tensor var_2596_cast_fp16 = mul(x = x_285_cast_fp16, y = var_2595_cast_fp16)[name = tensor("op_2596_cast_fp16")]; + tensor input_441_cast_fp16 = add(x = var_2596_cast_fp16, y = var_2578_cast_fp16_0)[name = tensor("input_441_cast_fp16")]; + tensor linear_169_cast_fp16 = linear(bias = flow_net_res_blocks_2_mlp_0_bias_to_fp16, weight = flow_net_res_blocks_2_mlp_0_weight_to_fp16, x = input_441_cast_fp16)[name = tensor("linear_169_cast_fp16")]; + tensor input_445_cast_fp16 = silu(x = linear_169_cast_fp16)[name = tensor("input_445_cast_fp16")]; + tensor linear_170_cast_fp16 = linear(bias = flow_net_res_blocks_2_mlp_2_bias_to_fp16, weight = flow_net_res_blocks_2_mlp_2_weight_to_fp16, x = input_445_cast_fp16)[name = tensor("linear_170_cast_fp16")]; + tensor var_2607_cast_fp16 = mul(x = var_2578_cast_fp16_2, y = linear_170_cast_fp16)[name = tensor("op_2607_cast_fp16")]; + tensor x_287_cast_fp16 = add(x = x_281_cast_fp16, y = var_2607_cast_fp16)[name = tensor("x_287_cast_fp16")]; + tensor linear_171_cast_fp16 = linear(bias = flow_net_res_blocks_3_adaLN_modulation_1_bias_to_fp16, weight = flow_net_res_blocks_3_adaLN_modulation_1_weight_to_fp16, x = input_423_cast_fp16)[name = tensor("linear_171_cast_fp16")]; + tensor var_2617_split_sizes_0 = const()[name = tensor("op_2617_split_sizes_0"), val = tensor([512, 512, 512])]; + tensor var_2617_axis_0 = const()[name = tensor("op_2617_axis_0"), val = tensor(-1)]; + tensor var_2617_cast_fp16_0, tensor var_2617_cast_fp16_1, tensor var_2617_cast_fp16_2 = split(axis = var_2617_axis_0, split_sizes = var_2617_split_sizes_0, x = linear_171_cast_fp16)[name = tensor("op_2617_cast_fp16")]; + tensor mean_91_axes_0 = const()[name = tensor("mean_91_axes_0"), val = tensor([-1])]; + tensor mean_91_keep_dims_0 = const()[name = tensor("mean_91_keep_dims_0"), val = tensor(true)]; + tensor mean_91_cast_fp16 = reduce_mean(axes = mean_91_axes_0, keep_dims = mean_91_keep_dims_0, x = x_287_cast_fp16)[name = tensor("mean_91_cast_fp16")]; + tensor sub_73_cast_fp16 = sub(x = x_287_cast_fp16, y = mean_91_cast_fp16)[name = tensor("sub_73_cast_fp16")]; + tensor square_59_cast_fp16 = square(x = sub_73_cast_fp16)[name = tensor("square_59_cast_fp16")]; + tensor reduce_mean_119_axes_0 = const()[name = tensor("reduce_mean_119_axes_0"), val = tensor([-1])]; + tensor reduce_mean_119_keep_dims_0 = const()[name = tensor("reduce_mean_119_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_119_cast_fp16 = reduce_mean(axes = reduce_mean_119_axes_0, keep_dims = reduce_mean_119_keep_dims_0, x = square_59_cast_fp16)[name = tensor("reduce_mean_119_cast_fp16")]; + tensor var_2627_to_fp16 = const()[name = tensor("op_2627_to_fp16"), val = tensor(0x1.1p-20)]; + tensor var_2628_cast_fp16 = add(x = reduce_mean_119_cast_fp16, y = var_2627_to_fp16)[name = tensor("op_2628_cast_fp16")]; + tensor var_2629_cast_fp16 = sqrt(x = var_2628_cast_fp16)[name = tensor("op_2629_cast_fp16")]; + tensor x_289_cast_fp16 = real_div(x = sub_73_cast_fp16, y = var_2629_cast_fp16)[name = tensor("x_289_cast_fp16")]; + tensor var_2631_cast_fp16 = mul(x = x_289_cast_fp16, y = flow_net_res_blocks_3_in_ln_weight_to_fp16)[name = tensor("op_2631_cast_fp16")]; + tensor x_291_cast_fp16 = add(x = var_2631_cast_fp16, y = flow_net_res_blocks_3_in_ln_bias_to_fp16)[name = tensor("x_291_cast_fp16")]; + tensor var_2633_promoted_to_fp16 = const()[name = tensor("op_2633_promoted_to_fp16"), val = tensor(0x1p+0)]; + tensor var_2634_cast_fp16 = add(x = var_2617_cast_fp16_1, y = var_2633_promoted_to_fp16)[name = tensor("op_2634_cast_fp16")]; + tensor var_2635_cast_fp16 = mul(x = x_291_cast_fp16, y = var_2634_cast_fp16)[name = tensor("op_2635_cast_fp16")]; + tensor input_449_cast_fp16 = add(x = var_2635_cast_fp16, y = var_2617_cast_fp16_0)[name = tensor("input_449_cast_fp16")]; + tensor linear_172_cast_fp16 = linear(bias = flow_net_res_blocks_3_mlp_0_bias_to_fp16, weight = flow_net_res_blocks_3_mlp_0_weight_to_fp16, x = input_449_cast_fp16)[name = tensor("linear_172_cast_fp16")]; + tensor input_453_cast_fp16 = silu(x = linear_172_cast_fp16)[name = tensor("input_453_cast_fp16")]; + tensor linear_173_cast_fp16 = linear(bias = flow_net_res_blocks_3_mlp_2_bias_to_fp16, weight = flow_net_res_blocks_3_mlp_2_weight_to_fp16, x = input_453_cast_fp16)[name = tensor("linear_173_cast_fp16")]; + tensor var_2646_cast_fp16 = mul(x = var_2617_cast_fp16_2, y = linear_173_cast_fp16)[name = tensor("op_2646_cast_fp16")]; + tensor x_293_cast_fp16 = add(x = x_287_cast_fp16, y = var_2646_cast_fp16)[name = tensor("x_293_cast_fp16")]; + tensor linear_174_cast_fp16 = linear(bias = flow_net_res_blocks_4_adaLN_modulation_1_bias_to_fp16, weight = flow_net_res_blocks_4_adaLN_modulation_1_weight_to_fp16, x = input_423_cast_fp16)[name = tensor("linear_174_cast_fp16")]; + tensor var_2656_split_sizes_0 = const()[name = tensor("op_2656_split_sizes_0"), val = tensor([512, 512, 512])]; + tensor var_2656_axis_0 = const()[name = tensor("op_2656_axis_0"), val = tensor(-1)]; + tensor var_2656_cast_fp16_0, tensor var_2656_cast_fp16_1, tensor var_2656_cast_fp16_2 = split(axis = var_2656_axis_0, split_sizes = var_2656_split_sizes_0, x = linear_174_cast_fp16)[name = tensor("op_2656_cast_fp16")]; + tensor mean_93_axes_0 = const()[name = tensor("mean_93_axes_0"), val = tensor([-1])]; + tensor mean_93_keep_dims_0 = const()[name = tensor("mean_93_keep_dims_0"), val = tensor(true)]; + tensor mean_93_cast_fp16 = reduce_mean(axes = mean_93_axes_0, keep_dims = mean_93_keep_dims_0, x = x_293_cast_fp16)[name = tensor("mean_93_cast_fp16")]; + tensor sub_74_cast_fp16 = sub(x = x_293_cast_fp16, y = mean_93_cast_fp16)[name = tensor("sub_74_cast_fp16")]; + tensor square_60_cast_fp16 = square(x = sub_74_cast_fp16)[name = tensor("square_60_cast_fp16")]; + tensor reduce_mean_121_axes_0 = const()[name = tensor("reduce_mean_121_axes_0"), val = tensor([-1])]; + tensor reduce_mean_121_keep_dims_0 = const()[name = tensor("reduce_mean_121_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_121_cast_fp16 = reduce_mean(axes = reduce_mean_121_axes_0, keep_dims = reduce_mean_121_keep_dims_0, x = square_60_cast_fp16)[name = tensor("reduce_mean_121_cast_fp16")]; + tensor var_2666_to_fp16 = const()[name = tensor("op_2666_to_fp16"), val = tensor(0x1.1p-20)]; + tensor var_2667_cast_fp16 = add(x = reduce_mean_121_cast_fp16, y = var_2666_to_fp16)[name = tensor("op_2667_cast_fp16")]; + tensor var_2668_cast_fp16 = sqrt(x = var_2667_cast_fp16)[name = tensor("op_2668_cast_fp16")]; + tensor x_295_cast_fp16 = real_div(x = sub_74_cast_fp16, y = var_2668_cast_fp16)[name = tensor("x_295_cast_fp16")]; + tensor var_2670_cast_fp16 = mul(x = x_295_cast_fp16, y = flow_net_res_blocks_4_in_ln_weight_to_fp16)[name = tensor("op_2670_cast_fp16")]; + tensor x_297_cast_fp16 = add(x = var_2670_cast_fp16, y = flow_net_res_blocks_4_in_ln_bias_to_fp16)[name = tensor("x_297_cast_fp16")]; + tensor var_2672_promoted_to_fp16 = const()[name = tensor("op_2672_promoted_to_fp16"), val = tensor(0x1p+0)]; + tensor var_2673_cast_fp16 = add(x = var_2656_cast_fp16_1, y = var_2672_promoted_to_fp16)[name = tensor("op_2673_cast_fp16")]; + tensor var_2674_cast_fp16 = mul(x = x_297_cast_fp16, y = var_2673_cast_fp16)[name = tensor("op_2674_cast_fp16")]; + tensor input_457_cast_fp16 = add(x = var_2674_cast_fp16, y = var_2656_cast_fp16_0)[name = tensor("input_457_cast_fp16")]; + tensor linear_175_cast_fp16 = linear(bias = flow_net_res_blocks_4_mlp_0_bias_to_fp16, weight = flow_net_res_blocks_4_mlp_0_weight_to_fp16, x = input_457_cast_fp16)[name = tensor("linear_175_cast_fp16")]; + tensor input_461_cast_fp16 = silu(x = linear_175_cast_fp16)[name = tensor("input_461_cast_fp16")]; + tensor linear_176_cast_fp16 = linear(bias = flow_net_res_blocks_4_mlp_2_bias_to_fp16, weight = flow_net_res_blocks_4_mlp_2_weight_to_fp16, x = input_461_cast_fp16)[name = tensor("linear_176_cast_fp16")]; + tensor var_2685_cast_fp16 = mul(x = var_2656_cast_fp16_2, y = linear_176_cast_fp16)[name = tensor("op_2685_cast_fp16")]; + tensor x_299_cast_fp16 = add(x = x_293_cast_fp16, y = var_2685_cast_fp16)[name = tensor("x_299_cast_fp16")]; + tensor linear_177_cast_fp16 = linear(bias = flow_net_res_blocks_5_adaLN_modulation_1_bias_to_fp16, weight = flow_net_res_blocks_5_adaLN_modulation_1_weight_to_fp16, x = input_423_cast_fp16)[name = tensor("linear_177_cast_fp16")]; + tensor var_2695_split_sizes_0 = const()[name = tensor("op_2695_split_sizes_0"), val = tensor([512, 512, 512])]; + tensor var_2695_axis_0 = const()[name = tensor("op_2695_axis_0"), val = tensor(-1)]; + tensor var_2695_cast_fp16_0, tensor var_2695_cast_fp16_1, tensor var_2695_cast_fp16_2 = split(axis = var_2695_axis_0, split_sizes = var_2695_split_sizes_0, x = linear_177_cast_fp16)[name = tensor("op_2695_cast_fp16")]; + tensor mean_95_axes_0 = const()[name = tensor("mean_95_axes_0"), val = tensor([-1])]; + tensor mean_95_keep_dims_0 = const()[name = tensor("mean_95_keep_dims_0"), val = tensor(true)]; + tensor mean_95_cast_fp16 = reduce_mean(axes = mean_95_axes_0, keep_dims = mean_95_keep_dims_0, x = x_299_cast_fp16)[name = tensor("mean_95_cast_fp16")]; + tensor sub_75_cast_fp16 = sub(x = x_299_cast_fp16, y = mean_95_cast_fp16)[name = tensor("sub_75_cast_fp16")]; + tensor square_61_cast_fp16 = square(x = sub_75_cast_fp16)[name = tensor("square_61_cast_fp16")]; + tensor reduce_mean_123_axes_0 = const()[name = tensor("reduce_mean_123_axes_0"), val = tensor([-1])]; + tensor reduce_mean_123_keep_dims_0 = const()[name = tensor("reduce_mean_123_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_123_cast_fp16 = reduce_mean(axes = reduce_mean_123_axes_0, keep_dims = reduce_mean_123_keep_dims_0, x = square_61_cast_fp16)[name = tensor("reduce_mean_123_cast_fp16")]; + tensor var_2705_to_fp16 = const()[name = tensor("op_2705_to_fp16"), val = tensor(0x1.1p-20)]; + tensor var_2706_cast_fp16 = add(x = reduce_mean_123_cast_fp16, y = var_2705_to_fp16)[name = tensor("op_2706_cast_fp16")]; + tensor var_2707_cast_fp16 = sqrt(x = var_2706_cast_fp16)[name = tensor("op_2707_cast_fp16")]; + tensor x_301_cast_fp16 = real_div(x = sub_75_cast_fp16, y = var_2707_cast_fp16)[name = tensor("x_301_cast_fp16")]; + tensor var_2709_cast_fp16 = mul(x = x_301_cast_fp16, y = flow_net_res_blocks_5_in_ln_weight_to_fp16)[name = tensor("op_2709_cast_fp16")]; + tensor x_303_cast_fp16 = add(x = var_2709_cast_fp16, y = flow_net_res_blocks_5_in_ln_bias_to_fp16)[name = tensor("x_303_cast_fp16")]; + tensor var_2711_promoted_to_fp16 = const()[name = tensor("op_2711_promoted_to_fp16"), val = tensor(0x1p+0)]; + tensor var_2712_cast_fp16 = add(x = var_2695_cast_fp16_1, y = var_2711_promoted_to_fp16)[name = tensor("op_2712_cast_fp16")]; + tensor var_2713_cast_fp16 = mul(x = x_303_cast_fp16, y = var_2712_cast_fp16)[name = tensor("op_2713_cast_fp16")]; + tensor input_465_cast_fp16 = add(x = var_2713_cast_fp16, y = var_2695_cast_fp16_0)[name = tensor("input_465_cast_fp16")]; + tensor linear_178_cast_fp16 = linear(bias = flow_net_res_blocks_5_mlp_0_bias_to_fp16, weight = flow_net_res_blocks_5_mlp_0_weight_to_fp16, x = input_465_cast_fp16)[name = tensor("linear_178_cast_fp16")]; + tensor input_469_cast_fp16 = silu(x = linear_178_cast_fp16)[name = tensor("input_469_cast_fp16")]; + tensor linear_179_cast_fp16 = linear(bias = flow_net_res_blocks_5_mlp_2_bias_to_fp16, weight = flow_net_res_blocks_5_mlp_2_weight_to_fp16, x = input_469_cast_fp16)[name = tensor("linear_179_cast_fp16")]; + tensor var_2724_cast_fp16 = mul(x = var_2695_cast_fp16_2, y = linear_179_cast_fp16)[name = tensor("op_2724_cast_fp16")]; + tensor x_305_cast_fp16 = add(x = x_299_cast_fp16, y = var_2724_cast_fp16)[name = tensor("x_305_cast_fp16")]; + tensor linear_180_cast_fp16 = linear(bias = flow_net_final_layer_adaLN_modulation_1_bias_to_fp16, weight = flow_net_final_layer_adaLN_modulation_1_weight_to_fp16, x = input_423_cast_fp16)[name = tensor("linear_180_cast_fp16")]; + tensor var_2733_split_sizes_0 = const()[name = tensor("op_2733_split_sizes_0"), val = tensor([512, 512])]; + tensor var_2733_axis_0 = const()[name = tensor("op_2733_axis_0"), val = tensor(-1)]; + tensor var_2733_cast_fp16_0, tensor var_2733_cast_fp16_1 = split(axis = var_2733_axis_0, split_sizes = var_2733_split_sizes_0, x = linear_180_cast_fp16)[name = tensor("op_2733_cast_fp16")]; + tensor mean_97_axes_0 = const()[name = tensor("mean_97_axes_0"), val = tensor([-1])]; + tensor mean_97_keep_dims_0 = const()[name = tensor("mean_97_keep_dims_0"), val = tensor(true)]; + tensor mean_97_cast_fp16 = reduce_mean(axes = mean_97_axes_0, keep_dims = mean_97_keep_dims_0, x = x_305_cast_fp16)[name = tensor("mean_97_cast_fp16")]; + tensor sub_76_cast_fp16 = sub(x = x_305_cast_fp16, y = mean_97_cast_fp16)[name = tensor("sub_76_cast_fp16")]; + tensor square_62_cast_fp16 = square(x = sub_76_cast_fp16)[name = tensor("square_62_cast_fp16")]; + tensor reduce_mean_125_axes_0 = const()[name = tensor("reduce_mean_125_axes_0"), val = tensor([-1])]; + tensor reduce_mean_125_keep_dims_0 = const()[name = tensor("reduce_mean_125_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_125_cast_fp16 = reduce_mean(axes = reduce_mean_125_axes_0, keep_dims = reduce_mean_125_keep_dims_0, x = square_62_cast_fp16)[name = tensor("reduce_mean_125_cast_fp16")]; + tensor var_2740_to_fp16 = const()[name = tensor("op_2740_to_fp16"), val = tensor(0x1.1p-20)]; + tensor var_2741_cast_fp16 = add(x = reduce_mean_125_cast_fp16, y = var_2740_to_fp16)[name = tensor("op_2741_cast_fp16")]; + tensor var_2742_cast_fp16 = sqrt(x = var_2741_cast_fp16)[name = tensor("op_2742_cast_fp16")]; + tensor x_307_cast_fp16 = real_div(x = sub_76_cast_fp16, y = var_2742_cast_fp16)[name = tensor("x_307_cast_fp16")]; + tensor var_2744_promoted_to_fp16 = const()[name = tensor("op_2744_promoted_to_fp16"), val = tensor(0x1p+0)]; + tensor var_2745_cast_fp16 = add(x = var_2733_cast_fp16_1, y = var_2744_promoted_to_fp16)[name = tensor("op_2745_cast_fp16")]; + tensor var_2746_cast_fp16 = mul(x = x_307_cast_fp16, y = var_2745_cast_fp16)[name = tensor("op_2746_cast_fp16")]; + tensor input_473_cast_fp16 = add(x = var_2746_cast_fp16, y = var_2733_cast_fp16_0)[name = tensor("input_473_cast_fp16")]; + tensor linear_181_cast_fp16 = linear(bias = flow_net_final_layer_linear_bias_to_fp16, weight = flow_net_final_layer_linear_weight_to_fp16, x = input_473_cast_fp16)[name = tensor("linear_181_cast_fp16")]; + tensor var_2757_to_fp16 = const()[name = tensor("op_2757_to_fp16"), val = tensor(0x1p-3)]; + tensor var_2758_cast_fp16 = mul(x = linear_181_cast_fp16, y = var_2757_to_fp16)[name = tensor("op_2758_cast_fp16")]; + tensor input_475_cast_fp16 = add(x = input_407_cast_fp16, y = var_2758_cast_fp16)[name = tensor("input_475_cast_fp16")]; + tensor linear_182_cast_fp16 = linear(bias = flow_net_input_proj_bias_to_fp16, weight = flow_net_input_proj_weight_to_fp16, x = input_475_cast_fp16)[name = tensor("linear_182_cast_fp16")]; + tensor input_479_to_fp16 = const()[name = tensor("input_479_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(19010432)))]; + tensor input_481_cast_fp16 = silu(x = input_479_to_fp16)[name = tensor("input_481_cast_fp16")]; + tensor linear_184_cast_fp16 = linear(bias = flow_net_time_embed_0_mlp_2_bias_to_fp16, weight = flow_net_time_embed_0_mlp_2_weight_to_fp16, x = input_481_cast_fp16)[name = tensor("linear_184_cast_fp16")]; + tensor reduce_mean_126_axes_0 = const()[name = tensor("reduce_mean_126_axes_0"), val = tensor([-1])]; + tensor reduce_mean_126_keep_dims_0 = const()[name = tensor("reduce_mean_126_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_126_cast_fp16 = reduce_mean(axes = reduce_mean_126_axes_0, keep_dims = reduce_mean_126_keep_dims_0, x = linear_184_cast_fp16)[name = tensor("reduce_mean_126_cast_fp16")]; + tensor sub_77_cast_fp16 = sub(x = linear_184_cast_fp16, y = reduce_mean_126_cast_fp16)[name = tensor("sub_77_cast_fp16")]; + tensor square_63_cast_fp16 = square(x = sub_77_cast_fp16)[name = tensor("square_63_cast_fp16")]; + tensor reduce_mean_127_axes_0 = const()[name = tensor("reduce_mean_127_axes_0"), val = tensor([-1])]; + tensor reduce_mean_127_keep_dims_0 = const()[name = tensor("reduce_mean_127_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_127_cast_fp16 = reduce_mean(axes = reduce_mean_127_axes_0, keep_dims = reduce_mean_127_keep_dims_0, x = square_63_cast_fp16)[name = tensor("reduce_mean_127_cast_fp16")]; + tensor real_div_14_to_fp16 = const()[name = tensor("real_div_14_to_fp16"), val = tensor(0x1.008p+0)]; + tensor mul_14_cast_fp16 = mul(x = reduce_mean_127_cast_fp16, y = real_div_14_to_fp16)[name = tensor("mul_14_cast_fp16")]; + tensor var_2824_to_fp16 = const()[name = tensor("op_2824_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_127_cast_fp16_0 = add(x = mul_14_cast_fp16, y = var_2824_to_fp16)[name = tensor("var_127_cast_fp16")]; + tensor var_2827_epsilon_0 = const()[name = tensor("op_2827_epsilon_0"), val = tensor(0x1.197998p-40)]; + tensor var_2827_cast_fp16 = rsqrt(epsilon = var_2827_epsilon_0, x = var_127_cast_fp16_0)[name = tensor("op_2827_cast_fp16")]; + tensor var_2828_cast_fp16 = mul(x = const_3_to_fp16, y = var_2827_cast_fp16)[name = tensor("op_2828_cast_fp16")]; + tensor var_2829_cast_fp16 = mul(x = linear_184_cast_fp16, y = var_2828_cast_fp16)[name = tensor("op_2829_cast_fp16")]; + tensor input_485_to_fp16 = const()[name = tensor("input_485_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(19011520)))]; + tensor input_487_cast_fp16 = silu(x = input_485_to_fp16)[name = tensor("input_487_cast_fp16")]; + tensor linear_186_cast_fp16 = linear(bias = flow_net_time_embed_1_mlp_2_bias_to_fp16, weight = flow_net_time_embed_1_mlp_2_weight_to_fp16, x = input_487_cast_fp16)[name = tensor("linear_186_cast_fp16")]; + tensor reduce_mean_128_axes_0 = const()[name = tensor("reduce_mean_128_axes_0"), val = tensor([-1])]; + tensor reduce_mean_128_keep_dims_0 = const()[name = tensor("reduce_mean_128_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_128_cast_fp16 = reduce_mean(axes = reduce_mean_128_axes_0, keep_dims = reduce_mean_128_keep_dims_0, x = linear_186_cast_fp16)[name = tensor("reduce_mean_128_cast_fp16")]; + tensor sub_79_cast_fp16 = sub(x = linear_186_cast_fp16, y = reduce_mean_128_cast_fp16)[name = tensor("sub_79_cast_fp16")]; + tensor square_64_cast_fp16 = square(x = sub_79_cast_fp16)[name = tensor("square_64_cast_fp16")]; + tensor reduce_mean_129_axes_0 = const()[name = tensor("reduce_mean_129_axes_0"), val = tensor([-1])]; + tensor reduce_mean_129_keep_dims_0 = const()[name = tensor("reduce_mean_129_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_129_cast_fp16 = reduce_mean(axes = reduce_mean_129_axes_0, keep_dims = reduce_mean_129_keep_dims_0, x = square_64_cast_fp16)[name = tensor("reduce_mean_129_cast_fp16")]; + tensor real_div_15_to_fp16 = const()[name = tensor("real_div_15_to_fp16"), val = tensor(0x1.008p+0)]; + tensor mul_15_cast_fp16 = mul(x = reduce_mean_129_cast_fp16, y = real_div_15_to_fp16)[name = tensor("mul_15_cast_fp16")]; + tensor var_2851_to_fp16 = const()[name = tensor("op_2851_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_129_cast_fp16 = add(x = mul_15_cast_fp16, y = var_2851_to_fp16)[name = tensor("var_129_cast_fp16")]; + tensor var_2854_epsilon_0 = const()[name = tensor("op_2854_epsilon_0"), val = tensor(0x1.197998p-40)]; + tensor var_2854_cast_fp16 = rsqrt(epsilon = var_2854_epsilon_0, x = var_129_cast_fp16)[name = tensor("op_2854_cast_fp16")]; + tensor var_2855_cast_fp16 = mul(x = const_5_to_fp16, y = var_2854_cast_fp16)[name = tensor("op_2855_cast_fp16")]; + tensor var_2856_cast_fp16 = mul(x = linear_186_cast_fp16, y = var_2855_cast_fp16)[name = tensor("op_2856_cast_fp16")]; + tensor var_2858_cast_fp16 = add(x = var_2829_cast_fp16, y = var_2856_cast_fp16)[name = tensor("op_2858_cast_fp16")]; + tensor _inversed_t_combined_y_0_to_fp16 = const()[name = tensor("_inversed_t_combined_y_0_to_fp16"), val = tensor(0x1p-1)]; + tensor _inversed_t_combined_cast_fp16 = mul(x = var_2858_cast_fp16, y = _inversed_t_combined_y_0_to_fp16)[name = tensor("_inversed_t_combined_cast_fp16")]; + tensor input_489_cast_fp16 = add(x = _inversed_t_combined_cast_fp16, y = linear_5_cast_fp16)[name = tensor("input_489_cast_fp16")]; + tensor input_491_cast_fp16 = silu(x = input_489_cast_fp16)[name = tensor("input_491_cast_fp16")]; + tensor linear_188_cast_fp16 = linear(bias = flow_net_res_blocks_0_adaLN_modulation_1_bias_to_fp16, weight = flow_net_res_blocks_0_adaLN_modulation_1_weight_to_fp16, x = input_491_cast_fp16)[name = tensor("linear_188_cast_fp16")]; + tensor var_2873_split_sizes_0 = const()[name = tensor("op_2873_split_sizes_0"), val = tensor([512, 512, 512])]; + tensor var_2873_axis_0 = const()[name = tensor("op_2873_axis_0"), val = tensor(-1)]; + tensor var_2873_cast_fp16_0, tensor var_2873_cast_fp16_1, tensor var_2873_cast_fp16_2 = split(axis = var_2873_axis_0, split_sizes = var_2873_split_sizes_0, x = linear_188_cast_fp16)[name = tensor("op_2873_cast_fp16")]; + tensor mean_99_axes_0 = const()[name = tensor("mean_99_axes_0"), val = tensor([-1])]; + tensor mean_99_keep_dims_0 = const()[name = tensor("mean_99_keep_dims_0"), val = tensor(true)]; + tensor mean_99_cast_fp16 = reduce_mean(axes = mean_99_axes_0, keep_dims = mean_99_keep_dims_0, x = linear_182_cast_fp16)[name = tensor("mean_99_cast_fp16")]; + tensor sub_81_cast_fp16 = sub(x = linear_182_cast_fp16, y = mean_99_cast_fp16)[name = tensor("sub_81_cast_fp16")]; + tensor square_65_cast_fp16 = square(x = sub_81_cast_fp16)[name = tensor("square_65_cast_fp16")]; + tensor reduce_mean_131_axes_0 = const()[name = tensor("reduce_mean_131_axes_0"), val = tensor([-1])]; + tensor reduce_mean_131_keep_dims_0 = const()[name = tensor("reduce_mean_131_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_131_cast_fp16 = reduce_mean(axes = reduce_mean_131_axes_0, keep_dims = reduce_mean_131_keep_dims_0, x = square_65_cast_fp16)[name = tensor("reduce_mean_131_cast_fp16")]; + tensor var_2883_to_fp16 = const()[name = tensor("op_2883_to_fp16"), val = tensor(0x1.1p-20)]; + tensor var_2884_cast_fp16 = add(x = reduce_mean_131_cast_fp16, y = var_2883_to_fp16)[name = tensor("op_2884_cast_fp16")]; + tensor var_2885_cast_fp16 = sqrt(x = var_2884_cast_fp16)[name = tensor("op_2885_cast_fp16")]; + tensor x_315_cast_fp16 = real_div(x = sub_81_cast_fp16, y = var_2885_cast_fp16)[name = tensor("x_315_cast_fp16")]; + tensor var_2887_cast_fp16 = mul(x = x_315_cast_fp16, y = flow_net_res_blocks_0_in_ln_weight_to_fp16)[name = tensor("op_2887_cast_fp16")]; + tensor x_317_cast_fp16 = add(x = var_2887_cast_fp16, y = flow_net_res_blocks_0_in_ln_bias_to_fp16)[name = tensor("x_317_cast_fp16")]; + tensor var_2889_promoted_to_fp16 = const()[name = tensor("op_2889_promoted_to_fp16"), val = tensor(0x1p+0)]; + tensor var_2890_cast_fp16 = add(x = var_2873_cast_fp16_1, y = var_2889_promoted_to_fp16)[name = tensor("op_2890_cast_fp16")]; + tensor var_2891_cast_fp16 = mul(x = x_317_cast_fp16, y = var_2890_cast_fp16)[name = tensor("op_2891_cast_fp16")]; + tensor input_493_cast_fp16 = add(x = var_2891_cast_fp16, y = var_2873_cast_fp16_0)[name = tensor("input_493_cast_fp16")]; + tensor linear_189_cast_fp16 = linear(bias = flow_net_res_blocks_0_mlp_0_bias_to_fp16, weight = flow_net_res_blocks_0_mlp_0_weight_to_fp16, x = input_493_cast_fp16)[name = tensor("linear_189_cast_fp16")]; + tensor input_497_cast_fp16 = silu(x = linear_189_cast_fp16)[name = tensor("input_497_cast_fp16")]; + tensor linear_190_cast_fp16 = linear(bias = flow_net_res_blocks_0_mlp_2_bias_to_fp16, weight = flow_net_res_blocks_0_mlp_2_weight_to_fp16, x = input_497_cast_fp16)[name = tensor("linear_190_cast_fp16")]; + tensor var_2902_cast_fp16 = mul(x = var_2873_cast_fp16_2, y = linear_190_cast_fp16)[name = tensor("op_2902_cast_fp16")]; + tensor x_319_cast_fp16 = add(x = linear_182_cast_fp16, y = var_2902_cast_fp16)[name = tensor("x_319_cast_fp16")]; + tensor linear_191_cast_fp16 = linear(bias = flow_net_res_blocks_1_adaLN_modulation_1_bias_to_fp16, weight = flow_net_res_blocks_1_adaLN_modulation_1_weight_to_fp16, x = input_491_cast_fp16)[name = tensor("linear_191_cast_fp16")]; + tensor var_2912_split_sizes_0 = const()[name = tensor("op_2912_split_sizes_0"), val = tensor([512, 512, 512])]; + tensor var_2912_axis_0 = const()[name = tensor("op_2912_axis_0"), val = tensor(-1)]; + tensor var_2912_cast_fp16_0, tensor var_2912_cast_fp16_1, tensor var_2912_cast_fp16_2 = split(axis = var_2912_axis_0, split_sizes = var_2912_split_sizes_0, x = linear_191_cast_fp16)[name = tensor("op_2912_cast_fp16")]; + tensor mean_101_axes_0 = const()[name = tensor("mean_101_axes_0"), val = tensor([-1])]; + tensor mean_101_keep_dims_0 = const()[name = tensor("mean_101_keep_dims_0"), val = tensor(true)]; + tensor mean_101_cast_fp16 = reduce_mean(axes = mean_101_axes_0, keep_dims = mean_101_keep_dims_0, x = x_319_cast_fp16)[name = tensor("mean_101_cast_fp16")]; + tensor sub_82_cast_fp16 = sub(x = x_319_cast_fp16, y = mean_101_cast_fp16)[name = tensor("sub_82_cast_fp16")]; + tensor square_66_cast_fp16 = square(x = sub_82_cast_fp16)[name = tensor("square_66_cast_fp16")]; + tensor reduce_mean_133_axes_0 = const()[name = tensor("reduce_mean_133_axes_0"), val = tensor([-1])]; + tensor reduce_mean_133_keep_dims_0 = const()[name = tensor("reduce_mean_133_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_133_cast_fp16 = reduce_mean(axes = reduce_mean_133_axes_0, keep_dims = reduce_mean_133_keep_dims_0, x = square_66_cast_fp16)[name = tensor("reduce_mean_133_cast_fp16")]; + tensor var_2922_to_fp16 = const()[name = tensor("op_2922_to_fp16"), val = tensor(0x1.1p-20)]; + tensor var_2923_cast_fp16 = add(x = reduce_mean_133_cast_fp16, y = var_2922_to_fp16)[name = tensor("op_2923_cast_fp16")]; + tensor var_2924_cast_fp16 = sqrt(x = var_2923_cast_fp16)[name = tensor("op_2924_cast_fp16")]; + tensor x_321_cast_fp16 = real_div(x = sub_82_cast_fp16, y = var_2924_cast_fp16)[name = tensor("x_321_cast_fp16")]; + tensor var_2926_cast_fp16 = mul(x = x_321_cast_fp16, y = flow_net_res_blocks_1_in_ln_weight_to_fp16)[name = tensor("op_2926_cast_fp16")]; + tensor x_323_cast_fp16 = add(x = var_2926_cast_fp16, y = flow_net_res_blocks_1_in_ln_bias_to_fp16)[name = tensor("x_323_cast_fp16")]; + tensor var_2928_promoted_to_fp16 = const()[name = tensor("op_2928_promoted_to_fp16"), val = tensor(0x1p+0)]; + tensor var_2929_cast_fp16 = add(x = var_2912_cast_fp16_1, y = var_2928_promoted_to_fp16)[name = tensor("op_2929_cast_fp16")]; + tensor var_2930_cast_fp16 = mul(x = x_323_cast_fp16, y = var_2929_cast_fp16)[name = tensor("op_2930_cast_fp16")]; + tensor input_501_cast_fp16 = add(x = var_2930_cast_fp16, y = var_2912_cast_fp16_0)[name = tensor("input_501_cast_fp16")]; + tensor linear_192_cast_fp16 = linear(bias = flow_net_res_blocks_1_mlp_0_bias_to_fp16, weight = flow_net_res_blocks_1_mlp_0_weight_to_fp16, x = input_501_cast_fp16)[name = tensor("linear_192_cast_fp16")]; + tensor input_505_cast_fp16 = silu(x = linear_192_cast_fp16)[name = tensor("input_505_cast_fp16")]; + tensor linear_193_cast_fp16 = linear(bias = flow_net_res_blocks_1_mlp_2_bias_to_fp16, weight = flow_net_res_blocks_1_mlp_2_weight_to_fp16, x = input_505_cast_fp16)[name = tensor("linear_193_cast_fp16")]; + tensor var_2941_cast_fp16 = mul(x = var_2912_cast_fp16_2, y = linear_193_cast_fp16)[name = tensor("op_2941_cast_fp16")]; + tensor x_325_cast_fp16 = add(x = x_319_cast_fp16, y = var_2941_cast_fp16)[name = tensor("x_325_cast_fp16")]; + tensor linear_194_cast_fp16 = linear(bias = flow_net_res_blocks_2_adaLN_modulation_1_bias_to_fp16, weight = flow_net_res_blocks_2_adaLN_modulation_1_weight_to_fp16, x = input_491_cast_fp16)[name = tensor("linear_194_cast_fp16")]; + tensor var_2951_split_sizes_0 = const()[name = tensor("op_2951_split_sizes_0"), val = tensor([512, 512, 512])]; + tensor var_2951_axis_0 = const()[name = tensor("op_2951_axis_0"), val = tensor(-1)]; + tensor var_2951_cast_fp16_0, tensor var_2951_cast_fp16_1, tensor var_2951_cast_fp16_2 = split(axis = var_2951_axis_0, split_sizes = var_2951_split_sizes_0, x = linear_194_cast_fp16)[name = tensor("op_2951_cast_fp16")]; + tensor mean_103_axes_0 = const()[name = tensor("mean_103_axes_0"), val = tensor([-1])]; + tensor mean_103_keep_dims_0 = const()[name = tensor("mean_103_keep_dims_0"), val = tensor(true)]; + tensor mean_103_cast_fp16 = reduce_mean(axes = mean_103_axes_0, keep_dims = mean_103_keep_dims_0, x = x_325_cast_fp16)[name = tensor("mean_103_cast_fp16")]; + tensor sub_83_cast_fp16 = sub(x = x_325_cast_fp16, y = mean_103_cast_fp16)[name = tensor("sub_83_cast_fp16")]; + tensor square_67_cast_fp16 = square(x = sub_83_cast_fp16)[name = tensor("square_67_cast_fp16")]; + tensor reduce_mean_135_axes_0 = const()[name = tensor("reduce_mean_135_axes_0"), val = tensor([-1])]; + tensor reduce_mean_135_keep_dims_0 = const()[name = tensor("reduce_mean_135_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_135_cast_fp16 = reduce_mean(axes = reduce_mean_135_axes_0, keep_dims = reduce_mean_135_keep_dims_0, x = square_67_cast_fp16)[name = tensor("reduce_mean_135_cast_fp16")]; + tensor var_2961_to_fp16 = const()[name = tensor("op_2961_to_fp16"), val = tensor(0x1.1p-20)]; + tensor var_2962_cast_fp16 = add(x = reduce_mean_135_cast_fp16, y = var_2961_to_fp16)[name = tensor("op_2962_cast_fp16")]; + tensor var_2963_cast_fp16 = sqrt(x = var_2962_cast_fp16)[name = tensor("op_2963_cast_fp16")]; + tensor x_327_cast_fp16 = real_div(x = sub_83_cast_fp16, y = var_2963_cast_fp16)[name = tensor("x_327_cast_fp16")]; + tensor var_2965_cast_fp16 = mul(x = x_327_cast_fp16, y = flow_net_res_blocks_2_in_ln_weight_to_fp16)[name = tensor("op_2965_cast_fp16")]; + tensor x_329_cast_fp16 = add(x = var_2965_cast_fp16, y = flow_net_res_blocks_2_in_ln_bias_to_fp16)[name = tensor("x_329_cast_fp16")]; + tensor var_2967_promoted_to_fp16 = const()[name = tensor("op_2967_promoted_to_fp16"), val = tensor(0x1p+0)]; + tensor var_2968_cast_fp16 = add(x = var_2951_cast_fp16_1, y = var_2967_promoted_to_fp16)[name = tensor("op_2968_cast_fp16")]; + tensor var_2969_cast_fp16 = mul(x = x_329_cast_fp16, y = var_2968_cast_fp16)[name = tensor("op_2969_cast_fp16")]; + tensor input_509_cast_fp16 = add(x = var_2969_cast_fp16, y = var_2951_cast_fp16_0)[name = tensor("input_509_cast_fp16")]; + tensor linear_195_cast_fp16 = linear(bias = flow_net_res_blocks_2_mlp_0_bias_to_fp16, weight = flow_net_res_blocks_2_mlp_0_weight_to_fp16, x = input_509_cast_fp16)[name = tensor("linear_195_cast_fp16")]; + tensor input_513_cast_fp16 = silu(x = linear_195_cast_fp16)[name = tensor("input_513_cast_fp16")]; + tensor linear_196_cast_fp16 = linear(bias = flow_net_res_blocks_2_mlp_2_bias_to_fp16, weight = flow_net_res_blocks_2_mlp_2_weight_to_fp16, x = input_513_cast_fp16)[name = tensor("linear_196_cast_fp16")]; + tensor var_2980_cast_fp16 = mul(x = var_2951_cast_fp16_2, y = linear_196_cast_fp16)[name = tensor("op_2980_cast_fp16")]; + tensor x_331_cast_fp16 = add(x = x_325_cast_fp16, y = var_2980_cast_fp16)[name = tensor("x_331_cast_fp16")]; + tensor linear_197_cast_fp16 = linear(bias = flow_net_res_blocks_3_adaLN_modulation_1_bias_to_fp16, weight = flow_net_res_blocks_3_adaLN_modulation_1_weight_to_fp16, x = input_491_cast_fp16)[name = tensor("linear_197_cast_fp16")]; + tensor var_2990_split_sizes_0 = const()[name = tensor("op_2990_split_sizes_0"), val = tensor([512, 512, 512])]; + tensor var_2990_axis_0 = const()[name = tensor("op_2990_axis_0"), val = tensor(-1)]; + tensor var_2990_cast_fp16_0, tensor var_2990_cast_fp16_1, tensor var_2990_cast_fp16_2 = split(axis = var_2990_axis_0, split_sizes = var_2990_split_sizes_0, x = linear_197_cast_fp16)[name = tensor("op_2990_cast_fp16")]; + tensor mean_105_axes_0 = const()[name = tensor("mean_105_axes_0"), val = tensor([-1])]; + tensor mean_105_keep_dims_0 = const()[name = tensor("mean_105_keep_dims_0"), val = tensor(true)]; + tensor mean_105_cast_fp16 = reduce_mean(axes = mean_105_axes_0, keep_dims = mean_105_keep_dims_0, x = x_331_cast_fp16)[name = tensor("mean_105_cast_fp16")]; + tensor sub_84_cast_fp16 = sub(x = x_331_cast_fp16, y = mean_105_cast_fp16)[name = tensor("sub_84_cast_fp16")]; + tensor square_68_cast_fp16 = square(x = sub_84_cast_fp16)[name = tensor("square_68_cast_fp16")]; + tensor reduce_mean_137_axes_0 = const()[name = tensor("reduce_mean_137_axes_0"), val = tensor([-1])]; + tensor reduce_mean_137_keep_dims_0 = const()[name = tensor("reduce_mean_137_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_137_cast_fp16 = reduce_mean(axes = reduce_mean_137_axes_0, keep_dims = reduce_mean_137_keep_dims_0, x = square_68_cast_fp16)[name = tensor("reduce_mean_137_cast_fp16")]; + tensor var_3000_to_fp16 = const()[name = tensor("op_3000_to_fp16"), val = tensor(0x1.1p-20)]; + tensor var_3001_cast_fp16 = add(x = reduce_mean_137_cast_fp16, y = var_3000_to_fp16)[name = tensor("op_3001_cast_fp16")]; + tensor var_3002_cast_fp16 = sqrt(x = var_3001_cast_fp16)[name = tensor("op_3002_cast_fp16")]; + tensor x_333_cast_fp16 = real_div(x = sub_84_cast_fp16, y = var_3002_cast_fp16)[name = tensor("x_333_cast_fp16")]; + tensor var_3004_cast_fp16 = mul(x = x_333_cast_fp16, y = flow_net_res_blocks_3_in_ln_weight_to_fp16)[name = tensor("op_3004_cast_fp16")]; + tensor x_335_cast_fp16 = add(x = var_3004_cast_fp16, y = flow_net_res_blocks_3_in_ln_bias_to_fp16)[name = tensor("x_335_cast_fp16")]; + tensor var_3006_promoted_to_fp16 = const()[name = tensor("op_3006_promoted_to_fp16"), val = tensor(0x1p+0)]; + tensor var_3007_cast_fp16 = add(x = var_2990_cast_fp16_1, y = var_3006_promoted_to_fp16)[name = tensor("op_3007_cast_fp16")]; + tensor var_3008_cast_fp16 = mul(x = x_335_cast_fp16, y = var_3007_cast_fp16)[name = tensor("op_3008_cast_fp16")]; + tensor input_517_cast_fp16 = add(x = var_3008_cast_fp16, y = var_2990_cast_fp16_0)[name = tensor("input_517_cast_fp16")]; + tensor linear_198_cast_fp16 = linear(bias = flow_net_res_blocks_3_mlp_0_bias_to_fp16, weight = flow_net_res_blocks_3_mlp_0_weight_to_fp16, x = input_517_cast_fp16)[name = tensor("linear_198_cast_fp16")]; + tensor input_521_cast_fp16 = silu(x = linear_198_cast_fp16)[name = tensor("input_521_cast_fp16")]; + tensor linear_199_cast_fp16 = linear(bias = flow_net_res_blocks_3_mlp_2_bias_to_fp16, weight = flow_net_res_blocks_3_mlp_2_weight_to_fp16, x = input_521_cast_fp16)[name = tensor("linear_199_cast_fp16")]; + tensor var_3019_cast_fp16 = mul(x = var_2990_cast_fp16_2, y = linear_199_cast_fp16)[name = tensor("op_3019_cast_fp16")]; + tensor x_337_cast_fp16 = add(x = x_331_cast_fp16, y = var_3019_cast_fp16)[name = tensor("x_337_cast_fp16")]; + tensor linear_200_cast_fp16 = linear(bias = flow_net_res_blocks_4_adaLN_modulation_1_bias_to_fp16, weight = flow_net_res_blocks_4_adaLN_modulation_1_weight_to_fp16, x = input_491_cast_fp16)[name = tensor("linear_200_cast_fp16")]; + tensor var_3029_split_sizes_0 = const()[name = tensor("op_3029_split_sizes_0"), val = tensor([512, 512, 512])]; + tensor var_3029_axis_0 = const()[name = tensor("op_3029_axis_0"), val = tensor(-1)]; + tensor var_3029_cast_fp16_0, tensor var_3029_cast_fp16_1, tensor var_3029_cast_fp16_2 = split(axis = var_3029_axis_0, split_sizes = var_3029_split_sizes_0, x = linear_200_cast_fp16)[name = tensor("op_3029_cast_fp16")]; + tensor mean_107_axes_0 = const()[name = tensor("mean_107_axes_0"), val = tensor([-1])]; + tensor mean_107_keep_dims_0 = const()[name = tensor("mean_107_keep_dims_0"), val = tensor(true)]; + tensor mean_107_cast_fp16 = reduce_mean(axes = mean_107_axes_0, keep_dims = mean_107_keep_dims_0, x = x_337_cast_fp16)[name = tensor("mean_107_cast_fp16")]; + tensor sub_85_cast_fp16 = sub(x = x_337_cast_fp16, y = mean_107_cast_fp16)[name = tensor("sub_85_cast_fp16")]; + tensor square_69_cast_fp16 = square(x = sub_85_cast_fp16)[name = tensor("square_69_cast_fp16")]; + tensor reduce_mean_139_axes_0 = const()[name = tensor("reduce_mean_139_axes_0"), val = tensor([-1])]; + tensor reduce_mean_139_keep_dims_0 = const()[name = tensor("reduce_mean_139_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_139_cast_fp16 = reduce_mean(axes = reduce_mean_139_axes_0, keep_dims = reduce_mean_139_keep_dims_0, x = square_69_cast_fp16)[name = tensor("reduce_mean_139_cast_fp16")]; + tensor var_3039_to_fp16 = const()[name = tensor("op_3039_to_fp16"), val = tensor(0x1.1p-20)]; + tensor var_3040_cast_fp16 = add(x = reduce_mean_139_cast_fp16, y = var_3039_to_fp16)[name = tensor("op_3040_cast_fp16")]; + tensor var_3041_cast_fp16 = sqrt(x = var_3040_cast_fp16)[name = tensor("op_3041_cast_fp16")]; + tensor x_339_cast_fp16 = real_div(x = sub_85_cast_fp16, y = var_3041_cast_fp16)[name = tensor("x_339_cast_fp16")]; + tensor var_3043_cast_fp16 = mul(x = x_339_cast_fp16, y = flow_net_res_blocks_4_in_ln_weight_to_fp16)[name = tensor("op_3043_cast_fp16")]; + tensor x_341_cast_fp16 = add(x = var_3043_cast_fp16, y = flow_net_res_blocks_4_in_ln_bias_to_fp16)[name = tensor("x_341_cast_fp16")]; + tensor var_3045_promoted_to_fp16 = const()[name = tensor("op_3045_promoted_to_fp16"), val = tensor(0x1p+0)]; + tensor var_3046_cast_fp16 = add(x = var_3029_cast_fp16_1, y = var_3045_promoted_to_fp16)[name = tensor("op_3046_cast_fp16")]; + tensor var_3047_cast_fp16 = mul(x = x_341_cast_fp16, y = var_3046_cast_fp16)[name = tensor("op_3047_cast_fp16")]; + tensor input_525_cast_fp16 = add(x = var_3047_cast_fp16, y = var_3029_cast_fp16_0)[name = tensor("input_525_cast_fp16")]; + tensor linear_201_cast_fp16 = linear(bias = flow_net_res_blocks_4_mlp_0_bias_to_fp16, weight = flow_net_res_blocks_4_mlp_0_weight_to_fp16, x = input_525_cast_fp16)[name = tensor("linear_201_cast_fp16")]; + tensor input_529_cast_fp16 = silu(x = linear_201_cast_fp16)[name = tensor("input_529_cast_fp16")]; + tensor linear_202_cast_fp16 = linear(bias = flow_net_res_blocks_4_mlp_2_bias_to_fp16, weight = flow_net_res_blocks_4_mlp_2_weight_to_fp16, x = input_529_cast_fp16)[name = tensor("linear_202_cast_fp16")]; + tensor var_3058_cast_fp16 = mul(x = var_3029_cast_fp16_2, y = linear_202_cast_fp16)[name = tensor("op_3058_cast_fp16")]; + tensor x_343_cast_fp16 = add(x = x_337_cast_fp16, y = var_3058_cast_fp16)[name = tensor("x_343_cast_fp16")]; + tensor linear_203_cast_fp16 = linear(bias = flow_net_res_blocks_5_adaLN_modulation_1_bias_to_fp16, weight = flow_net_res_blocks_5_adaLN_modulation_1_weight_to_fp16, x = input_491_cast_fp16)[name = tensor("linear_203_cast_fp16")]; + tensor var_3068_split_sizes_0 = const()[name = tensor("op_3068_split_sizes_0"), val = tensor([512, 512, 512])]; + tensor var_3068_axis_0 = const()[name = tensor("op_3068_axis_0"), val = tensor(-1)]; + tensor var_3068_cast_fp16_0, tensor var_3068_cast_fp16_1, tensor var_3068_cast_fp16_2 = split(axis = var_3068_axis_0, split_sizes = var_3068_split_sizes_0, x = linear_203_cast_fp16)[name = tensor("op_3068_cast_fp16")]; + tensor mean_109_axes_0 = const()[name = tensor("mean_109_axes_0"), val = tensor([-1])]; + tensor mean_109_keep_dims_0 = const()[name = tensor("mean_109_keep_dims_0"), val = tensor(true)]; + tensor mean_109_cast_fp16 = reduce_mean(axes = mean_109_axes_0, keep_dims = mean_109_keep_dims_0, x = x_343_cast_fp16)[name = tensor("mean_109_cast_fp16")]; + tensor sub_86_cast_fp16 = sub(x = x_343_cast_fp16, y = mean_109_cast_fp16)[name = tensor("sub_86_cast_fp16")]; + tensor square_70_cast_fp16 = square(x = sub_86_cast_fp16)[name = tensor("square_70_cast_fp16")]; + tensor reduce_mean_141_axes_0 = const()[name = tensor("reduce_mean_141_axes_0"), val = tensor([-1])]; + tensor reduce_mean_141_keep_dims_0 = const()[name = tensor("reduce_mean_141_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_141_cast_fp16 = reduce_mean(axes = reduce_mean_141_axes_0, keep_dims = reduce_mean_141_keep_dims_0, x = square_70_cast_fp16)[name = tensor("reduce_mean_141_cast_fp16")]; + tensor var_3078_to_fp16 = const()[name = tensor("op_3078_to_fp16"), val = tensor(0x1.1p-20)]; + tensor var_3079_cast_fp16 = add(x = reduce_mean_141_cast_fp16, y = var_3078_to_fp16)[name = tensor("op_3079_cast_fp16")]; + tensor var_3080_cast_fp16 = sqrt(x = var_3079_cast_fp16)[name = tensor("op_3080_cast_fp16")]; + tensor x_345_cast_fp16 = real_div(x = sub_86_cast_fp16, y = var_3080_cast_fp16)[name = tensor("x_345_cast_fp16")]; + tensor var_3082_cast_fp16 = mul(x = x_345_cast_fp16, y = flow_net_res_blocks_5_in_ln_weight_to_fp16)[name = tensor("op_3082_cast_fp16")]; + tensor x_347_cast_fp16 = add(x = var_3082_cast_fp16, y = flow_net_res_blocks_5_in_ln_bias_to_fp16)[name = tensor("x_347_cast_fp16")]; + tensor var_3084_promoted_to_fp16 = const()[name = tensor("op_3084_promoted_to_fp16"), val = tensor(0x1p+0)]; + tensor var_3085_cast_fp16 = add(x = var_3068_cast_fp16_1, y = var_3084_promoted_to_fp16)[name = tensor("op_3085_cast_fp16")]; + tensor var_3086_cast_fp16 = mul(x = x_347_cast_fp16, y = var_3085_cast_fp16)[name = tensor("op_3086_cast_fp16")]; + tensor input_533_cast_fp16 = add(x = var_3086_cast_fp16, y = var_3068_cast_fp16_0)[name = tensor("input_533_cast_fp16")]; + tensor linear_204_cast_fp16 = linear(bias = flow_net_res_blocks_5_mlp_0_bias_to_fp16, weight = flow_net_res_blocks_5_mlp_0_weight_to_fp16, x = input_533_cast_fp16)[name = tensor("linear_204_cast_fp16")]; + tensor input_537_cast_fp16 = silu(x = linear_204_cast_fp16)[name = tensor("input_537_cast_fp16")]; + tensor linear_205_cast_fp16 = linear(bias = flow_net_res_blocks_5_mlp_2_bias_to_fp16, weight = flow_net_res_blocks_5_mlp_2_weight_to_fp16, x = input_537_cast_fp16)[name = tensor("linear_205_cast_fp16")]; + tensor var_3097_cast_fp16 = mul(x = var_3068_cast_fp16_2, y = linear_205_cast_fp16)[name = tensor("op_3097_cast_fp16")]; + tensor x_349_cast_fp16 = add(x = x_343_cast_fp16, y = var_3097_cast_fp16)[name = tensor("x_349_cast_fp16")]; + tensor linear_206_cast_fp16 = linear(bias = flow_net_final_layer_adaLN_modulation_1_bias_to_fp16, weight = flow_net_final_layer_adaLN_modulation_1_weight_to_fp16, x = input_491_cast_fp16)[name = tensor("linear_206_cast_fp16")]; + tensor var_3106_split_sizes_0 = const()[name = tensor("op_3106_split_sizes_0"), val = tensor([512, 512])]; + tensor var_3106_axis_0 = const()[name = tensor("op_3106_axis_0"), val = tensor(-1)]; + tensor var_3106_cast_fp16_0, tensor var_3106_cast_fp16_1 = split(axis = var_3106_axis_0, split_sizes = var_3106_split_sizes_0, x = linear_206_cast_fp16)[name = tensor("op_3106_cast_fp16")]; + tensor mean_axes_0 = const()[name = tensor("mean_axes_0"), val = tensor([-1])]; + tensor mean_keep_dims_0 = const()[name = tensor("mean_keep_dims_0"), val = tensor(true)]; + tensor mean_cast_fp16 = reduce_mean(axes = mean_axes_0, keep_dims = mean_keep_dims_0, x = x_349_cast_fp16)[name = tensor("mean_cast_fp16")]; + tensor sub_87_cast_fp16 = sub(x = x_349_cast_fp16, y = mean_cast_fp16)[name = tensor("sub_87_cast_fp16")]; + tensor square_71_cast_fp16 = square(x = sub_87_cast_fp16)[name = tensor("square_71_cast_fp16")]; + tensor reduce_mean_143_axes_0 = const()[name = tensor("reduce_mean_143_axes_0"), val = tensor([-1])]; + tensor reduce_mean_143_keep_dims_0 = const()[name = tensor("reduce_mean_143_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_143_cast_fp16 = reduce_mean(axes = reduce_mean_143_axes_0, keep_dims = reduce_mean_143_keep_dims_0, x = square_71_cast_fp16)[name = tensor("reduce_mean_143_cast_fp16")]; + tensor var_3113_to_fp16 = const()[name = tensor("op_3113_to_fp16"), val = tensor(0x1.1p-20)]; + tensor var_3114_cast_fp16 = add(x = reduce_mean_143_cast_fp16, y = var_3113_to_fp16)[name = tensor("op_3114_cast_fp16")]; + tensor var_3115_cast_fp16 = sqrt(x = var_3114_cast_fp16)[name = tensor("op_3115_cast_fp16")]; + tensor x_cast_fp16 = real_div(x = sub_87_cast_fp16, y = var_3115_cast_fp16)[name = tensor("x_cast_fp16")]; + tensor var_3117_promoted_to_fp16 = const()[name = tensor("op_3117_promoted_to_fp16"), val = tensor(0x1p+0)]; + tensor var_3118_cast_fp16 = add(x = var_3106_cast_fp16_1, y = var_3117_promoted_to_fp16)[name = tensor("op_3118_cast_fp16")]; + tensor var_3119_cast_fp16 = mul(x = x_cast_fp16, y = var_3118_cast_fp16)[name = tensor("op_3119_cast_fp16")]; + tensor input_cast_fp16 = add(x = var_3119_cast_fp16, y = var_3106_cast_fp16_0)[name = tensor("input_cast_fp16")]; + tensor linear_207_cast_fp16 = linear(bias = flow_net_final_layer_linear_bias_to_fp16, weight = flow_net_final_layer_linear_weight_to_fp16, x = input_cast_fp16)[name = tensor("linear_207_cast_fp16")]; + tensor var_3124_to_fp16 = const()[name = tensor("op_3124_to_fp16"), val = tensor(0x1p-3)]; + tensor var_3125_cast_fp16 = mul(x = linear_207_cast_fp16, y = var_3124_to_fp16)[name = tensor("op_3125_cast_fp16")]; + tensor var_3127_cast_fp16 = add(x = input_475_cast_fp16, y = var_3125_cast_fp16)[name = tensor("op_3127_cast_fp16")]; + tensor var_3127_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_3127_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor latent_final = cast(dtype = var_3127_cast_fp16_to_fp32_dtype_0, x = var_3127_cast_fp16)[name = tensor("cast_144")]; + } -> (latent_final); +} \ No newline at end of file diff --git a/v2.1/english/flow_decoder_fused.mlmodelc/weights/weight.bin b/v2.1/english/flow_decoder_fused.mlmodelc/weights/weight.bin new file mode 100644 index 0000000000000000000000000000000000000000..6e75e3fb7a9763bb94b0c399b2ec940a3129eb4c --- /dev/null +++ b/v2.1/english/flow_decoder_fused.mlmodelc/weights/weight.bin @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:f28a1bd0af940a195802e9d702c9e050f3e69d265e582f3d0fdf4001709eb30c +size 19012608 diff --git a/v2.1/english/flow_decoder_fused.mlpackage/Data/com.apple.CoreML/model.mlmodel b/v2.1/english/flow_decoder_fused.mlpackage/Data/com.apple.CoreML/model.mlmodel new file mode 100644 index 0000000000000000000000000000000000000000..acf78b18b3fd0eb2f50aa2fd441d28e5ec15a9ef --- /dev/null +++ b/v2.1/english/flow_decoder_fused.mlpackage/Data/com.apple.CoreML/model.mlmodel @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:5ad334d91b3bd8d9febf29fafe47040ebe68d63dae5a77df2a2d8447b73c3888 +size 301733 diff --git a/v2.1/english/flow_decoder_fused.mlpackage/Data/com.apple.CoreML/weights/weight.bin b/v2.1/english/flow_decoder_fused.mlpackage/Data/com.apple.CoreML/weights/weight.bin new file mode 100644 index 0000000000000000000000000000000000000000..6e75e3fb7a9763bb94b0c399b2ec940a3129eb4c --- /dev/null +++ b/v2.1/english/flow_decoder_fused.mlpackage/Data/com.apple.CoreML/weights/weight.bin @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:f28a1bd0af940a195802e9d702c9e050f3e69d265e582f3d0fdf4001709eb30c +size 19012608 diff --git a/v2.1/english/flow_decoder_fused.mlpackage/Manifest.json b/v2.1/english/flow_decoder_fused.mlpackage/Manifest.json new file mode 100644 index 0000000000000000000000000000000000000000..4544cc85da738a2753635099042341e02f5d4208 --- /dev/null +++ b/v2.1/english/flow_decoder_fused.mlpackage/Manifest.json @@ -0,0 +1,18 @@ +{ + "fileFormatVersion": "1.0.0", + "itemInfoEntries": { + "4EF2F3E6-5A97-4A21-85B3-7F441F870501": { + "author": "com.apple.CoreML", + "description": "CoreML Model Weights", + "name": "weights", + "path": "com.apple.CoreML/weights" + }, + "6813D315-DF8C-4DDF-9A40-ECF63F127E28": { + "author": "com.apple.CoreML", + "description": "CoreML Model Specification", + "name": "model.mlmodel", + "path": "com.apple.CoreML/model.mlmodel" + } + }, + "rootModelIdentifier": "6813D315-DF8C-4DDF-9A40-ECF63F127E28" +} diff --git a/v2.1/english/flowlm_step.mlmodelc/analytics/coremldata.bin b/v2.1/english/flowlm_step.mlmodelc/analytics/coremldata.bin new file mode 100644 index 0000000000000000000000000000000000000000..25aaa311d0695e186248b8d9d556a44ebe7a1322 --- /dev/null +++ b/v2.1/english/flowlm_step.mlmodelc/analytics/coremldata.bin @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:fcc771807d303b09425871201b9c20fd82b92bdb869ca567e9011402b6b844b3 +size 243 diff --git a/v2.1/english/flowlm_step.mlmodelc/coremldata.bin b/v2.1/english/flowlm_step.mlmodelc/coremldata.bin new file mode 100644 index 0000000000000000000000000000000000000000..29ccd3814adfa9effb111fd40c4143cabe0191b3 --- /dev/null +++ b/v2.1/english/flowlm_step.mlmodelc/coremldata.bin @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:4cf9fcee17a690def93f29ebdffbe3fd4cbfda442666e282f2bc8673947c1707 +size 1191 diff --git a/v2.1/english/flowlm_step.mlmodelc/model.mil b/v2.1/english/flowlm_step.mlmodelc/model.mil new file mode 100644 index 0000000000000000000000000000000000000000..404944a1abd4f0a42b69a9515444e9b6f8e3d75e --- /dev/null +++ b/v2.1/english/flowlm_step.mlmodelc/model.mil @@ -0,0 +1,1359 @@ +program(1.0) +[buildInfo = dict, tensor>({{"coremlc-component-MIL", "3520.4.1"}, {"coremlc-version", "3520.5.1"}, {"coremltools-component-torch", "2.12.0"}, {"coremltools-source-dialect", "TorchScript"}, {"coremltools-version", "9.0"}})] +{ + func main(tensor bos_emb, tensor cache0, tensor cache1, tensor cache2, tensor cache3, tensor cache4, tensor cache5, tensor position0, tensor position1, tensor position2, tensor position3, tensor position4, tensor position5, tensor sequence) { + tensor sequence_to_fp16_dtype_0 = const()[name = tensor("sequence_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor sequence_to_fp16 = cast(dtype = sequence_to_fp16_dtype_0, x = sequence)[name = tensor("cast_141")]; + tensor var_54_cast_fp16 = not_equal(x = sequence_to_fp16, y = sequence_to_fp16)[name = tensor("op_54_cast_fp16")]; + tensor expand_dims_0_axes_0 = const()[name = tensor("expand_dims_0_axes_0"), val = tensor([0, 1])]; + tensor bos_emb_to_fp16_dtype_0 = const()[name = tensor("bos_emb_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor bos_emb_to_fp16 = cast(dtype = bos_emb_to_fp16_dtype_0, x = bos_emb)[name = tensor("cast_140")]; + tensor expand_dims_0_cast_fp16 = expand_dims(axes = expand_dims_0_axes_0, x = bos_emb_to_fp16)[name = tensor("expand_dims_0_cast_fp16")]; + tensor input_1_cast_fp16 = select(a = expand_dims_0_cast_fp16, b = sequence_to_fp16, cond = var_54_cast_fp16)[name = tensor("input_1_cast_fp16")]; + tensor input_linear_weight_to_fp16 = const()[name = tensor("input_linear_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(64)))]; + tensor linear_0_bias_0_to_fp16 = const()[name = tensor("linear_0_bias_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(65664)))]; + tensor linear_0_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = input_linear_weight_to_fp16, x = input_1_cast_fp16)[name = tensor("linear_0_cast_fp16")]; + tensor input_5_axes_0 = const()[name = tensor("input_5_axes_0"), val = tensor([-1])]; + tensor norm0_1_weight_to_fp16 = const()[name = tensor("norm0_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(67776)))]; + tensor norm0_1_bias_to_fp16 = const()[name = tensor("norm0_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(69888)))]; + tensor var_60_to_fp16 = const()[name = tensor("op_60_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_5_cast_fp16 = layer_norm(axes = input_5_axes_0, beta = norm0_1_bias_to_fp16, epsilon = var_60_to_fp16, gamma = norm0_1_weight_to_fp16, x = linear_0_cast_fp16)[name = tensor("input_5_cast_fp16")]; + tensor attn0_in_proj_weight_to_fp16 = const()[name = tensor("attn0_in_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(72000)))]; + tensor linear_1_bias_0_to_fp16 = const()[name = tensor("linear_1_bias_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(6363520)))]; + tensor linear_1_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = attn0_in_proj_weight_to_fp16, x = input_5_cast_fp16)[name = tensor("linear_1_cast_fp16")]; + tensor var_74 = const()[name = tensor("op_74"), val = tensor([1, 1, 3, 16, 64])]; + tensor qkv_1_cast_fp16 = reshape(shape = var_74, x = linear_1_cast_fp16)[name = tensor("qkv_1_cast_fp16")]; + tensor q_1_begin_0 = const()[name = tensor("q_1_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor q_1_end_0 = const()[name = tensor("q_1_end_0"), val = tensor([1, 1, 1, 16, 64])]; + tensor q_1_end_mask_0 = const()[name = tensor("q_1_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor q_1_squeeze_mask_0 = const()[name = tensor("q_1_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor q_1_cast_fp16 = slice_by_index(begin = q_1_begin_0, end = q_1_end_0, end_mask = q_1_end_mask_0, squeeze_mask = q_1_squeeze_mask_0, x = qkv_1_cast_fp16)[name = tensor("q_1_cast_fp16")]; + tensor k_1_begin_0 = const()[name = tensor("k_1_begin_0"), val = tensor([0, 0, 1, 0, 0])]; + tensor k_1_end_0 = const()[name = tensor("k_1_end_0"), val = tensor([1, 1, 2, 16, 64])]; + tensor k_1_end_mask_0 = const()[name = tensor("k_1_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor k_1_squeeze_mask_0 = const()[name = tensor("k_1_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor k_1_cast_fp16 = slice_by_index(begin = k_1_begin_0, end = k_1_end_0, end_mask = k_1_end_mask_0, squeeze_mask = k_1_squeeze_mask_0, x = qkv_1_cast_fp16)[name = tensor("k_1_cast_fp16")]; + tensor v_1_begin_0 = const()[name = tensor("v_1_begin_0"), val = tensor([0, 0, 2, 0, 0])]; + tensor v_1_end_0 = const()[name = tensor("v_1_end_0"), val = tensor([1, 1, 3, 16, 64])]; + tensor v_1_end_mask_0 = const()[name = tensor("v_1_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor v_1_squeeze_mask_0 = const()[name = tensor("v_1_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor v_1_cast_fp16 = slice_by_index(begin = v_1_begin_0, end = v_1_end_0, end_mask = v_1_end_mask_0, squeeze_mask = v_1_squeeze_mask_0, x = qkv_1_cast_fp16)[name = tensor("v_1_cast_fp16")]; + tensor var_140 = const()[name = tensor("op_140"), val = tensor([1, 1, 1, 1])]; + tensor position0_to_fp16_dtype_0 = const()[name = tensor("position0_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor position0_to_fp16 = cast(dtype = position0_to_fp16_dtype_0, x = position0)[name = tensor("cast_139")]; + tensor ts_5_cast_fp16 = reshape(shape = var_140, x = position0_to_fp16)[name = tensor("ts_5_cast_fp16")]; + tensor var_147 = const()[name = tensor("op_147"), val = tensor([1, 1, 16, 32, 2])]; + tensor q_complex_1_cast_fp16 = reshape(shape = var_147, x = q_1_cast_fp16)[name = tensor("q_complex_1_cast_fp16")]; + tensor var_154 = const()[name = tensor("op_154"), val = tensor([1, 1, 16, 32, 2])]; + tensor k_complex_1_cast_fp16 = reshape(shape = var_154, x = k_1_cast_fp16)[name = tensor("k_complex_1_cast_fp16")]; + tensor var_158_begin_0 = const()[name = tensor("op_158_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_158_end_0 = const()[name = tensor("op_158_end_0"), val = tensor([1, 1, 16, 32, 1])]; + tensor var_158_end_mask_0 = const()[name = tensor("op_158_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_158_squeeze_mask_0 = const()[name = tensor("op_158_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_158_cast_fp16 = slice_by_index(begin = var_158_begin_0, end = var_158_end_0, end_mask = var_158_end_mask_0, squeeze_mask = var_158_squeeze_mask_0, x = q_complex_1_cast_fp16)[name = tensor("op_158_cast_fp16")]; + tensor var_166_begin_0 = const()[name = tensor("op_166_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_166_end_0 = const()[name = tensor("op_166_end_0"), val = tensor([1, 1, 16, 32, 2])]; + tensor var_166_end_mask_0 = const()[name = tensor("op_166_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_166_squeeze_mask_0 = const()[name = tensor("op_166_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_166_cast_fp16 = slice_by_index(begin = var_166_begin_0, end = var_166_end_0, end_mask = var_166_end_mask_0, squeeze_mask = var_166_squeeze_mask_0, x = q_complex_1_cast_fp16)[name = tensor("op_166_cast_fp16")]; + tensor var_174_begin_0 = const()[name = tensor("op_174_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_174_end_0 = const()[name = tensor("op_174_end_0"), val = tensor([1, 1, 16, 32, 1])]; + tensor var_174_end_mask_0 = const()[name = tensor("op_174_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_174_squeeze_mask_0 = const()[name = tensor("op_174_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_174_cast_fp16 = slice_by_index(begin = var_174_begin_0, end = var_174_end_0, end_mask = var_174_end_mask_0, squeeze_mask = var_174_squeeze_mask_0, x = k_complex_1_cast_fp16)[name = tensor("op_174_cast_fp16")]; + tensor var_182_begin_0 = const()[name = tensor("op_182_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_182_end_0 = const()[name = tensor("op_182_end_0"), val = tensor([1, 1, 16, 32, 2])]; + tensor var_182_end_mask_0 = const()[name = tensor("op_182_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_182_squeeze_mask_0 = const()[name = tensor("op_182_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_182_cast_fp16 = slice_by_index(begin = var_182_begin_0, end = var_182_end_0, end_mask = var_182_end_mask_0, squeeze_mask = var_182_squeeze_mask_0, x = k_complex_1_cast_fp16)[name = tensor("op_182_cast_fp16")]; + tensor freqs_1_to_fp16 = const()[name = tensor("freqs_1_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(6369728)))]; + tensor var_188_cast_fp16 = mul(x = freqs_1_to_fp16, y = ts_5_cast_fp16)[name = tensor("op_188_cast_fp16")]; + tensor rotr_1_cast_fp16 = cos(x = var_188_cast_fp16)[name = tensor("rotr_1_cast_fp16")]; + tensor roti_1_cast_fp16 = sin(x = var_188_cast_fp16)[name = tensor("roti_1_cast_fp16")]; + tensor var_192_cast_fp16 = mul(x = var_158_cast_fp16, y = rotr_1_cast_fp16)[name = tensor("op_192_cast_fp16")]; + tensor var_193_cast_fp16 = mul(x = var_166_cast_fp16, y = roti_1_cast_fp16)[name = tensor("op_193_cast_fp16")]; + tensor qor_1_cast_fp16 = sub(x = var_192_cast_fp16, y = var_193_cast_fp16)[name = tensor("qor_1_cast_fp16")]; + tensor var_196_cast_fp16 = mul(x = var_158_cast_fp16, y = roti_1_cast_fp16)[name = tensor("op_196_cast_fp16")]; + tensor var_197_cast_fp16 = mul(x = var_166_cast_fp16, y = rotr_1_cast_fp16)[name = tensor("op_197_cast_fp16")]; + tensor qoi_1_cast_fp16 = add(x = var_196_cast_fp16, y = var_197_cast_fp16)[name = tensor("qoi_1_cast_fp16")]; + tensor var_200_cast_fp16 = mul(x = var_174_cast_fp16, y = rotr_1_cast_fp16)[name = tensor("op_200_cast_fp16")]; + tensor var_201_cast_fp16 = mul(x = var_182_cast_fp16, y = roti_1_cast_fp16)[name = tensor("op_201_cast_fp16")]; + tensor kor_1_cast_fp16 = sub(x = var_200_cast_fp16, y = var_201_cast_fp16)[name = tensor("kor_1_cast_fp16")]; + tensor var_204_cast_fp16 = mul(x = var_174_cast_fp16, y = roti_1_cast_fp16)[name = tensor("op_204_cast_fp16")]; + tensor var_205_cast_fp16 = mul(x = var_182_cast_fp16, y = rotr_1_cast_fp16)[name = tensor("op_205_cast_fp16")]; + tensor koi_1_cast_fp16 = add(x = var_204_cast_fp16, y = var_205_cast_fp16)[name = tensor("koi_1_cast_fp16")]; + tensor qo_1_axis_0 = const()[name = tensor("qo_1_axis_0"), val = tensor(-1)]; + tensor qo_1_cast_fp16 = stack(axis = qo_1_axis_0, values = (qor_1_cast_fp16, qoi_1_cast_fp16))[name = tensor("qo_1_cast_fp16")]; + tensor ko_1_axis_0 = const()[name = tensor("ko_1_axis_0"), val = tensor(-1)]; + tensor ko_1_cast_fp16 = stack(axis = ko_1_axis_0, values = (kor_1_cast_fp16, koi_1_cast_fp16))[name = tensor("ko_1_cast_fp16")]; + tensor var_238 = const()[name = tensor("op_238"), val = tensor([1, 1, 16, 64])]; + tensor q_3_cast_fp16 = reshape(shape = var_238, x = qo_1_cast_fp16)[name = tensor("q_3_cast_fp16")]; + tensor var_244 = const()[name = tensor("op_244"), val = tensor([1, 1, 16, 64])]; + tensor k_3_cast_fp16 = reshape(shape = var_244, x = ko_1_cast_fp16)[name = tensor("k_3_cast_fp16")]; + tensor _inversed_269_y_0_to_fp16 = const()[name = tensor("_inversed_269_y_0_to_fp16"), val = tensor(0x1p-9)]; + tensor _inversed_269_cast_fp16 = mul(x = ts_5_cast_fp16, y = _inversed_269_y_0_to_fp16)[name = tensor("_inversed_269_cast_fp16")]; + tensor var_270_cast_fp16 = floor(x = _inversed_269_cast_fp16)[name = tensor("op_270_cast_fp16")]; + tensor var_271_to_fp16 = const()[name = tensor("op_271_to_fp16"), val = tensor(0x1p+9)]; + tensor var_272_cast_fp16 = mul(x = var_270_cast_fp16, y = var_271_to_fp16)[name = tensor("op_272_cast_fp16")]; + tensor write_indices_float_3_cast_fp16 = sub(x = ts_5_cast_fp16, y = var_272_cast_fp16)[name = tensor("write_indices_float_3_cast_fp16")]; + tensor var_279_dtype_0 = const()[name = tensor("op_279_dtype_0"), val = tensor("int32")]; + tensor write_indices_1_reps_0 = const()[name = tensor("write_indices_1_reps_0"), val = tensor([1, 1, 16, 64])]; + tensor write_indices_float_3_cast_fp16_to_int32 = cast(dtype = var_279_dtype_0, x = write_indices_float_3_cast_fp16)[name = tensor("cast_138")]; + tensor write_indices_1 = tile(reps = write_indices_1_reps_0, x = write_indices_float_3_cast_fp16_to_int32)[name = tensor("write_indices_1")]; + tensor var_289_begin_0 = const()[name = tensor("op_289_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_289_end_0 = const()[name = tensor("op_289_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor var_289_end_mask_0 = const()[name = tensor("op_289_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_289_squeeze_mask_0 = const()[name = tensor("op_289_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor cache0_to_fp16_dtype_0 = const()[name = tensor("cache0_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor cache0_to_fp16 = cast(dtype = cache0_to_fp16_dtype_0, x = cache0)[name = tensor("cast_137")]; + tensor var_289_cast_fp16 = slice_by_index(begin = var_289_begin_0, end = var_289_end_0, end_mask = var_289_end_mask_0, squeeze_mask = var_289_squeeze_mask_0, x = cache0_to_fp16)[name = tensor("op_289_cast_fp16")]; + tensor var_291_axis_0 = const()[name = tensor("op_291_axis_0"), val = tensor(1)]; + tensor var_291_mode_0 = const()[name = tensor("op_291_mode_0"), val = tensor("update")]; + tensor var_291_validate_indices_0 = const()[name = tensor("op_291_validate_indices_0"), val = tensor(false)]; + tensor var_291_cast_fp16 = scatter_along_axis(axis = var_291_axis_0, data = var_289_cast_fp16, indices = write_indices_1, mode = var_291_mode_0, updates = k_3_cast_fp16, validate_indices = var_291_validate_indices_0)[name = tensor("op_291_cast_fp16")]; + tensor concat_2 = const()[name = tensor("concat_2"), val = tensor([0, 0, 0, 0, 0])]; + tensor concat_3 = const()[name = tensor("concat_3"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_1_internal_tensor_assign_1_stride_0 = const()[name = tensor("new_cache_1_internal_tensor_assign_1_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_1_internal_tensor_assign_1_begin_mask_0 = const()[name = tensor("new_cache_1_internal_tensor_assign_1_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_1_internal_tensor_assign_1_end_mask_0 = const()[name = tensor("new_cache_1_internal_tensor_assign_1_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_1_internal_tensor_assign_1_squeeze_mask_0 = const()[name = tensor("new_cache_1_internal_tensor_assign_1_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_12 = const()[name = tensor("shape_12"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_0 = const()[name = tensor("reduce_prod_0"), val = tensor(1048576)]; + tensor range_1d_0_start_0 = const()[name = tensor("range_1d_0_start_0"), val = tensor(0)]; + tensor range_1d_0_step_0 = const()[name = tensor("range_1d_0_step_0"), val = tensor(1)]; + tensor range_1d_0 = range_1d(end = reduce_prod_0, start = range_1d_0_start_0, step = range_1d_0_step_0)[name = tensor("range_1d_0")]; + tensor reshape_0 = reshape(shape = shape_12, x = range_1d_0)[name = tensor("reshape_0")]; + tensor slice_by_index_0 = slice_by_index(begin = concat_2, begin_mask = new_cache_1_internal_tensor_assign_1_begin_mask_0, end = concat_3, end_mask = new_cache_1_internal_tensor_assign_1_end_mask_0, squeeze_mask = new_cache_1_internal_tensor_assign_1_squeeze_mask_0, stride = new_cache_1_internal_tensor_assign_1_stride_0, x = reshape_0)[name = tensor("slice_by_index_0")]; + tensor reshape_1_shape_0 = const()[name = tensor("reshape_1_shape_0"), val = tensor([-1])]; + tensor reshape_1 = reshape(shape = reshape_1_shape_0, x = slice_by_index_0)[name = tensor("reshape_1")]; + tensor reshape_2_shape_0 = const()[name = tensor("reshape_2_shape_0"), val = tensor([-1])]; + tensor reshape_2_cast_fp16 = reshape(shape = reshape_2_shape_0, x = var_291_cast_fp16)[name = tensor("reshape_2_cast_fp16")]; + tensor reshape_3_shape_0 = const()[name = tensor("reshape_3_shape_0"), val = tensor([-1])]; + tensor reshape_3_cast_fp16 = reshape(shape = reshape_3_shape_0, x = cache0_to_fp16)[name = tensor("reshape_3_cast_fp16")]; + tensor scatter_0_mode_0 = const()[name = tensor("scatter_0_mode_0"), val = tensor("update")]; + tensor scatter_0_axis_0 = const()[name = tensor("scatter_0_axis_0"), val = tensor(0)]; + tensor scatter_0_validate_indices_0 = const()[name = tensor("scatter_0_validate_indices_0"), val = tensor(false)]; + tensor scatter_0_cast_fp16 = scatter(axis = scatter_0_axis_0, data = reshape_3_cast_fp16, indices = reshape_1, mode = scatter_0_mode_0, updates = reshape_2_cast_fp16, validate_indices = scatter_0_validate_indices_0)[name = tensor("scatter_0_cast_fp16")]; + tensor reshape_4_cast_fp16 = reshape(shape = shape_12, x = scatter_0_cast_fp16)[name = tensor("reshape_4_cast_fp16")]; + tensor var_299_begin_0 = const()[name = tensor("op_299_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor var_299_end_0 = const()[name = tensor("op_299_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor var_299_end_mask_0 = const()[name = tensor("op_299_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_299_squeeze_mask_0 = const()[name = tensor("op_299_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor var_299_cast_fp16 = slice_by_index(begin = var_299_begin_0, end = var_299_end_0, end_mask = var_299_end_mask_0, squeeze_mask = var_299_squeeze_mask_0, x = reshape_4_cast_fp16)[name = tensor("op_299_cast_fp16")]; + tensor var_301_axis_0 = const()[name = tensor("op_301_axis_0"), val = tensor(1)]; + tensor var_301_mode_0 = const()[name = tensor("op_301_mode_0"), val = tensor("update")]; + tensor var_301_validate_indices_0 = const()[name = tensor("op_301_validate_indices_0"), val = tensor(false)]; + tensor var_301_cast_fp16 = scatter_along_axis(axis = var_301_axis_0, data = var_299_cast_fp16, indices = write_indices_1, mode = var_301_mode_0, updates = v_1_cast_fp16, validate_indices = var_301_validate_indices_0)[name = tensor("op_301_cast_fp16")]; + tensor concat_4 = const()[name = tensor("concat_4"), val = tensor([1, 0, 0, 0, 0])]; + tensor concat_5 = const()[name = tensor("concat_5"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_1_internal_tensor_assign_2_stride_0 = const()[name = tensor("new_cache_1_internal_tensor_assign_2_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_1_internal_tensor_assign_2_begin_mask_0 = const()[name = tensor("new_cache_1_internal_tensor_assign_2_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_1_internal_tensor_assign_2_end_mask_0 = const()[name = tensor("new_cache_1_internal_tensor_assign_2_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_1_internal_tensor_assign_2_squeeze_mask_0 = const()[name = tensor("new_cache_1_internal_tensor_assign_2_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_13 = const()[name = tensor("shape_13"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_1 = const()[name = tensor("reduce_prod_1"), val = tensor(1048576)]; + tensor range_1d_1_start_0 = const()[name = tensor("range_1d_1_start_0"), val = tensor(0)]; + tensor range_1d_1_step_0 = const()[name = tensor("range_1d_1_step_0"), val = tensor(1)]; + tensor range_1d_1 = range_1d(end = reduce_prod_1, start = range_1d_1_start_0, step = range_1d_1_step_0)[name = tensor("range_1d_1")]; + tensor reshape_5 = reshape(shape = shape_13, x = range_1d_1)[name = tensor("reshape_5")]; + tensor slice_by_index_1 = slice_by_index(begin = concat_4, begin_mask = new_cache_1_internal_tensor_assign_2_begin_mask_0, end = concat_5, end_mask = new_cache_1_internal_tensor_assign_2_end_mask_0, squeeze_mask = new_cache_1_internal_tensor_assign_2_squeeze_mask_0, stride = new_cache_1_internal_tensor_assign_2_stride_0, x = reshape_5)[name = tensor("slice_by_index_1")]; + tensor reshape_6_shape_0 = const()[name = tensor("reshape_6_shape_0"), val = tensor([-1])]; + tensor reshape_6 = reshape(shape = reshape_6_shape_0, x = slice_by_index_1)[name = tensor("reshape_6")]; + tensor reshape_7_shape_0 = const()[name = tensor("reshape_7_shape_0"), val = tensor([-1])]; + tensor reshape_7_cast_fp16 = reshape(shape = reshape_7_shape_0, x = var_301_cast_fp16)[name = tensor("reshape_7_cast_fp16")]; + tensor reshape_8_shape_0 = const()[name = tensor("reshape_8_shape_0"), val = tensor([-1])]; + tensor reshape_8_cast_fp16 = reshape(shape = reshape_8_shape_0, x = reshape_4_cast_fp16)[name = tensor("reshape_8_cast_fp16")]; + tensor scatter_1_mode_0 = const()[name = tensor("scatter_1_mode_0"), val = tensor("update")]; + tensor scatter_1_axis_0 = const()[name = tensor("scatter_1_axis_0"), val = tensor(0)]; + tensor scatter_1_validate_indices_0 = const()[name = tensor("scatter_1_validate_indices_0"), val = tensor(false)]; + tensor scatter_1_cast_fp16 = scatter(axis = scatter_1_axis_0, data = reshape_8_cast_fp16, indices = reshape_6, mode = scatter_1_mode_0, updates = reshape_7_cast_fp16, validate_indices = scatter_1_validate_indices_0)[name = tensor("scatter_1_cast_fp16")]; + tensor reshape_9_cast_fp16 = reshape(shape = shape_13, x = scatter_1_cast_fp16)[name = tensor("reshape_9_cast_fp16")]; + tensor reshape_9_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("reshape_9_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor keys_1_begin_0 = const()[name = tensor("keys_1_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor keys_1_end_0 = const()[name = tensor("keys_1_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor keys_1_end_mask_0 = const()[name = tensor("keys_1_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor keys_1_squeeze_mask_0 = const()[name = tensor("keys_1_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor keys_1_cast_fp16 = slice_by_index(begin = keys_1_begin_0, end = keys_1_end_0, end_mask = keys_1_end_mask_0, squeeze_mask = keys_1_squeeze_mask_0, x = reshape_9_cast_fp16)[name = tensor("keys_1_cast_fp16")]; + tensor values_1_begin_0 = const()[name = tensor("values_1_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor values_1_end_0 = const()[name = tensor("values_1_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor values_1_end_mask_0 = const()[name = tensor("values_1_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor values_1_squeeze_mask_0 = const()[name = tensor("values_1_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor values_1_cast_fp16 = slice_by_index(begin = values_1_begin_0, end = values_1_end_0, end_mask = values_1_end_mask_0, squeeze_mask = values_1_squeeze_mask_0, x = reshape_9_cast_fp16)[name = tensor("values_1_cast_fp16")]; + tensor var_313_cast_fp16 = not_equal(x = keys_1_cast_fp16, y = keys_1_cast_fp16)[name = tensor("op_313_cast_fp16")]; + tensor var_319_to_fp16 = const()[name = tensor("op_319_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(6369856)))]; + tensor keys_3_cast_fp16 = select(a = var_319_to_fp16, b = keys_1_cast_fp16, cond = var_313_cast_fp16)[name = tensor("keys_3_cast_fp16")]; + tensor var_321_cast_fp16 = not_equal(x = values_1_cast_fp16, y = values_1_cast_fp16)[name = tensor("op_321_cast_fp16")]; + tensor values_3_cast_fp16 = select(a = var_319_to_fp16, b = values_1_cast_fp16, cond = var_321_cast_fp16)[name = tensor("values_3_cast_fp16")]; + tensor var_345 = const()[name = tensor("op_345"), val = tensor([0, 2, 1, 3])]; + tensor var_361 = const()[name = tensor("op_361"), val = tensor([1, 1, 1])]; + tensor var_362_cast_fp16 = reshape(shape = var_361, x = position0_to_fp16)[name = tensor("op_362_cast_fp16")]; + tensor var_382_to_fp16 = const()[name = tensor("op_382_to_fp16"), val = tensor(0x1p+0)]; + tensor valid_len_1_cast_fp16 = add(x = var_362_cast_fp16, y = var_382_to_fp16)[name = tensor("valid_len_1_cast_fp16")]; + tensor k_positions_1_promoted_to_fp16 = const()[name = tensor("k_positions_1_promoted_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(7418496)))]; + tensor valid_mask_1_cast_fp16 = less(x = k_positions_1_promoted_to_fp16, y = valid_len_1_cast_fp16)[name = tensor("valid_mask_1_cast_fp16")]; + tensor causal_mask_1_cast_fp16 = less_equal(x = k_positions_1_promoted_to_fp16, y = var_362_cast_fp16)[name = tensor("causal_mask_1_cast_fp16")]; + tensor attn_mask_1 = logical_and(x = valid_mask_1_cast_fp16, y = causal_mask_1_cast_fp16)[name = tensor("attn_mask_1")]; + tensor attn_mask_3_axes_0 = const()[name = tensor("attn_mask_3_axes_0"), val = tensor([1])]; + tensor attn_mask_3 = expand_dims(axes = attn_mask_3_axes_0, x = attn_mask_1)[name = tensor("attn_mask_3")]; + tensor var_400_transpose_x_0 = const()[name = tensor("op_400_transpose_x_0"), val = tensor(false)]; + tensor var_400_transpose_y_0 = const()[name = tensor("op_400_transpose_y_0"), val = tensor(false)]; + tensor transpose_18_perm_0 = const()[name = tensor("transpose_18_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_19_perm_0 = const()[name = tensor("transpose_19_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_19 = transpose(perm = transpose_19_perm_0, x = keys_3_cast_fp16)[name = tensor("transpose_51")]; + tensor transpose_18 = transpose(perm = transpose_18_perm_0, x = q_3_cast_fp16)[name = tensor("transpose_52")]; + tensor var_400_cast_fp16 = matmul(transpose_x = var_400_transpose_x_0, transpose_y = var_400_transpose_y_0, x = transpose_18, y = transpose_19)[name = tensor("op_400_cast_fp16")]; + tensor var_394_to_fp16 = const()[name = tensor("op_394_to_fp16"), val = tensor([0x1p-3])]; + tensor attn_weights_1_cast_fp16 = mul(x = var_400_cast_fp16, y = var_394_to_fp16)[name = tensor("attn_weights_1_cast_fp16")]; + tensor var_402 = logical_not(x = attn_mask_3)[name = tensor("op_402")]; + tensor var_403_to_fp16 = const()[name = tensor("op_403_to_fp16"), val = tensor(-inf)]; + tensor attn_weights_3_cast_fp16 = select(a = var_403_to_fp16, b = attn_weights_1_cast_fp16, cond = var_402)[name = tensor("attn_weights_3_cast_fp16")]; + tensor var_405 = const()[name = tensor("op_405"), val = tensor(-1)]; + tensor attn_weights_5_cast_fp16 = softmax(axis = var_405, x = attn_weights_3_cast_fp16)[name = tensor("attn_weights_5_cast_fp16")]; + tensor attn_output_1_transpose_x_0 = const()[name = tensor("attn_output_1_transpose_x_0"), val = tensor(false)]; + tensor attn_output_1_transpose_y_0 = const()[name = tensor("attn_output_1_transpose_y_0"), val = tensor(false)]; + tensor values_5_cast_fp16 = transpose(perm = var_345, x = values_3_cast_fp16)[name = tensor("transpose_53")]; + tensor attn_output_1_cast_fp16 = matmul(transpose_x = attn_output_1_transpose_x_0, transpose_y = attn_output_1_transpose_y_0, x = attn_weights_5_cast_fp16, y = values_5_cast_fp16)[name = tensor("attn_output_1_cast_fp16")]; + tensor var_413 = const()[name = tensor("op_413"), val = tensor([0, 2, 1, 3])]; + tensor var_418 = const()[name = tensor("op_418"), val = tensor([1, 1, 1024])]; + tensor var_414_cast_fp16 = transpose(perm = var_413, x = attn_output_1_cast_fp16)[name = tensor("transpose_50")]; + tensor input_7_cast_fp16 = reshape(shape = var_418, x = var_414_cast_fp16)[name = tensor("input_7_cast_fp16")]; + tensor attn0_out_proj_weight_to_fp16 = const()[name = tensor("attn0_out_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(7419584)))]; + tensor linear_2_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = attn0_out_proj_weight_to_fp16, x = input_7_cast_fp16)[name = tensor("linear_2_cast_fp16")]; + tensor var_424_to_fp16 = const()[name = tensor("op_424_to_fp16"), val = tensor(0x1p+0)]; + tensor var_425_cast_fp16 = add(x = position0_to_fp16, y = var_424_to_fp16)[name = tensor("op_425_cast_fp16")]; + tensor var_425_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_425_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor input_9_cast_fp16 = add(x = linear_0_cast_fp16, y = linear_2_cast_fp16)[name = tensor("input_9_cast_fp16")]; + tensor input_11_axes_0 = const()[name = tensor("input_11_axes_0"), val = tensor([-1])]; + tensor norm0_2_weight_to_fp16 = const()[name = tensor("norm0_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9516800)))]; + tensor norm0_2_bias_to_fp16 = const()[name = tensor("norm0_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9518912)))]; + tensor var_429_to_fp16 = const()[name = tensor("op_429_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_11_cast_fp16 = layer_norm(axes = input_11_axes_0, beta = norm0_2_bias_to_fp16, epsilon = var_429_to_fp16, gamma = norm0_2_weight_to_fp16, x = input_9_cast_fp16)[name = tensor("input_11_cast_fp16")]; + tensor linear0_1_weight_to_fp16 = const()[name = tensor("linear0_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9521024)))]; + tensor linear_3_bias_0_to_fp16 = const()[name = tensor("linear_3_bias_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(17909696)))]; + tensor linear_3_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = linear0_1_weight_to_fp16, x = input_11_cast_fp16)[name = tensor("linear_3_cast_fp16")]; + tensor input_13_mode_0 = const()[name = tensor("input_13_mode_0"), val = tensor("EXACT")]; + tensor input_13_cast_fp16 = gelu(mode = input_13_mode_0, x = linear_3_cast_fp16)[name = tensor("input_13_cast_fp16")]; + tensor linear0_2_weight_to_fp16 = const()[name = tensor("linear0_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(17917952)))]; + tensor linear_4_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = linear0_2_weight_to_fp16, x = input_13_cast_fp16)[name = tensor("linear_4_cast_fp16")]; + tensor input_15_cast_fp16 = add(x = input_9_cast_fp16, y = linear_4_cast_fp16)[name = tensor("input_15_cast_fp16")]; + tensor input_17_axes_0 = const()[name = tensor("input_17_axes_0"), val = tensor([-1])]; + tensor norm1_1_weight_to_fp16 = const()[name = tensor("norm1_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(26306624)))]; + tensor norm1_1_bias_to_fp16 = const()[name = tensor("norm1_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(26308736)))]; + tensor var_446_to_fp16 = const()[name = tensor("op_446_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_17_cast_fp16 = layer_norm(axes = input_17_axes_0, beta = norm1_1_bias_to_fp16, epsilon = var_446_to_fp16, gamma = norm1_1_weight_to_fp16, x = input_15_cast_fp16)[name = tensor("input_17_cast_fp16")]; + tensor attn1_in_proj_weight_to_fp16 = const()[name = tensor("attn1_in_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(26310848)))]; + tensor linear_5_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = attn1_in_proj_weight_to_fp16, x = input_17_cast_fp16)[name = tensor("linear_5_cast_fp16")]; + tensor var_460 = const()[name = tensor("op_460"), val = tensor([1, 1, 3, 16, 64])]; + tensor qkv_3_cast_fp16 = reshape(shape = var_460, x = linear_5_cast_fp16)[name = tensor("qkv_3_cast_fp16")]; + tensor q_7_begin_0 = const()[name = tensor("q_7_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor q_7_end_0 = const()[name = tensor("q_7_end_0"), val = tensor([1, 1, 1, 16, 64])]; + tensor q_7_end_mask_0 = const()[name = tensor("q_7_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor q_7_squeeze_mask_0 = const()[name = tensor("q_7_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor q_7_cast_fp16 = slice_by_index(begin = q_7_begin_0, end = q_7_end_0, end_mask = q_7_end_mask_0, squeeze_mask = q_7_squeeze_mask_0, x = qkv_3_cast_fp16)[name = tensor("q_7_cast_fp16")]; + tensor k_5_begin_0 = const()[name = tensor("k_5_begin_0"), val = tensor([0, 0, 1, 0, 0])]; + tensor k_5_end_0 = const()[name = tensor("k_5_end_0"), val = tensor([1, 1, 2, 16, 64])]; + tensor k_5_end_mask_0 = const()[name = tensor("k_5_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor k_5_squeeze_mask_0 = const()[name = tensor("k_5_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor k_5_cast_fp16 = slice_by_index(begin = k_5_begin_0, end = k_5_end_0, end_mask = k_5_end_mask_0, squeeze_mask = k_5_squeeze_mask_0, x = qkv_3_cast_fp16)[name = tensor("k_5_cast_fp16")]; + tensor v_3_begin_0 = const()[name = tensor("v_3_begin_0"), val = tensor([0, 0, 2, 0, 0])]; + tensor v_3_end_0 = const()[name = tensor("v_3_end_0"), val = tensor([1, 1, 3, 16, 64])]; + tensor v_3_end_mask_0 = const()[name = tensor("v_3_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor v_3_squeeze_mask_0 = const()[name = tensor("v_3_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor v_3_cast_fp16 = slice_by_index(begin = v_3_begin_0, end = v_3_end_0, end_mask = v_3_end_mask_0, squeeze_mask = v_3_squeeze_mask_0, x = qkv_3_cast_fp16)[name = tensor("v_3_cast_fp16")]; + tensor var_526 = const()[name = tensor("op_526"), val = tensor([1, 1, 1, 1])]; + tensor position1_to_fp16_dtype_0 = const()[name = tensor("position1_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor position1_to_fp16 = cast(dtype = position1_to_fp16_dtype_0, x = position1)[name = tensor("cast_134")]; + tensor ts_11_cast_fp16 = reshape(shape = var_526, x = position1_to_fp16)[name = tensor("ts_11_cast_fp16")]; + tensor var_533 = const()[name = tensor("op_533"), val = tensor([1, 1, 16, 32, 2])]; + tensor q_complex_3_cast_fp16 = reshape(shape = var_533, x = q_7_cast_fp16)[name = tensor("q_complex_3_cast_fp16")]; + tensor var_540 = const()[name = tensor("op_540"), val = tensor([1, 1, 16, 32, 2])]; + tensor k_complex_3_cast_fp16 = reshape(shape = var_540, x = k_5_cast_fp16)[name = tensor("k_complex_3_cast_fp16")]; + tensor var_544_begin_0 = const()[name = tensor("op_544_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_544_end_0 = const()[name = tensor("op_544_end_0"), val = tensor([1, 1, 16, 32, 1])]; + tensor var_544_end_mask_0 = const()[name = tensor("op_544_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_544_squeeze_mask_0 = const()[name = tensor("op_544_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_544_cast_fp16 = slice_by_index(begin = var_544_begin_0, end = var_544_end_0, end_mask = var_544_end_mask_0, squeeze_mask = var_544_squeeze_mask_0, x = q_complex_3_cast_fp16)[name = tensor("op_544_cast_fp16")]; + tensor var_552_begin_0 = const()[name = tensor("op_552_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_552_end_0 = const()[name = tensor("op_552_end_0"), val = tensor([1, 1, 16, 32, 2])]; + tensor var_552_end_mask_0 = const()[name = tensor("op_552_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_552_squeeze_mask_0 = const()[name = tensor("op_552_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_552_cast_fp16 = slice_by_index(begin = var_552_begin_0, end = var_552_end_0, end_mask = var_552_end_mask_0, squeeze_mask = var_552_squeeze_mask_0, x = q_complex_3_cast_fp16)[name = tensor("op_552_cast_fp16")]; + tensor var_560_begin_0 = const()[name = tensor("op_560_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_560_end_0 = const()[name = tensor("op_560_end_0"), val = tensor([1, 1, 16, 32, 1])]; + tensor var_560_end_mask_0 = const()[name = tensor("op_560_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_560_squeeze_mask_0 = const()[name = tensor("op_560_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_560_cast_fp16 = slice_by_index(begin = var_560_begin_0, end = var_560_end_0, end_mask = var_560_end_mask_0, squeeze_mask = var_560_squeeze_mask_0, x = k_complex_3_cast_fp16)[name = tensor("op_560_cast_fp16")]; + tensor var_568_begin_0 = const()[name = tensor("op_568_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_568_end_0 = const()[name = tensor("op_568_end_0"), val = tensor([1, 1, 16, 32, 2])]; + tensor var_568_end_mask_0 = const()[name = tensor("op_568_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_568_squeeze_mask_0 = const()[name = tensor("op_568_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_568_cast_fp16 = slice_by_index(begin = var_568_begin_0, end = var_568_end_0, end_mask = var_568_end_mask_0, squeeze_mask = var_568_squeeze_mask_0, x = k_complex_3_cast_fp16)[name = tensor("op_568_cast_fp16")]; + tensor freqs_3_to_fp16 = const()[name = tensor("freqs_3_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(32602368)))]; + tensor var_574_cast_fp16 = mul(x = freqs_3_to_fp16, y = ts_11_cast_fp16)[name = tensor("op_574_cast_fp16")]; + tensor rotr_3_cast_fp16 = cos(x = var_574_cast_fp16)[name = tensor("rotr_3_cast_fp16")]; + tensor roti_3_cast_fp16 = sin(x = var_574_cast_fp16)[name = tensor("roti_3_cast_fp16")]; + tensor var_578_cast_fp16 = mul(x = var_544_cast_fp16, y = rotr_3_cast_fp16)[name = tensor("op_578_cast_fp16")]; + tensor var_579_cast_fp16 = mul(x = var_552_cast_fp16, y = roti_3_cast_fp16)[name = tensor("op_579_cast_fp16")]; + tensor qor_5_cast_fp16 = sub(x = var_578_cast_fp16, y = var_579_cast_fp16)[name = tensor("qor_5_cast_fp16")]; + tensor var_582_cast_fp16 = mul(x = var_544_cast_fp16, y = roti_3_cast_fp16)[name = tensor("op_582_cast_fp16")]; + tensor var_583_cast_fp16 = mul(x = var_552_cast_fp16, y = rotr_3_cast_fp16)[name = tensor("op_583_cast_fp16")]; + tensor qoi_5_cast_fp16 = add(x = var_582_cast_fp16, y = var_583_cast_fp16)[name = tensor("qoi_5_cast_fp16")]; + tensor var_586_cast_fp16 = mul(x = var_560_cast_fp16, y = rotr_3_cast_fp16)[name = tensor("op_586_cast_fp16")]; + tensor var_587_cast_fp16 = mul(x = var_568_cast_fp16, y = roti_3_cast_fp16)[name = tensor("op_587_cast_fp16")]; + tensor kor_5_cast_fp16 = sub(x = var_586_cast_fp16, y = var_587_cast_fp16)[name = tensor("kor_5_cast_fp16")]; + tensor var_590_cast_fp16 = mul(x = var_560_cast_fp16, y = roti_3_cast_fp16)[name = tensor("op_590_cast_fp16")]; + tensor var_591_cast_fp16 = mul(x = var_568_cast_fp16, y = rotr_3_cast_fp16)[name = tensor("op_591_cast_fp16")]; + tensor koi_5_cast_fp16 = add(x = var_590_cast_fp16, y = var_591_cast_fp16)[name = tensor("koi_5_cast_fp16")]; + tensor qo_3_axis_0 = const()[name = tensor("qo_3_axis_0"), val = tensor(-1)]; + tensor qo_3_cast_fp16 = stack(axis = qo_3_axis_0, values = (qor_5_cast_fp16, qoi_5_cast_fp16))[name = tensor("qo_3_cast_fp16")]; + tensor ko_3_axis_0 = const()[name = tensor("ko_3_axis_0"), val = tensor(-1)]; + tensor ko_3_cast_fp16 = stack(axis = ko_3_axis_0, values = (kor_5_cast_fp16, koi_5_cast_fp16))[name = tensor("ko_3_cast_fp16")]; + tensor var_624 = const()[name = tensor("op_624"), val = tensor([1, 1, 16, 64])]; + tensor q_9_cast_fp16 = reshape(shape = var_624, x = qo_3_cast_fp16)[name = tensor("q_9_cast_fp16")]; + tensor var_630 = const()[name = tensor("op_630"), val = tensor([1, 1, 16, 64])]; + tensor k_7_cast_fp16 = reshape(shape = var_630, x = ko_3_cast_fp16)[name = tensor("k_7_cast_fp16")]; + tensor _inversed_655_y_0_to_fp16 = const()[name = tensor("_inversed_655_y_0_to_fp16"), val = tensor(0x1p-9)]; + tensor _inversed_655_cast_fp16 = mul(x = ts_11_cast_fp16, y = _inversed_655_y_0_to_fp16)[name = tensor("_inversed_655_cast_fp16")]; + tensor var_656_cast_fp16 = floor(x = _inversed_655_cast_fp16)[name = tensor("op_656_cast_fp16")]; + tensor var_657_to_fp16 = const()[name = tensor("op_657_to_fp16"), val = tensor(0x1p+9)]; + tensor var_658_cast_fp16 = mul(x = var_656_cast_fp16, y = var_657_to_fp16)[name = tensor("op_658_cast_fp16")]; + tensor write_indices_float_7_cast_fp16 = sub(x = ts_11_cast_fp16, y = var_658_cast_fp16)[name = tensor("write_indices_float_7_cast_fp16")]; + tensor var_665_dtype_0 = const()[name = tensor("op_665_dtype_0"), val = tensor("int32")]; + tensor write_indices_3_reps_0 = const()[name = tensor("write_indices_3_reps_0"), val = tensor([1, 1, 16, 64])]; + tensor write_indices_float_7_cast_fp16_to_int32 = cast(dtype = var_665_dtype_0, x = write_indices_float_7_cast_fp16)[name = tensor("cast_133")]; + tensor write_indices_3 = tile(reps = write_indices_3_reps_0, x = write_indices_float_7_cast_fp16_to_int32)[name = tensor("write_indices_3")]; + tensor var_675_begin_0 = const()[name = tensor("op_675_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_675_end_0 = const()[name = tensor("op_675_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor var_675_end_mask_0 = const()[name = tensor("op_675_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_675_squeeze_mask_0 = const()[name = tensor("op_675_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor cache1_to_fp16_dtype_0 = const()[name = tensor("cache1_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor cache1_to_fp16 = cast(dtype = cache1_to_fp16_dtype_0, x = cache1)[name = tensor("cast_132")]; + tensor var_675_cast_fp16 = slice_by_index(begin = var_675_begin_0, end = var_675_end_0, end_mask = var_675_end_mask_0, squeeze_mask = var_675_squeeze_mask_0, x = cache1_to_fp16)[name = tensor("op_675_cast_fp16")]; + tensor var_677_axis_0 = const()[name = tensor("op_677_axis_0"), val = tensor(1)]; + tensor var_677_mode_0 = const()[name = tensor("op_677_mode_0"), val = tensor("update")]; + tensor var_677_validate_indices_0 = const()[name = tensor("op_677_validate_indices_0"), val = tensor(false)]; + tensor var_677_cast_fp16 = scatter_along_axis(axis = var_677_axis_0, data = var_675_cast_fp16, indices = write_indices_3, mode = var_677_mode_0, updates = k_7_cast_fp16, validate_indices = var_677_validate_indices_0)[name = tensor("op_677_cast_fp16")]; + tensor concat_9 = const()[name = tensor("concat_9"), val = tensor([0, 0, 0, 0, 0])]; + tensor concat_10 = const()[name = tensor("concat_10"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_3_internal_tensor_assign_1_stride_0 = const()[name = tensor("new_cache_3_internal_tensor_assign_1_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_3_internal_tensor_assign_1_begin_mask_0 = const()[name = tensor("new_cache_3_internal_tensor_assign_1_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_3_internal_tensor_assign_1_end_mask_0 = const()[name = tensor("new_cache_3_internal_tensor_assign_1_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_3_internal_tensor_assign_1_squeeze_mask_0 = const()[name = tensor("new_cache_3_internal_tensor_assign_1_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_14 = const()[name = tensor("shape_14"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_2 = const()[name = tensor("reduce_prod_2"), val = tensor(1048576)]; + tensor range_1d_2_start_0 = const()[name = tensor("range_1d_2_start_0"), val = tensor(0)]; + tensor range_1d_2_step_0 = const()[name = tensor("range_1d_2_step_0"), val = tensor(1)]; + tensor range_1d_2 = range_1d(end = reduce_prod_2, start = range_1d_2_start_0, step = range_1d_2_step_0)[name = tensor("range_1d_2")]; + tensor reshape_10 = reshape(shape = shape_14, x = range_1d_2)[name = tensor("reshape_10")]; + tensor slice_by_index_2 = slice_by_index(begin = concat_9, begin_mask = new_cache_3_internal_tensor_assign_1_begin_mask_0, end = concat_10, end_mask = new_cache_3_internal_tensor_assign_1_end_mask_0, squeeze_mask = new_cache_3_internal_tensor_assign_1_squeeze_mask_0, stride = new_cache_3_internal_tensor_assign_1_stride_0, x = reshape_10)[name = tensor("slice_by_index_2")]; + tensor reshape_11_shape_0 = const()[name = tensor("reshape_11_shape_0"), val = tensor([-1])]; + tensor reshape_11 = reshape(shape = reshape_11_shape_0, x = slice_by_index_2)[name = tensor("reshape_11")]; + tensor reshape_12_shape_0 = const()[name = tensor("reshape_12_shape_0"), val = tensor([-1])]; + tensor reshape_12_cast_fp16 = reshape(shape = reshape_12_shape_0, x = var_677_cast_fp16)[name = tensor("reshape_12_cast_fp16")]; + tensor reshape_13_shape_0 = const()[name = tensor("reshape_13_shape_0"), val = tensor([-1])]; + tensor reshape_13_cast_fp16 = reshape(shape = reshape_13_shape_0, x = cache1_to_fp16)[name = tensor("reshape_13_cast_fp16")]; + tensor scatter_2_mode_0 = const()[name = tensor("scatter_2_mode_0"), val = tensor("update")]; + tensor scatter_2_axis_0 = const()[name = tensor("scatter_2_axis_0"), val = tensor(0)]; + tensor scatter_2_validate_indices_0 = const()[name = tensor("scatter_2_validate_indices_0"), val = tensor(false)]; + tensor scatter_2_cast_fp16 = scatter(axis = scatter_2_axis_0, data = reshape_13_cast_fp16, indices = reshape_11, mode = scatter_2_mode_0, updates = reshape_12_cast_fp16, validate_indices = scatter_2_validate_indices_0)[name = tensor("scatter_2_cast_fp16")]; + tensor reshape_14_cast_fp16 = reshape(shape = shape_14, x = scatter_2_cast_fp16)[name = tensor("reshape_14_cast_fp16")]; + tensor var_685_begin_0 = const()[name = tensor("op_685_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor var_685_end_0 = const()[name = tensor("op_685_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor var_685_end_mask_0 = const()[name = tensor("op_685_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_685_squeeze_mask_0 = const()[name = tensor("op_685_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor var_685_cast_fp16 = slice_by_index(begin = var_685_begin_0, end = var_685_end_0, end_mask = var_685_end_mask_0, squeeze_mask = var_685_squeeze_mask_0, x = reshape_14_cast_fp16)[name = tensor("op_685_cast_fp16")]; + tensor var_687_axis_0 = const()[name = tensor("op_687_axis_0"), val = tensor(1)]; + tensor var_687_mode_0 = const()[name = tensor("op_687_mode_0"), val = tensor("update")]; + tensor var_687_validate_indices_0 = const()[name = tensor("op_687_validate_indices_0"), val = tensor(false)]; + tensor var_687_cast_fp16 = scatter_along_axis(axis = var_687_axis_0, data = var_685_cast_fp16, indices = write_indices_3, mode = var_687_mode_0, updates = v_3_cast_fp16, validate_indices = var_687_validate_indices_0)[name = tensor("op_687_cast_fp16")]; + tensor concat_11 = const()[name = tensor("concat_11"), val = tensor([1, 0, 0, 0, 0])]; + tensor concat_12 = const()[name = tensor("concat_12"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_3_internal_tensor_assign_2_stride_0 = const()[name = tensor("new_cache_3_internal_tensor_assign_2_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_3_internal_tensor_assign_2_begin_mask_0 = const()[name = tensor("new_cache_3_internal_tensor_assign_2_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_3_internal_tensor_assign_2_end_mask_0 = const()[name = tensor("new_cache_3_internal_tensor_assign_2_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_3_internal_tensor_assign_2_squeeze_mask_0 = const()[name = tensor("new_cache_3_internal_tensor_assign_2_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_15 = const()[name = tensor("shape_15"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_3 = const()[name = tensor("reduce_prod_3"), val = tensor(1048576)]; + tensor range_1d_3_start_0 = const()[name = tensor("range_1d_3_start_0"), val = tensor(0)]; + tensor range_1d_3_step_0 = const()[name = tensor("range_1d_3_step_0"), val = tensor(1)]; + tensor range_1d_3 = range_1d(end = reduce_prod_3, start = range_1d_3_start_0, step = range_1d_3_step_0)[name = tensor("range_1d_3")]; + tensor reshape_15 = reshape(shape = shape_15, x = range_1d_3)[name = tensor("reshape_15")]; + tensor slice_by_index_3 = slice_by_index(begin = concat_11, begin_mask = new_cache_3_internal_tensor_assign_2_begin_mask_0, end = concat_12, end_mask = new_cache_3_internal_tensor_assign_2_end_mask_0, squeeze_mask = new_cache_3_internal_tensor_assign_2_squeeze_mask_0, stride = new_cache_3_internal_tensor_assign_2_stride_0, x = reshape_15)[name = tensor("slice_by_index_3")]; + tensor reshape_16_shape_0 = const()[name = tensor("reshape_16_shape_0"), val = tensor([-1])]; + tensor reshape_16 = reshape(shape = reshape_16_shape_0, x = slice_by_index_3)[name = tensor("reshape_16")]; + tensor reshape_17_shape_0 = const()[name = tensor("reshape_17_shape_0"), val = tensor([-1])]; + tensor reshape_17_cast_fp16 = reshape(shape = reshape_17_shape_0, x = var_687_cast_fp16)[name = tensor("reshape_17_cast_fp16")]; + tensor reshape_18_shape_0 = const()[name = tensor("reshape_18_shape_0"), val = tensor([-1])]; + tensor reshape_18_cast_fp16 = reshape(shape = reshape_18_shape_0, x = reshape_14_cast_fp16)[name = tensor("reshape_18_cast_fp16")]; + tensor scatter_3_mode_0 = const()[name = tensor("scatter_3_mode_0"), val = tensor("update")]; + tensor scatter_3_axis_0 = const()[name = tensor("scatter_3_axis_0"), val = tensor(0)]; + tensor scatter_3_validate_indices_0 = const()[name = tensor("scatter_3_validate_indices_0"), val = tensor(false)]; + tensor scatter_3_cast_fp16 = scatter(axis = scatter_3_axis_0, data = reshape_18_cast_fp16, indices = reshape_16, mode = scatter_3_mode_0, updates = reshape_17_cast_fp16, validate_indices = scatter_3_validate_indices_0)[name = tensor("scatter_3_cast_fp16")]; + tensor reshape_19_cast_fp16 = reshape(shape = shape_15, x = scatter_3_cast_fp16)[name = tensor("reshape_19_cast_fp16")]; + tensor reshape_19_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("reshape_19_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor keys_7_begin_0 = const()[name = tensor("keys_7_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor keys_7_end_0 = const()[name = tensor("keys_7_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor keys_7_end_mask_0 = const()[name = tensor("keys_7_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor keys_7_squeeze_mask_0 = const()[name = tensor("keys_7_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor keys_7_cast_fp16 = slice_by_index(begin = keys_7_begin_0, end = keys_7_end_0, end_mask = keys_7_end_mask_0, squeeze_mask = keys_7_squeeze_mask_0, x = reshape_19_cast_fp16)[name = tensor("keys_7_cast_fp16")]; + tensor values_7_begin_0 = const()[name = tensor("values_7_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor values_7_end_0 = const()[name = tensor("values_7_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor values_7_end_mask_0 = const()[name = tensor("values_7_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor values_7_squeeze_mask_0 = const()[name = tensor("values_7_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor values_7_cast_fp16 = slice_by_index(begin = values_7_begin_0, end = values_7_end_0, end_mask = values_7_end_mask_0, squeeze_mask = values_7_squeeze_mask_0, x = reshape_19_cast_fp16)[name = tensor("values_7_cast_fp16")]; + tensor var_699_cast_fp16 = not_equal(x = keys_7_cast_fp16, y = keys_7_cast_fp16)[name = tensor("op_699_cast_fp16")]; + tensor keys_9_cast_fp16 = select(a = var_319_to_fp16, b = keys_7_cast_fp16, cond = var_699_cast_fp16)[name = tensor("keys_9_cast_fp16")]; + tensor var_707_cast_fp16 = not_equal(x = values_7_cast_fp16, y = values_7_cast_fp16)[name = tensor("op_707_cast_fp16")]; + tensor values_9_cast_fp16 = select(a = var_319_to_fp16, b = values_7_cast_fp16, cond = var_707_cast_fp16)[name = tensor("values_9_cast_fp16")]; + tensor var_731 = const()[name = tensor("op_731"), val = tensor([0, 2, 1, 3])]; + tensor var_747 = const()[name = tensor("op_747"), val = tensor([1, 1, 1])]; + tensor var_748_cast_fp16 = reshape(shape = var_747, x = position1_to_fp16)[name = tensor("op_748_cast_fp16")]; + tensor var_768_to_fp16 = const()[name = tensor("op_768_to_fp16"), val = tensor(0x1p+0)]; + tensor valid_len_3_cast_fp16 = add(x = var_748_cast_fp16, y = var_768_to_fp16)[name = tensor("valid_len_3_cast_fp16")]; + tensor valid_mask_3_cast_fp16 = less(x = k_positions_1_promoted_to_fp16, y = valid_len_3_cast_fp16)[name = tensor("valid_mask_3_cast_fp16")]; + tensor causal_mask_3_cast_fp16 = less_equal(x = k_positions_1_promoted_to_fp16, y = var_748_cast_fp16)[name = tensor("causal_mask_3_cast_fp16")]; + tensor attn_mask_5 = logical_and(x = valid_mask_3_cast_fp16, y = causal_mask_3_cast_fp16)[name = tensor("attn_mask_5")]; + tensor attn_mask_7_axes_0 = const()[name = tensor("attn_mask_7_axes_0"), val = tensor([1])]; + tensor attn_mask_7 = expand_dims(axes = attn_mask_7_axes_0, x = attn_mask_5)[name = tensor("attn_mask_7")]; + tensor var_786_transpose_x_0 = const()[name = tensor("op_786_transpose_x_0"), val = tensor(false)]; + tensor var_786_transpose_y_0 = const()[name = tensor("op_786_transpose_y_0"), val = tensor(false)]; + tensor transpose_20_perm_0 = const()[name = tensor("transpose_20_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_21_perm_0 = const()[name = tensor("transpose_21_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_21 = transpose(perm = transpose_21_perm_0, x = keys_9_cast_fp16)[name = tensor("transpose_47")]; + tensor transpose_20 = transpose(perm = transpose_20_perm_0, x = q_9_cast_fp16)[name = tensor("transpose_48")]; + tensor var_786_cast_fp16 = matmul(transpose_x = var_786_transpose_x_0, transpose_y = var_786_transpose_y_0, x = transpose_20, y = transpose_21)[name = tensor("op_786_cast_fp16")]; + tensor var_780_to_fp16 = const()[name = tensor("op_780_to_fp16"), val = tensor([0x1p-3])]; + tensor attn_weights_7_cast_fp16 = mul(x = var_786_cast_fp16, y = var_780_to_fp16)[name = tensor("attn_weights_7_cast_fp16")]; + tensor var_788 = logical_not(x = attn_mask_7)[name = tensor("op_788")]; + tensor var_789_to_fp16 = const()[name = tensor("op_789_to_fp16"), val = tensor(-inf)]; + tensor attn_weights_9_cast_fp16 = select(a = var_789_to_fp16, b = attn_weights_7_cast_fp16, cond = var_788)[name = tensor("attn_weights_9_cast_fp16")]; + tensor var_791 = const()[name = tensor("op_791"), val = tensor(-1)]; + tensor attn_weights_11_cast_fp16 = softmax(axis = var_791, x = attn_weights_9_cast_fp16)[name = tensor("attn_weights_11_cast_fp16")]; + tensor attn_output_3_transpose_x_0 = const()[name = tensor("attn_output_3_transpose_x_0"), val = tensor(false)]; + tensor attn_output_3_transpose_y_0 = const()[name = tensor("attn_output_3_transpose_y_0"), val = tensor(false)]; + tensor values_11_cast_fp16 = transpose(perm = var_731, x = values_9_cast_fp16)[name = tensor("transpose_49")]; + tensor attn_output_3_cast_fp16 = matmul(transpose_x = attn_output_3_transpose_x_0, transpose_y = attn_output_3_transpose_y_0, x = attn_weights_11_cast_fp16, y = values_11_cast_fp16)[name = tensor("attn_output_3_cast_fp16")]; + tensor var_799 = const()[name = tensor("op_799"), val = tensor([0, 2, 1, 3])]; + tensor var_804 = const()[name = tensor("op_804"), val = tensor([1, 1, 1024])]; + tensor var_800_cast_fp16 = transpose(perm = var_799, x = attn_output_3_cast_fp16)[name = tensor("transpose_46")]; + tensor input_19_cast_fp16 = reshape(shape = var_804, x = var_800_cast_fp16)[name = tensor("input_19_cast_fp16")]; + tensor attn1_out_proj_weight_to_fp16 = const()[name = tensor("attn1_out_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(32602496)))]; + tensor linear_6_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = attn1_out_proj_weight_to_fp16, x = input_19_cast_fp16)[name = tensor("linear_6_cast_fp16")]; + tensor var_810_to_fp16 = const()[name = tensor("op_810_to_fp16"), val = tensor(0x1p+0)]; + tensor var_811_cast_fp16 = add(x = position1_to_fp16, y = var_810_to_fp16)[name = tensor("op_811_cast_fp16")]; + tensor var_811_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_811_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor input_21_cast_fp16 = add(x = input_15_cast_fp16, y = linear_6_cast_fp16)[name = tensor("input_21_cast_fp16")]; + tensor input_23_axes_0 = const()[name = tensor("input_23_axes_0"), val = tensor([-1])]; + tensor norm1_2_weight_to_fp16 = const()[name = tensor("norm1_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(34699712)))]; + tensor norm1_2_bias_to_fp16 = const()[name = tensor("norm1_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(34701824)))]; + tensor var_815_to_fp16 = const()[name = tensor("op_815_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_23_cast_fp16 = layer_norm(axes = input_23_axes_0, beta = norm1_2_bias_to_fp16, epsilon = var_815_to_fp16, gamma = norm1_2_weight_to_fp16, x = input_21_cast_fp16)[name = tensor("input_23_cast_fp16")]; + tensor linear1_1_weight_to_fp16 = const()[name = tensor("linear1_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(34703936)))]; + tensor linear_7_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = linear1_1_weight_to_fp16, x = input_23_cast_fp16)[name = tensor("linear_7_cast_fp16")]; + tensor input_25_mode_0 = const()[name = tensor("input_25_mode_0"), val = tensor("EXACT")]; + tensor input_25_cast_fp16 = gelu(mode = input_25_mode_0, x = linear_7_cast_fp16)[name = tensor("input_25_cast_fp16")]; + tensor linear1_2_weight_to_fp16 = const()[name = tensor("linear1_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(43092608)))]; + tensor linear_8_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = linear1_2_weight_to_fp16, x = input_25_cast_fp16)[name = tensor("linear_8_cast_fp16")]; + tensor input_27_cast_fp16 = add(x = input_21_cast_fp16, y = linear_8_cast_fp16)[name = tensor("input_27_cast_fp16")]; + tensor input_29_axes_0 = const()[name = tensor("input_29_axes_0"), val = tensor([-1])]; + tensor norm2_1_weight_to_fp16 = const()[name = tensor("norm2_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(51481280)))]; + tensor norm2_1_bias_to_fp16 = const()[name = tensor("norm2_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(51483392)))]; + tensor var_832_to_fp16 = const()[name = tensor("op_832_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_29_cast_fp16 = layer_norm(axes = input_29_axes_0, beta = norm2_1_bias_to_fp16, epsilon = var_832_to_fp16, gamma = norm2_1_weight_to_fp16, x = input_27_cast_fp16)[name = tensor("input_29_cast_fp16")]; + tensor attn2_in_proj_weight_to_fp16 = const()[name = tensor("attn2_in_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(51485504)))]; + tensor linear_9_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = attn2_in_proj_weight_to_fp16, x = input_29_cast_fp16)[name = tensor("linear_9_cast_fp16")]; + tensor var_846 = const()[name = tensor("op_846"), val = tensor([1, 1, 3, 16, 64])]; + tensor qkv_5_cast_fp16 = reshape(shape = var_846, x = linear_9_cast_fp16)[name = tensor("qkv_5_cast_fp16")]; + tensor q_13_begin_0 = const()[name = tensor("q_13_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor q_13_end_0 = const()[name = tensor("q_13_end_0"), val = tensor([1, 1, 1, 16, 64])]; + tensor q_13_end_mask_0 = const()[name = tensor("q_13_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor q_13_squeeze_mask_0 = const()[name = tensor("q_13_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor q_13_cast_fp16 = slice_by_index(begin = q_13_begin_0, end = q_13_end_0, end_mask = q_13_end_mask_0, squeeze_mask = q_13_squeeze_mask_0, x = qkv_5_cast_fp16)[name = tensor("q_13_cast_fp16")]; + tensor k_9_begin_0 = const()[name = tensor("k_9_begin_0"), val = tensor([0, 0, 1, 0, 0])]; + tensor k_9_end_0 = const()[name = tensor("k_9_end_0"), val = tensor([1, 1, 2, 16, 64])]; + tensor k_9_end_mask_0 = const()[name = tensor("k_9_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor k_9_squeeze_mask_0 = const()[name = tensor("k_9_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor k_9_cast_fp16 = slice_by_index(begin = k_9_begin_0, end = k_9_end_0, end_mask = k_9_end_mask_0, squeeze_mask = k_9_squeeze_mask_0, x = qkv_5_cast_fp16)[name = tensor("k_9_cast_fp16")]; + tensor v_5_begin_0 = const()[name = tensor("v_5_begin_0"), val = tensor([0, 0, 2, 0, 0])]; + tensor v_5_end_0 = const()[name = tensor("v_5_end_0"), val = tensor([1, 1, 3, 16, 64])]; + tensor v_5_end_mask_0 = const()[name = tensor("v_5_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor v_5_squeeze_mask_0 = const()[name = tensor("v_5_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor v_5_cast_fp16 = slice_by_index(begin = v_5_begin_0, end = v_5_end_0, end_mask = v_5_end_mask_0, squeeze_mask = v_5_squeeze_mask_0, x = qkv_5_cast_fp16)[name = tensor("v_5_cast_fp16")]; + tensor var_912 = const()[name = tensor("op_912"), val = tensor([1, 1, 1, 1])]; + tensor position2_to_fp16_dtype_0 = const()[name = tensor("position2_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor position2_to_fp16 = cast(dtype = position2_to_fp16_dtype_0, x = position2)[name = tensor("cast_129")]; + tensor ts_17_cast_fp16 = reshape(shape = var_912, x = position2_to_fp16)[name = tensor("ts_17_cast_fp16")]; + tensor var_919 = const()[name = tensor("op_919"), val = tensor([1, 1, 16, 32, 2])]; + tensor q_complex_5_cast_fp16 = reshape(shape = var_919, x = q_13_cast_fp16)[name = tensor("q_complex_5_cast_fp16")]; + tensor var_926 = const()[name = tensor("op_926"), val = tensor([1, 1, 16, 32, 2])]; + tensor k_complex_5_cast_fp16 = reshape(shape = var_926, x = k_9_cast_fp16)[name = tensor("k_complex_5_cast_fp16")]; + tensor var_930_begin_0 = const()[name = tensor("op_930_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_930_end_0 = const()[name = tensor("op_930_end_0"), val = tensor([1, 1, 16, 32, 1])]; + tensor var_930_end_mask_0 = const()[name = tensor("op_930_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_930_squeeze_mask_0 = const()[name = tensor("op_930_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_930_cast_fp16 = slice_by_index(begin = var_930_begin_0, end = var_930_end_0, end_mask = var_930_end_mask_0, squeeze_mask = var_930_squeeze_mask_0, x = q_complex_5_cast_fp16)[name = tensor("op_930_cast_fp16")]; + tensor var_938_begin_0 = const()[name = tensor("op_938_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_938_end_0 = const()[name = tensor("op_938_end_0"), val = tensor([1, 1, 16, 32, 2])]; + tensor var_938_end_mask_0 = const()[name = tensor("op_938_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_938_squeeze_mask_0 = const()[name = tensor("op_938_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_938_cast_fp16 = slice_by_index(begin = var_938_begin_0, end = var_938_end_0, end_mask = var_938_end_mask_0, squeeze_mask = var_938_squeeze_mask_0, x = q_complex_5_cast_fp16)[name = tensor("op_938_cast_fp16")]; + tensor var_946_begin_0 = const()[name = tensor("op_946_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_946_end_0 = const()[name = tensor("op_946_end_0"), val = tensor([1, 1, 16, 32, 1])]; + tensor var_946_end_mask_0 = const()[name = tensor("op_946_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_946_squeeze_mask_0 = const()[name = tensor("op_946_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_946_cast_fp16 = slice_by_index(begin = var_946_begin_0, end = var_946_end_0, end_mask = var_946_end_mask_0, squeeze_mask = var_946_squeeze_mask_0, x = k_complex_5_cast_fp16)[name = tensor("op_946_cast_fp16")]; + tensor var_954_begin_0 = const()[name = tensor("op_954_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_954_end_0 = const()[name = tensor("op_954_end_0"), val = tensor([1, 1, 16, 32, 2])]; + tensor var_954_end_mask_0 = const()[name = tensor("op_954_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_954_squeeze_mask_0 = const()[name = tensor("op_954_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_954_cast_fp16 = slice_by_index(begin = var_954_begin_0, end = var_954_end_0, end_mask = var_954_end_mask_0, squeeze_mask = var_954_squeeze_mask_0, x = k_complex_5_cast_fp16)[name = tensor("op_954_cast_fp16")]; + tensor freqs_5_to_fp16 = const()[name = tensor("freqs_5_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(57777024)))]; + tensor var_960_cast_fp16 = mul(x = freqs_5_to_fp16, y = ts_17_cast_fp16)[name = tensor("op_960_cast_fp16")]; + tensor rotr_5_cast_fp16 = cos(x = var_960_cast_fp16)[name = tensor("rotr_5_cast_fp16")]; + tensor roti_5_cast_fp16 = sin(x = var_960_cast_fp16)[name = tensor("roti_5_cast_fp16")]; + tensor var_964_cast_fp16 = mul(x = var_930_cast_fp16, y = rotr_5_cast_fp16)[name = tensor("op_964_cast_fp16")]; + tensor var_965_cast_fp16 = mul(x = var_938_cast_fp16, y = roti_5_cast_fp16)[name = tensor("op_965_cast_fp16")]; + tensor qor_9_cast_fp16 = sub(x = var_964_cast_fp16, y = var_965_cast_fp16)[name = tensor("qor_9_cast_fp16")]; + tensor var_968_cast_fp16 = mul(x = var_930_cast_fp16, y = roti_5_cast_fp16)[name = tensor("op_968_cast_fp16")]; + tensor var_969_cast_fp16 = mul(x = var_938_cast_fp16, y = rotr_5_cast_fp16)[name = tensor("op_969_cast_fp16")]; + tensor qoi_9_cast_fp16 = add(x = var_968_cast_fp16, y = var_969_cast_fp16)[name = tensor("qoi_9_cast_fp16")]; + tensor var_972_cast_fp16 = mul(x = var_946_cast_fp16, y = rotr_5_cast_fp16)[name = tensor("op_972_cast_fp16")]; + tensor var_973_cast_fp16 = mul(x = var_954_cast_fp16, y = roti_5_cast_fp16)[name = tensor("op_973_cast_fp16")]; + tensor kor_9_cast_fp16 = sub(x = var_972_cast_fp16, y = var_973_cast_fp16)[name = tensor("kor_9_cast_fp16")]; + tensor var_976_cast_fp16 = mul(x = var_946_cast_fp16, y = roti_5_cast_fp16)[name = tensor("op_976_cast_fp16")]; + tensor var_977_cast_fp16 = mul(x = var_954_cast_fp16, y = rotr_5_cast_fp16)[name = tensor("op_977_cast_fp16")]; + tensor koi_9_cast_fp16 = add(x = var_976_cast_fp16, y = var_977_cast_fp16)[name = tensor("koi_9_cast_fp16")]; + tensor qo_5_axis_0 = const()[name = tensor("qo_5_axis_0"), val = tensor(-1)]; + tensor qo_5_cast_fp16 = stack(axis = qo_5_axis_0, values = (qor_9_cast_fp16, qoi_9_cast_fp16))[name = tensor("qo_5_cast_fp16")]; + tensor ko_5_axis_0 = const()[name = tensor("ko_5_axis_0"), val = tensor(-1)]; + tensor ko_5_cast_fp16 = stack(axis = ko_5_axis_0, values = (kor_9_cast_fp16, koi_9_cast_fp16))[name = tensor("ko_5_cast_fp16")]; + tensor var_1010 = const()[name = tensor("op_1010"), val = tensor([1, 1, 16, 64])]; + tensor q_15_cast_fp16 = reshape(shape = var_1010, x = qo_5_cast_fp16)[name = tensor("q_15_cast_fp16")]; + tensor var_1016 = const()[name = tensor("op_1016"), val = tensor([1, 1, 16, 64])]; + tensor k_11_cast_fp16 = reshape(shape = var_1016, x = ko_5_cast_fp16)[name = tensor("k_11_cast_fp16")]; + tensor _inversed_1041_y_0_to_fp16 = const()[name = tensor("_inversed_1041_y_0_to_fp16"), val = tensor(0x1p-9)]; + tensor _inversed_1041_cast_fp16 = mul(x = ts_17_cast_fp16, y = _inversed_1041_y_0_to_fp16)[name = tensor("_inversed_1041_cast_fp16")]; + tensor var_1042_cast_fp16 = floor(x = _inversed_1041_cast_fp16)[name = tensor("op_1042_cast_fp16")]; + tensor var_1043_to_fp16 = const()[name = tensor("op_1043_to_fp16"), val = tensor(0x1p+9)]; + tensor var_1044_cast_fp16 = mul(x = var_1042_cast_fp16, y = var_1043_to_fp16)[name = tensor("op_1044_cast_fp16")]; + tensor write_indices_float_11_cast_fp16 = sub(x = ts_17_cast_fp16, y = var_1044_cast_fp16)[name = tensor("write_indices_float_11_cast_fp16")]; + tensor var_1051_dtype_0 = const()[name = tensor("op_1051_dtype_0"), val = tensor("int32")]; + tensor write_indices_5_reps_0 = const()[name = tensor("write_indices_5_reps_0"), val = tensor([1, 1, 16, 64])]; + tensor write_indices_float_11_cast_fp16_to_int32 = cast(dtype = var_1051_dtype_0, x = write_indices_float_11_cast_fp16)[name = tensor("cast_128")]; + tensor write_indices_5 = tile(reps = write_indices_5_reps_0, x = write_indices_float_11_cast_fp16_to_int32)[name = tensor("write_indices_5")]; + tensor var_1061_begin_0 = const()[name = tensor("op_1061_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_1061_end_0 = const()[name = tensor("op_1061_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor var_1061_end_mask_0 = const()[name = tensor("op_1061_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_1061_squeeze_mask_0 = const()[name = tensor("op_1061_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor cache2_to_fp16_dtype_0 = const()[name = tensor("cache2_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor cache2_to_fp16 = cast(dtype = cache2_to_fp16_dtype_0, x = cache2)[name = tensor("cast_127")]; + tensor var_1061_cast_fp16 = slice_by_index(begin = var_1061_begin_0, end = var_1061_end_0, end_mask = var_1061_end_mask_0, squeeze_mask = var_1061_squeeze_mask_0, x = cache2_to_fp16)[name = tensor("op_1061_cast_fp16")]; + tensor var_1063_axis_0 = const()[name = tensor("op_1063_axis_0"), val = tensor(1)]; + tensor var_1063_mode_0 = const()[name = tensor("op_1063_mode_0"), val = tensor("update")]; + tensor var_1063_validate_indices_0 = const()[name = tensor("op_1063_validate_indices_0"), val = tensor(false)]; + tensor var_1063_cast_fp16 = scatter_along_axis(axis = var_1063_axis_0, data = var_1061_cast_fp16, indices = write_indices_5, mode = var_1063_mode_0, updates = k_11_cast_fp16, validate_indices = var_1063_validate_indices_0)[name = tensor("op_1063_cast_fp16")]; + tensor concat_16 = const()[name = tensor("concat_16"), val = tensor([0, 0, 0, 0, 0])]; + tensor concat_17 = const()[name = tensor("concat_17"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_5_internal_tensor_assign_1_stride_0 = const()[name = tensor("new_cache_5_internal_tensor_assign_1_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_5_internal_tensor_assign_1_begin_mask_0 = const()[name = tensor("new_cache_5_internal_tensor_assign_1_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_5_internal_tensor_assign_1_end_mask_0 = const()[name = tensor("new_cache_5_internal_tensor_assign_1_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_5_internal_tensor_assign_1_squeeze_mask_0 = const()[name = tensor("new_cache_5_internal_tensor_assign_1_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_16 = const()[name = tensor("shape_16"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_4 = const()[name = tensor("reduce_prod_4"), val = tensor(1048576)]; + tensor range_1d_4_start_0 = const()[name = tensor("range_1d_4_start_0"), val = tensor(0)]; + tensor range_1d_4_step_0 = const()[name = tensor("range_1d_4_step_0"), val = tensor(1)]; + tensor range_1d_4 = range_1d(end = reduce_prod_4, start = range_1d_4_start_0, step = range_1d_4_step_0)[name = tensor("range_1d_4")]; + tensor reshape_20 = reshape(shape = shape_16, x = range_1d_4)[name = tensor("reshape_20")]; + tensor slice_by_index_4 = slice_by_index(begin = concat_16, begin_mask = new_cache_5_internal_tensor_assign_1_begin_mask_0, end = concat_17, end_mask = new_cache_5_internal_tensor_assign_1_end_mask_0, squeeze_mask = new_cache_5_internal_tensor_assign_1_squeeze_mask_0, stride = new_cache_5_internal_tensor_assign_1_stride_0, x = reshape_20)[name = tensor("slice_by_index_4")]; + tensor reshape_21_shape_0 = const()[name = tensor("reshape_21_shape_0"), val = tensor([-1])]; + tensor reshape_21 = reshape(shape = reshape_21_shape_0, x = slice_by_index_4)[name = tensor("reshape_21")]; + tensor reshape_22_shape_0 = const()[name = tensor("reshape_22_shape_0"), val = tensor([-1])]; + tensor reshape_22_cast_fp16 = reshape(shape = reshape_22_shape_0, x = var_1063_cast_fp16)[name = tensor("reshape_22_cast_fp16")]; + tensor reshape_23_shape_0 = const()[name = tensor("reshape_23_shape_0"), val = tensor([-1])]; + tensor reshape_23_cast_fp16 = reshape(shape = reshape_23_shape_0, x = cache2_to_fp16)[name = tensor("reshape_23_cast_fp16")]; + tensor scatter_4_mode_0 = const()[name = tensor("scatter_4_mode_0"), val = tensor("update")]; + tensor scatter_4_axis_0 = const()[name = tensor("scatter_4_axis_0"), val = tensor(0)]; + tensor scatter_4_validate_indices_0 = const()[name = tensor("scatter_4_validate_indices_0"), val = tensor(false)]; + tensor scatter_4_cast_fp16 = scatter(axis = scatter_4_axis_0, data = reshape_23_cast_fp16, indices = reshape_21, mode = scatter_4_mode_0, updates = reshape_22_cast_fp16, validate_indices = scatter_4_validate_indices_0)[name = tensor("scatter_4_cast_fp16")]; + tensor reshape_24_cast_fp16 = reshape(shape = shape_16, x = scatter_4_cast_fp16)[name = tensor("reshape_24_cast_fp16")]; + tensor var_1071_begin_0 = const()[name = tensor("op_1071_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor var_1071_end_0 = const()[name = tensor("op_1071_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor var_1071_end_mask_0 = const()[name = tensor("op_1071_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_1071_squeeze_mask_0 = const()[name = tensor("op_1071_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor var_1071_cast_fp16 = slice_by_index(begin = var_1071_begin_0, end = var_1071_end_0, end_mask = var_1071_end_mask_0, squeeze_mask = var_1071_squeeze_mask_0, x = reshape_24_cast_fp16)[name = tensor("op_1071_cast_fp16")]; + tensor var_1073_axis_0 = const()[name = tensor("op_1073_axis_0"), val = tensor(1)]; + tensor var_1073_mode_0 = const()[name = tensor("op_1073_mode_0"), val = tensor("update")]; + tensor var_1073_validate_indices_0 = const()[name = tensor("op_1073_validate_indices_0"), val = tensor(false)]; + tensor var_1073_cast_fp16 = scatter_along_axis(axis = var_1073_axis_0, data = var_1071_cast_fp16, indices = write_indices_5, mode = var_1073_mode_0, updates = v_5_cast_fp16, validate_indices = var_1073_validate_indices_0)[name = tensor("op_1073_cast_fp16")]; + tensor concat_18 = const()[name = tensor("concat_18"), val = tensor([1, 0, 0, 0, 0])]; + tensor concat_19 = const()[name = tensor("concat_19"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_5_internal_tensor_assign_2_stride_0 = const()[name = tensor("new_cache_5_internal_tensor_assign_2_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_5_internal_tensor_assign_2_begin_mask_0 = const()[name = tensor("new_cache_5_internal_tensor_assign_2_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_5_internal_tensor_assign_2_end_mask_0 = const()[name = tensor("new_cache_5_internal_tensor_assign_2_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_5_internal_tensor_assign_2_squeeze_mask_0 = const()[name = tensor("new_cache_5_internal_tensor_assign_2_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_17 = const()[name = tensor("shape_17"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_5 = const()[name = tensor("reduce_prod_5"), val = tensor(1048576)]; + tensor range_1d_5_start_0 = const()[name = tensor("range_1d_5_start_0"), val = tensor(0)]; + tensor range_1d_5_step_0 = const()[name = tensor("range_1d_5_step_0"), val = tensor(1)]; + tensor range_1d_5 = range_1d(end = reduce_prod_5, start = range_1d_5_start_0, step = range_1d_5_step_0)[name = tensor("range_1d_5")]; + tensor reshape_25 = reshape(shape = shape_17, x = range_1d_5)[name = tensor("reshape_25")]; + tensor slice_by_index_5 = slice_by_index(begin = concat_18, begin_mask = new_cache_5_internal_tensor_assign_2_begin_mask_0, end = concat_19, end_mask = new_cache_5_internal_tensor_assign_2_end_mask_0, squeeze_mask = new_cache_5_internal_tensor_assign_2_squeeze_mask_0, stride = new_cache_5_internal_tensor_assign_2_stride_0, x = reshape_25)[name = tensor("slice_by_index_5")]; + tensor reshape_26_shape_0 = const()[name = tensor("reshape_26_shape_0"), val = tensor([-1])]; + tensor reshape_26 = reshape(shape = reshape_26_shape_0, x = slice_by_index_5)[name = tensor("reshape_26")]; + tensor reshape_27_shape_0 = const()[name = tensor("reshape_27_shape_0"), val = tensor([-1])]; + tensor reshape_27_cast_fp16 = reshape(shape = reshape_27_shape_0, x = var_1073_cast_fp16)[name = tensor("reshape_27_cast_fp16")]; + tensor reshape_28_shape_0 = const()[name = tensor("reshape_28_shape_0"), val = tensor([-1])]; + tensor reshape_28_cast_fp16 = reshape(shape = reshape_28_shape_0, x = reshape_24_cast_fp16)[name = tensor("reshape_28_cast_fp16")]; + tensor scatter_5_mode_0 = const()[name = tensor("scatter_5_mode_0"), val = tensor("update")]; + tensor scatter_5_axis_0 = const()[name = tensor("scatter_5_axis_0"), val = tensor(0)]; + tensor scatter_5_validate_indices_0 = const()[name = tensor("scatter_5_validate_indices_0"), val = tensor(false)]; + tensor scatter_5_cast_fp16 = scatter(axis = scatter_5_axis_0, data = reshape_28_cast_fp16, indices = reshape_26, mode = scatter_5_mode_0, updates = reshape_27_cast_fp16, validate_indices = scatter_5_validate_indices_0)[name = tensor("scatter_5_cast_fp16")]; + tensor reshape_29_cast_fp16 = reshape(shape = shape_17, x = scatter_5_cast_fp16)[name = tensor("reshape_29_cast_fp16")]; + tensor reshape_29_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("reshape_29_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor keys_13_begin_0 = const()[name = tensor("keys_13_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor keys_13_end_0 = const()[name = tensor("keys_13_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor keys_13_end_mask_0 = const()[name = tensor("keys_13_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor keys_13_squeeze_mask_0 = const()[name = tensor("keys_13_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor keys_13_cast_fp16 = slice_by_index(begin = keys_13_begin_0, end = keys_13_end_0, end_mask = keys_13_end_mask_0, squeeze_mask = keys_13_squeeze_mask_0, x = reshape_29_cast_fp16)[name = tensor("keys_13_cast_fp16")]; + tensor values_13_begin_0 = const()[name = tensor("values_13_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor values_13_end_0 = const()[name = tensor("values_13_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor values_13_end_mask_0 = const()[name = tensor("values_13_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor values_13_squeeze_mask_0 = const()[name = tensor("values_13_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor values_13_cast_fp16 = slice_by_index(begin = values_13_begin_0, end = values_13_end_0, end_mask = values_13_end_mask_0, squeeze_mask = values_13_squeeze_mask_0, x = reshape_29_cast_fp16)[name = tensor("values_13_cast_fp16")]; + tensor var_1085_cast_fp16 = not_equal(x = keys_13_cast_fp16, y = keys_13_cast_fp16)[name = tensor("op_1085_cast_fp16")]; + tensor keys_15_cast_fp16 = select(a = var_319_to_fp16, b = keys_13_cast_fp16, cond = var_1085_cast_fp16)[name = tensor("keys_15_cast_fp16")]; + tensor var_1093_cast_fp16 = not_equal(x = values_13_cast_fp16, y = values_13_cast_fp16)[name = tensor("op_1093_cast_fp16")]; + tensor values_15_cast_fp16 = select(a = var_319_to_fp16, b = values_13_cast_fp16, cond = var_1093_cast_fp16)[name = tensor("values_15_cast_fp16")]; + tensor var_1117 = const()[name = tensor("op_1117"), val = tensor([0, 2, 1, 3])]; + tensor var_1133 = const()[name = tensor("op_1133"), val = tensor([1, 1, 1])]; + tensor var_1134_cast_fp16 = reshape(shape = var_1133, x = position2_to_fp16)[name = tensor("op_1134_cast_fp16")]; + tensor var_1154_to_fp16 = const()[name = tensor("op_1154_to_fp16"), val = tensor(0x1p+0)]; + tensor valid_len_5_cast_fp16 = add(x = var_1134_cast_fp16, y = var_1154_to_fp16)[name = tensor("valid_len_5_cast_fp16")]; + tensor valid_mask_5_cast_fp16 = less(x = k_positions_1_promoted_to_fp16, y = valid_len_5_cast_fp16)[name = tensor("valid_mask_5_cast_fp16")]; + tensor causal_mask_5_cast_fp16 = less_equal(x = k_positions_1_promoted_to_fp16, y = var_1134_cast_fp16)[name = tensor("causal_mask_5_cast_fp16")]; + tensor attn_mask_9 = logical_and(x = valid_mask_5_cast_fp16, y = causal_mask_5_cast_fp16)[name = tensor("attn_mask_9")]; + tensor attn_mask_11_axes_0 = const()[name = tensor("attn_mask_11_axes_0"), val = tensor([1])]; + tensor attn_mask_11 = expand_dims(axes = attn_mask_11_axes_0, x = attn_mask_9)[name = tensor("attn_mask_11")]; + tensor var_1172_transpose_x_0 = const()[name = tensor("op_1172_transpose_x_0"), val = tensor(false)]; + tensor var_1172_transpose_y_0 = const()[name = tensor("op_1172_transpose_y_0"), val = tensor(false)]; + tensor transpose_22_perm_0 = const()[name = tensor("transpose_22_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_23_perm_0 = const()[name = tensor("transpose_23_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_23 = transpose(perm = transpose_23_perm_0, x = keys_15_cast_fp16)[name = tensor("transpose_43")]; + tensor transpose_22 = transpose(perm = transpose_22_perm_0, x = q_15_cast_fp16)[name = tensor("transpose_44")]; + tensor var_1172_cast_fp16 = matmul(transpose_x = var_1172_transpose_x_0, transpose_y = var_1172_transpose_y_0, x = transpose_22, y = transpose_23)[name = tensor("op_1172_cast_fp16")]; + tensor var_1166_to_fp16 = const()[name = tensor("op_1166_to_fp16"), val = tensor([0x1p-3])]; + tensor attn_weights_13_cast_fp16 = mul(x = var_1172_cast_fp16, y = var_1166_to_fp16)[name = tensor("attn_weights_13_cast_fp16")]; + tensor var_1174 = logical_not(x = attn_mask_11)[name = tensor("op_1174")]; + tensor var_1175_to_fp16 = const()[name = tensor("op_1175_to_fp16"), val = tensor(-inf)]; + tensor attn_weights_15_cast_fp16 = select(a = var_1175_to_fp16, b = attn_weights_13_cast_fp16, cond = var_1174)[name = tensor("attn_weights_15_cast_fp16")]; + tensor var_1177 = const()[name = tensor("op_1177"), val = tensor(-1)]; + tensor attn_weights_17_cast_fp16 = softmax(axis = var_1177, x = attn_weights_15_cast_fp16)[name = tensor("attn_weights_17_cast_fp16")]; + tensor attn_output_5_transpose_x_0 = const()[name = tensor("attn_output_5_transpose_x_0"), val = tensor(false)]; + tensor attn_output_5_transpose_y_0 = const()[name = tensor("attn_output_5_transpose_y_0"), val = tensor(false)]; + tensor values_17_cast_fp16 = transpose(perm = var_1117, x = values_15_cast_fp16)[name = tensor("transpose_45")]; + tensor attn_output_5_cast_fp16 = matmul(transpose_x = attn_output_5_transpose_x_0, transpose_y = attn_output_5_transpose_y_0, x = attn_weights_17_cast_fp16, y = values_17_cast_fp16)[name = tensor("attn_output_5_cast_fp16")]; + tensor var_1185 = const()[name = tensor("op_1185"), val = tensor([0, 2, 1, 3])]; + tensor var_1190 = const()[name = tensor("op_1190"), val = tensor([1, 1, 1024])]; + tensor var_1186_cast_fp16 = transpose(perm = var_1185, x = attn_output_5_cast_fp16)[name = tensor("transpose_42")]; + tensor input_31_cast_fp16 = reshape(shape = var_1190, x = var_1186_cast_fp16)[name = tensor("input_31_cast_fp16")]; + tensor attn2_out_proj_weight_to_fp16 = const()[name = tensor("attn2_out_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(57777152)))]; + tensor linear_10_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = attn2_out_proj_weight_to_fp16, x = input_31_cast_fp16)[name = tensor("linear_10_cast_fp16")]; + tensor var_1196_to_fp16 = const()[name = tensor("op_1196_to_fp16"), val = tensor(0x1p+0)]; + tensor var_1197_cast_fp16 = add(x = position2_to_fp16, y = var_1196_to_fp16)[name = tensor("op_1197_cast_fp16")]; + tensor var_1197_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_1197_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor input_33_cast_fp16 = add(x = input_27_cast_fp16, y = linear_10_cast_fp16)[name = tensor("input_33_cast_fp16")]; + tensor input_35_axes_0 = const()[name = tensor("input_35_axes_0"), val = tensor([-1])]; + tensor norm2_2_weight_to_fp16 = const()[name = tensor("norm2_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(59874368)))]; + tensor norm2_2_bias_to_fp16 = const()[name = tensor("norm2_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(59876480)))]; + tensor var_1201_to_fp16 = const()[name = tensor("op_1201_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_35_cast_fp16 = layer_norm(axes = input_35_axes_0, beta = norm2_2_bias_to_fp16, epsilon = var_1201_to_fp16, gamma = norm2_2_weight_to_fp16, x = input_33_cast_fp16)[name = tensor("input_35_cast_fp16")]; + tensor linear2_1_weight_to_fp16 = const()[name = tensor("linear2_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(59878592)))]; + tensor linear_11_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = linear2_1_weight_to_fp16, x = input_35_cast_fp16)[name = tensor("linear_11_cast_fp16")]; + tensor input_37_mode_0 = const()[name = tensor("input_37_mode_0"), val = tensor("EXACT")]; + tensor input_37_cast_fp16 = gelu(mode = input_37_mode_0, x = linear_11_cast_fp16)[name = tensor("input_37_cast_fp16")]; + tensor linear2_2_weight_to_fp16 = const()[name = tensor("linear2_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(68267264)))]; + tensor linear_12_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = linear2_2_weight_to_fp16, x = input_37_cast_fp16)[name = tensor("linear_12_cast_fp16")]; + tensor input_39_cast_fp16 = add(x = input_33_cast_fp16, y = linear_12_cast_fp16)[name = tensor("input_39_cast_fp16")]; + tensor input_41_axes_0 = const()[name = tensor("input_41_axes_0"), val = tensor([-1])]; + tensor norm3_1_weight_to_fp16 = const()[name = tensor("norm3_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(76655936)))]; + tensor norm3_1_bias_to_fp16 = const()[name = tensor("norm3_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(76658048)))]; + tensor var_1218_to_fp16 = const()[name = tensor("op_1218_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_41_cast_fp16 = layer_norm(axes = input_41_axes_0, beta = norm3_1_bias_to_fp16, epsilon = var_1218_to_fp16, gamma = norm3_1_weight_to_fp16, x = input_39_cast_fp16)[name = tensor("input_41_cast_fp16")]; + tensor attn3_in_proj_weight_to_fp16 = const()[name = tensor("attn3_in_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(76660160)))]; + tensor linear_13_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = attn3_in_proj_weight_to_fp16, x = input_41_cast_fp16)[name = tensor("linear_13_cast_fp16")]; + tensor var_1232 = const()[name = tensor("op_1232"), val = tensor([1, 1, 3, 16, 64])]; + tensor qkv_7_cast_fp16 = reshape(shape = var_1232, x = linear_13_cast_fp16)[name = tensor("qkv_7_cast_fp16")]; + tensor q_19_begin_0 = const()[name = tensor("q_19_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor q_19_end_0 = const()[name = tensor("q_19_end_0"), val = tensor([1, 1, 1, 16, 64])]; + tensor q_19_end_mask_0 = const()[name = tensor("q_19_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor q_19_squeeze_mask_0 = const()[name = tensor("q_19_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor q_19_cast_fp16 = slice_by_index(begin = q_19_begin_0, end = q_19_end_0, end_mask = q_19_end_mask_0, squeeze_mask = q_19_squeeze_mask_0, x = qkv_7_cast_fp16)[name = tensor("q_19_cast_fp16")]; + tensor k_13_begin_0 = const()[name = tensor("k_13_begin_0"), val = tensor([0, 0, 1, 0, 0])]; + tensor k_13_end_0 = const()[name = tensor("k_13_end_0"), val = tensor([1, 1, 2, 16, 64])]; + tensor k_13_end_mask_0 = const()[name = tensor("k_13_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor k_13_squeeze_mask_0 = const()[name = tensor("k_13_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor k_13_cast_fp16 = slice_by_index(begin = k_13_begin_0, end = k_13_end_0, end_mask = k_13_end_mask_0, squeeze_mask = k_13_squeeze_mask_0, x = qkv_7_cast_fp16)[name = tensor("k_13_cast_fp16")]; + tensor v_7_begin_0 = const()[name = tensor("v_7_begin_0"), val = tensor([0, 0, 2, 0, 0])]; + tensor v_7_end_0 = const()[name = tensor("v_7_end_0"), val = tensor([1, 1, 3, 16, 64])]; + tensor v_7_end_mask_0 = const()[name = tensor("v_7_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor v_7_squeeze_mask_0 = const()[name = tensor("v_7_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor v_7_cast_fp16 = slice_by_index(begin = v_7_begin_0, end = v_7_end_0, end_mask = v_7_end_mask_0, squeeze_mask = v_7_squeeze_mask_0, x = qkv_7_cast_fp16)[name = tensor("v_7_cast_fp16")]; + tensor var_1298 = const()[name = tensor("op_1298"), val = tensor([1, 1, 1, 1])]; + tensor position3_to_fp16_dtype_0 = const()[name = tensor("position3_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor position3_to_fp16 = cast(dtype = position3_to_fp16_dtype_0, x = position3)[name = tensor("cast_124")]; + tensor ts_23_cast_fp16 = reshape(shape = var_1298, x = position3_to_fp16)[name = tensor("ts_23_cast_fp16")]; + tensor var_1305 = const()[name = tensor("op_1305"), val = tensor([1, 1, 16, 32, 2])]; + tensor q_complex_7_cast_fp16 = reshape(shape = var_1305, x = q_19_cast_fp16)[name = tensor("q_complex_7_cast_fp16")]; + tensor var_1312 = const()[name = tensor("op_1312"), val = tensor([1, 1, 16, 32, 2])]; + tensor k_complex_7_cast_fp16 = reshape(shape = var_1312, x = k_13_cast_fp16)[name = tensor("k_complex_7_cast_fp16")]; + tensor var_1316_begin_0 = const()[name = tensor("op_1316_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_1316_end_0 = const()[name = tensor("op_1316_end_0"), val = tensor([1, 1, 16, 32, 1])]; + tensor var_1316_end_mask_0 = const()[name = tensor("op_1316_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_1316_squeeze_mask_0 = const()[name = tensor("op_1316_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_1316_cast_fp16 = slice_by_index(begin = var_1316_begin_0, end = var_1316_end_0, end_mask = var_1316_end_mask_0, squeeze_mask = var_1316_squeeze_mask_0, x = q_complex_7_cast_fp16)[name = tensor("op_1316_cast_fp16")]; + tensor var_1324_begin_0 = const()[name = tensor("op_1324_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_1324_end_0 = const()[name = tensor("op_1324_end_0"), val = tensor([1, 1, 16, 32, 2])]; + tensor var_1324_end_mask_0 = const()[name = tensor("op_1324_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_1324_squeeze_mask_0 = const()[name = tensor("op_1324_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_1324_cast_fp16 = slice_by_index(begin = var_1324_begin_0, end = var_1324_end_0, end_mask = var_1324_end_mask_0, squeeze_mask = var_1324_squeeze_mask_0, x = q_complex_7_cast_fp16)[name = tensor("op_1324_cast_fp16")]; + tensor var_1332_begin_0 = const()[name = tensor("op_1332_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_1332_end_0 = const()[name = tensor("op_1332_end_0"), val = tensor([1, 1, 16, 32, 1])]; + tensor var_1332_end_mask_0 = const()[name = tensor("op_1332_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_1332_squeeze_mask_0 = const()[name = tensor("op_1332_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_1332_cast_fp16 = slice_by_index(begin = var_1332_begin_0, end = var_1332_end_0, end_mask = var_1332_end_mask_0, squeeze_mask = var_1332_squeeze_mask_0, x = k_complex_7_cast_fp16)[name = tensor("op_1332_cast_fp16")]; + tensor var_1340_begin_0 = const()[name = tensor("op_1340_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_1340_end_0 = const()[name = tensor("op_1340_end_0"), val = tensor([1, 1, 16, 32, 2])]; + tensor var_1340_end_mask_0 = const()[name = tensor("op_1340_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_1340_squeeze_mask_0 = const()[name = tensor("op_1340_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_1340_cast_fp16 = slice_by_index(begin = var_1340_begin_0, end = var_1340_end_0, end_mask = var_1340_end_mask_0, squeeze_mask = var_1340_squeeze_mask_0, x = k_complex_7_cast_fp16)[name = tensor("op_1340_cast_fp16")]; + tensor freqs_7_to_fp16 = const()[name = tensor("freqs_7_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(82951680)))]; + tensor var_1346_cast_fp16 = mul(x = freqs_7_to_fp16, y = ts_23_cast_fp16)[name = tensor("op_1346_cast_fp16")]; + tensor rotr_7_cast_fp16 = cos(x = var_1346_cast_fp16)[name = tensor("rotr_7_cast_fp16")]; + tensor roti_7_cast_fp16 = sin(x = var_1346_cast_fp16)[name = tensor("roti_7_cast_fp16")]; + tensor var_1350_cast_fp16 = mul(x = var_1316_cast_fp16, y = rotr_7_cast_fp16)[name = tensor("op_1350_cast_fp16")]; + tensor var_1351_cast_fp16 = mul(x = var_1324_cast_fp16, y = roti_7_cast_fp16)[name = tensor("op_1351_cast_fp16")]; + tensor qor_13_cast_fp16 = sub(x = var_1350_cast_fp16, y = var_1351_cast_fp16)[name = tensor("qor_13_cast_fp16")]; + tensor var_1354_cast_fp16 = mul(x = var_1316_cast_fp16, y = roti_7_cast_fp16)[name = tensor("op_1354_cast_fp16")]; + tensor var_1355_cast_fp16 = mul(x = var_1324_cast_fp16, y = rotr_7_cast_fp16)[name = tensor("op_1355_cast_fp16")]; + tensor qoi_13_cast_fp16 = add(x = var_1354_cast_fp16, y = var_1355_cast_fp16)[name = tensor("qoi_13_cast_fp16")]; + tensor var_1358_cast_fp16 = mul(x = var_1332_cast_fp16, y = rotr_7_cast_fp16)[name = tensor("op_1358_cast_fp16")]; + tensor var_1359_cast_fp16 = mul(x = var_1340_cast_fp16, y = roti_7_cast_fp16)[name = tensor("op_1359_cast_fp16")]; + tensor kor_13_cast_fp16 = sub(x = var_1358_cast_fp16, y = var_1359_cast_fp16)[name = tensor("kor_13_cast_fp16")]; + tensor var_1362_cast_fp16 = mul(x = var_1332_cast_fp16, y = roti_7_cast_fp16)[name = tensor("op_1362_cast_fp16")]; + tensor var_1363_cast_fp16 = mul(x = var_1340_cast_fp16, y = rotr_7_cast_fp16)[name = tensor("op_1363_cast_fp16")]; + tensor koi_13_cast_fp16 = add(x = var_1362_cast_fp16, y = var_1363_cast_fp16)[name = tensor("koi_13_cast_fp16")]; + tensor qo_7_axis_0 = const()[name = tensor("qo_7_axis_0"), val = tensor(-1)]; + tensor qo_7_cast_fp16 = stack(axis = qo_7_axis_0, values = (qor_13_cast_fp16, qoi_13_cast_fp16))[name = tensor("qo_7_cast_fp16")]; + tensor ko_7_axis_0 = const()[name = tensor("ko_7_axis_0"), val = tensor(-1)]; + tensor ko_7_cast_fp16 = stack(axis = ko_7_axis_0, values = (kor_13_cast_fp16, koi_13_cast_fp16))[name = tensor("ko_7_cast_fp16")]; + tensor var_1396 = const()[name = tensor("op_1396"), val = tensor([1, 1, 16, 64])]; + tensor q_21_cast_fp16 = reshape(shape = var_1396, x = qo_7_cast_fp16)[name = tensor("q_21_cast_fp16")]; + tensor var_1402 = const()[name = tensor("op_1402"), val = tensor([1, 1, 16, 64])]; + tensor k_15_cast_fp16 = reshape(shape = var_1402, x = ko_7_cast_fp16)[name = tensor("k_15_cast_fp16")]; + tensor _inversed_1427_y_0_to_fp16 = const()[name = tensor("_inversed_1427_y_0_to_fp16"), val = tensor(0x1p-9)]; + tensor _inversed_1427_cast_fp16 = mul(x = ts_23_cast_fp16, y = _inversed_1427_y_0_to_fp16)[name = tensor("_inversed_1427_cast_fp16")]; + tensor var_1428_cast_fp16 = floor(x = _inversed_1427_cast_fp16)[name = tensor("op_1428_cast_fp16")]; + tensor var_1429_to_fp16 = const()[name = tensor("op_1429_to_fp16"), val = tensor(0x1p+9)]; + tensor var_1430_cast_fp16 = mul(x = var_1428_cast_fp16, y = var_1429_to_fp16)[name = tensor("op_1430_cast_fp16")]; + tensor write_indices_float_15_cast_fp16 = sub(x = ts_23_cast_fp16, y = var_1430_cast_fp16)[name = tensor("write_indices_float_15_cast_fp16")]; + tensor var_1437_dtype_0 = const()[name = tensor("op_1437_dtype_0"), val = tensor("int32")]; + tensor write_indices_7_reps_0 = const()[name = tensor("write_indices_7_reps_0"), val = tensor([1, 1, 16, 64])]; + tensor write_indices_float_15_cast_fp16_to_int32 = cast(dtype = var_1437_dtype_0, x = write_indices_float_15_cast_fp16)[name = tensor("cast_123")]; + tensor write_indices_7 = tile(reps = write_indices_7_reps_0, x = write_indices_float_15_cast_fp16_to_int32)[name = tensor("write_indices_7")]; + tensor var_1447_begin_0 = const()[name = tensor("op_1447_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_1447_end_0 = const()[name = tensor("op_1447_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor var_1447_end_mask_0 = const()[name = tensor("op_1447_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_1447_squeeze_mask_0 = const()[name = tensor("op_1447_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor cache3_to_fp16_dtype_0 = const()[name = tensor("cache3_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor cache3_to_fp16 = cast(dtype = cache3_to_fp16_dtype_0, x = cache3)[name = tensor("cast_122")]; + tensor var_1447_cast_fp16 = slice_by_index(begin = var_1447_begin_0, end = var_1447_end_0, end_mask = var_1447_end_mask_0, squeeze_mask = var_1447_squeeze_mask_0, x = cache3_to_fp16)[name = tensor("op_1447_cast_fp16")]; + tensor var_1449_axis_0 = const()[name = tensor("op_1449_axis_0"), val = tensor(1)]; + tensor var_1449_mode_0 = const()[name = tensor("op_1449_mode_0"), val = tensor("update")]; + tensor var_1449_validate_indices_0 = const()[name = tensor("op_1449_validate_indices_0"), val = tensor(false)]; + tensor var_1449_cast_fp16 = scatter_along_axis(axis = var_1449_axis_0, data = var_1447_cast_fp16, indices = write_indices_7, mode = var_1449_mode_0, updates = k_15_cast_fp16, validate_indices = var_1449_validate_indices_0)[name = tensor("op_1449_cast_fp16")]; + tensor concat_23 = const()[name = tensor("concat_23"), val = tensor([0, 0, 0, 0, 0])]; + tensor concat_24 = const()[name = tensor("concat_24"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_7_internal_tensor_assign_1_stride_0 = const()[name = tensor("new_cache_7_internal_tensor_assign_1_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_7_internal_tensor_assign_1_begin_mask_0 = const()[name = tensor("new_cache_7_internal_tensor_assign_1_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_7_internal_tensor_assign_1_end_mask_0 = const()[name = tensor("new_cache_7_internal_tensor_assign_1_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_7_internal_tensor_assign_1_squeeze_mask_0 = const()[name = tensor("new_cache_7_internal_tensor_assign_1_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_18 = const()[name = tensor("shape_18"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_6 = const()[name = tensor("reduce_prod_6"), val = tensor(1048576)]; + tensor range_1d_6_start_0 = const()[name = tensor("range_1d_6_start_0"), val = tensor(0)]; + tensor range_1d_6_step_0 = const()[name = tensor("range_1d_6_step_0"), val = tensor(1)]; + tensor range_1d_6 = range_1d(end = reduce_prod_6, start = range_1d_6_start_0, step = range_1d_6_step_0)[name = tensor("range_1d_6")]; + tensor reshape_30 = reshape(shape = shape_18, x = range_1d_6)[name = tensor("reshape_30")]; + tensor slice_by_index_6 = slice_by_index(begin = concat_23, begin_mask = new_cache_7_internal_tensor_assign_1_begin_mask_0, end = concat_24, end_mask = new_cache_7_internal_tensor_assign_1_end_mask_0, squeeze_mask = new_cache_7_internal_tensor_assign_1_squeeze_mask_0, stride = new_cache_7_internal_tensor_assign_1_stride_0, x = reshape_30)[name = tensor("slice_by_index_6")]; + tensor reshape_31_shape_0 = const()[name = tensor("reshape_31_shape_0"), val = tensor([-1])]; + tensor reshape_31 = reshape(shape = reshape_31_shape_0, x = slice_by_index_6)[name = tensor("reshape_31")]; + tensor reshape_32_shape_0 = const()[name = tensor("reshape_32_shape_0"), val = tensor([-1])]; + tensor reshape_32_cast_fp16 = reshape(shape = reshape_32_shape_0, x = var_1449_cast_fp16)[name = tensor("reshape_32_cast_fp16")]; + tensor reshape_33_shape_0 = const()[name = tensor("reshape_33_shape_0"), val = tensor([-1])]; + tensor reshape_33_cast_fp16 = reshape(shape = reshape_33_shape_0, x = cache3_to_fp16)[name = tensor("reshape_33_cast_fp16")]; + tensor scatter_6_mode_0 = const()[name = tensor("scatter_6_mode_0"), val = tensor("update")]; + tensor scatter_6_axis_0 = const()[name = tensor("scatter_6_axis_0"), val = tensor(0)]; + tensor scatter_6_validate_indices_0 = const()[name = tensor("scatter_6_validate_indices_0"), val = tensor(false)]; + tensor scatter_6_cast_fp16 = scatter(axis = scatter_6_axis_0, data = reshape_33_cast_fp16, indices = reshape_31, mode = scatter_6_mode_0, updates = reshape_32_cast_fp16, validate_indices = scatter_6_validate_indices_0)[name = tensor("scatter_6_cast_fp16")]; + tensor reshape_34_cast_fp16 = reshape(shape = shape_18, x = scatter_6_cast_fp16)[name = tensor("reshape_34_cast_fp16")]; + tensor var_1457_begin_0 = const()[name = tensor("op_1457_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor var_1457_end_0 = const()[name = tensor("op_1457_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor var_1457_end_mask_0 = const()[name = tensor("op_1457_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_1457_squeeze_mask_0 = const()[name = tensor("op_1457_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor var_1457_cast_fp16 = slice_by_index(begin = var_1457_begin_0, end = var_1457_end_0, end_mask = var_1457_end_mask_0, squeeze_mask = var_1457_squeeze_mask_0, x = reshape_34_cast_fp16)[name = tensor("op_1457_cast_fp16")]; + tensor var_1459_axis_0 = const()[name = tensor("op_1459_axis_0"), val = tensor(1)]; + tensor var_1459_mode_0 = const()[name = tensor("op_1459_mode_0"), val = tensor("update")]; + tensor var_1459_validate_indices_0 = const()[name = tensor("op_1459_validate_indices_0"), val = tensor(false)]; + tensor var_1459_cast_fp16 = scatter_along_axis(axis = var_1459_axis_0, data = var_1457_cast_fp16, indices = write_indices_7, mode = var_1459_mode_0, updates = v_7_cast_fp16, validate_indices = var_1459_validate_indices_0)[name = tensor("op_1459_cast_fp16")]; + tensor concat_25 = const()[name = tensor("concat_25"), val = tensor([1, 0, 0, 0, 0])]; + tensor concat_26 = const()[name = tensor("concat_26"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_7_internal_tensor_assign_2_stride_0 = const()[name = tensor("new_cache_7_internal_tensor_assign_2_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_7_internal_tensor_assign_2_begin_mask_0 = const()[name = tensor("new_cache_7_internal_tensor_assign_2_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_7_internal_tensor_assign_2_end_mask_0 = const()[name = tensor("new_cache_7_internal_tensor_assign_2_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_7_internal_tensor_assign_2_squeeze_mask_0 = const()[name = tensor("new_cache_7_internal_tensor_assign_2_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_19 = const()[name = tensor("shape_19"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_7 = const()[name = tensor("reduce_prod_7"), val = tensor(1048576)]; + tensor range_1d_7_start_0 = const()[name = tensor("range_1d_7_start_0"), val = tensor(0)]; + tensor range_1d_7_step_0 = const()[name = tensor("range_1d_7_step_0"), val = tensor(1)]; + tensor range_1d_7 = range_1d(end = reduce_prod_7, start = range_1d_7_start_0, step = range_1d_7_step_0)[name = tensor("range_1d_7")]; + tensor reshape_35 = reshape(shape = shape_19, x = range_1d_7)[name = tensor("reshape_35")]; + tensor slice_by_index_7 = slice_by_index(begin = concat_25, begin_mask = new_cache_7_internal_tensor_assign_2_begin_mask_0, end = concat_26, end_mask = new_cache_7_internal_tensor_assign_2_end_mask_0, squeeze_mask = new_cache_7_internal_tensor_assign_2_squeeze_mask_0, stride = new_cache_7_internal_tensor_assign_2_stride_0, x = reshape_35)[name = tensor("slice_by_index_7")]; + tensor reshape_36_shape_0 = const()[name = tensor("reshape_36_shape_0"), val = tensor([-1])]; + tensor reshape_36 = reshape(shape = reshape_36_shape_0, x = slice_by_index_7)[name = tensor("reshape_36")]; + tensor reshape_37_shape_0 = const()[name = tensor("reshape_37_shape_0"), val = tensor([-1])]; + tensor reshape_37_cast_fp16 = reshape(shape = reshape_37_shape_0, x = var_1459_cast_fp16)[name = tensor("reshape_37_cast_fp16")]; + tensor reshape_38_shape_0 = const()[name = tensor("reshape_38_shape_0"), val = tensor([-1])]; + tensor reshape_38_cast_fp16 = reshape(shape = reshape_38_shape_0, x = reshape_34_cast_fp16)[name = tensor("reshape_38_cast_fp16")]; + tensor scatter_7_mode_0 = const()[name = tensor("scatter_7_mode_0"), val = tensor("update")]; + tensor scatter_7_axis_0 = const()[name = tensor("scatter_7_axis_0"), val = tensor(0)]; + tensor scatter_7_validate_indices_0 = const()[name = tensor("scatter_7_validate_indices_0"), val = tensor(false)]; + tensor scatter_7_cast_fp16 = scatter(axis = scatter_7_axis_0, data = reshape_38_cast_fp16, indices = reshape_36, mode = scatter_7_mode_0, updates = reshape_37_cast_fp16, validate_indices = scatter_7_validate_indices_0)[name = tensor("scatter_7_cast_fp16")]; + tensor reshape_39_cast_fp16 = reshape(shape = shape_19, x = scatter_7_cast_fp16)[name = tensor("reshape_39_cast_fp16")]; + tensor reshape_39_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("reshape_39_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor keys_19_begin_0 = const()[name = tensor("keys_19_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor keys_19_end_0 = const()[name = tensor("keys_19_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor keys_19_end_mask_0 = const()[name = tensor("keys_19_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor keys_19_squeeze_mask_0 = const()[name = tensor("keys_19_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor keys_19_cast_fp16 = slice_by_index(begin = keys_19_begin_0, end = keys_19_end_0, end_mask = keys_19_end_mask_0, squeeze_mask = keys_19_squeeze_mask_0, x = reshape_39_cast_fp16)[name = tensor("keys_19_cast_fp16")]; + tensor values_19_begin_0 = const()[name = tensor("values_19_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor values_19_end_0 = const()[name = tensor("values_19_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor values_19_end_mask_0 = const()[name = tensor("values_19_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor values_19_squeeze_mask_0 = const()[name = tensor("values_19_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor values_19_cast_fp16 = slice_by_index(begin = values_19_begin_0, end = values_19_end_0, end_mask = values_19_end_mask_0, squeeze_mask = values_19_squeeze_mask_0, x = reshape_39_cast_fp16)[name = tensor("values_19_cast_fp16")]; + tensor var_1471_cast_fp16 = not_equal(x = keys_19_cast_fp16, y = keys_19_cast_fp16)[name = tensor("op_1471_cast_fp16")]; + tensor keys_21_cast_fp16 = select(a = var_319_to_fp16, b = keys_19_cast_fp16, cond = var_1471_cast_fp16)[name = tensor("keys_21_cast_fp16")]; + tensor var_1479_cast_fp16 = not_equal(x = values_19_cast_fp16, y = values_19_cast_fp16)[name = tensor("op_1479_cast_fp16")]; + tensor values_21_cast_fp16 = select(a = var_319_to_fp16, b = values_19_cast_fp16, cond = var_1479_cast_fp16)[name = tensor("values_21_cast_fp16")]; + tensor var_1503 = const()[name = tensor("op_1503"), val = tensor([0, 2, 1, 3])]; + tensor var_1519 = const()[name = tensor("op_1519"), val = tensor([1, 1, 1])]; + tensor var_1520_cast_fp16 = reshape(shape = var_1519, x = position3_to_fp16)[name = tensor("op_1520_cast_fp16")]; + tensor var_1540_to_fp16 = const()[name = tensor("op_1540_to_fp16"), val = tensor(0x1p+0)]; + tensor valid_len_7_cast_fp16 = add(x = var_1520_cast_fp16, y = var_1540_to_fp16)[name = tensor("valid_len_7_cast_fp16")]; + tensor valid_mask_7_cast_fp16 = less(x = k_positions_1_promoted_to_fp16, y = valid_len_7_cast_fp16)[name = tensor("valid_mask_7_cast_fp16")]; + tensor causal_mask_7_cast_fp16 = less_equal(x = k_positions_1_promoted_to_fp16, y = var_1520_cast_fp16)[name = tensor("causal_mask_7_cast_fp16")]; + tensor attn_mask_13 = logical_and(x = valid_mask_7_cast_fp16, y = causal_mask_7_cast_fp16)[name = tensor("attn_mask_13")]; + tensor attn_mask_15_axes_0 = const()[name = tensor("attn_mask_15_axes_0"), val = tensor([1])]; + tensor attn_mask_15 = expand_dims(axes = attn_mask_15_axes_0, x = attn_mask_13)[name = tensor("attn_mask_15")]; + tensor var_1558_transpose_x_0 = const()[name = tensor("op_1558_transpose_x_0"), val = tensor(false)]; + tensor var_1558_transpose_y_0 = const()[name = tensor("op_1558_transpose_y_0"), val = tensor(false)]; + tensor transpose_24_perm_0 = const()[name = tensor("transpose_24_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_25_perm_0 = const()[name = tensor("transpose_25_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_25 = transpose(perm = transpose_25_perm_0, x = keys_21_cast_fp16)[name = tensor("transpose_39")]; + tensor transpose_24 = transpose(perm = transpose_24_perm_0, x = q_21_cast_fp16)[name = tensor("transpose_40")]; + tensor var_1558_cast_fp16 = matmul(transpose_x = var_1558_transpose_x_0, transpose_y = var_1558_transpose_y_0, x = transpose_24, y = transpose_25)[name = tensor("op_1558_cast_fp16")]; + tensor var_1552_to_fp16 = const()[name = tensor("op_1552_to_fp16"), val = tensor([0x1p-3])]; + tensor attn_weights_19_cast_fp16 = mul(x = var_1558_cast_fp16, y = var_1552_to_fp16)[name = tensor("attn_weights_19_cast_fp16")]; + tensor var_1560 = logical_not(x = attn_mask_15)[name = tensor("op_1560")]; + tensor var_1561_to_fp16 = const()[name = tensor("op_1561_to_fp16"), val = tensor(-inf)]; + tensor attn_weights_21_cast_fp16 = select(a = var_1561_to_fp16, b = attn_weights_19_cast_fp16, cond = var_1560)[name = tensor("attn_weights_21_cast_fp16")]; + tensor var_1563 = const()[name = tensor("op_1563"), val = tensor(-1)]; + tensor attn_weights_23_cast_fp16 = softmax(axis = var_1563, x = attn_weights_21_cast_fp16)[name = tensor("attn_weights_23_cast_fp16")]; + tensor attn_output_7_transpose_x_0 = const()[name = tensor("attn_output_7_transpose_x_0"), val = tensor(false)]; + tensor attn_output_7_transpose_y_0 = const()[name = tensor("attn_output_7_transpose_y_0"), val = tensor(false)]; + tensor values_23_cast_fp16 = transpose(perm = var_1503, x = values_21_cast_fp16)[name = tensor("transpose_41")]; + tensor attn_output_7_cast_fp16 = matmul(transpose_x = attn_output_7_transpose_x_0, transpose_y = attn_output_7_transpose_y_0, x = attn_weights_23_cast_fp16, y = values_23_cast_fp16)[name = tensor("attn_output_7_cast_fp16")]; + tensor var_1571 = const()[name = tensor("op_1571"), val = tensor([0, 2, 1, 3])]; + tensor var_1576 = const()[name = tensor("op_1576"), val = tensor([1, 1, 1024])]; + tensor var_1572_cast_fp16 = transpose(perm = var_1571, x = attn_output_7_cast_fp16)[name = tensor("transpose_38")]; + tensor input_43_cast_fp16 = reshape(shape = var_1576, x = var_1572_cast_fp16)[name = tensor("input_43_cast_fp16")]; + tensor attn3_out_proj_weight_to_fp16 = const()[name = tensor("attn3_out_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(82951808)))]; + tensor linear_14_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = attn3_out_proj_weight_to_fp16, x = input_43_cast_fp16)[name = tensor("linear_14_cast_fp16")]; + tensor var_1582_to_fp16 = const()[name = tensor("op_1582_to_fp16"), val = tensor(0x1p+0)]; + tensor var_1583_cast_fp16 = add(x = position3_to_fp16, y = var_1582_to_fp16)[name = tensor("op_1583_cast_fp16")]; + tensor var_1583_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_1583_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor input_45_cast_fp16 = add(x = input_39_cast_fp16, y = linear_14_cast_fp16)[name = tensor("input_45_cast_fp16")]; + tensor input_47_axes_0 = const()[name = tensor("input_47_axes_0"), val = tensor([-1])]; + tensor norm3_2_weight_to_fp16 = const()[name = tensor("norm3_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(85049024)))]; + tensor norm3_2_bias_to_fp16 = const()[name = tensor("norm3_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(85051136)))]; + tensor var_1587_to_fp16 = const()[name = tensor("op_1587_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_47_cast_fp16 = layer_norm(axes = input_47_axes_0, beta = norm3_2_bias_to_fp16, epsilon = var_1587_to_fp16, gamma = norm3_2_weight_to_fp16, x = input_45_cast_fp16)[name = tensor("input_47_cast_fp16")]; + tensor linear3_1_weight_to_fp16 = const()[name = tensor("linear3_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(85053248)))]; + tensor linear_15_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = linear3_1_weight_to_fp16, x = input_47_cast_fp16)[name = tensor("linear_15_cast_fp16")]; + tensor input_49_mode_0 = const()[name = tensor("input_49_mode_0"), val = tensor("EXACT")]; + tensor input_49_cast_fp16 = gelu(mode = input_49_mode_0, x = linear_15_cast_fp16)[name = tensor("input_49_cast_fp16")]; + tensor linear3_2_weight_to_fp16 = const()[name = tensor("linear3_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(93441920)))]; + tensor linear_16_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = linear3_2_weight_to_fp16, x = input_49_cast_fp16)[name = tensor("linear_16_cast_fp16")]; + tensor input_51_cast_fp16 = add(x = input_45_cast_fp16, y = linear_16_cast_fp16)[name = tensor("input_51_cast_fp16")]; + tensor input_53_axes_0 = const()[name = tensor("input_53_axes_0"), val = tensor([-1])]; + tensor norm4_1_weight_to_fp16 = const()[name = tensor("norm4_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(101830592)))]; + tensor norm4_1_bias_to_fp16 = const()[name = tensor("norm4_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(101832704)))]; + tensor var_1604_to_fp16 = const()[name = tensor("op_1604_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_53_cast_fp16 = layer_norm(axes = input_53_axes_0, beta = norm4_1_bias_to_fp16, epsilon = var_1604_to_fp16, gamma = norm4_1_weight_to_fp16, x = input_51_cast_fp16)[name = tensor("input_53_cast_fp16")]; + tensor attn4_in_proj_weight_to_fp16 = const()[name = tensor("attn4_in_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(101834816)))]; + tensor linear_17_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = attn4_in_proj_weight_to_fp16, x = input_53_cast_fp16)[name = tensor("linear_17_cast_fp16")]; + tensor var_1618 = const()[name = tensor("op_1618"), val = tensor([1, 1, 3, 16, 64])]; + tensor qkv_9_cast_fp16 = reshape(shape = var_1618, x = linear_17_cast_fp16)[name = tensor("qkv_9_cast_fp16")]; + tensor q_25_begin_0 = const()[name = tensor("q_25_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor q_25_end_0 = const()[name = tensor("q_25_end_0"), val = tensor([1, 1, 1, 16, 64])]; + tensor q_25_end_mask_0 = const()[name = tensor("q_25_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor q_25_squeeze_mask_0 = const()[name = tensor("q_25_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor q_25_cast_fp16 = slice_by_index(begin = q_25_begin_0, end = q_25_end_0, end_mask = q_25_end_mask_0, squeeze_mask = q_25_squeeze_mask_0, x = qkv_9_cast_fp16)[name = tensor("q_25_cast_fp16")]; + tensor k_17_begin_0 = const()[name = tensor("k_17_begin_0"), val = tensor([0, 0, 1, 0, 0])]; + tensor k_17_end_0 = const()[name = tensor("k_17_end_0"), val = tensor([1, 1, 2, 16, 64])]; + tensor k_17_end_mask_0 = const()[name = tensor("k_17_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor k_17_squeeze_mask_0 = const()[name = tensor("k_17_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor k_17_cast_fp16 = slice_by_index(begin = k_17_begin_0, end = k_17_end_0, end_mask = k_17_end_mask_0, squeeze_mask = k_17_squeeze_mask_0, x = qkv_9_cast_fp16)[name = tensor("k_17_cast_fp16")]; + tensor v_9_begin_0 = const()[name = tensor("v_9_begin_0"), val = tensor([0, 0, 2, 0, 0])]; + tensor v_9_end_0 = const()[name = tensor("v_9_end_0"), val = tensor([1, 1, 3, 16, 64])]; + tensor v_9_end_mask_0 = const()[name = tensor("v_9_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor v_9_squeeze_mask_0 = const()[name = tensor("v_9_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor v_9_cast_fp16 = slice_by_index(begin = v_9_begin_0, end = v_9_end_0, end_mask = v_9_end_mask_0, squeeze_mask = v_9_squeeze_mask_0, x = qkv_9_cast_fp16)[name = tensor("v_9_cast_fp16")]; + tensor var_1684 = const()[name = tensor("op_1684"), val = tensor([1, 1, 1, 1])]; + tensor position4_to_fp16_dtype_0 = const()[name = tensor("position4_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor position4_to_fp16 = cast(dtype = position4_to_fp16_dtype_0, x = position4)[name = tensor("cast_119")]; + tensor ts_29_cast_fp16 = reshape(shape = var_1684, x = position4_to_fp16)[name = tensor("ts_29_cast_fp16")]; + tensor var_1691 = const()[name = tensor("op_1691"), val = tensor([1, 1, 16, 32, 2])]; + tensor q_complex_9_cast_fp16 = reshape(shape = var_1691, x = q_25_cast_fp16)[name = tensor("q_complex_9_cast_fp16")]; + tensor var_1698 = const()[name = tensor("op_1698"), val = tensor([1, 1, 16, 32, 2])]; + tensor k_complex_9_cast_fp16 = reshape(shape = var_1698, x = k_17_cast_fp16)[name = tensor("k_complex_9_cast_fp16")]; + tensor var_1702_begin_0 = const()[name = tensor("op_1702_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_1702_end_0 = const()[name = tensor("op_1702_end_0"), val = tensor([1, 1, 16, 32, 1])]; + tensor var_1702_end_mask_0 = const()[name = tensor("op_1702_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_1702_squeeze_mask_0 = const()[name = tensor("op_1702_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_1702_cast_fp16 = slice_by_index(begin = var_1702_begin_0, end = var_1702_end_0, end_mask = var_1702_end_mask_0, squeeze_mask = var_1702_squeeze_mask_0, x = q_complex_9_cast_fp16)[name = tensor("op_1702_cast_fp16")]; + tensor var_1710_begin_0 = const()[name = tensor("op_1710_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_1710_end_0 = const()[name = tensor("op_1710_end_0"), val = tensor([1, 1, 16, 32, 2])]; + tensor var_1710_end_mask_0 = const()[name = tensor("op_1710_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_1710_squeeze_mask_0 = const()[name = tensor("op_1710_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_1710_cast_fp16 = slice_by_index(begin = var_1710_begin_0, end = var_1710_end_0, end_mask = var_1710_end_mask_0, squeeze_mask = var_1710_squeeze_mask_0, x = q_complex_9_cast_fp16)[name = tensor("op_1710_cast_fp16")]; + tensor var_1718_begin_0 = const()[name = tensor("op_1718_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_1718_end_0 = const()[name = tensor("op_1718_end_0"), val = tensor([1, 1, 16, 32, 1])]; + tensor var_1718_end_mask_0 = const()[name = tensor("op_1718_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_1718_squeeze_mask_0 = const()[name = tensor("op_1718_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_1718_cast_fp16 = slice_by_index(begin = var_1718_begin_0, end = var_1718_end_0, end_mask = var_1718_end_mask_0, squeeze_mask = var_1718_squeeze_mask_0, x = k_complex_9_cast_fp16)[name = tensor("op_1718_cast_fp16")]; + tensor var_1726_begin_0 = const()[name = tensor("op_1726_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_1726_end_0 = const()[name = tensor("op_1726_end_0"), val = tensor([1, 1, 16, 32, 2])]; + tensor var_1726_end_mask_0 = const()[name = tensor("op_1726_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_1726_squeeze_mask_0 = const()[name = tensor("op_1726_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_1726_cast_fp16 = slice_by_index(begin = var_1726_begin_0, end = var_1726_end_0, end_mask = var_1726_end_mask_0, squeeze_mask = var_1726_squeeze_mask_0, x = k_complex_9_cast_fp16)[name = tensor("op_1726_cast_fp16")]; + tensor freqs_9_to_fp16 = const()[name = tensor("freqs_9_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(108126336)))]; + tensor var_1732_cast_fp16 = mul(x = freqs_9_to_fp16, y = ts_29_cast_fp16)[name = tensor("op_1732_cast_fp16")]; + tensor rotr_9_cast_fp16 = cos(x = var_1732_cast_fp16)[name = tensor("rotr_9_cast_fp16")]; + tensor roti_9_cast_fp16 = sin(x = var_1732_cast_fp16)[name = tensor("roti_9_cast_fp16")]; + tensor var_1736_cast_fp16 = mul(x = var_1702_cast_fp16, y = rotr_9_cast_fp16)[name = tensor("op_1736_cast_fp16")]; + tensor var_1737_cast_fp16 = mul(x = var_1710_cast_fp16, y = roti_9_cast_fp16)[name = tensor("op_1737_cast_fp16")]; + tensor qor_17_cast_fp16 = sub(x = var_1736_cast_fp16, y = var_1737_cast_fp16)[name = tensor("qor_17_cast_fp16")]; + tensor var_1740_cast_fp16 = mul(x = var_1702_cast_fp16, y = roti_9_cast_fp16)[name = tensor("op_1740_cast_fp16")]; + tensor var_1741_cast_fp16 = mul(x = var_1710_cast_fp16, y = rotr_9_cast_fp16)[name = tensor("op_1741_cast_fp16")]; + tensor qoi_17_cast_fp16 = add(x = var_1740_cast_fp16, y = var_1741_cast_fp16)[name = tensor("qoi_17_cast_fp16")]; + tensor var_1744_cast_fp16 = mul(x = var_1718_cast_fp16, y = rotr_9_cast_fp16)[name = tensor("op_1744_cast_fp16")]; + tensor var_1745_cast_fp16 = mul(x = var_1726_cast_fp16, y = roti_9_cast_fp16)[name = tensor("op_1745_cast_fp16")]; + tensor kor_17_cast_fp16 = sub(x = var_1744_cast_fp16, y = var_1745_cast_fp16)[name = tensor("kor_17_cast_fp16")]; + tensor var_1748_cast_fp16 = mul(x = var_1718_cast_fp16, y = roti_9_cast_fp16)[name = tensor("op_1748_cast_fp16")]; + tensor var_1749_cast_fp16 = mul(x = var_1726_cast_fp16, y = rotr_9_cast_fp16)[name = tensor("op_1749_cast_fp16")]; + tensor koi_17_cast_fp16 = add(x = var_1748_cast_fp16, y = var_1749_cast_fp16)[name = tensor("koi_17_cast_fp16")]; + tensor qo_9_axis_0 = const()[name = tensor("qo_9_axis_0"), val = tensor(-1)]; + tensor qo_9_cast_fp16 = stack(axis = qo_9_axis_0, values = (qor_17_cast_fp16, qoi_17_cast_fp16))[name = tensor("qo_9_cast_fp16")]; + tensor ko_9_axis_0 = const()[name = tensor("ko_9_axis_0"), val = tensor(-1)]; + tensor ko_9_cast_fp16 = stack(axis = ko_9_axis_0, values = (kor_17_cast_fp16, koi_17_cast_fp16))[name = tensor("ko_9_cast_fp16")]; + tensor var_1782 = const()[name = tensor("op_1782"), val = tensor([1, 1, 16, 64])]; + tensor q_27_cast_fp16 = reshape(shape = var_1782, x = qo_9_cast_fp16)[name = tensor("q_27_cast_fp16")]; + tensor var_1788 = const()[name = tensor("op_1788"), val = tensor([1, 1, 16, 64])]; + tensor k_19_cast_fp16 = reshape(shape = var_1788, x = ko_9_cast_fp16)[name = tensor("k_19_cast_fp16")]; + tensor _inversed_1813_y_0_to_fp16 = const()[name = tensor("_inversed_1813_y_0_to_fp16"), val = tensor(0x1p-9)]; + tensor _inversed_1813_cast_fp16 = mul(x = ts_29_cast_fp16, y = _inversed_1813_y_0_to_fp16)[name = tensor("_inversed_1813_cast_fp16")]; + tensor var_1814_cast_fp16 = floor(x = _inversed_1813_cast_fp16)[name = tensor("op_1814_cast_fp16")]; + tensor var_1815_to_fp16 = const()[name = tensor("op_1815_to_fp16"), val = tensor(0x1p+9)]; + tensor var_1816_cast_fp16 = mul(x = var_1814_cast_fp16, y = var_1815_to_fp16)[name = tensor("op_1816_cast_fp16")]; + tensor write_indices_float_19_cast_fp16 = sub(x = ts_29_cast_fp16, y = var_1816_cast_fp16)[name = tensor("write_indices_float_19_cast_fp16")]; + tensor var_1823_dtype_0 = const()[name = tensor("op_1823_dtype_0"), val = tensor("int32")]; + tensor write_indices_9_reps_0 = const()[name = tensor("write_indices_9_reps_0"), val = tensor([1, 1, 16, 64])]; + tensor write_indices_float_19_cast_fp16_to_int32 = cast(dtype = var_1823_dtype_0, x = write_indices_float_19_cast_fp16)[name = tensor("cast_118")]; + tensor write_indices_9 = tile(reps = write_indices_9_reps_0, x = write_indices_float_19_cast_fp16_to_int32)[name = tensor("write_indices_9")]; + tensor var_1833_begin_0 = const()[name = tensor("op_1833_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_1833_end_0 = const()[name = tensor("op_1833_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor var_1833_end_mask_0 = const()[name = tensor("op_1833_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_1833_squeeze_mask_0 = const()[name = tensor("op_1833_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor cache4_to_fp16_dtype_0 = const()[name = tensor("cache4_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor cache4_to_fp16 = cast(dtype = cache4_to_fp16_dtype_0, x = cache4)[name = tensor("cast_117")]; + tensor var_1833_cast_fp16 = slice_by_index(begin = var_1833_begin_0, end = var_1833_end_0, end_mask = var_1833_end_mask_0, squeeze_mask = var_1833_squeeze_mask_0, x = cache4_to_fp16)[name = tensor("op_1833_cast_fp16")]; + tensor var_1835_axis_0 = const()[name = tensor("op_1835_axis_0"), val = tensor(1)]; + tensor var_1835_mode_0 = const()[name = tensor("op_1835_mode_0"), val = tensor("update")]; + tensor var_1835_validate_indices_0 = const()[name = tensor("op_1835_validate_indices_0"), val = tensor(false)]; + tensor var_1835_cast_fp16 = scatter_along_axis(axis = var_1835_axis_0, data = var_1833_cast_fp16, indices = write_indices_9, mode = var_1835_mode_0, updates = k_19_cast_fp16, validate_indices = var_1835_validate_indices_0)[name = tensor("op_1835_cast_fp16")]; + tensor concat_30 = const()[name = tensor("concat_30"), val = tensor([0, 0, 0, 0, 0])]; + tensor concat_31 = const()[name = tensor("concat_31"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_9_internal_tensor_assign_1_stride_0 = const()[name = tensor("new_cache_9_internal_tensor_assign_1_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_9_internal_tensor_assign_1_begin_mask_0 = const()[name = tensor("new_cache_9_internal_tensor_assign_1_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_9_internal_tensor_assign_1_end_mask_0 = const()[name = tensor("new_cache_9_internal_tensor_assign_1_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_9_internal_tensor_assign_1_squeeze_mask_0 = const()[name = tensor("new_cache_9_internal_tensor_assign_1_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_20 = const()[name = tensor("shape_20"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_8 = const()[name = tensor("reduce_prod_8"), val = tensor(1048576)]; + tensor range_1d_8_start_0 = const()[name = tensor("range_1d_8_start_0"), val = tensor(0)]; + tensor range_1d_8_step_0 = const()[name = tensor("range_1d_8_step_0"), val = tensor(1)]; + tensor range_1d_8 = range_1d(end = reduce_prod_8, start = range_1d_8_start_0, step = range_1d_8_step_0)[name = tensor("range_1d_8")]; + tensor reshape_40 = reshape(shape = shape_20, x = range_1d_8)[name = tensor("reshape_40")]; + tensor slice_by_index_8 = slice_by_index(begin = concat_30, begin_mask = new_cache_9_internal_tensor_assign_1_begin_mask_0, end = concat_31, end_mask = new_cache_9_internal_tensor_assign_1_end_mask_0, squeeze_mask = new_cache_9_internal_tensor_assign_1_squeeze_mask_0, stride = new_cache_9_internal_tensor_assign_1_stride_0, x = reshape_40)[name = tensor("slice_by_index_8")]; + tensor reshape_41_shape_0 = const()[name = tensor("reshape_41_shape_0"), val = tensor([-1])]; + tensor reshape_41 = reshape(shape = reshape_41_shape_0, x = slice_by_index_8)[name = tensor("reshape_41")]; + tensor reshape_42_shape_0 = const()[name = tensor("reshape_42_shape_0"), val = tensor([-1])]; + tensor reshape_42_cast_fp16 = reshape(shape = reshape_42_shape_0, x = var_1835_cast_fp16)[name = tensor("reshape_42_cast_fp16")]; + tensor reshape_43_shape_0 = const()[name = tensor("reshape_43_shape_0"), val = tensor([-1])]; + tensor reshape_43_cast_fp16 = reshape(shape = reshape_43_shape_0, x = cache4_to_fp16)[name = tensor("reshape_43_cast_fp16")]; + tensor scatter_8_mode_0 = const()[name = tensor("scatter_8_mode_0"), val = tensor("update")]; + tensor scatter_8_axis_0 = const()[name = tensor("scatter_8_axis_0"), val = tensor(0)]; + tensor scatter_8_validate_indices_0 = const()[name = tensor("scatter_8_validate_indices_0"), val = tensor(false)]; + tensor scatter_8_cast_fp16 = scatter(axis = scatter_8_axis_0, data = reshape_43_cast_fp16, indices = reshape_41, mode = scatter_8_mode_0, updates = reshape_42_cast_fp16, validate_indices = scatter_8_validate_indices_0)[name = tensor("scatter_8_cast_fp16")]; + tensor reshape_44_cast_fp16 = reshape(shape = shape_20, x = scatter_8_cast_fp16)[name = tensor("reshape_44_cast_fp16")]; + tensor var_1843_begin_0 = const()[name = tensor("op_1843_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor var_1843_end_0 = const()[name = tensor("op_1843_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor var_1843_end_mask_0 = const()[name = tensor("op_1843_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_1843_squeeze_mask_0 = const()[name = tensor("op_1843_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor var_1843_cast_fp16 = slice_by_index(begin = var_1843_begin_0, end = var_1843_end_0, end_mask = var_1843_end_mask_0, squeeze_mask = var_1843_squeeze_mask_0, x = reshape_44_cast_fp16)[name = tensor("op_1843_cast_fp16")]; + tensor var_1845_axis_0 = const()[name = tensor("op_1845_axis_0"), val = tensor(1)]; + tensor var_1845_mode_0 = const()[name = tensor("op_1845_mode_0"), val = tensor("update")]; + tensor var_1845_validate_indices_0 = const()[name = tensor("op_1845_validate_indices_0"), val = tensor(false)]; + tensor var_1845_cast_fp16 = scatter_along_axis(axis = var_1845_axis_0, data = var_1843_cast_fp16, indices = write_indices_9, mode = var_1845_mode_0, updates = v_9_cast_fp16, validate_indices = var_1845_validate_indices_0)[name = tensor("op_1845_cast_fp16")]; + tensor concat_32 = const()[name = tensor("concat_32"), val = tensor([1, 0, 0, 0, 0])]; + tensor concat_33 = const()[name = tensor("concat_33"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_9_internal_tensor_assign_2_stride_0 = const()[name = tensor("new_cache_9_internal_tensor_assign_2_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_9_internal_tensor_assign_2_begin_mask_0 = const()[name = tensor("new_cache_9_internal_tensor_assign_2_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_9_internal_tensor_assign_2_end_mask_0 = const()[name = tensor("new_cache_9_internal_tensor_assign_2_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_9_internal_tensor_assign_2_squeeze_mask_0 = const()[name = tensor("new_cache_9_internal_tensor_assign_2_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_21 = const()[name = tensor("shape_21"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_9 = const()[name = tensor("reduce_prod_9"), val = tensor(1048576)]; + tensor range_1d_9_start_0 = const()[name = tensor("range_1d_9_start_0"), val = tensor(0)]; + tensor range_1d_9_step_0 = const()[name = tensor("range_1d_9_step_0"), val = tensor(1)]; + tensor range_1d_9 = range_1d(end = reduce_prod_9, start = range_1d_9_start_0, step = range_1d_9_step_0)[name = tensor("range_1d_9")]; + tensor reshape_45 = reshape(shape = shape_21, x = range_1d_9)[name = tensor("reshape_45")]; + tensor slice_by_index_9 = slice_by_index(begin = concat_32, begin_mask = new_cache_9_internal_tensor_assign_2_begin_mask_0, end = concat_33, end_mask = new_cache_9_internal_tensor_assign_2_end_mask_0, squeeze_mask = new_cache_9_internal_tensor_assign_2_squeeze_mask_0, stride = new_cache_9_internal_tensor_assign_2_stride_0, x = reshape_45)[name = tensor("slice_by_index_9")]; + tensor reshape_46_shape_0 = const()[name = tensor("reshape_46_shape_0"), val = tensor([-1])]; + tensor reshape_46 = reshape(shape = reshape_46_shape_0, x = slice_by_index_9)[name = tensor("reshape_46")]; + tensor reshape_47_shape_0 = const()[name = tensor("reshape_47_shape_0"), val = tensor([-1])]; + tensor reshape_47_cast_fp16 = reshape(shape = reshape_47_shape_0, x = var_1845_cast_fp16)[name = tensor("reshape_47_cast_fp16")]; + tensor reshape_48_shape_0 = const()[name = tensor("reshape_48_shape_0"), val = tensor([-1])]; + tensor reshape_48_cast_fp16 = reshape(shape = reshape_48_shape_0, x = reshape_44_cast_fp16)[name = tensor("reshape_48_cast_fp16")]; + tensor scatter_9_mode_0 = const()[name = tensor("scatter_9_mode_0"), val = tensor("update")]; + tensor scatter_9_axis_0 = const()[name = tensor("scatter_9_axis_0"), val = tensor(0)]; + tensor scatter_9_validate_indices_0 = const()[name = tensor("scatter_9_validate_indices_0"), val = tensor(false)]; + tensor scatter_9_cast_fp16 = scatter(axis = scatter_9_axis_0, data = reshape_48_cast_fp16, indices = reshape_46, mode = scatter_9_mode_0, updates = reshape_47_cast_fp16, validate_indices = scatter_9_validate_indices_0)[name = tensor("scatter_9_cast_fp16")]; + tensor reshape_49_cast_fp16 = reshape(shape = shape_21, x = scatter_9_cast_fp16)[name = tensor("reshape_49_cast_fp16")]; + tensor reshape_49_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("reshape_49_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor keys_25_begin_0 = const()[name = tensor("keys_25_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor keys_25_end_0 = const()[name = tensor("keys_25_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor keys_25_end_mask_0 = const()[name = tensor("keys_25_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor keys_25_squeeze_mask_0 = const()[name = tensor("keys_25_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor keys_25_cast_fp16 = slice_by_index(begin = keys_25_begin_0, end = keys_25_end_0, end_mask = keys_25_end_mask_0, squeeze_mask = keys_25_squeeze_mask_0, x = reshape_49_cast_fp16)[name = tensor("keys_25_cast_fp16")]; + tensor values_25_begin_0 = const()[name = tensor("values_25_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor values_25_end_0 = const()[name = tensor("values_25_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor values_25_end_mask_0 = const()[name = tensor("values_25_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor values_25_squeeze_mask_0 = const()[name = tensor("values_25_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor values_25_cast_fp16 = slice_by_index(begin = values_25_begin_0, end = values_25_end_0, end_mask = values_25_end_mask_0, squeeze_mask = values_25_squeeze_mask_0, x = reshape_49_cast_fp16)[name = tensor("values_25_cast_fp16")]; + tensor var_1857_cast_fp16 = not_equal(x = keys_25_cast_fp16, y = keys_25_cast_fp16)[name = tensor("op_1857_cast_fp16")]; + tensor keys_27_cast_fp16 = select(a = var_319_to_fp16, b = keys_25_cast_fp16, cond = var_1857_cast_fp16)[name = tensor("keys_27_cast_fp16")]; + tensor var_1865_cast_fp16 = not_equal(x = values_25_cast_fp16, y = values_25_cast_fp16)[name = tensor("op_1865_cast_fp16")]; + tensor values_27_cast_fp16 = select(a = var_319_to_fp16, b = values_25_cast_fp16, cond = var_1865_cast_fp16)[name = tensor("values_27_cast_fp16")]; + tensor var_1889 = const()[name = tensor("op_1889"), val = tensor([0, 2, 1, 3])]; + tensor var_1905 = const()[name = tensor("op_1905"), val = tensor([1, 1, 1])]; + tensor var_1906_cast_fp16 = reshape(shape = var_1905, x = position4_to_fp16)[name = tensor("op_1906_cast_fp16")]; + tensor var_1926_to_fp16 = const()[name = tensor("op_1926_to_fp16"), val = tensor(0x1p+0)]; + tensor valid_len_9_cast_fp16 = add(x = var_1906_cast_fp16, y = var_1926_to_fp16)[name = tensor("valid_len_9_cast_fp16")]; + tensor valid_mask_9_cast_fp16 = less(x = k_positions_1_promoted_to_fp16, y = valid_len_9_cast_fp16)[name = tensor("valid_mask_9_cast_fp16")]; + tensor causal_mask_9_cast_fp16 = less_equal(x = k_positions_1_promoted_to_fp16, y = var_1906_cast_fp16)[name = tensor("causal_mask_9_cast_fp16")]; + tensor attn_mask_17 = logical_and(x = valid_mask_9_cast_fp16, y = causal_mask_9_cast_fp16)[name = tensor("attn_mask_17")]; + tensor attn_mask_19_axes_0 = const()[name = tensor("attn_mask_19_axes_0"), val = tensor([1])]; + tensor attn_mask_19 = expand_dims(axes = attn_mask_19_axes_0, x = attn_mask_17)[name = tensor("attn_mask_19")]; + tensor var_1944_transpose_x_0 = const()[name = tensor("op_1944_transpose_x_0"), val = tensor(false)]; + tensor var_1944_transpose_y_0 = const()[name = tensor("op_1944_transpose_y_0"), val = tensor(false)]; + tensor transpose_26_perm_0 = const()[name = tensor("transpose_26_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_27_perm_0 = const()[name = tensor("transpose_27_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_27 = transpose(perm = transpose_27_perm_0, x = keys_27_cast_fp16)[name = tensor("transpose_35")]; + tensor transpose_26 = transpose(perm = transpose_26_perm_0, x = q_27_cast_fp16)[name = tensor("transpose_36")]; + tensor var_1944_cast_fp16 = matmul(transpose_x = var_1944_transpose_x_0, transpose_y = var_1944_transpose_y_0, x = transpose_26, y = transpose_27)[name = tensor("op_1944_cast_fp16")]; + tensor var_1938_to_fp16 = const()[name = tensor("op_1938_to_fp16"), val = tensor([0x1p-3])]; + tensor attn_weights_25_cast_fp16 = mul(x = var_1944_cast_fp16, y = var_1938_to_fp16)[name = tensor("attn_weights_25_cast_fp16")]; + tensor var_1946 = logical_not(x = attn_mask_19)[name = tensor("op_1946")]; + tensor var_1947_to_fp16 = const()[name = tensor("op_1947_to_fp16"), val = tensor(-inf)]; + tensor attn_weights_27_cast_fp16 = select(a = var_1947_to_fp16, b = attn_weights_25_cast_fp16, cond = var_1946)[name = tensor("attn_weights_27_cast_fp16")]; + tensor var_1949 = const()[name = tensor("op_1949"), val = tensor(-1)]; + tensor attn_weights_29_cast_fp16 = softmax(axis = var_1949, x = attn_weights_27_cast_fp16)[name = tensor("attn_weights_29_cast_fp16")]; + tensor attn_output_9_transpose_x_0 = const()[name = tensor("attn_output_9_transpose_x_0"), val = tensor(false)]; + tensor attn_output_9_transpose_y_0 = const()[name = tensor("attn_output_9_transpose_y_0"), val = tensor(false)]; + tensor values_29_cast_fp16 = transpose(perm = var_1889, x = values_27_cast_fp16)[name = tensor("transpose_37")]; + tensor attn_output_9_cast_fp16 = matmul(transpose_x = attn_output_9_transpose_x_0, transpose_y = attn_output_9_transpose_y_0, x = attn_weights_29_cast_fp16, y = values_29_cast_fp16)[name = tensor("attn_output_9_cast_fp16")]; + tensor var_1957 = const()[name = tensor("op_1957"), val = tensor([0, 2, 1, 3])]; + tensor var_1962 = const()[name = tensor("op_1962"), val = tensor([1, 1, 1024])]; + tensor var_1958_cast_fp16 = transpose(perm = var_1957, x = attn_output_9_cast_fp16)[name = tensor("transpose_34")]; + tensor input_55_cast_fp16 = reshape(shape = var_1962, x = var_1958_cast_fp16)[name = tensor("input_55_cast_fp16")]; + tensor attn4_out_proj_weight_to_fp16 = const()[name = tensor("attn4_out_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(108126464)))]; + tensor linear_18_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = attn4_out_proj_weight_to_fp16, x = input_55_cast_fp16)[name = tensor("linear_18_cast_fp16")]; + tensor var_1968_to_fp16 = const()[name = tensor("op_1968_to_fp16"), val = tensor(0x1p+0)]; + tensor var_1969_cast_fp16 = add(x = position4_to_fp16, y = var_1968_to_fp16)[name = tensor("op_1969_cast_fp16")]; + tensor var_1969_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_1969_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor input_57_cast_fp16 = add(x = input_51_cast_fp16, y = linear_18_cast_fp16)[name = tensor("input_57_cast_fp16")]; + tensor input_59_axes_0 = const()[name = tensor("input_59_axes_0"), val = tensor([-1])]; + tensor norm4_2_weight_to_fp16 = const()[name = tensor("norm4_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(110223680)))]; + tensor norm4_2_bias_to_fp16 = const()[name = tensor("norm4_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(110225792)))]; + tensor var_1973_to_fp16 = const()[name = tensor("op_1973_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_59_cast_fp16 = layer_norm(axes = input_59_axes_0, beta = norm4_2_bias_to_fp16, epsilon = var_1973_to_fp16, gamma = norm4_2_weight_to_fp16, x = input_57_cast_fp16)[name = tensor("input_59_cast_fp16")]; + tensor linear4_1_weight_to_fp16 = const()[name = tensor("linear4_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(110227904)))]; + tensor linear_19_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = linear4_1_weight_to_fp16, x = input_59_cast_fp16)[name = tensor("linear_19_cast_fp16")]; + tensor input_61_mode_0 = const()[name = tensor("input_61_mode_0"), val = tensor("EXACT")]; + tensor input_61_cast_fp16 = gelu(mode = input_61_mode_0, x = linear_19_cast_fp16)[name = tensor("input_61_cast_fp16")]; + tensor linear4_2_weight_to_fp16 = const()[name = tensor("linear4_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(118616576)))]; + tensor linear_20_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = linear4_2_weight_to_fp16, x = input_61_cast_fp16)[name = tensor("linear_20_cast_fp16")]; + tensor input_63_cast_fp16 = add(x = input_57_cast_fp16, y = linear_20_cast_fp16)[name = tensor("input_63_cast_fp16")]; + tensor input_65_axes_0 = const()[name = tensor("input_65_axes_0"), val = tensor([-1])]; + tensor norm5_1_weight_to_fp16 = const()[name = tensor("norm5_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(127005248)))]; + tensor norm5_1_bias_to_fp16 = const()[name = tensor("norm5_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(127007360)))]; + tensor var_1990_to_fp16 = const()[name = tensor("op_1990_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_65_cast_fp16 = layer_norm(axes = input_65_axes_0, beta = norm5_1_bias_to_fp16, epsilon = var_1990_to_fp16, gamma = norm5_1_weight_to_fp16, x = input_63_cast_fp16)[name = tensor("input_65_cast_fp16")]; + tensor attn5_in_proj_weight_to_fp16 = const()[name = tensor("attn5_in_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(127009472)))]; + tensor linear_21_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = attn5_in_proj_weight_to_fp16, x = input_65_cast_fp16)[name = tensor("linear_21_cast_fp16")]; + tensor var_2004 = const()[name = tensor("op_2004"), val = tensor([1, 1, 3, 16, 64])]; + tensor qkv_cast_fp16 = reshape(shape = var_2004, x = linear_21_cast_fp16)[name = tensor("qkv_cast_fp16")]; + tensor q_31_begin_0 = const()[name = tensor("q_31_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor q_31_end_0 = const()[name = tensor("q_31_end_0"), val = tensor([1, 1, 1, 16, 64])]; + tensor q_31_end_mask_0 = const()[name = tensor("q_31_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor q_31_squeeze_mask_0 = const()[name = tensor("q_31_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor q_31_cast_fp16 = slice_by_index(begin = q_31_begin_0, end = q_31_end_0, end_mask = q_31_end_mask_0, squeeze_mask = q_31_squeeze_mask_0, x = qkv_cast_fp16)[name = tensor("q_31_cast_fp16")]; + tensor k_21_begin_0 = const()[name = tensor("k_21_begin_0"), val = tensor([0, 0, 1, 0, 0])]; + tensor k_21_end_0 = const()[name = tensor("k_21_end_0"), val = tensor([1, 1, 2, 16, 64])]; + tensor k_21_end_mask_0 = const()[name = tensor("k_21_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor k_21_squeeze_mask_0 = const()[name = tensor("k_21_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor k_21_cast_fp16 = slice_by_index(begin = k_21_begin_0, end = k_21_end_0, end_mask = k_21_end_mask_0, squeeze_mask = k_21_squeeze_mask_0, x = qkv_cast_fp16)[name = tensor("k_21_cast_fp16")]; + tensor v_begin_0 = const()[name = tensor("v_begin_0"), val = tensor([0, 0, 2, 0, 0])]; + tensor v_end_0 = const()[name = tensor("v_end_0"), val = tensor([1, 1, 3, 16, 64])]; + tensor v_end_mask_0 = const()[name = tensor("v_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor v_squeeze_mask_0 = const()[name = tensor("v_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor v_cast_fp16 = slice_by_index(begin = v_begin_0, end = v_end_0, end_mask = v_end_mask_0, squeeze_mask = v_squeeze_mask_0, x = qkv_cast_fp16)[name = tensor("v_cast_fp16")]; + tensor var_2070 = const()[name = tensor("op_2070"), val = tensor([1, 1, 1, 1])]; + tensor position5_to_fp16_dtype_0 = const()[name = tensor("position5_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor position5_to_fp16 = cast(dtype = position5_to_fp16_dtype_0, x = position5)[name = tensor("cast_114")]; + tensor ts_cast_fp16 = reshape(shape = var_2070, x = position5_to_fp16)[name = tensor("ts_cast_fp16")]; + tensor var_2077 = const()[name = tensor("op_2077"), val = tensor([1, 1, 16, 32, 2])]; + tensor q_complex_cast_fp16 = reshape(shape = var_2077, x = q_31_cast_fp16)[name = tensor("q_complex_cast_fp16")]; + tensor var_2084 = const()[name = tensor("op_2084"), val = tensor([1, 1, 16, 32, 2])]; + tensor k_complex_cast_fp16 = reshape(shape = var_2084, x = k_21_cast_fp16)[name = tensor("k_complex_cast_fp16")]; + tensor var_2088_begin_0 = const()[name = tensor("op_2088_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_2088_end_0 = const()[name = tensor("op_2088_end_0"), val = tensor([1, 1, 16, 32, 1])]; + tensor var_2088_end_mask_0 = const()[name = tensor("op_2088_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_2088_squeeze_mask_0 = const()[name = tensor("op_2088_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_2088_cast_fp16 = slice_by_index(begin = var_2088_begin_0, end = var_2088_end_0, end_mask = var_2088_end_mask_0, squeeze_mask = var_2088_squeeze_mask_0, x = q_complex_cast_fp16)[name = tensor("op_2088_cast_fp16")]; + tensor var_2096_begin_0 = const()[name = tensor("op_2096_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_2096_end_0 = const()[name = tensor("op_2096_end_0"), val = tensor([1, 1, 16, 32, 2])]; + tensor var_2096_end_mask_0 = const()[name = tensor("op_2096_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_2096_squeeze_mask_0 = const()[name = tensor("op_2096_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_2096_cast_fp16 = slice_by_index(begin = var_2096_begin_0, end = var_2096_end_0, end_mask = var_2096_end_mask_0, squeeze_mask = var_2096_squeeze_mask_0, x = q_complex_cast_fp16)[name = tensor("op_2096_cast_fp16")]; + tensor var_2104_begin_0 = const()[name = tensor("op_2104_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_2104_end_0 = const()[name = tensor("op_2104_end_0"), val = tensor([1, 1, 16, 32, 1])]; + tensor var_2104_end_mask_0 = const()[name = tensor("op_2104_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_2104_squeeze_mask_0 = const()[name = tensor("op_2104_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_2104_cast_fp16 = slice_by_index(begin = var_2104_begin_0, end = var_2104_end_0, end_mask = var_2104_end_mask_0, squeeze_mask = var_2104_squeeze_mask_0, x = k_complex_cast_fp16)[name = tensor("op_2104_cast_fp16")]; + tensor var_2112_begin_0 = const()[name = tensor("op_2112_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_2112_end_0 = const()[name = tensor("op_2112_end_0"), val = tensor([1, 1, 16, 32, 2])]; + tensor var_2112_end_mask_0 = const()[name = tensor("op_2112_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_2112_squeeze_mask_0 = const()[name = tensor("op_2112_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_2112_cast_fp16 = slice_by_index(begin = var_2112_begin_0, end = var_2112_end_0, end_mask = var_2112_end_mask_0, squeeze_mask = var_2112_squeeze_mask_0, x = k_complex_cast_fp16)[name = tensor("op_2112_cast_fp16")]; + tensor freqs_to_fp16 = const()[name = tensor("freqs_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(133300992)))]; + tensor var_2118_cast_fp16 = mul(x = freqs_to_fp16, y = ts_cast_fp16)[name = tensor("op_2118_cast_fp16")]; + tensor rotr_cast_fp16 = cos(x = var_2118_cast_fp16)[name = tensor("rotr_cast_fp16")]; + tensor roti_cast_fp16 = sin(x = var_2118_cast_fp16)[name = tensor("roti_cast_fp16")]; + tensor var_2122_cast_fp16 = mul(x = var_2088_cast_fp16, y = rotr_cast_fp16)[name = tensor("op_2122_cast_fp16")]; + tensor var_2123_cast_fp16 = mul(x = var_2096_cast_fp16, y = roti_cast_fp16)[name = tensor("op_2123_cast_fp16")]; + tensor qor_21_cast_fp16 = sub(x = var_2122_cast_fp16, y = var_2123_cast_fp16)[name = tensor("qor_21_cast_fp16")]; + tensor var_2126_cast_fp16 = mul(x = var_2088_cast_fp16, y = roti_cast_fp16)[name = tensor("op_2126_cast_fp16")]; + tensor var_2127_cast_fp16 = mul(x = var_2096_cast_fp16, y = rotr_cast_fp16)[name = tensor("op_2127_cast_fp16")]; + tensor qoi_21_cast_fp16 = add(x = var_2126_cast_fp16, y = var_2127_cast_fp16)[name = tensor("qoi_21_cast_fp16")]; + tensor var_2130_cast_fp16 = mul(x = var_2104_cast_fp16, y = rotr_cast_fp16)[name = tensor("op_2130_cast_fp16")]; + tensor var_2131_cast_fp16 = mul(x = var_2112_cast_fp16, y = roti_cast_fp16)[name = tensor("op_2131_cast_fp16")]; + tensor kor_21_cast_fp16 = sub(x = var_2130_cast_fp16, y = var_2131_cast_fp16)[name = tensor("kor_21_cast_fp16")]; + tensor var_2134_cast_fp16 = mul(x = var_2104_cast_fp16, y = roti_cast_fp16)[name = tensor("op_2134_cast_fp16")]; + tensor var_2135_cast_fp16 = mul(x = var_2112_cast_fp16, y = rotr_cast_fp16)[name = tensor("op_2135_cast_fp16")]; + tensor koi_21_cast_fp16 = add(x = var_2134_cast_fp16, y = var_2135_cast_fp16)[name = tensor("koi_21_cast_fp16")]; + tensor qo_axis_0 = const()[name = tensor("qo_axis_0"), val = tensor(-1)]; + tensor qo_cast_fp16 = stack(axis = qo_axis_0, values = (qor_21_cast_fp16, qoi_21_cast_fp16))[name = tensor("qo_cast_fp16")]; + tensor ko_axis_0 = const()[name = tensor("ko_axis_0"), val = tensor(-1)]; + tensor ko_cast_fp16 = stack(axis = ko_axis_0, values = (kor_21_cast_fp16, koi_21_cast_fp16))[name = tensor("ko_cast_fp16")]; + tensor var_2168 = const()[name = tensor("op_2168"), val = tensor([1, 1, 16, 64])]; + tensor q_33_cast_fp16 = reshape(shape = var_2168, x = qo_cast_fp16)[name = tensor("q_33_cast_fp16")]; + tensor var_2174 = const()[name = tensor("op_2174"), val = tensor([1, 1, 16, 64])]; + tensor k_cast_fp16 = reshape(shape = var_2174, x = ko_cast_fp16)[name = tensor("k_cast_fp16")]; + tensor _inversed_2199_y_0_to_fp16 = const()[name = tensor("_inversed_2199_y_0_to_fp16"), val = tensor(0x1p-9)]; + tensor _inversed_2199_cast_fp16 = mul(x = ts_cast_fp16, y = _inversed_2199_y_0_to_fp16)[name = tensor("_inversed_2199_cast_fp16")]; + tensor var_2200_cast_fp16 = floor(x = _inversed_2199_cast_fp16)[name = tensor("op_2200_cast_fp16")]; + tensor var_2201_to_fp16 = const()[name = tensor("op_2201_to_fp16"), val = tensor(0x1p+9)]; + tensor var_2202_cast_fp16 = mul(x = var_2200_cast_fp16, y = var_2201_to_fp16)[name = tensor("op_2202_cast_fp16")]; + tensor write_indices_float_cast_fp16 = sub(x = ts_cast_fp16, y = var_2202_cast_fp16)[name = tensor("write_indices_float_cast_fp16")]; + tensor var_2209_dtype_0 = const()[name = tensor("op_2209_dtype_0"), val = tensor("int32")]; + tensor write_indices_reps_0 = const()[name = tensor("write_indices_reps_0"), val = tensor([1, 1, 16, 64])]; + tensor write_indices_float_cast_fp16_to_int32 = cast(dtype = var_2209_dtype_0, x = write_indices_float_cast_fp16)[name = tensor("cast_113")]; + tensor write_indices = tile(reps = write_indices_reps_0, x = write_indices_float_cast_fp16_to_int32)[name = tensor("write_indices")]; + tensor var_2219_begin_0 = const()[name = tensor("op_2219_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_2219_end_0 = const()[name = tensor("op_2219_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor var_2219_end_mask_0 = const()[name = tensor("op_2219_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_2219_squeeze_mask_0 = const()[name = tensor("op_2219_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor cache5_to_fp16_dtype_0 = const()[name = tensor("cache5_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor cache5_to_fp16 = cast(dtype = cache5_to_fp16_dtype_0, x = cache5)[name = tensor("cast_112")]; + tensor var_2219_cast_fp16 = slice_by_index(begin = var_2219_begin_0, end = var_2219_end_0, end_mask = var_2219_end_mask_0, squeeze_mask = var_2219_squeeze_mask_0, x = cache5_to_fp16)[name = tensor("op_2219_cast_fp16")]; + tensor var_2221_axis_0 = const()[name = tensor("op_2221_axis_0"), val = tensor(1)]; + tensor var_2221_mode_0 = const()[name = tensor("op_2221_mode_0"), val = tensor("update")]; + tensor var_2221_validate_indices_0 = const()[name = tensor("op_2221_validate_indices_0"), val = tensor(false)]; + tensor var_2221_cast_fp16 = scatter_along_axis(axis = var_2221_axis_0, data = var_2219_cast_fp16, indices = write_indices, mode = var_2221_mode_0, updates = k_cast_fp16, validate_indices = var_2221_validate_indices_0)[name = tensor("op_2221_cast_fp16")]; + tensor concat_37 = const()[name = tensor("concat_37"), val = tensor([0, 0, 0, 0, 0])]; + tensor concat_38 = const()[name = tensor("concat_38"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_internal_tensor_assign_1_stride_0 = const()[name = tensor("new_cache_internal_tensor_assign_1_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_internal_tensor_assign_1_begin_mask_0 = const()[name = tensor("new_cache_internal_tensor_assign_1_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_internal_tensor_assign_1_end_mask_0 = const()[name = tensor("new_cache_internal_tensor_assign_1_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_internal_tensor_assign_1_squeeze_mask_0 = const()[name = tensor("new_cache_internal_tensor_assign_1_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_22 = const()[name = tensor("shape_22"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_10 = const()[name = tensor("reduce_prod_10"), val = tensor(1048576)]; + tensor range_1d_10_start_0 = const()[name = tensor("range_1d_10_start_0"), val = tensor(0)]; + tensor range_1d_10_step_0 = const()[name = tensor("range_1d_10_step_0"), val = tensor(1)]; + tensor range_1d_10 = range_1d(end = reduce_prod_10, start = range_1d_10_start_0, step = range_1d_10_step_0)[name = tensor("range_1d_10")]; + tensor reshape_50 = reshape(shape = shape_22, x = range_1d_10)[name = tensor("reshape_50")]; + tensor slice_by_index_10 = slice_by_index(begin = concat_37, begin_mask = new_cache_internal_tensor_assign_1_begin_mask_0, end = concat_38, end_mask = new_cache_internal_tensor_assign_1_end_mask_0, squeeze_mask = new_cache_internal_tensor_assign_1_squeeze_mask_0, stride = new_cache_internal_tensor_assign_1_stride_0, x = reshape_50)[name = tensor("slice_by_index_10")]; + tensor reshape_51_shape_0 = const()[name = tensor("reshape_51_shape_0"), val = tensor([-1])]; + tensor reshape_51 = reshape(shape = reshape_51_shape_0, x = slice_by_index_10)[name = tensor("reshape_51")]; + tensor reshape_52_shape_0 = const()[name = tensor("reshape_52_shape_0"), val = tensor([-1])]; + tensor reshape_52_cast_fp16 = reshape(shape = reshape_52_shape_0, x = var_2221_cast_fp16)[name = tensor("reshape_52_cast_fp16")]; + tensor reshape_53_shape_0 = const()[name = tensor("reshape_53_shape_0"), val = tensor([-1])]; + tensor reshape_53_cast_fp16 = reshape(shape = reshape_53_shape_0, x = cache5_to_fp16)[name = tensor("reshape_53_cast_fp16")]; + tensor scatter_10_mode_0 = const()[name = tensor("scatter_10_mode_0"), val = tensor("update")]; + tensor scatter_10_axis_0 = const()[name = tensor("scatter_10_axis_0"), val = tensor(0)]; + tensor scatter_10_validate_indices_0 = const()[name = tensor("scatter_10_validate_indices_0"), val = tensor(false)]; + tensor scatter_10_cast_fp16 = scatter(axis = scatter_10_axis_0, data = reshape_53_cast_fp16, indices = reshape_51, mode = scatter_10_mode_0, updates = reshape_52_cast_fp16, validate_indices = scatter_10_validate_indices_0)[name = tensor("scatter_10_cast_fp16")]; + tensor reshape_54_cast_fp16 = reshape(shape = shape_22, x = scatter_10_cast_fp16)[name = tensor("reshape_54_cast_fp16")]; + tensor var_2229_begin_0 = const()[name = tensor("op_2229_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor var_2229_end_0 = const()[name = tensor("op_2229_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor var_2229_end_mask_0 = const()[name = tensor("op_2229_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_2229_squeeze_mask_0 = const()[name = tensor("op_2229_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor var_2229_cast_fp16 = slice_by_index(begin = var_2229_begin_0, end = var_2229_end_0, end_mask = var_2229_end_mask_0, squeeze_mask = var_2229_squeeze_mask_0, x = reshape_54_cast_fp16)[name = tensor("op_2229_cast_fp16")]; + tensor var_2231_axis_0 = const()[name = tensor("op_2231_axis_0"), val = tensor(1)]; + tensor var_2231_mode_0 = const()[name = tensor("op_2231_mode_0"), val = tensor("update")]; + tensor var_2231_validate_indices_0 = const()[name = tensor("op_2231_validate_indices_0"), val = tensor(false)]; + tensor var_2231_cast_fp16 = scatter_along_axis(axis = var_2231_axis_0, data = var_2229_cast_fp16, indices = write_indices, mode = var_2231_mode_0, updates = v_cast_fp16, validate_indices = var_2231_validate_indices_0)[name = tensor("op_2231_cast_fp16")]; + tensor concat_39 = const()[name = tensor("concat_39"), val = tensor([1, 0, 0, 0, 0])]; + tensor concat_40 = const()[name = tensor("concat_40"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_internal_tensor_assign_2_stride_0 = const()[name = tensor("new_cache_internal_tensor_assign_2_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_internal_tensor_assign_2_begin_mask_0 = const()[name = tensor("new_cache_internal_tensor_assign_2_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_internal_tensor_assign_2_end_mask_0 = const()[name = tensor("new_cache_internal_tensor_assign_2_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_internal_tensor_assign_2_squeeze_mask_0 = const()[name = tensor("new_cache_internal_tensor_assign_2_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_23 = const()[name = tensor("shape_23"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_11 = const()[name = tensor("reduce_prod_11"), val = tensor(1048576)]; + tensor range_1d_11_start_0 = const()[name = tensor("range_1d_11_start_0"), val = tensor(0)]; + tensor range_1d_11_step_0 = const()[name = tensor("range_1d_11_step_0"), val = tensor(1)]; + tensor range_1d_11 = range_1d(end = reduce_prod_11, start = range_1d_11_start_0, step = range_1d_11_step_0)[name = tensor("range_1d_11")]; + tensor reshape_55 = reshape(shape = shape_23, x = range_1d_11)[name = tensor("reshape_55")]; + tensor slice_by_index_11 = slice_by_index(begin = concat_39, begin_mask = new_cache_internal_tensor_assign_2_begin_mask_0, end = concat_40, end_mask = new_cache_internal_tensor_assign_2_end_mask_0, squeeze_mask = new_cache_internal_tensor_assign_2_squeeze_mask_0, stride = new_cache_internal_tensor_assign_2_stride_0, x = reshape_55)[name = tensor("slice_by_index_11")]; + tensor reshape_56_shape_0 = const()[name = tensor("reshape_56_shape_0"), val = tensor([-1])]; + tensor reshape_56 = reshape(shape = reshape_56_shape_0, x = slice_by_index_11)[name = tensor("reshape_56")]; + tensor reshape_57_shape_0 = const()[name = tensor("reshape_57_shape_0"), val = tensor([-1])]; + tensor reshape_57_cast_fp16 = reshape(shape = reshape_57_shape_0, x = var_2231_cast_fp16)[name = tensor("reshape_57_cast_fp16")]; + tensor reshape_58_shape_0 = const()[name = tensor("reshape_58_shape_0"), val = tensor([-1])]; + tensor reshape_58_cast_fp16 = reshape(shape = reshape_58_shape_0, x = reshape_54_cast_fp16)[name = tensor("reshape_58_cast_fp16")]; + tensor scatter_11_mode_0 = const()[name = tensor("scatter_11_mode_0"), val = tensor("update")]; + tensor scatter_11_axis_0 = const()[name = tensor("scatter_11_axis_0"), val = tensor(0)]; + tensor scatter_11_validate_indices_0 = const()[name = tensor("scatter_11_validate_indices_0"), val = tensor(false)]; + tensor scatter_11_cast_fp16 = scatter(axis = scatter_11_axis_0, data = reshape_58_cast_fp16, indices = reshape_56, mode = scatter_11_mode_0, updates = reshape_57_cast_fp16, validate_indices = scatter_11_validate_indices_0)[name = tensor("scatter_11_cast_fp16")]; + tensor reshape_59_cast_fp16 = reshape(shape = shape_23, x = scatter_11_cast_fp16)[name = tensor("reshape_59_cast_fp16")]; + tensor reshape_59_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("reshape_59_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor keys_31_begin_0 = const()[name = tensor("keys_31_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor keys_31_end_0 = const()[name = tensor("keys_31_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor keys_31_end_mask_0 = const()[name = tensor("keys_31_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor keys_31_squeeze_mask_0 = const()[name = tensor("keys_31_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor keys_31_cast_fp16 = slice_by_index(begin = keys_31_begin_0, end = keys_31_end_0, end_mask = keys_31_end_mask_0, squeeze_mask = keys_31_squeeze_mask_0, x = reshape_59_cast_fp16)[name = tensor("keys_31_cast_fp16")]; + tensor values_31_begin_0 = const()[name = tensor("values_31_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor values_31_end_0 = const()[name = tensor("values_31_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor values_31_end_mask_0 = const()[name = tensor("values_31_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor values_31_squeeze_mask_0 = const()[name = tensor("values_31_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor values_31_cast_fp16 = slice_by_index(begin = values_31_begin_0, end = values_31_end_0, end_mask = values_31_end_mask_0, squeeze_mask = values_31_squeeze_mask_0, x = reshape_59_cast_fp16)[name = tensor("values_31_cast_fp16")]; + tensor var_2243_cast_fp16 = not_equal(x = keys_31_cast_fp16, y = keys_31_cast_fp16)[name = tensor("op_2243_cast_fp16")]; + tensor keys_33_cast_fp16 = select(a = var_319_to_fp16, b = keys_31_cast_fp16, cond = var_2243_cast_fp16)[name = tensor("keys_33_cast_fp16")]; + tensor var_2251_cast_fp16 = not_equal(x = values_31_cast_fp16, y = values_31_cast_fp16)[name = tensor("op_2251_cast_fp16")]; + tensor values_33_cast_fp16 = select(a = var_319_to_fp16, b = values_31_cast_fp16, cond = var_2251_cast_fp16)[name = tensor("values_33_cast_fp16")]; + tensor var_2275 = const()[name = tensor("op_2275"), val = tensor([0, 2, 1, 3])]; + tensor var_2291 = const()[name = tensor("op_2291"), val = tensor([1, 1, 1])]; + tensor var_2292_cast_fp16 = reshape(shape = var_2291, x = position5_to_fp16)[name = tensor("op_2292_cast_fp16")]; + tensor var_2312_to_fp16 = const()[name = tensor("op_2312_to_fp16"), val = tensor(0x1p+0)]; + tensor valid_len_cast_fp16 = add(x = var_2292_cast_fp16, y = var_2312_to_fp16)[name = tensor("valid_len_cast_fp16")]; + tensor valid_mask_cast_fp16 = less(x = k_positions_1_promoted_to_fp16, y = valid_len_cast_fp16)[name = tensor("valid_mask_cast_fp16")]; + tensor causal_mask_cast_fp16 = less_equal(x = k_positions_1_promoted_to_fp16, y = var_2292_cast_fp16)[name = tensor("causal_mask_cast_fp16")]; + tensor attn_mask_21 = logical_and(x = valid_mask_cast_fp16, y = causal_mask_cast_fp16)[name = tensor("attn_mask_21")]; + tensor attn_mask_axes_0 = const()[name = tensor("attn_mask_axes_0"), val = tensor([1])]; + tensor attn_mask = expand_dims(axes = attn_mask_axes_0, x = attn_mask_21)[name = tensor("attn_mask")]; + tensor var_2330_transpose_x_0 = const()[name = tensor("op_2330_transpose_x_0"), val = tensor(false)]; + tensor var_2330_transpose_y_0 = const()[name = tensor("op_2330_transpose_y_0"), val = tensor(false)]; + tensor transpose_28_perm_0 = const()[name = tensor("transpose_28_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_29_perm_0 = const()[name = tensor("transpose_29_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_29 = transpose(perm = transpose_29_perm_0, x = keys_33_cast_fp16)[name = tensor("transpose_31")]; + tensor transpose_28 = transpose(perm = transpose_28_perm_0, x = q_33_cast_fp16)[name = tensor("transpose_32")]; + tensor var_2330_cast_fp16 = matmul(transpose_x = var_2330_transpose_x_0, transpose_y = var_2330_transpose_y_0, x = transpose_28, y = transpose_29)[name = tensor("op_2330_cast_fp16")]; + tensor var_2324_to_fp16 = const()[name = tensor("op_2324_to_fp16"), val = tensor([0x1p-3])]; + tensor attn_weights_31_cast_fp16 = mul(x = var_2330_cast_fp16, y = var_2324_to_fp16)[name = tensor("attn_weights_31_cast_fp16")]; + tensor var_2332 = logical_not(x = attn_mask)[name = tensor("op_2332")]; + tensor var_2333_to_fp16 = const()[name = tensor("op_2333_to_fp16"), val = tensor(-inf)]; + tensor attn_weights_33_cast_fp16 = select(a = var_2333_to_fp16, b = attn_weights_31_cast_fp16, cond = var_2332)[name = tensor("attn_weights_33_cast_fp16")]; + tensor var_2335 = const()[name = tensor("op_2335"), val = tensor(-1)]; + tensor attn_weights_cast_fp16 = softmax(axis = var_2335, x = attn_weights_33_cast_fp16)[name = tensor("attn_weights_cast_fp16")]; + tensor attn_output_transpose_x_0 = const()[name = tensor("attn_output_transpose_x_0"), val = tensor(false)]; + tensor attn_output_transpose_y_0 = const()[name = tensor("attn_output_transpose_y_0"), val = tensor(false)]; + tensor values_cast_fp16 = transpose(perm = var_2275, x = values_33_cast_fp16)[name = tensor("transpose_33")]; + tensor attn_output_cast_fp16 = matmul(transpose_x = attn_output_transpose_x_0, transpose_y = attn_output_transpose_y_0, x = attn_weights_cast_fp16, y = values_cast_fp16)[name = tensor("attn_output_cast_fp16")]; + tensor var_2343 = const()[name = tensor("op_2343"), val = tensor([0, 2, 1, 3])]; + tensor var_2348 = const()[name = tensor("op_2348"), val = tensor([1, 1, 1024])]; + tensor var_2344_cast_fp16 = transpose(perm = var_2343, x = attn_output_cast_fp16)[name = tensor("transpose_30")]; + tensor input_67_cast_fp16 = reshape(shape = var_2348, x = var_2344_cast_fp16)[name = tensor("input_67_cast_fp16")]; + tensor attn5_out_proj_weight_to_fp16 = const()[name = tensor("attn5_out_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(133301120)))]; + tensor linear_22_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = attn5_out_proj_weight_to_fp16, x = input_67_cast_fp16)[name = tensor("linear_22_cast_fp16")]; + tensor var_2354_to_fp16 = const()[name = tensor("op_2354_to_fp16"), val = tensor(0x1p+0)]; + tensor var_2355_cast_fp16 = add(x = position5_to_fp16, y = var_2354_to_fp16)[name = tensor("op_2355_cast_fp16")]; + tensor var_2355_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_2355_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor input_69_cast_fp16 = add(x = input_63_cast_fp16, y = linear_22_cast_fp16)[name = tensor("input_69_cast_fp16")]; + tensor input_71_axes_0 = const()[name = tensor("input_71_axes_0"), val = tensor([-1])]; + tensor norm5_2_weight_to_fp16 = const()[name = tensor("norm5_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(135398336)))]; + tensor norm5_2_bias_to_fp16 = const()[name = tensor("norm5_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(135400448)))]; + tensor var_2359_to_fp16 = const()[name = tensor("op_2359_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_71_cast_fp16 = layer_norm(axes = input_71_axes_0, beta = norm5_2_bias_to_fp16, epsilon = var_2359_to_fp16, gamma = norm5_2_weight_to_fp16, x = input_69_cast_fp16)[name = tensor("input_71_cast_fp16")]; + tensor linear5_1_weight_to_fp16 = const()[name = tensor("linear5_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(135402560)))]; + tensor linear_23_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = linear5_1_weight_to_fp16, x = input_71_cast_fp16)[name = tensor("linear_23_cast_fp16")]; + tensor input_73_mode_0 = const()[name = tensor("input_73_mode_0"), val = tensor("EXACT")]; + tensor input_73_cast_fp16 = gelu(mode = input_73_mode_0, x = linear_23_cast_fp16)[name = tensor("input_73_cast_fp16")]; + tensor linear5_2_weight_to_fp16 = const()[name = tensor("linear5_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(143791232)))]; + tensor linear_24_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = linear5_2_weight_to_fp16, x = input_73_cast_fp16)[name = tensor("linear_24_cast_fp16")]; + tensor input_75_cast_fp16 = add(x = input_69_cast_fp16, y = linear_24_cast_fp16)[name = tensor("input_75_cast_fp16")]; + tensor input_axes_0 = const()[name = tensor("input_axes_0"), val = tensor([-1])]; + tensor out_norm_weight_to_fp16 = const()[name = tensor("out_norm_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(152179904)))]; + tensor out_norm_bias_to_fp16 = const()[name = tensor("out_norm_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(152182016)))]; + tensor var_2376_to_fp16 = const()[name = tensor("op_2376_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_cast_fp16 = layer_norm(axes = input_axes_0, beta = out_norm_bias_to_fp16, epsilon = var_2376_to_fp16, gamma = out_norm_weight_to_fp16, x = input_75_cast_fp16)[name = tensor("input_cast_fp16")]; + tensor input_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("input_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor out_eos_weight_to_fp16 = const()[name = tensor("out_eos_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(152184128)))]; + tensor out_eos_bias_to_fp16 = const()[name = tensor("out_eos_bias_to_fp16"), val = tensor([-0x1.36p-2])]; + tensor linear_25_cast_fp16 = linear(bias = out_eos_bias_to_fp16, weight = out_eos_weight_to_fp16, x = input_cast_fp16)[name = tensor("linear_25_cast_fp16")]; + tensor linear_25_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("linear_25_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor var_2384 = cast(dtype = linear_25_cast_fp16_to_fp32_dtype_0, x = linear_25_cast_fp16)[name = tensor("cast_108")]; + tensor input = cast(dtype = input_cast_fp16_to_fp32_dtype_0, x = input_cast_fp16)[name = tensor("cast_109")]; + tensor var_2355 = cast(dtype = var_2355_cast_fp16_to_fp32_dtype_0, x = var_2355_cast_fp16)[name = tensor("cast_110")]; + tensor new_cache_internal_tensor_assign_2 = cast(dtype = reshape_59_cast_fp16_to_fp32_dtype_0, x = reshape_59_cast_fp16)[name = tensor("cast_111")]; + tensor var_1969 = cast(dtype = var_1969_cast_fp16_to_fp32_dtype_0, x = var_1969_cast_fp16)[name = tensor("cast_115")]; + tensor new_cache_9_internal_tensor_assign_2 = cast(dtype = reshape_49_cast_fp16_to_fp32_dtype_0, x = reshape_49_cast_fp16)[name = tensor("cast_116")]; + tensor var_1583 = cast(dtype = var_1583_cast_fp16_to_fp32_dtype_0, x = var_1583_cast_fp16)[name = tensor("cast_120")]; + tensor new_cache_7_internal_tensor_assign_2 = cast(dtype = reshape_39_cast_fp16_to_fp32_dtype_0, x = reshape_39_cast_fp16)[name = tensor("cast_121")]; + tensor var_1197 = cast(dtype = var_1197_cast_fp16_to_fp32_dtype_0, x = var_1197_cast_fp16)[name = tensor("cast_125")]; + tensor new_cache_5_internal_tensor_assign_2 = cast(dtype = reshape_29_cast_fp16_to_fp32_dtype_0, x = reshape_29_cast_fp16)[name = tensor("cast_126")]; + tensor var_811 = cast(dtype = var_811_cast_fp16_to_fp32_dtype_0, x = var_811_cast_fp16)[name = tensor("cast_130")]; + tensor new_cache_3_internal_tensor_assign_2 = cast(dtype = reshape_19_cast_fp16_to_fp32_dtype_0, x = reshape_19_cast_fp16)[name = tensor("cast_131")]; + tensor var_425 = cast(dtype = var_425_cast_fp16_to_fp32_dtype_0, x = var_425_cast_fp16)[name = tensor("cast_135")]; + tensor new_cache_1_internal_tensor_assign_2 = cast(dtype = reshape_9_cast_fp16_to_fp32_dtype_0, x = reshape_9_cast_fp16)[name = tensor("cast_136")]; + } -> (input, var_2384, new_cache_1_internal_tensor_assign_2, var_425, new_cache_3_internal_tensor_assign_2, var_811, new_cache_5_internal_tensor_assign_2, var_1197, new_cache_7_internal_tensor_assign_2, var_1583, new_cache_9_internal_tensor_assign_2, var_1969, new_cache_internal_tensor_assign_2, var_2355); +} \ No newline at end of file diff --git a/v2.1/english/flowlm_step.mlmodelc/weights/weight.bin b/v2.1/english/flowlm_step.mlmodelc/weights/weight.bin new file mode 100644 index 0000000000000000000000000000000000000000..9d5ba58aff69bd315b18fc89d6a6df80bb9e481a --- /dev/null +++ b/v2.1/english/flowlm_step.mlmodelc/weights/weight.bin @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:5f63c2466dd9c1a29b074c467cc32b1fe47d77afa745e86f03ae7e862ac05aa7 +size 152186240 diff --git a/v2.1/english/flowlm_step.mlpackage/Data/com.apple.CoreML/model.mlmodel b/v2.1/english/flowlm_step.mlpackage/Data/com.apple.CoreML/model.mlmodel new file mode 100644 index 0000000000000000000000000000000000000000..23239f3a4eca12ef6a4605f242b88d9e3a5ab41d --- /dev/null +++ b/v2.1/english/flowlm_step.mlpackage/Data/com.apple.CoreML/model.mlmodel @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:28ab051ccca211c2a9f108a734a3efe0d1fa8b81b85ce7078fa11786837ee698 +size 207618 diff --git a/v2.1/english/flowlm_step.mlpackage/Data/com.apple.CoreML/weights/weight.bin b/v2.1/english/flowlm_step.mlpackage/Data/com.apple.CoreML/weights/weight.bin new file mode 100644 index 0000000000000000000000000000000000000000..9d5ba58aff69bd315b18fc89d6a6df80bb9e481a --- /dev/null +++ b/v2.1/english/flowlm_step.mlpackage/Data/com.apple.CoreML/weights/weight.bin @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:5f63c2466dd9c1a29b074c467cc32b1fe47d77afa745e86f03ae7e862ac05aa7 +size 152186240 diff --git a/v2.1/english/flowlm_step.mlpackage/Manifest.json b/v2.1/english/flowlm_step.mlpackage/Manifest.json new file mode 100644 index 0000000000000000000000000000000000000000..b75398afd1a5f97090dc7e59939a20d477c26265 --- /dev/null +++ b/v2.1/english/flowlm_step.mlpackage/Manifest.json @@ -0,0 +1,18 @@ +{ + "fileFormatVersion": "1.0.0", + "itemInfoEntries": { + "20655CC4-1FBD-4DDF-AFCB-3DEC555D8760": { + "author": "com.apple.CoreML", + "description": "CoreML Model Specification", + "name": "model.mlmodel", + "path": "com.apple.CoreML/model.mlmodel" + }, + "29F2DC2D-122F-4773-88C4-B0AA8422BCF0": { + "author": "com.apple.CoreML", + "description": "CoreML Model Weights", + "name": "weights", + "path": "com.apple.CoreML/weights" + } + }, + "rootModelIdentifier": "20655CC4-1FBD-4DDF-AFCB-3DEC555D8760" +} diff --git a/v2.1/english/flowlm_stepv2.mlmodelc/analytics/coremldata.bin b/v2.1/english/flowlm_stepv2.mlmodelc/analytics/coremldata.bin new file mode 100644 index 0000000000000000000000000000000000000000..d8cffcae9256919262daa96ae6be98d6a8863c23 --- /dev/null +++ b/v2.1/english/flowlm_stepv2.mlmodelc/analytics/coremldata.bin @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:f6c97b47c4dbb8ea1788be74e16913d4aa38ac64ee3d066f0edc1b32f9afb480 +size 243 diff --git a/v2.1/english/flowlm_stepv2.mlmodelc/coremldata.bin b/v2.1/english/flowlm_stepv2.mlmodelc/coremldata.bin new file mode 100644 index 0000000000000000000000000000000000000000..d41cf92ef91a29881b1b529b7ae40ba6b6f0e095 --- /dev/null +++ b/v2.1/english/flowlm_stepv2.mlmodelc/coremldata.bin @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:2ebbbc1c5a3adb00f2651b0babcf6c20554d56f921be5da5c0216154d949a3b2 +size 1190 diff --git a/v2.1/english/flowlm_stepv2.mlmodelc/metadata.json b/v2.1/english/flowlm_stepv2.mlmodelc/metadata.json new file mode 100644 index 0000000000000000000000000000000000000000..f2969c83e1cf7a6025c56eb3502b7ec456a89630 --- /dev/null +++ b/v2.1/english/flowlm_stepv2.mlmodelc/metadata.json @@ -0,0 +1,348 @@ +[ + { + "metadataOutputVersion" : "3.0", + "storagePrecision" : "Mixed (Float16, Int8)", + "outputSchema" : [ + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1 × 1 × 1024)", + "shortDescription" : "", + "shape" : "[1, 1, 1024]", + "name" : "input", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1 × 1 × 1)", + "shortDescription" : "", + "shape" : "[1, 1, 1]", + "name" : "var_2582", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 2 × 1 × 512 × 16 × 64)", + "shortDescription" : "", + "shape" : "[2, 1, 512, 16, 64]", + "name" : "new_cache_1_internal_tensor_assign_2", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1)", + "shortDescription" : "", + "shape" : "[1]", + "name" : "var_458", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 2 × 1 × 512 × 16 × 64)", + "shortDescription" : "", + "shape" : "[2, 1, 512, 16, 64]", + "name" : "new_cache_3_internal_tensor_assign_2", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1)", + "shortDescription" : "", + "shape" : "[1]", + "name" : "var_877", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 2 × 1 × 512 × 16 × 64)", + "shortDescription" : "", + "shape" : "[2, 1, 512, 16, 64]", + "name" : "new_cache_5_internal_tensor_assign_2", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1)", + "shortDescription" : "", + "shape" : "[1]", + "name" : "var_1296", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 2 × 1 × 512 × 16 × 64)", + "shortDescription" : "", + "shape" : "[2, 1, 512, 16, 64]", + "name" : "new_cache_7_internal_tensor_assign_2", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1)", + "shortDescription" : "", + "shape" : "[1]", + "name" : "var_1715", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 2 × 1 × 512 × 16 × 64)", + "shortDescription" : "", + "shape" : "[2, 1, 512, 16, 64]", + "name" : "new_cache_9_internal_tensor_assign_2", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1)", + "shortDescription" : "", + "shape" : "[1]", + "name" : "var_2134", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 2 × 1 × 512 × 16 × 64)", + "shortDescription" : "", + "shape" : "[2, 1, 512, 16, 64]", + "name" : "new_cache_internal_tensor_assign_2", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1)", + "shortDescription" : "", + "shape" : "[1]", + "name" : "var_2553", + "type" : "MultiArray" + } + ], + "modelParameters" : [ + + ], + "specificationVersion" : 8, + "mlProgramOperationTypeHistogram" : { + "Ios17.cos" : 6, + "Range1d" : 12, + "Ios17.notEqual" : 13, + "Ios17.floor" : 6, + "Ios17.reshape" : 108, + "Ios17.logicalAnd" : 6, + "Ios16.softmax" : 6, + "Ios17.scatter" : 12, + "Ios17.matmul" : 12, + "Ios17.sin" : 6, + "Ios17.transpose" : 24, + "Select" : 19, + "Ios17.expandDims" : 7, + "Ios17.add" : 36, + "Tile" : 6, + "Ios17.sliceByIndex" : 78, + "Ios17.lessEqual" : 6, + "Ios17.scatterAlongAxis" : 12, + "Ios17.logicalNot" : 6, + "Ios17.layerNorm" : 13, + "Ios16.constexprAffineDequantize" : 24, + "Ios17.cast" : 34, + "Ios17.sub" : 18, + "Ios17.less" : 6, + "Ios16.gelu" : 6, + "Ios17.linear" : 26, + "Stack" : 12, + "Ios17.mul" : 72 + }, + "computePrecision" : "Mixed (Float16, Float32, Int32)", + "isUpdatable" : "0", + "stateSchema" : [ + + ], + "availability" : { + "macOS" : "14.0", + "tvOS" : "17.0", + "visionOS" : "1.0", + "watchOS" : "10.0", + "iOS" : "17.0", + "macCatalyst" : "17.0" + }, + "modelType" : { + "name" : "MLModelType_mlProgram" + }, + "userDefinedMetadata" : { + "com.github.apple.coremltools.conversion_date" : "2026-04-27", + "com.github.apple.coremltools.source" : "torch==2.9.1", + "com.github.apple.coremltools.version" : "9.0", + "com.github.apple.coremltools.source_dialect" : "TorchScript" + }, + "inputSchema" : [ + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1 × 1 × 32)", + "shortDescription" : "", + "shape" : "[1, 1, 32]", + "name" : "sequence", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 32)", + "shortDescription" : "", + "shape" : "[32]", + "name" : "bos_emb", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 2 × 1 × 512 × 16 × 64)", + "shortDescription" : "", + "shape" : "[2, 1, 512, 16, 64]", + "name" : "cache0", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1)", + "shortDescription" : "", + "shape" : "[1]", + "name" : "position0", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 2 × 1 × 512 × 16 × 64)", + "shortDescription" : "", + "shape" : "[2, 1, 512, 16, 64]", + "name" : "cache1", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1)", + "shortDescription" : "", + "shape" : "[1]", + "name" : "position1", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 2 × 1 × 512 × 16 × 64)", + "shortDescription" : "", + "shape" : "[2, 1, 512, 16, 64]", + "name" : "cache2", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1)", + "shortDescription" : "", + "shape" : "[1]", + "name" : "position2", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 2 × 1 × 512 × 16 × 64)", + "shortDescription" : "", + "shape" : "[2, 1, 512, 16, 64]", + "name" : "cache3", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1)", + "shortDescription" : "", + "shape" : "[1]", + "name" : "position3", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 2 × 1 × 512 × 16 × 64)", + "shortDescription" : "", + "shape" : "[2, 1, 512, 16, 64]", + "name" : "cache4", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1)", + "shortDescription" : "", + "shape" : "[1]", + "name" : "position4", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 2 × 1 × 512 × 16 × 64)", + "shortDescription" : "", + "shape" : "[2, 1, 512, 16, 64]", + "name" : "cache5", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1)", + "shortDescription" : "", + "shape" : "[1]", + "name" : "position5", + "type" : "MultiArray" + } + ], + "generatedClassName" : "flowlm_stepv2", + "method" : "predict" + } +] \ No newline at end of file diff --git a/v2.1/english/flowlm_stepv2.mlmodelc/model.mil b/v2.1/english/flowlm_stepv2.mlmodelc/model.mil new file mode 100644 index 0000000000000000000000000000000000000000..c70a840a182ac8a48e2eb6dd454773a2197061fa --- /dev/null +++ b/v2.1/english/flowlm_stepv2.mlmodelc/model.mil @@ -0,0 +1,1359 @@ +program(1.0) +[buildInfo = dict, tensor>({{"coremlc-component-MIL", "3520.4.1"}, {"coremlc-version", "3520.5.1"}, {"coremltools-component-torch", "2.9.1"}, {"coremltools-source-dialect", "TorchScript"}, {"coremltools-version", "9.0"}})] +{ + func main(tensor bos_emb, tensor cache0, tensor cache1, tensor cache2, tensor cache3, tensor cache4, tensor cache5, tensor position0, tensor position1, tensor position2, tensor position3, tensor position4, tensor position5, tensor sequence) { + tensor attn0_in_proj_weight_cast_fp16 = constexpr_affine_dequantize()[axis = tensor(0), name = tensor("attn0_in_proj_weight_cast_fp16"), quantized_data = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(64))), scale = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(3148992))), zero_point = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(3145856)))]; + tensor attn0_out_proj_weight_cast_fp16 = constexpr_affine_dequantize()[axis = tensor(0), name = tensor("attn0_out_proj_weight_cast_fp16"), quantized_data = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(3155200))), scale = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(4204928))), zero_point = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(4203840)))]; + tensor linear0_1_weight_cast_fp16 = constexpr_affine_dequantize()[axis = tensor(0), name = tensor("linear0_1_weight_cast_fp16"), quantized_data = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(4207040))), scale = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8405568))), zero_point = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8401408)))]; + tensor linear0_2_weight_cast_fp16 = constexpr_affine_dequantize()[axis = tensor(0), name = tensor("linear0_2_weight_cast_fp16"), quantized_data = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8413824))), scale = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(12608192))), zero_point = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(4203840)))]; + tensor attn1_in_proj_weight_cast_fp16 = constexpr_affine_dequantize()[axis = tensor(0), name = tensor("attn1_in_proj_weight_cast_fp16"), quantized_data = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(12610304))), scale = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15756096))), zero_point = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(3145856)))]; + tensor attn1_out_proj_weight_cast_fp16 = constexpr_affine_dequantize()[axis = tensor(0), name = tensor("attn1_out_proj_weight_cast_fp16"), quantized_data = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15762304))), scale = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(16810944))), zero_point = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(4203840)))]; + tensor linear1_1_weight_cast_fp16 = constexpr_affine_dequantize()[axis = tensor(0), name = tensor("linear1_1_weight_cast_fp16"), quantized_data = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(16813056))), scale = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21007424))), zero_point = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8401408)))]; + tensor linear1_2_weight_cast_fp16 = constexpr_affine_dequantize()[axis = tensor(0), name = tensor("linear1_2_weight_cast_fp16"), quantized_data = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21015680))), scale = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(25210048))), zero_point = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(4203840)))]; + tensor attn2_in_proj_weight_cast_fp16 = constexpr_affine_dequantize()[axis = tensor(0), name = tensor("attn2_in_proj_weight_cast_fp16"), quantized_data = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(25212160))), scale = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28357952))), zero_point = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(3145856)))]; + tensor attn2_out_proj_weight_cast_fp16 = constexpr_affine_dequantize()[axis = tensor(0), name = tensor("attn2_out_proj_weight_cast_fp16"), quantized_data = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28364160))), scale = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(29412800))), zero_point = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(4203840)))]; + tensor linear2_1_weight_cast_fp16 = constexpr_affine_dequantize()[axis = tensor(0), name = tensor("linear2_1_weight_cast_fp16"), quantized_data = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(29414912))), scale = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(33609280))), zero_point = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8401408)))]; + tensor linear2_2_weight_cast_fp16 = constexpr_affine_dequantize()[axis = tensor(0), name = tensor("linear2_2_weight_cast_fp16"), quantized_data = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(33617536))), scale = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(37811904))), zero_point = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(4203840)))]; + tensor attn3_in_proj_weight_cast_fp16 = constexpr_affine_dequantize()[axis = tensor(0), name = tensor("attn3_in_proj_weight_cast_fp16"), quantized_data = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(37814016))), scale = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(40959808))), zero_point = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(3145856)))]; + tensor attn3_out_proj_weight_cast_fp16 = constexpr_affine_dequantize()[axis = tensor(0), name = tensor("attn3_out_proj_weight_cast_fp16"), quantized_data = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(40966016))), scale = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(42014656))), zero_point = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(4203840)))]; + tensor linear3_1_weight_cast_fp16 = constexpr_affine_dequantize()[axis = tensor(0), name = tensor("linear3_1_weight_cast_fp16"), quantized_data = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(42016768))), scale = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(46211136))), zero_point = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8401408)))]; + tensor linear3_2_weight_cast_fp16 = constexpr_affine_dequantize()[axis = tensor(0), name = tensor("linear3_2_weight_cast_fp16"), quantized_data = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(46219392))), scale = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(50413760))), zero_point = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(4203840)))]; + tensor attn4_in_proj_weight_cast_fp16 = constexpr_affine_dequantize()[axis = tensor(0), name = tensor("attn4_in_proj_weight_cast_fp16"), quantized_data = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(50415872))), scale = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(53561664))), zero_point = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(3145856)))]; + tensor attn4_out_proj_weight_cast_fp16 = constexpr_affine_dequantize()[axis = tensor(0), name = tensor("attn4_out_proj_weight_cast_fp16"), quantized_data = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(53567872))), scale = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(54616512))), zero_point = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(4203840)))]; + tensor linear4_1_weight_cast_fp16 = constexpr_affine_dequantize()[axis = tensor(0), name = tensor("linear4_1_weight_cast_fp16"), quantized_data = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(54618624))), scale = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(58812992))), zero_point = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8401408)))]; + tensor linear4_2_weight_cast_fp16 = constexpr_affine_dequantize()[axis = tensor(0), name = tensor("linear4_2_weight_cast_fp16"), quantized_data = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(58821248))), scale = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(63015616))), zero_point = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(4203840)))]; + tensor attn5_in_proj_weight_cast_fp16 = constexpr_affine_dequantize()[axis = tensor(0), name = tensor("attn5_in_proj_weight_cast_fp16"), quantized_data = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(63017728))), scale = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(66163520))), zero_point = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(3145856)))]; + tensor attn5_out_proj_weight_cast_fp16 = constexpr_affine_dequantize()[axis = tensor(0), name = tensor("attn5_out_proj_weight_cast_fp16"), quantized_data = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(66169728))), scale = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(67218368))), zero_point = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(4203840)))]; + tensor linear5_1_weight_cast_fp16 = constexpr_affine_dequantize()[axis = tensor(0), name = tensor("linear5_1_weight_cast_fp16"), quantized_data = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(67220480))), scale = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(71414848))), zero_point = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8401408)))]; + tensor linear5_2_weight_cast_fp16 = constexpr_affine_dequantize()[axis = tensor(0), name = tensor("linear5_2_weight_cast_fp16"), quantized_data = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(71423104))), scale = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(75617472))), zero_point = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(4203840)))]; + tensor sequence_to_fp16_dtype_0 = const()[name = tensor("sequence_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor sequence_to_fp16 = cast(dtype = sequence_to_fp16_dtype_0, x = sequence)[name = tensor("cast_141")]; + tensor var_54_cast_fp16 = not_equal(x = sequence_to_fp16, y = sequence_to_fp16)[name = tensor("op_54_cast_fp16")]; + tensor expand_dims_0_axes_0 = const()[name = tensor("expand_dims_0_axes_0"), val = tensor([0, 1])]; + tensor bos_emb_to_fp16_dtype_0 = const()[name = tensor("bos_emb_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor bos_emb_to_fp16 = cast(dtype = bos_emb_to_fp16_dtype_0, x = bos_emb)[name = tensor("cast_140")]; + tensor expand_dims_0_cast_fp16 = expand_dims(axes = expand_dims_0_axes_0, x = bos_emb_to_fp16)[name = tensor("expand_dims_0_cast_fp16")]; + tensor input_1_cast_fp16 = select(a = expand_dims_0_cast_fp16, b = sequence_to_fp16, cond = var_54_cast_fp16)[name = tensor("input_1_cast_fp16")]; + tensor input_linear_weight_to_fp16 = const()[name = tensor("input_linear_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(75619584)))]; + tensor linear_0_bias_0_to_fp16 = const()[name = tensor("linear_0_bias_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(75685184)))]; + tensor linear_0_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = input_linear_weight_to_fp16, x = input_1_cast_fp16)[name = tensor("linear_0_cast_fp16")]; + tensor x_1_axes_0 = const()[name = tensor("x_1_axes_0"), val = tensor([-1])]; + tensor norm0_1_weight_to_fp16 = const()[name = tensor("norm0_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(75687296)))]; + tensor norm0_1_bias_to_fp16 = const()[name = tensor("norm0_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(75689408)))]; + tensor var_60_to_fp16 = const()[name = tensor("op_60_to_fp16"), val = tensor(0x1.5p-17)]; + tensor x_1_cast_fp16 = layer_norm(axes = x_1_axes_0, beta = norm0_1_bias_to_fp16, epsilon = var_60_to_fp16, gamma = norm0_1_weight_to_fp16, x = linear_0_cast_fp16)[name = tensor("x_1_cast_fp16")]; + tensor linear_1_bias_0_to_fp16 = const()[name = tensor("linear_1_bias_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(75691520)))]; + tensor linear_1_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = attn0_in_proj_weight_cast_fp16, x = x_1_cast_fp16)[name = tensor("linear_1_cast_fp16")]; + tensor var_96 = const()[name = tensor("op_96"), val = tensor([1, 1, 3, 16, 64])]; + tensor qkv_1_cast_fp16 = reshape(shape = var_96, x = linear_1_cast_fp16)[name = tensor("qkv_1_cast_fp16")]; + tensor q_1_begin_0 = const()[name = tensor("q_1_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor q_1_end_0 = const()[name = tensor("q_1_end_0"), val = tensor([1, 1, 1, 16, 64])]; + tensor q_1_end_mask_0 = const()[name = tensor("q_1_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor q_1_squeeze_mask_0 = const()[name = tensor("q_1_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor q_1_cast_fp16 = slice_by_index(begin = q_1_begin_0, end = q_1_end_0, end_mask = q_1_end_mask_0, squeeze_mask = q_1_squeeze_mask_0, x = qkv_1_cast_fp16)[name = tensor("q_1_cast_fp16")]; + tensor k_1_begin_0 = const()[name = tensor("k_1_begin_0"), val = tensor([0, 0, 1, 0, 0])]; + tensor k_1_end_0 = const()[name = tensor("k_1_end_0"), val = tensor([1, 1, 2, 16, 64])]; + tensor k_1_end_mask_0 = const()[name = tensor("k_1_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor k_1_squeeze_mask_0 = const()[name = tensor("k_1_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor k_1_cast_fp16 = slice_by_index(begin = k_1_begin_0, end = k_1_end_0, end_mask = k_1_end_mask_0, squeeze_mask = k_1_squeeze_mask_0, x = qkv_1_cast_fp16)[name = tensor("k_1_cast_fp16")]; + tensor v_1_begin_0 = const()[name = tensor("v_1_begin_0"), val = tensor([0, 0, 2, 0, 0])]; + tensor v_1_end_0 = const()[name = tensor("v_1_end_0"), val = tensor([1, 1, 3, 16, 64])]; + tensor v_1_end_mask_0 = const()[name = tensor("v_1_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor v_1_squeeze_mask_0 = const()[name = tensor("v_1_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor v_1_cast_fp16 = slice_by_index(begin = v_1_begin_0, end = v_1_end_0, end_mask = v_1_end_mask_0, squeeze_mask = v_1_squeeze_mask_0, x = qkv_1_cast_fp16)[name = tensor("v_1_cast_fp16")]; + tensor var_200 = const()[name = tensor("op_200"), val = tensor([1, 1, 1, 1])]; + tensor position0_to_fp16_dtype_0 = const()[name = tensor("position0_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor position0_to_fp16 = cast(dtype = position0_to_fp16_dtype_0, x = position0)[name = tensor("cast_139")]; + tensor ts_5_cast_fp16 = reshape(shape = var_200, x = position0_to_fp16)[name = tensor("ts_5_cast_fp16")]; + tensor var_204 = const()[name = tensor("op_204"), val = tensor([1, 1, 16, 32, 2])]; + tensor q_complex_1_cast_fp16 = reshape(shape = var_204, x = q_1_cast_fp16)[name = tensor("q_complex_1_cast_fp16")]; + tensor var_208 = const()[name = tensor("op_208"), val = tensor([1, 1, 16, 32, 2])]; + tensor k_complex_1_cast_fp16 = reshape(shape = var_208, x = k_1_cast_fp16)[name = tensor("k_complex_1_cast_fp16")]; + tensor var_212_begin_0 = const()[name = tensor("op_212_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_212_end_0 = const()[name = tensor("op_212_end_0"), val = tensor([1, 1, 16, 32, 1])]; + tensor var_212_end_mask_0 = const()[name = tensor("op_212_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_212_squeeze_mask_0 = const()[name = tensor("op_212_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_212_cast_fp16 = slice_by_index(begin = var_212_begin_0, end = var_212_end_0, end_mask = var_212_end_mask_0, squeeze_mask = var_212_squeeze_mask_0, x = q_complex_1_cast_fp16)[name = tensor("op_212_cast_fp16")]; + tensor var_220_begin_0 = const()[name = tensor("op_220_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_220_end_0 = const()[name = tensor("op_220_end_0"), val = tensor([1, 1, 16, 32, 2])]; + tensor var_220_end_mask_0 = const()[name = tensor("op_220_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_220_squeeze_mask_0 = const()[name = tensor("op_220_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_220_cast_fp16 = slice_by_index(begin = var_220_begin_0, end = var_220_end_0, end_mask = var_220_end_mask_0, squeeze_mask = var_220_squeeze_mask_0, x = q_complex_1_cast_fp16)[name = tensor("op_220_cast_fp16")]; + tensor var_228_begin_0 = const()[name = tensor("op_228_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_228_end_0 = const()[name = tensor("op_228_end_0"), val = tensor([1, 1, 16, 32, 1])]; + tensor var_228_end_mask_0 = const()[name = tensor("op_228_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_228_squeeze_mask_0 = const()[name = tensor("op_228_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_228_cast_fp16 = slice_by_index(begin = var_228_begin_0, end = var_228_end_0, end_mask = var_228_end_mask_0, squeeze_mask = var_228_squeeze_mask_0, x = k_complex_1_cast_fp16)[name = tensor("op_228_cast_fp16")]; + tensor var_236_begin_0 = const()[name = tensor("op_236_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_236_end_0 = const()[name = tensor("op_236_end_0"), val = tensor([1, 1, 16, 32, 2])]; + tensor var_236_end_mask_0 = const()[name = tensor("op_236_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_236_squeeze_mask_0 = const()[name = tensor("op_236_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_236_cast_fp16 = slice_by_index(begin = var_236_begin_0, end = var_236_end_0, end_mask = var_236_end_mask_0, squeeze_mask = var_236_squeeze_mask_0, x = k_complex_1_cast_fp16)[name = tensor("op_236_cast_fp16")]; + tensor freqs_1_to_fp16 = const()[name = tensor("freqs_1_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(75697728)))]; + tensor var_242_cast_fp16 = mul(x = freqs_1_to_fp16, y = ts_5_cast_fp16)[name = tensor("op_242_cast_fp16")]; + tensor rotr_1_cast_fp16 = cos(x = var_242_cast_fp16)[name = tensor("rotr_1_cast_fp16")]; + tensor roti_1_cast_fp16 = sin(x = var_242_cast_fp16)[name = tensor("roti_1_cast_fp16")]; + tensor var_246_cast_fp16 = mul(x = var_212_cast_fp16, y = rotr_1_cast_fp16)[name = tensor("op_246_cast_fp16")]; + tensor var_247_cast_fp16 = mul(x = var_220_cast_fp16, y = roti_1_cast_fp16)[name = tensor("op_247_cast_fp16")]; + tensor qor_1_cast_fp16 = sub(x = var_246_cast_fp16, y = var_247_cast_fp16)[name = tensor("qor_1_cast_fp16")]; + tensor var_250_cast_fp16 = mul(x = var_212_cast_fp16, y = roti_1_cast_fp16)[name = tensor("op_250_cast_fp16")]; + tensor var_251_cast_fp16 = mul(x = var_220_cast_fp16, y = rotr_1_cast_fp16)[name = tensor("op_251_cast_fp16")]; + tensor qoi_1_cast_fp16 = add(x = var_250_cast_fp16, y = var_251_cast_fp16)[name = tensor("qoi_1_cast_fp16")]; + tensor var_254_cast_fp16 = mul(x = var_228_cast_fp16, y = rotr_1_cast_fp16)[name = tensor("op_254_cast_fp16")]; + tensor var_255_cast_fp16 = mul(x = var_236_cast_fp16, y = roti_1_cast_fp16)[name = tensor("op_255_cast_fp16")]; + tensor kor_1_cast_fp16 = sub(x = var_254_cast_fp16, y = var_255_cast_fp16)[name = tensor("kor_1_cast_fp16")]; + tensor var_258_cast_fp16 = mul(x = var_228_cast_fp16, y = roti_1_cast_fp16)[name = tensor("op_258_cast_fp16")]; + tensor var_259_cast_fp16 = mul(x = var_236_cast_fp16, y = rotr_1_cast_fp16)[name = tensor("op_259_cast_fp16")]; + tensor koi_1_cast_fp16 = add(x = var_258_cast_fp16, y = var_259_cast_fp16)[name = tensor("koi_1_cast_fp16")]; + tensor qo_1_axis_0 = const()[name = tensor("qo_1_axis_0"), val = tensor(-1)]; + tensor qo_1_cast_fp16 = stack(axis = qo_1_axis_0, values = (qor_1_cast_fp16, qoi_1_cast_fp16))[name = tensor("qo_1_cast_fp16")]; + tensor ko_1_axis_0 = const()[name = tensor("ko_1_axis_0"), val = tensor(-1)]; + tensor ko_1_cast_fp16 = stack(axis = ko_1_axis_0, values = (kor_1_cast_fp16, koi_1_cast_fp16))[name = tensor("ko_1_cast_fp16")]; + tensor var_288 = const()[name = tensor("op_288"), val = tensor([1, 1, 16, 64])]; + tensor q_3_cast_fp16 = reshape(shape = var_288, x = qo_1_cast_fp16)[name = tensor("q_3_cast_fp16")]; + tensor var_290 = const()[name = tensor("op_290"), val = tensor([1, 1, 16, 64])]; + tensor k_3_cast_fp16 = reshape(shape = var_290, x = ko_1_cast_fp16)[name = tensor("k_3_cast_fp16")]; + tensor _inversed_312_y_0_to_fp16 = const()[name = tensor("_inversed_312_y_0_to_fp16"), val = tensor(0x1p-9)]; + tensor _inversed_312_cast_fp16 = mul(x = ts_5_cast_fp16, y = _inversed_312_y_0_to_fp16)[name = tensor("_inversed_312_cast_fp16")]; + tensor var_313_cast_fp16 = floor(x = _inversed_312_cast_fp16)[name = tensor("op_313_cast_fp16")]; + tensor var_314_to_fp16 = const()[name = tensor("op_314_to_fp16"), val = tensor(0x1p+9)]; + tensor var_315_cast_fp16 = mul(x = var_313_cast_fp16, y = var_314_to_fp16)[name = tensor("op_315_cast_fp16")]; + tensor write_indices_float_3_cast_fp16 = sub(x = ts_5_cast_fp16, y = var_315_cast_fp16)[name = tensor("write_indices_float_3_cast_fp16")]; + tensor var_322_dtype_0 = const()[name = tensor("op_322_dtype_0"), val = tensor("int32")]; + tensor write_indices_1_reps_0 = const()[name = tensor("write_indices_1_reps_0"), val = tensor([1, 1, 16, 64])]; + tensor write_indices_float_3_cast_fp16_to_int32 = cast(dtype = var_322_dtype_0, x = write_indices_float_3_cast_fp16)[name = tensor("cast_138")]; + tensor write_indices_1 = tile(reps = write_indices_1_reps_0, x = write_indices_float_3_cast_fp16_to_int32)[name = tensor("write_indices_1")]; + tensor var_330_begin_0 = const()[name = tensor("op_330_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_330_end_0 = const()[name = tensor("op_330_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor var_330_end_mask_0 = const()[name = tensor("op_330_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_330_squeeze_mask_0 = const()[name = tensor("op_330_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor cache0_to_fp16_dtype_0 = const()[name = tensor("cache0_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor cache0_to_fp16 = cast(dtype = cache0_to_fp16_dtype_0, x = cache0)[name = tensor("cast_137")]; + tensor var_330_cast_fp16 = slice_by_index(begin = var_330_begin_0, end = var_330_end_0, end_mask = var_330_end_mask_0, squeeze_mask = var_330_squeeze_mask_0, x = cache0_to_fp16)[name = tensor("op_330_cast_fp16")]; + tensor var_332_axis_0 = const()[name = tensor("op_332_axis_0"), val = tensor(1)]; + tensor var_332_mode_0 = const()[name = tensor("op_332_mode_0"), val = tensor("update")]; + tensor var_332_validate_indices_0 = const()[name = tensor("op_332_validate_indices_0"), val = tensor(false)]; + tensor var_332_cast_fp16 = scatter_along_axis(axis = var_332_axis_0, data = var_330_cast_fp16, indices = write_indices_1, mode = var_332_mode_0, updates = k_3_cast_fp16, validate_indices = var_332_validate_indices_0)[name = tensor("op_332_cast_fp16")]; + tensor concat_2 = const()[name = tensor("concat_2"), val = tensor([0, 0, 0, 0, 0])]; + tensor concat_3 = const()[name = tensor("concat_3"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_1_internal_tensor_assign_1_stride_0 = const()[name = tensor("new_cache_1_internal_tensor_assign_1_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_1_internal_tensor_assign_1_begin_mask_0 = const()[name = tensor("new_cache_1_internal_tensor_assign_1_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_1_internal_tensor_assign_1_end_mask_0 = const()[name = tensor("new_cache_1_internal_tensor_assign_1_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_1_internal_tensor_assign_1_squeeze_mask_0 = const()[name = tensor("new_cache_1_internal_tensor_assign_1_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_12 = const()[name = tensor("shape_12"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_0 = const()[name = tensor("reduce_prod_0"), val = tensor(1048576)]; + tensor range_1d_0_start_0 = const()[name = tensor("range_1d_0_start_0"), val = tensor(0)]; + tensor range_1d_0_step_0 = const()[name = tensor("range_1d_0_step_0"), val = tensor(1)]; + tensor range_1d_0 = range_1d(end = reduce_prod_0, start = range_1d_0_start_0, step = range_1d_0_step_0)[name = tensor("range_1d_0")]; + tensor reshape_0 = reshape(shape = shape_12, x = range_1d_0)[name = tensor("reshape_0")]; + tensor slice_by_index_0 = slice_by_index(begin = concat_2, begin_mask = new_cache_1_internal_tensor_assign_1_begin_mask_0, end = concat_3, end_mask = new_cache_1_internal_tensor_assign_1_end_mask_0, squeeze_mask = new_cache_1_internal_tensor_assign_1_squeeze_mask_0, stride = new_cache_1_internal_tensor_assign_1_stride_0, x = reshape_0)[name = tensor("slice_by_index_0")]; + tensor reshape_1_shape_0 = const()[name = tensor("reshape_1_shape_0"), val = tensor([-1])]; + tensor reshape_1 = reshape(shape = reshape_1_shape_0, x = slice_by_index_0)[name = tensor("reshape_1")]; + tensor reshape_2_shape_0 = const()[name = tensor("reshape_2_shape_0"), val = tensor([-1])]; + tensor reshape_2_cast_fp16 = reshape(shape = reshape_2_shape_0, x = var_332_cast_fp16)[name = tensor("reshape_2_cast_fp16")]; + tensor reshape_3_shape_0 = const()[name = tensor("reshape_3_shape_0"), val = tensor([-1])]; + tensor reshape_3_cast_fp16 = reshape(shape = reshape_3_shape_0, x = cache0_to_fp16)[name = tensor("reshape_3_cast_fp16")]; + tensor scatter_0_mode_0 = const()[name = tensor("scatter_0_mode_0"), val = tensor("update")]; + tensor scatter_0_axis_0 = const()[name = tensor("scatter_0_axis_0"), val = tensor(0)]; + tensor scatter_0_validate_indices_0 = const()[name = tensor("scatter_0_validate_indices_0"), val = tensor(false)]; + tensor scatter_0_cast_fp16 = scatter(axis = scatter_0_axis_0, data = reshape_3_cast_fp16, indices = reshape_1, mode = scatter_0_mode_0, updates = reshape_2_cast_fp16, validate_indices = scatter_0_validate_indices_0)[name = tensor("scatter_0_cast_fp16")]; + tensor reshape_4_cast_fp16 = reshape(shape = shape_12, x = scatter_0_cast_fp16)[name = tensor("reshape_4_cast_fp16")]; + tensor var_340_begin_0 = const()[name = tensor("op_340_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor var_340_end_0 = const()[name = tensor("op_340_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor var_340_end_mask_0 = const()[name = tensor("op_340_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_340_squeeze_mask_0 = const()[name = tensor("op_340_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor var_340_cast_fp16 = slice_by_index(begin = var_340_begin_0, end = var_340_end_0, end_mask = var_340_end_mask_0, squeeze_mask = var_340_squeeze_mask_0, x = reshape_4_cast_fp16)[name = tensor("op_340_cast_fp16")]; + tensor var_342_axis_0 = const()[name = tensor("op_342_axis_0"), val = tensor(1)]; + tensor var_342_mode_0 = const()[name = tensor("op_342_mode_0"), val = tensor("update")]; + tensor var_342_validate_indices_0 = const()[name = tensor("op_342_validate_indices_0"), val = tensor(false)]; + tensor var_342_cast_fp16 = scatter_along_axis(axis = var_342_axis_0, data = var_340_cast_fp16, indices = write_indices_1, mode = var_342_mode_0, updates = v_1_cast_fp16, validate_indices = var_342_validate_indices_0)[name = tensor("op_342_cast_fp16")]; + tensor concat_4 = const()[name = tensor("concat_4"), val = tensor([1, 0, 0, 0, 0])]; + tensor concat_5 = const()[name = tensor("concat_5"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_1_internal_tensor_assign_2_stride_0 = const()[name = tensor("new_cache_1_internal_tensor_assign_2_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_1_internal_tensor_assign_2_begin_mask_0 = const()[name = tensor("new_cache_1_internal_tensor_assign_2_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_1_internal_tensor_assign_2_end_mask_0 = const()[name = tensor("new_cache_1_internal_tensor_assign_2_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_1_internal_tensor_assign_2_squeeze_mask_0 = const()[name = tensor("new_cache_1_internal_tensor_assign_2_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_13 = const()[name = tensor("shape_13"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_1 = const()[name = tensor("reduce_prod_1"), val = tensor(1048576)]; + tensor range_1d_1_start_0 = const()[name = tensor("range_1d_1_start_0"), val = tensor(0)]; + tensor range_1d_1_step_0 = const()[name = tensor("range_1d_1_step_0"), val = tensor(1)]; + tensor range_1d_1 = range_1d(end = reduce_prod_1, start = range_1d_1_start_0, step = range_1d_1_step_0)[name = tensor("range_1d_1")]; + tensor reshape_5 = reshape(shape = shape_13, x = range_1d_1)[name = tensor("reshape_5")]; + tensor slice_by_index_1 = slice_by_index(begin = concat_4, begin_mask = new_cache_1_internal_tensor_assign_2_begin_mask_0, end = concat_5, end_mask = new_cache_1_internal_tensor_assign_2_end_mask_0, squeeze_mask = new_cache_1_internal_tensor_assign_2_squeeze_mask_0, stride = new_cache_1_internal_tensor_assign_2_stride_0, x = reshape_5)[name = tensor("slice_by_index_1")]; + tensor reshape_6_shape_0 = const()[name = tensor("reshape_6_shape_0"), val = tensor([-1])]; + tensor reshape_6 = reshape(shape = reshape_6_shape_0, x = slice_by_index_1)[name = tensor("reshape_6")]; + tensor reshape_7_shape_0 = const()[name = tensor("reshape_7_shape_0"), val = tensor([-1])]; + tensor reshape_7_cast_fp16 = reshape(shape = reshape_7_shape_0, x = var_342_cast_fp16)[name = tensor("reshape_7_cast_fp16")]; + tensor reshape_8_shape_0 = const()[name = tensor("reshape_8_shape_0"), val = tensor([-1])]; + tensor reshape_8_cast_fp16 = reshape(shape = reshape_8_shape_0, x = reshape_4_cast_fp16)[name = tensor("reshape_8_cast_fp16")]; + tensor scatter_1_mode_0 = const()[name = tensor("scatter_1_mode_0"), val = tensor("update")]; + tensor scatter_1_axis_0 = const()[name = tensor("scatter_1_axis_0"), val = tensor(0)]; + tensor scatter_1_validate_indices_0 = const()[name = tensor("scatter_1_validate_indices_0"), val = tensor(false)]; + tensor scatter_1_cast_fp16 = scatter(axis = scatter_1_axis_0, data = reshape_8_cast_fp16, indices = reshape_6, mode = scatter_1_mode_0, updates = reshape_7_cast_fp16, validate_indices = scatter_1_validate_indices_0)[name = tensor("scatter_1_cast_fp16")]; + tensor reshape_9_cast_fp16 = reshape(shape = shape_13, x = scatter_1_cast_fp16)[name = tensor("reshape_9_cast_fp16")]; + tensor reshape_9_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("reshape_9_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor keys_1_begin_0 = const()[name = tensor("keys_1_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor keys_1_end_0 = const()[name = tensor("keys_1_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor keys_1_end_mask_0 = const()[name = tensor("keys_1_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor keys_1_squeeze_mask_0 = const()[name = tensor("keys_1_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor keys_1_cast_fp16 = slice_by_index(begin = keys_1_begin_0, end = keys_1_end_0, end_mask = keys_1_end_mask_0, squeeze_mask = keys_1_squeeze_mask_0, x = reshape_9_cast_fp16)[name = tensor("keys_1_cast_fp16")]; + tensor values_1_begin_0 = const()[name = tensor("values_1_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor values_1_end_0 = const()[name = tensor("values_1_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor values_1_end_mask_0 = const()[name = tensor("values_1_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor values_1_squeeze_mask_0 = const()[name = tensor("values_1_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor values_1_cast_fp16 = slice_by_index(begin = values_1_begin_0, end = values_1_end_0, end_mask = values_1_end_mask_0, squeeze_mask = values_1_squeeze_mask_0, x = reshape_9_cast_fp16)[name = tensor("values_1_cast_fp16")]; + tensor var_354_cast_fp16 = not_equal(x = keys_1_cast_fp16, y = keys_1_cast_fp16)[name = tensor("op_354_cast_fp16")]; + tensor var_360_to_fp16 = const()[name = tensor("op_360_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(75697856)))]; + tensor keys_3_cast_fp16 = select(a = var_360_to_fp16, b = keys_1_cast_fp16, cond = var_354_cast_fp16)[name = tensor("keys_3_cast_fp16")]; + tensor var_362_cast_fp16 = not_equal(x = values_1_cast_fp16, y = values_1_cast_fp16)[name = tensor("op_362_cast_fp16")]; + tensor values_3_cast_fp16 = select(a = var_360_to_fp16, b = values_1_cast_fp16, cond = var_362_cast_fp16)[name = tensor("values_3_cast_fp16")]; + tensor var_386 = const()[name = tensor("op_386"), val = tensor([0, 2, 1, 3])]; + tensor var_399 = const()[name = tensor("op_399"), val = tensor([1, 1, 1])]; + tensor var_400_cast_fp16 = reshape(shape = var_399, x = position0_to_fp16)[name = tensor("op_400_cast_fp16")]; + tensor var_417_to_fp16 = const()[name = tensor("op_417_to_fp16"), val = tensor(0x1p+0)]; + tensor valid_len_1_cast_fp16 = add(x = var_400_cast_fp16, y = var_417_to_fp16)[name = tensor("valid_len_1_cast_fp16")]; + tensor k_positions_1_promoted_to_fp16 = const()[name = tensor("k_positions_1_promoted_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(76746496)))]; + tensor valid_mask_1_cast_fp16 = less(x = k_positions_1_promoted_to_fp16, y = valid_len_1_cast_fp16)[name = tensor("valid_mask_1_cast_fp16")]; + tensor causal_mask_1_cast_fp16 = less_equal(x = k_positions_1_promoted_to_fp16, y = var_400_cast_fp16)[name = tensor("causal_mask_1_cast_fp16")]; + tensor attn_mask_1 = logical_and(x = valid_mask_1_cast_fp16, y = causal_mask_1_cast_fp16)[name = tensor("attn_mask_1")]; + tensor attn_mask_3_axes_0 = const()[name = tensor("attn_mask_3_axes_0"), val = tensor([1])]; + tensor attn_mask_3 = expand_dims(axes = attn_mask_3_axes_0, x = attn_mask_1)[name = tensor("attn_mask_3")]; + tensor var_435_transpose_x_0 = const()[name = tensor("op_435_transpose_x_0"), val = tensor(false)]; + tensor var_435_transpose_y_0 = const()[name = tensor("op_435_transpose_y_0"), val = tensor(false)]; + tensor transpose_18_perm_0 = const()[name = tensor("transpose_18_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_19_perm_0 = const()[name = tensor("transpose_19_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_19 = transpose(perm = transpose_19_perm_0, x = keys_3_cast_fp16)[name = tensor("transpose_51")]; + tensor transpose_18 = transpose(perm = transpose_18_perm_0, x = q_3_cast_fp16)[name = tensor("transpose_52")]; + tensor var_435_cast_fp16 = matmul(transpose_x = var_435_transpose_x_0, transpose_y = var_435_transpose_y_0, x = transpose_18, y = transpose_19)[name = tensor("op_435_cast_fp16")]; + tensor var_429_to_fp16 = const()[name = tensor("op_429_to_fp16"), val = tensor([0x1p-3])]; + tensor attn_weights_1_cast_fp16 = mul(x = var_435_cast_fp16, y = var_429_to_fp16)[name = tensor("attn_weights_1_cast_fp16")]; + tensor var_437 = logical_not(x = attn_mask_3)[name = tensor("op_437")]; + tensor var_438_to_fp16 = const()[name = tensor("op_438_to_fp16"), val = tensor(-inf)]; + tensor attn_weights_3_cast_fp16 = select(a = var_438_to_fp16, b = attn_weights_1_cast_fp16, cond = var_437)[name = tensor("attn_weights_3_cast_fp16")]; + tensor var_440 = const()[name = tensor("op_440"), val = tensor(-1)]; + tensor attn_weights_5_cast_fp16 = softmax(axis = var_440, x = attn_weights_3_cast_fp16)[name = tensor("attn_weights_5_cast_fp16")]; + tensor attn_output_1_transpose_x_0 = const()[name = tensor("attn_output_1_transpose_x_0"), val = tensor(false)]; + tensor attn_output_1_transpose_y_0 = const()[name = tensor("attn_output_1_transpose_y_0"), val = tensor(false)]; + tensor values_5_cast_fp16 = transpose(perm = var_386, x = values_3_cast_fp16)[name = tensor("transpose_53")]; + tensor attn_output_1_cast_fp16 = matmul(transpose_x = attn_output_1_transpose_x_0, transpose_y = attn_output_1_transpose_y_0, x = attn_weights_5_cast_fp16, y = values_5_cast_fp16)[name = tensor("attn_output_1_cast_fp16")]; + tensor var_448 = const()[name = tensor("op_448"), val = tensor([0, 2, 1, 3])]; + tensor var_451 = const()[name = tensor("op_451"), val = tensor([1, 1, 1024])]; + tensor var_449_cast_fp16 = transpose(perm = var_448, x = attn_output_1_cast_fp16)[name = tensor("transpose_50")]; + tensor input_5_cast_fp16 = reshape(shape = var_451, x = var_449_cast_fp16)[name = tensor("input_5_cast_fp16")]; + tensor linear_2_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = attn0_out_proj_weight_cast_fp16, x = input_5_cast_fp16)[name = tensor("linear_2_cast_fp16")]; + tensor var_457_to_fp16 = const()[name = tensor("op_457_to_fp16"), val = tensor(0x1p+0)]; + tensor var_458_cast_fp16 = add(x = position0_to_fp16, y = var_457_to_fp16)[name = tensor("op_458_cast_fp16")]; + tensor var_458_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_458_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor input_7_cast_fp16 = add(x = linear_0_cast_fp16, y = linear_2_cast_fp16)[name = tensor("input_7_cast_fp16")]; + tensor input_9_axes_0 = const()[name = tensor("input_9_axes_0"), val = tensor([-1])]; + tensor norm0_2_weight_to_fp16 = const()[name = tensor("norm0_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(76747584)))]; + tensor norm0_2_bias_to_fp16 = const()[name = tensor("norm0_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(76749696)))]; + tensor var_462_to_fp16 = const()[name = tensor("op_462_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_9_cast_fp16 = layer_norm(axes = input_9_axes_0, beta = norm0_2_bias_to_fp16, epsilon = var_462_to_fp16, gamma = norm0_2_weight_to_fp16, x = input_7_cast_fp16)[name = tensor("input_9_cast_fp16")]; + tensor linear_3_bias_0_to_fp16 = const()[name = tensor("linear_3_bias_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(76751808)))]; + tensor linear_3_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = linear0_1_weight_cast_fp16, x = input_9_cast_fp16)[name = tensor("linear_3_cast_fp16")]; + tensor input_11_mode_0 = const()[name = tensor("input_11_mode_0"), val = tensor("EXACT")]; + tensor input_11_cast_fp16 = gelu(mode = input_11_mode_0, x = linear_3_cast_fp16)[name = tensor("input_11_cast_fp16")]; + tensor linear_4_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = linear0_2_weight_cast_fp16, x = input_11_cast_fp16)[name = tensor("linear_4_cast_fp16")]; + tensor input_13_cast_fp16 = add(x = input_7_cast_fp16, y = linear_4_cast_fp16)[name = tensor("input_13_cast_fp16")]; + tensor x_3_axes_0 = const()[name = tensor("x_3_axes_0"), val = tensor([-1])]; + tensor norm1_1_weight_to_fp16 = const()[name = tensor("norm1_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(76760064)))]; + tensor norm1_1_bias_to_fp16 = const()[name = tensor("norm1_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(76762176)))]; + tensor var_479_to_fp16 = const()[name = tensor("op_479_to_fp16"), val = tensor(0x1.5p-17)]; + tensor x_3_cast_fp16 = layer_norm(axes = x_3_axes_0, beta = norm1_1_bias_to_fp16, epsilon = var_479_to_fp16, gamma = norm1_1_weight_to_fp16, x = input_13_cast_fp16)[name = tensor("x_3_cast_fp16")]; + tensor linear_5_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = attn1_in_proj_weight_cast_fp16, x = x_3_cast_fp16)[name = tensor("linear_5_cast_fp16")]; + tensor var_515 = const()[name = tensor("op_515"), val = tensor([1, 1, 3, 16, 64])]; + tensor qkv_3_cast_fp16 = reshape(shape = var_515, x = linear_5_cast_fp16)[name = tensor("qkv_3_cast_fp16")]; + tensor q_7_begin_0 = const()[name = tensor("q_7_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor q_7_end_0 = const()[name = tensor("q_7_end_0"), val = tensor([1, 1, 1, 16, 64])]; + tensor q_7_end_mask_0 = const()[name = tensor("q_7_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor q_7_squeeze_mask_0 = const()[name = tensor("q_7_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor q_7_cast_fp16 = slice_by_index(begin = q_7_begin_0, end = q_7_end_0, end_mask = q_7_end_mask_0, squeeze_mask = q_7_squeeze_mask_0, x = qkv_3_cast_fp16)[name = tensor("q_7_cast_fp16")]; + tensor k_5_begin_0 = const()[name = tensor("k_5_begin_0"), val = tensor([0, 0, 1, 0, 0])]; + tensor k_5_end_0 = const()[name = tensor("k_5_end_0"), val = tensor([1, 1, 2, 16, 64])]; + tensor k_5_end_mask_0 = const()[name = tensor("k_5_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor k_5_squeeze_mask_0 = const()[name = tensor("k_5_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor k_5_cast_fp16 = slice_by_index(begin = k_5_begin_0, end = k_5_end_0, end_mask = k_5_end_mask_0, squeeze_mask = k_5_squeeze_mask_0, x = qkv_3_cast_fp16)[name = tensor("k_5_cast_fp16")]; + tensor v_3_begin_0 = const()[name = tensor("v_3_begin_0"), val = tensor([0, 0, 2, 0, 0])]; + tensor v_3_end_0 = const()[name = tensor("v_3_end_0"), val = tensor([1, 1, 3, 16, 64])]; + tensor v_3_end_mask_0 = const()[name = tensor("v_3_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor v_3_squeeze_mask_0 = const()[name = tensor("v_3_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor v_3_cast_fp16 = slice_by_index(begin = v_3_begin_0, end = v_3_end_0, end_mask = v_3_end_mask_0, squeeze_mask = v_3_squeeze_mask_0, x = qkv_3_cast_fp16)[name = tensor("v_3_cast_fp16")]; + tensor var_619 = const()[name = tensor("op_619"), val = tensor([1, 1, 1, 1])]; + tensor position1_to_fp16_dtype_0 = const()[name = tensor("position1_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor position1_to_fp16 = cast(dtype = position1_to_fp16_dtype_0, x = position1)[name = tensor("cast_134")]; + tensor ts_11_cast_fp16 = reshape(shape = var_619, x = position1_to_fp16)[name = tensor("ts_11_cast_fp16")]; + tensor var_623 = const()[name = tensor("op_623"), val = tensor([1, 1, 16, 32, 2])]; + tensor q_complex_3_cast_fp16 = reshape(shape = var_623, x = q_7_cast_fp16)[name = tensor("q_complex_3_cast_fp16")]; + tensor var_627 = const()[name = tensor("op_627"), val = tensor([1, 1, 16, 32, 2])]; + tensor k_complex_3_cast_fp16 = reshape(shape = var_627, x = k_5_cast_fp16)[name = tensor("k_complex_3_cast_fp16")]; + tensor var_631_begin_0 = const()[name = tensor("op_631_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_631_end_0 = const()[name = tensor("op_631_end_0"), val = tensor([1, 1, 16, 32, 1])]; + tensor var_631_end_mask_0 = const()[name = tensor("op_631_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_631_squeeze_mask_0 = const()[name = tensor("op_631_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_631_cast_fp16 = slice_by_index(begin = var_631_begin_0, end = var_631_end_0, end_mask = var_631_end_mask_0, squeeze_mask = var_631_squeeze_mask_0, x = q_complex_3_cast_fp16)[name = tensor("op_631_cast_fp16")]; + tensor var_639_begin_0 = const()[name = tensor("op_639_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_639_end_0 = const()[name = tensor("op_639_end_0"), val = tensor([1, 1, 16, 32, 2])]; + tensor var_639_end_mask_0 = const()[name = tensor("op_639_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_639_squeeze_mask_0 = const()[name = tensor("op_639_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_639_cast_fp16 = slice_by_index(begin = var_639_begin_0, end = var_639_end_0, end_mask = var_639_end_mask_0, squeeze_mask = var_639_squeeze_mask_0, x = q_complex_3_cast_fp16)[name = tensor("op_639_cast_fp16")]; + tensor var_647_begin_0 = const()[name = tensor("op_647_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_647_end_0 = const()[name = tensor("op_647_end_0"), val = tensor([1, 1, 16, 32, 1])]; + tensor var_647_end_mask_0 = const()[name = tensor("op_647_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_647_squeeze_mask_0 = const()[name = tensor("op_647_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_647_cast_fp16 = slice_by_index(begin = var_647_begin_0, end = var_647_end_0, end_mask = var_647_end_mask_0, squeeze_mask = var_647_squeeze_mask_0, x = k_complex_3_cast_fp16)[name = tensor("op_647_cast_fp16")]; + tensor var_655_begin_0 = const()[name = tensor("op_655_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_655_end_0 = const()[name = tensor("op_655_end_0"), val = tensor([1, 1, 16, 32, 2])]; + tensor var_655_end_mask_0 = const()[name = tensor("op_655_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_655_squeeze_mask_0 = const()[name = tensor("op_655_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_655_cast_fp16 = slice_by_index(begin = var_655_begin_0, end = var_655_end_0, end_mask = var_655_end_mask_0, squeeze_mask = var_655_squeeze_mask_0, x = k_complex_3_cast_fp16)[name = tensor("op_655_cast_fp16")]; + tensor freqs_3_to_fp16 = const()[name = tensor("freqs_3_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(76764288)))]; + tensor var_661_cast_fp16 = mul(x = freqs_3_to_fp16, y = ts_11_cast_fp16)[name = tensor("op_661_cast_fp16")]; + tensor rotr_3_cast_fp16 = cos(x = var_661_cast_fp16)[name = tensor("rotr_3_cast_fp16")]; + tensor roti_3_cast_fp16 = sin(x = var_661_cast_fp16)[name = tensor("roti_3_cast_fp16")]; + tensor var_665_cast_fp16 = mul(x = var_631_cast_fp16, y = rotr_3_cast_fp16)[name = tensor("op_665_cast_fp16")]; + tensor var_666_cast_fp16 = mul(x = var_639_cast_fp16, y = roti_3_cast_fp16)[name = tensor("op_666_cast_fp16")]; + tensor qor_5_cast_fp16 = sub(x = var_665_cast_fp16, y = var_666_cast_fp16)[name = tensor("qor_5_cast_fp16")]; + tensor var_669_cast_fp16 = mul(x = var_631_cast_fp16, y = roti_3_cast_fp16)[name = tensor("op_669_cast_fp16")]; + tensor var_670_cast_fp16 = mul(x = var_639_cast_fp16, y = rotr_3_cast_fp16)[name = tensor("op_670_cast_fp16")]; + tensor qoi_5_cast_fp16 = add(x = var_669_cast_fp16, y = var_670_cast_fp16)[name = tensor("qoi_5_cast_fp16")]; + tensor var_673_cast_fp16 = mul(x = var_647_cast_fp16, y = rotr_3_cast_fp16)[name = tensor("op_673_cast_fp16")]; + tensor var_674_cast_fp16 = mul(x = var_655_cast_fp16, y = roti_3_cast_fp16)[name = tensor("op_674_cast_fp16")]; + tensor kor_5_cast_fp16 = sub(x = var_673_cast_fp16, y = var_674_cast_fp16)[name = tensor("kor_5_cast_fp16")]; + tensor var_677_cast_fp16 = mul(x = var_647_cast_fp16, y = roti_3_cast_fp16)[name = tensor("op_677_cast_fp16")]; + tensor var_678_cast_fp16 = mul(x = var_655_cast_fp16, y = rotr_3_cast_fp16)[name = tensor("op_678_cast_fp16")]; + tensor koi_5_cast_fp16 = add(x = var_677_cast_fp16, y = var_678_cast_fp16)[name = tensor("koi_5_cast_fp16")]; + tensor qo_3_axis_0 = const()[name = tensor("qo_3_axis_0"), val = tensor(-1)]; + tensor qo_3_cast_fp16 = stack(axis = qo_3_axis_0, values = (qor_5_cast_fp16, qoi_5_cast_fp16))[name = tensor("qo_3_cast_fp16")]; + tensor ko_3_axis_0 = const()[name = tensor("ko_3_axis_0"), val = tensor(-1)]; + tensor ko_3_cast_fp16 = stack(axis = ko_3_axis_0, values = (kor_5_cast_fp16, koi_5_cast_fp16))[name = tensor("ko_3_cast_fp16")]; + tensor var_707 = const()[name = tensor("op_707"), val = tensor([1, 1, 16, 64])]; + tensor q_9_cast_fp16 = reshape(shape = var_707, x = qo_3_cast_fp16)[name = tensor("q_9_cast_fp16")]; + tensor var_709 = const()[name = tensor("op_709"), val = tensor([1, 1, 16, 64])]; + tensor k_7_cast_fp16 = reshape(shape = var_709, x = ko_3_cast_fp16)[name = tensor("k_7_cast_fp16")]; + tensor _inversed_731_y_0_to_fp16 = const()[name = tensor("_inversed_731_y_0_to_fp16"), val = tensor(0x1p-9)]; + tensor _inversed_731_cast_fp16 = mul(x = ts_11_cast_fp16, y = _inversed_731_y_0_to_fp16)[name = tensor("_inversed_731_cast_fp16")]; + tensor var_732_cast_fp16 = floor(x = _inversed_731_cast_fp16)[name = tensor("op_732_cast_fp16")]; + tensor var_733_to_fp16 = const()[name = tensor("op_733_to_fp16"), val = tensor(0x1p+9)]; + tensor var_734_cast_fp16 = mul(x = var_732_cast_fp16, y = var_733_to_fp16)[name = tensor("op_734_cast_fp16")]; + tensor write_indices_float_7_cast_fp16 = sub(x = ts_11_cast_fp16, y = var_734_cast_fp16)[name = tensor("write_indices_float_7_cast_fp16")]; + tensor var_741_dtype_0 = const()[name = tensor("op_741_dtype_0"), val = tensor("int32")]; + tensor write_indices_3_reps_0 = const()[name = tensor("write_indices_3_reps_0"), val = tensor([1, 1, 16, 64])]; + tensor write_indices_float_7_cast_fp16_to_int32 = cast(dtype = var_741_dtype_0, x = write_indices_float_7_cast_fp16)[name = tensor("cast_133")]; + tensor write_indices_3 = tile(reps = write_indices_3_reps_0, x = write_indices_float_7_cast_fp16_to_int32)[name = tensor("write_indices_3")]; + tensor var_749_begin_0 = const()[name = tensor("op_749_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_749_end_0 = const()[name = tensor("op_749_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor var_749_end_mask_0 = const()[name = tensor("op_749_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_749_squeeze_mask_0 = const()[name = tensor("op_749_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor cache1_to_fp16_dtype_0 = const()[name = tensor("cache1_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor cache1_to_fp16 = cast(dtype = cache1_to_fp16_dtype_0, x = cache1)[name = tensor("cast_132")]; + tensor var_749_cast_fp16 = slice_by_index(begin = var_749_begin_0, end = var_749_end_0, end_mask = var_749_end_mask_0, squeeze_mask = var_749_squeeze_mask_0, x = cache1_to_fp16)[name = tensor("op_749_cast_fp16")]; + tensor var_751_axis_0 = const()[name = tensor("op_751_axis_0"), val = tensor(1)]; + tensor var_751_mode_0 = const()[name = tensor("op_751_mode_0"), val = tensor("update")]; + tensor var_751_validate_indices_0 = const()[name = tensor("op_751_validate_indices_0"), val = tensor(false)]; + tensor var_751_cast_fp16 = scatter_along_axis(axis = var_751_axis_0, data = var_749_cast_fp16, indices = write_indices_3, mode = var_751_mode_0, updates = k_7_cast_fp16, validate_indices = var_751_validate_indices_0)[name = tensor("op_751_cast_fp16")]; + tensor concat_9 = const()[name = tensor("concat_9"), val = tensor([0, 0, 0, 0, 0])]; + tensor concat_10 = const()[name = tensor("concat_10"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_3_internal_tensor_assign_1_stride_0 = const()[name = tensor("new_cache_3_internal_tensor_assign_1_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_3_internal_tensor_assign_1_begin_mask_0 = const()[name = tensor("new_cache_3_internal_tensor_assign_1_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_3_internal_tensor_assign_1_end_mask_0 = const()[name = tensor("new_cache_3_internal_tensor_assign_1_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_3_internal_tensor_assign_1_squeeze_mask_0 = const()[name = tensor("new_cache_3_internal_tensor_assign_1_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_14 = const()[name = tensor("shape_14"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_2 = const()[name = tensor("reduce_prod_2"), val = tensor(1048576)]; + tensor range_1d_2_start_0 = const()[name = tensor("range_1d_2_start_0"), val = tensor(0)]; + tensor range_1d_2_step_0 = const()[name = tensor("range_1d_2_step_0"), val = tensor(1)]; + tensor range_1d_2 = range_1d(end = reduce_prod_2, start = range_1d_2_start_0, step = range_1d_2_step_0)[name = tensor("range_1d_2")]; + tensor reshape_10 = reshape(shape = shape_14, x = range_1d_2)[name = tensor("reshape_10")]; + tensor slice_by_index_2 = slice_by_index(begin = concat_9, begin_mask = new_cache_3_internal_tensor_assign_1_begin_mask_0, end = concat_10, end_mask = new_cache_3_internal_tensor_assign_1_end_mask_0, squeeze_mask = new_cache_3_internal_tensor_assign_1_squeeze_mask_0, stride = new_cache_3_internal_tensor_assign_1_stride_0, x = reshape_10)[name = tensor("slice_by_index_2")]; + tensor reshape_11_shape_0 = const()[name = tensor("reshape_11_shape_0"), val = tensor([-1])]; + tensor reshape_11 = reshape(shape = reshape_11_shape_0, x = slice_by_index_2)[name = tensor("reshape_11")]; + tensor reshape_12_shape_0 = const()[name = tensor("reshape_12_shape_0"), val = tensor([-1])]; + tensor reshape_12_cast_fp16 = reshape(shape = reshape_12_shape_0, x = var_751_cast_fp16)[name = tensor("reshape_12_cast_fp16")]; + tensor reshape_13_shape_0 = const()[name = tensor("reshape_13_shape_0"), val = tensor([-1])]; + tensor reshape_13_cast_fp16 = reshape(shape = reshape_13_shape_0, x = cache1_to_fp16)[name = tensor("reshape_13_cast_fp16")]; + tensor scatter_2_mode_0 = const()[name = tensor("scatter_2_mode_0"), val = tensor("update")]; + tensor scatter_2_axis_0 = const()[name = tensor("scatter_2_axis_0"), val = tensor(0)]; + tensor scatter_2_validate_indices_0 = const()[name = tensor("scatter_2_validate_indices_0"), val = tensor(false)]; + tensor scatter_2_cast_fp16 = scatter(axis = scatter_2_axis_0, data = reshape_13_cast_fp16, indices = reshape_11, mode = scatter_2_mode_0, updates = reshape_12_cast_fp16, validate_indices = scatter_2_validate_indices_0)[name = tensor("scatter_2_cast_fp16")]; + tensor reshape_14_cast_fp16 = reshape(shape = shape_14, x = scatter_2_cast_fp16)[name = tensor("reshape_14_cast_fp16")]; + tensor var_759_begin_0 = const()[name = tensor("op_759_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor var_759_end_0 = const()[name = tensor("op_759_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor var_759_end_mask_0 = const()[name = tensor("op_759_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_759_squeeze_mask_0 = const()[name = tensor("op_759_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor var_759_cast_fp16 = slice_by_index(begin = var_759_begin_0, end = var_759_end_0, end_mask = var_759_end_mask_0, squeeze_mask = var_759_squeeze_mask_0, x = reshape_14_cast_fp16)[name = tensor("op_759_cast_fp16")]; + tensor var_761_axis_0 = const()[name = tensor("op_761_axis_0"), val = tensor(1)]; + tensor var_761_mode_0 = const()[name = tensor("op_761_mode_0"), val = tensor("update")]; + tensor var_761_validate_indices_0 = const()[name = tensor("op_761_validate_indices_0"), val = tensor(false)]; + tensor var_761_cast_fp16 = scatter_along_axis(axis = var_761_axis_0, data = var_759_cast_fp16, indices = write_indices_3, mode = var_761_mode_0, updates = v_3_cast_fp16, validate_indices = var_761_validate_indices_0)[name = tensor("op_761_cast_fp16")]; + tensor concat_11 = const()[name = tensor("concat_11"), val = tensor([1, 0, 0, 0, 0])]; + tensor concat_12 = const()[name = tensor("concat_12"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_3_internal_tensor_assign_2_stride_0 = const()[name = tensor("new_cache_3_internal_tensor_assign_2_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_3_internal_tensor_assign_2_begin_mask_0 = const()[name = tensor("new_cache_3_internal_tensor_assign_2_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_3_internal_tensor_assign_2_end_mask_0 = const()[name = tensor("new_cache_3_internal_tensor_assign_2_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_3_internal_tensor_assign_2_squeeze_mask_0 = const()[name = tensor("new_cache_3_internal_tensor_assign_2_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_15 = const()[name = tensor("shape_15"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_3 = const()[name = tensor("reduce_prod_3"), val = tensor(1048576)]; + tensor range_1d_3_start_0 = const()[name = tensor("range_1d_3_start_0"), val = tensor(0)]; + tensor range_1d_3_step_0 = const()[name = tensor("range_1d_3_step_0"), val = tensor(1)]; + tensor range_1d_3 = range_1d(end = reduce_prod_3, start = range_1d_3_start_0, step = range_1d_3_step_0)[name = tensor("range_1d_3")]; + tensor reshape_15 = reshape(shape = shape_15, x = range_1d_3)[name = tensor("reshape_15")]; + tensor slice_by_index_3 = slice_by_index(begin = concat_11, begin_mask = new_cache_3_internal_tensor_assign_2_begin_mask_0, end = concat_12, end_mask = new_cache_3_internal_tensor_assign_2_end_mask_0, squeeze_mask = new_cache_3_internal_tensor_assign_2_squeeze_mask_0, stride = new_cache_3_internal_tensor_assign_2_stride_0, x = reshape_15)[name = tensor("slice_by_index_3")]; + tensor reshape_16_shape_0 = const()[name = tensor("reshape_16_shape_0"), val = tensor([-1])]; + tensor reshape_16 = reshape(shape = reshape_16_shape_0, x = slice_by_index_3)[name = tensor("reshape_16")]; + tensor reshape_17_shape_0 = const()[name = tensor("reshape_17_shape_0"), val = tensor([-1])]; + tensor reshape_17_cast_fp16 = reshape(shape = reshape_17_shape_0, x = var_761_cast_fp16)[name = tensor("reshape_17_cast_fp16")]; + tensor reshape_18_shape_0 = const()[name = tensor("reshape_18_shape_0"), val = tensor([-1])]; + tensor reshape_18_cast_fp16 = reshape(shape = reshape_18_shape_0, x = reshape_14_cast_fp16)[name = tensor("reshape_18_cast_fp16")]; + tensor scatter_3_mode_0 = const()[name = tensor("scatter_3_mode_0"), val = tensor("update")]; + tensor scatter_3_axis_0 = const()[name = tensor("scatter_3_axis_0"), val = tensor(0)]; + tensor scatter_3_validate_indices_0 = const()[name = tensor("scatter_3_validate_indices_0"), val = tensor(false)]; + tensor scatter_3_cast_fp16 = scatter(axis = scatter_3_axis_0, data = reshape_18_cast_fp16, indices = reshape_16, mode = scatter_3_mode_0, updates = reshape_17_cast_fp16, validate_indices = scatter_3_validate_indices_0)[name = tensor("scatter_3_cast_fp16")]; + tensor reshape_19_cast_fp16 = reshape(shape = shape_15, x = scatter_3_cast_fp16)[name = tensor("reshape_19_cast_fp16")]; + tensor reshape_19_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("reshape_19_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor keys_7_begin_0 = const()[name = tensor("keys_7_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor keys_7_end_0 = const()[name = tensor("keys_7_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor keys_7_end_mask_0 = const()[name = tensor("keys_7_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor keys_7_squeeze_mask_0 = const()[name = tensor("keys_7_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor keys_7_cast_fp16 = slice_by_index(begin = keys_7_begin_0, end = keys_7_end_0, end_mask = keys_7_end_mask_0, squeeze_mask = keys_7_squeeze_mask_0, x = reshape_19_cast_fp16)[name = tensor("keys_7_cast_fp16")]; + tensor values_7_begin_0 = const()[name = tensor("values_7_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor values_7_end_0 = const()[name = tensor("values_7_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor values_7_end_mask_0 = const()[name = tensor("values_7_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor values_7_squeeze_mask_0 = const()[name = tensor("values_7_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor values_7_cast_fp16 = slice_by_index(begin = values_7_begin_0, end = values_7_end_0, end_mask = values_7_end_mask_0, squeeze_mask = values_7_squeeze_mask_0, x = reshape_19_cast_fp16)[name = tensor("values_7_cast_fp16")]; + tensor var_773_cast_fp16 = not_equal(x = keys_7_cast_fp16, y = keys_7_cast_fp16)[name = tensor("op_773_cast_fp16")]; + tensor keys_9_cast_fp16 = select(a = var_360_to_fp16, b = keys_7_cast_fp16, cond = var_773_cast_fp16)[name = tensor("keys_9_cast_fp16")]; + tensor var_781_cast_fp16 = not_equal(x = values_7_cast_fp16, y = values_7_cast_fp16)[name = tensor("op_781_cast_fp16")]; + tensor values_9_cast_fp16 = select(a = var_360_to_fp16, b = values_7_cast_fp16, cond = var_781_cast_fp16)[name = tensor("values_9_cast_fp16")]; + tensor var_805 = const()[name = tensor("op_805"), val = tensor([0, 2, 1, 3])]; + tensor var_818 = const()[name = tensor("op_818"), val = tensor([1, 1, 1])]; + tensor var_819_cast_fp16 = reshape(shape = var_818, x = position1_to_fp16)[name = tensor("op_819_cast_fp16")]; + tensor var_836_to_fp16 = const()[name = tensor("op_836_to_fp16"), val = tensor(0x1p+0)]; + tensor valid_len_3_cast_fp16 = add(x = var_819_cast_fp16, y = var_836_to_fp16)[name = tensor("valid_len_3_cast_fp16")]; + tensor valid_mask_3_cast_fp16 = less(x = k_positions_1_promoted_to_fp16, y = valid_len_3_cast_fp16)[name = tensor("valid_mask_3_cast_fp16")]; + tensor causal_mask_3_cast_fp16 = less_equal(x = k_positions_1_promoted_to_fp16, y = var_819_cast_fp16)[name = tensor("causal_mask_3_cast_fp16")]; + tensor attn_mask_5 = logical_and(x = valid_mask_3_cast_fp16, y = causal_mask_3_cast_fp16)[name = tensor("attn_mask_5")]; + tensor attn_mask_7_axes_0 = const()[name = tensor("attn_mask_7_axes_0"), val = tensor([1])]; + tensor attn_mask_7 = expand_dims(axes = attn_mask_7_axes_0, x = attn_mask_5)[name = tensor("attn_mask_7")]; + tensor var_854_transpose_x_0 = const()[name = tensor("op_854_transpose_x_0"), val = tensor(false)]; + tensor var_854_transpose_y_0 = const()[name = tensor("op_854_transpose_y_0"), val = tensor(false)]; + tensor transpose_20_perm_0 = const()[name = tensor("transpose_20_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_21_perm_0 = const()[name = tensor("transpose_21_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_21 = transpose(perm = transpose_21_perm_0, x = keys_9_cast_fp16)[name = tensor("transpose_47")]; + tensor transpose_20 = transpose(perm = transpose_20_perm_0, x = q_9_cast_fp16)[name = tensor("transpose_48")]; + tensor var_854_cast_fp16 = matmul(transpose_x = var_854_transpose_x_0, transpose_y = var_854_transpose_y_0, x = transpose_20, y = transpose_21)[name = tensor("op_854_cast_fp16")]; + tensor var_848_to_fp16 = const()[name = tensor("op_848_to_fp16"), val = tensor([0x1p-3])]; + tensor attn_weights_7_cast_fp16 = mul(x = var_854_cast_fp16, y = var_848_to_fp16)[name = tensor("attn_weights_7_cast_fp16")]; + tensor var_856 = logical_not(x = attn_mask_7)[name = tensor("op_856")]; + tensor var_857_to_fp16 = const()[name = tensor("op_857_to_fp16"), val = tensor(-inf)]; + tensor attn_weights_9_cast_fp16 = select(a = var_857_to_fp16, b = attn_weights_7_cast_fp16, cond = var_856)[name = tensor("attn_weights_9_cast_fp16")]; + tensor var_859 = const()[name = tensor("op_859"), val = tensor(-1)]; + tensor attn_weights_11_cast_fp16 = softmax(axis = var_859, x = attn_weights_9_cast_fp16)[name = tensor("attn_weights_11_cast_fp16")]; + tensor attn_output_3_transpose_x_0 = const()[name = tensor("attn_output_3_transpose_x_0"), val = tensor(false)]; + tensor attn_output_3_transpose_y_0 = const()[name = tensor("attn_output_3_transpose_y_0"), val = tensor(false)]; + tensor values_11_cast_fp16 = transpose(perm = var_805, x = values_9_cast_fp16)[name = tensor("transpose_49")]; + tensor attn_output_3_cast_fp16 = matmul(transpose_x = attn_output_3_transpose_x_0, transpose_y = attn_output_3_transpose_y_0, x = attn_weights_11_cast_fp16, y = values_11_cast_fp16)[name = tensor("attn_output_3_cast_fp16")]; + tensor var_867 = const()[name = tensor("op_867"), val = tensor([0, 2, 1, 3])]; + tensor var_870 = const()[name = tensor("op_870"), val = tensor([1, 1, 1024])]; + tensor var_868_cast_fp16 = transpose(perm = var_867, x = attn_output_3_cast_fp16)[name = tensor("transpose_46")]; + tensor input_15_cast_fp16 = reshape(shape = var_870, x = var_868_cast_fp16)[name = tensor("input_15_cast_fp16")]; + tensor linear_6_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = attn1_out_proj_weight_cast_fp16, x = input_15_cast_fp16)[name = tensor("linear_6_cast_fp16")]; + tensor var_876_to_fp16 = const()[name = tensor("op_876_to_fp16"), val = tensor(0x1p+0)]; + tensor var_877_cast_fp16 = add(x = position1_to_fp16, y = var_876_to_fp16)[name = tensor("op_877_cast_fp16")]; + tensor var_877_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_877_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor input_17_cast_fp16 = add(x = input_13_cast_fp16, y = linear_6_cast_fp16)[name = tensor("input_17_cast_fp16")]; + tensor input_19_axes_0 = const()[name = tensor("input_19_axes_0"), val = tensor([-1])]; + tensor norm1_2_weight_to_fp16 = const()[name = tensor("norm1_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(76764416)))]; + tensor norm1_2_bias_to_fp16 = const()[name = tensor("norm1_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(76766528)))]; + tensor var_881_to_fp16 = const()[name = tensor("op_881_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_19_cast_fp16 = layer_norm(axes = input_19_axes_0, beta = norm1_2_bias_to_fp16, epsilon = var_881_to_fp16, gamma = norm1_2_weight_to_fp16, x = input_17_cast_fp16)[name = tensor("input_19_cast_fp16")]; + tensor linear_7_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = linear1_1_weight_cast_fp16, x = input_19_cast_fp16)[name = tensor("linear_7_cast_fp16")]; + tensor input_21_mode_0 = const()[name = tensor("input_21_mode_0"), val = tensor("EXACT")]; + tensor input_21_cast_fp16 = gelu(mode = input_21_mode_0, x = linear_7_cast_fp16)[name = tensor("input_21_cast_fp16")]; + tensor linear_8_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = linear1_2_weight_cast_fp16, x = input_21_cast_fp16)[name = tensor("linear_8_cast_fp16")]; + tensor input_23_cast_fp16 = add(x = input_17_cast_fp16, y = linear_8_cast_fp16)[name = tensor("input_23_cast_fp16")]; + tensor x_5_axes_0 = const()[name = tensor("x_5_axes_0"), val = tensor([-1])]; + tensor norm2_1_weight_to_fp16 = const()[name = tensor("norm2_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(76768640)))]; + tensor norm2_1_bias_to_fp16 = const()[name = tensor("norm2_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(76770752)))]; + tensor var_898_to_fp16 = const()[name = tensor("op_898_to_fp16"), val = tensor(0x1.5p-17)]; + tensor x_5_cast_fp16 = layer_norm(axes = x_5_axes_0, beta = norm2_1_bias_to_fp16, epsilon = var_898_to_fp16, gamma = norm2_1_weight_to_fp16, x = input_23_cast_fp16)[name = tensor("x_5_cast_fp16")]; + tensor linear_9_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = attn2_in_proj_weight_cast_fp16, x = x_5_cast_fp16)[name = tensor("linear_9_cast_fp16")]; + tensor var_934 = const()[name = tensor("op_934"), val = tensor([1, 1, 3, 16, 64])]; + tensor qkv_5_cast_fp16 = reshape(shape = var_934, x = linear_9_cast_fp16)[name = tensor("qkv_5_cast_fp16")]; + tensor q_13_begin_0 = const()[name = tensor("q_13_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor q_13_end_0 = const()[name = tensor("q_13_end_0"), val = tensor([1, 1, 1, 16, 64])]; + tensor q_13_end_mask_0 = const()[name = tensor("q_13_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor q_13_squeeze_mask_0 = const()[name = tensor("q_13_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor q_13_cast_fp16 = slice_by_index(begin = q_13_begin_0, end = q_13_end_0, end_mask = q_13_end_mask_0, squeeze_mask = q_13_squeeze_mask_0, x = qkv_5_cast_fp16)[name = tensor("q_13_cast_fp16")]; + tensor k_9_begin_0 = const()[name = tensor("k_9_begin_0"), val = tensor([0, 0, 1, 0, 0])]; + tensor k_9_end_0 = const()[name = tensor("k_9_end_0"), val = tensor([1, 1, 2, 16, 64])]; + tensor k_9_end_mask_0 = const()[name = tensor("k_9_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor k_9_squeeze_mask_0 = const()[name = tensor("k_9_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor k_9_cast_fp16 = slice_by_index(begin = k_9_begin_0, end = k_9_end_0, end_mask = k_9_end_mask_0, squeeze_mask = k_9_squeeze_mask_0, x = qkv_5_cast_fp16)[name = tensor("k_9_cast_fp16")]; + tensor v_5_begin_0 = const()[name = tensor("v_5_begin_0"), val = tensor([0, 0, 2, 0, 0])]; + tensor v_5_end_0 = const()[name = tensor("v_5_end_0"), val = tensor([1, 1, 3, 16, 64])]; + tensor v_5_end_mask_0 = const()[name = tensor("v_5_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor v_5_squeeze_mask_0 = const()[name = tensor("v_5_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor v_5_cast_fp16 = slice_by_index(begin = v_5_begin_0, end = v_5_end_0, end_mask = v_5_end_mask_0, squeeze_mask = v_5_squeeze_mask_0, x = qkv_5_cast_fp16)[name = tensor("v_5_cast_fp16")]; + tensor var_1038 = const()[name = tensor("op_1038"), val = tensor([1, 1, 1, 1])]; + tensor position2_to_fp16_dtype_0 = const()[name = tensor("position2_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor position2_to_fp16 = cast(dtype = position2_to_fp16_dtype_0, x = position2)[name = tensor("cast_129")]; + tensor ts_17_cast_fp16 = reshape(shape = var_1038, x = position2_to_fp16)[name = tensor("ts_17_cast_fp16")]; + tensor var_1042 = const()[name = tensor("op_1042"), val = tensor([1, 1, 16, 32, 2])]; + tensor q_complex_5_cast_fp16 = reshape(shape = var_1042, x = q_13_cast_fp16)[name = tensor("q_complex_5_cast_fp16")]; + tensor var_1046 = const()[name = tensor("op_1046"), val = tensor([1, 1, 16, 32, 2])]; + tensor k_complex_5_cast_fp16 = reshape(shape = var_1046, x = k_9_cast_fp16)[name = tensor("k_complex_5_cast_fp16")]; + tensor var_1050_begin_0 = const()[name = tensor("op_1050_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_1050_end_0 = const()[name = tensor("op_1050_end_0"), val = tensor([1, 1, 16, 32, 1])]; + tensor var_1050_end_mask_0 = const()[name = tensor("op_1050_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_1050_squeeze_mask_0 = const()[name = tensor("op_1050_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_1050_cast_fp16 = slice_by_index(begin = var_1050_begin_0, end = var_1050_end_0, end_mask = var_1050_end_mask_0, squeeze_mask = var_1050_squeeze_mask_0, x = q_complex_5_cast_fp16)[name = tensor("op_1050_cast_fp16")]; + tensor var_1058_begin_0 = const()[name = tensor("op_1058_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_1058_end_0 = const()[name = tensor("op_1058_end_0"), val = tensor([1, 1, 16, 32, 2])]; + tensor var_1058_end_mask_0 = const()[name = tensor("op_1058_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_1058_squeeze_mask_0 = const()[name = tensor("op_1058_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_1058_cast_fp16 = slice_by_index(begin = var_1058_begin_0, end = var_1058_end_0, end_mask = var_1058_end_mask_0, squeeze_mask = var_1058_squeeze_mask_0, x = q_complex_5_cast_fp16)[name = tensor("op_1058_cast_fp16")]; + tensor var_1066_begin_0 = const()[name = tensor("op_1066_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_1066_end_0 = const()[name = tensor("op_1066_end_0"), val = tensor([1, 1, 16, 32, 1])]; + tensor var_1066_end_mask_0 = const()[name = tensor("op_1066_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_1066_squeeze_mask_0 = const()[name = tensor("op_1066_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_1066_cast_fp16 = slice_by_index(begin = var_1066_begin_0, end = var_1066_end_0, end_mask = var_1066_end_mask_0, squeeze_mask = var_1066_squeeze_mask_0, x = k_complex_5_cast_fp16)[name = tensor("op_1066_cast_fp16")]; + tensor var_1074_begin_0 = const()[name = tensor("op_1074_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_1074_end_0 = const()[name = tensor("op_1074_end_0"), val = tensor([1, 1, 16, 32, 2])]; + tensor var_1074_end_mask_0 = const()[name = tensor("op_1074_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_1074_squeeze_mask_0 = const()[name = tensor("op_1074_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_1074_cast_fp16 = slice_by_index(begin = var_1074_begin_0, end = var_1074_end_0, end_mask = var_1074_end_mask_0, squeeze_mask = var_1074_squeeze_mask_0, x = k_complex_5_cast_fp16)[name = tensor("op_1074_cast_fp16")]; + tensor freqs_5_to_fp16 = const()[name = tensor("freqs_5_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(76772864)))]; + tensor var_1080_cast_fp16 = mul(x = freqs_5_to_fp16, y = ts_17_cast_fp16)[name = tensor("op_1080_cast_fp16")]; + tensor rotr_5_cast_fp16 = cos(x = var_1080_cast_fp16)[name = tensor("rotr_5_cast_fp16")]; + tensor roti_5_cast_fp16 = sin(x = var_1080_cast_fp16)[name = tensor("roti_5_cast_fp16")]; + tensor var_1084_cast_fp16 = mul(x = var_1050_cast_fp16, y = rotr_5_cast_fp16)[name = tensor("op_1084_cast_fp16")]; + tensor var_1085_cast_fp16 = mul(x = var_1058_cast_fp16, y = roti_5_cast_fp16)[name = tensor("op_1085_cast_fp16")]; + tensor qor_9_cast_fp16 = sub(x = var_1084_cast_fp16, y = var_1085_cast_fp16)[name = tensor("qor_9_cast_fp16")]; + tensor var_1088_cast_fp16 = mul(x = var_1050_cast_fp16, y = roti_5_cast_fp16)[name = tensor("op_1088_cast_fp16")]; + tensor var_1089_cast_fp16 = mul(x = var_1058_cast_fp16, y = rotr_5_cast_fp16)[name = tensor("op_1089_cast_fp16")]; + tensor qoi_9_cast_fp16 = add(x = var_1088_cast_fp16, y = var_1089_cast_fp16)[name = tensor("qoi_9_cast_fp16")]; + tensor var_1092_cast_fp16 = mul(x = var_1066_cast_fp16, y = rotr_5_cast_fp16)[name = tensor("op_1092_cast_fp16")]; + tensor var_1093_cast_fp16 = mul(x = var_1074_cast_fp16, y = roti_5_cast_fp16)[name = tensor("op_1093_cast_fp16")]; + tensor kor_9_cast_fp16 = sub(x = var_1092_cast_fp16, y = var_1093_cast_fp16)[name = tensor("kor_9_cast_fp16")]; + tensor var_1096_cast_fp16 = mul(x = var_1066_cast_fp16, y = roti_5_cast_fp16)[name = tensor("op_1096_cast_fp16")]; + tensor var_1097_cast_fp16 = mul(x = var_1074_cast_fp16, y = rotr_5_cast_fp16)[name = tensor("op_1097_cast_fp16")]; + tensor koi_9_cast_fp16 = add(x = var_1096_cast_fp16, y = var_1097_cast_fp16)[name = tensor("koi_9_cast_fp16")]; + tensor qo_5_axis_0 = const()[name = tensor("qo_5_axis_0"), val = tensor(-1)]; + tensor qo_5_cast_fp16 = stack(axis = qo_5_axis_0, values = (qor_9_cast_fp16, qoi_9_cast_fp16))[name = tensor("qo_5_cast_fp16")]; + tensor ko_5_axis_0 = const()[name = tensor("ko_5_axis_0"), val = tensor(-1)]; + tensor ko_5_cast_fp16 = stack(axis = ko_5_axis_0, values = (kor_9_cast_fp16, koi_9_cast_fp16))[name = tensor("ko_5_cast_fp16")]; + tensor var_1126 = const()[name = tensor("op_1126"), val = tensor([1, 1, 16, 64])]; + tensor q_15_cast_fp16 = reshape(shape = var_1126, x = qo_5_cast_fp16)[name = tensor("q_15_cast_fp16")]; + tensor var_1128 = const()[name = tensor("op_1128"), val = tensor([1, 1, 16, 64])]; + tensor k_11_cast_fp16 = reshape(shape = var_1128, x = ko_5_cast_fp16)[name = tensor("k_11_cast_fp16")]; + tensor _inversed_1150_y_0_to_fp16 = const()[name = tensor("_inversed_1150_y_0_to_fp16"), val = tensor(0x1p-9)]; + tensor _inversed_1150_cast_fp16 = mul(x = ts_17_cast_fp16, y = _inversed_1150_y_0_to_fp16)[name = tensor("_inversed_1150_cast_fp16")]; + tensor var_1151_cast_fp16 = floor(x = _inversed_1150_cast_fp16)[name = tensor("op_1151_cast_fp16")]; + tensor var_1152_to_fp16 = const()[name = tensor("op_1152_to_fp16"), val = tensor(0x1p+9)]; + tensor var_1153_cast_fp16 = mul(x = var_1151_cast_fp16, y = var_1152_to_fp16)[name = tensor("op_1153_cast_fp16")]; + tensor write_indices_float_11_cast_fp16 = sub(x = ts_17_cast_fp16, y = var_1153_cast_fp16)[name = tensor("write_indices_float_11_cast_fp16")]; + tensor var_1160_dtype_0 = const()[name = tensor("op_1160_dtype_0"), val = tensor("int32")]; + tensor write_indices_5_reps_0 = const()[name = tensor("write_indices_5_reps_0"), val = tensor([1, 1, 16, 64])]; + tensor write_indices_float_11_cast_fp16_to_int32 = cast(dtype = var_1160_dtype_0, x = write_indices_float_11_cast_fp16)[name = tensor("cast_128")]; + tensor write_indices_5 = tile(reps = write_indices_5_reps_0, x = write_indices_float_11_cast_fp16_to_int32)[name = tensor("write_indices_5")]; + tensor var_1168_begin_0 = const()[name = tensor("op_1168_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_1168_end_0 = const()[name = tensor("op_1168_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor var_1168_end_mask_0 = const()[name = tensor("op_1168_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_1168_squeeze_mask_0 = const()[name = tensor("op_1168_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor cache2_to_fp16_dtype_0 = const()[name = tensor("cache2_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor cache2_to_fp16 = cast(dtype = cache2_to_fp16_dtype_0, x = cache2)[name = tensor("cast_127")]; + tensor var_1168_cast_fp16 = slice_by_index(begin = var_1168_begin_0, end = var_1168_end_0, end_mask = var_1168_end_mask_0, squeeze_mask = var_1168_squeeze_mask_0, x = cache2_to_fp16)[name = tensor("op_1168_cast_fp16")]; + tensor var_1170_axis_0 = const()[name = tensor("op_1170_axis_0"), val = tensor(1)]; + tensor var_1170_mode_0 = const()[name = tensor("op_1170_mode_0"), val = tensor("update")]; + tensor var_1170_validate_indices_0 = const()[name = tensor("op_1170_validate_indices_0"), val = tensor(false)]; + tensor var_1170_cast_fp16 = scatter_along_axis(axis = var_1170_axis_0, data = var_1168_cast_fp16, indices = write_indices_5, mode = var_1170_mode_0, updates = k_11_cast_fp16, validate_indices = var_1170_validate_indices_0)[name = tensor("op_1170_cast_fp16")]; + tensor concat_16 = const()[name = tensor("concat_16"), val = tensor([0, 0, 0, 0, 0])]; + tensor concat_17 = const()[name = tensor("concat_17"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_5_internal_tensor_assign_1_stride_0 = const()[name = tensor("new_cache_5_internal_tensor_assign_1_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_5_internal_tensor_assign_1_begin_mask_0 = const()[name = tensor("new_cache_5_internal_tensor_assign_1_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_5_internal_tensor_assign_1_end_mask_0 = const()[name = tensor("new_cache_5_internal_tensor_assign_1_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_5_internal_tensor_assign_1_squeeze_mask_0 = const()[name = tensor("new_cache_5_internal_tensor_assign_1_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_16 = const()[name = tensor("shape_16"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_4 = const()[name = tensor("reduce_prod_4"), val = tensor(1048576)]; + tensor range_1d_4_start_0 = const()[name = tensor("range_1d_4_start_0"), val = tensor(0)]; + tensor range_1d_4_step_0 = const()[name = tensor("range_1d_4_step_0"), val = tensor(1)]; + tensor range_1d_4 = range_1d(end = reduce_prod_4, start = range_1d_4_start_0, step = range_1d_4_step_0)[name = tensor("range_1d_4")]; + tensor reshape_20 = reshape(shape = shape_16, x = range_1d_4)[name = tensor("reshape_20")]; + tensor slice_by_index_4 = slice_by_index(begin = concat_16, begin_mask = new_cache_5_internal_tensor_assign_1_begin_mask_0, end = concat_17, end_mask = new_cache_5_internal_tensor_assign_1_end_mask_0, squeeze_mask = new_cache_5_internal_tensor_assign_1_squeeze_mask_0, stride = new_cache_5_internal_tensor_assign_1_stride_0, x = reshape_20)[name = tensor("slice_by_index_4")]; + tensor reshape_21_shape_0 = const()[name = tensor("reshape_21_shape_0"), val = tensor([-1])]; + tensor reshape_21 = reshape(shape = reshape_21_shape_0, x = slice_by_index_4)[name = tensor("reshape_21")]; + tensor reshape_22_shape_0 = const()[name = tensor("reshape_22_shape_0"), val = tensor([-1])]; + tensor reshape_22_cast_fp16 = reshape(shape = reshape_22_shape_0, x = var_1170_cast_fp16)[name = tensor("reshape_22_cast_fp16")]; + tensor reshape_23_shape_0 = const()[name = tensor("reshape_23_shape_0"), val = tensor([-1])]; + tensor reshape_23_cast_fp16 = reshape(shape = reshape_23_shape_0, x = cache2_to_fp16)[name = tensor("reshape_23_cast_fp16")]; + tensor scatter_4_mode_0 = const()[name = tensor("scatter_4_mode_0"), val = tensor("update")]; + tensor scatter_4_axis_0 = const()[name = tensor("scatter_4_axis_0"), val = tensor(0)]; + tensor scatter_4_validate_indices_0 = const()[name = tensor("scatter_4_validate_indices_0"), val = tensor(false)]; + tensor scatter_4_cast_fp16 = scatter(axis = scatter_4_axis_0, data = reshape_23_cast_fp16, indices = reshape_21, mode = scatter_4_mode_0, updates = reshape_22_cast_fp16, validate_indices = scatter_4_validate_indices_0)[name = tensor("scatter_4_cast_fp16")]; + tensor reshape_24_cast_fp16 = reshape(shape = shape_16, x = scatter_4_cast_fp16)[name = tensor("reshape_24_cast_fp16")]; + tensor var_1178_begin_0 = const()[name = tensor("op_1178_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor var_1178_end_0 = const()[name = tensor("op_1178_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor var_1178_end_mask_0 = const()[name = tensor("op_1178_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_1178_squeeze_mask_0 = const()[name = tensor("op_1178_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor var_1178_cast_fp16 = slice_by_index(begin = var_1178_begin_0, end = var_1178_end_0, end_mask = var_1178_end_mask_0, squeeze_mask = var_1178_squeeze_mask_0, x = reshape_24_cast_fp16)[name = tensor("op_1178_cast_fp16")]; + tensor var_1180_axis_0 = const()[name = tensor("op_1180_axis_0"), val = tensor(1)]; + tensor var_1180_mode_0 = const()[name = tensor("op_1180_mode_0"), val = tensor("update")]; + tensor var_1180_validate_indices_0 = const()[name = tensor("op_1180_validate_indices_0"), val = tensor(false)]; + tensor var_1180_cast_fp16 = scatter_along_axis(axis = var_1180_axis_0, data = var_1178_cast_fp16, indices = write_indices_5, mode = var_1180_mode_0, updates = v_5_cast_fp16, validate_indices = var_1180_validate_indices_0)[name = tensor("op_1180_cast_fp16")]; + tensor concat_18 = const()[name = tensor("concat_18"), val = tensor([1, 0, 0, 0, 0])]; + tensor concat_19 = const()[name = tensor("concat_19"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_5_internal_tensor_assign_2_stride_0 = const()[name = tensor("new_cache_5_internal_tensor_assign_2_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_5_internal_tensor_assign_2_begin_mask_0 = const()[name = tensor("new_cache_5_internal_tensor_assign_2_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_5_internal_tensor_assign_2_end_mask_0 = const()[name = tensor("new_cache_5_internal_tensor_assign_2_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_5_internal_tensor_assign_2_squeeze_mask_0 = const()[name = tensor("new_cache_5_internal_tensor_assign_2_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_17 = const()[name = tensor("shape_17"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_5 = const()[name = tensor("reduce_prod_5"), val = tensor(1048576)]; + tensor range_1d_5_start_0 = const()[name = tensor("range_1d_5_start_0"), val = tensor(0)]; + tensor range_1d_5_step_0 = const()[name = tensor("range_1d_5_step_0"), val = tensor(1)]; + tensor range_1d_5 = range_1d(end = reduce_prod_5, start = range_1d_5_start_0, step = range_1d_5_step_0)[name = tensor("range_1d_5")]; + tensor reshape_25 = reshape(shape = shape_17, x = range_1d_5)[name = tensor("reshape_25")]; + tensor slice_by_index_5 = slice_by_index(begin = concat_18, begin_mask = new_cache_5_internal_tensor_assign_2_begin_mask_0, end = concat_19, end_mask = new_cache_5_internal_tensor_assign_2_end_mask_0, squeeze_mask = new_cache_5_internal_tensor_assign_2_squeeze_mask_0, stride = new_cache_5_internal_tensor_assign_2_stride_0, x = reshape_25)[name = tensor("slice_by_index_5")]; + tensor reshape_26_shape_0 = const()[name = tensor("reshape_26_shape_0"), val = tensor([-1])]; + tensor reshape_26 = reshape(shape = reshape_26_shape_0, x = slice_by_index_5)[name = tensor("reshape_26")]; + tensor reshape_27_shape_0 = const()[name = tensor("reshape_27_shape_0"), val = tensor([-1])]; + tensor reshape_27_cast_fp16 = reshape(shape = reshape_27_shape_0, x = var_1180_cast_fp16)[name = tensor("reshape_27_cast_fp16")]; + tensor reshape_28_shape_0 = const()[name = tensor("reshape_28_shape_0"), val = tensor([-1])]; + tensor reshape_28_cast_fp16 = reshape(shape = reshape_28_shape_0, x = reshape_24_cast_fp16)[name = tensor("reshape_28_cast_fp16")]; + tensor scatter_5_mode_0 = const()[name = tensor("scatter_5_mode_0"), val = tensor("update")]; + tensor scatter_5_axis_0 = const()[name = tensor("scatter_5_axis_0"), val = tensor(0)]; + tensor scatter_5_validate_indices_0 = const()[name = tensor("scatter_5_validate_indices_0"), val = tensor(false)]; + tensor scatter_5_cast_fp16 = scatter(axis = scatter_5_axis_0, data = reshape_28_cast_fp16, indices = reshape_26, mode = scatter_5_mode_0, updates = reshape_27_cast_fp16, validate_indices = scatter_5_validate_indices_0)[name = tensor("scatter_5_cast_fp16")]; + tensor reshape_29_cast_fp16 = reshape(shape = shape_17, x = scatter_5_cast_fp16)[name = tensor("reshape_29_cast_fp16")]; + tensor reshape_29_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("reshape_29_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor keys_13_begin_0 = const()[name = tensor("keys_13_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor keys_13_end_0 = const()[name = tensor("keys_13_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor keys_13_end_mask_0 = const()[name = tensor("keys_13_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor keys_13_squeeze_mask_0 = const()[name = tensor("keys_13_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor keys_13_cast_fp16 = slice_by_index(begin = keys_13_begin_0, end = keys_13_end_0, end_mask = keys_13_end_mask_0, squeeze_mask = keys_13_squeeze_mask_0, x = reshape_29_cast_fp16)[name = tensor("keys_13_cast_fp16")]; + tensor values_13_begin_0 = const()[name = tensor("values_13_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor values_13_end_0 = const()[name = tensor("values_13_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor values_13_end_mask_0 = const()[name = tensor("values_13_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor values_13_squeeze_mask_0 = const()[name = tensor("values_13_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor values_13_cast_fp16 = slice_by_index(begin = values_13_begin_0, end = values_13_end_0, end_mask = values_13_end_mask_0, squeeze_mask = values_13_squeeze_mask_0, x = reshape_29_cast_fp16)[name = tensor("values_13_cast_fp16")]; + tensor var_1192_cast_fp16 = not_equal(x = keys_13_cast_fp16, y = keys_13_cast_fp16)[name = tensor("op_1192_cast_fp16")]; + tensor keys_15_cast_fp16 = select(a = var_360_to_fp16, b = keys_13_cast_fp16, cond = var_1192_cast_fp16)[name = tensor("keys_15_cast_fp16")]; + tensor var_1200_cast_fp16 = not_equal(x = values_13_cast_fp16, y = values_13_cast_fp16)[name = tensor("op_1200_cast_fp16")]; + tensor values_15_cast_fp16 = select(a = var_360_to_fp16, b = values_13_cast_fp16, cond = var_1200_cast_fp16)[name = tensor("values_15_cast_fp16")]; + tensor var_1224 = const()[name = tensor("op_1224"), val = tensor([0, 2, 1, 3])]; + tensor var_1237 = const()[name = tensor("op_1237"), val = tensor([1, 1, 1])]; + tensor var_1238_cast_fp16 = reshape(shape = var_1237, x = position2_to_fp16)[name = tensor("op_1238_cast_fp16")]; + tensor var_1255_to_fp16 = const()[name = tensor("op_1255_to_fp16"), val = tensor(0x1p+0)]; + tensor valid_len_5_cast_fp16 = add(x = var_1238_cast_fp16, y = var_1255_to_fp16)[name = tensor("valid_len_5_cast_fp16")]; + tensor valid_mask_5_cast_fp16 = less(x = k_positions_1_promoted_to_fp16, y = valid_len_5_cast_fp16)[name = tensor("valid_mask_5_cast_fp16")]; + tensor causal_mask_5_cast_fp16 = less_equal(x = k_positions_1_promoted_to_fp16, y = var_1238_cast_fp16)[name = tensor("causal_mask_5_cast_fp16")]; + tensor attn_mask_9 = logical_and(x = valid_mask_5_cast_fp16, y = causal_mask_5_cast_fp16)[name = tensor("attn_mask_9")]; + tensor attn_mask_11_axes_0 = const()[name = tensor("attn_mask_11_axes_0"), val = tensor([1])]; + tensor attn_mask_11 = expand_dims(axes = attn_mask_11_axes_0, x = attn_mask_9)[name = tensor("attn_mask_11")]; + tensor var_1273_transpose_x_0 = const()[name = tensor("op_1273_transpose_x_0"), val = tensor(false)]; + tensor var_1273_transpose_y_0 = const()[name = tensor("op_1273_transpose_y_0"), val = tensor(false)]; + tensor transpose_22_perm_0 = const()[name = tensor("transpose_22_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_23_perm_0 = const()[name = tensor("transpose_23_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_23 = transpose(perm = transpose_23_perm_0, x = keys_15_cast_fp16)[name = tensor("transpose_43")]; + tensor transpose_22 = transpose(perm = transpose_22_perm_0, x = q_15_cast_fp16)[name = tensor("transpose_44")]; + tensor var_1273_cast_fp16 = matmul(transpose_x = var_1273_transpose_x_0, transpose_y = var_1273_transpose_y_0, x = transpose_22, y = transpose_23)[name = tensor("op_1273_cast_fp16")]; + tensor var_1267_to_fp16 = const()[name = tensor("op_1267_to_fp16"), val = tensor([0x1p-3])]; + tensor attn_weights_13_cast_fp16 = mul(x = var_1273_cast_fp16, y = var_1267_to_fp16)[name = tensor("attn_weights_13_cast_fp16")]; + tensor var_1275 = logical_not(x = attn_mask_11)[name = tensor("op_1275")]; + tensor var_1276_to_fp16 = const()[name = tensor("op_1276_to_fp16"), val = tensor(-inf)]; + tensor attn_weights_15_cast_fp16 = select(a = var_1276_to_fp16, b = attn_weights_13_cast_fp16, cond = var_1275)[name = tensor("attn_weights_15_cast_fp16")]; + tensor var_1278 = const()[name = tensor("op_1278"), val = tensor(-1)]; + tensor attn_weights_17_cast_fp16 = softmax(axis = var_1278, x = attn_weights_15_cast_fp16)[name = tensor("attn_weights_17_cast_fp16")]; + tensor attn_output_5_transpose_x_0 = const()[name = tensor("attn_output_5_transpose_x_0"), val = tensor(false)]; + tensor attn_output_5_transpose_y_0 = const()[name = tensor("attn_output_5_transpose_y_0"), val = tensor(false)]; + tensor values_17_cast_fp16 = transpose(perm = var_1224, x = values_15_cast_fp16)[name = tensor("transpose_45")]; + tensor attn_output_5_cast_fp16 = matmul(transpose_x = attn_output_5_transpose_x_0, transpose_y = attn_output_5_transpose_y_0, x = attn_weights_17_cast_fp16, y = values_17_cast_fp16)[name = tensor("attn_output_5_cast_fp16")]; + tensor var_1286 = const()[name = tensor("op_1286"), val = tensor([0, 2, 1, 3])]; + tensor var_1289 = const()[name = tensor("op_1289"), val = tensor([1, 1, 1024])]; + tensor var_1287_cast_fp16 = transpose(perm = var_1286, x = attn_output_5_cast_fp16)[name = tensor("transpose_42")]; + tensor input_25_cast_fp16 = reshape(shape = var_1289, x = var_1287_cast_fp16)[name = tensor("input_25_cast_fp16")]; + tensor linear_10_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = attn2_out_proj_weight_cast_fp16, x = input_25_cast_fp16)[name = tensor("linear_10_cast_fp16")]; + tensor var_1295_to_fp16 = const()[name = tensor("op_1295_to_fp16"), val = tensor(0x1p+0)]; + tensor var_1296_cast_fp16 = add(x = position2_to_fp16, y = var_1295_to_fp16)[name = tensor("op_1296_cast_fp16")]; + tensor var_1296_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_1296_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor input_27_cast_fp16 = add(x = input_23_cast_fp16, y = linear_10_cast_fp16)[name = tensor("input_27_cast_fp16")]; + tensor input_29_axes_0 = const()[name = tensor("input_29_axes_0"), val = tensor([-1])]; + tensor norm2_2_weight_to_fp16 = const()[name = tensor("norm2_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(76772992)))]; + tensor norm2_2_bias_to_fp16 = const()[name = tensor("norm2_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(76775104)))]; + tensor var_1300_to_fp16 = const()[name = tensor("op_1300_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_29_cast_fp16 = layer_norm(axes = input_29_axes_0, beta = norm2_2_bias_to_fp16, epsilon = var_1300_to_fp16, gamma = norm2_2_weight_to_fp16, x = input_27_cast_fp16)[name = tensor("input_29_cast_fp16")]; + tensor linear_11_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = linear2_1_weight_cast_fp16, x = input_29_cast_fp16)[name = tensor("linear_11_cast_fp16")]; + tensor input_31_mode_0 = const()[name = tensor("input_31_mode_0"), val = tensor("EXACT")]; + tensor input_31_cast_fp16 = gelu(mode = input_31_mode_0, x = linear_11_cast_fp16)[name = tensor("input_31_cast_fp16")]; + tensor linear_12_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = linear2_2_weight_cast_fp16, x = input_31_cast_fp16)[name = tensor("linear_12_cast_fp16")]; + tensor input_33_cast_fp16 = add(x = input_27_cast_fp16, y = linear_12_cast_fp16)[name = tensor("input_33_cast_fp16")]; + tensor x_7_axes_0 = const()[name = tensor("x_7_axes_0"), val = tensor([-1])]; + tensor norm3_1_weight_to_fp16 = const()[name = tensor("norm3_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(76777216)))]; + tensor norm3_1_bias_to_fp16 = const()[name = tensor("norm3_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(76779328)))]; + tensor var_1317_to_fp16 = const()[name = tensor("op_1317_to_fp16"), val = tensor(0x1.5p-17)]; + tensor x_7_cast_fp16 = layer_norm(axes = x_7_axes_0, beta = norm3_1_bias_to_fp16, epsilon = var_1317_to_fp16, gamma = norm3_1_weight_to_fp16, x = input_33_cast_fp16)[name = tensor("x_7_cast_fp16")]; + tensor linear_13_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = attn3_in_proj_weight_cast_fp16, x = x_7_cast_fp16)[name = tensor("linear_13_cast_fp16")]; + tensor var_1353 = const()[name = tensor("op_1353"), val = tensor([1, 1, 3, 16, 64])]; + tensor qkv_7_cast_fp16 = reshape(shape = var_1353, x = linear_13_cast_fp16)[name = tensor("qkv_7_cast_fp16")]; + tensor q_19_begin_0 = const()[name = tensor("q_19_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor q_19_end_0 = const()[name = tensor("q_19_end_0"), val = tensor([1, 1, 1, 16, 64])]; + tensor q_19_end_mask_0 = const()[name = tensor("q_19_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor q_19_squeeze_mask_0 = const()[name = tensor("q_19_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor q_19_cast_fp16 = slice_by_index(begin = q_19_begin_0, end = q_19_end_0, end_mask = q_19_end_mask_0, squeeze_mask = q_19_squeeze_mask_0, x = qkv_7_cast_fp16)[name = tensor("q_19_cast_fp16")]; + tensor k_13_begin_0 = const()[name = tensor("k_13_begin_0"), val = tensor([0, 0, 1, 0, 0])]; + tensor k_13_end_0 = const()[name = tensor("k_13_end_0"), val = tensor([1, 1, 2, 16, 64])]; + tensor k_13_end_mask_0 = const()[name = tensor("k_13_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor k_13_squeeze_mask_0 = const()[name = tensor("k_13_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor k_13_cast_fp16 = slice_by_index(begin = k_13_begin_0, end = k_13_end_0, end_mask = k_13_end_mask_0, squeeze_mask = k_13_squeeze_mask_0, x = qkv_7_cast_fp16)[name = tensor("k_13_cast_fp16")]; + tensor v_7_begin_0 = const()[name = tensor("v_7_begin_0"), val = tensor([0, 0, 2, 0, 0])]; + tensor v_7_end_0 = const()[name = tensor("v_7_end_0"), val = tensor([1, 1, 3, 16, 64])]; + tensor v_7_end_mask_0 = const()[name = tensor("v_7_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor v_7_squeeze_mask_0 = const()[name = tensor("v_7_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor v_7_cast_fp16 = slice_by_index(begin = v_7_begin_0, end = v_7_end_0, end_mask = v_7_end_mask_0, squeeze_mask = v_7_squeeze_mask_0, x = qkv_7_cast_fp16)[name = tensor("v_7_cast_fp16")]; + tensor var_1457 = const()[name = tensor("op_1457"), val = tensor([1, 1, 1, 1])]; + tensor position3_to_fp16_dtype_0 = const()[name = tensor("position3_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor position3_to_fp16 = cast(dtype = position3_to_fp16_dtype_0, x = position3)[name = tensor("cast_124")]; + tensor ts_23_cast_fp16 = reshape(shape = var_1457, x = position3_to_fp16)[name = tensor("ts_23_cast_fp16")]; + tensor var_1461 = const()[name = tensor("op_1461"), val = tensor([1, 1, 16, 32, 2])]; + tensor q_complex_7_cast_fp16 = reshape(shape = var_1461, x = q_19_cast_fp16)[name = tensor("q_complex_7_cast_fp16")]; + tensor var_1465 = const()[name = tensor("op_1465"), val = tensor([1, 1, 16, 32, 2])]; + tensor k_complex_7_cast_fp16 = reshape(shape = var_1465, x = k_13_cast_fp16)[name = tensor("k_complex_7_cast_fp16")]; + tensor var_1469_begin_0 = const()[name = tensor("op_1469_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_1469_end_0 = const()[name = tensor("op_1469_end_0"), val = tensor([1, 1, 16, 32, 1])]; + tensor var_1469_end_mask_0 = const()[name = tensor("op_1469_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_1469_squeeze_mask_0 = const()[name = tensor("op_1469_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_1469_cast_fp16 = slice_by_index(begin = var_1469_begin_0, end = var_1469_end_0, end_mask = var_1469_end_mask_0, squeeze_mask = var_1469_squeeze_mask_0, x = q_complex_7_cast_fp16)[name = tensor("op_1469_cast_fp16")]; + tensor var_1477_begin_0 = const()[name = tensor("op_1477_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_1477_end_0 = const()[name = tensor("op_1477_end_0"), val = tensor([1, 1, 16, 32, 2])]; + tensor var_1477_end_mask_0 = const()[name = tensor("op_1477_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_1477_squeeze_mask_0 = const()[name = tensor("op_1477_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_1477_cast_fp16 = slice_by_index(begin = var_1477_begin_0, end = var_1477_end_0, end_mask = var_1477_end_mask_0, squeeze_mask = var_1477_squeeze_mask_0, x = q_complex_7_cast_fp16)[name = tensor("op_1477_cast_fp16")]; + tensor var_1485_begin_0 = const()[name = tensor("op_1485_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_1485_end_0 = const()[name = tensor("op_1485_end_0"), val = tensor([1, 1, 16, 32, 1])]; + tensor var_1485_end_mask_0 = const()[name = tensor("op_1485_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_1485_squeeze_mask_0 = const()[name = tensor("op_1485_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_1485_cast_fp16 = slice_by_index(begin = var_1485_begin_0, end = var_1485_end_0, end_mask = var_1485_end_mask_0, squeeze_mask = var_1485_squeeze_mask_0, x = k_complex_7_cast_fp16)[name = tensor("op_1485_cast_fp16")]; + tensor var_1493_begin_0 = const()[name = tensor("op_1493_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_1493_end_0 = const()[name = tensor("op_1493_end_0"), val = tensor([1, 1, 16, 32, 2])]; + tensor var_1493_end_mask_0 = const()[name = tensor("op_1493_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_1493_squeeze_mask_0 = const()[name = tensor("op_1493_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_1493_cast_fp16 = slice_by_index(begin = var_1493_begin_0, end = var_1493_end_0, end_mask = var_1493_end_mask_0, squeeze_mask = var_1493_squeeze_mask_0, x = k_complex_7_cast_fp16)[name = tensor("op_1493_cast_fp16")]; + tensor freqs_7_to_fp16 = const()[name = tensor("freqs_7_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(76781440)))]; + tensor var_1499_cast_fp16 = mul(x = freqs_7_to_fp16, y = ts_23_cast_fp16)[name = tensor("op_1499_cast_fp16")]; + tensor rotr_7_cast_fp16 = cos(x = var_1499_cast_fp16)[name = tensor("rotr_7_cast_fp16")]; + tensor roti_7_cast_fp16 = sin(x = var_1499_cast_fp16)[name = tensor("roti_7_cast_fp16")]; + tensor var_1503_cast_fp16 = mul(x = var_1469_cast_fp16, y = rotr_7_cast_fp16)[name = tensor("op_1503_cast_fp16")]; + tensor var_1504_cast_fp16 = mul(x = var_1477_cast_fp16, y = roti_7_cast_fp16)[name = tensor("op_1504_cast_fp16")]; + tensor qor_13_cast_fp16 = sub(x = var_1503_cast_fp16, y = var_1504_cast_fp16)[name = tensor("qor_13_cast_fp16")]; + tensor var_1507_cast_fp16 = mul(x = var_1469_cast_fp16, y = roti_7_cast_fp16)[name = tensor("op_1507_cast_fp16")]; + tensor var_1508_cast_fp16 = mul(x = var_1477_cast_fp16, y = rotr_7_cast_fp16)[name = tensor("op_1508_cast_fp16")]; + tensor qoi_13_cast_fp16 = add(x = var_1507_cast_fp16, y = var_1508_cast_fp16)[name = tensor("qoi_13_cast_fp16")]; + tensor var_1511_cast_fp16 = mul(x = var_1485_cast_fp16, y = rotr_7_cast_fp16)[name = tensor("op_1511_cast_fp16")]; + tensor var_1512_cast_fp16 = mul(x = var_1493_cast_fp16, y = roti_7_cast_fp16)[name = tensor("op_1512_cast_fp16")]; + tensor kor_13_cast_fp16 = sub(x = var_1511_cast_fp16, y = var_1512_cast_fp16)[name = tensor("kor_13_cast_fp16")]; + tensor var_1515_cast_fp16 = mul(x = var_1485_cast_fp16, y = roti_7_cast_fp16)[name = tensor("op_1515_cast_fp16")]; + tensor var_1516_cast_fp16 = mul(x = var_1493_cast_fp16, y = rotr_7_cast_fp16)[name = tensor("op_1516_cast_fp16")]; + tensor koi_13_cast_fp16 = add(x = var_1515_cast_fp16, y = var_1516_cast_fp16)[name = tensor("koi_13_cast_fp16")]; + tensor qo_7_axis_0 = const()[name = tensor("qo_7_axis_0"), val = tensor(-1)]; + tensor qo_7_cast_fp16 = stack(axis = qo_7_axis_0, values = (qor_13_cast_fp16, qoi_13_cast_fp16))[name = tensor("qo_7_cast_fp16")]; + tensor ko_7_axis_0 = const()[name = tensor("ko_7_axis_0"), val = tensor(-1)]; + tensor ko_7_cast_fp16 = stack(axis = ko_7_axis_0, values = (kor_13_cast_fp16, koi_13_cast_fp16))[name = tensor("ko_7_cast_fp16")]; + tensor var_1545 = const()[name = tensor("op_1545"), val = tensor([1, 1, 16, 64])]; + tensor q_21_cast_fp16 = reshape(shape = var_1545, x = qo_7_cast_fp16)[name = tensor("q_21_cast_fp16")]; + tensor var_1547 = const()[name = tensor("op_1547"), val = tensor([1, 1, 16, 64])]; + tensor k_15_cast_fp16 = reshape(shape = var_1547, x = ko_7_cast_fp16)[name = tensor("k_15_cast_fp16")]; + tensor _inversed_1569_y_0_to_fp16 = const()[name = tensor("_inversed_1569_y_0_to_fp16"), val = tensor(0x1p-9)]; + tensor _inversed_1569_cast_fp16 = mul(x = ts_23_cast_fp16, y = _inversed_1569_y_0_to_fp16)[name = tensor("_inversed_1569_cast_fp16")]; + tensor var_1570_cast_fp16 = floor(x = _inversed_1569_cast_fp16)[name = tensor("op_1570_cast_fp16")]; + tensor var_1571_to_fp16 = const()[name = tensor("op_1571_to_fp16"), val = tensor(0x1p+9)]; + tensor var_1572_cast_fp16 = mul(x = var_1570_cast_fp16, y = var_1571_to_fp16)[name = tensor("op_1572_cast_fp16")]; + tensor write_indices_float_15_cast_fp16 = sub(x = ts_23_cast_fp16, y = var_1572_cast_fp16)[name = tensor("write_indices_float_15_cast_fp16")]; + tensor var_1579_dtype_0 = const()[name = tensor("op_1579_dtype_0"), val = tensor("int32")]; + tensor write_indices_7_reps_0 = const()[name = tensor("write_indices_7_reps_0"), val = tensor([1, 1, 16, 64])]; + tensor write_indices_float_15_cast_fp16_to_int32 = cast(dtype = var_1579_dtype_0, x = write_indices_float_15_cast_fp16)[name = tensor("cast_123")]; + tensor write_indices_7 = tile(reps = write_indices_7_reps_0, x = write_indices_float_15_cast_fp16_to_int32)[name = tensor("write_indices_7")]; + tensor var_1587_begin_0 = const()[name = tensor("op_1587_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_1587_end_0 = const()[name = tensor("op_1587_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor var_1587_end_mask_0 = const()[name = tensor("op_1587_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_1587_squeeze_mask_0 = const()[name = tensor("op_1587_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor cache3_to_fp16_dtype_0 = const()[name = tensor("cache3_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor cache3_to_fp16 = cast(dtype = cache3_to_fp16_dtype_0, x = cache3)[name = tensor("cast_122")]; + tensor var_1587_cast_fp16 = slice_by_index(begin = var_1587_begin_0, end = var_1587_end_0, end_mask = var_1587_end_mask_0, squeeze_mask = var_1587_squeeze_mask_0, x = cache3_to_fp16)[name = tensor("op_1587_cast_fp16")]; + tensor var_1589_axis_0 = const()[name = tensor("op_1589_axis_0"), val = tensor(1)]; + tensor var_1589_mode_0 = const()[name = tensor("op_1589_mode_0"), val = tensor("update")]; + tensor var_1589_validate_indices_0 = const()[name = tensor("op_1589_validate_indices_0"), val = tensor(false)]; + tensor var_1589_cast_fp16 = scatter_along_axis(axis = var_1589_axis_0, data = var_1587_cast_fp16, indices = write_indices_7, mode = var_1589_mode_0, updates = k_15_cast_fp16, validate_indices = var_1589_validate_indices_0)[name = tensor("op_1589_cast_fp16")]; + tensor concat_23 = const()[name = tensor("concat_23"), val = tensor([0, 0, 0, 0, 0])]; + tensor concat_24 = const()[name = tensor("concat_24"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_7_internal_tensor_assign_1_stride_0 = const()[name = tensor("new_cache_7_internal_tensor_assign_1_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_7_internal_tensor_assign_1_begin_mask_0 = const()[name = tensor("new_cache_7_internal_tensor_assign_1_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_7_internal_tensor_assign_1_end_mask_0 = const()[name = tensor("new_cache_7_internal_tensor_assign_1_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_7_internal_tensor_assign_1_squeeze_mask_0 = const()[name = tensor("new_cache_7_internal_tensor_assign_1_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_18 = const()[name = tensor("shape_18"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_6 = const()[name = tensor("reduce_prod_6"), val = tensor(1048576)]; + tensor range_1d_6_start_0 = const()[name = tensor("range_1d_6_start_0"), val = tensor(0)]; + tensor range_1d_6_step_0 = const()[name = tensor("range_1d_6_step_0"), val = tensor(1)]; + tensor range_1d_6 = range_1d(end = reduce_prod_6, start = range_1d_6_start_0, step = range_1d_6_step_0)[name = tensor("range_1d_6")]; + tensor reshape_30 = reshape(shape = shape_18, x = range_1d_6)[name = tensor("reshape_30")]; + tensor slice_by_index_6 = slice_by_index(begin = concat_23, begin_mask = new_cache_7_internal_tensor_assign_1_begin_mask_0, end = concat_24, end_mask = new_cache_7_internal_tensor_assign_1_end_mask_0, squeeze_mask = new_cache_7_internal_tensor_assign_1_squeeze_mask_0, stride = new_cache_7_internal_tensor_assign_1_stride_0, x = reshape_30)[name = tensor("slice_by_index_6")]; + tensor reshape_31_shape_0 = const()[name = tensor("reshape_31_shape_0"), val = tensor([-1])]; + tensor reshape_31 = reshape(shape = reshape_31_shape_0, x = slice_by_index_6)[name = tensor("reshape_31")]; + tensor reshape_32_shape_0 = const()[name = tensor("reshape_32_shape_0"), val = tensor([-1])]; + tensor reshape_32_cast_fp16 = reshape(shape = reshape_32_shape_0, x = var_1589_cast_fp16)[name = tensor("reshape_32_cast_fp16")]; + tensor reshape_33_shape_0 = const()[name = tensor("reshape_33_shape_0"), val = tensor([-1])]; + tensor reshape_33_cast_fp16 = reshape(shape = reshape_33_shape_0, x = cache3_to_fp16)[name = tensor("reshape_33_cast_fp16")]; + tensor scatter_6_mode_0 = const()[name = tensor("scatter_6_mode_0"), val = tensor("update")]; + tensor scatter_6_axis_0 = const()[name = tensor("scatter_6_axis_0"), val = tensor(0)]; + tensor scatter_6_validate_indices_0 = const()[name = tensor("scatter_6_validate_indices_0"), val = tensor(false)]; + tensor scatter_6_cast_fp16 = scatter(axis = scatter_6_axis_0, data = reshape_33_cast_fp16, indices = reshape_31, mode = scatter_6_mode_0, updates = reshape_32_cast_fp16, validate_indices = scatter_6_validate_indices_0)[name = tensor("scatter_6_cast_fp16")]; + tensor reshape_34_cast_fp16 = reshape(shape = shape_18, x = scatter_6_cast_fp16)[name = tensor("reshape_34_cast_fp16")]; + tensor var_1597_begin_0 = const()[name = tensor("op_1597_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor var_1597_end_0 = const()[name = tensor("op_1597_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor var_1597_end_mask_0 = const()[name = tensor("op_1597_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_1597_squeeze_mask_0 = const()[name = tensor("op_1597_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor var_1597_cast_fp16 = slice_by_index(begin = var_1597_begin_0, end = var_1597_end_0, end_mask = var_1597_end_mask_0, squeeze_mask = var_1597_squeeze_mask_0, x = reshape_34_cast_fp16)[name = tensor("op_1597_cast_fp16")]; + tensor var_1599_axis_0 = const()[name = tensor("op_1599_axis_0"), val = tensor(1)]; + tensor var_1599_mode_0 = const()[name = tensor("op_1599_mode_0"), val = tensor("update")]; + tensor var_1599_validate_indices_0 = const()[name = tensor("op_1599_validate_indices_0"), val = tensor(false)]; + tensor var_1599_cast_fp16 = scatter_along_axis(axis = var_1599_axis_0, data = var_1597_cast_fp16, indices = write_indices_7, mode = var_1599_mode_0, updates = v_7_cast_fp16, validate_indices = var_1599_validate_indices_0)[name = tensor("op_1599_cast_fp16")]; + tensor concat_25 = const()[name = tensor("concat_25"), val = tensor([1, 0, 0, 0, 0])]; + tensor concat_26 = const()[name = tensor("concat_26"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_7_internal_tensor_assign_2_stride_0 = const()[name = tensor("new_cache_7_internal_tensor_assign_2_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_7_internal_tensor_assign_2_begin_mask_0 = const()[name = tensor("new_cache_7_internal_tensor_assign_2_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_7_internal_tensor_assign_2_end_mask_0 = const()[name = tensor("new_cache_7_internal_tensor_assign_2_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_7_internal_tensor_assign_2_squeeze_mask_0 = const()[name = tensor("new_cache_7_internal_tensor_assign_2_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_19 = const()[name = tensor("shape_19"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_7 = const()[name = tensor("reduce_prod_7"), val = tensor(1048576)]; + tensor range_1d_7_start_0 = const()[name = tensor("range_1d_7_start_0"), val = tensor(0)]; + tensor range_1d_7_step_0 = const()[name = tensor("range_1d_7_step_0"), val = tensor(1)]; + tensor range_1d_7 = range_1d(end = reduce_prod_7, start = range_1d_7_start_0, step = range_1d_7_step_0)[name = tensor("range_1d_7")]; + tensor reshape_35 = reshape(shape = shape_19, x = range_1d_7)[name = tensor("reshape_35")]; + tensor slice_by_index_7 = slice_by_index(begin = concat_25, begin_mask = new_cache_7_internal_tensor_assign_2_begin_mask_0, end = concat_26, end_mask = new_cache_7_internal_tensor_assign_2_end_mask_0, squeeze_mask = new_cache_7_internal_tensor_assign_2_squeeze_mask_0, stride = new_cache_7_internal_tensor_assign_2_stride_0, x = reshape_35)[name = tensor("slice_by_index_7")]; + tensor reshape_36_shape_0 = const()[name = tensor("reshape_36_shape_0"), val = tensor([-1])]; + tensor reshape_36 = reshape(shape = reshape_36_shape_0, x = slice_by_index_7)[name = tensor("reshape_36")]; + tensor reshape_37_shape_0 = const()[name = tensor("reshape_37_shape_0"), val = tensor([-1])]; + tensor reshape_37_cast_fp16 = reshape(shape = reshape_37_shape_0, x = var_1599_cast_fp16)[name = tensor("reshape_37_cast_fp16")]; + tensor reshape_38_shape_0 = const()[name = tensor("reshape_38_shape_0"), val = tensor([-1])]; + tensor reshape_38_cast_fp16 = reshape(shape = reshape_38_shape_0, x = reshape_34_cast_fp16)[name = tensor("reshape_38_cast_fp16")]; + tensor scatter_7_mode_0 = const()[name = tensor("scatter_7_mode_0"), val = tensor("update")]; + tensor scatter_7_axis_0 = const()[name = tensor("scatter_7_axis_0"), val = tensor(0)]; + tensor scatter_7_validate_indices_0 = const()[name = tensor("scatter_7_validate_indices_0"), val = tensor(false)]; + tensor scatter_7_cast_fp16 = scatter(axis = scatter_7_axis_0, data = reshape_38_cast_fp16, indices = reshape_36, mode = scatter_7_mode_0, updates = reshape_37_cast_fp16, validate_indices = scatter_7_validate_indices_0)[name = tensor("scatter_7_cast_fp16")]; + tensor reshape_39_cast_fp16 = reshape(shape = shape_19, x = scatter_7_cast_fp16)[name = tensor("reshape_39_cast_fp16")]; + tensor reshape_39_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("reshape_39_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor keys_19_begin_0 = const()[name = tensor("keys_19_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor keys_19_end_0 = const()[name = tensor("keys_19_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor keys_19_end_mask_0 = const()[name = tensor("keys_19_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor keys_19_squeeze_mask_0 = const()[name = tensor("keys_19_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor keys_19_cast_fp16 = slice_by_index(begin = keys_19_begin_0, end = keys_19_end_0, end_mask = keys_19_end_mask_0, squeeze_mask = keys_19_squeeze_mask_0, x = reshape_39_cast_fp16)[name = tensor("keys_19_cast_fp16")]; + tensor values_19_begin_0 = const()[name = tensor("values_19_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor values_19_end_0 = const()[name = tensor("values_19_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor values_19_end_mask_0 = const()[name = tensor("values_19_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor values_19_squeeze_mask_0 = const()[name = tensor("values_19_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor values_19_cast_fp16 = slice_by_index(begin = values_19_begin_0, end = values_19_end_0, end_mask = values_19_end_mask_0, squeeze_mask = values_19_squeeze_mask_0, x = reshape_39_cast_fp16)[name = tensor("values_19_cast_fp16")]; + tensor var_1611_cast_fp16 = not_equal(x = keys_19_cast_fp16, y = keys_19_cast_fp16)[name = tensor("op_1611_cast_fp16")]; + tensor keys_21_cast_fp16 = select(a = var_360_to_fp16, b = keys_19_cast_fp16, cond = var_1611_cast_fp16)[name = tensor("keys_21_cast_fp16")]; + tensor var_1619_cast_fp16 = not_equal(x = values_19_cast_fp16, y = values_19_cast_fp16)[name = tensor("op_1619_cast_fp16")]; + tensor values_21_cast_fp16 = select(a = var_360_to_fp16, b = values_19_cast_fp16, cond = var_1619_cast_fp16)[name = tensor("values_21_cast_fp16")]; + tensor var_1643 = const()[name = tensor("op_1643"), val = tensor([0, 2, 1, 3])]; + tensor var_1656 = const()[name = tensor("op_1656"), val = tensor([1, 1, 1])]; + tensor var_1657_cast_fp16 = reshape(shape = var_1656, x = position3_to_fp16)[name = tensor("op_1657_cast_fp16")]; + tensor var_1674_to_fp16 = const()[name = tensor("op_1674_to_fp16"), val = tensor(0x1p+0)]; + tensor valid_len_7_cast_fp16 = add(x = var_1657_cast_fp16, y = var_1674_to_fp16)[name = tensor("valid_len_7_cast_fp16")]; + tensor valid_mask_7_cast_fp16 = less(x = k_positions_1_promoted_to_fp16, y = valid_len_7_cast_fp16)[name = tensor("valid_mask_7_cast_fp16")]; + tensor causal_mask_7_cast_fp16 = less_equal(x = k_positions_1_promoted_to_fp16, y = var_1657_cast_fp16)[name = tensor("causal_mask_7_cast_fp16")]; + tensor attn_mask_13 = logical_and(x = valid_mask_7_cast_fp16, y = causal_mask_7_cast_fp16)[name = tensor("attn_mask_13")]; + tensor attn_mask_15_axes_0 = const()[name = tensor("attn_mask_15_axes_0"), val = tensor([1])]; + tensor attn_mask_15 = expand_dims(axes = attn_mask_15_axes_0, x = attn_mask_13)[name = tensor("attn_mask_15")]; + tensor var_1692_transpose_x_0 = const()[name = tensor("op_1692_transpose_x_0"), val = tensor(false)]; + tensor var_1692_transpose_y_0 = const()[name = tensor("op_1692_transpose_y_0"), val = tensor(false)]; + tensor transpose_24_perm_0 = const()[name = tensor("transpose_24_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_25_perm_0 = const()[name = tensor("transpose_25_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_25 = transpose(perm = transpose_25_perm_0, x = keys_21_cast_fp16)[name = tensor("transpose_39")]; + tensor transpose_24 = transpose(perm = transpose_24_perm_0, x = q_21_cast_fp16)[name = tensor("transpose_40")]; + tensor var_1692_cast_fp16 = matmul(transpose_x = var_1692_transpose_x_0, transpose_y = var_1692_transpose_y_0, x = transpose_24, y = transpose_25)[name = tensor("op_1692_cast_fp16")]; + tensor var_1686_to_fp16 = const()[name = tensor("op_1686_to_fp16"), val = tensor([0x1p-3])]; + tensor attn_weights_19_cast_fp16 = mul(x = var_1692_cast_fp16, y = var_1686_to_fp16)[name = tensor("attn_weights_19_cast_fp16")]; + tensor var_1694 = logical_not(x = attn_mask_15)[name = tensor("op_1694")]; + tensor var_1695_to_fp16 = const()[name = tensor("op_1695_to_fp16"), val = tensor(-inf)]; + tensor attn_weights_21_cast_fp16 = select(a = var_1695_to_fp16, b = attn_weights_19_cast_fp16, cond = var_1694)[name = tensor("attn_weights_21_cast_fp16")]; + tensor var_1697 = const()[name = tensor("op_1697"), val = tensor(-1)]; + tensor attn_weights_23_cast_fp16 = softmax(axis = var_1697, x = attn_weights_21_cast_fp16)[name = tensor("attn_weights_23_cast_fp16")]; + tensor attn_output_7_transpose_x_0 = const()[name = tensor("attn_output_7_transpose_x_0"), val = tensor(false)]; + tensor attn_output_7_transpose_y_0 = const()[name = tensor("attn_output_7_transpose_y_0"), val = tensor(false)]; + tensor values_23_cast_fp16 = transpose(perm = var_1643, x = values_21_cast_fp16)[name = tensor("transpose_41")]; + tensor attn_output_7_cast_fp16 = matmul(transpose_x = attn_output_7_transpose_x_0, transpose_y = attn_output_7_transpose_y_0, x = attn_weights_23_cast_fp16, y = values_23_cast_fp16)[name = tensor("attn_output_7_cast_fp16")]; + tensor var_1705 = const()[name = tensor("op_1705"), val = tensor([0, 2, 1, 3])]; + tensor var_1708 = const()[name = tensor("op_1708"), val = tensor([1, 1, 1024])]; + tensor var_1706_cast_fp16 = transpose(perm = var_1705, x = attn_output_7_cast_fp16)[name = tensor("transpose_38")]; + tensor input_35_cast_fp16 = reshape(shape = var_1708, x = var_1706_cast_fp16)[name = tensor("input_35_cast_fp16")]; + tensor linear_14_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = attn3_out_proj_weight_cast_fp16, x = input_35_cast_fp16)[name = tensor("linear_14_cast_fp16")]; + tensor var_1714_to_fp16 = const()[name = tensor("op_1714_to_fp16"), val = tensor(0x1p+0)]; + tensor var_1715_cast_fp16 = add(x = position3_to_fp16, y = var_1714_to_fp16)[name = tensor("op_1715_cast_fp16")]; + tensor var_1715_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_1715_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor input_37_cast_fp16 = add(x = input_33_cast_fp16, y = linear_14_cast_fp16)[name = tensor("input_37_cast_fp16")]; + tensor input_39_axes_0 = const()[name = tensor("input_39_axes_0"), val = tensor([-1])]; + tensor norm3_2_weight_to_fp16 = const()[name = tensor("norm3_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(76781568)))]; + tensor norm3_2_bias_to_fp16 = const()[name = tensor("norm3_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(76783680)))]; + tensor var_1719_to_fp16 = const()[name = tensor("op_1719_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_39_cast_fp16 = layer_norm(axes = input_39_axes_0, beta = norm3_2_bias_to_fp16, epsilon = var_1719_to_fp16, gamma = norm3_2_weight_to_fp16, x = input_37_cast_fp16)[name = tensor("input_39_cast_fp16")]; + tensor linear_15_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = linear3_1_weight_cast_fp16, x = input_39_cast_fp16)[name = tensor("linear_15_cast_fp16")]; + tensor input_41_mode_0 = const()[name = tensor("input_41_mode_0"), val = tensor("EXACT")]; + tensor input_41_cast_fp16 = gelu(mode = input_41_mode_0, x = linear_15_cast_fp16)[name = tensor("input_41_cast_fp16")]; + tensor linear_16_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = linear3_2_weight_cast_fp16, x = input_41_cast_fp16)[name = tensor("linear_16_cast_fp16")]; + tensor input_43_cast_fp16 = add(x = input_37_cast_fp16, y = linear_16_cast_fp16)[name = tensor("input_43_cast_fp16")]; + tensor x_9_axes_0 = const()[name = tensor("x_9_axes_0"), val = tensor([-1])]; + tensor norm4_1_weight_to_fp16 = const()[name = tensor("norm4_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(76785792)))]; + tensor norm4_1_bias_to_fp16 = const()[name = tensor("norm4_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(76787904)))]; + tensor var_1736_to_fp16 = const()[name = tensor("op_1736_to_fp16"), val = tensor(0x1.5p-17)]; + tensor x_9_cast_fp16 = layer_norm(axes = x_9_axes_0, beta = norm4_1_bias_to_fp16, epsilon = var_1736_to_fp16, gamma = norm4_1_weight_to_fp16, x = input_43_cast_fp16)[name = tensor("x_9_cast_fp16")]; + tensor linear_17_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = attn4_in_proj_weight_cast_fp16, x = x_9_cast_fp16)[name = tensor("linear_17_cast_fp16")]; + tensor var_1772 = const()[name = tensor("op_1772"), val = tensor([1, 1, 3, 16, 64])]; + tensor qkv_9_cast_fp16 = reshape(shape = var_1772, x = linear_17_cast_fp16)[name = tensor("qkv_9_cast_fp16")]; + tensor q_25_begin_0 = const()[name = tensor("q_25_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor q_25_end_0 = const()[name = tensor("q_25_end_0"), val = tensor([1, 1, 1, 16, 64])]; + tensor q_25_end_mask_0 = const()[name = tensor("q_25_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor q_25_squeeze_mask_0 = const()[name = tensor("q_25_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor q_25_cast_fp16 = slice_by_index(begin = q_25_begin_0, end = q_25_end_0, end_mask = q_25_end_mask_0, squeeze_mask = q_25_squeeze_mask_0, x = qkv_9_cast_fp16)[name = tensor("q_25_cast_fp16")]; + tensor k_17_begin_0 = const()[name = tensor("k_17_begin_0"), val = tensor([0, 0, 1, 0, 0])]; + tensor k_17_end_0 = const()[name = tensor("k_17_end_0"), val = tensor([1, 1, 2, 16, 64])]; + tensor k_17_end_mask_0 = const()[name = tensor("k_17_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor k_17_squeeze_mask_0 = const()[name = tensor("k_17_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor k_17_cast_fp16 = slice_by_index(begin = k_17_begin_0, end = k_17_end_0, end_mask = k_17_end_mask_0, squeeze_mask = k_17_squeeze_mask_0, x = qkv_9_cast_fp16)[name = tensor("k_17_cast_fp16")]; + tensor v_9_begin_0 = const()[name = tensor("v_9_begin_0"), val = tensor([0, 0, 2, 0, 0])]; + tensor v_9_end_0 = const()[name = tensor("v_9_end_0"), val = tensor([1, 1, 3, 16, 64])]; + tensor v_9_end_mask_0 = const()[name = tensor("v_9_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor v_9_squeeze_mask_0 = const()[name = tensor("v_9_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor v_9_cast_fp16 = slice_by_index(begin = v_9_begin_0, end = v_9_end_0, end_mask = v_9_end_mask_0, squeeze_mask = v_9_squeeze_mask_0, x = qkv_9_cast_fp16)[name = tensor("v_9_cast_fp16")]; + tensor var_1876 = const()[name = tensor("op_1876"), val = tensor([1, 1, 1, 1])]; + tensor position4_to_fp16_dtype_0 = const()[name = tensor("position4_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor position4_to_fp16 = cast(dtype = position4_to_fp16_dtype_0, x = position4)[name = tensor("cast_119")]; + tensor ts_29_cast_fp16 = reshape(shape = var_1876, x = position4_to_fp16)[name = tensor("ts_29_cast_fp16")]; + tensor var_1880 = const()[name = tensor("op_1880"), val = tensor([1, 1, 16, 32, 2])]; + tensor q_complex_9_cast_fp16 = reshape(shape = var_1880, x = q_25_cast_fp16)[name = tensor("q_complex_9_cast_fp16")]; + tensor var_1884 = const()[name = tensor("op_1884"), val = tensor([1, 1, 16, 32, 2])]; + tensor k_complex_9_cast_fp16 = reshape(shape = var_1884, x = k_17_cast_fp16)[name = tensor("k_complex_9_cast_fp16")]; + tensor var_1888_begin_0 = const()[name = tensor("op_1888_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_1888_end_0 = const()[name = tensor("op_1888_end_0"), val = tensor([1, 1, 16, 32, 1])]; + tensor var_1888_end_mask_0 = const()[name = tensor("op_1888_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_1888_squeeze_mask_0 = const()[name = tensor("op_1888_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_1888_cast_fp16 = slice_by_index(begin = var_1888_begin_0, end = var_1888_end_0, end_mask = var_1888_end_mask_0, squeeze_mask = var_1888_squeeze_mask_0, x = q_complex_9_cast_fp16)[name = tensor("op_1888_cast_fp16")]; + tensor var_1896_begin_0 = const()[name = tensor("op_1896_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_1896_end_0 = const()[name = tensor("op_1896_end_0"), val = tensor([1, 1, 16, 32, 2])]; + tensor var_1896_end_mask_0 = const()[name = tensor("op_1896_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_1896_squeeze_mask_0 = const()[name = tensor("op_1896_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_1896_cast_fp16 = slice_by_index(begin = var_1896_begin_0, end = var_1896_end_0, end_mask = var_1896_end_mask_0, squeeze_mask = var_1896_squeeze_mask_0, x = q_complex_9_cast_fp16)[name = tensor("op_1896_cast_fp16")]; + tensor var_1904_begin_0 = const()[name = tensor("op_1904_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_1904_end_0 = const()[name = tensor("op_1904_end_0"), val = tensor([1, 1, 16, 32, 1])]; + tensor var_1904_end_mask_0 = const()[name = tensor("op_1904_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_1904_squeeze_mask_0 = const()[name = tensor("op_1904_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_1904_cast_fp16 = slice_by_index(begin = var_1904_begin_0, end = var_1904_end_0, end_mask = var_1904_end_mask_0, squeeze_mask = var_1904_squeeze_mask_0, x = k_complex_9_cast_fp16)[name = tensor("op_1904_cast_fp16")]; + tensor var_1912_begin_0 = const()[name = tensor("op_1912_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_1912_end_0 = const()[name = tensor("op_1912_end_0"), val = tensor([1, 1, 16, 32, 2])]; + tensor var_1912_end_mask_0 = const()[name = tensor("op_1912_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_1912_squeeze_mask_0 = const()[name = tensor("op_1912_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_1912_cast_fp16 = slice_by_index(begin = var_1912_begin_0, end = var_1912_end_0, end_mask = var_1912_end_mask_0, squeeze_mask = var_1912_squeeze_mask_0, x = k_complex_9_cast_fp16)[name = tensor("op_1912_cast_fp16")]; + tensor freqs_9_to_fp16 = const()[name = tensor("freqs_9_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(76790016)))]; + tensor var_1918_cast_fp16 = mul(x = freqs_9_to_fp16, y = ts_29_cast_fp16)[name = tensor("op_1918_cast_fp16")]; + tensor rotr_9_cast_fp16 = cos(x = var_1918_cast_fp16)[name = tensor("rotr_9_cast_fp16")]; + tensor roti_9_cast_fp16 = sin(x = var_1918_cast_fp16)[name = tensor("roti_9_cast_fp16")]; + tensor var_1922_cast_fp16 = mul(x = var_1888_cast_fp16, y = rotr_9_cast_fp16)[name = tensor("op_1922_cast_fp16")]; + tensor var_1923_cast_fp16 = mul(x = var_1896_cast_fp16, y = roti_9_cast_fp16)[name = tensor("op_1923_cast_fp16")]; + tensor qor_17_cast_fp16 = sub(x = var_1922_cast_fp16, y = var_1923_cast_fp16)[name = tensor("qor_17_cast_fp16")]; + tensor var_1926_cast_fp16 = mul(x = var_1888_cast_fp16, y = roti_9_cast_fp16)[name = tensor("op_1926_cast_fp16")]; + tensor var_1927_cast_fp16 = mul(x = var_1896_cast_fp16, y = rotr_9_cast_fp16)[name = tensor("op_1927_cast_fp16")]; + tensor qoi_17_cast_fp16 = add(x = var_1926_cast_fp16, y = var_1927_cast_fp16)[name = tensor("qoi_17_cast_fp16")]; + tensor var_1930_cast_fp16 = mul(x = var_1904_cast_fp16, y = rotr_9_cast_fp16)[name = tensor("op_1930_cast_fp16")]; + tensor var_1931_cast_fp16 = mul(x = var_1912_cast_fp16, y = roti_9_cast_fp16)[name = tensor("op_1931_cast_fp16")]; + tensor kor_17_cast_fp16 = sub(x = var_1930_cast_fp16, y = var_1931_cast_fp16)[name = tensor("kor_17_cast_fp16")]; + tensor var_1934_cast_fp16 = mul(x = var_1904_cast_fp16, y = roti_9_cast_fp16)[name = tensor("op_1934_cast_fp16")]; + tensor var_1935_cast_fp16 = mul(x = var_1912_cast_fp16, y = rotr_9_cast_fp16)[name = tensor("op_1935_cast_fp16")]; + tensor koi_17_cast_fp16 = add(x = var_1934_cast_fp16, y = var_1935_cast_fp16)[name = tensor("koi_17_cast_fp16")]; + tensor qo_9_axis_0 = const()[name = tensor("qo_9_axis_0"), val = tensor(-1)]; + tensor qo_9_cast_fp16 = stack(axis = qo_9_axis_0, values = (qor_17_cast_fp16, qoi_17_cast_fp16))[name = tensor("qo_9_cast_fp16")]; + tensor ko_9_axis_0 = const()[name = tensor("ko_9_axis_0"), val = tensor(-1)]; + tensor ko_9_cast_fp16 = stack(axis = ko_9_axis_0, values = (kor_17_cast_fp16, koi_17_cast_fp16))[name = tensor("ko_9_cast_fp16")]; + tensor var_1964 = const()[name = tensor("op_1964"), val = tensor([1, 1, 16, 64])]; + tensor q_27_cast_fp16 = reshape(shape = var_1964, x = qo_9_cast_fp16)[name = tensor("q_27_cast_fp16")]; + tensor var_1966 = const()[name = tensor("op_1966"), val = tensor([1, 1, 16, 64])]; + tensor k_19_cast_fp16 = reshape(shape = var_1966, x = ko_9_cast_fp16)[name = tensor("k_19_cast_fp16")]; + tensor _inversed_1988_y_0_to_fp16 = const()[name = tensor("_inversed_1988_y_0_to_fp16"), val = tensor(0x1p-9)]; + tensor _inversed_1988_cast_fp16 = mul(x = ts_29_cast_fp16, y = _inversed_1988_y_0_to_fp16)[name = tensor("_inversed_1988_cast_fp16")]; + tensor var_1989_cast_fp16 = floor(x = _inversed_1988_cast_fp16)[name = tensor("op_1989_cast_fp16")]; + tensor var_1990_to_fp16 = const()[name = tensor("op_1990_to_fp16"), val = tensor(0x1p+9)]; + tensor var_1991_cast_fp16 = mul(x = var_1989_cast_fp16, y = var_1990_to_fp16)[name = tensor("op_1991_cast_fp16")]; + tensor write_indices_float_19_cast_fp16 = sub(x = ts_29_cast_fp16, y = var_1991_cast_fp16)[name = tensor("write_indices_float_19_cast_fp16")]; + tensor var_1998_dtype_0 = const()[name = tensor("op_1998_dtype_0"), val = tensor("int32")]; + tensor write_indices_9_reps_0 = const()[name = tensor("write_indices_9_reps_0"), val = tensor([1, 1, 16, 64])]; + tensor write_indices_float_19_cast_fp16_to_int32 = cast(dtype = var_1998_dtype_0, x = write_indices_float_19_cast_fp16)[name = tensor("cast_118")]; + tensor write_indices_9 = tile(reps = write_indices_9_reps_0, x = write_indices_float_19_cast_fp16_to_int32)[name = tensor("write_indices_9")]; + tensor var_2006_begin_0 = const()[name = tensor("op_2006_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_2006_end_0 = const()[name = tensor("op_2006_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor var_2006_end_mask_0 = const()[name = tensor("op_2006_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_2006_squeeze_mask_0 = const()[name = tensor("op_2006_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor cache4_to_fp16_dtype_0 = const()[name = tensor("cache4_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor cache4_to_fp16 = cast(dtype = cache4_to_fp16_dtype_0, x = cache4)[name = tensor("cast_117")]; + tensor var_2006_cast_fp16 = slice_by_index(begin = var_2006_begin_0, end = var_2006_end_0, end_mask = var_2006_end_mask_0, squeeze_mask = var_2006_squeeze_mask_0, x = cache4_to_fp16)[name = tensor("op_2006_cast_fp16")]; + tensor var_2008_axis_0 = const()[name = tensor("op_2008_axis_0"), val = tensor(1)]; + tensor var_2008_mode_0 = const()[name = tensor("op_2008_mode_0"), val = tensor("update")]; + tensor var_2008_validate_indices_0 = const()[name = tensor("op_2008_validate_indices_0"), val = tensor(false)]; + tensor var_2008_cast_fp16 = scatter_along_axis(axis = var_2008_axis_0, data = var_2006_cast_fp16, indices = write_indices_9, mode = var_2008_mode_0, updates = k_19_cast_fp16, validate_indices = var_2008_validate_indices_0)[name = tensor("op_2008_cast_fp16")]; + tensor concat_30 = const()[name = tensor("concat_30"), val = tensor([0, 0, 0, 0, 0])]; + tensor concat_31 = const()[name = tensor("concat_31"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_9_internal_tensor_assign_1_stride_0 = const()[name = tensor("new_cache_9_internal_tensor_assign_1_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_9_internal_tensor_assign_1_begin_mask_0 = const()[name = tensor("new_cache_9_internal_tensor_assign_1_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_9_internal_tensor_assign_1_end_mask_0 = const()[name = tensor("new_cache_9_internal_tensor_assign_1_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_9_internal_tensor_assign_1_squeeze_mask_0 = const()[name = tensor("new_cache_9_internal_tensor_assign_1_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_20 = const()[name = tensor("shape_20"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_8 = const()[name = tensor("reduce_prod_8"), val = tensor(1048576)]; + tensor range_1d_8_start_0 = const()[name = tensor("range_1d_8_start_0"), val = tensor(0)]; + tensor range_1d_8_step_0 = const()[name = tensor("range_1d_8_step_0"), val = tensor(1)]; + tensor range_1d_8 = range_1d(end = reduce_prod_8, start = range_1d_8_start_0, step = range_1d_8_step_0)[name = tensor("range_1d_8")]; + tensor reshape_40 = reshape(shape = shape_20, x = range_1d_8)[name = tensor("reshape_40")]; + tensor slice_by_index_8 = slice_by_index(begin = concat_30, begin_mask = new_cache_9_internal_tensor_assign_1_begin_mask_0, end = concat_31, end_mask = new_cache_9_internal_tensor_assign_1_end_mask_0, squeeze_mask = new_cache_9_internal_tensor_assign_1_squeeze_mask_0, stride = new_cache_9_internal_tensor_assign_1_stride_0, x = reshape_40)[name = tensor("slice_by_index_8")]; + tensor reshape_41_shape_0 = const()[name = tensor("reshape_41_shape_0"), val = tensor([-1])]; + tensor reshape_41 = reshape(shape = reshape_41_shape_0, x = slice_by_index_8)[name = tensor("reshape_41")]; + tensor reshape_42_shape_0 = const()[name = tensor("reshape_42_shape_0"), val = tensor([-1])]; + tensor reshape_42_cast_fp16 = reshape(shape = reshape_42_shape_0, x = var_2008_cast_fp16)[name = tensor("reshape_42_cast_fp16")]; + tensor reshape_43_shape_0 = const()[name = tensor("reshape_43_shape_0"), val = tensor([-1])]; + tensor reshape_43_cast_fp16 = reshape(shape = reshape_43_shape_0, x = cache4_to_fp16)[name = tensor("reshape_43_cast_fp16")]; + tensor scatter_8_mode_0 = const()[name = tensor("scatter_8_mode_0"), val = tensor("update")]; + tensor scatter_8_axis_0 = const()[name = tensor("scatter_8_axis_0"), val = tensor(0)]; + tensor scatter_8_validate_indices_0 = const()[name = tensor("scatter_8_validate_indices_0"), val = tensor(false)]; + tensor scatter_8_cast_fp16 = scatter(axis = scatter_8_axis_0, data = reshape_43_cast_fp16, indices = reshape_41, mode = scatter_8_mode_0, updates = reshape_42_cast_fp16, validate_indices = scatter_8_validate_indices_0)[name = tensor("scatter_8_cast_fp16")]; + tensor reshape_44_cast_fp16 = reshape(shape = shape_20, x = scatter_8_cast_fp16)[name = tensor("reshape_44_cast_fp16")]; + tensor var_2016_begin_0 = const()[name = tensor("op_2016_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor var_2016_end_0 = const()[name = tensor("op_2016_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor var_2016_end_mask_0 = const()[name = tensor("op_2016_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_2016_squeeze_mask_0 = const()[name = tensor("op_2016_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor var_2016_cast_fp16 = slice_by_index(begin = var_2016_begin_0, end = var_2016_end_0, end_mask = var_2016_end_mask_0, squeeze_mask = var_2016_squeeze_mask_0, x = reshape_44_cast_fp16)[name = tensor("op_2016_cast_fp16")]; + tensor var_2018_axis_0 = const()[name = tensor("op_2018_axis_0"), val = tensor(1)]; + tensor var_2018_mode_0 = const()[name = tensor("op_2018_mode_0"), val = tensor("update")]; + tensor var_2018_validate_indices_0 = const()[name = tensor("op_2018_validate_indices_0"), val = tensor(false)]; + tensor var_2018_cast_fp16 = scatter_along_axis(axis = var_2018_axis_0, data = var_2016_cast_fp16, indices = write_indices_9, mode = var_2018_mode_0, updates = v_9_cast_fp16, validate_indices = var_2018_validate_indices_0)[name = tensor("op_2018_cast_fp16")]; + tensor concat_32 = const()[name = tensor("concat_32"), val = tensor([1, 0, 0, 0, 0])]; + tensor concat_33 = const()[name = tensor("concat_33"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_9_internal_tensor_assign_2_stride_0 = const()[name = tensor("new_cache_9_internal_tensor_assign_2_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_9_internal_tensor_assign_2_begin_mask_0 = const()[name = tensor("new_cache_9_internal_tensor_assign_2_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_9_internal_tensor_assign_2_end_mask_0 = const()[name = tensor("new_cache_9_internal_tensor_assign_2_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_9_internal_tensor_assign_2_squeeze_mask_0 = const()[name = tensor("new_cache_9_internal_tensor_assign_2_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_21 = const()[name = tensor("shape_21"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_9 = const()[name = tensor("reduce_prod_9"), val = tensor(1048576)]; + tensor range_1d_9_start_0 = const()[name = tensor("range_1d_9_start_0"), val = tensor(0)]; + tensor range_1d_9_step_0 = const()[name = tensor("range_1d_9_step_0"), val = tensor(1)]; + tensor range_1d_9 = range_1d(end = reduce_prod_9, start = range_1d_9_start_0, step = range_1d_9_step_0)[name = tensor("range_1d_9")]; + tensor reshape_45 = reshape(shape = shape_21, x = range_1d_9)[name = tensor("reshape_45")]; + tensor slice_by_index_9 = slice_by_index(begin = concat_32, begin_mask = new_cache_9_internal_tensor_assign_2_begin_mask_0, end = concat_33, end_mask = new_cache_9_internal_tensor_assign_2_end_mask_0, squeeze_mask = new_cache_9_internal_tensor_assign_2_squeeze_mask_0, stride = new_cache_9_internal_tensor_assign_2_stride_0, x = reshape_45)[name = tensor("slice_by_index_9")]; + tensor reshape_46_shape_0 = const()[name = tensor("reshape_46_shape_0"), val = tensor([-1])]; + tensor reshape_46 = reshape(shape = reshape_46_shape_0, x = slice_by_index_9)[name = tensor("reshape_46")]; + tensor reshape_47_shape_0 = const()[name = tensor("reshape_47_shape_0"), val = tensor([-1])]; + tensor reshape_47_cast_fp16 = reshape(shape = reshape_47_shape_0, x = var_2018_cast_fp16)[name = tensor("reshape_47_cast_fp16")]; + tensor reshape_48_shape_0 = const()[name = tensor("reshape_48_shape_0"), val = tensor([-1])]; + tensor reshape_48_cast_fp16 = reshape(shape = reshape_48_shape_0, x = reshape_44_cast_fp16)[name = tensor("reshape_48_cast_fp16")]; + tensor scatter_9_mode_0 = const()[name = tensor("scatter_9_mode_0"), val = tensor("update")]; + tensor scatter_9_axis_0 = const()[name = tensor("scatter_9_axis_0"), val = tensor(0)]; + tensor scatter_9_validate_indices_0 = const()[name = tensor("scatter_9_validate_indices_0"), val = tensor(false)]; + tensor scatter_9_cast_fp16 = scatter(axis = scatter_9_axis_0, data = reshape_48_cast_fp16, indices = reshape_46, mode = scatter_9_mode_0, updates = reshape_47_cast_fp16, validate_indices = scatter_9_validate_indices_0)[name = tensor("scatter_9_cast_fp16")]; + tensor reshape_49_cast_fp16 = reshape(shape = shape_21, x = scatter_9_cast_fp16)[name = tensor("reshape_49_cast_fp16")]; + tensor reshape_49_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("reshape_49_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor keys_25_begin_0 = const()[name = tensor("keys_25_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor keys_25_end_0 = const()[name = tensor("keys_25_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor keys_25_end_mask_0 = const()[name = tensor("keys_25_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor keys_25_squeeze_mask_0 = const()[name = tensor("keys_25_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor keys_25_cast_fp16 = slice_by_index(begin = keys_25_begin_0, end = keys_25_end_0, end_mask = keys_25_end_mask_0, squeeze_mask = keys_25_squeeze_mask_0, x = reshape_49_cast_fp16)[name = tensor("keys_25_cast_fp16")]; + tensor values_25_begin_0 = const()[name = tensor("values_25_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor values_25_end_0 = const()[name = tensor("values_25_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor values_25_end_mask_0 = const()[name = tensor("values_25_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor values_25_squeeze_mask_0 = const()[name = tensor("values_25_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor values_25_cast_fp16 = slice_by_index(begin = values_25_begin_0, end = values_25_end_0, end_mask = values_25_end_mask_0, squeeze_mask = values_25_squeeze_mask_0, x = reshape_49_cast_fp16)[name = tensor("values_25_cast_fp16")]; + tensor var_2030_cast_fp16 = not_equal(x = keys_25_cast_fp16, y = keys_25_cast_fp16)[name = tensor("op_2030_cast_fp16")]; + tensor keys_27_cast_fp16 = select(a = var_360_to_fp16, b = keys_25_cast_fp16, cond = var_2030_cast_fp16)[name = tensor("keys_27_cast_fp16")]; + tensor var_2038_cast_fp16 = not_equal(x = values_25_cast_fp16, y = values_25_cast_fp16)[name = tensor("op_2038_cast_fp16")]; + tensor values_27_cast_fp16 = select(a = var_360_to_fp16, b = values_25_cast_fp16, cond = var_2038_cast_fp16)[name = tensor("values_27_cast_fp16")]; + tensor var_2062 = const()[name = tensor("op_2062"), val = tensor([0, 2, 1, 3])]; + tensor var_2075 = const()[name = tensor("op_2075"), val = tensor([1, 1, 1])]; + tensor var_2076_cast_fp16 = reshape(shape = var_2075, x = position4_to_fp16)[name = tensor("op_2076_cast_fp16")]; + tensor var_2093_to_fp16 = const()[name = tensor("op_2093_to_fp16"), val = tensor(0x1p+0)]; + tensor valid_len_9_cast_fp16 = add(x = var_2076_cast_fp16, y = var_2093_to_fp16)[name = tensor("valid_len_9_cast_fp16")]; + tensor valid_mask_9_cast_fp16 = less(x = k_positions_1_promoted_to_fp16, y = valid_len_9_cast_fp16)[name = tensor("valid_mask_9_cast_fp16")]; + tensor causal_mask_9_cast_fp16 = less_equal(x = k_positions_1_promoted_to_fp16, y = var_2076_cast_fp16)[name = tensor("causal_mask_9_cast_fp16")]; + tensor attn_mask_17 = logical_and(x = valid_mask_9_cast_fp16, y = causal_mask_9_cast_fp16)[name = tensor("attn_mask_17")]; + tensor attn_mask_19_axes_0 = const()[name = tensor("attn_mask_19_axes_0"), val = tensor([1])]; + tensor attn_mask_19 = expand_dims(axes = attn_mask_19_axes_0, x = attn_mask_17)[name = tensor("attn_mask_19")]; + tensor var_2111_transpose_x_0 = const()[name = tensor("op_2111_transpose_x_0"), val = tensor(false)]; + tensor var_2111_transpose_y_0 = const()[name = tensor("op_2111_transpose_y_0"), val = tensor(false)]; + tensor transpose_26_perm_0 = const()[name = tensor("transpose_26_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_27_perm_0 = const()[name = tensor("transpose_27_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_27 = transpose(perm = transpose_27_perm_0, x = keys_27_cast_fp16)[name = tensor("transpose_35")]; + tensor transpose_26 = transpose(perm = transpose_26_perm_0, x = q_27_cast_fp16)[name = tensor("transpose_36")]; + tensor var_2111_cast_fp16 = matmul(transpose_x = var_2111_transpose_x_0, transpose_y = var_2111_transpose_y_0, x = transpose_26, y = transpose_27)[name = tensor("op_2111_cast_fp16")]; + tensor var_2105_to_fp16 = const()[name = tensor("op_2105_to_fp16"), val = tensor([0x1p-3])]; + tensor attn_weights_25_cast_fp16 = mul(x = var_2111_cast_fp16, y = var_2105_to_fp16)[name = tensor("attn_weights_25_cast_fp16")]; + tensor var_2113 = logical_not(x = attn_mask_19)[name = tensor("op_2113")]; + tensor var_2114_to_fp16 = const()[name = tensor("op_2114_to_fp16"), val = tensor(-inf)]; + tensor attn_weights_27_cast_fp16 = select(a = var_2114_to_fp16, b = attn_weights_25_cast_fp16, cond = var_2113)[name = tensor("attn_weights_27_cast_fp16")]; + tensor var_2116 = const()[name = tensor("op_2116"), val = tensor(-1)]; + tensor attn_weights_29_cast_fp16 = softmax(axis = var_2116, x = attn_weights_27_cast_fp16)[name = tensor("attn_weights_29_cast_fp16")]; + tensor attn_output_9_transpose_x_0 = const()[name = tensor("attn_output_9_transpose_x_0"), val = tensor(false)]; + tensor attn_output_9_transpose_y_0 = const()[name = tensor("attn_output_9_transpose_y_0"), val = tensor(false)]; + tensor values_29_cast_fp16 = transpose(perm = var_2062, x = values_27_cast_fp16)[name = tensor("transpose_37")]; + tensor attn_output_9_cast_fp16 = matmul(transpose_x = attn_output_9_transpose_x_0, transpose_y = attn_output_9_transpose_y_0, x = attn_weights_29_cast_fp16, y = values_29_cast_fp16)[name = tensor("attn_output_9_cast_fp16")]; + tensor var_2124 = const()[name = tensor("op_2124"), val = tensor([0, 2, 1, 3])]; + tensor var_2127 = const()[name = tensor("op_2127"), val = tensor([1, 1, 1024])]; + tensor var_2125_cast_fp16 = transpose(perm = var_2124, x = attn_output_9_cast_fp16)[name = tensor("transpose_34")]; + tensor input_45_cast_fp16 = reshape(shape = var_2127, x = var_2125_cast_fp16)[name = tensor("input_45_cast_fp16")]; + tensor linear_18_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = attn4_out_proj_weight_cast_fp16, x = input_45_cast_fp16)[name = tensor("linear_18_cast_fp16")]; + tensor var_2133_to_fp16 = const()[name = tensor("op_2133_to_fp16"), val = tensor(0x1p+0)]; + tensor var_2134_cast_fp16 = add(x = position4_to_fp16, y = var_2133_to_fp16)[name = tensor("op_2134_cast_fp16")]; + tensor var_2134_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_2134_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor input_47_cast_fp16 = add(x = input_43_cast_fp16, y = linear_18_cast_fp16)[name = tensor("input_47_cast_fp16")]; + tensor input_49_axes_0 = const()[name = tensor("input_49_axes_0"), val = tensor([-1])]; + tensor norm4_2_weight_to_fp16 = const()[name = tensor("norm4_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(76790144)))]; + tensor norm4_2_bias_to_fp16 = const()[name = tensor("norm4_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(76792256)))]; + tensor var_2138_to_fp16 = const()[name = tensor("op_2138_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_49_cast_fp16 = layer_norm(axes = input_49_axes_0, beta = norm4_2_bias_to_fp16, epsilon = var_2138_to_fp16, gamma = norm4_2_weight_to_fp16, x = input_47_cast_fp16)[name = tensor("input_49_cast_fp16")]; + tensor linear_19_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = linear4_1_weight_cast_fp16, x = input_49_cast_fp16)[name = tensor("linear_19_cast_fp16")]; + tensor input_51_mode_0 = const()[name = tensor("input_51_mode_0"), val = tensor("EXACT")]; + tensor input_51_cast_fp16 = gelu(mode = input_51_mode_0, x = linear_19_cast_fp16)[name = tensor("input_51_cast_fp16")]; + tensor linear_20_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = linear4_2_weight_cast_fp16, x = input_51_cast_fp16)[name = tensor("linear_20_cast_fp16")]; + tensor input_53_cast_fp16 = add(x = input_47_cast_fp16, y = linear_20_cast_fp16)[name = tensor("input_53_cast_fp16")]; + tensor x_axes_0 = const()[name = tensor("x_axes_0"), val = tensor([-1])]; + tensor norm5_1_weight_to_fp16 = const()[name = tensor("norm5_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(76794368)))]; + tensor norm5_1_bias_to_fp16 = const()[name = tensor("norm5_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(76796480)))]; + tensor var_2155_to_fp16 = const()[name = tensor("op_2155_to_fp16"), val = tensor(0x1.5p-17)]; + tensor x_cast_fp16 = layer_norm(axes = x_axes_0, beta = norm5_1_bias_to_fp16, epsilon = var_2155_to_fp16, gamma = norm5_1_weight_to_fp16, x = input_53_cast_fp16)[name = tensor("x_cast_fp16")]; + tensor linear_21_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = attn5_in_proj_weight_cast_fp16, x = x_cast_fp16)[name = tensor("linear_21_cast_fp16")]; + tensor var_2191 = const()[name = tensor("op_2191"), val = tensor([1, 1, 3, 16, 64])]; + tensor qkv_cast_fp16 = reshape(shape = var_2191, x = linear_21_cast_fp16)[name = tensor("qkv_cast_fp16")]; + tensor q_31_begin_0 = const()[name = tensor("q_31_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor q_31_end_0 = const()[name = tensor("q_31_end_0"), val = tensor([1, 1, 1, 16, 64])]; + tensor q_31_end_mask_0 = const()[name = tensor("q_31_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor q_31_squeeze_mask_0 = const()[name = tensor("q_31_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor q_31_cast_fp16 = slice_by_index(begin = q_31_begin_0, end = q_31_end_0, end_mask = q_31_end_mask_0, squeeze_mask = q_31_squeeze_mask_0, x = qkv_cast_fp16)[name = tensor("q_31_cast_fp16")]; + tensor k_21_begin_0 = const()[name = tensor("k_21_begin_0"), val = tensor([0, 0, 1, 0, 0])]; + tensor k_21_end_0 = const()[name = tensor("k_21_end_0"), val = tensor([1, 1, 2, 16, 64])]; + tensor k_21_end_mask_0 = const()[name = tensor("k_21_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor k_21_squeeze_mask_0 = const()[name = tensor("k_21_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor k_21_cast_fp16 = slice_by_index(begin = k_21_begin_0, end = k_21_end_0, end_mask = k_21_end_mask_0, squeeze_mask = k_21_squeeze_mask_0, x = qkv_cast_fp16)[name = tensor("k_21_cast_fp16")]; + tensor v_begin_0 = const()[name = tensor("v_begin_0"), val = tensor([0, 0, 2, 0, 0])]; + tensor v_end_0 = const()[name = tensor("v_end_0"), val = tensor([1, 1, 3, 16, 64])]; + tensor v_end_mask_0 = const()[name = tensor("v_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor v_squeeze_mask_0 = const()[name = tensor("v_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor v_cast_fp16 = slice_by_index(begin = v_begin_0, end = v_end_0, end_mask = v_end_mask_0, squeeze_mask = v_squeeze_mask_0, x = qkv_cast_fp16)[name = tensor("v_cast_fp16")]; + tensor var_2295 = const()[name = tensor("op_2295"), val = tensor([1, 1, 1, 1])]; + tensor position5_to_fp16_dtype_0 = const()[name = tensor("position5_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor position5_to_fp16 = cast(dtype = position5_to_fp16_dtype_0, x = position5)[name = tensor("cast_114")]; + tensor ts_cast_fp16 = reshape(shape = var_2295, x = position5_to_fp16)[name = tensor("ts_cast_fp16")]; + tensor var_2299 = const()[name = tensor("op_2299"), val = tensor([1, 1, 16, 32, 2])]; + tensor q_complex_cast_fp16 = reshape(shape = var_2299, x = q_31_cast_fp16)[name = tensor("q_complex_cast_fp16")]; + tensor var_2303 = const()[name = tensor("op_2303"), val = tensor([1, 1, 16, 32, 2])]; + tensor k_complex_cast_fp16 = reshape(shape = var_2303, x = k_21_cast_fp16)[name = tensor("k_complex_cast_fp16")]; + tensor var_2307_begin_0 = const()[name = tensor("op_2307_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_2307_end_0 = const()[name = tensor("op_2307_end_0"), val = tensor([1, 1, 16, 32, 1])]; + tensor var_2307_end_mask_0 = const()[name = tensor("op_2307_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_2307_squeeze_mask_0 = const()[name = tensor("op_2307_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_2307_cast_fp16 = slice_by_index(begin = var_2307_begin_0, end = var_2307_end_0, end_mask = var_2307_end_mask_0, squeeze_mask = var_2307_squeeze_mask_0, x = q_complex_cast_fp16)[name = tensor("op_2307_cast_fp16")]; + tensor var_2315_begin_0 = const()[name = tensor("op_2315_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_2315_end_0 = const()[name = tensor("op_2315_end_0"), val = tensor([1, 1, 16, 32, 2])]; + tensor var_2315_end_mask_0 = const()[name = tensor("op_2315_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_2315_squeeze_mask_0 = const()[name = tensor("op_2315_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_2315_cast_fp16 = slice_by_index(begin = var_2315_begin_0, end = var_2315_end_0, end_mask = var_2315_end_mask_0, squeeze_mask = var_2315_squeeze_mask_0, x = q_complex_cast_fp16)[name = tensor("op_2315_cast_fp16")]; + tensor var_2323_begin_0 = const()[name = tensor("op_2323_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_2323_end_0 = const()[name = tensor("op_2323_end_0"), val = tensor([1, 1, 16, 32, 1])]; + tensor var_2323_end_mask_0 = const()[name = tensor("op_2323_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_2323_squeeze_mask_0 = const()[name = tensor("op_2323_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_2323_cast_fp16 = slice_by_index(begin = var_2323_begin_0, end = var_2323_end_0, end_mask = var_2323_end_mask_0, squeeze_mask = var_2323_squeeze_mask_0, x = k_complex_cast_fp16)[name = tensor("op_2323_cast_fp16")]; + tensor var_2331_begin_0 = const()[name = tensor("op_2331_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_2331_end_0 = const()[name = tensor("op_2331_end_0"), val = tensor([1, 1, 16, 32, 2])]; + tensor var_2331_end_mask_0 = const()[name = tensor("op_2331_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_2331_squeeze_mask_0 = const()[name = tensor("op_2331_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_2331_cast_fp16 = slice_by_index(begin = var_2331_begin_0, end = var_2331_end_0, end_mask = var_2331_end_mask_0, squeeze_mask = var_2331_squeeze_mask_0, x = k_complex_cast_fp16)[name = tensor("op_2331_cast_fp16")]; + tensor freqs_to_fp16 = const()[name = tensor("freqs_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(76798592)))]; + tensor var_2337_cast_fp16 = mul(x = freqs_to_fp16, y = ts_cast_fp16)[name = tensor("op_2337_cast_fp16")]; + tensor rotr_cast_fp16 = cos(x = var_2337_cast_fp16)[name = tensor("rotr_cast_fp16")]; + tensor roti_cast_fp16 = sin(x = var_2337_cast_fp16)[name = tensor("roti_cast_fp16")]; + tensor var_2341_cast_fp16 = mul(x = var_2307_cast_fp16, y = rotr_cast_fp16)[name = tensor("op_2341_cast_fp16")]; + tensor var_2342_cast_fp16 = mul(x = var_2315_cast_fp16, y = roti_cast_fp16)[name = tensor("op_2342_cast_fp16")]; + tensor qor_21_cast_fp16 = sub(x = var_2341_cast_fp16, y = var_2342_cast_fp16)[name = tensor("qor_21_cast_fp16")]; + tensor var_2345_cast_fp16 = mul(x = var_2307_cast_fp16, y = roti_cast_fp16)[name = tensor("op_2345_cast_fp16")]; + tensor var_2346_cast_fp16 = mul(x = var_2315_cast_fp16, y = rotr_cast_fp16)[name = tensor("op_2346_cast_fp16")]; + tensor qoi_21_cast_fp16 = add(x = var_2345_cast_fp16, y = var_2346_cast_fp16)[name = tensor("qoi_21_cast_fp16")]; + tensor var_2349_cast_fp16 = mul(x = var_2323_cast_fp16, y = rotr_cast_fp16)[name = tensor("op_2349_cast_fp16")]; + tensor var_2350_cast_fp16 = mul(x = var_2331_cast_fp16, y = roti_cast_fp16)[name = tensor("op_2350_cast_fp16")]; + tensor kor_21_cast_fp16 = sub(x = var_2349_cast_fp16, y = var_2350_cast_fp16)[name = tensor("kor_21_cast_fp16")]; + tensor var_2353_cast_fp16 = mul(x = var_2323_cast_fp16, y = roti_cast_fp16)[name = tensor("op_2353_cast_fp16")]; + tensor var_2354_cast_fp16 = mul(x = var_2331_cast_fp16, y = rotr_cast_fp16)[name = tensor("op_2354_cast_fp16")]; + tensor koi_21_cast_fp16 = add(x = var_2353_cast_fp16, y = var_2354_cast_fp16)[name = tensor("koi_21_cast_fp16")]; + tensor qo_axis_0 = const()[name = tensor("qo_axis_0"), val = tensor(-1)]; + tensor qo_cast_fp16 = stack(axis = qo_axis_0, values = (qor_21_cast_fp16, qoi_21_cast_fp16))[name = tensor("qo_cast_fp16")]; + tensor ko_axis_0 = const()[name = tensor("ko_axis_0"), val = tensor(-1)]; + tensor ko_cast_fp16 = stack(axis = ko_axis_0, values = (kor_21_cast_fp16, koi_21_cast_fp16))[name = tensor("ko_cast_fp16")]; + tensor var_2383 = const()[name = tensor("op_2383"), val = tensor([1, 1, 16, 64])]; + tensor q_33_cast_fp16 = reshape(shape = var_2383, x = qo_cast_fp16)[name = tensor("q_33_cast_fp16")]; + tensor var_2385 = const()[name = tensor("op_2385"), val = tensor([1, 1, 16, 64])]; + tensor k_cast_fp16 = reshape(shape = var_2385, x = ko_cast_fp16)[name = tensor("k_cast_fp16")]; + tensor _inversed_2407_y_0_to_fp16 = const()[name = tensor("_inversed_2407_y_0_to_fp16"), val = tensor(0x1p-9)]; + tensor _inversed_2407_cast_fp16 = mul(x = ts_cast_fp16, y = _inversed_2407_y_0_to_fp16)[name = tensor("_inversed_2407_cast_fp16")]; + tensor var_2408_cast_fp16 = floor(x = _inversed_2407_cast_fp16)[name = tensor("op_2408_cast_fp16")]; + tensor var_2409_to_fp16 = const()[name = tensor("op_2409_to_fp16"), val = tensor(0x1p+9)]; + tensor var_2410_cast_fp16 = mul(x = var_2408_cast_fp16, y = var_2409_to_fp16)[name = tensor("op_2410_cast_fp16")]; + tensor write_indices_float_cast_fp16 = sub(x = ts_cast_fp16, y = var_2410_cast_fp16)[name = tensor("write_indices_float_cast_fp16")]; + tensor var_2417_dtype_0 = const()[name = tensor("op_2417_dtype_0"), val = tensor("int32")]; + tensor write_indices_reps_0 = const()[name = tensor("write_indices_reps_0"), val = tensor([1, 1, 16, 64])]; + tensor write_indices_float_cast_fp16_to_int32 = cast(dtype = var_2417_dtype_0, x = write_indices_float_cast_fp16)[name = tensor("cast_113")]; + tensor write_indices = tile(reps = write_indices_reps_0, x = write_indices_float_cast_fp16_to_int32)[name = tensor("write_indices")]; + tensor var_2425_begin_0 = const()[name = tensor("op_2425_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_2425_end_0 = const()[name = tensor("op_2425_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor var_2425_end_mask_0 = const()[name = tensor("op_2425_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_2425_squeeze_mask_0 = const()[name = tensor("op_2425_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor cache5_to_fp16_dtype_0 = const()[name = tensor("cache5_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor cache5_to_fp16 = cast(dtype = cache5_to_fp16_dtype_0, x = cache5)[name = tensor("cast_112")]; + tensor var_2425_cast_fp16 = slice_by_index(begin = var_2425_begin_0, end = var_2425_end_0, end_mask = var_2425_end_mask_0, squeeze_mask = var_2425_squeeze_mask_0, x = cache5_to_fp16)[name = tensor("op_2425_cast_fp16")]; + tensor var_2427_axis_0 = const()[name = tensor("op_2427_axis_0"), val = tensor(1)]; + tensor var_2427_mode_0 = const()[name = tensor("op_2427_mode_0"), val = tensor("update")]; + tensor var_2427_validate_indices_0 = const()[name = tensor("op_2427_validate_indices_0"), val = tensor(false)]; + tensor var_2427_cast_fp16 = scatter_along_axis(axis = var_2427_axis_0, data = var_2425_cast_fp16, indices = write_indices, mode = var_2427_mode_0, updates = k_cast_fp16, validate_indices = var_2427_validate_indices_0)[name = tensor("op_2427_cast_fp16")]; + tensor concat_37 = const()[name = tensor("concat_37"), val = tensor([0, 0, 0, 0, 0])]; + tensor concat_38 = const()[name = tensor("concat_38"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_internal_tensor_assign_1_stride_0 = const()[name = tensor("new_cache_internal_tensor_assign_1_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_internal_tensor_assign_1_begin_mask_0 = const()[name = tensor("new_cache_internal_tensor_assign_1_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_internal_tensor_assign_1_end_mask_0 = const()[name = tensor("new_cache_internal_tensor_assign_1_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_internal_tensor_assign_1_squeeze_mask_0 = const()[name = tensor("new_cache_internal_tensor_assign_1_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_22 = const()[name = tensor("shape_22"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_10 = const()[name = tensor("reduce_prod_10"), val = tensor(1048576)]; + tensor range_1d_10_start_0 = const()[name = tensor("range_1d_10_start_0"), val = tensor(0)]; + tensor range_1d_10_step_0 = const()[name = tensor("range_1d_10_step_0"), val = tensor(1)]; + tensor range_1d_10 = range_1d(end = reduce_prod_10, start = range_1d_10_start_0, step = range_1d_10_step_0)[name = tensor("range_1d_10")]; + tensor reshape_50 = reshape(shape = shape_22, x = range_1d_10)[name = tensor("reshape_50")]; + tensor slice_by_index_10 = slice_by_index(begin = concat_37, begin_mask = new_cache_internal_tensor_assign_1_begin_mask_0, end = concat_38, end_mask = new_cache_internal_tensor_assign_1_end_mask_0, squeeze_mask = new_cache_internal_tensor_assign_1_squeeze_mask_0, stride = new_cache_internal_tensor_assign_1_stride_0, x = reshape_50)[name = tensor("slice_by_index_10")]; + tensor reshape_51_shape_0 = const()[name = tensor("reshape_51_shape_0"), val = tensor([-1])]; + tensor reshape_51 = reshape(shape = reshape_51_shape_0, x = slice_by_index_10)[name = tensor("reshape_51")]; + tensor reshape_52_shape_0 = const()[name = tensor("reshape_52_shape_0"), val = tensor([-1])]; + tensor reshape_52_cast_fp16 = reshape(shape = reshape_52_shape_0, x = var_2427_cast_fp16)[name = tensor("reshape_52_cast_fp16")]; + tensor reshape_53_shape_0 = const()[name = tensor("reshape_53_shape_0"), val = tensor([-1])]; + tensor reshape_53_cast_fp16 = reshape(shape = reshape_53_shape_0, x = cache5_to_fp16)[name = tensor("reshape_53_cast_fp16")]; + tensor scatter_10_mode_0 = const()[name = tensor("scatter_10_mode_0"), val = tensor("update")]; + tensor scatter_10_axis_0 = const()[name = tensor("scatter_10_axis_0"), val = tensor(0)]; + tensor scatter_10_validate_indices_0 = const()[name = tensor("scatter_10_validate_indices_0"), val = tensor(false)]; + tensor scatter_10_cast_fp16 = scatter(axis = scatter_10_axis_0, data = reshape_53_cast_fp16, indices = reshape_51, mode = scatter_10_mode_0, updates = reshape_52_cast_fp16, validate_indices = scatter_10_validate_indices_0)[name = tensor("scatter_10_cast_fp16")]; + tensor reshape_54_cast_fp16 = reshape(shape = shape_22, x = scatter_10_cast_fp16)[name = tensor("reshape_54_cast_fp16")]; + tensor var_2435_begin_0 = const()[name = tensor("op_2435_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor var_2435_end_0 = const()[name = tensor("op_2435_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor var_2435_end_mask_0 = const()[name = tensor("op_2435_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_2435_squeeze_mask_0 = const()[name = tensor("op_2435_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor var_2435_cast_fp16 = slice_by_index(begin = var_2435_begin_0, end = var_2435_end_0, end_mask = var_2435_end_mask_0, squeeze_mask = var_2435_squeeze_mask_0, x = reshape_54_cast_fp16)[name = tensor("op_2435_cast_fp16")]; + tensor var_2437_axis_0 = const()[name = tensor("op_2437_axis_0"), val = tensor(1)]; + tensor var_2437_mode_0 = const()[name = tensor("op_2437_mode_0"), val = tensor("update")]; + tensor var_2437_validate_indices_0 = const()[name = tensor("op_2437_validate_indices_0"), val = tensor(false)]; + tensor var_2437_cast_fp16 = scatter_along_axis(axis = var_2437_axis_0, data = var_2435_cast_fp16, indices = write_indices, mode = var_2437_mode_0, updates = v_cast_fp16, validate_indices = var_2437_validate_indices_0)[name = tensor("op_2437_cast_fp16")]; + tensor concat_39 = const()[name = tensor("concat_39"), val = tensor([1, 0, 0, 0, 0])]; + tensor concat_40 = const()[name = tensor("concat_40"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_internal_tensor_assign_2_stride_0 = const()[name = tensor("new_cache_internal_tensor_assign_2_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_internal_tensor_assign_2_begin_mask_0 = const()[name = tensor("new_cache_internal_tensor_assign_2_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_internal_tensor_assign_2_end_mask_0 = const()[name = tensor("new_cache_internal_tensor_assign_2_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_internal_tensor_assign_2_squeeze_mask_0 = const()[name = tensor("new_cache_internal_tensor_assign_2_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_23 = const()[name = tensor("shape_23"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_11 = const()[name = tensor("reduce_prod_11"), val = tensor(1048576)]; + tensor range_1d_11_start_0 = const()[name = tensor("range_1d_11_start_0"), val = tensor(0)]; + tensor range_1d_11_step_0 = const()[name = tensor("range_1d_11_step_0"), val = tensor(1)]; + tensor range_1d_11 = range_1d(end = reduce_prod_11, start = range_1d_11_start_0, step = range_1d_11_step_0)[name = tensor("range_1d_11")]; + tensor reshape_55 = reshape(shape = shape_23, x = range_1d_11)[name = tensor("reshape_55")]; + tensor slice_by_index_11 = slice_by_index(begin = concat_39, begin_mask = new_cache_internal_tensor_assign_2_begin_mask_0, end = concat_40, end_mask = new_cache_internal_tensor_assign_2_end_mask_0, squeeze_mask = new_cache_internal_tensor_assign_2_squeeze_mask_0, stride = new_cache_internal_tensor_assign_2_stride_0, x = reshape_55)[name = tensor("slice_by_index_11")]; + tensor reshape_56_shape_0 = const()[name = tensor("reshape_56_shape_0"), val = tensor([-1])]; + tensor reshape_56 = reshape(shape = reshape_56_shape_0, x = slice_by_index_11)[name = tensor("reshape_56")]; + tensor reshape_57_shape_0 = const()[name = tensor("reshape_57_shape_0"), val = tensor([-1])]; + tensor reshape_57_cast_fp16 = reshape(shape = reshape_57_shape_0, x = var_2437_cast_fp16)[name = tensor("reshape_57_cast_fp16")]; + tensor reshape_58_shape_0 = const()[name = tensor("reshape_58_shape_0"), val = tensor([-1])]; + tensor reshape_58_cast_fp16 = reshape(shape = reshape_58_shape_0, x = reshape_54_cast_fp16)[name = tensor("reshape_58_cast_fp16")]; + tensor scatter_11_mode_0 = const()[name = tensor("scatter_11_mode_0"), val = tensor("update")]; + tensor scatter_11_axis_0 = const()[name = tensor("scatter_11_axis_0"), val = tensor(0)]; + tensor scatter_11_validate_indices_0 = const()[name = tensor("scatter_11_validate_indices_0"), val = tensor(false)]; + tensor scatter_11_cast_fp16 = scatter(axis = scatter_11_axis_0, data = reshape_58_cast_fp16, indices = reshape_56, mode = scatter_11_mode_0, updates = reshape_57_cast_fp16, validate_indices = scatter_11_validate_indices_0)[name = tensor("scatter_11_cast_fp16")]; + tensor reshape_59_cast_fp16 = reshape(shape = shape_23, x = scatter_11_cast_fp16)[name = tensor("reshape_59_cast_fp16")]; + tensor reshape_59_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("reshape_59_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor keys_31_begin_0 = const()[name = tensor("keys_31_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor keys_31_end_0 = const()[name = tensor("keys_31_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor keys_31_end_mask_0 = const()[name = tensor("keys_31_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor keys_31_squeeze_mask_0 = const()[name = tensor("keys_31_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor keys_31_cast_fp16 = slice_by_index(begin = keys_31_begin_0, end = keys_31_end_0, end_mask = keys_31_end_mask_0, squeeze_mask = keys_31_squeeze_mask_0, x = reshape_59_cast_fp16)[name = tensor("keys_31_cast_fp16")]; + tensor values_31_begin_0 = const()[name = tensor("values_31_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor values_31_end_0 = const()[name = tensor("values_31_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor values_31_end_mask_0 = const()[name = tensor("values_31_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor values_31_squeeze_mask_0 = const()[name = tensor("values_31_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor values_31_cast_fp16 = slice_by_index(begin = values_31_begin_0, end = values_31_end_0, end_mask = values_31_end_mask_0, squeeze_mask = values_31_squeeze_mask_0, x = reshape_59_cast_fp16)[name = tensor("values_31_cast_fp16")]; + tensor var_2449_cast_fp16 = not_equal(x = keys_31_cast_fp16, y = keys_31_cast_fp16)[name = tensor("op_2449_cast_fp16")]; + tensor keys_33_cast_fp16 = select(a = var_360_to_fp16, b = keys_31_cast_fp16, cond = var_2449_cast_fp16)[name = tensor("keys_33_cast_fp16")]; + tensor var_2457_cast_fp16 = not_equal(x = values_31_cast_fp16, y = values_31_cast_fp16)[name = tensor("op_2457_cast_fp16")]; + tensor values_33_cast_fp16 = select(a = var_360_to_fp16, b = values_31_cast_fp16, cond = var_2457_cast_fp16)[name = tensor("values_33_cast_fp16")]; + tensor var_2481 = const()[name = tensor("op_2481"), val = tensor([0, 2, 1, 3])]; + tensor var_2494 = const()[name = tensor("op_2494"), val = tensor([1, 1, 1])]; + tensor var_2495_cast_fp16 = reshape(shape = var_2494, x = position5_to_fp16)[name = tensor("op_2495_cast_fp16")]; + tensor var_2512_to_fp16 = const()[name = tensor("op_2512_to_fp16"), val = tensor(0x1p+0)]; + tensor valid_len_cast_fp16 = add(x = var_2495_cast_fp16, y = var_2512_to_fp16)[name = tensor("valid_len_cast_fp16")]; + tensor valid_mask_cast_fp16 = less(x = k_positions_1_promoted_to_fp16, y = valid_len_cast_fp16)[name = tensor("valid_mask_cast_fp16")]; + tensor causal_mask_cast_fp16 = less_equal(x = k_positions_1_promoted_to_fp16, y = var_2495_cast_fp16)[name = tensor("causal_mask_cast_fp16")]; + tensor attn_mask_21 = logical_and(x = valid_mask_cast_fp16, y = causal_mask_cast_fp16)[name = tensor("attn_mask_21")]; + tensor attn_mask_axes_0 = const()[name = tensor("attn_mask_axes_0"), val = tensor([1])]; + tensor attn_mask = expand_dims(axes = attn_mask_axes_0, x = attn_mask_21)[name = tensor("attn_mask")]; + tensor var_2530_transpose_x_0 = const()[name = tensor("op_2530_transpose_x_0"), val = tensor(false)]; + tensor var_2530_transpose_y_0 = const()[name = tensor("op_2530_transpose_y_0"), val = tensor(false)]; + tensor transpose_28_perm_0 = const()[name = tensor("transpose_28_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_29_perm_0 = const()[name = tensor("transpose_29_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_29 = transpose(perm = transpose_29_perm_0, x = keys_33_cast_fp16)[name = tensor("transpose_31")]; + tensor transpose_28 = transpose(perm = transpose_28_perm_0, x = q_33_cast_fp16)[name = tensor("transpose_32")]; + tensor var_2530_cast_fp16 = matmul(transpose_x = var_2530_transpose_x_0, transpose_y = var_2530_transpose_y_0, x = transpose_28, y = transpose_29)[name = tensor("op_2530_cast_fp16")]; + tensor var_2524_to_fp16 = const()[name = tensor("op_2524_to_fp16"), val = tensor([0x1p-3])]; + tensor attn_weights_31_cast_fp16 = mul(x = var_2530_cast_fp16, y = var_2524_to_fp16)[name = tensor("attn_weights_31_cast_fp16")]; + tensor var_2532 = logical_not(x = attn_mask)[name = tensor("op_2532")]; + tensor var_2533_to_fp16 = const()[name = tensor("op_2533_to_fp16"), val = tensor(-inf)]; + tensor attn_weights_33_cast_fp16 = select(a = var_2533_to_fp16, b = attn_weights_31_cast_fp16, cond = var_2532)[name = tensor("attn_weights_33_cast_fp16")]; + tensor var_2535 = const()[name = tensor("op_2535"), val = tensor(-1)]; + tensor attn_weights_cast_fp16 = softmax(axis = var_2535, x = attn_weights_33_cast_fp16)[name = tensor("attn_weights_cast_fp16")]; + tensor attn_output_transpose_x_0 = const()[name = tensor("attn_output_transpose_x_0"), val = tensor(false)]; + tensor attn_output_transpose_y_0 = const()[name = tensor("attn_output_transpose_y_0"), val = tensor(false)]; + tensor values_cast_fp16 = transpose(perm = var_2481, x = values_33_cast_fp16)[name = tensor("transpose_33")]; + tensor attn_output_cast_fp16 = matmul(transpose_x = attn_output_transpose_x_0, transpose_y = attn_output_transpose_y_0, x = attn_weights_cast_fp16, y = values_cast_fp16)[name = tensor("attn_output_cast_fp16")]; + tensor var_2543 = const()[name = tensor("op_2543"), val = tensor([0, 2, 1, 3])]; + tensor var_2546 = const()[name = tensor("op_2546"), val = tensor([1, 1, 1024])]; + tensor var_2544_cast_fp16 = transpose(perm = var_2543, x = attn_output_cast_fp16)[name = tensor("transpose_30")]; + tensor input_55_cast_fp16 = reshape(shape = var_2546, x = var_2544_cast_fp16)[name = tensor("input_55_cast_fp16")]; + tensor linear_22_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = attn5_out_proj_weight_cast_fp16, x = input_55_cast_fp16)[name = tensor("linear_22_cast_fp16")]; + tensor var_2552_to_fp16 = const()[name = tensor("op_2552_to_fp16"), val = tensor(0x1p+0)]; + tensor var_2553_cast_fp16 = add(x = position5_to_fp16, y = var_2552_to_fp16)[name = tensor("op_2553_cast_fp16")]; + tensor var_2553_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_2553_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor input_57_cast_fp16 = add(x = input_53_cast_fp16, y = linear_22_cast_fp16)[name = tensor("input_57_cast_fp16")]; + tensor input_59_axes_0 = const()[name = tensor("input_59_axes_0"), val = tensor([-1])]; + tensor norm5_2_weight_to_fp16 = const()[name = tensor("norm5_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(76798720)))]; + tensor norm5_2_bias_to_fp16 = const()[name = tensor("norm5_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(76800832)))]; + tensor var_2557_to_fp16 = const()[name = tensor("op_2557_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_59_cast_fp16 = layer_norm(axes = input_59_axes_0, beta = norm5_2_bias_to_fp16, epsilon = var_2557_to_fp16, gamma = norm5_2_weight_to_fp16, x = input_57_cast_fp16)[name = tensor("input_59_cast_fp16")]; + tensor linear_23_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = linear5_1_weight_cast_fp16, x = input_59_cast_fp16)[name = tensor("linear_23_cast_fp16")]; + tensor input_61_mode_0 = const()[name = tensor("input_61_mode_0"), val = tensor("EXACT")]; + tensor input_61_cast_fp16 = gelu(mode = input_61_mode_0, x = linear_23_cast_fp16)[name = tensor("input_61_cast_fp16")]; + tensor linear_24_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = linear5_2_weight_cast_fp16, x = input_61_cast_fp16)[name = tensor("linear_24_cast_fp16")]; + tensor input_63_cast_fp16 = add(x = input_57_cast_fp16, y = linear_24_cast_fp16)[name = tensor("input_63_cast_fp16")]; + tensor input_axes_0 = const()[name = tensor("input_axes_0"), val = tensor([-1])]; + tensor out_norm_weight_to_fp16 = const()[name = tensor("out_norm_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(76802944)))]; + tensor out_norm_bias_to_fp16 = const()[name = tensor("out_norm_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(76805056)))]; + tensor var_2574_to_fp16 = const()[name = tensor("op_2574_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_cast_fp16 = layer_norm(axes = input_axes_0, beta = out_norm_bias_to_fp16, epsilon = var_2574_to_fp16, gamma = out_norm_weight_to_fp16, x = input_63_cast_fp16)[name = tensor("input_cast_fp16")]; + tensor input_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("input_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor out_eos_weight_to_fp16 = const()[name = tensor("out_eos_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(76807168)))]; + tensor out_eos_bias_to_fp16 = const()[name = tensor("out_eos_bias_to_fp16"), val = tensor([-0x1.36p-2])]; + tensor linear_25_cast_fp16 = linear(bias = out_eos_bias_to_fp16, weight = out_eos_weight_to_fp16, x = input_cast_fp16)[name = tensor("linear_25_cast_fp16")]; + tensor linear_25_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("linear_25_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor var_2582 = cast(dtype = linear_25_cast_fp16_to_fp32_dtype_0, x = linear_25_cast_fp16)[name = tensor("cast_108")]; + tensor input = cast(dtype = input_cast_fp16_to_fp32_dtype_0, x = input_cast_fp16)[name = tensor("cast_109")]; + tensor var_2553 = cast(dtype = var_2553_cast_fp16_to_fp32_dtype_0, x = var_2553_cast_fp16)[name = tensor("cast_110")]; + tensor new_cache_internal_tensor_assign_2 = cast(dtype = reshape_59_cast_fp16_to_fp32_dtype_0, x = reshape_59_cast_fp16)[name = tensor("cast_111")]; + tensor var_2134 = cast(dtype = var_2134_cast_fp16_to_fp32_dtype_0, x = var_2134_cast_fp16)[name = tensor("cast_115")]; + tensor new_cache_9_internal_tensor_assign_2 = cast(dtype = reshape_49_cast_fp16_to_fp32_dtype_0, x = reshape_49_cast_fp16)[name = tensor("cast_116")]; + tensor var_1715 = cast(dtype = var_1715_cast_fp16_to_fp32_dtype_0, x = var_1715_cast_fp16)[name = tensor("cast_120")]; + tensor new_cache_7_internal_tensor_assign_2 = cast(dtype = reshape_39_cast_fp16_to_fp32_dtype_0, x = reshape_39_cast_fp16)[name = tensor("cast_121")]; + tensor var_1296 = cast(dtype = var_1296_cast_fp16_to_fp32_dtype_0, x = var_1296_cast_fp16)[name = tensor("cast_125")]; + tensor new_cache_5_internal_tensor_assign_2 = cast(dtype = reshape_29_cast_fp16_to_fp32_dtype_0, x = reshape_29_cast_fp16)[name = tensor("cast_126")]; + tensor var_877 = cast(dtype = var_877_cast_fp16_to_fp32_dtype_0, x = var_877_cast_fp16)[name = tensor("cast_130")]; + tensor new_cache_3_internal_tensor_assign_2 = cast(dtype = reshape_19_cast_fp16_to_fp32_dtype_0, x = reshape_19_cast_fp16)[name = tensor("cast_131")]; + tensor var_458 = cast(dtype = var_458_cast_fp16_to_fp32_dtype_0, x = var_458_cast_fp16)[name = tensor("cast_135")]; + tensor new_cache_1_internal_tensor_assign_2 = cast(dtype = reshape_9_cast_fp16_to_fp32_dtype_0, x = reshape_9_cast_fp16)[name = tensor("cast_136")]; + } -> (input, var_2582, new_cache_1_internal_tensor_assign_2, var_458, new_cache_3_internal_tensor_assign_2, var_877, new_cache_5_internal_tensor_assign_2, var_1296, new_cache_7_internal_tensor_assign_2, var_1715, new_cache_9_internal_tensor_assign_2, var_2134, new_cache_internal_tensor_assign_2, var_2553); +} \ No newline at end of file diff --git a/v2.1/english/flowlm_stepv2.mlmodelc/weights/weight.bin b/v2.1/english/flowlm_stepv2.mlmodelc/weights/weight.bin new file mode 100644 index 0000000000000000000000000000000000000000..c03c6e39219c157dacc06300721fe1d76948f470 --- /dev/null +++ b/v2.1/english/flowlm_stepv2.mlmodelc/weights/weight.bin @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:5b7a9bbbc6ba5093cdfb55fb149c0d816191161ad866de5733406ebd7b01fba6 +size 76809280 diff --git a/v2.1/english/flowlm_stepv2.mlpackage/Data/com.apple.CoreML/model.mlmodel b/v2.1/english/flowlm_stepv2.mlpackage/Data/com.apple.CoreML/model.mlmodel new file mode 100644 index 0000000000000000000000000000000000000000..ca13c249ee1a9a9f8c33d261dc4d9ca0889c404c --- /dev/null +++ b/v2.1/english/flowlm_stepv2.mlpackage/Data/com.apple.CoreML/model.mlmodel @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:f106a8b4ceeb7a5485926fecf3df4060c552c9d0f20b7bb75934f98ea143bb29 +size 212305 diff --git a/v2.1/english/flowlm_stepv2.mlpackage/Data/com.apple.CoreML/weights/weight.bin b/v2.1/english/flowlm_stepv2.mlpackage/Data/com.apple.CoreML/weights/weight.bin new file mode 100644 index 0000000000000000000000000000000000000000..c03c6e39219c157dacc06300721fe1d76948f470 --- /dev/null +++ b/v2.1/english/flowlm_stepv2.mlpackage/Data/com.apple.CoreML/weights/weight.bin @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:5b7a9bbbc6ba5093cdfb55fb149c0d816191161ad866de5733406ebd7b01fba6 +size 76809280 diff --git a/v2.1/english/flowlm_stepv2.mlpackage/Manifest.json b/v2.1/english/flowlm_stepv2.mlpackage/Manifest.json new file mode 100644 index 0000000000000000000000000000000000000000..5375724f53cb1151346ef2dc6b79634a6ee7b03f --- /dev/null +++ b/v2.1/english/flowlm_stepv2.mlpackage/Manifest.json @@ -0,0 +1,18 @@ +{ + "fileFormatVersion": "1.0.0", + "itemInfoEntries": { + "1BF0A666-8AA0-4CCF-AA97-B101C6CAD73D": { + "author": "com.apple.CoreML", + "description": "CoreML Model Weights", + "name": "weights", + "path": "com.apple.CoreML/weights" + }, + "89D2A4C6-CD65-48B9-A789-DCC7BAB634F2": { + "author": "com.apple.CoreML", + "description": "CoreML Model Specification", + "name": "model.mlmodel", + "path": "com.apple.CoreML/model.mlmodel" + } + }, + "rootModelIdentifier": "89D2A4C6-CD65-48B9-A789-DCC7BAB634F2" +} diff --git a/v2.1/english/manifest.json b/v2.1/english/manifest.json new file mode 100644 index 0000000000000000000000000000000000000000..fc7da98920126d1261a9c47517b1303db1599db7 --- /dev/null +++ b/v2.1/english/manifest.json @@ -0,0 +1,27 @@ +{ + "version": "2.1", + "base": "v2", + "kind": "optimization-reconvert", + "self_contained": true, + "note": "Same weights as v2. Re-converted for speed: fused flow decoder (100% ANE), one-shot cond prefill, fp16 flowlm. NOT a finetune.", + "new": [ + "flow_decoder_fused", + "cond_prefill", + "flowlm_step" + ], + "copied_from_v2": [ + "flowlm_stepv2.mlpackage", + "flowlm_stepv2.mlmodelc", + "mimi_decoder.mlpackage", + "mimi_decoder.mlmodelc", + "constants", + "constants_bin" + ], + "compute_units": { + "flow_decoder_fused": "all (ANE)", + "cond_prefill": "all (GPU)", + "flowlm_step": "all (GPU, fp16)", + "flowlm_stepv2": "cpuAndGpu (GPU, int8 \u2014 fastest flowlm)", + "mimi_decoder": "cpuOnly" + } +} diff --git a/v2.1/english/mimi_decoder.mlmodelc/analytics/coremldata.bin b/v2.1/english/mimi_decoder.mlmodelc/analytics/coremldata.bin new file mode 100644 index 0000000000000000000000000000000000000000..54f1906017f5d5124751a6c5d39e72cf0ccc2e64 --- /dev/null +++ b/v2.1/english/mimi_decoder.mlmodelc/analytics/coremldata.bin @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:be4199b44838e2fb199bb44b7649531d28fe847a1a6633fd9f8ca89576c5a7f8 +size 243 diff --git a/v2.1/english/mimi_decoder.mlmodelc/coremldata.bin b/v2.1/english/mimi_decoder.mlmodelc/coremldata.bin new file mode 100644 index 0000000000000000000000000000000000000000..de964cb7c980cf8d8ede358a615d7047fcc1ee73 --- /dev/null +++ b/v2.1/english/mimi_decoder.mlmodelc/coremldata.bin @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:9d8644e2caad1513153ac28f1ba8090fcd2facdbd5d2621bf8016186fa932bf8 +size 1777 diff --git a/v2.1/english/mimi_decoder.mlmodelc/metadata.json b/v2.1/english/mimi_decoder.mlmodelc/metadata.json new file mode 100644 index 0000000000000000000000000000000000000000..f66f9e7ff5be2ffbc44c8300f3501f67a77afaf6 --- /dev/null +++ b/v2.1/english/mimi_decoder.mlmodelc/metadata.json @@ -0,0 +1,573 @@ +[ + { + "metadataOutputVersion" : "3.0", + "storagePrecision" : "Float32", + "outputSchema" : [ + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1 × 1 × 1920)", + "shortDescription" : "", + "shape" : "[1, 1, 1920]", + "name" : "var_724", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1 × 512 × 16)", + "shortDescription" : "", + "shape" : "[1, 512, 16]", + "name" : "var_77", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 2 × 1 × 256 × 8 × 64)", + "shortDescription" : "", + "shape" : "[2, 1, 256, 8, 64]", + "name" : "var_210", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1)", + "shortDescription" : "", + "shape" : "[1]", + "name" : "var_741", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 2 × 1 × 256 × 8 × 64)", + "shortDescription" : "", + "shape" : "[2, 1, 256, 8, 64]", + "name" : "var_400", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1)", + "shortDescription" : "", + "shape" : "[1]", + "name" : "var_744", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1 × 512 × 6)", + "shortDescription" : "", + "shape" : "[1, 512, 6]", + "name" : "var_542", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1)", + "shortDescription" : "", + "shape" : "[1]", + "name" : "conv0_first", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1 × 256 × 6)", + "shortDescription" : "", + "shape" : "[1, 256, 6]", + "name" : "var_565", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1 × 256 × 2)", + "shortDescription" : "", + "shape" : "[1, 256, 2]", + "name" : "var_585", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1)", + "shortDescription" : "", + "shape" : "[1]", + "name" : "res0_conv0_first", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1 × 128 × 0)", + "shortDescription" : "", + "shape" : "[1, 128, 0]", + "name" : "res0_conv1_prev", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1)", + "shortDescription" : "", + "shape" : "[1]", + "name" : "res0_conv1_first", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1 × 128 × 5)", + "shortDescription" : "", + "shape" : "[1, 128, 5]", + "name" : "var_621", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1 × 128 × 2)", + "shortDescription" : "", + "shape" : "[1, 128, 2]", + "name" : "var_641", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1)", + "shortDescription" : "", + "shape" : "[1]", + "name" : "res1_conv0_first", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1 × 64 × 0)", + "shortDescription" : "", + "shape" : "[1, 64, 0]", + "name" : "res1_conv1_prev", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1)", + "shortDescription" : "", + "shape" : "[1]", + "name" : "res1_conv1_first", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1 × 64 × 4)", + "shortDescription" : "", + "shape" : "[1, 64, 4]", + "name" : "var_677", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1 × 64 × 2)", + "shortDescription" : "", + "shape" : "[1, 64, 2]", + "name" : "var_697", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1)", + "shortDescription" : "", + "shape" : "[1]", + "name" : "res2_conv0_first", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1 × 32 × 0)", + "shortDescription" : "", + "shape" : "[1, 32, 0]", + "name" : "res2_conv1_prev", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1)", + "shortDescription" : "", + "shape" : "[1]", + "name" : "res2_conv1_first", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1 × 64 × 2)", + "shortDescription" : "", + "shape" : "[1, 64, 2]", + "name" : "var_725", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1)", + "shortDescription" : "", + "shape" : "[1]", + "name" : "conv_final_first", + "type" : "MultiArray" + } + ], + "modelParameters" : [ + + ], + "specificationVersion" : 8, + "mlProgramOperationTypeHistogram" : { + "Ios17.notEqual" : 4, + "Tile" : 2, + "Ios17.lessEqual" : 2, + "Ios17.layerNorm" : 4, + "Stack" : 6, + "Ios17.greaterEqual" : 4, + "Ios17.scatterAlongAxis" : 4, + "Ios17.sub" : 17, + "Identity" : 11, + "Ios17.elu" : 10, + "Ios17.logicalNot" : 2, + "Ios17.sliceByIndex" : 35, + "Ios17.concat" : 9, + "Ios17.convTranspose" : 4, + "Ios17.floorDiv" : 4, + "Ios17.expandDims" : 7, + "Ios17.less" : 2, + "Ios17.mul" : 29, + "Ios17.transpose" : 10, + "Ios17.matmul" : 4, + "Ios16.gelu" : 2, + "Ios17.conv" : 9, + "Ios17.reshape" : 22, + "Split" : 2, + "Ios17.squeeze" : 6, + "Ios17.cast" : 6, + "Ios17.sin" : 2, + "Ios17.linear" : 8, + "Select" : 6, + "Ios16.softmax" : 2, + "Ios17.add" : 28, + "Ios17.cos" : 2, + "Ios17.logicalAnd" : 6 + }, + "computePrecision" : "Mixed (Float32, Int32)", + "isUpdatable" : "0", + "stateSchema" : [ + + ], + "availability" : { + "macOS" : "14.0", + "tvOS" : "17.0", + "visionOS" : "1.0", + "watchOS" : "10.0", + "iOS" : "17.0", + "macCatalyst" : "17.0" + }, + "modelType" : { + "name" : "MLModelType_mlProgram" + }, + "userDefinedMetadata" : { + "com.github.apple.coremltools.conversion_date" : "2026-04-24", + "com.github.apple.coremltools.source" : "torch==2.9.1", + "com.github.apple.coremltools.version" : "9.0", + "com.github.apple.coremltools.source_dialect" : "TorchScript" + }, + "inputSchema" : [ + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1 × 32)", + "shortDescription" : "", + "shape" : "[1, 32]", + "name" : "latent", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1 × 512 × 16)", + "shortDescription" : "", + "shape" : "[1, 512, 16]", + "name" : "upsample_partial", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 2 × 1 × 256 × 8 × 64)", + "shortDescription" : "", + "shape" : "[2, 1, 256, 8, 64]", + "name" : "attn0_cache", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1)", + "shortDescription" : "", + "shape" : "[1]", + "name" : "attn0_offset", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 2 × 1 × 256 × 8 × 64)", + "shortDescription" : "", + "shape" : "[2, 1, 256, 8, 64]", + "name" : "attn1_cache", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1)", + "shortDescription" : "", + "shape" : "[1]", + "name" : "attn1_offset", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1 × 512 × 6)", + "shortDescription" : "", + "shape" : "[1, 512, 6]", + "name" : "conv0_prev", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1)", + "shortDescription" : "", + "shape" : "[1]", + "name" : "conv0_first", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1 × 256 × 6)", + "shortDescription" : "", + "shape" : "[1, 256, 6]", + "name" : "convtr0_partial", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1 × 256 × 2)", + "shortDescription" : "", + "shape" : "[1, 256, 2]", + "name" : "res0_conv0_prev", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1)", + "shortDescription" : "", + "shape" : "[1]", + "name" : "res0_conv0_first", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1 × 128 × 0)", + "shortDescription" : "", + "shape" : "[1, 128, 0]", + "name" : "res0_conv1_prev", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1)", + "shortDescription" : "", + "shape" : "[1]", + "name" : "res0_conv1_first", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1 × 128 × 5)", + "shortDescription" : "", + "shape" : "[1, 128, 5]", + "name" : "convtr1_partial", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1 × 128 × 2)", + "shortDescription" : "", + "shape" : "[1, 128, 2]", + "name" : "res1_conv0_prev", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1)", + "shortDescription" : "", + "shape" : "[1]", + "name" : "res1_conv0_first", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1 × 64 × 0)", + "shortDescription" : "", + "shape" : "[1, 64, 0]", + "name" : "res1_conv1_prev", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1)", + "shortDescription" : "", + "shape" : "[1]", + "name" : "res1_conv1_first", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1 × 64 × 4)", + "shortDescription" : "", + "shape" : "[1, 64, 4]", + "name" : "convtr2_partial", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1 × 64 × 2)", + "shortDescription" : "", + "shape" : "[1, 64, 2]", + "name" : "res2_conv0_prev", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1)", + "shortDescription" : "", + "shape" : "[1]", + "name" : "res2_conv0_first", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1 × 32 × 0)", + "shortDescription" : "", + "shape" : "[1, 32, 0]", + "name" : "res2_conv1_prev", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1)", + "shortDescription" : "", + "shape" : "[1]", + "name" : "res2_conv1_first", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1 × 64 × 2)", + "shortDescription" : "", + "shape" : "[1, 64, 2]", + "name" : "conv_final_prev", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1)", + "shortDescription" : "", + "shape" : "[1]", + "name" : "conv_final_first", + "type" : "MultiArray" + } + ], + "generatedClassName" : "mimi_decoder", + "method" : "predict" + } +] \ No newline at end of file diff --git a/v2.1/english/mimi_decoder.mlmodelc/model.mil b/v2.1/english/mimi_decoder.mlmodelc/model.mil new file mode 100644 index 0000000000000000000000000000000000000000..69ef039645ac6b4e76218d04f8327a3576d285bc --- /dev/null +++ b/v2.1/english/mimi_decoder.mlmodelc/model.mil @@ -0,0 +1,646 @@ +program(1.0) +[buildInfo = dict, tensor>({{"coremlc-component-MIL", "3520.4.1"}, {"coremlc-version", "3520.5.1"}, {"coremltools-component-torch", "2.9.1"}, {"coremltools-source-dialect", "TorchScript"}, {"coremltools-version", "9.0"}})] +{ + func main(tensor attn0_cache, tensor attn0_offset, tensor attn1_cache, tensor attn1_offset, tensor conv0_first, tensor conv0_prev, tensor conv_final_first, tensor conv_final_prev, tensor convtr0_partial, tensor convtr1_partial, tensor convtr2_partial, tensor latent, tensor res0_conv0_first, tensor res0_conv0_prev, tensor res0_conv1_first, tensor res0_conv1_prev, tensor res1_conv0_first, tensor res1_conv0_prev, tensor res1_conv1_first, tensor res1_conv1_prev, tensor res2_conv0_first, tensor res2_conv0_prev, tensor res2_conv1_first, tensor res2_conv1_prev, tensor upsample_partial) { + tensor emb_mean = const()[name = tensor("emb_mean"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(64)))]; + tensor emb_std = const()[name = tensor("emb_std"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(256)))]; + tensor mimi_quantizer_output_proj_weight = const()[name = tensor("mimi_quantizer_output_proj_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(448)))]; + tensor mimi_upsample_convtr_convtr_weight = const()[name = tensor("mimi_upsample_convtr_convtr_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(66048)))]; + tensor mimi_decoder_transformer_transformer_layers_0_norm1_bias = const()[name = tensor("mimi_decoder_transformer_transformer_layers_0_norm1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(131648)))]; + tensor mimi_decoder_transformer_transformer_layers_0_norm1_weight = const()[name = tensor("mimi_decoder_transformer_transformer_layers_0_norm1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(133760)))]; + tensor mimi_decoder_transformer_transformer_layers_0_self_attn_in_proj_weight = const()[name = tensor("mimi_decoder_transformer_transformer_layers_0_self_attn_in_proj_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(135872)))]; + tensor mimi_decoder_transformer_transformer_layers_0_self_attn_out_proj_weight = const()[name = tensor("mimi_decoder_transformer_transformer_layers_0_self_attn_out_proj_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(3281664)))]; + tensor mimi_decoder_transformer_transformer_layers_0_layer_scale_1_scale = const()[name = tensor("mimi_decoder_transformer_transformer_layers_0_layer_scale_1_scale"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(4330304)))]; + tensor mimi_decoder_transformer_transformer_layers_0_norm2_bias = const()[name = tensor("mimi_decoder_transformer_transformer_layers_0_norm2_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(4332416)))]; + tensor mimi_decoder_transformer_transformer_layers_0_norm2_weight = const()[name = tensor("mimi_decoder_transformer_transformer_layers_0_norm2_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(4334528)))]; + tensor mimi_decoder_transformer_transformer_layers_0_linear1_weight = const()[name = tensor("mimi_decoder_transformer_transformer_layers_0_linear1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(4336640)))]; + tensor mimi_decoder_transformer_transformer_layers_0_linear2_weight = const()[name = tensor("mimi_decoder_transformer_transformer_layers_0_linear2_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8531008)))]; + tensor mimi_decoder_transformer_transformer_layers_0_layer_scale_2_scale = const()[name = tensor("mimi_decoder_transformer_transformer_layers_0_layer_scale_2_scale"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(12725376)))]; + tensor mimi_decoder_transformer_transformer_layers_1_norm1_bias = const()[name = tensor("mimi_decoder_transformer_transformer_layers_1_norm1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(12727488)))]; + tensor mimi_decoder_transformer_transformer_layers_1_norm1_weight = const()[name = tensor("mimi_decoder_transformer_transformer_layers_1_norm1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(12729600)))]; + tensor mimi_decoder_transformer_transformer_layers_1_self_attn_in_proj_weight = const()[name = tensor("mimi_decoder_transformer_transformer_layers_1_self_attn_in_proj_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(12731712)))]; + tensor mimi_decoder_transformer_transformer_layers_1_self_attn_out_proj_weight = const()[name = tensor("mimi_decoder_transformer_transformer_layers_1_self_attn_out_proj_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15877504)))]; + tensor mimi_decoder_transformer_transformer_layers_1_layer_scale_1_scale = const()[name = tensor("mimi_decoder_transformer_transformer_layers_1_layer_scale_1_scale"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(16926144)))]; + tensor mimi_decoder_transformer_transformer_layers_1_norm2_bias = const()[name = tensor("mimi_decoder_transformer_transformer_layers_1_norm2_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(16928256)))]; + tensor mimi_decoder_transformer_transformer_layers_1_norm2_weight = const()[name = tensor("mimi_decoder_transformer_transformer_layers_1_norm2_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(16930368)))]; + tensor mimi_decoder_transformer_transformer_layers_1_linear1_weight = const()[name = tensor("mimi_decoder_transformer_transformer_layers_1_linear1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(16932480)))]; + tensor mimi_decoder_transformer_transformer_layers_1_linear2_weight = const()[name = tensor("mimi_decoder_transformer_transformer_layers_1_linear2_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21126848)))]; + tensor mimi_decoder_transformer_transformer_layers_1_layer_scale_2_scale = const()[name = tensor("mimi_decoder_transformer_transformer_layers_1_layer_scale_2_scale"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(25321216)))]; + tensor mimi_decoder_model_0_conv_bias = const()[name = tensor("mimi_decoder_model_0_conv_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(25323328)))]; + tensor mimi_decoder_model_0_conv_weight = const()[name = tensor("mimi_decoder_model_0_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(25325440)))]; + tensor mimi_decoder_model_2_convtr_bias = const()[name = tensor("mimi_decoder_model_2_convtr_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(32665536)))]; + tensor mimi_decoder_model_2_convtr_weight = const()[name = tensor("mimi_decoder_model_2_convtr_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(32666624)))]; + tensor mimi_decoder_model_3_block_1_conv_bias = const()[name = tensor("mimi_decoder_model_3_block_1_conv_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(38958144)))]; + tensor mimi_decoder_model_3_block_1_conv_weight = const()[name = tensor("mimi_decoder_model_3_block_1_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(38958720)))]; + tensor mimi_decoder_model_3_block_3_conv_bias = const()[name = tensor("mimi_decoder_model_3_block_3_conv_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(39352000)))]; + tensor mimi_decoder_model_3_block_3_conv_weight = const()[name = tensor("mimi_decoder_model_3_block_3_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(39353088)))]; + tensor mimi_decoder_model_5_convtr_bias = const()[name = tensor("mimi_decoder_model_5_convtr_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(39484224)))]; + tensor mimi_decoder_model_5_convtr_weight = const()[name = tensor("mimi_decoder_model_5_convtr_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(39484800)))]; + tensor mimi_decoder_model_6_block_1_conv_bias = const()[name = tensor("mimi_decoder_model_6_block_1_conv_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(40795584)))]; + tensor mimi_decoder_model_6_block_1_conv_weight = const()[name = tensor("mimi_decoder_model_6_block_1_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(40795904)))]; + tensor mimi_decoder_model_6_block_3_conv_bias = const()[name = tensor("mimi_decoder_model_6_block_3_conv_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(40894272)))]; + tensor mimi_decoder_model_6_block_3_conv_weight = const()[name = tensor("mimi_decoder_model_6_block_3_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(40894848)))]; + tensor mimi_decoder_model_8_convtr_bias = const()[name = tensor("mimi_decoder_model_8_convtr_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(40927680)))]; + tensor mimi_decoder_model_8_convtr_weight = const()[name = tensor("mimi_decoder_model_8_convtr_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(40928000)))]; + tensor mimi_decoder_model_9_block_1_conv_bias = const()[name = tensor("mimi_decoder_model_9_block_1_conv_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(41190208)))]; + tensor mimi_decoder_model_9_block_1_conv_weight = const()[name = tensor("mimi_decoder_model_9_block_1_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(41190400)))]; + tensor mimi_decoder_model_9_block_3_conv_bias = const()[name = tensor("mimi_decoder_model_9_block_3_conv_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(41215040)))]; + tensor mimi_decoder_model_9_block_3_conv_weight = const()[name = tensor("mimi_decoder_model_9_block_3_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(41215360)))]; + tensor mimi_decoder_model_11_conv_bias = const()[name = tensor("mimi_decoder_model_11_conv_bias"), val = tensor([0x1.38p-15])]; + tensor mimi_decoder_model_11_conv_weight = const()[name = tensor("mimi_decoder_model_11_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(41223616)))]; + tensor var_38 = mul(x = latent, y = emb_std)[name = tensor("op_38")]; + tensor denorm = add(x = var_38, y = emb_mean)[name = tensor("denorm")]; + tensor input_1_axes_0 = const()[name = tensor("input_1_axes_0"), val = tensor([-1])]; + tensor input_1 = expand_dims(axes = input_1_axes_0, x = denorm)[name = tensor("input_1")]; + tensor x_1_pad_type_0 = const()[name = tensor("x_1_pad_type_0"), val = tensor("valid")]; + tensor x_1_strides_0 = const()[name = tensor("x_1_strides_0"), val = tensor([1])]; + tensor x_1_pad_0 = const()[name = tensor("x_1_pad_0"), val = tensor([0, 0])]; + tensor x_1_dilations_0 = const()[name = tensor("x_1_dilations_0"), val = tensor([1])]; + tensor x_1_groups_0 = const()[name = tensor("x_1_groups_0"), val = tensor(1)]; + tensor x_1 = conv(dilations = x_1_dilations_0, groups = x_1_groups_0, pad = x_1_pad_0, pad_type = x_1_pad_type_0, strides = x_1_strides_0, weight = mimi_quantizer_output_proj_weight, x = input_1)[name = tensor("x_1")]; + tensor var_62 = const()[name = tensor("op_62"), val = tensor(-1)]; + tensor y_1_pad_type_0 = const()[name = tensor("y_1_pad_type_0"), val = tensor("valid")]; + tensor y_1_strides_0 = const()[name = tensor("y_1_strides_0"), val = tensor([16])]; + tensor y_1_groups_0 = const()[name = tensor("y_1_groups_0"), val = tensor(512)]; + tensor y_1_pad_0 = const()[name = tensor("y_1_pad_0"), val = tensor([0, 0])]; + tensor y_1_dilations_0 = const()[name = tensor("y_1_dilations_0"), val = tensor([1])]; + tensor y_1_has_output_shape_output_shape_0 = const()[name = tensor("y_1_has_output_shape_output_shape_0"), val = tensor([1, 512, 32])]; + tensor y_1_has_output_shape = conv_transpose(dilations = y_1_dilations_0, groups = y_1_groups_0, output_shape = y_1_has_output_shape_output_shape_0, pad = y_1_pad_0, pad_type = y_1_pad_type_0, strides = y_1_strides_0, weight = mimi_upsample_convtr_convtr_weight, x = x_1)[name = tensor("y_1_has_output_shape")]; + tensor var_72_begin_0 = const()[name = tensor("op_72_begin_0"), val = tensor([0, 0, 0])]; + tensor var_72_end_0 = const()[name = tensor("op_72_end_0"), val = tensor([1, 512, 16])]; + tensor var_72_end_mask_0 = const()[name = tensor("op_72_end_mask_0"), val = tensor([true, true, false])]; + tensor var_72 = slice_by_index(begin = var_72_begin_0, end = var_72_end_0, end_mask = var_72_end_mask_0, x = y_1_has_output_shape)[name = tensor("op_72")]; + tensor var_73 = add(x = var_72, y = upsample_partial)[name = tensor("op_73")]; + tensor var_74_begin_0 = const()[name = tensor("op_74_begin_0"), val = tensor([0, 0, 16])]; + tensor var_74_end_0 = const()[name = tensor("op_74_end_0"), val = tensor([1, 512, 32])]; + tensor var_74_end_mask_0 = const()[name = tensor("op_74_end_mask_0"), val = tensor([true, true, true])]; + tensor var_74 = slice_by_index(begin = var_74_begin_0, end = var_74_end_0, end_mask = var_74_end_mask_0, x = y_1_has_output_shape)[name = tensor("op_74")]; + tensor y_3_interleave_0 = const()[name = tensor("y_3_interleave_0"), val = tensor(false)]; + tensor y_3 = concat(axis = var_62, interleave = y_3_interleave_0, values = (var_73, var_74))[name = tensor("y_3")]; + tensor var_77_begin_0 = const()[name = tensor("op_77_begin_0"), val = tensor([0, 0, 16])]; + tensor var_77_end_0 = const()[name = tensor("op_77_end_0"), val = tensor([1, 512, 32])]; + tensor var_77_end_mask_0 = const()[name = tensor("op_77_end_mask_0"), val = tensor([true, true, true])]; + tensor var_77 = slice_by_index(begin = var_77_begin_0, end = var_77_end_0, end_mask = var_77_end_mask_0, x = y_3)[name = tensor("op_77")]; + tensor x_3_begin_0 = const()[name = tensor("x_3_begin_0"), val = tensor([0, 0, 0])]; + tensor x_3_end_0 = const()[name = tensor("x_3_end_0"), val = tensor([1, 512, 16])]; + tensor x_3_end_mask_0 = const()[name = tensor("x_3_end_mask_0"), val = tensor([true, true, false])]; + tensor x_3 = slice_by_index(begin = x_3_begin_0, end = x_3_end_0, end_mask = x_3_end_mask_0, x = y_3)[name = tensor("x_3")]; + tensor var_86 = const()[name = tensor("op_86"), val = tensor(0)]; + tensor var_91 = const()[name = tensor("op_91"), val = tensor(-1)]; + tensor var_100 = const()[name = tensor("op_100"), val = tensor(-0x1.ff933cp+127)]; + tensor var_102 = const()[name = tensor("op_102"), val = tensor(0x1.4f8b58p-17)]; + tensor input_3_perm_0 = const()[name = tensor("input_3_perm_0"), val = tensor([0, 2, 1])]; + tensor query_1_axes_0 = const()[name = tensor("query_1_axes_0"), val = tensor([-1])]; + tensor input_3 = transpose(perm = input_3_perm_0, x = x_3)[name = tensor("transpose_19")]; + tensor query_1 = layer_norm(axes = query_1_axes_0, beta = mimi_decoder_transformer_transformer_layers_0_norm1_bias, epsilon = var_102, gamma = mimi_decoder_transformer_transformer_layers_0_norm1_weight, x = input_3)[name = tensor("query_1")]; + tensor linear_0_bias_0 = const()[name = tensor("linear_0_bias_0"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(41224448)))]; + tensor projected_1 = linear(bias = linear_0_bias_0, weight = mimi_decoder_transformer_transformer_layers_0_self_attn_in_proj_weight, x = query_1)[name = tensor("linear_0")]; + tensor var_130 = const()[name = tensor("op_130"), val = tensor([1, 16, 3, 8, 64])]; + tensor packed_1 = reshape(shape = var_130, x = projected_1)[name = tensor("packed_1")]; + tensor var_132_split_sizes_0 = const()[name = tensor("op_132_split_sizes_0"), val = tensor([1, 1, 1])]; + tensor var_132_axis_0 = const()[name = tensor("op_132_axis_0"), val = tensor(2)]; + tensor var_132_0, tensor var_132_1, tensor var_132_2 = split(axis = var_132_axis_0, split_sizes = var_132_split_sizes_0, x = packed_1)[name = tensor("op_132")]; + tensor squeeze_0_axes_0 = const()[name = tensor("squeeze_0_axes_0"), val = tensor([2])]; + tensor squeeze_0 = squeeze(axes = squeeze_0_axes_0, x = var_132_0)[name = tensor("squeeze_0")]; + tensor squeeze_1_axes_0 = const()[name = tensor("squeeze_1_axes_0"), val = tensor([2])]; + tensor squeeze_1 = squeeze(axes = squeeze_1_axes_0, x = var_132_1)[name = tensor("squeeze_1")]; + tensor squeeze_2_axes_0 = const()[name = tensor("squeeze_2_axes_0"), val = tensor([2])]; + tensor squeeze_2 = squeeze(axes = squeeze_2_axes_0, x = var_132_2)[name = tensor("squeeze_2")]; + tensor offset_3_begin_0 = const()[name = tensor("offset_3_begin_0"), val = tensor([0])]; + tensor offset_3_end_0 = const()[name = tensor("offset_3_end_0"), val = tensor([1])]; + tensor offset_3_end_mask_0 = const()[name = tensor("offset_3_end_mask_0"), val = tensor([false])]; + tensor offset_3_squeeze_mask_0 = const()[name = tensor("offset_3_squeeze_mask_0"), val = tensor([true])]; + tensor offset_3 = slice_by_index(begin = offset_3_begin_0, end = offset_3_end_0, end_mask = offset_3_end_mask_0, squeeze_mask = offset_3_squeeze_mask_0, x = attn0_offset)[name = tensor("offset_3")]; + tensor freqs_1 = const()[name = tensor("freqs_1"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(41230656)))]; + tensor ts_1_promoted = const()[name = tensor("ts_1_promoted"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(41230848)))]; + tensor ts_3 = add(x = ts_1_promoted, y = offset_3)[name = tensor("ts_3")]; + tensor var_148 = const()[name = tensor("op_148"), val = tensor([-1, 1, 1])]; + tensor ts_5 = reshape(shape = var_148, x = ts_3)[name = tensor("ts_5")]; + tensor var_150 = const()[name = tensor("op_150"), val = tensor([1, 16, 8, 32, 2])]; + tensor q_3 = reshape(shape = var_150, x = squeeze_0)[name = tensor("q_3")]; + tensor var_152 = const()[name = tensor("op_152"), val = tensor([1, 16, 8, 32, 2])]; + tensor k_3 = reshape(shape = var_152, x = squeeze_1)[name = tensor("k_3")]; + tensor var_154_begin_0 = const()[name = tensor("op_154_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_154_end_0 = const()[name = tensor("op_154_end_0"), val = tensor([1, 16, 8, 32, 1])]; + tensor var_154_end_mask_0 = const()[name = tensor("op_154_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_154_squeeze_mask_0 = const()[name = tensor("op_154_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_154 = slice_by_index(begin = var_154_begin_0, end = var_154_end_0, end_mask = var_154_end_mask_0, squeeze_mask = var_154_squeeze_mask_0, x = q_3)[name = tensor("op_154")]; + tensor var_156_begin_0 = const()[name = tensor("op_156_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_156_end_0 = const()[name = tensor("op_156_end_0"), val = tensor([1, 16, 8, 32, 2])]; + tensor var_156_end_mask_0 = const()[name = tensor("op_156_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_156_squeeze_mask_0 = const()[name = tensor("op_156_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_156 = slice_by_index(begin = var_156_begin_0, end = var_156_end_0, end_mask = var_156_end_mask_0, squeeze_mask = var_156_squeeze_mask_0, x = q_3)[name = tensor("op_156")]; + tensor var_158_begin_0 = const()[name = tensor("op_158_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_158_end_0 = const()[name = tensor("op_158_end_0"), val = tensor([1, 16, 8, 32, 1])]; + tensor var_158_end_mask_0 = const()[name = tensor("op_158_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_158_squeeze_mask_0 = const()[name = tensor("op_158_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_158 = slice_by_index(begin = var_158_begin_0, end = var_158_end_0, end_mask = var_158_end_mask_0, squeeze_mask = var_158_squeeze_mask_0, x = k_3)[name = tensor("op_158")]; + tensor var_160_begin_0 = const()[name = tensor("op_160_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_160_end_0 = const()[name = tensor("op_160_end_0"), val = tensor([1, 16, 8, 32, 2])]; + tensor var_160_end_mask_0 = const()[name = tensor("op_160_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_160_squeeze_mask_0 = const()[name = tensor("op_160_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_160 = slice_by_index(begin = var_160_begin_0, end = var_160_end_0, end_mask = var_160_end_mask_0, squeeze_mask = var_160_squeeze_mask_0, x = k_3)[name = tensor("op_160")]; + tensor var_162 = mul(x = freqs_1, y = ts_5)[name = tensor("op_162")]; + tensor rotr_1 = cos(x = var_162)[name = tensor("rotr_1")]; + tensor roti_1 = sin(x = var_162)[name = tensor("roti_1")]; + tensor var_166 = mul(x = var_154, y = rotr_1)[name = tensor("op_166")]; + tensor var_167 = mul(x = var_156, y = roti_1)[name = tensor("op_167")]; + tensor qor_1 = sub(x = var_166, y = var_167)[name = tensor("qor_1")]; + tensor var_169 = mul(x = var_154, y = roti_1)[name = tensor("op_169")]; + tensor var_170 = mul(x = var_156, y = rotr_1)[name = tensor("op_170")]; + tensor qoi_1 = add(x = var_169, y = var_170)[name = tensor("qoi_1")]; + tensor var_172 = mul(x = var_158, y = rotr_1)[name = tensor("op_172")]; + tensor var_173 = mul(x = var_160, y = roti_1)[name = tensor("op_173")]; + tensor kor_1 = sub(x = var_172, y = var_173)[name = tensor("kor_1")]; + tensor var_175 = mul(x = var_158, y = roti_1)[name = tensor("op_175")]; + tensor var_176 = mul(x = var_160, y = rotr_1)[name = tensor("op_176")]; + tensor koi_1 = add(x = var_175, y = var_176)[name = tensor("koi_1")]; + tensor qo_1_axis_0 = const()[name = tensor("qo_1_axis_0"), val = tensor(-1)]; + tensor qo_1 = stack(axis = qo_1_axis_0, values = (qor_1, qoi_1))[name = tensor("qo_1")]; + tensor ko_1_axis_0 = const()[name = tensor("ko_1_axis_0"), val = tensor(-1)]; + tensor ko_1 = stack(axis = ko_1_axis_0, values = (kor_1, koi_1))[name = tensor("ko_1")]; + tensor var_186 = const()[name = tensor("op_186"), val = tensor([1, 16, 8, 64])]; + tensor q_5 = reshape(shape = var_186, x = qo_1)[name = tensor("q_5")]; + tensor var_188 = const()[name = tensor("op_188"), val = tensor([1, 16, 8, 64])]; + tensor k_5 = reshape(shape = var_188, x = ko_1)[name = tensor("k_5")]; + tensor capacity_1 = const()[name = tensor("capacity_1"), val = tensor([256])]; + tensor var_193_dtype_0 = const()[name = tensor("op_193_dtype_0"), val = tensor("int32")]; + tensor var_194 = const()[name = tensor("op_194"), val = tensor([1, 1])]; + tensor var_193 = cast(dtype = var_193_dtype_0, x = attn0_offset)[name = tensor("cast_49")]; + tensor write_base_1 = reshape(shape = var_194, x = var_193)[name = tensor("write_base_1")]; + tensor write_range_1 = const()[name = tensor("write_range_1"), val = tensor([[0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15]])]; + tensor abs_idx_1 = add(x = write_base_1, y = write_range_1)[name = tensor("abs_idx_1")]; + tensor wrapped_1_div = floor_div(x = abs_idx_1, y = capacity_1)[name = tensor("wrapped_1_div")]; + tensor wrapped_1_div_scaled = mul(x = wrapped_1_div, y = capacity_1)[name = tensor("wrapped_1_div_scaled")]; + tensor wrapped_1 = sub(x = abs_idx_1, y = wrapped_1_div_scaled)[name = tensor("wrapped_1")]; + tensor var_201 = const()[name = tensor("op_201"), val = tensor([1, 16, 1, 1])]; + tensor var_202 = reshape(shape = var_201, x = wrapped_1)[name = tensor("op_202")]; + tensor write_indexes_1_reps_0 = const()[name = tensor("write_indexes_1_reps_0"), val = tensor([1, 1, 8, 64])]; + tensor write_indexes_1 = tile(reps = write_indexes_1_reps_0, x = var_202)[name = tensor("write_indexes_1")]; + tensor var_205_begin_0 = const()[name = tensor("op_205_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_205_end_0 = const()[name = tensor("op_205_end_0"), val = tensor([1, 1, 256, 8, 64])]; + tensor var_205_end_mask_0 = const()[name = tensor("op_205_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_205_squeeze_mask_0 = const()[name = tensor("op_205_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor var_205 = slice_by_index(begin = var_205_begin_0, end = var_205_end_0, end_mask = var_205_end_mask_0, squeeze_mask = var_205_squeeze_mask_0, x = attn0_cache)[name = tensor("op_205")]; + tensor new_k_cache_1_axis_0 = const()[name = tensor("new_k_cache_1_axis_0"), val = tensor(1)]; + tensor new_k_cache_1_mode_0 = const()[name = tensor("new_k_cache_1_mode_0"), val = tensor("update")]; + tensor new_k_cache_1_validate_indices_0 = const()[name = tensor("new_k_cache_1_validate_indices_0"), val = tensor(false)]; + tensor new_k_cache_1 = scatter_along_axis(axis = new_k_cache_1_axis_0, data = var_205, indices = write_indexes_1, mode = new_k_cache_1_mode_0, updates = k_5, validate_indices = new_k_cache_1_validate_indices_0)[name = tensor("new_k_cache_1")]; + tensor var_207_begin_0 = const()[name = tensor("op_207_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor var_207_end_0 = const()[name = tensor("op_207_end_0"), val = tensor([2, 1, 256, 8, 64])]; + tensor var_207_end_mask_0 = const()[name = tensor("op_207_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_207_squeeze_mask_0 = const()[name = tensor("op_207_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor var_207 = slice_by_index(begin = var_207_begin_0, end = var_207_end_0, end_mask = var_207_end_mask_0, squeeze_mask = var_207_squeeze_mask_0, x = attn0_cache)[name = tensor("op_207")]; + tensor new_v_cache_1_axis_0 = const()[name = tensor("new_v_cache_1_axis_0"), val = tensor(1)]; + tensor new_v_cache_1_mode_0 = const()[name = tensor("new_v_cache_1_mode_0"), val = tensor("update")]; + tensor new_v_cache_1_validate_indices_0 = const()[name = tensor("new_v_cache_1_validate_indices_0"), val = tensor(false)]; + tensor new_v_cache_1 = scatter_along_axis(axis = new_v_cache_1_axis_0, data = var_207, indices = write_indexes_1, mode = new_v_cache_1_mode_0, updates = squeeze_2, validate_indices = new_v_cache_1_validate_indices_0)[name = tensor("new_v_cache_1")]; + tensor var_210_axis_0 = const()[name = tensor("op_210_axis_0"), val = tensor(0)]; + tensor var_210 = stack(axis = var_210_axis_0, values = (new_k_cache_1, new_v_cache_1))[name = tensor("op_210")]; + tensor var_211 = not_equal(x = new_k_cache_1, y = new_k_cache_1)[name = tensor("op_211")]; + tensor var_212 = const()[name = tensor("op_212"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(41230976)))]; + tensor new_k_cache_3 = select(a = var_212, b = new_k_cache_1, cond = var_211)[name = tensor("new_k_cache_3")]; + tensor var_214 = not_equal(x = new_v_cache_1, y = new_v_cache_1)[name = tensor("op_214")]; + tensor new_v_cache_3 = select(a = var_212, b = new_v_cache_1, cond = var_214)[name = tensor("new_v_cache_3")]; + tensor var_219 = const()[name = tensor("op_219"), val = tensor([0, 2, 1, 3])]; + tensor var_221 = const()[name = tensor("op_221"), val = tensor([1, 1])]; + tensor var_222 = reshape(shape = var_221, x = attn0_offset)[name = tensor("op_222")]; + tensor var_224_promoted = const()[name = tensor("op_224_promoted"), val = tensor([0x1.ep+3])]; + tensor var_225 = add(x = var_222, y = var_224_promoted)[name = tensor("op_225")]; + tensor last_pos_1_dtype_0 = const()[name = tensor("last_pos_1_dtype_0"), val = tensor("int32")]; + tensor slot_idx_1 = const()[name = tensor("slot_idx_1"), val = tensor([[0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19, 20, 21, 22, 23, 24, 25, 26, 27, 28, 29, 30, 31, 32, 33, 34, 35, 36, 37, 38, 39, 40, 41, 42, 43, 44, 45, 46, 47, 48, 49, 50, 51, 52, 53, 54, 55, 56, 57, 58, 59, 60, 61, 62, 63, 64, 65, 66, 67, 68, 69, 70, 71, 72, 73, 74, 75, 76, 77, 78, 79, 80, 81, 82, 83, 84, 85, 86, 87, 88, 89, 90, 91, 92, 93, 94, 95, 96, 97, 98, 99, 100, 101, 102, 103, 104, 105, 106, 107, 108, 109, 110, 111, 112, 113, 114, 115, 116, 117, 118, 119, 120, 121, 122, 123, 124, 125, 126, 127, 128, 129, 130, 131, 132, 133, 134, 135, 136, 137, 138, 139, 140, 141, 142, 143, 144, 145, 146, 147, 148, 149, 150, 151, 152, 153, 154, 155, 156, 157, 158, 159, 160, 161, 162, 163, 164, 165, 166, 167, 168, 169, 170, 171, 172, 173, 174, 175, 176, 177, 178, 179, 180, 181, 182, 183, 184, 185, 186, 187, 188, 189, 190, 191, 192, 193, 194, 195, 196, 197, 198, 199, 200, 201, 202, 203, 204, 205, 206, 207, 208, 209, 210, 211, 212, 213, 214, 215, 216, 217, 218, 219, 220, 221, 222, 223, 224, 225, 226, 227, 228, 229, 230, 231, 232, 233, 234, 235, 236, 237, 238, 239, 240, 241, 242, 243, 244, 245, 246, 247, 248, 249, 250, 251, 252, 253, 254, 255]])]; + tensor last_pos_1 = cast(dtype = last_pos_1_dtype_0, x = var_225)[name = tensor("cast_48")]; + tensor diff_1 = sub(x = last_pos_1, y = slot_idx_1)[name = tensor("diff_1")]; + tensor var_231_div = floor_div(x = diff_1, y = capacity_1)[name = tensor("op_231_div")]; + tensor var_231_div_scaled = mul(x = var_231_div, y = capacity_1)[name = tensor("op_231_div_scaled")]; + tensor var_231 = sub(x = diff_1, y = var_231_div_scaled)[name = tensor("op_231")]; + tensor pos_k_1 = sub(x = last_pos_1, y = var_231)[name = tensor("pos_k_1")]; + tensor var_237_promoted = const()[name = tensor("op_237_promoted"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(41755328)))]; + tensor pos_q_1 = add(x = var_222, y = var_237_promoted)[name = tensor("pos_q_1")]; + tensor var_241_axes_0 = const()[name = tensor("op_241_axes_0"), val = tensor([2])]; + tensor var_241 = expand_dims(axes = var_241_axes_0, x = pos_q_1)[name = tensor("op_241")]; + tensor var_243_axes_0 = const()[name = tensor("op_243_axes_0"), val = tensor([1])]; + tensor var_243 = expand_dims(axes = var_243_axes_0, x = pos_k_1)[name = tensor("op_243")]; + tensor var_244_promoted_dtype_0 = const()[name = tensor("op_244_promoted_dtype_0"), val = tensor("fp32")]; + tensor var_244_promoted = cast(dtype = var_244_promoted_dtype_0, x = var_243)[name = tensor("cast_47")]; + tensor delta_1 = sub(x = var_241, y = var_244_promoted)[name = tensor("delta_1")]; + tensor valid_1 = greater_equal(x = var_243, y = var_86)[name = tensor("valid_1")]; + tensor var_253 = const()[name = tensor("op_253"), val = tensor([1, 1, 1])]; + tensor var_254 = reshape(shape = var_253, x = attn0_offset)[name = tensor("op_254")]; + tensor var_256_promoted = const()[name = tensor("op_256_promoted"), val = tensor([0x1.ep+3])]; + tensor var_257 = add(x = var_254, y = var_256_promoted)[name = tensor("op_257")]; + tensor var_258 = less_equal(x = var_244_promoted, y = var_257)[name = tensor("op_258")]; + tensor valid_3 = logical_and(x = valid_1, y = var_258)[name = tensor("valid_3")]; + tensor var_86_promoted = const()[name = tensor("op_86_promoted"), val = tensor(0x0p+0)]; + tensor var_260 = greater_equal(x = delta_1, y = var_86_promoted)[name = tensor("op_260")]; + tensor attn_mask_1 = logical_and(x = valid_3, y = var_260)[name = tensor("attn_mask_1")]; + tensor var_98_promoted = const()[name = tensor("op_98_promoted"), val = tensor(0x1.f4p+7)]; + tensor var_262 = less(x = delta_1, y = var_98_promoted)[name = tensor("op_262")]; + tensor attn_mask_3 = logical_and(x = attn_mask_1, y = var_262)[name = tensor("attn_mask_3")]; + tensor attn_mask_5_axes_0 = const()[name = tensor("attn_mask_5_axes_0"), val = tensor([1])]; + tensor attn_mask_5 = expand_dims(axes = attn_mask_5_axes_0, x = attn_mask_3)[name = tensor("attn_mask_5")]; + tensor var_267_transpose_x_0 = const()[name = tensor("op_267_transpose_x_0"), val = tensor(false)]; + tensor var_267_transpose_y_0 = const()[name = tensor("op_267_transpose_y_0"), val = tensor(false)]; + tensor transpose_6_perm_0 = const()[name = tensor("transpose_6_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_7_perm_0 = const()[name = tensor("transpose_7_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_7 = transpose(perm = transpose_7_perm_0, x = new_k_cache_3)[name = tensor("transpose_16")]; + tensor transpose_6 = transpose(perm = transpose_6_perm_0, x = q_5)[name = tensor("transpose_17")]; + tensor var_267 = matmul(transpose_x = var_267_transpose_x_0, transpose_y = var_267_transpose_y_0, x = transpose_6, y = transpose_7)[name = tensor("op_267")]; + tensor var_268 = const()[name = tensor("op_268"), val = tensor(0x1p-3)]; + tensor attn_1 = mul(x = var_267, y = var_268)[name = tensor("attn_1")]; + tensor var_270 = logical_not(x = attn_mask_5)[name = tensor("op_270")]; + tensor attn_3 = select(a = var_100, b = attn_1, cond = var_270)[name = tensor("attn_3")]; + tensor attn_5 = softmax(axis = var_91, x = attn_3)[name = tensor("attn_5")]; + tensor x_5_transpose_x_0 = const()[name = tensor("x_5_transpose_x_0"), val = tensor(false)]; + tensor x_5_transpose_y_0 = const()[name = tensor("x_5_transpose_y_0"), val = tensor(false)]; + tensor v_attn_1 = transpose(perm = var_219, x = new_v_cache_3)[name = tensor("transpose_18")]; + tensor x_5 = matmul(transpose_x = x_5_transpose_x_0, transpose_y = x_5_transpose_y_0, x = attn_5, y = v_attn_1)[name = tensor("x_5")]; + tensor var_274_perm_0 = const()[name = tensor("op_274_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_275 = const()[name = tensor("op_275"), val = tensor([1, 16, 512])]; + tensor var_274 = transpose(perm = var_274_perm_0, x = x_5)[name = tensor("transpose_15")]; + tensor input_5 = reshape(shape = var_275, x = var_274)[name = tensor("input_5")]; + tensor linear_1_bias_0 = const()[name = tensor("linear_1_bias_0"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(41755456)))]; + tensor x_7 = linear(bias = linear_1_bias_0, weight = mimi_decoder_transformer_transformer_layers_0_self_attn_out_proj_weight, x = input_5)[name = tensor("linear_1")]; + tensor var_284 = mul(x = mimi_decoder_transformer_transformer_layers_0_layer_scale_1_scale, y = x_7)[name = tensor("op_284")]; + tensor input_7 = add(x = input_3, y = var_284)[name = tensor("input_7")]; + tensor input_9_axes_0 = const()[name = tensor("input_9_axes_0"), val = tensor([-1])]; + tensor input_9 = layer_norm(axes = input_9_axes_0, beta = mimi_decoder_transformer_transformer_layers_0_norm2_bias, epsilon = var_102, gamma = mimi_decoder_transformer_transformer_layers_0_norm2_weight, x = input_7)[name = tensor("input_9")]; + tensor linear_2_bias_0 = const()[name = tensor("linear_2_bias_0"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(41757568)))]; + tensor var_291 = linear(bias = linear_2_bias_0, weight = mimi_decoder_transformer_transformer_layers_0_linear1_weight, x = input_9)[name = tensor("linear_2")]; + tensor input_11_mode_0 = const()[name = tensor("input_11_mode_0"), val = tensor("EXACT")]; + tensor input_11 = gelu(mode = input_11_mode_0, x = var_291)[name = tensor("input_11")]; + tensor x_9 = linear(bias = linear_1_bias_0, weight = mimi_decoder_transformer_transformer_layers_0_linear2_weight, x = input_11)[name = tensor("linear_3")]; + tensor var_297 = mul(x = mimi_decoder_transformer_transformer_layers_0_layer_scale_2_scale, y = x_9)[name = tensor("op_297")]; + tensor input_13 = add(x = input_7, y = var_297)[name = tensor("input_13")]; + tensor query_axes_0 = const()[name = tensor("query_axes_0"), val = tensor([-1])]; + tensor query = layer_norm(axes = query_axes_0, beta = mimi_decoder_transformer_transformer_layers_1_norm1_bias, epsilon = var_102, gamma = mimi_decoder_transformer_transformer_layers_1_norm1_weight, x = input_13)[name = tensor("query")]; + tensor projected = linear(bias = linear_0_bias_0, weight = mimi_decoder_transformer_transformer_layers_1_self_attn_in_proj_weight, x = query)[name = tensor("linear_4")]; + tensor var_320 = const()[name = tensor("op_320"), val = tensor([1, 16, 3, 8, 64])]; + tensor packed = reshape(shape = var_320, x = projected)[name = tensor("packed")]; + tensor var_322_split_sizes_0 = const()[name = tensor("op_322_split_sizes_0"), val = tensor([1, 1, 1])]; + tensor var_322_axis_0 = const()[name = tensor("op_322_axis_0"), val = tensor(2)]; + tensor var_322_0, tensor var_322_1, tensor var_322_2 = split(axis = var_322_axis_0, split_sizes = var_322_split_sizes_0, x = packed)[name = tensor("op_322")]; + tensor squeeze_3_axes_0 = const()[name = tensor("squeeze_3_axes_0"), val = tensor([2])]; + tensor squeeze_3 = squeeze(axes = squeeze_3_axes_0, x = var_322_0)[name = tensor("squeeze_3")]; + tensor squeeze_4_axes_0 = const()[name = tensor("squeeze_4_axes_0"), val = tensor([2])]; + tensor squeeze_4 = squeeze(axes = squeeze_4_axes_0, x = var_322_1)[name = tensor("squeeze_4")]; + tensor squeeze_5_axes_0 = const()[name = tensor("squeeze_5_axes_0"), val = tensor([2])]; + tensor squeeze_5 = squeeze(axes = squeeze_5_axes_0, x = var_322_2)[name = tensor("squeeze_5")]; + tensor offset_begin_0 = const()[name = tensor("offset_begin_0"), val = tensor([0])]; + tensor offset_end_0 = const()[name = tensor("offset_end_0"), val = tensor([1])]; + tensor offset_end_mask_0 = const()[name = tensor("offset_end_mask_0"), val = tensor([false])]; + tensor offset_squeeze_mask_0 = const()[name = tensor("offset_squeeze_mask_0"), val = tensor([true])]; + tensor offset = slice_by_index(begin = offset_begin_0, end = offset_end_0, end_mask = offset_end_mask_0, squeeze_mask = offset_squeeze_mask_0, x = attn1_offset)[name = tensor("offset")]; + tensor freqs = const()[name = tensor("freqs"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(41765824)))]; + tensor ts_7_promoted = const()[name = tensor("ts_7_promoted"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(41766016)))]; + tensor ts_9 = add(x = ts_7_promoted, y = offset)[name = tensor("ts_9")]; + tensor var_338 = const()[name = tensor("op_338"), val = tensor([-1, 1, 1])]; + tensor ts = reshape(shape = var_338, x = ts_9)[name = tensor("ts")]; + tensor var_340 = const()[name = tensor("op_340"), val = tensor([1, 16, 8, 32, 2])]; + tensor q_9 = reshape(shape = var_340, x = squeeze_3)[name = tensor("q_9")]; + tensor var_342 = const()[name = tensor("op_342"), val = tensor([1, 16, 8, 32, 2])]; + tensor k_9 = reshape(shape = var_342, x = squeeze_4)[name = tensor("k_9")]; + tensor var_344_begin_0 = const()[name = tensor("op_344_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_344_end_0 = const()[name = tensor("op_344_end_0"), val = tensor([1, 16, 8, 32, 1])]; + tensor var_344_end_mask_0 = const()[name = tensor("op_344_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_344_squeeze_mask_0 = const()[name = tensor("op_344_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_344 = slice_by_index(begin = var_344_begin_0, end = var_344_end_0, end_mask = var_344_end_mask_0, squeeze_mask = var_344_squeeze_mask_0, x = q_9)[name = tensor("op_344")]; + tensor var_346_begin_0 = const()[name = tensor("op_346_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_346_end_0 = const()[name = tensor("op_346_end_0"), val = tensor([1, 16, 8, 32, 2])]; + tensor var_346_end_mask_0 = const()[name = tensor("op_346_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_346_squeeze_mask_0 = const()[name = tensor("op_346_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_346 = slice_by_index(begin = var_346_begin_0, end = var_346_end_0, end_mask = var_346_end_mask_0, squeeze_mask = var_346_squeeze_mask_0, x = q_9)[name = tensor("op_346")]; + tensor var_348_begin_0 = const()[name = tensor("op_348_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_348_end_0 = const()[name = tensor("op_348_end_0"), val = tensor([1, 16, 8, 32, 1])]; + tensor var_348_end_mask_0 = const()[name = tensor("op_348_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_348_squeeze_mask_0 = const()[name = tensor("op_348_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_348 = slice_by_index(begin = var_348_begin_0, end = var_348_end_0, end_mask = var_348_end_mask_0, squeeze_mask = var_348_squeeze_mask_0, x = k_9)[name = tensor("op_348")]; + tensor var_350_begin_0 = const()[name = tensor("op_350_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_350_end_0 = const()[name = tensor("op_350_end_0"), val = tensor([1, 16, 8, 32, 2])]; + tensor var_350_end_mask_0 = const()[name = tensor("op_350_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_350_squeeze_mask_0 = const()[name = tensor("op_350_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_350 = slice_by_index(begin = var_350_begin_0, end = var_350_end_0, end_mask = var_350_end_mask_0, squeeze_mask = var_350_squeeze_mask_0, x = k_9)[name = tensor("op_350")]; + tensor var_352 = mul(x = freqs, y = ts)[name = tensor("op_352")]; + tensor rotr = cos(x = var_352)[name = tensor("rotr")]; + tensor roti = sin(x = var_352)[name = tensor("roti")]; + tensor var_356 = mul(x = var_344, y = rotr)[name = tensor("op_356")]; + tensor var_357 = mul(x = var_346, y = roti)[name = tensor("op_357")]; + tensor qor_5 = sub(x = var_356, y = var_357)[name = tensor("qor_5")]; + tensor var_359 = mul(x = var_344, y = roti)[name = tensor("op_359")]; + tensor var_360 = mul(x = var_346, y = rotr)[name = tensor("op_360")]; + tensor qoi_5 = add(x = var_359, y = var_360)[name = tensor("qoi_5")]; + tensor var_362 = mul(x = var_348, y = rotr)[name = tensor("op_362")]; + tensor var_363 = mul(x = var_350, y = roti)[name = tensor("op_363")]; + tensor kor_5 = sub(x = var_362, y = var_363)[name = tensor("kor_5")]; + tensor var_365 = mul(x = var_348, y = roti)[name = tensor("op_365")]; + tensor var_366 = mul(x = var_350, y = rotr)[name = tensor("op_366")]; + tensor koi_5 = add(x = var_365, y = var_366)[name = tensor("koi_5")]; + tensor qo_axis_0 = const()[name = tensor("qo_axis_0"), val = tensor(-1)]; + tensor qo = stack(axis = qo_axis_0, values = (qor_5, qoi_5))[name = tensor("qo")]; + tensor ko_axis_0 = const()[name = tensor("ko_axis_0"), val = tensor(-1)]; + tensor ko = stack(axis = ko_axis_0, values = (kor_5, koi_5))[name = tensor("ko")]; + tensor var_376 = const()[name = tensor("op_376"), val = tensor([1, 16, 8, 64])]; + tensor q = reshape(shape = var_376, x = qo)[name = tensor("q")]; + tensor var_378 = const()[name = tensor("op_378"), val = tensor([1, 16, 8, 64])]; + tensor k = reshape(shape = var_378, x = ko)[name = tensor("k")]; + tensor capacity = const()[name = tensor("capacity"), val = tensor([256])]; + tensor var_383_dtype_0 = const()[name = tensor("op_383_dtype_0"), val = tensor("int32")]; + tensor var_384 = const()[name = tensor("op_384"), val = tensor([1, 1])]; + tensor var_383 = cast(dtype = var_383_dtype_0, x = attn1_offset)[name = tensor("cast_46")]; + tensor write_base = reshape(shape = var_384, x = var_383)[name = tensor("write_base")]; + tensor write_range = const()[name = tensor("write_range"), val = tensor([[0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15]])]; + tensor abs_idx = add(x = write_base, y = write_range)[name = tensor("abs_idx")]; + tensor wrapped_div = floor_div(x = abs_idx, y = capacity)[name = tensor("wrapped_div")]; + tensor wrapped_div_scaled = mul(x = wrapped_div, y = capacity)[name = tensor("wrapped_div_scaled")]; + tensor wrapped = sub(x = abs_idx, y = wrapped_div_scaled)[name = tensor("wrapped")]; + tensor var_391 = const()[name = tensor("op_391"), val = tensor([1, 16, 1, 1])]; + tensor var_392 = reshape(shape = var_391, x = wrapped)[name = tensor("op_392")]; + tensor write_indexes_reps_0 = const()[name = tensor("write_indexes_reps_0"), val = tensor([1, 1, 8, 64])]; + tensor write_indexes = tile(reps = write_indexes_reps_0, x = var_392)[name = tensor("write_indexes")]; + tensor var_395_begin_0 = const()[name = tensor("op_395_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_395_end_0 = const()[name = tensor("op_395_end_0"), val = tensor([1, 1, 256, 8, 64])]; + tensor var_395_end_mask_0 = const()[name = tensor("op_395_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_395_squeeze_mask_0 = const()[name = tensor("op_395_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor var_395 = slice_by_index(begin = var_395_begin_0, end = var_395_end_0, end_mask = var_395_end_mask_0, squeeze_mask = var_395_squeeze_mask_0, x = attn1_cache)[name = tensor("op_395")]; + tensor new_k_cache_5_axis_0 = const()[name = tensor("new_k_cache_5_axis_0"), val = tensor(1)]; + tensor new_k_cache_5_mode_0 = const()[name = tensor("new_k_cache_5_mode_0"), val = tensor("update")]; + tensor new_k_cache_5_validate_indices_0 = const()[name = tensor("new_k_cache_5_validate_indices_0"), val = tensor(false)]; + tensor new_k_cache_5 = scatter_along_axis(axis = new_k_cache_5_axis_0, data = var_395, indices = write_indexes, mode = new_k_cache_5_mode_0, updates = k, validate_indices = new_k_cache_5_validate_indices_0)[name = tensor("new_k_cache_5")]; + tensor var_397_begin_0 = const()[name = tensor("op_397_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor var_397_end_0 = const()[name = tensor("op_397_end_0"), val = tensor([2, 1, 256, 8, 64])]; + tensor var_397_end_mask_0 = const()[name = tensor("op_397_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_397_squeeze_mask_0 = const()[name = tensor("op_397_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor var_397 = slice_by_index(begin = var_397_begin_0, end = var_397_end_0, end_mask = var_397_end_mask_0, squeeze_mask = var_397_squeeze_mask_0, x = attn1_cache)[name = tensor("op_397")]; + tensor new_v_cache_5_axis_0 = const()[name = tensor("new_v_cache_5_axis_0"), val = tensor(1)]; + tensor new_v_cache_5_mode_0 = const()[name = tensor("new_v_cache_5_mode_0"), val = tensor("update")]; + tensor new_v_cache_5_validate_indices_0 = const()[name = tensor("new_v_cache_5_validate_indices_0"), val = tensor(false)]; + tensor new_v_cache_5 = scatter_along_axis(axis = new_v_cache_5_axis_0, data = var_397, indices = write_indexes, mode = new_v_cache_5_mode_0, updates = squeeze_5, validate_indices = new_v_cache_5_validate_indices_0)[name = tensor("new_v_cache_5")]; + tensor var_400_axis_0 = const()[name = tensor("op_400_axis_0"), val = tensor(0)]; + tensor var_400 = stack(axis = var_400_axis_0, values = (new_k_cache_5, new_v_cache_5))[name = tensor("op_400")]; + tensor var_401 = not_equal(x = new_k_cache_5, y = new_k_cache_5)[name = tensor("op_401")]; + tensor new_k_cache = select(a = var_212, b = new_k_cache_5, cond = var_401)[name = tensor("new_k_cache")]; + tensor var_404 = not_equal(x = new_v_cache_5, y = new_v_cache_5)[name = tensor("op_404")]; + tensor new_v_cache = select(a = var_212, b = new_v_cache_5, cond = var_404)[name = tensor("new_v_cache")]; + tensor var_409 = const()[name = tensor("op_409"), val = tensor([0, 2, 1, 3])]; + tensor var_411 = const()[name = tensor("op_411"), val = tensor([1, 1])]; + tensor var_412 = reshape(shape = var_411, x = attn1_offset)[name = tensor("op_412")]; + tensor var_414_promoted = const()[name = tensor("op_414_promoted"), val = tensor([0x1.ep+3])]; + tensor var_415 = add(x = var_412, y = var_414_promoted)[name = tensor("op_415")]; + tensor last_pos_dtype_0 = const()[name = tensor("last_pos_dtype_0"), val = tensor("int32")]; + tensor last_pos = cast(dtype = last_pos_dtype_0, x = var_415)[name = tensor("cast_45")]; + tensor diff = sub(x = last_pos, y = slot_idx_1)[name = tensor("diff")]; + tensor var_421_div = floor_div(x = diff, y = capacity)[name = tensor("op_421_div")]; + tensor var_421_div_scaled = mul(x = var_421_div, y = capacity)[name = tensor("op_421_div_scaled")]; + tensor var_421 = sub(x = diff, y = var_421_div_scaled)[name = tensor("op_421")]; + tensor pos_k = sub(x = last_pos, y = var_421)[name = tensor("pos_k")]; + tensor var_427_promoted = const()[name = tensor("op_427_promoted"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(41766144)))]; + tensor pos_q = add(x = var_412, y = var_427_promoted)[name = tensor("pos_q")]; + tensor var_431_axes_0 = const()[name = tensor("op_431_axes_0"), val = tensor([2])]; + tensor var_431 = expand_dims(axes = var_431_axes_0, x = pos_q)[name = tensor("op_431")]; + tensor var_433_axes_0 = const()[name = tensor("op_433_axes_0"), val = tensor([1])]; + tensor var_433 = expand_dims(axes = var_433_axes_0, x = pos_k)[name = tensor("op_433")]; + tensor var_434_promoted_dtype_0 = const()[name = tensor("op_434_promoted_dtype_0"), val = tensor("fp32")]; + tensor var_434_promoted = cast(dtype = var_434_promoted_dtype_0, x = var_433)[name = tensor("cast_44")]; + tensor delta = sub(x = var_431, y = var_434_promoted)[name = tensor("delta")]; + tensor valid_5 = greater_equal(x = var_433, y = var_86)[name = tensor("valid_5")]; + tensor var_443 = const()[name = tensor("op_443"), val = tensor([1, 1, 1])]; + tensor var_444 = reshape(shape = var_443, x = attn1_offset)[name = tensor("op_444")]; + tensor var_446_promoted = const()[name = tensor("op_446_promoted"), val = tensor([0x1.ep+3])]; + tensor var_447 = add(x = var_444, y = var_446_promoted)[name = tensor("op_447")]; + tensor var_448 = less_equal(x = var_434_promoted, y = var_447)[name = tensor("op_448")]; + tensor valid = logical_and(x = valid_5, y = var_448)[name = tensor("valid")]; + tensor var_86_promoted_1 = const()[name = tensor("op_86_promoted_1"), val = tensor(0x0p+0)]; + tensor var_450 = greater_equal(x = delta, y = var_86_promoted_1)[name = tensor("op_450")]; + tensor attn_mask_7 = logical_and(x = valid, y = var_450)[name = tensor("attn_mask_7")]; + tensor var_98_promoted_1 = const()[name = tensor("op_98_promoted_1"), val = tensor(0x1.f4p+7)]; + tensor var_452 = less(x = delta, y = var_98_promoted_1)[name = tensor("op_452")]; + tensor attn_mask_9 = logical_and(x = attn_mask_7, y = var_452)[name = tensor("attn_mask_9")]; + tensor attn_mask_axes_0 = const()[name = tensor("attn_mask_axes_0"), val = tensor([1])]; + tensor attn_mask = expand_dims(axes = attn_mask_axes_0, x = attn_mask_9)[name = tensor("attn_mask")]; + tensor var_457_transpose_x_0 = const()[name = tensor("op_457_transpose_x_0"), val = tensor(false)]; + tensor var_457_transpose_y_0 = const()[name = tensor("op_457_transpose_y_0"), val = tensor(false)]; + tensor transpose_8_perm_0 = const()[name = tensor("transpose_8_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_9_perm_0 = const()[name = tensor("transpose_9_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_9 = transpose(perm = transpose_9_perm_0, x = new_k_cache)[name = tensor("transpose_12")]; + tensor transpose_8 = transpose(perm = transpose_8_perm_0, x = q)[name = tensor("transpose_13")]; + tensor var_457 = matmul(transpose_x = var_457_transpose_x_0, transpose_y = var_457_transpose_y_0, x = transpose_8, y = transpose_9)[name = tensor("op_457")]; + tensor var_458 = const()[name = tensor("op_458"), val = tensor(0x1p-3)]; + tensor attn_7 = mul(x = var_457, y = var_458)[name = tensor("attn_7")]; + tensor var_460 = logical_not(x = attn_mask)[name = tensor("op_460")]; + tensor attn_9 = select(a = var_100, b = attn_7, cond = var_460)[name = tensor("attn_9")]; + tensor attn = softmax(axis = var_91, x = attn_9)[name = tensor("attn")]; + tensor x_11_transpose_x_0 = const()[name = tensor("x_11_transpose_x_0"), val = tensor(false)]; + tensor x_11_transpose_y_0 = const()[name = tensor("x_11_transpose_y_0"), val = tensor(false)]; + tensor v_attn = transpose(perm = var_409, x = new_v_cache)[name = tensor("transpose_14")]; + tensor x_11 = matmul(transpose_x = x_11_transpose_x_0, transpose_y = x_11_transpose_y_0, x = attn, y = v_attn)[name = tensor("x_11")]; + tensor var_464_perm_0 = const()[name = tensor("op_464_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_465 = const()[name = tensor("op_465"), val = tensor([1, 16, 512])]; + tensor var_464 = transpose(perm = var_464_perm_0, x = x_11)[name = tensor("transpose_11")]; + tensor input_15 = reshape(shape = var_465, x = var_464)[name = tensor("input_15")]; + tensor x_13 = linear(bias = linear_1_bias_0, weight = mimi_decoder_transformer_transformer_layers_1_self_attn_out_proj_weight, x = input_15)[name = tensor("linear_5")]; + tensor var_474 = mul(x = mimi_decoder_transformer_transformer_layers_1_layer_scale_1_scale, y = x_13)[name = tensor("op_474")]; + tensor input_17 = add(x = input_13, y = var_474)[name = tensor("input_17")]; + tensor input_19_axes_0 = const()[name = tensor("input_19_axes_0"), val = tensor([-1])]; + tensor input_19 = layer_norm(axes = input_19_axes_0, beta = mimi_decoder_transformer_transformer_layers_1_norm2_bias, epsilon = var_102, gamma = mimi_decoder_transformer_transformer_layers_1_norm2_weight, x = input_17)[name = tensor("input_19")]; + tensor var_481 = linear(bias = linear_2_bias_0, weight = mimi_decoder_transformer_transformer_layers_1_linear1_weight, x = input_19)[name = tensor("linear_6")]; + tensor input_21_mode_0 = const()[name = tensor("input_21_mode_0"), val = tensor("EXACT")]; + tensor input_21 = gelu(mode = input_21_mode_0, x = var_481)[name = tensor("input_21")]; + tensor x_15 = linear(bias = linear_1_bias_0, weight = mimi_decoder_transformer_transformer_layers_1_linear2_weight, x = input_21)[name = tensor("linear_7")]; + tensor var_487 = mul(x = mimi_decoder_transformer_transformer_layers_1_layer_scale_2_scale, y = x_15)[name = tensor("op_487")]; + tensor z = add(x = input_17, y = var_487)[name = tensor("z")]; + tensor x_17_perm_0 = const()[name = tensor("x_17_perm_0"), val = tensor([0, 2, 1])]; + tensor var_507 = const()[name = tensor("op_507"), val = tensor(0x1p+0)]; + tensor var_508 = const()[name = tensor("op_508"), val = tensor(-1)]; + tensor input_23_interleave_0 = const()[name = tensor("input_23_interleave_0"), val = tensor(false)]; + tensor x_17 = transpose(perm = x_17_perm_0, x = z)[name = tensor("transpose_10")]; + tensor input_23 = concat(axis = var_508, interleave = input_23_interleave_0, values = (conv0_prev, x_17))[name = tensor("input_23")]; + tensor input_25_pad_type_0 = const()[name = tensor("input_25_pad_type_0"), val = tensor("valid")]; + tensor input_25_strides_0 = const()[name = tensor("input_25_strides_0"), val = tensor([1])]; + tensor input_25_pad_0 = const()[name = tensor("input_25_pad_0"), val = tensor([0, 0])]; + tensor input_25_dilations_0 = const()[name = tensor("input_25_dilations_0"), val = tensor([1])]; + tensor input_25_groups_0 = const()[name = tensor("input_25_groups_0"), val = tensor(1)]; + tensor input_25 = conv(bias = mimi_decoder_model_0_conv_bias, dilations = input_25_dilations_0, groups = input_25_groups_0, pad = input_25_pad_0, pad_type = input_25_pad_type_0, strides = input_25_strides_0, weight = mimi_decoder_model_0_conv_weight, x = input_23)[name = tensor("input_25")]; + tensor var_542_begin_0 = const()[name = tensor("op_542_begin_0"), val = tensor([0, 0, 16])]; + tensor var_542_end_0 = const()[name = tensor("op_542_end_0"), val = tensor([1, 512, 22])]; + tensor var_542_end_mask_0 = const()[name = tensor("op_542_end_mask_0"), val = tensor([true, true, true])]; + tensor var_542 = slice_by_index(begin = var_542_begin_0, end = var_542_end_0, end_mask = var_542_end_mask_0, x = input_23)[name = tensor("op_542")]; + tensor input_27 = elu(alpha = var_507, x = input_25)[name = tensor("input_27")]; + tensor y_5_pad_type_0 = const()[name = tensor("y_5_pad_type_0"), val = tensor("valid")]; + tensor y_5_strides_0 = const()[name = tensor("y_5_strides_0"), val = tensor([6])]; + tensor y_5_pad_0 = const()[name = tensor("y_5_pad_0"), val = tensor([0, 0])]; + tensor y_5_dilations_0 = const()[name = tensor("y_5_dilations_0"), val = tensor([1])]; + tensor y_5_groups_0 = const()[name = tensor("y_5_groups_0"), val = tensor(1)]; + tensor y_5_has_output_shape_output_shape_0 = const()[name = tensor("y_5_has_output_shape_output_shape_0"), val = tensor([1, 256, 102])]; + tensor y_5_has_output_shape = conv_transpose(bias = mimi_decoder_model_2_convtr_bias, dilations = y_5_dilations_0, groups = y_5_groups_0, output_shape = y_5_has_output_shape_output_shape_0, pad = y_5_pad_0, pad_type = y_5_pad_type_0, strides = y_5_strides_0, weight = mimi_decoder_model_2_convtr_weight, x = input_27)[name = tensor("y_5_has_output_shape")]; + tensor var_557_begin_0 = const()[name = tensor("op_557_begin_0"), val = tensor([0, 0, 0])]; + tensor var_557_end_0 = const()[name = tensor("op_557_end_0"), val = tensor([1, 256, 6])]; + tensor var_557_end_mask_0 = const()[name = tensor("op_557_end_mask_0"), val = tensor([true, true, false])]; + tensor var_557 = slice_by_index(begin = var_557_begin_0, end = var_557_end_0, end_mask = var_557_end_mask_0, x = y_5_has_output_shape)[name = tensor("op_557")]; + tensor var_558 = add(x = var_557, y = convtr0_partial)[name = tensor("op_558")]; + tensor var_559_begin_0 = const()[name = tensor("op_559_begin_0"), val = tensor([0, 0, 6])]; + tensor var_559_end_0 = const()[name = tensor("op_559_end_0"), val = tensor([1, 256, 102])]; + tensor var_559_end_mask_0 = const()[name = tensor("op_559_end_mask_0"), val = tensor([true, true, true])]; + tensor var_559 = slice_by_index(begin = var_559_begin_0, end = var_559_end_0, end_mask = var_559_end_mask_0, x = y_5_has_output_shape)[name = tensor("op_559")]; + tensor y_7_interleave_0 = const()[name = tensor("y_7_interleave_0"), val = tensor(false)]; + tensor y_7 = concat(axis = var_508, interleave = y_7_interleave_0, values = (var_558, var_559))[name = tensor("y_7")]; + tensor new_partial_1_begin_0 = const()[name = tensor("new_partial_1_begin_0"), val = tensor([0, 0, 96])]; + tensor new_partial_1_end_0 = const()[name = tensor("new_partial_1_end_0"), val = tensor([1, 256, 102])]; + tensor new_partial_1_end_mask_0 = const()[name = tensor("new_partial_1_end_mask_0"), val = tensor([true, true, true])]; + tensor new_partial_1 = slice_by_index(begin = new_partial_1_begin_0, end = new_partial_1_end_0, end_mask = new_partial_1_end_mask_0, x = y_7)[name = tensor("new_partial_1")]; + tensor var_564 = const()[name = tensor("op_564"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(41766272)))]; + tensor var_565 = sub(x = new_partial_1, y = var_564)[name = tensor("op_565")]; + tensor input_29_begin_0 = const()[name = tensor("input_29_begin_0"), val = tensor([0, 0, 0])]; + tensor input_29_end_0 = const()[name = tensor("input_29_end_0"), val = tensor([1, 256, 96])]; + tensor input_29_end_mask_0 = const()[name = tensor("input_29_end_mask_0"), val = tensor([true, true, false])]; + tensor input_29 = slice_by_index(begin = input_29_begin_0, end = input_29_end_0, end_mask = input_29_end_mask_0, x = y_7)[name = tensor("input_29")]; + tensor x_19 = elu(alpha = var_507, x = input_29)[name = tensor("x_19")]; + tensor input_31_interleave_0 = const()[name = tensor("input_31_interleave_0"), val = tensor(false)]; + tensor input_31 = concat(axis = var_508, interleave = input_31_interleave_0, values = (res0_conv0_prev, x_19))[name = tensor("input_31")]; + tensor input_33_pad_type_0 = const()[name = tensor("input_33_pad_type_0"), val = tensor("valid")]; + tensor input_33_strides_0 = const()[name = tensor("input_33_strides_0"), val = tensor([1])]; + tensor input_33_pad_0 = const()[name = tensor("input_33_pad_0"), val = tensor([0, 0])]; + tensor input_33_dilations_0 = const()[name = tensor("input_33_dilations_0"), val = tensor([1])]; + tensor input_33_groups_0 = const()[name = tensor("input_33_groups_0"), val = tensor(1)]; + tensor input_33 = conv(bias = mimi_decoder_model_3_block_1_conv_bias, dilations = input_33_dilations_0, groups = input_33_groups_0, pad = input_33_pad_0, pad_type = input_33_pad_type_0, strides = input_33_strides_0, weight = mimi_decoder_model_3_block_1_conv_weight, x = input_31)[name = tensor("input_33")]; + tensor var_585_begin_0 = const()[name = tensor("op_585_begin_0"), val = tensor([0, 0, 96])]; + tensor var_585_end_0 = const()[name = tensor("op_585_end_0"), val = tensor([1, 256, 98])]; + tensor var_585_end_mask_0 = const()[name = tensor("op_585_end_mask_0"), val = tensor([true, true, true])]; + tensor var_585 = slice_by_index(begin = var_585_begin_0, end = var_585_end_0, end_mask = var_585_end_mask_0, x = input_31)[name = tensor("op_585")]; + tensor x_21 = elu(alpha = var_507, x = input_33)[name = tensor("x_21")]; + tensor v_5_pad_type_0 = const()[name = tensor("v_5_pad_type_0"), val = tensor("valid")]; + tensor v_5_strides_0 = const()[name = tensor("v_5_strides_0"), val = tensor([1])]; + tensor v_5_pad_0 = const()[name = tensor("v_5_pad_0"), val = tensor([0, 0])]; + tensor v_5_dilations_0 = const()[name = tensor("v_5_dilations_0"), val = tensor([1])]; + tensor v_5_groups_0 = const()[name = tensor("v_5_groups_0"), val = tensor(1)]; + tensor v_5 = conv(bias = mimi_decoder_model_3_block_3_conv_bias, dilations = v_5_dilations_0, groups = v_5_groups_0, pad = v_5_pad_0, pad_type = v_5_pad_type_0, strides = v_5_strides_0, weight = mimi_decoder_model_3_block_3_conv_weight, x = x_21)[name = tensor("v_5")]; + tensor input_35 = add(x = input_29, y = v_5)[name = tensor("input_35")]; + tensor input_37 = elu(alpha = var_507, x = input_35)[name = tensor("input_37")]; + tensor y_9_pad_type_0 = const()[name = tensor("y_9_pad_type_0"), val = tensor("valid")]; + tensor y_9_strides_0 = const()[name = tensor("y_9_strides_0"), val = tensor([5])]; + tensor y_9_pad_0 = const()[name = tensor("y_9_pad_0"), val = tensor([0, 0])]; + tensor y_9_dilations_0 = const()[name = tensor("y_9_dilations_0"), val = tensor([1])]; + tensor y_9_groups_0 = const()[name = tensor("y_9_groups_0"), val = tensor(1)]; + tensor y_9_has_output_shape_output_shape_0 = const()[name = tensor("y_9_has_output_shape_output_shape_0"), val = tensor([1, 128, 485])]; + tensor y_9_has_output_shape = conv_transpose(bias = mimi_decoder_model_5_convtr_bias, dilations = y_9_dilations_0, groups = y_9_groups_0, output_shape = y_9_has_output_shape_output_shape_0, pad = y_9_pad_0, pad_type = y_9_pad_type_0, strides = y_9_strides_0, weight = mimi_decoder_model_5_convtr_weight, x = input_37)[name = tensor("y_9_has_output_shape")]; + tensor var_613_begin_0 = const()[name = tensor("op_613_begin_0"), val = tensor([0, 0, 0])]; + tensor var_613_end_0 = const()[name = tensor("op_613_end_0"), val = tensor([1, 128, 5])]; + tensor var_613_end_mask_0 = const()[name = tensor("op_613_end_mask_0"), val = tensor([true, true, false])]; + tensor var_613 = slice_by_index(begin = var_613_begin_0, end = var_613_end_0, end_mask = var_613_end_mask_0, x = y_9_has_output_shape)[name = tensor("op_613")]; + tensor var_614 = add(x = var_613, y = convtr1_partial)[name = tensor("op_614")]; + tensor var_615_begin_0 = const()[name = tensor("op_615_begin_0"), val = tensor([0, 0, 5])]; + tensor var_615_end_0 = const()[name = tensor("op_615_end_0"), val = tensor([1, 128, 485])]; + tensor var_615_end_mask_0 = const()[name = tensor("op_615_end_mask_0"), val = tensor([true, true, true])]; + tensor var_615 = slice_by_index(begin = var_615_begin_0, end = var_615_end_0, end_mask = var_615_end_mask_0, x = y_9_has_output_shape)[name = tensor("op_615")]; + tensor y_11_interleave_0 = const()[name = tensor("y_11_interleave_0"), val = tensor(false)]; + tensor y_11 = concat(axis = var_508, interleave = y_11_interleave_0, values = (var_614, var_615))[name = tensor("y_11")]; + tensor new_partial_3_begin_0 = const()[name = tensor("new_partial_3_begin_0"), val = tensor([0, 0, 480])]; + tensor new_partial_3_end_0 = const()[name = tensor("new_partial_3_end_0"), val = tensor([1, 128, 485])]; + tensor new_partial_3_end_mask_0 = const()[name = tensor("new_partial_3_end_mask_0"), val = tensor([true, true, true])]; + tensor new_partial_3 = slice_by_index(begin = new_partial_3_begin_0, end = new_partial_3_end_0, end_mask = new_partial_3_end_mask_0, x = y_11)[name = tensor("new_partial_3")]; + tensor var_620 = const()[name = tensor("op_620"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(41767360)))]; + tensor var_621 = sub(x = new_partial_3, y = var_620)[name = tensor("op_621")]; + tensor input_39_begin_0 = const()[name = tensor("input_39_begin_0"), val = tensor([0, 0, 0])]; + tensor input_39_end_0 = const()[name = tensor("input_39_end_0"), val = tensor([1, 128, 480])]; + tensor input_39_end_mask_0 = const()[name = tensor("input_39_end_mask_0"), val = tensor([true, true, false])]; + tensor input_39 = slice_by_index(begin = input_39_begin_0, end = input_39_end_0, end_mask = input_39_end_mask_0, x = y_11)[name = tensor("input_39")]; + tensor x_23 = elu(alpha = var_507, x = input_39)[name = tensor("x_23")]; + tensor input_41_interleave_0 = const()[name = tensor("input_41_interleave_0"), val = tensor(false)]; + tensor input_41 = concat(axis = var_508, interleave = input_41_interleave_0, values = (res1_conv0_prev, x_23))[name = tensor("input_41")]; + tensor input_43_pad_type_0 = const()[name = tensor("input_43_pad_type_0"), val = tensor("valid")]; + tensor input_43_strides_0 = const()[name = tensor("input_43_strides_0"), val = tensor([1])]; + tensor input_43_pad_0 = const()[name = tensor("input_43_pad_0"), val = tensor([0, 0])]; + tensor input_43_dilations_0 = const()[name = tensor("input_43_dilations_0"), val = tensor([1])]; + tensor input_43_groups_0 = const()[name = tensor("input_43_groups_0"), val = tensor(1)]; + tensor input_43 = conv(bias = mimi_decoder_model_6_block_1_conv_bias, dilations = input_43_dilations_0, groups = input_43_groups_0, pad = input_43_pad_0, pad_type = input_43_pad_type_0, strides = input_43_strides_0, weight = mimi_decoder_model_6_block_1_conv_weight, x = input_41)[name = tensor("input_43")]; + tensor var_641_begin_0 = const()[name = tensor("op_641_begin_0"), val = tensor([0, 0, 480])]; + tensor var_641_end_0 = const()[name = tensor("op_641_end_0"), val = tensor([1, 128, 482])]; + tensor var_641_end_mask_0 = const()[name = tensor("op_641_end_mask_0"), val = tensor([true, true, true])]; + tensor var_641 = slice_by_index(begin = var_641_begin_0, end = var_641_end_0, end_mask = var_641_end_mask_0, x = input_41)[name = tensor("op_641")]; + tensor x_25 = elu(alpha = var_507, x = input_43)[name = tensor("x_25")]; + tensor v_7_pad_type_0 = const()[name = tensor("v_7_pad_type_0"), val = tensor("valid")]; + tensor v_7_strides_0 = const()[name = tensor("v_7_strides_0"), val = tensor([1])]; + tensor v_7_pad_0 = const()[name = tensor("v_7_pad_0"), val = tensor([0, 0])]; + tensor v_7_dilations_0 = const()[name = tensor("v_7_dilations_0"), val = tensor([1])]; + tensor v_7_groups_0 = const()[name = tensor("v_7_groups_0"), val = tensor(1)]; + tensor v_7 = conv(bias = mimi_decoder_model_6_block_3_conv_bias, dilations = v_7_dilations_0, groups = v_7_groups_0, pad = v_7_pad_0, pad_type = v_7_pad_type_0, strides = v_7_strides_0, weight = mimi_decoder_model_6_block_3_conv_weight, x = x_25)[name = tensor("v_7")]; + tensor input_45 = add(x = input_39, y = v_7)[name = tensor("input_45")]; + tensor input_47 = elu(alpha = var_507, x = input_45)[name = tensor("input_47")]; + tensor y_13_pad_type_0 = const()[name = tensor("y_13_pad_type_0"), val = tensor("valid")]; + tensor y_13_strides_0 = const()[name = tensor("y_13_strides_0"), val = tensor([4])]; + tensor y_13_pad_0 = const()[name = tensor("y_13_pad_0"), val = tensor([0, 0])]; + tensor y_13_dilations_0 = const()[name = tensor("y_13_dilations_0"), val = tensor([1])]; + tensor y_13_groups_0 = const()[name = tensor("y_13_groups_0"), val = tensor(1)]; + tensor y_13_has_output_shape_output_shape_0 = const()[name = tensor("y_13_has_output_shape_output_shape_0"), val = tensor([1, 64, 1924])]; + tensor y_13_has_output_shape = conv_transpose(bias = mimi_decoder_model_8_convtr_bias, dilations = y_13_dilations_0, groups = y_13_groups_0, output_shape = y_13_has_output_shape_output_shape_0, pad = y_13_pad_0, pad_type = y_13_pad_type_0, strides = y_13_strides_0, weight = mimi_decoder_model_8_convtr_weight, x = input_47)[name = tensor("y_13_has_output_shape")]; + tensor var_669_begin_0 = const()[name = tensor("op_669_begin_0"), val = tensor([0, 0, 0])]; + tensor var_669_end_0 = const()[name = tensor("op_669_end_0"), val = tensor([1, 64, 4])]; + tensor var_669_end_mask_0 = const()[name = tensor("op_669_end_mask_0"), val = tensor([true, true, false])]; + tensor var_669 = slice_by_index(begin = var_669_begin_0, end = var_669_end_0, end_mask = var_669_end_mask_0, x = y_13_has_output_shape)[name = tensor("op_669")]; + tensor var_670 = add(x = var_669, y = convtr2_partial)[name = tensor("op_670")]; + tensor var_671_begin_0 = const()[name = tensor("op_671_begin_0"), val = tensor([0, 0, 4])]; + tensor var_671_end_0 = const()[name = tensor("op_671_end_0"), val = tensor([1, 64, 1924])]; + tensor var_671_end_mask_0 = const()[name = tensor("op_671_end_mask_0"), val = tensor([true, true, true])]; + tensor var_671 = slice_by_index(begin = var_671_begin_0, end = var_671_end_0, end_mask = var_671_end_mask_0, x = y_13_has_output_shape)[name = tensor("op_671")]; + tensor y_interleave_0 = const()[name = tensor("y_interleave_0"), val = tensor(false)]; + tensor y = concat(axis = var_508, interleave = y_interleave_0, values = (var_670, var_671))[name = tensor("y")]; + tensor new_partial_begin_0 = const()[name = tensor("new_partial_begin_0"), val = tensor([0, 0, 1920])]; + tensor new_partial_end_0 = const()[name = tensor("new_partial_end_0"), val = tensor([1, 64, 1924])]; + tensor new_partial_end_mask_0 = const()[name = tensor("new_partial_end_mask_0"), val = tensor([true, true, true])]; + tensor new_partial = slice_by_index(begin = new_partial_begin_0, end = new_partial_end_0, end_mask = new_partial_end_mask_0, x = y)[name = tensor("new_partial")]; + tensor var_676 = const()[name = tensor("op_676"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(41767936)))]; + tensor var_677 = sub(x = new_partial, y = var_676)[name = tensor("op_677")]; + tensor input_49_begin_0 = const()[name = tensor("input_49_begin_0"), val = tensor([0, 0, 0])]; + tensor input_49_end_0 = const()[name = tensor("input_49_end_0"), val = tensor([1, 64, 1920])]; + tensor input_49_end_mask_0 = const()[name = tensor("input_49_end_mask_0"), val = tensor([true, true, false])]; + tensor input_49 = slice_by_index(begin = input_49_begin_0, end = input_49_end_0, end_mask = input_49_end_mask_0, x = y)[name = tensor("input_49")]; + tensor x_27 = elu(alpha = var_507, x = input_49)[name = tensor("x_27")]; + tensor input_51_interleave_0 = const()[name = tensor("input_51_interleave_0"), val = tensor(false)]; + tensor input_51 = concat(axis = var_508, interleave = input_51_interleave_0, values = (res2_conv0_prev, x_27))[name = tensor("input_51")]; + tensor input_53_pad_type_0 = const()[name = tensor("input_53_pad_type_0"), val = tensor("valid")]; + tensor input_53_strides_0 = const()[name = tensor("input_53_strides_0"), val = tensor([1])]; + tensor input_53_pad_0 = const()[name = tensor("input_53_pad_0"), val = tensor([0, 0])]; + tensor input_53_dilations_0 = const()[name = tensor("input_53_dilations_0"), val = tensor([1])]; + tensor input_53_groups_0 = const()[name = tensor("input_53_groups_0"), val = tensor(1)]; + tensor input_53 = conv(bias = mimi_decoder_model_9_block_1_conv_bias, dilations = input_53_dilations_0, groups = input_53_groups_0, pad = input_53_pad_0, pad_type = input_53_pad_type_0, strides = input_53_strides_0, weight = mimi_decoder_model_9_block_1_conv_weight, x = input_51)[name = tensor("input_53")]; + tensor var_697_begin_0 = const()[name = tensor("op_697_begin_0"), val = tensor([0, 0, 1920])]; + tensor var_697_end_0 = const()[name = tensor("op_697_end_0"), val = tensor([1, 64, 1922])]; + tensor var_697_end_mask_0 = const()[name = tensor("op_697_end_mask_0"), val = tensor([true, true, true])]; + tensor var_697 = slice_by_index(begin = var_697_begin_0, end = var_697_end_0, end_mask = var_697_end_mask_0, x = input_51)[name = tensor("op_697")]; + tensor x_29 = elu(alpha = var_507, x = input_53)[name = tensor("x_29")]; + tensor v_pad_type_0 = const()[name = tensor("v_pad_type_0"), val = tensor("valid")]; + tensor v_strides_0 = const()[name = tensor("v_strides_0"), val = tensor([1])]; + tensor v_pad_0 = const()[name = tensor("v_pad_0"), val = tensor([0, 0])]; + tensor v_dilations_0 = const()[name = tensor("v_dilations_0"), val = tensor([1])]; + tensor v_groups_0 = const()[name = tensor("v_groups_0"), val = tensor(1)]; + tensor v = conv(bias = mimi_decoder_model_9_block_3_conv_bias, dilations = v_dilations_0, groups = v_groups_0, pad = v_pad_0, pad_type = v_pad_type_0, strides = v_strides_0, weight = mimi_decoder_model_9_block_3_conv_weight, x = x_29)[name = tensor("v")]; + tensor input_55 = add(x = input_49, y = v)[name = tensor("input_55")]; + tensor x = elu(alpha = var_507, x = input_55)[name = tensor("x")]; + tensor input_interleave_0 = const()[name = tensor("input_interleave_0"), val = tensor(false)]; + tensor input = concat(axis = var_508, interleave = input_interleave_0, values = (conv_final_prev, x))[name = tensor("input")]; + tensor var_724_pad_type_0 = const()[name = tensor("op_724_pad_type_0"), val = tensor("valid")]; + tensor var_724_strides_0 = const()[name = tensor("op_724_strides_0"), val = tensor([1])]; + tensor var_724_pad_0 = const()[name = tensor("op_724_pad_0"), val = tensor([0, 0])]; + tensor var_724_dilations_0 = const()[name = tensor("op_724_dilations_0"), val = tensor([1])]; + tensor var_724_groups_0 = const()[name = tensor("op_724_groups_0"), val = tensor(1)]; + tensor var_724 = conv(bias = mimi_decoder_model_11_conv_bias, dilations = var_724_dilations_0, groups = var_724_groups_0, pad = var_724_pad_0, pad_type = var_724_pad_type_0, strides = var_724_strides_0, weight = mimi_decoder_model_11_conv_weight, x = input)[name = tensor("op_724")]; + tensor var_725_begin_0 = const()[name = tensor("op_725_begin_0"), val = tensor([0, 0, 1920])]; + tensor var_725_end_0 = const()[name = tensor("op_725_end_0"), val = tensor([1, 64, 1922])]; + tensor var_725_end_mask_0 = const()[name = tensor("op_725_end_mask_0"), val = tensor([true, true, true])]; + tensor var_725 = slice_by_index(begin = var_725_begin_0, end = var_725_end_0, end_mask = var_725_end_mask_0, x = input)[name = tensor("op_725")]; + tensor var_740_promoted = const()[name = tensor("op_740_promoted"), val = tensor(0x1p+4)]; + tensor var_741 = add(x = attn0_offset, y = var_740_promoted)[name = tensor("op_741")]; + tensor var_743_promoted = const()[name = tensor("op_743_promoted"), val = tensor(0x1p+4)]; + tensor var_744 = add(x = attn1_offset, y = var_743_promoted)[name = tensor("op_744")]; + tensor conv0_first_tmp = identity(x = conv0_first)[name = tensor("conv0_first_tmp")]; + tensor res0_conv0_first_tmp = identity(x = res0_conv0_first)[name = tensor("res0_conv0_first_tmp")]; + tensor res0_conv1_prev_tmp = identity(x = res0_conv1_prev)[name = tensor("res0_conv1_prev_tmp")]; + tensor res0_conv1_first_tmp = identity(x = res0_conv1_first)[name = tensor("res0_conv1_first_tmp")]; + tensor res1_conv0_first_tmp = identity(x = res1_conv0_first)[name = tensor("res1_conv0_first_tmp")]; + tensor res1_conv1_prev_tmp = identity(x = res1_conv1_prev)[name = tensor("res1_conv1_prev_tmp")]; + tensor res1_conv1_first_tmp = identity(x = res1_conv1_first)[name = tensor("res1_conv1_first_tmp")]; + tensor res2_conv0_first_tmp = identity(x = res2_conv0_first)[name = tensor("res2_conv0_first_tmp")]; + tensor res2_conv1_prev_tmp = identity(x = res2_conv1_prev)[name = tensor("res2_conv1_prev_tmp")]; + tensor res2_conv1_first_tmp = identity(x = res2_conv1_first)[name = tensor("res2_conv1_first_tmp")]; + tensor conv_final_first_tmp = identity(x = conv_final_first)[name = tensor("conv_final_first_tmp")]; + } -> (var_724, var_77, var_210, var_741, var_400, var_744, var_542, conv0_first, var_565, var_585, res0_conv0_first, res0_conv1_prev, res0_conv1_first, var_621, var_641, res1_conv0_first, res1_conv1_prev, res1_conv1_first, var_677, var_697, res2_conv0_first, res2_conv1_prev, res2_conv1_first, var_725, conv_final_first); +} \ No newline at end of file diff --git a/v2.1/english/mimi_decoder.mlmodelc/weights/weight.bin b/v2.1/english/mimi_decoder.mlmodelc/weights/weight.bin new file mode 100644 index 0000000000000000000000000000000000000000..190839671bb0332bcd8e7cc276be5dfe49c59320 --- /dev/null +++ b/v2.1/english/mimi_decoder.mlmodelc/weights/weight.bin @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:c93d182f9ad8c1042ec8302093aabb4e310d76616932dd12720e03b5d07f7c64 +size 41768256 diff --git a/v2.1/english/mimi_decoder.mlpackage/Data/com.apple.CoreML/model.mlmodel b/v2.1/english/mimi_decoder.mlpackage/Data/com.apple.CoreML/model.mlmodel new file mode 100644 index 0000000000000000000000000000000000000000..10ec1b130a0e006268e2a49ab7f34138623b5c31 --- /dev/null +++ b/v2.1/english/mimi_decoder.mlpackage/Data/com.apple.CoreML/model.mlmodel @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:8697d7a8b97c7ccc1ed77047a1de335d0d6648bb0dd9b8f3a941d8711e73f2c1 +size 93270 diff --git a/v2.1/english/mimi_decoder.mlpackage/Data/com.apple.CoreML/weights/weight.bin b/v2.1/english/mimi_decoder.mlpackage/Data/com.apple.CoreML/weights/weight.bin new file mode 100644 index 0000000000000000000000000000000000000000..190839671bb0332bcd8e7cc276be5dfe49c59320 --- /dev/null +++ b/v2.1/english/mimi_decoder.mlpackage/Data/com.apple.CoreML/weights/weight.bin @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:c93d182f9ad8c1042ec8302093aabb4e310d76616932dd12720e03b5d07f7c64 +size 41768256 diff --git a/v2.1/english/mimi_decoder.mlpackage/Manifest.json b/v2.1/english/mimi_decoder.mlpackage/Manifest.json new file mode 100644 index 0000000000000000000000000000000000000000..ea113ebd1a9ea27975299727adb546d87855bc71 --- /dev/null +++ b/v2.1/english/mimi_decoder.mlpackage/Manifest.json @@ -0,0 +1,18 @@ +{ + "fileFormatVersion": "1.0.0", + "itemInfoEntries": { + "5866D77B-1BEE-4A04-B630-3AED690E11D8": { + "author": "com.apple.CoreML", + "description": "CoreML Model Specification", + "name": "model.mlmodel", + "path": "com.apple.CoreML/model.mlmodel" + }, + "AD027508-2AB2-47F8-8219-42DAA70CD43E": { + "author": "com.apple.CoreML", + "description": "CoreML Model Weights", + "name": "weights", + "path": "com.apple.CoreML/weights" + } + }, + "rootModelIdentifier": "5866D77B-1BEE-4A04-B630-3AED690E11D8" +} diff --git a/v2.1/french_24l/cond_prefill.mlmodelc/analytics/coremldata.bin b/v2.1/french_24l/cond_prefill.mlmodelc/analytics/coremldata.bin new file mode 100644 index 0000000000000000000000000000000000000000..eb1d6c659fecc23af8b88d018bbac39bbde4fe3b --- /dev/null +++ b/v2.1/french_24l/cond_prefill.mlmodelc/analytics/coremldata.bin @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:133f57c0224610fb3a6ee728aa914049c20ca2ff6798778e4e3ccc25a0acb557 +size 243 diff --git a/v2.1/french_24l/cond_prefill.mlmodelc/coremldata.bin b/v2.1/french_24l/cond_prefill.mlmodelc/coremldata.bin new file mode 100644 index 0000000000000000000000000000000000000000..ae1b01d29e9d40d22fc361352af863b957d6805f --- /dev/null +++ b/v2.1/french_24l/cond_prefill.mlmodelc/coremldata.bin @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:851fd8269184aac7f2286d4439692c02f86ab392fe635ff45a39f8243e1fd44e +size 3519 diff --git a/v2.1/french_24l/cond_prefill.mlmodelc/model.mil b/v2.1/french_24l/cond_prefill.mlmodelc/model.mil new file mode 100644 index 0000000000000000000000000000000000000000..7028cd94630c33bb5143ac20b6b86ca016eb1d25 --- /dev/null +++ b/v2.1/french_24l/cond_prefill.mlmodelc/model.mil @@ -0,0 +1,5431 @@ +program(1.0) +[buildInfo = dict, tensor>({{"coremlc-component-MIL", "3520.4.1"}, {"coremlc-version", "3520.5.1"}, {"coremltools-component-torch", "2.12.0"}, {"coremltools-source-dialect", "TorchScript"}, {"coremltools-version", "9.0"}})] +{ + func main(tensor cache0, tensor cache1, tensor cache10, tensor cache11, tensor cache12, tensor cache13, tensor cache14, tensor cache15, tensor cache16, tensor cache17, tensor cache18, tensor cache19, tensor cache2, tensor cache20, tensor cache21, tensor cache22, tensor cache23, tensor cache3, tensor cache4, tensor cache5, tensor cache6, tensor cache7, tensor cache8, tensor cache9, tensor conditioning, tensor position0, tensor position1, tensor position10, tensor position11, tensor position12, tensor position13, tensor position14, tensor position15, tensor position16, tensor position17, tensor position18, tensor position19, tensor position2, tensor position20, tensor position21, tensor position22, tensor position23, tensor position3, tensor position4, tensor position5, tensor position6, tensor position7, tensor position8, tensor position9, tensor valid_len) { + tensor input_3_axes_0 = const()[name = tensor("input_3_axes_0"), val = tensor([-1])]; + tensor conditioning_to_fp16_dtype_0 = const()[name = tensor("conditioning_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor norm0_1_weight_to_fp16 = const()[name = tensor("norm0_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(64)))]; + tensor norm0_1_bias_to_fp16 = const()[name = tensor("norm0_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(2176)))]; + tensor var_192_to_fp16 = const()[name = tensor("op_192_to_fp16"), val = tensor(0x1.5p-17)]; + tensor conditioning_to_fp16 = cast(dtype = conditioning_to_fp16_dtype_0, x = conditioning)[name = tensor("cast_663")]; + tensor input_3_cast_fp16 = layer_norm(axes = input_3_axes_0, beta = norm0_1_bias_to_fp16, epsilon = var_192_to_fp16, gamma = norm0_1_weight_to_fp16, x = conditioning_to_fp16)[name = tensor("input_3_cast_fp16")]; + tensor attn0_in_proj_weight_to_fp16 = const()[name = tensor("attn0_in_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(4288)))]; + tensor linear_0_bias_0_to_fp16 = const()[name = tensor("linear_0_bias_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(6295808)))]; + tensor linear_0_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = attn0_in_proj_weight_to_fp16, x = input_3_cast_fp16)[name = tensor("linear_0_cast_fp16")]; + tensor var_206 = const()[name = tensor("op_206"), val = tensor([1, 256, 3, 16, 64])]; + tensor qkv_1_cast_fp16 = reshape(shape = var_206, x = linear_0_cast_fp16)[name = tensor("qkv_1_cast_fp16")]; + tensor q_1_begin_0 = const()[name = tensor("q_1_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor q_1_end_0 = const()[name = tensor("q_1_end_0"), val = tensor([1, 256, 1, 16, 64])]; + tensor q_1_end_mask_0 = const()[name = tensor("q_1_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor q_1_squeeze_mask_0 = const()[name = tensor("q_1_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor q_1_cast_fp16 = slice_by_index(begin = q_1_begin_0, end = q_1_end_0, end_mask = q_1_end_mask_0, squeeze_mask = q_1_squeeze_mask_0, x = qkv_1_cast_fp16)[name = tensor("q_1_cast_fp16")]; + tensor k_1_begin_0 = const()[name = tensor("k_1_begin_0"), val = tensor([0, 0, 1, 0, 0])]; + tensor k_1_end_0 = const()[name = tensor("k_1_end_0"), val = tensor([1, 256, 2, 16, 64])]; + tensor k_1_end_mask_0 = const()[name = tensor("k_1_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor k_1_squeeze_mask_0 = const()[name = tensor("k_1_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor k_1_cast_fp16 = slice_by_index(begin = k_1_begin_0, end = k_1_end_0, end_mask = k_1_end_mask_0, squeeze_mask = k_1_squeeze_mask_0, x = qkv_1_cast_fp16)[name = tensor("k_1_cast_fp16")]; + tensor v_1_begin_0 = const()[name = tensor("v_1_begin_0"), val = tensor([0, 0, 2, 0, 0])]; + tensor v_1_end_0 = const()[name = tensor("v_1_end_0"), val = tensor([1, 256, 3, 16, 64])]; + tensor v_1_end_mask_0 = const()[name = tensor("v_1_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor v_1_squeeze_mask_0 = const()[name = tensor("v_1_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor v_1_cast_fp16 = slice_by_index(begin = v_1_begin_0, end = v_1_end_0, end_mask = v_1_end_mask_0, squeeze_mask = v_1_squeeze_mask_0, x = qkv_1_cast_fp16)[name = tensor("v_1_cast_fp16")]; + tensor var_264 = const()[name = tensor("op_264"), val = tensor([1, 1])]; + tensor position0_to_fp16_dtype_0 = const()[name = tensor("position0_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor position0_to_fp16 = cast(dtype = position0_to_fp16_dtype_0, x = position0)[name = tensor("cast_662")]; + tensor var_265_cast_fp16 = reshape(shape = var_264, x = position0_to_fp16)[name = tensor("op_265_cast_fp16")]; + tensor ts_1_promoted_to_fp16 = const()[name = tensor("ts_1_promoted_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(6302016)))]; + tensor ts_3_cast_fp16 = add(x = ts_1_promoted_to_fp16, y = var_265_cast_fp16)[name = tensor("ts_3_cast_fp16")]; + tensor var_272 = const()[name = tensor("op_272"), val = tensor([1, 256, 1, 1])]; + tensor ts_5_cast_fp16 = reshape(shape = var_272, x = ts_3_cast_fp16)[name = tensor("ts_5_cast_fp16")]; + tensor var_279 = const()[name = tensor("op_279"), val = tensor([1, 256, 16, 32, 2])]; + tensor q_complex_1_cast_fp16 = reshape(shape = var_279, x = q_1_cast_fp16)[name = tensor("q_complex_1_cast_fp16")]; + tensor var_286 = const()[name = tensor("op_286"), val = tensor([1, 256, 16, 32, 2])]; + tensor k_complex_1_cast_fp16 = reshape(shape = var_286, x = k_1_cast_fp16)[name = tensor("k_complex_1_cast_fp16")]; + tensor var_290_begin_0 = const()[name = tensor("op_290_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_290_end_0 = const()[name = tensor("op_290_end_0"), val = tensor([1, 256, 16, 32, 1])]; + tensor var_290_end_mask_0 = const()[name = tensor("op_290_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_290_squeeze_mask_0 = const()[name = tensor("op_290_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_290_cast_fp16 = slice_by_index(begin = var_290_begin_0, end = var_290_end_0, end_mask = var_290_end_mask_0, squeeze_mask = var_290_squeeze_mask_0, x = q_complex_1_cast_fp16)[name = tensor("op_290_cast_fp16")]; + tensor var_298_begin_0 = const()[name = tensor("op_298_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_298_end_0 = const()[name = tensor("op_298_end_0"), val = tensor([1, 256, 16, 32, 2])]; + tensor var_298_end_mask_0 = const()[name = tensor("op_298_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_298_squeeze_mask_0 = const()[name = tensor("op_298_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_298_cast_fp16 = slice_by_index(begin = var_298_begin_0, end = var_298_end_0, end_mask = var_298_end_mask_0, squeeze_mask = var_298_squeeze_mask_0, x = q_complex_1_cast_fp16)[name = tensor("op_298_cast_fp16")]; + tensor var_306_begin_0 = const()[name = tensor("op_306_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_306_end_0 = const()[name = tensor("op_306_end_0"), val = tensor([1, 256, 16, 32, 1])]; + tensor var_306_end_mask_0 = const()[name = tensor("op_306_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_306_squeeze_mask_0 = const()[name = tensor("op_306_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_306_cast_fp16 = slice_by_index(begin = var_306_begin_0, end = var_306_end_0, end_mask = var_306_end_mask_0, squeeze_mask = var_306_squeeze_mask_0, x = k_complex_1_cast_fp16)[name = tensor("op_306_cast_fp16")]; + tensor var_314_begin_0 = const()[name = tensor("op_314_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_314_end_0 = const()[name = tensor("op_314_end_0"), val = tensor([1, 256, 16, 32, 2])]; + tensor var_314_end_mask_0 = const()[name = tensor("op_314_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_314_squeeze_mask_0 = const()[name = tensor("op_314_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_314_cast_fp16 = slice_by_index(begin = var_314_begin_0, end = var_314_end_0, end_mask = var_314_end_mask_0, squeeze_mask = var_314_squeeze_mask_0, x = k_complex_1_cast_fp16)[name = tensor("op_314_cast_fp16")]; + tensor freqs_1_to_fp16 = const()[name = tensor("freqs_1_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(6302592)))]; + tensor var_320_cast_fp16 = mul(x = freqs_1_to_fp16, y = ts_5_cast_fp16)[name = tensor("op_320_cast_fp16")]; + tensor rotr_1_cast_fp16 = cos(x = var_320_cast_fp16)[name = tensor("rotr_1_cast_fp16")]; + tensor roti_1_cast_fp16 = sin(x = var_320_cast_fp16)[name = tensor("roti_1_cast_fp16")]; + tensor var_324_cast_fp16 = mul(x = var_290_cast_fp16, y = rotr_1_cast_fp16)[name = tensor("op_324_cast_fp16")]; + tensor var_325_cast_fp16 = mul(x = var_298_cast_fp16, y = roti_1_cast_fp16)[name = tensor("op_325_cast_fp16")]; + tensor qor_1_cast_fp16 = sub(x = var_324_cast_fp16, y = var_325_cast_fp16)[name = tensor("qor_1_cast_fp16")]; + tensor var_328_cast_fp16 = mul(x = var_290_cast_fp16, y = roti_1_cast_fp16)[name = tensor("op_328_cast_fp16")]; + tensor var_329_cast_fp16 = mul(x = var_298_cast_fp16, y = rotr_1_cast_fp16)[name = tensor("op_329_cast_fp16")]; + tensor qoi_1_cast_fp16 = add(x = var_328_cast_fp16, y = var_329_cast_fp16)[name = tensor("qoi_1_cast_fp16")]; + tensor var_332_cast_fp16 = mul(x = var_306_cast_fp16, y = rotr_1_cast_fp16)[name = tensor("op_332_cast_fp16")]; + tensor var_333_cast_fp16 = mul(x = var_314_cast_fp16, y = roti_1_cast_fp16)[name = tensor("op_333_cast_fp16")]; + tensor kor_1_cast_fp16 = sub(x = var_332_cast_fp16, y = var_333_cast_fp16)[name = tensor("kor_1_cast_fp16")]; + tensor var_336_cast_fp16 = mul(x = var_306_cast_fp16, y = roti_1_cast_fp16)[name = tensor("op_336_cast_fp16")]; + tensor var_337_cast_fp16 = mul(x = var_314_cast_fp16, y = rotr_1_cast_fp16)[name = tensor("op_337_cast_fp16")]; + tensor koi_1_cast_fp16 = add(x = var_336_cast_fp16, y = var_337_cast_fp16)[name = tensor("koi_1_cast_fp16")]; + tensor qo_1_axis_0 = const()[name = tensor("qo_1_axis_0"), val = tensor(-1)]; + tensor qo_1_cast_fp16 = stack(axis = qo_1_axis_0, values = (qor_1_cast_fp16, qoi_1_cast_fp16))[name = tensor("qo_1_cast_fp16")]; + tensor ko_1_axis_0 = const()[name = tensor("ko_1_axis_0"), val = tensor(-1)]; + tensor ko_1_cast_fp16 = stack(axis = ko_1_axis_0, values = (kor_1_cast_fp16, koi_1_cast_fp16))[name = tensor("ko_1_cast_fp16")]; + tensor var_370 = const()[name = tensor("op_370"), val = tensor([1, 256, 16, 64])]; + tensor q_3_cast_fp16 = reshape(shape = var_370, x = qo_1_cast_fp16)[name = tensor("q_3_cast_fp16")]; + tensor var_376 = const()[name = tensor("op_376"), val = tensor([1, 256, 16, 64])]; + tensor k_3_cast_fp16 = reshape(shape = var_376, x = ko_1_cast_fp16)[name = tensor("k_3_cast_fp16")]; + tensor var_393_promoted_to_fp16 = const()[name = tensor("op_393_promoted_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(6302720)))]; + tensor write_indices_float_1_cast_fp16 = add(x = var_265_cast_fp16, y = var_393_promoted_to_fp16)[name = tensor("write_indices_float_1_cast_fp16")]; + tensor _inversed_397_y_0_to_fp16 = const()[name = tensor("_inversed_397_y_0_to_fp16"), val = tensor(0x1p-9)]; + tensor _inversed_397_cast_fp16 = mul(x = write_indices_float_1_cast_fp16, y = _inversed_397_y_0_to_fp16)[name = tensor("_inversed_397_cast_fp16")]; + tensor var_398_cast_fp16 = floor(x = _inversed_397_cast_fp16)[name = tensor("op_398_cast_fp16")]; + tensor var_399_to_fp16 = const()[name = tensor("op_399_to_fp16"), val = tensor(0x1p+9)]; + tensor var_400_cast_fp16 = mul(x = var_398_cast_fp16, y = var_399_to_fp16)[name = tensor("op_400_cast_fp16")]; + tensor write_indices_float_3_cast_fp16 = sub(x = write_indices_float_1_cast_fp16, y = var_400_cast_fp16)[name = tensor("write_indices_float_3_cast_fp16")]; + tensor var_409 = const()[name = tensor("op_409"), val = tensor([1, 1])]; + tensor valid_len_to_fp16_dtype_0 = const()[name = tensor("valid_len_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor valid_len_to_fp16 = cast(dtype = valid_len_to_fp16_dtype_0, x = valid_len)[name = tensor("cast_661")]; + tensor var_410_cast_fp16 = reshape(shape = var_409, x = valid_len_to_fp16)[name = tensor("op_410_cast_fp16")]; + tensor token_valid_1_cast_fp16 = less(x = var_393_promoted_to_fp16, y = var_410_cast_fp16)[name = tensor("token_valid_1_cast_fp16")]; + tensor var_418_value_0_to_fp16 = const()[name = tensor("op_418_value_0_to_fp16"), val = tensor(0x1.ffp+8)]; + tensor var_418_cast_fp16 = fill_like(ref_tensor = write_indices_float_3_cast_fp16, value = var_418_value_0_to_fp16)[name = tensor("op_418_cast_fp16")]; + tensor write_indices_float_5_cast_fp16 = select(a = write_indices_float_3_cast_fp16, b = var_418_cast_fp16, cond = token_valid_1_cast_fp16)[name = tensor("write_indices_float_5_cast_fp16")]; + tensor var_424_dtype_0 = const()[name = tensor("op_424_dtype_0"), val = tensor("int32")]; + tensor var_429 = const()[name = tensor("op_429"), val = tensor([1, 256, 1, 1])]; + tensor write_indices_float_5_cast_fp16_to_int32 = cast(dtype = var_424_dtype_0, x = write_indices_float_5_cast_fp16)[name = tensor("cast_660")]; + tensor var_430 = reshape(shape = var_429, x = write_indices_float_5_cast_fp16_to_int32)[name = tensor("op_430")]; + tensor write_indices_1_reps_0 = const()[name = tensor("write_indices_1_reps_0"), val = tensor([1, 1, 16, 64])]; + tensor write_indices_1 = tile(reps = write_indices_1_reps_0, x = var_430)[name = tensor("write_indices_1")]; + tensor var_440_begin_0 = const()[name = tensor("op_440_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_440_end_0 = const()[name = tensor("op_440_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor var_440_end_mask_0 = const()[name = tensor("op_440_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_440_squeeze_mask_0 = const()[name = tensor("op_440_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor cache0_to_fp16_dtype_0 = const()[name = tensor("cache0_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor cache0_to_fp16 = cast(dtype = cache0_to_fp16_dtype_0, x = cache0)[name = tensor("cast_659")]; + tensor var_440_cast_fp16 = slice_by_index(begin = var_440_begin_0, end = var_440_end_0, end_mask = var_440_end_mask_0, squeeze_mask = var_440_squeeze_mask_0, x = cache0_to_fp16)[name = tensor("op_440_cast_fp16")]; + tensor var_442_axis_0 = const()[name = tensor("op_442_axis_0"), val = tensor(1)]; + tensor var_442_mode_0 = const()[name = tensor("op_442_mode_0"), val = tensor("update")]; + tensor var_442_validate_indices_0 = const()[name = tensor("op_442_validate_indices_0"), val = tensor(false)]; + tensor var_442_cast_fp16 = scatter_along_axis(axis = var_442_axis_0, data = var_440_cast_fp16, indices = write_indices_1, mode = var_442_mode_0, updates = k_3_cast_fp16, validate_indices = var_442_validate_indices_0)[name = tensor("op_442_cast_fp16")]; + tensor concat_2 = const()[name = tensor("concat_2"), val = tensor([0, 0, 0, 0, 0])]; + tensor concat_3 = const()[name = tensor("concat_3"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_1_internal_tensor_assign_1_stride_0 = const()[name = tensor("new_cache_1_internal_tensor_assign_1_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_1_internal_tensor_assign_1_begin_mask_0 = const()[name = tensor("new_cache_1_internal_tensor_assign_1_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_1_internal_tensor_assign_1_end_mask_0 = const()[name = tensor("new_cache_1_internal_tensor_assign_1_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_1_internal_tensor_assign_1_squeeze_mask_0 = const()[name = tensor("new_cache_1_internal_tensor_assign_1_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_46 = const()[name = tensor("shape_46"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_0 = const()[name = tensor("reduce_prod_0"), val = tensor(1048576)]; + tensor range_1d_0_start_0 = const()[name = tensor("range_1d_0_start_0"), val = tensor(0)]; + tensor range_1d_0_step_0 = const()[name = tensor("range_1d_0_step_0"), val = tensor(1)]; + tensor range_1d_0 = range_1d(end = reduce_prod_0, start = range_1d_0_start_0, step = range_1d_0_step_0)[name = tensor("range_1d_0")]; + tensor reshape_0 = reshape(shape = shape_46, x = range_1d_0)[name = tensor("reshape_0")]; + tensor slice_by_index_0 = slice_by_index(begin = concat_2, begin_mask = new_cache_1_internal_tensor_assign_1_begin_mask_0, end = concat_3, end_mask = new_cache_1_internal_tensor_assign_1_end_mask_0, squeeze_mask = new_cache_1_internal_tensor_assign_1_squeeze_mask_0, stride = new_cache_1_internal_tensor_assign_1_stride_0, x = reshape_0)[name = tensor("slice_by_index_0")]; + tensor reshape_1_shape_0 = const()[name = tensor("reshape_1_shape_0"), val = tensor([-1])]; + tensor reshape_1 = reshape(shape = reshape_1_shape_0, x = slice_by_index_0)[name = tensor("reshape_1")]; + tensor reshape_2_shape_0 = const()[name = tensor("reshape_2_shape_0"), val = tensor([-1])]; + tensor reshape_2_cast_fp16 = reshape(shape = reshape_2_shape_0, x = var_442_cast_fp16)[name = tensor("reshape_2_cast_fp16")]; + tensor reshape_3_shape_0 = const()[name = tensor("reshape_3_shape_0"), val = tensor([-1])]; + tensor reshape_3_cast_fp16 = reshape(shape = reshape_3_shape_0, x = cache0_to_fp16)[name = tensor("reshape_3_cast_fp16")]; + tensor scatter_0_mode_0 = const()[name = tensor("scatter_0_mode_0"), val = tensor("update")]; + tensor scatter_0_axis_0 = const()[name = tensor("scatter_0_axis_0"), val = tensor(0)]; + tensor scatter_0_validate_indices_0 = const()[name = tensor("scatter_0_validate_indices_0"), val = tensor(false)]; + tensor scatter_0_cast_fp16 = scatter(axis = scatter_0_axis_0, data = reshape_3_cast_fp16, indices = reshape_1, mode = scatter_0_mode_0, updates = reshape_2_cast_fp16, validate_indices = scatter_0_validate_indices_0)[name = tensor("scatter_0_cast_fp16")]; + tensor reshape_4_cast_fp16 = reshape(shape = shape_46, x = scatter_0_cast_fp16)[name = tensor("reshape_4_cast_fp16")]; + tensor var_450_begin_0 = const()[name = tensor("op_450_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor var_450_end_0 = const()[name = tensor("op_450_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor var_450_end_mask_0 = const()[name = tensor("op_450_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_450_squeeze_mask_0 = const()[name = tensor("op_450_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor var_450_cast_fp16 = slice_by_index(begin = var_450_begin_0, end = var_450_end_0, end_mask = var_450_end_mask_0, squeeze_mask = var_450_squeeze_mask_0, x = reshape_4_cast_fp16)[name = tensor("op_450_cast_fp16")]; + tensor var_452_axis_0 = const()[name = tensor("op_452_axis_0"), val = tensor(1)]; + tensor var_452_mode_0 = const()[name = tensor("op_452_mode_0"), val = tensor("update")]; + tensor var_452_validate_indices_0 = const()[name = tensor("op_452_validate_indices_0"), val = tensor(false)]; + tensor var_452_cast_fp16 = scatter_along_axis(axis = var_452_axis_0, data = var_450_cast_fp16, indices = write_indices_1, mode = var_452_mode_0, updates = v_1_cast_fp16, validate_indices = var_452_validate_indices_0)[name = tensor("op_452_cast_fp16")]; + tensor concat_4 = const()[name = tensor("concat_4"), val = tensor([1, 0, 0, 0, 0])]; + tensor concat_5 = const()[name = tensor("concat_5"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_1_internal_tensor_assign_2_stride_0 = const()[name = tensor("new_cache_1_internal_tensor_assign_2_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_1_internal_tensor_assign_2_begin_mask_0 = const()[name = tensor("new_cache_1_internal_tensor_assign_2_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_1_internal_tensor_assign_2_end_mask_0 = const()[name = tensor("new_cache_1_internal_tensor_assign_2_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_1_internal_tensor_assign_2_squeeze_mask_0 = const()[name = tensor("new_cache_1_internal_tensor_assign_2_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_47 = const()[name = tensor("shape_47"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_1 = const()[name = tensor("reduce_prod_1"), val = tensor(1048576)]; + tensor range_1d_1_start_0 = const()[name = tensor("range_1d_1_start_0"), val = tensor(0)]; + tensor range_1d_1_step_0 = const()[name = tensor("range_1d_1_step_0"), val = tensor(1)]; + tensor range_1d_1 = range_1d(end = reduce_prod_1, start = range_1d_1_start_0, step = range_1d_1_step_0)[name = tensor("range_1d_1")]; + tensor reshape_5 = reshape(shape = shape_47, x = range_1d_1)[name = tensor("reshape_5")]; + tensor slice_by_index_1 = slice_by_index(begin = concat_4, begin_mask = new_cache_1_internal_tensor_assign_2_begin_mask_0, end = concat_5, end_mask = new_cache_1_internal_tensor_assign_2_end_mask_0, squeeze_mask = new_cache_1_internal_tensor_assign_2_squeeze_mask_0, stride = new_cache_1_internal_tensor_assign_2_stride_0, x = reshape_5)[name = tensor("slice_by_index_1")]; + tensor reshape_6_shape_0 = const()[name = tensor("reshape_6_shape_0"), val = tensor([-1])]; + tensor reshape_6 = reshape(shape = reshape_6_shape_0, x = slice_by_index_1)[name = tensor("reshape_6")]; + tensor reshape_7_shape_0 = const()[name = tensor("reshape_7_shape_0"), val = tensor([-1])]; + tensor reshape_7_cast_fp16 = reshape(shape = reshape_7_shape_0, x = var_452_cast_fp16)[name = tensor("reshape_7_cast_fp16")]; + tensor reshape_8_shape_0 = const()[name = tensor("reshape_8_shape_0"), val = tensor([-1])]; + tensor reshape_8_cast_fp16 = reshape(shape = reshape_8_shape_0, x = reshape_4_cast_fp16)[name = tensor("reshape_8_cast_fp16")]; + tensor scatter_1_mode_0 = const()[name = tensor("scatter_1_mode_0"), val = tensor("update")]; + tensor scatter_1_axis_0 = const()[name = tensor("scatter_1_axis_0"), val = tensor(0)]; + tensor scatter_1_validate_indices_0 = const()[name = tensor("scatter_1_validate_indices_0"), val = tensor(false)]; + tensor scatter_1_cast_fp16 = scatter(axis = scatter_1_axis_0, data = reshape_8_cast_fp16, indices = reshape_6, mode = scatter_1_mode_0, updates = reshape_7_cast_fp16, validate_indices = scatter_1_validate_indices_0)[name = tensor("scatter_1_cast_fp16")]; + tensor reshape_9_cast_fp16 = reshape(shape = shape_47, x = scatter_1_cast_fp16)[name = tensor("reshape_9_cast_fp16")]; + tensor reshape_9_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("reshape_9_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor keys_1_begin_0 = const()[name = tensor("keys_1_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor keys_1_end_0 = const()[name = tensor("keys_1_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor keys_1_end_mask_0 = const()[name = tensor("keys_1_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor keys_1_squeeze_mask_0 = const()[name = tensor("keys_1_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor keys_1_cast_fp16 = slice_by_index(begin = keys_1_begin_0, end = keys_1_end_0, end_mask = keys_1_end_mask_0, squeeze_mask = keys_1_squeeze_mask_0, x = reshape_9_cast_fp16)[name = tensor("keys_1_cast_fp16")]; + tensor values_1_begin_0 = const()[name = tensor("values_1_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor values_1_end_0 = const()[name = tensor("values_1_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor values_1_end_mask_0 = const()[name = tensor("values_1_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor values_1_squeeze_mask_0 = const()[name = tensor("values_1_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor values_1_cast_fp16 = slice_by_index(begin = values_1_begin_0, end = values_1_end_0, end_mask = values_1_end_mask_0, squeeze_mask = values_1_squeeze_mask_0, x = reshape_9_cast_fp16)[name = tensor("values_1_cast_fp16")]; + tensor var_464_cast_fp16 = not_equal(x = keys_1_cast_fp16, y = keys_1_cast_fp16)[name = tensor("op_464_cast_fp16")]; + tensor var_470_to_fp16 = const()[name = tensor("op_470_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(6303296)))]; + tensor keys_3_cast_fp16 = select(a = var_470_to_fp16, b = keys_1_cast_fp16, cond = var_464_cast_fp16)[name = tensor("keys_3_cast_fp16")]; + tensor var_472_cast_fp16 = not_equal(x = values_1_cast_fp16, y = values_1_cast_fp16)[name = tensor("op_472_cast_fp16")]; + tensor values_3_cast_fp16 = select(a = var_470_to_fp16, b = values_1_cast_fp16, cond = var_472_cast_fp16)[name = tensor("values_3_cast_fp16")]; + tensor var_496 = const()[name = tensor("op_496"), val = tensor([0, 2, 1, 3])]; + tensor var_512 = const()[name = tensor("op_512"), val = tensor([1, 1, 1])]; + tensor var_513_cast_fp16 = reshape(shape = var_512, x = position0_to_fp16)[name = tensor("op_513_cast_fp16")]; + tensor q_offsets_1_promoted_to_fp16 = const()[name = tensor("q_offsets_1_promoted_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(7351936)))]; + tensor q_positions_1_cast_fp16 = add(x = var_513_cast_fp16, y = q_offsets_1_promoted_to_fp16)[name = tensor("q_positions_1_cast_fp16")]; + tensor var_535 = const()[name = tensor("op_535"), val = tensor([1, 1, 1])]; + tensor var_536_cast_fp16 = reshape(shape = var_535, x = valid_len_to_fp16)[name = tensor("op_536_cast_fp16")]; + tensor valid_end_1_cast_fp16 = add(x = var_513_cast_fp16, y = var_536_cast_fp16)[name = tensor("valid_end_1_cast_fp16")]; + tensor k_positions_1_promoted_to_fp16 = const()[name = tensor("k_positions_1_promoted_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(7352512)))]; + tensor valid_mask_1_cast_fp16 = less(x = k_positions_1_promoted_to_fp16, y = valid_end_1_cast_fp16)[name = tensor("valid_mask_1_cast_fp16")]; + tensor causal_mask_1_cast_fp16 = less_equal(x = k_positions_1_promoted_to_fp16, y = q_positions_1_cast_fp16)[name = tensor("causal_mask_1_cast_fp16")]; + tensor var_541 = logical_and(x = valid_mask_1_cast_fp16, y = causal_mask_1_cast_fp16)[name = tensor("op_541")]; + tensor attn_mask_1_axes_0 = const()[name = tensor("attn_mask_1_axes_0"), val = tensor([1])]; + tensor attn_mask_1 = expand_dims(axes = attn_mask_1_axes_0, x = var_541)[name = tensor("attn_mask_1")]; + tensor var_555_transpose_x_0 = const()[name = tensor("op_555_transpose_x_0"), val = tensor(false)]; + tensor var_555_transpose_y_0 = const()[name = tensor("op_555_transpose_y_0"), val = tensor(false)]; + tensor transpose_69_perm_0 = const()[name = tensor("transpose_69_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_70_perm_0 = const()[name = tensor("transpose_70_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_70 = transpose(perm = transpose_70_perm_0, x = keys_3_cast_fp16)[name = tensor("transpose_204")]; + tensor transpose_69 = transpose(perm = transpose_69_perm_0, x = q_3_cast_fp16)[name = tensor("transpose_205")]; + tensor var_555_cast_fp16 = matmul(transpose_x = var_555_transpose_x_0, transpose_y = var_555_transpose_y_0, x = transpose_69, y = transpose_70)[name = tensor("op_555_cast_fp16")]; + tensor var_549_to_fp16 = const()[name = tensor("op_549_to_fp16"), val = tensor([0x1p-3])]; + tensor attn_weights_1_cast_fp16 = mul(x = var_555_cast_fp16, y = var_549_to_fp16)[name = tensor("attn_weights_1_cast_fp16")]; + tensor var_557 = logical_not(x = attn_mask_1)[name = tensor("op_557")]; + tensor var_558_to_fp16 = const()[name = tensor("op_558_to_fp16"), val = tensor(-inf)]; + tensor attn_weights_3_cast_fp16 = select(a = var_558_to_fp16, b = attn_weights_1_cast_fp16, cond = var_557)[name = tensor("attn_weights_3_cast_fp16")]; + tensor var_560 = const()[name = tensor("op_560"), val = tensor(-1)]; + tensor attn_weights_5_cast_fp16 = softmax(axis = var_560, x = attn_weights_3_cast_fp16)[name = tensor("attn_weights_5_cast_fp16")]; + tensor attn_output_1_transpose_x_0 = const()[name = tensor("attn_output_1_transpose_x_0"), val = tensor(false)]; + tensor attn_output_1_transpose_y_0 = const()[name = tensor("attn_output_1_transpose_y_0"), val = tensor(false)]; + tensor values_5_cast_fp16 = transpose(perm = var_496, x = values_3_cast_fp16)[name = tensor("transpose_206")]; + tensor attn_output_1_cast_fp16 = matmul(transpose_x = attn_output_1_transpose_x_0, transpose_y = attn_output_1_transpose_y_0, x = attn_weights_5_cast_fp16, y = values_5_cast_fp16)[name = tensor("attn_output_1_cast_fp16")]; + tensor var_568 = const()[name = tensor("op_568"), val = tensor([0, 2, 1, 3])]; + tensor var_573 = const()[name = tensor("op_573"), val = tensor([1, 256, 1024])]; + tensor var_569_cast_fp16 = transpose(perm = var_568, x = attn_output_1_cast_fp16)[name = tensor("transpose_203")]; + tensor input_5_cast_fp16 = reshape(shape = var_573, x = var_569_cast_fp16)[name = tensor("input_5_cast_fp16")]; + tensor attn0_out_proj_weight_to_fp16 = const()[name = tensor("attn0_out_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(7353600)))]; + tensor linear_1_bias_0_to_fp16 = const()[name = tensor("linear_1_bias_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9450816)))]; + tensor linear_1_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = attn0_out_proj_weight_to_fp16, x = input_5_cast_fp16)[name = tensor("linear_1_cast_fp16")]; + tensor var_582_cast_fp16 = add(x = position0_to_fp16, y = valid_len_to_fp16)[name = tensor("op_582_cast_fp16")]; + tensor var_582_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_582_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor input_7_cast_fp16 = add(x = conditioning_to_fp16, y = linear_1_cast_fp16)[name = tensor("input_7_cast_fp16")]; + tensor input_9_axes_0 = const()[name = tensor("input_9_axes_0"), val = tensor([-1])]; + tensor norm0_2_weight_to_fp16 = const()[name = tensor("norm0_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9452928)))]; + tensor norm0_2_bias_to_fp16 = const()[name = tensor("norm0_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9455040)))]; + tensor var_586_to_fp16 = const()[name = tensor("op_586_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_9_cast_fp16 = layer_norm(axes = input_9_axes_0, beta = norm0_2_bias_to_fp16, epsilon = var_586_to_fp16, gamma = norm0_2_weight_to_fp16, x = input_7_cast_fp16)[name = tensor("input_9_cast_fp16")]; + tensor linear0_1_weight_to_fp16 = const()[name = tensor("linear0_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9457152)))]; + tensor linear_2_bias_0_to_fp16 = const()[name = tensor("linear_2_bias_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(17845824)))]; + tensor linear_2_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = linear0_1_weight_to_fp16, x = input_9_cast_fp16)[name = tensor("linear_2_cast_fp16")]; + tensor input_11_mode_0 = const()[name = tensor("input_11_mode_0"), val = tensor("EXACT")]; + tensor input_11_cast_fp16 = gelu(mode = input_11_mode_0, x = linear_2_cast_fp16)[name = tensor("input_11_cast_fp16")]; + tensor linear0_2_weight_to_fp16 = const()[name = tensor("linear0_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(17854080)))]; + tensor linear_3_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = linear0_2_weight_to_fp16, x = input_11_cast_fp16)[name = tensor("linear_3_cast_fp16")]; + tensor input_13_cast_fp16 = add(x = input_7_cast_fp16, y = linear_3_cast_fp16)[name = tensor("input_13_cast_fp16")]; + tensor input_15_axes_0 = const()[name = tensor("input_15_axes_0"), val = tensor([-1])]; + tensor norm1_1_weight_to_fp16 = const()[name = tensor("norm1_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(26242752)))]; + tensor norm1_1_bias_to_fp16 = const()[name = tensor("norm1_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(26244864)))]; + tensor var_603_to_fp16 = const()[name = tensor("op_603_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_15_cast_fp16 = layer_norm(axes = input_15_axes_0, beta = norm1_1_bias_to_fp16, epsilon = var_603_to_fp16, gamma = norm1_1_weight_to_fp16, x = input_13_cast_fp16)[name = tensor("input_15_cast_fp16")]; + tensor attn1_in_proj_weight_to_fp16 = const()[name = tensor("attn1_in_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(26246976)))]; + tensor linear_4_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = attn1_in_proj_weight_to_fp16, x = input_15_cast_fp16)[name = tensor("linear_4_cast_fp16")]; + tensor var_617 = const()[name = tensor("op_617"), val = tensor([1, 256, 3, 16, 64])]; + tensor qkv_3_cast_fp16 = reshape(shape = var_617, x = linear_4_cast_fp16)[name = tensor("qkv_3_cast_fp16")]; + tensor q_7_begin_0 = const()[name = tensor("q_7_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor q_7_end_0 = const()[name = tensor("q_7_end_0"), val = tensor([1, 256, 1, 16, 64])]; + tensor q_7_end_mask_0 = const()[name = tensor("q_7_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor q_7_squeeze_mask_0 = const()[name = tensor("q_7_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor q_7_cast_fp16 = slice_by_index(begin = q_7_begin_0, end = q_7_end_0, end_mask = q_7_end_mask_0, squeeze_mask = q_7_squeeze_mask_0, x = qkv_3_cast_fp16)[name = tensor("q_7_cast_fp16")]; + tensor k_5_begin_0 = const()[name = tensor("k_5_begin_0"), val = tensor([0, 0, 1, 0, 0])]; + tensor k_5_end_0 = const()[name = tensor("k_5_end_0"), val = tensor([1, 256, 2, 16, 64])]; + tensor k_5_end_mask_0 = const()[name = tensor("k_5_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor k_5_squeeze_mask_0 = const()[name = tensor("k_5_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor k_5_cast_fp16 = slice_by_index(begin = k_5_begin_0, end = k_5_end_0, end_mask = k_5_end_mask_0, squeeze_mask = k_5_squeeze_mask_0, x = qkv_3_cast_fp16)[name = tensor("k_5_cast_fp16")]; + tensor v_3_begin_0 = const()[name = tensor("v_3_begin_0"), val = tensor([0, 0, 2, 0, 0])]; + tensor v_3_end_0 = const()[name = tensor("v_3_end_0"), val = tensor([1, 256, 3, 16, 64])]; + tensor v_3_end_mask_0 = const()[name = tensor("v_3_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor v_3_squeeze_mask_0 = const()[name = tensor("v_3_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor v_3_cast_fp16 = slice_by_index(begin = v_3_begin_0, end = v_3_end_0, end_mask = v_3_end_mask_0, squeeze_mask = v_3_squeeze_mask_0, x = qkv_3_cast_fp16)[name = tensor("v_3_cast_fp16")]; + tensor var_675 = const()[name = tensor("op_675"), val = tensor([1, 1])]; + tensor position1_to_fp16_dtype_0 = const()[name = tensor("position1_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor position1_to_fp16 = cast(dtype = position1_to_fp16_dtype_0, x = position1)[name = tensor("cast_656")]; + tensor var_676_cast_fp16 = reshape(shape = var_675, x = position1_to_fp16)[name = tensor("op_676_cast_fp16")]; + tensor ts_9_cast_fp16 = add(x = ts_1_promoted_to_fp16, y = var_676_cast_fp16)[name = tensor("ts_9_cast_fp16")]; + tensor var_683 = const()[name = tensor("op_683"), val = tensor([1, 256, 1, 1])]; + tensor ts_11_cast_fp16 = reshape(shape = var_683, x = ts_9_cast_fp16)[name = tensor("ts_11_cast_fp16")]; + tensor var_690 = const()[name = tensor("op_690"), val = tensor([1, 256, 16, 32, 2])]; + tensor q_complex_3_cast_fp16 = reshape(shape = var_690, x = q_7_cast_fp16)[name = tensor("q_complex_3_cast_fp16")]; + tensor var_697 = const()[name = tensor("op_697"), val = tensor([1, 256, 16, 32, 2])]; + tensor k_complex_3_cast_fp16 = reshape(shape = var_697, x = k_5_cast_fp16)[name = tensor("k_complex_3_cast_fp16")]; + tensor var_701_begin_0 = const()[name = tensor("op_701_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_701_end_0 = const()[name = tensor("op_701_end_0"), val = tensor([1, 256, 16, 32, 1])]; + tensor var_701_end_mask_0 = const()[name = tensor("op_701_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_701_squeeze_mask_0 = const()[name = tensor("op_701_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_701_cast_fp16 = slice_by_index(begin = var_701_begin_0, end = var_701_end_0, end_mask = var_701_end_mask_0, squeeze_mask = var_701_squeeze_mask_0, x = q_complex_3_cast_fp16)[name = tensor("op_701_cast_fp16")]; + tensor var_709_begin_0 = const()[name = tensor("op_709_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_709_end_0 = const()[name = tensor("op_709_end_0"), val = tensor([1, 256, 16, 32, 2])]; + tensor var_709_end_mask_0 = const()[name = tensor("op_709_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_709_squeeze_mask_0 = const()[name = tensor("op_709_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_709_cast_fp16 = slice_by_index(begin = var_709_begin_0, end = var_709_end_0, end_mask = var_709_end_mask_0, squeeze_mask = var_709_squeeze_mask_0, x = q_complex_3_cast_fp16)[name = tensor("op_709_cast_fp16")]; + tensor var_717_begin_0 = const()[name = tensor("op_717_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_717_end_0 = const()[name = tensor("op_717_end_0"), val = tensor([1, 256, 16, 32, 1])]; + tensor var_717_end_mask_0 = const()[name = tensor("op_717_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_717_squeeze_mask_0 = const()[name = tensor("op_717_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_717_cast_fp16 = slice_by_index(begin = var_717_begin_0, end = var_717_end_0, end_mask = var_717_end_mask_0, squeeze_mask = var_717_squeeze_mask_0, x = k_complex_3_cast_fp16)[name = tensor("op_717_cast_fp16")]; + tensor var_725_begin_0 = const()[name = tensor("op_725_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_725_end_0 = const()[name = tensor("op_725_end_0"), val = tensor([1, 256, 16, 32, 2])]; + tensor var_725_end_mask_0 = const()[name = tensor("op_725_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_725_squeeze_mask_0 = const()[name = tensor("op_725_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_725_cast_fp16 = slice_by_index(begin = var_725_begin_0, end = var_725_end_0, end_mask = var_725_end_mask_0, squeeze_mask = var_725_squeeze_mask_0, x = k_complex_3_cast_fp16)[name = tensor("op_725_cast_fp16")]; + tensor freqs_3_to_fp16 = const()[name = tensor("freqs_3_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(32538496)))]; + tensor var_731_cast_fp16 = mul(x = freqs_3_to_fp16, y = ts_11_cast_fp16)[name = tensor("op_731_cast_fp16")]; + tensor rotr_3_cast_fp16 = cos(x = var_731_cast_fp16)[name = tensor("rotr_3_cast_fp16")]; + tensor roti_3_cast_fp16 = sin(x = var_731_cast_fp16)[name = tensor("roti_3_cast_fp16")]; + tensor var_735_cast_fp16 = mul(x = var_701_cast_fp16, y = rotr_3_cast_fp16)[name = tensor("op_735_cast_fp16")]; + tensor var_736_cast_fp16 = mul(x = var_709_cast_fp16, y = roti_3_cast_fp16)[name = tensor("op_736_cast_fp16")]; + tensor qor_5_cast_fp16 = sub(x = var_735_cast_fp16, y = var_736_cast_fp16)[name = tensor("qor_5_cast_fp16")]; + tensor var_739_cast_fp16 = mul(x = var_701_cast_fp16, y = roti_3_cast_fp16)[name = tensor("op_739_cast_fp16")]; + tensor var_740_cast_fp16 = mul(x = var_709_cast_fp16, y = rotr_3_cast_fp16)[name = tensor("op_740_cast_fp16")]; + tensor qoi_5_cast_fp16 = add(x = var_739_cast_fp16, y = var_740_cast_fp16)[name = tensor("qoi_5_cast_fp16")]; + tensor var_743_cast_fp16 = mul(x = var_717_cast_fp16, y = rotr_3_cast_fp16)[name = tensor("op_743_cast_fp16")]; + tensor var_744_cast_fp16 = mul(x = var_725_cast_fp16, y = roti_3_cast_fp16)[name = tensor("op_744_cast_fp16")]; + tensor kor_5_cast_fp16 = sub(x = var_743_cast_fp16, y = var_744_cast_fp16)[name = tensor("kor_5_cast_fp16")]; + tensor var_747_cast_fp16 = mul(x = var_717_cast_fp16, y = roti_3_cast_fp16)[name = tensor("op_747_cast_fp16")]; + tensor var_748_cast_fp16 = mul(x = var_725_cast_fp16, y = rotr_3_cast_fp16)[name = tensor("op_748_cast_fp16")]; + tensor koi_5_cast_fp16 = add(x = var_747_cast_fp16, y = var_748_cast_fp16)[name = tensor("koi_5_cast_fp16")]; + tensor qo_3_axis_0 = const()[name = tensor("qo_3_axis_0"), val = tensor(-1)]; + tensor qo_3_cast_fp16 = stack(axis = qo_3_axis_0, values = (qor_5_cast_fp16, qoi_5_cast_fp16))[name = tensor("qo_3_cast_fp16")]; + tensor ko_3_axis_0 = const()[name = tensor("ko_3_axis_0"), val = tensor(-1)]; + tensor ko_3_cast_fp16 = stack(axis = ko_3_axis_0, values = (kor_5_cast_fp16, koi_5_cast_fp16))[name = tensor("ko_3_cast_fp16")]; + tensor var_781 = const()[name = tensor("op_781"), val = tensor([1, 256, 16, 64])]; + tensor q_9_cast_fp16 = reshape(shape = var_781, x = qo_3_cast_fp16)[name = tensor("q_9_cast_fp16")]; + tensor var_787 = const()[name = tensor("op_787"), val = tensor([1, 256, 16, 64])]; + tensor k_7_cast_fp16 = reshape(shape = var_787, x = ko_3_cast_fp16)[name = tensor("k_7_cast_fp16")]; + tensor write_indices_float_7_cast_fp16 = add(x = var_676_cast_fp16, y = var_393_promoted_to_fp16)[name = tensor("write_indices_float_7_cast_fp16")]; + tensor _inversed_808_y_0_to_fp16 = const()[name = tensor("_inversed_808_y_0_to_fp16"), val = tensor(0x1p-9)]; + tensor _inversed_808_cast_fp16 = mul(x = write_indices_float_7_cast_fp16, y = _inversed_808_y_0_to_fp16)[name = tensor("_inversed_808_cast_fp16")]; + tensor var_809_cast_fp16 = floor(x = _inversed_808_cast_fp16)[name = tensor("op_809_cast_fp16")]; + tensor var_810_to_fp16 = const()[name = tensor("op_810_to_fp16"), val = tensor(0x1p+9)]; + tensor var_811_cast_fp16 = mul(x = var_809_cast_fp16, y = var_810_to_fp16)[name = tensor("op_811_cast_fp16")]; + tensor write_indices_float_9_cast_fp16 = sub(x = write_indices_float_7_cast_fp16, y = var_811_cast_fp16)[name = tensor("write_indices_float_9_cast_fp16")]; + tensor var_829_value_0_to_fp16 = const()[name = tensor("op_829_value_0_to_fp16"), val = tensor(0x1.ffp+8)]; + tensor var_829_cast_fp16 = fill_like(ref_tensor = write_indices_float_9_cast_fp16, value = var_829_value_0_to_fp16)[name = tensor("op_829_cast_fp16")]; + tensor write_indices_float_11_cast_fp16 = select(a = write_indices_float_9_cast_fp16, b = var_829_cast_fp16, cond = token_valid_1_cast_fp16)[name = tensor("write_indices_float_11_cast_fp16")]; + tensor var_835_dtype_0 = const()[name = tensor("op_835_dtype_0"), val = tensor("int32")]; + tensor var_840 = const()[name = tensor("op_840"), val = tensor([1, 256, 1, 1])]; + tensor write_indices_float_11_cast_fp16_to_int32 = cast(dtype = var_835_dtype_0, x = write_indices_float_11_cast_fp16)[name = tensor("cast_655")]; + tensor var_841 = reshape(shape = var_840, x = write_indices_float_11_cast_fp16_to_int32)[name = tensor("op_841")]; + tensor write_indices_3_reps_0 = const()[name = tensor("write_indices_3_reps_0"), val = tensor([1, 1, 16, 64])]; + tensor write_indices_3 = tile(reps = write_indices_3_reps_0, x = var_841)[name = tensor("write_indices_3")]; + tensor var_851_begin_0 = const()[name = tensor("op_851_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_851_end_0 = const()[name = tensor("op_851_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor var_851_end_mask_0 = const()[name = tensor("op_851_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_851_squeeze_mask_0 = const()[name = tensor("op_851_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor cache1_to_fp16_dtype_0 = const()[name = tensor("cache1_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor cache1_to_fp16 = cast(dtype = cache1_to_fp16_dtype_0, x = cache1)[name = tensor("cast_654")]; + tensor var_851_cast_fp16 = slice_by_index(begin = var_851_begin_0, end = var_851_end_0, end_mask = var_851_end_mask_0, squeeze_mask = var_851_squeeze_mask_0, x = cache1_to_fp16)[name = tensor("op_851_cast_fp16")]; + tensor var_853_axis_0 = const()[name = tensor("op_853_axis_0"), val = tensor(1)]; + tensor var_853_mode_0 = const()[name = tensor("op_853_mode_0"), val = tensor("update")]; + tensor var_853_validate_indices_0 = const()[name = tensor("op_853_validate_indices_0"), val = tensor(false)]; + tensor var_853_cast_fp16 = scatter_along_axis(axis = var_853_axis_0, data = var_851_cast_fp16, indices = write_indices_3, mode = var_853_mode_0, updates = k_7_cast_fp16, validate_indices = var_853_validate_indices_0)[name = tensor("op_853_cast_fp16")]; + tensor concat_10 = const()[name = tensor("concat_10"), val = tensor([0, 0, 0, 0, 0])]; + tensor concat_11 = const()[name = tensor("concat_11"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_3_internal_tensor_assign_1_stride_0 = const()[name = tensor("new_cache_3_internal_tensor_assign_1_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_3_internal_tensor_assign_1_begin_mask_0 = const()[name = tensor("new_cache_3_internal_tensor_assign_1_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_3_internal_tensor_assign_1_end_mask_0 = const()[name = tensor("new_cache_3_internal_tensor_assign_1_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_3_internal_tensor_assign_1_squeeze_mask_0 = const()[name = tensor("new_cache_3_internal_tensor_assign_1_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_48 = const()[name = tensor("shape_48"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_2 = const()[name = tensor("reduce_prod_2"), val = tensor(1048576)]; + tensor range_1d_2_start_0 = const()[name = tensor("range_1d_2_start_0"), val = tensor(0)]; + tensor range_1d_2_step_0 = const()[name = tensor("range_1d_2_step_0"), val = tensor(1)]; + tensor range_1d_2 = range_1d(end = reduce_prod_2, start = range_1d_2_start_0, step = range_1d_2_step_0)[name = tensor("range_1d_2")]; + tensor reshape_10 = reshape(shape = shape_48, x = range_1d_2)[name = tensor("reshape_10")]; + tensor slice_by_index_2 = slice_by_index(begin = concat_10, begin_mask = new_cache_3_internal_tensor_assign_1_begin_mask_0, end = concat_11, end_mask = new_cache_3_internal_tensor_assign_1_end_mask_0, squeeze_mask = new_cache_3_internal_tensor_assign_1_squeeze_mask_0, stride = new_cache_3_internal_tensor_assign_1_stride_0, x = reshape_10)[name = tensor("slice_by_index_2")]; + tensor reshape_11_shape_0 = const()[name = tensor("reshape_11_shape_0"), val = tensor([-1])]; + tensor reshape_11 = reshape(shape = reshape_11_shape_0, x = slice_by_index_2)[name = tensor("reshape_11")]; + tensor reshape_12_shape_0 = const()[name = tensor("reshape_12_shape_0"), val = tensor([-1])]; + tensor reshape_12_cast_fp16 = reshape(shape = reshape_12_shape_0, x = var_853_cast_fp16)[name = tensor("reshape_12_cast_fp16")]; + tensor reshape_13_shape_0 = const()[name = tensor("reshape_13_shape_0"), val = tensor([-1])]; + tensor reshape_13_cast_fp16 = reshape(shape = reshape_13_shape_0, x = cache1_to_fp16)[name = tensor("reshape_13_cast_fp16")]; + tensor scatter_2_mode_0 = const()[name = tensor("scatter_2_mode_0"), val = tensor("update")]; + tensor scatter_2_axis_0 = const()[name = tensor("scatter_2_axis_0"), val = tensor(0)]; + tensor scatter_2_validate_indices_0 = const()[name = tensor("scatter_2_validate_indices_0"), val = tensor(false)]; + tensor scatter_2_cast_fp16 = scatter(axis = scatter_2_axis_0, data = reshape_13_cast_fp16, indices = reshape_11, mode = scatter_2_mode_0, updates = reshape_12_cast_fp16, validate_indices = scatter_2_validate_indices_0)[name = tensor("scatter_2_cast_fp16")]; + tensor reshape_14_cast_fp16 = reshape(shape = shape_48, x = scatter_2_cast_fp16)[name = tensor("reshape_14_cast_fp16")]; + tensor var_861_begin_0 = const()[name = tensor("op_861_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor var_861_end_0 = const()[name = tensor("op_861_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor var_861_end_mask_0 = const()[name = tensor("op_861_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_861_squeeze_mask_0 = const()[name = tensor("op_861_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor var_861_cast_fp16 = slice_by_index(begin = var_861_begin_0, end = var_861_end_0, end_mask = var_861_end_mask_0, squeeze_mask = var_861_squeeze_mask_0, x = reshape_14_cast_fp16)[name = tensor("op_861_cast_fp16")]; + tensor var_863_axis_0 = const()[name = tensor("op_863_axis_0"), val = tensor(1)]; + tensor var_863_mode_0 = const()[name = tensor("op_863_mode_0"), val = tensor("update")]; + tensor var_863_validate_indices_0 = const()[name = tensor("op_863_validate_indices_0"), val = tensor(false)]; + tensor var_863_cast_fp16 = scatter_along_axis(axis = var_863_axis_0, data = var_861_cast_fp16, indices = write_indices_3, mode = var_863_mode_0, updates = v_3_cast_fp16, validate_indices = var_863_validate_indices_0)[name = tensor("op_863_cast_fp16")]; + tensor concat_12 = const()[name = tensor("concat_12"), val = tensor([1, 0, 0, 0, 0])]; + tensor concat_13 = const()[name = tensor("concat_13"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_3_internal_tensor_assign_2_stride_0 = const()[name = tensor("new_cache_3_internal_tensor_assign_2_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_3_internal_tensor_assign_2_begin_mask_0 = const()[name = tensor("new_cache_3_internal_tensor_assign_2_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_3_internal_tensor_assign_2_end_mask_0 = const()[name = tensor("new_cache_3_internal_tensor_assign_2_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_3_internal_tensor_assign_2_squeeze_mask_0 = const()[name = tensor("new_cache_3_internal_tensor_assign_2_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_49 = const()[name = tensor("shape_49"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_3 = const()[name = tensor("reduce_prod_3"), val = tensor(1048576)]; + tensor range_1d_3_start_0 = const()[name = tensor("range_1d_3_start_0"), val = tensor(0)]; + tensor range_1d_3_step_0 = const()[name = tensor("range_1d_3_step_0"), val = tensor(1)]; + tensor range_1d_3 = range_1d(end = reduce_prod_3, start = range_1d_3_start_0, step = range_1d_3_step_0)[name = tensor("range_1d_3")]; + tensor reshape_15 = reshape(shape = shape_49, x = range_1d_3)[name = tensor("reshape_15")]; + tensor slice_by_index_3 = slice_by_index(begin = concat_12, begin_mask = new_cache_3_internal_tensor_assign_2_begin_mask_0, end = concat_13, end_mask = new_cache_3_internal_tensor_assign_2_end_mask_0, squeeze_mask = new_cache_3_internal_tensor_assign_2_squeeze_mask_0, stride = new_cache_3_internal_tensor_assign_2_stride_0, x = reshape_15)[name = tensor("slice_by_index_3")]; + tensor reshape_16_shape_0 = const()[name = tensor("reshape_16_shape_0"), val = tensor([-1])]; + tensor reshape_16 = reshape(shape = reshape_16_shape_0, x = slice_by_index_3)[name = tensor("reshape_16")]; + tensor reshape_17_shape_0 = const()[name = tensor("reshape_17_shape_0"), val = tensor([-1])]; + tensor reshape_17_cast_fp16 = reshape(shape = reshape_17_shape_0, x = var_863_cast_fp16)[name = tensor("reshape_17_cast_fp16")]; + tensor reshape_18_shape_0 = const()[name = tensor("reshape_18_shape_0"), val = tensor([-1])]; + tensor reshape_18_cast_fp16 = reshape(shape = reshape_18_shape_0, x = reshape_14_cast_fp16)[name = tensor("reshape_18_cast_fp16")]; + tensor scatter_3_mode_0 = const()[name = tensor("scatter_3_mode_0"), val = tensor("update")]; + tensor scatter_3_axis_0 = const()[name = tensor("scatter_3_axis_0"), val = tensor(0)]; + tensor scatter_3_validate_indices_0 = const()[name = tensor("scatter_3_validate_indices_0"), val = tensor(false)]; + tensor scatter_3_cast_fp16 = scatter(axis = scatter_3_axis_0, data = reshape_18_cast_fp16, indices = reshape_16, mode = scatter_3_mode_0, updates = reshape_17_cast_fp16, validate_indices = scatter_3_validate_indices_0)[name = tensor("scatter_3_cast_fp16")]; + tensor reshape_19_cast_fp16 = reshape(shape = shape_49, x = scatter_3_cast_fp16)[name = tensor("reshape_19_cast_fp16")]; + tensor reshape_19_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("reshape_19_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor keys_7_begin_0 = const()[name = tensor("keys_7_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor keys_7_end_0 = const()[name = tensor("keys_7_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor keys_7_end_mask_0 = const()[name = tensor("keys_7_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor keys_7_squeeze_mask_0 = const()[name = tensor("keys_7_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor keys_7_cast_fp16 = slice_by_index(begin = keys_7_begin_0, end = keys_7_end_0, end_mask = keys_7_end_mask_0, squeeze_mask = keys_7_squeeze_mask_0, x = reshape_19_cast_fp16)[name = tensor("keys_7_cast_fp16")]; + tensor values_7_begin_0 = const()[name = tensor("values_7_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor values_7_end_0 = const()[name = tensor("values_7_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor values_7_end_mask_0 = const()[name = tensor("values_7_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor values_7_squeeze_mask_0 = const()[name = tensor("values_7_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor values_7_cast_fp16 = slice_by_index(begin = values_7_begin_0, end = values_7_end_0, end_mask = values_7_end_mask_0, squeeze_mask = values_7_squeeze_mask_0, x = reshape_19_cast_fp16)[name = tensor("values_7_cast_fp16")]; + tensor var_875_cast_fp16 = not_equal(x = keys_7_cast_fp16, y = keys_7_cast_fp16)[name = tensor("op_875_cast_fp16")]; + tensor keys_9_cast_fp16 = select(a = var_470_to_fp16, b = keys_7_cast_fp16, cond = var_875_cast_fp16)[name = tensor("keys_9_cast_fp16")]; + tensor var_883_cast_fp16 = not_equal(x = values_7_cast_fp16, y = values_7_cast_fp16)[name = tensor("op_883_cast_fp16")]; + tensor values_9_cast_fp16 = select(a = var_470_to_fp16, b = values_7_cast_fp16, cond = var_883_cast_fp16)[name = tensor("values_9_cast_fp16")]; + tensor var_907 = const()[name = tensor("op_907"), val = tensor([0, 2, 1, 3])]; + tensor var_923 = const()[name = tensor("op_923"), val = tensor([1, 1, 1])]; + tensor var_924_cast_fp16 = reshape(shape = var_923, x = position1_to_fp16)[name = tensor("op_924_cast_fp16")]; + tensor q_positions_3_cast_fp16 = add(x = var_924_cast_fp16, y = q_offsets_1_promoted_to_fp16)[name = tensor("q_positions_3_cast_fp16")]; + tensor valid_end_3_cast_fp16 = add(x = var_924_cast_fp16, y = var_536_cast_fp16)[name = tensor("valid_end_3_cast_fp16")]; + tensor valid_mask_3_cast_fp16 = less(x = k_positions_1_promoted_to_fp16, y = valid_end_3_cast_fp16)[name = tensor("valid_mask_3_cast_fp16")]; + tensor causal_mask_3_cast_fp16 = less_equal(x = k_positions_1_promoted_to_fp16, y = q_positions_3_cast_fp16)[name = tensor("causal_mask_3_cast_fp16")]; + tensor var_952 = logical_and(x = valid_mask_3_cast_fp16, y = causal_mask_3_cast_fp16)[name = tensor("op_952")]; + tensor attn_mask_3_axes_0 = const()[name = tensor("attn_mask_3_axes_0"), val = tensor([1])]; + tensor attn_mask_3 = expand_dims(axes = attn_mask_3_axes_0, x = var_952)[name = tensor("attn_mask_3")]; + tensor var_966_transpose_x_0 = const()[name = tensor("op_966_transpose_x_0"), val = tensor(false)]; + tensor var_966_transpose_y_0 = const()[name = tensor("op_966_transpose_y_0"), val = tensor(false)]; + tensor transpose_71_perm_0 = const()[name = tensor("transpose_71_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_72_perm_0 = const()[name = tensor("transpose_72_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_72 = transpose(perm = transpose_72_perm_0, x = keys_9_cast_fp16)[name = tensor("transpose_200")]; + tensor transpose_71 = transpose(perm = transpose_71_perm_0, x = q_9_cast_fp16)[name = tensor("transpose_201")]; + tensor var_966_cast_fp16 = matmul(transpose_x = var_966_transpose_x_0, transpose_y = var_966_transpose_y_0, x = transpose_71, y = transpose_72)[name = tensor("op_966_cast_fp16")]; + tensor var_960_to_fp16 = const()[name = tensor("op_960_to_fp16"), val = tensor([0x1p-3])]; + tensor attn_weights_7_cast_fp16 = mul(x = var_966_cast_fp16, y = var_960_to_fp16)[name = tensor("attn_weights_7_cast_fp16")]; + tensor var_968 = logical_not(x = attn_mask_3)[name = tensor("op_968")]; + tensor var_969_to_fp16 = const()[name = tensor("op_969_to_fp16"), val = tensor(-inf)]; + tensor attn_weights_9_cast_fp16 = select(a = var_969_to_fp16, b = attn_weights_7_cast_fp16, cond = var_968)[name = tensor("attn_weights_9_cast_fp16")]; + tensor var_971 = const()[name = tensor("op_971"), val = tensor(-1)]; + tensor attn_weights_11_cast_fp16 = softmax(axis = var_971, x = attn_weights_9_cast_fp16)[name = tensor("attn_weights_11_cast_fp16")]; + tensor attn_output_3_transpose_x_0 = const()[name = tensor("attn_output_3_transpose_x_0"), val = tensor(false)]; + tensor attn_output_3_transpose_y_0 = const()[name = tensor("attn_output_3_transpose_y_0"), val = tensor(false)]; + tensor values_11_cast_fp16 = transpose(perm = var_907, x = values_9_cast_fp16)[name = tensor("transpose_202")]; + tensor attn_output_3_cast_fp16 = matmul(transpose_x = attn_output_3_transpose_x_0, transpose_y = attn_output_3_transpose_y_0, x = attn_weights_11_cast_fp16, y = values_11_cast_fp16)[name = tensor("attn_output_3_cast_fp16")]; + tensor var_979 = const()[name = tensor("op_979"), val = tensor([0, 2, 1, 3])]; + tensor var_984 = const()[name = tensor("op_984"), val = tensor([1, 256, 1024])]; + tensor var_980_cast_fp16 = transpose(perm = var_979, x = attn_output_3_cast_fp16)[name = tensor("transpose_199")]; + tensor input_17_cast_fp16 = reshape(shape = var_984, x = var_980_cast_fp16)[name = tensor("input_17_cast_fp16")]; + tensor attn1_out_proj_weight_to_fp16 = const()[name = tensor("attn1_out_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(32538624)))]; + tensor linear_5_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = attn1_out_proj_weight_to_fp16, x = input_17_cast_fp16)[name = tensor("linear_5_cast_fp16")]; + tensor var_993_cast_fp16 = add(x = position1_to_fp16, y = valid_len_to_fp16)[name = tensor("op_993_cast_fp16")]; + tensor var_993_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_993_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor input_19_cast_fp16 = add(x = input_13_cast_fp16, y = linear_5_cast_fp16)[name = tensor("input_19_cast_fp16")]; + tensor input_21_axes_0 = const()[name = tensor("input_21_axes_0"), val = tensor([-1])]; + tensor norm1_2_weight_to_fp16 = const()[name = tensor("norm1_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(34635840)))]; + tensor norm1_2_bias_to_fp16 = const()[name = tensor("norm1_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(34637952)))]; + tensor var_997_to_fp16 = const()[name = tensor("op_997_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_21_cast_fp16 = layer_norm(axes = input_21_axes_0, beta = norm1_2_bias_to_fp16, epsilon = var_997_to_fp16, gamma = norm1_2_weight_to_fp16, x = input_19_cast_fp16)[name = tensor("input_21_cast_fp16")]; + tensor linear1_1_weight_to_fp16 = const()[name = tensor("linear1_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(34640064)))]; + tensor linear_6_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = linear1_1_weight_to_fp16, x = input_21_cast_fp16)[name = tensor("linear_6_cast_fp16")]; + tensor input_23_mode_0 = const()[name = tensor("input_23_mode_0"), val = tensor("EXACT")]; + tensor input_23_cast_fp16 = gelu(mode = input_23_mode_0, x = linear_6_cast_fp16)[name = tensor("input_23_cast_fp16")]; + tensor linear1_2_weight_to_fp16 = const()[name = tensor("linear1_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(43028736)))]; + tensor linear_7_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = linear1_2_weight_to_fp16, x = input_23_cast_fp16)[name = tensor("linear_7_cast_fp16")]; + tensor input_25_cast_fp16 = add(x = input_19_cast_fp16, y = linear_7_cast_fp16)[name = tensor("input_25_cast_fp16")]; + tensor input_27_axes_0 = const()[name = tensor("input_27_axes_0"), val = tensor([-1])]; + tensor norm2_1_weight_to_fp16 = const()[name = tensor("norm2_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(51417408)))]; + tensor norm2_1_bias_to_fp16 = const()[name = tensor("norm2_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(51419520)))]; + tensor var_1014_to_fp16 = const()[name = tensor("op_1014_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_27_cast_fp16 = layer_norm(axes = input_27_axes_0, beta = norm2_1_bias_to_fp16, epsilon = var_1014_to_fp16, gamma = norm2_1_weight_to_fp16, x = input_25_cast_fp16)[name = tensor("input_27_cast_fp16")]; + tensor attn2_in_proj_weight_to_fp16 = const()[name = tensor("attn2_in_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(51421632)))]; + tensor linear_8_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = attn2_in_proj_weight_to_fp16, x = input_27_cast_fp16)[name = tensor("linear_8_cast_fp16")]; + tensor var_1028 = const()[name = tensor("op_1028"), val = tensor([1, 256, 3, 16, 64])]; + tensor qkv_5_cast_fp16 = reshape(shape = var_1028, x = linear_8_cast_fp16)[name = tensor("qkv_5_cast_fp16")]; + tensor q_13_begin_0 = const()[name = tensor("q_13_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor q_13_end_0 = const()[name = tensor("q_13_end_0"), val = tensor([1, 256, 1, 16, 64])]; + tensor q_13_end_mask_0 = const()[name = tensor("q_13_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor q_13_squeeze_mask_0 = const()[name = tensor("q_13_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor q_13_cast_fp16 = slice_by_index(begin = q_13_begin_0, end = q_13_end_0, end_mask = q_13_end_mask_0, squeeze_mask = q_13_squeeze_mask_0, x = qkv_5_cast_fp16)[name = tensor("q_13_cast_fp16")]; + tensor k_9_begin_0 = const()[name = tensor("k_9_begin_0"), val = tensor([0, 0, 1, 0, 0])]; + tensor k_9_end_0 = const()[name = tensor("k_9_end_0"), val = tensor([1, 256, 2, 16, 64])]; + tensor k_9_end_mask_0 = const()[name = tensor("k_9_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor k_9_squeeze_mask_0 = const()[name = tensor("k_9_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor k_9_cast_fp16 = slice_by_index(begin = k_9_begin_0, end = k_9_end_0, end_mask = k_9_end_mask_0, squeeze_mask = k_9_squeeze_mask_0, x = qkv_5_cast_fp16)[name = tensor("k_9_cast_fp16")]; + tensor v_5_begin_0 = const()[name = tensor("v_5_begin_0"), val = tensor([0, 0, 2, 0, 0])]; + tensor v_5_end_0 = const()[name = tensor("v_5_end_0"), val = tensor([1, 256, 3, 16, 64])]; + tensor v_5_end_mask_0 = const()[name = tensor("v_5_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor v_5_squeeze_mask_0 = const()[name = tensor("v_5_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor v_5_cast_fp16 = slice_by_index(begin = v_5_begin_0, end = v_5_end_0, end_mask = v_5_end_mask_0, squeeze_mask = v_5_squeeze_mask_0, x = qkv_5_cast_fp16)[name = tensor("v_5_cast_fp16")]; + tensor var_1086 = const()[name = tensor("op_1086"), val = tensor([1, 1])]; + tensor position2_to_fp16_dtype_0 = const()[name = tensor("position2_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor position2_to_fp16 = cast(dtype = position2_to_fp16_dtype_0, x = position2)[name = tensor("cast_651")]; + tensor var_1087_cast_fp16 = reshape(shape = var_1086, x = position2_to_fp16)[name = tensor("op_1087_cast_fp16")]; + tensor ts_15_cast_fp16 = add(x = ts_1_promoted_to_fp16, y = var_1087_cast_fp16)[name = tensor("ts_15_cast_fp16")]; + tensor var_1094 = const()[name = tensor("op_1094"), val = tensor([1, 256, 1, 1])]; + tensor ts_17_cast_fp16 = reshape(shape = var_1094, x = ts_15_cast_fp16)[name = tensor("ts_17_cast_fp16")]; + tensor var_1101 = const()[name = tensor("op_1101"), val = tensor([1, 256, 16, 32, 2])]; + tensor q_complex_5_cast_fp16 = reshape(shape = var_1101, x = q_13_cast_fp16)[name = tensor("q_complex_5_cast_fp16")]; + tensor var_1108 = const()[name = tensor("op_1108"), val = tensor([1, 256, 16, 32, 2])]; + tensor k_complex_5_cast_fp16 = reshape(shape = var_1108, x = k_9_cast_fp16)[name = tensor("k_complex_5_cast_fp16")]; + tensor var_1112_begin_0 = const()[name = tensor("op_1112_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_1112_end_0 = const()[name = tensor("op_1112_end_0"), val = tensor([1, 256, 16, 32, 1])]; + tensor var_1112_end_mask_0 = const()[name = tensor("op_1112_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_1112_squeeze_mask_0 = const()[name = tensor("op_1112_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_1112_cast_fp16 = slice_by_index(begin = var_1112_begin_0, end = var_1112_end_0, end_mask = var_1112_end_mask_0, squeeze_mask = var_1112_squeeze_mask_0, x = q_complex_5_cast_fp16)[name = tensor("op_1112_cast_fp16")]; + tensor var_1120_begin_0 = const()[name = tensor("op_1120_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_1120_end_0 = const()[name = tensor("op_1120_end_0"), val = tensor([1, 256, 16, 32, 2])]; + tensor var_1120_end_mask_0 = const()[name = tensor("op_1120_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_1120_squeeze_mask_0 = const()[name = tensor("op_1120_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_1120_cast_fp16 = slice_by_index(begin = var_1120_begin_0, end = var_1120_end_0, end_mask = var_1120_end_mask_0, squeeze_mask = var_1120_squeeze_mask_0, x = q_complex_5_cast_fp16)[name = tensor("op_1120_cast_fp16")]; + tensor var_1128_begin_0 = const()[name = tensor("op_1128_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_1128_end_0 = const()[name = tensor("op_1128_end_0"), val = tensor([1, 256, 16, 32, 1])]; + tensor var_1128_end_mask_0 = const()[name = tensor("op_1128_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_1128_squeeze_mask_0 = const()[name = tensor("op_1128_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_1128_cast_fp16 = slice_by_index(begin = var_1128_begin_0, end = var_1128_end_0, end_mask = var_1128_end_mask_0, squeeze_mask = var_1128_squeeze_mask_0, x = k_complex_5_cast_fp16)[name = tensor("op_1128_cast_fp16")]; + tensor var_1136_begin_0 = const()[name = tensor("op_1136_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_1136_end_0 = const()[name = tensor("op_1136_end_0"), val = tensor([1, 256, 16, 32, 2])]; + tensor var_1136_end_mask_0 = const()[name = tensor("op_1136_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_1136_squeeze_mask_0 = const()[name = tensor("op_1136_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_1136_cast_fp16 = slice_by_index(begin = var_1136_begin_0, end = var_1136_end_0, end_mask = var_1136_end_mask_0, squeeze_mask = var_1136_squeeze_mask_0, x = k_complex_5_cast_fp16)[name = tensor("op_1136_cast_fp16")]; + tensor freqs_5_to_fp16 = const()[name = tensor("freqs_5_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(57713152)))]; + tensor var_1142_cast_fp16 = mul(x = freqs_5_to_fp16, y = ts_17_cast_fp16)[name = tensor("op_1142_cast_fp16")]; + tensor rotr_5_cast_fp16 = cos(x = var_1142_cast_fp16)[name = tensor("rotr_5_cast_fp16")]; + tensor roti_5_cast_fp16 = sin(x = var_1142_cast_fp16)[name = tensor("roti_5_cast_fp16")]; + tensor var_1146_cast_fp16 = mul(x = var_1112_cast_fp16, y = rotr_5_cast_fp16)[name = tensor("op_1146_cast_fp16")]; + tensor var_1147_cast_fp16 = mul(x = var_1120_cast_fp16, y = roti_5_cast_fp16)[name = tensor("op_1147_cast_fp16")]; + tensor qor_9_cast_fp16 = sub(x = var_1146_cast_fp16, y = var_1147_cast_fp16)[name = tensor("qor_9_cast_fp16")]; + tensor var_1150_cast_fp16 = mul(x = var_1112_cast_fp16, y = roti_5_cast_fp16)[name = tensor("op_1150_cast_fp16")]; + tensor var_1151_cast_fp16 = mul(x = var_1120_cast_fp16, y = rotr_5_cast_fp16)[name = tensor("op_1151_cast_fp16")]; + tensor qoi_9_cast_fp16 = add(x = var_1150_cast_fp16, y = var_1151_cast_fp16)[name = tensor("qoi_9_cast_fp16")]; + tensor var_1154_cast_fp16 = mul(x = var_1128_cast_fp16, y = rotr_5_cast_fp16)[name = tensor("op_1154_cast_fp16")]; + tensor var_1155_cast_fp16 = mul(x = var_1136_cast_fp16, y = roti_5_cast_fp16)[name = tensor("op_1155_cast_fp16")]; + tensor kor_9_cast_fp16 = sub(x = var_1154_cast_fp16, y = var_1155_cast_fp16)[name = tensor("kor_9_cast_fp16")]; + tensor var_1158_cast_fp16 = mul(x = var_1128_cast_fp16, y = roti_5_cast_fp16)[name = tensor("op_1158_cast_fp16")]; + tensor var_1159_cast_fp16 = mul(x = var_1136_cast_fp16, y = rotr_5_cast_fp16)[name = tensor("op_1159_cast_fp16")]; + tensor koi_9_cast_fp16 = add(x = var_1158_cast_fp16, y = var_1159_cast_fp16)[name = tensor("koi_9_cast_fp16")]; + tensor qo_5_axis_0 = const()[name = tensor("qo_5_axis_0"), val = tensor(-1)]; + tensor qo_5_cast_fp16 = stack(axis = qo_5_axis_0, values = (qor_9_cast_fp16, qoi_9_cast_fp16))[name = tensor("qo_5_cast_fp16")]; + tensor ko_5_axis_0 = const()[name = tensor("ko_5_axis_0"), val = tensor(-1)]; + tensor ko_5_cast_fp16 = stack(axis = ko_5_axis_0, values = (kor_9_cast_fp16, koi_9_cast_fp16))[name = tensor("ko_5_cast_fp16")]; + tensor var_1192 = const()[name = tensor("op_1192"), val = tensor([1, 256, 16, 64])]; + tensor q_15_cast_fp16 = reshape(shape = var_1192, x = qo_5_cast_fp16)[name = tensor("q_15_cast_fp16")]; + tensor var_1198 = const()[name = tensor("op_1198"), val = tensor([1, 256, 16, 64])]; + tensor k_11_cast_fp16 = reshape(shape = var_1198, x = ko_5_cast_fp16)[name = tensor("k_11_cast_fp16")]; + tensor write_indices_float_13_cast_fp16 = add(x = var_1087_cast_fp16, y = var_393_promoted_to_fp16)[name = tensor("write_indices_float_13_cast_fp16")]; + tensor _inversed_1219_y_0_to_fp16 = const()[name = tensor("_inversed_1219_y_0_to_fp16"), val = tensor(0x1p-9)]; + tensor _inversed_1219_cast_fp16 = mul(x = write_indices_float_13_cast_fp16, y = _inversed_1219_y_0_to_fp16)[name = tensor("_inversed_1219_cast_fp16")]; + tensor var_1220_cast_fp16 = floor(x = _inversed_1219_cast_fp16)[name = tensor("op_1220_cast_fp16")]; + tensor var_1221_to_fp16 = const()[name = tensor("op_1221_to_fp16"), val = tensor(0x1p+9)]; + tensor var_1222_cast_fp16 = mul(x = var_1220_cast_fp16, y = var_1221_to_fp16)[name = tensor("op_1222_cast_fp16")]; + tensor write_indices_float_15_cast_fp16 = sub(x = write_indices_float_13_cast_fp16, y = var_1222_cast_fp16)[name = tensor("write_indices_float_15_cast_fp16")]; + tensor var_1240_value_0_to_fp16 = const()[name = tensor("op_1240_value_0_to_fp16"), val = tensor(0x1.ffp+8)]; + tensor var_1240_cast_fp16 = fill_like(ref_tensor = write_indices_float_15_cast_fp16, value = var_1240_value_0_to_fp16)[name = tensor("op_1240_cast_fp16")]; + tensor write_indices_float_17_cast_fp16 = select(a = write_indices_float_15_cast_fp16, b = var_1240_cast_fp16, cond = token_valid_1_cast_fp16)[name = tensor("write_indices_float_17_cast_fp16")]; + tensor var_1246_dtype_0 = const()[name = tensor("op_1246_dtype_0"), val = tensor("int32")]; + tensor var_1251 = const()[name = tensor("op_1251"), val = tensor([1, 256, 1, 1])]; + tensor write_indices_float_17_cast_fp16_to_int32 = cast(dtype = var_1246_dtype_0, x = write_indices_float_17_cast_fp16)[name = tensor("cast_650")]; + tensor var_1252 = reshape(shape = var_1251, x = write_indices_float_17_cast_fp16_to_int32)[name = tensor("op_1252")]; + tensor write_indices_5_reps_0 = const()[name = tensor("write_indices_5_reps_0"), val = tensor([1, 1, 16, 64])]; + tensor write_indices_5 = tile(reps = write_indices_5_reps_0, x = var_1252)[name = tensor("write_indices_5")]; + tensor var_1262_begin_0 = const()[name = tensor("op_1262_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_1262_end_0 = const()[name = tensor("op_1262_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor var_1262_end_mask_0 = const()[name = tensor("op_1262_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_1262_squeeze_mask_0 = const()[name = tensor("op_1262_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor cache2_to_fp16_dtype_0 = const()[name = tensor("cache2_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor cache2_to_fp16 = cast(dtype = cache2_to_fp16_dtype_0, x = cache2)[name = tensor("cast_649")]; + tensor var_1262_cast_fp16 = slice_by_index(begin = var_1262_begin_0, end = var_1262_end_0, end_mask = var_1262_end_mask_0, squeeze_mask = var_1262_squeeze_mask_0, x = cache2_to_fp16)[name = tensor("op_1262_cast_fp16")]; + tensor var_1264_axis_0 = const()[name = tensor("op_1264_axis_0"), val = tensor(1)]; + tensor var_1264_mode_0 = const()[name = tensor("op_1264_mode_0"), val = tensor("update")]; + tensor var_1264_validate_indices_0 = const()[name = tensor("op_1264_validate_indices_0"), val = tensor(false)]; + tensor var_1264_cast_fp16 = scatter_along_axis(axis = var_1264_axis_0, data = var_1262_cast_fp16, indices = write_indices_5, mode = var_1264_mode_0, updates = k_11_cast_fp16, validate_indices = var_1264_validate_indices_0)[name = tensor("op_1264_cast_fp16")]; + tensor concat_18 = const()[name = tensor("concat_18"), val = tensor([0, 0, 0, 0, 0])]; + tensor concat_19 = const()[name = tensor("concat_19"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_5_internal_tensor_assign_1_stride_0 = const()[name = tensor("new_cache_5_internal_tensor_assign_1_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_5_internal_tensor_assign_1_begin_mask_0 = const()[name = tensor("new_cache_5_internal_tensor_assign_1_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_5_internal_tensor_assign_1_end_mask_0 = const()[name = tensor("new_cache_5_internal_tensor_assign_1_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_5_internal_tensor_assign_1_squeeze_mask_0 = const()[name = tensor("new_cache_5_internal_tensor_assign_1_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_50 = const()[name = tensor("shape_50"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_4 = const()[name = tensor("reduce_prod_4"), val = tensor(1048576)]; + tensor range_1d_4_start_0 = const()[name = tensor("range_1d_4_start_0"), val = tensor(0)]; + tensor range_1d_4_step_0 = const()[name = tensor("range_1d_4_step_0"), val = tensor(1)]; + tensor range_1d_4 = range_1d(end = reduce_prod_4, start = range_1d_4_start_0, step = range_1d_4_step_0)[name = tensor("range_1d_4")]; + tensor reshape_20 = reshape(shape = shape_50, x = range_1d_4)[name = tensor("reshape_20")]; + tensor slice_by_index_4 = slice_by_index(begin = concat_18, begin_mask = new_cache_5_internal_tensor_assign_1_begin_mask_0, end = concat_19, end_mask = new_cache_5_internal_tensor_assign_1_end_mask_0, squeeze_mask = new_cache_5_internal_tensor_assign_1_squeeze_mask_0, stride = new_cache_5_internal_tensor_assign_1_stride_0, x = reshape_20)[name = tensor("slice_by_index_4")]; + tensor reshape_21_shape_0 = const()[name = tensor("reshape_21_shape_0"), val = tensor([-1])]; + tensor reshape_21 = reshape(shape = reshape_21_shape_0, x = slice_by_index_4)[name = tensor("reshape_21")]; + tensor reshape_22_shape_0 = const()[name = tensor("reshape_22_shape_0"), val = tensor([-1])]; + tensor reshape_22_cast_fp16 = reshape(shape = reshape_22_shape_0, x = var_1264_cast_fp16)[name = tensor("reshape_22_cast_fp16")]; + tensor reshape_23_shape_0 = const()[name = tensor("reshape_23_shape_0"), val = tensor([-1])]; + tensor reshape_23_cast_fp16 = reshape(shape = reshape_23_shape_0, x = cache2_to_fp16)[name = tensor("reshape_23_cast_fp16")]; + tensor scatter_4_mode_0 = const()[name = tensor("scatter_4_mode_0"), val = tensor("update")]; + tensor scatter_4_axis_0 = const()[name = tensor("scatter_4_axis_0"), val = tensor(0)]; + tensor scatter_4_validate_indices_0 = const()[name = tensor("scatter_4_validate_indices_0"), val = tensor(false)]; + tensor scatter_4_cast_fp16 = scatter(axis = scatter_4_axis_0, data = reshape_23_cast_fp16, indices = reshape_21, mode = scatter_4_mode_0, updates = reshape_22_cast_fp16, validate_indices = scatter_4_validate_indices_0)[name = tensor("scatter_4_cast_fp16")]; + tensor reshape_24_cast_fp16 = reshape(shape = shape_50, x = scatter_4_cast_fp16)[name = tensor("reshape_24_cast_fp16")]; + tensor var_1272_begin_0 = const()[name = tensor("op_1272_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor var_1272_end_0 = const()[name = tensor("op_1272_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor var_1272_end_mask_0 = const()[name = tensor("op_1272_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_1272_squeeze_mask_0 = const()[name = tensor("op_1272_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor var_1272_cast_fp16 = slice_by_index(begin = var_1272_begin_0, end = var_1272_end_0, end_mask = var_1272_end_mask_0, squeeze_mask = var_1272_squeeze_mask_0, x = reshape_24_cast_fp16)[name = tensor("op_1272_cast_fp16")]; + tensor var_1274_axis_0 = const()[name = tensor("op_1274_axis_0"), val = tensor(1)]; + tensor var_1274_mode_0 = const()[name = tensor("op_1274_mode_0"), val = tensor("update")]; + tensor var_1274_validate_indices_0 = const()[name = tensor("op_1274_validate_indices_0"), val = tensor(false)]; + tensor var_1274_cast_fp16 = scatter_along_axis(axis = var_1274_axis_0, data = var_1272_cast_fp16, indices = write_indices_5, mode = var_1274_mode_0, updates = v_5_cast_fp16, validate_indices = var_1274_validate_indices_0)[name = tensor("op_1274_cast_fp16")]; + tensor concat_20 = const()[name = tensor("concat_20"), val = tensor([1, 0, 0, 0, 0])]; + tensor concat_21 = const()[name = tensor("concat_21"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_5_internal_tensor_assign_2_stride_0 = const()[name = tensor("new_cache_5_internal_tensor_assign_2_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_5_internal_tensor_assign_2_begin_mask_0 = const()[name = tensor("new_cache_5_internal_tensor_assign_2_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_5_internal_tensor_assign_2_end_mask_0 = const()[name = tensor("new_cache_5_internal_tensor_assign_2_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_5_internal_tensor_assign_2_squeeze_mask_0 = const()[name = tensor("new_cache_5_internal_tensor_assign_2_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_51 = const()[name = tensor("shape_51"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_5 = const()[name = tensor("reduce_prod_5"), val = tensor(1048576)]; + tensor range_1d_5_start_0 = const()[name = tensor("range_1d_5_start_0"), val = tensor(0)]; + tensor range_1d_5_step_0 = const()[name = tensor("range_1d_5_step_0"), val = tensor(1)]; + tensor range_1d_5 = range_1d(end = reduce_prod_5, start = range_1d_5_start_0, step = range_1d_5_step_0)[name = tensor("range_1d_5")]; + tensor reshape_25 = reshape(shape = shape_51, x = range_1d_5)[name = tensor("reshape_25")]; + tensor slice_by_index_5 = slice_by_index(begin = concat_20, begin_mask = new_cache_5_internal_tensor_assign_2_begin_mask_0, end = concat_21, end_mask = new_cache_5_internal_tensor_assign_2_end_mask_0, squeeze_mask = new_cache_5_internal_tensor_assign_2_squeeze_mask_0, stride = new_cache_5_internal_tensor_assign_2_stride_0, x = reshape_25)[name = tensor("slice_by_index_5")]; + tensor reshape_26_shape_0 = const()[name = tensor("reshape_26_shape_0"), val = tensor([-1])]; + tensor reshape_26 = reshape(shape = reshape_26_shape_0, x = slice_by_index_5)[name = tensor("reshape_26")]; + tensor reshape_27_shape_0 = const()[name = tensor("reshape_27_shape_0"), val = tensor([-1])]; + tensor reshape_27_cast_fp16 = reshape(shape = reshape_27_shape_0, x = var_1274_cast_fp16)[name = tensor("reshape_27_cast_fp16")]; + tensor reshape_28_shape_0 = const()[name = tensor("reshape_28_shape_0"), val = tensor([-1])]; + tensor reshape_28_cast_fp16 = reshape(shape = reshape_28_shape_0, x = reshape_24_cast_fp16)[name = tensor("reshape_28_cast_fp16")]; + tensor scatter_5_mode_0 = const()[name = tensor("scatter_5_mode_0"), val = tensor("update")]; + tensor scatter_5_axis_0 = const()[name = tensor("scatter_5_axis_0"), val = tensor(0)]; + tensor scatter_5_validate_indices_0 = const()[name = tensor("scatter_5_validate_indices_0"), val = tensor(false)]; + tensor scatter_5_cast_fp16 = scatter(axis = scatter_5_axis_0, data = reshape_28_cast_fp16, indices = reshape_26, mode = scatter_5_mode_0, updates = reshape_27_cast_fp16, validate_indices = scatter_5_validate_indices_0)[name = tensor("scatter_5_cast_fp16")]; + tensor reshape_29_cast_fp16 = reshape(shape = shape_51, x = scatter_5_cast_fp16)[name = tensor("reshape_29_cast_fp16")]; + tensor reshape_29_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("reshape_29_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor keys_13_begin_0 = const()[name = tensor("keys_13_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor keys_13_end_0 = const()[name = tensor("keys_13_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor keys_13_end_mask_0 = const()[name = tensor("keys_13_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor keys_13_squeeze_mask_0 = const()[name = tensor("keys_13_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor keys_13_cast_fp16 = slice_by_index(begin = keys_13_begin_0, end = keys_13_end_0, end_mask = keys_13_end_mask_0, squeeze_mask = keys_13_squeeze_mask_0, x = reshape_29_cast_fp16)[name = tensor("keys_13_cast_fp16")]; + tensor values_13_begin_0 = const()[name = tensor("values_13_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor values_13_end_0 = const()[name = tensor("values_13_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor values_13_end_mask_0 = const()[name = tensor("values_13_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor values_13_squeeze_mask_0 = const()[name = tensor("values_13_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor values_13_cast_fp16 = slice_by_index(begin = values_13_begin_0, end = values_13_end_0, end_mask = values_13_end_mask_0, squeeze_mask = values_13_squeeze_mask_0, x = reshape_29_cast_fp16)[name = tensor("values_13_cast_fp16")]; + tensor var_1286_cast_fp16 = not_equal(x = keys_13_cast_fp16, y = keys_13_cast_fp16)[name = tensor("op_1286_cast_fp16")]; + tensor keys_15_cast_fp16 = select(a = var_470_to_fp16, b = keys_13_cast_fp16, cond = var_1286_cast_fp16)[name = tensor("keys_15_cast_fp16")]; + tensor var_1294_cast_fp16 = not_equal(x = values_13_cast_fp16, y = values_13_cast_fp16)[name = tensor("op_1294_cast_fp16")]; + tensor values_15_cast_fp16 = select(a = var_470_to_fp16, b = values_13_cast_fp16, cond = var_1294_cast_fp16)[name = tensor("values_15_cast_fp16")]; + tensor var_1318 = const()[name = tensor("op_1318"), val = tensor([0, 2, 1, 3])]; + tensor var_1334 = const()[name = tensor("op_1334"), val = tensor([1, 1, 1])]; + tensor var_1335_cast_fp16 = reshape(shape = var_1334, x = position2_to_fp16)[name = tensor("op_1335_cast_fp16")]; + tensor q_positions_5_cast_fp16 = add(x = var_1335_cast_fp16, y = q_offsets_1_promoted_to_fp16)[name = tensor("q_positions_5_cast_fp16")]; + tensor valid_end_5_cast_fp16 = add(x = var_1335_cast_fp16, y = var_536_cast_fp16)[name = tensor("valid_end_5_cast_fp16")]; + tensor valid_mask_5_cast_fp16 = less(x = k_positions_1_promoted_to_fp16, y = valid_end_5_cast_fp16)[name = tensor("valid_mask_5_cast_fp16")]; + tensor causal_mask_5_cast_fp16 = less_equal(x = k_positions_1_promoted_to_fp16, y = q_positions_5_cast_fp16)[name = tensor("causal_mask_5_cast_fp16")]; + tensor var_1363 = logical_and(x = valid_mask_5_cast_fp16, y = causal_mask_5_cast_fp16)[name = tensor("op_1363")]; + tensor attn_mask_5_axes_0 = const()[name = tensor("attn_mask_5_axes_0"), val = tensor([1])]; + tensor attn_mask_5 = expand_dims(axes = attn_mask_5_axes_0, x = var_1363)[name = tensor("attn_mask_5")]; + tensor var_1377_transpose_x_0 = const()[name = tensor("op_1377_transpose_x_0"), val = tensor(false)]; + tensor var_1377_transpose_y_0 = const()[name = tensor("op_1377_transpose_y_0"), val = tensor(false)]; + tensor transpose_73_perm_0 = const()[name = tensor("transpose_73_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_74_perm_0 = const()[name = tensor("transpose_74_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_74 = transpose(perm = transpose_74_perm_0, x = keys_15_cast_fp16)[name = tensor("transpose_196")]; + tensor transpose_73 = transpose(perm = transpose_73_perm_0, x = q_15_cast_fp16)[name = tensor("transpose_197")]; + tensor var_1377_cast_fp16 = matmul(transpose_x = var_1377_transpose_x_0, transpose_y = var_1377_transpose_y_0, x = transpose_73, y = transpose_74)[name = tensor("op_1377_cast_fp16")]; + tensor var_1371_to_fp16 = const()[name = tensor("op_1371_to_fp16"), val = tensor([0x1p-3])]; + tensor attn_weights_13_cast_fp16 = mul(x = var_1377_cast_fp16, y = var_1371_to_fp16)[name = tensor("attn_weights_13_cast_fp16")]; + tensor var_1379 = logical_not(x = attn_mask_5)[name = tensor("op_1379")]; + tensor var_1380_to_fp16 = const()[name = tensor("op_1380_to_fp16"), val = tensor(-inf)]; + tensor attn_weights_15_cast_fp16 = select(a = var_1380_to_fp16, b = attn_weights_13_cast_fp16, cond = var_1379)[name = tensor("attn_weights_15_cast_fp16")]; + tensor var_1382 = const()[name = tensor("op_1382"), val = tensor(-1)]; + tensor attn_weights_17_cast_fp16 = softmax(axis = var_1382, x = attn_weights_15_cast_fp16)[name = tensor("attn_weights_17_cast_fp16")]; + tensor attn_output_5_transpose_x_0 = const()[name = tensor("attn_output_5_transpose_x_0"), val = tensor(false)]; + tensor attn_output_5_transpose_y_0 = const()[name = tensor("attn_output_5_transpose_y_0"), val = tensor(false)]; + tensor values_17_cast_fp16 = transpose(perm = var_1318, x = values_15_cast_fp16)[name = tensor("transpose_198")]; + tensor attn_output_5_cast_fp16 = matmul(transpose_x = attn_output_5_transpose_x_0, transpose_y = attn_output_5_transpose_y_0, x = attn_weights_17_cast_fp16, y = values_17_cast_fp16)[name = tensor("attn_output_5_cast_fp16")]; + tensor var_1390 = const()[name = tensor("op_1390"), val = tensor([0, 2, 1, 3])]; + tensor var_1395 = const()[name = tensor("op_1395"), val = tensor([1, 256, 1024])]; + tensor var_1391_cast_fp16 = transpose(perm = var_1390, x = attn_output_5_cast_fp16)[name = tensor("transpose_195")]; + tensor input_29_cast_fp16 = reshape(shape = var_1395, x = var_1391_cast_fp16)[name = tensor("input_29_cast_fp16")]; + tensor attn2_out_proj_weight_to_fp16 = const()[name = tensor("attn2_out_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(57713280)))]; + tensor linear_9_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = attn2_out_proj_weight_to_fp16, x = input_29_cast_fp16)[name = tensor("linear_9_cast_fp16")]; + tensor var_1404_cast_fp16 = add(x = position2_to_fp16, y = valid_len_to_fp16)[name = tensor("op_1404_cast_fp16")]; + tensor var_1404_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_1404_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor input_31_cast_fp16 = add(x = input_25_cast_fp16, y = linear_9_cast_fp16)[name = tensor("input_31_cast_fp16")]; + tensor input_33_axes_0 = const()[name = tensor("input_33_axes_0"), val = tensor([-1])]; + tensor norm2_2_weight_to_fp16 = const()[name = tensor("norm2_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(59810496)))]; + tensor norm2_2_bias_to_fp16 = const()[name = tensor("norm2_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(59812608)))]; + tensor var_1408_to_fp16 = const()[name = tensor("op_1408_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_33_cast_fp16 = layer_norm(axes = input_33_axes_0, beta = norm2_2_bias_to_fp16, epsilon = var_1408_to_fp16, gamma = norm2_2_weight_to_fp16, x = input_31_cast_fp16)[name = tensor("input_33_cast_fp16")]; + tensor linear2_1_weight_to_fp16 = const()[name = tensor("linear2_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(59814720)))]; + tensor linear_10_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = linear2_1_weight_to_fp16, x = input_33_cast_fp16)[name = tensor("linear_10_cast_fp16")]; + tensor input_35_mode_0 = const()[name = tensor("input_35_mode_0"), val = tensor("EXACT")]; + tensor input_35_cast_fp16 = gelu(mode = input_35_mode_0, x = linear_10_cast_fp16)[name = tensor("input_35_cast_fp16")]; + tensor linear2_2_weight_to_fp16 = const()[name = tensor("linear2_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(68203392)))]; + tensor linear_11_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = linear2_2_weight_to_fp16, x = input_35_cast_fp16)[name = tensor("linear_11_cast_fp16")]; + tensor input_37_cast_fp16 = add(x = input_31_cast_fp16, y = linear_11_cast_fp16)[name = tensor("input_37_cast_fp16")]; + tensor input_39_axes_0 = const()[name = tensor("input_39_axes_0"), val = tensor([-1])]; + tensor norm3_1_weight_to_fp16 = const()[name = tensor("norm3_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(76592064)))]; + tensor norm3_1_bias_to_fp16 = const()[name = tensor("norm3_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(76594176)))]; + tensor var_1425_to_fp16 = const()[name = tensor("op_1425_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_39_cast_fp16 = layer_norm(axes = input_39_axes_0, beta = norm3_1_bias_to_fp16, epsilon = var_1425_to_fp16, gamma = norm3_1_weight_to_fp16, x = input_37_cast_fp16)[name = tensor("input_39_cast_fp16")]; + tensor attn3_in_proj_weight_to_fp16 = const()[name = tensor("attn3_in_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(76596288)))]; + tensor linear_12_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = attn3_in_proj_weight_to_fp16, x = input_39_cast_fp16)[name = tensor("linear_12_cast_fp16")]; + tensor var_1439 = const()[name = tensor("op_1439"), val = tensor([1, 256, 3, 16, 64])]; + tensor qkv_7_cast_fp16 = reshape(shape = var_1439, x = linear_12_cast_fp16)[name = tensor("qkv_7_cast_fp16")]; + tensor q_19_begin_0 = const()[name = tensor("q_19_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor q_19_end_0 = const()[name = tensor("q_19_end_0"), val = tensor([1, 256, 1, 16, 64])]; + tensor q_19_end_mask_0 = const()[name = tensor("q_19_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor q_19_squeeze_mask_0 = const()[name = tensor("q_19_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor q_19_cast_fp16 = slice_by_index(begin = q_19_begin_0, end = q_19_end_0, end_mask = q_19_end_mask_0, squeeze_mask = q_19_squeeze_mask_0, x = qkv_7_cast_fp16)[name = tensor("q_19_cast_fp16")]; + tensor k_13_begin_0 = const()[name = tensor("k_13_begin_0"), val = tensor([0, 0, 1, 0, 0])]; + tensor k_13_end_0 = const()[name = tensor("k_13_end_0"), val = tensor([1, 256, 2, 16, 64])]; + tensor k_13_end_mask_0 = const()[name = tensor("k_13_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor k_13_squeeze_mask_0 = const()[name = tensor("k_13_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor k_13_cast_fp16 = slice_by_index(begin = k_13_begin_0, end = k_13_end_0, end_mask = k_13_end_mask_0, squeeze_mask = k_13_squeeze_mask_0, x = qkv_7_cast_fp16)[name = tensor("k_13_cast_fp16")]; + tensor v_7_begin_0 = const()[name = tensor("v_7_begin_0"), val = tensor([0, 0, 2, 0, 0])]; + tensor v_7_end_0 = const()[name = tensor("v_7_end_0"), val = tensor([1, 256, 3, 16, 64])]; + tensor v_7_end_mask_0 = const()[name = tensor("v_7_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor v_7_squeeze_mask_0 = const()[name = tensor("v_7_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor v_7_cast_fp16 = slice_by_index(begin = v_7_begin_0, end = v_7_end_0, end_mask = v_7_end_mask_0, squeeze_mask = v_7_squeeze_mask_0, x = qkv_7_cast_fp16)[name = tensor("v_7_cast_fp16")]; + tensor var_1497 = const()[name = tensor("op_1497"), val = tensor([1, 1])]; + tensor position3_to_fp16_dtype_0 = const()[name = tensor("position3_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor position3_to_fp16 = cast(dtype = position3_to_fp16_dtype_0, x = position3)[name = tensor("cast_646")]; + tensor var_1498_cast_fp16 = reshape(shape = var_1497, x = position3_to_fp16)[name = tensor("op_1498_cast_fp16")]; + tensor ts_21_cast_fp16 = add(x = ts_1_promoted_to_fp16, y = var_1498_cast_fp16)[name = tensor("ts_21_cast_fp16")]; + tensor var_1505 = const()[name = tensor("op_1505"), val = tensor([1, 256, 1, 1])]; + tensor ts_23_cast_fp16 = reshape(shape = var_1505, x = ts_21_cast_fp16)[name = tensor("ts_23_cast_fp16")]; + tensor var_1512 = const()[name = tensor("op_1512"), val = tensor([1, 256, 16, 32, 2])]; + tensor q_complex_7_cast_fp16 = reshape(shape = var_1512, x = q_19_cast_fp16)[name = tensor("q_complex_7_cast_fp16")]; + tensor var_1519 = const()[name = tensor("op_1519"), val = tensor([1, 256, 16, 32, 2])]; + tensor k_complex_7_cast_fp16 = reshape(shape = var_1519, x = k_13_cast_fp16)[name = tensor("k_complex_7_cast_fp16")]; + tensor var_1523_begin_0 = const()[name = tensor("op_1523_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_1523_end_0 = const()[name = tensor("op_1523_end_0"), val = tensor([1, 256, 16, 32, 1])]; + tensor var_1523_end_mask_0 = const()[name = tensor("op_1523_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_1523_squeeze_mask_0 = const()[name = tensor("op_1523_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_1523_cast_fp16 = slice_by_index(begin = var_1523_begin_0, end = var_1523_end_0, end_mask = var_1523_end_mask_0, squeeze_mask = var_1523_squeeze_mask_0, x = q_complex_7_cast_fp16)[name = tensor("op_1523_cast_fp16")]; + tensor var_1531_begin_0 = const()[name = tensor("op_1531_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_1531_end_0 = const()[name = tensor("op_1531_end_0"), val = tensor([1, 256, 16, 32, 2])]; + tensor var_1531_end_mask_0 = const()[name = tensor("op_1531_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_1531_squeeze_mask_0 = const()[name = tensor("op_1531_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_1531_cast_fp16 = slice_by_index(begin = var_1531_begin_0, end = var_1531_end_0, end_mask = var_1531_end_mask_0, squeeze_mask = var_1531_squeeze_mask_0, x = q_complex_7_cast_fp16)[name = tensor("op_1531_cast_fp16")]; + tensor var_1539_begin_0 = const()[name = tensor("op_1539_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_1539_end_0 = const()[name = tensor("op_1539_end_0"), val = tensor([1, 256, 16, 32, 1])]; + tensor var_1539_end_mask_0 = const()[name = tensor("op_1539_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_1539_squeeze_mask_0 = const()[name = tensor("op_1539_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_1539_cast_fp16 = slice_by_index(begin = var_1539_begin_0, end = var_1539_end_0, end_mask = var_1539_end_mask_0, squeeze_mask = var_1539_squeeze_mask_0, x = k_complex_7_cast_fp16)[name = tensor("op_1539_cast_fp16")]; + tensor var_1547_begin_0 = const()[name = tensor("op_1547_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_1547_end_0 = const()[name = tensor("op_1547_end_0"), val = tensor([1, 256, 16, 32, 2])]; + tensor var_1547_end_mask_0 = const()[name = tensor("op_1547_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_1547_squeeze_mask_0 = const()[name = tensor("op_1547_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_1547_cast_fp16 = slice_by_index(begin = var_1547_begin_0, end = var_1547_end_0, end_mask = var_1547_end_mask_0, squeeze_mask = var_1547_squeeze_mask_0, x = k_complex_7_cast_fp16)[name = tensor("op_1547_cast_fp16")]; + tensor freqs_7_to_fp16 = const()[name = tensor("freqs_7_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(82887808)))]; + tensor var_1553_cast_fp16 = mul(x = freqs_7_to_fp16, y = ts_23_cast_fp16)[name = tensor("op_1553_cast_fp16")]; + tensor rotr_7_cast_fp16 = cos(x = var_1553_cast_fp16)[name = tensor("rotr_7_cast_fp16")]; + tensor roti_7_cast_fp16 = sin(x = var_1553_cast_fp16)[name = tensor("roti_7_cast_fp16")]; + tensor var_1557_cast_fp16 = mul(x = var_1523_cast_fp16, y = rotr_7_cast_fp16)[name = tensor("op_1557_cast_fp16")]; + tensor var_1558_cast_fp16 = mul(x = var_1531_cast_fp16, y = roti_7_cast_fp16)[name = tensor("op_1558_cast_fp16")]; + tensor qor_13_cast_fp16 = sub(x = var_1557_cast_fp16, y = var_1558_cast_fp16)[name = tensor("qor_13_cast_fp16")]; + tensor var_1561_cast_fp16 = mul(x = var_1523_cast_fp16, y = roti_7_cast_fp16)[name = tensor("op_1561_cast_fp16")]; + tensor var_1562_cast_fp16 = mul(x = var_1531_cast_fp16, y = rotr_7_cast_fp16)[name = tensor("op_1562_cast_fp16")]; + tensor qoi_13_cast_fp16 = add(x = var_1561_cast_fp16, y = var_1562_cast_fp16)[name = tensor("qoi_13_cast_fp16")]; + tensor var_1565_cast_fp16 = mul(x = var_1539_cast_fp16, y = rotr_7_cast_fp16)[name = tensor("op_1565_cast_fp16")]; + tensor var_1566_cast_fp16 = mul(x = var_1547_cast_fp16, y = roti_7_cast_fp16)[name = tensor("op_1566_cast_fp16")]; + tensor kor_13_cast_fp16 = sub(x = var_1565_cast_fp16, y = var_1566_cast_fp16)[name = tensor("kor_13_cast_fp16")]; + tensor var_1569_cast_fp16 = mul(x = var_1539_cast_fp16, y = roti_7_cast_fp16)[name = tensor("op_1569_cast_fp16")]; + tensor var_1570_cast_fp16 = mul(x = var_1547_cast_fp16, y = rotr_7_cast_fp16)[name = tensor("op_1570_cast_fp16")]; + tensor koi_13_cast_fp16 = add(x = var_1569_cast_fp16, y = var_1570_cast_fp16)[name = tensor("koi_13_cast_fp16")]; + tensor qo_7_axis_0 = const()[name = tensor("qo_7_axis_0"), val = tensor(-1)]; + tensor qo_7_cast_fp16 = stack(axis = qo_7_axis_0, values = (qor_13_cast_fp16, qoi_13_cast_fp16))[name = tensor("qo_7_cast_fp16")]; + tensor ko_7_axis_0 = const()[name = tensor("ko_7_axis_0"), val = tensor(-1)]; + tensor ko_7_cast_fp16 = stack(axis = ko_7_axis_0, values = (kor_13_cast_fp16, koi_13_cast_fp16))[name = tensor("ko_7_cast_fp16")]; + tensor var_1603 = const()[name = tensor("op_1603"), val = tensor([1, 256, 16, 64])]; + tensor q_21_cast_fp16 = reshape(shape = var_1603, x = qo_7_cast_fp16)[name = tensor("q_21_cast_fp16")]; + tensor var_1609 = const()[name = tensor("op_1609"), val = tensor([1, 256, 16, 64])]; + tensor k_15_cast_fp16 = reshape(shape = var_1609, x = ko_7_cast_fp16)[name = tensor("k_15_cast_fp16")]; + tensor write_indices_float_19_cast_fp16 = add(x = var_1498_cast_fp16, y = var_393_promoted_to_fp16)[name = tensor("write_indices_float_19_cast_fp16")]; + tensor _inversed_1630_y_0_to_fp16 = const()[name = tensor("_inversed_1630_y_0_to_fp16"), val = tensor(0x1p-9)]; + tensor _inversed_1630_cast_fp16 = mul(x = write_indices_float_19_cast_fp16, y = _inversed_1630_y_0_to_fp16)[name = tensor("_inversed_1630_cast_fp16")]; + tensor var_1631_cast_fp16 = floor(x = _inversed_1630_cast_fp16)[name = tensor("op_1631_cast_fp16")]; + tensor var_1632_to_fp16 = const()[name = tensor("op_1632_to_fp16"), val = tensor(0x1p+9)]; + tensor var_1633_cast_fp16 = mul(x = var_1631_cast_fp16, y = var_1632_to_fp16)[name = tensor("op_1633_cast_fp16")]; + tensor write_indices_float_21_cast_fp16 = sub(x = write_indices_float_19_cast_fp16, y = var_1633_cast_fp16)[name = tensor("write_indices_float_21_cast_fp16")]; + tensor var_1651_value_0_to_fp16 = const()[name = tensor("op_1651_value_0_to_fp16"), val = tensor(0x1.ffp+8)]; + tensor var_1651_cast_fp16 = fill_like(ref_tensor = write_indices_float_21_cast_fp16, value = var_1651_value_0_to_fp16)[name = tensor("op_1651_cast_fp16")]; + tensor write_indices_float_23_cast_fp16 = select(a = write_indices_float_21_cast_fp16, b = var_1651_cast_fp16, cond = token_valid_1_cast_fp16)[name = tensor("write_indices_float_23_cast_fp16")]; + tensor var_1657_dtype_0 = const()[name = tensor("op_1657_dtype_0"), val = tensor("int32")]; + tensor var_1662 = const()[name = tensor("op_1662"), val = tensor([1, 256, 1, 1])]; + tensor write_indices_float_23_cast_fp16_to_int32 = cast(dtype = var_1657_dtype_0, x = write_indices_float_23_cast_fp16)[name = tensor("cast_645")]; + tensor var_1663 = reshape(shape = var_1662, x = write_indices_float_23_cast_fp16_to_int32)[name = tensor("op_1663")]; + tensor write_indices_7_reps_0 = const()[name = tensor("write_indices_7_reps_0"), val = tensor([1, 1, 16, 64])]; + tensor write_indices_7 = tile(reps = write_indices_7_reps_0, x = var_1663)[name = tensor("write_indices_7")]; + tensor var_1673_begin_0 = const()[name = tensor("op_1673_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_1673_end_0 = const()[name = tensor("op_1673_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor var_1673_end_mask_0 = const()[name = tensor("op_1673_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_1673_squeeze_mask_0 = const()[name = tensor("op_1673_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor cache3_to_fp16_dtype_0 = const()[name = tensor("cache3_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor cache3_to_fp16 = cast(dtype = cache3_to_fp16_dtype_0, x = cache3)[name = tensor("cast_644")]; + tensor var_1673_cast_fp16 = slice_by_index(begin = var_1673_begin_0, end = var_1673_end_0, end_mask = var_1673_end_mask_0, squeeze_mask = var_1673_squeeze_mask_0, x = cache3_to_fp16)[name = tensor("op_1673_cast_fp16")]; + tensor var_1675_axis_0 = const()[name = tensor("op_1675_axis_0"), val = tensor(1)]; + tensor var_1675_mode_0 = const()[name = tensor("op_1675_mode_0"), val = tensor("update")]; + tensor var_1675_validate_indices_0 = const()[name = tensor("op_1675_validate_indices_0"), val = tensor(false)]; + tensor var_1675_cast_fp16 = scatter_along_axis(axis = var_1675_axis_0, data = var_1673_cast_fp16, indices = write_indices_7, mode = var_1675_mode_0, updates = k_15_cast_fp16, validate_indices = var_1675_validate_indices_0)[name = tensor("op_1675_cast_fp16")]; + tensor concat_26 = const()[name = tensor("concat_26"), val = tensor([0, 0, 0, 0, 0])]; + tensor concat_27 = const()[name = tensor("concat_27"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_7_internal_tensor_assign_1_stride_0 = const()[name = tensor("new_cache_7_internal_tensor_assign_1_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_7_internal_tensor_assign_1_begin_mask_0 = const()[name = tensor("new_cache_7_internal_tensor_assign_1_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_7_internal_tensor_assign_1_end_mask_0 = const()[name = tensor("new_cache_7_internal_tensor_assign_1_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_7_internal_tensor_assign_1_squeeze_mask_0 = const()[name = tensor("new_cache_7_internal_tensor_assign_1_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_52 = const()[name = tensor("shape_52"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_6 = const()[name = tensor("reduce_prod_6"), val = tensor(1048576)]; + tensor range_1d_6_start_0 = const()[name = tensor("range_1d_6_start_0"), val = tensor(0)]; + tensor range_1d_6_step_0 = const()[name = tensor("range_1d_6_step_0"), val = tensor(1)]; + tensor range_1d_6 = range_1d(end = reduce_prod_6, start = range_1d_6_start_0, step = range_1d_6_step_0)[name = tensor("range_1d_6")]; + tensor reshape_30 = reshape(shape = shape_52, x = range_1d_6)[name = tensor("reshape_30")]; + tensor slice_by_index_6 = slice_by_index(begin = concat_26, begin_mask = new_cache_7_internal_tensor_assign_1_begin_mask_0, end = concat_27, end_mask = new_cache_7_internal_tensor_assign_1_end_mask_0, squeeze_mask = new_cache_7_internal_tensor_assign_1_squeeze_mask_0, stride = new_cache_7_internal_tensor_assign_1_stride_0, x = reshape_30)[name = tensor("slice_by_index_6")]; + tensor reshape_31_shape_0 = const()[name = tensor("reshape_31_shape_0"), val = tensor([-1])]; + tensor reshape_31 = reshape(shape = reshape_31_shape_0, x = slice_by_index_6)[name = tensor("reshape_31")]; + tensor reshape_32_shape_0 = const()[name = tensor("reshape_32_shape_0"), val = tensor([-1])]; + tensor reshape_32_cast_fp16 = reshape(shape = reshape_32_shape_0, x = var_1675_cast_fp16)[name = tensor("reshape_32_cast_fp16")]; + tensor reshape_33_shape_0 = const()[name = tensor("reshape_33_shape_0"), val = tensor([-1])]; + tensor reshape_33_cast_fp16 = reshape(shape = reshape_33_shape_0, x = cache3_to_fp16)[name = tensor("reshape_33_cast_fp16")]; + tensor scatter_6_mode_0 = const()[name = tensor("scatter_6_mode_0"), val = tensor("update")]; + tensor scatter_6_axis_0 = const()[name = tensor("scatter_6_axis_0"), val = tensor(0)]; + tensor scatter_6_validate_indices_0 = const()[name = tensor("scatter_6_validate_indices_0"), val = tensor(false)]; + tensor scatter_6_cast_fp16 = scatter(axis = scatter_6_axis_0, data = reshape_33_cast_fp16, indices = reshape_31, mode = scatter_6_mode_0, updates = reshape_32_cast_fp16, validate_indices = scatter_6_validate_indices_0)[name = tensor("scatter_6_cast_fp16")]; + tensor reshape_34_cast_fp16 = reshape(shape = shape_52, x = scatter_6_cast_fp16)[name = tensor("reshape_34_cast_fp16")]; + tensor var_1683_begin_0 = const()[name = tensor("op_1683_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor var_1683_end_0 = const()[name = tensor("op_1683_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor var_1683_end_mask_0 = const()[name = tensor("op_1683_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_1683_squeeze_mask_0 = const()[name = tensor("op_1683_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor var_1683_cast_fp16 = slice_by_index(begin = var_1683_begin_0, end = var_1683_end_0, end_mask = var_1683_end_mask_0, squeeze_mask = var_1683_squeeze_mask_0, x = reshape_34_cast_fp16)[name = tensor("op_1683_cast_fp16")]; + tensor var_1685_axis_0 = const()[name = tensor("op_1685_axis_0"), val = tensor(1)]; + tensor var_1685_mode_0 = const()[name = tensor("op_1685_mode_0"), val = tensor("update")]; + tensor var_1685_validate_indices_0 = const()[name = tensor("op_1685_validate_indices_0"), val = tensor(false)]; + tensor var_1685_cast_fp16 = scatter_along_axis(axis = var_1685_axis_0, data = var_1683_cast_fp16, indices = write_indices_7, mode = var_1685_mode_0, updates = v_7_cast_fp16, validate_indices = var_1685_validate_indices_0)[name = tensor("op_1685_cast_fp16")]; + tensor concat_28 = const()[name = tensor("concat_28"), val = tensor([1, 0, 0, 0, 0])]; + tensor concat_29 = const()[name = tensor("concat_29"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_7_internal_tensor_assign_2_stride_0 = const()[name = tensor("new_cache_7_internal_tensor_assign_2_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_7_internal_tensor_assign_2_begin_mask_0 = const()[name = tensor("new_cache_7_internal_tensor_assign_2_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_7_internal_tensor_assign_2_end_mask_0 = const()[name = tensor("new_cache_7_internal_tensor_assign_2_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_7_internal_tensor_assign_2_squeeze_mask_0 = const()[name = tensor("new_cache_7_internal_tensor_assign_2_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_53 = const()[name = tensor("shape_53"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_7 = const()[name = tensor("reduce_prod_7"), val = tensor(1048576)]; + tensor range_1d_7_start_0 = const()[name = tensor("range_1d_7_start_0"), val = tensor(0)]; + tensor range_1d_7_step_0 = const()[name = tensor("range_1d_7_step_0"), val = tensor(1)]; + tensor range_1d_7 = range_1d(end = reduce_prod_7, start = range_1d_7_start_0, step = range_1d_7_step_0)[name = tensor("range_1d_7")]; + tensor reshape_35 = reshape(shape = shape_53, x = range_1d_7)[name = tensor("reshape_35")]; + tensor slice_by_index_7 = slice_by_index(begin = concat_28, begin_mask = new_cache_7_internal_tensor_assign_2_begin_mask_0, end = concat_29, end_mask = new_cache_7_internal_tensor_assign_2_end_mask_0, squeeze_mask = new_cache_7_internal_tensor_assign_2_squeeze_mask_0, stride = new_cache_7_internal_tensor_assign_2_stride_0, x = reshape_35)[name = tensor("slice_by_index_7")]; + tensor reshape_36_shape_0 = const()[name = tensor("reshape_36_shape_0"), val = tensor([-1])]; + tensor reshape_36 = reshape(shape = reshape_36_shape_0, x = slice_by_index_7)[name = tensor("reshape_36")]; + tensor reshape_37_shape_0 = const()[name = tensor("reshape_37_shape_0"), val = tensor([-1])]; + tensor reshape_37_cast_fp16 = reshape(shape = reshape_37_shape_0, x = var_1685_cast_fp16)[name = tensor("reshape_37_cast_fp16")]; + tensor reshape_38_shape_0 = const()[name = tensor("reshape_38_shape_0"), val = tensor([-1])]; + tensor reshape_38_cast_fp16 = reshape(shape = reshape_38_shape_0, x = reshape_34_cast_fp16)[name = tensor("reshape_38_cast_fp16")]; + tensor scatter_7_mode_0 = const()[name = tensor("scatter_7_mode_0"), val = tensor("update")]; + tensor scatter_7_axis_0 = const()[name = tensor("scatter_7_axis_0"), val = tensor(0)]; + tensor scatter_7_validate_indices_0 = const()[name = tensor("scatter_7_validate_indices_0"), val = tensor(false)]; + tensor scatter_7_cast_fp16 = scatter(axis = scatter_7_axis_0, data = reshape_38_cast_fp16, indices = reshape_36, mode = scatter_7_mode_0, updates = reshape_37_cast_fp16, validate_indices = scatter_7_validate_indices_0)[name = tensor("scatter_7_cast_fp16")]; + tensor reshape_39_cast_fp16 = reshape(shape = shape_53, x = scatter_7_cast_fp16)[name = tensor("reshape_39_cast_fp16")]; + tensor reshape_39_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("reshape_39_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor keys_19_begin_0 = const()[name = tensor("keys_19_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor keys_19_end_0 = const()[name = tensor("keys_19_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor keys_19_end_mask_0 = const()[name = tensor("keys_19_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor keys_19_squeeze_mask_0 = const()[name = tensor("keys_19_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor keys_19_cast_fp16 = slice_by_index(begin = keys_19_begin_0, end = keys_19_end_0, end_mask = keys_19_end_mask_0, squeeze_mask = keys_19_squeeze_mask_0, x = reshape_39_cast_fp16)[name = tensor("keys_19_cast_fp16")]; + tensor values_19_begin_0 = const()[name = tensor("values_19_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor values_19_end_0 = const()[name = tensor("values_19_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor values_19_end_mask_0 = const()[name = tensor("values_19_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor values_19_squeeze_mask_0 = const()[name = tensor("values_19_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor values_19_cast_fp16 = slice_by_index(begin = values_19_begin_0, end = values_19_end_0, end_mask = values_19_end_mask_0, squeeze_mask = values_19_squeeze_mask_0, x = reshape_39_cast_fp16)[name = tensor("values_19_cast_fp16")]; + tensor var_1697_cast_fp16 = not_equal(x = keys_19_cast_fp16, y = keys_19_cast_fp16)[name = tensor("op_1697_cast_fp16")]; + tensor keys_21_cast_fp16 = select(a = var_470_to_fp16, b = keys_19_cast_fp16, cond = var_1697_cast_fp16)[name = tensor("keys_21_cast_fp16")]; + tensor var_1705_cast_fp16 = not_equal(x = values_19_cast_fp16, y = values_19_cast_fp16)[name = tensor("op_1705_cast_fp16")]; + tensor values_21_cast_fp16 = select(a = var_470_to_fp16, b = values_19_cast_fp16, cond = var_1705_cast_fp16)[name = tensor("values_21_cast_fp16")]; + tensor var_1729 = const()[name = tensor("op_1729"), val = tensor([0, 2, 1, 3])]; + tensor var_1745 = const()[name = tensor("op_1745"), val = tensor([1, 1, 1])]; + tensor var_1746_cast_fp16 = reshape(shape = var_1745, x = position3_to_fp16)[name = tensor("op_1746_cast_fp16")]; + tensor q_positions_7_cast_fp16 = add(x = var_1746_cast_fp16, y = q_offsets_1_promoted_to_fp16)[name = tensor("q_positions_7_cast_fp16")]; + tensor valid_end_7_cast_fp16 = add(x = var_1746_cast_fp16, y = var_536_cast_fp16)[name = tensor("valid_end_7_cast_fp16")]; + tensor valid_mask_7_cast_fp16 = less(x = k_positions_1_promoted_to_fp16, y = valid_end_7_cast_fp16)[name = tensor("valid_mask_7_cast_fp16")]; + tensor causal_mask_7_cast_fp16 = less_equal(x = k_positions_1_promoted_to_fp16, y = q_positions_7_cast_fp16)[name = tensor("causal_mask_7_cast_fp16")]; + tensor var_1774 = logical_and(x = valid_mask_7_cast_fp16, y = causal_mask_7_cast_fp16)[name = tensor("op_1774")]; + tensor attn_mask_7_axes_0 = const()[name = tensor("attn_mask_7_axes_0"), val = tensor([1])]; + tensor attn_mask_7 = expand_dims(axes = attn_mask_7_axes_0, x = var_1774)[name = tensor("attn_mask_7")]; + tensor var_1788_transpose_x_0 = const()[name = tensor("op_1788_transpose_x_0"), val = tensor(false)]; + tensor var_1788_transpose_y_0 = const()[name = tensor("op_1788_transpose_y_0"), val = tensor(false)]; + tensor transpose_75_perm_0 = const()[name = tensor("transpose_75_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_76_perm_0 = const()[name = tensor("transpose_76_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_76 = transpose(perm = transpose_76_perm_0, x = keys_21_cast_fp16)[name = tensor("transpose_192")]; + tensor transpose_75 = transpose(perm = transpose_75_perm_0, x = q_21_cast_fp16)[name = tensor("transpose_193")]; + tensor var_1788_cast_fp16 = matmul(transpose_x = var_1788_transpose_x_0, transpose_y = var_1788_transpose_y_0, x = transpose_75, y = transpose_76)[name = tensor("op_1788_cast_fp16")]; + tensor var_1782_to_fp16 = const()[name = tensor("op_1782_to_fp16"), val = tensor([0x1p-3])]; + tensor attn_weights_19_cast_fp16 = mul(x = var_1788_cast_fp16, y = var_1782_to_fp16)[name = tensor("attn_weights_19_cast_fp16")]; + tensor var_1790 = logical_not(x = attn_mask_7)[name = tensor("op_1790")]; + tensor var_1791_to_fp16 = const()[name = tensor("op_1791_to_fp16"), val = tensor(-inf)]; + tensor attn_weights_21_cast_fp16 = select(a = var_1791_to_fp16, b = attn_weights_19_cast_fp16, cond = var_1790)[name = tensor("attn_weights_21_cast_fp16")]; + tensor var_1793 = const()[name = tensor("op_1793"), val = tensor(-1)]; + tensor attn_weights_23_cast_fp16 = softmax(axis = var_1793, x = attn_weights_21_cast_fp16)[name = tensor("attn_weights_23_cast_fp16")]; + tensor attn_output_7_transpose_x_0 = const()[name = tensor("attn_output_7_transpose_x_0"), val = tensor(false)]; + tensor attn_output_7_transpose_y_0 = const()[name = tensor("attn_output_7_transpose_y_0"), val = tensor(false)]; + tensor values_23_cast_fp16 = transpose(perm = var_1729, x = values_21_cast_fp16)[name = tensor("transpose_194")]; + tensor attn_output_7_cast_fp16 = matmul(transpose_x = attn_output_7_transpose_x_0, transpose_y = attn_output_7_transpose_y_0, x = attn_weights_23_cast_fp16, y = values_23_cast_fp16)[name = tensor("attn_output_7_cast_fp16")]; + tensor var_1801 = const()[name = tensor("op_1801"), val = tensor([0, 2, 1, 3])]; + tensor var_1806 = const()[name = tensor("op_1806"), val = tensor([1, 256, 1024])]; + tensor var_1802_cast_fp16 = transpose(perm = var_1801, x = attn_output_7_cast_fp16)[name = tensor("transpose_191")]; + tensor input_41_cast_fp16 = reshape(shape = var_1806, x = var_1802_cast_fp16)[name = tensor("input_41_cast_fp16")]; + tensor attn3_out_proj_weight_to_fp16 = const()[name = tensor("attn3_out_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(82887936)))]; + tensor linear_13_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = attn3_out_proj_weight_to_fp16, x = input_41_cast_fp16)[name = tensor("linear_13_cast_fp16")]; + tensor var_1815_cast_fp16 = add(x = position3_to_fp16, y = valid_len_to_fp16)[name = tensor("op_1815_cast_fp16")]; + tensor var_1815_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_1815_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor input_43_cast_fp16 = add(x = input_37_cast_fp16, y = linear_13_cast_fp16)[name = tensor("input_43_cast_fp16")]; + tensor input_45_axes_0 = const()[name = tensor("input_45_axes_0"), val = tensor([-1])]; + tensor norm3_2_weight_to_fp16 = const()[name = tensor("norm3_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(84985152)))]; + tensor norm3_2_bias_to_fp16 = const()[name = tensor("norm3_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(84987264)))]; + tensor var_1819_to_fp16 = const()[name = tensor("op_1819_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_45_cast_fp16 = layer_norm(axes = input_45_axes_0, beta = norm3_2_bias_to_fp16, epsilon = var_1819_to_fp16, gamma = norm3_2_weight_to_fp16, x = input_43_cast_fp16)[name = tensor("input_45_cast_fp16")]; + tensor linear3_1_weight_to_fp16 = const()[name = tensor("linear3_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(84989376)))]; + tensor linear_14_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = linear3_1_weight_to_fp16, x = input_45_cast_fp16)[name = tensor("linear_14_cast_fp16")]; + tensor input_47_mode_0 = const()[name = tensor("input_47_mode_0"), val = tensor("EXACT")]; + tensor input_47_cast_fp16 = gelu(mode = input_47_mode_0, x = linear_14_cast_fp16)[name = tensor("input_47_cast_fp16")]; + tensor linear3_2_weight_to_fp16 = const()[name = tensor("linear3_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(93378048)))]; + tensor linear_15_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = linear3_2_weight_to_fp16, x = input_47_cast_fp16)[name = tensor("linear_15_cast_fp16")]; + tensor input_49_cast_fp16 = add(x = input_43_cast_fp16, y = linear_15_cast_fp16)[name = tensor("input_49_cast_fp16")]; + tensor input_51_axes_0 = const()[name = tensor("input_51_axes_0"), val = tensor([-1])]; + tensor norm4_1_weight_to_fp16 = const()[name = tensor("norm4_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(101766720)))]; + tensor norm4_1_bias_to_fp16 = const()[name = tensor("norm4_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(101768832)))]; + tensor var_1836_to_fp16 = const()[name = tensor("op_1836_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_51_cast_fp16 = layer_norm(axes = input_51_axes_0, beta = norm4_1_bias_to_fp16, epsilon = var_1836_to_fp16, gamma = norm4_1_weight_to_fp16, x = input_49_cast_fp16)[name = tensor("input_51_cast_fp16")]; + tensor attn4_in_proj_weight_to_fp16 = const()[name = tensor("attn4_in_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(101770944)))]; + tensor linear_16_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = attn4_in_proj_weight_to_fp16, x = input_51_cast_fp16)[name = tensor("linear_16_cast_fp16")]; + tensor var_1850 = const()[name = tensor("op_1850"), val = tensor([1, 256, 3, 16, 64])]; + tensor qkv_9_cast_fp16 = reshape(shape = var_1850, x = linear_16_cast_fp16)[name = tensor("qkv_9_cast_fp16")]; + tensor q_25_begin_0 = const()[name = tensor("q_25_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor q_25_end_0 = const()[name = tensor("q_25_end_0"), val = tensor([1, 256, 1, 16, 64])]; + tensor q_25_end_mask_0 = const()[name = tensor("q_25_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor q_25_squeeze_mask_0 = const()[name = tensor("q_25_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor q_25_cast_fp16 = slice_by_index(begin = q_25_begin_0, end = q_25_end_0, end_mask = q_25_end_mask_0, squeeze_mask = q_25_squeeze_mask_0, x = qkv_9_cast_fp16)[name = tensor("q_25_cast_fp16")]; + tensor k_17_begin_0 = const()[name = tensor("k_17_begin_0"), val = tensor([0, 0, 1, 0, 0])]; + tensor k_17_end_0 = const()[name = tensor("k_17_end_0"), val = tensor([1, 256, 2, 16, 64])]; + tensor k_17_end_mask_0 = const()[name = tensor("k_17_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor k_17_squeeze_mask_0 = const()[name = tensor("k_17_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor k_17_cast_fp16 = slice_by_index(begin = k_17_begin_0, end = k_17_end_0, end_mask = k_17_end_mask_0, squeeze_mask = k_17_squeeze_mask_0, x = qkv_9_cast_fp16)[name = tensor("k_17_cast_fp16")]; + tensor v_9_begin_0 = const()[name = tensor("v_9_begin_0"), val = tensor([0, 0, 2, 0, 0])]; + tensor v_9_end_0 = const()[name = tensor("v_9_end_0"), val = tensor([1, 256, 3, 16, 64])]; + tensor v_9_end_mask_0 = const()[name = tensor("v_9_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor v_9_squeeze_mask_0 = const()[name = tensor("v_9_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor v_9_cast_fp16 = slice_by_index(begin = v_9_begin_0, end = v_9_end_0, end_mask = v_9_end_mask_0, squeeze_mask = v_9_squeeze_mask_0, x = qkv_9_cast_fp16)[name = tensor("v_9_cast_fp16")]; + tensor var_1908 = const()[name = tensor("op_1908"), val = tensor([1, 1])]; + tensor position4_to_fp16_dtype_0 = const()[name = tensor("position4_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor position4_to_fp16 = cast(dtype = position4_to_fp16_dtype_0, x = position4)[name = tensor("cast_641")]; + tensor var_1909_cast_fp16 = reshape(shape = var_1908, x = position4_to_fp16)[name = tensor("op_1909_cast_fp16")]; + tensor ts_27_cast_fp16 = add(x = ts_1_promoted_to_fp16, y = var_1909_cast_fp16)[name = tensor("ts_27_cast_fp16")]; + tensor var_1916 = const()[name = tensor("op_1916"), val = tensor([1, 256, 1, 1])]; + tensor ts_29_cast_fp16 = reshape(shape = var_1916, x = ts_27_cast_fp16)[name = tensor("ts_29_cast_fp16")]; + tensor var_1923 = const()[name = tensor("op_1923"), val = tensor([1, 256, 16, 32, 2])]; + tensor q_complex_9_cast_fp16 = reshape(shape = var_1923, x = q_25_cast_fp16)[name = tensor("q_complex_9_cast_fp16")]; + tensor var_1930 = const()[name = tensor("op_1930"), val = tensor([1, 256, 16, 32, 2])]; + tensor k_complex_9_cast_fp16 = reshape(shape = var_1930, x = k_17_cast_fp16)[name = tensor("k_complex_9_cast_fp16")]; + tensor var_1934_begin_0 = const()[name = tensor("op_1934_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_1934_end_0 = const()[name = tensor("op_1934_end_0"), val = tensor([1, 256, 16, 32, 1])]; + tensor var_1934_end_mask_0 = const()[name = tensor("op_1934_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_1934_squeeze_mask_0 = const()[name = tensor("op_1934_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_1934_cast_fp16 = slice_by_index(begin = var_1934_begin_0, end = var_1934_end_0, end_mask = var_1934_end_mask_0, squeeze_mask = var_1934_squeeze_mask_0, x = q_complex_9_cast_fp16)[name = tensor("op_1934_cast_fp16")]; + tensor var_1942_begin_0 = const()[name = tensor("op_1942_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_1942_end_0 = const()[name = tensor("op_1942_end_0"), val = tensor([1, 256, 16, 32, 2])]; + tensor var_1942_end_mask_0 = const()[name = tensor("op_1942_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_1942_squeeze_mask_0 = const()[name = tensor("op_1942_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_1942_cast_fp16 = slice_by_index(begin = var_1942_begin_0, end = var_1942_end_0, end_mask = var_1942_end_mask_0, squeeze_mask = var_1942_squeeze_mask_0, x = q_complex_9_cast_fp16)[name = tensor("op_1942_cast_fp16")]; + tensor var_1950_begin_0 = const()[name = tensor("op_1950_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_1950_end_0 = const()[name = tensor("op_1950_end_0"), val = tensor([1, 256, 16, 32, 1])]; + tensor var_1950_end_mask_0 = const()[name = tensor("op_1950_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_1950_squeeze_mask_0 = const()[name = tensor("op_1950_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_1950_cast_fp16 = slice_by_index(begin = var_1950_begin_0, end = var_1950_end_0, end_mask = var_1950_end_mask_0, squeeze_mask = var_1950_squeeze_mask_0, x = k_complex_9_cast_fp16)[name = tensor("op_1950_cast_fp16")]; + tensor var_1958_begin_0 = const()[name = tensor("op_1958_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_1958_end_0 = const()[name = tensor("op_1958_end_0"), val = tensor([1, 256, 16, 32, 2])]; + tensor var_1958_end_mask_0 = const()[name = tensor("op_1958_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_1958_squeeze_mask_0 = const()[name = tensor("op_1958_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_1958_cast_fp16 = slice_by_index(begin = var_1958_begin_0, end = var_1958_end_0, end_mask = var_1958_end_mask_0, squeeze_mask = var_1958_squeeze_mask_0, x = k_complex_9_cast_fp16)[name = tensor("op_1958_cast_fp16")]; + tensor freqs_9_to_fp16 = const()[name = tensor("freqs_9_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(108062464)))]; + tensor var_1964_cast_fp16 = mul(x = freqs_9_to_fp16, y = ts_29_cast_fp16)[name = tensor("op_1964_cast_fp16")]; + tensor rotr_9_cast_fp16 = cos(x = var_1964_cast_fp16)[name = tensor("rotr_9_cast_fp16")]; + tensor roti_9_cast_fp16 = sin(x = var_1964_cast_fp16)[name = tensor("roti_9_cast_fp16")]; + tensor var_1968_cast_fp16 = mul(x = var_1934_cast_fp16, y = rotr_9_cast_fp16)[name = tensor("op_1968_cast_fp16")]; + tensor var_1969_cast_fp16 = mul(x = var_1942_cast_fp16, y = roti_9_cast_fp16)[name = tensor("op_1969_cast_fp16")]; + tensor qor_17_cast_fp16 = sub(x = var_1968_cast_fp16, y = var_1969_cast_fp16)[name = tensor("qor_17_cast_fp16")]; + tensor var_1972_cast_fp16 = mul(x = var_1934_cast_fp16, y = roti_9_cast_fp16)[name = tensor("op_1972_cast_fp16")]; + tensor var_1973_cast_fp16 = mul(x = var_1942_cast_fp16, y = rotr_9_cast_fp16)[name = tensor("op_1973_cast_fp16")]; + tensor qoi_17_cast_fp16 = add(x = var_1972_cast_fp16, y = var_1973_cast_fp16)[name = tensor("qoi_17_cast_fp16")]; + tensor var_1976_cast_fp16 = mul(x = var_1950_cast_fp16, y = rotr_9_cast_fp16)[name = tensor("op_1976_cast_fp16")]; + tensor var_1977_cast_fp16 = mul(x = var_1958_cast_fp16, y = roti_9_cast_fp16)[name = tensor("op_1977_cast_fp16")]; + tensor kor_17_cast_fp16 = sub(x = var_1976_cast_fp16, y = var_1977_cast_fp16)[name = tensor("kor_17_cast_fp16")]; + tensor var_1980_cast_fp16 = mul(x = var_1950_cast_fp16, y = roti_9_cast_fp16)[name = tensor("op_1980_cast_fp16")]; + tensor var_1981_cast_fp16 = mul(x = var_1958_cast_fp16, y = rotr_9_cast_fp16)[name = tensor("op_1981_cast_fp16")]; + tensor koi_17_cast_fp16 = add(x = var_1980_cast_fp16, y = var_1981_cast_fp16)[name = tensor("koi_17_cast_fp16")]; + tensor qo_9_axis_0 = const()[name = tensor("qo_9_axis_0"), val = tensor(-1)]; + tensor qo_9_cast_fp16 = stack(axis = qo_9_axis_0, values = (qor_17_cast_fp16, qoi_17_cast_fp16))[name = tensor("qo_9_cast_fp16")]; + tensor ko_9_axis_0 = const()[name = tensor("ko_9_axis_0"), val = tensor(-1)]; + tensor ko_9_cast_fp16 = stack(axis = ko_9_axis_0, values = (kor_17_cast_fp16, koi_17_cast_fp16))[name = tensor("ko_9_cast_fp16")]; + tensor var_2014 = const()[name = tensor("op_2014"), val = tensor([1, 256, 16, 64])]; + tensor q_27_cast_fp16 = reshape(shape = var_2014, x = qo_9_cast_fp16)[name = tensor("q_27_cast_fp16")]; + tensor var_2020 = const()[name = tensor("op_2020"), val = tensor([1, 256, 16, 64])]; + tensor k_19_cast_fp16 = reshape(shape = var_2020, x = ko_9_cast_fp16)[name = tensor("k_19_cast_fp16")]; + tensor write_indices_float_25_cast_fp16 = add(x = var_1909_cast_fp16, y = var_393_promoted_to_fp16)[name = tensor("write_indices_float_25_cast_fp16")]; + tensor _inversed_2041_y_0_to_fp16 = const()[name = tensor("_inversed_2041_y_0_to_fp16"), val = tensor(0x1p-9)]; + tensor _inversed_2041_cast_fp16 = mul(x = write_indices_float_25_cast_fp16, y = _inversed_2041_y_0_to_fp16)[name = tensor("_inversed_2041_cast_fp16")]; + tensor var_2042_cast_fp16 = floor(x = _inversed_2041_cast_fp16)[name = tensor("op_2042_cast_fp16")]; + tensor var_2043_to_fp16 = const()[name = tensor("op_2043_to_fp16"), val = tensor(0x1p+9)]; + tensor var_2044_cast_fp16 = mul(x = var_2042_cast_fp16, y = var_2043_to_fp16)[name = tensor("op_2044_cast_fp16")]; + tensor write_indices_float_27_cast_fp16 = sub(x = write_indices_float_25_cast_fp16, y = var_2044_cast_fp16)[name = tensor("write_indices_float_27_cast_fp16")]; + tensor var_2062_value_0_to_fp16 = const()[name = tensor("op_2062_value_0_to_fp16"), val = tensor(0x1.ffp+8)]; + tensor var_2062_cast_fp16 = fill_like(ref_tensor = write_indices_float_27_cast_fp16, value = var_2062_value_0_to_fp16)[name = tensor("op_2062_cast_fp16")]; + tensor write_indices_float_29_cast_fp16 = select(a = write_indices_float_27_cast_fp16, b = var_2062_cast_fp16, cond = token_valid_1_cast_fp16)[name = tensor("write_indices_float_29_cast_fp16")]; + tensor var_2068_dtype_0 = const()[name = tensor("op_2068_dtype_0"), val = tensor("int32")]; + tensor var_2073 = const()[name = tensor("op_2073"), val = tensor([1, 256, 1, 1])]; + tensor write_indices_float_29_cast_fp16_to_int32 = cast(dtype = var_2068_dtype_0, x = write_indices_float_29_cast_fp16)[name = tensor("cast_640")]; + tensor var_2074 = reshape(shape = var_2073, x = write_indices_float_29_cast_fp16_to_int32)[name = tensor("op_2074")]; + tensor write_indices_9_reps_0 = const()[name = tensor("write_indices_9_reps_0"), val = tensor([1, 1, 16, 64])]; + tensor write_indices_9 = tile(reps = write_indices_9_reps_0, x = var_2074)[name = tensor("write_indices_9")]; + tensor var_2084_begin_0 = const()[name = tensor("op_2084_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_2084_end_0 = const()[name = tensor("op_2084_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor var_2084_end_mask_0 = const()[name = tensor("op_2084_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_2084_squeeze_mask_0 = const()[name = tensor("op_2084_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor cache4_to_fp16_dtype_0 = const()[name = tensor("cache4_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor cache4_to_fp16 = cast(dtype = cache4_to_fp16_dtype_0, x = cache4)[name = tensor("cast_639")]; + tensor var_2084_cast_fp16 = slice_by_index(begin = var_2084_begin_0, end = var_2084_end_0, end_mask = var_2084_end_mask_0, squeeze_mask = var_2084_squeeze_mask_0, x = cache4_to_fp16)[name = tensor("op_2084_cast_fp16")]; + tensor var_2086_axis_0 = const()[name = tensor("op_2086_axis_0"), val = tensor(1)]; + tensor var_2086_mode_0 = const()[name = tensor("op_2086_mode_0"), val = tensor("update")]; + tensor var_2086_validate_indices_0 = const()[name = tensor("op_2086_validate_indices_0"), val = tensor(false)]; + tensor var_2086_cast_fp16 = scatter_along_axis(axis = var_2086_axis_0, data = var_2084_cast_fp16, indices = write_indices_9, mode = var_2086_mode_0, updates = k_19_cast_fp16, validate_indices = var_2086_validate_indices_0)[name = tensor("op_2086_cast_fp16")]; + tensor concat_34 = const()[name = tensor("concat_34"), val = tensor([0, 0, 0, 0, 0])]; + tensor concat_35 = const()[name = tensor("concat_35"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_9_internal_tensor_assign_1_stride_0 = const()[name = tensor("new_cache_9_internal_tensor_assign_1_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_9_internal_tensor_assign_1_begin_mask_0 = const()[name = tensor("new_cache_9_internal_tensor_assign_1_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_9_internal_tensor_assign_1_end_mask_0 = const()[name = tensor("new_cache_9_internal_tensor_assign_1_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_9_internal_tensor_assign_1_squeeze_mask_0 = const()[name = tensor("new_cache_9_internal_tensor_assign_1_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_54 = const()[name = tensor("shape_54"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_8 = const()[name = tensor("reduce_prod_8"), val = tensor(1048576)]; + tensor range_1d_8_start_0 = const()[name = tensor("range_1d_8_start_0"), val = tensor(0)]; + tensor range_1d_8_step_0 = const()[name = tensor("range_1d_8_step_0"), val = tensor(1)]; + tensor range_1d_8 = range_1d(end = reduce_prod_8, start = range_1d_8_start_0, step = range_1d_8_step_0)[name = tensor("range_1d_8")]; + tensor reshape_40 = reshape(shape = shape_54, x = range_1d_8)[name = tensor("reshape_40")]; + tensor slice_by_index_8 = slice_by_index(begin = concat_34, begin_mask = new_cache_9_internal_tensor_assign_1_begin_mask_0, end = concat_35, end_mask = new_cache_9_internal_tensor_assign_1_end_mask_0, squeeze_mask = new_cache_9_internal_tensor_assign_1_squeeze_mask_0, stride = new_cache_9_internal_tensor_assign_1_stride_0, x = reshape_40)[name = tensor("slice_by_index_8")]; + tensor reshape_41_shape_0 = const()[name = tensor("reshape_41_shape_0"), val = tensor([-1])]; + tensor reshape_41 = reshape(shape = reshape_41_shape_0, x = slice_by_index_8)[name = tensor("reshape_41")]; + tensor reshape_42_shape_0 = const()[name = tensor("reshape_42_shape_0"), val = tensor([-1])]; + tensor reshape_42_cast_fp16 = reshape(shape = reshape_42_shape_0, x = var_2086_cast_fp16)[name = tensor("reshape_42_cast_fp16")]; + tensor reshape_43_shape_0 = const()[name = tensor("reshape_43_shape_0"), val = tensor([-1])]; + tensor reshape_43_cast_fp16 = reshape(shape = reshape_43_shape_0, x = cache4_to_fp16)[name = tensor("reshape_43_cast_fp16")]; + tensor scatter_8_mode_0 = const()[name = tensor("scatter_8_mode_0"), val = tensor("update")]; + tensor scatter_8_axis_0 = const()[name = tensor("scatter_8_axis_0"), val = tensor(0)]; + tensor scatter_8_validate_indices_0 = const()[name = tensor("scatter_8_validate_indices_0"), val = tensor(false)]; + tensor scatter_8_cast_fp16 = scatter(axis = scatter_8_axis_0, data = reshape_43_cast_fp16, indices = reshape_41, mode = scatter_8_mode_0, updates = reshape_42_cast_fp16, validate_indices = scatter_8_validate_indices_0)[name = tensor("scatter_8_cast_fp16")]; + tensor reshape_44_cast_fp16 = reshape(shape = shape_54, x = scatter_8_cast_fp16)[name = tensor("reshape_44_cast_fp16")]; + tensor var_2094_begin_0 = const()[name = tensor("op_2094_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor var_2094_end_0 = const()[name = tensor("op_2094_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor var_2094_end_mask_0 = const()[name = tensor("op_2094_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_2094_squeeze_mask_0 = const()[name = tensor("op_2094_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor var_2094_cast_fp16 = slice_by_index(begin = var_2094_begin_0, end = var_2094_end_0, end_mask = var_2094_end_mask_0, squeeze_mask = var_2094_squeeze_mask_0, x = reshape_44_cast_fp16)[name = tensor("op_2094_cast_fp16")]; + tensor var_2096_axis_0 = const()[name = tensor("op_2096_axis_0"), val = tensor(1)]; + tensor var_2096_mode_0 = const()[name = tensor("op_2096_mode_0"), val = tensor("update")]; + tensor var_2096_validate_indices_0 = const()[name = tensor("op_2096_validate_indices_0"), val = tensor(false)]; + tensor var_2096_cast_fp16 = scatter_along_axis(axis = var_2096_axis_0, data = var_2094_cast_fp16, indices = write_indices_9, mode = var_2096_mode_0, updates = v_9_cast_fp16, validate_indices = var_2096_validate_indices_0)[name = tensor("op_2096_cast_fp16")]; + tensor concat_36 = const()[name = tensor("concat_36"), val = tensor([1, 0, 0, 0, 0])]; + tensor concat_37 = const()[name = tensor("concat_37"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_9_internal_tensor_assign_2_stride_0 = const()[name = tensor("new_cache_9_internal_tensor_assign_2_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_9_internal_tensor_assign_2_begin_mask_0 = const()[name = tensor("new_cache_9_internal_tensor_assign_2_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_9_internal_tensor_assign_2_end_mask_0 = const()[name = tensor("new_cache_9_internal_tensor_assign_2_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_9_internal_tensor_assign_2_squeeze_mask_0 = const()[name = tensor("new_cache_9_internal_tensor_assign_2_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_55 = const()[name = tensor("shape_55"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_9 = const()[name = tensor("reduce_prod_9"), val = tensor(1048576)]; + tensor range_1d_9_start_0 = const()[name = tensor("range_1d_9_start_0"), val = tensor(0)]; + tensor range_1d_9_step_0 = const()[name = tensor("range_1d_9_step_0"), val = tensor(1)]; + tensor range_1d_9 = range_1d(end = reduce_prod_9, start = range_1d_9_start_0, step = range_1d_9_step_0)[name = tensor("range_1d_9")]; + tensor reshape_45 = reshape(shape = shape_55, x = range_1d_9)[name = tensor("reshape_45")]; + tensor slice_by_index_9 = slice_by_index(begin = concat_36, begin_mask = new_cache_9_internal_tensor_assign_2_begin_mask_0, end = concat_37, end_mask = new_cache_9_internal_tensor_assign_2_end_mask_0, squeeze_mask = new_cache_9_internal_tensor_assign_2_squeeze_mask_0, stride = new_cache_9_internal_tensor_assign_2_stride_0, x = reshape_45)[name = tensor("slice_by_index_9")]; + tensor reshape_46_shape_0 = const()[name = tensor("reshape_46_shape_0"), val = tensor([-1])]; + tensor reshape_46 = reshape(shape = reshape_46_shape_0, x = slice_by_index_9)[name = tensor("reshape_46")]; + tensor reshape_47_shape_0 = const()[name = tensor("reshape_47_shape_0"), val = tensor([-1])]; + tensor reshape_47_cast_fp16 = reshape(shape = reshape_47_shape_0, x = var_2096_cast_fp16)[name = tensor("reshape_47_cast_fp16")]; + tensor reshape_48_shape_0 = const()[name = tensor("reshape_48_shape_0"), val = tensor([-1])]; + tensor reshape_48_cast_fp16 = reshape(shape = reshape_48_shape_0, x = reshape_44_cast_fp16)[name = tensor("reshape_48_cast_fp16")]; + tensor scatter_9_mode_0 = const()[name = tensor("scatter_9_mode_0"), val = tensor("update")]; + tensor scatter_9_axis_0 = const()[name = tensor("scatter_9_axis_0"), val = tensor(0)]; + tensor scatter_9_validate_indices_0 = const()[name = tensor("scatter_9_validate_indices_0"), val = tensor(false)]; + tensor scatter_9_cast_fp16 = scatter(axis = scatter_9_axis_0, data = reshape_48_cast_fp16, indices = reshape_46, mode = scatter_9_mode_0, updates = reshape_47_cast_fp16, validate_indices = scatter_9_validate_indices_0)[name = tensor("scatter_9_cast_fp16")]; + tensor reshape_49_cast_fp16 = reshape(shape = shape_55, x = scatter_9_cast_fp16)[name = tensor("reshape_49_cast_fp16")]; + tensor reshape_49_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("reshape_49_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor keys_25_begin_0 = const()[name = tensor("keys_25_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor keys_25_end_0 = const()[name = tensor("keys_25_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor keys_25_end_mask_0 = const()[name = tensor("keys_25_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor keys_25_squeeze_mask_0 = const()[name = tensor("keys_25_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor keys_25_cast_fp16 = slice_by_index(begin = keys_25_begin_0, end = keys_25_end_0, end_mask = keys_25_end_mask_0, squeeze_mask = keys_25_squeeze_mask_0, x = reshape_49_cast_fp16)[name = tensor("keys_25_cast_fp16")]; + tensor values_25_begin_0 = const()[name = tensor("values_25_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor values_25_end_0 = const()[name = tensor("values_25_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor values_25_end_mask_0 = const()[name = tensor("values_25_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor values_25_squeeze_mask_0 = const()[name = tensor("values_25_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor values_25_cast_fp16 = slice_by_index(begin = values_25_begin_0, end = values_25_end_0, end_mask = values_25_end_mask_0, squeeze_mask = values_25_squeeze_mask_0, x = reshape_49_cast_fp16)[name = tensor("values_25_cast_fp16")]; + tensor var_2108_cast_fp16 = not_equal(x = keys_25_cast_fp16, y = keys_25_cast_fp16)[name = tensor("op_2108_cast_fp16")]; + tensor keys_27_cast_fp16 = select(a = var_470_to_fp16, b = keys_25_cast_fp16, cond = var_2108_cast_fp16)[name = tensor("keys_27_cast_fp16")]; + tensor var_2116_cast_fp16 = not_equal(x = values_25_cast_fp16, y = values_25_cast_fp16)[name = tensor("op_2116_cast_fp16")]; + tensor values_27_cast_fp16 = select(a = var_470_to_fp16, b = values_25_cast_fp16, cond = var_2116_cast_fp16)[name = tensor("values_27_cast_fp16")]; + tensor var_2140 = const()[name = tensor("op_2140"), val = tensor([0, 2, 1, 3])]; + tensor var_2156 = const()[name = tensor("op_2156"), val = tensor([1, 1, 1])]; + tensor var_2157_cast_fp16 = reshape(shape = var_2156, x = position4_to_fp16)[name = tensor("op_2157_cast_fp16")]; + tensor q_positions_9_cast_fp16 = add(x = var_2157_cast_fp16, y = q_offsets_1_promoted_to_fp16)[name = tensor("q_positions_9_cast_fp16")]; + tensor valid_end_9_cast_fp16 = add(x = var_2157_cast_fp16, y = var_536_cast_fp16)[name = tensor("valid_end_9_cast_fp16")]; + tensor valid_mask_9_cast_fp16 = less(x = k_positions_1_promoted_to_fp16, y = valid_end_9_cast_fp16)[name = tensor("valid_mask_9_cast_fp16")]; + tensor causal_mask_9_cast_fp16 = less_equal(x = k_positions_1_promoted_to_fp16, y = q_positions_9_cast_fp16)[name = tensor("causal_mask_9_cast_fp16")]; + tensor var_2185 = logical_and(x = valid_mask_9_cast_fp16, y = causal_mask_9_cast_fp16)[name = tensor("op_2185")]; + tensor attn_mask_9_axes_0 = const()[name = tensor("attn_mask_9_axes_0"), val = tensor([1])]; + tensor attn_mask_9 = expand_dims(axes = attn_mask_9_axes_0, x = var_2185)[name = tensor("attn_mask_9")]; + tensor var_2199_transpose_x_0 = const()[name = tensor("op_2199_transpose_x_0"), val = tensor(false)]; + tensor var_2199_transpose_y_0 = const()[name = tensor("op_2199_transpose_y_0"), val = tensor(false)]; + tensor transpose_77_perm_0 = const()[name = tensor("transpose_77_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_78_perm_0 = const()[name = tensor("transpose_78_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_78 = transpose(perm = transpose_78_perm_0, x = keys_27_cast_fp16)[name = tensor("transpose_188")]; + tensor transpose_77 = transpose(perm = transpose_77_perm_0, x = q_27_cast_fp16)[name = tensor("transpose_189")]; + tensor var_2199_cast_fp16 = matmul(transpose_x = var_2199_transpose_x_0, transpose_y = var_2199_transpose_y_0, x = transpose_77, y = transpose_78)[name = tensor("op_2199_cast_fp16")]; + tensor var_2193_to_fp16 = const()[name = tensor("op_2193_to_fp16"), val = tensor([0x1p-3])]; + tensor attn_weights_25_cast_fp16 = mul(x = var_2199_cast_fp16, y = var_2193_to_fp16)[name = tensor("attn_weights_25_cast_fp16")]; + tensor var_2201 = logical_not(x = attn_mask_9)[name = tensor("op_2201")]; + tensor var_2202_to_fp16 = const()[name = tensor("op_2202_to_fp16"), val = tensor(-inf)]; + tensor attn_weights_27_cast_fp16 = select(a = var_2202_to_fp16, b = attn_weights_25_cast_fp16, cond = var_2201)[name = tensor("attn_weights_27_cast_fp16")]; + tensor var_2204 = const()[name = tensor("op_2204"), val = tensor(-1)]; + tensor attn_weights_29_cast_fp16 = softmax(axis = var_2204, x = attn_weights_27_cast_fp16)[name = tensor("attn_weights_29_cast_fp16")]; + tensor attn_output_9_transpose_x_0 = const()[name = tensor("attn_output_9_transpose_x_0"), val = tensor(false)]; + tensor attn_output_9_transpose_y_0 = const()[name = tensor("attn_output_9_transpose_y_0"), val = tensor(false)]; + tensor values_29_cast_fp16 = transpose(perm = var_2140, x = values_27_cast_fp16)[name = tensor("transpose_190")]; + tensor attn_output_9_cast_fp16 = matmul(transpose_x = attn_output_9_transpose_x_0, transpose_y = attn_output_9_transpose_y_0, x = attn_weights_29_cast_fp16, y = values_29_cast_fp16)[name = tensor("attn_output_9_cast_fp16")]; + tensor var_2212 = const()[name = tensor("op_2212"), val = tensor([0, 2, 1, 3])]; + tensor var_2217 = const()[name = tensor("op_2217"), val = tensor([1, 256, 1024])]; + tensor var_2213_cast_fp16 = transpose(perm = var_2212, x = attn_output_9_cast_fp16)[name = tensor("transpose_187")]; + tensor input_53_cast_fp16 = reshape(shape = var_2217, x = var_2213_cast_fp16)[name = tensor("input_53_cast_fp16")]; + tensor attn4_out_proj_weight_to_fp16 = const()[name = tensor("attn4_out_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(108062592)))]; + tensor linear_17_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = attn4_out_proj_weight_to_fp16, x = input_53_cast_fp16)[name = tensor("linear_17_cast_fp16")]; + tensor var_2226_cast_fp16 = add(x = position4_to_fp16, y = valid_len_to_fp16)[name = tensor("op_2226_cast_fp16")]; + tensor var_2226_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_2226_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor input_55_cast_fp16 = add(x = input_49_cast_fp16, y = linear_17_cast_fp16)[name = tensor("input_55_cast_fp16")]; + tensor input_57_axes_0 = const()[name = tensor("input_57_axes_0"), val = tensor([-1])]; + tensor norm4_2_weight_to_fp16 = const()[name = tensor("norm4_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(110159808)))]; + tensor norm4_2_bias_to_fp16 = const()[name = tensor("norm4_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(110161920)))]; + tensor var_2230_to_fp16 = const()[name = tensor("op_2230_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_57_cast_fp16 = layer_norm(axes = input_57_axes_0, beta = norm4_2_bias_to_fp16, epsilon = var_2230_to_fp16, gamma = norm4_2_weight_to_fp16, x = input_55_cast_fp16)[name = tensor("input_57_cast_fp16")]; + tensor linear4_1_weight_to_fp16 = const()[name = tensor("linear4_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(110164032)))]; + tensor linear_18_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = linear4_1_weight_to_fp16, x = input_57_cast_fp16)[name = tensor("linear_18_cast_fp16")]; + tensor input_59_mode_0 = const()[name = tensor("input_59_mode_0"), val = tensor("EXACT")]; + tensor input_59_cast_fp16 = gelu(mode = input_59_mode_0, x = linear_18_cast_fp16)[name = tensor("input_59_cast_fp16")]; + tensor linear4_2_weight_to_fp16 = const()[name = tensor("linear4_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(118552704)))]; + tensor linear_19_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = linear4_2_weight_to_fp16, x = input_59_cast_fp16)[name = tensor("linear_19_cast_fp16")]; + tensor input_61_cast_fp16 = add(x = input_55_cast_fp16, y = linear_19_cast_fp16)[name = tensor("input_61_cast_fp16")]; + tensor input_63_axes_0 = const()[name = tensor("input_63_axes_0"), val = tensor([-1])]; + tensor norm5_1_weight_to_fp16 = const()[name = tensor("norm5_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(126941376)))]; + tensor norm5_1_bias_to_fp16 = const()[name = tensor("norm5_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(126943488)))]; + tensor var_2247_to_fp16 = const()[name = tensor("op_2247_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_63_cast_fp16 = layer_norm(axes = input_63_axes_0, beta = norm5_1_bias_to_fp16, epsilon = var_2247_to_fp16, gamma = norm5_1_weight_to_fp16, x = input_61_cast_fp16)[name = tensor("input_63_cast_fp16")]; + tensor attn5_in_proj_weight_to_fp16 = const()[name = tensor("attn5_in_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(126945600)))]; + tensor linear_20_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = attn5_in_proj_weight_to_fp16, x = input_63_cast_fp16)[name = tensor("linear_20_cast_fp16")]; + tensor var_2261 = const()[name = tensor("op_2261"), val = tensor([1, 256, 3, 16, 64])]; + tensor qkv_11_cast_fp16 = reshape(shape = var_2261, x = linear_20_cast_fp16)[name = tensor("qkv_11_cast_fp16")]; + tensor q_31_begin_0 = const()[name = tensor("q_31_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor q_31_end_0 = const()[name = tensor("q_31_end_0"), val = tensor([1, 256, 1, 16, 64])]; + tensor q_31_end_mask_0 = const()[name = tensor("q_31_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor q_31_squeeze_mask_0 = const()[name = tensor("q_31_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor q_31_cast_fp16 = slice_by_index(begin = q_31_begin_0, end = q_31_end_0, end_mask = q_31_end_mask_0, squeeze_mask = q_31_squeeze_mask_0, x = qkv_11_cast_fp16)[name = tensor("q_31_cast_fp16")]; + tensor k_21_begin_0 = const()[name = tensor("k_21_begin_0"), val = tensor([0, 0, 1, 0, 0])]; + tensor k_21_end_0 = const()[name = tensor("k_21_end_0"), val = tensor([1, 256, 2, 16, 64])]; + tensor k_21_end_mask_0 = const()[name = tensor("k_21_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor k_21_squeeze_mask_0 = const()[name = tensor("k_21_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor k_21_cast_fp16 = slice_by_index(begin = k_21_begin_0, end = k_21_end_0, end_mask = k_21_end_mask_0, squeeze_mask = k_21_squeeze_mask_0, x = qkv_11_cast_fp16)[name = tensor("k_21_cast_fp16")]; + tensor v_11_begin_0 = const()[name = tensor("v_11_begin_0"), val = tensor([0, 0, 2, 0, 0])]; + tensor v_11_end_0 = const()[name = tensor("v_11_end_0"), val = tensor([1, 256, 3, 16, 64])]; + tensor v_11_end_mask_0 = const()[name = tensor("v_11_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor v_11_squeeze_mask_0 = const()[name = tensor("v_11_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor v_11_cast_fp16 = slice_by_index(begin = v_11_begin_0, end = v_11_end_0, end_mask = v_11_end_mask_0, squeeze_mask = v_11_squeeze_mask_0, x = qkv_11_cast_fp16)[name = tensor("v_11_cast_fp16")]; + tensor var_2319 = const()[name = tensor("op_2319"), val = tensor([1, 1])]; + tensor position5_to_fp16_dtype_0 = const()[name = tensor("position5_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor position5_to_fp16 = cast(dtype = position5_to_fp16_dtype_0, x = position5)[name = tensor("cast_636")]; + tensor var_2320_cast_fp16 = reshape(shape = var_2319, x = position5_to_fp16)[name = tensor("op_2320_cast_fp16")]; + tensor ts_33_cast_fp16 = add(x = ts_1_promoted_to_fp16, y = var_2320_cast_fp16)[name = tensor("ts_33_cast_fp16")]; + tensor var_2327 = const()[name = tensor("op_2327"), val = tensor([1, 256, 1, 1])]; + tensor ts_35_cast_fp16 = reshape(shape = var_2327, x = ts_33_cast_fp16)[name = tensor("ts_35_cast_fp16")]; + tensor var_2334 = const()[name = tensor("op_2334"), val = tensor([1, 256, 16, 32, 2])]; + tensor q_complex_11_cast_fp16 = reshape(shape = var_2334, x = q_31_cast_fp16)[name = tensor("q_complex_11_cast_fp16")]; + tensor var_2341 = const()[name = tensor("op_2341"), val = tensor([1, 256, 16, 32, 2])]; + tensor k_complex_11_cast_fp16 = reshape(shape = var_2341, x = k_21_cast_fp16)[name = tensor("k_complex_11_cast_fp16")]; + tensor var_2345_begin_0 = const()[name = tensor("op_2345_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_2345_end_0 = const()[name = tensor("op_2345_end_0"), val = tensor([1, 256, 16, 32, 1])]; + tensor var_2345_end_mask_0 = const()[name = tensor("op_2345_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_2345_squeeze_mask_0 = const()[name = tensor("op_2345_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_2345_cast_fp16 = slice_by_index(begin = var_2345_begin_0, end = var_2345_end_0, end_mask = var_2345_end_mask_0, squeeze_mask = var_2345_squeeze_mask_0, x = q_complex_11_cast_fp16)[name = tensor("op_2345_cast_fp16")]; + tensor var_2353_begin_0 = const()[name = tensor("op_2353_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_2353_end_0 = const()[name = tensor("op_2353_end_0"), val = tensor([1, 256, 16, 32, 2])]; + tensor var_2353_end_mask_0 = const()[name = tensor("op_2353_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_2353_squeeze_mask_0 = const()[name = tensor("op_2353_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_2353_cast_fp16 = slice_by_index(begin = var_2353_begin_0, end = var_2353_end_0, end_mask = var_2353_end_mask_0, squeeze_mask = var_2353_squeeze_mask_0, x = q_complex_11_cast_fp16)[name = tensor("op_2353_cast_fp16")]; + tensor var_2361_begin_0 = const()[name = tensor("op_2361_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_2361_end_0 = const()[name = tensor("op_2361_end_0"), val = tensor([1, 256, 16, 32, 1])]; + tensor var_2361_end_mask_0 = const()[name = tensor("op_2361_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_2361_squeeze_mask_0 = const()[name = tensor("op_2361_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_2361_cast_fp16 = slice_by_index(begin = var_2361_begin_0, end = var_2361_end_0, end_mask = var_2361_end_mask_0, squeeze_mask = var_2361_squeeze_mask_0, x = k_complex_11_cast_fp16)[name = tensor("op_2361_cast_fp16")]; + tensor var_2369_begin_0 = const()[name = tensor("op_2369_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_2369_end_0 = const()[name = tensor("op_2369_end_0"), val = tensor([1, 256, 16, 32, 2])]; + tensor var_2369_end_mask_0 = const()[name = tensor("op_2369_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_2369_squeeze_mask_0 = const()[name = tensor("op_2369_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_2369_cast_fp16 = slice_by_index(begin = var_2369_begin_0, end = var_2369_end_0, end_mask = var_2369_end_mask_0, squeeze_mask = var_2369_squeeze_mask_0, x = k_complex_11_cast_fp16)[name = tensor("op_2369_cast_fp16")]; + tensor freqs_11_to_fp16 = const()[name = tensor("freqs_11_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(133237120)))]; + tensor var_2375_cast_fp16 = mul(x = freqs_11_to_fp16, y = ts_35_cast_fp16)[name = tensor("op_2375_cast_fp16")]; + tensor rotr_11_cast_fp16 = cos(x = var_2375_cast_fp16)[name = tensor("rotr_11_cast_fp16")]; + tensor roti_11_cast_fp16 = sin(x = var_2375_cast_fp16)[name = tensor("roti_11_cast_fp16")]; + tensor var_2379_cast_fp16 = mul(x = var_2345_cast_fp16, y = rotr_11_cast_fp16)[name = tensor("op_2379_cast_fp16")]; + tensor var_2380_cast_fp16 = mul(x = var_2353_cast_fp16, y = roti_11_cast_fp16)[name = tensor("op_2380_cast_fp16")]; + tensor qor_21_cast_fp16 = sub(x = var_2379_cast_fp16, y = var_2380_cast_fp16)[name = tensor("qor_21_cast_fp16")]; + tensor var_2383_cast_fp16 = mul(x = var_2345_cast_fp16, y = roti_11_cast_fp16)[name = tensor("op_2383_cast_fp16")]; + tensor var_2384_cast_fp16 = mul(x = var_2353_cast_fp16, y = rotr_11_cast_fp16)[name = tensor("op_2384_cast_fp16")]; + tensor qoi_21_cast_fp16 = add(x = var_2383_cast_fp16, y = var_2384_cast_fp16)[name = tensor("qoi_21_cast_fp16")]; + tensor var_2387_cast_fp16 = mul(x = var_2361_cast_fp16, y = rotr_11_cast_fp16)[name = tensor("op_2387_cast_fp16")]; + tensor var_2388_cast_fp16 = mul(x = var_2369_cast_fp16, y = roti_11_cast_fp16)[name = tensor("op_2388_cast_fp16")]; + tensor kor_21_cast_fp16 = sub(x = var_2387_cast_fp16, y = var_2388_cast_fp16)[name = tensor("kor_21_cast_fp16")]; + tensor var_2391_cast_fp16 = mul(x = var_2361_cast_fp16, y = roti_11_cast_fp16)[name = tensor("op_2391_cast_fp16")]; + tensor var_2392_cast_fp16 = mul(x = var_2369_cast_fp16, y = rotr_11_cast_fp16)[name = tensor("op_2392_cast_fp16")]; + tensor koi_21_cast_fp16 = add(x = var_2391_cast_fp16, y = var_2392_cast_fp16)[name = tensor("koi_21_cast_fp16")]; + tensor qo_11_axis_0 = const()[name = tensor("qo_11_axis_0"), val = tensor(-1)]; + tensor qo_11_cast_fp16 = stack(axis = qo_11_axis_0, values = (qor_21_cast_fp16, qoi_21_cast_fp16))[name = tensor("qo_11_cast_fp16")]; + tensor ko_11_axis_0 = const()[name = tensor("ko_11_axis_0"), val = tensor(-1)]; + tensor ko_11_cast_fp16 = stack(axis = ko_11_axis_0, values = (kor_21_cast_fp16, koi_21_cast_fp16))[name = tensor("ko_11_cast_fp16")]; + tensor var_2425 = const()[name = tensor("op_2425"), val = tensor([1, 256, 16, 64])]; + tensor q_33_cast_fp16 = reshape(shape = var_2425, x = qo_11_cast_fp16)[name = tensor("q_33_cast_fp16")]; + tensor var_2431 = const()[name = tensor("op_2431"), val = tensor([1, 256, 16, 64])]; + tensor k_23_cast_fp16 = reshape(shape = var_2431, x = ko_11_cast_fp16)[name = tensor("k_23_cast_fp16")]; + tensor write_indices_float_31_cast_fp16 = add(x = var_2320_cast_fp16, y = var_393_promoted_to_fp16)[name = tensor("write_indices_float_31_cast_fp16")]; + tensor _inversed_2452_y_0_to_fp16 = const()[name = tensor("_inversed_2452_y_0_to_fp16"), val = tensor(0x1p-9)]; + tensor _inversed_2452_cast_fp16 = mul(x = write_indices_float_31_cast_fp16, y = _inversed_2452_y_0_to_fp16)[name = tensor("_inversed_2452_cast_fp16")]; + tensor var_2453_cast_fp16 = floor(x = _inversed_2452_cast_fp16)[name = tensor("op_2453_cast_fp16")]; + tensor var_2454_to_fp16 = const()[name = tensor("op_2454_to_fp16"), val = tensor(0x1p+9)]; + tensor var_2455_cast_fp16 = mul(x = var_2453_cast_fp16, y = var_2454_to_fp16)[name = tensor("op_2455_cast_fp16")]; + tensor write_indices_float_33_cast_fp16 = sub(x = write_indices_float_31_cast_fp16, y = var_2455_cast_fp16)[name = tensor("write_indices_float_33_cast_fp16")]; + tensor var_2473_value_0_to_fp16 = const()[name = tensor("op_2473_value_0_to_fp16"), val = tensor(0x1.ffp+8)]; + tensor var_2473_cast_fp16 = fill_like(ref_tensor = write_indices_float_33_cast_fp16, value = var_2473_value_0_to_fp16)[name = tensor("op_2473_cast_fp16")]; + tensor write_indices_float_35_cast_fp16 = select(a = write_indices_float_33_cast_fp16, b = var_2473_cast_fp16, cond = token_valid_1_cast_fp16)[name = tensor("write_indices_float_35_cast_fp16")]; + tensor var_2479_dtype_0 = const()[name = tensor("op_2479_dtype_0"), val = tensor("int32")]; + tensor var_2484 = const()[name = tensor("op_2484"), val = tensor([1, 256, 1, 1])]; + tensor write_indices_float_35_cast_fp16_to_int32 = cast(dtype = var_2479_dtype_0, x = write_indices_float_35_cast_fp16)[name = tensor("cast_635")]; + tensor var_2485 = reshape(shape = var_2484, x = write_indices_float_35_cast_fp16_to_int32)[name = tensor("op_2485")]; + tensor write_indices_11_reps_0 = const()[name = tensor("write_indices_11_reps_0"), val = tensor([1, 1, 16, 64])]; + tensor write_indices_11 = tile(reps = write_indices_11_reps_0, x = var_2485)[name = tensor("write_indices_11")]; + tensor var_2495_begin_0 = const()[name = tensor("op_2495_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_2495_end_0 = const()[name = tensor("op_2495_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor var_2495_end_mask_0 = const()[name = tensor("op_2495_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_2495_squeeze_mask_0 = const()[name = tensor("op_2495_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor cache5_to_fp16_dtype_0 = const()[name = tensor("cache5_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor cache5_to_fp16 = cast(dtype = cache5_to_fp16_dtype_0, x = cache5)[name = tensor("cast_634")]; + tensor var_2495_cast_fp16 = slice_by_index(begin = var_2495_begin_0, end = var_2495_end_0, end_mask = var_2495_end_mask_0, squeeze_mask = var_2495_squeeze_mask_0, x = cache5_to_fp16)[name = tensor("op_2495_cast_fp16")]; + tensor var_2497_axis_0 = const()[name = tensor("op_2497_axis_0"), val = tensor(1)]; + tensor var_2497_mode_0 = const()[name = tensor("op_2497_mode_0"), val = tensor("update")]; + tensor var_2497_validate_indices_0 = const()[name = tensor("op_2497_validate_indices_0"), val = tensor(false)]; + tensor var_2497_cast_fp16 = scatter_along_axis(axis = var_2497_axis_0, data = var_2495_cast_fp16, indices = write_indices_11, mode = var_2497_mode_0, updates = k_23_cast_fp16, validate_indices = var_2497_validate_indices_0)[name = tensor("op_2497_cast_fp16")]; + tensor concat_42 = const()[name = tensor("concat_42"), val = tensor([0, 0, 0, 0, 0])]; + tensor concat_43 = const()[name = tensor("concat_43"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_11_internal_tensor_assign_1_stride_0 = const()[name = tensor("new_cache_11_internal_tensor_assign_1_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_11_internal_tensor_assign_1_begin_mask_0 = const()[name = tensor("new_cache_11_internal_tensor_assign_1_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_11_internal_tensor_assign_1_end_mask_0 = const()[name = tensor("new_cache_11_internal_tensor_assign_1_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_11_internal_tensor_assign_1_squeeze_mask_0 = const()[name = tensor("new_cache_11_internal_tensor_assign_1_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_56 = const()[name = tensor("shape_56"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_10 = const()[name = tensor("reduce_prod_10"), val = tensor(1048576)]; + tensor range_1d_10_start_0 = const()[name = tensor("range_1d_10_start_0"), val = tensor(0)]; + tensor range_1d_10_step_0 = const()[name = tensor("range_1d_10_step_0"), val = tensor(1)]; + tensor range_1d_10 = range_1d(end = reduce_prod_10, start = range_1d_10_start_0, step = range_1d_10_step_0)[name = tensor("range_1d_10")]; + tensor reshape_50 = reshape(shape = shape_56, x = range_1d_10)[name = tensor("reshape_50")]; + tensor slice_by_index_10 = slice_by_index(begin = concat_42, begin_mask = new_cache_11_internal_tensor_assign_1_begin_mask_0, end = concat_43, end_mask = new_cache_11_internal_tensor_assign_1_end_mask_0, squeeze_mask = new_cache_11_internal_tensor_assign_1_squeeze_mask_0, stride = new_cache_11_internal_tensor_assign_1_stride_0, x = reshape_50)[name = tensor("slice_by_index_10")]; + tensor reshape_51_shape_0 = const()[name = tensor("reshape_51_shape_0"), val = tensor([-1])]; + tensor reshape_51 = reshape(shape = reshape_51_shape_0, x = slice_by_index_10)[name = tensor("reshape_51")]; + tensor reshape_52_shape_0 = const()[name = tensor("reshape_52_shape_0"), val = tensor([-1])]; + tensor reshape_52_cast_fp16 = reshape(shape = reshape_52_shape_0, x = var_2497_cast_fp16)[name = tensor("reshape_52_cast_fp16")]; + tensor reshape_53_shape_0 = const()[name = tensor("reshape_53_shape_0"), val = tensor([-1])]; + tensor reshape_53_cast_fp16 = reshape(shape = reshape_53_shape_0, x = cache5_to_fp16)[name = tensor("reshape_53_cast_fp16")]; + tensor scatter_10_mode_0 = const()[name = tensor("scatter_10_mode_0"), val = tensor("update")]; + tensor scatter_10_axis_0 = const()[name = tensor("scatter_10_axis_0"), val = tensor(0)]; + tensor scatter_10_validate_indices_0 = const()[name = tensor("scatter_10_validate_indices_0"), val = tensor(false)]; + tensor scatter_10_cast_fp16 = scatter(axis = scatter_10_axis_0, data = reshape_53_cast_fp16, indices = reshape_51, mode = scatter_10_mode_0, updates = reshape_52_cast_fp16, validate_indices = scatter_10_validate_indices_0)[name = tensor("scatter_10_cast_fp16")]; + tensor reshape_54_cast_fp16 = reshape(shape = shape_56, x = scatter_10_cast_fp16)[name = tensor("reshape_54_cast_fp16")]; + tensor var_2505_begin_0 = const()[name = tensor("op_2505_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor var_2505_end_0 = const()[name = tensor("op_2505_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor var_2505_end_mask_0 = const()[name = tensor("op_2505_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_2505_squeeze_mask_0 = const()[name = tensor("op_2505_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor var_2505_cast_fp16 = slice_by_index(begin = var_2505_begin_0, end = var_2505_end_0, end_mask = var_2505_end_mask_0, squeeze_mask = var_2505_squeeze_mask_0, x = reshape_54_cast_fp16)[name = tensor("op_2505_cast_fp16")]; + tensor var_2507_axis_0 = const()[name = tensor("op_2507_axis_0"), val = tensor(1)]; + tensor var_2507_mode_0 = const()[name = tensor("op_2507_mode_0"), val = tensor("update")]; + tensor var_2507_validate_indices_0 = const()[name = tensor("op_2507_validate_indices_0"), val = tensor(false)]; + tensor var_2507_cast_fp16 = scatter_along_axis(axis = var_2507_axis_0, data = var_2505_cast_fp16, indices = write_indices_11, mode = var_2507_mode_0, updates = v_11_cast_fp16, validate_indices = var_2507_validate_indices_0)[name = tensor("op_2507_cast_fp16")]; + tensor concat_44 = const()[name = tensor("concat_44"), val = tensor([1, 0, 0, 0, 0])]; + tensor concat_45 = const()[name = tensor("concat_45"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_11_internal_tensor_assign_2_stride_0 = const()[name = tensor("new_cache_11_internal_tensor_assign_2_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_11_internal_tensor_assign_2_begin_mask_0 = const()[name = tensor("new_cache_11_internal_tensor_assign_2_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_11_internal_tensor_assign_2_end_mask_0 = const()[name = tensor("new_cache_11_internal_tensor_assign_2_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_11_internal_tensor_assign_2_squeeze_mask_0 = const()[name = tensor("new_cache_11_internal_tensor_assign_2_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_57 = const()[name = tensor("shape_57"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_11 = const()[name = tensor("reduce_prod_11"), val = tensor(1048576)]; + tensor range_1d_11_start_0 = const()[name = tensor("range_1d_11_start_0"), val = tensor(0)]; + tensor range_1d_11_step_0 = const()[name = tensor("range_1d_11_step_0"), val = tensor(1)]; + tensor range_1d_11 = range_1d(end = reduce_prod_11, start = range_1d_11_start_0, step = range_1d_11_step_0)[name = tensor("range_1d_11")]; + tensor reshape_55 = reshape(shape = shape_57, x = range_1d_11)[name = tensor("reshape_55")]; + tensor slice_by_index_11 = slice_by_index(begin = concat_44, begin_mask = new_cache_11_internal_tensor_assign_2_begin_mask_0, end = concat_45, end_mask = new_cache_11_internal_tensor_assign_2_end_mask_0, squeeze_mask = new_cache_11_internal_tensor_assign_2_squeeze_mask_0, stride = new_cache_11_internal_tensor_assign_2_stride_0, x = reshape_55)[name = tensor("slice_by_index_11")]; + tensor reshape_56_shape_0 = const()[name = tensor("reshape_56_shape_0"), val = tensor([-1])]; + tensor reshape_56 = reshape(shape = reshape_56_shape_0, x = slice_by_index_11)[name = tensor("reshape_56")]; + tensor reshape_57_shape_0 = const()[name = tensor("reshape_57_shape_0"), val = tensor([-1])]; + tensor reshape_57_cast_fp16 = reshape(shape = reshape_57_shape_0, x = var_2507_cast_fp16)[name = tensor("reshape_57_cast_fp16")]; + tensor reshape_58_shape_0 = const()[name = tensor("reshape_58_shape_0"), val = tensor([-1])]; + tensor reshape_58_cast_fp16 = reshape(shape = reshape_58_shape_0, x = reshape_54_cast_fp16)[name = tensor("reshape_58_cast_fp16")]; + tensor scatter_11_mode_0 = const()[name = tensor("scatter_11_mode_0"), val = tensor("update")]; + tensor scatter_11_axis_0 = const()[name = tensor("scatter_11_axis_0"), val = tensor(0)]; + tensor scatter_11_validate_indices_0 = const()[name = tensor("scatter_11_validate_indices_0"), val = tensor(false)]; + tensor scatter_11_cast_fp16 = scatter(axis = scatter_11_axis_0, data = reshape_58_cast_fp16, indices = reshape_56, mode = scatter_11_mode_0, updates = reshape_57_cast_fp16, validate_indices = scatter_11_validate_indices_0)[name = tensor("scatter_11_cast_fp16")]; + tensor reshape_59_cast_fp16 = reshape(shape = shape_57, x = scatter_11_cast_fp16)[name = tensor("reshape_59_cast_fp16")]; + tensor reshape_59_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("reshape_59_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor keys_31_begin_0 = const()[name = tensor("keys_31_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor keys_31_end_0 = const()[name = tensor("keys_31_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor keys_31_end_mask_0 = const()[name = tensor("keys_31_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor keys_31_squeeze_mask_0 = const()[name = tensor("keys_31_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor keys_31_cast_fp16 = slice_by_index(begin = keys_31_begin_0, end = keys_31_end_0, end_mask = keys_31_end_mask_0, squeeze_mask = keys_31_squeeze_mask_0, x = reshape_59_cast_fp16)[name = tensor("keys_31_cast_fp16")]; + tensor values_31_begin_0 = const()[name = tensor("values_31_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor values_31_end_0 = const()[name = tensor("values_31_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor values_31_end_mask_0 = const()[name = tensor("values_31_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor values_31_squeeze_mask_0 = const()[name = tensor("values_31_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor values_31_cast_fp16 = slice_by_index(begin = values_31_begin_0, end = values_31_end_0, end_mask = values_31_end_mask_0, squeeze_mask = values_31_squeeze_mask_0, x = reshape_59_cast_fp16)[name = tensor("values_31_cast_fp16")]; + tensor var_2519_cast_fp16 = not_equal(x = keys_31_cast_fp16, y = keys_31_cast_fp16)[name = tensor("op_2519_cast_fp16")]; + tensor keys_33_cast_fp16 = select(a = var_470_to_fp16, b = keys_31_cast_fp16, cond = var_2519_cast_fp16)[name = tensor("keys_33_cast_fp16")]; + tensor var_2527_cast_fp16 = not_equal(x = values_31_cast_fp16, y = values_31_cast_fp16)[name = tensor("op_2527_cast_fp16")]; + tensor values_33_cast_fp16 = select(a = var_470_to_fp16, b = values_31_cast_fp16, cond = var_2527_cast_fp16)[name = tensor("values_33_cast_fp16")]; + tensor var_2551 = const()[name = tensor("op_2551"), val = tensor([0, 2, 1, 3])]; + tensor var_2567 = const()[name = tensor("op_2567"), val = tensor([1, 1, 1])]; + tensor var_2568_cast_fp16 = reshape(shape = var_2567, x = position5_to_fp16)[name = tensor("op_2568_cast_fp16")]; + tensor q_positions_11_cast_fp16 = add(x = var_2568_cast_fp16, y = q_offsets_1_promoted_to_fp16)[name = tensor("q_positions_11_cast_fp16")]; + tensor valid_end_11_cast_fp16 = add(x = var_2568_cast_fp16, y = var_536_cast_fp16)[name = tensor("valid_end_11_cast_fp16")]; + tensor valid_mask_11_cast_fp16 = less(x = k_positions_1_promoted_to_fp16, y = valid_end_11_cast_fp16)[name = tensor("valid_mask_11_cast_fp16")]; + tensor causal_mask_11_cast_fp16 = less_equal(x = k_positions_1_promoted_to_fp16, y = q_positions_11_cast_fp16)[name = tensor("causal_mask_11_cast_fp16")]; + tensor var_2596 = logical_and(x = valid_mask_11_cast_fp16, y = causal_mask_11_cast_fp16)[name = tensor("op_2596")]; + tensor attn_mask_11_axes_0 = const()[name = tensor("attn_mask_11_axes_0"), val = tensor([1])]; + tensor attn_mask_11 = expand_dims(axes = attn_mask_11_axes_0, x = var_2596)[name = tensor("attn_mask_11")]; + tensor var_2610_transpose_x_0 = const()[name = tensor("op_2610_transpose_x_0"), val = tensor(false)]; + tensor var_2610_transpose_y_0 = const()[name = tensor("op_2610_transpose_y_0"), val = tensor(false)]; + tensor transpose_79_perm_0 = const()[name = tensor("transpose_79_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_80_perm_0 = const()[name = tensor("transpose_80_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_80 = transpose(perm = transpose_80_perm_0, x = keys_33_cast_fp16)[name = tensor("transpose_184")]; + tensor transpose_79 = transpose(perm = transpose_79_perm_0, x = q_33_cast_fp16)[name = tensor("transpose_185")]; + tensor var_2610_cast_fp16 = matmul(transpose_x = var_2610_transpose_x_0, transpose_y = var_2610_transpose_y_0, x = transpose_79, y = transpose_80)[name = tensor("op_2610_cast_fp16")]; + tensor var_2604_to_fp16 = const()[name = tensor("op_2604_to_fp16"), val = tensor([0x1p-3])]; + tensor attn_weights_31_cast_fp16 = mul(x = var_2610_cast_fp16, y = var_2604_to_fp16)[name = tensor("attn_weights_31_cast_fp16")]; + tensor var_2612 = logical_not(x = attn_mask_11)[name = tensor("op_2612")]; + tensor var_2613_to_fp16 = const()[name = tensor("op_2613_to_fp16"), val = tensor(-inf)]; + tensor attn_weights_33_cast_fp16 = select(a = var_2613_to_fp16, b = attn_weights_31_cast_fp16, cond = var_2612)[name = tensor("attn_weights_33_cast_fp16")]; + tensor var_2615 = const()[name = tensor("op_2615"), val = tensor(-1)]; + tensor attn_weights_35_cast_fp16 = softmax(axis = var_2615, x = attn_weights_33_cast_fp16)[name = tensor("attn_weights_35_cast_fp16")]; + tensor attn_output_11_transpose_x_0 = const()[name = tensor("attn_output_11_transpose_x_0"), val = tensor(false)]; + tensor attn_output_11_transpose_y_0 = const()[name = tensor("attn_output_11_transpose_y_0"), val = tensor(false)]; + tensor values_35_cast_fp16 = transpose(perm = var_2551, x = values_33_cast_fp16)[name = tensor("transpose_186")]; + tensor attn_output_11_cast_fp16 = matmul(transpose_x = attn_output_11_transpose_x_0, transpose_y = attn_output_11_transpose_y_0, x = attn_weights_35_cast_fp16, y = values_35_cast_fp16)[name = tensor("attn_output_11_cast_fp16")]; + tensor var_2623 = const()[name = tensor("op_2623"), val = tensor([0, 2, 1, 3])]; + tensor var_2628 = const()[name = tensor("op_2628"), val = tensor([1, 256, 1024])]; + tensor var_2624_cast_fp16 = transpose(perm = var_2623, x = attn_output_11_cast_fp16)[name = tensor("transpose_183")]; + tensor input_65_cast_fp16 = reshape(shape = var_2628, x = var_2624_cast_fp16)[name = tensor("input_65_cast_fp16")]; + tensor attn5_out_proj_weight_to_fp16 = const()[name = tensor("attn5_out_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(133237248)))]; + tensor linear_21_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = attn5_out_proj_weight_to_fp16, x = input_65_cast_fp16)[name = tensor("linear_21_cast_fp16")]; + tensor var_2637_cast_fp16 = add(x = position5_to_fp16, y = valid_len_to_fp16)[name = tensor("op_2637_cast_fp16")]; + tensor var_2637_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_2637_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor input_67_cast_fp16 = add(x = input_61_cast_fp16, y = linear_21_cast_fp16)[name = tensor("input_67_cast_fp16")]; + tensor input_69_axes_0 = const()[name = tensor("input_69_axes_0"), val = tensor([-1])]; + tensor norm5_2_weight_to_fp16 = const()[name = tensor("norm5_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(135334464)))]; + tensor norm5_2_bias_to_fp16 = const()[name = tensor("norm5_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(135336576)))]; + tensor var_2641_to_fp16 = const()[name = tensor("op_2641_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_69_cast_fp16 = layer_norm(axes = input_69_axes_0, beta = norm5_2_bias_to_fp16, epsilon = var_2641_to_fp16, gamma = norm5_2_weight_to_fp16, x = input_67_cast_fp16)[name = tensor("input_69_cast_fp16")]; + tensor linear5_1_weight_to_fp16 = const()[name = tensor("linear5_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(135338688)))]; + tensor linear_22_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = linear5_1_weight_to_fp16, x = input_69_cast_fp16)[name = tensor("linear_22_cast_fp16")]; + tensor input_71_mode_0 = const()[name = tensor("input_71_mode_0"), val = tensor("EXACT")]; + tensor input_71_cast_fp16 = gelu(mode = input_71_mode_0, x = linear_22_cast_fp16)[name = tensor("input_71_cast_fp16")]; + tensor linear5_2_weight_to_fp16 = const()[name = tensor("linear5_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(143727360)))]; + tensor linear_23_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = linear5_2_weight_to_fp16, x = input_71_cast_fp16)[name = tensor("linear_23_cast_fp16")]; + tensor input_73_cast_fp16 = add(x = input_67_cast_fp16, y = linear_23_cast_fp16)[name = tensor("input_73_cast_fp16")]; + tensor input_75_axes_0 = const()[name = tensor("input_75_axes_0"), val = tensor([-1])]; + tensor norm6_1_weight_to_fp16 = const()[name = tensor("norm6_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(152116032)))]; + tensor norm6_1_bias_to_fp16 = const()[name = tensor("norm6_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(152118144)))]; + tensor var_2658_to_fp16 = const()[name = tensor("op_2658_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_75_cast_fp16 = layer_norm(axes = input_75_axes_0, beta = norm6_1_bias_to_fp16, epsilon = var_2658_to_fp16, gamma = norm6_1_weight_to_fp16, x = input_73_cast_fp16)[name = tensor("input_75_cast_fp16")]; + tensor attn6_in_proj_weight_to_fp16 = const()[name = tensor("attn6_in_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(152120256)))]; + tensor linear_24_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = attn6_in_proj_weight_to_fp16, x = input_75_cast_fp16)[name = tensor("linear_24_cast_fp16")]; + tensor var_2672 = const()[name = tensor("op_2672"), val = tensor([1, 256, 3, 16, 64])]; + tensor qkv_13_cast_fp16 = reshape(shape = var_2672, x = linear_24_cast_fp16)[name = tensor("qkv_13_cast_fp16")]; + tensor q_37_begin_0 = const()[name = tensor("q_37_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor q_37_end_0 = const()[name = tensor("q_37_end_0"), val = tensor([1, 256, 1, 16, 64])]; + tensor q_37_end_mask_0 = const()[name = tensor("q_37_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor q_37_squeeze_mask_0 = const()[name = tensor("q_37_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor q_37_cast_fp16 = slice_by_index(begin = q_37_begin_0, end = q_37_end_0, end_mask = q_37_end_mask_0, squeeze_mask = q_37_squeeze_mask_0, x = qkv_13_cast_fp16)[name = tensor("q_37_cast_fp16")]; + tensor k_25_begin_0 = const()[name = tensor("k_25_begin_0"), val = tensor([0, 0, 1, 0, 0])]; + tensor k_25_end_0 = const()[name = tensor("k_25_end_0"), val = tensor([1, 256, 2, 16, 64])]; + tensor k_25_end_mask_0 = const()[name = tensor("k_25_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor k_25_squeeze_mask_0 = const()[name = tensor("k_25_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor k_25_cast_fp16 = slice_by_index(begin = k_25_begin_0, end = k_25_end_0, end_mask = k_25_end_mask_0, squeeze_mask = k_25_squeeze_mask_0, x = qkv_13_cast_fp16)[name = tensor("k_25_cast_fp16")]; + tensor v_13_begin_0 = const()[name = tensor("v_13_begin_0"), val = tensor([0, 0, 2, 0, 0])]; + tensor v_13_end_0 = const()[name = tensor("v_13_end_0"), val = tensor([1, 256, 3, 16, 64])]; + tensor v_13_end_mask_0 = const()[name = tensor("v_13_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor v_13_squeeze_mask_0 = const()[name = tensor("v_13_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor v_13_cast_fp16 = slice_by_index(begin = v_13_begin_0, end = v_13_end_0, end_mask = v_13_end_mask_0, squeeze_mask = v_13_squeeze_mask_0, x = qkv_13_cast_fp16)[name = tensor("v_13_cast_fp16")]; + tensor var_2730 = const()[name = tensor("op_2730"), val = tensor([1, 1])]; + tensor position6_to_fp16_dtype_0 = const()[name = tensor("position6_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor position6_to_fp16 = cast(dtype = position6_to_fp16_dtype_0, x = position6)[name = tensor("cast_631")]; + tensor var_2731_cast_fp16 = reshape(shape = var_2730, x = position6_to_fp16)[name = tensor("op_2731_cast_fp16")]; + tensor ts_39_cast_fp16 = add(x = ts_1_promoted_to_fp16, y = var_2731_cast_fp16)[name = tensor("ts_39_cast_fp16")]; + tensor var_2738 = const()[name = tensor("op_2738"), val = tensor([1, 256, 1, 1])]; + tensor ts_41_cast_fp16 = reshape(shape = var_2738, x = ts_39_cast_fp16)[name = tensor("ts_41_cast_fp16")]; + tensor var_2745 = const()[name = tensor("op_2745"), val = tensor([1, 256, 16, 32, 2])]; + tensor q_complex_13_cast_fp16 = reshape(shape = var_2745, x = q_37_cast_fp16)[name = tensor("q_complex_13_cast_fp16")]; + tensor var_2752 = const()[name = tensor("op_2752"), val = tensor([1, 256, 16, 32, 2])]; + tensor k_complex_13_cast_fp16 = reshape(shape = var_2752, x = k_25_cast_fp16)[name = tensor("k_complex_13_cast_fp16")]; + tensor var_2756_begin_0 = const()[name = tensor("op_2756_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_2756_end_0 = const()[name = tensor("op_2756_end_0"), val = tensor([1, 256, 16, 32, 1])]; + tensor var_2756_end_mask_0 = const()[name = tensor("op_2756_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_2756_squeeze_mask_0 = const()[name = tensor("op_2756_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_2756_cast_fp16 = slice_by_index(begin = var_2756_begin_0, end = var_2756_end_0, end_mask = var_2756_end_mask_0, squeeze_mask = var_2756_squeeze_mask_0, x = q_complex_13_cast_fp16)[name = tensor("op_2756_cast_fp16")]; + tensor var_2764_begin_0 = const()[name = tensor("op_2764_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_2764_end_0 = const()[name = tensor("op_2764_end_0"), val = tensor([1, 256, 16, 32, 2])]; + tensor var_2764_end_mask_0 = const()[name = tensor("op_2764_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_2764_squeeze_mask_0 = const()[name = tensor("op_2764_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_2764_cast_fp16 = slice_by_index(begin = var_2764_begin_0, end = var_2764_end_0, end_mask = var_2764_end_mask_0, squeeze_mask = var_2764_squeeze_mask_0, x = q_complex_13_cast_fp16)[name = tensor("op_2764_cast_fp16")]; + tensor var_2772_begin_0 = const()[name = tensor("op_2772_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_2772_end_0 = const()[name = tensor("op_2772_end_0"), val = tensor([1, 256, 16, 32, 1])]; + tensor var_2772_end_mask_0 = const()[name = tensor("op_2772_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_2772_squeeze_mask_0 = const()[name = tensor("op_2772_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_2772_cast_fp16 = slice_by_index(begin = var_2772_begin_0, end = var_2772_end_0, end_mask = var_2772_end_mask_0, squeeze_mask = var_2772_squeeze_mask_0, x = k_complex_13_cast_fp16)[name = tensor("op_2772_cast_fp16")]; + tensor var_2780_begin_0 = const()[name = tensor("op_2780_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_2780_end_0 = const()[name = tensor("op_2780_end_0"), val = tensor([1, 256, 16, 32, 2])]; + tensor var_2780_end_mask_0 = const()[name = tensor("op_2780_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_2780_squeeze_mask_0 = const()[name = tensor("op_2780_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_2780_cast_fp16 = slice_by_index(begin = var_2780_begin_0, end = var_2780_end_0, end_mask = var_2780_end_mask_0, squeeze_mask = var_2780_squeeze_mask_0, x = k_complex_13_cast_fp16)[name = tensor("op_2780_cast_fp16")]; + tensor freqs_13_to_fp16 = const()[name = tensor("freqs_13_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(158411776)))]; + tensor var_2786_cast_fp16 = mul(x = freqs_13_to_fp16, y = ts_41_cast_fp16)[name = tensor("op_2786_cast_fp16")]; + tensor rotr_13_cast_fp16 = cos(x = var_2786_cast_fp16)[name = tensor("rotr_13_cast_fp16")]; + tensor roti_13_cast_fp16 = sin(x = var_2786_cast_fp16)[name = tensor("roti_13_cast_fp16")]; + tensor var_2790_cast_fp16 = mul(x = var_2756_cast_fp16, y = rotr_13_cast_fp16)[name = tensor("op_2790_cast_fp16")]; + tensor var_2791_cast_fp16 = mul(x = var_2764_cast_fp16, y = roti_13_cast_fp16)[name = tensor("op_2791_cast_fp16")]; + tensor qor_25_cast_fp16 = sub(x = var_2790_cast_fp16, y = var_2791_cast_fp16)[name = tensor("qor_25_cast_fp16")]; + tensor var_2794_cast_fp16 = mul(x = var_2756_cast_fp16, y = roti_13_cast_fp16)[name = tensor("op_2794_cast_fp16")]; + tensor var_2795_cast_fp16 = mul(x = var_2764_cast_fp16, y = rotr_13_cast_fp16)[name = tensor("op_2795_cast_fp16")]; + tensor qoi_25_cast_fp16 = add(x = var_2794_cast_fp16, y = var_2795_cast_fp16)[name = tensor("qoi_25_cast_fp16")]; + tensor var_2798_cast_fp16 = mul(x = var_2772_cast_fp16, y = rotr_13_cast_fp16)[name = tensor("op_2798_cast_fp16")]; + tensor var_2799_cast_fp16 = mul(x = var_2780_cast_fp16, y = roti_13_cast_fp16)[name = tensor("op_2799_cast_fp16")]; + tensor kor_25_cast_fp16 = sub(x = var_2798_cast_fp16, y = var_2799_cast_fp16)[name = tensor("kor_25_cast_fp16")]; + tensor var_2802_cast_fp16 = mul(x = var_2772_cast_fp16, y = roti_13_cast_fp16)[name = tensor("op_2802_cast_fp16")]; + tensor var_2803_cast_fp16 = mul(x = var_2780_cast_fp16, y = rotr_13_cast_fp16)[name = tensor("op_2803_cast_fp16")]; + tensor koi_25_cast_fp16 = add(x = var_2802_cast_fp16, y = var_2803_cast_fp16)[name = tensor("koi_25_cast_fp16")]; + tensor qo_13_axis_0 = const()[name = tensor("qo_13_axis_0"), val = tensor(-1)]; + tensor qo_13_cast_fp16 = stack(axis = qo_13_axis_0, values = (qor_25_cast_fp16, qoi_25_cast_fp16))[name = tensor("qo_13_cast_fp16")]; + tensor ko_13_axis_0 = const()[name = tensor("ko_13_axis_0"), val = tensor(-1)]; + tensor ko_13_cast_fp16 = stack(axis = ko_13_axis_0, values = (kor_25_cast_fp16, koi_25_cast_fp16))[name = tensor("ko_13_cast_fp16")]; + tensor var_2836 = const()[name = tensor("op_2836"), val = tensor([1, 256, 16, 64])]; + tensor q_39_cast_fp16 = reshape(shape = var_2836, x = qo_13_cast_fp16)[name = tensor("q_39_cast_fp16")]; + tensor var_2842 = const()[name = tensor("op_2842"), val = tensor([1, 256, 16, 64])]; + tensor k_27_cast_fp16 = reshape(shape = var_2842, x = ko_13_cast_fp16)[name = tensor("k_27_cast_fp16")]; + tensor write_indices_float_37_cast_fp16 = add(x = var_2731_cast_fp16, y = var_393_promoted_to_fp16)[name = tensor("write_indices_float_37_cast_fp16")]; + tensor _inversed_2863_y_0_to_fp16 = const()[name = tensor("_inversed_2863_y_0_to_fp16"), val = tensor(0x1p-9)]; + tensor _inversed_2863_cast_fp16 = mul(x = write_indices_float_37_cast_fp16, y = _inversed_2863_y_0_to_fp16)[name = tensor("_inversed_2863_cast_fp16")]; + tensor var_2864_cast_fp16 = floor(x = _inversed_2863_cast_fp16)[name = tensor("op_2864_cast_fp16")]; + tensor var_2865_to_fp16 = const()[name = tensor("op_2865_to_fp16"), val = tensor(0x1p+9)]; + tensor var_2866_cast_fp16 = mul(x = var_2864_cast_fp16, y = var_2865_to_fp16)[name = tensor("op_2866_cast_fp16")]; + tensor write_indices_float_39_cast_fp16 = sub(x = write_indices_float_37_cast_fp16, y = var_2866_cast_fp16)[name = tensor("write_indices_float_39_cast_fp16")]; + tensor var_2884_value_0_to_fp16 = const()[name = tensor("op_2884_value_0_to_fp16"), val = tensor(0x1.ffp+8)]; + tensor var_2884_cast_fp16 = fill_like(ref_tensor = write_indices_float_39_cast_fp16, value = var_2884_value_0_to_fp16)[name = tensor("op_2884_cast_fp16")]; + tensor write_indices_float_41_cast_fp16 = select(a = write_indices_float_39_cast_fp16, b = var_2884_cast_fp16, cond = token_valid_1_cast_fp16)[name = tensor("write_indices_float_41_cast_fp16")]; + tensor var_2890_dtype_0 = const()[name = tensor("op_2890_dtype_0"), val = tensor("int32")]; + tensor var_2895 = const()[name = tensor("op_2895"), val = tensor([1, 256, 1, 1])]; + tensor write_indices_float_41_cast_fp16_to_int32 = cast(dtype = var_2890_dtype_0, x = write_indices_float_41_cast_fp16)[name = tensor("cast_630")]; + tensor var_2896 = reshape(shape = var_2895, x = write_indices_float_41_cast_fp16_to_int32)[name = tensor("op_2896")]; + tensor write_indices_13_reps_0 = const()[name = tensor("write_indices_13_reps_0"), val = tensor([1, 1, 16, 64])]; + tensor write_indices_13 = tile(reps = write_indices_13_reps_0, x = var_2896)[name = tensor("write_indices_13")]; + tensor var_2906_begin_0 = const()[name = tensor("op_2906_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_2906_end_0 = const()[name = tensor("op_2906_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor var_2906_end_mask_0 = const()[name = tensor("op_2906_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_2906_squeeze_mask_0 = const()[name = tensor("op_2906_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor cache6_to_fp16_dtype_0 = const()[name = tensor("cache6_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor cache6_to_fp16 = cast(dtype = cache6_to_fp16_dtype_0, x = cache6)[name = tensor("cast_629")]; + tensor var_2906_cast_fp16 = slice_by_index(begin = var_2906_begin_0, end = var_2906_end_0, end_mask = var_2906_end_mask_0, squeeze_mask = var_2906_squeeze_mask_0, x = cache6_to_fp16)[name = tensor("op_2906_cast_fp16")]; + tensor var_2908_axis_0 = const()[name = tensor("op_2908_axis_0"), val = tensor(1)]; + tensor var_2908_mode_0 = const()[name = tensor("op_2908_mode_0"), val = tensor("update")]; + tensor var_2908_validate_indices_0 = const()[name = tensor("op_2908_validate_indices_0"), val = tensor(false)]; + tensor var_2908_cast_fp16 = scatter_along_axis(axis = var_2908_axis_0, data = var_2906_cast_fp16, indices = write_indices_13, mode = var_2908_mode_0, updates = k_27_cast_fp16, validate_indices = var_2908_validate_indices_0)[name = tensor("op_2908_cast_fp16")]; + tensor concat_50 = const()[name = tensor("concat_50"), val = tensor([0, 0, 0, 0, 0])]; + tensor concat_51 = const()[name = tensor("concat_51"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_13_internal_tensor_assign_1_stride_0 = const()[name = tensor("new_cache_13_internal_tensor_assign_1_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_13_internal_tensor_assign_1_begin_mask_0 = const()[name = tensor("new_cache_13_internal_tensor_assign_1_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_13_internal_tensor_assign_1_end_mask_0 = const()[name = tensor("new_cache_13_internal_tensor_assign_1_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_13_internal_tensor_assign_1_squeeze_mask_0 = const()[name = tensor("new_cache_13_internal_tensor_assign_1_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_58 = const()[name = tensor("shape_58"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_12 = const()[name = tensor("reduce_prod_12"), val = tensor(1048576)]; + tensor range_1d_12_start_0 = const()[name = tensor("range_1d_12_start_0"), val = tensor(0)]; + tensor range_1d_12_step_0 = const()[name = tensor("range_1d_12_step_0"), val = tensor(1)]; + tensor range_1d_12 = range_1d(end = reduce_prod_12, start = range_1d_12_start_0, step = range_1d_12_step_0)[name = tensor("range_1d_12")]; + tensor reshape_60 = reshape(shape = shape_58, x = range_1d_12)[name = tensor("reshape_60")]; + tensor slice_by_index_12 = slice_by_index(begin = concat_50, begin_mask = new_cache_13_internal_tensor_assign_1_begin_mask_0, end = concat_51, end_mask = new_cache_13_internal_tensor_assign_1_end_mask_0, squeeze_mask = new_cache_13_internal_tensor_assign_1_squeeze_mask_0, stride = new_cache_13_internal_tensor_assign_1_stride_0, x = reshape_60)[name = tensor("slice_by_index_12")]; + tensor reshape_61_shape_0 = const()[name = tensor("reshape_61_shape_0"), val = tensor([-1])]; + tensor reshape_61 = reshape(shape = reshape_61_shape_0, x = slice_by_index_12)[name = tensor("reshape_61")]; + tensor reshape_62_shape_0 = const()[name = tensor("reshape_62_shape_0"), val = tensor([-1])]; + tensor reshape_62_cast_fp16 = reshape(shape = reshape_62_shape_0, x = var_2908_cast_fp16)[name = tensor("reshape_62_cast_fp16")]; + tensor reshape_63_shape_0 = const()[name = tensor("reshape_63_shape_0"), val = tensor([-1])]; + tensor reshape_63_cast_fp16 = reshape(shape = reshape_63_shape_0, x = cache6_to_fp16)[name = tensor("reshape_63_cast_fp16")]; + tensor scatter_12_mode_0 = const()[name = tensor("scatter_12_mode_0"), val = tensor("update")]; + tensor scatter_12_axis_0 = const()[name = tensor("scatter_12_axis_0"), val = tensor(0)]; + tensor scatter_12_validate_indices_0 = const()[name = tensor("scatter_12_validate_indices_0"), val = tensor(false)]; + tensor scatter_12_cast_fp16 = scatter(axis = scatter_12_axis_0, data = reshape_63_cast_fp16, indices = reshape_61, mode = scatter_12_mode_0, updates = reshape_62_cast_fp16, validate_indices = scatter_12_validate_indices_0)[name = tensor("scatter_12_cast_fp16")]; + tensor reshape_64_cast_fp16 = reshape(shape = shape_58, x = scatter_12_cast_fp16)[name = tensor("reshape_64_cast_fp16")]; + tensor var_2916_begin_0 = const()[name = tensor("op_2916_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor var_2916_end_0 = const()[name = tensor("op_2916_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor var_2916_end_mask_0 = const()[name = tensor("op_2916_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_2916_squeeze_mask_0 = const()[name = tensor("op_2916_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor var_2916_cast_fp16 = slice_by_index(begin = var_2916_begin_0, end = var_2916_end_0, end_mask = var_2916_end_mask_0, squeeze_mask = var_2916_squeeze_mask_0, x = reshape_64_cast_fp16)[name = tensor("op_2916_cast_fp16")]; + tensor var_2918_axis_0 = const()[name = tensor("op_2918_axis_0"), val = tensor(1)]; + tensor var_2918_mode_0 = const()[name = tensor("op_2918_mode_0"), val = tensor("update")]; + tensor var_2918_validate_indices_0 = const()[name = tensor("op_2918_validate_indices_0"), val = tensor(false)]; + tensor var_2918_cast_fp16 = scatter_along_axis(axis = var_2918_axis_0, data = var_2916_cast_fp16, indices = write_indices_13, mode = var_2918_mode_0, updates = v_13_cast_fp16, validate_indices = var_2918_validate_indices_0)[name = tensor("op_2918_cast_fp16")]; + tensor concat_52 = const()[name = tensor("concat_52"), val = tensor([1, 0, 0, 0, 0])]; + tensor concat_53 = const()[name = tensor("concat_53"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_13_internal_tensor_assign_2_stride_0 = const()[name = tensor("new_cache_13_internal_tensor_assign_2_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_13_internal_tensor_assign_2_begin_mask_0 = const()[name = tensor("new_cache_13_internal_tensor_assign_2_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_13_internal_tensor_assign_2_end_mask_0 = const()[name = tensor("new_cache_13_internal_tensor_assign_2_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_13_internal_tensor_assign_2_squeeze_mask_0 = const()[name = tensor("new_cache_13_internal_tensor_assign_2_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_59 = const()[name = tensor("shape_59"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_13 = const()[name = tensor("reduce_prod_13"), val = tensor(1048576)]; + tensor range_1d_13_start_0 = const()[name = tensor("range_1d_13_start_0"), val = tensor(0)]; + tensor range_1d_13_step_0 = const()[name = tensor("range_1d_13_step_0"), val = tensor(1)]; + tensor range_1d_13 = range_1d(end = reduce_prod_13, start = range_1d_13_start_0, step = range_1d_13_step_0)[name = tensor("range_1d_13")]; + tensor reshape_65 = reshape(shape = shape_59, x = range_1d_13)[name = tensor("reshape_65")]; + tensor slice_by_index_13 = slice_by_index(begin = concat_52, begin_mask = new_cache_13_internal_tensor_assign_2_begin_mask_0, end = concat_53, end_mask = new_cache_13_internal_tensor_assign_2_end_mask_0, squeeze_mask = new_cache_13_internal_tensor_assign_2_squeeze_mask_0, stride = new_cache_13_internal_tensor_assign_2_stride_0, x = reshape_65)[name = tensor("slice_by_index_13")]; + tensor reshape_66_shape_0 = const()[name = tensor("reshape_66_shape_0"), val = tensor([-1])]; + tensor reshape_66 = reshape(shape = reshape_66_shape_0, x = slice_by_index_13)[name = tensor("reshape_66")]; + tensor reshape_67_shape_0 = const()[name = tensor("reshape_67_shape_0"), val = tensor([-1])]; + tensor reshape_67_cast_fp16 = reshape(shape = reshape_67_shape_0, x = var_2918_cast_fp16)[name = tensor("reshape_67_cast_fp16")]; + tensor reshape_68_shape_0 = const()[name = tensor("reshape_68_shape_0"), val = tensor([-1])]; + tensor reshape_68_cast_fp16 = reshape(shape = reshape_68_shape_0, x = reshape_64_cast_fp16)[name = tensor("reshape_68_cast_fp16")]; + tensor scatter_13_mode_0 = const()[name = tensor("scatter_13_mode_0"), val = tensor("update")]; + tensor scatter_13_axis_0 = const()[name = tensor("scatter_13_axis_0"), val = tensor(0)]; + tensor scatter_13_validate_indices_0 = const()[name = tensor("scatter_13_validate_indices_0"), val = tensor(false)]; + tensor scatter_13_cast_fp16 = scatter(axis = scatter_13_axis_0, data = reshape_68_cast_fp16, indices = reshape_66, mode = scatter_13_mode_0, updates = reshape_67_cast_fp16, validate_indices = scatter_13_validate_indices_0)[name = tensor("scatter_13_cast_fp16")]; + tensor reshape_69_cast_fp16 = reshape(shape = shape_59, x = scatter_13_cast_fp16)[name = tensor("reshape_69_cast_fp16")]; + tensor reshape_69_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("reshape_69_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor keys_37_begin_0 = const()[name = tensor("keys_37_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor keys_37_end_0 = const()[name = tensor("keys_37_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor keys_37_end_mask_0 = const()[name = tensor("keys_37_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor keys_37_squeeze_mask_0 = const()[name = tensor("keys_37_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor keys_37_cast_fp16 = slice_by_index(begin = keys_37_begin_0, end = keys_37_end_0, end_mask = keys_37_end_mask_0, squeeze_mask = keys_37_squeeze_mask_0, x = reshape_69_cast_fp16)[name = tensor("keys_37_cast_fp16")]; + tensor values_37_begin_0 = const()[name = tensor("values_37_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor values_37_end_0 = const()[name = tensor("values_37_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor values_37_end_mask_0 = const()[name = tensor("values_37_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor values_37_squeeze_mask_0 = const()[name = tensor("values_37_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor values_37_cast_fp16 = slice_by_index(begin = values_37_begin_0, end = values_37_end_0, end_mask = values_37_end_mask_0, squeeze_mask = values_37_squeeze_mask_0, x = reshape_69_cast_fp16)[name = tensor("values_37_cast_fp16")]; + tensor var_2930_cast_fp16 = not_equal(x = keys_37_cast_fp16, y = keys_37_cast_fp16)[name = tensor("op_2930_cast_fp16")]; + tensor keys_39_cast_fp16 = select(a = var_470_to_fp16, b = keys_37_cast_fp16, cond = var_2930_cast_fp16)[name = tensor("keys_39_cast_fp16")]; + tensor var_2938_cast_fp16 = not_equal(x = values_37_cast_fp16, y = values_37_cast_fp16)[name = tensor("op_2938_cast_fp16")]; + tensor values_39_cast_fp16 = select(a = var_470_to_fp16, b = values_37_cast_fp16, cond = var_2938_cast_fp16)[name = tensor("values_39_cast_fp16")]; + tensor var_2962 = const()[name = tensor("op_2962"), val = tensor([0, 2, 1, 3])]; + tensor var_2978 = const()[name = tensor("op_2978"), val = tensor([1, 1, 1])]; + tensor var_2979_cast_fp16 = reshape(shape = var_2978, x = position6_to_fp16)[name = tensor("op_2979_cast_fp16")]; + tensor q_positions_13_cast_fp16 = add(x = var_2979_cast_fp16, y = q_offsets_1_promoted_to_fp16)[name = tensor("q_positions_13_cast_fp16")]; + tensor valid_end_13_cast_fp16 = add(x = var_2979_cast_fp16, y = var_536_cast_fp16)[name = tensor("valid_end_13_cast_fp16")]; + tensor valid_mask_13_cast_fp16 = less(x = k_positions_1_promoted_to_fp16, y = valid_end_13_cast_fp16)[name = tensor("valid_mask_13_cast_fp16")]; + tensor causal_mask_13_cast_fp16 = less_equal(x = k_positions_1_promoted_to_fp16, y = q_positions_13_cast_fp16)[name = tensor("causal_mask_13_cast_fp16")]; + tensor var_3007 = logical_and(x = valid_mask_13_cast_fp16, y = causal_mask_13_cast_fp16)[name = tensor("op_3007")]; + tensor attn_mask_13_axes_0 = const()[name = tensor("attn_mask_13_axes_0"), val = tensor([1])]; + tensor attn_mask_13 = expand_dims(axes = attn_mask_13_axes_0, x = var_3007)[name = tensor("attn_mask_13")]; + tensor var_3021_transpose_x_0 = const()[name = tensor("op_3021_transpose_x_0"), val = tensor(false)]; + tensor var_3021_transpose_y_0 = const()[name = tensor("op_3021_transpose_y_0"), val = tensor(false)]; + tensor transpose_81_perm_0 = const()[name = tensor("transpose_81_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_82_perm_0 = const()[name = tensor("transpose_82_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_82 = transpose(perm = transpose_82_perm_0, x = keys_39_cast_fp16)[name = tensor("transpose_180")]; + tensor transpose_81 = transpose(perm = transpose_81_perm_0, x = q_39_cast_fp16)[name = tensor("transpose_181")]; + tensor var_3021_cast_fp16 = matmul(transpose_x = var_3021_transpose_x_0, transpose_y = var_3021_transpose_y_0, x = transpose_81, y = transpose_82)[name = tensor("op_3021_cast_fp16")]; + tensor var_3015_to_fp16 = const()[name = tensor("op_3015_to_fp16"), val = tensor([0x1p-3])]; + tensor attn_weights_37_cast_fp16 = mul(x = var_3021_cast_fp16, y = var_3015_to_fp16)[name = tensor("attn_weights_37_cast_fp16")]; + tensor var_3023 = logical_not(x = attn_mask_13)[name = tensor("op_3023")]; + tensor var_3024_to_fp16 = const()[name = tensor("op_3024_to_fp16"), val = tensor(-inf)]; + tensor attn_weights_39_cast_fp16 = select(a = var_3024_to_fp16, b = attn_weights_37_cast_fp16, cond = var_3023)[name = tensor("attn_weights_39_cast_fp16")]; + tensor var_3026 = const()[name = tensor("op_3026"), val = tensor(-1)]; + tensor attn_weights_41_cast_fp16 = softmax(axis = var_3026, x = attn_weights_39_cast_fp16)[name = tensor("attn_weights_41_cast_fp16")]; + tensor attn_output_13_transpose_x_0 = const()[name = tensor("attn_output_13_transpose_x_0"), val = tensor(false)]; + tensor attn_output_13_transpose_y_0 = const()[name = tensor("attn_output_13_transpose_y_0"), val = tensor(false)]; + tensor values_41_cast_fp16 = transpose(perm = var_2962, x = values_39_cast_fp16)[name = tensor("transpose_182")]; + tensor attn_output_13_cast_fp16 = matmul(transpose_x = attn_output_13_transpose_x_0, transpose_y = attn_output_13_transpose_y_0, x = attn_weights_41_cast_fp16, y = values_41_cast_fp16)[name = tensor("attn_output_13_cast_fp16")]; + tensor var_3034 = const()[name = tensor("op_3034"), val = tensor([0, 2, 1, 3])]; + tensor var_3039 = const()[name = tensor("op_3039"), val = tensor([1, 256, 1024])]; + tensor var_3035_cast_fp16 = transpose(perm = var_3034, x = attn_output_13_cast_fp16)[name = tensor("transpose_179")]; + tensor input_77_cast_fp16 = reshape(shape = var_3039, x = var_3035_cast_fp16)[name = tensor("input_77_cast_fp16")]; + tensor attn6_out_proj_weight_to_fp16 = const()[name = tensor("attn6_out_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(158411904)))]; + tensor linear_25_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = attn6_out_proj_weight_to_fp16, x = input_77_cast_fp16)[name = tensor("linear_25_cast_fp16")]; + tensor var_3048_cast_fp16 = add(x = position6_to_fp16, y = valid_len_to_fp16)[name = tensor("op_3048_cast_fp16")]; + tensor var_3048_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_3048_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor input_79_cast_fp16 = add(x = input_73_cast_fp16, y = linear_25_cast_fp16)[name = tensor("input_79_cast_fp16")]; + tensor input_81_axes_0 = const()[name = tensor("input_81_axes_0"), val = tensor([-1])]; + tensor norm6_2_weight_to_fp16 = const()[name = tensor("norm6_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(160509120)))]; + tensor norm6_2_bias_to_fp16 = const()[name = tensor("norm6_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(160511232)))]; + tensor var_3052_to_fp16 = const()[name = tensor("op_3052_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_81_cast_fp16 = layer_norm(axes = input_81_axes_0, beta = norm6_2_bias_to_fp16, epsilon = var_3052_to_fp16, gamma = norm6_2_weight_to_fp16, x = input_79_cast_fp16)[name = tensor("input_81_cast_fp16")]; + tensor linear6_1_weight_to_fp16 = const()[name = tensor("linear6_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(160513344)))]; + tensor linear_26_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = linear6_1_weight_to_fp16, x = input_81_cast_fp16)[name = tensor("linear_26_cast_fp16")]; + tensor input_83_mode_0 = const()[name = tensor("input_83_mode_0"), val = tensor("EXACT")]; + tensor input_83_cast_fp16 = gelu(mode = input_83_mode_0, x = linear_26_cast_fp16)[name = tensor("input_83_cast_fp16")]; + tensor linear6_2_weight_to_fp16 = const()[name = tensor("linear6_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(168902016)))]; + tensor linear_27_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = linear6_2_weight_to_fp16, x = input_83_cast_fp16)[name = tensor("linear_27_cast_fp16")]; + tensor input_85_cast_fp16 = add(x = input_79_cast_fp16, y = linear_27_cast_fp16)[name = tensor("input_85_cast_fp16")]; + tensor input_87_axes_0 = const()[name = tensor("input_87_axes_0"), val = tensor([-1])]; + tensor norm7_1_weight_to_fp16 = const()[name = tensor("norm7_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(177290688)))]; + tensor norm7_1_bias_to_fp16 = const()[name = tensor("norm7_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(177292800)))]; + tensor var_3069_to_fp16 = const()[name = tensor("op_3069_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_87_cast_fp16 = layer_norm(axes = input_87_axes_0, beta = norm7_1_bias_to_fp16, epsilon = var_3069_to_fp16, gamma = norm7_1_weight_to_fp16, x = input_85_cast_fp16)[name = tensor("input_87_cast_fp16")]; + tensor attn7_in_proj_weight_to_fp16 = const()[name = tensor("attn7_in_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(177294912)))]; + tensor linear_28_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = attn7_in_proj_weight_to_fp16, x = input_87_cast_fp16)[name = tensor("linear_28_cast_fp16")]; + tensor var_3083 = const()[name = tensor("op_3083"), val = tensor([1, 256, 3, 16, 64])]; + tensor qkv_15_cast_fp16 = reshape(shape = var_3083, x = linear_28_cast_fp16)[name = tensor("qkv_15_cast_fp16")]; + tensor q_43_begin_0 = const()[name = tensor("q_43_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor q_43_end_0 = const()[name = tensor("q_43_end_0"), val = tensor([1, 256, 1, 16, 64])]; + tensor q_43_end_mask_0 = const()[name = tensor("q_43_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor q_43_squeeze_mask_0 = const()[name = tensor("q_43_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor q_43_cast_fp16 = slice_by_index(begin = q_43_begin_0, end = q_43_end_0, end_mask = q_43_end_mask_0, squeeze_mask = q_43_squeeze_mask_0, x = qkv_15_cast_fp16)[name = tensor("q_43_cast_fp16")]; + tensor k_29_begin_0 = const()[name = tensor("k_29_begin_0"), val = tensor([0, 0, 1, 0, 0])]; + tensor k_29_end_0 = const()[name = tensor("k_29_end_0"), val = tensor([1, 256, 2, 16, 64])]; + tensor k_29_end_mask_0 = const()[name = tensor("k_29_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor k_29_squeeze_mask_0 = const()[name = tensor("k_29_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor k_29_cast_fp16 = slice_by_index(begin = k_29_begin_0, end = k_29_end_0, end_mask = k_29_end_mask_0, squeeze_mask = k_29_squeeze_mask_0, x = qkv_15_cast_fp16)[name = tensor("k_29_cast_fp16")]; + tensor v_15_begin_0 = const()[name = tensor("v_15_begin_0"), val = tensor([0, 0, 2, 0, 0])]; + tensor v_15_end_0 = const()[name = tensor("v_15_end_0"), val = tensor([1, 256, 3, 16, 64])]; + tensor v_15_end_mask_0 = const()[name = tensor("v_15_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor v_15_squeeze_mask_0 = const()[name = tensor("v_15_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor v_15_cast_fp16 = slice_by_index(begin = v_15_begin_0, end = v_15_end_0, end_mask = v_15_end_mask_0, squeeze_mask = v_15_squeeze_mask_0, x = qkv_15_cast_fp16)[name = tensor("v_15_cast_fp16")]; + tensor var_3141 = const()[name = tensor("op_3141"), val = tensor([1, 1])]; + tensor position7_to_fp16_dtype_0 = const()[name = tensor("position7_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor position7_to_fp16 = cast(dtype = position7_to_fp16_dtype_0, x = position7)[name = tensor("cast_626")]; + tensor var_3142_cast_fp16 = reshape(shape = var_3141, x = position7_to_fp16)[name = tensor("op_3142_cast_fp16")]; + tensor ts_45_cast_fp16 = add(x = ts_1_promoted_to_fp16, y = var_3142_cast_fp16)[name = tensor("ts_45_cast_fp16")]; + tensor var_3149 = const()[name = tensor("op_3149"), val = tensor([1, 256, 1, 1])]; + tensor ts_47_cast_fp16 = reshape(shape = var_3149, x = ts_45_cast_fp16)[name = tensor("ts_47_cast_fp16")]; + tensor var_3156 = const()[name = tensor("op_3156"), val = tensor([1, 256, 16, 32, 2])]; + tensor q_complex_15_cast_fp16 = reshape(shape = var_3156, x = q_43_cast_fp16)[name = tensor("q_complex_15_cast_fp16")]; + tensor var_3163 = const()[name = tensor("op_3163"), val = tensor([1, 256, 16, 32, 2])]; + tensor k_complex_15_cast_fp16 = reshape(shape = var_3163, x = k_29_cast_fp16)[name = tensor("k_complex_15_cast_fp16")]; + tensor var_3167_begin_0 = const()[name = tensor("op_3167_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_3167_end_0 = const()[name = tensor("op_3167_end_0"), val = tensor([1, 256, 16, 32, 1])]; + tensor var_3167_end_mask_0 = const()[name = tensor("op_3167_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_3167_squeeze_mask_0 = const()[name = tensor("op_3167_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_3167_cast_fp16 = slice_by_index(begin = var_3167_begin_0, end = var_3167_end_0, end_mask = var_3167_end_mask_0, squeeze_mask = var_3167_squeeze_mask_0, x = q_complex_15_cast_fp16)[name = tensor("op_3167_cast_fp16")]; + tensor var_3175_begin_0 = const()[name = tensor("op_3175_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_3175_end_0 = const()[name = tensor("op_3175_end_0"), val = tensor([1, 256, 16, 32, 2])]; + tensor var_3175_end_mask_0 = const()[name = tensor("op_3175_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_3175_squeeze_mask_0 = const()[name = tensor("op_3175_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_3175_cast_fp16 = slice_by_index(begin = var_3175_begin_0, end = var_3175_end_0, end_mask = var_3175_end_mask_0, squeeze_mask = var_3175_squeeze_mask_0, x = q_complex_15_cast_fp16)[name = tensor("op_3175_cast_fp16")]; + tensor var_3183_begin_0 = const()[name = tensor("op_3183_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_3183_end_0 = const()[name = tensor("op_3183_end_0"), val = tensor([1, 256, 16, 32, 1])]; + tensor var_3183_end_mask_0 = const()[name = tensor("op_3183_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_3183_squeeze_mask_0 = const()[name = tensor("op_3183_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_3183_cast_fp16 = slice_by_index(begin = var_3183_begin_0, end = var_3183_end_0, end_mask = var_3183_end_mask_0, squeeze_mask = var_3183_squeeze_mask_0, x = k_complex_15_cast_fp16)[name = tensor("op_3183_cast_fp16")]; + tensor var_3191_begin_0 = const()[name = tensor("op_3191_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_3191_end_0 = const()[name = tensor("op_3191_end_0"), val = tensor([1, 256, 16, 32, 2])]; + tensor var_3191_end_mask_0 = const()[name = tensor("op_3191_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_3191_squeeze_mask_0 = const()[name = tensor("op_3191_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_3191_cast_fp16 = slice_by_index(begin = var_3191_begin_0, end = var_3191_end_0, end_mask = var_3191_end_mask_0, squeeze_mask = var_3191_squeeze_mask_0, x = k_complex_15_cast_fp16)[name = tensor("op_3191_cast_fp16")]; + tensor freqs_15_to_fp16 = const()[name = tensor("freqs_15_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(183586432)))]; + tensor var_3197_cast_fp16 = mul(x = freqs_15_to_fp16, y = ts_47_cast_fp16)[name = tensor("op_3197_cast_fp16")]; + tensor rotr_15_cast_fp16 = cos(x = var_3197_cast_fp16)[name = tensor("rotr_15_cast_fp16")]; + tensor roti_15_cast_fp16 = sin(x = var_3197_cast_fp16)[name = tensor("roti_15_cast_fp16")]; + tensor var_3201_cast_fp16 = mul(x = var_3167_cast_fp16, y = rotr_15_cast_fp16)[name = tensor("op_3201_cast_fp16")]; + tensor var_3202_cast_fp16 = mul(x = var_3175_cast_fp16, y = roti_15_cast_fp16)[name = tensor("op_3202_cast_fp16")]; + tensor qor_29_cast_fp16 = sub(x = var_3201_cast_fp16, y = var_3202_cast_fp16)[name = tensor("qor_29_cast_fp16")]; + tensor var_3205_cast_fp16 = mul(x = var_3167_cast_fp16, y = roti_15_cast_fp16)[name = tensor("op_3205_cast_fp16")]; + tensor var_3206_cast_fp16 = mul(x = var_3175_cast_fp16, y = rotr_15_cast_fp16)[name = tensor("op_3206_cast_fp16")]; + tensor qoi_29_cast_fp16 = add(x = var_3205_cast_fp16, y = var_3206_cast_fp16)[name = tensor("qoi_29_cast_fp16")]; + tensor var_3209_cast_fp16 = mul(x = var_3183_cast_fp16, y = rotr_15_cast_fp16)[name = tensor("op_3209_cast_fp16")]; + tensor var_3210_cast_fp16 = mul(x = var_3191_cast_fp16, y = roti_15_cast_fp16)[name = tensor("op_3210_cast_fp16")]; + tensor kor_29_cast_fp16 = sub(x = var_3209_cast_fp16, y = var_3210_cast_fp16)[name = tensor("kor_29_cast_fp16")]; + tensor var_3213_cast_fp16 = mul(x = var_3183_cast_fp16, y = roti_15_cast_fp16)[name = tensor("op_3213_cast_fp16")]; + tensor var_3214_cast_fp16 = mul(x = var_3191_cast_fp16, y = rotr_15_cast_fp16)[name = tensor("op_3214_cast_fp16")]; + tensor koi_29_cast_fp16 = add(x = var_3213_cast_fp16, y = var_3214_cast_fp16)[name = tensor("koi_29_cast_fp16")]; + tensor qo_15_axis_0 = const()[name = tensor("qo_15_axis_0"), val = tensor(-1)]; + tensor qo_15_cast_fp16 = stack(axis = qo_15_axis_0, values = (qor_29_cast_fp16, qoi_29_cast_fp16))[name = tensor("qo_15_cast_fp16")]; + tensor ko_15_axis_0 = const()[name = tensor("ko_15_axis_0"), val = tensor(-1)]; + tensor ko_15_cast_fp16 = stack(axis = ko_15_axis_0, values = (kor_29_cast_fp16, koi_29_cast_fp16))[name = tensor("ko_15_cast_fp16")]; + tensor var_3247 = const()[name = tensor("op_3247"), val = tensor([1, 256, 16, 64])]; + tensor q_45_cast_fp16 = reshape(shape = var_3247, x = qo_15_cast_fp16)[name = tensor("q_45_cast_fp16")]; + tensor var_3253 = const()[name = tensor("op_3253"), val = tensor([1, 256, 16, 64])]; + tensor k_31_cast_fp16 = reshape(shape = var_3253, x = ko_15_cast_fp16)[name = tensor("k_31_cast_fp16")]; + tensor write_indices_float_43_cast_fp16 = add(x = var_3142_cast_fp16, y = var_393_promoted_to_fp16)[name = tensor("write_indices_float_43_cast_fp16")]; + tensor _inversed_3274_y_0_to_fp16 = const()[name = tensor("_inversed_3274_y_0_to_fp16"), val = tensor(0x1p-9)]; + tensor _inversed_3274_cast_fp16 = mul(x = write_indices_float_43_cast_fp16, y = _inversed_3274_y_0_to_fp16)[name = tensor("_inversed_3274_cast_fp16")]; + tensor var_3275_cast_fp16 = floor(x = _inversed_3274_cast_fp16)[name = tensor("op_3275_cast_fp16")]; + tensor var_3276_to_fp16 = const()[name = tensor("op_3276_to_fp16"), val = tensor(0x1p+9)]; + tensor var_3277_cast_fp16 = mul(x = var_3275_cast_fp16, y = var_3276_to_fp16)[name = tensor("op_3277_cast_fp16")]; + tensor write_indices_float_45_cast_fp16 = sub(x = write_indices_float_43_cast_fp16, y = var_3277_cast_fp16)[name = tensor("write_indices_float_45_cast_fp16")]; + tensor var_3295_value_0_to_fp16 = const()[name = tensor("op_3295_value_0_to_fp16"), val = tensor(0x1.ffp+8)]; + tensor var_3295_cast_fp16 = fill_like(ref_tensor = write_indices_float_45_cast_fp16, value = var_3295_value_0_to_fp16)[name = tensor("op_3295_cast_fp16")]; + tensor write_indices_float_47_cast_fp16 = select(a = write_indices_float_45_cast_fp16, b = var_3295_cast_fp16, cond = token_valid_1_cast_fp16)[name = tensor("write_indices_float_47_cast_fp16")]; + tensor var_3301_dtype_0 = const()[name = tensor("op_3301_dtype_0"), val = tensor("int32")]; + tensor var_3306 = const()[name = tensor("op_3306"), val = tensor([1, 256, 1, 1])]; + tensor write_indices_float_47_cast_fp16_to_int32 = cast(dtype = var_3301_dtype_0, x = write_indices_float_47_cast_fp16)[name = tensor("cast_625")]; + tensor var_3307 = reshape(shape = var_3306, x = write_indices_float_47_cast_fp16_to_int32)[name = tensor("op_3307")]; + tensor write_indices_15_reps_0 = const()[name = tensor("write_indices_15_reps_0"), val = tensor([1, 1, 16, 64])]; + tensor write_indices_15 = tile(reps = write_indices_15_reps_0, x = var_3307)[name = tensor("write_indices_15")]; + tensor var_3317_begin_0 = const()[name = tensor("op_3317_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_3317_end_0 = const()[name = tensor("op_3317_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor var_3317_end_mask_0 = const()[name = tensor("op_3317_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_3317_squeeze_mask_0 = const()[name = tensor("op_3317_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor cache7_to_fp16_dtype_0 = const()[name = tensor("cache7_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor cache7_to_fp16 = cast(dtype = cache7_to_fp16_dtype_0, x = cache7)[name = tensor("cast_624")]; + tensor var_3317_cast_fp16 = slice_by_index(begin = var_3317_begin_0, end = var_3317_end_0, end_mask = var_3317_end_mask_0, squeeze_mask = var_3317_squeeze_mask_0, x = cache7_to_fp16)[name = tensor("op_3317_cast_fp16")]; + tensor var_3319_axis_0 = const()[name = tensor("op_3319_axis_0"), val = tensor(1)]; + tensor var_3319_mode_0 = const()[name = tensor("op_3319_mode_0"), val = tensor("update")]; + tensor var_3319_validate_indices_0 = const()[name = tensor("op_3319_validate_indices_0"), val = tensor(false)]; + tensor var_3319_cast_fp16 = scatter_along_axis(axis = var_3319_axis_0, data = var_3317_cast_fp16, indices = write_indices_15, mode = var_3319_mode_0, updates = k_31_cast_fp16, validate_indices = var_3319_validate_indices_0)[name = tensor("op_3319_cast_fp16")]; + tensor concat_58 = const()[name = tensor("concat_58"), val = tensor([0, 0, 0, 0, 0])]; + tensor concat_59 = const()[name = tensor("concat_59"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_15_internal_tensor_assign_1_stride_0 = const()[name = tensor("new_cache_15_internal_tensor_assign_1_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_15_internal_tensor_assign_1_begin_mask_0 = const()[name = tensor("new_cache_15_internal_tensor_assign_1_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_15_internal_tensor_assign_1_end_mask_0 = const()[name = tensor("new_cache_15_internal_tensor_assign_1_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_15_internal_tensor_assign_1_squeeze_mask_0 = const()[name = tensor("new_cache_15_internal_tensor_assign_1_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_60 = const()[name = tensor("shape_60"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_14 = const()[name = tensor("reduce_prod_14"), val = tensor(1048576)]; + tensor range_1d_14_start_0 = const()[name = tensor("range_1d_14_start_0"), val = tensor(0)]; + tensor range_1d_14_step_0 = const()[name = tensor("range_1d_14_step_0"), val = tensor(1)]; + tensor range_1d_14 = range_1d(end = reduce_prod_14, start = range_1d_14_start_0, step = range_1d_14_step_0)[name = tensor("range_1d_14")]; + tensor reshape_70 = reshape(shape = shape_60, x = range_1d_14)[name = tensor("reshape_70")]; + tensor slice_by_index_14 = slice_by_index(begin = concat_58, begin_mask = new_cache_15_internal_tensor_assign_1_begin_mask_0, end = concat_59, end_mask = new_cache_15_internal_tensor_assign_1_end_mask_0, squeeze_mask = new_cache_15_internal_tensor_assign_1_squeeze_mask_0, stride = new_cache_15_internal_tensor_assign_1_stride_0, x = reshape_70)[name = tensor("slice_by_index_14")]; + tensor reshape_71_shape_0 = const()[name = tensor("reshape_71_shape_0"), val = tensor([-1])]; + tensor reshape_71 = reshape(shape = reshape_71_shape_0, x = slice_by_index_14)[name = tensor("reshape_71")]; + tensor reshape_72_shape_0 = const()[name = tensor("reshape_72_shape_0"), val = tensor([-1])]; + tensor reshape_72_cast_fp16 = reshape(shape = reshape_72_shape_0, x = var_3319_cast_fp16)[name = tensor("reshape_72_cast_fp16")]; + tensor reshape_73_shape_0 = const()[name = tensor("reshape_73_shape_0"), val = tensor([-1])]; + tensor reshape_73_cast_fp16 = reshape(shape = reshape_73_shape_0, x = cache7_to_fp16)[name = tensor("reshape_73_cast_fp16")]; + tensor scatter_14_mode_0 = const()[name = tensor("scatter_14_mode_0"), val = tensor("update")]; + tensor scatter_14_axis_0 = const()[name = tensor("scatter_14_axis_0"), val = tensor(0)]; + tensor scatter_14_validate_indices_0 = const()[name = tensor("scatter_14_validate_indices_0"), val = tensor(false)]; + tensor scatter_14_cast_fp16 = scatter(axis = scatter_14_axis_0, data = reshape_73_cast_fp16, indices = reshape_71, mode = scatter_14_mode_0, updates = reshape_72_cast_fp16, validate_indices = scatter_14_validate_indices_0)[name = tensor("scatter_14_cast_fp16")]; + tensor reshape_74_cast_fp16 = reshape(shape = shape_60, x = scatter_14_cast_fp16)[name = tensor("reshape_74_cast_fp16")]; + tensor var_3327_begin_0 = const()[name = tensor("op_3327_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor var_3327_end_0 = const()[name = tensor("op_3327_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor var_3327_end_mask_0 = const()[name = tensor("op_3327_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_3327_squeeze_mask_0 = const()[name = tensor("op_3327_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor var_3327_cast_fp16 = slice_by_index(begin = var_3327_begin_0, end = var_3327_end_0, end_mask = var_3327_end_mask_0, squeeze_mask = var_3327_squeeze_mask_0, x = reshape_74_cast_fp16)[name = tensor("op_3327_cast_fp16")]; + tensor var_3329_axis_0 = const()[name = tensor("op_3329_axis_0"), val = tensor(1)]; + tensor var_3329_mode_0 = const()[name = tensor("op_3329_mode_0"), val = tensor("update")]; + tensor var_3329_validate_indices_0 = const()[name = tensor("op_3329_validate_indices_0"), val = tensor(false)]; + tensor var_3329_cast_fp16 = scatter_along_axis(axis = var_3329_axis_0, data = var_3327_cast_fp16, indices = write_indices_15, mode = var_3329_mode_0, updates = v_15_cast_fp16, validate_indices = var_3329_validate_indices_0)[name = tensor("op_3329_cast_fp16")]; + tensor concat_60 = const()[name = tensor("concat_60"), val = tensor([1, 0, 0, 0, 0])]; + tensor concat_61 = const()[name = tensor("concat_61"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_15_internal_tensor_assign_2_stride_0 = const()[name = tensor("new_cache_15_internal_tensor_assign_2_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_15_internal_tensor_assign_2_begin_mask_0 = const()[name = tensor("new_cache_15_internal_tensor_assign_2_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_15_internal_tensor_assign_2_end_mask_0 = const()[name = tensor("new_cache_15_internal_tensor_assign_2_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_15_internal_tensor_assign_2_squeeze_mask_0 = const()[name = tensor("new_cache_15_internal_tensor_assign_2_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_61 = const()[name = tensor("shape_61"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_15 = const()[name = tensor("reduce_prod_15"), val = tensor(1048576)]; + tensor range_1d_15_start_0 = const()[name = tensor("range_1d_15_start_0"), val = tensor(0)]; + tensor range_1d_15_step_0 = const()[name = tensor("range_1d_15_step_0"), val = tensor(1)]; + tensor range_1d_15 = range_1d(end = reduce_prod_15, start = range_1d_15_start_0, step = range_1d_15_step_0)[name = tensor("range_1d_15")]; + tensor reshape_75 = reshape(shape = shape_61, x = range_1d_15)[name = tensor("reshape_75")]; + tensor slice_by_index_15 = slice_by_index(begin = concat_60, begin_mask = new_cache_15_internal_tensor_assign_2_begin_mask_0, end = concat_61, end_mask = new_cache_15_internal_tensor_assign_2_end_mask_0, squeeze_mask = new_cache_15_internal_tensor_assign_2_squeeze_mask_0, stride = new_cache_15_internal_tensor_assign_2_stride_0, x = reshape_75)[name = tensor("slice_by_index_15")]; + tensor reshape_76_shape_0 = const()[name = tensor("reshape_76_shape_0"), val = tensor([-1])]; + tensor reshape_76 = reshape(shape = reshape_76_shape_0, x = slice_by_index_15)[name = tensor("reshape_76")]; + tensor reshape_77_shape_0 = const()[name = tensor("reshape_77_shape_0"), val = tensor([-1])]; + tensor reshape_77_cast_fp16 = reshape(shape = reshape_77_shape_0, x = var_3329_cast_fp16)[name = tensor("reshape_77_cast_fp16")]; + tensor reshape_78_shape_0 = const()[name = tensor("reshape_78_shape_0"), val = tensor([-1])]; + tensor reshape_78_cast_fp16 = reshape(shape = reshape_78_shape_0, x = reshape_74_cast_fp16)[name = tensor("reshape_78_cast_fp16")]; + tensor scatter_15_mode_0 = const()[name = tensor("scatter_15_mode_0"), val = tensor("update")]; + tensor scatter_15_axis_0 = const()[name = tensor("scatter_15_axis_0"), val = tensor(0)]; + tensor scatter_15_validate_indices_0 = const()[name = tensor("scatter_15_validate_indices_0"), val = tensor(false)]; + tensor scatter_15_cast_fp16 = scatter(axis = scatter_15_axis_0, data = reshape_78_cast_fp16, indices = reshape_76, mode = scatter_15_mode_0, updates = reshape_77_cast_fp16, validate_indices = scatter_15_validate_indices_0)[name = tensor("scatter_15_cast_fp16")]; + tensor reshape_79_cast_fp16 = reshape(shape = shape_61, x = scatter_15_cast_fp16)[name = tensor("reshape_79_cast_fp16")]; + tensor reshape_79_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("reshape_79_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor keys_43_begin_0 = const()[name = tensor("keys_43_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor keys_43_end_0 = const()[name = tensor("keys_43_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor keys_43_end_mask_0 = const()[name = tensor("keys_43_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor keys_43_squeeze_mask_0 = const()[name = tensor("keys_43_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor keys_43_cast_fp16 = slice_by_index(begin = keys_43_begin_0, end = keys_43_end_0, end_mask = keys_43_end_mask_0, squeeze_mask = keys_43_squeeze_mask_0, x = reshape_79_cast_fp16)[name = tensor("keys_43_cast_fp16")]; + tensor values_43_begin_0 = const()[name = tensor("values_43_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor values_43_end_0 = const()[name = tensor("values_43_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor values_43_end_mask_0 = const()[name = tensor("values_43_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor values_43_squeeze_mask_0 = const()[name = tensor("values_43_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor values_43_cast_fp16 = slice_by_index(begin = values_43_begin_0, end = values_43_end_0, end_mask = values_43_end_mask_0, squeeze_mask = values_43_squeeze_mask_0, x = reshape_79_cast_fp16)[name = tensor("values_43_cast_fp16")]; + tensor var_3341_cast_fp16 = not_equal(x = keys_43_cast_fp16, y = keys_43_cast_fp16)[name = tensor("op_3341_cast_fp16")]; + tensor keys_45_cast_fp16 = select(a = var_470_to_fp16, b = keys_43_cast_fp16, cond = var_3341_cast_fp16)[name = tensor("keys_45_cast_fp16")]; + tensor var_3349_cast_fp16 = not_equal(x = values_43_cast_fp16, y = values_43_cast_fp16)[name = tensor("op_3349_cast_fp16")]; + tensor values_45_cast_fp16 = select(a = var_470_to_fp16, b = values_43_cast_fp16, cond = var_3349_cast_fp16)[name = tensor("values_45_cast_fp16")]; + tensor var_3373 = const()[name = tensor("op_3373"), val = tensor([0, 2, 1, 3])]; + tensor var_3389 = const()[name = tensor("op_3389"), val = tensor([1, 1, 1])]; + tensor var_3390_cast_fp16 = reshape(shape = var_3389, x = position7_to_fp16)[name = tensor("op_3390_cast_fp16")]; + tensor q_positions_15_cast_fp16 = add(x = var_3390_cast_fp16, y = q_offsets_1_promoted_to_fp16)[name = tensor("q_positions_15_cast_fp16")]; + tensor valid_end_15_cast_fp16 = add(x = var_3390_cast_fp16, y = var_536_cast_fp16)[name = tensor("valid_end_15_cast_fp16")]; + tensor valid_mask_15_cast_fp16 = less(x = k_positions_1_promoted_to_fp16, y = valid_end_15_cast_fp16)[name = tensor("valid_mask_15_cast_fp16")]; + tensor causal_mask_15_cast_fp16 = less_equal(x = k_positions_1_promoted_to_fp16, y = q_positions_15_cast_fp16)[name = tensor("causal_mask_15_cast_fp16")]; + tensor var_3418 = logical_and(x = valid_mask_15_cast_fp16, y = causal_mask_15_cast_fp16)[name = tensor("op_3418")]; + tensor attn_mask_15_axes_0 = const()[name = tensor("attn_mask_15_axes_0"), val = tensor([1])]; + tensor attn_mask_15 = expand_dims(axes = attn_mask_15_axes_0, x = var_3418)[name = tensor("attn_mask_15")]; + tensor var_3432_transpose_x_0 = const()[name = tensor("op_3432_transpose_x_0"), val = tensor(false)]; + tensor var_3432_transpose_y_0 = const()[name = tensor("op_3432_transpose_y_0"), val = tensor(false)]; + tensor transpose_83_perm_0 = const()[name = tensor("transpose_83_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_84_perm_0 = const()[name = tensor("transpose_84_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_84 = transpose(perm = transpose_84_perm_0, x = keys_45_cast_fp16)[name = tensor("transpose_176")]; + tensor transpose_83 = transpose(perm = transpose_83_perm_0, x = q_45_cast_fp16)[name = tensor("transpose_177")]; + tensor var_3432_cast_fp16 = matmul(transpose_x = var_3432_transpose_x_0, transpose_y = var_3432_transpose_y_0, x = transpose_83, y = transpose_84)[name = tensor("op_3432_cast_fp16")]; + tensor var_3426_to_fp16 = const()[name = tensor("op_3426_to_fp16"), val = tensor([0x1p-3])]; + tensor attn_weights_43_cast_fp16 = mul(x = var_3432_cast_fp16, y = var_3426_to_fp16)[name = tensor("attn_weights_43_cast_fp16")]; + tensor var_3434 = logical_not(x = attn_mask_15)[name = tensor("op_3434")]; + tensor var_3435_to_fp16 = const()[name = tensor("op_3435_to_fp16"), val = tensor(-inf)]; + tensor attn_weights_45_cast_fp16 = select(a = var_3435_to_fp16, b = attn_weights_43_cast_fp16, cond = var_3434)[name = tensor("attn_weights_45_cast_fp16")]; + tensor var_3437 = const()[name = tensor("op_3437"), val = tensor(-1)]; + tensor attn_weights_47_cast_fp16 = softmax(axis = var_3437, x = attn_weights_45_cast_fp16)[name = tensor("attn_weights_47_cast_fp16")]; + tensor attn_output_15_transpose_x_0 = const()[name = tensor("attn_output_15_transpose_x_0"), val = tensor(false)]; + tensor attn_output_15_transpose_y_0 = const()[name = tensor("attn_output_15_transpose_y_0"), val = tensor(false)]; + tensor values_47_cast_fp16 = transpose(perm = var_3373, x = values_45_cast_fp16)[name = tensor("transpose_178")]; + tensor attn_output_15_cast_fp16 = matmul(transpose_x = attn_output_15_transpose_x_0, transpose_y = attn_output_15_transpose_y_0, x = attn_weights_47_cast_fp16, y = values_47_cast_fp16)[name = tensor("attn_output_15_cast_fp16")]; + tensor var_3445 = const()[name = tensor("op_3445"), val = tensor([0, 2, 1, 3])]; + tensor var_3450 = const()[name = tensor("op_3450"), val = tensor([1, 256, 1024])]; + tensor var_3446_cast_fp16 = transpose(perm = var_3445, x = attn_output_15_cast_fp16)[name = tensor("transpose_175")]; + tensor input_89_cast_fp16 = reshape(shape = var_3450, x = var_3446_cast_fp16)[name = tensor("input_89_cast_fp16")]; + tensor attn7_out_proj_weight_to_fp16 = const()[name = tensor("attn7_out_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(183586560)))]; + tensor linear_29_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = attn7_out_proj_weight_to_fp16, x = input_89_cast_fp16)[name = tensor("linear_29_cast_fp16")]; + tensor var_3459_cast_fp16 = add(x = position7_to_fp16, y = valid_len_to_fp16)[name = tensor("op_3459_cast_fp16")]; + tensor var_3459_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_3459_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor input_91_cast_fp16 = add(x = input_85_cast_fp16, y = linear_29_cast_fp16)[name = tensor("input_91_cast_fp16")]; + tensor input_93_axes_0 = const()[name = tensor("input_93_axes_0"), val = tensor([-1])]; + tensor norm7_2_weight_to_fp16 = const()[name = tensor("norm7_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(185683776)))]; + tensor norm7_2_bias_to_fp16 = const()[name = tensor("norm7_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(185685888)))]; + tensor var_3463_to_fp16 = const()[name = tensor("op_3463_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_93_cast_fp16 = layer_norm(axes = input_93_axes_0, beta = norm7_2_bias_to_fp16, epsilon = var_3463_to_fp16, gamma = norm7_2_weight_to_fp16, x = input_91_cast_fp16)[name = tensor("input_93_cast_fp16")]; + tensor linear7_1_weight_to_fp16 = const()[name = tensor("linear7_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(185688000)))]; + tensor linear_30_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = linear7_1_weight_to_fp16, x = input_93_cast_fp16)[name = tensor("linear_30_cast_fp16")]; + tensor input_95_mode_0 = const()[name = tensor("input_95_mode_0"), val = tensor("EXACT")]; + tensor input_95_cast_fp16 = gelu(mode = input_95_mode_0, x = linear_30_cast_fp16)[name = tensor("input_95_cast_fp16")]; + tensor linear7_2_weight_to_fp16 = const()[name = tensor("linear7_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(194076672)))]; + tensor linear_31_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = linear7_2_weight_to_fp16, x = input_95_cast_fp16)[name = tensor("linear_31_cast_fp16")]; + tensor input_97_cast_fp16 = add(x = input_91_cast_fp16, y = linear_31_cast_fp16)[name = tensor("input_97_cast_fp16")]; + tensor input_99_axes_0 = const()[name = tensor("input_99_axes_0"), val = tensor([-1])]; + tensor norm8_1_weight_to_fp16 = const()[name = tensor("norm8_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(202465344)))]; + tensor norm8_1_bias_to_fp16 = const()[name = tensor("norm8_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(202467456)))]; + tensor var_3480_to_fp16 = const()[name = tensor("op_3480_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_99_cast_fp16 = layer_norm(axes = input_99_axes_0, beta = norm8_1_bias_to_fp16, epsilon = var_3480_to_fp16, gamma = norm8_1_weight_to_fp16, x = input_97_cast_fp16)[name = tensor("input_99_cast_fp16")]; + tensor attn8_in_proj_weight_to_fp16 = const()[name = tensor("attn8_in_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(202469568)))]; + tensor linear_32_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = attn8_in_proj_weight_to_fp16, x = input_99_cast_fp16)[name = tensor("linear_32_cast_fp16")]; + tensor var_3494 = const()[name = tensor("op_3494"), val = tensor([1, 256, 3, 16, 64])]; + tensor qkv_17_cast_fp16 = reshape(shape = var_3494, x = linear_32_cast_fp16)[name = tensor("qkv_17_cast_fp16")]; + tensor q_49_begin_0 = const()[name = tensor("q_49_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor q_49_end_0 = const()[name = tensor("q_49_end_0"), val = tensor([1, 256, 1, 16, 64])]; + tensor q_49_end_mask_0 = const()[name = tensor("q_49_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor q_49_squeeze_mask_0 = const()[name = tensor("q_49_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor q_49_cast_fp16 = slice_by_index(begin = q_49_begin_0, end = q_49_end_0, end_mask = q_49_end_mask_0, squeeze_mask = q_49_squeeze_mask_0, x = qkv_17_cast_fp16)[name = tensor("q_49_cast_fp16")]; + tensor k_33_begin_0 = const()[name = tensor("k_33_begin_0"), val = tensor([0, 0, 1, 0, 0])]; + tensor k_33_end_0 = const()[name = tensor("k_33_end_0"), val = tensor([1, 256, 2, 16, 64])]; + tensor k_33_end_mask_0 = const()[name = tensor("k_33_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor k_33_squeeze_mask_0 = const()[name = tensor("k_33_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor k_33_cast_fp16 = slice_by_index(begin = k_33_begin_0, end = k_33_end_0, end_mask = k_33_end_mask_0, squeeze_mask = k_33_squeeze_mask_0, x = qkv_17_cast_fp16)[name = tensor("k_33_cast_fp16")]; + tensor v_17_begin_0 = const()[name = tensor("v_17_begin_0"), val = tensor([0, 0, 2, 0, 0])]; + tensor v_17_end_0 = const()[name = tensor("v_17_end_0"), val = tensor([1, 256, 3, 16, 64])]; + tensor v_17_end_mask_0 = const()[name = tensor("v_17_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor v_17_squeeze_mask_0 = const()[name = tensor("v_17_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor v_17_cast_fp16 = slice_by_index(begin = v_17_begin_0, end = v_17_end_0, end_mask = v_17_end_mask_0, squeeze_mask = v_17_squeeze_mask_0, x = qkv_17_cast_fp16)[name = tensor("v_17_cast_fp16")]; + tensor var_3552 = const()[name = tensor("op_3552"), val = tensor([1, 1])]; + tensor position8_to_fp16_dtype_0 = const()[name = tensor("position8_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor position8_to_fp16 = cast(dtype = position8_to_fp16_dtype_0, x = position8)[name = tensor("cast_621")]; + tensor var_3553_cast_fp16 = reshape(shape = var_3552, x = position8_to_fp16)[name = tensor("op_3553_cast_fp16")]; + tensor ts_51_cast_fp16 = add(x = ts_1_promoted_to_fp16, y = var_3553_cast_fp16)[name = tensor("ts_51_cast_fp16")]; + tensor var_3560 = const()[name = tensor("op_3560"), val = tensor([1, 256, 1, 1])]; + tensor ts_53_cast_fp16 = reshape(shape = var_3560, x = ts_51_cast_fp16)[name = tensor("ts_53_cast_fp16")]; + tensor var_3567 = const()[name = tensor("op_3567"), val = tensor([1, 256, 16, 32, 2])]; + tensor q_complex_17_cast_fp16 = reshape(shape = var_3567, x = q_49_cast_fp16)[name = tensor("q_complex_17_cast_fp16")]; + tensor var_3574 = const()[name = tensor("op_3574"), val = tensor([1, 256, 16, 32, 2])]; + tensor k_complex_17_cast_fp16 = reshape(shape = var_3574, x = k_33_cast_fp16)[name = tensor("k_complex_17_cast_fp16")]; + tensor var_3578_begin_0 = const()[name = tensor("op_3578_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_3578_end_0 = const()[name = tensor("op_3578_end_0"), val = tensor([1, 256, 16, 32, 1])]; + tensor var_3578_end_mask_0 = const()[name = tensor("op_3578_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_3578_squeeze_mask_0 = const()[name = tensor("op_3578_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_3578_cast_fp16 = slice_by_index(begin = var_3578_begin_0, end = var_3578_end_0, end_mask = var_3578_end_mask_0, squeeze_mask = var_3578_squeeze_mask_0, x = q_complex_17_cast_fp16)[name = tensor("op_3578_cast_fp16")]; + tensor var_3586_begin_0 = const()[name = tensor("op_3586_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_3586_end_0 = const()[name = tensor("op_3586_end_0"), val = tensor([1, 256, 16, 32, 2])]; + tensor var_3586_end_mask_0 = const()[name = tensor("op_3586_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_3586_squeeze_mask_0 = const()[name = tensor("op_3586_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_3586_cast_fp16 = slice_by_index(begin = var_3586_begin_0, end = var_3586_end_0, end_mask = var_3586_end_mask_0, squeeze_mask = var_3586_squeeze_mask_0, x = q_complex_17_cast_fp16)[name = tensor("op_3586_cast_fp16")]; + tensor var_3594_begin_0 = const()[name = tensor("op_3594_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_3594_end_0 = const()[name = tensor("op_3594_end_0"), val = tensor([1, 256, 16, 32, 1])]; + tensor var_3594_end_mask_0 = const()[name = tensor("op_3594_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_3594_squeeze_mask_0 = const()[name = tensor("op_3594_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_3594_cast_fp16 = slice_by_index(begin = var_3594_begin_0, end = var_3594_end_0, end_mask = var_3594_end_mask_0, squeeze_mask = var_3594_squeeze_mask_0, x = k_complex_17_cast_fp16)[name = tensor("op_3594_cast_fp16")]; + tensor var_3602_begin_0 = const()[name = tensor("op_3602_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_3602_end_0 = const()[name = tensor("op_3602_end_0"), val = tensor([1, 256, 16, 32, 2])]; + tensor var_3602_end_mask_0 = const()[name = tensor("op_3602_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_3602_squeeze_mask_0 = const()[name = tensor("op_3602_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_3602_cast_fp16 = slice_by_index(begin = var_3602_begin_0, end = var_3602_end_0, end_mask = var_3602_end_mask_0, squeeze_mask = var_3602_squeeze_mask_0, x = k_complex_17_cast_fp16)[name = tensor("op_3602_cast_fp16")]; + tensor freqs_17_to_fp16 = const()[name = tensor("freqs_17_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(208761088)))]; + tensor var_3608_cast_fp16 = mul(x = freqs_17_to_fp16, y = ts_53_cast_fp16)[name = tensor("op_3608_cast_fp16")]; + tensor rotr_17_cast_fp16 = cos(x = var_3608_cast_fp16)[name = tensor("rotr_17_cast_fp16")]; + tensor roti_17_cast_fp16 = sin(x = var_3608_cast_fp16)[name = tensor("roti_17_cast_fp16")]; + tensor var_3612_cast_fp16 = mul(x = var_3578_cast_fp16, y = rotr_17_cast_fp16)[name = tensor("op_3612_cast_fp16")]; + tensor var_3613_cast_fp16 = mul(x = var_3586_cast_fp16, y = roti_17_cast_fp16)[name = tensor("op_3613_cast_fp16")]; + tensor qor_33_cast_fp16 = sub(x = var_3612_cast_fp16, y = var_3613_cast_fp16)[name = tensor("qor_33_cast_fp16")]; + tensor var_3616_cast_fp16 = mul(x = var_3578_cast_fp16, y = roti_17_cast_fp16)[name = tensor("op_3616_cast_fp16")]; + tensor var_3617_cast_fp16 = mul(x = var_3586_cast_fp16, y = rotr_17_cast_fp16)[name = tensor("op_3617_cast_fp16")]; + tensor qoi_33_cast_fp16 = add(x = var_3616_cast_fp16, y = var_3617_cast_fp16)[name = tensor("qoi_33_cast_fp16")]; + tensor var_3620_cast_fp16 = mul(x = var_3594_cast_fp16, y = rotr_17_cast_fp16)[name = tensor("op_3620_cast_fp16")]; + tensor var_3621_cast_fp16 = mul(x = var_3602_cast_fp16, y = roti_17_cast_fp16)[name = tensor("op_3621_cast_fp16")]; + tensor kor_33_cast_fp16 = sub(x = var_3620_cast_fp16, y = var_3621_cast_fp16)[name = tensor("kor_33_cast_fp16")]; + tensor var_3624_cast_fp16 = mul(x = var_3594_cast_fp16, y = roti_17_cast_fp16)[name = tensor("op_3624_cast_fp16")]; + tensor var_3625_cast_fp16 = mul(x = var_3602_cast_fp16, y = rotr_17_cast_fp16)[name = tensor("op_3625_cast_fp16")]; + tensor koi_33_cast_fp16 = add(x = var_3624_cast_fp16, y = var_3625_cast_fp16)[name = tensor("koi_33_cast_fp16")]; + tensor qo_17_axis_0 = const()[name = tensor("qo_17_axis_0"), val = tensor(-1)]; + tensor qo_17_cast_fp16 = stack(axis = qo_17_axis_0, values = (qor_33_cast_fp16, qoi_33_cast_fp16))[name = tensor("qo_17_cast_fp16")]; + tensor ko_17_axis_0 = const()[name = tensor("ko_17_axis_0"), val = tensor(-1)]; + tensor ko_17_cast_fp16 = stack(axis = ko_17_axis_0, values = (kor_33_cast_fp16, koi_33_cast_fp16))[name = tensor("ko_17_cast_fp16")]; + tensor var_3658 = const()[name = tensor("op_3658"), val = tensor([1, 256, 16, 64])]; + tensor q_51_cast_fp16 = reshape(shape = var_3658, x = qo_17_cast_fp16)[name = tensor("q_51_cast_fp16")]; + tensor var_3664 = const()[name = tensor("op_3664"), val = tensor([1, 256, 16, 64])]; + tensor k_35_cast_fp16 = reshape(shape = var_3664, x = ko_17_cast_fp16)[name = tensor("k_35_cast_fp16")]; + tensor write_indices_float_49_cast_fp16 = add(x = var_3553_cast_fp16, y = var_393_promoted_to_fp16)[name = tensor("write_indices_float_49_cast_fp16")]; + tensor _inversed_3685_y_0_to_fp16 = const()[name = tensor("_inversed_3685_y_0_to_fp16"), val = tensor(0x1p-9)]; + tensor _inversed_3685_cast_fp16 = mul(x = write_indices_float_49_cast_fp16, y = _inversed_3685_y_0_to_fp16)[name = tensor("_inversed_3685_cast_fp16")]; + tensor var_3686_cast_fp16 = floor(x = _inversed_3685_cast_fp16)[name = tensor("op_3686_cast_fp16")]; + tensor var_3687_to_fp16 = const()[name = tensor("op_3687_to_fp16"), val = tensor(0x1p+9)]; + tensor var_3688_cast_fp16 = mul(x = var_3686_cast_fp16, y = var_3687_to_fp16)[name = tensor("op_3688_cast_fp16")]; + tensor write_indices_float_51_cast_fp16 = sub(x = write_indices_float_49_cast_fp16, y = var_3688_cast_fp16)[name = tensor("write_indices_float_51_cast_fp16")]; + tensor var_3706_value_0_to_fp16 = const()[name = tensor("op_3706_value_0_to_fp16"), val = tensor(0x1.ffp+8)]; + tensor var_3706_cast_fp16 = fill_like(ref_tensor = write_indices_float_51_cast_fp16, value = var_3706_value_0_to_fp16)[name = tensor("op_3706_cast_fp16")]; + tensor write_indices_float_53_cast_fp16 = select(a = write_indices_float_51_cast_fp16, b = var_3706_cast_fp16, cond = token_valid_1_cast_fp16)[name = tensor("write_indices_float_53_cast_fp16")]; + tensor var_3712_dtype_0 = const()[name = tensor("op_3712_dtype_0"), val = tensor("int32")]; + tensor var_3717 = const()[name = tensor("op_3717"), val = tensor([1, 256, 1, 1])]; + tensor write_indices_float_53_cast_fp16_to_int32 = cast(dtype = var_3712_dtype_0, x = write_indices_float_53_cast_fp16)[name = tensor("cast_620")]; + tensor var_3718 = reshape(shape = var_3717, x = write_indices_float_53_cast_fp16_to_int32)[name = tensor("op_3718")]; + tensor write_indices_17_reps_0 = const()[name = tensor("write_indices_17_reps_0"), val = tensor([1, 1, 16, 64])]; + tensor write_indices_17 = tile(reps = write_indices_17_reps_0, x = var_3718)[name = tensor("write_indices_17")]; + tensor var_3728_begin_0 = const()[name = tensor("op_3728_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_3728_end_0 = const()[name = tensor("op_3728_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor var_3728_end_mask_0 = const()[name = tensor("op_3728_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_3728_squeeze_mask_0 = const()[name = tensor("op_3728_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor cache8_to_fp16_dtype_0 = const()[name = tensor("cache8_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor cache8_to_fp16 = cast(dtype = cache8_to_fp16_dtype_0, x = cache8)[name = tensor("cast_619")]; + tensor var_3728_cast_fp16 = slice_by_index(begin = var_3728_begin_0, end = var_3728_end_0, end_mask = var_3728_end_mask_0, squeeze_mask = var_3728_squeeze_mask_0, x = cache8_to_fp16)[name = tensor("op_3728_cast_fp16")]; + tensor var_3730_axis_0 = const()[name = tensor("op_3730_axis_0"), val = tensor(1)]; + tensor var_3730_mode_0 = const()[name = tensor("op_3730_mode_0"), val = tensor("update")]; + tensor var_3730_validate_indices_0 = const()[name = tensor("op_3730_validate_indices_0"), val = tensor(false)]; + tensor var_3730_cast_fp16 = scatter_along_axis(axis = var_3730_axis_0, data = var_3728_cast_fp16, indices = write_indices_17, mode = var_3730_mode_0, updates = k_35_cast_fp16, validate_indices = var_3730_validate_indices_0)[name = tensor("op_3730_cast_fp16")]; + tensor concat_66 = const()[name = tensor("concat_66"), val = tensor([0, 0, 0, 0, 0])]; + tensor concat_67 = const()[name = tensor("concat_67"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_17_internal_tensor_assign_1_stride_0 = const()[name = tensor("new_cache_17_internal_tensor_assign_1_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_17_internal_tensor_assign_1_begin_mask_0 = const()[name = tensor("new_cache_17_internal_tensor_assign_1_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_17_internal_tensor_assign_1_end_mask_0 = const()[name = tensor("new_cache_17_internal_tensor_assign_1_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_17_internal_tensor_assign_1_squeeze_mask_0 = const()[name = tensor("new_cache_17_internal_tensor_assign_1_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_62 = const()[name = tensor("shape_62"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_16 = const()[name = tensor("reduce_prod_16"), val = tensor(1048576)]; + tensor range_1d_16_start_0 = const()[name = tensor("range_1d_16_start_0"), val = tensor(0)]; + tensor range_1d_16_step_0 = const()[name = tensor("range_1d_16_step_0"), val = tensor(1)]; + tensor range_1d_16 = range_1d(end = reduce_prod_16, start = range_1d_16_start_0, step = range_1d_16_step_0)[name = tensor("range_1d_16")]; + tensor reshape_80 = reshape(shape = shape_62, x = range_1d_16)[name = tensor("reshape_80")]; + tensor slice_by_index_16 = slice_by_index(begin = concat_66, begin_mask = new_cache_17_internal_tensor_assign_1_begin_mask_0, end = concat_67, end_mask = new_cache_17_internal_tensor_assign_1_end_mask_0, squeeze_mask = new_cache_17_internal_tensor_assign_1_squeeze_mask_0, stride = new_cache_17_internal_tensor_assign_1_stride_0, x = reshape_80)[name = tensor("slice_by_index_16")]; + tensor reshape_81_shape_0 = const()[name = tensor("reshape_81_shape_0"), val = tensor([-1])]; + tensor reshape_81 = reshape(shape = reshape_81_shape_0, x = slice_by_index_16)[name = tensor("reshape_81")]; + tensor reshape_82_shape_0 = const()[name = tensor("reshape_82_shape_0"), val = tensor([-1])]; + tensor reshape_82_cast_fp16 = reshape(shape = reshape_82_shape_0, x = var_3730_cast_fp16)[name = tensor("reshape_82_cast_fp16")]; + tensor reshape_83_shape_0 = const()[name = tensor("reshape_83_shape_0"), val = tensor([-1])]; + tensor reshape_83_cast_fp16 = reshape(shape = reshape_83_shape_0, x = cache8_to_fp16)[name = tensor("reshape_83_cast_fp16")]; + tensor scatter_16_mode_0 = const()[name = tensor("scatter_16_mode_0"), val = tensor("update")]; + tensor scatter_16_axis_0 = const()[name = tensor("scatter_16_axis_0"), val = tensor(0)]; + tensor scatter_16_validate_indices_0 = const()[name = tensor("scatter_16_validate_indices_0"), val = tensor(false)]; + tensor scatter_16_cast_fp16 = scatter(axis = scatter_16_axis_0, data = reshape_83_cast_fp16, indices = reshape_81, mode = scatter_16_mode_0, updates = reshape_82_cast_fp16, validate_indices = scatter_16_validate_indices_0)[name = tensor("scatter_16_cast_fp16")]; + tensor reshape_84_cast_fp16 = reshape(shape = shape_62, x = scatter_16_cast_fp16)[name = tensor("reshape_84_cast_fp16")]; + tensor var_3738_begin_0 = const()[name = tensor("op_3738_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor var_3738_end_0 = const()[name = tensor("op_3738_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor var_3738_end_mask_0 = const()[name = tensor("op_3738_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_3738_squeeze_mask_0 = const()[name = tensor("op_3738_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor var_3738_cast_fp16 = slice_by_index(begin = var_3738_begin_0, end = var_3738_end_0, end_mask = var_3738_end_mask_0, squeeze_mask = var_3738_squeeze_mask_0, x = reshape_84_cast_fp16)[name = tensor("op_3738_cast_fp16")]; + tensor var_3740_axis_0 = const()[name = tensor("op_3740_axis_0"), val = tensor(1)]; + tensor var_3740_mode_0 = const()[name = tensor("op_3740_mode_0"), val = tensor("update")]; + tensor var_3740_validate_indices_0 = const()[name = tensor("op_3740_validate_indices_0"), val = tensor(false)]; + tensor var_3740_cast_fp16 = scatter_along_axis(axis = var_3740_axis_0, data = var_3738_cast_fp16, indices = write_indices_17, mode = var_3740_mode_0, updates = v_17_cast_fp16, validate_indices = var_3740_validate_indices_0)[name = tensor("op_3740_cast_fp16")]; + tensor concat_68 = const()[name = tensor("concat_68"), val = tensor([1, 0, 0, 0, 0])]; + tensor concat_69 = const()[name = tensor("concat_69"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_17_internal_tensor_assign_2_stride_0 = const()[name = tensor("new_cache_17_internal_tensor_assign_2_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_17_internal_tensor_assign_2_begin_mask_0 = const()[name = tensor("new_cache_17_internal_tensor_assign_2_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_17_internal_tensor_assign_2_end_mask_0 = const()[name = tensor("new_cache_17_internal_tensor_assign_2_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_17_internal_tensor_assign_2_squeeze_mask_0 = const()[name = tensor("new_cache_17_internal_tensor_assign_2_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_63 = const()[name = tensor("shape_63"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_17 = const()[name = tensor("reduce_prod_17"), val = tensor(1048576)]; + tensor range_1d_17_start_0 = const()[name = tensor("range_1d_17_start_0"), val = tensor(0)]; + tensor range_1d_17_step_0 = const()[name = tensor("range_1d_17_step_0"), val = tensor(1)]; + tensor range_1d_17 = range_1d(end = reduce_prod_17, start = range_1d_17_start_0, step = range_1d_17_step_0)[name = tensor("range_1d_17")]; + tensor reshape_85 = reshape(shape = shape_63, x = range_1d_17)[name = tensor("reshape_85")]; + tensor slice_by_index_17 = slice_by_index(begin = concat_68, begin_mask = new_cache_17_internal_tensor_assign_2_begin_mask_0, end = concat_69, end_mask = new_cache_17_internal_tensor_assign_2_end_mask_0, squeeze_mask = new_cache_17_internal_tensor_assign_2_squeeze_mask_0, stride = new_cache_17_internal_tensor_assign_2_stride_0, x = reshape_85)[name = tensor("slice_by_index_17")]; + tensor reshape_86_shape_0 = const()[name = tensor("reshape_86_shape_0"), val = tensor([-1])]; + tensor reshape_86 = reshape(shape = reshape_86_shape_0, x = slice_by_index_17)[name = tensor("reshape_86")]; + tensor reshape_87_shape_0 = const()[name = tensor("reshape_87_shape_0"), val = tensor([-1])]; + tensor reshape_87_cast_fp16 = reshape(shape = reshape_87_shape_0, x = var_3740_cast_fp16)[name = tensor("reshape_87_cast_fp16")]; + tensor reshape_88_shape_0 = const()[name = tensor("reshape_88_shape_0"), val = tensor([-1])]; + tensor reshape_88_cast_fp16 = reshape(shape = reshape_88_shape_0, x = reshape_84_cast_fp16)[name = tensor("reshape_88_cast_fp16")]; + tensor scatter_17_mode_0 = const()[name = tensor("scatter_17_mode_0"), val = tensor("update")]; + tensor scatter_17_axis_0 = const()[name = tensor("scatter_17_axis_0"), val = tensor(0)]; + tensor scatter_17_validate_indices_0 = const()[name = tensor("scatter_17_validate_indices_0"), val = tensor(false)]; + tensor scatter_17_cast_fp16 = scatter(axis = scatter_17_axis_0, data = reshape_88_cast_fp16, indices = reshape_86, mode = scatter_17_mode_0, updates = reshape_87_cast_fp16, validate_indices = scatter_17_validate_indices_0)[name = tensor("scatter_17_cast_fp16")]; + tensor reshape_89_cast_fp16 = reshape(shape = shape_63, x = scatter_17_cast_fp16)[name = tensor("reshape_89_cast_fp16")]; + tensor reshape_89_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("reshape_89_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor keys_49_begin_0 = const()[name = tensor("keys_49_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor keys_49_end_0 = const()[name = tensor("keys_49_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor keys_49_end_mask_0 = const()[name = tensor("keys_49_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor keys_49_squeeze_mask_0 = const()[name = tensor("keys_49_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor keys_49_cast_fp16 = slice_by_index(begin = keys_49_begin_0, end = keys_49_end_0, end_mask = keys_49_end_mask_0, squeeze_mask = keys_49_squeeze_mask_0, x = reshape_89_cast_fp16)[name = tensor("keys_49_cast_fp16")]; + tensor values_49_begin_0 = const()[name = tensor("values_49_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor values_49_end_0 = const()[name = tensor("values_49_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor values_49_end_mask_0 = const()[name = tensor("values_49_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor values_49_squeeze_mask_0 = const()[name = tensor("values_49_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor values_49_cast_fp16 = slice_by_index(begin = values_49_begin_0, end = values_49_end_0, end_mask = values_49_end_mask_0, squeeze_mask = values_49_squeeze_mask_0, x = reshape_89_cast_fp16)[name = tensor("values_49_cast_fp16")]; + tensor var_3752_cast_fp16 = not_equal(x = keys_49_cast_fp16, y = keys_49_cast_fp16)[name = tensor("op_3752_cast_fp16")]; + tensor keys_51_cast_fp16 = select(a = var_470_to_fp16, b = keys_49_cast_fp16, cond = var_3752_cast_fp16)[name = tensor("keys_51_cast_fp16")]; + tensor var_3760_cast_fp16 = not_equal(x = values_49_cast_fp16, y = values_49_cast_fp16)[name = tensor("op_3760_cast_fp16")]; + tensor values_51_cast_fp16 = select(a = var_470_to_fp16, b = values_49_cast_fp16, cond = var_3760_cast_fp16)[name = tensor("values_51_cast_fp16")]; + tensor var_3784 = const()[name = tensor("op_3784"), val = tensor([0, 2, 1, 3])]; + tensor var_3800 = const()[name = tensor("op_3800"), val = tensor([1, 1, 1])]; + tensor var_3801_cast_fp16 = reshape(shape = var_3800, x = position8_to_fp16)[name = tensor("op_3801_cast_fp16")]; + tensor q_positions_17_cast_fp16 = add(x = var_3801_cast_fp16, y = q_offsets_1_promoted_to_fp16)[name = tensor("q_positions_17_cast_fp16")]; + tensor valid_end_17_cast_fp16 = add(x = var_3801_cast_fp16, y = var_536_cast_fp16)[name = tensor("valid_end_17_cast_fp16")]; + tensor valid_mask_17_cast_fp16 = less(x = k_positions_1_promoted_to_fp16, y = valid_end_17_cast_fp16)[name = tensor("valid_mask_17_cast_fp16")]; + tensor causal_mask_17_cast_fp16 = less_equal(x = k_positions_1_promoted_to_fp16, y = q_positions_17_cast_fp16)[name = tensor("causal_mask_17_cast_fp16")]; + tensor var_3829 = logical_and(x = valid_mask_17_cast_fp16, y = causal_mask_17_cast_fp16)[name = tensor("op_3829")]; + tensor attn_mask_17_axes_0 = const()[name = tensor("attn_mask_17_axes_0"), val = tensor([1])]; + tensor attn_mask_17 = expand_dims(axes = attn_mask_17_axes_0, x = var_3829)[name = tensor("attn_mask_17")]; + tensor var_3843_transpose_x_0 = const()[name = tensor("op_3843_transpose_x_0"), val = tensor(false)]; + tensor var_3843_transpose_y_0 = const()[name = tensor("op_3843_transpose_y_0"), val = tensor(false)]; + tensor transpose_85_perm_0 = const()[name = tensor("transpose_85_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_86_perm_0 = const()[name = tensor("transpose_86_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_86 = transpose(perm = transpose_86_perm_0, x = keys_51_cast_fp16)[name = tensor("transpose_172")]; + tensor transpose_85 = transpose(perm = transpose_85_perm_0, x = q_51_cast_fp16)[name = tensor("transpose_173")]; + tensor var_3843_cast_fp16 = matmul(transpose_x = var_3843_transpose_x_0, transpose_y = var_3843_transpose_y_0, x = transpose_85, y = transpose_86)[name = tensor("op_3843_cast_fp16")]; + tensor var_3837_to_fp16 = const()[name = tensor("op_3837_to_fp16"), val = tensor([0x1p-3])]; + tensor attn_weights_49_cast_fp16 = mul(x = var_3843_cast_fp16, y = var_3837_to_fp16)[name = tensor("attn_weights_49_cast_fp16")]; + tensor var_3845 = logical_not(x = attn_mask_17)[name = tensor("op_3845")]; + tensor var_3846_to_fp16 = const()[name = tensor("op_3846_to_fp16"), val = tensor(-inf)]; + tensor attn_weights_51_cast_fp16 = select(a = var_3846_to_fp16, b = attn_weights_49_cast_fp16, cond = var_3845)[name = tensor("attn_weights_51_cast_fp16")]; + tensor var_3848 = const()[name = tensor("op_3848"), val = tensor(-1)]; + tensor attn_weights_53_cast_fp16 = softmax(axis = var_3848, x = attn_weights_51_cast_fp16)[name = tensor("attn_weights_53_cast_fp16")]; + tensor attn_output_17_transpose_x_0 = const()[name = tensor("attn_output_17_transpose_x_0"), val = tensor(false)]; + tensor attn_output_17_transpose_y_0 = const()[name = tensor("attn_output_17_transpose_y_0"), val = tensor(false)]; + tensor values_53_cast_fp16 = transpose(perm = var_3784, x = values_51_cast_fp16)[name = tensor("transpose_174")]; + tensor attn_output_17_cast_fp16 = matmul(transpose_x = attn_output_17_transpose_x_0, transpose_y = attn_output_17_transpose_y_0, x = attn_weights_53_cast_fp16, y = values_53_cast_fp16)[name = tensor("attn_output_17_cast_fp16")]; + tensor var_3856 = const()[name = tensor("op_3856"), val = tensor([0, 2, 1, 3])]; + tensor var_3861 = const()[name = tensor("op_3861"), val = tensor([1, 256, 1024])]; + tensor var_3857_cast_fp16 = transpose(perm = var_3856, x = attn_output_17_cast_fp16)[name = tensor("transpose_171")]; + tensor input_101_cast_fp16 = reshape(shape = var_3861, x = var_3857_cast_fp16)[name = tensor("input_101_cast_fp16")]; + tensor attn8_out_proj_weight_to_fp16 = const()[name = tensor("attn8_out_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(208761216)))]; + tensor linear_33_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = attn8_out_proj_weight_to_fp16, x = input_101_cast_fp16)[name = tensor("linear_33_cast_fp16")]; + tensor var_3870_cast_fp16 = add(x = position8_to_fp16, y = valid_len_to_fp16)[name = tensor("op_3870_cast_fp16")]; + tensor var_3870_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_3870_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor input_103_cast_fp16 = add(x = input_97_cast_fp16, y = linear_33_cast_fp16)[name = tensor("input_103_cast_fp16")]; + tensor input_105_axes_0 = const()[name = tensor("input_105_axes_0"), val = tensor([-1])]; + tensor norm8_2_weight_to_fp16 = const()[name = tensor("norm8_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(210858432)))]; + tensor norm8_2_bias_to_fp16 = const()[name = tensor("norm8_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(210860544)))]; + tensor var_3874_to_fp16 = const()[name = tensor("op_3874_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_105_cast_fp16 = layer_norm(axes = input_105_axes_0, beta = norm8_2_bias_to_fp16, epsilon = var_3874_to_fp16, gamma = norm8_2_weight_to_fp16, x = input_103_cast_fp16)[name = tensor("input_105_cast_fp16")]; + tensor linear8_1_weight_to_fp16 = const()[name = tensor("linear8_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(210862656)))]; + tensor linear_34_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = linear8_1_weight_to_fp16, x = input_105_cast_fp16)[name = tensor("linear_34_cast_fp16")]; + tensor input_107_mode_0 = const()[name = tensor("input_107_mode_0"), val = tensor("EXACT")]; + tensor input_107_cast_fp16 = gelu(mode = input_107_mode_0, x = linear_34_cast_fp16)[name = tensor("input_107_cast_fp16")]; + tensor linear8_2_weight_to_fp16 = const()[name = tensor("linear8_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(219251328)))]; + tensor linear_35_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = linear8_2_weight_to_fp16, x = input_107_cast_fp16)[name = tensor("linear_35_cast_fp16")]; + tensor input_109_cast_fp16 = add(x = input_103_cast_fp16, y = linear_35_cast_fp16)[name = tensor("input_109_cast_fp16")]; + tensor input_111_axes_0 = const()[name = tensor("input_111_axes_0"), val = tensor([-1])]; + tensor norm9_1_weight_to_fp16 = const()[name = tensor("norm9_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(227640000)))]; + tensor norm9_1_bias_to_fp16 = const()[name = tensor("norm9_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(227642112)))]; + tensor var_3891_to_fp16 = const()[name = tensor("op_3891_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_111_cast_fp16 = layer_norm(axes = input_111_axes_0, beta = norm9_1_bias_to_fp16, epsilon = var_3891_to_fp16, gamma = norm9_1_weight_to_fp16, x = input_109_cast_fp16)[name = tensor("input_111_cast_fp16")]; + tensor attn9_in_proj_weight_to_fp16 = const()[name = tensor("attn9_in_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(227644224)))]; + tensor linear_36_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = attn9_in_proj_weight_to_fp16, x = input_111_cast_fp16)[name = tensor("linear_36_cast_fp16")]; + tensor var_3905 = const()[name = tensor("op_3905"), val = tensor([1, 256, 3, 16, 64])]; + tensor qkv_19_cast_fp16 = reshape(shape = var_3905, x = linear_36_cast_fp16)[name = tensor("qkv_19_cast_fp16")]; + tensor q_55_begin_0 = const()[name = tensor("q_55_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor q_55_end_0 = const()[name = tensor("q_55_end_0"), val = tensor([1, 256, 1, 16, 64])]; + tensor q_55_end_mask_0 = const()[name = tensor("q_55_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor q_55_squeeze_mask_0 = const()[name = tensor("q_55_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor q_55_cast_fp16 = slice_by_index(begin = q_55_begin_0, end = q_55_end_0, end_mask = q_55_end_mask_0, squeeze_mask = q_55_squeeze_mask_0, x = qkv_19_cast_fp16)[name = tensor("q_55_cast_fp16")]; + tensor k_37_begin_0 = const()[name = tensor("k_37_begin_0"), val = tensor([0, 0, 1, 0, 0])]; + tensor k_37_end_0 = const()[name = tensor("k_37_end_0"), val = tensor([1, 256, 2, 16, 64])]; + tensor k_37_end_mask_0 = const()[name = tensor("k_37_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor k_37_squeeze_mask_0 = const()[name = tensor("k_37_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor k_37_cast_fp16 = slice_by_index(begin = k_37_begin_0, end = k_37_end_0, end_mask = k_37_end_mask_0, squeeze_mask = k_37_squeeze_mask_0, x = qkv_19_cast_fp16)[name = tensor("k_37_cast_fp16")]; + tensor v_19_begin_0 = const()[name = tensor("v_19_begin_0"), val = tensor([0, 0, 2, 0, 0])]; + tensor v_19_end_0 = const()[name = tensor("v_19_end_0"), val = tensor([1, 256, 3, 16, 64])]; + tensor v_19_end_mask_0 = const()[name = tensor("v_19_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor v_19_squeeze_mask_0 = const()[name = tensor("v_19_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor v_19_cast_fp16 = slice_by_index(begin = v_19_begin_0, end = v_19_end_0, end_mask = v_19_end_mask_0, squeeze_mask = v_19_squeeze_mask_0, x = qkv_19_cast_fp16)[name = tensor("v_19_cast_fp16")]; + tensor var_3963 = const()[name = tensor("op_3963"), val = tensor([1, 1])]; + tensor position9_to_fp16_dtype_0 = const()[name = tensor("position9_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor position9_to_fp16 = cast(dtype = position9_to_fp16_dtype_0, x = position9)[name = tensor("cast_616")]; + tensor var_3964_cast_fp16 = reshape(shape = var_3963, x = position9_to_fp16)[name = tensor("op_3964_cast_fp16")]; + tensor ts_57_cast_fp16 = add(x = ts_1_promoted_to_fp16, y = var_3964_cast_fp16)[name = tensor("ts_57_cast_fp16")]; + tensor var_3971 = const()[name = tensor("op_3971"), val = tensor([1, 256, 1, 1])]; + tensor ts_59_cast_fp16 = reshape(shape = var_3971, x = ts_57_cast_fp16)[name = tensor("ts_59_cast_fp16")]; + tensor var_3978 = const()[name = tensor("op_3978"), val = tensor([1, 256, 16, 32, 2])]; + tensor q_complex_19_cast_fp16 = reshape(shape = var_3978, x = q_55_cast_fp16)[name = tensor("q_complex_19_cast_fp16")]; + tensor var_3985 = const()[name = tensor("op_3985"), val = tensor([1, 256, 16, 32, 2])]; + tensor k_complex_19_cast_fp16 = reshape(shape = var_3985, x = k_37_cast_fp16)[name = tensor("k_complex_19_cast_fp16")]; + tensor var_3989_begin_0 = const()[name = tensor("op_3989_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_3989_end_0 = const()[name = tensor("op_3989_end_0"), val = tensor([1, 256, 16, 32, 1])]; + tensor var_3989_end_mask_0 = const()[name = tensor("op_3989_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_3989_squeeze_mask_0 = const()[name = tensor("op_3989_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_3989_cast_fp16 = slice_by_index(begin = var_3989_begin_0, end = var_3989_end_0, end_mask = var_3989_end_mask_0, squeeze_mask = var_3989_squeeze_mask_0, x = q_complex_19_cast_fp16)[name = tensor("op_3989_cast_fp16")]; + tensor var_3997_begin_0 = const()[name = tensor("op_3997_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_3997_end_0 = const()[name = tensor("op_3997_end_0"), val = tensor([1, 256, 16, 32, 2])]; + tensor var_3997_end_mask_0 = const()[name = tensor("op_3997_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_3997_squeeze_mask_0 = const()[name = tensor("op_3997_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_3997_cast_fp16 = slice_by_index(begin = var_3997_begin_0, end = var_3997_end_0, end_mask = var_3997_end_mask_0, squeeze_mask = var_3997_squeeze_mask_0, x = q_complex_19_cast_fp16)[name = tensor("op_3997_cast_fp16")]; + tensor var_4005_begin_0 = const()[name = tensor("op_4005_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_4005_end_0 = const()[name = tensor("op_4005_end_0"), val = tensor([1, 256, 16, 32, 1])]; + tensor var_4005_end_mask_0 = const()[name = tensor("op_4005_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_4005_squeeze_mask_0 = const()[name = tensor("op_4005_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_4005_cast_fp16 = slice_by_index(begin = var_4005_begin_0, end = var_4005_end_0, end_mask = var_4005_end_mask_0, squeeze_mask = var_4005_squeeze_mask_0, x = k_complex_19_cast_fp16)[name = tensor("op_4005_cast_fp16")]; + tensor var_4013_begin_0 = const()[name = tensor("op_4013_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_4013_end_0 = const()[name = tensor("op_4013_end_0"), val = tensor([1, 256, 16, 32, 2])]; + tensor var_4013_end_mask_0 = const()[name = tensor("op_4013_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_4013_squeeze_mask_0 = const()[name = tensor("op_4013_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_4013_cast_fp16 = slice_by_index(begin = var_4013_begin_0, end = var_4013_end_0, end_mask = var_4013_end_mask_0, squeeze_mask = var_4013_squeeze_mask_0, x = k_complex_19_cast_fp16)[name = tensor("op_4013_cast_fp16")]; + tensor freqs_19_to_fp16 = const()[name = tensor("freqs_19_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(233935744)))]; + tensor var_4019_cast_fp16 = mul(x = freqs_19_to_fp16, y = ts_59_cast_fp16)[name = tensor("op_4019_cast_fp16")]; + tensor rotr_19_cast_fp16 = cos(x = var_4019_cast_fp16)[name = tensor("rotr_19_cast_fp16")]; + tensor roti_19_cast_fp16 = sin(x = var_4019_cast_fp16)[name = tensor("roti_19_cast_fp16")]; + tensor var_4023_cast_fp16 = mul(x = var_3989_cast_fp16, y = rotr_19_cast_fp16)[name = tensor("op_4023_cast_fp16")]; + tensor var_4024_cast_fp16 = mul(x = var_3997_cast_fp16, y = roti_19_cast_fp16)[name = tensor("op_4024_cast_fp16")]; + tensor qor_37_cast_fp16 = sub(x = var_4023_cast_fp16, y = var_4024_cast_fp16)[name = tensor("qor_37_cast_fp16")]; + tensor var_4027_cast_fp16 = mul(x = var_3989_cast_fp16, y = roti_19_cast_fp16)[name = tensor("op_4027_cast_fp16")]; + tensor var_4028_cast_fp16 = mul(x = var_3997_cast_fp16, y = rotr_19_cast_fp16)[name = tensor("op_4028_cast_fp16")]; + tensor qoi_37_cast_fp16 = add(x = var_4027_cast_fp16, y = var_4028_cast_fp16)[name = tensor("qoi_37_cast_fp16")]; + tensor var_4031_cast_fp16 = mul(x = var_4005_cast_fp16, y = rotr_19_cast_fp16)[name = tensor("op_4031_cast_fp16")]; + tensor var_4032_cast_fp16 = mul(x = var_4013_cast_fp16, y = roti_19_cast_fp16)[name = tensor("op_4032_cast_fp16")]; + tensor kor_37_cast_fp16 = sub(x = var_4031_cast_fp16, y = var_4032_cast_fp16)[name = tensor("kor_37_cast_fp16")]; + tensor var_4035_cast_fp16 = mul(x = var_4005_cast_fp16, y = roti_19_cast_fp16)[name = tensor("op_4035_cast_fp16")]; + tensor var_4036_cast_fp16 = mul(x = var_4013_cast_fp16, y = rotr_19_cast_fp16)[name = tensor("op_4036_cast_fp16")]; + tensor koi_37_cast_fp16 = add(x = var_4035_cast_fp16, y = var_4036_cast_fp16)[name = tensor("koi_37_cast_fp16")]; + tensor qo_19_axis_0 = const()[name = tensor("qo_19_axis_0"), val = tensor(-1)]; + tensor qo_19_cast_fp16 = stack(axis = qo_19_axis_0, values = (qor_37_cast_fp16, qoi_37_cast_fp16))[name = tensor("qo_19_cast_fp16")]; + tensor ko_19_axis_0 = const()[name = tensor("ko_19_axis_0"), val = tensor(-1)]; + tensor ko_19_cast_fp16 = stack(axis = ko_19_axis_0, values = (kor_37_cast_fp16, koi_37_cast_fp16))[name = tensor("ko_19_cast_fp16")]; + tensor var_4069 = const()[name = tensor("op_4069"), val = tensor([1, 256, 16, 64])]; + tensor q_57_cast_fp16 = reshape(shape = var_4069, x = qo_19_cast_fp16)[name = tensor("q_57_cast_fp16")]; + tensor var_4075 = const()[name = tensor("op_4075"), val = tensor([1, 256, 16, 64])]; + tensor k_39_cast_fp16 = reshape(shape = var_4075, x = ko_19_cast_fp16)[name = tensor("k_39_cast_fp16")]; + tensor write_indices_float_55_cast_fp16 = add(x = var_3964_cast_fp16, y = var_393_promoted_to_fp16)[name = tensor("write_indices_float_55_cast_fp16")]; + tensor _inversed_4096_y_0_to_fp16 = const()[name = tensor("_inversed_4096_y_0_to_fp16"), val = tensor(0x1p-9)]; + tensor _inversed_4096_cast_fp16 = mul(x = write_indices_float_55_cast_fp16, y = _inversed_4096_y_0_to_fp16)[name = tensor("_inversed_4096_cast_fp16")]; + tensor var_4097_cast_fp16 = floor(x = _inversed_4096_cast_fp16)[name = tensor("op_4097_cast_fp16")]; + tensor var_4098_to_fp16 = const()[name = tensor("op_4098_to_fp16"), val = tensor(0x1p+9)]; + tensor var_4099_cast_fp16 = mul(x = var_4097_cast_fp16, y = var_4098_to_fp16)[name = tensor("op_4099_cast_fp16")]; + tensor write_indices_float_57_cast_fp16 = sub(x = write_indices_float_55_cast_fp16, y = var_4099_cast_fp16)[name = tensor("write_indices_float_57_cast_fp16")]; + tensor var_4117_value_0_to_fp16 = const()[name = tensor("op_4117_value_0_to_fp16"), val = tensor(0x1.ffp+8)]; + tensor var_4117_cast_fp16 = fill_like(ref_tensor = write_indices_float_57_cast_fp16, value = var_4117_value_0_to_fp16)[name = tensor("op_4117_cast_fp16")]; + tensor write_indices_float_59_cast_fp16 = select(a = write_indices_float_57_cast_fp16, b = var_4117_cast_fp16, cond = token_valid_1_cast_fp16)[name = tensor("write_indices_float_59_cast_fp16")]; + tensor var_4123_dtype_0 = const()[name = tensor("op_4123_dtype_0"), val = tensor("int32")]; + tensor var_4128 = const()[name = tensor("op_4128"), val = tensor([1, 256, 1, 1])]; + tensor write_indices_float_59_cast_fp16_to_int32 = cast(dtype = var_4123_dtype_0, x = write_indices_float_59_cast_fp16)[name = tensor("cast_615")]; + tensor var_4129 = reshape(shape = var_4128, x = write_indices_float_59_cast_fp16_to_int32)[name = tensor("op_4129")]; + tensor write_indices_19_reps_0 = const()[name = tensor("write_indices_19_reps_0"), val = tensor([1, 1, 16, 64])]; + tensor write_indices_19 = tile(reps = write_indices_19_reps_0, x = var_4129)[name = tensor("write_indices_19")]; + tensor var_4139_begin_0 = const()[name = tensor("op_4139_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_4139_end_0 = const()[name = tensor("op_4139_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor var_4139_end_mask_0 = const()[name = tensor("op_4139_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_4139_squeeze_mask_0 = const()[name = tensor("op_4139_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor cache9_to_fp16_dtype_0 = const()[name = tensor("cache9_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor cache9_to_fp16 = cast(dtype = cache9_to_fp16_dtype_0, x = cache9)[name = tensor("cast_614")]; + tensor var_4139_cast_fp16 = slice_by_index(begin = var_4139_begin_0, end = var_4139_end_0, end_mask = var_4139_end_mask_0, squeeze_mask = var_4139_squeeze_mask_0, x = cache9_to_fp16)[name = tensor("op_4139_cast_fp16")]; + tensor var_4141_axis_0 = const()[name = tensor("op_4141_axis_0"), val = tensor(1)]; + tensor var_4141_mode_0 = const()[name = tensor("op_4141_mode_0"), val = tensor("update")]; + tensor var_4141_validate_indices_0 = const()[name = tensor("op_4141_validate_indices_0"), val = tensor(false)]; + tensor var_4141_cast_fp16 = scatter_along_axis(axis = var_4141_axis_0, data = var_4139_cast_fp16, indices = write_indices_19, mode = var_4141_mode_0, updates = k_39_cast_fp16, validate_indices = var_4141_validate_indices_0)[name = tensor("op_4141_cast_fp16")]; + tensor concat_74 = const()[name = tensor("concat_74"), val = tensor([0, 0, 0, 0, 0])]; + tensor concat_75 = const()[name = tensor("concat_75"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_19_internal_tensor_assign_1_stride_0 = const()[name = tensor("new_cache_19_internal_tensor_assign_1_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_19_internal_tensor_assign_1_begin_mask_0 = const()[name = tensor("new_cache_19_internal_tensor_assign_1_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_19_internal_tensor_assign_1_end_mask_0 = const()[name = tensor("new_cache_19_internal_tensor_assign_1_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_19_internal_tensor_assign_1_squeeze_mask_0 = const()[name = tensor("new_cache_19_internal_tensor_assign_1_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_64 = const()[name = tensor("shape_64"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_18 = const()[name = tensor("reduce_prod_18"), val = tensor(1048576)]; + tensor range_1d_18_start_0 = const()[name = tensor("range_1d_18_start_0"), val = tensor(0)]; + tensor range_1d_18_step_0 = const()[name = tensor("range_1d_18_step_0"), val = tensor(1)]; + tensor range_1d_18 = range_1d(end = reduce_prod_18, start = range_1d_18_start_0, step = range_1d_18_step_0)[name = tensor("range_1d_18")]; + tensor reshape_90 = reshape(shape = shape_64, x = range_1d_18)[name = tensor("reshape_90")]; + tensor slice_by_index_18 = slice_by_index(begin = concat_74, begin_mask = new_cache_19_internal_tensor_assign_1_begin_mask_0, end = concat_75, end_mask = new_cache_19_internal_tensor_assign_1_end_mask_0, squeeze_mask = new_cache_19_internal_tensor_assign_1_squeeze_mask_0, stride = new_cache_19_internal_tensor_assign_1_stride_0, x = reshape_90)[name = tensor("slice_by_index_18")]; + tensor reshape_91_shape_0 = const()[name = tensor("reshape_91_shape_0"), val = tensor([-1])]; + tensor reshape_91 = reshape(shape = reshape_91_shape_0, x = slice_by_index_18)[name = tensor("reshape_91")]; + tensor reshape_92_shape_0 = const()[name = tensor("reshape_92_shape_0"), val = tensor([-1])]; + tensor reshape_92_cast_fp16 = reshape(shape = reshape_92_shape_0, x = var_4141_cast_fp16)[name = tensor("reshape_92_cast_fp16")]; + tensor reshape_93_shape_0 = const()[name = tensor("reshape_93_shape_0"), val = tensor([-1])]; + tensor reshape_93_cast_fp16 = reshape(shape = reshape_93_shape_0, x = cache9_to_fp16)[name = tensor("reshape_93_cast_fp16")]; + tensor scatter_18_mode_0 = const()[name = tensor("scatter_18_mode_0"), val = tensor("update")]; + tensor scatter_18_axis_0 = const()[name = tensor("scatter_18_axis_0"), val = tensor(0)]; + tensor scatter_18_validate_indices_0 = const()[name = tensor("scatter_18_validate_indices_0"), val = tensor(false)]; + tensor scatter_18_cast_fp16 = scatter(axis = scatter_18_axis_0, data = reshape_93_cast_fp16, indices = reshape_91, mode = scatter_18_mode_0, updates = reshape_92_cast_fp16, validate_indices = scatter_18_validate_indices_0)[name = tensor("scatter_18_cast_fp16")]; + tensor reshape_94_cast_fp16 = reshape(shape = shape_64, x = scatter_18_cast_fp16)[name = tensor("reshape_94_cast_fp16")]; + tensor var_4149_begin_0 = const()[name = tensor("op_4149_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor var_4149_end_0 = const()[name = tensor("op_4149_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor var_4149_end_mask_0 = const()[name = tensor("op_4149_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_4149_squeeze_mask_0 = const()[name = tensor("op_4149_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor var_4149_cast_fp16 = slice_by_index(begin = var_4149_begin_0, end = var_4149_end_0, end_mask = var_4149_end_mask_0, squeeze_mask = var_4149_squeeze_mask_0, x = reshape_94_cast_fp16)[name = tensor("op_4149_cast_fp16")]; + tensor var_4151_axis_0 = const()[name = tensor("op_4151_axis_0"), val = tensor(1)]; + tensor var_4151_mode_0 = const()[name = tensor("op_4151_mode_0"), val = tensor("update")]; + tensor var_4151_validate_indices_0 = const()[name = tensor("op_4151_validate_indices_0"), val = tensor(false)]; + tensor var_4151_cast_fp16 = scatter_along_axis(axis = var_4151_axis_0, data = var_4149_cast_fp16, indices = write_indices_19, mode = var_4151_mode_0, updates = v_19_cast_fp16, validate_indices = var_4151_validate_indices_0)[name = tensor("op_4151_cast_fp16")]; + tensor concat_76 = const()[name = tensor("concat_76"), val = tensor([1, 0, 0, 0, 0])]; + tensor concat_77 = const()[name = tensor("concat_77"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_19_internal_tensor_assign_2_stride_0 = const()[name = tensor("new_cache_19_internal_tensor_assign_2_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_19_internal_tensor_assign_2_begin_mask_0 = const()[name = tensor("new_cache_19_internal_tensor_assign_2_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_19_internal_tensor_assign_2_end_mask_0 = const()[name = tensor("new_cache_19_internal_tensor_assign_2_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_19_internal_tensor_assign_2_squeeze_mask_0 = const()[name = tensor("new_cache_19_internal_tensor_assign_2_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_65 = const()[name = tensor("shape_65"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_19 = const()[name = tensor("reduce_prod_19"), val = tensor(1048576)]; + tensor range_1d_19_start_0 = const()[name = tensor("range_1d_19_start_0"), val = tensor(0)]; + tensor range_1d_19_step_0 = const()[name = tensor("range_1d_19_step_0"), val = tensor(1)]; + tensor range_1d_19 = range_1d(end = reduce_prod_19, start = range_1d_19_start_0, step = range_1d_19_step_0)[name = tensor("range_1d_19")]; + tensor reshape_95 = reshape(shape = shape_65, x = range_1d_19)[name = tensor("reshape_95")]; + tensor slice_by_index_19 = slice_by_index(begin = concat_76, begin_mask = new_cache_19_internal_tensor_assign_2_begin_mask_0, end = concat_77, end_mask = new_cache_19_internal_tensor_assign_2_end_mask_0, squeeze_mask = new_cache_19_internal_tensor_assign_2_squeeze_mask_0, stride = new_cache_19_internal_tensor_assign_2_stride_0, x = reshape_95)[name = tensor("slice_by_index_19")]; + tensor reshape_96_shape_0 = const()[name = tensor("reshape_96_shape_0"), val = tensor([-1])]; + tensor reshape_96 = reshape(shape = reshape_96_shape_0, x = slice_by_index_19)[name = tensor("reshape_96")]; + tensor reshape_97_shape_0 = const()[name = tensor("reshape_97_shape_0"), val = tensor([-1])]; + tensor reshape_97_cast_fp16 = reshape(shape = reshape_97_shape_0, x = var_4151_cast_fp16)[name = tensor("reshape_97_cast_fp16")]; + tensor reshape_98_shape_0 = const()[name = tensor("reshape_98_shape_0"), val = tensor([-1])]; + tensor reshape_98_cast_fp16 = reshape(shape = reshape_98_shape_0, x = reshape_94_cast_fp16)[name = tensor("reshape_98_cast_fp16")]; + tensor scatter_19_mode_0 = const()[name = tensor("scatter_19_mode_0"), val = tensor("update")]; + tensor scatter_19_axis_0 = const()[name = tensor("scatter_19_axis_0"), val = tensor(0)]; + tensor scatter_19_validate_indices_0 = const()[name = tensor("scatter_19_validate_indices_0"), val = tensor(false)]; + tensor scatter_19_cast_fp16 = scatter(axis = scatter_19_axis_0, data = reshape_98_cast_fp16, indices = reshape_96, mode = scatter_19_mode_0, updates = reshape_97_cast_fp16, validate_indices = scatter_19_validate_indices_0)[name = tensor("scatter_19_cast_fp16")]; + tensor reshape_99_cast_fp16 = reshape(shape = shape_65, x = scatter_19_cast_fp16)[name = tensor("reshape_99_cast_fp16")]; + tensor reshape_99_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("reshape_99_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor keys_55_begin_0 = const()[name = tensor("keys_55_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor keys_55_end_0 = const()[name = tensor("keys_55_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor keys_55_end_mask_0 = const()[name = tensor("keys_55_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor keys_55_squeeze_mask_0 = const()[name = tensor("keys_55_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor keys_55_cast_fp16 = slice_by_index(begin = keys_55_begin_0, end = keys_55_end_0, end_mask = keys_55_end_mask_0, squeeze_mask = keys_55_squeeze_mask_0, x = reshape_99_cast_fp16)[name = tensor("keys_55_cast_fp16")]; + tensor values_55_begin_0 = const()[name = tensor("values_55_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor values_55_end_0 = const()[name = tensor("values_55_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor values_55_end_mask_0 = const()[name = tensor("values_55_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor values_55_squeeze_mask_0 = const()[name = tensor("values_55_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor values_55_cast_fp16 = slice_by_index(begin = values_55_begin_0, end = values_55_end_0, end_mask = values_55_end_mask_0, squeeze_mask = values_55_squeeze_mask_0, x = reshape_99_cast_fp16)[name = tensor("values_55_cast_fp16")]; + tensor var_4163_cast_fp16 = not_equal(x = keys_55_cast_fp16, y = keys_55_cast_fp16)[name = tensor("op_4163_cast_fp16")]; + tensor keys_57_cast_fp16 = select(a = var_470_to_fp16, b = keys_55_cast_fp16, cond = var_4163_cast_fp16)[name = tensor("keys_57_cast_fp16")]; + tensor var_4171_cast_fp16 = not_equal(x = values_55_cast_fp16, y = values_55_cast_fp16)[name = tensor("op_4171_cast_fp16")]; + tensor values_57_cast_fp16 = select(a = var_470_to_fp16, b = values_55_cast_fp16, cond = var_4171_cast_fp16)[name = tensor("values_57_cast_fp16")]; + tensor var_4195 = const()[name = tensor("op_4195"), val = tensor([0, 2, 1, 3])]; + tensor var_4211 = const()[name = tensor("op_4211"), val = tensor([1, 1, 1])]; + tensor var_4212_cast_fp16 = reshape(shape = var_4211, x = position9_to_fp16)[name = tensor("op_4212_cast_fp16")]; + tensor q_positions_19_cast_fp16 = add(x = var_4212_cast_fp16, y = q_offsets_1_promoted_to_fp16)[name = tensor("q_positions_19_cast_fp16")]; + tensor valid_end_19_cast_fp16 = add(x = var_4212_cast_fp16, y = var_536_cast_fp16)[name = tensor("valid_end_19_cast_fp16")]; + tensor valid_mask_19_cast_fp16 = less(x = k_positions_1_promoted_to_fp16, y = valid_end_19_cast_fp16)[name = tensor("valid_mask_19_cast_fp16")]; + tensor causal_mask_19_cast_fp16 = less_equal(x = k_positions_1_promoted_to_fp16, y = q_positions_19_cast_fp16)[name = tensor("causal_mask_19_cast_fp16")]; + tensor var_4240 = logical_and(x = valid_mask_19_cast_fp16, y = causal_mask_19_cast_fp16)[name = tensor("op_4240")]; + tensor attn_mask_19_axes_0 = const()[name = tensor("attn_mask_19_axes_0"), val = tensor([1])]; + tensor attn_mask_19 = expand_dims(axes = attn_mask_19_axes_0, x = var_4240)[name = tensor("attn_mask_19")]; + tensor var_4254_transpose_x_0 = const()[name = tensor("op_4254_transpose_x_0"), val = tensor(false)]; + tensor var_4254_transpose_y_0 = const()[name = tensor("op_4254_transpose_y_0"), val = tensor(false)]; + tensor transpose_87_perm_0 = const()[name = tensor("transpose_87_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_88_perm_0 = const()[name = tensor("transpose_88_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_88 = transpose(perm = transpose_88_perm_0, x = keys_57_cast_fp16)[name = tensor("transpose_168")]; + tensor transpose_87 = transpose(perm = transpose_87_perm_0, x = q_57_cast_fp16)[name = tensor("transpose_169")]; + tensor var_4254_cast_fp16 = matmul(transpose_x = var_4254_transpose_x_0, transpose_y = var_4254_transpose_y_0, x = transpose_87, y = transpose_88)[name = tensor("op_4254_cast_fp16")]; + tensor var_4248_to_fp16 = const()[name = tensor("op_4248_to_fp16"), val = tensor([0x1p-3])]; + tensor attn_weights_55_cast_fp16 = mul(x = var_4254_cast_fp16, y = var_4248_to_fp16)[name = tensor("attn_weights_55_cast_fp16")]; + tensor var_4256 = logical_not(x = attn_mask_19)[name = tensor("op_4256")]; + tensor var_4257_to_fp16 = const()[name = tensor("op_4257_to_fp16"), val = tensor(-inf)]; + tensor attn_weights_57_cast_fp16 = select(a = var_4257_to_fp16, b = attn_weights_55_cast_fp16, cond = var_4256)[name = tensor("attn_weights_57_cast_fp16")]; + tensor var_4259 = const()[name = tensor("op_4259"), val = tensor(-1)]; + tensor attn_weights_59_cast_fp16 = softmax(axis = var_4259, x = attn_weights_57_cast_fp16)[name = tensor("attn_weights_59_cast_fp16")]; + tensor attn_output_19_transpose_x_0 = const()[name = tensor("attn_output_19_transpose_x_0"), val = tensor(false)]; + tensor attn_output_19_transpose_y_0 = const()[name = tensor("attn_output_19_transpose_y_0"), val = tensor(false)]; + tensor values_59_cast_fp16 = transpose(perm = var_4195, x = values_57_cast_fp16)[name = tensor("transpose_170")]; + tensor attn_output_19_cast_fp16 = matmul(transpose_x = attn_output_19_transpose_x_0, transpose_y = attn_output_19_transpose_y_0, x = attn_weights_59_cast_fp16, y = values_59_cast_fp16)[name = tensor("attn_output_19_cast_fp16")]; + tensor var_4267 = const()[name = tensor("op_4267"), val = tensor([0, 2, 1, 3])]; + tensor var_4272 = const()[name = tensor("op_4272"), val = tensor([1, 256, 1024])]; + tensor var_4268_cast_fp16 = transpose(perm = var_4267, x = attn_output_19_cast_fp16)[name = tensor("transpose_167")]; + tensor input_113_cast_fp16 = reshape(shape = var_4272, x = var_4268_cast_fp16)[name = tensor("input_113_cast_fp16")]; + tensor attn9_out_proj_weight_to_fp16 = const()[name = tensor("attn9_out_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(233935872)))]; + tensor linear_37_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = attn9_out_proj_weight_to_fp16, x = input_113_cast_fp16)[name = tensor("linear_37_cast_fp16")]; + tensor var_4281_cast_fp16 = add(x = position9_to_fp16, y = valid_len_to_fp16)[name = tensor("op_4281_cast_fp16")]; + tensor var_4281_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_4281_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor input_115_cast_fp16 = add(x = input_109_cast_fp16, y = linear_37_cast_fp16)[name = tensor("input_115_cast_fp16")]; + tensor input_117_axes_0 = const()[name = tensor("input_117_axes_0"), val = tensor([-1])]; + tensor norm9_2_weight_to_fp16 = const()[name = tensor("norm9_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(236033088)))]; + tensor norm9_2_bias_to_fp16 = const()[name = tensor("norm9_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(236035200)))]; + tensor var_4285_to_fp16 = const()[name = tensor("op_4285_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_117_cast_fp16 = layer_norm(axes = input_117_axes_0, beta = norm9_2_bias_to_fp16, epsilon = var_4285_to_fp16, gamma = norm9_2_weight_to_fp16, x = input_115_cast_fp16)[name = tensor("input_117_cast_fp16")]; + tensor linear9_1_weight_to_fp16 = const()[name = tensor("linear9_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(236037312)))]; + tensor linear_38_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = linear9_1_weight_to_fp16, x = input_117_cast_fp16)[name = tensor("linear_38_cast_fp16")]; + tensor input_119_mode_0 = const()[name = tensor("input_119_mode_0"), val = tensor("EXACT")]; + tensor input_119_cast_fp16 = gelu(mode = input_119_mode_0, x = linear_38_cast_fp16)[name = tensor("input_119_cast_fp16")]; + tensor linear9_2_weight_to_fp16 = const()[name = tensor("linear9_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(244425984)))]; + tensor linear_39_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = linear9_2_weight_to_fp16, x = input_119_cast_fp16)[name = tensor("linear_39_cast_fp16")]; + tensor input_121_cast_fp16 = add(x = input_115_cast_fp16, y = linear_39_cast_fp16)[name = tensor("input_121_cast_fp16")]; + tensor input_123_axes_0 = const()[name = tensor("input_123_axes_0"), val = tensor([-1])]; + tensor norm10_1_weight_to_fp16 = const()[name = tensor("norm10_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(252814656)))]; + tensor norm10_1_bias_to_fp16 = const()[name = tensor("norm10_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(252816768)))]; + tensor var_4302_to_fp16 = const()[name = tensor("op_4302_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_123_cast_fp16 = layer_norm(axes = input_123_axes_0, beta = norm10_1_bias_to_fp16, epsilon = var_4302_to_fp16, gamma = norm10_1_weight_to_fp16, x = input_121_cast_fp16)[name = tensor("input_123_cast_fp16")]; + tensor attn10_in_proj_weight_to_fp16 = const()[name = tensor("attn10_in_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(252818880)))]; + tensor linear_40_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = attn10_in_proj_weight_to_fp16, x = input_123_cast_fp16)[name = tensor("linear_40_cast_fp16")]; + tensor var_4316 = const()[name = tensor("op_4316"), val = tensor([1, 256, 3, 16, 64])]; + tensor qkv_21_cast_fp16 = reshape(shape = var_4316, x = linear_40_cast_fp16)[name = tensor("qkv_21_cast_fp16")]; + tensor q_61_begin_0 = const()[name = tensor("q_61_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor q_61_end_0 = const()[name = tensor("q_61_end_0"), val = tensor([1, 256, 1, 16, 64])]; + tensor q_61_end_mask_0 = const()[name = tensor("q_61_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor q_61_squeeze_mask_0 = const()[name = tensor("q_61_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor q_61_cast_fp16 = slice_by_index(begin = q_61_begin_0, end = q_61_end_0, end_mask = q_61_end_mask_0, squeeze_mask = q_61_squeeze_mask_0, x = qkv_21_cast_fp16)[name = tensor("q_61_cast_fp16")]; + tensor k_41_begin_0 = const()[name = tensor("k_41_begin_0"), val = tensor([0, 0, 1, 0, 0])]; + tensor k_41_end_0 = const()[name = tensor("k_41_end_0"), val = tensor([1, 256, 2, 16, 64])]; + tensor k_41_end_mask_0 = const()[name = tensor("k_41_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor k_41_squeeze_mask_0 = const()[name = tensor("k_41_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor k_41_cast_fp16 = slice_by_index(begin = k_41_begin_0, end = k_41_end_0, end_mask = k_41_end_mask_0, squeeze_mask = k_41_squeeze_mask_0, x = qkv_21_cast_fp16)[name = tensor("k_41_cast_fp16")]; + tensor v_21_begin_0 = const()[name = tensor("v_21_begin_0"), val = tensor([0, 0, 2, 0, 0])]; + tensor v_21_end_0 = const()[name = tensor("v_21_end_0"), val = tensor([1, 256, 3, 16, 64])]; + tensor v_21_end_mask_0 = const()[name = tensor("v_21_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor v_21_squeeze_mask_0 = const()[name = tensor("v_21_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor v_21_cast_fp16 = slice_by_index(begin = v_21_begin_0, end = v_21_end_0, end_mask = v_21_end_mask_0, squeeze_mask = v_21_squeeze_mask_0, x = qkv_21_cast_fp16)[name = tensor("v_21_cast_fp16")]; + tensor var_4374 = const()[name = tensor("op_4374"), val = tensor([1, 1])]; + tensor position10_to_fp16_dtype_0 = const()[name = tensor("position10_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor position10_to_fp16 = cast(dtype = position10_to_fp16_dtype_0, x = position10)[name = tensor("cast_611")]; + tensor var_4375_cast_fp16 = reshape(shape = var_4374, x = position10_to_fp16)[name = tensor("op_4375_cast_fp16")]; + tensor ts_63_cast_fp16 = add(x = ts_1_promoted_to_fp16, y = var_4375_cast_fp16)[name = tensor("ts_63_cast_fp16")]; + tensor var_4382 = const()[name = tensor("op_4382"), val = tensor([1, 256, 1, 1])]; + tensor ts_65_cast_fp16 = reshape(shape = var_4382, x = ts_63_cast_fp16)[name = tensor("ts_65_cast_fp16")]; + tensor var_4389 = const()[name = tensor("op_4389"), val = tensor([1, 256, 16, 32, 2])]; + tensor q_complex_21_cast_fp16 = reshape(shape = var_4389, x = q_61_cast_fp16)[name = tensor("q_complex_21_cast_fp16")]; + tensor var_4396 = const()[name = tensor("op_4396"), val = tensor([1, 256, 16, 32, 2])]; + tensor k_complex_21_cast_fp16 = reshape(shape = var_4396, x = k_41_cast_fp16)[name = tensor("k_complex_21_cast_fp16")]; + tensor var_4400_begin_0 = const()[name = tensor("op_4400_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_4400_end_0 = const()[name = tensor("op_4400_end_0"), val = tensor([1, 256, 16, 32, 1])]; + tensor var_4400_end_mask_0 = const()[name = tensor("op_4400_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_4400_squeeze_mask_0 = const()[name = tensor("op_4400_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_4400_cast_fp16 = slice_by_index(begin = var_4400_begin_0, end = var_4400_end_0, end_mask = var_4400_end_mask_0, squeeze_mask = var_4400_squeeze_mask_0, x = q_complex_21_cast_fp16)[name = tensor("op_4400_cast_fp16")]; + tensor var_4408_begin_0 = const()[name = tensor("op_4408_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_4408_end_0 = const()[name = tensor("op_4408_end_0"), val = tensor([1, 256, 16, 32, 2])]; + tensor var_4408_end_mask_0 = const()[name = tensor("op_4408_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_4408_squeeze_mask_0 = const()[name = tensor("op_4408_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_4408_cast_fp16 = slice_by_index(begin = var_4408_begin_0, end = var_4408_end_0, end_mask = var_4408_end_mask_0, squeeze_mask = var_4408_squeeze_mask_0, x = q_complex_21_cast_fp16)[name = tensor("op_4408_cast_fp16")]; + tensor var_4416_begin_0 = const()[name = tensor("op_4416_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_4416_end_0 = const()[name = tensor("op_4416_end_0"), val = tensor([1, 256, 16, 32, 1])]; + tensor var_4416_end_mask_0 = const()[name = tensor("op_4416_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_4416_squeeze_mask_0 = const()[name = tensor("op_4416_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_4416_cast_fp16 = slice_by_index(begin = var_4416_begin_0, end = var_4416_end_0, end_mask = var_4416_end_mask_0, squeeze_mask = var_4416_squeeze_mask_0, x = k_complex_21_cast_fp16)[name = tensor("op_4416_cast_fp16")]; + tensor var_4424_begin_0 = const()[name = tensor("op_4424_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_4424_end_0 = const()[name = tensor("op_4424_end_0"), val = tensor([1, 256, 16, 32, 2])]; + tensor var_4424_end_mask_0 = const()[name = tensor("op_4424_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_4424_squeeze_mask_0 = const()[name = tensor("op_4424_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_4424_cast_fp16 = slice_by_index(begin = var_4424_begin_0, end = var_4424_end_0, end_mask = var_4424_end_mask_0, squeeze_mask = var_4424_squeeze_mask_0, x = k_complex_21_cast_fp16)[name = tensor("op_4424_cast_fp16")]; + tensor freqs_21_to_fp16 = const()[name = tensor("freqs_21_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(259110400)))]; + tensor var_4430_cast_fp16 = mul(x = freqs_21_to_fp16, y = ts_65_cast_fp16)[name = tensor("op_4430_cast_fp16")]; + tensor rotr_21_cast_fp16 = cos(x = var_4430_cast_fp16)[name = tensor("rotr_21_cast_fp16")]; + tensor roti_21_cast_fp16 = sin(x = var_4430_cast_fp16)[name = tensor("roti_21_cast_fp16")]; + tensor var_4434_cast_fp16 = mul(x = var_4400_cast_fp16, y = rotr_21_cast_fp16)[name = tensor("op_4434_cast_fp16")]; + tensor var_4435_cast_fp16 = mul(x = var_4408_cast_fp16, y = roti_21_cast_fp16)[name = tensor("op_4435_cast_fp16")]; + tensor qor_41_cast_fp16 = sub(x = var_4434_cast_fp16, y = var_4435_cast_fp16)[name = tensor("qor_41_cast_fp16")]; + tensor var_4438_cast_fp16 = mul(x = var_4400_cast_fp16, y = roti_21_cast_fp16)[name = tensor("op_4438_cast_fp16")]; + tensor var_4439_cast_fp16 = mul(x = var_4408_cast_fp16, y = rotr_21_cast_fp16)[name = tensor("op_4439_cast_fp16")]; + tensor qoi_41_cast_fp16 = add(x = var_4438_cast_fp16, y = var_4439_cast_fp16)[name = tensor("qoi_41_cast_fp16")]; + tensor var_4442_cast_fp16 = mul(x = var_4416_cast_fp16, y = rotr_21_cast_fp16)[name = tensor("op_4442_cast_fp16")]; + tensor var_4443_cast_fp16 = mul(x = var_4424_cast_fp16, y = roti_21_cast_fp16)[name = tensor("op_4443_cast_fp16")]; + tensor kor_41_cast_fp16 = sub(x = var_4442_cast_fp16, y = var_4443_cast_fp16)[name = tensor("kor_41_cast_fp16")]; + tensor var_4446_cast_fp16 = mul(x = var_4416_cast_fp16, y = roti_21_cast_fp16)[name = tensor("op_4446_cast_fp16")]; + tensor var_4447_cast_fp16 = mul(x = var_4424_cast_fp16, y = rotr_21_cast_fp16)[name = tensor("op_4447_cast_fp16")]; + tensor koi_41_cast_fp16 = add(x = var_4446_cast_fp16, y = var_4447_cast_fp16)[name = tensor("koi_41_cast_fp16")]; + tensor qo_21_axis_0 = const()[name = tensor("qo_21_axis_0"), val = tensor(-1)]; + tensor qo_21_cast_fp16 = stack(axis = qo_21_axis_0, values = (qor_41_cast_fp16, qoi_41_cast_fp16))[name = tensor("qo_21_cast_fp16")]; + tensor ko_21_axis_0 = const()[name = tensor("ko_21_axis_0"), val = tensor(-1)]; + tensor ko_21_cast_fp16 = stack(axis = ko_21_axis_0, values = (kor_41_cast_fp16, koi_41_cast_fp16))[name = tensor("ko_21_cast_fp16")]; + tensor var_4480 = const()[name = tensor("op_4480"), val = tensor([1, 256, 16, 64])]; + tensor q_63_cast_fp16 = reshape(shape = var_4480, x = qo_21_cast_fp16)[name = tensor("q_63_cast_fp16")]; + tensor var_4486 = const()[name = tensor("op_4486"), val = tensor([1, 256, 16, 64])]; + tensor k_43_cast_fp16 = reshape(shape = var_4486, x = ko_21_cast_fp16)[name = tensor("k_43_cast_fp16")]; + tensor write_indices_float_61_cast_fp16 = add(x = var_4375_cast_fp16, y = var_393_promoted_to_fp16)[name = tensor("write_indices_float_61_cast_fp16")]; + tensor _inversed_4507_y_0_to_fp16 = const()[name = tensor("_inversed_4507_y_0_to_fp16"), val = tensor(0x1p-9)]; + tensor _inversed_4507_cast_fp16 = mul(x = write_indices_float_61_cast_fp16, y = _inversed_4507_y_0_to_fp16)[name = tensor("_inversed_4507_cast_fp16")]; + tensor var_4508_cast_fp16 = floor(x = _inversed_4507_cast_fp16)[name = tensor("op_4508_cast_fp16")]; + tensor var_4509_to_fp16 = const()[name = tensor("op_4509_to_fp16"), val = tensor(0x1p+9)]; + tensor var_4510_cast_fp16 = mul(x = var_4508_cast_fp16, y = var_4509_to_fp16)[name = tensor("op_4510_cast_fp16")]; + tensor write_indices_float_63_cast_fp16 = sub(x = write_indices_float_61_cast_fp16, y = var_4510_cast_fp16)[name = tensor("write_indices_float_63_cast_fp16")]; + tensor var_4528_value_0_to_fp16 = const()[name = tensor("op_4528_value_0_to_fp16"), val = tensor(0x1.ffp+8)]; + tensor var_4528_cast_fp16 = fill_like(ref_tensor = write_indices_float_63_cast_fp16, value = var_4528_value_0_to_fp16)[name = tensor("op_4528_cast_fp16")]; + tensor write_indices_float_65_cast_fp16 = select(a = write_indices_float_63_cast_fp16, b = var_4528_cast_fp16, cond = token_valid_1_cast_fp16)[name = tensor("write_indices_float_65_cast_fp16")]; + tensor var_4534_dtype_0 = const()[name = tensor("op_4534_dtype_0"), val = tensor("int32")]; + tensor var_4539 = const()[name = tensor("op_4539"), val = tensor([1, 256, 1, 1])]; + tensor write_indices_float_65_cast_fp16_to_int32 = cast(dtype = var_4534_dtype_0, x = write_indices_float_65_cast_fp16)[name = tensor("cast_610")]; + tensor var_4540 = reshape(shape = var_4539, x = write_indices_float_65_cast_fp16_to_int32)[name = tensor("op_4540")]; + tensor write_indices_21_reps_0 = const()[name = tensor("write_indices_21_reps_0"), val = tensor([1, 1, 16, 64])]; + tensor write_indices_21 = tile(reps = write_indices_21_reps_0, x = var_4540)[name = tensor("write_indices_21")]; + tensor var_4550_begin_0 = const()[name = tensor("op_4550_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_4550_end_0 = const()[name = tensor("op_4550_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor var_4550_end_mask_0 = const()[name = tensor("op_4550_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_4550_squeeze_mask_0 = const()[name = tensor("op_4550_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor cache10_to_fp16_dtype_0 = const()[name = tensor("cache10_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor cache10_to_fp16 = cast(dtype = cache10_to_fp16_dtype_0, x = cache10)[name = tensor("cast_609")]; + tensor var_4550_cast_fp16 = slice_by_index(begin = var_4550_begin_0, end = var_4550_end_0, end_mask = var_4550_end_mask_0, squeeze_mask = var_4550_squeeze_mask_0, x = cache10_to_fp16)[name = tensor("op_4550_cast_fp16")]; + tensor var_4552_axis_0 = const()[name = tensor("op_4552_axis_0"), val = tensor(1)]; + tensor var_4552_mode_0 = const()[name = tensor("op_4552_mode_0"), val = tensor("update")]; + tensor var_4552_validate_indices_0 = const()[name = tensor("op_4552_validate_indices_0"), val = tensor(false)]; + tensor var_4552_cast_fp16 = scatter_along_axis(axis = var_4552_axis_0, data = var_4550_cast_fp16, indices = write_indices_21, mode = var_4552_mode_0, updates = k_43_cast_fp16, validate_indices = var_4552_validate_indices_0)[name = tensor("op_4552_cast_fp16")]; + tensor concat_82 = const()[name = tensor("concat_82"), val = tensor([0, 0, 0, 0, 0])]; + tensor concat_83 = const()[name = tensor("concat_83"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_21_internal_tensor_assign_1_stride_0 = const()[name = tensor("new_cache_21_internal_tensor_assign_1_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_21_internal_tensor_assign_1_begin_mask_0 = const()[name = tensor("new_cache_21_internal_tensor_assign_1_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_21_internal_tensor_assign_1_end_mask_0 = const()[name = tensor("new_cache_21_internal_tensor_assign_1_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_21_internal_tensor_assign_1_squeeze_mask_0 = const()[name = tensor("new_cache_21_internal_tensor_assign_1_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_66 = const()[name = tensor("shape_66"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_20 = const()[name = tensor("reduce_prod_20"), val = tensor(1048576)]; + tensor range_1d_20_start_0 = const()[name = tensor("range_1d_20_start_0"), val = tensor(0)]; + tensor range_1d_20_step_0 = const()[name = tensor("range_1d_20_step_0"), val = tensor(1)]; + tensor range_1d_20 = range_1d(end = reduce_prod_20, start = range_1d_20_start_0, step = range_1d_20_step_0)[name = tensor("range_1d_20")]; + tensor reshape_100 = reshape(shape = shape_66, x = range_1d_20)[name = tensor("reshape_100")]; + tensor slice_by_index_20 = slice_by_index(begin = concat_82, begin_mask = new_cache_21_internal_tensor_assign_1_begin_mask_0, end = concat_83, end_mask = new_cache_21_internal_tensor_assign_1_end_mask_0, squeeze_mask = new_cache_21_internal_tensor_assign_1_squeeze_mask_0, stride = new_cache_21_internal_tensor_assign_1_stride_0, x = reshape_100)[name = tensor("slice_by_index_20")]; + tensor reshape_101_shape_0 = const()[name = tensor("reshape_101_shape_0"), val = tensor([-1])]; + tensor reshape_101 = reshape(shape = reshape_101_shape_0, x = slice_by_index_20)[name = tensor("reshape_101")]; + tensor reshape_102_shape_0 = const()[name = tensor("reshape_102_shape_0"), val = tensor([-1])]; + tensor reshape_102_cast_fp16 = reshape(shape = reshape_102_shape_0, x = var_4552_cast_fp16)[name = tensor("reshape_102_cast_fp16")]; + tensor reshape_103_shape_0 = const()[name = tensor("reshape_103_shape_0"), val = tensor([-1])]; + tensor reshape_103_cast_fp16 = reshape(shape = reshape_103_shape_0, x = cache10_to_fp16)[name = tensor("reshape_103_cast_fp16")]; + tensor scatter_20_mode_0 = const()[name = tensor("scatter_20_mode_0"), val = tensor("update")]; + tensor scatter_20_axis_0 = const()[name = tensor("scatter_20_axis_0"), val = tensor(0)]; + tensor scatter_20_validate_indices_0 = const()[name = tensor("scatter_20_validate_indices_0"), val = tensor(false)]; + tensor scatter_20_cast_fp16 = scatter(axis = scatter_20_axis_0, data = reshape_103_cast_fp16, indices = reshape_101, mode = scatter_20_mode_0, updates = reshape_102_cast_fp16, validate_indices = scatter_20_validate_indices_0)[name = tensor("scatter_20_cast_fp16")]; + tensor reshape_104_cast_fp16 = reshape(shape = shape_66, x = scatter_20_cast_fp16)[name = tensor("reshape_104_cast_fp16")]; + tensor var_4560_begin_0 = const()[name = tensor("op_4560_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor var_4560_end_0 = const()[name = tensor("op_4560_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor var_4560_end_mask_0 = const()[name = tensor("op_4560_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_4560_squeeze_mask_0 = const()[name = tensor("op_4560_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor var_4560_cast_fp16 = slice_by_index(begin = var_4560_begin_0, end = var_4560_end_0, end_mask = var_4560_end_mask_0, squeeze_mask = var_4560_squeeze_mask_0, x = reshape_104_cast_fp16)[name = tensor("op_4560_cast_fp16")]; + tensor var_4562_axis_0 = const()[name = tensor("op_4562_axis_0"), val = tensor(1)]; + tensor var_4562_mode_0 = const()[name = tensor("op_4562_mode_0"), val = tensor("update")]; + tensor var_4562_validate_indices_0 = const()[name = tensor("op_4562_validate_indices_0"), val = tensor(false)]; + tensor var_4562_cast_fp16 = scatter_along_axis(axis = var_4562_axis_0, data = var_4560_cast_fp16, indices = write_indices_21, mode = var_4562_mode_0, updates = v_21_cast_fp16, validate_indices = var_4562_validate_indices_0)[name = tensor("op_4562_cast_fp16")]; + tensor concat_84 = const()[name = tensor("concat_84"), val = tensor([1, 0, 0, 0, 0])]; + tensor concat_85 = const()[name = tensor("concat_85"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_21_internal_tensor_assign_2_stride_0 = const()[name = tensor("new_cache_21_internal_tensor_assign_2_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_21_internal_tensor_assign_2_begin_mask_0 = const()[name = tensor("new_cache_21_internal_tensor_assign_2_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_21_internal_tensor_assign_2_end_mask_0 = const()[name = tensor("new_cache_21_internal_tensor_assign_2_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_21_internal_tensor_assign_2_squeeze_mask_0 = const()[name = tensor("new_cache_21_internal_tensor_assign_2_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_67 = const()[name = tensor("shape_67"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_21 = const()[name = tensor("reduce_prod_21"), val = tensor(1048576)]; + tensor range_1d_21_start_0 = const()[name = tensor("range_1d_21_start_0"), val = tensor(0)]; + tensor range_1d_21_step_0 = const()[name = tensor("range_1d_21_step_0"), val = tensor(1)]; + tensor range_1d_21 = range_1d(end = reduce_prod_21, start = range_1d_21_start_0, step = range_1d_21_step_0)[name = tensor("range_1d_21")]; + tensor reshape_105 = reshape(shape = shape_67, x = range_1d_21)[name = tensor("reshape_105")]; + tensor slice_by_index_21 = slice_by_index(begin = concat_84, begin_mask = new_cache_21_internal_tensor_assign_2_begin_mask_0, end = concat_85, end_mask = new_cache_21_internal_tensor_assign_2_end_mask_0, squeeze_mask = new_cache_21_internal_tensor_assign_2_squeeze_mask_0, stride = new_cache_21_internal_tensor_assign_2_stride_0, x = reshape_105)[name = tensor("slice_by_index_21")]; + tensor reshape_106_shape_0 = const()[name = tensor("reshape_106_shape_0"), val = tensor([-1])]; + tensor reshape_106 = reshape(shape = reshape_106_shape_0, x = slice_by_index_21)[name = tensor("reshape_106")]; + tensor reshape_107_shape_0 = const()[name = tensor("reshape_107_shape_0"), val = tensor([-1])]; + tensor reshape_107_cast_fp16 = reshape(shape = reshape_107_shape_0, x = var_4562_cast_fp16)[name = tensor("reshape_107_cast_fp16")]; + tensor reshape_108_shape_0 = const()[name = tensor("reshape_108_shape_0"), val = tensor([-1])]; + tensor reshape_108_cast_fp16 = reshape(shape = reshape_108_shape_0, x = reshape_104_cast_fp16)[name = tensor("reshape_108_cast_fp16")]; + tensor scatter_21_mode_0 = const()[name = tensor("scatter_21_mode_0"), val = tensor("update")]; + tensor scatter_21_axis_0 = const()[name = tensor("scatter_21_axis_0"), val = tensor(0)]; + tensor scatter_21_validate_indices_0 = const()[name = tensor("scatter_21_validate_indices_0"), val = tensor(false)]; + tensor scatter_21_cast_fp16 = scatter(axis = scatter_21_axis_0, data = reshape_108_cast_fp16, indices = reshape_106, mode = scatter_21_mode_0, updates = reshape_107_cast_fp16, validate_indices = scatter_21_validate_indices_0)[name = tensor("scatter_21_cast_fp16")]; + tensor reshape_109_cast_fp16 = reshape(shape = shape_67, x = scatter_21_cast_fp16)[name = tensor("reshape_109_cast_fp16")]; + tensor reshape_109_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("reshape_109_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor keys_61_begin_0 = const()[name = tensor("keys_61_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor keys_61_end_0 = const()[name = tensor("keys_61_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor keys_61_end_mask_0 = const()[name = tensor("keys_61_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor keys_61_squeeze_mask_0 = const()[name = tensor("keys_61_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor keys_61_cast_fp16 = slice_by_index(begin = keys_61_begin_0, end = keys_61_end_0, end_mask = keys_61_end_mask_0, squeeze_mask = keys_61_squeeze_mask_0, x = reshape_109_cast_fp16)[name = tensor("keys_61_cast_fp16")]; + tensor values_61_begin_0 = const()[name = tensor("values_61_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor values_61_end_0 = const()[name = tensor("values_61_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor values_61_end_mask_0 = const()[name = tensor("values_61_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor values_61_squeeze_mask_0 = const()[name = tensor("values_61_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor values_61_cast_fp16 = slice_by_index(begin = values_61_begin_0, end = values_61_end_0, end_mask = values_61_end_mask_0, squeeze_mask = values_61_squeeze_mask_0, x = reshape_109_cast_fp16)[name = tensor("values_61_cast_fp16")]; + tensor var_4574_cast_fp16 = not_equal(x = keys_61_cast_fp16, y = keys_61_cast_fp16)[name = tensor("op_4574_cast_fp16")]; + tensor keys_63_cast_fp16 = select(a = var_470_to_fp16, b = keys_61_cast_fp16, cond = var_4574_cast_fp16)[name = tensor("keys_63_cast_fp16")]; + tensor var_4582_cast_fp16 = not_equal(x = values_61_cast_fp16, y = values_61_cast_fp16)[name = tensor("op_4582_cast_fp16")]; + tensor values_63_cast_fp16 = select(a = var_470_to_fp16, b = values_61_cast_fp16, cond = var_4582_cast_fp16)[name = tensor("values_63_cast_fp16")]; + tensor var_4606 = const()[name = tensor("op_4606"), val = tensor([0, 2, 1, 3])]; + tensor var_4622 = const()[name = tensor("op_4622"), val = tensor([1, 1, 1])]; + tensor var_4623_cast_fp16 = reshape(shape = var_4622, x = position10_to_fp16)[name = tensor("op_4623_cast_fp16")]; + tensor q_positions_21_cast_fp16 = add(x = var_4623_cast_fp16, y = q_offsets_1_promoted_to_fp16)[name = tensor("q_positions_21_cast_fp16")]; + tensor valid_end_21_cast_fp16 = add(x = var_4623_cast_fp16, y = var_536_cast_fp16)[name = tensor("valid_end_21_cast_fp16")]; + tensor valid_mask_21_cast_fp16 = less(x = k_positions_1_promoted_to_fp16, y = valid_end_21_cast_fp16)[name = tensor("valid_mask_21_cast_fp16")]; + tensor causal_mask_21_cast_fp16 = less_equal(x = k_positions_1_promoted_to_fp16, y = q_positions_21_cast_fp16)[name = tensor("causal_mask_21_cast_fp16")]; + tensor var_4651 = logical_and(x = valid_mask_21_cast_fp16, y = causal_mask_21_cast_fp16)[name = tensor("op_4651")]; + tensor attn_mask_21_axes_0 = const()[name = tensor("attn_mask_21_axes_0"), val = tensor([1])]; + tensor attn_mask_21 = expand_dims(axes = attn_mask_21_axes_0, x = var_4651)[name = tensor("attn_mask_21")]; + tensor var_4665_transpose_x_0 = const()[name = tensor("op_4665_transpose_x_0"), val = tensor(false)]; + tensor var_4665_transpose_y_0 = const()[name = tensor("op_4665_transpose_y_0"), val = tensor(false)]; + tensor transpose_89_perm_0 = const()[name = tensor("transpose_89_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_90_perm_0 = const()[name = tensor("transpose_90_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_90 = transpose(perm = transpose_90_perm_0, x = keys_63_cast_fp16)[name = tensor("transpose_164")]; + tensor transpose_89 = transpose(perm = transpose_89_perm_0, x = q_63_cast_fp16)[name = tensor("transpose_165")]; + tensor var_4665_cast_fp16 = matmul(transpose_x = var_4665_transpose_x_0, transpose_y = var_4665_transpose_y_0, x = transpose_89, y = transpose_90)[name = tensor("op_4665_cast_fp16")]; + tensor var_4659_to_fp16 = const()[name = tensor("op_4659_to_fp16"), val = tensor([0x1p-3])]; + tensor attn_weights_61_cast_fp16 = mul(x = var_4665_cast_fp16, y = var_4659_to_fp16)[name = tensor("attn_weights_61_cast_fp16")]; + tensor var_4667 = logical_not(x = attn_mask_21)[name = tensor("op_4667")]; + tensor var_4668_to_fp16 = const()[name = tensor("op_4668_to_fp16"), val = tensor(-inf)]; + tensor attn_weights_63_cast_fp16 = select(a = var_4668_to_fp16, b = attn_weights_61_cast_fp16, cond = var_4667)[name = tensor("attn_weights_63_cast_fp16")]; + tensor var_4670 = const()[name = tensor("op_4670"), val = tensor(-1)]; + tensor attn_weights_65_cast_fp16 = softmax(axis = var_4670, x = attn_weights_63_cast_fp16)[name = tensor("attn_weights_65_cast_fp16")]; + tensor attn_output_21_transpose_x_0 = const()[name = tensor("attn_output_21_transpose_x_0"), val = tensor(false)]; + tensor attn_output_21_transpose_y_0 = const()[name = tensor("attn_output_21_transpose_y_0"), val = tensor(false)]; + tensor values_65_cast_fp16 = transpose(perm = var_4606, x = values_63_cast_fp16)[name = tensor("transpose_166")]; + tensor attn_output_21_cast_fp16 = matmul(transpose_x = attn_output_21_transpose_x_0, transpose_y = attn_output_21_transpose_y_0, x = attn_weights_65_cast_fp16, y = values_65_cast_fp16)[name = tensor("attn_output_21_cast_fp16")]; + tensor var_4678 = const()[name = tensor("op_4678"), val = tensor([0, 2, 1, 3])]; + tensor var_4683 = const()[name = tensor("op_4683"), val = tensor([1, 256, 1024])]; + tensor var_4679_cast_fp16 = transpose(perm = var_4678, x = attn_output_21_cast_fp16)[name = tensor("transpose_163")]; + tensor input_125_cast_fp16 = reshape(shape = var_4683, x = var_4679_cast_fp16)[name = tensor("input_125_cast_fp16")]; + tensor attn10_out_proj_weight_to_fp16 = const()[name = tensor("attn10_out_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(259110528)))]; + tensor linear_41_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = attn10_out_proj_weight_to_fp16, x = input_125_cast_fp16)[name = tensor("linear_41_cast_fp16")]; + tensor var_4692_cast_fp16 = add(x = position10_to_fp16, y = valid_len_to_fp16)[name = tensor("op_4692_cast_fp16")]; + tensor var_4692_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_4692_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor input_127_cast_fp16 = add(x = input_121_cast_fp16, y = linear_41_cast_fp16)[name = tensor("input_127_cast_fp16")]; + tensor input_129_axes_0 = const()[name = tensor("input_129_axes_0"), val = tensor([-1])]; + tensor norm10_2_weight_to_fp16 = const()[name = tensor("norm10_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(261207744)))]; + tensor norm10_2_bias_to_fp16 = const()[name = tensor("norm10_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(261209856)))]; + tensor var_4696_to_fp16 = const()[name = tensor("op_4696_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_129_cast_fp16 = layer_norm(axes = input_129_axes_0, beta = norm10_2_bias_to_fp16, epsilon = var_4696_to_fp16, gamma = norm10_2_weight_to_fp16, x = input_127_cast_fp16)[name = tensor("input_129_cast_fp16")]; + tensor linear10_1_weight_to_fp16 = const()[name = tensor("linear10_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(261211968)))]; + tensor linear_42_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = linear10_1_weight_to_fp16, x = input_129_cast_fp16)[name = tensor("linear_42_cast_fp16")]; + tensor input_131_mode_0 = const()[name = tensor("input_131_mode_0"), val = tensor("EXACT")]; + tensor input_131_cast_fp16 = gelu(mode = input_131_mode_0, x = linear_42_cast_fp16)[name = tensor("input_131_cast_fp16")]; + tensor linear10_2_weight_to_fp16 = const()[name = tensor("linear10_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(269600640)))]; + tensor linear_43_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = linear10_2_weight_to_fp16, x = input_131_cast_fp16)[name = tensor("linear_43_cast_fp16")]; + tensor input_133_cast_fp16 = add(x = input_127_cast_fp16, y = linear_43_cast_fp16)[name = tensor("input_133_cast_fp16")]; + tensor input_135_axes_0 = const()[name = tensor("input_135_axes_0"), val = tensor([-1])]; + tensor norm11_1_weight_to_fp16 = const()[name = tensor("norm11_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(277989312)))]; + tensor norm11_1_bias_to_fp16 = const()[name = tensor("norm11_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(277991424)))]; + tensor var_4713_to_fp16 = const()[name = tensor("op_4713_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_135_cast_fp16 = layer_norm(axes = input_135_axes_0, beta = norm11_1_bias_to_fp16, epsilon = var_4713_to_fp16, gamma = norm11_1_weight_to_fp16, x = input_133_cast_fp16)[name = tensor("input_135_cast_fp16")]; + tensor attn11_in_proj_weight_to_fp16 = const()[name = tensor("attn11_in_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(277993536)))]; + tensor linear_44_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = attn11_in_proj_weight_to_fp16, x = input_135_cast_fp16)[name = tensor("linear_44_cast_fp16")]; + tensor var_4727 = const()[name = tensor("op_4727"), val = tensor([1, 256, 3, 16, 64])]; + tensor qkv_23_cast_fp16 = reshape(shape = var_4727, x = linear_44_cast_fp16)[name = tensor("qkv_23_cast_fp16")]; + tensor q_67_begin_0 = const()[name = tensor("q_67_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor q_67_end_0 = const()[name = tensor("q_67_end_0"), val = tensor([1, 256, 1, 16, 64])]; + tensor q_67_end_mask_0 = const()[name = tensor("q_67_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor q_67_squeeze_mask_0 = const()[name = tensor("q_67_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor q_67_cast_fp16 = slice_by_index(begin = q_67_begin_0, end = q_67_end_0, end_mask = q_67_end_mask_0, squeeze_mask = q_67_squeeze_mask_0, x = qkv_23_cast_fp16)[name = tensor("q_67_cast_fp16")]; + tensor k_45_begin_0 = const()[name = tensor("k_45_begin_0"), val = tensor([0, 0, 1, 0, 0])]; + tensor k_45_end_0 = const()[name = tensor("k_45_end_0"), val = tensor([1, 256, 2, 16, 64])]; + tensor k_45_end_mask_0 = const()[name = tensor("k_45_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor k_45_squeeze_mask_0 = const()[name = tensor("k_45_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor k_45_cast_fp16 = slice_by_index(begin = k_45_begin_0, end = k_45_end_0, end_mask = k_45_end_mask_0, squeeze_mask = k_45_squeeze_mask_0, x = qkv_23_cast_fp16)[name = tensor("k_45_cast_fp16")]; + tensor v_23_begin_0 = const()[name = tensor("v_23_begin_0"), val = tensor([0, 0, 2, 0, 0])]; + tensor v_23_end_0 = const()[name = tensor("v_23_end_0"), val = tensor([1, 256, 3, 16, 64])]; + tensor v_23_end_mask_0 = const()[name = tensor("v_23_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor v_23_squeeze_mask_0 = const()[name = tensor("v_23_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor v_23_cast_fp16 = slice_by_index(begin = v_23_begin_0, end = v_23_end_0, end_mask = v_23_end_mask_0, squeeze_mask = v_23_squeeze_mask_0, x = qkv_23_cast_fp16)[name = tensor("v_23_cast_fp16")]; + tensor var_4785 = const()[name = tensor("op_4785"), val = tensor([1, 1])]; + tensor position11_to_fp16_dtype_0 = const()[name = tensor("position11_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor position11_to_fp16 = cast(dtype = position11_to_fp16_dtype_0, x = position11)[name = tensor("cast_606")]; + tensor var_4786_cast_fp16 = reshape(shape = var_4785, x = position11_to_fp16)[name = tensor("op_4786_cast_fp16")]; + tensor ts_69_cast_fp16 = add(x = ts_1_promoted_to_fp16, y = var_4786_cast_fp16)[name = tensor("ts_69_cast_fp16")]; + tensor var_4793 = const()[name = tensor("op_4793"), val = tensor([1, 256, 1, 1])]; + tensor ts_71_cast_fp16 = reshape(shape = var_4793, x = ts_69_cast_fp16)[name = tensor("ts_71_cast_fp16")]; + tensor var_4800 = const()[name = tensor("op_4800"), val = tensor([1, 256, 16, 32, 2])]; + tensor q_complex_23_cast_fp16 = reshape(shape = var_4800, x = q_67_cast_fp16)[name = tensor("q_complex_23_cast_fp16")]; + tensor var_4807 = const()[name = tensor("op_4807"), val = tensor([1, 256, 16, 32, 2])]; + tensor k_complex_23_cast_fp16 = reshape(shape = var_4807, x = k_45_cast_fp16)[name = tensor("k_complex_23_cast_fp16")]; + tensor var_4811_begin_0 = const()[name = tensor("op_4811_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_4811_end_0 = const()[name = tensor("op_4811_end_0"), val = tensor([1, 256, 16, 32, 1])]; + tensor var_4811_end_mask_0 = const()[name = tensor("op_4811_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_4811_squeeze_mask_0 = const()[name = tensor("op_4811_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_4811_cast_fp16 = slice_by_index(begin = var_4811_begin_0, end = var_4811_end_0, end_mask = var_4811_end_mask_0, squeeze_mask = var_4811_squeeze_mask_0, x = q_complex_23_cast_fp16)[name = tensor("op_4811_cast_fp16")]; + tensor var_4819_begin_0 = const()[name = tensor("op_4819_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_4819_end_0 = const()[name = tensor("op_4819_end_0"), val = tensor([1, 256, 16, 32, 2])]; + tensor var_4819_end_mask_0 = const()[name = tensor("op_4819_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_4819_squeeze_mask_0 = const()[name = tensor("op_4819_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_4819_cast_fp16 = slice_by_index(begin = var_4819_begin_0, end = var_4819_end_0, end_mask = var_4819_end_mask_0, squeeze_mask = var_4819_squeeze_mask_0, x = q_complex_23_cast_fp16)[name = tensor("op_4819_cast_fp16")]; + tensor var_4827_begin_0 = const()[name = tensor("op_4827_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_4827_end_0 = const()[name = tensor("op_4827_end_0"), val = tensor([1, 256, 16, 32, 1])]; + tensor var_4827_end_mask_0 = const()[name = tensor("op_4827_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_4827_squeeze_mask_0 = const()[name = tensor("op_4827_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_4827_cast_fp16 = slice_by_index(begin = var_4827_begin_0, end = var_4827_end_0, end_mask = var_4827_end_mask_0, squeeze_mask = var_4827_squeeze_mask_0, x = k_complex_23_cast_fp16)[name = tensor("op_4827_cast_fp16")]; + tensor var_4835_begin_0 = const()[name = tensor("op_4835_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_4835_end_0 = const()[name = tensor("op_4835_end_0"), val = tensor([1, 256, 16, 32, 2])]; + tensor var_4835_end_mask_0 = const()[name = tensor("op_4835_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_4835_squeeze_mask_0 = const()[name = tensor("op_4835_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_4835_cast_fp16 = slice_by_index(begin = var_4835_begin_0, end = var_4835_end_0, end_mask = var_4835_end_mask_0, squeeze_mask = var_4835_squeeze_mask_0, x = k_complex_23_cast_fp16)[name = tensor("op_4835_cast_fp16")]; + tensor freqs_23_to_fp16 = const()[name = tensor("freqs_23_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(284285056)))]; + tensor var_4841_cast_fp16 = mul(x = freqs_23_to_fp16, y = ts_71_cast_fp16)[name = tensor("op_4841_cast_fp16")]; + tensor rotr_23_cast_fp16 = cos(x = var_4841_cast_fp16)[name = tensor("rotr_23_cast_fp16")]; + tensor roti_23_cast_fp16 = sin(x = var_4841_cast_fp16)[name = tensor("roti_23_cast_fp16")]; + tensor var_4845_cast_fp16 = mul(x = var_4811_cast_fp16, y = rotr_23_cast_fp16)[name = tensor("op_4845_cast_fp16")]; + tensor var_4846_cast_fp16 = mul(x = var_4819_cast_fp16, y = roti_23_cast_fp16)[name = tensor("op_4846_cast_fp16")]; + tensor qor_45_cast_fp16 = sub(x = var_4845_cast_fp16, y = var_4846_cast_fp16)[name = tensor("qor_45_cast_fp16")]; + tensor var_4849_cast_fp16 = mul(x = var_4811_cast_fp16, y = roti_23_cast_fp16)[name = tensor("op_4849_cast_fp16")]; + tensor var_4850_cast_fp16 = mul(x = var_4819_cast_fp16, y = rotr_23_cast_fp16)[name = tensor("op_4850_cast_fp16")]; + tensor qoi_45_cast_fp16 = add(x = var_4849_cast_fp16, y = var_4850_cast_fp16)[name = tensor("qoi_45_cast_fp16")]; + tensor var_4853_cast_fp16 = mul(x = var_4827_cast_fp16, y = rotr_23_cast_fp16)[name = tensor("op_4853_cast_fp16")]; + tensor var_4854_cast_fp16 = mul(x = var_4835_cast_fp16, y = roti_23_cast_fp16)[name = tensor("op_4854_cast_fp16")]; + tensor kor_45_cast_fp16 = sub(x = var_4853_cast_fp16, y = var_4854_cast_fp16)[name = tensor("kor_45_cast_fp16")]; + tensor var_4857_cast_fp16 = mul(x = var_4827_cast_fp16, y = roti_23_cast_fp16)[name = tensor("op_4857_cast_fp16")]; + tensor var_4858_cast_fp16 = mul(x = var_4835_cast_fp16, y = rotr_23_cast_fp16)[name = tensor("op_4858_cast_fp16")]; + tensor koi_45_cast_fp16 = add(x = var_4857_cast_fp16, y = var_4858_cast_fp16)[name = tensor("koi_45_cast_fp16")]; + tensor qo_23_axis_0 = const()[name = tensor("qo_23_axis_0"), val = tensor(-1)]; + tensor qo_23_cast_fp16 = stack(axis = qo_23_axis_0, values = (qor_45_cast_fp16, qoi_45_cast_fp16))[name = tensor("qo_23_cast_fp16")]; + tensor ko_23_axis_0 = const()[name = tensor("ko_23_axis_0"), val = tensor(-1)]; + tensor ko_23_cast_fp16 = stack(axis = ko_23_axis_0, values = (kor_45_cast_fp16, koi_45_cast_fp16))[name = tensor("ko_23_cast_fp16")]; + tensor var_4891 = const()[name = tensor("op_4891"), val = tensor([1, 256, 16, 64])]; + tensor q_69_cast_fp16 = reshape(shape = var_4891, x = qo_23_cast_fp16)[name = tensor("q_69_cast_fp16")]; + tensor var_4897 = const()[name = tensor("op_4897"), val = tensor([1, 256, 16, 64])]; + tensor k_47_cast_fp16 = reshape(shape = var_4897, x = ko_23_cast_fp16)[name = tensor("k_47_cast_fp16")]; + tensor write_indices_float_67_cast_fp16 = add(x = var_4786_cast_fp16, y = var_393_promoted_to_fp16)[name = tensor("write_indices_float_67_cast_fp16")]; + tensor _inversed_4918_y_0_to_fp16 = const()[name = tensor("_inversed_4918_y_0_to_fp16"), val = tensor(0x1p-9)]; + tensor _inversed_4918_cast_fp16 = mul(x = write_indices_float_67_cast_fp16, y = _inversed_4918_y_0_to_fp16)[name = tensor("_inversed_4918_cast_fp16")]; + tensor var_4919_cast_fp16 = floor(x = _inversed_4918_cast_fp16)[name = tensor("op_4919_cast_fp16")]; + tensor var_4920_to_fp16 = const()[name = tensor("op_4920_to_fp16"), val = tensor(0x1p+9)]; + tensor var_4921_cast_fp16 = mul(x = var_4919_cast_fp16, y = var_4920_to_fp16)[name = tensor("op_4921_cast_fp16")]; + tensor write_indices_float_69_cast_fp16 = sub(x = write_indices_float_67_cast_fp16, y = var_4921_cast_fp16)[name = tensor("write_indices_float_69_cast_fp16")]; + tensor var_4939_value_0_to_fp16 = const()[name = tensor("op_4939_value_0_to_fp16"), val = tensor(0x1.ffp+8)]; + tensor var_4939_cast_fp16 = fill_like(ref_tensor = write_indices_float_69_cast_fp16, value = var_4939_value_0_to_fp16)[name = tensor("op_4939_cast_fp16")]; + tensor write_indices_float_71_cast_fp16 = select(a = write_indices_float_69_cast_fp16, b = var_4939_cast_fp16, cond = token_valid_1_cast_fp16)[name = tensor("write_indices_float_71_cast_fp16")]; + tensor var_4945_dtype_0 = const()[name = tensor("op_4945_dtype_0"), val = tensor("int32")]; + tensor var_4950 = const()[name = tensor("op_4950"), val = tensor([1, 256, 1, 1])]; + tensor write_indices_float_71_cast_fp16_to_int32 = cast(dtype = var_4945_dtype_0, x = write_indices_float_71_cast_fp16)[name = tensor("cast_605")]; + tensor var_4951 = reshape(shape = var_4950, x = write_indices_float_71_cast_fp16_to_int32)[name = tensor("op_4951")]; + tensor write_indices_23_reps_0 = const()[name = tensor("write_indices_23_reps_0"), val = tensor([1, 1, 16, 64])]; + tensor write_indices_23 = tile(reps = write_indices_23_reps_0, x = var_4951)[name = tensor("write_indices_23")]; + tensor var_4961_begin_0 = const()[name = tensor("op_4961_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_4961_end_0 = const()[name = tensor("op_4961_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor var_4961_end_mask_0 = const()[name = tensor("op_4961_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_4961_squeeze_mask_0 = const()[name = tensor("op_4961_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor cache11_to_fp16_dtype_0 = const()[name = tensor("cache11_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor cache11_to_fp16 = cast(dtype = cache11_to_fp16_dtype_0, x = cache11)[name = tensor("cast_604")]; + tensor var_4961_cast_fp16 = slice_by_index(begin = var_4961_begin_0, end = var_4961_end_0, end_mask = var_4961_end_mask_0, squeeze_mask = var_4961_squeeze_mask_0, x = cache11_to_fp16)[name = tensor("op_4961_cast_fp16")]; + tensor var_4963_axis_0 = const()[name = tensor("op_4963_axis_0"), val = tensor(1)]; + tensor var_4963_mode_0 = const()[name = tensor("op_4963_mode_0"), val = tensor("update")]; + tensor var_4963_validate_indices_0 = const()[name = tensor("op_4963_validate_indices_0"), val = tensor(false)]; + tensor var_4963_cast_fp16 = scatter_along_axis(axis = var_4963_axis_0, data = var_4961_cast_fp16, indices = write_indices_23, mode = var_4963_mode_0, updates = k_47_cast_fp16, validate_indices = var_4963_validate_indices_0)[name = tensor("op_4963_cast_fp16")]; + tensor concat_90 = const()[name = tensor("concat_90"), val = tensor([0, 0, 0, 0, 0])]; + tensor concat_91 = const()[name = tensor("concat_91"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_23_internal_tensor_assign_1_stride_0 = const()[name = tensor("new_cache_23_internal_tensor_assign_1_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_23_internal_tensor_assign_1_begin_mask_0 = const()[name = tensor("new_cache_23_internal_tensor_assign_1_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_23_internal_tensor_assign_1_end_mask_0 = const()[name = tensor("new_cache_23_internal_tensor_assign_1_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_23_internal_tensor_assign_1_squeeze_mask_0 = const()[name = tensor("new_cache_23_internal_tensor_assign_1_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_68 = const()[name = tensor("shape_68"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_22 = const()[name = tensor("reduce_prod_22"), val = tensor(1048576)]; + tensor range_1d_22_start_0 = const()[name = tensor("range_1d_22_start_0"), val = tensor(0)]; + tensor range_1d_22_step_0 = const()[name = tensor("range_1d_22_step_0"), val = tensor(1)]; + tensor range_1d_22 = range_1d(end = reduce_prod_22, start = range_1d_22_start_0, step = range_1d_22_step_0)[name = tensor("range_1d_22")]; + tensor reshape_110 = reshape(shape = shape_68, x = range_1d_22)[name = tensor("reshape_110")]; + tensor slice_by_index_22 = slice_by_index(begin = concat_90, begin_mask = new_cache_23_internal_tensor_assign_1_begin_mask_0, end = concat_91, end_mask = new_cache_23_internal_tensor_assign_1_end_mask_0, squeeze_mask = new_cache_23_internal_tensor_assign_1_squeeze_mask_0, stride = new_cache_23_internal_tensor_assign_1_stride_0, x = reshape_110)[name = tensor("slice_by_index_22")]; + tensor reshape_111_shape_0 = const()[name = tensor("reshape_111_shape_0"), val = tensor([-1])]; + tensor reshape_111 = reshape(shape = reshape_111_shape_0, x = slice_by_index_22)[name = tensor("reshape_111")]; + tensor reshape_112_shape_0 = const()[name = tensor("reshape_112_shape_0"), val = tensor([-1])]; + tensor reshape_112_cast_fp16 = reshape(shape = reshape_112_shape_0, x = var_4963_cast_fp16)[name = tensor("reshape_112_cast_fp16")]; + tensor reshape_113_shape_0 = const()[name = tensor("reshape_113_shape_0"), val = tensor([-1])]; + tensor reshape_113_cast_fp16 = reshape(shape = reshape_113_shape_0, x = cache11_to_fp16)[name = tensor("reshape_113_cast_fp16")]; + tensor scatter_22_mode_0 = const()[name = tensor("scatter_22_mode_0"), val = tensor("update")]; + tensor scatter_22_axis_0 = const()[name = tensor("scatter_22_axis_0"), val = tensor(0)]; + tensor scatter_22_validate_indices_0 = const()[name = tensor("scatter_22_validate_indices_0"), val = tensor(false)]; + tensor scatter_22_cast_fp16 = scatter(axis = scatter_22_axis_0, data = reshape_113_cast_fp16, indices = reshape_111, mode = scatter_22_mode_0, updates = reshape_112_cast_fp16, validate_indices = scatter_22_validate_indices_0)[name = tensor("scatter_22_cast_fp16")]; + tensor reshape_114_cast_fp16 = reshape(shape = shape_68, x = scatter_22_cast_fp16)[name = tensor("reshape_114_cast_fp16")]; + tensor var_4971_begin_0 = const()[name = tensor("op_4971_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor var_4971_end_0 = const()[name = tensor("op_4971_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor var_4971_end_mask_0 = const()[name = tensor("op_4971_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_4971_squeeze_mask_0 = const()[name = tensor("op_4971_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor var_4971_cast_fp16 = slice_by_index(begin = var_4971_begin_0, end = var_4971_end_0, end_mask = var_4971_end_mask_0, squeeze_mask = var_4971_squeeze_mask_0, x = reshape_114_cast_fp16)[name = tensor("op_4971_cast_fp16")]; + tensor var_4973_axis_0 = const()[name = tensor("op_4973_axis_0"), val = tensor(1)]; + tensor var_4973_mode_0 = const()[name = tensor("op_4973_mode_0"), val = tensor("update")]; + tensor var_4973_validate_indices_0 = const()[name = tensor("op_4973_validate_indices_0"), val = tensor(false)]; + tensor var_4973_cast_fp16 = scatter_along_axis(axis = var_4973_axis_0, data = var_4971_cast_fp16, indices = write_indices_23, mode = var_4973_mode_0, updates = v_23_cast_fp16, validate_indices = var_4973_validate_indices_0)[name = tensor("op_4973_cast_fp16")]; + tensor concat_92 = const()[name = tensor("concat_92"), val = tensor([1, 0, 0, 0, 0])]; + tensor concat_93 = const()[name = tensor("concat_93"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_23_internal_tensor_assign_2_stride_0 = const()[name = tensor("new_cache_23_internal_tensor_assign_2_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_23_internal_tensor_assign_2_begin_mask_0 = const()[name = tensor("new_cache_23_internal_tensor_assign_2_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_23_internal_tensor_assign_2_end_mask_0 = const()[name = tensor("new_cache_23_internal_tensor_assign_2_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_23_internal_tensor_assign_2_squeeze_mask_0 = const()[name = tensor("new_cache_23_internal_tensor_assign_2_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_69 = const()[name = tensor("shape_69"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_23 = const()[name = tensor("reduce_prod_23"), val = tensor(1048576)]; + tensor range_1d_23_start_0 = const()[name = tensor("range_1d_23_start_0"), val = tensor(0)]; + tensor range_1d_23_step_0 = const()[name = tensor("range_1d_23_step_0"), val = tensor(1)]; + tensor range_1d_23 = range_1d(end = reduce_prod_23, start = range_1d_23_start_0, step = range_1d_23_step_0)[name = tensor("range_1d_23")]; + tensor reshape_115 = reshape(shape = shape_69, x = range_1d_23)[name = tensor("reshape_115")]; + tensor slice_by_index_23 = slice_by_index(begin = concat_92, begin_mask = new_cache_23_internal_tensor_assign_2_begin_mask_0, end = concat_93, end_mask = new_cache_23_internal_tensor_assign_2_end_mask_0, squeeze_mask = new_cache_23_internal_tensor_assign_2_squeeze_mask_0, stride = new_cache_23_internal_tensor_assign_2_stride_0, x = reshape_115)[name = tensor("slice_by_index_23")]; + tensor reshape_116_shape_0 = const()[name = tensor("reshape_116_shape_0"), val = tensor([-1])]; + tensor reshape_116 = reshape(shape = reshape_116_shape_0, x = slice_by_index_23)[name = tensor("reshape_116")]; + tensor reshape_117_shape_0 = const()[name = tensor("reshape_117_shape_0"), val = tensor([-1])]; + tensor reshape_117_cast_fp16 = reshape(shape = reshape_117_shape_0, x = var_4973_cast_fp16)[name = tensor("reshape_117_cast_fp16")]; + tensor reshape_118_shape_0 = const()[name = tensor("reshape_118_shape_0"), val = tensor([-1])]; + tensor reshape_118_cast_fp16 = reshape(shape = reshape_118_shape_0, x = reshape_114_cast_fp16)[name = tensor("reshape_118_cast_fp16")]; + tensor scatter_23_mode_0 = const()[name = tensor("scatter_23_mode_0"), val = tensor("update")]; + tensor scatter_23_axis_0 = const()[name = tensor("scatter_23_axis_0"), val = tensor(0)]; + tensor scatter_23_validate_indices_0 = const()[name = tensor("scatter_23_validate_indices_0"), val = tensor(false)]; + tensor scatter_23_cast_fp16 = scatter(axis = scatter_23_axis_0, data = reshape_118_cast_fp16, indices = reshape_116, mode = scatter_23_mode_0, updates = reshape_117_cast_fp16, validate_indices = scatter_23_validate_indices_0)[name = tensor("scatter_23_cast_fp16")]; + tensor reshape_119_cast_fp16 = reshape(shape = shape_69, x = scatter_23_cast_fp16)[name = tensor("reshape_119_cast_fp16")]; + tensor reshape_119_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("reshape_119_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor keys_67_begin_0 = const()[name = tensor("keys_67_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor keys_67_end_0 = const()[name = tensor("keys_67_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor keys_67_end_mask_0 = const()[name = tensor("keys_67_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor keys_67_squeeze_mask_0 = const()[name = tensor("keys_67_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor keys_67_cast_fp16 = slice_by_index(begin = keys_67_begin_0, end = keys_67_end_0, end_mask = keys_67_end_mask_0, squeeze_mask = keys_67_squeeze_mask_0, x = reshape_119_cast_fp16)[name = tensor("keys_67_cast_fp16")]; + tensor values_67_begin_0 = const()[name = tensor("values_67_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor values_67_end_0 = const()[name = tensor("values_67_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor values_67_end_mask_0 = const()[name = tensor("values_67_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor values_67_squeeze_mask_0 = const()[name = tensor("values_67_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor values_67_cast_fp16 = slice_by_index(begin = values_67_begin_0, end = values_67_end_0, end_mask = values_67_end_mask_0, squeeze_mask = values_67_squeeze_mask_0, x = reshape_119_cast_fp16)[name = tensor("values_67_cast_fp16")]; + tensor var_4985_cast_fp16 = not_equal(x = keys_67_cast_fp16, y = keys_67_cast_fp16)[name = tensor("op_4985_cast_fp16")]; + tensor keys_69_cast_fp16 = select(a = var_470_to_fp16, b = keys_67_cast_fp16, cond = var_4985_cast_fp16)[name = tensor("keys_69_cast_fp16")]; + tensor var_4993_cast_fp16 = not_equal(x = values_67_cast_fp16, y = values_67_cast_fp16)[name = tensor("op_4993_cast_fp16")]; + tensor values_69_cast_fp16 = select(a = var_470_to_fp16, b = values_67_cast_fp16, cond = var_4993_cast_fp16)[name = tensor("values_69_cast_fp16")]; + tensor var_5017 = const()[name = tensor("op_5017"), val = tensor([0, 2, 1, 3])]; + tensor var_5033 = const()[name = tensor("op_5033"), val = tensor([1, 1, 1])]; + tensor var_5034_cast_fp16 = reshape(shape = var_5033, x = position11_to_fp16)[name = tensor("op_5034_cast_fp16")]; + tensor q_positions_23_cast_fp16 = add(x = var_5034_cast_fp16, y = q_offsets_1_promoted_to_fp16)[name = tensor("q_positions_23_cast_fp16")]; + tensor valid_end_23_cast_fp16 = add(x = var_5034_cast_fp16, y = var_536_cast_fp16)[name = tensor("valid_end_23_cast_fp16")]; + tensor valid_mask_23_cast_fp16 = less(x = k_positions_1_promoted_to_fp16, y = valid_end_23_cast_fp16)[name = tensor("valid_mask_23_cast_fp16")]; + tensor causal_mask_23_cast_fp16 = less_equal(x = k_positions_1_promoted_to_fp16, y = q_positions_23_cast_fp16)[name = tensor("causal_mask_23_cast_fp16")]; + tensor var_5062 = logical_and(x = valid_mask_23_cast_fp16, y = causal_mask_23_cast_fp16)[name = tensor("op_5062")]; + tensor attn_mask_23_axes_0 = const()[name = tensor("attn_mask_23_axes_0"), val = tensor([1])]; + tensor attn_mask_23 = expand_dims(axes = attn_mask_23_axes_0, x = var_5062)[name = tensor("attn_mask_23")]; + tensor var_5076_transpose_x_0 = const()[name = tensor("op_5076_transpose_x_0"), val = tensor(false)]; + tensor var_5076_transpose_y_0 = const()[name = tensor("op_5076_transpose_y_0"), val = tensor(false)]; + tensor transpose_91_perm_0 = const()[name = tensor("transpose_91_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_92_perm_0 = const()[name = tensor("transpose_92_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_92 = transpose(perm = transpose_92_perm_0, x = keys_69_cast_fp16)[name = tensor("transpose_160")]; + tensor transpose_91 = transpose(perm = transpose_91_perm_0, x = q_69_cast_fp16)[name = tensor("transpose_161")]; + tensor var_5076_cast_fp16 = matmul(transpose_x = var_5076_transpose_x_0, transpose_y = var_5076_transpose_y_0, x = transpose_91, y = transpose_92)[name = tensor("op_5076_cast_fp16")]; + tensor var_5070_to_fp16 = const()[name = tensor("op_5070_to_fp16"), val = tensor([0x1p-3])]; + tensor attn_weights_67_cast_fp16 = mul(x = var_5076_cast_fp16, y = var_5070_to_fp16)[name = tensor("attn_weights_67_cast_fp16")]; + tensor var_5078 = logical_not(x = attn_mask_23)[name = tensor("op_5078")]; + tensor var_5079_to_fp16 = const()[name = tensor("op_5079_to_fp16"), val = tensor(-inf)]; + tensor attn_weights_69_cast_fp16 = select(a = var_5079_to_fp16, b = attn_weights_67_cast_fp16, cond = var_5078)[name = tensor("attn_weights_69_cast_fp16")]; + tensor var_5081 = const()[name = tensor("op_5081"), val = tensor(-1)]; + tensor attn_weights_71_cast_fp16 = softmax(axis = var_5081, x = attn_weights_69_cast_fp16)[name = tensor("attn_weights_71_cast_fp16")]; + tensor attn_output_23_transpose_x_0 = const()[name = tensor("attn_output_23_transpose_x_0"), val = tensor(false)]; + tensor attn_output_23_transpose_y_0 = const()[name = tensor("attn_output_23_transpose_y_0"), val = tensor(false)]; + tensor values_71_cast_fp16 = transpose(perm = var_5017, x = values_69_cast_fp16)[name = tensor("transpose_162")]; + tensor attn_output_23_cast_fp16 = matmul(transpose_x = attn_output_23_transpose_x_0, transpose_y = attn_output_23_transpose_y_0, x = attn_weights_71_cast_fp16, y = values_71_cast_fp16)[name = tensor("attn_output_23_cast_fp16")]; + tensor var_5089 = const()[name = tensor("op_5089"), val = tensor([0, 2, 1, 3])]; + tensor var_5094 = const()[name = tensor("op_5094"), val = tensor([1, 256, 1024])]; + tensor var_5090_cast_fp16 = transpose(perm = var_5089, x = attn_output_23_cast_fp16)[name = tensor("transpose_159")]; + tensor input_137_cast_fp16 = reshape(shape = var_5094, x = var_5090_cast_fp16)[name = tensor("input_137_cast_fp16")]; + tensor attn11_out_proj_weight_to_fp16 = const()[name = tensor("attn11_out_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(284285184)))]; + tensor linear_45_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = attn11_out_proj_weight_to_fp16, x = input_137_cast_fp16)[name = tensor("linear_45_cast_fp16")]; + tensor var_5103_cast_fp16 = add(x = position11_to_fp16, y = valid_len_to_fp16)[name = tensor("op_5103_cast_fp16")]; + tensor var_5103_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_5103_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor input_139_cast_fp16 = add(x = input_133_cast_fp16, y = linear_45_cast_fp16)[name = tensor("input_139_cast_fp16")]; + tensor input_141_axes_0 = const()[name = tensor("input_141_axes_0"), val = tensor([-1])]; + tensor norm11_2_weight_to_fp16 = const()[name = tensor("norm11_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(286382400)))]; + tensor norm11_2_bias_to_fp16 = const()[name = tensor("norm11_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(286384512)))]; + tensor var_5107_to_fp16 = const()[name = tensor("op_5107_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_141_cast_fp16 = layer_norm(axes = input_141_axes_0, beta = norm11_2_bias_to_fp16, epsilon = var_5107_to_fp16, gamma = norm11_2_weight_to_fp16, x = input_139_cast_fp16)[name = tensor("input_141_cast_fp16")]; + tensor linear11_1_weight_to_fp16 = const()[name = tensor("linear11_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(286386624)))]; + tensor linear_46_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = linear11_1_weight_to_fp16, x = input_141_cast_fp16)[name = tensor("linear_46_cast_fp16")]; + tensor input_143_mode_0 = const()[name = tensor("input_143_mode_0"), val = tensor("EXACT")]; + tensor input_143_cast_fp16 = gelu(mode = input_143_mode_0, x = linear_46_cast_fp16)[name = tensor("input_143_cast_fp16")]; + tensor linear11_2_weight_to_fp16 = const()[name = tensor("linear11_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(294775296)))]; + tensor linear_47_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = linear11_2_weight_to_fp16, x = input_143_cast_fp16)[name = tensor("linear_47_cast_fp16")]; + tensor input_145_cast_fp16 = add(x = input_139_cast_fp16, y = linear_47_cast_fp16)[name = tensor("input_145_cast_fp16")]; + tensor input_147_axes_0 = const()[name = tensor("input_147_axes_0"), val = tensor([-1])]; + tensor norm12_1_weight_to_fp16 = const()[name = tensor("norm12_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(303163968)))]; + tensor norm12_1_bias_to_fp16 = const()[name = tensor("norm12_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(303166080)))]; + tensor var_5124_to_fp16 = const()[name = tensor("op_5124_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_147_cast_fp16 = layer_norm(axes = input_147_axes_0, beta = norm12_1_bias_to_fp16, epsilon = var_5124_to_fp16, gamma = norm12_1_weight_to_fp16, x = input_145_cast_fp16)[name = tensor("input_147_cast_fp16")]; + tensor attn12_in_proj_weight_to_fp16 = const()[name = tensor("attn12_in_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(303168192)))]; + tensor linear_48_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = attn12_in_proj_weight_to_fp16, x = input_147_cast_fp16)[name = tensor("linear_48_cast_fp16")]; + tensor var_5138 = const()[name = tensor("op_5138"), val = tensor([1, 256, 3, 16, 64])]; + tensor qkv_25_cast_fp16 = reshape(shape = var_5138, x = linear_48_cast_fp16)[name = tensor("qkv_25_cast_fp16")]; + tensor q_73_begin_0 = const()[name = tensor("q_73_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor q_73_end_0 = const()[name = tensor("q_73_end_0"), val = tensor([1, 256, 1, 16, 64])]; + tensor q_73_end_mask_0 = const()[name = tensor("q_73_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor q_73_squeeze_mask_0 = const()[name = tensor("q_73_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor q_73_cast_fp16 = slice_by_index(begin = q_73_begin_0, end = q_73_end_0, end_mask = q_73_end_mask_0, squeeze_mask = q_73_squeeze_mask_0, x = qkv_25_cast_fp16)[name = tensor("q_73_cast_fp16")]; + tensor k_49_begin_0 = const()[name = tensor("k_49_begin_0"), val = tensor([0, 0, 1, 0, 0])]; + tensor k_49_end_0 = const()[name = tensor("k_49_end_0"), val = tensor([1, 256, 2, 16, 64])]; + tensor k_49_end_mask_0 = const()[name = tensor("k_49_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor k_49_squeeze_mask_0 = const()[name = tensor("k_49_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor k_49_cast_fp16 = slice_by_index(begin = k_49_begin_0, end = k_49_end_0, end_mask = k_49_end_mask_0, squeeze_mask = k_49_squeeze_mask_0, x = qkv_25_cast_fp16)[name = tensor("k_49_cast_fp16")]; + tensor v_25_begin_0 = const()[name = tensor("v_25_begin_0"), val = tensor([0, 0, 2, 0, 0])]; + tensor v_25_end_0 = const()[name = tensor("v_25_end_0"), val = tensor([1, 256, 3, 16, 64])]; + tensor v_25_end_mask_0 = const()[name = tensor("v_25_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor v_25_squeeze_mask_0 = const()[name = tensor("v_25_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor v_25_cast_fp16 = slice_by_index(begin = v_25_begin_0, end = v_25_end_0, end_mask = v_25_end_mask_0, squeeze_mask = v_25_squeeze_mask_0, x = qkv_25_cast_fp16)[name = tensor("v_25_cast_fp16")]; + tensor var_5196 = const()[name = tensor("op_5196"), val = tensor([1, 1])]; + tensor position12_to_fp16_dtype_0 = const()[name = tensor("position12_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor position12_to_fp16 = cast(dtype = position12_to_fp16_dtype_0, x = position12)[name = tensor("cast_601")]; + tensor var_5197_cast_fp16 = reshape(shape = var_5196, x = position12_to_fp16)[name = tensor("op_5197_cast_fp16")]; + tensor ts_75_cast_fp16 = add(x = ts_1_promoted_to_fp16, y = var_5197_cast_fp16)[name = tensor("ts_75_cast_fp16")]; + tensor var_5204 = const()[name = tensor("op_5204"), val = tensor([1, 256, 1, 1])]; + tensor ts_77_cast_fp16 = reshape(shape = var_5204, x = ts_75_cast_fp16)[name = tensor("ts_77_cast_fp16")]; + tensor var_5211 = const()[name = tensor("op_5211"), val = tensor([1, 256, 16, 32, 2])]; + tensor q_complex_25_cast_fp16 = reshape(shape = var_5211, x = q_73_cast_fp16)[name = tensor("q_complex_25_cast_fp16")]; + tensor var_5218 = const()[name = tensor("op_5218"), val = tensor([1, 256, 16, 32, 2])]; + tensor k_complex_25_cast_fp16 = reshape(shape = var_5218, x = k_49_cast_fp16)[name = tensor("k_complex_25_cast_fp16")]; + tensor var_5222_begin_0 = const()[name = tensor("op_5222_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_5222_end_0 = const()[name = tensor("op_5222_end_0"), val = tensor([1, 256, 16, 32, 1])]; + tensor var_5222_end_mask_0 = const()[name = tensor("op_5222_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_5222_squeeze_mask_0 = const()[name = tensor("op_5222_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_5222_cast_fp16 = slice_by_index(begin = var_5222_begin_0, end = var_5222_end_0, end_mask = var_5222_end_mask_0, squeeze_mask = var_5222_squeeze_mask_0, x = q_complex_25_cast_fp16)[name = tensor("op_5222_cast_fp16")]; + tensor var_5230_begin_0 = const()[name = tensor("op_5230_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_5230_end_0 = const()[name = tensor("op_5230_end_0"), val = tensor([1, 256, 16, 32, 2])]; + tensor var_5230_end_mask_0 = const()[name = tensor("op_5230_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_5230_squeeze_mask_0 = const()[name = tensor("op_5230_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_5230_cast_fp16 = slice_by_index(begin = var_5230_begin_0, end = var_5230_end_0, end_mask = var_5230_end_mask_0, squeeze_mask = var_5230_squeeze_mask_0, x = q_complex_25_cast_fp16)[name = tensor("op_5230_cast_fp16")]; + tensor var_5238_begin_0 = const()[name = tensor("op_5238_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_5238_end_0 = const()[name = tensor("op_5238_end_0"), val = tensor([1, 256, 16, 32, 1])]; + tensor var_5238_end_mask_0 = const()[name = tensor("op_5238_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_5238_squeeze_mask_0 = const()[name = tensor("op_5238_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_5238_cast_fp16 = slice_by_index(begin = var_5238_begin_0, end = var_5238_end_0, end_mask = var_5238_end_mask_0, squeeze_mask = var_5238_squeeze_mask_0, x = k_complex_25_cast_fp16)[name = tensor("op_5238_cast_fp16")]; + tensor var_5246_begin_0 = const()[name = tensor("op_5246_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_5246_end_0 = const()[name = tensor("op_5246_end_0"), val = tensor([1, 256, 16, 32, 2])]; + tensor var_5246_end_mask_0 = const()[name = tensor("op_5246_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_5246_squeeze_mask_0 = const()[name = tensor("op_5246_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_5246_cast_fp16 = slice_by_index(begin = var_5246_begin_0, end = var_5246_end_0, end_mask = var_5246_end_mask_0, squeeze_mask = var_5246_squeeze_mask_0, x = k_complex_25_cast_fp16)[name = tensor("op_5246_cast_fp16")]; + tensor freqs_25_to_fp16 = const()[name = tensor("freqs_25_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(309459712)))]; + tensor var_5252_cast_fp16 = mul(x = freqs_25_to_fp16, y = ts_77_cast_fp16)[name = tensor("op_5252_cast_fp16")]; + tensor rotr_25_cast_fp16 = cos(x = var_5252_cast_fp16)[name = tensor("rotr_25_cast_fp16")]; + tensor roti_25_cast_fp16 = sin(x = var_5252_cast_fp16)[name = tensor("roti_25_cast_fp16")]; + tensor var_5256_cast_fp16 = mul(x = var_5222_cast_fp16, y = rotr_25_cast_fp16)[name = tensor("op_5256_cast_fp16")]; + tensor var_5257_cast_fp16 = mul(x = var_5230_cast_fp16, y = roti_25_cast_fp16)[name = tensor("op_5257_cast_fp16")]; + tensor qor_49_cast_fp16 = sub(x = var_5256_cast_fp16, y = var_5257_cast_fp16)[name = tensor("qor_49_cast_fp16")]; + tensor var_5260_cast_fp16 = mul(x = var_5222_cast_fp16, y = roti_25_cast_fp16)[name = tensor("op_5260_cast_fp16")]; + tensor var_5261_cast_fp16 = mul(x = var_5230_cast_fp16, y = rotr_25_cast_fp16)[name = tensor("op_5261_cast_fp16")]; + tensor qoi_49_cast_fp16 = add(x = var_5260_cast_fp16, y = var_5261_cast_fp16)[name = tensor("qoi_49_cast_fp16")]; + tensor var_5264_cast_fp16 = mul(x = var_5238_cast_fp16, y = rotr_25_cast_fp16)[name = tensor("op_5264_cast_fp16")]; + tensor var_5265_cast_fp16 = mul(x = var_5246_cast_fp16, y = roti_25_cast_fp16)[name = tensor("op_5265_cast_fp16")]; + tensor kor_49_cast_fp16 = sub(x = var_5264_cast_fp16, y = var_5265_cast_fp16)[name = tensor("kor_49_cast_fp16")]; + tensor var_5268_cast_fp16 = mul(x = var_5238_cast_fp16, y = roti_25_cast_fp16)[name = tensor("op_5268_cast_fp16")]; + tensor var_5269_cast_fp16 = mul(x = var_5246_cast_fp16, y = rotr_25_cast_fp16)[name = tensor("op_5269_cast_fp16")]; + tensor koi_49_cast_fp16 = add(x = var_5268_cast_fp16, y = var_5269_cast_fp16)[name = tensor("koi_49_cast_fp16")]; + tensor qo_25_axis_0 = const()[name = tensor("qo_25_axis_0"), val = tensor(-1)]; + tensor qo_25_cast_fp16 = stack(axis = qo_25_axis_0, values = (qor_49_cast_fp16, qoi_49_cast_fp16))[name = tensor("qo_25_cast_fp16")]; + tensor ko_25_axis_0 = const()[name = tensor("ko_25_axis_0"), val = tensor(-1)]; + tensor ko_25_cast_fp16 = stack(axis = ko_25_axis_0, values = (kor_49_cast_fp16, koi_49_cast_fp16))[name = tensor("ko_25_cast_fp16")]; + tensor var_5302 = const()[name = tensor("op_5302"), val = tensor([1, 256, 16, 64])]; + tensor q_75_cast_fp16 = reshape(shape = var_5302, x = qo_25_cast_fp16)[name = tensor("q_75_cast_fp16")]; + tensor var_5308 = const()[name = tensor("op_5308"), val = tensor([1, 256, 16, 64])]; + tensor k_51_cast_fp16 = reshape(shape = var_5308, x = ko_25_cast_fp16)[name = tensor("k_51_cast_fp16")]; + tensor write_indices_float_73_cast_fp16 = add(x = var_5197_cast_fp16, y = var_393_promoted_to_fp16)[name = tensor("write_indices_float_73_cast_fp16")]; + tensor _inversed_5329_y_0_to_fp16 = const()[name = tensor("_inversed_5329_y_0_to_fp16"), val = tensor(0x1p-9)]; + tensor _inversed_5329_cast_fp16 = mul(x = write_indices_float_73_cast_fp16, y = _inversed_5329_y_0_to_fp16)[name = tensor("_inversed_5329_cast_fp16")]; + tensor var_5330_cast_fp16 = floor(x = _inversed_5329_cast_fp16)[name = tensor("op_5330_cast_fp16")]; + tensor var_5331_to_fp16 = const()[name = tensor("op_5331_to_fp16"), val = tensor(0x1p+9)]; + tensor var_5332_cast_fp16 = mul(x = var_5330_cast_fp16, y = var_5331_to_fp16)[name = tensor("op_5332_cast_fp16")]; + tensor write_indices_float_75_cast_fp16 = sub(x = write_indices_float_73_cast_fp16, y = var_5332_cast_fp16)[name = tensor("write_indices_float_75_cast_fp16")]; + tensor var_5350_value_0_to_fp16 = const()[name = tensor("op_5350_value_0_to_fp16"), val = tensor(0x1.ffp+8)]; + tensor var_5350_cast_fp16 = fill_like(ref_tensor = write_indices_float_75_cast_fp16, value = var_5350_value_0_to_fp16)[name = tensor("op_5350_cast_fp16")]; + tensor write_indices_float_77_cast_fp16 = select(a = write_indices_float_75_cast_fp16, b = var_5350_cast_fp16, cond = token_valid_1_cast_fp16)[name = tensor("write_indices_float_77_cast_fp16")]; + tensor var_5356_dtype_0 = const()[name = tensor("op_5356_dtype_0"), val = tensor("int32")]; + tensor var_5361 = const()[name = tensor("op_5361"), val = tensor([1, 256, 1, 1])]; + tensor write_indices_float_77_cast_fp16_to_int32 = cast(dtype = var_5356_dtype_0, x = write_indices_float_77_cast_fp16)[name = tensor("cast_600")]; + tensor var_5362 = reshape(shape = var_5361, x = write_indices_float_77_cast_fp16_to_int32)[name = tensor("op_5362")]; + tensor write_indices_25_reps_0 = const()[name = tensor("write_indices_25_reps_0"), val = tensor([1, 1, 16, 64])]; + tensor write_indices_25 = tile(reps = write_indices_25_reps_0, x = var_5362)[name = tensor("write_indices_25")]; + tensor var_5372_begin_0 = const()[name = tensor("op_5372_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_5372_end_0 = const()[name = tensor("op_5372_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor var_5372_end_mask_0 = const()[name = tensor("op_5372_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_5372_squeeze_mask_0 = const()[name = tensor("op_5372_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor cache12_to_fp16_dtype_0 = const()[name = tensor("cache12_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor cache12_to_fp16 = cast(dtype = cache12_to_fp16_dtype_0, x = cache12)[name = tensor("cast_599")]; + tensor var_5372_cast_fp16 = slice_by_index(begin = var_5372_begin_0, end = var_5372_end_0, end_mask = var_5372_end_mask_0, squeeze_mask = var_5372_squeeze_mask_0, x = cache12_to_fp16)[name = tensor("op_5372_cast_fp16")]; + tensor var_5374_axis_0 = const()[name = tensor("op_5374_axis_0"), val = tensor(1)]; + tensor var_5374_mode_0 = const()[name = tensor("op_5374_mode_0"), val = tensor("update")]; + tensor var_5374_validate_indices_0 = const()[name = tensor("op_5374_validate_indices_0"), val = tensor(false)]; + tensor var_5374_cast_fp16 = scatter_along_axis(axis = var_5374_axis_0, data = var_5372_cast_fp16, indices = write_indices_25, mode = var_5374_mode_0, updates = k_51_cast_fp16, validate_indices = var_5374_validate_indices_0)[name = tensor("op_5374_cast_fp16")]; + tensor concat_98 = const()[name = tensor("concat_98"), val = tensor([0, 0, 0, 0, 0])]; + tensor concat_99 = const()[name = tensor("concat_99"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_25_internal_tensor_assign_1_stride_0 = const()[name = tensor("new_cache_25_internal_tensor_assign_1_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_25_internal_tensor_assign_1_begin_mask_0 = const()[name = tensor("new_cache_25_internal_tensor_assign_1_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_25_internal_tensor_assign_1_end_mask_0 = const()[name = tensor("new_cache_25_internal_tensor_assign_1_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_25_internal_tensor_assign_1_squeeze_mask_0 = const()[name = tensor("new_cache_25_internal_tensor_assign_1_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_70 = const()[name = tensor("shape_70"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_24 = const()[name = tensor("reduce_prod_24"), val = tensor(1048576)]; + tensor range_1d_24_start_0 = const()[name = tensor("range_1d_24_start_0"), val = tensor(0)]; + tensor range_1d_24_step_0 = const()[name = tensor("range_1d_24_step_0"), val = tensor(1)]; + tensor range_1d_24 = range_1d(end = reduce_prod_24, start = range_1d_24_start_0, step = range_1d_24_step_0)[name = tensor("range_1d_24")]; + tensor reshape_120 = reshape(shape = shape_70, x = range_1d_24)[name = tensor("reshape_120")]; + tensor slice_by_index_24 = slice_by_index(begin = concat_98, begin_mask = new_cache_25_internal_tensor_assign_1_begin_mask_0, end = concat_99, end_mask = new_cache_25_internal_tensor_assign_1_end_mask_0, squeeze_mask = new_cache_25_internal_tensor_assign_1_squeeze_mask_0, stride = new_cache_25_internal_tensor_assign_1_stride_0, x = reshape_120)[name = tensor("slice_by_index_24")]; + tensor reshape_121_shape_0 = const()[name = tensor("reshape_121_shape_0"), val = tensor([-1])]; + tensor reshape_121 = reshape(shape = reshape_121_shape_0, x = slice_by_index_24)[name = tensor("reshape_121")]; + tensor reshape_122_shape_0 = const()[name = tensor("reshape_122_shape_0"), val = tensor([-1])]; + tensor reshape_122_cast_fp16 = reshape(shape = reshape_122_shape_0, x = var_5374_cast_fp16)[name = tensor("reshape_122_cast_fp16")]; + tensor reshape_123_shape_0 = const()[name = tensor("reshape_123_shape_0"), val = tensor([-1])]; + tensor reshape_123_cast_fp16 = reshape(shape = reshape_123_shape_0, x = cache12_to_fp16)[name = tensor("reshape_123_cast_fp16")]; + tensor scatter_24_mode_0 = const()[name = tensor("scatter_24_mode_0"), val = tensor("update")]; + tensor scatter_24_axis_0 = const()[name = tensor("scatter_24_axis_0"), val = tensor(0)]; + tensor scatter_24_validate_indices_0 = const()[name = tensor("scatter_24_validate_indices_0"), val = tensor(false)]; + tensor scatter_24_cast_fp16 = scatter(axis = scatter_24_axis_0, data = reshape_123_cast_fp16, indices = reshape_121, mode = scatter_24_mode_0, updates = reshape_122_cast_fp16, validate_indices = scatter_24_validate_indices_0)[name = tensor("scatter_24_cast_fp16")]; + tensor reshape_124_cast_fp16 = reshape(shape = shape_70, x = scatter_24_cast_fp16)[name = tensor("reshape_124_cast_fp16")]; + tensor var_5382_begin_0 = const()[name = tensor("op_5382_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor var_5382_end_0 = const()[name = tensor("op_5382_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor var_5382_end_mask_0 = const()[name = tensor("op_5382_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_5382_squeeze_mask_0 = const()[name = tensor("op_5382_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor var_5382_cast_fp16 = slice_by_index(begin = var_5382_begin_0, end = var_5382_end_0, end_mask = var_5382_end_mask_0, squeeze_mask = var_5382_squeeze_mask_0, x = reshape_124_cast_fp16)[name = tensor("op_5382_cast_fp16")]; + tensor var_5384_axis_0 = const()[name = tensor("op_5384_axis_0"), val = tensor(1)]; + tensor var_5384_mode_0 = const()[name = tensor("op_5384_mode_0"), val = tensor("update")]; + tensor var_5384_validate_indices_0 = const()[name = tensor("op_5384_validate_indices_0"), val = tensor(false)]; + tensor var_5384_cast_fp16 = scatter_along_axis(axis = var_5384_axis_0, data = var_5382_cast_fp16, indices = write_indices_25, mode = var_5384_mode_0, updates = v_25_cast_fp16, validate_indices = var_5384_validate_indices_0)[name = tensor("op_5384_cast_fp16")]; + tensor concat_100 = const()[name = tensor("concat_100"), val = tensor([1, 0, 0, 0, 0])]; + tensor concat_101 = const()[name = tensor("concat_101"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_25_internal_tensor_assign_2_stride_0 = const()[name = tensor("new_cache_25_internal_tensor_assign_2_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_25_internal_tensor_assign_2_begin_mask_0 = const()[name = tensor("new_cache_25_internal_tensor_assign_2_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_25_internal_tensor_assign_2_end_mask_0 = const()[name = tensor("new_cache_25_internal_tensor_assign_2_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_25_internal_tensor_assign_2_squeeze_mask_0 = const()[name = tensor("new_cache_25_internal_tensor_assign_2_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_71 = const()[name = tensor("shape_71"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_25 = const()[name = tensor("reduce_prod_25"), val = tensor(1048576)]; + tensor range_1d_25_start_0 = const()[name = tensor("range_1d_25_start_0"), val = tensor(0)]; + tensor range_1d_25_step_0 = const()[name = tensor("range_1d_25_step_0"), val = tensor(1)]; + tensor range_1d_25 = range_1d(end = reduce_prod_25, start = range_1d_25_start_0, step = range_1d_25_step_0)[name = tensor("range_1d_25")]; + tensor reshape_125 = reshape(shape = shape_71, x = range_1d_25)[name = tensor("reshape_125")]; + tensor slice_by_index_25 = slice_by_index(begin = concat_100, begin_mask = new_cache_25_internal_tensor_assign_2_begin_mask_0, end = concat_101, end_mask = new_cache_25_internal_tensor_assign_2_end_mask_0, squeeze_mask = new_cache_25_internal_tensor_assign_2_squeeze_mask_0, stride = new_cache_25_internal_tensor_assign_2_stride_0, x = reshape_125)[name = tensor("slice_by_index_25")]; + tensor reshape_126_shape_0 = const()[name = tensor("reshape_126_shape_0"), val = tensor([-1])]; + tensor reshape_126 = reshape(shape = reshape_126_shape_0, x = slice_by_index_25)[name = tensor("reshape_126")]; + tensor reshape_127_shape_0 = const()[name = tensor("reshape_127_shape_0"), val = tensor([-1])]; + tensor reshape_127_cast_fp16 = reshape(shape = reshape_127_shape_0, x = var_5384_cast_fp16)[name = tensor("reshape_127_cast_fp16")]; + tensor reshape_128_shape_0 = const()[name = tensor("reshape_128_shape_0"), val = tensor([-1])]; + tensor reshape_128_cast_fp16 = reshape(shape = reshape_128_shape_0, x = reshape_124_cast_fp16)[name = tensor("reshape_128_cast_fp16")]; + tensor scatter_25_mode_0 = const()[name = tensor("scatter_25_mode_0"), val = tensor("update")]; + tensor scatter_25_axis_0 = const()[name = tensor("scatter_25_axis_0"), val = tensor(0)]; + tensor scatter_25_validate_indices_0 = const()[name = tensor("scatter_25_validate_indices_0"), val = tensor(false)]; + tensor scatter_25_cast_fp16 = scatter(axis = scatter_25_axis_0, data = reshape_128_cast_fp16, indices = reshape_126, mode = scatter_25_mode_0, updates = reshape_127_cast_fp16, validate_indices = scatter_25_validate_indices_0)[name = tensor("scatter_25_cast_fp16")]; + tensor reshape_129_cast_fp16 = reshape(shape = shape_71, x = scatter_25_cast_fp16)[name = tensor("reshape_129_cast_fp16")]; + tensor reshape_129_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("reshape_129_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor keys_73_begin_0 = const()[name = tensor("keys_73_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor keys_73_end_0 = const()[name = tensor("keys_73_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor keys_73_end_mask_0 = const()[name = tensor("keys_73_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor keys_73_squeeze_mask_0 = const()[name = tensor("keys_73_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor keys_73_cast_fp16 = slice_by_index(begin = keys_73_begin_0, end = keys_73_end_0, end_mask = keys_73_end_mask_0, squeeze_mask = keys_73_squeeze_mask_0, x = reshape_129_cast_fp16)[name = tensor("keys_73_cast_fp16")]; + tensor values_73_begin_0 = const()[name = tensor("values_73_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor values_73_end_0 = const()[name = tensor("values_73_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor values_73_end_mask_0 = const()[name = tensor("values_73_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor values_73_squeeze_mask_0 = const()[name = tensor("values_73_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor values_73_cast_fp16 = slice_by_index(begin = values_73_begin_0, end = values_73_end_0, end_mask = values_73_end_mask_0, squeeze_mask = values_73_squeeze_mask_0, x = reshape_129_cast_fp16)[name = tensor("values_73_cast_fp16")]; + tensor var_5396_cast_fp16 = not_equal(x = keys_73_cast_fp16, y = keys_73_cast_fp16)[name = tensor("op_5396_cast_fp16")]; + tensor keys_75_cast_fp16 = select(a = var_470_to_fp16, b = keys_73_cast_fp16, cond = var_5396_cast_fp16)[name = tensor("keys_75_cast_fp16")]; + tensor var_5404_cast_fp16 = not_equal(x = values_73_cast_fp16, y = values_73_cast_fp16)[name = tensor("op_5404_cast_fp16")]; + tensor values_75_cast_fp16 = select(a = var_470_to_fp16, b = values_73_cast_fp16, cond = var_5404_cast_fp16)[name = tensor("values_75_cast_fp16")]; + tensor var_5428 = const()[name = tensor("op_5428"), val = tensor([0, 2, 1, 3])]; + tensor var_5444 = const()[name = tensor("op_5444"), val = tensor([1, 1, 1])]; + tensor var_5445_cast_fp16 = reshape(shape = var_5444, x = position12_to_fp16)[name = tensor("op_5445_cast_fp16")]; + tensor q_positions_25_cast_fp16 = add(x = var_5445_cast_fp16, y = q_offsets_1_promoted_to_fp16)[name = tensor("q_positions_25_cast_fp16")]; + tensor valid_end_25_cast_fp16 = add(x = var_5445_cast_fp16, y = var_536_cast_fp16)[name = tensor("valid_end_25_cast_fp16")]; + tensor valid_mask_25_cast_fp16 = less(x = k_positions_1_promoted_to_fp16, y = valid_end_25_cast_fp16)[name = tensor("valid_mask_25_cast_fp16")]; + tensor causal_mask_25_cast_fp16 = less_equal(x = k_positions_1_promoted_to_fp16, y = q_positions_25_cast_fp16)[name = tensor("causal_mask_25_cast_fp16")]; + tensor var_5473 = logical_and(x = valid_mask_25_cast_fp16, y = causal_mask_25_cast_fp16)[name = tensor("op_5473")]; + tensor attn_mask_25_axes_0 = const()[name = tensor("attn_mask_25_axes_0"), val = tensor([1])]; + tensor attn_mask_25 = expand_dims(axes = attn_mask_25_axes_0, x = var_5473)[name = tensor("attn_mask_25")]; + tensor var_5487_transpose_x_0 = const()[name = tensor("op_5487_transpose_x_0"), val = tensor(false)]; + tensor var_5487_transpose_y_0 = const()[name = tensor("op_5487_transpose_y_0"), val = tensor(false)]; + tensor transpose_93_perm_0 = const()[name = tensor("transpose_93_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_94_perm_0 = const()[name = tensor("transpose_94_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_94 = transpose(perm = transpose_94_perm_0, x = keys_75_cast_fp16)[name = tensor("transpose_156")]; + tensor transpose_93 = transpose(perm = transpose_93_perm_0, x = q_75_cast_fp16)[name = tensor("transpose_157")]; + tensor var_5487_cast_fp16 = matmul(transpose_x = var_5487_transpose_x_0, transpose_y = var_5487_transpose_y_0, x = transpose_93, y = transpose_94)[name = tensor("op_5487_cast_fp16")]; + tensor var_5481_to_fp16 = const()[name = tensor("op_5481_to_fp16"), val = tensor([0x1p-3])]; + tensor attn_weights_73_cast_fp16 = mul(x = var_5487_cast_fp16, y = var_5481_to_fp16)[name = tensor("attn_weights_73_cast_fp16")]; + tensor var_5489 = logical_not(x = attn_mask_25)[name = tensor("op_5489")]; + tensor var_5490_to_fp16 = const()[name = tensor("op_5490_to_fp16"), val = tensor(-inf)]; + tensor attn_weights_75_cast_fp16 = select(a = var_5490_to_fp16, b = attn_weights_73_cast_fp16, cond = var_5489)[name = tensor("attn_weights_75_cast_fp16")]; + tensor var_5492 = const()[name = tensor("op_5492"), val = tensor(-1)]; + tensor attn_weights_77_cast_fp16 = softmax(axis = var_5492, x = attn_weights_75_cast_fp16)[name = tensor("attn_weights_77_cast_fp16")]; + tensor attn_output_25_transpose_x_0 = const()[name = tensor("attn_output_25_transpose_x_0"), val = tensor(false)]; + tensor attn_output_25_transpose_y_0 = const()[name = tensor("attn_output_25_transpose_y_0"), val = tensor(false)]; + tensor values_77_cast_fp16 = transpose(perm = var_5428, x = values_75_cast_fp16)[name = tensor("transpose_158")]; + tensor attn_output_25_cast_fp16 = matmul(transpose_x = attn_output_25_transpose_x_0, transpose_y = attn_output_25_transpose_y_0, x = attn_weights_77_cast_fp16, y = values_77_cast_fp16)[name = tensor("attn_output_25_cast_fp16")]; + tensor var_5500 = const()[name = tensor("op_5500"), val = tensor([0, 2, 1, 3])]; + tensor var_5505 = const()[name = tensor("op_5505"), val = tensor([1, 256, 1024])]; + tensor var_5501_cast_fp16 = transpose(perm = var_5500, x = attn_output_25_cast_fp16)[name = tensor("transpose_155")]; + tensor input_149_cast_fp16 = reshape(shape = var_5505, x = var_5501_cast_fp16)[name = tensor("input_149_cast_fp16")]; + tensor attn12_out_proj_weight_to_fp16 = const()[name = tensor("attn12_out_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(309459840)))]; + tensor linear_49_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = attn12_out_proj_weight_to_fp16, x = input_149_cast_fp16)[name = tensor("linear_49_cast_fp16")]; + tensor var_5514_cast_fp16 = add(x = position12_to_fp16, y = valid_len_to_fp16)[name = tensor("op_5514_cast_fp16")]; + tensor var_5514_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_5514_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor input_151_cast_fp16 = add(x = input_145_cast_fp16, y = linear_49_cast_fp16)[name = tensor("input_151_cast_fp16")]; + tensor input_153_axes_0 = const()[name = tensor("input_153_axes_0"), val = tensor([-1])]; + tensor norm12_2_weight_to_fp16 = const()[name = tensor("norm12_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(311557056)))]; + tensor norm12_2_bias_to_fp16 = const()[name = tensor("norm12_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(311559168)))]; + tensor var_5518_to_fp16 = const()[name = tensor("op_5518_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_153_cast_fp16 = layer_norm(axes = input_153_axes_0, beta = norm12_2_bias_to_fp16, epsilon = var_5518_to_fp16, gamma = norm12_2_weight_to_fp16, x = input_151_cast_fp16)[name = tensor("input_153_cast_fp16")]; + tensor linear12_1_weight_to_fp16 = const()[name = tensor("linear12_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(311561280)))]; + tensor linear_50_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = linear12_1_weight_to_fp16, x = input_153_cast_fp16)[name = tensor("linear_50_cast_fp16")]; + tensor input_155_mode_0 = const()[name = tensor("input_155_mode_0"), val = tensor("EXACT")]; + tensor input_155_cast_fp16 = gelu(mode = input_155_mode_0, x = linear_50_cast_fp16)[name = tensor("input_155_cast_fp16")]; + tensor linear12_2_weight_to_fp16 = const()[name = tensor("linear12_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(319949952)))]; + tensor linear_51_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = linear12_2_weight_to_fp16, x = input_155_cast_fp16)[name = tensor("linear_51_cast_fp16")]; + tensor input_157_cast_fp16 = add(x = input_151_cast_fp16, y = linear_51_cast_fp16)[name = tensor("input_157_cast_fp16")]; + tensor input_159_axes_0 = const()[name = tensor("input_159_axes_0"), val = tensor([-1])]; + tensor norm13_1_weight_to_fp16 = const()[name = tensor("norm13_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(328338624)))]; + tensor norm13_1_bias_to_fp16 = const()[name = tensor("norm13_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(328340736)))]; + tensor var_5535_to_fp16 = const()[name = tensor("op_5535_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_159_cast_fp16 = layer_norm(axes = input_159_axes_0, beta = norm13_1_bias_to_fp16, epsilon = var_5535_to_fp16, gamma = norm13_1_weight_to_fp16, x = input_157_cast_fp16)[name = tensor("input_159_cast_fp16")]; + tensor attn13_in_proj_weight_to_fp16 = const()[name = tensor("attn13_in_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(328342848)))]; + tensor linear_52_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = attn13_in_proj_weight_to_fp16, x = input_159_cast_fp16)[name = tensor("linear_52_cast_fp16")]; + tensor var_5549 = const()[name = tensor("op_5549"), val = tensor([1, 256, 3, 16, 64])]; + tensor qkv_27_cast_fp16 = reshape(shape = var_5549, x = linear_52_cast_fp16)[name = tensor("qkv_27_cast_fp16")]; + tensor q_79_begin_0 = const()[name = tensor("q_79_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor q_79_end_0 = const()[name = tensor("q_79_end_0"), val = tensor([1, 256, 1, 16, 64])]; + tensor q_79_end_mask_0 = const()[name = tensor("q_79_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor q_79_squeeze_mask_0 = const()[name = tensor("q_79_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor q_79_cast_fp16 = slice_by_index(begin = q_79_begin_0, end = q_79_end_0, end_mask = q_79_end_mask_0, squeeze_mask = q_79_squeeze_mask_0, x = qkv_27_cast_fp16)[name = tensor("q_79_cast_fp16")]; + tensor k_53_begin_0 = const()[name = tensor("k_53_begin_0"), val = tensor([0, 0, 1, 0, 0])]; + tensor k_53_end_0 = const()[name = tensor("k_53_end_0"), val = tensor([1, 256, 2, 16, 64])]; + tensor k_53_end_mask_0 = const()[name = tensor("k_53_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor k_53_squeeze_mask_0 = const()[name = tensor("k_53_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor k_53_cast_fp16 = slice_by_index(begin = k_53_begin_0, end = k_53_end_0, end_mask = k_53_end_mask_0, squeeze_mask = k_53_squeeze_mask_0, x = qkv_27_cast_fp16)[name = tensor("k_53_cast_fp16")]; + tensor v_27_begin_0 = const()[name = tensor("v_27_begin_0"), val = tensor([0, 0, 2, 0, 0])]; + tensor v_27_end_0 = const()[name = tensor("v_27_end_0"), val = tensor([1, 256, 3, 16, 64])]; + tensor v_27_end_mask_0 = const()[name = tensor("v_27_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor v_27_squeeze_mask_0 = const()[name = tensor("v_27_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor v_27_cast_fp16 = slice_by_index(begin = v_27_begin_0, end = v_27_end_0, end_mask = v_27_end_mask_0, squeeze_mask = v_27_squeeze_mask_0, x = qkv_27_cast_fp16)[name = tensor("v_27_cast_fp16")]; + tensor var_5607 = const()[name = tensor("op_5607"), val = tensor([1, 1])]; + tensor position13_to_fp16_dtype_0 = const()[name = tensor("position13_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor position13_to_fp16 = cast(dtype = position13_to_fp16_dtype_0, x = position13)[name = tensor("cast_596")]; + tensor var_5608_cast_fp16 = reshape(shape = var_5607, x = position13_to_fp16)[name = tensor("op_5608_cast_fp16")]; + tensor ts_81_cast_fp16 = add(x = ts_1_promoted_to_fp16, y = var_5608_cast_fp16)[name = tensor("ts_81_cast_fp16")]; + tensor var_5615 = const()[name = tensor("op_5615"), val = tensor([1, 256, 1, 1])]; + tensor ts_83_cast_fp16 = reshape(shape = var_5615, x = ts_81_cast_fp16)[name = tensor("ts_83_cast_fp16")]; + tensor var_5622 = const()[name = tensor("op_5622"), val = tensor([1, 256, 16, 32, 2])]; + tensor q_complex_27_cast_fp16 = reshape(shape = var_5622, x = q_79_cast_fp16)[name = tensor("q_complex_27_cast_fp16")]; + tensor var_5629 = const()[name = tensor("op_5629"), val = tensor([1, 256, 16, 32, 2])]; + tensor k_complex_27_cast_fp16 = reshape(shape = var_5629, x = k_53_cast_fp16)[name = tensor("k_complex_27_cast_fp16")]; + tensor var_5633_begin_0 = const()[name = tensor("op_5633_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_5633_end_0 = const()[name = tensor("op_5633_end_0"), val = tensor([1, 256, 16, 32, 1])]; + tensor var_5633_end_mask_0 = const()[name = tensor("op_5633_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_5633_squeeze_mask_0 = const()[name = tensor("op_5633_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_5633_cast_fp16 = slice_by_index(begin = var_5633_begin_0, end = var_5633_end_0, end_mask = var_5633_end_mask_0, squeeze_mask = var_5633_squeeze_mask_0, x = q_complex_27_cast_fp16)[name = tensor("op_5633_cast_fp16")]; + tensor var_5641_begin_0 = const()[name = tensor("op_5641_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_5641_end_0 = const()[name = tensor("op_5641_end_0"), val = tensor([1, 256, 16, 32, 2])]; + tensor var_5641_end_mask_0 = const()[name = tensor("op_5641_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_5641_squeeze_mask_0 = const()[name = tensor("op_5641_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_5641_cast_fp16 = slice_by_index(begin = var_5641_begin_0, end = var_5641_end_0, end_mask = var_5641_end_mask_0, squeeze_mask = var_5641_squeeze_mask_0, x = q_complex_27_cast_fp16)[name = tensor("op_5641_cast_fp16")]; + tensor var_5649_begin_0 = const()[name = tensor("op_5649_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_5649_end_0 = const()[name = tensor("op_5649_end_0"), val = tensor([1, 256, 16, 32, 1])]; + tensor var_5649_end_mask_0 = const()[name = tensor("op_5649_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_5649_squeeze_mask_0 = const()[name = tensor("op_5649_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_5649_cast_fp16 = slice_by_index(begin = var_5649_begin_0, end = var_5649_end_0, end_mask = var_5649_end_mask_0, squeeze_mask = var_5649_squeeze_mask_0, x = k_complex_27_cast_fp16)[name = tensor("op_5649_cast_fp16")]; + tensor var_5657_begin_0 = const()[name = tensor("op_5657_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_5657_end_0 = const()[name = tensor("op_5657_end_0"), val = tensor([1, 256, 16, 32, 2])]; + tensor var_5657_end_mask_0 = const()[name = tensor("op_5657_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_5657_squeeze_mask_0 = const()[name = tensor("op_5657_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_5657_cast_fp16 = slice_by_index(begin = var_5657_begin_0, end = var_5657_end_0, end_mask = var_5657_end_mask_0, squeeze_mask = var_5657_squeeze_mask_0, x = k_complex_27_cast_fp16)[name = tensor("op_5657_cast_fp16")]; + tensor freqs_27_to_fp16 = const()[name = tensor("freqs_27_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(334634368)))]; + tensor var_5663_cast_fp16 = mul(x = freqs_27_to_fp16, y = ts_83_cast_fp16)[name = tensor("op_5663_cast_fp16")]; + tensor rotr_27_cast_fp16 = cos(x = var_5663_cast_fp16)[name = tensor("rotr_27_cast_fp16")]; + tensor roti_27_cast_fp16 = sin(x = var_5663_cast_fp16)[name = tensor("roti_27_cast_fp16")]; + tensor var_5667_cast_fp16 = mul(x = var_5633_cast_fp16, y = rotr_27_cast_fp16)[name = tensor("op_5667_cast_fp16")]; + tensor var_5668_cast_fp16 = mul(x = var_5641_cast_fp16, y = roti_27_cast_fp16)[name = tensor("op_5668_cast_fp16")]; + tensor qor_53_cast_fp16 = sub(x = var_5667_cast_fp16, y = var_5668_cast_fp16)[name = tensor("qor_53_cast_fp16")]; + tensor var_5671_cast_fp16 = mul(x = var_5633_cast_fp16, y = roti_27_cast_fp16)[name = tensor("op_5671_cast_fp16")]; + tensor var_5672_cast_fp16 = mul(x = var_5641_cast_fp16, y = rotr_27_cast_fp16)[name = tensor("op_5672_cast_fp16")]; + tensor qoi_53_cast_fp16 = add(x = var_5671_cast_fp16, y = var_5672_cast_fp16)[name = tensor("qoi_53_cast_fp16")]; + tensor var_5675_cast_fp16 = mul(x = var_5649_cast_fp16, y = rotr_27_cast_fp16)[name = tensor("op_5675_cast_fp16")]; + tensor var_5676_cast_fp16 = mul(x = var_5657_cast_fp16, y = roti_27_cast_fp16)[name = tensor("op_5676_cast_fp16")]; + tensor kor_53_cast_fp16 = sub(x = var_5675_cast_fp16, y = var_5676_cast_fp16)[name = tensor("kor_53_cast_fp16")]; + tensor var_5679_cast_fp16 = mul(x = var_5649_cast_fp16, y = roti_27_cast_fp16)[name = tensor("op_5679_cast_fp16")]; + tensor var_5680_cast_fp16 = mul(x = var_5657_cast_fp16, y = rotr_27_cast_fp16)[name = tensor("op_5680_cast_fp16")]; + tensor koi_53_cast_fp16 = add(x = var_5679_cast_fp16, y = var_5680_cast_fp16)[name = tensor("koi_53_cast_fp16")]; + tensor qo_27_axis_0 = const()[name = tensor("qo_27_axis_0"), val = tensor(-1)]; + tensor qo_27_cast_fp16 = stack(axis = qo_27_axis_0, values = (qor_53_cast_fp16, qoi_53_cast_fp16))[name = tensor("qo_27_cast_fp16")]; + tensor ko_27_axis_0 = const()[name = tensor("ko_27_axis_0"), val = tensor(-1)]; + tensor ko_27_cast_fp16 = stack(axis = ko_27_axis_0, values = (kor_53_cast_fp16, koi_53_cast_fp16))[name = tensor("ko_27_cast_fp16")]; + tensor var_5713 = const()[name = tensor("op_5713"), val = tensor([1, 256, 16, 64])]; + tensor q_81_cast_fp16 = reshape(shape = var_5713, x = qo_27_cast_fp16)[name = tensor("q_81_cast_fp16")]; + tensor var_5719 = const()[name = tensor("op_5719"), val = tensor([1, 256, 16, 64])]; + tensor k_55_cast_fp16 = reshape(shape = var_5719, x = ko_27_cast_fp16)[name = tensor("k_55_cast_fp16")]; + tensor write_indices_float_79_cast_fp16 = add(x = var_5608_cast_fp16, y = var_393_promoted_to_fp16)[name = tensor("write_indices_float_79_cast_fp16")]; + tensor _inversed_5740_y_0_to_fp16 = const()[name = tensor("_inversed_5740_y_0_to_fp16"), val = tensor(0x1p-9)]; + tensor _inversed_5740_cast_fp16 = mul(x = write_indices_float_79_cast_fp16, y = _inversed_5740_y_0_to_fp16)[name = tensor("_inversed_5740_cast_fp16")]; + tensor var_5741_cast_fp16 = floor(x = _inversed_5740_cast_fp16)[name = tensor("op_5741_cast_fp16")]; + tensor var_5742_to_fp16 = const()[name = tensor("op_5742_to_fp16"), val = tensor(0x1p+9)]; + tensor var_5743_cast_fp16 = mul(x = var_5741_cast_fp16, y = var_5742_to_fp16)[name = tensor("op_5743_cast_fp16")]; + tensor write_indices_float_81_cast_fp16 = sub(x = write_indices_float_79_cast_fp16, y = var_5743_cast_fp16)[name = tensor("write_indices_float_81_cast_fp16")]; + tensor var_5761_value_0_to_fp16 = const()[name = tensor("op_5761_value_0_to_fp16"), val = tensor(0x1.ffp+8)]; + tensor var_5761_cast_fp16 = fill_like(ref_tensor = write_indices_float_81_cast_fp16, value = var_5761_value_0_to_fp16)[name = tensor("op_5761_cast_fp16")]; + tensor write_indices_float_83_cast_fp16 = select(a = write_indices_float_81_cast_fp16, b = var_5761_cast_fp16, cond = token_valid_1_cast_fp16)[name = tensor("write_indices_float_83_cast_fp16")]; + tensor var_5767_dtype_0 = const()[name = tensor("op_5767_dtype_0"), val = tensor("int32")]; + tensor var_5772 = const()[name = tensor("op_5772"), val = tensor([1, 256, 1, 1])]; + tensor write_indices_float_83_cast_fp16_to_int32 = cast(dtype = var_5767_dtype_0, x = write_indices_float_83_cast_fp16)[name = tensor("cast_595")]; + tensor var_5773 = reshape(shape = var_5772, x = write_indices_float_83_cast_fp16_to_int32)[name = tensor("op_5773")]; + tensor write_indices_27_reps_0 = const()[name = tensor("write_indices_27_reps_0"), val = tensor([1, 1, 16, 64])]; + tensor write_indices_27 = tile(reps = write_indices_27_reps_0, x = var_5773)[name = tensor("write_indices_27")]; + tensor var_5783_begin_0 = const()[name = tensor("op_5783_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_5783_end_0 = const()[name = tensor("op_5783_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor var_5783_end_mask_0 = const()[name = tensor("op_5783_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_5783_squeeze_mask_0 = const()[name = tensor("op_5783_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor cache13_to_fp16_dtype_0 = const()[name = tensor("cache13_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor cache13_to_fp16 = cast(dtype = cache13_to_fp16_dtype_0, x = cache13)[name = tensor("cast_594")]; + tensor var_5783_cast_fp16 = slice_by_index(begin = var_5783_begin_0, end = var_5783_end_0, end_mask = var_5783_end_mask_0, squeeze_mask = var_5783_squeeze_mask_0, x = cache13_to_fp16)[name = tensor("op_5783_cast_fp16")]; + tensor var_5785_axis_0 = const()[name = tensor("op_5785_axis_0"), val = tensor(1)]; + tensor var_5785_mode_0 = const()[name = tensor("op_5785_mode_0"), val = tensor("update")]; + tensor var_5785_validate_indices_0 = const()[name = tensor("op_5785_validate_indices_0"), val = tensor(false)]; + tensor var_5785_cast_fp16 = scatter_along_axis(axis = var_5785_axis_0, data = var_5783_cast_fp16, indices = write_indices_27, mode = var_5785_mode_0, updates = k_55_cast_fp16, validate_indices = var_5785_validate_indices_0)[name = tensor("op_5785_cast_fp16")]; + tensor concat_106 = const()[name = tensor("concat_106"), val = tensor([0, 0, 0, 0, 0])]; + tensor concat_107 = const()[name = tensor("concat_107"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_27_internal_tensor_assign_1_stride_0 = const()[name = tensor("new_cache_27_internal_tensor_assign_1_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_27_internal_tensor_assign_1_begin_mask_0 = const()[name = tensor("new_cache_27_internal_tensor_assign_1_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_27_internal_tensor_assign_1_end_mask_0 = const()[name = tensor("new_cache_27_internal_tensor_assign_1_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_27_internal_tensor_assign_1_squeeze_mask_0 = const()[name = tensor("new_cache_27_internal_tensor_assign_1_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_72 = const()[name = tensor("shape_72"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_26 = const()[name = tensor("reduce_prod_26"), val = tensor(1048576)]; + tensor range_1d_26_start_0 = const()[name = tensor("range_1d_26_start_0"), val = tensor(0)]; + tensor range_1d_26_step_0 = const()[name = tensor("range_1d_26_step_0"), val = tensor(1)]; + tensor range_1d_26 = range_1d(end = reduce_prod_26, start = range_1d_26_start_0, step = range_1d_26_step_0)[name = tensor("range_1d_26")]; + tensor reshape_130 = reshape(shape = shape_72, x = range_1d_26)[name = tensor("reshape_130")]; + tensor slice_by_index_26 = slice_by_index(begin = concat_106, begin_mask = new_cache_27_internal_tensor_assign_1_begin_mask_0, end = concat_107, end_mask = new_cache_27_internal_tensor_assign_1_end_mask_0, squeeze_mask = new_cache_27_internal_tensor_assign_1_squeeze_mask_0, stride = new_cache_27_internal_tensor_assign_1_stride_0, x = reshape_130)[name = tensor("slice_by_index_26")]; + tensor reshape_131_shape_0 = const()[name = tensor("reshape_131_shape_0"), val = tensor([-1])]; + tensor reshape_131 = reshape(shape = reshape_131_shape_0, x = slice_by_index_26)[name = tensor("reshape_131")]; + tensor reshape_132_shape_0 = const()[name = tensor("reshape_132_shape_0"), val = tensor([-1])]; + tensor reshape_132_cast_fp16 = reshape(shape = reshape_132_shape_0, x = var_5785_cast_fp16)[name = tensor("reshape_132_cast_fp16")]; + tensor reshape_133_shape_0 = const()[name = tensor("reshape_133_shape_0"), val = tensor([-1])]; + tensor reshape_133_cast_fp16 = reshape(shape = reshape_133_shape_0, x = cache13_to_fp16)[name = tensor("reshape_133_cast_fp16")]; + tensor scatter_26_mode_0 = const()[name = tensor("scatter_26_mode_0"), val = tensor("update")]; + tensor scatter_26_axis_0 = const()[name = tensor("scatter_26_axis_0"), val = tensor(0)]; + tensor scatter_26_validate_indices_0 = const()[name = tensor("scatter_26_validate_indices_0"), val = tensor(false)]; + tensor scatter_26_cast_fp16 = scatter(axis = scatter_26_axis_0, data = reshape_133_cast_fp16, indices = reshape_131, mode = scatter_26_mode_0, updates = reshape_132_cast_fp16, validate_indices = scatter_26_validate_indices_0)[name = tensor("scatter_26_cast_fp16")]; + tensor reshape_134_cast_fp16 = reshape(shape = shape_72, x = scatter_26_cast_fp16)[name = tensor("reshape_134_cast_fp16")]; + tensor var_5793_begin_0 = const()[name = tensor("op_5793_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor var_5793_end_0 = const()[name = tensor("op_5793_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor var_5793_end_mask_0 = const()[name = tensor("op_5793_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_5793_squeeze_mask_0 = const()[name = tensor("op_5793_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor var_5793_cast_fp16 = slice_by_index(begin = var_5793_begin_0, end = var_5793_end_0, end_mask = var_5793_end_mask_0, squeeze_mask = var_5793_squeeze_mask_0, x = reshape_134_cast_fp16)[name = tensor("op_5793_cast_fp16")]; + tensor var_5795_axis_0 = const()[name = tensor("op_5795_axis_0"), val = tensor(1)]; + tensor var_5795_mode_0 = const()[name = tensor("op_5795_mode_0"), val = tensor("update")]; + tensor var_5795_validate_indices_0 = const()[name = tensor("op_5795_validate_indices_0"), val = tensor(false)]; + tensor var_5795_cast_fp16 = scatter_along_axis(axis = var_5795_axis_0, data = var_5793_cast_fp16, indices = write_indices_27, mode = var_5795_mode_0, updates = v_27_cast_fp16, validate_indices = var_5795_validate_indices_0)[name = tensor("op_5795_cast_fp16")]; + tensor concat_108 = const()[name = tensor("concat_108"), val = tensor([1, 0, 0, 0, 0])]; + tensor concat_109 = const()[name = tensor("concat_109"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_27_internal_tensor_assign_2_stride_0 = const()[name = tensor("new_cache_27_internal_tensor_assign_2_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_27_internal_tensor_assign_2_begin_mask_0 = const()[name = tensor("new_cache_27_internal_tensor_assign_2_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_27_internal_tensor_assign_2_end_mask_0 = const()[name = tensor("new_cache_27_internal_tensor_assign_2_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_27_internal_tensor_assign_2_squeeze_mask_0 = const()[name = tensor("new_cache_27_internal_tensor_assign_2_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_73 = const()[name = tensor("shape_73"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_27 = const()[name = tensor("reduce_prod_27"), val = tensor(1048576)]; + tensor range_1d_27_start_0 = const()[name = tensor("range_1d_27_start_0"), val = tensor(0)]; + tensor range_1d_27_step_0 = const()[name = tensor("range_1d_27_step_0"), val = tensor(1)]; + tensor range_1d_27 = range_1d(end = reduce_prod_27, start = range_1d_27_start_0, step = range_1d_27_step_0)[name = tensor("range_1d_27")]; + tensor reshape_135 = reshape(shape = shape_73, x = range_1d_27)[name = tensor("reshape_135")]; + tensor slice_by_index_27 = slice_by_index(begin = concat_108, begin_mask = new_cache_27_internal_tensor_assign_2_begin_mask_0, end = concat_109, end_mask = new_cache_27_internal_tensor_assign_2_end_mask_0, squeeze_mask = new_cache_27_internal_tensor_assign_2_squeeze_mask_0, stride = new_cache_27_internal_tensor_assign_2_stride_0, x = reshape_135)[name = tensor("slice_by_index_27")]; + tensor reshape_136_shape_0 = const()[name = tensor("reshape_136_shape_0"), val = tensor([-1])]; + tensor reshape_136 = reshape(shape = reshape_136_shape_0, x = slice_by_index_27)[name = tensor("reshape_136")]; + tensor reshape_137_shape_0 = const()[name = tensor("reshape_137_shape_0"), val = tensor([-1])]; + tensor reshape_137_cast_fp16 = reshape(shape = reshape_137_shape_0, x = var_5795_cast_fp16)[name = tensor("reshape_137_cast_fp16")]; + tensor reshape_138_shape_0 = const()[name = tensor("reshape_138_shape_0"), val = tensor([-1])]; + tensor reshape_138_cast_fp16 = reshape(shape = reshape_138_shape_0, x = reshape_134_cast_fp16)[name = tensor("reshape_138_cast_fp16")]; + tensor scatter_27_mode_0 = const()[name = tensor("scatter_27_mode_0"), val = tensor("update")]; + tensor scatter_27_axis_0 = const()[name = tensor("scatter_27_axis_0"), val = tensor(0)]; + tensor scatter_27_validate_indices_0 = const()[name = tensor("scatter_27_validate_indices_0"), val = tensor(false)]; + tensor scatter_27_cast_fp16 = scatter(axis = scatter_27_axis_0, data = reshape_138_cast_fp16, indices = reshape_136, mode = scatter_27_mode_0, updates = reshape_137_cast_fp16, validate_indices = scatter_27_validate_indices_0)[name = tensor("scatter_27_cast_fp16")]; + tensor reshape_139_cast_fp16 = reshape(shape = shape_73, x = scatter_27_cast_fp16)[name = tensor("reshape_139_cast_fp16")]; + tensor reshape_139_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("reshape_139_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor keys_79_begin_0 = const()[name = tensor("keys_79_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor keys_79_end_0 = const()[name = tensor("keys_79_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor keys_79_end_mask_0 = const()[name = tensor("keys_79_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor keys_79_squeeze_mask_0 = const()[name = tensor("keys_79_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor keys_79_cast_fp16 = slice_by_index(begin = keys_79_begin_0, end = keys_79_end_0, end_mask = keys_79_end_mask_0, squeeze_mask = keys_79_squeeze_mask_0, x = reshape_139_cast_fp16)[name = tensor("keys_79_cast_fp16")]; + tensor values_79_begin_0 = const()[name = tensor("values_79_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor values_79_end_0 = const()[name = tensor("values_79_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor values_79_end_mask_0 = const()[name = tensor("values_79_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor values_79_squeeze_mask_0 = const()[name = tensor("values_79_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor values_79_cast_fp16 = slice_by_index(begin = values_79_begin_0, end = values_79_end_0, end_mask = values_79_end_mask_0, squeeze_mask = values_79_squeeze_mask_0, x = reshape_139_cast_fp16)[name = tensor("values_79_cast_fp16")]; + tensor var_5807_cast_fp16 = not_equal(x = keys_79_cast_fp16, y = keys_79_cast_fp16)[name = tensor("op_5807_cast_fp16")]; + tensor keys_81_cast_fp16 = select(a = var_470_to_fp16, b = keys_79_cast_fp16, cond = var_5807_cast_fp16)[name = tensor("keys_81_cast_fp16")]; + tensor var_5815_cast_fp16 = not_equal(x = values_79_cast_fp16, y = values_79_cast_fp16)[name = tensor("op_5815_cast_fp16")]; + tensor values_81_cast_fp16 = select(a = var_470_to_fp16, b = values_79_cast_fp16, cond = var_5815_cast_fp16)[name = tensor("values_81_cast_fp16")]; + tensor var_5839 = const()[name = tensor("op_5839"), val = tensor([0, 2, 1, 3])]; + tensor var_5855 = const()[name = tensor("op_5855"), val = tensor([1, 1, 1])]; + tensor var_5856_cast_fp16 = reshape(shape = var_5855, x = position13_to_fp16)[name = tensor("op_5856_cast_fp16")]; + tensor q_positions_27_cast_fp16 = add(x = var_5856_cast_fp16, y = q_offsets_1_promoted_to_fp16)[name = tensor("q_positions_27_cast_fp16")]; + tensor valid_end_27_cast_fp16 = add(x = var_5856_cast_fp16, y = var_536_cast_fp16)[name = tensor("valid_end_27_cast_fp16")]; + tensor valid_mask_27_cast_fp16 = less(x = k_positions_1_promoted_to_fp16, y = valid_end_27_cast_fp16)[name = tensor("valid_mask_27_cast_fp16")]; + tensor causal_mask_27_cast_fp16 = less_equal(x = k_positions_1_promoted_to_fp16, y = q_positions_27_cast_fp16)[name = tensor("causal_mask_27_cast_fp16")]; + tensor var_5884 = logical_and(x = valid_mask_27_cast_fp16, y = causal_mask_27_cast_fp16)[name = tensor("op_5884")]; + tensor attn_mask_27_axes_0 = const()[name = tensor("attn_mask_27_axes_0"), val = tensor([1])]; + tensor attn_mask_27 = expand_dims(axes = attn_mask_27_axes_0, x = var_5884)[name = tensor("attn_mask_27")]; + tensor var_5898_transpose_x_0 = const()[name = tensor("op_5898_transpose_x_0"), val = tensor(false)]; + tensor var_5898_transpose_y_0 = const()[name = tensor("op_5898_transpose_y_0"), val = tensor(false)]; + tensor transpose_95_perm_0 = const()[name = tensor("transpose_95_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_96_perm_0 = const()[name = tensor("transpose_96_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_96 = transpose(perm = transpose_96_perm_0, x = keys_81_cast_fp16)[name = tensor("transpose_152")]; + tensor transpose_95 = transpose(perm = transpose_95_perm_0, x = q_81_cast_fp16)[name = tensor("transpose_153")]; + tensor var_5898_cast_fp16 = matmul(transpose_x = var_5898_transpose_x_0, transpose_y = var_5898_transpose_y_0, x = transpose_95, y = transpose_96)[name = tensor("op_5898_cast_fp16")]; + tensor var_5892_to_fp16 = const()[name = tensor("op_5892_to_fp16"), val = tensor([0x1p-3])]; + tensor attn_weights_79_cast_fp16 = mul(x = var_5898_cast_fp16, y = var_5892_to_fp16)[name = tensor("attn_weights_79_cast_fp16")]; + tensor var_5900 = logical_not(x = attn_mask_27)[name = tensor("op_5900")]; + tensor var_5901_to_fp16 = const()[name = tensor("op_5901_to_fp16"), val = tensor(-inf)]; + tensor attn_weights_81_cast_fp16 = select(a = var_5901_to_fp16, b = attn_weights_79_cast_fp16, cond = var_5900)[name = tensor("attn_weights_81_cast_fp16")]; + tensor var_5903 = const()[name = tensor("op_5903"), val = tensor(-1)]; + tensor attn_weights_83_cast_fp16 = softmax(axis = var_5903, x = attn_weights_81_cast_fp16)[name = tensor("attn_weights_83_cast_fp16")]; + tensor attn_output_27_transpose_x_0 = const()[name = tensor("attn_output_27_transpose_x_0"), val = tensor(false)]; + tensor attn_output_27_transpose_y_0 = const()[name = tensor("attn_output_27_transpose_y_0"), val = tensor(false)]; + tensor values_83_cast_fp16 = transpose(perm = var_5839, x = values_81_cast_fp16)[name = tensor("transpose_154")]; + tensor attn_output_27_cast_fp16 = matmul(transpose_x = attn_output_27_transpose_x_0, transpose_y = attn_output_27_transpose_y_0, x = attn_weights_83_cast_fp16, y = values_83_cast_fp16)[name = tensor("attn_output_27_cast_fp16")]; + tensor var_5911 = const()[name = tensor("op_5911"), val = tensor([0, 2, 1, 3])]; + tensor var_5916 = const()[name = tensor("op_5916"), val = tensor([1, 256, 1024])]; + tensor var_5912_cast_fp16 = transpose(perm = var_5911, x = attn_output_27_cast_fp16)[name = tensor("transpose_151")]; + tensor input_161_cast_fp16 = reshape(shape = var_5916, x = var_5912_cast_fp16)[name = tensor("input_161_cast_fp16")]; + tensor attn13_out_proj_weight_to_fp16 = const()[name = tensor("attn13_out_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(334634496)))]; + tensor linear_53_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = attn13_out_proj_weight_to_fp16, x = input_161_cast_fp16)[name = tensor("linear_53_cast_fp16")]; + tensor var_5925_cast_fp16 = add(x = position13_to_fp16, y = valid_len_to_fp16)[name = tensor("op_5925_cast_fp16")]; + tensor var_5925_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_5925_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor input_163_cast_fp16 = add(x = input_157_cast_fp16, y = linear_53_cast_fp16)[name = tensor("input_163_cast_fp16")]; + tensor input_165_axes_0 = const()[name = tensor("input_165_axes_0"), val = tensor([-1])]; + tensor norm13_2_weight_to_fp16 = const()[name = tensor("norm13_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(336731712)))]; + tensor norm13_2_bias_to_fp16 = const()[name = tensor("norm13_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(336733824)))]; + tensor var_5929_to_fp16 = const()[name = tensor("op_5929_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_165_cast_fp16 = layer_norm(axes = input_165_axes_0, beta = norm13_2_bias_to_fp16, epsilon = var_5929_to_fp16, gamma = norm13_2_weight_to_fp16, x = input_163_cast_fp16)[name = tensor("input_165_cast_fp16")]; + tensor linear13_1_weight_to_fp16 = const()[name = tensor("linear13_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(336735936)))]; + tensor linear_54_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = linear13_1_weight_to_fp16, x = input_165_cast_fp16)[name = tensor("linear_54_cast_fp16")]; + tensor input_167_mode_0 = const()[name = tensor("input_167_mode_0"), val = tensor("EXACT")]; + tensor input_167_cast_fp16 = gelu(mode = input_167_mode_0, x = linear_54_cast_fp16)[name = tensor("input_167_cast_fp16")]; + tensor linear13_2_weight_to_fp16 = const()[name = tensor("linear13_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(345124608)))]; + tensor linear_55_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = linear13_2_weight_to_fp16, x = input_167_cast_fp16)[name = tensor("linear_55_cast_fp16")]; + tensor input_169_cast_fp16 = add(x = input_163_cast_fp16, y = linear_55_cast_fp16)[name = tensor("input_169_cast_fp16")]; + tensor input_171_axes_0 = const()[name = tensor("input_171_axes_0"), val = tensor([-1])]; + tensor norm14_1_weight_to_fp16 = const()[name = tensor("norm14_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(353513280)))]; + tensor norm14_1_bias_to_fp16 = const()[name = tensor("norm14_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(353515392)))]; + tensor var_5946_to_fp16 = const()[name = tensor("op_5946_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_171_cast_fp16 = layer_norm(axes = input_171_axes_0, beta = norm14_1_bias_to_fp16, epsilon = var_5946_to_fp16, gamma = norm14_1_weight_to_fp16, x = input_169_cast_fp16)[name = tensor("input_171_cast_fp16")]; + tensor attn14_in_proj_weight_to_fp16 = const()[name = tensor("attn14_in_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(353517504)))]; + tensor linear_56_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = attn14_in_proj_weight_to_fp16, x = input_171_cast_fp16)[name = tensor("linear_56_cast_fp16")]; + tensor var_5960 = const()[name = tensor("op_5960"), val = tensor([1, 256, 3, 16, 64])]; + tensor qkv_29_cast_fp16 = reshape(shape = var_5960, x = linear_56_cast_fp16)[name = tensor("qkv_29_cast_fp16")]; + tensor q_85_begin_0 = const()[name = tensor("q_85_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor q_85_end_0 = const()[name = tensor("q_85_end_0"), val = tensor([1, 256, 1, 16, 64])]; + tensor q_85_end_mask_0 = const()[name = tensor("q_85_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor q_85_squeeze_mask_0 = const()[name = tensor("q_85_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor q_85_cast_fp16 = slice_by_index(begin = q_85_begin_0, end = q_85_end_0, end_mask = q_85_end_mask_0, squeeze_mask = q_85_squeeze_mask_0, x = qkv_29_cast_fp16)[name = tensor("q_85_cast_fp16")]; + tensor k_57_begin_0 = const()[name = tensor("k_57_begin_0"), val = tensor([0, 0, 1, 0, 0])]; + tensor k_57_end_0 = const()[name = tensor("k_57_end_0"), val = tensor([1, 256, 2, 16, 64])]; + tensor k_57_end_mask_0 = const()[name = tensor("k_57_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor k_57_squeeze_mask_0 = const()[name = tensor("k_57_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor k_57_cast_fp16 = slice_by_index(begin = k_57_begin_0, end = k_57_end_0, end_mask = k_57_end_mask_0, squeeze_mask = k_57_squeeze_mask_0, x = qkv_29_cast_fp16)[name = tensor("k_57_cast_fp16")]; + tensor v_29_begin_0 = const()[name = tensor("v_29_begin_0"), val = tensor([0, 0, 2, 0, 0])]; + tensor v_29_end_0 = const()[name = tensor("v_29_end_0"), val = tensor([1, 256, 3, 16, 64])]; + tensor v_29_end_mask_0 = const()[name = tensor("v_29_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor v_29_squeeze_mask_0 = const()[name = tensor("v_29_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor v_29_cast_fp16 = slice_by_index(begin = v_29_begin_0, end = v_29_end_0, end_mask = v_29_end_mask_0, squeeze_mask = v_29_squeeze_mask_0, x = qkv_29_cast_fp16)[name = tensor("v_29_cast_fp16")]; + tensor var_6018 = const()[name = tensor("op_6018"), val = tensor([1, 1])]; + tensor position14_to_fp16_dtype_0 = const()[name = tensor("position14_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor position14_to_fp16 = cast(dtype = position14_to_fp16_dtype_0, x = position14)[name = tensor("cast_591")]; + tensor var_6019_cast_fp16 = reshape(shape = var_6018, x = position14_to_fp16)[name = tensor("op_6019_cast_fp16")]; + tensor ts_87_cast_fp16 = add(x = ts_1_promoted_to_fp16, y = var_6019_cast_fp16)[name = tensor("ts_87_cast_fp16")]; + tensor var_6026 = const()[name = tensor("op_6026"), val = tensor([1, 256, 1, 1])]; + tensor ts_89_cast_fp16 = reshape(shape = var_6026, x = ts_87_cast_fp16)[name = tensor("ts_89_cast_fp16")]; + tensor var_6033 = const()[name = tensor("op_6033"), val = tensor([1, 256, 16, 32, 2])]; + tensor q_complex_29_cast_fp16 = reshape(shape = var_6033, x = q_85_cast_fp16)[name = tensor("q_complex_29_cast_fp16")]; + tensor var_6040 = const()[name = tensor("op_6040"), val = tensor([1, 256, 16, 32, 2])]; + tensor k_complex_29_cast_fp16 = reshape(shape = var_6040, x = k_57_cast_fp16)[name = tensor("k_complex_29_cast_fp16")]; + tensor var_6044_begin_0 = const()[name = tensor("op_6044_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_6044_end_0 = const()[name = tensor("op_6044_end_0"), val = tensor([1, 256, 16, 32, 1])]; + tensor var_6044_end_mask_0 = const()[name = tensor("op_6044_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_6044_squeeze_mask_0 = const()[name = tensor("op_6044_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_6044_cast_fp16 = slice_by_index(begin = var_6044_begin_0, end = var_6044_end_0, end_mask = var_6044_end_mask_0, squeeze_mask = var_6044_squeeze_mask_0, x = q_complex_29_cast_fp16)[name = tensor("op_6044_cast_fp16")]; + tensor var_6052_begin_0 = const()[name = tensor("op_6052_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_6052_end_0 = const()[name = tensor("op_6052_end_0"), val = tensor([1, 256, 16, 32, 2])]; + tensor var_6052_end_mask_0 = const()[name = tensor("op_6052_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_6052_squeeze_mask_0 = const()[name = tensor("op_6052_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_6052_cast_fp16 = slice_by_index(begin = var_6052_begin_0, end = var_6052_end_0, end_mask = var_6052_end_mask_0, squeeze_mask = var_6052_squeeze_mask_0, x = q_complex_29_cast_fp16)[name = tensor("op_6052_cast_fp16")]; + tensor var_6060_begin_0 = const()[name = tensor("op_6060_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_6060_end_0 = const()[name = tensor("op_6060_end_0"), val = tensor([1, 256, 16, 32, 1])]; + tensor var_6060_end_mask_0 = const()[name = tensor("op_6060_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_6060_squeeze_mask_0 = const()[name = tensor("op_6060_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_6060_cast_fp16 = slice_by_index(begin = var_6060_begin_0, end = var_6060_end_0, end_mask = var_6060_end_mask_0, squeeze_mask = var_6060_squeeze_mask_0, x = k_complex_29_cast_fp16)[name = tensor("op_6060_cast_fp16")]; + tensor var_6068_begin_0 = const()[name = tensor("op_6068_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_6068_end_0 = const()[name = tensor("op_6068_end_0"), val = tensor([1, 256, 16, 32, 2])]; + tensor var_6068_end_mask_0 = const()[name = tensor("op_6068_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_6068_squeeze_mask_0 = const()[name = tensor("op_6068_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_6068_cast_fp16 = slice_by_index(begin = var_6068_begin_0, end = var_6068_end_0, end_mask = var_6068_end_mask_0, squeeze_mask = var_6068_squeeze_mask_0, x = k_complex_29_cast_fp16)[name = tensor("op_6068_cast_fp16")]; + tensor freqs_29_to_fp16 = const()[name = tensor("freqs_29_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(359809024)))]; + tensor var_6074_cast_fp16 = mul(x = freqs_29_to_fp16, y = ts_89_cast_fp16)[name = tensor("op_6074_cast_fp16")]; + tensor rotr_29_cast_fp16 = cos(x = var_6074_cast_fp16)[name = tensor("rotr_29_cast_fp16")]; + tensor roti_29_cast_fp16 = sin(x = var_6074_cast_fp16)[name = tensor("roti_29_cast_fp16")]; + tensor var_6078_cast_fp16 = mul(x = var_6044_cast_fp16, y = rotr_29_cast_fp16)[name = tensor("op_6078_cast_fp16")]; + tensor var_6079_cast_fp16 = mul(x = var_6052_cast_fp16, y = roti_29_cast_fp16)[name = tensor("op_6079_cast_fp16")]; + tensor qor_57_cast_fp16 = sub(x = var_6078_cast_fp16, y = var_6079_cast_fp16)[name = tensor("qor_57_cast_fp16")]; + tensor var_6082_cast_fp16 = mul(x = var_6044_cast_fp16, y = roti_29_cast_fp16)[name = tensor("op_6082_cast_fp16")]; + tensor var_6083_cast_fp16 = mul(x = var_6052_cast_fp16, y = rotr_29_cast_fp16)[name = tensor("op_6083_cast_fp16")]; + tensor qoi_57_cast_fp16 = add(x = var_6082_cast_fp16, y = var_6083_cast_fp16)[name = tensor("qoi_57_cast_fp16")]; + tensor var_6086_cast_fp16 = mul(x = var_6060_cast_fp16, y = rotr_29_cast_fp16)[name = tensor("op_6086_cast_fp16")]; + tensor var_6087_cast_fp16 = mul(x = var_6068_cast_fp16, y = roti_29_cast_fp16)[name = tensor("op_6087_cast_fp16")]; + tensor kor_57_cast_fp16 = sub(x = var_6086_cast_fp16, y = var_6087_cast_fp16)[name = tensor("kor_57_cast_fp16")]; + tensor var_6090_cast_fp16 = mul(x = var_6060_cast_fp16, y = roti_29_cast_fp16)[name = tensor("op_6090_cast_fp16")]; + tensor var_6091_cast_fp16 = mul(x = var_6068_cast_fp16, y = rotr_29_cast_fp16)[name = tensor("op_6091_cast_fp16")]; + tensor koi_57_cast_fp16 = add(x = var_6090_cast_fp16, y = var_6091_cast_fp16)[name = tensor("koi_57_cast_fp16")]; + tensor qo_29_axis_0 = const()[name = tensor("qo_29_axis_0"), val = tensor(-1)]; + tensor qo_29_cast_fp16 = stack(axis = qo_29_axis_0, values = (qor_57_cast_fp16, qoi_57_cast_fp16))[name = tensor("qo_29_cast_fp16")]; + tensor ko_29_axis_0 = const()[name = tensor("ko_29_axis_0"), val = tensor(-1)]; + tensor ko_29_cast_fp16 = stack(axis = ko_29_axis_0, values = (kor_57_cast_fp16, koi_57_cast_fp16))[name = tensor("ko_29_cast_fp16")]; + tensor var_6124 = const()[name = tensor("op_6124"), val = tensor([1, 256, 16, 64])]; + tensor q_87_cast_fp16 = reshape(shape = var_6124, x = qo_29_cast_fp16)[name = tensor("q_87_cast_fp16")]; + tensor var_6130 = const()[name = tensor("op_6130"), val = tensor([1, 256, 16, 64])]; + tensor k_59_cast_fp16 = reshape(shape = var_6130, x = ko_29_cast_fp16)[name = tensor("k_59_cast_fp16")]; + tensor write_indices_float_85_cast_fp16 = add(x = var_6019_cast_fp16, y = var_393_promoted_to_fp16)[name = tensor("write_indices_float_85_cast_fp16")]; + tensor _inversed_6151_y_0_to_fp16 = const()[name = tensor("_inversed_6151_y_0_to_fp16"), val = tensor(0x1p-9)]; + tensor _inversed_6151_cast_fp16 = mul(x = write_indices_float_85_cast_fp16, y = _inversed_6151_y_0_to_fp16)[name = tensor("_inversed_6151_cast_fp16")]; + tensor var_6152_cast_fp16 = floor(x = _inversed_6151_cast_fp16)[name = tensor("op_6152_cast_fp16")]; + tensor var_6153_to_fp16 = const()[name = tensor("op_6153_to_fp16"), val = tensor(0x1p+9)]; + tensor var_6154_cast_fp16 = mul(x = var_6152_cast_fp16, y = var_6153_to_fp16)[name = tensor("op_6154_cast_fp16")]; + tensor write_indices_float_87_cast_fp16 = sub(x = write_indices_float_85_cast_fp16, y = var_6154_cast_fp16)[name = tensor("write_indices_float_87_cast_fp16")]; + tensor var_6172_value_0_to_fp16 = const()[name = tensor("op_6172_value_0_to_fp16"), val = tensor(0x1.ffp+8)]; + tensor var_6172_cast_fp16 = fill_like(ref_tensor = write_indices_float_87_cast_fp16, value = var_6172_value_0_to_fp16)[name = tensor("op_6172_cast_fp16")]; + tensor write_indices_float_89_cast_fp16 = select(a = write_indices_float_87_cast_fp16, b = var_6172_cast_fp16, cond = token_valid_1_cast_fp16)[name = tensor("write_indices_float_89_cast_fp16")]; + tensor var_6178_dtype_0 = const()[name = tensor("op_6178_dtype_0"), val = tensor("int32")]; + tensor var_6183 = const()[name = tensor("op_6183"), val = tensor([1, 256, 1, 1])]; + tensor write_indices_float_89_cast_fp16_to_int32 = cast(dtype = var_6178_dtype_0, x = write_indices_float_89_cast_fp16)[name = tensor("cast_590")]; + tensor var_6184 = reshape(shape = var_6183, x = write_indices_float_89_cast_fp16_to_int32)[name = tensor("op_6184")]; + tensor write_indices_29_reps_0 = const()[name = tensor("write_indices_29_reps_0"), val = tensor([1, 1, 16, 64])]; + tensor write_indices_29 = tile(reps = write_indices_29_reps_0, x = var_6184)[name = tensor("write_indices_29")]; + tensor var_6194_begin_0 = const()[name = tensor("op_6194_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_6194_end_0 = const()[name = tensor("op_6194_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor var_6194_end_mask_0 = const()[name = tensor("op_6194_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_6194_squeeze_mask_0 = const()[name = tensor("op_6194_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor cache14_to_fp16_dtype_0 = const()[name = tensor("cache14_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor cache14_to_fp16 = cast(dtype = cache14_to_fp16_dtype_0, x = cache14)[name = tensor("cast_589")]; + tensor var_6194_cast_fp16 = slice_by_index(begin = var_6194_begin_0, end = var_6194_end_0, end_mask = var_6194_end_mask_0, squeeze_mask = var_6194_squeeze_mask_0, x = cache14_to_fp16)[name = tensor("op_6194_cast_fp16")]; + tensor var_6196_axis_0 = const()[name = tensor("op_6196_axis_0"), val = tensor(1)]; + tensor var_6196_mode_0 = const()[name = tensor("op_6196_mode_0"), val = tensor("update")]; + tensor var_6196_validate_indices_0 = const()[name = tensor("op_6196_validate_indices_0"), val = tensor(false)]; + tensor var_6196_cast_fp16 = scatter_along_axis(axis = var_6196_axis_0, data = var_6194_cast_fp16, indices = write_indices_29, mode = var_6196_mode_0, updates = k_59_cast_fp16, validate_indices = var_6196_validate_indices_0)[name = tensor("op_6196_cast_fp16")]; + tensor concat_114 = const()[name = tensor("concat_114"), val = tensor([0, 0, 0, 0, 0])]; + tensor concat_115 = const()[name = tensor("concat_115"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_29_internal_tensor_assign_1_stride_0 = const()[name = tensor("new_cache_29_internal_tensor_assign_1_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_29_internal_tensor_assign_1_begin_mask_0 = const()[name = tensor("new_cache_29_internal_tensor_assign_1_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_29_internal_tensor_assign_1_end_mask_0 = const()[name = tensor("new_cache_29_internal_tensor_assign_1_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_29_internal_tensor_assign_1_squeeze_mask_0 = const()[name = tensor("new_cache_29_internal_tensor_assign_1_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_74 = const()[name = tensor("shape_74"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_28 = const()[name = tensor("reduce_prod_28"), val = tensor(1048576)]; + tensor range_1d_28_start_0 = const()[name = tensor("range_1d_28_start_0"), val = tensor(0)]; + tensor range_1d_28_step_0 = const()[name = tensor("range_1d_28_step_0"), val = tensor(1)]; + tensor range_1d_28 = range_1d(end = reduce_prod_28, start = range_1d_28_start_0, step = range_1d_28_step_0)[name = tensor("range_1d_28")]; + tensor reshape_140 = reshape(shape = shape_74, x = range_1d_28)[name = tensor("reshape_140")]; + tensor slice_by_index_28 = slice_by_index(begin = concat_114, begin_mask = new_cache_29_internal_tensor_assign_1_begin_mask_0, end = concat_115, end_mask = new_cache_29_internal_tensor_assign_1_end_mask_0, squeeze_mask = new_cache_29_internal_tensor_assign_1_squeeze_mask_0, stride = new_cache_29_internal_tensor_assign_1_stride_0, x = reshape_140)[name = tensor("slice_by_index_28")]; + tensor reshape_141_shape_0 = const()[name = tensor("reshape_141_shape_0"), val = tensor([-1])]; + tensor reshape_141 = reshape(shape = reshape_141_shape_0, x = slice_by_index_28)[name = tensor("reshape_141")]; + tensor reshape_142_shape_0 = const()[name = tensor("reshape_142_shape_0"), val = tensor([-1])]; + tensor reshape_142_cast_fp16 = reshape(shape = reshape_142_shape_0, x = var_6196_cast_fp16)[name = tensor("reshape_142_cast_fp16")]; + tensor reshape_143_shape_0 = const()[name = tensor("reshape_143_shape_0"), val = tensor([-1])]; + tensor reshape_143_cast_fp16 = reshape(shape = reshape_143_shape_0, x = cache14_to_fp16)[name = tensor("reshape_143_cast_fp16")]; + tensor scatter_28_mode_0 = const()[name = tensor("scatter_28_mode_0"), val = tensor("update")]; + tensor scatter_28_axis_0 = const()[name = tensor("scatter_28_axis_0"), val = tensor(0)]; + tensor scatter_28_validate_indices_0 = const()[name = tensor("scatter_28_validate_indices_0"), val = tensor(false)]; + tensor scatter_28_cast_fp16 = scatter(axis = scatter_28_axis_0, data = reshape_143_cast_fp16, indices = reshape_141, mode = scatter_28_mode_0, updates = reshape_142_cast_fp16, validate_indices = scatter_28_validate_indices_0)[name = tensor("scatter_28_cast_fp16")]; + tensor reshape_144_cast_fp16 = reshape(shape = shape_74, x = scatter_28_cast_fp16)[name = tensor("reshape_144_cast_fp16")]; + tensor var_6204_begin_0 = const()[name = tensor("op_6204_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor var_6204_end_0 = const()[name = tensor("op_6204_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor var_6204_end_mask_0 = const()[name = tensor("op_6204_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_6204_squeeze_mask_0 = const()[name = tensor("op_6204_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor var_6204_cast_fp16 = slice_by_index(begin = var_6204_begin_0, end = var_6204_end_0, end_mask = var_6204_end_mask_0, squeeze_mask = var_6204_squeeze_mask_0, x = reshape_144_cast_fp16)[name = tensor("op_6204_cast_fp16")]; + tensor var_6206_axis_0 = const()[name = tensor("op_6206_axis_0"), val = tensor(1)]; + tensor var_6206_mode_0 = const()[name = tensor("op_6206_mode_0"), val = tensor("update")]; + tensor var_6206_validate_indices_0 = const()[name = tensor("op_6206_validate_indices_0"), val = tensor(false)]; + tensor var_6206_cast_fp16 = scatter_along_axis(axis = var_6206_axis_0, data = var_6204_cast_fp16, indices = write_indices_29, mode = var_6206_mode_0, updates = v_29_cast_fp16, validate_indices = var_6206_validate_indices_0)[name = tensor("op_6206_cast_fp16")]; + tensor concat_116 = const()[name = tensor("concat_116"), val = tensor([1, 0, 0, 0, 0])]; + tensor concat_117 = const()[name = tensor("concat_117"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_29_internal_tensor_assign_2_stride_0 = const()[name = tensor("new_cache_29_internal_tensor_assign_2_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_29_internal_tensor_assign_2_begin_mask_0 = const()[name = tensor("new_cache_29_internal_tensor_assign_2_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_29_internal_tensor_assign_2_end_mask_0 = const()[name = tensor("new_cache_29_internal_tensor_assign_2_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_29_internal_tensor_assign_2_squeeze_mask_0 = const()[name = tensor("new_cache_29_internal_tensor_assign_2_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_75 = const()[name = tensor("shape_75"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_29 = const()[name = tensor("reduce_prod_29"), val = tensor(1048576)]; + tensor range_1d_29_start_0 = const()[name = tensor("range_1d_29_start_0"), val = tensor(0)]; + tensor range_1d_29_step_0 = const()[name = tensor("range_1d_29_step_0"), val = tensor(1)]; + tensor range_1d_29 = range_1d(end = reduce_prod_29, start = range_1d_29_start_0, step = range_1d_29_step_0)[name = tensor("range_1d_29")]; + tensor reshape_145 = reshape(shape = shape_75, x = range_1d_29)[name = tensor("reshape_145")]; + tensor slice_by_index_29 = slice_by_index(begin = concat_116, begin_mask = new_cache_29_internal_tensor_assign_2_begin_mask_0, end = concat_117, end_mask = new_cache_29_internal_tensor_assign_2_end_mask_0, squeeze_mask = new_cache_29_internal_tensor_assign_2_squeeze_mask_0, stride = new_cache_29_internal_tensor_assign_2_stride_0, x = reshape_145)[name = tensor("slice_by_index_29")]; + tensor reshape_146_shape_0 = const()[name = tensor("reshape_146_shape_0"), val = tensor([-1])]; + tensor reshape_146 = reshape(shape = reshape_146_shape_0, x = slice_by_index_29)[name = tensor("reshape_146")]; + tensor reshape_147_shape_0 = const()[name = tensor("reshape_147_shape_0"), val = tensor([-1])]; + tensor reshape_147_cast_fp16 = reshape(shape = reshape_147_shape_0, x = var_6206_cast_fp16)[name = tensor("reshape_147_cast_fp16")]; + tensor reshape_148_shape_0 = const()[name = tensor("reshape_148_shape_0"), val = tensor([-1])]; + tensor reshape_148_cast_fp16 = reshape(shape = reshape_148_shape_0, x = reshape_144_cast_fp16)[name = tensor("reshape_148_cast_fp16")]; + tensor scatter_29_mode_0 = const()[name = tensor("scatter_29_mode_0"), val = tensor("update")]; + tensor scatter_29_axis_0 = const()[name = tensor("scatter_29_axis_0"), val = tensor(0)]; + tensor scatter_29_validate_indices_0 = const()[name = tensor("scatter_29_validate_indices_0"), val = tensor(false)]; + tensor scatter_29_cast_fp16 = scatter(axis = scatter_29_axis_0, data = reshape_148_cast_fp16, indices = reshape_146, mode = scatter_29_mode_0, updates = reshape_147_cast_fp16, validate_indices = scatter_29_validate_indices_0)[name = tensor("scatter_29_cast_fp16")]; + tensor reshape_149_cast_fp16 = reshape(shape = shape_75, x = scatter_29_cast_fp16)[name = tensor("reshape_149_cast_fp16")]; + tensor reshape_149_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("reshape_149_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor keys_85_begin_0 = const()[name = tensor("keys_85_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor keys_85_end_0 = const()[name = tensor("keys_85_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor keys_85_end_mask_0 = const()[name = tensor("keys_85_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor keys_85_squeeze_mask_0 = const()[name = tensor("keys_85_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor keys_85_cast_fp16 = slice_by_index(begin = keys_85_begin_0, end = keys_85_end_0, end_mask = keys_85_end_mask_0, squeeze_mask = keys_85_squeeze_mask_0, x = reshape_149_cast_fp16)[name = tensor("keys_85_cast_fp16")]; + tensor values_85_begin_0 = const()[name = tensor("values_85_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor values_85_end_0 = const()[name = tensor("values_85_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor values_85_end_mask_0 = const()[name = tensor("values_85_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor values_85_squeeze_mask_0 = const()[name = tensor("values_85_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor values_85_cast_fp16 = slice_by_index(begin = values_85_begin_0, end = values_85_end_0, end_mask = values_85_end_mask_0, squeeze_mask = values_85_squeeze_mask_0, x = reshape_149_cast_fp16)[name = tensor("values_85_cast_fp16")]; + tensor var_6218_cast_fp16 = not_equal(x = keys_85_cast_fp16, y = keys_85_cast_fp16)[name = tensor("op_6218_cast_fp16")]; + tensor keys_87_cast_fp16 = select(a = var_470_to_fp16, b = keys_85_cast_fp16, cond = var_6218_cast_fp16)[name = tensor("keys_87_cast_fp16")]; + tensor var_6226_cast_fp16 = not_equal(x = values_85_cast_fp16, y = values_85_cast_fp16)[name = tensor("op_6226_cast_fp16")]; + tensor values_87_cast_fp16 = select(a = var_470_to_fp16, b = values_85_cast_fp16, cond = var_6226_cast_fp16)[name = tensor("values_87_cast_fp16")]; + tensor var_6250 = const()[name = tensor("op_6250"), val = tensor([0, 2, 1, 3])]; + tensor var_6266 = const()[name = tensor("op_6266"), val = tensor([1, 1, 1])]; + tensor var_6267_cast_fp16 = reshape(shape = var_6266, x = position14_to_fp16)[name = tensor("op_6267_cast_fp16")]; + tensor q_positions_29_cast_fp16 = add(x = var_6267_cast_fp16, y = q_offsets_1_promoted_to_fp16)[name = tensor("q_positions_29_cast_fp16")]; + tensor valid_end_29_cast_fp16 = add(x = var_6267_cast_fp16, y = var_536_cast_fp16)[name = tensor("valid_end_29_cast_fp16")]; + tensor valid_mask_29_cast_fp16 = less(x = k_positions_1_promoted_to_fp16, y = valid_end_29_cast_fp16)[name = tensor("valid_mask_29_cast_fp16")]; + tensor causal_mask_29_cast_fp16 = less_equal(x = k_positions_1_promoted_to_fp16, y = q_positions_29_cast_fp16)[name = tensor("causal_mask_29_cast_fp16")]; + tensor var_6295 = logical_and(x = valid_mask_29_cast_fp16, y = causal_mask_29_cast_fp16)[name = tensor("op_6295")]; + tensor attn_mask_29_axes_0 = const()[name = tensor("attn_mask_29_axes_0"), val = tensor([1])]; + tensor attn_mask_29 = expand_dims(axes = attn_mask_29_axes_0, x = var_6295)[name = tensor("attn_mask_29")]; + tensor var_6309_transpose_x_0 = const()[name = tensor("op_6309_transpose_x_0"), val = tensor(false)]; + tensor var_6309_transpose_y_0 = const()[name = tensor("op_6309_transpose_y_0"), val = tensor(false)]; + tensor transpose_97_perm_0 = const()[name = tensor("transpose_97_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_98_perm_0 = const()[name = tensor("transpose_98_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_98 = transpose(perm = transpose_98_perm_0, x = keys_87_cast_fp16)[name = tensor("transpose_148")]; + tensor transpose_97 = transpose(perm = transpose_97_perm_0, x = q_87_cast_fp16)[name = tensor("transpose_149")]; + tensor var_6309_cast_fp16 = matmul(transpose_x = var_6309_transpose_x_0, transpose_y = var_6309_transpose_y_0, x = transpose_97, y = transpose_98)[name = tensor("op_6309_cast_fp16")]; + tensor var_6303_to_fp16 = const()[name = tensor("op_6303_to_fp16"), val = tensor([0x1p-3])]; + tensor attn_weights_85_cast_fp16 = mul(x = var_6309_cast_fp16, y = var_6303_to_fp16)[name = tensor("attn_weights_85_cast_fp16")]; + tensor var_6311 = logical_not(x = attn_mask_29)[name = tensor("op_6311")]; + tensor var_6312_to_fp16 = const()[name = tensor("op_6312_to_fp16"), val = tensor(-inf)]; + tensor attn_weights_87_cast_fp16 = select(a = var_6312_to_fp16, b = attn_weights_85_cast_fp16, cond = var_6311)[name = tensor("attn_weights_87_cast_fp16")]; + tensor var_6314 = const()[name = tensor("op_6314"), val = tensor(-1)]; + tensor attn_weights_89_cast_fp16 = softmax(axis = var_6314, x = attn_weights_87_cast_fp16)[name = tensor("attn_weights_89_cast_fp16")]; + tensor attn_output_29_transpose_x_0 = const()[name = tensor("attn_output_29_transpose_x_0"), val = tensor(false)]; + tensor attn_output_29_transpose_y_0 = const()[name = tensor("attn_output_29_transpose_y_0"), val = tensor(false)]; + tensor values_89_cast_fp16 = transpose(perm = var_6250, x = values_87_cast_fp16)[name = tensor("transpose_150")]; + tensor attn_output_29_cast_fp16 = matmul(transpose_x = attn_output_29_transpose_x_0, transpose_y = attn_output_29_transpose_y_0, x = attn_weights_89_cast_fp16, y = values_89_cast_fp16)[name = tensor("attn_output_29_cast_fp16")]; + tensor var_6322 = const()[name = tensor("op_6322"), val = tensor([0, 2, 1, 3])]; + tensor var_6327 = const()[name = tensor("op_6327"), val = tensor([1, 256, 1024])]; + tensor var_6323_cast_fp16 = transpose(perm = var_6322, x = attn_output_29_cast_fp16)[name = tensor("transpose_147")]; + tensor input_173_cast_fp16 = reshape(shape = var_6327, x = var_6323_cast_fp16)[name = tensor("input_173_cast_fp16")]; + tensor attn14_out_proj_weight_to_fp16 = const()[name = tensor("attn14_out_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(359809152)))]; + tensor linear_57_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = attn14_out_proj_weight_to_fp16, x = input_173_cast_fp16)[name = tensor("linear_57_cast_fp16")]; + tensor var_6336_cast_fp16 = add(x = position14_to_fp16, y = valid_len_to_fp16)[name = tensor("op_6336_cast_fp16")]; + tensor var_6336_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_6336_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor input_175_cast_fp16 = add(x = input_169_cast_fp16, y = linear_57_cast_fp16)[name = tensor("input_175_cast_fp16")]; + tensor input_177_axes_0 = const()[name = tensor("input_177_axes_0"), val = tensor([-1])]; + tensor norm14_2_weight_to_fp16 = const()[name = tensor("norm14_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(361906368)))]; + tensor norm14_2_bias_to_fp16 = const()[name = tensor("norm14_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(361908480)))]; + tensor var_6340_to_fp16 = const()[name = tensor("op_6340_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_177_cast_fp16 = layer_norm(axes = input_177_axes_0, beta = norm14_2_bias_to_fp16, epsilon = var_6340_to_fp16, gamma = norm14_2_weight_to_fp16, x = input_175_cast_fp16)[name = tensor("input_177_cast_fp16")]; + tensor linear14_1_weight_to_fp16 = const()[name = tensor("linear14_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(361910592)))]; + tensor linear_58_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = linear14_1_weight_to_fp16, x = input_177_cast_fp16)[name = tensor("linear_58_cast_fp16")]; + tensor input_179_mode_0 = const()[name = tensor("input_179_mode_0"), val = tensor("EXACT")]; + tensor input_179_cast_fp16 = gelu(mode = input_179_mode_0, x = linear_58_cast_fp16)[name = tensor("input_179_cast_fp16")]; + tensor linear14_2_weight_to_fp16 = const()[name = tensor("linear14_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(370299264)))]; + tensor linear_59_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = linear14_2_weight_to_fp16, x = input_179_cast_fp16)[name = tensor("linear_59_cast_fp16")]; + tensor input_181_cast_fp16 = add(x = input_175_cast_fp16, y = linear_59_cast_fp16)[name = tensor("input_181_cast_fp16")]; + tensor input_183_axes_0 = const()[name = tensor("input_183_axes_0"), val = tensor([-1])]; + tensor norm15_1_weight_to_fp16 = const()[name = tensor("norm15_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(378687936)))]; + tensor norm15_1_bias_to_fp16 = const()[name = tensor("norm15_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(378690048)))]; + tensor var_6357_to_fp16 = const()[name = tensor("op_6357_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_183_cast_fp16 = layer_norm(axes = input_183_axes_0, beta = norm15_1_bias_to_fp16, epsilon = var_6357_to_fp16, gamma = norm15_1_weight_to_fp16, x = input_181_cast_fp16)[name = tensor("input_183_cast_fp16")]; + tensor attn15_in_proj_weight_to_fp16 = const()[name = tensor("attn15_in_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(378692160)))]; + tensor linear_60_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = attn15_in_proj_weight_to_fp16, x = input_183_cast_fp16)[name = tensor("linear_60_cast_fp16")]; + tensor var_6371 = const()[name = tensor("op_6371"), val = tensor([1, 256, 3, 16, 64])]; + tensor qkv_31_cast_fp16 = reshape(shape = var_6371, x = linear_60_cast_fp16)[name = tensor("qkv_31_cast_fp16")]; + tensor q_91_begin_0 = const()[name = tensor("q_91_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor q_91_end_0 = const()[name = tensor("q_91_end_0"), val = tensor([1, 256, 1, 16, 64])]; + tensor q_91_end_mask_0 = const()[name = tensor("q_91_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor q_91_squeeze_mask_0 = const()[name = tensor("q_91_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor q_91_cast_fp16 = slice_by_index(begin = q_91_begin_0, end = q_91_end_0, end_mask = q_91_end_mask_0, squeeze_mask = q_91_squeeze_mask_0, x = qkv_31_cast_fp16)[name = tensor("q_91_cast_fp16")]; + tensor k_61_begin_0 = const()[name = tensor("k_61_begin_0"), val = tensor([0, 0, 1, 0, 0])]; + tensor k_61_end_0 = const()[name = tensor("k_61_end_0"), val = tensor([1, 256, 2, 16, 64])]; + tensor k_61_end_mask_0 = const()[name = tensor("k_61_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor k_61_squeeze_mask_0 = const()[name = tensor("k_61_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor k_61_cast_fp16 = slice_by_index(begin = k_61_begin_0, end = k_61_end_0, end_mask = k_61_end_mask_0, squeeze_mask = k_61_squeeze_mask_0, x = qkv_31_cast_fp16)[name = tensor("k_61_cast_fp16")]; + tensor v_31_begin_0 = const()[name = tensor("v_31_begin_0"), val = tensor([0, 0, 2, 0, 0])]; + tensor v_31_end_0 = const()[name = tensor("v_31_end_0"), val = tensor([1, 256, 3, 16, 64])]; + tensor v_31_end_mask_0 = const()[name = tensor("v_31_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor v_31_squeeze_mask_0 = const()[name = tensor("v_31_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor v_31_cast_fp16 = slice_by_index(begin = v_31_begin_0, end = v_31_end_0, end_mask = v_31_end_mask_0, squeeze_mask = v_31_squeeze_mask_0, x = qkv_31_cast_fp16)[name = tensor("v_31_cast_fp16")]; + tensor var_6429 = const()[name = tensor("op_6429"), val = tensor([1, 1])]; + tensor position15_to_fp16_dtype_0 = const()[name = tensor("position15_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor position15_to_fp16 = cast(dtype = position15_to_fp16_dtype_0, x = position15)[name = tensor("cast_586")]; + tensor var_6430_cast_fp16 = reshape(shape = var_6429, x = position15_to_fp16)[name = tensor("op_6430_cast_fp16")]; + tensor ts_93_cast_fp16 = add(x = ts_1_promoted_to_fp16, y = var_6430_cast_fp16)[name = tensor("ts_93_cast_fp16")]; + tensor var_6437 = const()[name = tensor("op_6437"), val = tensor([1, 256, 1, 1])]; + tensor ts_95_cast_fp16 = reshape(shape = var_6437, x = ts_93_cast_fp16)[name = tensor("ts_95_cast_fp16")]; + tensor var_6444 = const()[name = tensor("op_6444"), val = tensor([1, 256, 16, 32, 2])]; + tensor q_complex_31_cast_fp16 = reshape(shape = var_6444, x = q_91_cast_fp16)[name = tensor("q_complex_31_cast_fp16")]; + tensor var_6451 = const()[name = tensor("op_6451"), val = tensor([1, 256, 16, 32, 2])]; + tensor k_complex_31_cast_fp16 = reshape(shape = var_6451, x = k_61_cast_fp16)[name = tensor("k_complex_31_cast_fp16")]; + tensor var_6455_begin_0 = const()[name = tensor("op_6455_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_6455_end_0 = const()[name = tensor("op_6455_end_0"), val = tensor([1, 256, 16, 32, 1])]; + tensor var_6455_end_mask_0 = const()[name = tensor("op_6455_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_6455_squeeze_mask_0 = const()[name = tensor("op_6455_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_6455_cast_fp16 = slice_by_index(begin = var_6455_begin_0, end = var_6455_end_0, end_mask = var_6455_end_mask_0, squeeze_mask = var_6455_squeeze_mask_0, x = q_complex_31_cast_fp16)[name = tensor("op_6455_cast_fp16")]; + tensor var_6463_begin_0 = const()[name = tensor("op_6463_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_6463_end_0 = const()[name = tensor("op_6463_end_0"), val = tensor([1, 256, 16, 32, 2])]; + tensor var_6463_end_mask_0 = const()[name = tensor("op_6463_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_6463_squeeze_mask_0 = const()[name = tensor("op_6463_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_6463_cast_fp16 = slice_by_index(begin = var_6463_begin_0, end = var_6463_end_0, end_mask = var_6463_end_mask_0, squeeze_mask = var_6463_squeeze_mask_0, x = q_complex_31_cast_fp16)[name = tensor("op_6463_cast_fp16")]; + tensor var_6471_begin_0 = const()[name = tensor("op_6471_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_6471_end_0 = const()[name = tensor("op_6471_end_0"), val = tensor([1, 256, 16, 32, 1])]; + tensor var_6471_end_mask_0 = const()[name = tensor("op_6471_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_6471_squeeze_mask_0 = const()[name = tensor("op_6471_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_6471_cast_fp16 = slice_by_index(begin = var_6471_begin_0, end = var_6471_end_0, end_mask = var_6471_end_mask_0, squeeze_mask = var_6471_squeeze_mask_0, x = k_complex_31_cast_fp16)[name = tensor("op_6471_cast_fp16")]; + tensor var_6479_begin_0 = const()[name = tensor("op_6479_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_6479_end_0 = const()[name = tensor("op_6479_end_0"), val = tensor([1, 256, 16, 32, 2])]; + tensor var_6479_end_mask_0 = const()[name = tensor("op_6479_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_6479_squeeze_mask_0 = const()[name = tensor("op_6479_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_6479_cast_fp16 = slice_by_index(begin = var_6479_begin_0, end = var_6479_end_0, end_mask = var_6479_end_mask_0, squeeze_mask = var_6479_squeeze_mask_0, x = k_complex_31_cast_fp16)[name = tensor("op_6479_cast_fp16")]; + tensor freqs_31_to_fp16 = const()[name = tensor("freqs_31_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(384983680)))]; + tensor var_6485_cast_fp16 = mul(x = freqs_31_to_fp16, y = ts_95_cast_fp16)[name = tensor("op_6485_cast_fp16")]; + tensor rotr_31_cast_fp16 = cos(x = var_6485_cast_fp16)[name = tensor("rotr_31_cast_fp16")]; + tensor roti_31_cast_fp16 = sin(x = var_6485_cast_fp16)[name = tensor("roti_31_cast_fp16")]; + tensor var_6489_cast_fp16 = mul(x = var_6455_cast_fp16, y = rotr_31_cast_fp16)[name = tensor("op_6489_cast_fp16")]; + tensor var_6490_cast_fp16 = mul(x = var_6463_cast_fp16, y = roti_31_cast_fp16)[name = tensor("op_6490_cast_fp16")]; + tensor qor_61_cast_fp16 = sub(x = var_6489_cast_fp16, y = var_6490_cast_fp16)[name = tensor("qor_61_cast_fp16")]; + tensor var_6493_cast_fp16 = mul(x = var_6455_cast_fp16, y = roti_31_cast_fp16)[name = tensor("op_6493_cast_fp16")]; + tensor var_6494_cast_fp16 = mul(x = var_6463_cast_fp16, y = rotr_31_cast_fp16)[name = tensor("op_6494_cast_fp16")]; + tensor qoi_61_cast_fp16 = add(x = var_6493_cast_fp16, y = var_6494_cast_fp16)[name = tensor("qoi_61_cast_fp16")]; + tensor var_6497_cast_fp16 = mul(x = var_6471_cast_fp16, y = rotr_31_cast_fp16)[name = tensor("op_6497_cast_fp16")]; + tensor var_6498_cast_fp16 = mul(x = var_6479_cast_fp16, y = roti_31_cast_fp16)[name = tensor("op_6498_cast_fp16")]; + tensor kor_61_cast_fp16 = sub(x = var_6497_cast_fp16, y = var_6498_cast_fp16)[name = tensor("kor_61_cast_fp16")]; + tensor var_6501_cast_fp16 = mul(x = var_6471_cast_fp16, y = roti_31_cast_fp16)[name = tensor("op_6501_cast_fp16")]; + tensor var_6502_cast_fp16 = mul(x = var_6479_cast_fp16, y = rotr_31_cast_fp16)[name = tensor("op_6502_cast_fp16")]; + tensor koi_61_cast_fp16 = add(x = var_6501_cast_fp16, y = var_6502_cast_fp16)[name = tensor("koi_61_cast_fp16")]; + tensor qo_31_axis_0 = const()[name = tensor("qo_31_axis_0"), val = tensor(-1)]; + tensor qo_31_cast_fp16 = stack(axis = qo_31_axis_0, values = (qor_61_cast_fp16, qoi_61_cast_fp16))[name = tensor("qo_31_cast_fp16")]; + tensor ko_31_axis_0 = const()[name = tensor("ko_31_axis_0"), val = tensor(-1)]; + tensor ko_31_cast_fp16 = stack(axis = ko_31_axis_0, values = (kor_61_cast_fp16, koi_61_cast_fp16))[name = tensor("ko_31_cast_fp16")]; + tensor var_6535 = const()[name = tensor("op_6535"), val = tensor([1, 256, 16, 64])]; + tensor q_93_cast_fp16 = reshape(shape = var_6535, x = qo_31_cast_fp16)[name = tensor("q_93_cast_fp16")]; + tensor var_6541 = const()[name = tensor("op_6541"), val = tensor([1, 256, 16, 64])]; + tensor k_63_cast_fp16 = reshape(shape = var_6541, x = ko_31_cast_fp16)[name = tensor("k_63_cast_fp16")]; + tensor write_indices_float_91_cast_fp16 = add(x = var_6430_cast_fp16, y = var_393_promoted_to_fp16)[name = tensor("write_indices_float_91_cast_fp16")]; + tensor _inversed_6562_y_0_to_fp16 = const()[name = tensor("_inversed_6562_y_0_to_fp16"), val = tensor(0x1p-9)]; + tensor _inversed_6562_cast_fp16 = mul(x = write_indices_float_91_cast_fp16, y = _inversed_6562_y_0_to_fp16)[name = tensor("_inversed_6562_cast_fp16")]; + tensor var_6563_cast_fp16 = floor(x = _inversed_6562_cast_fp16)[name = tensor("op_6563_cast_fp16")]; + tensor var_6564_to_fp16 = const()[name = tensor("op_6564_to_fp16"), val = tensor(0x1p+9)]; + tensor var_6565_cast_fp16 = mul(x = var_6563_cast_fp16, y = var_6564_to_fp16)[name = tensor("op_6565_cast_fp16")]; + tensor write_indices_float_93_cast_fp16 = sub(x = write_indices_float_91_cast_fp16, y = var_6565_cast_fp16)[name = tensor("write_indices_float_93_cast_fp16")]; + tensor var_6583_value_0_to_fp16 = const()[name = tensor("op_6583_value_0_to_fp16"), val = tensor(0x1.ffp+8)]; + tensor var_6583_cast_fp16 = fill_like(ref_tensor = write_indices_float_93_cast_fp16, value = var_6583_value_0_to_fp16)[name = tensor("op_6583_cast_fp16")]; + tensor write_indices_float_95_cast_fp16 = select(a = write_indices_float_93_cast_fp16, b = var_6583_cast_fp16, cond = token_valid_1_cast_fp16)[name = tensor("write_indices_float_95_cast_fp16")]; + tensor var_6589_dtype_0 = const()[name = tensor("op_6589_dtype_0"), val = tensor("int32")]; + tensor var_6594 = const()[name = tensor("op_6594"), val = tensor([1, 256, 1, 1])]; + tensor write_indices_float_95_cast_fp16_to_int32 = cast(dtype = var_6589_dtype_0, x = write_indices_float_95_cast_fp16)[name = tensor("cast_585")]; + tensor var_6595 = reshape(shape = var_6594, x = write_indices_float_95_cast_fp16_to_int32)[name = tensor("op_6595")]; + tensor write_indices_31_reps_0 = const()[name = tensor("write_indices_31_reps_0"), val = tensor([1, 1, 16, 64])]; + tensor write_indices_31 = tile(reps = write_indices_31_reps_0, x = var_6595)[name = tensor("write_indices_31")]; + tensor var_6605_begin_0 = const()[name = tensor("op_6605_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_6605_end_0 = const()[name = tensor("op_6605_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor var_6605_end_mask_0 = const()[name = tensor("op_6605_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_6605_squeeze_mask_0 = const()[name = tensor("op_6605_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor cache15_to_fp16_dtype_0 = const()[name = tensor("cache15_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor cache15_to_fp16 = cast(dtype = cache15_to_fp16_dtype_0, x = cache15)[name = tensor("cast_584")]; + tensor var_6605_cast_fp16 = slice_by_index(begin = var_6605_begin_0, end = var_6605_end_0, end_mask = var_6605_end_mask_0, squeeze_mask = var_6605_squeeze_mask_0, x = cache15_to_fp16)[name = tensor("op_6605_cast_fp16")]; + tensor var_6607_axis_0 = const()[name = tensor("op_6607_axis_0"), val = tensor(1)]; + tensor var_6607_mode_0 = const()[name = tensor("op_6607_mode_0"), val = tensor("update")]; + tensor var_6607_validate_indices_0 = const()[name = tensor("op_6607_validate_indices_0"), val = tensor(false)]; + tensor var_6607_cast_fp16 = scatter_along_axis(axis = var_6607_axis_0, data = var_6605_cast_fp16, indices = write_indices_31, mode = var_6607_mode_0, updates = k_63_cast_fp16, validate_indices = var_6607_validate_indices_0)[name = tensor("op_6607_cast_fp16")]; + tensor concat_122 = const()[name = tensor("concat_122"), val = tensor([0, 0, 0, 0, 0])]; + tensor concat_123 = const()[name = tensor("concat_123"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_31_internal_tensor_assign_1_stride_0 = const()[name = tensor("new_cache_31_internal_tensor_assign_1_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_31_internal_tensor_assign_1_begin_mask_0 = const()[name = tensor("new_cache_31_internal_tensor_assign_1_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_31_internal_tensor_assign_1_end_mask_0 = const()[name = tensor("new_cache_31_internal_tensor_assign_1_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_31_internal_tensor_assign_1_squeeze_mask_0 = const()[name = tensor("new_cache_31_internal_tensor_assign_1_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_76 = const()[name = tensor("shape_76"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_30 = const()[name = tensor("reduce_prod_30"), val = tensor(1048576)]; + tensor range_1d_30_start_0 = const()[name = tensor("range_1d_30_start_0"), val = tensor(0)]; + tensor range_1d_30_step_0 = const()[name = tensor("range_1d_30_step_0"), val = tensor(1)]; + tensor range_1d_30 = range_1d(end = reduce_prod_30, start = range_1d_30_start_0, step = range_1d_30_step_0)[name = tensor("range_1d_30")]; + tensor reshape_150 = reshape(shape = shape_76, x = range_1d_30)[name = tensor("reshape_150")]; + tensor slice_by_index_30 = slice_by_index(begin = concat_122, begin_mask = new_cache_31_internal_tensor_assign_1_begin_mask_0, end = concat_123, end_mask = new_cache_31_internal_tensor_assign_1_end_mask_0, squeeze_mask = new_cache_31_internal_tensor_assign_1_squeeze_mask_0, stride = new_cache_31_internal_tensor_assign_1_stride_0, x = reshape_150)[name = tensor("slice_by_index_30")]; + tensor reshape_151_shape_0 = const()[name = tensor("reshape_151_shape_0"), val = tensor([-1])]; + tensor reshape_151 = reshape(shape = reshape_151_shape_0, x = slice_by_index_30)[name = tensor("reshape_151")]; + tensor reshape_152_shape_0 = const()[name = tensor("reshape_152_shape_0"), val = tensor([-1])]; + tensor reshape_152_cast_fp16 = reshape(shape = reshape_152_shape_0, x = var_6607_cast_fp16)[name = tensor("reshape_152_cast_fp16")]; + tensor reshape_153_shape_0 = const()[name = tensor("reshape_153_shape_0"), val = tensor([-1])]; + tensor reshape_153_cast_fp16 = reshape(shape = reshape_153_shape_0, x = cache15_to_fp16)[name = tensor("reshape_153_cast_fp16")]; + tensor scatter_30_mode_0 = const()[name = tensor("scatter_30_mode_0"), val = tensor("update")]; + tensor scatter_30_axis_0 = const()[name = tensor("scatter_30_axis_0"), val = tensor(0)]; + tensor scatter_30_validate_indices_0 = const()[name = tensor("scatter_30_validate_indices_0"), val = tensor(false)]; + tensor scatter_30_cast_fp16 = scatter(axis = scatter_30_axis_0, data = reshape_153_cast_fp16, indices = reshape_151, mode = scatter_30_mode_0, updates = reshape_152_cast_fp16, validate_indices = scatter_30_validate_indices_0)[name = tensor("scatter_30_cast_fp16")]; + tensor reshape_154_cast_fp16 = reshape(shape = shape_76, x = scatter_30_cast_fp16)[name = tensor("reshape_154_cast_fp16")]; + tensor var_6615_begin_0 = const()[name = tensor("op_6615_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor var_6615_end_0 = const()[name = tensor("op_6615_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor var_6615_end_mask_0 = const()[name = tensor("op_6615_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_6615_squeeze_mask_0 = const()[name = tensor("op_6615_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor var_6615_cast_fp16 = slice_by_index(begin = var_6615_begin_0, end = var_6615_end_0, end_mask = var_6615_end_mask_0, squeeze_mask = var_6615_squeeze_mask_0, x = reshape_154_cast_fp16)[name = tensor("op_6615_cast_fp16")]; + tensor var_6617_axis_0 = const()[name = tensor("op_6617_axis_0"), val = tensor(1)]; + tensor var_6617_mode_0 = const()[name = tensor("op_6617_mode_0"), val = tensor("update")]; + tensor var_6617_validate_indices_0 = const()[name = tensor("op_6617_validate_indices_0"), val = tensor(false)]; + tensor var_6617_cast_fp16 = scatter_along_axis(axis = var_6617_axis_0, data = var_6615_cast_fp16, indices = write_indices_31, mode = var_6617_mode_0, updates = v_31_cast_fp16, validate_indices = var_6617_validate_indices_0)[name = tensor("op_6617_cast_fp16")]; + tensor concat_124 = const()[name = tensor("concat_124"), val = tensor([1, 0, 0, 0, 0])]; + tensor concat_125 = const()[name = tensor("concat_125"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_31_internal_tensor_assign_2_stride_0 = const()[name = tensor("new_cache_31_internal_tensor_assign_2_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_31_internal_tensor_assign_2_begin_mask_0 = const()[name = tensor("new_cache_31_internal_tensor_assign_2_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_31_internal_tensor_assign_2_end_mask_0 = const()[name = tensor("new_cache_31_internal_tensor_assign_2_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_31_internal_tensor_assign_2_squeeze_mask_0 = const()[name = tensor("new_cache_31_internal_tensor_assign_2_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_77 = const()[name = tensor("shape_77"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_31 = const()[name = tensor("reduce_prod_31"), val = tensor(1048576)]; + tensor range_1d_31_start_0 = const()[name = tensor("range_1d_31_start_0"), val = tensor(0)]; + tensor range_1d_31_step_0 = const()[name = tensor("range_1d_31_step_0"), val = tensor(1)]; + tensor range_1d_31 = range_1d(end = reduce_prod_31, start = range_1d_31_start_0, step = range_1d_31_step_0)[name = tensor("range_1d_31")]; + tensor reshape_155 = reshape(shape = shape_77, x = range_1d_31)[name = tensor("reshape_155")]; + tensor slice_by_index_31 = slice_by_index(begin = concat_124, begin_mask = new_cache_31_internal_tensor_assign_2_begin_mask_0, end = concat_125, end_mask = new_cache_31_internal_tensor_assign_2_end_mask_0, squeeze_mask = new_cache_31_internal_tensor_assign_2_squeeze_mask_0, stride = new_cache_31_internal_tensor_assign_2_stride_0, x = reshape_155)[name = tensor("slice_by_index_31")]; + tensor reshape_156_shape_0 = const()[name = tensor("reshape_156_shape_0"), val = tensor([-1])]; + tensor reshape_156 = reshape(shape = reshape_156_shape_0, x = slice_by_index_31)[name = tensor("reshape_156")]; + tensor reshape_157_shape_0 = const()[name = tensor("reshape_157_shape_0"), val = tensor([-1])]; + tensor reshape_157_cast_fp16 = reshape(shape = reshape_157_shape_0, x = var_6617_cast_fp16)[name = tensor("reshape_157_cast_fp16")]; + tensor reshape_158_shape_0 = const()[name = tensor("reshape_158_shape_0"), val = tensor([-1])]; + tensor reshape_158_cast_fp16 = reshape(shape = reshape_158_shape_0, x = reshape_154_cast_fp16)[name = tensor("reshape_158_cast_fp16")]; + tensor scatter_31_mode_0 = const()[name = tensor("scatter_31_mode_0"), val = tensor("update")]; + tensor scatter_31_axis_0 = const()[name = tensor("scatter_31_axis_0"), val = tensor(0)]; + tensor scatter_31_validate_indices_0 = const()[name = tensor("scatter_31_validate_indices_0"), val = tensor(false)]; + tensor scatter_31_cast_fp16 = scatter(axis = scatter_31_axis_0, data = reshape_158_cast_fp16, indices = reshape_156, mode = scatter_31_mode_0, updates = reshape_157_cast_fp16, validate_indices = scatter_31_validate_indices_0)[name = tensor("scatter_31_cast_fp16")]; + tensor reshape_159_cast_fp16 = reshape(shape = shape_77, x = scatter_31_cast_fp16)[name = tensor("reshape_159_cast_fp16")]; + tensor reshape_159_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("reshape_159_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor keys_91_begin_0 = const()[name = tensor("keys_91_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor keys_91_end_0 = const()[name = tensor("keys_91_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor keys_91_end_mask_0 = const()[name = tensor("keys_91_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor keys_91_squeeze_mask_0 = const()[name = tensor("keys_91_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor keys_91_cast_fp16 = slice_by_index(begin = keys_91_begin_0, end = keys_91_end_0, end_mask = keys_91_end_mask_0, squeeze_mask = keys_91_squeeze_mask_0, x = reshape_159_cast_fp16)[name = tensor("keys_91_cast_fp16")]; + tensor values_91_begin_0 = const()[name = tensor("values_91_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor values_91_end_0 = const()[name = tensor("values_91_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor values_91_end_mask_0 = const()[name = tensor("values_91_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor values_91_squeeze_mask_0 = const()[name = tensor("values_91_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor values_91_cast_fp16 = slice_by_index(begin = values_91_begin_0, end = values_91_end_0, end_mask = values_91_end_mask_0, squeeze_mask = values_91_squeeze_mask_0, x = reshape_159_cast_fp16)[name = tensor("values_91_cast_fp16")]; + tensor var_6629_cast_fp16 = not_equal(x = keys_91_cast_fp16, y = keys_91_cast_fp16)[name = tensor("op_6629_cast_fp16")]; + tensor keys_93_cast_fp16 = select(a = var_470_to_fp16, b = keys_91_cast_fp16, cond = var_6629_cast_fp16)[name = tensor("keys_93_cast_fp16")]; + tensor var_6637_cast_fp16 = not_equal(x = values_91_cast_fp16, y = values_91_cast_fp16)[name = tensor("op_6637_cast_fp16")]; + tensor values_93_cast_fp16 = select(a = var_470_to_fp16, b = values_91_cast_fp16, cond = var_6637_cast_fp16)[name = tensor("values_93_cast_fp16")]; + tensor var_6661 = const()[name = tensor("op_6661"), val = tensor([0, 2, 1, 3])]; + tensor var_6677 = const()[name = tensor("op_6677"), val = tensor([1, 1, 1])]; + tensor var_6678_cast_fp16 = reshape(shape = var_6677, x = position15_to_fp16)[name = tensor("op_6678_cast_fp16")]; + tensor q_positions_31_cast_fp16 = add(x = var_6678_cast_fp16, y = q_offsets_1_promoted_to_fp16)[name = tensor("q_positions_31_cast_fp16")]; + tensor valid_end_31_cast_fp16 = add(x = var_6678_cast_fp16, y = var_536_cast_fp16)[name = tensor("valid_end_31_cast_fp16")]; + tensor valid_mask_31_cast_fp16 = less(x = k_positions_1_promoted_to_fp16, y = valid_end_31_cast_fp16)[name = tensor("valid_mask_31_cast_fp16")]; + tensor causal_mask_31_cast_fp16 = less_equal(x = k_positions_1_promoted_to_fp16, y = q_positions_31_cast_fp16)[name = tensor("causal_mask_31_cast_fp16")]; + tensor var_6706 = logical_and(x = valid_mask_31_cast_fp16, y = causal_mask_31_cast_fp16)[name = tensor("op_6706")]; + tensor attn_mask_31_axes_0 = const()[name = tensor("attn_mask_31_axes_0"), val = tensor([1])]; + tensor attn_mask_31 = expand_dims(axes = attn_mask_31_axes_0, x = var_6706)[name = tensor("attn_mask_31")]; + tensor var_6720_transpose_x_0 = const()[name = tensor("op_6720_transpose_x_0"), val = tensor(false)]; + tensor var_6720_transpose_y_0 = const()[name = tensor("op_6720_transpose_y_0"), val = tensor(false)]; + tensor transpose_99_perm_0 = const()[name = tensor("transpose_99_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_100_perm_0 = const()[name = tensor("transpose_100_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_100 = transpose(perm = transpose_100_perm_0, x = keys_93_cast_fp16)[name = tensor("transpose_144")]; + tensor transpose_99 = transpose(perm = transpose_99_perm_0, x = q_93_cast_fp16)[name = tensor("transpose_145")]; + tensor var_6720_cast_fp16 = matmul(transpose_x = var_6720_transpose_x_0, transpose_y = var_6720_transpose_y_0, x = transpose_99, y = transpose_100)[name = tensor("op_6720_cast_fp16")]; + tensor var_6714_to_fp16 = const()[name = tensor("op_6714_to_fp16"), val = tensor([0x1p-3])]; + tensor attn_weights_91_cast_fp16 = mul(x = var_6720_cast_fp16, y = var_6714_to_fp16)[name = tensor("attn_weights_91_cast_fp16")]; + tensor var_6722 = logical_not(x = attn_mask_31)[name = tensor("op_6722")]; + tensor var_6723_to_fp16 = const()[name = tensor("op_6723_to_fp16"), val = tensor(-inf)]; + tensor attn_weights_93_cast_fp16 = select(a = var_6723_to_fp16, b = attn_weights_91_cast_fp16, cond = var_6722)[name = tensor("attn_weights_93_cast_fp16")]; + tensor var_6725 = const()[name = tensor("op_6725"), val = tensor(-1)]; + tensor attn_weights_95_cast_fp16 = softmax(axis = var_6725, x = attn_weights_93_cast_fp16)[name = tensor("attn_weights_95_cast_fp16")]; + tensor attn_output_31_transpose_x_0 = const()[name = tensor("attn_output_31_transpose_x_0"), val = tensor(false)]; + tensor attn_output_31_transpose_y_0 = const()[name = tensor("attn_output_31_transpose_y_0"), val = tensor(false)]; + tensor values_95_cast_fp16 = transpose(perm = var_6661, x = values_93_cast_fp16)[name = tensor("transpose_146")]; + tensor attn_output_31_cast_fp16 = matmul(transpose_x = attn_output_31_transpose_x_0, transpose_y = attn_output_31_transpose_y_0, x = attn_weights_95_cast_fp16, y = values_95_cast_fp16)[name = tensor("attn_output_31_cast_fp16")]; + tensor var_6733 = const()[name = tensor("op_6733"), val = tensor([0, 2, 1, 3])]; + tensor var_6738 = const()[name = tensor("op_6738"), val = tensor([1, 256, 1024])]; + tensor var_6734_cast_fp16 = transpose(perm = var_6733, x = attn_output_31_cast_fp16)[name = tensor("transpose_143")]; + tensor input_185_cast_fp16 = reshape(shape = var_6738, x = var_6734_cast_fp16)[name = tensor("input_185_cast_fp16")]; + tensor attn15_out_proj_weight_to_fp16 = const()[name = tensor("attn15_out_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(384983808)))]; + tensor linear_61_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = attn15_out_proj_weight_to_fp16, x = input_185_cast_fp16)[name = tensor("linear_61_cast_fp16")]; + tensor var_6747_cast_fp16 = add(x = position15_to_fp16, y = valid_len_to_fp16)[name = tensor("op_6747_cast_fp16")]; + tensor var_6747_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_6747_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor input_187_cast_fp16 = add(x = input_181_cast_fp16, y = linear_61_cast_fp16)[name = tensor("input_187_cast_fp16")]; + tensor input_189_axes_0 = const()[name = tensor("input_189_axes_0"), val = tensor([-1])]; + tensor norm15_2_weight_to_fp16 = const()[name = tensor("norm15_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(387081024)))]; + tensor norm15_2_bias_to_fp16 = const()[name = tensor("norm15_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(387083136)))]; + tensor var_6751_to_fp16 = const()[name = tensor("op_6751_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_189_cast_fp16 = layer_norm(axes = input_189_axes_0, beta = norm15_2_bias_to_fp16, epsilon = var_6751_to_fp16, gamma = norm15_2_weight_to_fp16, x = input_187_cast_fp16)[name = tensor("input_189_cast_fp16")]; + tensor linear15_1_weight_to_fp16 = const()[name = tensor("linear15_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(387085248)))]; + tensor linear_62_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = linear15_1_weight_to_fp16, x = input_189_cast_fp16)[name = tensor("linear_62_cast_fp16")]; + tensor input_191_mode_0 = const()[name = tensor("input_191_mode_0"), val = tensor("EXACT")]; + tensor input_191_cast_fp16 = gelu(mode = input_191_mode_0, x = linear_62_cast_fp16)[name = tensor("input_191_cast_fp16")]; + tensor linear15_2_weight_to_fp16 = const()[name = tensor("linear15_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(395473920)))]; + tensor linear_63_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = linear15_2_weight_to_fp16, x = input_191_cast_fp16)[name = tensor("linear_63_cast_fp16")]; + tensor input_193_cast_fp16 = add(x = input_187_cast_fp16, y = linear_63_cast_fp16)[name = tensor("input_193_cast_fp16")]; + tensor input_195_axes_0 = const()[name = tensor("input_195_axes_0"), val = tensor([-1])]; + tensor norm16_1_weight_to_fp16 = const()[name = tensor("norm16_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(403862592)))]; + tensor norm16_1_bias_to_fp16 = const()[name = tensor("norm16_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(403864704)))]; + tensor var_6768_to_fp16 = const()[name = tensor("op_6768_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_195_cast_fp16 = layer_norm(axes = input_195_axes_0, beta = norm16_1_bias_to_fp16, epsilon = var_6768_to_fp16, gamma = norm16_1_weight_to_fp16, x = input_193_cast_fp16)[name = tensor("input_195_cast_fp16")]; + tensor attn16_in_proj_weight_to_fp16 = const()[name = tensor("attn16_in_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(403866816)))]; + tensor linear_64_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = attn16_in_proj_weight_to_fp16, x = input_195_cast_fp16)[name = tensor("linear_64_cast_fp16")]; + tensor var_6782 = const()[name = tensor("op_6782"), val = tensor([1, 256, 3, 16, 64])]; + tensor qkv_33_cast_fp16 = reshape(shape = var_6782, x = linear_64_cast_fp16)[name = tensor("qkv_33_cast_fp16")]; + tensor q_97_begin_0 = const()[name = tensor("q_97_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor q_97_end_0 = const()[name = tensor("q_97_end_0"), val = tensor([1, 256, 1, 16, 64])]; + tensor q_97_end_mask_0 = const()[name = tensor("q_97_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor q_97_squeeze_mask_0 = const()[name = tensor("q_97_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor q_97_cast_fp16 = slice_by_index(begin = q_97_begin_0, end = q_97_end_0, end_mask = q_97_end_mask_0, squeeze_mask = q_97_squeeze_mask_0, x = qkv_33_cast_fp16)[name = tensor("q_97_cast_fp16")]; + tensor k_65_begin_0 = const()[name = tensor("k_65_begin_0"), val = tensor([0, 0, 1, 0, 0])]; + tensor k_65_end_0 = const()[name = tensor("k_65_end_0"), val = tensor([1, 256, 2, 16, 64])]; + tensor k_65_end_mask_0 = const()[name = tensor("k_65_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor k_65_squeeze_mask_0 = const()[name = tensor("k_65_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor k_65_cast_fp16 = slice_by_index(begin = k_65_begin_0, end = k_65_end_0, end_mask = k_65_end_mask_0, squeeze_mask = k_65_squeeze_mask_0, x = qkv_33_cast_fp16)[name = tensor("k_65_cast_fp16")]; + tensor v_33_begin_0 = const()[name = tensor("v_33_begin_0"), val = tensor([0, 0, 2, 0, 0])]; + tensor v_33_end_0 = const()[name = tensor("v_33_end_0"), val = tensor([1, 256, 3, 16, 64])]; + tensor v_33_end_mask_0 = const()[name = tensor("v_33_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor v_33_squeeze_mask_0 = const()[name = tensor("v_33_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor v_33_cast_fp16 = slice_by_index(begin = v_33_begin_0, end = v_33_end_0, end_mask = v_33_end_mask_0, squeeze_mask = v_33_squeeze_mask_0, x = qkv_33_cast_fp16)[name = tensor("v_33_cast_fp16")]; + tensor var_6840 = const()[name = tensor("op_6840"), val = tensor([1, 1])]; + tensor position16_to_fp16_dtype_0 = const()[name = tensor("position16_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor position16_to_fp16 = cast(dtype = position16_to_fp16_dtype_0, x = position16)[name = tensor("cast_581")]; + tensor var_6841_cast_fp16 = reshape(shape = var_6840, x = position16_to_fp16)[name = tensor("op_6841_cast_fp16")]; + tensor ts_99_cast_fp16 = add(x = ts_1_promoted_to_fp16, y = var_6841_cast_fp16)[name = tensor("ts_99_cast_fp16")]; + tensor var_6848 = const()[name = tensor("op_6848"), val = tensor([1, 256, 1, 1])]; + tensor ts_101_cast_fp16 = reshape(shape = var_6848, x = ts_99_cast_fp16)[name = tensor("ts_101_cast_fp16")]; + tensor var_6855 = const()[name = tensor("op_6855"), val = tensor([1, 256, 16, 32, 2])]; + tensor q_complex_33_cast_fp16 = reshape(shape = var_6855, x = q_97_cast_fp16)[name = tensor("q_complex_33_cast_fp16")]; + tensor var_6862 = const()[name = tensor("op_6862"), val = tensor([1, 256, 16, 32, 2])]; + tensor k_complex_33_cast_fp16 = reshape(shape = var_6862, x = k_65_cast_fp16)[name = tensor("k_complex_33_cast_fp16")]; + tensor var_6866_begin_0 = const()[name = tensor("op_6866_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_6866_end_0 = const()[name = tensor("op_6866_end_0"), val = tensor([1, 256, 16, 32, 1])]; + tensor var_6866_end_mask_0 = const()[name = tensor("op_6866_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_6866_squeeze_mask_0 = const()[name = tensor("op_6866_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_6866_cast_fp16 = slice_by_index(begin = var_6866_begin_0, end = var_6866_end_0, end_mask = var_6866_end_mask_0, squeeze_mask = var_6866_squeeze_mask_0, x = q_complex_33_cast_fp16)[name = tensor("op_6866_cast_fp16")]; + tensor var_6874_begin_0 = const()[name = tensor("op_6874_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_6874_end_0 = const()[name = tensor("op_6874_end_0"), val = tensor([1, 256, 16, 32, 2])]; + tensor var_6874_end_mask_0 = const()[name = tensor("op_6874_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_6874_squeeze_mask_0 = const()[name = tensor("op_6874_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_6874_cast_fp16 = slice_by_index(begin = var_6874_begin_0, end = var_6874_end_0, end_mask = var_6874_end_mask_0, squeeze_mask = var_6874_squeeze_mask_0, x = q_complex_33_cast_fp16)[name = tensor("op_6874_cast_fp16")]; + tensor var_6882_begin_0 = const()[name = tensor("op_6882_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_6882_end_0 = const()[name = tensor("op_6882_end_0"), val = tensor([1, 256, 16, 32, 1])]; + tensor var_6882_end_mask_0 = const()[name = tensor("op_6882_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_6882_squeeze_mask_0 = const()[name = tensor("op_6882_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_6882_cast_fp16 = slice_by_index(begin = var_6882_begin_0, end = var_6882_end_0, end_mask = var_6882_end_mask_0, squeeze_mask = var_6882_squeeze_mask_0, x = k_complex_33_cast_fp16)[name = tensor("op_6882_cast_fp16")]; + tensor var_6890_begin_0 = const()[name = tensor("op_6890_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_6890_end_0 = const()[name = tensor("op_6890_end_0"), val = tensor([1, 256, 16, 32, 2])]; + tensor var_6890_end_mask_0 = const()[name = tensor("op_6890_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_6890_squeeze_mask_0 = const()[name = tensor("op_6890_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_6890_cast_fp16 = slice_by_index(begin = var_6890_begin_0, end = var_6890_end_0, end_mask = var_6890_end_mask_0, squeeze_mask = var_6890_squeeze_mask_0, x = k_complex_33_cast_fp16)[name = tensor("op_6890_cast_fp16")]; + tensor freqs_33_to_fp16 = const()[name = tensor("freqs_33_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(410158336)))]; + tensor var_6896_cast_fp16 = mul(x = freqs_33_to_fp16, y = ts_101_cast_fp16)[name = tensor("op_6896_cast_fp16")]; + tensor rotr_33_cast_fp16 = cos(x = var_6896_cast_fp16)[name = tensor("rotr_33_cast_fp16")]; + tensor roti_33_cast_fp16 = sin(x = var_6896_cast_fp16)[name = tensor("roti_33_cast_fp16")]; + tensor var_6900_cast_fp16 = mul(x = var_6866_cast_fp16, y = rotr_33_cast_fp16)[name = tensor("op_6900_cast_fp16")]; + tensor var_6901_cast_fp16 = mul(x = var_6874_cast_fp16, y = roti_33_cast_fp16)[name = tensor("op_6901_cast_fp16")]; + tensor qor_65_cast_fp16 = sub(x = var_6900_cast_fp16, y = var_6901_cast_fp16)[name = tensor("qor_65_cast_fp16")]; + tensor var_6904_cast_fp16 = mul(x = var_6866_cast_fp16, y = roti_33_cast_fp16)[name = tensor("op_6904_cast_fp16")]; + tensor var_6905_cast_fp16 = mul(x = var_6874_cast_fp16, y = rotr_33_cast_fp16)[name = tensor("op_6905_cast_fp16")]; + tensor qoi_65_cast_fp16 = add(x = var_6904_cast_fp16, y = var_6905_cast_fp16)[name = tensor("qoi_65_cast_fp16")]; + tensor var_6908_cast_fp16 = mul(x = var_6882_cast_fp16, y = rotr_33_cast_fp16)[name = tensor("op_6908_cast_fp16")]; + tensor var_6909_cast_fp16 = mul(x = var_6890_cast_fp16, y = roti_33_cast_fp16)[name = tensor("op_6909_cast_fp16")]; + tensor kor_65_cast_fp16 = sub(x = var_6908_cast_fp16, y = var_6909_cast_fp16)[name = tensor("kor_65_cast_fp16")]; + tensor var_6912_cast_fp16 = mul(x = var_6882_cast_fp16, y = roti_33_cast_fp16)[name = tensor("op_6912_cast_fp16")]; + tensor var_6913_cast_fp16 = mul(x = var_6890_cast_fp16, y = rotr_33_cast_fp16)[name = tensor("op_6913_cast_fp16")]; + tensor koi_65_cast_fp16 = add(x = var_6912_cast_fp16, y = var_6913_cast_fp16)[name = tensor("koi_65_cast_fp16")]; + tensor qo_33_axis_0 = const()[name = tensor("qo_33_axis_0"), val = tensor(-1)]; + tensor qo_33_cast_fp16 = stack(axis = qo_33_axis_0, values = (qor_65_cast_fp16, qoi_65_cast_fp16))[name = tensor("qo_33_cast_fp16")]; + tensor ko_33_axis_0 = const()[name = tensor("ko_33_axis_0"), val = tensor(-1)]; + tensor ko_33_cast_fp16 = stack(axis = ko_33_axis_0, values = (kor_65_cast_fp16, koi_65_cast_fp16))[name = tensor("ko_33_cast_fp16")]; + tensor var_6946 = const()[name = tensor("op_6946"), val = tensor([1, 256, 16, 64])]; + tensor q_99_cast_fp16 = reshape(shape = var_6946, x = qo_33_cast_fp16)[name = tensor("q_99_cast_fp16")]; + tensor var_6952 = const()[name = tensor("op_6952"), val = tensor([1, 256, 16, 64])]; + tensor k_67_cast_fp16 = reshape(shape = var_6952, x = ko_33_cast_fp16)[name = tensor("k_67_cast_fp16")]; + tensor write_indices_float_97_cast_fp16 = add(x = var_6841_cast_fp16, y = var_393_promoted_to_fp16)[name = tensor("write_indices_float_97_cast_fp16")]; + tensor _inversed_6973_y_0_to_fp16 = const()[name = tensor("_inversed_6973_y_0_to_fp16"), val = tensor(0x1p-9)]; + tensor _inversed_6973_cast_fp16 = mul(x = write_indices_float_97_cast_fp16, y = _inversed_6973_y_0_to_fp16)[name = tensor("_inversed_6973_cast_fp16")]; + tensor var_6974_cast_fp16 = floor(x = _inversed_6973_cast_fp16)[name = tensor("op_6974_cast_fp16")]; + tensor var_6975_to_fp16 = const()[name = tensor("op_6975_to_fp16"), val = tensor(0x1p+9)]; + tensor var_6976_cast_fp16 = mul(x = var_6974_cast_fp16, y = var_6975_to_fp16)[name = tensor("op_6976_cast_fp16")]; + tensor write_indices_float_99_cast_fp16 = sub(x = write_indices_float_97_cast_fp16, y = var_6976_cast_fp16)[name = tensor("write_indices_float_99_cast_fp16")]; + tensor var_6994_value_0_to_fp16 = const()[name = tensor("op_6994_value_0_to_fp16"), val = tensor(0x1.ffp+8)]; + tensor var_6994_cast_fp16 = fill_like(ref_tensor = write_indices_float_99_cast_fp16, value = var_6994_value_0_to_fp16)[name = tensor("op_6994_cast_fp16")]; + tensor write_indices_float_101_cast_fp16 = select(a = write_indices_float_99_cast_fp16, b = var_6994_cast_fp16, cond = token_valid_1_cast_fp16)[name = tensor("write_indices_float_101_cast_fp16")]; + tensor var_7000_dtype_0 = const()[name = tensor("op_7000_dtype_0"), val = tensor("int32")]; + tensor var_7005 = const()[name = tensor("op_7005"), val = tensor([1, 256, 1, 1])]; + tensor write_indices_float_101_cast_fp16_to_int32 = cast(dtype = var_7000_dtype_0, x = write_indices_float_101_cast_fp16)[name = tensor("cast_580")]; + tensor var_7006 = reshape(shape = var_7005, x = write_indices_float_101_cast_fp16_to_int32)[name = tensor("op_7006")]; + tensor write_indices_33_reps_0 = const()[name = tensor("write_indices_33_reps_0"), val = tensor([1, 1, 16, 64])]; + tensor write_indices_33 = tile(reps = write_indices_33_reps_0, x = var_7006)[name = tensor("write_indices_33")]; + tensor var_7016_begin_0 = const()[name = tensor("op_7016_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_7016_end_0 = const()[name = tensor("op_7016_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor var_7016_end_mask_0 = const()[name = tensor("op_7016_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_7016_squeeze_mask_0 = const()[name = tensor("op_7016_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor cache16_to_fp16_dtype_0 = const()[name = tensor("cache16_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor cache16_to_fp16 = cast(dtype = cache16_to_fp16_dtype_0, x = cache16)[name = tensor("cast_579")]; + tensor var_7016_cast_fp16 = slice_by_index(begin = var_7016_begin_0, end = var_7016_end_0, end_mask = var_7016_end_mask_0, squeeze_mask = var_7016_squeeze_mask_0, x = cache16_to_fp16)[name = tensor("op_7016_cast_fp16")]; + tensor var_7018_axis_0 = const()[name = tensor("op_7018_axis_0"), val = tensor(1)]; + tensor var_7018_mode_0 = const()[name = tensor("op_7018_mode_0"), val = tensor("update")]; + tensor var_7018_validate_indices_0 = const()[name = tensor("op_7018_validate_indices_0"), val = tensor(false)]; + tensor var_7018_cast_fp16 = scatter_along_axis(axis = var_7018_axis_0, data = var_7016_cast_fp16, indices = write_indices_33, mode = var_7018_mode_0, updates = k_67_cast_fp16, validate_indices = var_7018_validate_indices_0)[name = tensor("op_7018_cast_fp16")]; + tensor concat_130 = const()[name = tensor("concat_130"), val = tensor([0, 0, 0, 0, 0])]; + tensor concat_131 = const()[name = tensor("concat_131"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_33_internal_tensor_assign_1_stride_0 = const()[name = tensor("new_cache_33_internal_tensor_assign_1_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_33_internal_tensor_assign_1_begin_mask_0 = const()[name = tensor("new_cache_33_internal_tensor_assign_1_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_33_internal_tensor_assign_1_end_mask_0 = const()[name = tensor("new_cache_33_internal_tensor_assign_1_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_33_internal_tensor_assign_1_squeeze_mask_0 = const()[name = tensor("new_cache_33_internal_tensor_assign_1_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_78 = const()[name = tensor("shape_78"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_32 = const()[name = tensor("reduce_prod_32"), val = tensor(1048576)]; + tensor range_1d_32_start_0 = const()[name = tensor("range_1d_32_start_0"), val = tensor(0)]; + tensor range_1d_32_step_0 = const()[name = tensor("range_1d_32_step_0"), val = tensor(1)]; + tensor range_1d_32 = range_1d(end = reduce_prod_32, start = range_1d_32_start_0, step = range_1d_32_step_0)[name = tensor("range_1d_32")]; + tensor reshape_160 = reshape(shape = shape_78, x = range_1d_32)[name = tensor("reshape_160")]; + tensor slice_by_index_32 = slice_by_index(begin = concat_130, begin_mask = new_cache_33_internal_tensor_assign_1_begin_mask_0, end = concat_131, end_mask = new_cache_33_internal_tensor_assign_1_end_mask_0, squeeze_mask = new_cache_33_internal_tensor_assign_1_squeeze_mask_0, stride = new_cache_33_internal_tensor_assign_1_stride_0, x = reshape_160)[name = tensor("slice_by_index_32")]; + tensor reshape_161_shape_0 = const()[name = tensor("reshape_161_shape_0"), val = tensor([-1])]; + tensor reshape_161 = reshape(shape = reshape_161_shape_0, x = slice_by_index_32)[name = tensor("reshape_161")]; + tensor reshape_162_shape_0 = const()[name = tensor("reshape_162_shape_0"), val = tensor([-1])]; + tensor reshape_162_cast_fp16 = reshape(shape = reshape_162_shape_0, x = var_7018_cast_fp16)[name = tensor("reshape_162_cast_fp16")]; + tensor reshape_163_shape_0 = const()[name = tensor("reshape_163_shape_0"), val = tensor([-1])]; + tensor reshape_163_cast_fp16 = reshape(shape = reshape_163_shape_0, x = cache16_to_fp16)[name = tensor("reshape_163_cast_fp16")]; + tensor scatter_32_mode_0 = const()[name = tensor("scatter_32_mode_0"), val = tensor("update")]; + tensor scatter_32_axis_0 = const()[name = tensor("scatter_32_axis_0"), val = tensor(0)]; + tensor scatter_32_validate_indices_0 = const()[name = tensor("scatter_32_validate_indices_0"), val = tensor(false)]; + tensor scatter_32_cast_fp16 = scatter(axis = scatter_32_axis_0, data = reshape_163_cast_fp16, indices = reshape_161, mode = scatter_32_mode_0, updates = reshape_162_cast_fp16, validate_indices = scatter_32_validate_indices_0)[name = tensor("scatter_32_cast_fp16")]; + tensor reshape_164_cast_fp16 = reshape(shape = shape_78, x = scatter_32_cast_fp16)[name = tensor("reshape_164_cast_fp16")]; + tensor var_7026_begin_0 = const()[name = tensor("op_7026_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor var_7026_end_0 = const()[name = tensor("op_7026_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor var_7026_end_mask_0 = const()[name = tensor("op_7026_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_7026_squeeze_mask_0 = const()[name = tensor("op_7026_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor var_7026_cast_fp16 = slice_by_index(begin = var_7026_begin_0, end = var_7026_end_0, end_mask = var_7026_end_mask_0, squeeze_mask = var_7026_squeeze_mask_0, x = reshape_164_cast_fp16)[name = tensor("op_7026_cast_fp16")]; + tensor var_7028_axis_0 = const()[name = tensor("op_7028_axis_0"), val = tensor(1)]; + tensor var_7028_mode_0 = const()[name = tensor("op_7028_mode_0"), val = tensor("update")]; + tensor var_7028_validate_indices_0 = const()[name = tensor("op_7028_validate_indices_0"), val = tensor(false)]; + tensor var_7028_cast_fp16 = scatter_along_axis(axis = var_7028_axis_0, data = var_7026_cast_fp16, indices = write_indices_33, mode = var_7028_mode_0, updates = v_33_cast_fp16, validate_indices = var_7028_validate_indices_0)[name = tensor("op_7028_cast_fp16")]; + tensor concat_132 = const()[name = tensor("concat_132"), val = tensor([1, 0, 0, 0, 0])]; + tensor concat_133 = const()[name = tensor("concat_133"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_33_internal_tensor_assign_2_stride_0 = const()[name = tensor("new_cache_33_internal_tensor_assign_2_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_33_internal_tensor_assign_2_begin_mask_0 = const()[name = tensor("new_cache_33_internal_tensor_assign_2_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_33_internal_tensor_assign_2_end_mask_0 = const()[name = tensor("new_cache_33_internal_tensor_assign_2_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_33_internal_tensor_assign_2_squeeze_mask_0 = const()[name = tensor("new_cache_33_internal_tensor_assign_2_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_79 = const()[name = tensor("shape_79"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_33 = const()[name = tensor("reduce_prod_33"), val = tensor(1048576)]; + tensor range_1d_33_start_0 = const()[name = tensor("range_1d_33_start_0"), val = tensor(0)]; + tensor range_1d_33_step_0 = const()[name = tensor("range_1d_33_step_0"), val = tensor(1)]; + tensor range_1d_33 = range_1d(end = reduce_prod_33, start = range_1d_33_start_0, step = range_1d_33_step_0)[name = tensor("range_1d_33")]; + tensor reshape_165 = reshape(shape = shape_79, x = range_1d_33)[name = tensor("reshape_165")]; + tensor slice_by_index_33 = slice_by_index(begin = concat_132, begin_mask = new_cache_33_internal_tensor_assign_2_begin_mask_0, end = concat_133, end_mask = new_cache_33_internal_tensor_assign_2_end_mask_0, squeeze_mask = new_cache_33_internal_tensor_assign_2_squeeze_mask_0, stride = new_cache_33_internal_tensor_assign_2_stride_0, x = reshape_165)[name = tensor("slice_by_index_33")]; + tensor reshape_166_shape_0 = const()[name = tensor("reshape_166_shape_0"), val = tensor([-1])]; + tensor reshape_166 = reshape(shape = reshape_166_shape_0, x = slice_by_index_33)[name = tensor("reshape_166")]; + tensor reshape_167_shape_0 = const()[name = tensor("reshape_167_shape_0"), val = tensor([-1])]; + tensor reshape_167_cast_fp16 = reshape(shape = reshape_167_shape_0, x = var_7028_cast_fp16)[name = tensor("reshape_167_cast_fp16")]; + tensor reshape_168_shape_0 = const()[name = tensor("reshape_168_shape_0"), val = tensor([-1])]; + tensor reshape_168_cast_fp16 = reshape(shape = reshape_168_shape_0, x = reshape_164_cast_fp16)[name = tensor("reshape_168_cast_fp16")]; + tensor scatter_33_mode_0 = const()[name = tensor("scatter_33_mode_0"), val = tensor("update")]; + tensor scatter_33_axis_0 = const()[name = tensor("scatter_33_axis_0"), val = tensor(0)]; + tensor scatter_33_validate_indices_0 = const()[name = tensor("scatter_33_validate_indices_0"), val = tensor(false)]; + tensor scatter_33_cast_fp16 = scatter(axis = scatter_33_axis_0, data = reshape_168_cast_fp16, indices = reshape_166, mode = scatter_33_mode_0, updates = reshape_167_cast_fp16, validate_indices = scatter_33_validate_indices_0)[name = tensor("scatter_33_cast_fp16")]; + tensor reshape_169_cast_fp16 = reshape(shape = shape_79, x = scatter_33_cast_fp16)[name = tensor("reshape_169_cast_fp16")]; + tensor reshape_169_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("reshape_169_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor keys_97_begin_0 = const()[name = tensor("keys_97_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor keys_97_end_0 = const()[name = tensor("keys_97_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor keys_97_end_mask_0 = const()[name = tensor("keys_97_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor keys_97_squeeze_mask_0 = const()[name = tensor("keys_97_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor keys_97_cast_fp16 = slice_by_index(begin = keys_97_begin_0, end = keys_97_end_0, end_mask = keys_97_end_mask_0, squeeze_mask = keys_97_squeeze_mask_0, x = reshape_169_cast_fp16)[name = tensor("keys_97_cast_fp16")]; + tensor values_97_begin_0 = const()[name = tensor("values_97_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor values_97_end_0 = const()[name = tensor("values_97_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor values_97_end_mask_0 = const()[name = tensor("values_97_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor values_97_squeeze_mask_0 = const()[name = tensor("values_97_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor values_97_cast_fp16 = slice_by_index(begin = values_97_begin_0, end = values_97_end_0, end_mask = values_97_end_mask_0, squeeze_mask = values_97_squeeze_mask_0, x = reshape_169_cast_fp16)[name = tensor("values_97_cast_fp16")]; + tensor var_7040_cast_fp16 = not_equal(x = keys_97_cast_fp16, y = keys_97_cast_fp16)[name = tensor("op_7040_cast_fp16")]; + tensor keys_99_cast_fp16 = select(a = var_470_to_fp16, b = keys_97_cast_fp16, cond = var_7040_cast_fp16)[name = tensor("keys_99_cast_fp16")]; + tensor var_7048_cast_fp16 = not_equal(x = values_97_cast_fp16, y = values_97_cast_fp16)[name = tensor("op_7048_cast_fp16")]; + tensor values_99_cast_fp16 = select(a = var_470_to_fp16, b = values_97_cast_fp16, cond = var_7048_cast_fp16)[name = tensor("values_99_cast_fp16")]; + tensor var_7072 = const()[name = tensor("op_7072"), val = tensor([0, 2, 1, 3])]; + tensor var_7088 = const()[name = tensor("op_7088"), val = tensor([1, 1, 1])]; + tensor var_7089_cast_fp16 = reshape(shape = var_7088, x = position16_to_fp16)[name = tensor("op_7089_cast_fp16")]; + tensor q_positions_33_cast_fp16 = add(x = var_7089_cast_fp16, y = q_offsets_1_promoted_to_fp16)[name = tensor("q_positions_33_cast_fp16")]; + tensor valid_end_33_cast_fp16 = add(x = var_7089_cast_fp16, y = var_536_cast_fp16)[name = tensor("valid_end_33_cast_fp16")]; + tensor valid_mask_33_cast_fp16 = less(x = k_positions_1_promoted_to_fp16, y = valid_end_33_cast_fp16)[name = tensor("valid_mask_33_cast_fp16")]; + tensor causal_mask_33_cast_fp16 = less_equal(x = k_positions_1_promoted_to_fp16, y = q_positions_33_cast_fp16)[name = tensor("causal_mask_33_cast_fp16")]; + tensor var_7117 = logical_and(x = valid_mask_33_cast_fp16, y = causal_mask_33_cast_fp16)[name = tensor("op_7117")]; + tensor attn_mask_33_axes_0 = const()[name = tensor("attn_mask_33_axes_0"), val = tensor([1])]; + tensor attn_mask_33 = expand_dims(axes = attn_mask_33_axes_0, x = var_7117)[name = tensor("attn_mask_33")]; + tensor var_7131_transpose_x_0 = const()[name = tensor("op_7131_transpose_x_0"), val = tensor(false)]; + tensor var_7131_transpose_y_0 = const()[name = tensor("op_7131_transpose_y_0"), val = tensor(false)]; + tensor transpose_101_perm_0 = const()[name = tensor("transpose_101_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_102_perm_0 = const()[name = tensor("transpose_102_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_102 = transpose(perm = transpose_102_perm_0, x = keys_99_cast_fp16)[name = tensor("transpose_140")]; + tensor transpose_101 = transpose(perm = transpose_101_perm_0, x = q_99_cast_fp16)[name = tensor("transpose_141")]; + tensor var_7131_cast_fp16 = matmul(transpose_x = var_7131_transpose_x_0, transpose_y = var_7131_transpose_y_0, x = transpose_101, y = transpose_102)[name = tensor("op_7131_cast_fp16")]; + tensor var_7125_to_fp16 = const()[name = tensor("op_7125_to_fp16"), val = tensor([0x1p-3])]; + tensor attn_weights_97_cast_fp16 = mul(x = var_7131_cast_fp16, y = var_7125_to_fp16)[name = tensor("attn_weights_97_cast_fp16")]; + tensor var_7133 = logical_not(x = attn_mask_33)[name = tensor("op_7133")]; + tensor var_7134_to_fp16 = const()[name = tensor("op_7134_to_fp16"), val = tensor(-inf)]; + tensor attn_weights_99_cast_fp16 = select(a = var_7134_to_fp16, b = attn_weights_97_cast_fp16, cond = var_7133)[name = tensor("attn_weights_99_cast_fp16")]; + tensor var_7136 = const()[name = tensor("op_7136"), val = tensor(-1)]; + tensor attn_weights_101_cast_fp16 = softmax(axis = var_7136, x = attn_weights_99_cast_fp16)[name = tensor("attn_weights_101_cast_fp16")]; + tensor attn_output_33_transpose_x_0 = const()[name = tensor("attn_output_33_transpose_x_0"), val = tensor(false)]; + tensor attn_output_33_transpose_y_0 = const()[name = tensor("attn_output_33_transpose_y_0"), val = tensor(false)]; + tensor values_101_cast_fp16 = transpose(perm = var_7072, x = values_99_cast_fp16)[name = tensor("transpose_142")]; + tensor attn_output_33_cast_fp16 = matmul(transpose_x = attn_output_33_transpose_x_0, transpose_y = attn_output_33_transpose_y_0, x = attn_weights_101_cast_fp16, y = values_101_cast_fp16)[name = tensor("attn_output_33_cast_fp16")]; + tensor var_7144 = const()[name = tensor("op_7144"), val = tensor([0, 2, 1, 3])]; + tensor var_7149 = const()[name = tensor("op_7149"), val = tensor([1, 256, 1024])]; + tensor var_7145_cast_fp16 = transpose(perm = var_7144, x = attn_output_33_cast_fp16)[name = tensor("transpose_139")]; + tensor input_197_cast_fp16 = reshape(shape = var_7149, x = var_7145_cast_fp16)[name = tensor("input_197_cast_fp16")]; + tensor attn16_out_proj_weight_to_fp16 = const()[name = tensor("attn16_out_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(410158464)))]; + tensor linear_65_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = attn16_out_proj_weight_to_fp16, x = input_197_cast_fp16)[name = tensor("linear_65_cast_fp16")]; + tensor var_7158_cast_fp16 = add(x = position16_to_fp16, y = valid_len_to_fp16)[name = tensor("op_7158_cast_fp16")]; + tensor var_7158_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_7158_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor input_199_cast_fp16 = add(x = input_193_cast_fp16, y = linear_65_cast_fp16)[name = tensor("input_199_cast_fp16")]; + tensor input_201_axes_0 = const()[name = tensor("input_201_axes_0"), val = tensor([-1])]; + tensor norm16_2_weight_to_fp16 = const()[name = tensor("norm16_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(412255680)))]; + tensor norm16_2_bias_to_fp16 = const()[name = tensor("norm16_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(412257792)))]; + tensor var_7162_to_fp16 = const()[name = tensor("op_7162_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_201_cast_fp16 = layer_norm(axes = input_201_axes_0, beta = norm16_2_bias_to_fp16, epsilon = var_7162_to_fp16, gamma = norm16_2_weight_to_fp16, x = input_199_cast_fp16)[name = tensor("input_201_cast_fp16")]; + tensor linear16_1_weight_to_fp16 = const()[name = tensor("linear16_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(412259904)))]; + tensor linear_66_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = linear16_1_weight_to_fp16, x = input_201_cast_fp16)[name = tensor("linear_66_cast_fp16")]; + tensor input_203_mode_0 = const()[name = tensor("input_203_mode_0"), val = tensor("EXACT")]; + tensor input_203_cast_fp16 = gelu(mode = input_203_mode_0, x = linear_66_cast_fp16)[name = tensor("input_203_cast_fp16")]; + tensor linear16_2_weight_to_fp16 = const()[name = tensor("linear16_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(420648576)))]; + tensor linear_67_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = linear16_2_weight_to_fp16, x = input_203_cast_fp16)[name = tensor("linear_67_cast_fp16")]; + tensor input_205_cast_fp16 = add(x = input_199_cast_fp16, y = linear_67_cast_fp16)[name = tensor("input_205_cast_fp16")]; + tensor input_207_axes_0 = const()[name = tensor("input_207_axes_0"), val = tensor([-1])]; + tensor norm17_1_weight_to_fp16 = const()[name = tensor("norm17_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(429037248)))]; + tensor norm17_1_bias_to_fp16 = const()[name = tensor("norm17_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(429039360)))]; + tensor var_7179_to_fp16 = const()[name = tensor("op_7179_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_207_cast_fp16 = layer_norm(axes = input_207_axes_0, beta = norm17_1_bias_to_fp16, epsilon = var_7179_to_fp16, gamma = norm17_1_weight_to_fp16, x = input_205_cast_fp16)[name = tensor("input_207_cast_fp16")]; + tensor attn17_in_proj_weight_to_fp16 = const()[name = tensor("attn17_in_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(429041472)))]; + tensor linear_68_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = attn17_in_proj_weight_to_fp16, x = input_207_cast_fp16)[name = tensor("linear_68_cast_fp16")]; + tensor var_7193 = const()[name = tensor("op_7193"), val = tensor([1, 256, 3, 16, 64])]; + tensor qkv_35_cast_fp16 = reshape(shape = var_7193, x = linear_68_cast_fp16)[name = tensor("qkv_35_cast_fp16")]; + tensor q_103_begin_0 = const()[name = tensor("q_103_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor q_103_end_0 = const()[name = tensor("q_103_end_0"), val = tensor([1, 256, 1, 16, 64])]; + tensor q_103_end_mask_0 = const()[name = tensor("q_103_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor q_103_squeeze_mask_0 = const()[name = tensor("q_103_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor q_103_cast_fp16 = slice_by_index(begin = q_103_begin_0, end = q_103_end_0, end_mask = q_103_end_mask_0, squeeze_mask = q_103_squeeze_mask_0, x = qkv_35_cast_fp16)[name = tensor("q_103_cast_fp16")]; + tensor k_69_begin_0 = const()[name = tensor("k_69_begin_0"), val = tensor([0, 0, 1, 0, 0])]; + tensor k_69_end_0 = const()[name = tensor("k_69_end_0"), val = tensor([1, 256, 2, 16, 64])]; + tensor k_69_end_mask_0 = const()[name = tensor("k_69_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor k_69_squeeze_mask_0 = const()[name = tensor("k_69_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor k_69_cast_fp16 = slice_by_index(begin = k_69_begin_0, end = k_69_end_0, end_mask = k_69_end_mask_0, squeeze_mask = k_69_squeeze_mask_0, x = qkv_35_cast_fp16)[name = tensor("k_69_cast_fp16")]; + tensor v_35_begin_0 = const()[name = tensor("v_35_begin_0"), val = tensor([0, 0, 2, 0, 0])]; + tensor v_35_end_0 = const()[name = tensor("v_35_end_0"), val = tensor([1, 256, 3, 16, 64])]; + tensor v_35_end_mask_0 = const()[name = tensor("v_35_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor v_35_squeeze_mask_0 = const()[name = tensor("v_35_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor v_35_cast_fp16 = slice_by_index(begin = v_35_begin_0, end = v_35_end_0, end_mask = v_35_end_mask_0, squeeze_mask = v_35_squeeze_mask_0, x = qkv_35_cast_fp16)[name = tensor("v_35_cast_fp16")]; + tensor var_7251 = const()[name = tensor("op_7251"), val = tensor([1, 1])]; + tensor position17_to_fp16_dtype_0 = const()[name = tensor("position17_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor position17_to_fp16 = cast(dtype = position17_to_fp16_dtype_0, x = position17)[name = tensor("cast_576")]; + tensor var_7252_cast_fp16 = reshape(shape = var_7251, x = position17_to_fp16)[name = tensor("op_7252_cast_fp16")]; + tensor ts_105_cast_fp16 = add(x = ts_1_promoted_to_fp16, y = var_7252_cast_fp16)[name = tensor("ts_105_cast_fp16")]; + tensor var_7259 = const()[name = tensor("op_7259"), val = tensor([1, 256, 1, 1])]; + tensor ts_107_cast_fp16 = reshape(shape = var_7259, x = ts_105_cast_fp16)[name = tensor("ts_107_cast_fp16")]; + tensor var_7266 = const()[name = tensor("op_7266"), val = tensor([1, 256, 16, 32, 2])]; + tensor q_complex_35_cast_fp16 = reshape(shape = var_7266, x = q_103_cast_fp16)[name = tensor("q_complex_35_cast_fp16")]; + tensor var_7273 = const()[name = tensor("op_7273"), val = tensor([1, 256, 16, 32, 2])]; + tensor k_complex_35_cast_fp16 = reshape(shape = var_7273, x = k_69_cast_fp16)[name = tensor("k_complex_35_cast_fp16")]; + tensor var_7277_begin_0 = const()[name = tensor("op_7277_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_7277_end_0 = const()[name = tensor("op_7277_end_0"), val = tensor([1, 256, 16, 32, 1])]; + tensor var_7277_end_mask_0 = const()[name = tensor("op_7277_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_7277_squeeze_mask_0 = const()[name = tensor("op_7277_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_7277_cast_fp16 = slice_by_index(begin = var_7277_begin_0, end = var_7277_end_0, end_mask = var_7277_end_mask_0, squeeze_mask = var_7277_squeeze_mask_0, x = q_complex_35_cast_fp16)[name = tensor("op_7277_cast_fp16")]; + tensor var_7285_begin_0 = const()[name = tensor("op_7285_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_7285_end_0 = const()[name = tensor("op_7285_end_0"), val = tensor([1, 256, 16, 32, 2])]; + tensor var_7285_end_mask_0 = const()[name = tensor("op_7285_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_7285_squeeze_mask_0 = const()[name = tensor("op_7285_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_7285_cast_fp16 = slice_by_index(begin = var_7285_begin_0, end = var_7285_end_0, end_mask = var_7285_end_mask_0, squeeze_mask = var_7285_squeeze_mask_0, x = q_complex_35_cast_fp16)[name = tensor("op_7285_cast_fp16")]; + tensor var_7293_begin_0 = const()[name = tensor("op_7293_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_7293_end_0 = const()[name = tensor("op_7293_end_0"), val = tensor([1, 256, 16, 32, 1])]; + tensor var_7293_end_mask_0 = const()[name = tensor("op_7293_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_7293_squeeze_mask_0 = const()[name = tensor("op_7293_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_7293_cast_fp16 = slice_by_index(begin = var_7293_begin_0, end = var_7293_end_0, end_mask = var_7293_end_mask_0, squeeze_mask = var_7293_squeeze_mask_0, x = k_complex_35_cast_fp16)[name = tensor("op_7293_cast_fp16")]; + tensor var_7301_begin_0 = const()[name = tensor("op_7301_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_7301_end_0 = const()[name = tensor("op_7301_end_0"), val = tensor([1, 256, 16, 32, 2])]; + tensor var_7301_end_mask_0 = const()[name = tensor("op_7301_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_7301_squeeze_mask_0 = const()[name = tensor("op_7301_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_7301_cast_fp16 = slice_by_index(begin = var_7301_begin_0, end = var_7301_end_0, end_mask = var_7301_end_mask_0, squeeze_mask = var_7301_squeeze_mask_0, x = k_complex_35_cast_fp16)[name = tensor("op_7301_cast_fp16")]; + tensor freqs_35_to_fp16 = const()[name = tensor("freqs_35_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(435332992)))]; + tensor var_7307_cast_fp16 = mul(x = freqs_35_to_fp16, y = ts_107_cast_fp16)[name = tensor("op_7307_cast_fp16")]; + tensor rotr_35_cast_fp16 = cos(x = var_7307_cast_fp16)[name = tensor("rotr_35_cast_fp16")]; + tensor roti_35_cast_fp16 = sin(x = var_7307_cast_fp16)[name = tensor("roti_35_cast_fp16")]; + tensor var_7311_cast_fp16 = mul(x = var_7277_cast_fp16, y = rotr_35_cast_fp16)[name = tensor("op_7311_cast_fp16")]; + tensor var_7312_cast_fp16 = mul(x = var_7285_cast_fp16, y = roti_35_cast_fp16)[name = tensor("op_7312_cast_fp16")]; + tensor qor_69_cast_fp16 = sub(x = var_7311_cast_fp16, y = var_7312_cast_fp16)[name = tensor("qor_69_cast_fp16")]; + tensor var_7315_cast_fp16 = mul(x = var_7277_cast_fp16, y = roti_35_cast_fp16)[name = tensor("op_7315_cast_fp16")]; + tensor var_7316_cast_fp16 = mul(x = var_7285_cast_fp16, y = rotr_35_cast_fp16)[name = tensor("op_7316_cast_fp16")]; + tensor qoi_69_cast_fp16 = add(x = var_7315_cast_fp16, y = var_7316_cast_fp16)[name = tensor("qoi_69_cast_fp16")]; + tensor var_7319_cast_fp16 = mul(x = var_7293_cast_fp16, y = rotr_35_cast_fp16)[name = tensor("op_7319_cast_fp16")]; + tensor var_7320_cast_fp16 = mul(x = var_7301_cast_fp16, y = roti_35_cast_fp16)[name = tensor("op_7320_cast_fp16")]; + tensor kor_69_cast_fp16 = sub(x = var_7319_cast_fp16, y = var_7320_cast_fp16)[name = tensor("kor_69_cast_fp16")]; + tensor var_7323_cast_fp16 = mul(x = var_7293_cast_fp16, y = roti_35_cast_fp16)[name = tensor("op_7323_cast_fp16")]; + tensor var_7324_cast_fp16 = mul(x = var_7301_cast_fp16, y = rotr_35_cast_fp16)[name = tensor("op_7324_cast_fp16")]; + tensor koi_69_cast_fp16 = add(x = var_7323_cast_fp16, y = var_7324_cast_fp16)[name = tensor("koi_69_cast_fp16")]; + tensor qo_35_axis_0 = const()[name = tensor("qo_35_axis_0"), val = tensor(-1)]; + tensor qo_35_cast_fp16 = stack(axis = qo_35_axis_0, values = (qor_69_cast_fp16, qoi_69_cast_fp16))[name = tensor("qo_35_cast_fp16")]; + tensor ko_35_axis_0 = const()[name = tensor("ko_35_axis_0"), val = tensor(-1)]; + tensor ko_35_cast_fp16 = stack(axis = ko_35_axis_0, values = (kor_69_cast_fp16, koi_69_cast_fp16))[name = tensor("ko_35_cast_fp16")]; + tensor var_7357 = const()[name = tensor("op_7357"), val = tensor([1, 256, 16, 64])]; + tensor q_105_cast_fp16 = reshape(shape = var_7357, x = qo_35_cast_fp16)[name = tensor("q_105_cast_fp16")]; + tensor var_7363 = const()[name = tensor("op_7363"), val = tensor([1, 256, 16, 64])]; + tensor k_71_cast_fp16 = reshape(shape = var_7363, x = ko_35_cast_fp16)[name = tensor("k_71_cast_fp16")]; + tensor write_indices_float_103_cast_fp16 = add(x = var_7252_cast_fp16, y = var_393_promoted_to_fp16)[name = tensor("write_indices_float_103_cast_fp16")]; + tensor _inversed_7384_y_0_to_fp16 = const()[name = tensor("_inversed_7384_y_0_to_fp16"), val = tensor(0x1p-9)]; + tensor _inversed_7384_cast_fp16 = mul(x = write_indices_float_103_cast_fp16, y = _inversed_7384_y_0_to_fp16)[name = tensor("_inversed_7384_cast_fp16")]; + tensor var_7385_cast_fp16 = floor(x = _inversed_7384_cast_fp16)[name = tensor("op_7385_cast_fp16")]; + tensor var_7386_to_fp16 = const()[name = tensor("op_7386_to_fp16"), val = tensor(0x1p+9)]; + tensor var_7387_cast_fp16 = mul(x = var_7385_cast_fp16, y = var_7386_to_fp16)[name = tensor("op_7387_cast_fp16")]; + tensor write_indices_float_105_cast_fp16 = sub(x = write_indices_float_103_cast_fp16, y = var_7387_cast_fp16)[name = tensor("write_indices_float_105_cast_fp16")]; + tensor var_7405_value_0_to_fp16 = const()[name = tensor("op_7405_value_0_to_fp16"), val = tensor(0x1.ffp+8)]; + tensor var_7405_cast_fp16 = fill_like(ref_tensor = write_indices_float_105_cast_fp16, value = var_7405_value_0_to_fp16)[name = tensor("op_7405_cast_fp16")]; + tensor write_indices_float_107_cast_fp16 = select(a = write_indices_float_105_cast_fp16, b = var_7405_cast_fp16, cond = token_valid_1_cast_fp16)[name = tensor("write_indices_float_107_cast_fp16")]; + tensor var_7411_dtype_0 = const()[name = tensor("op_7411_dtype_0"), val = tensor("int32")]; + tensor var_7416 = const()[name = tensor("op_7416"), val = tensor([1, 256, 1, 1])]; + tensor write_indices_float_107_cast_fp16_to_int32 = cast(dtype = var_7411_dtype_0, x = write_indices_float_107_cast_fp16)[name = tensor("cast_575")]; + tensor var_7417 = reshape(shape = var_7416, x = write_indices_float_107_cast_fp16_to_int32)[name = tensor("op_7417")]; + tensor write_indices_35_reps_0 = const()[name = tensor("write_indices_35_reps_0"), val = tensor([1, 1, 16, 64])]; + tensor write_indices_35 = tile(reps = write_indices_35_reps_0, x = var_7417)[name = tensor("write_indices_35")]; + tensor var_7427_begin_0 = const()[name = tensor("op_7427_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_7427_end_0 = const()[name = tensor("op_7427_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor var_7427_end_mask_0 = const()[name = tensor("op_7427_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_7427_squeeze_mask_0 = const()[name = tensor("op_7427_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor cache17_to_fp16_dtype_0 = const()[name = tensor("cache17_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor cache17_to_fp16 = cast(dtype = cache17_to_fp16_dtype_0, x = cache17)[name = tensor("cast_574")]; + tensor var_7427_cast_fp16 = slice_by_index(begin = var_7427_begin_0, end = var_7427_end_0, end_mask = var_7427_end_mask_0, squeeze_mask = var_7427_squeeze_mask_0, x = cache17_to_fp16)[name = tensor("op_7427_cast_fp16")]; + tensor var_7429_axis_0 = const()[name = tensor("op_7429_axis_0"), val = tensor(1)]; + tensor var_7429_mode_0 = const()[name = tensor("op_7429_mode_0"), val = tensor("update")]; + tensor var_7429_validate_indices_0 = const()[name = tensor("op_7429_validate_indices_0"), val = tensor(false)]; + tensor var_7429_cast_fp16 = scatter_along_axis(axis = var_7429_axis_0, data = var_7427_cast_fp16, indices = write_indices_35, mode = var_7429_mode_0, updates = k_71_cast_fp16, validate_indices = var_7429_validate_indices_0)[name = tensor("op_7429_cast_fp16")]; + tensor concat_138 = const()[name = tensor("concat_138"), val = tensor([0, 0, 0, 0, 0])]; + tensor concat_139 = const()[name = tensor("concat_139"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_35_internal_tensor_assign_1_stride_0 = const()[name = tensor("new_cache_35_internal_tensor_assign_1_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_35_internal_tensor_assign_1_begin_mask_0 = const()[name = tensor("new_cache_35_internal_tensor_assign_1_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_35_internal_tensor_assign_1_end_mask_0 = const()[name = tensor("new_cache_35_internal_tensor_assign_1_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_35_internal_tensor_assign_1_squeeze_mask_0 = const()[name = tensor("new_cache_35_internal_tensor_assign_1_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_80 = const()[name = tensor("shape_80"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_34 = const()[name = tensor("reduce_prod_34"), val = tensor(1048576)]; + tensor range_1d_34_start_0 = const()[name = tensor("range_1d_34_start_0"), val = tensor(0)]; + tensor range_1d_34_step_0 = const()[name = tensor("range_1d_34_step_0"), val = tensor(1)]; + tensor range_1d_34 = range_1d(end = reduce_prod_34, start = range_1d_34_start_0, step = range_1d_34_step_0)[name = tensor("range_1d_34")]; + tensor reshape_170 = reshape(shape = shape_80, x = range_1d_34)[name = tensor("reshape_170")]; + tensor slice_by_index_34 = slice_by_index(begin = concat_138, begin_mask = new_cache_35_internal_tensor_assign_1_begin_mask_0, end = concat_139, end_mask = new_cache_35_internal_tensor_assign_1_end_mask_0, squeeze_mask = new_cache_35_internal_tensor_assign_1_squeeze_mask_0, stride = new_cache_35_internal_tensor_assign_1_stride_0, x = reshape_170)[name = tensor("slice_by_index_34")]; + tensor reshape_171_shape_0 = const()[name = tensor("reshape_171_shape_0"), val = tensor([-1])]; + tensor reshape_171 = reshape(shape = reshape_171_shape_0, x = slice_by_index_34)[name = tensor("reshape_171")]; + tensor reshape_172_shape_0 = const()[name = tensor("reshape_172_shape_0"), val = tensor([-1])]; + tensor reshape_172_cast_fp16 = reshape(shape = reshape_172_shape_0, x = var_7429_cast_fp16)[name = tensor("reshape_172_cast_fp16")]; + tensor reshape_173_shape_0 = const()[name = tensor("reshape_173_shape_0"), val = tensor([-1])]; + tensor reshape_173_cast_fp16 = reshape(shape = reshape_173_shape_0, x = cache17_to_fp16)[name = tensor("reshape_173_cast_fp16")]; + tensor scatter_34_mode_0 = const()[name = tensor("scatter_34_mode_0"), val = tensor("update")]; + tensor scatter_34_axis_0 = const()[name = tensor("scatter_34_axis_0"), val = tensor(0)]; + tensor scatter_34_validate_indices_0 = const()[name = tensor("scatter_34_validate_indices_0"), val = tensor(false)]; + tensor scatter_34_cast_fp16 = scatter(axis = scatter_34_axis_0, data = reshape_173_cast_fp16, indices = reshape_171, mode = scatter_34_mode_0, updates = reshape_172_cast_fp16, validate_indices = scatter_34_validate_indices_0)[name = tensor("scatter_34_cast_fp16")]; + tensor reshape_174_cast_fp16 = reshape(shape = shape_80, x = scatter_34_cast_fp16)[name = tensor("reshape_174_cast_fp16")]; + tensor var_7437_begin_0 = const()[name = tensor("op_7437_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor var_7437_end_0 = const()[name = tensor("op_7437_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor var_7437_end_mask_0 = const()[name = tensor("op_7437_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_7437_squeeze_mask_0 = const()[name = tensor("op_7437_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor var_7437_cast_fp16 = slice_by_index(begin = var_7437_begin_0, end = var_7437_end_0, end_mask = var_7437_end_mask_0, squeeze_mask = var_7437_squeeze_mask_0, x = reshape_174_cast_fp16)[name = tensor("op_7437_cast_fp16")]; + tensor var_7439_axis_0 = const()[name = tensor("op_7439_axis_0"), val = tensor(1)]; + tensor var_7439_mode_0 = const()[name = tensor("op_7439_mode_0"), val = tensor("update")]; + tensor var_7439_validate_indices_0 = const()[name = tensor("op_7439_validate_indices_0"), val = tensor(false)]; + tensor var_7439_cast_fp16 = scatter_along_axis(axis = var_7439_axis_0, data = var_7437_cast_fp16, indices = write_indices_35, mode = var_7439_mode_0, updates = v_35_cast_fp16, validate_indices = var_7439_validate_indices_0)[name = tensor("op_7439_cast_fp16")]; + tensor concat_140 = const()[name = tensor("concat_140"), val = tensor([1, 0, 0, 0, 0])]; + tensor concat_141 = const()[name = tensor("concat_141"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_35_internal_tensor_assign_2_stride_0 = const()[name = tensor("new_cache_35_internal_tensor_assign_2_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_35_internal_tensor_assign_2_begin_mask_0 = const()[name = tensor("new_cache_35_internal_tensor_assign_2_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_35_internal_tensor_assign_2_end_mask_0 = const()[name = tensor("new_cache_35_internal_tensor_assign_2_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_35_internal_tensor_assign_2_squeeze_mask_0 = const()[name = tensor("new_cache_35_internal_tensor_assign_2_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_81 = const()[name = tensor("shape_81"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_35 = const()[name = tensor("reduce_prod_35"), val = tensor(1048576)]; + tensor range_1d_35_start_0 = const()[name = tensor("range_1d_35_start_0"), val = tensor(0)]; + tensor range_1d_35_step_0 = const()[name = tensor("range_1d_35_step_0"), val = tensor(1)]; + tensor range_1d_35 = range_1d(end = reduce_prod_35, start = range_1d_35_start_0, step = range_1d_35_step_0)[name = tensor("range_1d_35")]; + tensor reshape_175 = reshape(shape = shape_81, x = range_1d_35)[name = tensor("reshape_175")]; + tensor slice_by_index_35 = slice_by_index(begin = concat_140, begin_mask = new_cache_35_internal_tensor_assign_2_begin_mask_0, end = concat_141, end_mask = new_cache_35_internal_tensor_assign_2_end_mask_0, squeeze_mask = new_cache_35_internal_tensor_assign_2_squeeze_mask_0, stride = new_cache_35_internal_tensor_assign_2_stride_0, x = reshape_175)[name = tensor("slice_by_index_35")]; + tensor reshape_176_shape_0 = const()[name = tensor("reshape_176_shape_0"), val = tensor([-1])]; + tensor reshape_176 = reshape(shape = reshape_176_shape_0, x = slice_by_index_35)[name = tensor("reshape_176")]; + tensor reshape_177_shape_0 = const()[name = tensor("reshape_177_shape_0"), val = tensor([-1])]; + tensor reshape_177_cast_fp16 = reshape(shape = reshape_177_shape_0, x = var_7439_cast_fp16)[name = tensor("reshape_177_cast_fp16")]; + tensor reshape_178_shape_0 = const()[name = tensor("reshape_178_shape_0"), val = tensor([-1])]; + tensor reshape_178_cast_fp16 = reshape(shape = reshape_178_shape_0, x = reshape_174_cast_fp16)[name = tensor("reshape_178_cast_fp16")]; + tensor scatter_35_mode_0 = const()[name = tensor("scatter_35_mode_0"), val = tensor("update")]; + tensor scatter_35_axis_0 = const()[name = tensor("scatter_35_axis_0"), val = tensor(0)]; + tensor scatter_35_validate_indices_0 = const()[name = tensor("scatter_35_validate_indices_0"), val = tensor(false)]; + tensor scatter_35_cast_fp16 = scatter(axis = scatter_35_axis_0, data = reshape_178_cast_fp16, indices = reshape_176, mode = scatter_35_mode_0, updates = reshape_177_cast_fp16, validate_indices = scatter_35_validate_indices_0)[name = tensor("scatter_35_cast_fp16")]; + tensor reshape_179_cast_fp16 = reshape(shape = shape_81, x = scatter_35_cast_fp16)[name = tensor("reshape_179_cast_fp16")]; + tensor reshape_179_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("reshape_179_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor keys_103_begin_0 = const()[name = tensor("keys_103_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor keys_103_end_0 = const()[name = tensor("keys_103_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor keys_103_end_mask_0 = const()[name = tensor("keys_103_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor keys_103_squeeze_mask_0 = const()[name = tensor("keys_103_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor keys_103_cast_fp16 = slice_by_index(begin = keys_103_begin_0, end = keys_103_end_0, end_mask = keys_103_end_mask_0, squeeze_mask = keys_103_squeeze_mask_0, x = reshape_179_cast_fp16)[name = tensor("keys_103_cast_fp16")]; + tensor values_103_begin_0 = const()[name = tensor("values_103_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor values_103_end_0 = const()[name = tensor("values_103_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor values_103_end_mask_0 = const()[name = tensor("values_103_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor values_103_squeeze_mask_0 = const()[name = tensor("values_103_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor values_103_cast_fp16 = slice_by_index(begin = values_103_begin_0, end = values_103_end_0, end_mask = values_103_end_mask_0, squeeze_mask = values_103_squeeze_mask_0, x = reshape_179_cast_fp16)[name = tensor("values_103_cast_fp16")]; + tensor var_7451_cast_fp16 = not_equal(x = keys_103_cast_fp16, y = keys_103_cast_fp16)[name = tensor("op_7451_cast_fp16")]; + tensor keys_105_cast_fp16 = select(a = var_470_to_fp16, b = keys_103_cast_fp16, cond = var_7451_cast_fp16)[name = tensor("keys_105_cast_fp16")]; + tensor var_7459_cast_fp16 = not_equal(x = values_103_cast_fp16, y = values_103_cast_fp16)[name = tensor("op_7459_cast_fp16")]; + tensor values_105_cast_fp16 = select(a = var_470_to_fp16, b = values_103_cast_fp16, cond = var_7459_cast_fp16)[name = tensor("values_105_cast_fp16")]; + tensor var_7483 = const()[name = tensor("op_7483"), val = tensor([0, 2, 1, 3])]; + tensor var_7499 = const()[name = tensor("op_7499"), val = tensor([1, 1, 1])]; + tensor var_7500_cast_fp16 = reshape(shape = var_7499, x = position17_to_fp16)[name = tensor("op_7500_cast_fp16")]; + tensor q_positions_35_cast_fp16 = add(x = var_7500_cast_fp16, y = q_offsets_1_promoted_to_fp16)[name = tensor("q_positions_35_cast_fp16")]; + tensor valid_end_35_cast_fp16 = add(x = var_7500_cast_fp16, y = var_536_cast_fp16)[name = tensor("valid_end_35_cast_fp16")]; + tensor valid_mask_35_cast_fp16 = less(x = k_positions_1_promoted_to_fp16, y = valid_end_35_cast_fp16)[name = tensor("valid_mask_35_cast_fp16")]; + tensor causal_mask_35_cast_fp16 = less_equal(x = k_positions_1_promoted_to_fp16, y = q_positions_35_cast_fp16)[name = tensor("causal_mask_35_cast_fp16")]; + tensor var_7528 = logical_and(x = valid_mask_35_cast_fp16, y = causal_mask_35_cast_fp16)[name = tensor("op_7528")]; + tensor attn_mask_35_axes_0 = const()[name = tensor("attn_mask_35_axes_0"), val = tensor([1])]; + tensor attn_mask_35 = expand_dims(axes = attn_mask_35_axes_0, x = var_7528)[name = tensor("attn_mask_35")]; + tensor var_7542_transpose_x_0 = const()[name = tensor("op_7542_transpose_x_0"), val = tensor(false)]; + tensor var_7542_transpose_y_0 = const()[name = tensor("op_7542_transpose_y_0"), val = tensor(false)]; + tensor transpose_103_perm_0 = const()[name = tensor("transpose_103_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_104_perm_0 = const()[name = tensor("transpose_104_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_104 = transpose(perm = transpose_104_perm_0, x = keys_105_cast_fp16)[name = tensor("transpose_136")]; + tensor transpose_103 = transpose(perm = transpose_103_perm_0, x = q_105_cast_fp16)[name = tensor("transpose_137")]; + tensor var_7542_cast_fp16 = matmul(transpose_x = var_7542_transpose_x_0, transpose_y = var_7542_transpose_y_0, x = transpose_103, y = transpose_104)[name = tensor("op_7542_cast_fp16")]; + tensor var_7536_to_fp16 = const()[name = tensor("op_7536_to_fp16"), val = tensor([0x1p-3])]; + tensor attn_weights_103_cast_fp16 = mul(x = var_7542_cast_fp16, y = var_7536_to_fp16)[name = tensor("attn_weights_103_cast_fp16")]; + tensor var_7544 = logical_not(x = attn_mask_35)[name = tensor("op_7544")]; + tensor var_7545_to_fp16 = const()[name = tensor("op_7545_to_fp16"), val = tensor(-inf)]; + tensor attn_weights_105_cast_fp16 = select(a = var_7545_to_fp16, b = attn_weights_103_cast_fp16, cond = var_7544)[name = tensor("attn_weights_105_cast_fp16")]; + tensor var_7547 = const()[name = tensor("op_7547"), val = tensor(-1)]; + tensor attn_weights_107_cast_fp16 = softmax(axis = var_7547, x = attn_weights_105_cast_fp16)[name = tensor("attn_weights_107_cast_fp16")]; + tensor attn_output_35_transpose_x_0 = const()[name = tensor("attn_output_35_transpose_x_0"), val = tensor(false)]; + tensor attn_output_35_transpose_y_0 = const()[name = tensor("attn_output_35_transpose_y_0"), val = tensor(false)]; + tensor values_107_cast_fp16 = transpose(perm = var_7483, x = values_105_cast_fp16)[name = tensor("transpose_138")]; + tensor attn_output_35_cast_fp16 = matmul(transpose_x = attn_output_35_transpose_x_0, transpose_y = attn_output_35_transpose_y_0, x = attn_weights_107_cast_fp16, y = values_107_cast_fp16)[name = tensor("attn_output_35_cast_fp16")]; + tensor var_7555 = const()[name = tensor("op_7555"), val = tensor([0, 2, 1, 3])]; + tensor var_7560 = const()[name = tensor("op_7560"), val = tensor([1, 256, 1024])]; + tensor var_7556_cast_fp16 = transpose(perm = var_7555, x = attn_output_35_cast_fp16)[name = tensor("transpose_135")]; + tensor input_209_cast_fp16 = reshape(shape = var_7560, x = var_7556_cast_fp16)[name = tensor("input_209_cast_fp16")]; + tensor attn17_out_proj_weight_to_fp16 = const()[name = tensor("attn17_out_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(435333120)))]; + tensor linear_69_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = attn17_out_proj_weight_to_fp16, x = input_209_cast_fp16)[name = tensor("linear_69_cast_fp16")]; + tensor var_7569_cast_fp16 = add(x = position17_to_fp16, y = valid_len_to_fp16)[name = tensor("op_7569_cast_fp16")]; + tensor var_7569_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_7569_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor input_211_cast_fp16 = add(x = input_205_cast_fp16, y = linear_69_cast_fp16)[name = tensor("input_211_cast_fp16")]; + tensor input_213_axes_0 = const()[name = tensor("input_213_axes_0"), val = tensor([-1])]; + tensor norm17_2_weight_to_fp16 = const()[name = tensor("norm17_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(437430336)))]; + tensor norm17_2_bias_to_fp16 = const()[name = tensor("norm17_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(437432448)))]; + tensor var_7573_to_fp16 = const()[name = tensor("op_7573_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_213_cast_fp16 = layer_norm(axes = input_213_axes_0, beta = norm17_2_bias_to_fp16, epsilon = var_7573_to_fp16, gamma = norm17_2_weight_to_fp16, x = input_211_cast_fp16)[name = tensor("input_213_cast_fp16")]; + tensor linear17_1_weight_to_fp16 = const()[name = tensor("linear17_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(437434560)))]; + tensor linear_70_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = linear17_1_weight_to_fp16, x = input_213_cast_fp16)[name = tensor("linear_70_cast_fp16")]; + tensor input_215_mode_0 = const()[name = tensor("input_215_mode_0"), val = tensor("EXACT")]; + tensor input_215_cast_fp16 = gelu(mode = input_215_mode_0, x = linear_70_cast_fp16)[name = tensor("input_215_cast_fp16")]; + tensor linear17_2_weight_to_fp16 = const()[name = tensor("linear17_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(445823232)))]; + tensor linear_71_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = linear17_2_weight_to_fp16, x = input_215_cast_fp16)[name = tensor("linear_71_cast_fp16")]; + tensor input_217_cast_fp16 = add(x = input_211_cast_fp16, y = linear_71_cast_fp16)[name = tensor("input_217_cast_fp16")]; + tensor input_219_axes_0 = const()[name = tensor("input_219_axes_0"), val = tensor([-1])]; + tensor norm18_1_weight_to_fp16 = const()[name = tensor("norm18_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(454211904)))]; + tensor norm18_1_bias_to_fp16 = const()[name = tensor("norm18_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(454214016)))]; + tensor var_7590_to_fp16 = const()[name = tensor("op_7590_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_219_cast_fp16 = layer_norm(axes = input_219_axes_0, beta = norm18_1_bias_to_fp16, epsilon = var_7590_to_fp16, gamma = norm18_1_weight_to_fp16, x = input_217_cast_fp16)[name = tensor("input_219_cast_fp16")]; + tensor attn18_in_proj_weight_to_fp16 = const()[name = tensor("attn18_in_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(454216128)))]; + tensor linear_72_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = attn18_in_proj_weight_to_fp16, x = input_219_cast_fp16)[name = tensor("linear_72_cast_fp16")]; + tensor var_7604 = const()[name = tensor("op_7604"), val = tensor([1, 256, 3, 16, 64])]; + tensor qkv_37_cast_fp16 = reshape(shape = var_7604, x = linear_72_cast_fp16)[name = tensor("qkv_37_cast_fp16")]; + tensor q_109_begin_0 = const()[name = tensor("q_109_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor q_109_end_0 = const()[name = tensor("q_109_end_0"), val = tensor([1, 256, 1, 16, 64])]; + tensor q_109_end_mask_0 = const()[name = tensor("q_109_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor q_109_squeeze_mask_0 = const()[name = tensor("q_109_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor q_109_cast_fp16 = slice_by_index(begin = q_109_begin_0, end = q_109_end_0, end_mask = q_109_end_mask_0, squeeze_mask = q_109_squeeze_mask_0, x = qkv_37_cast_fp16)[name = tensor("q_109_cast_fp16")]; + tensor k_73_begin_0 = const()[name = tensor("k_73_begin_0"), val = tensor([0, 0, 1, 0, 0])]; + tensor k_73_end_0 = const()[name = tensor("k_73_end_0"), val = tensor([1, 256, 2, 16, 64])]; + tensor k_73_end_mask_0 = const()[name = tensor("k_73_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor k_73_squeeze_mask_0 = const()[name = tensor("k_73_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor k_73_cast_fp16 = slice_by_index(begin = k_73_begin_0, end = k_73_end_0, end_mask = k_73_end_mask_0, squeeze_mask = k_73_squeeze_mask_0, x = qkv_37_cast_fp16)[name = tensor("k_73_cast_fp16")]; + tensor v_37_begin_0 = const()[name = tensor("v_37_begin_0"), val = tensor([0, 0, 2, 0, 0])]; + tensor v_37_end_0 = const()[name = tensor("v_37_end_0"), val = tensor([1, 256, 3, 16, 64])]; + tensor v_37_end_mask_0 = const()[name = tensor("v_37_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor v_37_squeeze_mask_0 = const()[name = tensor("v_37_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor v_37_cast_fp16 = slice_by_index(begin = v_37_begin_0, end = v_37_end_0, end_mask = v_37_end_mask_0, squeeze_mask = v_37_squeeze_mask_0, x = qkv_37_cast_fp16)[name = tensor("v_37_cast_fp16")]; + tensor var_7662 = const()[name = tensor("op_7662"), val = tensor([1, 1])]; + tensor position18_to_fp16_dtype_0 = const()[name = tensor("position18_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor position18_to_fp16 = cast(dtype = position18_to_fp16_dtype_0, x = position18)[name = tensor("cast_571")]; + tensor var_7663_cast_fp16 = reshape(shape = var_7662, x = position18_to_fp16)[name = tensor("op_7663_cast_fp16")]; + tensor ts_111_cast_fp16 = add(x = ts_1_promoted_to_fp16, y = var_7663_cast_fp16)[name = tensor("ts_111_cast_fp16")]; + tensor var_7670 = const()[name = tensor("op_7670"), val = tensor([1, 256, 1, 1])]; + tensor ts_113_cast_fp16 = reshape(shape = var_7670, x = ts_111_cast_fp16)[name = tensor("ts_113_cast_fp16")]; + tensor var_7677 = const()[name = tensor("op_7677"), val = tensor([1, 256, 16, 32, 2])]; + tensor q_complex_37_cast_fp16 = reshape(shape = var_7677, x = q_109_cast_fp16)[name = tensor("q_complex_37_cast_fp16")]; + tensor var_7684 = const()[name = tensor("op_7684"), val = tensor([1, 256, 16, 32, 2])]; + tensor k_complex_37_cast_fp16 = reshape(shape = var_7684, x = k_73_cast_fp16)[name = tensor("k_complex_37_cast_fp16")]; + tensor var_7688_begin_0 = const()[name = tensor("op_7688_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_7688_end_0 = const()[name = tensor("op_7688_end_0"), val = tensor([1, 256, 16, 32, 1])]; + tensor var_7688_end_mask_0 = const()[name = tensor("op_7688_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_7688_squeeze_mask_0 = const()[name = tensor("op_7688_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_7688_cast_fp16 = slice_by_index(begin = var_7688_begin_0, end = var_7688_end_0, end_mask = var_7688_end_mask_0, squeeze_mask = var_7688_squeeze_mask_0, x = q_complex_37_cast_fp16)[name = tensor("op_7688_cast_fp16")]; + tensor var_7696_begin_0 = const()[name = tensor("op_7696_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_7696_end_0 = const()[name = tensor("op_7696_end_0"), val = tensor([1, 256, 16, 32, 2])]; + tensor var_7696_end_mask_0 = const()[name = tensor("op_7696_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_7696_squeeze_mask_0 = const()[name = tensor("op_7696_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_7696_cast_fp16 = slice_by_index(begin = var_7696_begin_0, end = var_7696_end_0, end_mask = var_7696_end_mask_0, squeeze_mask = var_7696_squeeze_mask_0, x = q_complex_37_cast_fp16)[name = tensor("op_7696_cast_fp16")]; + tensor var_7704_begin_0 = const()[name = tensor("op_7704_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_7704_end_0 = const()[name = tensor("op_7704_end_0"), val = tensor([1, 256, 16, 32, 1])]; + tensor var_7704_end_mask_0 = const()[name = tensor("op_7704_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_7704_squeeze_mask_0 = const()[name = tensor("op_7704_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_7704_cast_fp16 = slice_by_index(begin = var_7704_begin_0, end = var_7704_end_0, end_mask = var_7704_end_mask_0, squeeze_mask = var_7704_squeeze_mask_0, x = k_complex_37_cast_fp16)[name = tensor("op_7704_cast_fp16")]; + tensor var_7712_begin_0 = const()[name = tensor("op_7712_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_7712_end_0 = const()[name = tensor("op_7712_end_0"), val = tensor([1, 256, 16, 32, 2])]; + tensor var_7712_end_mask_0 = const()[name = tensor("op_7712_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_7712_squeeze_mask_0 = const()[name = tensor("op_7712_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_7712_cast_fp16 = slice_by_index(begin = var_7712_begin_0, end = var_7712_end_0, end_mask = var_7712_end_mask_0, squeeze_mask = var_7712_squeeze_mask_0, x = k_complex_37_cast_fp16)[name = tensor("op_7712_cast_fp16")]; + tensor freqs_37_to_fp16 = const()[name = tensor("freqs_37_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(460507648)))]; + tensor var_7718_cast_fp16 = mul(x = freqs_37_to_fp16, y = ts_113_cast_fp16)[name = tensor("op_7718_cast_fp16")]; + tensor rotr_37_cast_fp16 = cos(x = var_7718_cast_fp16)[name = tensor("rotr_37_cast_fp16")]; + tensor roti_37_cast_fp16 = sin(x = var_7718_cast_fp16)[name = tensor("roti_37_cast_fp16")]; + tensor var_7722_cast_fp16 = mul(x = var_7688_cast_fp16, y = rotr_37_cast_fp16)[name = tensor("op_7722_cast_fp16")]; + tensor var_7723_cast_fp16 = mul(x = var_7696_cast_fp16, y = roti_37_cast_fp16)[name = tensor("op_7723_cast_fp16")]; + tensor qor_73_cast_fp16 = sub(x = var_7722_cast_fp16, y = var_7723_cast_fp16)[name = tensor("qor_73_cast_fp16")]; + tensor var_7726_cast_fp16 = mul(x = var_7688_cast_fp16, y = roti_37_cast_fp16)[name = tensor("op_7726_cast_fp16")]; + tensor var_7727_cast_fp16 = mul(x = var_7696_cast_fp16, y = rotr_37_cast_fp16)[name = tensor("op_7727_cast_fp16")]; + tensor qoi_73_cast_fp16 = add(x = var_7726_cast_fp16, y = var_7727_cast_fp16)[name = tensor("qoi_73_cast_fp16")]; + tensor var_7730_cast_fp16 = mul(x = var_7704_cast_fp16, y = rotr_37_cast_fp16)[name = tensor("op_7730_cast_fp16")]; + tensor var_7731_cast_fp16 = mul(x = var_7712_cast_fp16, y = roti_37_cast_fp16)[name = tensor("op_7731_cast_fp16")]; + tensor kor_73_cast_fp16 = sub(x = var_7730_cast_fp16, y = var_7731_cast_fp16)[name = tensor("kor_73_cast_fp16")]; + tensor var_7734_cast_fp16 = mul(x = var_7704_cast_fp16, y = roti_37_cast_fp16)[name = tensor("op_7734_cast_fp16")]; + tensor var_7735_cast_fp16 = mul(x = var_7712_cast_fp16, y = rotr_37_cast_fp16)[name = tensor("op_7735_cast_fp16")]; + tensor koi_73_cast_fp16 = add(x = var_7734_cast_fp16, y = var_7735_cast_fp16)[name = tensor("koi_73_cast_fp16")]; + tensor qo_37_axis_0 = const()[name = tensor("qo_37_axis_0"), val = tensor(-1)]; + tensor qo_37_cast_fp16 = stack(axis = qo_37_axis_0, values = (qor_73_cast_fp16, qoi_73_cast_fp16))[name = tensor("qo_37_cast_fp16")]; + tensor ko_37_axis_0 = const()[name = tensor("ko_37_axis_0"), val = tensor(-1)]; + tensor ko_37_cast_fp16 = stack(axis = ko_37_axis_0, values = (kor_73_cast_fp16, koi_73_cast_fp16))[name = tensor("ko_37_cast_fp16")]; + tensor var_7768 = const()[name = tensor("op_7768"), val = tensor([1, 256, 16, 64])]; + tensor q_111_cast_fp16 = reshape(shape = var_7768, x = qo_37_cast_fp16)[name = tensor("q_111_cast_fp16")]; + tensor var_7774 = const()[name = tensor("op_7774"), val = tensor([1, 256, 16, 64])]; + tensor k_75_cast_fp16 = reshape(shape = var_7774, x = ko_37_cast_fp16)[name = tensor("k_75_cast_fp16")]; + tensor write_indices_float_109_cast_fp16 = add(x = var_7663_cast_fp16, y = var_393_promoted_to_fp16)[name = tensor("write_indices_float_109_cast_fp16")]; + tensor _inversed_7795_y_0_to_fp16 = const()[name = tensor("_inversed_7795_y_0_to_fp16"), val = tensor(0x1p-9)]; + tensor _inversed_7795_cast_fp16 = mul(x = write_indices_float_109_cast_fp16, y = _inversed_7795_y_0_to_fp16)[name = tensor("_inversed_7795_cast_fp16")]; + tensor var_7796_cast_fp16 = floor(x = _inversed_7795_cast_fp16)[name = tensor("op_7796_cast_fp16")]; + tensor var_7797_to_fp16 = const()[name = tensor("op_7797_to_fp16"), val = tensor(0x1p+9)]; + tensor var_7798_cast_fp16 = mul(x = var_7796_cast_fp16, y = var_7797_to_fp16)[name = tensor("op_7798_cast_fp16")]; + tensor write_indices_float_111_cast_fp16 = sub(x = write_indices_float_109_cast_fp16, y = var_7798_cast_fp16)[name = tensor("write_indices_float_111_cast_fp16")]; + tensor var_7816_value_0_to_fp16 = const()[name = tensor("op_7816_value_0_to_fp16"), val = tensor(0x1.ffp+8)]; + tensor var_7816_cast_fp16 = fill_like(ref_tensor = write_indices_float_111_cast_fp16, value = var_7816_value_0_to_fp16)[name = tensor("op_7816_cast_fp16")]; + tensor write_indices_float_113_cast_fp16 = select(a = write_indices_float_111_cast_fp16, b = var_7816_cast_fp16, cond = token_valid_1_cast_fp16)[name = tensor("write_indices_float_113_cast_fp16")]; + tensor var_7822_dtype_0 = const()[name = tensor("op_7822_dtype_0"), val = tensor("int32")]; + tensor var_7827 = const()[name = tensor("op_7827"), val = tensor([1, 256, 1, 1])]; + tensor write_indices_float_113_cast_fp16_to_int32 = cast(dtype = var_7822_dtype_0, x = write_indices_float_113_cast_fp16)[name = tensor("cast_570")]; + tensor var_7828 = reshape(shape = var_7827, x = write_indices_float_113_cast_fp16_to_int32)[name = tensor("op_7828")]; + tensor write_indices_37_reps_0 = const()[name = tensor("write_indices_37_reps_0"), val = tensor([1, 1, 16, 64])]; + tensor write_indices_37 = tile(reps = write_indices_37_reps_0, x = var_7828)[name = tensor("write_indices_37")]; + tensor var_7838_begin_0 = const()[name = tensor("op_7838_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_7838_end_0 = const()[name = tensor("op_7838_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor var_7838_end_mask_0 = const()[name = tensor("op_7838_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_7838_squeeze_mask_0 = const()[name = tensor("op_7838_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor cache18_to_fp16_dtype_0 = const()[name = tensor("cache18_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor cache18_to_fp16 = cast(dtype = cache18_to_fp16_dtype_0, x = cache18)[name = tensor("cast_569")]; + tensor var_7838_cast_fp16 = slice_by_index(begin = var_7838_begin_0, end = var_7838_end_0, end_mask = var_7838_end_mask_0, squeeze_mask = var_7838_squeeze_mask_0, x = cache18_to_fp16)[name = tensor("op_7838_cast_fp16")]; + tensor var_7840_axis_0 = const()[name = tensor("op_7840_axis_0"), val = tensor(1)]; + tensor var_7840_mode_0 = const()[name = tensor("op_7840_mode_0"), val = tensor("update")]; + tensor var_7840_validate_indices_0 = const()[name = tensor("op_7840_validate_indices_0"), val = tensor(false)]; + tensor var_7840_cast_fp16 = scatter_along_axis(axis = var_7840_axis_0, data = var_7838_cast_fp16, indices = write_indices_37, mode = var_7840_mode_0, updates = k_75_cast_fp16, validate_indices = var_7840_validate_indices_0)[name = tensor("op_7840_cast_fp16")]; + tensor concat_146 = const()[name = tensor("concat_146"), val = tensor([0, 0, 0, 0, 0])]; + tensor concat_147 = const()[name = tensor("concat_147"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_37_internal_tensor_assign_1_stride_0 = const()[name = tensor("new_cache_37_internal_tensor_assign_1_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_37_internal_tensor_assign_1_begin_mask_0 = const()[name = tensor("new_cache_37_internal_tensor_assign_1_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_37_internal_tensor_assign_1_end_mask_0 = const()[name = tensor("new_cache_37_internal_tensor_assign_1_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_37_internal_tensor_assign_1_squeeze_mask_0 = const()[name = tensor("new_cache_37_internal_tensor_assign_1_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_82 = const()[name = tensor("shape_82"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_36 = const()[name = tensor("reduce_prod_36"), val = tensor(1048576)]; + tensor range_1d_36_start_0 = const()[name = tensor("range_1d_36_start_0"), val = tensor(0)]; + tensor range_1d_36_step_0 = const()[name = tensor("range_1d_36_step_0"), val = tensor(1)]; + tensor range_1d_36 = range_1d(end = reduce_prod_36, start = range_1d_36_start_0, step = range_1d_36_step_0)[name = tensor("range_1d_36")]; + tensor reshape_180 = reshape(shape = shape_82, x = range_1d_36)[name = tensor("reshape_180")]; + tensor slice_by_index_36 = slice_by_index(begin = concat_146, begin_mask = new_cache_37_internal_tensor_assign_1_begin_mask_0, end = concat_147, end_mask = new_cache_37_internal_tensor_assign_1_end_mask_0, squeeze_mask = new_cache_37_internal_tensor_assign_1_squeeze_mask_0, stride = new_cache_37_internal_tensor_assign_1_stride_0, x = reshape_180)[name = tensor("slice_by_index_36")]; + tensor reshape_181_shape_0 = const()[name = tensor("reshape_181_shape_0"), val = tensor([-1])]; + tensor reshape_181 = reshape(shape = reshape_181_shape_0, x = slice_by_index_36)[name = tensor("reshape_181")]; + tensor reshape_182_shape_0 = const()[name = tensor("reshape_182_shape_0"), val = tensor([-1])]; + tensor reshape_182_cast_fp16 = reshape(shape = reshape_182_shape_0, x = var_7840_cast_fp16)[name = tensor("reshape_182_cast_fp16")]; + tensor reshape_183_shape_0 = const()[name = tensor("reshape_183_shape_0"), val = tensor([-1])]; + tensor reshape_183_cast_fp16 = reshape(shape = reshape_183_shape_0, x = cache18_to_fp16)[name = tensor("reshape_183_cast_fp16")]; + tensor scatter_36_mode_0 = const()[name = tensor("scatter_36_mode_0"), val = tensor("update")]; + tensor scatter_36_axis_0 = const()[name = tensor("scatter_36_axis_0"), val = tensor(0)]; + tensor scatter_36_validate_indices_0 = const()[name = tensor("scatter_36_validate_indices_0"), val = tensor(false)]; + tensor scatter_36_cast_fp16 = scatter(axis = scatter_36_axis_0, data = reshape_183_cast_fp16, indices = reshape_181, mode = scatter_36_mode_0, updates = reshape_182_cast_fp16, validate_indices = scatter_36_validate_indices_0)[name = tensor("scatter_36_cast_fp16")]; + tensor reshape_184_cast_fp16 = reshape(shape = shape_82, x = scatter_36_cast_fp16)[name = tensor("reshape_184_cast_fp16")]; + tensor var_7848_begin_0 = const()[name = tensor("op_7848_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor var_7848_end_0 = const()[name = tensor("op_7848_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor var_7848_end_mask_0 = const()[name = tensor("op_7848_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_7848_squeeze_mask_0 = const()[name = tensor("op_7848_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor var_7848_cast_fp16 = slice_by_index(begin = var_7848_begin_0, end = var_7848_end_0, end_mask = var_7848_end_mask_0, squeeze_mask = var_7848_squeeze_mask_0, x = reshape_184_cast_fp16)[name = tensor("op_7848_cast_fp16")]; + tensor var_7850_axis_0 = const()[name = tensor("op_7850_axis_0"), val = tensor(1)]; + tensor var_7850_mode_0 = const()[name = tensor("op_7850_mode_0"), val = tensor("update")]; + tensor var_7850_validate_indices_0 = const()[name = tensor("op_7850_validate_indices_0"), val = tensor(false)]; + tensor var_7850_cast_fp16 = scatter_along_axis(axis = var_7850_axis_0, data = var_7848_cast_fp16, indices = write_indices_37, mode = var_7850_mode_0, updates = v_37_cast_fp16, validate_indices = var_7850_validate_indices_0)[name = tensor("op_7850_cast_fp16")]; + tensor concat_148 = const()[name = tensor("concat_148"), val = tensor([1, 0, 0, 0, 0])]; + tensor concat_149 = const()[name = tensor("concat_149"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_37_internal_tensor_assign_2_stride_0 = const()[name = tensor("new_cache_37_internal_tensor_assign_2_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_37_internal_tensor_assign_2_begin_mask_0 = const()[name = tensor("new_cache_37_internal_tensor_assign_2_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_37_internal_tensor_assign_2_end_mask_0 = const()[name = tensor("new_cache_37_internal_tensor_assign_2_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_37_internal_tensor_assign_2_squeeze_mask_0 = const()[name = tensor("new_cache_37_internal_tensor_assign_2_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_83 = const()[name = tensor("shape_83"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_37 = const()[name = tensor("reduce_prod_37"), val = tensor(1048576)]; + tensor range_1d_37_start_0 = const()[name = tensor("range_1d_37_start_0"), val = tensor(0)]; + tensor range_1d_37_step_0 = const()[name = tensor("range_1d_37_step_0"), val = tensor(1)]; + tensor range_1d_37 = range_1d(end = reduce_prod_37, start = range_1d_37_start_0, step = range_1d_37_step_0)[name = tensor("range_1d_37")]; + tensor reshape_185 = reshape(shape = shape_83, x = range_1d_37)[name = tensor("reshape_185")]; + tensor slice_by_index_37 = slice_by_index(begin = concat_148, begin_mask = new_cache_37_internal_tensor_assign_2_begin_mask_0, end = concat_149, end_mask = new_cache_37_internal_tensor_assign_2_end_mask_0, squeeze_mask = new_cache_37_internal_tensor_assign_2_squeeze_mask_0, stride = new_cache_37_internal_tensor_assign_2_stride_0, x = reshape_185)[name = tensor("slice_by_index_37")]; + tensor reshape_186_shape_0 = const()[name = tensor("reshape_186_shape_0"), val = tensor([-1])]; + tensor reshape_186 = reshape(shape = reshape_186_shape_0, x = slice_by_index_37)[name = tensor("reshape_186")]; + tensor reshape_187_shape_0 = const()[name = tensor("reshape_187_shape_0"), val = tensor([-1])]; + tensor reshape_187_cast_fp16 = reshape(shape = reshape_187_shape_0, x = var_7850_cast_fp16)[name = tensor("reshape_187_cast_fp16")]; + tensor reshape_188_shape_0 = const()[name = tensor("reshape_188_shape_0"), val = tensor([-1])]; + tensor reshape_188_cast_fp16 = reshape(shape = reshape_188_shape_0, x = reshape_184_cast_fp16)[name = tensor("reshape_188_cast_fp16")]; + tensor scatter_37_mode_0 = const()[name = tensor("scatter_37_mode_0"), val = tensor("update")]; + tensor scatter_37_axis_0 = const()[name = tensor("scatter_37_axis_0"), val = tensor(0)]; + tensor scatter_37_validate_indices_0 = const()[name = tensor("scatter_37_validate_indices_0"), val = tensor(false)]; + tensor scatter_37_cast_fp16 = scatter(axis = scatter_37_axis_0, data = reshape_188_cast_fp16, indices = reshape_186, mode = scatter_37_mode_0, updates = reshape_187_cast_fp16, validate_indices = scatter_37_validate_indices_0)[name = tensor("scatter_37_cast_fp16")]; + tensor reshape_189_cast_fp16 = reshape(shape = shape_83, x = scatter_37_cast_fp16)[name = tensor("reshape_189_cast_fp16")]; + tensor reshape_189_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("reshape_189_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor keys_109_begin_0 = const()[name = tensor("keys_109_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor keys_109_end_0 = const()[name = tensor("keys_109_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor keys_109_end_mask_0 = const()[name = tensor("keys_109_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor keys_109_squeeze_mask_0 = const()[name = tensor("keys_109_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor keys_109_cast_fp16 = slice_by_index(begin = keys_109_begin_0, end = keys_109_end_0, end_mask = keys_109_end_mask_0, squeeze_mask = keys_109_squeeze_mask_0, x = reshape_189_cast_fp16)[name = tensor("keys_109_cast_fp16")]; + tensor values_109_begin_0 = const()[name = tensor("values_109_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor values_109_end_0 = const()[name = tensor("values_109_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor values_109_end_mask_0 = const()[name = tensor("values_109_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor values_109_squeeze_mask_0 = const()[name = tensor("values_109_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor values_109_cast_fp16 = slice_by_index(begin = values_109_begin_0, end = values_109_end_0, end_mask = values_109_end_mask_0, squeeze_mask = values_109_squeeze_mask_0, x = reshape_189_cast_fp16)[name = tensor("values_109_cast_fp16")]; + tensor var_7862_cast_fp16 = not_equal(x = keys_109_cast_fp16, y = keys_109_cast_fp16)[name = tensor("op_7862_cast_fp16")]; + tensor keys_111_cast_fp16 = select(a = var_470_to_fp16, b = keys_109_cast_fp16, cond = var_7862_cast_fp16)[name = tensor("keys_111_cast_fp16")]; + tensor var_7870_cast_fp16 = not_equal(x = values_109_cast_fp16, y = values_109_cast_fp16)[name = tensor("op_7870_cast_fp16")]; + tensor values_111_cast_fp16 = select(a = var_470_to_fp16, b = values_109_cast_fp16, cond = var_7870_cast_fp16)[name = tensor("values_111_cast_fp16")]; + tensor var_7894 = const()[name = tensor("op_7894"), val = tensor([0, 2, 1, 3])]; + tensor var_7910 = const()[name = tensor("op_7910"), val = tensor([1, 1, 1])]; + tensor var_7911_cast_fp16 = reshape(shape = var_7910, x = position18_to_fp16)[name = tensor("op_7911_cast_fp16")]; + tensor q_positions_37_cast_fp16 = add(x = var_7911_cast_fp16, y = q_offsets_1_promoted_to_fp16)[name = tensor("q_positions_37_cast_fp16")]; + tensor valid_end_37_cast_fp16 = add(x = var_7911_cast_fp16, y = var_536_cast_fp16)[name = tensor("valid_end_37_cast_fp16")]; + tensor valid_mask_37_cast_fp16 = less(x = k_positions_1_promoted_to_fp16, y = valid_end_37_cast_fp16)[name = tensor("valid_mask_37_cast_fp16")]; + tensor causal_mask_37_cast_fp16 = less_equal(x = k_positions_1_promoted_to_fp16, y = q_positions_37_cast_fp16)[name = tensor("causal_mask_37_cast_fp16")]; + tensor var_7939 = logical_and(x = valid_mask_37_cast_fp16, y = causal_mask_37_cast_fp16)[name = tensor("op_7939")]; + tensor attn_mask_37_axes_0 = const()[name = tensor("attn_mask_37_axes_0"), val = tensor([1])]; + tensor attn_mask_37 = expand_dims(axes = attn_mask_37_axes_0, x = var_7939)[name = tensor("attn_mask_37")]; + tensor var_7953_transpose_x_0 = const()[name = tensor("op_7953_transpose_x_0"), val = tensor(false)]; + tensor var_7953_transpose_y_0 = const()[name = tensor("op_7953_transpose_y_0"), val = tensor(false)]; + tensor transpose_105_perm_0 = const()[name = tensor("transpose_105_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_106_perm_0 = const()[name = tensor("transpose_106_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_106 = transpose(perm = transpose_106_perm_0, x = keys_111_cast_fp16)[name = tensor("transpose_132")]; + tensor transpose_105 = transpose(perm = transpose_105_perm_0, x = q_111_cast_fp16)[name = tensor("transpose_133")]; + tensor var_7953_cast_fp16 = matmul(transpose_x = var_7953_transpose_x_0, transpose_y = var_7953_transpose_y_0, x = transpose_105, y = transpose_106)[name = tensor("op_7953_cast_fp16")]; + tensor var_7947_to_fp16 = const()[name = tensor("op_7947_to_fp16"), val = tensor([0x1p-3])]; + tensor attn_weights_109_cast_fp16 = mul(x = var_7953_cast_fp16, y = var_7947_to_fp16)[name = tensor("attn_weights_109_cast_fp16")]; + tensor var_7955 = logical_not(x = attn_mask_37)[name = tensor("op_7955")]; + tensor var_7956_to_fp16 = const()[name = tensor("op_7956_to_fp16"), val = tensor(-inf)]; + tensor attn_weights_111_cast_fp16 = select(a = var_7956_to_fp16, b = attn_weights_109_cast_fp16, cond = var_7955)[name = tensor("attn_weights_111_cast_fp16")]; + tensor var_7958 = const()[name = tensor("op_7958"), val = tensor(-1)]; + tensor attn_weights_113_cast_fp16 = softmax(axis = var_7958, x = attn_weights_111_cast_fp16)[name = tensor("attn_weights_113_cast_fp16")]; + tensor attn_output_37_transpose_x_0 = const()[name = tensor("attn_output_37_transpose_x_0"), val = tensor(false)]; + tensor attn_output_37_transpose_y_0 = const()[name = tensor("attn_output_37_transpose_y_0"), val = tensor(false)]; + tensor values_113_cast_fp16 = transpose(perm = var_7894, x = values_111_cast_fp16)[name = tensor("transpose_134")]; + tensor attn_output_37_cast_fp16 = matmul(transpose_x = attn_output_37_transpose_x_0, transpose_y = attn_output_37_transpose_y_0, x = attn_weights_113_cast_fp16, y = values_113_cast_fp16)[name = tensor("attn_output_37_cast_fp16")]; + tensor var_7966 = const()[name = tensor("op_7966"), val = tensor([0, 2, 1, 3])]; + tensor var_7971 = const()[name = tensor("op_7971"), val = tensor([1, 256, 1024])]; + tensor var_7967_cast_fp16 = transpose(perm = var_7966, x = attn_output_37_cast_fp16)[name = tensor("transpose_131")]; + tensor input_221_cast_fp16 = reshape(shape = var_7971, x = var_7967_cast_fp16)[name = tensor("input_221_cast_fp16")]; + tensor attn18_out_proj_weight_to_fp16 = const()[name = tensor("attn18_out_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(460507776)))]; + tensor linear_73_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = attn18_out_proj_weight_to_fp16, x = input_221_cast_fp16)[name = tensor("linear_73_cast_fp16")]; + tensor var_7980_cast_fp16 = add(x = position18_to_fp16, y = valid_len_to_fp16)[name = tensor("op_7980_cast_fp16")]; + tensor var_7980_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_7980_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor input_223_cast_fp16 = add(x = input_217_cast_fp16, y = linear_73_cast_fp16)[name = tensor("input_223_cast_fp16")]; + tensor input_225_axes_0 = const()[name = tensor("input_225_axes_0"), val = tensor([-1])]; + tensor norm18_2_weight_to_fp16 = const()[name = tensor("norm18_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(462604992)))]; + tensor norm18_2_bias_to_fp16 = const()[name = tensor("norm18_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(462607104)))]; + tensor var_7984_to_fp16 = const()[name = tensor("op_7984_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_225_cast_fp16 = layer_norm(axes = input_225_axes_0, beta = norm18_2_bias_to_fp16, epsilon = var_7984_to_fp16, gamma = norm18_2_weight_to_fp16, x = input_223_cast_fp16)[name = tensor("input_225_cast_fp16")]; + tensor linear18_1_weight_to_fp16 = const()[name = tensor("linear18_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(462609216)))]; + tensor linear_74_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = linear18_1_weight_to_fp16, x = input_225_cast_fp16)[name = tensor("linear_74_cast_fp16")]; + tensor input_227_mode_0 = const()[name = tensor("input_227_mode_0"), val = tensor("EXACT")]; + tensor input_227_cast_fp16 = gelu(mode = input_227_mode_0, x = linear_74_cast_fp16)[name = tensor("input_227_cast_fp16")]; + tensor linear18_2_weight_to_fp16 = const()[name = tensor("linear18_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(470997888)))]; + tensor linear_75_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = linear18_2_weight_to_fp16, x = input_227_cast_fp16)[name = tensor("linear_75_cast_fp16")]; + tensor input_229_cast_fp16 = add(x = input_223_cast_fp16, y = linear_75_cast_fp16)[name = tensor("input_229_cast_fp16")]; + tensor input_231_axes_0 = const()[name = tensor("input_231_axes_0"), val = tensor([-1])]; + tensor norm19_1_weight_to_fp16 = const()[name = tensor("norm19_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(479386560)))]; + tensor norm19_1_bias_to_fp16 = const()[name = tensor("norm19_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(479388672)))]; + tensor var_8001_to_fp16 = const()[name = tensor("op_8001_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_231_cast_fp16 = layer_norm(axes = input_231_axes_0, beta = norm19_1_bias_to_fp16, epsilon = var_8001_to_fp16, gamma = norm19_1_weight_to_fp16, x = input_229_cast_fp16)[name = tensor("input_231_cast_fp16")]; + tensor attn19_in_proj_weight_to_fp16 = const()[name = tensor("attn19_in_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(479390784)))]; + tensor linear_76_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = attn19_in_proj_weight_to_fp16, x = input_231_cast_fp16)[name = tensor("linear_76_cast_fp16")]; + tensor var_8015 = const()[name = tensor("op_8015"), val = tensor([1, 256, 3, 16, 64])]; + tensor qkv_39_cast_fp16 = reshape(shape = var_8015, x = linear_76_cast_fp16)[name = tensor("qkv_39_cast_fp16")]; + tensor q_115_begin_0 = const()[name = tensor("q_115_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor q_115_end_0 = const()[name = tensor("q_115_end_0"), val = tensor([1, 256, 1, 16, 64])]; + tensor q_115_end_mask_0 = const()[name = tensor("q_115_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor q_115_squeeze_mask_0 = const()[name = tensor("q_115_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor q_115_cast_fp16 = slice_by_index(begin = q_115_begin_0, end = q_115_end_0, end_mask = q_115_end_mask_0, squeeze_mask = q_115_squeeze_mask_0, x = qkv_39_cast_fp16)[name = tensor("q_115_cast_fp16")]; + tensor k_77_begin_0 = const()[name = tensor("k_77_begin_0"), val = tensor([0, 0, 1, 0, 0])]; + tensor k_77_end_0 = const()[name = tensor("k_77_end_0"), val = tensor([1, 256, 2, 16, 64])]; + tensor k_77_end_mask_0 = const()[name = tensor("k_77_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor k_77_squeeze_mask_0 = const()[name = tensor("k_77_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor k_77_cast_fp16 = slice_by_index(begin = k_77_begin_0, end = k_77_end_0, end_mask = k_77_end_mask_0, squeeze_mask = k_77_squeeze_mask_0, x = qkv_39_cast_fp16)[name = tensor("k_77_cast_fp16")]; + tensor v_39_begin_0 = const()[name = tensor("v_39_begin_0"), val = tensor([0, 0, 2, 0, 0])]; + tensor v_39_end_0 = const()[name = tensor("v_39_end_0"), val = tensor([1, 256, 3, 16, 64])]; + tensor v_39_end_mask_0 = const()[name = tensor("v_39_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor v_39_squeeze_mask_0 = const()[name = tensor("v_39_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor v_39_cast_fp16 = slice_by_index(begin = v_39_begin_0, end = v_39_end_0, end_mask = v_39_end_mask_0, squeeze_mask = v_39_squeeze_mask_0, x = qkv_39_cast_fp16)[name = tensor("v_39_cast_fp16")]; + tensor var_8073 = const()[name = tensor("op_8073"), val = tensor([1, 1])]; + tensor position19_to_fp16_dtype_0 = const()[name = tensor("position19_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor position19_to_fp16 = cast(dtype = position19_to_fp16_dtype_0, x = position19)[name = tensor("cast_566")]; + tensor var_8074_cast_fp16 = reshape(shape = var_8073, x = position19_to_fp16)[name = tensor("op_8074_cast_fp16")]; + tensor ts_117_cast_fp16 = add(x = ts_1_promoted_to_fp16, y = var_8074_cast_fp16)[name = tensor("ts_117_cast_fp16")]; + tensor var_8081 = const()[name = tensor("op_8081"), val = tensor([1, 256, 1, 1])]; + tensor ts_119_cast_fp16 = reshape(shape = var_8081, x = ts_117_cast_fp16)[name = tensor("ts_119_cast_fp16")]; + tensor var_8088 = const()[name = tensor("op_8088"), val = tensor([1, 256, 16, 32, 2])]; + tensor q_complex_39_cast_fp16 = reshape(shape = var_8088, x = q_115_cast_fp16)[name = tensor("q_complex_39_cast_fp16")]; + tensor var_8095 = const()[name = tensor("op_8095"), val = tensor([1, 256, 16, 32, 2])]; + tensor k_complex_39_cast_fp16 = reshape(shape = var_8095, x = k_77_cast_fp16)[name = tensor("k_complex_39_cast_fp16")]; + tensor var_8099_begin_0 = const()[name = tensor("op_8099_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_8099_end_0 = const()[name = tensor("op_8099_end_0"), val = tensor([1, 256, 16, 32, 1])]; + tensor var_8099_end_mask_0 = const()[name = tensor("op_8099_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_8099_squeeze_mask_0 = const()[name = tensor("op_8099_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_8099_cast_fp16 = slice_by_index(begin = var_8099_begin_0, end = var_8099_end_0, end_mask = var_8099_end_mask_0, squeeze_mask = var_8099_squeeze_mask_0, x = q_complex_39_cast_fp16)[name = tensor("op_8099_cast_fp16")]; + tensor var_8107_begin_0 = const()[name = tensor("op_8107_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_8107_end_0 = const()[name = tensor("op_8107_end_0"), val = tensor([1, 256, 16, 32, 2])]; + tensor var_8107_end_mask_0 = const()[name = tensor("op_8107_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_8107_squeeze_mask_0 = const()[name = tensor("op_8107_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_8107_cast_fp16 = slice_by_index(begin = var_8107_begin_0, end = var_8107_end_0, end_mask = var_8107_end_mask_0, squeeze_mask = var_8107_squeeze_mask_0, x = q_complex_39_cast_fp16)[name = tensor("op_8107_cast_fp16")]; + tensor var_8115_begin_0 = const()[name = tensor("op_8115_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_8115_end_0 = const()[name = tensor("op_8115_end_0"), val = tensor([1, 256, 16, 32, 1])]; + tensor var_8115_end_mask_0 = const()[name = tensor("op_8115_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_8115_squeeze_mask_0 = const()[name = tensor("op_8115_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_8115_cast_fp16 = slice_by_index(begin = var_8115_begin_0, end = var_8115_end_0, end_mask = var_8115_end_mask_0, squeeze_mask = var_8115_squeeze_mask_0, x = k_complex_39_cast_fp16)[name = tensor("op_8115_cast_fp16")]; + tensor var_8123_begin_0 = const()[name = tensor("op_8123_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_8123_end_0 = const()[name = tensor("op_8123_end_0"), val = tensor([1, 256, 16, 32, 2])]; + tensor var_8123_end_mask_0 = const()[name = tensor("op_8123_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_8123_squeeze_mask_0 = const()[name = tensor("op_8123_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_8123_cast_fp16 = slice_by_index(begin = var_8123_begin_0, end = var_8123_end_0, end_mask = var_8123_end_mask_0, squeeze_mask = var_8123_squeeze_mask_0, x = k_complex_39_cast_fp16)[name = tensor("op_8123_cast_fp16")]; + tensor freqs_39_to_fp16 = const()[name = tensor("freqs_39_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(485682304)))]; + tensor var_8129_cast_fp16 = mul(x = freqs_39_to_fp16, y = ts_119_cast_fp16)[name = tensor("op_8129_cast_fp16")]; + tensor rotr_39_cast_fp16 = cos(x = var_8129_cast_fp16)[name = tensor("rotr_39_cast_fp16")]; + tensor roti_39_cast_fp16 = sin(x = var_8129_cast_fp16)[name = tensor("roti_39_cast_fp16")]; + tensor var_8133_cast_fp16 = mul(x = var_8099_cast_fp16, y = rotr_39_cast_fp16)[name = tensor("op_8133_cast_fp16")]; + tensor var_8134_cast_fp16 = mul(x = var_8107_cast_fp16, y = roti_39_cast_fp16)[name = tensor("op_8134_cast_fp16")]; + tensor qor_77_cast_fp16 = sub(x = var_8133_cast_fp16, y = var_8134_cast_fp16)[name = tensor("qor_77_cast_fp16")]; + tensor var_8137_cast_fp16 = mul(x = var_8099_cast_fp16, y = roti_39_cast_fp16)[name = tensor("op_8137_cast_fp16")]; + tensor var_8138_cast_fp16 = mul(x = var_8107_cast_fp16, y = rotr_39_cast_fp16)[name = tensor("op_8138_cast_fp16")]; + tensor qoi_77_cast_fp16 = add(x = var_8137_cast_fp16, y = var_8138_cast_fp16)[name = tensor("qoi_77_cast_fp16")]; + tensor var_8141_cast_fp16 = mul(x = var_8115_cast_fp16, y = rotr_39_cast_fp16)[name = tensor("op_8141_cast_fp16")]; + tensor var_8142_cast_fp16 = mul(x = var_8123_cast_fp16, y = roti_39_cast_fp16)[name = tensor("op_8142_cast_fp16")]; + tensor kor_77_cast_fp16 = sub(x = var_8141_cast_fp16, y = var_8142_cast_fp16)[name = tensor("kor_77_cast_fp16")]; + tensor var_8145_cast_fp16 = mul(x = var_8115_cast_fp16, y = roti_39_cast_fp16)[name = tensor("op_8145_cast_fp16")]; + tensor var_8146_cast_fp16 = mul(x = var_8123_cast_fp16, y = rotr_39_cast_fp16)[name = tensor("op_8146_cast_fp16")]; + tensor koi_77_cast_fp16 = add(x = var_8145_cast_fp16, y = var_8146_cast_fp16)[name = tensor("koi_77_cast_fp16")]; + tensor qo_39_axis_0 = const()[name = tensor("qo_39_axis_0"), val = tensor(-1)]; + tensor qo_39_cast_fp16 = stack(axis = qo_39_axis_0, values = (qor_77_cast_fp16, qoi_77_cast_fp16))[name = tensor("qo_39_cast_fp16")]; + tensor ko_39_axis_0 = const()[name = tensor("ko_39_axis_0"), val = tensor(-1)]; + tensor ko_39_cast_fp16 = stack(axis = ko_39_axis_0, values = (kor_77_cast_fp16, koi_77_cast_fp16))[name = tensor("ko_39_cast_fp16")]; + tensor var_8179 = const()[name = tensor("op_8179"), val = tensor([1, 256, 16, 64])]; + tensor q_117_cast_fp16 = reshape(shape = var_8179, x = qo_39_cast_fp16)[name = tensor("q_117_cast_fp16")]; + tensor var_8185 = const()[name = tensor("op_8185"), val = tensor([1, 256, 16, 64])]; + tensor k_79_cast_fp16 = reshape(shape = var_8185, x = ko_39_cast_fp16)[name = tensor("k_79_cast_fp16")]; + tensor write_indices_float_115_cast_fp16 = add(x = var_8074_cast_fp16, y = var_393_promoted_to_fp16)[name = tensor("write_indices_float_115_cast_fp16")]; + tensor _inversed_8206_y_0_to_fp16 = const()[name = tensor("_inversed_8206_y_0_to_fp16"), val = tensor(0x1p-9)]; + tensor _inversed_8206_cast_fp16 = mul(x = write_indices_float_115_cast_fp16, y = _inversed_8206_y_0_to_fp16)[name = tensor("_inversed_8206_cast_fp16")]; + tensor var_8207_cast_fp16 = floor(x = _inversed_8206_cast_fp16)[name = tensor("op_8207_cast_fp16")]; + tensor var_8208_to_fp16 = const()[name = tensor("op_8208_to_fp16"), val = tensor(0x1p+9)]; + tensor var_8209_cast_fp16 = mul(x = var_8207_cast_fp16, y = var_8208_to_fp16)[name = tensor("op_8209_cast_fp16")]; + tensor write_indices_float_117_cast_fp16 = sub(x = write_indices_float_115_cast_fp16, y = var_8209_cast_fp16)[name = tensor("write_indices_float_117_cast_fp16")]; + tensor var_8227_value_0_to_fp16 = const()[name = tensor("op_8227_value_0_to_fp16"), val = tensor(0x1.ffp+8)]; + tensor var_8227_cast_fp16 = fill_like(ref_tensor = write_indices_float_117_cast_fp16, value = var_8227_value_0_to_fp16)[name = tensor("op_8227_cast_fp16")]; + tensor write_indices_float_119_cast_fp16 = select(a = write_indices_float_117_cast_fp16, b = var_8227_cast_fp16, cond = token_valid_1_cast_fp16)[name = tensor("write_indices_float_119_cast_fp16")]; + tensor var_8233_dtype_0 = const()[name = tensor("op_8233_dtype_0"), val = tensor("int32")]; + tensor var_8238 = const()[name = tensor("op_8238"), val = tensor([1, 256, 1, 1])]; + tensor write_indices_float_119_cast_fp16_to_int32 = cast(dtype = var_8233_dtype_0, x = write_indices_float_119_cast_fp16)[name = tensor("cast_565")]; + tensor var_8239 = reshape(shape = var_8238, x = write_indices_float_119_cast_fp16_to_int32)[name = tensor("op_8239")]; + tensor write_indices_39_reps_0 = const()[name = tensor("write_indices_39_reps_0"), val = tensor([1, 1, 16, 64])]; + tensor write_indices_39 = tile(reps = write_indices_39_reps_0, x = var_8239)[name = tensor("write_indices_39")]; + tensor var_8249_begin_0 = const()[name = tensor("op_8249_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_8249_end_0 = const()[name = tensor("op_8249_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor var_8249_end_mask_0 = const()[name = tensor("op_8249_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_8249_squeeze_mask_0 = const()[name = tensor("op_8249_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor cache19_to_fp16_dtype_0 = const()[name = tensor("cache19_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor cache19_to_fp16 = cast(dtype = cache19_to_fp16_dtype_0, x = cache19)[name = tensor("cast_564")]; + tensor var_8249_cast_fp16 = slice_by_index(begin = var_8249_begin_0, end = var_8249_end_0, end_mask = var_8249_end_mask_0, squeeze_mask = var_8249_squeeze_mask_0, x = cache19_to_fp16)[name = tensor("op_8249_cast_fp16")]; + tensor var_8251_axis_0 = const()[name = tensor("op_8251_axis_0"), val = tensor(1)]; + tensor var_8251_mode_0 = const()[name = tensor("op_8251_mode_0"), val = tensor("update")]; + tensor var_8251_validate_indices_0 = const()[name = tensor("op_8251_validate_indices_0"), val = tensor(false)]; + tensor var_8251_cast_fp16 = scatter_along_axis(axis = var_8251_axis_0, data = var_8249_cast_fp16, indices = write_indices_39, mode = var_8251_mode_0, updates = k_79_cast_fp16, validate_indices = var_8251_validate_indices_0)[name = tensor("op_8251_cast_fp16")]; + tensor concat_154 = const()[name = tensor("concat_154"), val = tensor([0, 0, 0, 0, 0])]; + tensor concat_155 = const()[name = tensor("concat_155"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_39_internal_tensor_assign_1_stride_0 = const()[name = tensor("new_cache_39_internal_tensor_assign_1_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_39_internal_tensor_assign_1_begin_mask_0 = const()[name = tensor("new_cache_39_internal_tensor_assign_1_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_39_internal_tensor_assign_1_end_mask_0 = const()[name = tensor("new_cache_39_internal_tensor_assign_1_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_39_internal_tensor_assign_1_squeeze_mask_0 = const()[name = tensor("new_cache_39_internal_tensor_assign_1_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_84 = const()[name = tensor("shape_84"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_38 = const()[name = tensor("reduce_prod_38"), val = tensor(1048576)]; + tensor range_1d_38_start_0 = const()[name = tensor("range_1d_38_start_0"), val = tensor(0)]; + tensor range_1d_38_step_0 = const()[name = tensor("range_1d_38_step_0"), val = tensor(1)]; + tensor range_1d_38 = range_1d(end = reduce_prod_38, start = range_1d_38_start_0, step = range_1d_38_step_0)[name = tensor("range_1d_38")]; + tensor reshape_190 = reshape(shape = shape_84, x = range_1d_38)[name = tensor("reshape_190")]; + tensor slice_by_index_38 = slice_by_index(begin = concat_154, begin_mask = new_cache_39_internal_tensor_assign_1_begin_mask_0, end = concat_155, end_mask = new_cache_39_internal_tensor_assign_1_end_mask_0, squeeze_mask = new_cache_39_internal_tensor_assign_1_squeeze_mask_0, stride = new_cache_39_internal_tensor_assign_1_stride_0, x = reshape_190)[name = tensor("slice_by_index_38")]; + tensor reshape_191_shape_0 = const()[name = tensor("reshape_191_shape_0"), val = tensor([-1])]; + tensor reshape_191 = reshape(shape = reshape_191_shape_0, x = slice_by_index_38)[name = tensor("reshape_191")]; + tensor reshape_192_shape_0 = const()[name = tensor("reshape_192_shape_0"), val = tensor([-1])]; + tensor reshape_192_cast_fp16 = reshape(shape = reshape_192_shape_0, x = var_8251_cast_fp16)[name = tensor("reshape_192_cast_fp16")]; + tensor reshape_193_shape_0 = const()[name = tensor("reshape_193_shape_0"), val = tensor([-1])]; + tensor reshape_193_cast_fp16 = reshape(shape = reshape_193_shape_0, x = cache19_to_fp16)[name = tensor("reshape_193_cast_fp16")]; + tensor scatter_38_mode_0 = const()[name = tensor("scatter_38_mode_0"), val = tensor("update")]; + tensor scatter_38_axis_0 = const()[name = tensor("scatter_38_axis_0"), val = tensor(0)]; + tensor scatter_38_validate_indices_0 = const()[name = tensor("scatter_38_validate_indices_0"), val = tensor(false)]; + tensor scatter_38_cast_fp16 = scatter(axis = scatter_38_axis_0, data = reshape_193_cast_fp16, indices = reshape_191, mode = scatter_38_mode_0, updates = reshape_192_cast_fp16, validate_indices = scatter_38_validate_indices_0)[name = tensor("scatter_38_cast_fp16")]; + tensor reshape_194_cast_fp16 = reshape(shape = shape_84, x = scatter_38_cast_fp16)[name = tensor("reshape_194_cast_fp16")]; + tensor var_8259_begin_0 = const()[name = tensor("op_8259_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor var_8259_end_0 = const()[name = tensor("op_8259_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor var_8259_end_mask_0 = const()[name = tensor("op_8259_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_8259_squeeze_mask_0 = const()[name = tensor("op_8259_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor var_8259_cast_fp16 = slice_by_index(begin = var_8259_begin_0, end = var_8259_end_0, end_mask = var_8259_end_mask_0, squeeze_mask = var_8259_squeeze_mask_0, x = reshape_194_cast_fp16)[name = tensor("op_8259_cast_fp16")]; + tensor var_8261_axis_0 = const()[name = tensor("op_8261_axis_0"), val = tensor(1)]; + tensor var_8261_mode_0 = const()[name = tensor("op_8261_mode_0"), val = tensor("update")]; + tensor var_8261_validate_indices_0 = const()[name = tensor("op_8261_validate_indices_0"), val = tensor(false)]; + tensor var_8261_cast_fp16 = scatter_along_axis(axis = var_8261_axis_0, data = var_8259_cast_fp16, indices = write_indices_39, mode = var_8261_mode_0, updates = v_39_cast_fp16, validate_indices = var_8261_validate_indices_0)[name = tensor("op_8261_cast_fp16")]; + tensor concat_156 = const()[name = tensor("concat_156"), val = tensor([1, 0, 0, 0, 0])]; + tensor concat_157 = const()[name = tensor("concat_157"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_39_internal_tensor_assign_2_stride_0 = const()[name = tensor("new_cache_39_internal_tensor_assign_2_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_39_internal_tensor_assign_2_begin_mask_0 = const()[name = tensor("new_cache_39_internal_tensor_assign_2_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_39_internal_tensor_assign_2_end_mask_0 = const()[name = tensor("new_cache_39_internal_tensor_assign_2_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_39_internal_tensor_assign_2_squeeze_mask_0 = const()[name = tensor("new_cache_39_internal_tensor_assign_2_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_85 = const()[name = tensor("shape_85"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_39 = const()[name = tensor("reduce_prod_39"), val = tensor(1048576)]; + tensor range_1d_39_start_0 = const()[name = tensor("range_1d_39_start_0"), val = tensor(0)]; + tensor range_1d_39_step_0 = const()[name = tensor("range_1d_39_step_0"), val = tensor(1)]; + tensor range_1d_39 = range_1d(end = reduce_prod_39, start = range_1d_39_start_0, step = range_1d_39_step_0)[name = tensor("range_1d_39")]; + tensor reshape_195 = reshape(shape = shape_85, x = range_1d_39)[name = tensor("reshape_195")]; + tensor slice_by_index_39 = slice_by_index(begin = concat_156, begin_mask = new_cache_39_internal_tensor_assign_2_begin_mask_0, end = concat_157, end_mask = new_cache_39_internal_tensor_assign_2_end_mask_0, squeeze_mask = new_cache_39_internal_tensor_assign_2_squeeze_mask_0, stride = new_cache_39_internal_tensor_assign_2_stride_0, x = reshape_195)[name = tensor("slice_by_index_39")]; + tensor reshape_196_shape_0 = const()[name = tensor("reshape_196_shape_0"), val = tensor([-1])]; + tensor reshape_196 = reshape(shape = reshape_196_shape_0, x = slice_by_index_39)[name = tensor("reshape_196")]; + tensor reshape_197_shape_0 = const()[name = tensor("reshape_197_shape_0"), val = tensor([-1])]; + tensor reshape_197_cast_fp16 = reshape(shape = reshape_197_shape_0, x = var_8261_cast_fp16)[name = tensor("reshape_197_cast_fp16")]; + tensor reshape_198_shape_0 = const()[name = tensor("reshape_198_shape_0"), val = tensor([-1])]; + tensor reshape_198_cast_fp16 = reshape(shape = reshape_198_shape_0, x = reshape_194_cast_fp16)[name = tensor("reshape_198_cast_fp16")]; + tensor scatter_39_mode_0 = const()[name = tensor("scatter_39_mode_0"), val = tensor("update")]; + tensor scatter_39_axis_0 = const()[name = tensor("scatter_39_axis_0"), val = tensor(0)]; + tensor scatter_39_validate_indices_0 = const()[name = tensor("scatter_39_validate_indices_0"), val = tensor(false)]; + tensor scatter_39_cast_fp16 = scatter(axis = scatter_39_axis_0, data = reshape_198_cast_fp16, indices = reshape_196, mode = scatter_39_mode_0, updates = reshape_197_cast_fp16, validate_indices = scatter_39_validate_indices_0)[name = tensor("scatter_39_cast_fp16")]; + tensor reshape_199_cast_fp16 = reshape(shape = shape_85, x = scatter_39_cast_fp16)[name = tensor("reshape_199_cast_fp16")]; + tensor reshape_199_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("reshape_199_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor keys_115_begin_0 = const()[name = tensor("keys_115_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor keys_115_end_0 = const()[name = tensor("keys_115_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor keys_115_end_mask_0 = const()[name = tensor("keys_115_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor keys_115_squeeze_mask_0 = const()[name = tensor("keys_115_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor keys_115_cast_fp16 = slice_by_index(begin = keys_115_begin_0, end = keys_115_end_0, end_mask = keys_115_end_mask_0, squeeze_mask = keys_115_squeeze_mask_0, x = reshape_199_cast_fp16)[name = tensor("keys_115_cast_fp16")]; + tensor values_115_begin_0 = const()[name = tensor("values_115_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor values_115_end_0 = const()[name = tensor("values_115_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor values_115_end_mask_0 = const()[name = tensor("values_115_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor values_115_squeeze_mask_0 = const()[name = tensor("values_115_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor values_115_cast_fp16 = slice_by_index(begin = values_115_begin_0, end = values_115_end_0, end_mask = values_115_end_mask_0, squeeze_mask = values_115_squeeze_mask_0, x = reshape_199_cast_fp16)[name = tensor("values_115_cast_fp16")]; + tensor var_8273_cast_fp16 = not_equal(x = keys_115_cast_fp16, y = keys_115_cast_fp16)[name = tensor("op_8273_cast_fp16")]; + tensor keys_117_cast_fp16 = select(a = var_470_to_fp16, b = keys_115_cast_fp16, cond = var_8273_cast_fp16)[name = tensor("keys_117_cast_fp16")]; + tensor var_8281_cast_fp16 = not_equal(x = values_115_cast_fp16, y = values_115_cast_fp16)[name = tensor("op_8281_cast_fp16")]; + tensor values_117_cast_fp16 = select(a = var_470_to_fp16, b = values_115_cast_fp16, cond = var_8281_cast_fp16)[name = tensor("values_117_cast_fp16")]; + tensor var_8305 = const()[name = tensor("op_8305"), val = tensor([0, 2, 1, 3])]; + tensor var_8321 = const()[name = tensor("op_8321"), val = tensor([1, 1, 1])]; + tensor var_8322_cast_fp16 = reshape(shape = var_8321, x = position19_to_fp16)[name = tensor("op_8322_cast_fp16")]; + tensor q_positions_39_cast_fp16 = add(x = var_8322_cast_fp16, y = q_offsets_1_promoted_to_fp16)[name = tensor("q_positions_39_cast_fp16")]; + tensor valid_end_39_cast_fp16 = add(x = var_8322_cast_fp16, y = var_536_cast_fp16)[name = tensor("valid_end_39_cast_fp16")]; + tensor valid_mask_39_cast_fp16 = less(x = k_positions_1_promoted_to_fp16, y = valid_end_39_cast_fp16)[name = tensor("valid_mask_39_cast_fp16")]; + tensor causal_mask_39_cast_fp16 = less_equal(x = k_positions_1_promoted_to_fp16, y = q_positions_39_cast_fp16)[name = tensor("causal_mask_39_cast_fp16")]; + tensor var_8350 = logical_and(x = valid_mask_39_cast_fp16, y = causal_mask_39_cast_fp16)[name = tensor("op_8350")]; + tensor attn_mask_39_axes_0 = const()[name = tensor("attn_mask_39_axes_0"), val = tensor([1])]; + tensor attn_mask_39 = expand_dims(axes = attn_mask_39_axes_0, x = var_8350)[name = tensor("attn_mask_39")]; + tensor var_8364_transpose_x_0 = const()[name = tensor("op_8364_transpose_x_0"), val = tensor(false)]; + tensor var_8364_transpose_y_0 = const()[name = tensor("op_8364_transpose_y_0"), val = tensor(false)]; + tensor transpose_107_perm_0 = const()[name = tensor("transpose_107_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_108_perm_0 = const()[name = tensor("transpose_108_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_108 = transpose(perm = transpose_108_perm_0, x = keys_117_cast_fp16)[name = tensor("transpose_128")]; + tensor transpose_107 = transpose(perm = transpose_107_perm_0, x = q_117_cast_fp16)[name = tensor("transpose_129")]; + tensor var_8364_cast_fp16 = matmul(transpose_x = var_8364_transpose_x_0, transpose_y = var_8364_transpose_y_0, x = transpose_107, y = transpose_108)[name = tensor("op_8364_cast_fp16")]; + tensor var_8358_to_fp16 = const()[name = tensor("op_8358_to_fp16"), val = tensor([0x1p-3])]; + tensor attn_weights_115_cast_fp16 = mul(x = var_8364_cast_fp16, y = var_8358_to_fp16)[name = tensor("attn_weights_115_cast_fp16")]; + tensor var_8366 = logical_not(x = attn_mask_39)[name = tensor("op_8366")]; + tensor var_8367_to_fp16 = const()[name = tensor("op_8367_to_fp16"), val = tensor(-inf)]; + tensor attn_weights_117_cast_fp16 = select(a = var_8367_to_fp16, b = attn_weights_115_cast_fp16, cond = var_8366)[name = tensor("attn_weights_117_cast_fp16")]; + tensor var_8369 = const()[name = tensor("op_8369"), val = tensor(-1)]; + tensor attn_weights_119_cast_fp16 = softmax(axis = var_8369, x = attn_weights_117_cast_fp16)[name = tensor("attn_weights_119_cast_fp16")]; + tensor attn_output_39_transpose_x_0 = const()[name = tensor("attn_output_39_transpose_x_0"), val = tensor(false)]; + tensor attn_output_39_transpose_y_0 = const()[name = tensor("attn_output_39_transpose_y_0"), val = tensor(false)]; + tensor values_119_cast_fp16 = transpose(perm = var_8305, x = values_117_cast_fp16)[name = tensor("transpose_130")]; + tensor attn_output_39_cast_fp16 = matmul(transpose_x = attn_output_39_transpose_x_0, transpose_y = attn_output_39_transpose_y_0, x = attn_weights_119_cast_fp16, y = values_119_cast_fp16)[name = tensor("attn_output_39_cast_fp16")]; + tensor var_8377 = const()[name = tensor("op_8377"), val = tensor([0, 2, 1, 3])]; + tensor var_8382 = const()[name = tensor("op_8382"), val = tensor([1, 256, 1024])]; + tensor var_8378_cast_fp16 = transpose(perm = var_8377, x = attn_output_39_cast_fp16)[name = tensor("transpose_127")]; + tensor input_233_cast_fp16 = reshape(shape = var_8382, x = var_8378_cast_fp16)[name = tensor("input_233_cast_fp16")]; + tensor attn19_out_proj_weight_to_fp16 = const()[name = tensor("attn19_out_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(485682432)))]; + tensor linear_77_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = attn19_out_proj_weight_to_fp16, x = input_233_cast_fp16)[name = tensor("linear_77_cast_fp16")]; + tensor var_8391_cast_fp16 = add(x = position19_to_fp16, y = valid_len_to_fp16)[name = tensor("op_8391_cast_fp16")]; + tensor var_8391_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_8391_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor input_235_cast_fp16 = add(x = input_229_cast_fp16, y = linear_77_cast_fp16)[name = tensor("input_235_cast_fp16")]; + tensor input_237_axes_0 = const()[name = tensor("input_237_axes_0"), val = tensor([-1])]; + tensor norm19_2_weight_to_fp16 = const()[name = tensor("norm19_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(487779648)))]; + tensor norm19_2_bias_to_fp16 = const()[name = tensor("norm19_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(487781760)))]; + tensor var_8395_to_fp16 = const()[name = tensor("op_8395_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_237_cast_fp16 = layer_norm(axes = input_237_axes_0, beta = norm19_2_bias_to_fp16, epsilon = var_8395_to_fp16, gamma = norm19_2_weight_to_fp16, x = input_235_cast_fp16)[name = tensor("input_237_cast_fp16")]; + tensor linear19_1_weight_to_fp16 = const()[name = tensor("linear19_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(487783872)))]; + tensor linear_78_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = linear19_1_weight_to_fp16, x = input_237_cast_fp16)[name = tensor("linear_78_cast_fp16")]; + tensor input_239_mode_0 = const()[name = tensor("input_239_mode_0"), val = tensor("EXACT")]; + tensor input_239_cast_fp16 = gelu(mode = input_239_mode_0, x = linear_78_cast_fp16)[name = tensor("input_239_cast_fp16")]; + tensor linear19_2_weight_to_fp16 = const()[name = tensor("linear19_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(496172544)))]; + tensor linear_79_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = linear19_2_weight_to_fp16, x = input_239_cast_fp16)[name = tensor("linear_79_cast_fp16")]; + tensor input_241_cast_fp16 = add(x = input_235_cast_fp16, y = linear_79_cast_fp16)[name = tensor("input_241_cast_fp16")]; + tensor input_243_axes_0 = const()[name = tensor("input_243_axes_0"), val = tensor([-1])]; + tensor norm20_1_weight_to_fp16 = const()[name = tensor("norm20_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(504561216)))]; + tensor norm20_1_bias_to_fp16 = const()[name = tensor("norm20_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(504563328)))]; + tensor var_8412_to_fp16 = const()[name = tensor("op_8412_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_243_cast_fp16 = layer_norm(axes = input_243_axes_0, beta = norm20_1_bias_to_fp16, epsilon = var_8412_to_fp16, gamma = norm20_1_weight_to_fp16, x = input_241_cast_fp16)[name = tensor("input_243_cast_fp16")]; + tensor attn20_in_proj_weight_to_fp16 = const()[name = tensor("attn20_in_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(504565440)))]; + tensor linear_80_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = attn20_in_proj_weight_to_fp16, x = input_243_cast_fp16)[name = tensor("linear_80_cast_fp16")]; + tensor var_8426 = const()[name = tensor("op_8426"), val = tensor([1, 256, 3, 16, 64])]; + tensor qkv_41_cast_fp16 = reshape(shape = var_8426, x = linear_80_cast_fp16)[name = tensor("qkv_41_cast_fp16")]; + tensor q_121_begin_0 = const()[name = tensor("q_121_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor q_121_end_0 = const()[name = tensor("q_121_end_0"), val = tensor([1, 256, 1, 16, 64])]; + tensor q_121_end_mask_0 = const()[name = tensor("q_121_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor q_121_squeeze_mask_0 = const()[name = tensor("q_121_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor q_121_cast_fp16 = slice_by_index(begin = q_121_begin_0, end = q_121_end_0, end_mask = q_121_end_mask_0, squeeze_mask = q_121_squeeze_mask_0, x = qkv_41_cast_fp16)[name = tensor("q_121_cast_fp16")]; + tensor k_81_begin_0 = const()[name = tensor("k_81_begin_0"), val = tensor([0, 0, 1, 0, 0])]; + tensor k_81_end_0 = const()[name = tensor("k_81_end_0"), val = tensor([1, 256, 2, 16, 64])]; + tensor k_81_end_mask_0 = const()[name = tensor("k_81_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor k_81_squeeze_mask_0 = const()[name = tensor("k_81_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor k_81_cast_fp16 = slice_by_index(begin = k_81_begin_0, end = k_81_end_0, end_mask = k_81_end_mask_0, squeeze_mask = k_81_squeeze_mask_0, x = qkv_41_cast_fp16)[name = tensor("k_81_cast_fp16")]; + tensor v_41_begin_0 = const()[name = tensor("v_41_begin_0"), val = tensor([0, 0, 2, 0, 0])]; + tensor v_41_end_0 = const()[name = tensor("v_41_end_0"), val = tensor([1, 256, 3, 16, 64])]; + tensor v_41_end_mask_0 = const()[name = tensor("v_41_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor v_41_squeeze_mask_0 = const()[name = tensor("v_41_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor v_41_cast_fp16 = slice_by_index(begin = v_41_begin_0, end = v_41_end_0, end_mask = v_41_end_mask_0, squeeze_mask = v_41_squeeze_mask_0, x = qkv_41_cast_fp16)[name = tensor("v_41_cast_fp16")]; + tensor var_8484 = const()[name = tensor("op_8484"), val = tensor([1, 1])]; + tensor position20_to_fp16_dtype_0 = const()[name = tensor("position20_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor position20_to_fp16 = cast(dtype = position20_to_fp16_dtype_0, x = position20)[name = tensor("cast_561")]; + tensor var_8485_cast_fp16 = reshape(shape = var_8484, x = position20_to_fp16)[name = tensor("op_8485_cast_fp16")]; + tensor ts_123_cast_fp16 = add(x = ts_1_promoted_to_fp16, y = var_8485_cast_fp16)[name = tensor("ts_123_cast_fp16")]; + tensor var_8492 = const()[name = tensor("op_8492"), val = tensor([1, 256, 1, 1])]; + tensor ts_125_cast_fp16 = reshape(shape = var_8492, x = ts_123_cast_fp16)[name = tensor("ts_125_cast_fp16")]; + tensor var_8499 = const()[name = tensor("op_8499"), val = tensor([1, 256, 16, 32, 2])]; + tensor q_complex_41_cast_fp16 = reshape(shape = var_8499, x = q_121_cast_fp16)[name = tensor("q_complex_41_cast_fp16")]; + tensor var_8506 = const()[name = tensor("op_8506"), val = tensor([1, 256, 16, 32, 2])]; + tensor k_complex_41_cast_fp16 = reshape(shape = var_8506, x = k_81_cast_fp16)[name = tensor("k_complex_41_cast_fp16")]; + tensor var_8510_begin_0 = const()[name = tensor("op_8510_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_8510_end_0 = const()[name = tensor("op_8510_end_0"), val = tensor([1, 256, 16, 32, 1])]; + tensor var_8510_end_mask_0 = const()[name = tensor("op_8510_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_8510_squeeze_mask_0 = const()[name = tensor("op_8510_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_8510_cast_fp16 = slice_by_index(begin = var_8510_begin_0, end = var_8510_end_0, end_mask = var_8510_end_mask_0, squeeze_mask = var_8510_squeeze_mask_0, x = q_complex_41_cast_fp16)[name = tensor("op_8510_cast_fp16")]; + tensor var_8518_begin_0 = const()[name = tensor("op_8518_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_8518_end_0 = const()[name = tensor("op_8518_end_0"), val = tensor([1, 256, 16, 32, 2])]; + tensor var_8518_end_mask_0 = const()[name = tensor("op_8518_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_8518_squeeze_mask_0 = const()[name = tensor("op_8518_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_8518_cast_fp16 = slice_by_index(begin = var_8518_begin_0, end = var_8518_end_0, end_mask = var_8518_end_mask_0, squeeze_mask = var_8518_squeeze_mask_0, x = q_complex_41_cast_fp16)[name = tensor("op_8518_cast_fp16")]; + tensor var_8526_begin_0 = const()[name = tensor("op_8526_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_8526_end_0 = const()[name = tensor("op_8526_end_0"), val = tensor([1, 256, 16, 32, 1])]; + tensor var_8526_end_mask_0 = const()[name = tensor("op_8526_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_8526_squeeze_mask_0 = const()[name = tensor("op_8526_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_8526_cast_fp16 = slice_by_index(begin = var_8526_begin_0, end = var_8526_end_0, end_mask = var_8526_end_mask_0, squeeze_mask = var_8526_squeeze_mask_0, x = k_complex_41_cast_fp16)[name = tensor("op_8526_cast_fp16")]; + tensor var_8534_begin_0 = const()[name = tensor("op_8534_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_8534_end_0 = const()[name = tensor("op_8534_end_0"), val = tensor([1, 256, 16, 32, 2])]; + tensor var_8534_end_mask_0 = const()[name = tensor("op_8534_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_8534_squeeze_mask_0 = const()[name = tensor("op_8534_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_8534_cast_fp16 = slice_by_index(begin = var_8534_begin_0, end = var_8534_end_0, end_mask = var_8534_end_mask_0, squeeze_mask = var_8534_squeeze_mask_0, x = k_complex_41_cast_fp16)[name = tensor("op_8534_cast_fp16")]; + tensor freqs_41_to_fp16 = const()[name = tensor("freqs_41_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(510856960)))]; + tensor var_8540_cast_fp16 = mul(x = freqs_41_to_fp16, y = ts_125_cast_fp16)[name = tensor("op_8540_cast_fp16")]; + tensor rotr_41_cast_fp16 = cos(x = var_8540_cast_fp16)[name = tensor("rotr_41_cast_fp16")]; + tensor roti_41_cast_fp16 = sin(x = var_8540_cast_fp16)[name = tensor("roti_41_cast_fp16")]; + tensor var_8544_cast_fp16 = mul(x = var_8510_cast_fp16, y = rotr_41_cast_fp16)[name = tensor("op_8544_cast_fp16")]; + tensor var_8545_cast_fp16 = mul(x = var_8518_cast_fp16, y = roti_41_cast_fp16)[name = tensor("op_8545_cast_fp16")]; + tensor qor_81_cast_fp16 = sub(x = var_8544_cast_fp16, y = var_8545_cast_fp16)[name = tensor("qor_81_cast_fp16")]; + tensor var_8548_cast_fp16 = mul(x = var_8510_cast_fp16, y = roti_41_cast_fp16)[name = tensor("op_8548_cast_fp16")]; + tensor var_8549_cast_fp16 = mul(x = var_8518_cast_fp16, y = rotr_41_cast_fp16)[name = tensor("op_8549_cast_fp16")]; + tensor qoi_81_cast_fp16 = add(x = var_8548_cast_fp16, y = var_8549_cast_fp16)[name = tensor("qoi_81_cast_fp16")]; + tensor var_8552_cast_fp16 = mul(x = var_8526_cast_fp16, y = rotr_41_cast_fp16)[name = tensor("op_8552_cast_fp16")]; + tensor var_8553_cast_fp16 = mul(x = var_8534_cast_fp16, y = roti_41_cast_fp16)[name = tensor("op_8553_cast_fp16")]; + tensor kor_81_cast_fp16 = sub(x = var_8552_cast_fp16, y = var_8553_cast_fp16)[name = tensor("kor_81_cast_fp16")]; + tensor var_8556_cast_fp16 = mul(x = var_8526_cast_fp16, y = roti_41_cast_fp16)[name = tensor("op_8556_cast_fp16")]; + tensor var_8557_cast_fp16 = mul(x = var_8534_cast_fp16, y = rotr_41_cast_fp16)[name = tensor("op_8557_cast_fp16")]; + tensor koi_81_cast_fp16 = add(x = var_8556_cast_fp16, y = var_8557_cast_fp16)[name = tensor("koi_81_cast_fp16")]; + tensor qo_41_axis_0 = const()[name = tensor("qo_41_axis_0"), val = tensor(-1)]; + tensor qo_41_cast_fp16 = stack(axis = qo_41_axis_0, values = (qor_81_cast_fp16, qoi_81_cast_fp16))[name = tensor("qo_41_cast_fp16")]; + tensor ko_41_axis_0 = const()[name = tensor("ko_41_axis_0"), val = tensor(-1)]; + tensor ko_41_cast_fp16 = stack(axis = ko_41_axis_0, values = (kor_81_cast_fp16, koi_81_cast_fp16))[name = tensor("ko_41_cast_fp16")]; + tensor var_8590 = const()[name = tensor("op_8590"), val = tensor([1, 256, 16, 64])]; + tensor q_123_cast_fp16 = reshape(shape = var_8590, x = qo_41_cast_fp16)[name = tensor("q_123_cast_fp16")]; + tensor var_8596 = const()[name = tensor("op_8596"), val = tensor([1, 256, 16, 64])]; + tensor k_83_cast_fp16 = reshape(shape = var_8596, x = ko_41_cast_fp16)[name = tensor("k_83_cast_fp16")]; + tensor write_indices_float_121_cast_fp16 = add(x = var_8485_cast_fp16, y = var_393_promoted_to_fp16)[name = tensor("write_indices_float_121_cast_fp16")]; + tensor _inversed_8617_y_0_to_fp16 = const()[name = tensor("_inversed_8617_y_0_to_fp16"), val = tensor(0x1p-9)]; + tensor _inversed_8617_cast_fp16 = mul(x = write_indices_float_121_cast_fp16, y = _inversed_8617_y_0_to_fp16)[name = tensor("_inversed_8617_cast_fp16")]; + tensor var_8618_cast_fp16 = floor(x = _inversed_8617_cast_fp16)[name = tensor("op_8618_cast_fp16")]; + tensor var_8619_to_fp16 = const()[name = tensor("op_8619_to_fp16"), val = tensor(0x1p+9)]; + tensor var_8620_cast_fp16 = mul(x = var_8618_cast_fp16, y = var_8619_to_fp16)[name = tensor("op_8620_cast_fp16")]; + tensor write_indices_float_123_cast_fp16 = sub(x = write_indices_float_121_cast_fp16, y = var_8620_cast_fp16)[name = tensor("write_indices_float_123_cast_fp16")]; + tensor var_8638_value_0_to_fp16 = const()[name = tensor("op_8638_value_0_to_fp16"), val = tensor(0x1.ffp+8)]; + tensor var_8638_cast_fp16 = fill_like(ref_tensor = write_indices_float_123_cast_fp16, value = var_8638_value_0_to_fp16)[name = tensor("op_8638_cast_fp16")]; + tensor write_indices_float_125_cast_fp16 = select(a = write_indices_float_123_cast_fp16, b = var_8638_cast_fp16, cond = token_valid_1_cast_fp16)[name = tensor("write_indices_float_125_cast_fp16")]; + tensor var_8644_dtype_0 = const()[name = tensor("op_8644_dtype_0"), val = tensor("int32")]; + tensor var_8649 = const()[name = tensor("op_8649"), val = tensor([1, 256, 1, 1])]; + tensor write_indices_float_125_cast_fp16_to_int32 = cast(dtype = var_8644_dtype_0, x = write_indices_float_125_cast_fp16)[name = tensor("cast_560")]; + tensor var_8650 = reshape(shape = var_8649, x = write_indices_float_125_cast_fp16_to_int32)[name = tensor("op_8650")]; + tensor write_indices_41_reps_0 = const()[name = tensor("write_indices_41_reps_0"), val = tensor([1, 1, 16, 64])]; + tensor write_indices_41 = tile(reps = write_indices_41_reps_0, x = var_8650)[name = tensor("write_indices_41")]; + tensor var_8660_begin_0 = const()[name = tensor("op_8660_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_8660_end_0 = const()[name = tensor("op_8660_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor var_8660_end_mask_0 = const()[name = tensor("op_8660_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_8660_squeeze_mask_0 = const()[name = tensor("op_8660_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor cache20_to_fp16_dtype_0 = const()[name = tensor("cache20_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor cache20_to_fp16 = cast(dtype = cache20_to_fp16_dtype_0, x = cache20)[name = tensor("cast_559")]; + tensor var_8660_cast_fp16 = slice_by_index(begin = var_8660_begin_0, end = var_8660_end_0, end_mask = var_8660_end_mask_0, squeeze_mask = var_8660_squeeze_mask_0, x = cache20_to_fp16)[name = tensor("op_8660_cast_fp16")]; + tensor var_8662_axis_0 = const()[name = tensor("op_8662_axis_0"), val = tensor(1)]; + tensor var_8662_mode_0 = const()[name = tensor("op_8662_mode_0"), val = tensor("update")]; + tensor var_8662_validate_indices_0 = const()[name = tensor("op_8662_validate_indices_0"), val = tensor(false)]; + tensor var_8662_cast_fp16 = scatter_along_axis(axis = var_8662_axis_0, data = var_8660_cast_fp16, indices = write_indices_41, mode = var_8662_mode_0, updates = k_83_cast_fp16, validate_indices = var_8662_validate_indices_0)[name = tensor("op_8662_cast_fp16")]; + tensor concat_162 = const()[name = tensor("concat_162"), val = tensor([0, 0, 0, 0, 0])]; + tensor concat_163 = const()[name = tensor("concat_163"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_41_internal_tensor_assign_1_stride_0 = const()[name = tensor("new_cache_41_internal_tensor_assign_1_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_41_internal_tensor_assign_1_begin_mask_0 = const()[name = tensor("new_cache_41_internal_tensor_assign_1_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_41_internal_tensor_assign_1_end_mask_0 = const()[name = tensor("new_cache_41_internal_tensor_assign_1_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_41_internal_tensor_assign_1_squeeze_mask_0 = const()[name = tensor("new_cache_41_internal_tensor_assign_1_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_86 = const()[name = tensor("shape_86"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_40 = const()[name = tensor("reduce_prod_40"), val = tensor(1048576)]; + tensor range_1d_40_start_0 = const()[name = tensor("range_1d_40_start_0"), val = tensor(0)]; + tensor range_1d_40_step_0 = const()[name = tensor("range_1d_40_step_0"), val = tensor(1)]; + tensor range_1d_40 = range_1d(end = reduce_prod_40, start = range_1d_40_start_0, step = range_1d_40_step_0)[name = tensor("range_1d_40")]; + tensor reshape_200 = reshape(shape = shape_86, x = range_1d_40)[name = tensor("reshape_200")]; + tensor slice_by_index_40 = slice_by_index(begin = concat_162, begin_mask = new_cache_41_internal_tensor_assign_1_begin_mask_0, end = concat_163, end_mask = new_cache_41_internal_tensor_assign_1_end_mask_0, squeeze_mask = new_cache_41_internal_tensor_assign_1_squeeze_mask_0, stride = new_cache_41_internal_tensor_assign_1_stride_0, x = reshape_200)[name = tensor("slice_by_index_40")]; + tensor reshape_201_shape_0 = const()[name = tensor("reshape_201_shape_0"), val = tensor([-1])]; + tensor reshape_201 = reshape(shape = reshape_201_shape_0, x = slice_by_index_40)[name = tensor("reshape_201")]; + tensor reshape_202_shape_0 = const()[name = tensor("reshape_202_shape_0"), val = tensor([-1])]; + tensor reshape_202_cast_fp16 = reshape(shape = reshape_202_shape_0, x = var_8662_cast_fp16)[name = tensor("reshape_202_cast_fp16")]; + tensor reshape_203_shape_0 = const()[name = tensor("reshape_203_shape_0"), val = tensor([-1])]; + tensor reshape_203_cast_fp16 = reshape(shape = reshape_203_shape_0, x = cache20_to_fp16)[name = tensor("reshape_203_cast_fp16")]; + tensor scatter_40_mode_0 = const()[name = tensor("scatter_40_mode_0"), val = tensor("update")]; + tensor scatter_40_axis_0 = const()[name = tensor("scatter_40_axis_0"), val = tensor(0)]; + tensor scatter_40_validate_indices_0 = const()[name = tensor("scatter_40_validate_indices_0"), val = tensor(false)]; + tensor scatter_40_cast_fp16 = scatter(axis = scatter_40_axis_0, data = reshape_203_cast_fp16, indices = reshape_201, mode = scatter_40_mode_0, updates = reshape_202_cast_fp16, validate_indices = scatter_40_validate_indices_0)[name = tensor("scatter_40_cast_fp16")]; + tensor reshape_204_cast_fp16 = reshape(shape = shape_86, x = scatter_40_cast_fp16)[name = tensor("reshape_204_cast_fp16")]; + tensor var_8670_begin_0 = const()[name = tensor("op_8670_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor var_8670_end_0 = const()[name = tensor("op_8670_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor var_8670_end_mask_0 = const()[name = tensor("op_8670_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_8670_squeeze_mask_0 = const()[name = tensor("op_8670_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor var_8670_cast_fp16 = slice_by_index(begin = var_8670_begin_0, end = var_8670_end_0, end_mask = var_8670_end_mask_0, squeeze_mask = var_8670_squeeze_mask_0, x = reshape_204_cast_fp16)[name = tensor("op_8670_cast_fp16")]; + tensor var_8672_axis_0 = const()[name = tensor("op_8672_axis_0"), val = tensor(1)]; + tensor var_8672_mode_0 = const()[name = tensor("op_8672_mode_0"), val = tensor("update")]; + tensor var_8672_validate_indices_0 = const()[name = tensor("op_8672_validate_indices_0"), val = tensor(false)]; + tensor var_8672_cast_fp16 = scatter_along_axis(axis = var_8672_axis_0, data = var_8670_cast_fp16, indices = write_indices_41, mode = var_8672_mode_0, updates = v_41_cast_fp16, validate_indices = var_8672_validate_indices_0)[name = tensor("op_8672_cast_fp16")]; + tensor concat_164 = const()[name = tensor("concat_164"), val = tensor([1, 0, 0, 0, 0])]; + tensor concat_165 = const()[name = tensor("concat_165"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_41_internal_tensor_assign_2_stride_0 = const()[name = tensor("new_cache_41_internal_tensor_assign_2_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_41_internal_tensor_assign_2_begin_mask_0 = const()[name = tensor("new_cache_41_internal_tensor_assign_2_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_41_internal_tensor_assign_2_end_mask_0 = const()[name = tensor("new_cache_41_internal_tensor_assign_2_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_41_internal_tensor_assign_2_squeeze_mask_0 = const()[name = tensor("new_cache_41_internal_tensor_assign_2_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_87 = const()[name = tensor("shape_87"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_41 = const()[name = tensor("reduce_prod_41"), val = tensor(1048576)]; + tensor range_1d_41_start_0 = const()[name = tensor("range_1d_41_start_0"), val = tensor(0)]; + tensor range_1d_41_step_0 = const()[name = tensor("range_1d_41_step_0"), val = tensor(1)]; + tensor range_1d_41 = range_1d(end = reduce_prod_41, start = range_1d_41_start_0, step = range_1d_41_step_0)[name = tensor("range_1d_41")]; + tensor reshape_205 = reshape(shape = shape_87, x = range_1d_41)[name = tensor("reshape_205")]; + tensor slice_by_index_41 = slice_by_index(begin = concat_164, begin_mask = new_cache_41_internal_tensor_assign_2_begin_mask_0, end = concat_165, end_mask = new_cache_41_internal_tensor_assign_2_end_mask_0, squeeze_mask = new_cache_41_internal_tensor_assign_2_squeeze_mask_0, stride = new_cache_41_internal_tensor_assign_2_stride_0, x = reshape_205)[name = tensor("slice_by_index_41")]; + tensor reshape_206_shape_0 = const()[name = tensor("reshape_206_shape_0"), val = tensor([-1])]; + tensor reshape_206 = reshape(shape = reshape_206_shape_0, x = slice_by_index_41)[name = tensor("reshape_206")]; + tensor reshape_207_shape_0 = const()[name = tensor("reshape_207_shape_0"), val = tensor([-1])]; + tensor reshape_207_cast_fp16 = reshape(shape = reshape_207_shape_0, x = var_8672_cast_fp16)[name = tensor("reshape_207_cast_fp16")]; + tensor reshape_208_shape_0 = const()[name = tensor("reshape_208_shape_0"), val = tensor([-1])]; + tensor reshape_208_cast_fp16 = reshape(shape = reshape_208_shape_0, x = reshape_204_cast_fp16)[name = tensor("reshape_208_cast_fp16")]; + tensor scatter_41_mode_0 = const()[name = tensor("scatter_41_mode_0"), val = tensor("update")]; + tensor scatter_41_axis_0 = const()[name = tensor("scatter_41_axis_0"), val = tensor(0)]; + tensor scatter_41_validate_indices_0 = const()[name = tensor("scatter_41_validate_indices_0"), val = tensor(false)]; + tensor scatter_41_cast_fp16 = scatter(axis = scatter_41_axis_0, data = reshape_208_cast_fp16, indices = reshape_206, mode = scatter_41_mode_0, updates = reshape_207_cast_fp16, validate_indices = scatter_41_validate_indices_0)[name = tensor("scatter_41_cast_fp16")]; + tensor reshape_209_cast_fp16 = reshape(shape = shape_87, x = scatter_41_cast_fp16)[name = tensor("reshape_209_cast_fp16")]; + tensor reshape_209_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("reshape_209_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor keys_121_begin_0 = const()[name = tensor("keys_121_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor keys_121_end_0 = const()[name = tensor("keys_121_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor keys_121_end_mask_0 = const()[name = tensor("keys_121_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor keys_121_squeeze_mask_0 = const()[name = tensor("keys_121_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor keys_121_cast_fp16 = slice_by_index(begin = keys_121_begin_0, end = keys_121_end_0, end_mask = keys_121_end_mask_0, squeeze_mask = keys_121_squeeze_mask_0, x = reshape_209_cast_fp16)[name = tensor("keys_121_cast_fp16")]; + tensor values_121_begin_0 = const()[name = tensor("values_121_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor values_121_end_0 = const()[name = tensor("values_121_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor values_121_end_mask_0 = const()[name = tensor("values_121_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor values_121_squeeze_mask_0 = const()[name = tensor("values_121_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor values_121_cast_fp16 = slice_by_index(begin = values_121_begin_0, end = values_121_end_0, end_mask = values_121_end_mask_0, squeeze_mask = values_121_squeeze_mask_0, x = reshape_209_cast_fp16)[name = tensor("values_121_cast_fp16")]; + tensor var_8684_cast_fp16 = not_equal(x = keys_121_cast_fp16, y = keys_121_cast_fp16)[name = tensor("op_8684_cast_fp16")]; + tensor keys_123_cast_fp16 = select(a = var_470_to_fp16, b = keys_121_cast_fp16, cond = var_8684_cast_fp16)[name = tensor("keys_123_cast_fp16")]; + tensor var_8692_cast_fp16 = not_equal(x = values_121_cast_fp16, y = values_121_cast_fp16)[name = tensor("op_8692_cast_fp16")]; + tensor values_123_cast_fp16 = select(a = var_470_to_fp16, b = values_121_cast_fp16, cond = var_8692_cast_fp16)[name = tensor("values_123_cast_fp16")]; + tensor var_8716 = const()[name = tensor("op_8716"), val = tensor([0, 2, 1, 3])]; + tensor var_8732 = const()[name = tensor("op_8732"), val = tensor([1, 1, 1])]; + tensor var_8733_cast_fp16 = reshape(shape = var_8732, x = position20_to_fp16)[name = tensor("op_8733_cast_fp16")]; + tensor q_positions_41_cast_fp16 = add(x = var_8733_cast_fp16, y = q_offsets_1_promoted_to_fp16)[name = tensor("q_positions_41_cast_fp16")]; + tensor valid_end_41_cast_fp16 = add(x = var_8733_cast_fp16, y = var_536_cast_fp16)[name = tensor("valid_end_41_cast_fp16")]; + tensor valid_mask_41_cast_fp16 = less(x = k_positions_1_promoted_to_fp16, y = valid_end_41_cast_fp16)[name = tensor("valid_mask_41_cast_fp16")]; + tensor causal_mask_41_cast_fp16 = less_equal(x = k_positions_1_promoted_to_fp16, y = q_positions_41_cast_fp16)[name = tensor("causal_mask_41_cast_fp16")]; + tensor var_8761 = logical_and(x = valid_mask_41_cast_fp16, y = causal_mask_41_cast_fp16)[name = tensor("op_8761")]; + tensor attn_mask_41_axes_0 = const()[name = tensor("attn_mask_41_axes_0"), val = tensor([1])]; + tensor attn_mask_41 = expand_dims(axes = attn_mask_41_axes_0, x = var_8761)[name = tensor("attn_mask_41")]; + tensor var_8775_transpose_x_0 = const()[name = tensor("op_8775_transpose_x_0"), val = tensor(false)]; + tensor var_8775_transpose_y_0 = const()[name = tensor("op_8775_transpose_y_0"), val = tensor(false)]; + tensor transpose_109_perm_0 = const()[name = tensor("transpose_109_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_110_perm_0 = const()[name = tensor("transpose_110_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_110 = transpose(perm = transpose_110_perm_0, x = keys_123_cast_fp16)[name = tensor("transpose_124")]; + tensor transpose_109 = transpose(perm = transpose_109_perm_0, x = q_123_cast_fp16)[name = tensor("transpose_125")]; + tensor var_8775_cast_fp16 = matmul(transpose_x = var_8775_transpose_x_0, transpose_y = var_8775_transpose_y_0, x = transpose_109, y = transpose_110)[name = tensor("op_8775_cast_fp16")]; + tensor var_8769_to_fp16 = const()[name = tensor("op_8769_to_fp16"), val = tensor([0x1p-3])]; + tensor attn_weights_121_cast_fp16 = mul(x = var_8775_cast_fp16, y = var_8769_to_fp16)[name = tensor("attn_weights_121_cast_fp16")]; + tensor var_8777 = logical_not(x = attn_mask_41)[name = tensor("op_8777")]; + tensor var_8778_to_fp16 = const()[name = tensor("op_8778_to_fp16"), val = tensor(-inf)]; + tensor attn_weights_123_cast_fp16 = select(a = var_8778_to_fp16, b = attn_weights_121_cast_fp16, cond = var_8777)[name = tensor("attn_weights_123_cast_fp16")]; + tensor var_8780 = const()[name = tensor("op_8780"), val = tensor(-1)]; + tensor attn_weights_125_cast_fp16 = softmax(axis = var_8780, x = attn_weights_123_cast_fp16)[name = tensor("attn_weights_125_cast_fp16")]; + tensor attn_output_41_transpose_x_0 = const()[name = tensor("attn_output_41_transpose_x_0"), val = tensor(false)]; + tensor attn_output_41_transpose_y_0 = const()[name = tensor("attn_output_41_transpose_y_0"), val = tensor(false)]; + tensor values_125_cast_fp16 = transpose(perm = var_8716, x = values_123_cast_fp16)[name = tensor("transpose_126")]; + tensor attn_output_41_cast_fp16 = matmul(transpose_x = attn_output_41_transpose_x_0, transpose_y = attn_output_41_transpose_y_0, x = attn_weights_125_cast_fp16, y = values_125_cast_fp16)[name = tensor("attn_output_41_cast_fp16")]; + tensor var_8788 = const()[name = tensor("op_8788"), val = tensor([0, 2, 1, 3])]; + tensor var_8793 = const()[name = tensor("op_8793"), val = tensor([1, 256, 1024])]; + tensor var_8789_cast_fp16 = transpose(perm = var_8788, x = attn_output_41_cast_fp16)[name = tensor("transpose_123")]; + tensor input_245_cast_fp16 = reshape(shape = var_8793, x = var_8789_cast_fp16)[name = tensor("input_245_cast_fp16")]; + tensor attn20_out_proj_weight_to_fp16 = const()[name = tensor("attn20_out_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(510857088)))]; + tensor linear_81_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = attn20_out_proj_weight_to_fp16, x = input_245_cast_fp16)[name = tensor("linear_81_cast_fp16")]; + tensor var_8802_cast_fp16 = add(x = position20_to_fp16, y = valid_len_to_fp16)[name = tensor("op_8802_cast_fp16")]; + tensor var_8802_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_8802_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor input_247_cast_fp16 = add(x = input_241_cast_fp16, y = linear_81_cast_fp16)[name = tensor("input_247_cast_fp16")]; + tensor input_249_axes_0 = const()[name = tensor("input_249_axes_0"), val = tensor([-1])]; + tensor norm20_2_weight_to_fp16 = const()[name = tensor("norm20_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(512954304)))]; + tensor norm20_2_bias_to_fp16 = const()[name = tensor("norm20_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(512956416)))]; + tensor var_8806_to_fp16 = const()[name = tensor("op_8806_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_249_cast_fp16 = layer_norm(axes = input_249_axes_0, beta = norm20_2_bias_to_fp16, epsilon = var_8806_to_fp16, gamma = norm20_2_weight_to_fp16, x = input_247_cast_fp16)[name = tensor("input_249_cast_fp16")]; + tensor linear20_1_weight_to_fp16 = const()[name = tensor("linear20_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(512958528)))]; + tensor linear_82_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = linear20_1_weight_to_fp16, x = input_249_cast_fp16)[name = tensor("linear_82_cast_fp16")]; + tensor input_251_mode_0 = const()[name = tensor("input_251_mode_0"), val = tensor("EXACT")]; + tensor input_251_cast_fp16 = gelu(mode = input_251_mode_0, x = linear_82_cast_fp16)[name = tensor("input_251_cast_fp16")]; + tensor linear20_2_weight_to_fp16 = const()[name = tensor("linear20_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(521347200)))]; + tensor linear_83_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = linear20_2_weight_to_fp16, x = input_251_cast_fp16)[name = tensor("linear_83_cast_fp16")]; + tensor input_253_cast_fp16 = add(x = input_247_cast_fp16, y = linear_83_cast_fp16)[name = tensor("input_253_cast_fp16")]; + tensor input_255_axes_0 = const()[name = tensor("input_255_axes_0"), val = tensor([-1])]; + tensor norm21_1_weight_to_fp16 = const()[name = tensor("norm21_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(529735872)))]; + tensor norm21_1_bias_to_fp16 = const()[name = tensor("norm21_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(529737984)))]; + tensor var_8823_to_fp16 = const()[name = tensor("op_8823_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_255_cast_fp16 = layer_norm(axes = input_255_axes_0, beta = norm21_1_bias_to_fp16, epsilon = var_8823_to_fp16, gamma = norm21_1_weight_to_fp16, x = input_253_cast_fp16)[name = tensor("input_255_cast_fp16")]; + tensor attn21_in_proj_weight_to_fp16 = const()[name = tensor("attn21_in_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(529740096)))]; + tensor linear_84_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = attn21_in_proj_weight_to_fp16, x = input_255_cast_fp16)[name = tensor("linear_84_cast_fp16")]; + tensor var_8837 = const()[name = tensor("op_8837"), val = tensor([1, 256, 3, 16, 64])]; + tensor qkv_43_cast_fp16 = reshape(shape = var_8837, x = linear_84_cast_fp16)[name = tensor("qkv_43_cast_fp16")]; + tensor q_127_begin_0 = const()[name = tensor("q_127_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor q_127_end_0 = const()[name = tensor("q_127_end_0"), val = tensor([1, 256, 1, 16, 64])]; + tensor q_127_end_mask_0 = const()[name = tensor("q_127_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor q_127_squeeze_mask_0 = const()[name = tensor("q_127_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor q_127_cast_fp16 = slice_by_index(begin = q_127_begin_0, end = q_127_end_0, end_mask = q_127_end_mask_0, squeeze_mask = q_127_squeeze_mask_0, x = qkv_43_cast_fp16)[name = tensor("q_127_cast_fp16")]; + tensor k_85_begin_0 = const()[name = tensor("k_85_begin_0"), val = tensor([0, 0, 1, 0, 0])]; + tensor k_85_end_0 = const()[name = tensor("k_85_end_0"), val = tensor([1, 256, 2, 16, 64])]; + tensor k_85_end_mask_0 = const()[name = tensor("k_85_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor k_85_squeeze_mask_0 = const()[name = tensor("k_85_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor k_85_cast_fp16 = slice_by_index(begin = k_85_begin_0, end = k_85_end_0, end_mask = k_85_end_mask_0, squeeze_mask = k_85_squeeze_mask_0, x = qkv_43_cast_fp16)[name = tensor("k_85_cast_fp16")]; + tensor v_43_begin_0 = const()[name = tensor("v_43_begin_0"), val = tensor([0, 0, 2, 0, 0])]; + tensor v_43_end_0 = const()[name = tensor("v_43_end_0"), val = tensor([1, 256, 3, 16, 64])]; + tensor v_43_end_mask_0 = const()[name = tensor("v_43_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor v_43_squeeze_mask_0 = const()[name = tensor("v_43_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor v_43_cast_fp16 = slice_by_index(begin = v_43_begin_0, end = v_43_end_0, end_mask = v_43_end_mask_0, squeeze_mask = v_43_squeeze_mask_0, x = qkv_43_cast_fp16)[name = tensor("v_43_cast_fp16")]; + tensor var_8895 = const()[name = tensor("op_8895"), val = tensor([1, 1])]; + tensor position21_to_fp16_dtype_0 = const()[name = tensor("position21_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor position21_to_fp16 = cast(dtype = position21_to_fp16_dtype_0, x = position21)[name = tensor("cast_556")]; + tensor var_8896_cast_fp16 = reshape(shape = var_8895, x = position21_to_fp16)[name = tensor("op_8896_cast_fp16")]; + tensor ts_129_cast_fp16 = add(x = ts_1_promoted_to_fp16, y = var_8896_cast_fp16)[name = tensor("ts_129_cast_fp16")]; + tensor var_8903 = const()[name = tensor("op_8903"), val = tensor([1, 256, 1, 1])]; + tensor ts_131_cast_fp16 = reshape(shape = var_8903, x = ts_129_cast_fp16)[name = tensor("ts_131_cast_fp16")]; + tensor var_8910 = const()[name = tensor("op_8910"), val = tensor([1, 256, 16, 32, 2])]; + tensor q_complex_43_cast_fp16 = reshape(shape = var_8910, x = q_127_cast_fp16)[name = tensor("q_complex_43_cast_fp16")]; + tensor var_8917 = const()[name = tensor("op_8917"), val = tensor([1, 256, 16, 32, 2])]; + tensor k_complex_43_cast_fp16 = reshape(shape = var_8917, x = k_85_cast_fp16)[name = tensor("k_complex_43_cast_fp16")]; + tensor var_8921_begin_0 = const()[name = tensor("op_8921_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_8921_end_0 = const()[name = tensor("op_8921_end_0"), val = tensor([1, 256, 16, 32, 1])]; + tensor var_8921_end_mask_0 = const()[name = tensor("op_8921_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_8921_squeeze_mask_0 = const()[name = tensor("op_8921_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_8921_cast_fp16 = slice_by_index(begin = var_8921_begin_0, end = var_8921_end_0, end_mask = var_8921_end_mask_0, squeeze_mask = var_8921_squeeze_mask_0, x = q_complex_43_cast_fp16)[name = tensor("op_8921_cast_fp16")]; + tensor var_8929_begin_0 = const()[name = tensor("op_8929_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_8929_end_0 = const()[name = tensor("op_8929_end_0"), val = tensor([1, 256, 16, 32, 2])]; + tensor var_8929_end_mask_0 = const()[name = tensor("op_8929_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_8929_squeeze_mask_0 = const()[name = tensor("op_8929_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_8929_cast_fp16 = slice_by_index(begin = var_8929_begin_0, end = var_8929_end_0, end_mask = var_8929_end_mask_0, squeeze_mask = var_8929_squeeze_mask_0, x = q_complex_43_cast_fp16)[name = tensor("op_8929_cast_fp16")]; + tensor var_8937_begin_0 = const()[name = tensor("op_8937_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_8937_end_0 = const()[name = tensor("op_8937_end_0"), val = tensor([1, 256, 16, 32, 1])]; + tensor var_8937_end_mask_0 = const()[name = tensor("op_8937_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_8937_squeeze_mask_0 = const()[name = tensor("op_8937_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_8937_cast_fp16 = slice_by_index(begin = var_8937_begin_0, end = var_8937_end_0, end_mask = var_8937_end_mask_0, squeeze_mask = var_8937_squeeze_mask_0, x = k_complex_43_cast_fp16)[name = tensor("op_8937_cast_fp16")]; + tensor var_8945_begin_0 = const()[name = tensor("op_8945_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_8945_end_0 = const()[name = tensor("op_8945_end_0"), val = tensor([1, 256, 16, 32, 2])]; + tensor var_8945_end_mask_0 = const()[name = tensor("op_8945_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_8945_squeeze_mask_0 = const()[name = tensor("op_8945_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_8945_cast_fp16 = slice_by_index(begin = var_8945_begin_0, end = var_8945_end_0, end_mask = var_8945_end_mask_0, squeeze_mask = var_8945_squeeze_mask_0, x = k_complex_43_cast_fp16)[name = tensor("op_8945_cast_fp16")]; + tensor freqs_43_to_fp16 = const()[name = tensor("freqs_43_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(536031616)))]; + tensor var_8951_cast_fp16 = mul(x = freqs_43_to_fp16, y = ts_131_cast_fp16)[name = tensor("op_8951_cast_fp16")]; + tensor rotr_43_cast_fp16 = cos(x = var_8951_cast_fp16)[name = tensor("rotr_43_cast_fp16")]; + tensor roti_43_cast_fp16 = sin(x = var_8951_cast_fp16)[name = tensor("roti_43_cast_fp16")]; + tensor var_8955_cast_fp16 = mul(x = var_8921_cast_fp16, y = rotr_43_cast_fp16)[name = tensor("op_8955_cast_fp16")]; + tensor var_8956_cast_fp16 = mul(x = var_8929_cast_fp16, y = roti_43_cast_fp16)[name = tensor("op_8956_cast_fp16")]; + tensor qor_85_cast_fp16 = sub(x = var_8955_cast_fp16, y = var_8956_cast_fp16)[name = tensor("qor_85_cast_fp16")]; + tensor var_8959_cast_fp16 = mul(x = var_8921_cast_fp16, y = roti_43_cast_fp16)[name = tensor("op_8959_cast_fp16")]; + tensor var_8960_cast_fp16 = mul(x = var_8929_cast_fp16, y = rotr_43_cast_fp16)[name = tensor("op_8960_cast_fp16")]; + tensor qoi_85_cast_fp16 = add(x = var_8959_cast_fp16, y = var_8960_cast_fp16)[name = tensor("qoi_85_cast_fp16")]; + tensor var_8963_cast_fp16 = mul(x = var_8937_cast_fp16, y = rotr_43_cast_fp16)[name = tensor("op_8963_cast_fp16")]; + tensor var_8964_cast_fp16 = mul(x = var_8945_cast_fp16, y = roti_43_cast_fp16)[name = tensor("op_8964_cast_fp16")]; + tensor kor_85_cast_fp16 = sub(x = var_8963_cast_fp16, y = var_8964_cast_fp16)[name = tensor("kor_85_cast_fp16")]; + tensor var_8967_cast_fp16 = mul(x = var_8937_cast_fp16, y = roti_43_cast_fp16)[name = tensor("op_8967_cast_fp16")]; + tensor var_8968_cast_fp16 = mul(x = var_8945_cast_fp16, y = rotr_43_cast_fp16)[name = tensor("op_8968_cast_fp16")]; + tensor koi_85_cast_fp16 = add(x = var_8967_cast_fp16, y = var_8968_cast_fp16)[name = tensor("koi_85_cast_fp16")]; + tensor qo_43_axis_0 = const()[name = tensor("qo_43_axis_0"), val = tensor(-1)]; + tensor qo_43_cast_fp16 = stack(axis = qo_43_axis_0, values = (qor_85_cast_fp16, qoi_85_cast_fp16))[name = tensor("qo_43_cast_fp16")]; + tensor ko_43_axis_0 = const()[name = tensor("ko_43_axis_0"), val = tensor(-1)]; + tensor ko_43_cast_fp16 = stack(axis = ko_43_axis_0, values = (kor_85_cast_fp16, koi_85_cast_fp16))[name = tensor("ko_43_cast_fp16")]; + tensor var_9001 = const()[name = tensor("op_9001"), val = tensor([1, 256, 16, 64])]; + tensor q_129_cast_fp16 = reshape(shape = var_9001, x = qo_43_cast_fp16)[name = tensor("q_129_cast_fp16")]; + tensor var_9007 = const()[name = tensor("op_9007"), val = tensor([1, 256, 16, 64])]; + tensor k_87_cast_fp16 = reshape(shape = var_9007, x = ko_43_cast_fp16)[name = tensor("k_87_cast_fp16")]; + tensor write_indices_float_127_cast_fp16 = add(x = var_8896_cast_fp16, y = var_393_promoted_to_fp16)[name = tensor("write_indices_float_127_cast_fp16")]; + tensor _inversed_9028_y_0_to_fp16 = const()[name = tensor("_inversed_9028_y_0_to_fp16"), val = tensor(0x1p-9)]; + tensor _inversed_9028_cast_fp16 = mul(x = write_indices_float_127_cast_fp16, y = _inversed_9028_y_0_to_fp16)[name = tensor("_inversed_9028_cast_fp16")]; + tensor var_9029_cast_fp16 = floor(x = _inversed_9028_cast_fp16)[name = tensor("op_9029_cast_fp16")]; + tensor var_9030_to_fp16 = const()[name = tensor("op_9030_to_fp16"), val = tensor(0x1p+9)]; + tensor var_9031_cast_fp16 = mul(x = var_9029_cast_fp16, y = var_9030_to_fp16)[name = tensor("op_9031_cast_fp16")]; + tensor write_indices_float_129_cast_fp16 = sub(x = write_indices_float_127_cast_fp16, y = var_9031_cast_fp16)[name = tensor("write_indices_float_129_cast_fp16")]; + tensor var_9049_value_0_to_fp16 = const()[name = tensor("op_9049_value_0_to_fp16"), val = tensor(0x1.ffp+8)]; + tensor var_9049_cast_fp16 = fill_like(ref_tensor = write_indices_float_129_cast_fp16, value = var_9049_value_0_to_fp16)[name = tensor("op_9049_cast_fp16")]; + tensor write_indices_float_131_cast_fp16 = select(a = write_indices_float_129_cast_fp16, b = var_9049_cast_fp16, cond = token_valid_1_cast_fp16)[name = tensor("write_indices_float_131_cast_fp16")]; + tensor var_9055_dtype_0 = const()[name = tensor("op_9055_dtype_0"), val = tensor("int32")]; + tensor var_9060 = const()[name = tensor("op_9060"), val = tensor([1, 256, 1, 1])]; + tensor write_indices_float_131_cast_fp16_to_int32 = cast(dtype = var_9055_dtype_0, x = write_indices_float_131_cast_fp16)[name = tensor("cast_555")]; + tensor var_9061 = reshape(shape = var_9060, x = write_indices_float_131_cast_fp16_to_int32)[name = tensor("op_9061")]; + tensor write_indices_43_reps_0 = const()[name = tensor("write_indices_43_reps_0"), val = tensor([1, 1, 16, 64])]; + tensor write_indices_43 = tile(reps = write_indices_43_reps_0, x = var_9061)[name = tensor("write_indices_43")]; + tensor var_9071_begin_0 = const()[name = tensor("op_9071_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_9071_end_0 = const()[name = tensor("op_9071_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor var_9071_end_mask_0 = const()[name = tensor("op_9071_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_9071_squeeze_mask_0 = const()[name = tensor("op_9071_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor cache21_to_fp16_dtype_0 = const()[name = tensor("cache21_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor cache21_to_fp16 = cast(dtype = cache21_to_fp16_dtype_0, x = cache21)[name = tensor("cast_554")]; + tensor var_9071_cast_fp16 = slice_by_index(begin = var_9071_begin_0, end = var_9071_end_0, end_mask = var_9071_end_mask_0, squeeze_mask = var_9071_squeeze_mask_0, x = cache21_to_fp16)[name = tensor("op_9071_cast_fp16")]; + tensor var_9073_axis_0 = const()[name = tensor("op_9073_axis_0"), val = tensor(1)]; + tensor var_9073_mode_0 = const()[name = tensor("op_9073_mode_0"), val = tensor("update")]; + tensor var_9073_validate_indices_0 = const()[name = tensor("op_9073_validate_indices_0"), val = tensor(false)]; + tensor var_9073_cast_fp16 = scatter_along_axis(axis = var_9073_axis_0, data = var_9071_cast_fp16, indices = write_indices_43, mode = var_9073_mode_0, updates = k_87_cast_fp16, validate_indices = var_9073_validate_indices_0)[name = tensor("op_9073_cast_fp16")]; + tensor concat_170 = const()[name = tensor("concat_170"), val = tensor([0, 0, 0, 0, 0])]; + tensor concat_171 = const()[name = tensor("concat_171"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_43_internal_tensor_assign_1_stride_0 = const()[name = tensor("new_cache_43_internal_tensor_assign_1_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_43_internal_tensor_assign_1_begin_mask_0 = const()[name = tensor("new_cache_43_internal_tensor_assign_1_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_43_internal_tensor_assign_1_end_mask_0 = const()[name = tensor("new_cache_43_internal_tensor_assign_1_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_43_internal_tensor_assign_1_squeeze_mask_0 = const()[name = tensor("new_cache_43_internal_tensor_assign_1_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_88 = const()[name = tensor("shape_88"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_42 = const()[name = tensor("reduce_prod_42"), val = tensor(1048576)]; + tensor range_1d_42_start_0 = const()[name = tensor("range_1d_42_start_0"), val = tensor(0)]; + tensor range_1d_42_step_0 = const()[name = tensor("range_1d_42_step_0"), val = tensor(1)]; + tensor range_1d_42 = range_1d(end = reduce_prod_42, start = range_1d_42_start_0, step = range_1d_42_step_0)[name = tensor("range_1d_42")]; + tensor reshape_210 = reshape(shape = shape_88, x = range_1d_42)[name = tensor("reshape_210")]; + tensor slice_by_index_42 = slice_by_index(begin = concat_170, begin_mask = new_cache_43_internal_tensor_assign_1_begin_mask_0, end = concat_171, end_mask = new_cache_43_internal_tensor_assign_1_end_mask_0, squeeze_mask = new_cache_43_internal_tensor_assign_1_squeeze_mask_0, stride = new_cache_43_internal_tensor_assign_1_stride_0, x = reshape_210)[name = tensor("slice_by_index_42")]; + tensor reshape_211_shape_0 = const()[name = tensor("reshape_211_shape_0"), val = tensor([-1])]; + tensor reshape_211 = reshape(shape = reshape_211_shape_0, x = slice_by_index_42)[name = tensor("reshape_211")]; + tensor reshape_212_shape_0 = const()[name = tensor("reshape_212_shape_0"), val = tensor([-1])]; + tensor reshape_212_cast_fp16 = reshape(shape = reshape_212_shape_0, x = var_9073_cast_fp16)[name = tensor("reshape_212_cast_fp16")]; + tensor reshape_213_shape_0 = const()[name = tensor("reshape_213_shape_0"), val = tensor([-1])]; + tensor reshape_213_cast_fp16 = reshape(shape = reshape_213_shape_0, x = cache21_to_fp16)[name = tensor("reshape_213_cast_fp16")]; + tensor scatter_42_mode_0 = const()[name = tensor("scatter_42_mode_0"), val = tensor("update")]; + tensor scatter_42_axis_0 = const()[name = tensor("scatter_42_axis_0"), val = tensor(0)]; + tensor scatter_42_validate_indices_0 = const()[name = tensor("scatter_42_validate_indices_0"), val = tensor(false)]; + tensor scatter_42_cast_fp16 = scatter(axis = scatter_42_axis_0, data = reshape_213_cast_fp16, indices = reshape_211, mode = scatter_42_mode_0, updates = reshape_212_cast_fp16, validate_indices = scatter_42_validate_indices_0)[name = tensor("scatter_42_cast_fp16")]; + tensor reshape_214_cast_fp16 = reshape(shape = shape_88, x = scatter_42_cast_fp16)[name = tensor("reshape_214_cast_fp16")]; + tensor var_9081_begin_0 = const()[name = tensor("op_9081_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor var_9081_end_0 = const()[name = tensor("op_9081_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor var_9081_end_mask_0 = const()[name = tensor("op_9081_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_9081_squeeze_mask_0 = const()[name = tensor("op_9081_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor var_9081_cast_fp16 = slice_by_index(begin = var_9081_begin_0, end = var_9081_end_0, end_mask = var_9081_end_mask_0, squeeze_mask = var_9081_squeeze_mask_0, x = reshape_214_cast_fp16)[name = tensor("op_9081_cast_fp16")]; + tensor var_9083_axis_0 = const()[name = tensor("op_9083_axis_0"), val = tensor(1)]; + tensor var_9083_mode_0 = const()[name = tensor("op_9083_mode_0"), val = tensor("update")]; + tensor var_9083_validate_indices_0 = const()[name = tensor("op_9083_validate_indices_0"), val = tensor(false)]; + tensor var_9083_cast_fp16 = scatter_along_axis(axis = var_9083_axis_0, data = var_9081_cast_fp16, indices = write_indices_43, mode = var_9083_mode_0, updates = v_43_cast_fp16, validate_indices = var_9083_validate_indices_0)[name = tensor("op_9083_cast_fp16")]; + tensor concat_172 = const()[name = tensor("concat_172"), val = tensor([1, 0, 0, 0, 0])]; + tensor concat_173 = const()[name = tensor("concat_173"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_43_internal_tensor_assign_2_stride_0 = const()[name = tensor("new_cache_43_internal_tensor_assign_2_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_43_internal_tensor_assign_2_begin_mask_0 = const()[name = tensor("new_cache_43_internal_tensor_assign_2_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_43_internal_tensor_assign_2_end_mask_0 = const()[name = tensor("new_cache_43_internal_tensor_assign_2_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_43_internal_tensor_assign_2_squeeze_mask_0 = const()[name = tensor("new_cache_43_internal_tensor_assign_2_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_89 = const()[name = tensor("shape_89"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_43 = const()[name = tensor("reduce_prod_43"), val = tensor(1048576)]; + tensor range_1d_43_start_0 = const()[name = tensor("range_1d_43_start_0"), val = tensor(0)]; + tensor range_1d_43_step_0 = const()[name = tensor("range_1d_43_step_0"), val = tensor(1)]; + tensor range_1d_43 = range_1d(end = reduce_prod_43, start = range_1d_43_start_0, step = range_1d_43_step_0)[name = tensor("range_1d_43")]; + tensor reshape_215 = reshape(shape = shape_89, x = range_1d_43)[name = tensor("reshape_215")]; + tensor slice_by_index_43 = slice_by_index(begin = concat_172, begin_mask = new_cache_43_internal_tensor_assign_2_begin_mask_0, end = concat_173, end_mask = new_cache_43_internal_tensor_assign_2_end_mask_0, squeeze_mask = new_cache_43_internal_tensor_assign_2_squeeze_mask_0, stride = new_cache_43_internal_tensor_assign_2_stride_0, x = reshape_215)[name = tensor("slice_by_index_43")]; + tensor reshape_216_shape_0 = const()[name = tensor("reshape_216_shape_0"), val = tensor([-1])]; + tensor reshape_216 = reshape(shape = reshape_216_shape_0, x = slice_by_index_43)[name = tensor("reshape_216")]; + tensor reshape_217_shape_0 = const()[name = tensor("reshape_217_shape_0"), val = tensor([-1])]; + tensor reshape_217_cast_fp16 = reshape(shape = reshape_217_shape_0, x = var_9083_cast_fp16)[name = tensor("reshape_217_cast_fp16")]; + tensor reshape_218_shape_0 = const()[name = tensor("reshape_218_shape_0"), val = tensor([-1])]; + tensor reshape_218_cast_fp16 = reshape(shape = reshape_218_shape_0, x = reshape_214_cast_fp16)[name = tensor("reshape_218_cast_fp16")]; + tensor scatter_43_mode_0 = const()[name = tensor("scatter_43_mode_0"), val = tensor("update")]; + tensor scatter_43_axis_0 = const()[name = tensor("scatter_43_axis_0"), val = tensor(0)]; + tensor scatter_43_validate_indices_0 = const()[name = tensor("scatter_43_validate_indices_0"), val = tensor(false)]; + tensor scatter_43_cast_fp16 = scatter(axis = scatter_43_axis_0, data = reshape_218_cast_fp16, indices = reshape_216, mode = scatter_43_mode_0, updates = reshape_217_cast_fp16, validate_indices = scatter_43_validate_indices_0)[name = tensor("scatter_43_cast_fp16")]; + tensor reshape_219_cast_fp16 = reshape(shape = shape_89, x = scatter_43_cast_fp16)[name = tensor("reshape_219_cast_fp16")]; + tensor reshape_219_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("reshape_219_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor keys_127_begin_0 = const()[name = tensor("keys_127_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor keys_127_end_0 = const()[name = tensor("keys_127_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor keys_127_end_mask_0 = const()[name = tensor("keys_127_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor keys_127_squeeze_mask_0 = const()[name = tensor("keys_127_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor keys_127_cast_fp16 = slice_by_index(begin = keys_127_begin_0, end = keys_127_end_0, end_mask = keys_127_end_mask_0, squeeze_mask = keys_127_squeeze_mask_0, x = reshape_219_cast_fp16)[name = tensor("keys_127_cast_fp16")]; + tensor values_127_begin_0 = const()[name = tensor("values_127_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor values_127_end_0 = const()[name = tensor("values_127_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor values_127_end_mask_0 = const()[name = tensor("values_127_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor values_127_squeeze_mask_0 = const()[name = tensor("values_127_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor values_127_cast_fp16 = slice_by_index(begin = values_127_begin_0, end = values_127_end_0, end_mask = values_127_end_mask_0, squeeze_mask = values_127_squeeze_mask_0, x = reshape_219_cast_fp16)[name = tensor("values_127_cast_fp16")]; + tensor var_9095_cast_fp16 = not_equal(x = keys_127_cast_fp16, y = keys_127_cast_fp16)[name = tensor("op_9095_cast_fp16")]; + tensor keys_129_cast_fp16 = select(a = var_470_to_fp16, b = keys_127_cast_fp16, cond = var_9095_cast_fp16)[name = tensor("keys_129_cast_fp16")]; + tensor var_9103_cast_fp16 = not_equal(x = values_127_cast_fp16, y = values_127_cast_fp16)[name = tensor("op_9103_cast_fp16")]; + tensor values_129_cast_fp16 = select(a = var_470_to_fp16, b = values_127_cast_fp16, cond = var_9103_cast_fp16)[name = tensor("values_129_cast_fp16")]; + tensor var_9127 = const()[name = tensor("op_9127"), val = tensor([0, 2, 1, 3])]; + tensor var_9143 = const()[name = tensor("op_9143"), val = tensor([1, 1, 1])]; + tensor var_9144_cast_fp16 = reshape(shape = var_9143, x = position21_to_fp16)[name = tensor("op_9144_cast_fp16")]; + tensor q_positions_43_cast_fp16 = add(x = var_9144_cast_fp16, y = q_offsets_1_promoted_to_fp16)[name = tensor("q_positions_43_cast_fp16")]; + tensor valid_end_43_cast_fp16 = add(x = var_9144_cast_fp16, y = var_536_cast_fp16)[name = tensor("valid_end_43_cast_fp16")]; + tensor valid_mask_43_cast_fp16 = less(x = k_positions_1_promoted_to_fp16, y = valid_end_43_cast_fp16)[name = tensor("valid_mask_43_cast_fp16")]; + tensor causal_mask_43_cast_fp16 = less_equal(x = k_positions_1_promoted_to_fp16, y = q_positions_43_cast_fp16)[name = tensor("causal_mask_43_cast_fp16")]; + tensor var_9172 = logical_and(x = valid_mask_43_cast_fp16, y = causal_mask_43_cast_fp16)[name = tensor("op_9172")]; + tensor attn_mask_43_axes_0 = const()[name = tensor("attn_mask_43_axes_0"), val = tensor([1])]; + tensor attn_mask_43 = expand_dims(axes = attn_mask_43_axes_0, x = var_9172)[name = tensor("attn_mask_43")]; + tensor var_9186_transpose_x_0 = const()[name = tensor("op_9186_transpose_x_0"), val = tensor(false)]; + tensor var_9186_transpose_y_0 = const()[name = tensor("op_9186_transpose_y_0"), val = tensor(false)]; + tensor transpose_111_perm_0 = const()[name = tensor("transpose_111_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_112_perm_0 = const()[name = tensor("transpose_112_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_112 = transpose(perm = transpose_112_perm_0, x = keys_129_cast_fp16)[name = tensor("transpose_120")]; + tensor transpose_111 = transpose(perm = transpose_111_perm_0, x = q_129_cast_fp16)[name = tensor("transpose_121")]; + tensor var_9186_cast_fp16 = matmul(transpose_x = var_9186_transpose_x_0, transpose_y = var_9186_transpose_y_0, x = transpose_111, y = transpose_112)[name = tensor("op_9186_cast_fp16")]; + tensor var_9180_to_fp16 = const()[name = tensor("op_9180_to_fp16"), val = tensor([0x1p-3])]; + tensor attn_weights_127_cast_fp16 = mul(x = var_9186_cast_fp16, y = var_9180_to_fp16)[name = tensor("attn_weights_127_cast_fp16")]; + tensor var_9188 = logical_not(x = attn_mask_43)[name = tensor("op_9188")]; + tensor var_9189_to_fp16 = const()[name = tensor("op_9189_to_fp16"), val = tensor(-inf)]; + tensor attn_weights_129_cast_fp16 = select(a = var_9189_to_fp16, b = attn_weights_127_cast_fp16, cond = var_9188)[name = tensor("attn_weights_129_cast_fp16")]; + tensor var_9191 = const()[name = tensor("op_9191"), val = tensor(-1)]; + tensor attn_weights_131_cast_fp16 = softmax(axis = var_9191, x = attn_weights_129_cast_fp16)[name = tensor("attn_weights_131_cast_fp16")]; + tensor attn_output_43_transpose_x_0 = const()[name = tensor("attn_output_43_transpose_x_0"), val = tensor(false)]; + tensor attn_output_43_transpose_y_0 = const()[name = tensor("attn_output_43_transpose_y_0"), val = tensor(false)]; + tensor values_131_cast_fp16 = transpose(perm = var_9127, x = values_129_cast_fp16)[name = tensor("transpose_122")]; + tensor attn_output_43_cast_fp16 = matmul(transpose_x = attn_output_43_transpose_x_0, transpose_y = attn_output_43_transpose_y_0, x = attn_weights_131_cast_fp16, y = values_131_cast_fp16)[name = tensor("attn_output_43_cast_fp16")]; + tensor var_9199 = const()[name = tensor("op_9199"), val = tensor([0, 2, 1, 3])]; + tensor var_9204 = const()[name = tensor("op_9204"), val = tensor([1, 256, 1024])]; + tensor var_9200_cast_fp16 = transpose(perm = var_9199, x = attn_output_43_cast_fp16)[name = tensor("transpose_119")]; + tensor input_257_cast_fp16 = reshape(shape = var_9204, x = var_9200_cast_fp16)[name = tensor("input_257_cast_fp16")]; + tensor attn21_out_proj_weight_to_fp16 = const()[name = tensor("attn21_out_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(536031744)))]; + tensor linear_85_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = attn21_out_proj_weight_to_fp16, x = input_257_cast_fp16)[name = tensor("linear_85_cast_fp16")]; + tensor var_9213_cast_fp16 = add(x = position21_to_fp16, y = valid_len_to_fp16)[name = tensor("op_9213_cast_fp16")]; + tensor var_9213_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_9213_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor input_259_cast_fp16 = add(x = input_253_cast_fp16, y = linear_85_cast_fp16)[name = tensor("input_259_cast_fp16")]; + tensor input_261_axes_0 = const()[name = tensor("input_261_axes_0"), val = tensor([-1])]; + tensor norm21_2_weight_to_fp16 = const()[name = tensor("norm21_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(538128960)))]; + tensor norm21_2_bias_to_fp16 = const()[name = tensor("norm21_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(538131072)))]; + tensor var_9217_to_fp16 = const()[name = tensor("op_9217_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_261_cast_fp16 = layer_norm(axes = input_261_axes_0, beta = norm21_2_bias_to_fp16, epsilon = var_9217_to_fp16, gamma = norm21_2_weight_to_fp16, x = input_259_cast_fp16)[name = tensor("input_261_cast_fp16")]; + tensor linear21_1_weight_to_fp16 = const()[name = tensor("linear21_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(538133184)))]; + tensor linear_86_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = linear21_1_weight_to_fp16, x = input_261_cast_fp16)[name = tensor("linear_86_cast_fp16")]; + tensor input_263_mode_0 = const()[name = tensor("input_263_mode_0"), val = tensor("EXACT")]; + tensor input_263_cast_fp16 = gelu(mode = input_263_mode_0, x = linear_86_cast_fp16)[name = tensor("input_263_cast_fp16")]; + tensor linear21_2_weight_to_fp16 = const()[name = tensor("linear21_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(546521856)))]; + tensor linear_87_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = linear21_2_weight_to_fp16, x = input_263_cast_fp16)[name = tensor("linear_87_cast_fp16")]; + tensor input_265_cast_fp16 = add(x = input_259_cast_fp16, y = linear_87_cast_fp16)[name = tensor("input_265_cast_fp16")]; + tensor input_267_axes_0 = const()[name = tensor("input_267_axes_0"), val = tensor([-1])]; + tensor norm22_1_weight_to_fp16 = const()[name = tensor("norm22_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(554910528)))]; + tensor norm22_1_bias_to_fp16 = const()[name = tensor("norm22_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(554912640)))]; + tensor var_9234_to_fp16 = const()[name = tensor("op_9234_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_267_cast_fp16 = layer_norm(axes = input_267_axes_0, beta = norm22_1_bias_to_fp16, epsilon = var_9234_to_fp16, gamma = norm22_1_weight_to_fp16, x = input_265_cast_fp16)[name = tensor("input_267_cast_fp16")]; + tensor attn22_in_proj_weight_to_fp16 = const()[name = tensor("attn22_in_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(554914752)))]; + tensor linear_88_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = attn22_in_proj_weight_to_fp16, x = input_267_cast_fp16)[name = tensor("linear_88_cast_fp16")]; + tensor var_9248 = const()[name = tensor("op_9248"), val = tensor([1, 256, 3, 16, 64])]; + tensor qkv_45_cast_fp16 = reshape(shape = var_9248, x = linear_88_cast_fp16)[name = tensor("qkv_45_cast_fp16")]; + tensor q_133_begin_0 = const()[name = tensor("q_133_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor q_133_end_0 = const()[name = tensor("q_133_end_0"), val = tensor([1, 256, 1, 16, 64])]; + tensor q_133_end_mask_0 = const()[name = tensor("q_133_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor q_133_squeeze_mask_0 = const()[name = tensor("q_133_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor q_133_cast_fp16 = slice_by_index(begin = q_133_begin_0, end = q_133_end_0, end_mask = q_133_end_mask_0, squeeze_mask = q_133_squeeze_mask_0, x = qkv_45_cast_fp16)[name = tensor("q_133_cast_fp16")]; + tensor k_89_begin_0 = const()[name = tensor("k_89_begin_0"), val = tensor([0, 0, 1, 0, 0])]; + tensor k_89_end_0 = const()[name = tensor("k_89_end_0"), val = tensor([1, 256, 2, 16, 64])]; + tensor k_89_end_mask_0 = const()[name = tensor("k_89_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor k_89_squeeze_mask_0 = const()[name = tensor("k_89_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor k_89_cast_fp16 = slice_by_index(begin = k_89_begin_0, end = k_89_end_0, end_mask = k_89_end_mask_0, squeeze_mask = k_89_squeeze_mask_0, x = qkv_45_cast_fp16)[name = tensor("k_89_cast_fp16")]; + tensor v_45_begin_0 = const()[name = tensor("v_45_begin_0"), val = tensor([0, 0, 2, 0, 0])]; + tensor v_45_end_0 = const()[name = tensor("v_45_end_0"), val = tensor([1, 256, 3, 16, 64])]; + tensor v_45_end_mask_0 = const()[name = tensor("v_45_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor v_45_squeeze_mask_0 = const()[name = tensor("v_45_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor v_45_cast_fp16 = slice_by_index(begin = v_45_begin_0, end = v_45_end_0, end_mask = v_45_end_mask_0, squeeze_mask = v_45_squeeze_mask_0, x = qkv_45_cast_fp16)[name = tensor("v_45_cast_fp16")]; + tensor var_9306 = const()[name = tensor("op_9306"), val = tensor([1, 1])]; + tensor position22_to_fp16_dtype_0 = const()[name = tensor("position22_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor position22_to_fp16 = cast(dtype = position22_to_fp16_dtype_0, x = position22)[name = tensor("cast_551")]; + tensor var_9307_cast_fp16 = reshape(shape = var_9306, x = position22_to_fp16)[name = tensor("op_9307_cast_fp16")]; + tensor ts_135_cast_fp16 = add(x = ts_1_promoted_to_fp16, y = var_9307_cast_fp16)[name = tensor("ts_135_cast_fp16")]; + tensor var_9314 = const()[name = tensor("op_9314"), val = tensor([1, 256, 1, 1])]; + tensor ts_137_cast_fp16 = reshape(shape = var_9314, x = ts_135_cast_fp16)[name = tensor("ts_137_cast_fp16")]; + tensor var_9321 = const()[name = tensor("op_9321"), val = tensor([1, 256, 16, 32, 2])]; + tensor q_complex_45_cast_fp16 = reshape(shape = var_9321, x = q_133_cast_fp16)[name = tensor("q_complex_45_cast_fp16")]; + tensor var_9328 = const()[name = tensor("op_9328"), val = tensor([1, 256, 16, 32, 2])]; + tensor k_complex_45_cast_fp16 = reshape(shape = var_9328, x = k_89_cast_fp16)[name = tensor("k_complex_45_cast_fp16")]; + tensor var_9332_begin_0 = const()[name = tensor("op_9332_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_9332_end_0 = const()[name = tensor("op_9332_end_0"), val = tensor([1, 256, 16, 32, 1])]; + tensor var_9332_end_mask_0 = const()[name = tensor("op_9332_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_9332_squeeze_mask_0 = const()[name = tensor("op_9332_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_9332_cast_fp16 = slice_by_index(begin = var_9332_begin_0, end = var_9332_end_0, end_mask = var_9332_end_mask_0, squeeze_mask = var_9332_squeeze_mask_0, x = q_complex_45_cast_fp16)[name = tensor("op_9332_cast_fp16")]; + tensor var_9340_begin_0 = const()[name = tensor("op_9340_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_9340_end_0 = const()[name = tensor("op_9340_end_0"), val = tensor([1, 256, 16, 32, 2])]; + tensor var_9340_end_mask_0 = const()[name = tensor("op_9340_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_9340_squeeze_mask_0 = const()[name = tensor("op_9340_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_9340_cast_fp16 = slice_by_index(begin = var_9340_begin_0, end = var_9340_end_0, end_mask = var_9340_end_mask_0, squeeze_mask = var_9340_squeeze_mask_0, x = q_complex_45_cast_fp16)[name = tensor("op_9340_cast_fp16")]; + tensor var_9348_begin_0 = const()[name = tensor("op_9348_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_9348_end_0 = const()[name = tensor("op_9348_end_0"), val = tensor([1, 256, 16, 32, 1])]; + tensor var_9348_end_mask_0 = const()[name = tensor("op_9348_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_9348_squeeze_mask_0 = const()[name = tensor("op_9348_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_9348_cast_fp16 = slice_by_index(begin = var_9348_begin_0, end = var_9348_end_0, end_mask = var_9348_end_mask_0, squeeze_mask = var_9348_squeeze_mask_0, x = k_complex_45_cast_fp16)[name = tensor("op_9348_cast_fp16")]; + tensor var_9356_begin_0 = const()[name = tensor("op_9356_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_9356_end_0 = const()[name = tensor("op_9356_end_0"), val = tensor([1, 256, 16, 32, 2])]; + tensor var_9356_end_mask_0 = const()[name = tensor("op_9356_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_9356_squeeze_mask_0 = const()[name = tensor("op_9356_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_9356_cast_fp16 = slice_by_index(begin = var_9356_begin_0, end = var_9356_end_0, end_mask = var_9356_end_mask_0, squeeze_mask = var_9356_squeeze_mask_0, x = k_complex_45_cast_fp16)[name = tensor("op_9356_cast_fp16")]; + tensor freqs_45_to_fp16 = const()[name = tensor("freqs_45_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(561206272)))]; + tensor var_9362_cast_fp16 = mul(x = freqs_45_to_fp16, y = ts_137_cast_fp16)[name = tensor("op_9362_cast_fp16")]; + tensor rotr_45_cast_fp16 = cos(x = var_9362_cast_fp16)[name = tensor("rotr_45_cast_fp16")]; + tensor roti_45_cast_fp16 = sin(x = var_9362_cast_fp16)[name = tensor("roti_45_cast_fp16")]; + tensor var_9366_cast_fp16 = mul(x = var_9332_cast_fp16, y = rotr_45_cast_fp16)[name = tensor("op_9366_cast_fp16")]; + tensor var_9367_cast_fp16 = mul(x = var_9340_cast_fp16, y = roti_45_cast_fp16)[name = tensor("op_9367_cast_fp16")]; + tensor qor_89_cast_fp16 = sub(x = var_9366_cast_fp16, y = var_9367_cast_fp16)[name = tensor("qor_89_cast_fp16")]; + tensor var_9370_cast_fp16 = mul(x = var_9332_cast_fp16, y = roti_45_cast_fp16)[name = tensor("op_9370_cast_fp16")]; + tensor var_9371_cast_fp16 = mul(x = var_9340_cast_fp16, y = rotr_45_cast_fp16)[name = tensor("op_9371_cast_fp16")]; + tensor qoi_89_cast_fp16 = add(x = var_9370_cast_fp16, y = var_9371_cast_fp16)[name = tensor("qoi_89_cast_fp16")]; + tensor var_9374_cast_fp16 = mul(x = var_9348_cast_fp16, y = rotr_45_cast_fp16)[name = tensor("op_9374_cast_fp16")]; + tensor var_9375_cast_fp16 = mul(x = var_9356_cast_fp16, y = roti_45_cast_fp16)[name = tensor("op_9375_cast_fp16")]; + tensor kor_89_cast_fp16 = sub(x = var_9374_cast_fp16, y = var_9375_cast_fp16)[name = tensor("kor_89_cast_fp16")]; + tensor var_9378_cast_fp16 = mul(x = var_9348_cast_fp16, y = roti_45_cast_fp16)[name = tensor("op_9378_cast_fp16")]; + tensor var_9379_cast_fp16 = mul(x = var_9356_cast_fp16, y = rotr_45_cast_fp16)[name = tensor("op_9379_cast_fp16")]; + tensor koi_89_cast_fp16 = add(x = var_9378_cast_fp16, y = var_9379_cast_fp16)[name = tensor("koi_89_cast_fp16")]; + tensor qo_45_axis_0 = const()[name = tensor("qo_45_axis_0"), val = tensor(-1)]; + tensor qo_45_cast_fp16 = stack(axis = qo_45_axis_0, values = (qor_89_cast_fp16, qoi_89_cast_fp16))[name = tensor("qo_45_cast_fp16")]; + tensor ko_45_axis_0 = const()[name = tensor("ko_45_axis_0"), val = tensor(-1)]; + tensor ko_45_cast_fp16 = stack(axis = ko_45_axis_0, values = (kor_89_cast_fp16, koi_89_cast_fp16))[name = tensor("ko_45_cast_fp16")]; + tensor var_9412 = const()[name = tensor("op_9412"), val = tensor([1, 256, 16, 64])]; + tensor q_135_cast_fp16 = reshape(shape = var_9412, x = qo_45_cast_fp16)[name = tensor("q_135_cast_fp16")]; + tensor var_9418 = const()[name = tensor("op_9418"), val = tensor([1, 256, 16, 64])]; + tensor k_91_cast_fp16 = reshape(shape = var_9418, x = ko_45_cast_fp16)[name = tensor("k_91_cast_fp16")]; + tensor write_indices_float_133_cast_fp16 = add(x = var_9307_cast_fp16, y = var_393_promoted_to_fp16)[name = tensor("write_indices_float_133_cast_fp16")]; + tensor _inversed_9439_y_0_to_fp16 = const()[name = tensor("_inversed_9439_y_0_to_fp16"), val = tensor(0x1p-9)]; + tensor _inversed_9439_cast_fp16 = mul(x = write_indices_float_133_cast_fp16, y = _inversed_9439_y_0_to_fp16)[name = tensor("_inversed_9439_cast_fp16")]; + tensor var_9440_cast_fp16 = floor(x = _inversed_9439_cast_fp16)[name = tensor("op_9440_cast_fp16")]; + tensor var_9441_to_fp16 = const()[name = tensor("op_9441_to_fp16"), val = tensor(0x1p+9)]; + tensor var_9442_cast_fp16 = mul(x = var_9440_cast_fp16, y = var_9441_to_fp16)[name = tensor("op_9442_cast_fp16")]; + tensor write_indices_float_135_cast_fp16 = sub(x = write_indices_float_133_cast_fp16, y = var_9442_cast_fp16)[name = tensor("write_indices_float_135_cast_fp16")]; + tensor var_9460_value_0_to_fp16 = const()[name = tensor("op_9460_value_0_to_fp16"), val = tensor(0x1.ffp+8)]; + tensor var_9460_cast_fp16 = fill_like(ref_tensor = write_indices_float_135_cast_fp16, value = var_9460_value_0_to_fp16)[name = tensor("op_9460_cast_fp16")]; + tensor write_indices_float_137_cast_fp16 = select(a = write_indices_float_135_cast_fp16, b = var_9460_cast_fp16, cond = token_valid_1_cast_fp16)[name = tensor("write_indices_float_137_cast_fp16")]; + tensor var_9466_dtype_0 = const()[name = tensor("op_9466_dtype_0"), val = tensor("int32")]; + tensor var_9471 = const()[name = tensor("op_9471"), val = tensor([1, 256, 1, 1])]; + tensor write_indices_float_137_cast_fp16_to_int32 = cast(dtype = var_9466_dtype_0, x = write_indices_float_137_cast_fp16)[name = tensor("cast_550")]; + tensor var_9472 = reshape(shape = var_9471, x = write_indices_float_137_cast_fp16_to_int32)[name = tensor("op_9472")]; + tensor write_indices_45_reps_0 = const()[name = tensor("write_indices_45_reps_0"), val = tensor([1, 1, 16, 64])]; + tensor write_indices_45 = tile(reps = write_indices_45_reps_0, x = var_9472)[name = tensor("write_indices_45")]; + tensor var_9482_begin_0 = const()[name = tensor("op_9482_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_9482_end_0 = const()[name = tensor("op_9482_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor var_9482_end_mask_0 = const()[name = tensor("op_9482_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_9482_squeeze_mask_0 = const()[name = tensor("op_9482_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor cache22_to_fp16_dtype_0 = const()[name = tensor("cache22_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor cache22_to_fp16 = cast(dtype = cache22_to_fp16_dtype_0, x = cache22)[name = tensor("cast_549")]; + tensor var_9482_cast_fp16 = slice_by_index(begin = var_9482_begin_0, end = var_9482_end_0, end_mask = var_9482_end_mask_0, squeeze_mask = var_9482_squeeze_mask_0, x = cache22_to_fp16)[name = tensor("op_9482_cast_fp16")]; + tensor var_9484_axis_0 = const()[name = tensor("op_9484_axis_0"), val = tensor(1)]; + tensor var_9484_mode_0 = const()[name = tensor("op_9484_mode_0"), val = tensor("update")]; + tensor var_9484_validate_indices_0 = const()[name = tensor("op_9484_validate_indices_0"), val = tensor(false)]; + tensor var_9484_cast_fp16 = scatter_along_axis(axis = var_9484_axis_0, data = var_9482_cast_fp16, indices = write_indices_45, mode = var_9484_mode_0, updates = k_91_cast_fp16, validate_indices = var_9484_validate_indices_0)[name = tensor("op_9484_cast_fp16")]; + tensor concat_178 = const()[name = tensor("concat_178"), val = tensor([0, 0, 0, 0, 0])]; + tensor concat_179 = const()[name = tensor("concat_179"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_45_internal_tensor_assign_1_stride_0 = const()[name = tensor("new_cache_45_internal_tensor_assign_1_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_45_internal_tensor_assign_1_begin_mask_0 = const()[name = tensor("new_cache_45_internal_tensor_assign_1_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_45_internal_tensor_assign_1_end_mask_0 = const()[name = tensor("new_cache_45_internal_tensor_assign_1_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_45_internal_tensor_assign_1_squeeze_mask_0 = const()[name = tensor("new_cache_45_internal_tensor_assign_1_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_90 = const()[name = tensor("shape_90"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_44 = const()[name = tensor("reduce_prod_44"), val = tensor(1048576)]; + tensor range_1d_44_start_0 = const()[name = tensor("range_1d_44_start_0"), val = tensor(0)]; + tensor range_1d_44_step_0 = const()[name = tensor("range_1d_44_step_0"), val = tensor(1)]; + tensor range_1d_44 = range_1d(end = reduce_prod_44, start = range_1d_44_start_0, step = range_1d_44_step_0)[name = tensor("range_1d_44")]; + tensor reshape_220 = reshape(shape = shape_90, x = range_1d_44)[name = tensor("reshape_220")]; + tensor slice_by_index_44 = slice_by_index(begin = concat_178, begin_mask = new_cache_45_internal_tensor_assign_1_begin_mask_0, end = concat_179, end_mask = new_cache_45_internal_tensor_assign_1_end_mask_0, squeeze_mask = new_cache_45_internal_tensor_assign_1_squeeze_mask_0, stride = new_cache_45_internal_tensor_assign_1_stride_0, x = reshape_220)[name = tensor("slice_by_index_44")]; + tensor reshape_221_shape_0 = const()[name = tensor("reshape_221_shape_0"), val = tensor([-1])]; + tensor reshape_221 = reshape(shape = reshape_221_shape_0, x = slice_by_index_44)[name = tensor("reshape_221")]; + tensor reshape_222_shape_0 = const()[name = tensor("reshape_222_shape_0"), val = tensor([-1])]; + tensor reshape_222_cast_fp16 = reshape(shape = reshape_222_shape_0, x = var_9484_cast_fp16)[name = tensor("reshape_222_cast_fp16")]; + tensor reshape_223_shape_0 = const()[name = tensor("reshape_223_shape_0"), val = tensor([-1])]; + tensor reshape_223_cast_fp16 = reshape(shape = reshape_223_shape_0, x = cache22_to_fp16)[name = tensor("reshape_223_cast_fp16")]; + tensor scatter_44_mode_0 = const()[name = tensor("scatter_44_mode_0"), val = tensor("update")]; + tensor scatter_44_axis_0 = const()[name = tensor("scatter_44_axis_0"), val = tensor(0)]; + tensor scatter_44_validate_indices_0 = const()[name = tensor("scatter_44_validate_indices_0"), val = tensor(false)]; + tensor scatter_44_cast_fp16 = scatter(axis = scatter_44_axis_0, data = reshape_223_cast_fp16, indices = reshape_221, mode = scatter_44_mode_0, updates = reshape_222_cast_fp16, validate_indices = scatter_44_validate_indices_0)[name = tensor("scatter_44_cast_fp16")]; + tensor reshape_224_cast_fp16 = reshape(shape = shape_90, x = scatter_44_cast_fp16)[name = tensor("reshape_224_cast_fp16")]; + tensor var_9492_begin_0 = const()[name = tensor("op_9492_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor var_9492_end_0 = const()[name = tensor("op_9492_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor var_9492_end_mask_0 = const()[name = tensor("op_9492_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_9492_squeeze_mask_0 = const()[name = tensor("op_9492_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor var_9492_cast_fp16 = slice_by_index(begin = var_9492_begin_0, end = var_9492_end_0, end_mask = var_9492_end_mask_0, squeeze_mask = var_9492_squeeze_mask_0, x = reshape_224_cast_fp16)[name = tensor("op_9492_cast_fp16")]; + tensor var_9494_axis_0 = const()[name = tensor("op_9494_axis_0"), val = tensor(1)]; + tensor var_9494_mode_0 = const()[name = tensor("op_9494_mode_0"), val = tensor("update")]; + tensor var_9494_validate_indices_0 = const()[name = tensor("op_9494_validate_indices_0"), val = tensor(false)]; + tensor var_9494_cast_fp16 = scatter_along_axis(axis = var_9494_axis_0, data = var_9492_cast_fp16, indices = write_indices_45, mode = var_9494_mode_0, updates = v_45_cast_fp16, validate_indices = var_9494_validate_indices_0)[name = tensor("op_9494_cast_fp16")]; + tensor concat_180 = const()[name = tensor("concat_180"), val = tensor([1, 0, 0, 0, 0])]; + tensor concat_181 = const()[name = tensor("concat_181"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_45_internal_tensor_assign_2_stride_0 = const()[name = tensor("new_cache_45_internal_tensor_assign_2_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_45_internal_tensor_assign_2_begin_mask_0 = const()[name = tensor("new_cache_45_internal_tensor_assign_2_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_45_internal_tensor_assign_2_end_mask_0 = const()[name = tensor("new_cache_45_internal_tensor_assign_2_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_45_internal_tensor_assign_2_squeeze_mask_0 = const()[name = tensor("new_cache_45_internal_tensor_assign_2_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_91 = const()[name = tensor("shape_91"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_45 = const()[name = tensor("reduce_prod_45"), val = tensor(1048576)]; + tensor range_1d_45_start_0 = const()[name = tensor("range_1d_45_start_0"), val = tensor(0)]; + tensor range_1d_45_step_0 = const()[name = tensor("range_1d_45_step_0"), val = tensor(1)]; + tensor range_1d_45 = range_1d(end = reduce_prod_45, start = range_1d_45_start_0, step = range_1d_45_step_0)[name = tensor("range_1d_45")]; + tensor reshape_225 = reshape(shape = shape_91, x = range_1d_45)[name = tensor("reshape_225")]; + tensor slice_by_index_45 = slice_by_index(begin = concat_180, begin_mask = new_cache_45_internal_tensor_assign_2_begin_mask_0, end = concat_181, end_mask = new_cache_45_internal_tensor_assign_2_end_mask_0, squeeze_mask = new_cache_45_internal_tensor_assign_2_squeeze_mask_0, stride = new_cache_45_internal_tensor_assign_2_stride_0, x = reshape_225)[name = tensor("slice_by_index_45")]; + tensor reshape_226_shape_0 = const()[name = tensor("reshape_226_shape_0"), val = tensor([-1])]; + tensor reshape_226 = reshape(shape = reshape_226_shape_0, x = slice_by_index_45)[name = tensor("reshape_226")]; + tensor reshape_227_shape_0 = const()[name = tensor("reshape_227_shape_0"), val = tensor([-1])]; + tensor reshape_227_cast_fp16 = reshape(shape = reshape_227_shape_0, x = var_9494_cast_fp16)[name = tensor("reshape_227_cast_fp16")]; + tensor reshape_228_shape_0 = const()[name = tensor("reshape_228_shape_0"), val = tensor([-1])]; + tensor reshape_228_cast_fp16 = reshape(shape = reshape_228_shape_0, x = reshape_224_cast_fp16)[name = tensor("reshape_228_cast_fp16")]; + tensor scatter_45_mode_0 = const()[name = tensor("scatter_45_mode_0"), val = tensor("update")]; + tensor scatter_45_axis_0 = const()[name = tensor("scatter_45_axis_0"), val = tensor(0)]; + tensor scatter_45_validate_indices_0 = const()[name = tensor("scatter_45_validate_indices_0"), val = tensor(false)]; + tensor scatter_45_cast_fp16 = scatter(axis = scatter_45_axis_0, data = reshape_228_cast_fp16, indices = reshape_226, mode = scatter_45_mode_0, updates = reshape_227_cast_fp16, validate_indices = scatter_45_validate_indices_0)[name = tensor("scatter_45_cast_fp16")]; + tensor reshape_229_cast_fp16 = reshape(shape = shape_91, x = scatter_45_cast_fp16)[name = tensor("reshape_229_cast_fp16")]; + tensor reshape_229_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("reshape_229_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor keys_133_begin_0 = const()[name = tensor("keys_133_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor keys_133_end_0 = const()[name = tensor("keys_133_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor keys_133_end_mask_0 = const()[name = tensor("keys_133_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor keys_133_squeeze_mask_0 = const()[name = tensor("keys_133_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor keys_133_cast_fp16 = slice_by_index(begin = keys_133_begin_0, end = keys_133_end_0, end_mask = keys_133_end_mask_0, squeeze_mask = keys_133_squeeze_mask_0, x = reshape_229_cast_fp16)[name = tensor("keys_133_cast_fp16")]; + tensor values_133_begin_0 = const()[name = tensor("values_133_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor values_133_end_0 = const()[name = tensor("values_133_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor values_133_end_mask_0 = const()[name = tensor("values_133_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor values_133_squeeze_mask_0 = const()[name = tensor("values_133_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor values_133_cast_fp16 = slice_by_index(begin = values_133_begin_0, end = values_133_end_0, end_mask = values_133_end_mask_0, squeeze_mask = values_133_squeeze_mask_0, x = reshape_229_cast_fp16)[name = tensor("values_133_cast_fp16")]; + tensor var_9506_cast_fp16 = not_equal(x = keys_133_cast_fp16, y = keys_133_cast_fp16)[name = tensor("op_9506_cast_fp16")]; + tensor keys_135_cast_fp16 = select(a = var_470_to_fp16, b = keys_133_cast_fp16, cond = var_9506_cast_fp16)[name = tensor("keys_135_cast_fp16")]; + tensor var_9514_cast_fp16 = not_equal(x = values_133_cast_fp16, y = values_133_cast_fp16)[name = tensor("op_9514_cast_fp16")]; + tensor values_135_cast_fp16 = select(a = var_470_to_fp16, b = values_133_cast_fp16, cond = var_9514_cast_fp16)[name = tensor("values_135_cast_fp16")]; + tensor var_9538 = const()[name = tensor("op_9538"), val = tensor([0, 2, 1, 3])]; + tensor var_9554 = const()[name = tensor("op_9554"), val = tensor([1, 1, 1])]; + tensor var_9555_cast_fp16 = reshape(shape = var_9554, x = position22_to_fp16)[name = tensor("op_9555_cast_fp16")]; + tensor q_positions_45_cast_fp16 = add(x = var_9555_cast_fp16, y = q_offsets_1_promoted_to_fp16)[name = tensor("q_positions_45_cast_fp16")]; + tensor valid_end_45_cast_fp16 = add(x = var_9555_cast_fp16, y = var_536_cast_fp16)[name = tensor("valid_end_45_cast_fp16")]; + tensor valid_mask_45_cast_fp16 = less(x = k_positions_1_promoted_to_fp16, y = valid_end_45_cast_fp16)[name = tensor("valid_mask_45_cast_fp16")]; + tensor causal_mask_45_cast_fp16 = less_equal(x = k_positions_1_promoted_to_fp16, y = q_positions_45_cast_fp16)[name = tensor("causal_mask_45_cast_fp16")]; + tensor var_9583 = logical_and(x = valid_mask_45_cast_fp16, y = causal_mask_45_cast_fp16)[name = tensor("op_9583")]; + tensor attn_mask_45_axes_0 = const()[name = tensor("attn_mask_45_axes_0"), val = tensor([1])]; + tensor attn_mask_45 = expand_dims(axes = attn_mask_45_axes_0, x = var_9583)[name = tensor("attn_mask_45")]; + tensor var_9597_transpose_x_0 = const()[name = tensor("op_9597_transpose_x_0"), val = tensor(false)]; + tensor var_9597_transpose_y_0 = const()[name = tensor("op_9597_transpose_y_0"), val = tensor(false)]; + tensor transpose_113_perm_0 = const()[name = tensor("transpose_113_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_114_perm_0 = const()[name = tensor("transpose_114_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_114 = transpose(perm = transpose_114_perm_0, x = keys_135_cast_fp16)[name = tensor("transpose_116")]; + tensor transpose_113 = transpose(perm = transpose_113_perm_0, x = q_135_cast_fp16)[name = tensor("transpose_117")]; + tensor var_9597_cast_fp16 = matmul(transpose_x = var_9597_transpose_x_0, transpose_y = var_9597_transpose_y_0, x = transpose_113, y = transpose_114)[name = tensor("op_9597_cast_fp16")]; + tensor var_9591_to_fp16 = const()[name = tensor("op_9591_to_fp16"), val = tensor([0x1p-3])]; + tensor attn_weights_133_cast_fp16 = mul(x = var_9597_cast_fp16, y = var_9591_to_fp16)[name = tensor("attn_weights_133_cast_fp16")]; + tensor var_9599 = logical_not(x = attn_mask_45)[name = tensor("op_9599")]; + tensor var_9600_to_fp16 = const()[name = tensor("op_9600_to_fp16"), val = tensor(-inf)]; + tensor attn_weights_135_cast_fp16 = select(a = var_9600_to_fp16, b = attn_weights_133_cast_fp16, cond = var_9599)[name = tensor("attn_weights_135_cast_fp16")]; + tensor var_9602 = const()[name = tensor("op_9602"), val = tensor(-1)]; + tensor attn_weights_137_cast_fp16 = softmax(axis = var_9602, x = attn_weights_135_cast_fp16)[name = tensor("attn_weights_137_cast_fp16")]; + tensor attn_output_45_transpose_x_0 = const()[name = tensor("attn_output_45_transpose_x_0"), val = tensor(false)]; + tensor attn_output_45_transpose_y_0 = const()[name = tensor("attn_output_45_transpose_y_0"), val = tensor(false)]; + tensor values_137_cast_fp16 = transpose(perm = var_9538, x = values_135_cast_fp16)[name = tensor("transpose_118")]; + tensor attn_output_45_cast_fp16 = matmul(transpose_x = attn_output_45_transpose_x_0, transpose_y = attn_output_45_transpose_y_0, x = attn_weights_137_cast_fp16, y = values_137_cast_fp16)[name = tensor("attn_output_45_cast_fp16")]; + tensor var_9610 = const()[name = tensor("op_9610"), val = tensor([0, 2, 1, 3])]; + tensor var_9615 = const()[name = tensor("op_9615"), val = tensor([1, 256, 1024])]; + tensor var_9611_cast_fp16 = transpose(perm = var_9610, x = attn_output_45_cast_fp16)[name = tensor("transpose_115")]; + tensor input_269_cast_fp16 = reshape(shape = var_9615, x = var_9611_cast_fp16)[name = tensor("input_269_cast_fp16")]; + tensor attn22_out_proj_weight_to_fp16 = const()[name = tensor("attn22_out_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(561206400)))]; + tensor linear_89_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = attn22_out_proj_weight_to_fp16, x = input_269_cast_fp16)[name = tensor("linear_89_cast_fp16")]; + tensor var_9624_cast_fp16 = add(x = position22_to_fp16, y = valid_len_to_fp16)[name = tensor("op_9624_cast_fp16")]; + tensor var_9624_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_9624_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor input_271_cast_fp16 = add(x = input_265_cast_fp16, y = linear_89_cast_fp16)[name = tensor("input_271_cast_fp16")]; + tensor input_273_axes_0 = const()[name = tensor("input_273_axes_0"), val = tensor([-1])]; + tensor norm22_2_weight_to_fp16 = const()[name = tensor("norm22_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(563303616)))]; + tensor norm22_2_bias_to_fp16 = const()[name = tensor("norm22_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(563305728)))]; + tensor var_9628_to_fp16 = const()[name = tensor("op_9628_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_273_cast_fp16 = layer_norm(axes = input_273_axes_0, beta = norm22_2_bias_to_fp16, epsilon = var_9628_to_fp16, gamma = norm22_2_weight_to_fp16, x = input_271_cast_fp16)[name = tensor("input_273_cast_fp16")]; + tensor linear22_1_weight_to_fp16 = const()[name = tensor("linear22_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(563307840)))]; + tensor linear_90_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = linear22_1_weight_to_fp16, x = input_273_cast_fp16)[name = tensor("linear_90_cast_fp16")]; + tensor input_275_mode_0 = const()[name = tensor("input_275_mode_0"), val = tensor("EXACT")]; + tensor input_275_cast_fp16 = gelu(mode = input_275_mode_0, x = linear_90_cast_fp16)[name = tensor("input_275_cast_fp16")]; + tensor linear22_2_weight_to_fp16 = const()[name = tensor("linear22_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(571696512)))]; + tensor linear_91_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = linear22_2_weight_to_fp16, x = input_275_cast_fp16)[name = tensor("linear_91_cast_fp16")]; + tensor input_277_cast_fp16 = add(x = input_271_cast_fp16, y = linear_91_cast_fp16)[name = tensor("input_277_cast_fp16")]; + tensor input_279_axes_0 = const()[name = tensor("input_279_axes_0"), val = tensor([-1])]; + tensor norm23_1_weight_to_fp16 = const()[name = tensor("norm23_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(580085184)))]; + tensor norm23_1_bias_to_fp16 = const()[name = tensor("norm23_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(580087296)))]; + tensor var_9645_to_fp16 = const()[name = tensor("op_9645_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_279_cast_fp16 = layer_norm(axes = input_279_axes_0, beta = norm23_1_bias_to_fp16, epsilon = var_9645_to_fp16, gamma = norm23_1_weight_to_fp16, x = input_277_cast_fp16)[name = tensor("input_279_cast_fp16")]; + tensor attn23_in_proj_weight_to_fp16 = const()[name = tensor("attn23_in_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(580089408)))]; + tensor linear_92_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = attn23_in_proj_weight_to_fp16, x = input_279_cast_fp16)[name = tensor("linear_92_cast_fp16")]; + tensor var_9659 = const()[name = tensor("op_9659"), val = tensor([1, 256, 3, 16, 64])]; + tensor qkv_cast_fp16 = reshape(shape = var_9659, x = linear_92_cast_fp16)[name = tensor("qkv_cast_fp16")]; + tensor k_93_begin_0 = const()[name = tensor("k_93_begin_0"), val = tensor([0, 0, 1, 0, 0])]; + tensor k_93_end_0 = const()[name = tensor("k_93_end_0"), val = tensor([1, 256, 2, 16, 64])]; + tensor k_93_end_mask_0 = const()[name = tensor("k_93_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor k_93_squeeze_mask_0 = const()[name = tensor("k_93_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor k_93_cast_fp16 = slice_by_index(begin = k_93_begin_0, end = k_93_end_0, end_mask = k_93_end_mask_0, squeeze_mask = k_93_squeeze_mask_0, x = qkv_cast_fp16)[name = tensor("k_93_cast_fp16")]; + tensor v_begin_0 = const()[name = tensor("v_begin_0"), val = tensor([0, 0, 2, 0, 0])]; + tensor v_end_0 = const()[name = tensor("v_end_0"), val = tensor([1, 256, 3, 16, 64])]; + tensor v_end_mask_0 = const()[name = tensor("v_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor v_squeeze_mask_0 = const()[name = tensor("v_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor v_cast_fp16 = slice_by_index(begin = v_begin_0, end = v_end_0, end_mask = v_end_mask_0, squeeze_mask = v_squeeze_mask_0, x = qkv_cast_fp16)[name = tensor("v_cast_fp16")]; + tensor var_9704 = const()[name = tensor("op_9704"), val = tensor([1, 1])]; + tensor position23_to_fp16_dtype_0 = const()[name = tensor("position23_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor position23_to_fp16 = cast(dtype = position23_to_fp16_dtype_0, x = position23)[name = tensor("cast_546")]; + tensor var_9705_cast_fp16 = reshape(shape = var_9704, x = position23_to_fp16)[name = tensor("op_9705_cast_fp16")]; + tensor ts_141_cast_fp16 = add(x = ts_1_promoted_to_fp16, y = var_9705_cast_fp16)[name = tensor("ts_141_cast_fp16")]; + tensor var_9712 = const()[name = tensor("op_9712"), val = tensor([1, 256, 1, 1])]; + tensor ts_cast_fp16 = reshape(shape = var_9712, x = ts_141_cast_fp16)[name = tensor("ts_cast_fp16")]; + tensor var_9719 = const()[name = tensor("op_9719"), val = tensor([1, 256, 16, 32, 2])]; + tensor k_complex_cast_fp16 = reshape(shape = var_9719, x = k_93_cast_fp16)[name = tensor("k_complex_cast_fp16")]; + tensor var_9723_begin_0 = const()[name = tensor("op_9723_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_9723_end_0 = const()[name = tensor("op_9723_end_0"), val = tensor([1, 256, 16, 32, 1])]; + tensor var_9723_end_mask_0 = const()[name = tensor("op_9723_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_9723_squeeze_mask_0 = const()[name = tensor("op_9723_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_9723_cast_fp16 = slice_by_index(begin = var_9723_begin_0, end = var_9723_end_0, end_mask = var_9723_end_mask_0, squeeze_mask = var_9723_squeeze_mask_0, x = k_complex_cast_fp16)[name = tensor("op_9723_cast_fp16")]; + tensor var_9731_begin_0 = const()[name = tensor("op_9731_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_9731_end_0 = const()[name = tensor("op_9731_end_0"), val = tensor([1, 256, 16, 32, 2])]; + tensor var_9731_end_mask_0 = const()[name = tensor("op_9731_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_9731_squeeze_mask_0 = const()[name = tensor("op_9731_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_9731_cast_fp16 = slice_by_index(begin = var_9731_begin_0, end = var_9731_end_0, end_mask = var_9731_end_mask_0, squeeze_mask = var_9731_squeeze_mask_0, x = k_complex_cast_fp16)[name = tensor("op_9731_cast_fp16")]; + tensor freqs_to_fp16 = const()[name = tensor("freqs_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(586380928)))]; + tensor var_9737_cast_fp16 = mul(x = freqs_to_fp16, y = ts_cast_fp16)[name = tensor("op_9737_cast_fp16")]; + tensor rotr_cast_fp16 = cos(x = var_9737_cast_fp16)[name = tensor("rotr_cast_fp16")]; + tensor roti_cast_fp16 = sin(x = var_9737_cast_fp16)[name = tensor("roti_cast_fp16")]; + tensor var_9741_cast_fp16 = mul(x = var_9723_cast_fp16, y = rotr_cast_fp16)[name = tensor("op_9741_cast_fp16")]; + tensor var_9742_cast_fp16 = mul(x = var_9731_cast_fp16, y = roti_cast_fp16)[name = tensor("op_9742_cast_fp16")]; + tensor kor_93_cast_fp16 = sub(x = var_9741_cast_fp16, y = var_9742_cast_fp16)[name = tensor("kor_93_cast_fp16")]; + tensor var_9745_cast_fp16 = mul(x = var_9723_cast_fp16, y = roti_cast_fp16)[name = tensor("op_9745_cast_fp16")]; + tensor var_9746_cast_fp16 = mul(x = var_9731_cast_fp16, y = rotr_cast_fp16)[name = tensor("op_9746_cast_fp16")]; + tensor koi_93_cast_fp16 = add(x = var_9745_cast_fp16, y = var_9746_cast_fp16)[name = tensor("koi_93_cast_fp16")]; + tensor ko_axis_0 = const()[name = tensor("ko_axis_0"), val = tensor(-1)]; + tensor ko_cast_fp16 = stack(axis = ko_axis_0, values = (kor_93_cast_fp16, koi_93_cast_fp16))[name = tensor("ko_cast_fp16")]; + tensor var_9766 = const()[name = tensor("op_9766"), val = tensor([1, 256, 16, 64])]; + tensor k_cast_fp16 = reshape(shape = var_9766, x = ko_cast_fp16)[name = tensor("k_cast_fp16")]; + tensor write_indices_float_139_cast_fp16 = add(x = var_9705_cast_fp16, y = var_393_promoted_to_fp16)[name = tensor("write_indices_float_139_cast_fp16")]; + tensor _inversed_9787_y_0_to_fp16 = const()[name = tensor("_inversed_9787_y_0_to_fp16"), val = tensor(0x1p-9)]; + tensor _inversed_9787_cast_fp16 = mul(x = write_indices_float_139_cast_fp16, y = _inversed_9787_y_0_to_fp16)[name = tensor("_inversed_9787_cast_fp16")]; + tensor var_9788_cast_fp16 = floor(x = _inversed_9787_cast_fp16)[name = tensor("op_9788_cast_fp16")]; + tensor var_9789_to_fp16 = const()[name = tensor("op_9789_to_fp16"), val = tensor(0x1p+9)]; + tensor var_9790_cast_fp16 = mul(x = var_9788_cast_fp16, y = var_9789_to_fp16)[name = tensor("op_9790_cast_fp16")]; + tensor write_indices_float_141_cast_fp16 = sub(x = write_indices_float_139_cast_fp16, y = var_9790_cast_fp16)[name = tensor("write_indices_float_141_cast_fp16")]; + tensor var_9808_value_0_to_fp16 = const()[name = tensor("op_9808_value_0_to_fp16"), val = tensor(0x1.ffp+8)]; + tensor var_9808_cast_fp16 = fill_like(ref_tensor = write_indices_float_141_cast_fp16, value = var_9808_value_0_to_fp16)[name = tensor("op_9808_cast_fp16")]; + tensor write_indices_float_cast_fp16 = select(a = write_indices_float_141_cast_fp16, b = var_9808_cast_fp16, cond = token_valid_1_cast_fp16)[name = tensor("write_indices_float_cast_fp16")]; + tensor var_9814_dtype_0 = const()[name = tensor("op_9814_dtype_0"), val = tensor("int32")]; + tensor var_9819 = const()[name = tensor("op_9819"), val = tensor([1, 256, 1, 1])]; + tensor write_indices_float_cast_fp16_to_int32 = cast(dtype = var_9814_dtype_0, x = write_indices_float_cast_fp16)[name = tensor("cast_545")]; + tensor var_9820 = reshape(shape = var_9819, x = write_indices_float_cast_fp16_to_int32)[name = tensor("op_9820")]; + tensor write_indices_reps_0 = const()[name = tensor("write_indices_reps_0"), val = tensor([1, 1, 16, 64])]; + tensor write_indices = tile(reps = write_indices_reps_0, x = var_9820)[name = tensor("write_indices")]; + tensor var_9830_begin_0 = const()[name = tensor("op_9830_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_9830_end_0 = const()[name = tensor("op_9830_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor var_9830_end_mask_0 = const()[name = tensor("op_9830_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_9830_squeeze_mask_0 = const()[name = tensor("op_9830_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor cache23_to_fp16_dtype_0 = const()[name = tensor("cache23_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor cache23_to_fp16 = cast(dtype = cache23_to_fp16_dtype_0, x = cache23)[name = tensor("cast_544")]; + tensor var_9830_cast_fp16 = slice_by_index(begin = var_9830_begin_0, end = var_9830_end_0, end_mask = var_9830_end_mask_0, squeeze_mask = var_9830_squeeze_mask_0, x = cache23_to_fp16)[name = tensor("op_9830_cast_fp16")]; + tensor var_9832_axis_0 = const()[name = tensor("op_9832_axis_0"), val = tensor(1)]; + tensor var_9832_mode_0 = const()[name = tensor("op_9832_mode_0"), val = tensor("update")]; + tensor var_9832_validate_indices_0 = const()[name = tensor("op_9832_validate_indices_0"), val = tensor(false)]; + tensor var_9832_cast_fp16 = scatter_along_axis(axis = var_9832_axis_0, data = var_9830_cast_fp16, indices = write_indices, mode = var_9832_mode_0, updates = k_cast_fp16, validate_indices = var_9832_validate_indices_0)[name = tensor("op_9832_cast_fp16")]; + tensor concat_186 = const()[name = tensor("concat_186"), val = tensor([0, 0, 0, 0, 0])]; + tensor concat_187 = const()[name = tensor("concat_187"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_internal_tensor_assign_1_stride_0 = const()[name = tensor("new_cache_internal_tensor_assign_1_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_internal_tensor_assign_1_begin_mask_0 = const()[name = tensor("new_cache_internal_tensor_assign_1_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_internal_tensor_assign_1_end_mask_0 = const()[name = tensor("new_cache_internal_tensor_assign_1_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_internal_tensor_assign_1_squeeze_mask_0 = const()[name = tensor("new_cache_internal_tensor_assign_1_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_92 = const()[name = tensor("shape_92"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_46 = const()[name = tensor("reduce_prod_46"), val = tensor(1048576)]; + tensor range_1d_46_start_0 = const()[name = tensor("range_1d_46_start_0"), val = tensor(0)]; + tensor range_1d_46_step_0 = const()[name = tensor("range_1d_46_step_0"), val = tensor(1)]; + tensor range_1d_46 = range_1d(end = reduce_prod_46, start = range_1d_46_start_0, step = range_1d_46_step_0)[name = tensor("range_1d_46")]; + tensor reshape_230 = reshape(shape = shape_92, x = range_1d_46)[name = tensor("reshape_230")]; + tensor slice_by_index_46 = slice_by_index(begin = concat_186, begin_mask = new_cache_internal_tensor_assign_1_begin_mask_0, end = concat_187, end_mask = new_cache_internal_tensor_assign_1_end_mask_0, squeeze_mask = new_cache_internal_tensor_assign_1_squeeze_mask_0, stride = new_cache_internal_tensor_assign_1_stride_0, x = reshape_230)[name = tensor("slice_by_index_46")]; + tensor reshape_231_shape_0 = const()[name = tensor("reshape_231_shape_0"), val = tensor([-1])]; + tensor reshape_231 = reshape(shape = reshape_231_shape_0, x = slice_by_index_46)[name = tensor("reshape_231")]; + tensor reshape_232_shape_0 = const()[name = tensor("reshape_232_shape_0"), val = tensor([-1])]; + tensor reshape_232_cast_fp16 = reshape(shape = reshape_232_shape_0, x = var_9832_cast_fp16)[name = tensor("reshape_232_cast_fp16")]; + tensor reshape_233_shape_0 = const()[name = tensor("reshape_233_shape_0"), val = tensor([-1])]; + tensor reshape_233_cast_fp16 = reshape(shape = reshape_233_shape_0, x = cache23_to_fp16)[name = tensor("reshape_233_cast_fp16")]; + tensor scatter_46_mode_0 = const()[name = tensor("scatter_46_mode_0"), val = tensor("update")]; + tensor scatter_46_axis_0 = const()[name = tensor("scatter_46_axis_0"), val = tensor(0)]; + tensor scatter_46_validate_indices_0 = const()[name = tensor("scatter_46_validate_indices_0"), val = tensor(false)]; + tensor scatter_46_cast_fp16 = scatter(axis = scatter_46_axis_0, data = reshape_233_cast_fp16, indices = reshape_231, mode = scatter_46_mode_0, updates = reshape_232_cast_fp16, validate_indices = scatter_46_validate_indices_0)[name = tensor("scatter_46_cast_fp16")]; + tensor reshape_234_cast_fp16 = reshape(shape = shape_92, x = scatter_46_cast_fp16)[name = tensor("reshape_234_cast_fp16")]; + tensor var_9840_begin_0 = const()[name = tensor("op_9840_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor var_9840_end_0 = const()[name = tensor("op_9840_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor var_9840_end_mask_0 = const()[name = tensor("op_9840_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_9840_squeeze_mask_0 = const()[name = tensor("op_9840_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor var_9840_cast_fp16 = slice_by_index(begin = var_9840_begin_0, end = var_9840_end_0, end_mask = var_9840_end_mask_0, squeeze_mask = var_9840_squeeze_mask_0, x = reshape_234_cast_fp16)[name = tensor("op_9840_cast_fp16")]; + tensor var_9842_axis_0 = const()[name = tensor("op_9842_axis_0"), val = tensor(1)]; + tensor var_9842_mode_0 = const()[name = tensor("op_9842_mode_0"), val = tensor("update")]; + tensor var_9842_validate_indices_0 = const()[name = tensor("op_9842_validate_indices_0"), val = tensor(false)]; + tensor var_9842_cast_fp16 = scatter_along_axis(axis = var_9842_axis_0, data = var_9840_cast_fp16, indices = write_indices, mode = var_9842_mode_0, updates = v_cast_fp16, validate_indices = var_9842_validate_indices_0)[name = tensor("op_9842_cast_fp16")]; + tensor concat_188 = const()[name = tensor("concat_188"), val = tensor([1, 0, 0, 0, 0])]; + tensor concat_189 = const()[name = tensor("concat_189"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_internal_tensor_assign_2_stride_0 = const()[name = tensor("new_cache_internal_tensor_assign_2_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_internal_tensor_assign_2_begin_mask_0 = const()[name = tensor("new_cache_internal_tensor_assign_2_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_internal_tensor_assign_2_end_mask_0 = const()[name = tensor("new_cache_internal_tensor_assign_2_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_internal_tensor_assign_2_squeeze_mask_0 = const()[name = tensor("new_cache_internal_tensor_assign_2_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_93 = const()[name = tensor("shape_93"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_47 = const()[name = tensor("reduce_prod_47"), val = tensor(1048576)]; + tensor range_1d_47_start_0 = const()[name = tensor("range_1d_47_start_0"), val = tensor(0)]; + tensor range_1d_47_step_0 = const()[name = tensor("range_1d_47_step_0"), val = tensor(1)]; + tensor range_1d_47 = range_1d(end = reduce_prod_47, start = range_1d_47_start_0, step = range_1d_47_step_0)[name = tensor("range_1d_47")]; + tensor reshape_235 = reshape(shape = shape_93, x = range_1d_47)[name = tensor("reshape_235")]; + tensor slice_by_index_47 = slice_by_index(begin = concat_188, begin_mask = new_cache_internal_tensor_assign_2_begin_mask_0, end = concat_189, end_mask = new_cache_internal_tensor_assign_2_end_mask_0, squeeze_mask = new_cache_internal_tensor_assign_2_squeeze_mask_0, stride = new_cache_internal_tensor_assign_2_stride_0, x = reshape_235)[name = tensor("slice_by_index_47")]; + tensor reshape_236_shape_0 = const()[name = tensor("reshape_236_shape_0"), val = tensor([-1])]; + tensor reshape_236 = reshape(shape = reshape_236_shape_0, x = slice_by_index_47)[name = tensor("reshape_236")]; + tensor reshape_237_shape_0 = const()[name = tensor("reshape_237_shape_0"), val = tensor([-1])]; + tensor reshape_237_cast_fp16 = reshape(shape = reshape_237_shape_0, x = var_9842_cast_fp16)[name = tensor("reshape_237_cast_fp16")]; + tensor reshape_238_shape_0 = const()[name = tensor("reshape_238_shape_0"), val = tensor([-1])]; + tensor reshape_238_cast_fp16 = reshape(shape = reshape_238_shape_0, x = reshape_234_cast_fp16)[name = tensor("reshape_238_cast_fp16")]; + tensor scatter_47_mode_0 = const()[name = tensor("scatter_47_mode_0"), val = tensor("update")]; + tensor scatter_47_axis_0 = const()[name = tensor("scatter_47_axis_0"), val = tensor(0)]; + tensor scatter_47_validate_indices_0 = const()[name = tensor("scatter_47_validate_indices_0"), val = tensor(false)]; + tensor scatter_47_cast_fp16 = scatter(axis = scatter_47_axis_0, data = reshape_238_cast_fp16, indices = reshape_236, mode = scatter_47_mode_0, updates = reshape_237_cast_fp16, validate_indices = scatter_47_validate_indices_0)[name = tensor("scatter_47_cast_fp16")]; + tensor reshape_239_cast_fp16 = reshape(shape = shape_93, x = scatter_47_cast_fp16)[name = tensor("reshape_239_cast_fp16")]; + tensor reshape_239_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("reshape_239_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor var_9852_cast_fp16 = add(x = position23_to_fp16, y = valid_len_to_fp16)[name = tensor("op_9852_cast_fp16")]; + tensor var_9852_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_9852_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor var_9852 = cast(dtype = var_9852_cast_fp16_to_fp32_dtype_0, x = var_9852_cast_fp16)[name = tensor("cast_542")]; + tensor new_cache_internal_tensor_assign_2 = cast(dtype = reshape_239_cast_fp16_to_fp32_dtype_0, x = reshape_239_cast_fp16)[name = tensor("cast_543")]; + tensor var_9624 = cast(dtype = var_9624_cast_fp16_to_fp32_dtype_0, x = var_9624_cast_fp16)[name = tensor("cast_547")]; + tensor new_cache_45_internal_tensor_assign_2 = cast(dtype = reshape_229_cast_fp16_to_fp32_dtype_0, x = reshape_229_cast_fp16)[name = tensor("cast_548")]; + tensor var_9213 = cast(dtype = var_9213_cast_fp16_to_fp32_dtype_0, x = var_9213_cast_fp16)[name = tensor("cast_552")]; + tensor new_cache_43_internal_tensor_assign_2 = cast(dtype = reshape_219_cast_fp16_to_fp32_dtype_0, x = reshape_219_cast_fp16)[name = tensor("cast_553")]; + tensor var_8802 = cast(dtype = var_8802_cast_fp16_to_fp32_dtype_0, x = var_8802_cast_fp16)[name = tensor("cast_557")]; + tensor new_cache_41_internal_tensor_assign_2 = cast(dtype = reshape_209_cast_fp16_to_fp32_dtype_0, x = reshape_209_cast_fp16)[name = tensor("cast_558")]; + tensor var_8391 = cast(dtype = var_8391_cast_fp16_to_fp32_dtype_0, x = var_8391_cast_fp16)[name = tensor("cast_562")]; + tensor new_cache_39_internal_tensor_assign_2 = cast(dtype = reshape_199_cast_fp16_to_fp32_dtype_0, x = reshape_199_cast_fp16)[name = tensor("cast_563")]; + tensor var_7980 = cast(dtype = var_7980_cast_fp16_to_fp32_dtype_0, x = var_7980_cast_fp16)[name = tensor("cast_567")]; + tensor new_cache_37_internal_tensor_assign_2 = cast(dtype = reshape_189_cast_fp16_to_fp32_dtype_0, x = reshape_189_cast_fp16)[name = tensor("cast_568")]; + tensor var_7569 = cast(dtype = var_7569_cast_fp16_to_fp32_dtype_0, x = var_7569_cast_fp16)[name = tensor("cast_572")]; + tensor new_cache_35_internal_tensor_assign_2 = cast(dtype = reshape_179_cast_fp16_to_fp32_dtype_0, x = reshape_179_cast_fp16)[name = tensor("cast_573")]; + tensor var_7158 = cast(dtype = var_7158_cast_fp16_to_fp32_dtype_0, x = var_7158_cast_fp16)[name = tensor("cast_577")]; + tensor new_cache_33_internal_tensor_assign_2 = cast(dtype = reshape_169_cast_fp16_to_fp32_dtype_0, x = reshape_169_cast_fp16)[name = tensor("cast_578")]; + tensor var_6747 = cast(dtype = var_6747_cast_fp16_to_fp32_dtype_0, x = var_6747_cast_fp16)[name = tensor("cast_582")]; + tensor new_cache_31_internal_tensor_assign_2 = cast(dtype = reshape_159_cast_fp16_to_fp32_dtype_0, x = reshape_159_cast_fp16)[name = tensor("cast_583")]; + tensor var_6336 = cast(dtype = var_6336_cast_fp16_to_fp32_dtype_0, x = var_6336_cast_fp16)[name = tensor("cast_587")]; + tensor new_cache_29_internal_tensor_assign_2 = cast(dtype = reshape_149_cast_fp16_to_fp32_dtype_0, x = reshape_149_cast_fp16)[name = tensor("cast_588")]; + tensor var_5925 = cast(dtype = var_5925_cast_fp16_to_fp32_dtype_0, x = var_5925_cast_fp16)[name = tensor("cast_592")]; + tensor new_cache_27_internal_tensor_assign_2 = cast(dtype = reshape_139_cast_fp16_to_fp32_dtype_0, x = reshape_139_cast_fp16)[name = tensor("cast_593")]; + tensor var_5514 = cast(dtype = var_5514_cast_fp16_to_fp32_dtype_0, x = var_5514_cast_fp16)[name = tensor("cast_597")]; + tensor new_cache_25_internal_tensor_assign_2 = cast(dtype = reshape_129_cast_fp16_to_fp32_dtype_0, x = reshape_129_cast_fp16)[name = tensor("cast_598")]; + tensor var_5103 = cast(dtype = var_5103_cast_fp16_to_fp32_dtype_0, x = var_5103_cast_fp16)[name = tensor("cast_602")]; + tensor new_cache_23_internal_tensor_assign_2 = cast(dtype = reshape_119_cast_fp16_to_fp32_dtype_0, x = reshape_119_cast_fp16)[name = tensor("cast_603")]; + tensor var_4692 = cast(dtype = var_4692_cast_fp16_to_fp32_dtype_0, x = var_4692_cast_fp16)[name = tensor("cast_607")]; + tensor new_cache_21_internal_tensor_assign_2 = cast(dtype = reshape_109_cast_fp16_to_fp32_dtype_0, x = reshape_109_cast_fp16)[name = tensor("cast_608")]; + tensor var_4281 = cast(dtype = var_4281_cast_fp16_to_fp32_dtype_0, x = var_4281_cast_fp16)[name = tensor("cast_612")]; + tensor new_cache_19_internal_tensor_assign_2 = cast(dtype = reshape_99_cast_fp16_to_fp32_dtype_0, x = reshape_99_cast_fp16)[name = tensor("cast_613")]; + tensor var_3870 = cast(dtype = var_3870_cast_fp16_to_fp32_dtype_0, x = var_3870_cast_fp16)[name = tensor("cast_617")]; + tensor new_cache_17_internal_tensor_assign_2 = cast(dtype = reshape_89_cast_fp16_to_fp32_dtype_0, x = reshape_89_cast_fp16)[name = tensor("cast_618")]; + tensor var_3459 = cast(dtype = var_3459_cast_fp16_to_fp32_dtype_0, x = var_3459_cast_fp16)[name = tensor("cast_622")]; + tensor new_cache_15_internal_tensor_assign_2 = cast(dtype = reshape_79_cast_fp16_to_fp32_dtype_0, x = reshape_79_cast_fp16)[name = tensor("cast_623")]; + tensor var_3048 = cast(dtype = var_3048_cast_fp16_to_fp32_dtype_0, x = var_3048_cast_fp16)[name = tensor("cast_627")]; + tensor new_cache_13_internal_tensor_assign_2 = cast(dtype = reshape_69_cast_fp16_to_fp32_dtype_0, x = reshape_69_cast_fp16)[name = tensor("cast_628")]; + tensor var_2637 = cast(dtype = var_2637_cast_fp16_to_fp32_dtype_0, x = var_2637_cast_fp16)[name = tensor("cast_632")]; + tensor new_cache_11_internal_tensor_assign_2 = cast(dtype = reshape_59_cast_fp16_to_fp32_dtype_0, x = reshape_59_cast_fp16)[name = tensor("cast_633")]; + tensor var_2226 = cast(dtype = var_2226_cast_fp16_to_fp32_dtype_0, x = var_2226_cast_fp16)[name = tensor("cast_637")]; + tensor new_cache_9_internal_tensor_assign_2 = cast(dtype = reshape_49_cast_fp16_to_fp32_dtype_0, x = reshape_49_cast_fp16)[name = tensor("cast_638")]; + tensor var_1815 = cast(dtype = var_1815_cast_fp16_to_fp32_dtype_0, x = var_1815_cast_fp16)[name = tensor("cast_642")]; + tensor new_cache_7_internal_tensor_assign_2 = cast(dtype = reshape_39_cast_fp16_to_fp32_dtype_0, x = reshape_39_cast_fp16)[name = tensor("cast_643")]; + tensor var_1404 = cast(dtype = var_1404_cast_fp16_to_fp32_dtype_0, x = var_1404_cast_fp16)[name = tensor("cast_647")]; + tensor new_cache_5_internal_tensor_assign_2 = cast(dtype = reshape_29_cast_fp16_to_fp32_dtype_0, x = reshape_29_cast_fp16)[name = tensor("cast_648")]; + tensor var_993 = cast(dtype = var_993_cast_fp16_to_fp32_dtype_0, x = var_993_cast_fp16)[name = tensor("cast_652")]; + tensor new_cache_3_internal_tensor_assign_2 = cast(dtype = reshape_19_cast_fp16_to_fp32_dtype_0, x = reshape_19_cast_fp16)[name = tensor("cast_653")]; + tensor var_582 = cast(dtype = var_582_cast_fp16_to_fp32_dtype_0, x = var_582_cast_fp16)[name = tensor("cast_657")]; + tensor new_cache_1_internal_tensor_assign_2 = cast(dtype = reshape_9_cast_fp16_to_fp32_dtype_0, x = reshape_9_cast_fp16)[name = tensor("cast_658")]; + } -> (new_cache_1_internal_tensor_assign_2, var_582, new_cache_3_internal_tensor_assign_2, var_993, new_cache_5_internal_tensor_assign_2, var_1404, new_cache_7_internal_tensor_assign_2, var_1815, new_cache_9_internal_tensor_assign_2, var_2226, new_cache_11_internal_tensor_assign_2, var_2637, new_cache_13_internal_tensor_assign_2, var_3048, new_cache_15_internal_tensor_assign_2, var_3459, new_cache_17_internal_tensor_assign_2, var_3870, new_cache_19_internal_tensor_assign_2, var_4281, new_cache_21_internal_tensor_assign_2, var_4692, new_cache_23_internal_tensor_assign_2, var_5103, new_cache_25_internal_tensor_assign_2, var_5514, new_cache_27_internal_tensor_assign_2, var_5925, new_cache_29_internal_tensor_assign_2, var_6336, new_cache_31_internal_tensor_assign_2, var_6747, new_cache_33_internal_tensor_assign_2, var_7158, new_cache_35_internal_tensor_assign_2, var_7569, new_cache_37_internal_tensor_assign_2, var_7980, new_cache_39_internal_tensor_assign_2, var_8391, new_cache_41_internal_tensor_assign_2, var_8802, new_cache_43_internal_tensor_assign_2, var_9213, new_cache_45_internal_tensor_assign_2, var_9624, new_cache_internal_tensor_assign_2, var_9852); +} \ No newline at end of file diff --git a/v2.1/french_24l/cond_prefill.mlmodelc/weights/weight.bin b/v2.1/french_24l/cond_prefill.mlmodelc/weights/weight.bin new file mode 100644 index 0000000000000000000000000000000000000000..acb7413adee617034db507b72395e2593e5a63a6 --- /dev/null +++ b/v2.1/french_24l/cond_prefill.mlmodelc/weights/weight.bin @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:67d8d4f5360674983ff849bf4d218d657f2f21e81a0fa28007f12f98b46078c1 +size 586381056 diff --git a/v2.1/french_24l/cond_prefill.mlpackage/Data/com.apple.CoreML/model.mlmodel b/v2.1/french_24l/cond_prefill.mlpackage/Data/com.apple.CoreML/model.mlmodel new file mode 100644 index 0000000000000000000000000000000000000000..ec4a7a5a943ff99bbecf4c0f73770547256e66c0 --- /dev/null +++ b/v2.1/french_24l/cond_prefill.mlpackage/Data/com.apple.CoreML/model.mlmodel @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:3fe680b192e73def085b52aa6f5dc870cc6c3aded1ae1b355850e369bb01a7e7 +size 842494 diff --git a/v2.1/french_24l/cond_prefill.mlpackage/Data/com.apple.CoreML/weights/weight.bin b/v2.1/french_24l/cond_prefill.mlpackage/Data/com.apple.CoreML/weights/weight.bin new file mode 100644 index 0000000000000000000000000000000000000000..acb7413adee617034db507b72395e2593e5a63a6 --- /dev/null +++ b/v2.1/french_24l/cond_prefill.mlpackage/Data/com.apple.CoreML/weights/weight.bin @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:67d8d4f5360674983ff849bf4d218d657f2f21e81a0fa28007f12f98b46078c1 +size 586381056 diff --git a/v2.1/french_24l/cond_prefill.mlpackage/Manifest.json b/v2.1/french_24l/cond_prefill.mlpackage/Manifest.json new file mode 100644 index 0000000000000000000000000000000000000000..18caa8829c24e92769ad62ca1173d941d9976ac5 --- /dev/null +++ b/v2.1/french_24l/cond_prefill.mlpackage/Manifest.json @@ -0,0 +1,18 @@ +{ + "fileFormatVersion": "1.0.0", + "itemInfoEntries": { + "2A13BAFB-FA8C-48C0-A810-5C30F396D751": { + "author": "com.apple.CoreML", + "description": "CoreML Model Weights", + "name": "weights", + "path": "com.apple.CoreML/weights" + }, + "B850F1C7-33B8-4D1C-969C-FD22CD8D3570": { + "author": "com.apple.CoreML", + "description": "CoreML Model Specification", + "name": "model.mlmodel", + "path": "com.apple.CoreML/model.mlmodel" + } + }, + "rootModelIdentifier": "B850F1C7-33B8-4D1C-969C-FD22CD8D3570" +} diff --git a/v2.1/french_24l/constants/bos_emb.npy b/v2.1/french_24l/constants/bos_emb.npy new file mode 100644 index 0000000000000000000000000000000000000000..566d9b2901168cc1adecdfcc4d1773314ed585fb --- /dev/null +++ b/v2.1/french_24l/constants/bos_emb.npy @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:a83884c5cd8504340a1c8c2c0cfb528a68557be16e6583f5c1cbb127d7e17ced +size 256 diff --git a/v2.1/french_24l/constants/emb_mean.npy b/v2.1/french_24l/constants/emb_mean.npy new file mode 100644 index 0000000000000000000000000000000000000000..96d592c8bb9f0527a9ca83dc3a6e8d0a95e6395d --- /dev/null +++ b/v2.1/french_24l/constants/emb_mean.npy @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:97487bd8c7c9d520407abb440d18dffa8de4f96363113effcd5fd50fcd1e5e50 +size 256 diff --git a/v2.1/french_24l/constants/emb_std.npy b/v2.1/french_24l/constants/emb_std.npy new file mode 100644 index 0000000000000000000000000000000000000000..a2ae6cb7331ee625bdf997cdd274291dc962f8d4 --- /dev/null +++ b/v2.1/french_24l/constants/emb_std.npy @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:99d58fe567cb5f409bf17d5adfffbe8bade85ee866ea66d0a3e94a0cd3d59e0d +size 256 diff --git a/v2.1/french_24l/constants/mimi_init_state.npz b/v2.1/french_24l/constants/mimi_init_state.npz new file mode 100644 index 0000000000000000000000000000000000000000..3b52f383a86ef0439f9993506e4529c4c1234828 --- /dev/null +++ b/v2.1/french_24l/constants/mimi_init_state.npz @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:515588e10e9c14a6b009f1cfda9fcd592542b1640ad856d5fc69e50024702a39 +size 1083108 diff --git a/v2.1/french_24l/constants/quantizer_weight.npy b/v2.1/french_24l/constants/quantizer_weight.npy new file mode 100644 index 0000000000000000000000000000000000000000..cbd6c17eba9192dc343b255d61017d7c68c6f10c --- /dev/null +++ b/v2.1/french_24l/constants/quantizer_weight.npy @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:622ee2ef6db8e50658374e3b73971e9e80077e91e32a0dbdeba25f802fd05b04 +size 65664 diff --git a/v2.1/french_24l/constants/text_embed_table.npy b/v2.1/french_24l/constants/text_embed_table.npy new file mode 100644 index 0000000000000000000000000000000000000000..ead9739bc1a41c5300216a6fa4bcc0fd3587d1dc --- /dev/null +++ b/v2.1/french_24l/constants/text_embed_table.npy @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:88600b0444c5337e3205cb9dbe985ed40cb3b8fd3342f066a718540673684f9c +size 16388224 diff --git a/v2.1/french_24l/constants_bin/alba.safetensors b/v2.1/french_24l/constants_bin/alba.safetensors new file mode 100644 index 0000000000000000000000000000000000000000..1a5e4f0672c96c7e34a2988bc464a34a61b9bd04 --- /dev/null +++ b/v2.1/french_24l/constants_bin/alba.safetensors @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:81ef74464398d6db16c2734276293806845f755725a6b16feb4108b9ae781731 +size 24777760 diff --git a/v2.1/french_24l/constants_bin/anna.safetensors b/v2.1/french_24l/constants_bin/anna.safetensors new file mode 100644 index 0000000000000000000000000000000000000000..d934ee90da6bb7ef39e5b0d81cf2a4703127da49 --- /dev/null +++ b/v2.1/french_24l/constants_bin/anna.safetensors @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:8ec25e248cfe30c473bd3ceae408a7178eae072c40e45ccb1ed588b005d6c386 +size 31265824 diff --git a/v2.1/french_24l/constants_bin/azelma.safetensors b/v2.1/french_24l/constants_bin/azelma.safetensors new file mode 100644 index 0000000000000000000000000000000000000000..020a10f932de6c024cd67b916dca419a7d4d1330 --- /dev/null +++ b/v2.1/french_24l/constants_bin/azelma.safetensors @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:15db40136e4c33f8e4964ccf999b2bccba17f5897bd08406a4d789a33bb9d1b5 +size 31855648 diff --git a/v2.1/french_24l/constants_bin/bill_boerst.safetensors b/v2.1/french_24l/constants_bin/bill_boerst.safetensors new file mode 100644 index 0000000000000000000000000000000000000000..e68ebc15c9e81bdf9f1186e550db70f1c099b390 --- /dev/null +++ b/v2.1/french_24l/constants_bin/bill_boerst.safetensors @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:9a8ea532c38b4b5b346be3c5cb5468b47534ba5ba1e2e500bd1701f802665f90 +size 26940448 diff --git a/v2.1/french_24l/constants_bin/bos_before_voice.bin b/v2.1/french_24l/constants_bin/bos_before_voice.bin new file mode 100644 index 0000000000000000000000000000000000000000..f994ecdf54053809232e4b7f2bd713e63a2ae0eb --- /dev/null +++ b/v2.1/french_24l/constants_bin/bos_before_voice.bin @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:3db3c1dd0a65207fd53125249496a772147e2c41a0e7b0664ff71c5ced7729eb +size 4096 diff --git a/v2.1/french_24l/constants_bin/bos_emb.bin b/v2.1/french_24l/constants_bin/bos_emb.bin new file mode 100644 index 0000000000000000000000000000000000000000..fb15e235395641000579855a030588576e6ef9d5 --- /dev/null +++ b/v2.1/french_24l/constants_bin/bos_emb.bin @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:062f9c68cf2477920027c247050aaa5e6ad4a070192102a730ae6c327e7b7e76 +size 128 diff --git a/v2.1/french_24l/constants_bin/caro_davy.safetensors b/v2.1/french_24l/constants_bin/caro_davy.safetensors new file mode 100644 index 0000000000000000000000000000000000000000..ef0df52886ba1c07e973cfc610bcea346553c521 --- /dev/null +++ b/v2.1/french_24l/constants_bin/caro_davy.safetensors @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:1e645422e06d262288b42e51c873206ad365ceef9649d9c02d24231e6881f205 +size 21042200 diff --git a/v2.1/french_24l/constants_bin/charles.safetensors b/v2.1/french_24l/constants_bin/charles.safetensors new file mode 100644 index 0000000000000000000000000000000000000000..18f41f23a54b12d262c85493900fc89c1e3dddb9 --- /dev/null +++ b/v2.1/french_24l/constants_bin/charles.safetensors @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:5ba86d5829734b5f6ac378ea87e893d1d242cdafb8dfaa6c9bd14ac219e1fe47 +size 24777760 diff --git a/v2.1/french_24l/constants_bin/cosette.safetensors b/v2.1/french_24l/constants_bin/cosette.safetensors new file mode 100644 index 0000000000000000000000000000000000000000..ca6abb4c0ef71d2bfdc2ce9d59161b7323524538 --- /dev/null +++ b/v2.1/french_24l/constants_bin/cosette.safetensors @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:657e6a403de2cf159d1decdc6183fd3f8b12885885051c326061b4ee56ff492b +size 24777760 diff --git a/v2.1/french_24l/constants_bin/eponine.safetensors b/v2.1/french_24l/constants_bin/eponine.safetensors new file mode 100644 index 0000000000000000000000000000000000000000..603e554146079f01ee4d03dfd116dd0b8a1a70e3 --- /dev/null +++ b/v2.1/french_24l/constants_bin/eponine.safetensors @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:dc37497d3d2c6ebaf3b43390e4518c9995f0a9a34839d9b3cedac0505963e6fd +size 27726880 diff --git a/v2.1/french_24l/constants_bin/estelle.safetensors b/v2.1/french_24l/constants_bin/estelle.safetensors new file mode 100644 index 0000000000000000000000000000000000000000..b8e53fd59c3d3e59747f5f3c8d5bc62524b5ea61 --- /dev/null +++ b/v2.1/french_24l/constants_bin/estelle.safetensors @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:f1fa8ea7ea7f39f4d9a14c403ea3bd97af17e63e1e64ca9c8484efaedca93170 +size 33035296 diff --git a/v2.1/french_24l/constants_bin/eve.safetensors b/v2.1/french_24l/constants_bin/eve.safetensors new file mode 100644 index 0000000000000000000000000000000000000000..084ae92d93aebea783d52291298d6414b0a6057b --- /dev/null +++ b/v2.1/french_24l/constants_bin/eve.safetensors @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:caaa3ac419951d927ee40bdba22b3a5273f3d23564b2d573e1ec81a24c77051d +size 26154016 diff --git a/v2.1/french_24l/constants_bin/fantine.safetensors b/v2.1/french_24l/constants_bin/fantine.safetensors new file mode 100644 index 0000000000000000000000000000000000000000..457375010b081dd2dcff410b2b6ef9eb5300c508 --- /dev/null +++ b/v2.1/french_24l/constants_bin/fantine.safetensors @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:f19a024716b9132a310f9c1e1b750e11878a5079d6abbe9a01df6a78034f3db3 +size 26154016 diff --git a/v2.1/french_24l/constants_bin/george.safetensors b/v2.1/french_24l/constants_bin/george.safetensors new file mode 100644 index 0000000000000000000000000000000000000000..22adad628ab04e16de4b5ed5e1eea937db82665b --- /dev/null +++ b/v2.1/french_24l/constants_bin/george.safetensors @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:f98dd4400b5dca15a1a17c7ca7a9a31525c9e6acf03bd90e58da06bf2aec222e +size 24974368 diff --git a/v2.1/french_24l/constants_bin/giovanni.safetensors b/v2.1/french_24l/constants_bin/giovanni.safetensors new file mode 100644 index 0000000000000000000000000000000000000000..887d38f5049d82011f5af53658153a5c3256ad5f --- /dev/null +++ b/v2.1/french_24l/constants_bin/giovanni.safetensors @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:3c3a5311b3187f270883271e61f8fbd7a3a12c47140a759c0d05f83a37876de2 +size 18486272 diff --git a/v2.1/french_24l/constants_bin/jane.safetensors b/v2.1/french_24l/constants_bin/jane.safetensors new file mode 100644 index 0000000000000000000000000000000000000000..4df872d770d59329ed3c4fe5156801e1f50d1d0f --- /dev/null +++ b/v2.1/french_24l/constants_bin/jane.safetensors @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:799d8323c8808977f3d7f6faca208c29c8e737c2e16242d6a351a47997d6df89 +size 29496352 diff --git a/v2.1/french_24l/constants_bin/javert.safetensors b/v2.1/french_24l/constants_bin/javert.safetensors new file mode 100644 index 0000000000000000000000000000000000000000..bb68238d6b094576c8594e3562cc5859e9996e12 --- /dev/null +++ b/v2.1/french_24l/constants_bin/javert.safetensors @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:be49b3147a5a7a819fa190b5e44e471eef8b92cd18a6ed2594edff3996ba95bc +size 24777760 diff --git a/v2.1/french_24l/constants_bin/jean.safetensors b/v2.1/french_24l/constants_bin/jean.safetensors new file mode 100644 index 0000000000000000000000000000000000000000..339ed63b28a40363d2c5ba65638380896159ea79 --- /dev/null +++ b/v2.1/french_24l/constants_bin/jean.safetensors @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:13b53669e6ffed4e242044280fa8375dbd6c9002e2c6da5088d299e286bf5440 +size 24777760 diff --git a/v2.1/french_24l/constants_bin/juergen.safetensors b/v2.1/french_24l/constants_bin/juergen.safetensors new file mode 100644 index 0000000000000000000000000000000000000000..6ee366690dc12055f295f19b29551caf82364874 --- /dev/null +++ b/v2.1/french_24l/constants_bin/juergen.safetensors @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:7c695bf10e8ea551f3151fe3c5de787fa5a6353665aab190be14dd6cd1320c93 +size 24974368 diff --git a/v2.1/french_24l/constants_bin/lola.safetensors b/v2.1/french_24l/constants_bin/lola.safetensors new file mode 100644 index 0000000000000000000000000000000000000000..d2fed5807aa08a5fac002a87f14c7f5a322b0b2b --- /dev/null +++ b/v2.1/french_24l/constants_bin/lola.safetensors @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:b9fcbf24e36171e260bde32f2834421658fa68d4d19f50e013f44122b96ebc3f +size 23794712 diff --git a/v2.1/french_24l/constants_bin/marius.safetensors b/v2.1/french_24l/constants_bin/marius.safetensors new file mode 100644 index 0000000000000000000000000000000000000000..459749fa57b46208f8e3de209be65f0681d9aba9 --- /dev/null +++ b/v2.1/french_24l/constants_bin/marius.safetensors @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:6e2e12dc9a7650b3e3d5a28be9a2550ced91c5a11d5de9021efc6107dcc78587 +size 24777760 diff --git a/v2.1/french_24l/constants_bin/mary.safetensors b/v2.1/french_24l/constants_bin/mary.safetensors new file mode 100644 index 0000000000000000000000000000000000000000..5cdc033b6c99e1f5cfa6927ab82bd431a371a073 --- /dev/null +++ b/v2.1/french_24l/constants_bin/mary.safetensors @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:ad3d85829605e0b28e509c65350198c15f2a0592a39bf88063b4d01b4085c8b8 +size 24777760 diff --git a/v2.1/french_24l/constants_bin/michael.safetensors b/v2.1/french_24l/constants_bin/michael.safetensors new file mode 100644 index 0000000000000000000000000000000000000000..bfa8aaa68f3da9987bd1597ef38287e17905ec62 --- /dev/null +++ b/v2.1/french_24l/constants_bin/michael.safetensors @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:2d1dc47d55c149d79701c6d8e9c7f8efab7dfd397616526969b138a05619d0b4 +size 29103136 diff --git a/v2.1/french_24l/constants_bin/paul.safetensors b/v2.1/french_24l/constants_bin/paul.safetensors new file mode 100644 index 0000000000000000000000000000000000000000..ebf9797832025cbb53dca45a1f5e28d1a64e07d6 --- /dev/null +++ b/v2.1/french_24l/constants_bin/paul.safetensors @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:5dd222567a9c167ed3c39ea5f336e5ab0fcd51be653d2de68d9a8f9db19757c2 +size 27923488 diff --git a/v2.1/french_24l/constants_bin/peter_yearsley.safetensors b/v2.1/french_24l/constants_bin/peter_yearsley.safetensors new file mode 100644 index 0000000000000000000000000000000000000000..ca31112bf717586d8ae59bbb75ce251e810cf919 --- /dev/null +++ b/v2.1/french_24l/constants_bin/peter_yearsley.safetensors @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:997e18754aa28223e48fe19f8b4025b55cd5098b06cb36a4ab842c0325ad6ddc +size 14947320 diff --git a/v2.1/french_24l/constants_bin/rafael.safetensors b/v2.1/french_24l/constants_bin/rafael.safetensors new file mode 100644 index 0000000000000000000000000000000000000000..29bb57d4ee85f6de5687ca3b2cd6731a43ce107b --- /dev/null +++ b/v2.1/french_24l/constants_bin/rafael.safetensors @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:faf602de5907c0d022a551b97dd35d886cda71e4762afb5fd2db1a0bd3bb4326 +size 24777760 diff --git a/v2.1/french_24l/constants_bin/stuart_bell.safetensors b/v2.1/french_24l/constants_bin/stuart_bell.safetensors new file mode 100644 index 0000000000000000000000000000000000000000..3560d94e9f2aac8f45b550d0199715724a454a69 --- /dev/null +++ b/v2.1/french_24l/constants_bin/stuart_bell.safetensors @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:38793a0909ca62261d774d2ef09bba0120c552e67690153f4b061426fa36c791 +size 21042200 diff --git a/v2.1/french_24l/constants_bin/text_embed_table.bin b/v2.1/french_24l/constants_bin/text_embed_table.bin new file mode 100644 index 0000000000000000000000000000000000000000..bc9814fe9d60247a0e642fa669fd4ff2e8b55b13 --- /dev/null +++ b/v2.1/french_24l/constants_bin/text_embed_table.bin @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:8db4694230c0fd8e705dd731b57aa6e41d380c034b593f82b1560273c678d15b +size 16388096 diff --git a/v2.1/french_24l/constants_bin/tokenizer.model b/v2.1/french_24l/constants_bin/tokenizer.model new file mode 100644 index 0000000000000000000000000000000000000000..266f5edf9fe923eb2464e352c89e5e768bd77935 --- /dev/null +++ b/v2.1/french_24l/constants_bin/tokenizer.model @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:521c85bdb2da10618f4be52021ed1cb2a7a6299b040708487f133193f7b305e2 +size 60173 diff --git a/v2.1/french_24l/constants_bin/vera.safetensors b/v2.1/french_24l/constants_bin/vera.safetensors new file mode 100644 index 0000000000000000000000000000000000000000..44f3bdf2f5d360e789df469f84926d3f0efb06dc --- /dev/null +++ b/v2.1/french_24l/constants_bin/vera.safetensors @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:8533421ac10c6b169bb6271a16f50a1e851bddfe914ab8a7509a5b7908a9686e +size 26940448 diff --git a/v2.1/french_24l/flow_decoder_fused.mlmodelc/analytics/coremldata.bin b/v2.1/french_24l/flow_decoder_fused.mlmodelc/analytics/coremldata.bin new file mode 100644 index 0000000000000000000000000000000000000000..8e3376c668297f2b9b79eb3e6be1e2152e0a5069 --- /dev/null +++ b/v2.1/french_24l/flow_decoder_fused.mlmodelc/analytics/coremldata.bin @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:b83a8ae055446657c1bb78a915622efdec5d796d3438331e5879aa0050dd223b +size 243 diff --git a/v2.1/french_24l/flow_decoder_fused.mlmodelc/coremldata.bin b/v2.1/french_24l/flow_decoder_fused.mlmodelc/coremldata.bin new file mode 100644 index 0000000000000000000000000000000000000000..2193e383f54291453d675f31dd28a51c6a3b1f75 --- /dev/null +++ b/v2.1/french_24l/flow_decoder_fused.mlmodelc/coremldata.bin @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:b77b677b547d6ce2ee01d58dc20e75cd5a5ecf2ce86b0cb82a32bfc1c6aec496 +size 413 diff --git a/v2.1/french_24l/flow_decoder_fused.mlmodelc/model.mil b/v2.1/french_24l/flow_decoder_fused.mlmodelc/model.mil new file mode 100644 index 0000000000000000000000000000000000000000..c64a1f34555e7bbf9c8d4502d29f9ef52dd6b0ec --- /dev/null +++ b/v2.1/french_24l/flow_decoder_fused.mlmodelc/model.mil @@ -0,0 +1,1915 @@ +program(1.0) +[buildInfo = dict, tensor>({{"coremlc-component-MIL", "3520.4.1"}, {"coremlc-version", "3520.5.1"}, {"coremltools-component-torch", "2.12.0"}, {"coremltools-source-dialect", "TorchScript"}, {"coremltools-version", "9.0"}})] +{ + func main(tensor latent_init, tensor transformer_out) { + tensor latent_init_to_fp16_dtype_0 = const()[name = tensor("latent_init_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor flow_net_input_proj_weight_to_fp16 = const()[name = tensor("flow_net_input_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(64)))]; + tensor flow_net_input_proj_bias_to_fp16 = const()[name = tensor("flow_net_input_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(32896)))]; + tensor latent_init_to_fp16 = cast(dtype = latent_init_to_fp16_dtype_0, x = latent_init)[name = tensor("cast_146")]; + tensor linear_0_cast_fp16 = linear(bias = flow_net_input_proj_bias_to_fp16, weight = flow_net_input_proj_weight_to_fp16, x = latent_init_to_fp16)[name = tensor("linear_0_cast_fp16")]; + tensor input_3_to_fp16 = const()[name = tensor("input_3_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(33984)))]; + tensor input_5_cast_fp16 = silu(x = input_3_to_fp16)[name = tensor("input_5_cast_fp16")]; + tensor flow_net_time_embed_0_mlp_2_weight_to_fp16 = const()[name = tensor("flow_net_time_embed_0_mlp_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(35072)))]; + tensor flow_net_time_embed_0_mlp_2_bias_to_fp16 = const()[name = tensor("flow_net_time_embed_0_mlp_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(559424)))]; + tensor linear_2_cast_fp16 = linear(bias = flow_net_time_embed_0_mlp_2_bias_to_fp16, weight = flow_net_time_embed_0_mlp_2_weight_to_fp16, x = input_5_cast_fp16)[name = tensor("linear_2_cast_fp16")]; + tensor reduce_mean_0_axes_0 = const()[name = tensor("reduce_mean_0_axes_0"), val = tensor([-1])]; + tensor reduce_mean_0_keep_dims_0 = const()[name = tensor("reduce_mean_0_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_0_cast_fp16 = reduce_mean(axes = reduce_mean_0_axes_0, keep_dims = reduce_mean_0_keep_dims_0, x = linear_2_cast_fp16)[name = tensor("reduce_mean_0_cast_fp16")]; + tensor sub_0_cast_fp16 = sub(x = linear_2_cast_fp16, y = reduce_mean_0_cast_fp16)[name = tensor("sub_0_cast_fp16")]; + tensor square_0_cast_fp16 = square(x = sub_0_cast_fp16)[name = tensor("square_0_cast_fp16")]; + tensor reduce_mean_1_axes_0 = const()[name = tensor("reduce_mean_1_axes_0"), val = tensor([-1])]; + tensor reduce_mean_1_keep_dims_0 = const()[name = tensor("reduce_mean_1_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_1_cast_fp16 = reduce_mean(axes = reduce_mean_1_axes_0, keep_dims = reduce_mean_1_keep_dims_0, x = square_0_cast_fp16)[name = tensor("reduce_mean_1_cast_fp16")]; + tensor real_div_0_to_fp16 = const()[name = tensor("real_div_0_to_fp16"), val = tensor(0x1.008p+0)]; + tensor mul_0_cast_fp16 = mul(x = reduce_mean_1_cast_fp16, y = real_div_0_to_fp16)[name = tensor("mul_0_cast_fp16")]; + tensor var_70_to_fp16 = const()[name = tensor("op_70_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_1_cast_fp16 = add(x = mul_0_cast_fp16, y = var_70_to_fp16)[name = tensor("var_1_cast_fp16")]; + tensor var_73_epsilon_0 = const()[name = tensor("op_73_epsilon_0"), val = tensor(0x1.197998p-40)]; + tensor var_73_cast_fp16 = rsqrt(epsilon = var_73_epsilon_0, x = var_1_cast_fp16)[name = tensor("op_73_cast_fp16")]; + tensor const_3_to_fp16 = const()[name = tensor("const_3_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(560512)))]; + tensor var_74_cast_fp16 = mul(x = const_3_to_fp16, y = var_73_cast_fp16)[name = tensor("op_74_cast_fp16")]; + tensor var_75_cast_fp16 = mul(x = linear_2_cast_fp16, y = var_74_cast_fp16)[name = tensor("op_75_cast_fp16")]; + tensor input_9_to_fp16 = const()[name = tensor("input_9_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(561600)))]; + tensor input_11_cast_fp16 = silu(x = input_9_to_fp16)[name = tensor("input_11_cast_fp16")]; + tensor flow_net_time_embed_1_mlp_2_weight_to_fp16 = const()[name = tensor("flow_net_time_embed_1_mlp_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(562688)))]; + tensor flow_net_time_embed_1_mlp_2_bias_to_fp16 = const()[name = tensor("flow_net_time_embed_1_mlp_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1087040)))]; + tensor linear_4_cast_fp16 = linear(bias = flow_net_time_embed_1_mlp_2_bias_to_fp16, weight = flow_net_time_embed_1_mlp_2_weight_to_fp16, x = input_11_cast_fp16)[name = tensor("linear_4_cast_fp16")]; + tensor reduce_mean_2_axes_0 = const()[name = tensor("reduce_mean_2_axes_0"), val = tensor([-1])]; + tensor reduce_mean_2_keep_dims_0 = const()[name = tensor("reduce_mean_2_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_2_cast_fp16 = reduce_mean(axes = reduce_mean_2_axes_0, keep_dims = reduce_mean_2_keep_dims_0, x = linear_4_cast_fp16)[name = tensor("reduce_mean_2_cast_fp16")]; + tensor sub_2_cast_fp16 = sub(x = linear_4_cast_fp16, y = reduce_mean_2_cast_fp16)[name = tensor("sub_2_cast_fp16")]; + tensor square_1_cast_fp16 = square(x = sub_2_cast_fp16)[name = tensor("square_1_cast_fp16")]; + tensor reduce_mean_3_axes_0 = const()[name = tensor("reduce_mean_3_axes_0"), val = tensor([-1])]; + tensor reduce_mean_3_keep_dims_0 = const()[name = tensor("reduce_mean_3_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_3_cast_fp16 = reduce_mean(axes = reduce_mean_3_axes_0, keep_dims = reduce_mean_3_keep_dims_0, x = square_1_cast_fp16)[name = tensor("reduce_mean_3_cast_fp16")]; + tensor real_div_1_to_fp16 = const()[name = tensor("real_div_1_to_fp16"), val = tensor(0x1.008p+0)]; + tensor mul_1_cast_fp16 = mul(x = reduce_mean_3_cast_fp16, y = real_div_1_to_fp16)[name = tensor("mul_1_cast_fp16")]; + tensor var_110_to_fp16 = const()[name = tensor("op_110_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_3_cast_fp16 = add(x = mul_1_cast_fp16, y = var_110_to_fp16)[name = tensor("var_3_cast_fp16")]; + tensor var_113_epsilon_0 = const()[name = tensor("op_113_epsilon_0"), val = tensor(0x1.197998p-40)]; + tensor var_113_cast_fp16 = rsqrt(epsilon = var_113_epsilon_0, x = var_3_cast_fp16)[name = tensor("op_113_cast_fp16")]; + tensor const_5_to_fp16 = const()[name = tensor("const_5_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1088128)))]; + tensor var_114_cast_fp16 = mul(x = const_5_to_fp16, y = var_113_cast_fp16)[name = tensor("op_114_cast_fp16")]; + tensor var_115_cast_fp16 = mul(x = linear_4_cast_fp16, y = var_114_cast_fp16)[name = tensor("op_115_cast_fp16")]; + tensor var_127_cast_fp16 = add(x = var_75_cast_fp16, y = var_115_cast_fp16)[name = tensor("op_127_cast_fp16")]; + tensor _inversed_t_combined_1_y_0_to_fp16 = const()[name = tensor("_inversed_t_combined_1_y_0_to_fp16"), val = tensor(0x1p-1)]; + tensor _inversed_t_combined_1_cast_fp16 = mul(x = var_127_cast_fp16, y = _inversed_t_combined_1_y_0_to_fp16)[name = tensor("_inversed_t_combined_1_cast_fp16")]; + tensor transformer_out_to_fp16_dtype_0 = const()[name = tensor("transformer_out_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor flow_net_cond_embed_weight_to_fp16 = const()[name = tensor("flow_net_cond_embed_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1089216)))]; + tensor flow_net_cond_embed_bias_to_fp16 = const()[name = tensor("flow_net_cond_embed_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(2137856)))]; + tensor transformer_out_to_fp16 = cast(dtype = transformer_out_to_fp16_dtype_0, x = transformer_out)[name = tensor("cast_145")]; + tensor linear_5_cast_fp16 = linear(bias = flow_net_cond_embed_bias_to_fp16, weight = flow_net_cond_embed_weight_to_fp16, x = transformer_out_to_fp16)[name = tensor("linear_5_cast_fp16")]; + tensor input_13_cast_fp16 = add(x = _inversed_t_combined_1_cast_fp16, y = linear_5_cast_fp16)[name = tensor("input_13_cast_fp16")]; + tensor input_15_cast_fp16 = silu(x = input_13_cast_fp16)[name = tensor("input_15_cast_fp16")]; + tensor flow_net_res_blocks_0_adaLN_modulation_1_weight_to_fp16 = const()[name = tensor("flow_net_res_blocks_0_adaLN_modulation_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(2138944)))]; + tensor flow_net_res_blocks_0_adaLN_modulation_1_bias_to_fp16 = const()[name = tensor("flow_net_res_blocks_0_adaLN_modulation_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(3711872)))]; + tensor linear_6_cast_fp16 = linear(bias = flow_net_res_blocks_0_adaLN_modulation_1_bias_to_fp16, weight = flow_net_res_blocks_0_adaLN_modulation_1_weight_to_fp16, x = input_15_cast_fp16)[name = tensor("linear_6_cast_fp16")]; + tensor var_142_split_sizes_0 = const()[name = tensor("op_142_split_sizes_0"), val = tensor([512, 512, 512])]; + tensor var_142_axis_0 = const()[name = tensor("op_142_axis_0"), val = tensor(-1)]; + tensor var_142_cast_fp16_0, tensor var_142_cast_fp16_1, tensor var_142_cast_fp16_2 = split(axis = var_142_axis_0, split_sizes = var_142_split_sizes_0, x = linear_6_cast_fp16)[name = tensor("op_142_cast_fp16")]; + tensor mean_1_axes_0 = const()[name = tensor("mean_1_axes_0"), val = tensor([-1])]; + tensor mean_1_keep_dims_0 = const()[name = tensor("mean_1_keep_dims_0"), val = tensor(true)]; + tensor mean_1_cast_fp16 = reduce_mean(axes = mean_1_axes_0, keep_dims = mean_1_keep_dims_0, x = linear_0_cast_fp16)[name = tensor("mean_1_cast_fp16")]; + tensor sub_4_cast_fp16 = sub(x = linear_0_cast_fp16, y = mean_1_cast_fp16)[name = tensor("sub_4_cast_fp16")]; + tensor square_2_cast_fp16 = square(x = sub_4_cast_fp16)[name = tensor("square_2_cast_fp16")]; + tensor reduce_mean_5_axes_0 = const()[name = tensor("reduce_mean_5_axes_0"), val = tensor([-1])]; + tensor reduce_mean_5_keep_dims_0 = const()[name = tensor("reduce_mean_5_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_5_cast_fp16 = reduce_mean(axes = reduce_mean_5_axes_0, keep_dims = reduce_mean_5_keep_dims_0, x = square_2_cast_fp16)[name = tensor("reduce_mean_5_cast_fp16")]; + tensor var_152_to_fp16 = const()[name = tensor("op_152_to_fp16"), val = tensor(0x1.1p-20)]; + tensor var_153_cast_fp16 = add(x = reduce_mean_5_cast_fp16, y = var_152_to_fp16)[name = tensor("op_153_cast_fp16")]; + tensor var_154_cast_fp16 = sqrt(x = var_153_cast_fp16)[name = tensor("op_154_cast_fp16")]; + tensor x_7_cast_fp16 = real_div(x = sub_4_cast_fp16, y = var_154_cast_fp16)[name = tensor("x_7_cast_fp16")]; + tensor flow_net_res_blocks_0_in_ln_weight_to_fp16 = const()[name = tensor("flow_net_res_blocks_0_in_ln_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(3715008)))]; + tensor var_156_cast_fp16 = mul(x = x_7_cast_fp16, y = flow_net_res_blocks_0_in_ln_weight_to_fp16)[name = tensor("op_156_cast_fp16")]; + tensor flow_net_res_blocks_0_in_ln_bias_to_fp16 = const()[name = tensor("flow_net_res_blocks_0_in_ln_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(3716096)))]; + tensor x_9_cast_fp16 = add(x = var_156_cast_fp16, y = flow_net_res_blocks_0_in_ln_bias_to_fp16)[name = tensor("x_9_cast_fp16")]; + tensor var_158_promoted_to_fp16 = const()[name = tensor("op_158_promoted_to_fp16"), val = tensor(0x1p+0)]; + tensor var_159_cast_fp16 = add(x = var_142_cast_fp16_1, y = var_158_promoted_to_fp16)[name = tensor("op_159_cast_fp16")]; + tensor var_160_cast_fp16 = mul(x = x_9_cast_fp16, y = var_159_cast_fp16)[name = tensor("op_160_cast_fp16")]; + tensor input_17_cast_fp16 = add(x = var_160_cast_fp16, y = var_142_cast_fp16_0)[name = tensor("input_17_cast_fp16")]; + tensor flow_net_res_blocks_0_mlp_0_weight_to_fp16 = const()[name = tensor("flow_net_res_blocks_0_mlp_0_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(3717184)))]; + tensor flow_net_res_blocks_0_mlp_0_bias_to_fp16 = const()[name = tensor("flow_net_res_blocks_0_mlp_0_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(4241536)))]; + tensor linear_7_cast_fp16 = linear(bias = flow_net_res_blocks_0_mlp_0_bias_to_fp16, weight = flow_net_res_blocks_0_mlp_0_weight_to_fp16, x = input_17_cast_fp16)[name = tensor("linear_7_cast_fp16")]; + tensor input_21_cast_fp16 = silu(x = linear_7_cast_fp16)[name = tensor("input_21_cast_fp16")]; + tensor flow_net_res_blocks_0_mlp_2_weight_to_fp16 = const()[name = tensor("flow_net_res_blocks_0_mlp_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(4242624)))]; + tensor flow_net_res_blocks_0_mlp_2_bias_to_fp16 = const()[name = tensor("flow_net_res_blocks_0_mlp_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(4766976)))]; + tensor linear_8_cast_fp16 = linear(bias = flow_net_res_blocks_0_mlp_2_bias_to_fp16, weight = flow_net_res_blocks_0_mlp_2_weight_to_fp16, x = input_21_cast_fp16)[name = tensor("linear_8_cast_fp16")]; + tensor var_171_cast_fp16 = mul(x = var_142_cast_fp16_2, y = linear_8_cast_fp16)[name = tensor("op_171_cast_fp16")]; + tensor x_11_cast_fp16 = add(x = linear_0_cast_fp16, y = var_171_cast_fp16)[name = tensor("x_11_cast_fp16")]; + tensor flow_net_res_blocks_1_adaLN_modulation_1_weight_to_fp16 = const()[name = tensor("flow_net_res_blocks_1_adaLN_modulation_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(4768064)))]; + tensor flow_net_res_blocks_1_adaLN_modulation_1_bias_to_fp16 = const()[name = tensor("flow_net_res_blocks_1_adaLN_modulation_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(6340992)))]; + tensor linear_9_cast_fp16 = linear(bias = flow_net_res_blocks_1_adaLN_modulation_1_bias_to_fp16, weight = flow_net_res_blocks_1_adaLN_modulation_1_weight_to_fp16, x = input_15_cast_fp16)[name = tensor("linear_9_cast_fp16")]; + tensor var_181_split_sizes_0 = const()[name = tensor("op_181_split_sizes_0"), val = tensor([512, 512, 512])]; + tensor var_181_axis_0 = const()[name = tensor("op_181_axis_0"), val = tensor(-1)]; + tensor var_181_cast_fp16_0, tensor var_181_cast_fp16_1, tensor var_181_cast_fp16_2 = split(axis = var_181_axis_0, split_sizes = var_181_split_sizes_0, x = linear_9_cast_fp16)[name = tensor("op_181_cast_fp16")]; + tensor mean_3_axes_0 = const()[name = tensor("mean_3_axes_0"), val = tensor([-1])]; + tensor mean_3_keep_dims_0 = const()[name = tensor("mean_3_keep_dims_0"), val = tensor(true)]; + tensor mean_3_cast_fp16 = reduce_mean(axes = mean_3_axes_0, keep_dims = mean_3_keep_dims_0, x = x_11_cast_fp16)[name = tensor("mean_3_cast_fp16")]; + tensor sub_5_cast_fp16 = sub(x = x_11_cast_fp16, y = mean_3_cast_fp16)[name = tensor("sub_5_cast_fp16")]; + tensor square_3_cast_fp16 = square(x = sub_5_cast_fp16)[name = tensor("square_3_cast_fp16")]; + tensor reduce_mean_7_axes_0 = const()[name = tensor("reduce_mean_7_axes_0"), val = tensor([-1])]; + tensor reduce_mean_7_keep_dims_0 = const()[name = tensor("reduce_mean_7_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_7_cast_fp16 = reduce_mean(axes = reduce_mean_7_axes_0, keep_dims = reduce_mean_7_keep_dims_0, x = square_3_cast_fp16)[name = tensor("reduce_mean_7_cast_fp16")]; + tensor var_191_to_fp16 = const()[name = tensor("op_191_to_fp16"), val = tensor(0x1.1p-20)]; + tensor var_192_cast_fp16 = add(x = reduce_mean_7_cast_fp16, y = var_191_to_fp16)[name = tensor("op_192_cast_fp16")]; + tensor var_193_cast_fp16 = sqrt(x = var_192_cast_fp16)[name = tensor("op_193_cast_fp16")]; + tensor x_13_cast_fp16 = real_div(x = sub_5_cast_fp16, y = var_193_cast_fp16)[name = tensor("x_13_cast_fp16")]; + tensor flow_net_res_blocks_1_in_ln_weight_to_fp16 = const()[name = tensor("flow_net_res_blocks_1_in_ln_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(6344128)))]; + tensor var_195_cast_fp16 = mul(x = x_13_cast_fp16, y = flow_net_res_blocks_1_in_ln_weight_to_fp16)[name = tensor("op_195_cast_fp16")]; + tensor flow_net_res_blocks_1_in_ln_bias_to_fp16 = const()[name = tensor("flow_net_res_blocks_1_in_ln_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(6345216)))]; + tensor x_15_cast_fp16 = add(x = var_195_cast_fp16, y = flow_net_res_blocks_1_in_ln_bias_to_fp16)[name = tensor("x_15_cast_fp16")]; + tensor var_197_promoted_to_fp16 = const()[name = tensor("op_197_promoted_to_fp16"), val = tensor(0x1p+0)]; + tensor var_198_cast_fp16 = add(x = var_181_cast_fp16_1, y = var_197_promoted_to_fp16)[name = tensor("op_198_cast_fp16")]; + tensor var_199_cast_fp16 = mul(x = x_15_cast_fp16, y = var_198_cast_fp16)[name = tensor("op_199_cast_fp16")]; + tensor input_25_cast_fp16 = add(x = var_199_cast_fp16, y = var_181_cast_fp16_0)[name = tensor("input_25_cast_fp16")]; + tensor flow_net_res_blocks_1_mlp_0_weight_to_fp16 = const()[name = tensor("flow_net_res_blocks_1_mlp_0_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(6346304)))]; + tensor flow_net_res_blocks_1_mlp_0_bias_to_fp16 = const()[name = tensor("flow_net_res_blocks_1_mlp_0_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(6870656)))]; + tensor linear_10_cast_fp16 = linear(bias = flow_net_res_blocks_1_mlp_0_bias_to_fp16, weight = flow_net_res_blocks_1_mlp_0_weight_to_fp16, x = input_25_cast_fp16)[name = tensor("linear_10_cast_fp16")]; + tensor input_29_cast_fp16 = silu(x = linear_10_cast_fp16)[name = tensor("input_29_cast_fp16")]; + tensor flow_net_res_blocks_1_mlp_2_weight_to_fp16 = const()[name = tensor("flow_net_res_blocks_1_mlp_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(6871744)))]; + tensor flow_net_res_blocks_1_mlp_2_bias_to_fp16 = const()[name = tensor("flow_net_res_blocks_1_mlp_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(7396096)))]; + tensor linear_11_cast_fp16 = linear(bias = flow_net_res_blocks_1_mlp_2_bias_to_fp16, weight = flow_net_res_blocks_1_mlp_2_weight_to_fp16, x = input_29_cast_fp16)[name = tensor("linear_11_cast_fp16")]; + tensor var_210_cast_fp16 = mul(x = var_181_cast_fp16_2, y = linear_11_cast_fp16)[name = tensor("op_210_cast_fp16")]; + tensor x_17_cast_fp16 = add(x = x_11_cast_fp16, y = var_210_cast_fp16)[name = tensor("x_17_cast_fp16")]; + tensor flow_net_res_blocks_2_adaLN_modulation_1_weight_to_fp16 = const()[name = tensor("flow_net_res_blocks_2_adaLN_modulation_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(7397184)))]; + tensor flow_net_res_blocks_2_adaLN_modulation_1_bias_to_fp16 = const()[name = tensor("flow_net_res_blocks_2_adaLN_modulation_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8970112)))]; + tensor linear_12_cast_fp16 = linear(bias = flow_net_res_blocks_2_adaLN_modulation_1_bias_to_fp16, weight = flow_net_res_blocks_2_adaLN_modulation_1_weight_to_fp16, x = input_15_cast_fp16)[name = tensor("linear_12_cast_fp16")]; + tensor var_220_split_sizes_0 = const()[name = tensor("op_220_split_sizes_0"), val = tensor([512, 512, 512])]; + tensor var_220_axis_0 = const()[name = tensor("op_220_axis_0"), val = tensor(-1)]; + tensor var_220_cast_fp16_0, tensor var_220_cast_fp16_1, tensor var_220_cast_fp16_2 = split(axis = var_220_axis_0, split_sizes = var_220_split_sizes_0, x = linear_12_cast_fp16)[name = tensor("op_220_cast_fp16")]; + tensor mean_5_axes_0 = const()[name = tensor("mean_5_axes_0"), val = tensor([-1])]; + tensor mean_5_keep_dims_0 = const()[name = tensor("mean_5_keep_dims_0"), val = tensor(true)]; + tensor mean_5_cast_fp16 = reduce_mean(axes = mean_5_axes_0, keep_dims = mean_5_keep_dims_0, x = x_17_cast_fp16)[name = tensor("mean_5_cast_fp16")]; + tensor sub_6_cast_fp16 = sub(x = x_17_cast_fp16, y = mean_5_cast_fp16)[name = tensor("sub_6_cast_fp16")]; + tensor square_4_cast_fp16 = square(x = sub_6_cast_fp16)[name = tensor("square_4_cast_fp16")]; + tensor reduce_mean_9_axes_0 = const()[name = tensor("reduce_mean_9_axes_0"), val = tensor([-1])]; + tensor reduce_mean_9_keep_dims_0 = const()[name = tensor("reduce_mean_9_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_9_cast_fp16 = reduce_mean(axes = reduce_mean_9_axes_0, keep_dims = reduce_mean_9_keep_dims_0, x = square_4_cast_fp16)[name = tensor("reduce_mean_9_cast_fp16")]; + tensor var_230_to_fp16 = const()[name = tensor("op_230_to_fp16"), val = tensor(0x1.1p-20)]; + tensor var_231_cast_fp16 = add(x = reduce_mean_9_cast_fp16, y = var_230_to_fp16)[name = tensor("op_231_cast_fp16")]; + tensor var_232_cast_fp16 = sqrt(x = var_231_cast_fp16)[name = tensor("op_232_cast_fp16")]; + tensor x_19_cast_fp16 = real_div(x = sub_6_cast_fp16, y = var_232_cast_fp16)[name = tensor("x_19_cast_fp16")]; + tensor flow_net_res_blocks_2_in_ln_weight_to_fp16 = const()[name = tensor("flow_net_res_blocks_2_in_ln_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8973248)))]; + tensor var_234_cast_fp16 = mul(x = x_19_cast_fp16, y = flow_net_res_blocks_2_in_ln_weight_to_fp16)[name = tensor("op_234_cast_fp16")]; + tensor flow_net_res_blocks_2_in_ln_bias_to_fp16 = const()[name = tensor("flow_net_res_blocks_2_in_ln_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8974336)))]; + tensor x_21_cast_fp16 = add(x = var_234_cast_fp16, y = flow_net_res_blocks_2_in_ln_bias_to_fp16)[name = tensor("x_21_cast_fp16")]; + tensor var_236_promoted_to_fp16 = const()[name = tensor("op_236_promoted_to_fp16"), val = tensor(0x1p+0)]; + tensor var_237_cast_fp16 = add(x = var_220_cast_fp16_1, y = var_236_promoted_to_fp16)[name = tensor("op_237_cast_fp16")]; + tensor var_238_cast_fp16 = mul(x = x_21_cast_fp16, y = var_237_cast_fp16)[name = tensor("op_238_cast_fp16")]; + tensor input_33_cast_fp16 = add(x = var_238_cast_fp16, y = var_220_cast_fp16_0)[name = tensor("input_33_cast_fp16")]; + tensor flow_net_res_blocks_2_mlp_0_weight_to_fp16 = const()[name = tensor("flow_net_res_blocks_2_mlp_0_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8975424)))]; + tensor flow_net_res_blocks_2_mlp_0_bias_to_fp16 = const()[name = tensor("flow_net_res_blocks_2_mlp_0_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9499776)))]; + tensor linear_13_cast_fp16 = linear(bias = flow_net_res_blocks_2_mlp_0_bias_to_fp16, weight = flow_net_res_blocks_2_mlp_0_weight_to_fp16, x = input_33_cast_fp16)[name = tensor("linear_13_cast_fp16")]; + tensor input_37_cast_fp16 = silu(x = linear_13_cast_fp16)[name = tensor("input_37_cast_fp16")]; + tensor flow_net_res_blocks_2_mlp_2_weight_to_fp16 = const()[name = tensor("flow_net_res_blocks_2_mlp_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9500864)))]; + tensor flow_net_res_blocks_2_mlp_2_bias_to_fp16 = const()[name = tensor("flow_net_res_blocks_2_mlp_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(10025216)))]; + tensor linear_14_cast_fp16 = linear(bias = flow_net_res_blocks_2_mlp_2_bias_to_fp16, weight = flow_net_res_blocks_2_mlp_2_weight_to_fp16, x = input_37_cast_fp16)[name = tensor("linear_14_cast_fp16")]; + tensor var_249_cast_fp16 = mul(x = var_220_cast_fp16_2, y = linear_14_cast_fp16)[name = tensor("op_249_cast_fp16")]; + tensor x_23_cast_fp16 = add(x = x_17_cast_fp16, y = var_249_cast_fp16)[name = tensor("x_23_cast_fp16")]; + tensor flow_net_res_blocks_3_adaLN_modulation_1_weight_to_fp16 = const()[name = tensor("flow_net_res_blocks_3_adaLN_modulation_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(10026304)))]; + tensor flow_net_res_blocks_3_adaLN_modulation_1_bias_to_fp16 = const()[name = tensor("flow_net_res_blocks_3_adaLN_modulation_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(11599232)))]; + tensor linear_15_cast_fp16 = linear(bias = flow_net_res_blocks_3_adaLN_modulation_1_bias_to_fp16, weight = flow_net_res_blocks_3_adaLN_modulation_1_weight_to_fp16, x = input_15_cast_fp16)[name = tensor("linear_15_cast_fp16")]; + tensor var_259_split_sizes_0 = const()[name = tensor("op_259_split_sizes_0"), val = tensor([512, 512, 512])]; + tensor var_259_axis_0 = const()[name = tensor("op_259_axis_0"), val = tensor(-1)]; + tensor var_259_cast_fp16_0, tensor var_259_cast_fp16_1, tensor var_259_cast_fp16_2 = split(axis = var_259_axis_0, split_sizes = var_259_split_sizes_0, x = linear_15_cast_fp16)[name = tensor("op_259_cast_fp16")]; + tensor mean_7_axes_0 = const()[name = tensor("mean_7_axes_0"), val = tensor([-1])]; + tensor mean_7_keep_dims_0 = const()[name = tensor("mean_7_keep_dims_0"), val = tensor(true)]; + tensor mean_7_cast_fp16 = reduce_mean(axes = mean_7_axes_0, keep_dims = mean_7_keep_dims_0, x = x_23_cast_fp16)[name = tensor("mean_7_cast_fp16")]; + tensor sub_7_cast_fp16 = sub(x = x_23_cast_fp16, y = mean_7_cast_fp16)[name = tensor("sub_7_cast_fp16")]; + tensor square_5_cast_fp16 = square(x = sub_7_cast_fp16)[name = tensor("square_5_cast_fp16")]; + tensor reduce_mean_11_axes_0 = const()[name = tensor("reduce_mean_11_axes_0"), val = tensor([-1])]; + tensor reduce_mean_11_keep_dims_0 = const()[name = tensor("reduce_mean_11_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_11_cast_fp16 = reduce_mean(axes = reduce_mean_11_axes_0, keep_dims = reduce_mean_11_keep_dims_0, x = square_5_cast_fp16)[name = tensor("reduce_mean_11_cast_fp16")]; + tensor var_269_to_fp16 = const()[name = tensor("op_269_to_fp16"), val = tensor(0x1.1p-20)]; + tensor var_270_cast_fp16 = add(x = reduce_mean_11_cast_fp16, y = var_269_to_fp16)[name = tensor("op_270_cast_fp16")]; + tensor var_271_cast_fp16 = sqrt(x = var_270_cast_fp16)[name = tensor("op_271_cast_fp16")]; + tensor x_25_cast_fp16 = real_div(x = sub_7_cast_fp16, y = var_271_cast_fp16)[name = tensor("x_25_cast_fp16")]; + tensor flow_net_res_blocks_3_in_ln_weight_to_fp16 = const()[name = tensor("flow_net_res_blocks_3_in_ln_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(11602368)))]; + tensor var_273_cast_fp16 = mul(x = x_25_cast_fp16, y = flow_net_res_blocks_3_in_ln_weight_to_fp16)[name = tensor("op_273_cast_fp16")]; + tensor flow_net_res_blocks_3_in_ln_bias_to_fp16 = const()[name = tensor("flow_net_res_blocks_3_in_ln_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(11603456)))]; + tensor x_27_cast_fp16 = add(x = var_273_cast_fp16, y = flow_net_res_blocks_3_in_ln_bias_to_fp16)[name = tensor("x_27_cast_fp16")]; + tensor var_275_promoted_to_fp16 = const()[name = tensor("op_275_promoted_to_fp16"), val = tensor(0x1p+0)]; + tensor var_276_cast_fp16 = add(x = var_259_cast_fp16_1, y = var_275_promoted_to_fp16)[name = tensor("op_276_cast_fp16")]; + tensor var_277_cast_fp16 = mul(x = x_27_cast_fp16, y = var_276_cast_fp16)[name = tensor("op_277_cast_fp16")]; + tensor input_41_cast_fp16 = add(x = var_277_cast_fp16, y = var_259_cast_fp16_0)[name = tensor("input_41_cast_fp16")]; + tensor flow_net_res_blocks_3_mlp_0_weight_to_fp16 = const()[name = tensor("flow_net_res_blocks_3_mlp_0_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(11604544)))]; + tensor flow_net_res_blocks_3_mlp_0_bias_to_fp16 = const()[name = tensor("flow_net_res_blocks_3_mlp_0_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(12128896)))]; + tensor linear_16_cast_fp16 = linear(bias = flow_net_res_blocks_3_mlp_0_bias_to_fp16, weight = flow_net_res_blocks_3_mlp_0_weight_to_fp16, x = input_41_cast_fp16)[name = tensor("linear_16_cast_fp16")]; + tensor input_45_cast_fp16 = silu(x = linear_16_cast_fp16)[name = tensor("input_45_cast_fp16")]; + tensor flow_net_res_blocks_3_mlp_2_weight_to_fp16 = const()[name = tensor("flow_net_res_blocks_3_mlp_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(12129984)))]; + tensor flow_net_res_blocks_3_mlp_2_bias_to_fp16 = const()[name = tensor("flow_net_res_blocks_3_mlp_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(12654336)))]; + tensor linear_17_cast_fp16 = linear(bias = flow_net_res_blocks_3_mlp_2_bias_to_fp16, weight = flow_net_res_blocks_3_mlp_2_weight_to_fp16, x = input_45_cast_fp16)[name = tensor("linear_17_cast_fp16")]; + tensor var_288_cast_fp16 = mul(x = var_259_cast_fp16_2, y = linear_17_cast_fp16)[name = tensor("op_288_cast_fp16")]; + tensor x_29_cast_fp16 = add(x = x_23_cast_fp16, y = var_288_cast_fp16)[name = tensor("x_29_cast_fp16")]; + tensor flow_net_res_blocks_4_adaLN_modulation_1_weight_to_fp16 = const()[name = tensor("flow_net_res_blocks_4_adaLN_modulation_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(12655424)))]; + tensor flow_net_res_blocks_4_adaLN_modulation_1_bias_to_fp16 = const()[name = tensor("flow_net_res_blocks_4_adaLN_modulation_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(14228352)))]; + tensor linear_18_cast_fp16 = linear(bias = flow_net_res_blocks_4_adaLN_modulation_1_bias_to_fp16, weight = flow_net_res_blocks_4_adaLN_modulation_1_weight_to_fp16, x = input_15_cast_fp16)[name = tensor("linear_18_cast_fp16")]; + tensor var_298_split_sizes_0 = const()[name = tensor("op_298_split_sizes_0"), val = tensor([512, 512, 512])]; + tensor var_298_axis_0 = const()[name = tensor("op_298_axis_0"), val = tensor(-1)]; + tensor var_298_cast_fp16_0, tensor var_298_cast_fp16_1, tensor var_298_cast_fp16_2 = split(axis = var_298_axis_0, split_sizes = var_298_split_sizes_0, x = linear_18_cast_fp16)[name = tensor("op_298_cast_fp16")]; + tensor mean_9_axes_0 = const()[name = tensor("mean_9_axes_0"), val = tensor([-1])]; + tensor mean_9_keep_dims_0 = const()[name = tensor("mean_9_keep_dims_0"), val = tensor(true)]; + tensor mean_9_cast_fp16 = reduce_mean(axes = mean_9_axes_0, keep_dims = mean_9_keep_dims_0, x = x_29_cast_fp16)[name = tensor("mean_9_cast_fp16")]; + tensor sub_8_cast_fp16 = sub(x = x_29_cast_fp16, y = mean_9_cast_fp16)[name = tensor("sub_8_cast_fp16")]; + tensor square_6_cast_fp16 = square(x = sub_8_cast_fp16)[name = tensor("square_6_cast_fp16")]; + tensor reduce_mean_13_axes_0 = const()[name = tensor("reduce_mean_13_axes_0"), val = tensor([-1])]; + tensor reduce_mean_13_keep_dims_0 = const()[name = tensor("reduce_mean_13_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_13_cast_fp16 = reduce_mean(axes = reduce_mean_13_axes_0, keep_dims = reduce_mean_13_keep_dims_0, x = square_6_cast_fp16)[name = tensor("reduce_mean_13_cast_fp16")]; + tensor var_308_to_fp16 = const()[name = tensor("op_308_to_fp16"), val = tensor(0x1.1p-20)]; + tensor var_309_cast_fp16 = add(x = reduce_mean_13_cast_fp16, y = var_308_to_fp16)[name = tensor("op_309_cast_fp16")]; + tensor var_310_cast_fp16 = sqrt(x = var_309_cast_fp16)[name = tensor("op_310_cast_fp16")]; + tensor x_31_cast_fp16 = real_div(x = sub_8_cast_fp16, y = var_310_cast_fp16)[name = tensor("x_31_cast_fp16")]; + tensor flow_net_res_blocks_4_in_ln_weight_to_fp16 = const()[name = tensor("flow_net_res_blocks_4_in_ln_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(14231488)))]; + tensor var_312_cast_fp16 = mul(x = x_31_cast_fp16, y = flow_net_res_blocks_4_in_ln_weight_to_fp16)[name = tensor("op_312_cast_fp16")]; + tensor flow_net_res_blocks_4_in_ln_bias_to_fp16 = const()[name = tensor("flow_net_res_blocks_4_in_ln_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(14232576)))]; + tensor x_33_cast_fp16 = add(x = var_312_cast_fp16, y = flow_net_res_blocks_4_in_ln_bias_to_fp16)[name = tensor("x_33_cast_fp16")]; + tensor var_314_promoted_to_fp16 = const()[name = tensor("op_314_promoted_to_fp16"), val = tensor(0x1p+0)]; + tensor var_315_cast_fp16 = add(x = var_298_cast_fp16_1, y = var_314_promoted_to_fp16)[name = tensor("op_315_cast_fp16")]; + tensor var_316_cast_fp16 = mul(x = x_33_cast_fp16, y = var_315_cast_fp16)[name = tensor("op_316_cast_fp16")]; + tensor input_49_cast_fp16 = add(x = var_316_cast_fp16, y = var_298_cast_fp16_0)[name = tensor("input_49_cast_fp16")]; + tensor flow_net_res_blocks_4_mlp_0_weight_to_fp16 = const()[name = tensor("flow_net_res_blocks_4_mlp_0_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(14233664)))]; + tensor flow_net_res_blocks_4_mlp_0_bias_to_fp16 = const()[name = tensor("flow_net_res_blocks_4_mlp_0_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(14758016)))]; + tensor linear_19_cast_fp16 = linear(bias = flow_net_res_blocks_4_mlp_0_bias_to_fp16, weight = flow_net_res_blocks_4_mlp_0_weight_to_fp16, x = input_49_cast_fp16)[name = tensor("linear_19_cast_fp16")]; + tensor input_53_cast_fp16 = silu(x = linear_19_cast_fp16)[name = tensor("input_53_cast_fp16")]; + tensor flow_net_res_blocks_4_mlp_2_weight_to_fp16 = const()[name = tensor("flow_net_res_blocks_4_mlp_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(14759104)))]; + tensor flow_net_res_blocks_4_mlp_2_bias_to_fp16 = const()[name = tensor("flow_net_res_blocks_4_mlp_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15283456)))]; + tensor linear_20_cast_fp16 = linear(bias = flow_net_res_blocks_4_mlp_2_bias_to_fp16, weight = flow_net_res_blocks_4_mlp_2_weight_to_fp16, x = input_53_cast_fp16)[name = tensor("linear_20_cast_fp16")]; + tensor var_327_cast_fp16 = mul(x = var_298_cast_fp16_2, y = linear_20_cast_fp16)[name = tensor("op_327_cast_fp16")]; + tensor x_35_cast_fp16 = add(x = x_29_cast_fp16, y = var_327_cast_fp16)[name = tensor("x_35_cast_fp16")]; + tensor flow_net_res_blocks_5_adaLN_modulation_1_weight_to_fp16 = const()[name = tensor("flow_net_res_blocks_5_adaLN_modulation_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15284544)))]; + tensor flow_net_res_blocks_5_adaLN_modulation_1_bias_to_fp16 = const()[name = tensor("flow_net_res_blocks_5_adaLN_modulation_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(16857472)))]; + tensor linear_21_cast_fp16 = linear(bias = flow_net_res_blocks_5_adaLN_modulation_1_bias_to_fp16, weight = flow_net_res_blocks_5_adaLN_modulation_1_weight_to_fp16, x = input_15_cast_fp16)[name = tensor("linear_21_cast_fp16")]; + tensor var_337_split_sizes_0 = const()[name = tensor("op_337_split_sizes_0"), val = tensor([512, 512, 512])]; + tensor var_337_axis_0 = const()[name = tensor("op_337_axis_0"), val = tensor(-1)]; + tensor var_337_cast_fp16_0, tensor var_337_cast_fp16_1, tensor var_337_cast_fp16_2 = split(axis = var_337_axis_0, split_sizes = var_337_split_sizes_0, x = linear_21_cast_fp16)[name = tensor("op_337_cast_fp16")]; + tensor mean_11_axes_0 = const()[name = tensor("mean_11_axes_0"), val = tensor([-1])]; + tensor mean_11_keep_dims_0 = const()[name = tensor("mean_11_keep_dims_0"), val = tensor(true)]; + tensor mean_11_cast_fp16 = reduce_mean(axes = mean_11_axes_0, keep_dims = mean_11_keep_dims_0, x = x_35_cast_fp16)[name = tensor("mean_11_cast_fp16")]; + tensor sub_9_cast_fp16 = sub(x = x_35_cast_fp16, y = mean_11_cast_fp16)[name = tensor("sub_9_cast_fp16")]; + tensor square_7_cast_fp16 = square(x = sub_9_cast_fp16)[name = tensor("square_7_cast_fp16")]; + tensor reduce_mean_15_axes_0 = const()[name = tensor("reduce_mean_15_axes_0"), val = tensor([-1])]; + tensor reduce_mean_15_keep_dims_0 = const()[name = tensor("reduce_mean_15_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_15_cast_fp16 = reduce_mean(axes = reduce_mean_15_axes_0, keep_dims = reduce_mean_15_keep_dims_0, x = square_7_cast_fp16)[name = tensor("reduce_mean_15_cast_fp16")]; + tensor var_347_to_fp16 = const()[name = tensor("op_347_to_fp16"), val = tensor(0x1.1p-20)]; + tensor var_348_cast_fp16 = add(x = reduce_mean_15_cast_fp16, y = var_347_to_fp16)[name = tensor("op_348_cast_fp16")]; + tensor var_349_cast_fp16 = sqrt(x = var_348_cast_fp16)[name = tensor("op_349_cast_fp16")]; + tensor x_37_cast_fp16 = real_div(x = sub_9_cast_fp16, y = var_349_cast_fp16)[name = tensor("x_37_cast_fp16")]; + tensor flow_net_res_blocks_5_in_ln_weight_to_fp16 = const()[name = tensor("flow_net_res_blocks_5_in_ln_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(16860608)))]; + tensor var_351_cast_fp16 = mul(x = x_37_cast_fp16, y = flow_net_res_blocks_5_in_ln_weight_to_fp16)[name = tensor("op_351_cast_fp16")]; + tensor flow_net_res_blocks_5_in_ln_bias_to_fp16 = const()[name = tensor("flow_net_res_blocks_5_in_ln_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(16861696)))]; + tensor x_39_cast_fp16 = add(x = var_351_cast_fp16, y = flow_net_res_blocks_5_in_ln_bias_to_fp16)[name = tensor("x_39_cast_fp16")]; + tensor var_353_promoted_to_fp16 = const()[name = tensor("op_353_promoted_to_fp16"), val = tensor(0x1p+0)]; + tensor var_354_cast_fp16 = add(x = var_337_cast_fp16_1, y = var_353_promoted_to_fp16)[name = tensor("op_354_cast_fp16")]; + tensor var_355_cast_fp16 = mul(x = x_39_cast_fp16, y = var_354_cast_fp16)[name = tensor("op_355_cast_fp16")]; + tensor input_57_cast_fp16 = add(x = var_355_cast_fp16, y = var_337_cast_fp16_0)[name = tensor("input_57_cast_fp16")]; + tensor flow_net_res_blocks_5_mlp_0_weight_to_fp16 = const()[name = tensor("flow_net_res_blocks_5_mlp_0_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(16862784)))]; + tensor flow_net_res_blocks_5_mlp_0_bias_to_fp16 = const()[name = tensor("flow_net_res_blocks_5_mlp_0_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(17387136)))]; + tensor linear_22_cast_fp16 = linear(bias = flow_net_res_blocks_5_mlp_0_bias_to_fp16, weight = flow_net_res_blocks_5_mlp_0_weight_to_fp16, x = input_57_cast_fp16)[name = tensor("linear_22_cast_fp16")]; + tensor input_61_cast_fp16 = silu(x = linear_22_cast_fp16)[name = tensor("input_61_cast_fp16")]; + tensor flow_net_res_blocks_5_mlp_2_weight_to_fp16 = const()[name = tensor("flow_net_res_blocks_5_mlp_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(17388224)))]; + tensor flow_net_res_blocks_5_mlp_2_bias_to_fp16 = const()[name = tensor("flow_net_res_blocks_5_mlp_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(17912576)))]; + tensor linear_23_cast_fp16 = linear(bias = flow_net_res_blocks_5_mlp_2_bias_to_fp16, weight = flow_net_res_blocks_5_mlp_2_weight_to_fp16, x = input_61_cast_fp16)[name = tensor("linear_23_cast_fp16")]; + tensor var_366_cast_fp16 = mul(x = var_337_cast_fp16_2, y = linear_23_cast_fp16)[name = tensor("op_366_cast_fp16")]; + tensor x_41_cast_fp16 = add(x = x_35_cast_fp16, y = var_366_cast_fp16)[name = tensor("x_41_cast_fp16")]; + tensor flow_net_final_layer_adaLN_modulation_1_weight_to_fp16 = const()[name = tensor("flow_net_final_layer_adaLN_modulation_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(17913664)))]; + tensor flow_net_final_layer_adaLN_modulation_1_bias_to_fp16 = const()[name = tensor("flow_net_final_layer_adaLN_modulation_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(18962304)))]; + tensor linear_24_cast_fp16 = linear(bias = flow_net_final_layer_adaLN_modulation_1_bias_to_fp16, weight = flow_net_final_layer_adaLN_modulation_1_weight_to_fp16, x = input_15_cast_fp16)[name = tensor("linear_24_cast_fp16")]; + tensor var_375_split_sizes_0 = const()[name = tensor("op_375_split_sizes_0"), val = tensor([512, 512])]; + tensor var_375_axis_0 = const()[name = tensor("op_375_axis_0"), val = tensor(-1)]; + tensor var_375_cast_fp16_0, tensor var_375_cast_fp16_1 = split(axis = var_375_axis_0, split_sizes = var_375_split_sizes_0, x = linear_24_cast_fp16)[name = tensor("op_375_cast_fp16")]; + tensor mean_13_axes_0 = const()[name = tensor("mean_13_axes_0"), val = tensor([-1])]; + tensor mean_13_keep_dims_0 = const()[name = tensor("mean_13_keep_dims_0"), val = tensor(true)]; + tensor mean_13_cast_fp16 = reduce_mean(axes = mean_13_axes_0, keep_dims = mean_13_keep_dims_0, x = x_41_cast_fp16)[name = tensor("mean_13_cast_fp16")]; + tensor sub_10_cast_fp16 = sub(x = x_41_cast_fp16, y = mean_13_cast_fp16)[name = tensor("sub_10_cast_fp16")]; + tensor square_8_cast_fp16 = square(x = sub_10_cast_fp16)[name = tensor("square_8_cast_fp16")]; + tensor reduce_mean_17_axes_0 = const()[name = tensor("reduce_mean_17_axes_0"), val = tensor([-1])]; + tensor reduce_mean_17_keep_dims_0 = const()[name = tensor("reduce_mean_17_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_17_cast_fp16 = reduce_mean(axes = reduce_mean_17_axes_0, keep_dims = reduce_mean_17_keep_dims_0, x = square_8_cast_fp16)[name = tensor("reduce_mean_17_cast_fp16")]; + tensor var_382_to_fp16 = const()[name = tensor("op_382_to_fp16"), val = tensor(0x1.1p-20)]; + tensor var_383_cast_fp16 = add(x = reduce_mean_17_cast_fp16, y = var_382_to_fp16)[name = tensor("op_383_cast_fp16")]; + tensor var_384_cast_fp16 = sqrt(x = var_383_cast_fp16)[name = tensor("op_384_cast_fp16")]; + tensor x_43_cast_fp16 = real_div(x = sub_10_cast_fp16, y = var_384_cast_fp16)[name = tensor("x_43_cast_fp16")]; + tensor var_386_promoted_to_fp16 = const()[name = tensor("op_386_promoted_to_fp16"), val = tensor(0x1p+0)]; + tensor var_387_cast_fp16 = add(x = var_375_cast_fp16_1, y = var_386_promoted_to_fp16)[name = tensor("op_387_cast_fp16")]; + tensor var_388_cast_fp16 = mul(x = x_43_cast_fp16, y = var_387_cast_fp16)[name = tensor("op_388_cast_fp16")]; + tensor input_65_cast_fp16 = add(x = var_388_cast_fp16, y = var_375_cast_fp16_0)[name = tensor("input_65_cast_fp16")]; + tensor flow_net_final_layer_linear_weight_to_fp16 = const()[name = tensor("flow_net_final_layer_linear_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(18964416)))]; + tensor flow_net_final_layer_linear_bias_to_fp16 = const()[name = tensor("flow_net_final_layer_linear_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(18997248)))]; + tensor linear_25_cast_fp16 = linear(bias = flow_net_final_layer_linear_bias_to_fp16, weight = flow_net_final_layer_linear_weight_to_fp16, x = input_65_cast_fp16)[name = tensor("linear_25_cast_fp16")]; + tensor var_399_to_fp16 = const()[name = tensor("op_399_to_fp16"), val = tensor(0x1p-3)]; + tensor var_400_cast_fp16 = mul(x = linear_25_cast_fp16, y = var_399_to_fp16)[name = tensor("op_400_cast_fp16")]; + tensor input_67_cast_fp16 = add(x = latent_init_to_fp16, y = var_400_cast_fp16)[name = tensor("input_67_cast_fp16")]; + tensor linear_26_cast_fp16 = linear(bias = flow_net_input_proj_bias_to_fp16, weight = flow_net_input_proj_weight_to_fp16, x = input_67_cast_fp16)[name = tensor("linear_26_cast_fp16")]; + tensor input_71_to_fp16 = const()[name = tensor("input_71_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(18997376)))]; + tensor input_73_cast_fp16 = silu(x = input_71_to_fp16)[name = tensor("input_73_cast_fp16")]; + tensor linear_28_cast_fp16 = linear(bias = flow_net_time_embed_0_mlp_2_bias_to_fp16, weight = flow_net_time_embed_0_mlp_2_weight_to_fp16, x = input_73_cast_fp16)[name = tensor("linear_28_cast_fp16")]; + tensor reduce_mean_18_axes_0 = const()[name = tensor("reduce_mean_18_axes_0"), val = tensor([-1])]; + tensor reduce_mean_18_keep_dims_0 = const()[name = tensor("reduce_mean_18_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_18_cast_fp16 = reduce_mean(axes = reduce_mean_18_axes_0, keep_dims = reduce_mean_18_keep_dims_0, x = linear_28_cast_fp16)[name = tensor("reduce_mean_18_cast_fp16")]; + tensor sub_11_cast_fp16 = sub(x = linear_28_cast_fp16, y = reduce_mean_18_cast_fp16)[name = tensor("sub_11_cast_fp16")]; + tensor square_9_cast_fp16 = square(x = sub_11_cast_fp16)[name = tensor("square_9_cast_fp16")]; + tensor reduce_mean_19_axes_0 = const()[name = tensor("reduce_mean_19_axes_0"), val = tensor([-1])]; + tensor reduce_mean_19_keep_dims_0 = const()[name = tensor("reduce_mean_19_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_19_cast_fp16 = reduce_mean(axes = reduce_mean_19_axes_0, keep_dims = reduce_mean_19_keep_dims_0, x = square_9_cast_fp16)[name = tensor("reduce_mean_19_cast_fp16")]; + tensor real_div_2_to_fp16 = const()[name = tensor("real_div_2_to_fp16"), val = tensor(0x1.008p+0)]; + tensor mul_2_cast_fp16 = mul(x = reduce_mean_19_cast_fp16, y = real_div_2_to_fp16)[name = tensor("mul_2_cast_fp16")]; + tensor var_466_to_fp16 = const()[name = tensor("op_466_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_19_cast_fp16 = add(x = mul_2_cast_fp16, y = var_466_to_fp16)[name = tensor("var_19_cast_fp16")]; + tensor var_469_epsilon_0 = const()[name = tensor("op_469_epsilon_0"), val = tensor(0x1.197998p-40)]; + tensor var_469_cast_fp16 = rsqrt(epsilon = var_469_epsilon_0, x = var_19_cast_fp16)[name = tensor("op_469_cast_fp16")]; + tensor var_470_cast_fp16 = mul(x = const_3_to_fp16, y = var_469_cast_fp16)[name = tensor("op_470_cast_fp16")]; + tensor var_471_cast_fp16 = mul(x = linear_28_cast_fp16, y = var_470_cast_fp16)[name = tensor("op_471_cast_fp16")]; + tensor input_77_to_fp16 = const()[name = tensor("input_77_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(18998464)))]; + tensor input_79_cast_fp16 = silu(x = input_77_to_fp16)[name = tensor("input_79_cast_fp16")]; + tensor linear_30_cast_fp16 = linear(bias = flow_net_time_embed_1_mlp_2_bias_to_fp16, weight = flow_net_time_embed_1_mlp_2_weight_to_fp16, x = input_79_cast_fp16)[name = tensor("linear_30_cast_fp16")]; + tensor reduce_mean_20_axes_0 = const()[name = tensor("reduce_mean_20_axes_0"), val = tensor([-1])]; + tensor reduce_mean_20_keep_dims_0 = const()[name = tensor("reduce_mean_20_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_20_cast_fp16 = reduce_mean(axes = reduce_mean_20_axes_0, keep_dims = reduce_mean_20_keep_dims_0, x = linear_30_cast_fp16)[name = tensor("reduce_mean_20_cast_fp16")]; + tensor sub_13_cast_fp16 = sub(x = linear_30_cast_fp16, y = reduce_mean_20_cast_fp16)[name = tensor("sub_13_cast_fp16")]; + tensor square_10_cast_fp16 = square(x = sub_13_cast_fp16)[name = tensor("square_10_cast_fp16")]; + tensor reduce_mean_21_axes_0 = const()[name = tensor("reduce_mean_21_axes_0"), val = tensor([-1])]; + tensor reduce_mean_21_keep_dims_0 = const()[name = tensor("reduce_mean_21_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_21_cast_fp16 = reduce_mean(axes = reduce_mean_21_axes_0, keep_dims = reduce_mean_21_keep_dims_0, x = square_10_cast_fp16)[name = tensor("reduce_mean_21_cast_fp16")]; + tensor real_div_3_to_fp16 = const()[name = tensor("real_div_3_to_fp16"), val = tensor(0x1.008p+0)]; + tensor mul_3_cast_fp16 = mul(x = reduce_mean_21_cast_fp16, y = real_div_3_to_fp16)[name = tensor("mul_3_cast_fp16")]; + tensor var_503_to_fp16 = const()[name = tensor("op_503_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_21_cast_fp16 = add(x = mul_3_cast_fp16, y = var_503_to_fp16)[name = tensor("var_21_cast_fp16")]; + tensor var_506_epsilon_0 = const()[name = tensor("op_506_epsilon_0"), val = tensor(0x1.197998p-40)]; + tensor var_506_cast_fp16 = rsqrt(epsilon = var_506_epsilon_0, x = var_21_cast_fp16)[name = tensor("op_506_cast_fp16")]; + tensor var_507_cast_fp16 = mul(x = const_5_to_fp16, y = var_506_cast_fp16)[name = tensor("op_507_cast_fp16")]; + tensor var_508_cast_fp16 = mul(x = linear_30_cast_fp16, y = var_507_cast_fp16)[name = tensor("op_508_cast_fp16")]; + tensor var_520_cast_fp16 = add(x = var_471_cast_fp16, y = var_508_cast_fp16)[name = tensor("op_520_cast_fp16")]; + tensor _inversed_t_combined_3_y_0_to_fp16 = const()[name = tensor("_inversed_t_combined_3_y_0_to_fp16"), val = tensor(0x1p-1)]; + tensor _inversed_t_combined_3_cast_fp16 = mul(x = var_520_cast_fp16, y = _inversed_t_combined_3_y_0_to_fp16)[name = tensor("_inversed_t_combined_3_cast_fp16")]; + tensor input_81_cast_fp16 = add(x = _inversed_t_combined_3_cast_fp16, y = linear_5_cast_fp16)[name = tensor("input_81_cast_fp16")]; + tensor input_83_cast_fp16 = silu(x = input_81_cast_fp16)[name = tensor("input_83_cast_fp16")]; + tensor linear_32_cast_fp16 = linear(bias = flow_net_res_blocks_0_adaLN_modulation_1_bias_to_fp16, weight = flow_net_res_blocks_0_adaLN_modulation_1_weight_to_fp16, x = input_83_cast_fp16)[name = tensor("linear_32_cast_fp16")]; + tensor var_535_split_sizes_0 = const()[name = tensor("op_535_split_sizes_0"), val = tensor([512, 512, 512])]; + tensor var_535_axis_0 = const()[name = tensor("op_535_axis_0"), val = tensor(-1)]; + tensor var_535_cast_fp16_0, tensor var_535_cast_fp16_1, tensor var_535_cast_fp16_2 = split(axis = var_535_axis_0, split_sizes = var_535_split_sizes_0, x = linear_32_cast_fp16)[name = tensor("op_535_cast_fp16")]; + tensor mean_15_axes_0 = const()[name = tensor("mean_15_axes_0"), val = tensor([-1])]; + tensor mean_15_keep_dims_0 = const()[name = tensor("mean_15_keep_dims_0"), val = tensor(true)]; + tensor mean_15_cast_fp16 = reduce_mean(axes = mean_15_axes_0, keep_dims = mean_15_keep_dims_0, x = linear_26_cast_fp16)[name = tensor("mean_15_cast_fp16")]; + tensor sub_15_cast_fp16 = sub(x = linear_26_cast_fp16, y = mean_15_cast_fp16)[name = tensor("sub_15_cast_fp16")]; + tensor square_11_cast_fp16 = square(x = sub_15_cast_fp16)[name = tensor("square_11_cast_fp16")]; + tensor reduce_mean_23_axes_0 = const()[name = tensor("reduce_mean_23_axes_0"), val = tensor([-1])]; + tensor reduce_mean_23_keep_dims_0 = const()[name = tensor("reduce_mean_23_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_23_cast_fp16 = reduce_mean(axes = reduce_mean_23_axes_0, keep_dims = reduce_mean_23_keep_dims_0, x = square_11_cast_fp16)[name = tensor("reduce_mean_23_cast_fp16")]; + tensor var_545_to_fp16 = const()[name = tensor("op_545_to_fp16"), val = tensor(0x1.1p-20)]; + tensor var_546_cast_fp16 = add(x = reduce_mean_23_cast_fp16, y = var_545_to_fp16)[name = tensor("op_546_cast_fp16")]; + tensor var_547_cast_fp16 = sqrt(x = var_546_cast_fp16)[name = tensor("op_547_cast_fp16")]; + tensor x_51_cast_fp16 = real_div(x = sub_15_cast_fp16, y = var_547_cast_fp16)[name = tensor("x_51_cast_fp16")]; + tensor var_549_cast_fp16 = mul(x = x_51_cast_fp16, y = flow_net_res_blocks_0_in_ln_weight_to_fp16)[name = tensor("op_549_cast_fp16")]; + tensor x_53_cast_fp16 = add(x = var_549_cast_fp16, y = flow_net_res_blocks_0_in_ln_bias_to_fp16)[name = tensor("x_53_cast_fp16")]; + tensor var_551_promoted_to_fp16 = const()[name = tensor("op_551_promoted_to_fp16"), val = tensor(0x1p+0)]; + tensor var_552_cast_fp16 = add(x = var_535_cast_fp16_1, y = var_551_promoted_to_fp16)[name = tensor("op_552_cast_fp16")]; + tensor var_553_cast_fp16 = mul(x = x_53_cast_fp16, y = var_552_cast_fp16)[name = tensor("op_553_cast_fp16")]; + tensor input_85_cast_fp16 = add(x = var_553_cast_fp16, y = var_535_cast_fp16_0)[name = tensor("input_85_cast_fp16")]; + tensor linear_33_cast_fp16 = linear(bias = flow_net_res_blocks_0_mlp_0_bias_to_fp16, weight = flow_net_res_blocks_0_mlp_0_weight_to_fp16, x = input_85_cast_fp16)[name = tensor("linear_33_cast_fp16")]; + tensor input_89_cast_fp16 = silu(x = linear_33_cast_fp16)[name = tensor("input_89_cast_fp16")]; + tensor linear_34_cast_fp16 = linear(bias = flow_net_res_blocks_0_mlp_2_bias_to_fp16, weight = flow_net_res_blocks_0_mlp_2_weight_to_fp16, x = input_89_cast_fp16)[name = tensor("linear_34_cast_fp16")]; + tensor var_564_cast_fp16 = mul(x = var_535_cast_fp16_2, y = linear_34_cast_fp16)[name = tensor("op_564_cast_fp16")]; + tensor x_55_cast_fp16 = add(x = linear_26_cast_fp16, y = var_564_cast_fp16)[name = tensor("x_55_cast_fp16")]; + tensor linear_35_cast_fp16 = linear(bias = flow_net_res_blocks_1_adaLN_modulation_1_bias_to_fp16, weight = flow_net_res_blocks_1_adaLN_modulation_1_weight_to_fp16, x = input_83_cast_fp16)[name = tensor("linear_35_cast_fp16")]; + tensor var_574_split_sizes_0 = const()[name = tensor("op_574_split_sizes_0"), val = tensor([512, 512, 512])]; + tensor var_574_axis_0 = const()[name = tensor("op_574_axis_0"), val = tensor(-1)]; + tensor var_574_cast_fp16_0, tensor var_574_cast_fp16_1, tensor var_574_cast_fp16_2 = split(axis = var_574_axis_0, split_sizes = var_574_split_sizes_0, x = linear_35_cast_fp16)[name = tensor("op_574_cast_fp16")]; + tensor mean_17_axes_0 = const()[name = tensor("mean_17_axes_0"), val = tensor([-1])]; + tensor mean_17_keep_dims_0 = const()[name = tensor("mean_17_keep_dims_0"), val = tensor(true)]; + tensor mean_17_cast_fp16 = reduce_mean(axes = mean_17_axes_0, keep_dims = mean_17_keep_dims_0, x = x_55_cast_fp16)[name = tensor("mean_17_cast_fp16")]; + tensor sub_16_cast_fp16 = sub(x = x_55_cast_fp16, y = mean_17_cast_fp16)[name = tensor("sub_16_cast_fp16")]; + tensor square_12_cast_fp16 = square(x = sub_16_cast_fp16)[name = tensor("square_12_cast_fp16")]; + tensor reduce_mean_25_axes_0 = const()[name = tensor("reduce_mean_25_axes_0"), val = tensor([-1])]; + tensor reduce_mean_25_keep_dims_0 = const()[name = tensor("reduce_mean_25_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_25_cast_fp16 = reduce_mean(axes = reduce_mean_25_axes_0, keep_dims = reduce_mean_25_keep_dims_0, x = square_12_cast_fp16)[name = tensor("reduce_mean_25_cast_fp16")]; + tensor var_584_to_fp16 = const()[name = tensor("op_584_to_fp16"), val = tensor(0x1.1p-20)]; + tensor var_585_cast_fp16 = add(x = reduce_mean_25_cast_fp16, y = var_584_to_fp16)[name = tensor("op_585_cast_fp16")]; + tensor var_586_cast_fp16 = sqrt(x = var_585_cast_fp16)[name = tensor("op_586_cast_fp16")]; + tensor x_57_cast_fp16 = real_div(x = sub_16_cast_fp16, y = var_586_cast_fp16)[name = tensor("x_57_cast_fp16")]; + tensor var_588_cast_fp16 = mul(x = x_57_cast_fp16, y = flow_net_res_blocks_1_in_ln_weight_to_fp16)[name = tensor("op_588_cast_fp16")]; + tensor x_59_cast_fp16 = add(x = var_588_cast_fp16, y = flow_net_res_blocks_1_in_ln_bias_to_fp16)[name = tensor("x_59_cast_fp16")]; + tensor var_590_promoted_to_fp16 = const()[name = tensor("op_590_promoted_to_fp16"), val = tensor(0x1p+0)]; + tensor var_591_cast_fp16 = add(x = var_574_cast_fp16_1, y = var_590_promoted_to_fp16)[name = tensor("op_591_cast_fp16")]; + tensor var_592_cast_fp16 = mul(x = x_59_cast_fp16, y = var_591_cast_fp16)[name = tensor("op_592_cast_fp16")]; + tensor input_93_cast_fp16 = add(x = var_592_cast_fp16, y = var_574_cast_fp16_0)[name = tensor("input_93_cast_fp16")]; + tensor linear_36_cast_fp16 = linear(bias = flow_net_res_blocks_1_mlp_0_bias_to_fp16, weight = flow_net_res_blocks_1_mlp_0_weight_to_fp16, x = input_93_cast_fp16)[name = tensor("linear_36_cast_fp16")]; + tensor input_97_cast_fp16 = silu(x = linear_36_cast_fp16)[name = tensor("input_97_cast_fp16")]; + tensor linear_37_cast_fp16 = linear(bias = flow_net_res_blocks_1_mlp_2_bias_to_fp16, weight = flow_net_res_blocks_1_mlp_2_weight_to_fp16, x = input_97_cast_fp16)[name = tensor("linear_37_cast_fp16")]; + tensor var_603_cast_fp16 = mul(x = var_574_cast_fp16_2, y = linear_37_cast_fp16)[name = tensor("op_603_cast_fp16")]; + tensor x_61_cast_fp16 = add(x = x_55_cast_fp16, y = var_603_cast_fp16)[name = tensor("x_61_cast_fp16")]; + tensor linear_38_cast_fp16 = linear(bias = flow_net_res_blocks_2_adaLN_modulation_1_bias_to_fp16, weight = flow_net_res_blocks_2_adaLN_modulation_1_weight_to_fp16, x = input_83_cast_fp16)[name = tensor("linear_38_cast_fp16")]; + tensor var_613_split_sizes_0 = const()[name = tensor("op_613_split_sizes_0"), val = tensor([512, 512, 512])]; + tensor var_613_axis_0 = const()[name = tensor("op_613_axis_0"), val = tensor(-1)]; + tensor var_613_cast_fp16_0, tensor var_613_cast_fp16_1, tensor var_613_cast_fp16_2 = split(axis = var_613_axis_0, split_sizes = var_613_split_sizes_0, x = linear_38_cast_fp16)[name = tensor("op_613_cast_fp16")]; + tensor mean_19_axes_0 = const()[name = tensor("mean_19_axes_0"), val = tensor([-1])]; + tensor mean_19_keep_dims_0 = const()[name = tensor("mean_19_keep_dims_0"), val = tensor(true)]; + tensor mean_19_cast_fp16 = reduce_mean(axes = mean_19_axes_0, keep_dims = mean_19_keep_dims_0, x = x_61_cast_fp16)[name = tensor("mean_19_cast_fp16")]; + tensor sub_17_cast_fp16 = sub(x = x_61_cast_fp16, y = mean_19_cast_fp16)[name = tensor("sub_17_cast_fp16")]; + tensor square_13_cast_fp16 = square(x = sub_17_cast_fp16)[name = tensor("square_13_cast_fp16")]; + tensor reduce_mean_27_axes_0 = const()[name = tensor("reduce_mean_27_axes_0"), val = tensor([-1])]; + tensor reduce_mean_27_keep_dims_0 = const()[name = tensor("reduce_mean_27_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_27_cast_fp16 = reduce_mean(axes = reduce_mean_27_axes_0, keep_dims = reduce_mean_27_keep_dims_0, x = square_13_cast_fp16)[name = tensor("reduce_mean_27_cast_fp16")]; + tensor var_623_to_fp16 = const()[name = tensor("op_623_to_fp16"), val = tensor(0x1.1p-20)]; + tensor var_624_cast_fp16 = add(x = reduce_mean_27_cast_fp16, y = var_623_to_fp16)[name = tensor("op_624_cast_fp16")]; + tensor var_625_cast_fp16 = sqrt(x = var_624_cast_fp16)[name = tensor("op_625_cast_fp16")]; + tensor x_63_cast_fp16 = real_div(x = sub_17_cast_fp16, y = var_625_cast_fp16)[name = tensor("x_63_cast_fp16")]; + tensor var_627_cast_fp16 = mul(x = x_63_cast_fp16, y = flow_net_res_blocks_2_in_ln_weight_to_fp16)[name = tensor("op_627_cast_fp16")]; + tensor x_65_cast_fp16 = add(x = var_627_cast_fp16, y = flow_net_res_blocks_2_in_ln_bias_to_fp16)[name = tensor("x_65_cast_fp16")]; + tensor var_629_promoted_to_fp16 = const()[name = tensor("op_629_promoted_to_fp16"), val = tensor(0x1p+0)]; + tensor var_630_cast_fp16 = add(x = var_613_cast_fp16_1, y = var_629_promoted_to_fp16)[name = tensor("op_630_cast_fp16")]; + tensor var_631_cast_fp16 = mul(x = x_65_cast_fp16, y = var_630_cast_fp16)[name = tensor("op_631_cast_fp16")]; + tensor input_101_cast_fp16 = add(x = var_631_cast_fp16, y = var_613_cast_fp16_0)[name = tensor("input_101_cast_fp16")]; + tensor linear_39_cast_fp16 = linear(bias = flow_net_res_blocks_2_mlp_0_bias_to_fp16, weight = flow_net_res_blocks_2_mlp_0_weight_to_fp16, x = input_101_cast_fp16)[name = tensor("linear_39_cast_fp16")]; + tensor input_105_cast_fp16 = silu(x = linear_39_cast_fp16)[name = tensor("input_105_cast_fp16")]; + tensor linear_40_cast_fp16 = linear(bias = flow_net_res_blocks_2_mlp_2_bias_to_fp16, weight = flow_net_res_blocks_2_mlp_2_weight_to_fp16, x = input_105_cast_fp16)[name = tensor("linear_40_cast_fp16")]; + tensor var_642_cast_fp16 = mul(x = var_613_cast_fp16_2, y = linear_40_cast_fp16)[name = tensor("op_642_cast_fp16")]; + tensor x_67_cast_fp16 = add(x = x_61_cast_fp16, y = var_642_cast_fp16)[name = tensor("x_67_cast_fp16")]; + tensor linear_41_cast_fp16 = linear(bias = flow_net_res_blocks_3_adaLN_modulation_1_bias_to_fp16, weight = flow_net_res_blocks_3_adaLN_modulation_1_weight_to_fp16, x = input_83_cast_fp16)[name = tensor("linear_41_cast_fp16")]; + tensor var_652_split_sizes_0 = const()[name = tensor("op_652_split_sizes_0"), val = tensor([512, 512, 512])]; + tensor var_652_axis_0 = const()[name = tensor("op_652_axis_0"), val = tensor(-1)]; + tensor var_652_cast_fp16_0, tensor var_652_cast_fp16_1, tensor var_652_cast_fp16_2 = split(axis = var_652_axis_0, split_sizes = var_652_split_sizes_0, x = linear_41_cast_fp16)[name = tensor("op_652_cast_fp16")]; + tensor mean_21_axes_0 = const()[name = tensor("mean_21_axes_0"), val = tensor([-1])]; + tensor mean_21_keep_dims_0 = const()[name = tensor("mean_21_keep_dims_0"), val = tensor(true)]; + tensor mean_21_cast_fp16 = reduce_mean(axes = mean_21_axes_0, keep_dims = mean_21_keep_dims_0, x = x_67_cast_fp16)[name = tensor("mean_21_cast_fp16")]; + tensor sub_18_cast_fp16 = sub(x = x_67_cast_fp16, y = mean_21_cast_fp16)[name = tensor("sub_18_cast_fp16")]; + tensor square_14_cast_fp16 = square(x = sub_18_cast_fp16)[name = tensor("square_14_cast_fp16")]; + tensor reduce_mean_29_axes_0 = const()[name = tensor("reduce_mean_29_axes_0"), val = tensor([-1])]; + tensor reduce_mean_29_keep_dims_0 = const()[name = tensor("reduce_mean_29_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_29_cast_fp16 = reduce_mean(axes = reduce_mean_29_axes_0, keep_dims = reduce_mean_29_keep_dims_0, x = square_14_cast_fp16)[name = tensor("reduce_mean_29_cast_fp16")]; + tensor var_662_to_fp16 = const()[name = tensor("op_662_to_fp16"), val = tensor(0x1.1p-20)]; + tensor var_663_cast_fp16 = add(x = reduce_mean_29_cast_fp16, y = var_662_to_fp16)[name = tensor("op_663_cast_fp16")]; + tensor var_664_cast_fp16 = sqrt(x = var_663_cast_fp16)[name = tensor("op_664_cast_fp16")]; + tensor x_69_cast_fp16 = real_div(x = sub_18_cast_fp16, y = var_664_cast_fp16)[name = tensor("x_69_cast_fp16")]; + tensor var_666_cast_fp16 = mul(x = x_69_cast_fp16, y = flow_net_res_blocks_3_in_ln_weight_to_fp16)[name = tensor("op_666_cast_fp16")]; + tensor x_71_cast_fp16 = add(x = var_666_cast_fp16, y = flow_net_res_blocks_3_in_ln_bias_to_fp16)[name = tensor("x_71_cast_fp16")]; + tensor var_668_promoted_to_fp16 = const()[name = tensor("op_668_promoted_to_fp16"), val = tensor(0x1p+0)]; + tensor var_669_cast_fp16 = add(x = var_652_cast_fp16_1, y = var_668_promoted_to_fp16)[name = tensor("op_669_cast_fp16")]; + tensor var_670_cast_fp16 = mul(x = x_71_cast_fp16, y = var_669_cast_fp16)[name = tensor("op_670_cast_fp16")]; + tensor input_109_cast_fp16 = add(x = var_670_cast_fp16, y = var_652_cast_fp16_0)[name = tensor("input_109_cast_fp16")]; + tensor linear_42_cast_fp16 = linear(bias = flow_net_res_blocks_3_mlp_0_bias_to_fp16, weight = flow_net_res_blocks_3_mlp_0_weight_to_fp16, x = input_109_cast_fp16)[name = tensor("linear_42_cast_fp16")]; + tensor input_113_cast_fp16 = silu(x = linear_42_cast_fp16)[name = tensor("input_113_cast_fp16")]; + tensor linear_43_cast_fp16 = linear(bias = flow_net_res_blocks_3_mlp_2_bias_to_fp16, weight = flow_net_res_blocks_3_mlp_2_weight_to_fp16, x = input_113_cast_fp16)[name = tensor("linear_43_cast_fp16")]; + tensor var_681_cast_fp16 = mul(x = var_652_cast_fp16_2, y = linear_43_cast_fp16)[name = tensor("op_681_cast_fp16")]; + tensor x_73_cast_fp16 = add(x = x_67_cast_fp16, y = var_681_cast_fp16)[name = tensor("x_73_cast_fp16")]; + tensor linear_44_cast_fp16 = linear(bias = flow_net_res_blocks_4_adaLN_modulation_1_bias_to_fp16, weight = flow_net_res_blocks_4_adaLN_modulation_1_weight_to_fp16, x = input_83_cast_fp16)[name = tensor("linear_44_cast_fp16")]; + tensor var_691_split_sizes_0 = const()[name = tensor("op_691_split_sizes_0"), val = tensor([512, 512, 512])]; + tensor var_691_axis_0 = const()[name = tensor("op_691_axis_0"), val = tensor(-1)]; + tensor var_691_cast_fp16_0, tensor var_691_cast_fp16_1, tensor var_691_cast_fp16_2 = split(axis = var_691_axis_0, split_sizes = var_691_split_sizes_0, x = linear_44_cast_fp16)[name = tensor("op_691_cast_fp16")]; + tensor mean_23_axes_0 = const()[name = tensor("mean_23_axes_0"), val = tensor([-1])]; + tensor mean_23_keep_dims_0 = const()[name = tensor("mean_23_keep_dims_0"), val = tensor(true)]; + tensor mean_23_cast_fp16 = reduce_mean(axes = mean_23_axes_0, keep_dims = mean_23_keep_dims_0, x = x_73_cast_fp16)[name = tensor("mean_23_cast_fp16")]; + tensor sub_19_cast_fp16 = sub(x = x_73_cast_fp16, y = mean_23_cast_fp16)[name = tensor("sub_19_cast_fp16")]; + tensor square_15_cast_fp16 = square(x = sub_19_cast_fp16)[name = tensor("square_15_cast_fp16")]; + tensor reduce_mean_31_axes_0 = const()[name = tensor("reduce_mean_31_axes_0"), val = tensor([-1])]; + tensor reduce_mean_31_keep_dims_0 = const()[name = tensor("reduce_mean_31_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_31_cast_fp16 = reduce_mean(axes = reduce_mean_31_axes_0, keep_dims = reduce_mean_31_keep_dims_0, x = square_15_cast_fp16)[name = tensor("reduce_mean_31_cast_fp16")]; + tensor var_701_to_fp16 = const()[name = tensor("op_701_to_fp16"), val = tensor(0x1.1p-20)]; + tensor var_702_cast_fp16 = add(x = reduce_mean_31_cast_fp16, y = var_701_to_fp16)[name = tensor("op_702_cast_fp16")]; + tensor var_703_cast_fp16 = sqrt(x = var_702_cast_fp16)[name = tensor("op_703_cast_fp16")]; + tensor x_75_cast_fp16 = real_div(x = sub_19_cast_fp16, y = var_703_cast_fp16)[name = tensor("x_75_cast_fp16")]; + tensor var_705_cast_fp16 = mul(x = x_75_cast_fp16, y = flow_net_res_blocks_4_in_ln_weight_to_fp16)[name = tensor("op_705_cast_fp16")]; + tensor x_77_cast_fp16 = add(x = var_705_cast_fp16, y = flow_net_res_blocks_4_in_ln_bias_to_fp16)[name = tensor("x_77_cast_fp16")]; + tensor var_707_promoted_to_fp16 = const()[name = tensor("op_707_promoted_to_fp16"), val = tensor(0x1p+0)]; + tensor var_708_cast_fp16 = add(x = var_691_cast_fp16_1, y = var_707_promoted_to_fp16)[name = tensor("op_708_cast_fp16")]; + tensor var_709_cast_fp16 = mul(x = x_77_cast_fp16, y = var_708_cast_fp16)[name = tensor("op_709_cast_fp16")]; + tensor input_117_cast_fp16 = add(x = var_709_cast_fp16, y = var_691_cast_fp16_0)[name = tensor("input_117_cast_fp16")]; + tensor linear_45_cast_fp16 = linear(bias = flow_net_res_blocks_4_mlp_0_bias_to_fp16, weight = flow_net_res_blocks_4_mlp_0_weight_to_fp16, x = input_117_cast_fp16)[name = tensor("linear_45_cast_fp16")]; + tensor input_121_cast_fp16 = silu(x = linear_45_cast_fp16)[name = tensor("input_121_cast_fp16")]; + tensor linear_46_cast_fp16 = linear(bias = flow_net_res_blocks_4_mlp_2_bias_to_fp16, weight = flow_net_res_blocks_4_mlp_2_weight_to_fp16, x = input_121_cast_fp16)[name = tensor("linear_46_cast_fp16")]; + tensor var_720_cast_fp16 = mul(x = var_691_cast_fp16_2, y = linear_46_cast_fp16)[name = tensor("op_720_cast_fp16")]; + tensor x_79_cast_fp16 = add(x = x_73_cast_fp16, y = var_720_cast_fp16)[name = tensor("x_79_cast_fp16")]; + tensor linear_47_cast_fp16 = linear(bias = flow_net_res_blocks_5_adaLN_modulation_1_bias_to_fp16, weight = flow_net_res_blocks_5_adaLN_modulation_1_weight_to_fp16, x = input_83_cast_fp16)[name = tensor("linear_47_cast_fp16")]; + tensor var_730_split_sizes_0 = const()[name = tensor("op_730_split_sizes_0"), val = tensor([512, 512, 512])]; + tensor var_730_axis_0 = const()[name = tensor("op_730_axis_0"), val = tensor(-1)]; + tensor var_730_cast_fp16_0, tensor var_730_cast_fp16_1, tensor var_730_cast_fp16_2 = split(axis = var_730_axis_0, split_sizes = var_730_split_sizes_0, x = linear_47_cast_fp16)[name = tensor("op_730_cast_fp16")]; + tensor mean_25_axes_0 = const()[name = tensor("mean_25_axes_0"), val = tensor([-1])]; + tensor mean_25_keep_dims_0 = const()[name = tensor("mean_25_keep_dims_0"), val = tensor(true)]; + tensor mean_25_cast_fp16 = reduce_mean(axes = mean_25_axes_0, keep_dims = mean_25_keep_dims_0, x = x_79_cast_fp16)[name = tensor("mean_25_cast_fp16")]; + tensor sub_20_cast_fp16 = sub(x = x_79_cast_fp16, y = mean_25_cast_fp16)[name = tensor("sub_20_cast_fp16")]; + tensor square_16_cast_fp16 = square(x = sub_20_cast_fp16)[name = tensor("square_16_cast_fp16")]; + tensor reduce_mean_33_axes_0 = const()[name = tensor("reduce_mean_33_axes_0"), val = tensor([-1])]; + tensor reduce_mean_33_keep_dims_0 = const()[name = tensor("reduce_mean_33_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_33_cast_fp16 = reduce_mean(axes = reduce_mean_33_axes_0, keep_dims = reduce_mean_33_keep_dims_0, x = square_16_cast_fp16)[name = tensor("reduce_mean_33_cast_fp16")]; + tensor var_740_to_fp16 = const()[name = tensor("op_740_to_fp16"), val = tensor(0x1.1p-20)]; + tensor var_741_cast_fp16 = add(x = reduce_mean_33_cast_fp16, y = var_740_to_fp16)[name = tensor("op_741_cast_fp16")]; + tensor var_742_cast_fp16 = sqrt(x = var_741_cast_fp16)[name = tensor("op_742_cast_fp16")]; + tensor x_81_cast_fp16 = real_div(x = sub_20_cast_fp16, y = var_742_cast_fp16)[name = tensor("x_81_cast_fp16")]; + tensor var_744_cast_fp16 = mul(x = x_81_cast_fp16, y = flow_net_res_blocks_5_in_ln_weight_to_fp16)[name = tensor("op_744_cast_fp16")]; + tensor x_83_cast_fp16 = add(x = var_744_cast_fp16, y = flow_net_res_blocks_5_in_ln_bias_to_fp16)[name = tensor("x_83_cast_fp16")]; + tensor var_746_promoted_to_fp16 = const()[name = tensor("op_746_promoted_to_fp16"), val = tensor(0x1p+0)]; + tensor var_747_cast_fp16 = add(x = var_730_cast_fp16_1, y = var_746_promoted_to_fp16)[name = tensor("op_747_cast_fp16")]; + tensor var_748_cast_fp16 = mul(x = x_83_cast_fp16, y = var_747_cast_fp16)[name = tensor("op_748_cast_fp16")]; + tensor input_125_cast_fp16 = add(x = var_748_cast_fp16, y = var_730_cast_fp16_0)[name = tensor("input_125_cast_fp16")]; + tensor linear_48_cast_fp16 = linear(bias = flow_net_res_blocks_5_mlp_0_bias_to_fp16, weight = flow_net_res_blocks_5_mlp_0_weight_to_fp16, x = input_125_cast_fp16)[name = tensor("linear_48_cast_fp16")]; + tensor input_129_cast_fp16 = silu(x = linear_48_cast_fp16)[name = tensor("input_129_cast_fp16")]; + tensor linear_49_cast_fp16 = linear(bias = flow_net_res_blocks_5_mlp_2_bias_to_fp16, weight = flow_net_res_blocks_5_mlp_2_weight_to_fp16, x = input_129_cast_fp16)[name = tensor("linear_49_cast_fp16")]; + tensor var_759_cast_fp16 = mul(x = var_730_cast_fp16_2, y = linear_49_cast_fp16)[name = tensor("op_759_cast_fp16")]; + tensor x_85_cast_fp16 = add(x = x_79_cast_fp16, y = var_759_cast_fp16)[name = tensor("x_85_cast_fp16")]; + tensor linear_50_cast_fp16 = linear(bias = flow_net_final_layer_adaLN_modulation_1_bias_to_fp16, weight = flow_net_final_layer_adaLN_modulation_1_weight_to_fp16, x = input_83_cast_fp16)[name = tensor("linear_50_cast_fp16")]; + tensor var_768_split_sizes_0 = const()[name = tensor("op_768_split_sizes_0"), val = tensor([512, 512])]; + tensor var_768_axis_0 = const()[name = tensor("op_768_axis_0"), val = tensor(-1)]; + tensor var_768_cast_fp16_0, tensor var_768_cast_fp16_1 = split(axis = var_768_axis_0, split_sizes = var_768_split_sizes_0, x = linear_50_cast_fp16)[name = tensor("op_768_cast_fp16")]; + tensor mean_27_axes_0 = const()[name = tensor("mean_27_axes_0"), val = tensor([-1])]; + tensor mean_27_keep_dims_0 = const()[name = tensor("mean_27_keep_dims_0"), val = tensor(true)]; + tensor mean_27_cast_fp16 = reduce_mean(axes = mean_27_axes_0, keep_dims = mean_27_keep_dims_0, x = x_85_cast_fp16)[name = tensor("mean_27_cast_fp16")]; + tensor sub_21_cast_fp16 = sub(x = x_85_cast_fp16, y = mean_27_cast_fp16)[name = tensor("sub_21_cast_fp16")]; + tensor square_17_cast_fp16 = square(x = sub_21_cast_fp16)[name = tensor("square_17_cast_fp16")]; + tensor reduce_mean_35_axes_0 = const()[name = tensor("reduce_mean_35_axes_0"), val = tensor([-1])]; + tensor reduce_mean_35_keep_dims_0 = const()[name = tensor("reduce_mean_35_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_35_cast_fp16 = reduce_mean(axes = reduce_mean_35_axes_0, keep_dims = reduce_mean_35_keep_dims_0, x = square_17_cast_fp16)[name = tensor("reduce_mean_35_cast_fp16")]; + tensor var_775_to_fp16 = const()[name = tensor("op_775_to_fp16"), val = tensor(0x1.1p-20)]; + tensor var_776_cast_fp16 = add(x = reduce_mean_35_cast_fp16, y = var_775_to_fp16)[name = tensor("op_776_cast_fp16")]; + tensor var_777_cast_fp16 = sqrt(x = var_776_cast_fp16)[name = tensor("op_777_cast_fp16")]; + tensor x_87_cast_fp16 = real_div(x = sub_21_cast_fp16, y = var_777_cast_fp16)[name = tensor("x_87_cast_fp16")]; + tensor var_779_promoted_to_fp16 = const()[name = tensor("op_779_promoted_to_fp16"), val = tensor(0x1p+0)]; + tensor var_780_cast_fp16 = add(x = var_768_cast_fp16_1, y = var_779_promoted_to_fp16)[name = tensor("op_780_cast_fp16")]; + tensor var_781_cast_fp16 = mul(x = x_87_cast_fp16, y = var_780_cast_fp16)[name = tensor("op_781_cast_fp16")]; + tensor input_133_cast_fp16 = add(x = var_781_cast_fp16, y = var_768_cast_fp16_0)[name = tensor("input_133_cast_fp16")]; + tensor linear_51_cast_fp16 = linear(bias = flow_net_final_layer_linear_bias_to_fp16, weight = flow_net_final_layer_linear_weight_to_fp16, x = input_133_cast_fp16)[name = tensor("linear_51_cast_fp16")]; + tensor var_792_to_fp16 = const()[name = tensor("op_792_to_fp16"), val = tensor(0x1p-3)]; + tensor var_793_cast_fp16 = mul(x = linear_51_cast_fp16, y = var_792_to_fp16)[name = tensor("op_793_cast_fp16")]; + tensor input_135_cast_fp16 = add(x = input_67_cast_fp16, y = var_793_cast_fp16)[name = tensor("input_135_cast_fp16")]; + tensor linear_52_cast_fp16 = linear(bias = flow_net_input_proj_bias_to_fp16, weight = flow_net_input_proj_weight_to_fp16, x = input_135_cast_fp16)[name = tensor("linear_52_cast_fp16")]; + tensor input_139_to_fp16 = const()[name = tensor("input_139_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(18999552)))]; + tensor input_141_cast_fp16 = silu(x = input_139_to_fp16)[name = tensor("input_141_cast_fp16")]; + tensor linear_54_cast_fp16 = linear(bias = flow_net_time_embed_0_mlp_2_bias_to_fp16, weight = flow_net_time_embed_0_mlp_2_weight_to_fp16, x = input_141_cast_fp16)[name = tensor("linear_54_cast_fp16")]; + tensor reduce_mean_36_axes_0 = const()[name = tensor("reduce_mean_36_axes_0"), val = tensor([-1])]; + tensor reduce_mean_36_keep_dims_0 = const()[name = tensor("reduce_mean_36_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_36_cast_fp16 = reduce_mean(axes = reduce_mean_36_axes_0, keep_dims = reduce_mean_36_keep_dims_0, x = linear_54_cast_fp16)[name = tensor("reduce_mean_36_cast_fp16")]; + tensor sub_22_cast_fp16 = sub(x = linear_54_cast_fp16, y = reduce_mean_36_cast_fp16)[name = tensor("sub_22_cast_fp16")]; + tensor square_18_cast_fp16 = square(x = sub_22_cast_fp16)[name = tensor("square_18_cast_fp16")]; + tensor reduce_mean_37_axes_0 = const()[name = tensor("reduce_mean_37_axes_0"), val = tensor([-1])]; + tensor reduce_mean_37_keep_dims_0 = const()[name = tensor("reduce_mean_37_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_37_cast_fp16 = reduce_mean(axes = reduce_mean_37_axes_0, keep_dims = reduce_mean_37_keep_dims_0, x = square_18_cast_fp16)[name = tensor("reduce_mean_37_cast_fp16")]; + tensor real_div_4_to_fp16 = const()[name = tensor("real_div_4_to_fp16"), val = tensor(0x1.008p+0)]; + tensor mul_4_cast_fp16 = mul(x = reduce_mean_37_cast_fp16, y = real_div_4_to_fp16)[name = tensor("mul_4_cast_fp16")]; + tensor var_859_to_fp16 = const()[name = tensor("op_859_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_37_cast_fp16 = add(x = mul_4_cast_fp16, y = var_859_to_fp16)[name = tensor("var_37_cast_fp16")]; + tensor var_862_epsilon_0 = const()[name = tensor("op_862_epsilon_0"), val = tensor(0x1.197998p-40)]; + tensor var_862_cast_fp16 = rsqrt(epsilon = var_862_epsilon_0, x = var_37_cast_fp16)[name = tensor("op_862_cast_fp16")]; + tensor var_863_cast_fp16 = mul(x = const_3_to_fp16, y = var_862_cast_fp16)[name = tensor("op_863_cast_fp16")]; + tensor var_864_cast_fp16 = mul(x = linear_54_cast_fp16, y = var_863_cast_fp16)[name = tensor("op_864_cast_fp16")]; + tensor input_145_to_fp16 = const()[name = tensor("input_145_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(19000640)))]; + tensor input_147_cast_fp16 = silu(x = input_145_to_fp16)[name = tensor("input_147_cast_fp16")]; + tensor linear_56_cast_fp16 = linear(bias = flow_net_time_embed_1_mlp_2_bias_to_fp16, weight = flow_net_time_embed_1_mlp_2_weight_to_fp16, x = input_147_cast_fp16)[name = tensor("linear_56_cast_fp16")]; + tensor reduce_mean_38_axes_0 = const()[name = tensor("reduce_mean_38_axes_0"), val = tensor([-1])]; + tensor reduce_mean_38_keep_dims_0 = const()[name = tensor("reduce_mean_38_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_38_cast_fp16 = reduce_mean(axes = reduce_mean_38_axes_0, keep_dims = reduce_mean_38_keep_dims_0, x = linear_56_cast_fp16)[name = tensor("reduce_mean_38_cast_fp16")]; + tensor sub_24_cast_fp16 = sub(x = linear_56_cast_fp16, y = reduce_mean_38_cast_fp16)[name = tensor("sub_24_cast_fp16")]; + tensor square_19_cast_fp16 = square(x = sub_24_cast_fp16)[name = tensor("square_19_cast_fp16")]; + tensor reduce_mean_39_axes_0 = const()[name = tensor("reduce_mean_39_axes_0"), val = tensor([-1])]; + tensor reduce_mean_39_keep_dims_0 = const()[name = tensor("reduce_mean_39_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_39_cast_fp16 = reduce_mean(axes = reduce_mean_39_axes_0, keep_dims = reduce_mean_39_keep_dims_0, x = square_19_cast_fp16)[name = tensor("reduce_mean_39_cast_fp16")]; + tensor real_div_5_to_fp16 = const()[name = tensor("real_div_5_to_fp16"), val = tensor(0x1.008p+0)]; + tensor mul_5_cast_fp16 = mul(x = reduce_mean_39_cast_fp16, y = real_div_5_to_fp16)[name = tensor("mul_5_cast_fp16")]; + tensor var_896_to_fp16 = const()[name = tensor("op_896_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_39_cast_fp16 = add(x = mul_5_cast_fp16, y = var_896_to_fp16)[name = tensor("var_39_cast_fp16")]; + tensor var_899_epsilon_0 = const()[name = tensor("op_899_epsilon_0"), val = tensor(0x1.197998p-40)]; + tensor var_899_cast_fp16 = rsqrt(epsilon = var_899_epsilon_0, x = var_39_cast_fp16)[name = tensor("op_899_cast_fp16")]; + tensor var_900_cast_fp16 = mul(x = const_5_to_fp16, y = var_899_cast_fp16)[name = tensor("op_900_cast_fp16")]; + tensor var_901_cast_fp16 = mul(x = linear_56_cast_fp16, y = var_900_cast_fp16)[name = tensor("op_901_cast_fp16")]; + tensor var_913_cast_fp16 = add(x = var_864_cast_fp16, y = var_901_cast_fp16)[name = tensor("op_913_cast_fp16")]; + tensor _inversed_t_combined_5_y_0_to_fp16 = const()[name = tensor("_inversed_t_combined_5_y_0_to_fp16"), val = tensor(0x1p-1)]; + tensor _inversed_t_combined_5_cast_fp16 = mul(x = var_913_cast_fp16, y = _inversed_t_combined_5_y_0_to_fp16)[name = tensor("_inversed_t_combined_5_cast_fp16")]; + tensor input_149_cast_fp16 = add(x = _inversed_t_combined_5_cast_fp16, y = linear_5_cast_fp16)[name = tensor("input_149_cast_fp16")]; + tensor input_151_cast_fp16 = silu(x = input_149_cast_fp16)[name = tensor("input_151_cast_fp16")]; + tensor linear_58_cast_fp16 = linear(bias = flow_net_res_blocks_0_adaLN_modulation_1_bias_to_fp16, weight = flow_net_res_blocks_0_adaLN_modulation_1_weight_to_fp16, x = input_151_cast_fp16)[name = tensor("linear_58_cast_fp16")]; + tensor var_928_split_sizes_0 = const()[name = tensor("op_928_split_sizes_0"), val = tensor([512, 512, 512])]; + tensor var_928_axis_0 = const()[name = tensor("op_928_axis_0"), val = tensor(-1)]; + tensor var_928_cast_fp16_0, tensor var_928_cast_fp16_1, tensor var_928_cast_fp16_2 = split(axis = var_928_axis_0, split_sizes = var_928_split_sizes_0, x = linear_58_cast_fp16)[name = tensor("op_928_cast_fp16")]; + tensor mean_29_axes_0 = const()[name = tensor("mean_29_axes_0"), val = tensor([-1])]; + tensor mean_29_keep_dims_0 = const()[name = tensor("mean_29_keep_dims_0"), val = tensor(true)]; + tensor mean_29_cast_fp16 = reduce_mean(axes = mean_29_axes_0, keep_dims = mean_29_keep_dims_0, x = linear_52_cast_fp16)[name = tensor("mean_29_cast_fp16")]; + tensor sub_26_cast_fp16 = sub(x = linear_52_cast_fp16, y = mean_29_cast_fp16)[name = tensor("sub_26_cast_fp16")]; + tensor square_20_cast_fp16 = square(x = sub_26_cast_fp16)[name = tensor("square_20_cast_fp16")]; + tensor reduce_mean_41_axes_0 = const()[name = tensor("reduce_mean_41_axes_0"), val = tensor([-1])]; + tensor reduce_mean_41_keep_dims_0 = const()[name = tensor("reduce_mean_41_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_41_cast_fp16 = reduce_mean(axes = reduce_mean_41_axes_0, keep_dims = reduce_mean_41_keep_dims_0, x = square_20_cast_fp16)[name = tensor("reduce_mean_41_cast_fp16")]; + tensor var_938_to_fp16 = const()[name = tensor("op_938_to_fp16"), val = tensor(0x1.1p-20)]; + tensor var_939_cast_fp16 = add(x = reduce_mean_41_cast_fp16, y = var_938_to_fp16)[name = tensor("op_939_cast_fp16")]; + tensor var_940_cast_fp16 = sqrt(x = var_939_cast_fp16)[name = tensor("op_940_cast_fp16")]; + tensor x_95_cast_fp16 = real_div(x = sub_26_cast_fp16, y = var_940_cast_fp16)[name = tensor("x_95_cast_fp16")]; + tensor var_942_cast_fp16 = mul(x = x_95_cast_fp16, y = flow_net_res_blocks_0_in_ln_weight_to_fp16)[name = tensor("op_942_cast_fp16")]; + tensor x_97_cast_fp16 = add(x = var_942_cast_fp16, y = flow_net_res_blocks_0_in_ln_bias_to_fp16)[name = tensor("x_97_cast_fp16")]; + tensor var_944_promoted_to_fp16 = const()[name = tensor("op_944_promoted_to_fp16"), val = tensor(0x1p+0)]; + tensor var_945_cast_fp16 = add(x = var_928_cast_fp16_1, y = var_944_promoted_to_fp16)[name = tensor("op_945_cast_fp16")]; + tensor var_946_cast_fp16 = mul(x = x_97_cast_fp16, y = var_945_cast_fp16)[name = tensor("op_946_cast_fp16")]; + tensor input_153_cast_fp16 = add(x = var_946_cast_fp16, y = var_928_cast_fp16_0)[name = tensor("input_153_cast_fp16")]; + tensor linear_59_cast_fp16 = linear(bias = flow_net_res_blocks_0_mlp_0_bias_to_fp16, weight = flow_net_res_blocks_0_mlp_0_weight_to_fp16, x = input_153_cast_fp16)[name = tensor("linear_59_cast_fp16")]; + tensor input_157_cast_fp16 = silu(x = linear_59_cast_fp16)[name = tensor("input_157_cast_fp16")]; + tensor linear_60_cast_fp16 = linear(bias = flow_net_res_blocks_0_mlp_2_bias_to_fp16, weight = flow_net_res_blocks_0_mlp_2_weight_to_fp16, x = input_157_cast_fp16)[name = tensor("linear_60_cast_fp16")]; + tensor var_957_cast_fp16 = mul(x = var_928_cast_fp16_2, y = linear_60_cast_fp16)[name = tensor("op_957_cast_fp16")]; + tensor x_99_cast_fp16 = add(x = linear_52_cast_fp16, y = var_957_cast_fp16)[name = tensor("x_99_cast_fp16")]; + tensor linear_61_cast_fp16 = linear(bias = flow_net_res_blocks_1_adaLN_modulation_1_bias_to_fp16, weight = flow_net_res_blocks_1_adaLN_modulation_1_weight_to_fp16, x = input_151_cast_fp16)[name = tensor("linear_61_cast_fp16")]; + tensor var_967_split_sizes_0 = const()[name = tensor("op_967_split_sizes_0"), val = tensor([512, 512, 512])]; + tensor var_967_axis_0 = const()[name = tensor("op_967_axis_0"), val = tensor(-1)]; + tensor var_967_cast_fp16_0, tensor var_967_cast_fp16_1, tensor var_967_cast_fp16_2 = split(axis = var_967_axis_0, split_sizes = var_967_split_sizes_0, x = linear_61_cast_fp16)[name = tensor("op_967_cast_fp16")]; + tensor mean_31_axes_0 = const()[name = tensor("mean_31_axes_0"), val = tensor([-1])]; + tensor mean_31_keep_dims_0 = const()[name = tensor("mean_31_keep_dims_0"), val = tensor(true)]; + tensor mean_31_cast_fp16 = reduce_mean(axes = mean_31_axes_0, keep_dims = mean_31_keep_dims_0, x = x_99_cast_fp16)[name = tensor("mean_31_cast_fp16")]; + tensor sub_27_cast_fp16 = sub(x = x_99_cast_fp16, y = mean_31_cast_fp16)[name = tensor("sub_27_cast_fp16")]; + tensor square_21_cast_fp16 = square(x = sub_27_cast_fp16)[name = tensor("square_21_cast_fp16")]; + tensor reduce_mean_43_axes_0 = const()[name = tensor("reduce_mean_43_axes_0"), val = tensor([-1])]; + tensor reduce_mean_43_keep_dims_0 = const()[name = tensor("reduce_mean_43_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_43_cast_fp16 = reduce_mean(axes = reduce_mean_43_axes_0, keep_dims = reduce_mean_43_keep_dims_0, x = square_21_cast_fp16)[name = tensor("reduce_mean_43_cast_fp16")]; + tensor var_977_to_fp16 = const()[name = tensor("op_977_to_fp16"), val = tensor(0x1.1p-20)]; + tensor var_978_cast_fp16 = add(x = reduce_mean_43_cast_fp16, y = var_977_to_fp16)[name = tensor("op_978_cast_fp16")]; + tensor var_979_cast_fp16 = sqrt(x = var_978_cast_fp16)[name = tensor("op_979_cast_fp16")]; + tensor x_101_cast_fp16 = real_div(x = sub_27_cast_fp16, y = var_979_cast_fp16)[name = tensor("x_101_cast_fp16")]; + tensor var_981_cast_fp16 = mul(x = x_101_cast_fp16, y = flow_net_res_blocks_1_in_ln_weight_to_fp16)[name = tensor("op_981_cast_fp16")]; + tensor x_103_cast_fp16 = add(x = var_981_cast_fp16, y = flow_net_res_blocks_1_in_ln_bias_to_fp16)[name = tensor("x_103_cast_fp16")]; + tensor var_983_promoted_to_fp16 = const()[name = tensor("op_983_promoted_to_fp16"), val = tensor(0x1p+0)]; + tensor var_984_cast_fp16 = add(x = var_967_cast_fp16_1, y = var_983_promoted_to_fp16)[name = tensor("op_984_cast_fp16")]; + tensor var_985_cast_fp16 = mul(x = x_103_cast_fp16, y = var_984_cast_fp16)[name = tensor("op_985_cast_fp16")]; + tensor input_161_cast_fp16 = add(x = var_985_cast_fp16, y = var_967_cast_fp16_0)[name = tensor("input_161_cast_fp16")]; + tensor linear_62_cast_fp16 = linear(bias = flow_net_res_blocks_1_mlp_0_bias_to_fp16, weight = flow_net_res_blocks_1_mlp_0_weight_to_fp16, x = input_161_cast_fp16)[name = tensor("linear_62_cast_fp16")]; + tensor input_165_cast_fp16 = silu(x = linear_62_cast_fp16)[name = tensor("input_165_cast_fp16")]; + tensor linear_63_cast_fp16 = linear(bias = flow_net_res_blocks_1_mlp_2_bias_to_fp16, weight = flow_net_res_blocks_1_mlp_2_weight_to_fp16, x = input_165_cast_fp16)[name = tensor("linear_63_cast_fp16")]; + tensor var_996_cast_fp16 = mul(x = var_967_cast_fp16_2, y = linear_63_cast_fp16)[name = tensor("op_996_cast_fp16")]; + tensor x_105_cast_fp16 = add(x = x_99_cast_fp16, y = var_996_cast_fp16)[name = tensor("x_105_cast_fp16")]; + tensor linear_64_cast_fp16 = linear(bias = flow_net_res_blocks_2_adaLN_modulation_1_bias_to_fp16, weight = flow_net_res_blocks_2_adaLN_modulation_1_weight_to_fp16, x = input_151_cast_fp16)[name = tensor("linear_64_cast_fp16")]; + tensor var_1006_split_sizes_0 = const()[name = tensor("op_1006_split_sizes_0"), val = tensor([512, 512, 512])]; + tensor var_1006_axis_0 = const()[name = tensor("op_1006_axis_0"), val = tensor(-1)]; + tensor var_1006_cast_fp16_0, tensor var_1006_cast_fp16_1, tensor var_1006_cast_fp16_2 = split(axis = var_1006_axis_0, split_sizes = var_1006_split_sizes_0, x = linear_64_cast_fp16)[name = tensor("op_1006_cast_fp16")]; + tensor mean_33_axes_0 = const()[name = tensor("mean_33_axes_0"), val = tensor([-1])]; + tensor mean_33_keep_dims_0 = const()[name = tensor("mean_33_keep_dims_0"), val = tensor(true)]; + tensor mean_33_cast_fp16 = reduce_mean(axes = mean_33_axes_0, keep_dims = mean_33_keep_dims_0, x = x_105_cast_fp16)[name = tensor("mean_33_cast_fp16")]; + tensor sub_28_cast_fp16 = sub(x = x_105_cast_fp16, y = mean_33_cast_fp16)[name = tensor("sub_28_cast_fp16")]; + tensor square_22_cast_fp16 = square(x = sub_28_cast_fp16)[name = tensor("square_22_cast_fp16")]; + tensor reduce_mean_45_axes_0 = const()[name = tensor("reduce_mean_45_axes_0"), val = tensor([-1])]; + tensor reduce_mean_45_keep_dims_0 = const()[name = tensor("reduce_mean_45_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_45_cast_fp16 = reduce_mean(axes = reduce_mean_45_axes_0, keep_dims = reduce_mean_45_keep_dims_0, x = square_22_cast_fp16)[name = tensor("reduce_mean_45_cast_fp16")]; + tensor var_1016_to_fp16 = const()[name = tensor("op_1016_to_fp16"), val = tensor(0x1.1p-20)]; + tensor var_1017_cast_fp16 = add(x = reduce_mean_45_cast_fp16, y = var_1016_to_fp16)[name = tensor("op_1017_cast_fp16")]; + tensor var_1018_cast_fp16 = sqrt(x = var_1017_cast_fp16)[name = tensor("op_1018_cast_fp16")]; + tensor x_107_cast_fp16 = real_div(x = sub_28_cast_fp16, y = var_1018_cast_fp16)[name = tensor("x_107_cast_fp16")]; + tensor var_1020_cast_fp16 = mul(x = x_107_cast_fp16, y = flow_net_res_blocks_2_in_ln_weight_to_fp16)[name = tensor("op_1020_cast_fp16")]; + tensor x_109_cast_fp16 = add(x = var_1020_cast_fp16, y = flow_net_res_blocks_2_in_ln_bias_to_fp16)[name = tensor("x_109_cast_fp16")]; + tensor var_1022_promoted_to_fp16 = const()[name = tensor("op_1022_promoted_to_fp16"), val = tensor(0x1p+0)]; + tensor var_1023_cast_fp16 = add(x = var_1006_cast_fp16_1, y = var_1022_promoted_to_fp16)[name = tensor("op_1023_cast_fp16")]; + tensor var_1024_cast_fp16 = mul(x = x_109_cast_fp16, y = var_1023_cast_fp16)[name = tensor("op_1024_cast_fp16")]; + tensor input_169_cast_fp16 = add(x = var_1024_cast_fp16, y = var_1006_cast_fp16_0)[name = tensor("input_169_cast_fp16")]; + tensor linear_65_cast_fp16 = linear(bias = flow_net_res_blocks_2_mlp_0_bias_to_fp16, weight = flow_net_res_blocks_2_mlp_0_weight_to_fp16, x = input_169_cast_fp16)[name = tensor("linear_65_cast_fp16")]; + tensor input_173_cast_fp16 = silu(x = linear_65_cast_fp16)[name = tensor("input_173_cast_fp16")]; + tensor linear_66_cast_fp16 = linear(bias = flow_net_res_blocks_2_mlp_2_bias_to_fp16, weight = flow_net_res_blocks_2_mlp_2_weight_to_fp16, x = input_173_cast_fp16)[name = tensor("linear_66_cast_fp16")]; + tensor var_1035_cast_fp16 = mul(x = var_1006_cast_fp16_2, y = linear_66_cast_fp16)[name = tensor("op_1035_cast_fp16")]; + tensor x_111_cast_fp16 = add(x = x_105_cast_fp16, y = var_1035_cast_fp16)[name = tensor("x_111_cast_fp16")]; + tensor linear_67_cast_fp16 = linear(bias = flow_net_res_blocks_3_adaLN_modulation_1_bias_to_fp16, weight = flow_net_res_blocks_3_adaLN_modulation_1_weight_to_fp16, x = input_151_cast_fp16)[name = tensor("linear_67_cast_fp16")]; + tensor var_1045_split_sizes_0 = const()[name = tensor("op_1045_split_sizes_0"), val = tensor([512, 512, 512])]; + tensor var_1045_axis_0 = const()[name = tensor("op_1045_axis_0"), val = tensor(-1)]; + tensor var_1045_cast_fp16_0, tensor var_1045_cast_fp16_1, tensor var_1045_cast_fp16_2 = split(axis = var_1045_axis_0, split_sizes = var_1045_split_sizes_0, x = linear_67_cast_fp16)[name = tensor("op_1045_cast_fp16")]; + tensor mean_35_axes_0 = const()[name = tensor("mean_35_axes_0"), val = tensor([-1])]; + tensor mean_35_keep_dims_0 = const()[name = tensor("mean_35_keep_dims_0"), val = tensor(true)]; + tensor mean_35_cast_fp16 = reduce_mean(axes = mean_35_axes_0, keep_dims = mean_35_keep_dims_0, x = x_111_cast_fp16)[name = tensor("mean_35_cast_fp16")]; + tensor sub_29_cast_fp16 = sub(x = x_111_cast_fp16, y = mean_35_cast_fp16)[name = tensor("sub_29_cast_fp16")]; + tensor square_23_cast_fp16 = square(x = sub_29_cast_fp16)[name = tensor("square_23_cast_fp16")]; + tensor reduce_mean_47_axes_0 = const()[name = tensor("reduce_mean_47_axes_0"), val = tensor([-1])]; + tensor reduce_mean_47_keep_dims_0 = const()[name = tensor("reduce_mean_47_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_47_cast_fp16 = reduce_mean(axes = reduce_mean_47_axes_0, keep_dims = reduce_mean_47_keep_dims_0, x = square_23_cast_fp16)[name = tensor("reduce_mean_47_cast_fp16")]; + tensor var_1055_to_fp16 = const()[name = tensor("op_1055_to_fp16"), val = tensor(0x1.1p-20)]; + tensor var_1056_cast_fp16 = add(x = reduce_mean_47_cast_fp16, y = var_1055_to_fp16)[name = tensor("op_1056_cast_fp16")]; + tensor var_1057_cast_fp16 = sqrt(x = var_1056_cast_fp16)[name = tensor("op_1057_cast_fp16")]; + tensor x_113_cast_fp16 = real_div(x = sub_29_cast_fp16, y = var_1057_cast_fp16)[name = tensor("x_113_cast_fp16")]; + tensor var_1059_cast_fp16 = mul(x = x_113_cast_fp16, y = flow_net_res_blocks_3_in_ln_weight_to_fp16)[name = tensor("op_1059_cast_fp16")]; + tensor x_115_cast_fp16 = add(x = var_1059_cast_fp16, y = flow_net_res_blocks_3_in_ln_bias_to_fp16)[name = tensor("x_115_cast_fp16")]; + tensor var_1061_promoted_to_fp16 = const()[name = tensor("op_1061_promoted_to_fp16"), val = tensor(0x1p+0)]; + tensor var_1062_cast_fp16 = add(x = var_1045_cast_fp16_1, y = var_1061_promoted_to_fp16)[name = tensor("op_1062_cast_fp16")]; + tensor var_1063_cast_fp16 = mul(x = x_115_cast_fp16, y = var_1062_cast_fp16)[name = tensor("op_1063_cast_fp16")]; + tensor input_177_cast_fp16 = add(x = var_1063_cast_fp16, y = var_1045_cast_fp16_0)[name = tensor("input_177_cast_fp16")]; + tensor linear_68_cast_fp16 = linear(bias = flow_net_res_blocks_3_mlp_0_bias_to_fp16, weight = flow_net_res_blocks_3_mlp_0_weight_to_fp16, x = input_177_cast_fp16)[name = tensor("linear_68_cast_fp16")]; + tensor input_181_cast_fp16 = silu(x = linear_68_cast_fp16)[name = tensor("input_181_cast_fp16")]; + tensor linear_69_cast_fp16 = linear(bias = flow_net_res_blocks_3_mlp_2_bias_to_fp16, weight = flow_net_res_blocks_3_mlp_2_weight_to_fp16, x = input_181_cast_fp16)[name = tensor("linear_69_cast_fp16")]; + tensor var_1074_cast_fp16 = mul(x = var_1045_cast_fp16_2, y = linear_69_cast_fp16)[name = tensor("op_1074_cast_fp16")]; + tensor x_117_cast_fp16 = add(x = x_111_cast_fp16, y = var_1074_cast_fp16)[name = tensor("x_117_cast_fp16")]; + tensor linear_70_cast_fp16 = linear(bias = flow_net_res_blocks_4_adaLN_modulation_1_bias_to_fp16, weight = flow_net_res_blocks_4_adaLN_modulation_1_weight_to_fp16, x = input_151_cast_fp16)[name = tensor("linear_70_cast_fp16")]; + tensor var_1084_split_sizes_0 = const()[name = tensor("op_1084_split_sizes_0"), val = tensor([512, 512, 512])]; + tensor var_1084_axis_0 = const()[name = tensor("op_1084_axis_0"), val = tensor(-1)]; + tensor var_1084_cast_fp16_0, tensor var_1084_cast_fp16_1, tensor var_1084_cast_fp16_2 = split(axis = var_1084_axis_0, split_sizes = var_1084_split_sizes_0, x = linear_70_cast_fp16)[name = tensor("op_1084_cast_fp16")]; + tensor mean_37_axes_0 = const()[name = tensor("mean_37_axes_0"), val = tensor([-1])]; + tensor mean_37_keep_dims_0 = const()[name = tensor("mean_37_keep_dims_0"), val = tensor(true)]; + tensor mean_37_cast_fp16 = reduce_mean(axes = mean_37_axes_0, keep_dims = mean_37_keep_dims_0, x = x_117_cast_fp16)[name = tensor("mean_37_cast_fp16")]; + tensor sub_30_cast_fp16 = sub(x = x_117_cast_fp16, y = mean_37_cast_fp16)[name = tensor("sub_30_cast_fp16")]; + tensor square_24_cast_fp16 = square(x = sub_30_cast_fp16)[name = tensor("square_24_cast_fp16")]; + tensor reduce_mean_49_axes_0 = const()[name = tensor("reduce_mean_49_axes_0"), val = tensor([-1])]; + tensor reduce_mean_49_keep_dims_0 = const()[name = tensor("reduce_mean_49_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_49_cast_fp16 = reduce_mean(axes = reduce_mean_49_axes_0, keep_dims = reduce_mean_49_keep_dims_0, x = square_24_cast_fp16)[name = tensor("reduce_mean_49_cast_fp16")]; + tensor var_1094_to_fp16 = const()[name = tensor("op_1094_to_fp16"), val = tensor(0x1.1p-20)]; + tensor var_1095_cast_fp16 = add(x = reduce_mean_49_cast_fp16, y = var_1094_to_fp16)[name = tensor("op_1095_cast_fp16")]; + tensor var_1096_cast_fp16 = sqrt(x = var_1095_cast_fp16)[name = tensor("op_1096_cast_fp16")]; + tensor x_119_cast_fp16 = real_div(x = sub_30_cast_fp16, y = var_1096_cast_fp16)[name = tensor("x_119_cast_fp16")]; + tensor var_1098_cast_fp16 = mul(x = x_119_cast_fp16, y = flow_net_res_blocks_4_in_ln_weight_to_fp16)[name = tensor("op_1098_cast_fp16")]; + tensor x_121_cast_fp16 = add(x = var_1098_cast_fp16, y = flow_net_res_blocks_4_in_ln_bias_to_fp16)[name = tensor("x_121_cast_fp16")]; + tensor var_1100_promoted_to_fp16 = const()[name = tensor("op_1100_promoted_to_fp16"), val = tensor(0x1p+0)]; + tensor var_1101_cast_fp16 = add(x = var_1084_cast_fp16_1, y = var_1100_promoted_to_fp16)[name = tensor("op_1101_cast_fp16")]; + tensor var_1102_cast_fp16 = mul(x = x_121_cast_fp16, y = var_1101_cast_fp16)[name = tensor("op_1102_cast_fp16")]; + tensor input_185_cast_fp16 = add(x = var_1102_cast_fp16, y = var_1084_cast_fp16_0)[name = tensor("input_185_cast_fp16")]; + tensor linear_71_cast_fp16 = linear(bias = flow_net_res_blocks_4_mlp_0_bias_to_fp16, weight = flow_net_res_blocks_4_mlp_0_weight_to_fp16, x = input_185_cast_fp16)[name = tensor("linear_71_cast_fp16")]; + tensor input_189_cast_fp16 = silu(x = linear_71_cast_fp16)[name = tensor("input_189_cast_fp16")]; + tensor linear_72_cast_fp16 = linear(bias = flow_net_res_blocks_4_mlp_2_bias_to_fp16, weight = flow_net_res_blocks_4_mlp_2_weight_to_fp16, x = input_189_cast_fp16)[name = tensor("linear_72_cast_fp16")]; + tensor var_1113_cast_fp16 = mul(x = var_1084_cast_fp16_2, y = linear_72_cast_fp16)[name = tensor("op_1113_cast_fp16")]; + tensor x_123_cast_fp16 = add(x = x_117_cast_fp16, y = var_1113_cast_fp16)[name = tensor("x_123_cast_fp16")]; + tensor linear_73_cast_fp16 = linear(bias = flow_net_res_blocks_5_adaLN_modulation_1_bias_to_fp16, weight = flow_net_res_blocks_5_adaLN_modulation_1_weight_to_fp16, x = input_151_cast_fp16)[name = tensor("linear_73_cast_fp16")]; + tensor var_1123_split_sizes_0 = const()[name = tensor("op_1123_split_sizes_0"), val = tensor([512, 512, 512])]; + tensor var_1123_axis_0 = const()[name = tensor("op_1123_axis_0"), val = tensor(-1)]; + tensor var_1123_cast_fp16_0, tensor var_1123_cast_fp16_1, tensor var_1123_cast_fp16_2 = split(axis = var_1123_axis_0, split_sizes = var_1123_split_sizes_0, x = linear_73_cast_fp16)[name = tensor("op_1123_cast_fp16")]; + tensor mean_39_axes_0 = const()[name = tensor("mean_39_axes_0"), val = tensor([-1])]; + tensor mean_39_keep_dims_0 = const()[name = tensor("mean_39_keep_dims_0"), val = tensor(true)]; + tensor mean_39_cast_fp16 = reduce_mean(axes = mean_39_axes_0, keep_dims = mean_39_keep_dims_0, x = x_123_cast_fp16)[name = tensor("mean_39_cast_fp16")]; + tensor sub_31_cast_fp16 = sub(x = x_123_cast_fp16, y = mean_39_cast_fp16)[name = tensor("sub_31_cast_fp16")]; + tensor square_25_cast_fp16 = square(x = sub_31_cast_fp16)[name = tensor("square_25_cast_fp16")]; + tensor reduce_mean_51_axes_0 = const()[name = tensor("reduce_mean_51_axes_0"), val = tensor([-1])]; + tensor reduce_mean_51_keep_dims_0 = const()[name = tensor("reduce_mean_51_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_51_cast_fp16 = reduce_mean(axes = reduce_mean_51_axes_0, keep_dims = reduce_mean_51_keep_dims_0, x = square_25_cast_fp16)[name = tensor("reduce_mean_51_cast_fp16")]; + tensor var_1133_to_fp16 = const()[name = tensor("op_1133_to_fp16"), val = tensor(0x1.1p-20)]; + tensor var_1134_cast_fp16 = add(x = reduce_mean_51_cast_fp16, y = var_1133_to_fp16)[name = tensor("op_1134_cast_fp16")]; + tensor var_1135_cast_fp16 = sqrt(x = var_1134_cast_fp16)[name = tensor("op_1135_cast_fp16")]; + tensor x_125_cast_fp16 = real_div(x = sub_31_cast_fp16, y = var_1135_cast_fp16)[name = tensor("x_125_cast_fp16")]; + tensor var_1137_cast_fp16 = mul(x = x_125_cast_fp16, y = flow_net_res_blocks_5_in_ln_weight_to_fp16)[name = tensor("op_1137_cast_fp16")]; + tensor x_127_cast_fp16 = add(x = var_1137_cast_fp16, y = flow_net_res_blocks_5_in_ln_bias_to_fp16)[name = tensor("x_127_cast_fp16")]; + tensor var_1139_promoted_to_fp16 = const()[name = tensor("op_1139_promoted_to_fp16"), val = tensor(0x1p+0)]; + tensor var_1140_cast_fp16 = add(x = var_1123_cast_fp16_1, y = var_1139_promoted_to_fp16)[name = tensor("op_1140_cast_fp16")]; + tensor var_1141_cast_fp16 = mul(x = x_127_cast_fp16, y = var_1140_cast_fp16)[name = tensor("op_1141_cast_fp16")]; + tensor input_193_cast_fp16 = add(x = var_1141_cast_fp16, y = var_1123_cast_fp16_0)[name = tensor("input_193_cast_fp16")]; + tensor linear_74_cast_fp16 = linear(bias = flow_net_res_blocks_5_mlp_0_bias_to_fp16, weight = flow_net_res_blocks_5_mlp_0_weight_to_fp16, x = input_193_cast_fp16)[name = tensor("linear_74_cast_fp16")]; + tensor input_197_cast_fp16 = silu(x = linear_74_cast_fp16)[name = tensor("input_197_cast_fp16")]; + tensor linear_75_cast_fp16 = linear(bias = flow_net_res_blocks_5_mlp_2_bias_to_fp16, weight = flow_net_res_blocks_5_mlp_2_weight_to_fp16, x = input_197_cast_fp16)[name = tensor("linear_75_cast_fp16")]; + tensor var_1152_cast_fp16 = mul(x = var_1123_cast_fp16_2, y = linear_75_cast_fp16)[name = tensor("op_1152_cast_fp16")]; + tensor x_129_cast_fp16 = add(x = x_123_cast_fp16, y = var_1152_cast_fp16)[name = tensor("x_129_cast_fp16")]; + tensor linear_76_cast_fp16 = linear(bias = flow_net_final_layer_adaLN_modulation_1_bias_to_fp16, weight = flow_net_final_layer_adaLN_modulation_1_weight_to_fp16, x = input_151_cast_fp16)[name = tensor("linear_76_cast_fp16")]; + tensor var_1161_split_sizes_0 = const()[name = tensor("op_1161_split_sizes_0"), val = tensor([512, 512])]; + tensor var_1161_axis_0 = const()[name = tensor("op_1161_axis_0"), val = tensor(-1)]; + tensor var_1161_cast_fp16_0, tensor var_1161_cast_fp16_1 = split(axis = var_1161_axis_0, split_sizes = var_1161_split_sizes_0, x = linear_76_cast_fp16)[name = tensor("op_1161_cast_fp16")]; + tensor mean_41_axes_0 = const()[name = tensor("mean_41_axes_0"), val = tensor([-1])]; + tensor mean_41_keep_dims_0 = const()[name = tensor("mean_41_keep_dims_0"), val = tensor(true)]; + tensor mean_41_cast_fp16 = reduce_mean(axes = mean_41_axes_0, keep_dims = mean_41_keep_dims_0, x = x_129_cast_fp16)[name = tensor("mean_41_cast_fp16")]; + tensor sub_32_cast_fp16 = sub(x = x_129_cast_fp16, y = mean_41_cast_fp16)[name = tensor("sub_32_cast_fp16")]; + tensor square_26_cast_fp16 = square(x = sub_32_cast_fp16)[name = tensor("square_26_cast_fp16")]; + tensor reduce_mean_53_axes_0 = const()[name = tensor("reduce_mean_53_axes_0"), val = tensor([-1])]; + tensor reduce_mean_53_keep_dims_0 = const()[name = tensor("reduce_mean_53_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_53_cast_fp16 = reduce_mean(axes = reduce_mean_53_axes_0, keep_dims = reduce_mean_53_keep_dims_0, x = square_26_cast_fp16)[name = tensor("reduce_mean_53_cast_fp16")]; + tensor var_1168_to_fp16 = const()[name = tensor("op_1168_to_fp16"), val = tensor(0x1.1p-20)]; + tensor var_1169_cast_fp16 = add(x = reduce_mean_53_cast_fp16, y = var_1168_to_fp16)[name = tensor("op_1169_cast_fp16")]; + tensor var_1170_cast_fp16 = sqrt(x = var_1169_cast_fp16)[name = tensor("op_1170_cast_fp16")]; + tensor x_131_cast_fp16 = real_div(x = sub_32_cast_fp16, y = var_1170_cast_fp16)[name = tensor("x_131_cast_fp16")]; + tensor var_1172_promoted_to_fp16 = const()[name = tensor("op_1172_promoted_to_fp16"), val = tensor(0x1p+0)]; + tensor var_1173_cast_fp16 = add(x = var_1161_cast_fp16_1, y = var_1172_promoted_to_fp16)[name = tensor("op_1173_cast_fp16")]; + tensor var_1174_cast_fp16 = mul(x = x_131_cast_fp16, y = var_1173_cast_fp16)[name = tensor("op_1174_cast_fp16")]; + tensor input_201_cast_fp16 = add(x = var_1174_cast_fp16, y = var_1161_cast_fp16_0)[name = tensor("input_201_cast_fp16")]; + tensor linear_77_cast_fp16 = linear(bias = flow_net_final_layer_linear_bias_to_fp16, weight = flow_net_final_layer_linear_weight_to_fp16, x = input_201_cast_fp16)[name = tensor("linear_77_cast_fp16")]; + tensor var_1185_to_fp16 = const()[name = tensor("op_1185_to_fp16"), val = tensor(0x1p-3)]; + tensor var_1186_cast_fp16 = mul(x = linear_77_cast_fp16, y = var_1185_to_fp16)[name = tensor("op_1186_cast_fp16")]; + tensor input_203_cast_fp16 = add(x = input_135_cast_fp16, y = var_1186_cast_fp16)[name = tensor("input_203_cast_fp16")]; + tensor linear_78_cast_fp16 = linear(bias = flow_net_input_proj_bias_to_fp16, weight = flow_net_input_proj_weight_to_fp16, x = input_203_cast_fp16)[name = tensor("linear_78_cast_fp16")]; + tensor input_207_to_fp16 = const()[name = tensor("input_207_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(19001728)))]; + tensor input_209_cast_fp16 = silu(x = input_207_to_fp16)[name = tensor("input_209_cast_fp16")]; + tensor linear_80_cast_fp16 = linear(bias = flow_net_time_embed_0_mlp_2_bias_to_fp16, weight = flow_net_time_embed_0_mlp_2_weight_to_fp16, x = input_209_cast_fp16)[name = tensor("linear_80_cast_fp16")]; + tensor reduce_mean_54_axes_0 = const()[name = tensor("reduce_mean_54_axes_0"), val = tensor([-1])]; + tensor reduce_mean_54_keep_dims_0 = const()[name = tensor("reduce_mean_54_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_54_cast_fp16 = reduce_mean(axes = reduce_mean_54_axes_0, keep_dims = reduce_mean_54_keep_dims_0, x = linear_80_cast_fp16)[name = tensor("reduce_mean_54_cast_fp16")]; + tensor sub_33_cast_fp16 = sub(x = linear_80_cast_fp16, y = reduce_mean_54_cast_fp16)[name = tensor("sub_33_cast_fp16")]; + tensor square_27_cast_fp16 = square(x = sub_33_cast_fp16)[name = tensor("square_27_cast_fp16")]; + tensor reduce_mean_55_axes_0 = const()[name = tensor("reduce_mean_55_axes_0"), val = tensor([-1])]; + tensor reduce_mean_55_keep_dims_0 = const()[name = tensor("reduce_mean_55_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_55_cast_fp16 = reduce_mean(axes = reduce_mean_55_axes_0, keep_dims = reduce_mean_55_keep_dims_0, x = square_27_cast_fp16)[name = tensor("reduce_mean_55_cast_fp16")]; + tensor real_div_6_to_fp16 = const()[name = tensor("real_div_6_to_fp16"), val = tensor(0x1.008p+0)]; + tensor mul_6_cast_fp16 = mul(x = reduce_mean_55_cast_fp16, y = real_div_6_to_fp16)[name = tensor("mul_6_cast_fp16")]; + tensor var_1252_to_fp16 = const()[name = tensor("op_1252_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_55_cast_fp16 = add(x = mul_6_cast_fp16, y = var_1252_to_fp16)[name = tensor("var_55_cast_fp16")]; + tensor var_1255_epsilon_0 = const()[name = tensor("op_1255_epsilon_0"), val = tensor(0x1.197998p-40)]; + tensor var_1255_cast_fp16 = rsqrt(epsilon = var_1255_epsilon_0, x = var_55_cast_fp16)[name = tensor("op_1255_cast_fp16")]; + tensor var_1256_cast_fp16 = mul(x = const_3_to_fp16, y = var_1255_cast_fp16)[name = tensor("op_1256_cast_fp16")]; + tensor var_1257_cast_fp16 = mul(x = linear_80_cast_fp16, y = var_1256_cast_fp16)[name = tensor("op_1257_cast_fp16")]; + tensor input_213_to_fp16 = const()[name = tensor("input_213_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(19002816)))]; + tensor input_215_cast_fp16 = silu(x = input_213_to_fp16)[name = tensor("input_215_cast_fp16")]; + tensor linear_82_cast_fp16 = linear(bias = flow_net_time_embed_1_mlp_2_bias_to_fp16, weight = flow_net_time_embed_1_mlp_2_weight_to_fp16, x = input_215_cast_fp16)[name = tensor("linear_82_cast_fp16")]; + tensor reduce_mean_56_axes_0 = const()[name = tensor("reduce_mean_56_axes_0"), val = tensor([-1])]; + tensor reduce_mean_56_keep_dims_0 = const()[name = tensor("reduce_mean_56_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_56_cast_fp16 = reduce_mean(axes = reduce_mean_56_axes_0, keep_dims = reduce_mean_56_keep_dims_0, x = linear_82_cast_fp16)[name = tensor("reduce_mean_56_cast_fp16")]; + tensor sub_35_cast_fp16 = sub(x = linear_82_cast_fp16, y = reduce_mean_56_cast_fp16)[name = tensor("sub_35_cast_fp16")]; + tensor square_28_cast_fp16 = square(x = sub_35_cast_fp16)[name = tensor("square_28_cast_fp16")]; + tensor reduce_mean_57_axes_0 = const()[name = tensor("reduce_mean_57_axes_0"), val = tensor([-1])]; + tensor reduce_mean_57_keep_dims_0 = const()[name = tensor("reduce_mean_57_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_57_cast_fp16 = reduce_mean(axes = reduce_mean_57_axes_0, keep_dims = reduce_mean_57_keep_dims_0, x = square_28_cast_fp16)[name = tensor("reduce_mean_57_cast_fp16")]; + tensor real_div_7_to_fp16 = const()[name = tensor("real_div_7_to_fp16"), val = tensor(0x1.008p+0)]; + tensor mul_7_cast_fp16 = mul(x = reduce_mean_57_cast_fp16, y = real_div_7_to_fp16)[name = tensor("mul_7_cast_fp16")]; + tensor var_1289_to_fp16 = const()[name = tensor("op_1289_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_57_cast_fp16 = add(x = mul_7_cast_fp16, y = var_1289_to_fp16)[name = tensor("var_57_cast_fp16")]; + tensor var_1292_epsilon_0 = const()[name = tensor("op_1292_epsilon_0"), val = tensor(0x1.197998p-40)]; + tensor var_1292_cast_fp16 = rsqrt(epsilon = var_1292_epsilon_0, x = var_57_cast_fp16)[name = tensor("op_1292_cast_fp16")]; + tensor var_1293_cast_fp16 = mul(x = const_5_to_fp16, y = var_1292_cast_fp16)[name = tensor("op_1293_cast_fp16")]; + tensor var_1294_cast_fp16 = mul(x = linear_82_cast_fp16, y = var_1293_cast_fp16)[name = tensor("op_1294_cast_fp16")]; + tensor var_1306_cast_fp16 = add(x = var_1257_cast_fp16, y = var_1294_cast_fp16)[name = tensor("op_1306_cast_fp16")]; + tensor _inversed_t_combined_7_y_0_to_fp16 = const()[name = tensor("_inversed_t_combined_7_y_0_to_fp16"), val = tensor(0x1p-1)]; + tensor _inversed_t_combined_7_cast_fp16 = mul(x = var_1306_cast_fp16, y = _inversed_t_combined_7_y_0_to_fp16)[name = tensor("_inversed_t_combined_7_cast_fp16")]; + tensor input_217_cast_fp16 = add(x = _inversed_t_combined_7_cast_fp16, y = linear_5_cast_fp16)[name = tensor("input_217_cast_fp16")]; + tensor input_219_cast_fp16 = silu(x = input_217_cast_fp16)[name = tensor("input_219_cast_fp16")]; + tensor linear_84_cast_fp16 = linear(bias = flow_net_res_blocks_0_adaLN_modulation_1_bias_to_fp16, weight = flow_net_res_blocks_0_adaLN_modulation_1_weight_to_fp16, x = input_219_cast_fp16)[name = tensor("linear_84_cast_fp16")]; + tensor var_1321_split_sizes_0 = const()[name = tensor("op_1321_split_sizes_0"), val = tensor([512, 512, 512])]; + tensor var_1321_axis_0 = const()[name = tensor("op_1321_axis_0"), val = tensor(-1)]; + tensor var_1321_cast_fp16_0, tensor var_1321_cast_fp16_1, tensor var_1321_cast_fp16_2 = split(axis = var_1321_axis_0, split_sizes = var_1321_split_sizes_0, x = linear_84_cast_fp16)[name = tensor("op_1321_cast_fp16")]; + tensor mean_43_axes_0 = const()[name = tensor("mean_43_axes_0"), val = tensor([-1])]; + tensor mean_43_keep_dims_0 = const()[name = tensor("mean_43_keep_dims_0"), val = tensor(true)]; + tensor mean_43_cast_fp16 = reduce_mean(axes = mean_43_axes_0, keep_dims = mean_43_keep_dims_0, x = linear_78_cast_fp16)[name = tensor("mean_43_cast_fp16")]; + tensor sub_37_cast_fp16 = sub(x = linear_78_cast_fp16, y = mean_43_cast_fp16)[name = tensor("sub_37_cast_fp16")]; + tensor square_29_cast_fp16 = square(x = sub_37_cast_fp16)[name = tensor("square_29_cast_fp16")]; + tensor reduce_mean_59_axes_0 = const()[name = tensor("reduce_mean_59_axes_0"), val = tensor([-1])]; + tensor reduce_mean_59_keep_dims_0 = const()[name = tensor("reduce_mean_59_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_59_cast_fp16 = reduce_mean(axes = reduce_mean_59_axes_0, keep_dims = reduce_mean_59_keep_dims_0, x = square_29_cast_fp16)[name = tensor("reduce_mean_59_cast_fp16")]; + tensor var_1331_to_fp16 = const()[name = tensor("op_1331_to_fp16"), val = tensor(0x1.1p-20)]; + tensor var_1332_cast_fp16 = add(x = reduce_mean_59_cast_fp16, y = var_1331_to_fp16)[name = tensor("op_1332_cast_fp16")]; + tensor var_1333_cast_fp16 = sqrt(x = var_1332_cast_fp16)[name = tensor("op_1333_cast_fp16")]; + tensor x_139_cast_fp16 = real_div(x = sub_37_cast_fp16, y = var_1333_cast_fp16)[name = tensor("x_139_cast_fp16")]; + tensor var_1335_cast_fp16 = mul(x = x_139_cast_fp16, y = flow_net_res_blocks_0_in_ln_weight_to_fp16)[name = tensor("op_1335_cast_fp16")]; + tensor x_141_cast_fp16 = add(x = var_1335_cast_fp16, y = flow_net_res_blocks_0_in_ln_bias_to_fp16)[name = tensor("x_141_cast_fp16")]; + tensor var_1337_promoted_to_fp16 = const()[name = tensor("op_1337_promoted_to_fp16"), val = tensor(0x1p+0)]; + tensor var_1338_cast_fp16 = add(x = var_1321_cast_fp16_1, y = var_1337_promoted_to_fp16)[name = tensor("op_1338_cast_fp16")]; + tensor var_1339_cast_fp16 = mul(x = x_141_cast_fp16, y = var_1338_cast_fp16)[name = tensor("op_1339_cast_fp16")]; + tensor input_221_cast_fp16 = add(x = var_1339_cast_fp16, y = var_1321_cast_fp16_0)[name = tensor("input_221_cast_fp16")]; + tensor linear_85_cast_fp16 = linear(bias = flow_net_res_blocks_0_mlp_0_bias_to_fp16, weight = flow_net_res_blocks_0_mlp_0_weight_to_fp16, x = input_221_cast_fp16)[name = tensor("linear_85_cast_fp16")]; + tensor input_225_cast_fp16 = silu(x = linear_85_cast_fp16)[name = tensor("input_225_cast_fp16")]; + tensor linear_86_cast_fp16 = linear(bias = flow_net_res_blocks_0_mlp_2_bias_to_fp16, weight = flow_net_res_blocks_0_mlp_2_weight_to_fp16, x = input_225_cast_fp16)[name = tensor("linear_86_cast_fp16")]; + tensor var_1350_cast_fp16 = mul(x = var_1321_cast_fp16_2, y = linear_86_cast_fp16)[name = tensor("op_1350_cast_fp16")]; + tensor x_143_cast_fp16 = add(x = linear_78_cast_fp16, y = var_1350_cast_fp16)[name = tensor("x_143_cast_fp16")]; + tensor linear_87_cast_fp16 = linear(bias = flow_net_res_blocks_1_adaLN_modulation_1_bias_to_fp16, weight = flow_net_res_blocks_1_adaLN_modulation_1_weight_to_fp16, x = input_219_cast_fp16)[name = tensor("linear_87_cast_fp16")]; + tensor var_1360_split_sizes_0 = const()[name = tensor("op_1360_split_sizes_0"), val = tensor([512, 512, 512])]; + tensor var_1360_axis_0 = const()[name = tensor("op_1360_axis_0"), val = tensor(-1)]; + tensor var_1360_cast_fp16_0, tensor var_1360_cast_fp16_1, tensor var_1360_cast_fp16_2 = split(axis = var_1360_axis_0, split_sizes = var_1360_split_sizes_0, x = linear_87_cast_fp16)[name = tensor("op_1360_cast_fp16")]; + tensor mean_45_axes_0 = const()[name = tensor("mean_45_axes_0"), val = tensor([-1])]; + tensor mean_45_keep_dims_0 = const()[name = tensor("mean_45_keep_dims_0"), val = tensor(true)]; + tensor mean_45_cast_fp16 = reduce_mean(axes = mean_45_axes_0, keep_dims = mean_45_keep_dims_0, x = x_143_cast_fp16)[name = tensor("mean_45_cast_fp16")]; + tensor sub_38_cast_fp16 = sub(x = x_143_cast_fp16, y = mean_45_cast_fp16)[name = tensor("sub_38_cast_fp16")]; + tensor square_30_cast_fp16 = square(x = sub_38_cast_fp16)[name = tensor("square_30_cast_fp16")]; + tensor reduce_mean_61_axes_0 = const()[name = tensor("reduce_mean_61_axes_0"), val = tensor([-1])]; + tensor reduce_mean_61_keep_dims_0 = const()[name = tensor("reduce_mean_61_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_61_cast_fp16 = reduce_mean(axes = reduce_mean_61_axes_0, keep_dims = reduce_mean_61_keep_dims_0, x = square_30_cast_fp16)[name = tensor("reduce_mean_61_cast_fp16")]; + tensor var_1370_to_fp16 = const()[name = tensor("op_1370_to_fp16"), val = tensor(0x1.1p-20)]; + tensor var_1371_cast_fp16 = add(x = reduce_mean_61_cast_fp16, y = var_1370_to_fp16)[name = tensor("op_1371_cast_fp16")]; + tensor var_1372_cast_fp16 = sqrt(x = var_1371_cast_fp16)[name = tensor("op_1372_cast_fp16")]; + tensor x_145_cast_fp16 = real_div(x = sub_38_cast_fp16, y = var_1372_cast_fp16)[name = tensor("x_145_cast_fp16")]; + tensor var_1374_cast_fp16 = mul(x = x_145_cast_fp16, y = flow_net_res_blocks_1_in_ln_weight_to_fp16)[name = tensor("op_1374_cast_fp16")]; + tensor x_147_cast_fp16 = add(x = var_1374_cast_fp16, y = flow_net_res_blocks_1_in_ln_bias_to_fp16)[name = tensor("x_147_cast_fp16")]; + tensor var_1376_promoted_to_fp16 = const()[name = tensor("op_1376_promoted_to_fp16"), val = tensor(0x1p+0)]; + tensor var_1377_cast_fp16 = add(x = var_1360_cast_fp16_1, y = var_1376_promoted_to_fp16)[name = tensor("op_1377_cast_fp16")]; + tensor var_1378_cast_fp16 = mul(x = x_147_cast_fp16, y = var_1377_cast_fp16)[name = tensor("op_1378_cast_fp16")]; + tensor input_229_cast_fp16 = add(x = var_1378_cast_fp16, y = var_1360_cast_fp16_0)[name = tensor("input_229_cast_fp16")]; + tensor linear_88_cast_fp16 = linear(bias = flow_net_res_blocks_1_mlp_0_bias_to_fp16, weight = flow_net_res_blocks_1_mlp_0_weight_to_fp16, x = input_229_cast_fp16)[name = tensor("linear_88_cast_fp16")]; + tensor input_233_cast_fp16 = silu(x = linear_88_cast_fp16)[name = tensor("input_233_cast_fp16")]; + tensor linear_89_cast_fp16 = linear(bias = flow_net_res_blocks_1_mlp_2_bias_to_fp16, weight = flow_net_res_blocks_1_mlp_2_weight_to_fp16, x = input_233_cast_fp16)[name = tensor("linear_89_cast_fp16")]; + tensor var_1389_cast_fp16 = mul(x = var_1360_cast_fp16_2, y = linear_89_cast_fp16)[name = tensor("op_1389_cast_fp16")]; + tensor x_149_cast_fp16 = add(x = x_143_cast_fp16, y = var_1389_cast_fp16)[name = tensor("x_149_cast_fp16")]; + tensor linear_90_cast_fp16 = linear(bias = flow_net_res_blocks_2_adaLN_modulation_1_bias_to_fp16, weight = flow_net_res_blocks_2_adaLN_modulation_1_weight_to_fp16, x = input_219_cast_fp16)[name = tensor("linear_90_cast_fp16")]; + tensor var_1399_split_sizes_0 = const()[name = tensor("op_1399_split_sizes_0"), val = tensor([512, 512, 512])]; + tensor var_1399_axis_0 = const()[name = tensor("op_1399_axis_0"), val = tensor(-1)]; + tensor var_1399_cast_fp16_0, tensor var_1399_cast_fp16_1, tensor var_1399_cast_fp16_2 = split(axis = var_1399_axis_0, split_sizes = var_1399_split_sizes_0, x = linear_90_cast_fp16)[name = tensor("op_1399_cast_fp16")]; + tensor mean_47_axes_0 = const()[name = tensor("mean_47_axes_0"), val = tensor([-1])]; + tensor mean_47_keep_dims_0 = const()[name = tensor("mean_47_keep_dims_0"), val = tensor(true)]; + tensor mean_47_cast_fp16 = reduce_mean(axes = mean_47_axes_0, keep_dims = mean_47_keep_dims_0, x = x_149_cast_fp16)[name = tensor("mean_47_cast_fp16")]; + tensor sub_39_cast_fp16 = sub(x = x_149_cast_fp16, y = mean_47_cast_fp16)[name = tensor("sub_39_cast_fp16")]; + tensor square_31_cast_fp16 = square(x = sub_39_cast_fp16)[name = tensor("square_31_cast_fp16")]; + tensor reduce_mean_63_axes_0 = const()[name = tensor("reduce_mean_63_axes_0"), val = tensor([-1])]; + tensor reduce_mean_63_keep_dims_0 = const()[name = tensor("reduce_mean_63_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_63_cast_fp16 = reduce_mean(axes = reduce_mean_63_axes_0, keep_dims = reduce_mean_63_keep_dims_0, x = square_31_cast_fp16)[name = tensor("reduce_mean_63_cast_fp16")]; + tensor var_1409_to_fp16 = const()[name = tensor("op_1409_to_fp16"), val = tensor(0x1.1p-20)]; + tensor var_1410_cast_fp16 = add(x = reduce_mean_63_cast_fp16, y = var_1409_to_fp16)[name = tensor("op_1410_cast_fp16")]; + tensor var_1411_cast_fp16 = sqrt(x = var_1410_cast_fp16)[name = tensor("op_1411_cast_fp16")]; + tensor x_151_cast_fp16 = real_div(x = sub_39_cast_fp16, y = var_1411_cast_fp16)[name = tensor("x_151_cast_fp16")]; + tensor var_1413_cast_fp16 = mul(x = x_151_cast_fp16, y = flow_net_res_blocks_2_in_ln_weight_to_fp16)[name = tensor("op_1413_cast_fp16")]; + tensor x_153_cast_fp16 = add(x = var_1413_cast_fp16, y = flow_net_res_blocks_2_in_ln_bias_to_fp16)[name = tensor("x_153_cast_fp16")]; + tensor var_1415_promoted_to_fp16 = const()[name = tensor("op_1415_promoted_to_fp16"), val = tensor(0x1p+0)]; + tensor var_1416_cast_fp16 = add(x = var_1399_cast_fp16_1, y = var_1415_promoted_to_fp16)[name = tensor("op_1416_cast_fp16")]; + tensor var_1417_cast_fp16 = mul(x = x_153_cast_fp16, y = var_1416_cast_fp16)[name = tensor("op_1417_cast_fp16")]; + tensor input_237_cast_fp16 = add(x = var_1417_cast_fp16, y = var_1399_cast_fp16_0)[name = tensor("input_237_cast_fp16")]; + tensor linear_91_cast_fp16 = linear(bias = flow_net_res_blocks_2_mlp_0_bias_to_fp16, weight = flow_net_res_blocks_2_mlp_0_weight_to_fp16, x = input_237_cast_fp16)[name = tensor("linear_91_cast_fp16")]; + tensor input_241_cast_fp16 = silu(x = linear_91_cast_fp16)[name = tensor("input_241_cast_fp16")]; + tensor linear_92_cast_fp16 = linear(bias = flow_net_res_blocks_2_mlp_2_bias_to_fp16, weight = flow_net_res_blocks_2_mlp_2_weight_to_fp16, x = input_241_cast_fp16)[name = tensor("linear_92_cast_fp16")]; + tensor var_1428_cast_fp16 = mul(x = var_1399_cast_fp16_2, y = linear_92_cast_fp16)[name = tensor("op_1428_cast_fp16")]; + tensor x_155_cast_fp16 = add(x = x_149_cast_fp16, y = var_1428_cast_fp16)[name = tensor("x_155_cast_fp16")]; + tensor linear_93_cast_fp16 = linear(bias = flow_net_res_blocks_3_adaLN_modulation_1_bias_to_fp16, weight = flow_net_res_blocks_3_adaLN_modulation_1_weight_to_fp16, x = input_219_cast_fp16)[name = tensor("linear_93_cast_fp16")]; + tensor var_1438_split_sizes_0 = const()[name = tensor("op_1438_split_sizes_0"), val = tensor([512, 512, 512])]; + tensor var_1438_axis_0 = const()[name = tensor("op_1438_axis_0"), val = tensor(-1)]; + tensor var_1438_cast_fp16_0, tensor var_1438_cast_fp16_1, tensor var_1438_cast_fp16_2 = split(axis = var_1438_axis_0, split_sizes = var_1438_split_sizes_0, x = linear_93_cast_fp16)[name = tensor("op_1438_cast_fp16")]; + tensor mean_49_axes_0 = const()[name = tensor("mean_49_axes_0"), val = tensor([-1])]; + tensor mean_49_keep_dims_0 = const()[name = tensor("mean_49_keep_dims_0"), val = tensor(true)]; + tensor mean_49_cast_fp16 = reduce_mean(axes = mean_49_axes_0, keep_dims = mean_49_keep_dims_0, x = x_155_cast_fp16)[name = tensor("mean_49_cast_fp16")]; + tensor sub_40_cast_fp16 = sub(x = x_155_cast_fp16, y = mean_49_cast_fp16)[name = tensor("sub_40_cast_fp16")]; + tensor square_32_cast_fp16 = square(x = sub_40_cast_fp16)[name = tensor("square_32_cast_fp16")]; + tensor reduce_mean_65_axes_0 = const()[name = tensor("reduce_mean_65_axes_0"), val = tensor([-1])]; + tensor reduce_mean_65_keep_dims_0 = const()[name = tensor("reduce_mean_65_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_65_cast_fp16 = reduce_mean(axes = reduce_mean_65_axes_0, keep_dims = reduce_mean_65_keep_dims_0, x = square_32_cast_fp16)[name = tensor("reduce_mean_65_cast_fp16")]; + tensor var_1448_to_fp16 = const()[name = tensor("op_1448_to_fp16"), val = tensor(0x1.1p-20)]; + tensor var_1449_cast_fp16 = add(x = reduce_mean_65_cast_fp16, y = var_1448_to_fp16)[name = tensor("op_1449_cast_fp16")]; + tensor var_1450_cast_fp16 = sqrt(x = var_1449_cast_fp16)[name = tensor("op_1450_cast_fp16")]; + tensor x_157_cast_fp16 = real_div(x = sub_40_cast_fp16, y = var_1450_cast_fp16)[name = tensor("x_157_cast_fp16")]; + tensor var_1452_cast_fp16 = mul(x = x_157_cast_fp16, y = flow_net_res_blocks_3_in_ln_weight_to_fp16)[name = tensor("op_1452_cast_fp16")]; + tensor x_159_cast_fp16 = add(x = var_1452_cast_fp16, y = flow_net_res_blocks_3_in_ln_bias_to_fp16)[name = tensor("x_159_cast_fp16")]; + tensor var_1454_promoted_to_fp16 = const()[name = tensor("op_1454_promoted_to_fp16"), val = tensor(0x1p+0)]; + tensor var_1455_cast_fp16 = add(x = var_1438_cast_fp16_1, y = var_1454_promoted_to_fp16)[name = tensor("op_1455_cast_fp16")]; + tensor var_1456_cast_fp16 = mul(x = x_159_cast_fp16, y = var_1455_cast_fp16)[name = tensor("op_1456_cast_fp16")]; + tensor input_245_cast_fp16 = add(x = var_1456_cast_fp16, y = var_1438_cast_fp16_0)[name = tensor("input_245_cast_fp16")]; + tensor linear_94_cast_fp16 = linear(bias = flow_net_res_blocks_3_mlp_0_bias_to_fp16, weight = flow_net_res_blocks_3_mlp_0_weight_to_fp16, x = input_245_cast_fp16)[name = tensor("linear_94_cast_fp16")]; + tensor input_249_cast_fp16 = silu(x = linear_94_cast_fp16)[name = tensor("input_249_cast_fp16")]; + tensor linear_95_cast_fp16 = linear(bias = flow_net_res_blocks_3_mlp_2_bias_to_fp16, weight = flow_net_res_blocks_3_mlp_2_weight_to_fp16, x = input_249_cast_fp16)[name = tensor("linear_95_cast_fp16")]; + tensor var_1467_cast_fp16 = mul(x = var_1438_cast_fp16_2, y = linear_95_cast_fp16)[name = tensor("op_1467_cast_fp16")]; + tensor x_161_cast_fp16 = add(x = x_155_cast_fp16, y = var_1467_cast_fp16)[name = tensor("x_161_cast_fp16")]; + tensor linear_96_cast_fp16 = linear(bias = flow_net_res_blocks_4_adaLN_modulation_1_bias_to_fp16, weight = flow_net_res_blocks_4_adaLN_modulation_1_weight_to_fp16, x = input_219_cast_fp16)[name = tensor("linear_96_cast_fp16")]; + tensor var_1477_split_sizes_0 = const()[name = tensor("op_1477_split_sizes_0"), val = tensor([512, 512, 512])]; + tensor var_1477_axis_0 = const()[name = tensor("op_1477_axis_0"), val = tensor(-1)]; + tensor var_1477_cast_fp16_0, tensor var_1477_cast_fp16_1, tensor var_1477_cast_fp16_2 = split(axis = var_1477_axis_0, split_sizes = var_1477_split_sizes_0, x = linear_96_cast_fp16)[name = tensor("op_1477_cast_fp16")]; + tensor mean_51_axes_0 = const()[name = tensor("mean_51_axes_0"), val = tensor([-1])]; + tensor mean_51_keep_dims_0 = const()[name = tensor("mean_51_keep_dims_0"), val = tensor(true)]; + tensor mean_51_cast_fp16 = reduce_mean(axes = mean_51_axes_0, keep_dims = mean_51_keep_dims_0, x = x_161_cast_fp16)[name = tensor("mean_51_cast_fp16")]; + tensor sub_41_cast_fp16 = sub(x = x_161_cast_fp16, y = mean_51_cast_fp16)[name = tensor("sub_41_cast_fp16")]; + tensor square_33_cast_fp16 = square(x = sub_41_cast_fp16)[name = tensor("square_33_cast_fp16")]; + tensor reduce_mean_67_axes_0 = const()[name = tensor("reduce_mean_67_axes_0"), val = tensor([-1])]; + tensor reduce_mean_67_keep_dims_0 = const()[name = tensor("reduce_mean_67_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_67_cast_fp16 = reduce_mean(axes = reduce_mean_67_axes_0, keep_dims = reduce_mean_67_keep_dims_0, x = square_33_cast_fp16)[name = tensor("reduce_mean_67_cast_fp16")]; + tensor var_1487_to_fp16 = const()[name = tensor("op_1487_to_fp16"), val = tensor(0x1.1p-20)]; + tensor var_1488_cast_fp16 = add(x = reduce_mean_67_cast_fp16, y = var_1487_to_fp16)[name = tensor("op_1488_cast_fp16")]; + tensor var_1489_cast_fp16 = sqrt(x = var_1488_cast_fp16)[name = tensor("op_1489_cast_fp16")]; + tensor x_163_cast_fp16 = real_div(x = sub_41_cast_fp16, y = var_1489_cast_fp16)[name = tensor("x_163_cast_fp16")]; + tensor var_1491_cast_fp16 = mul(x = x_163_cast_fp16, y = flow_net_res_blocks_4_in_ln_weight_to_fp16)[name = tensor("op_1491_cast_fp16")]; + tensor x_165_cast_fp16 = add(x = var_1491_cast_fp16, y = flow_net_res_blocks_4_in_ln_bias_to_fp16)[name = tensor("x_165_cast_fp16")]; + tensor var_1493_promoted_to_fp16 = const()[name = tensor("op_1493_promoted_to_fp16"), val = tensor(0x1p+0)]; + tensor var_1494_cast_fp16 = add(x = var_1477_cast_fp16_1, y = var_1493_promoted_to_fp16)[name = tensor("op_1494_cast_fp16")]; + tensor var_1495_cast_fp16 = mul(x = x_165_cast_fp16, y = var_1494_cast_fp16)[name = tensor("op_1495_cast_fp16")]; + tensor input_253_cast_fp16 = add(x = var_1495_cast_fp16, y = var_1477_cast_fp16_0)[name = tensor("input_253_cast_fp16")]; + tensor linear_97_cast_fp16 = linear(bias = flow_net_res_blocks_4_mlp_0_bias_to_fp16, weight = flow_net_res_blocks_4_mlp_0_weight_to_fp16, x = input_253_cast_fp16)[name = tensor("linear_97_cast_fp16")]; + tensor input_257_cast_fp16 = silu(x = linear_97_cast_fp16)[name = tensor("input_257_cast_fp16")]; + tensor linear_98_cast_fp16 = linear(bias = flow_net_res_blocks_4_mlp_2_bias_to_fp16, weight = flow_net_res_blocks_4_mlp_2_weight_to_fp16, x = input_257_cast_fp16)[name = tensor("linear_98_cast_fp16")]; + tensor var_1506_cast_fp16 = mul(x = var_1477_cast_fp16_2, y = linear_98_cast_fp16)[name = tensor("op_1506_cast_fp16")]; + tensor x_167_cast_fp16 = add(x = x_161_cast_fp16, y = var_1506_cast_fp16)[name = tensor("x_167_cast_fp16")]; + tensor linear_99_cast_fp16 = linear(bias = flow_net_res_blocks_5_adaLN_modulation_1_bias_to_fp16, weight = flow_net_res_blocks_5_adaLN_modulation_1_weight_to_fp16, x = input_219_cast_fp16)[name = tensor("linear_99_cast_fp16")]; + tensor var_1516_split_sizes_0 = const()[name = tensor("op_1516_split_sizes_0"), val = tensor([512, 512, 512])]; + tensor var_1516_axis_0 = const()[name = tensor("op_1516_axis_0"), val = tensor(-1)]; + tensor var_1516_cast_fp16_0, tensor var_1516_cast_fp16_1, tensor var_1516_cast_fp16_2 = split(axis = var_1516_axis_0, split_sizes = var_1516_split_sizes_0, x = linear_99_cast_fp16)[name = tensor("op_1516_cast_fp16")]; + tensor mean_53_axes_0 = const()[name = tensor("mean_53_axes_0"), val = tensor([-1])]; + tensor mean_53_keep_dims_0 = const()[name = tensor("mean_53_keep_dims_0"), val = tensor(true)]; + tensor mean_53_cast_fp16 = reduce_mean(axes = mean_53_axes_0, keep_dims = mean_53_keep_dims_0, x = x_167_cast_fp16)[name = tensor("mean_53_cast_fp16")]; + tensor sub_42_cast_fp16 = sub(x = x_167_cast_fp16, y = mean_53_cast_fp16)[name = tensor("sub_42_cast_fp16")]; + tensor square_34_cast_fp16 = square(x = sub_42_cast_fp16)[name = tensor("square_34_cast_fp16")]; + tensor reduce_mean_69_axes_0 = const()[name = tensor("reduce_mean_69_axes_0"), val = tensor([-1])]; + tensor reduce_mean_69_keep_dims_0 = const()[name = tensor("reduce_mean_69_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_69_cast_fp16 = reduce_mean(axes = reduce_mean_69_axes_0, keep_dims = reduce_mean_69_keep_dims_0, x = square_34_cast_fp16)[name = tensor("reduce_mean_69_cast_fp16")]; + tensor var_1526_to_fp16 = const()[name = tensor("op_1526_to_fp16"), val = tensor(0x1.1p-20)]; + tensor var_1527_cast_fp16 = add(x = reduce_mean_69_cast_fp16, y = var_1526_to_fp16)[name = tensor("op_1527_cast_fp16")]; + tensor var_1528_cast_fp16 = sqrt(x = var_1527_cast_fp16)[name = tensor("op_1528_cast_fp16")]; + tensor x_169_cast_fp16 = real_div(x = sub_42_cast_fp16, y = var_1528_cast_fp16)[name = tensor("x_169_cast_fp16")]; + tensor var_1530_cast_fp16 = mul(x = x_169_cast_fp16, y = flow_net_res_blocks_5_in_ln_weight_to_fp16)[name = tensor("op_1530_cast_fp16")]; + tensor x_171_cast_fp16 = add(x = var_1530_cast_fp16, y = flow_net_res_blocks_5_in_ln_bias_to_fp16)[name = tensor("x_171_cast_fp16")]; + tensor var_1532_promoted_to_fp16 = const()[name = tensor("op_1532_promoted_to_fp16"), val = tensor(0x1p+0)]; + tensor var_1533_cast_fp16 = add(x = var_1516_cast_fp16_1, y = var_1532_promoted_to_fp16)[name = tensor("op_1533_cast_fp16")]; + tensor var_1534_cast_fp16 = mul(x = x_171_cast_fp16, y = var_1533_cast_fp16)[name = tensor("op_1534_cast_fp16")]; + tensor input_261_cast_fp16 = add(x = var_1534_cast_fp16, y = var_1516_cast_fp16_0)[name = tensor("input_261_cast_fp16")]; + tensor linear_100_cast_fp16 = linear(bias = flow_net_res_blocks_5_mlp_0_bias_to_fp16, weight = flow_net_res_blocks_5_mlp_0_weight_to_fp16, x = input_261_cast_fp16)[name = tensor("linear_100_cast_fp16")]; + tensor input_265_cast_fp16 = silu(x = linear_100_cast_fp16)[name = tensor("input_265_cast_fp16")]; + tensor linear_101_cast_fp16 = linear(bias = flow_net_res_blocks_5_mlp_2_bias_to_fp16, weight = flow_net_res_blocks_5_mlp_2_weight_to_fp16, x = input_265_cast_fp16)[name = tensor("linear_101_cast_fp16")]; + tensor var_1545_cast_fp16 = mul(x = var_1516_cast_fp16_2, y = linear_101_cast_fp16)[name = tensor("op_1545_cast_fp16")]; + tensor x_173_cast_fp16 = add(x = x_167_cast_fp16, y = var_1545_cast_fp16)[name = tensor("x_173_cast_fp16")]; + tensor linear_102_cast_fp16 = linear(bias = flow_net_final_layer_adaLN_modulation_1_bias_to_fp16, weight = flow_net_final_layer_adaLN_modulation_1_weight_to_fp16, x = input_219_cast_fp16)[name = tensor("linear_102_cast_fp16")]; + tensor var_1554_split_sizes_0 = const()[name = tensor("op_1554_split_sizes_0"), val = tensor([512, 512])]; + tensor var_1554_axis_0 = const()[name = tensor("op_1554_axis_0"), val = tensor(-1)]; + tensor var_1554_cast_fp16_0, tensor var_1554_cast_fp16_1 = split(axis = var_1554_axis_0, split_sizes = var_1554_split_sizes_0, x = linear_102_cast_fp16)[name = tensor("op_1554_cast_fp16")]; + tensor mean_55_axes_0 = const()[name = tensor("mean_55_axes_0"), val = tensor([-1])]; + tensor mean_55_keep_dims_0 = const()[name = tensor("mean_55_keep_dims_0"), val = tensor(true)]; + tensor mean_55_cast_fp16 = reduce_mean(axes = mean_55_axes_0, keep_dims = mean_55_keep_dims_0, x = x_173_cast_fp16)[name = tensor("mean_55_cast_fp16")]; + tensor sub_43_cast_fp16 = sub(x = x_173_cast_fp16, y = mean_55_cast_fp16)[name = tensor("sub_43_cast_fp16")]; + tensor square_35_cast_fp16 = square(x = sub_43_cast_fp16)[name = tensor("square_35_cast_fp16")]; + tensor reduce_mean_71_axes_0 = const()[name = tensor("reduce_mean_71_axes_0"), val = tensor([-1])]; + tensor reduce_mean_71_keep_dims_0 = const()[name = tensor("reduce_mean_71_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_71_cast_fp16 = reduce_mean(axes = reduce_mean_71_axes_0, keep_dims = reduce_mean_71_keep_dims_0, x = square_35_cast_fp16)[name = tensor("reduce_mean_71_cast_fp16")]; + tensor var_1561_to_fp16 = const()[name = tensor("op_1561_to_fp16"), val = tensor(0x1.1p-20)]; + tensor var_1562_cast_fp16 = add(x = reduce_mean_71_cast_fp16, y = var_1561_to_fp16)[name = tensor("op_1562_cast_fp16")]; + tensor var_1563_cast_fp16 = sqrt(x = var_1562_cast_fp16)[name = tensor("op_1563_cast_fp16")]; + tensor x_175_cast_fp16 = real_div(x = sub_43_cast_fp16, y = var_1563_cast_fp16)[name = tensor("x_175_cast_fp16")]; + tensor var_1565_promoted_to_fp16 = const()[name = tensor("op_1565_promoted_to_fp16"), val = tensor(0x1p+0)]; + tensor var_1566_cast_fp16 = add(x = var_1554_cast_fp16_1, y = var_1565_promoted_to_fp16)[name = tensor("op_1566_cast_fp16")]; + tensor var_1567_cast_fp16 = mul(x = x_175_cast_fp16, y = var_1566_cast_fp16)[name = tensor("op_1567_cast_fp16")]; + tensor input_269_cast_fp16 = add(x = var_1567_cast_fp16, y = var_1554_cast_fp16_0)[name = tensor("input_269_cast_fp16")]; + tensor linear_103_cast_fp16 = linear(bias = flow_net_final_layer_linear_bias_to_fp16, weight = flow_net_final_layer_linear_weight_to_fp16, x = input_269_cast_fp16)[name = tensor("linear_103_cast_fp16")]; + tensor var_1578_to_fp16 = const()[name = tensor("op_1578_to_fp16"), val = tensor(0x1p-3)]; + tensor var_1579_cast_fp16 = mul(x = linear_103_cast_fp16, y = var_1578_to_fp16)[name = tensor("op_1579_cast_fp16")]; + tensor input_271_cast_fp16 = add(x = input_203_cast_fp16, y = var_1579_cast_fp16)[name = tensor("input_271_cast_fp16")]; + tensor linear_104_cast_fp16 = linear(bias = flow_net_input_proj_bias_to_fp16, weight = flow_net_input_proj_weight_to_fp16, x = input_271_cast_fp16)[name = tensor("linear_104_cast_fp16")]; + tensor input_275_to_fp16 = const()[name = tensor("input_275_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(19003904)))]; + tensor input_277_cast_fp16 = silu(x = input_275_to_fp16)[name = tensor("input_277_cast_fp16")]; + tensor linear_106_cast_fp16 = linear(bias = flow_net_time_embed_0_mlp_2_bias_to_fp16, weight = flow_net_time_embed_0_mlp_2_weight_to_fp16, x = input_277_cast_fp16)[name = tensor("linear_106_cast_fp16")]; + tensor reduce_mean_72_axes_0 = const()[name = tensor("reduce_mean_72_axes_0"), val = tensor([-1])]; + tensor reduce_mean_72_keep_dims_0 = const()[name = tensor("reduce_mean_72_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_72_cast_fp16 = reduce_mean(axes = reduce_mean_72_axes_0, keep_dims = reduce_mean_72_keep_dims_0, x = linear_106_cast_fp16)[name = tensor("reduce_mean_72_cast_fp16")]; + tensor sub_44_cast_fp16 = sub(x = linear_106_cast_fp16, y = reduce_mean_72_cast_fp16)[name = tensor("sub_44_cast_fp16")]; + tensor square_36_cast_fp16 = square(x = sub_44_cast_fp16)[name = tensor("square_36_cast_fp16")]; + tensor reduce_mean_73_axes_0 = const()[name = tensor("reduce_mean_73_axes_0"), val = tensor([-1])]; + tensor reduce_mean_73_keep_dims_0 = const()[name = tensor("reduce_mean_73_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_73_cast_fp16 = reduce_mean(axes = reduce_mean_73_axes_0, keep_dims = reduce_mean_73_keep_dims_0, x = square_36_cast_fp16)[name = tensor("reduce_mean_73_cast_fp16")]; + tensor real_div_8_to_fp16 = const()[name = tensor("real_div_8_to_fp16"), val = tensor(0x1.008p+0)]; + tensor mul_8_cast_fp16 = mul(x = reduce_mean_73_cast_fp16, y = real_div_8_to_fp16)[name = tensor("mul_8_cast_fp16")]; + tensor var_1645_to_fp16 = const()[name = tensor("op_1645_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_73_cast_fp16_0 = add(x = mul_8_cast_fp16, y = var_1645_to_fp16)[name = tensor("var_73_cast_fp16")]; + tensor var_1648_epsilon_0 = const()[name = tensor("op_1648_epsilon_0"), val = tensor(0x1.197998p-40)]; + tensor var_1648_cast_fp16 = rsqrt(epsilon = var_1648_epsilon_0, x = var_73_cast_fp16_0)[name = tensor("op_1648_cast_fp16")]; + tensor var_1649_cast_fp16 = mul(x = const_3_to_fp16, y = var_1648_cast_fp16)[name = tensor("op_1649_cast_fp16")]; + tensor var_1650_cast_fp16 = mul(x = linear_106_cast_fp16, y = var_1649_cast_fp16)[name = tensor("op_1650_cast_fp16")]; + tensor input_281_to_fp16 = const()[name = tensor("input_281_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(19004992)))]; + tensor input_283_cast_fp16 = silu(x = input_281_to_fp16)[name = tensor("input_283_cast_fp16")]; + tensor linear_108_cast_fp16 = linear(bias = flow_net_time_embed_1_mlp_2_bias_to_fp16, weight = flow_net_time_embed_1_mlp_2_weight_to_fp16, x = input_283_cast_fp16)[name = tensor("linear_108_cast_fp16")]; + tensor reduce_mean_74_axes_0 = const()[name = tensor("reduce_mean_74_axes_0"), val = tensor([-1])]; + tensor reduce_mean_74_keep_dims_0 = const()[name = tensor("reduce_mean_74_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_74_cast_fp16 = reduce_mean(axes = reduce_mean_74_axes_0, keep_dims = reduce_mean_74_keep_dims_0, x = linear_108_cast_fp16)[name = tensor("reduce_mean_74_cast_fp16")]; + tensor sub_46_cast_fp16 = sub(x = linear_108_cast_fp16, y = reduce_mean_74_cast_fp16)[name = tensor("sub_46_cast_fp16")]; + tensor square_37_cast_fp16 = square(x = sub_46_cast_fp16)[name = tensor("square_37_cast_fp16")]; + tensor reduce_mean_75_axes_0 = const()[name = tensor("reduce_mean_75_axes_0"), val = tensor([-1])]; + tensor reduce_mean_75_keep_dims_0 = const()[name = tensor("reduce_mean_75_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_75_cast_fp16 = reduce_mean(axes = reduce_mean_75_axes_0, keep_dims = reduce_mean_75_keep_dims_0, x = square_37_cast_fp16)[name = tensor("reduce_mean_75_cast_fp16")]; + tensor real_div_9_to_fp16 = const()[name = tensor("real_div_9_to_fp16"), val = tensor(0x1.008p+0)]; + tensor mul_9_cast_fp16 = mul(x = reduce_mean_75_cast_fp16, y = real_div_9_to_fp16)[name = tensor("mul_9_cast_fp16")]; + tensor var_1682_to_fp16 = const()[name = tensor("op_1682_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_75_cast_fp16_0 = add(x = mul_9_cast_fp16, y = var_1682_to_fp16)[name = tensor("var_75_cast_fp16")]; + tensor var_1685_epsilon_0 = const()[name = tensor("op_1685_epsilon_0"), val = tensor(0x1.197998p-40)]; + tensor var_1685_cast_fp16 = rsqrt(epsilon = var_1685_epsilon_0, x = var_75_cast_fp16_0)[name = tensor("op_1685_cast_fp16")]; + tensor var_1686_cast_fp16 = mul(x = const_5_to_fp16, y = var_1685_cast_fp16)[name = tensor("op_1686_cast_fp16")]; + tensor var_1687_cast_fp16 = mul(x = linear_108_cast_fp16, y = var_1686_cast_fp16)[name = tensor("op_1687_cast_fp16")]; + tensor var_1699_cast_fp16 = add(x = var_1650_cast_fp16, y = var_1687_cast_fp16)[name = tensor("op_1699_cast_fp16")]; + tensor _inversed_t_combined_9_y_0_to_fp16 = const()[name = tensor("_inversed_t_combined_9_y_0_to_fp16"), val = tensor(0x1p-1)]; + tensor _inversed_t_combined_9_cast_fp16 = mul(x = var_1699_cast_fp16, y = _inversed_t_combined_9_y_0_to_fp16)[name = tensor("_inversed_t_combined_9_cast_fp16")]; + tensor input_285_cast_fp16 = add(x = _inversed_t_combined_9_cast_fp16, y = linear_5_cast_fp16)[name = tensor("input_285_cast_fp16")]; + tensor input_287_cast_fp16 = silu(x = input_285_cast_fp16)[name = tensor("input_287_cast_fp16")]; + tensor linear_110_cast_fp16 = linear(bias = flow_net_res_blocks_0_adaLN_modulation_1_bias_to_fp16, weight = flow_net_res_blocks_0_adaLN_modulation_1_weight_to_fp16, x = input_287_cast_fp16)[name = tensor("linear_110_cast_fp16")]; + tensor var_1714_split_sizes_0 = const()[name = tensor("op_1714_split_sizes_0"), val = tensor([512, 512, 512])]; + tensor var_1714_axis_0 = const()[name = tensor("op_1714_axis_0"), val = tensor(-1)]; + tensor var_1714_cast_fp16_0, tensor var_1714_cast_fp16_1, tensor var_1714_cast_fp16_2 = split(axis = var_1714_axis_0, split_sizes = var_1714_split_sizes_0, x = linear_110_cast_fp16)[name = tensor("op_1714_cast_fp16")]; + tensor mean_57_axes_0 = const()[name = tensor("mean_57_axes_0"), val = tensor([-1])]; + tensor mean_57_keep_dims_0 = const()[name = tensor("mean_57_keep_dims_0"), val = tensor(true)]; + tensor mean_57_cast_fp16 = reduce_mean(axes = mean_57_axes_0, keep_dims = mean_57_keep_dims_0, x = linear_104_cast_fp16)[name = tensor("mean_57_cast_fp16")]; + tensor sub_48_cast_fp16 = sub(x = linear_104_cast_fp16, y = mean_57_cast_fp16)[name = tensor("sub_48_cast_fp16")]; + tensor square_38_cast_fp16 = square(x = sub_48_cast_fp16)[name = tensor("square_38_cast_fp16")]; + tensor reduce_mean_77_axes_0 = const()[name = tensor("reduce_mean_77_axes_0"), val = tensor([-1])]; + tensor reduce_mean_77_keep_dims_0 = const()[name = tensor("reduce_mean_77_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_77_cast_fp16 = reduce_mean(axes = reduce_mean_77_axes_0, keep_dims = reduce_mean_77_keep_dims_0, x = square_38_cast_fp16)[name = tensor("reduce_mean_77_cast_fp16")]; + tensor var_1724_to_fp16 = const()[name = tensor("op_1724_to_fp16"), val = tensor(0x1.1p-20)]; + tensor var_1725_cast_fp16 = add(x = reduce_mean_77_cast_fp16, y = var_1724_to_fp16)[name = tensor("op_1725_cast_fp16")]; + tensor var_1726_cast_fp16 = sqrt(x = var_1725_cast_fp16)[name = tensor("op_1726_cast_fp16")]; + tensor x_183_cast_fp16 = real_div(x = sub_48_cast_fp16, y = var_1726_cast_fp16)[name = tensor("x_183_cast_fp16")]; + tensor var_1728_cast_fp16 = mul(x = x_183_cast_fp16, y = flow_net_res_blocks_0_in_ln_weight_to_fp16)[name = tensor("op_1728_cast_fp16")]; + tensor x_185_cast_fp16 = add(x = var_1728_cast_fp16, y = flow_net_res_blocks_0_in_ln_bias_to_fp16)[name = tensor("x_185_cast_fp16")]; + tensor var_1730_promoted_to_fp16 = const()[name = tensor("op_1730_promoted_to_fp16"), val = tensor(0x1p+0)]; + tensor var_1731_cast_fp16 = add(x = var_1714_cast_fp16_1, y = var_1730_promoted_to_fp16)[name = tensor("op_1731_cast_fp16")]; + tensor var_1732_cast_fp16 = mul(x = x_185_cast_fp16, y = var_1731_cast_fp16)[name = tensor("op_1732_cast_fp16")]; + tensor input_289_cast_fp16 = add(x = var_1732_cast_fp16, y = var_1714_cast_fp16_0)[name = tensor("input_289_cast_fp16")]; + tensor linear_111_cast_fp16 = linear(bias = flow_net_res_blocks_0_mlp_0_bias_to_fp16, weight = flow_net_res_blocks_0_mlp_0_weight_to_fp16, x = input_289_cast_fp16)[name = tensor("linear_111_cast_fp16")]; + tensor input_293_cast_fp16 = silu(x = linear_111_cast_fp16)[name = tensor("input_293_cast_fp16")]; + tensor linear_112_cast_fp16 = linear(bias = flow_net_res_blocks_0_mlp_2_bias_to_fp16, weight = flow_net_res_blocks_0_mlp_2_weight_to_fp16, x = input_293_cast_fp16)[name = tensor("linear_112_cast_fp16")]; + tensor var_1743_cast_fp16 = mul(x = var_1714_cast_fp16_2, y = linear_112_cast_fp16)[name = tensor("op_1743_cast_fp16")]; + tensor x_187_cast_fp16 = add(x = linear_104_cast_fp16, y = var_1743_cast_fp16)[name = tensor("x_187_cast_fp16")]; + tensor linear_113_cast_fp16 = linear(bias = flow_net_res_blocks_1_adaLN_modulation_1_bias_to_fp16, weight = flow_net_res_blocks_1_adaLN_modulation_1_weight_to_fp16, x = input_287_cast_fp16)[name = tensor("linear_113_cast_fp16")]; + tensor var_1753_split_sizes_0 = const()[name = tensor("op_1753_split_sizes_0"), val = tensor([512, 512, 512])]; + tensor var_1753_axis_0 = const()[name = tensor("op_1753_axis_0"), val = tensor(-1)]; + tensor var_1753_cast_fp16_0, tensor var_1753_cast_fp16_1, tensor var_1753_cast_fp16_2 = split(axis = var_1753_axis_0, split_sizes = var_1753_split_sizes_0, x = linear_113_cast_fp16)[name = tensor("op_1753_cast_fp16")]; + tensor mean_59_axes_0 = const()[name = tensor("mean_59_axes_0"), val = tensor([-1])]; + tensor mean_59_keep_dims_0 = const()[name = tensor("mean_59_keep_dims_0"), val = tensor(true)]; + tensor mean_59_cast_fp16 = reduce_mean(axes = mean_59_axes_0, keep_dims = mean_59_keep_dims_0, x = x_187_cast_fp16)[name = tensor("mean_59_cast_fp16")]; + tensor sub_49_cast_fp16 = sub(x = x_187_cast_fp16, y = mean_59_cast_fp16)[name = tensor("sub_49_cast_fp16")]; + tensor square_39_cast_fp16 = square(x = sub_49_cast_fp16)[name = tensor("square_39_cast_fp16")]; + tensor reduce_mean_79_axes_0 = const()[name = tensor("reduce_mean_79_axes_0"), val = tensor([-1])]; + tensor reduce_mean_79_keep_dims_0 = const()[name = tensor("reduce_mean_79_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_79_cast_fp16 = reduce_mean(axes = reduce_mean_79_axes_0, keep_dims = reduce_mean_79_keep_dims_0, x = square_39_cast_fp16)[name = tensor("reduce_mean_79_cast_fp16")]; + tensor var_1763_to_fp16 = const()[name = tensor("op_1763_to_fp16"), val = tensor(0x1.1p-20)]; + tensor var_1764_cast_fp16 = add(x = reduce_mean_79_cast_fp16, y = var_1763_to_fp16)[name = tensor("op_1764_cast_fp16")]; + tensor var_1765_cast_fp16 = sqrt(x = var_1764_cast_fp16)[name = tensor("op_1765_cast_fp16")]; + tensor x_189_cast_fp16 = real_div(x = sub_49_cast_fp16, y = var_1765_cast_fp16)[name = tensor("x_189_cast_fp16")]; + tensor var_1767_cast_fp16 = mul(x = x_189_cast_fp16, y = flow_net_res_blocks_1_in_ln_weight_to_fp16)[name = tensor("op_1767_cast_fp16")]; + tensor x_191_cast_fp16 = add(x = var_1767_cast_fp16, y = flow_net_res_blocks_1_in_ln_bias_to_fp16)[name = tensor("x_191_cast_fp16")]; + tensor var_1769_promoted_to_fp16 = const()[name = tensor("op_1769_promoted_to_fp16"), val = tensor(0x1p+0)]; + tensor var_1770_cast_fp16 = add(x = var_1753_cast_fp16_1, y = var_1769_promoted_to_fp16)[name = tensor("op_1770_cast_fp16")]; + tensor var_1771_cast_fp16 = mul(x = x_191_cast_fp16, y = var_1770_cast_fp16)[name = tensor("op_1771_cast_fp16")]; + tensor input_297_cast_fp16 = add(x = var_1771_cast_fp16, y = var_1753_cast_fp16_0)[name = tensor("input_297_cast_fp16")]; + tensor linear_114_cast_fp16 = linear(bias = flow_net_res_blocks_1_mlp_0_bias_to_fp16, weight = flow_net_res_blocks_1_mlp_0_weight_to_fp16, x = input_297_cast_fp16)[name = tensor("linear_114_cast_fp16")]; + tensor input_301_cast_fp16 = silu(x = linear_114_cast_fp16)[name = tensor("input_301_cast_fp16")]; + tensor linear_115_cast_fp16 = linear(bias = flow_net_res_blocks_1_mlp_2_bias_to_fp16, weight = flow_net_res_blocks_1_mlp_2_weight_to_fp16, x = input_301_cast_fp16)[name = tensor("linear_115_cast_fp16")]; + tensor var_1782_cast_fp16 = mul(x = var_1753_cast_fp16_2, y = linear_115_cast_fp16)[name = tensor("op_1782_cast_fp16")]; + tensor x_193_cast_fp16 = add(x = x_187_cast_fp16, y = var_1782_cast_fp16)[name = tensor("x_193_cast_fp16")]; + tensor linear_116_cast_fp16 = linear(bias = flow_net_res_blocks_2_adaLN_modulation_1_bias_to_fp16, weight = flow_net_res_blocks_2_adaLN_modulation_1_weight_to_fp16, x = input_287_cast_fp16)[name = tensor("linear_116_cast_fp16")]; + tensor var_1792_split_sizes_0 = const()[name = tensor("op_1792_split_sizes_0"), val = tensor([512, 512, 512])]; + tensor var_1792_axis_0 = const()[name = tensor("op_1792_axis_0"), val = tensor(-1)]; + tensor var_1792_cast_fp16_0, tensor var_1792_cast_fp16_1, tensor var_1792_cast_fp16_2 = split(axis = var_1792_axis_0, split_sizes = var_1792_split_sizes_0, x = linear_116_cast_fp16)[name = tensor("op_1792_cast_fp16")]; + tensor mean_61_axes_0 = const()[name = tensor("mean_61_axes_0"), val = tensor([-1])]; + tensor mean_61_keep_dims_0 = const()[name = tensor("mean_61_keep_dims_0"), val = tensor(true)]; + tensor mean_61_cast_fp16 = reduce_mean(axes = mean_61_axes_0, keep_dims = mean_61_keep_dims_0, x = x_193_cast_fp16)[name = tensor("mean_61_cast_fp16")]; + tensor sub_50_cast_fp16 = sub(x = x_193_cast_fp16, y = mean_61_cast_fp16)[name = tensor("sub_50_cast_fp16")]; + tensor square_40_cast_fp16 = square(x = sub_50_cast_fp16)[name = tensor("square_40_cast_fp16")]; + tensor reduce_mean_81_axes_0 = const()[name = tensor("reduce_mean_81_axes_0"), val = tensor([-1])]; + tensor reduce_mean_81_keep_dims_0 = const()[name = tensor("reduce_mean_81_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_81_cast_fp16 = reduce_mean(axes = reduce_mean_81_axes_0, keep_dims = reduce_mean_81_keep_dims_0, x = square_40_cast_fp16)[name = tensor("reduce_mean_81_cast_fp16")]; + tensor var_1802_to_fp16 = const()[name = tensor("op_1802_to_fp16"), val = tensor(0x1.1p-20)]; + tensor var_1803_cast_fp16 = add(x = reduce_mean_81_cast_fp16, y = var_1802_to_fp16)[name = tensor("op_1803_cast_fp16")]; + tensor var_1804_cast_fp16 = sqrt(x = var_1803_cast_fp16)[name = tensor("op_1804_cast_fp16")]; + tensor x_195_cast_fp16 = real_div(x = sub_50_cast_fp16, y = var_1804_cast_fp16)[name = tensor("x_195_cast_fp16")]; + tensor var_1806_cast_fp16 = mul(x = x_195_cast_fp16, y = flow_net_res_blocks_2_in_ln_weight_to_fp16)[name = tensor("op_1806_cast_fp16")]; + tensor x_197_cast_fp16 = add(x = var_1806_cast_fp16, y = flow_net_res_blocks_2_in_ln_bias_to_fp16)[name = tensor("x_197_cast_fp16")]; + tensor var_1808_promoted_to_fp16 = const()[name = tensor("op_1808_promoted_to_fp16"), val = tensor(0x1p+0)]; + tensor var_1809_cast_fp16 = add(x = var_1792_cast_fp16_1, y = var_1808_promoted_to_fp16)[name = tensor("op_1809_cast_fp16")]; + tensor var_1810_cast_fp16 = mul(x = x_197_cast_fp16, y = var_1809_cast_fp16)[name = tensor("op_1810_cast_fp16")]; + tensor input_305_cast_fp16 = add(x = var_1810_cast_fp16, y = var_1792_cast_fp16_0)[name = tensor("input_305_cast_fp16")]; + tensor linear_117_cast_fp16 = linear(bias = flow_net_res_blocks_2_mlp_0_bias_to_fp16, weight = flow_net_res_blocks_2_mlp_0_weight_to_fp16, x = input_305_cast_fp16)[name = tensor("linear_117_cast_fp16")]; + tensor input_309_cast_fp16 = silu(x = linear_117_cast_fp16)[name = tensor("input_309_cast_fp16")]; + tensor linear_118_cast_fp16 = linear(bias = flow_net_res_blocks_2_mlp_2_bias_to_fp16, weight = flow_net_res_blocks_2_mlp_2_weight_to_fp16, x = input_309_cast_fp16)[name = tensor("linear_118_cast_fp16")]; + tensor var_1821_cast_fp16 = mul(x = var_1792_cast_fp16_2, y = linear_118_cast_fp16)[name = tensor("op_1821_cast_fp16")]; + tensor x_199_cast_fp16 = add(x = x_193_cast_fp16, y = var_1821_cast_fp16)[name = tensor("x_199_cast_fp16")]; + tensor linear_119_cast_fp16 = linear(bias = flow_net_res_blocks_3_adaLN_modulation_1_bias_to_fp16, weight = flow_net_res_blocks_3_adaLN_modulation_1_weight_to_fp16, x = input_287_cast_fp16)[name = tensor("linear_119_cast_fp16")]; + tensor var_1831_split_sizes_0 = const()[name = tensor("op_1831_split_sizes_0"), val = tensor([512, 512, 512])]; + tensor var_1831_axis_0 = const()[name = tensor("op_1831_axis_0"), val = tensor(-1)]; + tensor var_1831_cast_fp16_0, tensor var_1831_cast_fp16_1, tensor var_1831_cast_fp16_2 = split(axis = var_1831_axis_0, split_sizes = var_1831_split_sizes_0, x = linear_119_cast_fp16)[name = tensor("op_1831_cast_fp16")]; + tensor mean_63_axes_0 = const()[name = tensor("mean_63_axes_0"), val = tensor([-1])]; + tensor mean_63_keep_dims_0 = const()[name = tensor("mean_63_keep_dims_0"), val = tensor(true)]; + tensor mean_63_cast_fp16 = reduce_mean(axes = mean_63_axes_0, keep_dims = mean_63_keep_dims_0, x = x_199_cast_fp16)[name = tensor("mean_63_cast_fp16")]; + tensor sub_51_cast_fp16 = sub(x = x_199_cast_fp16, y = mean_63_cast_fp16)[name = tensor("sub_51_cast_fp16")]; + tensor square_41_cast_fp16 = square(x = sub_51_cast_fp16)[name = tensor("square_41_cast_fp16")]; + tensor reduce_mean_83_axes_0 = const()[name = tensor("reduce_mean_83_axes_0"), val = tensor([-1])]; + tensor reduce_mean_83_keep_dims_0 = const()[name = tensor("reduce_mean_83_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_83_cast_fp16 = reduce_mean(axes = reduce_mean_83_axes_0, keep_dims = reduce_mean_83_keep_dims_0, x = square_41_cast_fp16)[name = tensor("reduce_mean_83_cast_fp16")]; + tensor var_1841_to_fp16 = const()[name = tensor("op_1841_to_fp16"), val = tensor(0x1.1p-20)]; + tensor var_1842_cast_fp16 = add(x = reduce_mean_83_cast_fp16, y = var_1841_to_fp16)[name = tensor("op_1842_cast_fp16")]; + tensor var_1843_cast_fp16 = sqrt(x = var_1842_cast_fp16)[name = tensor("op_1843_cast_fp16")]; + tensor x_201_cast_fp16 = real_div(x = sub_51_cast_fp16, y = var_1843_cast_fp16)[name = tensor("x_201_cast_fp16")]; + tensor var_1845_cast_fp16 = mul(x = x_201_cast_fp16, y = flow_net_res_blocks_3_in_ln_weight_to_fp16)[name = tensor("op_1845_cast_fp16")]; + tensor x_203_cast_fp16 = add(x = var_1845_cast_fp16, y = flow_net_res_blocks_3_in_ln_bias_to_fp16)[name = tensor("x_203_cast_fp16")]; + tensor var_1847_promoted_to_fp16 = const()[name = tensor("op_1847_promoted_to_fp16"), val = tensor(0x1p+0)]; + tensor var_1848_cast_fp16 = add(x = var_1831_cast_fp16_1, y = var_1847_promoted_to_fp16)[name = tensor("op_1848_cast_fp16")]; + tensor var_1849_cast_fp16 = mul(x = x_203_cast_fp16, y = var_1848_cast_fp16)[name = tensor("op_1849_cast_fp16")]; + tensor input_313_cast_fp16 = add(x = var_1849_cast_fp16, y = var_1831_cast_fp16_0)[name = tensor("input_313_cast_fp16")]; + tensor linear_120_cast_fp16 = linear(bias = flow_net_res_blocks_3_mlp_0_bias_to_fp16, weight = flow_net_res_blocks_3_mlp_0_weight_to_fp16, x = input_313_cast_fp16)[name = tensor("linear_120_cast_fp16")]; + tensor input_317_cast_fp16 = silu(x = linear_120_cast_fp16)[name = tensor("input_317_cast_fp16")]; + tensor linear_121_cast_fp16 = linear(bias = flow_net_res_blocks_3_mlp_2_bias_to_fp16, weight = flow_net_res_blocks_3_mlp_2_weight_to_fp16, x = input_317_cast_fp16)[name = tensor("linear_121_cast_fp16")]; + tensor var_1860_cast_fp16 = mul(x = var_1831_cast_fp16_2, y = linear_121_cast_fp16)[name = tensor("op_1860_cast_fp16")]; + tensor x_205_cast_fp16 = add(x = x_199_cast_fp16, y = var_1860_cast_fp16)[name = tensor("x_205_cast_fp16")]; + tensor linear_122_cast_fp16 = linear(bias = flow_net_res_blocks_4_adaLN_modulation_1_bias_to_fp16, weight = flow_net_res_blocks_4_adaLN_modulation_1_weight_to_fp16, x = input_287_cast_fp16)[name = tensor("linear_122_cast_fp16")]; + tensor var_1870_split_sizes_0 = const()[name = tensor("op_1870_split_sizes_0"), val = tensor([512, 512, 512])]; + tensor var_1870_axis_0 = const()[name = tensor("op_1870_axis_0"), val = tensor(-1)]; + tensor var_1870_cast_fp16_0, tensor var_1870_cast_fp16_1, tensor var_1870_cast_fp16_2 = split(axis = var_1870_axis_0, split_sizes = var_1870_split_sizes_0, x = linear_122_cast_fp16)[name = tensor("op_1870_cast_fp16")]; + tensor mean_65_axes_0 = const()[name = tensor("mean_65_axes_0"), val = tensor([-1])]; + tensor mean_65_keep_dims_0 = const()[name = tensor("mean_65_keep_dims_0"), val = tensor(true)]; + tensor mean_65_cast_fp16 = reduce_mean(axes = mean_65_axes_0, keep_dims = mean_65_keep_dims_0, x = x_205_cast_fp16)[name = tensor("mean_65_cast_fp16")]; + tensor sub_52_cast_fp16 = sub(x = x_205_cast_fp16, y = mean_65_cast_fp16)[name = tensor("sub_52_cast_fp16")]; + tensor square_42_cast_fp16 = square(x = sub_52_cast_fp16)[name = tensor("square_42_cast_fp16")]; + tensor reduce_mean_85_axes_0 = const()[name = tensor("reduce_mean_85_axes_0"), val = tensor([-1])]; + tensor reduce_mean_85_keep_dims_0 = const()[name = tensor("reduce_mean_85_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_85_cast_fp16 = reduce_mean(axes = reduce_mean_85_axes_0, keep_dims = reduce_mean_85_keep_dims_0, x = square_42_cast_fp16)[name = tensor("reduce_mean_85_cast_fp16")]; + tensor var_1880_to_fp16 = const()[name = tensor("op_1880_to_fp16"), val = tensor(0x1.1p-20)]; + tensor var_1881_cast_fp16 = add(x = reduce_mean_85_cast_fp16, y = var_1880_to_fp16)[name = tensor("op_1881_cast_fp16")]; + tensor var_1882_cast_fp16 = sqrt(x = var_1881_cast_fp16)[name = tensor("op_1882_cast_fp16")]; + tensor x_207_cast_fp16 = real_div(x = sub_52_cast_fp16, y = var_1882_cast_fp16)[name = tensor("x_207_cast_fp16")]; + tensor var_1884_cast_fp16 = mul(x = x_207_cast_fp16, y = flow_net_res_blocks_4_in_ln_weight_to_fp16)[name = tensor("op_1884_cast_fp16")]; + tensor x_209_cast_fp16 = add(x = var_1884_cast_fp16, y = flow_net_res_blocks_4_in_ln_bias_to_fp16)[name = tensor("x_209_cast_fp16")]; + tensor var_1886_promoted_to_fp16 = const()[name = tensor("op_1886_promoted_to_fp16"), val = tensor(0x1p+0)]; + tensor var_1887_cast_fp16 = add(x = var_1870_cast_fp16_1, y = var_1886_promoted_to_fp16)[name = tensor("op_1887_cast_fp16")]; + tensor var_1888_cast_fp16 = mul(x = x_209_cast_fp16, y = var_1887_cast_fp16)[name = tensor("op_1888_cast_fp16")]; + tensor input_321_cast_fp16 = add(x = var_1888_cast_fp16, y = var_1870_cast_fp16_0)[name = tensor("input_321_cast_fp16")]; + tensor linear_123_cast_fp16 = linear(bias = flow_net_res_blocks_4_mlp_0_bias_to_fp16, weight = flow_net_res_blocks_4_mlp_0_weight_to_fp16, x = input_321_cast_fp16)[name = tensor("linear_123_cast_fp16")]; + tensor input_325_cast_fp16 = silu(x = linear_123_cast_fp16)[name = tensor("input_325_cast_fp16")]; + tensor linear_124_cast_fp16 = linear(bias = flow_net_res_blocks_4_mlp_2_bias_to_fp16, weight = flow_net_res_blocks_4_mlp_2_weight_to_fp16, x = input_325_cast_fp16)[name = tensor("linear_124_cast_fp16")]; + tensor var_1899_cast_fp16 = mul(x = var_1870_cast_fp16_2, y = linear_124_cast_fp16)[name = tensor("op_1899_cast_fp16")]; + tensor x_211_cast_fp16 = add(x = x_205_cast_fp16, y = var_1899_cast_fp16)[name = tensor("x_211_cast_fp16")]; + tensor linear_125_cast_fp16 = linear(bias = flow_net_res_blocks_5_adaLN_modulation_1_bias_to_fp16, weight = flow_net_res_blocks_5_adaLN_modulation_1_weight_to_fp16, x = input_287_cast_fp16)[name = tensor("linear_125_cast_fp16")]; + tensor var_1909_split_sizes_0 = const()[name = tensor("op_1909_split_sizes_0"), val = tensor([512, 512, 512])]; + tensor var_1909_axis_0 = const()[name = tensor("op_1909_axis_0"), val = tensor(-1)]; + tensor var_1909_cast_fp16_0, tensor var_1909_cast_fp16_1, tensor var_1909_cast_fp16_2 = split(axis = var_1909_axis_0, split_sizes = var_1909_split_sizes_0, x = linear_125_cast_fp16)[name = tensor("op_1909_cast_fp16")]; + tensor mean_67_axes_0 = const()[name = tensor("mean_67_axes_0"), val = tensor([-1])]; + tensor mean_67_keep_dims_0 = const()[name = tensor("mean_67_keep_dims_0"), val = tensor(true)]; + tensor mean_67_cast_fp16 = reduce_mean(axes = mean_67_axes_0, keep_dims = mean_67_keep_dims_0, x = x_211_cast_fp16)[name = tensor("mean_67_cast_fp16")]; + tensor sub_53_cast_fp16 = sub(x = x_211_cast_fp16, y = mean_67_cast_fp16)[name = tensor("sub_53_cast_fp16")]; + tensor square_43_cast_fp16 = square(x = sub_53_cast_fp16)[name = tensor("square_43_cast_fp16")]; + tensor reduce_mean_87_axes_0 = const()[name = tensor("reduce_mean_87_axes_0"), val = tensor([-1])]; + tensor reduce_mean_87_keep_dims_0 = const()[name = tensor("reduce_mean_87_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_87_cast_fp16 = reduce_mean(axes = reduce_mean_87_axes_0, keep_dims = reduce_mean_87_keep_dims_0, x = square_43_cast_fp16)[name = tensor("reduce_mean_87_cast_fp16")]; + tensor var_1919_to_fp16 = const()[name = tensor("op_1919_to_fp16"), val = tensor(0x1.1p-20)]; + tensor var_1920_cast_fp16 = add(x = reduce_mean_87_cast_fp16, y = var_1919_to_fp16)[name = tensor("op_1920_cast_fp16")]; + tensor var_1921_cast_fp16 = sqrt(x = var_1920_cast_fp16)[name = tensor("op_1921_cast_fp16")]; + tensor x_213_cast_fp16 = real_div(x = sub_53_cast_fp16, y = var_1921_cast_fp16)[name = tensor("x_213_cast_fp16")]; + tensor var_1923_cast_fp16 = mul(x = x_213_cast_fp16, y = flow_net_res_blocks_5_in_ln_weight_to_fp16)[name = tensor("op_1923_cast_fp16")]; + tensor x_215_cast_fp16 = add(x = var_1923_cast_fp16, y = flow_net_res_blocks_5_in_ln_bias_to_fp16)[name = tensor("x_215_cast_fp16")]; + tensor var_1925_promoted_to_fp16 = const()[name = tensor("op_1925_promoted_to_fp16"), val = tensor(0x1p+0)]; + tensor var_1926_cast_fp16 = add(x = var_1909_cast_fp16_1, y = var_1925_promoted_to_fp16)[name = tensor("op_1926_cast_fp16")]; + tensor var_1927_cast_fp16 = mul(x = x_215_cast_fp16, y = var_1926_cast_fp16)[name = tensor("op_1927_cast_fp16")]; + tensor input_329_cast_fp16 = add(x = var_1927_cast_fp16, y = var_1909_cast_fp16_0)[name = tensor("input_329_cast_fp16")]; + tensor linear_126_cast_fp16 = linear(bias = flow_net_res_blocks_5_mlp_0_bias_to_fp16, weight = flow_net_res_blocks_5_mlp_0_weight_to_fp16, x = input_329_cast_fp16)[name = tensor("linear_126_cast_fp16")]; + tensor input_333_cast_fp16 = silu(x = linear_126_cast_fp16)[name = tensor("input_333_cast_fp16")]; + tensor linear_127_cast_fp16 = linear(bias = flow_net_res_blocks_5_mlp_2_bias_to_fp16, weight = flow_net_res_blocks_5_mlp_2_weight_to_fp16, x = input_333_cast_fp16)[name = tensor("linear_127_cast_fp16")]; + tensor var_1938_cast_fp16 = mul(x = var_1909_cast_fp16_2, y = linear_127_cast_fp16)[name = tensor("op_1938_cast_fp16")]; + tensor x_217_cast_fp16 = add(x = x_211_cast_fp16, y = var_1938_cast_fp16)[name = tensor("x_217_cast_fp16")]; + tensor linear_128_cast_fp16 = linear(bias = flow_net_final_layer_adaLN_modulation_1_bias_to_fp16, weight = flow_net_final_layer_adaLN_modulation_1_weight_to_fp16, x = input_287_cast_fp16)[name = tensor("linear_128_cast_fp16")]; + tensor var_1947_split_sizes_0 = const()[name = tensor("op_1947_split_sizes_0"), val = tensor([512, 512])]; + tensor var_1947_axis_0 = const()[name = tensor("op_1947_axis_0"), val = tensor(-1)]; + tensor var_1947_cast_fp16_0, tensor var_1947_cast_fp16_1 = split(axis = var_1947_axis_0, split_sizes = var_1947_split_sizes_0, x = linear_128_cast_fp16)[name = tensor("op_1947_cast_fp16")]; + tensor mean_69_axes_0 = const()[name = tensor("mean_69_axes_0"), val = tensor([-1])]; + tensor mean_69_keep_dims_0 = const()[name = tensor("mean_69_keep_dims_0"), val = tensor(true)]; + tensor mean_69_cast_fp16 = reduce_mean(axes = mean_69_axes_0, keep_dims = mean_69_keep_dims_0, x = x_217_cast_fp16)[name = tensor("mean_69_cast_fp16")]; + tensor sub_54_cast_fp16 = sub(x = x_217_cast_fp16, y = mean_69_cast_fp16)[name = tensor("sub_54_cast_fp16")]; + tensor square_44_cast_fp16 = square(x = sub_54_cast_fp16)[name = tensor("square_44_cast_fp16")]; + tensor reduce_mean_89_axes_0 = const()[name = tensor("reduce_mean_89_axes_0"), val = tensor([-1])]; + tensor reduce_mean_89_keep_dims_0 = const()[name = tensor("reduce_mean_89_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_89_cast_fp16 = reduce_mean(axes = reduce_mean_89_axes_0, keep_dims = reduce_mean_89_keep_dims_0, x = square_44_cast_fp16)[name = tensor("reduce_mean_89_cast_fp16")]; + tensor var_1954_to_fp16 = const()[name = tensor("op_1954_to_fp16"), val = tensor(0x1.1p-20)]; + tensor var_1955_cast_fp16 = add(x = reduce_mean_89_cast_fp16, y = var_1954_to_fp16)[name = tensor("op_1955_cast_fp16")]; + tensor var_1956_cast_fp16 = sqrt(x = var_1955_cast_fp16)[name = tensor("op_1956_cast_fp16")]; + tensor x_219_cast_fp16 = real_div(x = sub_54_cast_fp16, y = var_1956_cast_fp16)[name = tensor("x_219_cast_fp16")]; + tensor var_1958_promoted_to_fp16 = const()[name = tensor("op_1958_promoted_to_fp16"), val = tensor(0x1p+0)]; + tensor var_1959_cast_fp16 = add(x = var_1947_cast_fp16_1, y = var_1958_promoted_to_fp16)[name = tensor("op_1959_cast_fp16")]; + tensor var_1960_cast_fp16 = mul(x = x_219_cast_fp16, y = var_1959_cast_fp16)[name = tensor("op_1960_cast_fp16")]; + tensor input_337_cast_fp16 = add(x = var_1960_cast_fp16, y = var_1947_cast_fp16_0)[name = tensor("input_337_cast_fp16")]; + tensor linear_129_cast_fp16 = linear(bias = flow_net_final_layer_linear_bias_to_fp16, weight = flow_net_final_layer_linear_weight_to_fp16, x = input_337_cast_fp16)[name = tensor("linear_129_cast_fp16")]; + tensor var_1971_to_fp16 = const()[name = tensor("op_1971_to_fp16"), val = tensor(0x1p-3)]; + tensor var_1972_cast_fp16 = mul(x = linear_129_cast_fp16, y = var_1971_to_fp16)[name = tensor("op_1972_cast_fp16")]; + tensor input_339_cast_fp16 = add(x = input_271_cast_fp16, y = var_1972_cast_fp16)[name = tensor("input_339_cast_fp16")]; + tensor linear_130_cast_fp16 = linear(bias = flow_net_input_proj_bias_to_fp16, weight = flow_net_input_proj_weight_to_fp16, x = input_339_cast_fp16)[name = tensor("linear_130_cast_fp16")]; + tensor input_343_to_fp16 = const()[name = tensor("input_343_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(19006080)))]; + tensor input_345_cast_fp16 = silu(x = input_343_to_fp16)[name = tensor("input_345_cast_fp16")]; + tensor linear_132_cast_fp16 = linear(bias = flow_net_time_embed_0_mlp_2_bias_to_fp16, weight = flow_net_time_embed_0_mlp_2_weight_to_fp16, x = input_345_cast_fp16)[name = tensor("linear_132_cast_fp16")]; + tensor reduce_mean_90_axes_0 = const()[name = tensor("reduce_mean_90_axes_0"), val = tensor([-1])]; + tensor reduce_mean_90_keep_dims_0 = const()[name = tensor("reduce_mean_90_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_90_cast_fp16 = reduce_mean(axes = reduce_mean_90_axes_0, keep_dims = reduce_mean_90_keep_dims_0, x = linear_132_cast_fp16)[name = tensor("reduce_mean_90_cast_fp16")]; + tensor sub_55_cast_fp16 = sub(x = linear_132_cast_fp16, y = reduce_mean_90_cast_fp16)[name = tensor("sub_55_cast_fp16")]; + tensor square_45_cast_fp16 = square(x = sub_55_cast_fp16)[name = tensor("square_45_cast_fp16")]; + tensor reduce_mean_91_axes_0 = const()[name = tensor("reduce_mean_91_axes_0"), val = tensor([-1])]; + tensor reduce_mean_91_keep_dims_0 = const()[name = tensor("reduce_mean_91_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_91_cast_fp16 = reduce_mean(axes = reduce_mean_91_axes_0, keep_dims = reduce_mean_91_keep_dims_0, x = square_45_cast_fp16)[name = tensor("reduce_mean_91_cast_fp16")]; + tensor real_div_10_to_fp16 = const()[name = tensor("real_div_10_to_fp16"), val = tensor(0x1.008p+0)]; + tensor mul_10_cast_fp16 = mul(x = reduce_mean_91_cast_fp16, y = real_div_10_to_fp16)[name = tensor("mul_10_cast_fp16")]; + tensor var_2038_to_fp16 = const()[name = tensor("op_2038_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_91_cast_fp16 = add(x = mul_10_cast_fp16, y = var_2038_to_fp16)[name = tensor("var_91_cast_fp16")]; + tensor var_2041_epsilon_0 = const()[name = tensor("op_2041_epsilon_0"), val = tensor(0x1.197998p-40)]; + tensor var_2041_cast_fp16 = rsqrt(epsilon = var_2041_epsilon_0, x = var_91_cast_fp16)[name = tensor("op_2041_cast_fp16")]; + tensor var_2042_cast_fp16 = mul(x = const_3_to_fp16, y = var_2041_cast_fp16)[name = tensor("op_2042_cast_fp16")]; + tensor var_2043_cast_fp16 = mul(x = linear_132_cast_fp16, y = var_2042_cast_fp16)[name = tensor("op_2043_cast_fp16")]; + tensor input_349_to_fp16 = const()[name = tensor("input_349_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(19007168)))]; + tensor input_351_cast_fp16 = silu(x = input_349_to_fp16)[name = tensor("input_351_cast_fp16")]; + tensor linear_134_cast_fp16 = linear(bias = flow_net_time_embed_1_mlp_2_bias_to_fp16, weight = flow_net_time_embed_1_mlp_2_weight_to_fp16, x = input_351_cast_fp16)[name = tensor("linear_134_cast_fp16")]; + tensor reduce_mean_92_axes_0 = const()[name = tensor("reduce_mean_92_axes_0"), val = tensor([-1])]; + tensor reduce_mean_92_keep_dims_0 = const()[name = tensor("reduce_mean_92_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_92_cast_fp16 = reduce_mean(axes = reduce_mean_92_axes_0, keep_dims = reduce_mean_92_keep_dims_0, x = linear_134_cast_fp16)[name = tensor("reduce_mean_92_cast_fp16")]; + tensor sub_57_cast_fp16 = sub(x = linear_134_cast_fp16, y = reduce_mean_92_cast_fp16)[name = tensor("sub_57_cast_fp16")]; + tensor square_46_cast_fp16 = square(x = sub_57_cast_fp16)[name = tensor("square_46_cast_fp16")]; + tensor reduce_mean_93_axes_0 = const()[name = tensor("reduce_mean_93_axes_0"), val = tensor([-1])]; + tensor reduce_mean_93_keep_dims_0 = const()[name = tensor("reduce_mean_93_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_93_cast_fp16 = reduce_mean(axes = reduce_mean_93_axes_0, keep_dims = reduce_mean_93_keep_dims_0, x = square_46_cast_fp16)[name = tensor("reduce_mean_93_cast_fp16")]; + tensor real_div_11_to_fp16 = const()[name = tensor("real_div_11_to_fp16"), val = tensor(0x1.008p+0)]; + tensor mul_11_cast_fp16 = mul(x = reduce_mean_93_cast_fp16, y = real_div_11_to_fp16)[name = tensor("mul_11_cast_fp16")]; + tensor var_2075_to_fp16 = const()[name = tensor("op_2075_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_93_cast_fp16 = add(x = mul_11_cast_fp16, y = var_2075_to_fp16)[name = tensor("var_93_cast_fp16")]; + tensor var_2078_epsilon_0 = const()[name = tensor("op_2078_epsilon_0"), val = tensor(0x1.197998p-40)]; + tensor var_2078_cast_fp16 = rsqrt(epsilon = var_2078_epsilon_0, x = var_93_cast_fp16)[name = tensor("op_2078_cast_fp16")]; + tensor var_2079_cast_fp16 = mul(x = const_5_to_fp16, y = var_2078_cast_fp16)[name = tensor("op_2079_cast_fp16")]; + tensor var_2080_cast_fp16 = mul(x = linear_134_cast_fp16, y = var_2079_cast_fp16)[name = tensor("op_2080_cast_fp16")]; + tensor var_2092_cast_fp16 = add(x = var_2043_cast_fp16, y = var_2080_cast_fp16)[name = tensor("op_2092_cast_fp16")]; + tensor _inversed_t_combined_11_y_0_to_fp16 = const()[name = tensor("_inversed_t_combined_11_y_0_to_fp16"), val = tensor(0x1p-1)]; + tensor _inversed_t_combined_11_cast_fp16 = mul(x = var_2092_cast_fp16, y = _inversed_t_combined_11_y_0_to_fp16)[name = tensor("_inversed_t_combined_11_cast_fp16")]; + tensor input_353_cast_fp16 = add(x = _inversed_t_combined_11_cast_fp16, y = linear_5_cast_fp16)[name = tensor("input_353_cast_fp16")]; + tensor input_355_cast_fp16 = silu(x = input_353_cast_fp16)[name = tensor("input_355_cast_fp16")]; + tensor linear_136_cast_fp16 = linear(bias = flow_net_res_blocks_0_adaLN_modulation_1_bias_to_fp16, weight = flow_net_res_blocks_0_adaLN_modulation_1_weight_to_fp16, x = input_355_cast_fp16)[name = tensor("linear_136_cast_fp16")]; + tensor var_2107_split_sizes_0 = const()[name = tensor("op_2107_split_sizes_0"), val = tensor([512, 512, 512])]; + tensor var_2107_axis_0 = const()[name = tensor("op_2107_axis_0"), val = tensor(-1)]; + tensor var_2107_cast_fp16_0, tensor var_2107_cast_fp16_1, tensor var_2107_cast_fp16_2 = split(axis = var_2107_axis_0, split_sizes = var_2107_split_sizes_0, x = linear_136_cast_fp16)[name = tensor("op_2107_cast_fp16")]; + tensor mean_71_axes_0 = const()[name = tensor("mean_71_axes_0"), val = tensor([-1])]; + tensor mean_71_keep_dims_0 = const()[name = tensor("mean_71_keep_dims_0"), val = tensor(true)]; + tensor mean_71_cast_fp16 = reduce_mean(axes = mean_71_axes_0, keep_dims = mean_71_keep_dims_0, x = linear_130_cast_fp16)[name = tensor("mean_71_cast_fp16")]; + tensor sub_59_cast_fp16 = sub(x = linear_130_cast_fp16, y = mean_71_cast_fp16)[name = tensor("sub_59_cast_fp16")]; + tensor square_47_cast_fp16 = square(x = sub_59_cast_fp16)[name = tensor("square_47_cast_fp16")]; + tensor reduce_mean_95_axes_0 = const()[name = tensor("reduce_mean_95_axes_0"), val = tensor([-1])]; + tensor reduce_mean_95_keep_dims_0 = const()[name = tensor("reduce_mean_95_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_95_cast_fp16 = reduce_mean(axes = reduce_mean_95_axes_0, keep_dims = reduce_mean_95_keep_dims_0, x = square_47_cast_fp16)[name = tensor("reduce_mean_95_cast_fp16")]; + tensor var_2117_to_fp16 = const()[name = tensor("op_2117_to_fp16"), val = tensor(0x1.1p-20)]; + tensor var_2118_cast_fp16 = add(x = reduce_mean_95_cast_fp16, y = var_2117_to_fp16)[name = tensor("op_2118_cast_fp16")]; + tensor var_2119_cast_fp16 = sqrt(x = var_2118_cast_fp16)[name = tensor("op_2119_cast_fp16")]; + tensor x_227_cast_fp16 = real_div(x = sub_59_cast_fp16, y = var_2119_cast_fp16)[name = tensor("x_227_cast_fp16")]; + tensor var_2121_cast_fp16 = mul(x = x_227_cast_fp16, y = flow_net_res_blocks_0_in_ln_weight_to_fp16)[name = tensor("op_2121_cast_fp16")]; + tensor x_229_cast_fp16 = add(x = var_2121_cast_fp16, y = flow_net_res_blocks_0_in_ln_bias_to_fp16)[name = tensor("x_229_cast_fp16")]; + tensor var_2123_promoted_to_fp16 = const()[name = tensor("op_2123_promoted_to_fp16"), val = tensor(0x1p+0)]; + tensor var_2124_cast_fp16 = add(x = var_2107_cast_fp16_1, y = var_2123_promoted_to_fp16)[name = tensor("op_2124_cast_fp16")]; + tensor var_2125_cast_fp16 = mul(x = x_229_cast_fp16, y = var_2124_cast_fp16)[name = tensor("op_2125_cast_fp16")]; + tensor input_357_cast_fp16 = add(x = var_2125_cast_fp16, y = var_2107_cast_fp16_0)[name = tensor("input_357_cast_fp16")]; + tensor linear_137_cast_fp16 = linear(bias = flow_net_res_blocks_0_mlp_0_bias_to_fp16, weight = flow_net_res_blocks_0_mlp_0_weight_to_fp16, x = input_357_cast_fp16)[name = tensor("linear_137_cast_fp16")]; + tensor input_361_cast_fp16 = silu(x = linear_137_cast_fp16)[name = tensor("input_361_cast_fp16")]; + tensor linear_138_cast_fp16 = linear(bias = flow_net_res_blocks_0_mlp_2_bias_to_fp16, weight = flow_net_res_blocks_0_mlp_2_weight_to_fp16, x = input_361_cast_fp16)[name = tensor("linear_138_cast_fp16")]; + tensor var_2136_cast_fp16 = mul(x = var_2107_cast_fp16_2, y = linear_138_cast_fp16)[name = tensor("op_2136_cast_fp16")]; + tensor x_231_cast_fp16 = add(x = linear_130_cast_fp16, y = var_2136_cast_fp16)[name = tensor("x_231_cast_fp16")]; + tensor linear_139_cast_fp16 = linear(bias = flow_net_res_blocks_1_adaLN_modulation_1_bias_to_fp16, weight = flow_net_res_blocks_1_adaLN_modulation_1_weight_to_fp16, x = input_355_cast_fp16)[name = tensor("linear_139_cast_fp16")]; + tensor var_2146_split_sizes_0 = const()[name = tensor("op_2146_split_sizes_0"), val = tensor([512, 512, 512])]; + tensor var_2146_axis_0 = const()[name = tensor("op_2146_axis_0"), val = tensor(-1)]; + tensor var_2146_cast_fp16_0, tensor var_2146_cast_fp16_1, tensor var_2146_cast_fp16_2 = split(axis = var_2146_axis_0, split_sizes = var_2146_split_sizes_0, x = linear_139_cast_fp16)[name = tensor("op_2146_cast_fp16")]; + tensor mean_73_axes_0 = const()[name = tensor("mean_73_axes_0"), val = tensor([-1])]; + tensor mean_73_keep_dims_0 = const()[name = tensor("mean_73_keep_dims_0"), val = tensor(true)]; + tensor mean_73_cast_fp16 = reduce_mean(axes = mean_73_axes_0, keep_dims = mean_73_keep_dims_0, x = x_231_cast_fp16)[name = tensor("mean_73_cast_fp16")]; + tensor sub_60_cast_fp16 = sub(x = x_231_cast_fp16, y = mean_73_cast_fp16)[name = tensor("sub_60_cast_fp16")]; + tensor square_48_cast_fp16 = square(x = sub_60_cast_fp16)[name = tensor("square_48_cast_fp16")]; + tensor reduce_mean_97_axes_0 = const()[name = tensor("reduce_mean_97_axes_0"), val = tensor([-1])]; + tensor reduce_mean_97_keep_dims_0 = const()[name = tensor("reduce_mean_97_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_97_cast_fp16 = reduce_mean(axes = reduce_mean_97_axes_0, keep_dims = reduce_mean_97_keep_dims_0, x = square_48_cast_fp16)[name = tensor("reduce_mean_97_cast_fp16")]; + tensor var_2156_to_fp16 = const()[name = tensor("op_2156_to_fp16"), val = tensor(0x1.1p-20)]; + tensor var_2157_cast_fp16 = add(x = reduce_mean_97_cast_fp16, y = var_2156_to_fp16)[name = tensor("op_2157_cast_fp16")]; + tensor var_2158_cast_fp16 = sqrt(x = var_2157_cast_fp16)[name = tensor("op_2158_cast_fp16")]; + tensor x_233_cast_fp16 = real_div(x = sub_60_cast_fp16, y = var_2158_cast_fp16)[name = tensor("x_233_cast_fp16")]; + tensor var_2160_cast_fp16 = mul(x = x_233_cast_fp16, y = flow_net_res_blocks_1_in_ln_weight_to_fp16)[name = tensor("op_2160_cast_fp16")]; + tensor x_235_cast_fp16 = add(x = var_2160_cast_fp16, y = flow_net_res_blocks_1_in_ln_bias_to_fp16)[name = tensor("x_235_cast_fp16")]; + tensor var_2162_promoted_to_fp16 = const()[name = tensor("op_2162_promoted_to_fp16"), val = tensor(0x1p+0)]; + tensor var_2163_cast_fp16 = add(x = var_2146_cast_fp16_1, y = var_2162_promoted_to_fp16)[name = tensor("op_2163_cast_fp16")]; + tensor var_2164_cast_fp16 = mul(x = x_235_cast_fp16, y = var_2163_cast_fp16)[name = tensor("op_2164_cast_fp16")]; + tensor input_365_cast_fp16 = add(x = var_2164_cast_fp16, y = var_2146_cast_fp16_0)[name = tensor("input_365_cast_fp16")]; + tensor linear_140_cast_fp16 = linear(bias = flow_net_res_blocks_1_mlp_0_bias_to_fp16, weight = flow_net_res_blocks_1_mlp_0_weight_to_fp16, x = input_365_cast_fp16)[name = tensor("linear_140_cast_fp16")]; + tensor input_369_cast_fp16 = silu(x = linear_140_cast_fp16)[name = tensor("input_369_cast_fp16")]; + tensor linear_141_cast_fp16 = linear(bias = flow_net_res_blocks_1_mlp_2_bias_to_fp16, weight = flow_net_res_blocks_1_mlp_2_weight_to_fp16, x = input_369_cast_fp16)[name = tensor("linear_141_cast_fp16")]; + tensor var_2175_cast_fp16 = mul(x = var_2146_cast_fp16_2, y = linear_141_cast_fp16)[name = tensor("op_2175_cast_fp16")]; + tensor x_237_cast_fp16 = add(x = x_231_cast_fp16, y = var_2175_cast_fp16)[name = tensor("x_237_cast_fp16")]; + tensor linear_142_cast_fp16 = linear(bias = flow_net_res_blocks_2_adaLN_modulation_1_bias_to_fp16, weight = flow_net_res_blocks_2_adaLN_modulation_1_weight_to_fp16, x = input_355_cast_fp16)[name = tensor("linear_142_cast_fp16")]; + tensor var_2185_split_sizes_0 = const()[name = tensor("op_2185_split_sizes_0"), val = tensor([512, 512, 512])]; + tensor var_2185_axis_0 = const()[name = tensor("op_2185_axis_0"), val = tensor(-1)]; + tensor var_2185_cast_fp16_0, tensor var_2185_cast_fp16_1, tensor var_2185_cast_fp16_2 = split(axis = var_2185_axis_0, split_sizes = var_2185_split_sizes_0, x = linear_142_cast_fp16)[name = tensor("op_2185_cast_fp16")]; + tensor mean_75_axes_0 = const()[name = tensor("mean_75_axes_0"), val = tensor([-1])]; + tensor mean_75_keep_dims_0 = const()[name = tensor("mean_75_keep_dims_0"), val = tensor(true)]; + tensor mean_75_cast_fp16 = reduce_mean(axes = mean_75_axes_0, keep_dims = mean_75_keep_dims_0, x = x_237_cast_fp16)[name = tensor("mean_75_cast_fp16")]; + tensor sub_61_cast_fp16 = sub(x = x_237_cast_fp16, y = mean_75_cast_fp16)[name = tensor("sub_61_cast_fp16")]; + tensor square_49_cast_fp16 = square(x = sub_61_cast_fp16)[name = tensor("square_49_cast_fp16")]; + tensor reduce_mean_99_axes_0 = const()[name = tensor("reduce_mean_99_axes_0"), val = tensor([-1])]; + tensor reduce_mean_99_keep_dims_0 = const()[name = tensor("reduce_mean_99_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_99_cast_fp16 = reduce_mean(axes = reduce_mean_99_axes_0, keep_dims = reduce_mean_99_keep_dims_0, x = square_49_cast_fp16)[name = tensor("reduce_mean_99_cast_fp16")]; + tensor var_2195_to_fp16 = const()[name = tensor("op_2195_to_fp16"), val = tensor(0x1.1p-20)]; + tensor var_2196_cast_fp16 = add(x = reduce_mean_99_cast_fp16, y = var_2195_to_fp16)[name = tensor("op_2196_cast_fp16")]; + tensor var_2197_cast_fp16 = sqrt(x = var_2196_cast_fp16)[name = tensor("op_2197_cast_fp16")]; + tensor x_239_cast_fp16 = real_div(x = sub_61_cast_fp16, y = var_2197_cast_fp16)[name = tensor("x_239_cast_fp16")]; + tensor var_2199_cast_fp16 = mul(x = x_239_cast_fp16, y = flow_net_res_blocks_2_in_ln_weight_to_fp16)[name = tensor("op_2199_cast_fp16")]; + tensor x_241_cast_fp16 = add(x = var_2199_cast_fp16, y = flow_net_res_blocks_2_in_ln_bias_to_fp16)[name = tensor("x_241_cast_fp16")]; + tensor var_2201_promoted_to_fp16 = const()[name = tensor("op_2201_promoted_to_fp16"), val = tensor(0x1p+0)]; + tensor var_2202_cast_fp16 = add(x = var_2185_cast_fp16_1, y = var_2201_promoted_to_fp16)[name = tensor("op_2202_cast_fp16")]; + tensor var_2203_cast_fp16 = mul(x = x_241_cast_fp16, y = var_2202_cast_fp16)[name = tensor("op_2203_cast_fp16")]; + tensor input_373_cast_fp16 = add(x = var_2203_cast_fp16, y = var_2185_cast_fp16_0)[name = tensor("input_373_cast_fp16")]; + tensor linear_143_cast_fp16 = linear(bias = flow_net_res_blocks_2_mlp_0_bias_to_fp16, weight = flow_net_res_blocks_2_mlp_0_weight_to_fp16, x = input_373_cast_fp16)[name = tensor("linear_143_cast_fp16")]; + tensor input_377_cast_fp16 = silu(x = linear_143_cast_fp16)[name = tensor("input_377_cast_fp16")]; + tensor linear_144_cast_fp16 = linear(bias = flow_net_res_blocks_2_mlp_2_bias_to_fp16, weight = flow_net_res_blocks_2_mlp_2_weight_to_fp16, x = input_377_cast_fp16)[name = tensor("linear_144_cast_fp16")]; + tensor var_2214_cast_fp16 = mul(x = var_2185_cast_fp16_2, y = linear_144_cast_fp16)[name = tensor("op_2214_cast_fp16")]; + tensor x_243_cast_fp16 = add(x = x_237_cast_fp16, y = var_2214_cast_fp16)[name = tensor("x_243_cast_fp16")]; + tensor linear_145_cast_fp16 = linear(bias = flow_net_res_blocks_3_adaLN_modulation_1_bias_to_fp16, weight = flow_net_res_blocks_3_adaLN_modulation_1_weight_to_fp16, x = input_355_cast_fp16)[name = tensor("linear_145_cast_fp16")]; + tensor var_2224_split_sizes_0 = const()[name = tensor("op_2224_split_sizes_0"), val = tensor([512, 512, 512])]; + tensor var_2224_axis_0 = const()[name = tensor("op_2224_axis_0"), val = tensor(-1)]; + tensor var_2224_cast_fp16_0, tensor var_2224_cast_fp16_1, tensor var_2224_cast_fp16_2 = split(axis = var_2224_axis_0, split_sizes = var_2224_split_sizes_0, x = linear_145_cast_fp16)[name = tensor("op_2224_cast_fp16")]; + tensor mean_77_axes_0 = const()[name = tensor("mean_77_axes_0"), val = tensor([-1])]; + tensor mean_77_keep_dims_0 = const()[name = tensor("mean_77_keep_dims_0"), val = tensor(true)]; + tensor mean_77_cast_fp16 = reduce_mean(axes = mean_77_axes_0, keep_dims = mean_77_keep_dims_0, x = x_243_cast_fp16)[name = tensor("mean_77_cast_fp16")]; + tensor sub_62_cast_fp16 = sub(x = x_243_cast_fp16, y = mean_77_cast_fp16)[name = tensor("sub_62_cast_fp16")]; + tensor square_50_cast_fp16 = square(x = sub_62_cast_fp16)[name = tensor("square_50_cast_fp16")]; + tensor reduce_mean_101_axes_0 = const()[name = tensor("reduce_mean_101_axes_0"), val = tensor([-1])]; + tensor reduce_mean_101_keep_dims_0 = const()[name = tensor("reduce_mean_101_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_101_cast_fp16 = reduce_mean(axes = reduce_mean_101_axes_0, keep_dims = reduce_mean_101_keep_dims_0, x = square_50_cast_fp16)[name = tensor("reduce_mean_101_cast_fp16")]; + tensor var_2234_to_fp16 = const()[name = tensor("op_2234_to_fp16"), val = tensor(0x1.1p-20)]; + tensor var_2235_cast_fp16 = add(x = reduce_mean_101_cast_fp16, y = var_2234_to_fp16)[name = tensor("op_2235_cast_fp16")]; + tensor var_2236_cast_fp16 = sqrt(x = var_2235_cast_fp16)[name = tensor("op_2236_cast_fp16")]; + tensor x_245_cast_fp16 = real_div(x = sub_62_cast_fp16, y = var_2236_cast_fp16)[name = tensor("x_245_cast_fp16")]; + tensor var_2238_cast_fp16 = mul(x = x_245_cast_fp16, y = flow_net_res_blocks_3_in_ln_weight_to_fp16)[name = tensor("op_2238_cast_fp16")]; + tensor x_247_cast_fp16 = add(x = var_2238_cast_fp16, y = flow_net_res_blocks_3_in_ln_bias_to_fp16)[name = tensor("x_247_cast_fp16")]; + tensor var_2240_promoted_to_fp16 = const()[name = tensor("op_2240_promoted_to_fp16"), val = tensor(0x1p+0)]; + tensor var_2241_cast_fp16 = add(x = var_2224_cast_fp16_1, y = var_2240_promoted_to_fp16)[name = tensor("op_2241_cast_fp16")]; + tensor var_2242_cast_fp16 = mul(x = x_247_cast_fp16, y = var_2241_cast_fp16)[name = tensor("op_2242_cast_fp16")]; + tensor input_381_cast_fp16 = add(x = var_2242_cast_fp16, y = var_2224_cast_fp16_0)[name = tensor("input_381_cast_fp16")]; + tensor linear_146_cast_fp16 = linear(bias = flow_net_res_blocks_3_mlp_0_bias_to_fp16, weight = flow_net_res_blocks_3_mlp_0_weight_to_fp16, x = input_381_cast_fp16)[name = tensor("linear_146_cast_fp16")]; + tensor input_385_cast_fp16 = silu(x = linear_146_cast_fp16)[name = tensor("input_385_cast_fp16")]; + tensor linear_147_cast_fp16 = linear(bias = flow_net_res_blocks_3_mlp_2_bias_to_fp16, weight = flow_net_res_blocks_3_mlp_2_weight_to_fp16, x = input_385_cast_fp16)[name = tensor("linear_147_cast_fp16")]; + tensor var_2253_cast_fp16 = mul(x = var_2224_cast_fp16_2, y = linear_147_cast_fp16)[name = tensor("op_2253_cast_fp16")]; + tensor x_249_cast_fp16 = add(x = x_243_cast_fp16, y = var_2253_cast_fp16)[name = tensor("x_249_cast_fp16")]; + tensor linear_148_cast_fp16 = linear(bias = flow_net_res_blocks_4_adaLN_modulation_1_bias_to_fp16, weight = flow_net_res_blocks_4_adaLN_modulation_1_weight_to_fp16, x = input_355_cast_fp16)[name = tensor("linear_148_cast_fp16")]; + tensor var_2263_split_sizes_0 = const()[name = tensor("op_2263_split_sizes_0"), val = tensor([512, 512, 512])]; + tensor var_2263_axis_0 = const()[name = tensor("op_2263_axis_0"), val = tensor(-1)]; + tensor var_2263_cast_fp16_0, tensor var_2263_cast_fp16_1, tensor var_2263_cast_fp16_2 = split(axis = var_2263_axis_0, split_sizes = var_2263_split_sizes_0, x = linear_148_cast_fp16)[name = tensor("op_2263_cast_fp16")]; + tensor mean_79_axes_0 = const()[name = tensor("mean_79_axes_0"), val = tensor([-1])]; + tensor mean_79_keep_dims_0 = const()[name = tensor("mean_79_keep_dims_0"), val = tensor(true)]; + tensor mean_79_cast_fp16 = reduce_mean(axes = mean_79_axes_0, keep_dims = mean_79_keep_dims_0, x = x_249_cast_fp16)[name = tensor("mean_79_cast_fp16")]; + tensor sub_63_cast_fp16 = sub(x = x_249_cast_fp16, y = mean_79_cast_fp16)[name = tensor("sub_63_cast_fp16")]; + tensor square_51_cast_fp16 = square(x = sub_63_cast_fp16)[name = tensor("square_51_cast_fp16")]; + tensor reduce_mean_103_axes_0 = const()[name = tensor("reduce_mean_103_axes_0"), val = tensor([-1])]; + tensor reduce_mean_103_keep_dims_0 = const()[name = tensor("reduce_mean_103_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_103_cast_fp16 = reduce_mean(axes = reduce_mean_103_axes_0, keep_dims = reduce_mean_103_keep_dims_0, x = square_51_cast_fp16)[name = tensor("reduce_mean_103_cast_fp16")]; + tensor var_2273_to_fp16 = const()[name = tensor("op_2273_to_fp16"), val = tensor(0x1.1p-20)]; + tensor var_2274_cast_fp16 = add(x = reduce_mean_103_cast_fp16, y = var_2273_to_fp16)[name = tensor("op_2274_cast_fp16")]; + tensor var_2275_cast_fp16 = sqrt(x = var_2274_cast_fp16)[name = tensor("op_2275_cast_fp16")]; + tensor x_251_cast_fp16 = real_div(x = sub_63_cast_fp16, y = var_2275_cast_fp16)[name = tensor("x_251_cast_fp16")]; + tensor var_2277_cast_fp16 = mul(x = x_251_cast_fp16, y = flow_net_res_blocks_4_in_ln_weight_to_fp16)[name = tensor("op_2277_cast_fp16")]; + tensor x_253_cast_fp16 = add(x = var_2277_cast_fp16, y = flow_net_res_blocks_4_in_ln_bias_to_fp16)[name = tensor("x_253_cast_fp16")]; + tensor var_2279_promoted_to_fp16 = const()[name = tensor("op_2279_promoted_to_fp16"), val = tensor(0x1p+0)]; + tensor var_2280_cast_fp16 = add(x = var_2263_cast_fp16_1, y = var_2279_promoted_to_fp16)[name = tensor("op_2280_cast_fp16")]; + tensor var_2281_cast_fp16 = mul(x = x_253_cast_fp16, y = var_2280_cast_fp16)[name = tensor("op_2281_cast_fp16")]; + tensor input_389_cast_fp16 = add(x = var_2281_cast_fp16, y = var_2263_cast_fp16_0)[name = tensor("input_389_cast_fp16")]; + tensor linear_149_cast_fp16 = linear(bias = flow_net_res_blocks_4_mlp_0_bias_to_fp16, weight = flow_net_res_blocks_4_mlp_0_weight_to_fp16, x = input_389_cast_fp16)[name = tensor("linear_149_cast_fp16")]; + tensor input_393_cast_fp16 = silu(x = linear_149_cast_fp16)[name = tensor("input_393_cast_fp16")]; + tensor linear_150_cast_fp16 = linear(bias = flow_net_res_blocks_4_mlp_2_bias_to_fp16, weight = flow_net_res_blocks_4_mlp_2_weight_to_fp16, x = input_393_cast_fp16)[name = tensor("linear_150_cast_fp16")]; + tensor var_2292_cast_fp16 = mul(x = var_2263_cast_fp16_2, y = linear_150_cast_fp16)[name = tensor("op_2292_cast_fp16")]; + tensor x_255_cast_fp16 = add(x = x_249_cast_fp16, y = var_2292_cast_fp16)[name = tensor("x_255_cast_fp16")]; + tensor linear_151_cast_fp16 = linear(bias = flow_net_res_blocks_5_adaLN_modulation_1_bias_to_fp16, weight = flow_net_res_blocks_5_adaLN_modulation_1_weight_to_fp16, x = input_355_cast_fp16)[name = tensor("linear_151_cast_fp16")]; + tensor var_2302_split_sizes_0 = const()[name = tensor("op_2302_split_sizes_0"), val = tensor([512, 512, 512])]; + tensor var_2302_axis_0 = const()[name = tensor("op_2302_axis_0"), val = tensor(-1)]; + tensor var_2302_cast_fp16_0, tensor var_2302_cast_fp16_1, tensor var_2302_cast_fp16_2 = split(axis = var_2302_axis_0, split_sizes = var_2302_split_sizes_0, x = linear_151_cast_fp16)[name = tensor("op_2302_cast_fp16")]; + tensor mean_81_axes_0 = const()[name = tensor("mean_81_axes_0"), val = tensor([-1])]; + tensor mean_81_keep_dims_0 = const()[name = tensor("mean_81_keep_dims_0"), val = tensor(true)]; + tensor mean_81_cast_fp16 = reduce_mean(axes = mean_81_axes_0, keep_dims = mean_81_keep_dims_0, x = x_255_cast_fp16)[name = tensor("mean_81_cast_fp16")]; + tensor sub_64_cast_fp16 = sub(x = x_255_cast_fp16, y = mean_81_cast_fp16)[name = tensor("sub_64_cast_fp16")]; + tensor square_52_cast_fp16 = square(x = sub_64_cast_fp16)[name = tensor("square_52_cast_fp16")]; + tensor reduce_mean_105_axes_0 = const()[name = tensor("reduce_mean_105_axes_0"), val = tensor([-1])]; + tensor reduce_mean_105_keep_dims_0 = const()[name = tensor("reduce_mean_105_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_105_cast_fp16 = reduce_mean(axes = reduce_mean_105_axes_0, keep_dims = reduce_mean_105_keep_dims_0, x = square_52_cast_fp16)[name = tensor("reduce_mean_105_cast_fp16")]; + tensor var_2312_to_fp16 = const()[name = tensor("op_2312_to_fp16"), val = tensor(0x1.1p-20)]; + tensor var_2313_cast_fp16 = add(x = reduce_mean_105_cast_fp16, y = var_2312_to_fp16)[name = tensor("op_2313_cast_fp16")]; + tensor var_2314_cast_fp16 = sqrt(x = var_2313_cast_fp16)[name = tensor("op_2314_cast_fp16")]; + tensor x_257_cast_fp16 = real_div(x = sub_64_cast_fp16, y = var_2314_cast_fp16)[name = tensor("x_257_cast_fp16")]; + tensor var_2316_cast_fp16 = mul(x = x_257_cast_fp16, y = flow_net_res_blocks_5_in_ln_weight_to_fp16)[name = tensor("op_2316_cast_fp16")]; + tensor x_259_cast_fp16 = add(x = var_2316_cast_fp16, y = flow_net_res_blocks_5_in_ln_bias_to_fp16)[name = tensor("x_259_cast_fp16")]; + tensor var_2318_promoted_to_fp16 = const()[name = tensor("op_2318_promoted_to_fp16"), val = tensor(0x1p+0)]; + tensor var_2319_cast_fp16 = add(x = var_2302_cast_fp16_1, y = var_2318_promoted_to_fp16)[name = tensor("op_2319_cast_fp16")]; + tensor var_2320_cast_fp16 = mul(x = x_259_cast_fp16, y = var_2319_cast_fp16)[name = tensor("op_2320_cast_fp16")]; + tensor input_397_cast_fp16 = add(x = var_2320_cast_fp16, y = var_2302_cast_fp16_0)[name = tensor("input_397_cast_fp16")]; + tensor linear_152_cast_fp16 = linear(bias = flow_net_res_blocks_5_mlp_0_bias_to_fp16, weight = flow_net_res_blocks_5_mlp_0_weight_to_fp16, x = input_397_cast_fp16)[name = tensor("linear_152_cast_fp16")]; + tensor input_401_cast_fp16 = silu(x = linear_152_cast_fp16)[name = tensor("input_401_cast_fp16")]; + tensor linear_153_cast_fp16 = linear(bias = flow_net_res_blocks_5_mlp_2_bias_to_fp16, weight = flow_net_res_blocks_5_mlp_2_weight_to_fp16, x = input_401_cast_fp16)[name = tensor("linear_153_cast_fp16")]; + tensor var_2331_cast_fp16 = mul(x = var_2302_cast_fp16_2, y = linear_153_cast_fp16)[name = tensor("op_2331_cast_fp16")]; + tensor x_261_cast_fp16 = add(x = x_255_cast_fp16, y = var_2331_cast_fp16)[name = tensor("x_261_cast_fp16")]; + tensor linear_154_cast_fp16 = linear(bias = flow_net_final_layer_adaLN_modulation_1_bias_to_fp16, weight = flow_net_final_layer_adaLN_modulation_1_weight_to_fp16, x = input_355_cast_fp16)[name = tensor("linear_154_cast_fp16")]; + tensor var_2340_split_sizes_0 = const()[name = tensor("op_2340_split_sizes_0"), val = tensor([512, 512])]; + tensor var_2340_axis_0 = const()[name = tensor("op_2340_axis_0"), val = tensor(-1)]; + tensor var_2340_cast_fp16_0, tensor var_2340_cast_fp16_1 = split(axis = var_2340_axis_0, split_sizes = var_2340_split_sizes_0, x = linear_154_cast_fp16)[name = tensor("op_2340_cast_fp16")]; + tensor mean_83_axes_0 = const()[name = tensor("mean_83_axes_0"), val = tensor([-1])]; + tensor mean_83_keep_dims_0 = const()[name = tensor("mean_83_keep_dims_0"), val = tensor(true)]; + tensor mean_83_cast_fp16 = reduce_mean(axes = mean_83_axes_0, keep_dims = mean_83_keep_dims_0, x = x_261_cast_fp16)[name = tensor("mean_83_cast_fp16")]; + tensor sub_65_cast_fp16 = sub(x = x_261_cast_fp16, y = mean_83_cast_fp16)[name = tensor("sub_65_cast_fp16")]; + tensor square_53_cast_fp16 = square(x = sub_65_cast_fp16)[name = tensor("square_53_cast_fp16")]; + tensor reduce_mean_107_axes_0 = const()[name = tensor("reduce_mean_107_axes_0"), val = tensor([-1])]; + tensor reduce_mean_107_keep_dims_0 = const()[name = tensor("reduce_mean_107_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_107_cast_fp16 = reduce_mean(axes = reduce_mean_107_axes_0, keep_dims = reduce_mean_107_keep_dims_0, x = square_53_cast_fp16)[name = tensor("reduce_mean_107_cast_fp16")]; + tensor var_2347_to_fp16 = const()[name = tensor("op_2347_to_fp16"), val = tensor(0x1.1p-20)]; + tensor var_2348_cast_fp16 = add(x = reduce_mean_107_cast_fp16, y = var_2347_to_fp16)[name = tensor("op_2348_cast_fp16")]; + tensor var_2349_cast_fp16 = sqrt(x = var_2348_cast_fp16)[name = tensor("op_2349_cast_fp16")]; + tensor x_263_cast_fp16 = real_div(x = sub_65_cast_fp16, y = var_2349_cast_fp16)[name = tensor("x_263_cast_fp16")]; + tensor var_2351_promoted_to_fp16 = const()[name = tensor("op_2351_promoted_to_fp16"), val = tensor(0x1p+0)]; + tensor var_2352_cast_fp16 = add(x = var_2340_cast_fp16_1, y = var_2351_promoted_to_fp16)[name = tensor("op_2352_cast_fp16")]; + tensor var_2353_cast_fp16 = mul(x = x_263_cast_fp16, y = var_2352_cast_fp16)[name = tensor("op_2353_cast_fp16")]; + tensor input_405_cast_fp16 = add(x = var_2353_cast_fp16, y = var_2340_cast_fp16_0)[name = tensor("input_405_cast_fp16")]; + tensor linear_155_cast_fp16 = linear(bias = flow_net_final_layer_linear_bias_to_fp16, weight = flow_net_final_layer_linear_weight_to_fp16, x = input_405_cast_fp16)[name = tensor("linear_155_cast_fp16")]; + tensor var_2364_to_fp16 = const()[name = tensor("op_2364_to_fp16"), val = tensor(0x1p-3)]; + tensor var_2365_cast_fp16 = mul(x = linear_155_cast_fp16, y = var_2364_to_fp16)[name = tensor("op_2365_cast_fp16")]; + tensor input_407_cast_fp16 = add(x = input_339_cast_fp16, y = var_2365_cast_fp16)[name = tensor("input_407_cast_fp16")]; + tensor linear_156_cast_fp16 = linear(bias = flow_net_input_proj_bias_to_fp16, weight = flow_net_input_proj_weight_to_fp16, x = input_407_cast_fp16)[name = tensor("linear_156_cast_fp16")]; + tensor input_411_to_fp16 = const()[name = tensor("input_411_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(19008256)))]; + tensor input_413_cast_fp16 = silu(x = input_411_to_fp16)[name = tensor("input_413_cast_fp16")]; + tensor linear_158_cast_fp16 = linear(bias = flow_net_time_embed_0_mlp_2_bias_to_fp16, weight = flow_net_time_embed_0_mlp_2_weight_to_fp16, x = input_413_cast_fp16)[name = tensor("linear_158_cast_fp16")]; + tensor reduce_mean_108_axes_0 = const()[name = tensor("reduce_mean_108_axes_0"), val = tensor([-1])]; + tensor reduce_mean_108_keep_dims_0 = const()[name = tensor("reduce_mean_108_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_108_cast_fp16 = reduce_mean(axes = reduce_mean_108_axes_0, keep_dims = reduce_mean_108_keep_dims_0, x = linear_158_cast_fp16)[name = tensor("reduce_mean_108_cast_fp16")]; + tensor sub_66_cast_fp16 = sub(x = linear_158_cast_fp16, y = reduce_mean_108_cast_fp16)[name = tensor("sub_66_cast_fp16")]; + tensor square_54_cast_fp16 = square(x = sub_66_cast_fp16)[name = tensor("square_54_cast_fp16")]; + tensor reduce_mean_109_axes_0 = const()[name = tensor("reduce_mean_109_axes_0"), val = tensor([-1])]; + tensor reduce_mean_109_keep_dims_0 = const()[name = tensor("reduce_mean_109_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_109_cast_fp16 = reduce_mean(axes = reduce_mean_109_axes_0, keep_dims = reduce_mean_109_keep_dims_0, x = square_54_cast_fp16)[name = tensor("reduce_mean_109_cast_fp16")]; + tensor real_div_12_to_fp16 = const()[name = tensor("real_div_12_to_fp16"), val = tensor(0x1.008p+0)]; + tensor mul_12_cast_fp16 = mul(x = reduce_mean_109_cast_fp16, y = real_div_12_to_fp16)[name = tensor("mul_12_cast_fp16")]; + tensor var_2431_to_fp16 = const()[name = tensor("op_2431_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_109_cast_fp16 = add(x = mul_12_cast_fp16, y = var_2431_to_fp16)[name = tensor("var_109_cast_fp16")]; + tensor var_2434_epsilon_0 = const()[name = tensor("op_2434_epsilon_0"), val = tensor(0x1.197998p-40)]; + tensor var_2434_cast_fp16 = rsqrt(epsilon = var_2434_epsilon_0, x = var_109_cast_fp16)[name = tensor("op_2434_cast_fp16")]; + tensor var_2435_cast_fp16 = mul(x = const_3_to_fp16, y = var_2434_cast_fp16)[name = tensor("op_2435_cast_fp16")]; + tensor var_2436_cast_fp16 = mul(x = linear_158_cast_fp16, y = var_2435_cast_fp16)[name = tensor("op_2436_cast_fp16")]; + tensor input_417_to_fp16 = const()[name = tensor("input_417_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(19009344)))]; + tensor input_419_cast_fp16 = silu(x = input_417_to_fp16)[name = tensor("input_419_cast_fp16")]; + tensor linear_160_cast_fp16 = linear(bias = flow_net_time_embed_1_mlp_2_bias_to_fp16, weight = flow_net_time_embed_1_mlp_2_weight_to_fp16, x = input_419_cast_fp16)[name = tensor("linear_160_cast_fp16")]; + tensor reduce_mean_110_axes_0 = const()[name = tensor("reduce_mean_110_axes_0"), val = tensor([-1])]; + tensor reduce_mean_110_keep_dims_0 = const()[name = tensor("reduce_mean_110_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_110_cast_fp16 = reduce_mean(axes = reduce_mean_110_axes_0, keep_dims = reduce_mean_110_keep_dims_0, x = linear_160_cast_fp16)[name = tensor("reduce_mean_110_cast_fp16")]; + tensor sub_68_cast_fp16 = sub(x = linear_160_cast_fp16, y = reduce_mean_110_cast_fp16)[name = tensor("sub_68_cast_fp16")]; + tensor square_55_cast_fp16 = square(x = sub_68_cast_fp16)[name = tensor("square_55_cast_fp16")]; + tensor reduce_mean_111_axes_0 = const()[name = tensor("reduce_mean_111_axes_0"), val = tensor([-1])]; + tensor reduce_mean_111_keep_dims_0 = const()[name = tensor("reduce_mean_111_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_111_cast_fp16 = reduce_mean(axes = reduce_mean_111_axes_0, keep_dims = reduce_mean_111_keep_dims_0, x = square_55_cast_fp16)[name = tensor("reduce_mean_111_cast_fp16")]; + tensor real_div_13_to_fp16 = const()[name = tensor("real_div_13_to_fp16"), val = tensor(0x1.008p+0)]; + tensor mul_13_cast_fp16 = mul(x = reduce_mean_111_cast_fp16, y = real_div_13_to_fp16)[name = tensor("mul_13_cast_fp16")]; + tensor var_2468_to_fp16 = const()[name = tensor("op_2468_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_111_cast_fp16 = add(x = mul_13_cast_fp16, y = var_2468_to_fp16)[name = tensor("var_111_cast_fp16")]; + tensor var_2471_epsilon_0 = const()[name = tensor("op_2471_epsilon_0"), val = tensor(0x1.197998p-40)]; + tensor var_2471_cast_fp16 = rsqrt(epsilon = var_2471_epsilon_0, x = var_111_cast_fp16)[name = tensor("op_2471_cast_fp16")]; + tensor var_2472_cast_fp16 = mul(x = const_5_to_fp16, y = var_2471_cast_fp16)[name = tensor("op_2472_cast_fp16")]; + tensor var_2473_cast_fp16 = mul(x = linear_160_cast_fp16, y = var_2472_cast_fp16)[name = tensor("op_2473_cast_fp16")]; + tensor var_2485_cast_fp16 = add(x = var_2436_cast_fp16, y = var_2473_cast_fp16)[name = tensor("op_2485_cast_fp16")]; + tensor _inversed_t_combined_13_y_0_to_fp16 = const()[name = tensor("_inversed_t_combined_13_y_0_to_fp16"), val = tensor(0x1p-1)]; + tensor _inversed_t_combined_13_cast_fp16 = mul(x = var_2485_cast_fp16, y = _inversed_t_combined_13_y_0_to_fp16)[name = tensor("_inversed_t_combined_13_cast_fp16")]; + tensor input_421_cast_fp16 = add(x = _inversed_t_combined_13_cast_fp16, y = linear_5_cast_fp16)[name = tensor("input_421_cast_fp16")]; + tensor input_423_cast_fp16 = silu(x = input_421_cast_fp16)[name = tensor("input_423_cast_fp16")]; + tensor linear_162_cast_fp16 = linear(bias = flow_net_res_blocks_0_adaLN_modulation_1_bias_to_fp16, weight = flow_net_res_blocks_0_adaLN_modulation_1_weight_to_fp16, x = input_423_cast_fp16)[name = tensor("linear_162_cast_fp16")]; + tensor var_2500_split_sizes_0 = const()[name = tensor("op_2500_split_sizes_0"), val = tensor([512, 512, 512])]; + tensor var_2500_axis_0 = const()[name = tensor("op_2500_axis_0"), val = tensor(-1)]; + tensor var_2500_cast_fp16_0, tensor var_2500_cast_fp16_1, tensor var_2500_cast_fp16_2 = split(axis = var_2500_axis_0, split_sizes = var_2500_split_sizes_0, x = linear_162_cast_fp16)[name = tensor("op_2500_cast_fp16")]; + tensor mean_85_axes_0 = const()[name = tensor("mean_85_axes_0"), val = tensor([-1])]; + tensor mean_85_keep_dims_0 = const()[name = tensor("mean_85_keep_dims_0"), val = tensor(true)]; + tensor mean_85_cast_fp16 = reduce_mean(axes = mean_85_axes_0, keep_dims = mean_85_keep_dims_0, x = linear_156_cast_fp16)[name = tensor("mean_85_cast_fp16")]; + tensor sub_70_cast_fp16 = sub(x = linear_156_cast_fp16, y = mean_85_cast_fp16)[name = tensor("sub_70_cast_fp16")]; + tensor square_56_cast_fp16 = square(x = sub_70_cast_fp16)[name = tensor("square_56_cast_fp16")]; + tensor reduce_mean_113_axes_0 = const()[name = tensor("reduce_mean_113_axes_0"), val = tensor([-1])]; + tensor reduce_mean_113_keep_dims_0 = const()[name = tensor("reduce_mean_113_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_113_cast_fp16 = reduce_mean(axes = reduce_mean_113_axes_0, keep_dims = reduce_mean_113_keep_dims_0, x = square_56_cast_fp16)[name = tensor("reduce_mean_113_cast_fp16")]; + tensor var_2510_to_fp16 = const()[name = tensor("op_2510_to_fp16"), val = tensor(0x1.1p-20)]; + tensor var_2511_cast_fp16 = add(x = reduce_mean_113_cast_fp16, y = var_2510_to_fp16)[name = tensor("op_2511_cast_fp16")]; + tensor var_2512_cast_fp16 = sqrt(x = var_2511_cast_fp16)[name = tensor("op_2512_cast_fp16")]; + tensor x_271_cast_fp16 = real_div(x = sub_70_cast_fp16, y = var_2512_cast_fp16)[name = tensor("x_271_cast_fp16")]; + tensor var_2514_cast_fp16 = mul(x = x_271_cast_fp16, y = flow_net_res_blocks_0_in_ln_weight_to_fp16)[name = tensor("op_2514_cast_fp16")]; + tensor x_273_cast_fp16 = add(x = var_2514_cast_fp16, y = flow_net_res_blocks_0_in_ln_bias_to_fp16)[name = tensor("x_273_cast_fp16")]; + tensor var_2516_promoted_to_fp16 = const()[name = tensor("op_2516_promoted_to_fp16"), val = tensor(0x1p+0)]; + tensor var_2517_cast_fp16 = add(x = var_2500_cast_fp16_1, y = var_2516_promoted_to_fp16)[name = tensor("op_2517_cast_fp16")]; + tensor var_2518_cast_fp16 = mul(x = x_273_cast_fp16, y = var_2517_cast_fp16)[name = tensor("op_2518_cast_fp16")]; + tensor input_425_cast_fp16 = add(x = var_2518_cast_fp16, y = var_2500_cast_fp16_0)[name = tensor("input_425_cast_fp16")]; + tensor linear_163_cast_fp16 = linear(bias = flow_net_res_blocks_0_mlp_0_bias_to_fp16, weight = flow_net_res_blocks_0_mlp_0_weight_to_fp16, x = input_425_cast_fp16)[name = tensor("linear_163_cast_fp16")]; + tensor input_429_cast_fp16 = silu(x = linear_163_cast_fp16)[name = tensor("input_429_cast_fp16")]; + tensor linear_164_cast_fp16 = linear(bias = flow_net_res_blocks_0_mlp_2_bias_to_fp16, weight = flow_net_res_blocks_0_mlp_2_weight_to_fp16, x = input_429_cast_fp16)[name = tensor("linear_164_cast_fp16")]; + tensor var_2529_cast_fp16 = mul(x = var_2500_cast_fp16_2, y = linear_164_cast_fp16)[name = tensor("op_2529_cast_fp16")]; + tensor x_275_cast_fp16 = add(x = linear_156_cast_fp16, y = var_2529_cast_fp16)[name = tensor("x_275_cast_fp16")]; + tensor linear_165_cast_fp16 = linear(bias = flow_net_res_blocks_1_adaLN_modulation_1_bias_to_fp16, weight = flow_net_res_blocks_1_adaLN_modulation_1_weight_to_fp16, x = input_423_cast_fp16)[name = tensor("linear_165_cast_fp16")]; + tensor var_2539_split_sizes_0 = const()[name = tensor("op_2539_split_sizes_0"), val = tensor([512, 512, 512])]; + tensor var_2539_axis_0 = const()[name = tensor("op_2539_axis_0"), val = tensor(-1)]; + tensor var_2539_cast_fp16_0, tensor var_2539_cast_fp16_1, tensor var_2539_cast_fp16_2 = split(axis = var_2539_axis_0, split_sizes = var_2539_split_sizes_0, x = linear_165_cast_fp16)[name = tensor("op_2539_cast_fp16")]; + tensor mean_87_axes_0 = const()[name = tensor("mean_87_axes_0"), val = tensor([-1])]; + tensor mean_87_keep_dims_0 = const()[name = tensor("mean_87_keep_dims_0"), val = tensor(true)]; + tensor mean_87_cast_fp16 = reduce_mean(axes = mean_87_axes_0, keep_dims = mean_87_keep_dims_0, x = x_275_cast_fp16)[name = tensor("mean_87_cast_fp16")]; + tensor sub_71_cast_fp16 = sub(x = x_275_cast_fp16, y = mean_87_cast_fp16)[name = tensor("sub_71_cast_fp16")]; + tensor square_57_cast_fp16 = square(x = sub_71_cast_fp16)[name = tensor("square_57_cast_fp16")]; + tensor reduce_mean_115_axes_0 = const()[name = tensor("reduce_mean_115_axes_0"), val = tensor([-1])]; + tensor reduce_mean_115_keep_dims_0 = const()[name = tensor("reduce_mean_115_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_115_cast_fp16 = reduce_mean(axes = reduce_mean_115_axes_0, keep_dims = reduce_mean_115_keep_dims_0, x = square_57_cast_fp16)[name = tensor("reduce_mean_115_cast_fp16")]; + tensor var_2549_to_fp16 = const()[name = tensor("op_2549_to_fp16"), val = tensor(0x1.1p-20)]; + tensor var_2550_cast_fp16 = add(x = reduce_mean_115_cast_fp16, y = var_2549_to_fp16)[name = tensor("op_2550_cast_fp16")]; + tensor var_2551_cast_fp16 = sqrt(x = var_2550_cast_fp16)[name = tensor("op_2551_cast_fp16")]; + tensor x_277_cast_fp16 = real_div(x = sub_71_cast_fp16, y = var_2551_cast_fp16)[name = tensor("x_277_cast_fp16")]; + tensor var_2553_cast_fp16 = mul(x = x_277_cast_fp16, y = flow_net_res_blocks_1_in_ln_weight_to_fp16)[name = tensor("op_2553_cast_fp16")]; + tensor x_279_cast_fp16 = add(x = var_2553_cast_fp16, y = flow_net_res_blocks_1_in_ln_bias_to_fp16)[name = tensor("x_279_cast_fp16")]; + tensor var_2555_promoted_to_fp16 = const()[name = tensor("op_2555_promoted_to_fp16"), val = tensor(0x1p+0)]; + tensor var_2556_cast_fp16 = add(x = var_2539_cast_fp16_1, y = var_2555_promoted_to_fp16)[name = tensor("op_2556_cast_fp16")]; + tensor var_2557_cast_fp16 = mul(x = x_279_cast_fp16, y = var_2556_cast_fp16)[name = tensor("op_2557_cast_fp16")]; + tensor input_433_cast_fp16 = add(x = var_2557_cast_fp16, y = var_2539_cast_fp16_0)[name = tensor("input_433_cast_fp16")]; + tensor linear_166_cast_fp16 = linear(bias = flow_net_res_blocks_1_mlp_0_bias_to_fp16, weight = flow_net_res_blocks_1_mlp_0_weight_to_fp16, x = input_433_cast_fp16)[name = tensor("linear_166_cast_fp16")]; + tensor input_437_cast_fp16 = silu(x = linear_166_cast_fp16)[name = tensor("input_437_cast_fp16")]; + tensor linear_167_cast_fp16 = linear(bias = flow_net_res_blocks_1_mlp_2_bias_to_fp16, weight = flow_net_res_blocks_1_mlp_2_weight_to_fp16, x = input_437_cast_fp16)[name = tensor("linear_167_cast_fp16")]; + tensor var_2568_cast_fp16 = mul(x = var_2539_cast_fp16_2, y = linear_167_cast_fp16)[name = tensor("op_2568_cast_fp16")]; + tensor x_281_cast_fp16 = add(x = x_275_cast_fp16, y = var_2568_cast_fp16)[name = tensor("x_281_cast_fp16")]; + tensor linear_168_cast_fp16 = linear(bias = flow_net_res_blocks_2_adaLN_modulation_1_bias_to_fp16, weight = flow_net_res_blocks_2_adaLN_modulation_1_weight_to_fp16, x = input_423_cast_fp16)[name = tensor("linear_168_cast_fp16")]; + tensor var_2578_split_sizes_0 = const()[name = tensor("op_2578_split_sizes_0"), val = tensor([512, 512, 512])]; + tensor var_2578_axis_0 = const()[name = tensor("op_2578_axis_0"), val = tensor(-1)]; + tensor var_2578_cast_fp16_0, tensor var_2578_cast_fp16_1, tensor var_2578_cast_fp16_2 = split(axis = var_2578_axis_0, split_sizes = var_2578_split_sizes_0, x = linear_168_cast_fp16)[name = tensor("op_2578_cast_fp16")]; + tensor mean_89_axes_0 = const()[name = tensor("mean_89_axes_0"), val = tensor([-1])]; + tensor mean_89_keep_dims_0 = const()[name = tensor("mean_89_keep_dims_0"), val = tensor(true)]; + tensor mean_89_cast_fp16 = reduce_mean(axes = mean_89_axes_0, keep_dims = mean_89_keep_dims_0, x = x_281_cast_fp16)[name = tensor("mean_89_cast_fp16")]; + tensor sub_72_cast_fp16 = sub(x = x_281_cast_fp16, y = mean_89_cast_fp16)[name = tensor("sub_72_cast_fp16")]; + tensor square_58_cast_fp16 = square(x = sub_72_cast_fp16)[name = tensor("square_58_cast_fp16")]; + tensor reduce_mean_117_axes_0 = const()[name = tensor("reduce_mean_117_axes_0"), val = tensor([-1])]; + tensor reduce_mean_117_keep_dims_0 = const()[name = tensor("reduce_mean_117_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_117_cast_fp16 = reduce_mean(axes = reduce_mean_117_axes_0, keep_dims = reduce_mean_117_keep_dims_0, x = square_58_cast_fp16)[name = tensor("reduce_mean_117_cast_fp16")]; + tensor var_2588_to_fp16 = const()[name = tensor("op_2588_to_fp16"), val = tensor(0x1.1p-20)]; + tensor var_2589_cast_fp16 = add(x = reduce_mean_117_cast_fp16, y = var_2588_to_fp16)[name = tensor("op_2589_cast_fp16")]; + tensor var_2590_cast_fp16 = sqrt(x = var_2589_cast_fp16)[name = tensor("op_2590_cast_fp16")]; + tensor x_283_cast_fp16 = real_div(x = sub_72_cast_fp16, y = var_2590_cast_fp16)[name = tensor("x_283_cast_fp16")]; + tensor var_2592_cast_fp16 = mul(x = x_283_cast_fp16, y = flow_net_res_blocks_2_in_ln_weight_to_fp16)[name = tensor("op_2592_cast_fp16")]; + tensor x_285_cast_fp16 = add(x = var_2592_cast_fp16, y = flow_net_res_blocks_2_in_ln_bias_to_fp16)[name = tensor("x_285_cast_fp16")]; + tensor var_2594_promoted_to_fp16 = const()[name = tensor("op_2594_promoted_to_fp16"), val = tensor(0x1p+0)]; + tensor var_2595_cast_fp16 = add(x = var_2578_cast_fp16_1, y = var_2594_promoted_to_fp16)[name = tensor("op_2595_cast_fp16")]; + tensor var_2596_cast_fp16 = mul(x = x_285_cast_fp16, y = var_2595_cast_fp16)[name = tensor("op_2596_cast_fp16")]; + tensor input_441_cast_fp16 = add(x = var_2596_cast_fp16, y = var_2578_cast_fp16_0)[name = tensor("input_441_cast_fp16")]; + tensor linear_169_cast_fp16 = linear(bias = flow_net_res_blocks_2_mlp_0_bias_to_fp16, weight = flow_net_res_blocks_2_mlp_0_weight_to_fp16, x = input_441_cast_fp16)[name = tensor("linear_169_cast_fp16")]; + tensor input_445_cast_fp16 = silu(x = linear_169_cast_fp16)[name = tensor("input_445_cast_fp16")]; + tensor linear_170_cast_fp16 = linear(bias = flow_net_res_blocks_2_mlp_2_bias_to_fp16, weight = flow_net_res_blocks_2_mlp_2_weight_to_fp16, x = input_445_cast_fp16)[name = tensor("linear_170_cast_fp16")]; + tensor var_2607_cast_fp16 = mul(x = var_2578_cast_fp16_2, y = linear_170_cast_fp16)[name = tensor("op_2607_cast_fp16")]; + tensor x_287_cast_fp16 = add(x = x_281_cast_fp16, y = var_2607_cast_fp16)[name = tensor("x_287_cast_fp16")]; + tensor linear_171_cast_fp16 = linear(bias = flow_net_res_blocks_3_adaLN_modulation_1_bias_to_fp16, weight = flow_net_res_blocks_3_adaLN_modulation_1_weight_to_fp16, x = input_423_cast_fp16)[name = tensor("linear_171_cast_fp16")]; + tensor var_2617_split_sizes_0 = const()[name = tensor("op_2617_split_sizes_0"), val = tensor([512, 512, 512])]; + tensor var_2617_axis_0 = const()[name = tensor("op_2617_axis_0"), val = tensor(-1)]; + tensor var_2617_cast_fp16_0, tensor var_2617_cast_fp16_1, tensor var_2617_cast_fp16_2 = split(axis = var_2617_axis_0, split_sizes = var_2617_split_sizes_0, x = linear_171_cast_fp16)[name = tensor("op_2617_cast_fp16")]; + tensor mean_91_axes_0 = const()[name = tensor("mean_91_axes_0"), val = tensor([-1])]; + tensor mean_91_keep_dims_0 = const()[name = tensor("mean_91_keep_dims_0"), val = tensor(true)]; + tensor mean_91_cast_fp16 = reduce_mean(axes = mean_91_axes_0, keep_dims = mean_91_keep_dims_0, x = x_287_cast_fp16)[name = tensor("mean_91_cast_fp16")]; + tensor sub_73_cast_fp16 = sub(x = x_287_cast_fp16, y = mean_91_cast_fp16)[name = tensor("sub_73_cast_fp16")]; + tensor square_59_cast_fp16 = square(x = sub_73_cast_fp16)[name = tensor("square_59_cast_fp16")]; + tensor reduce_mean_119_axes_0 = const()[name = tensor("reduce_mean_119_axes_0"), val = tensor([-1])]; + tensor reduce_mean_119_keep_dims_0 = const()[name = tensor("reduce_mean_119_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_119_cast_fp16 = reduce_mean(axes = reduce_mean_119_axes_0, keep_dims = reduce_mean_119_keep_dims_0, x = square_59_cast_fp16)[name = tensor("reduce_mean_119_cast_fp16")]; + tensor var_2627_to_fp16 = const()[name = tensor("op_2627_to_fp16"), val = tensor(0x1.1p-20)]; + tensor var_2628_cast_fp16 = add(x = reduce_mean_119_cast_fp16, y = var_2627_to_fp16)[name = tensor("op_2628_cast_fp16")]; + tensor var_2629_cast_fp16 = sqrt(x = var_2628_cast_fp16)[name = tensor("op_2629_cast_fp16")]; + tensor x_289_cast_fp16 = real_div(x = sub_73_cast_fp16, y = var_2629_cast_fp16)[name = tensor("x_289_cast_fp16")]; + tensor var_2631_cast_fp16 = mul(x = x_289_cast_fp16, y = flow_net_res_blocks_3_in_ln_weight_to_fp16)[name = tensor("op_2631_cast_fp16")]; + tensor x_291_cast_fp16 = add(x = var_2631_cast_fp16, y = flow_net_res_blocks_3_in_ln_bias_to_fp16)[name = tensor("x_291_cast_fp16")]; + tensor var_2633_promoted_to_fp16 = const()[name = tensor("op_2633_promoted_to_fp16"), val = tensor(0x1p+0)]; + tensor var_2634_cast_fp16 = add(x = var_2617_cast_fp16_1, y = var_2633_promoted_to_fp16)[name = tensor("op_2634_cast_fp16")]; + tensor var_2635_cast_fp16 = mul(x = x_291_cast_fp16, y = var_2634_cast_fp16)[name = tensor("op_2635_cast_fp16")]; + tensor input_449_cast_fp16 = add(x = var_2635_cast_fp16, y = var_2617_cast_fp16_0)[name = tensor("input_449_cast_fp16")]; + tensor linear_172_cast_fp16 = linear(bias = flow_net_res_blocks_3_mlp_0_bias_to_fp16, weight = flow_net_res_blocks_3_mlp_0_weight_to_fp16, x = input_449_cast_fp16)[name = tensor("linear_172_cast_fp16")]; + tensor input_453_cast_fp16 = silu(x = linear_172_cast_fp16)[name = tensor("input_453_cast_fp16")]; + tensor linear_173_cast_fp16 = linear(bias = flow_net_res_blocks_3_mlp_2_bias_to_fp16, weight = flow_net_res_blocks_3_mlp_2_weight_to_fp16, x = input_453_cast_fp16)[name = tensor("linear_173_cast_fp16")]; + tensor var_2646_cast_fp16 = mul(x = var_2617_cast_fp16_2, y = linear_173_cast_fp16)[name = tensor("op_2646_cast_fp16")]; + tensor x_293_cast_fp16 = add(x = x_287_cast_fp16, y = var_2646_cast_fp16)[name = tensor("x_293_cast_fp16")]; + tensor linear_174_cast_fp16 = linear(bias = flow_net_res_blocks_4_adaLN_modulation_1_bias_to_fp16, weight = flow_net_res_blocks_4_adaLN_modulation_1_weight_to_fp16, x = input_423_cast_fp16)[name = tensor("linear_174_cast_fp16")]; + tensor var_2656_split_sizes_0 = const()[name = tensor("op_2656_split_sizes_0"), val = tensor([512, 512, 512])]; + tensor var_2656_axis_0 = const()[name = tensor("op_2656_axis_0"), val = tensor(-1)]; + tensor var_2656_cast_fp16_0, tensor var_2656_cast_fp16_1, tensor var_2656_cast_fp16_2 = split(axis = var_2656_axis_0, split_sizes = var_2656_split_sizes_0, x = linear_174_cast_fp16)[name = tensor("op_2656_cast_fp16")]; + tensor mean_93_axes_0 = const()[name = tensor("mean_93_axes_0"), val = tensor([-1])]; + tensor mean_93_keep_dims_0 = const()[name = tensor("mean_93_keep_dims_0"), val = tensor(true)]; + tensor mean_93_cast_fp16 = reduce_mean(axes = mean_93_axes_0, keep_dims = mean_93_keep_dims_0, x = x_293_cast_fp16)[name = tensor("mean_93_cast_fp16")]; + tensor sub_74_cast_fp16 = sub(x = x_293_cast_fp16, y = mean_93_cast_fp16)[name = tensor("sub_74_cast_fp16")]; + tensor square_60_cast_fp16 = square(x = sub_74_cast_fp16)[name = tensor("square_60_cast_fp16")]; + tensor reduce_mean_121_axes_0 = const()[name = tensor("reduce_mean_121_axes_0"), val = tensor([-1])]; + tensor reduce_mean_121_keep_dims_0 = const()[name = tensor("reduce_mean_121_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_121_cast_fp16 = reduce_mean(axes = reduce_mean_121_axes_0, keep_dims = reduce_mean_121_keep_dims_0, x = square_60_cast_fp16)[name = tensor("reduce_mean_121_cast_fp16")]; + tensor var_2666_to_fp16 = const()[name = tensor("op_2666_to_fp16"), val = tensor(0x1.1p-20)]; + tensor var_2667_cast_fp16 = add(x = reduce_mean_121_cast_fp16, y = var_2666_to_fp16)[name = tensor("op_2667_cast_fp16")]; + tensor var_2668_cast_fp16 = sqrt(x = var_2667_cast_fp16)[name = tensor("op_2668_cast_fp16")]; + tensor x_295_cast_fp16 = real_div(x = sub_74_cast_fp16, y = var_2668_cast_fp16)[name = tensor("x_295_cast_fp16")]; + tensor var_2670_cast_fp16 = mul(x = x_295_cast_fp16, y = flow_net_res_blocks_4_in_ln_weight_to_fp16)[name = tensor("op_2670_cast_fp16")]; + tensor x_297_cast_fp16 = add(x = var_2670_cast_fp16, y = flow_net_res_blocks_4_in_ln_bias_to_fp16)[name = tensor("x_297_cast_fp16")]; + tensor var_2672_promoted_to_fp16 = const()[name = tensor("op_2672_promoted_to_fp16"), val = tensor(0x1p+0)]; + tensor var_2673_cast_fp16 = add(x = var_2656_cast_fp16_1, y = var_2672_promoted_to_fp16)[name = tensor("op_2673_cast_fp16")]; + tensor var_2674_cast_fp16 = mul(x = x_297_cast_fp16, y = var_2673_cast_fp16)[name = tensor("op_2674_cast_fp16")]; + tensor input_457_cast_fp16 = add(x = var_2674_cast_fp16, y = var_2656_cast_fp16_0)[name = tensor("input_457_cast_fp16")]; + tensor linear_175_cast_fp16 = linear(bias = flow_net_res_blocks_4_mlp_0_bias_to_fp16, weight = flow_net_res_blocks_4_mlp_0_weight_to_fp16, x = input_457_cast_fp16)[name = tensor("linear_175_cast_fp16")]; + tensor input_461_cast_fp16 = silu(x = linear_175_cast_fp16)[name = tensor("input_461_cast_fp16")]; + tensor linear_176_cast_fp16 = linear(bias = flow_net_res_blocks_4_mlp_2_bias_to_fp16, weight = flow_net_res_blocks_4_mlp_2_weight_to_fp16, x = input_461_cast_fp16)[name = tensor("linear_176_cast_fp16")]; + tensor var_2685_cast_fp16 = mul(x = var_2656_cast_fp16_2, y = linear_176_cast_fp16)[name = tensor("op_2685_cast_fp16")]; + tensor x_299_cast_fp16 = add(x = x_293_cast_fp16, y = var_2685_cast_fp16)[name = tensor("x_299_cast_fp16")]; + tensor linear_177_cast_fp16 = linear(bias = flow_net_res_blocks_5_adaLN_modulation_1_bias_to_fp16, weight = flow_net_res_blocks_5_adaLN_modulation_1_weight_to_fp16, x = input_423_cast_fp16)[name = tensor("linear_177_cast_fp16")]; + tensor var_2695_split_sizes_0 = const()[name = tensor("op_2695_split_sizes_0"), val = tensor([512, 512, 512])]; + tensor var_2695_axis_0 = const()[name = tensor("op_2695_axis_0"), val = tensor(-1)]; + tensor var_2695_cast_fp16_0, tensor var_2695_cast_fp16_1, tensor var_2695_cast_fp16_2 = split(axis = var_2695_axis_0, split_sizes = var_2695_split_sizes_0, x = linear_177_cast_fp16)[name = tensor("op_2695_cast_fp16")]; + tensor mean_95_axes_0 = const()[name = tensor("mean_95_axes_0"), val = tensor([-1])]; + tensor mean_95_keep_dims_0 = const()[name = tensor("mean_95_keep_dims_0"), val = tensor(true)]; + tensor mean_95_cast_fp16 = reduce_mean(axes = mean_95_axes_0, keep_dims = mean_95_keep_dims_0, x = x_299_cast_fp16)[name = tensor("mean_95_cast_fp16")]; + tensor sub_75_cast_fp16 = sub(x = x_299_cast_fp16, y = mean_95_cast_fp16)[name = tensor("sub_75_cast_fp16")]; + tensor square_61_cast_fp16 = square(x = sub_75_cast_fp16)[name = tensor("square_61_cast_fp16")]; + tensor reduce_mean_123_axes_0 = const()[name = tensor("reduce_mean_123_axes_0"), val = tensor([-1])]; + tensor reduce_mean_123_keep_dims_0 = const()[name = tensor("reduce_mean_123_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_123_cast_fp16 = reduce_mean(axes = reduce_mean_123_axes_0, keep_dims = reduce_mean_123_keep_dims_0, x = square_61_cast_fp16)[name = tensor("reduce_mean_123_cast_fp16")]; + tensor var_2705_to_fp16 = const()[name = tensor("op_2705_to_fp16"), val = tensor(0x1.1p-20)]; + tensor var_2706_cast_fp16 = add(x = reduce_mean_123_cast_fp16, y = var_2705_to_fp16)[name = tensor("op_2706_cast_fp16")]; + tensor var_2707_cast_fp16 = sqrt(x = var_2706_cast_fp16)[name = tensor("op_2707_cast_fp16")]; + tensor x_301_cast_fp16 = real_div(x = sub_75_cast_fp16, y = var_2707_cast_fp16)[name = tensor("x_301_cast_fp16")]; + tensor var_2709_cast_fp16 = mul(x = x_301_cast_fp16, y = flow_net_res_blocks_5_in_ln_weight_to_fp16)[name = tensor("op_2709_cast_fp16")]; + tensor x_303_cast_fp16 = add(x = var_2709_cast_fp16, y = flow_net_res_blocks_5_in_ln_bias_to_fp16)[name = tensor("x_303_cast_fp16")]; + tensor var_2711_promoted_to_fp16 = const()[name = tensor("op_2711_promoted_to_fp16"), val = tensor(0x1p+0)]; + tensor var_2712_cast_fp16 = add(x = var_2695_cast_fp16_1, y = var_2711_promoted_to_fp16)[name = tensor("op_2712_cast_fp16")]; + tensor var_2713_cast_fp16 = mul(x = x_303_cast_fp16, y = var_2712_cast_fp16)[name = tensor("op_2713_cast_fp16")]; + tensor input_465_cast_fp16 = add(x = var_2713_cast_fp16, y = var_2695_cast_fp16_0)[name = tensor("input_465_cast_fp16")]; + tensor linear_178_cast_fp16 = linear(bias = flow_net_res_blocks_5_mlp_0_bias_to_fp16, weight = flow_net_res_blocks_5_mlp_0_weight_to_fp16, x = input_465_cast_fp16)[name = tensor("linear_178_cast_fp16")]; + tensor input_469_cast_fp16 = silu(x = linear_178_cast_fp16)[name = tensor("input_469_cast_fp16")]; + tensor linear_179_cast_fp16 = linear(bias = flow_net_res_blocks_5_mlp_2_bias_to_fp16, weight = flow_net_res_blocks_5_mlp_2_weight_to_fp16, x = input_469_cast_fp16)[name = tensor("linear_179_cast_fp16")]; + tensor var_2724_cast_fp16 = mul(x = var_2695_cast_fp16_2, y = linear_179_cast_fp16)[name = tensor("op_2724_cast_fp16")]; + tensor x_305_cast_fp16 = add(x = x_299_cast_fp16, y = var_2724_cast_fp16)[name = tensor("x_305_cast_fp16")]; + tensor linear_180_cast_fp16 = linear(bias = flow_net_final_layer_adaLN_modulation_1_bias_to_fp16, weight = flow_net_final_layer_adaLN_modulation_1_weight_to_fp16, x = input_423_cast_fp16)[name = tensor("linear_180_cast_fp16")]; + tensor var_2733_split_sizes_0 = const()[name = tensor("op_2733_split_sizes_0"), val = tensor([512, 512])]; + tensor var_2733_axis_0 = const()[name = tensor("op_2733_axis_0"), val = tensor(-1)]; + tensor var_2733_cast_fp16_0, tensor var_2733_cast_fp16_1 = split(axis = var_2733_axis_0, split_sizes = var_2733_split_sizes_0, x = linear_180_cast_fp16)[name = tensor("op_2733_cast_fp16")]; + tensor mean_97_axes_0 = const()[name = tensor("mean_97_axes_0"), val = tensor([-1])]; + tensor mean_97_keep_dims_0 = const()[name = tensor("mean_97_keep_dims_0"), val = tensor(true)]; + tensor mean_97_cast_fp16 = reduce_mean(axes = mean_97_axes_0, keep_dims = mean_97_keep_dims_0, x = x_305_cast_fp16)[name = tensor("mean_97_cast_fp16")]; + tensor sub_76_cast_fp16 = sub(x = x_305_cast_fp16, y = mean_97_cast_fp16)[name = tensor("sub_76_cast_fp16")]; + tensor square_62_cast_fp16 = square(x = sub_76_cast_fp16)[name = tensor("square_62_cast_fp16")]; + tensor reduce_mean_125_axes_0 = const()[name = tensor("reduce_mean_125_axes_0"), val = tensor([-1])]; + tensor reduce_mean_125_keep_dims_0 = const()[name = tensor("reduce_mean_125_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_125_cast_fp16 = reduce_mean(axes = reduce_mean_125_axes_0, keep_dims = reduce_mean_125_keep_dims_0, x = square_62_cast_fp16)[name = tensor("reduce_mean_125_cast_fp16")]; + tensor var_2740_to_fp16 = const()[name = tensor("op_2740_to_fp16"), val = tensor(0x1.1p-20)]; + tensor var_2741_cast_fp16 = add(x = reduce_mean_125_cast_fp16, y = var_2740_to_fp16)[name = tensor("op_2741_cast_fp16")]; + tensor var_2742_cast_fp16 = sqrt(x = var_2741_cast_fp16)[name = tensor("op_2742_cast_fp16")]; + tensor x_307_cast_fp16 = real_div(x = sub_76_cast_fp16, y = var_2742_cast_fp16)[name = tensor("x_307_cast_fp16")]; + tensor var_2744_promoted_to_fp16 = const()[name = tensor("op_2744_promoted_to_fp16"), val = tensor(0x1p+0)]; + tensor var_2745_cast_fp16 = add(x = var_2733_cast_fp16_1, y = var_2744_promoted_to_fp16)[name = tensor("op_2745_cast_fp16")]; + tensor var_2746_cast_fp16 = mul(x = x_307_cast_fp16, y = var_2745_cast_fp16)[name = tensor("op_2746_cast_fp16")]; + tensor input_473_cast_fp16 = add(x = var_2746_cast_fp16, y = var_2733_cast_fp16_0)[name = tensor("input_473_cast_fp16")]; + tensor linear_181_cast_fp16 = linear(bias = flow_net_final_layer_linear_bias_to_fp16, weight = flow_net_final_layer_linear_weight_to_fp16, x = input_473_cast_fp16)[name = tensor("linear_181_cast_fp16")]; + tensor var_2757_to_fp16 = const()[name = tensor("op_2757_to_fp16"), val = tensor(0x1p-3)]; + tensor var_2758_cast_fp16 = mul(x = linear_181_cast_fp16, y = var_2757_to_fp16)[name = tensor("op_2758_cast_fp16")]; + tensor input_475_cast_fp16 = add(x = input_407_cast_fp16, y = var_2758_cast_fp16)[name = tensor("input_475_cast_fp16")]; + tensor linear_182_cast_fp16 = linear(bias = flow_net_input_proj_bias_to_fp16, weight = flow_net_input_proj_weight_to_fp16, x = input_475_cast_fp16)[name = tensor("linear_182_cast_fp16")]; + tensor input_479_to_fp16 = const()[name = tensor("input_479_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(19010432)))]; + tensor input_481_cast_fp16 = silu(x = input_479_to_fp16)[name = tensor("input_481_cast_fp16")]; + tensor linear_184_cast_fp16 = linear(bias = flow_net_time_embed_0_mlp_2_bias_to_fp16, weight = flow_net_time_embed_0_mlp_2_weight_to_fp16, x = input_481_cast_fp16)[name = tensor("linear_184_cast_fp16")]; + tensor reduce_mean_126_axes_0 = const()[name = tensor("reduce_mean_126_axes_0"), val = tensor([-1])]; + tensor reduce_mean_126_keep_dims_0 = const()[name = tensor("reduce_mean_126_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_126_cast_fp16 = reduce_mean(axes = reduce_mean_126_axes_0, keep_dims = reduce_mean_126_keep_dims_0, x = linear_184_cast_fp16)[name = tensor("reduce_mean_126_cast_fp16")]; + tensor sub_77_cast_fp16 = sub(x = linear_184_cast_fp16, y = reduce_mean_126_cast_fp16)[name = tensor("sub_77_cast_fp16")]; + tensor square_63_cast_fp16 = square(x = sub_77_cast_fp16)[name = tensor("square_63_cast_fp16")]; + tensor reduce_mean_127_axes_0 = const()[name = tensor("reduce_mean_127_axes_0"), val = tensor([-1])]; + tensor reduce_mean_127_keep_dims_0 = const()[name = tensor("reduce_mean_127_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_127_cast_fp16 = reduce_mean(axes = reduce_mean_127_axes_0, keep_dims = reduce_mean_127_keep_dims_0, x = square_63_cast_fp16)[name = tensor("reduce_mean_127_cast_fp16")]; + tensor real_div_14_to_fp16 = const()[name = tensor("real_div_14_to_fp16"), val = tensor(0x1.008p+0)]; + tensor mul_14_cast_fp16 = mul(x = reduce_mean_127_cast_fp16, y = real_div_14_to_fp16)[name = tensor("mul_14_cast_fp16")]; + tensor var_2824_to_fp16 = const()[name = tensor("op_2824_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_127_cast_fp16_0 = add(x = mul_14_cast_fp16, y = var_2824_to_fp16)[name = tensor("var_127_cast_fp16")]; + tensor var_2827_epsilon_0 = const()[name = tensor("op_2827_epsilon_0"), val = tensor(0x1.197998p-40)]; + tensor var_2827_cast_fp16 = rsqrt(epsilon = var_2827_epsilon_0, x = var_127_cast_fp16_0)[name = tensor("op_2827_cast_fp16")]; + tensor var_2828_cast_fp16 = mul(x = const_3_to_fp16, y = var_2827_cast_fp16)[name = tensor("op_2828_cast_fp16")]; + tensor var_2829_cast_fp16 = mul(x = linear_184_cast_fp16, y = var_2828_cast_fp16)[name = tensor("op_2829_cast_fp16")]; + tensor input_485_to_fp16 = const()[name = tensor("input_485_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(19011520)))]; + tensor input_487_cast_fp16 = silu(x = input_485_to_fp16)[name = tensor("input_487_cast_fp16")]; + tensor linear_186_cast_fp16 = linear(bias = flow_net_time_embed_1_mlp_2_bias_to_fp16, weight = flow_net_time_embed_1_mlp_2_weight_to_fp16, x = input_487_cast_fp16)[name = tensor("linear_186_cast_fp16")]; + tensor reduce_mean_128_axes_0 = const()[name = tensor("reduce_mean_128_axes_0"), val = tensor([-1])]; + tensor reduce_mean_128_keep_dims_0 = const()[name = tensor("reduce_mean_128_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_128_cast_fp16 = reduce_mean(axes = reduce_mean_128_axes_0, keep_dims = reduce_mean_128_keep_dims_0, x = linear_186_cast_fp16)[name = tensor("reduce_mean_128_cast_fp16")]; + tensor sub_79_cast_fp16 = sub(x = linear_186_cast_fp16, y = reduce_mean_128_cast_fp16)[name = tensor("sub_79_cast_fp16")]; + tensor square_64_cast_fp16 = square(x = sub_79_cast_fp16)[name = tensor("square_64_cast_fp16")]; + tensor reduce_mean_129_axes_0 = const()[name = tensor("reduce_mean_129_axes_0"), val = tensor([-1])]; + tensor reduce_mean_129_keep_dims_0 = const()[name = tensor("reduce_mean_129_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_129_cast_fp16 = reduce_mean(axes = reduce_mean_129_axes_0, keep_dims = reduce_mean_129_keep_dims_0, x = square_64_cast_fp16)[name = tensor("reduce_mean_129_cast_fp16")]; + tensor real_div_15_to_fp16 = const()[name = tensor("real_div_15_to_fp16"), val = tensor(0x1.008p+0)]; + tensor mul_15_cast_fp16 = mul(x = reduce_mean_129_cast_fp16, y = real_div_15_to_fp16)[name = tensor("mul_15_cast_fp16")]; + tensor var_2851_to_fp16 = const()[name = tensor("op_2851_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_129_cast_fp16 = add(x = mul_15_cast_fp16, y = var_2851_to_fp16)[name = tensor("var_129_cast_fp16")]; + tensor var_2854_epsilon_0 = const()[name = tensor("op_2854_epsilon_0"), val = tensor(0x1.197998p-40)]; + tensor var_2854_cast_fp16 = rsqrt(epsilon = var_2854_epsilon_0, x = var_129_cast_fp16)[name = tensor("op_2854_cast_fp16")]; + tensor var_2855_cast_fp16 = mul(x = const_5_to_fp16, y = var_2854_cast_fp16)[name = tensor("op_2855_cast_fp16")]; + tensor var_2856_cast_fp16 = mul(x = linear_186_cast_fp16, y = var_2855_cast_fp16)[name = tensor("op_2856_cast_fp16")]; + tensor var_2858_cast_fp16 = add(x = var_2829_cast_fp16, y = var_2856_cast_fp16)[name = tensor("op_2858_cast_fp16")]; + tensor _inversed_t_combined_y_0_to_fp16 = const()[name = tensor("_inversed_t_combined_y_0_to_fp16"), val = tensor(0x1p-1)]; + tensor _inversed_t_combined_cast_fp16 = mul(x = var_2858_cast_fp16, y = _inversed_t_combined_y_0_to_fp16)[name = tensor("_inversed_t_combined_cast_fp16")]; + tensor input_489_cast_fp16 = add(x = _inversed_t_combined_cast_fp16, y = linear_5_cast_fp16)[name = tensor("input_489_cast_fp16")]; + tensor input_491_cast_fp16 = silu(x = input_489_cast_fp16)[name = tensor("input_491_cast_fp16")]; + tensor linear_188_cast_fp16 = linear(bias = flow_net_res_blocks_0_adaLN_modulation_1_bias_to_fp16, weight = flow_net_res_blocks_0_adaLN_modulation_1_weight_to_fp16, x = input_491_cast_fp16)[name = tensor("linear_188_cast_fp16")]; + tensor var_2873_split_sizes_0 = const()[name = tensor("op_2873_split_sizes_0"), val = tensor([512, 512, 512])]; + tensor var_2873_axis_0 = const()[name = tensor("op_2873_axis_0"), val = tensor(-1)]; + tensor var_2873_cast_fp16_0, tensor var_2873_cast_fp16_1, tensor var_2873_cast_fp16_2 = split(axis = var_2873_axis_0, split_sizes = var_2873_split_sizes_0, x = linear_188_cast_fp16)[name = tensor("op_2873_cast_fp16")]; + tensor mean_99_axes_0 = const()[name = tensor("mean_99_axes_0"), val = tensor([-1])]; + tensor mean_99_keep_dims_0 = const()[name = tensor("mean_99_keep_dims_0"), val = tensor(true)]; + tensor mean_99_cast_fp16 = reduce_mean(axes = mean_99_axes_0, keep_dims = mean_99_keep_dims_0, x = linear_182_cast_fp16)[name = tensor("mean_99_cast_fp16")]; + tensor sub_81_cast_fp16 = sub(x = linear_182_cast_fp16, y = mean_99_cast_fp16)[name = tensor("sub_81_cast_fp16")]; + tensor square_65_cast_fp16 = square(x = sub_81_cast_fp16)[name = tensor("square_65_cast_fp16")]; + tensor reduce_mean_131_axes_0 = const()[name = tensor("reduce_mean_131_axes_0"), val = tensor([-1])]; + tensor reduce_mean_131_keep_dims_0 = const()[name = tensor("reduce_mean_131_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_131_cast_fp16 = reduce_mean(axes = reduce_mean_131_axes_0, keep_dims = reduce_mean_131_keep_dims_0, x = square_65_cast_fp16)[name = tensor("reduce_mean_131_cast_fp16")]; + tensor var_2883_to_fp16 = const()[name = tensor("op_2883_to_fp16"), val = tensor(0x1.1p-20)]; + tensor var_2884_cast_fp16 = add(x = reduce_mean_131_cast_fp16, y = var_2883_to_fp16)[name = tensor("op_2884_cast_fp16")]; + tensor var_2885_cast_fp16 = sqrt(x = var_2884_cast_fp16)[name = tensor("op_2885_cast_fp16")]; + tensor x_315_cast_fp16 = real_div(x = sub_81_cast_fp16, y = var_2885_cast_fp16)[name = tensor("x_315_cast_fp16")]; + tensor var_2887_cast_fp16 = mul(x = x_315_cast_fp16, y = flow_net_res_blocks_0_in_ln_weight_to_fp16)[name = tensor("op_2887_cast_fp16")]; + tensor x_317_cast_fp16 = add(x = var_2887_cast_fp16, y = flow_net_res_blocks_0_in_ln_bias_to_fp16)[name = tensor("x_317_cast_fp16")]; + tensor var_2889_promoted_to_fp16 = const()[name = tensor("op_2889_promoted_to_fp16"), val = tensor(0x1p+0)]; + tensor var_2890_cast_fp16 = add(x = var_2873_cast_fp16_1, y = var_2889_promoted_to_fp16)[name = tensor("op_2890_cast_fp16")]; + tensor var_2891_cast_fp16 = mul(x = x_317_cast_fp16, y = var_2890_cast_fp16)[name = tensor("op_2891_cast_fp16")]; + tensor input_493_cast_fp16 = add(x = var_2891_cast_fp16, y = var_2873_cast_fp16_0)[name = tensor("input_493_cast_fp16")]; + tensor linear_189_cast_fp16 = linear(bias = flow_net_res_blocks_0_mlp_0_bias_to_fp16, weight = flow_net_res_blocks_0_mlp_0_weight_to_fp16, x = input_493_cast_fp16)[name = tensor("linear_189_cast_fp16")]; + tensor input_497_cast_fp16 = silu(x = linear_189_cast_fp16)[name = tensor("input_497_cast_fp16")]; + tensor linear_190_cast_fp16 = linear(bias = flow_net_res_blocks_0_mlp_2_bias_to_fp16, weight = flow_net_res_blocks_0_mlp_2_weight_to_fp16, x = input_497_cast_fp16)[name = tensor("linear_190_cast_fp16")]; + tensor var_2902_cast_fp16 = mul(x = var_2873_cast_fp16_2, y = linear_190_cast_fp16)[name = tensor("op_2902_cast_fp16")]; + tensor x_319_cast_fp16 = add(x = linear_182_cast_fp16, y = var_2902_cast_fp16)[name = tensor("x_319_cast_fp16")]; + tensor linear_191_cast_fp16 = linear(bias = flow_net_res_blocks_1_adaLN_modulation_1_bias_to_fp16, weight = flow_net_res_blocks_1_adaLN_modulation_1_weight_to_fp16, x = input_491_cast_fp16)[name = tensor("linear_191_cast_fp16")]; + tensor var_2912_split_sizes_0 = const()[name = tensor("op_2912_split_sizes_0"), val = tensor([512, 512, 512])]; + tensor var_2912_axis_0 = const()[name = tensor("op_2912_axis_0"), val = tensor(-1)]; + tensor var_2912_cast_fp16_0, tensor var_2912_cast_fp16_1, tensor var_2912_cast_fp16_2 = split(axis = var_2912_axis_0, split_sizes = var_2912_split_sizes_0, x = linear_191_cast_fp16)[name = tensor("op_2912_cast_fp16")]; + tensor mean_101_axes_0 = const()[name = tensor("mean_101_axes_0"), val = tensor([-1])]; + tensor mean_101_keep_dims_0 = const()[name = tensor("mean_101_keep_dims_0"), val = tensor(true)]; + tensor mean_101_cast_fp16 = reduce_mean(axes = mean_101_axes_0, keep_dims = mean_101_keep_dims_0, x = x_319_cast_fp16)[name = tensor("mean_101_cast_fp16")]; + tensor sub_82_cast_fp16 = sub(x = x_319_cast_fp16, y = mean_101_cast_fp16)[name = tensor("sub_82_cast_fp16")]; + tensor square_66_cast_fp16 = square(x = sub_82_cast_fp16)[name = tensor("square_66_cast_fp16")]; + tensor reduce_mean_133_axes_0 = const()[name = tensor("reduce_mean_133_axes_0"), val = tensor([-1])]; + tensor reduce_mean_133_keep_dims_0 = const()[name = tensor("reduce_mean_133_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_133_cast_fp16 = reduce_mean(axes = reduce_mean_133_axes_0, keep_dims = reduce_mean_133_keep_dims_0, x = square_66_cast_fp16)[name = tensor("reduce_mean_133_cast_fp16")]; + tensor var_2922_to_fp16 = const()[name = tensor("op_2922_to_fp16"), val = tensor(0x1.1p-20)]; + tensor var_2923_cast_fp16 = add(x = reduce_mean_133_cast_fp16, y = var_2922_to_fp16)[name = tensor("op_2923_cast_fp16")]; + tensor var_2924_cast_fp16 = sqrt(x = var_2923_cast_fp16)[name = tensor("op_2924_cast_fp16")]; + tensor x_321_cast_fp16 = real_div(x = sub_82_cast_fp16, y = var_2924_cast_fp16)[name = tensor("x_321_cast_fp16")]; + tensor var_2926_cast_fp16 = mul(x = x_321_cast_fp16, y = flow_net_res_blocks_1_in_ln_weight_to_fp16)[name = tensor("op_2926_cast_fp16")]; + tensor x_323_cast_fp16 = add(x = var_2926_cast_fp16, y = flow_net_res_blocks_1_in_ln_bias_to_fp16)[name = tensor("x_323_cast_fp16")]; + tensor var_2928_promoted_to_fp16 = const()[name = tensor("op_2928_promoted_to_fp16"), val = tensor(0x1p+0)]; + tensor var_2929_cast_fp16 = add(x = var_2912_cast_fp16_1, y = var_2928_promoted_to_fp16)[name = tensor("op_2929_cast_fp16")]; + tensor var_2930_cast_fp16 = mul(x = x_323_cast_fp16, y = var_2929_cast_fp16)[name = tensor("op_2930_cast_fp16")]; + tensor input_501_cast_fp16 = add(x = var_2930_cast_fp16, y = var_2912_cast_fp16_0)[name = tensor("input_501_cast_fp16")]; + tensor linear_192_cast_fp16 = linear(bias = flow_net_res_blocks_1_mlp_0_bias_to_fp16, weight = flow_net_res_blocks_1_mlp_0_weight_to_fp16, x = input_501_cast_fp16)[name = tensor("linear_192_cast_fp16")]; + tensor input_505_cast_fp16 = silu(x = linear_192_cast_fp16)[name = tensor("input_505_cast_fp16")]; + tensor linear_193_cast_fp16 = linear(bias = flow_net_res_blocks_1_mlp_2_bias_to_fp16, weight = flow_net_res_blocks_1_mlp_2_weight_to_fp16, x = input_505_cast_fp16)[name = tensor("linear_193_cast_fp16")]; + tensor var_2941_cast_fp16 = mul(x = var_2912_cast_fp16_2, y = linear_193_cast_fp16)[name = tensor("op_2941_cast_fp16")]; + tensor x_325_cast_fp16 = add(x = x_319_cast_fp16, y = var_2941_cast_fp16)[name = tensor("x_325_cast_fp16")]; + tensor linear_194_cast_fp16 = linear(bias = flow_net_res_blocks_2_adaLN_modulation_1_bias_to_fp16, weight = flow_net_res_blocks_2_adaLN_modulation_1_weight_to_fp16, x = input_491_cast_fp16)[name = tensor("linear_194_cast_fp16")]; + tensor var_2951_split_sizes_0 = const()[name = tensor("op_2951_split_sizes_0"), val = tensor([512, 512, 512])]; + tensor var_2951_axis_0 = const()[name = tensor("op_2951_axis_0"), val = tensor(-1)]; + tensor var_2951_cast_fp16_0, tensor var_2951_cast_fp16_1, tensor var_2951_cast_fp16_2 = split(axis = var_2951_axis_0, split_sizes = var_2951_split_sizes_0, x = linear_194_cast_fp16)[name = tensor("op_2951_cast_fp16")]; + tensor mean_103_axes_0 = const()[name = tensor("mean_103_axes_0"), val = tensor([-1])]; + tensor mean_103_keep_dims_0 = const()[name = tensor("mean_103_keep_dims_0"), val = tensor(true)]; + tensor mean_103_cast_fp16 = reduce_mean(axes = mean_103_axes_0, keep_dims = mean_103_keep_dims_0, x = x_325_cast_fp16)[name = tensor("mean_103_cast_fp16")]; + tensor sub_83_cast_fp16 = sub(x = x_325_cast_fp16, y = mean_103_cast_fp16)[name = tensor("sub_83_cast_fp16")]; + tensor square_67_cast_fp16 = square(x = sub_83_cast_fp16)[name = tensor("square_67_cast_fp16")]; + tensor reduce_mean_135_axes_0 = const()[name = tensor("reduce_mean_135_axes_0"), val = tensor([-1])]; + tensor reduce_mean_135_keep_dims_0 = const()[name = tensor("reduce_mean_135_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_135_cast_fp16 = reduce_mean(axes = reduce_mean_135_axes_0, keep_dims = reduce_mean_135_keep_dims_0, x = square_67_cast_fp16)[name = tensor("reduce_mean_135_cast_fp16")]; + tensor var_2961_to_fp16 = const()[name = tensor("op_2961_to_fp16"), val = tensor(0x1.1p-20)]; + tensor var_2962_cast_fp16 = add(x = reduce_mean_135_cast_fp16, y = var_2961_to_fp16)[name = tensor("op_2962_cast_fp16")]; + tensor var_2963_cast_fp16 = sqrt(x = var_2962_cast_fp16)[name = tensor("op_2963_cast_fp16")]; + tensor x_327_cast_fp16 = real_div(x = sub_83_cast_fp16, y = var_2963_cast_fp16)[name = tensor("x_327_cast_fp16")]; + tensor var_2965_cast_fp16 = mul(x = x_327_cast_fp16, y = flow_net_res_blocks_2_in_ln_weight_to_fp16)[name = tensor("op_2965_cast_fp16")]; + tensor x_329_cast_fp16 = add(x = var_2965_cast_fp16, y = flow_net_res_blocks_2_in_ln_bias_to_fp16)[name = tensor("x_329_cast_fp16")]; + tensor var_2967_promoted_to_fp16 = const()[name = tensor("op_2967_promoted_to_fp16"), val = tensor(0x1p+0)]; + tensor var_2968_cast_fp16 = add(x = var_2951_cast_fp16_1, y = var_2967_promoted_to_fp16)[name = tensor("op_2968_cast_fp16")]; + tensor var_2969_cast_fp16 = mul(x = x_329_cast_fp16, y = var_2968_cast_fp16)[name = tensor("op_2969_cast_fp16")]; + tensor input_509_cast_fp16 = add(x = var_2969_cast_fp16, y = var_2951_cast_fp16_0)[name = tensor("input_509_cast_fp16")]; + tensor linear_195_cast_fp16 = linear(bias = flow_net_res_blocks_2_mlp_0_bias_to_fp16, weight = flow_net_res_blocks_2_mlp_0_weight_to_fp16, x = input_509_cast_fp16)[name = tensor("linear_195_cast_fp16")]; + tensor input_513_cast_fp16 = silu(x = linear_195_cast_fp16)[name = tensor("input_513_cast_fp16")]; + tensor linear_196_cast_fp16 = linear(bias = flow_net_res_blocks_2_mlp_2_bias_to_fp16, weight = flow_net_res_blocks_2_mlp_2_weight_to_fp16, x = input_513_cast_fp16)[name = tensor("linear_196_cast_fp16")]; + tensor var_2980_cast_fp16 = mul(x = var_2951_cast_fp16_2, y = linear_196_cast_fp16)[name = tensor("op_2980_cast_fp16")]; + tensor x_331_cast_fp16 = add(x = x_325_cast_fp16, y = var_2980_cast_fp16)[name = tensor("x_331_cast_fp16")]; + tensor linear_197_cast_fp16 = linear(bias = flow_net_res_blocks_3_adaLN_modulation_1_bias_to_fp16, weight = flow_net_res_blocks_3_adaLN_modulation_1_weight_to_fp16, x = input_491_cast_fp16)[name = tensor("linear_197_cast_fp16")]; + tensor var_2990_split_sizes_0 = const()[name = tensor("op_2990_split_sizes_0"), val = tensor([512, 512, 512])]; + tensor var_2990_axis_0 = const()[name = tensor("op_2990_axis_0"), val = tensor(-1)]; + tensor var_2990_cast_fp16_0, tensor var_2990_cast_fp16_1, tensor var_2990_cast_fp16_2 = split(axis = var_2990_axis_0, split_sizes = var_2990_split_sizes_0, x = linear_197_cast_fp16)[name = tensor("op_2990_cast_fp16")]; + tensor mean_105_axes_0 = const()[name = tensor("mean_105_axes_0"), val = tensor([-1])]; + tensor mean_105_keep_dims_0 = const()[name = tensor("mean_105_keep_dims_0"), val = tensor(true)]; + tensor mean_105_cast_fp16 = reduce_mean(axes = mean_105_axes_0, keep_dims = mean_105_keep_dims_0, x = x_331_cast_fp16)[name = tensor("mean_105_cast_fp16")]; + tensor sub_84_cast_fp16 = sub(x = x_331_cast_fp16, y = mean_105_cast_fp16)[name = tensor("sub_84_cast_fp16")]; + tensor square_68_cast_fp16 = square(x = sub_84_cast_fp16)[name = tensor("square_68_cast_fp16")]; + tensor reduce_mean_137_axes_0 = const()[name = tensor("reduce_mean_137_axes_0"), val = tensor([-1])]; + tensor reduce_mean_137_keep_dims_0 = const()[name = tensor("reduce_mean_137_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_137_cast_fp16 = reduce_mean(axes = reduce_mean_137_axes_0, keep_dims = reduce_mean_137_keep_dims_0, x = square_68_cast_fp16)[name = tensor("reduce_mean_137_cast_fp16")]; + tensor var_3000_to_fp16 = const()[name = tensor("op_3000_to_fp16"), val = tensor(0x1.1p-20)]; + tensor var_3001_cast_fp16 = add(x = reduce_mean_137_cast_fp16, y = var_3000_to_fp16)[name = tensor("op_3001_cast_fp16")]; + tensor var_3002_cast_fp16 = sqrt(x = var_3001_cast_fp16)[name = tensor("op_3002_cast_fp16")]; + tensor x_333_cast_fp16 = real_div(x = sub_84_cast_fp16, y = var_3002_cast_fp16)[name = tensor("x_333_cast_fp16")]; + tensor var_3004_cast_fp16 = mul(x = x_333_cast_fp16, y = flow_net_res_blocks_3_in_ln_weight_to_fp16)[name = tensor("op_3004_cast_fp16")]; + tensor x_335_cast_fp16 = add(x = var_3004_cast_fp16, y = flow_net_res_blocks_3_in_ln_bias_to_fp16)[name = tensor("x_335_cast_fp16")]; + tensor var_3006_promoted_to_fp16 = const()[name = tensor("op_3006_promoted_to_fp16"), val = tensor(0x1p+0)]; + tensor var_3007_cast_fp16 = add(x = var_2990_cast_fp16_1, y = var_3006_promoted_to_fp16)[name = tensor("op_3007_cast_fp16")]; + tensor var_3008_cast_fp16 = mul(x = x_335_cast_fp16, y = var_3007_cast_fp16)[name = tensor("op_3008_cast_fp16")]; + tensor input_517_cast_fp16 = add(x = var_3008_cast_fp16, y = var_2990_cast_fp16_0)[name = tensor("input_517_cast_fp16")]; + tensor linear_198_cast_fp16 = linear(bias = flow_net_res_blocks_3_mlp_0_bias_to_fp16, weight = flow_net_res_blocks_3_mlp_0_weight_to_fp16, x = input_517_cast_fp16)[name = tensor("linear_198_cast_fp16")]; + tensor input_521_cast_fp16 = silu(x = linear_198_cast_fp16)[name = tensor("input_521_cast_fp16")]; + tensor linear_199_cast_fp16 = linear(bias = flow_net_res_blocks_3_mlp_2_bias_to_fp16, weight = flow_net_res_blocks_3_mlp_2_weight_to_fp16, x = input_521_cast_fp16)[name = tensor("linear_199_cast_fp16")]; + tensor var_3019_cast_fp16 = mul(x = var_2990_cast_fp16_2, y = linear_199_cast_fp16)[name = tensor("op_3019_cast_fp16")]; + tensor x_337_cast_fp16 = add(x = x_331_cast_fp16, y = var_3019_cast_fp16)[name = tensor("x_337_cast_fp16")]; + tensor linear_200_cast_fp16 = linear(bias = flow_net_res_blocks_4_adaLN_modulation_1_bias_to_fp16, weight = flow_net_res_blocks_4_adaLN_modulation_1_weight_to_fp16, x = input_491_cast_fp16)[name = tensor("linear_200_cast_fp16")]; + tensor var_3029_split_sizes_0 = const()[name = tensor("op_3029_split_sizes_0"), val = tensor([512, 512, 512])]; + tensor var_3029_axis_0 = const()[name = tensor("op_3029_axis_0"), val = tensor(-1)]; + tensor var_3029_cast_fp16_0, tensor var_3029_cast_fp16_1, tensor var_3029_cast_fp16_2 = split(axis = var_3029_axis_0, split_sizes = var_3029_split_sizes_0, x = linear_200_cast_fp16)[name = tensor("op_3029_cast_fp16")]; + tensor mean_107_axes_0 = const()[name = tensor("mean_107_axes_0"), val = tensor([-1])]; + tensor mean_107_keep_dims_0 = const()[name = tensor("mean_107_keep_dims_0"), val = tensor(true)]; + tensor mean_107_cast_fp16 = reduce_mean(axes = mean_107_axes_0, keep_dims = mean_107_keep_dims_0, x = x_337_cast_fp16)[name = tensor("mean_107_cast_fp16")]; + tensor sub_85_cast_fp16 = sub(x = x_337_cast_fp16, y = mean_107_cast_fp16)[name = tensor("sub_85_cast_fp16")]; + tensor square_69_cast_fp16 = square(x = sub_85_cast_fp16)[name = tensor("square_69_cast_fp16")]; + tensor reduce_mean_139_axes_0 = const()[name = tensor("reduce_mean_139_axes_0"), val = tensor([-1])]; + tensor reduce_mean_139_keep_dims_0 = const()[name = tensor("reduce_mean_139_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_139_cast_fp16 = reduce_mean(axes = reduce_mean_139_axes_0, keep_dims = reduce_mean_139_keep_dims_0, x = square_69_cast_fp16)[name = tensor("reduce_mean_139_cast_fp16")]; + tensor var_3039_to_fp16 = const()[name = tensor("op_3039_to_fp16"), val = tensor(0x1.1p-20)]; + tensor var_3040_cast_fp16 = add(x = reduce_mean_139_cast_fp16, y = var_3039_to_fp16)[name = tensor("op_3040_cast_fp16")]; + tensor var_3041_cast_fp16 = sqrt(x = var_3040_cast_fp16)[name = tensor("op_3041_cast_fp16")]; + tensor x_339_cast_fp16 = real_div(x = sub_85_cast_fp16, y = var_3041_cast_fp16)[name = tensor("x_339_cast_fp16")]; + tensor var_3043_cast_fp16 = mul(x = x_339_cast_fp16, y = flow_net_res_blocks_4_in_ln_weight_to_fp16)[name = tensor("op_3043_cast_fp16")]; + tensor x_341_cast_fp16 = add(x = var_3043_cast_fp16, y = flow_net_res_blocks_4_in_ln_bias_to_fp16)[name = tensor("x_341_cast_fp16")]; + tensor var_3045_promoted_to_fp16 = const()[name = tensor("op_3045_promoted_to_fp16"), val = tensor(0x1p+0)]; + tensor var_3046_cast_fp16 = add(x = var_3029_cast_fp16_1, y = var_3045_promoted_to_fp16)[name = tensor("op_3046_cast_fp16")]; + tensor var_3047_cast_fp16 = mul(x = x_341_cast_fp16, y = var_3046_cast_fp16)[name = tensor("op_3047_cast_fp16")]; + tensor input_525_cast_fp16 = add(x = var_3047_cast_fp16, y = var_3029_cast_fp16_0)[name = tensor("input_525_cast_fp16")]; + tensor linear_201_cast_fp16 = linear(bias = flow_net_res_blocks_4_mlp_0_bias_to_fp16, weight = flow_net_res_blocks_4_mlp_0_weight_to_fp16, x = input_525_cast_fp16)[name = tensor("linear_201_cast_fp16")]; + tensor input_529_cast_fp16 = silu(x = linear_201_cast_fp16)[name = tensor("input_529_cast_fp16")]; + tensor linear_202_cast_fp16 = linear(bias = flow_net_res_blocks_4_mlp_2_bias_to_fp16, weight = flow_net_res_blocks_4_mlp_2_weight_to_fp16, x = input_529_cast_fp16)[name = tensor("linear_202_cast_fp16")]; + tensor var_3058_cast_fp16 = mul(x = var_3029_cast_fp16_2, y = linear_202_cast_fp16)[name = tensor("op_3058_cast_fp16")]; + tensor x_343_cast_fp16 = add(x = x_337_cast_fp16, y = var_3058_cast_fp16)[name = tensor("x_343_cast_fp16")]; + tensor linear_203_cast_fp16 = linear(bias = flow_net_res_blocks_5_adaLN_modulation_1_bias_to_fp16, weight = flow_net_res_blocks_5_adaLN_modulation_1_weight_to_fp16, x = input_491_cast_fp16)[name = tensor("linear_203_cast_fp16")]; + tensor var_3068_split_sizes_0 = const()[name = tensor("op_3068_split_sizes_0"), val = tensor([512, 512, 512])]; + tensor var_3068_axis_0 = const()[name = tensor("op_3068_axis_0"), val = tensor(-1)]; + tensor var_3068_cast_fp16_0, tensor var_3068_cast_fp16_1, tensor var_3068_cast_fp16_2 = split(axis = var_3068_axis_0, split_sizes = var_3068_split_sizes_0, x = linear_203_cast_fp16)[name = tensor("op_3068_cast_fp16")]; + tensor mean_109_axes_0 = const()[name = tensor("mean_109_axes_0"), val = tensor([-1])]; + tensor mean_109_keep_dims_0 = const()[name = tensor("mean_109_keep_dims_0"), val = tensor(true)]; + tensor mean_109_cast_fp16 = reduce_mean(axes = mean_109_axes_0, keep_dims = mean_109_keep_dims_0, x = x_343_cast_fp16)[name = tensor("mean_109_cast_fp16")]; + tensor sub_86_cast_fp16 = sub(x = x_343_cast_fp16, y = mean_109_cast_fp16)[name = tensor("sub_86_cast_fp16")]; + tensor square_70_cast_fp16 = square(x = sub_86_cast_fp16)[name = tensor("square_70_cast_fp16")]; + tensor reduce_mean_141_axes_0 = const()[name = tensor("reduce_mean_141_axes_0"), val = tensor([-1])]; + tensor reduce_mean_141_keep_dims_0 = const()[name = tensor("reduce_mean_141_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_141_cast_fp16 = reduce_mean(axes = reduce_mean_141_axes_0, keep_dims = reduce_mean_141_keep_dims_0, x = square_70_cast_fp16)[name = tensor("reduce_mean_141_cast_fp16")]; + tensor var_3078_to_fp16 = const()[name = tensor("op_3078_to_fp16"), val = tensor(0x1.1p-20)]; + tensor var_3079_cast_fp16 = add(x = reduce_mean_141_cast_fp16, y = var_3078_to_fp16)[name = tensor("op_3079_cast_fp16")]; + tensor var_3080_cast_fp16 = sqrt(x = var_3079_cast_fp16)[name = tensor("op_3080_cast_fp16")]; + tensor x_345_cast_fp16 = real_div(x = sub_86_cast_fp16, y = var_3080_cast_fp16)[name = tensor("x_345_cast_fp16")]; + tensor var_3082_cast_fp16 = mul(x = x_345_cast_fp16, y = flow_net_res_blocks_5_in_ln_weight_to_fp16)[name = tensor("op_3082_cast_fp16")]; + tensor x_347_cast_fp16 = add(x = var_3082_cast_fp16, y = flow_net_res_blocks_5_in_ln_bias_to_fp16)[name = tensor("x_347_cast_fp16")]; + tensor var_3084_promoted_to_fp16 = const()[name = tensor("op_3084_promoted_to_fp16"), val = tensor(0x1p+0)]; + tensor var_3085_cast_fp16 = add(x = var_3068_cast_fp16_1, y = var_3084_promoted_to_fp16)[name = tensor("op_3085_cast_fp16")]; + tensor var_3086_cast_fp16 = mul(x = x_347_cast_fp16, y = var_3085_cast_fp16)[name = tensor("op_3086_cast_fp16")]; + tensor input_533_cast_fp16 = add(x = var_3086_cast_fp16, y = var_3068_cast_fp16_0)[name = tensor("input_533_cast_fp16")]; + tensor linear_204_cast_fp16 = linear(bias = flow_net_res_blocks_5_mlp_0_bias_to_fp16, weight = flow_net_res_blocks_5_mlp_0_weight_to_fp16, x = input_533_cast_fp16)[name = tensor("linear_204_cast_fp16")]; + tensor input_537_cast_fp16 = silu(x = linear_204_cast_fp16)[name = tensor("input_537_cast_fp16")]; + tensor linear_205_cast_fp16 = linear(bias = flow_net_res_blocks_5_mlp_2_bias_to_fp16, weight = flow_net_res_blocks_5_mlp_2_weight_to_fp16, x = input_537_cast_fp16)[name = tensor("linear_205_cast_fp16")]; + tensor var_3097_cast_fp16 = mul(x = var_3068_cast_fp16_2, y = linear_205_cast_fp16)[name = tensor("op_3097_cast_fp16")]; + tensor x_349_cast_fp16 = add(x = x_343_cast_fp16, y = var_3097_cast_fp16)[name = tensor("x_349_cast_fp16")]; + tensor linear_206_cast_fp16 = linear(bias = flow_net_final_layer_adaLN_modulation_1_bias_to_fp16, weight = flow_net_final_layer_adaLN_modulation_1_weight_to_fp16, x = input_491_cast_fp16)[name = tensor("linear_206_cast_fp16")]; + tensor var_3106_split_sizes_0 = const()[name = tensor("op_3106_split_sizes_0"), val = tensor([512, 512])]; + tensor var_3106_axis_0 = const()[name = tensor("op_3106_axis_0"), val = tensor(-1)]; + tensor var_3106_cast_fp16_0, tensor var_3106_cast_fp16_1 = split(axis = var_3106_axis_0, split_sizes = var_3106_split_sizes_0, x = linear_206_cast_fp16)[name = tensor("op_3106_cast_fp16")]; + tensor mean_axes_0 = const()[name = tensor("mean_axes_0"), val = tensor([-1])]; + tensor mean_keep_dims_0 = const()[name = tensor("mean_keep_dims_0"), val = tensor(true)]; + tensor mean_cast_fp16 = reduce_mean(axes = mean_axes_0, keep_dims = mean_keep_dims_0, x = x_349_cast_fp16)[name = tensor("mean_cast_fp16")]; + tensor sub_87_cast_fp16 = sub(x = x_349_cast_fp16, y = mean_cast_fp16)[name = tensor("sub_87_cast_fp16")]; + tensor square_71_cast_fp16 = square(x = sub_87_cast_fp16)[name = tensor("square_71_cast_fp16")]; + tensor reduce_mean_143_axes_0 = const()[name = tensor("reduce_mean_143_axes_0"), val = tensor([-1])]; + tensor reduce_mean_143_keep_dims_0 = const()[name = tensor("reduce_mean_143_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_143_cast_fp16 = reduce_mean(axes = reduce_mean_143_axes_0, keep_dims = reduce_mean_143_keep_dims_0, x = square_71_cast_fp16)[name = tensor("reduce_mean_143_cast_fp16")]; + tensor var_3113_to_fp16 = const()[name = tensor("op_3113_to_fp16"), val = tensor(0x1.1p-20)]; + tensor var_3114_cast_fp16 = add(x = reduce_mean_143_cast_fp16, y = var_3113_to_fp16)[name = tensor("op_3114_cast_fp16")]; + tensor var_3115_cast_fp16 = sqrt(x = var_3114_cast_fp16)[name = tensor("op_3115_cast_fp16")]; + tensor x_cast_fp16 = real_div(x = sub_87_cast_fp16, y = var_3115_cast_fp16)[name = tensor("x_cast_fp16")]; + tensor var_3117_promoted_to_fp16 = const()[name = tensor("op_3117_promoted_to_fp16"), val = tensor(0x1p+0)]; + tensor var_3118_cast_fp16 = add(x = var_3106_cast_fp16_1, y = var_3117_promoted_to_fp16)[name = tensor("op_3118_cast_fp16")]; + tensor var_3119_cast_fp16 = mul(x = x_cast_fp16, y = var_3118_cast_fp16)[name = tensor("op_3119_cast_fp16")]; + tensor input_cast_fp16 = add(x = var_3119_cast_fp16, y = var_3106_cast_fp16_0)[name = tensor("input_cast_fp16")]; + tensor linear_207_cast_fp16 = linear(bias = flow_net_final_layer_linear_bias_to_fp16, weight = flow_net_final_layer_linear_weight_to_fp16, x = input_cast_fp16)[name = tensor("linear_207_cast_fp16")]; + tensor var_3124_to_fp16 = const()[name = tensor("op_3124_to_fp16"), val = tensor(0x1p-3)]; + tensor var_3125_cast_fp16 = mul(x = linear_207_cast_fp16, y = var_3124_to_fp16)[name = tensor("op_3125_cast_fp16")]; + tensor var_3127_cast_fp16 = add(x = input_475_cast_fp16, y = var_3125_cast_fp16)[name = tensor("op_3127_cast_fp16")]; + tensor var_3127_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_3127_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor latent_final = cast(dtype = var_3127_cast_fp16_to_fp32_dtype_0, x = var_3127_cast_fp16)[name = tensor("cast_144")]; + } -> (latent_final); +} \ No newline at end of file diff --git a/v2.1/french_24l/flow_decoder_fused.mlmodelc/weights/weight.bin b/v2.1/french_24l/flow_decoder_fused.mlmodelc/weights/weight.bin new file mode 100644 index 0000000000000000000000000000000000000000..8c89b4fc5f8f7888ff89324e0d63102f6611fe96 --- /dev/null +++ b/v2.1/french_24l/flow_decoder_fused.mlmodelc/weights/weight.bin @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:7feb662b6fb8571b8699c2a70bfaa7b178b9d5dab8e52c2d2b435a68ce60861a +size 19012608 diff --git a/v2.1/french_24l/flow_decoder_fused.mlpackage/Data/com.apple.CoreML/model.mlmodel b/v2.1/french_24l/flow_decoder_fused.mlpackage/Data/com.apple.CoreML/model.mlmodel new file mode 100644 index 0000000000000000000000000000000000000000..9b4b449e90e77f954ec11180d263fdc812edd784 --- /dev/null +++ b/v2.1/french_24l/flow_decoder_fused.mlpackage/Data/com.apple.CoreML/model.mlmodel @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:58e18b6eee62b168bc4f5419f92380aa84b2bcfe8301ffed92bd5c0120e59be1 +size 301733 diff --git a/v2.1/french_24l/flow_decoder_fused.mlpackage/Data/com.apple.CoreML/weights/weight.bin b/v2.1/french_24l/flow_decoder_fused.mlpackage/Data/com.apple.CoreML/weights/weight.bin new file mode 100644 index 0000000000000000000000000000000000000000..8c89b4fc5f8f7888ff89324e0d63102f6611fe96 --- /dev/null +++ b/v2.1/french_24l/flow_decoder_fused.mlpackage/Data/com.apple.CoreML/weights/weight.bin @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:7feb662b6fb8571b8699c2a70bfaa7b178b9d5dab8e52c2d2b435a68ce60861a +size 19012608 diff --git a/v2.1/french_24l/flow_decoder_fused.mlpackage/Manifest.json b/v2.1/french_24l/flow_decoder_fused.mlpackage/Manifest.json new file mode 100644 index 0000000000000000000000000000000000000000..414acc47ee5a46e9f9399d1fa94f6c69e2c3562e --- /dev/null +++ b/v2.1/french_24l/flow_decoder_fused.mlpackage/Manifest.json @@ -0,0 +1,18 @@ +{ + "fileFormatVersion": "1.0.0", + "itemInfoEntries": { + "8044E13F-A888-4DA4-AA11-976315E3660F": { + "author": "com.apple.CoreML", + "description": "CoreML Model Specification", + "name": "model.mlmodel", + "path": "com.apple.CoreML/model.mlmodel" + }, + "F1A00FD8-1E17-454E-90D9-6A1E39FEA973": { + "author": "com.apple.CoreML", + "description": "CoreML Model Weights", + "name": "weights", + "path": "com.apple.CoreML/weights" + } + }, + "rootModelIdentifier": "8044E13F-A888-4DA4-AA11-976315E3660F" +} diff --git a/v2.1/french_24l/flowlm_step.mlmodelc/analytics/coremldata.bin b/v2.1/french_24l/flowlm_step.mlmodelc/analytics/coremldata.bin new file mode 100644 index 0000000000000000000000000000000000000000..d0d249c380d86328047cb957dc0ed514796cce36 --- /dev/null +++ b/v2.1/french_24l/flowlm_step.mlmodelc/analytics/coremldata.bin @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:b1907418a22a52cf4107624cfb8dd50e3faacc94df44edf7a9698e3633b1e31d +size 243 diff --git a/v2.1/french_24l/flowlm_step.mlmodelc/coremldata.bin b/v2.1/french_24l/flowlm_step.mlmodelc/coremldata.bin new file mode 100644 index 0000000000000000000000000000000000000000..e951008a779220c3c4680c05ef18bce2653857be --- /dev/null +++ b/v2.1/french_24l/flowlm_step.mlmodelc/coremldata.bin @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:46a80e790a9a3c3fee8aec899911032a4aef5a8f8904d61190ca71912d0d3035 +size 3559 diff --git a/v2.1/french_24l/flowlm_step.mlmodelc/model.mil b/v2.1/french_24l/flowlm_step.mlmodelc/model.mil new file mode 100644 index 0000000000000000000000000000000000000000..2a7cf641b17448f6d819faf9c83f92de50855d8e --- /dev/null +++ b/v2.1/french_24l/flowlm_step.mlmodelc/model.mil @@ -0,0 +1,5337 @@ +program(1.0) +[buildInfo = dict, tensor>({{"coremlc-component-MIL", "3520.4.1"}, {"coremlc-version", "3520.5.1"}, {"coremltools-component-torch", "2.12.0"}, {"coremltools-source-dialect", "TorchScript"}, {"coremltools-version", "9.0"}})] +{ + func main(tensor bos_emb, tensor cache0, tensor cache1, tensor cache10, tensor cache11, tensor cache12, tensor cache13, tensor cache14, tensor cache15, tensor cache16, tensor cache17, tensor cache18, tensor cache19, tensor cache2, tensor cache20, tensor cache21, tensor cache22, tensor cache23, tensor cache3, tensor cache4, tensor cache5, tensor cache6, tensor cache7, tensor cache8, tensor cache9, tensor position0, tensor position1, tensor position10, tensor position11, tensor position12, tensor position13, tensor position14, tensor position15, tensor position16, tensor position17, tensor position18, tensor position19, tensor position2, tensor position20, tensor position21, tensor position22, tensor position23, tensor position3, tensor position4, tensor position5, tensor position6, tensor position7, tensor position8, tensor position9, tensor sequence) { + tensor sequence_to_fp16_dtype_0 = const()[name = tensor("sequence_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor sequence_to_fp16 = cast(dtype = sequence_to_fp16_dtype_0, x = sequence)[name = tensor("cast_555")]; + tensor var_198_cast_fp16 = not_equal(x = sequence_to_fp16, y = sequence_to_fp16)[name = tensor("op_198_cast_fp16")]; + tensor expand_dims_0_axes_0 = const()[name = tensor("expand_dims_0_axes_0"), val = tensor([0, 1])]; + tensor bos_emb_to_fp16_dtype_0 = const()[name = tensor("bos_emb_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor bos_emb_to_fp16 = cast(dtype = bos_emb_to_fp16_dtype_0, x = bos_emb)[name = tensor("cast_554")]; + tensor expand_dims_0_cast_fp16 = expand_dims(axes = expand_dims_0_axes_0, x = bos_emb_to_fp16)[name = tensor("expand_dims_0_cast_fp16")]; + tensor input_1_cast_fp16 = select(a = expand_dims_0_cast_fp16, b = sequence_to_fp16, cond = var_198_cast_fp16)[name = tensor("input_1_cast_fp16")]; + tensor input_linear_weight_to_fp16 = const()[name = tensor("input_linear_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(64)))]; + tensor linear_0_bias_0_to_fp16 = const()[name = tensor("linear_0_bias_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(65664)))]; + tensor linear_0_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = input_linear_weight_to_fp16, x = input_1_cast_fp16)[name = tensor("linear_0_cast_fp16")]; + tensor input_5_axes_0 = const()[name = tensor("input_5_axes_0"), val = tensor([-1])]; + tensor norm0_1_weight_to_fp16 = const()[name = tensor("norm0_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(67776)))]; + tensor norm0_1_bias_to_fp16 = const()[name = tensor("norm0_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(69888)))]; + tensor var_204_to_fp16 = const()[name = tensor("op_204_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_5_cast_fp16 = layer_norm(axes = input_5_axes_0, beta = norm0_1_bias_to_fp16, epsilon = var_204_to_fp16, gamma = norm0_1_weight_to_fp16, x = linear_0_cast_fp16)[name = tensor("input_5_cast_fp16")]; + tensor attn0_in_proj_weight_to_fp16 = const()[name = tensor("attn0_in_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(72000)))]; + tensor linear_1_bias_0_to_fp16 = const()[name = tensor("linear_1_bias_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(6363520)))]; + tensor linear_1_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = attn0_in_proj_weight_to_fp16, x = input_5_cast_fp16)[name = tensor("linear_1_cast_fp16")]; + tensor var_218 = const()[name = tensor("op_218"), val = tensor([1, 1, 3, 16, 64])]; + tensor qkv_1_cast_fp16 = reshape(shape = var_218, x = linear_1_cast_fp16)[name = tensor("qkv_1_cast_fp16")]; + tensor q_1_begin_0 = const()[name = tensor("q_1_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor q_1_end_0 = const()[name = tensor("q_1_end_0"), val = tensor([1, 1, 1, 16, 64])]; + tensor q_1_end_mask_0 = const()[name = tensor("q_1_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor q_1_squeeze_mask_0 = const()[name = tensor("q_1_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor q_1_cast_fp16 = slice_by_index(begin = q_1_begin_0, end = q_1_end_0, end_mask = q_1_end_mask_0, squeeze_mask = q_1_squeeze_mask_0, x = qkv_1_cast_fp16)[name = tensor("q_1_cast_fp16")]; + tensor k_1_begin_0 = const()[name = tensor("k_1_begin_0"), val = tensor([0, 0, 1, 0, 0])]; + tensor k_1_end_0 = const()[name = tensor("k_1_end_0"), val = tensor([1, 1, 2, 16, 64])]; + tensor k_1_end_mask_0 = const()[name = tensor("k_1_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor k_1_squeeze_mask_0 = const()[name = tensor("k_1_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor k_1_cast_fp16 = slice_by_index(begin = k_1_begin_0, end = k_1_end_0, end_mask = k_1_end_mask_0, squeeze_mask = k_1_squeeze_mask_0, x = qkv_1_cast_fp16)[name = tensor("k_1_cast_fp16")]; + tensor v_1_begin_0 = const()[name = tensor("v_1_begin_0"), val = tensor([0, 0, 2, 0, 0])]; + tensor v_1_end_0 = const()[name = tensor("v_1_end_0"), val = tensor([1, 1, 3, 16, 64])]; + tensor v_1_end_mask_0 = const()[name = tensor("v_1_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor v_1_squeeze_mask_0 = const()[name = tensor("v_1_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor v_1_cast_fp16 = slice_by_index(begin = v_1_begin_0, end = v_1_end_0, end_mask = v_1_end_mask_0, squeeze_mask = v_1_squeeze_mask_0, x = qkv_1_cast_fp16)[name = tensor("v_1_cast_fp16")]; + tensor var_284 = const()[name = tensor("op_284"), val = tensor([1, 1, 1, 1])]; + tensor position0_to_fp16_dtype_0 = const()[name = tensor("position0_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor position0_to_fp16 = cast(dtype = position0_to_fp16_dtype_0, x = position0)[name = tensor("cast_553")]; + tensor ts_5_cast_fp16 = reshape(shape = var_284, x = position0_to_fp16)[name = tensor("ts_5_cast_fp16")]; + tensor var_291 = const()[name = tensor("op_291"), val = tensor([1, 1, 16, 32, 2])]; + tensor q_complex_1_cast_fp16 = reshape(shape = var_291, x = q_1_cast_fp16)[name = tensor("q_complex_1_cast_fp16")]; + tensor var_298 = const()[name = tensor("op_298"), val = tensor([1, 1, 16, 32, 2])]; + tensor k_complex_1_cast_fp16 = reshape(shape = var_298, x = k_1_cast_fp16)[name = tensor("k_complex_1_cast_fp16")]; + tensor var_302_begin_0 = const()[name = tensor("op_302_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_302_end_0 = const()[name = tensor("op_302_end_0"), val = tensor([1, 1, 16, 32, 1])]; + tensor var_302_end_mask_0 = const()[name = tensor("op_302_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_302_squeeze_mask_0 = const()[name = tensor("op_302_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_302_cast_fp16 = slice_by_index(begin = var_302_begin_0, end = var_302_end_0, end_mask = var_302_end_mask_0, squeeze_mask = var_302_squeeze_mask_0, x = q_complex_1_cast_fp16)[name = tensor("op_302_cast_fp16")]; + tensor var_310_begin_0 = const()[name = tensor("op_310_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_310_end_0 = const()[name = tensor("op_310_end_0"), val = tensor([1, 1, 16, 32, 2])]; + tensor var_310_end_mask_0 = const()[name = tensor("op_310_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_310_squeeze_mask_0 = const()[name = tensor("op_310_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_310_cast_fp16 = slice_by_index(begin = var_310_begin_0, end = var_310_end_0, end_mask = var_310_end_mask_0, squeeze_mask = var_310_squeeze_mask_0, x = q_complex_1_cast_fp16)[name = tensor("op_310_cast_fp16")]; + tensor var_318_begin_0 = const()[name = tensor("op_318_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_318_end_0 = const()[name = tensor("op_318_end_0"), val = tensor([1, 1, 16, 32, 1])]; + tensor var_318_end_mask_0 = const()[name = tensor("op_318_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_318_squeeze_mask_0 = const()[name = tensor("op_318_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_318_cast_fp16 = slice_by_index(begin = var_318_begin_0, end = var_318_end_0, end_mask = var_318_end_mask_0, squeeze_mask = var_318_squeeze_mask_0, x = k_complex_1_cast_fp16)[name = tensor("op_318_cast_fp16")]; + tensor var_326_begin_0 = const()[name = tensor("op_326_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_326_end_0 = const()[name = tensor("op_326_end_0"), val = tensor([1, 1, 16, 32, 2])]; + tensor var_326_end_mask_0 = const()[name = tensor("op_326_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_326_squeeze_mask_0 = const()[name = tensor("op_326_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_326_cast_fp16 = slice_by_index(begin = var_326_begin_0, end = var_326_end_0, end_mask = var_326_end_mask_0, squeeze_mask = var_326_squeeze_mask_0, x = k_complex_1_cast_fp16)[name = tensor("op_326_cast_fp16")]; + tensor freqs_1_to_fp16 = const()[name = tensor("freqs_1_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(6369728)))]; + tensor var_332_cast_fp16 = mul(x = freqs_1_to_fp16, y = ts_5_cast_fp16)[name = tensor("op_332_cast_fp16")]; + tensor rotr_1_cast_fp16 = cos(x = var_332_cast_fp16)[name = tensor("rotr_1_cast_fp16")]; + tensor roti_1_cast_fp16 = sin(x = var_332_cast_fp16)[name = tensor("roti_1_cast_fp16")]; + tensor var_336_cast_fp16 = mul(x = var_302_cast_fp16, y = rotr_1_cast_fp16)[name = tensor("op_336_cast_fp16")]; + tensor var_337_cast_fp16 = mul(x = var_310_cast_fp16, y = roti_1_cast_fp16)[name = tensor("op_337_cast_fp16")]; + tensor qor_1_cast_fp16 = sub(x = var_336_cast_fp16, y = var_337_cast_fp16)[name = tensor("qor_1_cast_fp16")]; + tensor var_340_cast_fp16 = mul(x = var_302_cast_fp16, y = roti_1_cast_fp16)[name = tensor("op_340_cast_fp16")]; + tensor var_341_cast_fp16 = mul(x = var_310_cast_fp16, y = rotr_1_cast_fp16)[name = tensor("op_341_cast_fp16")]; + tensor qoi_1_cast_fp16 = add(x = var_340_cast_fp16, y = var_341_cast_fp16)[name = tensor("qoi_1_cast_fp16")]; + tensor var_344_cast_fp16 = mul(x = var_318_cast_fp16, y = rotr_1_cast_fp16)[name = tensor("op_344_cast_fp16")]; + tensor var_345_cast_fp16 = mul(x = var_326_cast_fp16, y = roti_1_cast_fp16)[name = tensor("op_345_cast_fp16")]; + tensor kor_1_cast_fp16 = sub(x = var_344_cast_fp16, y = var_345_cast_fp16)[name = tensor("kor_1_cast_fp16")]; + tensor var_348_cast_fp16 = mul(x = var_318_cast_fp16, y = roti_1_cast_fp16)[name = tensor("op_348_cast_fp16")]; + tensor var_349_cast_fp16 = mul(x = var_326_cast_fp16, y = rotr_1_cast_fp16)[name = tensor("op_349_cast_fp16")]; + tensor koi_1_cast_fp16 = add(x = var_348_cast_fp16, y = var_349_cast_fp16)[name = tensor("koi_1_cast_fp16")]; + tensor qo_1_axis_0 = const()[name = tensor("qo_1_axis_0"), val = tensor(-1)]; + tensor qo_1_cast_fp16 = stack(axis = qo_1_axis_0, values = (qor_1_cast_fp16, qoi_1_cast_fp16))[name = tensor("qo_1_cast_fp16")]; + tensor ko_1_axis_0 = const()[name = tensor("ko_1_axis_0"), val = tensor(-1)]; + tensor ko_1_cast_fp16 = stack(axis = ko_1_axis_0, values = (kor_1_cast_fp16, koi_1_cast_fp16))[name = tensor("ko_1_cast_fp16")]; + tensor var_382 = const()[name = tensor("op_382"), val = tensor([1, 1, 16, 64])]; + tensor q_3_cast_fp16 = reshape(shape = var_382, x = qo_1_cast_fp16)[name = tensor("q_3_cast_fp16")]; + tensor var_388 = const()[name = tensor("op_388"), val = tensor([1, 1, 16, 64])]; + tensor k_3_cast_fp16 = reshape(shape = var_388, x = ko_1_cast_fp16)[name = tensor("k_3_cast_fp16")]; + tensor _inversed_413_y_0_to_fp16 = const()[name = tensor("_inversed_413_y_0_to_fp16"), val = tensor(0x1p-9)]; + tensor _inversed_413_cast_fp16 = mul(x = ts_5_cast_fp16, y = _inversed_413_y_0_to_fp16)[name = tensor("_inversed_413_cast_fp16")]; + tensor var_414_cast_fp16 = floor(x = _inversed_413_cast_fp16)[name = tensor("op_414_cast_fp16")]; + tensor var_415_to_fp16 = const()[name = tensor("op_415_to_fp16"), val = tensor(0x1p+9)]; + tensor var_416_cast_fp16 = mul(x = var_414_cast_fp16, y = var_415_to_fp16)[name = tensor("op_416_cast_fp16")]; + tensor write_indices_float_3_cast_fp16 = sub(x = ts_5_cast_fp16, y = var_416_cast_fp16)[name = tensor("write_indices_float_3_cast_fp16")]; + tensor var_423_dtype_0 = const()[name = tensor("op_423_dtype_0"), val = tensor("int32")]; + tensor write_indices_1_reps_0 = const()[name = tensor("write_indices_1_reps_0"), val = tensor([1, 1, 16, 64])]; + tensor write_indices_float_3_cast_fp16_to_int32 = cast(dtype = var_423_dtype_0, x = write_indices_float_3_cast_fp16)[name = tensor("cast_552")]; + tensor write_indices_1 = tile(reps = write_indices_1_reps_0, x = write_indices_float_3_cast_fp16_to_int32)[name = tensor("write_indices_1")]; + tensor var_433_begin_0 = const()[name = tensor("op_433_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_433_end_0 = const()[name = tensor("op_433_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor var_433_end_mask_0 = const()[name = tensor("op_433_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_433_squeeze_mask_0 = const()[name = tensor("op_433_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor cache0_to_fp16_dtype_0 = const()[name = tensor("cache0_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor cache0_to_fp16 = cast(dtype = cache0_to_fp16_dtype_0, x = cache0)[name = tensor("cast_551")]; + tensor var_433_cast_fp16 = slice_by_index(begin = var_433_begin_0, end = var_433_end_0, end_mask = var_433_end_mask_0, squeeze_mask = var_433_squeeze_mask_0, x = cache0_to_fp16)[name = tensor("op_433_cast_fp16")]; + tensor var_435_axis_0 = const()[name = tensor("op_435_axis_0"), val = tensor(1)]; + tensor var_435_mode_0 = const()[name = tensor("op_435_mode_0"), val = tensor("update")]; + tensor var_435_validate_indices_0 = const()[name = tensor("op_435_validate_indices_0"), val = tensor(false)]; + tensor var_435_cast_fp16 = scatter_along_axis(axis = var_435_axis_0, data = var_433_cast_fp16, indices = write_indices_1, mode = var_435_mode_0, updates = k_3_cast_fp16, validate_indices = var_435_validate_indices_0)[name = tensor("op_435_cast_fp16")]; + tensor concat_2 = const()[name = tensor("concat_2"), val = tensor([0, 0, 0, 0, 0])]; + tensor concat_3 = const()[name = tensor("concat_3"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_1_internal_tensor_assign_1_stride_0 = const()[name = tensor("new_cache_1_internal_tensor_assign_1_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_1_internal_tensor_assign_1_begin_mask_0 = const()[name = tensor("new_cache_1_internal_tensor_assign_1_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_1_internal_tensor_assign_1_end_mask_0 = const()[name = tensor("new_cache_1_internal_tensor_assign_1_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_1_internal_tensor_assign_1_squeeze_mask_0 = const()[name = tensor("new_cache_1_internal_tensor_assign_1_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_48 = const()[name = tensor("shape_48"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_0 = const()[name = tensor("reduce_prod_0"), val = tensor(1048576)]; + tensor range_1d_0_start_0 = const()[name = tensor("range_1d_0_start_0"), val = tensor(0)]; + tensor range_1d_0_step_0 = const()[name = tensor("range_1d_0_step_0"), val = tensor(1)]; + tensor range_1d_0 = range_1d(end = reduce_prod_0, start = range_1d_0_start_0, step = range_1d_0_step_0)[name = tensor("range_1d_0")]; + tensor reshape_0 = reshape(shape = shape_48, x = range_1d_0)[name = tensor("reshape_0")]; + tensor slice_by_index_0 = slice_by_index(begin = concat_2, begin_mask = new_cache_1_internal_tensor_assign_1_begin_mask_0, end = concat_3, end_mask = new_cache_1_internal_tensor_assign_1_end_mask_0, squeeze_mask = new_cache_1_internal_tensor_assign_1_squeeze_mask_0, stride = new_cache_1_internal_tensor_assign_1_stride_0, x = reshape_0)[name = tensor("slice_by_index_0")]; + tensor reshape_1_shape_0 = const()[name = tensor("reshape_1_shape_0"), val = tensor([-1])]; + tensor reshape_1 = reshape(shape = reshape_1_shape_0, x = slice_by_index_0)[name = tensor("reshape_1")]; + tensor reshape_2_shape_0 = const()[name = tensor("reshape_2_shape_0"), val = tensor([-1])]; + tensor reshape_2_cast_fp16 = reshape(shape = reshape_2_shape_0, x = var_435_cast_fp16)[name = tensor("reshape_2_cast_fp16")]; + tensor reshape_3_shape_0 = const()[name = tensor("reshape_3_shape_0"), val = tensor([-1])]; + tensor reshape_3_cast_fp16 = reshape(shape = reshape_3_shape_0, x = cache0_to_fp16)[name = tensor("reshape_3_cast_fp16")]; + tensor scatter_0_mode_0 = const()[name = tensor("scatter_0_mode_0"), val = tensor("update")]; + tensor scatter_0_axis_0 = const()[name = tensor("scatter_0_axis_0"), val = tensor(0)]; + tensor scatter_0_validate_indices_0 = const()[name = tensor("scatter_0_validate_indices_0"), val = tensor(false)]; + tensor scatter_0_cast_fp16 = scatter(axis = scatter_0_axis_0, data = reshape_3_cast_fp16, indices = reshape_1, mode = scatter_0_mode_0, updates = reshape_2_cast_fp16, validate_indices = scatter_0_validate_indices_0)[name = tensor("scatter_0_cast_fp16")]; + tensor reshape_4_cast_fp16 = reshape(shape = shape_48, x = scatter_0_cast_fp16)[name = tensor("reshape_4_cast_fp16")]; + tensor var_443_begin_0 = const()[name = tensor("op_443_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor var_443_end_0 = const()[name = tensor("op_443_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor var_443_end_mask_0 = const()[name = tensor("op_443_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_443_squeeze_mask_0 = const()[name = tensor("op_443_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor var_443_cast_fp16 = slice_by_index(begin = var_443_begin_0, end = var_443_end_0, end_mask = var_443_end_mask_0, squeeze_mask = var_443_squeeze_mask_0, x = reshape_4_cast_fp16)[name = tensor("op_443_cast_fp16")]; + tensor var_445_axis_0 = const()[name = tensor("op_445_axis_0"), val = tensor(1)]; + tensor var_445_mode_0 = const()[name = tensor("op_445_mode_0"), val = tensor("update")]; + tensor var_445_validate_indices_0 = const()[name = tensor("op_445_validate_indices_0"), val = tensor(false)]; + tensor var_445_cast_fp16 = scatter_along_axis(axis = var_445_axis_0, data = var_443_cast_fp16, indices = write_indices_1, mode = var_445_mode_0, updates = v_1_cast_fp16, validate_indices = var_445_validate_indices_0)[name = tensor("op_445_cast_fp16")]; + tensor concat_4 = const()[name = tensor("concat_4"), val = tensor([1, 0, 0, 0, 0])]; + tensor concat_5 = const()[name = tensor("concat_5"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_1_internal_tensor_assign_2_stride_0 = const()[name = tensor("new_cache_1_internal_tensor_assign_2_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_1_internal_tensor_assign_2_begin_mask_0 = const()[name = tensor("new_cache_1_internal_tensor_assign_2_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_1_internal_tensor_assign_2_end_mask_0 = const()[name = tensor("new_cache_1_internal_tensor_assign_2_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_1_internal_tensor_assign_2_squeeze_mask_0 = const()[name = tensor("new_cache_1_internal_tensor_assign_2_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_49 = const()[name = tensor("shape_49"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_1 = const()[name = tensor("reduce_prod_1"), val = tensor(1048576)]; + tensor range_1d_1_start_0 = const()[name = tensor("range_1d_1_start_0"), val = tensor(0)]; + tensor range_1d_1_step_0 = const()[name = tensor("range_1d_1_step_0"), val = tensor(1)]; + tensor range_1d_1 = range_1d(end = reduce_prod_1, start = range_1d_1_start_0, step = range_1d_1_step_0)[name = tensor("range_1d_1")]; + tensor reshape_5 = reshape(shape = shape_49, x = range_1d_1)[name = tensor("reshape_5")]; + tensor slice_by_index_1 = slice_by_index(begin = concat_4, begin_mask = new_cache_1_internal_tensor_assign_2_begin_mask_0, end = concat_5, end_mask = new_cache_1_internal_tensor_assign_2_end_mask_0, squeeze_mask = new_cache_1_internal_tensor_assign_2_squeeze_mask_0, stride = new_cache_1_internal_tensor_assign_2_stride_0, x = reshape_5)[name = tensor("slice_by_index_1")]; + tensor reshape_6_shape_0 = const()[name = tensor("reshape_6_shape_0"), val = tensor([-1])]; + tensor reshape_6 = reshape(shape = reshape_6_shape_0, x = slice_by_index_1)[name = tensor("reshape_6")]; + tensor reshape_7_shape_0 = const()[name = tensor("reshape_7_shape_0"), val = tensor([-1])]; + tensor reshape_7_cast_fp16 = reshape(shape = reshape_7_shape_0, x = var_445_cast_fp16)[name = tensor("reshape_7_cast_fp16")]; + tensor reshape_8_shape_0 = const()[name = tensor("reshape_8_shape_0"), val = tensor([-1])]; + tensor reshape_8_cast_fp16 = reshape(shape = reshape_8_shape_0, x = reshape_4_cast_fp16)[name = tensor("reshape_8_cast_fp16")]; + tensor scatter_1_mode_0 = const()[name = tensor("scatter_1_mode_0"), val = tensor("update")]; + tensor scatter_1_axis_0 = const()[name = tensor("scatter_1_axis_0"), val = tensor(0)]; + tensor scatter_1_validate_indices_0 = const()[name = tensor("scatter_1_validate_indices_0"), val = tensor(false)]; + tensor scatter_1_cast_fp16 = scatter(axis = scatter_1_axis_0, data = reshape_8_cast_fp16, indices = reshape_6, mode = scatter_1_mode_0, updates = reshape_7_cast_fp16, validate_indices = scatter_1_validate_indices_0)[name = tensor("scatter_1_cast_fp16")]; + tensor reshape_9_cast_fp16 = reshape(shape = shape_49, x = scatter_1_cast_fp16)[name = tensor("reshape_9_cast_fp16")]; + tensor reshape_9_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("reshape_9_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor keys_1_begin_0 = const()[name = tensor("keys_1_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor keys_1_end_0 = const()[name = tensor("keys_1_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor keys_1_end_mask_0 = const()[name = tensor("keys_1_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor keys_1_squeeze_mask_0 = const()[name = tensor("keys_1_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor keys_1_cast_fp16 = slice_by_index(begin = keys_1_begin_0, end = keys_1_end_0, end_mask = keys_1_end_mask_0, squeeze_mask = keys_1_squeeze_mask_0, x = reshape_9_cast_fp16)[name = tensor("keys_1_cast_fp16")]; + tensor values_1_begin_0 = const()[name = tensor("values_1_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor values_1_end_0 = const()[name = tensor("values_1_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor values_1_end_mask_0 = const()[name = tensor("values_1_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor values_1_squeeze_mask_0 = const()[name = tensor("values_1_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor values_1_cast_fp16 = slice_by_index(begin = values_1_begin_0, end = values_1_end_0, end_mask = values_1_end_mask_0, squeeze_mask = values_1_squeeze_mask_0, x = reshape_9_cast_fp16)[name = tensor("values_1_cast_fp16")]; + tensor var_457_cast_fp16 = not_equal(x = keys_1_cast_fp16, y = keys_1_cast_fp16)[name = tensor("op_457_cast_fp16")]; + tensor var_463_to_fp16 = const()[name = tensor("op_463_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(6369856)))]; + tensor keys_3_cast_fp16 = select(a = var_463_to_fp16, b = keys_1_cast_fp16, cond = var_457_cast_fp16)[name = tensor("keys_3_cast_fp16")]; + tensor var_465_cast_fp16 = not_equal(x = values_1_cast_fp16, y = values_1_cast_fp16)[name = tensor("op_465_cast_fp16")]; + tensor values_3_cast_fp16 = select(a = var_463_to_fp16, b = values_1_cast_fp16, cond = var_465_cast_fp16)[name = tensor("values_3_cast_fp16")]; + tensor var_489 = const()[name = tensor("op_489"), val = tensor([0, 2, 1, 3])]; + tensor var_505 = const()[name = tensor("op_505"), val = tensor([1, 1, 1])]; + tensor var_506_cast_fp16 = reshape(shape = var_505, x = position0_to_fp16)[name = tensor("op_506_cast_fp16")]; + tensor var_526_to_fp16 = const()[name = tensor("op_526_to_fp16"), val = tensor(0x1p+0)]; + tensor valid_len_1_cast_fp16 = add(x = var_506_cast_fp16, y = var_526_to_fp16)[name = tensor("valid_len_1_cast_fp16")]; + tensor k_positions_1_promoted_to_fp16 = const()[name = tensor("k_positions_1_promoted_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(7418496)))]; + tensor valid_mask_1_cast_fp16 = less(x = k_positions_1_promoted_to_fp16, y = valid_len_1_cast_fp16)[name = tensor("valid_mask_1_cast_fp16")]; + tensor causal_mask_1_cast_fp16 = less_equal(x = k_positions_1_promoted_to_fp16, y = var_506_cast_fp16)[name = tensor("causal_mask_1_cast_fp16")]; + tensor attn_mask_1 = logical_and(x = valid_mask_1_cast_fp16, y = causal_mask_1_cast_fp16)[name = tensor("attn_mask_1")]; + tensor attn_mask_3_axes_0 = const()[name = tensor("attn_mask_3_axes_0"), val = tensor([1])]; + tensor attn_mask_3 = expand_dims(axes = attn_mask_3_axes_0, x = attn_mask_1)[name = tensor("attn_mask_3")]; + tensor var_544_transpose_x_0 = const()[name = tensor("op_544_transpose_x_0"), val = tensor(false)]; + tensor var_544_transpose_y_0 = const()[name = tensor("op_544_transpose_y_0"), val = tensor(false)]; + tensor transpose_72_perm_0 = const()[name = tensor("transpose_72_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_73_perm_0 = const()[name = tensor("transpose_73_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_73 = transpose(perm = transpose_73_perm_0, x = keys_3_cast_fp16)[name = tensor("transpose_213")]; + tensor transpose_72 = transpose(perm = transpose_72_perm_0, x = q_3_cast_fp16)[name = tensor("transpose_214")]; + tensor var_544_cast_fp16 = matmul(transpose_x = var_544_transpose_x_0, transpose_y = var_544_transpose_y_0, x = transpose_72, y = transpose_73)[name = tensor("op_544_cast_fp16")]; + tensor var_538_to_fp16 = const()[name = tensor("op_538_to_fp16"), val = tensor([0x1p-3])]; + tensor attn_weights_1_cast_fp16 = mul(x = var_544_cast_fp16, y = var_538_to_fp16)[name = tensor("attn_weights_1_cast_fp16")]; + tensor var_546 = logical_not(x = attn_mask_3)[name = tensor("op_546")]; + tensor var_547_to_fp16 = const()[name = tensor("op_547_to_fp16"), val = tensor(-inf)]; + tensor attn_weights_3_cast_fp16 = select(a = var_547_to_fp16, b = attn_weights_1_cast_fp16, cond = var_546)[name = tensor("attn_weights_3_cast_fp16")]; + tensor var_549 = const()[name = tensor("op_549"), val = tensor(-1)]; + tensor attn_weights_5_cast_fp16 = softmax(axis = var_549, x = attn_weights_3_cast_fp16)[name = tensor("attn_weights_5_cast_fp16")]; + tensor attn_output_1_transpose_x_0 = const()[name = tensor("attn_output_1_transpose_x_0"), val = tensor(false)]; + tensor attn_output_1_transpose_y_0 = const()[name = tensor("attn_output_1_transpose_y_0"), val = tensor(false)]; + tensor values_5_cast_fp16 = transpose(perm = var_489, x = values_3_cast_fp16)[name = tensor("transpose_215")]; + tensor attn_output_1_cast_fp16 = matmul(transpose_x = attn_output_1_transpose_x_0, transpose_y = attn_output_1_transpose_y_0, x = attn_weights_5_cast_fp16, y = values_5_cast_fp16)[name = tensor("attn_output_1_cast_fp16")]; + tensor var_557 = const()[name = tensor("op_557"), val = tensor([0, 2, 1, 3])]; + tensor var_562 = const()[name = tensor("op_562"), val = tensor([1, 1, 1024])]; + tensor var_558_cast_fp16 = transpose(perm = var_557, x = attn_output_1_cast_fp16)[name = tensor("transpose_212")]; + tensor input_7_cast_fp16 = reshape(shape = var_562, x = var_558_cast_fp16)[name = tensor("input_7_cast_fp16")]; + tensor attn0_out_proj_weight_to_fp16 = const()[name = tensor("attn0_out_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(7419584)))]; + tensor linear_2_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = attn0_out_proj_weight_to_fp16, x = input_7_cast_fp16)[name = tensor("linear_2_cast_fp16")]; + tensor var_568_to_fp16 = const()[name = tensor("op_568_to_fp16"), val = tensor(0x1p+0)]; + tensor var_569_cast_fp16 = add(x = position0_to_fp16, y = var_568_to_fp16)[name = tensor("op_569_cast_fp16")]; + tensor var_569_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_569_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor input_9_cast_fp16 = add(x = linear_0_cast_fp16, y = linear_2_cast_fp16)[name = tensor("input_9_cast_fp16")]; + tensor input_11_axes_0 = const()[name = tensor("input_11_axes_0"), val = tensor([-1])]; + tensor norm0_2_weight_to_fp16 = const()[name = tensor("norm0_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9516800)))]; + tensor norm0_2_bias_to_fp16 = const()[name = tensor("norm0_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9518912)))]; + tensor var_573_to_fp16 = const()[name = tensor("op_573_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_11_cast_fp16 = layer_norm(axes = input_11_axes_0, beta = norm0_2_bias_to_fp16, epsilon = var_573_to_fp16, gamma = norm0_2_weight_to_fp16, x = input_9_cast_fp16)[name = tensor("input_11_cast_fp16")]; + tensor linear0_1_weight_to_fp16 = const()[name = tensor("linear0_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9521024)))]; + tensor linear_3_bias_0_to_fp16 = const()[name = tensor("linear_3_bias_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(17909696)))]; + tensor linear_3_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = linear0_1_weight_to_fp16, x = input_11_cast_fp16)[name = tensor("linear_3_cast_fp16")]; + tensor input_13_mode_0 = const()[name = tensor("input_13_mode_0"), val = tensor("EXACT")]; + tensor input_13_cast_fp16 = gelu(mode = input_13_mode_0, x = linear_3_cast_fp16)[name = tensor("input_13_cast_fp16")]; + tensor linear0_2_weight_to_fp16 = const()[name = tensor("linear0_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(17917952)))]; + tensor linear_4_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = linear0_2_weight_to_fp16, x = input_13_cast_fp16)[name = tensor("linear_4_cast_fp16")]; + tensor input_15_cast_fp16 = add(x = input_9_cast_fp16, y = linear_4_cast_fp16)[name = tensor("input_15_cast_fp16")]; + tensor input_17_axes_0 = const()[name = tensor("input_17_axes_0"), val = tensor([-1])]; + tensor norm1_1_weight_to_fp16 = const()[name = tensor("norm1_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(26306624)))]; + tensor norm1_1_bias_to_fp16 = const()[name = tensor("norm1_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(26308736)))]; + tensor var_590_to_fp16 = const()[name = tensor("op_590_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_17_cast_fp16 = layer_norm(axes = input_17_axes_0, beta = norm1_1_bias_to_fp16, epsilon = var_590_to_fp16, gamma = norm1_1_weight_to_fp16, x = input_15_cast_fp16)[name = tensor("input_17_cast_fp16")]; + tensor attn1_in_proj_weight_to_fp16 = const()[name = tensor("attn1_in_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(26310848)))]; + tensor linear_5_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = attn1_in_proj_weight_to_fp16, x = input_17_cast_fp16)[name = tensor("linear_5_cast_fp16")]; + tensor var_604 = const()[name = tensor("op_604"), val = tensor([1, 1, 3, 16, 64])]; + tensor qkv_3_cast_fp16 = reshape(shape = var_604, x = linear_5_cast_fp16)[name = tensor("qkv_3_cast_fp16")]; + tensor q_7_begin_0 = const()[name = tensor("q_7_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor q_7_end_0 = const()[name = tensor("q_7_end_0"), val = tensor([1, 1, 1, 16, 64])]; + tensor q_7_end_mask_0 = const()[name = tensor("q_7_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor q_7_squeeze_mask_0 = const()[name = tensor("q_7_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor q_7_cast_fp16 = slice_by_index(begin = q_7_begin_0, end = q_7_end_0, end_mask = q_7_end_mask_0, squeeze_mask = q_7_squeeze_mask_0, x = qkv_3_cast_fp16)[name = tensor("q_7_cast_fp16")]; + tensor k_5_begin_0 = const()[name = tensor("k_5_begin_0"), val = tensor([0, 0, 1, 0, 0])]; + tensor k_5_end_0 = const()[name = tensor("k_5_end_0"), val = tensor([1, 1, 2, 16, 64])]; + tensor k_5_end_mask_0 = const()[name = tensor("k_5_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor k_5_squeeze_mask_0 = const()[name = tensor("k_5_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor k_5_cast_fp16 = slice_by_index(begin = k_5_begin_0, end = k_5_end_0, end_mask = k_5_end_mask_0, squeeze_mask = k_5_squeeze_mask_0, x = qkv_3_cast_fp16)[name = tensor("k_5_cast_fp16")]; + tensor v_3_begin_0 = const()[name = tensor("v_3_begin_0"), val = tensor([0, 0, 2, 0, 0])]; + tensor v_3_end_0 = const()[name = tensor("v_3_end_0"), val = tensor([1, 1, 3, 16, 64])]; + tensor v_3_end_mask_0 = const()[name = tensor("v_3_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor v_3_squeeze_mask_0 = const()[name = tensor("v_3_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor v_3_cast_fp16 = slice_by_index(begin = v_3_begin_0, end = v_3_end_0, end_mask = v_3_end_mask_0, squeeze_mask = v_3_squeeze_mask_0, x = qkv_3_cast_fp16)[name = tensor("v_3_cast_fp16")]; + tensor var_670 = const()[name = tensor("op_670"), val = tensor([1, 1, 1, 1])]; + tensor position1_to_fp16_dtype_0 = const()[name = tensor("position1_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor position1_to_fp16 = cast(dtype = position1_to_fp16_dtype_0, x = position1)[name = tensor("cast_548")]; + tensor ts_11_cast_fp16 = reshape(shape = var_670, x = position1_to_fp16)[name = tensor("ts_11_cast_fp16")]; + tensor var_677 = const()[name = tensor("op_677"), val = tensor([1, 1, 16, 32, 2])]; + tensor q_complex_3_cast_fp16 = reshape(shape = var_677, x = q_7_cast_fp16)[name = tensor("q_complex_3_cast_fp16")]; + tensor var_684 = const()[name = tensor("op_684"), val = tensor([1, 1, 16, 32, 2])]; + tensor k_complex_3_cast_fp16 = reshape(shape = var_684, x = k_5_cast_fp16)[name = tensor("k_complex_3_cast_fp16")]; + tensor var_688_begin_0 = const()[name = tensor("op_688_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_688_end_0 = const()[name = tensor("op_688_end_0"), val = tensor([1, 1, 16, 32, 1])]; + tensor var_688_end_mask_0 = const()[name = tensor("op_688_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_688_squeeze_mask_0 = const()[name = tensor("op_688_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_688_cast_fp16 = slice_by_index(begin = var_688_begin_0, end = var_688_end_0, end_mask = var_688_end_mask_0, squeeze_mask = var_688_squeeze_mask_0, x = q_complex_3_cast_fp16)[name = tensor("op_688_cast_fp16")]; + tensor var_696_begin_0 = const()[name = tensor("op_696_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_696_end_0 = const()[name = tensor("op_696_end_0"), val = tensor([1, 1, 16, 32, 2])]; + tensor var_696_end_mask_0 = const()[name = tensor("op_696_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_696_squeeze_mask_0 = const()[name = tensor("op_696_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_696_cast_fp16 = slice_by_index(begin = var_696_begin_0, end = var_696_end_0, end_mask = var_696_end_mask_0, squeeze_mask = var_696_squeeze_mask_0, x = q_complex_3_cast_fp16)[name = tensor("op_696_cast_fp16")]; + tensor var_704_begin_0 = const()[name = tensor("op_704_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_704_end_0 = const()[name = tensor("op_704_end_0"), val = tensor([1, 1, 16, 32, 1])]; + tensor var_704_end_mask_0 = const()[name = tensor("op_704_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_704_squeeze_mask_0 = const()[name = tensor("op_704_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_704_cast_fp16 = slice_by_index(begin = var_704_begin_0, end = var_704_end_0, end_mask = var_704_end_mask_0, squeeze_mask = var_704_squeeze_mask_0, x = k_complex_3_cast_fp16)[name = tensor("op_704_cast_fp16")]; + tensor var_712_begin_0 = const()[name = tensor("op_712_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_712_end_0 = const()[name = tensor("op_712_end_0"), val = tensor([1, 1, 16, 32, 2])]; + tensor var_712_end_mask_0 = const()[name = tensor("op_712_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_712_squeeze_mask_0 = const()[name = tensor("op_712_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_712_cast_fp16 = slice_by_index(begin = var_712_begin_0, end = var_712_end_0, end_mask = var_712_end_mask_0, squeeze_mask = var_712_squeeze_mask_0, x = k_complex_3_cast_fp16)[name = tensor("op_712_cast_fp16")]; + tensor freqs_3_to_fp16 = const()[name = tensor("freqs_3_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(32602368)))]; + tensor var_718_cast_fp16 = mul(x = freqs_3_to_fp16, y = ts_11_cast_fp16)[name = tensor("op_718_cast_fp16")]; + tensor rotr_3_cast_fp16 = cos(x = var_718_cast_fp16)[name = tensor("rotr_3_cast_fp16")]; + tensor roti_3_cast_fp16 = sin(x = var_718_cast_fp16)[name = tensor("roti_3_cast_fp16")]; + tensor var_722_cast_fp16 = mul(x = var_688_cast_fp16, y = rotr_3_cast_fp16)[name = tensor("op_722_cast_fp16")]; + tensor var_723_cast_fp16 = mul(x = var_696_cast_fp16, y = roti_3_cast_fp16)[name = tensor("op_723_cast_fp16")]; + tensor qor_5_cast_fp16 = sub(x = var_722_cast_fp16, y = var_723_cast_fp16)[name = tensor("qor_5_cast_fp16")]; + tensor var_726_cast_fp16 = mul(x = var_688_cast_fp16, y = roti_3_cast_fp16)[name = tensor("op_726_cast_fp16")]; + tensor var_727_cast_fp16 = mul(x = var_696_cast_fp16, y = rotr_3_cast_fp16)[name = tensor("op_727_cast_fp16")]; + tensor qoi_5_cast_fp16 = add(x = var_726_cast_fp16, y = var_727_cast_fp16)[name = tensor("qoi_5_cast_fp16")]; + tensor var_730_cast_fp16 = mul(x = var_704_cast_fp16, y = rotr_3_cast_fp16)[name = tensor("op_730_cast_fp16")]; + tensor var_731_cast_fp16 = mul(x = var_712_cast_fp16, y = roti_3_cast_fp16)[name = tensor("op_731_cast_fp16")]; + tensor kor_5_cast_fp16 = sub(x = var_730_cast_fp16, y = var_731_cast_fp16)[name = tensor("kor_5_cast_fp16")]; + tensor var_734_cast_fp16 = mul(x = var_704_cast_fp16, y = roti_3_cast_fp16)[name = tensor("op_734_cast_fp16")]; + tensor var_735_cast_fp16 = mul(x = var_712_cast_fp16, y = rotr_3_cast_fp16)[name = tensor("op_735_cast_fp16")]; + tensor koi_5_cast_fp16 = add(x = var_734_cast_fp16, y = var_735_cast_fp16)[name = tensor("koi_5_cast_fp16")]; + tensor qo_3_axis_0 = const()[name = tensor("qo_3_axis_0"), val = tensor(-1)]; + tensor qo_3_cast_fp16 = stack(axis = qo_3_axis_0, values = (qor_5_cast_fp16, qoi_5_cast_fp16))[name = tensor("qo_3_cast_fp16")]; + tensor ko_3_axis_0 = const()[name = tensor("ko_3_axis_0"), val = tensor(-1)]; + tensor ko_3_cast_fp16 = stack(axis = ko_3_axis_0, values = (kor_5_cast_fp16, koi_5_cast_fp16))[name = tensor("ko_3_cast_fp16")]; + tensor var_768 = const()[name = tensor("op_768"), val = tensor([1, 1, 16, 64])]; + tensor q_9_cast_fp16 = reshape(shape = var_768, x = qo_3_cast_fp16)[name = tensor("q_9_cast_fp16")]; + tensor var_774 = const()[name = tensor("op_774"), val = tensor([1, 1, 16, 64])]; + tensor k_7_cast_fp16 = reshape(shape = var_774, x = ko_3_cast_fp16)[name = tensor("k_7_cast_fp16")]; + tensor _inversed_799_y_0_to_fp16 = const()[name = tensor("_inversed_799_y_0_to_fp16"), val = tensor(0x1p-9)]; + tensor _inversed_799_cast_fp16 = mul(x = ts_11_cast_fp16, y = _inversed_799_y_0_to_fp16)[name = tensor("_inversed_799_cast_fp16")]; + tensor var_800_cast_fp16 = floor(x = _inversed_799_cast_fp16)[name = tensor("op_800_cast_fp16")]; + tensor var_801_to_fp16 = const()[name = tensor("op_801_to_fp16"), val = tensor(0x1p+9)]; + tensor var_802_cast_fp16 = mul(x = var_800_cast_fp16, y = var_801_to_fp16)[name = tensor("op_802_cast_fp16")]; + tensor write_indices_float_7_cast_fp16 = sub(x = ts_11_cast_fp16, y = var_802_cast_fp16)[name = tensor("write_indices_float_7_cast_fp16")]; + tensor var_809_dtype_0 = const()[name = tensor("op_809_dtype_0"), val = tensor("int32")]; + tensor write_indices_3_reps_0 = const()[name = tensor("write_indices_3_reps_0"), val = tensor([1, 1, 16, 64])]; + tensor write_indices_float_7_cast_fp16_to_int32 = cast(dtype = var_809_dtype_0, x = write_indices_float_7_cast_fp16)[name = tensor("cast_547")]; + tensor write_indices_3 = tile(reps = write_indices_3_reps_0, x = write_indices_float_7_cast_fp16_to_int32)[name = tensor("write_indices_3")]; + tensor var_819_begin_0 = const()[name = tensor("op_819_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_819_end_0 = const()[name = tensor("op_819_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor var_819_end_mask_0 = const()[name = tensor("op_819_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_819_squeeze_mask_0 = const()[name = tensor("op_819_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor cache1_to_fp16_dtype_0 = const()[name = tensor("cache1_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor cache1_to_fp16 = cast(dtype = cache1_to_fp16_dtype_0, x = cache1)[name = tensor("cast_546")]; + tensor var_819_cast_fp16 = slice_by_index(begin = var_819_begin_0, end = var_819_end_0, end_mask = var_819_end_mask_0, squeeze_mask = var_819_squeeze_mask_0, x = cache1_to_fp16)[name = tensor("op_819_cast_fp16")]; + tensor var_821_axis_0 = const()[name = tensor("op_821_axis_0"), val = tensor(1)]; + tensor var_821_mode_0 = const()[name = tensor("op_821_mode_0"), val = tensor("update")]; + tensor var_821_validate_indices_0 = const()[name = tensor("op_821_validate_indices_0"), val = tensor(false)]; + tensor var_821_cast_fp16 = scatter_along_axis(axis = var_821_axis_0, data = var_819_cast_fp16, indices = write_indices_3, mode = var_821_mode_0, updates = k_7_cast_fp16, validate_indices = var_821_validate_indices_0)[name = tensor("op_821_cast_fp16")]; + tensor concat_9 = const()[name = tensor("concat_9"), val = tensor([0, 0, 0, 0, 0])]; + tensor concat_10 = const()[name = tensor("concat_10"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_3_internal_tensor_assign_1_stride_0 = const()[name = tensor("new_cache_3_internal_tensor_assign_1_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_3_internal_tensor_assign_1_begin_mask_0 = const()[name = tensor("new_cache_3_internal_tensor_assign_1_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_3_internal_tensor_assign_1_end_mask_0 = const()[name = tensor("new_cache_3_internal_tensor_assign_1_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_3_internal_tensor_assign_1_squeeze_mask_0 = const()[name = tensor("new_cache_3_internal_tensor_assign_1_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_50 = const()[name = tensor("shape_50"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_2 = const()[name = tensor("reduce_prod_2"), val = tensor(1048576)]; + tensor range_1d_2_start_0 = const()[name = tensor("range_1d_2_start_0"), val = tensor(0)]; + tensor range_1d_2_step_0 = const()[name = tensor("range_1d_2_step_0"), val = tensor(1)]; + tensor range_1d_2 = range_1d(end = reduce_prod_2, start = range_1d_2_start_0, step = range_1d_2_step_0)[name = tensor("range_1d_2")]; + tensor reshape_10 = reshape(shape = shape_50, x = range_1d_2)[name = tensor("reshape_10")]; + tensor slice_by_index_2 = slice_by_index(begin = concat_9, begin_mask = new_cache_3_internal_tensor_assign_1_begin_mask_0, end = concat_10, end_mask = new_cache_3_internal_tensor_assign_1_end_mask_0, squeeze_mask = new_cache_3_internal_tensor_assign_1_squeeze_mask_0, stride = new_cache_3_internal_tensor_assign_1_stride_0, x = reshape_10)[name = tensor("slice_by_index_2")]; + tensor reshape_11_shape_0 = const()[name = tensor("reshape_11_shape_0"), val = tensor([-1])]; + tensor reshape_11 = reshape(shape = reshape_11_shape_0, x = slice_by_index_2)[name = tensor("reshape_11")]; + tensor reshape_12_shape_0 = const()[name = tensor("reshape_12_shape_0"), val = tensor([-1])]; + tensor reshape_12_cast_fp16 = reshape(shape = reshape_12_shape_0, x = var_821_cast_fp16)[name = tensor("reshape_12_cast_fp16")]; + tensor reshape_13_shape_0 = const()[name = tensor("reshape_13_shape_0"), val = tensor([-1])]; + tensor reshape_13_cast_fp16 = reshape(shape = reshape_13_shape_0, x = cache1_to_fp16)[name = tensor("reshape_13_cast_fp16")]; + tensor scatter_2_mode_0 = const()[name = tensor("scatter_2_mode_0"), val = tensor("update")]; + tensor scatter_2_axis_0 = const()[name = tensor("scatter_2_axis_0"), val = tensor(0)]; + tensor scatter_2_validate_indices_0 = const()[name = tensor("scatter_2_validate_indices_0"), val = tensor(false)]; + tensor scatter_2_cast_fp16 = scatter(axis = scatter_2_axis_0, data = reshape_13_cast_fp16, indices = reshape_11, mode = scatter_2_mode_0, updates = reshape_12_cast_fp16, validate_indices = scatter_2_validate_indices_0)[name = tensor("scatter_2_cast_fp16")]; + tensor reshape_14_cast_fp16 = reshape(shape = shape_50, x = scatter_2_cast_fp16)[name = tensor("reshape_14_cast_fp16")]; + tensor var_829_begin_0 = const()[name = tensor("op_829_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor var_829_end_0 = const()[name = tensor("op_829_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor var_829_end_mask_0 = const()[name = tensor("op_829_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_829_squeeze_mask_0 = const()[name = tensor("op_829_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor var_829_cast_fp16 = slice_by_index(begin = var_829_begin_0, end = var_829_end_0, end_mask = var_829_end_mask_0, squeeze_mask = var_829_squeeze_mask_0, x = reshape_14_cast_fp16)[name = tensor("op_829_cast_fp16")]; + tensor var_831_axis_0 = const()[name = tensor("op_831_axis_0"), val = tensor(1)]; + tensor var_831_mode_0 = const()[name = tensor("op_831_mode_0"), val = tensor("update")]; + tensor var_831_validate_indices_0 = const()[name = tensor("op_831_validate_indices_0"), val = tensor(false)]; + tensor var_831_cast_fp16 = scatter_along_axis(axis = var_831_axis_0, data = var_829_cast_fp16, indices = write_indices_3, mode = var_831_mode_0, updates = v_3_cast_fp16, validate_indices = var_831_validate_indices_0)[name = tensor("op_831_cast_fp16")]; + tensor concat_11 = const()[name = tensor("concat_11"), val = tensor([1, 0, 0, 0, 0])]; + tensor concat_12 = const()[name = tensor("concat_12"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_3_internal_tensor_assign_2_stride_0 = const()[name = tensor("new_cache_3_internal_tensor_assign_2_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_3_internal_tensor_assign_2_begin_mask_0 = const()[name = tensor("new_cache_3_internal_tensor_assign_2_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_3_internal_tensor_assign_2_end_mask_0 = const()[name = tensor("new_cache_3_internal_tensor_assign_2_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_3_internal_tensor_assign_2_squeeze_mask_0 = const()[name = tensor("new_cache_3_internal_tensor_assign_2_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_51 = const()[name = tensor("shape_51"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_3 = const()[name = tensor("reduce_prod_3"), val = tensor(1048576)]; + tensor range_1d_3_start_0 = const()[name = tensor("range_1d_3_start_0"), val = tensor(0)]; + tensor range_1d_3_step_0 = const()[name = tensor("range_1d_3_step_0"), val = tensor(1)]; + tensor range_1d_3 = range_1d(end = reduce_prod_3, start = range_1d_3_start_0, step = range_1d_3_step_0)[name = tensor("range_1d_3")]; + tensor reshape_15 = reshape(shape = shape_51, x = range_1d_3)[name = tensor("reshape_15")]; + tensor slice_by_index_3 = slice_by_index(begin = concat_11, begin_mask = new_cache_3_internal_tensor_assign_2_begin_mask_0, end = concat_12, end_mask = new_cache_3_internal_tensor_assign_2_end_mask_0, squeeze_mask = new_cache_3_internal_tensor_assign_2_squeeze_mask_0, stride = new_cache_3_internal_tensor_assign_2_stride_0, x = reshape_15)[name = tensor("slice_by_index_3")]; + tensor reshape_16_shape_0 = const()[name = tensor("reshape_16_shape_0"), val = tensor([-1])]; + tensor reshape_16 = reshape(shape = reshape_16_shape_0, x = slice_by_index_3)[name = tensor("reshape_16")]; + tensor reshape_17_shape_0 = const()[name = tensor("reshape_17_shape_0"), val = tensor([-1])]; + tensor reshape_17_cast_fp16 = reshape(shape = reshape_17_shape_0, x = var_831_cast_fp16)[name = tensor("reshape_17_cast_fp16")]; + tensor reshape_18_shape_0 = const()[name = tensor("reshape_18_shape_0"), val = tensor([-1])]; + tensor reshape_18_cast_fp16 = reshape(shape = reshape_18_shape_0, x = reshape_14_cast_fp16)[name = tensor("reshape_18_cast_fp16")]; + tensor scatter_3_mode_0 = const()[name = tensor("scatter_3_mode_0"), val = tensor("update")]; + tensor scatter_3_axis_0 = const()[name = tensor("scatter_3_axis_0"), val = tensor(0)]; + tensor scatter_3_validate_indices_0 = const()[name = tensor("scatter_3_validate_indices_0"), val = tensor(false)]; + tensor scatter_3_cast_fp16 = scatter(axis = scatter_3_axis_0, data = reshape_18_cast_fp16, indices = reshape_16, mode = scatter_3_mode_0, updates = reshape_17_cast_fp16, validate_indices = scatter_3_validate_indices_0)[name = tensor("scatter_3_cast_fp16")]; + tensor reshape_19_cast_fp16 = reshape(shape = shape_51, x = scatter_3_cast_fp16)[name = tensor("reshape_19_cast_fp16")]; + tensor reshape_19_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("reshape_19_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor keys_7_begin_0 = const()[name = tensor("keys_7_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor keys_7_end_0 = const()[name = tensor("keys_7_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor keys_7_end_mask_0 = const()[name = tensor("keys_7_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor keys_7_squeeze_mask_0 = const()[name = tensor("keys_7_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor keys_7_cast_fp16 = slice_by_index(begin = keys_7_begin_0, end = keys_7_end_0, end_mask = keys_7_end_mask_0, squeeze_mask = keys_7_squeeze_mask_0, x = reshape_19_cast_fp16)[name = tensor("keys_7_cast_fp16")]; + tensor values_7_begin_0 = const()[name = tensor("values_7_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor values_7_end_0 = const()[name = tensor("values_7_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor values_7_end_mask_0 = const()[name = tensor("values_7_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor values_7_squeeze_mask_0 = const()[name = tensor("values_7_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor values_7_cast_fp16 = slice_by_index(begin = values_7_begin_0, end = values_7_end_0, end_mask = values_7_end_mask_0, squeeze_mask = values_7_squeeze_mask_0, x = reshape_19_cast_fp16)[name = tensor("values_7_cast_fp16")]; + tensor var_843_cast_fp16 = not_equal(x = keys_7_cast_fp16, y = keys_7_cast_fp16)[name = tensor("op_843_cast_fp16")]; + tensor keys_9_cast_fp16 = select(a = var_463_to_fp16, b = keys_7_cast_fp16, cond = var_843_cast_fp16)[name = tensor("keys_9_cast_fp16")]; + tensor var_851_cast_fp16 = not_equal(x = values_7_cast_fp16, y = values_7_cast_fp16)[name = tensor("op_851_cast_fp16")]; + tensor values_9_cast_fp16 = select(a = var_463_to_fp16, b = values_7_cast_fp16, cond = var_851_cast_fp16)[name = tensor("values_9_cast_fp16")]; + tensor var_875 = const()[name = tensor("op_875"), val = tensor([0, 2, 1, 3])]; + tensor var_891 = const()[name = tensor("op_891"), val = tensor([1, 1, 1])]; + tensor var_892_cast_fp16 = reshape(shape = var_891, x = position1_to_fp16)[name = tensor("op_892_cast_fp16")]; + tensor var_912_to_fp16 = const()[name = tensor("op_912_to_fp16"), val = tensor(0x1p+0)]; + tensor valid_len_3_cast_fp16 = add(x = var_892_cast_fp16, y = var_912_to_fp16)[name = tensor("valid_len_3_cast_fp16")]; + tensor valid_mask_3_cast_fp16 = less(x = k_positions_1_promoted_to_fp16, y = valid_len_3_cast_fp16)[name = tensor("valid_mask_3_cast_fp16")]; + tensor causal_mask_3_cast_fp16 = less_equal(x = k_positions_1_promoted_to_fp16, y = var_892_cast_fp16)[name = tensor("causal_mask_3_cast_fp16")]; + tensor attn_mask_5 = logical_and(x = valid_mask_3_cast_fp16, y = causal_mask_3_cast_fp16)[name = tensor("attn_mask_5")]; + tensor attn_mask_7_axes_0 = const()[name = tensor("attn_mask_7_axes_0"), val = tensor([1])]; + tensor attn_mask_7 = expand_dims(axes = attn_mask_7_axes_0, x = attn_mask_5)[name = tensor("attn_mask_7")]; + tensor var_930_transpose_x_0 = const()[name = tensor("op_930_transpose_x_0"), val = tensor(false)]; + tensor var_930_transpose_y_0 = const()[name = tensor("op_930_transpose_y_0"), val = tensor(false)]; + tensor transpose_74_perm_0 = const()[name = tensor("transpose_74_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_75_perm_0 = const()[name = tensor("transpose_75_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_75 = transpose(perm = transpose_75_perm_0, x = keys_9_cast_fp16)[name = tensor("transpose_209")]; + tensor transpose_74 = transpose(perm = transpose_74_perm_0, x = q_9_cast_fp16)[name = tensor("transpose_210")]; + tensor var_930_cast_fp16 = matmul(transpose_x = var_930_transpose_x_0, transpose_y = var_930_transpose_y_0, x = transpose_74, y = transpose_75)[name = tensor("op_930_cast_fp16")]; + tensor var_924_to_fp16 = const()[name = tensor("op_924_to_fp16"), val = tensor([0x1p-3])]; + tensor attn_weights_7_cast_fp16 = mul(x = var_930_cast_fp16, y = var_924_to_fp16)[name = tensor("attn_weights_7_cast_fp16")]; + tensor var_932 = logical_not(x = attn_mask_7)[name = tensor("op_932")]; + tensor var_933_to_fp16 = const()[name = tensor("op_933_to_fp16"), val = tensor(-inf)]; + tensor attn_weights_9_cast_fp16 = select(a = var_933_to_fp16, b = attn_weights_7_cast_fp16, cond = var_932)[name = tensor("attn_weights_9_cast_fp16")]; + tensor var_935 = const()[name = tensor("op_935"), val = tensor(-1)]; + tensor attn_weights_11_cast_fp16 = softmax(axis = var_935, x = attn_weights_9_cast_fp16)[name = tensor("attn_weights_11_cast_fp16")]; + tensor attn_output_3_transpose_x_0 = const()[name = tensor("attn_output_3_transpose_x_0"), val = tensor(false)]; + tensor attn_output_3_transpose_y_0 = const()[name = tensor("attn_output_3_transpose_y_0"), val = tensor(false)]; + tensor values_11_cast_fp16 = transpose(perm = var_875, x = values_9_cast_fp16)[name = tensor("transpose_211")]; + tensor attn_output_3_cast_fp16 = matmul(transpose_x = attn_output_3_transpose_x_0, transpose_y = attn_output_3_transpose_y_0, x = attn_weights_11_cast_fp16, y = values_11_cast_fp16)[name = tensor("attn_output_3_cast_fp16")]; + tensor var_943 = const()[name = tensor("op_943"), val = tensor([0, 2, 1, 3])]; + tensor var_948 = const()[name = tensor("op_948"), val = tensor([1, 1, 1024])]; + tensor var_944_cast_fp16 = transpose(perm = var_943, x = attn_output_3_cast_fp16)[name = tensor("transpose_208")]; + tensor input_19_cast_fp16 = reshape(shape = var_948, x = var_944_cast_fp16)[name = tensor("input_19_cast_fp16")]; + tensor attn1_out_proj_weight_to_fp16 = const()[name = tensor("attn1_out_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(32602496)))]; + tensor linear_6_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = attn1_out_proj_weight_to_fp16, x = input_19_cast_fp16)[name = tensor("linear_6_cast_fp16")]; + tensor var_954_to_fp16 = const()[name = tensor("op_954_to_fp16"), val = tensor(0x1p+0)]; + tensor var_955_cast_fp16 = add(x = position1_to_fp16, y = var_954_to_fp16)[name = tensor("op_955_cast_fp16")]; + tensor var_955_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_955_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor input_21_cast_fp16 = add(x = input_15_cast_fp16, y = linear_6_cast_fp16)[name = tensor("input_21_cast_fp16")]; + tensor input_23_axes_0 = const()[name = tensor("input_23_axes_0"), val = tensor([-1])]; + tensor norm1_2_weight_to_fp16 = const()[name = tensor("norm1_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(34699712)))]; + tensor norm1_2_bias_to_fp16 = const()[name = tensor("norm1_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(34701824)))]; + tensor var_959_to_fp16 = const()[name = tensor("op_959_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_23_cast_fp16 = layer_norm(axes = input_23_axes_0, beta = norm1_2_bias_to_fp16, epsilon = var_959_to_fp16, gamma = norm1_2_weight_to_fp16, x = input_21_cast_fp16)[name = tensor("input_23_cast_fp16")]; + tensor linear1_1_weight_to_fp16 = const()[name = tensor("linear1_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(34703936)))]; + tensor linear_7_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = linear1_1_weight_to_fp16, x = input_23_cast_fp16)[name = tensor("linear_7_cast_fp16")]; + tensor input_25_mode_0 = const()[name = tensor("input_25_mode_0"), val = tensor("EXACT")]; + tensor input_25_cast_fp16 = gelu(mode = input_25_mode_0, x = linear_7_cast_fp16)[name = tensor("input_25_cast_fp16")]; + tensor linear1_2_weight_to_fp16 = const()[name = tensor("linear1_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(43092608)))]; + tensor linear_8_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = linear1_2_weight_to_fp16, x = input_25_cast_fp16)[name = tensor("linear_8_cast_fp16")]; + tensor input_27_cast_fp16 = add(x = input_21_cast_fp16, y = linear_8_cast_fp16)[name = tensor("input_27_cast_fp16")]; + tensor input_29_axes_0 = const()[name = tensor("input_29_axes_0"), val = tensor([-1])]; + tensor norm2_1_weight_to_fp16 = const()[name = tensor("norm2_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(51481280)))]; + tensor norm2_1_bias_to_fp16 = const()[name = tensor("norm2_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(51483392)))]; + tensor var_976_to_fp16 = const()[name = tensor("op_976_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_29_cast_fp16 = layer_norm(axes = input_29_axes_0, beta = norm2_1_bias_to_fp16, epsilon = var_976_to_fp16, gamma = norm2_1_weight_to_fp16, x = input_27_cast_fp16)[name = tensor("input_29_cast_fp16")]; + tensor attn2_in_proj_weight_to_fp16 = const()[name = tensor("attn2_in_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(51485504)))]; + tensor linear_9_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = attn2_in_proj_weight_to_fp16, x = input_29_cast_fp16)[name = tensor("linear_9_cast_fp16")]; + tensor var_990 = const()[name = tensor("op_990"), val = tensor([1, 1, 3, 16, 64])]; + tensor qkv_5_cast_fp16 = reshape(shape = var_990, x = linear_9_cast_fp16)[name = tensor("qkv_5_cast_fp16")]; + tensor q_13_begin_0 = const()[name = tensor("q_13_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor q_13_end_0 = const()[name = tensor("q_13_end_0"), val = tensor([1, 1, 1, 16, 64])]; + tensor q_13_end_mask_0 = const()[name = tensor("q_13_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor q_13_squeeze_mask_0 = const()[name = tensor("q_13_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor q_13_cast_fp16 = slice_by_index(begin = q_13_begin_0, end = q_13_end_0, end_mask = q_13_end_mask_0, squeeze_mask = q_13_squeeze_mask_0, x = qkv_5_cast_fp16)[name = tensor("q_13_cast_fp16")]; + tensor k_9_begin_0 = const()[name = tensor("k_9_begin_0"), val = tensor([0, 0, 1, 0, 0])]; + tensor k_9_end_0 = const()[name = tensor("k_9_end_0"), val = tensor([1, 1, 2, 16, 64])]; + tensor k_9_end_mask_0 = const()[name = tensor("k_9_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor k_9_squeeze_mask_0 = const()[name = tensor("k_9_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor k_9_cast_fp16 = slice_by_index(begin = k_9_begin_0, end = k_9_end_0, end_mask = k_9_end_mask_0, squeeze_mask = k_9_squeeze_mask_0, x = qkv_5_cast_fp16)[name = tensor("k_9_cast_fp16")]; + tensor v_5_begin_0 = const()[name = tensor("v_5_begin_0"), val = tensor([0, 0, 2, 0, 0])]; + tensor v_5_end_0 = const()[name = tensor("v_5_end_0"), val = tensor([1, 1, 3, 16, 64])]; + tensor v_5_end_mask_0 = const()[name = tensor("v_5_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor v_5_squeeze_mask_0 = const()[name = tensor("v_5_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor v_5_cast_fp16 = slice_by_index(begin = v_5_begin_0, end = v_5_end_0, end_mask = v_5_end_mask_0, squeeze_mask = v_5_squeeze_mask_0, x = qkv_5_cast_fp16)[name = tensor("v_5_cast_fp16")]; + tensor var_1056 = const()[name = tensor("op_1056"), val = tensor([1, 1, 1, 1])]; + tensor position2_to_fp16_dtype_0 = const()[name = tensor("position2_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor position2_to_fp16 = cast(dtype = position2_to_fp16_dtype_0, x = position2)[name = tensor("cast_543")]; + tensor ts_17_cast_fp16 = reshape(shape = var_1056, x = position2_to_fp16)[name = tensor("ts_17_cast_fp16")]; + tensor var_1063 = const()[name = tensor("op_1063"), val = tensor([1, 1, 16, 32, 2])]; + tensor q_complex_5_cast_fp16 = reshape(shape = var_1063, x = q_13_cast_fp16)[name = tensor("q_complex_5_cast_fp16")]; + tensor var_1070 = const()[name = tensor("op_1070"), val = tensor([1, 1, 16, 32, 2])]; + tensor k_complex_5_cast_fp16 = reshape(shape = var_1070, x = k_9_cast_fp16)[name = tensor("k_complex_5_cast_fp16")]; + tensor var_1074_begin_0 = const()[name = tensor("op_1074_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_1074_end_0 = const()[name = tensor("op_1074_end_0"), val = tensor([1, 1, 16, 32, 1])]; + tensor var_1074_end_mask_0 = const()[name = tensor("op_1074_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_1074_squeeze_mask_0 = const()[name = tensor("op_1074_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_1074_cast_fp16 = slice_by_index(begin = var_1074_begin_0, end = var_1074_end_0, end_mask = var_1074_end_mask_0, squeeze_mask = var_1074_squeeze_mask_0, x = q_complex_5_cast_fp16)[name = tensor("op_1074_cast_fp16")]; + tensor var_1082_begin_0 = const()[name = tensor("op_1082_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_1082_end_0 = const()[name = tensor("op_1082_end_0"), val = tensor([1, 1, 16, 32, 2])]; + tensor var_1082_end_mask_0 = const()[name = tensor("op_1082_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_1082_squeeze_mask_0 = const()[name = tensor("op_1082_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_1082_cast_fp16 = slice_by_index(begin = var_1082_begin_0, end = var_1082_end_0, end_mask = var_1082_end_mask_0, squeeze_mask = var_1082_squeeze_mask_0, x = q_complex_5_cast_fp16)[name = tensor("op_1082_cast_fp16")]; + tensor var_1090_begin_0 = const()[name = tensor("op_1090_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_1090_end_0 = const()[name = tensor("op_1090_end_0"), val = tensor([1, 1, 16, 32, 1])]; + tensor var_1090_end_mask_0 = const()[name = tensor("op_1090_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_1090_squeeze_mask_0 = const()[name = tensor("op_1090_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_1090_cast_fp16 = slice_by_index(begin = var_1090_begin_0, end = var_1090_end_0, end_mask = var_1090_end_mask_0, squeeze_mask = var_1090_squeeze_mask_0, x = k_complex_5_cast_fp16)[name = tensor("op_1090_cast_fp16")]; + tensor var_1098_begin_0 = const()[name = tensor("op_1098_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_1098_end_0 = const()[name = tensor("op_1098_end_0"), val = tensor([1, 1, 16, 32, 2])]; + tensor var_1098_end_mask_0 = const()[name = tensor("op_1098_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_1098_squeeze_mask_0 = const()[name = tensor("op_1098_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_1098_cast_fp16 = slice_by_index(begin = var_1098_begin_0, end = var_1098_end_0, end_mask = var_1098_end_mask_0, squeeze_mask = var_1098_squeeze_mask_0, x = k_complex_5_cast_fp16)[name = tensor("op_1098_cast_fp16")]; + tensor freqs_5_to_fp16 = const()[name = tensor("freqs_5_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(57777024)))]; + tensor var_1104_cast_fp16 = mul(x = freqs_5_to_fp16, y = ts_17_cast_fp16)[name = tensor("op_1104_cast_fp16")]; + tensor rotr_5_cast_fp16 = cos(x = var_1104_cast_fp16)[name = tensor("rotr_5_cast_fp16")]; + tensor roti_5_cast_fp16 = sin(x = var_1104_cast_fp16)[name = tensor("roti_5_cast_fp16")]; + tensor var_1108_cast_fp16 = mul(x = var_1074_cast_fp16, y = rotr_5_cast_fp16)[name = tensor("op_1108_cast_fp16")]; + tensor var_1109_cast_fp16 = mul(x = var_1082_cast_fp16, y = roti_5_cast_fp16)[name = tensor("op_1109_cast_fp16")]; + tensor qor_9_cast_fp16 = sub(x = var_1108_cast_fp16, y = var_1109_cast_fp16)[name = tensor("qor_9_cast_fp16")]; + tensor var_1112_cast_fp16 = mul(x = var_1074_cast_fp16, y = roti_5_cast_fp16)[name = tensor("op_1112_cast_fp16")]; + tensor var_1113_cast_fp16 = mul(x = var_1082_cast_fp16, y = rotr_5_cast_fp16)[name = tensor("op_1113_cast_fp16")]; + tensor qoi_9_cast_fp16 = add(x = var_1112_cast_fp16, y = var_1113_cast_fp16)[name = tensor("qoi_9_cast_fp16")]; + tensor var_1116_cast_fp16 = mul(x = var_1090_cast_fp16, y = rotr_5_cast_fp16)[name = tensor("op_1116_cast_fp16")]; + tensor var_1117_cast_fp16 = mul(x = var_1098_cast_fp16, y = roti_5_cast_fp16)[name = tensor("op_1117_cast_fp16")]; + tensor kor_9_cast_fp16 = sub(x = var_1116_cast_fp16, y = var_1117_cast_fp16)[name = tensor("kor_9_cast_fp16")]; + tensor var_1120_cast_fp16 = mul(x = var_1090_cast_fp16, y = roti_5_cast_fp16)[name = tensor("op_1120_cast_fp16")]; + tensor var_1121_cast_fp16 = mul(x = var_1098_cast_fp16, y = rotr_5_cast_fp16)[name = tensor("op_1121_cast_fp16")]; + tensor koi_9_cast_fp16 = add(x = var_1120_cast_fp16, y = var_1121_cast_fp16)[name = tensor("koi_9_cast_fp16")]; + tensor qo_5_axis_0 = const()[name = tensor("qo_5_axis_0"), val = tensor(-1)]; + tensor qo_5_cast_fp16 = stack(axis = qo_5_axis_0, values = (qor_9_cast_fp16, qoi_9_cast_fp16))[name = tensor("qo_5_cast_fp16")]; + tensor ko_5_axis_0 = const()[name = tensor("ko_5_axis_0"), val = tensor(-1)]; + tensor ko_5_cast_fp16 = stack(axis = ko_5_axis_0, values = (kor_9_cast_fp16, koi_9_cast_fp16))[name = tensor("ko_5_cast_fp16")]; + tensor var_1154 = const()[name = tensor("op_1154"), val = tensor([1, 1, 16, 64])]; + tensor q_15_cast_fp16 = reshape(shape = var_1154, x = qo_5_cast_fp16)[name = tensor("q_15_cast_fp16")]; + tensor var_1160 = const()[name = tensor("op_1160"), val = tensor([1, 1, 16, 64])]; + tensor k_11_cast_fp16 = reshape(shape = var_1160, x = ko_5_cast_fp16)[name = tensor("k_11_cast_fp16")]; + tensor _inversed_1185_y_0_to_fp16 = const()[name = tensor("_inversed_1185_y_0_to_fp16"), val = tensor(0x1p-9)]; + tensor _inversed_1185_cast_fp16 = mul(x = ts_17_cast_fp16, y = _inversed_1185_y_0_to_fp16)[name = tensor("_inversed_1185_cast_fp16")]; + tensor var_1186_cast_fp16 = floor(x = _inversed_1185_cast_fp16)[name = tensor("op_1186_cast_fp16")]; + tensor var_1187_to_fp16 = const()[name = tensor("op_1187_to_fp16"), val = tensor(0x1p+9)]; + tensor var_1188_cast_fp16 = mul(x = var_1186_cast_fp16, y = var_1187_to_fp16)[name = tensor("op_1188_cast_fp16")]; + tensor write_indices_float_11_cast_fp16 = sub(x = ts_17_cast_fp16, y = var_1188_cast_fp16)[name = tensor("write_indices_float_11_cast_fp16")]; + tensor var_1195_dtype_0 = const()[name = tensor("op_1195_dtype_0"), val = tensor("int32")]; + tensor write_indices_5_reps_0 = const()[name = tensor("write_indices_5_reps_0"), val = tensor([1, 1, 16, 64])]; + tensor write_indices_float_11_cast_fp16_to_int32 = cast(dtype = var_1195_dtype_0, x = write_indices_float_11_cast_fp16)[name = tensor("cast_542")]; + tensor write_indices_5 = tile(reps = write_indices_5_reps_0, x = write_indices_float_11_cast_fp16_to_int32)[name = tensor("write_indices_5")]; + tensor var_1205_begin_0 = const()[name = tensor("op_1205_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_1205_end_0 = const()[name = tensor("op_1205_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor var_1205_end_mask_0 = const()[name = tensor("op_1205_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_1205_squeeze_mask_0 = const()[name = tensor("op_1205_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor cache2_to_fp16_dtype_0 = const()[name = tensor("cache2_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor cache2_to_fp16 = cast(dtype = cache2_to_fp16_dtype_0, x = cache2)[name = tensor("cast_541")]; + tensor var_1205_cast_fp16 = slice_by_index(begin = var_1205_begin_0, end = var_1205_end_0, end_mask = var_1205_end_mask_0, squeeze_mask = var_1205_squeeze_mask_0, x = cache2_to_fp16)[name = tensor("op_1205_cast_fp16")]; + tensor var_1207_axis_0 = const()[name = tensor("op_1207_axis_0"), val = tensor(1)]; + tensor var_1207_mode_0 = const()[name = tensor("op_1207_mode_0"), val = tensor("update")]; + tensor var_1207_validate_indices_0 = const()[name = tensor("op_1207_validate_indices_0"), val = tensor(false)]; + tensor var_1207_cast_fp16 = scatter_along_axis(axis = var_1207_axis_0, data = var_1205_cast_fp16, indices = write_indices_5, mode = var_1207_mode_0, updates = k_11_cast_fp16, validate_indices = var_1207_validate_indices_0)[name = tensor("op_1207_cast_fp16")]; + tensor concat_16 = const()[name = tensor("concat_16"), val = tensor([0, 0, 0, 0, 0])]; + tensor concat_17 = const()[name = tensor("concat_17"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_5_internal_tensor_assign_1_stride_0 = const()[name = tensor("new_cache_5_internal_tensor_assign_1_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_5_internal_tensor_assign_1_begin_mask_0 = const()[name = tensor("new_cache_5_internal_tensor_assign_1_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_5_internal_tensor_assign_1_end_mask_0 = const()[name = tensor("new_cache_5_internal_tensor_assign_1_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_5_internal_tensor_assign_1_squeeze_mask_0 = const()[name = tensor("new_cache_5_internal_tensor_assign_1_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_52 = const()[name = tensor("shape_52"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_4 = const()[name = tensor("reduce_prod_4"), val = tensor(1048576)]; + tensor range_1d_4_start_0 = const()[name = tensor("range_1d_4_start_0"), val = tensor(0)]; + tensor range_1d_4_step_0 = const()[name = tensor("range_1d_4_step_0"), val = tensor(1)]; + tensor range_1d_4 = range_1d(end = reduce_prod_4, start = range_1d_4_start_0, step = range_1d_4_step_0)[name = tensor("range_1d_4")]; + tensor reshape_20 = reshape(shape = shape_52, x = range_1d_4)[name = tensor("reshape_20")]; + tensor slice_by_index_4 = slice_by_index(begin = concat_16, begin_mask = new_cache_5_internal_tensor_assign_1_begin_mask_0, end = concat_17, end_mask = new_cache_5_internal_tensor_assign_1_end_mask_0, squeeze_mask = new_cache_5_internal_tensor_assign_1_squeeze_mask_0, stride = new_cache_5_internal_tensor_assign_1_stride_0, x = reshape_20)[name = tensor("slice_by_index_4")]; + tensor reshape_21_shape_0 = const()[name = tensor("reshape_21_shape_0"), val = tensor([-1])]; + tensor reshape_21 = reshape(shape = reshape_21_shape_0, x = slice_by_index_4)[name = tensor("reshape_21")]; + tensor reshape_22_shape_0 = const()[name = tensor("reshape_22_shape_0"), val = tensor([-1])]; + tensor reshape_22_cast_fp16 = reshape(shape = reshape_22_shape_0, x = var_1207_cast_fp16)[name = tensor("reshape_22_cast_fp16")]; + tensor reshape_23_shape_0 = const()[name = tensor("reshape_23_shape_0"), val = tensor([-1])]; + tensor reshape_23_cast_fp16 = reshape(shape = reshape_23_shape_0, x = cache2_to_fp16)[name = tensor("reshape_23_cast_fp16")]; + tensor scatter_4_mode_0 = const()[name = tensor("scatter_4_mode_0"), val = tensor("update")]; + tensor scatter_4_axis_0 = const()[name = tensor("scatter_4_axis_0"), val = tensor(0)]; + tensor scatter_4_validate_indices_0 = const()[name = tensor("scatter_4_validate_indices_0"), val = tensor(false)]; + tensor scatter_4_cast_fp16 = scatter(axis = scatter_4_axis_0, data = reshape_23_cast_fp16, indices = reshape_21, mode = scatter_4_mode_0, updates = reshape_22_cast_fp16, validate_indices = scatter_4_validate_indices_0)[name = tensor("scatter_4_cast_fp16")]; + tensor reshape_24_cast_fp16 = reshape(shape = shape_52, x = scatter_4_cast_fp16)[name = tensor("reshape_24_cast_fp16")]; + tensor var_1215_begin_0 = const()[name = tensor("op_1215_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor var_1215_end_0 = const()[name = tensor("op_1215_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor var_1215_end_mask_0 = const()[name = tensor("op_1215_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_1215_squeeze_mask_0 = const()[name = tensor("op_1215_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor var_1215_cast_fp16 = slice_by_index(begin = var_1215_begin_0, end = var_1215_end_0, end_mask = var_1215_end_mask_0, squeeze_mask = var_1215_squeeze_mask_0, x = reshape_24_cast_fp16)[name = tensor("op_1215_cast_fp16")]; + tensor var_1217_axis_0 = const()[name = tensor("op_1217_axis_0"), val = tensor(1)]; + tensor var_1217_mode_0 = const()[name = tensor("op_1217_mode_0"), val = tensor("update")]; + tensor var_1217_validate_indices_0 = const()[name = tensor("op_1217_validate_indices_0"), val = tensor(false)]; + tensor var_1217_cast_fp16 = scatter_along_axis(axis = var_1217_axis_0, data = var_1215_cast_fp16, indices = write_indices_5, mode = var_1217_mode_0, updates = v_5_cast_fp16, validate_indices = var_1217_validate_indices_0)[name = tensor("op_1217_cast_fp16")]; + tensor concat_18 = const()[name = tensor("concat_18"), val = tensor([1, 0, 0, 0, 0])]; + tensor concat_19 = const()[name = tensor("concat_19"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_5_internal_tensor_assign_2_stride_0 = const()[name = tensor("new_cache_5_internal_tensor_assign_2_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_5_internal_tensor_assign_2_begin_mask_0 = const()[name = tensor("new_cache_5_internal_tensor_assign_2_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_5_internal_tensor_assign_2_end_mask_0 = const()[name = tensor("new_cache_5_internal_tensor_assign_2_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_5_internal_tensor_assign_2_squeeze_mask_0 = const()[name = tensor("new_cache_5_internal_tensor_assign_2_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_53 = const()[name = tensor("shape_53"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_5 = const()[name = tensor("reduce_prod_5"), val = tensor(1048576)]; + tensor range_1d_5_start_0 = const()[name = tensor("range_1d_5_start_0"), val = tensor(0)]; + tensor range_1d_5_step_0 = const()[name = tensor("range_1d_5_step_0"), val = tensor(1)]; + tensor range_1d_5 = range_1d(end = reduce_prod_5, start = range_1d_5_start_0, step = range_1d_5_step_0)[name = tensor("range_1d_5")]; + tensor reshape_25 = reshape(shape = shape_53, x = range_1d_5)[name = tensor("reshape_25")]; + tensor slice_by_index_5 = slice_by_index(begin = concat_18, begin_mask = new_cache_5_internal_tensor_assign_2_begin_mask_0, end = concat_19, end_mask = new_cache_5_internal_tensor_assign_2_end_mask_0, squeeze_mask = new_cache_5_internal_tensor_assign_2_squeeze_mask_0, stride = new_cache_5_internal_tensor_assign_2_stride_0, x = reshape_25)[name = tensor("slice_by_index_5")]; + tensor reshape_26_shape_0 = const()[name = tensor("reshape_26_shape_0"), val = tensor([-1])]; + tensor reshape_26 = reshape(shape = reshape_26_shape_0, x = slice_by_index_5)[name = tensor("reshape_26")]; + tensor reshape_27_shape_0 = const()[name = tensor("reshape_27_shape_0"), val = tensor([-1])]; + tensor reshape_27_cast_fp16 = reshape(shape = reshape_27_shape_0, x = var_1217_cast_fp16)[name = tensor("reshape_27_cast_fp16")]; + tensor reshape_28_shape_0 = const()[name = tensor("reshape_28_shape_0"), val = tensor([-1])]; + tensor reshape_28_cast_fp16 = reshape(shape = reshape_28_shape_0, x = reshape_24_cast_fp16)[name = tensor("reshape_28_cast_fp16")]; + tensor scatter_5_mode_0 = const()[name = tensor("scatter_5_mode_0"), val = tensor("update")]; + tensor scatter_5_axis_0 = const()[name = tensor("scatter_5_axis_0"), val = tensor(0)]; + tensor scatter_5_validate_indices_0 = const()[name = tensor("scatter_5_validate_indices_0"), val = tensor(false)]; + tensor scatter_5_cast_fp16 = scatter(axis = scatter_5_axis_0, data = reshape_28_cast_fp16, indices = reshape_26, mode = scatter_5_mode_0, updates = reshape_27_cast_fp16, validate_indices = scatter_5_validate_indices_0)[name = tensor("scatter_5_cast_fp16")]; + tensor reshape_29_cast_fp16 = reshape(shape = shape_53, x = scatter_5_cast_fp16)[name = tensor("reshape_29_cast_fp16")]; + tensor reshape_29_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("reshape_29_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor keys_13_begin_0 = const()[name = tensor("keys_13_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor keys_13_end_0 = const()[name = tensor("keys_13_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor keys_13_end_mask_0 = const()[name = tensor("keys_13_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor keys_13_squeeze_mask_0 = const()[name = tensor("keys_13_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor keys_13_cast_fp16 = slice_by_index(begin = keys_13_begin_0, end = keys_13_end_0, end_mask = keys_13_end_mask_0, squeeze_mask = keys_13_squeeze_mask_0, x = reshape_29_cast_fp16)[name = tensor("keys_13_cast_fp16")]; + tensor values_13_begin_0 = const()[name = tensor("values_13_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor values_13_end_0 = const()[name = tensor("values_13_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor values_13_end_mask_0 = const()[name = tensor("values_13_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor values_13_squeeze_mask_0 = const()[name = tensor("values_13_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor values_13_cast_fp16 = slice_by_index(begin = values_13_begin_0, end = values_13_end_0, end_mask = values_13_end_mask_0, squeeze_mask = values_13_squeeze_mask_0, x = reshape_29_cast_fp16)[name = tensor("values_13_cast_fp16")]; + tensor var_1229_cast_fp16 = not_equal(x = keys_13_cast_fp16, y = keys_13_cast_fp16)[name = tensor("op_1229_cast_fp16")]; + tensor keys_15_cast_fp16 = select(a = var_463_to_fp16, b = keys_13_cast_fp16, cond = var_1229_cast_fp16)[name = tensor("keys_15_cast_fp16")]; + tensor var_1237_cast_fp16 = not_equal(x = values_13_cast_fp16, y = values_13_cast_fp16)[name = tensor("op_1237_cast_fp16")]; + tensor values_15_cast_fp16 = select(a = var_463_to_fp16, b = values_13_cast_fp16, cond = var_1237_cast_fp16)[name = tensor("values_15_cast_fp16")]; + tensor var_1261 = const()[name = tensor("op_1261"), val = tensor([0, 2, 1, 3])]; + tensor var_1277 = const()[name = tensor("op_1277"), val = tensor([1, 1, 1])]; + tensor var_1278_cast_fp16 = reshape(shape = var_1277, x = position2_to_fp16)[name = tensor("op_1278_cast_fp16")]; + tensor var_1298_to_fp16 = const()[name = tensor("op_1298_to_fp16"), val = tensor(0x1p+0)]; + tensor valid_len_5_cast_fp16 = add(x = var_1278_cast_fp16, y = var_1298_to_fp16)[name = tensor("valid_len_5_cast_fp16")]; + tensor valid_mask_5_cast_fp16 = less(x = k_positions_1_promoted_to_fp16, y = valid_len_5_cast_fp16)[name = tensor("valid_mask_5_cast_fp16")]; + tensor causal_mask_5_cast_fp16 = less_equal(x = k_positions_1_promoted_to_fp16, y = var_1278_cast_fp16)[name = tensor("causal_mask_5_cast_fp16")]; + tensor attn_mask_9 = logical_and(x = valid_mask_5_cast_fp16, y = causal_mask_5_cast_fp16)[name = tensor("attn_mask_9")]; + tensor attn_mask_11_axes_0 = const()[name = tensor("attn_mask_11_axes_0"), val = tensor([1])]; + tensor attn_mask_11 = expand_dims(axes = attn_mask_11_axes_0, x = attn_mask_9)[name = tensor("attn_mask_11")]; + tensor var_1316_transpose_x_0 = const()[name = tensor("op_1316_transpose_x_0"), val = tensor(false)]; + tensor var_1316_transpose_y_0 = const()[name = tensor("op_1316_transpose_y_0"), val = tensor(false)]; + tensor transpose_76_perm_0 = const()[name = tensor("transpose_76_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_77_perm_0 = const()[name = tensor("transpose_77_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_77 = transpose(perm = transpose_77_perm_0, x = keys_15_cast_fp16)[name = tensor("transpose_205")]; + tensor transpose_76 = transpose(perm = transpose_76_perm_0, x = q_15_cast_fp16)[name = tensor("transpose_206")]; + tensor var_1316_cast_fp16 = matmul(transpose_x = var_1316_transpose_x_0, transpose_y = var_1316_transpose_y_0, x = transpose_76, y = transpose_77)[name = tensor("op_1316_cast_fp16")]; + tensor var_1310_to_fp16 = const()[name = tensor("op_1310_to_fp16"), val = tensor([0x1p-3])]; + tensor attn_weights_13_cast_fp16 = mul(x = var_1316_cast_fp16, y = var_1310_to_fp16)[name = tensor("attn_weights_13_cast_fp16")]; + tensor var_1318 = logical_not(x = attn_mask_11)[name = tensor("op_1318")]; + tensor var_1319_to_fp16 = const()[name = tensor("op_1319_to_fp16"), val = tensor(-inf)]; + tensor attn_weights_15_cast_fp16 = select(a = var_1319_to_fp16, b = attn_weights_13_cast_fp16, cond = var_1318)[name = tensor("attn_weights_15_cast_fp16")]; + tensor var_1321 = const()[name = tensor("op_1321"), val = tensor(-1)]; + tensor attn_weights_17_cast_fp16 = softmax(axis = var_1321, x = attn_weights_15_cast_fp16)[name = tensor("attn_weights_17_cast_fp16")]; + tensor attn_output_5_transpose_x_0 = const()[name = tensor("attn_output_5_transpose_x_0"), val = tensor(false)]; + tensor attn_output_5_transpose_y_0 = const()[name = tensor("attn_output_5_transpose_y_0"), val = tensor(false)]; + tensor values_17_cast_fp16 = transpose(perm = var_1261, x = values_15_cast_fp16)[name = tensor("transpose_207")]; + tensor attn_output_5_cast_fp16 = matmul(transpose_x = attn_output_5_transpose_x_0, transpose_y = attn_output_5_transpose_y_0, x = attn_weights_17_cast_fp16, y = values_17_cast_fp16)[name = tensor("attn_output_5_cast_fp16")]; + tensor var_1329 = const()[name = tensor("op_1329"), val = tensor([0, 2, 1, 3])]; + tensor var_1334 = const()[name = tensor("op_1334"), val = tensor([1, 1, 1024])]; + tensor var_1330_cast_fp16 = transpose(perm = var_1329, x = attn_output_5_cast_fp16)[name = tensor("transpose_204")]; + tensor input_31_cast_fp16 = reshape(shape = var_1334, x = var_1330_cast_fp16)[name = tensor("input_31_cast_fp16")]; + tensor attn2_out_proj_weight_to_fp16 = const()[name = tensor("attn2_out_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(57777152)))]; + tensor linear_10_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = attn2_out_proj_weight_to_fp16, x = input_31_cast_fp16)[name = tensor("linear_10_cast_fp16")]; + tensor var_1340_to_fp16 = const()[name = tensor("op_1340_to_fp16"), val = tensor(0x1p+0)]; + tensor var_1341_cast_fp16 = add(x = position2_to_fp16, y = var_1340_to_fp16)[name = tensor("op_1341_cast_fp16")]; + tensor var_1341_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_1341_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor input_33_cast_fp16 = add(x = input_27_cast_fp16, y = linear_10_cast_fp16)[name = tensor("input_33_cast_fp16")]; + tensor input_35_axes_0 = const()[name = tensor("input_35_axes_0"), val = tensor([-1])]; + tensor norm2_2_weight_to_fp16 = const()[name = tensor("norm2_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(59874368)))]; + tensor norm2_2_bias_to_fp16 = const()[name = tensor("norm2_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(59876480)))]; + tensor var_1345_to_fp16 = const()[name = tensor("op_1345_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_35_cast_fp16 = layer_norm(axes = input_35_axes_0, beta = norm2_2_bias_to_fp16, epsilon = var_1345_to_fp16, gamma = norm2_2_weight_to_fp16, x = input_33_cast_fp16)[name = tensor("input_35_cast_fp16")]; + tensor linear2_1_weight_to_fp16 = const()[name = tensor("linear2_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(59878592)))]; + tensor linear_11_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = linear2_1_weight_to_fp16, x = input_35_cast_fp16)[name = tensor("linear_11_cast_fp16")]; + tensor input_37_mode_0 = const()[name = tensor("input_37_mode_0"), val = tensor("EXACT")]; + tensor input_37_cast_fp16 = gelu(mode = input_37_mode_0, x = linear_11_cast_fp16)[name = tensor("input_37_cast_fp16")]; + tensor linear2_2_weight_to_fp16 = const()[name = tensor("linear2_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(68267264)))]; + tensor linear_12_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = linear2_2_weight_to_fp16, x = input_37_cast_fp16)[name = tensor("linear_12_cast_fp16")]; + tensor input_39_cast_fp16 = add(x = input_33_cast_fp16, y = linear_12_cast_fp16)[name = tensor("input_39_cast_fp16")]; + tensor input_41_axes_0 = const()[name = tensor("input_41_axes_0"), val = tensor([-1])]; + tensor norm3_1_weight_to_fp16 = const()[name = tensor("norm3_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(76655936)))]; + tensor norm3_1_bias_to_fp16 = const()[name = tensor("norm3_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(76658048)))]; + tensor var_1362_to_fp16 = const()[name = tensor("op_1362_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_41_cast_fp16 = layer_norm(axes = input_41_axes_0, beta = norm3_1_bias_to_fp16, epsilon = var_1362_to_fp16, gamma = norm3_1_weight_to_fp16, x = input_39_cast_fp16)[name = tensor("input_41_cast_fp16")]; + tensor attn3_in_proj_weight_to_fp16 = const()[name = tensor("attn3_in_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(76660160)))]; + tensor linear_13_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = attn3_in_proj_weight_to_fp16, x = input_41_cast_fp16)[name = tensor("linear_13_cast_fp16")]; + tensor var_1376 = const()[name = tensor("op_1376"), val = tensor([1, 1, 3, 16, 64])]; + tensor qkv_7_cast_fp16 = reshape(shape = var_1376, x = linear_13_cast_fp16)[name = tensor("qkv_7_cast_fp16")]; + tensor q_19_begin_0 = const()[name = tensor("q_19_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor q_19_end_0 = const()[name = tensor("q_19_end_0"), val = tensor([1, 1, 1, 16, 64])]; + tensor q_19_end_mask_0 = const()[name = tensor("q_19_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor q_19_squeeze_mask_0 = const()[name = tensor("q_19_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor q_19_cast_fp16 = slice_by_index(begin = q_19_begin_0, end = q_19_end_0, end_mask = q_19_end_mask_0, squeeze_mask = q_19_squeeze_mask_0, x = qkv_7_cast_fp16)[name = tensor("q_19_cast_fp16")]; + tensor k_13_begin_0 = const()[name = tensor("k_13_begin_0"), val = tensor([0, 0, 1, 0, 0])]; + tensor k_13_end_0 = const()[name = tensor("k_13_end_0"), val = tensor([1, 1, 2, 16, 64])]; + tensor k_13_end_mask_0 = const()[name = tensor("k_13_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor k_13_squeeze_mask_0 = const()[name = tensor("k_13_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor k_13_cast_fp16 = slice_by_index(begin = k_13_begin_0, end = k_13_end_0, end_mask = k_13_end_mask_0, squeeze_mask = k_13_squeeze_mask_0, x = qkv_7_cast_fp16)[name = tensor("k_13_cast_fp16")]; + tensor v_7_begin_0 = const()[name = tensor("v_7_begin_0"), val = tensor([0, 0, 2, 0, 0])]; + tensor v_7_end_0 = const()[name = tensor("v_7_end_0"), val = tensor([1, 1, 3, 16, 64])]; + tensor v_7_end_mask_0 = const()[name = tensor("v_7_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor v_7_squeeze_mask_0 = const()[name = tensor("v_7_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor v_7_cast_fp16 = slice_by_index(begin = v_7_begin_0, end = v_7_end_0, end_mask = v_7_end_mask_0, squeeze_mask = v_7_squeeze_mask_0, x = qkv_7_cast_fp16)[name = tensor("v_7_cast_fp16")]; + tensor var_1442 = const()[name = tensor("op_1442"), val = tensor([1, 1, 1, 1])]; + tensor position3_to_fp16_dtype_0 = const()[name = tensor("position3_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor position3_to_fp16 = cast(dtype = position3_to_fp16_dtype_0, x = position3)[name = tensor("cast_538")]; + tensor ts_23_cast_fp16 = reshape(shape = var_1442, x = position3_to_fp16)[name = tensor("ts_23_cast_fp16")]; + tensor var_1449 = const()[name = tensor("op_1449"), val = tensor([1, 1, 16, 32, 2])]; + tensor q_complex_7_cast_fp16 = reshape(shape = var_1449, x = q_19_cast_fp16)[name = tensor("q_complex_7_cast_fp16")]; + tensor var_1456 = const()[name = tensor("op_1456"), val = tensor([1, 1, 16, 32, 2])]; + tensor k_complex_7_cast_fp16 = reshape(shape = var_1456, x = k_13_cast_fp16)[name = tensor("k_complex_7_cast_fp16")]; + tensor var_1460_begin_0 = const()[name = tensor("op_1460_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_1460_end_0 = const()[name = tensor("op_1460_end_0"), val = tensor([1, 1, 16, 32, 1])]; + tensor var_1460_end_mask_0 = const()[name = tensor("op_1460_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_1460_squeeze_mask_0 = const()[name = tensor("op_1460_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_1460_cast_fp16 = slice_by_index(begin = var_1460_begin_0, end = var_1460_end_0, end_mask = var_1460_end_mask_0, squeeze_mask = var_1460_squeeze_mask_0, x = q_complex_7_cast_fp16)[name = tensor("op_1460_cast_fp16")]; + tensor var_1468_begin_0 = const()[name = tensor("op_1468_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_1468_end_0 = const()[name = tensor("op_1468_end_0"), val = tensor([1, 1, 16, 32, 2])]; + tensor var_1468_end_mask_0 = const()[name = tensor("op_1468_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_1468_squeeze_mask_0 = const()[name = tensor("op_1468_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_1468_cast_fp16 = slice_by_index(begin = var_1468_begin_0, end = var_1468_end_0, end_mask = var_1468_end_mask_0, squeeze_mask = var_1468_squeeze_mask_0, x = q_complex_7_cast_fp16)[name = tensor("op_1468_cast_fp16")]; + tensor var_1476_begin_0 = const()[name = tensor("op_1476_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_1476_end_0 = const()[name = tensor("op_1476_end_0"), val = tensor([1, 1, 16, 32, 1])]; + tensor var_1476_end_mask_0 = const()[name = tensor("op_1476_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_1476_squeeze_mask_0 = const()[name = tensor("op_1476_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_1476_cast_fp16 = slice_by_index(begin = var_1476_begin_0, end = var_1476_end_0, end_mask = var_1476_end_mask_0, squeeze_mask = var_1476_squeeze_mask_0, x = k_complex_7_cast_fp16)[name = tensor("op_1476_cast_fp16")]; + tensor var_1484_begin_0 = const()[name = tensor("op_1484_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_1484_end_0 = const()[name = tensor("op_1484_end_0"), val = tensor([1, 1, 16, 32, 2])]; + tensor var_1484_end_mask_0 = const()[name = tensor("op_1484_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_1484_squeeze_mask_0 = const()[name = tensor("op_1484_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_1484_cast_fp16 = slice_by_index(begin = var_1484_begin_0, end = var_1484_end_0, end_mask = var_1484_end_mask_0, squeeze_mask = var_1484_squeeze_mask_0, x = k_complex_7_cast_fp16)[name = tensor("op_1484_cast_fp16")]; + tensor freqs_7_to_fp16 = const()[name = tensor("freqs_7_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(82951680)))]; + tensor var_1490_cast_fp16 = mul(x = freqs_7_to_fp16, y = ts_23_cast_fp16)[name = tensor("op_1490_cast_fp16")]; + tensor rotr_7_cast_fp16 = cos(x = var_1490_cast_fp16)[name = tensor("rotr_7_cast_fp16")]; + tensor roti_7_cast_fp16 = sin(x = var_1490_cast_fp16)[name = tensor("roti_7_cast_fp16")]; + tensor var_1494_cast_fp16 = mul(x = var_1460_cast_fp16, y = rotr_7_cast_fp16)[name = tensor("op_1494_cast_fp16")]; + tensor var_1495_cast_fp16 = mul(x = var_1468_cast_fp16, y = roti_7_cast_fp16)[name = tensor("op_1495_cast_fp16")]; + tensor qor_13_cast_fp16 = sub(x = var_1494_cast_fp16, y = var_1495_cast_fp16)[name = tensor("qor_13_cast_fp16")]; + tensor var_1498_cast_fp16 = mul(x = var_1460_cast_fp16, y = roti_7_cast_fp16)[name = tensor("op_1498_cast_fp16")]; + tensor var_1499_cast_fp16 = mul(x = var_1468_cast_fp16, y = rotr_7_cast_fp16)[name = tensor("op_1499_cast_fp16")]; + tensor qoi_13_cast_fp16 = add(x = var_1498_cast_fp16, y = var_1499_cast_fp16)[name = tensor("qoi_13_cast_fp16")]; + tensor var_1502_cast_fp16 = mul(x = var_1476_cast_fp16, y = rotr_7_cast_fp16)[name = tensor("op_1502_cast_fp16")]; + tensor var_1503_cast_fp16 = mul(x = var_1484_cast_fp16, y = roti_7_cast_fp16)[name = tensor("op_1503_cast_fp16")]; + tensor kor_13_cast_fp16 = sub(x = var_1502_cast_fp16, y = var_1503_cast_fp16)[name = tensor("kor_13_cast_fp16")]; + tensor var_1506_cast_fp16 = mul(x = var_1476_cast_fp16, y = roti_7_cast_fp16)[name = tensor("op_1506_cast_fp16")]; + tensor var_1507_cast_fp16 = mul(x = var_1484_cast_fp16, y = rotr_7_cast_fp16)[name = tensor("op_1507_cast_fp16")]; + tensor koi_13_cast_fp16 = add(x = var_1506_cast_fp16, y = var_1507_cast_fp16)[name = tensor("koi_13_cast_fp16")]; + tensor qo_7_axis_0 = const()[name = tensor("qo_7_axis_0"), val = tensor(-1)]; + tensor qo_7_cast_fp16 = stack(axis = qo_7_axis_0, values = (qor_13_cast_fp16, qoi_13_cast_fp16))[name = tensor("qo_7_cast_fp16")]; + tensor ko_7_axis_0 = const()[name = tensor("ko_7_axis_0"), val = tensor(-1)]; + tensor ko_7_cast_fp16 = stack(axis = ko_7_axis_0, values = (kor_13_cast_fp16, koi_13_cast_fp16))[name = tensor("ko_7_cast_fp16")]; + tensor var_1540 = const()[name = tensor("op_1540"), val = tensor([1, 1, 16, 64])]; + tensor q_21_cast_fp16 = reshape(shape = var_1540, x = qo_7_cast_fp16)[name = tensor("q_21_cast_fp16")]; + tensor var_1546 = const()[name = tensor("op_1546"), val = tensor([1, 1, 16, 64])]; + tensor k_15_cast_fp16 = reshape(shape = var_1546, x = ko_7_cast_fp16)[name = tensor("k_15_cast_fp16")]; + tensor _inversed_1571_y_0_to_fp16 = const()[name = tensor("_inversed_1571_y_0_to_fp16"), val = tensor(0x1p-9)]; + tensor _inversed_1571_cast_fp16 = mul(x = ts_23_cast_fp16, y = _inversed_1571_y_0_to_fp16)[name = tensor("_inversed_1571_cast_fp16")]; + tensor var_1572_cast_fp16 = floor(x = _inversed_1571_cast_fp16)[name = tensor("op_1572_cast_fp16")]; + tensor var_1573_to_fp16 = const()[name = tensor("op_1573_to_fp16"), val = tensor(0x1p+9)]; + tensor var_1574_cast_fp16 = mul(x = var_1572_cast_fp16, y = var_1573_to_fp16)[name = tensor("op_1574_cast_fp16")]; + tensor write_indices_float_15_cast_fp16 = sub(x = ts_23_cast_fp16, y = var_1574_cast_fp16)[name = tensor("write_indices_float_15_cast_fp16")]; + tensor var_1581_dtype_0 = const()[name = tensor("op_1581_dtype_0"), val = tensor("int32")]; + tensor write_indices_7_reps_0 = const()[name = tensor("write_indices_7_reps_0"), val = tensor([1, 1, 16, 64])]; + tensor write_indices_float_15_cast_fp16_to_int32 = cast(dtype = var_1581_dtype_0, x = write_indices_float_15_cast_fp16)[name = tensor("cast_537")]; + tensor write_indices_7 = tile(reps = write_indices_7_reps_0, x = write_indices_float_15_cast_fp16_to_int32)[name = tensor("write_indices_7")]; + tensor var_1591_begin_0 = const()[name = tensor("op_1591_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_1591_end_0 = const()[name = tensor("op_1591_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor var_1591_end_mask_0 = const()[name = tensor("op_1591_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_1591_squeeze_mask_0 = const()[name = tensor("op_1591_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor cache3_to_fp16_dtype_0 = const()[name = tensor("cache3_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor cache3_to_fp16 = cast(dtype = cache3_to_fp16_dtype_0, x = cache3)[name = tensor("cast_536")]; + tensor var_1591_cast_fp16 = slice_by_index(begin = var_1591_begin_0, end = var_1591_end_0, end_mask = var_1591_end_mask_0, squeeze_mask = var_1591_squeeze_mask_0, x = cache3_to_fp16)[name = tensor("op_1591_cast_fp16")]; + tensor var_1593_axis_0 = const()[name = tensor("op_1593_axis_0"), val = tensor(1)]; + tensor var_1593_mode_0 = const()[name = tensor("op_1593_mode_0"), val = tensor("update")]; + tensor var_1593_validate_indices_0 = const()[name = tensor("op_1593_validate_indices_0"), val = tensor(false)]; + tensor var_1593_cast_fp16 = scatter_along_axis(axis = var_1593_axis_0, data = var_1591_cast_fp16, indices = write_indices_7, mode = var_1593_mode_0, updates = k_15_cast_fp16, validate_indices = var_1593_validate_indices_0)[name = tensor("op_1593_cast_fp16")]; + tensor concat_23 = const()[name = tensor("concat_23"), val = tensor([0, 0, 0, 0, 0])]; + tensor concat_24 = const()[name = tensor("concat_24"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_7_internal_tensor_assign_1_stride_0 = const()[name = tensor("new_cache_7_internal_tensor_assign_1_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_7_internal_tensor_assign_1_begin_mask_0 = const()[name = tensor("new_cache_7_internal_tensor_assign_1_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_7_internal_tensor_assign_1_end_mask_0 = const()[name = tensor("new_cache_7_internal_tensor_assign_1_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_7_internal_tensor_assign_1_squeeze_mask_0 = const()[name = tensor("new_cache_7_internal_tensor_assign_1_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_54 = const()[name = tensor("shape_54"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_6 = const()[name = tensor("reduce_prod_6"), val = tensor(1048576)]; + tensor range_1d_6_start_0 = const()[name = tensor("range_1d_6_start_0"), val = tensor(0)]; + tensor range_1d_6_step_0 = const()[name = tensor("range_1d_6_step_0"), val = tensor(1)]; + tensor range_1d_6 = range_1d(end = reduce_prod_6, start = range_1d_6_start_0, step = range_1d_6_step_0)[name = tensor("range_1d_6")]; + tensor reshape_30 = reshape(shape = shape_54, x = range_1d_6)[name = tensor("reshape_30")]; + tensor slice_by_index_6 = slice_by_index(begin = concat_23, begin_mask = new_cache_7_internal_tensor_assign_1_begin_mask_0, end = concat_24, end_mask = new_cache_7_internal_tensor_assign_1_end_mask_0, squeeze_mask = new_cache_7_internal_tensor_assign_1_squeeze_mask_0, stride = new_cache_7_internal_tensor_assign_1_stride_0, x = reshape_30)[name = tensor("slice_by_index_6")]; + tensor reshape_31_shape_0 = const()[name = tensor("reshape_31_shape_0"), val = tensor([-1])]; + tensor reshape_31 = reshape(shape = reshape_31_shape_0, x = slice_by_index_6)[name = tensor("reshape_31")]; + tensor reshape_32_shape_0 = const()[name = tensor("reshape_32_shape_0"), val = tensor([-1])]; + tensor reshape_32_cast_fp16 = reshape(shape = reshape_32_shape_0, x = var_1593_cast_fp16)[name = tensor("reshape_32_cast_fp16")]; + tensor reshape_33_shape_0 = const()[name = tensor("reshape_33_shape_0"), val = tensor([-1])]; + tensor reshape_33_cast_fp16 = reshape(shape = reshape_33_shape_0, x = cache3_to_fp16)[name = tensor("reshape_33_cast_fp16")]; + tensor scatter_6_mode_0 = const()[name = tensor("scatter_6_mode_0"), val = tensor("update")]; + tensor scatter_6_axis_0 = const()[name = tensor("scatter_6_axis_0"), val = tensor(0)]; + tensor scatter_6_validate_indices_0 = const()[name = tensor("scatter_6_validate_indices_0"), val = tensor(false)]; + tensor scatter_6_cast_fp16 = scatter(axis = scatter_6_axis_0, data = reshape_33_cast_fp16, indices = reshape_31, mode = scatter_6_mode_0, updates = reshape_32_cast_fp16, validate_indices = scatter_6_validate_indices_0)[name = tensor("scatter_6_cast_fp16")]; + tensor reshape_34_cast_fp16 = reshape(shape = shape_54, x = scatter_6_cast_fp16)[name = tensor("reshape_34_cast_fp16")]; + tensor var_1601_begin_0 = const()[name = tensor("op_1601_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor var_1601_end_0 = const()[name = tensor("op_1601_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor var_1601_end_mask_0 = const()[name = tensor("op_1601_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_1601_squeeze_mask_0 = const()[name = tensor("op_1601_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor var_1601_cast_fp16 = slice_by_index(begin = var_1601_begin_0, end = var_1601_end_0, end_mask = var_1601_end_mask_0, squeeze_mask = var_1601_squeeze_mask_0, x = reshape_34_cast_fp16)[name = tensor("op_1601_cast_fp16")]; + tensor var_1603_axis_0 = const()[name = tensor("op_1603_axis_0"), val = tensor(1)]; + tensor var_1603_mode_0 = const()[name = tensor("op_1603_mode_0"), val = tensor("update")]; + tensor var_1603_validate_indices_0 = const()[name = tensor("op_1603_validate_indices_0"), val = tensor(false)]; + tensor var_1603_cast_fp16 = scatter_along_axis(axis = var_1603_axis_0, data = var_1601_cast_fp16, indices = write_indices_7, mode = var_1603_mode_0, updates = v_7_cast_fp16, validate_indices = var_1603_validate_indices_0)[name = tensor("op_1603_cast_fp16")]; + tensor concat_25 = const()[name = tensor("concat_25"), val = tensor([1, 0, 0, 0, 0])]; + tensor concat_26 = const()[name = tensor("concat_26"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_7_internal_tensor_assign_2_stride_0 = const()[name = tensor("new_cache_7_internal_tensor_assign_2_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_7_internal_tensor_assign_2_begin_mask_0 = const()[name = tensor("new_cache_7_internal_tensor_assign_2_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_7_internal_tensor_assign_2_end_mask_0 = const()[name = tensor("new_cache_7_internal_tensor_assign_2_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_7_internal_tensor_assign_2_squeeze_mask_0 = const()[name = tensor("new_cache_7_internal_tensor_assign_2_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_55 = const()[name = tensor("shape_55"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_7 = const()[name = tensor("reduce_prod_7"), val = tensor(1048576)]; + tensor range_1d_7_start_0 = const()[name = tensor("range_1d_7_start_0"), val = tensor(0)]; + tensor range_1d_7_step_0 = const()[name = tensor("range_1d_7_step_0"), val = tensor(1)]; + tensor range_1d_7 = range_1d(end = reduce_prod_7, start = range_1d_7_start_0, step = range_1d_7_step_0)[name = tensor("range_1d_7")]; + tensor reshape_35 = reshape(shape = shape_55, x = range_1d_7)[name = tensor("reshape_35")]; + tensor slice_by_index_7 = slice_by_index(begin = concat_25, begin_mask = new_cache_7_internal_tensor_assign_2_begin_mask_0, end = concat_26, end_mask = new_cache_7_internal_tensor_assign_2_end_mask_0, squeeze_mask = new_cache_7_internal_tensor_assign_2_squeeze_mask_0, stride = new_cache_7_internal_tensor_assign_2_stride_0, x = reshape_35)[name = tensor("slice_by_index_7")]; + tensor reshape_36_shape_0 = const()[name = tensor("reshape_36_shape_0"), val = tensor([-1])]; + tensor reshape_36 = reshape(shape = reshape_36_shape_0, x = slice_by_index_7)[name = tensor("reshape_36")]; + tensor reshape_37_shape_0 = const()[name = tensor("reshape_37_shape_0"), val = tensor([-1])]; + tensor reshape_37_cast_fp16 = reshape(shape = reshape_37_shape_0, x = var_1603_cast_fp16)[name = tensor("reshape_37_cast_fp16")]; + tensor reshape_38_shape_0 = const()[name = tensor("reshape_38_shape_0"), val = tensor([-1])]; + tensor reshape_38_cast_fp16 = reshape(shape = reshape_38_shape_0, x = reshape_34_cast_fp16)[name = tensor("reshape_38_cast_fp16")]; + tensor scatter_7_mode_0 = const()[name = tensor("scatter_7_mode_0"), val = tensor("update")]; + tensor scatter_7_axis_0 = const()[name = tensor("scatter_7_axis_0"), val = tensor(0)]; + tensor scatter_7_validate_indices_0 = const()[name = tensor("scatter_7_validate_indices_0"), val = tensor(false)]; + tensor scatter_7_cast_fp16 = scatter(axis = scatter_7_axis_0, data = reshape_38_cast_fp16, indices = reshape_36, mode = scatter_7_mode_0, updates = reshape_37_cast_fp16, validate_indices = scatter_7_validate_indices_0)[name = tensor("scatter_7_cast_fp16")]; + tensor reshape_39_cast_fp16 = reshape(shape = shape_55, x = scatter_7_cast_fp16)[name = tensor("reshape_39_cast_fp16")]; + tensor reshape_39_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("reshape_39_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor keys_19_begin_0 = const()[name = tensor("keys_19_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor keys_19_end_0 = const()[name = tensor("keys_19_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor keys_19_end_mask_0 = const()[name = tensor("keys_19_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor keys_19_squeeze_mask_0 = const()[name = tensor("keys_19_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor keys_19_cast_fp16 = slice_by_index(begin = keys_19_begin_0, end = keys_19_end_0, end_mask = keys_19_end_mask_0, squeeze_mask = keys_19_squeeze_mask_0, x = reshape_39_cast_fp16)[name = tensor("keys_19_cast_fp16")]; + tensor values_19_begin_0 = const()[name = tensor("values_19_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor values_19_end_0 = const()[name = tensor("values_19_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor values_19_end_mask_0 = const()[name = tensor("values_19_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor values_19_squeeze_mask_0 = const()[name = tensor("values_19_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor values_19_cast_fp16 = slice_by_index(begin = values_19_begin_0, end = values_19_end_0, end_mask = values_19_end_mask_0, squeeze_mask = values_19_squeeze_mask_0, x = reshape_39_cast_fp16)[name = tensor("values_19_cast_fp16")]; + tensor var_1615_cast_fp16 = not_equal(x = keys_19_cast_fp16, y = keys_19_cast_fp16)[name = tensor("op_1615_cast_fp16")]; + tensor keys_21_cast_fp16 = select(a = var_463_to_fp16, b = keys_19_cast_fp16, cond = var_1615_cast_fp16)[name = tensor("keys_21_cast_fp16")]; + tensor var_1623_cast_fp16 = not_equal(x = values_19_cast_fp16, y = values_19_cast_fp16)[name = tensor("op_1623_cast_fp16")]; + tensor values_21_cast_fp16 = select(a = var_463_to_fp16, b = values_19_cast_fp16, cond = var_1623_cast_fp16)[name = tensor("values_21_cast_fp16")]; + tensor var_1647 = const()[name = tensor("op_1647"), val = tensor([0, 2, 1, 3])]; + tensor var_1663 = const()[name = tensor("op_1663"), val = tensor([1, 1, 1])]; + tensor var_1664_cast_fp16 = reshape(shape = var_1663, x = position3_to_fp16)[name = tensor("op_1664_cast_fp16")]; + tensor var_1684_to_fp16 = const()[name = tensor("op_1684_to_fp16"), val = tensor(0x1p+0)]; + tensor valid_len_7_cast_fp16 = add(x = var_1664_cast_fp16, y = var_1684_to_fp16)[name = tensor("valid_len_7_cast_fp16")]; + tensor valid_mask_7_cast_fp16 = less(x = k_positions_1_promoted_to_fp16, y = valid_len_7_cast_fp16)[name = tensor("valid_mask_7_cast_fp16")]; + tensor causal_mask_7_cast_fp16 = less_equal(x = k_positions_1_promoted_to_fp16, y = var_1664_cast_fp16)[name = tensor("causal_mask_7_cast_fp16")]; + tensor attn_mask_13 = logical_and(x = valid_mask_7_cast_fp16, y = causal_mask_7_cast_fp16)[name = tensor("attn_mask_13")]; + tensor attn_mask_15_axes_0 = const()[name = tensor("attn_mask_15_axes_0"), val = tensor([1])]; + tensor attn_mask_15 = expand_dims(axes = attn_mask_15_axes_0, x = attn_mask_13)[name = tensor("attn_mask_15")]; + tensor var_1702_transpose_x_0 = const()[name = tensor("op_1702_transpose_x_0"), val = tensor(false)]; + tensor var_1702_transpose_y_0 = const()[name = tensor("op_1702_transpose_y_0"), val = tensor(false)]; + tensor transpose_78_perm_0 = const()[name = tensor("transpose_78_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_79_perm_0 = const()[name = tensor("transpose_79_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_79 = transpose(perm = transpose_79_perm_0, x = keys_21_cast_fp16)[name = tensor("transpose_201")]; + tensor transpose_78 = transpose(perm = transpose_78_perm_0, x = q_21_cast_fp16)[name = tensor("transpose_202")]; + tensor var_1702_cast_fp16 = matmul(transpose_x = var_1702_transpose_x_0, transpose_y = var_1702_transpose_y_0, x = transpose_78, y = transpose_79)[name = tensor("op_1702_cast_fp16")]; + tensor var_1696_to_fp16 = const()[name = tensor("op_1696_to_fp16"), val = tensor([0x1p-3])]; + tensor attn_weights_19_cast_fp16 = mul(x = var_1702_cast_fp16, y = var_1696_to_fp16)[name = tensor("attn_weights_19_cast_fp16")]; + tensor var_1704 = logical_not(x = attn_mask_15)[name = tensor("op_1704")]; + tensor var_1705_to_fp16 = const()[name = tensor("op_1705_to_fp16"), val = tensor(-inf)]; + tensor attn_weights_21_cast_fp16 = select(a = var_1705_to_fp16, b = attn_weights_19_cast_fp16, cond = var_1704)[name = tensor("attn_weights_21_cast_fp16")]; + tensor var_1707 = const()[name = tensor("op_1707"), val = tensor(-1)]; + tensor attn_weights_23_cast_fp16 = softmax(axis = var_1707, x = attn_weights_21_cast_fp16)[name = tensor("attn_weights_23_cast_fp16")]; + tensor attn_output_7_transpose_x_0 = const()[name = tensor("attn_output_7_transpose_x_0"), val = tensor(false)]; + tensor attn_output_7_transpose_y_0 = const()[name = tensor("attn_output_7_transpose_y_0"), val = tensor(false)]; + tensor values_23_cast_fp16 = transpose(perm = var_1647, x = values_21_cast_fp16)[name = tensor("transpose_203")]; + tensor attn_output_7_cast_fp16 = matmul(transpose_x = attn_output_7_transpose_x_0, transpose_y = attn_output_7_transpose_y_0, x = attn_weights_23_cast_fp16, y = values_23_cast_fp16)[name = tensor("attn_output_7_cast_fp16")]; + tensor var_1715 = const()[name = tensor("op_1715"), val = tensor([0, 2, 1, 3])]; + tensor var_1720 = const()[name = tensor("op_1720"), val = tensor([1, 1, 1024])]; + tensor var_1716_cast_fp16 = transpose(perm = var_1715, x = attn_output_7_cast_fp16)[name = tensor("transpose_200")]; + tensor input_43_cast_fp16 = reshape(shape = var_1720, x = var_1716_cast_fp16)[name = tensor("input_43_cast_fp16")]; + tensor attn3_out_proj_weight_to_fp16 = const()[name = tensor("attn3_out_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(82951808)))]; + tensor linear_14_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = attn3_out_proj_weight_to_fp16, x = input_43_cast_fp16)[name = tensor("linear_14_cast_fp16")]; + tensor var_1726_to_fp16 = const()[name = tensor("op_1726_to_fp16"), val = tensor(0x1p+0)]; + tensor var_1727_cast_fp16 = add(x = position3_to_fp16, y = var_1726_to_fp16)[name = tensor("op_1727_cast_fp16")]; + tensor var_1727_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_1727_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor input_45_cast_fp16 = add(x = input_39_cast_fp16, y = linear_14_cast_fp16)[name = tensor("input_45_cast_fp16")]; + tensor input_47_axes_0 = const()[name = tensor("input_47_axes_0"), val = tensor([-1])]; + tensor norm3_2_weight_to_fp16 = const()[name = tensor("norm3_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(85049024)))]; + tensor norm3_2_bias_to_fp16 = const()[name = tensor("norm3_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(85051136)))]; + tensor var_1731_to_fp16 = const()[name = tensor("op_1731_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_47_cast_fp16 = layer_norm(axes = input_47_axes_0, beta = norm3_2_bias_to_fp16, epsilon = var_1731_to_fp16, gamma = norm3_2_weight_to_fp16, x = input_45_cast_fp16)[name = tensor("input_47_cast_fp16")]; + tensor linear3_1_weight_to_fp16 = const()[name = tensor("linear3_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(85053248)))]; + tensor linear_15_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = linear3_1_weight_to_fp16, x = input_47_cast_fp16)[name = tensor("linear_15_cast_fp16")]; + tensor input_49_mode_0 = const()[name = tensor("input_49_mode_0"), val = tensor("EXACT")]; + tensor input_49_cast_fp16 = gelu(mode = input_49_mode_0, x = linear_15_cast_fp16)[name = tensor("input_49_cast_fp16")]; + tensor linear3_2_weight_to_fp16 = const()[name = tensor("linear3_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(93441920)))]; + tensor linear_16_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = linear3_2_weight_to_fp16, x = input_49_cast_fp16)[name = tensor("linear_16_cast_fp16")]; + tensor input_51_cast_fp16 = add(x = input_45_cast_fp16, y = linear_16_cast_fp16)[name = tensor("input_51_cast_fp16")]; + tensor input_53_axes_0 = const()[name = tensor("input_53_axes_0"), val = tensor([-1])]; + tensor norm4_1_weight_to_fp16 = const()[name = tensor("norm4_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(101830592)))]; + tensor norm4_1_bias_to_fp16 = const()[name = tensor("norm4_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(101832704)))]; + tensor var_1748_to_fp16 = const()[name = tensor("op_1748_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_53_cast_fp16 = layer_norm(axes = input_53_axes_0, beta = norm4_1_bias_to_fp16, epsilon = var_1748_to_fp16, gamma = norm4_1_weight_to_fp16, x = input_51_cast_fp16)[name = tensor("input_53_cast_fp16")]; + tensor attn4_in_proj_weight_to_fp16 = const()[name = tensor("attn4_in_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(101834816)))]; + tensor linear_17_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = attn4_in_proj_weight_to_fp16, x = input_53_cast_fp16)[name = tensor("linear_17_cast_fp16")]; + tensor var_1762 = const()[name = tensor("op_1762"), val = tensor([1, 1, 3, 16, 64])]; + tensor qkv_9_cast_fp16 = reshape(shape = var_1762, x = linear_17_cast_fp16)[name = tensor("qkv_9_cast_fp16")]; + tensor q_25_begin_0 = const()[name = tensor("q_25_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor q_25_end_0 = const()[name = tensor("q_25_end_0"), val = tensor([1, 1, 1, 16, 64])]; + tensor q_25_end_mask_0 = const()[name = tensor("q_25_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor q_25_squeeze_mask_0 = const()[name = tensor("q_25_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor q_25_cast_fp16 = slice_by_index(begin = q_25_begin_0, end = q_25_end_0, end_mask = q_25_end_mask_0, squeeze_mask = q_25_squeeze_mask_0, x = qkv_9_cast_fp16)[name = tensor("q_25_cast_fp16")]; + tensor k_17_begin_0 = const()[name = tensor("k_17_begin_0"), val = tensor([0, 0, 1, 0, 0])]; + tensor k_17_end_0 = const()[name = tensor("k_17_end_0"), val = tensor([1, 1, 2, 16, 64])]; + tensor k_17_end_mask_0 = const()[name = tensor("k_17_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor k_17_squeeze_mask_0 = const()[name = tensor("k_17_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor k_17_cast_fp16 = slice_by_index(begin = k_17_begin_0, end = k_17_end_0, end_mask = k_17_end_mask_0, squeeze_mask = k_17_squeeze_mask_0, x = qkv_9_cast_fp16)[name = tensor("k_17_cast_fp16")]; + tensor v_9_begin_0 = const()[name = tensor("v_9_begin_0"), val = tensor([0, 0, 2, 0, 0])]; + tensor v_9_end_0 = const()[name = tensor("v_9_end_0"), val = tensor([1, 1, 3, 16, 64])]; + tensor v_9_end_mask_0 = const()[name = tensor("v_9_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor v_9_squeeze_mask_0 = const()[name = tensor("v_9_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor v_9_cast_fp16 = slice_by_index(begin = v_9_begin_0, end = v_9_end_0, end_mask = v_9_end_mask_0, squeeze_mask = v_9_squeeze_mask_0, x = qkv_9_cast_fp16)[name = tensor("v_9_cast_fp16")]; + tensor var_1828 = const()[name = tensor("op_1828"), val = tensor([1, 1, 1, 1])]; + tensor position4_to_fp16_dtype_0 = const()[name = tensor("position4_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor position4_to_fp16 = cast(dtype = position4_to_fp16_dtype_0, x = position4)[name = tensor("cast_533")]; + tensor ts_29_cast_fp16 = reshape(shape = var_1828, x = position4_to_fp16)[name = tensor("ts_29_cast_fp16")]; + tensor var_1835 = const()[name = tensor("op_1835"), val = tensor([1, 1, 16, 32, 2])]; + tensor q_complex_9_cast_fp16 = reshape(shape = var_1835, x = q_25_cast_fp16)[name = tensor("q_complex_9_cast_fp16")]; + tensor var_1842 = const()[name = tensor("op_1842"), val = tensor([1, 1, 16, 32, 2])]; + tensor k_complex_9_cast_fp16 = reshape(shape = var_1842, x = k_17_cast_fp16)[name = tensor("k_complex_9_cast_fp16")]; + tensor var_1846_begin_0 = const()[name = tensor("op_1846_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_1846_end_0 = const()[name = tensor("op_1846_end_0"), val = tensor([1, 1, 16, 32, 1])]; + tensor var_1846_end_mask_0 = const()[name = tensor("op_1846_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_1846_squeeze_mask_0 = const()[name = tensor("op_1846_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_1846_cast_fp16 = slice_by_index(begin = var_1846_begin_0, end = var_1846_end_0, end_mask = var_1846_end_mask_0, squeeze_mask = var_1846_squeeze_mask_0, x = q_complex_9_cast_fp16)[name = tensor("op_1846_cast_fp16")]; + tensor var_1854_begin_0 = const()[name = tensor("op_1854_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_1854_end_0 = const()[name = tensor("op_1854_end_0"), val = tensor([1, 1, 16, 32, 2])]; + tensor var_1854_end_mask_0 = const()[name = tensor("op_1854_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_1854_squeeze_mask_0 = const()[name = tensor("op_1854_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_1854_cast_fp16 = slice_by_index(begin = var_1854_begin_0, end = var_1854_end_0, end_mask = var_1854_end_mask_0, squeeze_mask = var_1854_squeeze_mask_0, x = q_complex_9_cast_fp16)[name = tensor("op_1854_cast_fp16")]; + tensor var_1862_begin_0 = const()[name = tensor("op_1862_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_1862_end_0 = const()[name = tensor("op_1862_end_0"), val = tensor([1, 1, 16, 32, 1])]; + tensor var_1862_end_mask_0 = const()[name = tensor("op_1862_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_1862_squeeze_mask_0 = const()[name = tensor("op_1862_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_1862_cast_fp16 = slice_by_index(begin = var_1862_begin_0, end = var_1862_end_0, end_mask = var_1862_end_mask_0, squeeze_mask = var_1862_squeeze_mask_0, x = k_complex_9_cast_fp16)[name = tensor("op_1862_cast_fp16")]; + tensor var_1870_begin_0 = const()[name = tensor("op_1870_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_1870_end_0 = const()[name = tensor("op_1870_end_0"), val = tensor([1, 1, 16, 32, 2])]; + tensor var_1870_end_mask_0 = const()[name = tensor("op_1870_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_1870_squeeze_mask_0 = const()[name = tensor("op_1870_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_1870_cast_fp16 = slice_by_index(begin = var_1870_begin_0, end = var_1870_end_0, end_mask = var_1870_end_mask_0, squeeze_mask = var_1870_squeeze_mask_0, x = k_complex_9_cast_fp16)[name = tensor("op_1870_cast_fp16")]; + tensor freqs_9_to_fp16 = const()[name = tensor("freqs_9_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(108126336)))]; + tensor var_1876_cast_fp16 = mul(x = freqs_9_to_fp16, y = ts_29_cast_fp16)[name = tensor("op_1876_cast_fp16")]; + tensor rotr_9_cast_fp16 = cos(x = var_1876_cast_fp16)[name = tensor("rotr_9_cast_fp16")]; + tensor roti_9_cast_fp16 = sin(x = var_1876_cast_fp16)[name = tensor("roti_9_cast_fp16")]; + tensor var_1880_cast_fp16 = mul(x = var_1846_cast_fp16, y = rotr_9_cast_fp16)[name = tensor("op_1880_cast_fp16")]; + tensor var_1881_cast_fp16 = mul(x = var_1854_cast_fp16, y = roti_9_cast_fp16)[name = tensor("op_1881_cast_fp16")]; + tensor qor_17_cast_fp16 = sub(x = var_1880_cast_fp16, y = var_1881_cast_fp16)[name = tensor("qor_17_cast_fp16")]; + tensor var_1884_cast_fp16 = mul(x = var_1846_cast_fp16, y = roti_9_cast_fp16)[name = tensor("op_1884_cast_fp16")]; + tensor var_1885_cast_fp16 = mul(x = var_1854_cast_fp16, y = rotr_9_cast_fp16)[name = tensor("op_1885_cast_fp16")]; + tensor qoi_17_cast_fp16 = add(x = var_1884_cast_fp16, y = var_1885_cast_fp16)[name = tensor("qoi_17_cast_fp16")]; + tensor var_1888_cast_fp16 = mul(x = var_1862_cast_fp16, y = rotr_9_cast_fp16)[name = tensor("op_1888_cast_fp16")]; + tensor var_1889_cast_fp16 = mul(x = var_1870_cast_fp16, y = roti_9_cast_fp16)[name = tensor("op_1889_cast_fp16")]; + tensor kor_17_cast_fp16 = sub(x = var_1888_cast_fp16, y = var_1889_cast_fp16)[name = tensor("kor_17_cast_fp16")]; + tensor var_1892_cast_fp16 = mul(x = var_1862_cast_fp16, y = roti_9_cast_fp16)[name = tensor("op_1892_cast_fp16")]; + tensor var_1893_cast_fp16 = mul(x = var_1870_cast_fp16, y = rotr_9_cast_fp16)[name = tensor("op_1893_cast_fp16")]; + tensor koi_17_cast_fp16 = add(x = var_1892_cast_fp16, y = var_1893_cast_fp16)[name = tensor("koi_17_cast_fp16")]; + tensor qo_9_axis_0 = const()[name = tensor("qo_9_axis_0"), val = tensor(-1)]; + tensor qo_9_cast_fp16 = stack(axis = qo_9_axis_0, values = (qor_17_cast_fp16, qoi_17_cast_fp16))[name = tensor("qo_9_cast_fp16")]; + tensor ko_9_axis_0 = const()[name = tensor("ko_9_axis_0"), val = tensor(-1)]; + tensor ko_9_cast_fp16 = stack(axis = ko_9_axis_0, values = (kor_17_cast_fp16, koi_17_cast_fp16))[name = tensor("ko_9_cast_fp16")]; + tensor var_1926 = const()[name = tensor("op_1926"), val = tensor([1, 1, 16, 64])]; + tensor q_27_cast_fp16 = reshape(shape = var_1926, x = qo_9_cast_fp16)[name = tensor("q_27_cast_fp16")]; + tensor var_1932 = const()[name = tensor("op_1932"), val = tensor([1, 1, 16, 64])]; + tensor k_19_cast_fp16 = reshape(shape = var_1932, x = ko_9_cast_fp16)[name = tensor("k_19_cast_fp16")]; + tensor _inversed_1957_y_0_to_fp16 = const()[name = tensor("_inversed_1957_y_0_to_fp16"), val = tensor(0x1p-9)]; + tensor _inversed_1957_cast_fp16 = mul(x = ts_29_cast_fp16, y = _inversed_1957_y_0_to_fp16)[name = tensor("_inversed_1957_cast_fp16")]; + tensor var_1958_cast_fp16 = floor(x = _inversed_1957_cast_fp16)[name = tensor("op_1958_cast_fp16")]; + tensor var_1959_to_fp16 = const()[name = tensor("op_1959_to_fp16"), val = tensor(0x1p+9)]; + tensor var_1960_cast_fp16 = mul(x = var_1958_cast_fp16, y = var_1959_to_fp16)[name = tensor("op_1960_cast_fp16")]; + tensor write_indices_float_19_cast_fp16 = sub(x = ts_29_cast_fp16, y = var_1960_cast_fp16)[name = tensor("write_indices_float_19_cast_fp16")]; + tensor var_1967_dtype_0 = const()[name = tensor("op_1967_dtype_0"), val = tensor("int32")]; + tensor write_indices_9_reps_0 = const()[name = tensor("write_indices_9_reps_0"), val = tensor([1, 1, 16, 64])]; + tensor write_indices_float_19_cast_fp16_to_int32 = cast(dtype = var_1967_dtype_0, x = write_indices_float_19_cast_fp16)[name = tensor("cast_532")]; + tensor write_indices_9 = tile(reps = write_indices_9_reps_0, x = write_indices_float_19_cast_fp16_to_int32)[name = tensor("write_indices_9")]; + tensor var_1977_begin_0 = const()[name = tensor("op_1977_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_1977_end_0 = const()[name = tensor("op_1977_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor var_1977_end_mask_0 = const()[name = tensor("op_1977_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_1977_squeeze_mask_0 = const()[name = tensor("op_1977_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor cache4_to_fp16_dtype_0 = const()[name = tensor("cache4_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor cache4_to_fp16 = cast(dtype = cache4_to_fp16_dtype_0, x = cache4)[name = tensor("cast_531")]; + tensor var_1977_cast_fp16 = slice_by_index(begin = var_1977_begin_0, end = var_1977_end_0, end_mask = var_1977_end_mask_0, squeeze_mask = var_1977_squeeze_mask_0, x = cache4_to_fp16)[name = tensor("op_1977_cast_fp16")]; + tensor var_1979_axis_0 = const()[name = tensor("op_1979_axis_0"), val = tensor(1)]; + tensor var_1979_mode_0 = const()[name = tensor("op_1979_mode_0"), val = tensor("update")]; + tensor var_1979_validate_indices_0 = const()[name = tensor("op_1979_validate_indices_0"), val = tensor(false)]; + tensor var_1979_cast_fp16 = scatter_along_axis(axis = var_1979_axis_0, data = var_1977_cast_fp16, indices = write_indices_9, mode = var_1979_mode_0, updates = k_19_cast_fp16, validate_indices = var_1979_validate_indices_0)[name = tensor("op_1979_cast_fp16")]; + tensor concat_30 = const()[name = tensor("concat_30"), val = tensor([0, 0, 0, 0, 0])]; + tensor concat_31 = const()[name = tensor("concat_31"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_9_internal_tensor_assign_1_stride_0 = const()[name = tensor("new_cache_9_internal_tensor_assign_1_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_9_internal_tensor_assign_1_begin_mask_0 = const()[name = tensor("new_cache_9_internal_tensor_assign_1_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_9_internal_tensor_assign_1_end_mask_0 = const()[name = tensor("new_cache_9_internal_tensor_assign_1_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_9_internal_tensor_assign_1_squeeze_mask_0 = const()[name = tensor("new_cache_9_internal_tensor_assign_1_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_56 = const()[name = tensor("shape_56"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_8 = const()[name = tensor("reduce_prod_8"), val = tensor(1048576)]; + tensor range_1d_8_start_0 = const()[name = tensor("range_1d_8_start_0"), val = tensor(0)]; + tensor range_1d_8_step_0 = const()[name = tensor("range_1d_8_step_0"), val = tensor(1)]; + tensor range_1d_8 = range_1d(end = reduce_prod_8, start = range_1d_8_start_0, step = range_1d_8_step_0)[name = tensor("range_1d_8")]; + tensor reshape_40 = reshape(shape = shape_56, x = range_1d_8)[name = tensor("reshape_40")]; + tensor slice_by_index_8 = slice_by_index(begin = concat_30, begin_mask = new_cache_9_internal_tensor_assign_1_begin_mask_0, end = concat_31, end_mask = new_cache_9_internal_tensor_assign_1_end_mask_0, squeeze_mask = new_cache_9_internal_tensor_assign_1_squeeze_mask_0, stride = new_cache_9_internal_tensor_assign_1_stride_0, x = reshape_40)[name = tensor("slice_by_index_8")]; + tensor reshape_41_shape_0 = const()[name = tensor("reshape_41_shape_0"), val = tensor([-1])]; + tensor reshape_41 = reshape(shape = reshape_41_shape_0, x = slice_by_index_8)[name = tensor("reshape_41")]; + tensor reshape_42_shape_0 = const()[name = tensor("reshape_42_shape_0"), val = tensor([-1])]; + tensor reshape_42_cast_fp16 = reshape(shape = reshape_42_shape_0, x = var_1979_cast_fp16)[name = tensor("reshape_42_cast_fp16")]; + tensor reshape_43_shape_0 = const()[name = tensor("reshape_43_shape_0"), val = tensor([-1])]; + tensor reshape_43_cast_fp16 = reshape(shape = reshape_43_shape_0, x = cache4_to_fp16)[name = tensor("reshape_43_cast_fp16")]; + tensor scatter_8_mode_0 = const()[name = tensor("scatter_8_mode_0"), val = tensor("update")]; + tensor scatter_8_axis_0 = const()[name = tensor("scatter_8_axis_0"), val = tensor(0)]; + tensor scatter_8_validate_indices_0 = const()[name = tensor("scatter_8_validate_indices_0"), val = tensor(false)]; + tensor scatter_8_cast_fp16 = scatter(axis = scatter_8_axis_0, data = reshape_43_cast_fp16, indices = reshape_41, mode = scatter_8_mode_0, updates = reshape_42_cast_fp16, validate_indices = scatter_8_validate_indices_0)[name = tensor("scatter_8_cast_fp16")]; + tensor reshape_44_cast_fp16 = reshape(shape = shape_56, x = scatter_8_cast_fp16)[name = tensor("reshape_44_cast_fp16")]; + tensor var_1987_begin_0 = const()[name = tensor("op_1987_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor var_1987_end_0 = const()[name = tensor("op_1987_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor var_1987_end_mask_0 = const()[name = tensor("op_1987_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_1987_squeeze_mask_0 = const()[name = tensor("op_1987_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor var_1987_cast_fp16 = slice_by_index(begin = var_1987_begin_0, end = var_1987_end_0, end_mask = var_1987_end_mask_0, squeeze_mask = var_1987_squeeze_mask_0, x = reshape_44_cast_fp16)[name = tensor("op_1987_cast_fp16")]; + tensor var_1989_axis_0 = const()[name = tensor("op_1989_axis_0"), val = tensor(1)]; + tensor var_1989_mode_0 = const()[name = tensor("op_1989_mode_0"), val = tensor("update")]; + tensor var_1989_validate_indices_0 = const()[name = tensor("op_1989_validate_indices_0"), val = tensor(false)]; + tensor var_1989_cast_fp16 = scatter_along_axis(axis = var_1989_axis_0, data = var_1987_cast_fp16, indices = write_indices_9, mode = var_1989_mode_0, updates = v_9_cast_fp16, validate_indices = var_1989_validate_indices_0)[name = tensor("op_1989_cast_fp16")]; + tensor concat_32 = const()[name = tensor("concat_32"), val = tensor([1, 0, 0, 0, 0])]; + tensor concat_33 = const()[name = tensor("concat_33"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_9_internal_tensor_assign_2_stride_0 = const()[name = tensor("new_cache_9_internal_tensor_assign_2_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_9_internal_tensor_assign_2_begin_mask_0 = const()[name = tensor("new_cache_9_internal_tensor_assign_2_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_9_internal_tensor_assign_2_end_mask_0 = const()[name = tensor("new_cache_9_internal_tensor_assign_2_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_9_internal_tensor_assign_2_squeeze_mask_0 = const()[name = tensor("new_cache_9_internal_tensor_assign_2_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_57 = const()[name = tensor("shape_57"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_9 = const()[name = tensor("reduce_prod_9"), val = tensor(1048576)]; + tensor range_1d_9_start_0 = const()[name = tensor("range_1d_9_start_0"), val = tensor(0)]; + tensor range_1d_9_step_0 = const()[name = tensor("range_1d_9_step_0"), val = tensor(1)]; + tensor range_1d_9 = range_1d(end = reduce_prod_9, start = range_1d_9_start_0, step = range_1d_9_step_0)[name = tensor("range_1d_9")]; + tensor reshape_45 = reshape(shape = shape_57, x = range_1d_9)[name = tensor("reshape_45")]; + tensor slice_by_index_9 = slice_by_index(begin = concat_32, begin_mask = new_cache_9_internal_tensor_assign_2_begin_mask_0, end = concat_33, end_mask = new_cache_9_internal_tensor_assign_2_end_mask_0, squeeze_mask = new_cache_9_internal_tensor_assign_2_squeeze_mask_0, stride = new_cache_9_internal_tensor_assign_2_stride_0, x = reshape_45)[name = tensor("slice_by_index_9")]; + tensor reshape_46_shape_0 = const()[name = tensor("reshape_46_shape_0"), val = tensor([-1])]; + tensor reshape_46 = reshape(shape = reshape_46_shape_0, x = slice_by_index_9)[name = tensor("reshape_46")]; + tensor reshape_47_shape_0 = const()[name = tensor("reshape_47_shape_0"), val = tensor([-1])]; + tensor reshape_47_cast_fp16 = reshape(shape = reshape_47_shape_0, x = var_1989_cast_fp16)[name = tensor("reshape_47_cast_fp16")]; + tensor reshape_48_shape_0 = const()[name = tensor("reshape_48_shape_0"), val = tensor([-1])]; + tensor reshape_48_cast_fp16 = reshape(shape = reshape_48_shape_0, x = reshape_44_cast_fp16)[name = tensor("reshape_48_cast_fp16")]; + tensor scatter_9_mode_0 = const()[name = tensor("scatter_9_mode_0"), val = tensor("update")]; + tensor scatter_9_axis_0 = const()[name = tensor("scatter_9_axis_0"), val = tensor(0)]; + tensor scatter_9_validate_indices_0 = const()[name = tensor("scatter_9_validate_indices_0"), val = tensor(false)]; + tensor scatter_9_cast_fp16 = scatter(axis = scatter_9_axis_0, data = reshape_48_cast_fp16, indices = reshape_46, mode = scatter_9_mode_0, updates = reshape_47_cast_fp16, validate_indices = scatter_9_validate_indices_0)[name = tensor("scatter_9_cast_fp16")]; + tensor reshape_49_cast_fp16 = reshape(shape = shape_57, x = scatter_9_cast_fp16)[name = tensor("reshape_49_cast_fp16")]; + tensor reshape_49_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("reshape_49_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor keys_25_begin_0 = const()[name = tensor("keys_25_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor keys_25_end_0 = const()[name = tensor("keys_25_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor keys_25_end_mask_0 = const()[name = tensor("keys_25_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor keys_25_squeeze_mask_0 = const()[name = tensor("keys_25_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor keys_25_cast_fp16 = slice_by_index(begin = keys_25_begin_0, end = keys_25_end_0, end_mask = keys_25_end_mask_0, squeeze_mask = keys_25_squeeze_mask_0, x = reshape_49_cast_fp16)[name = tensor("keys_25_cast_fp16")]; + tensor values_25_begin_0 = const()[name = tensor("values_25_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor values_25_end_0 = const()[name = tensor("values_25_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor values_25_end_mask_0 = const()[name = tensor("values_25_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor values_25_squeeze_mask_0 = const()[name = tensor("values_25_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor values_25_cast_fp16 = slice_by_index(begin = values_25_begin_0, end = values_25_end_0, end_mask = values_25_end_mask_0, squeeze_mask = values_25_squeeze_mask_0, x = reshape_49_cast_fp16)[name = tensor("values_25_cast_fp16")]; + tensor var_2001_cast_fp16 = not_equal(x = keys_25_cast_fp16, y = keys_25_cast_fp16)[name = tensor("op_2001_cast_fp16")]; + tensor keys_27_cast_fp16 = select(a = var_463_to_fp16, b = keys_25_cast_fp16, cond = var_2001_cast_fp16)[name = tensor("keys_27_cast_fp16")]; + tensor var_2009_cast_fp16 = not_equal(x = values_25_cast_fp16, y = values_25_cast_fp16)[name = tensor("op_2009_cast_fp16")]; + tensor values_27_cast_fp16 = select(a = var_463_to_fp16, b = values_25_cast_fp16, cond = var_2009_cast_fp16)[name = tensor("values_27_cast_fp16")]; + tensor var_2033 = const()[name = tensor("op_2033"), val = tensor([0, 2, 1, 3])]; + tensor var_2049 = const()[name = tensor("op_2049"), val = tensor([1, 1, 1])]; + tensor var_2050_cast_fp16 = reshape(shape = var_2049, x = position4_to_fp16)[name = tensor("op_2050_cast_fp16")]; + tensor var_2070_to_fp16 = const()[name = tensor("op_2070_to_fp16"), val = tensor(0x1p+0)]; + tensor valid_len_9_cast_fp16 = add(x = var_2050_cast_fp16, y = var_2070_to_fp16)[name = tensor("valid_len_9_cast_fp16")]; + tensor valid_mask_9_cast_fp16 = less(x = k_positions_1_promoted_to_fp16, y = valid_len_9_cast_fp16)[name = tensor("valid_mask_9_cast_fp16")]; + tensor causal_mask_9_cast_fp16 = less_equal(x = k_positions_1_promoted_to_fp16, y = var_2050_cast_fp16)[name = tensor("causal_mask_9_cast_fp16")]; + tensor attn_mask_17 = logical_and(x = valid_mask_9_cast_fp16, y = causal_mask_9_cast_fp16)[name = tensor("attn_mask_17")]; + tensor attn_mask_19_axes_0 = const()[name = tensor("attn_mask_19_axes_0"), val = tensor([1])]; + tensor attn_mask_19 = expand_dims(axes = attn_mask_19_axes_0, x = attn_mask_17)[name = tensor("attn_mask_19")]; + tensor var_2088_transpose_x_0 = const()[name = tensor("op_2088_transpose_x_0"), val = tensor(false)]; + tensor var_2088_transpose_y_0 = const()[name = tensor("op_2088_transpose_y_0"), val = tensor(false)]; + tensor transpose_80_perm_0 = const()[name = tensor("transpose_80_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_81_perm_0 = const()[name = tensor("transpose_81_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_81 = transpose(perm = transpose_81_perm_0, x = keys_27_cast_fp16)[name = tensor("transpose_197")]; + tensor transpose_80 = transpose(perm = transpose_80_perm_0, x = q_27_cast_fp16)[name = tensor("transpose_198")]; + tensor var_2088_cast_fp16 = matmul(transpose_x = var_2088_transpose_x_0, transpose_y = var_2088_transpose_y_0, x = transpose_80, y = transpose_81)[name = tensor("op_2088_cast_fp16")]; + tensor var_2082_to_fp16 = const()[name = tensor("op_2082_to_fp16"), val = tensor([0x1p-3])]; + tensor attn_weights_25_cast_fp16 = mul(x = var_2088_cast_fp16, y = var_2082_to_fp16)[name = tensor("attn_weights_25_cast_fp16")]; + tensor var_2090 = logical_not(x = attn_mask_19)[name = tensor("op_2090")]; + tensor var_2091_to_fp16 = const()[name = tensor("op_2091_to_fp16"), val = tensor(-inf)]; + tensor attn_weights_27_cast_fp16 = select(a = var_2091_to_fp16, b = attn_weights_25_cast_fp16, cond = var_2090)[name = tensor("attn_weights_27_cast_fp16")]; + tensor var_2093 = const()[name = tensor("op_2093"), val = tensor(-1)]; + tensor attn_weights_29_cast_fp16 = softmax(axis = var_2093, x = attn_weights_27_cast_fp16)[name = tensor("attn_weights_29_cast_fp16")]; + tensor attn_output_9_transpose_x_0 = const()[name = tensor("attn_output_9_transpose_x_0"), val = tensor(false)]; + tensor attn_output_9_transpose_y_0 = const()[name = tensor("attn_output_9_transpose_y_0"), val = tensor(false)]; + tensor values_29_cast_fp16 = transpose(perm = var_2033, x = values_27_cast_fp16)[name = tensor("transpose_199")]; + tensor attn_output_9_cast_fp16 = matmul(transpose_x = attn_output_9_transpose_x_0, transpose_y = attn_output_9_transpose_y_0, x = attn_weights_29_cast_fp16, y = values_29_cast_fp16)[name = tensor("attn_output_9_cast_fp16")]; + tensor var_2101 = const()[name = tensor("op_2101"), val = tensor([0, 2, 1, 3])]; + tensor var_2106 = const()[name = tensor("op_2106"), val = tensor([1, 1, 1024])]; + tensor var_2102_cast_fp16 = transpose(perm = var_2101, x = attn_output_9_cast_fp16)[name = tensor("transpose_196")]; + tensor input_55_cast_fp16 = reshape(shape = var_2106, x = var_2102_cast_fp16)[name = tensor("input_55_cast_fp16")]; + tensor attn4_out_proj_weight_to_fp16 = const()[name = tensor("attn4_out_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(108126464)))]; + tensor linear_18_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = attn4_out_proj_weight_to_fp16, x = input_55_cast_fp16)[name = tensor("linear_18_cast_fp16")]; + tensor var_2112_to_fp16 = const()[name = tensor("op_2112_to_fp16"), val = tensor(0x1p+0)]; + tensor var_2113_cast_fp16 = add(x = position4_to_fp16, y = var_2112_to_fp16)[name = tensor("op_2113_cast_fp16")]; + tensor var_2113_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_2113_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor input_57_cast_fp16 = add(x = input_51_cast_fp16, y = linear_18_cast_fp16)[name = tensor("input_57_cast_fp16")]; + tensor input_59_axes_0 = const()[name = tensor("input_59_axes_0"), val = tensor([-1])]; + tensor norm4_2_weight_to_fp16 = const()[name = tensor("norm4_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(110223680)))]; + tensor norm4_2_bias_to_fp16 = const()[name = tensor("norm4_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(110225792)))]; + tensor var_2117_to_fp16 = const()[name = tensor("op_2117_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_59_cast_fp16 = layer_norm(axes = input_59_axes_0, beta = norm4_2_bias_to_fp16, epsilon = var_2117_to_fp16, gamma = norm4_2_weight_to_fp16, x = input_57_cast_fp16)[name = tensor("input_59_cast_fp16")]; + tensor linear4_1_weight_to_fp16 = const()[name = tensor("linear4_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(110227904)))]; + tensor linear_19_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = linear4_1_weight_to_fp16, x = input_59_cast_fp16)[name = tensor("linear_19_cast_fp16")]; + tensor input_61_mode_0 = const()[name = tensor("input_61_mode_0"), val = tensor("EXACT")]; + tensor input_61_cast_fp16 = gelu(mode = input_61_mode_0, x = linear_19_cast_fp16)[name = tensor("input_61_cast_fp16")]; + tensor linear4_2_weight_to_fp16 = const()[name = tensor("linear4_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(118616576)))]; + tensor linear_20_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = linear4_2_weight_to_fp16, x = input_61_cast_fp16)[name = tensor("linear_20_cast_fp16")]; + tensor input_63_cast_fp16 = add(x = input_57_cast_fp16, y = linear_20_cast_fp16)[name = tensor("input_63_cast_fp16")]; + tensor input_65_axes_0 = const()[name = tensor("input_65_axes_0"), val = tensor([-1])]; + tensor norm5_1_weight_to_fp16 = const()[name = tensor("norm5_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(127005248)))]; + tensor norm5_1_bias_to_fp16 = const()[name = tensor("norm5_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(127007360)))]; + tensor var_2134_to_fp16 = const()[name = tensor("op_2134_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_65_cast_fp16 = layer_norm(axes = input_65_axes_0, beta = norm5_1_bias_to_fp16, epsilon = var_2134_to_fp16, gamma = norm5_1_weight_to_fp16, x = input_63_cast_fp16)[name = tensor("input_65_cast_fp16")]; + tensor attn5_in_proj_weight_to_fp16 = const()[name = tensor("attn5_in_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(127009472)))]; + tensor linear_21_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = attn5_in_proj_weight_to_fp16, x = input_65_cast_fp16)[name = tensor("linear_21_cast_fp16")]; + tensor var_2148 = const()[name = tensor("op_2148"), val = tensor([1, 1, 3, 16, 64])]; + tensor qkv_11_cast_fp16 = reshape(shape = var_2148, x = linear_21_cast_fp16)[name = tensor("qkv_11_cast_fp16")]; + tensor q_31_begin_0 = const()[name = tensor("q_31_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor q_31_end_0 = const()[name = tensor("q_31_end_0"), val = tensor([1, 1, 1, 16, 64])]; + tensor q_31_end_mask_0 = const()[name = tensor("q_31_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor q_31_squeeze_mask_0 = const()[name = tensor("q_31_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor q_31_cast_fp16 = slice_by_index(begin = q_31_begin_0, end = q_31_end_0, end_mask = q_31_end_mask_0, squeeze_mask = q_31_squeeze_mask_0, x = qkv_11_cast_fp16)[name = tensor("q_31_cast_fp16")]; + tensor k_21_begin_0 = const()[name = tensor("k_21_begin_0"), val = tensor([0, 0, 1, 0, 0])]; + tensor k_21_end_0 = const()[name = tensor("k_21_end_0"), val = tensor([1, 1, 2, 16, 64])]; + tensor k_21_end_mask_0 = const()[name = tensor("k_21_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor k_21_squeeze_mask_0 = const()[name = tensor("k_21_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor k_21_cast_fp16 = slice_by_index(begin = k_21_begin_0, end = k_21_end_0, end_mask = k_21_end_mask_0, squeeze_mask = k_21_squeeze_mask_0, x = qkv_11_cast_fp16)[name = tensor("k_21_cast_fp16")]; + tensor v_11_begin_0 = const()[name = tensor("v_11_begin_0"), val = tensor([0, 0, 2, 0, 0])]; + tensor v_11_end_0 = const()[name = tensor("v_11_end_0"), val = tensor([1, 1, 3, 16, 64])]; + tensor v_11_end_mask_0 = const()[name = tensor("v_11_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor v_11_squeeze_mask_0 = const()[name = tensor("v_11_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor v_11_cast_fp16 = slice_by_index(begin = v_11_begin_0, end = v_11_end_0, end_mask = v_11_end_mask_0, squeeze_mask = v_11_squeeze_mask_0, x = qkv_11_cast_fp16)[name = tensor("v_11_cast_fp16")]; + tensor var_2214 = const()[name = tensor("op_2214"), val = tensor([1, 1, 1, 1])]; + tensor position5_to_fp16_dtype_0 = const()[name = tensor("position5_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor position5_to_fp16 = cast(dtype = position5_to_fp16_dtype_0, x = position5)[name = tensor("cast_528")]; + tensor ts_35_cast_fp16 = reshape(shape = var_2214, x = position5_to_fp16)[name = tensor("ts_35_cast_fp16")]; + tensor var_2221 = const()[name = tensor("op_2221"), val = tensor([1, 1, 16, 32, 2])]; + tensor q_complex_11_cast_fp16 = reshape(shape = var_2221, x = q_31_cast_fp16)[name = tensor("q_complex_11_cast_fp16")]; + tensor var_2228 = const()[name = tensor("op_2228"), val = tensor([1, 1, 16, 32, 2])]; + tensor k_complex_11_cast_fp16 = reshape(shape = var_2228, x = k_21_cast_fp16)[name = tensor("k_complex_11_cast_fp16")]; + tensor var_2232_begin_0 = const()[name = tensor("op_2232_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_2232_end_0 = const()[name = tensor("op_2232_end_0"), val = tensor([1, 1, 16, 32, 1])]; + tensor var_2232_end_mask_0 = const()[name = tensor("op_2232_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_2232_squeeze_mask_0 = const()[name = tensor("op_2232_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_2232_cast_fp16 = slice_by_index(begin = var_2232_begin_0, end = var_2232_end_0, end_mask = var_2232_end_mask_0, squeeze_mask = var_2232_squeeze_mask_0, x = q_complex_11_cast_fp16)[name = tensor("op_2232_cast_fp16")]; + tensor var_2240_begin_0 = const()[name = tensor("op_2240_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_2240_end_0 = const()[name = tensor("op_2240_end_0"), val = tensor([1, 1, 16, 32, 2])]; + tensor var_2240_end_mask_0 = const()[name = tensor("op_2240_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_2240_squeeze_mask_0 = const()[name = tensor("op_2240_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_2240_cast_fp16 = slice_by_index(begin = var_2240_begin_0, end = var_2240_end_0, end_mask = var_2240_end_mask_0, squeeze_mask = var_2240_squeeze_mask_0, x = q_complex_11_cast_fp16)[name = tensor("op_2240_cast_fp16")]; + tensor var_2248_begin_0 = const()[name = tensor("op_2248_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_2248_end_0 = const()[name = tensor("op_2248_end_0"), val = tensor([1, 1, 16, 32, 1])]; + tensor var_2248_end_mask_0 = const()[name = tensor("op_2248_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_2248_squeeze_mask_0 = const()[name = tensor("op_2248_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_2248_cast_fp16 = slice_by_index(begin = var_2248_begin_0, end = var_2248_end_0, end_mask = var_2248_end_mask_0, squeeze_mask = var_2248_squeeze_mask_0, x = k_complex_11_cast_fp16)[name = tensor("op_2248_cast_fp16")]; + tensor var_2256_begin_0 = const()[name = tensor("op_2256_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_2256_end_0 = const()[name = tensor("op_2256_end_0"), val = tensor([1, 1, 16, 32, 2])]; + tensor var_2256_end_mask_0 = const()[name = tensor("op_2256_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_2256_squeeze_mask_0 = const()[name = tensor("op_2256_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_2256_cast_fp16 = slice_by_index(begin = var_2256_begin_0, end = var_2256_end_0, end_mask = var_2256_end_mask_0, squeeze_mask = var_2256_squeeze_mask_0, x = k_complex_11_cast_fp16)[name = tensor("op_2256_cast_fp16")]; + tensor freqs_11_to_fp16 = const()[name = tensor("freqs_11_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(133300992)))]; + tensor var_2262_cast_fp16 = mul(x = freqs_11_to_fp16, y = ts_35_cast_fp16)[name = tensor("op_2262_cast_fp16")]; + tensor rotr_11_cast_fp16 = cos(x = var_2262_cast_fp16)[name = tensor("rotr_11_cast_fp16")]; + tensor roti_11_cast_fp16 = sin(x = var_2262_cast_fp16)[name = tensor("roti_11_cast_fp16")]; + tensor var_2266_cast_fp16 = mul(x = var_2232_cast_fp16, y = rotr_11_cast_fp16)[name = tensor("op_2266_cast_fp16")]; + tensor var_2267_cast_fp16 = mul(x = var_2240_cast_fp16, y = roti_11_cast_fp16)[name = tensor("op_2267_cast_fp16")]; + tensor qor_21_cast_fp16 = sub(x = var_2266_cast_fp16, y = var_2267_cast_fp16)[name = tensor("qor_21_cast_fp16")]; + tensor var_2270_cast_fp16 = mul(x = var_2232_cast_fp16, y = roti_11_cast_fp16)[name = tensor("op_2270_cast_fp16")]; + tensor var_2271_cast_fp16 = mul(x = var_2240_cast_fp16, y = rotr_11_cast_fp16)[name = tensor("op_2271_cast_fp16")]; + tensor qoi_21_cast_fp16 = add(x = var_2270_cast_fp16, y = var_2271_cast_fp16)[name = tensor("qoi_21_cast_fp16")]; + tensor var_2274_cast_fp16 = mul(x = var_2248_cast_fp16, y = rotr_11_cast_fp16)[name = tensor("op_2274_cast_fp16")]; + tensor var_2275_cast_fp16 = mul(x = var_2256_cast_fp16, y = roti_11_cast_fp16)[name = tensor("op_2275_cast_fp16")]; + tensor kor_21_cast_fp16 = sub(x = var_2274_cast_fp16, y = var_2275_cast_fp16)[name = tensor("kor_21_cast_fp16")]; + tensor var_2278_cast_fp16 = mul(x = var_2248_cast_fp16, y = roti_11_cast_fp16)[name = tensor("op_2278_cast_fp16")]; + tensor var_2279_cast_fp16 = mul(x = var_2256_cast_fp16, y = rotr_11_cast_fp16)[name = tensor("op_2279_cast_fp16")]; + tensor koi_21_cast_fp16 = add(x = var_2278_cast_fp16, y = var_2279_cast_fp16)[name = tensor("koi_21_cast_fp16")]; + tensor qo_11_axis_0 = const()[name = tensor("qo_11_axis_0"), val = tensor(-1)]; + tensor qo_11_cast_fp16 = stack(axis = qo_11_axis_0, values = (qor_21_cast_fp16, qoi_21_cast_fp16))[name = tensor("qo_11_cast_fp16")]; + tensor ko_11_axis_0 = const()[name = tensor("ko_11_axis_0"), val = tensor(-1)]; + tensor ko_11_cast_fp16 = stack(axis = ko_11_axis_0, values = (kor_21_cast_fp16, koi_21_cast_fp16))[name = tensor("ko_11_cast_fp16")]; + tensor var_2312 = const()[name = tensor("op_2312"), val = tensor([1, 1, 16, 64])]; + tensor q_33_cast_fp16 = reshape(shape = var_2312, x = qo_11_cast_fp16)[name = tensor("q_33_cast_fp16")]; + tensor var_2318 = const()[name = tensor("op_2318"), val = tensor([1, 1, 16, 64])]; + tensor k_23_cast_fp16 = reshape(shape = var_2318, x = ko_11_cast_fp16)[name = tensor("k_23_cast_fp16")]; + tensor _inversed_2343_y_0_to_fp16 = const()[name = tensor("_inversed_2343_y_0_to_fp16"), val = tensor(0x1p-9)]; + tensor _inversed_2343_cast_fp16 = mul(x = ts_35_cast_fp16, y = _inversed_2343_y_0_to_fp16)[name = tensor("_inversed_2343_cast_fp16")]; + tensor var_2344_cast_fp16 = floor(x = _inversed_2343_cast_fp16)[name = tensor("op_2344_cast_fp16")]; + tensor var_2345_to_fp16 = const()[name = tensor("op_2345_to_fp16"), val = tensor(0x1p+9)]; + tensor var_2346_cast_fp16 = mul(x = var_2344_cast_fp16, y = var_2345_to_fp16)[name = tensor("op_2346_cast_fp16")]; + tensor write_indices_float_23_cast_fp16 = sub(x = ts_35_cast_fp16, y = var_2346_cast_fp16)[name = tensor("write_indices_float_23_cast_fp16")]; + tensor var_2353_dtype_0 = const()[name = tensor("op_2353_dtype_0"), val = tensor("int32")]; + tensor write_indices_11_reps_0 = const()[name = tensor("write_indices_11_reps_0"), val = tensor([1, 1, 16, 64])]; + tensor write_indices_float_23_cast_fp16_to_int32 = cast(dtype = var_2353_dtype_0, x = write_indices_float_23_cast_fp16)[name = tensor("cast_527")]; + tensor write_indices_11 = tile(reps = write_indices_11_reps_0, x = write_indices_float_23_cast_fp16_to_int32)[name = tensor("write_indices_11")]; + tensor var_2363_begin_0 = const()[name = tensor("op_2363_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_2363_end_0 = const()[name = tensor("op_2363_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor var_2363_end_mask_0 = const()[name = tensor("op_2363_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_2363_squeeze_mask_0 = const()[name = tensor("op_2363_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor cache5_to_fp16_dtype_0 = const()[name = tensor("cache5_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor cache5_to_fp16 = cast(dtype = cache5_to_fp16_dtype_0, x = cache5)[name = tensor("cast_526")]; + tensor var_2363_cast_fp16 = slice_by_index(begin = var_2363_begin_0, end = var_2363_end_0, end_mask = var_2363_end_mask_0, squeeze_mask = var_2363_squeeze_mask_0, x = cache5_to_fp16)[name = tensor("op_2363_cast_fp16")]; + tensor var_2365_axis_0 = const()[name = tensor("op_2365_axis_0"), val = tensor(1)]; + tensor var_2365_mode_0 = const()[name = tensor("op_2365_mode_0"), val = tensor("update")]; + tensor var_2365_validate_indices_0 = const()[name = tensor("op_2365_validate_indices_0"), val = tensor(false)]; + tensor var_2365_cast_fp16 = scatter_along_axis(axis = var_2365_axis_0, data = var_2363_cast_fp16, indices = write_indices_11, mode = var_2365_mode_0, updates = k_23_cast_fp16, validate_indices = var_2365_validate_indices_0)[name = tensor("op_2365_cast_fp16")]; + tensor concat_37 = const()[name = tensor("concat_37"), val = tensor([0, 0, 0, 0, 0])]; + tensor concat_38 = const()[name = tensor("concat_38"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_11_internal_tensor_assign_1_stride_0 = const()[name = tensor("new_cache_11_internal_tensor_assign_1_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_11_internal_tensor_assign_1_begin_mask_0 = const()[name = tensor("new_cache_11_internal_tensor_assign_1_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_11_internal_tensor_assign_1_end_mask_0 = const()[name = tensor("new_cache_11_internal_tensor_assign_1_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_11_internal_tensor_assign_1_squeeze_mask_0 = const()[name = tensor("new_cache_11_internal_tensor_assign_1_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_58 = const()[name = tensor("shape_58"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_10 = const()[name = tensor("reduce_prod_10"), val = tensor(1048576)]; + tensor range_1d_10_start_0 = const()[name = tensor("range_1d_10_start_0"), val = tensor(0)]; + tensor range_1d_10_step_0 = const()[name = tensor("range_1d_10_step_0"), val = tensor(1)]; + tensor range_1d_10 = range_1d(end = reduce_prod_10, start = range_1d_10_start_0, step = range_1d_10_step_0)[name = tensor("range_1d_10")]; + tensor reshape_50 = reshape(shape = shape_58, x = range_1d_10)[name = tensor("reshape_50")]; + tensor slice_by_index_10 = slice_by_index(begin = concat_37, begin_mask = new_cache_11_internal_tensor_assign_1_begin_mask_0, end = concat_38, end_mask = new_cache_11_internal_tensor_assign_1_end_mask_0, squeeze_mask = new_cache_11_internal_tensor_assign_1_squeeze_mask_0, stride = new_cache_11_internal_tensor_assign_1_stride_0, x = reshape_50)[name = tensor("slice_by_index_10")]; + tensor reshape_51_shape_0 = const()[name = tensor("reshape_51_shape_0"), val = tensor([-1])]; + tensor reshape_51 = reshape(shape = reshape_51_shape_0, x = slice_by_index_10)[name = tensor("reshape_51")]; + tensor reshape_52_shape_0 = const()[name = tensor("reshape_52_shape_0"), val = tensor([-1])]; + tensor reshape_52_cast_fp16 = reshape(shape = reshape_52_shape_0, x = var_2365_cast_fp16)[name = tensor("reshape_52_cast_fp16")]; + tensor reshape_53_shape_0 = const()[name = tensor("reshape_53_shape_0"), val = tensor([-1])]; + tensor reshape_53_cast_fp16 = reshape(shape = reshape_53_shape_0, x = cache5_to_fp16)[name = tensor("reshape_53_cast_fp16")]; + tensor scatter_10_mode_0 = const()[name = tensor("scatter_10_mode_0"), val = tensor("update")]; + tensor scatter_10_axis_0 = const()[name = tensor("scatter_10_axis_0"), val = tensor(0)]; + tensor scatter_10_validate_indices_0 = const()[name = tensor("scatter_10_validate_indices_0"), val = tensor(false)]; + tensor scatter_10_cast_fp16 = scatter(axis = scatter_10_axis_0, data = reshape_53_cast_fp16, indices = reshape_51, mode = scatter_10_mode_0, updates = reshape_52_cast_fp16, validate_indices = scatter_10_validate_indices_0)[name = tensor("scatter_10_cast_fp16")]; + tensor reshape_54_cast_fp16 = reshape(shape = shape_58, x = scatter_10_cast_fp16)[name = tensor("reshape_54_cast_fp16")]; + tensor var_2373_begin_0 = const()[name = tensor("op_2373_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor var_2373_end_0 = const()[name = tensor("op_2373_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor var_2373_end_mask_0 = const()[name = tensor("op_2373_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_2373_squeeze_mask_0 = const()[name = tensor("op_2373_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor var_2373_cast_fp16 = slice_by_index(begin = var_2373_begin_0, end = var_2373_end_0, end_mask = var_2373_end_mask_0, squeeze_mask = var_2373_squeeze_mask_0, x = reshape_54_cast_fp16)[name = tensor("op_2373_cast_fp16")]; + tensor var_2375_axis_0 = const()[name = tensor("op_2375_axis_0"), val = tensor(1)]; + tensor var_2375_mode_0 = const()[name = tensor("op_2375_mode_0"), val = tensor("update")]; + tensor var_2375_validate_indices_0 = const()[name = tensor("op_2375_validate_indices_0"), val = tensor(false)]; + tensor var_2375_cast_fp16 = scatter_along_axis(axis = var_2375_axis_0, data = var_2373_cast_fp16, indices = write_indices_11, mode = var_2375_mode_0, updates = v_11_cast_fp16, validate_indices = var_2375_validate_indices_0)[name = tensor("op_2375_cast_fp16")]; + tensor concat_39 = const()[name = tensor("concat_39"), val = tensor([1, 0, 0, 0, 0])]; + tensor concat_40 = const()[name = tensor("concat_40"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_11_internal_tensor_assign_2_stride_0 = const()[name = tensor("new_cache_11_internal_tensor_assign_2_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_11_internal_tensor_assign_2_begin_mask_0 = const()[name = tensor("new_cache_11_internal_tensor_assign_2_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_11_internal_tensor_assign_2_end_mask_0 = const()[name = tensor("new_cache_11_internal_tensor_assign_2_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_11_internal_tensor_assign_2_squeeze_mask_0 = const()[name = tensor("new_cache_11_internal_tensor_assign_2_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_59 = const()[name = tensor("shape_59"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_11 = const()[name = tensor("reduce_prod_11"), val = tensor(1048576)]; + tensor range_1d_11_start_0 = const()[name = tensor("range_1d_11_start_0"), val = tensor(0)]; + tensor range_1d_11_step_0 = const()[name = tensor("range_1d_11_step_0"), val = tensor(1)]; + tensor range_1d_11 = range_1d(end = reduce_prod_11, start = range_1d_11_start_0, step = range_1d_11_step_0)[name = tensor("range_1d_11")]; + tensor reshape_55 = reshape(shape = shape_59, x = range_1d_11)[name = tensor("reshape_55")]; + tensor slice_by_index_11 = slice_by_index(begin = concat_39, begin_mask = new_cache_11_internal_tensor_assign_2_begin_mask_0, end = concat_40, end_mask = new_cache_11_internal_tensor_assign_2_end_mask_0, squeeze_mask = new_cache_11_internal_tensor_assign_2_squeeze_mask_0, stride = new_cache_11_internal_tensor_assign_2_stride_0, x = reshape_55)[name = tensor("slice_by_index_11")]; + tensor reshape_56_shape_0 = const()[name = tensor("reshape_56_shape_0"), val = tensor([-1])]; + tensor reshape_56 = reshape(shape = reshape_56_shape_0, x = slice_by_index_11)[name = tensor("reshape_56")]; + tensor reshape_57_shape_0 = const()[name = tensor("reshape_57_shape_0"), val = tensor([-1])]; + tensor reshape_57_cast_fp16 = reshape(shape = reshape_57_shape_0, x = var_2375_cast_fp16)[name = tensor("reshape_57_cast_fp16")]; + tensor reshape_58_shape_0 = const()[name = tensor("reshape_58_shape_0"), val = tensor([-1])]; + tensor reshape_58_cast_fp16 = reshape(shape = reshape_58_shape_0, x = reshape_54_cast_fp16)[name = tensor("reshape_58_cast_fp16")]; + tensor scatter_11_mode_0 = const()[name = tensor("scatter_11_mode_0"), val = tensor("update")]; + tensor scatter_11_axis_0 = const()[name = tensor("scatter_11_axis_0"), val = tensor(0)]; + tensor scatter_11_validate_indices_0 = const()[name = tensor("scatter_11_validate_indices_0"), val = tensor(false)]; + tensor scatter_11_cast_fp16 = scatter(axis = scatter_11_axis_0, data = reshape_58_cast_fp16, indices = reshape_56, mode = scatter_11_mode_0, updates = reshape_57_cast_fp16, validate_indices = scatter_11_validate_indices_0)[name = tensor("scatter_11_cast_fp16")]; + tensor reshape_59_cast_fp16 = reshape(shape = shape_59, x = scatter_11_cast_fp16)[name = tensor("reshape_59_cast_fp16")]; + tensor reshape_59_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("reshape_59_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor keys_31_begin_0 = const()[name = tensor("keys_31_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor keys_31_end_0 = const()[name = tensor("keys_31_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor keys_31_end_mask_0 = const()[name = tensor("keys_31_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor keys_31_squeeze_mask_0 = const()[name = tensor("keys_31_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor keys_31_cast_fp16 = slice_by_index(begin = keys_31_begin_0, end = keys_31_end_0, end_mask = keys_31_end_mask_0, squeeze_mask = keys_31_squeeze_mask_0, x = reshape_59_cast_fp16)[name = tensor("keys_31_cast_fp16")]; + tensor values_31_begin_0 = const()[name = tensor("values_31_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor values_31_end_0 = const()[name = tensor("values_31_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor values_31_end_mask_0 = const()[name = tensor("values_31_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor values_31_squeeze_mask_0 = const()[name = tensor("values_31_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor values_31_cast_fp16 = slice_by_index(begin = values_31_begin_0, end = values_31_end_0, end_mask = values_31_end_mask_0, squeeze_mask = values_31_squeeze_mask_0, x = reshape_59_cast_fp16)[name = tensor("values_31_cast_fp16")]; + tensor var_2387_cast_fp16 = not_equal(x = keys_31_cast_fp16, y = keys_31_cast_fp16)[name = tensor("op_2387_cast_fp16")]; + tensor keys_33_cast_fp16 = select(a = var_463_to_fp16, b = keys_31_cast_fp16, cond = var_2387_cast_fp16)[name = tensor("keys_33_cast_fp16")]; + tensor var_2395_cast_fp16 = not_equal(x = values_31_cast_fp16, y = values_31_cast_fp16)[name = tensor("op_2395_cast_fp16")]; + tensor values_33_cast_fp16 = select(a = var_463_to_fp16, b = values_31_cast_fp16, cond = var_2395_cast_fp16)[name = tensor("values_33_cast_fp16")]; + tensor var_2419 = const()[name = tensor("op_2419"), val = tensor([0, 2, 1, 3])]; + tensor var_2435 = const()[name = tensor("op_2435"), val = tensor([1, 1, 1])]; + tensor var_2436_cast_fp16 = reshape(shape = var_2435, x = position5_to_fp16)[name = tensor("op_2436_cast_fp16")]; + tensor var_2456_to_fp16 = const()[name = tensor("op_2456_to_fp16"), val = tensor(0x1p+0)]; + tensor valid_len_11_cast_fp16 = add(x = var_2436_cast_fp16, y = var_2456_to_fp16)[name = tensor("valid_len_11_cast_fp16")]; + tensor valid_mask_11_cast_fp16 = less(x = k_positions_1_promoted_to_fp16, y = valid_len_11_cast_fp16)[name = tensor("valid_mask_11_cast_fp16")]; + tensor causal_mask_11_cast_fp16 = less_equal(x = k_positions_1_promoted_to_fp16, y = var_2436_cast_fp16)[name = tensor("causal_mask_11_cast_fp16")]; + tensor attn_mask_21 = logical_and(x = valid_mask_11_cast_fp16, y = causal_mask_11_cast_fp16)[name = tensor("attn_mask_21")]; + tensor attn_mask_23_axes_0 = const()[name = tensor("attn_mask_23_axes_0"), val = tensor([1])]; + tensor attn_mask_23 = expand_dims(axes = attn_mask_23_axes_0, x = attn_mask_21)[name = tensor("attn_mask_23")]; + tensor var_2474_transpose_x_0 = const()[name = tensor("op_2474_transpose_x_0"), val = tensor(false)]; + tensor var_2474_transpose_y_0 = const()[name = tensor("op_2474_transpose_y_0"), val = tensor(false)]; + tensor transpose_82_perm_0 = const()[name = tensor("transpose_82_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_83_perm_0 = const()[name = tensor("transpose_83_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_83 = transpose(perm = transpose_83_perm_0, x = keys_33_cast_fp16)[name = tensor("transpose_193")]; + tensor transpose_82 = transpose(perm = transpose_82_perm_0, x = q_33_cast_fp16)[name = tensor("transpose_194")]; + tensor var_2474_cast_fp16 = matmul(transpose_x = var_2474_transpose_x_0, transpose_y = var_2474_transpose_y_0, x = transpose_82, y = transpose_83)[name = tensor("op_2474_cast_fp16")]; + tensor var_2468_to_fp16 = const()[name = tensor("op_2468_to_fp16"), val = tensor([0x1p-3])]; + tensor attn_weights_31_cast_fp16 = mul(x = var_2474_cast_fp16, y = var_2468_to_fp16)[name = tensor("attn_weights_31_cast_fp16")]; + tensor var_2476 = logical_not(x = attn_mask_23)[name = tensor("op_2476")]; + tensor var_2477_to_fp16 = const()[name = tensor("op_2477_to_fp16"), val = tensor(-inf)]; + tensor attn_weights_33_cast_fp16 = select(a = var_2477_to_fp16, b = attn_weights_31_cast_fp16, cond = var_2476)[name = tensor("attn_weights_33_cast_fp16")]; + tensor var_2479 = const()[name = tensor("op_2479"), val = tensor(-1)]; + tensor attn_weights_35_cast_fp16 = softmax(axis = var_2479, x = attn_weights_33_cast_fp16)[name = tensor("attn_weights_35_cast_fp16")]; + tensor attn_output_11_transpose_x_0 = const()[name = tensor("attn_output_11_transpose_x_0"), val = tensor(false)]; + tensor attn_output_11_transpose_y_0 = const()[name = tensor("attn_output_11_transpose_y_0"), val = tensor(false)]; + tensor values_35_cast_fp16 = transpose(perm = var_2419, x = values_33_cast_fp16)[name = tensor("transpose_195")]; + tensor attn_output_11_cast_fp16 = matmul(transpose_x = attn_output_11_transpose_x_0, transpose_y = attn_output_11_transpose_y_0, x = attn_weights_35_cast_fp16, y = values_35_cast_fp16)[name = tensor("attn_output_11_cast_fp16")]; + tensor var_2487 = const()[name = tensor("op_2487"), val = tensor([0, 2, 1, 3])]; + tensor var_2492 = const()[name = tensor("op_2492"), val = tensor([1, 1, 1024])]; + tensor var_2488_cast_fp16 = transpose(perm = var_2487, x = attn_output_11_cast_fp16)[name = tensor("transpose_192")]; + tensor input_67_cast_fp16 = reshape(shape = var_2492, x = var_2488_cast_fp16)[name = tensor("input_67_cast_fp16")]; + tensor attn5_out_proj_weight_to_fp16 = const()[name = tensor("attn5_out_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(133301120)))]; + tensor linear_22_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = attn5_out_proj_weight_to_fp16, x = input_67_cast_fp16)[name = tensor("linear_22_cast_fp16")]; + tensor var_2498_to_fp16 = const()[name = tensor("op_2498_to_fp16"), val = tensor(0x1p+0)]; + tensor var_2499_cast_fp16 = add(x = position5_to_fp16, y = var_2498_to_fp16)[name = tensor("op_2499_cast_fp16")]; + tensor var_2499_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_2499_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor input_69_cast_fp16 = add(x = input_63_cast_fp16, y = linear_22_cast_fp16)[name = tensor("input_69_cast_fp16")]; + tensor input_71_axes_0 = const()[name = tensor("input_71_axes_0"), val = tensor([-1])]; + tensor norm5_2_weight_to_fp16 = const()[name = tensor("norm5_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(135398336)))]; + tensor norm5_2_bias_to_fp16 = const()[name = tensor("norm5_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(135400448)))]; + tensor var_2503_to_fp16 = const()[name = tensor("op_2503_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_71_cast_fp16 = layer_norm(axes = input_71_axes_0, beta = norm5_2_bias_to_fp16, epsilon = var_2503_to_fp16, gamma = norm5_2_weight_to_fp16, x = input_69_cast_fp16)[name = tensor("input_71_cast_fp16")]; + tensor linear5_1_weight_to_fp16 = const()[name = tensor("linear5_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(135402560)))]; + tensor linear_23_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = linear5_1_weight_to_fp16, x = input_71_cast_fp16)[name = tensor("linear_23_cast_fp16")]; + tensor input_73_mode_0 = const()[name = tensor("input_73_mode_0"), val = tensor("EXACT")]; + tensor input_73_cast_fp16 = gelu(mode = input_73_mode_0, x = linear_23_cast_fp16)[name = tensor("input_73_cast_fp16")]; + tensor linear5_2_weight_to_fp16 = const()[name = tensor("linear5_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(143791232)))]; + tensor linear_24_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = linear5_2_weight_to_fp16, x = input_73_cast_fp16)[name = tensor("linear_24_cast_fp16")]; + tensor input_75_cast_fp16 = add(x = input_69_cast_fp16, y = linear_24_cast_fp16)[name = tensor("input_75_cast_fp16")]; + tensor input_77_axes_0 = const()[name = tensor("input_77_axes_0"), val = tensor([-1])]; + tensor norm6_1_weight_to_fp16 = const()[name = tensor("norm6_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(152179904)))]; + tensor norm6_1_bias_to_fp16 = const()[name = tensor("norm6_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(152182016)))]; + tensor var_2520_to_fp16 = const()[name = tensor("op_2520_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_77_cast_fp16 = layer_norm(axes = input_77_axes_0, beta = norm6_1_bias_to_fp16, epsilon = var_2520_to_fp16, gamma = norm6_1_weight_to_fp16, x = input_75_cast_fp16)[name = tensor("input_77_cast_fp16")]; + tensor attn6_in_proj_weight_to_fp16 = const()[name = tensor("attn6_in_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(152184128)))]; + tensor linear_25_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = attn6_in_proj_weight_to_fp16, x = input_77_cast_fp16)[name = tensor("linear_25_cast_fp16")]; + tensor var_2534 = const()[name = tensor("op_2534"), val = tensor([1, 1, 3, 16, 64])]; + tensor qkv_13_cast_fp16 = reshape(shape = var_2534, x = linear_25_cast_fp16)[name = tensor("qkv_13_cast_fp16")]; + tensor q_37_begin_0 = const()[name = tensor("q_37_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor q_37_end_0 = const()[name = tensor("q_37_end_0"), val = tensor([1, 1, 1, 16, 64])]; + tensor q_37_end_mask_0 = const()[name = tensor("q_37_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor q_37_squeeze_mask_0 = const()[name = tensor("q_37_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor q_37_cast_fp16 = slice_by_index(begin = q_37_begin_0, end = q_37_end_0, end_mask = q_37_end_mask_0, squeeze_mask = q_37_squeeze_mask_0, x = qkv_13_cast_fp16)[name = tensor("q_37_cast_fp16")]; + tensor k_25_begin_0 = const()[name = tensor("k_25_begin_0"), val = tensor([0, 0, 1, 0, 0])]; + tensor k_25_end_0 = const()[name = tensor("k_25_end_0"), val = tensor([1, 1, 2, 16, 64])]; + tensor k_25_end_mask_0 = const()[name = tensor("k_25_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor k_25_squeeze_mask_0 = const()[name = tensor("k_25_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor k_25_cast_fp16 = slice_by_index(begin = k_25_begin_0, end = k_25_end_0, end_mask = k_25_end_mask_0, squeeze_mask = k_25_squeeze_mask_0, x = qkv_13_cast_fp16)[name = tensor("k_25_cast_fp16")]; + tensor v_13_begin_0 = const()[name = tensor("v_13_begin_0"), val = tensor([0, 0, 2, 0, 0])]; + tensor v_13_end_0 = const()[name = tensor("v_13_end_0"), val = tensor([1, 1, 3, 16, 64])]; + tensor v_13_end_mask_0 = const()[name = tensor("v_13_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor v_13_squeeze_mask_0 = const()[name = tensor("v_13_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor v_13_cast_fp16 = slice_by_index(begin = v_13_begin_0, end = v_13_end_0, end_mask = v_13_end_mask_0, squeeze_mask = v_13_squeeze_mask_0, x = qkv_13_cast_fp16)[name = tensor("v_13_cast_fp16")]; + tensor var_2600 = const()[name = tensor("op_2600"), val = tensor([1, 1, 1, 1])]; + tensor position6_to_fp16_dtype_0 = const()[name = tensor("position6_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor position6_to_fp16 = cast(dtype = position6_to_fp16_dtype_0, x = position6)[name = tensor("cast_523")]; + tensor ts_41_cast_fp16 = reshape(shape = var_2600, x = position6_to_fp16)[name = tensor("ts_41_cast_fp16")]; + tensor var_2607 = const()[name = tensor("op_2607"), val = tensor([1, 1, 16, 32, 2])]; + tensor q_complex_13_cast_fp16 = reshape(shape = var_2607, x = q_37_cast_fp16)[name = tensor("q_complex_13_cast_fp16")]; + tensor var_2614 = const()[name = tensor("op_2614"), val = tensor([1, 1, 16, 32, 2])]; + tensor k_complex_13_cast_fp16 = reshape(shape = var_2614, x = k_25_cast_fp16)[name = tensor("k_complex_13_cast_fp16")]; + tensor var_2618_begin_0 = const()[name = tensor("op_2618_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_2618_end_0 = const()[name = tensor("op_2618_end_0"), val = tensor([1, 1, 16, 32, 1])]; + tensor var_2618_end_mask_0 = const()[name = tensor("op_2618_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_2618_squeeze_mask_0 = const()[name = tensor("op_2618_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_2618_cast_fp16 = slice_by_index(begin = var_2618_begin_0, end = var_2618_end_0, end_mask = var_2618_end_mask_0, squeeze_mask = var_2618_squeeze_mask_0, x = q_complex_13_cast_fp16)[name = tensor("op_2618_cast_fp16")]; + tensor var_2626_begin_0 = const()[name = tensor("op_2626_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_2626_end_0 = const()[name = tensor("op_2626_end_0"), val = tensor([1, 1, 16, 32, 2])]; + tensor var_2626_end_mask_0 = const()[name = tensor("op_2626_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_2626_squeeze_mask_0 = const()[name = tensor("op_2626_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_2626_cast_fp16 = slice_by_index(begin = var_2626_begin_0, end = var_2626_end_0, end_mask = var_2626_end_mask_0, squeeze_mask = var_2626_squeeze_mask_0, x = q_complex_13_cast_fp16)[name = tensor("op_2626_cast_fp16")]; + tensor var_2634_begin_0 = const()[name = tensor("op_2634_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_2634_end_0 = const()[name = tensor("op_2634_end_0"), val = tensor([1, 1, 16, 32, 1])]; + tensor var_2634_end_mask_0 = const()[name = tensor("op_2634_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_2634_squeeze_mask_0 = const()[name = tensor("op_2634_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_2634_cast_fp16 = slice_by_index(begin = var_2634_begin_0, end = var_2634_end_0, end_mask = var_2634_end_mask_0, squeeze_mask = var_2634_squeeze_mask_0, x = k_complex_13_cast_fp16)[name = tensor("op_2634_cast_fp16")]; + tensor var_2642_begin_0 = const()[name = tensor("op_2642_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_2642_end_0 = const()[name = tensor("op_2642_end_0"), val = tensor([1, 1, 16, 32, 2])]; + tensor var_2642_end_mask_0 = const()[name = tensor("op_2642_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_2642_squeeze_mask_0 = const()[name = tensor("op_2642_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_2642_cast_fp16 = slice_by_index(begin = var_2642_begin_0, end = var_2642_end_0, end_mask = var_2642_end_mask_0, squeeze_mask = var_2642_squeeze_mask_0, x = k_complex_13_cast_fp16)[name = tensor("op_2642_cast_fp16")]; + tensor freqs_13_to_fp16 = const()[name = tensor("freqs_13_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(158475648)))]; + tensor var_2648_cast_fp16 = mul(x = freqs_13_to_fp16, y = ts_41_cast_fp16)[name = tensor("op_2648_cast_fp16")]; + tensor rotr_13_cast_fp16 = cos(x = var_2648_cast_fp16)[name = tensor("rotr_13_cast_fp16")]; + tensor roti_13_cast_fp16 = sin(x = var_2648_cast_fp16)[name = tensor("roti_13_cast_fp16")]; + tensor var_2652_cast_fp16 = mul(x = var_2618_cast_fp16, y = rotr_13_cast_fp16)[name = tensor("op_2652_cast_fp16")]; + tensor var_2653_cast_fp16 = mul(x = var_2626_cast_fp16, y = roti_13_cast_fp16)[name = tensor("op_2653_cast_fp16")]; + tensor qor_25_cast_fp16 = sub(x = var_2652_cast_fp16, y = var_2653_cast_fp16)[name = tensor("qor_25_cast_fp16")]; + tensor var_2656_cast_fp16 = mul(x = var_2618_cast_fp16, y = roti_13_cast_fp16)[name = tensor("op_2656_cast_fp16")]; + tensor var_2657_cast_fp16 = mul(x = var_2626_cast_fp16, y = rotr_13_cast_fp16)[name = tensor("op_2657_cast_fp16")]; + tensor qoi_25_cast_fp16 = add(x = var_2656_cast_fp16, y = var_2657_cast_fp16)[name = tensor("qoi_25_cast_fp16")]; + tensor var_2660_cast_fp16 = mul(x = var_2634_cast_fp16, y = rotr_13_cast_fp16)[name = tensor("op_2660_cast_fp16")]; + tensor var_2661_cast_fp16 = mul(x = var_2642_cast_fp16, y = roti_13_cast_fp16)[name = tensor("op_2661_cast_fp16")]; + tensor kor_25_cast_fp16 = sub(x = var_2660_cast_fp16, y = var_2661_cast_fp16)[name = tensor("kor_25_cast_fp16")]; + tensor var_2664_cast_fp16 = mul(x = var_2634_cast_fp16, y = roti_13_cast_fp16)[name = tensor("op_2664_cast_fp16")]; + tensor var_2665_cast_fp16 = mul(x = var_2642_cast_fp16, y = rotr_13_cast_fp16)[name = tensor("op_2665_cast_fp16")]; + tensor koi_25_cast_fp16 = add(x = var_2664_cast_fp16, y = var_2665_cast_fp16)[name = tensor("koi_25_cast_fp16")]; + tensor qo_13_axis_0 = const()[name = tensor("qo_13_axis_0"), val = tensor(-1)]; + tensor qo_13_cast_fp16 = stack(axis = qo_13_axis_0, values = (qor_25_cast_fp16, qoi_25_cast_fp16))[name = tensor("qo_13_cast_fp16")]; + tensor ko_13_axis_0 = const()[name = tensor("ko_13_axis_0"), val = tensor(-1)]; + tensor ko_13_cast_fp16 = stack(axis = ko_13_axis_0, values = (kor_25_cast_fp16, koi_25_cast_fp16))[name = tensor("ko_13_cast_fp16")]; + tensor var_2698 = const()[name = tensor("op_2698"), val = tensor([1, 1, 16, 64])]; + tensor q_39_cast_fp16 = reshape(shape = var_2698, x = qo_13_cast_fp16)[name = tensor("q_39_cast_fp16")]; + tensor var_2704 = const()[name = tensor("op_2704"), val = tensor([1, 1, 16, 64])]; + tensor k_27_cast_fp16 = reshape(shape = var_2704, x = ko_13_cast_fp16)[name = tensor("k_27_cast_fp16")]; + tensor _inversed_2729_y_0_to_fp16 = const()[name = tensor("_inversed_2729_y_0_to_fp16"), val = tensor(0x1p-9)]; + tensor _inversed_2729_cast_fp16 = mul(x = ts_41_cast_fp16, y = _inversed_2729_y_0_to_fp16)[name = tensor("_inversed_2729_cast_fp16")]; + tensor var_2730_cast_fp16 = floor(x = _inversed_2729_cast_fp16)[name = tensor("op_2730_cast_fp16")]; + tensor var_2731_to_fp16 = const()[name = tensor("op_2731_to_fp16"), val = tensor(0x1p+9)]; + tensor var_2732_cast_fp16 = mul(x = var_2730_cast_fp16, y = var_2731_to_fp16)[name = tensor("op_2732_cast_fp16")]; + tensor write_indices_float_27_cast_fp16 = sub(x = ts_41_cast_fp16, y = var_2732_cast_fp16)[name = tensor("write_indices_float_27_cast_fp16")]; + tensor var_2739_dtype_0 = const()[name = tensor("op_2739_dtype_0"), val = tensor("int32")]; + tensor write_indices_13_reps_0 = const()[name = tensor("write_indices_13_reps_0"), val = tensor([1, 1, 16, 64])]; + tensor write_indices_float_27_cast_fp16_to_int32 = cast(dtype = var_2739_dtype_0, x = write_indices_float_27_cast_fp16)[name = tensor("cast_522")]; + tensor write_indices_13 = tile(reps = write_indices_13_reps_0, x = write_indices_float_27_cast_fp16_to_int32)[name = tensor("write_indices_13")]; + tensor var_2749_begin_0 = const()[name = tensor("op_2749_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_2749_end_0 = const()[name = tensor("op_2749_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor var_2749_end_mask_0 = const()[name = tensor("op_2749_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_2749_squeeze_mask_0 = const()[name = tensor("op_2749_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor cache6_to_fp16_dtype_0 = const()[name = tensor("cache6_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor cache6_to_fp16 = cast(dtype = cache6_to_fp16_dtype_0, x = cache6)[name = tensor("cast_521")]; + tensor var_2749_cast_fp16 = slice_by_index(begin = var_2749_begin_0, end = var_2749_end_0, end_mask = var_2749_end_mask_0, squeeze_mask = var_2749_squeeze_mask_0, x = cache6_to_fp16)[name = tensor("op_2749_cast_fp16")]; + tensor var_2751_axis_0 = const()[name = tensor("op_2751_axis_0"), val = tensor(1)]; + tensor var_2751_mode_0 = const()[name = tensor("op_2751_mode_0"), val = tensor("update")]; + tensor var_2751_validate_indices_0 = const()[name = tensor("op_2751_validate_indices_0"), val = tensor(false)]; + tensor var_2751_cast_fp16 = scatter_along_axis(axis = var_2751_axis_0, data = var_2749_cast_fp16, indices = write_indices_13, mode = var_2751_mode_0, updates = k_27_cast_fp16, validate_indices = var_2751_validate_indices_0)[name = tensor("op_2751_cast_fp16")]; + tensor concat_44 = const()[name = tensor("concat_44"), val = tensor([0, 0, 0, 0, 0])]; + tensor concat_45 = const()[name = tensor("concat_45"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_13_internal_tensor_assign_1_stride_0 = const()[name = tensor("new_cache_13_internal_tensor_assign_1_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_13_internal_tensor_assign_1_begin_mask_0 = const()[name = tensor("new_cache_13_internal_tensor_assign_1_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_13_internal_tensor_assign_1_end_mask_0 = const()[name = tensor("new_cache_13_internal_tensor_assign_1_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_13_internal_tensor_assign_1_squeeze_mask_0 = const()[name = tensor("new_cache_13_internal_tensor_assign_1_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_60 = const()[name = tensor("shape_60"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_12 = const()[name = tensor("reduce_prod_12"), val = tensor(1048576)]; + tensor range_1d_12_start_0 = const()[name = tensor("range_1d_12_start_0"), val = tensor(0)]; + tensor range_1d_12_step_0 = const()[name = tensor("range_1d_12_step_0"), val = tensor(1)]; + tensor range_1d_12 = range_1d(end = reduce_prod_12, start = range_1d_12_start_0, step = range_1d_12_step_0)[name = tensor("range_1d_12")]; + tensor reshape_60 = reshape(shape = shape_60, x = range_1d_12)[name = tensor("reshape_60")]; + tensor slice_by_index_12 = slice_by_index(begin = concat_44, begin_mask = new_cache_13_internal_tensor_assign_1_begin_mask_0, end = concat_45, end_mask = new_cache_13_internal_tensor_assign_1_end_mask_0, squeeze_mask = new_cache_13_internal_tensor_assign_1_squeeze_mask_0, stride = new_cache_13_internal_tensor_assign_1_stride_0, x = reshape_60)[name = tensor("slice_by_index_12")]; + tensor reshape_61_shape_0 = const()[name = tensor("reshape_61_shape_0"), val = tensor([-1])]; + tensor reshape_61 = reshape(shape = reshape_61_shape_0, x = slice_by_index_12)[name = tensor("reshape_61")]; + tensor reshape_62_shape_0 = const()[name = tensor("reshape_62_shape_0"), val = tensor([-1])]; + tensor reshape_62_cast_fp16 = reshape(shape = reshape_62_shape_0, x = var_2751_cast_fp16)[name = tensor("reshape_62_cast_fp16")]; + tensor reshape_63_shape_0 = const()[name = tensor("reshape_63_shape_0"), val = tensor([-1])]; + tensor reshape_63_cast_fp16 = reshape(shape = reshape_63_shape_0, x = cache6_to_fp16)[name = tensor("reshape_63_cast_fp16")]; + tensor scatter_12_mode_0 = const()[name = tensor("scatter_12_mode_0"), val = tensor("update")]; + tensor scatter_12_axis_0 = const()[name = tensor("scatter_12_axis_0"), val = tensor(0)]; + tensor scatter_12_validate_indices_0 = const()[name = tensor("scatter_12_validate_indices_0"), val = tensor(false)]; + tensor scatter_12_cast_fp16 = scatter(axis = scatter_12_axis_0, data = reshape_63_cast_fp16, indices = reshape_61, mode = scatter_12_mode_0, updates = reshape_62_cast_fp16, validate_indices = scatter_12_validate_indices_0)[name = tensor("scatter_12_cast_fp16")]; + tensor reshape_64_cast_fp16 = reshape(shape = shape_60, x = scatter_12_cast_fp16)[name = tensor("reshape_64_cast_fp16")]; + tensor var_2759_begin_0 = const()[name = tensor("op_2759_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor var_2759_end_0 = const()[name = tensor("op_2759_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor var_2759_end_mask_0 = const()[name = tensor("op_2759_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_2759_squeeze_mask_0 = const()[name = tensor("op_2759_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor var_2759_cast_fp16 = slice_by_index(begin = var_2759_begin_0, end = var_2759_end_0, end_mask = var_2759_end_mask_0, squeeze_mask = var_2759_squeeze_mask_0, x = reshape_64_cast_fp16)[name = tensor("op_2759_cast_fp16")]; + tensor var_2761_axis_0 = const()[name = tensor("op_2761_axis_0"), val = tensor(1)]; + tensor var_2761_mode_0 = const()[name = tensor("op_2761_mode_0"), val = tensor("update")]; + tensor var_2761_validate_indices_0 = const()[name = tensor("op_2761_validate_indices_0"), val = tensor(false)]; + tensor var_2761_cast_fp16 = scatter_along_axis(axis = var_2761_axis_0, data = var_2759_cast_fp16, indices = write_indices_13, mode = var_2761_mode_0, updates = v_13_cast_fp16, validate_indices = var_2761_validate_indices_0)[name = tensor("op_2761_cast_fp16")]; + tensor concat_46 = const()[name = tensor("concat_46"), val = tensor([1, 0, 0, 0, 0])]; + tensor concat_47 = const()[name = tensor("concat_47"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_13_internal_tensor_assign_2_stride_0 = const()[name = tensor("new_cache_13_internal_tensor_assign_2_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_13_internal_tensor_assign_2_begin_mask_0 = const()[name = tensor("new_cache_13_internal_tensor_assign_2_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_13_internal_tensor_assign_2_end_mask_0 = const()[name = tensor("new_cache_13_internal_tensor_assign_2_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_13_internal_tensor_assign_2_squeeze_mask_0 = const()[name = tensor("new_cache_13_internal_tensor_assign_2_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_61 = const()[name = tensor("shape_61"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_13 = const()[name = tensor("reduce_prod_13"), val = tensor(1048576)]; + tensor range_1d_13_start_0 = const()[name = tensor("range_1d_13_start_0"), val = tensor(0)]; + tensor range_1d_13_step_0 = const()[name = tensor("range_1d_13_step_0"), val = tensor(1)]; + tensor range_1d_13 = range_1d(end = reduce_prod_13, start = range_1d_13_start_0, step = range_1d_13_step_0)[name = tensor("range_1d_13")]; + tensor reshape_65 = reshape(shape = shape_61, x = range_1d_13)[name = tensor("reshape_65")]; + tensor slice_by_index_13 = slice_by_index(begin = concat_46, begin_mask = new_cache_13_internal_tensor_assign_2_begin_mask_0, end = concat_47, end_mask = new_cache_13_internal_tensor_assign_2_end_mask_0, squeeze_mask = new_cache_13_internal_tensor_assign_2_squeeze_mask_0, stride = new_cache_13_internal_tensor_assign_2_stride_0, x = reshape_65)[name = tensor("slice_by_index_13")]; + tensor reshape_66_shape_0 = const()[name = tensor("reshape_66_shape_0"), val = tensor([-1])]; + tensor reshape_66 = reshape(shape = reshape_66_shape_0, x = slice_by_index_13)[name = tensor("reshape_66")]; + tensor reshape_67_shape_0 = const()[name = tensor("reshape_67_shape_0"), val = tensor([-1])]; + tensor reshape_67_cast_fp16 = reshape(shape = reshape_67_shape_0, x = var_2761_cast_fp16)[name = tensor("reshape_67_cast_fp16")]; + tensor reshape_68_shape_0 = const()[name = tensor("reshape_68_shape_0"), val = tensor([-1])]; + tensor reshape_68_cast_fp16 = reshape(shape = reshape_68_shape_0, x = reshape_64_cast_fp16)[name = tensor("reshape_68_cast_fp16")]; + tensor scatter_13_mode_0 = const()[name = tensor("scatter_13_mode_0"), val = tensor("update")]; + tensor scatter_13_axis_0 = const()[name = tensor("scatter_13_axis_0"), val = tensor(0)]; + tensor scatter_13_validate_indices_0 = const()[name = tensor("scatter_13_validate_indices_0"), val = tensor(false)]; + tensor scatter_13_cast_fp16 = scatter(axis = scatter_13_axis_0, data = reshape_68_cast_fp16, indices = reshape_66, mode = scatter_13_mode_0, updates = reshape_67_cast_fp16, validate_indices = scatter_13_validate_indices_0)[name = tensor("scatter_13_cast_fp16")]; + tensor reshape_69_cast_fp16 = reshape(shape = shape_61, x = scatter_13_cast_fp16)[name = tensor("reshape_69_cast_fp16")]; + tensor reshape_69_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("reshape_69_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor keys_37_begin_0 = const()[name = tensor("keys_37_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor keys_37_end_0 = const()[name = tensor("keys_37_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor keys_37_end_mask_0 = const()[name = tensor("keys_37_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor keys_37_squeeze_mask_0 = const()[name = tensor("keys_37_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor keys_37_cast_fp16 = slice_by_index(begin = keys_37_begin_0, end = keys_37_end_0, end_mask = keys_37_end_mask_0, squeeze_mask = keys_37_squeeze_mask_0, x = reshape_69_cast_fp16)[name = tensor("keys_37_cast_fp16")]; + tensor values_37_begin_0 = const()[name = tensor("values_37_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor values_37_end_0 = const()[name = tensor("values_37_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor values_37_end_mask_0 = const()[name = tensor("values_37_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor values_37_squeeze_mask_0 = const()[name = tensor("values_37_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor values_37_cast_fp16 = slice_by_index(begin = values_37_begin_0, end = values_37_end_0, end_mask = values_37_end_mask_0, squeeze_mask = values_37_squeeze_mask_0, x = reshape_69_cast_fp16)[name = tensor("values_37_cast_fp16")]; + tensor var_2773_cast_fp16 = not_equal(x = keys_37_cast_fp16, y = keys_37_cast_fp16)[name = tensor("op_2773_cast_fp16")]; + tensor keys_39_cast_fp16 = select(a = var_463_to_fp16, b = keys_37_cast_fp16, cond = var_2773_cast_fp16)[name = tensor("keys_39_cast_fp16")]; + tensor var_2781_cast_fp16 = not_equal(x = values_37_cast_fp16, y = values_37_cast_fp16)[name = tensor("op_2781_cast_fp16")]; + tensor values_39_cast_fp16 = select(a = var_463_to_fp16, b = values_37_cast_fp16, cond = var_2781_cast_fp16)[name = tensor("values_39_cast_fp16")]; + tensor var_2805 = const()[name = tensor("op_2805"), val = tensor([0, 2, 1, 3])]; + tensor var_2821 = const()[name = tensor("op_2821"), val = tensor([1, 1, 1])]; + tensor var_2822_cast_fp16 = reshape(shape = var_2821, x = position6_to_fp16)[name = tensor("op_2822_cast_fp16")]; + tensor var_2842_to_fp16 = const()[name = tensor("op_2842_to_fp16"), val = tensor(0x1p+0)]; + tensor valid_len_13_cast_fp16 = add(x = var_2822_cast_fp16, y = var_2842_to_fp16)[name = tensor("valid_len_13_cast_fp16")]; + tensor valid_mask_13_cast_fp16 = less(x = k_positions_1_promoted_to_fp16, y = valid_len_13_cast_fp16)[name = tensor("valid_mask_13_cast_fp16")]; + tensor causal_mask_13_cast_fp16 = less_equal(x = k_positions_1_promoted_to_fp16, y = var_2822_cast_fp16)[name = tensor("causal_mask_13_cast_fp16")]; + tensor attn_mask_25 = logical_and(x = valid_mask_13_cast_fp16, y = causal_mask_13_cast_fp16)[name = tensor("attn_mask_25")]; + tensor attn_mask_27_axes_0 = const()[name = tensor("attn_mask_27_axes_0"), val = tensor([1])]; + tensor attn_mask_27 = expand_dims(axes = attn_mask_27_axes_0, x = attn_mask_25)[name = tensor("attn_mask_27")]; + tensor var_2860_transpose_x_0 = const()[name = tensor("op_2860_transpose_x_0"), val = tensor(false)]; + tensor var_2860_transpose_y_0 = const()[name = tensor("op_2860_transpose_y_0"), val = tensor(false)]; + tensor transpose_84_perm_0 = const()[name = tensor("transpose_84_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_85_perm_0 = const()[name = tensor("transpose_85_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_85 = transpose(perm = transpose_85_perm_0, x = keys_39_cast_fp16)[name = tensor("transpose_189")]; + tensor transpose_84 = transpose(perm = transpose_84_perm_0, x = q_39_cast_fp16)[name = tensor("transpose_190")]; + tensor var_2860_cast_fp16 = matmul(transpose_x = var_2860_transpose_x_0, transpose_y = var_2860_transpose_y_0, x = transpose_84, y = transpose_85)[name = tensor("op_2860_cast_fp16")]; + tensor var_2854_to_fp16 = const()[name = tensor("op_2854_to_fp16"), val = tensor([0x1p-3])]; + tensor attn_weights_37_cast_fp16 = mul(x = var_2860_cast_fp16, y = var_2854_to_fp16)[name = tensor("attn_weights_37_cast_fp16")]; + tensor var_2862 = logical_not(x = attn_mask_27)[name = tensor("op_2862")]; + tensor var_2863_to_fp16 = const()[name = tensor("op_2863_to_fp16"), val = tensor(-inf)]; + tensor attn_weights_39_cast_fp16 = select(a = var_2863_to_fp16, b = attn_weights_37_cast_fp16, cond = var_2862)[name = tensor("attn_weights_39_cast_fp16")]; + tensor var_2865 = const()[name = tensor("op_2865"), val = tensor(-1)]; + tensor attn_weights_41_cast_fp16 = softmax(axis = var_2865, x = attn_weights_39_cast_fp16)[name = tensor("attn_weights_41_cast_fp16")]; + tensor attn_output_13_transpose_x_0 = const()[name = tensor("attn_output_13_transpose_x_0"), val = tensor(false)]; + tensor attn_output_13_transpose_y_0 = const()[name = tensor("attn_output_13_transpose_y_0"), val = tensor(false)]; + tensor values_41_cast_fp16 = transpose(perm = var_2805, x = values_39_cast_fp16)[name = tensor("transpose_191")]; + tensor attn_output_13_cast_fp16 = matmul(transpose_x = attn_output_13_transpose_x_0, transpose_y = attn_output_13_transpose_y_0, x = attn_weights_41_cast_fp16, y = values_41_cast_fp16)[name = tensor("attn_output_13_cast_fp16")]; + tensor var_2873 = const()[name = tensor("op_2873"), val = tensor([0, 2, 1, 3])]; + tensor var_2878 = const()[name = tensor("op_2878"), val = tensor([1, 1, 1024])]; + tensor var_2874_cast_fp16 = transpose(perm = var_2873, x = attn_output_13_cast_fp16)[name = tensor("transpose_188")]; + tensor input_79_cast_fp16 = reshape(shape = var_2878, x = var_2874_cast_fp16)[name = tensor("input_79_cast_fp16")]; + tensor attn6_out_proj_weight_to_fp16 = const()[name = tensor("attn6_out_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(158475776)))]; + tensor linear_26_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = attn6_out_proj_weight_to_fp16, x = input_79_cast_fp16)[name = tensor("linear_26_cast_fp16")]; + tensor var_2884_to_fp16 = const()[name = tensor("op_2884_to_fp16"), val = tensor(0x1p+0)]; + tensor var_2885_cast_fp16 = add(x = position6_to_fp16, y = var_2884_to_fp16)[name = tensor("op_2885_cast_fp16")]; + tensor var_2885_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_2885_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor input_81_cast_fp16 = add(x = input_75_cast_fp16, y = linear_26_cast_fp16)[name = tensor("input_81_cast_fp16")]; + tensor input_83_axes_0 = const()[name = tensor("input_83_axes_0"), val = tensor([-1])]; + tensor norm6_2_weight_to_fp16 = const()[name = tensor("norm6_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(160572992)))]; + tensor norm6_2_bias_to_fp16 = const()[name = tensor("norm6_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(160575104)))]; + tensor var_2889_to_fp16 = const()[name = tensor("op_2889_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_83_cast_fp16 = layer_norm(axes = input_83_axes_0, beta = norm6_2_bias_to_fp16, epsilon = var_2889_to_fp16, gamma = norm6_2_weight_to_fp16, x = input_81_cast_fp16)[name = tensor("input_83_cast_fp16")]; + tensor linear6_1_weight_to_fp16 = const()[name = tensor("linear6_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(160577216)))]; + tensor linear_27_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = linear6_1_weight_to_fp16, x = input_83_cast_fp16)[name = tensor("linear_27_cast_fp16")]; + tensor input_85_mode_0 = const()[name = tensor("input_85_mode_0"), val = tensor("EXACT")]; + tensor input_85_cast_fp16 = gelu(mode = input_85_mode_0, x = linear_27_cast_fp16)[name = tensor("input_85_cast_fp16")]; + tensor linear6_2_weight_to_fp16 = const()[name = tensor("linear6_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(168965888)))]; + tensor linear_28_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = linear6_2_weight_to_fp16, x = input_85_cast_fp16)[name = tensor("linear_28_cast_fp16")]; + tensor input_87_cast_fp16 = add(x = input_81_cast_fp16, y = linear_28_cast_fp16)[name = tensor("input_87_cast_fp16")]; + tensor input_89_axes_0 = const()[name = tensor("input_89_axes_0"), val = tensor([-1])]; + tensor norm7_1_weight_to_fp16 = const()[name = tensor("norm7_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(177354560)))]; + tensor norm7_1_bias_to_fp16 = const()[name = tensor("norm7_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(177356672)))]; + tensor var_2906_to_fp16 = const()[name = tensor("op_2906_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_89_cast_fp16 = layer_norm(axes = input_89_axes_0, beta = norm7_1_bias_to_fp16, epsilon = var_2906_to_fp16, gamma = norm7_1_weight_to_fp16, x = input_87_cast_fp16)[name = tensor("input_89_cast_fp16")]; + tensor attn7_in_proj_weight_to_fp16 = const()[name = tensor("attn7_in_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(177358784)))]; + tensor linear_29_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = attn7_in_proj_weight_to_fp16, x = input_89_cast_fp16)[name = tensor("linear_29_cast_fp16")]; + tensor var_2920 = const()[name = tensor("op_2920"), val = tensor([1, 1, 3, 16, 64])]; + tensor qkv_15_cast_fp16 = reshape(shape = var_2920, x = linear_29_cast_fp16)[name = tensor("qkv_15_cast_fp16")]; + tensor q_43_begin_0 = const()[name = tensor("q_43_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor q_43_end_0 = const()[name = tensor("q_43_end_0"), val = tensor([1, 1, 1, 16, 64])]; + tensor q_43_end_mask_0 = const()[name = tensor("q_43_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor q_43_squeeze_mask_0 = const()[name = tensor("q_43_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor q_43_cast_fp16 = slice_by_index(begin = q_43_begin_0, end = q_43_end_0, end_mask = q_43_end_mask_0, squeeze_mask = q_43_squeeze_mask_0, x = qkv_15_cast_fp16)[name = tensor("q_43_cast_fp16")]; + tensor k_29_begin_0 = const()[name = tensor("k_29_begin_0"), val = tensor([0, 0, 1, 0, 0])]; + tensor k_29_end_0 = const()[name = tensor("k_29_end_0"), val = tensor([1, 1, 2, 16, 64])]; + tensor k_29_end_mask_0 = const()[name = tensor("k_29_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor k_29_squeeze_mask_0 = const()[name = tensor("k_29_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor k_29_cast_fp16 = slice_by_index(begin = k_29_begin_0, end = k_29_end_0, end_mask = k_29_end_mask_0, squeeze_mask = k_29_squeeze_mask_0, x = qkv_15_cast_fp16)[name = tensor("k_29_cast_fp16")]; + tensor v_15_begin_0 = const()[name = tensor("v_15_begin_0"), val = tensor([0, 0, 2, 0, 0])]; + tensor v_15_end_0 = const()[name = tensor("v_15_end_0"), val = tensor([1, 1, 3, 16, 64])]; + tensor v_15_end_mask_0 = const()[name = tensor("v_15_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor v_15_squeeze_mask_0 = const()[name = tensor("v_15_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor v_15_cast_fp16 = slice_by_index(begin = v_15_begin_0, end = v_15_end_0, end_mask = v_15_end_mask_0, squeeze_mask = v_15_squeeze_mask_0, x = qkv_15_cast_fp16)[name = tensor("v_15_cast_fp16")]; + tensor var_2986 = const()[name = tensor("op_2986"), val = tensor([1, 1, 1, 1])]; + tensor position7_to_fp16_dtype_0 = const()[name = tensor("position7_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor position7_to_fp16 = cast(dtype = position7_to_fp16_dtype_0, x = position7)[name = tensor("cast_518")]; + tensor ts_47_cast_fp16 = reshape(shape = var_2986, x = position7_to_fp16)[name = tensor("ts_47_cast_fp16")]; + tensor var_2993 = const()[name = tensor("op_2993"), val = tensor([1, 1, 16, 32, 2])]; + tensor q_complex_15_cast_fp16 = reshape(shape = var_2993, x = q_43_cast_fp16)[name = tensor("q_complex_15_cast_fp16")]; + tensor var_3000 = const()[name = tensor("op_3000"), val = tensor([1, 1, 16, 32, 2])]; + tensor k_complex_15_cast_fp16 = reshape(shape = var_3000, x = k_29_cast_fp16)[name = tensor("k_complex_15_cast_fp16")]; + tensor var_3004_begin_0 = const()[name = tensor("op_3004_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_3004_end_0 = const()[name = tensor("op_3004_end_0"), val = tensor([1, 1, 16, 32, 1])]; + tensor var_3004_end_mask_0 = const()[name = tensor("op_3004_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_3004_squeeze_mask_0 = const()[name = tensor("op_3004_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_3004_cast_fp16 = slice_by_index(begin = var_3004_begin_0, end = var_3004_end_0, end_mask = var_3004_end_mask_0, squeeze_mask = var_3004_squeeze_mask_0, x = q_complex_15_cast_fp16)[name = tensor("op_3004_cast_fp16")]; + tensor var_3012_begin_0 = const()[name = tensor("op_3012_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_3012_end_0 = const()[name = tensor("op_3012_end_0"), val = tensor([1, 1, 16, 32, 2])]; + tensor var_3012_end_mask_0 = const()[name = tensor("op_3012_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_3012_squeeze_mask_0 = const()[name = tensor("op_3012_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_3012_cast_fp16 = slice_by_index(begin = var_3012_begin_0, end = var_3012_end_0, end_mask = var_3012_end_mask_0, squeeze_mask = var_3012_squeeze_mask_0, x = q_complex_15_cast_fp16)[name = tensor("op_3012_cast_fp16")]; + tensor var_3020_begin_0 = const()[name = tensor("op_3020_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_3020_end_0 = const()[name = tensor("op_3020_end_0"), val = tensor([1, 1, 16, 32, 1])]; + tensor var_3020_end_mask_0 = const()[name = tensor("op_3020_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_3020_squeeze_mask_0 = const()[name = tensor("op_3020_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_3020_cast_fp16 = slice_by_index(begin = var_3020_begin_0, end = var_3020_end_0, end_mask = var_3020_end_mask_0, squeeze_mask = var_3020_squeeze_mask_0, x = k_complex_15_cast_fp16)[name = tensor("op_3020_cast_fp16")]; + tensor var_3028_begin_0 = const()[name = tensor("op_3028_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_3028_end_0 = const()[name = tensor("op_3028_end_0"), val = tensor([1, 1, 16, 32, 2])]; + tensor var_3028_end_mask_0 = const()[name = tensor("op_3028_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_3028_squeeze_mask_0 = const()[name = tensor("op_3028_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_3028_cast_fp16 = slice_by_index(begin = var_3028_begin_0, end = var_3028_end_0, end_mask = var_3028_end_mask_0, squeeze_mask = var_3028_squeeze_mask_0, x = k_complex_15_cast_fp16)[name = tensor("op_3028_cast_fp16")]; + tensor freqs_15_to_fp16 = const()[name = tensor("freqs_15_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(183650304)))]; + tensor var_3034_cast_fp16 = mul(x = freqs_15_to_fp16, y = ts_47_cast_fp16)[name = tensor("op_3034_cast_fp16")]; + tensor rotr_15_cast_fp16 = cos(x = var_3034_cast_fp16)[name = tensor("rotr_15_cast_fp16")]; + tensor roti_15_cast_fp16 = sin(x = var_3034_cast_fp16)[name = tensor("roti_15_cast_fp16")]; + tensor var_3038_cast_fp16 = mul(x = var_3004_cast_fp16, y = rotr_15_cast_fp16)[name = tensor("op_3038_cast_fp16")]; + tensor var_3039_cast_fp16 = mul(x = var_3012_cast_fp16, y = roti_15_cast_fp16)[name = tensor("op_3039_cast_fp16")]; + tensor qor_29_cast_fp16 = sub(x = var_3038_cast_fp16, y = var_3039_cast_fp16)[name = tensor("qor_29_cast_fp16")]; + tensor var_3042_cast_fp16 = mul(x = var_3004_cast_fp16, y = roti_15_cast_fp16)[name = tensor("op_3042_cast_fp16")]; + tensor var_3043_cast_fp16 = mul(x = var_3012_cast_fp16, y = rotr_15_cast_fp16)[name = tensor("op_3043_cast_fp16")]; + tensor qoi_29_cast_fp16 = add(x = var_3042_cast_fp16, y = var_3043_cast_fp16)[name = tensor("qoi_29_cast_fp16")]; + tensor var_3046_cast_fp16 = mul(x = var_3020_cast_fp16, y = rotr_15_cast_fp16)[name = tensor("op_3046_cast_fp16")]; + tensor var_3047_cast_fp16 = mul(x = var_3028_cast_fp16, y = roti_15_cast_fp16)[name = tensor("op_3047_cast_fp16")]; + tensor kor_29_cast_fp16 = sub(x = var_3046_cast_fp16, y = var_3047_cast_fp16)[name = tensor("kor_29_cast_fp16")]; + tensor var_3050_cast_fp16 = mul(x = var_3020_cast_fp16, y = roti_15_cast_fp16)[name = tensor("op_3050_cast_fp16")]; + tensor var_3051_cast_fp16 = mul(x = var_3028_cast_fp16, y = rotr_15_cast_fp16)[name = tensor("op_3051_cast_fp16")]; + tensor koi_29_cast_fp16 = add(x = var_3050_cast_fp16, y = var_3051_cast_fp16)[name = tensor("koi_29_cast_fp16")]; + tensor qo_15_axis_0 = const()[name = tensor("qo_15_axis_0"), val = tensor(-1)]; + tensor qo_15_cast_fp16 = stack(axis = qo_15_axis_0, values = (qor_29_cast_fp16, qoi_29_cast_fp16))[name = tensor("qo_15_cast_fp16")]; + tensor ko_15_axis_0 = const()[name = tensor("ko_15_axis_0"), val = tensor(-1)]; + tensor ko_15_cast_fp16 = stack(axis = ko_15_axis_0, values = (kor_29_cast_fp16, koi_29_cast_fp16))[name = tensor("ko_15_cast_fp16")]; + tensor var_3084 = const()[name = tensor("op_3084"), val = tensor([1, 1, 16, 64])]; + tensor q_45_cast_fp16 = reshape(shape = var_3084, x = qo_15_cast_fp16)[name = tensor("q_45_cast_fp16")]; + tensor var_3090 = const()[name = tensor("op_3090"), val = tensor([1, 1, 16, 64])]; + tensor k_31_cast_fp16 = reshape(shape = var_3090, x = ko_15_cast_fp16)[name = tensor("k_31_cast_fp16")]; + tensor _inversed_3115_y_0_to_fp16 = const()[name = tensor("_inversed_3115_y_0_to_fp16"), val = tensor(0x1p-9)]; + tensor _inversed_3115_cast_fp16 = mul(x = ts_47_cast_fp16, y = _inversed_3115_y_0_to_fp16)[name = tensor("_inversed_3115_cast_fp16")]; + tensor var_3116_cast_fp16 = floor(x = _inversed_3115_cast_fp16)[name = tensor("op_3116_cast_fp16")]; + tensor var_3117_to_fp16 = const()[name = tensor("op_3117_to_fp16"), val = tensor(0x1p+9)]; + tensor var_3118_cast_fp16 = mul(x = var_3116_cast_fp16, y = var_3117_to_fp16)[name = tensor("op_3118_cast_fp16")]; + tensor write_indices_float_31_cast_fp16 = sub(x = ts_47_cast_fp16, y = var_3118_cast_fp16)[name = tensor("write_indices_float_31_cast_fp16")]; + tensor var_3125_dtype_0 = const()[name = tensor("op_3125_dtype_0"), val = tensor("int32")]; + tensor write_indices_15_reps_0 = const()[name = tensor("write_indices_15_reps_0"), val = tensor([1, 1, 16, 64])]; + tensor write_indices_float_31_cast_fp16_to_int32 = cast(dtype = var_3125_dtype_0, x = write_indices_float_31_cast_fp16)[name = tensor("cast_517")]; + tensor write_indices_15 = tile(reps = write_indices_15_reps_0, x = write_indices_float_31_cast_fp16_to_int32)[name = tensor("write_indices_15")]; + tensor var_3135_begin_0 = const()[name = tensor("op_3135_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_3135_end_0 = const()[name = tensor("op_3135_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor var_3135_end_mask_0 = const()[name = tensor("op_3135_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_3135_squeeze_mask_0 = const()[name = tensor("op_3135_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor cache7_to_fp16_dtype_0 = const()[name = tensor("cache7_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor cache7_to_fp16 = cast(dtype = cache7_to_fp16_dtype_0, x = cache7)[name = tensor("cast_516")]; + tensor var_3135_cast_fp16 = slice_by_index(begin = var_3135_begin_0, end = var_3135_end_0, end_mask = var_3135_end_mask_0, squeeze_mask = var_3135_squeeze_mask_0, x = cache7_to_fp16)[name = tensor("op_3135_cast_fp16")]; + tensor var_3137_axis_0 = const()[name = tensor("op_3137_axis_0"), val = tensor(1)]; + tensor var_3137_mode_0 = const()[name = tensor("op_3137_mode_0"), val = tensor("update")]; + tensor var_3137_validate_indices_0 = const()[name = tensor("op_3137_validate_indices_0"), val = tensor(false)]; + tensor var_3137_cast_fp16 = scatter_along_axis(axis = var_3137_axis_0, data = var_3135_cast_fp16, indices = write_indices_15, mode = var_3137_mode_0, updates = k_31_cast_fp16, validate_indices = var_3137_validate_indices_0)[name = tensor("op_3137_cast_fp16")]; + tensor concat_51 = const()[name = tensor("concat_51"), val = tensor([0, 0, 0, 0, 0])]; + tensor concat_52 = const()[name = tensor("concat_52"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_15_internal_tensor_assign_1_stride_0 = const()[name = tensor("new_cache_15_internal_tensor_assign_1_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_15_internal_tensor_assign_1_begin_mask_0 = const()[name = tensor("new_cache_15_internal_tensor_assign_1_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_15_internal_tensor_assign_1_end_mask_0 = const()[name = tensor("new_cache_15_internal_tensor_assign_1_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_15_internal_tensor_assign_1_squeeze_mask_0 = const()[name = tensor("new_cache_15_internal_tensor_assign_1_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_62 = const()[name = tensor("shape_62"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_14 = const()[name = tensor("reduce_prod_14"), val = tensor(1048576)]; + tensor range_1d_14_start_0 = const()[name = tensor("range_1d_14_start_0"), val = tensor(0)]; + tensor range_1d_14_step_0 = const()[name = tensor("range_1d_14_step_0"), val = tensor(1)]; + tensor range_1d_14 = range_1d(end = reduce_prod_14, start = range_1d_14_start_0, step = range_1d_14_step_0)[name = tensor("range_1d_14")]; + tensor reshape_70 = reshape(shape = shape_62, x = range_1d_14)[name = tensor("reshape_70")]; + tensor slice_by_index_14 = slice_by_index(begin = concat_51, begin_mask = new_cache_15_internal_tensor_assign_1_begin_mask_0, end = concat_52, end_mask = new_cache_15_internal_tensor_assign_1_end_mask_0, squeeze_mask = new_cache_15_internal_tensor_assign_1_squeeze_mask_0, stride = new_cache_15_internal_tensor_assign_1_stride_0, x = reshape_70)[name = tensor("slice_by_index_14")]; + tensor reshape_71_shape_0 = const()[name = tensor("reshape_71_shape_0"), val = tensor([-1])]; + tensor reshape_71 = reshape(shape = reshape_71_shape_0, x = slice_by_index_14)[name = tensor("reshape_71")]; + tensor reshape_72_shape_0 = const()[name = tensor("reshape_72_shape_0"), val = tensor([-1])]; + tensor reshape_72_cast_fp16 = reshape(shape = reshape_72_shape_0, x = var_3137_cast_fp16)[name = tensor("reshape_72_cast_fp16")]; + tensor reshape_73_shape_0 = const()[name = tensor("reshape_73_shape_0"), val = tensor([-1])]; + tensor reshape_73_cast_fp16 = reshape(shape = reshape_73_shape_0, x = cache7_to_fp16)[name = tensor("reshape_73_cast_fp16")]; + tensor scatter_14_mode_0 = const()[name = tensor("scatter_14_mode_0"), val = tensor("update")]; + tensor scatter_14_axis_0 = const()[name = tensor("scatter_14_axis_0"), val = tensor(0)]; + tensor scatter_14_validate_indices_0 = const()[name = tensor("scatter_14_validate_indices_0"), val = tensor(false)]; + tensor scatter_14_cast_fp16 = scatter(axis = scatter_14_axis_0, data = reshape_73_cast_fp16, indices = reshape_71, mode = scatter_14_mode_0, updates = reshape_72_cast_fp16, validate_indices = scatter_14_validate_indices_0)[name = tensor("scatter_14_cast_fp16")]; + tensor reshape_74_cast_fp16 = reshape(shape = shape_62, x = scatter_14_cast_fp16)[name = tensor("reshape_74_cast_fp16")]; + tensor var_3145_begin_0 = const()[name = tensor("op_3145_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor var_3145_end_0 = const()[name = tensor("op_3145_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor var_3145_end_mask_0 = const()[name = tensor("op_3145_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_3145_squeeze_mask_0 = const()[name = tensor("op_3145_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor var_3145_cast_fp16 = slice_by_index(begin = var_3145_begin_0, end = var_3145_end_0, end_mask = var_3145_end_mask_0, squeeze_mask = var_3145_squeeze_mask_0, x = reshape_74_cast_fp16)[name = tensor("op_3145_cast_fp16")]; + tensor var_3147_axis_0 = const()[name = tensor("op_3147_axis_0"), val = tensor(1)]; + tensor var_3147_mode_0 = const()[name = tensor("op_3147_mode_0"), val = tensor("update")]; + tensor var_3147_validate_indices_0 = const()[name = tensor("op_3147_validate_indices_0"), val = tensor(false)]; + tensor var_3147_cast_fp16 = scatter_along_axis(axis = var_3147_axis_0, data = var_3145_cast_fp16, indices = write_indices_15, mode = var_3147_mode_0, updates = v_15_cast_fp16, validate_indices = var_3147_validate_indices_0)[name = tensor("op_3147_cast_fp16")]; + tensor concat_53 = const()[name = tensor("concat_53"), val = tensor([1, 0, 0, 0, 0])]; + tensor concat_54 = const()[name = tensor("concat_54"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_15_internal_tensor_assign_2_stride_0 = const()[name = tensor("new_cache_15_internal_tensor_assign_2_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_15_internal_tensor_assign_2_begin_mask_0 = const()[name = tensor("new_cache_15_internal_tensor_assign_2_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_15_internal_tensor_assign_2_end_mask_0 = const()[name = tensor("new_cache_15_internal_tensor_assign_2_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_15_internal_tensor_assign_2_squeeze_mask_0 = const()[name = tensor("new_cache_15_internal_tensor_assign_2_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_63 = const()[name = tensor("shape_63"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_15 = const()[name = tensor("reduce_prod_15"), val = tensor(1048576)]; + tensor range_1d_15_start_0 = const()[name = tensor("range_1d_15_start_0"), val = tensor(0)]; + tensor range_1d_15_step_0 = const()[name = tensor("range_1d_15_step_0"), val = tensor(1)]; + tensor range_1d_15 = range_1d(end = reduce_prod_15, start = range_1d_15_start_0, step = range_1d_15_step_0)[name = tensor("range_1d_15")]; + tensor reshape_75 = reshape(shape = shape_63, x = range_1d_15)[name = tensor("reshape_75")]; + tensor slice_by_index_15 = slice_by_index(begin = concat_53, begin_mask = new_cache_15_internal_tensor_assign_2_begin_mask_0, end = concat_54, end_mask = new_cache_15_internal_tensor_assign_2_end_mask_0, squeeze_mask = new_cache_15_internal_tensor_assign_2_squeeze_mask_0, stride = new_cache_15_internal_tensor_assign_2_stride_0, x = reshape_75)[name = tensor("slice_by_index_15")]; + tensor reshape_76_shape_0 = const()[name = tensor("reshape_76_shape_0"), val = tensor([-1])]; + tensor reshape_76 = reshape(shape = reshape_76_shape_0, x = slice_by_index_15)[name = tensor("reshape_76")]; + tensor reshape_77_shape_0 = const()[name = tensor("reshape_77_shape_0"), val = tensor([-1])]; + tensor reshape_77_cast_fp16 = reshape(shape = reshape_77_shape_0, x = var_3147_cast_fp16)[name = tensor("reshape_77_cast_fp16")]; + tensor reshape_78_shape_0 = const()[name = tensor("reshape_78_shape_0"), val = tensor([-1])]; + tensor reshape_78_cast_fp16 = reshape(shape = reshape_78_shape_0, x = reshape_74_cast_fp16)[name = tensor("reshape_78_cast_fp16")]; + tensor scatter_15_mode_0 = const()[name = tensor("scatter_15_mode_0"), val = tensor("update")]; + tensor scatter_15_axis_0 = const()[name = tensor("scatter_15_axis_0"), val = tensor(0)]; + tensor scatter_15_validate_indices_0 = const()[name = tensor("scatter_15_validate_indices_0"), val = tensor(false)]; + tensor scatter_15_cast_fp16 = scatter(axis = scatter_15_axis_0, data = reshape_78_cast_fp16, indices = reshape_76, mode = scatter_15_mode_0, updates = reshape_77_cast_fp16, validate_indices = scatter_15_validate_indices_0)[name = tensor("scatter_15_cast_fp16")]; + tensor reshape_79_cast_fp16 = reshape(shape = shape_63, x = scatter_15_cast_fp16)[name = tensor("reshape_79_cast_fp16")]; + tensor reshape_79_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("reshape_79_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor keys_43_begin_0 = const()[name = tensor("keys_43_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor keys_43_end_0 = const()[name = tensor("keys_43_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor keys_43_end_mask_0 = const()[name = tensor("keys_43_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor keys_43_squeeze_mask_0 = const()[name = tensor("keys_43_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor keys_43_cast_fp16 = slice_by_index(begin = keys_43_begin_0, end = keys_43_end_0, end_mask = keys_43_end_mask_0, squeeze_mask = keys_43_squeeze_mask_0, x = reshape_79_cast_fp16)[name = tensor("keys_43_cast_fp16")]; + tensor values_43_begin_0 = const()[name = tensor("values_43_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor values_43_end_0 = const()[name = tensor("values_43_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor values_43_end_mask_0 = const()[name = tensor("values_43_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor values_43_squeeze_mask_0 = const()[name = tensor("values_43_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor values_43_cast_fp16 = slice_by_index(begin = values_43_begin_0, end = values_43_end_0, end_mask = values_43_end_mask_0, squeeze_mask = values_43_squeeze_mask_0, x = reshape_79_cast_fp16)[name = tensor("values_43_cast_fp16")]; + tensor var_3159_cast_fp16 = not_equal(x = keys_43_cast_fp16, y = keys_43_cast_fp16)[name = tensor("op_3159_cast_fp16")]; + tensor keys_45_cast_fp16 = select(a = var_463_to_fp16, b = keys_43_cast_fp16, cond = var_3159_cast_fp16)[name = tensor("keys_45_cast_fp16")]; + tensor var_3167_cast_fp16 = not_equal(x = values_43_cast_fp16, y = values_43_cast_fp16)[name = tensor("op_3167_cast_fp16")]; + tensor values_45_cast_fp16 = select(a = var_463_to_fp16, b = values_43_cast_fp16, cond = var_3167_cast_fp16)[name = tensor("values_45_cast_fp16")]; + tensor var_3191 = const()[name = tensor("op_3191"), val = tensor([0, 2, 1, 3])]; + tensor var_3207 = const()[name = tensor("op_3207"), val = tensor([1, 1, 1])]; + tensor var_3208_cast_fp16 = reshape(shape = var_3207, x = position7_to_fp16)[name = tensor("op_3208_cast_fp16")]; + tensor var_3228_to_fp16 = const()[name = tensor("op_3228_to_fp16"), val = tensor(0x1p+0)]; + tensor valid_len_15_cast_fp16 = add(x = var_3208_cast_fp16, y = var_3228_to_fp16)[name = tensor("valid_len_15_cast_fp16")]; + tensor valid_mask_15_cast_fp16 = less(x = k_positions_1_promoted_to_fp16, y = valid_len_15_cast_fp16)[name = tensor("valid_mask_15_cast_fp16")]; + tensor causal_mask_15_cast_fp16 = less_equal(x = k_positions_1_promoted_to_fp16, y = var_3208_cast_fp16)[name = tensor("causal_mask_15_cast_fp16")]; + tensor attn_mask_29 = logical_and(x = valid_mask_15_cast_fp16, y = causal_mask_15_cast_fp16)[name = tensor("attn_mask_29")]; + tensor attn_mask_31_axes_0 = const()[name = tensor("attn_mask_31_axes_0"), val = tensor([1])]; + tensor attn_mask_31 = expand_dims(axes = attn_mask_31_axes_0, x = attn_mask_29)[name = tensor("attn_mask_31")]; + tensor var_3246_transpose_x_0 = const()[name = tensor("op_3246_transpose_x_0"), val = tensor(false)]; + tensor var_3246_transpose_y_0 = const()[name = tensor("op_3246_transpose_y_0"), val = tensor(false)]; + tensor transpose_86_perm_0 = const()[name = tensor("transpose_86_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_87_perm_0 = const()[name = tensor("transpose_87_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_87 = transpose(perm = transpose_87_perm_0, x = keys_45_cast_fp16)[name = tensor("transpose_185")]; + tensor transpose_86 = transpose(perm = transpose_86_perm_0, x = q_45_cast_fp16)[name = tensor("transpose_186")]; + tensor var_3246_cast_fp16 = matmul(transpose_x = var_3246_transpose_x_0, transpose_y = var_3246_transpose_y_0, x = transpose_86, y = transpose_87)[name = tensor("op_3246_cast_fp16")]; + tensor var_3240_to_fp16 = const()[name = tensor("op_3240_to_fp16"), val = tensor([0x1p-3])]; + tensor attn_weights_43_cast_fp16 = mul(x = var_3246_cast_fp16, y = var_3240_to_fp16)[name = tensor("attn_weights_43_cast_fp16")]; + tensor var_3248 = logical_not(x = attn_mask_31)[name = tensor("op_3248")]; + tensor var_3249_to_fp16 = const()[name = tensor("op_3249_to_fp16"), val = tensor(-inf)]; + tensor attn_weights_45_cast_fp16 = select(a = var_3249_to_fp16, b = attn_weights_43_cast_fp16, cond = var_3248)[name = tensor("attn_weights_45_cast_fp16")]; + tensor var_3251 = const()[name = tensor("op_3251"), val = tensor(-1)]; + tensor attn_weights_47_cast_fp16 = softmax(axis = var_3251, x = attn_weights_45_cast_fp16)[name = tensor("attn_weights_47_cast_fp16")]; + tensor attn_output_15_transpose_x_0 = const()[name = tensor("attn_output_15_transpose_x_0"), val = tensor(false)]; + tensor attn_output_15_transpose_y_0 = const()[name = tensor("attn_output_15_transpose_y_0"), val = tensor(false)]; + tensor values_47_cast_fp16 = transpose(perm = var_3191, x = values_45_cast_fp16)[name = tensor("transpose_187")]; + tensor attn_output_15_cast_fp16 = matmul(transpose_x = attn_output_15_transpose_x_0, transpose_y = attn_output_15_transpose_y_0, x = attn_weights_47_cast_fp16, y = values_47_cast_fp16)[name = tensor("attn_output_15_cast_fp16")]; + tensor var_3259 = const()[name = tensor("op_3259"), val = tensor([0, 2, 1, 3])]; + tensor var_3264 = const()[name = tensor("op_3264"), val = tensor([1, 1, 1024])]; + tensor var_3260_cast_fp16 = transpose(perm = var_3259, x = attn_output_15_cast_fp16)[name = tensor("transpose_184")]; + tensor input_91_cast_fp16 = reshape(shape = var_3264, x = var_3260_cast_fp16)[name = tensor("input_91_cast_fp16")]; + tensor attn7_out_proj_weight_to_fp16 = const()[name = tensor("attn7_out_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(183650432)))]; + tensor linear_30_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = attn7_out_proj_weight_to_fp16, x = input_91_cast_fp16)[name = tensor("linear_30_cast_fp16")]; + tensor var_3270_to_fp16 = const()[name = tensor("op_3270_to_fp16"), val = tensor(0x1p+0)]; + tensor var_3271_cast_fp16 = add(x = position7_to_fp16, y = var_3270_to_fp16)[name = tensor("op_3271_cast_fp16")]; + tensor var_3271_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_3271_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor input_93_cast_fp16 = add(x = input_87_cast_fp16, y = linear_30_cast_fp16)[name = tensor("input_93_cast_fp16")]; + tensor input_95_axes_0 = const()[name = tensor("input_95_axes_0"), val = tensor([-1])]; + tensor norm7_2_weight_to_fp16 = const()[name = tensor("norm7_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(185747648)))]; + tensor norm7_2_bias_to_fp16 = const()[name = tensor("norm7_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(185749760)))]; + tensor var_3275_to_fp16 = const()[name = tensor("op_3275_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_95_cast_fp16 = layer_norm(axes = input_95_axes_0, beta = norm7_2_bias_to_fp16, epsilon = var_3275_to_fp16, gamma = norm7_2_weight_to_fp16, x = input_93_cast_fp16)[name = tensor("input_95_cast_fp16")]; + tensor linear7_1_weight_to_fp16 = const()[name = tensor("linear7_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(185751872)))]; + tensor linear_31_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = linear7_1_weight_to_fp16, x = input_95_cast_fp16)[name = tensor("linear_31_cast_fp16")]; + tensor input_97_mode_0 = const()[name = tensor("input_97_mode_0"), val = tensor("EXACT")]; + tensor input_97_cast_fp16 = gelu(mode = input_97_mode_0, x = linear_31_cast_fp16)[name = tensor("input_97_cast_fp16")]; + tensor linear7_2_weight_to_fp16 = const()[name = tensor("linear7_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(194140544)))]; + tensor linear_32_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = linear7_2_weight_to_fp16, x = input_97_cast_fp16)[name = tensor("linear_32_cast_fp16")]; + tensor input_99_cast_fp16 = add(x = input_93_cast_fp16, y = linear_32_cast_fp16)[name = tensor("input_99_cast_fp16")]; + tensor input_101_axes_0 = const()[name = tensor("input_101_axes_0"), val = tensor([-1])]; + tensor norm8_1_weight_to_fp16 = const()[name = tensor("norm8_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(202529216)))]; + tensor norm8_1_bias_to_fp16 = const()[name = tensor("norm8_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(202531328)))]; + tensor var_3292_to_fp16 = const()[name = tensor("op_3292_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_101_cast_fp16 = layer_norm(axes = input_101_axes_0, beta = norm8_1_bias_to_fp16, epsilon = var_3292_to_fp16, gamma = norm8_1_weight_to_fp16, x = input_99_cast_fp16)[name = tensor("input_101_cast_fp16")]; + tensor attn8_in_proj_weight_to_fp16 = const()[name = tensor("attn8_in_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(202533440)))]; + tensor linear_33_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = attn8_in_proj_weight_to_fp16, x = input_101_cast_fp16)[name = tensor("linear_33_cast_fp16")]; + tensor var_3306 = const()[name = tensor("op_3306"), val = tensor([1, 1, 3, 16, 64])]; + tensor qkv_17_cast_fp16 = reshape(shape = var_3306, x = linear_33_cast_fp16)[name = tensor("qkv_17_cast_fp16")]; + tensor q_49_begin_0 = const()[name = tensor("q_49_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor q_49_end_0 = const()[name = tensor("q_49_end_0"), val = tensor([1, 1, 1, 16, 64])]; + tensor q_49_end_mask_0 = const()[name = tensor("q_49_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor q_49_squeeze_mask_0 = const()[name = tensor("q_49_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor q_49_cast_fp16 = slice_by_index(begin = q_49_begin_0, end = q_49_end_0, end_mask = q_49_end_mask_0, squeeze_mask = q_49_squeeze_mask_0, x = qkv_17_cast_fp16)[name = tensor("q_49_cast_fp16")]; + tensor k_33_begin_0 = const()[name = tensor("k_33_begin_0"), val = tensor([0, 0, 1, 0, 0])]; + tensor k_33_end_0 = const()[name = tensor("k_33_end_0"), val = tensor([1, 1, 2, 16, 64])]; + tensor k_33_end_mask_0 = const()[name = tensor("k_33_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor k_33_squeeze_mask_0 = const()[name = tensor("k_33_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor k_33_cast_fp16 = slice_by_index(begin = k_33_begin_0, end = k_33_end_0, end_mask = k_33_end_mask_0, squeeze_mask = k_33_squeeze_mask_0, x = qkv_17_cast_fp16)[name = tensor("k_33_cast_fp16")]; + tensor v_17_begin_0 = const()[name = tensor("v_17_begin_0"), val = tensor([0, 0, 2, 0, 0])]; + tensor v_17_end_0 = const()[name = tensor("v_17_end_0"), val = tensor([1, 1, 3, 16, 64])]; + tensor v_17_end_mask_0 = const()[name = tensor("v_17_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor v_17_squeeze_mask_0 = const()[name = tensor("v_17_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor v_17_cast_fp16 = slice_by_index(begin = v_17_begin_0, end = v_17_end_0, end_mask = v_17_end_mask_0, squeeze_mask = v_17_squeeze_mask_0, x = qkv_17_cast_fp16)[name = tensor("v_17_cast_fp16")]; + tensor var_3372 = const()[name = tensor("op_3372"), val = tensor([1, 1, 1, 1])]; + tensor position8_to_fp16_dtype_0 = const()[name = tensor("position8_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor position8_to_fp16 = cast(dtype = position8_to_fp16_dtype_0, x = position8)[name = tensor("cast_513")]; + tensor ts_53_cast_fp16 = reshape(shape = var_3372, x = position8_to_fp16)[name = tensor("ts_53_cast_fp16")]; + tensor var_3379 = const()[name = tensor("op_3379"), val = tensor([1, 1, 16, 32, 2])]; + tensor q_complex_17_cast_fp16 = reshape(shape = var_3379, x = q_49_cast_fp16)[name = tensor("q_complex_17_cast_fp16")]; + tensor var_3386 = const()[name = tensor("op_3386"), val = tensor([1, 1, 16, 32, 2])]; + tensor k_complex_17_cast_fp16 = reshape(shape = var_3386, x = k_33_cast_fp16)[name = tensor("k_complex_17_cast_fp16")]; + tensor var_3390_begin_0 = const()[name = tensor("op_3390_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_3390_end_0 = const()[name = tensor("op_3390_end_0"), val = tensor([1, 1, 16, 32, 1])]; + tensor var_3390_end_mask_0 = const()[name = tensor("op_3390_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_3390_squeeze_mask_0 = const()[name = tensor("op_3390_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_3390_cast_fp16 = slice_by_index(begin = var_3390_begin_0, end = var_3390_end_0, end_mask = var_3390_end_mask_0, squeeze_mask = var_3390_squeeze_mask_0, x = q_complex_17_cast_fp16)[name = tensor("op_3390_cast_fp16")]; + tensor var_3398_begin_0 = const()[name = tensor("op_3398_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_3398_end_0 = const()[name = tensor("op_3398_end_0"), val = tensor([1, 1, 16, 32, 2])]; + tensor var_3398_end_mask_0 = const()[name = tensor("op_3398_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_3398_squeeze_mask_0 = const()[name = tensor("op_3398_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_3398_cast_fp16 = slice_by_index(begin = var_3398_begin_0, end = var_3398_end_0, end_mask = var_3398_end_mask_0, squeeze_mask = var_3398_squeeze_mask_0, x = q_complex_17_cast_fp16)[name = tensor("op_3398_cast_fp16")]; + tensor var_3406_begin_0 = const()[name = tensor("op_3406_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_3406_end_0 = const()[name = tensor("op_3406_end_0"), val = tensor([1, 1, 16, 32, 1])]; + tensor var_3406_end_mask_0 = const()[name = tensor("op_3406_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_3406_squeeze_mask_0 = const()[name = tensor("op_3406_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_3406_cast_fp16 = slice_by_index(begin = var_3406_begin_0, end = var_3406_end_0, end_mask = var_3406_end_mask_0, squeeze_mask = var_3406_squeeze_mask_0, x = k_complex_17_cast_fp16)[name = tensor("op_3406_cast_fp16")]; + tensor var_3414_begin_0 = const()[name = tensor("op_3414_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_3414_end_0 = const()[name = tensor("op_3414_end_0"), val = tensor([1, 1, 16, 32, 2])]; + tensor var_3414_end_mask_0 = const()[name = tensor("op_3414_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_3414_squeeze_mask_0 = const()[name = tensor("op_3414_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_3414_cast_fp16 = slice_by_index(begin = var_3414_begin_0, end = var_3414_end_0, end_mask = var_3414_end_mask_0, squeeze_mask = var_3414_squeeze_mask_0, x = k_complex_17_cast_fp16)[name = tensor("op_3414_cast_fp16")]; + tensor freqs_17_to_fp16 = const()[name = tensor("freqs_17_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(208824960)))]; + tensor var_3420_cast_fp16 = mul(x = freqs_17_to_fp16, y = ts_53_cast_fp16)[name = tensor("op_3420_cast_fp16")]; + tensor rotr_17_cast_fp16 = cos(x = var_3420_cast_fp16)[name = tensor("rotr_17_cast_fp16")]; + tensor roti_17_cast_fp16 = sin(x = var_3420_cast_fp16)[name = tensor("roti_17_cast_fp16")]; + tensor var_3424_cast_fp16 = mul(x = var_3390_cast_fp16, y = rotr_17_cast_fp16)[name = tensor("op_3424_cast_fp16")]; + tensor var_3425_cast_fp16 = mul(x = var_3398_cast_fp16, y = roti_17_cast_fp16)[name = tensor("op_3425_cast_fp16")]; + tensor qor_33_cast_fp16 = sub(x = var_3424_cast_fp16, y = var_3425_cast_fp16)[name = tensor("qor_33_cast_fp16")]; + tensor var_3428_cast_fp16 = mul(x = var_3390_cast_fp16, y = roti_17_cast_fp16)[name = tensor("op_3428_cast_fp16")]; + tensor var_3429_cast_fp16 = mul(x = var_3398_cast_fp16, y = rotr_17_cast_fp16)[name = tensor("op_3429_cast_fp16")]; + tensor qoi_33_cast_fp16 = add(x = var_3428_cast_fp16, y = var_3429_cast_fp16)[name = tensor("qoi_33_cast_fp16")]; + tensor var_3432_cast_fp16 = mul(x = var_3406_cast_fp16, y = rotr_17_cast_fp16)[name = tensor("op_3432_cast_fp16")]; + tensor var_3433_cast_fp16 = mul(x = var_3414_cast_fp16, y = roti_17_cast_fp16)[name = tensor("op_3433_cast_fp16")]; + tensor kor_33_cast_fp16 = sub(x = var_3432_cast_fp16, y = var_3433_cast_fp16)[name = tensor("kor_33_cast_fp16")]; + tensor var_3436_cast_fp16 = mul(x = var_3406_cast_fp16, y = roti_17_cast_fp16)[name = tensor("op_3436_cast_fp16")]; + tensor var_3437_cast_fp16 = mul(x = var_3414_cast_fp16, y = rotr_17_cast_fp16)[name = tensor("op_3437_cast_fp16")]; + tensor koi_33_cast_fp16 = add(x = var_3436_cast_fp16, y = var_3437_cast_fp16)[name = tensor("koi_33_cast_fp16")]; + tensor qo_17_axis_0 = const()[name = tensor("qo_17_axis_0"), val = tensor(-1)]; + tensor qo_17_cast_fp16 = stack(axis = qo_17_axis_0, values = (qor_33_cast_fp16, qoi_33_cast_fp16))[name = tensor("qo_17_cast_fp16")]; + tensor ko_17_axis_0 = const()[name = tensor("ko_17_axis_0"), val = tensor(-1)]; + tensor ko_17_cast_fp16 = stack(axis = ko_17_axis_0, values = (kor_33_cast_fp16, koi_33_cast_fp16))[name = tensor("ko_17_cast_fp16")]; + tensor var_3470 = const()[name = tensor("op_3470"), val = tensor([1, 1, 16, 64])]; + tensor q_51_cast_fp16 = reshape(shape = var_3470, x = qo_17_cast_fp16)[name = tensor("q_51_cast_fp16")]; + tensor var_3476 = const()[name = tensor("op_3476"), val = tensor([1, 1, 16, 64])]; + tensor k_35_cast_fp16 = reshape(shape = var_3476, x = ko_17_cast_fp16)[name = tensor("k_35_cast_fp16")]; + tensor _inversed_3501_y_0_to_fp16 = const()[name = tensor("_inversed_3501_y_0_to_fp16"), val = tensor(0x1p-9)]; + tensor _inversed_3501_cast_fp16 = mul(x = ts_53_cast_fp16, y = _inversed_3501_y_0_to_fp16)[name = tensor("_inversed_3501_cast_fp16")]; + tensor var_3502_cast_fp16 = floor(x = _inversed_3501_cast_fp16)[name = tensor("op_3502_cast_fp16")]; + tensor var_3503_to_fp16 = const()[name = tensor("op_3503_to_fp16"), val = tensor(0x1p+9)]; + tensor var_3504_cast_fp16 = mul(x = var_3502_cast_fp16, y = var_3503_to_fp16)[name = tensor("op_3504_cast_fp16")]; + tensor write_indices_float_35_cast_fp16 = sub(x = ts_53_cast_fp16, y = var_3504_cast_fp16)[name = tensor("write_indices_float_35_cast_fp16")]; + tensor var_3511_dtype_0 = const()[name = tensor("op_3511_dtype_0"), val = tensor("int32")]; + tensor write_indices_17_reps_0 = const()[name = tensor("write_indices_17_reps_0"), val = tensor([1, 1, 16, 64])]; + tensor write_indices_float_35_cast_fp16_to_int32 = cast(dtype = var_3511_dtype_0, x = write_indices_float_35_cast_fp16)[name = tensor("cast_512")]; + tensor write_indices_17 = tile(reps = write_indices_17_reps_0, x = write_indices_float_35_cast_fp16_to_int32)[name = tensor("write_indices_17")]; + tensor var_3521_begin_0 = const()[name = tensor("op_3521_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_3521_end_0 = const()[name = tensor("op_3521_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor var_3521_end_mask_0 = const()[name = tensor("op_3521_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_3521_squeeze_mask_0 = const()[name = tensor("op_3521_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor cache8_to_fp16_dtype_0 = const()[name = tensor("cache8_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor cache8_to_fp16 = cast(dtype = cache8_to_fp16_dtype_0, x = cache8)[name = tensor("cast_511")]; + tensor var_3521_cast_fp16 = slice_by_index(begin = var_3521_begin_0, end = var_3521_end_0, end_mask = var_3521_end_mask_0, squeeze_mask = var_3521_squeeze_mask_0, x = cache8_to_fp16)[name = tensor("op_3521_cast_fp16")]; + tensor var_3523_axis_0 = const()[name = tensor("op_3523_axis_0"), val = tensor(1)]; + tensor var_3523_mode_0 = const()[name = tensor("op_3523_mode_0"), val = tensor("update")]; + tensor var_3523_validate_indices_0 = const()[name = tensor("op_3523_validate_indices_0"), val = tensor(false)]; + tensor var_3523_cast_fp16 = scatter_along_axis(axis = var_3523_axis_0, data = var_3521_cast_fp16, indices = write_indices_17, mode = var_3523_mode_0, updates = k_35_cast_fp16, validate_indices = var_3523_validate_indices_0)[name = tensor("op_3523_cast_fp16")]; + tensor concat_58 = const()[name = tensor("concat_58"), val = tensor([0, 0, 0, 0, 0])]; + tensor concat_59 = const()[name = tensor("concat_59"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_17_internal_tensor_assign_1_stride_0 = const()[name = tensor("new_cache_17_internal_tensor_assign_1_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_17_internal_tensor_assign_1_begin_mask_0 = const()[name = tensor("new_cache_17_internal_tensor_assign_1_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_17_internal_tensor_assign_1_end_mask_0 = const()[name = tensor("new_cache_17_internal_tensor_assign_1_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_17_internal_tensor_assign_1_squeeze_mask_0 = const()[name = tensor("new_cache_17_internal_tensor_assign_1_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_64 = const()[name = tensor("shape_64"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_16 = const()[name = tensor("reduce_prod_16"), val = tensor(1048576)]; + tensor range_1d_16_start_0 = const()[name = tensor("range_1d_16_start_0"), val = tensor(0)]; + tensor range_1d_16_step_0 = const()[name = tensor("range_1d_16_step_0"), val = tensor(1)]; + tensor range_1d_16 = range_1d(end = reduce_prod_16, start = range_1d_16_start_0, step = range_1d_16_step_0)[name = tensor("range_1d_16")]; + tensor reshape_80 = reshape(shape = shape_64, x = range_1d_16)[name = tensor("reshape_80")]; + tensor slice_by_index_16 = slice_by_index(begin = concat_58, begin_mask = new_cache_17_internal_tensor_assign_1_begin_mask_0, end = concat_59, end_mask = new_cache_17_internal_tensor_assign_1_end_mask_0, squeeze_mask = new_cache_17_internal_tensor_assign_1_squeeze_mask_0, stride = new_cache_17_internal_tensor_assign_1_stride_0, x = reshape_80)[name = tensor("slice_by_index_16")]; + tensor reshape_81_shape_0 = const()[name = tensor("reshape_81_shape_0"), val = tensor([-1])]; + tensor reshape_81 = reshape(shape = reshape_81_shape_0, x = slice_by_index_16)[name = tensor("reshape_81")]; + tensor reshape_82_shape_0 = const()[name = tensor("reshape_82_shape_0"), val = tensor([-1])]; + tensor reshape_82_cast_fp16 = reshape(shape = reshape_82_shape_0, x = var_3523_cast_fp16)[name = tensor("reshape_82_cast_fp16")]; + tensor reshape_83_shape_0 = const()[name = tensor("reshape_83_shape_0"), val = tensor([-1])]; + tensor reshape_83_cast_fp16 = reshape(shape = reshape_83_shape_0, x = cache8_to_fp16)[name = tensor("reshape_83_cast_fp16")]; + tensor scatter_16_mode_0 = const()[name = tensor("scatter_16_mode_0"), val = tensor("update")]; + tensor scatter_16_axis_0 = const()[name = tensor("scatter_16_axis_0"), val = tensor(0)]; + tensor scatter_16_validate_indices_0 = const()[name = tensor("scatter_16_validate_indices_0"), val = tensor(false)]; + tensor scatter_16_cast_fp16 = scatter(axis = scatter_16_axis_0, data = reshape_83_cast_fp16, indices = reshape_81, mode = scatter_16_mode_0, updates = reshape_82_cast_fp16, validate_indices = scatter_16_validate_indices_0)[name = tensor("scatter_16_cast_fp16")]; + tensor reshape_84_cast_fp16 = reshape(shape = shape_64, x = scatter_16_cast_fp16)[name = tensor("reshape_84_cast_fp16")]; + tensor var_3531_begin_0 = const()[name = tensor("op_3531_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor var_3531_end_0 = const()[name = tensor("op_3531_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor var_3531_end_mask_0 = const()[name = tensor("op_3531_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_3531_squeeze_mask_0 = const()[name = tensor("op_3531_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor var_3531_cast_fp16 = slice_by_index(begin = var_3531_begin_0, end = var_3531_end_0, end_mask = var_3531_end_mask_0, squeeze_mask = var_3531_squeeze_mask_0, x = reshape_84_cast_fp16)[name = tensor("op_3531_cast_fp16")]; + tensor var_3533_axis_0 = const()[name = tensor("op_3533_axis_0"), val = tensor(1)]; + tensor var_3533_mode_0 = const()[name = tensor("op_3533_mode_0"), val = tensor("update")]; + tensor var_3533_validate_indices_0 = const()[name = tensor("op_3533_validate_indices_0"), val = tensor(false)]; + tensor var_3533_cast_fp16 = scatter_along_axis(axis = var_3533_axis_0, data = var_3531_cast_fp16, indices = write_indices_17, mode = var_3533_mode_0, updates = v_17_cast_fp16, validate_indices = var_3533_validate_indices_0)[name = tensor("op_3533_cast_fp16")]; + tensor concat_60 = const()[name = tensor("concat_60"), val = tensor([1, 0, 0, 0, 0])]; + tensor concat_61 = const()[name = tensor("concat_61"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_17_internal_tensor_assign_2_stride_0 = const()[name = tensor("new_cache_17_internal_tensor_assign_2_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_17_internal_tensor_assign_2_begin_mask_0 = const()[name = tensor("new_cache_17_internal_tensor_assign_2_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_17_internal_tensor_assign_2_end_mask_0 = const()[name = tensor("new_cache_17_internal_tensor_assign_2_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_17_internal_tensor_assign_2_squeeze_mask_0 = const()[name = tensor("new_cache_17_internal_tensor_assign_2_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_65 = const()[name = tensor("shape_65"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_17 = const()[name = tensor("reduce_prod_17"), val = tensor(1048576)]; + tensor range_1d_17_start_0 = const()[name = tensor("range_1d_17_start_0"), val = tensor(0)]; + tensor range_1d_17_step_0 = const()[name = tensor("range_1d_17_step_0"), val = tensor(1)]; + tensor range_1d_17 = range_1d(end = reduce_prod_17, start = range_1d_17_start_0, step = range_1d_17_step_0)[name = tensor("range_1d_17")]; + tensor reshape_85 = reshape(shape = shape_65, x = range_1d_17)[name = tensor("reshape_85")]; + tensor slice_by_index_17 = slice_by_index(begin = concat_60, begin_mask = new_cache_17_internal_tensor_assign_2_begin_mask_0, end = concat_61, end_mask = new_cache_17_internal_tensor_assign_2_end_mask_0, squeeze_mask = new_cache_17_internal_tensor_assign_2_squeeze_mask_0, stride = new_cache_17_internal_tensor_assign_2_stride_0, x = reshape_85)[name = tensor("slice_by_index_17")]; + tensor reshape_86_shape_0 = const()[name = tensor("reshape_86_shape_0"), val = tensor([-1])]; + tensor reshape_86 = reshape(shape = reshape_86_shape_0, x = slice_by_index_17)[name = tensor("reshape_86")]; + tensor reshape_87_shape_0 = const()[name = tensor("reshape_87_shape_0"), val = tensor([-1])]; + tensor reshape_87_cast_fp16 = reshape(shape = reshape_87_shape_0, x = var_3533_cast_fp16)[name = tensor("reshape_87_cast_fp16")]; + tensor reshape_88_shape_0 = const()[name = tensor("reshape_88_shape_0"), val = tensor([-1])]; + tensor reshape_88_cast_fp16 = reshape(shape = reshape_88_shape_0, x = reshape_84_cast_fp16)[name = tensor("reshape_88_cast_fp16")]; + tensor scatter_17_mode_0 = const()[name = tensor("scatter_17_mode_0"), val = tensor("update")]; + tensor scatter_17_axis_0 = const()[name = tensor("scatter_17_axis_0"), val = tensor(0)]; + tensor scatter_17_validate_indices_0 = const()[name = tensor("scatter_17_validate_indices_0"), val = tensor(false)]; + tensor scatter_17_cast_fp16 = scatter(axis = scatter_17_axis_0, data = reshape_88_cast_fp16, indices = reshape_86, mode = scatter_17_mode_0, updates = reshape_87_cast_fp16, validate_indices = scatter_17_validate_indices_0)[name = tensor("scatter_17_cast_fp16")]; + tensor reshape_89_cast_fp16 = reshape(shape = shape_65, x = scatter_17_cast_fp16)[name = tensor("reshape_89_cast_fp16")]; + tensor reshape_89_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("reshape_89_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor keys_49_begin_0 = const()[name = tensor("keys_49_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor keys_49_end_0 = const()[name = tensor("keys_49_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor keys_49_end_mask_0 = const()[name = tensor("keys_49_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor keys_49_squeeze_mask_0 = const()[name = tensor("keys_49_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor keys_49_cast_fp16 = slice_by_index(begin = keys_49_begin_0, end = keys_49_end_0, end_mask = keys_49_end_mask_0, squeeze_mask = keys_49_squeeze_mask_0, x = reshape_89_cast_fp16)[name = tensor("keys_49_cast_fp16")]; + tensor values_49_begin_0 = const()[name = tensor("values_49_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor values_49_end_0 = const()[name = tensor("values_49_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor values_49_end_mask_0 = const()[name = tensor("values_49_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor values_49_squeeze_mask_0 = const()[name = tensor("values_49_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor values_49_cast_fp16 = slice_by_index(begin = values_49_begin_0, end = values_49_end_0, end_mask = values_49_end_mask_0, squeeze_mask = values_49_squeeze_mask_0, x = reshape_89_cast_fp16)[name = tensor("values_49_cast_fp16")]; + tensor var_3545_cast_fp16 = not_equal(x = keys_49_cast_fp16, y = keys_49_cast_fp16)[name = tensor("op_3545_cast_fp16")]; + tensor keys_51_cast_fp16 = select(a = var_463_to_fp16, b = keys_49_cast_fp16, cond = var_3545_cast_fp16)[name = tensor("keys_51_cast_fp16")]; + tensor var_3553_cast_fp16 = not_equal(x = values_49_cast_fp16, y = values_49_cast_fp16)[name = tensor("op_3553_cast_fp16")]; + tensor values_51_cast_fp16 = select(a = var_463_to_fp16, b = values_49_cast_fp16, cond = var_3553_cast_fp16)[name = tensor("values_51_cast_fp16")]; + tensor var_3577 = const()[name = tensor("op_3577"), val = tensor([0, 2, 1, 3])]; + tensor var_3593 = const()[name = tensor("op_3593"), val = tensor([1, 1, 1])]; + tensor var_3594_cast_fp16 = reshape(shape = var_3593, x = position8_to_fp16)[name = tensor("op_3594_cast_fp16")]; + tensor var_3614_to_fp16 = const()[name = tensor("op_3614_to_fp16"), val = tensor(0x1p+0)]; + tensor valid_len_17_cast_fp16 = add(x = var_3594_cast_fp16, y = var_3614_to_fp16)[name = tensor("valid_len_17_cast_fp16")]; + tensor valid_mask_17_cast_fp16 = less(x = k_positions_1_promoted_to_fp16, y = valid_len_17_cast_fp16)[name = tensor("valid_mask_17_cast_fp16")]; + tensor causal_mask_17_cast_fp16 = less_equal(x = k_positions_1_promoted_to_fp16, y = var_3594_cast_fp16)[name = tensor("causal_mask_17_cast_fp16")]; + tensor attn_mask_33 = logical_and(x = valid_mask_17_cast_fp16, y = causal_mask_17_cast_fp16)[name = tensor("attn_mask_33")]; + tensor attn_mask_35_axes_0 = const()[name = tensor("attn_mask_35_axes_0"), val = tensor([1])]; + tensor attn_mask_35 = expand_dims(axes = attn_mask_35_axes_0, x = attn_mask_33)[name = tensor("attn_mask_35")]; + tensor var_3632_transpose_x_0 = const()[name = tensor("op_3632_transpose_x_0"), val = tensor(false)]; + tensor var_3632_transpose_y_0 = const()[name = tensor("op_3632_transpose_y_0"), val = tensor(false)]; + tensor transpose_88_perm_0 = const()[name = tensor("transpose_88_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_89_perm_0 = const()[name = tensor("transpose_89_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_89 = transpose(perm = transpose_89_perm_0, x = keys_51_cast_fp16)[name = tensor("transpose_181")]; + tensor transpose_88 = transpose(perm = transpose_88_perm_0, x = q_51_cast_fp16)[name = tensor("transpose_182")]; + tensor var_3632_cast_fp16 = matmul(transpose_x = var_3632_transpose_x_0, transpose_y = var_3632_transpose_y_0, x = transpose_88, y = transpose_89)[name = tensor("op_3632_cast_fp16")]; + tensor var_3626_to_fp16 = const()[name = tensor("op_3626_to_fp16"), val = tensor([0x1p-3])]; + tensor attn_weights_49_cast_fp16 = mul(x = var_3632_cast_fp16, y = var_3626_to_fp16)[name = tensor("attn_weights_49_cast_fp16")]; + tensor var_3634 = logical_not(x = attn_mask_35)[name = tensor("op_3634")]; + tensor var_3635_to_fp16 = const()[name = tensor("op_3635_to_fp16"), val = tensor(-inf)]; + tensor attn_weights_51_cast_fp16 = select(a = var_3635_to_fp16, b = attn_weights_49_cast_fp16, cond = var_3634)[name = tensor("attn_weights_51_cast_fp16")]; + tensor var_3637 = const()[name = tensor("op_3637"), val = tensor(-1)]; + tensor attn_weights_53_cast_fp16 = softmax(axis = var_3637, x = attn_weights_51_cast_fp16)[name = tensor("attn_weights_53_cast_fp16")]; + tensor attn_output_17_transpose_x_0 = const()[name = tensor("attn_output_17_transpose_x_0"), val = tensor(false)]; + tensor attn_output_17_transpose_y_0 = const()[name = tensor("attn_output_17_transpose_y_0"), val = tensor(false)]; + tensor values_53_cast_fp16 = transpose(perm = var_3577, x = values_51_cast_fp16)[name = tensor("transpose_183")]; + tensor attn_output_17_cast_fp16 = matmul(transpose_x = attn_output_17_transpose_x_0, transpose_y = attn_output_17_transpose_y_0, x = attn_weights_53_cast_fp16, y = values_53_cast_fp16)[name = tensor("attn_output_17_cast_fp16")]; + tensor var_3645 = const()[name = tensor("op_3645"), val = tensor([0, 2, 1, 3])]; + tensor var_3650 = const()[name = tensor("op_3650"), val = tensor([1, 1, 1024])]; + tensor var_3646_cast_fp16 = transpose(perm = var_3645, x = attn_output_17_cast_fp16)[name = tensor("transpose_180")]; + tensor input_103_cast_fp16 = reshape(shape = var_3650, x = var_3646_cast_fp16)[name = tensor("input_103_cast_fp16")]; + tensor attn8_out_proj_weight_to_fp16 = const()[name = tensor("attn8_out_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(208825088)))]; + tensor linear_34_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = attn8_out_proj_weight_to_fp16, x = input_103_cast_fp16)[name = tensor("linear_34_cast_fp16")]; + tensor var_3656_to_fp16 = const()[name = tensor("op_3656_to_fp16"), val = tensor(0x1p+0)]; + tensor var_3657_cast_fp16 = add(x = position8_to_fp16, y = var_3656_to_fp16)[name = tensor("op_3657_cast_fp16")]; + tensor var_3657_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_3657_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor input_105_cast_fp16 = add(x = input_99_cast_fp16, y = linear_34_cast_fp16)[name = tensor("input_105_cast_fp16")]; + tensor input_107_axes_0 = const()[name = tensor("input_107_axes_0"), val = tensor([-1])]; + tensor norm8_2_weight_to_fp16 = const()[name = tensor("norm8_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(210922304)))]; + tensor norm8_2_bias_to_fp16 = const()[name = tensor("norm8_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(210924416)))]; + tensor var_3661_to_fp16 = const()[name = tensor("op_3661_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_107_cast_fp16 = layer_norm(axes = input_107_axes_0, beta = norm8_2_bias_to_fp16, epsilon = var_3661_to_fp16, gamma = norm8_2_weight_to_fp16, x = input_105_cast_fp16)[name = tensor("input_107_cast_fp16")]; + tensor linear8_1_weight_to_fp16 = const()[name = tensor("linear8_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(210926528)))]; + tensor linear_35_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = linear8_1_weight_to_fp16, x = input_107_cast_fp16)[name = tensor("linear_35_cast_fp16")]; + tensor input_109_mode_0 = const()[name = tensor("input_109_mode_0"), val = tensor("EXACT")]; + tensor input_109_cast_fp16 = gelu(mode = input_109_mode_0, x = linear_35_cast_fp16)[name = tensor("input_109_cast_fp16")]; + tensor linear8_2_weight_to_fp16 = const()[name = tensor("linear8_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(219315200)))]; + tensor linear_36_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = linear8_2_weight_to_fp16, x = input_109_cast_fp16)[name = tensor("linear_36_cast_fp16")]; + tensor input_111_cast_fp16 = add(x = input_105_cast_fp16, y = linear_36_cast_fp16)[name = tensor("input_111_cast_fp16")]; + tensor input_113_axes_0 = const()[name = tensor("input_113_axes_0"), val = tensor([-1])]; + tensor norm9_1_weight_to_fp16 = const()[name = tensor("norm9_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(227703872)))]; + tensor norm9_1_bias_to_fp16 = const()[name = tensor("norm9_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(227705984)))]; + tensor var_3678_to_fp16 = const()[name = tensor("op_3678_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_113_cast_fp16 = layer_norm(axes = input_113_axes_0, beta = norm9_1_bias_to_fp16, epsilon = var_3678_to_fp16, gamma = norm9_1_weight_to_fp16, x = input_111_cast_fp16)[name = tensor("input_113_cast_fp16")]; + tensor attn9_in_proj_weight_to_fp16 = const()[name = tensor("attn9_in_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(227708096)))]; + tensor linear_37_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = attn9_in_proj_weight_to_fp16, x = input_113_cast_fp16)[name = tensor("linear_37_cast_fp16")]; + tensor var_3692 = const()[name = tensor("op_3692"), val = tensor([1, 1, 3, 16, 64])]; + tensor qkv_19_cast_fp16 = reshape(shape = var_3692, x = linear_37_cast_fp16)[name = tensor("qkv_19_cast_fp16")]; + tensor q_55_begin_0 = const()[name = tensor("q_55_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor q_55_end_0 = const()[name = tensor("q_55_end_0"), val = tensor([1, 1, 1, 16, 64])]; + tensor q_55_end_mask_0 = const()[name = tensor("q_55_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor q_55_squeeze_mask_0 = const()[name = tensor("q_55_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor q_55_cast_fp16 = slice_by_index(begin = q_55_begin_0, end = q_55_end_0, end_mask = q_55_end_mask_0, squeeze_mask = q_55_squeeze_mask_0, x = qkv_19_cast_fp16)[name = tensor("q_55_cast_fp16")]; + tensor k_37_begin_0 = const()[name = tensor("k_37_begin_0"), val = tensor([0, 0, 1, 0, 0])]; + tensor k_37_end_0 = const()[name = tensor("k_37_end_0"), val = tensor([1, 1, 2, 16, 64])]; + tensor k_37_end_mask_0 = const()[name = tensor("k_37_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor k_37_squeeze_mask_0 = const()[name = tensor("k_37_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor k_37_cast_fp16 = slice_by_index(begin = k_37_begin_0, end = k_37_end_0, end_mask = k_37_end_mask_0, squeeze_mask = k_37_squeeze_mask_0, x = qkv_19_cast_fp16)[name = tensor("k_37_cast_fp16")]; + tensor v_19_begin_0 = const()[name = tensor("v_19_begin_0"), val = tensor([0, 0, 2, 0, 0])]; + tensor v_19_end_0 = const()[name = tensor("v_19_end_0"), val = tensor([1, 1, 3, 16, 64])]; + tensor v_19_end_mask_0 = const()[name = tensor("v_19_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor v_19_squeeze_mask_0 = const()[name = tensor("v_19_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor v_19_cast_fp16 = slice_by_index(begin = v_19_begin_0, end = v_19_end_0, end_mask = v_19_end_mask_0, squeeze_mask = v_19_squeeze_mask_0, x = qkv_19_cast_fp16)[name = tensor("v_19_cast_fp16")]; + tensor var_3758 = const()[name = tensor("op_3758"), val = tensor([1, 1, 1, 1])]; + tensor position9_to_fp16_dtype_0 = const()[name = tensor("position9_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor position9_to_fp16 = cast(dtype = position9_to_fp16_dtype_0, x = position9)[name = tensor("cast_508")]; + tensor ts_59_cast_fp16 = reshape(shape = var_3758, x = position9_to_fp16)[name = tensor("ts_59_cast_fp16")]; + tensor var_3765 = const()[name = tensor("op_3765"), val = tensor([1, 1, 16, 32, 2])]; + tensor q_complex_19_cast_fp16 = reshape(shape = var_3765, x = q_55_cast_fp16)[name = tensor("q_complex_19_cast_fp16")]; + tensor var_3772 = const()[name = tensor("op_3772"), val = tensor([1, 1, 16, 32, 2])]; + tensor k_complex_19_cast_fp16 = reshape(shape = var_3772, x = k_37_cast_fp16)[name = tensor("k_complex_19_cast_fp16")]; + tensor var_3776_begin_0 = const()[name = tensor("op_3776_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_3776_end_0 = const()[name = tensor("op_3776_end_0"), val = tensor([1, 1, 16, 32, 1])]; + tensor var_3776_end_mask_0 = const()[name = tensor("op_3776_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_3776_squeeze_mask_0 = const()[name = tensor("op_3776_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_3776_cast_fp16 = slice_by_index(begin = var_3776_begin_0, end = var_3776_end_0, end_mask = var_3776_end_mask_0, squeeze_mask = var_3776_squeeze_mask_0, x = q_complex_19_cast_fp16)[name = tensor("op_3776_cast_fp16")]; + tensor var_3784_begin_0 = const()[name = tensor("op_3784_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_3784_end_0 = const()[name = tensor("op_3784_end_0"), val = tensor([1, 1, 16, 32, 2])]; + tensor var_3784_end_mask_0 = const()[name = tensor("op_3784_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_3784_squeeze_mask_0 = const()[name = tensor("op_3784_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_3784_cast_fp16 = slice_by_index(begin = var_3784_begin_0, end = var_3784_end_0, end_mask = var_3784_end_mask_0, squeeze_mask = var_3784_squeeze_mask_0, x = q_complex_19_cast_fp16)[name = tensor("op_3784_cast_fp16")]; + tensor var_3792_begin_0 = const()[name = tensor("op_3792_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_3792_end_0 = const()[name = tensor("op_3792_end_0"), val = tensor([1, 1, 16, 32, 1])]; + tensor var_3792_end_mask_0 = const()[name = tensor("op_3792_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_3792_squeeze_mask_0 = const()[name = tensor("op_3792_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_3792_cast_fp16 = slice_by_index(begin = var_3792_begin_0, end = var_3792_end_0, end_mask = var_3792_end_mask_0, squeeze_mask = var_3792_squeeze_mask_0, x = k_complex_19_cast_fp16)[name = tensor("op_3792_cast_fp16")]; + tensor var_3800_begin_0 = const()[name = tensor("op_3800_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_3800_end_0 = const()[name = tensor("op_3800_end_0"), val = tensor([1, 1, 16, 32, 2])]; + tensor var_3800_end_mask_0 = const()[name = tensor("op_3800_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_3800_squeeze_mask_0 = const()[name = tensor("op_3800_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_3800_cast_fp16 = slice_by_index(begin = var_3800_begin_0, end = var_3800_end_0, end_mask = var_3800_end_mask_0, squeeze_mask = var_3800_squeeze_mask_0, x = k_complex_19_cast_fp16)[name = tensor("op_3800_cast_fp16")]; + tensor freqs_19_to_fp16 = const()[name = tensor("freqs_19_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(233999616)))]; + tensor var_3806_cast_fp16 = mul(x = freqs_19_to_fp16, y = ts_59_cast_fp16)[name = tensor("op_3806_cast_fp16")]; + tensor rotr_19_cast_fp16 = cos(x = var_3806_cast_fp16)[name = tensor("rotr_19_cast_fp16")]; + tensor roti_19_cast_fp16 = sin(x = var_3806_cast_fp16)[name = tensor("roti_19_cast_fp16")]; + tensor var_3810_cast_fp16 = mul(x = var_3776_cast_fp16, y = rotr_19_cast_fp16)[name = tensor("op_3810_cast_fp16")]; + tensor var_3811_cast_fp16 = mul(x = var_3784_cast_fp16, y = roti_19_cast_fp16)[name = tensor("op_3811_cast_fp16")]; + tensor qor_37_cast_fp16 = sub(x = var_3810_cast_fp16, y = var_3811_cast_fp16)[name = tensor("qor_37_cast_fp16")]; + tensor var_3814_cast_fp16 = mul(x = var_3776_cast_fp16, y = roti_19_cast_fp16)[name = tensor("op_3814_cast_fp16")]; + tensor var_3815_cast_fp16 = mul(x = var_3784_cast_fp16, y = rotr_19_cast_fp16)[name = tensor("op_3815_cast_fp16")]; + tensor qoi_37_cast_fp16 = add(x = var_3814_cast_fp16, y = var_3815_cast_fp16)[name = tensor("qoi_37_cast_fp16")]; + tensor var_3818_cast_fp16 = mul(x = var_3792_cast_fp16, y = rotr_19_cast_fp16)[name = tensor("op_3818_cast_fp16")]; + tensor var_3819_cast_fp16 = mul(x = var_3800_cast_fp16, y = roti_19_cast_fp16)[name = tensor("op_3819_cast_fp16")]; + tensor kor_37_cast_fp16 = sub(x = var_3818_cast_fp16, y = var_3819_cast_fp16)[name = tensor("kor_37_cast_fp16")]; + tensor var_3822_cast_fp16 = mul(x = var_3792_cast_fp16, y = roti_19_cast_fp16)[name = tensor("op_3822_cast_fp16")]; + tensor var_3823_cast_fp16 = mul(x = var_3800_cast_fp16, y = rotr_19_cast_fp16)[name = tensor("op_3823_cast_fp16")]; + tensor koi_37_cast_fp16 = add(x = var_3822_cast_fp16, y = var_3823_cast_fp16)[name = tensor("koi_37_cast_fp16")]; + tensor qo_19_axis_0 = const()[name = tensor("qo_19_axis_0"), val = tensor(-1)]; + tensor qo_19_cast_fp16 = stack(axis = qo_19_axis_0, values = (qor_37_cast_fp16, qoi_37_cast_fp16))[name = tensor("qo_19_cast_fp16")]; + tensor ko_19_axis_0 = const()[name = tensor("ko_19_axis_0"), val = tensor(-1)]; + tensor ko_19_cast_fp16 = stack(axis = ko_19_axis_0, values = (kor_37_cast_fp16, koi_37_cast_fp16))[name = tensor("ko_19_cast_fp16")]; + tensor var_3856 = const()[name = tensor("op_3856"), val = tensor([1, 1, 16, 64])]; + tensor q_57_cast_fp16 = reshape(shape = var_3856, x = qo_19_cast_fp16)[name = tensor("q_57_cast_fp16")]; + tensor var_3862 = const()[name = tensor("op_3862"), val = tensor([1, 1, 16, 64])]; + tensor k_39_cast_fp16 = reshape(shape = var_3862, x = ko_19_cast_fp16)[name = tensor("k_39_cast_fp16")]; + tensor _inversed_3887_y_0_to_fp16 = const()[name = tensor("_inversed_3887_y_0_to_fp16"), val = tensor(0x1p-9)]; + tensor _inversed_3887_cast_fp16 = mul(x = ts_59_cast_fp16, y = _inversed_3887_y_0_to_fp16)[name = tensor("_inversed_3887_cast_fp16")]; + tensor var_3888_cast_fp16 = floor(x = _inversed_3887_cast_fp16)[name = tensor("op_3888_cast_fp16")]; + tensor var_3889_to_fp16 = const()[name = tensor("op_3889_to_fp16"), val = tensor(0x1p+9)]; + tensor var_3890_cast_fp16 = mul(x = var_3888_cast_fp16, y = var_3889_to_fp16)[name = tensor("op_3890_cast_fp16")]; + tensor write_indices_float_39_cast_fp16 = sub(x = ts_59_cast_fp16, y = var_3890_cast_fp16)[name = tensor("write_indices_float_39_cast_fp16")]; + tensor var_3897_dtype_0 = const()[name = tensor("op_3897_dtype_0"), val = tensor("int32")]; + tensor write_indices_19_reps_0 = const()[name = tensor("write_indices_19_reps_0"), val = tensor([1, 1, 16, 64])]; + tensor write_indices_float_39_cast_fp16_to_int32 = cast(dtype = var_3897_dtype_0, x = write_indices_float_39_cast_fp16)[name = tensor("cast_507")]; + tensor write_indices_19 = tile(reps = write_indices_19_reps_0, x = write_indices_float_39_cast_fp16_to_int32)[name = tensor("write_indices_19")]; + tensor var_3907_begin_0 = const()[name = tensor("op_3907_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_3907_end_0 = const()[name = tensor("op_3907_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor var_3907_end_mask_0 = const()[name = tensor("op_3907_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_3907_squeeze_mask_0 = const()[name = tensor("op_3907_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor cache9_to_fp16_dtype_0 = const()[name = tensor("cache9_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor cache9_to_fp16 = cast(dtype = cache9_to_fp16_dtype_0, x = cache9)[name = tensor("cast_506")]; + tensor var_3907_cast_fp16 = slice_by_index(begin = var_3907_begin_0, end = var_3907_end_0, end_mask = var_3907_end_mask_0, squeeze_mask = var_3907_squeeze_mask_0, x = cache9_to_fp16)[name = tensor("op_3907_cast_fp16")]; + tensor var_3909_axis_0 = const()[name = tensor("op_3909_axis_0"), val = tensor(1)]; + tensor var_3909_mode_0 = const()[name = tensor("op_3909_mode_0"), val = tensor("update")]; + tensor var_3909_validate_indices_0 = const()[name = tensor("op_3909_validate_indices_0"), val = tensor(false)]; + tensor var_3909_cast_fp16 = scatter_along_axis(axis = var_3909_axis_0, data = var_3907_cast_fp16, indices = write_indices_19, mode = var_3909_mode_0, updates = k_39_cast_fp16, validate_indices = var_3909_validate_indices_0)[name = tensor("op_3909_cast_fp16")]; + tensor concat_65 = const()[name = tensor("concat_65"), val = tensor([0, 0, 0, 0, 0])]; + tensor concat_66 = const()[name = tensor("concat_66"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_19_internal_tensor_assign_1_stride_0 = const()[name = tensor("new_cache_19_internal_tensor_assign_1_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_19_internal_tensor_assign_1_begin_mask_0 = const()[name = tensor("new_cache_19_internal_tensor_assign_1_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_19_internal_tensor_assign_1_end_mask_0 = const()[name = tensor("new_cache_19_internal_tensor_assign_1_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_19_internal_tensor_assign_1_squeeze_mask_0 = const()[name = tensor("new_cache_19_internal_tensor_assign_1_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_66 = const()[name = tensor("shape_66"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_18 = const()[name = tensor("reduce_prod_18"), val = tensor(1048576)]; + tensor range_1d_18_start_0 = const()[name = tensor("range_1d_18_start_0"), val = tensor(0)]; + tensor range_1d_18_step_0 = const()[name = tensor("range_1d_18_step_0"), val = tensor(1)]; + tensor range_1d_18 = range_1d(end = reduce_prod_18, start = range_1d_18_start_0, step = range_1d_18_step_0)[name = tensor("range_1d_18")]; + tensor reshape_90 = reshape(shape = shape_66, x = range_1d_18)[name = tensor("reshape_90")]; + tensor slice_by_index_18 = slice_by_index(begin = concat_65, begin_mask = new_cache_19_internal_tensor_assign_1_begin_mask_0, end = concat_66, end_mask = new_cache_19_internal_tensor_assign_1_end_mask_0, squeeze_mask = new_cache_19_internal_tensor_assign_1_squeeze_mask_0, stride = new_cache_19_internal_tensor_assign_1_stride_0, x = reshape_90)[name = tensor("slice_by_index_18")]; + tensor reshape_91_shape_0 = const()[name = tensor("reshape_91_shape_0"), val = tensor([-1])]; + tensor reshape_91 = reshape(shape = reshape_91_shape_0, x = slice_by_index_18)[name = tensor("reshape_91")]; + tensor reshape_92_shape_0 = const()[name = tensor("reshape_92_shape_0"), val = tensor([-1])]; + tensor reshape_92_cast_fp16 = reshape(shape = reshape_92_shape_0, x = var_3909_cast_fp16)[name = tensor("reshape_92_cast_fp16")]; + tensor reshape_93_shape_0 = const()[name = tensor("reshape_93_shape_0"), val = tensor([-1])]; + tensor reshape_93_cast_fp16 = reshape(shape = reshape_93_shape_0, x = cache9_to_fp16)[name = tensor("reshape_93_cast_fp16")]; + tensor scatter_18_mode_0 = const()[name = tensor("scatter_18_mode_0"), val = tensor("update")]; + tensor scatter_18_axis_0 = const()[name = tensor("scatter_18_axis_0"), val = tensor(0)]; + tensor scatter_18_validate_indices_0 = const()[name = tensor("scatter_18_validate_indices_0"), val = tensor(false)]; + tensor scatter_18_cast_fp16 = scatter(axis = scatter_18_axis_0, data = reshape_93_cast_fp16, indices = reshape_91, mode = scatter_18_mode_0, updates = reshape_92_cast_fp16, validate_indices = scatter_18_validate_indices_0)[name = tensor("scatter_18_cast_fp16")]; + tensor reshape_94_cast_fp16 = reshape(shape = shape_66, x = scatter_18_cast_fp16)[name = tensor("reshape_94_cast_fp16")]; + tensor var_3917_begin_0 = const()[name = tensor("op_3917_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor var_3917_end_0 = const()[name = tensor("op_3917_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor var_3917_end_mask_0 = const()[name = tensor("op_3917_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_3917_squeeze_mask_0 = const()[name = tensor("op_3917_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor var_3917_cast_fp16 = slice_by_index(begin = var_3917_begin_0, end = var_3917_end_0, end_mask = var_3917_end_mask_0, squeeze_mask = var_3917_squeeze_mask_0, x = reshape_94_cast_fp16)[name = tensor("op_3917_cast_fp16")]; + tensor var_3919_axis_0 = const()[name = tensor("op_3919_axis_0"), val = tensor(1)]; + tensor var_3919_mode_0 = const()[name = tensor("op_3919_mode_0"), val = tensor("update")]; + tensor var_3919_validate_indices_0 = const()[name = tensor("op_3919_validate_indices_0"), val = tensor(false)]; + tensor var_3919_cast_fp16 = scatter_along_axis(axis = var_3919_axis_0, data = var_3917_cast_fp16, indices = write_indices_19, mode = var_3919_mode_0, updates = v_19_cast_fp16, validate_indices = var_3919_validate_indices_0)[name = tensor("op_3919_cast_fp16")]; + tensor concat_67 = const()[name = tensor("concat_67"), val = tensor([1, 0, 0, 0, 0])]; + tensor concat_68 = const()[name = tensor("concat_68"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_19_internal_tensor_assign_2_stride_0 = const()[name = tensor("new_cache_19_internal_tensor_assign_2_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_19_internal_tensor_assign_2_begin_mask_0 = const()[name = tensor("new_cache_19_internal_tensor_assign_2_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_19_internal_tensor_assign_2_end_mask_0 = const()[name = tensor("new_cache_19_internal_tensor_assign_2_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_19_internal_tensor_assign_2_squeeze_mask_0 = const()[name = tensor("new_cache_19_internal_tensor_assign_2_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_67 = const()[name = tensor("shape_67"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_19 = const()[name = tensor("reduce_prod_19"), val = tensor(1048576)]; + tensor range_1d_19_start_0 = const()[name = tensor("range_1d_19_start_0"), val = tensor(0)]; + tensor range_1d_19_step_0 = const()[name = tensor("range_1d_19_step_0"), val = tensor(1)]; + tensor range_1d_19 = range_1d(end = reduce_prod_19, start = range_1d_19_start_0, step = range_1d_19_step_0)[name = tensor("range_1d_19")]; + tensor reshape_95 = reshape(shape = shape_67, x = range_1d_19)[name = tensor("reshape_95")]; + tensor slice_by_index_19 = slice_by_index(begin = concat_67, begin_mask = new_cache_19_internal_tensor_assign_2_begin_mask_0, end = concat_68, end_mask = new_cache_19_internal_tensor_assign_2_end_mask_0, squeeze_mask = new_cache_19_internal_tensor_assign_2_squeeze_mask_0, stride = new_cache_19_internal_tensor_assign_2_stride_0, x = reshape_95)[name = tensor("slice_by_index_19")]; + tensor reshape_96_shape_0 = const()[name = tensor("reshape_96_shape_0"), val = tensor([-1])]; + tensor reshape_96 = reshape(shape = reshape_96_shape_0, x = slice_by_index_19)[name = tensor("reshape_96")]; + tensor reshape_97_shape_0 = const()[name = tensor("reshape_97_shape_0"), val = tensor([-1])]; + tensor reshape_97_cast_fp16 = reshape(shape = reshape_97_shape_0, x = var_3919_cast_fp16)[name = tensor("reshape_97_cast_fp16")]; + tensor reshape_98_shape_0 = const()[name = tensor("reshape_98_shape_0"), val = tensor([-1])]; + tensor reshape_98_cast_fp16 = reshape(shape = reshape_98_shape_0, x = reshape_94_cast_fp16)[name = tensor("reshape_98_cast_fp16")]; + tensor scatter_19_mode_0 = const()[name = tensor("scatter_19_mode_0"), val = tensor("update")]; + tensor scatter_19_axis_0 = const()[name = tensor("scatter_19_axis_0"), val = tensor(0)]; + tensor scatter_19_validate_indices_0 = const()[name = tensor("scatter_19_validate_indices_0"), val = tensor(false)]; + tensor scatter_19_cast_fp16 = scatter(axis = scatter_19_axis_0, data = reshape_98_cast_fp16, indices = reshape_96, mode = scatter_19_mode_0, updates = reshape_97_cast_fp16, validate_indices = scatter_19_validate_indices_0)[name = tensor("scatter_19_cast_fp16")]; + tensor reshape_99_cast_fp16 = reshape(shape = shape_67, x = scatter_19_cast_fp16)[name = tensor("reshape_99_cast_fp16")]; + tensor reshape_99_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("reshape_99_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor keys_55_begin_0 = const()[name = tensor("keys_55_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor keys_55_end_0 = const()[name = tensor("keys_55_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor keys_55_end_mask_0 = const()[name = tensor("keys_55_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor keys_55_squeeze_mask_0 = const()[name = tensor("keys_55_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor keys_55_cast_fp16 = slice_by_index(begin = keys_55_begin_0, end = keys_55_end_0, end_mask = keys_55_end_mask_0, squeeze_mask = keys_55_squeeze_mask_0, x = reshape_99_cast_fp16)[name = tensor("keys_55_cast_fp16")]; + tensor values_55_begin_0 = const()[name = tensor("values_55_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor values_55_end_0 = const()[name = tensor("values_55_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor values_55_end_mask_0 = const()[name = tensor("values_55_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor values_55_squeeze_mask_0 = const()[name = tensor("values_55_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor values_55_cast_fp16 = slice_by_index(begin = values_55_begin_0, end = values_55_end_0, end_mask = values_55_end_mask_0, squeeze_mask = values_55_squeeze_mask_0, x = reshape_99_cast_fp16)[name = tensor("values_55_cast_fp16")]; + tensor var_3931_cast_fp16 = not_equal(x = keys_55_cast_fp16, y = keys_55_cast_fp16)[name = tensor("op_3931_cast_fp16")]; + tensor keys_57_cast_fp16 = select(a = var_463_to_fp16, b = keys_55_cast_fp16, cond = var_3931_cast_fp16)[name = tensor("keys_57_cast_fp16")]; + tensor var_3939_cast_fp16 = not_equal(x = values_55_cast_fp16, y = values_55_cast_fp16)[name = tensor("op_3939_cast_fp16")]; + tensor values_57_cast_fp16 = select(a = var_463_to_fp16, b = values_55_cast_fp16, cond = var_3939_cast_fp16)[name = tensor("values_57_cast_fp16")]; + tensor var_3963 = const()[name = tensor("op_3963"), val = tensor([0, 2, 1, 3])]; + tensor var_3979 = const()[name = tensor("op_3979"), val = tensor([1, 1, 1])]; + tensor var_3980_cast_fp16 = reshape(shape = var_3979, x = position9_to_fp16)[name = tensor("op_3980_cast_fp16")]; + tensor var_4000_to_fp16 = const()[name = tensor("op_4000_to_fp16"), val = tensor(0x1p+0)]; + tensor valid_len_19_cast_fp16 = add(x = var_3980_cast_fp16, y = var_4000_to_fp16)[name = tensor("valid_len_19_cast_fp16")]; + tensor valid_mask_19_cast_fp16 = less(x = k_positions_1_promoted_to_fp16, y = valid_len_19_cast_fp16)[name = tensor("valid_mask_19_cast_fp16")]; + tensor causal_mask_19_cast_fp16 = less_equal(x = k_positions_1_promoted_to_fp16, y = var_3980_cast_fp16)[name = tensor("causal_mask_19_cast_fp16")]; + tensor attn_mask_37 = logical_and(x = valid_mask_19_cast_fp16, y = causal_mask_19_cast_fp16)[name = tensor("attn_mask_37")]; + tensor attn_mask_39_axes_0 = const()[name = tensor("attn_mask_39_axes_0"), val = tensor([1])]; + tensor attn_mask_39 = expand_dims(axes = attn_mask_39_axes_0, x = attn_mask_37)[name = tensor("attn_mask_39")]; + tensor var_4018_transpose_x_0 = const()[name = tensor("op_4018_transpose_x_0"), val = tensor(false)]; + tensor var_4018_transpose_y_0 = const()[name = tensor("op_4018_transpose_y_0"), val = tensor(false)]; + tensor transpose_90_perm_0 = const()[name = tensor("transpose_90_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_91_perm_0 = const()[name = tensor("transpose_91_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_91 = transpose(perm = transpose_91_perm_0, x = keys_57_cast_fp16)[name = tensor("transpose_177")]; + tensor transpose_90 = transpose(perm = transpose_90_perm_0, x = q_57_cast_fp16)[name = tensor("transpose_178")]; + tensor var_4018_cast_fp16 = matmul(transpose_x = var_4018_transpose_x_0, transpose_y = var_4018_transpose_y_0, x = transpose_90, y = transpose_91)[name = tensor("op_4018_cast_fp16")]; + tensor var_4012_to_fp16 = const()[name = tensor("op_4012_to_fp16"), val = tensor([0x1p-3])]; + tensor attn_weights_55_cast_fp16 = mul(x = var_4018_cast_fp16, y = var_4012_to_fp16)[name = tensor("attn_weights_55_cast_fp16")]; + tensor var_4020 = logical_not(x = attn_mask_39)[name = tensor("op_4020")]; + tensor var_4021_to_fp16 = const()[name = tensor("op_4021_to_fp16"), val = tensor(-inf)]; + tensor attn_weights_57_cast_fp16 = select(a = var_4021_to_fp16, b = attn_weights_55_cast_fp16, cond = var_4020)[name = tensor("attn_weights_57_cast_fp16")]; + tensor var_4023 = const()[name = tensor("op_4023"), val = tensor(-1)]; + tensor attn_weights_59_cast_fp16 = softmax(axis = var_4023, x = attn_weights_57_cast_fp16)[name = tensor("attn_weights_59_cast_fp16")]; + tensor attn_output_19_transpose_x_0 = const()[name = tensor("attn_output_19_transpose_x_0"), val = tensor(false)]; + tensor attn_output_19_transpose_y_0 = const()[name = tensor("attn_output_19_transpose_y_0"), val = tensor(false)]; + tensor values_59_cast_fp16 = transpose(perm = var_3963, x = values_57_cast_fp16)[name = tensor("transpose_179")]; + tensor attn_output_19_cast_fp16 = matmul(transpose_x = attn_output_19_transpose_x_0, transpose_y = attn_output_19_transpose_y_0, x = attn_weights_59_cast_fp16, y = values_59_cast_fp16)[name = tensor("attn_output_19_cast_fp16")]; + tensor var_4031 = const()[name = tensor("op_4031"), val = tensor([0, 2, 1, 3])]; + tensor var_4036 = const()[name = tensor("op_4036"), val = tensor([1, 1, 1024])]; + tensor var_4032_cast_fp16 = transpose(perm = var_4031, x = attn_output_19_cast_fp16)[name = tensor("transpose_176")]; + tensor input_115_cast_fp16 = reshape(shape = var_4036, x = var_4032_cast_fp16)[name = tensor("input_115_cast_fp16")]; + tensor attn9_out_proj_weight_to_fp16 = const()[name = tensor("attn9_out_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(233999744)))]; + tensor linear_38_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = attn9_out_proj_weight_to_fp16, x = input_115_cast_fp16)[name = tensor("linear_38_cast_fp16")]; + tensor var_4042_to_fp16 = const()[name = tensor("op_4042_to_fp16"), val = tensor(0x1p+0)]; + tensor var_4043_cast_fp16 = add(x = position9_to_fp16, y = var_4042_to_fp16)[name = tensor("op_4043_cast_fp16")]; + tensor var_4043_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_4043_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor input_117_cast_fp16 = add(x = input_111_cast_fp16, y = linear_38_cast_fp16)[name = tensor("input_117_cast_fp16")]; + tensor input_119_axes_0 = const()[name = tensor("input_119_axes_0"), val = tensor([-1])]; + tensor norm9_2_weight_to_fp16 = const()[name = tensor("norm9_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(236096960)))]; + tensor norm9_2_bias_to_fp16 = const()[name = tensor("norm9_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(236099072)))]; + tensor var_4047_to_fp16 = const()[name = tensor("op_4047_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_119_cast_fp16 = layer_norm(axes = input_119_axes_0, beta = norm9_2_bias_to_fp16, epsilon = var_4047_to_fp16, gamma = norm9_2_weight_to_fp16, x = input_117_cast_fp16)[name = tensor("input_119_cast_fp16")]; + tensor linear9_1_weight_to_fp16 = const()[name = tensor("linear9_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(236101184)))]; + tensor linear_39_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = linear9_1_weight_to_fp16, x = input_119_cast_fp16)[name = tensor("linear_39_cast_fp16")]; + tensor input_121_mode_0 = const()[name = tensor("input_121_mode_0"), val = tensor("EXACT")]; + tensor input_121_cast_fp16 = gelu(mode = input_121_mode_0, x = linear_39_cast_fp16)[name = tensor("input_121_cast_fp16")]; + tensor linear9_2_weight_to_fp16 = const()[name = tensor("linear9_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(244489856)))]; + tensor linear_40_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = linear9_2_weight_to_fp16, x = input_121_cast_fp16)[name = tensor("linear_40_cast_fp16")]; + tensor input_123_cast_fp16 = add(x = input_117_cast_fp16, y = linear_40_cast_fp16)[name = tensor("input_123_cast_fp16")]; + tensor input_125_axes_0 = const()[name = tensor("input_125_axes_0"), val = tensor([-1])]; + tensor norm10_1_weight_to_fp16 = const()[name = tensor("norm10_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(252878528)))]; + tensor norm10_1_bias_to_fp16 = const()[name = tensor("norm10_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(252880640)))]; + tensor var_4064_to_fp16 = const()[name = tensor("op_4064_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_125_cast_fp16 = layer_norm(axes = input_125_axes_0, beta = norm10_1_bias_to_fp16, epsilon = var_4064_to_fp16, gamma = norm10_1_weight_to_fp16, x = input_123_cast_fp16)[name = tensor("input_125_cast_fp16")]; + tensor attn10_in_proj_weight_to_fp16 = const()[name = tensor("attn10_in_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(252882752)))]; + tensor linear_41_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = attn10_in_proj_weight_to_fp16, x = input_125_cast_fp16)[name = tensor("linear_41_cast_fp16")]; + tensor var_4078 = const()[name = tensor("op_4078"), val = tensor([1, 1, 3, 16, 64])]; + tensor qkv_21_cast_fp16 = reshape(shape = var_4078, x = linear_41_cast_fp16)[name = tensor("qkv_21_cast_fp16")]; + tensor q_61_begin_0 = const()[name = tensor("q_61_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor q_61_end_0 = const()[name = tensor("q_61_end_0"), val = tensor([1, 1, 1, 16, 64])]; + tensor q_61_end_mask_0 = const()[name = tensor("q_61_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor q_61_squeeze_mask_0 = const()[name = tensor("q_61_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor q_61_cast_fp16 = slice_by_index(begin = q_61_begin_0, end = q_61_end_0, end_mask = q_61_end_mask_0, squeeze_mask = q_61_squeeze_mask_0, x = qkv_21_cast_fp16)[name = tensor("q_61_cast_fp16")]; + tensor k_41_begin_0 = const()[name = tensor("k_41_begin_0"), val = tensor([0, 0, 1, 0, 0])]; + tensor k_41_end_0 = const()[name = tensor("k_41_end_0"), val = tensor([1, 1, 2, 16, 64])]; + tensor k_41_end_mask_0 = const()[name = tensor("k_41_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor k_41_squeeze_mask_0 = const()[name = tensor("k_41_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor k_41_cast_fp16 = slice_by_index(begin = k_41_begin_0, end = k_41_end_0, end_mask = k_41_end_mask_0, squeeze_mask = k_41_squeeze_mask_0, x = qkv_21_cast_fp16)[name = tensor("k_41_cast_fp16")]; + tensor v_21_begin_0 = const()[name = tensor("v_21_begin_0"), val = tensor([0, 0, 2, 0, 0])]; + tensor v_21_end_0 = const()[name = tensor("v_21_end_0"), val = tensor([1, 1, 3, 16, 64])]; + tensor v_21_end_mask_0 = const()[name = tensor("v_21_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor v_21_squeeze_mask_0 = const()[name = tensor("v_21_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor v_21_cast_fp16 = slice_by_index(begin = v_21_begin_0, end = v_21_end_0, end_mask = v_21_end_mask_0, squeeze_mask = v_21_squeeze_mask_0, x = qkv_21_cast_fp16)[name = tensor("v_21_cast_fp16")]; + tensor var_4144 = const()[name = tensor("op_4144"), val = tensor([1, 1, 1, 1])]; + tensor position10_to_fp16_dtype_0 = const()[name = tensor("position10_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor position10_to_fp16 = cast(dtype = position10_to_fp16_dtype_0, x = position10)[name = tensor("cast_503")]; + tensor ts_65_cast_fp16 = reshape(shape = var_4144, x = position10_to_fp16)[name = tensor("ts_65_cast_fp16")]; + tensor var_4151 = const()[name = tensor("op_4151"), val = tensor([1, 1, 16, 32, 2])]; + tensor q_complex_21_cast_fp16 = reshape(shape = var_4151, x = q_61_cast_fp16)[name = tensor("q_complex_21_cast_fp16")]; + tensor var_4158 = const()[name = tensor("op_4158"), val = tensor([1, 1, 16, 32, 2])]; + tensor k_complex_21_cast_fp16 = reshape(shape = var_4158, x = k_41_cast_fp16)[name = tensor("k_complex_21_cast_fp16")]; + tensor var_4162_begin_0 = const()[name = tensor("op_4162_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_4162_end_0 = const()[name = tensor("op_4162_end_0"), val = tensor([1, 1, 16, 32, 1])]; + tensor var_4162_end_mask_0 = const()[name = tensor("op_4162_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_4162_squeeze_mask_0 = const()[name = tensor("op_4162_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_4162_cast_fp16 = slice_by_index(begin = var_4162_begin_0, end = var_4162_end_0, end_mask = var_4162_end_mask_0, squeeze_mask = var_4162_squeeze_mask_0, x = q_complex_21_cast_fp16)[name = tensor("op_4162_cast_fp16")]; + tensor var_4170_begin_0 = const()[name = tensor("op_4170_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_4170_end_0 = const()[name = tensor("op_4170_end_0"), val = tensor([1, 1, 16, 32, 2])]; + tensor var_4170_end_mask_0 = const()[name = tensor("op_4170_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_4170_squeeze_mask_0 = const()[name = tensor("op_4170_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_4170_cast_fp16 = slice_by_index(begin = var_4170_begin_0, end = var_4170_end_0, end_mask = var_4170_end_mask_0, squeeze_mask = var_4170_squeeze_mask_0, x = q_complex_21_cast_fp16)[name = tensor("op_4170_cast_fp16")]; + tensor var_4178_begin_0 = const()[name = tensor("op_4178_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_4178_end_0 = const()[name = tensor("op_4178_end_0"), val = tensor([1, 1, 16, 32, 1])]; + tensor var_4178_end_mask_0 = const()[name = tensor("op_4178_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_4178_squeeze_mask_0 = const()[name = tensor("op_4178_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_4178_cast_fp16 = slice_by_index(begin = var_4178_begin_0, end = var_4178_end_0, end_mask = var_4178_end_mask_0, squeeze_mask = var_4178_squeeze_mask_0, x = k_complex_21_cast_fp16)[name = tensor("op_4178_cast_fp16")]; + tensor var_4186_begin_0 = const()[name = tensor("op_4186_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_4186_end_0 = const()[name = tensor("op_4186_end_0"), val = tensor([1, 1, 16, 32, 2])]; + tensor var_4186_end_mask_0 = const()[name = tensor("op_4186_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_4186_squeeze_mask_0 = const()[name = tensor("op_4186_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_4186_cast_fp16 = slice_by_index(begin = var_4186_begin_0, end = var_4186_end_0, end_mask = var_4186_end_mask_0, squeeze_mask = var_4186_squeeze_mask_0, x = k_complex_21_cast_fp16)[name = tensor("op_4186_cast_fp16")]; + tensor freqs_21_to_fp16 = const()[name = tensor("freqs_21_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(259174272)))]; + tensor var_4192_cast_fp16 = mul(x = freqs_21_to_fp16, y = ts_65_cast_fp16)[name = tensor("op_4192_cast_fp16")]; + tensor rotr_21_cast_fp16 = cos(x = var_4192_cast_fp16)[name = tensor("rotr_21_cast_fp16")]; + tensor roti_21_cast_fp16 = sin(x = var_4192_cast_fp16)[name = tensor("roti_21_cast_fp16")]; + tensor var_4196_cast_fp16 = mul(x = var_4162_cast_fp16, y = rotr_21_cast_fp16)[name = tensor("op_4196_cast_fp16")]; + tensor var_4197_cast_fp16 = mul(x = var_4170_cast_fp16, y = roti_21_cast_fp16)[name = tensor("op_4197_cast_fp16")]; + tensor qor_41_cast_fp16 = sub(x = var_4196_cast_fp16, y = var_4197_cast_fp16)[name = tensor("qor_41_cast_fp16")]; + tensor var_4200_cast_fp16 = mul(x = var_4162_cast_fp16, y = roti_21_cast_fp16)[name = tensor("op_4200_cast_fp16")]; + tensor var_4201_cast_fp16 = mul(x = var_4170_cast_fp16, y = rotr_21_cast_fp16)[name = tensor("op_4201_cast_fp16")]; + tensor qoi_41_cast_fp16 = add(x = var_4200_cast_fp16, y = var_4201_cast_fp16)[name = tensor("qoi_41_cast_fp16")]; + tensor var_4204_cast_fp16 = mul(x = var_4178_cast_fp16, y = rotr_21_cast_fp16)[name = tensor("op_4204_cast_fp16")]; + tensor var_4205_cast_fp16 = mul(x = var_4186_cast_fp16, y = roti_21_cast_fp16)[name = tensor("op_4205_cast_fp16")]; + tensor kor_41_cast_fp16 = sub(x = var_4204_cast_fp16, y = var_4205_cast_fp16)[name = tensor("kor_41_cast_fp16")]; + tensor var_4208_cast_fp16 = mul(x = var_4178_cast_fp16, y = roti_21_cast_fp16)[name = tensor("op_4208_cast_fp16")]; + tensor var_4209_cast_fp16 = mul(x = var_4186_cast_fp16, y = rotr_21_cast_fp16)[name = tensor("op_4209_cast_fp16")]; + tensor koi_41_cast_fp16 = add(x = var_4208_cast_fp16, y = var_4209_cast_fp16)[name = tensor("koi_41_cast_fp16")]; + tensor qo_21_axis_0 = const()[name = tensor("qo_21_axis_0"), val = tensor(-1)]; + tensor qo_21_cast_fp16 = stack(axis = qo_21_axis_0, values = (qor_41_cast_fp16, qoi_41_cast_fp16))[name = tensor("qo_21_cast_fp16")]; + tensor ko_21_axis_0 = const()[name = tensor("ko_21_axis_0"), val = tensor(-1)]; + tensor ko_21_cast_fp16 = stack(axis = ko_21_axis_0, values = (kor_41_cast_fp16, koi_41_cast_fp16))[name = tensor("ko_21_cast_fp16")]; + tensor var_4242 = const()[name = tensor("op_4242"), val = tensor([1, 1, 16, 64])]; + tensor q_63_cast_fp16 = reshape(shape = var_4242, x = qo_21_cast_fp16)[name = tensor("q_63_cast_fp16")]; + tensor var_4248 = const()[name = tensor("op_4248"), val = tensor([1, 1, 16, 64])]; + tensor k_43_cast_fp16 = reshape(shape = var_4248, x = ko_21_cast_fp16)[name = tensor("k_43_cast_fp16")]; + tensor _inversed_4273_y_0_to_fp16 = const()[name = tensor("_inversed_4273_y_0_to_fp16"), val = tensor(0x1p-9)]; + tensor _inversed_4273_cast_fp16 = mul(x = ts_65_cast_fp16, y = _inversed_4273_y_0_to_fp16)[name = tensor("_inversed_4273_cast_fp16")]; + tensor var_4274_cast_fp16 = floor(x = _inversed_4273_cast_fp16)[name = tensor("op_4274_cast_fp16")]; + tensor var_4275_to_fp16 = const()[name = tensor("op_4275_to_fp16"), val = tensor(0x1p+9)]; + tensor var_4276_cast_fp16 = mul(x = var_4274_cast_fp16, y = var_4275_to_fp16)[name = tensor("op_4276_cast_fp16")]; + tensor write_indices_float_43_cast_fp16 = sub(x = ts_65_cast_fp16, y = var_4276_cast_fp16)[name = tensor("write_indices_float_43_cast_fp16")]; + tensor var_4283_dtype_0 = const()[name = tensor("op_4283_dtype_0"), val = tensor("int32")]; + tensor write_indices_21_reps_0 = const()[name = tensor("write_indices_21_reps_0"), val = tensor([1, 1, 16, 64])]; + tensor write_indices_float_43_cast_fp16_to_int32 = cast(dtype = var_4283_dtype_0, x = write_indices_float_43_cast_fp16)[name = tensor("cast_502")]; + tensor write_indices_21 = tile(reps = write_indices_21_reps_0, x = write_indices_float_43_cast_fp16_to_int32)[name = tensor("write_indices_21")]; + tensor var_4293_begin_0 = const()[name = tensor("op_4293_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_4293_end_0 = const()[name = tensor("op_4293_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor var_4293_end_mask_0 = const()[name = tensor("op_4293_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_4293_squeeze_mask_0 = const()[name = tensor("op_4293_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor cache10_to_fp16_dtype_0 = const()[name = tensor("cache10_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor cache10_to_fp16 = cast(dtype = cache10_to_fp16_dtype_0, x = cache10)[name = tensor("cast_501")]; + tensor var_4293_cast_fp16 = slice_by_index(begin = var_4293_begin_0, end = var_4293_end_0, end_mask = var_4293_end_mask_0, squeeze_mask = var_4293_squeeze_mask_0, x = cache10_to_fp16)[name = tensor("op_4293_cast_fp16")]; + tensor var_4295_axis_0 = const()[name = tensor("op_4295_axis_0"), val = tensor(1)]; + tensor var_4295_mode_0 = const()[name = tensor("op_4295_mode_0"), val = tensor("update")]; + tensor var_4295_validate_indices_0 = const()[name = tensor("op_4295_validate_indices_0"), val = tensor(false)]; + tensor var_4295_cast_fp16 = scatter_along_axis(axis = var_4295_axis_0, data = var_4293_cast_fp16, indices = write_indices_21, mode = var_4295_mode_0, updates = k_43_cast_fp16, validate_indices = var_4295_validate_indices_0)[name = tensor("op_4295_cast_fp16")]; + tensor concat_72 = const()[name = tensor("concat_72"), val = tensor([0, 0, 0, 0, 0])]; + tensor concat_73 = const()[name = tensor("concat_73"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_21_internal_tensor_assign_1_stride_0 = const()[name = tensor("new_cache_21_internal_tensor_assign_1_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_21_internal_tensor_assign_1_begin_mask_0 = const()[name = tensor("new_cache_21_internal_tensor_assign_1_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_21_internal_tensor_assign_1_end_mask_0 = const()[name = tensor("new_cache_21_internal_tensor_assign_1_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_21_internal_tensor_assign_1_squeeze_mask_0 = const()[name = tensor("new_cache_21_internal_tensor_assign_1_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_68 = const()[name = tensor("shape_68"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_20 = const()[name = tensor("reduce_prod_20"), val = tensor(1048576)]; + tensor range_1d_20_start_0 = const()[name = tensor("range_1d_20_start_0"), val = tensor(0)]; + tensor range_1d_20_step_0 = const()[name = tensor("range_1d_20_step_0"), val = tensor(1)]; + tensor range_1d_20 = range_1d(end = reduce_prod_20, start = range_1d_20_start_0, step = range_1d_20_step_0)[name = tensor("range_1d_20")]; + tensor reshape_100 = reshape(shape = shape_68, x = range_1d_20)[name = tensor("reshape_100")]; + tensor slice_by_index_20 = slice_by_index(begin = concat_72, begin_mask = new_cache_21_internal_tensor_assign_1_begin_mask_0, end = concat_73, end_mask = new_cache_21_internal_tensor_assign_1_end_mask_0, squeeze_mask = new_cache_21_internal_tensor_assign_1_squeeze_mask_0, stride = new_cache_21_internal_tensor_assign_1_stride_0, x = reshape_100)[name = tensor("slice_by_index_20")]; + tensor reshape_101_shape_0 = const()[name = tensor("reshape_101_shape_0"), val = tensor([-1])]; + tensor reshape_101 = reshape(shape = reshape_101_shape_0, x = slice_by_index_20)[name = tensor("reshape_101")]; + tensor reshape_102_shape_0 = const()[name = tensor("reshape_102_shape_0"), val = tensor([-1])]; + tensor reshape_102_cast_fp16 = reshape(shape = reshape_102_shape_0, x = var_4295_cast_fp16)[name = tensor("reshape_102_cast_fp16")]; + tensor reshape_103_shape_0 = const()[name = tensor("reshape_103_shape_0"), val = tensor([-1])]; + tensor reshape_103_cast_fp16 = reshape(shape = reshape_103_shape_0, x = cache10_to_fp16)[name = tensor("reshape_103_cast_fp16")]; + tensor scatter_20_mode_0 = const()[name = tensor("scatter_20_mode_0"), val = tensor("update")]; + tensor scatter_20_axis_0 = const()[name = tensor("scatter_20_axis_0"), val = tensor(0)]; + tensor scatter_20_validate_indices_0 = const()[name = tensor("scatter_20_validate_indices_0"), val = tensor(false)]; + tensor scatter_20_cast_fp16 = scatter(axis = scatter_20_axis_0, data = reshape_103_cast_fp16, indices = reshape_101, mode = scatter_20_mode_0, updates = reshape_102_cast_fp16, validate_indices = scatter_20_validate_indices_0)[name = tensor("scatter_20_cast_fp16")]; + tensor reshape_104_cast_fp16 = reshape(shape = shape_68, x = scatter_20_cast_fp16)[name = tensor("reshape_104_cast_fp16")]; + tensor var_4303_begin_0 = const()[name = tensor("op_4303_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor var_4303_end_0 = const()[name = tensor("op_4303_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor var_4303_end_mask_0 = const()[name = tensor("op_4303_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_4303_squeeze_mask_0 = const()[name = tensor("op_4303_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor var_4303_cast_fp16 = slice_by_index(begin = var_4303_begin_0, end = var_4303_end_0, end_mask = var_4303_end_mask_0, squeeze_mask = var_4303_squeeze_mask_0, x = reshape_104_cast_fp16)[name = tensor("op_4303_cast_fp16")]; + tensor var_4305_axis_0 = const()[name = tensor("op_4305_axis_0"), val = tensor(1)]; + tensor var_4305_mode_0 = const()[name = tensor("op_4305_mode_0"), val = tensor("update")]; + tensor var_4305_validate_indices_0 = const()[name = tensor("op_4305_validate_indices_0"), val = tensor(false)]; + tensor var_4305_cast_fp16 = scatter_along_axis(axis = var_4305_axis_0, data = var_4303_cast_fp16, indices = write_indices_21, mode = var_4305_mode_0, updates = v_21_cast_fp16, validate_indices = var_4305_validate_indices_0)[name = tensor("op_4305_cast_fp16")]; + tensor concat_74 = const()[name = tensor("concat_74"), val = tensor([1, 0, 0, 0, 0])]; + tensor concat_75 = const()[name = tensor("concat_75"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_21_internal_tensor_assign_2_stride_0 = const()[name = tensor("new_cache_21_internal_tensor_assign_2_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_21_internal_tensor_assign_2_begin_mask_0 = const()[name = tensor("new_cache_21_internal_tensor_assign_2_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_21_internal_tensor_assign_2_end_mask_0 = const()[name = tensor("new_cache_21_internal_tensor_assign_2_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_21_internal_tensor_assign_2_squeeze_mask_0 = const()[name = tensor("new_cache_21_internal_tensor_assign_2_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_69 = const()[name = tensor("shape_69"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_21 = const()[name = tensor("reduce_prod_21"), val = tensor(1048576)]; + tensor range_1d_21_start_0 = const()[name = tensor("range_1d_21_start_0"), val = tensor(0)]; + tensor range_1d_21_step_0 = const()[name = tensor("range_1d_21_step_0"), val = tensor(1)]; + tensor range_1d_21 = range_1d(end = reduce_prod_21, start = range_1d_21_start_0, step = range_1d_21_step_0)[name = tensor("range_1d_21")]; + tensor reshape_105 = reshape(shape = shape_69, x = range_1d_21)[name = tensor("reshape_105")]; + tensor slice_by_index_21 = slice_by_index(begin = concat_74, begin_mask = new_cache_21_internal_tensor_assign_2_begin_mask_0, end = concat_75, end_mask = new_cache_21_internal_tensor_assign_2_end_mask_0, squeeze_mask = new_cache_21_internal_tensor_assign_2_squeeze_mask_0, stride = new_cache_21_internal_tensor_assign_2_stride_0, x = reshape_105)[name = tensor("slice_by_index_21")]; + tensor reshape_106_shape_0 = const()[name = tensor("reshape_106_shape_0"), val = tensor([-1])]; + tensor reshape_106 = reshape(shape = reshape_106_shape_0, x = slice_by_index_21)[name = tensor("reshape_106")]; + tensor reshape_107_shape_0 = const()[name = tensor("reshape_107_shape_0"), val = tensor([-1])]; + tensor reshape_107_cast_fp16 = reshape(shape = reshape_107_shape_0, x = var_4305_cast_fp16)[name = tensor("reshape_107_cast_fp16")]; + tensor reshape_108_shape_0 = const()[name = tensor("reshape_108_shape_0"), val = tensor([-1])]; + tensor reshape_108_cast_fp16 = reshape(shape = reshape_108_shape_0, x = reshape_104_cast_fp16)[name = tensor("reshape_108_cast_fp16")]; + tensor scatter_21_mode_0 = const()[name = tensor("scatter_21_mode_0"), val = tensor("update")]; + tensor scatter_21_axis_0 = const()[name = tensor("scatter_21_axis_0"), val = tensor(0)]; + tensor scatter_21_validate_indices_0 = const()[name = tensor("scatter_21_validate_indices_0"), val = tensor(false)]; + tensor scatter_21_cast_fp16 = scatter(axis = scatter_21_axis_0, data = reshape_108_cast_fp16, indices = reshape_106, mode = scatter_21_mode_0, updates = reshape_107_cast_fp16, validate_indices = scatter_21_validate_indices_0)[name = tensor("scatter_21_cast_fp16")]; + tensor reshape_109_cast_fp16 = reshape(shape = shape_69, x = scatter_21_cast_fp16)[name = tensor("reshape_109_cast_fp16")]; + tensor reshape_109_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("reshape_109_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor keys_61_begin_0 = const()[name = tensor("keys_61_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor keys_61_end_0 = const()[name = tensor("keys_61_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor keys_61_end_mask_0 = const()[name = tensor("keys_61_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor keys_61_squeeze_mask_0 = const()[name = tensor("keys_61_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor keys_61_cast_fp16 = slice_by_index(begin = keys_61_begin_0, end = keys_61_end_0, end_mask = keys_61_end_mask_0, squeeze_mask = keys_61_squeeze_mask_0, x = reshape_109_cast_fp16)[name = tensor("keys_61_cast_fp16")]; + tensor values_61_begin_0 = const()[name = tensor("values_61_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor values_61_end_0 = const()[name = tensor("values_61_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor values_61_end_mask_0 = const()[name = tensor("values_61_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor values_61_squeeze_mask_0 = const()[name = tensor("values_61_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor values_61_cast_fp16 = slice_by_index(begin = values_61_begin_0, end = values_61_end_0, end_mask = values_61_end_mask_0, squeeze_mask = values_61_squeeze_mask_0, x = reshape_109_cast_fp16)[name = tensor("values_61_cast_fp16")]; + tensor var_4317_cast_fp16 = not_equal(x = keys_61_cast_fp16, y = keys_61_cast_fp16)[name = tensor("op_4317_cast_fp16")]; + tensor keys_63_cast_fp16 = select(a = var_463_to_fp16, b = keys_61_cast_fp16, cond = var_4317_cast_fp16)[name = tensor("keys_63_cast_fp16")]; + tensor var_4325_cast_fp16 = not_equal(x = values_61_cast_fp16, y = values_61_cast_fp16)[name = tensor("op_4325_cast_fp16")]; + tensor values_63_cast_fp16 = select(a = var_463_to_fp16, b = values_61_cast_fp16, cond = var_4325_cast_fp16)[name = tensor("values_63_cast_fp16")]; + tensor var_4349 = const()[name = tensor("op_4349"), val = tensor([0, 2, 1, 3])]; + tensor var_4365 = const()[name = tensor("op_4365"), val = tensor([1, 1, 1])]; + tensor var_4366_cast_fp16 = reshape(shape = var_4365, x = position10_to_fp16)[name = tensor("op_4366_cast_fp16")]; + tensor var_4386_to_fp16 = const()[name = tensor("op_4386_to_fp16"), val = tensor(0x1p+0)]; + tensor valid_len_21_cast_fp16 = add(x = var_4366_cast_fp16, y = var_4386_to_fp16)[name = tensor("valid_len_21_cast_fp16")]; + tensor valid_mask_21_cast_fp16 = less(x = k_positions_1_promoted_to_fp16, y = valid_len_21_cast_fp16)[name = tensor("valid_mask_21_cast_fp16")]; + tensor causal_mask_21_cast_fp16 = less_equal(x = k_positions_1_promoted_to_fp16, y = var_4366_cast_fp16)[name = tensor("causal_mask_21_cast_fp16")]; + tensor attn_mask_41 = logical_and(x = valid_mask_21_cast_fp16, y = causal_mask_21_cast_fp16)[name = tensor("attn_mask_41")]; + tensor attn_mask_43_axes_0 = const()[name = tensor("attn_mask_43_axes_0"), val = tensor([1])]; + tensor attn_mask_43 = expand_dims(axes = attn_mask_43_axes_0, x = attn_mask_41)[name = tensor("attn_mask_43")]; + tensor var_4404_transpose_x_0 = const()[name = tensor("op_4404_transpose_x_0"), val = tensor(false)]; + tensor var_4404_transpose_y_0 = const()[name = tensor("op_4404_transpose_y_0"), val = tensor(false)]; + tensor transpose_92_perm_0 = const()[name = tensor("transpose_92_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_93_perm_0 = const()[name = tensor("transpose_93_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_93 = transpose(perm = transpose_93_perm_0, x = keys_63_cast_fp16)[name = tensor("transpose_173")]; + tensor transpose_92 = transpose(perm = transpose_92_perm_0, x = q_63_cast_fp16)[name = tensor("transpose_174")]; + tensor var_4404_cast_fp16 = matmul(transpose_x = var_4404_transpose_x_0, transpose_y = var_4404_transpose_y_0, x = transpose_92, y = transpose_93)[name = tensor("op_4404_cast_fp16")]; + tensor var_4398_to_fp16 = const()[name = tensor("op_4398_to_fp16"), val = tensor([0x1p-3])]; + tensor attn_weights_61_cast_fp16 = mul(x = var_4404_cast_fp16, y = var_4398_to_fp16)[name = tensor("attn_weights_61_cast_fp16")]; + tensor var_4406 = logical_not(x = attn_mask_43)[name = tensor("op_4406")]; + tensor var_4407_to_fp16 = const()[name = tensor("op_4407_to_fp16"), val = tensor(-inf)]; + tensor attn_weights_63_cast_fp16 = select(a = var_4407_to_fp16, b = attn_weights_61_cast_fp16, cond = var_4406)[name = tensor("attn_weights_63_cast_fp16")]; + tensor var_4409 = const()[name = tensor("op_4409"), val = tensor(-1)]; + tensor attn_weights_65_cast_fp16 = softmax(axis = var_4409, x = attn_weights_63_cast_fp16)[name = tensor("attn_weights_65_cast_fp16")]; + tensor attn_output_21_transpose_x_0 = const()[name = tensor("attn_output_21_transpose_x_0"), val = tensor(false)]; + tensor attn_output_21_transpose_y_0 = const()[name = tensor("attn_output_21_transpose_y_0"), val = tensor(false)]; + tensor values_65_cast_fp16 = transpose(perm = var_4349, x = values_63_cast_fp16)[name = tensor("transpose_175")]; + tensor attn_output_21_cast_fp16 = matmul(transpose_x = attn_output_21_transpose_x_0, transpose_y = attn_output_21_transpose_y_0, x = attn_weights_65_cast_fp16, y = values_65_cast_fp16)[name = tensor("attn_output_21_cast_fp16")]; + tensor var_4417 = const()[name = tensor("op_4417"), val = tensor([0, 2, 1, 3])]; + tensor var_4422 = const()[name = tensor("op_4422"), val = tensor([1, 1, 1024])]; + tensor var_4418_cast_fp16 = transpose(perm = var_4417, x = attn_output_21_cast_fp16)[name = tensor("transpose_172")]; + tensor input_127_cast_fp16 = reshape(shape = var_4422, x = var_4418_cast_fp16)[name = tensor("input_127_cast_fp16")]; + tensor attn10_out_proj_weight_to_fp16 = const()[name = tensor("attn10_out_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(259174400)))]; + tensor linear_42_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = attn10_out_proj_weight_to_fp16, x = input_127_cast_fp16)[name = tensor("linear_42_cast_fp16")]; + tensor var_4428_to_fp16 = const()[name = tensor("op_4428_to_fp16"), val = tensor(0x1p+0)]; + tensor var_4429_cast_fp16 = add(x = position10_to_fp16, y = var_4428_to_fp16)[name = tensor("op_4429_cast_fp16")]; + tensor var_4429_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_4429_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor input_129_cast_fp16 = add(x = input_123_cast_fp16, y = linear_42_cast_fp16)[name = tensor("input_129_cast_fp16")]; + tensor input_131_axes_0 = const()[name = tensor("input_131_axes_0"), val = tensor([-1])]; + tensor norm10_2_weight_to_fp16 = const()[name = tensor("norm10_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(261271616)))]; + tensor norm10_2_bias_to_fp16 = const()[name = tensor("norm10_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(261273728)))]; + tensor var_4433_to_fp16 = const()[name = tensor("op_4433_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_131_cast_fp16 = layer_norm(axes = input_131_axes_0, beta = norm10_2_bias_to_fp16, epsilon = var_4433_to_fp16, gamma = norm10_2_weight_to_fp16, x = input_129_cast_fp16)[name = tensor("input_131_cast_fp16")]; + tensor linear10_1_weight_to_fp16 = const()[name = tensor("linear10_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(261275840)))]; + tensor linear_43_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = linear10_1_weight_to_fp16, x = input_131_cast_fp16)[name = tensor("linear_43_cast_fp16")]; + tensor input_133_mode_0 = const()[name = tensor("input_133_mode_0"), val = tensor("EXACT")]; + tensor input_133_cast_fp16 = gelu(mode = input_133_mode_0, x = linear_43_cast_fp16)[name = tensor("input_133_cast_fp16")]; + tensor linear10_2_weight_to_fp16 = const()[name = tensor("linear10_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(269664512)))]; + tensor linear_44_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = linear10_2_weight_to_fp16, x = input_133_cast_fp16)[name = tensor("linear_44_cast_fp16")]; + tensor input_135_cast_fp16 = add(x = input_129_cast_fp16, y = linear_44_cast_fp16)[name = tensor("input_135_cast_fp16")]; + tensor input_137_axes_0 = const()[name = tensor("input_137_axes_0"), val = tensor([-1])]; + tensor norm11_1_weight_to_fp16 = const()[name = tensor("norm11_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(278053184)))]; + tensor norm11_1_bias_to_fp16 = const()[name = tensor("norm11_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(278055296)))]; + tensor var_4450_to_fp16 = const()[name = tensor("op_4450_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_137_cast_fp16 = layer_norm(axes = input_137_axes_0, beta = norm11_1_bias_to_fp16, epsilon = var_4450_to_fp16, gamma = norm11_1_weight_to_fp16, x = input_135_cast_fp16)[name = tensor("input_137_cast_fp16")]; + tensor attn11_in_proj_weight_to_fp16 = const()[name = tensor("attn11_in_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(278057408)))]; + tensor linear_45_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = attn11_in_proj_weight_to_fp16, x = input_137_cast_fp16)[name = tensor("linear_45_cast_fp16")]; + tensor var_4464 = const()[name = tensor("op_4464"), val = tensor([1, 1, 3, 16, 64])]; + tensor qkv_23_cast_fp16 = reshape(shape = var_4464, x = linear_45_cast_fp16)[name = tensor("qkv_23_cast_fp16")]; + tensor q_67_begin_0 = const()[name = tensor("q_67_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor q_67_end_0 = const()[name = tensor("q_67_end_0"), val = tensor([1, 1, 1, 16, 64])]; + tensor q_67_end_mask_0 = const()[name = tensor("q_67_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor q_67_squeeze_mask_0 = const()[name = tensor("q_67_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor q_67_cast_fp16 = slice_by_index(begin = q_67_begin_0, end = q_67_end_0, end_mask = q_67_end_mask_0, squeeze_mask = q_67_squeeze_mask_0, x = qkv_23_cast_fp16)[name = tensor("q_67_cast_fp16")]; + tensor k_45_begin_0 = const()[name = tensor("k_45_begin_0"), val = tensor([0, 0, 1, 0, 0])]; + tensor k_45_end_0 = const()[name = tensor("k_45_end_0"), val = tensor([1, 1, 2, 16, 64])]; + tensor k_45_end_mask_0 = const()[name = tensor("k_45_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor k_45_squeeze_mask_0 = const()[name = tensor("k_45_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor k_45_cast_fp16 = slice_by_index(begin = k_45_begin_0, end = k_45_end_0, end_mask = k_45_end_mask_0, squeeze_mask = k_45_squeeze_mask_0, x = qkv_23_cast_fp16)[name = tensor("k_45_cast_fp16")]; + tensor v_23_begin_0 = const()[name = tensor("v_23_begin_0"), val = tensor([0, 0, 2, 0, 0])]; + tensor v_23_end_0 = const()[name = tensor("v_23_end_0"), val = tensor([1, 1, 3, 16, 64])]; + tensor v_23_end_mask_0 = const()[name = tensor("v_23_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor v_23_squeeze_mask_0 = const()[name = tensor("v_23_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor v_23_cast_fp16 = slice_by_index(begin = v_23_begin_0, end = v_23_end_0, end_mask = v_23_end_mask_0, squeeze_mask = v_23_squeeze_mask_0, x = qkv_23_cast_fp16)[name = tensor("v_23_cast_fp16")]; + tensor var_4530 = const()[name = tensor("op_4530"), val = tensor([1, 1, 1, 1])]; + tensor position11_to_fp16_dtype_0 = const()[name = tensor("position11_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor position11_to_fp16 = cast(dtype = position11_to_fp16_dtype_0, x = position11)[name = tensor("cast_498")]; + tensor ts_71_cast_fp16 = reshape(shape = var_4530, x = position11_to_fp16)[name = tensor("ts_71_cast_fp16")]; + tensor var_4537 = const()[name = tensor("op_4537"), val = tensor([1, 1, 16, 32, 2])]; + tensor q_complex_23_cast_fp16 = reshape(shape = var_4537, x = q_67_cast_fp16)[name = tensor("q_complex_23_cast_fp16")]; + tensor var_4544 = const()[name = tensor("op_4544"), val = tensor([1, 1, 16, 32, 2])]; + tensor k_complex_23_cast_fp16 = reshape(shape = var_4544, x = k_45_cast_fp16)[name = tensor("k_complex_23_cast_fp16")]; + tensor var_4548_begin_0 = const()[name = tensor("op_4548_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_4548_end_0 = const()[name = tensor("op_4548_end_0"), val = tensor([1, 1, 16, 32, 1])]; + tensor var_4548_end_mask_0 = const()[name = tensor("op_4548_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_4548_squeeze_mask_0 = const()[name = tensor("op_4548_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_4548_cast_fp16 = slice_by_index(begin = var_4548_begin_0, end = var_4548_end_0, end_mask = var_4548_end_mask_0, squeeze_mask = var_4548_squeeze_mask_0, x = q_complex_23_cast_fp16)[name = tensor("op_4548_cast_fp16")]; + tensor var_4556_begin_0 = const()[name = tensor("op_4556_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_4556_end_0 = const()[name = tensor("op_4556_end_0"), val = tensor([1, 1, 16, 32, 2])]; + tensor var_4556_end_mask_0 = const()[name = tensor("op_4556_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_4556_squeeze_mask_0 = const()[name = tensor("op_4556_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_4556_cast_fp16 = slice_by_index(begin = var_4556_begin_0, end = var_4556_end_0, end_mask = var_4556_end_mask_0, squeeze_mask = var_4556_squeeze_mask_0, x = q_complex_23_cast_fp16)[name = tensor("op_4556_cast_fp16")]; + tensor var_4564_begin_0 = const()[name = tensor("op_4564_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_4564_end_0 = const()[name = tensor("op_4564_end_0"), val = tensor([1, 1, 16, 32, 1])]; + tensor var_4564_end_mask_0 = const()[name = tensor("op_4564_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_4564_squeeze_mask_0 = const()[name = tensor("op_4564_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_4564_cast_fp16 = slice_by_index(begin = var_4564_begin_0, end = var_4564_end_0, end_mask = var_4564_end_mask_0, squeeze_mask = var_4564_squeeze_mask_0, x = k_complex_23_cast_fp16)[name = tensor("op_4564_cast_fp16")]; + tensor var_4572_begin_0 = const()[name = tensor("op_4572_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_4572_end_0 = const()[name = tensor("op_4572_end_0"), val = tensor([1, 1, 16, 32, 2])]; + tensor var_4572_end_mask_0 = const()[name = tensor("op_4572_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_4572_squeeze_mask_0 = const()[name = tensor("op_4572_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_4572_cast_fp16 = slice_by_index(begin = var_4572_begin_0, end = var_4572_end_0, end_mask = var_4572_end_mask_0, squeeze_mask = var_4572_squeeze_mask_0, x = k_complex_23_cast_fp16)[name = tensor("op_4572_cast_fp16")]; + tensor freqs_23_to_fp16 = const()[name = tensor("freqs_23_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(284348928)))]; + tensor var_4578_cast_fp16 = mul(x = freqs_23_to_fp16, y = ts_71_cast_fp16)[name = tensor("op_4578_cast_fp16")]; + tensor rotr_23_cast_fp16 = cos(x = var_4578_cast_fp16)[name = tensor("rotr_23_cast_fp16")]; + tensor roti_23_cast_fp16 = sin(x = var_4578_cast_fp16)[name = tensor("roti_23_cast_fp16")]; + tensor var_4582_cast_fp16 = mul(x = var_4548_cast_fp16, y = rotr_23_cast_fp16)[name = tensor("op_4582_cast_fp16")]; + tensor var_4583_cast_fp16 = mul(x = var_4556_cast_fp16, y = roti_23_cast_fp16)[name = tensor("op_4583_cast_fp16")]; + tensor qor_45_cast_fp16 = sub(x = var_4582_cast_fp16, y = var_4583_cast_fp16)[name = tensor("qor_45_cast_fp16")]; + tensor var_4586_cast_fp16 = mul(x = var_4548_cast_fp16, y = roti_23_cast_fp16)[name = tensor("op_4586_cast_fp16")]; + tensor var_4587_cast_fp16 = mul(x = var_4556_cast_fp16, y = rotr_23_cast_fp16)[name = tensor("op_4587_cast_fp16")]; + tensor qoi_45_cast_fp16 = add(x = var_4586_cast_fp16, y = var_4587_cast_fp16)[name = tensor("qoi_45_cast_fp16")]; + tensor var_4590_cast_fp16 = mul(x = var_4564_cast_fp16, y = rotr_23_cast_fp16)[name = tensor("op_4590_cast_fp16")]; + tensor var_4591_cast_fp16 = mul(x = var_4572_cast_fp16, y = roti_23_cast_fp16)[name = tensor("op_4591_cast_fp16")]; + tensor kor_45_cast_fp16 = sub(x = var_4590_cast_fp16, y = var_4591_cast_fp16)[name = tensor("kor_45_cast_fp16")]; + tensor var_4594_cast_fp16 = mul(x = var_4564_cast_fp16, y = roti_23_cast_fp16)[name = tensor("op_4594_cast_fp16")]; + tensor var_4595_cast_fp16 = mul(x = var_4572_cast_fp16, y = rotr_23_cast_fp16)[name = tensor("op_4595_cast_fp16")]; + tensor koi_45_cast_fp16 = add(x = var_4594_cast_fp16, y = var_4595_cast_fp16)[name = tensor("koi_45_cast_fp16")]; + tensor qo_23_axis_0 = const()[name = tensor("qo_23_axis_0"), val = tensor(-1)]; + tensor qo_23_cast_fp16 = stack(axis = qo_23_axis_0, values = (qor_45_cast_fp16, qoi_45_cast_fp16))[name = tensor("qo_23_cast_fp16")]; + tensor ko_23_axis_0 = const()[name = tensor("ko_23_axis_0"), val = tensor(-1)]; + tensor ko_23_cast_fp16 = stack(axis = ko_23_axis_0, values = (kor_45_cast_fp16, koi_45_cast_fp16))[name = tensor("ko_23_cast_fp16")]; + tensor var_4628 = const()[name = tensor("op_4628"), val = tensor([1, 1, 16, 64])]; + tensor q_69_cast_fp16 = reshape(shape = var_4628, x = qo_23_cast_fp16)[name = tensor("q_69_cast_fp16")]; + tensor var_4634 = const()[name = tensor("op_4634"), val = tensor([1, 1, 16, 64])]; + tensor k_47_cast_fp16 = reshape(shape = var_4634, x = ko_23_cast_fp16)[name = tensor("k_47_cast_fp16")]; + tensor _inversed_4659_y_0_to_fp16 = const()[name = tensor("_inversed_4659_y_0_to_fp16"), val = tensor(0x1p-9)]; + tensor _inversed_4659_cast_fp16 = mul(x = ts_71_cast_fp16, y = _inversed_4659_y_0_to_fp16)[name = tensor("_inversed_4659_cast_fp16")]; + tensor var_4660_cast_fp16 = floor(x = _inversed_4659_cast_fp16)[name = tensor("op_4660_cast_fp16")]; + tensor var_4661_to_fp16 = const()[name = tensor("op_4661_to_fp16"), val = tensor(0x1p+9)]; + tensor var_4662_cast_fp16 = mul(x = var_4660_cast_fp16, y = var_4661_to_fp16)[name = tensor("op_4662_cast_fp16")]; + tensor write_indices_float_47_cast_fp16 = sub(x = ts_71_cast_fp16, y = var_4662_cast_fp16)[name = tensor("write_indices_float_47_cast_fp16")]; + tensor var_4669_dtype_0 = const()[name = tensor("op_4669_dtype_0"), val = tensor("int32")]; + tensor write_indices_23_reps_0 = const()[name = tensor("write_indices_23_reps_0"), val = tensor([1, 1, 16, 64])]; + tensor write_indices_float_47_cast_fp16_to_int32 = cast(dtype = var_4669_dtype_0, x = write_indices_float_47_cast_fp16)[name = tensor("cast_497")]; + tensor write_indices_23 = tile(reps = write_indices_23_reps_0, x = write_indices_float_47_cast_fp16_to_int32)[name = tensor("write_indices_23")]; + tensor var_4679_begin_0 = const()[name = tensor("op_4679_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_4679_end_0 = const()[name = tensor("op_4679_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor var_4679_end_mask_0 = const()[name = tensor("op_4679_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_4679_squeeze_mask_0 = const()[name = tensor("op_4679_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor cache11_to_fp16_dtype_0 = const()[name = tensor("cache11_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor cache11_to_fp16 = cast(dtype = cache11_to_fp16_dtype_0, x = cache11)[name = tensor("cast_496")]; + tensor var_4679_cast_fp16 = slice_by_index(begin = var_4679_begin_0, end = var_4679_end_0, end_mask = var_4679_end_mask_0, squeeze_mask = var_4679_squeeze_mask_0, x = cache11_to_fp16)[name = tensor("op_4679_cast_fp16")]; + tensor var_4681_axis_0 = const()[name = tensor("op_4681_axis_0"), val = tensor(1)]; + tensor var_4681_mode_0 = const()[name = tensor("op_4681_mode_0"), val = tensor("update")]; + tensor var_4681_validate_indices_0 = const()[name = tensor("op_4681_validate_indices_0"), val = tensor(false)]; + tensor var_4681_cast_fp16 = scatter_along_axis(axis = var_4681_axis_0, data = var_4679_cast_fp16, indices = write_indices_23, mode = var_4681_mode_0, updates = k_47_cast_fp16, validate_indices = var_4681_validate_indices_0)[name = tensor("op_4681_cast_fp16")]; + tensor concat_79 = const()[name = tensor("concat_79"), val = tensor([0, 0, 0, 0, 0])]; + tensor concat_80 = const()[name = tensor("concat_80"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_23_internal_tensor_assign_1_stride_0 = const()[name = tensor("new_cache_23_internal_tensor_assign_1_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_23_internal_tensor_assign_1_begin_mask_0 = const()[name = tensor("new_cache_23_internal_tensor_assign_1_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_23_internal_tensor_assign_1_end_mask_0 = const()[name = tensor("new_cache_23_internal_tensor_assign_1_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_23_internal_tensor_assign_1_squeeze_mask_0 = const()[name = tensor("new_cache_23_internal_tensor_assign_1_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_70 = const()[name = tensor("shape_70"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_22 = const()[name = tensor("reduce_prod_22"), val = tensor(1048576)]; + tensor range_1d_22_start_0 = const()[name = tensor("range_1d_22_start_0"), val = tensor(0)]; + tensor range_1d_22_step_0 = const()[name = tensor("range_1d_22_step_0"), val = tensor(1)]; + tensor range_1d_22 = range_1d(end = reduce_prod_22, start = range_1d_22_start_0, step = range_1d_22_step_0)[name = tensor("range_1d_22")]; + tensor reshape_110 = reshape(shape = shape_70, x = range_1d_22)[name = tensor("reshape_110")]; + tensor slice_by_index_22 = slice_by_index(begin = concat_79, begin_mask = new_cache_23_internal_tensor_assign_1_begin_mask_0, end = concat_80, end_mask = new_cache_23_internal_tensor_assign_1_end_mask_0, squeeze_mask = new_cache_23_internal_tensor_assign_1_squeeze_mask_0, stride = new_cache_23_internal_tensor_assign_1_stride_0, x = reshape_110)[name = tensor("slice_by_index_22")]; + tensor reshape_111_shape_0 = const()[name = tensor("reshape_111_shape_0"), val = tensor([-1])]; + tensor reshape_111 = reshape(shape = reshape_111_shape_0, x = slice_by_index_22)[name = tensor("reshape_111")]; + tensor reshape_112_shape_0 = const()[name = tensor("reshape_112_shape_0"), val = tensor([-1])]; + tensor reshape_112_cast_fp16 = reshape(shape = reshape_112_shape_0, x = var_4681_cast_fp16)[name = tensor("reshape_112_cast_fp16")]; + tensor reshape_113_shape_0 = const()[name = tensor("reshape_113_shape_0"), val = tensor([-1])]; + tensor reshape_113_cast_fp16 = reshape(shape = reshape_113_shape_0, x = cache11_to_fp16)[name = tensor("reshape_113_cast_fp16")]; + tensor scatter_22_mode_0 = const()[name = tensor("scatter_22_mode_0"), val = tensor("update")]; + tensor scatter_22_axis_0 = const()[name = tensor("scatter_22_axis_0"), val = tensor(0)]; + tensor scatter_22_validate_indices_0 = const()[name = tensor("scatter_22_validate_indices_0"), val = tensor(false)]; + tensor scatter_22_cast_fp16 = scatter(axis = scatter_22_axis_0, data = reshape_113_cast_fp16, indices = reshape_111, mode = scatter_22_mode_0, updates = reshape_112_cast_fp16, validate_indices = scatter_22_validate_indices_0)[name = tensor("scatter_22_cast_fp16")]; + tensor reshape_114_cast_fp16 = reshape(shape = shape_70, x = scatter_22_cast_fp16)[name = tensor("reshape_114_cast_fp16")]; + tensor var_4689_begin_0 = const()[name = tensor("op_4689_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor var_4689_end_0 = const()[name = tensor("op_4689_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor var_4689_end_mask_0 = const()[name = tensor("op_4689_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_4689_squeeze_mask_0 = const()[name = tensor("op_4689_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor var_4689_cast_fp16 = slice_by_index(begin = var_4689_begin_0, end = var_4689_end_0, end_mask = var_4689_end_mask_0, squeeze_mask = var_4689_squeeze_mask_0, x = reshape_114_cast_fp16)[name = tensor("op_4689_cast_fp16")]; + tensor var_4691_axis_0 = const()[name = tensor("op_4691_axis_0"), val = tensor(1)]; + tensor var_4691_mode_0 = const()[name = tensor("op_4691_mode_0"), val = tensor("update")]; + tensor var_4691_validate_indices_0 = const()[name = tensor("op_4691_validate_indices_0"), val = tensor(false)]; + tensor var_4691_cast_fp16 = scatter_along_axis(axis = var_4691_axis_0, data = var_4689_cast_fp16, indices = write_indices_23, mode = var_4691_mode_0, updates = v_23_cast_fp16, validate_indices = var_4691_validate_indices_0)[name = tensor("op_4691_cast_fp16")]; + tensor concat_81 = const()[name = tensor("concat_81"), val = tensor([1, 0, 0, 0, 0])]; + tensor concat_82 = const()[name = tensor("concat_82"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_23_internal_tensor_assign_2_stride_0 = const()[name = tensor("new_cache_23_internal_tensor_assign_2_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_23_internal_tensor_assign_2_begin_mask_0 = const()[name = tensor("new_cache_23_internal_tensor_assign_2_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_23_internal_tensor_assign_2_end_mask_0 = const()[name = tensor("new_cache_23_internal_tensor_assign_2_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_23_internal_tensor_assign_2_squeeze_mask_0 = const()[name = tensor("new_cache_23_internal_tensor_assign_2_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_71 = const()[name = tensor("shape_71"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_23 = const()[name = tensor("reduce_prod_23"), val = tensor(1048576)]; + tensor range_1d_23_start_0 = const()[name = tensor("range_1d_23_start_0"), val = tensor(0)]; + tensor range_1d_23_step_0 = const()[name = tensor("range_1d_23_step_0"), val = tensor(1)]; + tensor range_1d_23 = range_1d(end = reduce_prod_23, start = range_1d_23_start_0, step = range_1d_23_step_0)[name = tensor("range_1d_23")]; + tensor reshape_115 = reshape(shape = shape_71, x = range_1d_23)[name = tensor("reshape_115")]; + tensor slice_by_index_23 = slice_by_index(begin = concat_81, begin_mask = new_cache_23_internal_tensor_assign_2_begin_mask_0, end = concat_82, end_mask = new_cache_23_internal_tensor_assign_2_end_mask_0, squeeze_mask = new_cache_23_internal_tensor_assign_2_squeeze_mask_0, stride = new_cache_23_internal_tensor_assign_2_stride_0, x = reshape_115)[name = tensor("slice_by_index_23")]; + tensor reshape_116_shape_0 = const()[name = tensor("reshape_116_shape_0"), val = tensor([-1])]; + tensor reshape_116 = reshape(shape = reshape_116_shape_0, x = slice_by_index_23)[name = tensor("reshape_116")]; + tensor reshape_117_shape_0 = const()[name = tensor("reshape_117_shape_0"), val = tensor([-1])]; + tensor reshape_117_cast_fp16 = reshape(shape = reshape_117_shape_0, x = var_4691_cast_fp16)[name = tensor("reshape_117_cast_fp16")]; + tensor reshape_118_shape_0 = const()[name = tensor("reshape_118_shape_0"), val = tensor([-1])]; + tensor reshape_118_cast_fp16 = reshape(shape = reshape_118_shape_0, x = reshape_114_cast_fp16)[name = tensor("reshape_118_cast_fp16")]; + tensor scatter_23_mode_0 = const()[name = tensor("scatter_23_mode_0"), val = tensor("update")]; + tensor scatter_23_axis_0 = const()[name = tensor("scatter_23_axis_0"), val = tensor(0)]; + tensor scatter_23_validate_indices_0 = const()[name = tensor("scatter_23_validate_indices_0"), val = tensor(false)]; + tensor scatter_23_cast_fp16 = scatter(axis = scatter_23_axis_0, data = reshape_118_cast_fp16, indices = reshape_116, mode = scatter_23_mode_0, updates = reshape_117_cast_fp16, validate_indices = scatter_23_validate_indices_0)[name = tensor("scatter_23_cast_fp16")]; + tensor reshape_119_cast_fp16 = reshape(shape = shape_71, x = scatter_23_cast_fp16)[name = tensor("reshape_119_cast_fp16")]; + tensor reshape_119_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("reshape_119_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor keys_67_begin_0 = const()[name = tensor("keys_67_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor keys_67_end_0 = const()[name = tensor("keys_67_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor keys_67_end_mask_0 = const()[name = tensor("keys_67_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor keys_67_squeeze_mask_0 = const()[name = tensor("keys_67_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor keys_67_cast_fp16 = slice_by_index(begin = keys_67_begin_0, end = keys_67_end_0, end_mask = keys_67_end_mask_0, squeeze_mask = keys_67_squeeze_mask_0, x = reshape_119_cast_fp16)[name = tensor("keys_67_cast_fp16")]; + tensor values_67_begin_0 = const()[name = tensor("values_67_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor values_67_end_0 = const()[name = tensor("values_67_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor values_67_end_mask_0 = const()[name = tensor("values_67_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor values_67_squeeze_mask_0 = const()[name = tensor("values_67_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor values_67_cast_fp16 = slice_by_index(begin = values_67_begin_0, end = values_67_end_0, end_mask = values_67_end_mask_0, squeeze_mask = values_67_squeeze_mask_0, x = reshape_119_cast_fp16)[name = tensor("values_67_cast_fp16")]; + tensor var_4703_cast_fp16 = not_equal(x = keys_67_cast_fp16, y = keys_67_cast_fp16)[name = tensor("op_4703_cast_fp16")]; + tensor keys_69_cast_fp16 = select(a = var_463_to_fp16, b = keys_67_cast_fp16, cond = var_4703_cast_fp16)[name = tensor("keys_69_cast_fp16")]; + tensor var_4711_cast_fp16 = not_equal(x = values_67_cast_fp16, y = values_67_cast_fp16)[name = tensor("op_4711_cast_fp16")]; + tensor values_69_cast_fp16 = select(a = var_463_to_fp16, b = values_67_cast_fp16, cond = var_4711_cast_fp16)[name = tensor("values_69_cast_fp16")]; + tensor var_4735 = const()[name = tensor("op_4735"), val = tensor([0, 2, 1, 3])]; + tensor var_4751 = const()[name = tensor("op_4751"), val = tensor([1, 1, 1])]; + tensor var_4752_cast_fp16 = reshape(shape = var_4751, x = position11_to_fp16)[name = tensor("op_4752_cast_fp16")]; + tensor var_4772_to_fp16 = const()[name = tensor("op_4772_to_fp16"), val = tensor(0x1p+0)]; + tensor valid_len_23_cast_fp16 = add(x = var_4752_cast_fp16, y = var_4772_to_fp16)[name = tensor("valid_len_23_cast_fp16")]; + tensor valid_mask_23_cast_fp16 = less(x = k_positions_1_promoted_to_fp16, y = valid_len_23_cast_fp16)[name = tensor("valid_mask_23_cast_fp16")]; + tensor causal_mask_23_cast_fp16 = less_equal(x = k_positions_1_promoted_to_fp16, y = var_4752_cast_fp16)[name = tensor("causal_mask_23_cast_fp16")]; + tensor attn_mask_45 = logical_and(x = valid_mask_23_cast_fp16, y = causal_mask_23_cast_fp16)[name = tensor("attn_mask_45")]; + tensor attn_mask_47_axes_0 = const()[name = tensor("attn_mask_47_axes_0"), val = tensor([1])]; + tensor attn_mask_47 = expand_dims(axes = attn_mask_47_axes_0, x = attn_mask_45)[name = tensor("attn_mask_47")]; + tensor var_4790_transpose_x_0 = const()[name = tensor("op_4790_transpose_x_0"), val = tensor(false)]; + tensor var_4790_transpose_y_0 = const()[name = tensor("op_4790_transpose_y_0"), val = tensor(false)]; + tensor transpose_94_perm_0 = const()[name = tensor("transpose_94_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_95_perm_0 = const()[name = tensor("transpose_95_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_95 = transpose(perm = transpose_95_perm_0, x = keys_69_cast_fp16)[name = tensor("transpose_169")]; + tensor transpose_94 = transpose(perm = transpose_94_perm_0, x = q_69_cast_fp16)[name = tensor("transpose_170")]; + tensor var_4790_cast_fp16 = matmul(transpose_x = var_4790_transpose_x_0, transpose_y = var_4790_transpose_y_0, x = transpose_94, y = transpose_95)[name = tensor("op_4790_cast_fp16")]; + tensor var_4784_to_fp16 = const()[name = tensor("op_4784_to_fp16"), val = tensor([0x1p-3])]; + tensor attn_weights_67_cast_fp16 = mul(x = var_4790_cast_fp16, y = var_4784_to_fp16)[name = tensor("attn_weights_67_cast_fp16")]; + tensor var_4792 = logical_not(x = attn_mask_47)[name = tensor("op_4792")]; + tensor var_4793_to_fp16 = const()[name = tensor("op_4793_to_fp16"), val = tensor(-inf)]; + tensor attn_weights_69_cast_fp16 = select(a = var_4793_to_fp16, b = attn_weights_67_cast_fp16, cond = var_4792)[name = tensor("attn_weights_69_cast_fp16")]; + tensor var_4795 = const()[name = tensor("op_4795"), val = tensor(-1)]; + tensor attn_weights_71_cast_fp16 = softmax(axis = var_4795, x = attn_weights_69_cast_fp16)[name = tensor("attn_weights_71_cast_fp16")]; + tensor attn_output_23_transpose_x_0 = const()[name = tensor("attn_output_23_transpose_x_0"), val = tensor(false)]; + tensor attn_output_23_transpose_y_0 = const()[name = tensor("attn_output_23_transpose_y_0"), val = tensor(false)]; + tensor values_71_cast_fp16 = transpose(perm = var_4735, x = values_69_cast_fp16)[name = tensor("transpose_171")]; + tensor attn_output_23_cast_fp16 = matmul(transpose_x = attn_output_23_transpose_x_0, transpose_y = attn_output_23_transpose_y_0, x = attn_weights_71_cast_fp16, y = values_71_cast_fp16)[name = tensor("attn_output_23_cast_fp16")]; + tensor var_4803 = const()[name = tensor("op_4803"), val = tensor([0, 2, 1, 3])]; + tensor var_4808 = const()[name = tensor("op_4808"), val = tensor([1, 1, 1024])]; + tensor var_4804_cast_fp16 = transpose(perm = var_4803, x = attn_output_23_cast_fp16)[name = tensor("transpose_168")]; + tensor input_139_cast_fp16 = reshape(shape = var_4808, x = var_4804_cast_fp16)[name = tensor("input_139_cast_fp16")]; + tensor attn11_out_proj_weight_to_fp16 = const()[name = tensor("attn11_out_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(284349056)))]; + tensor linear_46_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = attn11_out_proj_weight_to_fp16, x = input_139_cast_fp16)[name = tensor("linear_46_cast_fp16")]; + tensor var_4814_to_fp16 = const()[name = tensor("op_4814_to_fp16"), val = tensor(0x1p+0)]; + tensor var_4815_cast_fp16 = add(x = position11_to_fp16, y = var_4814_to_fp16)[name = tensor("op_4815_cast_fp16")]; + tensor var_4815_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_4815_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor input_141_cast_fp16 = add(x = input_135_cast_fp16, y = linear_46_cast_fp16)[name = tensor("input_141_cast_fp16")]; + tensor input_143_axes_0 = const()[name = tensor("input_143_axes_0"), val = tensor([-1])]; + tensor norm11_2_weight_to_fp16 = const()[name = tensor("norm11_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(286446272)))]; + tensor norm11_2_bias_to_fp16 = const()[name = tensor("norm11_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(286448384)))]; + tensor var_4819_to_fp16 = const()[name = tensor("op_4819_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_143_cast_fp16 = layer_norm(axes = input_143_axes_0, beta = norm11_2_bias_to_fp16, epsilon = var_4819_to_fp16, gamma = norm11_2_weight_to_fp16, x = input_141_cast_fp16)[name = tensor("input_143_cast_fp16")]; + tensor linear11_1_weight_to_fp16 = const()[name = tensor("linear11_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(286450496)))]; + tensor linear_47_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = linear11_1_weight_to_fp16, x = input_143_cast_fp16)[name = tensor("linear_47_cast_fp16")]; + tensor input_145_mode_0 = const()[name = tensor("input_145_mode_0"), val = tensor("EXACT")]; + tensor input_145_cast_fp16 = gelu(mode = input_145_mode_0, x = linear_47_cast_fp16)[name = tensor("input_145_cast_fp16")]; + tensor linear11_2_weight_to_fp16 = const()[name = tensor("linear11_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(294839168)))]; + tensor linear_48_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = linear11_2_weight_to_fp16, x = input_145_cast_fp16)[name = tensor("linear_48_cast_fp16")]; + tensor input_147_cast_fp16 = add(x = input_141_cast_fp16, y = linear_48_cast_fp16)[name = tensor("input_147_cast_fp16")]; + tensor input_149_axes_0 = const()[name = tensor("input_149_axes_0"), val = tensor([-1])]; + tensor norm12_1_weight_to_fp16 = const()[name = tensor("norm12_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(303227840)))]; + tensor norm12_1_bias_to_fp16 = const()[name = tensor("norm12_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(303229952)))]; + tensor var_4836_to_fp16 = const()[name = tensor("op_4836_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_149_cast_fp16 = layer_norm(axes = input_149_axes_0, beta = norm12_1_bias_to_fp16, epsilon = var_4836_to_fp16, gamma = norm12_1_weight_to_fp16, x = input_147_cast_fp16)[name = tensor("input_149_cast_fp16")]; + tensor attn12_in_proj_weight_to_fp16 = const()[name = tensor("attn12_in_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(303232064)))]; + tensor linear_49_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = attn12_in_proj_weight_to_fp16, x = input_149_cast_fp16)[name = tensor("linear_49_cast_fp16")]; + tensor var_4850 = const()[name = tensor("op_4850"), val = tensor([1, 1, 3, 16, 64])]; + tensor qkv_25_cast_fp16 = reshape(shape = var_4850, x = linear_49_cast_fp16)[name = tensor("qkv_25_cast_fp16")]; + tensor q_73_begin_0 = const()[name = tensor("q_73_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor q_73_end_0 = const()[name = tensor("q_73_end_0"), val = tensor([1, 1, 1, 16, 64])]; + tensor q_73_end_mask_0 = const()[name = tensor("q_73_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor q_73_squeeze_mask_0 = const()[name = tensor("q_73_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor q_73_cast_fp16 = slice_by_index(begin = q_73_begin_0, end = q_73_end_0, end_mask = q_73_end_mask_0, squeeze_mask = q_73_squeeze_mask_0, x = qkv_25_cast_fp16)[name = tensor("q_73_cast_fp16")]; + tensor k_49_begin_0 = const()[name = tensor("k_49_begin_0"), val = tensor([0, 0, 1, 0, 0])]; + tensor k_49_end_0 = const()[name = tensor("k_49_end_0"), val = tensor([1, 1, 2, 16, 64])]; + tensor k_49_end_mask_0 = const()[name = tensor("k_49_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor k_49_squeeze_mask_0 = const()[name = tensor("k_49_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor k_49_cast_fp16 = slice_by_index(begin = k_49_begin_0, end = k_49_end_0, end_mask = k_49_end_mask_0, squeeze_mask = k_49_squeeze_mask_0, x = qkv_25_cast_fp16)[name = tensor("k_49_cast_fp16")]; + tensor v_25_begin_0 = const()[name = tensor("v_25_begin_0"), val = tensor([0, 0, 2, 0, 0])]; + tensor v_25_end_0 = const()[name = tensor("v_25_end_0"), val = tensor([1, 1, 3, 16, 64])]; + tensor v_25_end_mask_0 = const()[name = tensor("v_25_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor v_25_squeeze_mask_0 = const()[name = tensor("v_25_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor v_25_cast_fp16 = slice_by_index(begin = v_25_begin_0, end = v_25_end_0, end_mask = v_25_end_mask_0, squeeze_mask = v_25_squeeze_mask_0, x = qkv_25_cast_fp16)[name = tensor("v_25_cast_fp16")]; + tensor var_4916 = const()[name = tensor("op_4916"), val = tensor([1, 1, 1, 1])]; + tensor position12_to_fp16_dtype_0 = const()[name = tensor("position12_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor position12_to_fp16 = cast(dtype = position12_to_fp16_dtype_0, x = position12)[name = tensor("cast_493")]; + tensor ts_77_cast_fp16 = reshape(shape = var_4916, x = position12_to_fp16)[name = tensor("ts_77_cast_fp16")]; + tensor var_4923 = const()[name = tensor("op_4923"), val = tensor([1, 1, 16, 32, 2])]; + tensor q_complex_25_cast_fp16 = reshape(shape = var_4923, x = q_73_cast_fp16)[name = tensor("q_complex_25_cast_fp16")]; + tensor var_4930 = const()[name = tensor("op_4930"), val = tensor([1, 1, 16, 32, 2])]; + tensor k_complex_25_cast_fp16 = reshape(shape = var_4930, x = k_49_cast_fp16)[name = tensor("k_complex_25_cast_fp16")]; + tensor var_4934_begin_0 = const()[name = tensor("op_4934_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_4934_end_0 = const()[name = tensor("op_4934_end_0"), val = tensor([1, 1, 16, 32, 1])]; + tensor var_4934_end_mask_0 = const()[name = tensor("op_4934_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_4934_squeeze_mask_0 = const()[name = tensor("op_4934_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_4934_cast_fp16 = slice_by_index(begin = var_4934_begin_0, end = var_4934_end_0, end_mask = var_4934_end_mask_0, squeeze_mask = var_4934_squeeze_mask_0, x = q_complex_25_cast_fp16)[name = tensor("op_4934_cast_fp16")]; + tensor var_4942_begin_0 = const()[name = tensor("op_4942_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_4942_end_0 = const()[name = tensor("op_4942_end_0"), val = tensor([1, 1, 16, 32, 2])]; + tensor var_4942_end_mask_0 = const()[name = tensor("op_4942_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_4942_squeeze_mask_0 = const()[name = tensor("op_4942_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_4942_cast_fp16 = slice_by_index(begin = var_4942_begin_0, end = var_4942_end_0, end_mask = var_4942_end_mask_0, squeeze_mask = var_4942_squeeze_mask_0, x = q_complex_25_cast_fp16)[name = tensor("op_4942_cast_fp16")]; + tensor var_4950_begin_0 = const()[name = tensor("op_4950_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_4950_end_0 = const()[name = tensor("op_4950_end_0"), val = tensor([1, 1, 16, 32, 1])]; + tensor var_4950_end_mask_0 = const()[name = tensor("op_4950_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_4950_squeeze_mask_0 = const()[name = tensor("op_4950_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_4950_cast_fp16 = slice_by_index(begin = var_4950_begin_0, end = var_4950_end_0, end_mask = var_4950_end_mask_0, squeeze_mask = var_4950_squeeze_mask_0, x = k_complex_25_cast_fp16)[name = tensor("op_4950_cast_fp16")]; + tensor var_4958_begin_0 = const()[name = tensor("op_4958_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_4958_end_0 = const()[name = tensor("op_4958_end_0"), val = tensor([1, 1, 16, 32, 2])]; + tensor var_4958_end_mask_0 = const()[name = tensor("op_4958_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_4958_squeeze_mask_0 = const()[name = tensor("op_4958_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_4958_cast_fp16 = slice_by_index(begin = var_4958_begin_0, end = var_4958_end_0, end_mask = var_4958_end_mask_0, squeeze_mask = var_4958_squeeze_mask_0, x = k_complex_25_cast_fp16)[name = tensor("op_4958_cast_fp16")]; + tensor freqs_25_to_fp16 = const()[name = tensor("freqs_25_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(309523584)))]; + tensor var_4964_cast_fp16 = mul(x = freqs_25_to_fp16, y = ts_77_cast_fp16)[name = tensor("op_4964_cast_fp16")]; + tensor rotr_25_cast_fp16 = cos(x = var_4964_cast_fp16)[name = tensor("rotr_25_cast_fp16")]; + tensor roti_25_cast_fp16 = sin(x = var_4964_cast_fp16)[name = tensor("roti_25_cast_fp16")]; + tensor var_4968_cast_fp16 = mul(x = var_4934_cast_fp16, y = rotr_25_cast_fp16)[name = tensor("op_4968_cast_fp16")]; + tensor var_4969_cast_fp16 = mul(x = var_4942_cast_fp16, y = roti_25_cast_fp16)[name = tensor("op_4969_cast_fp16")]; + tensor qor_49_cast_fp16 = sub(x = var_4968_cast_fp16, y = var_4969_cast_fp16)[name = tensor("qor_49_cast_fp16")]; + tensor var_4972_cast_fp16 = mul(x = var_4934_cast_fp16, y = roti_25_cast_fp16)[name = tensor("op_4972_cast_fp16")]; + tensor var_4973_cast_fp16 = mul(x = var_4942_cast_fp16, y = rotr_25_cast_fp16)[name = tensor("op_4973_cast_fp16")]; + tensor qoi_49_cast_fp16 = add(x = var_4972_cast_fp16, y = var_4973_cast_fp16)[name = tensor("qoi_49_cast_fp16")]; + tensor var_4976_cast_fp16 = mul(x = var_4950_cast_fp16, y = rotr_25_cast_fp16)[name = tensor("op_4976_cast_fp16")]; + tensor var_4977_cast_fp16 = mul(x = var_4958_cast_fp16, y = roti_25_cast_fp16)[name = tensor("op_4977_cast_fp16")]; + tensor kor_49_cast_fp16 = sub(x = var_4976_cast_fp16, y = var_4977_cast_fp16)[name = tensor("kor_49_cast_fp16")]; + tensor var_4980_cast_fp16 = mul(x = var_4950_cast_fp16, y = roti_25_cast_fp16)[name = tensor("op_4980_cast_fp16")]; + tensor var_4981_cast_fp16 = mul(x = var_4958_cast_fp16, y = rotr_25_cast_fp16)[name = tensor("op_4981_cast_fp16")]; + tensor koi_49_cast_fp16 = add(x = var_4980_cast_fp16, y = var_4981_cast_fp16)[name = tensor("koi_49_cast_fp16")]; + tensor qo_25_axis_0 = const()[name = tensor("qo_25_axis_0"), val = tensor(-1)]; + tensor qo_25_cast_fp16 = stack(axis = qo_25_axis_0, values = (qor_49_cast_fp16, qoi_49_cast_fp16))[name = tensor("qo_25_cast_fp16")]; + tensor ko_25_axis_0 = const()[name = tensor("ko_25_axis_0"), val = tensor(-1)]; + tensor ko_25_cast_fp16 = stack(axis = ko_25_axis_0, values = (kor_49_cast_fp16, koi_49_cast_fp16))[name = tensor("ko_25_cast_fp16")]; + tensor var_5014 = const()[name = tensor("op_5014"), val = tensor([1, 1, 16, 64])]; + tensor q_75_cast_fp16 = reshape(shape = var_5014, x = qo_25_cast_fp16)[name = tensor("q_75_cast_fp16")]; + tensor var_5020 = const()[name = tensor("op_5020"), val = tensor([1, 1, 16, 64])]; + tensor k_51_cast_fp16 = reshape(shape = var_5020, x = ko_25_cast_fp16)[name = tensor("k_51_cast_fp16")]; + tensor _inversed_5045_y_0_to_fp16 = const()[name = tensor("_inversed_5045_y_0_to_fp16"), val = tensor(0x1p-9)]; + tensor _inversed_5045_cast_fp16 = mul(x = ts_77_cast_fp16, y = _inversed_5045_y_0_to_fp16)[name = tensor("_inversed_5045_cast_fp16")]; + tensor var_5046_cast_fp16 = floor(x = _inversed_5045_cast_fp16)[name = tensor("op_5046_cast_fp16")]; + tensor var_5047_to_fp16 = const()[name = tensor("op_5047_to_fp16"), val = tensor(0x1p+9)]; + tensor var_5048_cast_fp16 = mul(x = var_5046_cast_fp16, y = var_5047_to_fp16)[name = tensor("op_5048_cast_fp16")]; + tensor write_indices_float_51_cast_fp16 = sub(x = ts_77_cast_fp16, y = var_5048_cast_fp16)[name = tensor("write_indices_float_51_cast_fp16")]; + tensor var_5055_dtype_0 = const()[name = tensor("op_5055_dtype_0"), val = tensor("int32")]; + tensor write_indices_25_reps_0 = const()[name = tensor("write_indices_25_reps_0"), val = tensor([1, 1, 16, 64])]; + tensor write_indices_float_51_cast_fp16_to_int32 = cast(dtype = var_5055_dtype_0, x = write_indices_float_51_cast_fp16)[name = tensor("cast_492")]; + tensor write_indices_25 = tile(reps = write_indices_25_reps_0, x = write_indices_float_51_cast_fp16_to_int32)[name = tensor("write_indices_25")]; + tensor var_5065_begin_0 = const()[name = tensor("op_5065_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_5065_end_0 = const()[name = tensor("op_5065_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor var_5065_end_mask_0 = const()[name = tensor("op_5065_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_5065_squeeze_mask_0 = const()[name = tensor("op_5065_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor cache12_to_fp16_dtype_0 = const()[name = tensor("cache12_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor cache12_to_fp16 = cast(dtype = cache12_to_fp16_dtype_0, x = cache12)[name = tensor("cast_491")]; + tensor var_5065_cast_fp16 = slice_by_index(begin = var_5065_begin_0, end = var_5065_end_0, end_mask = var_5065_end_mask_0, squeeze_mask = var_5065_squeeze_mask_0, x = cache12_to_fp16)[name = tensor("op_5065_cast_fp16")]; + tensor var_5067_axis_0 = const()[name = tensor("op_5067_axis_0"), val = tensor(1)]; + tensor var_5067_mode_0 = const()[name = tensor("op_5067_mode_0"), val = tensor("update")]; + tensor var_5067_validate_indices_0 = const()[name = tensor("op_5067_validate_indices_0"), val = tensor(false)]; + tensor var_5067_cast_fp16 = scatter_along_axis(axis = var_5067_axis_0, data = var_5065_cast_fp16, indices = write_indices_25, mode = var_5067_mode_0, updates = k_51_cast_fp16, validate_indices = var_5067_validate_indices_0)[name = tensor("op_5067_cast_fp16")]; + tensor concat_86 = const()[name = tensor("concat_86"), val = tensor([0, 0, 0, 0, 0])]; + tensor concat_87 = const()[name = tensor("concat_87"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_25_internal_tensor_assign_1_stride_0 = const()[name = tensor("new_cache_25_internal_tensor_assign_1_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_25_internal_tensor_assign_1_begin_mask_0 = const()[name = tensor("new_cache_25_internal_tensor_assign_1_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_25_internal_tensor_assign_1_end_mask_0 = const()[name = tensor("new_cache_25_internal_tensor_assign_1_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_25_internal_tensor_assign_1_squeeze_mask_0 = const()[name = tensor("new_cache_25_internal_tensor_assign_1_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_72 = const()[name = tensor("shape_72"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_24 = const()[name = tensor("reduce_prod_24"), val = tensor(1048576)]; + tensor range_1d_24_start_0 = const()[name = tensor("range_1d_24_start_0"), val = tensor(0)]; + tensor range_1d_24_step_0 = const()[name = tensor("range_1d_24_step_0"), val = tensor(1)]; + tensor range_1d_24 = range_1d(end = reduce_prod_24, start = range_1d_24_start_0, step = range_1d_24_step_0)[name = tensor("range_1d_24")]; + tensor reshape_120 = reshape(shape = shape_72, x = range_1d_24)[name = tensor("reshape_120")]; + tensor slice_by_index_24 = slice_by_index(begin = concat_86, begin_mask = new_cache_25_internal_tensor_assign_1_begin_mask_0, end = concat_87, end_mask = new_cache_25_internal_tensor_assign_1_end_mask_0, squeeze_mask = new_cache_25_internal_tensor_assign_1_squeeze_mask_0, stride = new_cache_25_internal_tensor_assign_1_stride_0, x = reshape_120)[name = tensor("slice_by_index_24")]; + tensor reshape_121_shape_0 = const()[name = tensor("reshape_121_shape_0"), val = tensor([-1])]; + tensor reshape_121 = reshape(shape = reshape_121_shape_0, x = slice_by_index_24)[name = tensor("reshape_121")]; + tensor reshape_122_shape_0 = const()[name = tensor("reshape_122_shape_0"), val = tensor([-1])]; + tensor reshape_122_cast_fp16 = reshape(shape = reshape_122_shape_0, x = var_5067_cast_fp16)[name = tensor("reshape_122_cast_fp16")]; + tensor reshape_123_shape_0 = const()[name = tensor("reshape_123_shape_0"), val = tensor([-1])]; + tensor reshape_123_cast_fp16 = reshape(shape = reshape_123_shape_0, x = cache12_to_fp16)[name = tensor("reshape_123_cast_fp16")]; + tensor scatter_24_mode_0 = const()[name = tensor("scatter_24_mode_0"), val = tensor("update")]; + tensor scatter_24_axis_0 = const()[name = tensor("scatter_24_axis_0"), val = tensor(0)]; + tensor scatter_24_validate_indices_0 = const()[name = tensor("scatter_24_validate_indices_0"), val = tensor(false)]; + tensor scatter_24_cast_fp16 = scatter(axis = scatter_24_axis_0, data = reshape_123_cast_fp16, indices = reshape_121, mode = scatter_24_mode_0, updates = reshape_122_cast_fp16, validate_indices = scatter_24_validate_indices_0)[name = tensor("scatter_24_cast_fp16")]; + tensor reshape_124_cast_fp16 = reshape(shape = shape_72, x = scatter_24_cast_fp16)[name = tensor("reshape_124_cast_fp16")]; + tensor var_5075_begin_0 = const()[name = tensor("op_5075_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor var_5075_end_0 = const()[name = tensor("op_5075_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor var_5075_end_mask_0 = const()[name = tensor("op_5075_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_5075_squeeze_mask_0 = const()[name = tensor("op_5075_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor var_5075_cast_fp16 = slice_by_index(begin = var_5075_begin_0, end = var_5075_end_0, end_mask = var_5075_end_mask_0, squeeze_mask = var_5075_squeeze_mask_0, x = reshape_124_cast_fp16)[name = tensor("op_5075_cast_fp16")]; + tensor var_5077_axis_0 = const()[name = tensor("op_5077_axis_0"), val = tensor(1)]; + tensor var_5077_mode_0 = const()[name = tensor("op_5077_mode_0"), val = tensor("update")]; + tensor var_5077_validate_indices_0 = const()[name = tensor("op_5077_validate_indices_0"), val = tensor(false)]; + tensor var_5077_cast_fp16 = scatter_along_axis(axis = var_5077_axis_0, data = var_5075_cast_fp16, indices = write_indices_25, mode = var_5077_mode_0, updates = v_25_cast_fp16, validate_indices = var_5077_validate_indices_0)[name = tensor("op_5077_cast_fp16")]; + tensor concat_88 = const()[name = tensor("concat_88"), val = tensor([1, 0, 0, 0, 0])]; + tensor concat_89 = const()[name = tensor("concat_89"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_25_internal_tensor_assign_2_stride_0 = const()[name = tensor("new_cache_25_internal_tensor_assign_2_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_25_internal_tensor_assign_2_begin_mask_0 = const()[name = tensor("new_cache_25_internal_tensor_assign_2_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_25_internal_tensor_assign_2_end_mask_0 = const()[name = tensor("new_cache_25_internal_tensor_assign_2_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_25_internal_tensor_assign_2_squeeze_mask_0 = const()[name = tensor("new_cache_25_internal_tensor_assign_2_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_73 = const()[name = tensor("shape_73"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_25 = const()[name = tensor("reduce_prod_25"), val = tensor(1048576)]; + tensor range_1d_25_start_0 = const()[name = tensor("range_1d_25_start_0"), val = tensor(0)]; + tensor range_1d_25_step_0 = const()[name = tensor("range_1d_25_step_0"), val = tensor(1)]; + tensor range_1d_25 = range_1d(end = reduce_prod_25, start = range_1d_25_start_0, step = range_1d_25_step_0)[name = tensor("range_1d_25")]; + tensor reshape_125 = reshape(shape = shape_73, x = range_1d_25)[name = tensor("reshape_125")]; + tensor slice_by_index_25 = slice_by_index(begin = concat_88, begin_mask = new_cache_25_internal_tensor_assign_2_begin_mask_0, end = concat_89, end_mask = new_cache_25_internal_tensor_assign_2_end_mask_0, squeeze_mask = new_cache_25_internal_tensor_assign_2_squeeze_mask_0, stride = new_cache_25_internal_tensor_assign_2_stride_0, x = reshape_125)[name = tensor("slice_by_index_25")]; + tensor reshape_126_shape_0 = const()[name = tensor("reshape_126_shape_0"), val = tensor([-1])]; + tensor reshape_126 = reshape(shape = reshape_126_shape_0, x = slice_by_index_25)[name = tensor("reshape_126")]; + tensor reshape_127_shape_0 = const()[name = tensor("reshape_127_shape_0"), val = tensor([-1])]; + tensor reshape_127_cast_fp16 = reshape(shape = reshape_127_shape_0, x = var_5077_cast_fp16)[name = tensor("reshape_127_cast_fp16")]; + tensor reshape_128_shape_0 = const()[name = tensor("reshape_128_shape_0"), val = tensor([-1])]; + tensor reshape_128_cast_fp16 = reshape(shape = reshape_128_shape_0, x = reshape_124_cast_fp16)[name = tensor("reshape_128_cast_fp16")]; + tensor scatter_25_mode_0 = const()[name = tensor("scatter_25_mode_0"), val = tensor("update")]; + tensor scatter_25_axis_0 = const()[name = tensor("scatter_25_axis_0"), val = tensor(0)]; + tensor scatter_25_validate_indices_0 = const()[name = tensor("scatter_25_validate_indices_0"), val = tensor(false)]; + tensor scatter_25_cast_fp16 = scatter(axis = scatter_25_axis_0, data = reshape_128_cast_fp16, indices = reshape_126, mode = scatter_25_mode_0, updates = reshape_127_cast_fp16, validate_indices = scatter_25_validate_indices_0)[name = tensor("scatter_25_cast_fp16")]; + tensor reshape_129_cast_fp16 = reshape(shape = shape_73, x = scatter_25_cast_fp16)[name = tensor("reshape_129_cast_fp16")]; + tensor reshape_129_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("reshape_129_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor keys_73_begin_0 = const()[name = tensor("keys_73_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor keys_73_end_0 = const()[name = tensor("keys_73_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor keys_73_end_mask_0 = const()[name = tensor("keys_73_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor keys_73_squeeze_mask_0 = const()[name = tensor("keys_73_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor keys_73_cast_fp16 = slice_by_index(begin = keys_73_begin_0, end = keys_73_end_0, end_mask = keys_73_end_mask_0, squeeze_mask = keys_73_squeeze_mask_0, x = reshape_129_cast_fp16)[name = tensor("keys_73_cast_fp16")]; + tensor values_73_begin_0 = const()[name = tensor("values_73_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor values_73_end_0 = const()[name = tensor("values_73_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor values_73_end_mask_0 = const()[name = tensor("values_73_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor values_73_squeeze_mask_0 = const()[name = tensor("values_73_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor values_73_cast_fp16 = slice_by_index(begin = values_73_begin_0, end = values_73_end_0, end_mask = values_73_end_mask_0, squeeze_mask = values_73_squeeze_mask_0, x = reshape_129_cast_fp16)[name = tensor("values_73_cast_fp16")]; + tensor var_5089_cast_fp16 = not_equal(x = keys_73_cast_fp16, y = keys_73_cast_fp16)[name = tensor("op_5089_cast_fp16")]; + tensor keys_75_cast_fp16 = select(a = var_463_to_fp16, b = keys_73_cast_fp16, cond = var_5089_cast_fp16)[name = tensor("keys_75_cast_fp16")]; + tensor var_5097_cast_fp16 = not_equal(x = values_73_cast_fp16, y = values_73_cast_fp16)[name = tensor("op_5097_cast_fp16")]; + tensor values_75_cast_fp16 = select(a = var_463_to_fp16, b = values_73_cast_fp16, cond = var_5097_cast_fp16)[name = tensor("values_75_cast_fp16")]; + tensor var_5121 = const()[name = tensor("op_5121"), val = tensor([0, 2, 1, 3])]; + tensor var_5137 = const()[name = tensor("op_5137"), val = tensor([1, 1, 1])]; + tensor var_5138_cast_fp16 = reshape(shape = var_5137, x = position12_to_fp16)[name = tensor("op_5138_cast_fp16")]; + tensor var_5158_to_fp16 = const()[name = tensor("op_5158_to_fp16"), val = tensor(0x1p+0)]; + tensor valid_len_25_cast_fp16 = add(x = var_5138_cast_fp16, y = var_5158_to_fp16)[name = tensor("valid_len_25_cast_fp16")]; + tensor valid_mask_25_cast_fp16 = less(x = k_positions_1_promoted_to_fp16, y = valid_len_25_cast_fp16)[name = tensor("valid_mask_25_cast_fp16")]; + tensor causal_mask_25_cast_fp16 = less_equal(x = k_positions_1_promoted_to_fp16, y = var_5138_cast_fp16)[name = tensor("causal_mask_25_cast_fp16")]; + tensor attn_mask_49 = logical_and(x = valid_mask_25_cast_fp16, y = causal_mask_25_cast_fp16)[name = tensor("attn_mask_49")]; + tensor attn_mask_51_axes_0 = const()[name = tensor("attn_mask_51_axes_0"), val = tensor([1])]; + tensor attn_mask_51 = expand_dims(axes = attn_mask_51_axes_0, x = attn_mask_49)[name = tensor("attn_mask_51")]; + tensor var_5176_transpose_x_0 = const()[name = tensor("op_5176_transpose_x_0"), val = tensor(false)]; + tensor var_5176_transpose_y_0 = const()[name = tensor("op_5176_transpose_y_0"), val = tensor(false)]; + tensor transpose_96_perm_0 = const()[name = tensor("transpose_96_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_97_perm_0 = const()[name = tensor("transpose_97_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_97 = transpose(perm = transpose_97_perm_0, x = keys_75_cast_fp16)[name = tensor("transpose_165")]; + tensor transpose_96 = transpose(perm = transpose_96_perm_0, x = q_75_cast_fp16)[name = tensor("transpose_166")]; + tensor var_5176_cast_fp16 = matmul(transpose_x = var_5176_transpose_x_0, transpose_y = var_5176_transpose_y_0, x = transpose_96, y = transpose_97)[name = tensor("op_5176_cast_fp16")]; + tensor var_5170_to_fp16 = const()[name = tensor("op_5170_to_fp16"), val = tensor([0x1p-3])]; + tensor attn_weights_73_cast_fp16 = mul(x = var_5176_cast_fp16, y = var_5170_to_fp16)[name = tensor("attn_weights_73_cast_fp16")]; + tensor var_5178 = logical_not(x = attn_mask_51)[name = tensor("op_5178")]; + tensor var_5179_to_fp16 = const()[name = tensor("op_5179_to_fp16"), val = tensor(-inf)]; + tensor attn_weights_75_cast_fp16 = select(a = var_5179_to_fp16, b = attn_weights_73_cast_fp16, cond = var_5178)[name = tensor("attn_weights_75_cast_fp16")]; + tensor var_5181 = const()[name = tensor("op_5181"), val = tensor(-1)]; + tensor attn_weights_77_cast_fp16 = softmax(axis = var_5181, x = attn_weights_75_cast_fp16)[name = tensor("attn_weights_77_cast_fp16")]; + tensor attn_output_25_transpose_x_0 = const()[name = tensor("attn_output_25_transpose_x_0"), val = tensor(false)]; + tensor attn_output_25_transpose_y_0 = const()[name = tensor("attn_output_25_transpose_y_0"), val = tensor(false)]; + tensor values_77_cast_fp16 = transpose(perm = var_5121, x = values_75_cast_fp16)[name = tensor("transpose_167")]; + tensor attn_output_25_cast_fp16 = matmul(transpose_x = attn_output_25_transpose_x_0, transpose_y = attn_output_25_transpose_y_0, x = attn_weights_77_cast_fp16, y = values_77_cast_fp16)[name = tensor("attn_output_25_cast_fp16")]; + tensor var_5189 = const()[name = tensor("op_5189"), val = tensor([0, 2, 1, 3])]; + tensor var_5194 = const()[name = tensor("op_5194"), val = tensor([1, 1, 1024])]; + tensor var_5190_cast_fp16 = transpose(perm = var_5189, x = attn_output_25_cast_fp16)[name = tensor("transpose_164")]; + tensor input_151_cast_fp16 = reshape(shape = var_5194, x = var_5190_cast_fp16)[name = tensor("input_151_cast_fp16")]; + tensor attn12_out_proj_weight_to_fp16 = const()[name = tensor("attn12_out_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(309523712)))]; + tensor linear_50_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = attn12_out_proj_weight_to_fp16, x = input_151_cast_fp16)[name = tensor("linear_50_cast_fp16")]; + tensor var_5200_to_fp16 = const()[name = tensor("op_5200_to_fp16"), val = tensor(0x1p+0)]; + tensor var_5201_cast_fp16 = add(x = position12_to_fp16, y = var_5200_to_fp16)[name = tensor("op_5201_cast_fp16")]; + tensor var_5201_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_5201_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor input_153_cast_fp16 = add(x = input_147_cast_fp16, y = linear_50_cast_fp16)[name = tensor("input_153_cast_fp16")]; + tensor input_155_axes_0 = const()[name = tensor("input_155_axes_0"), val = tensor([-1])]; + tensor norm12_2_weight_to_fp16 = const()[name = tensor("norm12_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(311620928)))]; + tensor norm12_2_bias_to_fp16 = const()[name = tensor("norm12_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(311623040)))]; + tensor var_5205_to_fp16 = const()[name = tensor("op_5205_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_155_cast_fp16 = layer_norm(axes = input_155_axes_0, beta = norm12_2_bias_to_fp16, epsilon = var_5205_to_fp16, gamma = norm12_2_weight_to_fp16, x = input_153_cast_fp16)[name = tensor("input_155_cast_fp16")]; + tensor linear12_1_weight_to_fp16 = const()[name = tensor("linear12_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(311625152)))]; + tensor linear_51_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = linear12_1_weight_to_fp16, x = input_155_cast_fp16)[name = tensor("linear_51_cast_fp16")]; + tensor input_157_mode_0 = const()[name = tensor("input_157_mode_0"), val = tensor("EXACT")]; + tensor input_157_cast_fp16 = gelu(mode = input_157_mode_0, x = linear_51_cast_fp16)[name = tensor("input_157_cast_fp16")]; + tensor linear12_2_weight_to_fp16 = const()[name = tensor("linear12_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(320013824)))]; + tensor linear_52_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = linear12_2_weight_to_fp16, x = input_157_cast_fp16)[name = tensor("linear_52_cast_fp16")]; + tensor input_159_cast_fp16 = add(x = input_153_cast_fp16, y = linear_52_cast_fp16)[name = tensor("input_159_cast_fp16")]; + tensor input_161_axes_0 = const()[name = tensor("input_161_axes_0"), val = tensor([-1])]; + tensor norm13_1_weight_to_fp16 = const()[name = tensor("norm13_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(328402496)))]; + tensor norm13_1_bias_to_fp16 = const()[name = tensor("norm13_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(328404608)))]; + tensor var_5222_to_fp16 = const()[name = tensor("op_5222_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_161_cast_fp16 = layer_norm(axes = input_161_axes_0, beta = norm13_1_bias_to_fp16, epsilon = var_5222_to_fp16, gamma = norm13_1_weight_to_fp16, x = input_159_cast_fp16)[name = tensor("input_161_cast_fp16")]; + tensor attn13_in_proj_weight_to_fp16 = const()[name = tensor("attn13_in_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(328406720)))]; + tensor linear_53_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = attn13_in_proj_weight_to_fp16, x = input_161_cast_fp16)[name = tensor("linear_53_cast_fp16")]; + tensor var_5236 = const()[name = tensor("op_5236"), val = tensor([1, 1, 3, 16, 64])]; + tensor qkv_27_cast_fp16 = reshape(shape = var_5236, x = linear_53_cast_fp16)[name = tensor("qkv_27_cast_fp16")]; + tensor q_79_begin_0 = const()[name = tensor("q_79_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor q_79_end_0 = const()[name = tensor("q_79_end_0"), val = tensor([1, 1, 1, 16, 64])]; + tensor q_79_end_mask_0 = const()[name = tensor("q_79_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor q_79_squeeze_mask_0 = const()[name = tensor("q_79_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor q_79_cast_fp16 = slice_by_index(begin = q_79_begin_0, end = q_79_end_0, end_mask = q_79_end_mask_0, squeeze_mask = q_79_squeeze_mask_0, x = qkv_27_cast_fp16)[name = tensor("q_79_cast_fp16")]; + tensor k_53_begin_0 = const()[name = tensor("k_53_begin_0"), val = tensor([0, 0, 1, 0, 0])]; + tensor k_53_end_0 = const()[name = tensor("k_53_end_0"), val = tensor([1, 1, 2, 16, 64])]; + tensor k_53_end_mask_0 = const()[name = tensor("k_53_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor k_53_squeeze_mask_0 = const()[name = tensor("k_53_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor k_53_cast_fp16 = slice_by_index(begin = k_53_begin_0, end = k_53_end_0, end_mask = k_53_end_mask_0, squeeze_mask = k_53_squeeze_mask_0, x = qkv_27_cast_fp16)[name = tensor("k_53_cast_fp16")]; + tensor v_27_begin_0 = const()[name = tensor("v_27_begin_0"), val = tensor([0, 0, 2, 0, 0])]; + tensor v_27_end_0 = const()[name = tensor("v_27_end_0"), val = tensor([1, 1, 3, 16, 64])]; + tensor v_27_end_mask_0 = const()[name = tensor("v_27_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor v_27_squeeze_mask_0 = const()[name = tensor("v_27_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor v_27_cast_fp16 = slice_by_index(begin = v_27_begin_0, end = v_27_end_0, end_mask = v_27_end_mask_0, squeeze_mask = v_27_squeeze_mask_0, x = qkv_27_cast_fp16)[name = tensor("v_27_cast_fp16")]; + tensor var_5302 = const()[name = tensor("op_5302"), val = tensor([1, 1, 1, 1])]; + tensor position13_to_fp16_dtype_0 = const()[name = tensor("position13_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor position13_to_fp16 = cast(dtype = position13_to_fp16_dtype_0, x = position13)[name = tensor("cast_488")]; + tensor ts_83_cast_fp16 = reshape(shape = var_5302, x = position13_to_fp16)[name = tensor("ts_83_cast_fp16")]; + tensor var_5309 = const()[name = tensor("op_5309"), val = tensor([1, 1, 16, 32, 2])]; + tensor q_complex_27_cast_fp16 = reshape(shape = var_5309, x = q_79_cast_fp16)[name = tensor("q_complex_27_cast_fp16")]; + tensor var_5316 = const()[name = tensor("op_5316"), val = tensor([1, 1, 16, 32, 2])]; + tensor k_complex_27_cast_fp16 = reshape(shape = var_5316, x = k_53_cast_fp16)[name = tensor("k_complex_27_cast_fp16")]; + tensor var_5320_begin_0 = const()[name = tensor("op_5320_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_5320_end_0 = const()[name = tensor("op_5320_end_0"), val = tensor([1, 1, 16, 32, 1])]; + tensor var_5320_end_mask_0 = const()[name = tensor("op_5320_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_5320_squeeze_mask_0 = const()[name = tensor("op_5320_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_5320_cast_fp16 = slice_by_index(begin = var_5320_begin_0, end = var_5320_end_0, end_mask = var_5320_end_mask_0, squeeze_mask = var_5320_squeeze_mask_0, x = q_complex_27_cast_fp16)[name = tensor("op_5320_cast_fp16")]; + tensor var_5328_begin_0 = const()[name = tensor("op_5328_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_5328_end_0 = const()[name = tensor("op_5328_end_0"), val = tensor([1, 1, 16, 32, 2])]; + tensor var_5328_end_mask_0 = const()[name = tensor("op_5328_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_5328_squeeze_mask_0 = const()[name = tensor("op_5328_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_5328_cast_fp16 = slice_by_index(begin = var_5328_begin_0, end = var_5328_end_0, end_mask = var_5328_end_mask_0, squeeze_mask = var_5328_squeeze_mask_0, x = q_complex_27_cast_fp16)[name = tensor("op_5328_cast_fp16")]; + tensor var_5336_begin_0 = const()[name = tensor("op_5336_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_5336_end_0 = const()[name = tensor("op_5336_end_0"), val = tensor([1, 1, 16, 32, 1])]; + tensor var_5336_end_mask_0 = const()[name = tensor("op_5336_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_5336_squeeze_mask_0 = const()[name = tensor("op_5336_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_5336_cast_fp16 = slice_by_index(begin = var_5336_begin_0, end = var_5336_end_0, end_mask = var_5336_end_mask_0, squeeze_mask = var_5336_squeeze_mask_0, x = k_complex_27_cast_fp16)[name = tensor("op_5336_cast_fp16")]; + tensor var_5344_begin_0 = const()[name = tensor("op_5344_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_5344_end_0 = const()[name = tensor("op_5344_end_0"), val = tensor([1, 1, 16, 32, 2])]; + tensor var_5344_end_mask_0 = const()[name = tensor("op_5344_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_5344_squeeze_mask_0 = const()[name = tensor("op_5344_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_5344_cast_fp16 = slice_by_index(begin = var_5344_begin_0, end = var_5344_end_0, end_mask = var_5344_end_mask_0, squeeze_mask = var_5344_squeeze_mask_0, x = k_complex_27_cast_fp16)[name = tensor("op_5344_cast_fp16")]; + tensor freqs_27_to_fp16 = const()[name = tensor("freqs_27_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(334698240)))]; + tensor var_5350_cast_fp16 = mul(x = freqs_27_to_fp16, y = ts_83_cast_fp16)[name = tensor("op_5350_cast_fp16")]; + tensor rotr_27_cast_fp16 = cos(x = var_5350_cast_fp16)[name = tensor("rotr_27_cast_fp16")]; + tensor roti_27_cast_fp16 = sin(x = var_5350_cast_fp16)[name = tensor("roti_27_cast_fp16")]; + tensor var_5354_cast_fp16 = mul(x = var_5320_cast_fp16, y = rotr_27_cast_fp16)[name = tensor("op_5354_cast_fp16")]; + tensor var_5355_cast_fp16 = mul(x = var_5328_cast_fp16, y = roti_27_cast_fp16)[name = tensor("op_5355_cast_fp16")]; + tensor qor_53_cast_fp16 = sub(x = var_5354_cast_fp16, y = var_5355_cast_fp16)[name = tensor("qor_53_cast_fp16")]; + tensor var_5358_cast_fp16 = mul(x = var_5320_cast_fp16, y = roti_27_cast_fp16)[name = tensor("op_5358_cast_fp16")]; + tensor var_5359_cast_fp16 = mul(x = var_5328_cast_fp16, y = rotr_27_cast_fp16)[name = tensor("op_5359_cast_fp16")]; + tensor qoi_53_cast_fp16 = add(x = var_5358_cast_fp16, y = var_5359_cast_fp16)[name = tensor("qoi_53_cast_fp16")]; + tensor var_5362_cast_fp16 = mul(x = var_5336_cast_fp16, y = rotr_27_cast_fp16)[name = tensor("op_5362_cast_fp16")]; + tensor var_5363_cast_fp16 = mul(x = var_5344_cast_fp16, y = roti_27_cast_fp16)[name = tensor("op_5363_cast_fp16")]; + tensor kor_53_cast_fp16 = sub(x = var_5362_cast_fp16, y = var_5363_cast_fp16)[name = tensor("kor_53_cast_fp16")]; + tensor var_5366_cast_fp16 = mul(x = var_5336_cast_fp16, y = roti_27_cast_fp16)[name = tensor("op_5366_cast_fp16")]; + tensor var_5367_cast_fp16 = mul(x = var_5344_cast_fp16, y = rotr_27_cast_fp16)[name = tensor("op_5367_cast_fp16")]; + tensor koi_53_cast_fp16 = add(x = var_5366_cast_fp16, y = var_5367_cast_fp16)[name = tensor("koi_53_cast_fp16")]; + tensor qo_27_axis_0 = const()[name = tensor("qo_27_axis_0"), val = tensor(-1)]; + tensor qo_27_cast_fp16 = stack(axis = qo_27_axis_0, values = (qor_53_cast_fp16, qoi_53_cast_fp16))[name = tensor("qo_27_cast_fp16")]; + tensor ko_27_axis_0 = const()[name = tensor("ko_27_axis_0"), val = tensor(-1)]; + tensor ko_27_cast_fp16 = stack(axis = ko_27_axis_0, values = (kor_53_cast_fp16, koi_53_cast_fp16))[name = tensor("ko_27_cast_fp16")]; + tensor var_5400 = const()[name = tensor("op_5400"), val = tensor([1, 1, 16, 64])]; + tensor q_81_cast_fp16 = reshape(shape = var_5400, x = qo_27_cast_fp16)[name = tensor("q_81_cast_fp16")]; + tensor var_5406 = const()[name = tensor("op_5406"), val = tensor([1, 1, 16, 64])]; + tensor k_55_cast_fp16 = reshape(shape = var_5406, x = ko_27_cast_fp16)[name = tensor("k_55_cast_fp16")]; + tensor _inversed_5431_y_0_to_fp16 = const()[name = tensor("_inversed_5431_y_0_to_fp16"), val = tensor(0x1p-9)]; + tensor _inversed_5431_cast_fp16 = mul(x = ts_83_cast_fp16, y = _inversed_5431_y_0_to_fp16)[name = tensor("_inversed_5431_cast_fp16")]; + tensor var_5432_cast_fp16 = floor(x = _inversed_5431_cast_fp16)[name = tensor("op_5432_cast_fp16")]; + tensor var_5433_to_fp16 = const()[name = tensor("op_5433_to_fp16"), val = tensor(0x1p+9)]; + tensor var_5434_cast_fp16 = mul(x = var_5432_cast_fp16, y = var_5433_to_fp16)[name = tensor("op_5434_cast_fp16")]; + tensor write_indices_float_55_cast_fp16 = sub(x = ts_83_cast_fp16, y = var_5434_cast_fp16)[name = tensor("write_indices_float_55_cast_fp16")]; + tensor var_5441_dtype_0 = const()[name = tensor("op_5441_dtype_0"), val = tensor("int32")]; + tensor write_indices_27_reps_0 = const()[name = tensor("write_indices_27_reps_0"), val = tensor([1, 1, 16, 64])]; + tensor write_indices_float_55_cast_fp16_to_int32 = cast(dtype = var_5441_dtype_0, x = write_indices_float_55_cast_fp16)[name = tensor("cast_487")]; + tensor write_indices_27 = tile(reps = write_indices_27_reps_0, x = write_indices_float_55_cast_fp16_to_int32)[name = tensor("write_indices_27")]; + tensor var_5451_begin_0 = const()[name = tensor("op_5451_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_5451_end_0 = const()[name = tensor("op_5451_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor var_5451_end_mask_0 = const()[name = tensor("op_5451_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_5451_squeeze_mask_0 = const()[name = tensor("op_5451_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor cache13_to_fp16_dtype_0 = const()[name = tensor("cache13_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor cache13_to_fp16 = cast(dtype = cache13_to_fp16_dtype_0, x = cache13)[name = tensor("cast_486")]; + tensor var_5451_cast_fp16 = slice_by_index(begin = var_5451_begin_0, end = var_5451_end_0, end_mask = var_5451_end_mask_0, squeeze_mask = var_5451_squeeze_mask_0, x = cache13_to_fp16)[name = tensor("op_5451_cast_fp16")]; + tensor var_5453_axis_0 = const()[name = tensor("op_5453_axis_0"), val = tensor(1)]; + tensor var_5453_mode_0 = const()[name = tensor("op_5453_mode_0"), val = tensor("update")]; + tensor var_5453_validate_indices_0 = const()[name = tensor("op_5453_validate_indices_0"), val = tensor(false)]; + tensor var_5453_cast_fp16 = scatter_along_axis(axis = var_5453_axis_0, data = var_5451_cast_fp16, indices = write_indices_27, mode = var_5453_mode_0, updates = k_55_cast_fp16, validate_indices = var_5453_validate_indices_0)[name = tensor("op_5453_cast_fp16")]; + tensor concat_93 = const()[name = tensor("concat_93"), val = tensor([0, 0, 0, 0, 0])]; + tensor concat_94 = const()[name = tensor("concat_94"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_27_internal_tensor_assign_1_stride_0 = const()[name = tensor("new_cache_27_internal_tensor_assign_1_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_27_internal_tensor_assign_1_begin_mask_0 = const()[name = tensor("new_cache_27_internal_tensor_assign_1_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_27_internal_tensor_assign_1_end_mask_0 = const()[name = tensor("new_cache_27_internal_tensor_assign_1_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_27_internal_tensor_assign_1_squeeze_mask_0 = const()[name = tensor("new_cache_27_internal_tensor_assign_1_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_74 = const()[name = tensor("shape_74"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_26 = const()[name = tensor("reduce_prod_26"), val = tensor(1048576)]; + tensor range_1d_26_start_0 = const()[name = tensor("range_1d_26_start_0"), val = tensor(0)]; + tensor range_1d_26_step_0 = const()[name = tensor("range_1d_26_step_0"), val = tensor(1)]; + tensor range_1d_26 = range_1d(end = reduce_prod_26, start = range_1d_26_start_0, step = range_1d_26_step_0)[name = tensor("range_1d_26")]; + tensor reshape_130 = reshape(shape = shape_74, x = range_1d_26)[name = tensor("reshape_130")]; + tensor slice_by_index_26 = slice_by_index(begin = concat_93, begin_mask = new_cache_27_internal_tensor_assign_1_begin_mask_0, end = concat_94, end_mask = new_cache_27_internal_tensor_assign_1_end_mask_0, squeeze_mask = new_cache_27_internal_tensor_assign_1_squeeze_mask_0, stride = new_cache_27_internal_tensor_assign_1_stride_0, x = reshape_130)[name = tensor("slice_by_index_26")]; + tensor reshape_131_shape_0 = const()[name = tensor("reshape_131_shape_0"), val = tensor([-1])]; + tensor reshape_131 = reshape(shape = reshape_131_shape_0, x = slice_by_index_26)[name = tensor("reshape_131")]; + tensor reshape_132_shape_0 = const()[name = tensor("reshape_132_shape_0"), val = tensor([-1])]; + tensor reshape_132_cast_fp16 = reshape(shape = reshape_132_shape_0, x = var_5453_cast_fp16)[name = tensor("reshape_132_cast_fp16")]; + tensor reshape_133_shape_0 = const()[name = tensor("reshape_133_shape_0"), val = tensor([-1])]; + tensor reshape_133_cast_fp16 = reshape(shape = reshape_133_shape_0, x = cache13_to_fp16)[name = tensor("reshape_133_cast_fp16")]; + tensor scatter_26_mode_0 = const()[name = tensor("scatter_26_mode_0"), val = tensor("update")]; + tensor scatter_26_axis_0 = const()[name = tensor("scatter_26_axis_0"), val = tensor(0)]; + tensor scatter_26_validate_indices_0 = const()[name = tensor("scatter_26_validate_indices_0"), val = tensor(false)]; + tensor scatter_26_cast_fp16 = scatter(axis = scatter_26_axis_0, data = reshape_133_cast_fp16, indices = reshape_131, mode = scatter_26_mode_0, updates = reshape_132_cast_fp16, validate_indices = scatter_26_validate_indices_0)[name = tensor("scatter_26_cast_fp16")]; + tensor reshape_134_cast_fp16 = reshape(shape = shape_74, x = scatter_26_cast_fp16)[name = tensor("reshape_134_cast_fp16")]; + tensor var_5461_begin_0 = const()[name = tensor("op_5461_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor var_5461_end_0 = const()[name = tensor("op_5461_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor var_5461_end_mask_0 = const()[name = tensor("op_5461_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_5461_squeeze_mask_0 = const()[name = tensor("op_5461_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor var_5461_cast_fp16 = slice_by_index(begin = var_5461_begin_0, end = var_5461_end_0, end_mask = var_5461_end_mask_0, squeeze_mask = var_5461_squeeze_mask_0, x = reshape_134_cast_fp16)[name = tensor("op_5461_cast_fp16")]; + tensor var_5463_axis_0 = const()[name = tensor("op_5463_axis_0"), val = tensor(1)]; + tensor var_5463_mode_0 = const()[name = tensor("op_5463_mode_0"), val = tensor("update")]; + tensor var_5463_validate_indices_0 = const()[name = tensor("op_5463_validate_indices_0"), val = tensor(false)]; + tensor var_5463_cast_fp16 = scatter_along_axis(axis = var_5463_axis_0, data = var_5461_cast_fp16, indices = write_indices_27, mode = var_5463_mode_0, updates = v_27_cast_fp16, validate_indices = var_5463_validate_indices_0)[name = tensor("op_5463_cast_fp16")]; + tensor concat_95 = const()[name = tensor("concat_95"), val = tensor([1, 0, 0, 0, 0])]; + tensor concat_96 = const()[name = tensor("concat_96"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_27_internal_tensor_assign_2_stride_0 = const()[name = tensor("new_cache_27_internal_tensor_assign_2_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_27_internal_tensor_assign_2_begin_mask_0 = const()[name = tensor("new_cache_27_internal_tensor_assign_2_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_27_internal_tensor_assign_2_end_mask_0 = const()[name = tensor("new_cache_27_internal_tensor_assign_2_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_27_internal_tensor_assign_2_squeeze_mask_0 = const()[name = tensor("new_cache_27_internal_tensor_assign_2_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_75 = const()[name = tensor("shape_75"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_27 = const()[name = tensor("reduce_prod_27"), val = tensor(1048576)]; + tensor range_1d_27_start_0 = const()[name = tensor("range_1d_27_start_0"), val = tensor(0)]; + tensor range_1d_27_step_0 = const()[name = tensor("range_1d_27_step_0"), val = tensor(1)]; + tensor range_1d_27 = range_1d(end = reduce_prod_27, start = range_1d_27_start_0, step = range_1d_27_step_0)[name = tensor("range_1d_27")]; + tensor reshape_135 = reshape(shape = shape_75, x = range_1d_27)[name = tensor("reshape_135")]; + tensor slice_by_index_27 = slice_by_index(begin = concat_95, begin_mask = new_cache_27_internal_tensor_assign_2_begin_mask_0, end = concat_96, end_mask = new_cache_27_internal_tensor_assign_2_end_mask_0, squeeze_mask = new_cache_27_internal_tensor_assign_2_squeeze_mask_0, stride = new_cache_27_internal_tensor_assign_2_stride_0, x = reshape_135)[name = tensor("slice_by_index_27")]; + tensor reshape_136_shape_0 = const()[name = tensor("reshape_136_shape_0"), val = tensor([-1])]; + tensor reshape_136 = reshape(shape = reshape_136_shape_0, x = slice_by_index_27)[name = tensor("reshape_136")]; + tensor reshape_137_shape_0 = const()[name = tensor("reshape_137_shape_0"), val = tensor([-1])]; + tensor reshape_137_cast_fp16 = reshape(shape = reshape_137_shape_0, x = var_5463_cast_fp16)[name = tensor("reshape_137_cast_fp16")]; + tensor reshape_138_shape_0 = const()[name = tensor("reshape_138_shape_0"), val = tensor([-1])]; + tensor reshape_138_cast_fp16 = reshape(shape = reshape_138_shape_0, x = reshape_134_cast_fp16)[name = tensor("reshape_138_cast_fp16")]; + tensor scatter_27_mode_0 = const()[name = tensor("scatter_27_mode_0"), val = tensor("update")]; + tensor scatter_27_axis_0 = const()[name = tensor("scatter_27_axis_0"), val = tensor(0)]; + tensor scatter_27_validate_indices_0 = const()[name = tensor("scatter_27_validate_indices_0"), val = tensor(false)]; + tensor scatter_27_cast_fp16 = scatter(axis = scatter_27_axis_0, data = reshape_138_cast_fp16, indices = reshape_136, mode = scatter_27_mode_0, updates = reshape_137_cast_fp16, validate_indices = scatter_27_validate_indices_0)[name = tensor("scatter_27_cast_fp16")]; + tensor reshape_139_cast_fp16 = reshape(shape = shape_75, x = scatter_27_cast_fp16)[name = tensor("reshape_139_cast_fp16")]; + tensor reshape_139_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("reshape_139_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor keys_79_begin_0 = const()[name = tensor("keys_79_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor keys_79_end_0 = const()[name = tensor("keys_79_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor keys_79_end_mask_0 = const()[name = tensor("keys_79_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor keys_79_squeeze_mask_0 = const()[name = tensor("keys_79_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor keys_79_cast_fp16 = slice_by_index(begin = keys_79_begin_0, end = keys_79_end_0, end_mask = keys_79_end_mask_0, squeeze_mask = keys_79_squeeze_mask_0, x = reshape_139_cast_fp16)[name = tensor("keys_79_cast_fp16")]; + tensor values_79_begin_0 = const()[name = tensor("values_79_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor values_79_end_0 = const()[name = tensor("values_79_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor values_79_end_mask_0 = const()[name = tensor("values_79_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor values_79_squeeze_mask_0 = const()[name = tensor("values_79_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor values_79_cast_fp16 = slice_by_index(begin = values_79_begin_0, end = values_79_end_0, end_mask = values_79_end_mask_0, squeeze_mask = values_79_squeeze_mask_0, x = reshape_139_cast_fp16)[name = tensor("values_79_cast_fp16")]; + tensor var_5475_cast_fp16 = not_equal(x = keys_79_cast_fp16, y = keys_79_cast_fp16)[name = tensor("op_5475_cast_fp16")]; + tensor keys_81_cast_fp16 = select(a = var_463_to_fp16, b = keys_79_cast_fp16, cond = var_5475_cast_fp16)[name = tensor("keys_81_cast_fp16")]; + tensor var_5483_cast_fp16 = not_equal(x = values_79_cast_fp16, y = values_79_cast_fp16)[name = tensor("op_5483_cast_fp16")]; + tensor values_81_cast_fp16 = select(a = var_463_to_fp16, b = values_79_cast_fp16, cond = var_5483_cast_fp16)[name = tensor("values_81_cast_fp16")]; + tensor var_5507 = const()[name = tensor("op_5507"), val = tensor([0, 2, 1, 3])]; + tensor var_5523 = const()[name = tensor("op_5523"), val = tensor([1, 1, 1])]; + tensor var_5524_cast_fp16 = reshape(shape = var_5523, x = position13_to_fp16)[name = tensor("op_5524_cast_fp16")]; + tensor var_5544_to_fp16 = const()[name = tensor("op_5544_to_fp16"), val = tensor(0x1p+0)]; + tensor valid_len_27_cast_fp16 = add(x = var_5524_cast_fp16, y = var_5544_to_fp16)[name = tensor("valid_len_27_cast_fp16")]; + tensor valid_mask_27_cast_fp16 = less(x = k_positions_1_promoted_to_fp16, y = valid_len_27_cast_fp16)[name = tensor("valid_mask_27_cast_fp16")]; + tensor causal_mask_27_cast_fp16 = less_equal(x = k_positions_1_promoted_to_fp16, y = var_5524_cast_fp16)[name = tensor("causal_mask_27_cast_fp16")]; + tensor attn_mask_53 = logical_and(x = valid_mask_27_cast_fp16, y = causal_mask_27_cast_fp16)[name = tensor("attn_mask_53")]; + tensor attn_mask_55_axes_0 = const()[name = tensor("attn_mask_55_axes_0"), val = tensor([1])]; + tensor attn_mask_55 = expand_dims(axes = attn_mask_55_axes_0, x = attn_mask_53)[name = tensor("attn_mask_55")]; + tensor var_5562_transpose_x_0 = const()[name = tensor("op_5562_transpose_x_0"), val = tensor(false)]; + tensor var_5562_transpose_y_0 = const()[name = tensor("op_5562_transpose_y_0"), val = tensor(false)]; + tensor transpose_98_perm_0 = const()[name = tensor("transpose_98_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_99_perm_0 = const()[name = tensor("transpose_99_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_99 = transpose(perm = transpose_99_perm_0, x = keys_81_cast_fp16)[name = tensor("transpose_161")]; + tensor transpose_98 = transpose(perm = transpose_98_perm_0, x = q_81_cast_fp16)[name = tensor("transpose_162")]; + tensor var_5562_cast_fp16 = matmul(transpose_x = var_5562_transpose_x_0, transpose_y = var_5562_transpose_y_0, x = transpose_98, y = transpose_99)[name = tensor("op_5562_cast_fp16")]; + tensor var_5556_to_fp16 = const()[name = tensor("op_5556_to_fp16"), val = tensor([0x1p-3])]; + tensor attn_weights_79_cast_fp16 = mul(x = var_5562_cast_fp16, y = var_5556_to_fp16)[name = tensor("attn_weights_79_cast_fp16")]; + tensor var_5564 = logical_not(x = attn_mask_55)[name = tensor("op_5564")]; + tensor var_5565_to_fp16 = const()[name = tensor("op_5565_to_fp16"), val = tensor(-inf)]; + tensor attn_weights_81_cast_fp16 = select(a = var_5565_to_fp16, b = attn_weights_79_cast_fp16, cond = var_5564)[name = tensor("attn_weights_81_cast_fp16")]; + tensor var_5567 = const()[name = tensor("op_5567"), val = tensor(-1)]; + tensor attn_weights_83_cast_fp16 = softmax(axis = var_5567, x = attn_weights_81_cast_fp16)[name = tensor("attn_weights_83_cast_fp16")]; + tensor attn_output_27_transpose_x_0 = const()[name = tensor("attn_output_27_transpose_x_0"), val = tensor(false)]; + tensor attn_output_27_transpose_y_0 = const()[name = tensor("attn_output_27_transpose_y_0"), val = tensor(false)]; + tensor values_83_cast_fp16 = transpose(perm = var_5507, x = values_81_cast_fp16)[name = tensor("transpose_163")]; + tensor attn_output_27_cast_fp16 = matmul(transpose_x = attn_output_27_transpose_x_0, transpose_y = attn_output_27_transpose_y_0, x = attn_weights_83_cast_fp16, y = values_83_cast_fp16)[name = tensor("attn_output_27_cast_fp16")]; + tensor var_5575 = const()[name = tensor("op_5575"), val = tensor([0, 2, 1, 3])]; + tensor var_5580 = const()[name = tensor("op_5580"), val = tensor([1, 1, 1024])]; + tensor var_5576_cast_fp16 = transpose(perm = var_5575, x = attn_output_27_cast_fp16)[name = tensor("transpose_160")]; + tensor input_163_cast_fp16 = reshape(shape = var_5580, x = var_5576_cast_fp16)[name = tensor("input_163_cast_fp16")]; + tensor attn13_out_proj_weight_to_fp16 = const()[name = tensor("attn13_out_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(334698368)))]; + tensor linear_54_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = attn13_out_proj_weight_to_fp16, x = input_163_cast_fp16)[name = tensor("linear_54_cast_fp16")]; + tensor var_5586_to_fp16 = const()[name = tensor("op_5586_to_fp16"), val = tensor(0x1p+0)]; + tensor var_5587_cast_fp16 = add(x = position13_to_fp16, y = var_5586_to_fp16)[name = tensor("op_5587_cast_fp16")]; + tensor var_5587_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_5587_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor input_165_cast_fp16 = add(x = input_159_cast_fp16, y = linear_54_cast_fp16)[name = tensor("input_165_cast_fp16")]; + tensor input_167_axes_0 = const()[name = tensor("input_167_axes_0"), val = tensor([-1])]; + tensor norm13_2_weight_to_fp16 = const()[name = tensor("norm13_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(336795584)))]; + tensor norm13_2_bias_to_fp16 = const()[name = tensor("norm13_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(336797696)))]; + tensor var_5591_to_fp16 = const()[name = tensor("op_5591_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_167_cast_fp16 = layer_norm(axes = input_167_axes_0, beta = norm13_2_bias_to_fp16, epsilon = var_5591_to_fp16, gamma = norm13_2_weight_to_fp16, x = input_165_cast_fp16)[name = tensor("input_167_cast_fp16")]; + tensor linear13_1_weight_to_fp16 = const()[name = tensor("linear13_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(336799808)))]; + tensor linear_55_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = linear13_1_weight_to_fp16, x = input_167_cast_fp16)[name = tensor("linear_55_cast_fp16")]; + tensor input_169_mode_0 = const()[name = tensor("input_169_mode_0"), val = tensor("EXACT")]; + tensor input_169_cast_fp16 = gelu(mode = input_169_mode_0, x = linear_55_cast_fp16)[name = tensor("input_169_cast_fp16")]; + tensor linear13_2_weight_to_fp16 = const()[name = tensor("linear13_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(345188480)))]; + tensor linear_56_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = linear13_2_weight_to_fp16, x = input_169_cast_fp16)[name = tensor("linear_56_cast_fp16")]; + tensor input_171_cast_fp16 = add(x = input_165_cast_fp16, y = linear_56_cast_fp16)[name = tensor("input_171_cast_fp16")]; + tensor input_173_axes_0 = const()[name = tensor("input_173_axes_0"), val = tensor([-1])]; + tensor norm14_1_weight_to_fp16 = const()[name = tensor("norm14_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(353577152)))]; + tensor norm14_1_bias_to_fp16 = const()[name = tensor("norm14_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(353579264)))]; + tensor var_5608_to_fp16 = const()[name = tensor("op_5608_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_173_cast_fp16 = layer_norm(axes = input_173_axes_0, beta = norm14_1_bias_to_fp16, epsilon = var_5608_to_fp16, gamma = norm14_1_weight_to_fp16, x = input_171_cast_fp16)[name = tensor("input_173_cast_fp16")]; + tensor attn14_in_proj_weight_to_fp16 = const()[name = tensor("attn14_in_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(353581376)))]; + tensor linear_57_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = attn14_in_proj_weight_to_fp16, x = input_173_cast_fp16)[name = tensor("linear_57_cast_fp16")]; + tensor var_5622 = const()[name = tensor("op_5622"), val = tensor([1, 1, 3, 16, 64])]; + tensor qkv_29_cast_fp16 = reshape(shape = var_5622, x = linear_57_cast_fp16)[name = tensor("qkv_29_cast_fp16")]; + tensor q_85_begin_0 = const()[name = tensor("q_85_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor q_85_end_0 = const()[name = tensor("q_85_end_0"), val = tensor([1, 1, 1, 16, 64])]; + tensor q_85_end_mask_0 = const()[name = tensor("q_85_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor q_85_squeeze_mask_0 = const()[name = tensor("q_85_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor q_85_cast_fp16 = slice_by_index(begin = q_85_begin_0, end = q_85_end_0, end_mask = q_85_end_mask_0, squeeze_mask = q_85_squeeze_mask_0, x = qkv_29_cast_fp16)[name = tensor("q_85_cast_fp16")]; + tensor k_57_begin_0 = const()[name = tensor("k_57_begin_0"), val = tensor([0, 0, 1, 0, 0])]; + tensor k_57_end_0 = const()[name = tensor("k_57_end_0"), val = tensor([1, 1, 2, 16, 64])]; + tensor k_57_end_mask_0 = const()[name = tensor("k_57_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor k_57_squeeze_mask_0 = const()[name = tensor("k_57_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor k_57_cast_fp16 = slice_by_index(begin = k_57_begin_0, end = k_57_end_0, end_mask = k_57_end_mask_0, squeeze_mask = k_57_squeeze_mask_0, x = qkv_29_cast_fp16)[name = tensor("k_57_cast_fp16")]; + tensor v_29_begin_0 = const()[name = tensor("v_29_begin_0"), val = tensor([0, 0, 2, 0, 0])]; + tensor v_29_end_0 = const()[name = tensor("v_29_end_0"), val = tensor([1, 1, 3, 16, 64])]; + tensor v_29_end_mask_0 = const()[name = tensor("v_29_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor v_29_squeeze_mask_0 = const()[name = tensor("v_29_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor v_29_cast_fp16 = slice_by_index(begin = v_29_begin_0, end = v_29_end_0, end_mask = v_29_end_mask_0, squeeze_mask = v_29_squeeze_mask_0, x = qkv_29_cast_fp16)[name = tensor("v_29_cast_fp16")]; + tensor var_5688 = const()[name = tensor("op_5688"), val = tensor([1, 1, 1, 1])]; + tensor position14_to_fp16_dtype_0 = const()[name = tensor("position14_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor position14_to_fp16 = cast(dtype = position14_to_fp16_dtype_0, x = position14)[name = tensor("cast_483")]; + tensor ts_89_cast_fp16 = reshape(shape = var_5688, x = position14_to_fp16)[name = tensor("ts_89_cast_fp16")]; + tensor var_5695 = const()[name = tensor("op_5695"), val = tensor([1, 1, 16, 32, 2])]; + tensor q_complex_29_cast_fp16 = reshape(shape = var_5695, x = q_85_cast_fp16)[name = tensor("q_complex_29_cast_fp16")]; + tensor var_5702 = const()[name = tensor("op_5702"), val = tensor([1, 1, 16, 32, 2])]; + tensor k_complex_29_cast_fp16 = reshape(shape = var_5702, x = k_57_cast_fp16)[name = tensor("k_complex_29_cast_fp16")]; + tensor var_5706_begin_0 = const()[name = tensor("op_5706_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_5706_end_0 = const()[name = tensor("op_5706_end_0"), val = tensor([1, 1, 16, 32, 1])]; + tensor var_5706_end_mask_0 = const()[name = tensor("op_5706_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_5706_squeeze_mask_0 = const()[name = tensor("op_5706_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_5706_cast_fp16 = slice_by_index(begin = var_5706_begin_0, end = var_5706_end_0, end_mask = var_5706_end_mask_0, squeeze_mask = var_5706_squeeze_mask_0, x = q_complex_29_cast_fp16)[name = tensor("op_5706_cast_fp16")]; + tensor var_5714_begin_0 = const()[name = tensor("op_5714_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_5714_end_0 = const()[name = tensor("op_5714_end_0"), val = tensor([1, 1, 16, 32, 2])]; + tensor var_5714_end_mask_0 = const()[name = tensor("op_5714_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_5714_squeeze_mask_0 = const()[name = tensor("op_5714_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_5714_cast_fp16 = slice_by_index(begin = var_5714_begin_0, end = var_5714_end_0, end_mask = var_5714_end_mask_0, squeeze_mask = var_5714_squeeze_mask_0, x = q_complex_29_cast_fp16)[name = tensor("op_5714_cast_fp16")]; + tensor var_5722_begin_0 = const()[name = tensor("op_5722_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_5722_end_0 = const()[name = tensor("op_5722_end_0"), val = tensor([1, 1, 16, 32, 1])]; + tensor var_5722_end_mask_0 = const()[name = tensor("op_5722_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_5722_squeeze_mask_0 = const()[name = tensor("op_5722_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_5722_cast_fp16 = slice_by_index(begin = var_5722_begin_0, end = var_5722_end_0, end_mask = var_5722_end_mask_0, squeeze_mask = var_5722_squeeze_mask_0, x = k_complex_29_cast_fp16)[name = tensor("op_5722_cast_fp16")]; + tensor var_5730_begin_0 = const()[name = tensor("op_5730_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_5730_end_0 = const()[name = tensor("op_5730_end_0"), val = tensor([1, 1, 16, 32, 2])]; + tensor var_5730_end_mask_0 = const()[name = tensor("op_5730_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_5730_squeeze_mask_0 = const()[name = tensor("op_5730_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_5730_cast_fp16 = slice_by_index(begin = var_5730_begin_0, end = var_5730_end_0, end_mask = var_5730_end_mask_0, squeeze_mask = var_5730_squeeze_mask_0, x = k_complex_29_cast_fp16)[name = tensor("op_5730_cast_fp16")]; + tensor freqs_29_to_fp16 = const()[name = tensor("freqs_29_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(359872896)))]; + tensor var_5736_cast_fp16 = mul(x = freqs_29_to_fp16, y = ts_89_cast_fp16)[name = tensor("op_5736_cast_fp16")]; + tensor rotr_29_cast_fp16 = cos(x = var_5736_cast_fp16)[name = tensor("rotr_29_cast_fp16")]; + tensor roti_29_cast_fp16 = sin(x = var_5736_cast_fp16)[name = tensor("roti_29_cast_fp16")]; + tensor var_5740_cast_fp16 = mul(x = var_5706_cast_fp16, y = rotr_29_cast_fp16)[name = tensor("op_5740_cast_fp16")]; + tensor var_5741_cast_fp16 = mul(x = var_5714_cast_fp16, y = roti_29_cast_fp16)[name = tensor("op_5741_cast_fp16")]; + tensor qor_57_cast_fp16 = sub(x = var_5740_cast_fp16, y = var_5741_cast_fp16)[name = tensor("qor_57_cast_fp16")]; + tensor var_5744_cast_fp16 = mul(x = var_5706_cast_fp16, y = roti_29_cast_fp16)[name = tensor("op_5744_cast_fp16")]; + tensor var_5745_cast_fp16 = mul(x = var_5714_cast_fp16, y = rotr_29_cast_fp16)[name = tensor("op_5745_cast_fp16")]; + tensor qoi_57_cast_fp16 = add(x = var_5744_cast_fp16, y = var_5745_cast_fp16)[name = tensor("qoi_57_cast_fp16")]; + tensor var_5748_cast_fp16 = mul(x = var_5722_cast_fp16, y = rotr_29_cast_fp16)[name = tensor("op_5748_cast_fp16")]; + tensor var_5749_cast_fp16 = mul(x = var_5730_cast_fp16, y = roti_29_cast_fp16)[name = tensor("op_5749_cast_fp16")]; + tensor kor_57_cast_fp16 = sub(x = var_5748_cast_fp16, y = var_5749_cast_fp16)[name = tensor("kor_57_cast_fp16")]; + tensor var_5752_cast_fp16 = mul(x = var_5722_cast_fp16, y = roti_29_cast_fp16)[name = tensor("op_5752_cast_fp16")]; + tensor var_5753_cast_fp16 = mul(x = var_5730_cast_fp16, y = rotr_29_cast_fp16)[name = tensor("op_5753_cast_fp16")]; + tensor koi_57_cast_fp16 = add(x = var_5752_cast_fp16, y = var_5753_cast_fp16)[name = tensor("koi_57_cast_fp16")]; + tensor qo_29_axis_0 = const()[name = tensor("qo_29_axis_0"), val = tensor(-1)]; + tensor qo_29_cast_fp16 = stack(axis = qo_29_axis_0, values = (qor_57_cast_fp16, qoi_57_cast_fp16))[name = tensor("qo_29_cast_fp16")]; + tensor ko_29_axis_0 = const()[name = tensor("ko_29_axis_0"), val = tensor(-1)]; + tensor ko_29_cast_fp16 = stack(axis = ko_29_axis_0, values = (kor_57_cast_fp16, koi_57_cast_fp16))[name = tensor("ko_29_cast_fp16")]; + tensor var_5786 = const()[name = tensor("op_5786"), val = tensor([1, 1, 16, 64])]; + tensor q_87_cast_fp16 = reshape(shape = var_5786, x = qo_29_cast_fp16)[name = tensor("q_87_cast_fp16")]; + tensor var_5792 = const()[name = tensor("op_5792"), val = tensor([1, 1, 16, 64])]; + tensor k_59_cast_fp16 = reshape(shape = var_5792, x = ko_29_cast_fp16)[name = tensor("k_59_cast_fp16")]; + tensor _inversed_5817_y_0_to_fp16 = const()[name = tensor("_inversed_5817_y_0_to_fp16"), val = tensor(0x1p-9)]; + tensor _inversed_5817_cast_fp16 = mul(x = ts_89_cast_fp16, y = _inversed_5817_y_0_to_fp16)[name = tensor("_inversed_5817_cast_fp16")]; + tensor var_5818_cast_fp16 = floor(x = _inversed_5817_cast_fp16)[name = tensor("op_5818_cast_fp16")]; + tensor var_5819_to_fp16 = const()[name = tensor("op_5819_to_fp16"), val = tensor(0x1p+9)]; + tensor var_5820_cast_fp16 = mul(x = var_5818_cast_fp16, y = var_5819_to_fp16)[name = tensor("op_5820_cast_fp16")]; + tensor write_indices_float_59_cast_fp16 = sub(x = ts_89_cast_fp16, y = var_5820_cast_fp16)[name = tensor("write_indices_float_59_cast_fp16")]; + tensor var_5827_dtype_0 = const()[name = tensor("op_5827_dtype_0"), val = tensor("int32")]; + tensor write_indices_29_reps_0 = const()[name = tensor("write_indices_29_reps_0"), val = tensor([1, 1, 16, 64])]; + tensor write_indices_float_59_cast_fp16_to_int32 = cast(dtype = var_5827_dtype_0, x = write_indices_float_59_cast_fp16)[name = tensor("cast_482")]; + tensor write_indices_29 = tile(reps = write_indices_29_reps_0, x = write_indices_float_59_cast_fp16_to_int32)[name = tensor("write_indices_29")]; + tensor var_5837_begin_0 = const()[name = tensor("op_5837_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_5837_end_0 = const()[name = tensor("op_5837_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor var_5837_end_mask_0 = const()[name = tensor("op_5837_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_5837_squeeze_mask_0 = const()[name = tensor("op_5837_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor cache14_to_fp16_dtype_0 = const()[name = tensor("cache14_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor cache14_to_fp16 = cast(dtype = cache14_to_fp16_dtype_0, x = cache14)[name = tensor("cast_481")]; + tensor var_5837_cast_fp16 = slice_by_index(begin = var_5837_begin_0, end = var_5837_end_0, end_mask = var_5837_end_mask_0, squeeze_mask = var_5837_squeeze_mask_0, x = cache14_to_fp16)[name = tensor("op_5837_cast_fp16")]; + tensor var_5839_axis_0 = const()[name = tensor("op_5839_axis_0"), val = tensor(1)]; + tensor var_5839_mode_0 = const()[name = tensor("op_5839_mode_0"), val = tensor("update")]; + tensor var_5839_validate_indices_0 = const()[name = tensor("op_5839_validate_indices_0"), val = tensor(false)]; + tensor var_5839_cast_fp16 = scatter_along_axis(axis = var_5839_axis_0, data = var_5837_cast_fp16, indices = write_indices_29, mode = var_5839_mode_0, updates = k_59_cast_fp16, validate_indices = var_5839_validate_indices_0)[name = tensor("op_5839_cast_fp16")]; + tensor concat_100 = const()[name = tensor("concat_100"), val = tensor([0, 0, 0, 0, 0])]; + tensor concat_101 = const()[name = tensor("concat_101"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_29_internal_tensor_assign_1_stride_0 = const()[name = tensor("new_cache_29_internal_tensor_assign_1_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_29_internal_tensor_assign_1_begin_mask_0 = const()[name = tensor("new_cache_29_internal_tensor_assign_1_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_29_internal_tensor_assign_1_end_mask_0 = const()[name = tensor("new_cache_29_internal_tensor_assign_1_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_29_internal_tensor_assign_1_squeeze_mask_0 = const()[name = tensor("new_cache_29_internal_tensor_assign_1_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_76 = const()[name = tensor("shape_76"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_28 = const()[name = tensor("reduce_prod_28"), val = tensor(1048576)]; + tensor range_1d_28_start_0 = const()[name = tensor("range_1d_28_start_0"), val = tensor(0)]; + tensor range_1d_28_step_0 = const()[name = tensor("range_1d_28_step_0"), val = tensor(1)]; + tensor range_1d_28 = range_1d(end = reduce_prod_28, start = range_1d_28_start_0, step = range_1d_28_step_0)[name = tensor("range_1d_28")]; + tensor reshape_140 = reshape(shape = shape_76, x = range_1d_28)[name = tensor("reshape_140")]; + tensor slice_by_index_28 = slice_by_index(begin = concat_100, begin_mask = new_cache_29_internal_tensor_assign_1_begin_mask_0, end = concat_101, end_mask = new_cache_29_internal_tensor_assign_1_end_mask_0, squeeze_mask = new_cache_29_internal_tensor_assign_1_squeeze_mask_0, stride = new_cache_29_internal_tensor_assign_1_stride_0, x = reshape_140)[name = tensor("slice_by_index_28")]; + tensor reshape_141_shape_0 = const()[name = tensor("reshape_141_shape_0"), val = tensor([-1])]; + tensor reshape_141 = reshape(shape = reshape_141_shape_0, x = slice_by_index_28)[name = tensor("reshape_141")]; + tensor reshape_142_shape_0 = const()[name = tensor("reshape_142_shape_0"), val = tensor([-1])]; + tensor reshape_142_cast_fp16 = reshape(shape = reshape_142_shape_0, x = var_5839_cast_fp16)[name = tensor("reshape_142_cast_fp16")]; + tensor reshape_143_shape_0 = const()[name = tensor("reshape_143_shape_0"), val = tensor([-1])]; + tensor reshape_143_cast_fp16 = reshape(shape = reshape_143_shape_0, x = cache14_to_fp16)[name = tensor("reshape_143_cast_fp16")]; + tensor scatter_28_mode_0 = const()[name = tensor("scatter_28_mode_0"), val = tensor("update")]; + tensor scatter_28_axis_0 = const()[name = tensor("scatter_28_axis_0"), val = tensor(0)]; + tensor scatter_28_validate_indices_0 = const()[name = tensor("scatter_28_validate_indices_0"), val = tensor(false)]; + tensor scatter_28_cast_fp16 = scatter(axis = scatter_28_axis_0, data = reshape_143_cast_fp16, indices = reshape_141, mode = scatter_28_mode_0, updates = reshape_142_cast_fp16, validate_indices = scatter_28_validate_indices_0)[name = tensor("scatter_28_cast_fp16")]; + tensor reshape_144_cast_fp16 = reshape(shape = shape_76, x = scatter_28_cast_fp16)[name = tensor("reshape_144_cast_fp16")]; + tensor var_5847_begin_0 = const()[name = tensor("op_5847_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor var_5847_end_0 = const()[name = tensor("op_5847_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor var_5847_end_mask_0 = const()[name = tensor("op_5847_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_5847_squeeze_mask_0 = const()[name = tensor("op_5847_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor var_5847_cast_fp16 = slice_by_index(begin = var_5847_begin_0, end = var_5847_end_0, end_mask = var_5847_end_mask_0, squeeze_mask = var_5847_squeeze_mask_0, x = reshape_144_cast_fp16)[name = tensor("op_5847_cast_fp16")]; + tensor var_5849_axis_0 = const()[name = tensor("op_5849_axis_0"), val = tensor(1)]; + tensor var_5849_mode_0 = const()[name = tensor("op_5849_mode_0"), val = tensor("update")]; + tensor var_5849_validate_indices_0 = const()[name = tensor("op_5849_validate_indices_0"), val = tensor(false)]; + tensor var_5849_cast_fp16 = scatter_along_axis(axis = var_5849_axis_0, data = var_5847_cast_fp16, indices = write_indices_29, mode = var_5849_mode_0, updates = v_29_cast_fp16, validate_indices = var_5849_validate_indices_0)[name = tensor("op_5849_cast_fp16")]; + tensor concat_102 = const()[name = tensor("concat_102"), val = tensor([1, 0, 0, 0, 0])]; + tensor concat_103 = const()[name = tensor("concat_103"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_29_internal_tensor_assign_2_stride_0 = const()[name = tensor("new_cache_29_internal_tensor_assign_2_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_29_internal_tensor_assign_2_begin_mask_0 = const()[name = tensor("new_cache_29_internal_tensor_assign_2_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_29_internal_tensor_assign_2_end_mask_0 = const()[name = tensor("new_cache_29_internal_tensor_assign_2_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_29_internal_tensor_assign_2_squeeze_mask_0 = const()[name = tensor("new_cache_29_internal_tensor_assign_2_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_77 = const()[name = tensor("shape_77"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_29 = const()[name = tensor("reduce_prod_29"), val = tensor(1048576)]; + tensor range_1d_29_start_0 = const()[name = tensor("range_1d_29_start_0"), val = tensor(0)]; + tensor range_1d_29_step_0 = const()[name = tensor("range_1d_29_step_0"), val = tensor(1)]; + tensor range_1d_29 = range_1d(end = reduce_prod_29, start = range_1d_29_start_0, step = range_1d_29_step_0)[name = tensor("range_1d_29")]; + tensor reshape_145 = reshape(shape = shape_77, x = range_1d_29)[name = tensor("reshape_145")]; + tensor slice_by_index_29 = slice_by_index(begin = concat_102, begin_mask = new_cache_29_internal_tensor_assign_2_begin_mask_0, end = concat_103, end_mask = new_cache_29_internal_tensor_assign_2_end_mask_0, squeeze_mask = new_cache_29_internal_tensor_assign_2_squeeze_mask_0, stride = new_cache_29_internal_tensor_assign_2_stride_0, x = reshape_145)[name = tensor("slice_by_index_29")]; + tensor reshape_146_shape_0 = const()[name = tensor("reshape_146_shape_0"), val = tensor([-1])]; + tensor reshape_146 = reshape(shape = reshape_146_shape_0, x = slice_by_index_29)[name = tensor("reshape_146")]; + tensor reshape_147_shape_0 = const()[name = tensor("reshape_147_shape_0"), val = tensor([-1])]; + tensor reshape_147_cast_fp16 = reshape(shape = reshape_147_shape_0, x = var_5849_cast_fp16)[name = tensor("reshape_147_cast_fp16")]; + tensor reshape_148_shape_0 = const()[name = tensor("reshape_148_shape_0"), val = tensor([-1])]; + tensor reshape_148_cast_fp16 = reshape(shape = reshape_148_shape_0, x = reshape_144_cast_fp16)[name = tensor("reshape_148_cast_fp16")]; + tensor scatter_29_mode_0 = const()[name = tensor("scatter_29_mode_0"), val = tensor("update")]; + tensor scatter_29_axis_0 = const()[name = tensor("scatter_29_axis_0"), val = tensor(0)]; + tensor scatter_29_validate_indices_0 = const()[name = tensor("scatter_29_validate_indices_0"), val = tensor(false)]; + tensor scatter_29_cast_fp16 = scatter(axis = scatter_29_axis_0, data = reshape_148_cast_fp16, indices = reshape_146, mode = scatter_29_mode_0, updates = reshape_147_cast_fp16, validate_indices = scatter_29_validate_indices_0)[name = tensor("scatter_29_cast_fp16")]; + tensor reshape_149_cast_fp16 = reshape(shape = shape_77, x = scatter_29_cast_fp16)[name = tensor("reshape_149_cast_fp16")]; + tensor reshape_149_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("reshape_149_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor keys_85_begin_0 = const()[name = tensor("keys_85_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor keys_85_end_0 = const()[name = tensor("keys_85_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor keys_85_end_mask_0 = const()[name = tensor("keys_85_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor keys_85_squeeze_mask_0 = const()[name = tensor("keys_85_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor keys_85_cast_fp16 = slice_by_index(begin = keys_85_begin_0, end = keys_85_end_0, end_mask = keys_85_end_mask_0, squeeze_mask = keys_85_squeeze_mask_0, x = reshape_149_cast_fp16)[name = tensor("keys_85_cast_fp16")]; + tensor values_85_begin_0 = const()[name = tensor("values_85_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor values_85_end_0 = const()[name = tensor("values_85_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor values_85_end_mask_0 = const()[name = tensor("values_85_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor values_85_squeeze_mask_0 = const()[name = tensor("values_85_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor values_85_cast_fp16 = slice_by_index(begin = values_85_begin_0, end = values_85_end_0, end_mask = values_85_end_mask_0, squeeze_mask = values_85_squeeze_mask_0, x = reshape_149_cast_fp16)[name = tensor("values_85_cast_fp16")]; + tensor var_5861_cast_fp16 = not_equal(x = keys_85_cast_fp16, y = keys_85_cast_fp16)[name = tensor("op_5861_cast_fp16")]; + tensor keys_87_cast_fp16 = select(a = var_463_to_fp16, b = keys_85_cast_fp16, cond = var_5861_cast_fp16)[name = tensor("keys_87_cast_fp16")]; + tensor var_5869_cast_fp16 = not_equal(x = values_85_cast_fp16, y = values_85_cast_fp16)[name = tensor("op_5869_cast_fp16")]; + tensor values_87_cast_fp16 = select(a = var_463_to_fp16, b = values_85_cast_fp16, cond = var_5869_cast_fp16)[name = tensor("values_87_cast_fp16")]; + tensor var_5893 = const()[name = tensor("op_5893"), val = tensor([0, 2, 1, 3])]; + tensor var_5909 = const()[name = tensor("op_5909"), val = tensor([1, 1, 1])]; + tensor var_5910_cast_fp16 = reshape(shape = var_5909, x = position14_to_fp16)[name = tensor("op_5910_cast_fp16")]; + tensor var_5930_to_fp16 = const()[name = tensor("op_5930_to_fp16"), val = tensor(0x1p+0)]; + tensor valid_len_29_cast_fp16 = add(x = var_5910_cast_fp16, y = var_5930_to_fp16)[name = tensor("valid_len_29_cast_fp16")]; + tensor valid_mask_29_cast_fp16 = less(x = k_positions_1_promoted_to_fp16, y = valid_len_29_cast_fp16)[name = tensor("valid_mask_29_cast_fp16")]; + tensor causal_mask_29_cast_fp16 = less_equal(x = k_positions_1_promoted_to_fp16, y = var_5910_cast_fp16)[name = tensor("causal_mask_29_cast_fp16")]; + tensor attn_mask_57 = logical_and(x = valid_mask_29_cast_fp16, y = causal_mask_29_cast_fp16)[name = tensor("attn_mask_57")]; + tensor attn_mask_59_axes_0 = const()[name = tensor("attn_mask_59_axes_0"), val = tensor([1])]; + tensor attn_mask_59 = expand_dims(axes = attn_mask_59_axes_0, x = attn_mask_57)[name = tensor("attn_mask_59")]; + tensor var_5948_transpose_x_0 = const()[name = tensor("op_5948_transpose_x_0"), val = tensor(false)]; + tensor var_5948_transpose_y_0 = const()[name = tensor("op_5948_transpose_y_0"), val = tensor(false)]; + tensor transpose_100_perm_0 = const()[name = tensor("transpose_100_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_101_perm_0 = const()[name = tensor("transpose_101_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_101 = transpose(perm = transpose_101_perm_0, x = keys_87_cast_fp16)[name = tensor("transpose_157")]; + tensor transpose_100 = transpose(perm = transpose_100_perm_0, x = q_87_cast_fp16)[name = tensor("transpose_158")]; + tensor var_5948_cast_fp16 = matmul(transpose_x = var_5948_transpose_x_0, transpose_y = var_5948_transpose_y_0, x = transpose_100, y = transpose_101)[name = tensor("op_5948_cast_fp16")]; + tensor var_5942_to_fp16 = const()[name = tensor("op_5942_to_fp16"), val = tensor([0x1p-3])]; + tensor attn_weights_85_cast_fp16 = mul(x = var_5948_cast_fp16, y = var_5942_to_fp16)[name = tensor("attn_weights_85_cast_fp16")]; + tensor var_5950 = logical_not(x = attn_mask_59)[name = tensor("op_5950")]; + tensor var_5951_to_fp16 = const()[name = tensor("op_5951_to_fp16"), val = tensor(-inf)]; + tensor attn_weights_87_cast_fp16 = select(a = var_5951_to_fp16, b = attn_weights_85_cast_fp16, cond = var_5950)[name = tensor("attn_weights_87_cast_fp16")]; + tensor var_5953 = const()[name = tensor("op_5953"), val = tensor(-1)]; + tensor attn_weights_89_cast_fp16 = softmax(axis = var_5953, x = attn_weights_87_cast_fp16)[name = tensor("attn_weights_89_cast_fp16")]; + tensor attn_output_29_transpose_x_0 = const()[name = tensor("attn_output_29_transpose_x_0"), val = tensor(false)]; + tensor attn_output_29_transpose_y_0 = const()[name = tensor("attn_output_29_transpose_y_0"), val = tensor(false)]; + tensor values_89_cast_fp16 = transpose(perm = var_5893, x = values_87_cast_fp16)[name = tensor("transpose_159")]; + tensor attn_output_29_cast_fp16 = matmul(transpose_x = attn_output_29_transpose_x_0, transpose_y = attn_output_29_transpose_y_0, x = attn_weights_89_cast_fp16, y = values_89_cast_fp16)[name = tensor("attn_output_29_cast_fp16")]; + tensor var_5961 = const()[name = tensor("op_5961"), val = tensor([0, 2, 1, 3])]; + tensor var_5966 = const()[name = tensor("op_5966"), val = tensor([1, 1, 1024])]; + tensor var_5962_cast_fp16 = transpose(perm = var_5961, x = attn_output_29_cast_fp16)[name = tensor("transpose_156")]; + tensor input_175_cast_fp16 = reshape(shape = var_5966, x = var_5962_cast_fp16)[name = tensor("input_175_cast_fp16")]; + tensor attn14_out_proj_weight_to_fp16 = const()[name = tensor("attn14_out_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(359873024)))]; + tensor linear_58_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = attn14_out_proj_weight_to_fp16, x = input_175_cast_fp16)[name = tensor("linear_58_cast_fp16")]; + tensor var_5972_to_fp16 = const()[name = tensor("op_5972_to_fp16"), val = tensor(0x1p+0)]; + tensor var_5973_cast_fp16 = add(x = position14_to_fp16, y = var_5972_to_fp16)[name = tensor("op_5973_cast_fp16")]; + tensor var_5973_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_5973_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor input_177_cast_fp16 = add(x = input_171_cast_fp16, y = linear_58_cast_fp16)[name = tensor("input_177_cast_fp16")]; + tensor input_179_axes_0 = const()[name = tensor("input_179_axes_0"), val = tensor([-1])]; + tensor norm14_2_weight_to_fp16 = const()[name = tensor("norm14_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(361970240)))]; + tensor norm14_2_bias_to_fp16 = const()[name = tensor("norm14_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(361972352)))]; + tensor var_5977_to_fp16 = const()[name = tensor("op_5977_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_179_cast_fp16 = layer_norm(axes = input_179_axes_0, beta = norm14_2_bias_to_fp16, epsilon = var_5977_to_fp16, gamma = norm14_2_weight_to_fp16, x = input_177_cast_fp16)[name = tensor("input_179_cast_fp16")]; + tensor linear14_1_weight_to_fp16 = const()[name = tensor("linear14_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(361974464)))]; + tensor linear_59_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = linear14_1_weight_to_fp16, x = input_179_cast_fp16)[name = tensor("linear_59_cast_fp16")]; + tensor input_181_mode_0 = const()[name = tensor("input_181_mode_0"), val = tensor("EXACT")]; + tensor input_181_cast_fp16 = gelu(mode = input_181_mode_0, x = linear_59_cast_fp16)[name = tensor("input_181_cast_fp16")]; + tensor linear14_2_weight_to_fp16 = const()[name = tensor("linear14_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(370363136)))]; + tensor linear_60_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = linear14_2_weight_to_fp16, x = input_181_cast_fp16)[name = tensor("linear_60_cast_fp16")]; + tensor input_183_cast_fp16 = add(x = input_177_cast_fp16, y = linear_60_cast_fp16)[name = tensor("input_183_cast_fp16")]; + tensor input_185_axes_0 = const()[name = tensor("input_185_axes_0"), val = tensor([-1])]; + tensor norm15_1_weight_to_fp16 = const()[name = tensor("norm15_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(378751808)))]; + tensor norm15_1_bias_to_fp16 = const()[name = tensor("norm15_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(378753920)))]; + tensor var_5994_to_fp16 = const()[name = tensor("op_5994_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_185_cast_fp16 = layer_norm(axes = input_185_axes_0, beta = norm15_1_bias_to_fp16, epsilon = var_5994_to_fp16, gamma = norm15_1_weight_to_fp16, x = input_183_cast_fp16)[name = tensor("input_185_cast_fp16")]; + tensor attn15_in_proj_weight_to_fp16 = const()[name = tensor("attn15_in_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(378756032)))]; + tensor linear_61_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = attn15_in_proj_weight_to_fp16, x = input_185_cast_fp16)[name = tensor("linear_61_cast_fp16")]; + tensor var_6008 = const()[name = tensor("op_6008"), val = tensor([1, 1, 3, 16, 64])]; + tensor qkv_31_cast_fp16 = reshape(shape = var_6008, x = linear_61_cast_fp16)[name = tensor("qkv_31_cast_fp16")]; + tensor q_91_begin_0 = const()[name = tensor("q_91_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor q_91_end_0 = const()[name = tensor("q_91_end_0"), val = tensor([1, 1, 1, 16, 64])]; + tensor q_91_end_mask_0 = const()[name = tensor("q_91_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor q_91_squeeze_mask_0 = const()[name = tensor("q_91_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor q_91_cast_fp16 = slice_by_index(begin = q_91_begin_0, end = q_91_end_0, end_mask = q_91_end_mask_0, squeeze_mask = q_91_squeeze_mask_0, x = qkv_31_cast_fp16)[name = tensor("q_91_cast_fp16")]; + tensor k_61_begin_0 = const()[name = tensor("k_61_begin_0"), val = tensor([0, 0, 1, 0, 0])]; + tensor k_61_end_0 = const()[name = tensor("k_61_end_0"), val = tensor([1, 1, 2, 16, 64])]; + tensor k_61_end_mask_0 = const()[name = tensor("k_61_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor k_61_squeeze_mask_0 = const()[name = tensor("k_61_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor k_61_cast_fp16 = slice_by_index(begin = k_61_begin_0, end = k_61_end_0, end_mask = k_61_end_mask_0, squeeze_mask = k_61_squeeze_mask_0, x = qkv_31_cast_fp16)[name = tensor("k_61_cast_fp16")]; + tensor v_31_begin_0 = const()[name = tensor("v_31_begin_0"), val = tensor([0, 0, 2, 0, 0])]; + tensor v_31_end_0 = const()[name = tensor("v_31_end_0"), val = tensor([1, 1, 3, 16, 64])]; + tensor v_31_end_mask_0 = const()[name = tensor("v_31_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor v_31_squeeze_mask_0 = const()[name = tensor("v_31_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor v_31_cast_fp16 = slice_by_index(begin = v_31_begin_0, end = v_31_end_0, end_mask = v_31_end_mask_0, squeeze_mask = v_31_squeeze_mask_0, x = qkv_31_cast_fp16)[name = tensor("v_31_cast_fp16")]; + tensor var_6074 = const()[name = tensor("op_6074"), val = tensor([1, 1, 1, 1])]; + tensor position15_to_fp16_dtype_0 = const()[name = tensor("position15_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor position15_to_fp16 = cast(dtype = position15_to_fp16_dtype_0, x = position15)[name = tensor("cast_478")]; + tensor ts_95_cast_fp16 = reshape(shape = var_6074, x = position15_to_fp16)[name = tensor("ts_95_cast_fp16")]; + tensor var_6081 = const()[name = tensor("op_6081"), val = tensor([1, 1, 16, 32, 2])]; + tensor q_complex_31_cast_fp16 = reshape(shape = var_6081, x = q_91_cast_fp16)[name = tensor("q_complex_31_cast_fp16")]; + tensor var_6088 = const()[name = tensor("op_6088"), val = tensor([1, 1, 16, 32, 2])]; + tensor k_complex_31_cast_fp16 = reshape(shape = var_6088, x = k_61_cast_fp16)[name = tensor("k_complex_31_cast_fp16")]; + tensor var_6092_begin_0 = const()[name = tensor("op_6092_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_6092_end_0 = const()[name = tensor("op_6092_end_0"), val = tensor([1, 1, 16, 32, 1])]; + tensor var_6092_end_mask_0 = const()[name = tensor("op_6092_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_6092_squeeze_mask_0 = const()[name = tensor("op_6092_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_6092_cast_fp16 = slice_by_index(begin = var_6092_begin_0, end = var_6092_end_0, end_mask = var_6092_end_mask_0, squeeze_mask = var_6092_squeeze_mask_0, x = q_complex_31_cast_fp16)[name = tensor("op_6092_cast_fp16")]; + tensor var_6100_begin_0 = const()[name = tensor("op_6100_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_6100_end_0 = const()[name = tensor("op_6100_end_0"), val = tensor([1, 1, 16, 32, 2])]; + tensor var_6100_end_mask_0 = const()[name = tensor("op_6100_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_6100_squeeze_mask_0 = const()[name = tensor("op_6100_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_6100_cast_fp16 = slice_by_index(begin = var_6100_begin_0, end = var_6100_end_0, end_mask = var_6100_end_mask_0, squeeze_mask = var_6100_squeeze_mask_0, x = q_complex_31_cast_fp16)[name = tensor("op_6100_cast_fp16")]; + tensor var_6108_begin_0 = const()[name = tensor("op_6108_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_6108_end_0 = const()[name = tensor("op_6108_end_0"), val = tensor([1, 1, 16, 32, 1])]; + tensor var_6108_end_mask_0 = const()[name = tensor("op_6108_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_6108_squeeze_mask_0 = const()[name = tensor("op_6108_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_6108_cast_fp16 = slice_by_index(begin = var_6108_begin_0, end = var_6108_end_0, end_mask = var_6108_end_mask_0, squeeze_mask = var_6108_squeeze_mask_0, x = k_complex_31_cast_fp16)[name = tensor("op_6108_cast_fp16")]; + tensor var_6116_begin_0 = const()[name = tensor("op_6116_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_6116_end_0 = const()[name = tensor("op_6116_end_0"), val = tensor([1, 1, 16, 32, 2])]; + tensor var_6116_end_mask_0 = const()[name = tensor("op_6116_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_6116_squeeze_mask_0 = const()[name = tensor("op_6116_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_6116_cast_fp16 = slice_by_index(begin = var_6116_begin_0, end = var_6116_end_0, end_mask = var_6116_end_mask_0, squeeze_mask = var_6116_squeeze_mask_0, x = k_complex_31_cast_fp16)[name = tensor("op_6116_cast_fp16")]; + tensor freqs_31_to_fp16 = const()[name = tensor("freqs_31_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(385047552)))]; + tensor var_6122_cast_fp16 = mul(x = freqs_31_to_fp16, y = ts_95_cast_fp16)[name = tensor("op_6122_cast_fp16")]; + tensor rotr_31_cast_fp16 = cos(x = var_6122_cast_fp16)[name = tensor("rotr_31_cast_fp16")]; + tensor roti_31_cast_fp16 = sin(x = var_6122_cast_fp16)[name = tensor("roti_31_cast_fp16")]; + tensor var_6126_cast_fp16 = mul(x = var_6092_cast_fp16, y = rotr_31_cast_fp16)[name = tensor("op_6126_cast_fp16")]; + tensor var_6127_cast_fp16 = mul(x = var_6100_cast_fp16, y = roti_31_cast_fp16)[name = tensor("op_6127_cast_fp16")]; + tensor qor_61_cast_fp16 = sub(x = var_6126_cast_fp16, y = var_6127_cast_fp16)[name = tensor("qor_61_cast_fp16")]; + tensor var_6130_cast_fp16 = mul(x = var_6092_cast_fp16, y = roti_31_cast_fp16)[name = tensor("op_6130_cast_fp16")]; + tensor var_6131_cast_fp16 = mul(x = var_6100_cast_fp16, y = rotr_31_cast_fp16)[name = tensor("op_6131_cast_fp16")]; + tensor qoi_61_cast_fp16 = add(x = var_6130_cast_fp16, y = var_6131_cast_fp16)[name = tensor("qoi_61_cast_fp16")]; + tensor var_6134_cast_fp16 = mul(x = var_6108_cast_fp16, y = rotr_31_cast_fp16)[name = tensor("op_6134_cast_fp16")]; + tensor var_6135_cast_fp16 = mul(x = var_6116_cast_fp16, y = roti_31_cast_fp16)[name = tensor("op_6135_cast_fp16")]; + tensor kor_61_cast_fp16 = sub(x = var_6134_cast_fp16, y = var_6135_cast_fp16)[name = tensor("kor_61_cast_fp16")]; + tensor var_6138_cast_fp16 = mul(x = var_6108_cast_fp16, y = roti_31_cast_fp16)[name = tensor("op_6138_cast_fp16")]; + tensor var_6139_cast_fp16 = mul(x = var_6116_cast_fp16, y = rotr_31_cast_fp16)[name = tensor("op_6139_cast_fp16")]; + tensor koi_61_cast_fp16 = add(x = var_6138_cast_fp16, y = var_6139_cast_fp16)[name = tensor("koi_61_cast_fp16")]; + tensor qo_31_axis_0 = const()[name = tensor("qo_31_axis_0"), val = tensor(-1)]; + tensor qo_31_cast_fp16 = stack(axis = qo_31_axis_0, values = (qor_61_cast_fp16, qoi_61_cast_fp16))[name = tensor("qo_31_cast_fp16")]; + tensor ko_31_axis_0 = const()[name = tensor("ko_31_axis_0"), val = tensor(-1)]; + tensor ko_31_cast_fp16 = stack(axis = ko_31_axis_0, values = (kor_61_cast_fp16, koi_61_cast_fp16))[name = tensor("ko_31_cast_fp16")]; + tensor var_6172 = const()[name = tensor("op_6172"), val = tensor([1, 1, 16, 64])]; + tensor q_93_cast_fp16 = reshape(shape = var_6172, x = qo_31_cast_fp16)[name = tensor("q_93_cast_fp16")]; + tensor var_6178 = const()[name = tensor("op_6178"), val = tensor([1, 1, 16, 64])]; + tensor k_63_cast_fp16 = reshape(shape = var_6178, x = ko_31_cast_fp16)[name = tensor("k_63_cast_fp16")]; + tensor _inversed_6203_y_0_to_fp16 = const()[name = tensor("_inversed_6203_y_0_to_fp16"), val = tensor(0x1p-9)]; + tensor _inversed_6203_cast_fp16 = mul(x = ts_95_cast_fp16, y = _inversed_6203_y_0_to_fp16)[name = tensor("_inversed_6203_cast_fp16")]; + tensor var_6204_cast_fp16 = floor(x = _inversed_6203_cast_fp16)[name = tensor("op_6204_cast_fp16")]; + tensor var_6205_to_fp16 = const()[name = tensor("op_6205_to_fp16"), val = tensor(0x1p+9)]; + tensor var_6206_cast_fp16 = mul(x = var_6204_cast_fp16, y = var_6205_to_fp16)[name = tensor("op_6206_cast_fp16")]; + tensor write_indices_float_63_cast_fp16 = sub(x = ts_95_cast_fp16, y = var_6206_cast_fp16)[name = tensor("write_indices_float_63_cast_fp16")]; + tensor var_6213_dtype_0 = const()[name = tensor("op_6213_dtype_0"), val = tensor("int32")]; + tensor write_indices_31_reps_0 = const()[name = tensor("write_indices_31_reps_0"), val = tensor([1, 1, 16, 64])]; + tensor write_indices_float_63_cast_fp16_to_int32 = cast(dtype = var_6213_dtype_0, x = write_indices_float_63_cast_fp16)[name = tensor("cast_477")]; + tensor write_indices_31 = tile(reps = write_indices_31_reps_0, x = write_indices_float_63_cast_fp16_to_int32)[name = tensor("write_indices_31")]; + tensor var_6223_begin_0 = const()[name = tensor("op_6223_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_6223_end_0 = const()[name = tensor("op_6223_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor var_6223_end_mask_0 = const()[name = tensor("op_6223_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_6223_squeeze_mask_0 = const()[name = tensor("op_6223_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor cache15_to_fp16_dtype_0 = const()[name = tensor("cache15_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor cache15_to_fp16 = cast(dtype = cache15_to_fp16_dtype_0, x = cache15)[name = tensor("cast_476")]; + tensor var_6223_cast_fp16 = slice_by_index(begin = var_6223_begin_0, end = var_6223_end_0, end_mask = var_6223_end_mask_0, squeeze_mask = var_6223_squeeze_mask_0, x = cache15_to_fp16)[name = tensor("op_6223_cast_fp16")]; + tensor var_6225_axis_0 = const()[name = tensor("op_6225_axis_0"), val = tensor(1)]; + tensor var_6225_mode_0 = const()[name = tensor("op_6225_mode_0"), val = tensor("update")]; + tensor var_6225_validate_indices_0 = const()[name = tensor("op_6225_validate_indices_0"), val = tensor(false)]; + tensor var_6225_cast_fp16 = scatter_along_axis(axis = var_6225_axis_0, data = var_6223_cast_fp16, indices = write_indices_31, mode = var_6225_mode_0, updates = k_63_cast_fp16, validate_indices = var_6225_validate_indices_0)[name = tensor("op_6225_cast_fp16")]; + tensor concat_107 = const()[name = tensor("concat_107"), val = tensor([0, 0, 0, 0, 0])]; + tensor concat_108 = const()[name = tensor("concat_108"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_31_internal_tensor_assign_1_stride_0 = const()[name = tensor("new_cache_31_internal_tensor_assign_1_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_31_internal_tensor_assign_1_begin_mask_0 = const()[name = tensor("new_cache_31_internal_tensor_assign_1_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_31_internal_tensor_assign_1_end_mask_0 = const()[name = tensor("new_cache_31_internal_tensor_assign_1_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_31_internal_tensor_assign_1_squeeze_mask_0 = const()[name = tensor("new_cache_31_internal_tensor_assign_1_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_78 = const()[name = tensor("shape_78"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_30 = const()[name = tensor("reduce_prod_30"), val = tensor(1048576)]; + tensor range_1d_30_start_0 = const()[name = tensor("range_1d_30_start_0"), val = tensor(0)]; + tensor range_1d_30_step_0 = const()[name = tensor("range_1d_30_step_0"), val = tensor(1)]; + tensor range_1d_30 = range_1d(end = reduce_prod_30, start = range_1d_30_start_0, step = range_1d_30_step_0)[name = tensor("range_1d_30")]; + tensor reshape_150 = reshape(shape = shape_78, x = range_1d_30)[name = tensor("reshape_150")]; + tensor slice_by_index_30 = slice_by_index(begin = concat_107, begin_mask = new_cache_31_internal_tensor_assign_1_begin_mask_0, end = concat_108, end_mask = new_cache_31_internal_tensor_assign_1_end_mask_0, squeeze_mask = new_cache_31_internal_tensor_assign_1_squeeze_mask_0, stride = new_cache_31_internal_tensor_assign_1_stride_0, x = reshape_150)[name = tensor("slice_by_index_30")]; + tensor reshape_151_shape_0 = const()[name = tensor("reshape_151_shape_0"), val = tensor([-1])]; + tensor reshape_151 = reshape(shape = reshape_151_shape_0, x = slice_by_index_30)[name = tensor("reshape_151")]; + tensor reshape_152_shape_0 = const()[name = tensor("reshape_152_shape_0"), val = tensor([-1])]; + tensor reshape_152_cast_fp16 = reshape(shape = reshape_152_shape_0, x = var_6225_cast_fp16)[name = tensor("reshape_152_cast_fp16")]; + tensor reshape_153_shape_0 = const()[name = tensor("reshape_153_shape_0"), val = tensor([-1])]; + tensor reshape_153_cast_fp16 = reshape(shape = reshape_153_shape_0, x = cache15_to_fp16)[name = tensor("reshape_153_cast_fp16")]; + tensor scatter_30_mode_0 = const()[name = tensor("scatter_30_mode_0"), val = tensor("update")]; + tensor scatter_30_axis_0 = const()[name = tensor("scatter_30_axis_0"), val = tensor(0)]; + tensor scatter_30_validate_indices_0 = const()[name = tensor("scatter_30_validate_indices_0"), val = tensor(false)]; + tensor scatter_30_cast_fp16 = scatter(axis = scatter_30_axis_0, data = reshape_153_cast_fp16, indices = reshape_151, mode = scatter_30_mode_0, updates = reshape_152_cast_fp16, validate_indices = scatter_30_validate_indices_0)[name = tensor("scatter_30_cast_fp16")]; + tensor reshape_154_cast_fp16 = reshape(shape = shape_78, x = scatter_30_cast_fp16)[name = tensor("reshape_154_cast_fp16")]; + tensor var_6233_begin_0 = const()[name = tensor("op_6233_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor var_6233_end_0 = const()[name = tensor("op_6233_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor var_6233_end_mask_0 = const()[name = tensor("op_6233_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_6233_squeeze_mask_0 = const()[name = tensor("op_6233_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor var_6233_cast_fp16 = slice_by_index(begin = var_6233_begin_0, end = var_6233_end_0, end_mask = var_6233_end_mask_0, squeeze_mask = var_6233_squeeze_mask_0, x = reshape_154_cast_fp16)[name = tensor("op_6233_cast_fp16")]; + tensor var_6235_axis_0 = const()[name = tensor("op_6235_axis_0"), val = tensor(1)]; + tensor var_6235_mode_0 = const()[name = tensor("op_6235_mode_0"), val = tensor("update")]; + tensor var_6235_validate_indices_0 = const()[name = tensor("op_6235_validate_indices_0"), val = tensor(false)]; + tensor var_6235_cast_fp16 = scatter_along_axis(axis = var_6235_axis_0, data = var_6233_cast_fp16, indices = write_indices_31, mode = var_6235_mode_0, updates = v_31_cast_fp16, validate_indices = var_6235_validate_indices_0)[name = tensor("op_6235_cast_fp16")]; + tensor concat_109 = const()[name = tensor("concat_109"), val = tensor([1, 0, 0, 0, 0])]; + tensor concat_110 = const()[name = tensor("concat_110"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_31_internal_tensor_assign_2_stride_0 = const()[name = tensor("new_cache_31_internal_tensor_assign_2_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_31_internal_tensor_assign_2_begin_mask_0 = const()[name = tensor("new_cache_31_internal_tensor_assign_2_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_31_internal_tensor_assign_2_end_mask_0 = const()[name = tensor("new_cache_31_internal_tensor_assign_2_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_31_internal_tensor_assign_2_squeeze_mask_0 = const()[name = tensor("new_cache_31_internal_tensor_assign_2_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_79 = const()[name = tensor("shape_79"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_31 = const()[name = tensor("reduce_prod_31"), val = tensor(1048576)]; + tensor range_1d_31_start_0 = const()[name = tensor("range_1d_31_start_0"), val = tensor(0)]; + tensor range_1d_31_step_0 = const()[name = tensor("range_1d_31_step_0"), val = tensor(1)]; + tensor range_1d_31 = range_1d(end = reduce_prod_31, start = range_1d_31_start_0, step = range_1d_31_step_0)[name = tensor("range_1d_31")]; + tensor reshape_155 = reshape(shape = shape_79, x = range_1d_31)[name = tensor("reshape_155")]; + tensor slice_by_index_31 = slice_by_index(begin = concat_109, begin_mask = new_cache_31_internal_tensor_assign_2_begin_mask_0, end = concat_110, end_mask = new_cache_31_internal_tensor_assign_2_end_mask_0, squeeze_mask = new_cache_31_internal_tensor_assign_2_squeeze_mask_0, stride = new_cache_31_internal_tensor_assign_2_stride_0, x = reshape_155)[name = tensor("slice_by_index_31")]; + tensor reshape_156_shape_0 = const()[name = tensor("reshape_156_shape_0"), val = tensor([-1])]; + tensor reshape_156 = reshape(shape = reshape_156_shape_0, x = slice_by_index_31)[name = tensor("reshape_156")]; + tensor reshape_157_shape_0 = const()[name = tensor("reshape_157_shape_0"), val = tensor([-1])]; + tensor reshape_157_cast_fp16 = reshape(shape = reshape_157_shape_0, x = var_6235_cast_fp16)[name = tensor("reshape_157_cast_fp16")]; + tensor reshape_158_shape_0 = const()[name = tensor("reshape_158_shape_0"), val = tensor([-1])]; + tensor reshape_158_cast_fp16 = reshape(shape = reshape_158_shape_0, x = reshape_154_cast_fp16)[name = tensor("reshape_158_cast_fp16")]; + tensor scatter_31_mode_0 = const()[name = tensor("scatter_31_mode_0"), val = tensor("update")]; + tensor scatter_31_axis_0 = const()[name = tensor("scatter_31_axis_0"), val = tensor(0)]; + tensor scatter_31_validate_indices_0 = const()[name = tensor("scatter_31_validate_indices_0"), val = tensor(false)]; + tensor scatter_31_cast_fp16 = scatter(axis = scatter_31_axis_0, data = reshape_158_cast_fp16, indices = reshape_156, mode = scatter_31_mode_0, updates = reshape_157_cast_fp16, validate_indices = scatter_31_validate_indices_0)[name = tensor("scatter_31_cast_fp16")]; + tensor reshape_159_cast_fp16 = reshape(shape = shape_79, x = scatter_31_cast_fp16)[name = tensor("reshape_159_cast_fp16")]; + tensor reshape_159_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("reshape_159_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor keys_91_begin_0 = const()[name = tensor("keys_91_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor keys_91_end_0 = const()[name = tensor("keys_91_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor keys_91_end_mask_0 = const()[name = tensor("keys_91_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor keys_91_squeeze_mask_0 = const()[name = tensor("keys_91_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor keys_91_cast_fp16 = slice_by_index(begin = keys_91_begin_0, end = keys_91_end_0, end_mask = keys_91_end_mask_0, squeeze_mask = keys_91_squeeze_mask_0, x = reshape_159_cast_fp16)[name = tensor("keys_91_cast_fp16")]; + tensor values_91_begin_0 = const()[name = tensor("values_91_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor values_91_end_0 = const()[name = tensor("values_91_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor values_91_end_mask_0 = const()[name = tensor("values_91_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor values_91_squeeze_mask_0 = const()[name = tensor("values_91_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor values_91_cast_fp16 = slice_by_index(begin = values_91_begin_0, end = values_91_end_0, end_mask = values_91_end_mask_0, squeeze_mask = values_91_squeeze_mask_0, x = reshape_159_cast_fp16)[name = tensor("values_91_cast_fp16")]; + tensor var_6247_cast_fp16 = not_equal(x = keys_91_cast_fp16, y = keys_91_cast_fp16)[name = tensor("op_6247_cast_fp16")]; + tensor keys_93_cast_fp16 = select(a = var_463_to_fp16, b = keys_91_cast_fp16, cond = var_6247_cast_fp16)[name = tensor("keys_93_cast_fp16")]; + tensor var_6255_cast_fp16 = not_equal(x = values_91_cast_fp16, y = values_91_cast_fp16)[name = tensor("op_6255_cast_fp16")]; + tensor values_93_cast_fp16 = select(a = var_463_to_fp16, b = values_91_cast_fp16, cond = var_6255_cast_fp16)[name = tensor("values_93_cast_fp16")]; + tensor var_6279 = const()[name = tensor("op_6279"), val = tensor([0, 2, 1, 3])]; + tensor var_6295 = const()[name = tensor("op_6295"), val = tensor([1, 1, 1])]; + tensor var_6296_cast_fp16 = reshape(shape = var_6295, x = position15_to_fp16)[name = tensor("op_6296_cast_fp16")]; + tensor var_6316_to_fp16 = const()[name = tensor("op_6316_to_fp16"), val = tensor(0x1p+0)]; + tensor valid_len_31_cast_fp16 = add(x = var_6296_cast_fp16, y = var_6316_to_fp16)[name = tensor("valid_len_31_cast_fp16")]; + tensor valid_mask_31_cast_fp16 = less(x = k_positions_1_promoted_to_fp16, y = valid_len_31_cast_fp16)[name = tensor("valid_mask_31_cast_fp16")]; + tensor causal_mask_31_cast_fp16 = less_equal(x = k_positions_1_promoted_to_fp16, y = var_6296_cast_fp16)[name = tensor("causal_mask_31_cast_fp16")]; + tensor attn_mask_61 = logical_and(x = valid_mask_31_cast_fp16, y = causal_mask_31_cast_fp16)[name = tensor("attn_mask_61")]; + tensor attn_mask_63_axes_0 = const()[name = tensor("attn_mask_63_axes_0"), val = tensor([1])]; + tensor attn_mask_63 = expand_dims(axes = attn_mask_63_axes_0, x = attn_mask_61)[name = tensor("attn_mask_63")]; + tensor var_6334_transpose_x_0 = const()[name = tensor("op_6334_transpose_x_0"), val = tensor(false)]; + tensor var_6334_transpose_y_0 = const()[name = tensor("op_6334_transpose_y_0"), val = tensor(false)]; + tensor transpose_102_perm_0 = const()[name = tensor("transpose_102_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_103_perm_0 = const()[name = tensor("transpose_103_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_103 = transpose(perm = transpose_103_perm_0, x = keys_93_cast_fp16)[name = tensor("transpose_153")]; + tensor transpose_102 = transpose(perm = transpose_102_perm_0, x = q_93_cast_fp16)[name = tensor("transpose_154")]; + tensor var_6334_cast_fp16 = matmul(transpose_x = var_6334_transpose_x_0, transpose_y = var_6334_transpose_y_0, x = transpose_102, y = transpose_103)[name = tensor("op_6334_cast_fp16")]; + tensor var_6328_to_fp16 = const()[name = tensor("op_6328_to_fp16"), val = tensor([0x1p-3])]; + tensor attn_weights_91_cast_fp16 = mul(x = var_6334_cast_fp16, y = var_6328_to_fp16)[name = tensor("attn_weights_91_cast_fp16")]; + tensor var_6336 = logical_not(x = attn_mask_63)[name = tensor("op_6336")]; + tensor var_6337_to_fp16 = const()[name = tensor("op_6337_to_fp16"), val = tensor(-inf)]; + tensor attn_weights_93_cast_fp16 = select(a = var_6337_to_fp16, b = attn_weights_91_cast_fp16, cond = var_6336)[name = tensor("attn_weights_93_cast_fp16")]; + tensor var_6339 = const()[name = tensor("op_6339"), val = tensor(-1)]; + tensor attn_weights_95_cast_fp16 = softmax(axis = var_6339, x = attn_weights_93_cast_fp16)[name = tensor("attn_weights_95_cast_fp16")]; + tensor attn_output_31_transpose_x_0 = const()[name = tensor("attn_output_31_transpose_x_0"), val = tensor(false)]; + tensor attn_output_31_transpose_y_0 = const()[name = tensor("attn_output_31_transpose_y_0"), val = tensor(false)]; + tensor values_95_cast_fp16 = transpose(perm = var_6279, x = values_93_cast_fp16)[name = tensor("transpose_155")]; + tensor attn_output_31_cast_fp16 = matmul(transpose_x = attn_output_31_transpose_x_0, transpose_y = attn_output_31_transpose_y_0, x = attn_weights_95_cast_fp16, y = values_95_cast_fp16)[name = tensor("attn_output_31_cast_fp16")]; + tensor var_6347 = const()[name = tensor("op_6347"), val = tensor([0, 2, 1, 3])]; + tensor var_6352 = const()[name = tensor("op_6352"), val = tensor([1, 1, 1024])]; + tensor var_6348_cast_fp16 = transpose(perm = var_6347, x = attn_output_31_cast_fp16)[name = tensor("transpose_152")]; + tensor input_187_cast_fp16 = reshape(shape = var_6352, x = var_6348_cast_fp16)[name = tensor("input_187_cast_fp16")]; + tensor attn15_out_proj_weight_to_fp16 = const()[name = tensor("attn15_out_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(385047680)))]; + tensor linear_62_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = attn15_out_proj_weight_to_fp16, x = input_187_cast_fp16)[name = tensor("linear_62_cast_fp16")]; + tensor var_6358_to_fp16 = const()[name = tensor("op_6358_to_fp16"), val = tensor(0x1p+0)]; + tensor var_6359_cast_fp16 = add(x = position15_to_fp16, y = var_6358_to_fp16)[name = tensor("op_6359_cast_fp16")]; + tensor var_6359_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_6359_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor input_189_cast_fp16 = add(x = input_183_cast_fp16, y = linear_62_cast_fp16)[name = tensor("input_189_cast_fp16")]; + tensor input_191_axes_0 = const()[name = tensor("input_191_axes_0"), val = tensor([-1])]; + tensor norm15_2_weight_to_fp16 = const()[name = tensor("norm15_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(387144896)))]; + tensor norm15_2_bias_to_fp16 = const()[name = tensor("norm15_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(387147008)))]; + tensor var_6363_to_fp16 = const()[name = tensor("op_6363_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_191_cast_fp16 = layer_norm(axes = input_191_axes_0, beta = norm15_2_bias_to_fp16, epsilon = var_6363_to_fp16, gamma = norm15_2_weight_to_fp16, x = input_189_cast_fp16)[name = tensor("input_191_cast_fp16")]; + tensor linear15_1_weight_to_fp16 = const()[name = tensor("linear15_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(387149120)))]; + tensor linear_63_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = linear15_1_weight_to_fp16, x = input_191_cast_fp16)[name = tensor("linear_63_cast_fp16")]; + tensor input_193_mode_0 = const()[name = tensor("input_193_mode_0"), val = tensor("EXACT")]; + tensor input_193_cast_fp16 = gelu(mode = input_193_mode_0, x = linear_63_cast_fp16)[name = tensor("input_193_cast_fp16")]; + tensor linear15_2_weight_to_fp16 = const()[name = tensor("linear15_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(395537792)))]; + tensor linear_64_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = linear15_2_weight_to_fp16, x = input_193_cast_fp16)[name = tensor("linear_64_cast_fp16")]; + tensor input_195_cast_fp16 = add(x = input_189_cast_fp16, y = linear_64_cast_fp16)[name = tensor("input_195_cast_fp16")]; + tensor input_197_axes_0 = const()[name = tensor("input_197_axes_0"), val = tensor([-1])]; + tensor norm16_1_weight_to_fp16 = const()[name = tensor("norm16_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(403926464)))]; + tensor norm16_1_bias_to_fp16 = const()[name = tensor("norm16_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(403928576)))]; + tensor var_6380_to_fp16 = const()[name = tensor("op_6380_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_197_cast_fp16 = layer_norm(axes = input_197_axes_0, beta = norm16_1_bias_to_fp16, epsilon = var_6380_to_fp16, gamma = norm16_1_weight_to_fp16, x = input_195_cast_fp16)[name = tensor("input_197_cast_fp16")]; + tensor attn16_in_proj_weight_to_fp16 = const()[name = tensor("attn16_in_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(403930688)))]; + tensor linear_65_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = attn16_in_proj_weight_to_fp16, x = input_197_cast_fp16)[name = tensor("linear_65_cast_fp16")]; + tensor var_6394 = const()[name = tensor("op_6394"), val = tensor([1, 1, 3, 16, 64])]; + tensor qkv_33_cast_fp16 = reshape(shape = var_6394, x = linear_65_cast_fp16)[name = tensor("qkv_33_cast_fp16")]; + tensor q_97_begin_0 = const()[name = tensor("q_97_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor q_97_end_0 = const()[name = tensor("q_97_end_0"), val = tensor([1, 1, 1, 16, 64])]; + tensor q_97_end_mask_0 = const()[name = tensor("q_97_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor q_97_squeeze_mask_0 = const()[name = tensor("q_97_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor q_97_cast_fp16 = slice_by_index(begin = q_97_begin_0, end = q_97_end_0, end_mask = q_97_end_mask_0, squeeze_mask = q_97_squeeze_mask_0, x = qkv_33_cast_fp16)[name = tensor("q_97_cast_fp16")]; + tensor k_65_begin_0 = const()[name = tensor("k_65_begin_0"), val = tensor([0, 0, 1, 0, 0])]; + tensor k_65_end_0 = const()[name = tensor("k_65_end_0"), val = tensor([1, 1, 2, 16, 64])]; + tensor k_65_end_mask_0 = const()[name = tensor("k_65_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor k_65_squeeze_mask_0 = const()[name = tensor("k_65_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor k_65_cast_fp16 = slice_by_index(begin = k_65_begin_0, end = k_65_end_0, end_mask = k_65_end_mask_0, squeeze_mask = k_65_squeeze_mask_0, x = qkv_33_cast_fp16)[name = tensor("k_65_cast_fp16")]; + tensor v_33_begin_0 = const()[name = tensor("v_33_begin_0"), val = tensor([0, 0, 2, 0, 0])]; + tensor v_33_end_0 = const()[name = tensor("v_33_end_0"), val = tensor([1, 1, 3, 16, 64])]; + tensor v_33_end_mask_0 = const()[name = tensor("v_33_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor v_33_squeeze_mask_0 = const()[name = tensor("v_33_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor v_33_cast_fp16 = slice_by_index(begin = v_33_begin_0, end = v_33_end_0, end_mask = v_33_end_mask_0, squeeze_mask = v_33_squeeze_mask_0, x = qkv_33_cast_fp16)[name = tensor("v_33_cast_fp16")]; + tensor var_6460 = const()[name = tensor("op_6460"), val = tensor([1, 1, 1, 1])]; + tensor position16_to_fp16_dtype_0 = const()[name = tensor("position16_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor position16_to_fp16 = cast(dtype = position16_to_fp16_dtype_0, x = position16)[name = tensor("cast_473")]; + tensor ts_101_cast_fp16 = reshape(shape = var_6460, x = position16_to_fp16)[name = tensor("ts_101_cast_fp16")]; + tensor var_6467 = const()[name = tensor("op_6467"), val = tensor([1, 1, 16, 32, 2])]; + tensor q_complex_33_cast_fp16 = reshape(shape = var_6467, x = q_97_cast_fp16)[name = tensor("q_complex_33_cast_fp16")]; + tensor var_6474 = const()[name = tensor("op_6474"), val = tensor([1, 1, 16, 32, 2])]; + tensor k_complex_33_cast_fp16 = reshape(shape = var_6474, x = k_65_cast_fp16)[name = tensor("k_complex_33_cast_fp16")]; + tensor var_6478_begin_0 = const()[name = tensor("op_6478_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_6478_end_0 = const()[name = tensor("op_6478_end_0"), val = tensor([1, 1, 16, 32, 1])]; + tensor var_6478_end_mask_0 = const()[name = tensor("op_6478_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_6478_squeeze_mask_0 = const()[name = tensor("op_6478_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_6478_cast_fp16 = slice_by_index(begin = var_6478_begin_0, end = var_6478_end_0, end_mask = var_6478_end_mask_0, squeeze_mask = var_6478_squeeze_mask_0, x = q_complex_33_cast_fp16)[name = tensor("op_6478_cast_fp16")]; + tensor var_6486_begin_0 = const()[name = tensor("op_6486_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_6486_end_0 = const()[name = tensor("op_6486_end_0"), val = tensor([1, 1, 16, 32, 2])]; + tensor var_6486_end_mask_0 = const()[name = tensor("op_6486_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_6486_squeeze_mask_0 = const()[name = tensor("op_6486_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_6486_cast_fp16 = slice_by_index(begin = var_6486_begin_0, end = var_6486_end_0, end_mask = var_6486_end_mask_0, squeeze_mask = var_6486_squeeze_mask_0, x = q_complex_33_cast_fp16)[name = tensor("op_6486_cast_fp16")]; + tensor var_6494_begin_0 = const()[name = tensor("op_6494_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_6494_end_0 = const()[name = tensor("op_6494_end_0"), val = tensor([1, 1, 16, 32, 1])]; + tensor var_6494_end_mask_0 = const()[name = tensor("op_6494_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_6494_squeeze_mask_0 = const()[name = tensor("op_6494_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_6494_cast_fp16 = slice_by_index(begin = var_6494_begin_0, end = var_6494_end_0, end_mask = var_6494_end_mask_0, squeeze_mask = var_6494_squeeze_mask_0, x = k_complex_33_cast_fp16)[name = tensor("op_6494_cast_fp16")]; + tensor var_6502_begin_0 = const()[name = tensor("op_6502_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_6502_end_0 = const()[name = tensor("op_6502_end_0"), val = tensor([1, 1, 16, 32, 2])]; + tensor var_6502_end_mask_0 = const()[name = tensor("op_6502_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_6502_squeeze_mask_0 = const()[name = tensor("op_6502_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_6502_cast_fp16 = slice_by_index(begin = var_6502_begin_0, end = var_6502_end_0, end_mask = var_6502_end_mask_0, squeeze_mask = var_6502_squeeze_mask_0, x = k_complex_33_cast_fp16)[name = tensor("op_6502_cast_fp16")]; + tensor freqs_33_to_fp16 = const()[name = tensor("freqs_33_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(410222208)))]; + tensor var_6508_cast_fp16 = mul(x = freqs_33_to_fp16, y = ts_101_cast_fp16)[name = tensor("op_6508_cast_fp16")]; + tensor rotr_33_cast_fp16 = cos(x = var_6508_cast_fp16)[name = tensor("rotr_33_cast_fp16")]; + tensor roti_33_cast_fp16 = sin(x = var_6508_cast_fp16)[name = tensor("roti_33_cast_fp16")]; + tensor var_6512_cast_fp16 = mul(x = var_6478_cast_fp16, y = rotr_33_cast_fp16)[name = tensor("op_6512_cast_fp16")]; + tensor var_6513_cast_fp16 = mul(x = var_6486_cast_fp16, y = roti_33_cast_fp16)[name = tensor("op_6513_cast_fp16")]; + tensor qor_65_cast_fp16 = sub(x = var_6512_cast_fp16, y = var_6513_cast_fp16)[name = tensor("qor_65_cast_fp16")]; + tensor var_6516_cast_fp16 = mul(x = var_6478_cast_fp16, y = roti_33_cast_fp16)[name = tensor("op_6516_cast_fp16")]; + tensor var_6517_cast_fp16 = mul(x = var_6486_cast_fp16, y = rotr_33_cast_fp16)[name = tensor("op_6517_cast_fp16")]; + tensor qoi_65_cast_fp16 = add(x = var_6516_cast_fp16, y = var_6517_cast_fp16)[name = tensor("qoi_65_cast_fp16")]; + tensor var_6520_cast_fp16 = mul(x = var_6494_cast_fp16, y = rotr_33_cast_fp16)[name = tensor("op_6520_cast_fp16")]; + tensor var_6521_cast_fp16 = mul(x = var_6502_cast_fp16, y = roti_33_cast_fp16)[name = tensor("op_6521_cast_fp16")]; + tensor kor_65_cast_fp16 = sub(x = var_6520_cast_fp16, y = var_6521_cast_fp16)[name = tensor("kor_65_cast_fp16")]; + tensor var_6524_cast_fp16 = mul(x = var_6494_cast_fp16, y = roti_33_cast_fp16)[name = tensor("op_6524_cast_fp16")]; + tensor var_6525_cast_fp16 = mul(x = var_6502_cast_fp16, y = rotr_33_cast_fp16)[name = tensor("op_6525_cast_fp16")]; + tensor koi_65_cast_fp16 = add(x = var_6524_cast_fp16, y = var_6525_cast_fp16)[name = tensor("koi_65_cast_fp16")]; + tensor qo_33_axis_0 = const()[name = tensor("qo_33_axis_0"), val = tensor(-1)]; + tensor qo_33_cast_fp16 = stack(axis = qo_33_axis_0, values = (qor_65_cast_fp16, qoi_65_cast_fp16))[name = tensor("qo_33_cast_fp16")]; + tensor ko_33_axis_0 = const()[name = tensor("ko_33_axis_0"), val = tensor(-1)]; + tensor ko_33_cast_fp16 = stack(axis = ko_33_axis_0, values = (kor_65_cast_fp16, koi_65_cast_fp16))[name = tensor("ko_33_cast_fp16")]; + tensor var_6558 = const()[name = tensor("op_6558"), val = tensor([1, 1, 16, 64])]; + tensor q_99_cast_fp16 = reshape(shape = var_6558, x = qo_33_cast_fp16)[name = tensor("q_99_cast_fp16")]; + tensor var_6564 = const()[name = tensor("op_6564"), val = tensor([1, 1, 16, 64])]; + tensor k_67_cast_fp16 = reshape(shape = var_6564, x = ko_33_cast_fp16)[name = tensor("k_67_cast_fp16")]; + tensor _inversed_6589_y_0_to_fp16 = const()[name = tensor("_inversed_6589_y_0_to_fp16"), val = tensor(0x1p-9)]; + tensor _inversed_6589_cast_fp16 = mul(x = ts_101_cast_fp16, y = _inversed_6589_y_0_to_fp16)[name = tensor("_inversed_6589_cast_fp16")]; + tensor var_6590_cast_fp16 = floor(x = _inversed_6589_cast_fp16)[name = tensor("op_6590_cast_fp16")]; + tensor var_6591_to_fp16 = const()[name = tensor("op_6591_to_fp16"), val = tensor(0x1p+9)]; + tensor var_6592_cast_fp16 = mul(x = var_6590_cast_fp16, y = var_6591_to_fp16)[name = tensor("op_6592_cast_fp16")]; + tensor write_indices_float_67_cast_fp16 = sub(x = ts_101_cast_fp16, y = var_6592_cast_fp16)[name = tensor("write_indices_float_67_cast_fp16")]; + tensor var_6599_dtype_0 = const()[name = tensor("op_6599_dtype_0"), val = tensor("int32")]; + tensor write_indices_33_reps_0 = const()[name = tensor("write_indices_33_reps_0"), val = tensor([1, 1, 16, 64])]; + tensor write_indices_float_67_cast_fp16_to_int32 = cast(dtype = var_6599_dtype_0, x = write_indices_float_67_cast_fp16)[name = tensor("cast_472")]; + tensor write_indices_33 = tile(reps = write_indices_33_reps_0, x = write_indices_float_67_cast_fp16_to_int32)[name = tensor("write_indices_33")]; + tensor var_6609_begin_0 = const()[name = tensor("op_6609_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_6609_end_0 = const()[name = tensor("op_6609_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor var_6609_end_mask_0 = const()[name = tensor("op_6609_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_6609_squeeze_mask_0 = const()[name = tensor("op_6609_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor cache16_to_fp16_dtype_0 = const()[name = tensor("cache16_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor cache16_to_fp16 = cast(dtype = cache16_to_fp16_dtype_0, x = cache16)[name = tensor("cast_471")]; + tensor var_6609_cast_fp16 = slice_by_index(begin = var_6609_begin_0, end = var_6609_end_0, end_mask = var_6609_end_mask_0, squeeze_mask = var_6609_squeeze_mask_0, x = cache16_to_fp16)[name = tensor("op_6609_cast_fp16")]; + tensor var_6611_axis_0 = const()[name = tensor("op_6611_axis_0"), val = tensor(1)]; + tensor var_6611_mode_0 = const()[name = tensor("op_6611_mode_0"), val = tensor("update")]; + tensor var_6611_validate_indices_0 = const()[name = tensor("op_6611_validate_indices_0"), val = tensor(false)]; + tensor var_6611_cast_fp16 = scatter_along_axis(axis = var_6611_axis_0, data = var_6609_cast_fp16, indices = write_indices_33, mode = var_6611_mode_0, updates = k_67_cast_fp16, validate_indices = var_6611_validate_indices_0)[name = tensor("op_6611_cast_fp16")]; + tensor concat_114 = const()[name = tensor("concat_114"), val = tensor([0, 0, 0, 0, 0])]; + tensor concat_115 = const()[name = tensor("concat_115"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_33_internal_tensor_assign_1_stride_0 = const()[name = tensor("new_cache_33_internal_tensor_assign_1_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_33_internal_tensor_assign_1_begin_mask_0 = const()[name = tensor("new_cache_33_internal_tensor_assign_1_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_33_internal_tensor_assign_1_end_mask_0 = const()[name = tensor("new_cache_33_internal_tensor_assign_1_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_33_internal_tensor_assign_1_squeeze_mask_0 = const()[name = tensor("new_cache_33_internal_tensor_assign_1_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_80 = const()[name = tensor("shape_80"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_32 = const()[name = tensor("reduce_prod_32"), val = tensor(1048576)]; + tensor range_1d_32_start_0 = const()[name = tensor("range_1d_32_start_0"), val = tensor(0)]; + tensor range_1d_32_step_0 = const()[name = tensor("range_1d_32_step_0"), val = tensor(1)]; + tensor range_1d_32 = range_1d(end = reduce_prod_32, start = range_1d_32_start_0, step = range_1d_32_step_0)[name = tensor("range_1d_32")]; + tensor reshape_160 = reshape(shape = shape_80, x = range_1d_32)[name = tensor("reshape_160")]; + tensor slice_by_index_32 = slice_by_index(begin = concat_114, begin_mask = new_cache_33_internal_tensor_assign_1_begin_mask_0, end = concat_115, end_mask = new_cache_33_internal_tensor_assign_1_end_mask_0, squeeze_mask = new_cache_33_internal_tensor_assign_1_squeeze_mask_0, stride = new_cache_33_internal_tensor_assign_1_stride_0, x = reshape_160)[name = tensor("slice_by_index_32")]; + tensor reshape_161_shape_0 = const()[name = tensor("reshape_161_shape_0"), val = tensor([-1])]; + tensor reshape_161 = reshape(shape = reshape_161_shape_0, x = slice_by_index_32)[name = tensor("reshape_161")]; + tensor reshape_162_shape_0 = const()[name = tensor("reshape_162_shape_0"), val = tensor([-1])]; + tensor reshape_162_cast_fp16 = reshape(shape = reshape_162_shape_0, x = var_6611_cast_fp16)[name = tensor("reshape_162_cast_fp16")]; + tensor reshape_163_shape_0 = const()[name = tensor("reshape_163_shape_0"), val = tensor([-1])]; + tensor reshape_163_cast_fp16 = reshape(shape = reshape_163_shape_0, x = cache16_to_fp16)[name = tensor("reshape_163_cast_fp16")]; + tensor scatter_32_mode_0 = const()[name = tensor("scatter_32_mode_0"), val = tensor("update")]; + tensor scatter_32_axis_0 = const()[name = tensor("scatter_32_axis_0"), val = tensor(0)]; + tensor scatter_32_validate_indices_0 = const()[name = tensor("scatter_32_validate_indices_0"), val = tensor(false)]; + tensor scatter_32_cast_fp16 = scatter(axis = scatter_32_axis_0, data = reshape_163_cast_fp16, indices = reshape_161, mode = scatter_32_mode_0, updates = reshape_162_cast_fp16, validate_indices = scatter_32_validate_indices_0)[name = tensor("scatter_32_cast_fp16")]; + tensor reshape_164_cast_fp16 = reshape(shape = shape_80, x = scatter_32_cast_fp16)[name = tensor("reshape_164_cast_fp16")]; + tensor var_6619_begin_0 = const()[name = tensor("op_6619_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor var_6619_end_0 = const()[name = tensor("op_6619_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor var_6619_end_mask_0 = const()[name = tensor("op_6619_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_6619_squeeze_mask_0 = const()[name = tensor("op_6619_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor var_6619_cast_fp16 = slice_by_index(begin = var_6619_begin_0, end = var_6619_end_0, end_mask = var_6619_end_mask_0, squeeze_mask = var_6619_squeeze_mask_0, x = reshape_164_cast_fp16)[name = tensor("op_6619_cast_fp16")]; + tensor var_6621_axis_0 = const()[name = tensor("op_6621_axis_0"), val = tensor(1)]; + tensor var_6621_mode_0 = const()[name = tensor("op_6621_mode_0"), val = tensor("update")]; + tensor var_6621_validate_indices_0 = const()[name = tensor("op_6621_validate_indices_0"), val = tensor(false)]; + tensor var_6621_cast_fp16 = scatter_along_axis(axis = var_6621_axis_0, data = var_6619_cast_fp16, indices = write_indices_33, mode = var_6621_mode_0, updates = v_33_cast_fp16, validate_indices = var_6621_validate_indices_0)[name = tensor("op_6621_cast_fp16")]; + tensor concat_116 = const()[name = tensor("concat_116"), val = tensor([1, 0, 0, 0, 0])]; + tensor concat_117 = const()[name = tensor("concat_117"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_33_internal_tensor_assign_2_stride_0 = const()[name = tensor("new_cache_33_internal_tensor_assign_2_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_33_internal_tensor_assign_2_begin_mask_0 = const()[name = tensor("new_cache_33_internal_tensor_assign_2_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_33_internal_tensor_assign_2_end_mask_0 = const()[name = tensor("new_cache_33_internal_tensor_assign_2_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_33_internal_tensor_assign_2_squeeze_mask_0 = const()[name = tensor("new_cache_33_internal_tensor_assign_2_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_81 = const()[name = tensor("shape_81"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_33 = const()[name = tensor("reduce_prod_33"), val = tensor(1048576)]; + tensor range_1d_33_start_0 = const()[name = tensor("range_1d_33_start_0"), val = tensor(0)]; + tensor range_1d_33_step_0 = const()[name = tensor("range_1d_33_step_0"), val = tensor(1)]; + tensor range_1d_33 = range_1d(end = reduce_prod_33, start = range_1d_33_start_0, step = range_1d_33_step_0)[name = tensor("range_1d_33")]; + tensor reshape_165 = reshape(shape = shape_81, x = range_1d_33)[name = tensor("reshape_165")]; + tensor slice_by_index_33 = slice_by_index(begin = concat_116, begin_mask = new_cache_33_internal_tensor_assign_2_begin_mask_0, end = concat_117, end_mask = new_cache_33_internal_tensor_assign_2_end_mask_0, squeeze_mask = new_cache_33_internal_tensor_assign_2_squeeze_mask_0, stride = new_cache_33_internal_tensor_assign_2_stride_0, x = reshape_165)[name = tensor("slice_by_index_33")]; + tensor reshape_166_shape_0 = const()[name = tensor("reshape_166_shape_0"), val = tensor([-1])]; + tensor reshape_166 = reshape(shape = reshape_166_shape_0, x = slice_by_index_33)[name = tensor("reshape_166")]; + tensor reshape_167_shape_0 = const()[name = tensor("reshape_167_shape_0"), val = tensor([-1])]; + tensor reshape_167_cast_fp16 = reshape(shape = reshape_167_shape_0, x = var_6621_cast_fp16)[name = tensor("reshape_167_cast_fp16")]; + tensor reshape_168_shape_0 = const()[name = tensor("reshape_168_shape_0"), val = tensor([-1])]; + tensor reshape_168_cast_fp16 = reshape(shape = reshape_168_shape_0, x = reshape_164_cast_fp16)[name = tensor("reshape_168_cast_fp16")]; + tensor scatter_33_mode_0 = const()[name = tensor("scatter_33_mode_0"), val = tensor("update")]; + tensor scatter_33_axis_0 = const()[name = tensor("scatter_33_axis_0"), val = tensor(0)]; + tensor scatter_33_validate_indices_0 = const()[name = tensor("scatter_33_validate_indices_0"), val = tensor(false)]; + tensor scatter_33_cast_fp16 = scatter(axis = scatter_33_axis_0, data = reshape_168_cast_fp16, indices = reshape_166, mode = scatter_33_mode_0, updates = reshape_167_cast_fp16, validate_indices = scatter_33_validate_indices_0)[name = tensor("scatter_33_cast_fp16")]; + tensor reshape_169_cast_fp16 = reshape(shape = shape_81, x = scatter_33_cast_fp16)[name = tensor("reshape_169_cast_fp16")]; + tensor reshape_169_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("reshape_169_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor keys_97_begin_0 = const()[name = tensor("keys_97_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor keys_97_end_0 = const()[name = tensor("keys_97_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor keys_97_end_mask_0 = const()[name = tensor("keys_97_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor keys_97_squeeze_mask_0 = const()[name = tensor("keys_97_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor keys_97_cast_fp16 = slice_by_index(begin = keys_97_begin_0, end = keys_97_end_0, end_mask = keys_97_end_mask_0, squeeze_mask = keys_97_squeeze_mask_0, x = reshape_169_cast_fp16)[name = tensor("keys_97_cast_fp16")]; + tensor values_97_begin_0 = const()[name = tensor("values_97_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor values_97_end_0 = const()[name = tensor("values_97_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor values_97_end_mask_0 = const()[name = tensor("values_97_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor values_97_squeeze_mask_0 = const()[name = tensor("values_97_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor values_97_cast_fp16 = slice_by_index(begin = values_97_begin_0, end = values_97_end_0, end_mask = values_97_end_mask_0, squeeze_mask = values_97_squeeze_mask_0, x = reshape_169_cast_fp16)[name = tensor("values_97_cast_fp16")]; + tensor var_6633_cast_fp16 = not_equal(x = keys_97_cast_fp16, y = keys_97_cast_fp16)[name = tensor("op_6633_cast_fp16")]; + tensor keys_99_cast_fp16 = select(a = var_463_to_fp16, b = keys_97_cast_fp16, cond = var_6633_cast_fp16)[name = tensor("keys_99_cast_fp16")]; + tensor var_6641_cast_fp16 = not_equal(x = values_97_cast_fp16, y = values_97_cast_fp16)[name = tensor("op_6641_cast_fp16")]; + tensor values_99_cast_fp16 = select(a = var_463_to_fp16, b = values_97_cast_fp16, cond = var_6641_cast_fp16)[name = tensor("values_99_cast_fp16")]; + tensor var_6665 = const()[name = tensor("op_6665"), val = tensor([0, 2, 1, 3])]; + tensor var_6681 = const()[name = tensor("op_6681"), val = tensor([1, 1, 1])]; + tensor var_6682_cast_fp16 = reshape(shape = var_6681, x = position16_to_fp16)[name = tensor("op_6682_cast_fp16")]; + tensor var_6702_to_fp16 = const()[name = tensor("op_6702_to_fp16"), val = tensor(0x1p+0)]; + tensor valid_len_33_cast_fp16 = add(x = var_6682_cast_fp16, y = var_6702_to_fp16)[name = tensor("valid_len_33_cast_fp16")]; + tensor valid_mask_33_cast_fp16 = less(x = k_positions_1_promoted_to_fp16, y = valid_len_33_cast_fp16)[name = tensor("valid_mask_33_cast_fp16")]; + tensor causal_mask_33_cast_fp16 = less_equal(x = k_positions_1_promoted_to_fp16, y = var_6682_cast_fp16)[name = tensor("causal_mask_33_cast_fp16")]; + tensor attn_mask_65 = logical_and(x = valid_mask_33_cast_fp16, y = causal_mask_33_cast_fp16)[name = tensor("attn_mask_65")]; + tensor attn_mask_67_axes_0 = const()[name = tensor("attn_mask_67_axes_0"), val = tensor([1])]; + tensor attn_mask_67 = expand_dims(axes = attn_mask_67_axes_0, x = attn_mask_65)[name = tensor("attn_mask_67")]; + tensor var_6720_transpose_x_0 = const()[name = tensor("op_6720_transpose_x_0"), val = tensor(false)]; + tensor var_6720_transpose_y_0 = const()[name = tensor("op_6720_transpose_y_0"), val = tensor(false)]; + tensor transpose_104_perm_0 = const()[name = tensor("transpose_104_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_105_perm_0 = const()[name = tensor("transpose_105_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_105 = transpose(perm = transpose_105_perm_0, x = keys_99_cast_fp16)[name = tensor("transpose_149")]; + tensor transpose_104 = transpose(perm = transpose_104_perm_0, x = q_99_cast_fp16)[name = tensor("transpose_150")]; + tensor var_6720_cast_fp16 = matmul(transpose_x = var_6720_transpose_x_0, transpose_y = var_6720_transpose_y_0, x = transpose_104, y = transpose_105)[name = tensor("op_6720_cast_fp16")]; + tensor var_6714_to_fp16 = const()[name = tensor("op_6714_to_fp16"), val = tensor([0x1p-3])]; + tensor attn_weights_97_cast_fp16 = mul(x = var_6720_cast_fp16, y = var_6714_to_fp16)[name = tensor("attn_weights_97_cast_fp16")]; + tensor var_6722 = logical_not(x = attn_mask_67)[name = tensor("op_6722")]; + tensor var_6723_to_fp16 = const()[name = tensor("op_6723_to_fp16"), val = tensor(-inf)]; + tensor attn_weights_99_cast_fp16 = select(a = var_6723_to_fp16, b = attn_weights_97_cast_fp16, cond = var_6722)[name = tensor("attn_weights_99_cast_fp16")]; + tensor var_6725 = const()[name = tensor("op_6725"), val = tensor(-1)]; + tensor attn_weights_101_cast_fp16 = softmax(axis = var_6725, x = attn_weights_99_cast_fp16)[name = tensor("attn_weights_101_cast_fp16")]; + tensor attn_output_33_transpose_x_0 = const()[name = tensor("attn_output_33_transpose_x_0"), val = tensor(false)]; + tensor attn_output_33_transpose_y_0 = const()[name = tensor("attn_output_33_transpose_y_0"), val = tensor(false)]; + tensor values_101_cast_fp16 = transpose(perm = var_6665, x = values_99_cast_fp16)[name = tensor("transpose_151")]; + tensor attn_output_33_cast_fp16 = matmul(transpose_x = attn_output_33_transpose_x_0, transpose_y = attn_output_33_transpose_y_0, x = attn_weights_101_cast_fp16, y = values_101_cast_fp16)[name = tensor("attn_output_33_cast_fp16")]; + tensor var_6733 = const()[name = tensor("op_6733"), val = tensor([0, 2, 1, 3])]; + tensor var_6738 = const()[name = tensor("op_6738"), val = tensor([1, 1, 1024])]; + tensor var_6734_cast_fp16 = transpose(perm = var_6733, x = attn_output_33_cast_fp16)[name = tensor("transpose_148")]; + tensor input_199_cast_fp16 = reshape(shape = var_6738, x = var_6734_cast_fp16)[name = tensor("input_199_cast_fp16")]; + tensor attn16_out_proj_weight_to_fp16 = const()[name = tensor("attn16_out_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(410222336)))]; + tensor linear_66_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = attn16_out_proj_weight_to_fp16, x = input_199_cast_fp16)[name = tensor("linear_66_cast_fp16")]; + tensor var_6744_to_fp16 = const()[name = tensor("op_6744_to_fp16"), val = tensor(0x1p+0)]; + tensor var_6745_cast_fp16 = add(x = position16_to_fp16, y = var_6744_to_fp16)[name = tensor("op_6745_cast_fp16")]; + tensor var_6745_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_6745_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor input_201_cast_fp16 = add(x = input_195_cast_fp16, y = linear_66_cast_fp16)[name = tensor("input_201_cast_fp16")]; + tensor input_203_axes_0 = const()[name = tensor("input_203_axes_0"), val = tensor([-1])]; + tensor norm16_2_weight_to_fp16 = const()[name = tensor("norm16_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(412319552)))]; + tensor norm16_2_bias_to_fp16 = const()[name = tensor("norm16_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(412321664)))]; + tensor var_6749_to_fp16 = const()[name = tensor("op_6749_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_203_cast_fp16 = layer_norm(axes = input_203_axes_0, beta = norm16_2_bias_to_fp16, epsilon = var_6749_to_fp16, gamma = norm16_2_weight_to_fp16, x = input_201_cast_fp16)[name = tensor("input_203_cast_fp16")]; + tensor linear16_1_weight_to_fp16 = const()[name = tensor("linear16_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(412323776)))]; + tensor linear_67_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = linear16_1_weight_to_fp16, x = input_203_cast_fp16)[name = tensor("linear_67_cast_fp16")]; + tensor input_205_mode_0 = const()[name = tensor("input_205_mode_0"), val = tensor("EXACT")]; + tensor input_205_cast_fp16 = gelu(mode = input_205_mode_0, x = linear_67_cast_fp16)[name = tensor("input_205_cast_fp16")]; + tensor linear16_2_weight_to_fp16 = const()[name = tensor("linear16_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(420712448)))]; + tensor linear_68_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = linear16_2_weight_to_fp16, x = input_205_cast_fp16)[name = tensor("linear_68_cast_fp16")]; + tensor input_207_cast_fp16 = add(x = input_201_cast_fp16, y = linear_68_cast_fp16)[name = tensor("input_207_cast_fp16")]; + tensor input_209_axes_0 = const()[name = tensor("input_209_axes_0"), val = tensor([-1])]; + tensor norm17_1_weight_to_fp16 = const()[name = tensor("norm17_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(429101120)))]; + tensor norm17_1_bias_to_fp16 = const()[name = tensor("norm17_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(429103232)))]; + tensor var_6766_to_fp16 = const()[name = tensor("op_6766_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_209_cast_fp16 = layer_norm(axes = input_209_axes_0, beta = norm17_1_bias_to_fp16, epsilon = var_6766_to_fp16, gamma = norm17_1_weight_to_fp16, x = input_207_cast_fp16)[name = tensor("input_209_cast_fp16")]; + tensor attn17_in_proj_weight_to_fp16 = const()[name = tensor("attn17_in_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(429105344)))]; + tensor linear_69_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = attn17_in_proj_weight_to_fp16, x = input_209_cast_fp16)[name = tensor("linear_69_cast_fp16")]; + tensor var_6780 = const()[name = tensor("op_6780"), val = tensor([1, 1, 3, 16, 64])]; + tensor qkv_35_cast_fp16 = reshape(shape = var_6780, x = linear_69_cast_fp16)[name = tensor("qkv_35_cast_fp16")]; + tensor q_103_begin_0 = const()[name = tensor("q_103_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor q_103_end_0 = const()[name = tensor("q_103_end_0"), val = tensor([1, 1, 1, 16, 64])]; + tensor q_103_end_mask_0 = const()[name = tensor("q_103_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor q_103_squeeze_mask_0 = const()[name = tensor("q_103_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor q_103_cast_fp16 = slice_by_index(begin = q_103_begin_0, end = q_103_end_0, end_mask = q_103_end_mask_0, squeeze_mask = q_103_squeeze_mask_0, x = qkv_35_cast_fp16)[name = tensor("q_103_cast_fp16")]; + tensor k_69_begin_0 = const()[name = tensor("k_69_begin_0"), val = tensor([0, 0, 1, 0, 0])]; + tensor k_69_end_0 = const()[name = tensor("k_69_end_0"), val = tensor([1, 1, 2, 16, 64])]; + tensor k_69_end_mask_0 = const()[name = tensor("k_69_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor k_69_squeeze_mask_0 = const()[name = tensor("k_69_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor k_69_cast_fp16 = slice_by_index(begin = k_69_begin_0, end = k_69_end_0, end_mask = k_69_end_mask_0, squeeze_mask = k_69_squeeze_mask_0, x = qkv_35_cast_fp16)[name = tensor("k_69_cast_fp16")]; + tensor v_35_begin_0 = const()[name = tensor("v_35_begin_0"), val = tensor([0, 0, 2, 0, 0])]; + tensor v_35_end_0 = const()[name = tensor("v_35_end_0"), val = tensor([1, 1, 3, 16, 64])]; + tensor v_35_end_mask_0 = const()[name = tensor("v_35_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor v_35_squeeze_mask_0 = const()[name = tensor("v_35_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor v_35_cast_fp16 = slice_by_index(begin = v_35_begin_0, end = v_35_end_0, end_mask = v_35_end_mask_0, squeeze_mask = v_35_squeeze_mask_0, x = qkv_35_cast_fp16)[name = tensor("v_35_cast_fp16")]; + tensor var_6846 = const()[name = tensor("op_6846"), val = tensor([1, 1, 1, 1])]; + tensor position17_to_fp16_dtype_0 = const()[name = tensor("position17_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor position17_to_fp16 = cast(dtype = position17_to_fp16_dtype_0, x = position17)[name = tensor("cast_468")]; + tensor ts_107_cast_fp16 = reshape(shape = var_6846, x = position17_to_fp16)[name = tensor("ts_107_cast_fp16")]; + tensor var_6853 = const()[name = tensor("op_6853"), val = tensor([1, 1, 16, 32, 2])]; + tensor q_complex_35_cast_fp16 = reshape(shape = var_6853, x = q_103_cast_fp16)[name = tensor("q_complex_35_cast_fp16")]; + tensor var_6860 = const()[name = tensor("op_6860"), val = tensor([1, 1, 16, 32, 2])]; + tensor k_complex_35_cast_fp16 = reshape(shape = var_6860, x = k_69_cast_fp16)[name = tensor("k_complex_35_cast_fp16")]; + tensor var_6864_begin_0 = const()[name = tensor("op_6864_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_6864_end_0 = const()[name = tensor("op_6864_end_0"), val = tensor([1, 1, 16, 32, 1])]; + tensor var_6864_end_mask_0 = const()[name = tensor("op_6864_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_6864_squeeze_mask_0 = const()[name = tensor("op_6864_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_6864_cast_fp16 = slice_by_index(begin = var_6864_begin_0, end = var_6864_end_0, end_mask = var_6864_end_mask_0, squeeze_mask = var_6864_squeeze_mask_0, x = q_complex_35_cast_fp16)[name = tensor("op_6864_cast_fp16")]; + tensor var_6872_begin_0 = const()[name = tensor("op_6872_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_6872_end_0 = const()[name = tensor("op_6872_end_0"), val = tensor([1, 1, 16, 32, 2])]; + tensor var_6872_end_mask_0 = const()[name = tensor("op_6872_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_6872_squeeze_mask_0 = const()[name = tensor("op_6872_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_6872_cast_fp16 = slice_by_index(begin = var_6872_begin_0, end = var_6872_end_0, end_mask = var_6872_end_mask_0, squeeze_mask = var_6872_squeeze_mask_0, x = q_complex_35_cast_fp16)[name = tensor("op_6872_cast_fp16")]; + tensor var_6880_begin_0 = const()[name = tensor("op_6880_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_6880_end_0 = const()[name = tensor("op_6880_end_0"), val = tensor([1, 1, 16, 32, 1])]; + tensor var_6880_end_mask_0 = const()[name = tensor("op_6880_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_6880_squeeze_mask_0 = const()[name = tensor("op_6880_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_6880_cast_fp16 = slice_by_index(begin = var_6880_begin_0, end = var_6880_end_0, end_mask = var_6880_end_mask_0, squeeze_mask = var_6880_squeeze_mask_0, x = k_complex_35_cast_fp16)[name = tensor("op_6880_cast_fp16")]; + tensor var_6888_begin_0 = const()[name = tensor("op_6888_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_6888_end_0 = const()[name = tensor("op_6888_end_0"), val = tensor([1, 1, 16, 32, 2])]; + tensor var_6888_end_mask_0 = const()[name = tensor("op_6888_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_6888_squeeze_mask_0 = const()[name = tensor("op_6888_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_6888_cast_fp16 = slice_by_index(begin = var_6888_begin_0, end = var_6888_end_0, end_mask = var_6888_end_mask_0, squeeze_mask = var_6888_squeeze_mask_0, x = k_complex_35_cast_fp16)[name = tensor("op_6888_cast_fp16")]; + tensor freqs_35_to_fp16 = const()[name = tensor("freqs_35_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(435396864)))]; + tensor var_6894_cast_fp16 = mul(x = freqs_35_to_fp16, y = ts_107_cast_fp16)[name = tensor("op_6894_cast_fp16")]; + tensor rotr_35_cast_fp16 = cos(x = var_6894_cast_fp16)[name = tensor("rotr_35_cast_fp16")]; + tensor roti_35_cast_fp16 = sin(x = var_6894_cast_fp16)[name = tensor("roti_35_cast_fp16")]; + tensor var_6898_cast_fp16 = mul(x = var_6864_cast_fp16, y = rotr_35_cast_fp16)[name = tensor("op_6898_cast_fp16")]; + tensor var_6899_cast_fp16 = mul(x = var_6872_cast_fp16, y = roti_35_cast_fp16)[name = tensor("op_6899_cast_fp16")]; + tensor qor_69_cast_fp16 = sub(x = var_6898_cast_fp16, y = var_6899_cast_fp16)[name = tensor("qor_69_cast_fp16")]; + tensor var_6902_cast_fp16 = mul(x = var_6864_cast_fp16, y = roti_35_cast_fp16)[name = tensor("op_6902_cast_fp16")]; + tensor var_6903_cast_fp16 = mul(x = var_6872_cast_fp16, y = rotr_35_cast_fp16)[name = tensor("op_6903_cast_fp16")]; + tensor qoi_69_cast_fp16 = add(x = var_6902_cast_fp16, y = var_6903_cast_fp16)[name = tensor("qoi_69_cast_fp16")]; + tensor var_6906_cast_fp16 = mul(x = var_6880_cast_fp16, y = rotr_35_cast_fp16)[name = tensor("op_6906_cast_fp16")]; + tensor var_6907_cast_fp16 = mul(x = var_6888_cast_fp16, y = roti_35_cast_fp16)[name = tensor("op_6907_cast_fp16")]; + tensor kor_69_cast_fp16 = sub(x = var_6906_cast_fp16, y = var_6907_cast_fp16)[name = tensor("kor_69_cast_fp16")]; + tensor var_6910_cast_fp16 = mul(x = var_6880_cast_fp16, y = roti_35_cast_fp16)[name = tensor("op_6910_cast_fp16")]; + tensor var_6911_cast_fp16 = mul(x = var_6888_cast_fp16, y = rotr_35_cast_fp16)[name = tensor("op_6911_cast_fp16")]; + tensor koi_69_cast_fp16 = add(x = var_6910_cast_fp16, y = var_6911_cast_fp16)[name = tensor("koi_69_cast_fp16")]; + tensor qo_35_axis_0 = const()[name = tensor("qo_35_axis_0"), val = tensor(-1)]; + tensor qo_35_cast_fp16 = stack(axis = qo_35_axis_0, values = (qor_69_cast_fp16, qoi_69_cast_fp16))[name = tensor("qo_35_cast_fp16")]; + tensor ko_35_axis_0 = const()[name = tensor("ko_35_axis_0"), val = tensor(-1)]; + tensor ko_35_cast_fp16 = stack(axis = ko_35_axis_0, values = (kor_69_cast_fp16, koi_69_cast_fp16))[name = tensor("ko_35_cast_fp16")]; + tensor var_6944 = const()[name = tensor("op_6944"), val = tensor([1, 1, 16, 64])]; + tensor q_105_cast_fp16 = reshape(shape = var_6944, x = qo_35_cast_fp16)[name = tensor("q_105_cast_fp16")]; + tensor var_6950 = const()[name = tensor("op_6950"), val = tensor([1, 1, 16, 64])]; + tensor k_71_cast_fp16 = reshape(shape = var_6950, x = ko_35_cast_fp16)[name = tensor("k_71_cast_fp16")]; + tensor _inversed_6975_y_0_to_fp16 = const()[name = tensor("_inversed_6975_y_0_to_fp16"), val = tensor(0x1p-9)]; + tensor _inversed_6975_cast_fp16 = mul(x = ts_107_cast_fp16, y = _inversed_6975_y_0_to_fp16)[name = tensor("_inversed_6975_cast_fp16")]; + tensor var_6976_cast_fp16 = floor(x = _inversed_6975_cast_fp16)[name = tensor("op_6976_cast_fp16")]; + tensor var_6977_to_fp16 = const()[name = tensor("op_6977_to_fp16"), val = tensor(0x1p+9)]; + tensor var_6978_cast_fp16 = mul(x = var_6976_cast_fp16, y = var_6977_to_fp16)[name = tensor("op_6978_cast_fp16")]; + tensor write_indices_float_71_cast_fp16 = sub(x = ts_107_cast_fp16, y = var_6978_cast_fp16)[name = tensor("write_indices_float_71_cast_fp16")]; + tensor var_6985_dtype_0 = const()[name = tensor("op_6985_dtype_0"), val = tensor("int32")]; + tensor write_indices_35_reps_0 = const()[name = tensor("write_indices_35_reps_0"), val = tensor([1, 1, 16, 64])]; + tensor write_indices_float_71_cast_fp16_to_int32 = cast(dtype = var_6985_dtype_0, x = write_indices_float_71_cast_fp16)[name = tensor("cast_467")]; + tensor write_indices_35 = tile(reps = write_indices_35_reps_0, x = write_indices_float_71_cast_fp16_to_int32)[name = tensor("write_indices_35")]; + tensor var_6995_begin_0 = const()[name = tensor("op_6995_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_6995_end_0 = const()[name = tensor("op_6995_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor var_6995_end_mask_0 = const()[name = tensor("op_6995_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_6995_squeeze_mask_0 = const()[name = tensor("op_6995_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor cache17_to_fp16_dtype_0 = const()[name = tensor("cache17_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor cache17_to_fp16 = cast(dtype = cache17_to_fp16_dtype_0, x = cache17)[name = tensor("cast_466")]; + tensor var_6995_cast_fp16 = slice_by_index(begin = var_6995_begin_0, end = var_6995_end_0, end_mask = var_6995_end_mask_0, squeeze_mask = var_6995_squeeze_mask_0, x = cache17_to_fp16)[name = tensor("op_6995_cast_fp16")]; + tensor var_6997_axis_0 = const()[name = tensor("op_6997_axis_0"), val = tensor(1)]; + tensor var_6997_mode_0 = const()[name = tensor("op_6997_mode_0"), val = tensor("update")]; + tensor var_6997_validate_indices_0 = const()[name = tensor("op_6997_validate_indices_0"), val = tensor(false)]; + tensor var_6997_cast_fp16 = scatter_along_axis(axis = var_6997_axis_0, data = var_6995_cast_fp16, indices = write_indices_35, mode = var_6997_mode_0, updates = k_71_cast_fp16, validate_indices = var_6997_validate_indices_0)[name = tensor("op_6997_cast_fp16")]; + tensor concat_121 = const()[name = tensor("concat_121"), val = tensor([0, 0, 0, 0, 0])]; + tensor concat_122 = const()[name = tensor("concat_122"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_35_internal_tensor_assign_1_stride_0 = const()[name = tensor("new_cache_35_internal_tensor_assign_1_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_35_internal_tensor_assign_1_begin_mask_0 = const()[name = tensor("new_cache_35_internal_tensor_assign_1_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_35_internal_tensor_assign_1_end_mask_0 = const()[name = tensor("new_cache_35_internal_tensor_assign_1_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_35_internal_tensor_assign_1_squeeze_mask_0 = const()[name = tensor("new_cache_35_internal_tensor_assign_1_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_82 = const()[name = tensor("shape_82"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_34 = const()[name = tensor("reduce_prod_34"), val = tensor(1048576)]; + tensor range_1d_34_start_0 = const()[name = tensor("range_1d_34_start_0"), val = tensor(0)]; + tensor range_1d_34_step_0 = const()[name = tensor("range_1d_34_step_0"), val = tensor(1)]; + tensor range_1d_34 = range_1d(end = reduce_prod_34, start = range_1d_34_start_0, step = range_1d_34_step_0)[name = tensor("range_1d_34")]; + tensor reshape_170 = reshape(shape = shape_82, x = range_1d_34)[name = tensor("reshape_170")]; + tensor slice_by_index_34 = slice_by_index(begin = concat_121, begin_mask = new_cache_35_internal_tensor_assign_1_begin_mask_0, end = concat_122, end_mask = new_cache_35_internal_tensor_assign_1_end_mask_0, squeeze_mask = new_cache_35_internal_tensor_assign_1_squeeze_mask_0, stride = new_cache_35_internal_tensor_assign_1_stride_0, x = reshape_170)[name = tensor("slice_by_index_34")]; + tensor reshape_171_shape_0 = const()[name = tensor("reshape_171_shape_0"), val = tensor([-1])]; + tensor reshape_171 = reshape(shape = reshape_171_shape_0, x = slice_by_index_34)[name = tensor("reshape_171")]; + tensor reshape_172_shape_0 = const()[name = tensor("reshape_172_shape_0"), val = tensor([-1])]; + tensor reshape_172_cast_fp16 = reshape(shape = reshape_172_shape_0, x = var_6997_cast_fp16)[name = tensor("reshape_172_cast_fp16")]; + tensor reshape_173_shape_0 = const()[name = tensor("reshape_173_shape_0"), val = tensor([-1])]; + tensor reshape_173_cast_fp16 = reshape(shape = reshape_173_shape_0, x = cache17_to_fp16)[name = tensor("reshape_173_cast_fp16")]; + tensor scatter_34_mode_0 = const()[name = tensor("scatter_34_mode_0"), val = tensor("update")]; + tensor scatter_34_axis_0 = const()[name = tensor("scatter_34_axis_0"), val = tensor(0)]; + tensor scatter_34_validate_indices_0 = const()[name = tensor("scatter_34_validate_indices_0"), val = tensor(false)]; + tensor scatter_34_cast_fp16 = scatter(axis = scatter_34_axis_0, data = reshape_173_cast_fp16, indices = reshape_171, mode = scatter_34_mode_0, updates = reshape_172_cast_fp16, validate_indices = scatter_34_validate_indices_0)[name = tensor("scatter_34_cast_fp16")]; + tensor reshape_174_cast_fp16 = reshape(shape = shape_82, x = scatter_34_cast_fp16)[name = tensor("reshape_174_cast_fp16")]; + tensor var_7005_begin_0 = const()[name = tensor("op_7005_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor var_7005_end_0 = const()[name = tensor("op_7005_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor var_7005_end_mask_0 = const()[name = tensor("op_7005_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_7005_squeeze_mask_0 = const()[name = tensor("op_7005_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor var_7005_cast_fp16 = slice_by_index(begin = var_7005_begin_0, end = var_7005_end_0, end_mask = var_7005_end_mask_0, squeeze_mask = var_7005_squeeze_mask_0, x = reshape_174_cast_fp16)[name = tensor("op_7005_cast_fp16")]; + tensor var_7007_axis_0 = const()[name = tensor("op_7007_axis_0"), val = tensor(1)]; + tensor var_7007_mode_0 = const()[name = tensor("op_7007_mode_0"), val = tensor("update")]; + tensor var_7007_validate_indices_0 = const()[name = tensor("op_7007_validate_indices_0"), val = tensor(false)]; + tensor var_7007_cast_fp16 = scatter_along_axis(axis = var_7007_axis_0, data = var_7005_cast_fp16, indices = write_indices_35, mode = var_7007_mode_0, updates = v_35_cast_fp16, validate_indices = var_7007_validate_indices_0)[name = tensor("op_7007_cast_fp16")]; + tensor concat_123 = const()[name = tensor("concat_123"), val = tensor([1, 0, 0, 0, 0])]; + tensor concat_124 = const()[name = tensor("concat_124"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_35_internal_tensor_assign_2_stride_0 = const()[name = tensor("new_cache_35_internal_tensor_assign_2_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_35_internal_tensor_assign_2_begin_mask_0 = const()[name = tensor("new_cache_35_internal_tensor_assign_2_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_35_internal_tensor_assign_2_end_mask_0 = const()[name = tensor("new_cache_35_internal_tensor_assign_2_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_35_internal_tensor_assign_2_squeeze_mask_0 = const()[name = tensor("new_cache_35_internal_tensor_assign_2_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_83 = const()[name = tensor("shape_83"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_35 = const()[name = tensor("reduce_prod_35"), val = tensor(1048576)]; + tensor range_1d_35_start_0 = const()[name = tensor("range_1d_35_start_0"), val = tensor(0)]; + tensor range_1d_35_step_0 = const()[name = tensor("range_1d_35_step_0"), val = tensor(1)]; + tensor range_1d_35 = range_1d(end = reduce_prod_35, start = range_1d_35_start_0, step = range_1d_35_step_0)[name = tensor("range_1d_35")]; + tensor reshape_175 = reshape(shape = shape_83, x = range_1d_35)[name = tensor("reshape_175")]; + tensor slice_by_index_35 = slice_by_index(begin = concat_123, begin_mask = new_cache_35_internal_tensor_assign_2_begin_mask_0, end = concat_124, end_mask = new_cache_35_internal_tensor_assign_2_end_mask_0, squeeze_mask = new_cache_35_internal_tensor_assign_2_squeeze_mask_0, stride = new_cache_35_internal_tensor_assign_2_stride_0, x = reshape_175)[name = tensor("slice_by_index_35")]; + tensor reshape_176_shape_0 = const()[name = tensor("reshape_176_shape_0"), val = tensor([-1])]; + tensor reshape_176 = reshape(shape = reshape_176_shape_0, x = slice_by_index_35)[name = tensor("reshape_176")]; + tensor reshape_177_shape_0 = const()[name = tensor("reshape_177_shape_0"), val = tensor([-1])]; + tensor reshape_177_cast_fp16 = reshape(shape = reshape_177_shape_0, x = var_7007_cast_fp16)[name = tensor("reshape_177_cast_fp16")]; + tensor reshape_178_shape_0 = const()[name = tensor("reshape_178_shape_0"), val = tensor([-1])]; + tensor reshape_178_cast_fp16 = reshape(shape = reshape_178_shape_0, x = reshape_174_cast_fp16)[name = tensor("reshape_178_cast_fp16")]; + tensor scatter_35_mode_0 = const()[name = tensor("scatter_35_mode_0"), val = tensor("update")]; + tensor scatter_35_axis_0 = const()[name = tensor("scatter_35_axis_0"), val = tensor(0)]; + tensor scatter_35_validate_indices_0 = const()[name = tensor("scatter_35_validate_indices_0"), val = tensor(false)]; + tensor scatter_35_cast_fp16 = scatter(axis = scatter_35_axis_0, data = reshape_178_cast_fp16, indices = reshape_176, mode = scatter_35_mode_0, updates = reshape_177_cast_fp16, validate_indices = scatter_35_validate_indices_0)[name = tensor("scatter_35_cast_fp16")]; + tensor reshape_179_cast_fp16 = reshape(shape = shape_83, x = scatter_35_cast_fp16)[name = tensor("reshape_179_cast_fp16")]; + tensor reshape_179_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("reshape_179_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor keys_103_begin_0 = const()[name = tensor("keys_103_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor keys_103_end_0 = const()[name = tensor("keys_103_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor keys_103_end_mask_0 = const()[name = tensor("keys_103_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor keys_103_squeeze_mask_0 = const()[name = tensor("keys_103_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor keys_103_cast_fp16 = slice_by_index(begin = keys_103_begin_0, end = keys_103_end_0, end_mask = keys_103_end_mask_0, squeeze_mask = keys_103_squeeze_mask_0, x = reshape_179_cast_fp16)[name = tensor("keys_103_cast_fp16")]; + tensor values_103_begin_0 = const()[name = tensor("values_103_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor values_103_end_0 = const()[name = tensor("values_103_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor values_103_end_mask_0 = const()[name = tensor("values_103_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor values_103_squeeze_mask_0 = const()[name = tensor("values_103_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor values_103_cast_fp16 = slice_by_index(begin = values_103_begin_0, end = values_103_end_0, end_mask = values_103_end_mask_0, squeeze_mask = values_103_squeeze_mask_0, x = reshape_179_cast_fp16)[name = tensor("values_103_cast_fp16")]; + tensor var_7019_cast_fp16 = not_equal(x = keys_103_cast_fp16, y = keys_103_cast_fp16)[name = tensor("op_7019_cast_fp16")]; + tensor keys_105_cast_fp16 = select(a = var_463_to_fp16, b = keys_103_cast_fp16, cond = var_7019_cast_fp16)[name = tensor("keys_105_cast_fp16")]; + tensor var_7027_cast_fp16 = not_equal(x = values_103_cast_fp16, y = values_103_cast_fp16)[name = tensor("op_7027_cast_fp16")]; + tensor values_105_cast_fp16 = select(a = var_463_to_fp16, b = values_103_cast_fp16, cond = var_7027_cast_fp16)[name = tensor("values_105_cast_fp16")]; + tensor var_7051 = const()[name = tensor("op_7051"), val = tensor([0, 2, 1, 3])]; + tensor var_7067 = const()[name = tensor("op_7067"), val = tensor([1, 1, 1])]; + tensor var_7068_cast_fp16 = reshape(shape = var_7067, x = position17_to_fp16)[name = tensor("op_7068_cast_fp16")]; + tensor var_7088_to_fp16 = const()[name = tensor("op_7088_to_fp16"), val = tensor(0x1p+0)]; + tensor valid_len_35_cast_fp16 = add(x = var_7068_cast_fp16, y = var_7088_to_fp16)[name = tensor("valid_len_35_cast_fp16")]; + tensor valid_mask_35_cast_fp16 = less(x = k_positions_1_promoted_to_fp16, y = valid_len_35_cast_fp16)[name = tensor("valid_mask_35_cast_fp16")]; + tensor causal_mask_35_cast_fp16 = less_equal(x = k_positions_1_promoted_to_fp16, y = var_7068_cast_fp16)[name = tensor("causal_mask_35_cast_fp16")]; + tensor attn_mask_69 = logical_and(x = valid_mask_35_cast_fp16, y = causal_mask_35_cast_fp16)[name = tensor("attn_mask_69")]; + tensor attn_mask_71_axes_0 = const()[name = tensor("attn_mask_71_axes_0"), val = tensor([1])]; + tensor attn_mask_71 = expand_dims(axes = attn_mask_71_axes_0, x = attn_mask_69)[name = tensor("attn_mask_71")]; + tensor var_7106_transpose_x_0 = const()[name = tensor("op_7106_transpose_x_0"), val = tensor(false)]; + tensor var_7106_transpose_y_0 = const()[name = tensor("op_7106_transpose_y_0"), val = tensor(false)]; + tensor transpose_106_perm_0 = const()[name = tensor("transpose_106_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_107_perm_0 = const()[name = tensor("transpose_107_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_107 = transpose(perm = transpose_107_perm_0, x = keys_105_cast_fp16)[name = tensor("transpose_145")]; + tensor transpose_106 = transpose(perm = transpose_106_perm_0, x = q_105_cast_fp16)[name = tensor("transpose_146")]; + tensor var_7106_cast_fp16 = matmul(transpose_x = var_7106_transpose_x_0, transpose_y = var_7106_transpose_y_0, x = transpose_106, y = transpose_107)[name = tensor("op_7106_cast_fp16")]; + tensor var_7100_to_fp16 = const()[name = tensor("op_7100_to_fp16"), val = tensor([0x1p-3])]; + tensor attn_weights_103_cast_fp16 = mul(x = var_7106_cast_fp16, y = var_7100_to_fp16)[name = tensor("attn_weights_103_cast_fp16")]; + tensor var_7108 = logical_not(x = attn_mask_71)[name = tensor("op_7108")]; + tensor var_7109_to_fp16 = const()[name = tensor("op_7109_to_fp16"), val = tensor(-inf)]; + tensor attn_weights_105_cast_fp16 = select(a = var_7109_to_fp16, b = attn_weights_103_cast_fp16, cond = var_7108)[name = tensor("attn_weights_105_cast_fp16")]; + tensor var_7111 = const()[name = tensor("op_7111"), val = tensor(-1)]; + tensor attn_weights_107_cast_fp16 = softmax(axis = var_7111, x = attn_weights_105_cast_fp16)[name = tensor("attn_weights_107_cast_fp16")]; + tensor attn_output_35_transpose_x_0 = const()[name = tensor("attn_output_35_transpose_x_0"), val = tensor(false)]; + tensor attn_output_35_transpose_y_0 = const()[name = tensor("attn_output_35_transpose_y_0"), val = tensor(false)]; + tensor values_107_cast_fp16 = transpose(perm = var_7051, x = values_105_cast_fp16)[name = tensor("transpose_147")]; + tensor attn_output_35_cast_fp16 = matmul(transpose_x = attn_output_35_transpose_x_0, transpose_y = attn_output_35_transpose_y_0, x = attn_weights_107_cast_fp16, y = values_107_cast_fp16)[name = tensor("attn_output_35_cast_fp16")]; + tensor var_7119 = const()[name = tensor("op_7119"), val = tensor([0, 2, 1, 3])]; + tensor var_7124 = const()[name = tensor("op_7124"), val = tensor([1, 1, 1024])]; + tensor var_7120_cast_fp16 = transpose(perm = var_7119, x = attn_output_35_cast_fp16)[name = tensor("transpose_144")]; + tensor input_211_cast_fp16 = reshape(shape = var_7124, x = var_7120_cast_fp16)[name = tensor("input_211_cast_fp16")]; + tensor attn17_out_proj_weight_to_fp16 = const()[name = tensor("attn17_out_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(435396992)))]; + tensor linear_70_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = attn17_out_proj_weight_to_fp16, x = input_211_cast_fp16)[name = tensor("linear_70_cast_fp16")]; + tensor var_7130_to_fp16 = const()[name = tensor("op_7130_to_fp16"), val = tensor(0x1p+0)]; + tensor var_7131_cast_fp16 = add(x = position17_to_fp16, y = var_7130_to_fp16)[name = tensor("op_7131_cast_fp16")]; + tensor var_7131_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_7131_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor input_213_cast_fp16 = add(x = input_207_cast_fp16, y = linear_70_cast_fp16)[name = tensor("input_213_cast_fp16")]; + tensor input_215_axes_0 = const()[name = tensor("input_215_axes_0"), val = tensor([-1])]; + tensor norm17_2_weight_to_fp16 = const()[name = tensor("norm17_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(437494208)))]; + tensor norm17_2_bias_to_fp16 = const()[name = tensor("norm17_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(437496320)))]; + tensor var_7135_to_fp16 = const()[name = tensor("op_7135_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_215_cast_fp16 = layer_norm(axes = input_215_axes_0, beta = norm17_2_bias_to_fp16, epsilon = var_7135_to_fp16, gamma = norm17_2_weight_to_fp16, x = input_213_cast_fp16)[name = tensor("input_215_cast_fp16")]; + tensor linear17_1_weight_to_fp16 = const()[name = tensor("linear17_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(437498432)))]; + tensor linear_71_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = linear17_1_weight_to_fp16, x = input_215_cast_fp16)[name = tensor("linear_71_cast_fp16")]; + tensor input_217_mode_0 = const()[name = tensor("input_217_mode_0"), val = tensor("EXACT")]; + tensor input_217_cast_fp16 = gelu(mode = input_217_mode_0, x = linear_71_cast_fp16)[name = tensor("input_217_cast_fp16")]; + tensor linear17_2_weight_to_fp16 = const()[name = tensor("linear17_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(445887104)))]; + tensor linear_72_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = linear17_2_weight_to_fp16, x = input_217_cast_fp16)[name = tensor("linear_72_cast_fp16")]; + tensor input_219_cast_fp16 = add(x = input_213_cast_fp16, y = linear_72_cast_fp16)[name = tensor("input_219_cast_fp16")]; + tensor input_221_axes_0 = const()[name = tensor("input_221_axes_0"), val = tensor([-1])]; + tensor norm18_1_weight_to_fp16 = const()[name = tensor("norm18_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(454275776)))]; + tensor norm18_1_bias_to_fp16 = const()[name = tensor("norm18_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(454277888)))]; + tensor var_7152_to_fp16 = const()[name = tensor("op_7152_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_221_cast_fp16 = layer_norm(axes = input_221_axes_0, beta = norm18_1_bias_to_fp16, epsilon = var_7152_to_fp16, gamma = norm18_1_weight_to_fp16, x = input_219_cast_fp16)[name = tensor("input_221_cast_fp16")]; + tensor attn18_in_proj_weight_to_fp16 = const()[name = tensor("attn18_in_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(454280000)))]; + tensor linear_73_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = attn18_in_proj_weight_to_fp16, x = input_221_cast_fp16)[name = tensor("linear_73_cast_fp16")]; + tensor var_7166 = const()[name = tensor("op_7166"), val = tensor([1, 1, 3, 16, 64])]; + tensor qkv_37_cast_fp16 = reshape(shape = var_7166, x = linear_73_cast_fp16)[name = tensor("qkv_37_cast_fp16")]; + tensor q_109_begin_0 = const()[name = tensor("q_109_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor q_109_end_0 = const()[name = tensor("q_109_end_0"), val = tensor([1, 1, 1, 16, 64])]; + tensor q_109_end_mask_0 = const()[name = tensor("q_109_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor q_109_squeeze_mask_0 = const()[name = tensor("q_109_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor q_109_cast_fp16 = slice_by_index(begin = q_109_begin_0, end = q_109_end_0, end_mask = q_109_end_mask_0, squeeze_mask = q_109_squeeze_mask_0, x = qkv_37_cast_fp16)[name = tensor("q_109_cast_fp16")]; + tensor k_73_begin_0 = const()[name = tensor("k_73_begin_0"), val = tensor([0, 0, 1, 0, 0])]; + tensor k_73_end_0 = const()[name = tensor("k_73_end_0"), val = tensor([1, 1, 2, 16, 64])]; + tensor k_73_end_mask_0 = const()[name = tensor("k_73_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor k_73_squeeze_mask_0 = const()[name = tensor("k_73_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor k_73_cast_fp16 = slice_by_index(begin = k_73_begin_0, end = k_73_end_0, end_mask = k_73_end_mask_0, squeeze_mask = k_73_squeeze_mask_0, x = qkv_37_cast_fp16)[name = tensor("k_73_cast_fp16")]; + tensor v_37_begin_0 = const()[name = tensor("v_37_begin_0"), val = tensor([0, 0, 2, 0, 0])]; + tensor v_37_end_0 = const()[name = tensor("v_37_end_0"), val = tensor([1, 1, 3, 16, 64])]; + tensor v_37_end_mask_0 = const()[name = tensor("v_37_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor v_37_squeeze_mask_0 = const()[name = tensor("v_37_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor v_37_cast_fp16 = slice_by_index(begin = v_37_begin_0, end = v_37_end_0, end_mask = v_37_end_mask_0, squeeze_mask = v_37_squeeze_mask_0, x = qkv_37_cast_fp16)[name = tensor("v_37_cast_fp16")]; + tensor var_7232 = const()[name = tensor("op_7232"), val = tensor([1, 1, 1, 1])]; + tensor position18_to_fp16_dtype_0 = const()[name = tensor("position18_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor position18_to_fp16 = cast(dtype = position18_to_fp16_dtype_0, x = position18)[name = tensor("cast_463")]; + tensor ts_113_cast_fp16 = reshape(shape = var_7232, x = position18_to_fp16)[name = tensor("ts_113_cast_fp16")]; + tensor var_7239 = const()[name = tensor("op_7239"), val = tensor([1, 1, 16, 32, 2])]; + tensor q_complex_37_cast_fp16 = reshape(shape = var_7239, x = q_109_cast_fp16)[name = tensor("q_complex_37_cast_fp16")]; + tensor var_7246 = const()[name = tensor("op_7246"), val = tensor([1, 1, 16, 32, 2])]; + tensor k_complex_37_cast_fp16 = reshape(shape = var_7246, x = k_73_cast_fp16)[name = tensor("k_complex_37_cast_fp16")]; + tensor var_7250_begin_0 = const()[name = tensor("op_7250_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_7250_end_0 = const()[name = tensor("op_7250_end_0"), val = tensor([1, 1, 16, 32, 1])]; + tensor var_7250_end_mask_0 = const()[name = tensor("op_7250_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_7250_squeeze_mask_0 = const()[name = tensor("op_7250_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_7250_cast_fp16 = slice_by_index(begin = var_7250_begin_0, end = var_7250_end_0, end_mask = var_7250_end_mask_0, squeeze_mask = var_7250_squeeze_mask_0, x = q_complex_37_cast_fp16)[name = tensor("op_7250_cast_fp16")]; + tensor var_7258_begin_0 = const()[name = tensor("op_7258_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_7258_end_0 = const()[name = tensor("op_7258_end_0"), val = tensor([1, 1, 16, 32, 2])]; + tensor var_7258_end_mask_0 = const()[name = tensor("op_7258_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_7258_squeeze_mask_0 = const()[name = tensor("op_7258_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_7258_cast_fp16 = slice_by_index(begin = var_7258_begin_0, end = var_7258_end_0, end_mask = var_7258_end_mask_0, squeeze_mask = var_7258_squeeze_mask_0, x = q_complex_37_cast_fp16)[name = tensor("op_7258_cast_fp16")]; + tensor var_7266_begin_0 = const()[name = tensor("op_7266_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_7266_end_0 = const()[name = tensor("op_7266_end_0"), val = tensor([1, 1, 16, 32, 1])]; + tensor var_7266_end_mask_0 = const()[name = tensor("op_7266_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_7266_squeeze_mask_0 = const()[name = tensor("op_7266_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_7266_cast_fp16 = slice_by_index(begin = var_7266_begin_0, end = var_7266_end_0, end_mask = var_7266_end_mask_0, squeeze_mask = var_7266_squeeze_mask_0, x = k_complex_37_cast_fp16)[name = tensor("op_7266_cast_fp16")]; + tensor var_7274_begin_0 = const()[name = tensor("op_7274_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_7274_end_0 = const()[name = tensor("op_7274_end_0"), val = tensor([1, 1, 16, 32, 2])]; + tensor var_7274_end_mask_0 = const()[name = tensor("op_7274_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_7274_squeeze_mask_0 = const()[name = tensor("op_7274_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_7274_cast_fp16 = slice_by_index(begin = var_7274_begin_0, end = var_7274_end_0, end_mask = var_7274_end_mask_0, squeeze_mask = var_7274_squeeze_mask_0, x = k_complex_37_cast_fp16)[name = tensor("op_7274_cast_fp16")]; + tensor freqs_37_to_fp16 = const()[name = tensor("freqs_37_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(460571520)))]; + tensor var_7280_cast_fp16 = mul(x = freqs_37_to_fp16, y = ts_113_cast_fp16)[name = tensor("op_7280_cast_fp16")]; + tensor rotr_37_cast_fp16 = cos(x = var_7280_cast_fp16)[name = tensor("rotr_37_cast_fp16")]; + tensor roti_37_cast_fp16 = sin(x = var_7280_cast_fp16)[name = tensor("roti_37_cast_fp16")]; + tensor var_7284_cast_fp16 = mul(x = var_7250_cast_fp16, y = rotr_37_cast_fp16)[name = tensor("op_7284_cast_fp16")]; + tensor var_7285_cast_fp16 = mul(x = var_7258_cast_fp16, y = roti_37_cast_fp16)[name = tensor("op_7285_cast_fp16")]; + tensor qor_73_cast_fp16 = sub(x = var_7284_cast_fp16, y = var_7285_cast_fp16)[name = tensor("qor_73_cast_fp16")]; + tensor var_7288_cast_fp16 = mul(x = var_7250_cast_fp16, y = roti_37_cast_fp16)[name = tensor("op_7288_cast_fp16")]; + tensor var_7289_cast_fp16 = mul(x = var_7258_cast_fp16, y = rotr_37_cast_fp16)[name = tensor("op_7289_cast_fp16")]; + tensor qoi_73_cast_fp16 = add(x = var_7288_cast_fp16, y = var_7289_cast_fp16)[name = tensor("qoi_73_cast_fp16")]; + tensor var_7292_cast_fp16 = mul(x = var_7266_cast_fp16, y = rotr_37_cast_fp16)[name = tensor("op_7292_cast_fp16")]; + tensor var_7293_cast_fp16 = mul(x = var_7274_cast_fp16, y = roti_37_cast_fp16)[name = tensor("op_7293_cast_fp16")]; + tensor kor_73_cast_fp16 = sub(x = var_7292_cast_fp16, y = var_7293_cast_fp16)[name = tensor("kor_73_cast_fp16")]; + tensor var_7296_cast_fp16 = mul(x = var_7266_cast_fp16, y = roti_37_cast_fp16)[name = tensor("op_7296_cast_fp16")]; + tensor var_7297_cast_fp16 = mul(x = var_7274_cast_fp16, y = rotr_37_cast_fp16)[name = tensor("op_7297_cast_fp16")]; + tensor koi_73_cast_fp16 = add(x = var_7296_cast_fp16, y = var_7297_cast_fp16)[name = tensor("koi_73_cast_fp16")]; + tensor qo_37_axis_0 = const()[name = tensor("qo_37_axis_0"), val = tensor(-1)]; + tensor qo_37_cast_fp16 = stack(axis = qo_37_axis_0, values = (qor_73_cast_fp16, qoi_73_cast_fp16))[name = tensor("qo_37_cast_fp16")]; + tensor ko_37_axis_0 = const()[name = tensor("ko_37_axis_0"), val = tensor(-1)]; + tensor ko_37_cast_fp16 = stack(axis = ko_37_axis_0, values = (kor_73_cast_fp16, koi_73_cast_fp16))[name = tensor("ko_37_cast_fp16")]; + tensor var_7330 = const()[name = tensor("op_7330"), val = tensor([1, 1, 16, 64])]; + tensor q_111_cast_fp16 = reshape(shape = var_7330, x = qo_37_cast_fp16)[name = tensor("q_111_cast_fp16")]; + tensor var_7336 = const()[name = tensor("op_7336"), val = tensor([1, 1, 16, 64])]; + tensor k_75_cast_fp16 = reshape(shape = var_7336, x = ko_37_cast_fp16)[name = tensor("k_75_cast_fp16")]; + tensor _inversed_7361_y_0_to_fp16 = const()[name = tensor("_inversed_7361_y_0_to_fp16"), val = tensor(0x1p-9)]; + tensor _inversed_7361_cast_fp16 = mul(x = ts_113_cast_fp16, y = _inversed_7361_y_0_to_fp16)[name = tensor("_inversed_7361_cast_fp16")]; + tensor var_7362_cast_fp16 = floor(x = _inversed_7361_cast_fp16)[name = tensor("op_7362_cast_fp16")]; + tensor var_7363_to_fp16 = const()[name = tensor("op_7363_to_fp16"), val = tensor(0x1p+9)]; + tensor var_7364_cast_fp16 = mul(x = var_7362_cast_fp16, y = var_7363_to_fp16)[name = tensor("op_7364_cast_fp16")]; + tensor write_indices_float_75_cast_fp16 = sub(x = ts_113_cast_fp16, y = var_7364_cast_fp16)[name = tensor("write_indices_float_75_cast_fp16")]; + tensor var_7371_dtype_0 = const()[name = tensor("op_7371_dtype_0"), val = tensor("int32")]; + tensor write_indices_37_reps_0 = const()[name = tensor("write_indices_37_reps_0"), val = tensor([1, 1, 16, 64])]; + tensor write_indices_float_75_cast_fp16_to_int32 = cast(dtype = var_7371_dtype_0, x = write_indices_float_75_cast_fp16)[name = tensor("cast_462")]; + tensor write_indices_37 = tile(reps = write_indices_37_reps_0, x = write_indices_float_75_cast_fp16_to_int32)[name = tensor("write_indices_37")]; + tensor var_7381_begin_0 = const()[name = tensor("op_7381_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_7381_end_0 = const()[name = tensor("op_7381_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor var_7381_end_mask_0 = const()[name = tensor("op_7381_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_7381_squeeze_mask_0 = const()[name = tensor("op_7381_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor cache18_to_fp16_dtype_0 = const()[name = tensor("cache18_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor cache18_to_fp16 = cast(dtype = cache18_to_fp16_dtype_0, x = cache18)[name = tensor("cast_461")]; + tensor var_7381_cast_fp16 = slice_by_index(begin = var_7381_begin_0, end = var_7381_end_0, end_mask = var_7381_end_mask_0, squeeze_mask = var_7381_squeeze_mask_0, x = cache18_to_fp16)[name = tensor("op_7381_cast_fp16")]; + tensor var_7383_axis_0 = const()[name = tensor("op_7383_axis_0"), val = tensor(1)]; + tensor var_7383_mode_0 = const()[name = tensor("op_7383_mode_0"), val = tensor("update")]; + tensor var_7383_validate_indices_0 = const()[name = tensor("op_7383_validate_indices_0"), val = tensor(false)]; + tensor var_7383_cast_fp16 = scatter_along_axis(axis = var_7383_axis_0, data = var_7381_cast_fp16, indices = write_indices_37, mode = var_7383_mode_0, updates = k_75_cast_fp16, validate_indices = var_7383_validate_indices_0)[name = tensor("op_7383_cast_fp16")]; + tensor concat_128 = const()[name = tensor("concat_128"), val = tensor([0, 0, 0, 0, 0])]; + tensor concat_129 = const()[name = tensor("concat_129"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_37_internal_tensor_assign_1_stride_0 = const()[name = tensor("new_cache_37_internal_tensor_assign_1_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_37_internal_tensor_assign_1_begin_mask_0 = const()[name = tensor("new_cache_37_internal_tensor_assign_1_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_37_internal_tensor_assign_1_end_mask_0 = const()[name = tensor("new_cache_37_internal_tensor_assign_1_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_37_internal_tensor_assign_1_squeeze_mask_0 = const()[name = tensor("new_cache_37_internal_tensor_assign_1_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_84 = const()[name = tensor("shape_84"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_36 = const()[name = tensor("reduce_prod_36"), val = tensor(1048576)]; + tensor range_1d_36_start_0 = const()[name = tensor("range_1d_36_start_0"), val = tensor(0)]; + tensor range_1d_36_step_0 = const()[name = tensor("range_1d_36_step_0"), val = tensor(1)]; + tensor range_1d_36 = range_1d(end = reduce_prod_36, start = range_1d_36_start_0, step = range_1d_36_step_0)[name = tensor("range_1d_36")]; + tensor reshape_180 = reshape(shape = shape_84, x = range_1d_36)[name = tensor("reshape_180")]; + tensor slice_by_index_36 = slice_by_index(begin = concat_128, begin_mask = new_cache_37_internal_tensor_assign_1_begin_mask_0, end = concat_129, end_mask = new_cache_37_internal_tensor_assign_1_end_mask_0, squeeze_mask = new_cache_37_internal_tensor_assign_1_squeeze_mask_0, stride = new_cache_37_internal_tensor_assign_1_stride_0, x = reshape_180)[name = tensor("slice_by_index_36")]; + tensor reshape_181_shape_0 = const()[name = tensor("reshape_181_shape_0"), val = tensor([-1])]; + tensor reshape_181 = reshape(shape = reshape_181_shape_0, x = slice_by_index_36)[name = tensor("reshape_181")]; + tensor reshape_182_shape_0 = const()[name = tensor("reshape_182_shape_0"), val = tensor([-1])]; + tensor reshape_182_cast_fp16 = reshape(shape = reshape_182_shape_0, x = var_7383_cast_fp16)[name = tensor("reshape_182_cast_fp16")]; + tensor reshape_183_shape_0 = const()[name = tensor("reshape_183_shape_0"), val = tensor([-1])]; + tensor reshape_183_cast_fp16 = reshape(shape = reshape_183_shape_0, x = cache18_to_fp16)[name = tensor("reshape_183_cast_fp16")]; + tensor scatter_36_mode_0 = const()[name = tensor("scatter_36_mode_0"), val = tensor("update")]; + tensor scatter_36_axis_0 = const()[name = tensor("scatter_36_axis_0"), val = tensor(0)]; + tensor scatter_36_validate_indices_0 = const()[name = tensor("scatter_36_validate_indices_0"), val = tensor(false)]; + tensor scatter_36_cast_fp16 = scatter(axis = scatter_36_axis_0, data = reshape_183_cast_fp16, indices = reshape_181, mode = scatter_36_mode_0, updates = reshape_182_cast_fp16, validate_indices = scatter_36_validate_indices_0)[name = tensor("scatter_36_cast_fp16")]; + tensor reshape_184_cast_fp16 = reshape(shape = shape_84, x = scatter_36_cast_fp16)[name = tensor("reshape_184_cast_fp16")]; + tensor var_7391_begin_0 = const()[name = tensor("op_7391_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor var_7391_end_0 = const()[name = tensor("op_7391_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor var_7391_end_mask_0 = const()[name = tensor("op_7391_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_7391_squeeze_mask_0 = const()[name = tensor("op_7391_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor var_7391_cast_fp16 = slice_by_index(begin = var_7391_begin_0, end = var_7391_end_0, end_mask = var_7391_end_mask_0, squeeze_mask = var_7391_squeeze_mask_0, x = reshape_184_cast_fp16)[name = tensor("op_7391_cast_fp16")]; + tensor var_7393_axis_0 = const()[name = tensor("op_7393_axis_0"), val = tensor(1)]; + tensor var_7393_mode_0 = const()[name = tensor("op_7393_mode_0"), val = tensor("update")]; + tensor var_7393_validate_indices_0 = const()[name = tensor("op_7393_validate_indices_0"), val = tensor(false)]; + tensor var_7393_cast_fp16 = scatter_along_axis(axis = var_7393_axis_0, data = var_7391_cast_fp16, indices = write_indices_37, mode = var_7393_mode_0, updates = v_37_cast_fp16, validate_indices = var_7393_validate_indices_0)[name = tensor("op_7393_cast_fp16")]; + tensor concat_130 = const()[name = tensor("concat_130"), val = tensor([1, 0, 0, 0, 0])]; + tensor concat_131 = const()[name = tensor("concat_131"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_37_internal_tensor_assign_2_stride_0 = const()[name = tensor("new_cache_37_internal_tensor_assign_2_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_37_internal_tensor_assign_2_begin_mask_0 = const()[name = tensor("new_cache_37_internal_tensor_assign_2_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_37_internal_tensor_assign_2_end_mask_0 = const()[name = tensor("new_cache_37_internal_tensor_assign_2_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_37_internal_tensor_assign_2_squeeze_mask_0 = const()[name = tensor("new_cache_37_internal_tensor_assign_2_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_85 = const()[name = tensor("shape_85"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_37 = const()[name = tensor("reduce_prod_37"), val = tensor(1048576)]; + tensor range_1d_37_start_0 = const()[name = tensor("range_1d_37_start_0"), val = tensor(0)]; + tensor range_1d_37_step_0 = const()[name = tensor("range_1d_37_step_0"), val = tensor(1)]; + tensor range_1d_37 = range_1d(end = reduce_prod_37, start = range_1d_37_start_0, step = range_1d_37_step_0)[name = tensor("range_1d_37")]; + tensor reshape_185 = reshape(shape = shape_85, x = range_1d_37)[name = tensor("reshape_185")]; + tensor slice_by_index_37 = slice_by_index(begin = concat_130, begin_mask = new_cache_37_internal_tensor_assign_2_begin_mask_0, end = concat_131, end_mask = new_cache_37_internal_tensor_assign_2_end_mask_0, squeeze_mask = new_cache_37_internal_tensor_assign_2_squeeze_mask_0, stride = new_cache_37_internal_tensor_assign_2_stride_0, x = reshape_185)[name = tensor("slice_by_index_37")]; + tensor reshape_186_shape_0 = const()[name = tensor("reshape_186_shape_0"), val = tensor([-1])]; + tensor reshape_186 = reshape(shape = reshape_186_shape_0, x = slice_by_index_37)[name = tensor("reshape_186")]; + tensor reshape_187_shape_0 = const()[name = tensor("reshape_187_shape_0"), val = tensor([-1])]; + tensor reshape_187_cast_fp16 = reshape(shape = reshape_187_shape_0, x = var_7393_cast_fp16)[name = tensor("reshape_187_cast_fp16")]; + tensor reshape_188_shape_0 = const()[name = tensor("reshape_188_shape_0"), val = tensor([-1])]; + tensor reshape_188_cast_fp16 = reshape(shape = reshape_188_shape_0, x = reshape_184_cast_fp16)[name = tensor("reshape_188_cast_fp16")]; + tensor scatter_37_mode_0 = const()[name = tensor("scatter_37_mode_0"), val = tensor("update")]; + tensor scatter_37_axis_0 = const()[name = tensor("scatter_37_axis_0"), val = tensor(0)]; + tensor scatter_37_validate_indices_0 = const()[name = tensor("scatter_37_validate_indices_0"), val = tensor(false)]; + tensor scatter_37_cast_fp16 = scatter(axis = scatter_37_axis_0, data = reshape_188_cast_fp16, indices = reshape_186, mode = scatter_37_mode_0, updates = reshape_187_cast_fp16, validate_indices = scatter_37_validate_indices_0)[name = tensor("scatter_37_cast_fp16")]; + tensor reshape_189_cast_fp16 = reshape(shape = shape_85, x = scatter_37_cast_fp16)[name = tensor("reshape_189_cast_fp16")]; + tensor reshape_189_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("reshape_189_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor keys_109_begin_0 = const()[name = tensor("keys_109_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor keys_109_end_0 = const()[name = tensor("keys_109_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor keys_109_end_mask_0 = const()[name = tensor("keys_109_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor keys_109_squeeze_mask_0 = const()[name = tensor("keys_109_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor keys_109_cast_fp16 = slice_by_index(begin = keys_109_begin_0, end = keys_109_end_0, end_mask = keys_109_end_mask_0, squeeze_mask = keys_109_squeeze_mask_0, x = reshape_189_cast_fp16)[name = tensor("keys_109_cast_fp16")]; + tensor values_109_begin_0 = const()[name = tensor("values_109_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor values_109_end_0 = const()[name = tensor("values_109_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor values_109_end_mask_0 = const()[name = tensor("values_109_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor values_109_squeeze_mask_0 = const()[name = tensor("values_109_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor values_109_cast_fp16 = slice_by_index(begin = values_109_begin_0, end = values_109_end_0, end_mask = values_109_end_mask_0, squeeze_mask = values_109_squeeze_mask_0, x = reshape_189_cast_fp16)[name = tensor("values_109_cast_fp16")]; + tensor var_7405_cast_fp16 = not_equal(x = keys_109_cast_fp16, y = keys_109_cast_fp16)[name = tensor("op_7405_cast_fp16")]; + tensor keys_111_cast_fp16 = select(a = var_463_to_fp16, b = keys_109_cast_fp16, cond = var_7405_cast_fp16)[name = tensor("keys_111_cast_fp16")]; + tensor var_7413_cast_fp16 = not_equal(x = values_109_cast_fp16, y = values_109_cast_fp16)[name = tensor("op_7413_cast_fp16")]; + tensor values_111_cast_fp16 = select(a = var_463_to_fp16, b = values_109_cast_fp16, cond = var_7413_cast_fp16)[name = tensor("values_111_cast_fp16")]; + tensor var_7437 = const()[name = tensor("op_7437"), val = tensor([0, 2, 1, 3])]; + tensor var_7453 = const()[name = tensor("op_7453"), val = tensor([1, 1, 1])]; + tensor var_7454_cast_fp16 = reshape(shape = var_7453, x = position18_to_fp16)[name = tensor("op_7454_cast_fp16")]; + tensor var_7474_to_fp16 = const()[name = tensor("op_7474_to_fp16"), val = tensor(0x1p+0)]; + tensor valid_len_37_cast_fp16 = add(x = var_7454_cast_fp16, y = var_7474_to_fp16)[name = tensor("valid_len_37_cast_fp16")]; + tensor valid_mask_37_cast_fp16 = less(x = k_positions_1_promoted_to_fp16, y = valid_len_37_cast_fp16)[name = tensor("valid_mask_37_cast_fp16")]; + tensor causal_mask_37_cast_fp16 = less_equal(x = k_positions_1_promoted_to_fp16, y = var_7454_cast_fp16)[name = tensor("causal_mask_37_cast_fp16")]; + tensor attn_mask_73 = logical_and(x = valid_mask_37_cast_fp16, y = causal_mask_37_cast_fp16)[name = tensor("attn_mask_73")]; + tensor attn_mask_75_axes_0 = const()[name = tensor("attn_mask_75_axes_0"), val = tensor([1])]; + tensor attn_mask_75 = expand_dims(axes = attn_mask_75_axes_0, x = attn_mask_73)[name = tensor("attn_mask_75")]; + tensor var_7492_transpose_x_0 = const()[name = tensor("op_7492_transpose_x_0"), val = tensor(false)]; + tensor var_7492_transpose_y_0 = const()[name = tensor("op_7492_transpose_y_0"), val = tensor(false)]; + tensor transpose_108_perm_0 = const()[name = tensor("transpose_108_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_109_perm_0 = const()[name = tensor("transpose_109_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_109 = transpose(perm = transpose_109_perm_0, x = keys_111_cast_fp16)[name = tensor("transpose_141")]; + tensor transpose_108 = transpose(perm = transpose_108_perm_0, x = q_111_cast_fp16)[name = tensor("transpose_142")]; + tensor var_7492_cast_fp16 = matmul(transpose_x = var_7492_transpose_x_0, transpose_y = var_7492_transpose_y_0, x = transpose_108, y = transpose_109)[name = tensor("op_7492_cast_fp16")]; + tensor var_7486_to_fp16 = const()[name = tensor("op_7486_to_fp16"), val = tensor([0x1p-3])]; + tensor attn_weights_109_cast_fp16 = mul(x = var_7492_cast_fp16, y = var_7486_to_fp16)[name = tensor("attn_weights_109_cast_fp16")]; + tensor var_7494 = logical_not(x = attn_mask_75)[name = tensor("op_7494")]; + tensor var_7495_to_fp16 = const()[name = tensor("op_7495_to_fp16"), val = tensor(-inf)]; + tensor attn_weights_111_cast_fp16 = select(a = var_7495_to_fp16, b = attn_weights_109_cast_fp16, cond = var_7494)[name = tensor("attn_weights_111_cast_fp16")]; + tensor var_7497 = const()[name = tensor("op_7497"), val = tensor(-1)]; + tensor attn_weights_113_cast_fp16 = softmax(axis = var_7497, x = attn_weights_111_cast_fp16)[name = tensor("attn_weights_113_cast_fp16")]; + tensor attn_output_37_transpose_x_0 = const()[name = tensor("attn_output_37_transpose_x_0"), val = tensor(false)]; + tensor attn_output_37_transpose_y_0 = const()[name = tensor("attn_output_37_transpose_y_0"), val = tensor(false)]; + tensor values_113_cast_fp16 = transpose(perm = var_7437, x = values_111_cast_fp16)[name = tensor("transpose_143")]; + tensor attn_output_37_cast_fp16 = matmul(transpose_x = attn_output_37_transpose_x_0, transpose_y = attn_output_37_transpose_y_0, x = attn_weights_113_cast_fp16, y = values_113_cast_fp16)[name = tensor("attn_output_37_cast_fp16")]; + tensor var_7505 = const()[name = tensor("op_7505"), val = tensor([0, 2, 1, 3])]; + tensor var_7510 = const()[name = tensor("op_7510"), val = tensor([1, 1, 1024])]; + tensor var_7506_cast_fp16 = transpose(perm = var_7505, x = attn_output_37_cast_fp16)[name = tensor("transpose_140")]; + tensor input_223_cast_fp16 = reshape(shape = var_7510, x = var_7506_cast_fp16)[name = tensor("input_223_cast_fp16")]; + tensor attn18_out_proj_weight_to_fp16 = const()[name = tensor("attn18_out_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(460571648)))]; + tensor linear_74_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = attn18_out_proj_weight_to_fp16, x = input_223_cast_fp16)[name = tensor("linear_74_cast_fp16")]; + tensor var_7516_to_fp16 = const()[name = tensor("op_7516_to_fp16"), val = tensor(0x1p+0)]; + tensor var_7517_cast_fp16 = add(x = position18_to_fp16, y = var_7516_to_fp16)[name = tensor("op_7517_cast_fp16")]; + tensor var_7517_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_7517_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor input_225_cast_fp16 = add(x = input_219_cast_fp16, y = linear_74_cast_fp16)[name = tensor("input_225_cast_fp16")]; + tensor input_227_axes_0 = const()[name = tensor("input_227_axes_0"), val = tensor([-1])]; + tensor norm18_2_weight_to_fp16 = const()[name = tensor("norm18_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(462668864)))]; + tensor norm18_2_bias_to_fp16 = const()[name = tensor("norm18_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(462670976)))]; + tensor var_7521_to_fp16 = const()[name = tensor("op_7521_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_227_cast_fp16 = layer_norm(axes = input_227_axes_0, beta = norm18_2_bias_to_fp16, epsilon = var_7521_to_fp16, gamma = norm18_2_weight_to_fp16, x = input_225_cast_fp16)[name = tensor("input_227_cast_fp16")]; + tensor linear18_1_weight_to_fp16 = const()[name = tensor("linear18_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(462673088)))]; + tensor linear_75_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = linear18_1_weight_to_fp16, x = input_227_cast_fp16)[name = tensor("linear_75_cast_fp16")]; + tensor input_229_mode_0 = const()[name = tensor("input_229_mode_0"), val = tensor("EXACT")]; + tensor input_229_cast_fp16 = gelu(mode = input_229_mode_0, x = linear_75_cast_fp16)[name = tensor("input_229_cast_fp16")]; + tensor linear18_2_weight_to_fp16 = const()[name = tensor("linear18_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(471061760)))]; + tensor linear_76_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = linear18_2_weight_to_fp16, x = input_229_cast_fp16)[name = tensor("linear_76_cast_fp16")]; + tensor input_231_cast_fp16 = add(x = input_225_cast_fp16, y = linear_76_cast_fp16)[name = tensor("input_231_cast_fp16")]; + tensor input_233_axes_0 = const()[name = tensor("input_233_axes_0"), val = tensor([-1])]; + tensor norm19_1_weight_to_fp16 = const()[name = tensor("norm19_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(479450432)))]; + tensor norm19_1_bias_to_fp16 = const()[name = tensor("norm19_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(479452544)))]; + tensor var_7538_to_fp16 = const()[name = tensor("op_7538_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_233_cast_fp16 = layer_norm(axes = input_233_axes_0, beta = norm19_1_bias_to_fp16, epsilon = var_7538_to_fp16, gamma = norm19_1_weight_to_fp16, x = input_231_cast_fp16)[name = tensor("input_233_cast_fp16")]; + tensor attn19_in_proj_weight_to_fp16 = const()[name = tensor("attn19_in_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(479454656)))]; + tensor linear_77_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = attn19_in_proj_weight_to_fp16, x = input_233_cast_fp16)[name = tensor("linear_77_cast_fp16")]; + tensor var_7552 = const()[name = tensor("op_7552"), val = tensor([1, 1, 3, 16, 64])]; + tensor qkv_39_cast_fp16 = reshape(shape = var_7552, x = linear_77_cast_fp16)[name = tensor("qkv_39_cast_fp16")]; + tensor q_115_begin_0 = const()[name = tensor("q_115_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor q_115_end_0 = const()[name = tensor("q_115_end_0"), val = tensor([1, 1, 1, 16, 64])]; + tensor q_115_end_mask_0 = const()[name = tensor("q_115_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor q_115_squeeze_mask_0 = const()[name = tensor("q_115_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor q_115_cast_fp16 = slice_by_index(begin = q_115_begin_0, end = q_115_end_0, end_mask = q_115_end_mask_0, squeeze_mask = q_115_squeeze_mask_0, x = qkv_39_cast_fp16)[name = tensor("q_115_cast_fp16")]; + tensor k_77_begin_0 = const()[name = tensor("k_77_begin_0"), val = tensor([0, 0, 1, 0, 0])]; + tensor k_77_end_0 = const()[name = tensor("k_77_end_0"), val = tensor([1, 1, 2, 16, 64])]; + tensor k_77_end_mask_0 = const()[name = tensor("k_77_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor k_77_squeeze_mask_0 = const()[name = tensor("k_77_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor k_77_cast_fp16 = slice_by_index(begin = k_77_begin_0, end = k_77_end_0, end_mask = k_77_end_mask_0, squeeze_mask = k_77_squeeze_mask_0, x = qkv_39_cast_fp16)[name = tensor("k_77_cast_fp16")]; + tensor v_39_begin_0 = const()[name = tensor("v_39_begin_0"), val = tensor([0, 0, 2, 0, 0])]; + tensor v_39_end_0 = const()[name = tensor("v_39_end_0"), val = tensor([1, 1, 3, 16, 64])]; + tensor v_39_end_mask_0 = const()[name = tensor("v_39_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor v_39_squeeze_mask_0 = const()[name = tensor("v_39_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor v_39_cast_fp16 = slice_by_index(begin = v_39_begin_0, end = v_39_end_0, end_mask = v_39_end_mask_0, squeeze_mask = v_39_squeeze_mask_0, x = qkv_39_cast_fp16)[name = tensor("v_39_cast_fp16")]; + tensor var_7618 = const()[name = tensor("op_7618"), val = tensor([1, 1, 1, 1])]; + tensor position19_to_fp16_dtype_0 = const()[name = tensor("position19_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor position19_to_fp16 = cast(dtype = position19_to_fp16_dtype_0, x = position19)[name = tensor("cast_458")]; + tensor ts_119_cast_fp16 = reshape(shape = var_7618, x = position19_to_fp16)[name = tensor("ts_119_cast_fp16")]; + tensor var_7625 = const()[name = tensor("op_7625"), val = tensor([1, 1, 16, 32, 2])]; + tensor q_complex_39_cast_fp16 = reshape(shape = var_7625, x = q_115_cast_fp16)[name = tensor("q_complex_39_cast_fp16")]; + tensor var_7632 = const()[name = tensor("op_7632"), val = tensor([1, 1, 16, 32, 2])]; + tensor k_complex_39_cast_fp16 = reshape(shape = var_7632, x = k_77_cast_fp16)[name = tensor("k_complex_39_cast_fp16")]; + tensor var_7636_begin_0 = const()[name = tensor("op_7636_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_7636_end_0 = const()[name = tensor("op_7636_end_0"), val = tensor([1, 1, 16, 32, 1])]; + tensor var_7636_end_mask_0 = const()[name = tensor("op_7636_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_7636_squeeze_mask_0 = const()[name = tensor("op_7636_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_7636_cast_fp16 = slice_by_index(begin = var_7636_begin_0, end = var_7636_end_0, end_mask = var_7636_end_mask_0, squeeze_mask = var_7636_squeeze_mask_0, x = q_complex_39_cast_fp16)[name = tensor("op_7636_cast_fp16")]; + tensor var_7644_begin_0 = const()[name = tensor("op_7644_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_7644_end_0 = const()[name = tensor("op_7644_end_0"), val = tensor([1, 1, 16, 32, 2])]; + tensor var_7644_end_mask_0 = const()[name = tensor("op_7644_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_7644_squeeze_mask_0 = const()[name = tensor("op_7644_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_7644_cast_fp16 = slice_by_index(begin = var_7644_begin_0, end = var_7644_end_0, end_mask = var_7644_end_mask_0, squeeze_mask = var_7644_squeeze_mask_0, x = q_complex_39_cast_fp16)[name = tensor("op_7644_cast_fp16")]; + tensor var_7652_begin_0 = const()[name = tensor("op_7652_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_7652_end_0 = const()[name = tensor("op_7652_end_0"), val = tensor([1, 1, 16, 32, 1])]; + tensor var_7652_end_mask_0 = const()[name = tensor("op_7652_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_7652_squeeze_mask_0 = const()[name = tensor("op_7652_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_7652_cast_fp16 = slice_by_index(begin = var_7652_begin_0, end = var_7652_end_0, end_mask = var_7652_end_mask_0, squeeze_mask = var_7652_squeeze_mask_0, x = k_complex_39_cast_fp16)[name = tensor("op_7652_cast_fp16")]; + tensor var_7660_begin_0 = const()[name = tensor("op_7660_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_7660_end_0 = const()[name = tensor("op_7660_end_0"), val = tensor([1, 1, 16, 32, 2])]; + tensor var_7660_end_mask_0 = const()[name = tensor("op_7660_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_7660_squeeze_mask_0 = const()[name = tensor("op_7660_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_7660_cast_fp16 = slice_by_index(begin = var_7660_begin_0, end = var_7660_end_0, end_mask = var_7660_end_mask_0, squeeze_mask = var_7660_squeeze_mask_0, x = k_complex_39_cast_fp16)[name = tensor("op_7660_cast_fp16")]; + tensor freqs_39_to_fp16 = const()[name = tensor("freqs_39_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(485746176)))]; + tensor var_7666_cast_fp16 = mul(x = freqs_39_to_fp16, y = ts_119_cast_fp16)[name = tensor("op_7666_cast_fp16")]; + tensor rotr_39_cast_fp16 = cos(x = var_7666_cast_fp16)[name = tensor("rotr_39_cast_fp16")]; + tensor roti_39_cast_fp16 = sin(x = var_7666_cast_fp16)[name = tensor("roti_39_cast_fp16")]; + tensor var_7670_cast_fp16 = mul(x = var_7636_cast_fp16, y = rotr_39_cast_fp16)[name = tensor("op_7670_cast_fp16")]; + tensor var_7671_cast_fp16 = mul(x = var_7644_cast_fp16, y = roti_39_cast_fp16)[name = tensor("op_7671_cast_fp16")]; + tensor qor_77_cast_fp16 = sub(x = var_7670_cast_fp16, y = var_7671_cast_fp16)[name = tensor("qor_77_cast_fp16")]; + tensor var_7674_cast_fp16 = mul(x = var_7636_cast_fp16, y = roti_39_cast_fp16)[name = tensor("op_7674_cast_fp16")]; + tensor var_7675_cast_fp16 = mul(x = var_7644_cast_fp16, y = rotr_39_cast_fp16)[name = tensor("op_7675_cast_fp16")]; + tensor qoi_77_cast_fp16 = add(x = var_7674_cast_fp16, y = var_7675_cast_fp16)[name = tensor("qoi_77_cast_fp16")]; + tensor var_7678_cast_fp16 = mul(x = var_7652_cast_fp16, y = rotr_39_cast_fp16)[name = tensor("op_7678_cast_fp16")]; + tensor var_7679_cast_fp16 = mul(x = var_7660_cast_fp16, y = roti_39_cast_fp16)[name = tensor("op_7679_cast_fp16")]; + tensor kor_77_cast_fp16 = sub(x = var_7678_cast_fp16, y = var_7679_cast_fp16)[name = tensor("kor_77_cast_fp16")]; + tensor var_7682_cast_fp16 = mul(x = var_7652_cast_fp16, y = roti_39_cast_fp16)[name = tensor("op_7682_cast_fp16")]; + tensor var_7683_cast_fp16 = mul(x = var_7660_cast_fp16, y = rotr_39_cast_fp16)[name = tensor("op_7683_cast_fp16")]; + tensor koi_77_cast_fp16 = add(x = var_7682_cast_fp16, y = var_7683_cast_fp16)[name = tensor("koi_77_cast_fp16")]; + tensor qo_39_axis_0 = const()[name = tensor("qo_39_axis_0"), val = tensor(-1)]; + tensor qo_39_cast_fp16 = stack(axis = qo_39_axis_0, values = (qor_77_cast_fp16, qoi_77_cast_fp16))[name = tensor("qo_39_cast_fp16")]; + tensor ko_39_axis_0 = const()[name = tensor("ko_39_axis_0"), val = tensor(-1)]; + tensor ko_39_cast_fp16 = stack(axis = ko_39_axis_0, values = (kor_77_cast_fp16, koi_77_cast_fp16))[name = tensor("ko_39_cast_fp16")]; + tensor var_7716 = const()[name = tensor("op_7716"), val = tensor([1, 1, 16, 64])]; + tensor q_117_cast_fp16 = reshape(shape = var_7716, x = qo_39_cast_fp16)[name = tensor("q_117_cast_fp16")]; + tensor var_7722 = const()[name = tensor("op_7722"), val = tensor([1, 1, 16, 64])]; + tensor k_79_cast_fp16 = reshape(shape = var_7722, x = ko_39_cast_fp16)[name = tensor("k_79_cast_fp16")]; + tensor _inversed_7747_y_0_to_fp16 = const()[name = tensor("_inversed_7747_y_0_to_fp16"), val = tensor(0x1p-9)]; + tensor _inversed_7747_cast_fp16 = mul(x = ts_119_cast_fp16, y = _inversed_7747_y_0_to_fp16)[name = tensor("_inversed_7747_cast_fp16")]; + tensor var_7748_cast_fp16 = floor(x = _inversed_7747_cast_fp16)[name = tensor("op_7748_cast_fp16")]; + tensor var_7749_to_fp16 = const()[name = tensor("op_7749_to_fp16"), val = tensor(0x1p+9)]; + tensor var_7750_cast_fp16 = mul(x = var_7748_cast_fp16, y = var_7749_to_fp16)[name = tensor("op_7750_cast_fp16")]; + tensor write_indices_float_79_cast_fp16 = sub(x = ts_119_cast_fp16, y = var_7750_cast_fp16)[name = tensor("write_indices_float_79_cast_fp16")]; + tensor var_7757_dtype_0 = const()[name = tensor("op_7757_dtype_0"), val = tensor("int32")]; + tensor write_indices_39_reps_0 = const()[name = tensor("write_indices_39_reps_0"), val = tensor([1, 1, 16, 64])]; + tensor write_indices_float_79_cast_fp16_to_int32 = cast(dtype = var_7757_dtype_0, x = write_indices_float_79_cast_fp16)[name = tensor("cast_457")]; + tensor write_indices_39 = tile(reps = write_indices_39_reps_0, x = write_indices_float_79_cast_fp16_to_int32)[name = tensor("write_indices_39")]; + tensor var_7767_begin_0 = const()[name = tensor("op_7767_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_7767_end_0 = const()[name = tensor("op_7767_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor var_7767_end_mask_0 = const()[name = tensor("op_7767_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_7767_squeeze_mask_0 = const()[name = tensor("op_7767_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor cache19_to_fp16_dtype_0 = const()[name = tensor("cache19_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor cache19_to_fp16 = cast(dtype = cache19_to_fp16_dtype_0, x = cache19)[name = tensor("cast_456")]; + tensor var_7767_cast_fp16 = slice_by_index(begin = var_7767_begin_0, end = var_7767_end_0, end_mask = var_7767_end_mask_0, squeeze_mask = var_7767_squeeze_mask_0, x = cache19_to_fp16)[name = tensor("op_7767_cast_fp16")]; + tensor var_7769_axis_0 = const()[name = tensor("op_7769_axis_0"), val = tensor(1)]; + tensor var_7769_mode_0 = const()[name = tensor("op_7769_mode_0"), val = tensor("update")]; + tensor var_7769_validate_indices_0 = const()[name = tensor("op_7769_validate_indices_0"), val = tensor(false)]; + tensor var_7769_cast_fp16 = scatter_along_axis(axis = var_7769_axis_0, data = var_7767_cast_fp16, indices = write_indices_39, mode = var_7769_mode_0, updates = k_79_cast_fp16, validate_indices = var_7769_validate_indices_0)[name = tensor("op_7769_cast_fp16")]; + tensor concat_135 = const()[name = tensor("concat_135"), val = tensor([0, 0, 0, 0, 0])]; + tensor concat_136 = const()[name = tensor("concat_136"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_39_internal_tensor_assign_1_stride_0 = const()[name = tensor("new_cache_39_internal_tensor_assign_1_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_39_internal_tensor_assign_1_begin_mask_0 = const()[name = tensor("new_cache_39_internal_tensor_assign_1_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_39_internal_tensor_assign_1_end_mask_0 = const()[name = tensor("new_cache_39_internal_tensor_assign_1_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_39_internal_tensor_assign_1_squeeze_mask_0 = const()[name = tensor("new_cache_39_internal_tensor_assign_1_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_86 = const()[name = tensor("shape_86"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_38 = const()[name = tensor("reduce_prod_38"), val = tensor(1048576)]; + tensor range_1d_38_start_0 = const()[name = tensor("range_1d_38_start_0"), val = tensor(0)]; + tensor range_1d_38_step_0 = const()[name = tensor("range_1d_38_step_0"), val = tensor(1)]; + tensor range_1d_38 = range_1d(end = reduce_prod_38, start = range_1d_38_start_0, step = range_1d_38_step_0)[name = tensor("range_1d_38")]; + tensor reshape_190 = reshape(shape = shape_86, x = range_1d_38)[name = tensor("reshape_190")]; + tensor slice_by_index_38 = slice_by_index(begin = concat_135, begin_mask = new_cache_39_internal_tensor_assign_1_begin_mask_0, end = concat_136, end_mask = new_cache_39_internal_tensor_assign_1_end_mask_0, squeeze_mask = new_cache_39_internal_tensor_assign_1_squeeze_mask_0, stride = new_cache_39_internal_tensor_assign_1_stride_0, x = reshape_190)[name = tensor("slice_by_index_38")]; + tensor reshape_191_shape_0 = const()[name = tensor("reshape_191_shape_0"), val = tensor([-1])]; + tensor reshape_191 = reshape(shape = reshape_191_shape_0, x = slice_by_index_38)[name = tensor("reshape_191")]; + tensor reshape_192_shape_0 = const()[name = tensor("reshape_192_shape_0"), val = tensor([-1])]; + tensor reshape_192_cast_fp16 = reshape(shape = reshape_192_shape_0, x = var_7769_cast_fp16)[name = tensor("reshape_192_cast_fp16")]; + tensor reshape_193_shape_0 = const()[name = tensor("reshape_193_shape_0"), val = tensor([-1])]; + tensor reshape_193_cast_fp16 = reshape(shape = reshape_193_shape_0, x = cache19_to_fp16)[name = tensor("reshape_193_cast_fp16")]; + tensor scatter_38_mode_0 = const()[name = tensor("scatter_38_mode_0"), val = tensor("update")]; + tensor scatter_38_axis_0 = const()[name = tensor("scatter_38_axis_0"), val = tensor(0)]; + tensor scatter_38_validate_indices_0 = const()[name = tensor("scatter_38_validate_indices_0"), val = tensor(false)]; + tensor scatter_38_cast_fp16 = scatter(axis = scatter_38_axis_0, data = reshape_193_cast_fp16, indices = reshape_191, mode = scatter_38_mode_0, updates = reshape_192_cast_fp16, validate_indices = scatter_38_validate_indices_0)[name = tensor("scatter_38_cast_fp16")]; + tensor reshape_194_cast_fp16 = reshape(shape = shape_86, x = scatter_38_cast_fp16)[name = tensor("reshape_194_cast_fp16")]; + tensor var_7777_begin_0 = const()[name = tensor("op_7777_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor var_7777_end_0 = const()[name = tensor("op_7777_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor var_7777_end_mask_0 = const()[name = tensor("op_7777_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_7777_squeeze_mask_0 = const()[name = tensor("op_7777_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor var_7777_cast_fp16 = slice_by_index(begin = var_7777_begin_0, end = var_7777_end_0, end_mask = var_7777_end_mask_0, squeeze_mask = var_7777_squeeze_mask_0, x = reshape_194_cast_fp16)[name = tensor("op_7777_cast_fp16")]; + tensor var_7779_axis_0 = const()[name = tensor("op_7779_axis_0"), val = tensor(1)]; + tensor var_7779_mode_0 = const()[name = tensor("op_7779_mode_0"), val = tensor("update")]; + tensor var_7779_validate_indices_0 = const()[name = tensor("op_7779_validate_indices_0"), val = tensor(false)]; + tensor var_7779_cast_fp16 = scatter_along_axis(axis = var_7779_axis_0, data = var_7777_cast_fp16, indices = write_indices_39, mode = var_7779_mode_0, updates = v_39_cast_fp16, validate_indices = var_7779_validate_indices_0)[name = tensor("op_7779_cast_fp16")]; + tensor concat_137 = const()[name = tensor("concat_137"), val = tensor([1, 0, 0, 0, 0])]; + tensor concat_138 = const()[name = tensor("concat_138"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_39_internal_tensor_assign_2_stride_0 = const()[name = tensor("new_cache_39_internal_tensor_assign_2_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_39_internal_tensor_assign_2_begin_mask_0 = const()[name = tensor("new_cache_39_internal_tensor_assign_2_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_39_internal_tensor_assign_2_end_mask_0 = const()[name = tensor("new_cache_39_internal_tensor_assign_2_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_39_internal_tensor_assign_2_squeeze_mask_0 = const()[name = tensor("new_cache_39_internal_tensor_assign_2_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_87 = const()[name = tensor("shape_87"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_39 = const()[name = tensor("reduce_prod_39"), val = tensor(1048576)]; + tensor range_1d_39_start_0 = const()[name = tensor("range_1d_39_start_0"), val = tensor(0)]; + tensor range_1d_39_step_0 = const()[name = tensor("range_1d_39_step_0"), val = tensor(1)]; + tensor range_1d_39 = range_1d(end = reduce_prod_39, start = range_1d_39_start_0, step = range_1d_39_step_0)[name = tensor("range_1d_39")]; + tensor reshape_195 = reshape(shape = shape_87, x = range_1d_39)[name = tensor("reshape_195")]; + tensor slice_by_index_39 = slice_by_index(begin = concat_137, begin_mask = new_cache_39_internal_tensor_assign_2_begin_mask_0, end = concat_138, end_mask = new_cache_39_internal_tensor_assign_2_end_mask_0, squeeze_mask = new_cache_39_internal_tensor_assign_2_squeeze_mask_0, stride = new_cache_39_internal_tensor_assign_2_stride_0, x = reshape_195)[name = tensor("slice_by_index_39")]; + tensor reshape_196_shape_0 = const()[name = tensor("reshape_196_shape_0"), val = tensor([-1])]; + tensor reshape_196 = reshape(shape = reshape_196_shape_0, x = slice_by_index_39)[name = tensor("reshape_196")]; + tensor reshape_197_shape_0 = const()[name = tensor("reshape_197_shape_0"), val = tensor([-1])]; + tensor reshape_197_cast_fp16 = reshape(shape = reshape_197_shape_0, x = var_7779_cast_fp16)[name = tensor("reshape_197_cast_fp16")]; + tensor reshape_198_shape_0 = const()[name = tensor("reshape_198_shape_0"), val = tensor([-1])]; + tensor reshape_198_cast_fp16 = reshape(shape = reshape_198_shape_0, x = reshape_194_cast_fp16)[name = tensor("reshape_198_cast_fp16")]; + tensor scatter_39_mode_0 = const()[name = tensor("scatter_39_mode_0"), val = tensor("update")]; + tensor scatter_39_axis_0 = const()[name = tensor("scatter_39_axis_0"), val = tensor(0)]; + tensor scatter_39_validate_indices_0 = const()[name = tensor("scatter_39_validate_indices_0"), val = tensor(false)]; + tensor scatter_39_cast_fp16 = scatter(axis = scatter_39_axis_0, data = reshape_198_cast_fp16, indices = reshape_196, mode = scatter_39_mode_0, updates = reshape_197_cast_fp16, validate_indices = scatter_39_validate_indices_0)[name = tensor("scatter_39_cast_fp16")]; + tensor reshape_199_cast_fp16 = reshape(shape = shape_87, x = scatter_39_cast_fp16)[name = tensor("reshape_199_cast_fp16")]; + tensor reshape_199_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("reshape_199_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor keys_115_begin_0 = const()[name = tensor("keys_115_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor keys_115_end_0 = const()[name = tensor("keys_115_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor keys_115_end_mask_0 = const()[name = tensor("keys_115_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor keys_115_squeeze_mask_0 = const()[name = tensor("keys_115_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor keys_115_cast_fp16 = slice_by_index(begin = keys_115_begin_0, end = keys_115_end_0, end_mask = keys_115_end_mask_0, squeeze_mask = keys_115_squeeze_mask_0, x = reshape_199_cast_fp16)[name = tensor("keys_115_cast_fp16")]; + tensor values_115_begin_0 = const()[name = tensor("values_115_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor values_115_end_0 = const()[name = tensor("values_115_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor values_115_end_mask_0 = const()[name = tensor("values_115_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor values_115_squeeze_mask_0 = const()[name = tensor("values_115_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor values_115_cast_fp16 = slice_by_index(begin = values_115_begin_0, end = values_115_end_0, end_mask = values_115_end_mask_0, squeeze_mask = values_115_squeeze_mask_0, x = reshape_199_cast_fp16)[name = tensor("values_115_cast_fp16")]; + tensor var_7791_cast_fp16 = not_equal(x = keys_115_cast_fp16, y = keys_115_cast_fp16)[name = tensor("op_7791_cast_fp16")]; + tensor keys_117_cast_fp16 = select(a = var_463_to_fp16, b = keys_115_cast_fp16, cond = var_7791_cast_fp16)[name = tensor("keys_117_cast_fp16")]; + tensor var_7799_cast_fp16 = not_equal(x = values_115_cast_fp16, y = values_115_cast_fp16)[name = tensor("op_7799_cast_fp16")]; + tensor values_117_cast_fp16 = select(a = var_463_to_fp16, b = values_115_cast_fp16, cond = var_7799_cast_fp16)[name = tensor("values_117_cast_fp16")]; + tensor var_7823 = const()[name = tensor("op_7823"), val = tensor([0, 2, 1, 3])]; + tensor var_7839 = const()[name = tensor("op_7839"), val = tensor([1, 1, 1])]; + tensor var_7840_cast_fp16 = reshape(shape = var_7839, x = position19_to_fp16)[name = tensor("op_7840_cast_fp16")]; + tensor var_7860_to_fp16 = const()[name = tensor("op_7860_to_fp16"), val = tensor(0x1p+0)]; + tensor valid_len_39_cast_fp16 = add(x = var_7840_cast_fp16, y = var_7860_to_fp16)[name = tensor("valid_len_39_cast_fp16")]; + tensor valid_mask_39_cast_fp16 = less(x = k_positions_1_promoted_to_fp16, y = valid_len_39_cast_fp16)[name = tensor("valid_mask_39_cast_fp16")]; + tensor causal_mask_39_cast_fp16 = less_equal(x = k_positions_1_promoted_to_fp16, y = var_7840_cast_fp16)[name = tensor("causal_mask_39_cast_fp16")]; + tensor attn_mask_77 = logical_and(x = valid_mask_39_cast_fp16, y = causal_mask_39_cast_fp16)[name = tensor("attn_mask_77")]; + tensor attn_mask_79_axes_0 = const()[name = tensor("attn_mask_79_axes_0"), val = tensor([1])]; + tensor attn_mask_79 = expand_dims(axes = attn_mask_79_axes_0, x = attn_mask_77)[name = tensor("attn_mask_79")]; + tensor var_7878_transpose_x_0 = const()[name = tensor("op_7878_transpose_x_0"), val = tensor(false)]; + tensor var_7878_transpose_y_0 = const()[name = tensor("op_7878_transpose_y_0"), val = tensor(false)]; + tensor transpose_110_perm_0 = const()[name = tensor("transpose_110_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_111_perm_0 = const()[name = tensor("transpose_111_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_111 = transpose(perm = transpose_111_perm_0, x = keys_117_cast_fp16)[name = tensor("transpose_137")]; + tensor transpose_110 = transpose(perm = transpose_110_perm_0, x = q_117_cast_fp16)[name = tensor("transpose_138")]; + tensor var_7878_cast_fp16 = matmul(transpose_x = var_7878_transpose_x_0, transpose_y = var_7878_transpose_y_0, x = transpose_110, y = transpose_111)[name = tensor("op_7878_cast_fp16")]; + tensor var_7872_to_fp16 = const()[name = tensor("op_7872_to_fp16"), val = tensor([0x1p-3])]; + tensor attn_weights_115_cast_fp16 = mul(x = var_7878_cast_fp16, y = var_7872_to_fp16)[name = tensor("attn_weights_115_cast_fp16")]; + tensor var_7880 = logical_not(x = attn_mask_79)[name = tensor("op_7880")]; + tensor var_7881_to_fp16 = const()[name = tensor("op_7881_to_fp16"), val = tensor(-inf)]; + tensor attn_weights_117_cast_fp16 = select(a = var_7881_to_fp16, b = attn_weights_115_cast_fp16, cond = var_7880)[name = tensor("attn_weights_117_cast_fp16")]; + tensor var_7883 = const()[name = tensor("op_7883"), val = tensor(-1)]; + tensor attn_weights_119_cast_fp16 = softmax(axis = var_7883, x = attn_weights_117_cast_fp16)[name = tensor("attn_weights_119_cast_fp16")]; + tensor attn_output_39_transpose_x_0 = const()[name = tensor("attn_output_39_transpose_x_0"), val = tensor(false)]; + tensor attn_output_39_transpose_y_0 = const()[name = tensor("attn_output_39_transpose_y_0"), val = tensor(false)]; + tensor values_119_cast_fp16 = transpose(perm = var_7823, x = values_117_cast_fp16)[name = tensor("transpose_139")]; + tensor attn_output_39_cast_fp16 = matmul(transpose_x = attn_output_39_transpose_x_0, transpose_y = attn_output_39_transpose_y_0, x = attn_weights_119_cast_fp16, y = values_119_cast_fp16)[name = tensor("attn_output_39_cast_fp16")]; + tensor var_7891 = const()[name = tensor("op_7891"), val = tensor([0, 2, 1, 3])]; + tensor var_7896 = const()[name = tensor("op_7896"), val = tensor([1, 1, 1024])]; + tensor var_7892_cast_fp16 = transpose(perm = var_7891, x = attn_output_39_cast_fp16)[name = tensor("transpose_136")]; + tensor input_235_cast_fp16 = reshape(shape = var_7896, x = var_7892_cast_fp16)[name = tensor("input_235_cast_fp16")]; + tensor attn19_out_proj_weight_to_fp16 = const()[name = tensor("attn19_out_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(485746304)))]; + tensor linear_78_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = attn19_out_proj_weight_to_fp16, x = input_235_cast_fp16)[name = tensor("linear_78_cast_fp16")]; + tensor var_7902_to_fp16 = const()[name = tensor("op_7902_to_fp16"), val = tensor(0x1p+0)]; + tensor var_7903_cast_fp16 = add(x = position19_to_fp16, y = var_7902_to_fp16)[name = tensor("op_7903_cast_fp16")]; + tensor var_7903_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_7903_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor input_237_cast_fp16 = add(x = input_231_cast_fp16, y = linear_78_cast_fp16)[name = tensor("input_237_cast_fp16")]; + tensor input_239_axes_0 = const()[name = tensor("input_239_axes_0"), val = tensor([-1])]; + tensor norm19_2_weight_to_fp16 = const()[name = tensor("norm19_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(487843520)))]; + tensor norm19_2_bias_to_fp16 = const()[name = tensor("norm19_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(487845632)))]; + tensor var_7907_to_fp16 = const()[name = tensor("op_7907_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_239_cast_fp16 = layer_norm(axes = input_239_axes_0, beta = norm19_2_bias_to_fp16, epsilon = var_7907_to_fp16, gamma = norm19_2_weight_to_fp16, x = input_237_cast_fp16)[name = tensor("input_239_cast_fp16")]; + tensor linear19_1_weight_to_fp16 = const()[name = tensor("linear19_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(487847744)))]; + tensor linear_79_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = linear19_1_weight_to_fp16, x = input_239_cast_fp16)[name = tensor("linear_79_cast_fp16")]; + tensor input_241_mode_0 = const()[name = tensor("input_241_mode_0"), val = tensor("EXACT")]; + tensor input_241_cast_fp16 = gelu(mode = input_241_mode_0, x = linear_79_cast_fp16)[name = tensor("input_241_cast_fp16")]; + tensor linear19_2_weight_to_fp16 = const()[name = tensor("linear19_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(496236416)))]; + tensor linear_80_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = linear19_2_weight_to_fp16, x = input_241_cast_fp16)[name = tensor("linear_80_cast_fp16")]; + tensor input_243_cast_fp16 = add(x = input_237_cast_fp16, y = linear_80_cast_fp16)[name = tensor("input_243_cast_fp16")]; + tensor input_245_axes_0 = const()[name = tensor("input_245_axes_0"), val = tensor([-1])]; + tensor norm20_1_weight_to_fp16 = const()[name = tensor("norm20_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(504625088)))]; + tensor norm20_1_bias_to_fp16 = const()[name = tensor("norm20_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(504627200)))]; + tensor var_7924_to_fp16 = const()[name = tensor("op_7924_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_245_cast_fp16 = layer_norm(axes = input_245_axes_0, beta = norm20_1_bias_to_fp16, epsilon = var_7924_to_fp16, gamma = norm20_1_weight_to_fp16, x = input_243_cast_fp16)[name = tensor("input_245_cast_fp16")]; + tensor attn20_in_proj_weight_to_fp16 = const()[name = tensor("attn20_in_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(504629312)))]; + tensor linear_81_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = attn20_in_proj_weight_to_fp16, x = input_245_cast_fp16)[name = tensor("linear_81_cast_fp16")]; + tensor var_7938 = const()[name = tensor("op_7938"), val = tensor([1, 1, 3, 16, 64])]; + tensor qkv_41_cast_fp16 = reshape(shape = var_7938, x = linear_81_cast_fp16)[name = tensor("qkv_41_cast_fp16")]; + tensor q_121_begin_0 = const()[name = tensor("q_121_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor q_121_end_0 = const()[name = tensor("q_121_end_0"), val = tensor([1, 1, 1, 16, 64])]; + tensor q_121_end_mask_0 = const()[name = tensor("q_121_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor q_121_squeeze_mask_0 = const()[name = tensor("q_121_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor q_121_cast_fp16 = slice_by_index(begin = q_121_begin_0, end = q_121_end_0, end_mask = q_121_end_mask_0, squeeze_mask = q_121_squeeze_mask_0, x = qkv_41_cast_fp16)[name = tensor("q_121_cast_fp16")]; + tensor k_81_begin_0 = const()[name = tensor("k_81_begin_0"), val = tensor([0, 0, 1, 0, 0])]; + tensor k_81_end_0 = const()[name = tensor("k_81_end_0"), val = tensor([1, 1, 2, 16, 64])]; + tensor k_81_end_mask_0 = const()[name = tensor("k_81_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor k_81_squeeze_mask_0 = const()[name = tensor("k_81_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor k_81_cast_fp16 = slice_by_index(begin = k_81_begin_0, end = k_81_end_0, end_mask = k_81_end_mask_0, squeeze_mask = k_81_squeeze_mask_0, x = qkv_41_cast_fp16)[name = tensor("k_81_cast_fp16")]; + tensor v_41_begin_0 = const()[name = tensor("v_41_begin_0"), val = tensor([0, 0, 2, 0, 0])]; + tensor v_41_end_0 = const()[name = tensor("v_41_end_0"), val = tensor([1, 1, 3, 16, 64])]; + tensor v_41_end_mask_0 = const()[name = tensor("v_41_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor v_41_squeeze_mask_0 = const()[name = tensor("v_41_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor v_41_cast_fp16 = slice_by_index(begin = v_41_begin_0, end = v_41_end_0, end_mask = v_41_end_mask_0, squeeze_mask = v_41_squeeze_mask_0, x = qkv_41_cast_fp16)[name = tensor("v_41_cast_fp16")]; + tensor var_8004 = const()[name = tensor("op_8004"), val = tensor([1, 1, 1, 1])]; + tensor position20_to_fp16_dtype_0 = const()[name = tensor("position20_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor position20_to_fp16 = cast(dtype = position20_to_fp16_dtype_0, x = position20)[name = tensor("cast_453")]; + tensor ts_125_cast_fp16 = reshape(shape = var_8004, x = position20_to_fp16)[name = tensor("ts_125_cast_fp16")]; + tensor var_8011 = const()[name = tensor("op_8011"), val = tensor([1, 1, 16, 32, 2])]; + tensor q_complex_41_cast_fp16 = reshape(shape = var_8011, x = q_121_cast_fp16)[name = tensor("q_complex_41_cast_fp16")]; + tensor var_8018 = const()[name = tensor("op_8018"), val = tensor([1, 1, 16, 32, 2])]; + tensor k_complex_41_cast_fp16 = reshape(shape = var_8018, x = k_81_cast_fp16)[name = tensor("k_complex_41_cast_fp16")]; + tensor var_8022_begin_0 = const()[name = tensor("op_8022_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_8022_end_0 = const()[name = tensor("op_8022_end_0"), val = tensor([1, 1, 16, 32, 1])]; + tensor var_8022_end_mask_0 = const()[name = tensor("op_8022_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_8022_squeeze_mask_0 = const()[name = tensor("op_8022_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_8022_cast_fp16 = slice_by_index(begin = var_8022_begin_0, end = var_8022_end_0, end_mask = var_8022_end_mask_0, squeeze_mask = var_8022_squeeze_mask_0, x = q_complex_41_cast_fp16)[name = tensor("op_8022_cast_fp16")]; + tensor var_8030_begin_0 = const()[name = tensor("op_8030_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_8030_end_0 = const()[name = tensor("op_8030_end_0"), val = tensor([1, 1, 16, 32, 2])]; + tensor var_8030_end_mask_0 = const()[name = tensor("op_8030_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_8030_squeeze_mask_0 = const()[name = tensor("op_8030_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_8030_cast_fp16 = slice_by_index(begin = var_8030_begin_0, end = var_8030_end_0, end_mask = var_8030_end_mask_0, squeeze_mask = var_8030_squeeze_mask_0, x = q_complex_41_cast_fp16)[name = tensor("op_8030_cast_fp16")]; + tensor var_8038_begin_0 = const()[name = tensor("op_8038_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_8038_end_0 = const()[name = tensor("op_8038_end_0"), val = tensor([1, 1, 16, 32, 1])]; + tensor var_8038_end_mask_0 = const()[name = tensor("op_8038_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_8038_squeeze_mask_0 = const()[name = tensor("op_8038_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_8038_cast_fp16 = slice_by_index(begin = var_8038_begin_0, end = var_8038_end_0, end_mask = var_8038_end_mask_0, squeeze_mask = var_8038_squeeze_mask_0, x = k_complex_41_cast_fp16)[name = tensor("op_8038_cast_fp16")]; + tensor var_8046_begin_0 = const()[name = tensor("op_8046_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_8046_end_0 = const()[name = tensor("op_8046_end_0"), val = tensor([1, 1, 16, 32, 2])]; + tensor var_8046_end_mask_0 = const()[name = tensor("op_8046_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_8046_squeeze_mask_0 = const()[name = tensor("op_8046_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_8046_cast_fp16 = slice_by_index(begin = var_8046_begin_0, end = var_8046_end_0, end_mask = var_8046_end_mask_0, squeeze_mask = var_8046_squeeze_mask_0, x = k_complex_41_cast_fp16)[name = tensor("op_8046_cast_fp16")]; + tensor freqs_41_to_fp16 = const()[name = tensor("freqs_41_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(510920832)))]; + tensor var_8052_cast_fp16 = mul(x = freqs_41_to_fp16, y = ts_125_cast_fp16)[name = tensor("op_8052_cast_fp16")]; + tensor rotr_41_cast_fp16 = cos(x = var_8052_cast_fp16)[name = tensor("rotr_41_cast_fp16")]; + tensor roti_41_cast_fp16 = sin(x = var_8052_cast_fp16)[name = tensor("roti_41_cast_fp16")]; + tensor var_8056_cast_fp16 = mul(x = var_8022_cast_fp16, y = rotr_41_cast_fp16)[name = tensor("op_8056_cast_fp16")]; + tensor var_8057_cast_fp16 = mul(x = var_8030_cast_fp16, y = roti_41_cast_fp16)[name = tensor("op_8057_cast_fp16")]; + tensor qor_81_cast_fp16 = sub(x = var_8056_cast_fp16, y = var_8057_cast_fp16)[name = tensor("qor_81_cast_fp16")]; + tensor var_8060_cast_fp16 = mul(x = var_8022_cast_fp16, y = roti_41_cast_fp16)[name = tensor("op_8060_cast_fp16")]; + tensor var_8061_cast_fp16 = mul(x = var_8030_cast_fp16, y = rotr_41_cast_fp16)[name = tensor("op_8061_cast_fp16")]; + tensor qoi_81_cast_fp16 = add(x = var_8060_cast_fp16, y = var_8061_cast_fp16)[name = tensor("qoi_81_cast_fp16")]; + tensor var_8064_cast_fp16 = mul(x = var_8038_cast_fp16, y = rotr_41_cast_fp16)[name = tensor("op_8064_cast_fp16")]; + tensor var_8065_cast_fp16 = mul(x = var_8046_cast_fp16, y = roti_41_cast_fp16)[name = tensor("op_8065_cast_fp16")]; + tensor kor_81_cast_fp16 = sub(x = var_8064_cast_fp16, y = var_8065_cast_fp16)[name = tensor("kor_81_cast_fp16")]; + tensor var_8068_cast_fp16 = mul(x = var_8038_cast_fp16, y = roti_41_cast_fp16)[name = tensor("op_8068_cast_fp16")]; + tensor var_8069_cast_fp16 = mul(x = var_8046_cast_fp16, y = rotr_41_cast_fp16)[name = tensor("op_8069_cast_fp16")]; + tensor koi_81_cast_fp16 = add(x = var_8068_cast_fp16, y = var_8069_cast_fp16)[name = tensor("koi_81_cast_fp16")]; + tensor qo_41_axis_0 = const()[name = tensor("qo_41_axis_0"), val = tensor(-1)]; + tensor qo_41_cast_fp16 = stack(axis = qo_41_axis_0, values = (qor_81_cast_fp16, qoi_81_cast_fp16))[name = tensor("qo_41_cast_fp16")]; + tensor ko_41_axis_0 = const()[name = tensor("ko_41_axis_0"), val = tensor(-1)]; + tensor ko_41_cast_fp16 = stack(axis = ko_41_axis_0, values = (kor_81_cast_fp16, koi_81_cast_fp16))[name = tensor("ko_41_cast_fp16")]; + tensor var_8102 = const()[name = tensor("op_8102"), val = tensor([1, 1, 16, 64])]; + tensor q_123_cast_fp16 = reshape(shape = var_8102, x = qo_41_cast_fp16)[name = tensor("q_123_cast_fp16")]; + tensor var_8108 = const()[name = tensor("op_8108"), val = tensor([1, 1, 16, 64])]; + tensor k_83_cast_fp16 = reshape(shape = var_8108, x = ko_41_cast_fp16)[name = tensor("k_83_cast_fp16")]; + tensor _inversed_8133_y_0_to_fp16 = const()[name = tensor("_inversed_8133_y_0_to_fp16"), val = tensor(0x1p-9)]; + tensor _inversed_8133_cast_fp16 = mul(x = ts_125_cast_fp16, y = _inversed_8133_y_0_to_fp16)[name = tensor("_inversed_8133_cast_fp16")]; + tensor var_8134_cast_fp16 = floor(x = _inversed_8133_cast_fp16)[name = tensor("op_8134_cast_fp16")]; + tensor var_8135_to_fp16 = const()[name = tensor("op_8135_to_fp16"), val = tensor(0x1p+9)]; + tensor var_8136_cast_fp16 = mul(x = var_8134_cast_fp16, y = var_8135_to_fp16)[name = tensor("op_8136_cast_fp16")]; + tensor write_indices_float_83_cast_fp16 = sub(x = ts_125_cast_fp16, y = var_8136_cast_fp16)[name = tensor("write_indices_float_83_cast_fp16")]; + tensor var_8143_dtype_0 = const()[name = tensor("op_8143_dtype_0"), val = tensor("int32")]; + tensor write_indices_41_reps_0 = const()[name = tensor("write_indices_41_reps_0"), val = tensor([1, 1, 16, 64])]; + tensor write_indices_float_83_cast_fp16_to_int32 = cast(dtype = var_8143_dtype_0, x = write_indices_float_83_cast_fp16)[name = tensor("cast_452")]; + tensor write_indices_41 = tile(reps = write_indices_41_reps_0, x = write_indices_float_83_cast_fp16_to_int32)[name = tensor("write_indices_41")]; + tensor var_8153_begin_0 = const()[name = tensor("op_8153_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_8153_end_0 = const()[name = tensor("op_8153_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor var_8153_end_mask_0 = const()[name = tensor("op_8153_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_8153_squeeze_mask_0 = const()[name = tensor("op_8153_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor cache20_to_fp16_dtype_0 = const()[name = tensor("cache20_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor cache20_to_fp16 = cast(dtype = cache20_to_fp16_dtype_0, x = cache20)[name = tensor("cast_451")]; + tensor var_8153_cast_fp16 = slice_by_index(begin = var_8153_begin_0, end = var_8153_end_0, end_mask = var_8153_end_mask_0, squeeze_mask = var_8153_squeeze_mask_0, x = cache20_to_fp16)[name = tensor("op_8153_cast_fp16")]; + tensor var_8155_axis_0 = const()[name = tensor("op_8155_axis_0"), val = tensor(1)]; + tensor var_8155_mode_0 = const()[name = tensor("op_8155_mode_0"), val = tensor("update")]; + tensor var_8155_validate_indices_0 = const()[name = tensor("op_8155_validate_indices_0"), val = tensor(false)]; + tensor var_8155_cast_fp16 = scatter_along_axis(axis = var_8155_axis_0, data = var_8153_cast_fp16, indices = write_indices_41, mode = var_8155_mode_0, updates = k_83_cast_fp16, validate_indices = var_8155_validate_indices_0)[name = tensor("op_8155_cast_fp16")]; + tensor concat_142 = const()[name = tensor("concat_142"), val = tensor([0, 0, 0, 0, 0])]; + tensor concat_143 = const()[name = tensor("concat_143"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_41_internal_tensor_assign_1_stride_0 = const()[name = tensor("new_cache_41_internal_tensor_assign_1_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_41_internal_tensor_assign_1_begin_mask_0 = const()[name = tensor("new_cache_41_internal_tensor_assign_1_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_41_internal_tensor_assign_1_end_mask_0 = const()[name = tensor("new_cache_41_internal_tensor_assign_1_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_41_internal_tensor_assign_1_squeeze_mask_0 = const()[name = tensor("new_cache_41_internal_tensor_assign_1_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_88 = const()[name = tensor("shape_88"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_40 = const()[name = tensor("reduce_prod_40"), val = tensor(1048576)]; + tensor range_1d_40_start_0 = const()[name = tensor("range_1d_40_start_0"), val = tensor(0)]; + tensor range_1d_40_step_0 = const()[name = tensor("range_1d_40_step_0"), val = tensor(1)]; + tensor range_1d_40 = range_1d(end = reduce_prod_40, start = range_1d_40_start_0, step = range_1d_40_step_0)[name = tensor("range_1d_40")]; + tensor reshape_200 = reshape(shape = shape_88, x = range_1d_40)[name = tensor("reshape_200")]; + tensor slice_by_index_40 = slice_by_index(begin = concat_142, begin_mask = new_cache_41_internal_tensor_assign_1_begin_mask_0, end = concat_143, end_mask = new_cache_41_internal_tensor_assign_1_end_mask_0, squeeze_mask = new_cache_41_internal_tensor_assign_1_squeeze_mask_0, stride = new_cache_41_internal_tensor_assign_1_stride_0, x = reshape_200)[name = tensor("slice_by_index_40")]; + tensor reshape_201_shape_0 = const()[name = tensor("reshape_201_shape_0"), val = tensor([-1])]; + tensor reshape_201 = reshape(shape = reshape_201_shape_0, x = slice_by_index_40)[name = tensor("reshape_201")]; + tensor reshape_202_shape_0 = const()[name = tensor("reshape_202_shape_0"), val = tensor([-1])]; + tensor reshape_202_cast_fp16 = reshape(shape = reshape_202_shape_0, x = var_8155_cast_fp16)[name = tensor("reshape_202_cast_fp16")]; + tensor reshape_203_shape_0 = const()[name = tensor("reshape_203_shape_0"), val = tensor([-1])]; + tensor reshape_203_cast_fp16 = reshape(shape = reshape_203_shape_0, x = cache20_to_fp16)[name = tensor("reshape_203_cast_fp16")]; + tensor scatter_40_mode_0 = const()[name = tensor("scatter_40_mode_0"), val = tensor("update")]; + tensor scatter_40_axis_0 = const()[name = tensor("scatter_40_axis_0"), val = tensor(0)]; + tensor scatter_40_validate_indices_0 = const()[name = tensor("scatter_40_validate_indices_0"), val = tensor(false)]; + tensor scatter_40_cast_fp16 = scatter(axis = scatter_40_axis_0, data = reshape_203_cast_fp16, indices = reshape_201, mode = scatter_40_mode_0, updates = reshape_202_cast_fp16, validate_indices = scatter_40_validate_indices_0)[name = tensor("scatter_40_cast_fp16")]; + tensor reshape_204_cast_fp16 = reshape(shape = shape_88, x = scatter_40_cast_fp16)[name = tensor("reshape_204_cast_fp16")]; + tensor var_8163_begin_0 = const()[name = tensor("op_8163_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor var_8163_end_0 = const()[name = tensor("op_8163_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor var_8163_end_mask_0 = const()[name = tensor("op_8163_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_8163_squeeze_mask_0 = const()[name = tensor("op_8163_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor var_8163_cast_fp16 = slice_by_index(begin = var_8163_begin_0, end = var_8163_end_0, end_mask = var_8163_end_mask_0, squeeze_mask = var_8163_squeeze_mask_0, x = reshape_204_cast_fp16)[name = tensor("op_8163_cast_fp16")]; + tensor var_8165_axis_0 = const()[name = tensor("op_8165_axis_0"), val = tensor(1)]; + tensor var_8165_mode_0 = const()[name = tensor("op_8165_mode_0"), val = tensor("update")]; + tensor var_8165_validate_indices_0 = const()[name = tensor("op_8165_validate_indices_0"), val = tensor(false)]; + tensor var_8165_cast_fp16 = scatter_along_axis(axis = var_8165_axis_0, data = var_8163_cast_fp16, indices = write_indices_41, mode = var_8165_mode_0, updates = v_41_cast_fp16, validate_indices = var_8165_validate_indices_0)[name = tensor("op_8165_cast_fp16")]; + tensor concat_144 = const()[name = tensor("concat_144"), val = tensor([1, 0, 0, 0, 0])]; + tensor concat_145 = const()[name = tensor("concat_145"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_41_internal_tensor_assign_2_stride_0 = const()[name = tensor("new_cache_41_internal_tensor_assign_2_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_41_internal_tensor_assign_2_begin_mask_0 = const()[name = tensor("new_cache_41_internal_tensor_assign_2_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_41_internal_tensor_assign_2_end_mask_0 = const()[name = tensor("new_cache_41_internal_tensor_assign_2_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_41_internal_tensor_assign_2_squeeze_mask_0 = const()[name = tensor("new_cache_41_internal_tensor_assign_2_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_89 = const()[name = tensor("shape_89"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_41 = const()[name = tensor("reduce_prod_41"), val = tensor(1048576)]; + tensor range_1d_41_start_0 = const()[name = tensor("range_1d_41_start_0"), val = tensor(0)]; + tensor range_1d_41_step_0 = const()[name = tensor("range_1d_41_step_0"), val = tensor(1)]; + tensor range_1d_41 = range_1d(end = reduce_prod_41, start = range_1d_41_start_0, step = range_1d_41_step_0)[name = tensor("range_1d_41")]; + tensor reshape_205 = reshape(shape = shape_89, x = range_1d_41)[name = tensor("reshape_205")]; + tensor slice_by_index_41 = slice_by_index(begin = concat_144, begin_mask = new_cache_41_internal_tensor_assign_2_begin_mask_0, end = concat_145, end_mask = new_cache_41_internal_tensor_assign_2_end_mask_0, squeeze_mask = new_cache_41_internal_tensor_assign_2_squeeze_mask_0, stride = new_cache_41_internal_tensor_assign_2_stride_0, x = reshape_205)[name = tensor("slice_by_index_41")]; + tensor reshape_206_shape_0 = const()[name = tensor("reshape_206_shape_0"), val = tensor([-1])]; + tensor reshape_206 = reshape(shape = reshape_206_shape_0, x = slice_by_index_41)[name = tensor("reshape_206")]; + tensor reshape_207_shape_0 = const()[name = tensor("reshape_207_shape_0"), val = tensor([-1])]; + tensor reshape_207_cast_fp16 = reshape(shape = reshape_207_shape_0, x = var_8165_cast_fp16)[name = tensor("reshape_207_cast_fp16")]; + tensor reshape_208_shape_0 = const()[name = tensor("reshape_208_shape_0"), val = tensor([-1])]; + tensor reshape_208_cast_fp16 = reshape(shape = reshape_208_shape_0, x = reshape_204_cast_fp16)[name = tensor("reshape_208_cast_fp16")]; + tensor scatter_41_mode_0 = const()[name = tensor("scatter_41_mode_0"), val = tensor("update")]; + tensor scatter_41_axis_0 = const()[name = tensor("scatter_41_axis_0"), val = tensor(0)]; + tensor scatter_41_validate_indices_0 = const()[name = tensor("scatter_41_validate_indices_0"), val = tensor(false)]; + tensor scatter_41_cast_fp16 = scatter(axis = scatter_41_axis_0, data = reshape_208_cast_fp16, indices = reshape_206, mode = scatter_41_mode_0, updates = reshape_207_cast_fp16, validate_indices = scatter_41_validate_indices_0)[name = tensor("scatter_41_cast_fp16")]; + tensor reshape_209_cast_fp16 = reshape(shape = shape_89, x = scatter_41_cast_fp16)[name = tensor("reshape_209_cast_fp16")]; + tensor reshape_209_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("reshape_209_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor keys_121_begin_0 = const()[name = tensor("keys_121_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor keys_121_end_0 = const()[name = tensor("keys_121_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor keys_121_end_mask_0 = const()[name = tensor("keys_121_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor keys_121_squeeze_mask_0 = const()[name = tensor("keys_121_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor keys_121_cast_fp16 = slice_by_index(begin = keys_121_begin_0, end = keys_121_end_0, end_mask = keys_121_end_mask_0, squeeze_mask = keys_121_squeeze_mask_0, x = reshape_209_cast_fp16)[name = tensor("keys_121_cast_fp16")]; + tensor values_121_begin_0 = const()[name = tensor("values_121_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor values_121_end_0 = const()[name = tensor("values_121_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor values_121_end_mask_0 = const()[name = tensor("values_121_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor values_121_squeeze_mask_0 = const()[name = tensor("values_121_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor values_121_cast_fp16 = slice_by_index(begin = values_121_begin_0, end = values_121_end_0, end_mask = values_121_end_mask_0, squeeze_mask = values_121_squeeze_mask_0, x = reshape_209_cast_fp16)[name = tensor("values_121_cast_fp16")]; + tensor var_8177_cast_fp16 = not_equal(x = keys_121_cast_fp16, y = keys_121_cast_fp16)[name = tensor("op_8177_cast_fp16")]; + tensor keys_123_cast_fp16 = select(a = var_463_to_fp16, b = keys_121_cast_fp16, cond = var_8177_cast_fp16)[name = tensor("keys_123_cast_fp16")]; + tensor var_8185_cast_fp16 = not_equal(x = values_121_cast_fp16, y = values_121_cast_fp16)[name = tensor("op_8185_cast_fp16")]; + tensor values_123_cast_fp16 = select(a = var_463_to_fp16, b = values_121_cast_fp16, cond = var_8185_cast_fp16)[name = tensor("values_123_cast_fp16")]; + tensor var_8209 = const()[name = tensor("op_8209"), val = tensor([0, 2, 1, 3])]; + tensor var_8225 = const()[name = tensor("op_8225"), val = tensor([1, 1, 1])]; + tensor var_8226_cast_fp16 = reshape(shape = var_8225, x = position20_to_fp16)[name = tensor("op_8226_cast_fp16")]; + tensor var_8246_to_fp16 = const()[name = tensor("op_8246_to_fp16"), val = tensor(0x1p+0)]; + tensor valid_len_41_cast_fp16 = add(x = var_8226_cast_fp16, y = var_8246_to_fp16)[name = tensor("valid_len_41_cast_fp16")]; + tensor valid_mask_41_cast_fp16 = less(x = k_positions_1_promoted_to_fp16, y = valid_len_41_cast_fp16)[name = tensor("valid_mask_41_cast_fp16")]; + tensor causal_mask_41_cast_fp16 = less_equal(x = k_positions_1_promoted_to_fp16, y = var_8226_cast_fp16)[name = tensor("causal_mask_41_cast_fp16")]; + tensor attn_mask_81 = logical_and(x = valid_mask_41_cast_fp16, y = causal_mask_41_cast_fp16)[name = tensor("attn_mask_81")]; + tensor attn_mask_83_axes_0 = const()[name = tensor("attn_mask_83_axes_0"), val = tensor([1])]; + tensor attn_mask_83 = expand_dims(axes = attn_mask_83_axes_0, x = attn_mask_81)[name = tensor("attn_mask_83")]; + tensor var_8264_transpose_x_0 = const()[name = tensor("op_8264_transpose_x_0"), val = tensor(false)]; + tensor var_8264_transpose_y_0 = const()[name = tensor("op_8264_transpose_y_0"), val = tensor(false)]; + tensor transpose_112_perm_0 = const()[name = tensor("transpose_112_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_113_perm_0 = const()[name = tensor("transpose_113_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_113 = transpose(perm = transpose_113_perm_0, x = keys_123_cast_fp16)[name = tensor("transpose_133")]; + tensor transpose_112 = transpose(perm = transpose_112_perm_0, x = q_123_cast_fp16)[name = tensor("transpose_134")]; + tensor var_8264_cast_fp16 = matmul(transpose_x = var_8264_transpose_x_0, transpose_y = var_8264_transpose_y_0, x = transpose_112, y = transpose_113)[name = tensor("op_8264_cast_fp16")]; + tensor var_8258_to_fp16 = const()[name = tensor("op_8258_to_fp16"), val = tensor([0x1p-3])]; + tensor attn_weights_121_cast_fp16 = mul(x = var_8264_cast_fp16, y = var_8258_to_fp16)[name = tensor("attn_weights_121_cast_fp16")]; + tensor var_8266 = logical_not(x = attn_mask_83)[name = tensor("op_8266")]; + tensor var_8267_to_fp16 = const()[name = tensor("op_8267_to_fp16"), val = tensor(-inf)]; + tensor attn_weights_123_cast_fp16 = select(a = var_8267_to_fp16, b = attn_weights_121_cast_fp16, cond = var_8266)[name = tensor("attn_weights_123_cast_fp16")]; + tensor var_8269 = const()[name = tensor("op_8269"), val = tensor(-1)]; + tensor attn_weights_125_cast_fp16 = softmax(axis = var_8269, x = attn_weights_123_cast_fp16)[name = tensor("attn_weights_125_cast_fp16")]; + tensor attn_output_41_transpose_x_0 = const()[name = tensor("attn_output_41_transpose_x_0"), val = tensor(false)]; + tensor attn_output_41_transpose_y_0 = const()[name = tensor("attn_output_41_transpose_y_0"), val = tensor(false)]; + tensor values_125_cast_fp16 = transpose(perm = var_8209, x = values_123_cast_fp16)[name = tensor("transpose_135")]; + tensor attn_output_41_cast_fp16 = matmul(transpose_x = attn_output_41_transpose_x_0, transpose_y = attn_output_41_transpose_y_0, x = attn_weights_125_cast_fp16, y = values_125_cast_fp16)[name = tensor("attn_output_41_cast_fp16")]; + tensor var_8277 = const()[name = tensor("op_8277"), val = tensor([0, 2, 1, 3])]; + tensor var_8282 = const()[name = tensor("op_8282"), val = tensor([1, 1, 1024])]; + tensor var_8278_cast_fp16 = transpose(perm = var_8277, x = attn_output_41_cast_fp16)[name = tensor("transpose_132")]; + tensor input_247_cast_fp16 = reshape(shape = var_8282, x = var_8278_cast_fp16)[name = tensor("input_247_cast_fp16")]; + tensor attn20_out_proj_weight_to_fp16 = const()[name = tensor("attn20_out_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(510920960)))]; + tensor linear_82_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = attn20_out_proj_weight_to_fp16, x = input_247_cast_fp16)[name = tensor("linear_82_cast_fp16")]; + tensor var_8288_to_fp16 = const()[name = tensor("op_8288_to_fp16"), val = tensor(0x1p+0)]; + tensor var_8289_cast_fp16 = add(x = position20_to_fp16, y = var_8288_to_fp16)[name = tensor("op_8289_cast_fp16")]; + tensor var_8289_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_8289_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor input_249_cast_fp16 = add(x = input_243_cast_fp16, y = linear_82_cast_fp16)[name = tensor("input_249_cast_fp16")]; + tensor input_251_axes_0 = const()[name = tensor("input_251_axes_0"), val = tensor([-1])]; + tensor norm20_2_weight_to_fp16 = const()[name = tensor("norm20_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(513018176)))]; + tensor norm20_2_bias_to_fp16 = const()[name = tensor("norm20_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(513020288)))]; + tensor var_8293_to_fp16 = const()[name = tensor("op_8293_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_251_cast_fp16 = layer_norm(axes = input_251_axes_0, beta = norm20_2_bias_to_fp16, epsilon = var_8293_to_fp16, gamma = norm20_2_weight_to_fp16, x = input_249_cast_fp16)[name = tensor("input_251_cast_fp16")]; + tensor linear20_1_weight_to_fp16 = const()[name = tensor("linear20_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(513022400)))]; + tensor linear_83_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = linear20_1_weight_to_fp16, x = input_251_cast_fp16)[name = tensor("linear_83_cast_fp16")]; + tensor input_253_mode_0 = const()[name = tensor("input_253_mode_0"), val = tensor("EXACT")]; + tensor input_253_cast_fp16 = gelu(mode = input_253_mode_0, x = linear_83_cast_fp16)[name = tensor("input_253_cast_fp16")]; + tensor linear20_2_weight_to_fp16 = const()[name = tensor("linear20_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(521411072)))]; + tensor linear_84_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = linear20_2_weight_to_fp16, x = input_253_cast_fp16)[name = tensor("linear_84_cast_fp16")]; + tensor input_255_cast_fp16 = add(x = input_249_cast_fp16, y = linear_84_cast_fp16)[name = tensor("input_255_cast_fp16")]; + tensor input_257_axes_0 = const()[name = tensor("input_257_axes_0"), val = tensor([-1])]; + tensor norm21_1_weight_to_fp16 = const()[name = tensor("norm21_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(529799744)))]; + tensor norm21_1_bias_to_fp16 = const()[name = tensor("norm21_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(529801856)))]; + tensor var_8310_to_fp16 = const()[name = tensor("op_8310_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_257_cast_fp16 = layer_norm(axes = input_257_axes_0, beta = norm21_1_bias_to_fp16, epsilon = var_8310_to_fp16, gamma = norm21_1_weight_to_fp16, x = input_255_cast_fp16)[name = tensor("input_257_cast_fp16")]; + tensor attn21_in_proj_weight_to_fp16 = const()[name = tensor("attn21_in_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(529803968)))]; + tensor linear_85_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = attn21_in_proj_weight_to_fp16, x = input_257_cast_fp16)[name = tensor("linear_85_cast_fp16")]; + tensor var_8324 = const()[name = tensor("op_8324"), val = tensor([1, 1, 3, 16, 64])]; + tensor qkv_43_cast_fp16 = reshape(shape = var_8324, x = linear_85_cast_fp16)[name = tensor("qkv_43_cast_fp16")]; + tensor q_127_begin_0 = const()[name = tensor("q_127_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor q_127_end_0 = const()[name = tensor("q_127_end_0"), val = tensor([1, 1, 1, 16, 64])]; + tensor q_127_end_mask_0 = const()[name = tensor("q_127_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor q_127_squeeze_mask_0 = const()[name = tensor("q_127_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor q_127_cast_fp16 = slice_by_index(begin = q_127_begin_0, end = q_127_end_0, end_mask = q_127_end_mask_0, squeeze_mask = q_127_squeeze_mask_0, x = qkv_43_cast_fp16)[name = tensor("q_127_cast_fp16")]; + tensor k_85_begin_0 = const()[name = tensor("k_85_begin_0"), val = tensor([0, 0, 1, 0, 0])]; + tensor k_85_end_0 = const()[name = tensor("k_85_end_0"), val = tensor([1, 1, 2, 16, 64])]; + tensor k_85_end_mask_0 = const()[name = tensor("k_85_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor k_85_squeeze_mask_0 = const()[name = tensor("k_85_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor k_85_cast_fp16 = slice_by_index(begin = k_85_begin_0, end = k_85_end_0, end_mask = k_85_end_mask_0, squeeze_mask = k_85_squeeze_mask_0, x = qkv_43_cast_fp16)[name = tensor("k_85_cast_fp16")]; + tensor v_43_begin_0 = const()[name = tensor("v_43_begin_0"), val = tensor([0, 0, 2, 0, 0])]; + tensor v_43_end_0 = const()[name = tensor("v_43_end_0"), val = tensor([1, 1, 3, 16, 64])]; + tensor v_43_end_mask_0 = const()[name = tensor("v_43_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor v_43_squeeze_mask_0 = const()[name = tensor("v_43_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor v_43_cast_fp16 = slice_by_index(begin = v_43_begin_0, end = v_43_end_0, end_mask = v_43_end_mask_0, squeeze_mask = v_43_squeeze_mask_0, x = qkv_43_cast_fp16)[name = tensor("v_43_cast_fp16")]; + tensor var_8390 = const()[name = tensor("op_8390"), val = tensor([1, 1, 1, 1])]; + tensor position21_to_fp16_dtype_0 = const()[name = tensor("position21_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor position21_to_fp16 = cast(dtype = position21_to_fp16_dtype_0, x = position21)[name = tensor("cast_448")]; + tensor ts_131_cast_fp16 = reshape(shape = var_8390, x = position21_to_fp16)[name = tensor("ts_131_cast_fp16")]; + tensor var_8397 = const()[name = tensor("op_8397"), val = tensor([1, 1, 16, 32, 2])]; + tensor q_complex_43_cast_fp16 = reshape(shape = var_8397, x = q_127_cast_fp16)[name = tensor("q_complex_43_cast_fp16")]; + tensor var_8404 = const()[name = tensor("op_8404"), val = tensor([1, 1, 16, 32, 2])]; + tensor k_complex_43_cast_fp16 = reshape(shape = var_8404, x = k_85_cast_fp16)[name = tensor("k_complex_43_cast_fp16")]; + tensor var_8408_begin_0 = const()[name = tensor("op_8408_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_8408_end_0 = const()[name = tensor("op_8408_end_0"), val = tensor([1, 1, 16, 32, 1])]; + tensor var_8408_end_mask_0 = const()[name = tensor("op_8408_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_8408_squeeze_mask_0 = const()[name = tensor("op_8408_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_8408_cast_fp16 = slice_by_index(begin = var_8408_begin_0, end = var_8408_end_0, end_mask = var_8408_end_mask_0, squeeze_mask = var_8408_squeeze_mask_0, x = q_complex_43_cast_fp16)[name = tensor("op_8408_cast_fp16")]; + tensor var_8416_begin_0 = const()[name = tensor("op_8416_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_8416_end_0 = const()[name = tensor("op_8416_end_0"), val = tensor([1, 1, 16, 32, 2])]; + tensor var_8416_end_mask_0 = const()[name = tensor("op_8416_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_8416_squeeze_mask_0 = const()[name = tensor("op_8416_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_8416_cast_fp16 = slice_by_index(begin = var_8416_begin_0, end = var_8416_end_0, end_mask = var_8416_end_mask_0, squeeze_mask = var_8416_squeeze_mask_0, x = q_complex_43_cast_fp16)[name = tensor("op_8416_cast_fp16")]; + tensor var_8424_begin_0 = const()[name = tensor("op_8424_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_8424_end_0 = const()[name = tensor("op_8424_end_0"), val = tensor([1, 1, 16, 32, 1])]; + tensor var_8424_end_mask_0 = const()[name = tensor("op_8424_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_8424_squeeze_mask_0 = const()[name = tensor("op_8424_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_8424_cast_fp16 = slice_by_index(begin = var_8424_begin_0, end = var_8424_end_0, end_mask = var_8424_end_mask_0, squeeze_mask = var_8424_squeeze_mask_0, x = k_complex_43_cast_fp16)[name = tensor("op_8424_cast_fp16")]; + tensor var_8432_begin_0 = const()[name = tensor("op_8432_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_8432_end_0 = const()[name = tensor("op_8432_end_0"), val = tensor([1, 1, 16, 32, 2])]; + tensor var_8432_end_mask_0 = const()[name = tensor("op_8432_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_8432_squeeze_mask_0 = const()[name = tensor("op_8432_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_8432_cast_fp16 = slice_by_index(begin = var_8432_begin_0, end = var_8432_end_0, end_mask = var_8432_end_mask_0, squeeze_mask = var_8432_squeeze_mask_0, x = k_complex_43_cast_fp16)[name = tensor("op_8432_cast_fp16")]; + tensor freqs_43_to_fp16 = const()[name = tensor("freqs_43_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(536095488)))]; + tensor var_8438_cast_fp16 = mul(x = freqs_43_to_fp16, y = ts_131_cast_fp16)[name = tensor("op_8438_cast_fp16")]; + tensor rotr_43_cast_fp16 = cos(x = var_8438_cast_fp16)[name = tensor("rotr_43_cast_fp16")]; + tensor roti_43_cast_fp16 = sin(x = var_8438_cast_fp16)[name = tensor("roti_43_cast_fp16")]; + tensor var_8442_cast_fp16 = mul(x = var_8408_cast_fp16, y = rotr_43_cast_fp16)[name = tensor("op_8442_cast_fp16")]; + tensor var_8443_cast_fp16 = mul(x = var_8416_cast_fp16, y = roti_43_cast_fp16)[name = tensor("op_8443_cast_fp16")]; + tensor qor_85_cast_fp16 = sub(x = var_8442_cast_fp16, y = var_8443_cast_fp16)[name = tensor("qor_85_cast_fp16")]; + tensor var_8446_cast_fp16 = mul(x = var_8408_cast_fp16, y = roti_43_cast_fp16)[name = tensor("op_8446_cast_fp16")]; + tensor var_8447_cast_fp16 = mul(x = var_8416_cast_fp16, y = rotr_43_cast_fp16)[name = tensor("op_8447_cast_fp16")]; + tensor qoi_85_cast_fp16 = add(x = var_8446_cast_fp16, y = var_8447_cast_fp16)[name = tensor("qoi_85_cast_fp16")]; + tensor var_8450_cast_fp16 = mul(x = var_8424_cast_fp16, y = rotr_43_cast_fp16)[name = tensor("op_8450_cast_fp16")]; + tensor var_8451_cast_fp16 = mul(x = var_8432_cast_fp16, y = roti_43_cast_fp16)[name = tensor("op_8451_cast_fp16")]; + tensor kor_85_cast_fp16 = sub(x = var_8450_cast_fp16, y = var_8451_cast_fp16)[name = tensor("kor_85_cast_fp16")]; + tensor var_8454_cast_fp16 = mul(x = var_8424_cast_fp16, y = roti_43_cast_fp16)[name = tensor("op_8454_cast_fp16")]; + tensor var_8455_cast_fp16 = mul(x = var_8432_cast_fp16, y = rotr_43_cast_fp16)[name = tensor("op_8455_cast_fp16")]; + tensor koi_85_cast_fp16 = add(x = var_8454_cast_fp16, y = var_8455_cast_fp16)[name = tensor("koi_85_cast_fp16")]; + tensor qo_43_axis_0 = const()[name = tensor("qo_43_axis_0"), val = tensor(-1)]; + tensor qo_43_cast_fp16 = stack(axis = qo_43_axis_0, values = (qor_85_cast_fp16, qoi_85_cast_fp16))[name = tensor("qo_43_cast_fp16")]; + tensor ko_43_axis_0 = const()[name = tensor("ko_43_axis_0"), val = tensor(-1)]; + tensor ko_43_cast_fp16 = stack(axis = ko_43_axis_0, values = (kor_85_cast_fp16, koi_85_cast_fp16))[name = tensor("ko_43_cast_fp16")]; + tensor var_8488 = const()[name = tensor("op_8488"), val = tensor([1, 1, 16, 64])]; + tensor q_129_cast_fp16 = reshape(shape = var_8488, x = qo_43_cast_fp16)[name = tensor("q_129_cast_fp16")]; + tensor var_8494 = const()[name = tensor("op_8494"), val = tensor([1, 1, 16, 64])]; + tensor k_87_cast_fp16 = reshape(shape = var_8494, x = ko_43_cast_fp16)[name = tensor("k_87_cast_fp16")]; + tensor _inversed_8519_y_0_to_fp16 = const()[name = tensor("_inversed_8519_y_0_to_fp16"), val = tensor(0x1p-9)]; + tensor _inversed_8519_cast_fp16 = mul(x = ts_131_cast_fp16, y = _inversed_8519_y_0_to_fp16)[name = tensor("_inversed_8519_cast_fp16")]; + tensor var_8520_cast_fp16 = floor(x = _inversed_8519_cast_fp16)[name = tensor("op_8520_cast_fp16")]; + tensor var_8521_to_fp16 = const()[name = tensor("op_8521_to_fp16"), val = tensor(0x1p+9)]; + tensor var_8522_cast_fp16 = mul(x = var_8520_cast_fp16, y = var_8521_to_fp16)[name = tensor("op_8522_cast_fp16")]; + tensor write_indices_float_87_cast_fp16 = sub(x = ts_131_cast_fp16, y = var_8522_cast_fp16)[name = tensor("write_indices_float_87_cast_fp16")]; + tensor var_8529_dtype_0 = const()[name = tensor("op_8529_dtype_0"), val = tensor("int32")]; + tensor write_indices_43_reps_0 = const()[name = tensor("write_indices_43_reps_0"), val = tensor([1, 1, 16, 64])]; + tensor write_indices_float_87_cast_fp16_to_int32 = cast(dtype = var_8529_dtype_0, x = write_indices_float_87_cast_fp16)[name = tensor("cast_447")]; + tensor write_indices_43 = tile(reps = write_indices_43_reps_0, x = write_indices_float_87_cast_fp16_to_int32)[name = tensor("write_indices_43")]; + tensor var_8539_begin_0 = const()[name = tensor("op_8539_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_8539_end_0 = const()[name = tensor("op_8539_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor var_8539_end_mask_0 = const()[name = tensor("op_8539_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_8539_squeeze_mask_0 = const()[name = tensor("op_8539_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor cache21_to_fp16_dtype_0 = const()[name = tensor("cache21_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor cache21_to_fp16 = cast(dtype = cache21_to_fp16_dtype_0, x = cache21)[name = tensor("cast_446")]; + tensor var_8539_cast_fp16 = slice_by_index(begin = var_8539_begin_0, end = var_8539_end_0, end_mask = var_8539_end_mask_0, squeeze_mask = var_8539_squeeze_mask_0, x = cache21_to_fp16)[name = tensor("op_8539_cast_fp16")]; + tensor var_8541_axis_0 = const()[name = tensor("op_8541_axis_0"), val = tensor(1)]; + tensor var_8541_mode_0 = const()[name = tensor("op_8541_mode_0"), val = tensor("update")]; + tensor var_8541_validate_indices_0 = const()[name = tensor("op_8541_validate_indices_0"), val = tensor(false)]; + tensor var_8541_cast_fp16 = scatter_along_axis(axis = var_8541_axis_0, data = var_8539_cast_fp16, indices = write_indices_43, mode = var_8541_mode_0, updates = k_87_cast_fp16, validate_indices = var_8541_validate_indices_0)[name = tensor("op_8541_cast_fp16")]; + tensor concat_149 = const()[name = tensor("concat_149"), val = tensor([0, 0, 0, 0, 0])]; + tensor concat_150 = const()[name = tensor("concat_150"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_43_internal_tensor_assign_1_stride_0 = const()[name = tensor("new_cache_43_internal_tensor_assign_1_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_43_internal_tensor_assign_1_begin_mask_0 = const()[name = tensor("new_cache_43_internal_tensor_assign_1_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_43_internal_tensor_assign_1_end_mask_0 = const()[name = tensor("new_cache_43_internal_tensor_assign_1_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_43_internal_tensor_assign_1_squeeze_mask_0 = const()[name = tensor("new_cache_43_internal_tensor_assign_1_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_90 = const()[name = tensor("shape_90"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_42 = const()[name = tensor("reduce_prod_42"), val = tensor(1048576)]; + tensor range_1d_42_start_0 = const()[name = tensor("range_1d_42_start_0"), val = tensor(0)]; + tensor range_1d_42_step_0 = const()[name = tensor("range_1d_42_step_0"), val = tensor(1)]; + tensor range_1d_42 = range_1d(end = reduce_prod_42, start = range_1d_42_start_0, step = range_1d_42_step_0)[name = tensor("range_1d_42")]; + tensor reshape_210 = reshape(shape = shape_90, x = range_1d_42)[name = tensor("reshape_210")]; + tensor slice_by_index_42 = slice_by_index(begin = concat_149, begin_mask = new_cache_43_internal_tensor_assign_1_begin_mask_0, end = concat_150, end_mask = new_cache_43_internal_tensor_assign_1_end_mask_0, squeeze_mask = new_cache_43_internal_tensor_assign_1_squeeze_mask_0, stride = new_cache_43_internal_tensor_assign_1_stride_0, x = reshape_210)[name = tensor("slice_by_index_42")]; + tensor reshape_211_shape_0 = const()[name = tensor("reshape_211_shape_0"), val = tensor([-1])]; + tensor reshape_211 = reshape(shape = reshape_211_shape_0, x = slice_by_index_42)[name = tensor("reshape_211")]; + tensor reshape_212_shape_0 = const()[name = tensor("reshape_212_shape_0"), val = tensor([-1])]; + tensor reshape_212_cast_fp16 = reshape(shape = reshape_212_shape_0, x = var_8541_cast_fp16)[name = tensor("reshape_212_cast_fp16")]; + tensor reshape_213_shape_0 = const()[name = tensor("reshape_213_shape_0"), val = tensor([-1])]; + tensor reshape_213_cast_fp16 = reshape(shape = reshape_213_shape_0, x = cache21_to_fp16)[name = tensor("reshape_213_cast_fp16")]; + tensor scatter_42_mode_0 = const()[name = tensor("scatter_42_mode_0"), val = tensor("update")]; + tensor scatter_42_axis_0 = const()[name = tensor("scatter_42_axis_0"), val = tensor(0)]; + tensor scatter_42_validate_indices_0 = const()[name = tensor("scatter_42_validate_indices_0"), val = tensor(false)]; + tensor scatter_42_cast_fp16 = scatter(axis = scatter_42_axis_0, data = reshape_213_cast_fp16, indices = reshape_211, mode = scatter_42_mode_0, updates = reshape_212_cast_fp16, validate_indices = scatter_42_validate_indices_0)[name = tensor("scatter_42_cast_fp16")]; + tensor reshape_214_cast_fp16 = reshape(shape = shape_90, x = scatter_42_cast_fp16)[name = tensor("reshape_214_cast_fp16")]; + tensor var_8549_begin_0 = const()[name = tensor("op_8549_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor var_8549_end_0 = const()[name = tensor("op_8549_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor var_8549_end_mask_0 = const()[name = tensor("op_8549_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_8549_squeeze_mask_0 = const()[name = tensor("op_8549_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor var_8549_cast_fp16 = slice_by_index(begin = var_8549_begin_0, end = var_8549_end_0, end_mask = var_8549_end_mask_0, squeeze_mask = var_8549_squeeze_mask_0, x = reshape_214_cast_fp16)[name = tensor("op_8549_cast_fp16")]; + tensor var_8551_axis_0 = const()[name = tensor("op_8551_axis_0"), val = tensor(1)]; + tensor var_8551_mode_0 = const()[name = tensor("op_8551_mode_0"), val = tensor("update")]; + tensor var_8551_validate_indices_0 = const()[name = tensor("op_8551_validate_indices_0"), val = tensor(false)]; + tensor var_8551_cast_fp16 = scatter_along_axis(axis = var_8551_axis_0, data = var_8549_cast_fp16, indices = write_indices_43, mode = var_8551_mode_0, updates = v_43_cast_fp16, validate_indices = var_8551_validate_indices_0)[name = tensor("op_8551_cast_fp16")]; + tensor concat_151 = const()[name = tensor("concat_151"), val = tensor([1, 0, 0, 0, 0])]; + tensor concat_152 = const()[name = tensor("concat_152"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_43_internal_tensor_assign_2_stride_0 = const()[name = tensor("new_cache_43_internal_tensor_assign_2_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_43_internal_tensor_assign_2_begin_mask_0 = const()[name = tensor("new_cache_43_internal_tensor_assign_2_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_43_internal_tensor_assign_2_end_mask_0 = const()[name = tensor("new_cache_43_internal_tensor_assign_2_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_43_internal_tensor_assign_2_squeeze_mask_0 = const()[name = tensor("new_cache_43_internal_tensor_assign_2_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_91 = const()[name = tensor("shape_91"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_43 = const()[name = tensor("reduce_prod_43"), val = tensor(1048576)]; + tensor range_1d_43_start_0 = const()[name = tensor("range_1d_43_start_0"), val = tensor(0)]; + tensor range_1d_43_step_0 = const()[name = tensor("range_1d_43_step_0"), val = tensor(1)]; + tensor range_1d_43 = range_1d(end = reduce_prod_43, start = range_1d_43_start_0, step = range_1d_43_step_0)[name = tensor("range_1d_43")]; + tensor reshape_215 = reshape(shape = shape_91, x = range_1d_43)[name = tensor("reshape_215")]; + tensor slice_by_index_43 = slice_by_index(begin = concat_151, begin_mask = new_cache_43_internal_tensor_assign_2_begin_mask_0, end = concat_152, end_mask = new_cache_43_internal_tensor_assign_2_end_mask_0, squeeze_mask = new_cache_43_internal_tensor_assign_2_squeeze_mask_0, stride = new_cache_43_internal_tensor_assign_2_stride_0, x = reshape_215)[name = tensor("slice_by_index_43")]; + tensor reshape_216_shape_0 = const()[name = tensor("reshape_216_shape_0"), val = tensor([-1])]; + tensor reshape_216 = reshape(shape = reshape_216_shape_0, x = slice_by_index_43)[name = tensor("reshape_216")]; + tensor reshape_217_shape_0 = const()[name = tensor("reshape_217_shape_0"), val = tensor([-1])]; + tensor reshape_217_cast_fp16 = reshape(shape = reshape_217_shape_0, x = var_8551_cast_fp16)[name = tensor("reshape_217_cast_fp16")]; + tensor reshape_218_shape_0 = const()[name = tensor("reshape_218_shape_0"), val = tensor([-1])]; + tensor reshape_218_cast_fp16 = reshape(shape = reshape_218_shape_0, x = reshape_214_cast_fp16)[name = tensor("reshape_218_cast_fp16")]; + tensor scatter_43_mode_0 = const()[name = tensor("scatter_43_mode_0"), val = tensor("update")]; + tensor scatter_43_axis_0 = const()[name = tensor("scatter_43_axis_0"), val = tensor(0)]; + tensor scatter_43_validate_indices_0 = const()[name = tensor("scatter_43_validate_indices_0"), val = tensor(false)]; + tensor scatter_43_cast_fp16 = scatter(axis = scatter_43_axis_0, data = reshape_218_cast_fp16, indices = reshape_216, mode = scatter_43_mode_0, updates = reshape_217_cast_fp16, validate_indices = scatter_43_validate_indices_0)[name = tensor("scatter_43_cast_fp16")]; + tensor reshape_219_cast_fp16 = reshape(shape = shape_91, x = scatter_43_cast_fp16)[name = tensor("reshape_219_cast_fp16")]; + tensor reshape_219_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("reshape_219_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor keys_127_begin_0 = const()[name = tensor("keys_127_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor keys_127_end_0 = const()[name = tensor("keys_127_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor keys_127_end_mask_0 = const()[name = tensor("keys_127_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor keys_127_squeeze_mask_0 = const()[name = tensor("keys_127_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor keys_127_cast_fp16 = slice_by_index(begin = keys_127_begin_0, end = keys_127_end_0, end_mask = keys_127_end_mask_0, squeeze_mask = keys_127_squeeze_mask_0, x = reshape_219_cast_fp16)[name = tensor("keys_127_cast_fp16")]; + tensor values_127_begin_0 = const()[name = tensor("values_127_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor values_127_end_0 = const()[name = tensor("values_127_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor values_127_end_mask_0 = const()[name = tensor("values_127_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor values_127_squeeze_mask_0 = const()[name = tensor("values_127_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor values_127_cast_fp16 = slice_by_index(begin = values_127_begin_0, end = values_127_end_0, end_mask = values_127_end_mask_0, squeeze_mask = values_127_squeeze_mask_0, x = reshape_219_cast_fp16)[name = tensor("values_127_cast_fp16")]; + tensor var_8563_cast_fp16 = not_equal(x = keys_127_cast_fp16, y = keys_127_cast_fp16)[name = tensor("op_8563_cast_fp16")]; + tensor keys_129_cast_fp16 = select(a = var_463_to_fp16, b = keys_127_cast_fp16, cond = var_8563_cast_fp16)[name = tensor("keys_129_cast_fp16")]; + tensor var_8571_cast_fp16 = not_equal(x = values_127_cast_fp16, y = values_127_cast_fp16)[name = tensor("op_8571_cast_fp16")]; + tensor values_129_cast_fp16 = select(a = var_463_to_fp16, b = values_127_cast_fp16, cond = var_8571_cast_fp16)[name = tensor("values_129_cast_fp16")]; + tensor var_8595 = const()[name = tensor("op_8595"), val = tensor([0, 2, 1, 3])]; + tensor var_8611 = const()[name = tensor("op_8611"), val = tensor([1, 1, 1])]; + tensor var_8612_cast_fp16 = reshape(shape = var_8611, x = position21_to_fp16)[name = tensor("op_8612_cast_fp16")]; + tensor var_8632_to_fp16 = const()[name = tensor("op_8632_to_fp16"), val = tensor(0x1p+0)]; + tensor valid_len_43_cast_fp16 = add(x = var_8612_cast_fp16, y = var_8632_to_fp16)[name = tensor("valid_len_43_cast_fp16")]; + tensor valid_mask_43_cast_fp16 = less(x = k_positions_1_promoted_to_fp16, y = valid_len_43_cast_fp16)[name = tensor("valid_mask_43_cast_fp16")]; + tensor causal_mask_43_cast_fp16 = less_equal(x = k_positions_1_promoted_to_fp16, y = var_8612_cast_fp16)[name = tensor("causal_mask_43_cast_fp16")]; + tensor attn_mask_85 = logical_and(x = valid_mask_43_cast_fp16, y = causal_mask_43_cast_fp16)[name = tensor("attn_mask_85")]; + tensor attn_mask_87_axes_0 = const()[name = tensor("attn_mask_87_axes_0"), val = tensor([1])]; + tensor attn_mask_87 = expand_dims(axes = attn_mask_87_axes_0, x = attn_mask_85)[name = tensor("attn_mask_87")]; + tensor var_8650_transpose_x_0 = const()[name = tensor("op_8650_transpose_x_0"), val = tensor(false)]; + tensor var_8650_transpose_y_0 = const()[name = tensor("op_8650_transpose_y_0"), val = tensor(false)]; + tensor transpose_114_perm_0 = const()[name = tensor("transpose_114_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_115_perm_0 = const()[name = tensor("transpose_115_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_115 = transpose(perm = transpose_115_perm_0, x = keys_129_cast_fp16)[name = tensor("transpose_129")]; + tensor transpose_114 = transpose(perm = transpose_114_perm_0, x = q_129_cast_fp16)[name = tensor("transpose_130")]; + tensor var_8650_cast_fp16 = matmul(transpose_x = var_8650_transpose_x_0, transpose_y = var_8650_transpose_y_0, x = transpose_114, y = transpose_115)[name = tensor("op_8650_cast_fp16")]; + tensor var_8644_to_fp16 = const()[name = tensor("op_8644_to_fp16"), val = tensor([0x1p-3])]; + tensor attn_weights_127_cast_fp16 = mul(x = var_8650_cast_fp16, y = var_8644_to_fp16)[name = tensor("attn_weights_127_cast_fp16")]; + tensor var_8652 = logical_not(x = attn_mask_87)[name = tensor("op_8652")]; + tensor var_8653_to_fp16 = const()[name = tensor("op_8653_to_fp16"), val = tensor(-inf)]; + tensor attn_weights_129_cast_fp16 = select(a = var_8653_to_fp16, b = attn_weights_127_cast_fp16, cond = var_8652)[name = tensor("attn_weights_129_cast_fp16")]; + tensor var_8655 = const()[name = tensor("op_8655"), val = tensor(-1)]; + tensor attn_weights_131_cast_fp16 = softmax(axis = var_8655, x = attn_weights_129_cast_fp16)[name = tensor("attn_weights_131_cast_fp16")]; + tensor attn_output_43_transpose_x_0 = const()[name = tensor("attn_output_43_transpose_x_0"), val = tensor(false)]; + tensor attn_output_43_transpose_y_0 = const()[name = tensor("attn_output_43_transpose_y_0"), val = tensor(false)]; + tensor values_131_cast_fp16 = transpose(perm = var_8595, x = values_129_cast_fp16)[name = tensor("transpose_131")]; + tensor attn_output_43_cast_fp16 = matmul(transpose_x = attn_output_43_transpose_x_0, transpose_y = attn_output_43_transpose_y_0, x = attn_weights_131_cast_fp16, y = values_131_cast_fp16)[name = tensor("attn_output_43_cast_fp16")]; + tensor var_8663 = const()[name = tensor("op_8663"), val = tensor([0, 2, 1, 3])]; + tensor var_8668 = const()[name = tensor("op_8668"), val = tensor([1, 1, 1024])]; + tensor var_8664_cast_fp16 = transpose(perm = var_8663, x = attn_output_43_cast_fp16)[name = tensor("transpose_128")]; + tensor input_259_cast_fp16 = reshape(shape = var_8668, x = var_8664_cast_fp16)[name = tensor("input_259_cast_fp16")]; + tensor attn21_out_proj_weight_to_fp16 = const()[name = tensor("attn21_out_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(536095616)))]; + tensor linear_86_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = attn21_out_proj_weight_to_fp16, x = input_259_cast_fp16)[name = tensor("linear_86_cast_fp16")]; + tensor var_8674_to_fp16 = const()[name = tensor("op_8674_to_fp16"), val = tensor(0x1p+0)]; + tensor var_8675_cast_fp16 = add(x = position21_to_fp16, y = var_8674_to_fp16)[name = tensor("op_8675_cast_fp16")]; + tensor var_8675_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_8675_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor input_261_cast_fp16 = add(x = input_255_cast_fp16, y = linear_86_cast_fp16)[name = tensor("input_261_cast_fp16")]; + tensor input_263_axes_0 = const()[name = tensor("input_263_axes_0"), val = tensor([-1])]; + tensor norm21_2_weight_to_fp16 = const()[name = tensor("norm21_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(538192832)))]; + tensor norm21_2_bias_to_fp16 = const()[name = tensor("norm21_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(538194944)))]; + tensor var_8679_to_fp16 = const()[name = tensor("op_8679_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_263_cast_fp16 = layer_norm(axes = input_263_axes_0, beta = norm21_2_bias_to_fp16, epsilon = var_8679_to_fp16, gamma = norm21_2_weight_to_fp16, x = input_261_cast_fp16)[name = tensor("input_263_cast_fp16")]; + tensor linear21_1_weight_to_fp16 = const()[name = tensor("linear21_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(538197056)))]; + tensor linear_87_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = linear21_1_weight_to_fp16, x = input_263_cast_fp16)[name = tensor("linear_87_cast_fp16")]; + tensor input_265_mode_0 = const()[name = tensor("input_265_mode_0"), val = tensor("EXACT")]; + tensor input_265_cast_fp16 = gelu(mode = input_265_mode_0, x = linear_87_cast_fp16)[name = tensor("input_265_cast_fp16")]; + tensor linear21_2_weight_to_fp16 = const()[name = tensor("linear21_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(546585728)))]; + tensor linear_88_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = linear21_2_weight_to_fp16, x = input_265_cast_fp16)[name = tensor("linear_88_cast_fp16")]; + tensor input_267_cast_fp16 = add(x = input_261_cast_fp16, y = linear_88_cast_fp16)[name = tensor("input_267_cast_fp16")]; + tensor input_269_axes_0 = const()[name = tensor("input_269_axes_0"), val = tensor([-1])]; + tensor norm22_1_weight_to_fp16 = const()[name = tensor("norm22_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(554974400)))]; + tensor norm22_1_bias_to_fp16 = const()[name = tensor("norm22_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(554976512)))]; + tensor var_8696_to_fp16 = const()[name = tensor("op_8696_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_269_cast_fp16 = layer_norm(axes = input_269_axes_0, beta = norm22_1_bias_to_fp16, epsilon = var_8696_to_fp16, gamma = norm22_1_weight_to_fp16, x = input_267_cast_fp16)[name = tensor("input_269_cast_fp16")]; + tensor attn22_in_proj_weight_to_fp16 = const()[name = tensor("attn22_in_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(554978624)))]; + tensor linear_89_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = attn22_in_proj_weight_to_fp16, x = input_269_cast_fp16)[name = tensor("linear_89_cast_fp16")]; + tensor var_8710 = const()[name = tensor("op_8710"), val = tensor([1, 1, 3, 16, 64])]; + tensor qkv_45_cast_fp16 = reshape(shape = var_8710, x = linear_89_cast_fp16)[name = tensor("qkv_45_cast_fp16")]; + tensor q_133_begin_0 = const()[name = tensor("q_133_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor q_133_end_0 = const()[name = tensor("q_133_end_0"), val = tensor([1, 1, 1, 16, 64])]; + tensor q_133_end_mask_0 = const()[name = tensor("q_133_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor q_133_squeeze_mask_0 = const()[name = tensor("q_133_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor q_133_cast_fp16 = slice_by_index(begin = q_133_begin_0, end = q_133_end_0, end_mask = q_133_end_mask_0, squeeze_mask = q_133_squeeze_mask_0, x = qkv_45_cast_fp16)[name = tensor("q_133_cast_fp16")]; + tensor k_89_begin_0 = const()[name = tensor("k_89_begin_0"), val = tensor([0, 0, 1, 0, 0])]; + tensor k_89_end_0 = const()[name = tensor("k_89_end_0"), val = tensor([1, 1, 2, 16, 64])]; + tensor k_89_end_mask_0 = const()[name = tensor("k_89_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor k_89_squeeze_mask_0 = const()[name = tensor("k_89_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor k_89_cast_fp16 = slice_by_index(begin = k_89_begin_0, end = k_89_end_0, end_mask = k_89_end_mask_0, squeeze_mask = k_89_squeeze_mask_0, x = qkv_45_cast_fp16)[name = tensor("k_89_cast_fp16")]; + tensor v_45_begin_0 = const()[name = tensor("v_45_begin_0"), val = tensor([0, 0, 2, 0, 0])]; + tensor v_45_end_0 = const()[name = tensor("v_45_end_0"), val = tensor([1, 1, 3, 16, 64])]; + tensor v_45_end_mask_0 = const()[name = tensor("v_45_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor v_45_squeeze_mask_0 = const()[name = tensor("v_45_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor v_45_cast_fp16 = slice_by_index(begin = v_45_begin_0, end = v_45_end_0, end_mask = v_45_end_mask_0, squeeze_mask = v_45_squeeze_mask_0, x = qkv_45_cast_fp16)[name = tensor("v_45_cast_fp16")]; + tensor var_8776 = const()[name = tensor("op_8776"), val = tensor([1, 1, 1, 1])]; + tensor position22_to_fp16_dtype_0 = const()[name = tensor("position22_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor position22_to_fp16 = cast(dtype = position22_to_fp16_dtype_0, x = position22)[name = tensor("cast_443")]; + tensor ts_137_cast_fp16 = reshape(shape = var_8776, x = position22_to_fp16)[name = tensor("ts_137_cast_fp16")]; + tensor var_8783 = const()[name = tensor("op_8783"), val = tensor([1, 1, 16, 32, 2])]; + tensor q_complex_45_cast_fp16 = reshape(shape = var_8783, x = q_133_cast_fp16)[name = tensor("q_complex_45_cast_fp16")]; + tensor var_8790 = const()[name = tensor("op_8790"), val = tensor([1, 1, 16, 32, 2])]; + tensor k_complex_45_cast_fp16 = reshape(shape = var_8790, x = k_89_cast_fp16)[name = tensor("k_complex_45_cast_fp16")]; + tensor var_8794_begin_0 = const()[name = tensor("op_8794_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_8794_end_0 = const()[name = tensor("op_8794_end_0"), val = tensor([1, 1, 16, 32, 1])]; + tensor var_8794_end_mask_0 = const()[name = tensor("op_8794_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_8794_squeeze_mask_0 = const()[name = tensor("op_8794_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_8794_cast_fp16 = slice_by_index(begin = var_8794_begin_0, end = var_8794_end_0, end_mask = var_8794_end_mask_0, squeeze_mask = var_8794_squeeze_mask_0, x = q_complex_45_cast_fp16)[name = tensor("op_8794_cast_fp16")]; + tensor var_8802_begin_0 = const()[name = tensor("op_8802_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_8802_end_0 = const()[name = tensor("op_8802_end_0"), val = tensor([1, 1, 16, 32, 2])]; + tensor var_8802_end_mask_0 = const()[name = tensor("op_8802_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_8802_squeeze_mask_0 = const()[name = tensor("op_8802_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_8802_cast_fp16 = slice_by_index(begin = var_8802_begin_0, end = var_8802_end_0, end_mask = var_8802_end_mask_0, squeeze_mask = var_8802_squeeze_mask_0, x = q_complex_45_cast_fp16)[name = tensor("op_8802_cast_fp16")]; + tensor var_8810_begin_0 = const()[name = tensor("op_8810_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_8810_end_0 = const()[name = tensor("op_8810_end_0"), val = tensor([1, 1, 16, 32, 1])]; + tensor var_8810_end_mask_0 = const()[name = tensor("op_8810_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_8810_squeeze_mask_0 = const()[name = tensor("op_8810_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_8810_cast_fp16 = slice_by_index(begin = var_8810_begin_0, end = var_8810_end_0, end_mask = var_8810_end_mask_0, squeeze_mask = var_8810_squeeze_mask_0, x = k_complex_45_cast_fp16)[name = tensor("op_8810_cast_fp16")]; + tensor var_8818_begin_0 = const()[name = tensor("op_8818_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_8818_end_0 = const()[name = tensor("op_8818_end_0"), val = tensor([1, 1, 16, 32, 2])]; + tensor var_8818_end_mask_0 = const()[name = tensor("op_8818_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_8818_squeeze_mask_0 = const()[name = tensor("op_8818_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_8818_cast_fp16 = slice_by_index(begin = var_8818_begin_0, end = var_8818_end_0, end_mask = var_8818_end_mask_0, squeeze_mask = var_8818_squeeze_mask_0, x = k_complex_45_cast_fp16)[name = tensor("op_8818_cast_fp16")]; + tensor freqs_45_to_fp16 = const()[name = tensor("freqs_45_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(561270144)))]; + tensor var_8824_cast_fp16 = mul(x = freqs_45_to_fp16, y = ts_137_cast_fp16)[name = tensor("op_8824_cast_fp16")]; + tensor rotr_45_cast_fp16 = cos(x = var_8824_cast_fp16)[name = tensor("rotr_45_cast_fp16")]; + tensor roti_45_cast_fp16 = sin(x = var_8824_cast_fp16)[name = tensor("roti_45_cast_fp16")]; + tensor var_8828_cast_fp16 = mul(x = var_8794_cast_fp16, y = rotr_45_cast_fp16)[name = tensor("op_8828_cast_fp16")]; + tensor var_8829_cast_fp16 = mul(x = var_8802_cast_fp16, y = roti_45_cast_fp16)[name = tensor("op_8829_cast_fp16")]; + tensor qor_89_cast_fp16 = sub(x = var_8828_cast_fp16, y = var_8829_cast_fp16)[name = tensor("qor_89_cast_fp16")]; + tensor var_8832_cast_fp16 = mul(x = var_8794_cast_fp16, y = roti_45_cast_fp16)[name = tensor("op_8832_cast_fp16")]; + tensor var_8833_cast_fp16 = mul(x = var_8802_cast_fp16, y = rotr_45_cast_fp16)[name = tensor("op_8833_cast_fp16")]; + tensor qoi_89_cast_fp16 = add(x = var_8832_cast_fp16, y = var_8833_cast_fp16)[name = tensor("qoi_89_cast_fp16")]; + tensor var_8836_cast_fp16 = mul(x = var_8810_cast_fp16, y = rotr_45_cast_fp16)[name = tensor("op_8836_cast_fp16")]; + tensor var_8837_cast_fp16 = mul(x = var_8818_cast_fp16, y = roti_45_cast_fp16)[name = tensor("op_8837_cast_fp16")]; + tensor kor_89_cast_fp16 = sub(x = var_8836_cast_fp16, y = var_8837_cast_fp16)[name = tensor("kor_89_cast_fp16")]; + tensor var_8840_cast_fp16 = mul(x = var_8810_cast_fp16, y = roti_45_cast_fp16)[name = tensor("op_8840_cast_fp16")]; + tensor var_8841_cast_fp16 = mul(x = var_8818_cast_fp16, y = rotr_45_cast_fp16)[name = tensor("op_8841_cast_fp16")]; + tensor koi_89_cast_fp16 = add(x = var_8840_cast_fp16, y = var_8841_cast_fp16)[name = tensor("koi_89_cast_fp16")]; + tensor qo_45_axis_0 = const()[name = tensor("qo_45_axis_0"), val = tensor(-1)]; + tensor qo_45_cast_fp16 = stack(axis = qo_45_axis_0, values = (qor_89_cast_fp16, qoi_89_cast_fp16))[name = tensor("qo_45_cast_fp16")]; + tensor ko_45_axis_0 = const()[name = tensor("ko_45_axis_0"), val = tensor(-1)]; + tensor ko_45_cast_fp16 = stack(axis = ko_45_axis_0, values = (kor_89_cast_fp16, koi_89_cast_fp16))[name = tensor("ko_45_cast_fp16")]; + tensor var_8874 = const()[name = tensor("op_8874"), val = tensor([1, 1, 16, 64])]; + tensor q_135_cast_fp16 = reshape(shape = var_8874, x = qo_45_cast_fp16)[name = tensor("q_135_cast_fp16")]; + tensor var_8880 = const()[name = tensor("op_8880"), val = tensor([1, 1, 16, 64])]; + tensor k_91_cast_fp16 = reshape(shape = var_8880, x = ko_45_cast_fp16)[name = tensor("k_91_cast_fp16")]; + tensor _inversed_8905_y_0_to_fp16 = const()[name = tensor("_inversed_8905_y_0_to_fp16"), val = tensor(0x1p-9)]; + tensor _inversed_8905_cast_fp16 = mul(x = ts_137_cast_fp16, y = _inversed_8905_y_0_to_fp16)[name = tensor("_inversed_8905_cast_fp16")]; + tensor var_8906_cast_fp16 = floor(x = _inversed_8905_cast_fp16)[name = tensor("op_8906_cast_fp16")]; + tensor var_8907_to_fp16 = const()[name = tensor("op_8907_to_fp16"), val = tensor(0x1p+9)]; + tensor var_8908_cast_fp16 = mul(x = var_8906_cast_fp16, y = var_8907_to_fp16)[name = tensor("op_8908_cast_fp16")]; + tensor write_indices_float_91_cast_fp16 = sub(x = ts_137_cast_fp16, y = var_8908_cast_fp16)[name = tensor("write_indices_float_91_cast_fp16")]; + tensor var_8915_dtype_0 = const()[name = tensor("op_8915_dtype_0"), val = tensor("int32")]; + tensor write_indices_45_reps_0 = const()[name = tensor("write_indices_45_reps_0"), val = tensor([1, 1, 16, 64])]; + tensor write_indices_float_91_cast_fp16_to_int32 = cast(dtype = var_8915_dtype_0, x = write_indices_float_91_cast_fp16)[name = tensor("cast_442")]; + tensor write_indices_45 = tile(reps = write_indices_45_reps_0, x = write_indices_float_91_cast_fp16_to_int32)[name = tensor("write_indices_45")]; + tensor var_8925_begin_0 = const()[name = tensor("op_8925_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_8925_end_0 = const()[name = tensor("op_8925_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor var_8925_end_mask_0 = const()[name = tensor("op_8925_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_8925_squeeze_mask_0 = const()[name = tensor("op_8925_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor cache22_to_fp16_dtype_0 = const()[name = tensor("cache22_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor cache22_to_fp16 = cast(dtype = cache22_to_fp16_dtype_0, x = cache22)[name = tensor("cast_441")]; + tensor var_8925_cast_fp16 = slice_by_index(begin = var_8925_begin_0, end = var_8925_end_0, end_mask = var_8925_end_mask_0, squeeze_mask = var_8925_squeeze_mask_0, x = cache22_to_fp16)[name = tensor("op_8925_cast_fp16")]; + tensor var_8927_axis_0 = const()[name = tensor("op_8927_axis_0"), val = tensor(1)]; + tensor var_8927_mode_0 = const()[name = tensor("op_8927_mode_0"), val = tensor("update")]; + tensor var_8927_validate_indices_0 = const()[name = tensor("op_8927_validate_indices_0"), val = tensor(false)]; + tensor var_8927_cast_fp16 = scatter_along_axis(axis = var_8927_axis_0, data = var_8925_cast_fp16, indices = write_indices_45, mode = var_8927_mode_0, updates = k_91_cast_fp16, validate_indices = var_8927_validate_indices_0)[name = tensor("op_8927_cast_fp16")]; + tensor concat_156 = const()[name = tensor("concat_156"), val = tensor([0, 0, 0, 0, 0])]; + tensor concat_157 = const()[name = tensor("concat_157"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_45_internal_tensor_assign_1_stride_0 = const()[name = tensor("new_cache_45_internal_tensor_assign_1_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_45_internal_tensor_assign_1_begin_mask_0 = const()[name = tensor("new_cache_45_internal_tensor_assign_1_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_45_internal_tensor_assign_1_end_mask_0 = const()[name = tensor("new_cache_45_internal_tensor_assign_1_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_45_internal_tensor_assign_1_squeeze_mask_0 = const()[name = tensor("new_cache_45_internal_tensor_assign_1_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_92 = const()[name = tensor("shape_92"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_44 = const()[name = tensor("reduce_prod_44"), val = tensor(1048576)]; + tensor range_1d_44_start_0 = const()[name = tensor("range_1d_44_start_0"), val = tensor(0)]; + tensor range_1d_44_step_0 = const()[name = tensor("range_1d_44_step_0"), val = tensor(1)]; + tensor range_1d_44 = range_1d(end = reduce_prod_44, start = range_1d_44_start_0, step = range_1d_44_step_0)[name = tensor("range_1d_44")]; + tensor reshape_220 = reshape(shape = shape_92, x = range_1d_44)[name = tensor("reshape_220")]; + tensor slice_by_index_44 = slice_by_index(begin = concat_156, begin_mask = new_cache_45_internal_tensor_assign_1_begin_mask_0, end = concat_157, end_mask = new_cache_45_internal_tensor_assign_1_end_mask_0, squeeze_mask = new_cache_45_internal_tensor_assign_1_squeeze_mask_0, stride = new_cache_45_internal_tensor_assign_1_stride_0, x = reshape_220)[name = tensor("slice_by_index_44")]; + tensor reshape_221_shape_0 = const()[name = tensor("reshape_221_shape_0"), val = tensor([-1])]; + tensor reshape_221 = reshape(shape = reshape_221_shape_0, x = slice_by_index_44)[name = tensor("reshape_221")]; + tensor reshape_222_shape_0 = const()[name = tensor("reshape_222_shape_0"), val = tensor([-1])]; + tensor reshape_222_cast_fp16 = reshape(shape = reshape_222_shape_0, x = var_8927_cast_fp16)[name = tensor("reshape_222_cast_fp16")]; + tensor reshape_223_shape_0 = const()[name = tensor("reshape_223_shape_0"), val = tensor([-1])]; + tensor reshape_223_cast_fp16 = reshape(shape = reshape_223_shape_0, x = cache22_to_fp16)[name = tensor("reshape_223_cast_fp16")]; + tensor scatter_44_mode_0 = const()[name = tensor("scatter_44_mode_0"), val = tensor("update")]; + tensor scatter_44_axis_0 = const()[name = tensor("scatter_44_axis_0"), val = tensor(0)]; + tensor scatter_44_validate_indices_0 = const()[name = tensor("scatter_44_validate_indices_0"), val = tensor(false)]; + tensor scatter_44_cast_fp16 = scatter(axis = scatter_44_axis_0, data = reshape_223_cast_fp16, indices = reshape_221, mode = scatter_44_mode_0, updates = reshape_222_cast_fp16, validate_indices = scatter_44_validate_indices_0)[name = tensor("scatter_44_cast_fp16")]; + tensor reshape_224_cast_fp16 = reshape(shape = shape_92, x = scatter_44_cast_fp16)[name = tensor("reshape_224_cast_fp16")]; + tensor var_8935_begin_0 = const()[name = tensor("op_8935_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor var_8935_end_0 = const()[name = tensor("op_8935_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor var_8935_end_mask_0 = const()[name = tensor("op_8935_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_8935_squeeze_mask_0 = const()[name = tensor("op_8935_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor var_8935_cast_fp16 = slice_by_index(begin = var_8935_begin_0, end = var_8935_end_0, end_mask = var_8935_end_mask_0, squeeze_mask = var_8935_squeeze_mask_0, x = reshape_224_cast_fp16)[name = tensor("op_8935_cast_fp16")]; + tensor var_8937_axis_0 = const()[name = tensor("op_8937_axis_0"), val = tensor(1)]; + tensor var_8937_mode_0 = const()[name = tensor("op_8937_mode_0"), val = tensor("update")]; + tensor var_8937_validate_indices_0 = const()[name = tensor("op_8937_validate_indices_0"), val = tensor(false)]; + tensor var_8937_cast_fp16 = scatter_along_axis(axis = var_8937_axis_0, data = var_8935_cast_fp16, indices = write_indices_45, mode = var_8937_mode_0, updates = v_45_cast_fp16, validate_indices = var_8937_validate_indices_0)[name = tensor("op_8937_cast_fp16")]; + tensor concat_158 = const()[name = tensor("concat_158"), val = tensor([1, 0, 0, 0, 0])]; + tensor concat_159 = const()[name = tensor("concat_159"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_45_internal_tensor_assign_2_stride_0 = const()[name = tensor("new_cache_45_internal_tensor_assign_2_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_45_internal_tensor_assign_2_begin_mask_0 = const()[name = tensor("new_cache_45_internal_tensor_assign_2_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_45_internal_tensor_assign_2_end_mask_0 = const()[name = tensor("new_cache_45_internal_tensor_assign_2_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_45_internal_tensor_assign_2_squeeze_mask_0 = const()[name = tensor("new_cache_45_internal_tensor_assign_2_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_93 = const()[name = tensor("shape_93"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_45 = const()[name = tensor("reduce_prod_45"), val = tensor(1048576)]; + tensor range_1d_45_start_0 = const()[name = tensor("range_1d_45_start_0"), val = tensor(0)]; + tensor range_1d_45_step_0 = const()[name = tensor("range_1d_45_step_0"), val = tensor(1)]; + tensor range_1d_45 = range_1d(end = reduce_prod_45, start = range_1d_45_start_0, step = range_1d_45_step_0)[name = tensor("range_1d_45")]; + tensor reshape_225 = reshape(shape = shape_93, x = range_1d_45)[name = tensor("reshape_225")]; + tensor slice_by_index_45 = slice_by_index(begin = concat_158, begin_mask = new_cache_45_internal_tensor_assign_2_begin_mask_0, end = concat_159, end_mask = new_cache_45_internal_tensor_assign_2_end_mask_0, squeeze_mask = new_cache_45_internal_tensor_assign_2_squeeze_mask_0, stride = new_cache_45_internal_tensor_assign_2_stride_0, x = reshape_225)[name = tensor("slice_by_index_45")]; + tensor reshape_226_shape_0 = const()[name = tensor("reshape_226_shape_0"), val = tensor([-1])]; + tensor reshape_226 = reshape(shape = reshape_226_shape_0, x = slice_by_index_45)[name = tensor("reshape_226")]; + tensor reshape_227_shape_0 = const()[name = tensor("reshape_227_shape_0"), val = tensor([-1])]; + tensor reshape_227_cast_fp16 = reshape(shape = reshape_227_shape_0, x = var_8937_cast_fp16)[name = tensor("reshape_227_cast_fp16")]; + tensor reshape_228_shape_0 = const()[name = tensor("reshape_228_shape_0"), val = tensor([-1])]; + tensor reshape_228_cast_fp16 = reshape(shape = reshape_228_shape_0, x = reshape_224_cast_fp16)[name = tensor("reshape_228_cast_fp16")]; + tensor scatter_45_mode_0 = const()[name = tensor("scatter_45_mode_0"), val = tensor("update")]; + tensor scatter_45_axis_0 = const()[name = tensor("scatter_45_axis_0"), val = tensor(0)]; + tensor scatter_45_validate_indices_0 = const()[name = tensor("scatter_45_validate_indices_0"), val = tensor(false)]; + tensor scatter_45_cast_fp16 = scatter(axis = scatter_45_axis_0, data = reshape_228_cast_fp16, indices = reshape_226, mode = scatter_45_mode_0, updates = reshape_227_cast_fp16, validate_indices = scatter_45_validate_indices_0)[name = tensor("scatter_45_cast_fp16")]; + tensor reshape_229_cast_fp16 = reshape(shape = shape_93, x = scatter_45_cast_fp16)[name = tensor("reshape_229_cast_fp16")]; + tensor reshape_229_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("reshape_229_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor keys_133_begin_0 = const()[name = tensor("keys_133_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor keys_133_end_0 = const()[name = tensor("keys_133_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor keys_133_end_mask_0 = const()[name = tensor("keys_133_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor keys_133_squeeze_mask_0 = const()[name = tensor("keys_133_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor keys_133_cast_fp16 = slice_by_index(begin = keys_133_begin_0, end = keys_133_end_0, end_mask = keys_133_end_mask_0, squeeze_mask = keys_133_squeeze_mask_0, x = reshape_229_cast_fp16)[name = tensor("keys_133_cast_fp16")]; + tensor values_133_begin_0 = const()[name = tensor("values_133_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor values_133_end_0 = const()[name = tensor("values_133_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor values_133_end_mask_0 = const()[name = tensor("values_133_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor values_133_squeeze_mask_0 = const()[name = tensor("values_133_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor values_133_cast_fp16 = slice_by_index(begin = values_133_begin_0, end = values_133_end_0, end_mask = values_133_end_mask_0, squeeze_mask = values_133_squeeze_mask_0, x = reshape_229_cast_fp16)[name = tensor("values_133_cast_fp16")]; + tensor var_8949_cast_fp16 = not_equal(x = keys_133_cast_fp16, y = keys_133_cast_fp16)[name = tensor("op_8949_cast_fp16")]; + tensor keys_135_cast_fp16 = select(a = var_463_to_fp16, b = keys_133_cast_fp16, cond = var_8949_cast_fp16)[name = tensor("keys_135_cast_fp16")]; + tensor var_8957_cast_fp16 = not_equal(x = values_133_cast_fp16, y = values_133_cast_fp16)[name = tensor("op_8957_cast_fp16")]; + tensor values_135_cast_fp16 = select(a = var_463_to_fp16, b = values_133_cast_fp16, cond = var_8957_cast_fp16)[name = tensor("values_135_cast_fp16")]; + tensor var_8981 = const()[name = tensor("op_8981"), val = tensor([0, 2, 1, 3])]; + tensor var_8997 = const()[name = tensor("op_8997"), val = tensor([1, 1, 1])]; + tensor var_8998_cast_fp16 = reshape(shape = var_8997, x = position22_to_fp16)[name = tensor("op_8998_cast_fp16")]; + tensor var_9018_to_fp16 = const()[name = tensor("op_9018_to_fp16"), val = tensor(0x1p+0)]; + tensor valid_len_45_cast_fp16 = add(x = var_8998_cast_fp16, y = var_9018_to_fp16)[name = tensor("valid_len_45_cast_fp16")]; + tensor valid_mask_45_cast_fp16 = less(x = k_positions_1_promoted_to_fp16, y = valid_len_45_cast_fp16)[name = tensor("valid_mask_45_cast_fp16")]; + tensor causal_mask_45_cast_fp16 = less_equal(x = k_positions_1_promoted_to_fp16, y = var_8998_cast_fp16)[name = tensor("causal_mask_45_cast_fp16")]; + tensor attn_mask_89 = logical_and(x = valid_mask_45_cast_fp16, y = causal_mask_45_cast_fp16)[name = tensor("attn_mask_89")]; + tensor attn_mask_91_axes_0 = const()[name = tensor("attn_mask_91_axes_0"), val = tensor([1])]; + tensor attn_mask_91 = expand_dims(axes = attn_mask_91_axes_0, x = attn_mask_89)[name = tensor("attn_mask_91")]; + tensor var_9036_transpose_x_0 = const()[name = tensor("op_9036_transpose_x_0"), val = tensor(false)]; + tensor var_9036_transpose_y_0 = const()[name = tensor("op_9036_transpose_y_0"), val = tensor(false)]; + tensor transpose_116_perm_0 = const()[name = tensor("transpose_116_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_117_perm_0 = const()[name = tensor("transpose_117_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_117 = transpose(perm = transpose_117_perm_0, x = keys_135_cast_fp16)[name = tensor("transpose_125")]; + tensor transpose_116 = transpose(perm = transpose_116_perm_0, x = q_135_cast_fp16)[name = tensor("transpose_126")]; + tensor var_9036_cast_fp16 = matmul(transpose_x = var_9036_transpose_x_0, transpose_y = var_9036_transpose_y_0, x = transpose_116, y = transpose_117)[name = tensor("op_9036_cast_fp16")]; + tensor var_9030_to_fp16 = const()[name = tensor("op_9030_to_fp16"), val = tensor([0x1p-3])]; + tensor attn_weights_133_cast_fp16 = mul(x = var_9036_cast_fp16, y = var_9030_to_fp16)[name = tensor("attn_weights_133_cast_fp16")]; + tensor var_9038 = logical_not(x = attn_mask_91)[name = tensor("op_9038")]; + tensor var_9039_to_fp16 = const()[name = tensor("op_9039_to_fp16"), val = tensor(-inf)]; + tensor attn_weights_135_cast_fp16 = select(a = var_9039_to_fp16, b = attn_weights_133_cast_fp16, cond = var_9038)[name = tensor("attn_weights_135_cast_fp16")]; + tensor var_9041 = const()[name = tensor("op_9041"), val = tensor(-1)]; + tensor attn_weights_137_cast_fp16 = softmax(axis = var_9041, x = attn_weights_135_cast_fp16)[name = tensor("attn_weights_137_cast_fp16")]; + tensor attn_output_45_transpose_x_0 = const()[name = tensor("attn_output_45_transpose_x_0"), val = tensor(false)]; + tensor attn_output_45_transpose_y_0 = const()[name = tensor("attn_output_45_transpose_y_0"), val = tensor(false)]; + tensor values_137_cast_fp16 = transpose(perm = var_8981, x = values_135_cast_fp16)[name = tensor("transpose_127")]; + tensor attn_output_45_cast_fp16 = matmul(transpose_x = attn_output_45_transpose_x_0, transpose_y = attn_output_45_transpose_y_0, x = attn_weights_137_cast_fp16, y = values_137_cast_fp16)[name = tensor("attn_output_45_cast_fp16")]; + tensor var_9049 = const()[name = tensor("op_9049"), val = tensor([0, 2, 1, 3])]; + tensor var_9054 = const()[name = tensor("op_9054"), val = tensor([1, 1, 1024])]; + tensor var_9050_cast_fp16 = transpose(perm = var_9049, x = attn_output_45_cast_fp16)[name = tensor("transpose_124")]; + tensor input_271_cast_fp16 = reshape(shape = var_9054, x = var_9050_cast_fp16)[name = tensor("input_271_cast_fp16")]; + tensor attn22_out_proj_weight_to_fp16 = const()[name = tensor("attn22_out_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(561270272)))]; + tensor linear_90_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = attn22_out_proj_weight_to_fp16, x = input_271_cast_fp16)[name = tensor("linear_90_cast_fp16")]; + tensor var_9060_to_fp16 = const()[name = tensor("op_9060_to_fp16"), val = tensor(0x1p+0)]; + tensor var_9061_cast_fp16 = add(x = position22_to_fp16, y = var_9060_to_fp16)[name = tensor("op_9061_cast_fp16")]; + tensor var_9061_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_9061_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor input_273_cast_fp16 = add(x = input_267_cast_fp16, y = linear_90_cast_fp16)[name = tensor("input_273_cast_fp16")]; + tensor input_275_axes_0 = const()[name = tensor("input_275_axes_0"), val = tensor([-1])]; + tensor norm22_2_weight_to_fp16 = const()[name = tensor("norm22_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(563367488)))]; + tensor norm22_2_bias_to_fp16 = const()[name = tensor("norm22_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(563369600)))]; + tensor var_9065_to_fp16 = const()[name = tensor("op_9065_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_275_cast_fp16 = layer_norm(axes = input_275_axes_0, beta = norm22_2_bias_to_fp16, epsilon = var_9065_to_fp16, gamma = norm22_2_weight_to_fp16, x = input_273_cast_fp16)[name = tensor("input_275_cast_fp16")]; + tensor linear22_1_weight_to_fp16 = const()[name = tensor("linear22_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(563371712)))]; + tensor linear_91_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = linear22_1_weight_to_fp16, x = input_275_cast_fp16)[name = tensor("linear_91_cast_fp16")]; + tensor input_277_mode_0 = const()[name = tensor("input_277_mode_0"), val = tensor("EXACT")]; + tensor input_277_cast_fp16 = gelu(mode = input_277_mode_0, x = linear_91_cast_fp16)[name = tensor("input_277_cast_fp16")]; + tensor linear22_2_weight_to_fp16 = const()[name = tensor("linear22_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(571760384)))]; + tensor linear_92_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = linear22_2_weight_to_fp16, x = input_277_cast_fp16)[name = tensor("linear_92_cast_fp16")]; + tensor input_279_cast_fp16 = add(x = input_273_cast_fp16, y = linear_92_cast_fp16)[name = tensor("input_279_cast_fp16")]; + tensor input_281_axes_0 = const()[name = tensor("input_281_axes_0"), val = tensor([-1])]; + tensor norm23_1_weight_to_fp16 = const()[name = tensor("norm23_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(580149056)))]; + tensor norm23_1_bias_to_fp16 = const()[name = tensor("norm23_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(580151168)))]; + tensor var_9082_to_fp16 = const()[name = tensor("op_9082_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_281_cast_fp16 = layer_norm(axes = input_281_axes_0, beta = norm23_1_bias_to_fp16, epsilon = var_9082_to_fp16, gamma = norm23_1_weight_to_fp16, x = input_279_cast_fp16)[name = tensor("input_281_cast_fp16")]; + tensor attn23_in_proj_weight_to_fp16 = const()[name = tensor("attn23_in_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(580153280)))]; + tensor linear_93_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = attn23_in_proj_weight_to_fp16, x = input_281_cast_fp16)[name = tensor("linear_93_cast_fp16")]; + tensor var_9096 = const()[name = tensor("op_9096"), val = tensor([1, 1, 3, 16, 64])]; + tensor qkv_cast_fp16 = reshape(shape = var_9096, x = linear_93_cast_fp16)[name = tensor("qkv_cast_fp16")]; + tensor q_139_begin_0 = const()[name = tensor("q_139_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor q_139_end_0 = const()[name = tensor("q_139_end_0"), val = tensor([1, 1, 1, 16, 64])]; + tensor q_139_end_mask_0 = const()[name = tensor("q_139_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor q_139_squeeze_mask_0 = const()[name = tensor("q_139_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor q_139_cast_fp16 = slice_by_index(begin = q_139_begin_0, end = q_139_end_0, end_mask = q_139_end_mask_0, squeeze_mask = q_139_squeeze_mask_0, x = qkv_cast_fp16)[name = tensor("q_139_cast_fp16")]; + tensor k_93_begin_0 = const()[name = tensor("k_93_begin_0"), val = tensor([0, 0, 1, 0, 0])]; + tensor k_93_end_0 = const()[name = tensor("k_93_end_0"), val = tensor([1, 1, 2, 16, 64])]; + tensor k_93_end_mask_0 = const()[name = tensor("k_93_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor k_93_squeeze_mask_0 = const()[name = tensor("k_93_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor k_93_cast_fp16 = slice_by_index(begin = k_93_begin_0, end = k_93_end_0, end_mask = k_93_end_mask_0, squeeze_mask = k_93_squeeze_mask_0, x = qkv_cast_fp16)[name = tensor("k_93_cast_fp16")]; + tensor v_begin_0 = const()[name = tensor("v_begin_0"), val = tensor([0, 0, 2, 0, 0])]; + tensor v_end_0 = const()[name = tensor("v_end_0"), val = tensor([1, 1, 3, 16, 64])]; + tensor v_end_mask_0 = const()[name = tensor("v_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor v_squeeze_mask_0 = const()[name = tensor("v_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor v_cast_fp16 = slice_by_index(begin = v_begin_0, end = v_end_0, end_mask = v_end_mask_0, squeeze_mask = v_squeeze_mask_0, x = qkv_cast_fp16)[name = tensor("v_cast_fp16")]; + tensor var_9162 = const()[name = tensor("op_9162"), val = tensor([1, 1, 1, 1])]; + tensor position23_to_fp16_dtype_0 = const()[name = tensor("position23_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor position23_to_fp16 = cast(dtype = position23_to_fp16_dtype_0, x = position23)[name = tensor("cast_438")]; + tensor ts_cast_fp16 = reshape(shape = var_9162, x = position23_to_fp16)[name = tensor("ts_cast_fp16")]; + tensor var_9169 = const()[name = tensor("op_9169"), val = tensor([1, 1, 16, 32, 2])]; + tensor q_complex_cast_fp16 = reshape(shape = var_9169, x = q_139_cast_fp16)[name = tensor("q_complex_cast_fp16")]; + tensor var_9176 = const()[name = tensor("op_9176"), val = tensor([1, 1, 16, 32, 2])]; + tensor k_complex_cast_fp16 = reshape(shape = var_9176, x = k_93_cast_fp16)[name = tensor("k_complex_cast_fp16")]; + tensor var_9180_begin_0 = const()[name = tensor("op_9180_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_9180_end_0 = const()[name = tensor("op_9180_end_0"), val = tensor([1, 1, 16, 32, 1])]; + tensor var_9180_end_mask_0 = const()[name = tensor("op_9180_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_9180_squeeze_mask_0 = const()[name = tensor("op_9180_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_9180_cast_fp16 = slice_by_index(begin = var_9180_begin_0, end = var_9180_end_0, end_mask = var_9180_end_mask_0, squeeze_mask = var_9180_squeeze_mask_0, x = q_complex_cast_fp16)[name = tensor("op_9180_cast_fp16")]; + tensor var_9188_begin_0 = const()[name = tensor("op_9188_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_9188_end_0 = const()[name = tensor("op_9188_end_0"), val = tensor([1, 1, 16, 32, 2])]; + tensor var_9188_end_mask_0 = const()[name = tensor("op_9188_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_9188_squeeze_mask_0 = const()[name = tensor("op_9188_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_9188_cast_fp16 = slice_by_index(begin = var_9188_begin_0, end = var_9188_end_0, end_mask = var_9188_end_mask_0, squeeze_mask = var_9188_squeeze_mask_0, x = q_complex_cast_fp16)[name = tensor("op_9188_cast_fp16")]; + tensor var_9196_begin_0 = const()[name = tensor("op_9196_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_9196_end_0 = const()[name = tensor("op_9196_end_0"), val = tensor([1, 1, 16, 32, 1])]; + tensor var_9196_end_mask_0 = const()[name = tensor("op_9196_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_9196_squeeze_mask_0 = const()[name = tensor("op_9196_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_9196_cast_fp16 = slice_by_index(begin = var_9196_begin_0, end = var_9196_end_0, end_mask = var_9196_end_mask_0, squeeze_mask = var_9196_squeeze_mask_0, x = k_complex_cast_fp16)[name = tensor("op_9196_cast_fp16")]; + tensor var_9204_begin_0 = const()[name = tensor("op_9204_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_9204_end_0 = const()[name = tensor("op_9204_end_0"), val = tensor([1, 1, 16, 32, 2])]; + tensor var_9204_end_mask_0 = const()[name = tensor("op_9204_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_9204_squeeze_mask_0 = const()[name = tensor("op_9204_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_9204_cast_fp16 = slice_by_index(begin = var_9204_begin_0, end = var_9204_end_0, end_mask = var_9204_end_mask_0, squeeze_mask = var_9204_squeeze_mask_0, x = k_complex_cast_fp16)[name = tensor("op_9204_cast_fp16")]; + tensor freqs_to_fp16 = const()[name = tensor("freqs_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(586444800)))]; + tensor var_9210_cast_fp16 = mul(x = freqs_to_fp16, y = ts_cast_fp16)[name = tensor("op_9210_cast_fp16")]; + tensor rotr_cast_fp16 = cos(x = var_9210_cast_fp16)[name = tensor("rotr_cast_fp16")]; + tensor roti_cast_fp16 = sin(x = var_9210_cast_fp16)[name = tensor("roti_cast_fp16")]; + tensor var_9214_cast_fp16 = mul(x = var_9180_cast_fp16, y = rotr_cast_fp16)[name = tensor("op_9214_cast_fp16")]; + tensor var_9215_cast_fp16 = mul(x = var_9188_cast_fp16, y = roti_cast_fp16)[name = tensor("op_9215_cast_fp16")]; + tensor qor_93_cast_fp16 = sub(x = var_9214_cast_fp16, y = var_9215_cast_fp16)[name = tensor("qor_93_cast_fp16")]; + tensor var_9218_cast_fp16 = mul(x = var_9180_cast_fp16, y = roti_cast_fp16)[name = tensor("op_9218_cast_fp16")]; + tensor var_9219_cast_fp16 = mul(x = var_9188_cast_fp16, y = rotr_cast_fp16)[name = tensor("op_9219_cast_fp16")]; + tensor qoi_93_cast_fp16 = add(x = var_9218_cast_fp16, y = var_9219_cast_fp16)[name = tensor("qoi_93_cast_fp16")]; + tensor var_9222_cast_fp16 = mul(x = var_9196_cast_fp16, y = rotr_cast_fp16)[name = tensor("op_9222_cast_fp16")]; + tensor var_9223_cast_fp16 = mul(x = var_9204_cast_fp16, y = roti_cast_fp16)[name = tensor("op_9223_cast_fp16")]; + tensor kor_93_cast_fp16 = sub(x = var_9222_cast_fp16, y = var_9223_cast_fp16)[name = tensor("kor_93_cast_fp16")]; + tensor var_9226_cast_fp16 = mul(x = var_9196_cast_fp16, y = roti_cast_fp16)[name = tensor("op_9226_cast_fp16")]; + tensor var_9227_cast_fp16 = mul(x = var_9204_cast_fp16, y = rotr_cast_fp16)[name = tensor("op_9227_cast_fp16")]; + tensor koi_93_cast_fp16 = add(x = var_9226_cast_fp16, y = var_9227_cast_fp16)[name = tensor("koi_93_cast_fp16")]; + tensor qo_axis_0 = const()[name = tensor("qo_axis_0"), val = tensor(-1)]; + tensor qo_cast_fp16 = stack(axis = qo_axis_0, values = (qor_93_cast_fp16, qoi_93_cast_fp16))[name = tensor("qo_cast_fp16")]; + tensor ko_axis_0 = const()[name = tensor("ko_axis_0"), val = tensor(-1)]; + tensor ko_cast_fp16 = stack(axis = ko_axis_0, values = (kor_93_cast_fp16, koi_93_cast_fp16))[name = tensor("ko_cast_fp16")]; + tensor var_9260 = const()[name = tensor("op_9260"), val = tensor([1, 1, 16, 64])]; + tensor q_141_cast_fp16 = reshape(shape = var_9260, x = qo_cast_fp16)[name = tensor("q_141_cast_fp16")]; + tensor var_9266 = const()[name = tensor("op_9266"), val = tensor([1, 1, 16, 64])]; + tensor k_cast_fp16 = reshape(shape = var_9266, x = ko_cast_fp16)[name = tensor("k_cast_fp16")]; + tensor _inversed_9291_y_0_to_fp16 = const()[name = tensor("_inversed_9291_y_0_to_fp16"), val = tensor(0x1p-9)]; + tensor _inversed_9291_cast_fp16 = mul(x = ts_cast_fp16, y = _inversed_9291_y_0_to_fp16)[name = tensor("_inversed_9291_cast_fp16")]; + tensor var_9292_cast_fp16 = floor(x = _inversed_9291_cast_fp16)[name = tensor("op_9292_cast_fp16")]; + tensor var_9293_to_fp16 = const()[name = tensor("op_9293_to_fp16"), val = tensor(0x1p+9)]; + tensor var_9294_cast_fp16 = mul(x = var_9292_cast_fp16, y = var_9293_to_fp16)[name = tensor("op_9294_cast_fp16")]; + tensor write_indices_float_cast_fp16 = sub(x = ts_cast_fp16, y = var_9294_cast_fp16)[name = tensor("write_indices_float_cast_fp16")]; + tensor var_9301_dtype_0 = const()[name = tensor("op_9301_dtype_0"), val = tensor("int32")]; + tensor write_indices_reps_0 = const()[name = tensor("write_indices_reps_0"), val = tensor([1, 1, 16, 64])]; + tensor write_indices_float_cast_fp16_to_int32 = cast(dtype = var_9301_dtype_0, x = write_indices_float_cast_fp16)[name = tensor("cast_437")]; + tensor write_indices = tile(reps = write_indices_reps_0, x = write_indices_float_cast_fp16_to_int32)[name = tensor("write_indices")]; + tensor var_9311_begin_0 = const()[name = tensor("op_9311_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_9311_end_0 = const()[name = tensor("op_9311_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor var_9311_end_mask_0 = const()[name = tensor("op_9311_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_9311_squeeze_mask_0 = const()[name = tensor("op_9311_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor cache23_to_fp16_dtype_0 = const()[name = tensor("cache23_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor cache23_to_fp16 = cast(dtype = cache23_to_fp16_dtype_0, x = cache23)[name = tensor("cast_436")]; + tensor var_9311_cast_fp16 = slice_by_index(begin = var_9311_begin_0, end = var_9311_end_0, end_mask = var_9311_end_mask_0, squeeze_mask = var_9311_squeeze_mask_0, x = cache23_to_fp16)[name = tensor("op_9311_cast_fp16")]; + tensor var_9313_axis_0 = const()[name = tensor("op_9313_axis_0"), val = tensor(1)]; + tensor var_9313_mode_0 = const()[name = tensor("op_9313_mode_0"), val = tensor("update")]; + tensor var_9313_validate_indices_0 = const()[name = tensor("op_9313_validate_indices_0"), val = tensor(false)]; + tensor var_9313_cast_fp16 = scatter_along_axis(axis = var_9313_axis_0, data = var_9311_cast_fp16, indices = write_indices, mode = var_9313_mode_0, updates = k_cast_fp16, validate_indices = var_9313_validate_indices_0)[name = tensor("op_9313_cast_fp16")]; + tensor concat_163 = const()[name = tensor("concat_163"), val = tensor([0, 0, 0, 0, 0])]; + tensor concat_164 = const()[name = tensor("concat_164"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_internal_tensor_assign_1_stride_0 = const()[name = tensor("new_cache_internal_tensor_assign_1_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_internal_tensor_assign_1_begin_mask_0 = const()[name = tensor("new_cache_internal_tensor_assign_1_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_internal_tensor_assign_1_end_mask_0 = const()[name = tensor("new_cache_internal_tensor_assign_1_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_internal_tensor_assign_1_squeeze_mask_0 = const()[name = tensor("new_cache_internal_tensor_assign_1_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_94 = const()[name = tensor("shape_94"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_46 = const()[name = tensor("reduce_prod_46"), val = tensor(1048576)]; + tensor range_1d_46_start_0 = const()[name = tensor("range_1d_46_start_0"), val = tensor(0)]; + tensor range_1d_46_step_0 = const()[name = tensor("range_1d_46_step_0"), val = tensor(1)]; + tensor range_1d_46 = range_1d(end = reduce_prod_46, start = range_1d_46_start_0, step = range_1d_46_step_0)[name = tensor("range_1d_46")]; + tensor reshape_230 = reshape(shape = shape_94, x = range_1d_46)[name = tensor("reshape_230")]; + tensor slice_by_index_46 = slice_by_index(begin = concat_163, begin_mask = new_cache_internal_tensor_assign_1_begin_mask_0, end = concat_164, end_mask = new_cache_internal_tensor_assign_1_end_mask_0, squeeze_mask = new_cache_internal_tensor_assign_1_squeeze_mask_0, stride = new_cache_internal_tensor_assign_1_stride_0, x = reshape_230)[name = tensor("slice_by_index_46")]; + tensor reshape_231_shape_0 = const()[name = tensor("reshape_231_shape_0"), val = tensor([-1])]; + tensor reshape_231 = reshape(shape = reshape_231_shape_0, x = slice_by_index_46)[name = tensor("reshape_231")]; + tensor reshape_232_shape_0 = const()[name = tensor("reshape_232_shape_0"), val = tensor([-1])]; + tensor reshape_232_cast_fp16 = reshape(shape = reshape_232_shape_0, x = var_9313_cast_fp16)[name = tensor("reshape_232_cast_fp16")]; + tensor reshape_233_shape_0 = const()[name = tensor("reshape_233_shape_0"), val = tensor([-1])]; + tensor reshape_233_cast_fp16 = reshape(shape = reshape_233_shape_0, x = cache23_to_fp16)[name = tensor("reshape_233_cast_fp16")]; + tensor scatter_46_mode_0 = const()[name = tensor("scatter_46_mode_0"), val = tensor("update")]; + tensor scatter_46_axis_0 = const()[name = tensor("scatter_46_axis_0"), val = tensor(0)]; + tensor scatter_46_validate_indices_0 = const()[name = tensor("scatter_46_validate_indices_0"), val = tensor(false)]; + tensor scatter_46_cast_fp16 = scatter(axis = scatter_46_axis_0, data = reshape_233_cast_fp16, indices = reshape_231, mode = scatter_46_mode_0, updates = reshape_232_cast_fp16, validate_indices = scatter_46_validate_indices_0)[name = tensor("scatter_46_cast_fp16")]; + tensor reshape_234_cast_fp16 = reshape(shape = shape_94, x = scatter_46_cast_fp16)[name = tensor("reshape_234_cast_fp16")]; + tensor var_9321_begin_0 = const()[name = tensor("op_9321_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor var_9321_end_0 = const()[name = tensor("op_9321_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor var_9321_end_mask_0 = const()[name = tensor("op_9321_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_9321_squeeze_mask_0 = const()[name = tensor("op_9321_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor var_9321_cast_fp16 = slice_by_index(begin = var_9321_begin_0, end = var_9321_end_0, end_mask = var_9321_end_mask_0, squeeze_mask = var_9321_squeeze_mask_0, x = reshape_234_cast_fp16)[name = tensor("op_9321_cast_fp16")]; + tensor var_9323_axis_0 = const()[name = tensor("op_9323_axis_0"), val = tensor(1)]; + tensor var_9323_mode_0 = const()[name = tensor("op_9323_mode_0"), val = tensor("update")]; + tensor var_9323_validate_indices_0 = const()[name = tensor("op_9323_validate_indices_0"), val = tensor(false)]; + tensor var_9323_cast_fp16 = scatter_along_axis(axis = var_9323_axis_0, data = var_9321_cast_fp16, indices = write_indices, mode = var_9323_mode_0, updates = v_cast_fp16, validate_indices = var_9323_validate_indices_0)[name = tensor("op_9323_cast_fp16")]; + tensor concat_165 = const()[name = tensor("concat_165"), val = tensor([1, 0, 0, 0, 0])]; + tensor concat_166 = const()[name = tensor("concat_166"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_internal_tensor_assign_2_stride_0 = const()[name = tensor("new_cache_internal_tensor_assign_2_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_internal_tensor_assign_2_begin_mask_0 = const()[name = tensor("new_cache_internal_tensor_assign_2_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_internal_tensor_assign_2_end_mask_0 = const()[name = tensor("new_cache_internal_tensor_assign_2_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_internal_tensor_assign_2_squeeze_mask_0 = const()[name = tensor("new_cache_internal_tensor_assign_2_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_95 = const()[name = tensor("shape_95"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_47 = const()[name = tensor("reduce_prod_47"), val = tensor(1048576)]; + tensor range_1d_47_start_0 = const()[name = tensor("range_1d_47_start_0"), val = tensor(0)]; + tensor range_1d_47_step_0 = const()[name = tensor("range_1d_47_step_0"), val = tensor(1)]; + tensor range_1d_47 = range_1d(end = reduce_prod_47, start = range_1d_47_start_0, step = range_1d_47_step_0)[name = tensor("range_1d_47")]; + tensor reshape_235 = reshape(shape = shape_95, x = range_1d_47)[name = tensor("reshape_235")]; + tensor slice_by_index_47 = slice_by_index(begin = concat_165, begin_mask = new_cache_internal_tensor_assign_2_begin_mask_0, end = concat_166, end_mask = new_cache_internal_tensor_assign_2_end_mask_0, squeeze_mask = new_cache_internal_tensor_assign_2_squeeze_mask_0, stride = new_cache_internal_tensor_assign_2_stride_0, x = reshape_235)[name = tensor("slice_by_index_47")]; + tensor reshape_236_shape_0 = const()[name = tensor("reshape_236_shape_0"), val = tensor([-1])]; + tensor reshape_236 = reshape(shape = reshape_236_shape_0, x = slice_by_index_47)[name = tensor("reshape_236")]; + tensor reshape_237_shape_0 = const()[name = tensor("reshape_237_shape_0"), val = tensor([-1])]; + tensor reshape_237_cast_fp16 = reshape(shape = reshape_237_shape_0, x = var_9323_cast_fp16)[name = tensor("reshape_237_cast_fp16")]; + tensor reshape_238_shape_0 = const()[name = tensor("reshape_238_shape_0"), val = tensor([-1])]; + tensor reshape_238_cast_fp16 = reshape(shape = reshape_238_shape_0, x = reshape_234_cast_fp16)[name = tensor("reshape_238_cast_fp16")]; + tensor scatter_47_mode_0 = const()[name = tensor("scatter_47_mode_0"), val = tensor("update")]; + tensor scatter_47_axis_0 = const()[name = tensor("scatter_47_axis_0"), val = tensor(0)]; + tensor scatter_47_validate_indices_0 = const()[name = tensor("scatter_47_validate_indices_0"), val = tensor(false)]; + tensor scatter_47_cast_fp16 = scatter(axis = scatter_47_axis_0, data = reshape_238_cast_fp16, indices = reshape_236, mode = scatter_47_mode_0, updates = reshape_237_cast_fp16, validate_indices = scatter_47_validate_indices_0)[name = tensor("scatter_47_cast_fp16")]; + tensor reshape_239_cast_fp16 = reshape(shape = shape_95, x = scatter_47_cast_fp16)[name = tensor("reshape_239_cast_fp16")]; + tensor reshape_239_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("reshape_239_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor keys_139_begin_0 = const()[name = tensor("keys_139_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor keys_139_end_0 = const()[name = tensor("keys_139_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor keys_139_end_mask_0 = const()[name = tensor("keys_139_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor keys_139_squeeze_mask_0 = const()[name = tensor("keys_139_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor keys_139_cast_fp16 = slice_by_index(begin = keys_139_begin_0, end = keys_139_end_0, end_mask = keys_139_end_mask_0, squeeze_mask = keys_139_squeeze_mask_0, x = reshape_239_cast_fp16)[name = tensor("keys_139_cast_fp16")]; + tensor values_139_begin_0 = const()[name = tensor("values_139_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor values_139_end_0 = const()[name = tensor("values_139_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor values_139_end_mask_0 = const()[name = tensor("values_139_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor values_139_squeeze_mask_0 = const()[name = tensor("values_139_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor values_139_cast_fp16 = slice_by_index(begin = values_139_begin_0, end = values_139_end_0, end_mask = values_139_end_mask_0, squeeze_mask = values_139_squeeze_mask_0, x = reshape_239_cast_fp16)[name = tensor("values_139_cast_fp16")]; + tensor var_9335_cast_fp16 = not_equal(x = keys_139_cast_fp16, y = keys_139_cast_fp16)[name = tensor("op_9335_cast_fp16")]; + tensor keys_141_cast_fp16 = select(a = var_463_to_fp16, b = keys_139_cast_fp16, cond = var_9335_cast_fp16)[name = tensor("keys_141_cast_fp16")]; + tensor var_9343_cast_fp16 = not_equal(x = values_139_cast_fp16, y = values_139_cast_fp16)[name = tensor("op_9343_cast_fp16")]; + tensor values_141_cast_fp16 = select(a = var_463_to_fp16, b = values_139_cast_fp16, cond = var_9343_cast_fp16)[name = tensor("values_141_cast_fp16")]; + tensor var_9367 = const()[name = tensor("op_9367"), val = tensor([0, 2, 1, 3])]; + tensor var_9383 = const()[name = tensor("op_9383"), val = tensor([1, 1, 1])]; + tensor var_9384_cast_fp16 = reshape(shape = var_9383, x = position23_to_fp16)[name = tensor("op_9384_cast_fp16")]; + tensor var_9404_to_fp16 = const()[name = tensor("op_9404_to_fp16"), val = tensor(0x1p+0)]; + tensor valid_len_cast_fp16 = add(x = var_9384_cast_fp16, y = var_9404_to_fp16)[name = tensor("valid_len_cast_fp16")]; + tensor valid_mask_cast_fp16 = less(x = k_positions_1_promoted_to_fp16, y = valid_len_cast_fp16)[name = tensor("valid_mask_cast_fp16")]; + tensor causal_mask_cast_fp16 = less_equal(x = k_positions_1_promoted_to_fp16, y = var_9384_cast_fp16)[name = tensor("causal_mask_cast_fp16")]; + tensor attn_mask_93 = logical_and(x = valid_mask_cast_fp16, y = causal_mask_cast_fp16)[name = tensor("attn_mask_93")]; + tensor attn_mask_axes_0 = const()[name = tensor("attn_mask_axes_0"), val = tensor([1])]; + tensor attn_mask = expand_dims(axes = attn_mask_axes_0, x = attn_mask_93)[name = tensor("attn_mask")]; + tensor var_9422_transpose_x_0 = const()[name = tensor("op_9422_transpose_x_0"), val = tensor(false)]; + tensor var_9422_transpose_y_0 = const()[name = tensor("op_9422_transpose_y_0"), val = tensor(false)]; + tensor transpose_118_perm_0 = const()[name = tensor("transpose_118_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_119_perm_0 = const()[name = tensor("transpose_119_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_119 = transpose(perm = transpose_119_perm_0, x = keys_141_cast_fp16)[name = tensor("transpose_121")]; + tensor transpose_118 = transpose(perm = transpose_118_perm_0, x = q_141_cast_fp16)[name = tensor("transpose_122")]; + tensor var_9422_cast_fp16 = matmul(transpose_x = var_9422_transpose_x_0, transpose_y = var_9422_transpose_y_0, x = transpose_118, y = transpose_119)[name = tensor("op_9422_cast_fp16")]; + tensor var_9416_to_fp16 = const()[name = tensor("op_9416_to_fp16"), val = tensor([0x1p-3])]; + tensor attn_weights_139_cast_fp16 = mul(x = var_9422_cast_fp16, y = var_9416_to_fp16)[name = tensor("attn_weights_139_cast_fp16")]; + tensor var_9424 = logical_not(x = attn_mask)[name = tensor("op_9424")]; + tensor var_9425_to_fp16 = const()[name = tensor("op_9425_to_fp16"), val = tensor(-inf)]; + tensor attn_weights_141_cast_fp16 = select(a = var_9425_to_fp16, b = attn_weights_139_cast_fp16, cond = var_9424)[name = tensor("attn_weights_141_cast_fp16")]; + tensor var_9427 = const()[name = tensor("op_9427"), val = tensor(-1)]; + tensor attn_weights_cast_fp16 = softmax(axis = var_9427, x = attn_weights_141_cast_fp16)[name = tensor("attn_weights_cast_fp16")]; + tensor attn_output_transpose_x_0 = const()[name = tensor("attn_output_transpose_x_0"), val = tensor(false)]; + tensor attn_output_transpose_y_0 = const()[name = tensor("attn_output_transpose_y_0"), val = tensor(false)]; + tensor values_cast_fp16 = transpose(perm = var_9367, x = values_141_cast_fp16)[name = tensor("transpose_123")]; + tensor attn_output_cast_fp16 = matmul(transpose_x = attn_output_transpose_x_0, transpose_y = attn_output_transpose_y_0, x = attn_weights_cast_fp16, y = values_cast_fp16)[name = tensor("attn_output_cast_fp16")]; + tensor var_9435 = const()[name = tensor("op_9435"), val = tensor([0, 2, 1, 3])]; + tensor var_9440 = const()[name = tensor("op_9440"), val = tensor([1, 1, 1024])]; + tensor var_9436_cast_fp16 = transpose(perm = var_9435, x = attn_output_cast_fp16)[name = tensor("transpose_120")]; + tensor input_283_cast_fp16 = reshape(shape = var_9440, x = var_9436_cast_fp16)[name = tensor("input_283_cast_fp16")]; + tensor attn23_out_proj_weight_to_fp16 = const()[name = tensor("attn23_out_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(586444928)))]; + tensor linear_94_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = attn23_out_proj_weight_to_fp16, x = input_283_cast_fp16)[name = tensor("linear_94_cast_fp16")]; + tensor var_9446_to_fp16 = const()[name = tensor("op_9446_to_fp16"), val = tensor(0x1p+0)]; + tensor var_9447_cast_fp16 = add(x = position23_to_fp16, y = var_9446_to_fp16)[name = tensor("op_9447_cast_fp16")]; + tensor var_9447_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_9447_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor input_285_cast_fp16 = add(x = input_279_cast_fp16, y = linear_94_cast_fp16)[name = tensor("input_285_cast_fp16")]; + tensor input_287_axes_0 = const()[name = tensor("input_287_axes_0"), val = tensor([-1])]; + tensor norm23_2_weight_to_fp16 = const()[name = tensor("norm23_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(588542144)))]; + tensor norm23_2_bias_to_fp16 = const()[name = tensor("norm23_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(588544256)))]; + tensor var_9451_to_fp16 = const()[name = tensor("op_9451_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_287_cast_fp16 = layer_norm(axes = input_287_axes_0, beta = norm23_2_bias_to_fp16, epsilon = var_9451_to_fp16, gamma = norm23_2_weight_to_fp16, x = input_285_cast_fp16)[name = tensor("input_287_cast_fp16")]; + tensor linear23_1_weight_to_fp16 = const()[name = tensor("linear23_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(588546368)))]; + tensor linear_95_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = linear23_1_weight_to_fp16, x = input_287_cast_fp16)[name = tensor("linear_95_cast_fp16")]; + tensor input_289_mode_0 = const()[name = tensor("input_289_mode_0"), val = tensor("EXACT")]; + tensor input_289_cast_fp16 = gelu(mode = input_289_mode_0, x = linear_95_cast_fp16)[name = tensor("input_289_cast_fp16")]; + tensor linear23_2_weight_to_fp16 = const()[name = tensor("linear23_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(596935040)))]; + tensor linear_96_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = linear23_2_weight_to_fp16, x = input_289_cast_fp16)[name = tensor("linear_96_cast_fp16")]; + tensor input_291_cast_fp16 = add(x = input_285_cast_fp16, y = linear_96_cast_fp16)[name = tensor("input_291_cast_fp16")]; + tensor input_axes_0 = const()[name = tensor("input_axes_0"), val = tensor([-1])]; + tensor out_norm_weight_to_fp16 = const()[name = tensor("out_norm_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(605323712)))]; + tensor out_norm_bias_to_fp16 = const()[name = tensor("out_norm_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(605325824)))]; + tensor var_9468_to_fp16 = const()[name = tensor("op_9468_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_cast_fp16 = layer_norm(axes = input_axes_0, beta = out_norm_bias_to_fp16, epsilon = var_9468_to_fp16, gamma = out_norm_weight_to_fp16, x = input_291_cast_fp16)[name = tensor("input_cast_fp16")]; + tensor input_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("input_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor out_eos_weight_to_fp16 = const()[name = tensor("out_eos_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(605327936)))]; + tensor out_eos_bias_to_fp16 = const()[name = tensor("out_eos_bias_to_fp16"), val = tensor([-0x1.8ap-3])]; + tensor linear_97_cast_fp16 = linear(bias = out_eos_bias_to_fp16, weight = out_eos_weight_to_fp16, x = input_cast_fp16)[name = tensor("linear_97_cast_fp16")]; + tensor linear_97_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("linear_97_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor var_9476 = cast(dtype = linear_97_cast_fp16_to_fp32_dtype_0, x = linear_97_cast_fp16)[name = tensor("cast_432")]; + tensor input = cast(dtype = input_cast_fp16_to_fp32_dtype_0, x = input_cast_fp16)[name = tensor("cast_433")]; + tensor var_9447 = cast(dtype = var_9447_cast_fp16_to_fp32_dtype_0, x = var_9447_cast_fp16)[name = tensor("cast_434")]; + tensor new_cache_internal_tensor_assign_2 = cast(dtype = reshape_239_cast_fp16_to_fp32_dtype_0, x = reshape_239_cast_fp16)[name = tensor("cast_435")]; + tensor var_9061 = cast(dtype = var_9061_cast_fp16_to_fp32_dtype_0, x = var_9061_cast_fp16)[name = tensor("cast_439")]; + tensor new_cache_45_internal_tensor_assign_2 = cast(dtype = reshape_229_cast_fp16_to_fp32_dtype_0, x = reshape_229_cast_fp16)[name = tensor("cast_440")]; + tensor var_8675 = cast(dtype = var_8675_cast_fp16_to_fp32_dtype_0, x = var_8675_cast_fp16)[name = tensor("cast_444")]; + tensor new_cache_43_internal_tensor_assign_2 = cast(dtype = reshape_219_cast_fp16_to_fp32_dtype_0, x = reshape_219_cast_fp16)[name = tensor("cast_445")]; + tensor var_8289 = cast(dtype = var_8289_cast_fp16_to_fp32_dtype_0, x = var_8289_cast_fp16)[name = tensor("cast_449")]; + tensor new_cache_41_internal_tensor_assign_2 = cast(dtype = reshape_209_cast_fp16_to_fp32_dtype_0, x = reshape_209_cast_fp16)[name = tensor("cast_450")]; + tensor var_7903 = cast(dtype = var_7903_cast_fp16_to_fp32_dtype_0, x = var_7903_cast_fp16)[name = tensor("cast_454")]; + tensor new_cache_39_internal_tensor_assign_2 = cast(dtype = reshape_199_cast_fp16_to_fp32_dtype_0, x = reshape_199_cast_fp16)[name = tensor("cast_455")]; + tensor var_7517 = cast(dtype = var_7517_cast_fp16_to_fp32_dtype_0, x = var_7517_cast_fp16)[name = tensor("cast_459")]; + tensor new_cache_37_internal_tensor_assign_2 = cast(dtype = reshape_189_cast_fp16_to_fp32_dtype_0, x = reshape_189_cast_fp16)[name = tensor("cast_460")]; + tensor var_7131 = cast(dtype = var_7131_cast_fp16_to_fp32_dtype_0, x = var_7131_cast_fp16)[name = tensor("cast_464")]; + tensor new_cache_35_internal_tensor_assign_2 = cast(dtype = reshape_179_cast_fp16_to_fp32_dtype_0, x = reshape_179_cast_fp16)[name = tensor("cast_465")]; + tensor var_6745 = cast(dtype = var_6745_cast_fp16_to_fp32_dtype_0, x = var_6745_cast_fp16)[name = tensor("cast_469")]; + tensor new_cache_33_internal_tensor_assign_2 = cast(dtype = reshape_169_cast_fp16_to_fp32_dtype_0, x = reshape_169_cast_fp16)[name = tensor("cast_470")]; + tensor var_6359 = cast(dtype = var_6359_cast_fp16_to_fp32_dtype_0, x = var_6359_cast_fp16)[name = tensor("cast_474")]; + tensor new_cache_31_internal_tensor_assign_2 = cast(dtype = reshape_159_cast_fp16_to_fp32_dtype_0, x = reshape_159_cast_fp16)[name = tensor("cast_475")]; + tensor var_5973 = cast(dtype = var_5973_cast_fp16_to_fp32_dtype_0, x = var_5973_cast_fp16)[name = tensor("cast_479")]; + tensor new_cache_29_internal_tensor_assign_2 = cast(dtype = reshape_149_cast_fp16_to_fp32_dtype_0, x = reshape_149_cast_fp16)[name = tensor("cast_480")]; + tensor var_5587 = cast(dtype = var_5587_cast_fp16_to_fp32_dtype_0, x = var_5587_cast_fp16)[name = tensor("cast_484")]; + tensor new_cache_27_internal_tensor_assign_2 = cast(dtype = reshape_139_cast_fp16_to_fp32_dtype_0, x = reshape_139_cast_fp16)[name = tensor("cast_485")]; + tensor var_5201 = cast(dtype = var_5201_cast_fp16_to_fp32_dtype_0, x = var_5201_cast_fp16)[name = tensor("cast_489")]; + tensor new_cache_25_internal_tensor_assign_2 = cast(dtype = reshape_129_cast_fp16_to_fp32_dtype_0, x = reshape_129_cast_fp16)[name = tensor("cast_490")]; + tensor var_4815 = cast(dtype = var_4815_cast_fp16_to_fp32_dtype_0, x = var_4815_cast_fp16)[name = tensor("cast_494")]; + tensor new_cache_23_internal_tensor_assign_2 = cast(dtype = reshape_119_cast_fp16_to_fp32_dtype_0, x = reshape_119_cast_fp16)[name = tensor("cast_495")]; + tensor var_4429 = cast(dtype = var_4429_cast_fp16_to_fp32_dtype_0, x = var_4429_cast_fp16)[name = tensor("cast_499")]; + tensor new_cache_21_internal_tensor_assign_2 = cast(dtype = reshape_109_cast_fp16_to_fp32_dtype_0, x = reshape_109_cast_fp16)[name = tensor("cast_500")]; + tensor var_4043 = cast(dtype = var_4043_cast_fp16_to_fp32_dtype_0, x = var_4043_cast_fp16)[name = tensor("cast_504")]; + tensor new_cache_19_internal_tensor_assign_2 = cast(dtype = reshape_99_cast_fp16_to_fp32_dtype_0, x = reshape_99_cast_fp16)[name = tensor("cast_505")]; + tensor var_3657 = cast(dtype = var_3657_cast_fp16_to_fp32_dtype_0, x = var_3657_cast_fp16)[name = tensor("cast_509")]; + tensor new_cache_17_internal_tensor_assign_2 = cast(dtype = reshape_89_cast_fp16_to_fp32_dtype_0, x = reshape_89_cast_fp16)[name = tensor("cast_510")]; + tensor var_3271 = cast(dtype = var_3271_cast_fp16_to_fp32_dtype_0, x = var_3271_cast_fp16)[name = tensor("cast_514")]; + tensor new_cache_15_internal_tensor_assign_2 = cast(dtype = reshape_79_cast_fp16_to_fp32_dtype_0, x = reshape_79_cast_fp16)[name = tensor("cast_515")]; + tensor var_2885 = cast(dtype = var_2885_cast_fp16_to_fp32_dtype_0, x = var_2885_cast_fp16)[name = tensor("cast_519")]; + tensor new_cache_13_internal_tensor_assign_2 = cast(dtype = reshape_69_cast_fp16_to_fp32_dtype_0, x = reshape_69_cast_fp16)[name = tensor("cast_520")]; + tensor var_2499 = cast(dtype = var_2499_cast_fp16_to_fp32_dtype_0, x = var_2499_cast_fp16)[name = tensor("cast_524")]; + tensor new_cache_11_internal_tensor_assign_2 = cast(dtype = reshape_59_cast_fp16_to_fp32_dtype_0, x = reshape_59_cast_fp16)[name = tensor("cast_525")]; + tensor var_2113 = cast(dtype = var_2113_cast_fp16_to_fp32_dtype_0, x = var_2113_cast_fp16)[name = tensor("cast_529")]; + tensor new_cache_9_internal_tensor_assign_2 = cast(dtype = reshape_49_cast_fp16_to_fp32_dtype_0, x = reshape_49_cast_fp16)[name = tensor("cast_530")]; + tensor var_1727 = cast(dtype = var_1727_cast_fp16_to_fp32_dtype_0, x = var_1727_cast_fp16)[name = tensor("cast_534")]; + tensor new_cache_7_internal_tensor_assign_2 = cast(dtype = reshape_39_cast_fp16_to_fp32_dtype_0, x = reshape_39_cast_fp16)[name = tensor("cast_535")]; + tensor var_1341 = cast(dtype = var_1341_cast_fp16_to_fp32_dtype_0, x = var_1341_cast_fp16)[name = tensor("cast_539")]; + tensor new_cache_5_internal_tensor_assign_2 = cast(dtype = reshape_29_cast_fp16_to_fp32_dtype_0, x = reshape_29_cast_fp16)[name = tensor("cast_540")]; + tensor var_955 = cast(dtype = var_955_cast_fp16_to_fp32_dtype_0, x = var_955_cast_fp16)[name = tensor("cast_544")]; + tensor new_cache_3_internal_tensor_assign_2 = cast(dtype = reshape_19_cast_fp16_to_fp32_dtype_0, x = reshape_19_cast_fp16)[name = tensor("cast_545")]; + tensor var_569 = cast(dtype = var_569_cast_fp16_to_fp32_dtype_0, x = var_569_cast_fp16)[name = tensor("cast_549")]; + tensor new_cache_1_internal_tensor_assign_2 = cast(dtype = reshape_9_cast_fp16_to_fp32_dtype_0, x = reshape_9_cast_fp16)[name = tensor("cast_550")]; + } -> (input, var_9476, new_cache_1_internal_tensor_assign_2, var_569, new_cache_3_internal_tensor_assign_2, var_955, new_cache_5_internal_tensor_assign_2, var_1341, new_cache_7_internal_tensor_assign_2, var_1727, new_cache_9_internal_tensor_assign_2, var_2113, new_cache_11_internal_tensor_assign_2, var_2499, new_cache_13_internal_tensor_assign_2, var_2885, new_cache_15_internal_tensor_assign_2, var_3271, new_cache_17_internal_tensor_assign_2, var_3657, new_cache_19_internal_tensor_assign_2, var_4043, new_cache_21_internal_tensor_assign_2, var_4429, new_cache_23_internal_tensor_assign_2, var_4815, new_cache_25_internal_tensor_assign_2, var_5201, new_cache_27_internal_tensor_assign_2, var_5587, new_cache_29_internal_tensor_assign_2, var_5973, new_cache_31_internal_tensor_assign_2, var_6359, new_cache_33_internal_tensor_assign_2, var_6745, new_cache_35_internal_tensor_assign_2, var_7131, new_cache_37_internal_tensor_assign_2, var_7517, new_cache_39_internal_tensor_assign_2, var_7903, new_cache_41_internal_tensor_assign_2, var_8289, new_cache_43_internal_tensor_assign_2, var_8675, new_cache_45_internal_tensor_assign_2, var_9061, new_cache_internal_tensor_assign_2, var_9447); +} \ No newline at end of file diff --git a/v2.1/french_24l/flowlm_step.mlmodelc/weights/weight.bin b/v2.1/french_24l/flowlm_step.mlmodelc/weights/weight.bin new file mode 100644 index 0000000000000000000000000000000000000000..8ebff19aa2e04fd03eced8bf393064c8bc1bbb1c --- /dev/null +++ b/v2.1/french_24l/flowlm_step.mlmodelc/weights/weight.bin @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:2fa4d2a41cae9ffc4f8b00437d4b79f8e224144e976ff911c93abe83561514d3 +size 605330048 diff --git a/v2.1/french_24l/flowlm_step.mlpackage/Data/com.apple.CoreML/model.mlmodel b/v2.1/french_24l/flowlm_step.mlpackage/Data/com.apple.CoreML/model.mlmodel new file mode 100644 index 0000000000000000000000000000000000000000..53d94bd34cf8d4d90d46b323d196104206f9e17a --- /dev/null +++ b/v2.1/french_24l/flowlm_step.mlpackage/Data/com.apple.CoreML/model.mlmodel @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:390d877aef92a06187578fdacd0f1a28d4e0f7955d680429be1862f9f388ad1b +size 825108 diff --git a/v2.1/french_24l/flowlm_step.mlpackage/Data/com.apple.CoreML/weights/weight.bin b/v2.1/french_24l/flowlm_step.mlpackage/Data/com.apple.CoreML/weights/weight.bin new file mode 100644 index 0000000000000000000000000000000000000000..8ebff19aa2e04fd03eced8bf393064c8bc1bbb1c --- /dev/null +++ b/v2.1/french_24l/flowlm_step.mlpackage/Data/com.apple.CoreML/weights/weight.bin @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:2fa4d2a41cae9ffc4f8b00437d4b79f8e224144e976ff911c93abe83561514d3 +size 605330048 diff --git a/v2.1/french_24l/flowlm_step.mlpackage/Manifest.json b/v2.1/french_24l/flowlm_step.mlpackage/Manifest.json new file mode 100644 index 0000000000000000000000000000000000000000..a46003438116fe015b89507936e88d6009454173 --- /dev/null +++ b/v2.1/french_24l/flowlm_step.mlpackage/Manifest.json @@ -0,0 +1,18 @@ +{ + "fileFormatVersion": "1.0.0", + "itemInfoEntries": { + "18B268EF-BB43-4C12-9D86-95D927ED4EA2": { + "author": "com.apple.CoreML", + "description": "CoreML Model Weights", + "name": "weights", + "path": "com.apple.CoreML/weights" + }, + "3B05E3B5-7ADD-4DDE-82B2-07C458A4687A": { + "author": "com.apple.CoreML", + "description": "CoreML Model Specification", + "name": "model.mlmodel", + "path": "com.apple.CoreML/model.mlmodel" + } + }, + "rootModelIdentifier": "3B05E3B5-7ADD-4DDE-82B2-07C458A4687A" +} diff --git a/v2.1/french_24l/flowlm_stepv2.mlmodelc/analytics/coremldata.bin b/v2.1/french_24l/flowlm_stepv2.mlmodelc/analytics/coremldata.bin new file mode 100644 index 0000000000000000000000000000000000000000..d27ea490f41b845413534db869eab6fcaef53369 --- /dev/null +++ b/v2.1/french_24l/flowlm_stepv2.mlmodelc/analytics/coremldata.bin @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:d20222b9d25c929804d534f19b9f51daad549805cb010d9b5f0dd970272ef3f6 +size 243 diff --git a/v2.1/french_24l/flowlm_stepv2.mlmodelc/coremldata.bin b/v2.1/french_24l/flowlm_stepv2.mlmodelc/coremldata.bin new file mode 100644 index 0000000000000000000000000000000000000000..98660effe1f7c37734f5b9cc13e60ed7e219a1bf --- /dev/null +++ b/v2.1/french_24l/flowlm_stepv2.mlmodelc/coremldata.bin @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:4e5930318538862799470bd2cc790a6cc6079e9ba9d35249d74fc0850217b096 +size 3561 diff --git a/v2.1/french_24l/flowlm_stepv2.mlmodelc/metadata.json b/v2.1/french_24l/flowlm_stepv2.mlmodelc/metadata.json new file mode 100644 index 0000000000000000000000000000000000000000..629da9a69819375413a6c11a7789eb2ca4f04e28 --- /dev/null +++ b/v2.1/french_24l/flowlm_stepv2.mlmodelc/metadata.json @@ -0,0 +1,1068 @@ +[ + { + "metadataOutputVersion" : "3.0", + "storagePrecision" : "Mixed (Float16, Int8)", + "outputSchema" : [ + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1 × 1 × 1024)", + "shortDescription" : "", + "shape" : "[1, 1, 1024]", + "name" : "input", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1 × 1 × 1)", + "shortDescription" : "", + "shape" : "[1, 1, 1]", + "name" : "var_10268", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 2 × 1 × 512 × 16 × 64)", + "shortDescription" : "", + "shape" : "[2, 1, 512, 16, 64]", + "name" : "new_cache_1_internal_tensor_assign_2", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1)", + "shortDescription" : "", + "shape" : "[1]", + "name" : "var_602", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 2 × 1 × 512 × 16 × 64)", + "shortDescription" : "", + "shape" : "[2, 1, 512, 16, 64]", + "name" : "new_cache_3_internal_tensor_assign_2", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1)", + "shortDescription" : "", + "shape" : "[1]", + "name" : "var_1021", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 2 × 1 × 512 × 16 × 64)", + "shortDescription" : "", + "shape" : "[2, 1, 512, 16, 64]", + "name" : "new_cache_5_internal_tensor_assign_2", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1)", + "shortDescription" : "", + "shape" : "[1]", + "name" : "var_1440", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 2 × 1 × 512 × 16 × 64)", + "shortDescription" : "", + "shape" : "[2, 1, 512, 16, 64]", + "name" : "new_cache_7_internal_tensor_assign_2", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1)", + "shortDescription" : "", + "shape" : "[1]", + "name" : "var_1859", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 2 × 1 × 512 × 16 × 64)", + "shortDescription" : "", + "shape" : "[2, 1, 512, 16, 64]", + "name" : "new_cache_9_internal_tensor_assign_2", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1)", + "shortDescription" : "", + "shape" : "[1]", + "name" : "var_2278", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 2 × 1 × 512 × 16 × 64)", + "shortDescription" : "", + "shape" : "[2, 1, 512, 16, 64]", + "name" : "new_cache_11_internal_tensor_assign_2", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1)", + "shortDescription" : "", + "shape" : "[1]", + "name" : "var_2697", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 2 × 1 × 512 × 16 × 64)", + "shortDescription" : "", + "shape" : "[2, 1, 512, 16, 64]", + "name" : "new_cache_13_internal_tensor_assign_2", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1)", + "shortDescription" : "", + "shape" : "[1]", + "name" : "var_3116", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 2 × 1 × 512 × 16 × 64)", + "shortDescription" : "", + "shape" : "[2, 1, 512, 16, 64]", + "name" : "new_cache_15_internal_tensor_assign_2", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1)", + "shortDescription" : "", + "shape" : "[1]", + "name" : "var_3535", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 2 × 1 × 512 × 16 × 64)", + "shortDescription" : "", + "shape" : "[2, 1, 512, 16, 64]", + "name" : "new_cache_17_internal_tensor_assign_2", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1)", + "shortDescription" : "", + "shape" : "[1]", + "name" : "var_3954", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 2 × 1 × 512 × 16 × 64)", + "shortDescription" : "", + "shape" : "[2, 1, 512, 16, 64]", + "name" : "new_cache_19_internal_tensor_assign_2", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1)", + "shortDescription" : "", + "shape" : "[1]", + "name" : "var_4373", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 2 × 1 × 512 × 16 × 64)", + "shortDescription" : "", + "shape" : "[2, 1, 512, 16, 64]", + "name" : "new_cache_21_internal_tensor_assign_2", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1)", + "shortDescription" : "", + "shape" : "[1]", + "name" : "var_4792", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 2 × 1 × 512 × 16 × 64)", + "shortDescription" : "", + "shape" : "[2, 1, 512, 16, 64]", + "name" : "new_cache_23_internal_tensor_assign_2", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1)", + "shortDescription" : "", + "shape" : "[1]", + "name" : "var_5211", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 2 × 1 × 512 × 16 × 64)", + "shortDescription" : "", + "shape" : "[2, 1, 512, 16, 64]", + "name" : "new_cache_25_internal_tensor_assign_2", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1)", + "shortDescription" : "", + "shape" : "[1]", + "name" : "var_5630", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 2 × 1 × 512 × 16 × 64)", + "shortDescription" : "", + "shape" : "[2, 1, 512, 16, 64]", + "name" : "new_cache_27_internal_tensor_assign_2", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1)", + "shortDescription" : "", + "shape" : "[1]", + "name" : "var_6049", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 2 × 1 × 512 × 16 × 64)", + "shortDescription" : "", + "shape" : "[2, 1, 512, 16, 64]", + "name" : "new_cache_29_internal_tensor_assign_2", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1)", + "shortDescription" : "", + "shape" : "[1]", + "name" : "var_6468", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 2 × 1 × 512 × 16 × 64)", + "shortDescription" : "", + "shape" : "[2, 1, 512, 16, 64]", + "name" : "new_cache_31_internal_tensor_assign_2", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1)", + "shortDescription" : "", + "shape" : "[1]", + "name" : "var_6887", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 2 × 1 × 512 × 16 × 64)", + "shortDescription" : "", + "shape" : "[2, 1, 512, 16, 64]", + "name" : "new_cache_33_internal_tensor_assign_2", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1)", + "shortDescription" : "", + "shape" : "[1]", + "name" : "var_7306", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 2 × 1 × 512 × 16 × 64)", + "shortDescription" : "", + "shape" : "[2, 1, 512, 16, 64]", + "name" : "new_cache_35_internal_tensor_assign_2", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1)", + "shortDescription" : "", + "shape" : "[1]", + "name" : "var_7725", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 2 × 1 × 512 × 16 × 64)", + "shortDescription" : "", + "shape" : "[2, 1, 512, 16, 64]", + "name" : "new_cache_37_internal_tensor_assign_2", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1)", + "shortDescription" : "", + "shape" : "[1]", + "name" : "var_8144", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 2 × 1 × 512 × 16 × 64)", + "shortDescription" : "", + "shape" : "[2, 1, 512, 16, 64]", + "name" : "new_cache_39_internal_tensor_assign_2", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1)", + "shortDescription" : "", + "shape" : "[1]", + "name" : "var_8563", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 2 × 1 × 512 × 16 × 64)", + "shortDescription" : "", + "shape" : "[2, 1, 512, 16, 64]", + "name" : "new_cache_41_internal_tensor_assign_2", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1)", + "shortDescription" : "", + "shape" : "[1]", + "name" : "var_8982", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 2 × 1 × 512 × 16 × 64)", + "shortDescription" : "", + "shape" : "[2, 1, 512, 16, 64]", + "name" : "new_cache_43_internal_tensor_assign_2", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1)", + "shortDescription" : "", + "shape" : "[1]", + "name" : "var_9401", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 2 × 1 × 512 × 16 × 64)", + "shortDescription" : "", + "shape" : "[2, 1, 512, 16, 64]", + "name" : "new_cache_45_internal_tensor_assign_2", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1)", + "shortDescription" : "", + "shape" : "[1]", + "name" : "var_9820", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 2 × 1 × 512 × 16 × 64)", + "shortDescription" : "", + "shape" : "[2, 1, 512, 16, 64]", + "name" : "new_cache_internal_tensor_assign_2", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1)", + "shortDescription" : "", + "shape" : "[1]", + "name" : "var_10239", + "type" : "MultiArray" + } + ], + "modelParameters" : [ + + ], + "specificationVersion" : 8, + "mlProgramOperationTypeHistogram" : { + "Ios17.cos" : 24, + "Range1d" : 48, + "Ios17.notEqual" : 49, + "Ios17.floor" : 24, + "Ios17.reshape" : 432, + "Ios17.logicalAnd" : 24, + "Ios16.softmax" : 24, + "Ios17.scatter" : 48, + "Ios17.matmul" : 48, + "Ios17.sin" : 24, + "Ios17.transpose" : 96, + "Select" : 73, + "Ios17.expandDims" : 25, + "Ios17.add" : 144, + "Tile" : 24, + "Ios17.sliceByIndex" : 312, + "Ios17.lessEqual" : 24, + "Ios17.scatterAlongAxis" : 48, + "Ios17.logicalNot" : 24, + "Ios17.layerNorm" : 49, + "Ios16.constexprAffineDequantize" : 96, + "Ios17.cast" : 124, + "Ios17.sub" : 72, + "Ios17.less" : 24, + "Ios16.gelu" : 24, + "Ios17.linear" : 98, + "Stack" : 48, + "Ios17.mul" : 288 + }, + "computePrecision" : "Mixed (Float16, Float32, Int32)", + "isUpdatable" : "0", + "stateSchema" : [ + + ], + "availability" : { + "macOS" : "14.0", + "tvOS" : "17.0", + "visionOS" : "1.0", + "watchOS" : "10.0", + "iOS" : "17.0", + "macCatalyst" : "17.0" + }, + "modelType" : { + "name" : "MLModelType_mlProgram" + }, + "userDefinedMetadata" : { + "com.github.apple.coremltools.conversion_date" : "2026-04-27", + "com.github.apple.coremltools.source" : "torch==2.9.1", + "com.github.apple.coremltools.version" : "9.0", + "com.github.apple.coremltools.source_dialect" : "TorchScript" + }, + "inputSchema" : [ + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1 × 1 × 32)", + "shortDescription" : "", + "shape" : "[1, 1, 32]", + "name" : "sequence", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 32)", + "shortDescription" : "", + "shape" : "[32]", + "name" : "bos_emb", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 2 × 1 × 512 × 16 × 64)", + "shortDescription" : "", + "shape" : "[2, 1, 512, 16, 64]", + "name" : "cache0", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1)", + "shortDescription" : "", + "shape" : "[1]", + "name" : "position0", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 2 × 1 × 512 × 16 × 64)", + "shortDescription" : "", + "shape" : "[2, 1, 512, 16, 64]", + "name" : "cache1", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1)", + "shortDescription" : "", + "shape" : "[1]", + "name" : "position1", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 2 × 1 × 512 × 16 × 64)", + "shortDescription" : "", + "shape" : "[2, 1, 512, 16, 64]", + "name" : "cache2", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1)", + "shortDescription" : "", + "shape" : "[1]", + "name" : "position2", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 2 × 1 × 512 × 16 × 64)", + "shortDescription" : "", + "shape" : "[2, 1, 512, 16, 64]", + "name" : "cache3", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1)", + "shortDescription" : "", + "shape" : "[1]", + "name" : "position3", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 2 × 1 × 512 × 16 × 64)", + "shortDescription" : "", + "shape" : "[2, 1, 512, 16, 64]", + "name" : "cache4", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1)", + "shortDescription" : "", + "shape" : "[1]", + "name" : "position4", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 2 × 1 × 512 × 16 × 64)", + "shortDescription" : "", + "shape" : "[2, 1, 512, 16, 64]", + "name" : "cache5", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1)", + "shortDescription" : "", + "shape" : "[1]", + "name" : "position5", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 2 × 1 × 512 × 16 × 64)", + "shortDescription" : "", + "shape" : "[2, 1, 512, 16, 64]", + "name" : "cache6", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1)", + "shortDescription" : "", + "shape" : "[1]", + "name" : "position6", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 2 × 1 × 512 × 16 × 64)", + "shortDescription" : "", + "shape" : "[2, 1, 512, 16, 64]", + "name" : "cache7", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1)", + "shortDescription" : "", + "shape" : "[1]", + "name" : "position7", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 2 × 1 × 512 × 16 × 64)", + "shortDescription" : "", + "shape" : "[2, 1, 512, 16, 64]", + "name" : "cache8", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1)", + "shortDescription" : "", + "shape" : "[1]", + "name" : "position8", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 2 × 1 × 512 × 16 × 64)", + "shortDescription" : "", + "shape" : "[2, 1, 512, 16, 64]", + "name" : "cache9", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1)", + "shortDescription" : "", + "shape" : "[1]", + "name" : "position9", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 2 × 1 × 512 × 16 × 64)", + "shortDescription" : "", + "shape" : "[2, 1, 512, 16, 64]", + "name" : "cache10", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1)", + "shortDescription" : "", + "shape" : "[1]", + "name" : "position10", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 2 × 1 × 512 × 16 × 64)", + "shortDescription" : "", + "shape" : "[2, 1, 512, 16, 64]", + "name" : "cache11", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1)", + "shortDescription" : "", + "shape" : "[1]", + "name" : "position11", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 2 × 1 × 512 × 16 × 64)", + "shortDescription" : "", + "shape" : "[2, 1, 512, 16, 64]", + "name" : "cache12", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1)", + "shortDescription" : "", + "shape" : "[1]", + "name" : "position12", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 2 × 1 × 512 × 16 × 64)", + "shortDescription" : "", + "shape" : "[2, 1, 512, 16, 64]", + "name" : "cache13", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1)", + "shortDescription" : "", + "shape" : "[1]", + "name" : "position13", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 2 × 1 × 512 × 16 × 64)", + "shortDescription" : "", + "shape" : "[2, 1, 512, 16, 64]", + "name" : "cache14", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1)", + "shortDescription" : "", + "shape" : "[1]", + "name" : "position14", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 2 × 1 × 512 × 16 × 64)", + "shortDescription" : "", + "shape" : "[2, 1, 512, 16, 64]", + "name" : "cache15", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1)", + "shortDescription" : "", + "shape" : "[1]", + "name" : "position15", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 2 × 1 × 512 × 16 × 64)", + "shortDescription" : "", + "shape" : "[2, 1, 512, 16, 64]", + "name" : "cache16", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1)", + "shortDescription" : "", + "shape" : "[1]", + "name" : "position16", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 2 × 1 × 512 × 16 × 64)", + "shortDescription" : "", + "shape" : "[2, 1, 512, 16, 64]", + "name" : "cache17", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1)", + "shortDescription" : "", + "shape" : "[1]", + "name" : "position17", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 2 × 1 × 512 × 16 × 64)", + "shortDescription" : "", + "shape" : "[2, 1, 512, 16, 64]", + "name" : "cache18", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1)", + "shortDescription" : "", + "shape" : "[1]", + "name" : "position18", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 2 × 1 × 512 × 16 × 64)", + "shortDescription" : "", + "shape" : "[2, 1, 512, 16, 64]", + "name" : "cache19", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1)", + "shortDescription" : "", + "shape" : "[1]", + "name" : "position19", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 2 × 1 × 512 × 16 × 64)", + "shortDescription" : "", + "shape" : "[2, 1, 512, 16, 64]", + "name" : "cache20", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1)", + "shortDescription" : "", + "shape" : "[1]", + "name" : "position20", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 2 × 1 × 512 × 16 × 64)", + "shortDescription" : "", + "shape" : "[2, 1, 512, 16, 64]", + "name" : "cache21", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1)", + "shortDescription" : "", + "shape" : "[1]", + "name" : "position21", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 2 × 1 × 512 × 16 × 64)", + "shortDescription" : "", + "shape" : "[2, 1, 512, 16, 64]", + "name" : "cache22", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1)", + "shortDescription" : "", + "shape" : "[1]", + "name" : "position22", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 2 × 1 × 512 × 16 × 64)", + "shortDescription" : "", + "shape" : "[2, 1, 512, 16, 64]", + "name" : "cache23", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1)", + "shortDescription" : "", + "shape" : "[1]", + "name" : "position23", + "type" : "MultiArray" + } + ], + "generatedClassName" : "flowlm_stepv2", + "method" : "predict" + } +] \ No newline at end of file diff --git a/v2.1/french_24l/flowlm_stepv2.mlmodelc/model.mil b/v2.1/french_24l/flowlm_stepv2.mlmodelc/model.mil new file mode 100644 index 0000000000000000000000000000000000000000..abdebe0ef22e1c1977cf3ba08456089a8623718d --- /dev/null +++ b/v2.1/french_24l/flowlm_stepv2.mlmodelc/model.mil @@ -0,0 +1,5337 @@ +program(1.0) +[buildInfo = dict, tensor>({{"coremlc-component-MIL", "3520.4.1"}, {"coremlc-version", "3520.5.1"}, {"coremltools-component-torch", "2.9.1"}, {"coremltools-source-dialect", "TorchScript"}, {"coremltools-version", "9.0"}})] +{ + func main(tensor bos_emb, tensor cache0, tensor cache1, tensor cache10, tensor cache11, tensor cache12, tensor cache13, tensor cache14, tensor cache15, tensor cache16, tensor cache17, tensor cache18, tensor cache19, tensor cache2, tensor cache20, tensor cache21, tensor cache22, tensor cache23, tensor cache3, tensor cache4, tensor cache5, tensor cache6, tensor cache7, tensor cache8, tensor cache9, tensor position0, tensor position1, tensor position10, tensor position11, tensor position12, tensor position13, tensor position14, tensor position15, tensor position16, tensor position17, tensor position18, tensor position19, tensor position2, tensor position20, tensor position21, tensor position22, tensor position23, tensor position3, tensor position4, tensor position5, tensor position6, tensor position7, tensor position8, tensor position9, tensor sequence) { + tensor attn0_in_proj_weight_cast_fp16 = constexpr_affine_dequantize()[axis = tensor(0), name = tensor("attn0_in_proj_weight_cast_fp16"), quantized_data = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(64))), scale = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(3148992))), zero_point = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(3145856)))]; + tensor attn0_out_proj_weight_cast_fp16 = constexpr_affine_dequantize()[axis = tensor(0), name = tensor("attn0_out_proj_weight_cast_fp16"), quantized_data = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(3155200))), scale = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(4204928))), zero_point = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(4203840)))]; + tensor linear0_1_weight_cast_fp16 = constexpr_affine_dequantize()[axis = tensor(0), name = tensor("linear0_1_weight_cast_fp16"), quantized_data = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(4207040))), scale = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8405568))), zero_point = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8401408)))]; + tensor linear0_2_weight_cast_fp16 = constexpr_affine_dequantize()[axis = tensor(0), name = tensor("linear0_2_weight_cast_fp16"), quantized_data = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8413824))), scale = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(12608192))), zero_point = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(4203840)))]; + tensor attn1_in_proj_weight_cast_fp16 = constexpr_affine_dequantize()[axis = tensor(0), name = tensor("attn1_in_proj_weight_cast_fp16"), quantized_data = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(12610304))), scale = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15756096))), zero_point = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(3145856)))]; + tensor attn1_out_proj_weight_cast_fp16 = constexpr_affine_dequantize()[axis = tensor(0), name = tensor("attn1_out_proj_weight_cast_fp16"), quantized_data = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15762304))), scale = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(16810944))), zero_point = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(4203840)))]; + tensor linear1_1_weight_cast_fp16 = constexpr_affine_dequantize()[axis = tensor(0), name = tensor("linear1_1_weight_cast_fp16"), quantized_data = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(16813056))), scale = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21007424))), zero_point = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8401408)))]; + tensor linear1_2_weight_cast_fp16 = constexpr_affine_dequantize()[axis = tensor(0), name = tensor("linear1_2_weight_cast_fp16"), quantized_data = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21015680))), scale = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(25210048))), zero_point = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(4203840)))]; + tensor attn2_in_proj_weight_cast_fp16 = constexpr_affine_dequantize()[axis = tensor(0), name = tensor("attn2_in_proj_weight_cast_fp16"), quantized_data = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(25212160))), scale = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28357952))), zero_point = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(3145856)))]; + tensor attn2_out_proj_weight_cast_fp16 = constexpr_affine_dequantize()[axis = tensor(0), name = tensor("attn2_out_proj_weight_cast_fp16"), quantized_data = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28364160))), scale = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(29412800))), zero_point = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(4203840)))]; + tensor linear2_1_weight_cast_fp16 = constexpr_affine_dequantize()[axis = tensor(0), name = tensor("linear2_1_weight_cast_fp16"), quantized_data = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(29414912))), scale = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(33609280))), zero_point = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8401408)))]; + tensor linear2_2_weight_cast_fp16 = constexpr_affine_dequantize()[axis = tensor(0), name = tensor("linear2_2_weight_cast_fp16"), quantized_data = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(33617536))), scale = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(37811904))), zero_point = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(4203840)))]; + tensor attn3_in_proj_weight_cast_fp16 = constexpr_affine_dequantize()[axis = tensor(0), name = tensor("attn3_in_proj_weight_cast_fp16"), quantized_data = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(37814016))), scale = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(40959808))), zero_point = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(3145856)))]; + tensor attn3_out_proj_weight_cast_fp16 = constexpr_affine_dequantize()[axis = tensor(0), name = tensor("attn3_out_proj_weight_cast_fp16"), quantized_data = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(40966016))), scale = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(42014656))), zero_point = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(4203840)))]; + tensor linear3_1_weight_cast_fp16 = constexpr_affine_dequantize()[axis = tensor(0), name = tensor("linear3_1_weight_cast_fp16"), quantized_data = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(42016768))), scale = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(46211136))), zero_point = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8401408)))]; + tensor linear3_2_weight_cast_fp16 = constexpr_affine_dequantize()[axis = tensor(0), name = tensor("linear3_2_weight_cast_fp16"), quantized_data = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(46219392))), scale = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(50413760))), zero_point = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(4203840)))]; + tensor attn4_in_proj_weight_cast_fp16 = constexpr_affine_dequantize()[axis = tensor(0), name = tensor("attn4_in_proj_weight_cast_fp16"), quantized_data = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(50415872))), scale = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(53561664))), zero_point = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(3145856)))]; + tensor attn4_out_proj_weight_cast_fp16 = constexpr_affine_dequantize()[axis = tensor(0), name = tensor("attn4_out_proj_weight_cast_fp16"), quantized_data = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(53567872))), scale = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(54616512))), zero_point = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(4203840)))]; + tensor linear4_1_weight_cast_fp16 = constexpr_affine_dequantize()[axis = tensor(0), name = tensor("linear4_1_weight_cast_fp16"), quantized_data = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(54618624))), scale = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(58812992))), zero_point = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8401408)))]; + tensor linear4_2_weight_cast_fp16 = constexpr_affine_dequantize()[axis = tensor(0), name = tensor("linear4_2_weight_cast_fp16"), quantized_data = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(58821248))), scale = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(63015616))), zero_point = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(4203840)))]; + tensor attn5_in_proj_weight_cast_fp16 = constexpr_affine_dequantize()[axis = tensor(0), name = tensor("attn5_in_proj_weight_cast_fp16"), quantized_data = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(63017728))), scale = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(66163520))), zero_point = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(3145856)))]; + tensor attn5_out_proj_weight_cast_fp16 = constexpr_affine_dequantize()[axis = tensor(0), name = tensor("attn5_out_proj_weight_cast_fp16"), quantized_data = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(66169728))), scale = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(67218368))), zero_point = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(4203840)))]; + tensor linear5_1_weight_cast_fp16 = constexpr_affine_dequantize()[axis = tensor(0), name = tensor("linear5_1_weight_cast_fp16"), quantized_data = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(67220480))), scale = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(71414848))), zero_point = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8401408)))]; + tensor linear5_2_weight_cast_fp16 = constexpr_affine_dequantize()[axis = tensor(0), name = tensor("linear5_2_weight_cast_fp16"), quantized_data = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(71423104))), scale = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(75617472))), zero_point = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(4203840)))]; + tensor attn6_in_proj_weight_cast_fp16 = constexpr_affine_dequantize()[axis = tensor(0), name = tensor("attn6_in_proj_weight_cast_fp16"), quantized_data = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(75619584))), scale = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(78765376))), zero_point = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(3145856)))]; + tensor attn6_out_proj_weight_cast_fp16 = constexpr_affine_dequantize()[axis = tensor(0), name = tensor("attn6_out_proj_weight_cast_fp16"), quantized_data = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(78771584))), scale = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(79820224))), zero_point = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(4203840)))]; + tensor linear6_1_weight_cast_fp16 = constexpr_affine_dequantize()[axis = tensor(0), name = tensor("linear6_1_weight_cast_fp16"), quantized_data = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(79822336))), scale = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(84016704))), zero_point = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8401408)))]; + tensor linear6_2_weight_cast_fp16 = constexpr_affine_dequantize()[axis = tensor(0), name = tensor("linear6_2_weight_cast_fp16"), quantized_data = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(84024960))), scale = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(88219328))), zero_point = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(4203840)))]; + tensor attn7_in_proj_weight_cast_fp16 = constexpr_affine_dequantize()[axis = tensor(0), name = tensor("attn7_in_proj_weight_cast_fp16"), quantized_data = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(88221440))), scale = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(91367232))), zero_point = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(3145856)))]; + tensor attn7_out_proj_weight_cast_fp16 = constexpr_affine_dequantize()[axis = tensor(0), name = tensor("attn7_out_proj_weight_cast_fp16"), quantized_data = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(91373440))), scale = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(92422080))), zero_point = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(4203840)))]; + tensor linear7_1_weight_cast_fp16 = constexpr_affine_dequantize()[axis = tensor(0), name = tensor("linear7_1_weight_cast_fp16"), quantized_data = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(92424192))), scale = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(96618560))), zero_point = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8401408)))]; + tensor linear7_2_weight_cast_fp16 = constexpr_affine_dequantize()[axis = tensor(0), name = tensor("linear7_2_weight_cast_fp16"), quantized_data = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(96626816))), scale = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(100821184))), zero_point = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(4203840)))]; + tensor attn8_in_proj_weight_cast_fp16 = constexpr_affine_dequantize()[axis = tensor(0), name = tensor("attn8_in_proj_weight_cast_fp16"), quantized_data = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(100823296))), scale = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(103969088))), zero_point = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(3145856)))]; + tensor attn8_out_proj_weight_cast_fp16 = constexpr_affine_dequantize()[axis = tensor(0), name = tensor("attn8_out_proj_weight_cast_fp16"), quantized_data = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(103975296))), scale = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(105023936))), zero_point = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(4203840)))]; + tensor linear8_1_weight_cast_fp16 = constexpr_affine_dequantize()[axis = tensor(0), name = tensor("linear8_1_weight_cast_fp16"), quantized_data = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(105026048))), scale = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(109220416))), zero_point = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8401408)))]; + tensor linear8_2_weight_cast_fp16 = constexpr_affine_dequantize()[axis = tensor(0), name = tensor("linear8_2_weight_cast_fp16"), quantized_data = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(109228672))), scale = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(113423040))), zero_point = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(4203840)))]; + tensor attn9_in_proj_weight_cast_fp16 = constexpr_affine_dequantize()[axis = tensor(0), name = tensor("attn9_in_proj_weight_cast_fp16"), quantized_data = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(113425152))), scale = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(116570944))), zero_point = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(3145856)))]; + tensor attn9_out_proj_weight_cast_fp16 = constexpr_affine_dequantize()[axis = tensor(0), name = tensor("attn9_out_proj_weight_cast_fp16"), quantized_data = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(116577152))), scale = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(117625792))), zero_point = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(4203840)))]; + tensor linear9_1_weight_cast_fp16 = constexpr_affine_dequantize()[axis = tensor(0), name = tensor("linear9_1_weight_cast_fp16"), quantized_data = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(117627904))), scale = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(121822272))), zero_point = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8401408)))]; + tensor linear9_2_weight_cast_fp16 = constexpr_affine_dequantize()[axis = tensor(0), name = tensor("linear9_2_weight_cast_fp16"), quantized_data = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(121830528))), scale = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(126024896))), zero_point = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(4203840)))]; + tensor attn10_in_proj_weight_cast_fp16 = constexpr_affine_dequantize()[axis = tensor(0), name = tensor("attn10_in_proj_weight_cast_fp16"), quantized_data = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(126027008))), scale = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(129172800))), zero_point = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(3145856)))]; + tensor attn10_out_proj_weight_cast_fp16 = constexpr_affine_dequantize()[axis = tensor(0), name = tensor("attn10_out_proj_weight_cast_fp16"), quantized_data = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(129179008))), scale = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(130227648))), zero_point = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(4203840)))]; + tensor linear10_1_weight_cast_fp16 = constexpr_affine_dequantize()[axis = tensor(0), name = tensor("linear10_1_weight_cast_fp16"), quantized_data = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(130229760))), scale = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(134424128))), zero_point = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8401408)))]; + tensor linear10_2_weight_cast_fp16 = constexpr_affine_dequantize()[axis = tensor(0), name = tensor("linear10_2_weight_cast_fp16"), quantized_data = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(134432384))), scale = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(138626752))), zero_point = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(4203840)))]; + tensor attn11_in_proj_weight_cast_fp16 = constexpr_affine_dequantize()[axis = tensor(0), name = tensor("attn11_in_proj_weight_cast_fp16"), quantized_data = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(138628864))), scale = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(141774656))), zero_point = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(3145856)))]; + tensor attn11_out_proj_weight_cast_fp16 = constexpr_affine_dequantize()[axis = tensor(0), name = tensor("attn11_out_proj_weight_cast_fp16"), quantized_data = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(141780864))), scale = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(142829504))), zero_point = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(4203840)))]; + tensor linear11_1_weight_cast_fp16 = constexpr_affine_dequantize()[axis = tensor(0), name = tensor("linear11_1_weight_cast_fp16"), quantized_data = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(142831616))), scale = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(147025984))), zero_point = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8401408)))]; + tensor linear11_2_weight_cast_fp16 = constexpr_affine_dequantize()[axis = tensor(0), name = tensor("linear11_2_weight_cast_fp16"), quantized_data = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(147034240))), scale = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(151228608))), zero_point = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(4203840)))]; + tensor attn12_in_proj_weight_cast_fp16 = constexpr_affine_dequantize()[axis = tensor(0), name = tensor("attn12_in_proj_weight_cast_fp16"), quantized_data = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(151230720))), scale = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(154376512))), zero_point = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(3145856)))]; + tensor attn12_out_proj_weight_cast_fp16 = constexpr_affine_dequantize()[axis = tensor(0), name = tensor("attn12_out_proj_weight_cast_fp16"), quantized_data = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(154382720))), scale = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(155431360))), zero_point = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(4203840)))]; + tensor linear12_1_weight_cast_fp16 = constexpr_affine_dequantize()[axis = tensor(0), name = tensor("linear12_1_weight_cast_fp16"), quantized_data = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(155433472))), scale = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(159627840))), zero_point = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8401408)))]; + tensor linear12_2_weight_cast_fp16 = constexpr_affine_dequantize()[axis = tensor(0), name = tensor("linear12_2_weight_cast_fp16"), quantized_data = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(159636096))), scale = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(163830464))), zero_point = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(4203840)))]; + tensor attn13_in_proj_weight_cast_fp16 = constexpr_affine_dequantize()[axis = tensor(0), name = tensor("attn13_in_proj_weight_cast_fp16"), quantized_data = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(163832576))), scale = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(166978368))), zero_point = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(3145856)))]; + tensor attn13_out_proj_weight_cast_fp16 = constexpr_affine_dequantize()[axis = tensor(0), name = tensor("attn13_out_proj_weight_cast_fp16"), quantized_data = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(166984576))), scale = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(168033216))), zero_point = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(4203840)))]; + tensor linear13_1_weight_cast_fp16 = constexpr_affine_dequantize()[axis = tensor(0), name = tensor("linear13_1_weight_cast_fp16"), quantized_data = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(168035328))), scale = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(172229696))), zero_point = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8401408)))]; + tensor linear13_2_weight_cast_fp16 = constexpr_affine_dequantize()[axis = tensor(0), name = tensor("linear13_2_weight_cast_fp16"), quantized_data = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(172237952))), scale = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(176432320))), zero_point = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(4203840)))]; + tensor attn14_in_proj_weight_cast_fp16 = constexpr_affine_dequantize()[axis = tensor(0), name = tensor("attn14_in_proj_weight_cast_fp16"), quantized_data = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(176434432))), scale = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(179580224))), zero_point = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(3145856)))]; + tensor attn14_out_proj_weight_cast_fp16 = constexpr_affine_dequantize()[axis = tensor(0), name = tensor("attn14_out_proj_weight_cast_fp16"), quantized_data = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(179586432))), scale = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(180635072))), zero_point = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(4203840)))]; + tensor linear14_1_weight_cast_fp16 = constexpr_affine_dequantize()[axis = tensor(0), name = tensor("linear14_1_weight_cast_fp16"), quantized_data = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(180637184))), scale = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(184831552))), zero_point = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8401408)))]; + tensor linear14_2_weight_cast_fp16 = constexpr_affine_dequantize()[axis = tensor(0), name = tensor("linear14_2_weight_cast_fp16"), quantized_data = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(184839808))), scale = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(189034176))), zero_point = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(4203840)))]; + tensor attn15_in_proj_weight_cast_fp16 = constexpr_affine_dequantize()[axis = tensor(0), name = tensor("attn15_in_proj_weight_cast_fp16"), quantized_data = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(189036288))), scale = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(192182080))), zero_point = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(3145856)))]; + tensor attn15_out_proj_weight_cast_fp16 = constexpr_affine_dequantize()[axis = tensor(0), name = tensor("attn15_out_proj_weight_cast_fp16"), quantized_data = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(192188288))), scale = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(193236928))), zero_point = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(4203840)))]; + tensor linear15_1_weight_cast_fp16 = constexpr_affine_dequantize()[axis = tensor(0), name = tensor("linear15_1_weight_cast_fp16"), quantized_data = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(193239040))), scale = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(197433408))), zero_point = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8401408)))]; + tensor linear15_2_weight_cast_fp16 = constexpr_affine_dequantize()[axis = tensor(0), name = tensor("linear15_2_weight_cast_fp16"), quantized_data = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(197441664))), scale = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(201636032))), zero_point = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(4203840)))]; + tensor attn16_in_proj_weight_cast_fp16 = constexpr_affine_dequantize()[axis = tensor(0), name = tensor("attn16_in_proj_weight_cast_fp16"), quantized_data = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(201638144))), scale = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(204783936))), zero_point = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(3145856)))]; + tensor attn16_out_proj_weight_cast_fp16 = constexpr_affine_dequantize()[axis = tensor(0), name = tensor("attn16_out_proj_weight_cast_fp16"), quantized_data = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(204790144))), scale = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(205838784))), zero_point = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(4203840)))]; + tensor linear16_1_weight_cast_fp16 = constexpr_affine_dequantize()[axis = tensor(0), name = tensor("linear16_1_weight_cast_fp16"), quantized_data = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(205840896))), scale = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(210035264))), zero_point = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8401408)))]; + tensor linear16_2_weight_cast_fp16 = constexpr_affine_dequantize()[axis = tensor(0), name = tensor("linear16_2_weight_cast_fp16"), quantized_data = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(210043520))), scale = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(214237888))), zero_point = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(4203840)))]; + tensor attn17_in_proj_weight_cast_fp16 = constexpr_affine_dequantize()[axis = tensor(0), name = tensor("attn17_in_proj_weight_cast_fp16"), quantized_data = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(214240000))), scale = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(217385792))), zero_point = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(3145856)))]; + tensor attn17_out_proj_weight_cast_fp16 = constexpr_affine_dequantize()[axis = tensor(0), name = tensor("attn17_out_proj_weight_cast_fp16"), quantized_data = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(217392000))), scale = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(218440640))), zero_point = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(4203840)))]; + tensor linear17_1_weight_cast_fp16 = constexpr_affine_dequantize()[axis = tensor(0), name = tensor("linear17_1_weight_cast_fp16"), quantized_data = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(218442752))), scale = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(222637120))), zero_point = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8401408)))]; + tensor linear17_2_weight_cast_fp16 = constexpr_affine_dequantize()[axis = tensor(0), name = tensor("linear17_2_weight_cast_fp16"), quantized_data = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(222645376))), scale = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(226839744))), zero_point = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(4203840)))]; + tensor attn18_in_proj_weight_cast_fp16 = constexpr_affine_dequantize()[axis = tensor(0), name = tensor("attn18_in_proj_weight_cast_fp16"), quantized_data = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(226841856))), scale = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(229987648))), zero_point = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(3145856)))]; + tensor attn18_out_proj_weight_cast_fp16 = constexpr_affine_dequantize()[axis = tensor(0), name = tensor("attn18_out_proj_weight_cast_fp16"), quantized_data = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(229993856))), scale = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(231042496))), zero_point = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(4203840)))]; + tensor linear18_1_weight_cast_fp16 = constexpr_affine_dequantize()[axis = tensor(0), name = tensor("linear18_1_weight_cast_fp16"), quantized_data = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(231044608))), scale = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(235238976))), zero_point = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8401408)))]; + tensor linear18_2_weight_cast_fp16 = constexpr_affine_dequantize()[axis = tensor(0), name = tensor("linear18_2_weight_cast_fp16"), quantized_data = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(235247232))), scale = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(239441600))), zero_point = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(4203840)))]; + tensor attn19_in_proj_weight_cast_fp16 = constexpr_affine_dequantize()[axis = tensor(0), name = tensor("attn19_in_proj_weight_cast_fp16"), quantized_data = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(239443712))), scale = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(242589504))), zero_point = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(3145856)))]; + tensor attn19_out_proj_weight_cast_fp16 = constexpr_affine_dequantize()[axis = tensor(0), name = tensor("attn19_out_proj_weight_cast_fp16"), quantized_data = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(242595712))), scale = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(243644352))), zero_point = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(4203840)))]; + tensor linear19_1_weight_cast_fp16 = constexpr_affine_dequantize()[axis = tensor(0), name = tensor("linear19_1_weight_cast_fp16"), quantized_data = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(243646464))), scale = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(247840832))), zero_point = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8401408)))]; + tensor linear19_2_weight_cast_fp16 = constexpr_affine_dequantize()[axis = tensor(0), name = tensor("linear19_2_weight_cast_fp16"), quantized_data = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(247849088))), scale = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(252043456))), zero_point = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(4203840)))]; + tensor attn20_in_proj_weight_cast_fp16 = constexpr_affine_dequantize()[axis = tensor(0), name = tensor("attn20_in_proj_weight_cast_fp16"), quantized_data = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(252045568))), scale = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(255191360))), zero_point = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(3145856)))]; + tensor attn20_out_proj_weight_cast_fp16 = constexpr_affine_dequantize()[axis = tensor(0), name = tensor("attn20_out_proj_weight_cast_fp16"), quantized_data = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(255197568))), scale = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(256246208))), zero_point = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(4203840)))]; + tensor linear20_1_weight_cast_fp16 = constexpr_affine_dequantize()[axis = tensor(0), name = tensor("linear20_1_weight_cast_fp16"), quantized_data = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(256248320))), scale = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(260442688))), zero_point = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8401408)))]; + tensor linear20_2_weight_cast_fp16 = constexpr_affine_dequantize()[axis = tensor(0), name = tensor("linear20_2_weight_cast_fp16"), quantized_data = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(260450944))), scale = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(264645312))), zero_point = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(4203840)))]; + tensor attn21_in_proj_weight_cast_fp16 = constexpr_affine_dequantize()[axis = tensor(0), name = tensor("attn21_in_proj_weight_cast_fp16"), quantized_data = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(264647424))), scale = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(267793216))), zero_point = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(3145856)))]; + tensor attn21_out_proj_weight_cast_fp16 = constexpr_affine_dequantize()[axis = tensor(0), name = tensor("attn21_out_proj_weight_cast_fp16"), quantized_data = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(267799424))), scale = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(268848064))), zero_point = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(4203840)))]; + tensor linear21_1_weight_cast_fp16 = constexpr_affine_dequantize()[axis = tensor(0), name = tensor("linear21_1_weight_cast_fp16"), quantized_data = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(268850176))), scale = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(273044544))), zero_point = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8401408)))]; + tensor linear21_2_weight_cast_fp16 = constexpr_affine_dequantize()[axis = tensor(0), name = tensor("linear21_2_weight_cast_fp16"), quantized_data = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(273052800))), scale = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(277247168))), zero_point = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(4203840)))]; + tensor attn22_in_proj_weight_cast_fp16 = constexpr_affine_dequantize()[axis = tensor(0), name = tensor("attn22_in_proj_weight_cast_fp16"), quantized_data = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(277249280))), scale = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(280395072))), zero_point = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(3145856)))]; + tensor attn22_out_proj_weight_cast_fp16 = constexpr_affine_dequantize()[axis = tensor(0), name = tensor("attn22_out_proj_weight_cast_fp16"), quantized_data = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(280401280))), scale = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(281449920))), zero_point = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(4203840)))]; + tensor linear22_1_weight_cast_fp16 = constexpr_affine_dequantize()[axis = tensor(0), name = tensor("linear22_1_weight_cast_fp16"), quantized_data = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(281452032))), scale = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(285646400))), zero_point = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8401408)))]; + tensor linear22_2_weight_cast_fp16 = constexpr_affine_dequantize()[axis = tensor(0), name = tensor("linear22_2_weight_cast_fp16"), quantized_data = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(285654656))), scale = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(289849024))), zero_point = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(4203840)))]; + tensor attn23_in_proj_weight_cast_fp16 = constexpr_affine_dequantize()[axis = tensor(0), name = tensor("attn23_in_proj_weight_cast_fp16"), quantized_data = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(289851136))), scale = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(292996928))), zero_point = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(3145856)))]; + tensor attn23_out_proj_weight_cast_fp16 = constexpr_affine_dequantize()[axis = tensor(0), name = tensor("attn23_out_proj_weight_cast_fp16"), quantized_data = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(293003136))), scale = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(294051776))), zero_point = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(4203840)))]; + tensor linear23_1_weight_cast_fp16 = constexpr_affine_dequantize()[axis = tensor(0), name = tensor("linear23_1_weight_cast_fp16"), quantized_data = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(294053888))), scale = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(298248256))), zero_point = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8401408)))]; + tensor linear23_2_weight_cast_fp16 = constexpr_affine_dequantize()[axis = tensor(0), name = tensor("linear23_2_weight_cast_fp16"), quantized_data = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(298256512))), scale = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(302450880))), zero_point = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(4203840)))]; + tensor sequence_to_fp16_dtype_0 = const()[name = tensor("sequence_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor sequence_to_fp16 = cast(dtype = sequence_to_fp16_dtype_0, x = sequence)[name = tensor("cast_555")]; + tensor var_198_cast_fp16 = not_equal(x = sequence_to_fp16, y = sequence_to_fp16)[name = tensor("op_198_cast_fp16")]; + tensor expand_dims_0_axes_0 = const()[name = tensor("expand_dims_0_axes_0"), val = tensor([0, 1])]; + tensor bos_emb_to_fp16_dtype_0 = const()[name = tensor("bos_emb_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor bos_emb_to_fp16 = cast(dtype = bos_emb_to_fp16_dtype_0, x = bos_emb)[name = tensor("cast_554")]; + tensor expand_dims_0_cast_fp16 = expand_dims(axes = expand_dims_0_axes_0, x = bos_emb_to_fp16)[name = tensor("expand_dims_0_cast_fp16")]; + tensor input_1_cast_fp16 = select(a = expand_dims_0_cast_fp16, b = sequence_to_fp16, cond = var_198_cast_fp16)[name = tensor("input_1_cast_fp16")]; + tensor input_linear_weight_to_fp16 = const()[name = tensor("input_linear_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(302452992)))]; + tensor linear_0_bias_0_to_fp16 = const()[name = tensor("linear_0_bias_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(302518592)))]; + tensor linear_0_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = input_linear_weight_to_fp16, x = input_1_cast_fp16)[name = tensor("linear_0_cast_fp16")]; + tensor x_1_axes_0 = const()[name = tensor("x_1_axes_0"), val = tensor([-1])]; + tensor norm0_1_weight_to_fp16 = const()[name = tensor("norm0_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(302520704)))]; + tensor norm0_1_bias_to_fp16 = const()[name = tensor("norm0_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(302522816)))]; + tensor var_204_to_fp16 = const()[name = tensor("op_204_to_fp16"), val = tensor(0x1.5p-17)]; + tensor x_1_cast_fp16 = layer_norm(axes = x_1_axes_0, beta = norm0_1_bias_to_fp16, epsilon = var_204_to_fp16, gamma = norm0_1_weight_to_fp16, x = linear_0_cast_fp16)[name = tensor("x_1_cast_fp16")]; + tensor linear_1_bias_0_to_fp16 = const()[name = tensor("linear_1_bias_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(302524928)))]; + tensor linear_1_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = attn0_in_proj_weight_cast_fp16, x = x_1_cast_fp16)[name = tensor("linear_1_cast_fp16")]; + tensor var_240 = const()[name = tensor("op_240"), val = tensor([1, 1, 3, 16, 64])]; + tensor qkv_1_cast_fp16 = reshape(shape = var_240, x = linear_1_cast_fp16)[name = tensor("qkv_1_cast_fp16")]; + tensor q_1_begin_0 = const()[name = tensor("q_1_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor q_1_end_0 = const()[name = tensor("q_1_end_0"), val = tensor([1, 1, 1, 16, 64])]; + tensor q_1_end_mask_0 = const()[name = tensor("q_1_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor q_1_squeeze_mask_0 = const()[name = tensor("q_1_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor q_1_cast_fp16 = slice_by_index(begin = q_1_begin_0, end = q_1_end_0, end_mask = q_1_end_mask_0, squeeze_mask = q_1_squeeze_mask_0, x = qkv_1_cast_fp16)[name = tensor("q_1_cast_fp16")]; + tensor k_1_begin_0 = const()[name = tensor("k_1_begin_0"), val = tensor([0, 0, 1, 0, 0])]; + tensor k_1_end_0 = const()[name = tensor("k_1_end_0"), val = tensor([1, 1, 2, 16, 64])]; + tensor k_1_end_mask_0 = const()[name = tensor("k_1_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor k_1_squeeze_mask_0 = const()[name = tensor("k_1_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor k_1_cast_fp16 = slice_by_index(begin = k_1_begin_0, end = k_1_end_0, end_mask = k_1_end_mask_0, squeeze_mask = k_1_squeeze_mask_0, x = qkv_1_cast_fp16)[name = tensor("k_1_cast_fp16")]; + tensor v_1_begin_0 = const()[name = tensor("v_1_begin_0"), val = tensor([0, 0, 2, 0, 0])]; + tensor v_1_end_0 = const()[name = tensor("v_1_end_0"), val = tensor([1, 1, 3, 16, 64])]; + tensor v_1_end_mask_0 = const()[name = tensor("v_1_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor v_1_squeeze_mask_0 = const()[name = tensor("v_1_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor v_1_cast_fp16 = slice_by_index(begin = v_1_begin_0, end = v_1_end_0, end_mask = v_1_end_mask_0, squeeze_mask = v_1_squeeze_mask_0, x = qkv_1_cast_fp16)[name = tensor("v_1_cast_fp16")]; + tensor var_344 = const()[name = tensor("op_344"), val = tensor([1, 1, 1, 1])]; + tensor position0_to_fp16_dtype_0 = const()[name = tensor("position0_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor position0_to_fp16 = cast(dtype = position0_to_fp16_dtype_0, x = position0)[name = tensor("cast_553")]; + tensor ts_5_cast_fp16 = reshape(shape = var_344, x = position0_to_fp16)[name = tensor("ts_5_cast_fp16")]; + tensor var_348 = const()[name = tensor("op_348"), val = tensor([1, 1, 16, 32, 2])]; + tensor q_complex_1_cast_fp16 = reshape(shape = var_348, x = q_1_cast_fp16)[name = tensor("q_complex_1_cast_fp16")]; + tensor var_352 = const()[name = tensor("op_352"), val = tensor([1, 1, 16, 32, 2])]; + tensor k_complex_1_cast_fp16 = reshape(shape = var_352, x = k_1_cast_fp16)[name = tensor("k_complex_1_cast_fp16")]; + tensor var_356_begin_0 = const()[name = tensor("op_356_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_356_end_0 = const()[name = tensor("op_356_end_0"), val = tensor([1, 1, 16, 32, 1])]; + tensor var_356_end_mask_0 = const()[name = tensor("op_356_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_356_squeeze_mask_0 = const()[name = tensor("op_356_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_356_cast_fp16 = slice_by_index(begin = var_356_begin_0, end = var_356_end_0, end_mask = var_356_end_mask_0, squeeze_mask = var_356_squeeze_mask_0, x = q_complex_1_cast_fp16)[name = tensor("op_356_cast_fp16")]; + tensor var_364_begin_0 = const()[name = tensor("op_364_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_364_end_0 = const()[name = tensor("op_364_end_0"), val = tensor([1, 1, 16, 32, 2])]; + tensor var_364_end_mask_0 = const()[name = tensor("op_364_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_364_squeeze_mask_0 = const()[name = tensor("op_364_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_364_cast_fp16 = slice_by_index(begin = var_364_begin_0, end = var_364_end_0, end_mask = var_364_end_mask_0, squeeze_mask = var_364_squeeze_mask_0, x = q_complex_1_cast_fp16)[name = tensor("op_364_cast_fp16")]; + tensor var_372_begin_0 = const()[name = tensor("op_372_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_372_end_0 = const()[name = tensor("op_372_end_0"), val = tensor([1, 1, 16, 32, 1])]; + tensor var_372_end_mask_0 = const()[name = tensor("op_372_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_372_squeeze_mask_0 = const()[name = tensor("op_372_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_372_cast_fp16 = slice_by_index(begin = var_372_begin_0, end = var_372_end_0, end_mask = var_372_end_mask_0, squeeze_mask = var_372_squeeze_mask_0, x = k_complex_1_cast_fp16)[name = tensor("op_372_cast_fp16")]; + tensor var_380_begin_0 = const()[name = tensor("op_380_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_380_end_0 = const()[name = tensor("op_380_end_0"), val = tensor([1, 1, 16, 32, 2])]; + tensor var_380_end_mask_0 = const()[name = tensor("op_380_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_380_squeeze_mask_0 = const()[name = tensor("op_380_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_380_cast_fp16 = slice_by_index(begin = var_380_begin_0, end = var_380_end_0, end_mask = var_380_end_mask_0, squeeze_mask = var_380_squeeze_mask_0, x = k_complex_1_cast_fp16)[name = tensor("op_380_cast_fp16")]; + tensor freqs_1_to_fp16 = const()[name = tensor("freqs_1_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(302531136)))]; + tensor var_386_cast_fp16 = mul(x = freqs_1_to_fp16, y = ts_5_cast_fp16)[name = tensor("op_386_cast_fp16")]; + tensor rotr_1_cast_fp16 = cos(x = var_386_cast_fp16)[name = tensor("rotr_1_cast_fp16")]; + tensor roti_1_cast_fp16 = sin(x = var_386_cast_fp16)[name = tensor("roti_1_cast_fp16")]; + tensor var_390_cast_fp16 = mul(x = var_356_cast_fp16, y = rotr_1_cast_fp16)[name = tensor("op_390_cast_fp16")]; + tensor var_391_cast_fp16 = mul(x = var_364_cast_fp16, y = roti_1_cast_fp16)[name = tensor("op_391_cast_fp16")]; + tensor qor_1_cast_fp16 = sub(x = var_390_cast_fp16, y = var_391_cast_fp16)[name = tensor("qor_1_cast_fp16")]; + tensor var_394_cast_fp16 = mul(x = var_356_cast_fp16, y = roti_1_cast_fp16)[name = tensor("op_394_cast_fp16")]; + tensor var_395_cast_fp16 = mul(x = var_364_cast_fp16, y = rotr_1_cast_fp16)[name = tensor("op_395_cast_fp16")]; + tensor qoi_1_cast_fp16 = add(x = var_394_cast_fp16, y = var_395_cast_fp16)[name = tensor("qoi_1_cast_fp16")]; + tensor var_398_cast_fp16 = mul(x = var_372_cast_fp16, y = rotr_1_cast_fp16)[name = tensor("op_398_cast_fp16")]; + tensor var_399_cast_fp16 = mul(x = var_380_cast_fp16, y = roti_1_cast_fp16)[name = tensor("op_399_cast_fp16")]; + tensor kor_1_cast_fp16 = sub(x = var_398_cast_fp16, y = var_399_cast_fp16)[name = tensor("kor_1_cast_fp16")]; + tensor var_402_cast_fp16 = mul(x = var_372_cast_fp16, y = roti_1_cast_fp16)[name = tensor("op_402_cast_fp16")]; + tensor var_403_cast_fp16 = mul(x = var_380_cast_fp16, y = rotr_1_cast_fp16)[name = tensor("op_403_cast_fp16")]; + tensor koi_1_cast_fp16 = add(x = var_402_cast_fp16, y = var_403_cast_fp16)[name = tensor("koi_1_cast_fp16")]; + tensor qo_1_axis_0 = const()[name = tensor("qo_1_axis_0"), val = tensor(-1)]; + tensor qo_1_cast_fp16 = stack(axis = qo_1_axis_0, values = (qor_1_cast_fp16, qoi_1_cast_fp16))[name = tensor("qo_1_cast_fp16")]; + tensor ko_1_axis_0 = const()[name = tensor("ko_1_axis_0"), val = tensor(-1)]; + tensor ko_1_cast_fp16 = stack(axis = ko_1_axis_0, values = (kor_1_cast_fp16, koi_1_cast_fp16))[name = tensor("ko_1_cast_fp16")]; + tensor var_432 = const()[name = tensor("op_432"), val = tensor([1, 1, 16, 64])]; + tensor q_3_cast_fp16 = reshape(shape = var_432, x = qo_1_cast_fp16)[name = tensor("q_3_cast_fp16")]; + tensor var_434 = const()[name = tensor("op_434"), val = tensor([1, 1, 16, 64])]; + tensor k_3_cast_fp16 = reshape(shape = var_434, x = ko_1_cast_fp16)[name = tensor("k_3_cast_fp16")]; + tensor _inversed_456_y_0_to_fp16 = const()[name = tensor("_inversed_456_y_0_to_fp16"), val = tensor(0x1p-9)]; + tensor _inversed_456_cast_fp16 = mul(x = ts_5_cast_fp16, y = _inversed_456_y_0_to_fp16)[name = tensor("_inversed_456_cast_fp16")]; + tensor var_457_cast_fp16 = floor(x = _inversed_456_cast_fp16)[name = tensor("op_457_cast_fp16")]; + tensor var_458_to_fp16 = const()[name = tensor("op_458_to_fp16"), val = tensor(0x1p+9)]; + tensor var_459_cast_fp16 = mul(x = var_457_cast_fp16, y = var_458_to_fp16)[name = tensor("op_459_cast_fp16")]; + tensor write_indices_float_3_cast_fp16 = sub(x = ts_5_cast_fp16, y = var_459_cast_fp16)[name = tensor("write_indices_float_3_cast_fp16")]; + tensor var_466_dtype_0 = const()[name = tensor("op_466_dtype_0"), val = tensor("int32")]; + tensor write_indices_1_reps_0 = const()[name = tensor("write_indices_1_reps_0"), val = tensor([1, 1, 16, 64])]; + tensor write_indices_float_3_cast_fp16_to_int32 = cast(dtype = var_466_dtype_0, x = write_indices_float_3_cast_fp16)[name = tensor("cast_552")]; + tensor write_indices_1 = tile(reps = write_indices_1_reps_0, x = write_indices_float_3_cast_fp16_to_int32)[name = tensor("write_indices_1")]; + tensor var_474_begin_0 = const()[name = tensor("op_474_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_474_end_0 = const()[name = tensor("op_474_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor var_474_end_mask_0 = const()[name = tensor("op_474_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_474_squeeze_mask_0 = const()[name = tensor("op_474_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor cache0_to_fp16_dtype_0 = const()[name = tensor("cache0_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor cache0_to_fp16 = cast(dtype = cache0_to_fp16_dtype_0, x = cache0)[name = tensor("cast_551")]; + tensor var_474_cast_fp16 = slice_by_index(begin = var_474_begin_0, end = var_474_end_0, end_mask = var_474_end_mask_0, squeeze_mask = var_474_squeeze_mask_0, x = cache0_to_fp16)[name = tensor("op_474_cast_fp16")]; + tensor var_476_axis_0 = const()[name = tensor("op_476_axis_0"), val = tensor(1)]; + tensor var_476_mode_0 = const()[name = tensor("op_476_mode_0"), val = tensor("update")]; + tensor var_476_validate_indices_0 = const()[name = tensor("op_476_validate_indices_0"), val = tensor(false)]; + tensor var_476_cast_fp16 = scatter_along_axis(axis = var_476_axis_0, data = var_474_cast_fp16, indices = write_indices_1, mode = var_476_mode_0, updates = k_3_cast_fp16, validate_indices = var_476_validate_indices_0)[name = tensor("op_476_cast_fp16")]; + tensor concat_2 = const()[name = tensor("concat_2"), val = tensor([0, 0, 0, 0, 0])]; + tensor concat_3 = const()[name = tensor("concat_3"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_1_internal_tensor_assign_1_stride_0 = const()[name = tensor("new_cache_1_internal_tensor_assign_1_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_1_internal_tensor_assign_1_begin_mask_0 = const()[name = tensor("new_cache_1_internal_tensor_assign_1_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_1_internal_tensor_assign_1_end_mask_0 = const()[name = tensor("new_cache_1_internal_tensor_assign_1_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_1_internal_tensor_assign_1_squeeze_mask_0 = const()[name = tensor("new_cache_1_internal_tensor_assign_1_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_48 = const()[name = tensor("shape_48"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_0 = const()[name = tensor("reduce_prod_0"), val = tensor(1048576)]; + tensor range_1d_0_start_0 = const()[name = tensor("range_1d_0_start_0"), val = tensor(0)]; + tensor range_1d_0_step_0 = const()[name = tensor("range_1d_0_step_0"), val = tensor(1)]; + tensor range_1d_0 = range_1d(end = reduce_prod_0, start = range_1d_0_start_0, step = range_1d_0_step_0)[name = tensor("range_1d_0")]; + tensor reshape_0 = reshape(shape = shape_48, x = range_1d_0)[name = tensor("reshape_0")]; + tensor slice_by_index_0 = slice_by_index(begin = concat_2, begin_mask = new_cache_1_internal_tensor_assign_1_begin_mask_0, end = concat_3, end_mask = new_cache_1_internal_tensor_assign_1_end_mask_0, squeeze_mask = new_cache_1_internal_tensor_assign_1_squeeze_mask_0, stride = new_cache_1_internal_tensor_assign_1_stride_0, x = reshape_0)[name = tensor("slice_by_index_0")]; + tensor reshape_1_shape_0 = const()[name = tensor("reshape_1_shape_0"), val = tensor([-1])]; + tensor reshape_1 = reshape(shape = reshape_1_shape_0, x = slice_by_index_0)[name = tensor("reshape_1")]; + tensor reshape_2_shape_0 = const()[name = tensor("reshape_2_shape_0"), val = tensor([-1])]; + tensor reshape_2_cast_fp16 = reshape(shape = reshape_2_shape_0, x = var_476_cast_fp16)[name = tensor("reshape_2_cast_fp16")]; + tensor reshape_3_shape_0 = const()[name = tensor("reshape_3_shape_0"), val = tensor([-1])]; + tensor reshape_3_cast_fp16 = reshape(shape = reshape_3_shape_0, x = cache0_to_fp16)[name = tensor("reshape_3_cast_fp16")]; + tensor scatter_0_mode_0 = const()[name = tensor("scatter_0_mode_0"), val = tensor("update")]; + tensor scatter_0_axis_0 = const()[name = tensor("scatter_0_axis_0"), val = tensor(0)]; + tensor scatter_0_validate_indices_0 = const()[name = tensor("scatter_0_validate_indices_0"), val = tensor(false)]; + tensor scatter_0_cast_fp16 = scatter(axis = scatter_0_axis_0, data = reshape_3_cast_fp16, indices = reshape_1, mode = scatter_0_mode_0, updates = reshape_2_cast_fp16, validate_indices = scatter_0_validate_indices_0)[name = tensor("scatter_0_cast_fp16")]; + tensor reshape_4_cast_fp16 = reshape(shape = shape_48, x = scatter_0_cast_fp16)[name = tensor("reshape_4_cast_fp16")]; + tensor var_484_begin_0 = const()[name = tensor("op_484_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor var_484_end_0 = const()[name = tensor("op_484_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor var_484_end_mask_0 = const()[name = tensor("op_484_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_484_squeeze_mask_0 = const()[name = tensor("op_484_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor var_484_cast_fp16 = slice_by_index(begin = var_484_begin_0, end = var_484_end_0, end_mask = var_484_end_mask_0, squeeze_mask = var_484_squeeze_mask_0, x = reshape_4_cast_fp16)[name = tensor("op_484_cast_fp16")]; + tensor var_486_axis_0 = const()[name = tensor("op_486_axis_0"), val = tensor(1)]; + tensor var_486_mode_0 = const()[name = tensor("op_486_mode_0"), val = tensor("update")]; + tensor var_486_validate_indices_0 = const()[name = tensor("op_486_validate_indices_0"), val = tensor(false)]; + tensor var_486_cast_fp16 = scatter_along_axis(axis = var_486_axis_0, data = var_484_cast_fp16, indices = write_indices_1, mode = var_486_mode_0, updates = v_1_cast_fp16, validate_indices = var_486_validate_indices_0)[name = tensor("op_486_cast_fp16")]; + tensor concat_4 = const()[name = tensor("concat_4"), val = tensor([1, 0, 0, 0, 0])]; + tensor concat_5 = const()[name = tensor("concat_5"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_1_internal_tensor_assign_2_stride_0 = const()[name = tensor("new_cache_1_internal_tensor_assign_2_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_1_internal_tensor_assign_2_begin_mask_0 = const()[name = tensor("new_cache_1_internal_tensor_assign_2_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_1_internal_tensor_assign_2_end_mask_0 = const()[name = tensor("new_cache_1_internal_tensor_assign_2_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_1_internal_tensor_assign_2_squeeze_mask_0 = const()[name = tensor("new_cache_1_internal_tensor_assign_2_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_49 = const()[name = tensor("shape_49"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_1 = const()[name = tensor("reduce_prod_1"), val = tensor(1048576)]; + tensor range_1d_1_start_0 = const()[name = tensor("range_1d_1_start_0"), val = tensor(0)]; + tensor range_1d_1_step_0 = const()[name = tensor("range_1d_1_step_0"), val = tensor(1)]; + tensor range_1d_1 = range_1d(end = reduce_prod_1, start = range_1d_1_start_0, step = range_1d_1_step_0)[name = tensor("range_1d_1")]; + tensor reshape_5 = reshape(shape = shape_49, x = range_1d_1)[name = tensor("reshape_5")]; + tensor slice_by_index_1 = slice_by_index(begin = concat_4, begin_mask = new_cache_1_internal_tensor_assign_2_begin_mask_0, end = concat_5, end_mask = new_cache_1_internal_tensor_assign_2_end_mask_0, squeeze_mask = new_cache_1_internal_tensor_assign_2_squeeze_mask_0, stride = new_cache_1_internal_tensor_assign_2_stride_0, x = reshape_5)[name = tensor("slice_by_index_1")]; + tensor reshape_6_shape_0 = const()[name = tensor("reshape_6_shape_0"), val = tensor([-1])]; + tensor reshape_6 = reshape(shape = reshape_6_shape_0, x = slice_by_index_1)[name = tensor("reshape_6")]; + tensor reshape_7_shape_0 = const()[name = tensor("reshape_7_shape_0"), val = tensor([-1])]; + tensor reshape_7_cast_fp16 = reshape(shape = reshape_7_shape_0, x = var_486_cast_fp16)[name = tensor("reshape_7_cast_fp16")]; + tensor reshape_8_shape_0 = const()[name = tensor("reshape_8_shape_0"), val = tensor([-1])]; + tensor reshape_8_cast_fp16 = reshape(shape = reshape_8_shape_0, x = reshape_4_cast_fp16)[name = tensor("reshape_8_cast_fp16")]; + tensor scatter_1_mode_0 = const()[name = tensor("scatter_1_mode_0"), val = tensor("update")]; + tensor scatter_1_axis_0 = const()[name = tensor("scatter_1_axis_0"), val = tensor(0)]; + tensor scatter_1_validate_indices_0 = const()[name = tensor("scatter_1_validate_indices_0"), val = tensor(false)]; + tensor scatter_1_cast_fp16 = scatter(axis = scatter_1_axis_0, data = reshape_8_cast_fp16, indices = reshape_6, mode = scatter_1_mode_0, updates = reshape_7_cast_fp16, validate_indices = scatter_1_validate_indices_0)[name = tensor("scatter_1_cast_fp16")]; + tensor reshape_9_cast_fp16 = reshape(shape = shape_49, x = scatter_1_cast_fp16)[name = tensor("reshape_9_cast_fp16")]; + tensor reshape_9_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("reshape_9_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor keys_1_begin_0 = const()[name = tensor("keys_1_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor keys_1_end_0 = const()[name = tensor("keys_1_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor keys_1_end_mask_0 = const()[name = tensor("keys_1_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor keys_1_squeeze_mask_0 = const()[name = tensor("keys_1_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor keys_1_cast_fp16 = slice_by_index(begin = keys_1_begin_0, end = keys_1_end_0, end_mask = keys_1_end_mask_0, squeeze_mask = keys_1_squeeze_mask_0, x = reshape_9_cast_fp16)[name = tensor("keys_1_cast_fp16")]; + tensor values_1_begin_0 = const()[name = tensor("values_1_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor values_1_end_0 = const()[name = tensor("values_1_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor values_1_end_mask_0 = const()[name = tensor("values_1_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor values_1_squeeze_mask_0 = const()[name = tensor("values_1_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor values_1_cast_fp16 = slice_by_index(begin = values_1_begin_0, end = values_1_end_0, end_mask = values_1_end_mask_0, squeeze_mask = values_1_squeeze_mask_0, x = reshape_9_cast_fp16)[name = tensor("values_1_cast_fp16")]; + tensor var_498_cast_fp16 = not_equal(x = keys_1_cast_fp16, y = keys_1_cast_fp16)[name = tensor("op_498_cast_fp16")]; + tensor var_504_to_fp16 = const()[name = tensor("op_504_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(302531264)))]; + tensor keys_3_cast_fp16 = select(a = var_504_to_fp16, b = keys_1_cast_fp16, cond = var_498_cast_fp16)[name = tensor("keys_3_cast_fp16")]; + tensor var_506_cast_fp16 = not_equal(x = values_1_cast_fp16, y = values_1_cast_fp16)[name = tensor("op_506_cast_fp16")]; + tensor values_3_cast_fp16 = select(a = var_504_to_fp16, b = values_1_cast_fp16, cond = var_506_cast_fp16)[name = tensor("values_3_cast_fp16")]; + tensor var_530 = const()[name = tensor("op_530"), val = tensor([0, 2, 1, 3])]; + tensor var_543 = const()[name = tensor("op_543"), val = tensor([1, 1, 1])]; + tensor var_544_cast_fp16 = reshape(shape = var_543, x = position0_to_fp16)[name = tensor("op_544_cast_fp16")]; + tensor var_561_to_fp16 = const()[name = tensor("op_561_to_fp16"), val = tensor(0x1p+0)]; + tensor valid_len_1_cast_fp16 = add(x = var_544_cast_fp16, y = var_561_to_fp16)[name = tensor("valid_len_1_cast_fp16")]; + tensor k_positions_1_promoted_to_fp16 = const()[name = tensor("k_positions_1_promoted_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(303579904)))]; + tensor valid_mask_1_cast_fp16 = less(x = k_positions_1_promoted_to_fp16, y = valid_len_1_cast_fp16)[name = tensor("valid_mask_1_cast_fp16")]; + tensor causal_mask_1_cast_fp16 = less_equal(x = k_positions_1_promoted_to_fp16, y = var_544_cast_fp16)[name = tensor("causal_mask_1_cast_fp16")]; + tensor attn_mask_1 = logical_and(x = valid_mask_1_cast_fp16, y = causal_mask_1_cast_fp16)[name = tensor("attn_mask_1")]; + tensor attn_mask_3_axes_0 = const()[name = tensor("attn_mask_3_axes_0"), val = tensor([1])]; + tensor attn_mask_3 = expand_dims(axes = attn_mask_3_axes_0, x = attn_mask_1)[name = tensor("attn_mask_3")]; + tensor var_579_transpose_x_0 = const()[name = tensor("op_579_transpose_x_0"), val = tensor(false)]; + tensor var_579_transpose_y_0 = const()[name = tensor("op_579_transpose_y_0"), val = tensor(false)]; + tensor transpose_72_perm_0 = const()[name = tensor("transpose_72_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_73_perm_0 = const()[name = tensor("transpose_73_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_73 = transpose(perm = transpose_73_perm_0, x = keys_3_cast_fp16)[name = tensor("transpose_213")]; + tensor transpose_72 = transpose(perm = transpose_72_perm_0, x = q_3_cast_fp16)[name = tensor("transpose_214")]; + tensor var_579_cast_fp16 = matmul(transpose_x = var_579_transpose_x_0, transpose_y = var_579_transpose_y_0, x = transpose_72, y = transpose_73)[name = tensor("op_579_cast_fp16")]; + tensor var_573_to_fp16 = const()[name = tensor("op_573_to_fp16"), val = tensor([0x1p-3])]; + tensor attn_weights_1_cast_fp16 = mul(x = var_579_cast_fp16, y = var_573_to_fp16)[name = tensor("attn_weights_1_cast_fp16")]; + tensor var_581 = logical_not(x = attn_mask_3)[name = tensor("op_581")]; + tensor var_582_to_fp16 = const()[name = tensor("op_582_to_fp16"), val = tensor(-inf)]; + tensor attn_weights_3_cast_fp16 = select(a = var_582_to_fp16, b = attn_weights_1_cast_fp16, cond = var_581)[name = tensor("attn_weights_3_cast_fp16")]; + tensor var_584 = const()[name = tensor("op_584"), val = tensor(-1)]; + tensor attn_weights_5_cast_fp16 = softmax(axis = var_584, x = attn_weights_3_cast_fp16)[name = tensor("attn_weights_5_cast_fp16")]; + tensor attn_output_1_transpose_x_0 = const()[name = tensor("attn_output_1_transpose_x_0"), val = tensor(false)]; + tensor attn_output_1_transpose_y_0 = const()[name = tensor("attn_output_1_transpose_y_0"), val = tensor(false)]; + tensor values_5_cast_fp16 = transpose(perm = var_530, x = values_3_cast_fp16)[name = tensor("transpose_215")]; + tensor attn_output_1_cast_fp16 = matmul(transpose_x = attn_output_1_transpose_x_0, transpose_y = attn_output_1_transpose_y_0, x = attn_weights_5_cast_fp16, y = values_5_cast_fp16)[name = tensor("attn_output_1_cast_fp16")]; + tensor var_592 = const()[name = tensor("op_592"), val = tensor([0, 2, 1, 3])]; + tensor var_595 = const()[name = tensor("op_595"), val = tensor([1, 1, 1024])]; + tensor var_593_cast_fp16 = transpose(perm = var_592, x = attn_output_1_cast_fp16)[name = tensor("transpose_212")]; + tensor input_5_cast_fp16 = reshape(shape = var_595, x = var_593_cast_fp16)[name = tensor("input_5_cast_fp16")]; + tensor linear_2_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = attn0_out_proj_weight_cast_fp16, x = input_5_cast_fp16)[name = tensor("linear_2_cast_fp16")]; + tensor var_601_to_fp16 = const()[name = tensor("op_601_to_fp16"), val = tensor(0x1p+0)]; + tensor var_602_cast_fp16 = add(x = position0_to_fp16, y = var_601_to_fp16)[name = tensor("op_602_cast_fp16")]; + tensor var_602_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_602_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor input_7_cast_fp16 = add(x = linear_0_cast_fp16, y = linear_2_cast_fp16)[name = tensor("input_7_cast_fp16")]; + tensor input_9_axes_0 = const()[name = tensor("input_9_axes_0"), val = tensor([-1])]; + tensor norm0_2_weight_to_fp16 = const()[name = tensor("norm0_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(303580992)))]; + tensor norm0_2_bias_to_fp16 = const()[name = tensor("norm0_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(303583104)))]; + tensor var_606_to_fp16 = const()[name = tensor("op_606_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_9_cast_fp16 = layer_norm(axes = input_9_axes_0, beta = norm0_2_bias_to_fp16, epsilon = var_606_to_fp16, gamma = norm0_2_weight_to_fp16, x = input_7_cast_fp16)[name = tensor("input_9_cast_fp16")]; + tensor linear_3_bias_0_to_fp16 = const()[name = tensor("linear_3_bias_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(303585216)))]; + tensor linear_3_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = linear0_1_weight_cast_fp16, x = input_9_cast_fp16)[name = tensor("linear_3_cast_fp16")]; + tensor input_11_mode_0 = const()[name = tensor("input_11_mode_0"), val = tensor("EXACT")]; + tensor input_11_cast_fp16 = gelu(mode = input_11_mode_0, x = linear_3_cast_fp16)[name = tensor("input_11_cast_fp16")]; + tensor linear_4_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = linear0_2_weight_cast_fp16, x = input_11_cast_fp16)[name = tensor("linear_4_cast_fp16")]; + tensor input_13_cast_fp16 = add(x = input_7_cast_fp16, y = linear_4_cast_fp16)[name = tensor("input_13_cast_fp16")]; + tensor x_3_axes_0 = const()[name = tensor("x_3_axes_0"), val = tensor([-1])]; + tensor norm1_1_weight_to_fp16 = const()[name = tensor("norm1_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(303593472)))]; + tensor norm1_1_bias_to_fp16 = const()[name = tensor("norm1_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(303595584)))]; + tensor var_623_to_fp16 = const()[name = tensor("op_623_to_fp16"), val = tensor(0x1.5p-17)]; + tensor x_3_cast_fp16 = layer_norm(axes = x_3_axes_0, beta = norm1_1_bias_to_fp16, epsilon = var_623_to_fp16, gamma = norm1_1_weight_to_fp16, x = input_13_cast_fp16)[name = tensor("x_3_cast_fp16")]; + tensor linear_5_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = attn1_in_proj_weight_cast_fp16, x = x_3_cast_fp16)[name = tensor("linear_5_cast_fp16")]; + tensor var_659 = const()[name = tensor("op_659"), val = tensor([1, 1, 3, 16, 64])]; + tensor qkv_3_cast_fp16 = reshape(shape = var_659, x = linear_5_cast_fp16)[name = tensor("qkv_3_cast_fp16")]; + tensor q_7_begin_0 = const()[name = tensor("q_7_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor q_7_end_0 = const()[name = tensor("q_7_end_0"), val = tensor([1, 1, 1, 16, 64])]; + tensor q_7_end_mask_0 = const()[name = tensor("q_7_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor q_7_squeeze_mask_0 = const()[name = tensor("q_7_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor q_7_cast_fp16 = slice_by_index(begin = q_7_begin_0, end = q_7_end_0, end_mask = q_7_end_mask_0, squeeze_mask = q_7_squeeze_mask_0, x = qkv_3_cast_fp16)[name = tensor("q_7_cast_fp16")]; + tensor k_5_begin_0 = const()[name = tensor("k_5_begin_0"), val = tensor([0, 0, 1, 0, 0])]; + tensor k_5_end_0 = const()[name = tensor("k_5_end_0"), val = tensor([1, 1, 2, 16, 64])]; + tensor k_5_end_mask_0 = const()[name = tensor("k_5_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor k_5_squeeze_mask_0 = const()[name = tensor("k_5_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor k_5_cast_fp16 = slice_by_index(begin = k_5_begin_0, end = k_5_end_0, end_mask = k_5_end_mask_0, squeeze_mask = k_5_squeeze_mask_0, x = qkv_3_cast_fp16)[name = tensor("k_5_cast_fp16")]; + tensor v_3_begin_0 = const()[name = tensor("v_3_begin_0"), val = tensor([0, 0, 2, 0, 0])]; + tensor v_3_end_0 = const()[name = tensor("v_3_end_0"), val = tensor([1, 1, 3, 16, 64])]; + tensor v_3_end_mask_0 = const()[name = tensor("v_3_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor v_3_squeeze_mask_0 = const()[name = tensor("v_3_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor v_3_cast_fp16 = slice_by_index(begin = v_3_begin_0, end = v_3_end_0, end_mask = v_3_end_mask_0, squeeze_mask = v_3_squeeze_mask_0, x = qkv_3_cast_fp16)[name = tensor("v_3_cast_fp16")]; + tensor var_763 = const()[name = tensor("op_763"), val = tensor([1, 1, 1, 1])]; + tensor position1_to_fp16_dtype_0 = const()[name = tensor("position1_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor position1_to_fp16 = cast(dtype = position1_to_fp16_dtype_0, x = position1)[name = tensor("cast_548")]; + tensor ts_11_cast_fp16 = reshape(shape = var_763, x = position1_to_fp16)[name = tensor("ts_11_cast_fp16")]; + tensor var_767 = const()[name = tensor("op_767"), val = tensor([1, 1, 16, 32, 2])]; + tensor q_complex_3_cast_fp16 = reshape(shape = var_767, x = q_7_cast_fp16)[name = tensor("q_complex_3_cast_fp16")]; + tensor var_771 = const()[name = tensor("op_771"), val = tensor([1, 1, 16, 32, 2])]; + tensor k_complex_3_cast_fp16 = reshape(shape = var_771, x = k_5_cast_fp16)[name = tensor("k_complex_3_cast_fp16")]; + tensor var_775_begin_0 = const()[name = tensor("op_775_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_775_end_0 = const()[name = tensor("op_775_end_0"), val = tensor([1, 1, 16, 32, 1])]; + tensor var_775_end_mask_0 = const()[name = tensor("op_775_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_775_squeeze_mask_0 = const()[name = tensor("op_775_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_775_cast_fp16 = slice_by_index(begin = var_775_begin_0, end = var_775_end_0, end_mask = var_775_end_mask_0, squeeze_mask = var_775_squeeze_mask_0, x = q_complex_3_cast_fp16)[name = tensor("op_775_cast_fp16")]; + tensor var_783_begin_0 = const()[name = tensor("op_783_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_783_end_0 = const()[name = tensor("op_783_end_0"), val = tensor([1, 1, 16, 32, 2])]; + tensor var_783_end_mask_0 = const()[name = tensor("op_783_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_783_squeeze_mask_0 = const()[name = tensor("op_783_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_783_cast_fp16 = slice_by_index(begin = var_783_begin_0, end = var_783_end_0, end_mask = var_783_end_mask_0, squeeze_mask = var_783_squeeze_mask_0, x = q_complex_3_cast_fp16)[name = tensor("op_783_cast_fp16")]; + tensor var_791_begin_0 = const()[name = tensor("op_791_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_791_end_0 = const()[name = tensor("op_791_end_0"), val = tensor([1, 1, 16, 32, 1])]; + tensor var_791_end_mask_0 = const()[name = tensor("op_791_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_791_squeeze_mask_0 = const()[name = tensor("op_791_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_791_cast_fp16 = slice_by_index(begin = var_791_begin_0, end = var_791_end_0, end_mask = var_791_end_mask_0, squeeze_mask = var_791_squeeze_mask_0, x = k_complex_3_cast_fp16)[name = tensor("op_791_cast_fp16")]; + tensor var_799_begin_0 = const()[name = tensor("op_799_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_799_end_0 = const()[name = tensor("op_799_end_0"), val = tensor([1, 1, 16, 32, 2])]; + tensor var_799_end_mask_0 = const()[name = tensor("op_799_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_799_squeeze_mask_0 = const()[name = tensor("op_799_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_799_cast_fp16 = slice_by_index(begin = var_799_begin_0, end = var_799_end_0, end_mask = var_799_end_mask_0, squeeze_mask = var_799_squeeze_mask_0, x = k_complex_3_cast_fp16)[name = tensor("op_799_cast_fp16")]; + tensor freqs_3_to_fp16 = const()[name = tensor("freqs_3_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(303597696)))]; + tensor var_805_cast_fp16 = mul(x = freqs_3_to_fp16, y = ts_11_cast_fp16)[name = tensor("op_805_cast_fp16")]; + tensor rotr_3_cast_fp16 = cos(x = var_805_cast_fp16)[name = tensor("rotr_3_cast_fp16")]; + tensor roti_3_cast_fp16 = sin(x = var_805_cast_fp16)[name = tensor("roti_3_cast_fp16")]; + tensor var_809_cast_fp16 = mul(x = var_775_cast_fp16, y = rotr_3_cast_fp16)[name = tensor("op_809_cast_fp16")]; + tensor var_810_cast_fp16 = mul(x = var_783_cast_fp16, y = roti_3_cast_fp16)[name = tensor("op_810_cast_fp16")]; + tensor qor_5_cast_fp16 = sub(x = var_809_cast_fp16, y = var_810_cast_fp16)[name = tensor("qor_5_cast_fp16")]; + tensor var_813_cast_fp16 = mul(x = var_775_cast_fp16, y = roti_3_cast_fp16)[name = tensor("op_813_cast_fp16")]; + tensor var_814_cast_fp16 = mul(x = var_783_cast_fp16, y = rotr_3_cast_fp16)[name = tensor("op_814_cast_fp16")]; + tensor qoi_5_cast_fp16 = add(x = var_813_cast_fp16, y = var_814_cast_fp16)[name = tensor("qoi_5_cast_fp16")]; + tensor var_817_cast_fp16 = mul(x = var_791_cast_fp16, y = rotr_3_cast_fp16)[name = tensor("op_817_cast_fp16")]; + tensor var_818_cast_fp16 = mul(x = var_799_cast_fp16, y = roti_3_cast_fp16)[name = tensor("op_818_cast_fp16")]; + tensor kor_5_cast_fp16 = sub(x = var_817_cast_fp16, y = var_818_cast_fp16)[name = tensor("kor_5_cast_fp16")]; + tensor var_821_cast_fp16 = mul(x = var_791_cast_fp16, y = roti_3_cast_fp16)[name = tensor("op_821_cast_fp16")]; + tensor var_822_cast_fp16 = mul(x = var_799_cast_fp16, y = rotr_3_cast_fp16)[name = tensor("op_822_cast_fp16")]; + tensor koi_5_cast_fp16 = add(x = var_821_cast_fp16, y = var_822_cast_fp16)[name = tensor("koi_5_cast_fp16")]; + tensor qo_3_axis_0 = const()[name = tensor("qo_3_axis_0"), val = tensor(-1)]; + tensor qo_3_cast_fp16 = stack(axis = qo_3_axis_0, values = (qor_5_cast_fp16, qoi_5_cast_fp16))[name = tensor("qo_3_cast_fp16")]; + tensor ko_3_axis_0 = const()[name = tensor("ko_3_axis_0"), val = tensor(-1)]; + tensor ko_3_cast_fp16 = stack(axis = ko_3_axis_0, values = (kor_5_cast_fp16, koi_5_cast_fp16))[name = tensor("ko_3_cast_fp16")]; + tensor var_851 = const()[name = tensor("op_851"), val = tensor([1, 1, 16, 64])]; + tensor q_9_cast_fp16 = reshape(shape = var_851, x = qo_3_cast_fp16)[name = tensor("q_9_cast_fp16")]; + tensor var_853 = const()[name = tensor("op_853"), val = tensor([1, 1, 16, 64])]; + tensor k_7_cast_fp16 = reshape(shape = var_853, x = ko_3_cast_fp16)[name = tensor("k_7_cast_fp16")]; + tensor _inversed_875_y_0_to_fp16 = const()[name = tensor("_inversed_875_y_0_to_fp16"), val = tensor(0x1p-9)]; + tensor _inversed_875_cast_fp16 = mul(x = ts_11_cast_fp16, y = _inversed_875_y_0_to_fp16)[name = tensor("_inversed_875_cast_fp16")]; + tensor var_876_cast_fp16 = floor(x = _inversed_875_cast_fp16)[name = tensor("op_876_cast_fp16")]; + tensor var_877_to_fp16 = const()[name = tensor("op_877_to_fp16"), val = tensor(0x1p+9)]; + tensor var_878_cast_fp16 = mul(x = var_876_cast_fp16, y = var_877_to_fp16)[name = tensor("op_878_cast_fp16")]; + tensor write_indices_float_7_cast_fp16 = sub(x = ts_11_cast_fp16, y = var_878_cast_fp16)[name = tensor("write_indices_float_7_cast_fp16")]; + tensor var_885_dtype_0 = const()[name = tensor("op_885_dtype_0"), val = tensor("int32")]; + tensor write_indices_3_reps_0 = const()[name = tensor("write_indices_3_reps_0"), val = tensor([1, 1, 16, 64])]; + tensor write_indices_float_7_cast_fp16_to_int32 = cast(dtype = var_885_dtype_0, x = write_indices_float_7_cast_fp16)[name = tensor("cast_547")]; + tensor write_indices_3 = tile(reps = write_indices_3_reps_0, x = write_indices_float_7_cast_fp16_to_int32)[name = tensor("write_indices_3")]; + tensor var_893_begin_0 = const()[name = tensor("op_893_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_893_end_0 = const()[name = tensor("op_893_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor var_893_end_mask_0 = const()[name = tensor("op_893_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_893_squeeze_mask_0 = const()[name = tensor("op_893_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor cache1_to_fp16_dtype_0 = const()[name = tensor("cache1_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor cache1_to_fp16 = cast(dtype = cache1_to_fp16_dtype_0, x = cache1)[name = tensor("cast_546")]; + tensor var_893_cast_fp16 = slice_by_index(begin = var_893_begin_0, end = var_893_end_0, end_mask = var_893_end_mask_0, squeeze_mask = var_893_squeeze_mask_0, x = cache1_to_fp16)[name = tensor("op_893_cast_fp16")]; + tensor var_895_axis_0 = const()[name = tensor("op_895_axis_0"), val = tensor(1)]; + tensor var_895_mode_0 = const()[name = tensor("op_895_mode_0"), val = tensor("update")]; + tensor var_895_validate_indices_0 = const()[name = tensor("op_895_validate_indices_0"), val = tensor(false)]; + tensor var_895_cast_fp16 = scatter_along_axis(axis = var_895_axis_0, data = var_893_cast_fp16, indices = write_indices_3, mode = var_895_mode_0, updates = k_7_cast_fp16, validate_indices = var_895_validate_indices_0)[name = tensor("op_895_cast_fp16")]; + tensor concat_9 = const()[name = tensor("concat_9"), val = tensor([0, 0, 0, 0, 0])]; + tensor concat_10 = const()[name = tensor("concat_10"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_3_internal_tensor_assign_1_stride_0 = const()[name = tensor("new_cache_3_internal_tensor_assign_1_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_3_internal_tensor_assign_1_begin_mask_0 = const()[name = tensor("new_cache_3_internal_tensor_assign_1_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_3_internal_tensor_assign_1_end_mask_0 = const()[name = tensor("new_cache_3_internal_tensor_assign_1_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_3_internal_tensor_assign_1_squeeze_mask_0 = const()[name = tensor("new_cache_3_internal_tensor_assign_1_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_50 = const()[name = tensor("shape_50"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_2 = const()[name = tensor("reduce_prod_2"), val = tensor(1048576)]; + tensor range_1d_2_start_0 = const()[name = tensor("range_1d_2_start_0"), val = tensor(0)]; + tensor range_1d_2_step_0 = const()[name = tensor("range_1d_2_step_0"), val = tensor(1)]; + tensor range_1d_2 = range_1d(end = reduce_prod_2, start = range_1d_2_start_0, step = range_1d_2_step_0)[name = tensor("range_1d_2")]; + tensor reshape_10 = reshape(shape = shape_50, x = range_1d_2)[name = tensor("reshape_10")]; + tensor slice_by_index_2 = slice_by_index(begin = concat_9, begin_mask = new_cache_3_internal_tensor_assign_1_begin_mask_0, end = concat_10, end_mask = new_cache_3_internal_tensor_assign_1_end_mask_0, squeeze_mask = new_cache_3_internal_tensor_assign_1_squeeze_mask_0, stride = new_cache_3_internal_tensor_assign_1_stride_0, x = reshape_10)[name = tensor("slice_by_index_2")]; + tensor reshape_11_shape_0 = const()[name = tensor("reshape_11_shape_0"), val = tensor([-1])]; + tensor reshape_11 = reshape(shape = reshape_11_shape_0, x = slice_by_index_2)[name = tensor("reshape_11")]; + tensor reshape_12_shape_0 = const()[name = tensor("reshape_12_shape_0"), val = tensor([-1])]; + tensor reshape_12_cast_fp16 = reshape(shape = reshape_12_shape_0, x = var_895_cast_fp16)[name = tensor("reshape_12_cast_fp16")]; + tensor reshape_13_shape_0 = const()[name = tensor("reshape_13_shape_0"), val = tensor([-1])]; + tensor reshape_13_cast_fp16 = reshape(shape = reshape_13_shape_0, x = cache1_to_fp16)[name = tensor("reshape_13_cast_fp16")]; + tensor scatter_2_mode_0 = const()[name = tensor("scatter_2_mode_0"), val = tensor("update")]; + tensor scatter_2_axis_0 = const()[name = tensor("scatter_2_axis_0"), val = tensor(0)]; + tensor scatter_2_validate_indices_0 = const()[name = tensor("scatter_2_validate_indices_0"), val = tensor(false)]; + tensor scatter_2_cast_fp16 = scatter(axis = scatter_2_axis_0, data = reshape_13_cast_fp16, indices = reshape_11, mode = scatter_2_mode_0, updates = reshape_12_cast_fp16, validate_indices = scatter_2_validate_indices_0)[name = tensor("scatter_2_cast_fp16")]; + tensor reshape_14_cast_fp16 = reshape(shape = shape_50, x = scatter_2_cast_fp16)[name = tensor("reshape_14_cast_fp16")]; + tensor var_903_begin_0 = const()[name = tensor("op_903_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor var_903_end_0 = const()[name = tensor("op_903_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor var_903_end_mask_0 = const()[name = tensor("op_903_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_903_squeeze_mask_0 = const()[name = tensor("op_903_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor var_903_cast_fp16 = slice_by_index(begin = var_903_begin_0, end = var_903_end_0, end_mask = var_903_end_mask_0, squeeze_mask = var_903_squeeze_mask_0, x = reshape_14_cast_fp16)[name = tensor("op_903_cast_fp16")]; + tensor var_905_axis_0 = const()[name = tensor("op_905_axis_0"), val = tensor(1)]; + tensor var_905_mode_0 = const()[name = tensor("op_905_mode_0"), val = tensor("update")]; + tensor var_905_validate_indices_0 = const()[name = tensor("op_905_validate_indices_0"), val = tensor(false)]; + tensor var_905_cast_fp16 = scatter_along_axis(axis = var_905_axis_0, data = var_903_cast_fp16, indices = write_indices_3, mode = var_905_mode_0, updates = v_3_cast_fp16, validate_indices = var_905_validate_indices_0)[name = tensor("op_905_cast_fp16")]; + tensor concat_11 = const()[name = tensor("concat_11"), val = tensor([1, 0, 0, 0, 0])]; + tensor concat_12 = const()[name = tensor("concat_12"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_3_internal_tensor_assign_2_stride_0 = const()[name = tensor("new_cache_3_internal_tensor_assign_2_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_3_internal_tensor_assign_2_begin_mask_0 = const()[name = tensor("new_cache_3_internal_tensor_assign_2_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_3_internal_tensor_assign_2_end_mask_0 = const()[name = tensor("new_cache_3_internal_tensor_assign_2_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_3_internal_tensor_assign_2_squeeze_mask_0 = const()[name = tensor("new_cache_3_internal_tensor_assign_2_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_51 = const()[name = tensor("shape_51"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_3 = const()[name = tensor("reduce_prod_3"), val = tensor(1048576)]; + tensor range_1d_3_start_0 = const()[name = tensor("range_1d_3_start_0"), val = tensor(0)]; + tensor range_1d_3_step_0 = const()[name = tensor("range_1d_3_step_0"), val = tensor(1)]; + tensor range_1d_3 = range_1d(end = reduce_prod_3, start = range_1d_3_start_0, step = range_1d_3_step_0)[name = tensor("range_1d_3")]; + tensor reshape_15 = reshape(shape = shape_51, x = range_1d_3)[name = tensor("reshape_15")]; + tensor slice_by_index_3 = slice_by_index(begin = concat_11, begin_mask = new_cache_3_internal_tensor_assign_2_begin_mask_0, end = concat_12, end_mask = new_cache_3_internal_tensor_assign_2_end_mask_0, squeeze_mask = new_cache_3_internal_tensor_assign_2_squeeze_mask_0, stride = new_cache_3_internal_tensor_assign_2_stride_0, x = reshape_15)[name = tensor("slice_by_index_3")]; + tensor reshape_16_shape_0 = const()[name = tensor("reshape_16_shape_0"), val = tensor([-1])]; + tensor reshape_16 = reshape(shape = reshape_16_shape_0, x = slice_by_index_3)[name = tensor("reshape_16")]; + tensor reshape_17_shape_0 = const()[name = tensor("reshape_17_shape_0"), val = tensor([-1])]; + tensor reshape_17_cast_fp16 = reshape(shape = reshape_17_shape_0, x = var_905_cast_fp16)[name = tensor("reshape_17_cast_fp16")]; + tensor reshape_18_shape_0 = const()[name = tensor("reshape_18_shape_0"), val = tensor([-1])]; + tensor reshape_18_cast_fp16 = reshape(shape = reshape_18_shape_0, x = reshape_14_cast_fp16)[name = tensor("reshape_18_cast_fp16")]; + tensor scatter_3_mode_0 = const()[name = tensor("scatter_3_mode_0"), val = tensor("update")]; + tensor scatter_3_axis_0 = const()[name = tensor("scatter_3_axis_0"), val = tensor(0)]; + tensor scatter_3_validate_indices_0 = const()[name = tensor("scatter_3_validate_indices_0"), val = tensor(false)]; + tensor scatter_3_cast_fp16 = scatter(axis = scatter_3_axis_0, data = reshape_18_cast_fp16, indices = reshape_16, mode = scatter_3_mode_0, updates = reshape_17_cast_fp16, validate_indices = scatter_3_validate_indices_0)[name = tensor("scatter_3_cast_fp16")]; + tensor reshape_19_cast_fp16 = reshape(shape = shape_51, x = scatter_3_cast_fp16)[name = tensor("reshape_19_cast_fp16")]; + tensor reshape_19_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("reshape_19_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor keys_7_begin_0 = const()[name = tensor("keys_7_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor keys_7_end_0 = const()[name = tensor("keys_7_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor keys_7_end_mask_0 = const()[name = tensor("keys_7_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor keys_7_squeeze_mask_0 = const()[name = tensor("keys_7_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor keys_7_cast_fp16 = slice_by_index(begin = keys_7_begin_0, end = keys_7_end_0, end_mask = keys_7_end_mask_0, squeeze_mask = keys_7_squeeze_mask_0, x = reshape_19_cast_fp16)[name = tensor("keys_7_cast_fp16")]; + tensor values_7_begin_0 = const()[name = tensor("values_7_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor values_7_end_0 = const()[name = tensor("values_7_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor values_7_end_mask_0 = const()[name = tensor("values_7_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor values_7_squeeze_mask_0 = const()[name = tensor("values_7_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor values_7_cast_fp16 = slice_by_index(begin = values_7_begin_0, end = values_7_end_0, end_mask = values_7_end_mask_0, squeeze_mask = values_7_squeeze_mask_0, x = reshape_19_cast_fp16)[name = tensor("values_7_cast_fp16")]; + tensor var_917_cast_fp16 = not_equal(x = keys_7_cast_fp16, y = keys_7_cast_fp16)[name = tensor("op_917_cast_fp16")]; + tensor keys_9_cast_fp16 = select(a = var_504_to_fp16, b = keys_7_cast_fp16, cond = var_917_cast_fp16)[name = tensor("keys_9_cast_fp16")]; + tensor var_925_cast_fp16 = not_equal(x = values_7_cast_fp16, y = values_7_cast_fp16)[name = tensor("op_925_cast_fp16")]; + tensor values_9_cast_fp16 = select(a = var_504_to_fp16, b = values_7_cast_fp16, cond = var_925_cast_fp16)[name = tensor("values_9_cast_fp16")]; + tensor var_949 = const()[name = tensor("op_949"), val = tensor([0, 2, 1, 3])]; + tensor var_962 = const()[name = tensor("op_962"), val = tensor([1, 1, 1])]; + tensor var_963_cast_fp16 = reshape(shape = var_962, x = position1_to_fp16)[name = tensor("op_963_cast_fp16")]; + tensor var_980_to_fp16 = const()[name = tensor("op_980_to_fp16"), val = tensor(0x1p+0)]; + tensor valid_len_3_cast_fp16 = add(x = var_963_cast_fp16, y = var_980_to_fp16)[name = tensor("valid_len_3_cast_fp16")]; + tensor valid_mask_3_cast_fp16 = less(x = k_positions_1_promoted_to_fp16, y = valid_len_3_cast_fp16)[name = tensor("valid_mask_3_cast_fp16")]; + tensor causal_mask_3_cast_fp16 = less_equal(x = k_positions_1_promoted_to_fp16, y = var_963_cast_fp16)[name = tensor("causal_mask_3_cast_fp16")]; + tensor attn_mask_5 = logical_and(x = valid_mask_3_cast_fp16, y = causal_mask_3_cast_fp16)[name = tensor("attn_mask_5")]; + tensor attn_mask_7_axes_0 = const()[name = tensor("attn_mask_7_axes_0"), val = tensor([1])]; + tensor attn_mask_7 = expand_dims(axes = attn_mask_7_axes_0, x = attn_mask_5)[name = tensor("attn_mask_7")]; + tensor var_998_transpose_x_0 = const()[name = tensor("op_998_transpose_x_0"), val = tensor(false)]; + tensor var_998_transpose_y_0 = const()[name = tensor("op_998_transpose_y_0"), val = tensor(false)]; + tensor transpose_74_perm_0 = const()[name = tensor("transpose_74_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_75_perm_0 = const()[name = tensor("transpose_75_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_75 = transpose(perm = transpose_75_perm_0, x = keys_9_cast_fp16)[name = tensor("transpose_209")]; + tensor transpose_74 = transpose(perm = transpose_74_perm_0, x = q_9_cast_fp16)[name = tensor("transpose_210")]; + tensor var_998_cast_fp16 = matmul(transpose_x = var_998_transpose_x_0, transpose_y = var_998_transpose_y_0, x = transpose_74, y = transpose_75)[name = tensor("op_998_cast_fp16")]; + tensor var_992_to_fp16 = const()[name = tensor("op_992_to_fp16"), val = tensor([0x1p-3])]; + tensor attn_weights_7_cast_fp16 = mul(x = var_998_cast_fp16, y = var_992_to_fp16)[name = tensor("attn_weights_7_cast_fp16")]; + tensor var_1000 = logical_not(x = attn_mask_7)[name = tensor("op_1000")]; + tensor var_1001_to_fp16 = const()[name = tensor("op_1001_to_fp16"), val = tensor(-inf)]; + tensor attn_weights_9_cast_fp16 = select(a = var_1001_to_fp16, b = attn_weights_7_cast_fp16, cond = var_1000)[name = tensor("attn_weights_9_cast_fp16")]; + tensor var_1003 = const()[name = tensor("op_1003"), val = tensor(-1)]; + tensor attn_weights_11_cast_fp16 = softmax(axis = var_1003, x = attn_weights_9_cast_fp16)[name = tensor("attn_weights_11_cast_fp16")]; + tensor attn_output_3_transpose_x_0 = const()[name = tensor("attn_output_3_transpose_x_0"), val = tensor(false)]; + tensor attn_output_3_transpose_y_0 = const()[name = tensor("attn_output_3_transpose_y_0"), val = tensor(false)]; + tensor values_11_cast_fp16 = transpose(perm = var_949, x = values_9_cast_fp16)[name = tensor("transpose_211")]; + tensor attn_output_3_cast_fp16 = matmul(transpose_x = attn_output_3_transpose_x_0, transpose_y = attn_output_3_transpose_y_0, x = attn_weights_11_cast_fp16, y = values_11_cast_fp16)[name = tensor("attn_output_3_cast_fp16")]; + tensor var_1011 = const()[name = tensor("op_1011"), val = tensor([0, 2, 1, 3])]; + tensor var_1014 = const()[name = tensor("op_1014"), val = tensor([1, 1, 1024])]; + tensor var_1012_cast_fp16 = transpose(perm = var_1011, x = attn_output_3_cast_fp16)[name = tensor("transpose_208")]; + tensor input_15_cast_fp16 = reshape(shape = var_1014, x = var_1012_cast_fp16)[name = tensor("input_15_cast_fp16")]; + tensor linear_6_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = attn1_out_proj_weight_cast_fp16, x = input_15_cast_fp16)[name = tensor("linear_6_cast_fp16")]; + tensor var_1020_to_fp16 = const()[name = tensor("op_1020_to_fp16"), val = tensor(0x1p+0)]; + tensor var_1021_cast_fp16 = add(x = position1_to_fp16, y = var_1020_to_fp16)[name = tensor("op_1021_cast_fp16")]; + tensor var_1021_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_1021_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor input_17_cast_fp16 = add(x = input_13_cast_fp16, y = linear_6_cast_fp16)[name = tensor("input_17_cast_fp16")]; + tensor input_19_axes_0 = const()[name = tensor("input_19_axes_0"), val = tensor([-1])]; + tensor norm1_2_weight_to_fp16 = const()[name = tensor("norm1_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(303597824)))]; + tensor norm1_2_bias_to_fp16 = const()[name = tensor("norm1_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(303599936)))]; + tensor var_1025_to_fp16 = const()[name = tensor("op_1025_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_19_cast_fp16 = layer_norm(axes = input_19_axes_0, beta = norm1_2_bias_to_fp16, epsilon = var_1025_to_fp16, gamma = norm1_2_weight_to_fp16, x = input_17_cast_fp16)[name = tensor("input_19_cast_fp16")]; + tensor linear_7_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = linear1_1_weight_cast_fp16, x = input_19_cast_fp16)[name = tensor("linear_7_cast_fp16")]; + tensor input_21_mode_0 = const()[name = tensor("input_21_mode_0"), val = tensor("EXACT")]; + tensor input_21_cast_fp16 = gelu(mode = input_21_mode_0, x = linear_7_cast_fp16)[name = tensor("input_21_cast_fp16")]; + tensor linear_8_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = linear1_2_weight_cast_fp16, x = input_21_cast_fp16)[name = tensor("linear_8_cast_fp16")]; + tensor input_23_cast_fp16 = add(x = input_17_cast_fp16, y = linear_8_cast_fp16)[name = tensor("input_23_cast_fp16")]; + tensor x_5_axes_0 = const()[name = tensor("x_5_axes_0"), val = tensor([-1])]; + tensor norm2_1_weight_to_fp16 = const()[name = tensor("norm2_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(303602048)))]; + tensor norm2_1_bias_to_fp16 = const()[name = tensor("norm2_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(303604160)))]; + tensor var_1042_to_fp16 = const()[name = tensor("op_1042_to_fp16"), val = tensor(0x1.5p-17)]; + tensor x_5_cast_fp16 = layer_norm(axes = x_5_axes_0, beta = norm2_1_bias_to_fp16, epsilon = var_1042_to_fp16, gamma = norm2_1_weight_to_fp16, x = input_23_cast_fp16)[name = tensor("x_5_cast_fp16")]; + tensor linear_9_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = attn2_in_proj_weight_cast_fp16, x = x_5_cast_fp16)[name = tensor("linear_9_cast_fp16")]; + tensor var_1078 = const()[name = tensor("op_1078"), val = tensor([1, 1, 3, 16, 64])]; + tensor qkv_5_cast_fp16 = reshape(shape = var_1078, x = linear_9_cast_fp16)[name = tensor("qkv_5_cast_fp16")]; + tensor q_13_begin_0 = const()[name = tensor("q_13_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor q_13_end_0 = const()[name = tensor("q_13_end_0"), val = tensor([1, 1, 1, 16, 64])]; + tensor q_13_end_mask_0 = const()[name = tensor("q_13_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor q_13_squeeze_mask_0 = const()[name = tensor("q_13_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor q_13_cast_fp16 = slice_by_index(begin = q_13_begin_0, end = q_13_end_0, end_mask = q_13_end_mask_0, squeeze_mask = q_13_squeeze_mask_0, x = qkv_5_cast_fp16)[name = tensor("q_13_cast_fp16")]; + tensor k_9_begin_0 = const()[name = tensor("k_9_begin_0"), val = tensor([0, 0, 1, 0, 0])]; + tensor k_9_end_0 = const()[name = tensor("k_9_end_0"), val = tensor([1, 1, 2, 16, 64])]; + tensor k_9_end_mask_0 = const()[name = tensor("k_9_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor k_9_squeeze_mask_0 = const()[name = tensor("k_9_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor k_9_cast_fp16 = slice_by_index(begin = k_9_begin_0, end = k_9_end_0, end_mask = k_9_end_mask_0, squeeze_mask = k_9_squeeze_mask_0, x = qkv_5_cast_fp16)[name = tensor("k_9_cast_fp16")]; + tensor v_5_begin_0 = const()[name = tensor("v_5_begin_0"), val = tensor([0, 0, 2, 0, 0])]; + tensor v_5_end_0 = const()[name = tensor("v_5_end_0"), val = tensor([1, 1, 3, 16, 64])]; + tensor v_5_end_mask_0 = const()[name = tensor("v_5_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor v_5_squeeze_mask_0 = const()[name = tensor("v_5_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor v_5_cast_fp16 = slice_by_index(begin = v_5_begin_0, end = v_5_end_0, end_mask = v_5_end_mask_0, squeeze_mask = v_5_squeeze_mask_0, x = qkv_5_cast_fp16)[name = tensor("v_5_cast_fp16")]; + tensor var_1182 = const()[name = tensor("op_1182"), val = tensor([1, 1, 1, 1])]; + tensor position2_to_fp16_dtype_0 = const()[name = tensor("position2_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor position2_to_fp16 = cast(dtype = position2_to_fp16_dtype_0, x = position2)[name = tensor("cast_543")]; + tensor ts_17_cast_fp16 = reshape(shape = var_1182, x = position2_to_fp16)[name = tensor("ts_17_cast_fp16")]; + tensor var_1186 = const()[name = tensor("op_1186"), val = tensor([1, 1, 16, 32, 2])]; + tensor q_complex_5_cast_fp16 = reshape(shape = var_1186, x = q_13_cast_fp16)[name = tensor("q_complex_5_cast_fp16")]; + tensor var_1190 = const()[name = tensor("op_1190"), val = tensor([1, 1, 16, 32, 2])]; + tensor k_complex_5_cast_fp16 = reshape(shape = var_1190, x = k_9_cast_fp16)[name = tensor("k_complex_5_cast_fp16")]; + tensor var_1194_begin_0 = const()[name = tensor("op_1194_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_1194_end_0 = const()[name = tensor("op_1194_end_0"), val = tensor([1, 1, 16, 32, 1])]; + tensor var_1194_end_mask_0 = const()[name = tensor("op_1194_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_1194_squeeze_mask_0 = const()[name = tensor("op_1194_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_1194_cast_fp16 = slice_by_index(begin = var_1194_begin_0, end = var_1194_end_0, end_mask = var_1194_end_mask_0, squeeze_mask = var_1194_squeeze_mask_0, x = q_complex_5_cast_fp16)[name = tensor("op_1194_cast_fp16")]; + tensor var_1202_begin_0 = const()[name = tensor("op_1202_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_1202_end_0 = const()[name = tensor("op_1202_end_0"), val = tensor([1, 1, 16, 32, 2])]; + tensor var_1202_end_mask_0 = const()[name = tensor("op_1202_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_1202_squeeze_mask_0 = const()[name = tensor("op_1202_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_1202_cast_fp16 = slice_by_index(begin = var_1202_begin_0, end = var_1202_end_0, end_mask = var_1202_end_mask_0, squeeze_mask = var_1202_squeeze_mask_0, x = q_complex_5_cast_fp16)[name = tensor("op_1202_cast_fp16")]; + tensor var_1210_begin_0 = const()[name = tensor("op_1210_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_1210_end_0 = const()[name = tensor("op_1210_end_0"), val = tensor([1, 1, 16, 32, 1])]; + tensor var_1210_end_mask_0 = const()[name = tensor("op_1210_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_1210_squeeze_mask_0 = const()[name = tensor("op_1210_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_1210_cast_fp16 = slice_by_index(begin = var_1210_begin_0, end = var_1210_end_0, end_mask = var_1210_end_mask_0, squeeze_mask = var_1210_squeeze_mask_0, x = k_complex_5_cast_fp16)[name = tensor("op_1210_cast_fp16")]; + tensor var_1218_begin_0 = const()[name = tensor("op_1218_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_1218_end_0 = const()[name = tensor("op_1218_end_0"), val = tensor([1, 1, 16, 32, 2])]; + tensor var_1218_end_mask_0 = const()[name = tensor("op_1218_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_1218_squeeze_mask_0 = const()[name = tensor("op_1218_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_1218_cast_fp16 = slice_by_index(begin = var_1218_begin_0, end = var_1218_end_0, end_mask = var_1218_end_mask_0, squeeze_mask = var_1218_squeeze_mask_0, x = k_complex_5_cast_fp16)[name = tensor("op_1218_cast_fp16")]; + tensor freqs_5_to_fp16 = const()[name = tensor("freqs_5_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(303606272)))]; + tensor var_1224_cast_fp16 = mul(x = freqs_5_to_fp16, y = ts_17_cast_fp16)[name = tensor("op_1224_cast_fp16")]; + tensor rotr_5_cast_fp16 = cos(x = var_1224_cast_fp16)[name = tensor("rotr_5_cast_fp16")]; + tensor roti_5_cast_fp16 = sin(x = var_1224_cast_fp16)[name = tensor("roti_5_cast_fp16")]; + tensor var_1228_cast_fp16 = mul(x = var_1194_cast_fp16, y = rotr_5_cast_fp16)[name = tensor("op_1228_cast_fp16")]; + tensor var_1229_cast_fp16 = mul(x = var_1202_cast_fp16, y = roti_5_cast_fp16)[name = tensor("op_1229_cast_fp16")]; + tensor qor_9_cast_fp16 = sub(x = var_1228_cast_fp16, y = var_1229_cast_fp16)[name = tensor("qor_9_cast_fp16")]; + tensor var_1232_cast_fp16 = mul(x = var_1194_cast_fp16, y = roti_5_cast_fp16)[name = tensor("op_1232_cast_fp16")]; + tensor var_1233_cast_fp16 = mul(x = var_1202_cast_fp16, y = rotr_5_cast_fp16)[name = tensor("op_1233_cast_fp16")]; + tensor qoi_9_cast_fp16 = add(x = var_1232_cast_fp16, y = var_1233_cast_fp16)[name = tensor("qoi_9_cast_fp16")]; + tensor var_1236_cast_fp16 = mul(x = var_1210_cast_fp16, y = rotr_5_cast_fp16)[name = tensor("op_1236_cast_fp16")]; + tensor var_1237_cast_fp16 = mul(x = var_1218_cast_fp16, y = roti_5_cast_fp16)[name = tensor("op_1237_cast_fp16")]; + tensor kor_9_cast_fp16 = sub(x = var_1236_cast_fp16, y = var_1237_cast_fp16)[name = tensor("kor_9_cast_fp16")]; + tensor var_1240_cast_fp16 = mul(x = var_1210_cast_fp16, y = roti_5_cast_fp16)[name = tensor("op_1240_cast_fp16")]; + tensor var_1241_cast_fp16 = mul(x = var_1218_cast_fp16, y = rotr_5_cast_fp16)[name = tensor("op_1241_cast_fp16")]; + tensor koi_9_cast_fp16 = add(x = var_1240_cast_fp16, y = var_1241_cast_fp16)[name = tensor("koi_9_cast_fp16")]; + tensor qo_5_axis_0 = const()[name = tensor("qo_5_axis_0"), val = tensor(-1)]; + tensor qo_5_cast_fp16 = stack(axis = qo_5_axis_0, values = (qor_9_cast_fp16, qoi_9_cast_fp16))[name = tensor("qo_5_cast_fp16")]; + tensor ko_5_axis_0 = const()[name = tensor("ko_5_axis_0"), val = tensor(-1)]; + tensor ko_5_cast_fp16 = stack(axis = ko_5_axis_0, values = (kor_9_cast_fp16, koi_9_cast_fp16))[name = tensor("ko_5_cast_fp16")]; + tensor var_1270 = const()[name = tensor("op_1270"), val = tensor([1, 1, 16, 64])]; + tensor q_15_cast_fp16 = reshape(shape = var_1270, x = qo_5_cast_fp16)[name = tensor("q_15_cast_fp16")]; + tensor var_1272 = const()[name = tensor("op_1272"), val = tensor([1, 1, 16, 64])]; + tensor k_11_cast_fp16 = reshape(shape = var_1272, x = ko_5_cast_fp16)[name = tensor("k_11_cast_fp16")]; + tensor _inversed_1294_y_0_to_fp16 = const()[name = tensor("_inversed_1294_y_0_to_fp16"), val = tensor(0x1p-9)]; + tensor _inversed_1294_cast_fp16 = mul(x = ts_17_cast_fp16, y = _inversed_1294_y_0_to_fp16)[name = tensor("_inversed_1294_cast_fp16")]; + tensor var_1295_cast_fp16 = floor(x = _inversed_1294_cast_fp16)[name = tensor("op_1295_cast_fp16")]; + tensor var_1296_to_fp16 = const()[name = tensor("op_1296_to_fp16"), val = tensor(0x1p+9)]; + tensor var_1297_cast_fp16 = mul(x = var_1295_cast_fp16, y = var_1296_to_fp16)[name = tensor("op_1297_cast_fp16")]; + tensor write_indices_float_11_cast_fp16 = sub(x = ts_17_cast_fp16, y = var_1297_cast_fp16)[name = tensor("write_indices_float_11_cast_fp16")]; + tensor var_1304_dtype_0 = const()[name = tensor("op_1304_dtype_0"), val = tensor("int32")]; + tensor write_indices_5_reps_0 = const()[name = tensor("write_indices_5_reps_0"), val = tensor([1, 1, 16, 64])]; + tensor write_indices_float_11_cast_fp16_to_int32 = cast(dtype = var_1304_dtype_0, x = write_indices_float_11_cast_fp16)[name = tensor("cast_542")]; + tensor write_indices_5 = tile(reps = write_indices_5_reps_0, x = write_indices_float_11_cast_fp16_to_int32)[name = tensor("write_indices_5")]; + tensor var_1312_begin_0 = const()[name = tensor("op_1312_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_1312_end_0 = const()[name = tensor("op_1312_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor var_1312_end_mask_0 = const()[name = tensor("op_1312_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_1312_squeeze_mask_0 = const()[name = tensor("op_1312_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor cache2_to_fp16_dtype_0 = const()[name = tensor("cache2_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor cache2_to_fp16 = cast(dtype = cache2_to_fp16_dtype_0, x = cache2)[name = tensor("cast_541")]; + tensor var_1312_cast_fp16 = slice_by_index(begin = var_1312_begin_0, end = var_1312_end_0, end_mask = var_1312_end_mask_0, squeeze_mask = var_1312_squeeze_mask_0, x = cache2_to_fp16)[name = tensor("op_1312_cast_fp16")]; + tensor var_1314_axis_0 = const()[name = tensor("op_1314_axis_0"), val = tensor(1)]; + tensor var_1314_mode_0 = const()[name = tensor("op_1314_mode_0"), val = tensor("update")]; + tensor var_1314_validate_indices_0 = const()[name = tensor("op_1314_validate_indices_0"), val = tensor(false)]; + tensor var_1314_cast_fp16 = scatter_along_axis(axis = var_1314_axis_0, data = var_1312_cast_fp16, indices = write_indices_5, mode = var_1314_mode_0, updates = k_11_cast_fp16, validate_indices = var_1314_validate_indices_0)[name = tensor("op_1314_cast_fp16")]; + tensor concat_16 = const()[name = tensor("concat_16"), val = tensor([0, 0, 0, 0, 0])]; + tensor concat_17 = const()[name = tensor("concat_17"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_5_internal_tensor_assign_1_stride_0 = const()[name = tensor("new_cache_5_internal_tensor_assign_1_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_5_internal_tensor_assign_1_begin_mask_0 = const()[name = tensor("new_cache_5_internal_tensor_assign_1_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_5_internal_tensor_assign_1_end_mask_0 = const()[name = tensor("new_cache_5_internal_tensor_assign_1_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_5_internal_tensor_assign_1_squeeze_mask_0 = const()[name = tensor("new_cache_5_internal_tensor_assign_1_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_52 = const()[name = tensor("shape_52"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_4 = const()[name = tensor("reduce_prod_4"), val = tensor(1048576)]; + tensor range_1d_4_start_0 = const()[name = tensor("range_1d_4_start_0"), val = tensor(0)]; + tensor range_1d_4_step_0 = const()[name = tensor("range_1d_4_step_0"), val = tensor(1)]; + tensor range_1d_4 = range_1d(end = reduce_prod_4, start = range_1d_4_start_0, step = range_1d_4_step_0)[name = tensor("range_1d_4")]; + tensor reshape_20 = reshape(shape = shape_52, x = range_1d_4)[name = tensor("reshape_20")]; + tensor slice_by_index_4 = slice_by_index(begin = concat_16, begin_mask = new_cache_5_internal_tensor_assign_1_begin_mask_0, end = concat_17, end_mask = new_cache_5_internal_tensor_assign_1_end_mask_0, squeeze_mask = new_cache_5_internal_tensor_assign_1_squeeze_mask_0, stride = new_cache_5_internal_tensor_assign_1_stride_0, x = reshape_20)[name = tensor("slice_by_index_4")]; + tensor reshape_21_shape_0 = const()[name = tensor("reshape_21_shape_0"), val = tensor([-1])]; + tensor reshape_21 = reshape(shape = reshape_21_shape_0, x = slice_by_index_4)[name = tensor("reshape_21")]; + tensor reshape_22_shape_0 = const()[name = tensor("reshape_22_shape_0"), val = tensor([-1])]; + tensor reshape_22_cast_fp16 = reshape(shape = reshape_22_shape_0, x = var_1314_cast_fp16)[name = tensor("reshape_22_cast_fp16")]; + tensor reshape_23_shape_0 = const()[name = tensor("reshape_23_shape_0"), val = tensor([-1])]; + tensor reshape_23_cast_fp16 = reshape(shape = reshape_23_shape_0, x = cache2_to_fp16)[name = tensor("reshape_23_cast_fp16")]; + tensor scatter_4_mode_0 = const()[name = tensor("scatter_4_mode_0"), val = tensor("update")]; + tensor scatter_4_axis_0 = const()[name = tensor("scatter_4_axis_0"), val = tensor(0)]; + tensor scatter_4_validate_indices_0 = const()[name = tensor("scatter_4_validate_indices_0"), val = tensor(false)]; + tensor scatter_4_cast_fp16 = scatter(axis = scatter_4_axis_0, data = reshape_23_cast_fp16, indices = reshape_21, mode = scatter_4_mode_0, updates = reshape_22_cast_fp16, validate_indices = scatter_4_validate_indices_0)[name = tensor("scatter_4_cast_fp16")]; + tensor reshape_24_cast_fp16 = reshape(shape = shape_52, x = scatter_4_cast_fp16)[name = tensor("reshape_24_cast_fp16")]; + tensor var_1322_begin_0 = const()[name = tensor("op_1322_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor var_1322_end_0 = const()[name = tensor("op_1322_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor var_1322_end_mask_0 = const()[name = tensor("op_1322_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_1322_squeeze_mask_0 = const()[name = tensor("op_1322_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor var_1322_cast_fp16 = slice_by_index(begin = var_1322_begin_0, end = var_1322_end_0, end_mask = var_1322_end_mask_0, squeeze_mask = var_1322_squeeze_mask_0, x = reshape_24_cast_fp16)[name = tensor("op_1322_cast_fp16")]; + tensor var_1324_axis_0 = const()[name = tensor("op_1324_axis_0"), val = tensor(1)]; + tensor var_1324_mode_0 = const()[name = tensor("op_1324_mode_0"), val = tensor("update")]; + tensor var_1324_validate_indices_0 = const()[name = tensor("op_1324_validate_indices_0"), val = tensor(false)]; + tensor var_1324_cast_fp16 = scatter_along_axis(axis = var_1324_axis_0, data = var_1322_cast_fp16, indices = write_indices_5, mode = var_1324_mode_0, updates = v_5_cast_fp16, validate_indices = var_1324_validate_indices_0)[name = tensor("op_1324_cast_fp16")]; + tensor concat_18 = const()[name = tensor("concat_18"), val = tensor([1, 0, 0, 0, 0])]; + tensor concat_19 = const()[name = tensor("concat_19"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_5_internal_tensor_assign_2_stride_0 = const()[name = tensor("new_cache_5_internal_tensor_assign_2_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_5_internal_tensor_assign_2_begin_mask_0 = const()[name = tensor("new_cache_5_internal_tensor_assign_2_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_5_internal_tensor_assign_2_end_mask_0 = const()[name = tensor("new_cache_5_internal_tensor_assign_2_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_5_internal_tensor_assign_2_squeeze_mask_0 = const()[name = tensor("new_cache_5_internal_tensor_assign_2_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_53 = const()[name = tensor("shape_53"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_5 = const()[name = tensor("reduce_prod_5"), val = tensor(1048576)]; + tensor range_1d_5_start_0 = const()[name = tensor("range_1d_5_start_0"), val = tensor(0)]; + tensor range_1d_5_step_0 = const()[name = tensor("range_1d_5_step_0"), val = tensor(1)]; + tensor range_1d_5 = range_1d(end = reduce_prod_5, start = range_1d_5_start_0, step = range_1d_5_step_0)[name = tensor("range_1d_5")]; + tensor reshape_25 = reshape(shape = shape_53, x = range_1d_5)[name = tensor("reshape_25")]; + tensor slice_by_index_5 = slice_by_index(begin = concat_18, begin_mask = new_cache_5_internal_tensor_assign_2_begin_mask_0, end = concat_19, end_mask = new_cache_5_internal_tensor_assign_2_end_mask_0, squeeze_mask = new_cache_5_internal_tensor_assign_2_squeeze_mask_0, stride = new_cache_5_internal_tensor_assign_2_stride_0, x = reshape_25)[name = tensor("slice_by_index_5")]; + tensor reshape_26_shape_0 = const()[name = tensor("reshape_26_shape_0"), val = tensor([-1])]; + tensor reshape_26 = reshape(shape = reshape_26_shape_0, x = slice_by_index_5)[name = tensor("reshape_26")]; + tensor reshape_27_shape_0 = const()[name = tensor("reshape_27_shape_0"), val = tensor([-1])]; + tensor reshape_27_cast_fp16 = reshape(shape = reshape_27_shape_0, x = var_1324_cast_fp16)[name = tensor("reshape_27_cast_fp16")]; + tensor reshape_28_shape_0 = const()[name = tensor("reshape_28_shape_0"), val = tensor([-1])]; + tensor reshape_28_cast_fp16 = reshape(shape = reshape_28_shape_0, x = reshape_24_cast_fp16)[name = tensor("reshape_28_cast_fp16")]; + tensor scatter_5_mode_0 = const()[name = tensor("scatter_5_mode_0"), val = tensor("update")]; + tensor scatter_5_axis_0 = const()[name = tensor("scatter_5_axis_0"), val = tensor(0)]; + tensor scatter_5_validate_indices_0 = const()[name = tensor("scatter_5_validate_indices_0"), val = tensor(false)]; + tensor scatter_5_cast_fp16 = scatter(axis = scatter_5_axis_0, data = reshape_28_cast_fp16, indices = reshape_26, mode = scatter_5_mode_0, updates = reshape_27_cast_fp16, validate_indices = scatter_5_validate_indices_0)[name = tensor("scatter_5_cast_fp16")]; + tensor reshape_29_cast_fp16 = reshape(shape = shape_53, x = scatter_5_cast_fp16)[name = tensor("reshape_29_cast_fp16")]; + tensor reshape_29_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("reshape_29_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor keys_13_begin_0 = const()[name = tensor("keys_13_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor keys_13_end_0 = const()[name = tensor("keys_13_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor keys_13_end_mask_0 = const()[name = tensor("keys_13_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor keys_13_squeeze_mask_0 = const()[name = tensor("keys_13_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor keys_13_cast_fp16 = slice_by_index(begin = keys_13_begin_0, end = keys_13_end_0, end_mask = keys_13_end_mask_0, squeeze_mask = keys_13_squeeze_mask_0, x = reshape_29_cast_fp16)[name = tensor("keys_13_cast_fp16")]; + tensor values_13_begin_0 = const()[name = tensor("values_13_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor values_13_end_0 = const()[name = tensor("values_13_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor values_13_end_mask_0 = const()[name = tensor("values_13_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor values_13_squeeze_mask_0 = const()[name = tensor("values_13_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor values_13_cast_fp16 = slice_by_index(begin = values_13_begin_0, end = values_13_end_0, end_mask = values_13_end_mask_0, squeeze_mask = values_13_squeeze_mask_0, x = reshape_29_cast_fp16)[name = tensor("values_13_cast_fp16")]; + tensor var_1336_cast_fp16 = not_equal(x = keys_13_cast_fp16, y = keys_13_cast_fp16)[name = tensor("op_1336_cast_fp16")]; + tensor keys_15_cast_fp16 = select(a = var_504_to_fp16, b = keys_13_cast_fp16, cond = var_1336_cast_fp16)[name = tensor("keys_15_cast_fp16")]; + tensor var_1344_cast_fp16 = not_equal(x = values_13_cast_fp16, y = values_13_cast_fp16)[name = tensor("op_1344_cast_fp16")]; + tensor values_15_cast_fp16 = select(a = var_504_to_fp16, b = values_13_cast_fp16, cond = var_1344_cast_fp16)[name = tensor("values_15_cast_fp16")]; + tensor var_1368 = const()[name = tensor("op_1368"), val = tensor([0, 2, 1, 3])]; + tensor var_1381 = const()[name = tensor("op_1381"), val = tensor([1, 1, 1])]; + tensor var_1382_cast_fp16 = reshape(shape = var_1381, x = position2_to_fp16)[name = tensor("op_1382_cast_fp16")]; + tensor var_1399_to_fp16 = const()[name = tensor("op_1399_to_fp16"), val = tensor(0x1p+0)]; + tensor valid_len_5_cast_fp16 = add(x = var_1382_cast_fp16, y = var_1399_to_fp16)[name = tensor("valid_len_5_cast_fp16")]; + tensor valid_mask_5_cast_fp16 = less(x = k_positions_1_promoted_to_fp16, y = valid_len_5_cast_fp16)[name = tensor("valid_mask_5_cast_fp16")]; + tensor causal_mask_5_cast_fp16 = less_equal(x = k_positions_1_promoted_to_fp16, y = var_1382_cast_fp16)[name = tensor("causal_mask_5_cast_fp16")]; + tensor attn_mask_9 = logical_and(x = valid_mask_5_cast_fp16, y = causal_mask_5_cast_fp16)[name = tensor("attn_mask_9")]; + tensor attn_mask_11_axes_0 = const()[name = tensor("attn_mask_11_axes_0"), val = tensor([1])]; + tensor attn_mask_11 = expand_dims(axes = attn_mask_11_axes_0, x = attn_mask_9)[name = tensor("attn_mask_11")]; + tensor var_1417_transpose_x_0 = const()[name = tensor("op_1417_transpose_x_0"), val = tensor(false)]; + tensor var_1417_transpose_y_0 = const()[name = tensor("op_1417_transpose_y_0"), val = tensor(false)]; + tensor transpose_76_perm_0 = const()[name = tensor("transpose_76_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_77_perm_0 = const()[name = tensor("transpose_77_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_77 = transpose(perm = transpose_77_perm_0, x = keys_15_cast_fp16)[name = tensor("transpose_205")]; + tensor transpose_76 = transpose(perm = transpose_76_perm_0, x = q_15_cast_fp16)[name = tensor("transpose_206")]; + tensor var_1417_cast_fp16 = matmul(transpose_x = var_1417_transpose_x_0, transpose_y = var_1417_transpose_y_0, x = transpose_76, y = transpose_77)[name = tensor("op_1417_cast_fp16")]; + tensor var_1411_to_fp16 = const()[name = tensor("op_1411_to_fp16"), val = tensor([0x1p-3])]; + tensor attn_weights_13_cast_fp16 = mul(x = var_1417_cast_fp16, y = var_1411_to_fp16)[name = tensor("attn_weights_13_cast_fp16")]; + tensor var_1419 = logical_not(x = attn_mask_11)[name = tensor("op_1419")]; + tensor var_1420_to_fp16 = const()[name = tensor("op_1420_to_fp16"), val = tensor(-inf)]; + tensor attn_weights_15_cast_fp16 = select(a = var_1420_to_fp16, b = attn_weights_13_cast_fp16, cond = var_1419)[name = tensor("attn_weights_15_cast_fp16")]; + tensor var_1422 = const()[name = tensor("op_1422"), val = tensor(-1)]; + tensor attn_weights_17_cast_fp16 = softmax(axis = var_1422, x = attn_weights_15_cast_fp16)[name = tensor("attn_weights_17_cast_fp16")]; + tensor attn_output_5_transpose_x_0 = const()[name = tensor("attn_output_5_transpose_x_0"), val = tensor(false)]; + tensor attn_output_5_transpose_y_0 = const()[name = tensor("attn_output_5_transpose_y_0"), val = tensor(false)]; + tensor values_17_cast_fp16 = transpose(perm = var_1368, x = values_15_cast_fp16)[name = tensor("transpose_207")]; + tensor attn_output_5_cast_fp16 = matmul(transpose_x = attn_output_5_transpose_x_0, transpose_y = attn_output_5_transpose_y_0, x = attn_weights_17_cast_fp16, y = values_17_cast_fp16)[name = tensor("attn_output_5_cast_fp16")]; + tensor var_1430 = const()[name = tensor("op_1430"), val = tensor([0, 2, 1, 3])]; + tensor var_1433 = const()[name = tensor("op_1433"), val = tensor([1, 1, 1024])]; + tensor var_1431_cast_fp16 = transpose(perm = var_1430, x = attn_output_5_cast_fp16)[name = tensor("transpose_204")]; + tensor input_25_cast_fp16 = reshape(shape = var_1433, x = var_1431_cast_fp16)[name = tensor("input_25_cast_fp16")]; + tensor linear_10_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = attn2_out_proj_weight_cast_fp16, x = input_25_cast_fp16)[name = tensor("linear_10_cast_fp16")]; + tensor var_1439_to_fp16 = const()[name = tensor("op_1439_to_fp16"), val = tensor(0x1p+0)]; + tensor var_1440_cast_fp16 = add(x = position2_to_fp16, y = var_1439_to_fp16)[name = tensor("op_1440_cast_fp16")]; + tensor var_1440_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_1440_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor input_27_cast_fp16 = add(x = input_23_cast_fp16, y = linear_10_cast_fp16)[name = tensor("input_27_cast_fp16")]; + tensor input_29_axes_0 = const()[name = tensor("input_29_axes_0"), val = tensor([-1])]; + tensor norm2_2_weight_to_fp16 = const()[name = tensor("norm2_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(303606400)))]; + tensor norm2_2_bias_to_fp16 = const()[name = tensor("norm2_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(303608512)))]; + tensor var_1444_to_fp16 = const()[name = tensor("op_1444_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_29_cast_fp16 = layer_norm(axes = input_29_axes_0, beta = norm2_2_bias_to_fp16, epsilon = var_1444_to_fp16, gamma = norm2_2_weight_to_fp16, x = input_27_cast_fp16)[name = tensor("input_29_cast_fp16")]; + tensor linear_11_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = linear2_1_weight_cast_fp16, x = input_29_cast_fp16)[name = tensor("linear_11_cast_fp16")]; + tensor input_31_mode_0 = const()[name = tensor("input_31_mode_0"), val = tensor("EXACT")]; + tensor input_31_cast_fp16 = gelu(mode = input_31_mode_0, x = linear_11_cast_fp16)[name = tensor("input_31_cast_fp16")]; + tensor linear_12_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = linear2_2_weight_cast_fp16, x = input_31_cast_fp16)[name = tensor("linear_12_cast_fp16")]; + tensor input_33_cast_fp16 = add(x = input_27_cast_fp16, y = linear_12_cast_fp16)[name = tensor("input_33_cast_fp16")]; + tensor x_7_axes_0 = const()[name = tensor("x_7_axes_0"), val = tensor([-1])]; + tensor norm3_1_weight_to_fp16 = const()[name = tensor("norm3_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(303610624)))]; + tensor norm3_1_bias_to_fp16 = const()[name = tensor("norm3_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(303612736)))]; + tensor var_1461_to_fp16 = const()[name = tensor("op_1461_to_fp16"), val = tensor(0x1.5p-17)]; + tensor x_7_cast_fp16 = layer_norm(axes = x_7_axes_0, beta = norm3_1_bias_to_fp16, epsilon = var_1461_to_fp16, gamma = norm3_1_weight_to_fp16, x = input_33_cast_fp16)[name = tensor("x_7_cast_fp16")]; + tensor linear_13_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = attn3_in_proj_weight_cast_fp16, x = x_7_cast_fp16)[name = tensor("linear_13_cast_fp16")]; + tensor var_1497 = const()[name = tensor("op_1497"), val = tensor([1, 1, 3, 16, 64])]; + tensor qkv_7_cast_fp16 = reshape(shape = var_1497, x = linear_13_cast_fp16)[name = tensor("qkv_7_cast_fp16")]; + tensor q_19_begin_0 = const()[name = tensor("q_19_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor q_19_end_0 = const()[name = tensor("q_19_end_0"), val = tensor([1, 1, 1, 16, 64])]; + tensor q_19_end_mask_0 = const()[name = tensor("q_19_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor q_19_squeeze_mask_0 = const()[name = tensor("q_19_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor q_19_cast_fp16 = slice_by_index(begin = q_19_begin_0, end = q_19_end_0, end_mask = q_19_end_mask_0, squeeze_mask = q_19_squeeze_mask_0, x = qkv_7_cast_fp16)[name = tensor("q_19_cast_fp16")]; + tensor k_13_begin_0 = const()[name = tensor("k_13_begin_0"), val = tensor([0, 0, 1, 0, 0])]; + tensor k_13_end_0 = const()[name = tensor("k_13_end_0"), val = tensor([1, 1, 2, 16, 64])]; + tensor k_13_end_mask_0 = const()[name = tensor("k_13_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor k_13_squeeze_mask_0 = const()[name = tensor("k_13_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor k_13_cast_fp16 = slice_by_index(begin = k_13_begin_0, end = k_13_end_0, end_mask = k_13_end_mask_0, squeeze_mask = k_13_squeeze_mask_0, x = qkv_7_cast_fp16)[name = tensor("k_13_cast_fp16")]; + tensor v_7_begin_0 = const()[name = tensor("v_7_begin_0"), val = tensor([0, 0, 2, 0, 0])]; + tensor v_7_end_0 = const()[name = tensor("v_7_end_0"), val = tensor([1, 1, 3, 16, 64])]; + tensor v_7_end_mask_0 = const()[name = tensor("v_7_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor v_7_squeeze_mask_0 = const()[name = tensor("v_7_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor v_7_cast_fp16 = slice_by_index(begin = v_7_begin_0, end = v_7_end_0, end_mask = v_7_end_mask_0, squeeze_mask = v_7_squeeze_mask_0, x = qkv_7_cast_fp16)[name = tensor("v_7_cast_fp16")]; + tensor var_1601 = const()[name = tensor("op_1601"), val = tensor([1, 1, 1, 1])]; + tensor position3_to_fp16_dtype_0 = const()[name = tensor("position3_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor position3_to_fp16 = cast(dtype = position3_to_fp16_dtype_0, x = position3)[name = tensor("cast_538")]; + tensor ts_23_cast_fp16 = reshape(shape = var_1601, x = position3_to_fp16)[name = tensor("ts_23_cast_fp16")]; + tensor var_1605 = const()[name = tensor("op_1605"), val = tensor([1, 1, 16, 32, 2])]; + tensor q_complex_7_cast_fp16 = reshape(shape = var_1605, x = q_19_cast_fp16)[name = tensor("q_complex_7_cast_fp16")]; + tensor var_1609 = const()[name = tensor("op_1609"), val = tensor([1, 1, 16, 32, 2])]; + tensor k_complex_7_cast_fp16 = reshape(shape = var_1609, x = k_13_cast_fp16)[name = tensor("k_complex_7_cast_fp16")]; + tensor var_1613_begin_0 = const()[name = tensor("op_1613_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_1613_end_0 = const()[name = tensor("op_1613_end_0"), val = tensor([1, 1, 16, 32, 1])]; + tensor var_1613_end_mask_0 = const()[name = tensor("op_1613_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_1613_squeeze_mask_0 = const()[name = tensor("op_1613_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_1613_cast_fp16 = slice_by_index(begin = var_1613_begin_0, end = var_1613_end_0, end_mask = var_1613_end_mask_0, squeeze_mask = var_1613_squeeze_mask_0, x = q_complex_7_cast_fp16)[name = tensor("op_1613_cast_fp16")]; + tensor var_1621_begin_0 = const()[name = tensor("op_1621_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_1621_end_0 = const()[name = tensor("op_1621_end_0"), val = tensor([1, 1, 16, 32, 2])]; + tensor var_1621_end_mask_0 = const()[name = tensor("op_1621_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_1621_squeeze_mask_0 = const()[name = tensor("op_1621_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_1621_cast_fp16 = slice_by_index(begin = var_1621_begin_0, end = var_1621_end_0, end_mask = var_1621_end_mask_0, squeeze_mask = var_1621_squeeze_mask_0, x = q_complex_7_cast_fp16)[name = tensor("op_1621_cast_fp16")]; + tensor var_1629_begin_0 = const()[name = tensor("op_1629_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_1629_end_0 = const()[name = tensor("op_1629_end_0"), val = tensor([1, 1, 16, 32, 1])]; + tensor var_1629_end_mask_0 = const()[name = tensor("op_1629_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_1629_squeeze_mask_0 = const()[name = tensor("op_1629_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_1629_cast_fp16 = slice_by_index(begin = var_1629_begin_0, end = var_1629_end_0, end_mask = var_1629_end_mask_0, squeeze_mask = var_1629_squeeze_mask_0, x = k_complex_7_cast_fp16)[name = tensor("op_1629_cast_fp16")]; + tensor var_1637_begin_0 = const()[name = tensor("op_1637_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_1637_end_0 = const()[name = tensor("op_1637_end_0"), val = tensor([1, 1, 16, 32, 2])]; + tensor var_1637_end_mask_0 = const()[name = tensor("op_1637_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_1637_squeeze_mask_0 = const()[name = tensor("op_1637_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_1637_cast_fp16 = slice_by_index(begin = var_1637_begin_0, end = var_1637_end_0, end_mask = var_1637_end_mask_0, squeeze_mask = var_1637_squeeze_mask_0, x = k_complex_7_cast_fp16)[name = tensor("op_1637_cast_fp16")]; + tensor freqs_7_to_fp16 = const()[name = tensor("freqs_7_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(303614848)))]; + tensor var_1643_cast_fp16 = mul(x = freqs_7_to_fp16, y = ts_23_cast_fp16)[name = tensor("op_1643_cast_fp16")]; + tensor rotr_7_cast_fp16 = cos(x = var_1643_cast_fp16)[name = tensor("rotr_7_cast_fp16")]; + tensor roti_7_cast_fp16 = sin(x = var_1643_cast_fp16)[name = tensor("roti_7_cast_fp16")]; + tensor var_1647_cast_fp16 = mul(x = var_1613_cast_fp16, y = rotr_7_cast_fp16)[name = tensor("op_1647_cast_fp16")]; + tensor var_1648_cast_fp16 = mul(x = var_1621_cast_fp16, y = roti_7_cast_fp16)[name = tensor("op_1648_cast_fp16")]; + tensor qor_13_cast_fp16 = sub(x = var_1647_cast_fp16, y = var_1648_cast_fp16)[name = tensor("qor_13_cast_fp16")]; + tensor var_1651_cast_fp16 = mul(x = var_1613_cast_fp16, y = roti_7_cast_fp16)[name = tensor("op_1651_cast_fp16")]; + tensor var_1652_cast_fp16 = mul(x = var_1621_cast_fp16, y = rotr_7_cast_fp16)[name = tensor("op_1652_cast_fp16")]; + tensor qoi_13_cast_fp16 = add(x = var_1651_cast_fp16, y = var_1652_cast_fp16)[name = tensor("qoi_13_cast_fp16")]; + tensor var_1655_cast_fp16 = mul(x = var_1629_cast_fp16, y = rotr_7_cast_fp16)[name = tensor("op_1655_cast_fp16")]; + tensor var_1656_cast_fp16 = mul(x = var_1637_cast_fp16, y = roti_7_cast_fp16)[name = tensor("op_1656_cast_fp16")]; + tensor kor_13_cast_fp16 = sub(x = var_1655_cast_fp16, y = var_1656_cast_fp16)[name = tensor("kor_13_cast_fp16")]; + tensor var_1659_cast_fp16 = mul(x = var_1629_cast_fp16, y = roti_7_cast_fp16)[name = tensor("op_1659_cast_fp16")]; + tensor var_1660_cast_fp16 = mul(x = var_1637_cast_fp16, y = rotr_7_cast_fp16)[name = tensor("op_1660_cast_fp16")]; + tensor koi_13_cast_fp16 = add(x = var_1659_cast_fp16, y = var_1660_cast_fp16)[name = tensor("koi_13_cast_fp16")]; + tensor qo_7_axis_0 = const()[name = tensor("qo_7_axis_0"), val = tensor(-1)]; + tensor qo_7_cast_fp16 = stack(axis = qo_7_axis_0, values = (qor_13_cast_fp16, qoi_13_cast_fp16))[name = tensor("qo_7_cast_fp16")]; + tensor ko_7_axis_0 = const()[name = tensor("ko_7_axis_0"), val = tensor(-1)]; + tensor ko_7_cast_fp16 = stack(axis = ko_7_axis_0, values = (kor_13_cast_fp16, koi_13_cast_fp16))[name = tensor("ko_7_cast_fp16")]; + tensor var_1689 = const()[name = tensor("op_1689"), val = tensor([1, 1, 16, 64])]; + tensor q_21_cast_fp16 = reshape(shape = var_1689, x = qo_7_cast_fp16)[name = tensor("q_21_cast_fp16")]; + tensor var_1691 = const()[name = tensor("op_1691"), val = tensor([1, 1, 16, 64])]; + tensor k_15_cast_fp16 = reshape(shape = var_1691, x = ko_7_cast_fp16)[name = tensor("k_15_cast_fp16")]; + tensor _inversed_1713_y_0_to_fp16 = const()[name = tensor("_inversed_1713_y_0_to_fp16"), val = tensor(0x1p-9)]; + tensor _inversed_1713_cast_fp16 = mul(x = ts_23_cast_fp16, y = _inversed_1713_y_0_to_fp16)[name = tensor("_inversed_1713_cast_fp16")]; + tensor var_1714_cast_fp16 = floor(x = _inversed_1713_cast_fp16)[name = tensor("op_1714_cast_fp16")]; + tensor var_1715_to_fp16 = const()[name = tensor("op_1715_to_fp16"), val = tensor(0x1p+9)]; + tensor var_1716_cast_fp16 = mul(x = var_1714_cast_fp16, y = var_1715_to_fp16)[name = tensor("op_1716_cast_fp16")]; + tensor write_indices_float_15_cast_fp16 = sub(x = ts_23_cast_fp16, y = var_1716_cast_fp16)[name = tensor("write_indices_float_15_cast_fp16")]; + tensor var_1723_dtype_0 = const()[name = tensor("op_1723_dtype_0"), val = tensor("int32")]; + tensor write_indices_7_reps_0 = const()[name = tensor("write_indices_7_reps_0"), val = tensor([1, 1, 16, 64])]; + tensor write_indices_float_15_cast_fp16_to_int32 = cast(dtype = var_1723_dtype_0, x = write_indices_float_15_cast_fp16)[name = tensor("cast_537")]; + tensor write_indices_7 = tile(reps = write_indices_7_reps_0, x = write_indices_float_15_cast_fp16_to_int32)[name = tensor("write_indices_7")]; + tensor var_1731_begin_0 = const()[name = tensor("op_1731_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_1731_end_0 = const()[name = tensor("op_1731_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor var_1731_end_mask_0 = const()[name = tensor("op_1731_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_1731_squeeze_mask_0 = const()[name = tensor("op_1731_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor cache3_to_fp16_dtype_0 = const()[name = tensor("cache3_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor cache3_to_fp16 = cast(dtype = cache3_to_fp16_dtype_0, x = cache3)[name = tensor("cast_536")]; + tensor var_1731_cast_fp16 = slice_by_index(begin = var_1731_begin_0, end = var_1731_end_0, end_mask = var_1731_end_mask_0, squeeze_mask = var_1731_squeeze_mask_0, x = cache3_to_fp16)[name = tensor("op_1731_cast_fp16")]; + tensor var_1733_axis_0 = const()[name = tensor("op_1733_axis_0"), val = tensor(1)]; + tensor var_1733_mode_0 = const()[name = tensor("op_1733_mode_0"), val = tensor("update")]; + tensor var_1733_validate_indices_0 = const()[name = tensor("op_1733_validate_indices_0"), val = tensor(false)]; + tensor var_1733_cast_fp16 = scatter_along_axis(axis = var_1733_axis_0, data = var_1731_cast_fp16, indices = write_indices_7, mode = var_1733_mode_0, updates = k_15_cast_fp16, validate_indices = var_1733_validate_indices_0)[name = tensor("op_1733_cast_fp16")]; + tensor concat_23 = const()[name = tensor("concat_23"), val = tensor([0, 0, 0, 0, 0])]; + tensor concat_24 = const()[name = tensor("concat_24"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_7_internal_tensor_assign_1_stride_0 = const()[name = tensor("new_cache_7_internal_tensor_assign_1_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_7_internal_tensor_assign_1_begin_mask_0 = const()[name = tensor("new_cache_7_internal_tensor_assign_1_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_7_internal_tensor_assign_1_end_mask_0 = const()[name = tensor("new_cache_7_internal_tensor_assign_1_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_7_internal_tensor_assign_1_squeeze_mask_0 = const()[name = tensor("new_cache_7_internal_tensor_assign_1_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_54 = const()[name = tensor("shape_54"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_6 = const()[name = tensor("reduce_prod_6"), val = tensor(1048576)]; + tensor range_1d_6_start_0 = const()[name = tensor("range_1d_6_start_0"), val = tensor(0)]; + tensor range_1d_6_step_0 = const()[name = tensor("range_1d_6_step_0"), val = tensor(1)]; + tensor range_1d_6 = range_1d(end = reduce_prod_6, start = range_1d_6_start_0, step = range_1d_6_step_0)[name = tensor("range_1d_6")]; + tensor reshape_30 = reshape(shape = shape_54, x = range_1d_6)[name = tensor("reshape_30")]; + tensor slice_by_index_6 = slice_by_index(begin = concat_23, begin_mask = new_cache_7_internal_tensor_assign_1_begin_mask_0, end = concat_24, end_mask = new_cache_7_internal_tensor_assign_1_end_mask_0, squeeze_mask = new_cache_7_internal_tensor_assign_1_squeeze_mask_0, stride = new_cache_7_internal_tensor_assign_1_stride_0, x = reshape_30)[name = tensor("slice_by_index_6")]; + tensor reshape_31_shape_0 = const()[name = tensor("reshape_31_shape_0"), val = tensor([-1])]; + tensor reshape_31 = reshape(shape = reshape_31_shape_0, x = slice_by_index_6)[name = tensor("reshape_31")]; + tensor reshape_32_shape_0 = const()[name = tensor("reshape_32_shape_0"), val = tensor([-1])]; + tensor reshape_32_cast_fp16 = reshape(shape = reshape_32_shape_0, x = var_1733_cast_fp16)[name = tensor("reshape_32_cast_fp16")]; + tensor reshape_33_shape_0 = const()[name = tensor("reshape_33_shape_0"), val = tensor([-1])]; + tensor reshape_33_cast_fp16 = reshape(shape = reshape_33_shape_0, x = cache3_to_fp16)[name = tensor("reshape_33_cast_fp16")]; + tensor scatter_6_mode_0 = const()[name = tensor("scatter_6_mode_0"), val = tensor("update")]; + tensor scatter_6_axis_0 = const()[name = tensor("scatter_6_axis_0"), val = tensor(0)]; + tensor scatter_6_validate_indices_0 = const()[name = tensor("scatter_6_validate_indices_0"), val = tensor(false)]; + tensor scatter_6_cast_fp16 = scatter(axis = scatter_6_axis_0, data = reshape_33_cast_fp16, indices = reshape_31, mode = scatter_6_mode_0, updates = reshape_32_cast_fp16, validate_indices = scatter_6_validate_indices_0)[name = tensor("scatter_6_cast_fp16")]; + tensor reshape_34_cast_fp16 = reshape(shape = shape_54, x = scatter_6_cast_fp16)[name = tensor("reshape_34_cast_fp16")]; + tensor var_1741_begin_0 = const()[name = tensor("op_1741_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor var_1741_end_0 = const()[name = tensor("op_1741_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor var_1741_end_mask_0 = const()[name = tensor("op_1741_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_1741_squeeze_mask_0 = const()[name = tensor("op_1741_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor var_1741_cast_fp16 = slice_by_index(begin = var_1741_begin_0, end = var_1741_end_0, end_mask = var_1741_end_mask_0, squeeze_mask = var_1741_squeeze_mask_0, x = reshape_34_cast_fp16)[name = tensor("op_1741_cast_fp16")]; + tensor var_1743_axis_0 = const()[name = tensor("op_1743_axis_0"), val = tensor(1)]; + tensor var_1743_mode_0 = const()[name = tensor("op_1743_mode_0"), val = tensor("update")]; + tensor var_1743_validate_indices_0 = const()[name = tensor("op_1743_validate_indices_0"), val = tensor(false)]; + tensor var_1743_cast_fp16 = scatter_along_axis(axis = var_1743_axis_0, data = var_1741_cast_fp16, indices = write_indices_7, mode = var_1743_mode_0, updates = v_7_cast_fp16, validate_indices = var_1743_validate_indices_0)[name = tensor("op_1743_cast_fp16")]; + tensor concat_25 = const()[name = tensor("concat_25"), val = tensor([1, 0, 0, 0, 0])]; + tensor concat_26 = const()[name = tensor("concat_26"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_7_internal_tensor_assign_2_stride_0 = const()[name = tensor("new_cache_7_internal_tensor_assign_2_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_7_internal_tensor_assign_2_begin_mask_0 = const()[name = tensor("new_cache_7_internal_tensor_assign_2_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_7_internal_tensor_assign_2_end_mask_0 = const()[name = tensor("new_cache_7_internal_tensor_assign_2_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_7_internal_tensor_assign_2_squeeze_mask_0 = const()[name = tensor("new_cache_7_internal_tensor_assign_2_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_55 = const()[name = tensor("shape_55"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_7 = const()[name = tensor("reduce_prod_7"), val = tensor(1048576)]; + tensor range_1d_7_start_0 = const()[name = tensor("range_1d_7_start_0"), val = tensor(0)]; + tensor range_1d_7_step_0 = const()[name = tensor("range_1d_7_step_0"), val = tensor(1)]; + tensor range_1d_7 = range_1d(end = reduce_prod_7, start = range_1d_7_start_0, step = range_1d_7_step_0)[name = tensor("range_1d_7")]; + tensor reshape_35 = reshape(shape = shape_55, x = range_1d_7)[name = tensor("reshape_35")]; + tensor slice_by_index_7 = slice_by_index(begin = concat_25, begin_mask = new_cache_7_internal_tensor_assign_2_begin_mask_0, end = concat_26, end_mask = new_cache_7_internal_tensor_assign_2_end_mask_0, squeeze_mask = new_cache_7_internal_tensor_assign_2_squeeze_mask_0, stride = new_cache_7_internal_tensor_assign_2_stride_0, x = reshape_35)[name = tensor("slice_by_index_7")]; + tensor reshape_36_shape_0 = const()[name = tensor("reshape_36_shape_0"), val = tensor([-1])]; + tensor reshape_36 = reshape(shape = reshape_36_shape_0, x = slice_by_index_7)[name = tensor("reshape_36")]; + tensor reshape_37_shape_0 = const()[name = tensor("reshape_37_shape_0"), val = tensor([-1])]; + tensor reshape_37_cast_fp16 = reshape(shape = reshape_37_shape_0, x = var_1743_cast_fp16)[name = tensor("reshape_37_cast_fp16")]; + tensor reshape_38_shape_0 = const()[name = tensor("reshape_38_shape_0"), val = tensor([-1])]; + tensor reshape_38_cast_fp16 = reshape(shape = reshape_38_shape_0, x = reshape_34_cast_fp16)[name = tensor("reshape_38_cast_fp16")]; + tensor scatter_7_mode_0 = const()[name = tensor("scatter_7_mode_0"), val = tensor("update")]; + tensor scatter_7_axis_0 = const()[name = tensor("scatter_7_axis_0"), val = tensor(0)]; + tensor scatter_7_validate_indices_0 = const()[name = tensor("scatter_7_validate_indices_0"), val = tensor(false)]; + tensor scatter_7_cast_fp16 = scatter(axis = scatter_7_axis_0, data = reshape_38_cast_fp16, indices = reshape_36, mode = scatter_7_mode_0, updates = reshape_37_cast_fp16, validate_indices = scatter_7_validate_indices_0)[name = tensor("scatter_7_cast_fp16")]; + tensor reshape_39_cast_fp16 = reshape(shape = shape_55, x = scatter_7_cast_fp16)[name = tensor("reshape_39_cast_fp16")]; + tensor reshape_39_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("reshape_39_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor keys_19_begin_0 = const()[name = tensor("keys_19_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor keys_19_end_0 = const()[name = tensor("keys_19_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor keys_19_end_mask_0 = const()[name = tensor("keys_19_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor keys_19_squeeze_mask_0 = const()[name = tensor("keys_19_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor keys_19_cast_fp16 = slice_by_index(begin = keys_19_begin_0, end = keys_19_end_0, end_mask = keys_19_end_mask_0, squeeze_mask = keys_19_squeeze_mask_0, x = reshape_39_cast_fp16)[name = tensor("keys_19_cast_fp16")]; + tensor values_19_begin_0 = const()[name = tensor("values_19_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor values_19_end_0 = const()[name = tensor("values_19_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor values_19_end_mask_0 = const()[name = tensor("values_19_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor values_19_squeeze_mask_0 = const()[name = tensor("values_19_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor values_19_cast_fp16 = slice_by_index(begin = values_19_begin_0, end = values_19_end_0, end_mask = values_19_end_mask_0, squeeze_mask = values_19_squeeze_mask_0, x = reshape_39_cast_fp16)[name = tensor("values_19_cast_fp16")]; + tensor var_1755_cast_fp16 = not_equal(x = keys_19_cast_fp16, y = keys_19_cast_fp16)[name = tensor("op_1755_cast_fp16")]; + tensor keys_21_cast_fp16 = select(a = var_504_to_fp16, b = keys_19_cast_fp16, cond = var_1755_cast_fp16)[name = tensor("keys_21_cast_fp16")]; + tensor var_1763_cast_fp16 = not_equal(x = values_19_cast_fp16, y = values_19_cast_fp16)[name = tensor("op_1763_cast_fp16")]; + tensor values_21_cast_fp16 = select(a = var_504_to_fp16, b = values_19_cast_fp16, cond = var_1763_cast_fp16)[name = tensor("values_21_cast_fp16")]; + tensor var_1787 = const()[name = tensor("op_1787"), val = tensor([0, 2, 1, 3])]; + tensor var_1800 = const()[name = tensor("op_1800"), val = tensor([1, 1, 1])]; + tensor var_1801_cast_fp16 = reshape(shape = var_1800, x = position3_to_fp16)[name = tensor("op_1801_cast_fp16")]; + tensor var_1818_to_fp16 = const()[name = tensor("op_1818_to_fp16"), val = tensor(0x1p+0)]; + tensor valid_len_7_cast_fp16 = add(x = var_1801_cast_fp16, y = var_1818_to_fp16)[name = tensor("valid_len_7_cast_fp16")]; + tensor valid_mask_7_cast_fp16 = less(x = k_positions_1_promoted_to_fp16, y = valid_len_7_cast_fp16)[name = tensor("valid_mask_7_cast_fp16")]; + tensor causal_mask_7_cast_fp16 = less_equal(x = k_positions_1_promoted_to_fp16, y = var_1801_cast_fp16)[name = tensor("causal_mask_7_cast_fp16")]; + tensor attn_mask_13 = logical_and(x = valid_mask_7_cast_fp16, y = causal_mask_7_cast_fp16)[name = tensor("attn_mask_13")]; + tensor attn_mask_15_axes_0 = const()[name = tensor("attn_mask_15_axes_0"), val = tensor([1])]; + tensor attn_mask_15 = expand_dims(axes = attn_mask_15_axes_0, x = attn_mask_13)[name = tensor("attn_mask_15")]; + tensor var_1836_transpose_x_0 = const()[name = tensor("op_1836_transpose_x_0"), val = tensor(false)]; + tensor var_1836_transpose_y_0 = const()[name = tensor("op_1836_transpose_y_0"), val = tensor(false)]; + tensor transpose_78_perm_0 = const()[name = tensor("transpose_78_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_79_perm_0 = const()[name = tensor("transpose_79_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_79 = transpose(perm = transpose_79_perm_0, x = keys_21_cast_fp16)[name = tensor("transpose_201")]; + tensor transpose_78 = transpose(perm = transpose_78_perm_0, x = q_21_cast_fp16)[name = tensor("transpose_202")]; + tensor var_1836_cast_fp16 = matmul(transpose_x = var_1836_transpose_x_0, transpose_y = var_1836_transpose_y_0, x = transpose_78, y = transpose_79)[name = tensor("op_1836_cast_fp16")]; + tensor var_1830_to_fp16 = const()[name = tensor("op_1830_to_fp16"), val = tensor([0x1p-3])]; + tensor attn_weights_19_cast_fp16 = mul(x = var_1836_cast_fp16, y = var_1830_to_fp16)[name = tensor("attn_weights_19_cast_fp16")]; + tensor var_1838 = logical_not(x = attn_mask_15)[name = tensor("op_1838")]; + tensor var_1839_to_fp16 = const()[name = tensor("op_1839_to_fp16"), val = tensor(-inf)]; + tensor attn_weights_21_cast_fp16 = select(a = var_1839_to_fp16, b = attn_weights_19_cast_fp16, cond = var_1838)[name = tensor("attn_weights_21_cast_fp16")]; + tensor var_1841 = const()[name = tensor("op_1841"), val = tensor(-1)]; + tensor attn_weights_23_cast_fp16 = softmax(axis = var_1841, x = attn_weights_21_cast_fp16)[name = tensor("attn_weights_23_cast_fp16")]; + tensor attn_output_7_transpose_x_0 = const()[name = tensor("attn_output_7_transpose_x_0"), val = tensor(false)]; + tensor attn_output_7_transpose_y_0 = const()[name = tensor("attn_output_7_transpose_y_0"), val = tensor(false)]; + tensor values_23_cast_fp16 = transpose(perm = var_1787, x = values_21_cast_fp16)[name = tensor("transpose_203")]; + tensor attn_output_7_cast_fp16 = matmul(transpose_x = attn_output_7_transpose_x_0, transpose_y = attn_output_7_transpose_y_0, x = attn_weights_23_cast_fp16, y = values_23_cast_fp16)[name = tensor("attn_output_7_cast_fp16")]; + tensor var_1849 = const()[name = tensor("op_1849"), val = tensor([0, 2, 1, 3])]; + tensor var_1852 = const()[name = tensor("op_1852"), val = tensor([1, 1, 1024])]; + tensor var_1850_cast_fp16 = transpose(perm = var_1849, x = attn_output_7_cast_fp16)[name = tensor("transpose_200")]; + tensor input_35_cast_fp16 = reshape(shape = var_1852, x = var_1850_cast_fp16)[name = tensor("input_35_cast_fp16")]; + tensor linear_14_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = attn3_out_proj_weight_cast_fp16, x = input_35_cast_fp16)[name = tensor("linear_14_cast_fp16")]; + tensor var_1858_to_fp16 = const()[name = tensor("op_1858_to_fp16"), val = tensor(0x1p+0)]; + tensor var_1859_cast_fp16 = add(x = position3_to_fp16, y = var_1858_to_fp16)[name = tensor("op_1859_cast_fp16")]; + tensor var_1859_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_1859_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor input_37_cast_fp16 = add(x = input_33_cast_fp16, y = linear_14_cast_fp16)[name = tensor("input_37_cast_fp16")]; + tensor input_39_axes_0 = const()[name = tensor("input_39_axes_0"), val = tensor([-1])]; + tensor norm3_2_weight_to_fp16 = const()[name = tensor("norm3_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(303614976)))]; + tensor norm3_2_bias_to_fp16 = const()[name = tensor("norm3_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(303617088)))]; + tensor var_1863_to_fp16 = const()[name = tensor("op_1863_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_39_cast_fp16 = layer_norm(axes = input_39_axes_0, beta = norm3_2_bias_to_fp16, epsilon = var_1863_to_fp16, gamma = norm3_2_weight_to_fp16, x = input_37_cast_fp16)[name = tensor("input_39_cast_fp16")]; + tensor linear_15_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = linear3_1_weight_cast_fp16, x = input_39_cast_fp16)[name = tensor("linear_15_cast_fp16")]; + tensor input_41_mode_0 = const()[name = tensor("input_41_mode_0"), val = tensor("EXACT")]; + tensor input_41_cast_fp16 = gelu(mode = input_41_mode_0, x = linear_15_cast_fp16)[name = tensor("input_41_cast_fp16")]; + tensor linear_16_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = linear3_2_weight_cast_fp16, x = input_41_cast_fp16)[name = tensor("linear_16_cast_fp16")]; + tensor input_43_cast_fp16 = add(x = input_37_cast_fp16, y = linear_16_cast_fp16)[name = tensor("input_43_cast_fp16")]; + tensor x_9_axes_0 = const()[name = tensor("x_9_axes_0"), val = tensor([-1])]; + tensor norm4_1_weight_to_fp16 = const()[name = tensor("norm4_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(303619200)))]; + tensor norm4_1_bias_to_fp16 = const()[name = tensor("norm4_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(303621312)))]; + tensor var_1880_to_fp16 = const()[name = tensor("op_1880_to_fp16"), val = tensor(0x1.5p-17)]; + tensor x_9_cast_fp16 = layer_norm(axes = x_9_axes_0, beta = norm4_1_bias_to_fp16, epsilon = var_1880_to_fp16, gamma = norm4_1_weight_to_fp16, x = input_43_cast_fp16)[name = tensor("x_9_cast_fp16")]; + tensor linear_17_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = attn4_in_proj_weight_cast_fp16, x = x_9_cast_fp16)[name = tensor("linear_17_cast_fp16")]; + tensor var_1916 = const()[name = tensor("op_1916"), val = tensor([1, 1, 3, 16, 64])]; + tensor qkv_9_cast_fp16 = reshape(shape = var_1916, x = linear_17_cast_fp16)[name = tensor("qkv_9_cast_fp16")]; + tensor q_25_begin_0 = const()[name = tensor("q_25_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor q_25_end_0 = const()[name = tensor("q_25_end_0"), val = tensor([1, 1, 1, 16, 64])]; + tensor q_25_end_mask_0 = const()[name = tensor("q_25_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor q_25_squeeze_mask_0 = const()[name = tensor("q_25_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor q_25_cast_fp16 = slice_by_index(begin = q_25_begin_0, end = q_25_end_0, end_mask = q_25_end_mask_0, squeeze_mask = q_25_squeeze_mask_0, x = qkv_9_cast_fp16)[name = tensor("q_25_cast_fp16")]; + tensor k_17_begin_0 = const()[name = tensor("k_17_begin_0"), val = tensor([0, 0, 1, 0, 0])]; + tensor k_17_end_0 = const()[name = tensor("k_17_end_0"), val = tensor([1, 1, 2, 16, 64])]; + tensor k_17_end_mask_0 = const()[name = tensor("k_17_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor k_17_squeeze_mask_0 = const()[name = tensor("k_17_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor k_17_cast_fp16 = slice_by_index(begin = k_17_begin_0, end = k_17_end_0, end_mask = k_17_end_mask_0, squeeze_mask = k_17_squeeze_mask_0, x = qkv_9_cast_fp16)[name = tensor("k_17_cast_fp16")]; + tensor v_9_begin_0 = const()[name = tensor("v_9_begin_0"), val = tensor([0, 0, 2, 0, 0])]; + tensor v_9_end_0 = const()[name = tensor("v_9_end_0"), val = tensor([1, 1, 3, 16, 64])]; + tensor v_9_end_mask_0 = const()[name = tensor("v_9_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor v_9_squeeze_mask_0 = const()[name = tensor("v_9_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor v_9_cast_fp16 = slice_by_index(begin = v_9_begin_0, end = v_9_end_0, end_mask = v_9_end_mask_0, squeeze_mask = v_9_squeeze_mask_0, x = qkv_9_cast_fp16)[name = tensor("v_9_cast_fp16")]; + tensor var_2020 = const()[name = tensor("op_2020"), val = tensor([1, 1, 1, 1])]; + tensor position4_to_fp16_dtype_0 = const()[name = tensor("position4_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor position4_to_fp16 = cast(dtype = position4_to_fp16_dtype_0, x = position4)[name = tensor("cast_533")]; + tensor ts_29_cast_fp16 = reshape(shape = var_2020, x = position4_to_fp16)[name = tensor("ts_29_cast_fp16")]; + tensor var_2024 = const()[name = tensor("op_2024"), val = tensor([1, 1, 16, 32, 2])]; + tensor q_complex_9_cast_fp16 = reshape(shape = var_2024, x = q_25_cast_fp16)[name = tensor("q_complex_9_cast_fp16")]; + tensor var_2028 = const()[name = tensor("op_2028"), val = tensor([1, 1, 16, 32, 2])]; + tensor k_complex_9_cast_fp16 = reshape(shape = var_2028, x = k_17_cast_fp16)[name = tensor("k_complex_9_cast_fp16")]; + tensor var_2032_begin_0 = const()[name = tensor("op_2032_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_2032_end_0 = const()[name = tensor("op_2032_end_0"), val = tensor([1, 1, 16, 32, 1])]; + tensor var_2032_end_mask_0 = const()[name = tensor("op_2032_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_2032_squeeze_mask_0 = const()[name = tensor("op_2032_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_2032_cast_fp16 = slice_by_index(begin = var_2032_begin_0, end = var_2032_end_0, end_mask = var_2032_end_mask_0, squeeze_mask = var_2032_squeeze_mask_0, x = q_complex_9_cast_fp16)[name = tensor("op_2032_cast_fp16")]; + tensor var_2040_begin_0 = const()[name = tensor("op_2040_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_2040_end_0 = const()[name = tensor("op_2040_end_0"), val = tensor([1, 1, 16, 32, 2])]; + tensor var_2040_end_mask_0 = const()[name = tensor("op_2040_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_2040_squeeze_mask_0 = const()[name = tensor("op_2040_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_2040_cast_fp16 = slice_by_index(begin = var_2040_begin_0, end = var_2040_end_0, end_mask = var_2040_end_mask_0, squeeze_mask = var_2040_squeeze_mask_0, x = q_complex_9_cast_fp16)[name = tensor("op_2040_cast_fp16")]; + tensor var_2048_begin_0 = const()[name = tensor("op_2048_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_2048_end_0 = const()[name = tensor("op_2048_end_0"), val = tensor([1, 1, 16, 32, 1])]; + tensor var_2048_end_mask_0 = const()[name = tensor("op_2048_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_2048_squeeze_mask_0 = const()[name = tensor("op_2048_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_2048_cast_fp16 = slice_by_index(begin = var_2048_begin_0, end = var_2048_end_0, end_mask = var_2048_end_mask_0, squeeze_mask = var_2048_squeeze_mask_0, x = k_complex_9_cast_fp16)[name = tensor("op_2048_cast_fp16")]; + tensor var_2056_begin_0 = const()[name = tensor("op_2056_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_2056_end_0 = const()[name = tensor("op_2056_end_0"), val = tensor([1, 1, 16, 32, 2])]; + tensor var_2056_end_mask_0 = const()[name = tensor("op_2056_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_2056_squeeze_mask_0 = const()[name = tensor("op_2056_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_2056_cast_fp16 = slice_by_index(begin = var_2056_begin_0, end = var_2056_end_0, end_mask = var_2056_end_mask_0, squeeze_mask = var_2056_squeeze_mask_0, x = k_complex_9_cast_fp16)[name = tensor("op_2056_cast_fp16")]; + tensor freqs_9_to_fp16 = const()[name = tensor("freqs_9_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(303623424)))]; + tensor var_2062_cast_fp16 = mul(x = freqs_9_to_fp16, y = ts_29_cast_fp16)[name = tensor("op_2062_cast_fp16")]; + tensor rotr_9_cast_fp16 = cos(x = var_2062_cast_fp16)[name = tensor("rotr_9_cast_fp16")]; + tensor roti_9_cast_fp16 = sin(x = var_2062_cast_fp16)[name = tensor("roti_9_cast_fp16")]; + tensor var_2066_cast_fp16 = mul(x = var_2032_cast_fp16, y = rotr_9_cast_fp16)[name = tensor("op_2066_cast_fp16")]; + tensor var_2067_cast_fp16 = mul(x = var_2040_cast_fp16, y = roti_9_cast_fp16)[name = tensor("op_2067_cast_fp16")]; + tensor qor_17_cast_fp16 = sub(x = var_2066_cast_fp16, y = var_2067_cast_fp16)[name = tensor("qor_17_cast_fp16")]; + tensor var_2070_cast_fp16 = mul(x = var_2032_cast_fp16, y = roti_9_cast_fp16)[name = tensor("op_2070_cast_fp16")]; + tensor var_2071_cast_fp16 = mul(x = var_2040_cast_fp16, y = rotr_9_cast_fp16)[name = tensor("op_2071_cast_fp16")]; + tensor qoi_17_cast_fp16 = add(x = var_2070_cast_fp16, y = var_2071_cast_fp16)[name = tensor("qoi_17_cast_fp16")]; + tensor var_2074_cast_fp16 = mul(x = var_2048_cast_fp16, y = rotr_9_cast_fp16)[name = tensor("op_2074_cast_fp16")]; + tensor var_2075_cast_fp16 = mul(x = var_2056_cast_fp16, y = roti_9_cast_fp16)[name = tensor("op_2075_cast_fp16")]; + tensor kor_17_cast_fp16 = sub(x = var_2074_cast_fp16, y = var_2075_cast_fp16)[name = tensor("kor_17_cast_fp16")]; + tensor var_2078_cast_fp16 = mul(x = var_2048_cast_fp16, y = roti_9_cast_fp16)[name = tensor("op_2078_cast_fp16")]; + tensor var_2079_cast_fp16 = mul(x = var_2056_cast_fp16, y = rotr_9_cast_fp16)[name = tensor("op_2079_cast_fp16")]; + tensor koi_17_cast_fp16 = add(x = var_2078_cast_fp16, y = var_2079_cast_fp16)[name = tensor("koi_17_cast_fp16")]; + tensor qo_9_axis_0 = const()[name = tensor("qo_9_axis_0"), val = tensor(-1)]; + tensor qo_9_cast_fp16 = stack(axis = qo_9_axis_0, values = (qor_17_cast_fp16, qoi_17_cast_fp16))[name = tensor("qo_9_cast_fp16")]; + tensor ko_9_axis_0 = const()[name = tensor("ko_9_axis_0"), val = tensor(-1)]; + tensor ko_9_cast_fp16 = stack(axis = ko_9_axis_0, values = (kor_17_cast_fp16, koi_17_cast_fp16))[name = tensor("ko_9_cast_fp16")]; + tensor var_2108 = const()[name = tensor("op_2108"), val = tensor([1, 1, 16, 64])]; + tensor q_27_cast_fp16 = reshape(shape = var_2108, x = qo_9_cast_fp16)[name = tensor("q_27_cast_fp16")]; + tensor var_2110 = const()[name = tensor("op_2110"), val = tensor([1, 1, 16, 64])]; + tensor k_19_cast_fp16 = reshape(shape = var_2110, x = ko_9_cast_fp16)[name = tensor("k_19_cast_fp16")]; + tensor _inversed_2132_y_0_to_fp16 = const()[name = tensor("_inversed_2132_y_0_to_fp16"), val = tensor(0x1p-9)]; + tensor _inversed_2132_cast_fp16 = mul(x = ts_29_cast_fp16, y = _inversed_2132_y_0_to_fp16)[name = tensor("_inversed_2132_cast_fp16")]; + tensor var_2133_cast_fp16 = floor(x = _inversed_2132_cast_fp16)[name = tensor("op_2133_cast_fp16")]; + tensor var_2134_to_fp16 = const()[name = tensor("op_2134_to_fp16"), val = tensor(0x1p+9)]; + tensor var_2135_cast_fp16 = mul(x = var_2133_cast_fp16, y = var_2134_to_fp16)[name = tensor("op_2135_cast_fp16")]; + tensor write_indices_float_19_cast_fp16 = sub(x = ts_29_cast_fp16, y = var_2135_cast_fp16)[name = tensor("write_indices_float_19_cast_fp16")]; + tensor var_2142_dtype_0 = const()[name = tensor("op_2142_dtype_0"), val = tensor("int32")]; + tensor write_indices_9_reps_0 = const()[name = tensor("write_indices_9_reps_0"), val = tensor([1, 1, 16, 64])]; + tensor write_indices_float_19_cast_fp16_to_int32 = cast(dtype = var_2142_dtype_0, x = write_indices_float_19_cast_fp16)[name = tensor("cast_532")]; + tensor write_indices_9 = tile(reps = write_indices_9_reps_0, x = write_indices_float_19_cast_fp16_to_int32)[name = tensor("write_indices_9")]; + tensor var_2150_begin_0 = const()[name = tensor("op_2150_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_2150_end_0 = const()[name = tensor("op_2150_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor var_2150_end_mask_0 = const()[name = tensor("op_2150_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_2150_squeeze_mask_0 = const()[name = tensor("op_2150_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor cache4_to_fp16_dtype_0 = const()[name = tensor("cache4_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor cache4_to_fp16 = cast(dtype = cache4_to_fp16_dtype_0, x = cache4)[name = tensor("cast_531")]; + tensor var_2150_cast_fp16 = slice_by_index(begin = var_2150_begin_0, end = var_2150_end_0, end_mask = var_2150_end_mask_0, squeeze_mask = var_2150_squeeze_mask_0, x = cache4_to_fp16)[name = tensor("op_2150_cast_fp16")]; + tensor var_2152_axis_0 = const()[name = tensor("op_2152_axis_0"), val = tensor(1)]; + tensor var_2152_mode_0 = const()[name = tensor("op_2152_mode_0"), val = tensor("update")]; + tensor var_2152_validate_indices_0 = const()[name = tensor("op_2152_validate_indices_0"), val = tensor(false)]; + tensor var_2152_cast_fp16 = scatter_along_axis(axis = var_2152_axis_0, data = var_2150_cast_fp16, indices = write_indices_9, mode = var_2152_mode_0, updates = k_19_cast_fp16, validate_indices = var_2152_validate_indices_0)[name = tensor("op_2152_cast_fp16")]; + tensor concat_30 = const()[name = tensor("concat_30"), val = tensor([0, 0, 0, 0, 0])]; + tensor concat_31 = const()[name = tensor("concat_31"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_9_internal_tensor_assign_1_stride_0 = const()[name = tensor("new_cache_9_internal_tensor_assign_1_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_9_internal_tensor_assign_1_begin_mask_0 = const()[name = tensor("new_cache_9_internal_tensor_assign_1_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_9_internal_tensor_assign_1_end_mask_0 = const()[name = tensor("new_cache_9_internal_tensor_assign_1_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_9_internal_tensor_assign_1_squeeze_mask_0 = const()[name = tensor("new_cache_9_internal_tensor_assign_1_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_56 = const()[name = tensor("shape_56"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_8 = const()[name = tensor("reduce_prod_8"), val = tensor(1048576)]; + tensor range_1d_8_start_0 = const()[name = tensor("range_1d_8_start_0"), val = tensor(0)]; + tensor range_1d_8_step_0 = const()[name = tensor("range_1d_8_step_0"), val = tensor(1)]; + tensor range_1d_8 = range_1d(end = reduce_prod_8, start = range_1d_8_start_0, step = range_1d_8_step_0)[name = tensor("range_1d_8")]; + tensor reshape_40 = reshape(shape = shape_56, x = range_1d_8)[name = tensor("reshape_40")]; + tensor slice_by_index_8 = slice_by_index(begin = concat_30, begin_mask = new_cache_9_internal_tensor_assign_1_begin_mask_0, end = concat_31, end_mask = new_cache_9_internal_tensor_assign_1_end_mask_0, squeeze_mask = new_cache_9_internal_tensor_assign_1_squeeze_mask_0, stride = new_cache_9_internal_tensor_assign_1_stride_0, x = reshape_40)[name = tensor("slice_by_index_8")]; + tensor reshape_41_shape_0 = const()[name = tensor("reshape_41_shape_0"), val = tensor([-1])]; + tensor reshape_41 = reshape(shape = reshape_41_shape_0, x = slice_by_index_8)[name = tensor("reshape_41")]; + tensor reshape_42_shape_0 = const()[name = tensor("reshape_42_shape_0"), val = tensor([-1])]; + tensor reshape_42_cast_fp16 = reshape(shape = reshape_42_shape_0, x = var_2152_cast_fp16)[name = tensor("reshape_42_cast_fp16")]; + tensor reshape_43_shape_0 = const()[name = tensor("reshape_43_shape_0"), val = tensor([-1])]; + tensor reshape_43_cast_fp16 = reshape(shape = reshape_43_shape_0, x = cache4_to_fp16)[name = tensor("reshape_43_cast_fp16")]; + tensor scatter_8_mode_0 = const()[name = tensor("scatter_8_mode_0"), val = tensor("update")]; + tensor scatter_8_axis_0 = const()[name = tensor("scatter_8_axis_0"), val = tensor(0)]; + tensor scatter_8_validate_indices_0 = const()[name = tensor("scatter_8_validate_indices_0"), val = tensor(false)]; + tensor scatter_8_cast_fp16 = scatter(axis = scatter_8_axis_0, data = reshape_43_cast_fp16, indices = reshape_41, mode = scatter_8_mode_0, updates = reshape_42_cast_fp16, validate_indices = scatter_8_validate_indices_0)[name = tensor("scatter_8_cast_fp16")]; + tensor reshape_44_cast_fp16 = reshape(shape = shape_56, x = scatter_8_cast_fp16)[name = tensor("reshape_44_cast_fp16")]; + tensor var_2160_begin_0 = const()[name = tensor("op_2160_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor var_2160_end_0 = const()[name = tensor("op_2160_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor var_2160_end_mask_0 = const()[name = tensor("op_2160_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_2160_squeeze_mask_0 = const()[name = tensor("op_2160_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor var_2160_cast_fp16 = slice_by_index(begin = var_2160_begin_0, end = var_2160_end_0, end_mask = var_2160_end_mask_0, squeeze_mask = var_2160_squeeze_mask_0, x = reshape_44_cast_fp16)[name = tensor("op_2160_cast_fp16")]; + tensor var_2162_axis_0 = const()[name = tensor("op_2162_axis_0"), val = tensor(1)]; + tensor var_2162_mode_0 = const()[name = tensor("op_2162_mode_0"), val = tensor("update")]; + tensor var_2162_validate_indices_0 = const()[name = tensor("op_2162_validate_indices_0"), val = tensor(false)]; + tensor var_2162_cast_fp16 = scatter_along_axis(axis = var_2162_axis_0, data = var_2160_cast_fp16, indices = write_indices_9, mode = var_2162_mode_0, updates = v_9_cast_fp16, validate_indices = var_2162_validate_indices_0)[name = tensor("op_2162_cast_fp16")]; + tensor concat_32 = const()[name = tensor("concat_32"), val = tensor([1, 0, 0, 0, 0])]; + tensor concat_33 = const()[name = tensor("concat_33"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_9_internal_tensor_assign_2_stride_0 = const()[name = tensor("new_cache_9_internal_tensor_assign_2_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_9_internal_tensor_assign_2_begin_mask_0 = const()[name = tensor("new_cache_9_internal_tensor_assign_2_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_9_internal_tensor_assign_2_end_mask_0 = const()[name = tensor("new_cache_9_internal_tensor_assign_2_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_9_internal_tensor_assign_2_squeeze_mask_0 = const()[name = tensor("new_cache_9_internal_tensor_assign_2_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_57 = const()[name = tensor("shape_57"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_9 = const()[name = tensor("reduce_prod_9"), val = tensor(1048576)]; + tensor range_1d_9_start_0 = const()[name = tensor("range_1d_9_start_0"), val = tensor(0)]; + tensor range_1d_9_step_0 = const()[name = tensor("range_1d_9_step_0"), val = tensor(1)]; + tensor range_1d_9 = range_1d(end = reduce_prod_9, start = range_1d_9_start_0, step = range_1d_9_step_0)[name = tensor("range_1d_9")]; + tensor reshape_45 = reshape(shape = shape_57, x = range_1d_9)[name = tensor("reshape_45")]; + tensor slice_by_index_9 = slice_by_index(begin = concat_32, begin_mask = new_cache_9_internal_tensor_assign_2_begin_mask_0, end = concat_33, end_mask = new_cache_9_internal_tensor_assign_2_end_mask_0, squeeze_mask = new_cache_9_internal_tensor_assign_2_squeeze_mask_0, stride = new_cache_9_internal_tensor_assign_2_stride_0, x = reshape_45)[name = tensor("slice_by_index_9")]; + tensor reshape_46_shape_0 = const()[name = tensor("reshape_46_shape_0"), val = tensor([-1])]; + tensor reshape_46 = reshape(shape = reshape_46_shape_0, x = slice_by_index_9)[name = tensor("reshape_46")]; + tensor reshape_47_shape_0 = const()[name = tensor("reshape_47_shape_0"), val = tensor([-1])]; + tensor reshape_47_cast_fp16 = reshape(shape = reshape_47_shape_0, x = var_2162_cast_fp16)[name = tensor("reshape_47_cast_fp16")]; + tensor reshape_48_shape_0 = const()[name = tensor("reshape_48_shape_0"), val = tensor([-1])]; + tensor reshape_48_cast_fp16 = reshape(shape = reshape_48_shape_0, x = reshape_44_cast_fp16)[name = tensor("reshape_48_cast_fp16")]; + tensor scatter_9_mode_0 = const()[name = tensor("scatter_9_mode_0"), val = tensor("update")]; + tensor scatter_9_axis_0 = const()[name = tensor("scatter_9_axis_0"), val = tensor(0)]; + tensor scatter_9_validate_indices_0 = const()[name = tensor("scatter_9_validate_indices_0"), val = tensor(false)]; + tensor scatter_9_cast_fp16 = scatter(axis = scatter_9_axis_0, data = reshape_48_cast_fp16, indices = reshape_46, mode = scatter_9_mode_0, updates = reshape_47_cast_fp16, validate_indices = scatter_9_validate_indices_0)[name = tensor("scatter_9_cast_fp16")]; + tensor reshape_49_cast_fp16 = reshape(shape = shape_57, x = scatter_9_cast_fp16)[name = tensor("reshape_49_cast_fp16")]; + tensor reshape_49_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("reshape_49_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor keys_25_begin_0 = const()[name = tensor("keys_25_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor keys_25_end_0 = const()[name = tensor("keys_25_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor keys_25_end_mask_0 = const()[name = tensor("keys_25_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor keys_25_squeeze_mask_0 = const()[name = tensor("keys_25_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor keys_25_cast_fp16 = slice_by_index(begin = keys_25_begin_0, end = keys_25_end_0, end_mask = keys_25_end_mask_0, squeeze_mask = keys_25_squeeze_mask_0, x = reshape_49_cast_fp16)[name = tensor("keys_25_cast_fp16")]; + tensor values_25_begin_0 = const()[name = tensor("values_25_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor values_25_end_0 = const()[name = tensor("values_25_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor values_25_end_mask_0 = const()[name = tensor("values_25_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor values_25_squeeze_mask_0 = const()[name = tensor("values_25_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor values_25_cast_fp16 = slice_by_index(begin = values_25_begin_0, end = values_25_end_0, end_mask = values_25_end_mask_0, squeeze_mask = values_25_squeeze_mask_0, x = reshape_49_cast_fp16)[name = tensor("values_25_cast_fp16")]; + tensor var_2174_cast_fp16 = not_equal(x = keys_25_cast_fp16, y = keys_25_cast_fp16)[name = tensor("op_2174_cast_fp16")]; + tensor keys_27_cast_fp16 = select(a = var_504_to_fp16, b = keys_25_cast_fp16, cond = var_2174_cast_fp16)[name = tensor("keys_27_cast_fp16")]; + tensor var_2182_cast_fp16 = not_equal(x = values_25_cast_fp16, y = values_25_cast_fp16)[name = tensor("op_2182_cast_fp16")]; + tensor values_27_cast_fp16 = select(a = var_504_to_fp16, b = values_25_cast_fp16, cond = var_2182_cast_fp16)[name = tensor("values_27_cast_fp16")]; + tensor var_2206 = const()[name = tensor("op_2206"), val = tensor([0, 2, 1, 3])]; + tensor var_2219 = const()[name = tensor("op_2219"), val = tensor([1, 1, 1])]; + tensor var_2220_cast_fp16 = reshape(shape = var_2219, x = position4_to_fp16)[name = tensor("op_2220_cast_fp16")]; + tensor var_2237_to_fp16 = const()[name = tensor("op_2237_to_fp16"), val = tensor(0x1p+0)]; + tensor valid_len_9_cast_fp16 = add(x = var_2220_cast_fp16, y = var_2237_to_fp16)[name = tensor("valid_len_9_cast_fp16")]; + tensor valid_mask_9_cast_fp16 = less(x = k_positions_1_promoted_to_fp16, y = valid_len_9_cast_fp16)[name = tensor("valid_mask_9_cast_fp16")]; + tensor causal_mask_9_cast_fp16 = less_equal(x = k_positions_1_promoted_to_fp16, y = var_2220_cast_fp16)[name = tensor("causal_mask_9_cast_fp16")]; + tensor attn_mask_17 = logical_and(x = valid_mask_9_cast_fp16, y = causal_mask_9_cast_fp16)[name = tensor("attn_mask_17")]; + tensor attn_mask_19_axes_0 = const()[name = tensor("attn_mask_19_axes_0"), val = tensor([1])]; + tensor attn_mask_19 = expand_dims(axes = attn_mask_19_axes_0, x = attn_mask_17)[name = tensor("attn_mask_19")]; + tensor var_2255_transpose_x_0 = const()[name = tensor("op_2255_transpose_x_0"), val = tensor(false)]; + tensor var_2255_transpose_y_0 = const()[name = tensor("op_2255_transpose_y_0"), val = tensor(false)]; + tensor transpose_80_perm_0 = const()[name = tensor("transpose_80_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_81_perm_0 = const()[name = tensor("transpose_81_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_81 = transpose(perm = transpose_81_perm_0, x = keys_27_cast_fp16)[name = tensor("transpose_197")]; + tensor transpose_80 = transpose(perm = transpose_80_perm_0, x = q_27_cast_fp16)[name = tensor("transpose_198")]; + tensor var_2255_cast_fp16 = matmul(transpose_x = var_2255_transpose_x_0, transpose_y = var_2255_transpose_y_0, x = transpose_80, y = transpose_81)[name = tensor("op_2255_cast_fp16")]; + tensor var_2249_to_fp16 = const()[name = tensor("op_2249_to_fp16"), val = tensor([0x1p-3])]; + tensor attn_weights_25_cast_fp16 = mul(x = var_2255_cast_fp16, y = var_2249_to_fp16)[name = tensor("attn_weights_25_cast_fp16")]; + tensor var_2257 = logical_not(x = attn_mask_19)[name = tensor("op_2257")]; + tensor var_2258_to_fp16 = const()[name = tensor("op_2258_to_fp16"), val = tensor(-inf)]; + tensor attn_weights_27_cast_fp16 = select(a = var_2258_to_fp16, b = attn_weights_25_cast_fp16, cond = var_2257)[name = tensor("attn_weights_27_cast_fp16")]; + tensor var_2260 = const()[name = tensor("op_2260"), val = tensor(-1)]; + tensor attn_weights_29_cast_fp16 = softmax(axis = var_2260, x = attn_weights_27_cast_fp16)[name = tensor("attn_weights_29_cast_fp16")]; + tensor attn_output_9_transpose_x_0 = const()[name = tensor("attn_output_9_transpose_x_0"), val = tensor(false)]; + tensor attn_output_9_transpose_y_0 = const()[name = tensor("attn_output_9_transpose_y_0"), val = tensor(false)]; + tensor values_29_cast_fp16 = transpose(perm = var_2206, x = values_27_cast_fp16)[name = tensor("transpose_199")]; + tensor attn_output_9_cast_fp16 = matmul(transpose_x = attn_output_9_transpose_x_0, transpose_y = attn_output_9_transpose_y_0, x = attn_weights_29_cast_fp16, y = values_29_cast_fp16)[name = tensor("attn_output_9_cast_fp16")]; + tensor var_2268 = const()[name = tensor("op_2268"), val = tensor([0, 2, 1, 3])]; + tensor var_2271 = const()[name = tensor("op_2271"), val = tensor([1, 1, 1024])]; + tensor var_2269_cast_fp16 = transpose(perm = var_2268, x = attn_output_9_cast_fp16)[name = tensor("transpose_196")]; + tensor input_45_cast_fp16 = reshape(shape = var_2271, x = var_2269_cast_fp16)[name = tensor("input_45_cast_fp16")]; + tensor linear_18_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = attn4_out_proj_weight_cast_fp16, x = input_45_cast_fp16)[name = tensor("linear_18_cast_fp16")]; + tensor var_2277_to_fp16 = const()[name = tensor("op_2277_to_fp16"), val = tensor(0x1p+0)]; + tensor var_2278_cast_fp16 = add(x = position4_to_fp16, y = var_2277_to_fp16)[name = tensor("op_2278_cast_fp16")]; + tensor var_2278_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_2278_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor input_47_cast_fp16 = add(x = input_43_cast_fp16, y = linear_18_cast_fp16)[name = tensor("input_47_cast_fp16")]; + tensor input_49_axes_0 = const()[name = tensor("input_49_axes_0"), val = tensor([-1])]; + tensor norm4_2_weight_to_fp16 = const()[name = tensor("norm4_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(303623552)))]; + tensor norm4_2_bias_to_fp16 = const()[name = tensor("norm4_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(303625664)))]; + tensor var_2282_to_fp16 = const()[name = tensor("op_2282_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_49_cast_fp16 = layer_norm(axes = input_49_axes_0, beta = norm4_2_bias_to_fp16, epsilon = var_2282_to_fp16, gamma = norm4_2_weight_to_fp16, x = input_47_cast_fp16)[name = tensor("input_49_cast_fp16")]; + tensor linear_19_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = linear4_1_weight_cast_fp16, x = input_49_cast_fp16)[name = tensor("linear_19_cast_fp16")]; + tensor input_51_mode_0 = const()[name = tensor("input_51_mode_0"), val = tensor("EXACT")]; + tensor input_51_cast_fp16 = gelu(mode = input_51_mode_0, x = linear_19_cast_fp16)[name = tensor("input_51_cast_fp16")]; + tensor linear_20_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = linear4_2_weight_cast_fp16, x = input_51_cast_fp16)[name = tensor("linear_20_cast_fp16")]; + tensor input_53_cast_fp16 = add(x = input_47_cast_fp16, y = linear_20_cast_fp16)[name = tensor("input_53_cast_fp16")]; + tensor x_11_axes_0 = const()[name = tensor("x_11_axes_0"), val = tensor([-1])]; + tensor norm5_1_weight_to_fp16 = const()[name = tensor("norm5_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(303627776)))]; + tensor norm5_1_bias_to_fp16 = const()[name = tensor("norm5_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(303629888)))]; + tensor var_2299_to_fp16 = const()[name = tensor("op_2299_to_fp16"), val = tensor(0x1.5p-17)]; + tensor x_11_cast_fp16 = layer_norm(axes = x_11_axes_0, beta = norm5_1_bias_to_fp16, epsilon = var_2299_to_fp16, gamma = norm5_1_weight_to_fp16, x = input_53_cast_fp16)[name = tensor("x_11_cast_fp16")]; + tensor linear_21_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = attn5_in_proj_weight_cast_fp16, x = x_11_cast_fp16)[name = tensor("linear_21_cast_fp16")]; + tensor var_2335 = const()[name = tensor("op_2335"), val = tensor([1, 1, 3, 16, 64])]; + tensor qkv_11_cast_fp16 = reshape(shape = var_2335, x = linear_21_cast_fp16)[name = tensor("qkv_11_cast_fp16")]; + tensor q_31_begin_0 = const()[name = tensor("q_31_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor q_31_end_0 = const()[name = tensor("q_31_end_0"), val = tensor([1, 1, 1, 16, 64])]; + tensor q_31_end_mask_0 = const()[name = tensor("q_31_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor q_31_squeeze_mask_0 = const()[name = tensor("q_31_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor q_31_cast_fp16 = slice_by_index(begin = q_31_begin_0, end = q_31_end_0, end_mask = q_31_end_mask_0, squeeze_mask = q_31_squeeze_mask_0, x = qkv_11_cast_fp16)[name = tensor("q_31_cast_fp16")]; + tensor k_21_begin_0 = const()[name = tensor("k_21_begin_0"), val = tensor([0, 0, 1, 0, 0])]; + tensor k_21_end_0 = const()[name = tensor("k_21_end_0"), val = tensor([1, 1, 2, 16, 64])]; + tensor k_21_end_mask_0 = const()[name = tensor("k_21_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor k_21_squeeze_mask_0 = const()[name = tensor("k_21_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor k_21_cast_fp16 = slice_by_index(begin = k_21_begin_0, end = k_21_end_0, end_mask = k_21_end_mask_0, squeeze_mask = k_21_squeeze_mask_0, x = qkv_11_cast_fp16)[name = tensor("k_21_cast_fp16")]; + tensor v_11_begin_0 = const()[name = tensor("v_11_begin_0"), val = tensor([0, 0, 2, 0, 0])]; + tensor v_11_end_0 = const()[name = tensor("v_11_end_0"), val = tensor([1, 1, 3, 16, 64])]; + tensor v_11_end_mask_0 = const()[name = tensor("v_11_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor v_11_squeeze_mask_0 = const()[name = tensor("v_11_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor v_11_cast_fp16 = slice_by_index(begin = v_11_begin_0, end = v_11_end_0, end_mask = v_11_end_mask_0, squeeze_mask = v_11_squeeze_mask_0, x = qkv_11_cast_fp16)[name = tensor("v_11_cast_fp16")]; + tensor var_2439 = const()[name = tensor("op_2439"), val = tensor([1, 1, 1, 1])]; + tensor position5_to_fp16_dtype_0 = const()[name = tensor("position5_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor position5_to_fp16 = cast(dtype = position5_to_fp16_dtype_0, x = position5)[name = tensor("cast_528")]; + tensor ts_35_cast_fp16 = reshape(shape = var_2439, x = position5_to_fp16)[name = tensor("ts_35_cast_fp16")]; + tensor var_2443 = const()[name = tensor("op_2443"), val = tensor([1, 1, 16, 32, 2])]; + tensor q_complex_11_cast_fp16 = reshape(shape = var_2443, x = q_31_cast_fp16)[name = tensor("q_complex_11_cast_fp16")]; + tensor var_2447 = const()[name = tensor("op_2447"), val = tensor([1, 1, 16, 32, 2])]; + tensor k_complex_11_cast_fp16 = reshape(shape = var_2447, x = k_21_cast_fp16)[name = tensor("k_complex_11_cast_fp16")]; + tensor var_2451_begin_0 = const()[name = tensor("op_2451_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_2451_end_0 = const()[name = tensor("op_2451_end_0"), val = tensor([1, 1, 16, 32, 1])]; + tensor var_2451_end_mask_0 = const()[name = tensor("op_2451_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_2451_squeeze_mask_0 = const()[name = tensor("op_2451_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_2451_cast_fp16 = slice_by_index(begin = var_2451_begin_0, end = var_2451_end_0, end_mask = var_2451_end_mask_0, squeeze_mask = var_2451_squeeze_mask_0, x = q_complex_11_cast_fp16)[name = tensor("op_2451_cast_fp16")]; + tensor var_2459_begin_0 = const()[name = tensor("op_2459_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_2459_end_0 = const()[name = tensor("op_2459_end_0"), val = tensor([1, 1, 16, 32, 2])]; + tensor var_2459_end_mask_0 = const()[name = tensor("op_2459_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_2459_squeeze_mask_0 = const()[name = tensor("op_2459_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_2459_cast_fp16 = slice_by_index(begin = var_2459_begin_0, end = var_2459_end_0, end_mask = var_2459_end_mask_0, squeeze_mask = var_2459_squeeze_mask_0, x = q_complex_11_cast_fp16)[name = tensor("op_2459_cast_fp16")]; + tensor var_2467_begin_0 = const()[name = tensor("op_2467_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_2467_end_0 = const()[name = tensor("op_2467_end_0"), val = tensor([1, 1, 16, 32, 1])]; + tensor var_2467_end_mask_0 = const()[name = tensor("op_2467_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_2467_squeeze_mask_0 = const()[name = tensor("op_2467_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_2467_cast_fp16 = slice_by_index(begin = var_2467_begin_0, end = var_2467_end_0, end_mask = var_2467_end_mask_0, squeeze_mask = var_2467_squeeze_mask_0, x = k_complex_11_cast_fp16)[name = tensor("op_2467_cast_fp16")]; + tensor var_2475_begin_0 = const()[name = tensor("op_2475_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_2475_end_0 = const()[name = tensor("op_2475_end_0"), val = tensor([1, 1, 16, 32, 2])]; + tensor var_2475_end_mask_0 = const()[name = tensor("op_2475_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_2475_squeeze_mask_0 = const()[name = tensor("op_2475_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_2475_cast_fp16 = slice_by_index(begin = var_2475_begin_0, end = var_2475_end_0, end_mask = var_2475_end_mask_0, squeeze_mask = var_2475_squeeze_mask_0, x = k_complex_11_cast_fp16)[name = tensor("op_2475_cast_fp16")]; + tensor freqs_11_to_fp16 = const()[name = tensor("freqs_11_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(303632000)))]; + tensor var_2481_cast_fp16 = mul(x = freqs_11_to_fp16, y = ts_35_cast_fp16)[name = tensor("op_2481_cast_fp16")]; + tensor rotr_11_cast_fp16 = cos(x = var_2481_cast_fp16)[name = tensor("rotr_11_cast_fp16")]; + tensor roti_11_cast_fp16 = sin(x = var_2481_cast_fp16)[name = tensor("roti_11_cast_fp16")]; + tensor var_2485_cast_fp16 = mul(x = var_2451_cast_fp16, y = rotr_11_cast_fp16)[name = tensor("op_2485_cast_fp16")]; + tensor var_2486_cast_fp16 = mul(x = var_2459_cast_fp16, y = roti_11_cast_fp16)[name = tensor("op_2486_cast_fp16")]; + tensor qor_21_cast_fp16 = sub(x = var_2485_cast_fp16, y = var_2486_cast_fp16)[name = tensor("qor_21_cast_fp16")]; + tensor var_2489_cast_fp16 = mul(x = var_2451_cast_fp16, y = roti_11_cast_fp16)[name = tensor("op_2489_cast_fp16")]; + tensor var_2490_cast_fp16 = mul(x = var_2459_cast_fp16, y = rotr_11_cast_fp16)[name = tensor("op_2490_cast_fp16")]; + tensor qoi_21_cast_fp16 = add(x = var_2489_cast_fp16, y = var_2490_cast_fp16)[name = tensor("qoi_21_cast_fp16")]; + tensor var_2493_cast_fp16 = mul(x = var_2467_cast_fp16, y = rotr_11_cast_fp16)[name = tensor("op_2493_cast_fp16")]; + tensor var_2494_cast_fp16 = mul(x = var_2475_cast_fp16, y = roti_11_cast_fp16)[name = tensor("op_2494_cast_fp16")]; + tensor kor_21_cast_fp16 = sub(x = var_2493_cast_fp16, y = var_2494_cast_fp16)[name = tensor("kor_21_cast_fp16")]; + tensor var_2497_cast_fp16 = mul(x = var_2467_cast_fp16, y = roti_11_cast_fp16)[name = tensor("op_2497_cast_fp16")]; + tensor var_2498_cast_fp16 = mul(x = var_2475_cast_fp16, y = rotr_11_cast_fp16)[name = tensor("op_2498_cast_fp16")]; + tensor koi_21_cast_fp16 = add(x = var_2497_cast_fp16, y = var_2498_cast_fp16)[name = tensor("koi_21_cast_fp16")]; + tensor qo_11_axis_0 = const()[name = tensor("qo_11_axis_0"), val = tensor(-1)]; + tensor qo_11_cast_fp16 = stack(axis = qo_11_axis_0, values = (qor_21_cast_fp16, qoi_21_cast_fp16))[name = tensor("qo_11_cast_fp16")]; + tensor ko_11_axis_0 = const()[name = tensor("ko_11_axis_0"), val = tensor(-1)]; + tensor ko_11_cast_fp16 = stack(axis = ko_11_axis_0, values = (kor_21_cast_fp16, koi_21_cast_fp16))[name = tensor("ko_11_cast_fp16")]; + tensor var_2527 = const()[name = tensor("op_2527"), val = tensor([1, 1, 16, 64])]; + tensor q_33_cast_fp16 = reshape(shape = var_2527, x = qo_11_cast_fp16)[name = tensor("q_33_cast_fp16")]; + tensor var_2529 = const()[name = tensor("op_2529"), val = tensor([1, 1, 16, 64])]; + tensor k_23_cast_fp16 = reshape(shape = var_2529, x = ko_11_cast_fp16)[name = tensor("k_23_cast_fp16")]; + tensor _inversed_2551_y_0_to_fp16 = const()[name = tensor("_inversed_2551_y_0_to_fp16"), val = tensor(0x1p-9)]; + tensor _inversed_2551_cast_fp16 = mul(x = ts_35_cast_fp16, y = _inversed_2551_y_0_to_fp16)[name = tensor("_inversed_2551_cast_fp16")]; + tensor var_2552_cast_fp16 = floor(x = _inversed_2551_cast_fp16)[name = tensor("op_2552_cast_fp16")]; + tensor var_2553_to_fp16 = const()[name = tensor("op_2553_to_fp16"), val = tensor(0x1p+9)]; + tensor var_2554_cast_fp16 = mul(x = var_2552_cast_fp16, y = var_2553_to_fp16)[name = tensor("op_2554_cast_fp16")]; + tensor write_indices_float_23_cast_fp16 = sub(x = ts_35_cast_fp16, y = var_2554_cast_fp16)[name = tensor("write_indices_float_23_cast_fp16")]; + tensor var_2561_dtype_0 = const()[name = tensor("op_2561_dtype_0"), val = tensor("int32")]; + tensor write_indices_11_reps_0 = const()[name = tensor("write_indices_11_reps_0"), val = tensor([1, 1, 16, 64])]; + tensor write_indices_float_23_cast_fp16_to_int32 = cast(dtype = var_2561_dtype_0, x = write_indices_float_23_cast_fp16)[name = tensor("cast_527")]; + tensor write_indices_11 = tile(reps = write_indices_11_reps_0, x = write_indices_float_23_cast_fp16_to_int32)[name = tensor("write_indices_11")]; + tensor var_2569_begin_0 = const()[name = tensor("op_2569_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_2569_end_0 = const()[name = tensor("op_2569_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor var_2569_end_mask_0 = const()[name = tensor("op_2569_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_2569_squeeze_mask_0 = const()[name = tensor("op_2569_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor cache5_to_fp16_dtype_0 = const()[name = tensor("cache5_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor cache5_to_fp16 = cast(dtype = cache5_to_fp16_dtype_0, x = cache5)[name = tensor("cast_526")]; + tensor var_2569_cast_fp16 = slice_by_index(begin = var_2569_begin_0, end = var_2569_end_0, end_mask = var_2569_end_mask_0, squeeze_mask = var_2569_squeeze_mask_0, x = cache5_to_fp16)[name = tensor("op_2569_cast_fp16")]; + tensor var_2571_axis_0 = const()[name = tensor("op_2571_axis_0"), val = tensor(1)]; + tensor var_2571_mode_0 = const()[name = tensor("op_2571_mode_0"), val = tensor("update")]; + tensor var_2571_validate_indices_0 = const()[name = tensor("op_2571_validate_indices_0"), val = tensor(false)]; + tensor var_2571_cast_fp16 = scatter_along_axis(axis = var_2571_axis_0, data = var_2569_cast_fp16, indices = write_indices_11, mode = var_2571_mode_0, updates = k_23_cast_fp16, validate_indices = var_2571_validate_indices_0)[name = tensor("op_2571_cast_fp16")]; + tensor concat_37 = const()[name = tensor("concat_37"), val = tensor([0, 0, 0, 0, 0])]; + tensor concat_38 = const()[name = tensor("concat_38"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_11_internal_tensor_assign_1_stride_0 = const()[name = tensor("new_cache_11_internal_tensor_assign_1_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_11_internal_tensor_assign_1_begin_mask_0 = const()[name = tensor("new_cache_11_internal_tensor_assign_1_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_11_internal_tensor_assign_1_end_mask_0 = const()[name = tensor("new_cache_11_internal_tensor_assign_1_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_11_internal_tensor_assign_1_squeeze_mask_0 = const()[name = tensor("new_cache_11_internal_tensor_assign_1_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_58 = const()[name = tensor("shape_58"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_10 = const()[name = tensor("reduce_prod_10"), val = tensor(1048576)]; + tensor range_1d_10_start_0 = const()[name = tensor("range_1d_10_start_0"), val = tensor(0)]; + tensor range_1d_10_step_0 = const()[name = tensor("range_1d_10_step_0"), val = tensor(1)]; + tensor range_1d_10 = range_1d(end = reduce_prod_10, start = range_1d_10_start_0, step = range_1d_10_step_0)[name = tensor("range_1d_10")]; + tensor reshape_50 = reshape(shape = shape_58, x = range_1d_10)[name = tensor("reshape_50")]; + tensor slice_by_index_10 = slice_by_index(begin = concat_37, begin_mask = new_cache_11_internal_tensor_assign_1_begin_mask_0, end = concat_38, end_mask = new_cache_11_internal_tensor_assign_1_end_mask_0, squeeze_mask = new_cache_11_internal_tensor_assign_1_squeeze_mask_0, stride = new_cache_11_internal_tensor_assign_1_stride_0, x = reshape_50)[name = tensor("slice_by_index_10")]; + tensor reshape_51_shape_0 = const()[name = tensor("reshape_51_shape_0"), val = tensor([-1])]; + tensor reshape_51 = reshape(shape = reshape_51_shape_0, x = slice_by_index_10)[name = tensor("reshape_51")]; + tensor reshape_52_shape_0 = const()[name = tensor("reshape_52_shape_0"), val = tensor([-1])]; + tensor reshape_52_cast_fp16 = reshape(shape = reshape_52_shape_0, x = var_2571_cast_fp16)[name = tensor("reshape_52_cast_fp16")]; + tensor reshape_53_shape_0 = const()[name = tensor("reshape_53_shape_0"), val = tensor([-1])]; + tensor reshape_53_cast_fp16 = reshape(shape = reshape_53_shape_0, x = cache5_to_fp16)[name = tensor("reshape_53_cast_fp16")]; + tensor scatter_10_mode_0 = const()[name = tensor("scatter_10_mode_0"), val = tensor("update")]; + tensor scatter_10_axis_0 = const()[name = tensor("scatter_10_axis_0"), val = tensor(0)]; + tensor scatter_10_validate_indices_0 = const()[name = tensor("scatter_10_validate_indices_0"), val = tensor(false)]; + tensor scatter_10_cast_fp16 = scatter(axis = scatter_10_axis_0, data = reshape_53_cast_fp16, indices = reshape_51, mode = scatter_10_mode_0, updates = reshape_52_cast_fp16, validate_indices = scatter_10_validate_indices_0)[name = tensor("scatter_10_cast_fp16")]; + tensor reshape_54_cast_fp16 = reshape(shape = shape_58, x = scatter_10_cast_fp16)[name = tensor("reshape_54_cast_fp16")]; + tensor var_2579_begin_0 = const()[name = tensor("op_2579_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor var_2579_end_0 = const()[name = tensor("op_2579_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor var_2579_end_mask_0 = const()[name = tensor("op_2579_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_2579_squeeze_mask_0 = const()[name = tensor("op_2579_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor var_2579_cast_fp16 = slice_by_index(begin = var_2579_begin_0, end = var_2579_end_0, end_mask = var_2579_end_mask_0, squeeze_mask = var_2579_squeeze_mask_0, x = reshape_54_cast_fp16)[name = tensor("op_2579_cast_fp16")]; + tensor var_2581_axis_0 = const()[name = tensor("op_2581_axis_0"), val = tensor(1)]; + tensor var_2581_mode_0 = const()[name = tensor("op_2581_mode_0"), val = tensor("update")]; + tensor var_2581_validate_indices_0 = const()[name = tensor("op_2581_validate_indices_0"), val = tensor(false)]; + tensor var_2581_cast_fp16 = scatter_along_axis(axis = var_2581_axis_0, data = var_2579_cast_fp16, indices = write_indices_11, mode = var_2581_mode_0, updates = v_11_cast_fp16, validate_indices = var_2581_validate_indices_0)[name = tensor("op_2581_cast_fp16")]; + tensor concat_39 = const()[name = tensor("concat_39"), val = tensor([1, 0, 0, 0, 0])]; + tensor concat_40 = const()[name = tensor("concat_40"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_11_internal_tensor_assign_2_stride_0 = const()[name = tensor("new_cache_11_internal_tensor_assign_2_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_11_internal_tensor_assign_2_begin_mask_0 = const()[name = tensor("new_cache_11_internal_tensor_assign_2_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_11_internal_tensor_assign_2_end_mask_0 = const()[name = tensor("new_cache_11_internal_tensor_assign_2_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_11_internal_tensor_assign_2_squeeze_mask_0 = const()[name = tensor("new_cache_11_internal_tensor_assign_2_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_59 = const()[name = tensor("shape_59"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_11 = const()[name = tensor("reduce_prod_11"), val = tensor(1048576)]; + tensor range_1d_11_start_0 = const()[name = tensor("range_1d_11_start_0"), val = tensor(0)]; + tensor range_1d_11_step_0 = const()[name = tensor("range_1d_11_step_0"), val = tensor(1)]; + tensor range_1d_11 = range_1d(end = reduce_prod_11, start = range_1d_11_start_0, step = range_1d_11_step_0)[name = tensor("range_1d_11")]; + tensor reshape_55 = reshape(shape = shape_59, x = range_1d_11)[name = tensor("reshape_55")]; + tensor slice_by_index_11 = slice_by_index(begin = concat_39, begin_mask = new_cache_11_internal_tensor_assign_2_begin_mask_0, end = concat_40, end_mask = new_cache_11_internal_tensor_assign_2_end_mask_0, squeeze_mask = new_cache_11_internal_tensor_assign_2_squeeze_mask_0, stride = new_cache_11_internal_tensor_assign_2_stride_0, x = reshape_55)[name = tensor("slice_by_index_11")]; + tensor reshape_56_shape_0 = const()[name = tensor("reshape_56_shape_0"), val = tensor([-1])]; + tensor reshape_56 = reshape(shape = reshape_56_shape_0, x = slice_by_index_11)[name = tensor("reshape_56")]; + tensor reshape_57_shape_0 = const()[name = tensor("reshape_57_shape_0"), val = tensor([-1])]; + tensor reshape_57_cast_fp16 = reshape(shape = reshape_57_shape_0, x = var_2581_cast_fp16)[name = tensor("reshape_57_cast_fp16")]; + tensor reshape_58_shape_0 = const()[name = tensor("reshape_58_shape_0"), val = tensor([-1])]; + tensor reshape_58_cast_fp16 = reshape(shape = reshape_58_shape_0, x = reshape_54_cast_fp16)[name = tensor("reshape_58_cast_fp16")]; + tensor scatter_11_mode_0 = const()[name = tensor("scatter_11_mode_0"), val = tensor("update")]; + tensor scatter_11_axis_0 = const()[name = tensor("scatter_11_axis_0"), val = tensor(0)]; + tensor scatter_11_validate_indices_0 = const()[name = tensor("scatter_11_validate_indices_0"), val = tensor(false)]; + tensor scatter_11_cast_fp16 = scatter(axis = scatter_11_axis_0, data = reshape_58_cast_fp16, indices = reshape_56, mode = scatter_11_mode_0, updates = reshape_57_cast_fp16, validate_indices = scatter_11_validate_indices_0)[name = tensor("scatter_11_cast_fp16")]; + tensor reshape_59_cast_fp16 = reshape(shape = shape_59, x = scatter_11_cast_fp16)[name = tensor("reshape_59_cast_fp16")]; + tensor reshape_59_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("reshape_59_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor keys_31_begin_0 = const()[name = tensor("keys_31_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor keys_31_end_0 = const()[name = tensor("keys_31_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor keys_31_end_mask_0 = const()[name = tensor("keys_31_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor keys_31_squeeze_mask_0 = const()[name = tensor("keys_31_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor keys_31_cast_fp16 = slice_by_index(begin = keys_31_begin_0, end = keys_31_end_0, end_mask = keys_31_end_mask_0, squeeze_mask = keys_31_squeeze_mask_0, x = reshape_59_cast_fp16)[name = tensor("keys_31_cast_fp16")]; + tensor values_31_begin_0 = const()[name = tensor("values_31_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor values_31_end_0 = const()[name = tensor("values_31_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor values_31_end_mask_0 = const()[name = tensor("values_31_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor values_31_squeeze_mask_0 = const()[name = tensor("values_31_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor values_31_cast_fp16 = slice_by_index(begin = values_31_begin_0, end = values_31_end_0, end_mask = values_31_end_mask_0, squeeze_mask = values_31_squeeze_mask_0, x = reshape_59_cast_fp16)[name = tensor("values_31_cast_fp16")]; + tensor var_2593_cast_fp16 = not_equal(x = keys_31_cast_fp16, y = keys_31_cast_fp16)[name = tensor("op_2593_cast_fp16")]; + tensor keys_33_cast_fp16 = select(a = var_504_to_fp16, b = keys_31_cast_fp16, cond = var_2593_cast_fp16)[name = tensor("keys_33_cast_fp16")]; + tensor var_2601_cast_fp16 = not_equal(x = values_31_cast_fp16, y = values_31_cast_fp16)[name = tensor("op_2601_cast_fp16")]; + tensor values_33_cast_fp16 = select(a = var_504_to_fp16, b = values_31_cast_fp16, cond = var_2601_cast_fp16)[name = tensor("values_33_cast_fp16")]; + tensor var_2625 = const()[name = tensor("op_2625"), val = tensor([0, 2, 1, 3])]; + tensor var_2638 = const()[name = tensor("op_2638"), val = tensor([1, 1, 1])]; + tensor var_2639_cast_fp16 = reshape(shape = var_2638, x = position5_to_fp16)[name = tensor("op_2639_cast_fp16")]; + tensor var_2656_to_fp16 = const()[name = tensor("op_2656_to_fp16"), val = tensor(0x1p+0)]; + tensor valid_len_11_cast_fp16 = add(x = var_2639_cast_fp16, y = var_2656_to_fp16)[name = tensor("valid_len_11_cast_fp16")]; + tensor valid_mask_11_cast_fp16 = less(x = k_positions_1_promoted_to_fp16, y = valid_len_11_cast_fp16)[name = tensor("valid_mask_11_cast_fp16")]; + tensor causal_mask_11_cast_fp16 = less_equal(x = k_positions_1_promoted_to_fp16, y = var_2639_cast_fp16)[name = tensor("causal_mask_11_cast_fp16")]; + tensor attn_mask_21 = logical_and(x = valid_mask_11_cast_fp16, y = causal_mask_11_cast_fp16)[name = tensor("attn_mask_21")]; + tensor attn_mask_23_axes_0 = const()[name = tensor("attn_mask_23_axes_0"), val = tensor([1])]; + tensor attn_mask_23 = expand_dims(axes = attn_mask_23_axes_0, x = attn_mask_21)[name = tensor("attn_mask_23")]; + tensor var_2674_transpose_x_0 = const()[name = tensor("op_2674_transpose_x_0"), val = tensor(false)]; + tensor var_2674_transpose_y_0 = const()[name = tensor("op_2674_transpose_y_0"), val = tensor(false)]; + tensor transpose_82_perm_0 = const()[name = tensor("transpose_82_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_83_perm_0 = const()[name = tensor("transpose_83_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_83 = transpose(perm = transpose_83_perm_0, x = keys_33_cast_fp16)[name = tensor("transpose_193")]; + tensor transpose_82 = transpose(perm = transpose_82_perm_0, x = q_33_cast_fp16)[name = tensor("transpose_194")]; + tensor var_2674_cast_fp16 = matmul(transpose_x = var_2674_transpose_x_0, transpose_y = var_2674_transpose_y_0, x = transpose_82, y = transpose_83)[name = tensor("op_2674_cast_fp16")]; + tensor var_2668_to_fp16 = const()[name = tensor("op_2668_to_fp16"), val = tensor([0x1p-3])]; + tensor attn_weights_31_cast_fp16 = mul(x = var_2674_cast_fp16, y = var_2668_to_fp16)[name = tensor("attn_weights_31_cast_fp16")]; + tensor var_2676 = logical_not(x = attn_mask_23)[name = tensor("op_2676")]; + tensor var_2677_to_fp16 = const()[name = tensor("op_2677_to_fp16"), val = tensor(-inf)]; + tensor attn_weights_33_cast_fp16 = select(a = var_2677_to_fp16, b = attn_weights_31_cast_fp16, cond = var_2676)[name = tensor("attn_weights_33_cast_fp16")]; + tensor var_2679 = const()[name = tensor("op_2679"), val = tensor(-1)]; + tensor attn_weights_35_cast_fp16 = softmax(axis = var_2679, x = attn_weights_33_cast_fp16)[name = tensor("attn_weights_35_cast_fp16")]; + tensor attn_output_11_transpose_x_0 = const()[name = tensor("attn_output_11_transpose_x_0"), val = tensor(false)]; + tensor attn_output_11_transpose_y_0 = const()[name = tensor("attn_output_11_transpose_y_0"), val = tensor(false)]; + tensor values_35_cast_fp16 = transpose(perm = var_2625, x = values_33_cast_fp16)[name = tensor("transpose_195")]; + tensor attn_output_11_cast_fp16 = matmul(transpose_x = attn_output_11_transpose_x_0, transpose_y = attn_output_11_transpose_y_0, x = attn_weights_35_cast_fp16, y = values_35_cast_fp16)[name = tensor("attn_output_11_cast_fp16")]; + tensor var_2687 = const()[name = tensor("op_2687"), val = tensor([0, 2, 1, 3])]; + tensor var_2690 = const()[name = tensor("op_2690"), val = tensor([1, 1, 1024])]; + tensor var_2688_cast_fp16 = transpose(perm = var_2687, x = attn_output_11_cast_fp16)[name = tensor("transpose_192")]; + tensor input_55_cast_fp16 = reshape(shape = var_2690, x = var_2688_cast_fp16)[name = tensor("input_55_cast_fp16")]; + tensor linear_22_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = attn5_out_proj_weight_cast_fp16, x = input_55_cast_fp16)[name = tensor("linear_22_cast_fp16")]; + tensor var_2696_to_fp16 = const()[name = tensor("op_2696_to_fp16"), val = tensor(0x1p+0)]; + tensor var_2697_cast_fp16 = add(x = position5_to_fp16, y = var_2696_to_fp16)[name = tensor("op_2697_cast_fp16")]; + tensor var_2697_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_2697_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor input_57_cast_fp16 = add(x = input_53_cast_fp16, y = linear_22_cast_fp16)[name = tensor("input_57_cast_fp16")]; + tensor input_59_axes_0 = const()[name = tensor("input_59_axes_0"), val = tensor([-1])]; + tensor norm5_2_weight_to_fp16 = const()[name = tensor("norm5_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(303632128)))]; + tensor norm5_2_bias_to_fp16 = const()[name = tensor("norm5_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(303634240)))]; + tensor var_2701_to_fp16 = const()[name = tensor("op_2701_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_59_cast_fp16 = layer_norm(axes = input_59_axes_0, beta = norm5_2_bias_to_fp16, epsilon = var_2701_to_fp16, gamma = norm5_2_weight_to_fp16, x = input_57_cast_fp16)[name = tensor("input_59_cast_fp16")]; + tensor linear_23_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = linear5_1_weight_cast_fp16, x = input_59_cast_fp16)[name = tensor("linear_23_cast_fp16")]; + tensor input_61_mode_0 = const()[name = tensor("input_61_mode_0"), val = tensor("EXACT")]; + tensor input_61_cast_fp16 = gelu(mode = input_61_mode_0, x = linear_23_cast_fp16)[name = tensor("input_61_cast_fp16")]; + tensor linear_24_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = linear5_2_weight_cast_fp16, x = input_61_cast_fp16)[name = tensor("linear_24_cast_fp16")]; + tensor input_63_cast_fp16 = add(x = input_57_cast_fp16, y = linear_24_cast_fp16)[name = tensor("input_63_cast_fp16")]; + tensor x_13_axes_0 = const()[name = tensor("x_13_axes_0"), val = tensor([-1])]; + tensor norm6_1_weight_to_fp16 = const()[name = tensor("norm6_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(303636352)))]; + tensor norm6_1_bias_to_fp16 = const()[name = tensor("norm6_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(303638464)))]; + tensor var_2718_to_fp16 = const()[name = tensor("op_2718_to_fp16"), val = tensor(0x1.5p-17)]; + tensor x_13_cast_fp16 = layer_norm(axes = x_13_axes_0, beta = norm6_1_bias_to_fp16, epsilon = var_2718_to_fp16, gamma = norm6_1_weight_to_fp16, x = input_63_cast_fp16)[name = tensor("x_13_cast_fp16")]; + tensor linear_25_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = attn6_in_proj_weight_cast_fp16, x = x_13_cast_fp16)[name = tensor("linear_25_cast_fp16")]; + tensor var_2754 = const()[name = tensor("op_2754"), val = tensor([1, 1, 3, 16, 64])]; + tensor qkv_13_cast_fp16 = reshape(shape = var_2754, x = linear_25_cast_fp16)[name = tensor("qkv_13_cast_fp16")]; + tensor q_37_begin_0 = const()[name = tensor("q_37_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor q_37_end_0 = const()[name = tensor("q_37_end_0"), val = tensor([1, 1, 1, 16, 64])]; + tensor q_37_end_mask_0 = const()[name = tensor("q_37_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor q_37_squeeze_mask_0 = const()[name = tensor("q_37_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor q_37_cast_fp16 = slice_by_index(begin = q_37_begin_0, end = q_37_end_0, end_mask = q_37_end_mask_0, squeeze_mask = q_37_squeeze_mask_0, x = qkv_13_cast_fp16)[name = tensor("q_37_cast_fp16")]; + tensor k_25_begin_0 = const()[name = tensor("k_25_begin_0"), val = tensor([0, 0, 1, 0, 0])]; + tensor k_25_end_0 = const()[name = tensor("k_25_end_0"), val = tensor([1, 1, 2, 16, 64])]; + tensor k_25_end_mask_0 = const()[name = tensor("k_25_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor k_25_squeeze_mask_0 = const()[name = tensor("k_25_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor k_25_cast_fp16 = slice_by_index(begin = k_25_begin_0, end = k_25_end_0, end_mask = k_25_end_mask_0, squeeze_mask = k_25_squeeze_mask_0, x = qkv_13_cast_fp16)[name = tensor("k_25_cast_fp16")]; + tensor v_13_begin_0 = const()[name = tensor("v_13_begin_0"), val = tensor([0, 0, 2, 0, 0])]; + tensor v_13_end_0 = const()[name = tensor("v_13_end_0"), val = tensor([1, 1, 3, 16, 64])]; + tensor v_13_end_mask_0 = const()[name = tensor("v_13_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor v_13_squeeze_mask_0 = const()[name = tensor("v_13_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor v_13_cast_fp16 = slice_by_index(begin = v_13_begin_0, end = v_13_end_0, end_mask = v_13_end_mask_0, squeeze_mask = v_13_squeeze_mask_0, x = qkv_13_cast_fp16)[name = tensor("v_13_cast_fp16")]; + tensor var_2858 = const()[name = tensor("op_2858"), val = tensor([1, 1, 1, 1])]; + tensor position6_to_fp16_dtype_0 = const()[name = tensor("position6_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor position6_to_fp16 = cast(dtype = position6_to_fp16_dtype_0, x = position6)[name = tensor("cast_523")]; + tensor ts_41_cast_fp16 = reshape(shape = var_2858, x = position6_to_fp16)[name = tensor("ts_41_cast_fp16")]; + tensor var_2862 = const()[name = tensor("op_2862"), val = tensor([1, 1, 16, 32, 2])]; + tensor q_complex_13_cast_fp16 = reshape(shape = var_2862, x = q_37_cast_fp16)[name = tensor("q_complex_13_cast_fp16")]; + tensor var_2866 = const()[name = tensor("op_2866"), val = tensor([1, 1, 16, 32, 2])]; + tensor k_complex_13_cast_fp16 = reshape(shape = var_2866, x = k_25_cast_fp16)[name = tensor("k_complex_13_cast_fp16")]; + tensor var_2870_begin_0 = const()[name = tensor("op_2870_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_2870_end_0 = const()[name = tensor("op_2870_end_0"), val = tensor([1, 1, 16, 32, 1])]; + tensor var_2870_end_mask_0 = const()[name = tensor("op_2870_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_2870_squeeze_mask_0 = const()[name = tensor("op_2870_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_2870_cast_fp16 = slice_by_index(begin = var_2870_begin_0, end = var_2870_end_0, end_mask = var_2870_end_mask_0, squeeze_mask = var_2870_squeeze_mask_0, x = q_complex_13_cast_fp16)[name = tensor("op_2870_cast_fp16")]; + tensor var_2878_begin_0 = const()[name = tensor("op_2878_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_2878_end_0 = const()[name = tensor("op_2878_end_0"), val = tensor([1, 1, 16, 32, 2])]; + tensor var_2878_end_mask_0 = const()[name = tensor("op_2878_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_2878_squeeze_mask_0 = const()[name = tensor("op_2878_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_2878_cast_fp16 = slice_by_index(begin = var_2878_begin_0, end = var_2878_end_0, end_mask = var_2878_end_mask_0, squeeze_mask = var_2878_squeeze_mask_0, x = q_complex_13_cast_fp16)[name = tensor("op_2878_cast_fp16")]; + tensor var_2886_begin_0 = const()[name = tensor("op_2886_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_2886_end_0 = const()[name = tensor("op_2886_end_0"), val = tensor([1, 1, 16, 32, 1])]; + tensor var_2886_end_mask_0 = const()[name = tensor("op_2886_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_2886_squeeze_mask_0 = const()[name = tensor("op_2886_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_2886_cast_fp16 = slice_by_index(begin = var_2886_begin_0, end = var_2886_end_0, end_mask = var_2886_end_mask_0, squeeze_mask = var_2886_squeeze_mask_0, x = k_complex_13_cast_fp16)[name = tensor("op_2886_cast_fp16")]; + tensor var_2894_begin_0 = const()[name = tensor("op_2894_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_2894_end_0 = const()[name = tensor("op_2894_end_0"), val = tensor([1, 1, 16, 32, 2])]; + tensor var_2894_end_mask_0 = const()[name = tensor("op_2894_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_2894_squeeze_mask_0 = const()[name = tensor("op_2894_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_2894_cast_fp16 = slice_by_index(begin = var_2894_begin_0, end = var_2894_end_0, end_mask = var_2894_end_mask_0, squeeze_mask = var_2894_squeeze_mask_0, x = k_complex_13_cast_fp16)[name = tensor("op_2894_cast_fp16")]; + tensor freqs_13_to_fp16 = const()[name = tensor("freqs_13_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(303640576)))]; + tensor var_2900_cast_fp16 = mul(x = freqs_13_to_fp16, y = ts_41_cast_fp16)[name = tensor("op_2900_cast_fp16")]; + tensor rotr_13_cast_fp16 = cos(x = var_2900_cast_fp16)[name = tensor("rotr_13_cast_fp16")]; + tensor roti_13_cast_fp16 = sin(x = var_2900_cast_fp16)[name = tensor("roti_13_cast_fp16")]; + tensor var_2904_cast_fp16 = mul(x = var_2870_cast_fp16, y = rotr_13_cast_fp16)[name = tensor("op_2904_cast_fp16")]; + tensor var_2905_cast_fp16 = mul(x = var_2878_cast_fp16, y = roti_13_cast_fp16)[name = tensor("op_2905_cast_fp16")]; + tensor qor_25_cast_fp16 = sub(x = var_2904_cast_fp16, y = var_2905_cast_fp16)[name = tensor("qor_25_cast_fp16")]; + tensor var_2908_cast_fp16 = mul(x = var_2870_cast_fp16, y = roti_13_cast_fp16)[name = tensor("op_2908_cast_fp16")]; + tensor var_2909_cast_fp16 = mul(x = var_2878_cast_fp16, y = rotr_13_cast_fp16)[name = tensor("op_2909_cast_fp16")]; + tensor qoi_25_cast_fp16 = add(x = var_2908_cast_fp16, y = var_2909_cast_fp16)[name = tensor("qoi_25_cast_fp16")]; + tensor var_2912_cast_fp16 = mul(x = var_2886_cast_fp16, y = rotr_13_cast_fp16)[name = tensor("op_2912_cast_fp16")]; + tensor var_2913_cast_fp16 = mul(x = var_2894_cast_fp16, y = roti_13_cast_fp16)[name = tensor("op_2913_cast_fp16")]; + tensor kor_25_cast_fp16 = sub(x = var_2912_cast_fp16, y = var_2913_cast_fp16)[name = tensor("kor_25_cast_fp16")]; + tensor var_2916_cast_fp16 = mul(x = var_2886_cast_fp16, y = roti_13_cast_fp16)[name = tensor("op_2916_cast_fp16")]; + tensor var_2917_cast_fp16 = mul(x = var_2894_cast_fp16, y = rotr_13_cast_fp16)[name = tensor("op_2917_cast_fp16")]; + tensor koi_25_cast_fp16 = add(x = var_2916_cast_fp16, y = var_2917_cast_fp16)[name = tensor("koi_25_cast_fp16")]; + tensor qo_13_axis_0 = const()[name = tensor("qo_13_axis_0"), val = tensor(-1)]; + tensor qo_13_cast_fp16 = stack(axis = qo_13_axis_0, values = (qor_25_cast_fp16, qoi_25_cast_fp16))[name = tensor("qo_13_cast_fp16")]; + tensor ko_13_axis_0 = const()[name = tensor("ko_13_axis_0"), val = tensor(-1)]; + tensor ko_13_cast_fp16 = stack(axis = ko_13_axis_0, values = (kor_25_cast_fp16, koi_25_cast_fp16))[name = tensor("ko_13_cast_fp16")]; + tensor var_2946 = const()[name = tensor("op_2946"), val = tensor([1, 1, 16, 64])]; + tensor q_39_cast_fp16 = reshape(shape = var_2946, x = qo_13_cast_fp16)[name = tensor("q_39_cast_fp16")]; + tensor var_2948 = const()[name = tensor("op_2948"), val = tensor([1, 1, 16, 64])]; + tensor k_27_cast_fp16 = reshape(shape = var_2948, x = ko_13_cast_fp16)[name = tensor("k_27_cast_fp16")]; + tensor _inversed_2970_y_0_to_fp16 = const()[name = tensor("_inversed_2970_y_0_to_fp16"), val = tensor(0x1p-9)]; + tensor _inversed_2970_cast_fp16 = mul(x = ts_41_cast_fp16, y = _inversed_2970_y_0_to_fp16)[name = tensor("_inversed_2970_cast_fp16")]; + tensor var_2971_cast_fp16 = floor(x = _inversed_2970_cast_fp16)[name = tensor("op_2971_cast_fp16")]; + tensor var_2972_to_fp16 = const()[name = tensor("op_2972_to_fp16"), val = tensor(0x1p+9)]; + tensor var_2973_cast_fp16 = mul(x = var_2971_cast_fp16, y = var_2972_to_fp16)[name = tensor("op_2973_cast_fp16")]; + tensor write_indices_float_27_cast_fp16 = sub(x = ts_41_cast_fp16, y = var_2973_cast_fp16)[name = tensor("write_indices_float_27_cast_fp16")]; + tensor var_2980_dtype_0 = const()[name = tensor("op_2980_dtype_0"), val = tensor("int32")]; + tensor write_indices_13_reps_0 = const()[name = tensor("write_indices_13_reps_0"), val = tensor([1, 1, 16, 64])]; + tensor write_indices_float_27_cast_fp16_to_int32 = cast(dtype = var_2980_dtype_0, x = write_indices_float_27_cast_fp16)[name = tensor("cast_522")]; + tensor write_indices_13 = tile(reps = write_indices_13_reps_0, x = write_indices_float_27_cast_fp16_to_int32)[name = tensor("write_indices_13")]; + tensor var_2988_begin_0 = const()[name = tensor("op_2988_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_2988_end_0 = const()[name = tensor("op_2988_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor var_2988_end_mask_0 = const()[name = tensor("op_2988_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_2988_squeeze_mask_0 = const()[name = tensor("op_2988_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor cache6_to_fp16_dtype_0 = const()[name = tensor("cache6_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor cache6_to_fp16 = cast(dtype = cache6_to_fp16_dtype_0, x = cache6)[name = tensor("cast_521")]; + tensor var_2988_cast_fp16 = slice_by_index(begin = var_2988_begin_0, end = var_2988_end_0, end_mask = var_2988_end_mask_0, squeeze_mask = var_2988_squeeze_mask_0, x = cache6_to_fp16)[name = tensor("op_2988_cast_fp16")]; + tensor var_2990_axis_0 = const()[name = tensor("op_2990_axis_0"), val = tensor(1)]; + tensor var_2990_mode_0 = const()[name = tensor("op_2990_mode_0"), val = tensor("update")]; + tensor var_2990_validate_indices_0 = const()[name = tensor("op_2990_validate_indices_0"), val = tensor(false)]; + tensor var_2990_cast_fp16 = scatter_along_axis(axis = var_2990_axis_0, data = var_2988_cast_fp16, indices = write_indices_13, mode = var_2990_mode_0, updates = k_27_cast_fp16, validate_indices = var_2990_validate_indices_0)[name = tensor("op_2990_cast_fp16")]; + tensor concat_44 = const()[name = tensor("concat_44"), val = tensor([0, 0, 0, 0, 0])]; + tensor concat_45 = const()[name = tensor("concat_45"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_13_internal_tensor_assign_1_stride_0 = const()[name = tensor("new_cache_13_internal_tensor_assign_1_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_13_internal_tensor_assign_1_begin_mask_0 = const()[name = tensor("new_cache_13_internal_tensor_assign_1_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_13_internal_tensor_assign_1_end_mask_0 = const()[name = tensor("new_cache_13_internal_tensor_assign_1_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_13_internal_tensor_assign_1_squeeze_mask_0 = const()[name = tensor("new_cache_13_internal_tensor_assign_1_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_60 = const()[name = tensor("shape_60"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_12 = const()[name = tensor("reduce_prod_12"), val = tensor(1048576)]; + tensor range_1d_12_start_0 = const()[name = tensor("range_1d_12_start_0"), val = tensor(0)]; + tensor range_1d_12_step_0 = const()[name = tensor("range_1d_12_step_0"), val = tensor(1)]; + tensor range_1d_12 = range_1d(end = reduce_prod_12, start = range_1d_12_start_0, step = range_1d_12_step_0)[name = tensor("range_1d_12")]; + tensor reshape_60 = reshape(shape = shape_60, x = range_1d_12)[name = tensor("reshape_60")]; + tensor slice_by_index_12 = slice_by_index(begin = concat_44, begin_mask = new_cache_13_internal_tensor_assign_1_begin_mask_0, end = concat_45, end_mask = new_cache_13_internal_tensor_assign_1_end_mask_0, squeeze_mask = new_cache_13_internal_tensor_assign_1_squeeze_mask_0, stride = new_cache_13_internal_tensor_assign_1_stride_0, x = reshape_60)[name = tensor("slice_by_index_12")]; + tensor reshape_61_shape_0 = const()[name = tensor("reshape_61_shape_0"), val = tensor([-1])]; + tensor reshape_61 = reshape(shape = reshape_61_shape_0, x = slice_by_index_12)[name = tensor("reshape_61")]; + tensor reshape_62_shape_0 = const()[name = tensor("reshape_62_shape_0"), val = tensor([-1])]; + tensor reshape_62_cast_fp16 = reshape(shape = reshape_62_shape_0, x = var_2990_cast_fp16)[name = tensor("reshape_62_cast_fp16")]; + tensor reshape_63_shape_0 = const()[name = tensor("reshape_63_shape_0"), val = tensor([-1])]; + tensor reshape_63_cast_fp16 = reshape(shape = reshape_63_shape_0, x = cache6_to_fp16)[name = tensor("reshape_63_cast_fp16")]; + tensor scatter_12_mode_0 = const()[name = tensor("scatter_12_mode_0"), val = tensor("update")]; + tensor scatter_12_axis_0 = const()[name = tensor("scatter_12_axis_0"), val = tensor(0)]; + tensor scatter_12_validate_indices_0 = const()[name = tensor("scatter_12_validate_indices_0"), val = tensor(false)]; + tensor scatter_12_cast_fp16 = scatter(axis = scatter_12_axis_0, data = reshape_63_cast_fp16, indices = reshape_61, mode = scatter_12_mode_0, updates = reshape_62_cast_fp16, validate_indices = scatter_12_validate_indices_0)[name = tensor("scatter_12_cast_fp16")]; + tensor reshape_64_cast_fp16 = reshape(shape = shape_60, x = scatter_12_cast_fp16)[name = tensor("reshape_64_cast_fp16")]; + tensor var_2998_begin_0 = const()[name = tensor("op_2998_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor var_2998_end_0 = const()[name = tensor("op_2998_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor var_2998_end_mask_0 = const()[name = tensor("op_2998_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_2998_squeeze_mask_0 = const()[name = tensor("op_2998_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor var_2998_cast_fp16 = slice_by_index(begin = var_2998_begin_0, end = var_2998_end_0, end_mask = var_2998_end_mask_0, squeeze_mask = var_2998_squeeze_mask_0, x = reshape_64_cast_fp16)[name = tensor("op_2998_cast_fp16")]; + tensor var_3000_axis_0 = const()[name = tensor("op_3000_axis_0"), val = tensor(1)]; + tensor var_3000_mode_0 = const()[name = tensor("op_3000_mode_0"), val = tensor("update")]; + tensor var_3000_validate_indices_0 = const()[name = tensor("op_3000_validate_indices_0"), val = tensor(false)]; + tensor var_3000_cast_fp16 = scatter_along_axis(axis = var_3000_axis_0, data = var_2998_cast_fp16, indices = write_indices_13, mode = var_3000_mode_0, updates = v_13_cast_fp16, validate_indices = var_3000_validate_indices_0)[name = tensor("op_3000_cast_fp16")]; + tensor concat_46 = const()[name = tensor("concat_46"), val = tensor([1, 0, 0, 0, 0])]; + tensor concat_47 = const()[name = tensor("concat_47"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_13_internal_tensor_assign_2_stride_0 = const()[name = tensor("new_cache_13_internal_tensor_assign_2_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_13_internal_tensor_assign_2_begin_mask_0 = const()[name = tensor("new_cache_13_internal_tensor_assign_2_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_13_internal_tensor_assign_2_end_mask_0 = const()[name = tensor("new_cache_13_internal_tensor_assign_2_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_13_internal_tensor_assign_2_squeeze_mask_0 = const()[name = tensor("new_cache_13_internal_tensor_assign_2_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_61 = const()[name = tensor("shape_61"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_13 = const()[name = tensor("reduce_prod_13"), val = tensor(1048576)]; + tensor range_1d_13_start_0 = const()[name = tensor("range_1d_13_start_0"), val = tensor(0)]; + tensor range_1d_13_step_0 = const()[name = tensor("range_1d_13_step_0"), val = tensor(1)]; + tensor range_1d_13 = range_1d(end = reduce_prod_13, start = range_1d_13_start_0, step = range_1d_13_step_0)[name = tensor("range_1d_13")]; + tensor reshape_65 = reshape(shape = shape_61, x = range_1d_13)[name = tensor("reshape_65")]; + tensor slice_by_index_13 = slice_by_index(begin = concat_46, begin_mask = new_cache_13_internal_tensor_assign_2_begin_mask_0, end = concat_47, end_mask = new_cache_13_internal_tensor_assign_2_end_mask_0, squeeze_mask = new_cache_13_internal_tensor_assign_2_squeeze_mask_0, stride = new_cache_13_internal_tensor_assign_2_stride_0, x = reshape_65)[name = tensor("slice_by_index_13")]; + tensor reshape_66_shape_0 = const()[name = tensor("reshape_66_shape_0"), val = tensor([-1])]; + tensor reshape_66 = reshape(shape = reshape_66_shape_0, x = slice_by_index_13)[name = tensor("reshape_66")]; + tensor reshape_67_shape_0 = const()[name = tensor("reshape_67_shape_0"), val = tensor([-1])]; + tensor reshape_67_cast_fp16 = reshape(shape = reshape_67_shape_0, x = var_3000_cast_fp16)[name = tensor("reshape_67_cast_fp16")]; + tensor reshape_68_shape_0 = const()[name = tensor("reshape_68_shape_0"), val = tensor([-1])]; + tensor reshape_68_cast_fp16 = reshape(shape = reshape_68_shape_0, x = reshape_64_cast_fp16)[name = tensor("reshape_68_cast_fp16")]; + tensor scatter_13_mode_0 = const()[name = tensor("scatter_13_mode_0"), val = tensor("update")]; + tensor scatter_13_axis_0 = const()[name = tensor("scatter_13_axis_0"), val = tensor(0)]; + tensor scatter_13_validate_indices_0 = const()[name = tensor("scatter_13_validate_indices_0"), val = tensor(false)]; + tensor scatter_13_cast_fp16 = scatter(axis = scatter_13_axis_0, data = reshape_68_cast_fp16, indices = reshape_66, mode = scatter_13_mode_0, updates = reshape_67_cast_fp16, validate_indices = scatter_13_validate_indices_0)[name = tensor("scatter_13_cast_fp16")]; + tensor reshape_69_cast_fp16 = reshape(shape = shape_61, x = scatter_13_cast_fp16)[name = tensor("reshape_69_cast_fp16")]; + tensor reshape_69_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("reshape_69_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor keys_37_begin_0 = const()[name = tensor("keys_37_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor keys_37_end_0 = const()[name = tensor("keys_37_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor keys_37_end_mask_0 = const()[name = tensor("keys_37_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor keys_37_squeeze_mask_0 = const()[name = tensor("keys_37_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor keys_37_cast_fp16 = slice_by_index(begin = keys_37_begin_0, end = keys_37_end_0, end_mask = keys_37_end_mask_0, squeeze_mask = keys_37_squeeze_mask_0, x = reshape_69_cast_fp16)[name = tensor("keys_37_cast_fp16")]; + tensor values_37_begin_0 = const()[name = tensor("values_37_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor values_37_end_0 = const()[name = tensor("values_37_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor values_37_end_mask_0 = const()[name = tensor("values_37_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor values_37_squeeze_mask_0 = const()[name = tensor("values_37_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor values_37_cast_fp16 = slice_by_index(begin = values_37_begin_0, end = values_37_end_0, end_mask = values_37_end_mask_0, squeeze_mask = values_37_squeeze_mask_0, x = reshape_69_cast_fp16)[name = tensor("values_37_cast_fp16")]; + tensor var_3012_cast_fp16 = not_equal(x = keys_37_cast_fp16, y = keys_37_cast_fp16)[name = tensor("op_3012_cast_fp16")]; + tensor keys_39_cast_fp16 = select(a = var_504_to_fp16, b = keys_37_cast_fp16, cond = var_3012_cast_fp16)[name = tensor("keys_39_cast_fp16")]; + tensor var_3020_cast_fp16 = not_equal(x = values_37_cast_fp16, y = values_37_cast_fp16)[name = tensor("op_3020_cast_fp16")]; + tensor values_39_cast_fp16 = select(a = var_504_to_fp16, b = values_37_cast_fp16, cond = var_3020_cast_fp16)[name = tensor("values_39_cast_fp16")]; + tensor var_3044 = const()[name = tensor("op_3044"), val = tensor([0, 2, 1, 3])]; + tensor var_3057 = const()[name = tensor("op_3057"), val = tensor([1, 1, 1])]; + tensor var_3058_cast_fp16 = reshape(shape = var_3057, x = position6_to_fp16)[name = tensor("op_3058_cast_fp16")]; + tensor var_3075_to_fp16 = const()[name = tensor("op_3075_to_fp16"), val = tensor(0x1p+0)]; + tensor valid_len_13_cast_fp16 = add(x = var_3058_cast_fp16, y = var_3075_to_fp16)[name = tensor("valid_len_13_cast_fp16")]; + tensor valid_mask_13_cast_fp16 = less(x = k_positions_1_promoted_to_fp16, y = valid_len_13_cast_fp16)[name = tensor("valid_mask_13_cast_fp16")]; + tensor causal_mask_13_cast_fp16 = less_equal(x = k_positions_1_promoted_to_fp16, y = var_3058_cast_fp16)[name = tensor("causal_mask_13_cast_fp16")]; + tensor attn_mask_25 = logical_and(x = valid_mask_13_cast_fp16, y = causal_mask_13_cast_fp16)[name = tensor("attn_mask_25")]; + tensor attn_mask_27_axes_0 = const()[name = tensor("attn_mask_27_axes_0"), val = tensor([1])]; + tensor attn_mask_27 = expand_dims(axes = attn_mask_27_axes_0, x = attn_mask_25)[name = tensor("attn_mask_27")]; + tensor var_3093_transpose_x_0 = const()[name = tensor("op_3093_transpose_x_0"), val = tensor(false)]; + tensor var_3093_transpose_y_0 = const()[name = tensor("op_3093_transpose_y_0"), val = tensor(false)]; + tensor transpose_84_perm_0 = const()[name = tensor("transpose_84_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_85_perm_0 = const()[name = tensor("transpose_85_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_85 = transpose(perm = transpose_85_perm_0, x = keys_39_cast_fp16)[name = tensor("transpose_189")]; + tensor transpose_84 = transpose(perm = transpose_84_perm_0, x = q_39_cast_fp16)[name = tensor("transpose_190")]; + tensor var_3093_cast_fp16 = matmul(transpose_x = var_3093_transpose_x_0, transpose_y = var_3093_transpose_y_0, x = transpose_84, y = transpose_85)[name = tensor("op_3093_cast_fp16")]; + tensor var_3087_to_fp16 = const()[name = tensor("op_3087_to_fp16"), val = tensor([0x1p-3])]; + tensor attn_weights_37_cast_fp16 = mul(x = var_3093_cast_fp16, y = var_3087_to_fp16)[name = tensor("attn_weights_37_cast_fp16")]; + tensor var_3095 = logical_not(x = attn_mask_27)[name = tensor("op_3095")]; + tensor var_3096_to_fp16 = const()[name = tensor("op_3096_to_fp16"), val = tensor(-inf)]; + tensor attn_weights_39_cast_fp16 = select(a = var_3096_to_fp16, b = attn_weights_37_cast_fp16, cond = var_3095)[name = tensor("attn_weights_39_cast_fp16")]; + tensor var_3098 = const()[name = tensor("op_3098"), val = tensor(-1)]; + tensor attn_weights_41_cast_fp16 = softmax(axis = var_3098, x = attn_weights_39_cast_fp16)[name = tensor("attn_weights_41_cast_fp16")]; + tensor attn_output_13_transpose_x_0 = const()[name = tensor("attn_output_13_transpose_x_0"), val = tensor(false)]; + tensor attn_output_13_transpose_y_0 = const()[name = tensor("attn_output_13_transpose_y_0"), val = tensor(false)]; + tensor values_41_cast_fp16 = transpose(perm = var_3044, x = values_39_cast_fp16)[name = tensor("transpose_191")]; + tensor attn_output_13_cast_fp16 = matmul(transpose_x = attn_output_13_transpose_x_0, transpose_y = attn_output_13_transpose_y_0, x = attn_weights_41_cast_fp16, y = values_41_cast_fp16)[name = tensor("attn_output_13_cast_fp16")]; + tensor var_3106 = const()[name = tensor("op_3106"), val = tensor([0, 2, 1, 3])]; + tensor var_3109 = const()[name = tensor("op_3109"), val = tensor([1, 1, 1024])]; + tensor var_3107_cast_fp16 = transpose(perm = var_3106, x = attn_output_13_cast_fp16)[name = tensor("transpose_188")]; + tensor input_65_cast_fp16 = reshape(shape = var_3109, x = var_3107_cast_fp16)[name = tensor("input_65_cast_fp16")]; + tensor linear_26_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = attn6_out_proj_weight_cast_fp16, x = input_65_cast_fp16)[name = tensor("linear_26_cast_fp16")]; + tensor var_3115_to_fp16 = const()[name = tensor("op_3115_to_fp16"), val = tensor(0x1p+0)]; + tensor var_3116_cast_fp16 = add(x = position6_to_fp16, y = var_3115_to_fp16)[name = tensor("op_3116_cast_fp16")]; + tensor var_3116_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_3116_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor input_67_cast_fp16 = add(x = input_63_cast_fp16, y = linear_26_cast_fp16)[name = tensor("input_67_cast_fp16")]; + tensor input_69_axes_0 = const()[name = tensor("input_69_axes_0"), val = tensor([-1])]; + tensor norm6_2_weight_to_fp16 = const()[name = tensor("norm6_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(303640704)))]; + tensor norm6_2_bias_to_fp16 = const()[name = tensor("norm6_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(303642816)))]; + tensor var_3120_to_fp16 = const()[name = tensor("op_3120_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_69_cast_fp16 = layer_norm(axes = input_69_axes_0, beta = norm6_2_bias_to_fp16, epsilon = var_3120_to_fp16, gamma = norm6_2_weight_to_fp16, x = input_67_cast_fp16)[name = tensor("input_69_cast_fp16")]; + tensor linear_27_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = linear6_1_weight_cast_fp16, x = input_69_cast_fp16)[name = tensor("linear_27_cast_fp16")]; + tensor input_71_mode_0 = const()[name = tensor("input_71_mode_0"), val = tensor("EXACT")]; + tensor input_71_cast_fp16 = gelu(mode = input_71_mode_0, x = linear_27_cast_fp16)[name = tensor("input_71_cast_fp16")]; + tensor linear_28_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = linear6_2_weight_cast_fp16, x = input_71_cast_fp16)[name = tensor("linear_28_cast_fp16")]; + tensor input_73_cast_fp16 = add(x = input_67_cast_fp16, y = linear_28_cast_fp16)[name = tensor("input_73_cast_fp16")]; + tensor x_15_axes_0 = const()[name = tensor("x_15_axes_0"), val = tensor([-1])]; + tensor norm7_1_weight_to_fp16 = const()[name = tensor("norm7_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(303644928)))]; + tensor norm7_1_bias_to_fp16 = const()[name = tensor("norm7_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(303647040)))]; + tensor var_3137_to_fp16 = const()[name = tensor("op_3137_to_fp16"), val = tensor(0x1.5p-17)]; + tensor x_15_cast_fp16 = layer_norm(axes = x_15_axes_0, beta = norm7_1_bias_to_fp16, epsilon = var_3137_to_fp16, gamma = norm7_1_weight_to_fp16, x = input_73_cast_fp16)[name = tensor("x_15_cast_fp16")]; + tensor linear_29_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = attn7_in_proj_weight_cast_fp16, x = x_15_cast_fp16)[name = tensor("linear_29_cast_fp16")]; + tensor var_3173 = const()[name = tensor("op_3173"), val = tensor([1, 1, 3, 16, 64])]; + tensor qkv_15_cast_fp16 = reshape(shape = var_3173, x = linear_29_cast_fp16)[name = tensor("qkv_15_cast_fp16")]; + tensor q_43_begin_0 = const()[name = tensor("q_43_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor q_43_end_0 = const()[name = tensor("q_43_end_0"), val = tensor([1, 1, 1, 16, 64])]; + tensor q_43_end_mask_0 = const()[name = tensor("q_43_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor q_43_squeeze_mask_0 = const()[name = tensor("q_43_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor q_43_cast_fp16 = slice_by_index(begin = q_43_begin_0, end = q_43_end_0, end_mask = q_43_end_mask_0, squeeze_mask = q_43_squeeze_mask_0, x = qkv_15_cast_fp16)[name = tensor("q_43_cast_fp16")]; + tensor k_29_begin_0 = const()[name = tensor("k_29_begin_0"), val = tensor([0, 0, 1, 0, 0])]; + tensor k_29_end_0 = const()[name = tensor("k_29_end_0"), val = tensor([1, 1, 2, 16, 64])]; + tensor k_29_end_mask_0 = const()[name = tensor("k_29_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor k_29_squeeze_mask_0 = const()[name = tensor("k_29_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor k_29_cast_fp16 = slice_by_index(begin = k_29_begin_0, end = k_29_end_0, end_mask = k_29_end_mask_0, squeeze_mask = k_29_squeeze_mask_0, x = qkv_15_cast_fp16)[name = tensor("k_29_cast_fp16")]; + tensor v_15_begin_0 = const()[name = tensor("v_15_begin_0"), val = tensor([0, 0, 2, 0, 0])]; + tensor v_15_end_0 = const()[name = tensor("v_15_end_0"), val = tensor([1, 1, 3, 16, 64])]; + tensor v_15_end_mask_0 = const()[name = tensor("v_15_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor v_15_squeeze_mask_0 = const()[name = tensor("v_15_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor v_15_cast_fp16 = slice_by_index(begin = v_15_begin_0, end = v_15_end_0, end_mask = v_15_end_mask_0, squeeze_mask = v_15_squeeze_mask_0, x = qkv_15_cast_fp16)[name = tensor("v_15_cast_fp16")]; + tensor var_3277 = const()[name = tensor("op_3277"), val = tensor([1, 1, 1, 1])]; + tensor position7_to_fp16_dtype_0 = const()[name = tensor("position7_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor position7_to_fp16 = cast(dtype = position7_to_fp16_dtype_0, x = position7)[name = tensor("cast_518")]; + tensor ts_47_cast_fp16 = reshape(shape = var_3277, x = position7_to_fp16)[name = tensor("ts_47_cast_fp16")]; + tensor var_3281 = const()[name = tensor("op_3281"), val = tensor([1, 1, 16, 32, 2])]; + tensor q_complex_15_cast_fp16 = reshape(shape = var_3281, x = q_43_cast_fp16)[name = tensor("q_complex_15_cast_fp16")]; + tensor var_3285 = const()[name = tensor("op_3285"), val = tensor([1, 1, 16, 32, 2])]; + tensor k_complex_15_cast_fp16 = reshape(shape = var_3285, x = k_29_cast_fp16)[name = tensor("k_complex_15_cast_fp16")]; + tensor var_3289_begin_0 = const()[name = tensor("op_3289_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_3289_end_0 = const()[name = tensor("op_3289_end_0"), val = tensor([1, 1, 16, 32, 1])]; + tensor var_3289_end_mask_0 = const()[name = tensor("op_3289_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_3289_squeeze_mask_0 = const()[name = tensor("op_3289_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_3289_cast_fp16 = slice_by_index(begin = var_3289_begin_0, end = var_3289_end_0, end_mask = var_3289_end_mask_0, squeeze_mask = var_3289_squeeze_mask_0, x = q_complex_15_cast_fp16)[name = tensor("op_3289_cast_fp16")]; + tensor var_3297_begin_0 = const()[name = tensor("op_3297_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_3297_end_0 = const()[name = tensor("op_3297_end_0"), val = tensor([1, 1, 16, 32, 2])]; + tensor var_3297_end_mask_0 = const()[name = tensor("op_3297_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_3297_squeeze_mask_0 = const()[name = tensor("op_3297_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_3297_cast_fp16 = slice_by_index(begin = var_3297_begin_0, end = var_3297_end_0, end_mask = var_3297_end_mask_0, squeeze_mask = var_3297_squeeze_mask_0, x = q_complex_15_cast_fp16)[name = tensor("op_3297_cast_fp16")]; + tensor var_3305_begin_0 = const()[name = tensor("op_3305_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_3305_end_0 = const()[name = tensor("op_3305_end_0"), val = tensor([1, 1, 16, 32, 1])]; + tensor var_3305_end_mask_0 = const()[name = tensor("op_3305_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_3305_squeeze_mask_0 = const()[name = tensor("op_3305_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_3305_cast_fp16 = slice_by_index(begin = var_3305_begin_0, end = var_3305_end_0, end_mask = var_3305_end_mask_0, squeeze_mask = var_3305_squeeze_mask_0, x = k_complex_15_cast_fp16)[name = tensor("op_3305_cast_fp16")]; + tensor var_3313_begin_0 = const()[name = tensor("op_3313_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_3313_end_0 = const()[name = tensor("op_3313_end_0"), val = tensor([1, 1, 16, 32, 2])]; + tensor var_3313_end_mask_0 = const()[name = tensor("op_3313_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_3313_squeeze_mask_0 = const()[name = tensor("op_3313_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_3313_cast_fp16 = slice_by_index(begin = var_3313_begin_0, end = var_3313_end_0, end_mask = var_3313_end_mask_0, squeeze_mask = var_3313_squeeze_mask_0, x = k_complex_15_cast_fp16)[name = tensor("op_3313_cast_fp16")]; + tensor freqs_15_to_fp16 = const()[name = tensor("freqs_15_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(303649152)))]; + tensor var_3319_cast_fp16 = mul(x = freqs_15_to_fp16, y = ts_47_cast_fp16)[name = tensor("op_3319_cast_fp16")]; + tensor rotr_15_cast_fp16 = cos(x = var_3319_cast_fp16)[name = tensor("rotr_15_cast_fp16")]; + tensor roti_15_cast_fp16 = sin(x = var_3319_cast_fp16)[name = tensor("roti_15_cast_fp16")]; + tensor var_3323_cast_fp16 = mul(x = var_3289_cast_fp16, y = rotr_15_cast_fp16)[name = tensor("op_3323_cast_fp16")]; + tensor var_3324_cast_fp16 = mul(x = var_3297_cast_fp16, y = roti_15_cast_fp16)[name = tensor("op_3324_cast_fp16")]; + tensor qor_29_cast_fp16 = sub(x = var_3323_cast_fp16, y = var_3324_cast_fp16)[name = tensor("qor_29_cast_fp16")]; + tensor var_3327_cast_fp16 = mul(x = var_3289_cast_fp16, y = roti_15_cast_fp16)[name = tensor("op_3327_cast_fp16")]; + tensor var_3328_cast_fp16 = mul(x = var_3297_cast_fp16, y = rotr_15_cast_fp16)[name = tensor("op_3328_cast_fp16")]; + tensor qoi_29_cast_fp16 = add(x = var_3327_cast_fp16, y = var_3328_cast_fp16)[name = tensor("qoi_29_cast_fp16")]; + tensor var_3331_cast_fp16 = mul(x = var_3305_cast_fp16, y = rotr_15_cast_fp16)[name = tensor("op_3331_cast_fp16")]; + tensor var_3332_cast_fp16 = mul(x = var_3313_cast_fp16, y = roti_15_cast_fp16)[name = tensor("op_3332_cast_fp16")]; + tensor kor_29_cast_fp16 = sub(x = var_3331_cast_fp16, y = var_3332_cast_fp16)[name = tensor("kor_29_cast_fp16")]; + tensor var_3335_cast_fp16 = mul(x = var_3305_cast_fp16, y = roti_15_cast_fp16)[name = tensor("op_3335_cast_fp16")]; + tensor var_3336_cast_fp16 = mul(x = var_3313_cast_fp16, y = rotr_15_cast_fp16)[name = tensor("op_3336_cast_fp16")]; + tensor koi_29_cast_fp16 = add(x = var_3335_cast_fp16, y = var_3336_cast_fp16)[name = tensor("koi_29_cast_fp16")]; + tensor qo_15_axis_0 = const()[name = tensor("qo_15_axis_0"), val = tensor(-1)]; + tensor qo_15_cast_fp16 = stack(axis = qo_15_axis_0, values = (qor_29_cast_fp16, qoi_29_cast_fp16))[name = tensor("qo_15_cast_fp16")]; + tensor ko_15_axis_0 = const()[name = tensor("ko_15_axis_0"), val = tensor(-1)]; + tensor ko_15_cast_fp16 = stack(axis = ko_15_axis_0, values = (kor_29_cast_fp16, koi_29_cast_fp16))[name = tensor("ko_15_cast_fp16")]; + tensor var_3365 = const()[name = tensor("op_3365"), val = tensor([1, 1, 16, 64])]; + tensor q_45_cast_fp16 = reshape(shape = var_3365, x = qo_15_cast_fp16)[name = tensor("q_45_cast_fp16")]; + tensor var_3367 = const()[name = tensor("op_3367"), val = tensor([1, 1, 16, 64])]; + tensor k_31_cast_fp16 = reshape(shape = var_3367, x = ko_15_cast_fp16)[name = tensor("k_31_cast_fp16")]; + tensor _inversed_3389_y_0_to_fp16 = const()[name = tensor("_inversed_3389_y_0_to_fp16"), val = tensor(0x1p-9)]; + tensor _inversed_3389_cast_fp16 = mul(x = ts_47_cast_fp16, y = _inversed_3389_y_0_to_fp16)[name = tensor("_inversed_3389_cast_fp16")]; + tensor var_3390_cast_fp16 = floor(x = _inversed_3389_cast_fp16)[name = tensor("op_3390_cast_fp16")]; + tensor var_3391_to_fp16 = const()[name = tensor("op_3391_to_fp16"), val = tensor(0x1p+9)]; + tensor var_3392_cast_fp16 = mul(x = var_3390_cast_fp16, y = var_3391_to_fp16)[name = tensor("op_3392_cast_fp16")]; + tensor write_indices_float_31_cast_fp16 = sub(x = ts_47_cast_fp16, y = var_3392_cast_fp16)[name = tensor("write_indices_float_31_cast_fp16")]; + tensor var_3399_dtype_0 = const()[name = tensor("op_3399_dtype_0"), val = tensor("int32")]; + tensor write_indices_15_reps_0 = const()[name = tensor("write_indices_15_reps_0"), val = tensor([1, 1, 16, 64])]; + tensor write_indices_float_31_cast_fp16_to_int32 = cast(dtype = var_3399_dtype_0, x = write_indices_float_31_cast_fp16)[name = tensor("cast_517")]; + tensor write_indices_15 = tile(reps = write_indices_15_reps_0, x = write_indices_float_31_cast_fp16_to_int32)[name = tensor("write_indices_15")]; + tensor var_3407_begin_0 = const()[name = tensor("op_3407_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_3407_end_0 = const()[name = tensor("op_3407_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor var_3407_end_mask_0 = const()[name = tensor("op_3407_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_3407_squeeze_mask_0 = const()[name = tensor("op_3407_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor cache7_to_fp16_dtype_0 = const()[name = tensor("cache7_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor cache7_to_fp16 = cast(dtype = cache7_to_fp16_dtype_0, x = cache7)[name = tensor("cast_516")]; + tensor var_3407_cast_fp16 = slice_by_index(begin = var_3407_begin_0, end = var_3407_end_0, end_mask = var_3407_end_mask_0, squeeze_mask = var_3407_squeeze_mask_0, x = cache7_to_fp16)[name = tensor("op_3407_cast_fp16")]; + tensor var_3409_axis_0 = const()[name = tensor("op_3409_axis_0"), val = tensor(1)]; + tensor var_3409_mode_0 = const()[name = tensor("op_3409_mode_0"), val = tensor("update")]; + tensor var_3409_validate_indices_0 = const()[name = tensor("op_3409_validate_indices_0"), val = tensor(false)]; + tensor var_3409_cast_fp16 = scatter_along_axis(axis = var_3409_axis_0, data = var_3407_cast_fp16, indices = write_indices_15, mode = var_3409_mode_0, updates = k_31_cast_fp16, validate_indices = var_3409_validate_indices_0)[name = tensor("op_3409_cast_fp16")]; + tensor concat_51 = const()[name = tensor("concat_51"), val = tensor([0, 0, 0, 0, 0])]; + tensor concat_52 = const()[name = tensor("concat_52"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_15_internal_tensor_assign_1_stride_0 = const()[name = tensor("new_cache_15_internal_tensor_assign_1_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_15_internal_tensor_assign_1_begin_mask_0 = const()[name = tensor("new_cache_15_internal_tensor_assign_1_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_15_internal_tensor_assign_1_end_mask_0 = const()[name = tensor("new_cache_15_internal_tensor_assign_1_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_15_internal_tensor_assign_1_squeeze_mask_0 = const()[name = tensor("new_cache_15_internal_tensor_assign_1_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_62 = const()[name = tensor("shape_62"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_14 = const()[name = tensor("reduce_prod_14"), val = tensor(1048576)]; + tensor range_1d_14_start_0 = const()[name = tensor("range_1d_14_start_0"), val = tensor(0)]; + tensor range_1d_14_step_0 = const()[name = tensor("range_1d_14_step_0"), val = tensor(1)]; + tensor range_1d_14 = range_1d(end = reduce_prod_14, start = range_1d_14_start_0, step = range_1d_14_step_0)[name = tensor("range_1d_14")]; + tensor reshape_70 = reshape(shape = shape_62, x = range_1d_14)[name = tensor("reshape_70")]; + tensor slice_by_index_14 = slice_by_index(begin = concat_51, begin_mask = new_cache_15_internal_tensor_assign_1_begin_mask_0, end = concat_52, end_mask = new_cache_15_internal_tensor_assign_1_end_mask_0, squeeze_mask = new_cache_15_internal_tensor_assign_1_squeeze_mask_0, stride = new_cache_15_internal_tensor_assign_1_stride_0, x = reshape_70)[name = tensor("slice_by_index_14")]; + tensor reshape_71_shape_0 = const()[name = tensor("reshape_71_shape_0"), val = tensor([-1])]; + tensor reshape_71 = reshape(shape = reshape_71_shape_0, x = slice_by_index_14)[name = tensor("reshape_71")]; + tensor reshape_72_shape_0 = const()[name = tensor("reshape_72_shape_0"), val = tensor([-1])]; + tensor reshape_72_cast_fp16 = reshape(shape = reshape_72_shape_0, x = var_3409_cast_fp16)[name = tensor("reshape_72_cast_fp16")]; + tensor reshape_73_shape_0 = const()[name = tensor("reshape_73_shape_0"), val = tensor([-1])]; + tensor reshape_73_cast_fp16 = reshape(shape = reshape_73_shape_0, x = cache7_to_fp16)[name = tensor("reshape_73_cast_fp16")]; + tensor scatter_14_mode_0 = const()[name = tensor("scatter_14_mode_0"), val = tensor("update")]; + tensor scatter_14_axis_0 = const()[name = tensor("scatter_14_axis_0"), val = tensor(0)]; + tensor scatter_14_validate_indices_0 = const()[name = tensor("scatter_14_validate_indices_0"), val = tensor(false)]; + tensor scatter_14_cast_fp16 = scatter(axis = scatter_14_axis_0, data = reshape_73_cast_fp16, indices = reshape_71, mode = scatter_14_mode_0, updates = reshape_72_cast_fp16, validate_indices = scatter_14_validate_indices_0)[name = tensor("scatter_14_cast_fp16")]; + tensor reshape_74_cast_fp16 = reshape(shape = shape_62, x = scatter_14_cast_fp16)[name = tensor("reshape_74_cast_fp16")]; + tensor var_3417_begin_0 = const()[name = tensor("op_3417_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor var_3417_end_0 = const()[name = tensor("op_3417_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor var_3417_end_mask_0 = const()[name = tensor("op_3417_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_3417_squeeze_mask_0 = const()[name = tensor("op_3417_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor var_3417_cast_fp16 = slice_by_index(begin = var_3417_begin_0, end = var_3417_end_0, end_mask = var_3417_end_mask_0, squeeze_mask = var_3417_squeeze_mask_0, x = reshape_74_cast_fp16)[name = tensor("op_3417_cast_fp16")]; + tensor var_3419_axis_0 = const()[name = tensor("op_3419_axis_0"), val = tensor(1)]; + tensor var_3419_mode_0 = const()[name = tensor("op_3419_mode_0"), val = tensor("update")]; + tensor var_3419_validate_indices_0 = const()[name = tensor("op_3419_validate_indices_0"), val = tensor(false)]; + tensor var_3419_cast_fp16 = scatter_along_axis(axis = var_3419_axis_0, data = var_3417_cast_fp16, indices = write_indices_15, mode = var_3419_mode_0, updates = v_15_cast_fp16, validate_indices = var_3419_validate_indices_0)[name = tensor("op_3419_cast_fp16")]; + tensor concat_53 = const()[name = tensor("concat_53"), val = tensor([1, 0, 0, 0, 0])]; + tensor concat_54 = const()[name = tensor("concat_54"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_15_internal_tensor_assign_2_stride_0 = const()[name = tensor("new_cache_15_internal_tensor_assign_2_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_15_internal_tensor_assign_2_begin_mask_0 = const()[name = tensor("new_cache_15_internal_tensor_assign_2_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_15_internal_tensor_assign_2_end_mask_0 = const()[name = tensor("new_cache_15_internal_tensor_assign_2_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_15_internal_tensor_assign_2_squeeze_mask_0 = const()[name = tensor("new_cache_15_internal_tensor_assign_2_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_63 = const()[name = tensor("shape_63"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_15 = const()[name = tensor("reduce_prod_15"), val = tensor(1048576)]; + tensor range_1d_15_start_0 = const()[name = tensor("range_1d_15_start_0"), val = tensor(0)]; + tensor range_1d_15_step_0 = const()[name = tensor("range_1d_15_step_0"), val = tensor(1)]; + tensor range_1d_15 = range_1d(end = reduce_prod_15, start = range_1d_15_start_0, step = range_1d_15_step_0)[name = tensor("range_1d_15")]; + tensor reshape_75 = reshape(shape = shape_63, x = range_1d_15)[name = tensor("reshape_75")]; + tensor slice_by_index_15 = slice_by_index(begin = concat_53, begin_mask = new_cache_15_internal_tensor_assign_2_begin_mask_0, end = concat_54, end_mask = new_cache_15_internal_tensor_assign_2_end_mask_0, squeeze_mask = new_cache_15_internal_tensor_assign_2_squeeze_mask_0, stride = new_cache_15_internal_tensor_assign_2_stride_0, x = reshape_75)[name = tensor("slice_by_index_15")]; + tensor reshape_76_shape_0 = const()[name = tensor("reshape_76_shape_0"), val = tensor([-1])]; + tensor reshape_76 = reshape(shape = reshape_76_shape_0, x = slice_by_index_15)[name = tensor("reshape_76")]; + tensor reshape_77_shape_0 = const()[name = tensor("reshape_77_shape_0"), val = tensor([-1])]; + tensor reshape_77_cast_fp16 = reshape(shape = reshape_77_shape_0, x = var_3419_cast_fp16)[name = tensor("reshape_77_cast_fp16")]; + tensor reshape_78_shape_0 = const()[name = tensor("reshape_78_shape_0"), val = tensor([-1])]; + tensor reshape_78_cast_fp16 = reshape(shape = reshape_78_shape_0, x = reshape_74_cast_fp16)[name = tensor("reshape_78_cast_fp16")]; + tensor scatter_15_mode_0 = const()[name = tensor("scatter_15_mode_0"), val = tensor("update")]; + tensor scatter_15_axis_0 = const()[name = tensor("scatter_15_axis_0"), val = tensor(0)]; + tensor scatter_15_validate_indices_0 = const()[name = tensor("scatter_15_validate_indices_0"), val = tensor(false)]; + tensor scatter_15_cast_fp16 = scatter(axis = scatter_15_axis_0, data = reshape_78_cast_fp16, indices = reshape_76, mode = scatter_15_mode_0, updates = reshape_77_cast_fp16, validate_indices = scatter_15_validate_indices_0)[name = tensor("scatter_15_cast_fp16")]; + tensor reshape_79_cast_fp16 = reshape(shape = shape_63, x = scatter_15_cast_fp16)[name = tensor("reshape_79_cast_fp16")]; + tensor reshape_79_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("reshape_79_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor keys_43_begin_0 = const()[name = tensor("keys_43_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor keys_43_end_0 = const()[name = tensor("keys_43_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor keys_43_end_mask_0 = const()[name = tensor("keys_43_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor keys_43_squeeze_mask_0 = const()[name = tensor("keys_43_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor keys_43_cast_fp16 = slice_by_index(begin = keys_43_begin_0, end = keys_43_end_0, end_mask = keys_43_end_mask_0, squeeze_mask = keys_43_squeeze_mask_0, x = reshape_79_cast_fp16)[name = tensor("keys_43_cast_fp16")]; + tensor values_43_begin_0 = const()[name = tensor("values_43_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor values_43_end_0 = const()[name = tensor("values_43_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor values_43_end_mask_0 = const()[name = tensor("values_43_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor values_43_squeeze_mask_0 = const()[name = tensor("values_43_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor values_43_cast_fp16 = slice_by_index(begin = values_43_begin_0, end = values_43_end_0, end_mask = values_43_end_mask_0, squeeze_mask = values_43_squeeze_mask_0, x = reshape_79_cast_fp16)[name = tensor("values_43_cast_fp16")]; + tensor var_3431_cast_fp16 = not_equal(x = keys_43_cast_fp16, y = keys_43_cast_fp16)[name = tensor("op_3431_cast_fp16")]; + tensor keys_45_cast_fp16 = select(a = var_504_to_fp16, b = keys_43_cast_fp16, cond = var_3431_cast_fp16)[name = tensor("keys_45_cast_fp16")]; + tensor var_3439_cast_fp16 = not_equal(x = values_43_cast_fp16, y = values_43_cast_fp16)[name = tensor("op_3439_cast_fp16")]; + tensor values_45_cast_fp16 = select(a = var_504_to_fp16, b = values_43_cast_fp16, cond = var_3439_cast_fp16)[name = tensor("values_45_cast_fp16")]; + tensor var_3463 = const()[name = tensor("op_3463"), val = tensor([0, 2, 1, 3])]; + tensor var_3476 = const()[name = tensor("op_3476"), val = tensor([1, 1, 1])]; + tensor var_3477_cast_fp16 = reshape(shape = var_3476, x = position7_to_fp16)[name = tensor("op_3477_cast_fp16")]; + tensor var_3494_to_fp16 = const()[name = tensor("op_3494_to_fp16"), val = tensor(0x1p+0)]; + tensor valid_len_15_cast_fp16 = add(x = var_3477_cast_fp16, y = var_3494_to_fp16)[name = tensor("valid_len_15_cast_fp16")]; + tensor valid_mask_15_cast_fp16 = less(x = k_positions_1_promoted_to_fp16, y = valid_len_15_cast_fp16)[name = tensor("valid_mask_15_cast_fp16")]; + tensor causal_mask_15_cast_fp16 = less_equal(x = k_positions_1_promoted_to_fp16, y = var_3477_cast_fp16)[name = tensor("causal_mask_15_cast_fp16")]; + tensor attn_mask_29 = logical_and(x = valid_mask_15_cast_fp16, y = causal_mask_15_cast_fp16)[name = tensor("attn_mask_29")]; + tensor attn_mask_31_axes_0 = const()[name = tensor("attn_mask_31_axes_0"), val = tensor([1])]; + tensor attn_mask_31 = expand_dims(axes = attn_mask_31_axes_0, x = attn_mask_29)[name = tensor("attn_mask_31")]; + tensor var_3512_transpose_x_0 = const()[name = tensor("op_3512_transpose_x_0"), val = tensor(false)]; + tensor var_3512_transpose_y_0 = const()[name = tensor("op_3512_transpose_y_0"), val = tensor(false)]; + tensor transpose_86_perm_0 = const()[name = tensor("transpose_86_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_87_perm_0 = const()[name = tensor("transpose_87_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_87 = transpose(perm = transpose_87_perm_0, x = keys_45_cast_fp16)[name = tensor("transpose_185")]; + tensor transpose_86 = transpose(perm = transpose_86_perm_0, x = q_45_cast_fp16)[name = tensor("transpose_186")]; + tensor var_3512_cast_fp16 = matmul(transpose_x = var_3512_transpose_x_0, transpose_y = var_3512_transpose_y_0, x = transpose_86, y = transpose_87)[name = tensor("op_3512_cast_fp16")]; + tensor var_3506_to_fp16 = const()[name = tensor("op_3506_to_fp16"), val = tensor([0x1p-3])]; + tensor attn_weights_43_cast_fp16 = mul(x = var_3512_cast_fp16, y = var_3506_to_fp16)[name = tensor("attn_weights_43_cast_fp16")]; + tensor var_3514 = logical_not(x = attn_mask_31)[name = tensor("op_3514")]; + tensor var_3515_to_fp16 = const()[name = tensor("op_3515_to_fp16"), val = tensor(-inf)]; + tensor attn_weights_45_cast_fp16 = select(a = var_3515_to_fp16, b = attn_weights_43_cast_fp16, cond = var_3514)[name = tensor("attn_weights_45_cast_fp16")]; + tensor var_3517 = const()[name = tensor("op_3517"), val = tensor(-1)]; + tensor attn_weights_47_cast_fp16 = softmax(axis = var_3517, x = attn_weights_45_cast_fp16)[name = tensor("attn_weights_47_cast_fp16")]; + tensor attn_output_15_transpose_x_0 = const()[name = tensor("attn_output_15_transpose_x_0"), val = tensor(false)]; + tensor attn_output_15_transpose_y_0 = const()[name = tensor("attn_output_15_transpose_y_0"), val = tensor(false)]; + tensor values_47_cast_fp16 = transpose(perm = var_3463, x = values_45_cast_fp16)[name = tensor("transpose_187")]; + tensor attn_output_15_cast_fp16 = matmul(transpose_x = attn_output_15_transpose_x_0, transpose_y = attn_output_15_transpose_y_0, x = attn_weights_47_cast_fp16, y = values_47_cast_fp16)[name = tensor("attn_output_15_cast_fp16")]; + tensor var_3525 = const()[name = tensor("op_3525"), val = tensor([0, 2, 1, 3])]; + tensor var_3528 = const()[name = tensor("op_3528"), val = tensor([1, 1, 1024])]; + tensor var_3526_cast_fp16 = transpose(perm = var_3525, x = attn_output_15_cast_fp16)[name = tensor("transpose_184")]; + tensor input_75_cast_fp16 = reshape(shape = var_3528, x = var_3526_cast_fp16)[name = tensor("input_75_cast_fp16")]; + tensor linear_30_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = attn7_out_proj_weight_cast_fp16, x = input_75_cast_fp16)[name = tensor("linear_30_cast_fp16")]; + tensor var_3534_to_fp16 = const()[name = tensor("op_3534_to_fp16"), val = tensor(0x1p+0)]; + tensor var_3535_cast_fp16 = add(x = position7_to_fp16, y = var_3534_to_fp16)[name = tensor("op_3535_cast_fp16")]; + tensor var_3535_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_3535_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor input_77_cast_fp16 = add(x = input_73_cast_fp16, y = linear_30_cast_fp16)[name = tensor("input_77_cast_fp16")]; + tensor input_79_axes_0 = const()[name = tensor("input_79_axes_0"), val = tensor([-1])]; + tensor norm7_2_weight_to_fp16 = const()[name = tensor("norm7_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(303649280)))]; + tensor norm7_2_bias_to_fp16 = const()[name = tensor("norm7_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(303651392)))]; + tensor var_3539_to_fp16 = const()[name = tensor("op_3539_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_79_cast_fp16 = layer_norm(axes = input_79_axes_0, beta = norm7_2_bias_to_fp16, epsilon = var_3539_to_fp16, gamma = norm7_2_weight_to_fp16, x = input_77_cast_fp16)[name = tensor("input_79_cast_fp16")]; + tensor linear_31_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = linear7_1_weight_cast_fp16, x = input_79_cast_fp16)[name = tensor("linear_31_cast_fp16")]; + tensor input_81_mode_0 = const()[name = tensor("input_81_mode_0"), val = tensor("EXACT")]; + tensor input_81_cast_fp16 = gelu(mode = input_81_mode_0, x = linear_31_cast_fp16)[name = tensor("input_81_cast_fp16")]; + tensor linear_32_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = linear7_2_weight_cast_fp16, x = input_81_cast_fp16)[name = tensor("linear_32_cast_fp16")]; + tensor input_83_cast_fp16 = add(x = input_77_cast_fp16, y = linear_32_cast_fp16)[name = tensor("input_83_cast_fp16")]; + tensor x_17_axes_0 = const()[name = tensor("x_17_axes_0"), val = tensor([-1])]; + tensor norm8_1_weight_to_fp16 = const()[name = tensor("norm8_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(303653504)))]; + tensor norm8_1_bias_to_fp16 = const()[name = tensor("norm8_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(303655616)))]; + tensor var_3556_to_fp16 = const()[name = tensor("op_3556_to_fp16"), val = tensor(0x1.5p-17)]; + tensor x_17_cast_fp16 = layer_norm(axes = x_17_axes_0, beta = norm8_1_bias_to_fp16, epsilon = var_3556_to_fp16, gamma = norm8_1_weight_to_fp16, x = input_83_cast_fp16)[name = tensor("x_17_cast_fp16")]; + tensor linear_33_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = attn8_in_proj_weight_cast_fp16, x = x_17_cast_fp16)[name = tensor("linear_33_cast_fp16")]; + tensor var_3592 = const()[name = tensor("op_3592"), val = tensor([1, 1, 3, 16, 64])]; + tensor qkv_17_cast_fp16 = reshape(shape = var_3592, x = linear_33_cast_fp16)[name = tensor("qkv_17_cast_fp16")]; + tensor q_49_begin_0 = const()[name = tensor("q_49_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor q_49_end_0 = const()[name = tensor("q_49_end_0"), val = tensor([1, 1, 1, 16, 64])]; + tensor q_49_end_mask_0 = const()[name = tensor("q_49_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor q_49_squeeze_mask_0 = const()[name = tensor("q_49_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor q_49_cast_fp16 = slice_by_index(begin = q_49_begin_0, end = q_49_end_0, end_mask = q_49_end_mask_0, squeeze_mask = q_49_squeeze_mask_0, x = qkv_17_cast_fp16)[name = tensor("q_49_cast_fp16")]; + tensor k_33_begin_0 = const()[name = tensor("k_33_begin_0"), val = tensor([0, 0, 1, 0, 0])]; + tensor k_33_end_0 = const()[name = tensor("k_33_end_0"), val = tensor([1, 1, 2, 16, 64])]; + tensor k_33_end_mask_0 = const()[name = tensor("k_33_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor k_33_squeeze_mask_0 = const()[name = tensor("k_33_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor k_33_cast_fp16 = slice_by_index(begin = k_33_begin_0, end = k_33_end_0, end_mask = k_33_end_mask_0, squeeze_mask = k_33_squeeze_mask_0, x = qkv_17_cast_fp16)[name = tensor("k_33_cast_fp16")]; + tensor v_17_begin_0 = const()[name = tensor("v_17_begin_0"), val = tensor([0, 0, 2, 0, 0])]; + tensor v_17_end_0 = const()[name = tensor("v_17_end_0"), val = tensor([1, 1, 3, 16, 64])]; + tensor v_17_end_mask_0 = const()[name = tensor("v_17_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor v_17_squeeze_mask_0 = const()[name = tensor("v_17_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor v_17_cast_fp16 = slice_by_index(begin = v_17_begin_0, end = v_17_end_0, end_mask = v_17_end_mask_0, squeeze_mask = v_17_squeeze_mask_0, x = qkv_17_cast_fp16)[name = tensor("v_17_cast_fp16")]; + tensor var_3696 = const()[name = tensor("op_3696"), val = tensor([1, 1, 1, 1])]; + tensor position8_to_fp16_dtype_0 = const()[name = tensor("position8_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor position8_to_fp16 = cast(dtype = position8_to_fp16_dtype_0, x = position8)[name = tensor("cast_513")]; + tensor ts_53_cast_fp16 = reshape(shape = var_3696, x = position8_to_fp16)[name = tensor("ts_53_cast_fp16")]; + tensor var_3700 = const()[name = tensor("op_3700"), val = tensor([1, 1, 16, 32, 2])]; + tensor q_complex_17_cast_fp16 = reshape(shape = var_3700, x = q_49_cast_fp16)[name = tensor("q_complex_17_cast_fp16")]; + tensor var_3704 = const()[name = tensor("op_3704"), val = tensor([1, 1, 16, 32, 2])]; + tensor k_complex_17_cast_fp16 = reshape(shape = var_3704, x = k_33_cast_fp16)[name = tensor("k_complex_17_cast_fp16")]; + tensor var_3708_begin_0 = const()[name = tensor("op_3708_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_3708_end_0 = const()[name = tensor("op_3708_end_0"), val = tensor([1, 1, 16, 32, 1])]; + tensor var_3708_end_mask_0 = const()[name = tensor("op_3708_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_3708_squeeze_mask_0 = const()[name = tensor("op_3708_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_3708_cast_fp16 = slice_by_index(begin = var_3708_begin_0, end = var_3708_end_0, end_mask = var_3708_end_mask_0, squeeze_mask = var_3708_squeeze_mask_0, x = q_complex_17_cast_fp16)[name = tensor("op_3708_cast_fp16")]; + tensor var_3716_begin_0 = const()[name = tensor("op_3716_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_3716_end_0 = const()[name = tensor("op_3716_end_0"), val = tensor([1, 1, 16, 32, 2])]; + tensor var_3716_end_mask_0 = const()[name = tensor("op_3716_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_3716_squeeze_mask_0 = const()[name = tensor("op_3716_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_3716_cast_fp16 = slice_by_index(begin = var_3716_begin_0, end = var_3716_end_0, end_mask = var_3716_end_mask_0, squeeze_mask = var_3716_squeeze_mask_0, x = q_complex_17_cast_fp16)[name = tensor("op_3716_cast_fp16")]; + tensor var_3724_begin_0 = const()[name = tensor("op_3724_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_3724_end_0 = const()[name = tensor("op_3724_end_0"), val = tensor([1, 1, 16, 32, 1])]; + tensor var_3724_end_mask_0 = const()[name = tensor("op_3724_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_3724_squeeze_mask_0 = const()[name = tensor("op_3724_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_3724_cast_fp16 = slice_by_index(begin = var_3724_begin_0, end = var_3724_end_0, end_mask = var_3724_end_mask_0, squeeze_mask = var_3724_squeeze_mask_0, x = k_complex_17_cast_fp16)[name = tensor("op_3724_cast_fp16")]; + tensor var_3732_begin_0 = const()[name = tensor("op_3732_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_3732_end_0 = const()[name = tensor("op_3732_end_0"), val = tensor([1, 1, 16, 32, 2])]; + tensor var_3732_end_mask_0 = const()[name = tensor("op_3732_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_3732_squeeze_mask_0 = const()[name = tensor("op_3732_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_3732_cast_fp16 = slice_by_index(begin = var_3732_begin_0, end = var_3732_end_0, end_mask = var_3732_end_mask_0, squeeze_mask = var_3732_squeeze_mask_0, x = k_complex_17_cast_fp16)[name = tensor("op_3732_cast_fp16")]; + tensor freqs_17_to_fp16 = const()[name = tensor("freqs_17_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(303657728)))]; + tensor var_3738_cast_fp16 = mul(x = freqs_17_to_fp16, y = ts_53_cast_fp16)[name = tensor("op_3738_cast_fp16")]; + tensor rotr_17_cast_fp16 = cos(x = var_3738_cast_fp16)[name = tensor("rotr_17_cast_fp16")]; + tensor roti_17_cast_fp16 = sin(x = var_3738_cast_fp16)[name = tensor("roti_17_cast_fp16")]; + tensor var_3742_cast_fp16 = mul(x = var_3708_cast_fp16, y = rotr_17_cast_fp16)[name = tensor("op_3742_cast_fp16")]; + tensor var_3743_cast_fp16 = mul(x = var_3716_cast_fp16, y = roti_17_cast_fp16)[name = tensor("op_3743_cast_fp16")]; + tensor qor_33_cast_fp16 = sub(x = var_3742_cast_fp16, y = var_3743_cast_fp16)[name = tensor("qor_33_cast_fp16")]; + tensor var_3746_cast_fp16 = mul(x = var_3708_cast_fp16, y = roti_17_cast_fp16)[name = tensor("op_3746_cast_fp16")]; + tensor var_3747_cast_fp16 = mul(x = var_3716_cast_fp16, y = rotr_17_cast_fp16)[name = tensor("op_3747_cast_fp16")]; + tensor qoi_33_cast_fp16 = add(x = var_3746_cast_fp16, y = var_3747_cast_fp16)[name = tensor("qoi_33_cast_fp16")]; + tensor var_3750_cast_fp16 = mul(x = var_3724_cast_fp16, y = rotr_17_cast_fp16)[name = tensor("op_3750_cast_fp16")]; + tensor var_3751_cast_fp16 = mul(x = var_3732_cast_fp16, y = roti_17_cast_fp16)[name = tensor("op_3751_cast_fp16")]; + tensor kor_33_cast_fp16 = sub(x = var_3750_cast_fp16, y = var_3751_cast_fp16)[name = tensor("kor_33_cast_fp16")]; + tensor var_3754_cast_fp16 = mul(x = var_3724_cast_fp16, y = roti_17_cast_fp16)[name = tensor("op_3754_cast_fp16")]; + tensor var_3755_cast_fp16 = mul(x = var_3732_cast_fp16, y = rotr_17_cast_fp16)[name = tensor("op_3755_cast_fp16")]; + tensor koi_33_cast_fp16 = add(x = var_3754_cast_fp16, y = var_3755_cast_fp16)[name = tensor("koi_33_cast_fp16")]; + tensor qo_17_axis_0 = const()[name = tensor("qo_17_axis_0"), val = tensor(-1)]; + tensor qo_17_cast_fp16 = stack(axis = qo_17_axis_0, values = (qor_33_cast_fp16, qoi_33_cast_fp16))[name = tensor("qo_17_cast_fp16")]; + tensor ko_17_axis_0 = const()[name = tensor("ko_17_axis_0"), val = tensor(-1)]; + tensor ko_17_cast_fp16 = stack(axis = ko_17_axis_0, values = (kor_33_cast_fp16, koi_33_cast_fp16))[name = tensor("ko_17_cast_fp16")]; + tensor var_3784 = const()[name = tensor("op_3784"), val = tensor([1, 1, 16, 64])]; + tensor q_51_cast_fp16 = reshape(shape = var_3784, x = qo_17_cast_fp16)[name = tensor("q_51_cast_fp16")]; + tensor var_3786 = const()[name = tensor("op_3786"), val = tensor([1, 1, 16, 64])]; + tensor k_35_cast_fp16 = reshape(shape = var_3786, x = ko_17_cast_fp16)[name = tensor("k_35_cast_fp16")]; + tensor _inversed_3808_y_0_to_fp16 = const()[name = tensor("_inversed_3808_y_0_to_fp16"), val = tensor(0x1p-9)]; + tensor _inversed_3808_cast_fp16 = mul(x = ts_53_cast_fp16, y = _inversed_3808_y_0_to_fp16)[name = tensor("_inversed_3808_cast_fp16")]; + tensor var_3809_cast_fp16 = floor(x = _inversed_3808_cast_fp16)[name = tensor("op_3809_cast_fp16")]; + tensor var_3810_to_fp16 = const()[name = tensor("op_3810_to_fp16"), val = tensor(0x1p+9)]; + tensor var_3811_cast_fp16 = mul(x = var_3809_cast_fp16, y = var_3810_to_fp16)[name = tensor("op_3811_cast_fp16")]; + tensor write_indices_float_35_cast_fp16 = sub(x = ts_53_cast_fp16, y = var_3811_cast_fp16)[name = tensor("write_indices_float_35_cast_fp16")]; + tensor var_3818_dtype_0 = const()[name = tensor("op_3818_dtype_0"), val = tensor("int32")]; + tensor write_indices_17_reps_0 = const()[name = tensor("write_indices_17_reps_0"), val = tensor([1, 1, 16, 64])]; + tensor write_indices_float_35_cast_fp16_to_int32 = cast(dtype = var_3818_dtype_0, x = write_indices_float_35_cast_fp16)[name = tensor("cast_512")]; + tensor write_indices_17 = tile(reps = write_indices_17_reps_0, x = write_indices_float_35_cast_fp16_to_int32)[name = tensor("write_indices_17")]; + tensor var_3826_begin_0 = const()[name = tensor("op_3826_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_3826_end_0 = const()[name = tensor("op_3826_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor var_3826_end_mask_0 = const()[name = tensor("op_3826_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_3826_squeeze_mask_0 = const()[name = tensor("op_3826_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor cache8_to_fp16_dtype_0 = const()[name = tensor("cache8_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor cache8_to_fp16 = cast(dtype = cache8_to_fp16_dtype_0, x = cache8)[name = tensor("cast_511")]; + tensor var_3826_cast_fp16 = slice_by_index(begin = var_3826_begin_0, end = var_3826_end_0, end_mask = var_3826_end_mask_0, squeeze_mask = var_3826_squeeze_mask_0, x = cache8_to_fp16)[name = tensor("op_3826_cast_fp16")]; + tensor var_3828_axis_0 = const()[name = tensor("op_3828_axis_0"), val = tensor(1)]; + tensor var_3828_mode_0 = const()[name = tensor("op_3828_mode_0"), val = tensor("update")]; + tensor var_3828_validate_indices_0 = const()[name = tensor("op_3828_validate_indices_0"), val = tensor(false)]; + tensor var_3828_cast_fp16 = scatter_along_axis(axis = var_3828_axis_0, data = var_3826_cast_fp16, indices = write_indices_17, mode = var_3828_mode_0, updates = k_35_cast_fp16, validate_indices = var_3828_validate_indices_0)[name = tensor("op_3828_cast_fp16")]; + tensor concat_58 = const()[name = tensor("concat_58"), val = tensor([0, 0, 0, 0, 0])]; + tensor concat_59 = const()[name = tensor("concat_59"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_17_internal_tensor_assign_1_stride_0 = const()[name = tensor("new_cache_17_internal_tensor_assign_1_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_17_internal_tensor_assign_1_begin_mask_0 = const()[name = tensor("new_cache_17_internal_tensor_assign_1_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_17_internal_tensor_assign_1_end_mask_0 = const()[name = tensor("new_cache_17_internal_tensor_assign_1_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_17_internal_tensor_assign_1_squeeze_mask_0 = const()[name = tensor("new_cache_17_internal_tensor_assign_1_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_64 = const()[name = tensor("shape_64"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_16 = const()[name = tensor("reduce_prod_16"), val = tensor(1048576)]; + tensor range_1d_16_start_0 = const()[name = tensor("range_1d_16_start_0"), val = tensor(0)]; + tensor range_1d_16_step_0 = const()[name = tensor("range_1d_16_step_0"), val = tensor(1)]; + tensor range_1d_16 = range_1d(end = reduce_prod_16, start = range_1d_16_start_0, step = range_1d_16_step_0)[name = tensor("range_1d_16")]; + tensor reshape_80 = reshape(shape = shape_64, x = range_1d_16)[name = tensor("reshape_80")]; + tensor slice_by_index_16 = slice_by_index(begin = concat_58, begin_mask = new_cache_17_internal_tensor_assign_1_begin_mask_0, end = concat_59, end_mask = new_cache_17_internal_tensor_assign_1_end_mask_0, squeeze_mask = new_cache_17_internal_tensor_assign_1_squeeze_mask_0, stride = new_cache_17_internal_tensor_assign_1_stride_0, x = reshape_80)[name = tensor("slice_by_index_16")]; + tensor reshape_81_shape_0 = const()[name = tensor("reshape_81_shape_0"), val = tensor([-1])]; + tensor reshape_81 = reshape(shape = reshape_81_shape_0, x = slice_by_index_16)[name = tensor("reshape_81")]; + tensor reshape_82_shape_0 = const()[name = tensor("reshape_82_shape_0"), val = tensor([-1])]; + tensor reshape_82_cast_fp16 = reshape(shape = reshape_82_shape_0, x = var_3828_cast_fp16)[name = tensor("reshape_82_cast_fp16")]; + tensor reshape_83_shape_0 = const()[name = tensor("reshape_83_shape_0"), val = tensor([-1])]; + tensor reshape_83_cast_fp16 = reshape(shape = reshape_83_shape_0, x = cache8_to_fp16)[name = tensor("reshape_83_cast_fp16")]; + tensor scatter_16_mode_0 = const()[name = tensor("scatter_16_mode_0"), val = tensor("update")]; + tensor scatter_16_axis_0 = const()[name = tensor("scatter_16_axis_0"), val = tensor(0)]; + tensor scatter_16_validate_indices_0 = const()[name = tensor("scatter_16_validate_indices_0"), val = tensor(false)]; + tensor scatter_16_cast_fp16 = scatter(axis = scatter_16_axis_0, data = reshape_83_cast_fp16, indices = reshape_81, mode = scatter_16_mode_0, updates = reshape_82_cast_fp16, validate_indices = scatter_16_validate_indices_0)[name = tensor("scatter_16_cast_fp16")]; + tensor reshape_84_cast_fp16 = reshape(shape = shape_64, x = scatter_16_cast_fp16)[name = tensor("reshape_84_cast_fp16")]; + tensor var_3836_begin_0 = const()[name = tensor("op_3836_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor var_3836_end_0 = const()[name = tensor("op_3836_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor var_3836_end_mask_0 = const()[name = tensor("op_3836_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_3836_squeeze_mask_0 = const()[name = tensor("op_3836_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor var_3836_cast_fp16 = slice_by_index(begin = var_3836_begin_0, end = var_3836_end_0, end_mask = var_3836_end_mask_0, squeeze_mask = var_3836_squeeze_mask_0, x = reshape_84_cast_fp16)[name = tensor("op_3836_cast_fp16")]; + tensor var_3838_axis_0 = const()[name = tensor("op_3838_axis_0"), val = tensor(1)]; + tensor var_3838_mode_0 = const()[name = tensor("op_3838_mode_0"), val = tensor("update")]; + tensor var_3838_validate_indices_0 = const()[name = tensor("op_3838_validate_indices_0"), val = tensor(false)]; + tensor var_3838_cast_fp16 = scatter_along_axis(axis = var_3838_axis_0, data = var_3836_cast_fp16, indices = write_indices_17, mode = var_3838_mode_0, updates = v_17_cast_fp16, validate_indices = var_3838_validate_indices_0)[name = tensor("op_3838_cast_fp16")]; + tensor concat_60 = const()[name = tensor("concat_60"), val = tensor([1, 0, 0, 0, 0])]; + tensor concat_61 = const()[name = tensor("concat_61"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_17_internal_tensor_assign_2_stride_0 = const()[name = tensor("new_cache_17_internal_tensor_assign_2_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_17_internal_tensor_assign_2_begin_mask_0 = const()[name = tensor("new_cache_17_internal_tensor_assign_2_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_17_internal_tensor_assign_2_end_mask_0 = const()[name = tensor("new_cache_17_internal_tensor_assign_2_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_17_internal_tensor_assign_2_squeeze_mask_0 = const()[name = tensor("new_cache_17_internal_tensor_assign_2_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_65 = const()[name = tensor("shape_65"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_17 = const()[name = tensor("reduce_prod_17"), val = tensor(1048576)]; + tensor range_1d_17_start_0 = const()[name = tensor("range_1d_17_start_0"), val = tensor(0)]; + tensor range_1d_17_step_0 = const()[name = tensor("range_1d_17_step_0"), val = tensor(1)]; + tensor range_1d_17 = range_1d(end = reduce_prod_17, start = range_1d_17_start_0, step = range_1d_17_step_0)[name = tensor("range_1d_17")]; + tensor reshape_85 = reshape(shape = shape_65, x = range_1d_17)[name = tensor("reshape_85")]; + tensor slice_by_index_17 = slice_by_index(begin = concat_60, begin_mask = new_cache_17_internal_tensor_assign_2_begin_mask_0, end = concat_61, end_mask = new_cache_17_internal_tensor_assign_2_end_mask_0, squeeze_mask = new_cache_17_internal_tensor_assign_2_squeeze_mask_0, stride = new_cache_17_internal_tensor_assign_2_stride_0, x = reshape_85)[name = tensor("slice_by_index_17")]; + tensor reshape_86_shape_0 = const()[name = tensor("reshape_86_shape_0"), val = tensor([-1])]; + tensor reshape_86 = reshape(shape = reshape_86_shape_0, x = slice_by_index_17)[name = tensor("reshape_86")]; + tensor reshape_87_shape_0 = const()[name = tensor("reshape_87_shape_0"), val = tensor([-1])]; + tensor reshape_87_cast_fp16 = reshape(shape = reshape_87_shape_0, x = var_3838_cast_fp16)[name = tensor("reshape_87_cast_fp16")]; + tensor reshape_88_shape_0 = const()[name = tensor("reshape_88_shape_0"), val = tensor([-1])]; + tensor reshape_88_cast_fp16 = reshape(shape = reshape_88_shape_0, x = reshape_84_cast_fp16)[name = tensor("reshape_88_cast_fp16")]; + tensor scatter_17_mode_0 = const()[name = tensor("scatter_17_mode_0"), val = tensor("update")]; + tensor scatter_17_axis_0 = const()[name = tensor("scatter_17_axis_0"), val = tensor(0)]; + tensor scatter_17_validate_indices_0 = const()[name = tensor("scatter_17_validate_indices_0"), val = tensor(false)]; + tensor scatter_17_cast_fp16 = scatter(axis = scatter_17_axis_0, data = reshape_88_cast_fp16, indices = reshape_86, mode = scatter_17_mode_0, updates = reshape_87_cast_fp16, validate_indices = scatter_17_validate_indices_0)[name = tensor("scatter_17_cast_fp16")]; + tensor reshape_89_cast_fp16 = reshape(shape = shape_65, x = scatter_17_cast_fp16)[name = tensor("reshape_89_cast_fp16")]; + tensor reshape_89_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("reshape_89_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor keys_49_begin_0 = const()[name = tensor("keys_49_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor keys_49_end_0 = const()[name = tensor("keys_49_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor keys_49_end_mask_0 = const()[name = tensor("keys_49_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor keys_49_squeeze_mask_0 = const()[name = tensor("keys_49_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor keys_49_cast_fp16 = slice_by_index(begin = keys_49_begin_0, end = keys_49_end_0, end_mask = keys_49_end_mask_0, squeeze_mask = keys_49_squeeze_mask_0, x = reshape_89_cast_fp16)[name = tensor("keys_49_cast_fp16")]; + tensor values_49_begin_0 = const()[name = tensor("values_49_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor values_49_end_0 = const()[name = tensor("values_49_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor values_49_end_mask_0 = const()[name = tensor("values_49_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor values_49_squeeze_mask_0 = const()[name = tensor("values_49_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor values_49_cast_fp16 = slice_by_index(begin = values_49_begin_0, end = values_49_end_0, end_mask = values_49_end_mask_0, squeeze_mask = values_49_squeeze_mask_0, x = reshape_89_cast_fp16)[name = tensor("values_49_cast_fp16")]; + tensor var_3850_cast_fp16 = not_equal(x = keys_49_cast_fp16, y = keys_49_cast_fp16)[name = tensor("op_3850_cast_fp16")]; + tensor keys_51_cast_fp16 = select(a = var_504_to_fp16, b = keys_49_cast_fp16, cond = var_3850_cast_fp16)[name = tensor("keys_51_cast_fp16")]; + tensor var_3858_cast_fp16 = not_equal(x = values_49_cast_fp16, y = values_49_cast_fp16)[name = tensor("op_3858_cast_fp16")]; + tensor values_51_cast_fp16 = select(a = var_504_to_fp16, b = values_49_cast_fp16, cond = var_3858_cast_fp16)[name = tensor("values_51_cast_fp16")]; + tensor var_3882 = const()[name = tensor("op_3882"), val = tensor([0, 2, 1, 3])]; + tensor var_3895 = const()[name = tensor("op_3895"), val = tensor([1, 1, 1])]; + tensor var_3896_cast_fp16 = reshape(shape = var_3895, x = position8_to_fp16)[name = tensor("op_3896_cast_fp16")]; + tensor var_3913_to_fp16 = const()[name = tensor("op_3913_to_fp16"), val = tensor(0x1p+0)]; + tensor valid_len_17_cast_fp16 = add(x = var_3896_cast_fp16, y = var_3913_to_fp16)[name = tensor("valid_len_17_cast_fp16")]; + tensor valid_mask_17_cast_fp16 = less(x = k_positions_1_promoted_to_fp16, y = valid_len_17_cast_fp16)[name = tensor("valid_mask_17_cast_fp16")]; + tensor causal_mask_17_cast_fp16 = less_equal(x = k_positions_1_promoted_to_fp16, y = var_3896_cast_fp16)[name = tensor("causal_mask_17_cast_fp16")]; + tensor attn_mask_33 = logical_and(x = valid_mask_17_cast_fp16, y = causal_mask_17_cast_fp16)[name = tensor("attn_mask_33")]; + tensor attn_mask_35_axes_0 = const()[name = tensor("attn_mask_35_axes_0"), val = tensor([1])]; + tensor attn_mask_35 = expand_dims(axes = attn_mask_35_axes_0, x = attn_mask_33)[name = tensor("attn_mask_35")]; + tensor var_3931_transpose_x_0 = const()[name = tensor("op_3931_transpose_x_0"), val = tensor(false)]; + tensor var_3931_transpose_y_0 = const()[name = tensor("op_3931_transpose_y_0"), val = tensor(false)]; + tensor transpose_88_perm_0 = const()[name = tensor("transpose_88_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_89_perm_0 = const()[name = tensor("transpose_89_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_89 = transpose(perm = transpose_89_perm_0, x = keys_51_cast_fp16)[name = tensor("transpose_181")]; + tensor transpose_88 = transpose(perm = transpose_88_perm_0, x = q_51_cast_fp16)[name = tensor("transpose_182")]; + tensor var_3931_cast_fp16 = matmul(transpose_x = var_3931_transpose_x_0, transpose_y = var_3931_transpose_y_0, x = transpose_88, y = transpose_89)[name = tensor("op_3931_cast_fp16")]; + tensor var_3925_to_fp16 = const()[name = tensor("op_3925_to_fp16"), val = tensor([0x1p-3])]; + tensor attn_weights_49_cast_fp16 = mul(x = var_3931_cast_fp16, y = var_3925_to_fp16)[name = tensor("attn_weights_49_cast_fp16")]; + tensor var_3933 = logical_not(x = attn_mask_35)[name = tensor("op_3933")]; + tensor var_3934_to_fp16 = const()[name = tensor("op_3934_to_fp16"), val = tensor(-inf)]; + tensor attn_weights_51_cast_fp16 = select(a = var_3934_to_fp16, b = attn_weights_49_cast_fp16, cond = var_3933)[name = tensor("attn_weights_51_cast_fp16")]; + tensor var_3936 = const()[name = tensor("op_3936"), val = tensor(-1)]; + tensor attn_weights_53_cast_fp16 = softmax(axis = var_3936, x = attn_weights_51_cast_fp16)[name = tensor("attn_weights_53_cast_fp16")]; + tensor attn_output_17_transpose_x_0 = const()[name = tensor("attn_output_17_transpose_x_0"), val = tensor(false)]; + tensor attn_output_17_transpose_y_0 = const()[name = tensor("attn_output_17_transpose_y_0"), val = tensor(false)]; + tensor values_53_cast_fp16 = transpose(perm = var_3882, x = values_51_cast_fp16)[name = tensor("transpose_183")]; + tensor attn_output_17_cast_fp16 = matmul(transpose_x = attn_output_17_transpose_x_0, transpose_y = attn_output_17_transpose_y_0, x = attn_weights_53_cast_fp16, y = values_53_cast_fp16)[name = tensor("attn_output_17_cast_fp16")]; + tensor var_3944 = const()[name = tensor("op_3944"), val = tensor([0, 2, 1, 3])]; + tensor var_3947 = const()[name = tensor("op_3947"), val = tensor([1, 1, 1024])]; + tensor var_3945_cast_fp16 = transpose(perm = var_3944, x = attn_output_17_cast_fp16)[name = tensor("transpose_180")]; + tensor input_85_cast_fp16 = reshape(shape = var_3947, x = var_3945_cast_fp16)[name = tensor("input_85_cast_fp16")]; + tensor linear_34_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = attn8_out_proj_weight_cast_fp16, x = input_85_cast_fp16)[name = tensor("linear_34_cast_fp16")]; + tensor var_3953_to_fp16 = const()[name = tensor("op_3953_to_fp16"), val = tensor(0x1p+0)]; + tensor var_3954_cast_fp16 = add(x = position8_to_fp16, y = var_3953_to_fp16)[name = tensor("op_3954_cast_fp16")]; + tensor var_3954_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_3954_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor input_87_cast_fp16 = add(x = input_83_cast_fp16, y = linear_34_cast_fp16)[name = tensor("input_87_cast_fp16")]; + tensor input_89_axes_0 = const()[name = tensor("input_89_axes_0"), val = tensor([-1])]; + tensor norm8_2_weight_to_fp16 = const()[name = tensor("norm8_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(303657856)))]; + tensor norm8_2_bias_to_fp16 = const()[name = tensor("norm8_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(303659968)))]; + tensor var_3958_to_fp16 = const()[name = tensor("op_3958_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_89_cast_fp16 = layer_norm(axes = input_89_axes_0, beta = norm8_2_bias_to_fp16, epsilon = var_3958_to_fp16, gamma = norm8_2_weight_to_fp16, x = input_87_cast_fp16)[name = tensor("input_89_cast_fp16")]; + tensor linear_35_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = linear8_1_weight_cast_fp16, x = input_89_cast_fp16)[name = tensor("linear_35_cast_fp16")]; + tensor input_91_mode_0 = const()[name = tensor("input_91_mode_0"), val = tensor("EXACT")]; + tensor input_91_cast_fp16 = gelu(mode = input_91_mode_0, x = linear_35_cast_fp16)[name = tensor("input_91_cast_fp16")]; + tensor linear_36_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = linear8_2_weight_cast_fp16, x = input_91_cast_fp16)[name = tensor("linear_36_cast_fp16")]; + tensor input_93_cast_fp16 = add(x = input_87_cast_fp16, y = linear_36_cast_fp16)[name = tensor("input_93_cast_fp16")]; + tensor x_19_axes_0 = const()[name = tensor("x_19_axes_0"), val = tensor([-1])]; + tensor norm9_1_weight_to_fp16 = const()[name = tensor("norm9_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(303662080)))]; + tensor norm9_1_bias_to_fp16 = const()[name = tensor("norm9_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(303664192)))]; + tensor var_3975_to_fp16 = const()[name = tensor("op_3975_to_fp16"), val = tensor(0x1.5p-17)]; + tensor x_19_cast_fp16 = layer_norm(axes = x_19_axes_0, beta = norm9_1_bias_to_fp16, epsilon = var_3975_to_fp16, gamma = norm9_1_weight_to_fp16, x = input_93_cast_fp16)[name = tensor("x_19_cast_fp16")]; + tensor linear_37_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = attn9_in_proj_weight_cast_fp16, x = x_19_cast_fp16)[name = tensor("linear_37_cast_fp16")]; + tensor var_4011 = const()[name = tensor("op_4011"), val = tensor([1, 1, 3, 16, 64])]; + tensor qkv_19_cast_fp16 = reshape(shape = var_4011, x = linear_37_cast_fp16)[name = tensor("qkv_19_cast_fp16")]; + tensor q_55_begin_0 = const()[name = tensor("q_55_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor q_55_end_0 = const()[name = tensor("q_55_end_0"), val = tensor([1, 1, 1, 16, 64])]; + tensor q_55_end_mask_0 = const()[name = tensor("q_55_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor q_55_squeeze_mask_0 = const()[name = tensor("q_55_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor q_55_cast_fp16 = slice_by_index(begin = q_55_begin_0, end = q_55_end_0, end_mask = q_55_end_mask_0, squeeze_mask = q_55_squeeze_mask_0, x = qkv_19_cast_fp16)[name = tensor("q_55_cast_fp16")]; + tensor k_37_begin_0 = const()[name = tensor("k_37_begin_0"), val = tensor([0, 0, 1, 0, 0])]; + tensor k_37_end_0 = const()[name = tensor("k_37_end_0"), val = tensor([1, 1, 2, 16, 64])]; + tensor k_37_end_mask_0 = const()[name = tensor("k_37_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor k_37_squeeze_mask_0 = const()[name = tensor("k_37_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor k_37_cast_fp16 = slice_by_index(begin = k_37_begin_0, end = k_37_end_0, end_mask = k_37_end_mask_0, squeeze_mask = k_37_squeeze_mask_0, x = qkv_19_cast_fp16)[name = tensor("k_37_cast_fp16")]; + tensor v_19_begin_0 = const()[name = tensor("v_19_begin_0"), val = tensor([0, 0, 2, 0, 0])]; + tensor v_19_end_0 = const()[name = tensor("v_19_end_0"), val = tensor([1, 1, 3, 16, 64])]; + tensor v_19_end_mask_0 = const()[name = tensor("v_19_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor v_19_squeeze_mask_0 = const()[name = tensor("v_19_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor v_19_cast_fp16 = slice_by_index(begin = v_19_begin_0, end = v_19_end_0, end_mask = v_19_end_mask_0, squeeze_mask = v_19_squeeze_mask_0, x = qkv_19_cast_fp16)[name = tensor("v_19_cast_fp16")]; + tensor var_4115 = const()[name = tensor("op_4115"), val = tensor([1, 1, 1, 1])]; + tensor position9_to_fp16_dtype_0 = const()[name = tensor("position9_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor position9_to_fp16 = cast(dtype = position9_to_fp16_dtype_0, x = position9)[name = tensor("cast_508")]; + tensor ts_59_cast_fp16 = reshape(shape = var_4115, x = position9_to_fp16)[name = tensor("ts_59_cast_fp16")]; + tensor var_4119 = const()[name = tensor("op_4119"), val = tensor([1, 1, 16, 32, 2])]; + tensor q_complex_19_cast_fp16 = reshape(shape = var_4119, x = q_55_cast_fp16)[name = tensor("q_complex_19_cast_fp16")]; + tensor var_4123 = const()[name = tensor("op_4123"), val = tensor([1, 1, 16, 32, 2])]; + tensor k_complex_19_cast_fp16 = reshape(shape = var_4123, x = k_37_cast_fp16)[name = tensor("k_complex_19_cast_fp16")]; + tensor var_4127_begin_0 = const()[name = tensor("op_4127_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_4127_end_0 = const()[name = tensor("op_4127_end_0"), val = tensor([1, 1, 16, 32, 1])]; + tensor var_4127_end_mask_0 = const()[name = tensor("op_4127_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_4127_squeeze_mask_0 = const()[name = tensor("op_4127_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_4127_cast_fp16 = slice_by_index(begin = var_4127_begin_0, end = var_4127_end_0, end_mask = var_4127_end_mask_0, squeeze_mask = var_4127_squeeze_mask_0, x = q_complex_19_cast_fp16)[name = tensor("op_4127_cast_fp16")]; + tensor var_4135_begin_0 = const()[name = tensor("op_4135_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_4135_end_0 = const()[name = tensor("op_4135_end_0"), val = tensor([1, 1, 16, 32, 2])]; + tensor var_4135_end_mask_0 = const()[name = tensor("op_4135_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_4135_squeeze_mask_0 = const()[name = tensor("op_4135_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_4135_cast_fp16 = slice_by_index(begin = var_4135_begin_0, end = var_4135_end_0, end_mask = var_4135_end_mask_0, squeeze_mask = var_4135_squeeze_mask_0, x = q_complex_19_cast_fp16)[name = tensor("op_4135_cast_fp16")]; + tensor var_4143_begin_0 = const()[name = tensor("op_4143_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_4143_end_0 = const()[name = tensor("op_4143_end_0"), val = tensor([1, 1, 16, 32, 1])]; + tensor var_4143_end_mask_0 = const()[name = tensor("op_4143_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_4143_squeeze_mask_0 = const()[name = tensor("op_4143_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_4143_cast_fp16 = slice_by_index(begin = var_4143_begin_0, end = var_4143_end_0, end_mask = var_4143_end_mask_0, squeeze_mask = var_4143_squeeze_mask_0, x = k_complex_19_cast_fp16)[name = tensor("op_4143_cast_fp16")]; + tensor var_4151_begin_0 = const()[name = tensor("op_4151_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_4151_end_0 = const()[name = tensor("op_4151_end_0"), val = tensor([1, 1, 16, 32, 2])]; + tensor var_4151_end_mask_0 = const()[name = tensor("op_4151_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_4151_squeeze_mask_0 = const()[name = tensor("op_4151_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_4151_cast_fp16 = slice_by_index(begin = var_4151_begin_0, end = var_4151_end_0, end_mask = var_4151_end_mask_0, squeeze_mask = var_4151_squeeze_mask_0, x = k_complex_19_cast_fp16)[name = tensor("op_4151_cast_fp16")]; + tensor freqs_19_to_fp16 = const()[name = tensor("freqs_19_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(303666304)))]; + tensor var_4157_cast_fp16 = mul(x = freqs_19_to_fp16, y = ts_59_cast_fp16)[name = tensor("op_4157_cast_fp16")]; + tensor rotr_19_cast_fp16 = cos(x = var_4157_cast_fp16)[name = tensor("rotr_19_cast_fp16")]; + tensor roti_19_cast_fp16 = sin(x = var_4157_cast_fp16)[name = tensor("roti_19_cast_fp16")]; + tensor var_4161_cast_fp16 = mul(x = var_4127_cast_fp16, y = rotr_19_cast_fp16)[name = tensor("op_4161_cast_fp16")]; + tensor var_4162_cast_fp16 = mul(x = var_4135_cast_fp16, y = roti_19_cast_fp16)[name = tensor("op_4162_cast_fp16")]; + tensor qor_37_cast_fp16 = sub(x = var_4161_cast_fp16, y = var_4162_cast_fp16)[name = tensor("qor_37_cast_fp16")]; + tensor var_4165_cast_fp16 = mul(x = var_4127_cast_fp16, y = roti_19_cast_fp16)[name = tensor("op_4165_cast_fp16")]; + tensor var_4166_cast_fp16 = mul(x = var_4135_cast_fp16, y = rotr_19_cast_fp16)[name = tensor("op_4166_cast_fp16")]; + tensor qoi_37_cast_fp16 = add(x = var_4165_cast_fp16, y = var_4166_cast_fp16)[name = tensor("qoi_37_cast_fp16")]; + tensor var_4169_cast_fp16 = mul(x = var_4143_cast_fp16, y = rotr_19_cast_fp16)[name = tensor("op_4169_cast_fp16")]; + tensor var_4170_cast_fp16 = mul(x = var_4151_cast_fp16, y = roti_19_cast_fp16)[name = tensor("op_4170_cast_fp16")]; + tensor kor_37_cast_fp16 = sub(x = var_4169_cast_fp16, y = var_4170_cast_fp16)[name = tensor("kor_37_cast_fp16")]; + tensor var_4173_cast_fp16 = mul(x = var_4143_cast_fp16, y = roti_19_cast_fp16)[name = tensor("op_4173_cast_fp16")]; + tensor var_4174_cast_fp16 = mul(x = var_4151_cast_fp16, y = rotr_19_cast_fp16)[name = tensor("op_4174_cast_fp16")]; + tensor koi_37_cast_fp16 = add(x = var_4173_cast_fp16, y = var_4174_cast_fp16)[name = tensor("koi_37_cast_fp16")]; + tensor qo_19_axis_0 = const()[name = tensor("qo_19_axis_0"), val = tensor(-1)]; + tensor qo_19_cast_fp16 = stack(axis = qo_19_axis_0, values = (qor_37_cast_fp16, qoi_37_cast_fp16))[name = tensor("qo_19_cast_fp16")]; + tensor ko_19_axis_0 = const()[name = tensor("ko_19_axis_0"), val = tensor(-1)]; + tensor ko_19_cast_fp16 = stack(axis = ko_19_axis_0, values = (kor_37_cast_fp16, koi_37_cast_fp16))[name = tensor("ko_19_cast_fp16")]; + tensor var_4203 = const()[name = tensor("op_4203"), val = tensor([1, 1, 16, 64])]; + tensor q_57_cast_fp16 = reshape(shape = var_4203, x = qo_19_cast_fp16)[name = tensor("q_57_cast_fp16")]; + tensor var_4205 = const()[name = tensor("op_4205"), val = tensor([1, 1, 16, 64])]; + tensor k_39_cast_fp16 = reshape(shape = var_4205, x = ko_19_cast_fp16)[name = tensor("k_39_cast_fp16")]; + tensor _inversed_4227_y_0_to_fp16 = const()[name = tensor("_inversed_4227_y_0_to_fp16"), val = tensor(0x1p-9)]; + tensor _inversed_4227_cast_fp16 = mul(x = ts_59_cast_fp16, y = _inversed_4227_y_0_to_fp16)[name = tensor("_inversed_4227_cast_fp16")]; + tensor var_4228_cast_fp16 = floor(x = _inversed_4227_cast_fp16)[name = tensor("op_4228_cast_fp16")]; + tensor var_4229_to_fp16 = const()[name = tensor("op_4229_to_fp16"), val = tensor(0x1p+9)]; + tensor var_4230_cast_fp16 = mul(x = var_4228_cast_fp16, y = var_4229_to_fp16)[name = tensor("op_4230_cast_fp16")]; + tensor write_indices_float_39_cast_fp16 = sub(x = ts_59_cast_fp16, y = var_4230_cast_fp16)[name = tensor("write_indices_float_39_cast_fp16")]; + tensor var_4237_dtype_0 = const()[name = tensor("op_4237_dtype_0"), val = tensor("int32")]; + tensor write_indices_19_reps_0 = const()[name = tensor("write_indices_19_reps_0"), val = tensor([1, 1, 16, 64])]; + tensor write_indices_float_39_cast_fp16_to_int32 = cast(dtype = var_4237_dtype_0, x = write_indices_float_39_cast_fp16)[name = tensor("cast_507")]; + tensor write_indices_19 = tile(reps = write_indices_19_reps_0, x = write_indices_float_39_cast_fp16_to_int32)[name = tensor("write_indices_19")]; + tensor var_4245_begin_0 = const()[name = tensor("op_4245_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_4245_end_0 = const()[name = tensor("op_4245_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor var_4245_end_mask_0 = const()[name = tensor("op_4245_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_4245_squeeze_mask_0 = const()[name = tensor("op_4245_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor cache9_to_fp16_dtype_0 = const()[name = tensor("cache9_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor cache9_to_fp16 = cast(dtype = cache9_to_fp16_dtype_0, x = cache9)[name = tensor("cast_506")]; + tensor var_4245_cast_fp16 = slice_by_index(begin = var_4245_begin_0, end = var_4245_end_0, end_mask = var_4245_end_mask_0, squeeze_mask = var_4245_squeeze_mask_0, x = cache9_to_fp16)[name = tensor("op_4245_cast_fp16")]; + tensor var_4247_axis_0 = const()[name = tensor("op_4247_axis_0"), val = tensor(1)]; + tensor var_4247_mode_0 = const()[name = tensor("op_4247_mode_0"), val = tensor("update")]; + tensor var_4247_validate_indices_0 = const()[name = tensor("op_4247_validate_indices_0"), val = tensor(false)]; + tensor var_4247_cast_fp16 = scatter_along_axis(axis = var_4247_axis_0, data = var_4245_cast_fp16, indices = write_indices_19, mode = var_4247_mode_0, updates = k_39_cast_fp16, validate_indices = var_4247_validate_indices_0)[name = tensor("op_4247_cast_fp16")]; + tensor concat_65 = const()[name = tensor("concat_65"), val = tensor([0, 0, 0, 0, 0])]; + tensor concat_66 = const()[name = tensor("concat_66"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_19_internal_tensor_assign_1_stride_0 = const()[name = tensor("new_cache_19_internal_tensor_assign_1_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_19_internal_tensor_assign_1_begin_mask_0 = const()[name = tensor("new_cache_19_internal_tensor_assign_1_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_19_internal_tensor_assign_1_end_mask_0 = const()[name = tensor("new_cache_19_internal_tensor_assign_1_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_19_internal_tensor_assign_1_squeeze_mask_0 = const()[name = tensor("new_cache_19_internal_tensor_assign_1_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_66 = const()[name = tensor("shape_66"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_18 = const()[name = tensor("reduce_prod_18"), val = tensor(1048576)]; + tensor range_1d_18_start_0 = const()[name = tensor("range_1d_18_start_0"), val = tensor(0)]; + tensor range_1d_18_step_0 = const()[name = tensor("range_1d_18_step_0"), val = tensor(1)]; + tensor range_1d_18 = range_1d(end = reduce_prod_18, start = range_1d_18_start_0, step = range_1d_18_step_0)[name = tensor("range_1d_18")]; + tensor reshape_90 = reshape(shape = shape_66, x = range_1d_18)[name = tensor("reshape_90")]; + tensor slice_by_index_18 = slice_by_index(begin = concat_65, begin_mask = new_cache_19_internal_tensor_assign_1_begin_mask_0, end = concat_66, end_mask = new_cache_19_internal_tensor_assign_1_end_mask_0, squeeze_mask = new_cache_19_internal_tensor_assign_1_squeeze_mask_0, stride = new_cache_19_internal_tensor_assign_1_stride_0, x = reshape_90)[name = tensor("slice_by_index_18")]; + tensor reshape_91_shape_0 = const()[name = tensor("reshape_91_shape_0"), val = tensor([-1])]; + tensor reshape_91 = reshape(shape = reshape_91_shape_0, x = slice_by_index_18)[name = tensor("reshape_91")]; + tensor reshape_92_shape_0 = const()[name = tensor("reshape_92_shape_0"), val = tensor([-1])]; + tensor reshape_92_cast_fp16 = reshape(shape = reshape_92_shape_0, x = var_4247_cast_fp16)[name = tensor("reshape_92_cast_fp16")]; + tensor reshape_93_shape_0 = const()[name = tensor("reshape_93_shape_0"), val = tensor([-1])]; + tensor reshape_93_cast_fp16 = reshape(shape = reshape_93_shape_0, x = cache9_to_fp16)[name = tensor("reshape_93_cast_fp16")]; + tensor scatter_18_mode_0 = const()[name = tensor("scatter_18_mode_0"), val = tensor("update")]; + tensor scatter_18_axis_0 = const()[name = tensor("scatter_18_axis_0"), val = tensor(0)]; + tensor scatter_18_validate_indices_0 = const()[name = tensor("scatter_18_validate_indices_0"), val = tensor(false)]; + tensor scatter_18_cast_fp16 = scatter(axis = scatter_18_axis_0, data = reshape_93_cast_fp16, indices = reshape_91, mode = scatter_18_mode_0, updates = reshape_92_cast_fp16, validate_indices = scatter_18_validate_indices_0)[name = tensor("scatter_18_cast_fp16")]; + tensor reshape_94_cast_fp16 = reshape(shape = shape_66, x = scatter_18_cast_fp16)[name = tensor("reshape_94_cast_fp16")]; + tensor var_4255_begin_0 = const()[name = tensor("op_4255_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor var_4255_end_0 = const()[name = tensor("op_4255_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor var_4255_end_mask_0 = const()[name = tensor("op_4255_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_4255_squeeze_mask_0 = const()[name = tensor("op_4255_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor var_4255_cast_fp16 = slice_by_index(begin = var_4255_begin_0, end = var_4255_end_0, end_mask = var_4255_end_mask_0, squeeze_mask = var_4255_squeeze_mask_0, x = reshape_94_cast_fp16)[name = tensor("op_4255_cast_fp16")]; + tensor var_4257_axis_0 = const()[name = tensor("op_4257_axis_0"), val = tensor(1)]; + tensor var_4257_mode_0 = const()[name = tensor("op_4257_mode_0"), val = tensor("update")]; + tensor var_4257_validate_indices_0 = const()[name = tensor("op_4257_validate_indices_0"), val = tensor(false)]; + tensor var_4257_cast_fp16 = scatter_along_axis(axis = var_4257_axis_0, data = var_4255_cast_fp16, indices = write_indices_19, mode = var_4257_mode_0, updates = v_19_cast_fp16, validate_indices = var_4257_validate_indices_0)[name = tensor("op_4257_cast_fp16")]; + tensor concat_67 = const()[name = tensor("concat_67"), val = tensor([1, 0, 0, 0, 0])]; + tensor concat_68 = const()[name = tensor("concat_68"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_19_internal_tensor_assign_2_stride_0 = const()[name = tensor("new_cache_19_internal_tensor_assign_2_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_19_internal_tensor_assign_2_begin_mask_0 = const()[name = tensor("new_cache_19_internal_tensor_assign_2_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_19_internal_tensor_assign_2_end_mask_0 = const()[name = tensor("new_cache_19_internal_tensor_assign_2_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_19_internal_tensor_assign_2_squeeze_mask_0 = const()[name = tensor("new_cache_19_internal_tensor_assign_2_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_67 = const()[name = tensor("shape_67"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_19 = const()[name = tensor("reduce_prod_19"), val = tensor(1048576)]; + tensor range_1d_19_start_0 = const()[name = tensor("range_1d_19_start_0"), val = tensor(0)]; + tensor range_1d_19_step_0 = const()[name = tensor("range_1d_19_step_0"), val = tensor(1)]; + tensor range_1d_19 = range_1d(end = reduce_prod_19, start = range_1d_19_start_0, step = range_1d_19_step_0)[name = tensor("range_1d_19")]; + tensor reshape_95 = reshape(shape = shape_67, x = range_1d_19)[name = tensor("reshape_95")]; + tensor slice_by_index_19 = slice_by_index(begin = concat_67, begin_mask = new_cache_19_internal_tensor_assign_2_begin_mask_0, end = concat_68, end_mask = new_cache_19_internal_tensor_assign_2_end_mask_0, squeeze_mask = new_cache_19_internal_tensor_assign_2_squeeze_mask_0, stride = new_cache_19_internal_tensor_assign_2_stride_0, x = reshape_95)[name = tensor("slice_by_index_19")]; + tensor reshape_96_shape_0 = const()[name = tensor("reshape_96_shape_0"), val = tensor([-1])]; + tensor reshape_96 = reshape(shape = reshape_96_shape_0, x = slice_by_index_19)[name = tensor("reshape_96")]; + tensor reshape_97_shape_0 = const()[name = tensor("reshape_97_shape_0"), val = tensor([-1])]; + tensor reshape_97_cast_fp16 = reshape(shape = reshape_97_shape_0, x = var_4257_cast_fp16)[name = tensor("reshape_97_cast_fp16")]; + tensor reshape_98_shape_0 = const()[name = tensor("reshape_98_shape_0"), val = tensor([-1])]; + tensor reshape_98_cast_fp16 = reshape(shape = reshape_98_shape_0, x = reshape_94_cast_fp16)[name = tensor("reshape_98_cast_fp16")]; + tensor scatter_19_mode_0 = const()[name = tensor("scatter_19_mode_0"), val = tensor("update")]; + tensor scatter_19_axis_0 = const()[name = tensor("scatter_19_axis_0"), val = tensor(0)]; + tensor scatter_19_validate_indices_0 = const()[name = tensor("scatter_19_validate_indices_0"), val = tensor(false)]; + tensor scatter_19_cast_fp16 = scatter(axis = scatter_19_axis_0, data = reshape_98_cast_fp16, indices = reshape_96, mode = scatter_19_mode_0, updates = reshape_97_cast_fp16, validate_indices = scatter_19_validate_indices_0)[name = tensor("scatter_19_cast_fp16")]; + tensor reshape_99_cast_fp16 = reshape(shape = shape_67, x = scatter_19_cast_fp16)[name = tensor("reshape_99_cast_fp16")]; + tensor reshape_99_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("reshape_99_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor keys_55_begin_0 = const()[name = tensor("keys_55_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor keys_55_end_0 = const()[name = tensor("keys_55_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor keys_55_end_mask_0 = const()[name = tensor("keys_55_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor keys_55_squeeze_mask_0 = const()[name = tensor("keys_55_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor keys_55_cast_fp16 = slice_by_index(begin = keys_55_begin_0, end = keys_55_end_0, end_mask = keys_55_end_mask_0, squeeze_mask = keys_55_squeeze_mask_0, x = reshape_99_cast_fp16)[name = tensor("keys_55_cast_fp16")]; + tensor values_55_begin_0 = const()[name = tensor("values_55_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor values_55_end_0 = const()[name = tensor("values_55_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor values_55_end_mask_0 = const()[name = tensor("values_55_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor values_55_squeeze_mask_0 = const()[name = tensor("values_55_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor values_55_cast_fp16 = slice_by_index(begin = values_55_begin_0, end = values_55_end_0, end_mask = values_55_end_mask_0, squeeze_mask = values_55_squeeze_mask_0, x = reshape_99_cast_fp16)[name = tensor("values_55_cast_fp16")]; + tensor var_4269_cast_fp16 = not_equal(x = keys_55_cast_fp16, y = keys_55_cast_fp16)[name = tensor("op_4269_cast_fp16")]; + tensor keys_57_cast_fp16 = select(a = var_504_to_fp16, b = keys_55_cast_fp16, cond = var_4269_cast_fp16)[name = tensor("keys_57_cast_fp16")]; + tensor var_4277_cast_fp16 = not_equal(x = values_55_cast_fp16, y = values_55_cast_fp16)[name = tensor("op_4277_cast_fp16")]; + tensor values_57_cast_fp16 = select(a = var_504_to_fp16, b = values_55_cast_fp16, cond = var_4277_cast_fp16)[name = tensor("values_57_cast_fp16")]; + tensor var_4301 = const()[name = tensor("op_4301"), val = tensor([0, 2, 1, 3])]; + tensor var_4314 = const()[name = tensor("op_4314"), val = tensor([1, 1, 1])]; + tensor var_4315_cast_fp16 = reshape(shape = var_4314, x = position9_to_fp16)[name = tensor("op_4315_cast_fp16")]; + tensor var_4332_to_fp16 = const()[name = tensor("op_4332_to_fp16"), val = tensor(0x1p+0)]; + tensor valid_len_19_cast_fp16 = add(x = var_4315_cast_fp16, y = var_4332_to_fp16)[name = tensor("valid_len_19_cast_fp16")]; + tensor valid_mask_19_cast_fp16 = less(x = k_positions_1_promoted_to_fp16, y = valid_len_19_cast_fp16)[name = tensor("valid_mask_19_cast_fp16")]; + tensor causal_mask_19_cast_fp16 = less_equal(x = k_positions_1_promoted_to_fp16, y = var_4315_cast_fp16)[name = tensor("causal_mask_19_cast_fp16")]; + tensor attn_mask_37 = logical_and(x = valid_mask_19_cast_fp16, y = causal_mask_19_cast_fp16)[name = tensor("attn_mask_37")]; + tensor attn_mask_39_axes_0 = const()[name = tensor("attn_mask_39_axes_0"), val = tensor([1])]; + tensor attn_mask_39 = expand_dims(axes = attn_mask_39_axes_0, x = attn_mask_37)[name = tensor("attn_mask_39")]; + tensor var_4350_transpose_x_0 = const()[name = tensor("op_4350_transpose_x_0"), val = tensor(false)]; + tensor var_4350_transpose_y_0 = const()[name = tensor("op_4350_transpose_y_0"), val = tensor(false)]; + tensor transpose_90_perm_0 = const()[name = tensor("transpose_90_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_91_perm_0 = const()[name = tensor("transpose_91_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_91 = transpose(perm = transpose_91_perm_0, x = keys_57_cast_fp16)[name = tensor("transpose_177")]; + tensor transpose_90 = transpose(perm = transpose_90_perm_0, x = q_57_cast_fp16)[name = tensor("transpose_178")]; + tensor var_4350_cast_fp16 = matmul(transpose_x = var_4350_transpose_x_0, transpose_y = var_4350_transpose_y_0, x = transpose_90, y = transpose_91)[name = tensor("op_4350_cast_fp16")]; + tensor var_4344_to_fp16 = const()[name = tensor("op_4344_to_fp16"), val = tensor([0x1p-3])]; + tensor attn_weights_55_cast_fp16 = mul(x = var_4350_cast_fp16, y = var_4344_to_fp16)[name = tensor("attn_weights_55_cast_fp16")]; + tensor var_4352 = logical_not(x = attn_mask_39)[name = tensor("op_4352")]; + tensor var_4353_to_fp16 = const()[name = tensor("op_4353_to_fp16"), val = tensor(-inf)]; + tensor attn_weights_57_cast_fp16 = select(a = var_4353_to_fp16, b = attn_weights_55_cast_fp16, cond = var_4352)[name = tensor("attn_weights_57_cast_fp16")]; + tensor var_4355 = const()[name = tensor("op_4355"), val = tensor(-1)]; + tensor attn_weights_59_cast_fp16 = softmax(axis = var_4355, x = attn_weights_57_cast_fp16)[name = tensor("attn_weights_59_cast_fp16")]; + tensor attn_output_19_transpose_x_0 = const()[name = tensor("attn_output_19_transpose_x_0"), val = tensor(false)]; + tensor attn_output_19_transpose_y_0 = const()[name = tensor("attn_output_19_transpose_y_0"), val = tensor(false)]; + tensor values_59_cast_fp16 = transpose(perm = var_4301, x = values_57_cast_fp16)[name = tensor("transpose_179")]; + tensor attn_output_19_cast_fp16 = matmul(transpose_x = attn_output_19_transpose_x_0, transpose_y = attn_output_19_transpose_y_0, x = attn_weights_59_cast_fp16, y = values_59_cast_fp16)[name = tensor("attn_output_19_cast_fp16")]; + tensor var_4363 = const()[name = tensor("op_4363"), val = tensor([0, 2, 1, 3])]; + tensor var_4366 = const()[name = tensor("op_4366"), val = tensor([1, 1, 1024])]; + tensor var_4364_cast_fp16 = transpose(perm = var_4363, x = attn_output_19_cast_fp16)[name = tensor("transpose_176")]; + tensor input_95_cast_fp16 = reshape(shape = var_4366, x = var_4364_cast_fp16)[name = tensor("input_95_cast_fp16")]; + tensor linear_38_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = attn9_out_proj_weight_cast_fp16, x = input_95_cast_fp16)[name = tensor("linear_38_cast_fp16")]; + tensor var_4372_to_fp16 = const()[name = tensor("op_4372_to_fp16"), val = tensor(0x1p+0)]; + tensor var_4373_cast_fp16 = add(x = position9_to_fp16, y = var_4372_to_fp16)[name = tensor("op_4373_cast_fp16")]; + tensor var_4373_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_4373_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor input_97_cast_fp16 = add(x = input_93_cast_fp16, y = linear_38_cast_fp16)[name = tensor("input_97_cast_fp16")]; + tensor input_99_axes_0 = const()[name = tensor("input_99_axes_0"), val = tensor([-1])]; + tensor norm9_2_weight_to_fp16 = const()[name = tensor("norm9_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(303666432)))]; + tensor norm9_2_bias_to_fp16 = const()[name = tensor("norm9_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(303668544)))]; + tensor var_4377_to_fp16 = const()[name = tensor("op_4377_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_99_cast_fp16 = layer_norm(axes = input_99_axes_0, beta = norm9_2_bias_to_fp16, epsilon = var_4377_to_fp16, gamma = norm9_2_weight_to_fp16, x = input_97_cast_fp16)[name = tensor("input_99_cast_fp16")]; + tensor linear_39_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = linear9_1_weight_cast_fp16, x = input_99_cast_fp16)[name = tensor("linear_39_cast_fp16")]; + tensor input_101_mode_0 = const()[name = tensor("input_101_mode_0"), val = tensor("EXACT")]; + tensor input_101_cast_fp16 = gelu(mode = input_101_mode_0, x = linear_39_cast_fp16)[name = tensor("input_101_cast_fp16")]; + tensor linear_40_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = linear9_2_weight_cast_fp16, x = input_101_cast_fp16)[name = tensor("linear_40_cast_fp16")]; + tensor input_103_cast_fp16 = add(x = input_97_cast_fp16, y = linear_40_cast_fp16)[name = tensor("input_103_cast_fp16")]; + tensor x_21_axes_0 = const()[name = tensor("x_21_axes_0"), val = tensor([-1])]; + tensor norm10_1_weight_to_fp16 = const()[name = tensor("norm10_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(303670656)))]; + tensor norm10_1_bias_to_fp16 = const()[name = tensor("norm10_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(303672768)))]; + tensor var_4394_to_fp16 = const()[name = tensor("op_4394_to_fp16"), val = tensor(0x1.5p-17)]; + tensor x_21_cast_fp16 = layer_norm(axes = x_21_axes_0, beta = norm10_1_bias_to_fp16, epsilon = var_4394_to_fp16, gamma = norm10_1_weight_to_fp16, x = input_103_cast_fp16)[name = tensor("x_21_cast_fp16")]; + tensor linear_41_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = attn10_in_proj_weight_cast_fp16, x = x_21_cast_fp16)[name = tensor("linear_41_cast_fp16")]; + tensor var_4430 = const()[name = tensor("op_4430"), val = tensor([1, 1, 3, 16, 64])]; + tensor qkv_21_cast_fp16 = reshape(shape = var_4430, x = linear_41_cast_fp16)[name = tensor("qkv_21_cast_fp16")]; + tensor q_61_begin_0 = const()[name = tensor("q_61_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor q_61_end_0 = const()[name = tensor("q_61_end_0"), val = tensor([1, 1, 1, 16, 64])]; + tensor q_61_end_mask_0 = const()[name = tensor("q_61_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor q_61_squeeze_mask_0 = const()[name = tensor("q_61_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor q_61_cast_fp16 = slice_by_index(begin = q_61_begin_0, end = q_61_end_0, end_mask = q_61_end_mask_0, squeeze_mask = q_61_squeeze_mask_0, x = qkv_21_cast_fp16)[name = tensor("q_61_cast_fp16")]; + tensor k_41_begin_0 = const()[name = tensor("k_41_begin_0"), val = tensor([0, 0, 1, 0, 0])]; + tensor k_41_end_0 = const()[name = tensor("k_41_end_0"), val = tensor([1, 1, 2, 16, 64])]; + tensor k_41_end_mask_0 = const()[name = tensor("k_41_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor k_41_squeeze_mask_0 = const()[name = tensor("k_41_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor k_41_cast_fp16 = slice_by_index(begin = k_41_begin_0, end = k_41_end_0, end_mask = k_41_end_mask_0, squeeze_mask = k_41_squeeze_mask_0, x = qkv_21_cast_fp16)[name = tensor("k_41_cast_fp16")]; + tensor v_21_begin_0 = const()[name = tensor("v_21_begin_0"), val = tensor([0, 0, 2, 0, 0])]; + tensor v_21_end_0 = const()[name = tensor("v_21_end_0"), val = tensor([1, 1, 3, 16, 64])]; + tensor v_21_end_mask_0 = const()[name = tensor("v_21_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor v_21_squeeze_mask_0 = const()[name = tensor("v_21_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor v_21_cast_fp16 = slice_by_index(begin = v_21_begin_0, end = v_21_end_0, end_mask = v_21_end_mask_0, squeeze_mask = v_21_squeeze_mask_0, x = qkv_21_cast_fp16)[name = tensor("v_21_cast_fp16")]; + tensor var_4534 = const()[name = tensor("op_4534"), val = tensor([1, 1, 1, 1])]; + tensor position10_to_fp16_dtype_0 = const()[name = tensor("position10_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor position10_to_fp16 = cast(dtype = position10_to_fp16_dtype_0, x = position10)[name = tensor("cast_503")]; + tensor ts_65_cast_fp16 = reshape(shape = var_4534, x = position10_to_fp16)[name = tensor("ts_65_cast_fp16")]; + tensor var_4538 = const()[name = tensor("op_4538"), val = tensor([1, 1, 16, 32, 2])]; + tensor q_complex_21_cast_fp16 = reshape(shape = var_4538, x = q_61_cast_fp16)[name = tensor("q_complex_21_cast_fp16")]; + tensor var_4542 = const()[name = tensor("op_4542"), val = tensor([1, 1, 16, 32, 2])]; + tensor k_complex_21_cast_fp16 = reshape(shape = var_4542, x = k_41_cast_fp16)[name = tensor("k_complex_21_cast_fp16")]; + tensor var_4546_begin_0 = const()[name = tensor("op_4546_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_4546_end_0 = const()[name = tensor("op_4546_end_0"), val = tensor([1, 1, 16, 32, 1])]; + tensor var_4546_end_mask_0 = const()[name = tensor("op_4546_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_4546_squeeze_mask_0 = const()[name = tensor("op_4546_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_4546_cast_fp16 = slice_by_index(begin = var_4546_begin_0, end = var_4546_end_0, end_mask = var_4546_end_mask_0, squeeze_mask = var_4546_squeeze_mask_0, x = q_complex_21_cast_fp16)[name = tensor("op_4546_cast_fp16")]; + tensor var_4554_begin_0 = const()[name = tensor("op_4554_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_4554_end_0 = const()[name = tensor("op_4554_end_0"), val = tensor([1, 1, 16, 32, 2])]; + tensor var_4554_end_mask_0 = const()[name = tensor("op_4554_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_4554_squeeze_mask_0 = const()[name = tensor("op_4554_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_4554_cast_fp16 = slice_by_index(begin = var_4554_begin_0, end = var_4554_end_0, end_mask = var_4554_end_mask_0, squeeze_mask = var_4554_squeeze_mask_0, x = q_complex_21_cast_fp16)[name = tensor("op_4554_cast_fp16")]; + tensor var_4562_begin_0 = const()[name = tensor("op_4562_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_4562_end_0 = const()[name = tensor("op_4562_end_0"), val = tensor([1, 1, 16, 32, 1])]; + tensor var_4562_end_mask_0 = const()[name = tensor("op_4562_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_4562_squeeze_mask_0 = const()[name = tensor("op_4562_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_4562_cast_fp16 = slice_by_index(begin = var_4562_begin_0, end = var_4562_end_0, end_mask = var_4562_end_mask_0, squeeze_mask = var_4562_squeeze_mask_0, x = k_complex_21_cast_fp16)[name = tensor("op_4562_cast_fp16")]; + tensor var_4570_begin_0 = const()[name = tensor("op_4570_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_4570_end_0 = const()[name = tensor("op_4570_end_0"), val = tensor([1, 1, 16, 32, 2])]; + tensor var_4570_end_mask_0 = const()[name = tensor("op_4570_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_4570_squeeze_mask_0 = const()[name = tensor("op_4570_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_4570_cast_fp16 = slice_by_index(begin = var_4570_begin_0, end = var_4570_end_0, end_mask = var_4570_end_mask_0, squeeze_mask = var_4570_squeeze_mask_0, x = k_complex_21_cast_fp16)[name = tensor("op_4570_cast_fp16")]; + tensor freqs_21_to_fp16 = const()[name = tensor("freqs_21_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(303674880)))]; + tensor var_4576_cast_fp16 = mul(x = freqs_21_to_fp16, y = ts_65_cast_fp16)[name = tensor("op_4576_cast_fp16")]; + tensor rotr_21_cast_fp16 = cos(x = var_4576_cast_fp16)[name = tensor("rotr_21_cast_fp16")]; + tensor roti_21_cast_fp16 = sin(x = var_4576_cast_fp16)[name = tensor("roti_21_cast_fp16")]; + tensor var_4580_cast_fp16 = mul(x = var_4546_cast_fp16, y = rotr_21_cast_fp16)[name = tensor("op_4580_cast_fp16")]; + tensor var_4581_cast_fp16 = mul(x = var_4554_cast_fp16, y = roti_21_cast_fp16)[name = tensor("op_4581_cast_fp16")]; + tensor qor_41_cast_fp16 = sub(x = var_4580_cast_fp16, y = var_4581_cast_fp16)[name = tensor("qor_41_cast_fp16")]; + tensor var_4584_cast_fp16 = mul(x = var_4546_cast_fp16, y = roti_21_cast_fp16)[name = tensor("op_4584_cast_fp16")]; + tensor var_4585_cast_fp16 = mul(x = var_4554_cast_fp16, y = rotr_21_cast_fp16)[name = tensor("op_4585_cast_fp16")]; + tensor qoi_41_cast_fp16 = add(x = var_4584_cast_fp16, y = var_4585_cast_fp16)[name = tensor("qoi_41_cast_fp16")]; + tensor var_4588_cast_fp16 = mul(x = var_4562_cast_fp16, y = rotr_21_cast_fp16)[name = tensor("op_4588_cast_fp16")]; + tensor var_4589_cast_fp16 = mul(x = var_4570_cast_fp16, y = roti_21_cast_fp16)[name = tensor("op_4589_cast_fp16")]; + tensor kor_41_cast_fp16 = sub(x = var_4588_cast_fp16, y = var_4589_cast_fp16)[name = tensor("kor_41_cast_fp16")]; + tensor var_4592_cast_fp16 = mul(x = var_4562_cast_fp16, y = roti_21_cast_fp16)[name = tensor("op_4592_cast_fp16")]; + tensor var_4593_cast_fp16 = mul(x = var_4570_cast_fp16, y = rotr_21_cast_fp16)[name = tensor("op_4593_cast_fp16")]; + tensor koi_41_cast_fp16 = add(x = var_4592_cast_fp16, y = var_4593_cast_fp16)[name = tensor("koi_41_cast_fp16")]; + tensor qo_21_axis_0 = const()[name = tensor("qo_21_axis_0"), val = tensor(-1)]; + tensor qo_21_cast_fp16 = stack(axis = qo_21_axis_0, values = (qor_41_cast_fp16, qoi_41_cast_fp16))[name = tensor("qo_21_cast_fp16")]; + tensor ko_21_axis_0 = const()[name = tensor("ko_21_axis_0"), val = tensor(-1)]; + tensor ko_21_cast_fp16 = stack(axis = ko_21_axis_0, values = (kor_41_cast_fp16, koi_41_cast_fp16))[name = tensor("ko_21_cast_fp16")]; + tensor var_4622 = const()[name = tensor("op_4622"), val = tensor([1, 1, 16, 64])]; + tensor q_63_cast_fp16 = reshape(shape = var_4622, x = qo_21_cast_fp16)[name = tensor("q_63_cast_fp16")]; + tensor var_4624 = const()[name = tensor("op_4624"), val = tensor([1, 1, 16, 64])]; + tensor k_43_cast_fp16 = reshape(shape = var_4624, x = ko_21_cast_fp16)[name = tensor("k_43_cast_fp16")]; + tensor _inversed_4646_y_0_to_fp16 = const()[name = tensor("_inversed_4646_y_0_to_fp16"), val = tensor(0x1p-9)]; + tensor _inversed_4646_cast_fp16 = mul(x = ts_65_cast_fp16, y = _inversed_4646_y_0_to_fp16)[name = tensor("_inversed_4646_cast_fp16")]; + tensor var_4647_cast_fp16 = floor(x = _inversed_4646_cast_fp16)[name = tensor("op_4647_cast_fp16")]; + tensor var_4648_to_fp16 = const()[name = tensor("op_4648_to_fp16"), val = tensor(0x1p+9)]; + tensor var_4649_cast_fp16 = mul(x = var_4647_cast_fp16, y = var_4648_to_fp16)[name = tensor("op_4649_cast_fp16")]; + tensor write_indices_float_43_cast_fp16 = sub(x = ts_65_cast_fp16, y = var_4649_cast_fp16)[name = tensor("write_indices_float_43_cast_fp16")]; + tensor var_4656_dtype_0 = const()[name = tensor("op_4656_dtype_0"), val = tensor("int32")]; + tensor write_indices_21_reps_0 = const()[name = tensor("write_indices_21_reps_0"), val = tensor([1, 1, 16, 64])]; + tensor write_indices_float_43_cast_fp16_to_int32 = cast(dtype = var_4656_dtype_0, x = write_indices_float_43_cast_fp16)[name = tensor("cast_502")]; + tensor write_indices_21 = tile(reps = write_indices_21_reps_0, x = write_indices_float_43_cast_fp16_to_int32)[name = tensor("write_indices_21")]; + tensor var_4664_begin_0 = const()[name = tensor("op_4664_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_4664_end_0 = const()[name = tensor("op_4664_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor var_4664_end_mask_0 = const()[name = tensor("op_4664_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_4664_squeeze_mask_0 = const()[name = tensor("op_4664_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor cache10_to_fp16_dtype_0 = const()[name = tensor("cache10_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor cache10_to_fp16 = cast(dtype = cache10_to_fp16_dtype_0, x = cache10)[name = tensor("cast_501")]; + tensor var_4664_cast_fp16 = slice_by_index(begin = var_4664_begin_0, end = var_4664_end_0, end_mask = var_4664_end_mask_0, squeeze_mask = var_4664_squeeze_mask_0, x = cache10_to_fp16)[name = tensor("op_4664_cast_fp16")]; + tensor var_4666_axis_0 = const()[name = tensor("op_4666_axis_0"), val = tensor(1)]; + tensor var_4666_mode_0 = const()[name = tensor("op_4666_mode_0"), val = tensor("update")]; + tensor var_4666_validate_indices_0 = const()[name = tensor("op_4666_validate_indices_0"), val = tensor(false)]; + tensor var_4666_cast_fp16 = scatter_along_axis(axis = var_4666_axis_0, data = var_4664_cast_fp16, indices = write_indices_21, mode = var_4666_mode_0, updates = k_43_cast_fp16, validate_indices = var_4666_validate_indices_0)[name = tensor("op_4666_cast_fp16")]; + tensor concat_72 = const()[name = tensor("concat_72"), val = tensor([0, 0, 0, 0, 0])]; + tensor concat_73 = const()[name = tensor("concat_73"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_21_internal_tensor_assign_1_stride_0 = const()[name = tensor("new_cache_21_internal_tensor_assign_1_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_21_internal_tensor_assign_1_begin_mask_0 = const()[name = tensor("new_cache_21_internal_tensor_assign_1_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_21_internal_tensor_assign_1_end_mask_0 = const()[name = tensor("new_cache_21_internal_tensor_assign_1_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_21_internal_tensor_assign_1_squeeze_mask_0 = const()[name = tensor("new_cache_21_internal_tensor_assign_1_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_68 = const()[name = tensor("shape_68"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_20 = const()[name = tensor("reduce_prod_20"), val = tensor(1048576)]; + tensor range_1d_20_start_0 = const()[name = tensor("range_1d_20_start_0"), val = tensor(0)]; + tensor range_1d_20_step_0 = const()[name = tensor("range_1d_20_step_0"), val = tensor(1)]; + tensor range_1d_20 = range_1d(end = reduce_prod_20, start = range_1d_20_start_0, step = range_1d_20_step_0)[name = tensor("range_1d_20")]; + tensor reshape_100 = reshape(shape = shape_68, x = range_1d_20)[name = tensor("reshape_100")]; + tensor slice_by_index_20 = slice_by_index(begin = concat_72, begin_mask = new_cache_21_internal_tensor_assign_1_begin_mask_0, end = concat_73, end_mask = new_cache_21_internal_tensor_assign_1_end_mask_0, squeeze_mask = new_cache_21_internal_tensor_assign_1_squeeze_mask_0, stride = new_cache_21_internal_tensor_assign_1_stride_0, x = reshape_100)[name = tensor("slice_by_index_20")]; + tensor reshape_101_shape_0 = const()[name = tensor("reshape_101_shape_0"), val = tensor([-1])]; + tensor reshape_101 = reshape(shape = reshape_101_shape_0, x = slice_by_index_20)[name = tensor("reshape_101")]; + tensor reshape_102_shape_0 = const()[name = tensor("reshape_102_shape_0"), val = tensor([-1])]; + tensor reshape_102_cast_fp16 = reshape(shape = reshape_102_shape_0, x = var_4666_cast_fp16)[name = tensor("reshape_102_cast_fp16")]; + tensor reshape_103_shape_0 = const()[name = tensor("reshape_103_shape_0"), val = tensor([-1])]; + tensor reshape_103_cast_fp16 = reshape(shape = reshape_103_shape_0, x = cache10_to_fp16)[name = tensor("reshape_103_cast_fp16")]; + tensor scatter_20_mode_0 = const()[name = tensor("scatter_20_mode_0"), val = tensor("update")]; + tensor scatter_20_axis_0 = const()[name = tensor("scatter_20_axis_0"), val = tensor(0)]; + tensor scatter_20_validate_indices_0 = const()[name = tensor("scatter_20_validate_indices_0"), val = tensor(false)]; + tensor scatter_20_cast_fp16 = scatter(axis = scatter_20_axis_0, data = reshape_103_cast_fp16, indices = reshape_101, mode = scatter_20_mode_0, updates = reshape_102_cast_fp16, validate_indices = scatter_20_validate_indices_0)[name = tensor("scatter_20_cast_fp16")]; + tensor reshape_104_cast_fp16 = reshape(shape = shape_68, x = scatter_20_cast_fp16)[name = tensor("reshape_104_cast_fp16")]; + tensor var_4674_begin_0 = const()[name = tensor("op_4674_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor var_4674_end_0 = const()[name = tensor("op_4674_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor var_4674_end_mask_0 = const()[name = tensor("op_4674_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_4674_squeeze_mask_0 = const()[name = tensor("op_4674_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor var_4674_cast_fp16 = slice_by_index(begin = var_4674_begin_0, end = var_4674_end_0, end_mask = var_4674_end_mask_0, squeeze_mask = var_4674_squeeze_mask_0, x = reshape_104_cast_fp16)[name = tensor("op_4674_cast_fp16")]; + tensor var_4676_axis_0 = const()[name = tensor("op_4676_axis_0"), val = tensor(1)]; + tensor var_4676_mode_0 = const()[name = tensor("op_4676_mode_0"), val = tensor("update")]; + tensor var_4676_validate_indices_0 = const()[name = tensor("op_4676_validate_indices_0"), val = tensor(false)]; + tensor var_4676_cast_fp16 = scatter_along_axis(axis = var_4676_axis_0, data = var_4674_cast_fp16, indices = write_indices_21, mode = var_4676_mode_0, updates = v_21_cast_fp16, validate_indices = var_4676_validate_indices_0)[name = tensor("op_4676_cast_fp16")]; + tensor concat_74 = const()[name = tensor("concat_74"), val = tensor([1, 0, 0, 0, 0])]; + tensor concat_75 = const()[name = tensor("concat_75"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_21_internal_tensor_assign_2_stride_0 = const()[name = tensor("new_cache_21_internal_tensor_assign_2_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_21_internal_tensor_assign_2_begin_mask_0 = const()[name = tensor("new_cache_21_internal_tensor_assign_2_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_21_internal_tensor_assign_2_end_mask_0 = const()[name = tensor("new_cache_21_internal_tensor_assign_2_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_21_internal_tensor_assign_2_squeeze_mask_0 = const()[name = tensor("new_cache_21_internal_tensor_assign_2_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_69 = const()[name = tensor("shape_69"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_21 = const()[name = tensor("reduce_prod_21"), val = tensor(1048576)]; + tensor range_1d_21_start_0 = const()[name = tensor("range_1d_21_start_0"), val = tensor(0)]; + tensor range_1d_21_step_0 = const()[name = tensor("range_1d_21_step_0"), val = tensor(1)]; + tensor range_1d_21 = range_1d(end = reduce_prod_21, start = range_1d_21_start_0, step = range_1d_21_step_0)[name = tensor("range_1d_21")]; + tensor reshape_105 = reshape(shape = shape_69, x = range_1d_21)[name = tensor("reshape_105")]; + tensor slice_by_index_21 = slice_by_index(begin = concat_74, begin_mask = new_cache_21_internal_tensor_assign_2_begin_mask_0, end = concat_75, end_mask = new_cache_21_internal_tensor_assign_2_end_mask_0, squeeze_mask = new_cache_21_internal_tensor_assign_2_squeeze_mask_0, stride = new_cache_21_internal_tensor_assign_2_stride_0, x = reshape_105)[name = tensor("slice_by_index_21")]; + tensor reshape_106_shape_0 = const()[name = tensor("reshape_106_shape_0"), val = tensor([-1])]; + tensor reshape_106 = reshape(shape = reshape_106_shape_0, x = slice_by_index_21)[name = tensor("reshape_106")]; + tensor reshape_107_shape_0 = const()[name = tensor("reshape_107_shape_0"), val = tensor([-1])]; + tensor reshape_107_cast_fp16 = reshape(shape = reshape_107_shape_0, x = var_4676_cast_fp16)[name = tensor("reshape_107_cast_fp16")]; + tensor reshape_108_shape_0 = const()[name = tensor("reshape_108_shape_0"), val = tensor([-1])]; + tensor reshape_108_cast_fp16 = reshape(shape = reshape_108_shape_0, x = reshape_104_cast_fp16)[name = tensor("reshape_108_cast_fp16")]; + tensor scatter_21_mode_0 = const()[name = tensor("scatter_21_mode_0"), val = tensor("update")]; + tensor scatter_21_axis_0 = const()[name = tensor("scatter_21_axis_0"), val = tensor(0)]; + tensor scatter_21_validate_indices_0 = const()[name = tensor("scatter_21_validate_indices_0"), val = tensor(false)]; + tensor scatter_21_cast_fp16 = scatter(axis = scatter_21_axis_0, data = reshape_108_cast_fp16, indices = reshape_106, mode = scatter_21_mode_0, updates = reshape_107_cast_fp16, validate_indices = scatter_21_validate_indices_0)[name = tensor("scatter_21_cast_fp16")]; + tensor reshape_109_cast_fp16 = reshape(shape = shape_69, x = scatter_21_cast_fp16)[name = tensor("reshape_109_cast_fp16")]; + tensor reshape_109_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("reshape_109_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor keys_61_begin_0 = const()[name = tensor("keys_61_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor keys_61_end_0 = const()[name = tensor("keys_61_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor keys_61_end_mask_0 = const()[name = tensor("keys_61_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor keys_61_squeeze_mask_0 = const()[name = tensor("keys_61_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor keys_61_cast_fp16 = slice_by_index(begin = keys_61_begin_0, end = keys_61_end_0, end_mask = keys_61_end_mask_0, squeeze_mask = keys_61_squeeze_mask_0, x = reshape_109_cast_fp16)[name = tensor("keys_61_cast_fp16")]; + tensor values_61_begin_0 = const()[name = tensor("values_61_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor values_61_end_0 = const()[name = tensor("values_61_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor values_61_end_mask_0 = const()[name = tensor("values_61_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor values_61_squeeze_mask_0 = const()[name = tensor("values_61_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor values_61_cast_fp16 = slice_by_index(begin = values_61_begin_0, end = values_61_end_0, end_mask = values_61_end_mask_0, squeeze_mask = values_61_squeeze_mask_0, x = reshape_109_cast_fp16)[name = tensor("values_61_cast_fp16")]; + tensor var_4688_cast_fp16 = not_equal(x = keys_61_cast_fp16, y = keys_61_cast_fp16)[name = tensor("op_4688_cast_fp16")]; + tensor keys_63_cast_fp16 = select(a = var_504_to_fp16, b = keys_61_cast_fp16, cond = var_4688_cast_fp16)[name = tensor("keys_63_cast_fp16")]; + tensor var_4696_cast_fp16 = not_equal(x = values_61_cast_fp16, y = values_61_cast_fp16)[name = tensor("op_4696_cast_fp16")]; + tensor values_63_cast_fp16 = select(a = var_504_to_fp16, b = values_61_cast_fp16, cond = var_4696_cast_fp16)[name = tensor("values_63_cast_fp16")]; + tensor var_4720 = const()[name = tensor("op_4720"), val = tensor([0, 2, 1, 3])]; + tensor var_4733 = const()[name = tensor("op_4733"), val = tensor([1, 1, 1])]; + tensor var_4734_cast_fp16 = reshape(shape = var_4733, x = position10_to_fp16)[name = tensor("op_4734_cast_fp16")]; + tensor var_4751_to_fp16 = const()[name = tensor("op_4751_to_fp16"), val = tensor(0x1p+0)]; + tensor valid_len_21_cast_fp16 = add(x = var_4734_cast_fp16, y = var_4751_to_fp16)[name = tensor("valid_len_21_cast_fp16")]; + tensor valid_mask_21_cast_fp16 = less(x = k_positions_1_promoted_to_fp16, y = valid_len_21_cast_fp16)[name = tensor("valid_mask_21_cast_fp16")]; + tensor causal_mask_21_cast_fp16 = less_equal(x = k_positions_1_promoted_to_fp16, y = var_4734_cast_fp16)[name = tensor("causal_mask_21_cast_fp16")]; + tensor attn_mask_41 = logical_and(x = valid_mask_21_cast_fp16, y = causal_mask_21_cast_fp16)[name = tensor("attn_mask_41")]; + tensor attn_mask_43_axes_0 = const()[name = tensor("attn_mask_43_axes_0"), val = tensor([1])]; + tensor attn_mask_43 = expand_dims(axes = attn_mask_43_axes_0, x = attn_mask_41)[name = tensor("attn_mask_43")]; + tensor var_4769_transpose_x_0 = const()[name = tensor("op_4769_transpose_x_0"), val = tensor(false)]; + tensor var_4769_transpose_y_0 = const()[name = tensor("op_4769_transpose_y_0"), val = tensor(false)]; + tensor transpose_92_perm_0 = const()[name = tensor("transpose_92_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_93_perm_0 = const()[name = tensor("transpose_93_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_93 = transpose(perm = transpose_93_perm_0, x = keys_63_cast_fp16)[name = tensor("transpose_173")]; + tensor transpose_92 = transpose(perm = transpose_92_perm_0, x = q_63_cast_fp16)[name = tensor("transpose_174")]; + tensor var_4769_cast_fp16 = matmul(transpose_x = var_4769_transpose_x_0, transpose_y = var_4769_transpose_y_0, x = transpose_92, y = transpose_93)[name = tensor("op_4769_cast_fp16")]; + tensor var_4763_to_fp16 = const()[name = tensor("op_4763_to_fp16"), val = tensor([0x1p-3])]; + tensor attn_weights_61_cast_fp16 = mul(x = var_4769_cast_fp16, y = var_4763_to_fp16)[name = tensor("attn_weights_61_cast_fp16")]; + tensor var_4771 = logical_not(x = attn_mask_43)[name = tensor("op_4771")]; + tensor var_4772_to_fp16 = const()[name = tensor("op_4772_to_fp16"), val = tensor(-inf)]; + tensor attn_weights_63_cast_fp16 = select(a = var_4772_to_fp16, b = attn_weights_61_cast_fp16, cond = var_4771)[name = tensor("attn_weights_63_cast_fp16")]; + tensor var_4774 = const()[name = tensor("op_4774"), val = tensor(-1)]; + tensor attn_weights_65_cast_fp16 = softmax(axis = var_4774, x = attn_weights_63_cast_fp16)[name = tensor("attn_weights_65_cast_fp16")]; + tensor attn_output_21_transpose_x_0 = const()[name = tensor("attn_output_21_transpose_x_0"), val = tensor(false)]; + tensor attn_output_21_transpose_y_0 = const()[name = tensor("attn_output_21_transpose_y_0"), val = tensor(false)]; + tensor values_65_cast_fp16 = transpose(perm = var_4720, x = values_63_cast_fp16)[name = tensor("transpose_175")]; + tensor attn_output_21_cast_fp16 = matmul(transpose_x = attn_output_21_transpose_x_0, transpose_y = attn_output_21_transpose_y_0, x = attn_weights_65_cast_fp16, y = values_65_cast_fp16)[name = tensor("attn_output_21_cast_fp16")]; + tensor var_4782 = const()[name = tensor("op_4782"), val = tensor([0, 2, 1, 3])]; + tensor var_4785 = const()[name = tensor("op_4785"), val = tensor([1, 1, 1024])]; + tensor var_4783_cast_fp16 = transpose(perm = var_4782, x = attn_output_21_cast_fp16)[name = tensor("transpose_172")]; + tensor input_105_cast_fp16 = reshape(shape = var_4785, x = var_4783_cast_fp16)[name = tensor("input_105_cast_fp16")]; + tensor linear_42_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = attn10_out_proj_weight_cast_fp16, x = input_105_cast_fp16)[name = tensor("linear_42_cast_fp16")]; + tensor var_4791_to_fp16 = const()[name = tensor("op_4791_to_fp16"), val = tensor(0x1p+0)]; + tensor var_4792_cast_fp16 = add(x = position10_to_fp16, y = var_4791_to_fp16)[name = tensor("op_4792_cast_fp16")]; + tensor var_4792_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_4792_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor input_107_cast_fp16 = add(x = input_103_cast_fp16, y = linear_42_cast_fp16)[name = tensor("input_107_cast_fp16")]; + tensor input_109_axes_0 = const()[name = tensor("input_109_axes_0"), val = tensor([-1])]; + tensor norm10_2_weight_to_fp16 = const()[name = tensor("norm10_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(303675008)))]; + tensor norm10_2_bias_to_fp16 = const()[name = tensor("norm10_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(303677120)))]; + tensor var_4796_to_fp16 = const()[name = tensor("op_4796_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_109_cast_fp16 = layer_norm(axes = input_109_axes_0, beta = norm10_2_bias_to_fp16, epsilon = var_4796_to_fp16, gamma = norm10_2_weight_to_fp16, x = input_107_cast_fp16)[name = tensor("input_109_cast_fp16")]; + tensor linear_43_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = linear10_1_weight_cast_fp16, x = input_109_cast_fp16)[name = tensor("linear_43_cast_fp16")]; + tensor input_111_mode_0 = const()[name = tensor("input_111_mode_0"), val = tensor("EXACT")]; + tensor input_111_cast_fp16 = gelu(mode = input_111_mode_0, x = linear_43_cast_fp16)[name = tensor("input_111_cast_fp16")]; + tensor linear_44_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = linear10_2_weight_cast_fp16, x = input_111_cast_fp16)[name = tensor("linear_44_cast_fp16")]; + tensor input_113_cast_fp16 = add(x = input_107_cast_fp16, y = linear_44_cast_fp16)[name = tensor("input_113_cast_fp16")]; + tensor x_23_axes_0 = const()[name = tensor("x_23_axes_0"), val = tensor([-1])]; + tensor norm11_1_weight_to_fp16 = const()[name = tensor("norm11_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(303679232)))]; + tensor norm11_1_bias_to_fp16 = const()[name = tensor("norm11_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(303681344)))]; + tensor var_4813_to_fp16 = const()[name = tensor("op_4813_to_fp16"), val = tensor(0x1.5p-17)]; + tensor x_23_cast_fp16 = layer_norm(axes = x_23_axes_0, beta = norm11_1_bias_to_fp16, epsilon = var_4813_to_fp16, gamma = norm11_1_weight_to_fp16, x = input_113_cast_fp16)[name = tensor("x_23_cast_fp16")]; + tensor linear_45_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = attn11_in_proj_weight_cast_fp16, x = x_23_cast_fp16)[name = tensor("linear_45_cast_fp16")]; + tensor var_4849 = const()[name = tensor("op_4849"), val = tensor([1, 1, 3, 16, 64])]; + tensor qkv_23_cast_fp16 = reshape(shape = var_4849, x = linear_45_cast_fp16)[name = tensor("qkv_23_cast_fp16")]; + tensor q_67_begin_0 = const()[name = tensor("q_67_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor q_67_end_0 = const()[name = tensor("q_67_end_0"), val = tensor([1, 1, 1, 16, 64])]; + tensor q_67_end_mask_0 = const()[name = tensor("q_67_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor q_67_squeeze_mask_0 = const()[name = tensor("q_67_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor q_67_cast_fp16 = slice_by_index(begin = q_67_begin_0, end = q_67_end_0, end_mask = q_67_end_mask_0, squeeze_mask = q_67_squeeze_mask_0, x = qkv_23_cast_fp16)[name = tensor("q_67_cast_fp16")]; + tensor k_45_begin_0 = const()[name = tensor("k_45_begin_0"), val = tensor([0, 0, 1, 0, 0])]; + tensor k_45_end_0 = const()[name = tensor("k_45_end_0"), val = tensor([1, 1, 2, 16, 64])]; + tensor k_45_end_mask_0 = const()[name = tensor("k_45_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor k_45_squeeze_mask_0 = const()[name = tensor("k_45_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor k_45_cast_fp16 = slice_by_index(begin = k_45_begin_0, end = k_45_end_0, end_mask = k_45_end_mask_0, squeeze_mask = k_45_squeeze_mask_0, x = qkv_23_cast_fp16)[name = tensor("k_45_cast_fp16")]; + tensor v_23_begin_0 = const()[name = tensor("v_23_begin_0"), val = tensor([0, 0, 2, 0, 0])]; + tensor v_23_end_0 = const()[name = tensor("v_23_end_0"), val = tensor([1, 1, 3, 16, 64])]; + tensor v_23_end_mask_0 = const()[name = tensor("v_23_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor v_23_squeeze_mask_0 = const()[name = tensor("v_23_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor v_23_cast_fp16 = slice_by_index(begin = v_23_begin_0, end = v_23_end_0, end_mask = v_23_end_mask_0, squeeze_mask = v_23_squeeze_mask_0, x = qkv_23_cast_fp16)[name = tensor("v_23_cast_fp16")]; + tensor var_4953 = const()[name = tensor("op_4953"), val = tensor([1, 1, 1, 1])]; + tensor position11_to_fp16_dtype_0 = const()[name = tensor("position11_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor position11_to_fp16 = cast(dtype = position11_to_fp16_dtype_0, x = position11)[name = tensor("cast_498")]; + tensor ts_71_cast_fp16 = reshape(shape = var_4953, x = position11_to_fp16)[name = tensor("ts_71_cast_fp16")]; + tensor var_4957 = const()[name = tensor("op_4957"), val = tensor([1, 1, 16, 32, 2])]; + tensor q_complex_23_cast_fp16 = reshape(shape = var_4957, x = q_67_cast_fp16)[name = tensor("q_complex_23_cast_fp16")]; + tensor var_4961 = const()[name = tensor("op_4961"), val = tensor([1, 1, 16, 32, 2])]; + tensor k_complex_23_cast_fp16 = reshape(shape = var_4961, x = k_45_cast_fp16)[name = tensor("k_complex_23_cast_fp16")]; + tensor var_4965_begin_0 = const()[name = tensor("op_4965_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_4965_end_0 = const()[name = tensor("op_4965_end_0"), val = tensor([1, 1, 16, 32, 1])]; + tensor var_4965_end_mask_0 = const()[name = tensor("op_4965_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_4965_squeeze_mask_0 = const()[name = tensor("op_4965_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_4965_cast_fp16 = slice_by_index(begin = var_4965_begin_0, end = var_4965_end_0, end_mask = var_4965_end_mask_0, squeeze_mask = var_4965_squeeze_mask_0, x = q_complex_23_cast_fp16)[name = tensor("op_4965_cast_fp16")]; + tensor var_4973_begin_0 = const()[name = tensor("op_4973_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_4973_end_0 = const()[name = tensor("op_4973_end_0"), val = tensor([1, 1, 16, 32, 2])]; + tensor var_4973_end_mask_0 = const()[name = tensor("op_4973_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_4973_squeeze_mask_0 = const()[name = tensor("op_4973_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_4973_cast_fp16 = slice_by_index(begin = var_4973_begin_0, end = var_4973_end_0, end_mask = var_4973_end_mask_0, squeeze_mask = var_4973_squeeze_mask_0, x = q_complex_23_cast_fp16)[name = tensor("op_4973_cast_fp16")]; + tensor var_4981_begin_0 = const()[name = tensor("op_4981_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_4981_end_0 = const()[name = tensor("op_4981_end_0"), val = tensor([1, 1, 16, 32, 1])]; + tensor var_4981_end_mask_0 = const()[name = tensor("op_4981_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_4981_squeeze_mask_0 = const()[name = tensor("op_4981_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_4981_cast_fp16 = slice_by_index(begin = var_4981_begin_0, end = var_4981_end_0, end_mask = var_4981_end_mask_0, squeeze_mask = var_4981_squeeze_mask_0, x = k_complex_23_cast_fp16)[name = tensor("op_4981_cast_fp16")]; + tensor var_4989_begin_0 = const()[name = tensor("op_4989_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_4989_end_0 = const()[name = tensor("op_4989_end_0"), val = tensor([1, 1, 16, 32, 2])]; + tensor var_4989_end_mask_0 = const()[name = tensor("op_4989_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_4989_squeeze_mask_0 = const()[name = tensor("op_4989_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_4989_cast_fp16 = slice_by_index(begin = var_4989_begin_0, end = var_4989_end_0, end_mask = var_4989_end_mask_0, squeeze_mask = var_4989_squeeze_mask_0, x = k_complex_23_cast_fp16)[name = tensor("op_4989_cast_fp16")]; + tensor freqs_23_to_fp16 = const()[name = tensor("freqs_23_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(303683456)))]; + tensor var_4995_cast_fp16 = mul(x = freqs_23_to_fp16, y = ts_71_cast_fp16)[name = tensor("op_4995_cast_fp16")]; + tensor rotr_23_cast_fp16 = cos(x = var_4995_cast_fp16)[name = tensor("rotr_23_cast_fp16")]; + tensor roti_23_cast_fp16 = sin(x = var_4995_cast_fp16)[name = tensor("roti_23_cast_fp16")]; + tensor var_4999_cast_fp16 = mul(x = var_4965_cast_fp16, y = rotr_23_cast_fp16)[name = tensor("op_4999_cast_fp16")]; + tensor var_5000_cast_fp16 = mul(x = var_4973_cast_fp16, y = roti_23_cast_fp16)[name = tensor("op_5000_cast_fp16")]; + tensor qor_45_cast_fp16 = sub(x = var_4999_cast_fp16, y = var_5000_cast_fp16)[name = tensor("qor_45_cast_fp16")]; + tensor var_5003_cast_fp16 = mul(x = var_4965_cast_fp16, y = roti_23_cast_fp16)[name = tensor("op_5003_cast_fp16")]; + tensor var_5004_cast_fp16 = mul(x = var_4973_cast_fp16, y = rotr_23_cast_fp16)[name = tensor("op_5004_cast_fp16")]; + tensor qoi_45_cast_fp16 = add(x = var_5003_cast_fp16, y = var_5004_cast_fp16)[name = tensor("qoi_45_cast_fp16")]; + tensor var_5007_cast_fp16 = mul(x = var_4981_cast_fp16, y = rotr_23_cast_fp16)[name = tensor("op_5007_cast_fp16")]; + tensor var_5008_cast_fp16 = mul(x = var_4989_cast_fp16, y = roti_23_cast_fp16)[name = tensor("op_5008_cast_fp16")]; + tensor kor_45_cast_fp16 = sub(x = var_5007_cast_fp16, y = var_5008_cast_fp16)[name = tensor("kor_45_cast_fp16")]; + tensor var_5011_cast_fp16 = mul(x = var_4981_cast_fp16, y = roti_23_cast_fp16)[name = tensor("op_5011_cast_fp16")]; + tensor var_5012_cast_fp16 = mul(x = var_4989_cast_fp16, y = rotr_23_cast_fp16)[name = tensor("op_5012_cast_fp16")]; + tensor koi_45_cast_fp16 = add(x = var_5011_cast_fp16, y = var_5012_cast_fp16)[name = tensor("koi_45_cast_fp16")]; + tensor qo_23_axis_0 = const()[name = tensor("qo_23_axis_0"), val = tensor(-1)]; + tensor qo_23_cast_fp16 = stack(axis = qo_23_axis_0, values = (qor_45_cast_fp16, qoi_45_cast_fp16))[name = tensor("qo_23_cast_fp16")]; + tensor ko_23_axis_0 = const()[name = tensor("ko_23_axis_0"), val = tensor(-1)]; + tensor ko_23_cast_fp16 = stack(axis = ko_23_axis_0, values = (kor_45_cast_fp16, koi_45_cast_fp16))[name = tensor("ko_23_cast_fp16")]; + tensor var_5041 = const()[name = tensor("op_5041"), val = tensor([1, 1, 16, 64])]; + tensor q_69_cast_fp16 = reshape(shape = var_5041, x = qo_23_cast_fp16)[name = tensor("q_69_cast_fp16")]; + tensor var_5043 = const()[name = tensor("op_5043"), val = tensor([1, 1, 16, 64])]; + tensor k_47_cast_fp16 = reshape(shape = var_5043, x = ko_23_cast_fp16)[name = tensor("k_47_cast_fp16")]; + tensor _inversed_5065_y_0_to_fp16 = const()[name = tensor("_inversed_5065_y_0_to_fp16"), val = tensor(0x1p-9)]; + tensor _inversed_5065_cast_fp16 = mul(x = ts_71_cast_fp16, y = _inversed_5065_y_0_to_fp16)[name = tensor("_inversed_5065_cast_fp16")]; + tensor var_5066_cast_fp16 = floor(x = _inversed_5065_cast_fp16)[name = tensor("op_5066_cast_fp16")]; + tensor var_5067_to_fp16 = const()[name = tensor("op_5067_to_fp16"), val = tensor(0x1p+9)]; + tensor var_5068_cast_fp16 = mul(x = var_5066_cast_fp16, y = var_5067_to_fp16)[name = tensor("op_5068_cast_fp16")]; + tensor write_indices_float_47_cast_fp16 = sub(x = ts_71_cast_fp16, y = var_5068_cast_fp16)[name = tensor("write_indices_float_47_cast_fp16")]; + tensor var_5075_dtype_0 = const()[name = tensor("op_5075_dtype_0"), val = tensor("int32")]; + tensor write_indices_23_reps_0 = const()[name = tensor("write_indices_23_reps_0"), val = tensor([1, 1, 16, 64])]; + tensor write_indices_float_47_cast_fp16_to_int32 = cast(dtype = var_5075_dtype_0, x = write_indices_float_47_cast_fp16)[name = tensor("cast_497")]; + tensor write_indices_23 = tile(reps = write_indices_23_reps_0, x = write_indices_float_47_cast_fp16_to_int32)[name = tensor("write_indices_23")]; + tensor var_5083_begin_0 = const()[name = tensor("op_5083_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_5083_end_0 = const()[name = tensor("op_5083_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor var_5083_end_mask_0 = const()[name = tensor("op_5083_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_5083_squeeze_mask_0 = const()[name = tensor("op_5083_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor cache11_to_fp16_dtype_0 = const()[name = tensor("cache11_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor cache11_to_fp16 = cast(dtype = cache11_to_fp16_dtype_0, x = cache11)[name = tensor("cast_496")]; + tensor var_5083_cast_fp16 = slice_by_index(begin = var_5083_begin_0, end = var_5083_end_0, end_mask = var_5083_end_mask_0, squeeze_mask = var_5083_squeeze_mask_0, x = cache11_to_fp16)[name = tensor("op_5083_cast_fp16")]; + tensor var_5085_axis_0 = const()[name = tensor("op_5085_axis_0"), val = tensor(1)]; + tensor var_5085_mode_0 = const()[name = tensor("op_5085_mode_0"), val = tensor("update")]; + tensor var_5085_validate_indices_0 = const()[name = tensor("op_5085_validate_indices_0"), val = tensor(false)]; + tensor var_5085_cast_fp16 = scatter_along_axis(axis = var_5085_axis_0, data = var_5083_cast_fp16, indices = write_indices_23, mode = var_5085_mode_0, updates = k_47_cast_fp16, validate_indices = var_5085_validate_indices_0)[name = tensor("op_5085_cast_fp16")]; + tensor concat_79 = const()[name = tensor("concat_79"), val = tensor([0, 0, 0, 0, 0])]; + tensor concat_80 = const()[name = tensor("concat_80"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_23_internal_tensor_assign_1_stride_0 = const()[name = tensor("new_cache_23_internal_tensor_assign_1_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_23_internal_tensor_assign_1_begin_mask_0 = const()[name = tensor("new_cache_23_internal_tensor_assign_1_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_23_internal_tensor_assign_1_end_mask_0 = const()[name = tensor("new_cache_23_internal_tensor_assign_1_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_23_internal_tensor_assign_1_squeeze_mask_0 = const()[name = tensor("new_cache_23_internal_tensor_assign_1_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_70 = const()[name = tensor("shape_70"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_22 = const()[name = tensor("reduce_prod_22"), val = tensor(1048576)]; + tensor range_1d_22_start_0 = const()[name = tensor("range_1d_22_start_0"), val = tensor(0)]; + tensor range_1d_22_step_0 = const()[name = tensor("range_1d_22_step_0"), val = tensor(1)]; + tensor range_1d_22 = range_1d(end = reduce_prod_22, start = range_1d_22_start_0, step = range_1d_22_step_0)[name = tensor("range_1d_22")]; + tensor reshape_110 = reshape(shape = shape_70, x = range_1d_22)[name = tensor("reshape_110")]; + tensor slice_by_index_22 = slice_by_index(begin = concat_79, begin_mask = new_cache_23_internal_tensor_assign_1_begin_mask_0, end = concat_80, end_mask = new_cache_23_internal_tensor_assign_1_end_mask_0, squeeze_mask = new_cache_23_internal_tensor_assign_1_squeeze_mask_0, stride = new_cache_23_internal_tensor_assign_1_stride_0, x = reshape_110)[name = tensor("slice_by_index_22")]; + tensor reshape_111_shape_0 = const()[name = tensor("reshape_111_shape_0"), val = tensor([-1])]; + tensor reshape_111 = reshape(shape = reshape_111_shape_0, x = slice_by_index_22)[name = tensor("reshape_111")]; + tensor reshape_112_shape_0 = const()[name = tensor("reshape_112_shape_0"), val = tensor([-1])]; + tensor reshape_112_cast_fp16 = reshape(shape = reshape_112_shape_0, x = var_5085_cast_fp16)[name = tensor("reshape_112_cast_fp16")]; + tensor reshape_113_shape_0 = const()[name = tensor("reshape_113_shape_0"), val = tensor([-1])]; + tensor reshape_113_cast_fp16 = reshape(shape = reshape_113_shape_0, x = cache11_to_fp16)[name = tensor("reshape_113_cast_fp16")]; + tensor scatter_22_mode_0 = const()[name = tensor("scatter_22_mode_0"), val = tensor("update")]; + tensor scatter_22_axis_0 = const()[name = tensor("scatter_22_axis_0"), val = tensor(0)]; + tensor scatter_22_validate_indices_0 = const()[name = tensor("scatter_22_validate_indices_0"), val = tensor(false)]; + tensor scatter_22_cast_fp16 = scatter(axis = scatter_22_axis_0, data = reshape_113_cast_fp16, indices = reshape_111, mode = scatter_22_mode_0, updates = reshape_112_cast_fp16, validate_indices = scatter_22_validate_indices_0)[name = tensor("scatter_22_cast_fp16")]; + tensor reshape_114_cast_fp16 = reshape(shape = shape_70, x = scatter_22_cast_fp16)[name = tensor("reshape_114_cast_fp16")]; + tensor var_5093_begin_0 = const()[name = tensor("op_5093_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor var_5093_end_0 = const()[name = tensor("op_5093_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor var_5093_end_mask_0 = const()[name = tensor("op_5093_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_5093_squeeze_mask_0 = const()[name = tensor("op_5093_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor var_5093_cast_fp16 = slice_by_index(begin = var_5093_begin_0, end = var_5093_end_0, end_mask = var_5093_end_mask_0, squeeze_mask = var_5093_squeeze_mask_0, x = reshape_114_cast_fp16)[name = tensor("op_5093_cast_fp16")]; + tensor var_5095_axis_0 = const()[name = tensor("op_5095_axis_0"), val = tensor(1)]; + tensor var_5095_mode_0 = const()[name = tensor("op_5095_mode_0"), val = tensor("update")]; + tensor var_5095_validate_indices_0 = const()[name = tensor("op_5095_validate_indices_0"), val = tensor(false)]; + tensor var_5095_cast_fp16 = scatter_along_axis(axis = var_5095_axis_0, data = var_5093_cast_fp16, indices = write_indices_23, mode = var_5095_mode_0, updates = v_23_cast_fp16, validate_indices = var_5095_validate_indices_0)[name = tensor("op_5095_cast_fp16")]; + tensor concat_81 = const()[name = tensor("concat_81"), val = tensor([1, 0, 0, 0, 0])]; + tensor concat_82 = const()[name = tensor("concat_82"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_23_internal_tensor_assign_2_stride_0 = const()[name = tensor("new_cache_23_internal_tensor_assign_2_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_23_internal_tensor_assign_2_begin_mask_0 = const()[name = tensor("new_cache_23_internal_tensor_assign_2_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_23_internal_tensor_assign_2_end_mask_0 = const()[name = tensor("new_cache_23_internal_tensor_assign_2_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_23_internal_tensor_assign_2_squeeze_mask_0 = const()[name = tensor("new_cache_23_internal_tensor_assign_2_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_71 = const()[name = tensor("shape_71"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_23 = const()[name = tensor("reduce_prod_23"), val = tensor(1048576)]; + tensor range_1d_23_start_0 = const()[name = tensor("range_1d_23_start_0"), val = tensor(0)]; + tensor range_1d_23_step_0 = const()[name = tensor("range_1d_23_step_0"), val = tensor(1)]; + tensor range_1d_23 = range_1d(end = reduce_prod_23, start = range_1d_23_start_0, step = range_1d_23_step_0)[name = tensor("range_1d_23")]; + tensor reshape_115 = reshape(shape = shape_71, x = range_1d_23)[name = tensor("reshape_115")]; + tensor slice_by_index_23 = slice_by_index(begin = concat_81, begin_mask = new_cache_23_internal_tensor_assign_2_begin_mask_0, end = concat_82, end_mask = new_cache_23_internal_tensor_assign_2_end_mask_0, squeeze_mask = new_cache_23_internal_tensor_assign_2_squeeze_mask_0, stride = new_cache_23_internal_tensor_assign_2_stride_0, x = reshape_115)[name = tensor("slice_by_index_23")]; + tensor reshape_116_shape_0 = const()[name = tensor("reshape_116_shape_0"), val = tensor([-1])]; + tensor reshape_116 = reshape(shape = reshape_116_shape_0, x = slice_by_index_23)[name = tensor("reshape_116")]; + tensor reshape_117_shape_0 = const()[name = tensor("reshape_117_shape_0"), val = tensor([-1])]; + tensor reshape_117_cast_fp16 = reshape(shape = reshape_117_shape_0, x = var_5095_cast_fp16)[name = tensor("reshape_117_cast_fp16")]; + tensor reshape_118_shape_0 = const()[name = tensor("reshape_118_shape_0"), val = tensor([-1])]; + tensor reshape_118_cast_fp16 = reshape(shape = reshape_118_shape_0, x = reshape_114_cast_fp16)[name = tensor("reshape_118_cast_fp16")]; + tensor scatter_23_mode_0 = const()[name = tensor("scatter_23_mode_0"), val = tensor("update")]; + tensor scatter_23_axis_0 = const()[name = tensor("scatter_23_axis_0"), val = tensor(0)]; + tensor scatter_23_validate_indices_0 = const()[name = tensor("scatter_23_validate_indices_0"), val = tensor(false)]; + tensor scatter_23_cast_fp16 = scatter(axis = scatter_23_axis_0, data = reshape_118_cast_fp16, indices = reshape_116, mode = scatter_23_mode_0, updates = reshape_117_cast_fp16, validate_indices = scatter_23_validate_indices_0)[name = tensor("scatter_23_cast_fp16")]; + tensor reshape_119_cast_fp16 = reshape(shape = shape_71, x = scatter_23_cast_fp16)[name = tensor("reshape_119_cast_fp16")]; + tensor reshape_119_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("reshape_119_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor keys_67_begin_0 = const()[name = tensor("keys_67_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor keys_67_end_0 = const()[name = tensor("keys_67_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor keys_67_end_mask_0 = const()[name = tensor("keys_67_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor keys_67_squeeze_mask_0 = const()[name = tensor("keys_67_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor keys_67_cast_fp16 = slice_by_index(begin = keys_67_begin_0, end = keys_67_end_0, end_mask = keys_67_end_mask_0, squeeze_mask = keys_67_squeeze_mask_0, x = reshape_119_cast_fp16)[name = tensor("keys_67_cast_fp16")]; + tensor values_67_begin_0 = const()[name = tensor("values_67_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor values_67_end_0 = const()[name = tensor("values_67_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor values_67_end_mask_0 = const()[name = tensor("values_67_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor values_67_squeeze_mask_0 = const()[name = tensor("values_67_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor values_67_cast_fp16 = slice_by_index(begin = values_67_begin_0, end = values_67_end_0, end_mask = values_67_end_mask_0, squeeze_mask = values_67_squeeze_mask_0, x = reshape_119_cast_fp16)[name = tensor("values_67_cast_fp16")]; + tensor var_5107_cast_fp16 = not_equal(x = keys_67_cast_fp16, y = keys_67_cast_fp16)[name = tensor("op_5107_cast_fp16")]; + tensor keys_69_cast_fp16 = select(a = var_504_to_fp16, b = keys_67_cast_fp16, cond = var_5107_cast_fp16)[name = tensor("keys_69_cast_fp16")]; + tensor var_5115_cast_fp16 = not_equal(x = values_67_cast_fp16, y = values_67_cast_fp16)[name = tensor("op_5115_cast_fp16")]; + tensor values_69_cast_fp16 = select(a = var_504_to_fp16, b = values_67_cast_fp16, cond = var_5115_cast_fp16)[name = tensor("values_69_cast_fp16")]; + tensor var_5139 = const()[name = tensor("op_5139"), val = tensor([0, 2, 1, 3])]; + tensor var_5152 = const()[name = tensor("op_5152"), val = tensor([1, 1, 1])]; + tensor var_5153_cast_fp16 = reshape(shape = var_5152, x = position11_to_fp16)[name = tensor("op_5153_cast_fp16")]; + tensor var_5170_to_fp16 = const()[name = tensor("op_5170_to_fp16"), val = tensor(0x1p+0)]; + tensor valid_len_23_cast_fp16 = add(x = var_5153_cast_fp16, y = var_5170_to_fp16)[name = tensor("valid_len_23_cast_fp16")]; + tensor valid_mask_23_cast_fp16 = less(x = k_positions_1_promoted_to_fp16, y = valid_len_23_cast_fp16)[name = tensor("valid_mask_23_cast_fp16")]; + tensor causal_mask_23_cast_fp16 = less_equal(x = k_positions_1_promoted_to_fp16, y = var_5153_cast_fp16)[name = tensor("causal_mask_23_cast_fp16")]; + tensor attn_mask_45 = logical_and(x = valid_mask_23_cast_fp16, y = causal_mask_23_cast_fp16)[name = tensor("attn_mask_45")]; + tensor attn_mask_47_axes_0 = const()[name = tensor("attn_mask_47_axes_0"), val = tensor([1])]; + tensor attn_mask_47 = expand_dims(axes = attn_mask_47_axes_0, x = attn_mask_45)[name = tensor("attn_mask_47")]; + tensor var_5188_transpose_x_0 = const()[name = tensor("op_5188_transpose_x_0"), val = tensor(false)]; + tensor var_5188_transpose_y_0 = const()[name = tensor("op_5188_transpose_y_0"), val = tensor(false)]; + tensor transpose_94_perm_0 = const()[name = tensor("transpose_94_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_95_perm_0 = const()[name = tensor("transpose_95_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_95 = transpose(perm = transpose_95_perm_0, x = keys_69_cast_fp16)[name = tensor("transpose_169")]; + tensor transpose_94 = transpose(perm = transpose_94_perm_0, x = q_69_cast_fp16)[name = tensor("transpose_170")]; + tensor var_5188_cast_fp16 = matmul(transpose_x = var_5188_transpose_x_0, transpose_y = var_5188_transpose_y_0, x = transpose_94, y = transpose_95)[name = tensor("op_5188_cast_fp16")]; + tensor var_5182_to_fp16 = const()[name = tensor("op_5182_to_fp16"), val = tensor([0x1p-3])]; + tensor attn_weights_67_cast_fp16 = mul(x = var_5188_cast_fp16, y = var_5182_to_fp16)[name = tensor("attn_weights_67_cast_fp16")]; + tensor var_5190 = logical_not(x = attn_mask_47)[name = tensor("op_5190")]; + tensor var_5191_to_fp16 = const()[name = tensor("op_5191_to_fp16"), val = tensor(-inf)]; + tensor attn_weights_69_cast_fp16 = select(a = var_5191_to_fp16, b = attn_weights_67_cast_fp16, cond = var_5190)[name = tensor("attn_weights_69_cast_fp16")]; + tensor var_5193 = const()[name = tensor("op_5193"), val = tensor(-1)]; + tensor attn_weights_71_cast_fp16 = softmax(axis = var_5193, x = attn_weights_69_cast_fp16)[name = tensor("attn_weights_71_cast_fp16")]; + tensor attn_output_23_transpose_x_0 = const()[name = tensor("attn_output_23_transpose_x_0"), val = tensor(false)]; + tensor attn_output_23_transpose_y_0 = const()[name = tensor("attn_output_23_transpose_y_0"), val = tensor(false)]; + tensor values_71_cast_fp16 = transpose(perm = var_5139, x = values_69_cast_fp16)[name = tensor("transpose_171")]; + tensor attn_output_23_cast_fp16 = matmul(transpose_x = attn_output_23_transpose_x_0, transpose_y = attn_output_23_transpose_y_0, x = attn_weights_71_cast_fp16, y = values_71_cast_fp16)[name = tensor("attn_output_23_cast_fp16")]; + tensor var_5201 = const()[name = tensor("op_5201"), val = tensor([0, 2, 1, 3])]; + tensor var_5204 = const()[name = tensor("op_5204"), val = tensor([1, 1, 1024])]; + tensor var_5202_cast_fp16 = transpose(perm = var_5201, x = attn_output_23_cast_fp16)[name = tensor("transpose_168")]; + tensor input_115_cast_fp16 = reshape(shape = var_5204, x = var_5202_cast_fp16)[name = tensor("input_115_cast_fp16")]; + tensor linear_46_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = attn11_out_proj_weight_cast_fp16, x = input_115_cast_fp16)[name = tensor("linear_46_cast_fp16")]; + tensor var_5210_to_fp16 = const()[name = tensor("op_5210_to_fp16"), val = tensor(0x1p+0)]; + tensor var_5211_cast_fp16 = add(x = position11_to_fp16, y = var_5210_to_fp16)[name = tensor("op_5211_cast_fp16")]; + tensor var_5211_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_5211_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor input_117_cast_fp16 = add(x = input_113_cast_fp16, y = linear_46_cast_fp16)[name = tensor("input_117_cast_fp16")]; + tensor input_119_axes_0 = const()[name = tensor("input_119_axes_0"), val = tensor([-1])]; + tensor norm11_2_weight_to_fp16 = const()[name = tensor("norm11_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(303683584)))]; + tensor norm11_2_bias_to_fp16 = const()[name = tensor("norm11_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(303685696)))]; + tensor var_5215_to_fp16 = const()[name = tensor("op_5215_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_119_cast_fp16 = layer_norm(axes = input_119_axes_0, beta = norm11_2_bias_to_fp16, epsilon = var_5215_to_fp16, gamma = norm11_2_weight_to_fp16, x = input_117_cast_fp16)[name = tensor("input_119_cast_fp16")]; + tensor linear_47_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = linear11_1_weight_cast_fp16, x = input_119_cast_fp16)[name = tensor("linear_47_cast_fp16")]; + tensor input_121_mode_0 = const()[name = tensor("input_121_mode_0"), val = tensor("EXACT")]; + tensor input_121_cast_fp16 = gelu(mode = input_121_mode_0, x = linear_47_cast_fp16)[name = tensor("input_121_cast_fp16")]; + tensor linear_48_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = linear11_2_weight_cast_fp16, x = input_121_cast_fp16)[name = tensor("linear_48_cast_fp16")]; + tensor input_123_cast_fp16 = add(x = input_117_cast_fp16, y = linear_48_cast_fp16)[name = tensor("input_123_cast_fp16")]; + tensor x_25_axes_0 = const()[name = tensor("x_25_axes_0"), val = tensor([-1])]; + tensor norm12_1_weight_to_fp16 = const()[name = tensor("norm12_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(303687808)))]; + tensor norm12_1_bias_to_fp16 = const()[name = tensor("norm12_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(303689920)))]; + tensor var_5232_to_fp16 = const()[name = tensor("op_5232_to_fp16"), val = tensor(0x1.5p-17)]; + tensor x_25_cast_fp16 = layer_norm(axes = x_25_axes_0, beta = norm12_1_bias_to_fp16, epsilon = var_5232_to_fp16, gamma = norm12_1_weight_to_fp16, x = input_123_cast_fp16)[name = tensor("x_25_cast_fp16")]; + tensor linear_49_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = attn12_in_proj_weight_cast_fp16, x = x_25_cast_fp16)[name = tensor("linear_49_cast_fp16")]; + tensor var_5268 = const()[name = tensor("op_5268"), val = tensor([1, 1, 3, 16, 64])]; + tensor qkv_25_cast_fp16 = reshape(shape = var_5268, x = linear_49_cast_fp16)[name = tensor("qkv_25_cast_fp16")]; + tensor q_73_begin_0 = const()[name = tensor("q_73_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor q_73_end_0 = const()[name = tensor("q_73_end_0"), val = tensor([1, 1, 1, 16, 64])]; + tensor q_73_end_mask_0 = const()[name = tensor("q_73_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor q_73_squeeze_mask_0 = const()[name = tensor("q_73_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor q_73_cast_fp16 = slice_by_index(begin = q_73_begin_0, end = q_73_end_0, end_mask = q_73_end_mask_0, squeeze_mask = q_73_squeeze_mask_0, x = qkv_25_cast_fp16)[name = tensor("q_73_cast_fp16")]; + tensor k_49_begin_0 = const()[name = tensor("k_49_begin_0"), val = tensor([0, 0, 1, 0, 0])]; + tensor k_49_end_0 = const()[name = tensor("k_49_end_0"), val = tensor([1, 1, 2, 16, 64])]; + tensor k_49_end_mask_0 = const()[name = tensor("k_49_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor k_49_squeeze_mask_0 = const()[name = tensor("k_49_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor k_49_cast_fp16 = slice_by_index(begin = k_49_begin_0, end = k_49_end_0, end_mask = k_49_end_mask_0, squeeze_mask = k_49_squeeze_mask_0, x = qkv_25_cast_fp16)[name = tensor("k_49_cast_fp16")]; + tensor v_25_begin_0 = const()[name = tensor("v_25_begin_0"), val = tensor([0, 0, 2, 0, 0])]; + tensor v_25_end_0 = const()[name = tensor("v_25_end_0"), val = tensor([1, 1, 3, 16, 64])]; + tensor v_25_end_mask_0 = const()[name = tensor("v_25_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor v_25_squeeze_mask_0 = const()[name = tensor("v_25_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor v_25_cast_fp16 = slice_by_index(begin = v_25_begin_0, end = v_25_end_0, end_mask = v_25_end_mask_0, squeeze_mask = v_25_squeeze_mask_0, x = qkv_25_cast_fp16)[name = tensor("v_25_cast_fp16")]; + tensor var_5372 = const()[name = tensor("op_5372"), val = tensor([1, 1, 1, 1])]; + tensor position12_to_fp16_dtype_0 = const()[name = tensor("position12_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor position12_to_fp16 = cast(dtype = position12_to_fp16_dtype_0, x = position12)[name = tensor("cast_493")]; + tensor ts_77_cast_fp16 = reshape(shape = var_5372, x = position12_to_fp16)[name = tensor("ts_77_cast_fp16")]; + tensor var_5376 = const()[name = tensor("op_5376"), val = tensor([1, 1, 16, 32, 2])]; + tensor q_complex_25_cast_fp16 = reshape(shape = var_5376, x = q_73_cast_fp16)[name = tensor("q_complex_25_cast_fp16")]; + tensor var_5380 = const()[name = tensor("op_5380"), val = tensor([1, 1, 16, 32, 2])]; + tensor k_complex_25_cast_fp16 = reshape(shape = var_5380, x = k_49_cast_fp16)[name = tensor("k_complex_25_cast_fp16")]; + tensor var_5384_begin_0 = const()[name = tensor("op_5384_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_5384_end_0 = const()[name = tensor("op_5384_end_0"), val = tensor([1, 1, 16, 32, 1])]; + tensor var_5384_end_mask_0 = const()[name = tensor("op_5384_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_5384_squeeze_mask_0 = const()[name = tensor("op_5384_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_5384_cast_fp16 = slice_by_index(begin = var_5384_begin_0, end = var_5384_end_0, end_mask = var_5384_end_mask_0, squeeze_mask = var_5384_squeeze_mask_0, x = q_complex_25_cast_fp16)[name = tensor("op_5384_cast_fp16")]; + tensor var_5392_begin_0 = const()[name = tensor("op_5392_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_5392_end_0 = const()[name = tensor("op_5392_end_0"), val = tensor([1, 1, 16, 32, 2])]; + tensor var_5392_end_mask_0 = const()[name = tensor("op_5392_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_5392_squeeze_mask_0 = const()[name = tensor("op_5392_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_5392_cast_fp16 = slice_by_index(begin = var_5392_begin_0, end = var_5392_end_0, end_mask = var_5392_end_mask_0, squeeze_mask = var_5392_squeeze_mask_0, x = q_complex_25_cast_fp16)[name = tensor("op_5392_cast_fp16")]; + tensor var_5400_begin_0 = const()[name = tensor("op_5400_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_5400_end_0 = const()[name = tensor("op_5400_end_0"), val = tensor([1, 1, 16, 32, 1])]; + tensor var_5400_end_mask_0 = const()[name = tensor("op_5400_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_5400_squeeze_mask_0 = const()[name = tensor("op_5400_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_5400_cast_fp16 = slice_by_index(begin = var_5400_begin_0, end = var_5400_end_0, end_mask = var_5400_end_mask_0, squeeze_mask = var_5400_squeeze_mask_0, x = k_complex_25_cast_fp16)[name = tensor("op_5400_cast_fp16")]; + tensor var_5408_begin_0 = const()[name = tensor("op_5408_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_5408_end_0 = const()[name = tensor("op_5408_end_0"), val = tensor([1, 1, 16, 32, 2])]; + tensor var_5408_end_mask_0 = const()[name = tensor("op_5408_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_5408_squeeze_mask_0 = const()[name = tensor("op_5408_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_5408_cast_fp16 = slice_by_index(begin = var_5408_begin_0, end = var_5408_end_0, end_mask = var_5408_end_mask_0, squeeze_mask = var_5408_squeeze_mask_0, x = k_complex_25_cast_fp16)[name = tensor("op_5408_cast_fp16")]; + tensor freqs_25_to_fp16 = const()[name = tensor("freqs_25_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(303692032)))]; + tensor var_5414_cast_fp16 = mul(x = freqs_25_to_fp16, y = ts_77_cast_fp16)[name = tensor("op_5414_cast_fp16")]; + tensor rotr_25_cast_fp16 = cos(x = var_5414_cast_fp16)[name = tensor("rotr_25_cast_fp16")]; + tensor roti_25_cast_fp16 = sin(x = var_5414_cast_fp16)[name = tensor("roti_25_cast_fp16")]; + tensor var_5418_cast_fp16 = mul(x = var_5384_cast_fp16, y = rotr_25_cast_fp16)[name = tensor("op_5418_cast_fp16")]; + tensor var_5419_cast_fp16 = mul(x = var_5392_cast_fp16, y = roti_25_cast_fp16)[name = tensor("op_5419_cast_fp16")]; + tensor qor_49_cast_fp16 = sub(x = var_5418_cast_fp16, y = var_5419_cast_fp16)[name = tensor("qor_49_cast_fp16")]; + tensor var_5422_cast_fp16 = mul(x = var_5384_cast_fp16, y = roti_25_cast_fp16)[name = tensor("op_5422_cast_fp16")]; + tensor var_5423_cast_fp16 = mul(x = var_5392_cast_fp16, y = rotr_25_cast_fp16)[name = tensor("op_5423_cast_fp16")]; + tensor qoi_49_cast_fp16 = add(x = var_5422_cast_fp16, y = var_5423_cast_fp16)[name = tensor("qoi_49_cast_fp16")]; + tensor var_5426_cast_fp16 = mul(x = var_5400_cast_fp16, y = rotr_25_cast_fp16)[name = tensor("op_5426_cast_fp16")]; + tensor var_5427_cast_fp16 = mul(x = var_5408_cast_fp16, y = roti_25_cast_fp16)[name = tensor("op_5427_cast_fp16")]; + tensor kor_49_cast_fp16 = sub(x = var_5426_cast_fp16, y = var_5427_cast_fp16)[name = tensor("kor_49_cast_fp16")]; + tensor var_5430_cast_fp16 = mul(x = var_5400_cast_fp16, y = roti_25_cast_fp16)[name = tensor("op_5430_cast_fp16")]; + tensor var_5431_cast_fp16 = mul(x = var_5408_cast_fp16, y = rotr_25_cast_fp16)[name = tensor("op_5431_cast_fp16")]; + tensor koi_49_cast_fp16 = add(x = var_5430_cast_fp16, y = var_5431_cast_fp16)[name = tensor("koi_49_cast_fp16")]; + tensor qo_25_axis_0 = const()[name = tensor("qo_25_axis_0"), val = tensor(-1)]; + tensor qo_25_cast_fp16 = stack(axis = qo_25_axis_0, values = (qor_49_cast_fp16, qoi_49_cast_fp16))[name = tensor("qo_25_cast_fp16")]; + tensor ko_25_axis_0 = const()[name = tensor("ko_25_axis_0"), val = tensor(-1)]; + tensor ko_25_cast_fp16 = stack(axis = ko_25_axis_0, values = (kor_49_cast_fp16, koi_49_cast_fp16))[name = tensor("ko_25_cast_fp16")]; + tensor var_5460 = const()[name = tensor("op_5460"), val = tensor([1, 1, 16, 64])]; + tensor q_75_cast_fp16 = reshape(shape = var_5460, x = qo_25_cast_fp16)[name = tensor("q_75_cast_fp16")]; + tensor var_5462 = const()[name = tensor("op_5462"), val = tensor([1, 1, 16, 64])]; + tensor k_51_cast_fp16 = reshape(shape = var_5462, x = ko_25_cast_fp16)[name = tensor("k_51_cast_fp16")]; + tensor _inversed_5484_y_0_to_fp16 = const()[name = tensor("_inversed_5484_y_0_to_fp16"), val = tensor(0x1p-9)]; + tensor _inversed_5484_cast_fp16 = mul(x = ts_77_cast_fp16, y = _inversed_5484_y_0_to_fp16)[name = tensor("_inversed_5484_cast_fp16")]; + tensor var_5485_cast_fp16 = floor(x = _inversed_5484_cast_fp16)[name = tensor("op_5485_cast_fp16")]; + tensor var_5486_to_fp16 = const()[name = tensor("op_5486_to_fp16"), val = tensor(0x1p+9)]; + tensor var_5487_cast_fp16 = mul(x = var_5485_cast_fp16, y = var_5486_to_fp16)[name = tensor("op_5487_cast_fp16")]; + tensor write_indices_float_51_cast_fp16 = sub(x = ts_77_cast_fp16, y = var_5487_cast_fp16)[name = tensor("write_indices_float_51_cast_fp16")]; + tensor var_5494_dtype_0 = const()[name = tensor("op_5494_dtype_0"), val = tensor("int32")]; + tensor write_indices_25_reps_0 = const()[name = tensor("write_indices_25_reps_0"), val = tensor([1, 1, 16, 64])]; + tensor write_indices_float_51_cast_fp16_to_int32 = cast(dtype = var_5494_dtype_0, x = write_indices_float_51_cast_fp16)[name = tensor("cast_492")]; + tensor write_indices_25 = tile(reps = write_indices_25_reps_0, x = write_indices_float_51_cast_fp16_to_int32)[name = tensor("write_indices_25")]; + tensor var_5502_begin_0 = const()[name = tensor("op_5502_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_5502_end_0 = const()[name = tensor("op_5502_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor var_5502_end_mask_0 = const()[name = tensor("op_5502_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_5502_squeeze_mask_0 = const()[name = tensor("op_5502_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor cache12_to_fp16_dtype_0 = const()[name = tensor("cache12_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor cache12_to_fp16 = cast(dtype = cache12_to_fp16_dtype_0, x = cache12)[name = tensor("cast_491")]; + tensor var_5502_cast_fp16 = slice_by_index(begin = var_5502_begin_0, end = var_5502_end_0, end_mask = var_5502_end_mask_0, squeeze_mask = var_5502_squeeze_mask_0, x = cache12_to_fp16)[name = tensor("op_5502_cast_fp16")]; + tensor var_5504_axis_0 = const()[name = tensor("op_5504_axis_0"), val = tensor(1)]; + tensor var_5504_mode_0 = const()[name = tensor("op_5504_mode_0"), val = tensor("update")]; + tensor var_5504_validate_indices_0 = const()[name = tensor("op_5504_validate_indices_0"), val = tensor(false)]; + tensor var_5504_cast_fp16 = scatter_along_axis(axis = var_5504_axis_0, data = var_5502_cast_fp16, indices = write_indices_25, mode = var_5504_mode_0, updates = k_51_cast_fp16, validate_indices = var_5504_validate_indices_0)[name = tensor("op_5504_cast_fp16")]; + tensor concat_86 = const()[name = tensor("concat_86"), val = tensor([0, 0, 0, 0, 0])]; + tensor concat_87 = const()[name = tensor("concat_87"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_25_internal_tensor_assign_1_stride_0 = const()[name = tensor("new_cache_25_internal_tensor_assign_1_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_25_internal_tensor_assign_1_begin_mask_0 = const()[name = tensor("new_cache_25_internal_tensor_assign_1_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_25_internal_tensor_assign_1_end_mask_0 = const()[name = tensor("new_cache_25_internal_tensor_assign_1_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_25_internal_tensor_assign_1_squeeze_mask_0 = const()[name = tensor("new_cache_25_internal_tensor_assign_1_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_72 = const()[name = tensor("shape_72"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_24 = const()[name = tensor("reduce_prod_24"), val = tensor(1048576)]; + tensor range_1d_24_start_0 = const()[name = tensor("range_1d_24_start_0"), val = tensor(0)]; + tensor range_1d_24_step_0 = const()[name = tensor("range_1d_24_step_0"), val = tensor(1)]; + tensor range_1d_24 = range_1d(end = reduce_prod_24, start = range_1d_24_start_0, step = range_1d_24_step_0)[name = tensor("range_1d_24")]; + tensor reshape_120 = reshape(shape = shape_72, x = range_1d_24)[name = tensor("reshape_120")]; + tensor slice_by_index_24 = slice_by_index(begin = concat_86, begin_mask = new_cache_25_internal_tensor_assign_1_begin_mask_0, end = concat_87, end_mask = new_cache_25_internal_tensor_assign_1_end_mask_0, squeeze_mask = new_cache_25_internal_tensor_assign_1_squeeze_mask_0, stride = new_cache_25_internal_tensor_assign_1_stride_0, x = reshape_120)[name = tensor("slice_by_index_24")]; + tensor reshape_121_shape_0 = const()[name = tensor("reshape_121_shape_0"), val = tensor([-1])]; + tensor reshape_121 = reshape(shape = reshape_121_shape_0, x = slice_by_index_24)[name = tensor("reshape_121")]; + tensor reshape_122_shape_0 = const()[name = tensor("reshape_122_shape_0"), val = tensor([-1])]; + tensor reshape_122_cast_fp16 = reshape(shape = reshape_122_shape_0, x = var_5504_cast_fp16)[name = tensor("reshape_122_cast_fp16")]; + tensor reshape_123_shape_0 = const()[name = tensor("reshape_123_shape_0"), val = tensor([-1])]; + tensor reshape_123_cast_fp16 = reshape(shape = reshape_123_shape_0, x = cache12_to_fp16)[name = tensor("reshape_123_cast_fp16")]; + tensor scatter_24_mode_0 = const()[name = tensor("scatter_24_mode_0"), val = tensor("update")]; + tensor scatter_24_axis_0 = const()[name = tensor("scatter_24_axis_0"), val = tensor(0)]; + tensor scatter_24_validate_indices_0 = const()[name = tensor("scatter_24_validate_indices_0"), val = tensor(false)]; + tensor scatter_24_cast_fp16 = scatter(axis = scatter_24_axis_0, data = reshape_123_cast_fp16, indices = reshape_121, mode = scatter_24_mode_0, updates = reshape_122_cast_fp16, validate_indices = scatter_24_validate_indices_0)[name = tensor("scatter_24_cast_fp16")]; + tensor reshape_124_cast_fp16 = reshape(shape = shape_72, x = scatter_24_cast_fp16)[name = tensor("reshape_124_cast_fp16")]; + tensor var_5512_begin_0 = const()[name = tensor("op_5512_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor var_5512_end_0 = const()[name = tensor("op_5512_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor var_5512_end_mask_0 = const()[name = tensor("op_5512_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_5512_squeeze_mask_0 = const()[name = tensor("op_5512_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor var_5512_cast_fp16 = slice_by_index(begin = var_5512_begin_0, end = var_5512_end_0, end_mask = var_5512_end_mask_0, squeeze_mask = var_5512_squeeze_mask_0, x = reshape_124_cast_fp16)[name = tensor("op_5512_cast_fp16")]; + tensor var_5514_axis_0 = const()[name = tensor("op_5514_axis_0"), val = tensor(1)]; + tensor var_5514_mode_0 = const()[name = tensor("op_5514_mode_0"), val = tensor("update")]; + tensor var_5514_validate_indices_0 = const()[name = tensor("op_5514_validate_indices_0"), val = tensor(false)]; + tensor var_5514_cast_fp16 = scatter_along_axis(axis = var_5514_axis_0, data = var_5512_cast_fp16, indices = write_indices_25, mode = var_5514_mode_0, updates = v_25_cast_fp16, validate_indices = var_5514_validate_indices_0)[name = tensor("op_5514_cast_fp16")]; + tensor concat_88 = const()[name = tensor("concat_88"), val = tensor([1, 0, 0, 0, 0])]; + tensor concat_89 = const()[name = tensor("concat_89"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_25_internal_tensor_assign_2_stride_0 = const()[name = tensor("new_cache_25_internal_tensor_assign_2_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_25_internal_tensor_assign_2_begin_mask_0 = const()[name = tensor("new_cache_25_internal_tensor_assign_2_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_25_internal_tensor_assign_2_end_mask_0 = const()[name = tensor("new_cache_25_internal_tensor_assign_2_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_25_internal_tensor_assign_2_squeeze_mask_0 = const()[name = tensor("new_cache_25_internal_tensor_assign_2_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_73 = const()[name = tensor("shape_73"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_25 = const()[name = tensor("reduce_prod_25"), val = tensor(1048576)]; + tensor range_1d_25_start_0 = const()[name = tensor("range_1d_25_start_0"), val = tensor(0)]; + tensor range_1d_25_step_0 = const()[name = tensor("range_1d_25_step_0"), val = tensor(1)]; + tensor range_1d_25 = range_1d(end = reduce_prod_25, start = range_1d_25_start_0, step = range_1d_25_step_0)[name = tensor("range_1d_25")]; + tensor reshape_125 = reshape(shape = shape_73, x = range_1d_25)[name = tensor("reshape_125")]; + tensor slice_by_index_25 = slice_by_index(begin = concat_88, begin_mask = new_cache_25_internal_tensor_assign_2_begin_mask_0, end = concat_89, end_mask = new_cache_25_internal_tensor_assign_2_end_mask_0, squeeze_mask = new_cache_25_internal_tensor_assign_2_squeeze_mask_0, stride = new_cache_25_internal_tensor_assign_2_stride_0, x = reshape_125)[name = tensor("slice_by_index_25")]; + tensor reshape_126_shape_0 = const()[name = tensor("reshape_126_shape_0"), val = tensor([-1])]; + tensor reshape_126 = reshape(shape = reshape_126_shape_0, x = slice_by_index_25)[name = tensor("reshape_126")]; + tensor reshape_127_shape_0 = const()[name = tensor("reshape_127_shape_0"), val = tensor([-1])]; + tensor reshape_127_cast_fp16 = reshape(shape = reshape_127_shape_0, x = var_5514_cast_fp16)[name = tensor("reshape_127_cast_fp16")]; + tensor reshape_128_shape_0 = const()[name = tensor("reshape_128_shape_0"), val = tensor([-1])]; + tensor reshape_128_cast_fp16 = reshape(shape = reshape_128_shape_0, x = reshape_124_cast_fp16)[name = tensor("reshape_128_cast_fp16")]; + tensor scatter_25_mode_0 = const()[name = tensor("scatter_25_mode_0"), val = tensor("update")]; + tensor scatter_25_axis_0 = const()[name = tensor("scatter_25_axis_0"), val = tensor(0)]; + tensor scatter_25_validate_indices_0 = const()[name = tensor("scatter_25_validate_indices_0"), val = tensor(false)]; + tensor scatter_25_cast_fp16 = scatter(axis = scatter_25_axis_0, data = reshape_128_cast_fp16, indices = reshape_126, mode = scatter_25_mode_0, updates = reshape_127_cast_fp16, validate_indices = scatter_25_validate_indices_0)[name = tensor("scatter_25_cast_fp16")]; + tensor reshape_129_cast_fp16 = reshape(shape = shape_73, x = scatter_25_cast_fp16)[name = tensor("reshape_129_cast_fp16")]; + tensor reshape_129_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("reshape_129_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor keys_73_begin_0 = const()[name = tensor("keys_73_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor keys_73_end_0 = const()[name = tensor("keys_73_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor keys_73_end_mask_0 = const()[name = tensor("keys_73_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor keys_73_squeeze_mask_0 = const()[name = tensor("keys_73_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor keys_73_cast_fp16 = slice_by_index(begin = keys_73_begin_0, end = keys_73_end_0, end_mask = keys_73_end_mask_0, squeeze_mask = keys_73_squeeze_mask_0, x = reshape_129_cast_fp16)[name = tensor("keys_73_cast_fp16")]; + tensor values_73_begin_0 = const()[name = tensor("values_73_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor values_73_end_0 = const()[name = tensor("values_73_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor values_73_end_mask_0 = const()[name = tensor("values_73_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor values_73_squeeze_mask_0 = const()[name = tensor("values_73_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor values_73_cast_fp16 = slice_by_index(begin = values_73_begin_0, end = values_73_end_0, end_mask = values_73_end_mask_0, squeeze_mask = values_73_squeeze_mask_0, x = reshape_129_cast_fp16)[name = tensor("values_73_cast_fp16")]; + tensor var_5526_cast_fp16 = not_equal(x = keys_73_cast_fp16, y = keys_73_cast_fp16)[name = tensor("op_5526_cast_fp16")]; + tensor keys_75_cast_fp16 = select(a = var_504_to_fp16, b = keys_73_cast_fp16, cond = var_5526_cast_fp16)[name = tensor("keys_75_cast_fp16")]; + tensor var_5534_cast_fp16 = not_equal(x = values_73_cast_fp16, y = values_73_cast_fp16)[name = tensor("op_5534_cast_fp16")]; + tensor values_75_cast_fp16 = select(a = var_504_to_fp16, b = values_73_cast_fp16, cond = var_5534_cast_fp16)[name = tensor("values_75_cast_fp16")]; + tensor var_5558 = const()[name = tensor("op_5558"), val = tensor([0, 2, 1, 3])]; + tensor var_5571 = const()[name = tensor("op_5571"), val = tensor([1, 1, 1])]; + tensor var_5572_cast_fp16 = reshape(shape = var_5571, x = position12_to_fp16)[name = tensor("op_5572_cast_fp16")]; + tensor var_5589_to_fp16 = const()[name = tensor("op_5589_to_fp16"), val = tensor(0x1p+0)]; + tensor valid_len_25_cast_fp16 = add(x = var_5572_cast_fp16, y = var_5589_to_fp16)[name = tensor("valid_len_25_cast_fp16")]; + tensor valid_mask_25_cast_fp16 = less(x = k_positions_1_promoted_to_fp16, y = valid_len_25_cast_fp16)[name = tensor("valid_mask_25_cast_fp16")]; + tensor causal_mask_25_cast_fp16 = less_equal(x = k_positions_1_promoted_to_fp16, y = var_5572_cast_fp16)[name = tensor("causal_mask_25_cast_fp16")]; + tensor attn_mask_49 = logical_and(x = valid_mask_25_cast_fp16, y = causal_mask_25_cast_fp16)[name = tensor("attn_mask_49")]; + tensor attn_mask_51_axes_0 = const()[name = tensor("attn_mask_51_axes_0"), val = tensor([1])]; + tensor attn_mask_51 = expand_dims(axes = attn_mask_51_axes_0, x = attn_mask_49)[name = tensor("attn_mask_51")]; + tensor var_5607_transpose_x_0 = const()[name = tensor("op_5607_transpose_x_0"), val = tensor(false)]; + tensor var_5607_transpose_y_0 = const()[name = tensor("op_5607_transpose_y_0"), val = tensor(false)]; + tensor transpose_96_perm_0 = const()[name = tensor("transpose_96_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_97_perm_0 = const()[name = tensor("transpose_97_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_97 = transpose(perm = transpose_97_perm_0, x = keys_75_cast_fp16)[name = tensor("transpose_165")]; + tensor transpose_96 = transpose(perm = transpose_96_perm_0, x = q_75_cast_fp16)[name = tensor("transpose_166")]; + tensor var_5607_cast_fp16 = matmul(transpose_x = var_5607_transpose_x_0, transpose_y = var_5607_transpose_y_0, x = transpose_96, y = transpose_97)[name = tensor("op_5607_cast_fp16")]; + tensor var_5601_to_fp16 = const()[name = tensor("op_5601_to_fp16"), val = tensor([0x1p-3])]; + tensor attn_weights_73_cast_fp16 = mul(x = var_5607_cast_fp16, y = var_5601_to_fp16)[name = tensor("attn_weights_73_cast_fp16")]; + tensor var_5609 = logical_not(x = attn_mask_51)[name = tensor("op_5609")]; + tensor var_5610_to_fp16 = const()[name = tensor("op_5610_to_fp16"), val = tensor(-inf)]; + tensor attn_weights_75_cast_fp16 = select(a = var_5610_to_fp16, b = attn_weights_73_cast_fp16, cond = var_5609)[name = tensor("attn_weights_75_cast_fp16")]; + tensor var_5612 = const()[name = tensor("op_5612"), val = tensor(-1)]; + tensor attn_weights_77_cast_fp16 = softmax(axis = var_5612, x = attn_weights_75_cast_fp16)[name = tensor("attn_weights_77_cast_fp16")]; + tensor attn_output_25_transpose_x_0 = const()[name = tensor("attn_output_25_transpose_x_0"), val = tensor(false)]; + tensor attn_output_25_transpose_y_0 = const()[name = tensor("attn_output_25_transpose_y_0"), val = tensor(false)]; + tensor values_77_cast_fp16 = transpose(perm = var_5558, x = values_75_cast_fp16)[name = tensor("transpose_167")]; + tensor attn_output_25_cast_fp16 = matmul(transpose_x = attn_output_25_transpose_x_0, transpose_y = attn_output_25_transpose_y_0, x = attn_weights_77_cast_fp16, y = values_77_cast_fp16)[name = tensor("attn_output_25_cast_fp16")]; + tensor var_5620 = const()[name = tensor("op_5620"), val = tensor([0, 2, 1, 3])]; + tensor var_5623 = const()[name = tensor("op_5623"), val = tensor([1, 1, 1024])]; + tensor var_5621_cast_fp16 = transpose(perm = var_5620, x = attn_output_25_cast_fp16)[name = tensor("transpose_164")]; + tensor input_125_cast_fp16 = reshape(shape = var_5623, x = var_5621_cast_fp16)[name = tensor("input_125_cast_fp16")]; + tensor linear_50_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = attn12_out_proj_weight_cast_fp16, x = input_125_cast_fp16)[name = tensor("linear_50_cast_fp16")]; + tensor var_5629_to_fp16 = const()[name = tensor("op_5629_to_fp16"), val = tensor(0x1p+0)]; + tensor var_5630_cast_fp16 = add(x = position12_to_fp16, y = var_5629_to_fp16)[name = tensor("op_5630_cast_fp16")]; + tensor var_5630_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_5630_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor input_127_cast_fp16 = add(x = input_123_cast_fp16, y = linear_50_cast_fp16)[name = tensor("input_127_cast_fp16")]; + tensor input_129_axes_0 = const()[name = tensor("input_129_axes_0"), val = tensor([-1])]; + tensor norm12_2_weight_to_fp16 = const()[name = tensor("norm12_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(303692160)))]; + tensor norm12_2_bias_to_fp16 = const()[name = tensor("norm12_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(303694272)))]; + tensor var_5634_to_fp16 = const()[name = tensor("op_5634_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_129_cast_fp16 = layer_norm(axes = input_129_axes_0, beta = norm12_2_bias_to_fp16, epsilon = var_5634_to_fp16, gamma = norm12_2_weight_to_fp16, x = input_127_cast_fp16)[name = tensor("input_129_cast_fp16")]; + tensor linear_51_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = linear12_1_weight_cast_fp16, x = input_129_cast_fp16)[name = tensor("linear_51_cast_fp16")]; + tensor input_131_mode_0 = const()[name = tensor("input_131_mode_0"), val = tensor("EXACT")]; + tensor input_131_cast_fp16 = gelu(mode = input_131_mode_0, x = linear_51_cast_fp16)[name = tensor("input_131_cast_fp16")]; + tensor linear_52_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = linear12_2_weight_cast_fp16, x = input_131_cast_fp16)[name = tensor("linear_52_cast_fp16")]; + tensor input_133_cast_fp16 = add(x = input_127_cast_fp16, y = linear_52_cast_fp16)[name = tensor("input_133_cast_fp16")]; + tensor x_27_axes_0 = const()[name = tensor("x_27_axes_0"), val = tensor([-1])]; + tensor norm13_1_weight_to_fp16 = const()[name = tensor("norm13_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(303696384)))]; + tensor norm13_1_bias_to_fp16 = const()[name = tensor("norm13_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(303698496)))]; + tensor var_5651_to_fp16 = const()[name = tensor("op_5651_to_fp16"), val = tensor(0x1.5p-17)]; + tensor x_27_cast_fp16 = layer_norm(axes = x_27_axes_0, beta = norm13_1_bias_to_fp16, epsilon = var_5651_to_fp16, gamma = norm13_1_weight_to_fp16, x = input_133_cast_fp16)[name = tensor("x_27_cast_fp16")]; + tensor linear_53_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = attn13_in_proj_weight_cast_fp16, x = x_27_cast_fp16)[name = tensor("linear_53_cast_fp16")]; + tensor var_5687 = const()[name = tensor("op_5687"), val = tensor([1, 1, 3, 16, 64])]; + tensor qkv_27_cast_fp16 = reshape(shape = var_5687, x = linear_53_cast_fp16)[name = tensor("qkv_27_cast_fp16")]; + tensor q_79_begin_0 = const()[name = tensor("q_79_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor q_79_end_0 = const()[name = tensor("q_79_end_0"), val = tensor([1, 1, 1, 16, 64])]; + tensor q_79_end_mask_0 = const()[name = tensor("q_79_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor q_79_squeeze_mask_0 = const()[name = tensor("q_79_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor q_79_cast_fp16 = slice_by_index(begin = q_79_begin_0, end = q_79_end_0, end_mask = q_79_end_mask_0, squeeze_mask = q_79_squeeze_mask_0, x = qkv_27_cast_fp16)[name = tensor("q_79_cast_fp16")]; + tensor k_53_begin_0 = const()[name = tensor("k_53_begin_0"), val = tensor([0, 0, 1, 0, 0])]; + tensor k_53_end_0 = const()[name = tensor("k_53_end_0"), val = tensor([1, 1, 2, 16, 64])]; + tensor k_53_end_mask_0 = const()[name = tensor("k_53_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor k_53_squeeze_mask_0 = const()[name = tensor("k_53_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor k_53_cast_fp16 = slice_by_index(begin = k_53_begin_0, end = k_53_end_0, end_mask = k_53_end_mask_0, squeeze_mask = k_53_squeeze_mask_0, x = qkv_27_cast_fp16)[name = tensor("k_53_cast_fp16")]; + tensor v_27_begin_0 = const()[name = tensor("v_27_begin_0"), val = tensor([0, 0, 2, 0, 0])]; + tensor v_27_end_0 = const()[name = tensor("v_27_end_0"), val = tensor([1, 1, 3, 16, 64])]; + tensor v_27_end_mask_0 = const()[name = tensor("v_27_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor v_27_squeeze_mask_0 = const()[name = tensor("v_27_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor v_27_cast_fp16 = slice_by_index(begin = v_27_begin_0, end = v_27_end_0, end_mask = v_27_end_mask_0, squeeze_mask = v_27_squeeze_mask_0, x = qkv_27_cast_fp16)[name = tensor("v_27_cast_fp16")]; + tensor var_5791 = const()[name = tensor("op_5791"), val = tensor([1, 1, 1, 1])]; + tensor position13_to_fp16_dtype_0 = const()[name = tensor("position13_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor position13_to_fp16 = cast(dtype = position13_to_fp16_dtype_0, x = position13)[name = tensor("cast_488")]; + tensor ts_83_cast_fp16 = reshape(shape = var_5791, x = position13_to_fp16)[name = tensor("ts_83_cast_fp16")]; + tensor var_5795 = const()[name = tensor("op_5795"), val = tensor([1, 1, 16, 32, 2])]; + tensor q_complex_27_cast_fp16 = reshape(shape = var_5795, x = q_79_cast_fp16)[name = tensor("q_complex_27_cast_fp16")]; + tensor var_5799 = const()[name = tensor("op_5799"), val = tensor([1, 1, 16, 32, 2])]; + tensor k_complex_27_cast_fp16 = reshape(shape = var_5799, x = k_53_cast_fp16)[name = tensor("k_complex_27_cast_fp16")]; + tensor var_5803_begin_0 = const()[name = tensor("op_5803_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_5803_end_0 = const()[name = tensor("op_5803_end_0"), val = tensor([1, 1, 16, 32, 1])]; + tensor var_5803_end_mask_0 = const()[name = tensor("op_5803_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_5803_squeeze_mask_0 = const()[name = tensor("op_5803_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_5803_cast_fp16 = slice_by_index(begin = var_5803_begin_0, end = var_5803_end_0, end_mask = var_5803_end_mask_0, squeeze_mask = var_5803_squeeze_mask_0, x = q_complex_27_cast_fp16)[name = tensor("op_5803_cast_fp16")]; + tensor var_5811_begin_0 = const()[name = tensor("op_5811_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_5811_end_0 = const()[name = tensor("op_5811_end_0"), val = tensor([1, 1, 16, 32, 2])]; + tensor var_5811_end_mask_0 = const()[name = tensor("op_5811_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_5811_squeeze_mask_0 = const()[name = tensor("op_5811_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_5811_cast_fp16 = slice_by_index(begin = var_5811_begin_0, end = var_5811_end_0, end_mask = var_5811_end_mask_0, squeeze_mask = var_5811_squeeze_mask_0, x = q_complex_27_cast_fp16)[name = tensor("op_5811_cast_fp16")]; + tensor var_5819_begin_0 = const()[name = tensor("op_5819_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_5819_end_0 = const()[name = tensor("op_5819_end_0"), val = tensor([1, 1, 16, 32, 1])]; + tensor var_5819_end_mask_0 = const()[name = tensor("op_5819_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_5819_squeeze_mask_0 = const()[name = tensor("op_5819_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_5819_cast_fp16 = slice_by_index(begin = var_5819_begin_0, end = var_5819_end_0, end_mask = var_5819_end_mask_0, squeeze_mask = var_5819_squeeze_mask_0, x = k_complex_27_cast_fp16)[name = tensor("op_5819_cast_fp16")]; + tensor var_5827_begin_0 = const()[name = tensor("op_5827_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_5827_end_0 = const()[name = tensor("op_5827_end_0"), val = tensor([1, 1, 16, 32, 2])]; + tensor var_5827_end_mask_0 = const()[name = tensor("op_5827_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_5827_squeeze_mask_0 = const()[name = tensor("op_5827_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_5827_cast_fp16 = slice_by_index(begin = var_5827_begin_0, end = var_5827_end_0, end_mask = var_5827_end_mask_0, squeeze_mask = var_5827_squeeze_mask_0, x = k_complex_27_cast_fp16)[name = tensor("op_5827_cast_fp16")]; + tensor freqs_27_to_fp16 = const()[name = tensor("freqs_27_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(303700608)))]; + tensor var_5833_cast_fp16 = mul(x = freqs_27_to_fp16, y = ts_83_cast_fp16)[name = tensor("op_5833_cast_fp16")]; + tensor rotr_27_cast_fp16 = cos(x = var_5833_cast_fp16)[name = tensor("rotr_27_cast_fp16")]; + tensor roti_27_cast_fp16 = sin(x = var_5833_cast_fp16)[name = tensor("roti_27_cast_fp16")]; + tensor var_5837_cast_fp16 = mul(x = var_5803_cast_fp16, y = rotr_27_cast_fp16)[name = tensor("op_5837_cast_fp16")]; + tensor var_5838_cast_fp16 = mul(x = var_5811_cast_fp16, y = roti_27_cast_fp16)[name = tensor("op_5838_cast_fp16")]; + tensor qor_53_cast_fp16 = sub(x = var_5837_cast_fp16, y = var_5838_cast_fp16)[name = tensor("qor_53_cast_fp16")]; + tensor var_5841_cast_fp16 = mul(x = var_5803_cast_fp16, y = roti_27_cast_fp16)[name = tensor("op_5841_cast_fp16")]; + tensor var_5842_cast_fp16 = mul(x = var_5811_cast_fp16, y = rotr_27_cast_fp16)[name = tensor("op_5842_cast_fp16")]; + tensor qoi_53_cast_fp16 = add(x = var_5841_cast_fp16, y = var_5842_cast_fp16)[name = tensor("qoi_53_cast_fp16")]; + tensor var_5845_cast_fp16 = mul(x = var_5819_cast_fp16, y = rotr_27_cast_fp16)[name = tensor("op_5845_cast_fp16")]; + tensor var_5846_cast_fp16 = mul(x = var_5827_cast_fp16, y = roti_27_cast_fp16)[name = tensor("op_5846_cast_fp16")]; + tensor kor_53_cast_fp16 = sub(x = var_5845_cast_fp16, y = var_5846_cast_fp16)[name = tensor("kor_53_cast_fp16")]; + tensor var_5849_cast_fp16 = mul(x = var_5819_cast_fp16, y = roti_27_cast_fp16)[name = tensor("op_5849_cast_fp16")]; + tensor var_5850_cast_fp16 = mul(x = var_5827_cast_fp16, y = rotr_27_cast_fp16)[name = tensor("op_5850_cast_fp16")]; + tensor koi_53_cast_fp16 = add(x = var_5849_cast_fp16, y = var_5850_cast_fp16)[name = tensor("koi_53_cast_fp16")]; + tensor qo_27_axis_0 = const()[name = tensor("qo_27_axis_0"), val = tensor(-1)]; + tensor qo_27_cast_fp16 = stack(axis = qo_27_axis_0, values = (qor_53_cast_fp16, qoi_53_cast_fp16))[name = tensor("qo_27_cast_fp16")]; + tensor ko_27_axis_0 = const()[name = tensor("ko_27_axis_0"), val = tensor(-1)]; + tensor ko_27_cast_fp16 = stack(axis = ko_27_axis_0, values = (kor_53_cast_fp16, koi_53_cast_fp16))[name = tensor("ko_27_cast_fp16")]; + tensor var_5879 = const()[name = tensor("op_5879"), val = tensor([1, 1, 16, 64])]; + tensor q_81_cast_fp16 = reshape(shape = var_5879, x = qo_27_cast_fp16)[name = tensor("q_81_cast_fp16")]; + tensor var_5881 = const()[name = tensor("op_5881"), val = tensor([1, 1, 16, 64])]; + tensor k_55_cast_fp16 = reshape(shape = var_5881, x = ko_27_cast_fp16)[name = tensor("k_55_cast_fp16")]; + tensor _inversed_5903_y_0_to_fp16 = const()[name = tensor("_inversed_5903_y_0_to_fp16"), val = tensor(0x1p-9)]; + tensor _inversed_5903_cast_fp16 = mul(x = ts_83_cast_fp16, y = _inversed_5903_y_0_to_fp16)[name = tensor("_inversed_5903_cast_fp16")]; + tensor var_5904_cast_fp16 = floor(x = _inversed_5903_cast_fp16)[name = tensor("op_5904_cast_fp16")]; + tensor var_5905_to_fp16 = const()[name = tensor("op_5905_to_fp16"), val = tensor(0x1p+9)]; + tensor var_5906_cast_fp16 = mul(x = var_5904_cast_fp16, y = var_5905_to_fp16)[name = tensor("op_5906_cast_fp16")]; + tensor write_indices_float_55_cast_fp16 = sub(x = ts_83_cast_fp16, y = var_5906_cast_fp16)[name = tensor("write_indices_float_55_cast_fp16")]; + tensor var_5913_dtype_0 = const()[name = tensor("op_5913_dtype_0"), val = tensor("int32")]; + tensor write_indices_27_reps_0 = const()[name = tensor("write_indices_27_reps_0"), val = tensor([1, 1, 16, 64])]; + tensor write_indices_float_55_cast_fp16_to_int32 = cast(dtype = var_5913_dtype_0, x = write_indices_float_55_cast_fp16)[name = tensor("cast_487")]; + tensor write_indices_27 = tile(reps = write_indices_27_reps_0, x = write_indices_float_55_cast_fp16_to_int32)[name = tensor("write_indices_27")]; + tensor var_5921_begin_0 = const()[name = tensor("op_5921_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_5921_end_0 = const()[name = tensor("op_5921_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor var_5921_end_mask_0 = const()[name = tensor("op_5921_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_5921_squeeze_mask_0 = const()[name = tensor("op_5921_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor cache13_to_fp16_dtype_0 = const()[name = tensor("cache13_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor cache13_to_fp16 = cast(dtype = cache13_to_fp16_dtype_0, x = cache13)[name = tensor("cast_486")]; + tensor var_5921_cast_fp16 = slice_by_index(begin = var_5921_begin_0, end = var_5921_end_0, end_mask = var_5921_end_mask_0, squeeze_mask = var_5921_squeeze_mask_0, x = cache13_to_fp16)[name = tensor("op_5921_cast_fp16")]; + tensor var_5923_axis_0 = const()[name = tensor("op_5923_axis_0"), val = tensor(1)]; + tensor var_5923_mode_0 = const()[name = tensor("op_5923_mode_0"), val = tensor("update")]; + tensor var_5923_validate_indices_0 = const()[name = tensor("op_5923_validate_indices_0"), val = tensor(false)]; + tensor var_5923_cast_fp16 = scatter_along_axis(axis = var_5923_axis_0, data = var_5921_cast_fp16, indices = write_indices_27, mode = var_5923_mode_0, updates = k_55_cast_fp16, validate_indices = var_5923_validate_indices_0)[name = tensor("op_5923_cast_fp16")]; + tensor concat_93 = const()[name = tensor("concat_93"), val = tensor([0, 0, 0, 0, 0])]; + tensor concat_94 = const()[name = tensor("concat_94"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_27_internal_tensor_assign_1_stride_0 = const()[name = tensor("new_cache_27_internal_tensor_assign_1_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_27_internal_tensor_assign_1_begin_mask_0 = const()[name = tensor("new_cache_27_internal_tensor_assign_1_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_27_internal_tensor_assign_1_end_mask_0 = const()[name = tensor("new_cache_27_internal_tensor_assign_1_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_27_internal_tensor_assign_1_squeeze_mask_0 = const()[name = tensor("new_cache_27_internal_tensor_assign_1_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_74 = const()[name = tensor("shape_74"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_26 = const()[name = tensor("reduce_prod_26"), val = tensor(1048576)]; + tensor range_1d_26_start_0 = const()[name = tensor("range_1d_26_start_0"), val = tensor(0)]; + tensor range_1d_26_step_0 = const()[name = tensor("range_1d_26_step_0"), val = tensor(1)]; + tensor range_1d_26 = range_1d(end = reduce_prod_26, start = range_1d_26_start_0, step = range_1d_26_step_0)[name = tensor("range_1d_26")]; + tensor reshape_130 = reshape(shape = shape_74, x = range_1d_26)[name = tensor("reshape_130")]; + tensor slice_by_index_26 = slice_by_index(begin = concat_93, begin_mask = new_cache_27_internal_tensor_assign_1_begin_mask_0, end = concat_94, end_mask = new_cache_27_internal_tensor_assign_1_end_mask_0, squeeze_mask = new_cache_27_internal_tensor_assign_1_squeeze_mask_0, stride = new_cache_27_internal_tensor_assign_1_stride_0, x = reshape_130)[name = tensor("slice_by_index_26")]; + tensor reshape_131_shape_0 = const()[name = tensor("reshape_131_shape_0"), val = tensor([-1])]; + tensor reshape_131 = reshape(shape = reshape_131_shape_0, x = slice_by_index_26)[name = tensor("reshape_131")]; + tensor reshape_132_shape_0 = const()[name = tensor("reshape_132_shape_0"), val = tensor([-1])]; + tensor reshape_132_cast_fp16 = reshape(shape = reshape_132_shape_0, x = var_5923_cast_fp16)[name = tensor("reshape_132_cast_fp16")]; + tensor reshape_133_shape_0 = const()[name = tensor("reshape_133_shape_0"), val = tensor([-1])]; + tensor reshape_133_cast_fp16 = reshape(shape = reshape_133_shape_0, x = cache13_to_fp16)[name = tensor("reshape_133_cast_fp16")]; + tensor scatter_26_mode_0 = const()[name = tensor("scatter_26_mode_0"), val = tensor("update")]; + tensor scatter_26_axis_0 = const()[name = tensor("scatter_26_axis_0"), val = tensor(0)]; + tensor scatter_26_validate_indices_0 = const()[name = tensor("scatter_26_validate_indices_0"), val = tensor(false)]; + tensor scatter_26_cast_fp16 = scatter(axis = scatter_26_axis_0, data = reshape_133_cast_fp16, indices = reshape_131, mode = scatter_26_mode_0, updates = reshape_132_cast_fp16, validate_indices = scatter_26_validate_indices_0)[name = tensor("scatter_26_cast_fp16")]; + tensor reshape_134_cast_fp16 = reshape(shape = shape_74, x = scatter_26_cast_fp16)[name = tensor("reshape_134_cast_fp16")]; + tensor var_5931_begin_0 = const()[name = tensor("op_5931_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor var_5931_end_0 = const()[name = tensor("op_5931_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor var_5931_end_mask_0 = const()[name = tensor("op_5931_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_5931_squeeze_mask_0 = const()[name = tensor("op_5931_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor var_5931_cast_fp16 = slice_by_index(begin = var_5931_begin_0, end = var_5931_end_0, end_mask = var_5931_end_mask_0, squeeze_mask = var_5931_squeeze_mask_0, x = reshape_134_cast_fp16)[name = tensor("op_5931_cast_fp16")]; + tensor var_5933_axis_0 = const()[name = tensor("op_5933_axis_0"), val = tensor(1)]; + tensor var_5933_mode_0 = const()[name = tensor("op_5933_mode_0"), val = tensor("update")]; + tensor var_5933_validate_indices_0 = const()[name = tensor("op_5933_validate_indices_0"), val = tensor(false)]; + tensor var_5933_cast_fp16 = scatter_along_axis(axis = var_5933_axis_0, data = var_5931_cast_fp16, indices = write_indices_27, mode = var_5933_mode_0, updates = v_27_cast_fp16, validate_indices = var_5933_validate_indices_0)[name = tensor("op_5933_cast_fp16")]; + tensor concat_95 = const()[name = tensor("concat_95"), val = tensor([1, 0, 0, 0, 0])]; + tensor concat_96 = const()[name = tensor("concat_96"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_27_internal_tensor_assign_2_stride_0 = const()[name = tensor("new_cache_27_internal_tensor_assign_2_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_27_internal_tensor_assign_2_begin_mask_0 = const()[name = tensor("new_cache_27_internal_tensor_assign_2_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_27_internal_tensor_assign_2_end_mask_0 = const()[name = tensor("new_cache_27_internal_tensor_assign_2_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_27_internal_tensor_assign_2_squeeze_mask_0 = const()[name = tensor("new_cache_27_internal_tensor_assign_2_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_75 = const()[name = tensor("shape_75"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_27 = const()[name = tensor("reduce_prod_27"), val = tensor(1048576)]; + tensor range_1d_27_start_0 = const()[name = tensor("range_1d_27_start_0"), val = tensor(0)]; + tensor range_1d_27_step_0 = const()[name = tensor("range_1d_27_step_0"), val = tensor(1)]; + tensor range_1d_27 = range_1d(end = reduce_prod_27, start = range_1d_27_start_0, step = range_1d_27_step_0)[name = tensor("range_1d_27")]; + tensor reshape_135 = reshape(shape = shape_75, x = range_1d_27)[name = tensor("reshape_135")]; + tensor slice_by_index_27 = slice_by_index(begin = concat_95, begin_mask = new_cache_27_internal_tensor_assign_2_begin_mask_0, end = concat_96, end_mask = new_cache_27_internal_tensor_assign_2_end_mask_0, squeeze_mask = new_cache_27_internal_tensor_assign_2_squeeze_mask_0, stride = new_cache_27_internal_tensor_assign_2_stride_0, x = reshape_135)[name = tensor("slice_by_index_27")]; + tensor reshape_136_shape_0 = const()[name = tensor("reshape_136_shape_0"), val = tensor([-1])]; + tensor reshape_136 = reshape(shape = reshape_136_shape_0, x = slice_by_index_27)[name = tensor("reshape_136")]; + tensor reshape_137_shape_0 = const()[name = tensor("reshape_137_shape_0"), val = tensor([-1])]; + tensor reshape_137_cast_fp16 = reshape(shape = reshape_137_shape_0, x = var_5933_cast_fp16)[name = tensor("reshape_137_cast_fp16")]; + tensor reshape_138_shape_0 = const()[name = tensor("reshape_138_shape_0"), val = tensor([-1])]; + tensor reshape_138_cast_fp16 = reshape(shape = reshape_138_shape_0, x = reshape_134_cast_fp16)[name = tensor("reshape_138_cast_fp16")]; + tensor scatter_27_mode_0 = const()[name = tensor("scatter_27_mode_0"), val = tensor("update")]; + tensor scatter_27_axis_0 = const()[name = tensor("scatter_27_axis_0"), val = tensor(0)]; + tensor scatter_27_validate_indices_0 = const()[name = tensor("scatter_27_validate_indices_0"), val = tensor(false)]; + tensor scatter_27_cast_fp16 = scatter(axis = scatter_27_axis_0, data = reshape_138_cast_fp16, indices = reshape_136, mode = scatter_27_mode_0, updates = reshape_137_cast_fp16, validate_indices = scatter_27_validate_indices_0)[name = tensor("scatter_27_cast_fp16")]; + tensor reshape_139_cast_fp16 = reshape(shape = shape_75, x = scatter_27_cast_fp16)[name = tensor("reshape_139_cast_fp16")]; + tensor reshape_139_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("reshape_139_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor keys_79_begin_0 = const()[name = tensor("keys_79_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor keys_79_end_0 = const()[name = tensor("keys_79_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor keys_79_end_mask_0 = const()[name = tensor("keys_79_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor keys_79_squeeze_mask_0 = const()[name = tensor("keys_79_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor keys_79_cast_fp16 = slice_by_index(begin = keys_79_begin_0, end = keys_79_end_0, end_mask = keys_79_end_mask_0, squeeze_mask = keys_79_squeeze_mask_0, x = reshape_139_cast_fp16)[name = tensor("keys_79_cast_fp16")]; + tensor values_79_begin_0 = const()[name = tensor("values_79_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor values_79_end_0 = const()[name = tensor("values_79_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor values_79_end_mask_0 = const()[name = tensor("values_79_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor values_79_squeeze_mask_0 = const()[name = tensor("values_79_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor values_79_cast_fp16 = slice_by_index(begin = values_79_begin_0, end = values_79_end_0, end_mask = values_79_end_mask_0, squeeze_mask = values_79_squeeze_mask_0, x = reshape_139_cast_fp16)[name = tensor("values_79_cast_fp16")]; + tensor var_5945_cast_fp16 = not_equal(x = keys_79_cast_fp16, y = keys_79_cast_fp16)[name = tensor("op_5945_cast_fp16")]; + tensor keys_81_cast_fp16 = select(a = var_504_to_fp16, b = keys_79_cast_fp16, cond = var_5945_cast_fp16)[name = tensor("keys_81_cast_fp16")]; + tensor var_5953_cast_fp16 = not_equal(x = values_79_cast_fp16, y = values_79_cast_fp16)[name = tensor("op_5953_cast_fp16")]; + tensor values_81_cast_fp16 = select(a = var_504_to_fp16, b = values_79_cast_fp16, cond = var_5953_cast_fp16)[name = tensor("values_81_cast_fp16")]; + tensor var_5977 = const()[name = tensor("op_5977"), val = tensor([0, 2, 1, 3])]; + tensor var_5990 = const()[name = tensor("op_5990"), val = tensor([1, 1, 1])]; + tensor var_5991_cast_fp16 = reshape(shape = var_5990, x = position13_to_fp16)[name = tensor("op_5991_cast_fp16")]; + tensor var_6008_to_fp16 = const()[name = tensor("op_6008_to_fp16"), val = tensor(0x1p+0)]; + tensor valid_len_27_cast_fp16 = add(x = var_5991_cast_fp16, y = var_6008_to_fp16)[name = tensor("valid_len_27_cast_fp16")]; + tensor valid_mask_27_cast_fp16 = less(x = k_positions_1_promoted_to_fp16, y = valid_len_27_cast_fp16)[name = tensor("valid_mask_27_cast_fp16")]; + tensor causal_mask_27_cast_fp16 = less_equal(x = k_positions_1_promoted_to_fp16, y = var_5991_cast_fp16)[name = tensor("causal_mask_27_cast_fp16")]; + tensor attn_mask_53 = logical_and(x = valid_mask_27_cast_fp16, y = causal_mask_27_cast_fp16)[name = tensor("attn_mask_53")]; + tensor attn_mask_55_axes_0 = const()[name = tensor("attn_mask_55_axes_0"), val = tensor([1])]; + tensor attn_mask_55 = expand_dims(axes = attn_mask_55_axes_0, x = attn_mask_53)[name = tensor("attn_mask_55")]; + tensor var_6026_transpose_x_0 = const()[name = tensor("op_6026_transpose_x_0"), val = tensor(false)]; + tensor var_6026_transpose_y_0 = const()[name = tensor("op_6026_transpose_y_0"), val = tensor(false)]; + tensor transpose_98_perm_0 = const()[name = tensor("transpose_98_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_99_perm_0 = const()[name = tensor("transpose_99_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_99 = transpose(perm = transpose_99_perm_0, x = keys_81_cast_fp16)[name = tensor("transpose_161")]; + tensor transpose_98 = transpose(perm = transpose_98_perm_0, x = q_81_cast_fp16)[name = tensor("transpose_162")]; + tensor var_6026_cast_fp16 = matmul(transpose_x = var_6026_transpose_x_0, transpose_y = var_6026_transpose_y_0, x = transpose_98, y = transpose_99)[name = tensor("op_6026_cast_fp16")]; + tensor var_6020_to_fp16 = const()[name = tensor("op_6020_to_fp16"), val = tensor([0x1p-3])]; + tensor attn_weights_79_cast_fp16 = mul(x = var_6026_cast_fp16, y = var_6020_to_fp16)[name = tensor("attn_weights_79_cast_fp16")]; + tensor var_6028 = logical_not(x = attn_mask_55)[name = tensor("op_6028")]; + tensor var_6029_to_fp16 = const()[name = tensor("op_6029_to_fp16"), val = tensor(-inf)]; + tensor attn_weights_81_cast_fp16 = select(a = var_6029_to_fp16, b = attn_weights_79_cast_fp16, cond = var_6028)[name = tensor("attn_weights_81_cast_fp16")]; + tensor var_6031 = const()[name = tensor("op_6031"), val = tensor(-1)]; + tensor attn_weights_83_cast_fp16 = softmax(axis = var_6031, x = attn_weights_81_cast_fp16)[name = tensor("attn_weights_83_cast_fp16")]; + tensor attn_output_27_transpose_x_0 = const()[name = tensor("attn_output_27_transpose_x_0"), val = tensor(false)]; + tensor attn_output_27_transpose_y_0 = const()[name = tensor("attn_output_27_transpose_y_0"), val = tensor(false)]; + tensor values_83_cast_fp16 = transpose(perm = var_5977, x = values_81_cast_fp16)[name = tensor("transpose_163")]; + tensor attn_output_27_cast_fp16 = matmul(transpose_x = attn_output_27_transpose_x_0, transpose_y = attn_output_27_transpose_y_0, x = attn_weights_83_cast_fp16, y = values_83_cast_fp16)[name = tensor("attn_output_27_cast_fp16")]; + tensor var_6039 = const()[name = tensor("op_6039"), val = tensor([0, 2, 1, 3])]; + tensor var_6042 = const()[name = tensor("op_6042"), val = tensor([1, 1, 1024])]; + tensor var_6040_cast_fp16 = transpose(perm = var_6039, x = attn_output_27_cast_fp16)[name = tensor("transpose_160")]; + tensor input_135_cast_fp16 = reshape(shape = var_6042, x = var_6040_cast_fp16)[name = tensor("input_135_cast_fp16")]; + tensor linear_54_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = attn13_out_proj_weight_cast_fp16, x = input_135_cast_fp16)[name = tensor("linear_54_cast_fp16")]; + tensor var_6048_to_fp16 = const()[name = tensor("op_6048_to_fp16"), val = tensor(0x1p+0)]; + tensor var_6049_cast_fp16 = add(x = position13_to_fp16, y = var_6048_to_fp16)[name = tensor("op_6049_cast_fp16")]; + tensor var_6049_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_6049_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor input_137_cast_fp16 = add(x = input_133_cast_fp16, y = linear_54_cast_fp16)[name = tensor("input_137_cast_fp16")]; + tensor input_139_axes_0 = const()[name = tensor("input_139_axes_0"), val = tensor([-1])]; + tensor norm13_2_weight_to_fp16 = const()[name = tensor("norm13_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(303700736)))]; + tensor norm13_2_bias_to_fp16 = const()[name = tensor("norm13_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(303702848)))]; + tensor var_6053_to_fp16 = const()[name = tensor("op_6053_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_139_cast_fp16 = layer_norm(axes = input_139_axes_0, beta = norm13_2_bias_to_fp16, epsilon = var_6053_to_fp16, gamma = norm13_2_weight_to_fp16, x = input_137_cast_fp16)[name = tensor("input_139_cast_fp16")]; + tensor linear_55_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = linear13_1_weight_cast_fp16, x = input_139_cast_fp16)[name = tensor("linear_55_cast_fp16")]; + tensor input_141_mode_0 = const()[name = tensor("input_141_mode_0"), val = tensor("EXACT")]; + tensor input_141_cast_fp16 = gelu(mode = input_141_mode_0, x = linear_55_cast_fp16)[name = tensor("input_141_cast_fp16")]; + tensor linear_56_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = linear13_2_weight_cast_fp16, x = input_141_cast_fp16)[name = tensor("linear_56_cast_fp16")]; + tensor input_143_cast_fp16 = add(x = input_137_cast_fp16, y = linear_56_cast_fp16)[name = tensor("input_143_cast_fp16")]; + tensor x_29_axes_0 = const()[name = tensor("x_29_axes_0"), val = tensor([-1])]; + tensor norm14_1_weight_to_fp16 = const()[name = tensor("norm14_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(303704960)))]; + tensor norm14_1_bias_to_fp16 = const()[name = tensor("norm14_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(303707072)))]; + tensor var_6070_to_fp16 = const()[name = tensor("op_6070_to_fp16"), val = tensor(0x1.5p-17)]; + tensor x_29_cast_fp16 = layer_norm(axes = x_29_axes_0, beta = norm14_1_bias_to_fp16, epsilon = var_6070_to_fp16, gamma = norm14_1_weight_to_fp16, x = input_143_cast_fp16)[name = tensor("x_29_cast_fp16")]; + tensor linear_57_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = attn14_in_proj_weight_cast_fp16, x = x_29_cast_fp16)[name = tensor("linear_57_cast_fp16")]; + tensor var_6106 = const()[name = tensor("op_6106"), val = tensor([1, 1, 3, 16, 64])]; + tensor qkv_29_cast_fp16 = reshape(shape = var_6106, x = linear_57_cast_fp16)[name = tensor("qkv_29_cast_fp16")]; + tensor q_85_begin_0 = const()[name = tensor("q_85_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor q_85_end_0 = const()[name = tensor("q_85_end_0"), val = tensor([1, 1, 1, 16, 64])]; + tensor q_85_end_mask_0 = const()[name = tensor("q_85_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor q_85_squeeze_mask_0 = const()[name = tensor("q_85_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor q_85_cast_fp16 = slice_by_index(begin = q_85_begin_0, end = q_85_end_0, end_mask = q_85_end_mask_0, squeeze_mask = q_85_squeeze_mask_0, x = qkv_29_cast_fp16)[name = tensor("q_85_cast_fp16")]; + tensor k_57_begin_0 = const()[name = tensor("k_57_begin_0"), val = tensor([0, 0, 1, 0, 0])]; + tensor k_57_end_0 = const()[name = tensor("k_57_end_0"), val = tensor([1, 1, 2, 16, 64])]; + tensor k_57_end_mask_0 = const()[name = tensor("k_57_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor k_57_squeeze_mask_0 = const()[name = tensor("k_57_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor k_57_cast_fp16 = slice_by_index(begin = k_57_begin_0, end = k_57_end_0, end_mask = k_57_end_mask_0, squeeze_mask = k_57_squeeze_mask_0, x = qkv_29_cast_fp16)[name = tensor("k_57_cast_fp16")]; + tensor v_29_begin_0 = const()[name = tensor("v_29_begin_0"), val = tensor([0, 0, 2, 0, 0])]; + tensor v_29_end_0 = const()[name = tensor("v_29_end_0"), val = tensor([1, 1, 3, 16, 64])]; + tensor v_29_end_mask_0 = const()[name = tensor("v_29_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor v_29_squeeze_mask_0 = const()[name = tensor("v_29_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor v_29_cast_fp16 = slice_by_index(begin = v_29_begin_0, end = v_29_end_0, end_mask = v_29_end_mask_0, squeeze_mask = v_29_squeeze_mask_0, x = qkv_29_cast_fp16)[name = tensor("v_29_cast_fp16")]; + tensor var_6210 = const()[name = tensor("op_6210"), val = tensor([1, 1, 1, 1])]; + tensor position14_to_fp16_dtype_0 = const()[name = tensor("position14_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor position14_to_fp16 = cast(dtype = position14_to_fp16_dtype_0, x = position14)[name = tensor("cast_483")]; + tensor ts_89_cast_fp16 = reshape(shape = var_6210, x = position14_to_fp16)[name = tensor("ts_89_cast_fp16")]; + tensor var_6214 = const()[name = tensor("op_6214"), val = tensor([1, 1, 16, 32, 2])]; + tensor q_complex_29_cast_fp16 = reshape(shape = var_6214, x = q_85_cast_fp16)[name = tensor("q_complex_29_cast_fp16")]; + tensor var_6218 = const()[name = tensor("op_6218"), val = tensor([1, 1, 16, 32, 2])]; + tensor k_complex_29_cast_fp16 = reshape(shape = var_6218, x = k_57_cast_fp16)[name = tensor("k_complex_29_cast_fp16")]; + tensor var_6222_begin_0 = const()[name = tensor("op_6222_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_6222_end_0 = const()[name = tensor("op_6222_end_0"), val = tensor([1, 1, 16, 32, 1])]; + tensor var_6222_end_mask_0 = const()[name = tensor("op_6222_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_6222_squeeze_mask_0 = const()[name = tensor("op_6222_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_6222_cast_fp16 = slice_by_index(begin = var_6222_begin_0, end = var_6222_end_0, end_mask = var_6222_end_mask_0, squeeze_mask = var_6222_squeeze_mask_0, x = q_complex_29_cast_fp16)[name = tensor("op_6222_cast_fp16")]; + tensor var_6230_begin_0 = const()[name = tensor("op_6230_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_6230_end_0 = const()[name = tensor("op_6230_end_0"), val = tensor([1, 1, 16, 32, 2])]; + tensor var_6230_end_mask_0 = const()[name = tensor("op_6230_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_6230_squeeze_mask_0 = const()[name = tensor("op_6230_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_6230_cast_fp16 = slice_by_index(begin = var_6230_begin_0, end = var_6230_end_0, end_mask = var_6230_end_mask_0, squeeze_mask = var_6230_squeeze_mask_0, x = q_complex_29_cast_fp16)[name = tensor("op_6230_cast_fp16")]; + tensor var_6238_begin_0 = const()[name = tensor("op_6238_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_6238_end_0 = const()[name = tensor("op_6238_end_0"), val = tensor([1, 1, 16, 32, 1])]; + tensor var_6238_end_mask_0 = const()[name = tensor("op_6238_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_6238_squeeze_mask_0 = const()[name = tensor("op_6238_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_6238_cast_fp16 = slice_by_index(begin = var_6238_begin_0, end = var_6238_end_0, end_mask = var_6238_end_mask_0, squeeze_mask = var_6238_squeeze_mask_0, x = k_complex_29_cast_fp16)[name = tensor("op_6238_cast_fp16")]; + tensor var_6246_begin_0 = const()[name = tensor("op_6246_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_6246_end_0 = const()[name = tensor("op_6246_end_0"), val = tensor([1, 1, 16, 32, 2])]; + tensor var_6246_end_mask_0 = const()[name = tensor("op_6246_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_6246_squeeze_mask_0 = const()[name = tensor("op_6246_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_6246_cast_fp16 = slice_by_index(begin = var_6246_begin_0, end = var_6246_end_0, end_mask = var_6246_end_mask_0, squeeze_mask = var_6246_squeeze_mask_0, x = k_complex_29_cast_fp16)[name = tensor("op_6246_cast_fp16")]; + tensor freqs_29_to_fp16 = const()[name = tensor("freqs_29_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(303709184)))]; + tensor var_6252_cast_fp16 = mul(x = freqs_29_to_fp16, y = ts_89_cast_fp16)[name = tensor("op_6252_cast_fp16")]; + tensor rotr_29_cast_fp16 = cos(x = var_6252_cast_fp16)[name = tensor("rotr_29_cast_fp16")]; + tensor roti_29_cast_fp16 = sin(x = var_6252_cast_fp16)[name = tensor("roti_29_cast_fp16")]; + tensor var_6256_cast_fp16 = mul(x = var_6222_cast_fp16, y = rotr_29_cast_fp16)[name = tensor("op_6256_cast_fp16")]; + tensor var_6257_cast_fp16 = mul(x = var_6230_cast_fp16, y = roti_29_cast_fp16)[name = tensor("op_6257_cast_fp16")]; + tensor qor_57_cast_fp16 = sub(x = var_6256_cast_fp16, y = var_6257_cast_fp16)[name = tensor("qor_57_cast_fp16")]; + tensor var_6260_cast_fp16 = mul(x = var_6222_cast_fp16, y = roti_29_cast_fp16)[name = tensor("op_6260_cast_fp16")]; + tensor var_6261_cast_fp16 = mul(x = var_6230_cast_fp16, y = rotr_29_cast_fp16)[name = tensor("op_6261_cast_fp16")]; + tensor qoi_57_cast_fp16 = add(x = var_6260_cast_fp16, y = var_6261_cast_fp16)[name = tensor("qoi_57_cast_fp16")]; + tensor var_6264_cast_fp16 = mul(x = var_6238_cast_fp16, y = rotr_29_cast_fp16)[name = tensor("op_6264_cast_fp16")]; + tensor var_6265_cast_fp16 = mul(x = var_6246_cast_fp16, y = roti_29_cast_fp16)[name = tensor("op_6265_cast_fp16")]; + tensor kor_57_cast_fp16 = sub(x = var_6264_cast_fp16, y = var_6265_cast_fp16)[name = tensor("kor_57_cast_fp16")]; + tensor var_6268_cast_fp16 = mul(x = var_6238_cast_fp16, y = roti_29_cast_fp16)[name = tensor("op_6268_cast_fp16")]; + tensor var_6269_cast_fp16 = mul(x = var_6246_cast_fp16, y = rotr_29_cast_fp16)[name = tensor("op_6269_cast_fp16")]; + tensor koi_57_cast_fp16 = add(x = var_6268_cast_fp16, y = var_6269_cast_fp16)[name = tensor("koi_57_cast_fp16")]; + tensor qo_29_axis_0 = const()[name = tensor("qo_29_axis_0"), val = tensor(-1)]; + tensor qo_29_cast_fp16 = stack(axis = qo_29_axis_0, values = (qor_57_cast_fp16, qoi_57_cast_fp16))[name = tensor("qo_29_cast_fp16")]; + tensor ko_29_axis_0 = const()[name = tensor("ko_29_axis_0"), val = tensor(-1)]; + tensor ko_29_cast_fp16 = stack(axis = ko_29_axis_0, values = (kor_57_cast_fp16, koi_57_cast_fp16))[name = tensor("ko_29_cast_fp16")]; + tensor var_6298 = const()[name = tensor("op_6298"), val = tensor([1, 1, 16, 64])]; + tensor q_87_cast_fp16 = reshape(shape = var_6298, x = qo_29_cast_fp16)[name = tensor("q_87_cast_fp16")]; + tensor var_6300 = const()[name = tensor("op_6300"), val = tensor([1, 1, 16, 64])]; + tensor k_59_cast_fp16 = reshape(shape = var_6300, x = ko_29_cast_fp16)[name = tensor("k_59_cast_fp16")]; + tensor _inversed_6322_y_0_to_fp16 = const()[name = tensor("_inversed_6322_y_0_to_fp16"), val = tensor(0x1p-9)]; + tensor _inversed_6322_cast_fp16 = mul(x = ts_89_cast_fp16, y = _inversed_6322_y_0_to_fp16)[name = tensor("_inversed_6322_cast_fp16")]; + tensor var_6323_cast_fp16 = floor(x = _inversed_6322_cast_fp16)[name = tensor("op_6323_cast_fp16")]; + tensor var_6324_to_fp16 = const()[name = tensor("op_6324_to_fp16"), val = tensor(0x1p+9)]; + tensor var_6325_cast_fp16 = mul(x = var_6323_cast_fp16, y = var_6324_to_fp16)[name = tensor("op_6325_cast_fp16")]; + tensor write_indices_float_59_cast_fp16 = sub(x = ts_89_cast_fp16, y = var_6325_cast_fp16)[name = tensor("write_indices_float_59_cast_fp16")]; + tensor var_6332_dtype_0 = const()[name = tensor("op_6332_dtype_0"), val = tensor("int32")]; + tensor write_indices_29_reps_0 = const()[name = tensor("write_indices_29_reps_0"), val = tensor([1, 1, 16, 64])]; + tensor write_indices_float_59_cast_fp16_to_int32 = cast(dtype = var_6332_dtype_0, x = write_indices_float_59_cast_fp16)[name = tensor("cast_482")]; + tensor write_indices_29 = tile(reps = write_indices_29_reps_0, x = write_indices_float_59_cast_fp16_to_int32)[name = tensor("write_indices_29")]; + tensor var_6340_begin_0 = const()[name = tensor("op_6340_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_6340_end_0 = const()[name = tensor("op_6340_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor var_6340_end_mask_0 = const()[name = tensor("op_6340_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_6340_squeeze_mask_0 = const()[name = tensor("op_6340_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor cache14_to_fp16_dtype_0 = const()[name = tensor("cache14_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor cache14_to_fp16 = cast(dtype = cache14_to_fp16_dtype_0, x = cache14)[name = tensor("cast_481")]; + tensor var_6340_cast_fp16 = slice_by_index(begin = var_6340_begin_0, end = var_6340_end_0, end_mask = var_6340_end_mask_0, squeeze_mask = var_6340_squeeze_mask_0, x = cache14_to_fp16)[name = tensor("op_6340_cast_fp16")]; + tensor var_6342_axis_0 = const()[name = tensor("op_6342_axis_0"), val = tensor(1)]; + tensor var_6342_mode_0 = const()[name = tensor("op_6342_mode_0"), val = tensor("update")]; + tensor var_6342_validate_indices_0 = const()[name = tensor("op_6342_validate_indices_0"), val = tensor(false)]; + tensor var_6342_cast_fp16 = scatter_along_axis(axis = var_6342_axis_0, data = var_6340_cast_fp16, indices = write_indices_29, mode = var_6342_mode_0, updates = k_59_cast_fp16, validate_indices = var_6342_validate_indices_0)[name = tensor("op_6342_cast_fp16")]; + tensor concat_100 = const()[name = tensor("concat_100"), val = tensor([0, 0, 0, 0, 0])]; + tensor concat_101 = const()[name = tensor("concat_101"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_29_internal_tensor_assign_1_stride_0 = const()[name = tensor("new_cache_29_internal_tensor_assign_1_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_29_internal_tensor_assign_1_begin_mask_0 = const()[name = tensor("new_cache_29_internal_tensor_assign_1_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_29_internal_tensor_assign_1_end_mask_0 = const()[name = tensor("new_cache_29_internal_tensor_assign_1_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_29_internal_tensor_assign_1_squeeze_mask_0 = const()[name = tensor("new_cache_29_internal_tensor_assign_1_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_76 = const()[name = tensor("shape_76"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_28 = const()[name = tensor("reduce_prod_28"), val = tensor(1048576)]; + tensor range_1d_28_start_0 = const()[name = tensor("range_1d_28_start_0"), val = tensor(0)]; + tensor range_1d_28_step_0 = const()[name = tensor("range_1d_28_step_0"), val = tensor(1)]; + tensor range_1d_28 = range_1d(end = reduce_prod_28, start = range_1d_28_start_0, step = range_1d_28_step_0)[name = tensor("range_1d_28")]; + tensor reshape_140 = reshape(shape = shape_76, x = range_1d_28)[name = tensor("reshape_140")]; + tensor slice_by_index_28 = slice_by_index(begin = concat_100, begin_mask = new_cache_29_internal_tensor_assign_1_begin_mask_0, end = concat_101, end_mask = new_cache_29_internal_tensor_assign_1_end_mask_0, squeeze_mask = new_cache_29_internal_tensor_assign_1_squeeze_mask_0, stride = new_cache_29_internal_tensor_assign_1_stride_0, x = reshape_140)[name = tensor("slice_by_index_28")]; + tensor reshape_141_shape_0 = const()[name = tensor("reshape_141_shape_0"), val = tensor([-1])]; + tensor reshape_141 = reshape(shape = reshape_141_shape_0, x = slice_by_index_28)[name = tensor("reshape_141")]; + tensor reshape_142_shape_0 = const()[name = tensor("reshape_142_shape_0"), val = tensor([-1])]; + tensor reshape_142_cast_fp16 = reshape(shape = reshape_142_shape_0, x = var_6342_cast_fp16)[name = tensor("reshape_142_cast_fp16")]; + tensor reshape_143_shape_0 = const()[name = tensor("reshape_143_shape_0"), val = tensor([-1])]; + tensor reshape_143_cast_fp16 = reshape(shape = reshape_143_shape_0, x = cache14_to_fp16)[name = tensor("reshape_143_cast_fp16")]; + tensor scatter_28_mode_0 = const()[name = tensor("scatter_28_mode_0"), val = tensor("update")]; + tensor scatter_28_axis_0 = const()[name = tensor("scatter_28_axis_0"), val = tensor(0)]; + tensor scatter_28_validate_indices_0 = const()[name = tensor("scatter_28_validate_indices_0"), val = tensor(false)]; + tensor scatter_28_cast_fp16 = scatter(axis = scatter_28_axis_0, data = reshape_143_cast_fp16, indices = reshape_141, mode = scatter_28_mode_0, updates = reshape_142_cast_fp16, validate_indices = scatter_28_validate_indices_0)[name = tensor("scatter_28_cast_fp16")]; + tensor reshape_144_cast_fp16 = reshape(shape = shape_76, x = scatter_28_cast_fp16)[name = tensor("reshape_144_cast_fp16")]; + tensor var_6350_begin_0 = const()[name = tensor("op_6350_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor var_6350_end_0 = const()[name = tensor("op_6350_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor var_6350_end_mask_0 = const()[name = tensor("op_6350_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_6350_squeeze_mask_0 = const()[name = tensor("op_6350_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor var_6350_cast_fp16 = slice_by_index(begin = var_6350_begin_0, end = var_6350_end_0, end_mask = var_6350_end_mask_0, squeeze_mask = var_6350_squeeze_mask_0, x = reshape_144_cast_fp16)[name = tensor("op_6350_cast_fp16")]; + tensor var_6352_axis_0 = const()[name = tensor("op_6352_axis_0"), val = tensor(1)]; + tensor var_6352_mode_0 = const()[name = tensor("op_6352_mode_0"), val = tensor("update")]; + tensor var_6352_validate_indices_0 = const()[name = tensor("op_6352_validate_indices_0"), val = tensor(false)]; + tensor var_6352_cast_fp16 = scatter_along_axis(axis = var_6352_axis_0, data = var_6350_cast_fp16, indices = write_indices_29, mode = var_6352_mode_0, updates = v_29_cast_fp16, validate_indices = var_6352_validate_indices_0)[name = tensor("op_6352_cast_fp16")]; + tensor concat_102 = const()[name = tensor("concat_102"), val = tensor([1, 0, 0, 0, 0])]; + tensor concat_103 = const()[name = tensor("concat_103"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_29_internal_tensor_assign_2_stride_0 = const()[name = tensor("new_cache_29_internal_tensor_assign_2_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_29_internal_tensor_assign_2_begin_mask_0 = const()[name = tensor("new_cache_29_internal_tensor_assign_2_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_29_internal_tensor_assign_2_end_mask_0 = const()[name = tensor("new_cache_29_internal_tensor_assign_2_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_29_internal_tensor_assign_2_squeeze_mask_0 = const()[name = tensor("new_cache_29_internal_tensor_assign_2_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_77 = const()[name = tensor("shape_77"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_29 = const()[name = tensor("reduce_prod_29"), val = tensor(1048576)]; + tensor range_1d_29_start_0 = const()[name = tensor("range_1d_29_start_0"), val = tensor(0)]; + tensor range_1d_29_step_0 = const()[name = tensor("range_1d_29_step_0"), val = tensor(1)]; + tensor range_1d_29 = range_1d(end = reduce_prod_29, start = range_1d_29_start_0, step = range_1d_29_step_0)[name = tensor("range_1d_29")]; + tensor reshape_145 = reshape(shape = shape_77, x = range_1d_29)[name = tensor("reshape_145")]; + tensor slice_by_index_29 = slice_by_index(begin = concat_102, begin_mask = new_cache_29_internal_tensor_assign_2_begin_mask_0, end = concat_103, end_mask = new_cache_29_internal_tensor_assign_2_end_mask_0, squeeze_mask = new_cache_29_internal_tensor_assign_2_squeeze_mask_0, stride = new_cache_29_internal_tensor_assign_2_stride_0, x = reshape_145)[name = tensor("slice_by_index_29")]; + tensor reshape_146_shape_0 = const()[name = tensor("reshape_146_shape_0"), val = tensor([-1])]; + tensor reshape_146 = reshape(shape = reshape_146_shape_0, x = slice_by_index_29)[name = tensor("reshape_146")]; + tensor reshape_147_shape_0 = const()[name = tensor("reshape_147_shape_0"), val = tensor([-1])]; + tensor reshape_147_cast_fp16 = reshape(shape = reshape_147_shape_0, x = var_6352_cast_fp16)[name = tensor("reshape_147_cast_fp16")]; + tensor reshape_148_shape_0 = const()[name = tensor("reshape_148_shape_0"), val = tensor([-1])]; + tensor reshape_148_cast_fp16 = reshape(shape = reshape_148_shape_0, x = reshape_144_cast_fp16)[name = tensor("reshape_148_cast_fp16")]; + tensor scatter_29_mode_0 = const()[name = tensor("scatter_29_mode_0"), val = tensor("update")]; + tensor scatter_29_axis_0 = const()[name = tensor("scatter_29_axis_0"), val = tensor(0)]; + tensor scatter_29_validate_indices_0 = const()[name = tensor("scatter_29_validate_indices_0"), val = tensor(false)]; + tensor scatter_29_cast_fp16 = scatter(axis = scatter_29_axis_0, data = reshape_148_cast_fp16, indices = reshape_146, mode = scatter_29_mode_0, updates = reshape_147_cast_fp16, validate_indices = scatter_29_validate_indices_0)[name = tensor("scatter_29_cast_fp16")]; + tensor reshape_149_cast_fp16 = reshape(shape = shape_77, x = scatter_29_cast_fp16)[name = tensor("reshape_149_cast_fp16")]; + tensor reshape_149_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("reshape_149_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor keys_85_begin_0 = const()[name = tensor("keys_85_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor keys_85_end_0 = const()[name = tensor("keys_85_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor keys_85_end_mask_0 = const()[name = tensor("keys_85_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor keys_85_squeeze_mask_0 = const()[name = tensor("keys_85_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor keys_85_cast_fp16 = slice_by_index(begin = keys_85_begin_0, end = keys_85_end_0, end_mask = keys_85_end_mask_0, squeeze_mask = keys_85_squeeze_mask_0, x = reshape_149_cast_fp16)[name = tensor("keys_85_cast_fp16")]; + tensor values_85_begin_0 = const()[name = tensor("values_85_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor values_85_end_0 = const()[name = tensor("values_85_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor values_85_end_mask_0 = const()[name = tensor("values_85_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor values_85_squeeze_mask_0 = const()[name = tensor("values_85_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor values_85_cast_fp16 = slice_by_index(begin = values_85_begin_0, end = values_85_end_0, end_mask = values_85_end_mask_0, squeeze_mask = values_85_squeeze_mask_0, x = reshape_149_cast_fp16)[name = tensor("values_85_cast_fp16")]; + tensor var_6364_cast_fp16 = not_equal(x = keys_85_cast_fp16, y = keys_85_cast_fp16)[name = tensor("op_6364_cast_fp16")]; + tensor keys_87_cast_fp16 = select(a = var_504_to_fp16, b = keys_85_cast_fp16, cond = var_6364_cast_fp16)[name = tensor("keys_87_cast_fp16")]; + tensor var_6372_cast_fp16 = not_equal(x = values_85_cast_fp16, y = values_85_cast_fp16)[name = tensor("op_6372_cast_fp16")]; + tensor values_87_cast_fp16 = select(a = var_504_to_fp16, b = values_85_cast_fp16, cond = var_6372_cast_fp16)[name = tensor("values_87_cast_fp16")]; + tensor var_6396 = const()[name = tensor("op_6396"), val = tensor([0, 2, 1, 3])]; + tensor var_6409 = const()[name = tensor("op_6409"), val = tensor([1, 1, 1])]; + tensor var_6410_cast_fp16 = reshape(shape = var_6409, x = position14_to_fp16)[name = tensor("op_6410_cast_fp16")]; + tensor var_6427_to_fp16 = const()[name = tensor("op_6427_to_fp16"), val = tensor(0x1p+0)]; + tensor valid_len_29_cast_fp16 = add(x = var_6410_cast_fp16, y = var_6427_to_fp16)[name = tensor("valid_len_29_cast_fp16")]; + tensor valid_mask_29_cast_fp16 = less(x = k_positions_1_promoted_to_fp16, y = valid_len_29_cast_fp16)[name = tensor("valid_mask_29_cast_fp16")]; + tensor causal_mask_29_cast_fp16 = less_equal(x = k_positions_1_promoted_to_fp16, y = var_6410_cast_fp16)[name = tensor("causal_mask_29_cast_fp16")]; + tensor attn_mask_57 = logical_and(x = valid_mask_29_cast_fp16, y = causal_mask_29_cast_fp16)[name = tensor("attn_mask_57")]; + tensor attn_mask_59_axes_0 = const()[name = tensor("attn_mask_59_axes_0"), val = tensor([1])]; + tensor attn_mask_59 = expand_dims(axes = attn_mask_59_axes_0, x = attn_mask_57)[name = tensor("attn_mask_59")]; + tensor var_6445_transpose_x_0 = const()[name = tensor("op_6445_transpose_x_0"), val = tensor(false)]; + tensor var_6445_transpose_y_0 = const()[name = tensor("op_6445_transpose_y_0"), val = tensor(false)]; + tensor transpose_100_perm_0 = const()[name = tensor("transpose_100_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_101_perm_0 = const()[name = tensor("transpose_101_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_101 = transpose(perm = transpose_101_perm_0, x = keys_87_cast_fp16)[name = tensor("transpose_157")]; + tensor transpose_100 = transpose(perm = transpose_100_perm_0, x = q_87_cast_fp16)[name = tensor("transpose_158")]; + tensor var_6445_cast_fp16 = matmul(transpose_x = var_6445_transpose_x_0, transpose_y = var_6445_transpose_y_0, x = transpose_100, y = transpose_101)[name = tensor("op_6445_cast_fp16")]; + tensor var_6439_to_fp16 = const()[name = tensor("op_6439_to_fp16"), val = tensor([0x1p-3])]; + tensor attn_weights_85_cast_fp16 = mul(x = var_6445_cast_fp16, y = var_6439_to_fp16)[name = tensor("attn_weights_85_cast_fp16")]; + tensor var_6447 = logical_not(x = attn_mask_59)[name = tensor("op_6447")]; + tensor var_6448_to_fp16 = const()[name = tensor("op_6448_to_fp16"), val = tensor(-inf)]; + tensor attn_weights_87_cast_fp16 = select(a = var_6448_to_fp16, b = attn_weights_85_cast_fp16, cond = var_6447)[name = tensor("attn_weights_87_cast_fp16")]; + tensor var_6450 = const()[name = tensor("op_6450"), val = tensor(-1)]; + tensor attn_weights_89_cast_fp16 = softmax(axis = var_6450, x = attn_weights_87_cast_fp16)[name = tensor("attn_weights_89_cast_fp16")]; + tensor attn_output_29_transpose_x_0 = const()[name = tensor("attn_output_29_transpose_x_0"), val = tensor(false)]; + tensor attn_output_29_transpose_y_0 = const()[name = tensor("attn_output_29_transpose_y_0"), val = tensor(false)]; + tensor values_89_cast_fp16 = transpose(perm = var_6396, x = values_87_cast_fp16)[name = tensor("transpose_159")]; + tensor attn_output_29_cast_fp16 = matmul(transpose_x = attn_output_29_transpose_x_0, transpose_y = attn_output_29_transpose_y_0, x = attn_weights_89_cast_fp16, y = values_89_cast_fp16)[name = tensor("attn_output_29_cast_fp16")]; + tensor var_6458 = const()[name = tensor("op_6458"), val = tensor([0, 2, 1, 3])]; + tensor var_6461 = const()[name = tensor("op_6461"), val = tensor([1, 1, 1024])]; + tensor var_6459_cast_fp16 = transpose(perm = var_6458, x = attn_output_29_cast_fp16)[name = tensor("transpose_156")]; + tensor input_145_cast_fp16 = reshape(shape = var_6461, x = var_6459_cast_fp16)[name = tensor("input_145_cast_fp16")]; + tensor linear_58_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = attn14_out_proj_weight_cast_fp16, x = input_145_cast_fp16)[name = tensor("linear_58_cast_fp16")]; + tensor var_6467_to_fp16 = const()[name = tensor("op_6467_to_fp16"), val = tensor(0x1p+0)]; + tensor var_6468_cast_fp16 = add(x = position14_to_fp16, y = var_6467_to_fp16)[name = tensor("op_6468_cast_fp16")]; + tensor var_6468_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_6468_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor input_147_cast_fp16 = add(x = input_143_cast_fp16, y = linear_58_cast_fp16)[name = tensor("input_147_cast_fp16")]; + tensor input_149_axes_0 = const()[name = tensor("input_149_axes_0"), val = tensor([-1])]; + tensor norm14_2_weight_to_fp16 = const()[name = tensor("norm14_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(303709312)))]; + tensor norm14_2_bias_to_fp16 = const()[name = tensor("norm14_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(303711424)))]; + tensor var_6472_to_fp16 = const()[name = tensor("op_6472_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_149_cast_fp16 = layer_norm(axes = input_149_axes_0, beta = norm14_2_bias_to_fp16, epsilon = var_6472_to_fp16, gamma = norm14_2_weight_to_fp16, x = input_147_cast_fp16)[name = tensor("input_149_cast_fp16")]; + tensor linear_59_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = linear14_1_weight_cast_fp16, x = input_149_cast_fp16)[name = tensor("linear_59_cast_fp16")]; + tensor input_151_mode_0 = const()[name = tensor("input_151_mode_0"), val = tensor("EXACT")]; + tensor input_151_cast_fp16 = gelu(mode = input_151_mode_0, x = linear_59_cast_fp16)[name = tensor("input_151_cast_fp16")]; + tensor linear_60_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = linear14_2_weight_cast_fp16, x = input_151_cast_fp16)[name = tensor("linear_60_cast_fp16")]; + tensor input_153_cast_fp16 = add(x = input_147_cast_fp16, y = linear_60_cast_fp16)[name = tensor("input_153_cast_fp16")]; + tensor x_31_axes_0 = const()[name = tensor("x_31_axes_0"), val = tensor([-1])]; + tensor norm15_1_weight_to_fp16 = const()[name = tensor("norm15_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(303713536)))]; + tensor norm15_1_bias_to_fp16 = const()[name = tensor("norm15_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(303715648)))]; + tensor var_6489_to_fp16 = const()[name = tensor("op_6489_to_fp16"), val = tensor(0x1.5p-17)]; + tensor x_31_cast_fp16 = layer_norm(axes = x_31_axes_0, beta = norm15_1_bias_to_fp16, epsilon = var_6489_to_fp16, gamma = norm15_1_weight_to_fp16, x = input_153_cast_fp16)[name = tensor("x_31_cast_fp16")]; + tensor linear_61_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = attn15_in_proj_weight_cast_fp16, x = x_31_cast_fp16)[name = tensor("linear_61_cast_fp16")]; + tensor var_6525 = const()[name = tensor("op_6525"), val = tensor([1, 1, 3, 16, 64])]; + tensor qkv_31_cast_fp16 = reshape(shape = var_6525, x = linear_61_cast_fp16)[name = tensor("qkv_31_cast_fp16")]; + tensor q_91_begin_0 = const()[name = tensor("q_91_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor q_91_end_0 = const()[name = tensor("q_91_end_0"), val = tensor([1, 1, 1, 16, 64])]; + tensor q_91_end_mask_0 = const()[name = tensor("q_91_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor q_91_squeeze_mask_0 = const()[name = tensor("q_91_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor q_91_cast_fp16 = slice_by_index(begin = q_91_begin_0, end = q_91_end_0, end_mask = q_91_end_mask_0, squeeze_mask = q_91_squeeze_mask_0, x = qkv_31_cast_fp16)[name = tensor("q_91_cast_fp16")]; + tensor k_61_begin_0 = const()[name = tensor("k_61_begin_0"), val = tensor([0, 0, 1, 0, 0])]; + tensor k_61_end_0 = const()[name = tensor("k_61_end_0"), val = tensor([1, 1, 2, 16, 64])]; + tensor k_61_end_mask_0 = const()[name = tensor("k_61_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor k_61_squeeze_mask_0 = const()[name = tensor("k_61_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor k_61_cast_fp16 = slice_by_index(begin = k_61_begin_0, end = k_61_end_0, end_mask = k_61_end_mask_0, squeeze_mask = k_61_squeeze_mask_0, x = qkv_31_cast_fp16)[name = tensor("k_61_cast_fp16")]; + tensor v_31_begin_0 = const()[name = tensor("v_31_begin_0"), val = tensor([0, 0, 2, 0, 0])]; + tensor v_31_end_0 = const()[name = tensor("v_31_end_0"), val = tensor([1, 1, 3, 16, 64])]; + tensor v_31_end_mask_0 = const()[name = tensor("v_31_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor v_31_squeeze_mask_0 = const()[name = tensor("v_31_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor v_31_cast_fp16 = slice_by_index(begin = v_31_begin_0, end = v_31_end_0, end_mask = v_31_end_mask_0, squeeze_mask = v_31_squeeze_mask_0, x = qkv_31_cast_fp16)[name = tensor("v_31_cast_fp16")]; + tensor var_6629 = const()[name = tensor("op_6629"), val = tensor([1, 1, 1, 1])]; + tensor position15_to_fp16_dtype_0 = const()[name = tensor("position15_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor position15_to_fp16 = cast(dtype = position15_to_fp16_dtype_0, x = position15)[name = tensor("cast_478")]; + tensor ts_95_cast_fp16 = reshape(shape = var_6629, x = position15_to_fp16)[name = tensor("ts_95_cast_fp16")]; + tensor var_6633 = const()[name = tensor("op_6633"), val = tensor([1, 1, 16, 32, 2])]; + tensor q_complex_31_cast_fp16 = reshape(shape = var_6633, x = q_91_cast_fp16)[name = tensor("q_complex_31_cast_fp16")]; + tensor var_6637 = const()[name = tensor("op_6637"), val = tensor([1, 1, 16, 32, 2])]; + tensor k_complex_31_cast_fp16 = reshape(shape = var_6637, x = k_61_cast_fp16)[name = tensor("k_complex_31_cast_fp16")]; + tensor var_6641_begin_0 = const()[name = tensor("op_6641_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_6641_end_0 = const()[name = tensor("op_6641_end_0"), val = tensor([1, 1, 16, 32, 1])]; + tensor var_6641_end_mask_0 = const()[name = tensor("op_6641_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_6641_squeeze_mask_0 = const()[name = tensor("op_6641_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_6641_cast_fp16 = slice_by_index(begin = var_6641_begin_0, end = var_6641_end_0, end_mask = var_6641_end_mask_0, squeeze_mask = var_6641_squeeze_mask_0, x = q_complex_31_cast_fp16)[name = tensor("op_6641_cast_fp16")]; + tensor var_6649_begin_0 = const()[name = tensor("op_6649_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_6649_end_0 = const()[name = tensor("op_6649_end_0"), val = tensor([1, 1, 16, 32, 2])]; + tensor var_6649_end_mask_0 = const()[name = tensor("op_6649_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_6649_squeeze_mask_0 = const()[name = tensor("op_6649_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_6649_cast_fp16 = slice_by_index(begin = var_6649_begin_0, end = var_6649_end_0, end_mask = var_6649_end_mask_0, squeeze_mask = var_6649_squeeze_mask_0, x = q_complex_31_cast_fp16)[name = tensor("op_6649_cast_fp16")]; + tensor var_6657_begin_0 = const()[name = tensor("op_6657_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_6657_end_0 = const()[name = tensor("op_6657_end_0"), val = tensor([1, 1, 16, 32, 1])]; + tensor var_6657_end_mask_0 = const()[name = tensor("op_6657_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_6657_squeeze_mask_0 = const()[name = tensor("op_6657_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_6657_cast_fp16 = slice_by_index(begin = var_6657_begin_0, end = var_6657_end_0, end_mask = var_6657_end_mask_0, squeeze_mask = var_6657_squeeze_mask_0, x = k_complex_31_cast_fp16)[name = tensor("op_6657_cast_fp16")]; + tensor var_6665_begin_0 = const()[name = tensor("op_6665_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_6665_end_0 = const()[name = tensor("op_6665_end_0"), val = tensor([1, 1, 16, 32, 2])]; + tensor var_6665_end_mask_0 = const()[name = tensor("op_6665_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_6665_squeeze_mask_0 = const()[name = tensor("op_6665_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_6665_cast_fp16 = slice_by_index(begin = var_6665_begin_0, end = var_6665_end_0, end_mask = var_6665_end_mask_0, squeeze_mask = var_6665_squeeze_mask_0, x = k_complex_31_cast_fp16)[name = tensor("op_6665_cast_fp16")]; + tensor freqs_31_to_fp16 = const()[name = tensor("freqs_31_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(303717760)))]; + tensor var_6671_cast_fp16 = mul(x = freqs_31_to_fp16, y = ts_95_cast_fp16)[name = tensor("op_6671_cast_fp16")]; + tensor rotr_31_cast_fp16 = cos(x = var_6671_cast_fp16)[name = tensor("rotr_31_cast_fp16")]; + tensor roti_31_cast_fp16 = sin(x = var_6671_cast_fp16)[name = tensor("roti_31_cast_fp16")]; + tensor var_6675_cast_fp16 = mul(x = var_6641_cast_fp16, y = rotr_31_cast_fp16)[name = tensor("op_6675_cast_fp16")]; + tensor var_6676_cast_fp16 = mul(x = var_6649_cast_fp16, y = roti_31_cast_fp16)[name = tensor("op_6676_cast_fp16")]; + tensor qor_61_cast_fp16 = sub(x = var_6675_cast_fp16, y = var_6676_cast_fp16)[name = tensor("qor_61_cast_fp16")]; + tensor var_6679_cast_fp16 = mul(x = var_6641_cast_fp16, y = roti_31_cast_fp16)[name = tensor("op_6679_cast_fp16")]; + tensor var_6680_cast_fp16 = mul(x = var_6649_cast_fp16, y = rotr_31_cast_fp16)[name = tensor("op_6680_cast_fp16")]; + tensor qoi_61_cast_fp16 = add(x = var_6679_cast_fp16, y = var_6680_cast_fp16)[name = tensor("qoi_61_cast_fp16")]; + tensor var_6683_cast_fp16 = mul(x = var_6657_cast_fp16, y = rotr_31_cast_fp16)[name = tensor("op_6683_cast_fp16")]; + tensor var_6684_cast_fp16 = mul(x = var_6665_cast_fp16, y = roti_31_cast_fp16)[name = tensor("op_6684_cast_fp16")]; + tensor kor_61_cast_fp16 = sub(x = var_6683_cast_fp16, y = var_6684_cast_fp16)[name = tensor("kor_61_cast_fp16")]; + tensor var_6687_cast_fp16 = mul(x = var_6657_cast_fp16, y = roti_31_cast_fp16)[name = tensor("op_6687_cast_fp16")]; + tensor var_6688_cast_fp16 = mul(x = var_6665_cast_fp16, y = rotr_31_cast_fp16)[name = tensor("op_6688_cast_fp16")]; + tensor koi_61_cast_fp16 = add(x = var_6687_cast_fp16, y = var_6688_cast_fp16)[name = tensor("koi_61_cast_fp16")]; + tensor qo_31_axis_0 = const()[name = tensor("qo_31_axis_0"), val = tensor(-1)]; + tensor qo_31_cast_fp16 = stack(axis = qo_31_axis_0, values = (qor_61_cast_fp16, qoi_61_cast_fp16))[name = tensor("qo_31_cast_fp16")]; + tensor ko_31_axis_0 = const()[name = tensor("ko_31_axis_0"), val = tensor(-1)]; + tensor ko_31_cast_fp16 = stack(axis = ko_31_axis_0, values = (kor_61_cast_fp16, koi_61_cast_fp16))[name = tensor("ko_31_cast_fp16")]; + tensor var_6717 = const()[name = tensor("op_6717"), val = tensor([1, 1, 16, 64])]; + tensor q_93_cast_fp16 = reshape(shape = var_6717, x = qo_31_cast_fp16)[name = tensor("q_93_cast_fp16")]; + tensor var_6719 = const()[name = tensor("op_6719"), val = tensor([1, 1, 16, 64])]; + tensor k_63_cast_fp16 = reshape(shape = var_6719, x = ko_31_cast_fp16)[name = tensor("k_63_cast_fp16")]; + tensor _inversed_6741_y_0_to_fp16 = const()[name = tensor("_inversed_6741_y_0_to_fp16"), val = tensor(0x1p-9)]; + tensor _inversed_6741_cast_fp16 = mul(x = ts_95_cast_fp16, y = _inversed_6741_y_0_to_fp16)[name = tensor("_inversed_6741_cast_fp16")]; + tensor var_6742_cast_fp16 = floor(x = _inversed_6741_cast_fp16)[name = tensor("op_6742_cast_fp16")]; + tensor var_6743_to_fp16 = const()[name = tensor("op_6743_to_fp16"), val = tensor(0x1p+9)]; + tensor var_6744_cast_fp16 = mul(x = var_6742_cast_fp16, y = var_6743_to_fp16)[name = tensor("op_6744_cast_fp16")]; + tensor write_indices_float_63_cast_fp16 = sub(x = ts_95_cast_fp16, y = var_6744_cast_fp16)[name = tensor("write_indices_float_63_cast_fp16")]; + tensor var_6751_dtype_0 = const()[name = tensor("op_6751_dtype_0"), val = tensor("int32")]; + tensor write_indices_31_reps_0 = const()[name = tensor("write_indices_31_reps_0"), val = tensor([1, 1, 16, 64])]; + tensor write_indices_float_63_cast_fp16_to_int32 = cast(dtype = var_6751_dtype_0, x = write_indices_float_63_cast_fp16)[name = tensor("cast_477")]; + tensor write_indices_31 = tile(reps = write_indices_31_reps_0, x = write_indices_float_63_cast_fp16_to_int32)[name = tensor("write_indices_31")]; + tensor var_6759_begin_0 = const()[name = tensor("op_6759_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_6759_end_0 = const()[name = tensor("op_6759_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor var_6759_end_mask_0 = const()[name = tensor("op_6759_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_6759_squeeze_mask_0 = const()[name = tensor("op_6759_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor cache15_to_fp16_dtype_0 = const()[name = tensor("cache15_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor cache15_to_fp16 = cast(dtype = cache15_to_fp16_dtype_0, x = cache15)[name = tensor("cast_476")]; + tensor var_6759_cast_fp16 = slice_by_index(begin = var_6759_begin_0, end = var_6759_end_0, end_mask = var_6759_end_mask_0, squeeze_mask = var_6759_squeeze_mask_0, x = cache15_to_fp16)[name = tensor("op_6759_cast_fp16")]; + tensor var_6761_axis_0 = const()[name = tensor("op_6761_axis_0"), val = tensor(1)]; + tensor var_6761_mode_0 = const()[name = tensor("op_6761_mode_0"), val = tensor("update")]; + tensor var_6761_validate_indices_0 = const()[name = tensor("op_6761_validate_indices_0"), val = tensor(false)]; + tensor var_6761_cast_fp16 = scatter_along_axis(axis = var_6761_axis_0, data = var_6759_cast_fp16, indices = write_indices_31, mode = var_6761_mode_0, updates = k_63_cast_fp16, validate_indices = var_6761_validate_indices_0)[name = tensor("op_6761_cast_fp16")]; + tensor concat_107 = const()[name = tensor("concat_107"), val = tensor([0, 0, 0, 0, 0])]; + tensor concat_108 = const()[name = tensor("concat_108"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_31_internal_tensor_assign_1_stride_0 = const()[name = tensor("new_cache_31_internal_tensor_assign_1_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_31_internal_tensor_assign_1_begin_mask_0 = const()[name = tensor("new_cache_31_internal_tensor_assign_1_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_31_internal_tensor_assign_1_end_mask_0 = const()[name = tensor("new_cache_31_internal_tensor_assign_1_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_31_internal_tensor_assign_1_squeeze_mask_0 = const()[name = tensor("new_cache_31_internal_tensor_assign_1_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_78 = const()[name = tensor("shape_78"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_30 = const()[name = tensor("reduce_prod_30"), val = tensor(1048576)]; + tensor range_1d_30_start_0 = const()[name = tensor("range_1d_30_start_0"), val = tensor(0)]; + tensor range_1d_30_step_0 = const()[name = tensor("range_1d_30_step_0"), val = tensor(1)]; + tensor range_1d_30 = range_1d(end = reduce_prod_30, start = range_1d_30_start_0, step = range_1d_30_step_0)[name = tensor("range_1d_30")]; + tensor reshape_150 = reshape(shape = shape_78, x = range_1d_30)[name = tensor("reshape_150")]; + tensor slice_by_index_30 = slice_by_index(begin = concat_107, begin_mask = new_cache_31_internal_tensor_assign_1_begin_mask_0, end = concat_108, end_mask = new_cache_31_internal_tensor_assign_1_end_mask_0, squeeze_mask = new_cache_31_internal_tensor_assign_1_squeeze_mask_0, stride = new_cache_31_internal_tensor_assign_1_stride_0, x = reshape_150)[name = tensor("slice_by_index_30")]; + tensor reshape_151_shape_0 = const()[name = tensor("reshape_151_shape_0"), val = tensor([-1])]; + tensor reshape_151 = reshape(shape = reshape_151_shape_0, x = slice_by_index_30)[name = tensor("reshape_151")]; + tensor reshape_152_shape_0 = const()[name = tensor("reshape_152_shape_0"), val = tensor([-1])]; + tensor reshape_152_cast_fp16 = reshape(shape = reshape_152_shape_0, x = var_6761_cast_fp16)[name = tensor("reshape_152_cast_fp16")]; + tensor reshape_153_shape_0 = const()[name = tensor("reshape_153_shape_0"), val = tensor([-1])]; + tensor reshape_153_cast_fp16 = reshape(shape = reshape_153_shape_0, x = cache15_to_fp16)[name = tensor("reshape_153_cast_fp16")]; + tensor scatter_30_mode_0 = const()[name = tensor("scatter_30_mode_0"), val = tensor("update")]; + tensor scatter_30_axis_0 = const()[name = tensor("scatter_30_axis_0"), val = tensor(0)]; + tensor scatter_30_validate_indices_0 = const()[name = tensor("scatter_30_validate_indices_0"), val = tensor(false)]; + tensor scatter_30_cast_fp16 = scatter(axis = scatter_30_axis_0, data = reshape_153_cast_fp16, indices = reshape_151, mode = scatter_30_mode_0, updates = reshape_152_cast_fp16, validate_indices = scatter_30_validate_indices_0)[name = tensor("scatter_30_cast_fp16")]; + tensor reshape_154_cast_fp16 = reshape(shape = shape_78, x = scatter_30_cast_fp16)[name = tensor("reshape_154_cast_fp16")]; + tensor var_6769_begin_0 = const()[name = tensor("op_6769_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor var_6769_end_0 = const()[name = tensor("op_6769_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor var_6769_end_mask_0 = const()[name = tensor("op_6769_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_6769_squeeze_mask_0 = const()[name = tensor("op_6769_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor var_6769_cast_fp16 = slice_by_index(begin = var_6769_begin_0, end = var_6769_end_0, end_mask = var_6769_end_mask_0, squeeze_mask = var_6769_squeeze_mask_0, x = reshape_154_cast_fp16)[name = tensor("op_6769_cast_fp16")]; + tensor var_6771_axis_0 = const()[name = tensor("op_6771_axis_0"), val = tensor(1)]; + tensor var_6771_mode_0 = const()[name = tensor("op_6771_mode_0"), val = tensor("update")]; + tensor var_6771_validate_indices_0 = const()[name = tensor("op_6771_validate_indices_0"), val = tensor(false)]; + tensor var_6771_cast_fp16 = scatter_along_axis(axis = var_6771_axis_0, data = var_6769_cast_fp16, indices = write_indices_31, mode = var_6771_mode_0, updates = v_31_cast_fp16, validate_indices = var_6771_validate_indices_0)[name = tensor("op_6771_cast_fp16")]; + tensor concat_109 = const()[name = tensor("concat_109"), val = tensor([1, 0, 0, 0, 0])]; + tensor concat_110 = const()[name = tensor("concat_110"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_31_internal_tensor_assign_2_stride_0 = const()[name = tensor("new_cache_31_internal_tensor_assign_2_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_31_internal_tensor_assign_2_begin_mask_0 = const()[name = tensor("new_cache_31_internal_tensor_assign_2_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_31_internal_tensor_assign_2_end_mask_0 = const()[name = tensor("new_cache_31_internal_tensor_assign_2_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_31_internal_tensor_assign_2_squeeze_mask_0 = const()[name = tensor("new_cache_31_internal_tensor_assign_2_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_79 = const()[name = tensor("shape_79"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_31 = const()[name = tensor("reduce_prod_31"), val = tensor(1048576)]; + tensor range_1d_31_start_0 = const()[name = tensor("range_1d_31_start_0"), val = tensor(0)]; + tensor range_1d_31_step_0 = const()[name = tensor("range_1d_31_step_0"), val = tensor(1)]; + tensor range_1d_31 = range_1d(end = reduce_prod_31, start = range_1d_31_start_0, step = range_1d_31_step_0)[name = tensor("range_1d_31")]; + tensor reshape_155 = reshape(shape = shape_79, x = range_1d_31)[name = tensor("reshape_155")]; + tensor slice_by_index_31 = slice_by_index(begin = concat_109, begin_mask = new_cache_31_internal_tensor_assign_2_begin_mask_0, end = concat_110, end_mask = new_cache_31_internal_tensor_assign_2_end_mask_0, squeeze_mask = new_cache_31_internal_tensor_assign_2_squeeze_mask_0, stride = new_cache_31_internal_tensor_assign_2_stride_0, x = reshape_155)[name = tensor("slice_by_index_31")]; + tensor reshape_156_shape_0 = const()[name = tensor("reshape_156_shape_0"), val = tensor([-1])]; + tensor reshape_156 = reshape(shape = reshape_156_shape_0, x = slice_by_index_31)[name = tensor("reshape_156")]; + tensor reshape_157_shape_0 = const()[name = tensor("reshape_157_shape_0"), val = tensor([-1])]; + tensor reshape_157_cast_fp16 = reshape(shape = reshape_157_shape_0, x = var_6771_cast_fp16)[name = tensor("reshape_157_cast_fp16")]; + tensor reshape_158_shape_0 = const()[name = tensor("reshape_158_shape_0"), val = tensor([-1])]; + tensor reshape_158_cast_fp16 = reshape(shape = reshape_158_shape_0, x = reshape_154_cast_fp16)[name = tensor("reshape_158_cast_fp16")]; + tensor scatter_31_mode_0 = const()[name = tensor("scatter_31_mode_0"), val = tensor("update")]; + tensor scatter_31_axis_0 = const()[name = tensor("scatter_31_axis_0"), val = tensor(0)]; + tensor scatter_31_validate_indices_0 = const()[name = tensor("scatter_31_validate_indices_0"), val = tensor(false)]; + tensor scatter_31_cast_fp16 = scatter(axis = scatter_31_axis_0, data = reshape_158_cast_fp16, indices = reshape_156, mode = scatter_31_mode_0, updates = reshape_157_cast_fp16, validate_indices = scatter_31_validate_indices_0)[name = tensor("scatter_31_cast_fp16")]; + tensor reshape_159_cast_fp16 = reshape(shape = shape_79, x = scatter_31_cast_fp16)[name = tensor("reshape_159_cast_fp16")]; + tensor reshape_159_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("reshape_159_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor keys_91_begin_0 = const()[name = tensor("keys_91_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor keys_91_end_0 = const()[name = tensor("keys_91_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor keys_91_end_mask_0 = const()[name = tensor("keys_91_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor keys_91_squeeze_mask_0 = const()[name = tensor("keys_91_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor keys_91_cast_fp16 = slice_by_index(begin = keys_91_begin_0, end = keys_91_end_0, end_mask = keys_91_end_mask_0, squeeze_mask = keys_91_squeeze_mask_0, x = reshape_159_cast_fp16)[name = tensor("keys_91_cast_fp16")]; + tensor values_91_begin_0 = const()[name = tensor("values_91_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor values_91_end_0 = const()[name = tensor("values_91_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor values_91_end_mask_0 = const()[name = tensor("values_91_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor values_91_squeeze_mask_0 = const()[name = tensor("values_91_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor values_91_cast_fp16 = slice_by_index(begin = values_91_begin_0, end = values_91_end_0, end_mask = values_91_end_mask_0, squeeze_mask = values_91_squeeze_mask_0, x = reshape_159_cast_fp16)[name = tensor("values_91_cast_fp16")]; + tensor var_6783_cast_fp16 = not_equal(x = keys_91_cast_fp16, y = keys_91_cast_fp16)[name = tensor("op_6783_cast_fp16")]; + tensor keys_93_cast_fp16 = select(a = var_504_to_fp16, b = keys_91_cast_fp16, cond = var_6783_cast_fp16)[name = tensor("keys_93_cast_fp16")]; + tensor var_6791_cast_fp16 = not_equal(x = values_91_cast_fp16, y = values_91_cast_fp16)[name = tensor("op_6791_cast_fp16")]; + tensor values_93_cast_fp16 = select(a = var_504_to_fp16, b = values_91_cast_fp16, cond = var_6791_cast_fp16)[name = tensor("values_93_cast_fp16")]; + tensor var_6815 = const()[name = tensor("op_6815"), val = tensor([0, 2, 1, 3])]; + tensor var_6828 = const()[name = tensor("op_6828"), val = tensor([1, 1, 1])]; + tensor var_6829_cast_fp16 = reshape(shape = var_6828, x = position15_to_fp16)[name = tensor("op_6829_cast_fp16")]; + tensor var_6846_to_fp16 = const()[name = tensor("op_6846_to_fp16"), val = tensor(0x1p+0)]; + tensor valid_len_31_cast_fp16 = add(x = var_6829_cast_fp16, y = var_6846_to_fp16)[name = tensor("valid_len_31_cast_fp16")]; + tensor valid_mask_31_cast_fp16 = less(x = k_positions_1_promoted_to_fp16, y = valid_len_31_cast_fp16)[name = tensor("valid_mask_31_cast_fp16")]; + tensor causal_mask_31_cast_fp16 = less_equal(x = k_positions_1_promoted_to_fp16, y = var_6829_cast_fp16)[name = tensor("causal_mask_31_cast_fp16")]; + tensor attn_mask_61 = logical_and(x = valid_mask_31_cast_fp16, y = causal_mask_31_cast_fp16)[name = tensor("attn_mask_61")]; + tensor attn_mask_63_axes_0 = const()[name = tensor("attn_mask_63_axes_0"), val = tensor([1])]; + tensor attn_mask_63 = expand_dims(axes = attn_mask_63_axes_0, x = attn_mask_61)[name = tensor("attn_mask_63")]; + tensor var_6864_transpose_x_0 = const()[name = tensor("op_6864_transpose_x_0"), val = tensor(false)]; + tensor var_6864_transpose_y_0 = const()[name = tensor("op_6864_transpose_y_0"), val = tensor(false)]; + tensor transpose_102_perm_0 = const()[name = tensor("transpose_102_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_103_perm_0 = const()[name = tensor("transpose_103_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_103 = transpose(perm = transpose_103_perm_0, x = keys_93_cast_fp16)[name = tensor("transpose_153")]; + tensor transpose_102 = transpose(perm = transpose_102_perm_0, x = q_93_cast_fp16)[name = tensor("transpose_154")]; + tensor var_6864_cast_fp16 = matmul(transpose_x = var_6864_transpose_x_0, transpose_y = var_6864_transpose_y_0, x = transpose_102, y = transpose_103)[name = tensor("op_6864_cast_fp16")]; + tensor var_6858_to_fp16 = const()[name = tensor("op_6858_to_fp16"), val = tensor([0x1p-3])]; + tensor attn_weights_91_cast_fp16 = mul(x = var_6864_cast_fp16, y = var_6858_to_fp16)[name = tensor("attn_weights_91_cast_fp16")]; + tensor var_6866 = logical_not(x = attn_mask_63)[name = tensor("op_6866")]; + tensor var_6867_to_fp16 = const()[name = tensor("op_6867_to_fp16"), val = tensor(-inf)]; + tensor attn_weights_93_cast_fp16 = select(a = var_6867_to_fp16, b = attn_weights_91_cast_fp16, cond = var_6866)[name = tensor("attn_weights_93_cast_fp16")]; + tensor var_6869 = const()[name = tensor("op_6869"), val = tensor(-1)]; + tensor attn_weights_95_cast_fp16 = softmax(axis = var_6869, x = attn_weights_93_cast_fp16)[name = tensor("attn_weights_95_cast_fp16")]; + tensor attn_output_31_transpose_x_0 = const()[name = tensor("attn_output_31_transpose_x_0"), val = tensor(false)]; + tensor attn_output_31_transpose_y_0 = const()[name = tensor("attn_output_31_transpose_y_0"), val = tensor(false)]; + tensor values_95_cast_fp16 = transpose(perm = var_6815, x = values_93_cast_fp16)[name = tensor("transpose_155")]; + tensor attn_output_31_cast_fp16 = matmul(transpose_x = attn_output_31_transpose_x_0, transpose_y = attn_output_31_transpose_y_0, x = attn_weights_95_cast_fp16, y = values_95_cast_fp16)[name = tensor("attn_output_31_cast_fp16")]; + tensor var_6877 = const()[name = tensor("op_6877"), val = tensor([0, 2, 1, 3])]; + tensor var_6880 = const()[name = tensor("op_6880"), val = tensor([1, 1, 1024])]; + tensor var_6878_cast_fp16 = transpose(perm = var_6877, x = attn_output_31_cast_fp16)[name = tensor("transpose_152")]; + tensor input_155_cast_fp16 = reshape(shape = var_6880, x = var_6878_cast_fp16)[name = tensor("input_155_cast_fp16")]; + tensor linear_62_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = attn15_out_proj_weight_cast_fp16, x = input_155_cast_fp16)[name = tensor("linear_62_cast_fp16")]; + tensor var_6886_to_fp16 = const()[name = tensor("op_6886_to_fp16"), val = tensor(0x1p+0)]; + tensor var_6887_cast_fp16 = add(x = position15_to_fp16, y = var_6886_to_fp16)[name = tensor("op_6887_cast_fp16")]; + tensor var_6887_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_6887_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor input_157_cast_fp16 = add(x = input_153_cast_fp16, y = linear_62_cast_fp16)[name = tensor("input_157_cast_fp16")]; + tensor input_159_axes_0 = const()[name = tensor("input_159_axes_0"), val = tensor([-1])]; + tensor norm15_2_weight_to_fp16 = const()[name = tensor("norm15_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(303717888)))]; + tensor norm15_2_bias_to_fp16 = const()[name = tensor("norm15_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(303720000)))]; + tensor var_6891_to_fp16 = const()[name = tensor("op_6891_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_159_cast_fp16 = layer_norm(axes = input_159_axes_0, beta = norm15_2_bias_to_fp16, epsilon = var_6891_to_fp16, gamma = norm15_2_weight_to_fp16, x = input_157_cast_fp16)[name = tensor("input_159_cast_fp16")]; + tensor linear_63_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = linear15_1_weight_cast_fp16, x = input_159_cast_fp16)[name = tensor("linear_63_cast_fp16")]; + tensor input_161_mode_0 = const()[name = tensor("input_161_mode_0"), val = tensor("EXACT")]; + tensor input_161_cast_fp16 = gelu(mode = input_161_mode_0, x = linear_63_cast_fp16)[name = tensor("input_161_cast_fp16")]; + tensor linear_64_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = linear15_2_weight_cast_fp16, x = input_161_cast_fp16)[name = tensor("linear_64_cast_fp16")]; + tensor input_163_cast_fp16 = add(x = input_157_cast_fp16, y = linear_64_cast_fp16)[name = tensor("input_163_cast_fp16")]; + tensor x_33_axes_0 = const()[name = tensor("x_33_axes_0"), val = tensor([-1])]; + tensor norm16_1_weight_to_fp16 = const()[name = tensor("norm16_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(303722112)))]; + tensor norm16_1_bias_to_fp16 = const()[name = tensor("norm16_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(303724224)))]; + tensor var_6908_to_fp16 = const()[name = tensor("op_6908_to_fp16"), val = tensor(0x1.5p-17)]; + tensor x_33_cast_fp16 = layer_norm(axes = x_33_axes_0, beta = norm16_1_bias_to_fp16, epsilon = var_6908_to_fp16, gamma = norm16_1_weight_to_fp16, x = input_163_cast_fp16)[name = tensor("x_33_cast_fp16")]; + tensor linear_65_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = attn16_in_proj_weight_cast_fp16, x = x_33_cast_fp16)[name = tensor("linear_65_cast_fp16")]; + tensor var_6944 = const()[name = tensor("op_6944"), val = tensor([1, 1, 3, 16, 64])]; + tensor qkv_33_cast_fp16 = reshape(shape = var_6944, x = linear_65_cast_fp16)[name = tensor("qkv_33_cast_fp16")]; + tensor q_97_begin_0 = const()[name = tensor("q_97_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor q_97_end_0 = const()[name = tensor("q_97_end_0"), val = tensor([1, 1, 1, 16, 64])]; + tensor q_97_end_mask_0 = const()[name = tensor("q_97_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor q_97_squeeze_mask_0 = const()[name = tensor("q_97_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor q_97_cast_fp16 = slice_by_index(begin = q_97_begin_0, end = q_97_end_0, end_mask = q_97_end_mask_0, squeeze_mask = q_97_squeeze_mask_0, x = qkv_33_cast_fp16)[name = tensor("q_97_cast_fp16")]; + tensor k_65_begin_0 = const()[name = tensor("k_65_begin_0"), val = tensor([0, 0, 1, 0, 0])]; + tensor k_65_end_0 = const()[name = tensor("k_65_end_0"), val = tensor([1, 1, 2, 16, 64])]; + tensor k_65_end_mask_0 = const()[name = tensor("k_65_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor k_65_squeeze_mask_0 = const()[name = tensor("k_65_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor k_65_cast_fp16 = slice_by_index(begin = k_65_begin_0, end = k_65_end_0, end_mask = k_65_end_mask_0, squeeze_mask = k_65_squeeze_mask_0, x = qkv_33_cast_fp16)[name = tensor("k_65_cast_fp16")]; + tensor v_33_begin_0 = const()[name = tensor("v_33_begin_0"), val = tensor([0, 0, 2, 0, 0])]; + tensor v_33_end_0 = const()[name = tensor("v_33_end_0"), val = tensor([1, 1, 3, 16, 64])]; + tensor v_33_end_mask_0 = const()[name = tensor("v_33_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor v_33_squeeze_mask_0 = const()[name = tensor("v_33_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor v_33_cast_fp16 = slice_by_index(begin = v_33_begin_0, end = v_33_end_0, end_mask = v_33_end_mask_0, squeeze_mask = v_33_squeeze_mask_0, x = qkv_33_cast_fp16)[name = tensor("v_33_cast_fp16")]; + tensor var_7048 = const()[name = tensor("op_7048"), val = tensor([1, 1, 1, 1])]; + tensor position16_to_fp16_dtype_0 = const()[name = tensor("position16_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor position16_to_fp16 = cast(dtype = position16_to_fp16_dtype_0, x = position16)[name = tensor("cast_473")]; + tensor ts_101_cast_fp16 = reshape(shape = var_7048, x = position16_to_fp16)[name = tensor("ts_101_cast_fp16")]; + tensor var_7052 = const()[name = tensor("op_7052"), val = tensor([1, 1, 16, 32, 2])]; + tensor q_complex_33_cast_fp16 = reshape(shape = var_7052, x = q_97_cast_fp16)[name = tensor("q_complex_33_cast_fp16")]; + tensor var_7056 = const()[name = tensor("op_7056"), val = tensor([1, 1, 16, 32, 2])]; + tensor k_complex_33_cast_fp16 = reshape(shape = var_7056, x = k_65_cast_fp16)[name = tensor("k_complex_33_cast_fp16")]; + tensor var_7060_begin_0 = const()[name = tensor("op_7060_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_7060_end_0 = const()[name = tensor("op_7060_end_0"), val = tensor([1, 1, 16, 32, 1])]; + tensor var_7060_end_mask_0 = const()[name = tensor("op_7060_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_7060_squeeze_mask_0 = const()[name = tensor("op_7060_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_7060_cast_fp16 = slice_by_index(begin = var_7060_begin_0, end = var_7060_end_0, end_mask = var_7060_end_mask_0, squeeze_mask = var_7060_squeeze_mask_0, x = q_complex_33_cast_fp16)[name = tensor("op_7060_cast_fp16")]; + tensor var_7068_begin_0 = const()[name = tensor("op_7068_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_7068_end_0 = const()[name = tensor("op_7068_end_0"), val = tensor([1, 1, 16, 32, 2])]; + tensor var_7068_end_mask_0 = const()[name = tensor("op_7068_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_7068_squeeze_mask_0 = const()[name = tensor("op_7068_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_7068_cast_fp16 = slice_by_index(begin = var_7068_begin_0, end = var_7068_end_0, end_mask = var_7068_end_mask_0, squeeze_mask = var_7068_squeeze_mask_0, x = q_complex_33_cast_fp16)[name = tensor("op_7068_cast_fp16")]; + tensor var_7076_begin_0 = const()[name = tensor("op_7076_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_7076_end_0 = const()[name = tensor("op_7076_end_0"), val = tensor([1, 1, 16, 32, 1])]; + tensor var_7076_end_mask_0 = const()[name = tensor("op_7076_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_7076_squeeze_mask_0 = const()[name = tensor("op_7076_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_7076_cast_fp16 = slice_by_index(begin = var_7076_begin_0, end = var_7076_end_0, end_mask = var_7076_end_mask_0, squeeze_mask = var_7076_squeeze_mask_0, x = k_complex_33_cast_fp16)[name = tensor("op_7076_cast_fp16")]; + tensor var_7084_begin_0 = const()[name = tensor("op_7084_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_7084_end_0 = const()[name = tensor("op_7084_end_0"), val = tensor([1, 1, 16, 32, 2])]; + tensor var_7084_end_mask_0 = const()[name = tensor("op_7084_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_7084_squeeze_mask_0 = const()[name = tensor("op_7084_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_7084_cast_fp16 = slice_by_index(begin = var_7084_begin_0, end = var_7084_end_0, end_mask = var_7084_end_mask_0, squeeze_mask = var_7084_squeeze_mask_0, x = k_complex_33_cast_fp16)[name = tensor("op_7084_cast_fp16")]; + tensor freqs_33_to_fp16 = const()[name = tensor("freqs_33_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(303726336)))]; + tensor var_7090_cast_fp16 = mul(x = freqs_33_to_fp16, y = ts_101_cast_fp16)[name = tensor("op_7090_cast_fp16")]; + tensor rotr_33_cast_fp16 = cos(x = var_7090_cast_fp16)[name = tensor("rotr_33_cast_fp16")]; + tensor roti_33_cast_fp16 = sin(x = var_7090_cast_fp16)[name = tensor("roti_33_cast_fp16")]; + tensor var_7094_cast_fp16 = mul(x = var_7060_cast_fp16, y = rotr_33_cast_fp16)[name = tensor("op_7094_cast_fp16")]; + tensor var_7095_cast_fp16 = mul(x = var_7068_cast_fp16, y = roti_33_cast_fp16)[name = tensor("op_7095_cast_fp16")]; + tensor qor_65_cast_fp16 = sub(x = var_7094_cast_fp16, y = var_7095_cast_fp16)[name = tensor("qor_65_cast_fp16")]; + tensor var_7098_cast_fp16 = mul(x = var_7060_cast_fp16, y = roti_33_cast_fp16)[name = tensor("op_7098_cast_fp16")]; + tensor var_7099_cast_fp16 = mul(x = var_7068_cast_fp16, y = rotr_33_cast_fp16)[name = tensor("op_7099_cast_fp16")]; + tensor qoi_65_cast_fp16 = add(x = var_7098_cast_fp16, y = var_7099_cast_fp16)[name = tensor("qoi_65_cast_fp16")]; + tensor var_7102_cast_fp16 = mul(x = var_7076_cast_fp16, y = rotr_33_cast_fp16)[name = tensor("op_7102_cast_fp16")]; + tensor var_7103_cast_fp16 = mul(x = var_7084_cast_fp16, y = roti_33_cast_fp16)[name = tensor("op_7103_cast_fp16")]; + tensor kor_65_cast_fp16 = sub(x = var_7102_cast_fp16, y = var_7103_cast_fp16)[name = tensor("kor_65_cast_fp16")]; + tensor var_7106_cast_fp16 = mul(x = var_7076_cast_fp16, y = roti_33_cast_fp16)[name = tensor("op_7106_cast_fp16")]; + tensor var_7107_cast_fp16 = mul(x = var_7084_cast_fp16, y = rotr_33_cast_fp16)[name = tensor("op_7107_cast_fp16")]; + tensor koi_65_cast_fp16 = add(x = var_7106_cast_fp16, y = var_7107_cast_fp16)[name = tensor("koi_65_cast_fp16")]; + tensor qo_33_axis_0 = const()[name = tensor("qo_33_axis_0"), val = tensor(-1)]; + tensor qo_33_cast_fp16 = stack(axis = qo_33_axis_0, values = (qor_65_cast_fp16, qoi_65_cast_fp16))[name = tensor("qo_33_cast_fp16")]; + tensor ko_33_axis_0 = const()[name = tensor("ko_33_axis_0"), val = tensor(-1)]; + tensor ko_33_cast_fp16 = stack(axis = ko_33_axis_0, values = (kor_65_cast_fp16, koi_65_cast_fp16))[name = tensor("ko_33_cast_fp16")]; + tensor var_7136 = const()[name = tensor("op_7136"), val = tensor([1, 1, 16, 64])]; + tensor q_99_cast_fp16 = reshape(shape = var_7136, x = qo_33_cast_fp16)[name = tensor("q_99_cast_fp16")]; + tensor var_7138 = const()[name = tensor("op_7138"), val = tensor([1, 1, 16, 64])]; + tensor k_67_cast_fp16 = reshape(shape = var_7138, x = ko_33_cast_fp16)[name = tensor("k_67_cast_fp16")]; + tensor _inversed_7160_y_0_to_fp16 = const()[name = tensor("_inversed_7160_y_0_to_fp16"), val = tensor(0x1p-9)]; + tensor _inversed_7160_cast_fp16 = mul(x = ts_101_cast_fp16, y = _inversed_7160_y_0_to_fp16)[name = tensor("_inversed_7160_cast_fp16")]; + tensor var_7161_cast_fp16 = floor(x = _inversed_7160_cast_fp16)[name = tensor("op_7161_cast_fp16")]; + tensor var_7162_to_fp16 = const()[name = tensor("op_7162_to_fp16"), val = tensor(0x1p+9)]; + tensor var_7163_cast_fp16 = mul(x = var_7161_cast_fp16, y = var_7162_to_fp16)[name = tensor("op_7163_cast_fp16")]; + tensor write_indices_float_67_cast_fp16 = sub(x = ts_101_cast_fp16, y = var_7163_cast_fp16)[name = tensor("write_indices_float_67_cast_fp16")]; + tensor var_7170_dtype_0 = const()[name = tensor("op_7170_dtype_0"), val = tensor("int32")]; + tensor write_indices_33_reps_0 = const()[name = tensor("write_indices_33_reps_0"), val = tensor([1, 1, 16, 64])]; + tensor write_indices_float_67_cast_fp16_to_int32 = cast(dtype = var_7170_dtype_0, x = write_indices_float_67_cast_fp16)[name = tensor("cast_472")]; + tensor write_indices_33 = tile(reps = write_indices_33_reps_0, x = write_indices_float_67_cast_fp16_to_int32)[name = tensor("write_indices_33")]; + tensor var_7178_begin_0 = const()[name = tensor("op_7178_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_7178_end_0 = const()[name = tensor("op_7178_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor var_7178_end_mask_0 = const()[name = tensor("op_7178_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_7178_squeeze_mask_0 = const()[name = tensor("op_7178_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor cache16_to_fp16_dtype_0 = const()[name = tensor("cache16_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor cache16_to_fp16 = cast(dtype = cache16_to_fp16_dtype_0, x = cache16)[name = tensor("cast_471")]; + tensor var_7178_cast_fp16 = slice_by_index(begin = var_7178_begin_0, end = var_7178_end_0, end_mask = var_7178_end_mask_0, squeeze_mask = var_7178_squeeze_mask_0, x = cache16_to_fp16)[name = tensor("op_7178_cast_fp16")]; + tensor var_7180_axis_0 = const()[name = tensor("op_7180_axis_0"), val = tensor(1)]; + tensor var_7180_mode_0 = const()[name = tensor("op_7180_mode_0"), val = tensor("update")]; + tensor var_7180_validate_indices_0 = const()[name = tensor("op_7180_validate_indices_0"), val = tensor(false)]; + tensor var_7180_cast_fp16 = scatter_along_axis(axis = var_7180_axis_0, data = var_7178_cast_fp16, indices = write_indices_33, mode = var_7180_mode_0, updates = k_67_cast_fp16, validate_indices = var_7180_validate_indices_0)[name = tensor("op_7180_cast_fp16")]; + tensor concat_114 = const()[name = tensor("concat_114"), val = tensor([0, 0, 0, 0, 0])]; + tensor concat_115 = const()[name = tensor("concat_115"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_33_internal_tensor_assign_1_stride_0 = const()[name = tensor("new_cache_33_internal_tensor_assign_1_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_33_internal_tensor_assign_1_begin_mask_0 = const()[name = tensor("new_cache_33_internal_tensor_assign_1_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_33_internal_tensor_assign_1_end_mask_0 = const()[name = tensor("new_cache_33_internal_tensor_assign_1_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_33_internal_tensor_assign_1_squeeze_mask_0 = const()[name = tensor("new_cache_33_internal_tensor_assign_1_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_80 = const()[name = tensor("shape_80"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_32 = const()[name = tensor("reduce_prod_32"), val = tensor(1048576)]; + tensor range_1d_32_start_0 = const()[name = tensor("range_1d_32_start_0"), val = tensor(0)]; + tensor range_1d_32_step_0 = const()[name = tensor("range_1d_32_step_0"), val = tensor(1)]; + tensor range_1d_32 = range_1d(end = reduce_prod_32, start = range_1d_32_start_0, step = range_1d_32_step_0)[name = tensor("range_1d_32")]; + tensor reshape_160 = reshape(shape = shape_80, x = range_1d_32)[name = tensor("reshape_160")]; + tensor slice_by_index_32 = slice_by_index(begin = concat_114, begin_mask = new_cache_33_internal_tensor_assign_1_begin_mask_0, end = concat_115, end_mask = new_cache_33_internal_tensor_assign_1_end_mask_0, squeeze_mask = new_cache_33_internal_tensor_assign_1_squeeze_mask_0, stride = new_cache_33_internal_tensor_assign_1_stride_0, x = reshape_160)[name = tensor("slice_by_index_32")]; + tensor reshape_161_shape_0 = const()[name = tensor("reshape_161_shape_0"), val = tensor([-1])]; + tensor reshape_161 = reshape(shape = reshape_161_shape_0, x = slice_by_index_32)[name = tensor("reshape_161")]; + tensor reshape_162_shape_0 = const()[name = tensor("reshape_162_shape_0"), val = tensor([-1])]; + tensor reshape_162_cast_fp16 = reshape(shape = reshape_162_shape_0, x = var_7180_cast_fp16)[name = tensor("reshape_162_cast_fp16")]; + tensor reshape_163_shape_0 = const()[name = tensor("reshape_163_shape_0"), val = tensor([-1])]; + tensor reshape_163_cast_fp16 = reshape(shape = reshape_163_shape_0, x = cache16_to_fp16)[name = tensor("reshape_163_cast_fp16")]; + tensor scatter_32_mode_0 = const()[name = tensor("scatter_32_mode_0"), val = tensor("update")]; + tensor scatter_32_axis_0 = const()[name = tensor("scatter_32_axis_0"), val = tensor(0)]; + tensor scatter_32_validate_indices_0 = const()[name = tensor("scatter_32_validate_indices_0"), val = tensor(false)]; + tensor scatter_32_cast_fp16 = scatter(axis = scatter_32_axis_0, data = reshape_163_cast_fp16, indices = reshape_161, mode = scatter_32_mode_0, updates = reshape_162_cast_fp16, validate_indices = scatter_32_validate_indices_0)[name = tensor("scatter_32_cast_fp16")]; + tensor reshape_164_cast_fp16 = reshape(shape = shape_80, x = scatter_32_cast_fp16)[name = tensor("reshape_164_cast_fp16")]; + tensor var_7188_begin_0 = const()[name = tensor("op_7188_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor var_7188_end_0 = const()[name = tensor("op_7188_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor var_7188_end_mask_0 = const()[name = tensor("op_7188_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_7188_squeeze_mask_0 = const()[name = tensor("op_7188_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor var_7188_cast_fp16 = slice_by_index(begin = var_7188_begin_0, end = var_7188_end_0, end_mask = var_7188_end_mask_0, squeeze_mask = var_7188_squeeze_mask_0, x = reshape_164_cast_fp16)[name = tensor("op_7188_cast_fp16")]; + tensor var_7190_axis_0 = const()[name = tensor("op_7190_axis_0"), val = tensor(1)]; + tensor var_7190_mode_0 = const()[name = tensor("op_7190_mode_0"), val = tensor("update")]; + tensor var_7190_validate_indices_0 = const()[name = tensor("op_7190_validate_indices_0"), val = tensor(false)]; + tensor var_7190_cast_fp16 = scatter_along_axis(axis = var_7190_axis_0, data = var_7188_cast_fp16, indices = write_indices_33, mode = var_7190_mode_0, updates = v_33_cast_fp16, validate_indices = var_7190_validate_indices_0)[name = tensor("op_7190_cast_fp16")]; + tensor concat_116 = const()[name = tensor("concat_116"), val = tensor([1, 0, 0, 0, 0])]; + tensor concat_117 = const()[name = tensor("concat_117"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_33_internal_tensor_assign_2_stride_0 = const()[name = tensor("new_cache_33_internal_tensor_assign_2_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_33_internal_tensor_assign_2_begin_mask_0 = const()[name = tensor("new_cache_33_internal_tensor_assign_2_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_33_internal_tensor_assign_2_end_mask_0 = const()[name = tensor("new_cache_33_internal_tensor_assign_2_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_33_internal_tensor_assign_2_squeeze_mask_0 = const()[name = tensor("new_cache_33_internal_tensor_assign_2_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_81 = const()[name = tensor("shape_81"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_33 = const()[name = tensor("reduce_prod_33"), val = tensor(1048576)]; + tensor range_1d_33_start_0 = const()[name = tensor("range_1d_33_start_0"), val = tensor(0)]; + tensor range_1d_33_step_0 = const()[name = tensor("range_1d_33_step_0"), val = tensor(1)]; + tensor range_1d_33 = range_1d(end = reduce_prod_33, start = range_1d_33_start_0, step = range_1d_33_step_0)[name = tensor("range_1d_33")]; + tensor reshape_165 = reshape(shape = shape_81, x = range_1d_33)[name = tensor("reshape_165")]; + tensor slice_by_index_33 = slice_by_index(begin = concat_116, begin_mask = new_cache_33_internal_tensor_assign_2_begin_mask_0, end = concat_117, end_mask = new_cache_33_internal_tensor_assign_2_end_mask_0, squeeze_mask = new_cache_33_internal_tensor_assign_2_squeeze_mask_0, stride = new_cache_33_internal_tensor_assign_2_stride_0, x = reshape_165)[name = tensor("slice_by_index_33")]; + tensor reshape_166_shape_0 = const()[name = tensor("reshape_166_shape_0"), val = tensor([-1])]; + tensor reshape_166 = reshape(shape = reshape_166_shape_0, x = slice_by_index_33)[name = tensor("reshape_166")]; + tensor reshape_167_shape_0 = const()[name = tensor("reshape_167_shape_0"), val = tensor([-1])]; + tensor reshape_167_cast_fp16 = reshape(shape = reshape_167_shape_0, x = var_7190_cast_fp16)[name = tensor("reshape_167_cast_fp16")]; + tensor reshape_168_shape_0 = const()[name = tensor("reshape_168_shape_0"), val = tensor([-1])]; + tensor reshape_168_cast_fp16 = reshape(shape = reshape_168_shape_0, x = reshape_164_cast_fp16)[name = tensor("reshape_168_cast_fp16")]; + tensor scatter_33_mode_0 = const()[name = tensor("scatter_33_mode_0"), val = tensor("update")]; + tensor scatter_33_axis_0 = const()[name = tensor("scatter_33_axis_0"), val = tensor(0)]; + tensor scatter_33_validate_indices_0 = const()[name = tensor("scatter_33_validate_indices_0"), val = tensor(false)]; + tensor scatter_33_cast_fp16 = scatter(axis = scatter_33_axis_0, data = reshape_168_cast_fp16, indices = reshape_166, mode = scatter_33_mode_0, updates = reshape_167_cast_fp16, validate_indices = scatter_33_validate_indices_0)[name = tensor("scatter_33_cast_fp16")]; + tensor reshape_169_cast_fp16 = reshape(shape = shape_81, x = scatter_33_cast_fp16)[name = tensor("reshape_169_cast_fp16")]; + tensor reshape_169_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("reshape_169_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor keys_97_begin_0 = const()[name = tensor("keys_97_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor keys_97_end_0 = const()[name = tensor("keys_97_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor keys_97_end_mask_0 = const()[name = tensor("keys_97_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor keys_97_squeeze_mask_0 = const()[name = tensor("keys_97_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor keys_97_cast_fp16 = slice_by_index(begin = keys_97_begin_0, end = keys_97_end_0, end_mask = keys_97_end_mask_0, squeeze_mask = keys_97_squeeze_mask_0, x = reshape_169_cast_fp16)[name = tensor("keys_97_cast_fp16")]; + tensor values_97_begin_0 = const()[name = tensor("values_97_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor values_97_end_0 = const()[name = tensor("values_97_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor values_97_end_mask_0 = const()[name = tensor("values_97_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor values_97_squeeze_mask_0 = const()[name = tensor("values_97_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor values_97_cast_fp16 = slice_by_index(begin = values_97_begin_0, end = values_97_end_0, end_mask = values_97_end_mask_0, squeeze_mask = values_97_squeeze_mask_0, x = reshape_169_cast_fp16)[name = tensor("values_97_cast_fp16")]; + tensor var_7202_cast_fp16 = not_equal(x = keys_97_cast_fp16, y = keys_97_cast_fp16)[name = tensor("op_7202_cast_fp16")]; + tensor keys_99_cast_fp16 = select(a = var_504_to_fp16, b = keys_97_cast_fp16, cond = var_7202_cast_fp16)[name = tensor("keys_99_cast_fp16")]; + tensor var_7210_cast_fp16 = not_equal(x = values_97_cast_fp16, y = values_97_cast_fp16)[name = tensor("op_7210_cast_fp16")]; + tensor values_99_cast_fp16 = select(a = var_504_to_fp16, b = values_97_cast_fp16, cond = var_7210_cast_fp16)[name = tensor("values_99_cast_fp16")]; + tensor var_7234 = const()[name = tensor("op_7234"), val = tensor([0, 2, 1, 3])]; + tensor var_7247 = const()[name = tensor("op_7247"), val = tensor([1, 1, 1])]; + tensor var_7248_cast_fp16 = reshape(shape = var_7247, x = position16_to_fp16)[name = tensor("op_7248_cast_fp16")]; + tensor var_7265_to_fp16 = const()[name = tensor("op_7265_to_fp16"), val = tensor(0x1p+0)]; + tensor valid_len_33_cast_fp16 = add(x = var_7248_cast_fp16, y = var_7265_to_fp16)[name = tensor("valid_len_33_cast_fp16")]; + tensor valid_mask_33_cast_fp16 = less(x = k_positions_1_promoted_to_fp16, y = valid_len_33_cast_fp16)[name = tensor("valid_mask_33_cast_fp16")]; + tensor causal_mask_33_cast_fp16 = less_equal(x = k_positions_1_promoted_to_fp16, y = var_7248_cast_fp16)[name = tensor("causal_mask_33_cast_fp16")]; + tensor attn_mask_65 = logical_and(x = valid_mask_33_cast_fp16, y = causal_mask_33_cast_fp16)[name = tensor("attn_mask_65")]; + tensor attn_mask_67_axes_0 = const()[name = tensor("attn_mask_67_axes_0"), val = tensor([1])]; + tensor attn_mask_67 = expand_dims(axes = attn_mask_67_axes_0, x = attn_mask_65)[name = tensor("attn_mask_67")]; + tensor var_7283_transpose_x_0 = const()[name = tensor("op_7283_transpose_x_0"), val = tensor(false)]; + tensor var_7283_transpose_y_0 = const()[name = tensor("op_7283_transpose_y_0"), val = tensor(false)]; + tensor transpose_104_perm_0 = const()[name = tensor("transpose_104_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_105_perm_0 = const()[name = tensor("transpose_105_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_105 = transpose(perm = transpose_105_perm_0, x = keys_99_cast_fp16)[name = tensor("transpose_149")]; + tensor transpose_104 = transpose(perm = transpose_104_perm_0, x = q_99_cast_fp16)[name = tensor("transpose_150")]; + tensor var_7283_cast_fp16 = matmul(transpose_x = var_7283_transpose_x_0, transpose_y = var_7283_transpose_y_0, x = transpose_104, y = transpose_105)[name = tensor("op_7283_cast_fp16")]; + tensor var_7277_to_fp16 = const()[name = tensor("op_7277_to_fp16"), val = tensor([0x1p-3])]; + tensor attn_weights_97_cast_fp16 = mul(x = var_7283_cast_fp16, y = var_7277_to_fp16)[name = tensor("attn_weights_97_cast_fp16")]; + tensor var_7285 = logical_not(x = attn_mask_67)[name = tensor("op_7285")]; + tensor var_7286_to_fp16 = const()[name = tensor("op_7286_to_fp16"), val = tensor(-inf)]; + tensor attn_weights_99_cast_fp16 = select(a = var_7286_to_fp16, b = attn_weights_97_cast_fp16, cond = var_7285)[name = tensor("attn_weights_99_cast_fp16")]; + tensor var_7288 = const()[name = tensor("op_7288"), val = tensor(-1)]; + tensor attn_weights_101_cast_fp16 = softmax(axis = var_7288, x = attn_weights_99_cast_fp16)[name = tensor("attn_weights_101_cast_fp16")]; + tensor attn_output_33_transpose_x_0 = const()[name = tensor("attn_output_33_transpose_x_0"), val = tensor(false)]; + tensor attn_output_33_transpose_y_0 = const()[name = tensor("attn_output_33_transpose_y_0"), val = tensor(false)]; + tensor values_101_cast_fp16 = transpose(perm = var_7234, x = values_99_cast_fp16)[name = tensor("transpose_151")]; + tensor attn_output_33_cast_fp16 = matmul(transpose_x = attn_output_33_transpose_x_0, transpose_y = attn_output_33_transpose_y_0, x = attn_weights_101_cast_fp16, y = values_101_cast_fp16)[name = tensor("attn_output_33_cast_fp16")]; + tensor var_7296 = const()[name = tensor("op_7296"), val = tensor([0, 2, 1, 3])]; + tensor var_7299 = const()[name = tensor("op_7299"), val = tensor([1, 1, 1024])]; + tensor var_7297_cast_fp16 = transpose(perm = var_7296, x = attn_output_33_cast_fp16)[name = tensor("transpose_148")]; + tensor input_165_cast_fp16 = reshape(shape = var_7299, x = var_7297_cast_fp16)[name = tensor("input_165_cast_fp16")]; + tensor linear_66_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = attn16_out_proj_weight_cast_fp16, x = input_165_cast_fp16)[name = tensor("linear_66_cast_fp16")]; + tensor var_7305_to_fp16 = const()[name = tensor("op_7305_to_fp16"), val = tensor(0x1p+0)]; + tensor var_7306_cast_fp16 = add(x = position16_to_fp16, y = var_7305_to_fp16)[name = tensor("op_7306_cast_fp16")]; + tensor var_7306_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_7306_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor input_167_cast_fp16 = add(x = input_163_cast_fp16, y = linear_66_cast_fp16)[name = tensor("input_167_cast_fp16")]; + tensor input_169_axes_0 = const()[name = tensor("input_169_axes_0"), val = tensor([-1])]; + tensor norm16_2_weight_to_fp16 = const()[name = tensor("norm16_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(303726464)))]; + tensor norm16_2_bias_to_fp16 = const()[name = tensor("norm16_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(303728576)))]; + tensor var_7310_to_fp16 = const()[name = tensor("op_7310_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_169_cast_fp16 = layer_norm(axes = input_169_axes_0, beta = norm16_2_bias_to_fp16, epsilon = var_7310_to_fp16, gamma = norm16_2_weight_to_fp16, x = input_167_cast_fp16)[name = tensor("input_169_cast_fp16")]; + tensor linear_67_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = linear16_1_weight_cast_fp16, x = input_169_cast_fp16)[name = tensor("linear_67_cast_fp16")]; + tensor input_171_mode_0 = const()[name = tensor("input_171_mode_0"), val = tensor("EXACT")]; + tensor input_171_cast_fp16 = gelu(mode = input_171_mode_0, x = linear_67_cast_fp16)[name = tensor("input_171_cast_fp16")]; + tensor linear_68_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = linear16_2_weight_cast_fp16, x = input_171_cast_fp16)[name = tensor("linear_68_cast_fp16")]; + tensor input_173_cast_fp16 = add(x = input_167_cast_fp16, y = linear_68_cast_fp16)[name = tensor("input_173_cast_fp16")]; + tensor x_35_axes_0 = const()[name = tensor("x_35_axes_0"), val = tensor([-1])]; + tensor norm17_1_weight_to_fp16 = const()[name = tensor("norm17_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(303730688)))]; + tensor norm17_1_bias_to_fp16 = const()[name = tensor("norm17_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(303732800)))]; + tensor var_7327_to_fp16 = const()[name = tensor("op_7327_to_fp16"), val = tensor(0x1.5p-17)]; + tensor x_35_cast_fp16 = layer_norm(axes = x_35_axes_0, beta = norm17_1_bias_to_fp16, epsilon = var_7327_to_fp16, gamma = norm17_1_weight_to_fp16, x = input_173_cast_fp16)[name = tensor("x_35_cast_fp16")]; + tensor linear_69_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = attn17_in_proj_weight_cast_fp16, x = x_35_cast_fp16)[name = tensor("linear_69_cast_fp16")]; + tensor var_7363 = const()[name = tensor("op_7363"), val = tensor([1, 1, 3, 16, 64])]; + tensor qkv_35_cast_fp16 = reshape(shape = var_7363, x = linear_69_cast_fp16)[name = tensor("qkv_35_cast_fp16")]; + tensor q_103_begin_0 = const()[name = tensor("q_103_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor q_103_end_0 = const()[name = tensor("q_103_end_0"), val = tensor([1, 1, 1, 16, 64])]; + tensor q_103_end_mask_0 = const()[name = tensor("q_103_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor q_103_squeeze_mask_0 = const()[name = tensor("q_103_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor q_103_cast_fp16 = slice_by_index(begin = q_103_begin_0, end = q_103_end_0, end_mask = q_103_end_mask_0, squeeze_mask = q_103_squeeze_mask_0, x = qkv_35_cast_fp16)[name = tensor("q_103_cast_fp16")]; + tensor k_69_begin_0 = const()[name = tensor("k_69_begin_0"), val = tensor([0, 0, 1, 0, 0])]; + tensor k_69_end_0 = const()[name = tensor("k_69_end_0"), val = tensor([1, 1, 2, 16, 64])]; + tensor k_69_end_mask_0 = const()[name = tensor("k_69_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor k_69_squeeze_mask_0 = const()[name = tensor("k_69_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor k_69_cast_fp16 = slice_by_index(begin = k_69_begin_0, end = k_69_end_0, end_mask = k_69_end_mask_0, squeeze_mask = k_69_squeeze_mask_0, x = qkv_35_cast_fp16)[name = tensor("k_69_cast_fp16")]; + tensor v_35_begin_0 = const()[name = tensor("v_35_begin_0"), val = tensor([0, 0, 2, 0, 0])]; + tensor v_35_end_0 = const()[name = tensor("v_35_end_0"), val = tensor([1, 1, 3, 16, 64])]; + tensor v_35_end_mask_0 = const()[name = tensor("v_35_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor v_35_squeeze_mask_0 = const()[name = tensor("v_35_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor v_35_cast_fp16 = slice_by_index(begin = v_35_begin_0, end = v_35_end_0, end_mask = v_35_end_mask_0, squeeze_mask = v_35_squeeze_mask_0, x = qkv_35_cast_fp16)[name = tensor("v_35_cast_fp16")]; + tensor var_7467 = const()[name = tensor("op_7467"), val = tensor([1, 1, 1, 1])]; + tensor position17_to_fp16_dtype_0 = const()[name = tensor("position17_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor position17_to_fp16 = cast(dtype = position17_to_fp16_dtype_0, x = position17)[name = tensor("cast_468")]; + tensor ts_107_cast_fp16 = reshape(shape = var_7467, x = position17_to_fp16)[name = tensor("ts_107_cast_fp16")]; + tensor var_7471 = const()[name = tensor("op_7471"), val = tensor([1, 1, 16, 32, 2])]; + tensor q_complex_35_cast_fp16 = reshape(shape = var_7471, x = q_103_cast_fp16)[name = tensor("q_complex_35_cast_fp16")]; + tensor var_7475 = const()[name = tensor("op_7475"), val = tensor([1, 1, 16, 32, 2])]; + tensor k_complex_35_cast_fp16 = reshape(shape = var_7475, x = k_69_cast_fp16)[name = tensor("k_complex_35_cast_fp16")]; + tensor var_7479_begin_0 = const()[name = tensor("op_7479_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_7479_end_0 = const()[name = tensor("op_7479_end_0"), val = tensor([1, 1, 16, 32, 1])]; + tensor var_7479_end_mask_0 = const()[name = tensor("op_7479_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_7479_squeeze_mask_0 = const()[name = tensor("op_7479_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_7479_cast_fp16 = slice_by_index(begin = var_7479_begin_0, end = var_7479_end_0, end_mask = var_7479_end_mask_0, squeeze_mask = var_7479_squeeze_mask_0, x = q_complex_35_cast_fp16)[name = tensor("op_7479_cast_fp16")]; + tensor var_7487_begin_0 = const()[name = tensor("op_7487_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_7487_end_0 = const()[name = tensor("op_7487_end_0"), val = tensor([1, 1, 16, 32, 2])]; + tensor var_7487_end_mask_0 = const()[name = tensor("op_7487_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_7487_squeeze_mask_0 = const()[name = tensor("op_7487_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_7487_cast_fp16 = slice_by_index(begin = var_7487_begin_0, end = var_7487_end_0, end_mask = var_7487_end_mask_0, squeeze_mask = var_7487_squeeze_mask_0, x = q_complex_35_cast_fp16)[name = tensor("op_7487_cast_fp16")]; + tensor var_7495_begin_0 = const()[name = tensor("op_7495_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_7495_end_0 = const()[name = tensor("op_7495_end_0"), val = tensor([1, 1, 16, 32, 1])]; + tensor var_7495_end_mask_0 = const()[name = tensor("op_7495_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_7495_squeeze_mask_0 = const()[name = tensor("op_7495_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_7495_cast_fp16 = slice_by_index(begin = var_7495_begin_0, end = var_7495_end_0, end_mask = var_7495_end_mask_0, squeeze_mask = var_7495_squeeze_mask_0, x = k_complex_35_cast_fp16)[name = tensor("op_7495_cast_fp16")]; + tensor var_7503_begin_0 = const()[name = tensor("op_7503_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_7503_end_0 = const()[name = tensor("op_7503_end_0"), val = tensor([1, 1, 16, 32, 2])]; + tensor var_7503_end_mask_0 = const()[name = tensor("op_7503_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_7503_squeeze_mask_0 = const()[name = tensor("op_7503_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_7503_cast_fp16 = slice_by_index(begin = var_7503_begin_0, end = var_7503_end_0, end_mask = var_7503_end_mask_0, squeeze_mask = var_7503_squeeze_mask_0, x = k_complex_35_cast_fp16)[name = tensor("op_7503_cast_fp16")]; + tensor freqs_35_to_fp16 = const()[name = tensor("freqs_35_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(303734912)))]; + tensor var_7509_cast_fp16 = mul(x = freqs_35_to_fp16, y = ts_107_cast_fp16)[name = tensor("op_7509_cast_fp16")]; + tensor rotr_35_cast_fp16 = cos(x = var_7509_cast_fp16)[name = tensor("rotr_35_cast_fp16")]; + tensor roti_35_cast_fp16 = sin(x = var_7509_cast_fp16)[name = tensor("roti_35_cast_fp16")]; + tensor var_7513_cast_fp16 = mul(x = var_7479_cast_fp16, y = rotr_35_cast_fp16)[name = tensor("op_7513_cast_fp16")]; + tensor var_7514_cast_fp16 = mul(x = var_7487_cast_fp16, y = roti_35_cast_fp16)[name = tensor("op_7514_cast_fp16")]; + tensor qor_69_cast_fp16 = sub(x = var_7513_cast_fp16, y = var_7514_cast_fp16)[name = tensor("qor_69_cast_fp16")]; + tensor var_7517_cast_fp16 = mul(x = var_7479_cast_fp16, y = roti_35_cast_fp16)[name = tensor("op_7517_cast_fp16")]; + tensor var_7518_cast_fp16 = mul(x = var_7487_cast_fp16, y = rotr_35_cast_fp16)[name = tensor("op_7518_cast_fp16")]; + tensor qoi_69_cast_fp16 = add(x = var_7517_cast_fp16, y = var_7518_cast_fp16)[name = tensor("qoi_69_cast_fp16")]; + tensor var_7521_cast_fp16 = mul(x = var_7495_cast_fp16, y = rotr_35_cast_fp16)[name = tensor("op_7521_cast_fp16")]; + tensor var_7522_cast_fp16 = mul(x = var_7503_cast_fp16, y = roti_35_cast_fp16)[name = tensor("op_7522_cast_fp16")]; + tensor kor_69_cast_fp16 = sub(x = var_7521_cast_fp16, y = var_7522_cast_fp16)[name = tensor("kor_69_cast_fp16")]; + tensor var_7525_cast_fp16 = mul(x = var_7495_cast_fp16, y = roti_35_cast_fp16)[name = tensor("op_7525_cast_fp16")]; + tensor var_7526_cast_fp16 = mul(x = var_7503_cast_fp16, y = rotr_35_cast_fp16)[name = tensor("op_7526_cast_fp16")]; + tensor koi_69_cast_fp16 = add(x = var_7525_cast_fp16, y = var_7526_cast_fp16)[name = tensor("koi_69_cast_fp16")]; + tensor qo_35_axis_0 = const()[name = tensor("qo_35_axis_0"), val = tensor(-1)]; + tensor qo_35_cast_fp16 = stack(axis = qo_35_axis_0, values = (qor_69_cast_fp16, qoi_69_cast_fp16))[name = tensor("qo_35_cast_fp16")]; + tensor ko_35_axis_0 = const()[name = tensor("ko_35_axis_0"), val = tensor(-1)]; + tensor ko_35_cast_fp16 = stack(axis = ko_35_axis_0, values = (kor_69_cast_fp16, koi_69_cast_fp16))[name = tensor("ko_35_cast_fp16")]; + tensor var_7555 = const()[name = tensor("op_7555"), val = tensor([1, 1, 16, 64])]; + tensor q_105_cast_fp16 = reshape(shape = var_7555, x = qo_35_cast_fp16)[name = tensor("q_105_cast_fp16")]; + tensor var_7557 = const()[name = tensor("op_7557"), val = tensor([1, 1, 16, 64])]; + tensor k_71_cast_fp16 = reshape(shape = var_7557, x = ko_35_cast_fp16)[name = tensor("k_71_cast_fp16")]; + tensor _inversed_7579_y_0_to_fp16 = const()[name = tensor("_inversed_7579_y_0_to_fp16"), val = tensor(0x1p-9)]; + tensor _inversed_7579_cast_fp16 = mul(x = ts_107_cast_fp16, y = _inversed_7579_y_0_to_fp16)[name = tensor("_inversed_7579_cast_fp16")]; + tensor var_7580_cast_fp16 = floor(x = _inversed_7579_cast_fp16)[name = tensor("op_7580_cast_fp16")]; + tensor var_7581_to_fp16 = const()[name = tensor("op_7581_to_fp16"), val = tensor(0x1p+9)]; + tensor var_7582_cast_fp16 = mul(x = var_7580_cast_fp16, y = var_7581_to_fp16)[name = tensor("op_7582_cast_fp16")]; + tensor write_indices_float_71_cast_fp16 = sub(x = ts_107_cast_fp16, y = var_7582_cast_fp16)[name = tensor("write_indices_float_71_cast_fp16")]; + tensor var_7589_dtype_0 = const()[name = tensor("op_7589_dtype_0"), val = tensor("int32")]; + tensor write_indices_35_reps_0 = const()[name = tensor("write_indices_35_reps_0"), val = tensor([1, 1, 16, 64])]; + tensor write_indices_float_71_cast_fp16_to_int32 = cast(dtype = var_7589_dtype_0, x = write_indices_float_71_cast_fp16)[name = tensor("cast_467")]; + tensor write_indices_35 = tile(reps = write_indices_35_reps_0, x = write_indices_float_71_cast_fp16_to_int32)[name = tensor("write_indices_35")]; + tensor var_7597_begin_0 = const()[name = tensor("op_7597_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_7597_end_0 = const()[name = tensor("op_7597_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor var_7597_end_mask_0 = const()[name = tensor("op_7597_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_7597_squeeze_mask_0 = const()[name = tensor("op_7597_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor cache17_to_fp16_dtype_0 = const()[name = tensor("cache17_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor cache17_to_fp16 = cast(dtype = cache17_to_fp16_dtype_0, x = cache17)[name = tensor("cast_466")]; + tensor var_7597_cast_fp16 = slice_by_index(begin = var_7597_begin_0, end = var_7597_end_0, end_mask = var_7597_end_mask_0, squeeze_mask = var_7597_squeeze_mask_0, x = cache17_to_fp16)[name = tensor("op_7597_cast_fp16")]; + tensor var_7599_axis_0 = const()[name = tensor("op_7599_axis_0"), val = tensor(1)]; + tensor var_7599_mode_0 = const()[name = tensor("op_7599_mode_0"), val = tensor("update")]; + tensor var_7599_validate_indices_0 = const()[name = tensor("op_7599_validate_indices_0"), val = tensor(false)]; + tensor var_7599_cast_fp16 = scatter_along_axis(axis = var_7599_axis_0, data = var_7597_cast_fp16, indices = write_indices_35, mode = var_7599_mode_0, updates = k_71_cast_fp16, validate_indices = var_7599_validate_indices_0)[name = tensor("op_7599_cast_fp16")]; + tensor concat_121 = const()[name = tensor("concat_121"), val = tensor([0, 0, 0, 0, 0])]; + tensor concat_122 = const()[name = tensor("concat_122"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_35_internal_tensor_assign_1_stride_0 = const()[name = tensor("new_cache_35_internal_tensor_assign_1_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_35_internal_tensor_assign_1_begin_mask_0 = const()[name = tensor("new_cache_35_internal_tensor_assign_1_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_35_internal_tensor_assign_1_end_mask_0 = const()[name = tensor("new_cache_35_internal_tensor_assign_1_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_35_internal_tensor_assign_1_squeeze_mask_0 = const()[name = tensor("new_cache_35_internal_tensor_assign_1_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_82 = const()[name = tensor("shape_82"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_34 = const()[name = tensor("reduce_prod_34"), val = tensor(1048576)]; + tensor range_1d_34_start_0 = const()[name = tensor("range_1d_34_start_0"), val = tensor(0)]; + tensor range_1d_34_step_0 = const()[name = tensor("range_1d_34_step_0"), val = tensor(1)]; + tensor range_1d_34 = range_1d(end = reduce_prod_34, start = range_1d_34_start_0, step = range_1d_34_step_0)[name = tensor("range_1d_34")]; + tensor reshape_170 = reshape(shape = shape_82, x = range_1d_34)[name = tensor("reshape_170")]; + tensor slice_by_index_34 = slice_by_index(begin = concat_121, begin_mask = new_cache_35_internal_tensor_assign_1_begin_mask_0, end = concat_122, end_mask = new_cache_35_internal_tensor_assign_1_end_mask_0, squeeze_mask = new_cache_35_internal_tensor_assign_1_squeeze_mask_0, stride = new_cache_35_internal_tensor_assign_1_stride_0, x = reshape_170)[name = tensor("slice_by_index_34")]; + tensor reshape_171_shape_0 = const()[name = tensor("reshape_171_shape_0"), val = tensor([-1])]; + tensor reshape_171 = reshape(shape = reshape_171_shape_0, x = slice_by_index_34)[name = tensor("reshape_171")]; + tensor reshape_172_shape_0 = const()[name = tensor("reshape_172_shape_0"), val = tensor([-1])]; + tensor reshape_172_cast_fp16 = reshape(shape = reshape_172_shape_0, x = var_7599_cast_fp16)[name = tensor("reshape_172_cast_fp16")]; + tensor reshape_173_shape_0 = const()[name = tensor("reshape_173_shape_0"), val = tensor([-1])]; + tensor reshape_173_cast_fp16 = reshape(shape = reshape_173_shape_0, x = cache17_to_fp16)[name = tensor("reshape_173_cast_fp16")]; + tensor scatter_34_mode_0 = const()[name = tensor("scatter_34_mode_0"), val = tensor("update")]; + tensor scatter_34_axis_0 = const()[name = tensor("scatter_34_axis_0"), val = tensor(0)]; + tensor scatter_34_validate_indices_0 = const()[name = tensor("scatter_34_validate_indices_0"), val = tensor(false)]; + tensor scatter_34_cast_fp16 = scatter(axis = scatter_34_axis_0, data = reshape_173_cast_fp16, indices = reshape_171, mode = scatter_34_mode_0, updates = reshape_172_cast_fp16, validate_indices = scatter_34_validate_indices_0)[name = tensor("scatter_34_cast_fp16")]; + tensor reshape_174_cast_fp16 = reshape(shape = shape_82, x = scatter_34_cast_fp16)[name = tensor("reshape_174_cast_fp16")]; + tensor var_7607_begin_0 = const()[name = tensor("op_7607_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor var_7607_end_0 = const()[name = tensor("op_7607_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor var_7607_end_mask_0 = const()[name = tensor("op_7607_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_7607_squeeze_mask_0 = const()[name = tensor("op_7607_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor var_7607_cast_fp16 = slice_by_index(begin = var_7607_begin_0, end = var_7607_end_0, end_mask = var_7607_end_mask_0, squeeze_mask = var_7607_squeeze_mask_0, x = reshape_174_cast_fp16)[name = tensor("op_7607_cast_fp16")]; + tensor var_7609_axis_0 = const()[name = tensor("op_7609_axis_0"), val = tensor(1)]; + tensor var_7609_mode_0 = const()[name = tensor("op_7609_mode_0"), val = tensor("update")]; + tensor var_7609_validate_indices_0 = const()[name = tensor("op_7609_validate_indices_0"), val = tensor(false)]; + tensor var_7609_cast_fp16 = scatter_along_axis(axis = var_7609_axis_0, data = var_7607_cast_fp16, indices = write_indices_35, mode = var_7609_mode_0, updates = v_35_cast_fp16, validate_indices = var_7609_validate_indices_0)[name = tensor("op_7609_cast_fp16")]; + tensor concat_123 = const()[name = tensor("concat_123"), val = tensor([1, 0, 0, 0, 0])]; + tensor concat_124 = const()[name = tensor("concat_124"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_35_internal_tensor_assign_2_stride_0 = const()[name = tensor("new_cache_35_internal_tensor_assign_2_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_35_internal_tensor_assign_2_begin_mask_0 = const()[name = tensor("new_cache_35_internal_tensor_assign_2_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_35_internal_tensor_assign_2_end_mask_0 = const()[name = tensor("new_cache_35_internal_tensor_assign_2_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_35_internal_tensor_assign_2_squeeze_mask_0 = const()[name = tensor("new_cache_35_internal_tensor_assign_2_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_83 = const()[name = tensor("shape_83"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_35 = const()[name = tensor("reduce_prod_35"), val = tensor(1048576)]; + tensor range_1d_35_start_0 = const()[name = tensor("range_1d_35_start_0"), val = tensor(0)]; + tensor range_1d_35_step_0 = const()[name = tensor("range_1d_35_step_0"), val = tensor(1)]; + tensor range_1d_35 = range_1d(end = reduce_prod_35, start = range_1d_35_start_0, step = range_1d_35_step_0)[name = tensor("range_1d_35")]; + tensor reshape_175 = reshape(shape = shape_83, x = range_1d_35)[name = tensor("reshape_175")]; + tensor slice_by_index_35 = slice_by_index(begin = concat_123, begin_mask = new_cache_35_internal_tensor_assign_2_begin_mask_0, end = concat_124, end_mask = new_cache_35_internal_tensor_assign_2_end_mask_0, squeeze_mask = new_cache_35_internal_tensor_assign_2_squeeze_mask_0, stride = new_cache_35_internal_tensor_assign_2_stride_0, x = reshape_175)[name = tensor("slice_by_index_35")]; + tensor reshape_176_shape_0 = const()[name = tensor("reshape_176_shape_0"), val = tensor([-1])]; + tensor reshape_176 = reshape(shape = reshape_176_shape_0, x = slice_by_index_35)[name = tensor("reshape_176")]; + tensor reshape_177_shape_0 = const()[name = tensor("reshape_177_shape_0"), val = tensor([-1])]; + tensor reshape_177_cast_fp16 = reshape(shape = reshape_177_shape_0, x = var_7609_cast_fp16)[name = tensor("reshape_177_cast_fp16")]; + tensor reshape_178_shape_0 = const()[name = tensor("reshape_178_shape_0"), val = tensor([-1])]; + tensor reshape_178_cast_fp16 = reshape(shape = reshape_178_shape_0, x = reshape_174_cast_fp16)[name = tensor("reshape_178_cast_fp16")]; + tensor scatter_35_mode_0 = const()[name = tensor("scatter_35_mode_0"), val = tensor("update")]; + tensor scatter_35_axis_0 = const()[name = tensor("scatter_35_axis_0"), val = tensor(0)]; + tensor scatter_35_validate_indices_0 = const()[name = tensor("scatter_35_validate_indices_0"), val = tensor(false)]; + tensor scatter_35_cast_fp16 = scatter(axis = scatter_35_axis_0, data = reshape_178_cast_fp16, indices = reshape_176, mode = scatter_35_mode_0, updates = reshape_177_cast_fp16, validate_indices = scatter_35_validate_indices_0)[name = tensor("scatter_35_cast_fp16")]; + tensor reshape_179_cast_fp16 = reshape(shape = shape_83, x = scatter_35_cast_fp16)[name = tensor("reshape_179_cast_fp16")]; + tensor reshape_179_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("reshape_179_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor keys_103_begin_0 = const()[name = tensor("keys_103_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor keys_103_end_0 = const()[name = tensor("keys_103_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor keys_103_end_mask_0 = const()[name = tensor("keys_103_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor keys_103_squeeze_mask_0 = const()[name = tensor("keys_103_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor keys_103_cast_fp16 = slice_by_index(begin = keys_103_begin_0, end = keys_103_end_0, end_mask = keys_103_end_mask_0, squeeze_mask = keys_103_squeeze_mask_0, x = reshape_179_cast_fp16)[name = tensor("keys_103_cast_fp16")]; + tensor values_103_begin_0 = const()[name = tensor("values_103_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor values_103_end_0 = const()[name = tensor("values_103_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor values_103_end_mask_0 = const()[name = tensor("values_103_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor values_103_squeeze_mask_0 = const()[name = tensor("values_103_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor values_103_cast_fp16 = slice_by_index(begin = values_103_begin_0, end = values_103_end_0, end_mask = values_103_end_mask_0, squeeze_mask = values_103_squeeze_mask_0, x = reshape_179_cast_fp16)[name = tensor("values_103_cast_fp16")]; + tensor var_7621_cast_fp16 = not_equal(x = keys_103_cast_fp16, y = keys_103_cast_fp16)[name = tensor("op_7621_cast_fp16")]; + tensor keys_105_cast_fp16 = select(a = var_504_to_fp16, b = keys_103_cast_fp16, cond = var_7621_cast_fp16)[name = tensor("keys_105_cast_fp16")]; + tensor var_7629_cast_fp16 = not_equal(x = values_103_cast_fp16, y = values_103_cast_fp16)[name = tensor("op_7629_cast_fp16")]; + tensor values_105_cast_fp16 = select(a = var_504_to_fp16, b = values_103_cast_fp16, cond = var_7629_cast_fp16)[name = tensor("values_105_cast_fp16")]; + tensor var_7653 = const()[name = tensor("op_7653"), val = tensor([0, 2, 1, 3])]; + tensor var_7666 = const()[name = tensor("op_7666"), val = tensor([1, 1, 1])]; + tensor var_7667_cast_fp16 = reshape(shape = var_7666, x = position17_to_fp16)[name = tensor("op_7667_cast_fp16")]; + tensor var_7684_to_fp16 = const()[name = tensor("op_7684_to_fp16"), val = tensor(0x1p+0)]; + tensor valid_len_35_cast_fp16 = add(x = var_7667_cast_fp16, y = var_7684_to_fp16)[name = tensor("valid_len_35_cast_fp16")]; + tensor valid_mask_35_cast_fp16 = less(x = k_positions_1_promoted_to_fp16, y = valid_len_35_cast_fp16)[name = tensor("valid_mask_35_cast_fp16")]; + tensor causal_mask_35_cast_fp16 = less_equal(x = k_positions_1_promoted_to_fp16, y = var_7667_cast_fp16)[name = tensor("causal_mask_35_cast_fp16")]; + tensor attn_mask_69 = logical_and(x = valid_mask_35_cast_fp16, y = causal_mask_35_cast_fp16)[name = tensor("attn_mask_69")]; + tensor attn_mask_71_axes_0 = const()[name = tensor("attn_mask_71_axes_0"), val = tensor([1])]; + tensor attn_mask_71 = expand_dims(axes = attn_mask_71_axes_0, x = attn_mask_69)[name = tensor("attn_mask_71")]; + tensor var_7702_transpose_x_0 = const()[name = tensor("op_7702_transpose_x_0"), val = tensor(false)]; + tensor var_7702_transpose_y_0 = const()[name = tensor("op_7702_transpose_y_0"), val = tensor(false)]; + tensor transpose_106_perm_0 = const()[name = tensor("transpose_106_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_107_perm_0 = const()[name = tensor("transpose_107_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_107 = transpose(perm = transpose_107_perm_0, x = keys_105_cast_fp16)[name = tensor("transpose_145")]; + tensor transpose_106 = transpose(perm = transpose_106_perm_0, x = q_105_cast_fp16)[name = tensor("transpose_146")]; + tensor var_7702_cast_fp16 = matmul(transpose_x = var_7702_transpose_x_0, transpose_y = var_7702_transpose_y_0, x = transpose_106, y = transpose_107)[name = tensor("op_7702_cast_fp16")]; + tensor var_7696_to_fp16 = const()[name = tensor("op_7696_to_fp16"), val = tensor([0x1p-3])]; + tensor attn_weights_103_cast_fp16 = mul(x = var_7702_cast_fp16, y = var_7696_to_fp16)[name = tensor("attn_weights_103_cast_fp16")]; + tensor var_7704 = logical_not(x = attn_mask_71)[name = tensor("op_7704")]; + tensor var_7705_to_fp16 = const()[name = tensor("op_7705_to_fp16"), val = tensor(-inf)]; + tensor attn_weights_105_cast_fp16 = select(a = var_7705_to_fp16, b = attn_weights_103_cast_fp16, cond = var_7704)[name = tensor("attn_weights_105_cast_fp16")]; + tensor var_7707 = const()[name = tensor("op_7707"), val = tensor(-1)]; + tensor attn_weights_107_cast_fp16 = softmax(axis = var_7707, x = attn_weights_105_cast_fp16)[name = tensor("attn_weights_107_cast_fp16")]; + tensor attn_output_35_transpose_x_0 = const()[name = tensor("attn_output_35_transpose_x_0"), val = tensor(false)]; + tensor attn_output_35_transpose_y_0 = const()[name = tensor("attn_output_35_transpose_y_0"), val = tensor(false)]; + tensor values_107_cast_fp16 = transpose(perm = var_7653, x = values_105_cast_fp16)[name = tensor("transpose_147")]; + tensor attn_output_35_cast_fp16 = matmul(transpose_x = attn_output_35_transpose_x_0, transpose_y = attn_output_35_transpose_y_0, x = attn_weights_107_cast_fp16, y = values_107_cast_fp16)[name = tensor("attn_output_35_cast_fp16")]; + tensor var_7715 = const()[name = tensor("op_7715"), val = tensor([0, 2, 1, 3])]; + tensor var_7718 = const()[name = tensor("op_7718"), val = tensor([1, 1, 1024])]; + tensor var_7716_cast_fp16 = transpose(perm = var_7715, x = attn_output_35_cast_fp16)[name = tensor("transpose_144")]; + tensor input_175_cast_fp16 = reshape(shape = var_7718, x = var_7716_cast_fp16)[name = tensor("input_175_cast_fp16")]; + tensor linear_70_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = attn17_out_proj_weight_cast_fp16, x = input_175_cast_fp16)[name = tensor("linear_70_cast_fp16")]; + tensor var_7724_to_fp16 = const()[name = tensor("op_7724_to_fp16"), val = tensor(0x1p+0)]; + tensor var_7725_cast_fp16 = add(x = position17_to_fp16, y = var_7724_to_fp16)[name = tensor("op_7725_cast_fp16")]; + tensor var_7725_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_7725_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor input_177_cast_fp16 = add(x = input_173_cast_fp16, y = linear_70_cast_fp16)[name = tensor("input_177_cast_fp16")]; + tensor input_179_axes_0 = const()[name = tensor("input_179_axes_0"), val = tensor([-1])]; + tensor norm17_2_weight_to_fp16 = const()[name = tensor("norm17_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(303735040)))]; + tensor norm17_2_bias_to_fp16 = const()[name = tensor("norm17_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(303737152)))]; + tensor var_7729_to_fp16 = const()[name = tensor("op_7729_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_179_cast_fp16 = layer_norm(axes = input_179_axes_0, beta = norm17_2_bias_to_fp16, epsilon = var_7729_to_fp16, gamma = norm17_2_weight_to_fp16, x = input_177_cast_fp16)[name = tensor("input_179_cast_fp16")]; + tensor linear_71_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = linear17_1_weight_cast_fp16, x = input_179_cast_fp16)[name = tensor("linear_71_cast_fp16")]; + tensor input_181_mode_0 = const()[name = tensor("input_181_mode_0"), val = tensor("EXACT")]; + tensor input_181_cast_fp16 = gelu(mode = input_181_mode_0, x = linear_71_cast_fp16)[name = tensor("input_181_cast_fp16")]; + tensor linear_72_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = linear17_2_weight_cast_fp16, x = input_181_cast_fp16)[name = tensor("linear_72_cast_fp16")]; + tensor input_183_cast_fp16 = add(x = input_177_cast_fp16, y = linear_72_cast_fp16)[name = tensor("input_183_cast_fp16")]; + tensor x_37_axes_0 = const()[name = tensor("x_37_axes_0"), val = tensor([-1])]; + tensor norm18_1_weight_to_fp16 = const()[name = tensor("norm18_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(303739264)))]; + tensor norm18_1_bias_to_fp16 = const()[name = tensor("norm18_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(303741376)))]; + tensor var_7746_to_fp16 = const()[name = tensor("op_7746_to_fp16"), val = tensor(0x1.5p-17)]; + tensor x_37_cast_fp16 = layer_norm(axes = x_37_axes_0, beta = norm18_1_bias_to_fp16, epsilon = var_7746_to_fp16, gamma = norm18_1_weight_to_fp16, x = input_183_cast_fp16)[name = tensor("x_37_cast_fp16")]; + tensor linear_73_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = attn18_in_proj_weight_cast_fp16, x = x_37_cast_fp16)[name = tensor("linear_73_cast_fp16")]; + tensor var_7782 = const()[name = tensor("op_7782"), val = tensor([1, 1, 3, 16, 64])]; + tensor qkv_37_cast_fp16 = reshape(shape = var_7782, x = linear_73_cast_fp16)[name = tensor("qkv_37_cast_fp16")]; + tensor q_109_begin_0 = const()[name = tensor("q_109_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor q_109_end_0 = const()[name = tensor("q_109_end_0"), val = tensor([1, 1, 1, 16, 64])]; + tensor q_109_end_mask_0 = const()[name = tensor("q_109_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor q_109_squeeze_mask_0 = const()[name = tensor("q_109_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor q_109_cast_fp16 = slice_by_index(begin = q_109_begin_0, end = q_109_end_0, end_mask = q_109_end_mask_0, squeeze_mask = q_109_squeeze_mask_0, x = qkv_37_cast_fp16)[name = tensor("q_109_cast_fp16")]; + tensor k_73_begin_0 = const()[name = tensor("k_73_begin_0"), val = tensor([0, 0, 1, 0, 0])]; + tensor k_73_end_0 = const()[name = tensor("k_73_end_0"), val = tensor([1, 1, 2, 16, 64])]; + tensor k_73_end_mask_0 = const()[name = tensor("k_73_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor k_73_squeeze_mask_0 = const()[name = tensor("k_73_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor k_73_cast_fp16 = slice_by_index(begin = k_73_begin_0, end = k_73_end_0, end_mask = k_73_end_mask_0, squeeze_mask = k_73_squeeze_mask_0, x = qkv_37_cast_fp16)[name = tensor("k_73_cast_fp16")]; + tensor v_37_begin_0 = const()[name = tensor("v_37_begin_0"), val = tensor([0, 0, 2, 0, 0])]; + tensor v_37_end_0 = const()[name = tensor("v_37_end_0"), val = tensor([1, 1, 3, 16, 64])]; + tensor v_37_end_mask_0 = const()[name = tensor("v_37_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor v_37_squeeze_mask_0 = const()[name = tensor("v_37_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor v_37_cast_fp16 = slice_by_index(begin = v_37_begin_0, end = v_37_end_0, end_mask = v_37_end_mask_0, squeeze_mask = v_37_squeeze_mask_0, x = qkv_37_cast_fp16)[name = tensor("v_37_cast_fp16")]; + tensor var_7886 = const()[name = tensor("op_7886"), val = tensor([1, 1, 1, 1])]; + tensor position18_to_fp16_dtype_0 = const()[name = tensor("position18_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor position18_to_fp16 = cast(dtype = position18_to_fp16_dtype_0, x = position18)[name = tensor("cast_463")]; + tensor ts_113_cast_fp16 = reshape(shape = var_7886, x = position18_to_fp16)[name = tensor("ts_113_cast_fp16")]; + tensor var_7890 = const()[name = tensor("op_7890"), val = tensor([1, 1, 16, 32, 2])]; + tensor q_complex_37_cast_fp16 = reshape(shape = var_7890, x = q_109_cast_fp16)[name = tensor("q_complex_37_cast_fp16")]; + tensor var_7894 = const()[name = tensor("op_7894"), val = tensor([1, 1, 16, 32, 2])]; + tensor k_complex_37_cast_fp16 = reshape(shape = var_7894, x = k_73_cast_fp16)[name = tensor("k_complex_37_cast_fp16")]; + tensor var_7898_begin_0 = const()[name = tensor("op_7898_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_7898_end_0 = const()[name = tensor("op_7898_end_0"), val = tensor([1, 1, 16, 32, 1])]; + tensor var_7898_end_mask_0 = const()[name = tensor("op_7898_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_7898_squeeze_mask_0 = const()[name = tensor("op_7898_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_7898_cast_fp16 = slice_by_index(begin = var_7898_begin_0, end = var_7898_end_0, end_mask = var_7898_end_mask_0, squeeze_mask = var_7898_squeeze_mask_0, x = q_complex_37_cast_fp16)[name = tensor("op_7898_cast_fp16")]; + tensor var_7906_begin_0 = const()[name = tensor("op_7906_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_7906_end_0 = const()[name = tensor("op_7906_end_0"), val = tensor([1, 1, 16, 32, 2])]; + tensor var_7906_end_mask_0 = const()[name = tensor("op_7906_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_7906_squeeze_mask_0 = const()[name = tensor("op_7906_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_7906_cast_fp16 = slice_by_index(begin = var_7906_begin_0, end = var_7906_end_0, end_mask = var_7906_end_mask_0, squeeze_mask = var_7906_squeeze_mask_0, x = q_complex_37_cast_fp16)[name = tensor("op_7906_cast_fp16")]; + tensor var_7914_begin_0 = const()[name = tensor("op_7914_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_7914_end_0 = const()[name = tensor("op_7914_end_0"), val = tensor([1, 1, 16, 32, 1])]; + tensor var_7914_end_mask_0 = const()[name = tensor("op_7914_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_7914_squeeze_mask_0 = const()[name = tensor("op_7914_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_7914_cast_fp16 = slice_by_index(begin = var_7914_begin_0, end = var_7914_end_0, end_mask = var_7914_end_mask_0, squeeze_mask = var_7914_squeeze_mask_0, x = k_complex_37_cast_fp16)[name = tensor("op_7914_cast_fp16")]; + tensor var_7922_begin_0 = const()[name = tensor("op_7922_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_7922_end_0 = const()[name = tensor("op_7922_end_0"), val = tensor([1, 1, 16, 32, 2])]; + tensor var_7922_end_mask_0 = const()[name = tensor("op_7922_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_7922_squeeze_mask_0 = const()[name = tensor("op_7922_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_7922_cast_fp16 = slice_by_index(begin = var_7922_begin_0, end = var_7922_end_0, end_mask = var_7922_end_mask_0, squeeze_mask = var_7922_squeeze_mask_0, x = k_complex_37_cast_fp16)[name = tensor("op_7922_cast_fp16")]; + tensor freqs_37_to_fp16 = const()[name = tensor("freqs_37_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(303743488)))]; + tensor var_7928_cast_fp16 = mul(x = freqs_37_to_fp16, y = ts_113_cast_fp16)[name = tensor("op_7928_cast_fp16")]; + tensor rotr_37_cast_fp16 = cos(x = var_7928_cast_fp16)[name = tensor("rotr_37_cast_fp16")]; + tensor roti_37_cast_fp16 = sin(x = var_7928_cast_fp16)[name = tensor("roti_37_cast_fp16")]; + tensor var_7932_cast_fp16 = mul(x = var_7898_cast_fp16, y = rotr_37_cast_fp16)[name = tensor("op_7932_cast_fp16")]; + tensor var_7933_cast_fp16 = mul(x = var_7906_cast_fp16, y = roti_37_cast_fp16)[name = tensor("op_7933_cast_fp16")]; + tensor qor_73_cast_fp16 = sub(x = var_7932_cast_fp16, y = var_7933_cast_fp16)[name = tensor("qor_73_cast_fp16")]; + tensor var_7936_cast_fp16 = mul(x = var_7898_cast_fp16, y = roti_37_cast_fp16)[name = tensor("op_7936_cast_fp16")]; + tensor var_7937_cast_fp16 = mul(x = var_7906_cast_fp16, y = rotr_37_cast_fp16)[name = tensor("op_7937_cast_fp16")]; + tensor qoi_73_cast_fp16 = add(x = var_7936_cast_fp16, y = var_7937_cast_fp16)[name = tensor("qoi_73_cast_fp16")]; + tensor var_7940_cast_fp16 = mul(x = var_7914_cast_fp16, y = rotr_37_cast_fp16)[name = tensor("op_7940_cast_fp16")]; + tensor var_7941_cast_fp16 = mul(x = var_7922_cast_fp16, y = roti_37_cast_fp16)[name = tensor("op_7941_cast_fp16")]; + tensor kor_73_cast_fp16 = sub(x = var_7940_cast_fp16, y = var_7941_cast_fp16)[name = tensor("kor_73_cast_fp16")]; + tensor var_7944_cast_fp16 = mul(x = var_7914_cast_fp16, y = roti_37_cast_fp16)[name = tensor("op_7944_cast_fp16")]; + tensor var_7945_cast_fp16 = mul(x = var_7922_cast_fp16, y = rotr_37_cast_fp16)[name = tensor("op_7945_cast_fp16")]; + tensor koi_73_cast_fp16 = add(x = var_7944_cast_fp16, y = var_7945_cast_fp16)[name = tensor("koi_73_cast_fp16")]; + tensor qo_37_axis_0 = const()[name = tensor("qo_37_axis_0"), val = tensor(-1)]; + tensor qo_37_cast_fp16 = stack(axis = qo_37_axis_0, values = (qor_73_cast_fp16, qoi_73_cast_fp16))[name = tensor("qo_37_cast_fp16")]; + tensor ko_37_axis_0 = const()[name = tensor("ko_37_axis_0"), val = tensor(-1)]; + tensor ko_37_cast_fp16 = stack(axis = ko_37_axis_0, values = (kor_73_cast_fp16, koi_73_cast_fp16))[name = tensor("ko_37_cast_fp16")]; + tensor var_7974 = const()[name = tensor("op_7974"), val = tensor([1, 1, 16, 64])]; + tensor q_111_cast_fp16 = reshape(shape = var_7974, x = qo_37_cast_fp16)[name = tensor("q_111_cast_fp16")]; + tensor var_7976 = const()[name = tensor("op_7976"), val = tensor([1, 1, 16, 64])]; + tensor k_75_cast_fp16 = reshape(shape = var_7976, x = ko_37_cast_fp16)[name = tensor("k_75_cast_fp16")]; + tensor _inversed_7998_y_0_to_fp16 = const()[name = tensor("_inversed_7998_y_0_to_fp16"), val = tensor(0x1p-9)]; + tensor _inversed_7998_cast_fp16 = mul(x = ts_113_cast_fp16, y = _inversed_7998_y_0_to_fp16)[name = tensor("_inversed_7998_cast_fp16")]; + tensor var_7999_cast_fp16 = floor(x = _inversed_7998_cast_fp16)[name = tensor("op_7999_cast_fp16")]; + tensor var_8000_to_fp16 = const()[name = tensor("op_8000_to_fp16"), val = tensor(0x1p+9)]; + tensor var_8001_cast_fp16 = mul(x = var_7999_cast_fp16, y = var_8000_to_fp16)[name = tensor("op_8001_cast_fp16")]; + tensor write_indices_float_75_cast_fp16 = sub(x = ts_113_cast_fp16, y = var_8001_cast_fp16)[name = tensor("write_indices_float_75_cast_fp16")]; + tensor var_8008_dtype_0 = const()[name = tensor("op_8008_dtype_0"), val = tensor("int32")]; + tensor write_indices_37_reps_0 = const()[name = tensor("write_indices_37_reps_0"), val = tensor([1, 1, 16, 64])]; + tensor write_indices_float_75_cast_fp16_to_int32 = cast(dtype = var_8008_dtype_0, x = write_indices_float_75_cast_fp16)[name = tensor("cast_462")]; + tensor write_indices_37 = tile(reps = write_indices_37_reps_0, x = write_indices_float_75_cast_fp16_to_int32)[name = tensor("write_indices_37")]; + tensor var_8016_begin_0 = const()[name = tensor("op_8016_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_8016_end_0 = const()[name = tensor("op_8016_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor var_8016_end_mask_0 = const()[name = tensor("op_8016_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_8016_squeeze_mask_0 = const()[name = tensor("op_8016_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor cache18_to_fp16_dtype_0 = const()[name = tensor("cache18_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor cache18_to_fp16 = cast(dtype = cache18_to_fp16_dtype_0, x = cache18)[name = tensor("cast_461")]; + tensor var_8016_cast_fp16 = slice_by_index(begin = var_8016_begin_0, end = var_8016_end_0, end_mask = var_8016_end_mask_0, squeeze_mask = var_8016_squeeze_mask_0, x = cache18_to_fp16)[name = tensor("op_8016_cast_fp16")]; + tensor var_8018_axis_0 = const()[name = tensor("op_8018_axis_0"), val = tensor(1)]; + tensor var_8018_mode_0 = const()[name = tensor("op_8018_mode_0"), val = tensor("update")]; + tensor var_8018_validate_indices_0 = const()[name = tensor("op_8018_validate_indices_0"), val = tensor(false)]; + tensor var_8018_cast_fp16 = scatter_along_axis(axis = var_8018_axis_0, data = var_8016_cast_fp16, indices = write_indices_37, mode = var_8018_mode_0, updates = k_75_cast_fp16, validate_indices = var_8018_validate_indices_0)[name = tensor("op_8018_cast_fp16")]; + tensor concat_128 = const()[name = tensor("concat_128"), val = tensor([0, 0, 0, 0, 0])]; + tensor concat_129 = const()[name = tensor("concat_129"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_37_internal_tensor_assign_1_stride_0 = const()[name = tensor("new_cache_37_internal_tensor_assign_1_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_37_internal_tensor_assign_1_begin_mask_0 = const()[name = tensor("new_cache_37_internal_tensor_assign_1_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_37_internal_tensor_assign_1_end_mask_0 = const()[name = tensor("new_cache_37_internal_tensor_assign_1_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_37_internal_tensor_assign_1_squeeze_mask_0 = const()[name = tensor("new_cache_37_internal_tensor_assign_1_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_84 = const()[name = tensor("shape_84"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_36 = const()[name = tensor("reduce_prod_36"), val = tensor(1048576)]; + tensor range_1d_36_start_0 = const()[name = tensor("range_1d_36_start_0"), val = tensor(0)]; + tensor range_1d_36_step_0 = const()[name = tensor("range_1d_36_step_0"), val = tensor(1)]; + tensor range_1d_36 = range_1d(end = reduce_prod_36, start = range_1d_36_start_0, step = range_1d_36_step_0)[name = tensor("range_1d_36")]; + tensor reshape_180 = reshape(shape = shape_84, x = range_1d_36)[name = tensor("reshape_180")]; + tensor slice_by_index_36 = slice_by_index(begin = concat_128, begin_mask = new_cache_37_internal_tensor_assign_1_begin_mask_0, end = concat_129, end_mask = new_cache_37_internal_tensor_assign_1_end_mask_0, squeeze_mask = new_cache_37_internal_tensor_assign_1_squeeze_mask_0, stride = new_cache_37_internal_tensor_assign_1_stride_0, x = reshape_180)[name = tensor("slice_by_index_36")]; + tensor reshape_181_shape_0 = const()[name = tensor("reshape_181_shape_0"), val = tensor([-1])]; + tensor reshape_181 = reshape(shape = reshape_181_shape_0, x = slice_by_index_36)[name = tensor("reshape_181")]; + tensor reshape_182_shape_0 = const()[name = tensor("reshape_182_shape_0"), val = tensor([-1])]; + tensor reshape_182_cast_fp16 = reshape(shape = reshape_182_shape_0, x = var_8018_cast_fp16)[name = tensor("reshape_182_cast_fp16")]; + tensor reshape_183_shape_0 = const()[name = tensor("reshape_183_shape_0"), val = tensor([-1])]; + tensor reshape_183_cast_fp16 = reshape(shape = reshape_183_shape_0, x = cache18_to_fp16)[name = tensor("reshape_183_cast_fp16")]; + tensor scatter_36_mode_0 = const()[name = tensor("scatter_36_mode_0"), val = tensor("update")]; + tensor scatter_36_axis_0 = const()[name = tensor("scatter_36_axis_0"), val = tensor(0)]; + tensor scatter_36_validate_indices_0 = const()[name = tensor("scatter_36_validate_indices_0"), val = tensor(false)]; + tensor scatter_36_cast_fp16 = scatter(axis = scatter_36_axis_0, data = reshape_183_cast_fp16, indices = reshape_181, mode = scatter_36_mode_0, updates = reshape_182_cast_fp16, validate_indices = scatter_36_validate_indices_0)[name = tensor("scatter_36_cast_fp16")]; + tensor reshape_184_cast_fp16 = reshape(shape = shape_84, x = scatter_36_cast_fp16)[name = tensor("reshape_184_cast_fp16")]; + tensor var_8026_begin_0 = const()[name = tensor("op_8026_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor var_8026_end_0 = const()[name = tensor("op_8026_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor var_8026_end_mask_0 = const()[name = tensor("op_8026_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_8026_squeeze_mask_0 = const()[name = tensor("op_8026_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor var_8026_cast_fp16 = slice_by_index(begin = var_8026_begin_0, end = var_8026_end_0, end_mask = var_8026_end_mask_0, squeeze_mask = var_8026_squeeze_mask_0, x = reshape_184_cast_fp16)[name = tensor("op_8026_cast_fp16")]; + tensor var_8028_axis_0 = const()[name = tensor("op_8028_axis_0"), val = tensor(1)]; + tensor var_8028_mode_0 = const()[name = tensor("op_8028_mode_0"), val = tensor("update")]; + tensor var_8028_validate_indices_0 = const()[name = tensor("op_8028_validate_indices_0"), val = tensor(false)]; + tensor var_8028_cast_fp16 = scatter_along_axis(axis = var_8028_axis_0, data = var_8026_cast_fp16, indices = write_indices_37, mode = var_8028_mode_0, updates = v_37_cast_fp16, validate_indices = var_8028_validate_indices_0)[name = tensor("op_8028_cast_fp16")]; + tensor concat_130 = const()[name = tensor("concat_130"), val = tensor([1, 0, 0, 0, 0])]; + tensor concat_131 = const()[name = tensor("concat_131"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_37_internal_tensor_assign_2_stride_0 = const()[name = tensor("new_cache_37_internal_tensor_assign_2_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_37_internal_tensor_assign_2_begin_mask_0 = const()[name = tensor("new_cache_37_internal_tensor_assign_2_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_37_internal_tensor_assign_2_end_mask_0 = const()[name = tensor("new_cache_37_internal_tensor_assign_2_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_37_internal_tensor_assign_2_squeeze_mask_0 = const()[name = tensor("new_cache_37_internal_tensor_assign_2_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_85 = const()[name = tensor("shape_85"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_37 = const()[name = tensor("reduce_prod_37"), val = tensor(1048576)]; + tensor range_1d_37_start_0 = const()[name = tensor("range_1d_37_start_0"), val = tensor(0)]; + tensor range_1d_37_step_0 = const()[name = tensor("range_1d_37_step_0"), val = tensor(1)]; + tensor range_1d_37 = range_1d(end = reduce_prod_37, start = range_1d_37_start_0, step = range_1d_37_step_0)[name = tensor("range_1d_37")]; + tensor reshape_185 = reshape(shape = shape_85, x = range_1d_37)[name = tensor("reshape_185")]; + tensor slice_by_index_37 = slice_by_index(begin = concat_130, begin_mask = new_cache_37_internal_tensor_assign_2_begin_mask_0, end = concat_131, end_mask = new_cache_37_internal_tensor_assign_2_end_mask_0, squeeze_mask = new_cache_37_internal_tensor_assign_2_squeeze_mask_0, stride = new_cache_37_internal_tensor_assign_2_stride_0, x = reshape_185)[name = tensor("slice_by_index_37")]; + tensor reshape_186_shape_0 = const()[name = tensor("reshape_186_shape_0"), val = tensor([-1])]; + tensor reshape_186 = reshape(shape = reshape_186_shape_0, x = slice_by_index_37)[name = tensor("reshape_186")]; + tensor reshape_187_shape_0 = const()[name = tensor("reshape_187_shape_0"), val = tensor([-1])]; + tensor reshape_187_cast_fp16 = reshape(shape = reshape_187_shape_0, x = var_8028_cast_fp16)[name = tensor("reshape_187_cast_fp16")]; + tensor reshape_188_shape_0 = const()[name = tensor("reshape_188_shape_0"), val = tensor([-1])]; + tensor reshape_188_cast_fp16 = reshape(shape = reshape_188_shape_0, x = reshape_184_cast_fp16)[name = tensor("reshape_188_cast_fp16")]; + tensor scatter_37_mode_0 = const()[name = tensor("scatter_37_mode_0"), val = tensor("update")]; + tensor scatter_37_axis_0 = const()[name = tensor("scatter_37_axis_0"), val = tensor(0)]; + tensor scatter_37_validate_indices_0 = const()[name = tensor("scatter_37_validate_indices_0"), val = tensor(false)]; + tensor scatter_37_cast_fp16 = scatter(axis = scatter_37_axis_0, data = reshape_188_cast_fp16, indices = reshape_186, mode = scatter_37_mode_0, updates = reshape_187_cast_fp16, validate_indices = scatter_37_validate_indices_0)[name = tensor("scatter_37_cast_fp16")]; + tensor reshape_189_cast_fp16 = reshape(shape = shape_85, x = scatter_37_cast_fp16)[name = tensor("reshape_189_cast_fp16")]; + tensor reshape_189_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("reshape_189_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor keys_109_begin_0 = const()[name = tensor("keys_109_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor keys_109_end_0 = const()[name = tensor("keys_109_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor keys_109_end_mask_0 = const()[name = tensor("keys_109_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor keys_109_squeeze_mask_0 = const()[name = tensor("keys_109_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor keys_109_cast_fp16 = slice_by_index(begin = keys_109_begin_0, end = keys_109_end_0, end_mask = keys_109_end_mask_0, squeeze_mask = keys_109_squeeze_mask_0, x = reshape_189_cast_fp16)[name = tensor("keys_109_cast_fp16")]; + tensor values_109_begin_0 = const()[name = tensor("values_109_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor values_109_end_0 = const()[name = tensor("values_109_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor values_109_end_mask_0 = const()[name = tensor("values_109_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor values_109_squeeze_mask_0 = const()[name = tensor("values_109_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor values_109_cast_fp16 = slice_by_index(begin = values_109_begin_0, end = values_109_end_0, end_mask = values_109_end_mask_0, squeeze_mask = values_109_squeeze_mask_0, x = reshape_189_cast_fp16)[name = tensor("values_109_cast_fp16")]; + tensor var_8040_cast_fp16 = not_equal(x = keys_109_cast_fp16, y = keys_109_cast_fp16)[name = tensor("op_8040_cast_fp16")]; + tensor keys_111_cast_fp16 = select(a = var_504_to_fp16, b = keys_109_cast_fp16, cond = var_8040_cast_fp16)[name = tensor("keys_111_cast_fp16")]; + tensor var_8048_cast_fp16 = not_equal(x = values_109_cast_fp16, y = values_109_cast_fp16)[name = tensor("op_8048_cast_fp16")]; + tensor values_111_cast_fp16 = select(a = var_504_to_fp16, b = values_109_cast_fp16, cond = var_8048_cast_fp16)[name = tensor("values_111_cast_fp16")]; + tensor var_8072 = const()[name = tensor("op_8072"), val = tensor([0, 2, 1, 3])]; + tensor var_8085 = const()[name = tensor("op_8085"), val = tensor([1, 1, 1])]; + tensor var_8086_cast_fp16 = reshape(shape = var_8085, x = position18_to_fp16)[name = tensor("op_8086_cast_fp16")]; + tensor var_8103_to_fp16 = const()[name = tensor("op_8103_to_fp16"), val = tensor(0x1p+0)]; + tensor valid_len_37_cast_fp16 = add(x = var_8086_cast_fp16, y = var_8103_to_fp16)[name = tensor("valid_len_37_cast_fp16")]; + tensor valid_mask_37_cast_fp16 = less(x = k_positions_1_promoted_to_fp16, y = valid_len_37_cast_fp16)[name = tensor("valid_mask_37_cast_fp16")]; + tensor causal_mask_37_cast_fp16 = less_equal(x = k_positions_1_promoted_to_fp16, y = var_8086_cast_fp16)[name = tensor("causal_mask_37_cast_fp16")]; + tensor attn_mask_73 = logical_and(x = valid_mask_37_cast_fp16, y = causal_mask_37_cast_fp16)[name = tensor("attn_mask_73")]; + tensor attn_mask_75_axes_0 = const()[name = tensor("attn_mask_75_axes_0"), val = tensor([1])]; + tensor attn_mask_75 = expand_dims(axes = attn_mask_75_axes_0, x = attn_mask_73)[name = tensor("attn_mask_75")]; + tensor var_8121_transpose_x_0 = const()[name = tensor("op_8121_transpose_x_0"), val = tensor(false)]; + tensor var_8121_transpose_y_0 = const()[name = tensor("op_8121_transpose_y_0"), val = tensor(false)]; + tensor transpose_108_perm_0 = const()[name = tensor("transpose_108_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_109_perm_0 = const()[name = tensor("transpose_109_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_109 = transpose(perm = transpose_109_perm_0, x = keys_111_cast_fp16)[name = tensor("transpose_141")]; + tensor transpose_108 = transpose(perm = transpose_108_perm_0, x = q_111_cast_fp16)[name = tensor("transpose_142")]; + tensor var_8121_cast_fp16 = matmul(transpose_x = var_8121_transpose_x_0, transpose_y = var_8121_transpose_y_0, x = transpose_108, y = transpose_109)[name = tensor("op_8121_cast_fp16")]; + tensor var_8115_to_fp16 = const()[name = tensor("op_8115_to_fp16"), val = tensor([0x1p-3])]; + tensor attn_weights_109_cast_fp16 = mul(x = var_8121_cast_fp16, y = var_8115_to_fp16)[name = tensor("attn_weights_109_cast_fp16")]; + tensor var_8123 = logical_not(x = attn_mask_75)[name = tensor("op_8123")]; + tensor var_8124_to_fp16 = const()[name = tensor("op_8124_to_fp16"), val = tensor(-inf)]; + tensor attn_weights_111_cast_fp16 = select(a = var_8124_to_fp16, b = attn_weights_109_cast_fp16, cond = var_8123)[name = tensor("attn_weights_111_cast_fp16")]; + tensor var_8126 = const()[name = tensor("op_8126"), val = tensor(-1)]; + tensor attn_weights_113_cast_fp16 = softmax(axis = var_8126, x = attn_weights_111_cast_fp16)[name = tensor("attn_weights_113_cast_fp16")]; + tensor attn_output_37_transpose_x_0 = const()[name = tensor("attn_output_37_transpose_x_0"), val = tensor(false)]; + tensor attn_output_37_transpose_y_0 = const()[name = tensor("attn_output_37_transpose_y_0"), val = tensor(false)]; + tensor values_113_cast_fp16 = transpose(perm = var_8072, x = values_111_cast_fp16)[name = tensor("transpose_143")]; + tensor attn_output_37_cast_fp16 = matmul(transpose_x = attn_output_37_transpose_x_0, transpose_y = attn_output_37_transpose_y_0, x = attn_weights_113_cast_fp16, y = values_113_cast_fp16)[name = tensor("attn_output_37_cast_fp16")]; + tensor var_8134 = const()[name = tensor("op_8134"), val = tensor([0, 2, 1, 3])]; + tensor var_8137 = const()[name = tensor("op_8137"), val = tensor([1, 1, 1024])]; + tensor var_8135_cast_fp16 = transpose(perm = var_8134, x = attn_output_37_cast_fp16)[name = tensor("transpose_140")]; + tensor input_185_cast_fp16 = reshape(shape = var_8137, x = var_8135_cast_fp16)[name = tensor("input_185_cast_fp16")]; + tensor linear_74_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = attn18_out_proj_weight_cast_fp16, x = input_185_cast_fp16)[name = tensor("linear_74_cast_fp16")]; + tensor var_8143_to_fp16 = const()[name = tensor("op_8143_to_fp16"), val = tensor(0x1p+0)]; + tensor var_8144_cast_fp16 = add(x = position18_to_fp16, y = var_8143_to_fp16)[name = tensor("op_8144_cast_fp16")]; + tensor var_8144_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_8144_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor input_187_cast_fp16 = add(x = input_183_cast_fp16, y = linear_74_cast_fp16)[name = tensor("input_187_cast_fp16")]; + tensor input_189_axes_0 = const()[name = tensor("input_189_axes_0"), val = tensor([-1])]; + tensor norm18_2_weight_to_fp16 = const()[name = tensor("norm18_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(303743616)))]; + tensor norm18_2_bias_to_fp16 = const()[name = tensor("norm18_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(303745728)))]; + tensor var_8148_to_fp16 = const()[name = tensor("op_8148_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_189_cast_fp16 = layer_norm(axes = input_189_axes_0, beta = norm18_2_bias_to_fp16, epsilon = var_8148_to_fp16, gamma = norm18_2_weight_to_fp16, x = input_187_cast_fp16)[name = tensor("input_189_cast_fp16")]; + tensor linear_75_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = linear18_1_weight_cast_fp16, x = input_189_cast_fp16)[name = tensor("linear_75_cast_fp16")]; + tensor input_191_mode_0 = const()[name = tensor("input_191_mode_0"), val = tensor("EXACT")]; + tensor input_191_cast_fp16 = gelu(mode = input_191_mode_0, x = linear_75_cast_fp16)[name = tensor("input_191_cast_fp16")]; + tensor linear_76_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = linear18_2_weight_cast_fp16, x = input_191_cast_fp16)[name = tensor("linear_76_cast_fp16")]; + tensor input_193_cast_fp16 = add(x = input_187_cast_fp16, y = linear_76_cast_fp16)[name = tensor("input_193_cast_fp16")]; + tensor x_39_axes_0 = const()[name = tensor("x_39_axes_0"), val = tensor([-1])]; + tensor norm19_1_weight_to_fp16 = const()[name = tensor("norm19_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(303747840)))]; + tensor norm19_1_bias_to_fp16 = const()[name = tensor("norm19_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(303749952)))]; + tensor var_8165_to_fp16 = const()[name = tensor("op_8165_to_fp16"), val = tensor(0x1.5p-17)]; + tensor x_39_cast_fp16 = layer_norm(axes = x_39_axes_0, beta = norm19_1_bias_to_fp16, epsilon = var_8165_to_fp16, gamma = norm19_1_weight_to_fp16, x = input_193_cast_fp16)[name = tensor("x_39_cast_fp16")]; + tensor linear_77_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = attn19_in_proj_weight_cast_fp16, x = x_39_cast_fp16)[name = tensor("linear_77_cast_fp16")]; + tensor var_8201 = const()[name = tensor("op_8201"), val = tensor([1, 1, 3, 16, 64])]; + tensor qkv_39_cast_fp16 = reshape(shape = var_8201, x = linear_77_cast_fp16)[name = tensor("qkv_39_cast_fp16")]; + tensor q_115_begin_0 = const()[name = tensor("q_115_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor q_115_end_0 = const()[name = tensor("q_115_end_0"), val = tensor([1, 1, 1, 16, 64])]; + tensor q_115_end_mask_0 = const()[name = tensor("q_115_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor q_115_squeeze_mask_0 = const()[name = tensor("q_115_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor q_115_cast_fp16 = slice_by_index(begin = q_115_begin_0, end = q_115_end_0, end_mask = q_115_end_mask_0, squeeze_mask = q_115_squeeze_mask_0, x = qkv_39_cast_fp16)[name = tensor("q_115_cast_fp16")]; + tensor k_77_begin_0 = const()[name = tensor("k_77_begin_0"), val = tensor([0, 0, 1, 0, 0])]; + tensor k_77_end_0 = const()[name = tensor("k_77_end_0"), val = tensor([1, 1, 2, 16, 64])]; + tensor k_77_end_mask_0 = const()[name = tensor("k_77_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor k_77_squeeze_mask_0 = const()[name = tensor("k_77_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor k_77_cast_fp16 = slice_by_index(begin = k_77_begin_0, end = k_77_end_0, end_mask = k_77_end_mask_0, squeeze_mask = k_77_squeeze_mask_0, x = qkv_39_cast_fp16)[name = tensor("k_77_cast_fp16")]; + tensor v_39_begin_0 = const()[name = tensor("v_39_begin_0"), val = tensor([0, 0, 2, 0, 0])]; + tensor v_39_end_0 = const()[name = tensor("v_39_end_0"), val = tensor([1, 1, 3, 16, 64])]; + tensor v_39_end_mask_0 = const()[name = tensor("v_39_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor v_39_squeeze_mask_0 = const()[name = tensor("v_39_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor v_39_cast_fp16 = slice_by_index(begin = v_39_begin_0, end = v_39_end_0, end_mask = v_39_end_mask_0, squeeze_mask = v_39_squeeze_mask_0, x = qkv_39_cast_fp16)[name = tensor("v_39_cast_fp16")]; + tensor var_8305 = const()[name = tensor("op_8305"), val = tensor([1, 1, 1, 1])]; + tensor position19_to_fp16_dtype_0 = const()[name = tensor("position19_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor position19_to_fp16 = cast(dtype = position19_to_fp16_dtype_0, x = position19)[name = tensor("cast_458")]; + tensor ts_119_cast_fp16 = reshape(shape = var_8305, x = position19_to_fp16)[name = tensor("ts_119_cast_fp16")]; + tensor var_8309 = const()[name = tensor("op_8309"), val = tensor([1, 1, 16, 32, 2])]; + tensor q_complex_39_cast_fp16 = reshape(shape = var_8309, x = q_115_cast_fp16)[name = tensor("q_complex_39_cast_fp16")]; + tensor var_8313 = const()[name = tensor("op_8313"), val = tensor([1, 1, 16, 32, 2])]; + tensor k_complex_39_cast_fp16 = reshape(shape = var_8313, x = k_77_cast_fp16)[name = tensor("k_complex_39_cast_fp16")]; + tensor var_8317_begin_0 = const()[name = tensor("op_8317_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_8317_end_0 = const()[name = tensor("op_8317_end_0"), val = tensor([1, 1, 16, 32, 1])]; + tensor var_8317_end_mask_0 = const()[name = tensor("op_8317_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_8317_squeeze_mask_0 = const()[name = tensor("op_8317_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_8317_cast_fp16 = slice_by_index(begin = var_8317_begin_0, end = var_8317_end_0, end_mask = var_8317_end_mask_0, squeeze_mask = var_8317_squeeze_mask_0, x = q_complex_39_cast_fp16)[name = tensor("op_8317_cast_fp16")]; + tensor var_8325_begin_0 = const()[name = tensor("op_8325_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_8325_end_0 = const()[name = tensor("op_8325_end_0"), val = tensor([1, 1, 16, 32, 2])]; + tensor var_8325_end_mask_0 = const()[name = tensor("op_8325_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_8325_squeeze_mask_0 = const()[name = tensor("op_8325_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_8325_cast_fp16 = slice_by_index(begin = var_8325_begin_0, end = var_8325_end_0, end_mask = var_8325_end_mask_0, squeeze_mask = var_8325_squeeze_mask_0, x = q_complex_39_cast_fp16)[name = tensor("op_8325_cast_fp16")]; + tensor var_8333_begin_0 = const()[name = tensor("op_8333_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_8333_end_0 = const()[name = tensor("op_8333_end_0"), val = tensor([1, 1, 16, 32, 1])]; + tensor var_8333_end_mask_0 = const()[name = tensor("op_8333_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_8333_squeeze_mask_0 = const()[name = tensor("op_8333_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_8333_cast_fp16 = slice_by_index(begin = var_8333_begin_0, end = var_8333_end_0, end_mask = var_8333_end_mask_0, squeeze_mask = var_8333_squeeze_mask_0, x = k_complex_39_cast_fp16)[name = tensor("op_8333_cast_fp16")]; + tensor var_8341_begin_0 = const()[name = tensor("op_8341_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_8341_end_0 = const()[name = tensor("op_8341_end_0"), val = tensor([1, 1, 16, 32, 2])]; + tensor var_8341_end_mask_0 = const()[name = tensor("op_8341_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_8341_squeeze_mask_0 = const()[name = tensor("op_8341_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_8341_cast_fp16 = slice_by_index(begin = var_8341_begin_0, end = var_8341_end_0, end_mask = var_8341_end_mask_0, squeeze_mask = var_8341_squeeze_mask_0, x = k_complex_39_cast_fp16)[name = tensor("op_8341_cast_fp16")]; + tensor freqs_39_to_fp16 = const()[name = tensor("freqs_39_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(303752064)))]; + tensor var_8347_cast_fp16 = mul(x = freqs_39_to_fp16, y = ts_119_cast_fp16)[name = tensor("op_8347_cast_fp16")]; + tensor rotr_39_cast_fp16 = cos(x = var_8347_cast_fp16)[name = tensor("rotr_39_cast_fp16")]; + tensor roti_39_cast_fp16 = sin(x = var_8347_cast_fp16)[name = tensor("roti_39_cast_fp16")]; + tensor var_8351_cast_fp16 = mul(x = var_8317_cast_fp16, y = rotr_39_cast_fp16)[name = tensor("op_8351_cast_fp16")]; + tensor var_8352_cast_fp16 = mul(x = var_8325_cast_fp16, y = roti_39_cast_fp16)[name = tensor("op_8352_cast_fp16")]; + tensor qor_77_cast_fp16 = sub(x = var_8351_cast_fp16, y = var_8352_cast_fp16)[name = tensor("qor_77_cast_fp16")]; + tensor var_8355_cast_fp16 = mul(x = var_8317_cast_fp16, y = roti_39_cast_fp16)[name = tensor("op_8355_cast_fp16")]; + tensor var_8356_cast_fp16 = mul(x = var_8325_cast_fp16, y = rotr_39_cast_fp16)[name = tensor("op_8356_cast_fp16")]; + tensor qoi_77_cast_fp16 = add(x = var_8355_cast_fp16, y = var_8356_cast_fp16)[name = tensor("qoi_77_cast_fp16")]; + tensor var_8359_cast_fp16 = mul(x = var_8333_cast_fp16, y = rotr_39_cast_fp16)[name = tensor("op_8359_cast_fp16")]; + tensor var_8360_cast_fp16 = mul(x = var_8341_cast_fp16, y = roti_39_cast_fp16)[name = tensor("op_8360_cast_fp16")]; + tensor kor_77_cast_fp16 = sub(x = var_8359_cast_fp16, y = var_8360_cast_fp16)[name = tensor("kor_77_cast_fp16")]; + tensor var_8363_cast_fp16 = mul(x = var_8333_cast_fp16, y = roti_39_cast_fp16)[name = tensor("op_8363_cast_fp16")]; + tensor var_8364_cast_fp16 = mul(x = var_8341_cast_fp16, y = rotr_39_cast_fp16)[name = tensor("op_8364_cast_fp16")]; + tensor koi_77_cast_fp16 = add(x = var_8363_cast_fp16, y = var_8364_cast_fp16)[name = tensor("koi_77_cast_fp16")]; + tensor qo_39_axis_0 = const()[name = tensor("qo_39_axis_0"), val = tensor(-1)]; + tensor qo_39_cast_fp16 = stack(axis = qo_39_axis_0, values = (qor_77_cast_fp16, qoi_77_cast_fp16))[name = tensor("qo_39_cast_fp16")]; + tensor ko_39_axis_0 = const()[name = tensor("ko_39_axis_0"), val = tensor(-1)]; + tensor ko_39_cast_fp16 = stack(axis = ko_39_axis_0, values = (kor_77_cast_fp16, koi_77_cast_fp16))[name = tensor("ko_39_cast_fp16")]; + tensor var_8393 = const()[name = tensor("op_8393"), val = tensor([1, 1, 16, 64])]; + tensor q_117_cast_fp16 = reshape(shape = var_8393, x = qo_39_cast_fp16)[name = tensor("q_117_cast_fp16")]; + tensor var_8395 = const()[name = tensor("op_8395"), val = tensor([1, 1, 16, 64])]; + tensor k_79_cast_fp16 = reshape(shape = var_8395, x = ko_39_cast_fp16)[name = tensor("k_79_cast_fp16")]; + tensor _inversed_8417_y_0_to_fp16 = const()[name = tensor("_inversed_8417_y_0_to_fp16"), val = tensor(0x1p-9)]; + tensor _inversed_8417_cast_fp16 = mul(x = ts_119_cast_fp16, y = _inversed_8417_y_0_to_fp16)[name = tensor("_inversed_8417_cast_fp16")]; + tensor var_8418_cast_fp16 = floor(x = _inversed_8417_cast_fp16)[name = tensor("op_8418_cast_fp16")]; + tensor var_8419_to_fp16 = const()[name = tensor("op_8419_to_fp16"), val = tensor(0x1p+9)]; + tensor var_8420_cast_fp16 = mul(x = var_8418_cast_fp16, y = var_8419_to_fp16)[name = tensor("op_8420_cast_fp16")]; + tensor write_indices_float_79_cast_fp16 = sub(x = ts_119_cast_fp16, y = var_8420_cast_fp16)[name = tensor("write_indices_float_79_cast_fp16")]; + tensor var_8427_dtype_0 = const()[name = tensor("op_8427_dtype_0"), val = tensor("int32")]; + tensor write_indices_39_reps_0 = const()[name = tensor("write_indices_39_reps_0"), val = tensor([1, 1, 16, 64])]; + tensor write_indices_float_79_cast_fp16_to_int32 = cast(dtype = var_8427_dtype_0, x = write_indices_float_79_cast_fp16)[name = tensor("cast_457")]; + tensor write_indices_39 = tile(reps = write_indices_39_reps_0, x = write_indices_float_79_cast_fp16_to_int32)[name = tensor("write_indices_39")]; + tensor var_8435_begin_0 = const()[name = tensor("op_8435_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_8435_end_0 = const()[name = tensor("op_8435_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor var_8435_end_mask_0 = const()[name = tensor("op_8435_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_8435_squeeze_mask_0 = const()[name = tensor("op_8435_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor cache19_to_fp16_dtype_0 = const()[name = tensor("cache19_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor cache19_to_fp16 = cast(dtype = cache19_to_fp16_dtype_0, x = cache19)[name = tensor("cast_456")]; + tensor var_8435_cast_fp16 = slice_by_index(begin = var_8435_begin_0, end = var_8435_end_0, end_mask = var_8435_end_mask_0, squeeze_mask = var_8435_squeeze_mask_0, x = cache19_to_fp16)[name = tensor("op_8435_cast_fp16")]; + tensor var_8437_axis_0 = const()[name = tensor("op_8437_axis_0"), val = tensor(1)]; + tensor var_8437_mode_0 = const()[name = tensor("op_8437_mode_0"), val = tensor("update")]; + tensor var_8437_validate_indices_0 = const()[name = tensor("op_8437_validate_indices_0"), val = tensor(false)]; + tensor var_8437_cast_fp16 = scatter_along_axis(axis = var_8437_axis_0, data = var_8435_cast_fp16, indices = write_indices_39, mode = var_8437_mode_0, updates = k_79_cast_fp16, validate_indices = var_8437_validate_indices_0)[name = tensor("op_8437_cast_fp16")]; + tensor concat_135 = const()[name = tensor("concat_135"), val = tensor([0, 0, 0, 0, 0])]; + tensor concat_136 = const()[name = tensor("concat_136"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_39_internal_tensor_assign_1_stride_0 = const()[name = tensor("new_cache_39_internal_tensor_assign_1_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_39_internal_tensor_assign_1_begin_mask_0 = const()[name = tensor("new_cache_39_internal_tensor_assign_1_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_39_internal_tensor_assign_1_end_mask_0 = const()[name = tensor("new_cache_39_internal_tensor_assign_1_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_39_internal_tensor_assign_1_squeeze_mask_0 = const()[name = tensor("new_cache_39_internal_tensor_assign_1_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_86 = const()[name = tensor("shape_86"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_38 = const()[name = tensor("reduce_prod_38"), val = tensor(1048576)]; + tensor range_1d_38_start_0 = const()[name = tensor("range_1d_38_start_0"), val = tensor(0)]; + tensor range_1d_38_step_0 = const()[name = tensor("range_1d_38_step_0"), val = tensor(1)]; + tensor range_1d_38 = range_1d(end = reduce_prod_38, start = range_1d_38_start_0, step = range_1d_38_step_0)[name = tensor("range_1d_38")]; + tensor reshape_190 = reshape(shape = shape_86, x = range_1d_38)[name = tensor("reshape_190")]; + tensor slice_by_index_38 = slice_by_index(begin = concat_135, begin_mask = new_cache_39_internal_tensor_assign_1_begin_mask_0, end = concat_136, end_mask = new_cache_39_internal_tensor_assign_1_end_mask_0, squeeze_mask = new_cache_39_internal_tensor_assign_1_squeeze_mask_0, stride = new_cache_39_internal_tensor_assign_1_stride_0, x = reshape_190)[name = tensor("slice_by_index_38")]; + tensor reshape_191_shape_0 = const()[name = tensor("reshape_191_shape_0"), val = tensor([-1])]; + tensor reshape_191 = reshape(shape = reshape_191_shape_0, x = slice_by_index_38)[name = tensor("reshape_191")]; + tensor reshape_192_shape_0 = const()[name = tensor("reshape_192_shape_0"), val = tensor([-1])]; + tensor reshape_192_cast_fp16 = reshape(shape = reshape_192_shape_0, x = var_8437_cast_fp16)[name = tensor("reshape_192_cast_fp16")]; + tensor reshape_193_shape_0 = const()[name = tensor("reshape_193_shape_0"), val = tensor([-1])]; + tensor reshape_193_cast_fp16 = reshape(shape = reshape_193_shape_0, x = cache19_to_fp16)[name = tensor("reshape_193_cast_fp16")]; + tensor scatter_38_mode_0 = const()[name = tensor("scatter_38_mode_0"), val = tensor("update")]; + tensor scatter_38_axis_0 = const()[name = tensor("scatter_38_axis_0"), val = tensor(0)]; + tensor scatter_38_validate_indices_0 = const()[name = tensor("scatter_38_validate_indices_0"), val = tensor(false)]; + tensor scatter_38_cast_fp16 = scatter(axis = scatter_38_axis_0, data = reshape_193_cast_fp16, indices = reshape_191, mode = scatter_38_mode_0, updates = reshape_192_cast_fp16, validate_indices = scatter_38_validate_indices_0)[name = tensor("scatter_38_cast_fp16")]; + tensor reshape_194_cast_fp16 = reshape(shape = shape_86, x = scatter_38_cast_fp16)[name = tensor("reshape_194_cast_fp16")]; + tensor var_8445_begin_0 = const()[name = tensor("op_8445_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor var_8445_end_0 = const()[name = tensor("op_8445_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor var_8445_end_mask_0 = const()[name = tensor("op_8445_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_8445_squeeze_mask_0 = const()[name = tensor("op_8445_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor var_8445_cast_fp16 = slice_by_index(begin = var_8445_begin_0, end = var_8445_end_0, end_mask = var_8445_end_mask_0, squeeze_mask = var_8445_squeeze_mask_0, x = reshape_194_cast_fp16)[name = tensor("op_8445_cast_fp16")]; + tensor var_8447_axis_0 = const()[name = tensor("op_8447_axis_0"), val = tensor(1)]; + tensor var_8447_mode_0 = const()[name = tensor("op_8447_mode_0"), val = tensor("update")]; + tensor var_8447_validate_indices_0 = const()[name = tensor("op_8447_validate_indices_0"), val = tensor(false)]; + tensor var_8447_cast_fp16 = scatter_along_axis(axis = var_8447_axis_0, data = var_8445_cast_fp16, indices = write_indices_39, mode = var_8447_mode_0, updates = v_39_cast_fp16, validate_indices = var_8447_validate_indices_0)[name = tensor("op_8447_cast_fp16")]; + tensor concat_137 = const()[name = tensor("concat_137"), val = tensor([1, 0, 0, 0, 0])]; + tensor concat_138 = const()[name = tensor("concat_138"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_39_internal_tensor_assign_2_stride_0 = const()[name = tensor("new_cache_39_internal_tensor_assign_2_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_39_internal_tensor_assign_2_begin_mask_0 = const()[name = tensor("new_cache_39_internal_tensor_assign_2_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_39_internal_tensor_assign_2_end_mask_0 = const()[name = tensor("new_cache_39_internal_tensor_assign_2_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_39_internal_tensor_assign_2_squeeze_mask_0 = const()[name = tensor("new_cache_39_internal_tensor_assign_2_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_87 = const()[name = tensor("shape_87"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_39 = const()[name = tensor("reduce_prod_39"), val = tensor(1048576)]; + tensor range_1d_39_start_0 = const()[name = tensor("range_1d_39_start_0"), val = tensor(0)]; + tensor range_1d_39_step_0 = const()[name = tensor("range_1d_39_step_0"), val = tensor(1)]; + tensor range_1d_39 = range_1d(end = reduce_prod_39, start = range_1d_39_start_0, step = range_1d_39_step_0)[name = tensor("range_1d_39")]; + tensor reshape_195 = reshape(shape = shape_87, x = range_1d_39)[name = tensor("reshape_195")]; + tensor slice_by_index_39 = slice_by_index(begin = concat_137, begin_mask = new_cache_39_internal_tensor_assign_2_begin_mask_0, end = concat_138, end_mask = new_cache_39_internal_tensor_assign_2_end_mask_0, squeeze_mask = new_cache_39_internal_tensor_assign_2_squeeze_mask_0, stride = new_cache_39_internal_tensor_assign_2_stride_0, x = reshape_195)[name = tensor("slice_by_index_39")]; + tensor reshape_196_shape_0 = const()[name = tensor("reshape_196_shape_0"), val = tensor([-1])]; + tensor reshape_196 = reshape(shape = reshape_196_shape_0, x = slice_by_index_39)[name = tensor("reshape_196")]; + tensor reshape_197_shape_0 = const()[name = tensor("reshape_197_shape_0"), val = tensor([-1])]; + tensor reshape_197_cast_fp16 = reshape(shape = reshape_197_shape_0, x = var_8447_cast_fp16)[name = tensor("reshape_197_cast_fp16")]; + tensor reshape_198_shape_0 = const()[name = tensor("reshape_198_shape_0"), val = tensor([-1])]; + tensor reshape_198_cast_fp16 = reshape(shape = reshape_198_shape_0, x = reshape_194_cast_fp16)[name = tensor("reshape_198_cast_fp16")]; + tensor scatter_39_mode_0 = const()[name = tensor("scatter_39_mode_0"), val = tensor("update")]; + tensor scatter_39_axis_0 = const()[name = tensor("scatter_39_axis_0"), val = tensor(0)]; + tensor scatter_39_validate_indices_0 = const()[name = tensor("scatter_39_validate_indices_0"), val = tensor(false)]; + tensor scatter_39_cast_fp16 = scatter(axis = scatter_39_axis_0, data = reshape_198_cast_fp16, indices = reshape_196, mode = scatter_39_mode_0, updates = reshape_197_cast_fp16, validate_indices = scatter_39_validate_indices_0)[name = tensor("scatter_39_cast_fp16")]; + tensor reshape_199_cast_fp16 = reshape(shape = shape_87, x = scatter_39_cast_fp16)[name = tensor("reshape_199_cast_fp16")]; + tensor reshape_199_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("reshape_199_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor keys_115_begin_0 = const()[name = tensor("keys_115_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor keys_115_end_0 = const()[name = tensor("keys_115_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor keys_115_end_mask_0 = const()[name = tensor("keys_115_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor keys_115_squeeze_mask_0 = const()[name = tensor("keys_115_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor keys_115_cast_fp16 = slice_by_index(begin = keys_115_begin_0, end = keys_115_end_0, end_mask = keys_115_end_mask_0, squeeze_mask = keys_115_squeeze_mask_0, x = reshape_199_cast_fp16)[name = tensor("keys_115_cast_fp16")]; + tensor values_115_begin_0 = const()[name = tensor("values_115_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor values_115_end_0 = const()[name = tensor("values_115_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor values_115_end_mask_0 = const()[name = tensor("values_115_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor values_115_squeeze_mask_0 = const()[name = tensor("values_115_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor values_115_cast_fp16 = slice_by_index(begin = values_115_begin_0, end = values_115_end_0, end_mask = values_115_end_mask_0, squeeze_mask = values_115_squeeze_mask_0, x = reshape_199_cast_fp16)[name = tensor("values_115_cast_fp16")]; + tensor var_8459_cast_fp16 = not_equal(x = keys_115_cast_fp16, y = keys_115_cast_fp16)[name = tensor("op_8459_cast_fp16")]; + tensor keys_117_cast_fp16 = select(a = var_504_to_fp16, b = keys_115_cast_fp16, cond = var_8459_cast_fp16)[name = tensor("keys_117_cast_fp16")]; + tensor var_8467_cast_fp16 = not_equal(x = values_115_cast_fp16, y = values_115_cast_fp16)[name = tensor("op_8467_cast_fp16")]; + tensor values_117_cast_fp16 = select(a = var_504_to_fp16, b = values_115_cast_fp16, cond = var_8467_cast_fp16)[name = tensor("values_117_cast_fp16")]; + tensor var_8491 = const()[name = tensor("op_8491"), val = tensor([0, 2, 1, 3])]; + tensor var_8504 = const()[name = tensor("op_8504"), val = tensor([1, 1, 1])]; + tensor var_8505_cast_fp16 = reshape(shape = var_8504, x = position19_to_fp16)[name = tensor("op_8505_cast_fp16")]; + tensor var_8522_to_fp16 = const()[name = tensor("op_8522_to_fp16"), val = tensor(0x1p+0)]; + tensor valid_len_39_cast_fp16 = add(x = var_8505_cast_fp16, y = var_8522_to_fp16)[name = tensor("valid_len_39_cast_fp16")]; + tensor valid_mask_39_cast_fp16 = less(x = k_positions_1_promoted_to_fp16, y = valid_len_39_cast_fp16)[name = tensor("valid_mask_39_cast_fp16")]; + tensor causal_mask_39_cast_fp16 = less_equal(x = k_positions_1_promoted_to_fp16, y = var_8505_cast_fp16)[name = tensor("causal_mask_39_cast_fp16")]; + tensor attn_mask_77 = logical_and(x = valid_mask_39_cast_fp16, y = causal_mask_39_cast_fp16)[name = tensor("attn_mask_77")]; + tensor attn_mask_79_axes_0 = const()[name = tensor("attn_mask_79_axes_0"), val = tensor([1])]; + tensor attn_mask_79 = expand_dims(axes = attn_mask_79_axes_0, x = attn_mask_77)[name = tensor("attn_mask_79")]; + tensor var_8540_transpose_x_0 = const()[name = tensor("op_8540_transpose_x_0"), val = tensor(false)]; + tensor var_8540_transpose_y_0 = const()[name = tensor("op_8540_transpose_y_0"), val = tensor(false)]; + tensor transpose_110_perm_0 = const()[name = tensor("transpose_110_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_111_perm_0 = const()[name = tensor("transpose_111_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_111 = transpose(perm = transpose_111_perm_0, x = keys_117_cast_fp16)[name = tensor("transpose_137")]; + tensor transpose_110 = transpose(perm = transpose_110_perm_0, x = q_117_cast_fp16)[name = tensor("transpose_138")]; + tensor var_8540_cast_fp16 = matmul(transpose_x = var_8540_transpose_x_0, transpose_y = var_8540_transpose_y_0, x = transpose_110, y = transpose_111)[name = tensor("op_8540_cast_fp16")]; + tensor var_8534_to_fp16 = const()[name = tensor("op_8534_to_fp16"), val = tensor([0x1p-3])]; + tensor attn_weights_115_cast_fp16 = mul(x = var_8540_cast_fp16, y = var_8534_to_fp16)[name = tensor("attn_weights_115_cast_fp16")]; + tensor var_8542 = logical_not(x = attn_mask_79)[name = tensor("op_8542")]; + tensor var_8543_to_fp16 = const()[name = tensor("op_8543_to_fp16"), val = tensor(-inf)]; + tensor attn_weights_117_cast_fp16 = select(a = var_8543_to_fp16, b = attn_weights_115_cast_fp16, cond = var_8542)[name = tensor("attn_weights_117_cast_fp16")]; + tensor var_8545 = const()[name = tensor("op_8545"), val = tensor(-1)]; + tensor attn_weights_119_cast_fp16 = softmax(axis = var_8545, x = attn_weights_117_cast_fp16)[name = tensor("attn_weights_119_cast_fp16")]; + tensor attn_output_39_transpose_x_0 = const()[name = tensor("attn_output_39_transpose_x_0"), val = tensor(false)]; + tensor attn_output_39_transpose_y_0 = const()[name = tensor("attn_output_39_transpose_y_0"), val = tensor(false)]; + tensor values_119_cast_fp16 = transpose(perm = var_8491, x = values_117_cast_fp16)[name = tensor("transpose_139")]; + tensor attn_output_39_cast_fp16 = matmul(transpose_x = attn_output_39_transpose_x_0, transpose_y = attn_output_39_transpose_y_0, x = attn_weights_119_cast_fp16, y = values_119_cast_fp16)[name = tensor("attn_output_39_cast_fp16")]; + tensor var_8553 = const()[name = tensor("op_8553"), val = tensor([0, 2, 1, 3])]; + tensor var_8556 = const()[name = tensor("op_8556"), val = tensor([1, 1, 1024])]; + tensor var_8554_cast_fp16 = transpose(perm = var_8553, x = attn_output_39_cast_fp16)[name = tensor("transpose_136")]; + tensor input_195_cast_fp16 = reshape(shape = var_8556, x = var_8554_cast_fp16)[name = tensor("input_195_cast_fp16")]; + tensor linear_78_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = attn19_out_proj_weight_cast_fp16, x = input_195_cast_fp16)[name = tensor("linear_78_cast_fp16")]; + tensor var_8562_to_fp16 = const()[name = tensor("op_8562_to_fp16"), val = tensor(0x1p+0)]; + tensor var_8563_cast_fp16 = add(x = position19_to_fp16, y = var_8562_to_fp16)[name = tensor("op_8563_cast_fp16")]; + tensor var_8563_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_8563_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor input_197_cast_fp16 = add(x = input_193_cast_fp16, y = linear_78_cast_fp16)[name = tensor("input_197_cast_fp16")]; + tensor input_199_axes_0 = const()[name = tensor("input_199_axes_0"), val = tensor([-1])]; + tensor norm19_2_weight_to_fp16 = const()[name = tensor("norm19_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(303752192)))]; + tensor norm19_2_bias_to_fp16 = const()[name = tensor("norm19_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(303754304)))]; + tensor var_8567_to_fp16 = const()[name = tensor("op_8567_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_199_cast_fp16 = layer_norm(axes = input_199_axes_0, beta = norm19_2_bias_to_fp16, epsilon = var_8567_to_fp16, gamma = norm19_2_weight_to_fp16, x = input_197_cast_fp16)[name = tensor("input_199_cast_fp16")]; + tensor linear_79_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = linear19_1_weight_cast_fp16, x = input_199_cast_fp16)[name = tensor("linear_79_cast_fp16")]; + tensor input_201_mode_0 = const()[name = tensor("input_201_mode_0"), val = tensor("EXACT")]; + tensor input_201_cast_fp16 = gelu(mode = input_201_mode_0, x = linear_79_cast_fp16)[name = tensor("input_201_cast_fp16")]; + tensor linear_80_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = linear19_2_weight_cast_fp16, x = input_201_cast_fp16)[name = tensor("linear_80_cast_fp16")]; + tensor input_203_cast_fp16 = add(x = input_197_cast_fp16, y = linear_80_cast_fp16)[name = tensor("input_203_cast_fp16")]; + tensor x_41_axes_0 = const()[name = tensor("x_41_axes_0"), val = tensor([-1])]; + tensor norm20_1_weight_to_fp16 = const()[name = tensor("norm20_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(303756416)))]; + tensor norm20_1_bias_to_fp16 = const()[name = tensor("norm20_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(303758528)))]; + tensor var_8584_to_fp16 = const()[name = tensor("op_8584_to_fp16"), val = tensor(0x1.5p-17)]; + tensor x_41_cast_fp16 = layer_norm(axes = x_41_axes_0, beta = norm20_1_bias_to_fp16, epsilon = var_8584_to_fp16, gamma = norm20_1_weight_to_fp16, x = input_203_cast_fp16)[name = tensor("x_41_cast_fp16")]; + tensor linear_81_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = attn20_in_proj_weight_cast_fp16, x = x_41_cast_fp16)[name = tensor("linear_81_cast_fp16")]; + tensor var_8620 = const()[name = tensor("op_8620"), val = tensor([1, 1, 3, 16, 64])]; + tensor qkv_41_cast_fp16 = reshape(shape = var_8620, x = linear_81_cast_fp16)[name = tensor("qkv_41_cast_fp16")]; + tensor q_121_begin_0 = const()[name = tensor("q_121_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor q_121_end_0 = const()[name = tensor("q_121_end_0"), val = tensor([1, 1, 1, 16, 64])]; + tensor q_121_end_mask_0 = const()[name = tensor("q_121_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor q_121_squeeze_mask_0 = const()[name = tensor("q_121_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor q_121_cast_fp16 = slice_by_index(begin = q_121_begin_0, end = q_121_end_0, end_mask = q_121_end_mask_0, squeeze_mask = q_121_squeeze_mask_0, x = qkv_41_cast_fp16)[name = tensor("q_121_cast_fp16")]; + tensor k_81_begin_0 = const()[name = tensor("k_81_begin_0"), val = tensor([0, 0, 1, 0, 0])]; + tensor k_81_end_0 = const()[name = tensor("k_81_end_0"), val = tensor([1, 1, 2, 16, 64])]; + tensor k_81_end_mask_0 = const()[name = tensor("k_81_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor k_81_squeeze_mask_0 = const()[name = tensor("k_81_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor k_81_cast_fp16 = slice_by_index(begin = k_81_begin_0, end = k_81_end_0, end_mask = k_81_end_mask_0, squeeze_mask = k_81_squeeze_mask_0, x = qkv_41_cast_fp16)[name = tensor("k_81_cast_fp16")]; + tensor v_41_begin_0 = const()[name = tensor("v_41_begin_0"), val = tensor([0, 0, 2, 0, 0])]; + tensor v_41_end_0 = const()[name = tensor("v_41_end_0"), val = tensor([1, 1, 3, 16, 64])]; + tensor v_41_end_mask_0 = const()[name = tensor("v_41_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor v_41_squeeze_mask_0 = const()[name = tensor("v_41_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor v_41_cast_fp16 = slice_by_index(begin = v_41_begin_0, end = v_41_end_0, end_mask = v_41_end_mask_0, squeeze_mask = v_41_squeeze_mask_0, x = qkv_41_cast_fp16)[name = tensor("v_41_cast_fp16")]; + tensor var_8724 = const()[name = tensor("op_8724"), val = tensor([1, 1, 1, 1])]; + tensor position20_to_fp16_dtype_0 = const()[name = tensor("position20_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor position20_to_fp16 = cast(dtype = position20_to_fp16_dtype_0, x = position20)[name = tensor("cast_453")]; + tensor ts_125_cast_fp16 = reshape(shape = var_8724, x = position20_to_fp16)[name = tensor("ts_125_cast_fp16")]; + tensor var_8728 = const()[name = tensor("op_8728"), val = tensor([1, 1, 16, 32, 2])]; + tensor q_complex_41_cast_fp16 = reshape(shape = var_8728, x = q_121_cast_fp16)[name = tensor("q_complex_41_cast_fp16")]; + tensor var_8732 = const()[name = tensor("op_8732"), val = tensor([1, 1, 16, 32, 2])]; + tensor k_complex_41_cast_fp16 = reshape(shape = var_8732, x = k_81_cast_fp16)[name = tensor("k_complex_41_cast_fp16")]; + tensor var_8736_begin_0 = const()[name = tensor("op_8736_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_8736_end_0 = const()[name = tensor("op_8736_end_0"), val = tensor([1, 1, 16, 32, 1])]; + tensor var_8736_end_mask_0 = const()[name = tensor("op_8736_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_8736_squeeze_mask_0 = const()[name = tensor("op_8736_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_8736_cast_fp16 = slice_by_index(begin = var_8736_begin_0, end = var_8736_end_0, end_mask = var_8736_end_mask_0, squeeze_mask = var_8736_squeeze_mask_0, x = q_complex_41_cast_fp16)[name = tensor("op_8736_cast_fp16")]; + tensor var_8744_begin_0 = const()[name = tensor("op_8744_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_8744_end_0 = const()[name = tensor("op_8744_end_0"), val = tensor([1, 1, 16, 32, 2])]; + tensor var_8744_end_mask_0 = const()[name = tensor("op_8744_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_8744_squeeze_mask_0 = const()[name = tensor("op_8744_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_8744_cast_fp16 = slice_by_index(begin = var_8744_begin_0, end = var_8744_end_0, end_mask = var_8744_end_mask_0, squeeze_mask = var_8744_squeeze_mask_0, x = q_complex_41_cast_fp16)[name = tensor("op_8744_cast_fp16")]; + tensor var_8752_begin_0 = const()[name = tensor("op_8752_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_8752_end_0 = const()[name = tensor("op_8752_end_0"), val = tensor([1, 1, 16, 32, 1])]; + tensor var_8752_end_mask_0 = const()[name = tensor("op_8752_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_8752_squeeze_mask_0 = const()[name = tensor("op_8752_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_8752_cast_fp16 = slice_by_index(begin = var_8752_begin_0, end = var_8752_end_0, end_mask = var_8752_end_mask_0, squeeze_mask = var_8752_squeeze_mask_0, x = k_complex_41_cast_fp16)[name = tensor("op_8752_cast_fp16")]; + tensor var_8760_begin_0 = const()[name = tensor("op_8760_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_8760_end_0 = const()[name = tensor("op_8760_end_0"), val = tensor([1, 1, 16, 32, 2])]; + tensor var_8760_end_mask_0 = const()[name = tensor("op_8760_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_8760_squeeze_mask_0 = const()[name = tensor("op_8760_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_8760_cast_fp16 = slice_by_index(begin = var_8760_begin_0, end = var_8760_end_0, end_mask = var_8760_end_mask_0, squeeze_mask = var_8760_squeeze_mask_0, x = k_complex_41_cast_fp16)[name = tensor("op_8760_cast_fp16")]; + tensor freqs_41_to_fp16 = const()[name = tensor("freqs_41_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(303760640)))]; + tensor var_8766_cast_fp16 = mul(x = freqs_41_to_fp16, y = ts_125_cast_fp16)[name = tensor("op_8766_cast_fp16")]; + tensor rotr_41_cast_fp16 = cos(x = var_8766_cast_fp16)[name = tensor("rotr_41_cast_fp16")]; + tensor roti_41_cast_fp16 = sin(x = var_8766_cast_fp16)[name = tensor("roti_41_cast_fp16")]; + tensor var_8770_cast_fp16 = mul(x = var_8736_cast_fp16, y = rotr_41_cast_fp16)[name = tensor("op_8770_cast_fp16")]; + tensor var_8771_cast_fp16 = mul(x = var_8744_cast_fp16, y = roti_41_cast_fp16)[name = tensor("op_8771_cast_fp16")]; + tensor qor_81_cast_fp16 = sub(x = var_8770_cast_fp16, y = var_8771_cast_fp16)[name = tensor("qor_81_cast_fp16")]; + tensor var_8774_cast_fp16 = mul(x = var_8736_cast_fp16, y = roti_41_cast_fp16)[name = tensor("op_8774_cast_fp16")]; + tensor var_8775_cast_fp16 = mul(x = var_8744_cast_fp16, y = rotr_41_cast_fp16)[name = tensor("op_8775_cast_fp16")]; + tensor qoi_81_cast_fp16 = add(x = var_8774_cast_fp16, y = var_8775_cast_fp16)[name = tensor("qoi_81_cast_fp16")]; + tensor var_8778_cast_fp16 = mul(x = var_8752_cast_fp16, y = rotr_41_cast_fp16)[name = tensor("op_8778_cast_fp16")]; + tensor var_8779_cast_fp16 = mul(x = var_8760_cast_fp16, y = roti_41_cast_fp16)[name = tensor("op_8779_cast_fp16")]; + tensor kor_81_cast_fp16 = sub(x = var_8778_cast_fp16, y = var_8779_cast_fp16)[name = tensor("kor_81_cast_fp16")]; + tensor var_8782_cast_fp16 = mul(x = var_8752_cast_fp16, y = roti_41_cast_fp16)[name = tensor("op_8782_cast_fp16")]; + tensor var_8783_cast_fp16 = mul(x = var_8760_cast_fp16, y = rotr_41_cast_fp16)[name = tensor("op_8783_cast_fp16")]; + tensor koi_81_cast_fp16 = add(x = var_8782_cast_fp16, y = var_8783_cast_fp16)[name = tensor("koi_81_cast_fp16")]; + tensor qo_41_axis_0 = const()[name = tensor("qo_41_axis_0"), val = tensor(-1)]; + tensor qo_41_cast_fp16 = stack(axis = qo_41_axis_0, values = (qor_81_cast_fp16, qoi_81_cast_fp16))[name = tensor("qo_41_cast_fp16")]; + tensor ko_41_axis_0 = const()[name = tensor("ko_41_axis_0"), val = tensor(-1)]; + tensor ko_41_cast_fp16 = stack(axis = ko_41_axis_0, values = (kor_81_cast_fp16, koi_81_cast_fp16))[name = tensor("ko_41_cast_fp16")]; + tensor var_8812 = const()[name = tensor("op_8812"), val = tensor([1, 1, 16, 64])]; + tensor q_123_cast_fp16 = reshape(shape = var_8812, x = qo_41_cast_fp16)[name = tensor("q_123_cast_fp16")]; + tensor var_8814 = const()[name = tensor("op_8814"), val = tensor([1, 1, 16, 64])]; + tensor k_83_cast_fp16 = reshape(shape = var_8814, x = ko_41_cast_fp16)[name = tensor("k_83_cast_fp16")]; + tensor _inversed_8836_y_0_to_fp16 = const()[name = tensor("_inversed_8836_y_0_to_fp16"), val = tensor(0x1p-9)]; + tensor _inversed_8836_cast_fp16 = mul(x = ts_125_cast_fp16, y = _inversed_8836_y_0_to_fp16)[name = tensor("_inversed_8836_cast_fp16")]; + tensor var_8837_cast_fp16 = floor(x = _inversed_8836_cast_fp16)[name = tensor("op_8837_cast_fp16")]; + tensor var_8838_to_fp16 = const()[name = tensor("op_8838_to_fp16"), val = tensor(0x1p+9)]; + tensor var_8839_cast_fp16 = mul(x = var_8837_cast_fp16, y = var_8838_to_fp16)[name = tensor("op_8839_cast_fp16")]; + tensor write_indices_float_83_cast_fp16 = sub(x = ts_125_cast_fp16, y = var_8839_cast_fp16)[name = tensor("write_indices_float_83_cast_fp16")]; + tensor var_8846_dtype_0 = const()[name = tensor("op_8846_dtype_0"), val = tensor("int32")]; + tensor write_indices_41_reps_0 = const()[name = tensor("write_indices_41_reps_0"), val = tensor([1, 1, 16, 64])]; + tensor write_indices_float_83_cast_fp16_to_int32 = cast(dtype = var_8846_dtype_0, x = write_indices_float_83_cast_fp16)[name = tensor("cast_452")]; + tensor write_indices_41 = tile(reps = write_indices_41_reps_0, x = write_indices_float_83_cast_fp16_to_int32)[name = tensor("write_indices_41")]; + tensor var_8854_begin_0 = const()[name = tensor("op_8854_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_8854_end_0 = const()[name = tensor("op_8854_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor var_8854_end_mask_0 = const()[name = tensor("op_8854_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_8854_squeeze_mask_0 = const()[name = tensor("op_8854_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor cache20_to_fp16_dtype_0 = const()[name = tensor("cache20_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor cache20_to_fp16 = cast(dtype = cache20_to_fp16_dtype_0, x = cache20)[name = tensor("cast_451")]; + tensor var_8854_cast_fp16 = slice_by_index(begin = var_8854_begin_0, end = var_8854_end_0, end_mask = var_8854_end_mask_0, squeeze_mask = var_8854_squeeze_mask_0, x = cache20_to_fp16)[name = tensor("op_8854_cast_fp16")]; + tensor var_8856_axis_0 = const()[name = tensor("op_8856_axis_0"), val = tensor(1)]; + tensor var_8856_mode_0 = const()[name = tensor("op_8856_mode_0"), val = tensor("update")]; + tensor var_8856_validate_indices_0 = const()[name = tensor("op_8856_validate_indices_0"), val = tensor(false)]; + tensor var_8856_cast_fp16 = scatter_along_axis(axis = var_8856_axis_0, data = var_8854_cast_fp16, indices = write_indices_41, mode = var_8856_mode_0, updates = k_83_cast_fp16, validate_indices = var_8856_validate_indices_0)[name = tensor("op_8856_cast_fp16")]; + tensor concat_142 = const()[name = tensor("concat_142"), val = tensor([0, 0, 0, 0, 0])]; + tensor concat_143 = const()[name = tensor("concat_143"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_41_internal_tensor_assign_1_stride_0 = const()[name = tensor("new_cache_41_internal_tensor_assign_1_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_41_internal_tensor_assign_1_begin_mask_0 = const()[name = tensor("new_cache_41_internal_tensor_assign_1_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_41_internal_tensor_assign_1_end_mask_0 = const()[name = tensor("new_cache_41_internal_tensor_assign_1_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_41_internal_tensor_assign_1_squeeze_mask_0 = const()[name = tensor("new_cache_41_internal_tensor_assign_1_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_88 = const()[name = tensor("shape_88"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_40 = const()[name = tensor("reduce_prod_40"), val = tensor(1048576)]; + tensor range_1d_40_start_0 = const()[name = tensor("range_1d_40_start_0"), val = tensor(0)]; + tensor range_1d_40_step_0 = const()[name = tensor("range_1d_40_step_0"), val = tensor(1)]; + tensor range_1d_40 = range_1d(end = reduce_prod_40, start = range_1d_40_start_0, step = range_1d_40_step_0)[name = tensor("range_1d_40")]; + tensor reshape_200 = reshape(shape = shape_88, x = range_1d_40)[name = tensor("reshape_200")]; + tensor slice_by_index_40 = slice_by_index(begin = concat_142, begin_mask = new_cache_41_internal_tensor_assign_1_begin_mask_0, end = concat_143, end_mask = new_cache_41_internal_tensor_assign_1_end_mask_0, squeeze_mask = new_cache_41_internal_tensor_assign_1_squeeze_mask_0, stride = new_cache_41_internal_tensor_assign_1_stride_0, x = reshape_200)[name = tensor("slice_by_index_40")]; + tensor reshape_201_shape_0 = const()[name = tensor("reshape_201_shape_0"), val = tensor([-1])]; + tensor reshape_201 = reshape(shape = reshape_201_shape_0, x = slice_by_index_40)[name = tensor("reshape_201")]; + tensor reshape_202_shape_0 = const()[name = tensor("reshape_202_shape_0"), val = tensor([-1])]; + tensor reshape_202_cast_fp16 = reshape(shape = reshape_202_shape_0, x = var_8856_cast_fp16)[name = tensor("reshape_202_cast_fp16")]; + tensor reshape_203_shape_0 = const()[name = tensor("reshape_203_shape_0"), val = tensor([-1])]; + tensor reshape_203_cast_fp16 = reshape(shape = reshape_203_shape_0, x = cache20_to_fp16)[name = tensor("reshape_203_cast_fp16")]; + tensor scatter_40_mode_0 = const()[name = tensor("scatter_40_mode_0"), val = tensor("update")]; + tensor scatter_40_axis_0 = const()[name = tensor("scatter_40_axis_0"), val = tensor(0)]; + tensor scatter_40_validate_indices_0 = const()[name = tensor("scatter_40_validate_indices_0"), val = tensor(false)]; + tensor scatter_40_cast_fp16 = scatter(axis = scatter_40_axis_0, data = reshape_203_cast_fp16, indices = reshape_201, mode = scatter_40_mode_0, updates = reshape_202_cast_fp16, validate_indices = scatter_40_validate_indices_0)[name = tensor("scatter_40_cast_fp16")]; + tensor reshape_204_cast_fp16 = reshape(shape = shape_88, x = scatter_40_cast_fp16)[name = tensor("reshape_204_cast_fp16")]; + tensor var_8864_begin_0 = const()[name = tensor("op_8864_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor var_8864_end_0 = const()[name = tensor("op_8864_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor var_8864_end_mask_0 = const()[name = tensor("op_8864_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_8864_squeeze_mask_0 = const()[name = tensor("op_8864_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor var_8864_cast_fp16 = slice_by_index(begin = var_8864_begin_0, end = var_8864_end_0, end_mask = var_8864_end_mask_0, squeeze_mask = var_8864_squeeze_mask_0, x = reshape_204_cast_fp16)[name = tensor("op_8864_cast_fp16")]; + tensor var_8866_axis_0 = const()[name = tensor("op_8866_axis_0"), val = tensor(1)]; + tensor var_8866_mode_0 = const()[name = tensor("op_8866_mode_0"), val = tensor("update")]; + tensor var_8866_validate_indices_0 = const()[name = tensor("op_8866_validate_indices_0"), val = tensor(false)]; + tensor var_8866_cast_fp16 = scatter_along_axis(axis = var_8866_axis_0, data = var_8864_cast_fp16, indices = write_indices_41, mode = var_8866_mode_0, updates = v_41_cast_fp16, validate_indices = var_8866_validate_indices_0)[name = tensor("op_8866_cast_fp16")]; + tensor concat_144 = const()[name = tensor("concat_144"), val = tensor([1, 0, 0, 0, 0])]; + tensor concat_145 = const()[name = tensor("concat_145"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_41_internal_tensor_assign_2_stride_0 = const()[name = tensor("new_cache_41_internal_tensor_assign_2_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_41_internal_tensor_assign_2_begin_mask_0 = const()[name = tensor("new_cache_41_internal_tensor_assign_2_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_41_internal_tensor_assign_2_end_mask_0 = const()[name = tensor("new_cache_41_internal_tensor_assign_2_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_41_internal_tensor_assign_2_squeeze_mask_0 = const()[name = tensor("new_cache_41_internal_tensor_assign_2_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_89 = const()[name = tensor("shape_89"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_41 = const()[name = tensor("reduce_prod_41"), val = tensor(1048576)]; + tensor range_1d_41_start_0 = const()[name = tensor("range_1d_41_start_0"), val = tensor(0)]; + tensor range_1d_41_step_0 = const()[name = tensor("range_1d_41_step_0"), val = tensor(1)]; + tensor range_1d_41 = range_1d(end = reduce_prod_41, start = range_1d_41_start_0, step = range_1d_41_step_0)[name = tensor("range_1d_41")]; + tensor reshape_205 = reshape(shape = shape_89, x = range_1d_41)[name = tensor("reshape_205")]; + tensor slice_by_index_41 = slice_by_index(begin = concat_144, begin_mask = new_cache_41_internal_tensor_assign_2_begin_mask_0, end = concat_145, end_mask = new_cache_41_internal_tensor_assign_2_end_mask_0, squeeze_mask = new_cache_41_internal_tensor_assign_2_squeeze_mask_0, stride = new_cache_41_internal_tensor_assign_2_stride_0, x = reshape_205)[name = tensor("slice_by_index_41")]; + tensor reshape_206_shape_0 = const()[name = tensor("reshape_206_shape_0"), val = tensor([-1])]; + tensor reshape_206 = reshape(shape = reshape_206_shape_0, x = slice_by_index_41)[name = tensor("reshape_206")]; + tensor reshape_207_shape_0 = const()[name = tensor("reshape_207_shape_0"), val = tensor([-1])]; + tensor reshape_207_cast_fp16 = reshape(shape = reshape_207_shape_0, x = var_8866_cast_fp16)[name = tensor("reshape_207_cast_fp16")]; + tensor reshape_208_shape_0 = const()[name = tensor("reshape_208_shape_0"), val = tensor([-1])]; + tensor reshape_208_cast_fp16 = reshape(shape = reshape_208_shape_0, x = reshape_204_cast_fp16)[name = tensor("reshape_208_cast_fp16")]; + tensor scatter_41_mode_0 = const()[name = tensor("scatter_41_mode_0"), val = tensor("update")]; + tensor scatter_41_axis_0 = const()[name = tensor("scatter_41_axis_0"), val = tensor(0)]; + tensor scatter_41_validate_indices_0 = const()[name = tensor("scatter_41_validate_indices_0"), val = tensor(false)]; + tensor scatter_41_cast_fp16 = scatter(axis = scatter_41_axis_0, data = reshape_208_cast_fp16, indices = reshape_206, mode = scatter_41_mode_0, updates = reshape_207_cast_fp16, validate_indices = scatter_41_validate_indices_0)[name = tensor("scatter_41_cast_fp16")]; + tensor reshape_209_cast_fp16 = reshape(shape = shape_89, x = scatter_41_cast_fp16)[name = tensor("reshape_209_cast_fp16")]; + tensor reshape_209_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("reshape_209_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor keys_121_begin_0 = const()[name = tensor("keys_121_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor keys_121_end_0 = const()[name = tensor("keys_121_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor keys_121_end_mask_0 = const()[name = tensor("keys_121_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor keys_121_squeeze_mask_0 = const()[name = tensor("keys_121_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor keys_121_cast_fp16 = slice_by_index(begin = keys_121_begin_0, end = keys_121_end_0, end_mask = keys_121_end_mask_0, squeeze_mask = keys_121_squeeze_mask_0, x = reshape_209_cast_fp16)[name = tensor("keys_121_cast_fp16")]; + tensor values_121_begin_0 = const()[name = tensor("values_121_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor values_121_end_0 = const()[name = tensor("values_121_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor values_121_end_mask_0 = const()[name = tensor("values_121_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor values_121_squeeze_mask_0 = const()[name = tensor("values_121_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor values_121_cast_fp16 = slice_by_index(begin = values_121_begin_0, end = values_121_end_0, end_mask = values_121_end_mask_0, squeeze_mask = values_121_squeeze_mask_0, x = reshape_209_cast_fp16)[name = tensor("values_121_cast_fp16")]; + tensor var_8878_cast_fp16 = not_equal(x = keys_121_cast_fp16, y = keys_121_cast_fp16)[name = tensor("op_8878_cast_fp16")]; + tensor keys_123_cast_fp16 = select(a = var_504_to_fp16, b = keys_121_cast_fp16, cond = var_8878_cast_fp16)[name = tensor("keys_123_cast_fp16")]; + tensor var_8886_cast_fp16 = not_equal(x = values_121_cast_fp16, y = values_121_cast_fp16)[name = tensor("op_8886_cast_fp16")]; + tensor values_123_cast_fp16 = select(a = var_504_to_fp16, b = values_121_cast_fp16, cond = var_8886_cast_fp16)[name = tensor("values_123_cast_fp16")]; + tensor var_8910 = const()[name = tensor("op_8910"), val = tensor([0, 2, 1, 3])]; + tensor var_8923 = const()[name = tensor("op_8923"), val = tensor([1, 1, 1])]; + tensor var_8924_cast_fp16 = reshape(shape = var_8923, x = position20_to_fp16)[name = tensor("op_8924_cast_fp16")]; + tensor var_8941_to_fp16 = const()[name = tensor("op_8941_to_fp16"), val = tensor(0x1p+0)]; + tensor valid_len_41_cast_fp16 = add(x = var_8924_cast_fp16, y = var_8941_to_fp16)[name = tensor("valid_len_41_cast_fp16")]; + tensor valid_mask_41_cast_fp16 = less(x = k_positions_1_promoted_to_fp16, y = valid_len_41_cast_fp16)[name = tensor("valid_mask_41_cast_fp16")]; + tensor causal_mask_41_cast_fp16 = less_equal(x = k_positions_1_promoted_to_fp16, y = var_8924_cast_fp16)[name = tensor("causal_mask_41_cast_fp16")]; + tensor attn_mask_81 = logical_and(x = valid_mask_41_cast_fp16, y = causal_mask_41_cast_fp16)[name = tensor("attn_mask_81")]; + tensor attn_mask_83_axes_0 = const()[name = tensor("attn_mask_83_axes_0"), val = tensor([1])]; + tensor attn_mask_83 = expand_dims(axes = attn_mask_83_axes_0, x = attn_mask_81)[name = tensor("attn_mask_83")]; + tensor var_8959_transpose_x_0 = const()[name = tensor("op_8959_transpose_x_0"), val = tensor(false)]; + tensor var_8959_transpose_y_0 = const()[name = tensor("op_8959_transpose_y_0"), val = tensor(false)]; + tensor transpose_112_perm_0 = const()[name = tensor("transpose_112_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_113_perm_0 = const()[name = tensor("transpose_113_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_113 = transpose(perm = transpose_113_perm_0, x = keys_123_cast_fp16)[name = tensor("transpose_133")]; + tensor transpose_112 = transpose(perm = transpose_112_perm_0, x = q_123_cast_fp16)[name = tensor("transpose_134")]; + tensor var_8959_cast_fp16 = matmul(transpose_x = var_8959_transpose_x_0, transpose_y = var_8959_transpose_y_0, x = transpose_112, y = transpose_113)[name = tensor("op_8959_cast_fp16")]; + tensor var_8953_to_fp16 = const()[name = tensor("op_8953_to_fp16"), val = tensor([0x1p-3])]; + tensor attn_weights_121_cast_fp16 = mul(x = var_8959_cast_fp16, y = var_8953_to_fp16)[name = tensor("attn_weights_121_cast_fp16")]; + tensor var_8961 = logical_not(x = attn_mask_83)[name = tensor("op_8961")]; + tensor var_8962_to_fp16 = const()[name = tensor("op_8962_to_fp16"), val = tensor(-inf)]; + tensor attn_weights_123_cast_fp16 = select(a = var_8962_to_fp16, b = attn_weights_121_cast_fp16, cond = var_8961)[name = tensor("attn_weights_123_cast_fp16")]; + tensor var_8964 = const()[name = tensor("op_8964"), val = tensor(-1)]; + tensor attn_weights_125_cast_fp16 = softmax(axis = var_8964, x = attn_weights_123_cast_fp16)[name = tensor("attn_weights_125_cast_fp16")]; + tensor attn_output_41_transpose_x_0 = const()[name = tensor("attn_output_41_transpose_x_0"), val = tensor(false)]; + tensor attn_output_41_transpose_y_0 = const()[name = tensor("attn_output_41_transpose_y_0"), val = tensor(false)]; + tensor values_125_cast_fp16 = transpose(perm = var_8910, x = values_123_cast_fp16)[name = tensor("transpose_135")]; + tensor attn_output_41_cast_fp16 = matmul(transpose_x = attn_output_41_transpose_x_0, transpose_y = attn_output_41_transpose_y_0, x = attn_weights_125_cast_fp16, y = values_125_cast_fp16)[name = tensor("attn_output_41_cast_fp16")]; + tensor var_8972 = const()[name = tensor("op_8972"), val = tensor([0, 2, 1, 3])]; + tensor var_8975 = const()[name = tensor("op_8975"), val = tensor([1, 1, 1024])]; + tensor var_8973_cast_fp16 = transpose(perm = var_8972, x = attn_output_41_cast_fp16)[name = tensor("transpose_132")]; + tensor input_205_cast_fp16 = reshape(shape = var_8975, x = var_8973_cast_fp16)[name = tensor("input_205_cast_fp16")]; + tensor linear_82_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = attn20_out_proj_weight_cast_fp16, x = input_205_cast_fp16)[name = tensor("linear_82_cast_fp16")]; + tensor var_8981_to_fp16 = const()[name = tensor("op_8981_to_fp16"), val = tensor(0x1p+0)]; + tensor var_8982_cast_fp16 = add(x = position20_to_fp16, y = var_8981_to_fp16)[name = tensor("op_8982_cast_fp16")]; + tensor var_8982_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_8982_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor input_207_cast_fp16 = add(x = input_203_cast_fp16, y = linear_82_cast_fp16)[name = tensor("input_207_cast_fp16")]; + tensor input_209_axes_0 = const()[name = tensor("input_209_axes_0"), val = tensor([-1])]; + tensor norm20_2_weight_to_fp16 = const()[name = tensor("norm20_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(303760768)))]; + tensor norm20_2_bias_to_fp16 = const()[name = tensor("norm20_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(303762880)))]; + tensor var_8986_to_fp16 = const()[name = tensor("op_8986_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_209_cast_fp16 = layer_norm(axes = input_209_axes_0, beta = norm20_2_bias_to_fp16, epsilon = var_8986_to_fp16, gamma = norm20_2_weight_to_fp16, x = input_207_cast_fp16)[name = tensor("input_209_cast_fp16")]; + tensor linear_83_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = linear20_1_weight_cast_fp16, x = input_209_cast_fp16)[name = tensor("linear_83_cast_fp16")]; + tensor input_211_mode_0 = const()[name = tensor("input_211_mode_0"), val = tensor("EXACT")]; + tensor input_211_cast_fp16 = gelu(mode = input_211_mode_0, x = linear_83_cast_fp16)[name = tensor("input_211_cast_fp16")]; + tensor linear_84_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = linear20_2_weight_cast_fp16, x = input_211_cast_fp16)[name = tensor("linear_84_cast_fp16")]; + tensor input_213_cast_fp16 = add(x = input_207_cast_fp16, y = linear_84_cast_fp16)[name = tensor("input_213_cast_fp16")]; + tensor x_43_axes_0 = const()[name = tensor("x_43_axes_0"), val = tensor([-1])]; + tensor norm21_1_weight_to_fp16 = const()[name = tensor("norm21_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(303764992)))]; + tensor norm21_1_bias_to_fp16 = const()[name = tensor("norm21_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(303767104)))]; + tensor var_9003_to_fp16 = const()[name = tensor("op_9003_to_fp16"), val = tensor(0x1.5p-17)]; + tensor x_43_cast_fp16 = layer_norm(axes = x_43_axes_0, beta = norm21_1_bias_to_fp16, epsilon = var_9003_to_fp16, gamma = norm21_1_weight_to_fp16, x = input_213_cast_fp16)[name = tensor("x_43_cast_fp16")]; + tensor linear_85_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = attn21_in_proj_weight_cast_fp16, x = x_43_cast_fp16)[name = tensor("linear_85_cast_fp16")]; + tensor var_9039 = const()[name = tensor("op_9039"), val = tensor([1, 1, 3, 16, 64])]; + tensor qkv_43_cast_fp16 = reshape(shape = var_9039, x = linear_85_cast_fp16)[name = tensor("qkv_43_cast_fp16")]; + tensor q_127_begin_0 = const()[name = tensor("q_127_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor q_127_end_0 = const()[name = tensor("q_127_end_0"), val = tensor([1, 1, 1, 16, 64])]; + tensor q_127_end_mask_0 = const()[name = tensor("q_127_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor q_127_squeeze_mask_0 = const()[name = tensor("q_127_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor q_127_cast_fp16 = slice_by_index(begin = q_127_begin_0, end = q_127_end_0, end_mask = q_127_end_mask_0, squeeze_mask = q_127_squeeze_mask_0, x = qkv_43_cast_fp16)[name = tensor("q_127_cast_fp16")]; + tensor k_85_begin_0 = const()[name = tensor("k_85_begin_0"), val = tensor([0, 0, 1, 0, 0])]; + tensor k_85_end_0 = const()[name = tensor("k_85_end_0"), val = tensor([1, 1, 2, 16, 64])]; + tensor k_85_end_mask_0 = const()[name = tensor("k_85_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor k_85_squeeze_mask_0 = const()[name = tensor("k_85_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor k_85_cast_fp16 = slice_by_index(begin = k_85_begin_0, end = k_85_end_0, end_mask = k_85_end_mask_0, squeeze_mask = k_85_squeeze_mask_0, x = qkv_43_cast_fp16)[name = tensor("k_85_cast_fp16")]; + tensor v_43_begin_0 = const()[name = tensor("v_43_begin_0"), val = tensor([0, 0, 2, 0, 0])]; + tensor v_43_end_0 = const()[name = tensor("v_43_end_0"), val = tensor([1, 1, 3, 16, 64])]; + tensor v_43_end_mask_0 = const()[name = tensor("v_43_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor v_43_squeeze_mask_0 = const()[name = tensor("v_43_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor v_43_cast_fp16 = slice_by_index(begin = v_43_begin_0, end = v_43_end_0, end_mask = v_43_end_mask_0, squeeze_mask = v_43_squeeze_mask_0, x = qkv_43_cast_fp16)[name = tensor("v_43_cast_fp16")]; + tensor var_9143 = const()[name = tensor("op_9143"), val = tensor([1, 1, 1, 1])]; + tensor position21_to_fp16_dtype_0 = const()[name = tensor("position21_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor position21_to_fp16 = cast(dtype = position21_to_fp16_dtype_0, x = position21)[name = tensor("cast_448")]; + tensor ts_131_cast_fp16 = reshape(shape = var_9143, x = position21_to_fp16)[name = tensor("ts_131_cast_fp16")]; + tensor var_9147 = const()[name = tensor("op_9147"), val = tensor([1, 1, 16, 32, 2])]; + tensor q_complex_43_cast_fp16 = reshape(shape = var_9147, x = q_127_cast_fp16)[name = tensor("q_complex_43_cast_fp16")]; + tensor var_9151 = const()[name = tensor("op_9151"), val = tensor([1, 1, 16, 32, 2])]; + tensor k_complex_43_cast_fp16 = reshape(shape = var_9151, x = k_85_cast_fp16)[name = tensor("k_complex_43_cast_fp16")]; + tensor var_9155_begin_0 = const()[name = tensor("op_9155_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_9155_end_0 = const()[name = tensor("op_9155_end_0"), val = tensor([1, 1, 16, 32, 1])]; + tensor var_9155_end_mask_0 = const()[name = tensor("op_9155_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_9155_squeeze_mask_0 = const()[name = tensor("op_9155_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_9155_cast_fp16 = slice_by_index(begin = var_9155_begin_0, end = var_9155_end_0, end_mask = var_9155_end_mask_0, squeeze_mask = var_9155_squeeze_mask_0, x = q_complex_43_cast_fp16)[name = tensor("op_9155_cast_fp16")]; + tensor var_9163_begin_0 = const()[name = tensor("op_9163_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_9163_end_0 = const()[name = tensor("op_9163_end_0"), val = tensor([1, 1, 16, 32, 2])]; + tensor var_9163_end_mask_0 = const()[name = tensor("op_9163_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_9163_squeeze_mask_0 = const()[name = tensor("op_9163_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_9163_cast_fp16 = slice_by_index(begin = var_9163_begin_0, end = var_9163_end_0, end_mask = var_9163_end_mask_0, squeeze_mask = var_9163_squeeze_mask_0, x = q_complex_43_cast_fp16)[name = tensor("op_9163_cast_fp16")]; + tensor var_9171_begin_0 = const()[name = tensor("op_9171_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_9171_end_0 = const()[name = tensor("op_9171_end_0"), val = tensor([1, 1, 16, 32, 1])]; + tensor var_9171_end_mask_0 = const()[name = tensor("op_9171_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_9171_squeeze_mask_0 = const()[name = tensor("op_9171_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_9171_cast_fp16 = slice_by_index(begin = var_9171_begin_0, end = var_9171_end_0, end_mask = var_9171_end_mask_0, squeeze_mask = var_9171_squeeze_mask_0, x = k_complex_43_cast_fp16)[name = tensor("op_9171_cast_fp16")]; + tensor var_9179_begin_0 = const()[name = tensor("op_9179_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_9179_end_0 = const()[name = tensor("op_9179_end_0"), val = tensor([1, 1, 16, 32, 2])]; + tensor var_9179_end_mask_0 = const()[name = tensor("op_9179_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_9179_squeeze_mask_0 = const()[name = tensor("op_9179_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_9179_cast_fp16 = slice_by_index(begin = var_9179_begin_0, end = var_9179_end_0, end_mask = var_9179_end_mask_0, squeeze_mask = var_9179_squeeze_mask_0, x = k_complex_43_cast_fp16)[name = tensor("op_9179_cast_fp16")]; + tensor freqs_43_to_fp16 = const()[name = tensor("freqs_43_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(303769216)))]; + tensor var_9185_cast_fp16 = mul(x = freqs_43_to_fp16, y = ts_131_cast_fp16)[name = tensor("op_9185_cast_fp16")]; + tensor rotr_43_cast_fp16 = cos(x = var_9185_cast_fp16)[name = tensor("rotr_43_cast_fp16")]; + tensor roti_43_cast_fp16 = sin(x = var_9185_cast_fp16)[name = tensor("roti_43_cast_fp16")]; + tensor var_9189_cast_fp16 = mul(x = var_9155_cast_fp16, y = rotr_43_cast_fp16)[name = tensor("op_9189_cast_fp16")]; + tensor var_9190_cast_fp16 = mul(x = var_9163_cast_fp16, y = roti_43_cast_fp16)[name = tensor("op_9190_cast_fp16")]; + tensor qor_85_cast_fp16 = sub(x = var_9189_cast_fp16, y = var_9190_cast_fp16)[name = tensor("qor_85_cast_fp16")]; + tensor var_9193_cast_fp16 = mul(x = var_9155_cast_fp16, y = roti_43_cast_fp16)[name = tensor("op_9193_cast_fp16")]; + tensor var_9194_cast_fp16 = mul(x = var_9163_cast_fp16, y = rotr_43_cast_fp16)[name = tensor("op_9194_cast_fp16")]; + tensor qoi_85_cast_fp16 = add(x = var_9193_cast_fp16, y = var_9194_cast_fp16)[name = tensor("qoi_85_cast_fp16")]; + tensor var_9197_cast_fp16 = mul(x = var_9171_cast_fp16, y = rotr_43_cast_fp16)[name = tensor("op_9197_cast_fp16")]; + tensor var_9198_cast_fp16 = mul(x = var_9179_cast_fp16, y = roti_43_cast_fp16)[name = tensor("op_9198_cast_fp16")]; + tensor kor_85_cast_fp16 = sub(x = var_9197_cast_fp16, y = var_9198_cast_fp16)[name = tensor("kor_85_cast_fp16")]; + tensor var_9201_cast_fp16 = mul(x = var_9171_cast_fp16, y = roti_43_cast_fp16)[name = tensor("op_9201_cast_fp16")]; + tensor var_9202_cast_fp16 = mul(x = var_9179_cast_fp16, y = rotr_43_cast_fp16)[name = tensor("op_9202_cast_fp16")]; + tensor koi_85_cast_fp16 = add(x = var_9201_cast_fp16, y = var_9202_cast_fp16)[name = tensor("koi_85_cast_fp16")]; + tensor qo_43_axis_0 = const()[name = tensor("qo_43_axis_0"), val = tensor(-1)]; + tensor qo_43_cast_fp16 = stack(axis = qo_43_axis_0, values = (qor_85_cast_fp16, qoi_85_cast_fp16))[name = tensor("qo_43_cast_fp16")]; + tensor ko_43_axis_0 = const()[name = tensor("ko_43_axis_0"), val = tensor(-1)]; + tensor ko_43_cast_fp16 = stack(axis = ko_43_axis_0, values = (kor_85_cast_fp16, koi_85_cast_fp16))[name = tensor("ko_43_cast_fp16")]; + tensor var_9231 = const()[name = tensor("op_9231"), val = tensor([1, 1, 16, 64])]; + tensor q_129_cast_fp16 = reshape(shape = var_9231, x = qo_43_cast_fp16)[name = tensor("q_129_cast_fp16")]; + tensor var_9233 = const()[name = tensor("op_9233"), val = tensor([1, 1, 16, 64])]; + tensor k_87_cast_fp16 = reshape(shape = var_9233, x = ko_43_cast_fp16)[name = tensor("k_87_cast_fp16")]; + tensor _inversed_9255_y_0_to_fp16 = const()[name = tensor("_inversed_9255_y_0_to_fp16"), val = tensor(0x1p-9)]; + tensor _inversed_9255_cast_fp16 = mul(x = ts_131_cast_fp16, y = _inversed_9255_y_0_to_fp16)[name = tensor("_inversed_9255_cast_fp16")]; + tensor var_9256_cast_fp16 = floor(x = _inversed_9255_cast_fp16)[name = tensor("op_9256_cast_fp16")]; + tensor var_9257_to_fp16 = const()[name = tensor("op_9257_to_fp16"), val = tensor(0x1p+9)]; + tensor var_9258_cast_fp16 = mul(x = var_9256_cast_fp16, y = var_9257_to_fp16)[name = tensor("op_9258_cast_fp16")]; + tensor write_indices_float_87_cast_fp16 = sub(x = ts_131_cast_fp16, y = var_9258_cast_fp16)[name = tensor("write_indices_float_87_cast_fp16")]; + tensor var_9265_dtype_0 = const()[name = tensor("op_9265_dtype_0"), val = tensor("int32")]; + tensor write_indices_43_reps_0 = const()[name = tensor("write_indices_43_reps_0"), val = tensor([1, 1, 16, 64])]; + tensor write_indices_float_87_cast_fp16_to_int32 = cast(dtype = var_9265_dtype_0, x = write_indices_float_87_cast_fp16)[name = tensor("cast_447")]; + tensor write_indices_43 = tile(reps = write_indices_43_reps_0, x = write_indices_float_87_cast_fp16_to_int32)[name = tensor("write_indices_43")]; + tensor var_9273_begin_0 = const()[name = tensor("op_9273_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_9273_end_0 = const()[name = tensor("op_9273_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor var_9273_end_mask_0 = const()[name = tensor("op_9273_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_9273_squeeze_mask_0 = const()[name = tensor("op_9273_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor cache21_to_fp16_dtype_0 = const()[name = tensor("cache21_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor cache21_to_fp16 = cast(dtype = cache21_to_fp16_dtype_0, x = cache21)[name = tensor("cast_446")]; + tensor var_9273_cast_fp16 = slice_by_index(begin = var_9273_begin_0, end = var_9273_end_0, end_mask = var_9273_end_mask_0, squeeze_mask = var_9273_squeeze_mask_0, x = cache21_to_fp16)[name = tensor("op_9273_cast_fp16")]; + tensor var_9275_axis_0 = const()[name = tensor("op_9275_axis_0"), val = tensor(1)]; + tensor var_9275_mode_0 = const()[name = tensor("op_9275_mode_0"), val = tensor("update")]; + tensor var_9275_validate_indices_0 = const()[name = tensor("op_9275_validate_indices_0"), val = tensor(false)]; + tensor var_9275_cast_fp16 = scatter_along_axis(axis = var_9275_axis_0, data = var_9273_cast_fp16, indices = write_indices_43, mode = var_9275_mode_0, updates = k_87_cast_fp16, validate_indices = var_9275_validate_indices_0)[name = tensor("op_9275_cast_fp16")]; + tensor concat_149 = const()[name = tensor("concat_149"), val = tensor([0, 0, 0, 0, 0])]; + tensor concat_150 = const()[name = tensor("concat_150"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_43_internal_tensor_assign_1_stride_0 = const()[name = tensor("new_cache_43_internal_tensor_assign_1_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_43_internal_tensor_assign_1_begin_mask_0 = const()[name = tensor("new_cache_43_internal_tensor_assign_1_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_43_internal_tensor_assign_1_end_mask_0 = const()[name = tensor("new_cache_43_internal_tensor_assign_1_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_43_internal_tensor_assign_1_squeeze_mask_0 = const()[name = tensor("new_cache_43_internal_tensor_assign_1_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_90 = const()[name = tensor("shape_90"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_42 = const()[name = tensor("reduce_prod_42"), val = tensor(1048576)]; + tensor range_1d_42_start_0 = const()[name = tensor("range_1d_42_start_0"), val = tensor(0)]; + tensor range_1d_42_step_0 = const()[name = tensor("range_1d_42_step_0"), val = tensor(1)]; + tensor range_1d_42 = range_1d(end = reduce_prod_42, start = range_1d_42_start_0, step = range_1d_42_step_0)[name = tensor("range_1d_42")]; + tensor reshape_210 = reshape(shape = shape_90, x = range_1d_42)[name = tensor("reshape_210")]; + tensor slice_by_index_42 = slice_by_index(begin = concat_149, begin_mask = new_cache_43_internal_tensor_assign_1_begin_mask_0, end = concat_150, end_mask = new_cache_43_internal_tensor_assign_1_end_mask_0, squeeze_mask = new_cache_43_internal_tensor_assign_1_squeeze_mask_0, stride = new_cache_43_internal_tensor_assign_1_stride_0, x = reshape_210)[name = tensor("slice_by_index_42")]; + tensor reshape_211_shape_0 = const()[name = tensor("reshape_211_shape_0"), val = tensor([-1])]; + tensor reshape_211 = reshape(shape = reshape_211_shape_0, x = slice_by_index_42)[name = tensor("reshape_211")]; + tensor reshape_212_shape_0 = const()[name = tensor("reshape_212_shape_0"), val = tensor([-1])]; + tensor reshape_212_cast_fp16 = reshape(shape = reshape_212_shape_0, x = var_9275_cast_fp16)[name = tensor("reshape_212_cast_fp16")]; + tensor reshape_213_shape_0 = const()[name = tensor("reshape_213_shape_0"), val = tensor([-1])]; + tensor reshape_213_cast_fp16 = reshape(shape = reshape_213_shape_0, x = cache21_to_fp16)[name = tensor("reshape_213_cast_fp16")]; + tensor scatter_42_mode_0 = const()[name = tensor("scatter_42_mode_0"), val = tensor("update")]; + tensor scatter_42_axis_0 = const()[name = tensor("scatter_42_axis_0"), val = tensor(0)]; + tensor scatter_42_validate_indices_0 = const()[name = tensor("scatter_42_validate_indices_0"), val = tensor(false)]; + tensor scatter_42_cast_fp16 = scatter(axis = scatter_42_axis_0, data = reshape_213_cast_fp16, indices = reshape_211, mode = scatter_42_mode_0, updates = reshape_212_cast_fp16, validate_indices = scatter_42_validate_indices_0)[name = tensor("scatter_42_cast_fp16")]; + tensor reshape_214_cast_fp16 = reshape(shape = shape_90, x = scatter_42_cast_fp16)[name = tensor("reshape_214_cast_fp16")]; + tensor var_9283_begin_0 = const()[name = tensor("op_9283_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor var_9283_end_0 = const()[name = tensor("op_9283_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor var_9283_end_mask_0 = const()[name = tensor("op_9283_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_9283_squeeze_mask_0 = const()[name = tensor("op_9283_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor var_9283_cast_fp16 = slice_by_index(begin = var_9283_begin_0, end = var_9283_end_0, end_mask = var_9283_end_mask_0, squeeze_mask = var_9283_squeeze_mask_0, x = reshape_214_cast_fp16)[name = tensor("op_9283_cast_fp16")]; + tensor var_9285_axis_0 = const()[name = tensor("op_9285_axis_0"), val = tensor(1)]; + tensor var_9285_mode_0 = const()[name = tensor("op_9285_mode_0"), val = tensor("update")]; + tensor var_9285_validate_indices_0 = const()[name = tensor("op_9285_validate_indices_0"), val = tensor(false)]; + tensor var_9285_cast_fp16 = scatter_along_axis(axis = var_9285_axis_0, data = var_9283_cast_fp16, indices = write_indices_43, mode = var_9285_mode_0, updates = v_43_cast_fp16, validate_indices = var_9285_validate_indices_0)[name = tensor("op_9285_cast_fp16")]; + tensor concat_151 = const()[name = tensor("concat_151"), val = tensor([1, 0, 0, 0, 0])]; + tensor concat_152 = const()[name = tensor("concat_152"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_43_internal_tensor_assign_2_stride_0 = const()[name = tensor("new_cache_43_internal_tensor_assign_2_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_43_internal_tensor_assign_2_begin_mask_0 = const()[name = tensor("new_cache_43_internal_tensor_assign_2_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_43_internal_tensor_assign_2_end_mask_0 = const()[name = tensor("new_cache_43_internal_tensor_assign_2_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_43_internal_tensor_assign_2_squeeze_mask_0 = const()[name = tensor("new_cache_43_internal_tensor_assign_2_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_91 = const()[name = tensor("shape_91"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_43 = const()[name = tensor("reduce_prod_43"), val = tensor(1048576)]; + tensor range_1d_43_start_0 = const()[name = tensor("range_1d_43_start_0"), val = tensor(0)]; + tensor range_1d_43_step_0 = const()[name = tensor("range_1d_43_step_0"), val = tensor(1)]; + tensor range_1d_43 = range_1d(end = reduce_prod_43, start = range_1d_43_start_0, step = range_1d_43_step_0)[name = tensor("range_1d_43")]; + tensor reshape_215 = reshape(shape = shape_91, x = range_1d_43)[name = tensor("reshape_215")]; + tensor slice_by_index_43 = slice_by_index(begin = concat_151, begin_mask = new_cache_43_internal_tensor_assign_2_begin_mask_0, end = concat_152, end_mask = new_cache_43_internal_tensor_assign_2_end_mask_0, squeeze_mask = new_cache_43_internal_tensor_assign_2_squeeze_mask_0, stride = new_cache_43_internal_tensor_assign_2_stride_0, x = reshape_215)[name = tensor("slice_by_index_43")]; + tensor reshape_216_shape_0 = const()[name = tensor("reshape_216_shape_0"), val = tensor([-1])]; + tensor reshape_216 = reshape(shape = reshape_216_shape_0, x = slice_by_index_43)[name = tensor("reshape_216")]; + tensor reshape_217_shape_0 = const()[name = tensor("reshape_217_shape_0"), val = tensor([-1])]; + tensor reshape_217_cast_fp16 = reshape(shape = reshape_217_shape_0, x = var_9285_cast_fp16)[name = tensor("reshape_217_cast_fp16")]; + tensor reshape_218_shape_0 = const()[name = tensor("reshape_218_shape_0"), val = tensor([-1])]; + tensor reshape_218_cast_fp16 = reshape(shape = reshape_218_shape_0, x = reshape_214_cast_fp16)[name = tensor("reshape_218_cast_fp16")]; + tensor scatter_43_mode_0 = const()[name = tensor("scatter_43_mode_0"), val = tensor("update")]; + tensor scatter_43_axis_0 = const()[name = tensor("scatter_43_axis_0"), val = tensor(0)]; + tensor scatter_43_validate_indices_0 = const()[name = tensor("scatter_43_validate_indices_0"), val = tensor(false)]; + tensor scatter_43_cast_fp16 = scatter(axis = scatter_43_axis_0, data = reshape_218_cast_fp16, indices = reshape_216, mode = scatter_43_mode_0, updates = reshape_217_cast_fp16, validate_indices = scatter_43_validate_indices_0)[name = tensor("scatter_43_cast_fp16")]; + tensor reshape_219_cast_fp16 = reshape(shape = shape_91, x = scatter_43_cast_fp16)[name = tensor("reshape_219_cast_fp16")]; + tensor reshape_219_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("reshape_219_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor keys_127_begin_0 = const()[name = tensor("keys_127_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor keys_127_end_0 = const()[name = tensor("keys_127_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor keys_127_end_mask_0 = const()[name = tensor("keys_127_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor keys_127_squeeze_mask_0 = const()[name = tensor("keys_127_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor keys_127_cast_fp16 = slice_by_index(begin = keys_127_begin_0, end = keys_127_end_0, end_mask = keys_127_end_mask_0, squeeze_mask = keys_127_squeeze_mask_0, x = reshape_219_cast_fp16)[name = tensor("keys_127_cast_fp16")]; + tensor values_127_begin_0 = const()[name = tensor("values_127_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor values_127_end_0 = const()[name = tensor("values_127_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor values_127_end_mask_0 = const()[name = tensor("values_127_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor values_127_squeeze_mask_0 = const()[name = tensor("values_127_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor values_127_cast_fp16 = slice_by_index(begin = values_127_begin_0, end = values_127_end_0, end_mask = values_127_end_mask_0, squeeze_mask = values_127_squeeze_mask_0, x = reshape_219_cast_fp16)[name = tensor("values_127_cast_fp16")]; + tensor var_9297_cast_fp16 = not_equal(x = keys_127_cast_fp16, y = keys_127_cast_fp16)[name = tensor("op_9297_cast_fp16")]; + tensor keys_129_cast_fp16 = select(a = var_504_to_fp16, b = keys_127_cast_fp16, cond = var_9297_cast_fp16)[name = tensor("keys_129_cast_fp16")]; + tensor var_9305_cast_fp16 = not_equal(x = values_127_cast_fp16, y = values_127_cast_fp16)[name = tensor("op_9305_cast_fp16")]; + tensor values_129_cast_fp16 = select(a = var_504_to_fp16, b = values_127_cast_fp16, cond = var_9305_cast_fp16)[name = tensor("values_129_cast_fp16")]; + tensor var_9329 = const()[name = tensor("op_9329"), val = tensor([0, 2, 1, 3])]; + tensor var_9342 = const()[name = tensor("op_9342"), val = tensor([1, 1, 1])]; + tensor var_9343_cast_fp16 = reshape(shape = var_9342, x = position21_to_fp16)[name = tensor("op_9343_cast_fp16")]; + tensor var_9360_to_fp16 = const()[name = tensor("op_9360_to_fp16"), val = tensor(0x1p+0)]; + tensor valid_len_43_cast_fp16 = add(x = var_9343_cast_fp16, y = var_9360_to_fp16)[name = tensor("valid_len_43_cast_fp16")]; + tensor valid_mask_43_cast_fp16 = less(x = k_positions_1_promoted_to_fp16, y = valid_len_43_cast_fp16)[name = tensor("valid_mask_43_cast_fp16")]; + tensor causal_mask_43_cast_fp16 = less_equal(x = k_positions_1_promoted_to_fp16, y = var_9343_cast_fp16)[name = tensor("causal_mask_43_cast_fp16")]; + tensor attn_mask_85 = logical_and(x = valid_mask_43_cast_fp16, y = causal_mask_43_cast_fp16)[name = tensor("attn_mask_85")]; + tensor attn_mask_87_axes_0 = const()[name = tensor("attn_mask_87_axes_0"), val = tensor([1])]; + tensor attn_mask_87 = expand_dims(axes = attn_mask_87_axes_0, x = attn_mask_85)[name = tensor("attn_mask_87")]; + tensor var_9378_transpose_x_0 = const()[name = tensor("op_9378_transpose_x_0"), val = tensor(false)]; + tensor var_9378_transpose_y_0 = const()[name = tensor("op_9378_transpose_y_0"), val = tensor(false)]; + tensor transpose_114_perm_0 = const()[name = tensor("transpose_114_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_115_perm_0 = const()[name = tensor("transpose_115_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_115 = transpose(perm = transpose_115_perm_0, x = keys_129_cast_fp16)[name = tensor("transpose_129")]; + tensor transpose_114 = transpose(perm = transpose_114_perm_0, x = q_129_cast_fp16)[name = tensor("transpose_130")]; + tensor var_9378_cast_fp16 = matmul(transpose_x = var_9378_transpose_x_0, transpose_y = var_9378_transpose_y_0, x = transpose_114, y = transpose_115)[name = tensor("op_9378_cast_fp16")]; + tensor var_9372_to_fp16 = const()[name = tensor("op_9372_to_fp16"), val = tensor([0x1p-3])]; + tensor attn_weights_127_cast_fp16 = mul(x = var_9378_cast_fp16, y = var_9372_to_fp16)[name = tensor("attn_weights_127_cast_fp16")]; + tensor var_9380 = logical_not(x = attn_mask_87)[name = tensor("op_9380")]; + tensor var_9381_to_fp16 = const()[name = tensor("op_9381_to_fp16"), val = tensor(-inf)]; + tensor attn_weights_129_cast_fp16 = select(a = var_9381_to_fp16, b = attn_weights_127_cast_fp16, cond = var_9380)[name = tensor("attn_weights_129_cast_fp16")]; + tensor var_9383 = const()[name = tensor("op_9383"), val = tensor(-1)]; + tensor attn_weights_131_cast_fp16 = softmax(axis = var_9383, x = attn_weights_129_cast_fp16)[name = tensor("attn_weights_131_cast_fp16")]; + tensor attn_output_43_transpose_x_0 = const()[name = tensor("attn_output_43_transpose_x_0"), val = tensor(false)]; + tensor attn_output_43_transpose_y_0 = const()[name = tensor("attn_output_43_transpose_y_0"), val = tensor(false)]; + tensor values_131_cast_fp16 = transpose(perm = var_9329, x = values_129_cast_fp16)[name = tensor("transpose_131")]; + tensor attn_output_43_cast_fp16 = matmul(transpose_x = attn_output_43_transpose_x_0, transpose_y = attn_output_43_transpose_y_0, x = attn_weights_131_cast_fp16, y = values_131_cast_fp16)[name = tensor("attn_output_43_cast_fp16")]; + tensor var_9391 = const()[name = tensor("op_9391"), val = tensor([0, 2, 1, 3])]; + tensor var_9394 = const()[name = tensor("op_9394"), val = tensor([1, 1, 1024])]; + tensor var_9392_cast_fp16 = transpose(perm = var_9391, x = attn_output_43_cast_fp16)[name = tensor("transpose_128")]; + tensor input_215_cast_fp16 = reshape(shape = var_9394, x = var_9392_cast_fp16)[name = tensor("input_215_cast_fp16")]; + tensor linear_86_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = attn21_out_proj_weight_cast_fp16, x = input_215_cast_fp16)[name = tensor("linear_86_cast_fp16")]; + tensor var_9400_to_fp16 = const()[name = tensor("op_9400_to_fp16"), val = tensor(0x1p+0)]; + tensor var_9401_cast_fp16 = add(x = position21_to_fp16, y = var_9400_to_fp16)[name = tensor("op_9401_cast_fp16")]; + tensor var_9401_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_9401_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor input_217_cast_fp16 = add(x = input_213_cast_fp16, y = linear_86_cast_fp16)[name = tensor("input_217_cast_fp16")]; + tensor input_219_axes_0 = const()[name = tensor("input_219_axes_0"), val = tensor([-1])]; + tensor norm21_2_weight_to_fp16 = const()[name = tensor("norm21_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(303769344)))]; + tensor norm21_2_bias_to_fp16 = const()[name = tensor("norm21_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(303771456)))]; + tensor var_9405_to_fp16 = const()[name = tensor("op_9405_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_219_cast_fp16 = layer_norm(axes = input_219_axes_0, beta = norm21_2_bias_to_fp16, epsilon = var_9405_to_fp16, gamma = norm21_2_weight_to_fp16, x = input_217_cast_fp16)[name = tensor("input_219_cast_fp16")]; + tensor linear_87_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = linear21_1_weight_cast_fp16, x = input_219_cast_fp16)[name = tensor("linear_87_cast_fp16")]; + tensor input_221_mode_0 = const()[name = tensor("input_221_mode_0"), val = tensor("EXACT")]; + tensor input_221_cast_fp16 = gelu(mode = input_221_mode_0, x = linear_87_cast_fp16)[name = tensor("input_221_cast_fp16")]; + tensor linear_88_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = linear21_2_weight_cast_fp16, x = input_221_cast_fp16)[name = tensor("linear_88_cast_fp16")]; + tensor input_223_cast_fp16 = add(x = input_217_cast_fp16, y = linear_88_cast_fp16)[name = tensor("input_223_cast_fp16")]; + tensor x_45_axes_0 = const()[name = tensor("x_45_axes_0"), val = tensor([-1])]; + tensor norm22_1_weight_to_fp16 = const()[name = tensor("norm22_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(303773568)))]; + tensor norm22_1_bias_to_fp16 = const()[name = tensor("norm22_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(303775680)))]; + tensor var_9422_to_fp16 = const()[name = tensor("op_9422_to_fp16"), val = tensor(0x1.5p-17)]; + tensor x_45_cast_fp16 = layer_norm(axes = x_45_axes_0, beta = norm22_1_bias_to_fp16, epsilon = var_9422_to_fp16, gamma = norm22_1_weight_to_fp16, x = input_223_cast_fp16)[name = tensor("x_45_cast_fp16")]; + tensor linear_89_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = attn22_in_proj_weight_cast_fp16, x = x_45_cast_fp16)[name = tensor("linear_89_cast_fp16")]; + tensor var_9458 = const()[name = tensor("op_9458"), val = tensor([1, 1, 3, 16, 64])]; + tensor qkv_45_cast_fp16 = reshape(shape = var_9458, x = linear_89_cast_fp16)[name = tensor("qkv_45_cast_fp16")]; + tensor q_133_begin_0 = const()[name = tensor("q_133_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor q_133_end_0 = const()[name = tensor("q_133_end_0"), val = tensor([1, 1, 1, 16, 64])]; + tensor q_133_end_mask_0 = const()[name = tensor("q_133_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor q_133_squeeze_mask_0 = const()[name = tensor("q_133_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor q_133_cast_fp16 = slice_by_index(begin = q_133_begin_0, end = q_133_end_0, end_mask = q_133_end_mask_0, squeeze_mask = q_133_squeeze_mask_0, x = qkv_45_cast_fp16)[name = tensor("q_133_cast_fp16")]; + tensor k_89_begin_0 = const()[name = tensor("k_89_begin_0"), val = tensor([0, 0, 1, 0, 0])]; + tensor k_89_end_0 = const()[name = tensor("k_89_end_0"), val = tensor([1, 1, 2, 16, 64])]; + tensor k_89_end_mask_0 = const()[name = tensor("k_89_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor k_89_squeeze_mask_0 = const()[name = tensor("k_89_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor k_89_cast_fp16 = slice_by_index(begin = k_89_begin_0, end = k_89_end_0, end_mask = k_89_end_mask_0, squeeze_mask = k_89_squeeze_mask_0, x = qkv_45_cast_fp16)[name = tensor("k_89_cast_fp16")]; + tensor v_45_begin_0 = const()[name = tensor("v_45_begin_0"), val = tensor([0, 0, 2, 0, 0])]; + tensor v_45_end_0 = const()[name = tensor("v_45_end_0"), val = tensor([1, 1, 3, 16, 64])]; + tensor v_45_end_mask_0 = const()[name = tensor("v_45_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor v_45_squeeze_mask_0 = const()[name = tensor("v_45_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor v_45_cast_fp16 = slice_by_index(begin = v_45_begin_0, end = v_45_end_0, end_mask = v_45_end_mask_0, squeeze_mask = v_45_squeeze_mask_0, x = qkv_45_cast_fp16)[name = tensor("v_45_cast_fp16")]; + tensor var_9562 = const()[name = tensor("op_9562"), val = tensor([1, 1, 1, 1])]; + tensor position22_to_fp16_dtype_0 = const()[name = tensor("position22_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor position22_to_fp16 = cast(dtype = position22_to_fp16_dtype_0, x = position22)[name = tensor("cast_443")]; + tensor ts_137_cast_fp16 = reshape(shape = var_9562, x = position22_to_fp16)[name = tensor("ts_137_cast_fp16")]; + tensor var_9566 = const()[name = tensor("op_9566"), val = tensor([1, 1, 16, 32, 2])]; + tensor q_complex_45_cast_fp16 = reshape(shape = var_9566, x = q_133_cast_fp16)[name = tensor("q_complex_45_cast_fp16")]; + tensor var_9570 = const()[name = tensor("op_9570"), val = tensor([1, 1, 16, 32, 2])]; + tensor k_complex_45_cast_fp16 = reshape(shape = var_9570, x = k_89_cast_fp16)[name = tensor("k_complex_45_cast_fp16")]; + tensor var_9574_begin_0 = const()[name = tensor("op_9574_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_9574_end_0 = const()[name = tensor("op_9574_end_0"), val = tensor([1, 1, 16, 32, 1])]; + tensor var_9574_end_mask_0 = const()[name = tensor("op_9574_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_9574_squeeze_mask_0 = const()[name = tensor("op_9574_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_9574_cast_fp16 = slice_by_index(begin = var_9574_begin_0, end = var_9574_end_0, end_mask = var_9574_end_mask_0, squeeze_mask = var_9574_squeeze_mask_0, x = q_complex_45_cast_fp16)[name = tensor("op_9574_cast_fp16")]; + tensor var_9582_begin_0 = const()[name = tensor("op_9582_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_9582_end_0 = const()[name = tensor("op_9582_end_0"), val = tensor([1, 1, 16, 32, 2])]; + tensor var_9582_end_mask_0 = const()[name = tensor("op_9582_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_9582_squeeze_mask_0 = const()[name = tensor("op_9582_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_9582_cast_fp16 = slice_by_index(begin = var_9582_begin_0, end = var_9582_end_0, end_mask = var_9582_end_mask_0, squeeze_mask = var_9582_squeeze_mask_0, x = q_complex_45_cast_fp16)[name = tensor("op_9582_cast_fp16")]; + tensor var_9590_begin_0 = const()[name = tensor("op_9590_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_9590_end_0 = const()[name = tensor("op_9590_end_0"), val = tensor([1, 1, 16, 32, 1])]; + tensor var_9590_end_mask_0 = const()[name = tensor("op_9590_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_9590_squeeze_mask_0 = const()[name = tensor("op_9590_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_9590_cast_fp16 = slice_by_index(begin = var_9590_begin_0, end = var_9590_end_0, end_mask = var_9590_end_mask_0, squeeze_mask = var_9590_squeeze_mask_0, x = k_complex_45_cast_fp16)[name = tensor("op_9590_cast_fp16")]; + tensor var_9598_begin_0 = const()[name = tensor("op_9598_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_9598_end_0 = const()[name = tensor("op_9598_end_0"), val = tensor([1, 1, 16, 32, 2])]; + tensor var_9598_end_mask_0 = const()[name = tensor("op_9598_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_9598_squeeze_mask_0 = const()[name = tensor("op_9598_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_9598_cast_fp16 = slice_by_index(begin = var_9598_begin_0, end = var_9598_end_0, end_mask = var_9598_end_mask_0, squeeze_mask = var_9598_squeeze_mask_0, x = k_complex_45_cast_fp16)[name = tensor("op_9598_cast_fp16")]; + tensor freqs_45_to_fp16 = const()[name = tensor("freqs_45_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(303777792)))]; + tensor var_9604_cast_fp16 = mul(x = freqs_45_to_fp16, y = ts_137_cast_fp16)[name = tensor("op_9604_cast_fp16")]; + tensor rotr_45_cast_fp16 = cos(x = var_9604_cast_fp16)[name = tensor("rotr_45_cast_fp16")]; + tensor roti_45_cast_fp16 = sin(x = var_9604_cast_fp16)[name = tensor("roti_45_cast_fp16")]; + tensor var_9608_cast_fp16 = mul(x = var_9574_cast_fp16, y = rotr_45_cast_fp16)[name = tensor("op_9608_cast_fp16")]; + tensor var_9609_cast_fp16 = mul(x = var_9582_cast_fp16, y = roti_45_cast_fp16)[name = tensor("op_9609_cast_fp16")]; + tensor qor_89_cast_fp16 = sub(x = var_9608_cast_fp16, y = var_9609_cast_fp16)[name = tensor("qor_89_cast_fp16")]; + tensor var_9612_cast_fp16 = mul(x = var_9574_cast_fp16, y = roti_45_cast_fp16)[name = tensor("op_9612_cast_fp16")]; + tensor var_9613_cast_fp16 = mul(x = var_9582_cast_fp16, y = rotr_45_cast_fp16)[name = tensor("op_9613_cast_fp16")]; + tensor qoi_89_cast_fp16 = add(x = var_9612_cast_fp16, y = var_9613_cast_fp16)[name = tensor("qoi_89_cast_fp16")]; + tensor var_9616_cast_fp16 = mul(x = var_9590_cast_fp16, y = rotr_45_cast_fp16)[name = tensor("op_9616_cast_fp16")]; + tensor var_9617_cast_fp16 = mul(x = var_9598_cast_fp16, y = roti_45_cast_fp16)[name = tensor("op_9617_cast_fp16")]; + tensor kor_89_cast_fp16 = sub(x = var_9616_cast_fp16, y = var_9617_cast_fp16)[name = tensor("kor_89_cast_fp16")]; + tensor var_9620_cast_fp16 = mul(x = var_9590_cast_fp16, y = roti_45_cast_fp16)[name = tensor("op_9620_cast_fp16")]; + tensor var_9621_cast_fp16 = mul(x = var_9598_cast_fp16, y = rotr_45_cast_fp16)[name = tensor("op_9621_cast_fp16")]; + tensor koi_89_cast_fp16 = add(x = var_9620_cast_fp16, y = var_9621_cast_fp16)[name = tensor("koi_89_cast_fp16")]; + tensor qo_45_axis_0 = const()[name = tensor("qo_45_axis_0"), val = tensor(-1)]; + tensor qo_45_cast_fp16 = stack(axis = qo_45_axis_0, values = (qor_89_cast_fp16, qoi_89_cast_fp16))[name = tensor("qo_45_cast_fp16")]; + tensor ko_45_axis_0 = const()[name = tensor("ko_45_axis_0"), val = tensor(-1)]; + tensor ko_45_cast_fp16 = stack(axis = ko_45_axis_0, values = (kor_89_cast_fp16, koi_89_cast_fp16))[name = tensor("ko_45_cast_fp16")]; + tensor var_9650 = const()[name = tensor("op_9650"), val = tensor([1, 1, 16, 64])]; + tensor q_135_cast_fp16 = reshape(shape = var_9650, x = qo_45_cast_fp16)[name = tensor("q_135_cast_fp16")]; + tensor var_9652 = const()[name = tensor("op_9652"), val = tensor([1, 1, 16, 64])]; + tensor k_91_cast_fp16 = reshape(shape = var_9652, x = ko_45_cast_fp16)[name = tensor("k_91_cast_fp16")]; + tensor _inversed_9674_y_0_to_fp16 = const()[name = tensor("_inversed_9674_y_0_to_fp16"), val = tensor(0x1p-9)]; + tensor _inversed_9674_cast_fp16 = mul(x = ts_137_cast_fp16, y = _inversed_9674_y_0_to_fp16)[name = tensor("_inversed_9674_cast_fp16")]; + tensor var_9675_cast_fp16 = floor(x = _inversed_9674_cast_fp16)[name = tensor("op_9675_cast_fp16")]; + tensor var_9676_to_fp16 = const()[name = tensor("op_9676_to_fp16"), val = tensor(0x1p+9)]; + tensor var_9677_cast_fp16 = mul(x = var_9675_cast_fp16, y = var_9676_to_fp16)[name = tensor("op_9677_cast_fp16")]; + tensor write_indices_float_91_cast_fp16 = sub(x = ts_137_cast_fp16, y = var_9677_cast_fp16)[name = tensor("write_indices_float_91_cast_fp16")]; + tensor var_9684_dtype_0 = const()[name = tensor("op_9684_dtype_0"), val = tensor("int32")]; + tensor write_indices_45_reps_0 = const()[name = tensor("write_indices_45_reps_0"), val = tensor([1, 1, 16, 64])]; + tensor write_indices_float_91_cast_fp16_to_int32 = cast(dtype = var_9684_dtype_0, x = write_indices_float_91_cast_fp16)[name = tensor("cast_442")]; + tensor write_indices_45 = tile(reps = write_indices_45_reps_0, x = write_indices_float_91_cast_fp16_to_int32)[name = tensor("write_indices_45")]; + tensor var_9692_begin_0 = const()[name = tensor("op_9692_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_9692_end_0 = const()[name = tensor("op_9692_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor var_9692_end_mask_0 = const()[name = tensor("op_9692_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_9692_squeeze_mask_0 = const()[name = tensor("op_9692_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor cache22_to_fp16_dtype_0 = const()[name = tensor("cache22_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor cache22_to_fp16 = cast(dtype = cache22_to_fp16_dtype_0, x = cache22)[name = tensor("cast_441")]; + tensor var_9692_cast_fp16 = slice_by_index(begin = var_9692_begin_0, end = var_9692_end_0, end_mask = var_9692_end_mask_0, squeeze_mask = var_9692_squeeze_mask_0, x = cache22_to_fp16)[name = tensor("op_9692_cast_fp16")]; + tensor var_9694_axis_0 = const()[name = tensor("op_9694_axis_0"), val = tensor(1)]; + tensor var_9694_mode_0 = const()[name = tensor("op_9694_mode_0"), val = tensor("update")]; + tensor var_9694_validate_indices_0 = const()[name = tensor("op_9694_validate_indices_0"), val = tensor(false)]; + tensor var_9694_cast_fp16 = scatter_along_axis(axis = var_9694_axis_0, data = var_9692_cast_fp16, indices = write_indices_45, mode = var_9694_mode_0, updates = k_91_cast_fp16, validate_indices = var_9694_validate_indices_0)[name = tensor("op_9694_cast_fp16")]; + tensor concat_156 = const()[name = tensor("concat_156"), val = tensor([0, 0, 0, 0, 0])]; + tensor concat_157 = const()[name = tensor("concat_157"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_45_internal_tensor_assign_1_stride_0 = const()[name = tensor("new_cache_45_internal_tensor_assign_1_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_45_internal_tensor_assign_1_begin_mask_0 = const()[name = tensor("new_cache_45_internal_tensor_assign_1_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_45_internal_tensor_assign_1_end_mask_0 = const()[name = tensor("new_cache_45_internal_tensor_assign_1_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_45_internal_tensor_assign_1_squeeze_mask_0 = const()[name = tensor("new_cache_45_internal_tensor_assign_1_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_92 = const()[name = tensor("shape_92"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_44 = const()[name = tensor("reduce_prod_44"), val = tensor(1048576)]; + tensor range_1d_44_start_0 = const()[name = tensor("range_1d_44_start_0"), val = tensor(0)]; + tensor range_1d_44_step_0 = const()[name = tensor("range_1d_44_step_0"), val = tensor(1)]; + tensor range_1d_44 = range_1d(end = reduce_prod_44, start = range_1d_44_start_0, step = range_1d_44_step_0)[name = tensor("range_1d_44")]; + tensor reshape_220 = reshape(shape = shape_92, x = range_1d_44)[name = tensor("reshape_220")]; + tensor slice_by_index_44 = slice_by_index(begin = concat_156, begin_mask = new_cache_45_internal_tensor_assign_1_begin_mask_0, end = concat_157, end_mask = new_cache_45_internal_tensor_assign_1_end_mask_0, squeeze_mask = new_cache_45_internal_tensor_assign_1_squeeze_mask_0, stride = new_cache_45_internal_tensor_assign_1_stride_0, x = reshape_220)[name = tensor("slice_by_index_44")]; + tensor reshape_221_shape_0 = const()[name = tensor("reshape_221_shape_0"), val = tensor([-1])]; + tensor reshape_221 = reshape(shape = reshape_221_shape_0, x = slice_by_index_44)[name = tensor("reshape_221")]; + tensor reshape_222_shape_0 = const()[name = tensor("reshape_222_shape_0"), val = tensor([-1])]; + tensor reshape_222_cast_fp16 = reshape(shape = reshape_222_shape_0, x = var_9694_cast_fp16)[name = tensor("reshape_222_cast_fp16")]; + tensor reshape_223_shape_0 = const()[name = tensor("reshape_223_shape_0"), val = tensor([-1])]; + tensor reshape_223_cast_fp16 = reshape(shape = reshape_223_shape_0, x = cache22_to_fp16)[name = tensor("reshape_223_cast_fp16")]; + tensor scatter_44_mode_0 = const()[name = tensor("scatter_44_mode_0"), val = tensor("update")]; + tensor scatter_44_axis_0 = const()[name = tensor("scatter_44_axis_0"), val = tensor(0)]; + tensor scatter_44_validate_indices_0 = const()[name = tensor("scatter_44_validate_indices_0"), val = tensor(false)]; + tensor scatter_44_cast_fp16 = scatter(axis = scatter_44_axis_0, data = reshape_223_cast_fp16, indices = reshape_221, mode = scatter_44_mode_0, updates = reshape_222_cast_fp16, validate_indices = scatter_44_validate_indices_0)[name = tensor("scatter_44_cast_fp16")]; + tensor reshape_224_cast_fp16 = reshape(shape = shape_92, x = scatter_44_cast_fp16)[name = tensor("reshape_224_cast_fp16")]; + tensor var_9702_begin_0 = const()[name = tensor("op_9702_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor var_9702_end_0 = const()[name = tensor("op_9702_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor var_9702_end_mask_0 = const()[name = tensor("op_9702_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_9702_squeeze_mask_0 = const()[name = tensor("op_9702_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor var_9702_cast_fp16 = slice_by_index(begin = var_9702_begin_0, end = var_9702_end_0, end_mask = var_9702_end_mask_0, squeeze_mask = var_9702_squeeze_mask_0, x = reshape_224_cast_fp16)[name = tensor("op_9702_cast_fp16")]; + tensor var_9704_axis_0 = const()[name = tensor("op_9704_axis_0"), val = tensor(1)]; + tensor var_9704_mode_0 = const()[name = tensor("op_9704_mode_0"), val = tensor("update")]; + tensor var_9704_validate_indices_0 = const()[name = tensor("op_9704_validate_indices_0"), val = tensor(false)]; + tensor var_9704_cast_fp16 = scatter_along_axis(axis = var_9704_axis_0, data = var_9702_cast_fp16, indices = write_indices_45, mode = var_9704_mode_0, updates = v_45_cast_fp16, validate_indices = var_9704_validate_indices_0)[name = tensor("op_9704_cast_fp16")]; + tensor concat_158 = const()[name = tensor("concat_158"), val = tensor([1, 0, 0, 0, 0])]; + tensor concat_159 = const()[name = tensor("concat_159"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_45_internal_tensor_assign_2_stride_0 = const()[name = tensor("new_cache_45_internal_tensor_assign_2_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_45_internal_tensor_assign_2_begin_mask_0 = const()[name = tensor("new_cache_45_internal_tensor_assign_2_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_45_internal_tensor_assign_2_end_mask_0 = const()[name = tensor("new_cache_45_internal_tensor_assign_2_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_45_internal_tensor_assign_2_squeeze_mask_0 = const()[name = tensor("new_cache_45_internal_tensor_assign_2_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_93 = const()[name = tensor("shape_93"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_45 = const()[name = tensor("reduce_prod_45"), val = tensor(1048576)]; + tensor range_1d_45_start_0 = const()[name = tensor("range_1d_45_start_0"), val = tensor(0)]; + tensor range_1d_45_step_0 = const()[name = tensor("range_1d_45_step_0"), val = tensor(1)]; + tensor range_1d_45 = range_1d(end = reduce_prod_45, start = range_1d_45_start_0, step = range_1d_45_step_0)[name = tensor("range_1d_45")]; + tensor reshape_225 = reshape(shape = shape_93, x = range_1d_45)[name = tensor("reshape_225")]; + tensor slice_by_index_45 = slice_by_index(begin = concat_158, begin_mask = new_cache_45_internal_tensor_assign_2_begin_mask_0, end = concat_159, end_mask = new_cache_45_internal_tensor_assign_2_end_mask_0, squeeze_mask = new_cache_45_internal_tensor_assign_2_squeeze_mask_0, stride = new_cache_45_internal_tensor_assign_2_stride_0, x = reshape_225)[name = tensor("slice_by_index_45")]; + tensor reshape_226_shape_0 = const()[name = tensor("reshape_226_shape_0"), val = tensor([-1])]; + tensor reshape_226 = reshape(shape = reshape_226_shape_0, x = slice_by_index_45)[name = tensor("reshape_226")]; + tensor reshape_227_shape_0 = const()[name = tensor("reshape_227_shape_0"), val = tensor([-1])]; + tensor reshape_227_cast_fp16 = reshape(shape = reshape_227_shape_0, x = var_9704_cast_fp16)[name = tensor("reshape_227_cast_fp16")]; + tensor reshape_228_shape_0 = const()[name = tensor("reshape_228_shape_0"), val = tensor([-1])]; + tensor reshape_228_cast_fp16 = reshape(shape = reshape_228_shape_0, x = reshape_224_cast_fp16)[name = tensor("reshape_228_cast_fp16")]; + tensor scatter_45_mode_0 = const()[name = tensor("scatter_45_mode_0"), val = tensor("update")]; + tensor scatter_45_axis_0 = const()[name = tensor("scatter_45_axis_0"), val = tensor(0)]; + tensor scatter_45_validate_indices_0 = const()[name = tensor("scatter_45_validate_indices_0"), val = tensor(false)]; + tensor scatter_45_cast_fp16 = scatter(axis = scatter_45_axis_0, data = reshape_228_cast_fp16, indices = reshape_226, mode = scatter_45_mode_0, updates = reshape_227_cast_fp16, validate_indices = scatter_45_validate_indices_0)[name = tensor("scatter_45_cast_fp16")]; + tensor reshape_229_cast_fp16 = reshape(shape = shape_93, x = scatter_45_cast_fp16)[name = tensor("reshape_229_cast_fp16")]; + tensor reshape_229_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("reshape_229_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor keys_133_begin_0 = const()[name = tensor("keys_133_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor keys_133_end_0 = const()[name = tensor("keys_133_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor keys_133_end_mask_0 = const()[name = tensor("keys_133_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor keys_133_squeeze_mask_0 = const()[name = tensor("keys_133_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor keys_133_cast_fp16 = slice_by_index(begin = keys_133_begin_0, end = keys_133_end_0, end_mask = keys_133_end_mask_0, squeeze_mask = keys_133_squeeze_mask_0, x = reshape_229_cast_fp16)[name = tensor("keys_133_cast_fp16")]; + tensor values_133_begin_0 = const()[name = tensor("values_133_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor values_133_end_0 = const()[name = tensor("values_133_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor values_133_end_mask_0 = const()[name = tensor("values_133_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor values_133_squeeze_mask_0 = const()[name = tensor("values_133_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor values_133_cast_fp16 = slice_by_index(begin = values_133_begin_0, end = values_133_end_0, end_mask = values_133_end_mask_0, squeeze_mask = values_133_squeeze_mask_0, x = reshape_229_cast_fp16)[name = tensor("values_133_cast_fp16")]; + tensor var_9716_cast_fp16 = not_equal(x = keys_133_cast_fp16, y = keys_133_cast_fp16)[name = tensor("op_9716_cast_fp16")]; + tensor keys_135_cast_fp16 = select(a = var_504_to_fp16, b = keys_133_cast_fp16, cond = var_9716_cast_fp16)[name = tensor("keys_135_cast_fp16")]; + tensor var_9724_cast_fp16 = not_equal(x = values_133_cast_fp16, y = values_133_cast_fp16)[name = tensor("op_9724_cast_fp16")]; + tensor values_135_cast_fp16 = select(a = var_504_to_fp16, b = values_133_cast_fp16, cond = var_9724_cast_fp16)[name = tensor("values_135_cast_fp16")]; + tensor var_9748 = const()[name = tensor("op_9748"), val = tensor([0, 2, 1, 3])]; + tensor var_9761 = const()[name = tensor("op_9761"), val = tensor([1, 1, 1])]; + tensor var_9762_cast_fp16 = reshape(shape = var_9761, x = position22_to_fp16)[name = tensor("op_9762_cast_fp16")]; + tensor var_9779_to_fp16 = const()[name = tensor("op_9779_to_fp16"), val = tensor(0x1p+0)]; + tensor valid_len_45_cast_fp16 = add(x = var_9762_cast_fp16, y = var_9779_to_fp16)[name = tensor("valid_len_45_cast_fp16")]; + tensor valid_mask_45_cast_fp16 = less(x = k_positions_1_promoted_to_fp16, y = valid_len_45_cast_fp16)[name = tensor("valid_mask_45_cast_fp16")]; + tensor causal_mask_45_cast_fp16 = less_equal(x = k_positions_1_promoted_to_fp16, y = var_9762_cast_fp16)[name = tensor("causal_mask_45_cast_fp16")]; + tensor attn_mask_89 = logical_and(x = valid_mask_45_cast_fp16, y = causal_mask_45_cast_fp16)[name = tensor("attn_mask_89")]; + tensor attn_mask_91_axes_0 = const()[name = tensor("attn_mask_91_axes_0"), val = tensor([1])]; + tensor attn_mask_91 = expand_dims(axes = attn_mask_91_axes_0, x = attn_mask_89)[name = tensor("attn_mask_91")]; + tensor var_9797_transpose_x_0 = const()[name = tensor("op_9797_transpose_x_0"), val = tensor(false)]; + tensor var_9797_transpose_y_0 = const()[name = tensor("op_9797_transpose_y_0"), val = tensor(false)]; + tensor transpose_116_perm_0 = const()[name = tensor("transpose_116_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_117_perm_0 = const()[name = tensor("transpose_117_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_117 = transpose(perm = transpose_117_perm_0, x = keys_135_cast_fp16)[name = tensor("transpose_125")]; + tensor transpose_116 = transpose(perm = transpose_116_perm_0, x = q_135_cast_fp16)[name = tensor("transpose_126")]; + tensor var_9797_cast_fp16 = matmul(transpose_x = var_9797_transpose_x_0, transpose_y = var_9797_transpose_y_0, x = transpose_116, y = transpose_117)[name = tensor("op_9797_cast_fp16")]; + tensor var_9791_to_fp16 = const()[name = tensor("op_9791_to_fp16"), val = tensor([0x1p-3])]; + tensor attn_weights_133_cast_fp16 = mul(x = var_9797_cast_fp16, y = var_9791_to_fp16)[name = tensor("attn_weights_133_cast_fp16")]; + tensor var_9799 = logical_not(x = attn_mask_91)[name = tensor("op_9799")]; + tensor var_9800_to_fp16 = const()[name = tensor("op_9800_to_fp16"), val = tensor(-inf)]; + tensor attn_weights_135_cast_fp16 = select(a = var_9800_to_fp16, b = attn_weights_133_cast_fp16, cond = var_9799)[name = tensor("attn_weights_135_cast_fp16")]; + tensor var_9802 = const()[name = tensor("op_9802"), val = tensor(-1)]; + tensor attn_weights_137_cast_fp16 = softmax(axis = var_9802, x = attn_weights_135_cast_fp16)[name = tensor("attn_weights_137_cast_fp16")]; + tensor attn_output_45_transpose_x_0 = const()[name = tensor("attn_output_45_transpose_x_0"), val = tensor(false)]; + tensor attn_output_45_transpose_y_0 = const()[name = tensor("attn_output_45_transpose_y_0"), val = tensor(false)]; + tensor values_137_cast_fp16 = transpose(perm = var_9748, x = values_135_cast_fp16)[name = tensor("transpose_127")]; + tensor attn_output_45_cast_fp16 = matmul(transpose_x = attn_output_45_transpose_x_0, transpose_y = attn_output_45_transpose_y_0, x = attn_weights_137_cast_fp16, y = values_137_cast_fp16)[name = tensor("attn_output_45_cast_fp16")]; + tensor var_9810 = const()[name = tensor("op_9810"), val = tensor([0, 2, 1, 3])]; + tensor var_9813 = const()[name = tensor("op_9813"), val = tensor([1, 1, 1024])]; + tensor var_9811_cast_fp16 = transpose(perm = var_9810, x = attn_output_45_cast_fp16)[name = tensor("transpose_124")]; + tensor input_225_cast_fp16 = reshape(shape = var_9813, x = var_9811_cast_fp16)[name = tensor("input_225_cast_fp16")]; + tensor linear_90_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = attn22_out_proj_weight_cast_fp16, x = input_225_cast_fp16)[name = tensor("linear_90_cast_fp16")]; + tensor var_9819_to_fp16 = const()[name = tensor("op_9819_to_fp16"), val = tensor(0x1p+0)]; + tensor var_9820_cast_fp16 = add(x = position22_to_fp16, y = var_9819_to_fp16)[name = tensor("op_9820_cast_fp16")]; + tensor var_9820_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_9820_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor input_227_cast_fp16 = add(x = input_223_cast_fp16, y = linear_90_cast_fp16)[name = tensor("input_227_cast_fp16")]; + tensor input_229_axes_0 = const()[name = tensor("input_229_axes_0"), val = tensor([-1])]; + tensor norm22_2_weight_to_fp16 = const()[name = tensor("norm22_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(303777920)))]; + tensor norm22_2_bias_to_fp16 = const()[name = tensor("norm22_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(303780032)))]; + tensor var_9824_to_fp16 = const()[name = tensor("op_9824_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_229_cast_fp16 = layer_norm(axes = input_229_axes_0, beta = norm22_2_bias_to_fp16, epsilon = var_9824_to_fp16, gamma = norm22_2_weight_to_fp16, x = input_227_cast_fp16)[name = tensor("input_229_cast_fp16")]; + tensor linear_91_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = linear22_1_weight_cast_fp16, x = input_229_cast_fp16)[name = tensor("linear_91_cast_fp16")]; + tensor input_231_mode_0 = const()[name = tensor("input_231_mode_0"), val = tensor("EXACT")]; + tensor input_231_cast_fp16 = gelu(mode = input_231_mode_0, x = linear_91_cast_fp16)[name = tensor("input_231_cast_fp16")]; + tensor linear_92_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = linear22_2_weight_cast_fp16, x = input_231_cast_fp16)[name = tensor("linear_92_cast_fp16")]; + tensor input_233_cast_fp16 = add(x = input_227_cast_fp16, y = linear_92_cast_fp16)[name = tensor("input_233_cast_fp16")]; + tensor x_axes_0 = const()[name = tensor("x_axes_0"), val = tensor([-1])]; + tensor norm23_1_weight_to_fp16 = const()[name = tensor("norm23_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(303782144)))]; + tensor norm23_1_bias_to_fp16 = const()[name = tensor("norm23_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(303784256)))]; + tensor var_9841_to_fp16 = const()[name = tensor("op_9841_to_fp16"), val = tensor(0x1.5p-17)]; + tensor x_cast_fp16 = layer_norm(axes = x_axes_0, beta = norm23_1_bias_to_fp16, epsilon = var_9841_to_fp16, gamma = norm23_1_weight_to_fp16, x = input_233_cast_fp16)[name = tensor("x_cast_fp16")]; + tensor linear_93_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = attn23_in_proj_weight_cast_fp16, x = x_cast_fp16)[name = tensor("linear_93_cast_fp16")]; + tensor var_9877 = const()[name = tensor("op_9877"), val = tensor([1, 1, 3, 16, 64])]; + tensor qkv_cast_fp16 = reshape(shape = var_9877, x = linear_93_cast_fp16)[name = tensor("qkv_cast_fp16")]; + tensor q_139_begin_0 = const()[name = tensor("q_139_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor q_139_end_0 = const()[name = tensor("q_139_end_0"), val = tensor([1, 1, 1, 16, 64])]; + tensor q_139_end_mask_0 = const()[name = tensor("q_139_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor q_139_squeeze_mask_0 = const()[name = tensor("q_139_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor q_139_cast_fp16 = slice_by_index(begin = q_139_begin_0, end = q_139_end_0, end_mask = q_139_end_mask_0, squeeze_mask = q_139_squeeze_mask_0, x = qkv_cast_fp16)[name = tensor("q_139_cast_fp16")]; + tensor k_93_begin_0 = const()[name = tensor("k_93_begin_0"), val = tensor([0, 0, 1, 0, 0])]; + tensor k_93_end_0 = const()[name = tensor("k_93_end_0"), val = tensor([1, 1, 2, 16, 64])]; + tensor k_93_end_mask_0 = const()[name = tensor("k_93_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor k_93_squeeze_mask_0 = const()[name = tensor("k_93_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor k_93_cast_fp16 = slice_by_index(begin = k_93_begin_0, end = k_93_end_0, end_mask = k_93_end_mask_0, squeeze_mask = k_93_squeeze_mask_0, x = qkv_cast_fp16)[name = tensor("k_93_cast_fp16")]; + tensor v_begin_0 = const()[name = tensor("v_begin_0"), val = tensor([0, 0, 2, 0, 0])]; + tensor v_end_0 = const()[name = tensor("v_end_0"), val = tensor([1, 1, 3, 16, 64])]; + tensor v_end_mask_0 = const()[name = tensor("v_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor v_squeeze_mask_0 = const()[name = tensor("v_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor v_cast_fp16 = slice_by_index(begin = v_begin_0, end = v_end_0, end_mask = v_end_mask_0, squeeze_mask = v_squeeze_mask_0, x = qkv_cast_fp16)[name = tensor("v_cast_fp16")]; + tensor var_9981 = const()[name = tensor("op_9981"), val = tensor([1, 1, 1, 1])]; + tensor position23_to_fp16_dtype_0 = const()[name = tensor("position23_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor position23_to_fp16 = cast(dtype = position23_to_fp16_dtype_0, x = position23)[name = tensor("cast_438")]; + tensor ts_cast_fp16 = reshape(shape = var_9981, x = position23_to_fp16)[name = tensor("ts_cast_fp16")]; + tensor var_9985 = const()[name = tensor("op_9985"), val = tensor([1, 1, 16, 32, 2])]; + tensor q_complex_cast_fp16 = reshape(shape = var_9985, x = q_139_cast_fp16)[name = tensor("q_complex_cast_fp16")]; + tensor var_9989 = const()[name = tensor("op_9989"), val = tensor([1, 1, 16, 32, 2])]; + tensor k_complex_cast_fp16 = reshape(shape = var_9989, x = k_93_cast_fp16)[name = tensor("k_complex_cast_fp16")]; + tensor var_9993_begin_0 = const()[name = tensor("op_9993_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_9993_end_0 = const()[name = tensor("op_9993_end_0"), val = tensor([1, 1, 16, 32, 1])]; + tensor var_9993_end_mask_0 = const()[name = tensor("op_9993_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_9993_squeeze_mask_0 = const()[name = tensor("op_9993_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_9993_cast_fp16 = slice_by_index(begin = var_9993_begin_0, end = var_9993_end_0, end_mask = var_9993_end_mask_0, squeeze_mask = var_9993_squeeze_mask_0, x = q_complex_cast_fp16)[name = tensor("op_9993_cast_fp16")]; + tensor var_10001_begin_0 = const()[name = tensor("op_10001_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_10001_end_0 = const()[name = tensor("op_10001_end_0"), val = tensor([1, 1, 16, 32, 2])]; + tensor var_10001_end_mask_0 = const()[name = tensor("op_10001_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_10001_squeeze_mask_0 = const()[name = tensor("op_10001_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_10001_cast_fp16 = slice_by_index(begin = var_10001_begin_0, end = var_10001_end_0, end_mask = var_10001_end_mask_0, squeeze_mask = var_10001_squeeze_mask_0, x = q_complex_cast_fp16)[name = tensor("op_10001_cast_fp16")]; + tensor var_10009_begin_0 = const()[name = tensor("op_10009_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_10009_end_0 = const()[name = tensor("op_10009_end_0"), val = tensor([1, 1, 16, 32, 1])]; + tensor var_10009_end_mask_0 = const()[name = tensor("op_10009_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_10009_squeeze_mask_0 = const()[name = tensor("op_10009_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_10009_cast_fp16 = slice_by_index(begin = var_10009_begin_0, end = var_10009_end_0, end_mask = var_10009_end_mask_0, squeeze_mask = var_10009_squeeze_mask_0, x = k_complex_cast_fp16)[name = tensor("op_10009_cast_fp16")]; + tensor var_10017_begin_0 = const()[name = tensor("op_10017_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_10017_end_0 = const()[name = tensor("op_10017_end_0"), val = tensor([1, 1, 16, 32, 2])]; + tensor var_10017_end_mask_0 = const()[name = tensor("op_10017_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_10017_squeeze_mask_0 = const()[name = tensor("op_10017_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_10017_cast_fp16 = slice_by_index(begin = var_10017_begin_0, end = var_10017_end_0, end_mask = var_10017_end_mask_0, squeeze_mask = var_10017_squeeze_mask_0, x = k_complex_cast_fp16)[name = tensor("op_10017_cast_fp16")]; + tensor freqs_to_fp16 = const()[name = tensor("freqs_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(303786368)))]; + tensor var_10023_cast_fp16 = mul(x = freqs_to_fp16, y = ts_cast_fp16)[name = tensor("op_10023_cast_fp16")]; + tensor rotr_cast_fp16 = cos(x = var_10023_cast_fp16)[name = tensor("rotr_cast_fp16")]; + tensor roti_cast_fp16 = sin(x = var_10023_cast_fp16)[name = tensor("roti_cast_fp16")]; + tensor var_10027_cast_fp16 = mul(x = var_9993_cast_fp16, y = rotr_cast_fp16)[name = tensor("op_10027_cast_fp16")]; + tensor var_10028_cast_fp16 = mul(x = var_10001_cast_fp16, y = roti_cast_fp16)[name = tensor("op_10028_cast_fp16")]; + tensor qor_93_cast_fp16 = sub(x = var_10027_cast_fp16, y = var_10028_cast_fp16)[name = tensor("qor_93_cast_fp16")]; + tensor var_10031_cast_fp16 = mul(x = var_9993_cast_fp16, y = roti_cast_fp16)[name = tensor("op_10031_cast_fp16")]; + tensor var_10032_cast_fp16 = mul(x = var_10001_cast_fp16, y = rotr_cast_fp16)[name = tensor("op_10032_cast_fp16")]; + tensor qoi_93_cast_fp16 = add(x = var_10031_cast_fp16, y = var_10032_cast_fp16)[name = tensor("qoi_93_cast_fp16")]; + tensor var_10035_cast_fp16 = mul(x = var_10009_cast_fp16, y = rotr_cast_fp16)[name = tensor("op_10035_cast_fp16")]; + tensor var_10036_cast_fp16 = mul(x = var_10017_cast_fp16, y = roti_cast_fp16)[name = tensor("op_10036_cast_fp16")]; + tensor kor_93_cast_fp16 = sub(x = var_10035_cast_fp16, y = var_10036_cast_fp16)[name = tensor("kor_93_cast_fp16")]; + tensor var_10039_cast_fp16 = mul(x = var_10009_cast_fp16, y = roti_cast_fp16)[name = tensor("op_10039_cast_fp16")]; + tensor var_10040_cast_fp16 = mul(x = var_10017_cast_fp16, y = rotr_cast_fp16)[name = tensor("op_10040_cast_fp16")]; + tensor koi_93_cast_fp16 = add(x = var_10039_cast_fp16, y = var_10040_cast_fp16)[name = tensor("koi_93_cast_fp16")]; + tensor qo_axis_0 = const()[name = tensor("qo_axis_0"), val = tensor(-1)]; + tensor qo_cast_fp16 = stack(axis = qo_axis_0, values = (qor_93_cast_fp16, qoi_93_cast_fp16))[name = tensor("qo_cast_fp16")]; + tensor ko_axis_0 = const()[name = tensor("ko_axis_0"), val = tensor(-1)]; + tensor ko_cast_fp16 = stack(axis = ko_axis_0, values = (kor_93_cast_fp16, koi_93_cast_fp16))[name = tensor("ko_cast_fp16")]; + tensor var_10069 = const()[name = tensor("op_10069"), val = tensor([1, 1, 16, 64])]; + tensor q_141_cast_fp16 = reshape(shape = var_10069, x = qo_cast_fp16)[name = tensor("q_141_cast_fp16")]; + tensor var_10071 = const()[name = tensor("op_10071"), val = tensor([1, 1, 16, 64])]; + tensor k_cast_fp16 = reshape(shape = var_10071, x = ko_cast_fp16)[name = tensor("k_cast_fp16")]; + tensor _inversed_10093_y_0_to_fp16 = const()[name = tensor("_inversed_10093_y_0_to_fp16"), val = tensor(0x1p-9)]; + tensor _inversed_10093_cast_fp16 = mul(x = ts_cast_fp16, y = _inversed_10093_y_0_to_fp16)[name = tensor("_inversed_10093_cast_fp16")]; + tensor var_10094_cast_fp16 = floor(x = _inversed_10093_cast_fp16)[name = tensor("op_10094_cast_fp16")]; + tensor var_10095_to_fp16 = const()[name = tensor("op_10095_to_fp16"), val = tensor(0x1p+9)]; + tensor var_10096_cast_fp16 = mul(x = var_10094_cast_fp16, y = var_10095_to_fp16)[name = tensor("op_10096_cast_fp16")]; + tensor write_indices_float_cast_fp16 = sub(x = ts_cast_fp16, y = var_10096_cast_fp16)[name = tensor("write_indices_float_cast_fp16")]; + tensor var_10103_dtype_0 = const()[name = tensor("op_10103_dtype_0"), val = tensor("int32")]; + tensor write_indices_reps_0 = const()[name = tensor("write_indices_reps_0"), val = tensor([1, 1, 16, 64])]; + tensor write_indices_float_cast_fp16_to_int32 = cast(dtype = var_10103_dtype_0, x = write_indices_float_cast_fp16)[name = tensor("cast_437")]; + tensor write_indices = tile(reps = write_indices_reps_0, x = write_indices_float_cast_fp16_to_int32)[name = tensor("write_indices")]; + tensor var_10111_begin_0 = const()[name = tensor("op_10111_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_10111_end_0 = const()[name = tensor("op_10111_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor var_10111_end_mask_0 = const()[name = tensor("op_10111_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_10111_squeeze_mask_0 = const()[name = tensor("op_10111_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor cache23_to_fp16_dtype_0 = const()[name = tensor("cache23_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor cache23_to_fp16 = cast(dtype = cache23_to_fp16_dtype_0, x = cache23)[name = tensor("cast_436")]; + tensor var_10111_cast_fp16 = slice_by_index(begin = var_10111_begin_0, end = var_10111_end_0, end_mask = var_10111_end_mask_0, squeeze_mask = var_10111_squeeze_mask_0, x = cache23_to_fp16)[name = tensor("op_10111_cast_fp16")]; + tensor var_10113_axis_0 = const()[name = tensor("op_10113_axis_0"), val = tensor(1)]; + tensor var_10113_mode_0 = const()[name = tensor("op_10113_mode_0"), val = tensor("update")]; + tensor var_10113_validate_indices_0 = const()[name = tensor("op_10113_validate_indices_0"), val = tensor(false)]; + tensor var_10113_cast_fp16 = scatter_along_axis(axis = var_10113_axis_0, data = var_10111_cast_fp16, indices = write_indices, mode = var_10113_mode_0, updates = k_cast_fp16, validate_indices = var_10113_validate_indices_0)[name = tensor("op_10113_cast_fp16")]; + tensor concat_163 = const()[name = tensor("concat_163"), val = tensor([0, 0, 0, 0, 0])]; + tensor concat_164 = const()[name = tensor("concat_164"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_internal_tensor_assign_1_stride_0 = const()[name = tensor("new_cache_internal_tensor_assign_1_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_internal_tensor_assign_1_begin_mask_0 = const()[name = tensor("new_cache_internal_tensor_assign_1_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_internal_tensor_assign_1_end_mask_0 = const()[name = tensor("new_cache_internal_tensor_assign_1_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_internal_tensor_assign_1_squeeze_mask_0 = const()[name = tensor("new_cache_internal_tensor_assign_1_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_94 = const()[name = tensor("shape_94"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_46 = const()[name = tensor("reduce_prod_46"), val = tensor(1048576)]; + tensor range_1d_46_start_0 = const()[name = tensor("range_1d_46_start_0"), val = tensor(0)]; + tensor range_1d_46_step_0 = const()[name = tensor("range_1d_46_step_0"), val = tensor(1)]; + tensor range_1d_46 = range_1d(end = reduce_prod_46, start = range_1d_46_start_0, step = range_1d_46_step_0)[name = tensor("range_1d_46")]; + tensor reshape_230 = reshape(shape = shape_94, x = range_1d_46)[name = tensor("reshape_230")]; + tensor slice_by_index_46 = slice_by_index(begin = concat_163, begin_mask = new_cache_internal_tensor_assign_1_begin_mask_0, end = concat_164, end_mask = new_cache_internal_tensor_assign_1_end_mask_0, squeeze_mask = new_cache_internal_tensor_assign_1_squeeze_mask_0, stride = new_cache_internal_tensor_assign_1_stride_0, x = reshape_230)[name = tensor("slice_by_index_46")]; + tensor reshape_231_shape_0 = const()[name = tensor("reshape_231_shape_0"), val = tensor([-1])]; + tensor reshape_231 = reshape(shape = reshape_231_shape_0, x = slice_by_index_46)[name = tensor("reshape_231")]; + tensor reshape_232_shape_0 = const()[name = tensor("reshape_232_shape_0"), val = tensor([-1])]; + tensor reshape_232_cast_fp16 = reshape(shape = reshape_232_shape_0, x = var_10113_cast_fp16)[name = tensor("reshape_232_cast_fp16")]; + tensor reshape_233_shape_0 = const()[name = tensor("reshape_233_shape_0"), val = tensor([-1])]; + tensor reshape_233_cast_fp16 = reshape(shape = reshape_233_shape_0, x = cache23_to_fp16)[name = tensor("reshape_233_cast_fp16")]; + tensor scatter_46_mode_0 = const()[name = tensor("scatter_46_mode_0"), val = tensor("update")]; + tensor scatter_46_axis_0 = const()[name = tensor("scatter_46_axis_0"), val = tensor(0)]; + tensor scatter_46_validate_indices_0 = const()[name = tensor("scatter_46_validate_indices_0"), val = tensor(false)]; + tensor scatter_46_cast_fp16 = scatter(axis = scatter_46_axis_0, data = reshape_233_cast_fp16, indices = reshape_231, mode = scatter_46_mode_0, updates = reshape_232_cast_fp16, validate_indices = scatter_46_validate_indices_0)[name = tensor("scatter_46_cast_fp16")]; + tensor reshape_234_cast_fp16 = reshape(shape = shape_94, x = scatter_46_cast_fp16)[name = tensor("reshape_234_cast_fp16")]; + tensor var_10121_begin_0 = const()[name = tensor("op_10121_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor var_10121_end_0 = const()[name = tensor("op_10121_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor var_10121_end_mask_0 = const()[name = tensor("op_10121_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_10121_squeeze_mask_0 = const()[name = tensor("op_10121_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor var_10121_cast_fp16 = slice_by_index(begin = var_10121_begin_0, end = var_10121_end_0, end_mask = var_10121_end_mask_0, squeeze_mask = var_10121_squeeze_mask_0, x = reshape_234_cast_fp16)[name = tensor("op_10121_cast_fp16")]; + tensor var_10123_axis_0 = const()[name = tensor("op_10123_axis_0"), val = tensor(1)]; + tensor var_10123_mode_0 = const()[name = tensor("op_10123_mode_0"), val = tensor("update")]; + tensor var_10123_validate_indices_0 = const()[name = tensor("op_10123_validate_indices_0"), val = tensor(false)]; + tensor var_10123_cast_fp16 = scatter_along_axis(axis = var_10123_axis_0, data = var_10121_cast_fp16, indices = write_indices, mode = var_10123_mode_0, updates = v_cast_fp16, validate_indices = var_10123_validate_indices_0)[name = tensor("op_10123_cast_fp16")]; + tensor concat_165 = const()[name = tensor("concat_165"), val = tensor([1, 0, 0, 0, 0])]; + tensor concat_166 = const()[name = tensor("concat_166"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_internal_tensor_assign_2_stride_0 = const()[name = tensor("new_cache_internal_tensor_assign_2_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_internal_tensor_assign_2_begin_mask_0 = const()[name = tensor("new_cache_internal_tensor_assign_2_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_internal_tensor_assign_2_end_mask_0 = const()[name = tensor("new_cache_internal_tensor_assign_2_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_internal_tensor_assign_2_squeeze_mask_0 = const()[name = tensor("new_cache_internal_tensor_assign_2_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_95 = const()[name = tensor("shape_95"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_47 = const()[name = tensor("reduce_prod_47"), val = tensor(1048576)]; + tensor range_1d_47_start_0 = const()[name = tensor("range_1d_47_start_0"), val = tensor(0)]; + tensor range_1d_47_step_0 = const()[name = tensor("range_1d_47_step_0"), val = tensor(1)]; + tensor range_1d_47 = range_1d(end = reduce_prod_47, start = range_1d_47_start_0, step = range_1d_47_step_0)[name = tensor("range_1d_47")]; + tensor reshape_235 = reshape(shape = shape_95, x = range_1d_47)[name = tensor("reshape_235")]; + tensor slice_by_index_47 = slice_by_index(begin = concat_165, begin_mask = new_cache_internal_tensor_assign_2_begin_mask_0, end = concat_166, end_mask = new_cache_internal_tensor_assign_2_end_mask_0, squeeze_mask = new_cache_internal_tensor_assign_2_squeeze_mask_0, stride = new_cache_internal_tensor_assign_2_stride_0, x = reshape_235)[name = tensor("slice_by_index_47")]; + tensor reshape_236_shape_0 = const()[name = tensor("reshape_236_shape_0"), val = tensor([-1])]; + tensor reshape_236 = reshape(shape = reshape_236_shape_0, x = slice_by_index_47)[name = tensor("reshape_236")]; + tensor reshape_237_shape_0 = const()[name = tensor("reshape_237_shape_0"), val = tensor([-1])]; + tensor reshape_237_cast_fp16 = reshape(shape = reshape_237_shape_0, x = var_10123_cast_fp16)[name = tensor("reshape_237_cast_fp16")]; + tensor reshape_238_shape_0 = const()[name = tensor("reshape_238_shape_0"), val = tensor([-1])]; + tensor reshape_238_cast_fp16 = reshape(shape = reshape_238_shape_0, x = reshape_234_cast_fp16)[name = tensor("reshape_238_cast_fp16")]; + tensor scatter_47_mode_0 = const()[name = tensor("scatter_47_mode_0"), val = tensor("update")]; + tensor scatter_47_axis_0 = const()[name = tensor("scatter_47_axis_0"), val = tensor(0)]; + tensor scatter_47_validate_indices_0 = const()[name = tensor("scatter_47_validate_indices_0"), val = tensor(false)]; + tensor scatter_47_cast_fp16 = scatter(axis = scatter_47_axis_0, data = reshape_238_cast_fp16, indices = reshape_236, mode = scatter_47_mode_0, updates = reshape_237_cast_fp16, validate_indices = scatter_47_validate_indices_0)[name = tensor("scatter_47_cast_fp16")]; + tensor reshape_239_cast_fp16 = reshape(shape = shape_95, x = scatter_47_cast_fp16)[name = tensor("reshape_239_cast_fp16")]; + tensor reshape_239_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("reshape_239_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor keys_139_begin_0 = const()[name = tensor("keys_139_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor keys_139_end_0 = const()[name = tensor("keys_139_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor keys_139_end_mask_0 = const()[name = tensor("keys_139_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor keys_139_squeeze_mask_0 = const()[name = tensor("keys_139_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor keys_139_cast_fp16 = slice_by_index(begin = keys_139_begin_0, end = keys_139_end_0, end_mask = keys_139_end_mask_0, squeeze_mask = keys_139_squeeze_mask_0, x = reshape_239_cast_fp16)[name = tensor("keys_139_cast_fp16")]; + tensor values_139_begin_0 = const()[name = tensor("values_139_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor values_139_end_0 = const()[name = tensor("values_139_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor values_139_end_mask_0 = const()[name = tensor("values_139_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor values_139_squeeze_mask_0 = const()[name = tensor("values_139_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor values_139_cast_fp16 = slice_by_index(begin = values_139_begin_0, end = values_139_end_0, end_mask = values_139_end_mask_0, squeeze_mask = values_139_squeeze_mask_0, x = reshape_239_cast_fp16)[name = tensor("values_139_cast_fp16")]; + tensor var_10135_cast_fp16 = not_equal(x = keys_139_cast_fp16, y = keys_139_cast_fp16)[name = tensor("op_10135_cast_fp16")]; + tensor keys_141_cast_fp16 = select(a = var_504_to_fp16, b = keys_139_cast_fp16, cond = var_10135_cast_fp16)[name = tensor("keys_141_cast_fp16")]; + tensor var_10143_cast_fp16 = not_equal(x = values_139_cast_fp16, y = values_139_cast_fp16)[name = tensor("op_10143_cast_fp16")]; + tensor values_141_cast_fp16 = select(a = var_504_to_fp16, b = values_139_cast_fp16, cond = var_10143_cast_fp16)[name = tensor("values_141_cast_fp16")]; + tensor var_10167 = const()[name = tensor("op_10167"), val = tensor([0, 2, 1, 3])]; + tensor var_10180 = const()[name = tensor("op_10180"), val = tensor([1, 1, 1])]; + tensor var_10181_cast_fp16 = reshape(shape = var_10180, x = position23_to_fp16)[name = tensor("op_10181_cast_fp16")]; + tensor var_10198_to_fp16 = const()[name = tensor("op_10198_to_fp16"), val = tensor(0x1p+0)]; + tensor valid_len_cast_fp16 = add(x = var_10181_cast_fp16, y = var_10198_to_fp16)[name = tensor("valid_len_cast_fp16")]; + tensor valid_mask_cast_fp16 = less(x = k_positions_1_promoted_to_fp16, y = valid_len_cast_fp16)[name = tensor("valid_mask_cast_fp16")]; + tensor causal_mask_cast_fp16 = less_equal(x = k_positions_1_promoted_to_fp16, y = var_10181_cast_fp16)[name = tensor("causal_mask_cast_fp16")]; + tensor attn_mask_93 = logical_and(x = valid_mask_cast_fp16, y = causal_mask_cast_fp16)[name = tensor("attn_mask_93")]; + tensor attn_mask_axes_0 = const()[name = tensor("attn_mask_axes_0"), val = tensor([1])]; + tensor attn_mask = expand_dims(axes = attn_mask_axes_0, x = attn_mask_93)[name = tensor("attn_mask")]; + tensor var_10216_transpose_x_0 = const()[name = tensor("op_10216_transpose_x_0"), val = tensor(false)]; + tensor var_10216_transpose_y_0 = const()[name = tensor("op_10216_transpose_y_0"), val = tensor(false)]; + tensor transpose_118_perm_0 = const()[name = tensor("transpose_118_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_119_perm_0 = const()[name = tensor("transpose_119_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_119 = transpose(perm = transpose_119_perm_0, x = keys_141_cast_fp16)[name = tensor("transpose_121")]; + tensor transpose_118 = transpose(perm = transpose_118_perm_0, x = q_141_cast_fp16)[name = tensor("transpose_122")]; + tensor var_10216_cast_fp16 = matmul(transpose_x = var_10216_transpose_x_0, transpose_y = var_10216_transpose_y_0, x = transpose_118, y = transpose_119)[name = tensor("op_10216_cast_fp16")]; + tensor var_10210_to_fp16 = const()[name = tensor("op_10210_to_fp16"), val = tensor([0x1p-3])]; + tensor attn_weights_139_cast_fp16 = mul(x = var_10216_cast_fp16, y = var_10210_to_fp16)[name = tensor("attn_weights_139_cast_fp16")]; + tensor var_10218 = logical_not(x = attn_mask)[name = tensor("op_10218")]; + tensor var_10219_to_fp16 = const()[name = tensor("op_10219_to_fp16"), val = tensor(-inf)]; + tensor attn_weights_141_cast_fp16 = select(a = var_10219_to_fp16, b = attn_weights_139_cast_fp16, cond = var_10218)[name = tensor("attn_weights_141_cast_fp16")]; + tensor var_10221 = const()[name = tensor("op_10221"), val = tensor(-1)]; + tensor attn_weights_cast_fp16 = softmax(axis = var_10221, x = attn_weights_141_cast_fp16)[name = tensor("attn_weights_cast_fp16")]; + tensor attn_output_transpose_x_0 = const()[name = tensor("attn_output_transpose_x_0"), val = tensor(false)]; + tensor attn_output_transpose_y_0 = const()[name = tensor("attn_output_transpose_y_0"), val = tensor(false)]; + tensor values_cast_fp16 = transpose(perm = var_10167, x = values_141_cast_fp16)[name = tensor("transpose_123")]; + tensor attn_output_cast_fp16 = matmul(transpose_x = attn_output_transpose_x_0, transpose_y = attn_output_transpose_y_0, x = attn_weights_cast_fp16, y = values_cast_fp16)[name = tensor("attn_output_cast_fp16")]; + tensor var_10229 = const()[name = tensor("op_10229"), val = tensor([0, 2, 1, 3])]; + tensor var_10232 = const()[name = tensor("op_10232"), val = tensor([1, 1, 1024])]; + tensor var_10230_cast_fp16 = transpose(perm = var_10229, x = attn_output_cast_fp16)[name = tensor("transpose_120")]; + tensor input_235_cast_fp16 = reshape(shape = var_10232, x = var_10230_cast_fp16)[name = tensor("input_235_cast_fp16")]; + tensor linear_94_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = attn23_out_proj_weight_cast_fp16, x = input_235_cast_fp16)[name = tensor("linear_94_cast_fp16")]; + tensor var_10238_to_fp16 = const()[name = tensor("op_10238_to_fp16"), val = tensor(0x1p+0)]; + tensor var_10239_cast_fp16 = add(x = position23_to_fp16, y = var_10238_to_fp16)[name = tensor("op_10239_cast_fp16")]; + tensor var_10239_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_10239_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor input_237_cast_fp16 = add(x = input_233_cast_fp16, y = linear_94_cast_fp16)[name = tensor("input_237_cast_fp16")]; + tensor input_239_axes_0 = const()[name = tensor("input_239_axes_0"), val = tensor([-1])]; + tensor norm23_2_weight_to_fp16 = const()[name = tensor("norm23_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(303786496)))]; + tensor norm23_2_bias_to_fp16 = const()[name = tensor("norm23_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(303788608)))]; + tensor var_10243_to_fp16 = const()[name = tensor("op_10243_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_239_cast_fp16 = layer_norm(axes = input_239_axes_0, beta = norm23_2_bias_to_fp16, epsilon = var_10243_to_fp16, gamma = norm23_2_weight_to_fp16, x = input_237_cast_fp16)[name = tensor("input_239_cast_fp16")]; + tensor linear_95_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = linear23_1_weight_cast_fp16, x = input_239_cast_fp16)[name = tensor("linear_95_cast_fp16")]; + tensor input_241_mode_0 = const()[name = tensor("input_241_mode_0"), val = tensor("EXACT")]; + tensor input_241_cast_fp16 = gelu(mode = input_241_mode_0, x = linear_95_cast_fp16)[name = tensor("input_241_cast_fp16")]; + tensor linear_96_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = linear23_2_weight_cast_fp16, x = input_241_cast_fp16)[name = tensor("linear_96_cast_fp16")]; + tensor input_243_cast_fp16 = add(x = input_237_cast_fp16, y = linear_96_cast_fp16)[name = tensor("input_243_cast_fp16")]; + tensor input_axes_0 = const()[name = tensor("input_axes_0"), val = tensor([-1])]; + tensor out_norm_weight_to_fp16 = const()[name = tensor("out_norm_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(303790720)))]; + tensor out_norm_bias_to_fp16 = const()[name = tensor("out_norm_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(303792832)))]; + tensor var_10260_to_fp16 = const()[name = tensor("op_10260_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_cast_fp16 = layer_norm(axes = input_axes_0, beta = out_norm_bias_to_fp16, epsilon = var_10260_to_fp16, gamma = out_norm_weight_to_fp16, x = input_243_cast_fp16)[name = tensor("input_cast_fp16")]; + tensor input_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("input_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor out_eos_weight_to_fp16 = const()[name = tensor("out_eos_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(303794944)))]; + tensor out_eos_bias_to_fp16 = const()[name = tensor("out_eos_bias_to_fp16"), val = tensor([-0x1.8ap-3])]; + tensor linear_97_cast_fp16 = linear(bias = out_eos_bias_to_fp16, weight = out_eos_weight_to_fp16, x = input_cast_fp16)[name = tensor("linear_97_cast_fp16")]; + tensor linear_97_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("linear_97_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor var_10268 = cast(dtype = linear_97_cast_fp16_to_fp32_dtype_0, x = linear_97_cast_fp16)[name = tensor("cast_432")]; + tensor input = cast(dtype = input_cast_fp16_to_fp32_dtype_0, x = input_cast_fp16)[name = tensor("cast_433")]; + tensor var_10239 = cast(dtype = var_10239_cast_fp16_to_fp32_dtype_0, x = var_10239_cast_fp16)[name = tensor("cast_434")]; + tensor new_cache_internal_tensor_assign_2 = cast(dtype = reshape_239_cast_fp16_to_fp32_dtype_0, x = reshape_239_cast_fp16)[name = tensor("cast_435")]; + tensor var_9820 = cast(dtype = var_9820_cast_fp16_to_fp32_dtype_0, x = var_9820_cast_fp16)[name = tensor("cast_439")]; + tensor new_cache_45_internal_tensor_assign_2 = cast(dtype = reshape_229_cast_fp16_to_fp32_dtype_0, x = reshape_229_cast_fp16)[name = tensor("cast_440")]; + tensor var_9401 = cast(dtype = var_9401_cast_fp16_to_fp32_dtype_0, x = var_9401_cast_fp16)[name = tensor("cast_444")]; + tensor new_cache_43_internal_tensor_assign_2 = cast(dtype = reshape_219_cast_fp16_to_fp32_dtype_0, x = reshape_219_cast_fp16)[name = tensor("cast_445")]; + tensor var_8982 = cast(dtype = var_8982_cast_fp16_to_fp32_dtype_0, x = var_8982_cast_fp16)[name = tensor("cast_449")]; + tensor new_cache_41_internal_tensor_assign_2 = cast(dtype = reshape_209_cast_fp16_to_fp32_dtype_0, x = reshape_209_cast_fp16)[name = tensor("cast_450")]; + tensor var_8563 = cast(dtype = var_8563_cast_fp16_to_fp32_dtype_0, x = var_8563_cast_fp16)[name = tensor("cast_454")]; + tensor new_cache_39_internal_tensor_assign_2 = cast(dtype = reshape_199_cast_fp16_to_fp32_dtype_0, x = reshape_199_cast_fp16)[name = tensor("cast_455")]; + tensor var_8144 = cast(dtype = var_8144_cast_fp16_to_fp32_dtype_0, x = var_8144_cast_fp16)[name = tensor("cast_459")]; + tensor new_cache_37_internal_tensor_assign_2 = cast(dtype = reshape_189_cast_fp16_to_fp32_dtype_0, x = reshape_189_cast_fp16)[name = tensor("cast_460")]; + tensor var_7725 = cast(dtype = var_7725_cast_fp16_to_fp32_dtype_0, x = var_7725_cast_fp16)[name = tensor("cast_464")]; + tensor new_cache_35_internal_tensor_assign_2 = cast(dtype = reshape_179_cast_fp16_to_fp32_dtype_0, x = reshape_179_cast_fp16)[name = tensor("cast_465")]; + tensor var_7306 = cast(dtype = var_7306_cast_fp16_to_fp32_dtype_0, x = var_7306_cast_fp16)[name = tensor("cast_469")]; + tensor new_cache_33_internal_tensor_assign_2 = cast(dtype = reshape_169_cast_fp16_to_fp32_dtype_0, x = reshape_169_cast_fp16)[name = tensor("cast_470")]; + tensor var_6887 = cast(dtype = var_6887_cast_fp16_to_fp32_dtype_0, x = var_6887_cast_fp16)[name = tensor("cast_474")]; + tensor new_cache_31_internal_tensor_assign_2 = cast(dtype = reshape_159_cast_fp16_to_fp32_dtype_0, x = reshape_159_cast_fp16)[name = tensor("cast_475")]; + tensor var_6468 = cast(dtype = var_6468_cast_fp16_to_fp32_dtype_0, x = var_6468_cast_fp16)[name = tensor("cast_479")]; + tensor new_cache_29_internal_tensor_assign_2 = cast(dtype = reshape_149_cast_fp16_to_fp32_dtype_0, x = reshape_149_cast_fp16)[name = tensor("cast_480")]; + tensor var_6049 = cast(dtype = var_6049_cast_fp16_to_fp32_dtype_0, x = var_6049_cast_fp16)[name = tensor("cast_484")]; + tensor new_cache_27_internal_tensor_assign_2 = cast(dtype = reshape_139_cast_fp16_to_fp32_dtype_0, x = reshape_139_cast_fp16)[name = tensor("cast_485")]; + tensor var_5630 = cast(dtype = var_5630_cast_fp16_to_fp32_dtype_0, x = var_5630_cast_fp16)[name = tensor("cast_489")]; + tensor new_cache_25_internal_tensor_assign_2 = cast(dtype = reshape_129_cast_fp16_to_fp32_dtype_0, x = reshape_129_cast_fp16)[name = tensor("cast_490")]; + tensor var_5211 = cast(dtype = var_5211_cast_fp16_to_fp32_dtype_0, x = var_5211_cast_fp16)[name = tensor("cast_494")]; + tensor new_cache_23_internal_tensor_assign_2 = cast(dtype = reshape_119_cast_fp16_to_fp32_dtype_0, x = reshape_119_cast_fp16)[name = tensor("cast_495")]; + tensor var_4792 = cast(dtype = var_4792_cast_fp16_to_fp32_dtype_0, x = var_4792_cast_fp16)[name = tensor("cast_499")]; + tensor new_cache_21_internal_tensor_assign_2 = cast(dtype = reshape_109_cast_fp16_to_fp32_dtype_0, x = reshape_109_cast_fp16)[name = tensor("cast_500")]; + tensor var_4373 = cast(dtype = var_4373_cast_fp16_to_fp32_dtype_0, x = var_4373_cast_fp16)[name = tensor("cast_504")]; + tensor new_cache_19_internal_tensor_assign_2 = cast(dtype = reshape_99_cast_fp16_to_fp32_dtype_0, x = reshape_99_cast_fp16)[name = tensor("cast_505")]; + tensor var_3954 = cast(dtype = var_3954_cast_fp16_to_fp32_dtype_0, x = var_3954_cast_fp16)[name = tensor("cast_509")]; + tensor new_cache_17_internal_tensor_assign_2 = cast(dtype = reshape_89_cast_fp16_to_fp32_dtype_0, x = reshape_89_cast_fp16)[name = tensor("cast_510")]; + tensor var_3535 = cast(dtype = var_3535_cast_fp16_to_fp32_dtype_0, x = var_3535_cast_fp16)[name = tensor("cast_514")]; + tensor new_cache_15_internal_tensor_assign_2 = cast(dtype = reshape_79_cast_fp16_to_fp32_dtype_0, x = reshape_79_cast_fp16)[name = tensor("cast_515")]; + tensor var_3116 = cast(dtype = var_3116_cast_fp16_to_fp32_dtype_0, x = var_3116_cast_fp16)[name = tensor("cast_519")]; + tensor new_cache_13_internal_tensor_assign_2 = cast(dtype = reshape_69_cast_fp16_to_fp32_dtype_0, x = reshape_69_cast_fp16)[name = tensor("cast_520")]; + tensor var_2697 = cast(dtype = var_2697_cast_fp16_to_fp32_dtype_0, x = var_2697_cast_fp16)[name = tensor("cast_524")]; + tensor new_cache_11_internal_tensor_assign_2 = cast(dtype = reshape_59_cast_fp16_to_fp32_dtype_0, x = reshape_59_cast_fp16)[name = tensor("cast_525")]; + tensor var_2278 = cast(dtype = var_2278_cast_fp16_to_fp32_dtype_0, x = var_2278_cast_fp16)[name = tensor("cast_529")]; + tensor new_cache_9_internal_tensor_assign_2 = cast(dtype = reshape_49_cast_fp16_to_fp32_dtype_0, x = reshape_49_cast_fp16)[name = tensor("cast_530")]; + tensor var_1859 = cast(dtype = var_1859_cast_fp16_to_fp32_dtype_0, x = var_1859_cast_fp16)[name = tensor("cast_534")]; + tensor new_cache_7_internal_tensor_assign_2 = cast(dtype = reshape_39_cast_fp16_to_fp32_dtype_0, x = reshape_39_cast_fp16)[name = tensor("cast_535")]; + tensor var_1440 = cast(dtype = var_1440_cast_fp16_to_fp32_dtype_0, x = var_1440_cast_fp16)[name = tensor("cast_539")]; + tensor new_cache_5_internal_tensor_assign_2 = cast(dtype = reshape_29_cast_fp16_to_fp32_dtype_0, x = reshape_29_cast_fp16)[name = tensor("cast_540")]; + tensor var_1021 = cast(dtype = var_1021_cast_fp16_to_fp32_dtype_0, x = var_1021_cast_fp16)[name = tensor("cast_544")]; + tensor new_cache_3_internal_tensor_assign_2 = cast(dtype = reshape_19_cast_fp16_to_fp32_dtype_0, x = reshape_19_cast_fp16)[name = tensor("cast_545")]; + tensor var_602 = cast(dtype = var_602_cast_fp16_to_fp32_dtype_0, x = var_602_cast_fp16)[name = tensor("cast_549")]; + tensor new_cache_1_internal_tensor_assign_2 = cast(dtype = reshape_9_cast_fp16_to_fp32_dtype_0, x = reshape_9_cast_fp16)[name = tensor("cast_550")]; + } -> (input, var_10268, new_cache_1_internal_tensor_assign_2, var_602, new_cache_3_internal_tensor_assign_2, var_1021, new_cache_5_internal_tensor_assign_2, var_1440, new_cache_7_internal_tensor_assign_2, var_1859, new_cache_9_internal_tensor_assign_2, var_2278, new_cache_11_internal_tensor_assign_2, var_2697, new_cache_13_internal_tensor_assign_2, var_3116, new_cache_15_internal_tensor_assign_2, var_3535, new_cache_17_internal_tensor_assign_2, var_3954, new_cache_19_internal_tensor_assign_2, var_4373, new_cache_21_internal_tensor_assign_2, var_4792, new_cache_23_internal_tensor_assign_2, var_5211, new_cache_25_internal_tensor_assign_2, var_5630, new_cache_27_internal_tensor_assign_2, var_6049, new_cache_29_internal_tensor_assign_2, var_6468, new_cache_31_internal_tensor_assign_2, var_6887, new_cache_33_internal_tensor_assign_2, var_7306, new_cache_35_internal_tensor_assign_2, var_7725, new_cache_37_internal_tensor_assign_2, var_8144, new_cache_39_internal_tensor_assign_2, var_8563, new_cache_41_internal_tensor_assign_2, var_8982, new_cache_43_internal_tensor_assign_2, var_9401, new_cache_45_internal_tensor_assign_2, var_9820, new_cache_internal_tensor_assign_2, var_10239); +} \ No newline at end of file diff --git a/v2.1/french_24l/flowlm_stepv2.mlmodelc/weights/weight.bin b/v2.1/french_24l/flowlm_stepv2.mlmodelc/weights/weight.bin new file mode 100644 index 0000000000000000000000000000000000000000..6ea4cd4188379e886f01f9d419c796230e490f59 --- /dev/null +++ b/v2.1/french_24l/flowlm_stepv2.mlmodelc/weights/weight.bin @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:a9bca994f15518be8b91c0c0cae1fa01dfcd9aa1ed03467210be5941ef040576 +size 303797056 diff --git a/v2.1/french_24l/flowlm_stepv2.mlpackage/Data/com.apple.CoreML/model.mlmodel b/v2.1/french_24l/flowlm_stepv2.mlpackage/Data/com.apple.CoreML/model.mlmodel new file mode 100644 index 0000000000000000000000000000000000000000..13cf63ec193ec020d6c4cd23e3698c3287f50951 --- /dev/null +++ b/v2.1/french_24l/flowlm_stepv2.mlpackage/Data/com.apple.CoreML/model.mlmodel @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:cdedef0c579c157948f4860fd7db8fd7aed703e6e16b882c82e3e00cbccbc90a +size 843748 diff --git a/v2.1/french_24l/flowlm_stepv2.mlpackage/Data/com.apple.CoreML/weights/weight.bin b/v2.1/french_24l/flowlm_stepv2.mlpackage/Data/com.apple.CoreML/weights/weight.bin new file mode 100644 index 0000000000000000000000000000000000000000..6ea4cd4188379e886f01f9d419c796230e490f59 --- /dev/null +++ b/v2.1/french_24l/flowlm_stepv2.mlpackage/Data/com.apple.CoreML/weights/weight.bin @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:a9bca994f15518be8b91c0c0cae1fa01dfcd9aa1ed03467210be5941ef040576 +size 303797056 diff --git a/v2.1/french_24l/flowlm_stepv2.mlpackage/Manifest.json b/v2.1/french_24l/flowlm_stepv2.mlpackage/Manifest.json new file mode 100644 index 0000000000000000000000000000000000000000..c7b669028ebd77b53732512d446ec8596b8017bd --- /dev/null +++ b/v2.1/french_24l/flowlm_stepv2.mlpackage/Manifest.json @@ -0,0 +1,18 @@ +{ + "fileFormatVersion": "1.0.0", + "itemInfoEntries": { + "944AE485-E21D-4BE3-B83E-11AED94E28B6": { + "author": "com.apple.CoreML", + "description": "CoreML Model Specification", + "name": "model.mlmodel", + "path": "com.apple.CoreML/model.mlmodel" + }, + "F6581A40-6B82-4A38-87F7-8E00A377C040": { + "author": "com.apple.CoreML", + "description": "CoreML Model Weights", + "name": "weights", + "path": "com.apple.CoreML/weights" + } + }, + "rootModelIdentifier": "944AE485-E21D-4BE3-B83E-11AED94E28B6" +} diff --git a/v2.1/french_24l/manifest.json b/v2.1/french_24l/manifest.json new file mode 100644 index 0000000000000000000000000000000000000000..fc7da98920126d1261a9c47517b1303db1599db7 --- /dev/null +++ b/v2.1/french_24l/manifest.json @@ -0,0 +1,27 @@ +{ + "version": "2.1", + "base": "v2", + "kind": "optimization-reconvert", + "self_contained": true, + "note": "Same weights as v2. Re-converted for speed: fused flow decoder (100% ANE), one-shot cond prefill, fp16 flowlm. NOT a finetune.", + "new": [ + "flow_decoder_fused", + "cond_prefill", + "flowlm_step" + ], + "copied_from_v2": [ + "flowlm_stepv2.mlpackage", + "flowlm_stepv2.mlmodelc", + "mimi_decoder.mlpackage", + "mimi_decoder.mlmodelc", + "constants", + "constants_bin" + ], + "compute_units": { + "flow_decoder_fused": "all (ANE)", + "cond_prefill": "all (GPU)", + "flowlm_step": "all (GPU, fp16)", + "flowlm_stepv2": "cpuAndGpu (GPU, int8 \u2014 fastest flowlm)", + "mimi_decoder": "cpuOnly" + } +} diff --git a/v2.1/french_24l/mimi_decoder.mlmodelc/analytics/coremldata.bin b/v2.1/french_24l/mimi_decoder.mlmodelc/analytics/coremldata.bin new file mode 100644 index 0000000000000000000000000000000000000000..07a58b2acc832853ff18b57a6d67023283dba57d --- /dev/null +++ b/v2.1/french_24l/mimi_decoder.mlmodelc/analytics/coremldata.bin @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:53b41de86399328f57b5db0ac2cbfda4f9016f07b484adf9ec050f0a5ff07b23 +size 243 diff --git a/v2.1/french_24l/mimi_decoder.mlmodelc/coremldata.bin b/v2.1/french_24l/mimi_decoder.mlmodelc/coremldata.bin new file mode 100644 index 0000000000000000000000000000000000000000..7c5ff02e7badf2b13da255cad0b766f1765665c5 --- /dev/null +++ b/v2.1/french_24l/mimi_decoder.mlmodelc/coremldata.bin @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:de15f7cc799bcdefd8e4474a0befe1161788a3e9aa68b4b448af10939cf35e46 +size 1777 diff --git a/v2.1/french_24l/mimi_decoder.mlmodelc/metadata.json b/v2.1/french_24l/mimi_decoder.mlmodelc/metadata.json new file mode 100644 index 0000000000000000000000000000000000000000..f66f9e7ff5be2ffbc44c8300f3501f67a77afaf6 --- /dev/null +++ b/v2.1/french_24l/mimi_decoder.mlmodelc/metadata.json @@ -0,0 +1,573 @@ +[ + { + "metadataOutputVersion" : "3.0", + "storagePrecision" : "Float32", + "outputSchema" : [ + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1 × 1 × 1920)", + "shortDescription" : "", + "shape" : "[1, 1, 1920]", + "name" : "var_724", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1 × 512 × 16)", + "shortDescription" : "", + "shape" : "[1, 512, 16]", + "name" : "var_77", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 2 × 1 × 256 × 8 × 64)", + "shortDescription" : "", + "shape" : "[2, 1, 256, 8, 64]", + "name" : "var_210", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1)", + "shortDescription" : "", + "shape" : "[1]", + "name" : "var_741", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 2 × 1 × 256 × 8 × 64)", + "shortDescription" : "", + "shape" : "[2, 1, 256, 8, 64]", + "name" : "var_400", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1)", + "shortDescription" : "", + "shape" : "[1]", + "name" : "var_744", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1 × 512 × 6)", + "shortDescription" : "", + "shape" : "[1, 512, 6]", + "name" : "var_542", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1)", + "shortDescription" : "", + "shape" : "[1]", + "name" : "conv0_first", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1 × 256 × 6)", + "shortDescription" : "", + "shape" : "[1, 256, 6]", + "name" : "var_565", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1 × 256 × 2)", + "shortDescription" : "", + "shape" : "[1, 256, 2]", + "name" : "var_585", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1)", + "shortDescription" : "", + "shape" : "[1]", + "name" : "res0_conv0_first", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1 × 128 × 0)", + "shortDescription" : "", + "shape" : "[1, 128, 0]", + "name" : "res0_conv1_prev", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1)", + "shortDescription" : "", + "shape" : "[1]", + "name" : "res0_conv1_first", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1 × 128 × 5)", + "shortDescription" : "", + "shape" : "[1, 128, 5]", + "name" : "var_621", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1 × 128 × 2)", + "shortDescription" : "", + "shape" : "[1, 128, 2]", + "name" : "var_641", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1)", + "shortDescription" : "", + "shape" : "[1]", + "name" : "res1_conv0_first", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1 × 64 × 0)", + "shortDescription" : "", + "shape" : "[1, 64, 0]", + "name" : "res1_conv1_prev", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1)", + "shortDescription" : "", + "shape" : "[1]", + "name" : "res1_conv1_first", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1 × 64 × 4)", + "shortDescription" : "", + "shape" : "[1, 64, 4]", + "name" : "var_677", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1 × 64 × 2)", + "shortDescription" : "", + "shape" : "[1, 64, 2]", + "name" : "var_697", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1)", + "shortDescription" : "", + "shape" : "[1]", + "name" : "res2_conv0_first", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1 × 32 × 0)", + "shortDescription" : "", + "shape" : "[1, 32, 0]", + "name" : "res2_conv1_prev", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1)", + "shortDescription" : "", + "shape" : "[1]", + "name" : "res2_conv1_first", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1 × 64 × 2)", + "shortDescription" : "", + "shape" : "[1, 64, 2]", + "name" : "var_725", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1)", + "shortDescription" : "", + "shape" : "[1]", + "name" : "conv_final_first", + "type" : "MultiArray" + } + ], + "modelParameters" : [ + + ], + "specificationVersion" : 8, + "mlProgramOperationTypeHistogram" : { + "Ios17.notEqual" : 4, + "Tile" : 2, + "Ios17.lessEqual" : 2, + "Ios17.layerNorm" : 4, + "Stack" : 6, + "Ios17.greaterEqual" : 4, + "Ios17.scatterAlongAxis" : 4, + "Ios17.sub" : 17, + "Identity" : 11, + "Ios17.elu" : 10, + "Ios17.logicalNot" : 2, + "Ios17.sliceByIndex" : 35, + "Ios17.concat" : 9, + "Ios17.convTranspose" : 4, + "Ios17.floorDiv" : 4, + "Ios17.expandDims" : 7, + "Ios17.less" : 2, + "Ios17.mul" : 29, + "Ios17.transpose" : 10, + "Ios17.matmul" : 4, + "Ios16.gelu" : 2, + "Ios17.conv" : 9, + "Ios17.reshape" : 22, + "Split" : 2, + "Ios17.squeeze" : 6, + "Ios17.cast" : 6, + "Ios17.sin" : 2, + "Ios17.linear" : 8, + "Select" : 6, + "Ios16.softmax" : 2, + "Ios17.add" : 28, + "Ios17.cos" : 2, + "Ios17.logicalAnd" : 6 + }, + "computePrecision" : "Mixed (Float32, Int32)", + "isUpdatable" : "0", + "stateSchema" : [ + + ], + "availability" : { + "macOS" : "14.0", + "tvOS" : "17.0", + "visionOS" : "1.0", + "watchOS" : "10.0", + "iOS" : "17.0", + "macCatalyst" : "17.0" + }, + "modelType" : { + "name" : "MLModelType_mlProgram" + }, + "userDefinedMetadata" : { + "com.github.apple.coremltools.conversion_date" : "2026-04-24", + "com.github.apple.coremltools.source" : "torch==2.9.1", + "com.github.apple.coremltools.version" : "9.0", + "com.github.apple.coremltools.source_dialect" : "TorchScript" + }, + "inputSchema" : [ + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1 × 32)", + "shortDescription" : "", + "shape" : "[1, 32]", + "name" : "latent", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1 × 512 × 16)", + "shortDescription" : "", + "shape" : "[1, 512, 16]", + "name" : "upsample_partial", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 2 × 1 × 256 × 8 × 64)", + "shortDescription" : "", + "shape" : "[2, 1, 256, 8, 64]", + "name" : "attn0_cache", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1)", + "shortDescription" : "", + "shape" : "[1]", + "name" : "attn0_offset", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 2 × 1 × 256 × 8 × 64)", + "shortDescription" : "", + "shape" : "[2, 1, 256, 8, 64]", + "name" : "attn1_cache", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1)", + "shortDescription" : "", + "shape" : "[1]", + "name" : "attn1_offset", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1 × 512 × 6)", + "shortDescription" : "", + "shape" : "[1, 512, 6]", + "name" : "conv0_prev", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1)", + "shortDescription" : "", + "shape" : "[1]", + "name" : "conv0_first", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1 × 256 × 6)", + "shortDescription" : "", + "shape" : "[1, 256, 6]", + "name" : "convtr0_partial", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1 × 256 × 2)", + "shortDescription" : "", + "shape" : "[1, 256, 2]", + "name" : "res0_conv0_prev", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1)", + "shortDescription" : "", + "shape" : "[1]", + "name" : "res0_conv0_first", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1 × 128 × 0)", + "shortDescription" : "", + "shape" : "[1, 128, 0]", + "name" : "res0_conv1_prev", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1)", + "shortDescription" : "", + "shape" : "[1]", + "name" : "res0_conv1_first", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1 × 128 × 5)", + "shortDescription" : "", + "shape" : "[1, 128, 5]", + "name" : "convtr1_partial", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1 × 128 × 2)", + "shortDescription" : "", + "shape" : "[1, 128, 2]", + "name" : "res1_conv0_prev", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1)", + "shortDescription" : "", + "shape" : "[1]", + "name" : "res1_conv0_first", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1 × 64 × 0)", + "shortDescription" : "", + "shape" : "[1, 64, 0]", + "name" : "res1_conv1_prev", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1)", + "shortDescription" : "", + "shape" : "[1]", + "name" : "res1_conv1_first", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1 × 64 × 4)", + "shortDescription" : "", + "shape" : "[1, 64, 4]", + "name" : "convtr2_partial", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1 × 64 × 2)", + "shortDescription" : "", + "shape" : "[1, 64, 2]", + "name" : "res2_conv0_prev", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1)", + "shortDescription" : "", + "shape" : "[1]", + "name" : "res2_conv0_first", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1 × 32 × 0)", + "shortDescription" : "", + "shape" : "[1, 32, 0]", + "name" : "res2_conv1_prev", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1)", + "shortDescription" : "", + "shape" : "[1]", + "name" : "res2_conv1_first", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1 × 64 × 2)", + "shortDescription" : "", + "shape" : "[1, 64, 2]", + "name" : "conv_final_prev", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1)", + "shortDescription" : "", + "shape" : "[1]", + "name" : "conv_final_first", + "type" : "MultiArray" + } + ], + "generatedClassName" : "mimi_decoder", + "method" : "predict" + } +] \ No newline at end of file diff --git a/v2.1/french_24l/mimi_decoder.mlmodelc/model.mil b/v2.1/french_24l/mimi_decoder.mlmodelc/model.mil new file mode 100644 index 0000000000000000000000000000000000000000..c96d4668e23dbb8f1f9e9d85b199ffa213c5c6b1 --- /dev/null +++ b/v2.1/french_24l/mimi_decoder.mlmodelc/model.mil @@ -0,0 +1,646 @@ +program(1.0) +[buildInfo = dict, tensor>({{"coremlc-component-MIL", "3520.4.1"}, {"coremlc-version", "3520.5.1"}, {"coremltools-component-torch", "2.9.1"}, {"coremltools-source-dialect", "TorchScript"}, {"coremltools-version", "9.0"}})] +{ + func main(tensor attn0_cache, tensor attn0_offset, tensor attn1_cache, tensor attn1_offset, tensor conv0_first, tensor conv0_prev, tensor conv_final_first, tensor conv_final_prev, tensor convtr0_partial, tensor convtr1_partial, tensor convtr2_partial, tensor latent, tensor res0_conv0_first, tensor res0_conv0_prev, tensor res0_conv1_first, tensor res0_conv1_prev, tensor res1_conv0_first, tensor res1_conv0_prev, tensor res1_conv1_first, tensor res1_conv1_prev, tensor res2_conv0_first, tensor res2_conv0_prev, tensor res2_conv1_first, tensor res2_conv1_prev, tensor upsample_partial) { + tensor emb_mean = const()[name = tensor("emb_mean"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(64)))]; + tensor emb_std = const()[name = tensor("emb_std"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(256)))]; + tensor mimi_quantizer_output_proj_weight = const()[name = tensor("mimi_quantizer_output_proj_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(448)))]; + tensor mimi_upsample_convtr_convtr_weight = const()[name = tensor("mimi_upsample_convtr_convtr_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(66048)))]; + tensor mimi_decoder_transformer_transformer_layers_0_norm1_bias = const()[name = tensor("mimi_decoder_transformer_transformer_layers_0_norm1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(131648)))]; + tensor mimi_decoder_transformer_transformer_layers_0_norm1_weight = const()[name = tensor("mimi_decoder_transformer_transformer_layers_0_norm1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(133760)))]; + tensor mimi_decoder_transformer_transformer_layers_0_self_attn_in_proj_weight = const()[name = tensor("mimi_decoder_transformer_transformer_layers_0_self_attn_in_proj_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(135872)))]; + tensor mimi_decoder_transformer_transformer_layers_0_self_attn_out_proj_weight = const()[name = tensor("mimi_decoder_transformer_transformer_layers_0_self_attn_out_proj_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(3281664)))]; + tensor mimi_decoder_transformer_transformer_layers_0_layer_scale_1_scale = const()[name = tensor("mimi_decoder_transformer_transformer_layers_0_layer_scale_1_scale"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(4330304)))]; + tensor mimi_decoder_transformer_transformer_layers_0_norm2_bias = const()[name = tensor("mimi_decoder_transformer_transformer_layers_0_norm2_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(4332416)))]; + tensor mimi_decoder_transformer_transformer_layers_0_norm2_weight = const()[name = tensor("mimi_decoder_transformer_transformer_layers_0_norm2_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(4334528)))]; + tensor mimi_decoder_transformer_transformer_layers_0_linear1_weight = const()[name = tensor("mimi_decoder_transformer_transformer_layers_0_linear1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(4336640)))]; + tensor mimi_decoder_transformer_transformer_layers_0_linear2_weight = const()[name = tensor("mimi_decoder_transformer_transformer_layers_0_linear2_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8531008)))]; + tensor mimi_decoder_transformer_transformer_layers_0_layer_scale_2_scale = const()[name = tensor("mimi_decoder_transformer_transformer_layers_0_layer_scale_2_scale"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(12725376)))]; + tensor mimi_decoder_transformer_transformer_layers_1_norm1_bias = const()[name = tensor("mimi_decoder_transformer_transformer_layers_1_norm1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(12727488)))]; + tensor mimi_decoder_transformer_transformer_layers_1_norm1_weight = const()[name = tensor("mimi_decoder_transformer_transformer_layers_1_norm1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(12729600)))]; + tensor mimi_decoder_transformer_transformer_layers_1_self_attn_in_proj_weight = const()[name = tensor("mimi_decoder_transformer_transformer_layers_1_self_attn_in_proj_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(12731712)))]; + tensor mimi_decoder_transformer_transformer_layers_1_self_attn_out_proj_weight = const()[name = tensor("mimi_decoder_transformer_transformer_layers_1_self_attn_out_proj_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15877504)))]; + tensor mimi_decoder_transformer_transformer_layers_1_layer_scale_1_scale = const()[name = tensor("mimi_decoder_transformer_transformer_layers_1_layer_scale_1_scale"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(16926144)))]; + tensor mimi_decoder_transformer_transformer_layers_1_norm2_bias = const()[name = tensor("mimi_decoder_transformer_transformer_layers_1_norm2_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(16928256)))]; + tensor mimi_decoder_transformer_transformer_layers_1_norm2_weight = const()[name = tensor("mimi_decoder_transformer_transformer_layers_1_norm2_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(16930368)))]; + tensor mimi_decoder_transformer_transformer_layers_1_linear1_weight = const()[name = tensor("mimi_decoder_transformer_transformer_layers_1_linear1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(16932480)))]; + tensor mimi_decoder_transformer_transformer_layers_1_linear2_weight = const()[name = tensor("mimi_decoder_transformer_transformer_layers_1_linear2_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21126848)))]; + tensor mimi_decoder_transformer_transformer_layers_1_layer_scale_2_scale = const()[name = tensor("mimi_decoder_transformer_transformer_layers_1_layer_scale_2_scale"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(25321216)))]; + tensor mimi_decoder_model_0_conv_bias = const()[name = tensor("mimi_decoder_model_0_conv_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(25323328)))]; + tensor mimi_decoder_model_0_conv_weight = const()[name = tensor("mimi_decoder_model_0_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(25325440)))]; + tensor mimi_decoder_model_2_convtr_bias = const()[name = tensor("mimi_decoder_model_2_convtr_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(32665536)))]; + tensor mimi_decoder_model_2_convtr_weight = const()[name = tensor("mimi_decoder_model_2_convtr_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(32666624)))]; + tensor mimi_decoder_model_3_block_1_conv_bias = const()[name = tensor("mimi_decoder_model_3_block_1_conv_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(38958144)))]; + tensor mimi_decoder_model_3_block_1_conv_weight = const()[name = tensor("mimi_decoder_model_3_block_1_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(38958720)))]; + tensor mimi_decoder_model_3_block_3_conv_bias = const()[name = tensor("mimi_decoder_model_3_block_3_conv_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(39352000)))]; + tensor mimi_decoder_model_3_block_3_conv_weight = const()[name = tensor("mimi_decoder_model_3_block_3_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(39353088)))]; + tensor mimi_decoder_model_5_convtr_bias = const()[name = tensor("mimi_decoder_model_5_convtr_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(39484224)))]; + tensor mimi_decoder_model_5_convtr_weight = const()[name = tensor("mimi_decoder_model_5_convtr_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(39484800)))]; + tensor mimi_decoder_model_6_block_1_conv_bias = const()[name = tensor("mimi_decoder_model_6_block_1_conv_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(40795584)))]; + tensor mimi_decoder_model_6_block_1_conv_weight = const()[name = tensor("mimi_decoder_model_6_block_1_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(40795904)))]; + tensor mimi_decoder_model_6_block_3_conv_bias = const()[name = tensor("mimi_decoder_model_6_block_3_conv_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(40894272)))]; + tensor mimi_decoder_model_6_block_3_conv_weight = const()[name = tensor("mimi_decoder_model_6_block_3_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(40894848)))]; + tensor mimi_decoder_model_8_convtr_bias = const()[name = tensor("mimi_decoder_model_8_convtr_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(40927680)))]; + tensor mimi_decoder_model_8_convtr_weight = const()[name = tensor("mimi_decoder_model_8_convtr_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(40928000)))]; + tensor mimi_decoder_model_9_block_1_conv_bias = const()[name = tensor("mimi_decoder_model_9_block_1_conv_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(41190208)))]; + tensor mimi_decoder_model_9_block_1_conv_weight = const()[name = tensor("mimi_decoder_model_9_block_1_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(41190400)))]; + tensor mimi_decoder_model_9_block_3_conv_bias = const()[name = tensor("mimi_decoder_model_9_block_3_conv_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(41215040)))]; + tensor mimi_decoder_model_9_block_3_conv_weight = const()[name = tensor("mimi_decoder_model_9_block_3_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(41215360)))]; + tensor mimi_decoder_model_11_conv_bias = const()[name = tensor("mimi_decoder_model_11_conv_bias"), val = tensor([0x1.16p-13])]; + tensor mimi_decoder_model_11_conv_weight = const()[name = tensor("mimi_decoder_model_11_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(41223616)))]; + tensor var_38 = mul(x = latent, y = emb_std)[name = tensor("op_38")]; + tensor denorm = add(x = var_38, y = emb_mean)[name = tensor("denorm")]; + tensor input_1_axes_0 = const()[name = tensor("input_1_axes_0"), val = tensor([-1])]; + tensor input_1 = expand_dims(axes = input_1_axes_0, x = denorm)[name = tensor("input_1")]; + tensor x_1_pad_type_0 = const()[name = tensor("x_1_pad_type_0"), val = tensor("valid")]; + tensor x_1_strides_0 = const()[name = tensor("x_1_strides_0"), val = tensor([1])]; + tensor x_1_pad_0 = const()[name = tensor("x_1_pad_0"), val = tensor([0, 0])]; + tensor x_1_dilations_0 = const()[name = tensor("x_1_dilations_0"), val = tensor([1])]; + tensor x_1_groups_0 = const()[name = tensor("x_1_groups_0"), val = tensor(1)]; + tensor x_1 = conv(dilations = x_1_dilations_0, groups = x_1_groups_0, pad = x_1_pad_0, pad_type = x_1_pad_type_0, strides = x_1_strides_0, weight = mimi_quantizer_output_proj_weight, x = input_1)[name = tensor("x_1")]; + tensor var_62 = const()[name = tensor("op_62"), val = tensor(-1)]; + tensor y_1_pad_type_0 = const()[name = tensor("y_1_pad_type_0"), val = tensor("valid")]; + tensor y_1_strides_0 = const()[name = tensor("y_1_strides_0"), val = tensor([16])]; + tensor y_1_groups_0 = const()[name = tensor("y_1_groups_0"), val = tensor(512)]; + tensor y_1_pad_0 = const()[name = tensor("y_1_pad_0"), val = tensor([0, 0])]; + tensor y_1_dilations_0 = const()[name = tensor("y_1_dilations_0"), val = tensor([1])]; + tensor y_1_has_output_shape_output_shape_0 = const()[name = tensor("y_1_has_output_shape_output_shape_0"), val = tensor([1, 512, 32])]; + tensor y_1_has_output_shape = conv_transpose(dilations = y_1_dilations_0, groups = y_1_groups_0, output_shape = y_1_has_output_shape_output_shape_0, pad = y_1_pad_0, pad_type = y_1_pad_type_0, strides = y_1_strides_0, weight = mimi_upsample_convtr_convtr_weight, x = x_1)[name = tensor("y_1_has_output_shape")]; + tensor var_72_begin_0 = const()[name = tensor("op_72_begin_0"), val = tensor([0, 0, 0])]; + tensor var_72_end_0 = const()[name = tensor("op_72_end_0"), val = tensor([1, 512, 16])]; + tensor var_72_end_mask_0 = const()[name = tensor("op_72_end_mask_0"), val = tensor([true, true, false])]; + tensor var_72 = slice_by_index(begin = var_72_begin_0, end = var_72_end_0, end_mask = var_72_end_mask_0, x = y_1_has_output_shape)[name = tensor("op_72")]; + tensor var_73 = add(x = var_72, y = upsample_partial)[name = tensor("op_73")]; + tensor var_74_begin_0 = const()[name = tensor("op_74_begin_0"), val = tensor([0, 0, 16])]; + tensor var_74_end_0 = const()[name = tensor("op_74_end_0"), val = tensor([1, 512, 32])]; + tensor var_74_end_mask_0 = const()[name = tensor("op_74_end_mask_0"), val = tensor([true, true, true])]; + tensor var_74 = slice_by_index(begin = var_74_begin_0, end = var_74_end_0, end_mask = var_74_end_mask_0, x = y_1_has_output_shape)[name = tensor("op_74")]; + tensor y_3_interleave_0 = const()[name = tensor("y_3_interleave_0"), val = tensor(false)]; + tensor y_3 = concat(axis = var_62, interleave = y_3_interleave_0, values = (var_73, var_74))[name = tensor("y_3")]; + tensor var_77_begin_0 = const()[name = tensor("op_77_begin_0"), val = tensor([0, 0, 16])]; + tensor var_77_end_0 = const()[name = tensor("op_77_end_0"), val = tensor([1, 512, 32])]; + tensor var_77_end_mask_0 = const()[name = tensor("op_77_end_mask_0"), val = tensor([true, true, true])]; + tensor var_77 = slice_by_index(begin = var_77_begin_0, end = var_77_end_0, end_mask = var_77_end_mask_0, x = y_3)[name = tensor("op_77")]; + tensor x_3_begin_0 = const()[name = tensor("x_3_begin_0"), val = tensor([0, 0, 0])]; + tensor x_3_end_0 = const()[name = tensor("x_3_end_0"), val = tensor([1, 512, 16])]; + tensor x_3_end_mask_0 = const()[name = tensor("x_3_end_mask_0"), val = tensor([true, true, false])]; + tensor x_3 = slice_by_index(begin = x_3_begin_0, end = x_3_end_0, end_mask = x_3_end_mask_0, x = y_3)[name = tensor("x_3")]; + tensor var_86 = const()[name = tensor("op_86"), val = tensor(0)]; + tensor var_91 = const()[name = tensor("op_91"), val = tensor(-1)]; + tensor var_100 = const()[name = tensor("op_100"), val = tensor(-0x1.ff933cp+127)]; + tensor var_102 = const()[name = tensor("op_102"), val = tensor(0x1.4f8b58p-17)]; + tensor input_3_perm_0 = const()[name = tensor("input_3_perm_0"), val = tensor([0, 2, 1])]; + tensor query_1_axes_0 = const()[name = tensor("query_1_axes_0"), val = tensor([-1])]; + tensor input_3 = transpose(perm = input_3_perm_0, x = x_3)[name = tensor("transpose_19")]; + tensor query_1 = layer_norm(axes = query_1_axes_0, beta = mimi_decoder_transformer_transformer_layers_0_norm1_bias, epsilon = var_102, gamma = mimi_decoder_transformer_transformer_layers_0_norm1_weight, x = input_3)[name = tensor("query_1")]; + tensor linear_0_bias_0 = const()[name = tensor("linear_0_bias_0"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(41224448)))]; + tensor projected_1 = linear(bias = linear_0_bias_0, weight = mimi_decoder_transformer_transformer_layers_0_self_attn_in_proj_weight, x = query_1)[name = tensor("linear_0")]; + tensor var_130 = const()[name = tensor("op_130"), val = tensor([1, 16, 3, 8, 64])]; + tensor packed_1 = reshape(shape = var_130, x = projected_1)[name = tensor("packed_1")]; + tensor var_132_split_sizes_0 = const()[name = tensor("op_132_split_sizes_0"), val = tensor([1, 1, 1])]; + tensor var_132_axis_0 = const()[name = tensor("op_132_axis_0"), val = tensor(2)]; + tensor var_132_0, tensor var_132_1, tensor var_132_2 = split(axis = var_132_axis_0, split_sizes = var_132_split_sizes_0, x = packed_1)[name = tensor("op_132")]; + tensor squeeze_0_axes_0 = const()[name = tensor("squeeze_0_axes_0"), val = tensor([2])]; + tensor squeeze_0 = squeeze(axes = squeeze_0_axes_0, x = var_132_0)[name = tensor("squeeze_0")]; + tensor squeeze_1_axes_0 = const()[name = tensor("squeeze_1_axes_0"), val = tensor([2])]; + tensor squeeze_1 = squeeze(axes = squeeze_1_axes_0, x = var_132_1)[name = tensor("squeeze_1")]; + tensor squeeze_2_axes_0 = const()[name = tensor("squeeze_2_axes_0"), val = tensor([2])]; + tensor squeeze_2 = squeeze(axes = squeeze_2_axes_0, x = var_132_2)[name = tensor("squeeze_2")]; + tensor offset_3_begin_0 = const()[name = tensor("offset_3_begin_0"), val = tensor([0])]; + tensor offset_3_end_0 = const()[name = tensor("offset_3_end_0"), val = tensor([1])]; + tensor offset_3_end_mask_0 = const()[name = tensor("offset_3_end_mask_0"), val = tensor([false])]; + tensor offset_3_squeeze_mask_0 = const()[name = tensor("offset_3_squeeze_mask_0"), val = tensor([true])]; + tensor offset_3 = slice_by_index(begin = offset_3_begin_0, end = offset_3_end_0, end_mask = offset_3_end_mask_0, squeeze_mask = offset_3_squeeze_mask_0, x = attn0_offset)[name = tensor("offset_3")]; + tensor freqs_1 = const()[name = tensor("freqs_1"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(41230656)))]; + tensor ts_1_promoted = const()[name = tensor("ts_1_promoted"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(41230848)))]; + tensor ts_3 = add(x = ts_1_promoted, y = offset_3)[name = tensor("ts_3")]; + tensor var_148 = const()[name = tensor("op_148"), val = tensor([-1, 1, 1])]; + tensor ts_5 = reshape(shape = var_148, x = ts_3)[name = tensor("ts_5")]; + tensor var_150 = const()[name = tensor("op_150"), val = tensor([1, 16, 8, 32, 2])]; + tensor q_3 = reshape(shape = var_150, x = squeeze_0)[name = tensor("q_3")]; + tensor var_152 = const()[name = tensor("op_152"), val = tensor([1, 16, 8, 32, 2])]; + tensor k_3 = reshape(shape = var_152, x = squeeze_1)[name = tensor("k_3")]; + tensor var_154_begin_0 = const()[name = tensor("op_154_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_154_end_0 = const()[name = tensor("op_154_end_0"), val = tensor([1, 16, 8, 32, 1])]; + tensor var_154_end_mask_0 = const()[name = tensor("op_154_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_154_squeeze_mask_0 = const()[name = tensor("op_154_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_154 = slice_by_index(begin = var_154_begin_0, end = var_154_end_0, end_mask = var_154_end_mask_0, squeeze_mask = var_154_squeeze_mask_0, x = q_3)[name = tensor("op_154")]; + tensor var_156_begin_0 = const()[name = tensor("op_156_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_156_end_0 = const()[name = tensor("op_156_end_0"), val = tensor([1, 16, 8, 32, 2])]; + tensor var_156_end_mask_0 = const()[name = tensor("op_156_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_156_squeeze_mask_0 = const()[name = tensor("op_156_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_156 = slice_by_index(begin = var_156_begin_0, end = var_156_end_0, end_mask = var_156_end_mask_0, squeeze_mask = var_156_squeeze_mask_0, x = q_3)[name = tensor("op_156")]; + tensor var_158_begin_0 = const()[name = tensor("op_158_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_158_end_0 = const()[name = tensor("op_158_end_0"), val = tensor([1, 16, 8, 32, 1])]; + tensor var_158_end_mask_0 = const()[name = tensor("op_158_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_158_squeeze_mask_0 = const()[name = tensor("op_158_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_158 = slice_by_index(begin = var_158_begin_0, end = var_158_end_0, end_mask = var_158_end_mask_0, squeeze_mask = var_158_squeeze_mask_0, x = k_3)[name = tensor("op_158")]; + tensor var_160_begin_0 = const()[name = tensor("op_160_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_160_end_0 = const()[name = tensor("op_160_end_0"), val = tensor([1, 16, 8, 32, 2])]; + tensor var_160_end_mask_0 = const()[name = tensor("op_160_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_160_squeeze_mask_0 = const()[name = tensor("op_160_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_160 = slice_by_index(begin = var_160_begin_0, end = var_160_end_0, end_mask = var_160_end_mask_0, squeeze_mask = var_160_squeeze_mask_0, x = k_3)[name = tensor("op_160")]; + tensor var_162 = mul(x = freqs_1, y = ts_5)[name = tensor("op_162")]; + tensor rotr_1 = cos(x = var_162)[name = tensor("rotr_1")]; + tensor roti_1 = sin(x = var_162)[name = tensor("roti_1")]; + tensor var_166 = mul(x = var_154, y = rotr_1)[name = tensor("op_166")]; + tensor var_167 = mul(x = var_156, y = roti_1)[name = tensor("op_167")]; + tensor qor_1 = sub(x = var_166, y = var_167)[name = tensor("qor_1")]; + tensor var_169 = mul(x = var_154, y = roti_1)[name = tensor("op_169")]; + tensor var_170 = mul(x = var_156, y = rotr_1)[name = tensor("op_170")]; + tensor qoi_1 = add(x = var_169, y = var_170)[name = tensor("qoi_1")]; + tensor var_172 = mul(x = var_158, y = rotr_1)[name = tensor("op_172")]; + tensor var_173 = mul(x = var_160, y = roti_1)[name = tensor("op_173")]; + tensor kor_1 = sub(x = var_172, y = var_173)[name = tensor("kor_1")]; + tensor var_175 = mul(x = var_158, y = roti_1)[name = tensor("op_175")]; + tensor var_176 = mul(x = var_160, y = rotr_1)[name = tensor("op_176")]; + tensor koi_1 = add(x = var_175, y = var_176)[name = tensor("koi_1")]; + tensor qo_1_axis_0 = const()[name = tensor("qo_1_axis_0"), val = tensor(-1)]; + tensor qo_1 = stack(axis = qo_1_axis_0, values = (qor_1, qoi_1))[name = tensor("qo_1")]; + tensor ko_1_axis_0 = const()[name = tensor("ko_1_axis_0"), val = tensor(-1)]; + tensor ko_1 = stack(axis = ko_1_axis_0, values = (kor_1, koi_1))[name = tensor("ko_1")]; + tensor var_186 = const()[name = tensor("op_186"), val = tensor([1, 16, 8, 64])]; + tensor q_5 = reshape(shape = var_186, x = qo_1)[name = tensor("q_5")]; + tensor var_188 = const()[name = tensor("op_188"), val = tensor([1, 16, 8, 64])]; + tensor k_5 = reshape(shape = var_188, x = ko_1)[name = tensor("k_5")]; + tensor capacity_1 = const()[name = tensor("capacity_1"), val = tensor([256])]; + tensor var_193_dtype_0 = const()[name = tensor("op_193_dtype_0"), val = tensor("int32")]; + tensor var_194 = const()[name = tensor("op_194"), val = tensor([1, 1])]; + tensor var_193 = cast(dtype = var_193_dtype_0, x = attn0_offset)[name = tensor("cast_49")]; + tensor write_base_1 = reshape(shape = var_194, x = var_193)[name = tensor("write_base_1")]; + tensor write_range_1 = const()[name = tensor("write_range_1"), val = tensor([[0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15]])]; + tensor abs_idx_1 = add(x = write_base_1, y = write_range_1)[name = tensor("abs_idx_1")]; + tensor wrapped_1_div = floor_div(x = abs_idx_1, y = capacity_1)[name = tensor("wrapped_1_div")]; + tensor wrapped_1_div_scaled = mul(x = wrapped_1_div, y = capacity_1)[name = tensor("wrapped_1_div_scaled")]; + tensor wrapped_1 = sub(x = abs_idx_1, y = wrapped_1_div_scaled)[name = tensor("wrapped_1")]; + tensor var_201 = const()[name = tensor("op_201"), val = tensor([1, 16, 1, 1])]; + tensor var_202 = reshape(shape = var_201, x = wrapped_1)[name = tensor("op_202")]; + tensor write_indexes_1_reps_0 = const()[name = tensor("write_indexes_1_reps_0"), val = tensor([1, 1, 8, 64])]; + tensor write_indexes_1 = tile(reps = write_indexes_1_reps_0, x = var_202)[name = tensor("write_indexes_1")]; + tensor var_205_begin_0 = const()[name = tensor("op_205_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_205_end_0 = const()[name = tensor("op_205_end_0"), val = tensor([1, 1, 256, 8, 64])]; + tensor var_205_end_mask_0 = const()[name = tensor("op_205_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_205_squeeze_mask_0 = const()[name = tensor("op_205_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor var_205 = slice_by_index(begin = var_205_begin_0, end = var_205_end_0, end_mask = var_205_end_mask_0, squeeze_mask = var_205_squeeze_mask_0, x = attn0_cache)[name = tensor("op_205")]; + tensor new_k_cache_1_axis_0 = const()[name = tensor("new_k_cache_1_axis_0"), val = tensor(1)]; + tensor new_k_cache_1_mode_0 = const()[name = tensor("new_k_cache_1_mode_0"), val = tensor("update")]; + tensor new_k_cache_1_validate_indices_0 = const()[name = tensor("new_k_cache_1_validate_indices_0"), val = tensor(false)]; + tensor new_k_cache_1 = scatter_along_axis(axis = new_k_cache_1_axis_0, data = var_205, indices = write_indexes_1, mode = new_k_cache_1_mode_0, updates = k_5, validate_indices = new_k_cache_1_validate_indices_0)[name = tensor("new_k_cache_1")]; + tensor var_207_begin_0 = const()[name = tensor("op_207_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor var_207_end_0 = const()[name = tensor("op_207_end_0"), val = tensor([2, 1, 256, 8, 64])]; + tensor var_207_end_mask_0 = const()[name = tensor("op_207_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_207_squeeze_mask_0 = const()[name = tensor("op_207_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor var_207 = slice_by_index(begin = var_207_begin_0, end = var_207_end_0, end_mask = var_207_end_mask_0, squeeze_mask = var_207_squeeze_mask_0, x = attn0_cache)[name = tensor("op_207")]; + tensor new_v_cache_1_axis_0 = const()[name = tensor("new_v_cache_1_axis_0"), val = tensor(1)]; + tensor new_v_cache_1_mode_0 = const()[name = tensor("new_v_cache_1_mode_0"), val = tensor("update")]; + tensor new_v_cache_1_validate_indices_0 = const()[name = tensor("new_v_cache_1_validate_indices_0"), val = tensor(false)]; + tensor new_v_cache_1 = scatter_along_axis(axis = new_v_cache_1_axis_0, data = var_207, indices = write_indexes_1, mode = new_v_cache_1_mode_0, updates = squeeze_2, validate_indices = new_v_cache_1_validate_indices_0)[name = tensor("new_v_cache_1")]; + tensor var_210_axis_0 = const()[name = tensor("op_210_axis_0"), val = tensor(0)]; + tensor var_210 = stack(axis = var_210_axis_0, values = (new_k_cache_1, new_v_cache_1))[name = tensor("op_210")]; + tensor var_211 = not_equal(x = new_k_cache_1, y = new_k_cache_1)[name = tensor("op_211")]; + tensor var_212 = const()[name = tensor("op_212"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(41230976)))]; + tensor new_k_cache_3 = select(a = var_212, b = new_k_cache_1, cond = var_211)[name = tensor("new_k_cache_3")]; + tensor var_214 = not_equal(x = new_v_cache_1, y = new_v_cache_1)[name = tensor("op_214")]; + tensor new_v_cache_3 = select(a = var_212, b = new_v_cache_1, cond = var_214)[name = tensor("new_v_cache_3")]; + tensor var_219 = const()[name = tensor("op_219"), val = tensor([0, 2, 1, 3])]; + tensor var_221 = const()[name = tensor("op_221"), val = tensor([1, 1])]; + tensor var_222 = reshape(shape = var_221, x = attn0_offset)[name = tensor("op_222")]; + tensor var_224_promoted = const()[name = tensor("op_224_promoted"), val = tensor([0x1.ep+3])]; + tensor var_225 = add(x = var_222, y = var_224_promoted)[name = tensor("op_225")]; + tensor last_pos_1_dtype_0 = const()[name = tensor("last_pos_1_dtype_0"), val = tensor("int32")]; + tensor slot_idx_1 = const()[name = tensor("slot_idx_1"), val = tensor([[0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19, 20, 21, 22, 23, 24, 25, 26, 27, 28, 29, 30, 31, 32, 33, 34, 35, 36, 37, 38, 39, 40, 41, 42, 43, 44, 45, 46, 47, 48, 49, 50, 51, 52, 53, 54, 55, 56, 57, 58, 59, 60, 61, 62, 63, 64, 65, 66, 67, 68, 69, 70, 71, 72, 73, 74, 75, 76, 77, 78, 79, 80, 81, 82, 83, 84, 85, 86, 87, 88, 89, 90, 91, 92, 93, 94, 95, 96, 97, 98, 99, 100, 101, 102, 103, 104, 105, 106, 107, 108, 109, 110, 111, 112, 113, 114, 115, 116, 117, 118, 119, 120, 121, 122, 123, 124, 125, 126, 127, 128, 129, 130, 131, 132, 133, 134, 135, 136, 137, 138, 139, 140, 141, 142, 143, 144, 145, 146, 147, 148, 149, 150, 151, 152, 153, 154, 155, 156, 157, 158, 159, 160, 161, 162, 163, 164, 165, 166, 167, 168, 169, 170, 171, 172, 173, 174, 175, 176, 177, 178, 179, 180, 181, 182, 183, 184, 185, 186, 187, 188, 189, 190, 191, 192, 193, 194, 195, 196, 197, 198, 199, 200, 201, 202, 203, 204, 205, 206, 207, 208, 209, 210, 211, 212, 213, 214, 215, 216, 217, 218, 219, 220, 221, 222, 223, 224, 225, 226, 227, 228, 229, 230, 231, 232, 233, 234, 235, 236, 237, 238, 239, 240, 241, 242, 243, 244, 245, 246, 247, 248, 249, 250, 251, 252, 253, 254, 255]])]; + tensor last_pos_1 = cast(dtype = last_pos_1_dtype_0, x = var_225)[name = tensor("cast_48")]; + tensor diff_1 = sub(x = last_pos_1, y = slot_idx_1)[name = tensor("diff_1")]; + tensor var_231_div = floor_div(x = diff_1, y = capacity_1)[name = tensor("op_231_div")]; + tensor var_231_div_scaled = mul(x = var_231_div, y = capacity_1)[name = tensor("op_231_div_scaled")]; + tensor var_231 = sub(x = diff_1, y = var_231_div_scaled)[name = tensor("op_231")]; + tensor pos_k_1 = sub(x = last_pos_1, y = var_231)[name = tensor("pos_k_1")]; + tensor var_237_promoted = const()[name = tensor("op_237_promoted"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(41755328)))]; + tensor pos_q_1 = add(x = var_222, y = var_237_promoted)[name = tensor("pos_q_1")]; + tensor var_241_axes_0 = const()[name = tensor("op_241_axes_0"), val = tensor([2])]; + tensor var_241 = expand_dims(axes = var_241_axes_0, x = pos_q_1)[name = tensor("op_241")]; + tensor var_243_axes_0 = const()[name = tensor("op_243_axes_0"), val = tensor([1])]; + tensor var_243 = expand_dims(axes = var_243_axes_0, x = pos_k_1)[name = tensor("op_243")]; + tensor var_244_promoted_dtype_0 = const()[name = tensor("op_244_promoted_dtype_0"), val = tensor("fp32")]; + tensor var_244_promoted = cast(dtype = var_244_promoted_dtype_0, x = var_243)[name = tensor("cast_47")]; + tensor delta_1 = sub(x = var_241, y = var_244_promoted)[name = tensor("delta_1")]; + tensor valid_1 = greater_equal(x = var_243, y = var_86)[name = tensor("valid_1")]; + tensor var_253 = const()[name = tensor("op_253"), val = tensor([1, 1, 1])]; + tensor var_254 = reshape(shape = var_253, x = attn0_offset)[name = tensor("op_254")]; + tensor var_256_promoted = const()[name = tensor("op_256_promoted"), val = tensor([0x1.ep+3])]; + tensor var_257 = add(x = var_254, y = var_256_promoted)[name = tensor("op_257")]; + tensor var_258 = less_equal(x = var_244_promoted, y = var_257)[name = tensor("op_258")]; + tensor valid_3 = logical_and(x = valid_1, y = var_258)[name = tensor("valid_3")]; + tensor var_86_promoted = const()[name = tensor("op_86_promoted"), val = tensor(0x0p+0)]; + tensor var_260 = greater_equal(x = delta_1, y = var_86_promoted)[name = tensor("op_260")]; + tensor attn_mask_1 = logical_and(x = valid_3, y = var_260)[name = tensor("attn_mask_1")]; + tensor var_98_promoted = const()[name = tensor("op_98_promoted"), val = tensor(0x1.f4p+7)]; + tensor var_262 = less(x = delta_1, y = var_98_promoted)[name = tensor("op_262")]; + tensor attn_mask_3 = logical_and(x = attn_mask_1, y = var_262)[name = tensor("attn_mask_3")]; + tensor attn_mask_5_axes_0 = const()[name = tensor("attn_mask_5_axes_0"), val = tensor([1])]; + tensor attn_mask_5 = expand_dims(axes = attn_mask_5_axes_0, x = attn_mask_3)[name = tensor("attn_mask_5")]; + tensor var_267_transpose_x_0 = const()[name = tensor("op_267_transpose_x_0"), val = tensor(false)]; + tensor var_267_transpose_y_0 = const()[name = tensor("op_267_transpose_y_0"), val = tensor(false)]; + tensor transpose_6_perm_0 = const()[name = tensor("transpose_6_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_7_perm_0 = const()[name = tensor("transpose_7_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_7 = transpose(perm = transpose_7_perm_0, x = new_k_cache_3)[name = tensor("transpose_16")]; + tensor transpose_6 = transpose(perm = transpose_6_perm_0, x = q_5)[name = tensor("transpose_17")]; + tensor var_267 = matmul(transpose_x = var_267_transpose_x_0, transpose_y = var_267_transpose_y_0, x = transpose_6, y = transpose_7)[name = tensor("op_267")]; + tensor var_268 = const()[name = tensor("op_268"), val = tensor(0x1p-3)]; + tensor attn_1 = mul(x = var_267, y = var_268)[name = tensor("attn_1")]; + tensor var_270 = logical_not(x = attn_mask_5)[name = tensor("op_270")]; + tensor attn_3 = select(a = var_100, b = attn_1, cond = var_270)[name = tensor("attn_3")]; + tensor attn_5 = softmax(axis = var_91, x = attn_3)[name = tensor("attn_5")]; + tensor x_5_transpose_x_0 = const()[name = tensor("x_5_transpose_x_0"), val = tensor(false)]; + tensor x_5_transpose_y_0 = const()[name = tensor("x_5_transpose_y_0"), val = tensor(false)]; + tensor v_attn_1 = transpose(perm = var_219, x = new_v_cache_3)[name = tensor("transpose_18")]; + tensor x_5 = matmul(transpose_x = x_5_transpose_x_0, transpose_y = x_5_transpose_y_0, x = attn_5, y = v_attn_1)[name = tensor("x_5")]; + tensor var_274_perm_0 = const()[name = tensor("op_274_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_275 = const()[name = tensor("op_275"), val = tensor([1, 16, 512])]; + tensor var_274 = transpose(perm = var_274_perm_0, x = x_5)[name = tensor("transpose_15")]; + tensor input_5 = reshape(shape = var_275, x = var_274)[name = tensor("input_5")]; + tensor linear_1_bias_0 = const()[name = tensor("linear_1_bias_0"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(41755456)))]; + tensor x_7 = linear(bias = linear_1_bias_0, weight = mimi_decoder_transformer_transformer_layers_0_self_attn_out_proj_weight, x = input_5)[name = tensor("linear_1")]; + tensor var_284 = mul(x = mimi_decoder_transformer_transformer_layers_0_layer_scale_1_scale, y = x_7)[name = tensor("op_284")]; + tensor input_7 = add(x = input_3, y = var_284)[name = tensor("input_7")]; + tensor input_9_axes_0 = const()[name = tensor("input_9_axes_0"), val = tensor([-1])]; + tensor input_9 = layer_norm(axes = input_9_axes_0, beta = mimi_decoder_transformer_transformer_layers_0_norm2_bias, epsilon = var_102, gamma = mimi_decoder_transformer_transformer_layers_0_norm2_weight, x = input_7)[name = tensor("input_9")]; + tensor linear_2_bias_0 = const()[name = tensor("linear_2_bias_0"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(41757568)))]; + tensor var_291 = linear(bias = linear_2_bias_0, weight = mimi_decoder_transformer_transformer_layers_0_linear1_weight, x = input_9)[name = tensor("linear_2")]; + tensor input_11_mode_0 = const()[name = tensor("input_11_mode_0"), val = tensor("EXACT")]; + tensor input_11 = gelu(mode = input_11_mode_0, x = var_291)[name = tensor("input_11")]; + tensor x_9 = linear(bias = linear_1_bias_0, weight = mimi_decoder_transformer_transformer_layers_0_linear2_weight, x = input_11)[name = tensor("linear_3")]; + tensor var_297 = mul(x = mimi_decoder_transformer_transformer_layers_0_layer_scale_2_scale, y = x_9)[name = tensor("op_297")]; + tensor input_13 = add(x = input_7, y = var_297)[name = tensor("input_13")]; + tensor query_axes_0 = const()[name = tensor("query_axes_0"), val = tensor([-1])]; + tensor query = layer_norm(axes = query_axes_0, beta = mimi_decoder_transformer_transformer_layers_1_norm1_bias, epsilon = var_102, gamma = mimi_decoder_transformer_transformer_layers_1_norm1_weight, x = input_13)[name = tensor("query")]; + tensor projected = linear(bias = linear_0_bias_0, weight = mimi_decoder_transformer_transformer_layers_1_self_attn_in_proj_weight, x = query)[name = tensor("linear_4")]; + tensor var_320 = const()[name = tensor("op_320"), val = tensor([1, 16, 3, 8, 64])]; + tensor packed = reshape(shape = var_320, x = projected)[name = tensor("packed")]; + tensor var_322_split_sizes_0 = const()[name = tensor("op_322_split_sizes_0"), val = tensor([1, 1, 1])]; + tensor var_322_axis_0 = const()[name = tensor("op_322_axis_0"), val = tensor(2)]; + tensor var_322_0, tensor var_322_1, tensor var_322_2 = split(axis = var_322_axis_0, split_sizes = var_322_split_sizes_0, x = packed)[name = tensor("op_322")]; + tensor squeeze_3_axes_0 = const()[name = tensor("squeeze_3_axes_0"), val = tensor([2])]; + tensor squeeze_3 = squeeze(axes = squeeze_3_axes_0, x = var_322_0)[name = tensor("squeeze_3")]; + tensor squeeze_4_axes_0 = const()[name = tensor("squeeze_4_axes_0"), val = tensor([2])]; + tensor squeeze_4 = squeeze(axes = squeeze_4_axes_0, x = var_322_1)[name = tensor("squeeze_4")]; + tensor squeeze_5_axes_0 = const()[name = tensor("squeeze_5_axes_0"), val = tensor([2])]; + tensor squeeze_5 = squeeze(axes = squeeze_5_axes_0, x = var_322_2)[name = tensor("squeeze_5")]; + tensor offset_begin_0 = const()[name = tensor("offset_begin_0"), val = tensor([0])]; + tensor offset_end_0 = const()[name = tensor("offset_end_0"), val = tensor([1])]; + tensor offset_end_mask_0 = const()[name = tensor("offset_end_mask_0"), val = tensor([false])]; + tensor offset_squeeze_mask_0 = const()[name = tensor("offset_squeeze_mask_0"), val = tensor([true])]; + tensor offset = slice_by_index(begin = offset_begin_0, end = offset_end_0, end_mask = offset_end_mask_0, squeeze_mask = offset_squeeze_mask_0, x = attn1_offset)[name = tensor("offset")]; + tensor freqs = const()[name = tensor("freqs"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(41765824)))]; + tensor ts_7_promoted = const()[name = tensor("ts_7_promoted"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(41766016)))]; + tensor ts_9 = add(x = ts_7_promoted, y = offset)[name = tensor("ts_9")]; + tensor var_338 = const()[name = tensor("op_338"), val = tensor([-1, 1, 1])]; + tensor ts = reshape(shape = var_338, x = ts_9)[name = tensor("ts")]; + tensor var_340 = const()[name = tensor("op_340"), val = tensor([1, 16, 8, 32, 2])]; + tensor q_9 = reshape(shape = var_340, x = squeeze_3)[name = tensor("q_9")]; + tensor var_342 = const()[name = tensor("op_342"), val = tensor([1, 16, 8, 32, 2])]; + tensor k_9 = reshape(shape = var_342, x = squeeze_4)[name = tensor("k_9")]; + tensor var_344_begin_0 = const()[name = tensor("op_344_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_344_end_0 = const()[name = tensor("op_344_end_0"), val = tensor([1, 16, 8, 32, 1])]; + tensor var_344_end_mask_0 = const()[name = tensor("op_344_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_344_squeeze_mask_0 = const()[name = tensor("op_344_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_344 = slice_by_index(begin = var_344_begin_0, end = var_344_end_0, end_mask = var_344_end_mask_0, squeeze_mask = var_344_squeeze_mask_0, x = q_9)[name = tensor("op_344")]; + tensor var_346_begin_0 = const()[name = tensor("op_346_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_346_end_0 = const()[name = tensor("op_346_end_0"), val = tensor([1, 16, 8, 32, 2])]; + tensor var_346_end_mask_0 = const()[name = tensor("op_346_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_346_squeeze_mask_0 = const()[name = tensor("op_346_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_346 = slice_by_index(begin = var_346_begin_0, end = var_346_end_0, end_mask = var_346_end_mask_0, squeeze_mask = var_346_squeeze_mask_0, x = q_9)[name = tensor("op_346")]; + tensor var_348_begin_0 = const()[name = tensor("op_348_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_348_end_0 = const()[name = tensor("op_348_end_0"), val = tensor([1, 16, 8, 32, 1])]; + tensor var_348_end_mask_0 = const()[name = tensor("op_348_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_348_squeeze_mask_0 = const()[name = tensor("op_348_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_348 = slice_by_index(begin = var_348_begin_0, end = var_348_end_0, end_mask = var_348_end_mask_0, squeeze_mask = var_348_squeeze_mask_0, x = k_9)[name = tensor("op_348")]; + tensor var_350_begin_0 = const()[name = tensor("op_350_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_350_end_0 = const()[name = tensor("op_350_end_0"), val = tensor([1, 16, 8, 32, 2])]; + tensor var_350_end_mask_0 = const()[name = tensor("op_350_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_350_squeeze_mask_0 = const()[name = tensor("op_350_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_350 = slice_by_index(begin = var_350_begin_0, end = var_350_end_0, end_mask = var_350_end_mask_0, squeeze_mask = var_350_squeeze_mask_0, x = k_9)[name = tensor("op_350")]; + tensor var_352 = mul(x = freqs, y = ts)[name = tensor("op_352")]; + tensor rotr = cos(x = var_352)[name = tensor("rotr")]; + tensor roti = sin(x = var_352)[name = tensor("roti")]; + tensor var_356 = mul(x = var_344, y = rotr)[name = tensor("op_356")]; + tensor var_357 = mul(x = var_346, y = roti)[name = tensor("op_357")]; + tensor qor_5 = sub(x = var_356, y = var_357)[name = tensor("qor_5")]; + tensor var_359 = mul(x = var_344, y = roti)[name = tensor("op_359")]; + tensor var_360 = mul(x = var_346, y = rotr)[name = tensor("op_360")]; + tensor qoi_5 = add(x = var_359, y = var_360)[name = tensor("qoi_5")]; + tensor var_362 = mul(x = var_348, y = rotr)[name = tensor("op_362")]; + tensor var_363 = mul(x = var_350, y = roti)[name = tensor("op_363")]; + tensor kor_5 = sub(x = var_362, y = var_363)[name = tensor("kor_5")]; + tensor var_365 = mul(x = var_348, y = roti)[name = tensor("op_365")]; + tensor var_366 = mul(x = var_350, y = rotr)[name = tensor("op_366")]; + tensor koi_5 = add(x = var_365, y = var_366)[name = tensor("koi_5")]; + tensor qo_axis_0 = const()[name = tensor("qo_axis_0"), val = tensor(-1)]; + tensor qo = stack(axis = qo_axis_0, values = (qor_5, qoi_5))[name = tensor("qo")]; + tensor ko_axis_0 = const()[name = tensor("ko_axis_0"), val = tensor(-1)]; + tensor ko = stack(axis = ko_axis_0, values = (kor_5, koi_5))[name = tensor("ko")]; + tensor var_376 = const()[name = tensor("op_376"), val = tensor([1, 16, 8, 64])]; + tensor q = reshape(shape = var_376, x = qo)[name = tensor("q")]; + tensor var_378 = const()[name = tensor("op_378"), val = tensor([1, 16, 8, 64])]; + tensor k = reshape(shape = var_378, x = ko)[name = tensor("k")]; + tensor capacity = const()[name = tensor("capacity"), val = tensor([256])]; + tensor var_383_dtype_0 = const()[name = tensor("op_383_dtype_0"), val = tensor("int32")]; + tensor var_384 = const()[name = tensor("op_384"), val = tensor([1, 1])]; + tensor var_383 = cast(dtype = var_383_dtype_0, x = attn1_offset)[name = tensor("cast_46")]; + tensor write_base = reshape(shape = var_384, x = var_383)[name = tensor("write_base")]; + tensor write_range = const()[name = tensor("write_range"), val = tensor([[0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15]])]; + tensor abs_idx = add(x = write_base, y = write_range)[name = tensor("abs_idx")]; + tensor wrapped_div = floor_div(x = abs_idx, y = capacity)[name = tensor("wrapped_div")]; + tensor wrapped_div_scaled = mul(x = wrapped_div, y = capacity)[name = tensor("wrapped_div_scaled")]; + tensor wrapped = sub(x = abs_idx, y = wrapped_div_scaled)[name = tensor("wrapped")]; + tensor var_391 = const()[name = tensor("op_391"), val = tensor([1, 16, 1, 1])]; + tensor var_392 = reshape(shape = var_391, x = wrapped)[name = tensor("op_392")]; + tensor write_indexes_reps_0 = const()[name = tensor("write_indexes_reps_0"), val = tensor([1, 1, 8, 64])]; + tensor write_indexes = tile(reps = write_indexes_reps_0, x = var_392)[name = tensor("write_indexes")]; + tensor var_395_begin_0 = const()[name = tensor("op_395_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_395_end_0 = const()[name = tensor("op_395_end_0"), val = tensor([1, 1, 256, 8, 64])]; + tensor var_395_end_mask_0 = const()[name = tensor("op_395_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_395_squeeze_mask_0 = const()[name = tensor("op_395_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor var_395 = slice_by_index(begin = var_395_begin_0, end = var_395_end_0, end_mask = var_395_end_mask_0, squeeze_mask = var_395_squeeze_mask_0, x = attn1_cache)[name = tensor("op_395")]; + tensor new_k_cache_5_axis_0 = const()[name = tensor("new_k_cache_5_axis_0"), val = tensor(1)]; + tensor new_k_cache_5_mode_0 = const()[name = tensor("new_k_cache_5_mode_0"), val = tensor("update")]; + tensor new_k_cache_5_validate_indices_0 = const()[name = tensor("new_k_cache_5_validate_indices_0"), val = tensor(false)]; + tensor new_k_cache_5 = scatter_along_axis(axis = new_k_cache_5_axis_0, data = var_395, indices = write_indexes, mode = new_k_cache_5_mode_0, updates = k, validate_indices = new_k_cache_5_validate_indices_0)[name = tensor("new_k_cache_5")]; + tensor var_397_begin_0 = const()[name = tensor("op_397_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor var_397_end_0 = const()[name = tensor("op_397_end_0"), val = tensor([2, 1, 256, 8, 64])]; + tensor var_397_end_mask_0 = const()[name = tensor("op_397_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_397_squeeze_mask_0 = const()[name = tensor("op_397_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor var_397 = slice_by_index(begin = var_397_begin_0, end = var_397_end_0, end_mask = var_397_end_mask_0, squeeze_mask = var_397_squeeze_mask_0, x = attn1_cache)[name = tensor("op_397")]; + tensor new_v_cache_5_axis_0 = const()[name = tensor("new_v_cache_5_axis_0"), val = tensor(1)]; + tensor new_v_cache_5_mode_0 = const()[name = tensor("new_v_cache_5_mode_0"), val = tensor("update")]; + tensor new_v_cache_5_validate_indices_0 = const()[name = tensor("new_v_cache_5_validate_indices_0"), val = tensor(false)]; + tensor new_v_cache_5 = scatter_along_axis(axis = new_v_cache_5_axis_0, data = var_397, indices = write_indexes, mode = new_v_cache_5_mode_0, updates = squeeze_5, validate_indices = new_v_cache_5_validate_indices_0)[name = tensor("new_v_cache_5")]; + tensor var_400_axis_0 = const()[name = tensor("op_400_axis_0"), val = tensor(0)]; + tensor var_400 = stack(axis = var_400_axis_0, values = (new_k_cache_5, new_v_cache_5))[name = tensor("op_400")]; + tensor var_401 = not_equal(x = new_k_cache_5, y = new_k_cache_5)[name = tensor("op_401")]; + tensor new_k_cache = select(a = var_212, b = new_k_cache_5, cond = var_401)[name = tensor("new_k_cache")]; + tensor var_404 = not_equal(x = new_v_cache_5, y = new_v_cache_5)[name = tensor("op_404")]; + tensor new_v_cache = select(a = var_212, b = new_v_cache_5, cond = var_404)[name = tensor("new_v_cache")]; + tensor var_409 = const()[name = tensor("op_409"), val = tensor([0, 2, 1, 3])]; + tensor var_411 = const()[name = tensor("op_411"), val = tensor([1, 1])]; + tensor var_412 = reshape(shape = var_411, x = attn1_offset)[name = tensor("op_412")]; + tensor var_414_promoted = const()[name = tensor("op_414_promoted"), val = tensor([0x1.ep+3])]; + tensor var_415 = add(x = var_412, y = var_414_promoted)[name = tensor("op_415")]; + tensor last_pos_dtype_0 = const()[name = tensor("last_pos_dtype_0"), val = tensor("int32")]; + tensor last_pos = cast(dtype = last_pos_dtype_0, x = var_415)[name = tensor("cast_45")]; + tensor diff = sub(x = last_pos, y = slot_idx_1)[name = tensor("diff")]; + tensor var_421_div = floor_div(x = diff, y = capacity)[name = tensor("op_421_div")]; + tensor var_421_div_scaled = mul(x = var_421_div, y = capacity)[name = tensor("op_421_div_scaled")]; + tensor var_421 = sub(x = diff, y = var_421_div_scaled)[name = tensor("op_421")]; + tensor pos_k = sub(x = last_pos, y = var_421)[name = tensor("pos_k")]; + tensor var_427_promoted = const()[name = tensor("op_427_promoted"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(41766144)))]; + tensor pos_q = add(x = var_412, y = var_427_promoted)[name = tensor("pos_q")]; + tensor var_431_axes_0 = const()[name = tensor("op_431_axes_0"), val = tensor([2])]; + tensor var_431 = expand_dims(axes = var_431_axes_0, x = pos_q)[name = tensor("op_431")]; + tensor var_433_axes_0 = const()[name = tensor("op_433_axes_0"), val = tensor([1])]; + tensor var_433 = expand_dims(axes = var_433_axes_0, x = pos_k)[name = tensor("op_433")]; + tensor var_434_promoted_dtype_0 = const()[name = tensor("op_434_promoted_dtype_0"), val = tensor("fp32")]; + tensor var_434_promoted = cast(dtype = var_434_promoted_dtype_0, x = var_433)[name = tensor("cast_44")]; + tensor delta = sub(x = var_431, y = var_434_promoted)[name = tensor("delta")]; + tensor valid_5 = greater_equal(x = var_433, y = var_86)[name = tensor("valid_5")]; + tensor var_443 = const()[name = tensor("op_443"), val = tensor([1, 1, 1])]; + tensor var_444 = reshape(shape = var_443, x = attn1_offset)[name = tensor("op_444")]; + tensor var_446_promoted = const()[name = tensor("op_446_promoted"), val = tensor([0x1.ep+3])]; + tensor var_447 = add(x = var_444, y = var_446_promoted)[name = tensor("op_447")]; + tensor var_448 = less_equal(x = var_434_promoted, y = var_447)[name = tensor("op_448")]; + tensor valid = logical_and(x = valid_5, y = var_448)[name = tensor("valid")]; + tensor var_86_promoted_1 = const()[name = tensor("op_86_promoted_1"), val = tensor(0x0p+0)]; + tensor var_450 = greater_equal(x = delta, y = var_86_promoted_1)[name = tensor("op_450")]; + tensor attn_mask_7 = logical_and(x = valid, y = var_450)[name = tensor("attn_mask_7")]; + tensor var_98_promoted_1 = const()[name = tensor("op_98_promoted_1"), val = tensor(0x1.f4p+7)]; + tensor var_452 = less(x = delta, y = var_98_promoted_1)[name = tensor("op_452")]; + tensor attn_mask_9 = logical_and(x = attn_mask_7, y = var_452)[name = tensor("attn_mask_9")]; + tensor attn_mask_axes_0 = const()[name = tensor("attn_mask_axes_0"), val = tensor([1])]; + tensor attn_mask = expand_dims(axes = attn_mask_axes_0, x = attn_mask_9)[name = tensor("attn_mask")]; + tensor var_457_transpose_x_0 = const()[name = tensor("op_457_transpose_x_0"), val = tensor(false)]; + tensor var_457_transpose_y_0 = const()[name = tensor("op_457_transpose_y_0"), val = tensor(false)]; + tensor transpose_8_perm_0 = const()[name = tensor("transpose_8_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_9_perm_0 = const()[name = tensor("transpose_9_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_9 = transpose(perm = transpose_9_perm_0, x = new_k_cache)[name = tensor("transpose_12")]; + tensor transpose_8 = transpose(perm = transpose_8_perm_0, x = q)[name = tensor("transpose_13")]; + tensor var_457 = matmul(transpose_x = var_457_transpose_x_0, transpose_y = var_457_transpose_y_0, x = transpose_8, y = transpose_9)[name = tensor("op_457")]; + tensor var_458 = const()[name = tensor("op_458"), val = tensor(0x1p-3)]; + tensor attn_7 = mul(x = var_457, y = var_458)[name = tensor("attn_7")]; + tensor var_460 = logical_not(x = attn_mask)[name = tensor("op_460")]; + tensor attn_9 = select(a = var_100, b = attn_7, cond = var_460)[name = tensor("attn_9")]; + tensor attn = softmax(axis = var_91, x = attn_9)[name = tensor("attn")]; + tensor x_11_transpose_x_0 = const()[name = tensor("x_11_transpose_x_0"), val = tensor(false)]; + tensor x_11_transpose_y_0 = const()[name = tensor("x_11_transpose_y_0"), val = tensor(false)]; + tensor v_attn = transpose(perm = var_409, x = new_v_cache)[name = tensor("transpose_14")]; + tensor x_11 = matmul(transpose_x = x_11_transpose_x_0, transpose_y = x_11_transpose_y_0, x = attn, y = v_attn)[name = tensor("x_11")]; + tensor var_464_perm_0 = const()[name = tensor("op_464_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_465 = const()[name = tensor("op_465"), val = tensor([1, 16, 512])]; + tensor var_464 = transpose(perm = var_464_perm_0, x = x_11)[name = tensor("transpose_11")]; + tensor input_15 = reshape(shape = var_465, x = var_464)[name = tensor("input_15")]; + tensor x_13 = linear(bias = linear_1_bias_0, weight = mimi_decoder_transformer_transformer_layers_1_self_attn_out_proj_weight, x = input_15)[name = tensor("linear_5")]; + tensor var_474 = mul(x = mimi_decoder_transformer_transformer_layers_1_layer_scale_1_scale, y = x_13)[name = tensor("op_474")]; + tensor input_17 = add(x = input_13, y = var_474)[name = tensor("input_17")]; + tensor input_19_axes_0 = const()[name = tensor("input_19_axes_0"), val = tensor([-1])]; + tensor input_19 = layer_norm(axes = input_19_axes_0, beta = mimi_decoder_transformer_transformer_layers_1_norm2_bias, epsilon = var_102, gamma = mimi_decoder_transformer_transformer_layers_1_norm2_weight, x = input_17)[name = tensor("input_19")]; + tensor var_481 = linear(bias = linear_2_bias_0, weight = mimi_decoder_transformer_transformer_layers_1_linear1_weight, x = input_19)[name = tensor("linear_6")]; + tensor input_21_mode_0 = const()[name = tensor("input_21_mode_0"), val = tensor("EXACT")]; + tensor input_21 = gelu(mode = input_21_mode_0, x = var_481)[name = tensor("input_21")]; + tensor x_15 = linear(bias = linear_1_bias_0, weight = mimi_decoder_transformer_transformer_layers_1_linear2_weight, x = input_21)[name = tensor("linear_7")]; + tensor var_487 = mul(x = mimi_decoder_transformer_transformer_layers_1_layer_scale_2_scale, y = x_15)[name = tensor("op_487")]; + tensor z = add(x = input_17, y = var_487)[name = tensor("z")]; + tensor x_17_perm_0 = const()[name = tensor("x_17_perm_0"), val = tensor([0, 2, 1])]; + tensor var_507 = const()[name = tensor("op_507"), val = tensor(0x1p+0)]; + tensor var_508 = const()[name = tensor("op_508"), val = tensor(-1)]; + tensor input_23_interleave_0 = const()[name = tensor("input_23_interleave_0"), val = tensor(false)]; + tensor x_17 = transpose(perm = x_17_perm_0, x = z)[name = tensor("transpose_10")]; + tensor input_23 = concat(axis = var_508, interleave = input_23_interleave_0, values = (conv0_prev, x_17))[name = tensor("input_23")]; + tensor input_25_pad_type_0 = const()[name = tensor("input_25_pad_type_0"), val = tensor("valid")]; + tensor input_25_strides_0 = const()[name = tensor("input_25_strides_0"), val = tensor([1])]; + tensor input_25_pad_0 = const()[name = tensor("input_25_pad_0"), val = tensor([0, 0])]; + tensor input_25_dilations_0 = const()[name = tensor("input_25_dilations_0"), val = tensor([1])]; + tensor input_25_groups_0 = const()[name = tensor("input_25_groups_0"), val = tensor(1)]; + tensor input_25 = conv(bias = mimi_decoder_model_0_conv_bias, dilations = input_25_dilations_0, groups = input_25_groups_0, pad = input_25_pad_0, pad_type = input_25_pad_type_0, strides = input_25_strides_0, weight = mimi_decoder_model_0_conv_weight, x = input_23)[name = tensor("input_25")]; + tensor var_542_begin_0 = const()[name = tensor("op_542_begin_0"), val = tensor([0, 0, 16])]; + tensor var_542_end_0 = const()[name = tensor("op_542_end_0"), val = tensor([1, 512, 22])]; + tensor var_542_end_mask_0 = const()[name = tensor("op_542_end_mask_0"), val = tensor([true, true, true])]; + tensor var_542 = slice_by_index(begin = var_542_begin_0, end = var_542_end_0, end_mask = var_542_end_mask_0, x = input_23)[name = tensor("op_542")]; + tensor input_27 = elu(alpha = var_507, x = input_25)[name = tensor("input_27")]; + tensor y_5_pad_type_0 = const()[name = tensor("y_5_pad_type_0"), val = tensor("valid")]; + tensor y_5_strides_0 = const()[name = tensor("y_5_strides_0"), val = tensor([6])]; + tensor y_5_pad_0 = const()[name = tensor("y_5_pad_0"), val = tensor([0, 0])]; + tensor y_5_dilations_0 = const()[name = tensor("y_5_dilations_0"), val = tensor([1])]; + tensor y_5_groups_0 = const()[name = tensor("y_5_groups_0"), val = tensor(1)]; + tensor y_5_has_output_shape_output_shape_0 = const()[name = tensor("y_5_has_output_shape_output_shape_0"), val = tensor([1, 256, 102])]; + tensor y_5_has_output_shape = conv_transpose(bias = mimi_decoder_model_2_convtr_bias, dilations = y_5_dilations_0, groups = y_5_groups_0, output_shape = y_5_has_output_shape_output_shape_0, pad = y_5_pad_0, pad_type = y_5_pad_type_0, strides = y_5_strides_0, weight = mimi_decoder_model_2_convtr_weight, x = input_27)[name = tensor("y_5_has_output_shape")]; + tensor var_557_begin_0 = const()[name = tensor("op_557_begin_0"), val = tensor([0, 0, 0])]; + tensor var_557_end_0 = const()[name = tensor("op_557_end_0"), val = tensor([1, 256, 6])]; + tensor var_557_end_mask_0 = const()[name = tensor("op_557_end_mask_0"), val = tensor([true, true, false])]; + tensor var_557 = slice_by_index(begin = var_557_begin_0, end = var_557_end_0, end_mask = var_557_end_mask_0, x = y_5_has_output_shape)[name = tensor("op_557")]; + tensor var_558 = add(x = var_557, y = convtr0_partial)[name = tensor("op_558")]; + tensor var_559_begin_0 = const()[name = tensor("op_559_begin_0"), val = tensor([0, 0, 6])]; + tensor var_559_end_0 = const()[name = tensor("op_559_end_0"), val = tensor([1, 256, 102])]; + tensor var_559_end_mask_0 = const()[name = tensor("op_559_end_mask_0"), val = tensor([true, true, true])]; + tensor var_559 = slice_by_index(begin = var_559_begin_0, end = var_559_end_0, end_mask = var_559_end_mask_0, x = y_5_has_output_shape)[name = tensor("op_559")]; + tensor y_7_interleave_0 = const()[name = tensor("y_7_interleave_0"), val = tensor(false)]; + tensor y_7 = concat(axis = var_508, interleave = y_7_interleave_0, values = (var_558, var_559))[name = tensor("y_7")]; + tensor new_partial_1_begin_0 = const()[name = tensor("new_partial_1_begin_0"), val = tensor([0, 0, 96])]; + tensor new_partial_1_end_0 = const()[name = tensor("new_partial_1_end_0"), val = tensor([1, 256, 102])]; + tensor new_partial_1_end_mask_0 = const()[name = tensor("new_partial_1_end_mask_0"), val = tensor([true, true, true])]; + tensor new_partial_1 = slice_by_index(begin = new_partial_1_begin_0, end = new_partial_1_end_0, end_mask = new_partial_1_end_mask_0, x = y_7)[name = tensor("new_partial_1")]; + tensor var_564 = const()[name = tensor("op_564"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(41766272)))]; + tensor var_565 = sub(x = new_partial_1, y = var_564)[name = tensor("op_565")]; + tensor input_29_begin_0 = const()[name = tensor("input_29_begin_0"), val = tensor([0, 0, 0])]; + tensor input_29_end_0 = const()[name = tensor("input_29_end_0"), val = tensor([1, 256, 96])]; + tensor input_29_end_mask_0 = const()[name = tensor("input_29_end_mask_0"), val = tensor([true, true, false])]; + tensor input_29 = slice_by_index(begin = input_29_begin_0, end = input_29_end_0, end_mask = input_29_end_mask_0, x = y_7)[name = tensor("input_29")]; + tensor x_19 = elu(alpha = var_507, x = input_29)[name = tensor("x_19")]; + tensor input_31_interleave_0 = const()[name = tensor("input_31_interleave_0"), val = tensor(false)]; + tensor input_31 = concat(axis = var_508, interleave = input_31_interleave_0, values = (res0_conv0_prev, x_19))[name = tensor("input_31")]; + tensor input_33_pad_type_0 = const()[name = tensor("input_33_pad_type_0"), val = tensor("valid")]; + tensor input_33_strides_0 = const()[name = tensor("input_33_strides_0"), val = tensor([1])]; + tensor input_33_pad_0 = const()[name = tensor("input_33_pad_0"), val = tensor([0, 0])]; + tensor input_33_dilations_0 = const()[name = tensor("input_33_dilations_0"), val = tensor([1])]; + tensor input_33_groups_0 = const()[name = tensor("input_33_groups_0"), val = tensor(1)]; + tensor input_33 = conv(bias = mimi_decoder_model_3_block_1_conv_bias, dilations = input_33_dilations_0, groups = input_33_groups_0, pad = input_33_pad_0, pad_type = input_33_pad_type_0, strides = input_33_strides_0, weight = mimi_decoder_model_3_block_1_conv_weight, x = input_31)[name = tensor("input_33")]; + tensor var_585_begin_0 = const()[name = tensor("op_585_begin_0"), val = tensor([0, 0, 96])]; + tensor var_585_end_0 = const()[name = tensor("op_585_end_0"), val = tensor([1, 256, 98])]; + tensor var_585_end_mask_0 = const()[name = tensor("op_585_end_mask_0"), val = tensor([true, true, true])]; + tensor var_585 = slice_by_index(begin = var_585_begin_0, end = var_585_end_0, end_mask = var_585_end_mask_0, x = input_31)[name = tensor("op_585")]; + tensor x_21 = elu(alpha = var_507, x = input_33)[name = tensor("x_21")]; + tensor v_5_pad_type_0 = const()[name = tensor("v_5_pad_type_0"), val = tensor("valid")]; + tensor v_5_strides_0 = const()[name = tensor("v_5_strides_0"), val = tensor([1])]; + tensor v_5_pad_0 = const()[name = tensor("v_5_pad_0"), val = tensor([0, 0])]; + tensor v_5_dilations_0 = const()[name = tensor("v_5_dilations_0"), val = tensor([1])]; + tensor v_5_groups_0 = const()[name = tensor("v_5_groups_0"), val = tensor(1)]; + tensor v_5 = conv(bias = mimi_decoder_model_3_block_3_conv_bias, dilations = v_5_dilations_0, groups = v_5_groups_0, pad = v_5_pad_0, pad_type = v_5_pad_type_0, strides = v_5_strides_0, weight = mimi_decoder_model_3_block_3_conv_weight, x = x_21)[name = tensor("v_5")]; + tensor input_35 = add(x = input_29, y = v_5)[name = tensor("input_35")]; + tensor input_37 = elu(alpha = var_507, x = input_35)[name = tensor("input_37")]; + tensor y_9_pad_type_0 = const()[name = tensor("y_9_pad_type_0"), val = tensor("valid")]; + tensor y_9_strides_0 = const()[name = tensor("y_9_strides_0"), val = tensor([5])]; + tensor y_9_pad_0 = const()[name = tensor("y_9_pad_0"), val = tensor([0, 0])]; + tensor y_9_dilations_0 = const()[name = tensor("y_9_dilations_0"), val = tensor([1])]; + tensor y_9_groups_0 = const()[name = tensor("y_9_groups_0"), val = tensor(1)]; + tensor y_9_has_output_shape_output_shape_0 = const()[name = tensor("y_9_has_output_shape_output_shape_0"), val = tensor([1, 128, 485])]; + tensor y_9_has_output_shape = conv_transpose(bias = mimi_decoder_model_5_convtr_bias, dilations = y_9_dilations_0, groups = y_9_groups_0, output_shape = y_9_has_output_shape_output_shape_0, pad = y_9_pad_0, pad_type = y_9_pad_type_0, strides = y_9_strides_0, weight = mimi_decoder_model_5_convtr_weight, x = input_37)[name = tensor("y_9_has_output_shape")]; + tensor var_613_begin_0 = const()[name = tensor("op_613_begin_0"), val = tensor([0, 0, 0])]; + tensor var_613_end_0 = const()[name = tensor("op_613_end_0"), val = tensor([1, 128, 5])]; + tensor var_613_end_mask_0 = const()[name = tensor("op_613_end_mask_0"), val = tensor([true, true, false])]; + tensor var_613 = slice_by_index(begin = var_613_begin_0, end = var_613_end_0, end_mask = var_613_end_mask_0, x = y_9_has_output_shape)[name = tensor("op_613")]; + tensor var_614 = add(x = var_613, y = convtr1_partial)[name = tensor("op_614")]; + tensor var_615_begin_0 = const()[name = tensor("op_615_begin_0"), val = tensor([0, 0, 5])]; + tensor var_615_end_0 = const()[name = tensor("op_615_end_0"), val = tensor([1, 128, 485])]; + tensor var_615_end_mask_0 = const()[name = tensor("op_615_end_mask_0"), val = tensor([true, true, true])]; + tensor var_615 = slice_by_index(begin = var_615_begin_0, end = var_615_end_0, end_mask = var_615_end_mask_0, x = y_9_has_output_shape)[name = tensor("op_615")]; + tensor y_11_interleave_0 = const()[name = tensor("y_11_interleave_0"), val = tensor(false)]; + tensor y_11 = concat(axis = var_508, interleave = y_11_interleave_0, values = (var_614, var_615))[name = tensor("y_11")]; + tensor new_partial_3_begin_0 = const()[name = tensor("new_partial_3_begin_0"), val = tensor([0, 0, 480])]; + tensor new_partial_3_end_0 = const()[name = tensor("new_partial_3_end_0"), val = tensor([1, 128, 485])]; + tensor new_partial_3_end_mask_0 = const()[name = tensor("new_partial_3_end_mask_0"), val = tensor([true, true, true])]; + tensor new_partial_3 = slice_by_index(begin = new_partial_3_begin_0, end = new_partial_3_end_0, end_mask = new_partial_3_end_mask_0, x = y_11)[name = tensor("new_partial_3")]; + tensor var_620 = const()[name = tensor("op_620"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(41767360)))]; + tensor var_621 = sub(x = new_partial_3, y = var_620)[name = tensor("op_621")]; + tensor input_39_begin_0 = const()[name = tensor("input_39_begin_0"), val = tensor([0, 0, 0])]; + tensor input_39_end_0 = const()[name = tensor("input_39_end_0"), val = tensor([1, 128, 480])]; + tensor input_39_end_mask_0 = const()[name = tensor("input_39_end_mask_0"), val = tensor([true, true, false])]; + tensor input_39 = slice_by_index(begin = input_39_begin_0, end = input_39_end_0, end_mask = input_39_end_mask_0, x = y_11)[name = tensor("input_39")]; + tensor x_23 = elu(alpha = var_507, x = input_39)[name = tensor("x_23")]; + tensor input_41_interleave_0 = const()[name = tensor("input_41_interleave_0"), val = tensor(false)]; + tensor input_41 = concat(axis = var_508, interleave = input_41_interleave_0, values = (res1_conv0_prev, x_23))[name = tensor("input_41")]; + tensor input_43_pad_type_0 = const()[name = tensor("input_43_pad_type_0"), val = tensor("valid")]; + tensor input_43_strides_0 = const()[name = tensor("input_43_strides_0"), val = tensor([1])]; + tensor input_43_pad_0 = const()[name = tensor("input_43_pad_0"), val = tensor([0, 0])]; + tensor input_43_dilations_0 = const()[name = tensor("input_43_dilations_0"), val = tensor([1])]; + tensor input_43_groups_0 = const()[name = tensor("input_43_groups_0"), val = tensor(1)]; + tensor input_43 = conv(bias = mimi_decoder_model_6_block_1_conv_bias, dilations = input_43_dilations_0, groups = input_43_groups_0, pad = input_43_pad_0, pad_type = input_43_pad_type_0, strides = input_43_strides_0, weight = mimi_decoder_model_6_block_1_conv_weight, x = input_41)[name = tensor("input_43")]; + tensor var_641_begin_0 = const()[name = tensor("op_641_begin_0"), val = tensor([0, 0, 480])]; + tensor var_641_end_0 = const()[name = tensor("op_641_end_0"), val = tensor([1, 128, 482])]; + tensor var_641_end_mask_0 = const()[name = tensor("op_641_end_mask_0"), val = tensor([true, true, true])]; + tensor var_641 = slice_by_index(begin = var_641_begin_0, end = var_641_end_0, end_mask = var_641_end_mask_0, x = input_41)[name = tensor("op_641")]; + tensor x_25 = elu(alpha = var_507, x = input_43)[name = tensor("x_25")]; + tensor v_7_pad_type_0 = const()[name = tensor("v_7_pad_type_0"), val = tensor("valid")]; + tensor v_7_strides_0 = const()[name = tensor("v_7_strides_0"), val = tensor([1])]; + tensor v_7_pad_0 = const()[name = tensor("v_7_pad_0"), val = tensor([0, 0])]; + tensor v_7_dilations_0 = const()[name = tensor("v_7_dilations_0"), val = tensor([1])]; + tensor v_7_groups_0 = const()[name = tensor("v_7_groups_0"), val = tensor(1)]; + tensor v_7 = conv(bias = mimi_decoder_model_6_block_3_conv_bias, dilations = v_7_dilations_0, groups = v_7_groups_0, pad = v_7_pad_0, pad_type = v_7_pad_type_0, strides = v_7_strides_0, weight = mimi_decoder_model_6_block_3_conv_weight, x = x_25)[name = tensor("v_7")]; + tensor input_45 = add(x = input_39, y = v_7)[name = tensor("input_45")]; + tensor input_47 = elu(alpha = var_507, x = input_45)[name = tensor("input_47")]; + tensor y_13_pad_type_0 = const()[name = tensor("y_13_pad_type_0"), val = tensor("valid")]; + tensor y_13_strides_0 = const()[name = tensor("y_13_strides_0"), val = tensor([4])]; + tensor y_13_pad_0 = const()[name = tensor("y_13_pad_0"), val = tensor([0, 0])]; + tensor y_13_dilations_0 = const()[name = tensor("y_13_dilations_0"), val = tensor([1])]; + tensor y_13_groups_0 = const()[name = tensor("y_13_groups_0"), val = tensor(1)]; + tensor y_13_has_output_shape_output_shape_0 = const()[name = tensor("y_13_has_output_shape_output_shape_0"), val = tensor([1, 64, 1924])]; + tensor y_13_has_output_shape = conv_transpose(bias = mimi_decoder_model_8_convtr_bias, dilations = y_13_dilations_0, groups = y_13_groups_0, output_shape = y_13_has_output_shape_output_shape_0, pad = y_13_pad_0, pad_type = y_13_pad_type_0, strides = y_13_strides_0, weight = mimi_decoder_model_8_convtr_weight, x = input_47)[name = tensor("y_13_has_output_shape")]; + tensor var_669_begin_0 = const()[name = tensor("op_669_begin_0"), val = tensor([0, 0, 0])]; + tensor var_669_end_0 = const()[name = tensor("op_669_end_0"), val = tensor([1, 64, 4])]; + tensor var_669_end_mask_0 = const()[name = tensor("op_669_end_mask_0"), val = tensor([true, true, false])]; + tensor var_669 = slice_by_index(begin = var_669_begin_0, end = var_669_end_0, end_mask = var_669_end_mask_0, x = y_13_has_output_shape)[name = tensor("op_669")]; + tensor var_670 = add(x = var_669, y = convtr2_partial)[name = tensor("op_670")]; + tensor var_671_begin_0 = const()[name = tensor("op_671_begin_0"), val = tensor([0, 0, 4])]; + tensor var_671_end_0 = const()[name = tensor("op_671_end_0"), val = tensor([1, 64, 1924])]; + tensor var_671_end_mask_0 = const()[name = tensor("op_671_end_mask_0"), val = tensor([true, true, true])]; + tensor var_671 = slice_by_index(begin = var_671_begin_0, end = var_671_end_0, end_mask = var_671_end_mask_0, x = y_13_has_output_shape)[name = tensor("op_671")]; + tensor y_interleave_0 = const()[name = tensor("y_interleave_0"), val = tensor(false)]; + tensor y = concat(axis = var_508, interleave = y_interleave_0, values = (var_670, var_671))[name = tensor("y")]; + tensor new_partial_begin_0 = const()[name = tensor("new_partial_begin_0"), val = tensor([0, 0, 1920])]; + tensor new_partial_end_0 = const()[name = tensor("new_partial_end_0"), val = tensor([1, 64, 1924])]; + tensor new_partial_end_mask_0 = const()[name = tensor("new_partial_end_mask_0"), val = tensor([true, true, true])]; + tensor new_partial = slice_by_index(begin = new_partial_begin_0, end = new_partial_end_0, end_mask = new_partial_end_mask_0, x = y)[name = tensor("new_partial")]; + tensor var_676 = const()[name = tensor("op_676"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(41767936)))]; + tensor var_677 = sub(x = new_partial, y = var_676)[name = tensor("op_677")]; + tensor input_49_begin_0 = const()[name = tensor("input_49_begin_0"), val = tensor([0, 0, 0])]; + tensor input_49_end_0 = const()[name = tensor("input_49_end_0"), val = tensor([1, 64, 1920])]; + tensor input_49_end_mask_0 = const()[name = tensor("input_49_end_mask_0"), val = tensor([true, true, false])]; + tensor input_49 = slice_by_index(begin = input_49_begin_0, end = input_49_end_0, end_mask = input_49_end_mask_0, x = y)[name = tensor("input_49")]; + tensor x_27 = elu(alpha = var_507, x = input_49)[name = tensor("x_27")]; + tensor input_51_interleave_0 = const()[name = tensor("input_51_interleave_0"), val = tensor(false)]; + tensor input_51 = concat(axis = var_508, interleave = input_51_interleave_0, values = (res2_conv0_prev, x_27))[name = tensor("input_51")]; + tensor input_53_pad_type_0 = const()[name = tensor("input_53_pad_type_0"), val = tensor("valid")]; + tensor input_53_strides_0 = const()[name = tensor("input_53_strides_0"), val = tensor([1])]; + tensor input_53_pad_0 = const()[name = tensor("input_53_pad_0"), val = tensor([0, 0])]; + tensor input_53_dilations_0 = const()[name = tensor("input_53_dilations_0"), val = tensor([1])]; + tensor input_53_groups_0 = const()[name = tensor("input_53_groups_0"), val = tensor(1)]; + tensor input_53 = conv(bias = mimi_decoder_model_9_block_1_conv_bias, dilations = input_53_dilations_0, groups = input_53_groups_0, pad = input_53_pad_0, pad_type = input_53_pad_type_0, strides = input_53_strides_0, weight = mimi_decoder_model_9_block_1_conv_weight, x = input_51)[name = tensor("input_53")]; + tensor var_697_begin_0 = const()[name = tensor("op_697_begin_0"), val = tensor([0, 0, 1920])]; + tensor var_697_end_0 = const()[name = tensor("op_697_end_0"), val = tensor([1, 64, 1922])]; + tensor var_697_end_mask_0 = const()[name = tensor("op_697_end_mask_0"), val = tensor([true, true, true])]; + tensor var_697 = slice_by_index(begin = var_697_begin_0, end = var_697_end_0, end_mask = var_697_end_mask_0, x = input_51)[name = tensor("op_697")]; + tensor x_29 = elu(alpha = var_507, x = input_53)[name = tensor("x_29")]; + tensor v_pad_type_0 = const()[name = tensor("v_pad_type_0"), val = tensor("valid")]; + tensor v_strides_0 = const()[name = tensor("v_strides_0"), val = tensor([1])]; + tensor v_pad_0 = const()[name = tensor("v_pad_0"), val = tensor([0, 0])]; + tensor v_dilations_0 = const()[name = tensor("v_dilations_0"), val = tensor([1])]; + tensor v_groups_0 = const()[name = tensor("v_groups_0"), val = tensor(1)]; + tensor v = conv(bias = mimi_decoder_model_9_block_3_conv_bias, dilations = v_dilations_0, groups = v_groups_0, pad = v_pad_0, pad_type = v_pad_type_0, strides = v_strides_0, weight = mimi_decoder_model_9_block_3_conv_weight, x = x_29)[name = tensor("v")]; + tensor input_55 = add(x = input_49, y = v)[name = tensor("input_55")]; + tensor x = elu(alpha = var_507, x = input_55)[name = tensor("x")]; + tensor input_interleave_0 = const()[name = tensor("input_interleave_0"), val = tensor(false)]; + tensor input = concat(axis = var_508, interleave = input_interleave_0, values = (conv_final_prev, x))[name = tensor("input")]; + tensor var_724_pad_type_0 = const()[name = tensor("op_724_pad_type_0"), val = tensor("valid")]; + tensor var_724_strides_0 = const()[name = tensor("op_724_strides_0"), val = tensor([1])]; + tensor var_724_pad_0 = const()[name = tensor("op_724_pad_0"), val = tensor([0, 0])]; + tensor var_724_dilations_0 = const()[name = tensor("op_724_dilations_0"), val = tensor([1])]; + tensor var_724_groups_0 = const()[name = tensor("op_724_groups_0"), val = tensor(1)]; + tensor var_724 = conv(bias = mimi_decoder_model_11_conv_bias, dilations = var_724_dilations_0, groups = var_724_groups_0, pad = var_724_pad_0, pad_type = var_724_pad_type_0, strides = var_724_strides_0, weight = mimi_decoder_model_11_conv_weight, x = input)[name = tensor("op_724")]; + tensor var_725_begin_0 = const()[name = tensor("op_725_begin_0"), val = tensor([0, 0, 1920])]; + tensor var_725_end_0 = const()[name = tensor("op_725_end_0"), val = tensor([1, 64, 1922])]; + tensor var_725_end_mask_0 = const()[name = tensor("op_725_end_mask_0"), val = tensor([true, true, true])]; + tensor var_725 = slice_by_index(begin = var_725_begin_0, end = var_725_end_0, end_mask = var_725_end_mask_0, x = input)[name = tensor("op_725")]; + tensor var_740_promoted = const()[name = tensor("op_740_promoted"), val = tensor(0x1p+4)]; + tensor var_741 = add(x = attn0_offset, y = var_740_promoted)[name = tensor("op_741")]; + tensor var_743_promoted = const()[name = tensor("op_743_promoted"), val = tensor(0x1p+4)]; + tensor var_744 = add(x = attn1_offset, y = var_743_promoted)[name = tensor("op_744")]; + tensor conv0_first_tmp = identity(x = conv0_first)[name = tensor("conv0_first_tmp")]; + tensor res0_conv0_first_tmp = identity(x = res0_conv0_first)[name = tensor("res0_conv0_first_tmp")]; + tensor res0_conv1_prev_tmp = identity(x = res0_conv1_prev)[name = tensor("res0_conv1_prev_tmp")]; + tensor res0_conv1_first_tmp = identity(x = res0_conv1_first)[name = tensor("res0_conv1_first_tmp")]; + tensor res1_conv0_first_tmp = identity(x = res1_conv0_first)[name = tensor("res1_conv0_first_tmp")]; + tensor res1_conv1_prev_tmp = identity(x = res1_conv1_prev)[name = tensor("res1_conv1_prev_tmp")]; + tensor res1_conv1_first_tmp = identity(x = res1_conv1_first)[name = tensor("res1_conv1_first_tmp")]; + tensor res2_conv0_first_tmp = identity(x = res2_conv0_first)[name = tensor("res2_conv0_first_tmp")]; + tensor res2_conv1_prev_tmp = identity(x = res2_conv1_prev)[name = tensor("res2_conv1_prev_tmp")]; + tensor res2_conv1_first_tmp = identity(x = res2_conv1_first)[name = tensor("res2_conv1_first_tmp")]; + tensor conv_final_first_tmp = identity(x = conv_final_first)[name = tensor("conv_final_first_tmp")]; + } -> (var_724, var_77, var_210, var_741, var_400, var_744, var_542, conv0_first, var_565, var_585, res0_conv0_first, res0_conv1_prev, res0_conv1_first, var_621, var_641, res1_conv0_first, res1_conv1_prev, res1_conv1_first, var_677, var_697, res2_conv0_first, res2_conv1_prev, res2_conv1_first, var_725, conv_final_first); +} \ No newline at end of file diff --git a/v2.1/french_24l/mimi_decoder.mlmodelc/weights/weight.bin b/v2.1/french_24l/mimi_decoder.mlmodelc/weights/weight.bin new file mode 100644 index 0000000000000000000000000000000000000000..bf37b9e4f126187f3237bb76752d699c30997f93 --- /dev/null +++ b/v2.1/french_24l/mimi_decoder.mlmodelc/weights/weight.bin @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:06dbd8b6f8fd4ffefd443aa5bb8b1927cb6d71f7bbac1117ada13938964d2b0b +size 41768256 diff --git a/v2.1/french_24l/mimi_decoder.mlpackage/Data/com.apple.CoreML/model.mlmodel b/v2.1/french_24l/mimi_decoder.mlpackage/Data/com.apple.CoreML/model.mlmodel new file mode 100644 index 0000000000000000000000000000000000000000..1ce9c1af0059c79e89e547ac21f66ffb02b089ae --- /dev/null +++ b/v2.1/french_24l/mimi_decoder.mlpackage/Data/com.apple.CoreML/model.mlmodel @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:4f8b565a3c44e437cabde4993a5584e25e709d57106dfa249995214654d86ff2 +size 93270 diff --git a/v2.1/french_24l/mimi_decoder.mlpackage/Data/com.apple.CoreML/weights/weight.bin b/v2.1/french_24l/mimi_decoder.mlpackage/Data/com.apple.CoreML/weights/weight.bin new file mode 100644 index 0000000000000000000000000000000000000000..bf37b9e4f126187f3237bb76752d699c30997f93 --- /dev/null +++ b/v2.1/french_24l/mimi_decoder.mlpackage/Data/com.apple.CoreML/weights/weight.bin @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:06dbd8b6f8fd4ffefd443aa5bb8b1927cb6d71f7bbac1117ada13938964d2b0b +size 41768256 diff --git a/v2.1/french_24l/mimi_decoder.mlpackage/Manifest.json b/v2.1/french_24l/mimi_decoder.mlpackage/Manifest.json new file mode 100644 index 0000000000000000000000000000000000000000..ff424991bb9591338b9a2b90d2253ce912b3003a --- /dev/null +++ b/v2.1/french_24l/mimi_decoder.mlpackage/Manifest.json @@ -0,0 +1,18 @@ +{ + "fileFormatVersion": "1.0.0", + "itemInfoEntries": { + "E5741DA4-2CE8-4E1C-9BF0-D706DFC67953": { + "author": "com.apple.CoreML", + "description": "CoreML Model Weights", + "name": "weights", + "path": "com.apple.CoreML/weights" + }, + "E63F397C-889D-4991-9184-8D402F5AD988": { + "author": "com.apple.CoreML", + "description": "CoreML Model Specification", + "name": "model.mlmodel", + "path": "com.apple.CoreML/model.mlmodel" + } + }, + "rootModelIdentifier": "E63F397C-889D-4991-9184-8D402F5AD988" +} diff --git a/v2.1/german/cond_prefill.mlmodelc/analytics/coremldata.bin b/v2.1/german/cond_prefill.mlmodelc/analytics/coremldata.bin new file mode 100644 index 0000000000000000000000000000000000000000..1bead8f21ebb3366187e056b119c3524b3e5df49 --- /dev/null +++ b/v2.1/german/cond_prefill.mlmodelc/analytics/coremldata.bin @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:4775182e4d9f214ad95408328db188a92eff8313c22dfa751a7a344fa97379e9 +size 243 diff --git a/v2.1/german/cond_prefill.mlmodelc/coremldata.bin b/v2.1/german/cond_prefill.mlmodelc/coremldata.bin new file mode 100644 index 0000000000000000000000000000000000000000..ed1e169674df036e35958d4708d675837f7551ae --- /dev/null +++ b/v2.1/german/cond_prefill.mlmodelc/coremldata.bin @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:94e7823168aa983079bac808a03e88f6fa1eda74fcc6fdaf0efad40489ba6c83 +size 1151 diff --git a/v2.1/german/cond_prefill.mlmodelc/model.mil b/v2.1/german/cond_prefill.mlmodelc/model.mil new file mode 100644 index 0000000000000000000000000000000000000000..203b091a2c5043d4b8c07f4402d5a5f0319b3c3f --- /dev/null +++ b/v2.1/german/cond_prefill.mlmodelc/model.mil @@ -0,0 +1,1309 @@ +program(1.0) +[buildInfo = dict, tensor>({{"coremlc-component-MIL", "3520.4.1"}, {"coremlc-version", "3520.5.1"}, {"coremltools-component-torch", "2.12.0"}, {"coremltools-source-dialect", "TorchScript"}, {"coremltools-version", "9.0"}})] +{ + func main(tensor cache0, tensor cache1, tensor cache2, tensor cache3, tensor cache4, tensor cache5, tensor conditioning, tensor position0, tensor position1, tensor position2, tensor position3, tensor position4, tensor position5, tensor valid_len) { + tensor input_3_axes_0 = const()[name = tensor("input_3_axes_0"), val = tensor([-1])]; + tensor conditioning_to_fp16_dtype_0 = const()[name = tensor("conditioning_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor norm0_1_weight_to_fp16 = const()[name = tensor("norm0_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(64)))]; + tensor norm0_1_bias_to_fp16 = const()[name = tensor("norm0_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(2176)))]; + tensor var_48_to_fp16 = const()[name = tensor("op_48_to_fp16"), val = tensor(0x1.5p-17)]; + tensor conditioning_to_fp16 = cast(dtype = conditioning_to_fp16_dtype_0, x = conditioning)[name = tensor("cast_159")]; + tensor input_3_cast_fp16 = layer_norm(axes = input_3_axes_0, beta = norm0_1_bias_to_fp16, epsilon = var_48_to_fp16, gamma = norm0_1_weight_to_fp16, x = conditioning_to_fp16)[name = tensor("input_3_cast_fp16")]; + tensor attn0_in_proj_weight_to_fp16 = const()[name = tensor("attn0_in_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(4288)))]; + tensor linear_0_bias_0_to_fp16 = const()[name = tensor("linear_0_bias_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(6295808)))]; + tensor linear_0_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = attn0_in_proj_weight_to_fp16, x = input_3_cast_fp16)[name = tensor("linear_0_cast_fp16")]; + tensor var_62 = const()[name = tensor("op_62"), val = tensor([1, 256, 3, 16, 64])]; + tensor qkv_1_cast_fp16 = reshape(shape = var_62, x = linear_0_cast_fp16)[name = tensor("qkv_1_cast_fp16")]; + tensor q_1_begin_0 = const()[name = tensor("q_1_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor q_1_end_0 = const()[name = tensor("q_1_end_0"), val = tensor([1, 256, 1, 16, 64])]; + tensor q_1_end_mask_0 = const()[name = tensor("q_1_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor q_1_squeeze_mask_0 = const()[name = tensor("q_1_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor q_1_cast_fp16 = slice_by_index(begin = q_1_begin_0, end = q_1_end_0, end_mask = q_1_end_mask_0, squeeze_mask = q_1_squeeze_mask_0, x = qkv_1_cast_fp16)[name = tensor("q_1_cast_fp16")]; + tensor k_1_begin_0 = const()[name = tensor("k_1_begin_0"), val = tensor([0, 0, 1, 0, 0])]; + tensor k_1_end_0 = const()[name = tensor("k_1_end_0"), val = tensor([1, 256, 2, 16, 64])]; + tensor k_1_end_mask_0 = const()[name = tensor("k_1_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor k_1_squeeze_mask_0 = const()[name = tensor("k_1_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor k_1_cast_fp16 = slice_by_index(begin = k_1_begin_0, end = k_1_end_0, end_mask = k_1_end_mask_0, squeeze_mask = k_1_squeeze_mask_0, x = qkv_1_cast_fp16)[name = tensor("k_1_cast_fp16")]; + tensor v_1_begin_0 = const()[name = tensor("v_1_begin_0"), val = tensor([0, 0, 2, 0, 0])]; + tensor v_1_end_0 = const()[name = tensor("v_1_end_0"), val = tensor([1, 256, 3, 16, 64])]; + tensor v_1_end_mask_0 = const()[name = tensor("v_1_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor v_1_squeeze_mask_0 = const()[name = tensor("v_1_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor v_1_cast_fp16 = slice_by_index(begin = v_1_begin_0, end = v_1_end_0, end_mask = v_1_end_mask_0, squeeze_mask = v_1_squeeze_mask_0, x = qkv_1_cast_fp16)[name = tensor("v_1_cast_fp16")]; + tensor var_120 = const()[name = tensor("op_120"), val = tensor([1, 1])]; + tensor position0_to_fp16_dtype_0 = const()[name = tensor("position0_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor position0_to_fp16 = cast(dtype = position0_to_fp16_dtype_0, x = position0)[name = tensor("cast_158")]; + tensor var_121_cast_fp16 = reshape(shape = var_120, x = position0_to_fp16)[name = tensor("op_121_cast_fp16")]; + tensor ts_1_promoted_to_fp16 = const()[name = tensor("ts_1_promoted_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(6302016)))]; + tensor ts_3_cast_fp16 = add(x = ts_1_promoted_to_fp16, y = var_121_cast_fp16)[name = tensor("ts_3_cast_fp16")]; + tensor var_128 = const()[name = tensor("op_128"), val = tensor([1, 256, 1, 1])]; + tensor ts_5_cast_fp16 = reshape(shape = var_128, x = ts_3_cast_fp16)[name = tensor("ts_5_cast_fp16")]; + tensor var_135 = const()[name = tensor("op_135"), val = tensor([1, 256, 16, 32, 2])]; + tensor q_complex_1_cast_fp16 = reshape(shape = var_135, x = q_1_cast_fp16)[name = tensor("q_complex_1_cast_fp16")]; + tensor var_142 = const()[name = tensor("op_142"), val = tensor([1, 256, 16, 32, 2])]; + tensor k_complex_1_cast_fp16 = reshape(shape = var_142, x = k_1_cast_fp16)[name = tensor("k_complex_1_cast_fp16")]; + tensor var_146_begin_0 = const()[name = tensor("op_146_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_146_end_0 = const()[name = tensor("op_146_end_0"), val = tensor([1, 256, 16, 32, 1])]; + tensor var_146_end_mask_0 = const()[name = tensor("op_146_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_146_squeeze_mask_0 = const()[name = tensor("op_146_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_146_cast_fp16 = slice_by_index(begin = var_146_begin_0, end = var_146_end_0, end_mask = var_146_end_mask_0, squeeze_mask = var_146_squeeze_mask_0, x = q_complex_1_cast_fp16)[name = tensor("op_146_cast_fp16")]; + tensor var_154_begin_0 = const()[name = tensor("op_154_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_154_end_0 = const()[name = tensor("op_154_end_0"), val = tensor([1, 256, 16, 32, 2])]; + tensor var_154_end_mask_0 = const()[name = tensor("op_154_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_154_squeeze_mask_0 = const()[name = tensor("op_154_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_154_cast_fp16 = slice_by_index(begin = var_154_begin_0, end = var_154_end_0, end_mask = var_154_end_mask_0, squeeze_mask = var_154_squeeze_mask_0, x = q_complex_1_cast_fp16)[name = tensor("op_154_cast_fp16")]; + tensor var_162_begin_0 = const()[name = tensor("op_162_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_162_end_0 = const()[name = tensor("op_162_end_0"), val = tensor([1, 256, 16, 32, 1])]; + tensor var_162_end_mask_0 = const()[name = tensor("op_162_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_162_squeeze_mask_0 = const()[name = tensor("op_162_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_162_cast_fp16 = slice_by_index(begin = var_162_begin_0, end = var_162_end_0, end_mask = var_162_end_mask_0, squeeze_mask = var_162_squeeze_mask_0, x = k_complex_1_cast_fp16)[name = tensor("op_162_cast_fp16")]; + tensor var_170_begin_0 = const()[name = tensor("op_170_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_170_end_0 = const()[name = tensor("op_170_end_0"), val = tensor([1, 256, 16, 32, 2])]; + tensor var_170_end_mask_0 = const()[name = tensor("op_170_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_170_squeeze_mask_0 = const()[name = tensor("op_170_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_170_cast_fp16 = slice_by_index(begin = var_170_begin_0, end = var_170_end_0, end_mask = var_170_end_mask_0, squeeze_mask = var_170_squeeze_mask_0, x = k_complex_1_cast_fp16)[name = tensor("op_170_cast_fp16")]; + tensor freqs_1_to_fp16 = const()[name = tensor("freqs_1_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(6302592)))]; + tensor var_176_cast_fp16 = mul(x = freqs_1_to_fp16, y = ts_5_cast_fp16)[name = tensor("op_176_cast_fp16")]; + tensor rotr_1_cast_fp16 = cos(x = var_176_cast_fp16)[name = tensor("rotr_1_cast_fp16")]; + tensor roti_1_cast_fp16 = sin(x = var_176_cast_fp16)[name = tensor("roti_1_cast_fp16")]; + tensor var_180_cast_fp16 = mul(x = var_146_cast_fp16, y = rotr_1_cast_fp16)[name = tensor("op_180_cast_fp16")]; + tensor var_181_cast_fp16 = mul(x = var_154_cast_fp16, y = roti_1_cast_fp16)[name = tensor("op_181_cast_fp16")]; + tensor qor_1_cast_fp16 = sub(x = var_180_cast_fp16, y = var_181_cast_fp16)[name = tensor("qor_1_cast_fp16")]; + tensor var_184_cast_fp16 = mul(x = var_146_cast_fp16, y = roti_1_cast_fp16)[name = tensor("op_184_cast_fp16")]; + tensor var_185_cast_fp16 = mul(x = var_154_cast_fp16, y = rotr_1_cast_fp16)[name = tensor("op_185_cast_fp16")]; + tensor qoi_1_cast_fp16 = add(x = var_184_cast_fp16, y = var_185_cast_fp16)[name = tensor("qoi_1_cast_fp16")]; + tensor var_188_cast_fp16 = mul(x = var_162_cast_fp16, y = rotr_1_cast_fp16)[name = tensor("op_188_cast_fp16")]; + tensor var_189_cast_fp16 = mul(x = var_170_cast_fp16, y = roti_1_cast_fp16)[name = tensor("op_189_cast_fp16")]; + tensor kor_1_cast_fp16 = sub(x = var_188_cast_fp16, y = var_189_cast_fp16)[name = tensor("kor_1_cast_fp16")]; + tensor var_192_cast_fp16 = mul(x = var_162_cast_fp16, y = roti_1_cast_fp16)[name = tensor("op_192_cast_fp16")]; + tensor var_193_cast_fp16 = mul(x = var_170_cast_fp16, y = rotr_1_cast_fp16)[name = tensor("op_193_cast_fp16")]; + tensor koi_1_cast_fp16 = add(x = var_192_cast_fp16, y = var_193_cast_fp16)[name = tensor("koi_1_cast_fp16")]; + tensor qo_1_axis_0 = const()[name = tensor("qo_1_axis_0"), val = tensor(-1)]; + tensor qo_1_cast_fp16 = stack(axis = qo_1_axis_0, values = (qor_1_cast_fp16, qoi_1_cast_fp16))[name = tensor("qo_1_cast_fp16")]; + tensor ko_1_axis_0 = const()[name = tensor("ko_1_axis_0"), val = tensor(-1)]; + tensor ko_1_cast_fp16 = stack(axis = ko_1_axis_0, values = (kor_1_cast_fp16, koi_1_cast_fp16))[name = tensor("ko_1_cast_fp16")]; + tensor var_226 = const()[name = tensor("op_226"), val = tensor([1, 256, 16, 64])]; + tensor q_3_cast_fp16 = reshape(shape = var_226, x = qo_1_cast_fp16)[name = tensor("q_3_cast_fp16")]; + tensor var_232 = const()[name = tensor("op_232"), val = tensor([1, 256, 16, 64])]; + tensor k_3_cast_fp16 = reshape(shape = var_232, x = ko_1_cast_fp16)[name = tensor("k_3_cast_fp16")]; + tensor var_249_promoted_to_fp16 = const()[name = tensor("op_249_promoted_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(6302720)))]; + tensor write_indices_float_1_cast_fp16 = add(x = var_121_cast_fp16, y = var_249_promoted_to_fp16)[name = tensor("write_indices_float_1_cast_fp16")]; + tensor _inversed_253_y_0_to_fp16 = const()[name = tensor("_inversed_253_y_0_to_fp16"), val = tensor(0x1p-9)]; + tensor _inversed_253_cast_fp16 = mul(x = write_indices_float_1_cast_fp16, y = _inversed_253_y_0_to_fp16)[name = tensor("_inversed_253_cast_fp16")]; + tensor var_254_cast_fp16 = floor(x = _inversed_253_cast_fp16)[name = tensor("op_254_cast_fp16")]; + tensor var_255_to_fp16 = const()[name = tensor("op_255_to_fp16"), val = tensor(0x1p+9)]; + tensor var_256_cast_fp16 = mul(x = var_254_cast_fp16, y = var_255_to_fp16)[name = tensor("op_256_cast_fp16")]; + tensor write_indices_float_3_cast_fp16 = sub(x = write_indices_float_1_cast_fp16, y = var_256_cast_fp16)[name = tensor("write_indices_float_3_cast_fp16")]; + tensor var_265 = const()[name = tensor("op_265"), val = tensor([1, 1])]; + tensor valid_len_to_fp16_dtype_0 = const()[name = tensor("valid_len_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor valid_len_to_fp16 = cast(dtype = valid_len_to_fp16_dtype_0, x = valid_len)[name = tensor("cast_157")]; + tensor var_266_cast_fp16 = reshape(shape = var_265, x = valid_len_to_fp16)[name = tensor("op_266_cast_fp16")]; + tensor token_valid_1_cast_fp16 = less(x = var_249_promoted_to_fp16, y = var_266_cast_fp16)[name = tensor("token_valid_1_cast_fp16")]; + tensor var_274_value_0_to_fp16 = const()[name = tensor("op_274_value_0_to_fp16"), val = tensor(0x1.ffp+8)]; + tensor var_274_cast_fp16 = fill_like(ref_tensor = write_indices_float_3_cast_fp16, value = var_274_value_0_to_fp16)[name = tensor("op_274_cast_fp16")]; + tensor write_indices_float_5_cast_fp16 = select(a = write_indices_float_3_cast_fp16, b = var_274_cast_fp16, cond = token_valid_1_cast_fp16)[name = tensor("write_indices_float_5_cast_fp16")]; + tensor var_280_dtype_0 = const()[name = tensor("op_280_dtype_0"), val = tensor("int32")]; + tensor var_285 = const()[name = tensor("op_285"), val = tensor([1, 256, 1, 1])]; + tensor write_indices_float_5_cast_fp16_to_int32 = cast(dtype = var_280_dtype_0, x = write_indices_float_5_cast_fp16)[name = tensor("cast_156")]; + tensor var_286 = reshape(shape = var_285, x = write_indices_float_5_cast_fp16_to_int32)[name = tensor("op_286")]; + tensor write_indices_1_reps_0 = const()[name = tensor("write_indices_1_reps_0"), val = tensor([1, 1, 16, 64])]; + tensor write_indices_1 = tile(reps = write_indices_1_reps_0, x = var_286)[name = tensor("write_indices_1")]; + tensor var_296_begin_0 = const()[name = tensor("op_296_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_296_end_0 = const()[name = tensor("op_296_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor var_296_end_mask_0 = const()[name = tensor("op_296_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_296_squeeze_mask_0 = const()[name = tensor("op_296_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor cache0_to_fp16_dtype_0 = const()[name = tensor("cache0_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor cache0_to_fp16 = cast(dtype = cache0_to_fp16_dtype_0, x = cache0)[name = tensor("cast_155")]; + tensor var_296_cast_fp16 = slice_by_index(begin = var_296_begin_0, end = var_296_end_0, end_mask = var_296_end_mask_0, squeeze_mask = var_296_squeeze_mask_0, x = cache0_to_fp16)[name = tensor("op_296_cast_fp16")]; + tensor var_298_axis_0 = const()[name = tensor("op_298_axis_0"), val = tensor(1)]; + tensor var_298_mode_0 = const()[name = tensor("op_298_mode_0"), val = tensor("update")]; + tensor var_298_validate_indices_0 = const()[name = tensor("op_298_validate_indices_0"), val = tensor(false)]; + tensor var_298_cast_fp16 = scatter_along_axis(axis = var_298_axis_0, data = var_296_cast_fp16, indices = write_indices_1, mode = var_298_mode_0, updates = k_3_cast_fp16, validate_indices = var_298_validate_indices_0)[name = tensor("op_298_cast_fp16")]; + tensor concat_2 = const()[name = tensor("concat_2"), val = tensor([0, 0, 0, 0, 0])]; + tensor concat_3 = const()[name = tensor("concat_3"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_1_internal_tensor_assign_1_stride_0 = const()[name = tensor("new_cache_1_internal_tensor_assign_1_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_1_internal_tensor_assign_1_begin_mask_0 = const()[name = tensor("new_cache_1_internal_tensor_assign_1_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_1_internal_tensor_assign_1_end_mask_0 = const()[name = tensor("new_cache_1_internal_tensor_assign_1_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_1_internal_tensor_assign_1_squeeze_mask_0 = const()[name = tensor("new_cache_1_internal_tensor_assign_1_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_10 = const()[name = tensor("shape_10"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_0 = const()[name = tensor("reduce_prod_0"), val = tensor(1048576)]; + tensor range_1d_0_start_0 = const()[name = tensor("range_1d_0_start_0"), val = tensor(0)]; + tensor range_1d_0_step_0 = const()[name = tensor("range_1d_0_step_0"), val = tensor(1)]; + tensor range_1d_0 = range_1d(end = reduce_prod_0, start = range_1d_0_start_0, step = range_1d_0_step_0)[name = tensor("range_1d_0")]; + tensor reshape_0 = reshape(shape = shape_10, x = range_1d_0)[name = tensor("reshape_0")]; + tensor slice_by_index_0 = slice_by_index(begin = concat_2, begin_mask = new_cache_1_internal_tensor_assign_1_begin_mask_0, end = concat_3, end_mask = new_cache_1_internal_tensor_assign_1_end_mask_0, squeeze_mask = new_cache_1_internal_tensor_assign_1_squeeze_mask_0, stride = new_cache_1_internal_tensor_assign_1_stride_0, x = reshape_0)[name = tensor("slice_by_index_0")]; + tensor reshape_1_shape_0 = const()[name = tensor("reshape_1_shape_0"), val = tensor([-1])]; + tensor reshape_1 = reshape(shape = reshape_1_shape_0, x = slice_by_index_0)[name = tensor("reshape_1")]; + tensor reshape_2_shape_0 = const()[name = tensor("reshape_2_shape_0"), val = tensor([-1])]; + tensor reshape_2_cast_fp16 = reshape(shape = reshape_2_shape_0, x = var_298_cast_fp16)[name = tensor("reshape_2_cast_fp16")]; + tensor reshape_3_shape_0 = const()[name = tensor("reshape_3_shape_0"), val = tensor([-1])]; + tensor reshape_3_cast_fp16 = reshape(shape = reshape_3_shape_0, x = cache0_to_fp16)[name = tensor("reshape_3_cast_fp16")]; + tensor scatter_0_mode_0 = const()[name = tensor("scatter_0_mode_0"), val = tensor("update")]; + tensor scatter_0_axis_0 = const()[name = tensor("scatter_0_axis_0"), val = tensor(0)]; + tensor scatter_0_validate_indices_0 = const()[name = tensor("scatter_0_validate_indices_0"), val = tensor(false)]; + tensor scatter_0_cast_fp16 = scatter(axis = scatter_0_axis_0, data = reshape_3_cast_fp16, indices = reshape_1, mode = scatter_0_mode_0, updates = reshape_2_cast_fp16, validate_indices = scatter_0_validate_indices_0)[name = tensor("scatter_0_cast_fp16")]; + tensor reshape_4_cast_fp16 = reshape(shape = shape_10, x = scatter_0_cast_fp16)[name = tensor("reshape_4_cast_fp16")]; + tensor var_306_begin_0 = const()[name = tensor("op_306_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor var_306_end_0 = const()[name = tensor("op_306_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor var_306_end_mask_0 = const()[name = tensor("op_306_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_306_squeeze_mask_0 = const()[name = tensor("op_306_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor var_306_cast_fp16 = slice_by_index(begin = var_306_begin_0, end = var_306_end_0, end_mask = var_306_end_mask_0, squeeze_mask = var_306_squeeze_mask_0, x = reshape_4_cast_fp16)[name = tensor("op_306_cast_fp16")]; + tensor var_308_axis_0 = const()[name = tensor("op_308_axis_0"), val = tensor(1)]; + tensor var_308_mode_0 = const()[name = tensor("op_308_mode_0"), val = tensor("update")]; + tensor var_308_validate_indices_0 = const()[name = tensor("op_308_validate_indices_0"), val = tensor(false)]; + tensor var_308_cast_fp16 = scatter_along_axis(axis = var_308_axis_0, data = var_306_cast_fp16, indices = write_indices_1, mode = var_308_mode_0, updates = v_1_cast_fp16, validate_indices = var_308_validate_indices_0)[name = tensor("op_308_cast_fp16")]; + tensor concat_4 = const()[name = tensor("concat_4"), val = tensor([1, 0, 0, 0, 0])]; + tensor concat_5 = const()[name = tensor("concat_5"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_1_internal_tensor_assign_2_stride_0 = const()[name = tensor("new_cache_1_internal_tensor_assign_2_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_1_internal_tensor_assign_2_begin_mask_0 = const()[name = tensor("new_cache_1_internal_tensor_assign_2_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_1_internal_tensor_assign_2_end_mask_0 = const()[name = tensor("new_cache_1_internal_tensor_assign_2_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_1_internal_tensor_assign_2_squeeze_mask_0 = const()[name = tensor("new_cache_1_internal_tensor_assign_2_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_11 = const()[name = tensor("shape_11"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_1 = const()[name = tensor("reduce_prod_1"), val = tensor(1048576)]; + tensor range_1d_1_start_0 = const()[name = tensor("range_1d_1_start_0"), val = tensor(0)]; + tensor range_1d_1_step_0 = const()[name = tensor("range_1d_1_step_0"), val = tensor(1)]; + tensor range_1d_1 = range_1d(end = reduce_prod_1, start = range_1d_1_start_0, step = range_1d_1_step_0)[name = tensor("range_1d_1")]; + tensor reshape_5 = reshape(shape = shape_11, x = range_1d_1)[name = tensor("reshape_5")]; + tensor slice_by_index_1 = slice_by_index(begin = concat_4, begin_mask = new_cache_1_internal_tensor_assign_2_begin_mask_0, end = concat_5, end_mask = new_cache_1_internal_tensor_assign_2_end_mask_0, squeeze_mask = new_cache_1_internal_tensor_assign_2_squeeze_mask_0, stride = new_cache_1_internal_tensor_assign_2_stride_0, x = reshape_5)[name = tensor("slice_by_index_1")]; + tensor reshape_6_shape_0 = const()[name = tensor("reshape_6_shape_0"), val = tensor([-1])]; + tensor reshape_6 = reshape(shape = reshape_6_shape_0, x = slice_by_index_1)[name = tensor("reshape_6")]; + tensor reshape_7_shape_0 = const()[name = tensor("reshape_7_shape_0"), val = tensor([-1])]; + tensor reshape_7_cast_fp16 = reshape(shape = reshape_7_shape_0, x = var_308_cast_fp16)[name = tensor("reshape_7_cast_fp16")]; + tensor reshape_8_shape_0 = const()[name = tensor("reshape_8_shape_0"), val = tensor([-1])]; + tensor reshape_8_cast_fp16 = reshape(shape = reshape_8_shape_0, x = reshape_4_cast_fp16)[name = tensor("reshape_8_cast_fp16")]; + tensor scatter_1_mode_0 = const()[name = tensor("scatter_1_mode_0"), val = tensor("update")]; + tensor scatter_1_axis_0 = const()[name = tensor("scatter_1_axis_0"), val = tensor(0)]; + tensor scatter_1_validate_indices_0 = const()[name = tensor("scatter_1_validate_indices_0"), val = tensor(false)]; + tensor scatter_1_cast_fp16 = scatter(axis = scatter_1_axis_0, data = reshape_8_cast_fp16, indices = reshape_6, mode = scatter_1_mode_0, updates = reshape_7_cast_fp16, validate_indices = scatter_1_validate_indices_0)[name = tensor("scatter_1_cast_fp16")]; + tensor reshape_9_cast_fp16 = reshape(shape = shape_11, x = scatter_1_cast_fp16)[name = tensor("reshape_9_cast_fp16")]; + tensor reshape_9_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("reshape_9_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor keys_1_begin_0 = const()[name = tensor("keys_1_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor keys_1_end_0 = const()[name = tensor("keys_1_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor keys_1_end_mask_0 = const()[name = tensor("keys_1_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor keys_1_squeeze_mask_0 = const()[name = tensor("keys_1_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor keys_1_cast_fp16 = slice_by_index(begin = keys_1_begin_0, end = keys_1_end_0, end_mask = keys_1_end_mask_0, squeeze_mask = keys_1_squeeze_mask_0, x = reshape_9_cast_fp16)[name = tensor("keys_1_cast_fp16")]; + tensor values_1_begin_0 = const()[name = tensor("values_1_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor values_1_end_0 = const()[name = tensor("values_1_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor values_1_end_mask_0 = const()[name = tensor("values_1_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor values_1_squeeze_mask_0 = const()[name = tensor("values_1_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor values_1_cast_fp16 = slice_by_index(begin = values_1_begin_0, end = values_1_end_0, end_mask = values_1_end_mask_0, squeeze_mask = values_1_squeeze_mask_0, x = reshape_9_cast_fp16)[name = tensor("values_1_cast_fp16")]; + tensor var_320_cast_fp16 = not_equal(x = keys_1_cast_fp16, y = keys_1_cast_fp16)[name = tensor("op_320_cast_fp16")]; + tensor var_326_to_fp16 = const()[name = tensor("op_326_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(6303296)))]; + tensor keys_3_cast_fp16 = select(a = var_326_to_fp16, b = keys_1_cast_fp16, cond = var_320_cast_fp16)[name = tensor("keys_3_cast_fp16")]; + tensor var_328_cast_fp16 = not_equal(x = values_1_cast_fp16, y = values_1_cast_fp16)[name = tensor("op_328_cast_fp16")]; + tensor values_3_cast_fp16 = select(a = var_326_to_fp16, b = values_1_cast_fp16, cond = var_328_cast_fp16)[name = tensor("values_3_cast_fp16")]; + tensor var_352 = const()[name = tensor("op_352"), val = tensor([0, 2, 1, 3])]; + tensor var_368 = const()[name = tensor("op_368"), val = tensor([1, 1, 1])]; + tensor var_369_cast_fp16 = reshape(shape = var_368, x = position0_to_fp16)[name = tensor("op_369_cast_fp16")]; + tensor q_offsets_1_promoted_to_fp16 = const()[name = tensor("q_offsets_1_promoted_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(7351936)))]; + tensor q_positions_1_cast_fp16 = add(x = var_369_cast_fp16, y = q_offsets_1_promoted_to_fp16)[name = tensor("q_positions_1_cast_fp16")]; + tensor var_391 = const()[name = tensor("op_391"), val = tensor([1, 1, 1])]; + tensor var_392_cast_fp16 = reshape(shape = var_391, x = valid_len_to_fp16)[name = tensor("op_392_cast_fp16")]; + tensor valid_end_1_cast_fp16 = add(x = var_369_cast_fp16, y = var_392_cast_fp16)[name = tensor("valid_end_1_cast_fp16")]; + tensor k_positions_1_promoted_to_fp16 = const()[name = tensor("k_positions_1_promoted_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(7352512)))]; + tensor valid_mask_1_cast_fp16 = less(x = k_positions_1_promoted_to_fp16, y = valid_end_1_cast_fp16)[name = tensor("valid_mask_1_cast_fp16")]; + tensor causal_mask_1_cast_fp16 = less_equal(x = k_positions_1_promoted_to_fp16, y = q_positions_1_cast_fp16)[name = tensor("causal_mask_1_cast_fp16")]; + tensor var_397 = logical_and(x = valid_mask_1_cast_fp16, y = causal_mask_1_cast_fp16)[name = tensor("op_397")]; + tensor attn_mask_1_axes_0 = const()[name = tensor("attn_mask_1_axes_0"), val = tensor([1])]; + tensor attn_mask_1 = expand_dims(axes = attn_mask_1_axes_0, x = var_397)[name = tensor("attn_mask_1")]; + tensor var_411_transpose_x_0 = const()[name = tensor("op_411_transpose_x_0"), val = tensor(false)]; + tensor var_411_transpose_y_0 = const()[name = tensor("op_411_transpose_y_0"), val = tensor(false)]; + tensor transpose_15_perm_0 = const()[name = tensor("transpose_15_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_16_perm_0 = const()[name = tensor("transpose_16_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_16 = transpose(perm = transpose_16_perm_0, x = keys_3_cast_fp16)[name = tensor("transpose_42")]; + tensor transpose_15 = transpose(perm = transpose_15_perm_0, x = q_3_cast_fp16)[name = tensor("transpose_43")]; + tensor var_411_cast_fp16 = matmul(transpose_x = var_411_transpose_x_0, transpose_y = var_411_transpose_y_0, x = transpose_15, y = transpose_16)[name = tensor("op_411_cast_fp16")]; + tensor var_405_to_fp16 = const()[name = tensor("op_405_to_fp16"), val = tensor([0x1p-3])]; + tensor attn_weights_1_cast_fp16 = mul(x = var_411_cast_fp16, y = var_405_to_fp16)[name = tensor("attn_weights_1_cast_fp16")]; + tensor var_413 = logical_not(x = attn_mask_1)[name = tensor("op_413")]; + tensor var_414_to_fp16 = const()[name = tensor("op_414_to_fp16"), val = tensor(-inf)]; + tensor attn_weights_3_cast_fp16 = select(a = var_414_to_fp16, b = attn_weights_1_cast_fp16, cond = var_413)[name = tensor("attn_weights_3_cast_fp16")]; + tensor var_416 = const()[name = tensor("op_416"), val = tensor(-1)]; + tensor attn_weights_5_cast_fp16 = softmax(axis = var_416, x = attn_weights_3_cast_fp16)[name = tensor("attn_weights_5_cast_fp16")]; + tensor attn_output_1_transpose_x_0 = const()[name = tensor("attn_output_1_transpose_x_0"), val = tensor(false)]; + tensor attn_output_1_transpose_y_0 = const()[name = tensor("attn_output_1_transpose_y_0"), val = tensor(false)]; + tensor values_5_cast_fp16 = transpose(perm = var_352, x = values_3_cast_fp16)[name = tensor("transpose_44")]; + tensor attn_output_1_cast_fp16 = matmul(transpose_x = attn_output_1_transpose_x_0, transpose_y = attn_output_1_transpose_y_0, x = attn_weights_5_cast_fp16, y = values_5_cast_fp16)[name = tensor("attn_output_1_cast_fp16")]; + tensor var_424 = const()[name = tensor("op_424"), val = tensor([0, 2, 1, 3])]; + tensor var_429 = const()[name = tensor("op_429"), val = tensor([1, 256, 1024])]; + tensor var_425_cast_fp16 = transpose(perm = var_424, x = attn_output_1_cast_fp16)[name = tensor("transpose_41")]; + tensor input_5_cast_fp16 = reshape(shape = var_429, x = var_425_cast_fp16)[name = tensor("input_5_cast_fp16")]; + tensor attn0_out_proj_weight_to_fp16 = const()[name = tensor("attn0_out_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(7353600)))]; + tensor linear_1_bias_0_to_fp16 = const()[name = tensor("linear_1_bias_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9450816)))]; + tensor linear_1_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = attn0_out_proj_weight_to_fp16, x = input_5_cast_fp16)[name = tensor("linear_1_cast_fp16")]; + tensor var_438_cast_fp16 = add(x = position0_to_fp16, y = valid_len_to_fp16)[name = tensor("op_438_cast_fp16")]; + tensor var_438_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_438_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor input_7_cast_fp16 = add(x = conditioning_to_fp16, y = linear_1_cast_fp16)[name = tensor("input_7_cast_fp16")]; + tensor input_9_axes_0 = const()[name = tensor("input_9_axes_0"), val = tensor([-1])]; + tensor norm0_2_weight_to_fp16 = const()[name = tensor("norm0_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9452928)))]; + tensor norm0_2_bias_to_fp16 = const()[name = tensor("norm0_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9455040)))]; + tensor var_442_to_fp16 = const()[name = tensor("op_442_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_9_cast_fp16 = layer_norm(axes = input_9_axes_0, beta = norm0_2_bias_to_fp16, epsilon = var_442_to_fp16, gamma = norm0_2_weight_to_fp16, x = input_7_cast_fp16)[name = tensor("input_9_cast_fp16")]; + tensor linear0_1_weight_to_fp16 = const()[name = tensor("linear0_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9457152)))]; + tensor linear_2_bias_0_to_fp16 = const()[name = tensor("linear_2_bias_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(17845824)))]; + tensor linear_2_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = linear0_1_weight_to_fp16, x = input_9_cast_fp16)[name = tensor("linear_2_cast_fp16")]; + tensor input_11_mode_0 = const()[name = tensor("input_11_mode_0"), val = tensor("EXACT")]; + tensor input_11_cast_fp16 = gelu(mode = input_11_mode_0, x = linear_2_cast_fp16)[name = tensor("input_11_cast_fp16")]; + tensor linear0_2_weight_to_fp16 = const()[name = tensor("linear0_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(17854080)))]; + tensor linear_3_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = linear0_2_weight_to_fp16, x = input_11_cast_fp16)[name = tensor("linear_3_cast_fp16")]; + tensor input_13_cast_fp16 = add(x = input_7_cast_fp16, y = linear_3_cast_fp16)[name = tensor("input_13_cast_fp16")]; + tensor input_15_axes_0 = const()[name = tensor("input_15_axes_0"), val = tensor([-1])]; + tensor norm1_1_weight_to_fp16 = const()[name = tensor("norm1_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(26242752)))]; + tensor norm1_1_bias_to_fp16 = const()[name = tensor("norm1_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(26244864)))]; + tensor var_459_to_fp16 = const()[name = tensor("op_459_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_15_cast_fp16 = layer_norm(axes = input_15_axes_0, beta = norm1_1_bias_to_fp16, epsilon = var_459_to_fp16, gamma = norm1_1_weight_to_fp16, x = input_13_cast_fp16)[name = tensor("input_15_cast_fp16")]; + tensor attn1_in_proj_weight_to_fp16 = const()[name = tensor("attn1_in_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(26246976)))]; + tensor linear_4_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = attn1_in_proj_weight_to_fp16, x = input_15_cast_fp16)[name = tensor("linear_4_cast_fp16")]; + tensor var_473 = const()[name = tensor("op_473"), val = tensor([1, 256, 3, 16, 64])]; + tensor qkv_3_cast_fp16 = reshape(shape = var_473, x = linear_4_cast_fp16)[name = tensor("qkv_3_cast_fp16")]; + tensor q_7_begin_0 = const()[name = tensor("q_7_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor q_7_end_0 = const()[name = tensor("q_7_end_0"), val = tensor([1, 256, 1, 16, 64])]; + tensor q_7_end_mask_0 = const()[name = tensor("q_7_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor q_7_squeeze_mask_0 = const()[name = tensor("q_7_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor q_7_cast_fp16 = slice_by_index(begin = q_7_begin_0, end = q_7_end_0, end_mask = q_7_end_mask_0, squeeze_mask = q_7_squeeze_mask_0, x = qkv_3_cast_fp16)[name = tensor("q_7_cast_fp16")]; + tensor k_5_begin_0 = const()[name = tensor("k_5_begin_0"), val = tensor([0, 0, 1, 0, 0])]; + tensor k_5_end_0 = const()[name = tensor("k_5_end_0"), val = tensor([1, 256, 2, 16, 64])]; + tensor k_5_end_mask_0 = const()[name = tensor("k_5_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor k_5_squeeze_mask_0 = const()[name = tensor("k_5_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor k_5_cast_fp16 = slice_by_index(begin = k_5_begin_0, end = k_5_end_0, end_mask = k_5_end_mask_0, squeeze_mask = k_5_squeeze_mask_0, x = qkv_3_cast_fp16)[name = tensor("k_5_cast_fp16")]; + tensor v_3_begin_0 = const()[name = tensor("v_3_begin_0"), val = tensor([0, 0, 2, 0, 0])]; + tensor v_3_end_0 = const()[name = tensor("v_3_end_0"), val = tensor([1, 256, 3, 16, 64])]; + tensor v_3_end_mask_0 = const()[name = tensor("v_3_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor v_3_squeeze_mask_0 = const()[name = tensor("v_3_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor v_3_cast_fp16 = slice_by_index(begin = v_3_begin_0, end = v_3_end_0, end_mask = v_3_end_mask_0, squeeze_mask = v_3_squeeze_mask_0, x = qkv_3_cast_fp16)[name = tensor("v_3_cast_fp16")]; + tensor var_531 = const()[name = tensor("op_531"), val = tensor([1, 1])]; + tensor position1_to_fp16_dtype_0 = const()[name = tensor("position1_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor position1_to_fp16 = cast(dtype = position1_to_fp16_dtype_0, x = position1)[name = tensor("cast_152")]; + tensor var_532_cast_fp16 = reshape(shape = var_531, x = position1_to_fp16)[name = tensor("op_532_cast_fp16")]; + tensor ts_9_cast_fp16 = add(x = ts_1_promoted_to_fp16, y = var_532_cast_fp16)[name = tensor("ts_9_cast_fp16")]; + tensor var_539 = const()[name = tensor("op_539"), val = tensor([1, 256, 1, 1])]; + tensor ts_11_cast_fp16 = reshape(shape = var_539, x = ts_9_cast_fp16)[name = tensor("ts_11_cast_fp16")]; + tensor var_546 = const()[name = tensor("op_546"), val = tensor([1, 256, 16, 32, 2])]; + tensor q_complex_3_cast_fp16 = reshape(shape = var_546, x = q_7_cast_fp16)[name = tensor("q_complex_3_cast_fp16")]; + tensor var_553 = const()[name = tensor("op_553"), val = tensor([1, 256, 16, 32, 2])]; + tensor k_complex_3_cast_fp16 = reshape(shape = var_553, x = k_5_cast_fp16)[name = tensor("k_complex_3_cast_fp16")]; + tensor var_557_begin_0 = const()[name = tensor("op_557_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_557_end_0 = const()[name = tensor("op_557_end_0"), val = tensor([1, 256, 16, 32, 1])]; + tensor var_557_end_mask_0 = const()[name = tensor("op_557_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_557_squeeze_mask_0 = const()[name = tensor("op_557_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_557_cast_fp16 = slice_by_index(begin = var_557_begin_0, end = var_557_end_0, end_mask = var_557_end_mask_0, squeeze_mask = var_557_squeeze_mask_0, x = q_complex_3_cast_fp16)[name = tensor("op_557_cast_fp16")]; + tensor var_565_begin_0 = const()[name = tensor("op_565_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_565_end_0 = const()[name = tensor("op_565_end_0"), val = tensor([1, 256, 16, 32, 2])]; + tensor var_565_end_mask_0 = const()[name = tensor("op_565_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_565_squeeze_mask_0 = const()[name = tensor("op_565_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_565_cast_fp16 = slice_by_index(begin = var_565_begin_0, end = var_565_end_0, end_mask = var_565_end_mask_0, squeeze_mask = var_565_squeeze_mask_0, x = q_complex_3_cast_fp16)[name = tensor("op_565_cast_fp16")]; + tensor var_573_begin_0 = const()[name = tensor("op_573_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_573_end_0 = const()[name = tensor("op_573_end_0"), val = tensor([1, 256, 16, 32, 1])]; + tensor var_573_end_mask_0 = const()[name = tensor("op_573_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_573_squeeze_mask_0 = const()[name = tensor("op_573_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_573_cast_fp16 = slice_by_index(begin = var_573_begin_0, end = var_573_end_0, end_mask = var_573_end_mask_0, squeeze_mask = var_573_squeeze_mask_0, x = k_complex_3_cast_fp16)[name = tensor("op_573_cast_fp16")]; + tensor var_581_begin_0 = const()[name = tensor("op_581_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_581_end_0 = const()[name = tensor("op_581_end_0"), val = tensor([1, 256, 16, 32, 2])]; + tensor var_581_end_mask_0 = const()[name = tensor("op_581_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_581_squeeze_mask_0 = const()[name = tensor("op_581_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_581_cast_fp16 = slice_by_index(begin = var_581_begin_0, end = var_581_end_0, end_mask = var_581_end_mask_0, squeeze_mask = var_581_squeeze_mask_0, x = k_complex_3_cast_fp16)[name = tensor("op_581_cast_fp16")]; + tensor freqs_3_to_fp16 = const()[name = tensor("freqs_3_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(32538496)))]; + tensor var_587_cast_fp16 = mul(x = freqs_3_to_fp16, y = ts_11_cast_fp16)[name = tensor("op_587_cast_fp16")]; + tensor rotr_3_cast_fp16 = cos(x = var_587_cast_fp16)[name = tensor("rotr_3_cast_fp16")]; + tensor roti_3_cast_fp16 = sin(x = var_587_cast_fp16)[name = tensor("roti_3_cast_fp16")]; + tensor var_591_cast_fp16 = mul(x = var_557_cast_fp16, y = rotr_3_cast_fp16)[name = tensor("op_591_cast_fp16")]; + tensor var_592_cast_fp16 = mul(x = var_565_cast_fp16, y = roti_3_cast_fp16)[name = tensor("op_592_cast_fp16")]; + tensor qor_5_cast_fp16 = sub(x = var_591_cast_fp16, y = var_592_cast_fp16)[name = tensor("qor_5_cast_fp16")]; + tensor var_595_cast_fp16 = mul(x = var_557_cast_fp16, y = roti_3_cast_fp16)[name = tensor("op_595_cast_fp16")]; + tensor var_596_cast_fp16 = mul(x = var_565_cast_fp16, y = rotr_3_cast_fp16)[name = tensor("op_596_cast_fp16")]; + tensor qoi_5_cast_fp16 = add(x = var_595_cast_fp16, y = var_596_cast_fp16)[name = tensor("qoi_5_cast_fp16")]; + tensor var_599_cast_fp16 = mul(x = var_573_cast_fp16, y = rotr_3_cast_fp16)[name = tensor("op_599_cast_fp16")]; + tensor var_600_cast_fp16 = mul(x = var_581_cast_fp16, y = roti_3_cast_fp16)[name = tensor("op_600_cast_fp16")]; + tensor kor_5_cast_fp16 = sub(x = var_599_cast_fp16, y = var_600_cast_fp16)[name = tensor("kor_5_cast_fp16")]; + tensor var_603_cast_fp16 = mul(x = var_573_cast_fp16, y = roti_3_cast_fp16)[name = tensor("op_603_cast_fp16")]; + tensor var_604_cast_fp16 = mul(x = var_581_cast_fp16, y = rotr_3_cast_fp16)[name = tensor("op_604_cast_fp16")]; + tensor koi_5_cast_fp16 = add(x = var_603_cast_fp16, y = var_604_cast_fp16)[name = tensor("koi_5_cast_fp16")]; + tensor qo_3_axis_0 = const()[name = tensor("qo_3_axis_0"), val = tensor(-1)]; + tensor qo_3_cast_fp16 = stack(axis = qo_3_axis_0, values = (qor_5_cast_fp16, qoi_5_cast_fp16))[name = tensor("qo_3_cast_fp16")]; + tensor ko_3_axis_0 = const()[name = tensor("ko_3_axis_0"), val = tensor(-1)]; + tensor ko_3_cast_fp16 = stack(axis = ko_3_axis_0, values = (kor_5_cast_fp16, koi_5_cast_fp16))[name = tensor("ko_3_cast_fp16")]; + tensor var_637 = const()[name = tensor("op_637"), val = tensor([1, 256, 16, 64])]; + tensor q_9_cast_fp16 = reshape(shape = var_637, x = qo_3_cast_fp16)[name = tensor("q_9_cast_fp16")]; + tensor var_643 = const()[name = tensor("op_643"), val = tensor([1, 256, 16, 64])]; + tensor k_7_cast_fp16 = reshape(shape = var_643, x = ko_3_cast_fp16)[name = tensor("k_7_cast_fp16")]; + tensor write_indices_float_7_cast_fp16 = add(x = var_532_cast_fp16, y = var_249_promoted_to_fp16)[name = tensor("write_indices_float_7_cast_fp16")]; + tensor _inversed_664_y_0_to_fp16 = const()[name = tensor("_inversed_664_y_0_to_fp16"), val = tensor(0x1p-9)]; + tensor _inversed_664_cast_fp16 = mul(x = write_indices_float_7_cast_fp16, y = _inversed_664_y_0_to_fp16)[name = tensor("_inversed_664_cast_fp16")]; + tensor var_665_cast_fp16 = floor(x = _inversed_664_cast_fp16)[name = tensor("op_665_cast_fp16")]; + tensor var_666_to_fp16 = const()[name = tensor("op_666_to_fp16"), val = tensor(0x1p+9)]; + tensor var_667_cast_fp16 = mul(x = var_665_cast_fp16, y = var_666_to_fp16)[name = tensor("op_667_cast_fp16")]; + tensor write_indices_float_9_cast_fp16 = sub(x = write_indices_float_7_cast_fp16, y = var_667_cast_fp16)[name = tensor("write_indices_float_9_cast_fp16")]; + tensor var_685_value_0_to_fp16 = const()[name = tensor("op_685_value_0_to_fp16"), val = tensor(0x1.ffp+8)]; + tensor var_685_cast_fp16 = fill_like(ref_tensor = write_indices_float_9_cast_fp16, value = var_685_value_0_to_fp16)[name = tensor("op_685_cast_fp16")]; + tensor write_indices_float_11_cast_fp16 = select(a = write_indices_float_9_cast_fp16, b = var_685_cast_fp16, cond = token_valid_1_cast_fp16)[name = tensor("write_indices_float_11_cast_fp16")]; + tensor var_691_dtype_0 = const()[name = tensor("op_691_dtype_0"), val = tensor("int32")]; + tensor var_696 = const()[name = tensor("op_696"), val = tensor([1, 256, 1, 1])]; + tensor write_indices_float_11_cast_fp16_to_int32 = cast(dtype = var_691_dtype_0, x = write_indices_float_11_cast_fp16)[name = tensor("cast_151")]; + tensor var_697 = reshape(shape = var_696, x = write_indices_float_11_cast_fp16_to_int32)[name = tensor("op_697")]; + tensor write_indices_3_reps_0 = const()[name = tensor("write_indices_3_reps_0"), val = tensor([1, 1, 16, 64])]; + tensor write_indices_3 = tile(reps = write_indices_3_reps_0, x = var_697)[name = tensor("write_indices_3")]; + tensor var_707_begin_0 = const()[name = tensor("op_707_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_707_end_0 = const()[name = tensor("op_707_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor var_707_end_mask_0 = const()[name = tensor("op_707_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_707_squeeze_mask_0 = const()[name = tensor("op_707_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor cache1_to_fp16_dtype_0 = const()[name = tensor("cache1_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor cache1_to_fp16 = cast(dtype = cache1_to_fp16_dtype_0, x = cache1)[name = tensor("cast_150")]; + tensor var_707_cast_fp16 = slice_by_index(begin = var_707_begin_0, end = var_707_end_0, end_mask = var_707_end_mask_0, squeeze_mask = var_707_squeeze_mask_0, x = cache1_to_fp16)[name = tensor("op_707_cast_fp16")]; + tensor var_709_axis_0 = const()[name = tensor("op_709_axis_0"), val = tensor(1)]; + tensor var_709_mode_0 = const()[name = tensor("op_709_mode_0"), val = tensor("update")]; + tensor var_709_validate_indices_0 = const()[name = tensor("op_709_validate_indices_0"), val = tensor(false)]; + tensor var_709_cast_fp16 = scatter_along_axis(axis = var_709_axis_0, data = var_707_cast_fp16, indices = write_indices_3, mode = var_709_mode_0, updates = k_7_cast_fp16, validate_indices = var_709_validate_indices_0)[name = tensor("op_709_cast_fp16")]; + tensor concat_10 = const()[name = tensor("concat_10"), val = tensor([0, 0, 0, 0, 0])]; + tensor concat_11 = const()[name = tensor("concat_11"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_3_internal_tensor_assign_1_stride_0 = const()[name = tensor("new_cache_3_internal_tensor_assign_1_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_3_internal_tensor_assign_1_begin_mask_0 = const()[name = tensor("new_cache_3_internal_tensor_assign_1_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_3_internal_tensor_assign_1_end_mask_0 = const()[name = tensor("new_cache_3_internal_tensor_assign_1_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_3_internal_tensor_assign_1_squeeze_mask_0 = const()[name = tensor("new_cache_3_internal_tensor_assign_1_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_12 = const()[name = tensor("shape_12"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_2 = const()[name = tensor("reduce_prod_2"), val = tensor(1048576)]; + tensor range_1d_2_start_0 = const()[name = tensor("range_1d_2_start_0"), val = tensor(0)]; + tensor range_1d_2_step_0 = const()[name = tensor("range_1d_2_step_0"), val = tensor(1)]; + tensor range_1d_2 = range_1d(end = reduce_prod_2, start = range_1d_2_start_0, step = range_1d_2_step_0)[name = tensor("range_1d_2")]; + tensor reshape_10 = reshape(shape = shape_12, x = range_1d_2)[name = tensor("reshape_10")]; + tensor slice_by_index_2 = slice_by_index(begin = concat_10, begin_mask = new_cache_3_internal_tensor_assign_1_begin_mask_0, end = concat_11, end_mask = new_cache_3_internal_tensor_assign_1_end_mask_0, squeeze_mask = new_cache_3_internal_tensor_assign_1_squeeze_mask_0, stride = new_cache_3_internal_tensor_assign_1_stride_0, x = reshape_10)[name = tensor("slice_by_index_2")]; + tensor reshape_11_shape_0 = const()[name = tensor("reshape_11_shape_0"), val = tensor([-1])]; + tensor reshape_11 = reshape(shape = reshape_11_shape_0, x = slice_by_index_2)[name = tensor("reshape_11")]; + tensor reshape_12_shape_0 = const()[name = tensor("reshape_12_shape_0"), val = tensor([-1])]; + tensor reshape_12_cast_fp16 = reshape(shape = reshape_12_shape_0, x = var_709_cast_fp16)[name = tensor("reshape_12_cast_fp16")]; + tensor reshape_13_shape_0 = const()[name = tensor("reshape_13_shape_0"), val = tensor([-1])]; + tensor reshape_13_cast_fp16 = reshape(shape = reshape_13_shape_0, x = cache1_to_fp16)[name = tensor("reshape_13_cast_fp16")]; + tensor scatter_2_mode_0 = const()[name = tensor("scatter_2_mode_0"), val = tensor("update")]; + tensor scatter_2_axis_0 = const()[name = tensor("scatter_2_axis_0"), val = tensor(0)]; + tensor scatter_2_validate_indices_0 = const()[name = tensor("scatter_2_validate_indices_0"), val = tensor(false)]; + tensor scatter_2_cast_fp16 = scatter(axis = scatter_2_axis_0, data = reshape_13_cast_fp16, indices = reshape_11, mode = scatter_2_mode_0, updates = reshape_12_cast_fp16, validate_indices = scatter_2_validate_indices_0)[name = tensor("scatter_2_cast_fp16")]; + tensor reshape_14_cast_fp16 = reshape(shape = shape_12, x = scatter_2_cast_fp16)[name = tensor("reshape_14_cast_fp16")]; + tensor var_717_begin_0 = const()[name = tensor("op_717_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor var_717_end_0 = const()[name = tensor("op_717_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor var_717_end_mask_0 = const()[name = tensor("op_717_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_717_squeeze_mask_0 = const()[name = tensor("op_717_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor var_717_cast_fp16 = slice_by_index(begin = var_717_begin_0, end = var_717_end_0, end_mask = var_717_end_mask_0, squeeze_mask = var_717_squeeze_mask_0, x = reshape_14_cast_fp16)[name = tensor("op_717_cast_fp16")]; + tensor var_719_axis_0 = const()[name = tensor("op_719_axis_0"), val = tensor(1)]; + tensor var_719_mode_0 = const()[name = tensor("op_719_mode_0"), val = tensor("update")]; + tensor var_719_validate_indices_0 = const()[name = tensor("op_719_validate_indices_0"), val = tensor(false)]; + tensor var_719_cast_fp16 = scatter_along_axis(axis = var_719_axis_0, data = var_717_cast_fp16, indices = write_indices_3, mode = var_719_mode_0, updates = v_3_cast_fp16, validate_indices = var_719_validate_indices_0)[name = tensor("op_719_cast_fp16")]; + tensor concat_12 = const()[name = tensor("concat_12"), val = tensor([1, 0, 0, 0, 0])]; + tensor concat_13 = const()[name = tensor("concat_13"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_3_internal_tensor_assign_2_stride_0 = const()[name = tensor("new_cache_3_internal_tensor_assign_2_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_3_internal_tensor_assign_2_begin_mask_0 = const()[name = tensor("new_cache_3_internal_tensor_assign_2_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_3_internal_tensor_assign_2_end_mask_0 = const()[name = tensor("new_cache_3_internal_tensor_assign_2_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_3_internal_tensor_assign_2_squeeze_mask_0 = const()[name = tensor("new_cache_3_internal_tensor_assign_2_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_13 = const()[name = tensor("shape_13"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_3 = const()[name = tensor("reduce_prod_3"), val = tensor(1048576)]; + tensor range_1d_3_start_0 = const()[name = tensor("range_1d_3_start_0"), val = tensor(0)]; + tensor range_1d_3_step_0 = const()[name = tensor("range_1d_3_step_0"), val = tensor(1)]; + tensor range_1d_3 = range_1d(end = reduce_prod_3, start = range_1d_3_start_0, step = range_1d_3_step_0)[name = tensor("range_1d_3")]; + tensor reshape_15 = reshape(shape = shape_13, x = range_1d_3)[name = tensor("reshape_15")]; + tensor slice_by_index_3 = slice_by_index(begin = concat_12, begin_mask = new_cache_3_internal_tensor_assign_2_begin_mask_0, end = concat_13, end_mask = new_cache_3_internal_tensor_assign_2_end_mask_0, squeeze_mask = new_cache_3_internal_tensor_assign_2_squeeze_mask_0, stride = new_cache_3_internal_tensor_assign_2_stride_0, x = reshape_15)[name = tensor("slice_by_index_3")]; + tensor reshape_16_shape_0 = const()[name = tensor("reshape_16_shape_0"), val = tensor([-1])]; + tensor reshape_16 = reshape(shape = reshape_16_shape_0, x = slice_by_index_3)[name = tensor("reshape_16")]; + tensor reshape_17_shape_0 = const()[name = tensor("reshape_17_shape_0"), val = tensor([-1])]; + tensor reshape_17_cast_fp16 = reshape(shape = reshape_17_shape_0, x = var_719_cast_fp16)[name = tensor("reshape_17_cast_fp16")]; + tensor reshape_18_shape_0 = const()[name = tensor("reshape_18_shape_0"), val = tensor([-1])]; + tensor reshape_18_cast_fp16 = reshape(shape = reshape_18_shape_0, x = reshape_14_cast_fp16)[name = tensor("reshape_18_cast_fp16")]; + tensor scatter_3_mode_0 = const()[name = tensor("scatter_3_mode_0"), val = tensor("update")]; + tensor scatter_3_axis_0 = const()[name = tensor("scatter_3_axis_0"), val = tensor(0)]; + tensor scatter_3_validate_indices_0 = const()[name = tensor("scatter_3_validate_indices_0"), val = tensor(false)]; + tensor scatter_3_cast_fp16 = scatter(axis = scatter_3_axis_0, data = reshape_18_cast_fp16, indices = reshape_16, mode = scatter_3_mode_0, updates = reshape_17_cast_fp16, validate_indices = scatter_3_validate_indices_0)[name = tensor("scatter_3_cast_fp16")]; + tensor reshape_19_cast_fp16 = reshape(shape = shape_13, x = scatter_3_cast_fp16)[name = tensor("reshape_19_cast_fp16")]; + tensor reshape_19_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("reshape_19_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor keys_7_begin_0 = const()[name = tensor("keys_7_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor keys_7_end_0 = const()[name = tensor("keys_7_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor keys_7_end_mask_0 = const()[name = tensor("keys_7_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor keys_7_squeeze_mask_0 = const()[name = tensor("keys_7_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor keys_7_cast_fp16 = slice_by_index(begin = keys_7_begin_0, end = keys_7_end_0, end_mask = keys_7_end_mask_0, squeeze_mask = keys_7_squeeze_mask_0, x = reshape_19_cast_fp16)[name = tensor("keys_7_cast_fp16")]; + tensor values_7_begin_0 = const()[name = tensor("values_7_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor values_7_end_0 = const()[name = tensor("values_7_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor values_7_end_mask_0 = const()[name = tensor("values_7_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor values_7_squeeze_mask_0 = const()[name = tensor("values_7_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor values_7_cast_fp16 = slice_by_index(begin = values_7_begin_0, end = values_7_end_0, end_mask = values_7_end_mask_0, squeeze_mask = values_7_squeeze_mask_0, x = reshape_19_cast_fp16)[name = tensor("values_7_cast_fp16")]; + tensor var_731_cast_fp16 = not_equal(x = keys_7_cast_fp16, y = keys_7_cast_fp16)[name = tensor("op_731_cast_fp16")]; + tensor keys_9_cast_fp16 = select(a = var_326_to_fp16, b = keys_7_cast_fp16, cond = var_731_cast_fp16)[name = tensor("keys_9_cast_fp16")]; + tensor var_739_cast_fp16 = not_equal(x = values_7_cast_fp16, y = values_7_cast_fp16)[name = tensor("op_739_cast_fp16")]; + tensor values_9_cast_fp16 = select(a = var_326_to_fp16, b = values_7_cast_fp16, cond = var_739_cast_fp16)[name = tensor("values_9_cast_fp16")]; + tensor var_763 = const()[name = tensor("op_763"), val = tensor([0, 2, 1, 3])]; + tensor var_779 = const()[name = tensor("op_779"), val = tensor([1, 1, 1])]; + tensor var_780_cast_fp16 = reshape(shape = var_779, x = position1_to_fp16)[name = tensor("op_780_cast_fp16")]; + tensor q_positions_3_cast_fp16 = add(x = var_780_cast_fp16, y = q_offsets_1_promoted_to_fp16)[name = tensor("q_positions_3_cast_fp16")]; + tensor valid_end_3_cast_fp16 = add(x = var_780_cast_fp16, y = var_392_cast_fp16)[name = tensor("valid_end_3_cast_fp16")]; + tensor valid_mask_3_cast_fp16 = less(x = k_positions_1_promoted_to_fp16, y = valid_end_3_cast_fp16)[name = tensor("valid_mask_3_cast_fp16")]; + tensor causal_mask_3_cast_fp16 = less_equal(x = k_positions_1_promoted_to_fp16, y = q_positions_3_cast_fp16)[name = tensor("causal_mask_3_cast_fp16")]; + tensor var_808 = logical_and(x = valid_mask_3_cast_fp16, y = causal_mask_3_cast_fp16)[name = tensor("op_808")]; + tensor attn_mask_3_axes_0 = const()[name = tensor("attn_mask_3_axes_0"), val = tensor([1])]; + tensor attn_mask_3 = expand_dims(axes = attn_mask_3_axes_0, x = var_808)[name = tensor("attn_mask_3")]; + tensor var_822_transpose_x_0 = const()[name = tensor("op_822_transpose_x_0"), val = tensor(false)]; + tensor var_822_transpose_y_0 = const()[name = tensor("op_822_transpose_y_0"), val = tensor(false)]; + tensor transpose_17_perm_0 = const()[name = tensor("transpose_17_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_18_perm_0 = const()[name = tensor("transpose_18_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_18 = transpose(perm = transpose_18_perm_0, x = keys_9_cast_fp16)[name = tensor("transpose_38")]; + tensor transpose_17 = transpose(perm = transpose_17_perm_0, x = q_9_cast_fp16)[name = tensor("transpose_39")]; + tensor var_822_cast_fp16 = matmul(transpose_x = var_822_transpose_x_0, transpose_y = var_822_transpose_y_0, x = transpose_17, y = transpose_18)[name = tensor("op_822_cast_fp16")]; + tensor var_816_to_fp16 = const()[name = tensor("op_816_to_fp16"), val = tensor([0x1p-3])]; + tensor attn_weights_7_cast_fp16 = mul(x = var_822_cast_fp16, y = var_816_to_fp16)[name = tensor("attn_weights_7_cast_fp16")]; + tensor var_824 = logical_not(x = attn_mask_3)[name = tensor("op_824")]; + tensor var_825_to_fp16 = const()[name = tensor("op_825_to_fp16"), val = tensor(-inf)]; + tensor attn_weights_9_cast_fp16 = select(a = var_825_to_fp16, b = attn_weights_7_cast_fp16, cond = var_824)[name = tensor("attn_weights_9_cast_fp16")]; + tensor var_827 = const()[name = tensor("op_827"), val = tensor(-1)]; + tensor attn_weights_11_cast_fp16 = softmax(axis = var_827, x = attn_weights_9_cast_fp16)[name = tensor("attn_weights_11_cast_fp16")]; + tensor attn_output_3_transpose_x_0 = const()[name = tensor("attn_output_3_transpose_x_0"), val = tensor(false)]; + tensor attn_output_3_transpose_y_0 = const()[name = tensor("attn_output_3_transpose_y_0"), val = tensor(false)]; + tensor values_11_cast_fp16 = transpose(perm = var_763, x = values_9_cast_fp16)[name = tensor("transpose_40")]; + tensor attn_output_3_cast_fp16 = matmul(transpose_x = attn_output_3_transpose_x_0, transpose_y = attn_output_3_transpose_y_0, x = attn_weights_11_cast_fp16, y = values_11_cast_fp16)[name = tensor("attn_output_3_cast_fp16")]; + tensor var_835 = const()[name = tensor("op_835"), val = tensor([0, 2, 1, 3])]; + tensor var_840 = const()[name = tensor("op_840"), val = tensor([1, 256, 1024])]; + tensor var_836_cast_fp16 = transpose(perm = var_835, x = attn_output_3_cast_fp16)[name = tensor("transpose_37")]; + tensor input_17_cast_fp16 = reshape(shape = var_840, x = var_836_cast_fp16)[name = tensor("input_17_cast_fp16")]; + tensor attn1_out_proj_weight_to_fp16 = const()[name = tensor("attn1_out_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(32538624)))]; + tensor linear_5_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = attn1_out_proj_weight_to_fp16, x = input_17_cast_fp16)[name = tensor("linear_5_cast_fp16")]; + tensor var_849_cast_fp16 = add(x = position1_to_fp16, y = valid_len_to_fp16)[name = tensor("op_849_cast_fp16")]; + tensor var_849_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_849_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor input_19_cast_fp16 = add(x = input_13_cast_fp16, y = linear_5_cast_fp16)[name = tensor("input_19_cast_fp16")]; + tensor input_21_axes_0 = const()[name = tensor("input_21_axes_0"), val = tensor([-1])]; + tensor norm1_2_weight_to_fp16 = const()[name = tensor("norm1_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(34635840)))]; + tensor norm1_2_bias_to_fp16 = const()[name = tensor("norm1_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(34637952)))]; + tensor var_853_to_fp16 = const()[name = tensor("op_853_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_21_cast_fp16 = layer_norm(axes = input_21_axes_0, beta = norm1_2_bias_to_fp16, epsilon = var_853_to_fp16, gamma = norm1_2_weight_to_fp16, x = input_19_cast_fp16)[name = tensor("input_21_cast_fp16")]; + tensor linear1_1_weight_to_fp16 = const()[name = tensor("linear1_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(34640064)))]; + tensor linear_6_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = linear1_1_weight_to_fp16, x = input_21_cast_fp16)[name = tensor("linear_6_cast_fp16")]; + tensor input_23_mode_0 = const()[name = tensor("input_23_mode_0"), val = tensor("EXACT")]; + tensor input_23_cast_fp16 = gelu(mode = input_23_mode_0, x = linear_6_cast_fp16)[name = tensor("input_23_cast_fp16")]; + tensor linear1_2_weight_to_fp16 = const()[name = tensor("linear1_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(43028736)))]; + tensor linear_7_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = linear1_2_weight_to_fp16, x = input_23_cast_fp16)[name = tensor("linear_7_cast_fp16")]; + tensor input_25_cast_fp16 = add(x = input_19_cast_fp16, y = linear_7_cast_fp16)[name = tensor("input_25_cast_fp16")]; + tensor input_27_axes_0 = const()[name = tensor("input_27_axes_0"), val = tensor([-1])]; + tensor norm2_1_weight_to_fp16 = const()[name = tensor("norm2_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(51417408)))]; + tensor norm2_1_bias_to_fp16 = const()[name = tensor("norm2_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(51419520)))]; + tensor var_870_to_fp16 = const()[name = tensor("op_870_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_27_cast_fp16 = layer_norm(axes = input_27_axes_0, beta = norm2_1_bias_to_fp16, epsilon = var_870_to_fp16, gamma = norm2_1_weight_to_fp16, x = input_25_cast_fp16)[name = tensor("input_27_cast_fp16")]; + tensor attn2_in_proj_weight_to_fp16 = const()[name = tensor("attn2_in_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(51421632)))]; + tensor linear_8_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = attn2_in_proj_weight_to_fp16, x = input_27_cast_fp16)[name = tensor("linear_8_cast_fp16")]; + tensor var_884 = const()[name = tensor("op_884"), val = tensor([1, 256, 3, 16, 64])]; + tensor qkv_5_cast_fp16 = reshape(shape = var_884, x = linear_8_cast_fp16)[name = tensor("qkv_5_cast_fp16")]; + tensor q_13_begin_0 = const()[name = tensor("q_13_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor q_13_end_0 = const()[name = tensor("q_13_end_0"), val = tensor([1, 256, 1, 16, 64])]; + tensor q_13_end_mask_0 = const()[name = tensor("q_13_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor q_13_squeeze_mask_0 = const()[name = tensor("q_13_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor q_13_cast_fp16 = slice_by_index(begin = q_13_begin_0, end = q_13_end_0, end_mask = q_13_end_mask_0, squeeze_mask = q_13_squeeze_mask_0, x = qkv_5_cast_fp16)[name = tensor("q_13_cast_fp16")]; + tensor k_9_begin_0 = const()[name = tensor("k_9_begin_0"), val = tensor([0, 0, 1, 0, 0])]; + tensor k_9_end_0 = const()[name = tensor("k_9_end_0"), val = tensor([1, 256, 2, 16, 64])]; + tensor k_9_end_mask_0 = const()[name = tensor("k_9_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor k_9_squeeze_mask_0 = const()[name = tensor("k_9_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor k_9_cast_fp16 = slice_by_index(begin = k_9_begin_0, end = k_9_end_0, end_mask = k_9_end_mask_0, squeeze_mask = k_9_squeeze_mask_0, x = qkv_5_cast_fp16)[name = tensor("k_9_cast_fp16")]; + tensor v_5_begin_0 = const()[name = tensor("v_5_begin_0"), val = tensor([0, 0, 2, 0, 0])]; + tensor v_5_end_0 = const()[name = tensor("v_5_end_0"), val = tensor([1, 256, 3, 16, 64])]; + tensor v_5_end_mask_0 = const()[name = tensor("v_5_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor v_5_squeeze_mask_0 = const()[name = tensor("v_5_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor v_5_cast_fp16 = slice_by_index(begin = v_5_begin_0, end = v_5_end_0, end_mask = v_5_end_mask_0, squeeze_mask = v_5_squeeze_mask_0, x = qkv_5_cast_fp16)[name = tensor("v_5_cast_fp16")]; + tensor var_942 = const()[name = tensor("op_942"), val = tensor([1, 1])]; + tensor position2_to_fp16_dtype_0 = const()[name = tensor("position2_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor position2_to_fp16 = cast(dtype = position2_to_fp16_dtype_0, x = position2)[name = tensor("cast_147")]; + tensor var_943_cast_fp16 = reshape(shape = var_942, x = position2_to_fp16)[name = tensor("op_943_cast_fp16")]; + tensor ts_15_cast_fp16 = add(x = ts_1_promoted_to_fp16, y = var_943_cast_fp16)[name = tensor("ts_15_cast_fp16")]; + tensor var_950 = const()[name = tensor("op_950"), val = tensor([1, 256, 1, 1])]; + tensor ts_17_cast_fp16 = reshape(shape = var_950, x = ts_15_cast_fp16)[name = tensor("ts_17_cast_fp16")]; + tensor var_957 = const()[name = tensor("op_957"), val = tensor([1, 256, 16, 32, 2])]; + tensor q_complex_5_cast_fp16 = reshape(shape = var_957, x = q_13_cast_fp16)[name = tensor("q_complex_5_cast_fp16")]; + tensor var_964 = const()[name = tensor("op_964"), val = tensor([1, 256, 16, 32, 2])]; + tensor k_complex_5_cast_fp16 = reshape(shape = var_964, x = k_9_cast_fp16)[name = tensor("k_complex_5_cast_fp16")]; + tensor var_968_begin_0 = const()[name = tensor("op_968_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_968_end_0 = const()[name = tensor("op_968_end_0"), val = tensor([1, 256, 16, 32, 1])]; + tensor var_968_end_mask_0 = const()[name = tensor("op_968_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_968_squeeze_mask_0 = const()[name = tensor("op_968_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_968_cast_fp16 = slice_by_index(begin = var_968_begin_0, end = var_968_end_0, end_mask = var_968_end_mask_0, squeeze_mask = var_968_squeeze_mask_0, x = q_complex_5_cast_fp16)[name = tensor("op_968_cast_fp16")]; + tensor var_976_begin_0 = const()[name = tensor("op_976_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_976_end_0 = const()[name = tensor("op_976_end_0"), val = tensor([1, 256, 16, 32, 2])]; + tensor var_976_end_mask_0 = const()[name = tensor("op_976_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_976_squeeze_mask_0 = const()[name = tensor("op_976_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_976_cast_fp16 = slice_by_index(begin = var_976_begin_0, end = var_976_end_0, end_mask = var_976_end_mask_0, squeeze_mask = var_976_squeeze_mask_0, x = q_complex_5_cast_fp16)[name = tensor("op_976_cast_fp16")]; + tensor var_984_begin_0 = const()[name = tensor("op_984_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_984_end_0 = const()[name = tensor("op_984_end_0"), val = tensor([1, 256, 16, 32, 1])]; + tensor var_984_end_mask_0 = const()[name = tensor("op_984_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_984_squeeze_mask_0 = const()[name = tensor("op_984_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_984_cast_fp16 = slice_by_index(begin = var_984_begin_0, end = var_984_end_0, end_mask = var_984_end_mask_0, squeeze_mask = var_984_squeeze_mask_0, x = k_complex_5_cast_fp16)[name = tensor("op_984_cast_fp16")]; + tensor var_992_begin_0 = const()[name = tensor("op_992_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_992_end_0 = const()[name = tensor("op_992_end_0"), val = tensor([1, 256, 16, 32, 2])]; + tensor var_992_end_mask_0 = const()[name = tensor("op_992_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_992_squeeze_mask_0 = const()[name = tensor("op_992_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_992_cast_fp16 = slice_by_index(begin = var_992_begin_0, end = var_992_end_0, end_mask = var_992_end_mask_0, squeeze_mask = var_992_squeeze_mask_0, x = k_complex_5_cast_fp16)[name = tensor("op_992_cast_fp16")]; + tensor freqs_5_to_fp16 = const()[name = tensor("freqs_5_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(57713152)))]; + tensor var_998_cast_fp16 = mul(x = freqs_5_to_fp16, y = ts_17_cast_fp16)[name = tensor("op_998_cast_fp16")]; + tensor rotr_5_cast_fp16 = cos(x = var_998_cast_fp16)[name = tensor("rotr_5_cast_fp16")]; + tensor roti_5_cast_fp16 = sin(x = var_998_cast_fp16)[name = tensor("roti_5_cast_fp16")]; + tensor var_1002_cast_fp16 = mul(x = var_968_cast_fp16, y = rotr_5_cast_fp16)[name = tensor("op_1002_cast_fp16")]; + tensor var_1003_cast_fp16 = mul(x = var_976_cast_fp16, y = roti_5_cast_fp16)[name = tensor("op_1003_cast_fp16")]; + tensor qor_9_cast_fp16 = sub(x = var_1002_cast_fp16, y = var_1003_cast_fp16)[name = tensor("qor_9_cast_fp16")]; + tensor var_1006_cast_fp16 = mul(x = var_968_cast_fp16, y = roti_5_cast_fp16)[name = tensor("op_1006_cast_fp16")]; + tensor var_1007_cast_fp16 = mul(x = var_976_cast_fp16, y = rotr_5_cast_fp16)[name = tensor("op_1007_cast_fp16")]; + tensor qoi_9_cast_fp16 = add(x = var_1006_cast_fp16, y = var_1007_cast_fp16)[name = tensor("qoi_9_cast_fp16")]; + tensor var_1010_cast_fp16 = mul(x = var_984_cast_fp16, y = rotr_5_cast_fp16)[name = tensor("op_1010_cast_fp16")]; + tensor var_1011_cast_fp16 = mul(x = var_992_cast_fp16, y = roti_5_cast_fp16)[name = tensor("op_1011_cast_fp16")]; + tensor kor_9_cast_fp16 = sub(x = var_1010_cast_fp16, y = var_1011_cast_fp16)[name = tensor("kor_9_cast_fp16")]; + tensor var_1014_cast_fp16 = mul(x = var_984_cast_fp16, y = roti_5_cast_fp16)[name = tensor("op_1014_cast_fp16")]; + tensor var_1015_cast_fp16 = mul(x = var_992_cast_fp16, y = rotr_5_cast_fp16)[name = tensor("op_1015_cast_fp16")]; + tensor koi_9_cast_fp16 = add(x = var_1014_cast_fp16, y = var_1015_cast_fp16)[name = tensor("koi_9_cast_fp16")]; + tensor qo_5_axis_0 = const()[name = tensor("qo_5_axis_0"), val = tensor(-1)]; + tensor qo_5_cast_fp16 = stack(axis = qo_5_axis_0, values = (qor_9_cast_fp16, qoi_9_cast_fp16))[name = tensor("qo_5_cast_fp16")]; + tensor ko_5_axis_0 = const()[name = tensor("ko_5_axis_0"), val = tensor(-1)]; + tensor ko_5_cast_fp16 = stack(axis = ko_5_axis_0, values = (kor_9_cast_fp16, koi_9_cast_fp16))[name = tensor("ko_5_cast_fp16")]; + tensor var_1048 = const()[name = tensor("op_1048"), val = tensor([1, 256, 16, 64])]; + tensor q_15_cast_fp16 = reshape(shape = var_1048, x = qo_5_cast_fp16)[name = tensor("q_15_cast_fp16")]; + tensor var_1054 = const()[name = tensor("op_1054"), val = tensor([1, 256, 16, 64])]; + tensor k_11_cast_fp16 = reshape(shape = var_1054, x = ko_5_cast_fp16)[name = tensor("k_11_cast_fp16")]; + tensor write_indices_float_13_cast_fp16 = add(x = var_943_cast_fp16, y = var_249_promoted_to_fp16)[name = tensor("write_indices_float_13_cast_fp16")]; + tensor _inversed_1075_y_0_to_fp16 = const()[name = tensor("_inversed_1075_y_0_to_fp16"), val = tensor(0x1p-9)]; + tensor _inversed_1075_cast_fp16 = mul(x = write_indices_float_13_cast_fp16, y = _inversed_1075_y_0_to_fp16)[name = tensor("_inversed_1075_cast_fp16")]; + tensor var_1076_cast_fp16 = floor(x = _inversed_1075_cast_fp16)[name = tensor("op_1076_cast_fp16")]; + tensor var_1077_to_fp16 = const()[name = tensor("op_1077_to_fp16"), val = tensor(0x1p+9)]; + tensor var_1078_cast_fp16 = mul(x = var_1076_cast_fp16, y = var_1077_to_fp16)[name = tensor("op_1078_cast_fp16")]; + tensor write_indices_float_15_cast_fp16 = sub(x = write_indices_float_13_cast_fp16, y = var_1078_cast_fp16)[name = tensor("write_indices_float_15_cast_fp16")]; + tensor var_1096_value_0_to_fp16 = const()[name = tensor("op_1096_value_0_to_fp16"), val = tensor(0x1.ffp+8)]; + tensor var_1096_cast_fp16 = fill_like(ref_tensor = write_indices_float_15_cast_fp16, value = var_1096_value_0_to_fp16)[name = tensor("op_1096_cast_fp16")]; + tensor write_indices_float_17_cast_fp16 = select(a = write_indices_float_15_cast_fp16, b = var_1096_cast_fp16, cond = token_valid_1_cast_fp16)[name = tensor("write_indices_float_17_cast_fp16")]; + tensor var_1102_dtype_0 = const()[name = tensor("op_1102_dtype_0"), val = tensor("int32")]; + tensor var_1107 = const()[name = tensor("op_1107"), val = tensor([1, 256, 1, 1])]; + tensor write_indices_float_17_cast_fp16_to_int32 = cast(dtype = var_1102_dtype_0, x = write_indices_float_17_cast_fp16)[name = tensor("cast_146")]; + tensor var_1108 = reshape(shape = var_1107, x = write_indices_float_17_cast_fp16_to_int32)[name = tensor("op_1108")]; + tensor write_indices_5_reps_0 = const()[name = tensor("write_indices_5_reps_0"), val = tensor([1, 1, 16, 64])]; + tensor write_indices_5 = tile(reps = write_indices_5_reps_0, x = var_1108)[name = tensor("write_indices_5")]; + tensor var_1118_begin_0 = const()[name = tensor("op_1118_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_1118_end_0 = const()[name = tensor("op_1118_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor var_1118_end_mask_0 = const()[name = tensor("op_1118_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_1118_squeeze_mask_0 = const()[name = tensor("op_1118_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor cache2_to_fp16_dtype_0 = const()[name = tensor("cache2_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor cache2_to_fp16 = cast(dtype = cache2_to_fp16_dtype_0, x = cache2)[name = tensor("cast_145")]; + tensor var_1118_cast_fp16 = slice_by_index(begin = var_1118_begin_0, end = var_1118_end_0, end_mask = var_1118_end_mask_0, squeeze_mask = var_1118_squeeze_mask_0, x = cache2_to_fp16)[name = tensor("op_1118_cast_fp16")]; + tensor var_1120_axis_0 = const()[name = tensor("op_1120_axis_0"), val = tensor(1)]; + tensor var_1120_mode_0 = const()[name = tensor("op_1120_mode_0"), val = tensor("update")]; + tensor var_1120_validate_indices_0 = const()[name = tensor("op_1120_validate_indices_0"), val = tensor(false)]; + tensor var_1120_cast_fp16 = scatter_along_axis(axis = var_1120_axis_0, data = var_1118_cast_fp16, indices = write_indices_5, mode = var_1120_mode_0, updates = k_11_cast_fp16, validate_indices = var_1120_validate_indices_0)[name = tensor("op_1120_cast_fp16")]; + tensor concat_18 = const()[name = tensor("concat_18"), val = tensor([0, 0, 0, 0, 0])]; + tensor concat_19 = const()[name = tensor("concat_19"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_5_internal_tensor_assign_1_stride_0 = const()[name = tensor("new_cache_5_internal_tensor_assign_1_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_5_internal_tensor_assign_1_begin_mask_0 = const()[name = tensor("new_cache_5_internal_tensor_assign_1_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_5_internal_tensor_assign_1_end_mask_0 = const()[name = tensor("new_cache_5_internal_tensor_assign_1_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_5_internal_tensor_assign_1_squeeze_mask_0 = const()[name = tensor("new_cache_5_internal_tensor_assign_1_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_14 = const()[name = tensor("shape_14"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_4 = const()[name = tensor("reduce_prod_4"), val = tensor(1048576)]; + tensor range_1d_4_start_0 = const()[name = tensor("range_1d_4_start_0"), val = tensor(0)]; + tensor range_1d_4_step_0 = const()[name = tensor("range_1d_4_step_0"), val = tensor(1)]; + tensor range_1d_4 = range_1d(end = reduce_prod_4, start = range_1d_4_start_0, step = range_1d_4_step_0)[name = tensor("range_1d_4")]; + tensor reshape_20 = reshape(shape = shape_14, x = range_1d_4)[name = tensor("reshape_20")]; + tensor slice_by_index_4 = slice_by_index(begin = concat_18, begin_mask = new_cache_5_internal_tensor_assign_1_begin_mask_0, end = concat_19, end_mask = new_cache_5_internal_tensor_assign_1_end_mask_0, squeeze_mask = new_cache_5_internal_tensor_assign_1_squeeze_mask_0, stride = new_cache_5_internal_tensor_assign_1_stride_0, x = reshape_20)[name = tensor("slice_by_index_4")]; + tensor reshape_21_shape_0 = const()[name = tensor("reshape_21_shape_0"), val = tensor([-1])]; + tensor reshape_21 = reshape(shape = reshape_21_shape_0, x = slice_by_index_4)[name = tensor("reshape_21")]; + tensor reshape_22_shape_0 = const()[name = tensor("reshape_22_shape_0"), val = tensor([-1])]; + tensor reshape_22_cast_fp16 = reshape(shape = reshape_22_shape_0, x = var_1120_cast_fp16)[name = tensor("reshape_22_cast_fp16")]; + tensor reshape_23_shape_0 = const()[name = tensor("reshape_23_shape_0"), val = tensor([-1])]; + tensor reshape_23_cast_fp16 = reshape(shape = reshape_23_shape_0, x = cache2_to_fp16)[name = tensor("reshape_23_cast_fp16")]; + tensor scatter_4_mode_0 = const()[name = tensor("scatter_4_mode_0"), val = tensor("update")]; + tensor scatter_4_axis_0 = const()[name = tensor("scatter_4_axis_0"), val = tensor(0)]; + tensor scatter_4_validate_indices_0 = const()[name = tensor("scatter_4_validate_indices_0"), val = tensor(false)]; + tensor scatter_4_cast_fp16 = scatter(axis = scatter_4_axis_0, data = reshape_23_cast_fp16, indices = reshape_21, mode = scatter_4_mode_0, updates = reshape_22_cast_fp16, validate_indices = scatter_4_validate_indices_0)[name = tensor("scatter_4_cast_fp16")]; + tensor reshape_24_cast_fp16 = reshape(shape = shape_14, x = scatter_4_cast_fp16)[name = tensor("reshape_24_cast_fp16")]; + tensor var_1128_begin_0 = const()[name = tensor("op_1128_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor var_1128_end_0 = const()[name = tensor("op_1128_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor var_1128_end_mask_0 = const()[name = tensor("op_1128_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_1128_squeeze_mask_0 = const()[name = tensor("op_1128_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor var_1128_cast_fp16 = slice_by_index(begin = var_1128_begin_0, end = var_1128_end_0, end_mask = var_1128_end_mask_0, squeeze_mask = var_1128_squeeze_mask_0, x = reshape_24_cast_fp16)[name = tensor("op_1128_cast_fp16")]; + tensor var_1130_axis_0 = const()[name = tensor("op_1130_axis_0"), val = tensor(1)]; + tensor var_1130_mode_0 = const()[name = tensor("op_1130_mode_0"), val = tensor("update")]; + tensor var_1130_validate_indices_0 = const()[name = tensor("op_1130_validate_indices_0"), val = tensor(false)]; + tensor var_1130_cast_fp16 = scatter_along_axis(axis = var_1130_axis_0, data = var_1128_cast_fp16, indices = write_indices_5, mode = var_1130_mode_0, updates = v_5_cast_fp16, validate_indices = var_1130_validate_indices_0)[name = tensor("op_1130_cast_fp16")]; + tensor concat_20 = const()[name = tensor("concat_20"), val = tensor([1, 0, 0, 0, 0])]; + tensor concat_21 = const()[name = tensor("concat_21"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_5_internal_tensor_assign_2_stride_0 = const()[name = tensor("new_cache_5_internal_tensor_assign_2_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_5_internal_tensor_assign_2_begin_mask_0 = const()[name = tensor("new_cache_5_internal_tensor_assign_2_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_5_internal_tensor_assign_2_end_mask_0 = const()[name = tensor("new_cache_5_internal_tensor_assign_2_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_5_internal_tensor_assign_2_squeeze_mask_0 = const()[name = tensor("new_cache_5_internal_tensor_assign_2_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_15 = const()[name = tensor("shape_15"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_5 = const()[name = tensor("reduce_prod_5"), val = tensor(1048576)]; + tensor range_1d_5_start_0 = const()[name = tensor("range_1d_5_start_0"), val = tensor(0)]; + tensor range_1d_5_step_0 = const()[name = tensor("range_1d_5_step_0"), val = tensor(1)]; + tensor range_1d_5 = range_1d(end = reduce_prod_5, start = range_1d_5_start_0, step = range_1d_5_step_0)[name = tensor("range_1d_5")]; + tensor reshape_25 = reshape(shape = shape_15, x = range_1d_5)[name = tensor("reshape_25")]; + tensor slice_by_index_5 = slice_by_index(begin = concat_20, begin_mask = new_cache_5_internal_tensor_assign_2_begin_mask_0, end = concat_21, end_mask = new_cache_5_internal_tensor_assign_2_end_mask_0, squeeze_mask = new_cache_5_internal_tensor_assign_2_squeeze_mask_0, stride = new_cache_5_internal_tensor_assign_2_stride_0, x = reshape_25)[name = tensor("slice_by_index_5")]; + tensor reshape_26_shape_0 = const()[name = tensor("reshape_26_shape_0"), val = tensor([-1])]; + tensor reshape_26 = reshape(shape = reshape_26_shape_0, x = slice_by_index_5)[name = tensor("reshape_26")]; + tensor reshape_27_shape_0 = const()[name = tensor("reshape_27_shape_0"), val = tensor([-1])]; + tensor reshape_27_cast_fp16 = reshape(shape = reshape_27_shape_0, x = var_1130_cast_fp16)[name = tensor("reshape_27_cast_fp16")]; + tensor reshape_28_shape_0 = const()[name = tensor("reshape_28_shape_0"), val = tensor([-1])]; + tensor reshape_28_cast_fp16 = reshape(shape = reshape_28_shape_0, x = reshape_24_cast_fp16)[name = tensor("reshape_28_cast_fp16")]; + tensor scatter_5_mode_0 = const()[name = tensor("scatter_5_mode_0"), val = tensor("update")]; + tensor scatter_5_axis_0 = const()[name = tensor("scatter_5_axis_0"), val = tensor(0)]; + tensor scatter_5_validate_indices_0 = const()[name = tensor("scatter_5_validate_indices_0"), val = tensor(false)]; + tensor scatter_5_cast_fp16 = scatter(axis = scatter_5_axis_0, data = reshape_28_cast_fp16, indices = reshape_26, mode = scatter_5_mode_0, updates = reshape_27_cast_fp16, validate_indices = scatter_5_validate_indices_0)[name = tensor("scatter_5_cast_fp16")]; + tensor reshape_29_cast_fp16 = reshape(shape = shape_15, x = scatter_5_cast_fp16)[name = tensor("reshape_29_cast_fp16")]; + tensor reshape_29_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("reshape_29_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor keys_13_begin_0 = const()[name = tensor("keys_13_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor keys_13_end_0 = const()[name = tensor("keys_13_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor keys_13_end_mask_0 = const()[name = tensor("keys_13_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor keys_13_squeeze_mask_0 = const()[name = tensor("keys_13_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor keys_13_cast_fp16 = slice_by_index(begin = keys_13_begin_0, end = keys_13_end_0, end_mask = keys_13_end_mask_0, squeeze_mask = keys_13_squeeze_mask_0, x = reshape_29_cast_fp16)[name = tensor("keys_13_cast_fp16")]; + tensor values_13_begin_0 = const()[name = tensor("values_13_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor values_13_end_0 = const()[name = tensor("values_13_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor values_13_end_mask_0 = const()[name = tensor("values_13_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor values_13_squeeze_mask_0 = const()[name = tensor("values_13_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor values_13_cast_fp16 = slice_by_index(begin = values_13_begin_0, end = values_13_end_0, end_mask = values_13_end_mask_0, squeeze_mask = values_13_squeeze_mask_0, x = reshape_29_cast_fp16)[name = tensor("values_13_cast_fp16")]; + tensor var_1142_cast_fp16 = not_equal(x = keys_13_cast_fp16, y = keys_13_cast_fp16)[name = tensor("op_1142_cast_fp16")]; + tensor keys_15_cast_fp16 = select(a = var_326_to_fp16, b = keys_13_cast_fp16, cond = var_1142_cast_fp16)[name = tensor("keys_15_cast_fp16")]; + tensor var_1150_cast_fp16 = not_equal(x = values_13_cast_fp16, y = values_13_cast_fp16)[name = tensor("op_1150_cast_fp16")]; + tensor values_15_cast_fp16 = select(a = var_326_to_fp16, b = values_13_cast_fp16, cond = var_1150_cast_fp16)[name = tensor("values_15_cast_fp16")]; + tensor var_1174 = const()[name = tensor("op_1174"), val = tensor([0, 2, 1, 3])]; + tensor var_1190 = const()[name = tensor("op_1190"), val = tensor([1, 1, 1])]; + tensor var_1191_cast_fp16 = reshape(shape = var_1190, x = position2_to_fp16)[name = tensor("op_1191_cast_fp16")]; + tensor q_positions_5_cast_fp16 = add(x = var_1191_cast_fp16, y = q_offsets_1_promoted_to_fp16)[name = tensor("q_positions_5_cast_fp16")]; + tensor valid_end_5_cast_fp16 = add(x = var_1191_cast_fp16, y = var_392_cast_fp16)[name = tensor("valid_end_5_cast_fp16")]; + tensor valid_mask_5_cast_fp16 = less(x = k_positions_1_promoted_to_fp16, y = valid_end_5_cast_fp16)[name = tensor("valid_mask_5_cast_fp16")]; + tensor causal_mask_5_cast_fp16 = less_equal(x = k_positions_1_promoted_to_fp16, y = q_positions_5_cast_fp16)[name = tensor("causal_mask_5_cast_fp16")]; + tensor var_1219 = logical_and(x = valid_mask_5_cast_fp16, y = causal_mask_5_cast_fp16)[name = tensor("op_1219")]; + tensor attn_mask_5_axes_0 = const()[name = tensor("attn_mask_5_axes_0"), val = tensor([1])]; + tensor attn_mask_5 = expand_dims(axes = attn_mask_5_axes_0, x = var_1219)[name = tensor("attn_mask_5")]; + tensor var_1233_transpose_x_0 = const()[name = tensor("op_1233_transpose_x_0"), val = tensor(false)]; + tensor var_1233_transpose_y_0 = const()[name = tensor("op_1233_transpose_y_0"), val = tensor(false)]; + tensor transpose_19_perm_0 = const()[name = tensor("transpose_19_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_20_perm_0 = const()[name = tensor("transpose_20_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_20 = transpose(perm = transpose_20_perm_0, x = keys_15_cast_fp16)[name = tensor("transpose_34")]; + tensor transpose_19 = transpose(perm = transpose_19_perm_0, x = q_15_cast_fp16)[name = tensor("transpose_35")]; + tensor var_1233_cast_fp16 = matmul(transpose_x = var_1233_transpose_x_0, transpose_y = var_1233_transpose_y_0, x = transpose_19, y = transpose_20)[name = tensor("op_1233_cast_fp16")]; + tensor var_1227_to_fp16 = const()[name = tensor("op_1227_to_fp16"), val = tensor([0x1p-3])]; + tensor attn_weights_13_cast_fp16 = mul(x = var_1233_cast_fp16, y = var_1227_to_fp16)[name = tensor("attn_weights_13_cast_fp16")]; + tensor var_1235 = logical_not(x = attn_mask_5)[name = tensor("op_1235")]; + tensor var_1236_to_fp16 = const()[name = tensor("op_1236_to_fp16"), val = tensor(-inf)]; + tensor attn_weights_15_cast_fp16 = select(a = var_1236_to_fp16, b = attn_weights_13_cast_fp16, cond = var_1235)[name = tensor("attn_weights_15_cast_fp16")]; + tensor var_1238 = const()[name = tensor("op_1238"), val = tensor(-1)]; + tensor attn_weights_17_cast_fp16 = softmax(axis = var_1238, x = attn_weights_15_cast_fp16)[name = tensor("attn_weights_17_cast_fp16")]; + tensor attn_output_5_transpose_x_0 = const()[name = tensor("attn_output_5_transpose_x_0"), val = tensor(false)]; + tensor attn_output_5_transpose_y_0 = const()[name = tensor("attn_output_5_transpose_y_0"), val = tensor(false)]; + tensor values_17_cast_fp16 = transpose(perm = var_1174, x = values_15_cast_fp16)[name = tensor("transpose_36")]; + tensor attn_output_5_cast_fp16 = matmul(transpose_x = attn_output_5_transpose_x_0, transpose_y = attn_output_5_transpose_y_0, x = attn_weights_17_cast_fp16, y = values_17_cast_fp16)[name = tensor("attn_output_5_cast_fp16")]; + tensor var_1246 = const()[name = tensor("op_1246"), val = tensor([0, 2, 1, 3])]; + tensor var_1251 = const()[name = tensor("op_1251"), val = tensor([1, 256, 1024])]; + tensor var_1247_cast_fp16 = transpose(perm = var_1246, x = attn_output_5_cast_fp16)[name = tensor("transpose_33")]; + tensor input_29_cast_fp16 = reshape(shape = var_1251, x = var_1247_cast_fp16)[name = tensor("input_29_cast_fp16")]; + tensor attn2_out_proj_weight_to_fp16 = const()[name = tensor("attn2_out_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(57713280)))]; + tensor linear_9_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = attn2_out_proj_weight_to_fp16, x = input_29_cast_fp16)[name = tensor("linear_9_cast_fp16")]; + tensor var_1260_cast_fp16 = add(x = position2_to_fp16, y = valid_len_to_fp16)[name = tensor("op_1260_cast_fp16")]; + tensor var_1260_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_1260_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor input_31_cast_fp16 = add(x = input_25_cast_fp16, y = linear_9_cast_fp16)[name = tensor("input_31_cast_fp16")]; + tensor input_33_axes_0 = const()[name = tensor("input_33_axes_0"), val = tensor([-1])]; + tensor norm2_2_weight_to_fp16 = const()[name = tensor("norm2_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(59810496)))]; + tensor norm2_2_bias_to_fp16 = const()[name = tensor("norm2_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(59812608)))]; + tensor var_1264_to_fp16 = const()[name = tensor("op_1264_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_33_cast_fp16 = layer_norm(axes = input_33_axes_0, beta = norm2_2_bias_to_fp16, epsilon = var_1264_to_fp16, gamma = norm2_2_weight_to_fp16, x = input_31_cast_fp16)[name = tensor("input_33_cast_fp16")]; + tensor linear2_1_weight_to_fp16 = const()[name = tensor("linear2_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(59814720)))]; + tensor linear_10_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = linear2_1_weight_to_fp16, x = input_33_cast_fp16)[name = tensor("linear_10_cast_fp16")]; + tensor input_35_mode_0 = const()[name = tensor("input_35_mode_0"), val = tensor("EXACT")]; + tensor input_35_cast_fp16 = gelu(mode = input_35_mode_0, x = linear_10_cast_fp16)[name = tensor("input_35_cast_fp16")]; + tensor linear2_2_weight_to_fp16 = const()[name = tensor("linear2_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(68203392)))]; + tensor linear_11_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = linear2_2_weight_to_fp16, x = input_35_cast_fp16)[name = tensor("linear_11_cast_fp16")]; + tensor input_37_cast_fp16 = add(x = input_31_cast_fp16, y = linear_11_cast_fp16)[name = tensor("input_37_cast_fp16")]; + tensor input_39_axes_0 = const()[name = tensor("input_39_axes_0"), val = tensor([-1])]; + tensor norm3_1_weight_to_fp16 = const()[name = tensor("norm3_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(76592064)))]; + tensor norm3_1_bias_to_fp16 = const()[name = tensor("norm3_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(76594176)))]; + tensor var_1281_to_fp16 = const()[name = tensor("op_1281_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_39_cast_fp16 = layer_norm(axes = input_39_axes_0, beta = norm3_1_bias_to_fp16, epsilon = var_1281_to_fp16, gamma = norm3_1_weight_to_fp16, x = input_37_cast_fp16)[name = tensor("input_39_cast_fp16")]; + tensor attn3_in_proj_weight_to_fp16 = const()[name = tensor("attn3_in_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(76596288)))]; + tensor linear_12_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = attn3_in_proj_weight_to_fp16, x = input_39_cast_fp16)[name = tensor("linear_12_cast_fp16")]; + tensor var_1295 = const()[name = tensor("op_1295"), val = tensor([1, 256, 3, 16, 64])]; + tensor qkv_7_cast_fp16 = reshape(shape = var_1295, x = linear_12_cast_fp16)[name = tensor("qkv_7_cast_fp16")]; + tensor q_19_begin_0 = const()[name = tensor("q_19_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor q_19_end_0 = const()[name = tensor("q_19_end_0"), val = tensor([1, 256, 1, 16, 64])]; + tensor q_19_end_mask_0 = const()[name = tensor("q_19_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor q_19_squeeze_mask_0 = const()[name = tensor("q_19_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor q_19_cast_fp16 = slice_by_index(begin = q_19_begin_0, end = q_19_end_0, end_mask = q_19_end_mask_0, squeeze_mask = q_19_squeeze_mask_0, x = qkv_7_cast_fp16)[name = tensor("q_19_cast_fp16")]; + tensor k_13_begin_0 = const()[name = tensor("k_13_begin_0"), val = tensor([0, 0, 1, 0, 0])]; + tensor k_13_end_0 = const()[name = tensor("k_13_end_0"), val = tensor([1, 256, 2, 16, 64])]; + tensor k_13_end_mask_0 = const()[name = tensor("k_13_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor k_13_squeeze_mask_0 = const()[name = tensor("k_13_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor k_13_cast_fp16 = slice_by_index(begin = k_13_begin_0, end = k_13_end_0, end_mask = k_13_end_mask_0, squeeze_mask = k_13_squeeze_mask_0, x = qkv_7_cast_fp16)[name = tensor("k_13_cast_fp16")]; + tensor v_7_begin_0 = const()[name = tensor("v_7_begin_0"), val = tensor([0, 0, 2, 0, 0])]; + tensor v_7_end_0 = const()[name = tensor("v_7_end_0"), val = tensor([1, 256, 3, 16, 64])]; + tensor v_7_end_mask_0 = const()[name = tensor("v_7_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor v_7_squeeze_mask_0 = const()[name = tensor("v_7_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor v_7_cast_fp16 = slice_by_index(begin = v_7_begin_0, end = v_7_end_0, end_mask = v_7_end_mask_0, squeeze_mask = v_7_squeeze_mask_0, x = qkv_7_cast_fp16)[name = tensor("v_7_cast_fp16")]; + tensor var_1353 = const()[name = tensor("op_1353"), val = tensor([1, 1])]; + tensor position3_to_fp16_dtype_0 = const()[name = tensor("position3_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor position3_to_fp16 = cast(dtype = position3_to_fp16_dtype_0, x = position3)[name = tensor("cast_142")]; + tensor var_1354_cast_fp16 = reshape(shape = var_1353, x = position3_to_fp16)[name = tensor("op_1354_cast_fp16")]; + tensor ts_21_cast_fp16 = add(x = ts_1_promoted_to_fp16, y = var_1354_cast_fp16)[name = tensor("ts_21_cast_fp16")]; + tensor var_1361 = const()[name = tensor("op_1361"), val = tensor([1, 256, 1, 1])]; + tensor ts_23_cast_fp16 = reshape(shape = var_1361, x = ts_21_cast_fp16)[name = tensor("ts_23_cast_fp16")]; + tensor var_1368 = const()[name = tensor("op_1368"), val = tensor([1, 256, 16, 32, 2])]; + tensor q_complex_7_cast_fp16 = reshape(shape = var_1368, x = q_19_cast_fp16)[name = tensor("q_complex_7_cast_fp16")]; + tensor var_1375 = const()[name = tensor("op_1375"), val = tensor([1, 256, 16, 32, 2])]; + tensor k_complex_7_cast_fp16 = reshape(shape = var_1375, x = k_13_cast_fp16)[name = tensor("k_complex_7_cast_fp16")]; + tensor var_1379_begin_0 = const()[name = tensor("op_1379_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_1379_end_0 = const()[name = tensor("op_1379_end_0"), val = tensor([1, 256, 16, 32, 1])]; + tensor var_1379_end_mask_0 = const()[name = tensor("op_1379_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_1379_squeeze_mask_0 = const()[name = tensor("op_1379_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_1379_cast_fp16 = slice_by_index(begin = var_1379_begin_0, end = var_1379_end_0, end_mask = var_1379_end_mask_0, squeeze_mask = var_1379_squeeze_mask_0, x = q_complex_7_cast_fp16)[name = tensor("op_1379_cast_fp16")]; + tensor var_1387_begin_0 = const()[name = tensor("op_1387_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_1387_end_0 = const()[name = tensor("op_1387_end_0"), val = tensor([1, 256, 16, 32, 2])]; + tensor var_1387_end_mask_0 = const()[name = tensor("op_1387_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_1387_squeeze_mask_0 = const()[name = tensor("op_1387_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_1387_cast_fp16 = slice_by_index(begin = var_1387_begin_0, end = var_1387_end_0, end_mask = var_1387_end_mask_0, squeeze_mask = var_1387_squeeze_mask_0, x = q_complex_7_cast_fp16)[name = tensor("op_1387_cast_fp16")]; + tensor var_1395_begin_0 = const()[name = tensor("op_1395_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_1395_end_0 = const()[name = tensor("op_1395_end_0"), val = tensor([1, 256, 16, 32, 1])]; + tensor var_1395_end_mask_0 = const()[name = tensor("op_1395_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_1395_squeeze_mask_0 = const()[name = tensor("op_1395_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_1395_cast_fp16 = slice_by_index(begin = var_1395_begin_0, end = var_1395_end_0, end_mask = var_1395_end_mask_0, squeeze_mask = var_1395_squeeze_mask_0, x = k_complex_7_cast_fp16)[name = tensor("op_1395_cast_fp16")]; + tensor var_1403_begin_0 = const()[name = tensor("op_1403_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_1403_end_0 = const()[name = tensor("op_1403_end_0"), val = tensor([1, 256, 16, 32, 2])]; + tensor var_1403_end_mask_0 = const()[name = tensor("op_1403_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_1403_squeeze_mask_0 = const()[name = tensor("op_1403_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_1403_cast_fp16 = slice_by_index(begin = var_1403_begin_0, end = var_1403_end_0, end_mask = var_1403_end_mask_0, squeeze_mask = var_1403_squeeze_mask_0, x = k_complex_7_cast_fp16)[name = tensor("op_1403_cast_fp16")]; + tensor freqs_7_to_fp16 = const()[name = tensor("freqs_7_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(82887808)))]; + tensor var_1409_cast_fp16 = mul(x = freqs_7_to_fp16, y = ts_23_cast_fp16)[name = tensor("op_1409_cast_fp16")]; + tensor rotr_7_cast_fp16 = cos(x = var_1409_cast_fp16)[name = tensor("rotr_7_cast_fp16")]; + tensor roti_7_cast_fp16 = sin(x = var_1409_cast_fp16)[name = tensor("roti_7_cast_fp16")]; + tensor var_1413_cast_fp16 = mul(x = var_1379_cast_fp16, y = rotr_7_cast_fp16)[name = tensor("op_1413_cast_fp16")]; + tensor var_1414_cast_fp16 = mul(x = var_1387_cast_fp16, y = roti_7_cast_fp16)[name = tensor("op_1414_cast_fp16")]; + tensor qor_13_cast_fp16 = sub(x = var_1413_cast_fp16, y = var_1414_cast_fp16)[name = tensor("qor_13_cast_fp16")]; + tensor var_1417_cast_fp16 = mul(x = var_1379_cast_fp16, y = roti_7_cast_fp16)[name = tensor("op_1417_cast_fp16")]; + tensor var_1418_cast_fp16 = mul(x = var_1387_cast_fp16, y = rotr_7_cast_fp16)[name = tensor("op_1418_cast_fp16")]; + tensor qoi_13_cast_fp16 = add(x = var_1417_cast_fp16, y = var_1418_cast_fp16)[name = tensor("qoi_13_cast_fp16")]; + tensor var_1421_cast_fp16 = mul(x = var_1395_cast_fp16, y = rotr_7_cast_fp16)[name = tensor("op_1421_cast_fp16")]; + tensor var_1422_cast_fp16 = mul(x = var_1403_cast_fp16, y = roti_7_cast_fp16)[name = tensor("op_1422_cast_fp16")]; + tensor kor_13_cast_fp16 = sub(x = var_1421_cast_fp16, y = var_1422_cast_fp16)[name = tensor("kor_13_cast_fp16")]; + tensor var_1425_cast_fp16 = mul(x = var_1395_cast_fp16, y = roti_7_cast_fp16)[name = tensor("op_1425_cast_fp16")]; + tensor var_1426_cast_fp16 = mul(x = var_1403_cast_fp16, y = rotr_7_cast_fp16)[name = tensor("op_1426_cast_fp16")]; + tensor koi_13_cast_fp16 = add(x = var_1425_cast_fp16, y = var_1426_cast_fp16)[name = tensor("koi_13_cast_fp16")]; + tensor qo_7_axis_0 = const()[name = tensor("qo_7_axis_0"), val = tensor(-1)]; + tensor qo_7_cast_fp16 = stack(axis = qo_7_axis_0, values = (qor_13_cast_fp16, qoi_13_cast_fp16))[name = tensor("qo_7_cast_fp16")]; + tensor ko_7_axis_0 = const()[name = tensor("ko_7_axis_0"), val = tensor(-1)]; + tensor ko_7_cast_fp16 = stack(axis = ko_7_axis_0, values = (kor_13_cast_fp16, koi_13_cast_fp16))[name = tensor("ko_7_cast_fp16")]; + tensor var_1459 = const()[name = tensor("op_1459"), val = tensor([1, 256, 16, 64])]; + tensor q_21_cast_fp16 = reshape(shape = var_1459, x = qo_7_cast_fp16)[name = tensor("q_21_cast_fp16")]; + tensor var_1465 = const()[name = tensor("op_1465"), val = tensor([1, 256, 16, 64])]; + tensor k_15_cast_fp16 = reshape(shape = var_1465, x = ko_7_cast_fp16)[name = tensor("k_15_cast_fp16")]; + tensor write_indices_float_19_cast_fp16 = add(x = var_1354_cast_fp16, y = var_249_promoted_to_fp16)[name = tensor("write_indices_float_19_cast_fp16")]; + tensor _inversed_1486_y_0_to_fp16 = const()[name = tensor("_inversed_1486_y_0_to_fp16"), val = tensor(0x1p-9)]; + tensor _inversed_1486_cast_fp16 = mul(x = write_indices_float_19_cast_fp16, y = _inversed_1486_y_0_to_fp16)[name = tensor("_inversed_1486_cast_fp16")]; + tensor var_1487_cast_fp16 = floor(x = _inversed_1486_cast_fp16)[name = tensor("op_1487_cast_fp16")]; + tensor var_1488_to_fp16 = const()[name = tensor("op_1488_to_fp16"), val = tensor(0x1p+9)]; + tensor var_1489_cast_fp16 = mul(x = var_1487_cast_fp16, y = var_1488_to_fp16)[name = tensor("op_1489_cast_fp16")]; + tensor write_indices_float_21_cast_fp16 = sub(x = write_indices_float_19_cast_fp16, y = var_1489_cast_fp16)[name = tensor("write_indices_float_21_cast_fp16")]; + tensor var_1507_value_0_to_fp16 = const()[name = tensor("op_1507_value_0_to_fp16"), val = tensor(0x1.ffp+8)]; + tensor var_1507_cast_fp16 = fill_like(ref_tensor = write_indices_float_21_cast_fp16, value = var_1507_value_0_to_fp16)[name = tensor("op_1507_cast_fp16")]; + tensor write_indices_float_23_cast_fp16 = select(a = write_indices_float_21_cast_fp16, b = var_1507_cast_fp16, cond = token_valid_1_cast_fp16)[name = tensor("write_indices_float_23_cast_fp16")]; + tensor var_1513_dtype_0 = const()[name = tensor("op_1513_dtype_0"), val = tensor("int32")]; + tensor var_1518 = const()[name = tensor("op_1518"), val = tensor([1, 256, 1, 1])]; + tensor write_indices_float_23_cast_fp16_to_int32 = cast(dtype = var_1513_dtype_0, x = write_indices_float_23_cast_fp16)[name = tensor("cast_141")]; + tensor var_1519 = reshape(shape = var_1518, x = write_indices_float_23_cast_fp16_to_int32)[name = tensor("op_1519")]; + tensor write_indices_7_reps_0 = const()[name = tensor("write_indices_7_reps_0"), val = tensor([1, 1, 16, 64])]; + tensor write_indices_7 = tile(reps = write_indices_7_reps_0, x = var_1519)[name = tensor("write_indices_7")]; + tensor var_1529_begin_0 = const()[name = tensor("op_1529_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_1529_end_0 = const()[name = tensor("op_1529_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor var_1529_end_mask_0 = const()[name = tensor("op_1529_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_1529_squeeze_mask_0 = const()[name = tensor("op_1529_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor cache3_to_fp16_dtype_0 = const()[name = tensor("cache3_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor cache3_to_fp16 = cast(dtype = cache3_to_fp16_dtype_0, x = cache3)[name = tensor("cast_140")]; + tensor var_1529_cast_fp16 = slice_by_index(begin = var_1529_begin_0, end = var_1529_end_0, end_mask = var_1529_end_mask_0, squeeze_mask = var_1529_squeeze_mask_0, x = cache3_to_fp16)[name = tensor("op_1529_cast_fp16")]; + tensor var_1531_axis_0 = const()[name = tensor("op_1531_axis_0"), val = tensor(1)]; + tensor var_1531_mode_0 = const()[name = tensor("op_1531_mode_0"), val = tensor("update")]; + tensor var_1531_validate_indices_0 = const()[name = tensor("op_1531_validate_indices_0"), val = tensor(false)]; + tensor var_1531_cast_fp16 = scatter_along_axis(axis = var_1531_axis_0, data = var_1529_cast_fp16, indices = write_indices_7, mode = var_1531_mode_0, updates = k_15_cast_fp16, validate_indices = var_1531_validate_indices_0)[name = tensor("op_1531_cast_fp16")]; + tensor concat_26 = const()[name = tensor("concat_26"), val = tensor([0, 0, 0, 0, 0])]; + tensor concat_27 = const()[name = tensor("concat_27"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_7_internal_tensor_assign_1_stride_0 = const()[name = tensor("new_cache_7_internal_tensor_assign_1_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_7_internal_tensor_assign_1_begin_mask_0 = const()[name = tensor("new_cache_7_internal_tensor_assign_1_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_7_internal_tensor_assign_1_end_mask_0 = const()[name = tensor("new_cache_7_internal_tensor_assign_1_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_7_internal_tensor_assign_1_squeeze_mask_0 = const()[name = tensor("new_cache_7_internal_tensor_assign_1_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_16 = const()[name = tensor("shape_16"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_6 = const()[name = tensor("reduce_prod_6"), val = tensor(1048576)]; + tensor range_1d_6_start_0 = const()[name = tensor("range_1d_6_start_0"), val = tensor(0)]; + tensor range_1d_6_step_0 = const()[name = tensor("range_1d_6_step_0"), val = tensor(1)]; + tensor range_1d_6 = range_1d(end = reduce_prod_6, start = range_1d_6_start_0, step = range_1d_6_step_0)[name = tensor("range_1d_6")]; + tensor reshape_30 = reshape(shape = shape_16, x = range_1d_6)[name = tensor("reshape_30")]; + tensor slice_by_index_6 = slice_by_index(begin = concat_26, begin_mask = new_cache_7_internal_tensor_assign_1_begin_mask_0, end = concat_27, end_mask = new_cache_7_internal_tensor_assign_1_end_mask_0, squeeze_mask = new_cache_7_internal_tensor_assign_1_squeeze_mask_0, stride = new_cache_7_internal_tensor_assign_1_stride_0, x = reshape_30)[name = tensor("slice_by_index_6")]; + tensor reshape_31_shape_0 = const()[name = tensor("reshape_31_shape_0"), val = tensor([-1])]; + tensor reshape_31 = reshape(shape = reshape_31_shape_0, x = slice_by_index_6)[name = tensor("reshape_31")]; + tensor reshape_32_shape_0 = const()[name = tensor("reshape_32_shape_0"), val = tensor([-1])]; + tensor reshape_32_cast_fp16 = reshape(shape = reshape_32_shape_0, x = var_1531_cast_fp16)[name = tensor("reshape_32_cast_fp16")]; + tensor reshape_33_shape_0 = const()[name = tensor("reshape_33_shape_0"), val = tensor([-1])]; + tensor reshape_33_cast_fp16 = reshape(shape = reshape_33_shape_0, x = cache3_to_fp16)[name = tensor("reshape_33_cast_fp16")]; + tensor scatter_6_mode_0 = const()[name = tensor("scatter_6_mode_0"), val = tensor("update")]; + tensor scatter_6_axis_0 = const()[name = tensor("scatter_6_axis_0"), val = tensor(0)]; + tensor scatter_6_validate_indices_0 = const()[name = tensor("scatter_6_validate_indices_0"), val = tensor(false)]; + tensor scatter_6_cast_fp16 = scatter(axis = scatter_6_axis_0, data = reshape_33_cast_fp16, indices = reshape_31, mode = scatter_6_mode_0, updates = reshape_32_cast_fp16, validate_indices = scatter_6_validate_indices_0)[name = tensor("scatter_6_cast_fp16")]; + tensor reshape_34_cast_fp16 = reshape(shape = shape_16, x = scatter_6_cast_fp16)[name = tensor("reshape_34_cast_fp16")]; + tensor var_1539_begin_0 = const()[name = tensor("op_1539_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor var_1539_end_0 = const()[name = tensor("op_1539_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor var_1539_end_mask_0 = const()[name = tensor("op_1539_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_1539_squeeze_mask_0 = const()[name = tensor("op_1539_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor var_1539_cast_fp16 = slice_by_index(begin = var_1539_begin_0, end = var_1539_end_0, end_mask = var_1539_end_mask_0, squeeze_mask = var_1539_squeeze_mask_0, x = reshape_34_cast_fp16)[name = tensor("op_1539_cast_fp16")]; + tensor var_1541_axis_0 = const()[name = tensor("op_1541_axis_0"), val = tensor(1)]; + tensor var_1541_mode_0 = const()[name = tensor("op_1541_mode_0"), val = tensor("update")]; + tensor var_1541_validate_indices_0 = const()[name = tensor("op_1541_validate_indices_0"), val = tensor(false)]; + tensor var_1541_cast_fp16 = scatter_along_axis(axis = var_1541_axis_0, data = var_1539_cast_fp16, indices = write_indices_7, mode = var_1541_mode_0, updates = v_7_cast_fp16, validate_indices = var_1541_validate_indices_0)[name = tensor("op_1541_cast_fp16")]; + tensor concat_28 = const()[name = tensor("concat_28"), val = tensor([1, 0, 0, 0, 0])]; + tensor concat_29 = const()[name = tensor("concat_29"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_7_internal_tensor_assign_2_stride_0 = const()[name = tensor("new_cache_7_internal_tensor_assign_2_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_7_internal_tensor_assign_2_begin_mask_0 = const()[name = tensor("new_cache_7_internal_tensor_assign_2_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_7_internal_tensor_assign_2_end_mask_0 = const()[name = tensor("new_cache_7_internal_tensor_assign_2_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_7_internal_tensor_assign_2_squeeze_mask_0 = const()[name = tensor("new_cache_7_internal_tensor_assign_2_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_17 = const()[name = tensor("shape_17"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_7 = const()[name = tensor("reduce_prod_7"), val = tensor(1048576)]; + tensor range_1d_7_start_0 = const()[name = tensor("range_1d_7_start_0"), val = tensor(0)]; + tensor range_1d_7_step_0 = const()[name = tensor("range_1d_7_step_0"), val = tensor(1)]; + tensor range_1d_7 = range_1d(end = reduce_prod_7, start = range_1d_7_start_0, step = range_1d_7_step_0)[name = tensor("range_1d_7")]; + tensor reshape_35 = reshape(shape = shape_17, x = range_1d_7)[name = tensor("reshape_35")]; + tensor slice_by_index_7 = slice_by_index(begin = concat_28, begin_mask = new_cache_7_internal_tensor_assign_2_begin_mask_0, end = concat_29, end_mask = new_cache_7_internal_tensor_assign_2_end_mask_0, squeeze_mask = new_cache_7_internal_tensor_assign_2_squeeze_mask_0, stride = new_cache_7_internal_tensor_assign_2_stride_0, x = reshape_35)[name = tensor("slice_by_index_7")]; + tensor reshape_36_shape_0 = const()[name = tensor("reshape_36_shape_0"), val = tensor([-1])]; + tensor reshape_36 = reshape(shape = reshape_36_shape_0, x = slice_by_index_7)[name = tensor("reshape_36")]; + tensor reshape_37_shape_0 = const()[name = tensor("reshape_37_shape_0"), val = tensor([-1])]; + tensor reshape_37_cast_fp16 = reshape(shape = reshape_37_shape_0, x = var_1541_cast_fp16)[name = tensor("reshape_37_cast_fp16")]; + tensor reshape_38_shape_0 = const()[name = tensor("reshape_38_shape_0"), val = tensor([-1])]; + tensor reshape_38_cast_fp16 = reshape(shape = reshape_38_shape_0, x = reshape_34_cast_fp16)[name = tensor("reshape_38_cast_fp16")]; + tensor scatter_7_mode_0 = const()[name = tensor("scatter_7_mode_0"), val = tensor("update")]; + tensor scatter_7_axis_0 = const()[name = tensor("scatter_7_axis_0"), val = tensor(0)]; + tensor scatter_7_validate_indices_0 = const()[name = tensor("scatter_7_validate_indices_0"), val = tensor(false)]; + tensor scatter_7_cast_fp16 = scatter(axis = scatter_7_axis_0, data = reshape_38_cast_fp16, indices = reshape_36, mode = scatter_7_mode_0, updates = reshape_37_cast_fp16, validate_indices = scatter_7_validate_indices_0)[name = tensor("scatter_7_cast_fp16")]; + tensor reshape_39_cast_fp16 = reshape(shape = shape_17, x = scatter_7_cast_fp16)[name = tensor("reshape_39_cast_fp16")]; + tensor reshape_39_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("reshape_39_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor keys_19_begin_0 = const()[name = tensor("keys_19_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor keys_19_end_0 = const()[name = tensor("keys_19_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor keys_19_end_mask_0 = const()[name = tensor("keys_19_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor keys_19_squeeze_mask_0 = const()[name = tensor("keys_19_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor keys_19_cast_fp16 = slice_by_index(begin = keys_19_begin_0, end = keys_19_end_0, end_mask = keys_19_end_mask_0, squeeze_mask = keys_19_squeeze_mask_0, x = reshape_39_cast_fp16)[name = tensor("keys_19_cast_fp16")]; + tensor values_19_begin_0 = const()[name = tensor("values_19_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor values_19_end_0 = const()[name = tensor("values_19_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor values_19_end_mask_0 = const()[name = tensor("values_19_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor values_19_squeeze_mask_0 = const()[name = tensor("values_19_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor values_19_cast_fp16 = slice_by_index(begin = values_19_begin_0, end = values_19_end_0, end_mask = values_19_end_mask_0, squeeze_mask = values_19_squeeze_mask_0, x = reshape_39_cast_fp16)[name = tensor("values_19_cast_fp16")]; + tensor var_1553_cast_fp16 = not_equal(x = keys_19_cast_fp16, y = keys_19_cast_fp16)[name = tensor("op_1553_cast_fp16")]; + tensor keys_21_cast_fp16 = select(a = var_326_to_fp16, b = keys_19_cast_fp16, cond = var_1553_cast_fp16)[name = tensor("keys_21_cast_fp16")]; + tensor var_1561_cast_fp16 = not_equal(x = values_19_cast_fp16, y = values_19_cast_fp16)[name = tensor("op_1561_cast_fp16")]; + tensor values_21_cast_fp16 = select(a = var_326_to_fp16, b = values_19_cast_fp16, cond = var_1561_cast_fp16)[name = tensor("values_21_cast_fp16")]; + tensor var_1585 = const()[name = tensor("op_1585"), val = tensor([0, 2, 1, 3])]; + tensor var_1601 = const()[name = tensor("op_1601"), val = tensor([1, 1, 1])]; + tensor var_1602_cast_fp16 = reshape(shape = var_1601, x = position3_to_fp16)[name = tensor("op_1602_cast_fp16")]; + tensor q_positions_7_cast_fp16 = add(x = var_1602_cast_fp16, y = q_offsets_1_promoted_to_fp16)[name = tensor("q_positions_7_cast_fp16")]; + tensor valid_end_7_cast_fp16 = add(x = var_1602_cast_fp16, y = var_392_cast_fp16)[name = tensor("valid_end_7_cast_fp16")]; + tensor valid_mask_7_cast_fp16 = less(x = k_positions_1_promoted_to_fp16, y = valid_end_7_cast_fp16)[name = tensor("valid_mask_7_cast_fp16")]; + tensor causal_mask_7_cast_fp16 = less_equal(x = k_positions_1_promoted_to_fp16, y = q_positions_7_cast_fp16)[name = tensor("causal_mask_7_cast_fp16")]; + tensor var_1630 = logical_and(x = valid_mask_7_cast_fp16, y = causal_mask_7_cast_fp16)[name = tensor("op_1630")]; + tensor attn_mask_7_axes_0 = const()[name = tensor("attn_mask_7_axes_0"), val = tensor([1])]; + tensor attn_mask_7 = expand_dims(axes = attn_mask_7_axes_0, x = var_1630)[name = tensor("attn_mask_7")]; + tensor var_1644_transpose_x_0 = const()[name = tensor("op_1644_transpose_x_0"), val = tensor(false)]; + tensor var_1644_transpose_y_0 = const()[name = tensor("op_1644_transpose_y_0"), val = tensor(false)]; + tensor transpose_21_perm_0 = const()[name = tensor("transpose_21_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_22_perm_0 = const()[name = tensor("transpose_22_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_22 = transpose(perm = transpose_22_perm_0, x = keys_21_cast_fp16)[name = tensor("transpose_30")]; + tensor transpose_21 = transpose(perm = transpose_21_perm_0, x = q_21_cast_fp16)[name = tensor("transpose_31")]; + tensor var_1644_cast_fp16 = matmul(transpose_x = var_1644_transpose_x_0, transpose_y = var_1644_transpose_y_0, x = transpose_21, y = transpose_22)[name = tensor("op_1644_cast_fp16")]; + tensor var_1638_to_fp16 = const()[name = tensor("op_1638_to_fp16"), val = tensor([0x1p-3])]; + tensor attn_weights_19_cast_fp16 = mul(x = var_1644_cast_fp16, y = var_1638_to_fp16)[name = tensor("attn_weights_19_cast_fp16")]; + tensor var_1646 = logical_not(x = attn_mask_7)[name = tensor("op_1646")]; + tensor var_1647_to_fp16 = const()[name = tensor("op_1647_to_fp16"), val = tensor(-inf)]; + tensor attn_weights_21_cast_fp16 = select(a = var_1647_to_fp16, b = attn_weights_19_cast_fp16, cond = var_1646)[name = tensor("attn_weights_21_cast_fp16")]; + tensor var_1649 = const()[name = tensor("op_1649"), val = tensor(-1)]; + tensor attn_weights_23_cast_fp16 = softmax(axis = var_1649, x = attn_weights_21_cast_fp16)[name = tensor("attn_weights_23_cast_fp16")]; + tensor attn_output_7_transpose_x_0 = const()[name = tensor("attn_output_7_transpose_x_0"), val = tensor(false)]; + tensor attn_output_7_transpose_y_0 = const()[name = tensor("attn_output_7_transpose_y_0"), val = tensor(false)]; + tensor values_23_cast_fp16 = transpose(perm = var_1585, x = values_21_cast_fp16)[name = tensor("transpose_32")]; + tensor attn_output_7_cast_fp16 = matmul(transpose_x = attn_output_7_transpose_x_0, transpose_y = attn_output_7_transpose_y_0, x = attn_weights_23_cast_fp16, y = values_23_cast_fp16)[name = tensor("attn_output_7_cast_fp16")]; + tensor var_1657 = const()[name = tensor("op_1657"), val = tensor([0, 2, 1, 3])]; + tensor var_1662 = const()[name = tensor("op_1662"), val = tensor([1, 256, 1024])]; + tensor var_1658_cast_fp16 = transpose(perm = var_1657, x = attn_output_7_cast_fp16)[name = tensor("transpose_29")]; + tensor input_41_cast_fp16 = reshape(shape = var_1662, x = var_1658_cast_fp16)[name = tensor("input_41_cast_fp16")]; + tensor attn3_out_proj_weight_to_fp16 = const()[name = tensor("attn3_out_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(82887936)))]; + tensor linear_13_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = attn3_out_proj_weight_to_fp16, x = input_41_cast_fp16)[name = tensor("linear_13_cast_fp16")]; + tensor var_1671_cast_fp16 = add(x = position3_to_fp16, y = valid_len_to_fp16)[name = tensor("op_1671_cast_fp16")]; + tensor var_1671_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_1671_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor input_43_cast_fp16 = add(x = input_37_cast_fp16, y = linear_13_cast_fp16)[name = tensor("input_43_cast_fp16")]; + tensor input_45_axes_0 = const()[name = tensor("input_45_axes_0"), val = tensor([-1])]; + tensor norm3_2_weight_to_fp16 = const()[name = tensor("norm3_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(84985152)))]; + tensor norm3_2_bias_to_fp16 = const()[name = tensor("norm3_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(84987264)))]; + tensor var_1675_to_fp16 = const()[name = tensor("op_1675_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_45_cast_fp16 = layer_norm(axes = input_45_axes_0, beta = norm3_2_bias_to_fp16, epsilon = var_1675_to_fp16, gamma = norm3_2_weight_to_fp16, x = input_43_cast_fp16)[name = tensor("input_45_cast_fp16")]; + tensor linear3_1_weight_to_fp16 = const()[name = tensor("linear3_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(84989376)))]; + tensor linear_14_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = linear3_1_weight_to_fp16, x = input_45_cast_fp16)[name = tensor("linear_14_cast_fp16")]; + tensor input_47_mode_0 = const()[name = tensor("input_47_mode_0"), val = tensor("EXACT")]; + tensor input_47_cast_fp16 = gelu(mode = input_47_mode_0, x = linear_14_cast_fp16)[name = tensor("input_47_cast_fp16")]; + tensor linear3_2_weight_to_fp16 = const()[name = tensor("linear3_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(93378048)))]; + tensor linear_15_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = linear3_2_weight_to_fp16, x = input_47_cast_fp16)[name = tensor("linear_15_cast_fp16")]; + tensor input_49_cast_fp16 = add(x = input_43_cast_fp16, y = linear_15_cast_fp16)[name = tensor("input_49_cast_fp16")]; + tensor input_51_axes_0 = const()[name = tensor("input_51_axes_0"), val = tensor([-1])]; + tensor norm4_1_weight_to_fp16 = const()[name = tensor("norm4_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(101766720)))]; + tensor norm4_1_bias_to_fp16 = const()[name = tensor("norm4_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(101768832)))]; + tensor var_1692_to_fp16 = const()[name = tensor("op_1692_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_51_cast_fp16 = layer_norm(axes = input_51_axes_0, beta = norm4_1_bias_to_fp16, epsilon = var_1692_to_fp16, gamma = norm4_1_weight_to_fp16, x = input_49_cast_fp16)[name = tensor("input_51_cast_fp16")]; + tensor attn4_in_proj_weight_to_fp16 = const()[name = tensor("attn4_in_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(101770944)))]; + tensor linear_16_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = attn4_in_proj_weight_to_fp16, x = input_51_cast_fp16)[name = tensor("linear_16_cast_fp16")]; + tensor var_1706 = const()[name = tensor("op_1706"), val = tensor([1, 256, 3, 16, 64])]; + tensor qkv_9_cast_fp16 = reshape(shape = var_1706, x = linear_16_cast_fp16)[name = tensor("qkv_9_cast_fp16")]; + tensor q_25_begin_0 = const()[name = tensor("q_25_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor q_25_end_0 = const()[name = tensor("q_25_end_0"), val = tensor([1, 256, 1, 16, 64])]; + tensor q_25_end_mask_0 = const()[name = tensor("q_25_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor q_25_squeeze_mask_0 = const()[name = tensor("q_25_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor q_25_cast_fp16 = slice_by_index(begin = q_25_begin_0, end = q_25_end_0, end_mask = q_25_end_mask_0, squeeze_mask = q_25_squeeze_mask_0, x = qkv_9_cast_fp16)[name = tensor("q_25_cast_fp16")]; + tensor k_17_begin_0 = const()[name = tensor("k_17_begin_0"), val = tensor([0, 0, 1, 0, 0])]; + tensor k_17_end_0 = const()[name = tensor("k_17_end_0"), val = tensor([1, 256, 2, 16, 64])]; + tensor k_17_end_mask_0 = const()[name = tensor("k_17_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor k_17_squeeze_mask_0 = const()[name = tensor("k_17_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor k_17_cast_fp16 = slice_by_index(begin = k_17_begin_0, end = k_17_end_0, end_mask = k_17_end_mask_0, squeeze_mask = k_17_squeeze_mask_0, x = qkv_9_cast_fp16)[name = tensor("k_17_cast_fp16")]; + tensor v_9_begin_0 = const()[name = tensor("v_9_begin_0"), val = tensor([0, 0, 2, 0, 0])]; + tensor v_9_end_0 = const()[name = tensor("v_9_end_0"), val = tensor([1, 256, 3, 16, 64])]; + tensor v_9_end_mask_0 = const()[name = tensor("v_9_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor v_9_squeeze_mask_0 = const()[name = tensor("v_9_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor v_9_cast_fp16 = slice_by_index(begin = v_9_begin_0, end = v_9_end_0, end_mask = v_9_end_mask_0, squeeze_mask = v_9_squeeze_mask_0, x = qkv_9_cast_fp16)[name = tensor("v_9_cast_fp16")]; + tensor var_1764 = const()[name = tensor("op_1764"), val = tensor([1, 1])]; + tensor position4_to_fp16_dtype_0 = const()[name = tensor("position4_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor position4_to_fp16 = cast(dtype = position4_to_fp16_dtype_0, x = position4)[name = tensor("cast_137")]; + tensor var_1765_cast_fp16 = reshape(shape = var_1764, x = position4_to_fp16)[name = tensor("op_1765_cast_fp16")]; + tensor ts_27_cast_fp16 = add(x = ts_1_promoted_to_fp16, y = var_1765_cast_fp16)[name = tensor("ts_27_cast_fp16")]; + tensor var_1772 = const()[name = tensor("op_1772"), val = tensor([1, 256, 1, 1])]; + tensor ts_29_cast_fp16 = reshape(shape = var_1772, x = ts_27_cast_fp16)[name = tensor("ts_29_cast_fp16")]; + tensor var_1779 = const()[name = tensor("op_1779"), val = tensor([1, 256, 16, 32, 2])]; + tensor q_complex_9_cast_fp16 = reshape(shape = var_1779, x = q_25_cast_fp16)[name = tensor("q_complex_9_cast_fp16")]; + tensor var_1786 = const()[name = tensor("op_1786"), val = tensor([1, 256, 16, 32, 2])]; + tensor k_complex_9_cast_fp16 = reshape(shape = var_1786, x = k_17_cast_fp16)[name = tensor("k_complex_9_cast_fp16")]; + tensor var_1790_begin_0 = const()[name = tensor("op_1790_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_1790_end_0 = const()[name = tensor("op_1790_end_0"), val = tensor([1, 256, 16, 32, 1])]; + tensor var_1790_end_mask_0 = const()[name = tensor("op_1790_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_1790_squeeze_mask_0 = const()[name = tensor("op_1790_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_1790_cast_fp16 = slice_by_index(begin = var_1790_begin_0, end = var_1790_end_0, end_mask = var_1790_end_mask_0, squeeze_mask = var_1790_squeeze_mask_0, x = q_complex_9_cast_fp16)[name = tensor("op_1790_cast_fp16")]; + tensor var_1798_begin_0 = const()[name = tensor("op_1798_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_1798_end_0 = const()[name = tensor("op_1798_end_0"), val = tensor([1, 256, 16, 32, 2])]; + tensor var_1798_end_mask_0 = const()[name = tensor("op_1798_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_1798_squeeze_mask_0 = const()[name = tensor("op_1798_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_1798_cast_fp16 = slice_by_index(begin = var_1798_begin_0, end = var_1798_end_0, end_mask = var_1798_end_mask_0, squeeze_mask = var_1798_squeeze_mask_0, x = q_complex_9_cast_fp16)[name = tensor("op_1798_cast_fp16")]; + tensor var_1806_begin_0 = const()[name = tensor("op_1806_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_1806_end_0 = const()[name = tensor("op_1806_end_0"), val = tensor([1, 256, 16, 32, 1])]; + tensor var_1806_end_mask_0 = const()[name = tensor("op_1806_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_1806_squeeze_mask_0 = const()[name = tensor("op_1806_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_1806_cast_fp16 = slice_by_index(begin = var_1806_begin_0, end = var_1806_end_0, end_mask = var_1806_end_mask_0, squeeze_mask = var_1806_squeeze_mask_0, x = k_complex_9_cast_fp16)[name = tensor("op_1806_cast_fp16")]; + tensor var_1814_begin_0 = const()[name = tensor("op_1814_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_1814_end_0 = const()[name = tensor("op_1814_end_0"), val = tensor([1, 256, 16, 32, 2])]; + tensor var_1814_end_mask_0 = const()[name = tensor("op_1814_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_1814_squeeze_mask_0 = const()[name = tensor("op_1814_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_1814_cast_fp16 = slice_by_index(begin = var_1814_begin_0, end = var_1814_end_0, end_mask = var_1814_end_mask_0, squeeze_mask = var_1814_squeeze_mask_0, x = k_complex_9_cast_fp16)[name = tensor("op_1814_cast_fp16")]; + tensor freqs_9_to_fp16 = const()[name = tensor("freqs_9_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(108062464)))]; + tensor var_1820_cast_fp16 = mul(x = freqs_9_to_fp16, y = ts_29_cast_fp16)[name = tensor("op_1820_cast_fp16")]; + tensor rotr_9_cast_fp16 = cos(x = var_1820_cast_fp16)[name = tensor("rotr_9_cast_fp16")]; + tensor roti_9_cast_fp16 = sin(x = var_1820_cast_fp16)[name = tensor("roti_9_cast_fp16")]; + tensor var_1824_cast_fp16 = mul(x = var_1790_cast_fp16, y = rotr_9_cast_fp16)[name = tensor("op_1824_cast_fp16")]; + tensor var_1825_cast_fp16 = mul(x = var_1798_cast_fp16, y = roti_9_cast_fp16)[name = tensor("op_1825_cast_fp16")]; + tensor qor_17_cast_fp16 = sub(x = var_1824_cast_fp16, y = var_1825_cast_fp16)[name = tensor("qor_17_cast_fp16")]; + tensor var_1828_cast_fp16 = mul(x = var_1790_cast_fp16, y = roti_9_cast_fp16)[name = tensor("op_1828_cast_fp16")]; + tensor var_1829_cast_fp16 = mul(x = var_1798_cast_fp16, y = rotr_9_cast_fp16)[name = tensor("op_1829_cast_fp16")]; + tensor qoi_17_cast_fp16 = add(x = var_1828_cast_fp16, y = var_1829_cast_fp16)[name = tensor("qoi_17_cast_fp16")]; + tensor var_1832_cast_fp16 = mul(x = var_1806_cast_fp16, y = rotr_9_cast_fp16)[name = tensor("op_1832_cast_fp16")]; + tensor var_1833_cast_fp16 = mul(x = var_1814_cast_fp16, y = roti_9_cast_fp16)[name = tensor("op_1833_cast_fp16")]; + tensor kor_17_cast_fp16 = sub(x = var_1832_cast_fp16, y = var_1833_cast_fp16)[name = tensor("kor_17_cast_fp16")]; + tensor var_1836_cast_fp16 = mul(x = var_1806_cast_fp16, y = roti_9_cast_fp16)[name = tensor("op_1836_cast_fp16")]; + tensor var_1837_cast_fp16 = mul(x = var_1814_cast_fp16, y = rotr_9_cast_fp16)[name = tensor("op_1837_cast_fp16")]; + tensor koi_17_cast_fp16 = add(x = var_1836_cast_fp16, y = var_1837_cast_fp16)[name = tensor("koi_17_cast_fp16")]; + tensor qo_9_axis_0 = const()[name = tensor("qo_9_axis_0"), val = tensor(-1)]; + tensor qo_9_cast_fp16 = stack(axis = qo_9_axis_0, values = (qor_17_cast_fp16, qoi_17_cast_fp16))[name = tensor("qo_9_cast_fp16")]; + tensor ko_9_axis_0 = const()[name = tensor("ko_9_axis_0"), val = tensor(-1)]; + tensor ko_9_cast_fp16 = stack(axis = ko_9_axis_0, values = (kor_17_cast_fp16, koi_17_cast_fp16))[name = tensor("ko_9_cast_fp16")]; + tensor var_1870 = const()[name = tensor("op_1870"), val = tensor([1, 256, 16, 64])]; + tensor q_27_cast_fp16 = reshape(shape = var_1870, x = qo_9_cast_fp16)[name = tensor("q_27_cast_fp16")]; + tensor var_1876 = const()[name = tensor("op_1876"), val = tensor([1, 256, 16, 64])]; + tensor k_19_cast_fp16 = reshape(shape = var_1876, x = ko_9_cast_fp16)[name = tensor("k_19_cast_fp16")]; + tensor write_indices_float_25_cast_fp16 = add(x = var_1765_cast_fp16, y = var_249_promoted_to_fp16)[name = tensor("write_indices_float_25_cast_fp16")]; + tensor _inversed_1897_y_0_to_fp16 = const()[name = tensor("_inversed_1897_y_0_to_fp16"), val = tensor(0x1p-9)]; + tensor _inversed_1897_cast_fp16 = mul(x = write_indices_float_25_cast_fp16, y = _inversed_1897_y_0_to_fp16)[name = tensor("_inversed_1897_cast_fp16")]; + tensor var_1898_cast_fp16 = floor(x = _inversed_1897_cast_fp16)[name = tensor("op_1898_cast_fp16")]; + tensor var_1899_to_fp16 = const()[name = tensor("op_1899_to_fp16"), val = tensor(0x1p+9)]; + tensor var_1900_cast_fp16 = mul(x = var_1898_cast_fp16, y = var_1899_to_fp16)[name = tensor("op_1900_cast_fp16")]; + tensor write_indices_float_27_cast_fp16 = sub(x = write_indices_float_25_cast_fp16, y = var_1900_cast_fp16)[name = tensor("write_indices_float_27_cast_fp16")]; + tensor var_1918_value_0_to_fp16 = const()[name = tensor("op_1918_value_0_to_fp16"), val = tensor(0x1.ffp+8)]; + tensor var_1918_cast_fp16 = fill_like(ref_tensor = write_indices_float_27_cast_fp16, value = var_1918_value_0_to_fp16)[name = tensor("op_1918_cast_fp16")]; + tensor write_indices_float_29_cast_fp16 = select(a = write_indices_float_27_cast_fp16, b = var_1918_cast_fp16, cond = token_valid_1_cast_fp16)[name = tensor("write_indices_float_29_cast_fp16")]; + tensor var_1924_dtype_0 = const()[name = tensor("op_1924_dtype_0"), val = tensor("int32")]; + tensor var_1929 = const()[name = tensor("op_1929"), val = tensor([1, 256, 1, 1])]; + tensor write_indices_float_29_cast_fp16_to_int32 = cast(dtype = var_1924_dtype_0, x = write_indices_float_29_cast_fp16)[name = tensor("cast_136")]; + tensor var_1930 = reshape(shape = var_1929, x = write_indices_float_29_cast_fp16_to_int32)[name = tensor("op_1930")]; + tensor write_indices_9_reps_0 = const()[name = tensor("write_indices_9_reps_0"), val = tensor([1, 1, 16, 64])]; + tensor write_indices_9 = tile(reps = write_indices_9_reps_0, x = var_1930)[name = tensor("write_indices_9")]; + tensor var_1940_begin_0 = const()[name = tensor("op_1940_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_1940_end_0 = const()[name = tensor("op_1940_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor var_1940_end_mask_0 = const()[name = tensor("op_1940_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_1940_squeeze_mask_0 = const()[name = tensor("op_1940_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor cache4_to_fp16_dtype_0 = const()[name = tensor("cache4_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor cache4_to_fp16 = cast(dtype = cache4_to_fp16_dtype_0, x = cache4)[name = tensor("cast_135")]; + tensor var_1940_cast_fp16 = slice_by_index(begin = var_1940_begin_0, end = var_1940_end_0, end_mask = var_1940_end_mask_0, squeeze_mask = var_1940_squeeze_mask_0, x = cache4_to_fp16)[name = tensor("op_1940_cast_fp16")]; + tensor var_1942_axis_0 = const()[name = tensor("op_1942_axis_0"), val = tensor(1)]; + tensor var_1942_mode_0 = const()[name = tensor("op_1942_mode_0"), val = tensor("update")]; + tensor var_1942_validate_indices_0 = const()[name = tensor("op_1942_validate_indices_0"), val = tensor(false)]; + tensor var_1942_cast_fp16 = scatter_along_axis(axis = var_1942_axis_0, data = var_1940_cast_fp16, indices = write_indices_9, mode = var_1942_mode_0, updates = k_19_cast_fp16, validate_indices = var_1942_validate_indices_0)[name = tensor("op_1942_cast_fp16")]; + tensor concat_34 = const()[name = tensor("concat_34"), val = tensor([0, 0, 0, 0, 0])]; + tensor concat_35 = const()[name = tensor("concat_35"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_9_internal_tensor_assign_1_stride_0 = const()[name = tensor("new_cache_9_internal_tensor_assign_1_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_9_internal_tensor_assign_1_begin_mask_0 = const()[name = tensor("new_cache_9_internal_tensor_assign_1_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_9_internal_tensor_assign_1_end_mask_0 = const()[name = tensor("new_cache_9_internal_tensor_assign_1_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_9_internal_tensor_assign_1_squeeze_mask_0 = const()[name = tensor("new_cache_9_internal_tensor_assign_1_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_18 = const()[name = tensor("shape_18"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_8 = const()[name = tensor("reduce_prod_8"), val = tensor(1048576)]; + tensor range_1d_8_start_0 = const()[name = tensor("range_1d_8_start_0"), val = tensor(0)]; + tensor range_1d_8_step_0 = const()[name = tensor("range_1d_8_step_0"), val = tensor(1)]; + tensor range_1d_8 = range_1d(end = reduce_prod_8, start = range_1d_8_start_0, step = range_1d_8_step_0)[name = tensor("range_1d_8")]; + tensor reshape_40 = reshape(shape = shape_18, x = range_1d_8)[name = tensor("reshape_40")]; + tensor slice_by_index_8 = slice_by_index(begin = concat_34, begin_mask = new_cache_9_internal_tensor_assign_1_begin_mask_0, end = concat_35, end_mask = new_cache_9_internal_tensor_assign_1_end_mask_0, squeeze_mask = new_cache_9_internal_tensor_assign_1_squeeze_mask_0, stride = new_cache_9_internal_tensor_assign_1_stride_0, x = reshape_40)[name = tensor("slice_by_index_8")]; + tensor reshape_41_shape_0 = const()[name = tensor("reshape_41_shape_0"), val = tensor([-1])]; + tensor reshape_41 = reshape(shape = reshape_41_shape_0, x = slice_by_index_8)[name = tensor("reshape_41")]; + tensor reshape_42_shape_0 = const()[name = tensor("reshape_42_shape_0"), val = tensor([-1])]; + tensor reshape_42_cast_fp16 = reshape(shape = reshape_42_shape_0, x = var_1942_cast_fp16)[name = tensor("reshape_42_cast_fp16")]; + tensor reshape_43_shape_0 = const()[name = tensor("reshape_43_shape_0"), val = tensor([-1])]; + tensor reshape_43_cast_fp16 = reshape(shape = reshape_43_shape_0, x = cache4_to_fp16)[name = tensor("reshape_43_cast_fp16")]; + tensor scatter_8_mode_0 = const()[name = tensor("scatter_8_mode_0"), val = tensor("update")]; + tensor scatter_8_axis_0 = const()[name = tensor("scatter_8_axis_0"), val = tensor(0)]; + tensor scatter_8_validate_indices_0 = const()[name = tensor("scatter_8_validate_indices_0"), val = tensor(false)]; + tensor scatter_8_cast_fp16 = scatter(axis = scatter_8_axis_0, data = reshape_43_cast_fp16, indices = reshape_41, mode = scatter_8_mode_0, updates = reshape_42_cast_fp16, validate_indices = scatter_8_validate_indices_0)[name = tensor("scatter_8_cast_fp16")]; + tensor reshape_44_cast_fp16 = reshape(shape = shape_18, x = scatter_8_cast_fp16)[name = tensor("reshape_44_cast_fp16")]; + tensor var_1950_begin_0 = const()[name = tensor("op_1950_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor var_1950_end_0 = const()[name = tensor("op_1950_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor var_1950_end_mask_0 = const()[name = tensor("op_1950_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_1950_squeeze_mask_0 = const()[name = tensor("op_1950_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor var_1950_cast_fp16 = slice_by_index(begin = var_1950_begin_0, end = var_1950_end_0, end_mask = var_1950_end_mask_0, squeeze_mask = var_1950_squeeze_mask_0, x = reshape_44_cast_fp16)[name = tensor("op_1950_cast_fp16")]; + tensor var_1952_axis_0 = const()[name = tensor("op_1952_axis_0"), val = tensor(1)]; + tensor var_1952_mode_0 = const()[name = tensor("op_1952_mode_0"), val = tensor("update")]; + tensor var_1952_validate_indices_0 = const()[name = tensor("op_1952_validate_indices_0"), val = tensor(false)]; + tensor var_1952_cast_fp16 = scatter_along_axis(axis = var_1952_axis_0, data = var_1950_cast_fp16, indices = write_indices_9, mode = var_1952_mode_0, updates = v_9_cast_fp16, validate_indices = var_1952_validate_indices_0)[name = tensor("op_1952_cast_fp16")]; + tensor concat_36 = const()[name = tensor("concat_36"), val = tensor([1, 0, 0, 0, 0])]; + tensor concat_37 = const()[name = tensor("concat_37"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_9_internal_tensor_assign_2_stride_0 = const()[name = tensor("new_cache_9_internal_tensor_assign_2_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_9_internal_tensor_assign_2_begin_mask_0 = const()[name = tensor("new_cache_9_internal_tensor_assign_2_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_9_internal_tensor_assign_2_end_mask_0 = const()[name = tensor("new_cache_9_internal_tensor_assign_2_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_9_internal_tensor_assign_2_squeeze_mask_0 = const()[name = tensor("new_cache_9_internal_tensor_assign_2_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_19 = const()[name = tensor("shape_19"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_9 = const()[name = tensor("reduce_prod_9"), val = tensor(1048576)]; + tensor range_1d_9_start_0 = const()[name = tensor("range_1d_9_start_0"), val = tensor(0)]; + tensor range_1d_9_step_0 = const()[name = tensor("range_1d_9_step_0"), val = tensor(1)]; + tensor range_1d_9 = range_1d(end = reduce_prod_9, start = range_1d_9_start_0, step = range_1d_9_step_0)[name = tensor("range_1d_9")]; + tensor reshape_45 = reshape(shape = shape_19, x = range_1d_9)[name = tensor("reshape_45")]; + tensor slice_by_index_9 = slice_by_index(begin = concat_36, begin_mask = new_cache_9_internal_tensor_assign_2_begin_mask_0, end = concat_37, end_mask = new_cache_9_internal_tensor_assign_2_end_mask_0, squeeze_mask = new_cache_9_internal_tensor_assign_2_squeeze_mask_0, stride = new_cache_9_internal_tensor_assign_2_stride_0, x = reshape_45)[name = tensor("slice_by_index_9")]; + tensor reshape_46_shape_0 = const()[name = tensor("reshape_46_shape_0"), val = tensor([-1])]; + tensor reshape_46 = reshape(shape = reshape_46_shape_0, x = slice_by_index_9)[name = tensor("reshape_46")]; + tensor reshape_47_shape_0 = const()[name = tensor("reshape_47_shape_0"), val = tensor([-1])]; + tensor reshape_47_cast_fp16 = reshape(shape = reshape_47_shape_0, x = var_1952_cast_fp16)[name = tensor("reshape_47_cast_fp16")]; + tensor reshape_48_shape_0 = const()[name = tensor("reshape_48_shape_0"), val = tensor([-1])]; + tensor reshape_48_cast_fp16 = reshape(shape = reshape_48_shape_0, x = reshape_44_cast_fp16)[name = tensor("reshape_48_cast_fp16")]; + tensor scatter_9_mode_0 = const()[name = tensor("scatter_9_mode_0"), val = tensor("update")]; + tensor scatter_9_axis_0 = const()[name = tensor("scatter_9_axis_0"), val = tensor(0)]; + tensor scatter_9_validate_indices_0 = const()[name = tensor("scatter_9_validate_indices_0"), val = tensor(false)]; + tensor scatter_9_cast_fp16 = scatter(axis = scatter_9_axis_0, data = reshape_48_cast_fp16, indices = reshape_46, mode = scatter_9_mode_0, updates = reshape_47_cast_fp16, validate_indices = scatter_9_validate_indices_0)[name = tensor("scatter_9_cast_fp16")]; + tensor reshape_49_cast_fp16 = reshape(shape = shape_19, x = scatter_9_cast_fp16)[name = tensor("reshape_49_cast_fp16")]; + tensor reshape_49_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("reshape_49_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor keys_25_begin_0 = const()[name = tensor("keys_25_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor keys_25_end_0 = const()[name = tensor("keys_25_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor keys_25_end_mask_0 = const()[name = tensor("keys_25_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor keys_25_squeeze_mask_0 = const()[name = tensor("keys_25_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor keys_25_cast_fp16 = slice_by_index(begin = keys_25_begin_0, end = keys_25_end_0, end_mask = keys_25_end_mask_0, squeeze_mask = keys_25_squeeze_mask_0, x = reshape_49_cast_fp16)[name = tensor("keys_25_cast_fp16")]; + tensor values_25_begin_0 = const()[name = tensor("values_25_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor values_25_end_0 = const()[name = tensor("values_25_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor values_25_end_mask_0 = const()[name = tensor("values_25_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor values_25_squeeze_mask_0 = const()[name = tensor("values_25_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor values_25_cast_fp16 = slice_by_index(begin = values_25_begin_0, end = values_25_end_0, end_mask = values_25_end_mask_0, squeeze_mask = values_25_squeeze_mask_0, x = reshape_49_cast_fp16)[name = tensor("values_25_cast_fp16")]; + tensor var_1964_cast_fp16 = not_equal(x = keys_25_cast_fp16, y = keys_25_cast_fp16)[name = tensor("op_1964_cast_fp16")]; + tensor keys_27_cast_fp16 = select(a = var_326_to_fp16, b = keys_25_cast_fp16, cond = var_1964_cast_fp16)[name = tensor("keys_27_cast_fp16")]; + tensor var_1972_cast_fp16 = not_equal(x = values_25_cast_fp16, y = values_25_cast_fp16)[name = tensor("op_1972_cast_fp16")]; + tensor values_27_cast_fp16 = select(a = var_326_to_fp16, b = values_25_cast_fp16, cond = var_1972_cast_fp16)[name = tensor("values_27_cast_fp16")]; + tensor var_1996 = const()[name = tensor("op_1996"), val = tensor([0, 2, 1, 3])]; + tensor var_2012 = const()[name = tensor("op_2012"), val = tensor([1, 1, 1])]; + tensor var_2013_cast_fp16 = reshape(shape = var_2012, x = position4_to_fp16)[name = tensor("op_2013_cast_fp16")]; + tensor q_positions_9_cast_fp16 = add(x = var_2013_cast_fp16, y = q_offsets_1_promoted_to_fp16)[name = tensor("q_positions_9_cast_fp16")]; + tensor valid_end_9_cast_fp16 = add(x = var_2013_cast_fp16, y = var_392_cast_fp16)[name = tensor("valid_end_9_cast_fp16")]; + tensor valid_mask_9_cast_fp16 = less(x = k_positions_1_promoted_to_fp16, y = valid_end_9_cast_fp16)[name = tensor("valid_mask_9_cast_fp16")]; + tensor causal_mask_9_cast_fp16 = less_equal(x = k_positions_1_promoted_to_fp16, y = q_positions_9_cast_fp16)[name = tensor("causal_mask_9_cast_fp16")]; + tensor var_2041 = logical_and(x = valid_mask_9_cast_fp16, y = causal_mask_9_cast_fp16)[name = tensor("op_2041")]; + tensor attn_mask_9_axes_0 = const()[name = tensor("attn_mask_9_axes_0"), val = tensor([1])]; + tensor attn_mask_9 = expand_dims(axes = attn_mask_9_axes_0, x = var_2041)[name = tensor("attn_mask_9")]; + tensor var_2055_transpose_x_0 = const()[name = tensor("op_2055_transpose_x_0"), val = tensor(false)]; + tensor var_2055_transpose_y_0 = const()[name = tensor("op_2055_transpose_y_0"), val = tensor(false)]; + tensor transpose_23_perm_0 = const()[name = tensor("transpose_23_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_24_perm_0 = const()[name = tensor("transpose_24_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_24 = transpose(perm = transpose_24_perm_0, x = keys_27_cast_fp16)[name = tensor("transpose_26")]; + tensor transpose_23 = transpose(perm = transpose_23_perm_0, x = q_27_cast_fp16)[name = tensor("transpose_27")]; + tensor var_2055_cast_fp16 = matmul(transpose_x = var_2055_transpose_x_0, transpose_y = var_2055_transpose_y_0, x = transpose_23, y = transpose_24)[name = tensor("op_2055_cast_fp16")]; + tensor var_2049_to_fp16 = const()[name = tensor("op_2049_to_fp16"), val = tensor([0x1p-3])]; + tensor attn_weights_25_cast_fp16 = mul(x = var_2055_cast_fp16, y = var_2049_to_fp16)[name = tensor("attn_weights_25_cast_fp16")]; + tensor var_2057 = logical_not(x = attn_mask_9)[name = tensor("op_2057")]; + tensor var_2058_to_fp16 = const()[name = tensor("op_2058_to_fp16"), val = tensor(-inf)]; + tensor attn_weights_27_cast_fp16 = select(a = var_2058_to_fp16, b = attn_weights_25_cast_fp16, cond = var_2057)[name = tensor("attn_weights_27_cast_fp16")]; + tensor var_2060 = const()[name = tensor("op_2060"), val = tensor(-1)]; + tensor attn_weights_29_cast_fp16 = softmax(axis = var_2060, x = attn_weights_27_cast_fp16)[name = tensor("attn_weights_29_cast_fp16")]; + tensor attn_output_9_transpose_x_0 = const()[name = tensor("attn_output_9_transpose_x_0"), val = tensor(false)]; + tensor attn_output_9_transpose_y_0 = const()[name = tensor("attn_output_9_transpose_y_0"), val = tensor(false)]; + tensor values_29_cast_fp16 = transpose(perm = var_1996, x = values_27_cast_fp16)[name = tensor("transpose_28")]; + tensor attn_output_9_cast_fp16 = matmul(transpose_x = attn_output_9_transpose_x_0, transpose_y = attn_output_9_transpose_y_0, x = attn_weights_29_cast_fp16, y = values_29_cast_fp16)[name = tensor("attn_output_9_cast_fp16")]; + tensor var_2068 = const()[name = tensor("op_2068"), val = tensor([0, 2, 1, 3])]; + tensor var_2073 = const()[name = tensor("op_2073"), val = tensor([1, 256, 1024])]; + tensor var_2069_cast_fp16 = transpose(perm = var_2068, x = attn_output_9_cast_fp16)[name = tensor("transpose_25")]; + tensor input_53_cast_fp16 = reshape(shape = var_2073, x = var_2069_cast_fp16)[name = tensor("input_53_cast_fp16")]; + tensor attn4_out_proj_weight_to_fp16 = const()[name = tensor("attn4_out_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(108062592)))]; + tensor linear_17_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = attn4_out_proj_weight_to_fp16, x = input_53_cast_fp16)[name = tensor("linear_17_cast_fp16")]; + tensor var_2082_cast_fp16 = add(x = position4_to_fp16, y = valid_len_to_fp16)[name = tensor("op_2082_cast_fp16")]; + tensor var_2082_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_2082_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor input_55_cast_fp16 = add(x = input_49_cast_fp16, y = linear_17_cast_fp16)[name = tensor("input_55_cast_fp16")]; + tensor input_57_axes_0 = const()[name = tensor("input_57_axes_0"), val = tensor([-1])]; + tensor norm4_2_weight_to_fp16 = const()[name = tensor("norm4_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(110159808)))]; + tensor norm4_2_bias_to_fp16 = const()[name = tensor("norm4_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(110161920)))]; + tensor var_2086_to_fp16 = const()[name = tensor("op_2086_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_57_cast_fp16 = layer_norm(axes = input_57_axes_0, beta = norm4_2_bias_to_fp16, epsilon = var_2086_to_fp16, gamma = norm4_2_weight_to_fp16, x = input_55_cast_fp16)[name = tensor("input_57_cast_fp16")]; + tensor linear4_1_weight_to_fp16 = const()[name = tensor("linear4_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(110164032)))]; + tensor linear_18_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = linear4_1_weight_to_fp16, x = input_57_cast_fp16)[name = tensor("linear_18_cast_fp16")]; + tensor input_59_mode_0 = const()[name = tensor("input_59_mode_0"), val = tensor("EXACT")]; + tensor input_59_cast_fp16 = gelu(mode = input_59_mode_0, x = linear_18_cast_fp16)[name = tensor("input_59_cast_fp16")]; + tensor linear4_2_weight_to_fp16 = const()[name = tensor("linear4_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(118552704)))]; + tensor linear_19_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = linear4_2_weight_to_fp16, x = input_59_cast_fp16)[name = tensor("linear_19_cast_fp16")]; + tensor input_61_cast_fp16 = add(x = input_55_cast_fp16, y = linear_19_cast_fp16)[name = tensor("input_61_cast_fp16")]; + tensor input_63_axes_0 = const()[name = tensor("input_63_axes_0"), val = tensor([-1])]; + tensor norm5_1_weight_to_fp16 = const()[name = tensor("norm5_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(126941376)))]; + tensor norm5_1_bias_to_fp16 = const()[name = tensor("norm5_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(126943488)))]; + tensor var_2103_to_fp16 = const()[name = tensor("op_2103_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_63_cast_fp16 = layer_norm(axes = input_63_axes_0, beta = norm5_1_bias_to_fp16, epsilon = var_2103_to_fp16, gamma = norm5_1_weight_to_fp16, x = input_61_cast_fp16)[name = tensor("input_63_cast_fp16")]; + tensor attn5_in_proj_weight_to_fp16 = const()[name = tensor("attn5_in_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(126945600)))]; + tensor linear_20_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = attn5_in_proj_weight_to_fp16, x = input_63_cast_fp16)[name = tensor("linear_20_cast_fp16")]; + tensor var_2117 = const()[name = tensor("op_2117"), val = tensor([1, 256, 3, 16, 64])]; + tensor qkv_cast_fp16 = reshape(shape = var_2117, x = linear_20_cast_fp16)[name = tensor("qkv_cast_fp16")]; + tensor k_21_begin_0 = const()[name = tensor("k_21_begin_0"), val = tensor([0, 0, 1, 0, 0])]; + tensor k_21_end_0 = const()[name = tensor("k_21_end_0"), val = tensor([1, 256, 2, 16, 64])]; + tensor k_21_end_mask_0 = const()[name = tensor("k_21_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor k_21_squeeze_mask_0 = const()[name = tensor("k_21_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor k_21_cast_fp16 = slice_by_index(begin = k_21_begin_0, end = k_21_end_0, end_mask = k_21_end_mask_0, squeeze_mask = k_21_squeeze_mask_0, x = qkv_cast_fp16)[name = tensor("k_21_cast_fp16")]; + tensor v_begin_0 = const()[name = tensor("v_begin_0"), val = tensor([0, 0, 2, 0, 0])]; + tensor v_end_0 = const()[name = tensor("v_end_0"), val = tensor([1, 256, 3, 16, 64])]; + tensor v_end_mask_0 = const()[name = tensor("v_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor v_squeeze_mask_0 = const()[name = tensor("v_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor v_cast_fp16 = slice_by_index(begin = v_begin_0, end = v_end_0, end_mask = v_end_mask_0, squeeze_mask = v_squeeze_mask_0, x = qkv_cast_fp16)[name = tensor("v_cast_fp16")]; + tensor var_2162 = const()[name = tensor("op_2162"), val = tensor([1, 1])]; + tensor position5_to_fp16_dtype_0 = const()[name = tensor("position5_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor position5_to_fp16 = cast(dtype = position5_to_fp16_dtype_0, x = position5)[name = tensor("cast_132")]; + tensor var_2163_cast_fp16 = reshape(shape = var_2162, x = position5_to_fp16)[name = tensor("op_2163_cast_fp16")]; + tensor ts_33_cast_fp16 = add(x = ts_1_promoted_to_fp16, y = var_2163_cast_fp16)[name = tensor("ts_33_cast_fp16")]; + tensor var_2170 = const()[name = tensor("op_2170"), val = tensor([1, 256, 1, 1])]; + tensor ts_cast_fp16 = reshape(shape = var_2170, x = ts_33_cast_fp16)[name = tensor("ts_cast_fp16")]; + tensor var_2177 = const()[name = tensor("op_2177"), val = tensor([1, 256, 16, 32, 2])]; + tensor k_complex_cast_fp16 = reshape(shape = var_2177, x = k_21_cast_fp16)[name = tensor("k_complex_cast_fp16")]; + tensor var_2181_begin_0 = const()[name = tensor("op_2181_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_2181_end_0 = const()[name = tensor("op_2181_end_0"), val = tensor([1, 256, 16, 32, 1])]; + tensor var_2181_end_mask_0 = const()[name = tensor("op_2181_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_2181_squeeze_mask_0 = const()[name = tensor("op_2181_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_2181_cast_fp16 = slice_by_index(begin = var_2181_begin_0, end = var_2181_end_0, end_mask = var_2181_end_mask_0, squeeze_mask = var_2181_squeeze_mask_0, x = k_complex_cast_fp16)[name = tensor("op_2181_cast_fp16")]; + tensor var_2189_begin_0 = const()[name = tensor("op_2189_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_2189_end_0 = const()[name = tensor("op_2189_end_0"), val = tensor([1, 256, 16, 32, 2])]; + tensor var_2189_end_mask_0 = const()[name = tensor("op_2189_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_2189_squeeze_mask_0 = const()[name = tensor("op_2189_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_2189_cast_fp16 = slice_by_index(begin = var_2189_begin_0, end = var_2189_end_0, end_mask = var_2189_end_mask_0, squeeze_mask = var_2189_squeeze_mask_0, x = k_complex_cast_fp16)[name = tensor("op_2189_cast_fp16")]; + tensor freqs_to_fp16 = const()[name = tensor("freqs_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(133237120)))]; + tensor var_2195_cast_fp16 = mul(x = freqs_to_fp16, y = ts_cast_fp16)[name = tensor("op_2195_cast_fp16")]; + tensor rotr_cast_fp16 = cos(x = var_2195_cast_fp16)[name = tensor("rotr_cast_fp16")]; + tensor roti_cast_fp16 = sin(x = var_2195_cast_fp16)[name = tensor("roti_cast_fp16")]; + tensor var_2199_cast_fp16 = mul(x = var_2181_cast_fp16, y = rotr_cast_fp16)[name = tensor("op_2199_cast_fp16")]; + tensor var_2200_cast_fp16 = mul(x = var_2189_cast_fp16, y = roti_cast_fp16)[name = tensor("op_2200_cast_fp16")]; + tensor kor_21_cast_fp16 = sub(x = var_2199_cast_fp16, y = var_2200_cast_fp16)[name = tensor("kor_21_cast_fp16")]; + tensor var_2203_cast_fp16 = mul(x = var_2181_cast_fp16, y = roti_cast_fp16)[name = tensor("op_2203_cast_fp16")]; + tensor var_2204_cast_fp16 = mul(x = var_2189_cast_fp16, y = rotr_cast_fp16)[name = tensor("op_2204_cast_fp16")]; + tensor koi_21_cast_fp16 = add(x = var_2203_cast_fp16, y = var_2204_cast_fp16)[name = tensor("koi_21_cast_fp16")]; + tensor ko_axis_0 = const()[name = tensor("ko_axis_0"), val = tensor(-1)]; + tensor ko_cast_fp16 = stack(axis = ko_axis_0, values = (kor_21_cast_fp16, koi_21_cast_fp16))[name = tensor("ko_cast_fp16")]; + tensor var_2224 = const()[name = tensor("op_2224"), val = tensor([1, 256, 16, 64])]; + tensor k_cast_fp16 = reshape(shape = var_2224, x = ko_cast_fp16)[name = tensor("k_cast_fp16")]; + tensor write_indices_float_31_cast_fp16 = add(x = var_2163_cast_fp16, y = var_249_promoted_to_fp16)[name = tensor("write_indices_float_31_cast_fp16")]; + tensor _inversed_2245_y_0_to_fp16 = const()[name = tensor("_inversed_2245_y_0_to_fp16"), val = tensor(0x1p-9)]; + tensor _inversed_2245_cast_fp16 = mul(x = write_indices_float_31_cast_fp16, y = _inversed_2245_y_0_to_fp16)[name = tensor("_inversed_2245_cast_fp16")]; + tensor var_2246_cast_fp16 = floor(x = _inversed_2245_cast_fp16)[name = tensor("op_2246_cast_fp16")]; + tensor var_2247_to_fp16 = const()[name = tensor("op_2247_to_fp16"), val = tensor(0x1p+9)]; + tensor var_2248_cast_fp16 = mul(x = var_2246_cast_fp16, y = var_2247_to_fp16)[name = tensor("op_2248_cast_fp16")]; + tensor write_indices_float_33_cast_fp16 = sub(x = write_indices_float_31_cast_fp16, y = var_2248_cast_fp16)[name = tensor("write_indices_float_33_cast_fp16")]; + tensor var_2266_value_0_to_fp16 = const()[name = tensor("op_2266_value_0_to_fp16"), val = tensor(0x1.ffp+8)]; + tensor var_2266_cast_fp16 = fill_like(ref_tensor = write_indices_float_33_cast_fp16, value = var_2266_value_0_to_fp16)[name = tensor("op_2266_cast_fp16")]; + tensor write_indices_float_cast_fp16 = select(a = write_indices_float_33_cast_fp16, b = var_2266_cast_fp16, cond = token_valid_1_cast_fp16)[name = tensor("write_indices_float_cast_fp16")]; + tensor var_2272_dtype_0 = const()[name = tensor("op_2272_dtype_0"), val = tensor("int32")]; + tensor var_2277 = const()[name = tensor("op_2277"), val = tensor([1, 256, 1, 1])]; + tensor write_indices_float_cast_fp16_to_int32 = cast(dtype = var_2272_dtype_0, x = write_indices_float_cast_fp16)[name = tensor("cast_131")]; + tensor var_2278 = reshape(shape = var_2277, x = write_indices_float_cast_fp16_to_int32)[name = tensor("op_2278")]; + tensor write_indices_reps_0 = const()[name = tensor("write_indices_reps_0"), val = tensor([1, 1, 16, 64])]; + tensor write_indices = tile(reps = write_indices_reps_0, x = var_2278)[name = tensor("write_indices")]; + tensor var_2288_begin_0 = const()[name = tensor("op_2288_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_2288_end_0 = const()[name = tensor("op_2288_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor var_2288_end_mask_0 = const()[name = tensor("op_2288_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_2288_squeeze_mask_0 = const()[name = tensor("op_2288_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor cache5_to_fp16_dtype_0 = const()[name = tensor("cache5_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor cache5_to_fp16 = cast(dtype = cache5_to_fp16_dtype_0, x = cache5)[name = tensor("cast_130")]; + tensor var_2288_cast_fp16 = slice_by_index(begin = var_2288_begin_0, end = var_2288_end_0, end_mask = var_2288_end_mask_0, squeeze_mask = var_2288_squeeze_mask_0, x = cache5_to_fp16)[name = tensor("op_2288_cast_fp16")]; + tensor var_2290_axis_0 = const()[name = tensor("op_2290_axis_0"), val = tensor(1)]; + tensor var_2290_mode_0 = const()[name = tensor("op_2290_mode_0"), val = tensor("update")]; + tensor var_2290_validate_indices_0 = const()[name = tensor("op_2290_validate_indices_0"), val = tensor(false)]; + tensor var_2290_cast_fp16 = scatter_along_axis(axis = var_2290_axis_0, data = var_2288_cast_fp16, indices = write_indices, mode = var_2290_mode_0, updates = k_cast_fp16, validate_indices = var_2290_validate_indices_0)[name = tensor("op_2290_cast_fp16")]; + tensor concat_42 = const()[name = tensor("concat_42"), val = tensor([0, 0, 0, 0, 0])]; + tensor concat_43 = const()[name = tensor("concat_43"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_internal_tensor_assign_1_stride_0 = const()[name = tensor("new_cache_internal_tensor_assign_1_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_internal_tensor_assign_1_begin_mask_0 = const()[name = tensor("new_cache_internal_tensor_assign_1_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_internal_tensor_assign_1_end_mask_0 = const()[name = tensor("new_cache_internal_tensor_assign_1_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_internal_tensor_assign_1_squeeze_mask_0 = const()[name = tensor("new_cache_internal_tensor_assign_1_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_20 = const()[name = tensor("shape_20"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_10 = const()[name = tensor("reduce_prod_10"), val = tensor(1048576)]; + tensor range_1d_10_start_0 = const()[name = tensor("range_1d_10_start_0"), val = tensor(0)]; + tensor range_1d_10_step_0 = const()[name = tensor("range_1d_10_step_0"), val = tensor(1)]; + tensor range_1d_10 = range_1d(end = reduce_prod_10, start = range_1d_10_start_0, step = range_1d_10_step_0)[name = tensor("range_1d_10")]; + tensor reshape_50 = reshape(shape = shape_20, x = range_1d_10)[name = tensor("reshape_50")]; + tensor slice_by_index_10 = slice_by_index(begin = concat_42, begin_mask = new_cache_internal_tensor_assign_1_begin_mask_0, end = concat_43, end_mask = new_cache_internal_tensor_assign_1_end_mask_0, squeeze_mask = new_cache_internal_tensor_assign_1_squeeze_mask_0, stride = new_cache_internal_tensor_assign_1_stride_0, x = reshape_50)[name = tensor("slice_by_index_10")]; + tensor reshape_51_shape_0 = const()[name = tensor("reshape_51_shape_0"), val = tensor([-1])]; + tensor reshape_51 = reshape(shape = reshape_51_shape_0, x = slice_by_index_10)[name = tensor("reshape_51")]; + tensor reshape_52_shape_0 = const()[name = tensor("reshape_52_shape_0"), val = tensor([-1])]; + tensor reshape_52_cast_fp16 = reshape(shape = reshape_52_shape_0, x = var_2290_cast_fp16)[name = tensor("reshape_52_cast_fp16")]; + tensor reshape_53_shape_0 = const()[name = tensor("reshape_53_shape_0"), val = tensor([-1])]; + tensor reshape_53_cast_fp16 = reshape(shape = reshape_53_shape_0, x = cache5_to_fp16)[name = tensor("reshape_53_cast_fp16")]; + tensor scatter_10_mode_0 = const()[name = tensor("scatter_10_mode_0"), val = tensor("update")]; + tensor scatter_10_axis_0 = const()[name = tensor("scatter_10_axis_0"), val = tensor(0)]; + tensor scatter_10_validate_indices_0 = const()[name = tensor("scatter_10_validate_indices_0"), val = tensor(false)]; + tensor scatter_10_cast_fp16 = scatter(axis = scatter_10_axis_0, data = reshape_53_cast_fp16, indices = reshape_51, mode = scatter_10_mode_0, updates = reshape_52_cast_fp16, validate_indices = scatter_10_validate_indices_0)[name = tensor("scatter_10_cast_fp16")]; + tensor reshape_54_cast_fp16 = reshape(shape = shape_20, x = scatter_10_cast_fp16)[name = tensor("reshape_54_cast_fp16")]; + tensor var_2298_begin_0 = const()[name = tensor("op_2298_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor var_2298_end_0 = const()[name = tensor("op_2298_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor var_2298_end_mask_0 = const()[name = tensor("op_2298_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_2298_squeeze_mask_0 = const()[name = tensor("op_2298_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor var_2298_cast_fp16 = slice_by_index(begin = var_2298_begin_0, end = var_2298_end_0, end_mask = var_2298_end_mask_0, squeeze_mask = var_2298_squeeze_mask_0, x = reshape_54_cast_fp16)[name = tensor("op_2298_cast_fp16")]; + tensor var_2300_axis_0 = const()[name = tensor("op_2300_axis_0"), val = tensor(1)]; + tensor var_2300_mode_0 = const()[name = tensor("op_2300_mode_0"), val = tensor("update")]; + tensor var_2300_validate_indices_0 = const()[name = tensor("op_2300_validate_indices_0"), val = tensor(false)]; + tensor var_2300_cast_fp16 = scatter_along_axis(axis = var_2300_axis_0, data = var_2298_cast_fp16, indices = write_indices, mode = var_2300_mode_0, updates = v_cast_fp16, validate_indices = var_2300_validate_indices_0)[name = tensor("op_2300_cast_fp16")]; + tensor concat_44 = const()[name = tensor("concat_44"), val = tensor([1, 0, 0, 0, 0])]; + tensor concat_45 = const()[name = tensor("concat_45"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_internal_tensor_assign_2_stride_0 = const()[name = tensor("new_cache_internal_tensor_assign_2_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_internal_tensor_assign_2_begin_mask_0 = const()[name = tensor("new_cache_internal_tensor_assign_2_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_internal_tensor_assign_2_end_mask_0 = const()[name = tensor("new_cache_internal_tensor_assign_2_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_internal_tensor_assign_2_squeeze_mask_0 = const()[name = tensor("new_cache_internal_tensor_assign_2_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_21 = const()[name = tensor("shape_21"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_11 = const()[name = tensor("reduce_prod_11"), val = tensor(1048576)]; + tensor range_1d_11_start_0 = const()[name = tensor("range_1d_11_start_0"), val = tensor(0)]; + tensor range_1d_11_step_0 = const()[name = tensor("range_1d_11_step_0"), val = tensor(1)]; + tensor range_1d_11 = range_1d(end = reduce_prod_11, start = range_1d_11_start_0, step = range_1d_11_step_0)[name = tensor("range_1d_11")]; + tensor reshape_55 = reshape(shape = shape_21, x = range_1d_11)[name = tensor("reshape_55")]; + tensor slice_by_index_11 = slice_by_index(begin = concat_44, begin_mask = new_cache_internal_tensor_assign_2_begin_mask_0, end = concat_45, end_mask = new_cache_internal_tensor_assign_2_end_mask_0, squeeze_mask = new_cache_internal_tensor_assign_2_squeeze_mask_0, stride = new_cache_internal_tensor_assign_2_stride_0, x = reshape_55)[name = tensor("slice_by_index_11")]; + tensor reshape_56_shape_0 = const()[name = tensor("reshape_56_shape_0"), val = tensor([-1])]; + tensor reshape_56 = reshape(shape = reshape_56_shape_0, x = slice_by_index_11)[name = tensor("reshape_56")]; + tensor reshape_57_shape_0 = const()[name = tensor("reshape_57_shape_0"), val = tensor([-1])]; + tensor reshape_57_cast_fp16 = reshape(shape = reshape_57_shape_0, x = var_2300_cast_fp16)[name = tensor("reshape_57_cast_fp16")]; + tensor reshape_58_shape_0 = const()[name = tensor("reshape_58_shape_0"), val = tensor([-1])]; + tensor reshape_58_cast_fp16 = reshape(shape = reshape_58_shape_0, x = reshape_54_cast_fp16)[name = tensor("reshape_58_cast_fp16")]; + tensor scatter_11_mode_0 = const()[name = tensor("scatter_11_mode_0"), val = tensor("update")]; + tensor scatter_11_axis_0 = const()[name = tensor("scatter_11_axis_0"), val = tensor(0)]; + tensor scatter_11_validate_indices_0 = const()[name = tensor("scatter_11_validate_indices_0"), val = tensor(false)]; + tensor scatter_11_cast_fp16 = scatter(axis = scatter_11_axis_0, data = reshape_58_cast_fp16, indices = reshape_56, mode = scatter_11_mode_0, updates = reshape_57_cast_fp16, validate_indices = scatter_11_validate_indices_0)[name = tensor("scatter_11_cast_fp16")]; + tensor reshape_59_cast_fp16 = reshape(shape = shape_21, x = scatter_11_cast_fp16)[name = tensor("reshape_59_cast_fp16")]; + tensor reshape_59_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("reshape_59_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor var_2310_cast_fp16 = add(x = position5_to_fp16, y = valid_len_to_fp16)[name = tensor("op_2310_cast_fp16")]; + tensor var_2310_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_2310_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor var_2310 = cast(dtype = var_2310_cast_fp16_to_fp32_dtype_0, x = var_2310_cast_fp16)[name = tensor("cast_128")]; + tensor new_cache_internal_tensor_assign_2 = cast(dtype = reshape_59_cast_fp16_to_fp32_dtype_0, x = reshape_59_cast_fp16)[name = tensor("cast_129")]; + tensor var_2082 = cast(dtype = var_2082_cast_fp16_to_fp32_dtype_0, x = var_2082_cast_fp16)[name = tensor("cast_133")]; + tensor new_cache_9_internal_tensor_assign_2 = cast(dtype = reshape_49_cast_fp16_to_fp32_dtype_0, x = reshape_49_cast_fp16)[name = tensor("cast_134")]; + tensor var_1671 = cast(dtype = var_1671_cast_fp16_to_fp32_dtype_0, x = var_1671_cast_fp16)[name = tensor("cast_138")]; + tensor new_cache_7_internal_tensor_assign_2 = cast(dtype = reshape_39_cast_fp16_to_fp32_dtype_0, x = reshape_39_cast_fp16)[name = tensor("cast_139")]; + tensor var_1260 = cast(dtype = var_1260_cast_fp16_to_fp32_dtype_0, x = var_1260_cast_fp16)[name = tensor("cast_143")]; + tensor new_cache_5_internal_tensor_assign_2 = cast(dtype = reshape_29_cast_fp16_to_fp32_dtype_0, x = reshape_29_cast_fp16)[name = tensor("cast_144")]; + tensor var_849 = cast(dtype = var_849_cast_fp16_to_fp32_dtype_0, x = var_849_cast_fp16)[name = tensor("cast_148")]; + tensor new_cache_3_internal_tensor_assign_2 = cast(dtype = reshape_19_cast_fp16_to_fp32_dtype_0, x = reshape_19_cast_fp16)[name = tensor("cast_149")]; + tensor var_438 = cast(dtype = var_438_cast_fp16_to_fp32_dtype_0, x = var_438_cast_fp16)[name = tensor("cast_153")]; + tensor new_cache_1_internal_tensor_assign_2 = cast(dtype = reshape_9_cast_fp16_to_fp32_dtype_0, x = reshape_9_cast_fp16)[name = tensor("cast_154")]; + } -> (new_cache_1_internal_tensor_assign_2, var_438, new_cache_3_internal_tensor_assign_2, var_849, new_cache_5_internal_tensor_assign_2, var_1260, new_cache_7_internal_tensor_assign_2, var_1671, new_cache_9_internal_tensor_assign_2, var_2082, new_cache_internal_tensor_assign_2, var_2310); +} \ No newline at end of file diff --git a/v2.1/german/cond_prefill.mlmodelc/weights/weight.bin b/v2.1/german/cond_prefill.mlmodelc/weights/weight.bin new file mode 100644 index 0000000000000000000000000000000000000000..a0fd6fe62ca73eb0c865bef7c5ef386f3b340870 --- /dev/null +++ b/v2.1/german/cond_prefill.mlmodelc/weights/weight.bin @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:446a487be4e656586fa2a8ffa75b792fe06591398080ea594e92b413b2b965c2 +size 133237248 diff --git a/v2.1/german/cond_prefill.mlpackage/Data/com.apple.CoreML/model.mlmodel b/v2.1/german/cond_prefill.mlpackage/Data/com.apple.CoreML/model.mlmodel new file mode 100644 index 0000000000000000000000000000000000000000..0d3b5ebe81e280bb4bca0c918fd596b273a34847 --- /dev/null +++ b/v2.1/german/cond_prefill.mlpackage/Data/com.apple.CoreML/model.mlmodel @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:13d896df1b404a640ec656527d6bc07a582a9c7426279e978833edef3b5b2244 +size 200480 diff --git a/v2.1/german/cond_prefill.mlpackage/Data/com.apple.CoreML/weights/weight.bin b/v2.1/german/cond_prefill.mlpackage/Data/com.apple.CoreML/weights/weight.bin new file mode 100644 index 0000000000000000000000000000000000000000..a0fd6fe62ca73eb0c865bef7c5ef386f3b340870 --- /dev/null +++ b/v2.1/german/cond_prefill.mlpackage/Data/com.apple.CoreML/weights/weight.bin @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:446a487be4e656586fa2a8ffa75b792fe06591398080ea594e92b413b2b965c2 +size 133237248 diff --git a/v2.1/german/cond_prefill.mlpackage/Manifest.json b/v2.1/german/cond_prefill.mlpackage/Manifest.json new file mode 100644 index 0000000000000000000000000000000000000000..e42d67694dbf132d8774a444e15fbf6737f1d58d --- /dev/null +++ b/v2.1/german/cond_prefill.mlpackage/Manifest.json @@ -0,0 +1,18 @@ +{ + "fileFormatVersion": "1.0.0", + "itemInfoEntries": { + "833F2891-4A6F-4651-836B-281AF27086E5": { + "author": "com.apple.CoreML", + "description": "CoreML Model Specification", + "name": "model.mlmodel", + "path": "com.apple.CoreML/model.mlmodel" + }, + "ECB607D3-9E6F-469B-A234-319BADFA2875": { + "author": "com.apple.CoreML", + "description": "CoreML Model Weights", + "name": "weights", + "path": "com.apple.CoreML/weights" + } + }, + "rootModelIdentifier": "833F2891-4A6F-4651-836B-281AF27086E5" +} diff --git a/v2.1/german/constants/bos_emb.npy b/v2.1/german/constants/bos_emb.npy new file mode 100644 index 0000000000000000000000000000000000000000..e5da9ee48e2ca6431746c58edc2100bcc2142df0 --- /dev/null +++ b/v2.1/german/constants/bos_emb.npy @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:51471d57a7093bc6b1e17c28ea223bbcab7671e1ac613cc7a6189fd0c8f09778 +size 256 diff --git a/v2.1/german/constants/emb_mean.npy b/v2.1/german/constants/emb_mean.npy new file mode 100644 index 0000000000000000000000000000000000000000..0466a09c1264dc319a854a8ccc370ca251302e20 --- /dev/null +++ b/v2.1/german/constants/emb_mean.npy @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:00204e3732d8ade46adeba6f24922470c703638da7dc3fb4788f95ea012873a0 +size 256 diff --git a/v2.1/german/constants/emb_std.npy b/v2.1/german/constants/emb_std.npy new file mode 100644 index 0000000000000000000000000000000000000000..afdbc50ed0cfe53be8b317c2f24f6a674ddab536 --- /dev/null +++ b/v2.1/german/constants/emb_std.npy @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:11ae1576585fff8884e2c1f1d12427479594e634ddc6a8c7209e8bf44020b5a5 +size 256 diff --git a/v2.1/german/constants/mimi_init_state.npz b/v2.1/german/constants/mimi_init_state.npz new file mode 100644 index 0000000000000000000000000000000000000000..3b52f383a86ef0439f9993506e4529c4c1234828 --- /dev/null +++ b/v2.1/german/constants/mimi_init_state.npz @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:515588e10e9c14a6b009f1cfda9fcd592542b1640ad856d5fc69e50024702a39 +size 1083108 diff --git a/v2.1/german/constants/quantizer_weight.npy b/v2.1/german/constants/quantizer_weight.npy new file mode 100644 index 0000000000000000000000000000000000000000..2d53e5ced6278307fbff45e258705b0439fa985c --- /dev/null +++ b/v2.1/german/constants/quantizer_weight.npy @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:b9686cb8cb675e4c94515f5bee5268858e8a455fbc5934aaec4677be73fa97a1 +size 65664 diff --git a/v2.1/german/constants/text_embed_table.npy b/v2.1/german/constants/text_embed_table.npy new file mode 100644 index 0000000000000000000000000000000000000000..eefb1fb5490c6f73f6c8f731511db6fc73917e6a --- /dev/null +++ b/v2.1/german/constants/text_embed_table.npy @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:6a30609e0b62d44d941ecfe00268bfddce42527f29577cf1a9fd8e7641174a78 +size 16388224 diff --git a/v2.1/german/constants_bin/alba.safetensors b/v2.1/german/constants_bin/alba.safetensors new file mode 100644 index 0000000000000000000000000000000000000000..c964699c696d56209fa83b89732bd42480f49099 --- /dev/null +++ b/v2.1/german/constants_bin/alba.safetensors @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:b5a1933388634f63f0f709d3139f64d3c78281160d4a3bf6ea3bd1abc3ba94b6 +size 6194424 diff --git a/v2.1/german/constants_bin/anna.safetensors b/v2.1/german/constants_bin/anna.safetensors new file mode 100644 index 0000000000000000000000000000000000000000..ce76c20e6b7d9956d87381680346a22949f0a695 --- /dev/null +++ b/v2.1/german/constants_bin/anna.safetensors @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:6aa8e46421e4a8a602976918ed407fde55e7603b313acbe2e4432905b8bee7bc +size 7816440 diff --git a/v2.1/german/constants_bin/azelma.safetensors b/v2.1/german/constants_bin/azelma.safetensors new file mode 100644 index 0000000000000000000000000000000000000000..09e540d0977b25014f4829e1f2665f26b8bae587 --- /dev/null +++ b/v2.1/german/constants_bin/azelma.safetensors @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:e8fab25e856720159754ecda80a8791d6da09f70cdda32be2eda887e0ff1dcf6 +size 7963896 diff --git a/v2.1/german/constants_bin/bill_boerst.safetensors b/v2.1/german/constants_bin/bill_boerst.safetensors new file mode 100644 index 0000000000000000000000000000000000000000..4ef67662af7e81b7fa8780df382c377ae039d42e --- /dev/null +++ b/v2.1/german/constants_bin/bill_boerst.safetensors @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:b95ed4aab6539920466ed8b8ed1e4f3f21899079345d6f415f48b44faf0e18d3 +size 6735096 diff --git a/v2.1/german/constants_bin/bos_before_voice.bin b/v2.1/german/constants_bin/bos_before_voice.bin new file mode 100644 index 0000000000000000000000000000000000000000..7e4f2329afc45e7b60d0dd35608c74375211ded8 --- /dev/null +++ b/v2.1/german/constants_bin/bos_before_voice.bin @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:4b8b306d01ae74f9c843da41be0e7b23b32c98e82c2380a5fa77cef2338a3d34 +size 4096 diff --git a/v2.1/german/constants_bin/bos_emb.bin b/v2.1/german/constants_bin/bos_emb.bin new file mode 100644 index 0000000000000000000000000000000000000000..c583c1be6428d3ee1410e6465b883e6f775f706e --- /dev/null +++ b/v2.1/german/constants_bin/bos_emb.bin @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:be9f77dd0de8b5af969a0ddc5d57522e99df7e4134c40e55766104f53eeff5ad +size 128 diff --git a/v2.1/german/constants_bin/caro_davy.safetensors b/v2.1/german/constants_bin/caro_davy.safetensors new file mode 100644 index 0000000000000000000000000000000000000000..d1bebdae8d739295ed406ab76e2728571997691f --- /dev/null +++ b/v2.1/german/constants_bin/caro_davy.safetensors @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:caada25a417c0cbdf48a78ae4b00480e1dcd9d7487f41dcd569376e71b9cf245 +size 5260536 diff --git a/v2.1/german/constants_bin/charles.safetensors b/v2.1/german/constants_bin/charles.safetensors new file mode 100644 index 0000000000000000000000000000000000000000..f062b1a93164d75bf64185032cbe59bfafbf7a10 --- /dev/null +++ b/v2.1/german/constants_bin/charles.safetensors @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:c8bbc4a0e46cd0f16df2aaed7dfe811a0ca2d93799837ca4200851db906530b4 +size 6194424 diff --git a/v2.1/german/constants_bin/cosette.safetensors b/v2.1/german/constants_bin/cosette.safetensors new file mode 100644 index 0000000000000000000000000000000000000000..6098a08bc84dd3b0f5d6602f2cb3e93d58ec1c96 --- /dev/null +++ b/v2.1/german/constants_bin/cosette.safetensors @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:e9b82e86a807783bf1f17fd5ad9fdedee1af2e111ef3c7789d10a9a76c89a32a +size 6194424 diff --git a/v2.1/german/constants_bin/eponine.safetensors b/v2.1/german/constants_bin/eponine.safetensors new file mode 100644 index 0000000000000000000000000000000000000000..5794b4574a5608605345e6c5f22a0c5d284887f1 --- /dev/null +++ b/v2.1/german/constants_bin/eponine.safetensors @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:111350eaf68881da45a08239541b445481756365ac03308234a863f070334f33 +size 6931704 diff --git a/v2.1/german/constants_bin/estelle.safetensors b/v2.1/german/constants_bin/estelle.safetensors new file mode 100644 index 0000000000000000000000000000000000000000..6f5f1ecda885ac5f3c3728bb11a6d94719a49442 --- /dev/null +++ b/v2.1/german/constants_bin/estelle.safetensors @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:61285c542f94b345dae0f810a540d4baaca2b0ceb9cd741cd3bd1af993ec74f4 +size 8258808 diff --git a/v2.1/german/constants_bin/eve.safetensors b/v2.1/german/constants_bin/eve.safetensors new file mode 100644 index 0000000000000000000000000000000000000000..994bc2f823bf3824cfcba8224984689516bad36b --- /dev/null +++ b/v2.1/german/constants_bin/eve.safetensors @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:223f60a58a28236b27ca361e01f182fb919e72f0e3dceffb3bf18f006a0d0097 +size 6538488 diff --git a/v2.1/german/constants_bin/fantine.safetensors b/v2.1/german/constants_bin/fantine.safetensors new file mode 100644 index 0000000000000000000000000000000000000000..2c1be45cac50cd89f011821d85c3ed8e12cf31ad --- /dev/null +++ b/v2.1/german/constants_bin/fantine.safetensors @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:f0c2c0d22f5835c8cf8698cc7cb6ab93778229163eb1c55c86fe7e4ec512501a +size 6538488 diff --git a/v2.1/german/constants_bin/george.safetensors b/v2.1/german/constants_bin/george.safetensors new file mode 100644 index 0000000000000000000000000000000000000000..e0afbfe068e87f4551da99f96ccc8b0c3e3284f2 --- /dev/null +++ b/v2.1/german/constants_bin/george.safetensors @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:bb0ecc9d5524fe64102f44c190be74d2dbee33bcf51f44dbb16e9b6b046496e1 +size 6243576 diff --git a/v2.1/german/constants_bin/giovanni.safetensors b/v2.1/german/constants_bin/giovanni.safetensors new file mode 100644 index 0000000000000000000000000000000000000000..fdda973e640245e29bfb7daa1096136b2807f495 --- /dev/null +++ b/v2.1/german/constants_bin/giovanni.safetensors @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:35127fade02a96edca0f0ff6ec9f129e935d39bc774d0245e6e3ea2da2276db9 +size 4621552 diff --git a/v2.1/german/constants_bin/jane.safetensors b/v2.1/german/constants_bin/jane.safetensors new file mode 100644 index 0000000000000000000000000000000000000000..f86eb4ba2bf8f8e5cc36b01d69dac80e1d95446e --- /dev/null +++ b/v2.1/german/constants_bin/jane.safetensors @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:4572c0d06ce058ecb4f54d5059dcb872b3084d0cef95958faaf73bd4e1ccaf28 +size 7374072 diff --git a/v2.1/german/constants_bin/javert.safetensors b/v2.1/german/constants_bin/javert.safetensors new file mode 100644 index 0000000000000000000000000000000000000000..e5902ebfb5a507b53c404d2816e192c23829606e --- /dev/null +++ b/v2.1/german/constants_bin/javert.safetensors @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:71eed75d77eb0c3f902ef0900401530c06983aa100ea4cdbec036ad607778a16 +size 6194424 diff --git a/v2.1/german/constants_bin/jean.safetensors b/v2.1/german/constants_bin/jean.safetensors new file mode 100644 index 0000000000000000000000000000000000000000..776cc77c76db917fe42a640ef4e7ff4afa81afbd --- /dev/null +++ b/v2.1/german/constants_bin/jean.safetensors @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:43be39db618d1b632d879459e9c1dc618258593eb97c45d75cb98ae47920d1c7 +size 6194424 diff --git a/v2.1/german/constants_bin/juergen.safetensors b/v2.1/german/constants_bin/juergen.safetensors new file mode 100644 index 0000000000000000000000000000000000000000..5ff8b06812afbb9de5a8181da540b69cb815a43f --- /dev/null +++ b/v2.1/german/constants_bin/juergen.safetensors @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:826efa0f02676cb34743e329ff3042f406b9aa05e6110e7ce9a342e289e5f29c +size 6243576 diff --git a/v2.1/german/constants_bin/lola.safetensors b/v2.1/german/constants_bin/lola.safetensors new file mode 100644 index 0000000000000000000000000000000000000000..130952a7ece9611146fc276454ed79a5339c2043 --- /dev/null +++ b/v2.1/german/constants_bin/lola.safetensors @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:4fdba576772c7b1c90e8e37ec0b9bd45d8b37d1d01b2060caea100e482b60394 +size 5948664 diff --git a/v2.1/german/constants_bin/marius.safetensors b/v2.1/german/constants_bin/marius.safetensors new file mode 100644 index 0000000000000000000000000000000000000000..fdc908d7878bba0215514cd40cf98614f1ddd490 --- /dev/null +++ b/v2.1/german/constants_bin/marius.safetensors @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:7f74e9727ee0ad2698113e07b75e24519c7a180f9ba3b033bd54bc1ef18c5d0f +size 6194424 diff --git a/v2.1/german/constants_bin/mary.safetensors b/v2.1/german/constants_bin/mary.safetensors new file mode 100644 index 0000000000000000000000000000000000000000..9d7f62dd126fdb374fcfe0bad343b5a1ae5caeb4 --- /dev/null +++ b/v2.1/german/constants_bin/mary.safetensors @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:476f100e38206c5dff2cb22d582857241b7b718376f65d0ac69b9b0ea9dc1ebb +size 6194424 diff --git a/v2.1/german/constants_bin/michael.safetensors b/v2.1/german/constants_bin/michael.safetensors new file mode 100644 index 0000000000000000000000000000000000000000..9941acd1008fc154521dcdbcbb2a368e2fe607d3 --- /dev/null +++ b/v2.1/german/constants_bin/michael.safetensors @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:6a491e111f277fab05bc05696691d621fd0847ae3b3be9271beb64453970d0f2 +size 7275768 diff --git a/v2.1/german/constants_bin/paul.safetensors b/v2.1/german/constants_bin/paul.safetensors new file mode 100644 index 0000000000000000000000000000000000000000..c6038fd800529640667ae3a8c91805f3f2f4dbda --- /dev/null +++ b/v2.1/german/constants_bin/paul.safetensors @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:3f74b464af87e9f8ee59895d806d6a2a05b5e7c345491d03177a92bc66a0eede +size 6980856 diff --git a/v2.1/german/constants_bin/peter_yearsley.safetensors b/v2.1/german/constants_bin/peter_yearsley.safetensors new file mode 100644 index 0000000000000000000000000000000000000000..0f84b2f96ff3be7a09192a927cc2b0b739529e0d --- /dev/null +++ b/v2.1/german/constants_bin/peter_yearsley.safetensors @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:1cc888975274ecb60968e5c6ffd653f9bc3e9fe5b2527f71df52b1a7b3efba09 +size 3736816 diff --git a/v2.1/german/constants_bin/rafael.safetensors b/v2.1/german/constants_bin/rafael.safetensors new file mode 100644 index 0000000000000000000000000000000000000000..1f15797fabcd69cd1f93e3c6a7bb75de23dd2c42 --- /dev/null +++ b/v2.1/german/constants_bin/rafael.safetensors @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:0c8af3f86d3d90906ce4bfdc02179f2ff59457ef0fc20d7ca6f49745f5876213 +size 6194424 diff --git a/v2.1/german/constants_bin/stuart_bell.safetensors b/v2.1/german/constants_bin/stuart_bell.safetensors new file mode 100644 index 0000000000000000000000000000000000000000..f072fc8d453c30d331581013ec676038bc702eb0 --- /dev/null +++ b/v2.1/german/constants_bin/stuart_bell.safetensors @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:08d8f183ef1d300acad1b6018d102e862b90d4df64bc79c77e619b99f7a172ce +size 5260536 diff --git a/v2.1/german/constants_bin/text_embed_table.bin b/v2.1/german/constants_bin/text_embed_table.bin new file mode 100644 index 0000000000000000000000000000000000000000..49df6e3f71e602b07dd020ced3f47cf0b409ed5c --- /dev/null +++ b/v2.1/german/constants_bin/text_embed_table.bin @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:a733347daa714c28d83c36907b7962fb68e58f72025cb426eedc330663ab6b70 +size 16388096 diff --git a/v2.1/german/constants_bin/tokenizer.model b/v2.1/german/constants_bin/tokenizer.model new file mode 100644 index 0000000000000000000000000000000000000000..40e0004b17234ee1071f728c7780d7edcc176bdd --- /dev/null +++ b/v2.1/german/constants_bin/tokenizer.model @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:389079b9c67cd17ff1c565d53bc695df390b4ca8abf2300947e659585edd9d2a +size 59837 diff --git a/v2.1/german/constants_bin/vera.safetensors b/v2.1/german/constants_bin/vera.safetensors new file mode 100644 index 0000000000000000000000000000000000000000..2a264ee7c52560d6f3d1a3e5e0c3a17d9d6be9ad --- /dev/null +++ b/v2.1/german/constants_bin/vera.safetensors @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:53727005ebf05cac548aa372e26f93ca89b883210b904b58449c247a4f5444d1 +size 6735096 diff --git a/v2.1/german/flow_decoder_fused.mlmodelc/analytics/coremldata.bin b/v2.1/german/flow_decoder_fused.mlmodelc/analytics/coremldata.bin new file mode 100644 index 0000000000000000000000000000000000000000..f98f5718160c834d0d2add4a4e63152a8d409ebe --- /dev/null +++ b/v2.1/german/flow_decoder_fused.mlmodelc/analytics/coremldata.bin @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:8bcd0c77ddf388bc5e36a4c39f5649cab6e040fe7dde06c55c16c4afaf535943 +size 243 diff --git a/v2.1/german/flow_decoder_fused.mlmodelc/coremldata.bin b/v2.1/german/flow_decoder_fused.mlmodelc/coremldata.bin new file mode 100644 index 0000000000000000000000000000000000000000..48d17a84f8eff292be995d39150618a51b091326 --- /dev/null +++ b/v2.1/german/flow_decoder_fused.mlmodelc/coremldata.bin @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:28cfa0328e31d406f42b645e7a86f29c216315287a58e1626f80584fc4bc6d17 +size 413 diff --git a/v2.1/german/flow_decoder_fused.mlmodelc/model.mil b/v2.1/german/flow_decoder_fused.mlmodelc/model.mil new file mode 100644 index 0000000000000000000000000000000000000000..c64a1f34555e7bbf9c8d4502d29f9ef52dd6b0ec --- /dev/null +++ b/v2.1/german/flow_decoder_fused.mlmodelc/model.mil @@ -0,0 +1,1915 @@ +program(1.0) +[buildInfo = dict, tensor>({{"coremlc-component-MIL", "3520.4.1"}, {"coremlc-version", "3520.5.1"}, {"coremltools-component-torch", "2.12.0"}, {"coremltools-source-dialect", "TorchScript"}, {"coremltools-version", "9.0"}})] +{ + func main(tensor latent_init, tensor transformer_out) { + tensor latent_init_to_fp16_dtype_0 = const()[name = tensor("latent_init_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor flow_net_input_proj_weight_to_fp16 = const()[name = tensor("flow_net_input_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(64)))]; + tensor flow_net_input_proj_bias_to_fp16 = const()[name = tensor("flow_net_input_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(32896)))]; + tensor latent_init_to_fp16 = cast(dtype = latent_init_to_fp16_dtype_0, x = latent_init)[name = tensor("cast_146")]; + tensor linear_0_cast_fp16 = linear(bias = flow_net_input_proj_bias_to_fp16, weight = flow_net_input_proj_weight_to_fp16, x = latent_init_to_fp16)[name = tensor("linear_0_cast_fp16")]; + tensor input_3_to_fp16 = const()[name = tensor("input_3_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(33984)))]; + tensor input_5_cast_fp16 = silu(x = input_3_to_fp16)[name = tensor("input_5_cast_fp16")]; + tensor flow_net_time_embed_0_mlp_2_weight_to_fp16 = const()[name = tensor("flow_net_time_embed_0_mlp_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(35072)))]; + tensor flow_net_time_embed_0_mlp_2_bias_to_fp16 = const()[name = tensor("flow_net_time_embed_0_mlp_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(559424)))]; + tensor linear_2_cast_fp16 = linear(bias = flow_net_time_embed_0_mlp_2_bias_to_fp16, weight = flow_net_time_embed_0_mlp_2_weight_to_fp16, x = input_5_cast_fp16)[name = tensor("linear_2_cast_fp16")]; + tensor reduce_mean_0_axes_0 = const()[name = tensor("reduce_mean_0_axes_0"), val = tensor([-1])]; + tensor reduce_mean_0_keep_dims_0 = const()[name = tensor("reduce_mean_0_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_0_cast_fp16 = reduce_mean(axes = reduce_mean_0_axes_0, keep_dims = reduce_mean_0_keep_dims_0, x = linear_2_cast_fp16)[name = tensor("reduce_mean_0_cast_fp16")]; + tensor sub_0_cast_fp16 = sub(x = linear_2_cast_fp16, y = reduce_mean_0_cast_fp16)[name = tensor("sub_0_cast_fp16")]; + tensor square_0_cast_fp16 = square(x = sub_0_cast_fp16)[name = tensor("square_0_cast_fp16")]; + tensor reduce_mean_1_axes_0 = const()[name = tensor("reduce_mean_1_axes_0"), val = tensor([-1])]; + tensor reduce_mean_1_keep_dims_0 = const()[name = tensor("reduce_mean_1_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_1_cast_fp16 = reduce_mean(axes = reduce_mean_1_axes_0, keep_dims = reduce_mean_1_keep_dims_0, x = square_0_cast_fp16)[name = tensor("reduce_mean_1_cast_fp16")]; + tensor real_div_0_to_fp16 = const()[name = tensor("real_div_0_to_fp16"), val = tensor(0x1.008p+0)]; + tensor mul_0_cast_fp16 = mul(x = reduce_mean_1_cast_fp16, y = real_div_0_to_fp16)[name = tensor("mul_0_cast_fp16")]; + tensor var_70_to_fp16 = const()[name = tensor("op_70_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_1_cast_fp16 = add(x = mul_0_cast_fp16, y = var_70_to_fp16)[name = tensor("var_1_cast_fp16")]; + tensor var_73_epsilon_0 = const()[name = tensor("op_73_epsilon_0"), val = tensor(0x1.197998p-40)]; + tensor var_73_cast_fp16 = rsqrt(epsilon = var_73_epsilon_0, x = var_1_cast_fp16)[name = tensor("op_73_cast_fp16")]; + tensor const_3_to_fp16 = const()[name = tensor("const_3_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(560512)))]; + tensor var_74_cast_fp16 = mul(x = const_3_to_fp16, y = var_73_cast_fp16)[name = tensor("op_74_cast_fp16")]; + tensor var_75_cast_fp16 = mul(x = linear_2_cast_fp16, y = var_74_cast_fp16)[name = tensor("op_75_cast_fp16")]; + tensor input_9_to_fp16 = const()[name = tensor("input_9_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(561600)))]; + tensor input_11_cast_fp16 = silu(x = input_9_to_fp16)[name = tensor("input_11_cast_fp16")]; + tensor flow_net_time_embed_1_mlp_2_weight_to_fp16 = const()[name = tensor("flow_net_time_embed_1_mlp_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(562688)))]; + tensor flow_net_time_embed_1_mlp_2_bias_to_fp16 = const()[name = tensor("flow_net_time_embed_1_mlp_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1087040)))]; + tensor linear_4_cast_fp16 = linear(bias = flow_net_time_embed_1_mlp_2_bias_to_fp16, weight = flow_net_time_embed_1_mlp_2_weight_to_fp16, x = input_11_cast_fp16)[name = tensor("linear_4_cast_fp16")]; + tensor reduce_mean_2_axes_0 = const()[name = tensor("reduce_mean_2_axes_0"), val = tensor([-1])]; + tensor reduce_mean_2_keep_dims_0 = const()[name = tensor("reduce_mean_2_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_2_cast_fp16 = reduce_mean(axes = reduce_mean_2_axes_0, keep_dims = reduce_mean_2_keep_dims_0, x = linear_4_cast_fp16)[name = tensor("reduce_mean_2_cast_fp16")]; + tensor sub_2_cast_fp16 = sub(x = linear_4_cast_fp16, y = reduce_mean_2_cast_fp16)[name = tensor("sub_2_cast_fp16")]; + tensor square_1_cast_fp16 = square(x = sub_2_cast_fp16)[name = tensor("square_1_cast_fp16")]; + tensor reduce_mean_3_axes_0 = const()[name = tensor("reduce_mean_3_axes_0"), val = tensor([-1])]; + tensor reduce_mean_3_keep_dims_0 = const()[name = tensor("reduce_mean_3_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_3_cast_fp16 = reduce_mean(axes = reduce_mean_3_axes_0, keep_dims = reduce_mean_3_keep_dims_0, x = square_1_cast_fp16)[name = tensor("reduce_mean_3_cast_fp16")]; + tensor real_div_1_to_fp16 = const()[name = tensor("real_div_1_to_fp16"), val = tensor(0x1.008p+0)]; + tensor mul_1_cast_fp16 = mul(x = reduce_mean_3_cast_fp16, y = real_div_1_to_fp16)[name = tensor("mul_1_cast_fp16")]; + tensor var_110_to_fp16 = const()[name = tensor("op_110_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_3_cast_fp16 = add(x = mul_1_cast_fp16, y = var_110_to_fp16)[name = tensor("var_3_cast_fp16")]; + tensor var_113_epsilon_0 = const()[name = tensor("op_113_epsilon_0"), val = tensor(0x1.197998p-40)]; + tensor var_113_cast_fp16 = rsqrt(epsilon = var_113_epsilon_0, x = var_3_cast_fp16)[name = tensor("op_113_cast_fp16")]; + tensor const_5_to_fp16 = const()[name = tensor("const_5_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1088128)))]; + tensor var_114_cast_fp16 = mul(x = const_5_to_fp16, y = var_113_cast_fp16)[name = tensor("op_114_cast_fp16")]; + tensor var_115_cast_fp16 = mul(x = linear_4_cast_fp16, y = var_114_cast_fp16)[name = tensor("op_115_cast_fp16")]; + tensor var_127_cast_fp16 = add(x = var_75_cast_fp16, y = var_115_cast_fp16)[name = tensor("op_127_cast_fp16")]; + tensor _inversed_t_combined_1_y_0_to_fp16 = const()[name = tensor("_inversed_t_combined_1_y_0_to_fp16"), val = tensor(0x1p-1)]; + tensor _inversed_t_combined_1_cast_fp16 = mul(x = var_127_cast_fp16, y = _inversed_t_combined_1_y_0_to_fp16)[name = tensor("_inversed_t_combined_1_cast_fp16")]; + tensor transformer_out_to_fp16_dtype_0 = const()[name = tensor("transformer_out_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor flow_net_cond_embed_weight_to_fp16 = const()[name = tensor("flow_net_cond_embed_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1089216)))]; + tensor flow_net_cond_embed_bias_to_fp16 = const()[name = tensor("flow_net_cond_embed_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(2137856)))]; + tensor transformer_out_to_fp16 = cast(dtype = transformer_out_to_fp16_dtype_0, x = transformer_out)[name = tensor("cast_145")]; + tensor linear_5_cast_fp16 = linear(bias = flow_net_cond_embed_bias_to_fp16, weight = flow_net_cond_embed_weight_to_fp16, x = transformer_out_to_fp16)[name = tensor("linear_5_cast_fp16")]; + tensor input_13_cast_fp16 = add(x = _inversed_t_combined_1_cast_fp16, y = linear_5_cast_fp16)[name = tensor("input_13_cast_fp16")]; + tensor input_15_cast_fp16 = silu(x = input_13_cast_fp16)[name = tensor("input_15_cast_fp16")]; + tensor flow_net_res_blocks_0_adaLN_modulation_1_weight_to_fp16 = const()[name = tensor("flow_net_res_blocks_0_adaLN_modulation_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(2138944)))]; + tensor flow_net_res_blocks_0_adaLN_modulation_1_bias_to_fp16 = const()[name = tensor("flow_net_res_blocks_0_adaLN_modulation_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(3711872)))]; + tensor linear_6_cast_fp16 = linear(bias = flow_net_res_blocks_0_adaLN_modulation_1_bias_to_fp16, weight = flow_net_res_blocks_0_adaLN_modulation_1_weight_to_fp16, x = input_15_cast_fp16)[name = tensor("linear_6_cast_fp16")]; + tensor var_142_split_sizes_0 = const()[name = tensor("op_142_split_sizes_0"), val = tensor([512, 512, 512])]; + tensor var_142_axis_0 = const()[name = tensor("op_142_axis_0"), val = tensor(-1)]; + tensor var_142_cast_fp16_0, tensor var_142_cast_fp16_1, tensor var_142_cast_fp16_2 = split(axis = var_142_axis_0, split_sizes = var_142_split_sizes_0, x = linear_6_cast_fp16)[name = tensor("op_142_cast_fp16")]; + tensor mean_1_axes_0 = const()[name = tensor("mean_1_axes_0"), val = tensor([-1])]; + tensor mean_1_keep_dims_0 = const()[name = tensor("mean_1_keep_dims_0"), val = tensor(true)]; + tensor mean_1_cast_fp16 = reduce_mean(axes = mean_1_axes_0, keep_dims = mean_1_keep_dims_0, x = linear_0_cast_fp16)[name = tensor("mean_1_cast_fp16")]; + tensor sub_4_cast_fp16 = sub(x = linear_0_cast_fp16, y = mean_1_cast_fp16)[name = tensor("sub_4_cast_fp16")]; + tensor square_2_cast_fp16 = square(x = sub_4_cast_fp16)[name = tensor("square_2_cast_fp16")]; + tensor reduce_mean_5_axes_0 = const()[name = tensor("reduce_mean_5_axes_0"), val = tensor([-1])]; + tensor reduce_mean_5_keep_dims_0 = const()[name = tensor("reduce_mean_5_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_5_cast_fp16 = reduce_mean(axes = reduce_mean_5_axes_0, keep_dims = reduce_mean_5_keep_dims_0, x = square_2_cast_fp16)[name = tensor("reduce_mean_5_cast_fp16")]; + tensor var_152_to_fp16 = const()[name = tensor("op_152_to_fp16"), val = tensor(0x1.1p-20)]; + tensor var_153_cast_fp16 = add(x = reduce_mean_5_cast_fp16, y = var_152_to_fp16)[name = tensor("op_153_cast_fp16")]; + tensor var_154_cast_fp16 = sqrt(x = var_153_cast_fp16)[name = tensor("op_154_cast_fp16")]; + tensor x_7_cast_fp16 = real_div(x = sub_4_cast_fp16, y = var_154_cast_fp16)[name = tensor("x_7_cast_fp16")]; + tensor flow_net_res_blocks_0_in_ln_weight_to_fp16 = const()[name = tensor("flow_net_res_blocks_0_in_ln_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(3715008)))]; + tensor var_156_cast_fp16 = mul(x = x_7_cast_fp16, y = flow_net_res_blocks_0_in_ln_weight_to_fp16)[name = tensor("op_156_cast_fp16")]; + tensor flow_net_res_blocks_0_in_ln_bias_to_fp16 = const()[name = tensor("flow_net_res_blocks_0_in_ln_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(3716096)))]; + tensor x_9_cast_fp16 = add(x = var_156_cast_fp16, y = flow_net_res_blocks_0_in_ln_bias_to_fp16)[name = tensor("x_9_cast_fp16")]; + tensor var_158_promoted_to_fp16 = const()[name = tensor("op_158_promoted_to_fp16"), val = tensor(0x1p+0)]; + tensor var_159_cast_fp16 = add(x = var_142_cast_fp16_1, y = var_158_promoted_to_fp16)[name = tensor("op_159_cast_fp16")]; + tensor var_160_cast_fp16 = mul(x = x_9_cast_fp16, y = var_159_cast_fp16)[name = tensor("op_160_cast_fp16")]; + tensor input_17_cast_fp16 = add(x = var_160_cast_fp16, y = var_142_cast_fp16_0)[name = tensor("input_17_cast_fp16")]; + tensor flow_net_res_blocks_0_mlp_0_weight_to_fp16 = const()[name = tensor("flow_net_res_blocks_0_mlp_0_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(3717184)))]; + tensor flow_net_res_blocks_0_mlp_0_bias_to_fp16 = const()[name = tensor("flow_net_res_blocks_0_mlp_0_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(4241536)))]; + tensor linear_7_cast_fp16 = linear(bias = flow_net_res_blocks_0_mlp_0_bias_to_fp16, weight = flow_net_res_blocks_0_mlp_0_weight_to_fp16, x = input_17_cast_fp16)[name = tensor("linear_7_cast_fp16")]; + tensor input_21_cast_fp16 = silu(x = linear_7_cast_fp16)[name = tensor("input_21_cast_fp16")]; + tensor flow_net_res_blocks_0_mlp_2_weight_to_fp16 = const()[name = tensor("flow_net_res_blocks_0_mlp_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(4242624)))]; + tensor flow_net_res_blocks_0_mlp_2_bias_to_fp16 = const()[name = tensor("flow_net_res_blocks_0_mlp_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(4766976)))]; + tensor linear_8_cast_fp16 = linear(bias = flow_net_res_blocks_0_mlp_2_bias_to_fp16, weight = flow_net_res_blocks_0_mlp_2_weight_to_fp16, x = input_21_cast_fp16)[name = tensor("linear_8_cast_fp16")]; + tensor var_171_cast_fp16 = mul(x = var_142_cast_fp16_2, y = linear_8_cast_fp16)[name = tensor("op_171_cast_fp16")]; + tensor x_11_cast_fp16 = add(x = linear_0_cast_fp16, y = var_171_cast_fp16)[name = tensor("x_11_cast_fp16")]; + tensor flow_net_res_blocks_1_adaLN_modulation_1_weight_to_fp16 = const()[name = tensor("flow_net_res_blocks_1_adaLN_modulation_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(4768064)))]; + tensor flow_net_res_blocks_1_adaLN_modulation_1_bias_to_fp16 = const()[name = tensor("flow_net_res_blocks_1_adaLN_modulation_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(6340992)))]; + tensor linear_9_cast_fp16 = linear(bias = flow_net_res_blocks_1_adaLN_modulation_1_bias_to_fp16, weight = flow_net_res_blocks_1_adaLN_modulation_1_weight_to_fp16, x = input_15_cast_fp16)[name = tensor("linear_9_cast_fp16")]; + tensor var_181_split_sizes_0 = const()[name = tensor("op_181_split_sizes_0"), val = tensor([512, 512, 512])]; + tensor var_181_axis_0 = const()[name = tensor("op_181_axis_0"), val = tensor(-1)]; + tensor var_181_cast_fp16_0, tensor var_181_cast_fp16_1, tensor var_181_cast_fp16_2 = split(axis = var_181_axis_0, split_sizes = var_181_split_sizes_0, x = linear_9_cast_fp16)[name = tensor("op_181_cast_fp16")]; + tensor mean_3_axes_0 = const()[name = tensor("mean_3_axes_0"), val = tensor([-1])]; + tensor mean_3_keep_dims_0 = const()[name = tensor("mean_3_keep_dims_0"), val = tensor(true)]; + tensor mean_3_cast_fp16 = reduce_mean(axes = mean_3_axes_0, keep_dims = mean_3_keep_dims_0, x = x_11_cast_fp16)[name = tensor("mean_3_cast_fp16")]; + tensor sub_5_cast_fp16 = sub(x = x_11_cast_fp16, y = mean_3_cast_fp16)[name = tensor("sub_5_cast_fp16")]; + tensor square_3_cast_fp16 = square(x = sub_5_cast_fp16)[name = tensor("square_3_cast_fp16")]; + tensor reduce_mean_7_axes_0 = const()[name = tensor("reduce_mean_7_axes_0"), val = tensor([-1])]; + tensor reduce_mean_7_keep_dims_0 = const()[name = tensor("reduce_mean_7_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_7_cast_fp16 = reduce_mean(axes = reduce_mean_7_axes_0, keep_dims = reduce_mean_7_keep_dims_0, x = square_3_cast_fp16)[name = tensor("reduce_mean_7_cast_fp16")]; + tensor var_191_to_fp16 = const()[name = tensor("op_191_to_fp16"), val = tensor(0x1.1p-20)]; + tensor var_192_cast_fp16 = add(x = reduce_mean_7_cast_fp16, y = var_191_to_fp16)[name = tensor("op_192_cast_fp16")]; + tensor var_193_cast_fp16 = sqrt(x = var_192_cast_fp16)[name = tensor("op_193_cast_fp16")]; + tensor x_13_cast_fp16 = real_div(x = sub_5_cast_fp16, y = var_193_cast_fp16)[name = tensor("x_13_cast_fp16")]; + tensor flow_net_res_blocks_1_in_ln_weight_to_fp16 = const()[name = tensor("flow_net_res_blocks_1_in_ln_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(6344128)))]; + tensor var_195_cast_fp16 = mul(x = x_13_cast_fp16, y = flow_net_res_blocks_1_in_ln_weight_to_fp16)[name = tensor("op_195_cast_fp16")]; + tensor flow_net_res_blocks_1_in_ln_bias_to_fp16 = const()[name = tensor("flow_net_res_blocks_1_in_ln_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(6345216)))]; + tensor x_15_cast_fp16 = add(x = var_195_cast_fp16, y = flow_net_res_blocks_1_in_ln_bias_to_fp16)[name = tensor("x_15_cast_fp16")]; + tensor var_197_promoted_to_fp16 = const()[name = tensor("op_197_promoted_to_fp16"), val = tensor(0x1p+0)]; + tensor var_198_cast_fp16 = add(x = var_181_cast_fp16_1, y = var_197_promoted_to_fp16)[name = tensor("op_198_cast_fp16")]; + tensor var_199_cast_fp16 = mul(x = x_15_cast_fp16, y = var_198_cast_fp16)[name = tensor("op_199_cast_fp16")]; + tensor input_25_cast_fp16 = add(x = var_199_cast_fp16, y = var_181_cast_fp16_0)[name = tensor("input_25_cast_fp16")]; + tensor flow_net_res_blocks_1_mlp_0_weight_to_fp16 = const()[name = tensor("flow_net_res_blocks_1_mlp_0_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(6346304)))]; + tensor flow_net_res_blocks_1_mlp_0_bias_to_fp16 = const()[name = tensor("flow_net_res_blocks_1_mlp_0_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(6870656)))]; + tensor linear_10_cast_fp16 = linear(bias = flow_net_res_blocks_1_mlp_0_bias_to_fp16, weight = flow_net_res_blocks_1_mlp_0_weight_to_fp16, x = input_25_cast_fp16)[name = tensor("linear_10_cast_fp16")]; + tensor input_29_cast_fp16 = silu(x = linear_10_cast_fp16)[name = tensor("input_29_cast_fp16")]; + tensor flow_net_res_blocks_1_mlp_2_weight_to_fp16 = const()[name = tensor("flow_net_res_blocks_1_mlp_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(6871744)))]; + tensor flow_net_res_blocks_1_mlp_2_bias_to_fp16 = const()[name = tensor("flow_net_res_blocks_1_mlp_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(7396096)))]; + tensor linear_11_cast_fp16 = linear(bias = flow_net_res_blocks_1_mlp_2_bias_to_fp16, weight = flow_net_res_blocks_1_mlp_2_weight_to_fp16, x = input_29_cast_fp16)[name = tensor("linear_11_cast_fp16")]; + tensor var_210_cast_fp16 = mul(x = var_181_cast_fp16_2, y = linear_11_cast_fp16)[name = tensor("op_210_cast_fp16")]; + tensor x_17_cast_fp16 = add(x = x_11_cast_fp16, y = var_210_cast_fp16)[name = tensor("x_17_cast_fp16")]; + tensor flow_net_res_blocks_2_adaLN_modulation_1_weight_to_fp16 = const()[name = tensor("flow_net_res_blocks_2_adaLN_modulation_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(7397184)))]; + tensor flow_net_res_blocks_2_adaLN_modulation_1_bias_to_fp16 = const()[name = tensor("flow_net_res_blocks_2_adaLN_modulation_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8970112)))]; + tensor linear_12_cast_fp16 = linear(bias = flow_net_res_blocks_2_adaLN_modulation_1_bias_to_fp16, weight = flow_net_res_blocks_2_adaLN_modulation_1_weight_to_fp16, x = input_15_cast_fp16)[name = tensor("linear_12_cast_fp16")]; + tensor var_220_split_sizes_0 = const()[name = tensor("op_220_split_sizes_0"), val = tensor([512, 512, 512])]; + tensor var_220_axis_0 = const()[name = tensor("op_220_axis_0"), val = tensor(-1)]; + tensor var_220_cast_fp16_0, tensor var_220_cast_fp16_1, tensor var_220_cast_fp16_2 = split(axis = var_220_axis_0, split_sizes = var_220_split_sizes_0, x = linear_12_cast_fp16)[name = tensor("op_220_cast_fp16")]; + tensor mean_5_axes_0 = const()[name = tensor("mean_5_axes_0"), val = tensor([-1])]; + tensor mean_5_keep_dims_0 = const()[name = tensor("mean_5_keep_dims_0"), val = tensor(true)]; + tensor mean_5_cast_fp16 = reduce_mean(axes = mean_5_axes_0, keep_dims = mean_5_keep_dims_0, x = x_17_cast_fp16)[name = tensor("mean_5_cast_fp16")]; + tensor sub_6_cast_fp16 = sub(x = x_17_cast_fp16, y = mean_5_cast_fp16)[name = tensor("sub_6_cast_fp16")]; + tensor square_4_cast_fp16 = square(x = sub_6_cast_fp16)[name = tensor("square_4_cast_fp16")]; + tensor reduce_mean_9_axes_0 = const()[name = tensor("reduce_mean_9_axes_0"), val = tensor([-1])]; + tensor reduce_mean_9_keep_dims_0 = const()[name = tensor("reduce_mean_9_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_9_cast_fp16 = reduce_mean(axes = reduce_mean_9_axes_0, keep_dims = reduce_mean_9_keep_dims_0, x = square_4_cast_fp16)[name = tensor("reduce_mean_9_cast_fp16")]; + tensor var_230_to_fp16 = const()[name = tensor("op_230_to_fp16"), val = tensor(0x1.1p-20)]; + tensor var_231_cast_fp16 = add(x = reduce_mean_9_cast_fp16, y = var_230_to_fp16)[name = tensor("op_231_cast_fp16")]; + tensor var_232_cast_fp16 = sqrt(x = var_231_cast_fp16)[name = tensor("op_232_cast_fp16")]; + tensor x_19_cast_fp16 = real_div(x = sub_6_cast_fp16, y = var_232_cast_fp16)[name = tensor("x_19_cast_fp16")]; + tensor flow_net_res_blocks_2_in_ln_weight_to_fp16 = const()[name = tensor("flow_net_res_blocks_2_in_ln_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8973248)))]; + tensor var_234_cast_fp16 = mul(x = x_19_cast_fp16, y = flow_net_res_blocks_2_in_ln_weight_to_fp16)[name = tensor("op_234_cast_fp16")]; + tensor flow_net_res_blocks_2_in_ln_bias_to_fp16 = const()[name = tensor("flow_net_res_blocks_2_in_ln_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8974336)))]; + tensor x_21_cast_fp16 = add(x = var_234_cast_fp16, y = flow_net_res_blocks_2_in_ln_bias_to_fp16)[name = tensor("x_21_cast_fp16")]; + tensor var_236_promoted_to_fp16 = const()[name = tensor("op_236_promoted_to_fp16"), val = tensor(0x1p+0)]; + tensor var_237_cast_fp16 = add(x = var_220_cast_fp16_1, y = var_236_promoted_to_fp16)[name = tensor("op_237_cast_fp16")]; + tensor var_238_cast_fp16 = mul(x = x_21_cast_fp16, y = var_237_cast_fp16)[name = tensor("op_238_cast_fp16")]; + tensor input_33_cast_fp16 = add(x = var_238_cast_fp16, y = var_220_cast_fp16_0)[name = tensor("input_33_cast_fp16")]; + tensor flow_net_res_blocks_2_mlp_0_weight_to_fp16 = const()[name = tensor("flow_net_res_blocks_2_mlp_0_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8975424)))]; + tensor flow_net_res_blocks_2_mlp_0_bias_to_fp16 = const()[name = tensor("flow_net_res_blocks_2_mlp_0_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9499776)))]; + tensor linear_13_cast_fp16 = linear(bias = flow_net_res_blocks_2_mlp_0_bias_to_fp16, weight = flow_net_res_blocks_2_mlp_0_weight_to_fp16, x = input_33_cast_fp16)[name = tensor("linear_13_cast_fp16")]; + tensor input_37_cast_fp16 = silu(x = linear_13_cast_fp16)[name = tensor("input_37_cast_fp16")]; + tensor flow_net_res_blocks_2_mlp_2_weight_to_fp16 = const()[name = tensor("flow_net_res_blocks_2_mlp_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9500864)))]; + tensor flow_net_res_blocks_2_mlp_2_bias_to_fp16 = const()[name = tensor("flow_net_res_blocks_2_mlp_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(10025216)))]; + tensor linear_14_cast_fp16 = linear(bias = flow_net_res_blocks_2_mlp_2_bias_to_fp16, weight = flow_net_res_blocks_2_mlp_2_weight_to_fp16, x = input_37_cast_fp16)[name = tensor("linear_14_cast_fp16")]; + tensor var_249_cast_fp16 = mul(x = var_220_cast_fp16_2, y = linear_14_cast_fp16)[name = tensor("op_249_cast_fp16")]; + tensor x_23_cast_fp16 = add(x = x_17_cast_fp16, y = var_249_cast_fp16)[name = tensor("x_23_cast_fp16")]; + tensor flow_net_res_blocks_3_adaLN_modulation_1_weight_to_fp16 = const()[name = tensor("flow_net_res_blocks_3_adaLN_modulation_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(10026304)))]; + tensor flow_net_res_blocks_3_adaLN_modulation_1_bias_to_fp16 = const()[name = tensor("flow_net_res_blocks_3_adaLN_modulation_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(11599232)))]; + tensor linear_15_cast_fp16 = linear(bias = flow_net_res_blocks_3_adaLN_modulation_1_bias_to_fp16, weight = flow_net_res_blocks_3_adaLN_modulation_1_weight_to_fp16, x = input_15_cast_fp16)[name = tensor("linear_15_cast_fp16")]; + tensor var_259_split_sizes_0 = const()[name = tensor("op_259_split_sizes_0"), val = tensor([512, 512, 512])]; + tensor var_259_axis_0 = const()[name = tensor("op_259_axis_0"), val = tensor(-1)]; + tensor var_259_cast_fp16_0, tensor var_259_cast_fp16_1, tensor var_259_cast_fp16_2 = split(axis = var_259_axis_0, split_sizes = var_259_split_sizes_0, x = linear_15_cast_fp16)[name = tensor("op_259_cast_fp16")]; + tensor mean_7_axes_0 = const()[name = tensor("mean_7_axes_0"), val = tensor([-1])]; + tensor mean_7_keep_dims_0 = const()[name = tensor("mean_7_keep_dims_0"), val = tensor(true)]; + tensor mean_7_cast_fp16 = reduce_mean(axes = mean_7_axes_0, keep_dims = mean_7_keep_dims_0, x = x_23_cast_fp16)[name = tensor("mean_7_cast_fp16")]; + tensor sub_7_cast_fp16 = sub(x = x_23_cast_fp16, y = mean_7_cast_fp16)[name = tensor("sub_7_cast_fp16")]; + tensor square_5_cast_fp16 = square(x = sub_7_cast_fp16)[name = tensor("square_5_cast_fp16")]; + tensor reduce_mean_11_axes_0 = const()[name = tensor("reduce_mean_11_axes_0"), val = tensor([-1])]; + tensor reduce_mean_11_keep_dims_0 = const()[name = tensor("reduce_mean_11_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_11_cast_fp16 = reduce_mean(axes = reduce_mean_11_axes_0, keep_dims = reduce_mean_11_keep_dims_0, x = square_5_cast_fp16)[name = tensor("reduce_mean_11_cast_fp16")]; + tensor var_269_to_fp16 = const()[name = tensor("op_269_to_fp16"), val = tensor(0x1.1p-20)]; + tensor var_270_cast_fp16 = add(x = reduce_mean_11_cast_fp16, y = var_269_to_fp16)[name = tensor("op_270_cast_fp16")]; + tensor var_271_cast_fp16 = sqrt(x = var_270_cast_fp16)[name = tensor("op_271_cast_fp16")]; + tensor x_25_cast_fp16 = real_div(x = sub_7_cast_fp16, y = var_271_cast_fp16)[name = tensor("x_25_cast_fp16")]; + tensor flow_net_res_blocks_3_in_ln_weight_to_fp16 = const()[name = tensor("flow_net_res_blocks_3_in_ln_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(11602368)))]; + tensor var_273_cast_fp16 = mul(x = x_25_cast_fp16, y = flow_net_res_blocks_3_in_ln_weight_to_fp16)[name = tensor("op_273_cast_fp16")]; + tensor flow_net_res_blocks_3_in_ln_bias_to_fp16 = const()[name = tensor("flow_net_res_blocks_3_in_ln_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(11603456)))]; + tensor x_27_cast_fp16 = add(x = var_273_cast_fp16, y = flow_net_res_blocks_3_in_ln_bias_to_fp16)[name = tensor("x_27_cast_fp16")]; + tensor var_275_promoted_to_fp16 = const()[name = tensor("op_275_promoted_to_fp16"), val = tensor(0x1p+0)]; + tensor var_276_cast_fp16 = add(x = var_259_cast_fp16_1, y = var_275_promoted_to_fp16)[name = tensor("op_276_cast_fp16")]; + tensor var_277_cast_fp16 = mul(x = x_27_cast_fp16, y = var_276_cast_fp16)[name = tensor("op_277_cast_fp16")]; + tensor input_41_cast_fp16 = add(x = var_277_cast_fp16, y = var_259_cast_fp16_0)[name = tensor("input_41_cast_fp16")]; + tensor flow_net_res_blocks_3_mlp_0_weight_to_fp16 = const()[name = tensor("flow_net_res_blocks_3_mlp_0_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(11604544)))]; + tensor flow_net_res_blocks_3_mlp_0_bias_to_fp16 = const()[name = tensor("flow_net_res_blocks_3_mlp_0_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(12128896)))]; + tensor linear_16_cast_fp16 = linear(bias = flow_net_res_blocks_3_mlp_0_bias_to_fp16, weight = flow_net_res_blocks_3_mlp_0_weight_to_fp16, x = input_41_cast_fp16)[name = tensor("linear_16_cast_fp16")]; + tensor input_45_cast_fp16 = silu(x = linear_16_cast_fp16)[name = tensor("input_45_cast_fp16")]; + tensor flow_net_res_blocks_3_mlp_2_weight_to_fp16 = const()[name = tensor("flow_net_res_blocks_3_mlp_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(12129984)))]; + tensor flow_net_res_blocks_3_mlp_2_bias_to_fp16 = const()[name = tensor("flow_net_res_blocks_3_mlp_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(12654336)))]; + tensor linear_17_cast_fp16 = linear(bias = flow_net_res_blocks_3_mlp_2_bias_to_fp16, weight = flow_net_res_blocks_3_mlp_2_weight_to_fp16, x = input_45_cast_fp16)[name = tensor("linear_17_cast_fp16")]; + tensor var_288_cast_fp16 = mul(x = var_259_cast_fp16_2, y = linear_17_cast_fp16)[name = tensor("op_288_cast_fp16")]; + tensor x_29_cast_fp16 = add(x = x_23_cast_fp16, y = var_288_cast_fp16)[name = tensor("x_29_cast_fp16")]; + tensor flow_net_res_blocks_4_adaLN_modulation_1_weight_to_fp16 = const()[name = tensor("flow_net_res_blocks_4_adaLN_modulation_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(12655424)))]; + tensor flow_net_res_blocks_4_adaLN_modulation_1_bias_to_fp16 = const()[name = tensor("flow_net_res_blocks_4_adaLN_modulation_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(14228352)))]; + tensor linear_18_cast_fp16 = linear(bias = flow_net_res_blocks_4_adaLN_modulation_1_bias_to_fp16, weight = flow_net_res_blocks_4_adaLN_modulation_1_weight_to_fp16, x = input_15_cast_fp16)[name = tensor("linear_18_cast_fp16")]; + tensor var_298_split_sizes_0 = const()[name = tensor("op_298_split_sizes_0"), val = tensor([512, 512, 512])]; + tensor var_298_axis_0 = const()[name = tensor("op_298_axis_0"), val = tensor(-1)]; + tensor var_298_cast_fp16_0, tensor var_298_cast_fp16_1, tensor var_298_cast_fp16_2 = split(axis = var_298_axis_0, split_sizes = var_298_split_sizes_0, x = linear_18_cast_fp16)[name = tensor("op_298_cast_fp16")]; + tensor mean_9_axes_0 = const()[name = tensor("mean_9_axes_0"), val = tensor([-1])]; + tensor mean_9_keep_dims_0 = const()[name = tensor("mean_9_keep_dims_0"), val = tensor(true)]; + tensor mean_9_cast_fp16 = reduce_mean(axes = mean_9_axes_0, keep_dims = mean_9_keep_dims_0, x = x_29_cast_fp16)[name = tensor("mean_9_cast_fp16")]; + tensor sub_8_cast_fp16 = sub(x = x_29_cast_fp16, y = mean_9_cast_fp16)[name = tensor("sub_8_cast_fp16")]; + tensor square_6_cast_fp16 = square(x = sub_8_cast_fp16)[name = tensor("square_6_cast_fp16")]; + tensor reduce_mean_13_axes_0 = const()[name = tensor("reduce_mean_13_axes_0"), val = tensor([-1])]; + tensor reduce_mean_13_keep_dims_0 = const()[name = tensor("reduce_mean_13_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_13_cast_fp16 = reduce_mean(axes = reduce_mean_13_axes_0, keep_dims = reduce_mean_13_keep_dims_0, x = square_6_cast_fp16)[name = tensor("reduce_mean_13_cast_fp16")]; + tensor var_308_to_fp16 = const()[name = tensor("op_308_to_fp16"), val = tensor(0x1.1p-20)]; + tensor var_309_cast_fp16 = add(x = reduce_mean_13_cast_fp16, y = var_308_to_fp16)[name = tensor("op_309_cast_fp16")]; + tensor var_310_cast_fp16 = sqrt(x = var_309_cast_fp16)[name = tensor("op_310_cast_fp16")]; + tensor x_31_cast_fp16 = real_div(x = sub_8_cast_fp16, y = var_310_cast_fp16)[name = tensor("x_31_cast_fp16")]; + tensor flow_net_res_blocks_4_in_ln_weight_to_fp16 = const()[name = tensor("flow_net_res_blocks_4_in_ln_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(14231488)))]; + tensor var_312_cast_fp16 = mul(x = x_31_cast_fp16, y = flow_net_res_blocks_4_in_ln_weight_to_fp16)[name = tensor("op_312_cast_fp16")]; + tensor flow_net_res_blocks_4_in_ln_bias_to_fp16 = const()[name = tensor("flow_net_res_blocks_4_in_ln_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(14232576)))]; + tensor x_33_cast_fp16 = add(x = var_312_cast_fp16, y = flow_net_res_blocks_4_in_ln_bias_to_fp16)[name = tensor("x_33_cast_fp16")]; + tensor var_314_promoted_to_fp16 = const()[name = tensor("op_314_promoted_to_fp16"), val = tensor(0x1p+0)]; + tensor var_315_cast_fp16 = add(x = var_298_cast_fp16_1, y = var_314_promoted_to_fp16)[name = tensor("op_315_cast_fp16")]; + tensor var_316_cast_fp16 = mul(x = x_33_cast_fp16, y = var_315_cast_fp16)[name = tensor("op_316_cast_fp16")]; + tensor input_49_cast_fp16 = add(x = var_316_cast_fp16, y = var_298_cast_fp16_0)[name = tensor("input_49_cast_fp16")]; + tensor flow_net_res_blocks_4_mlp_0_weight_to_fp16 = const()[name = tensor("flow_net_res_blocks_4_mlp_0_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(14233664)))]; + tensor flow_net_res_blocks_4_mlp_0_bias_to_fp16 = const()[name = tensor("flow_net_res_blocks_4_mlp_0_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(14758016)))]; + tensor linear_19_cast_fp16 = linear(bias = flow_net_res_blocks_4_mlp_0_bias_to_fp16, weight = flow_net_res_blocks_4_mlp_0_weight_to_fp16, x = input_49_cast_fp16)[name = tensor("linear_19_cast_fp16")]; + tensor input_53_cast_fp16 = silu(x = linear_19_cast_fp16)[name = tensor("input_53_cast_fp16")]; + tensor flow_net_res_blocks_4_mlp_2_weight_to_fp16 = const()[name = tensor("flow_net_res_blocks_4_mlp_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(14759104)))]; + tensor flow_net_res_blocks_4_mlp_2_bias_to_fp16 = const()[name = tensor("flow_net_res_blocks_4_mlp_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15283456)))]; + tensor linear_20_cast_fp16 = linear(bias = flow_net_res_blocks_4_mlp_2_bias_to_fp16, weight = flow_net_res_blocks_4_mlp_2_weight_to_fp16, x = input_53_cast_fp16)[name = tensor("linear_20_cast_fp16")]; + tensor var_327_cast_fp16 = mul(x = var_298_cast_fp16_2, y = linear_20_cast_fp16)[name = tensor("op_327_cast_fp16")]; + tensor x_35_cast_fp16 = add(x = x_29_cast_fp16, y = var_327_cast_fp16)[name = tensor("x_35_cast_fp16")]; + tensor flow_net_res_blocks_5_adaLN_modulation_1_weight_to_fp16 = const()[name = tensor("flow_net_res_blocks_5_adaLN_modulation_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15284544)))]; + tensor flow_net_res_blocks_5_adaLN_modulation_1_bias_to_fp16 = const()[name = tensor("flow_net_res_blocks_5_adaLN_modulation_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(16857472)))]; + tensor linear_21_cast_fp16 = linear(bias = flow_net_res_blocks_5_adaLN_modulation_1_bias_to_fp16, weight = flow_net_res_blocks_5_adaLN_modulation_1_weight_to_fp16, x = input_15_cast_fp16)[name = tensor("linear_21_cast_fp16")]; + tensor var_337_split_sizes_0 = const()[name = tensor("op_337_split_sizes_0"), val = tensor([512, 512, 512])]; + tensor var_337_axis_0 = const()[name = tensor("op_337_axis_0"), val = tensor(-1)]; + tensor var_337_cast_fp16_0, tensor var_337_cast_fp16_1, tensor var_337_cast_fp16_2 = split(axis = var_337_axis_0, split_sizes = var_337_split_sizes_0, x = linear_21_cast_fp16)[name = tensor("op_337_cast_fp16")]; + tensor mean_11_axes_0 = const()[name = tensor("mean_11_axes_0"), val = tensor([-1])]; + tensor mean_11_keep_dims_0 = const()[name = tensor("mean_11_keep_dims_0"), val = tensor(true)]; + tensor mean_11_cast_fp16 = reduce_mean(axes = mean_11_axes_0, keep_dims = mean_11_keep_dims_0, x = x_35_cast_fp16)[name = tensor("mean_11_cast_fp16")]; + tensor sub_9_cast_fp16 = sub(x = x_35_cast_fp16, y = mean_11_cast_fp16)[name = tensor("sub_9_cast_fp16")]; + tensor square_7_cast_fp16 = square(x = sub_9_cast_fp16)[name = tensor("square_7_cast_fp16")]; + tensor reduce_mean_15_axes_0 = const()[name = tensor("reduce_mean_15_axes_0"), val = tensor([-1])]; + tensor reduce_mean_15_keep_dims_0 = const()[name = tensor("reduce_mean_15_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_15_cast_fp16 = reduce_mean(axes = reduce_mean_15_axes_0, keep_dims = reduce_mean_15_keep_dims_0, x = square_7_cast_fp16)[name = tensor("reduce_mean_15_cast_fp16")]; + tensor var_347_to_fp16 = const()[name = tensor("op_347_to_fp16"), val = tensor(0x1.1p-20)]; + tensor var_348_cast_fp16 = add(x = reduce_mean_15_cast_fp16, y = var_347_to_fp16)[name = tensor("op_348_cast_fp16")]; + tensor var_349_cast_fp16 = sqrt(x = var_348_cast_fp16)[name = tensor("op_349_cast_fp16")]; + tensor x_37_cast_fp16 = real_div(x = sub_9_cast_fp16, y = var_349_cast_fp16)[name = tensor("x_37_cast_fp16")]; + tensor flow_net_res_blocks_5_in_ln_weight_to_fp16 = const()[name = tensor("flow_net_res_blocks_5_in_ln_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(16860608)))]; + tensor var_351_cast_fp16 = mul(x = x_37_cast_fp16, y = flow_net_res_blocks_5_in_ln_weight_to_fp16)[name = tensor("op_351_cast_fp16")]; + tensor flow_net_res_blocks_5_in_ln_bias_to_fp16 = const()[name = tensor("flow_net_res_blocks_5_in_ln_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(16861696)))]; + tensor x_39_cast_fp16 = add(x = var_351_cast_fp16, y = flow_net_res_blocks_5_in_ln_bias_to_fp16)[name = tensor("x_39_cast_fp16")]; + tensor var_353_promoted_to_fp16 = const()[name = tensor("op_353_promoted_to_fp16"), val = tensor(0x1p+0)]; + tensor var_354_cast_fp16 = add(x = var_337_cast_fp16_1, y = var_353_promoted_to_fp16)[name = tensor("op_354_cast_fp16")]; + tensor var_355_cast_fp16 = mul(x = x_39_cast_fp16, y = var_354_cast_fp16)[name = tensor("op_355_cast_fp16")]; + tensor input_57_cast_fp16 = add(x = var_355_cast_fp16, y = var_337_cast_fp16_0)[name = tensor("input_57_cast_fp16")]; + tensor flow_net_res_blocks_5_mlp_0_weight_to_fp16 = const()[name = tensor("flow_net_res_blocks_5_mlp_0_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(16862784)))]; + tensor flow_net_res_blocks_5_mlp_0_bias_to_fp16 = const()[name = tensor("flow_net_res_blocks_5_mlp_0_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(17387136)))]; + tensor linear_22_cast_fp16 = linear(bias = flow_net_res_blocks_5_mlp_0_bias_to_fp16, weight = flow_net_res_blocks_5_mlp_0_weight_to_fp16, x = input_57_cast_fp16)[name = tensor("linear_22_cast_fp16")]; + tensor input_61_cast_fp16 = silu(x = linear_22_cast_fp16)[name = tensor("input_61_cast_fp16")]; + tensor flow_net_res_blocks_5_mlp_2_weight_to_fp16 = const()[name = tensor("flow_net_res_blocks_5_mlp_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(17388224)))]; + tensor flow_net_res_blocks_5_mlp_2_bias_to_fp16 = const()[name = tensor("flow_net_res_blocks_5_mlp_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(17912576)))]; + tensor linear_23_cast_fp16 = linear(bias = flow_net_res_blocks_5_mlp_2_bias_to_fp16, weight = flow_net_res_blocks_5_mlp_2_weight_to_fp16, x = input_61_cast_fp16)[name = tensor("linear_23_cast_fp16")]; + tensor var_366_cast_fp16 = mul(x = var_337_cast_fp16_2, y = linear_23_cast_fp16)[name = tensor("op_366_cast_fp16")]; + tensor x_41_cast_fp16 = add(x = x_35_cast_fp16, y = var_366_cast_fp16)[name = tensor("x_41_cast_fp16")]; + tensor flow_net_final_layer_adaLN_modulation_1_weight_to_fp16 = const()[name = tensor("flow_net_final_layer_adaLN_modulation_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(17913664)))]; + tensor flow_net_final_layer_adaLN_modulation_1_bias_to_fp16 = const()[name = tensor("flow_net_final_layer_adaLN_modulation_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(18962304)))]; + tensor linear_24_cast_fp16 = linear(bias = flow_net_final_layer_adaLN_modulation_1_bias_to_fp16, weight = flow_net_final_layer_adaLN_modulation_1_weight_to_fp16, x = input_15_cast_fp16)[name = tensor("linear_24_cast_fp16")]; + tensor var_375_split_sizes_0 = const()[name = tensor("op_375_split_sizes_0"), val = tensor([512, 512])]; + tensor var_375_axis_0 = const()[name = tensor("op_375_axis_0"), val = tensor(-1)]; + tensor var_375_cast_fp16_0, tensor var_375_cast_fp16_1 = split(axis = var_375_axis_0, split_sizes = var_375_split_sizes_0, x = linear_24_cast_fp16)[name = tensor("op_375_cast_fp16")]; + tensor mean_13_axes_0 = const()[name = tensor("mean_13_axes_0"), val = tensor([-1])]; + tensor mean_13_keep_dims_0 = const()[name = tensor("mean_13_keep_dims_0"), val = tensor(true)]; + tensor mean_13_cast_fp16 = reduce_mean(axes = mean_13_axes_0, keep_dims = mean_13_keep_dims_0, x = x_41_cast_fp16)[name = tensor("mean_13_cast_fp16")]; + tensor sub_10_cast_fp16 = sub(x = x_41_cast_fp16, y = mean_13_cast_fp16)[name = tensor("sub_10_cast_fp16")]; + tensor square_8_cast_fp16 = square(x = sub_10_cast_fp16)[name = tensor("square_8_cast_fp16")]; + tensor reduce_mean_17_axes_0 = const()[name = tensor("reduce_mean_17_axes_0"), val = tensor([-1])]; + tensor reduce_mean_17_keep_dims_0 = const()[name = tensor("reduce_mean_17_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_17_cast_fp16 = reduce_mean(axes = reduce_mean_17_axes_0, keep_dims = reduce_mean_17_keep_dims_0, x = square_8_cast_fp16)[name = tensor("reduce_mean_17_cast_fp16")]; + tensor var_382_to_fp16 = const()[name = tensor("op_382_to_fp16"), val = tensor(0x1.1p-20)]; + tensor var_383_cast_fp16 = add(x = reduce_mean_17_cast_fp16, y = var_382_to_fp16)[name = tensor("op_383_cast_fp16")]; + tensor var_384_cast_fp16 = sqrt(x = var_383_cast_fp16)[name = tensor("op_384_cast_fp16")]; + tensor x_43_cast_fp16 = real_div(x = sub_10_cast_fp16, y = var_384_cast_fp16)[name = tensor("x_43_cast_fp16")]; + tensor var_386_promoted_to_fp16 = const()[name = tensor("op_386_promoted_to_fp16"), val = tensor(0x1p+0)]; + tensor var_387_cast_fp16 = add(x = var_375_cast_fp16_1, y = var_386_promoted_to_fp16)[name = tensor("op_387_cast_fp16")]; + tensor var_388_cast_fp16 = mul(x = x_43_cast_fp16, y = var_387_cast_fp16)[name = tensor("op_388_cast_fp16")]; + tensor input_65_cast_fp16 = add(x = var_388_cast_fp16, y = var_375_cast_fp16_0)[name = tensor("input_65_cast_fp16")]; + tensor flow_net_final_layer_linear_weight_to_fp16 = const()[name = tensor("flow_net_final_layer_linear_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(18964416)))]; + tensor flow_net_final_layer_linear_bias_to_fp16 = const()[name = tensor("flow_net_final_layer_linear_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(18997248)))]; + tensor linear_25_cast_fp16 = linear(bias = flow_net_final_layer_linear_bias_to_fp16, weight = flow_net_final_layer_linear_weight_to_fp16, x = input_65_cast_fp16)[name = tensor("linear_25_cast_fp16")]; + tensor var_399_to_fp16 = const()[name = tensor("op_399_to_fp16"), val = tensor(0x1p-3)]; + tensor var_400_cast_fp16 = mul(x = linear_25_cast_fp16, y = var_399_to_fp16)[name = tensor("op_400_cast_fp16")]; + tensor input_67_cast_fp16 = add(x = latent_init_to_fp16, y = var_400_cast_fp16)[name = tensor("input_67_cast_fp16")]; + tensor linear_26_cast_fp16 = linear(bias = flow_net_input_proj_bias_to_fp16, weight = flow_net_input_proj_weight_to_fp16, x = input_67_cast_fp16)[name = tensor("linear_26_cast_fp16")]; + tensor input_71_to_fp16 = const()[name = tensor("input_71_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(18997376)))]; + tensor input_73_cast_fp16 = silu(x = input_71_to_fp16)[name = tensor("input_73_cast_fp16")]; + tensor linear_28_cast_fp16 = linear(bias = flow_net_time_embed_0_mlp_2_bias_to_fp16, weight = flow_net_time_embed_0_mlp_2_weight_to_fp16, x = input_73_cast_fp16)[name = tensor("linear_28_cast_fp16")]; + tensor reduce_mean_18_axes_0 = const()[name = tensor("reduce_mean_18_axes_0"), val = tensor([-1])]; + tensor reduce_mean_18_keep_dims_0 = const()[name = tensor("reduce_mean_18_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_18_cast_fp16 = reduce_mean(axes = reduce_mean_18_axes_0, keep_dims = reduce_mean_18_keep_dims_0, x = linear_28_cast_fp16)[name = tensor("reduce_mean_18_cast_fp16")]; + tensor sub_11_cast_fp16 = sub(x = linear_28_cast_fp16, y = reduce_mean_18_cast_fp16)[name = tensor("sub_11_cast_fp16")]; + tensor square_9_cast_fp16 = square(x = sub_11_cast_fp16)[name = tensor("square_9_cast_fp16")]; + tensor reduce_mean_19_axes_0 = const()[name = tensor("reduce_mean_19_axes_0"), val = tensor([-1])]; + tensor reduce_mean_19_keep_dims_0 = const()[name = tensor("reduce_mean_19_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_19_cast_fp16 = reduce_mean(axes = reduce_mean_19_axes_0, keep_dims = reduce_mean_19_keep_dims_0, x = square_9_cast_fp16)[name = tensor("reduce_mean_19_cast_fp16")]; + tensor real_div_2_to_fp16 = const()[name = tensor("real_div_2_to_fp16"), val = tensor(0x1.008p+0)]; + tensor mul_2_cast_fp16 = mul(x = reduce_mean_19_cast_fp16, y = real_div_2_to_fp16)[name = tensor("mul_2_cast_fp16")]; + tensor var_466_to_fp16 = const()[name = tensor("op_466_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_19_cast_fp16 = add(x = mul_2_cast_fp16, y = var_466_to_fp16)[name = tensor("var_19_cast_fp16")]; + tensor var_469_epsilon_0 = const()[name = tensor("op_469_epsilon_0"), val = tensor(0x1.197998p-40)]; + tensor var_469_cast_fp16 = rsqrt(epsilon = var_469_epsilon_0, x = var_19_cast_fp16)[name = tensor("op_469_cast_fp16")]; + tensor var_470_cast_fp16 = mul(x = const_3_to_fp16, y = var_469_cast_fp16)[name = tensor("op_470_cast_fp16")]; + tensor var_471_cast_fp16 = mul(x = linear_28_cast_fp16, y = var_470_cast_fp16)[name = tensor("op_471_cast_fp16")]; + tensor input_77_to_fp16 = const()[name = tensor("input_77_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(18998464)))]; + tensor input_79_cast_fp16 = silu(x = input_77_to_fp16)[name = tensor("input_79_cast_fp16")]; + tensor linear_30_cast_fp16 = linear(bias = flow_net_time_embed_1_mlp_2_bias_to_fp16, weight = flow_net_time_embed_1_mlp_2_weight_to_fp16, x = input_79_cast_fp16)[name = tensor("linear_30_cast_fp16")]; + tensor reduce_mean_20_axes_0 = const()[name = tensor("reduce_mean_20_axes_0"), val = tensor([-1])]; + tensor reduce_mean_20_keep_dims_0 = const()[name = tensor("reduce_mean_20_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_20_cast_fp16 = reduce_mean(axes = reduce_mean_20_axes_0, keep_dims = reduce_mean_20_keep_dims_0, x = linear_30_cast_fp16)[name = tensor("reduce_mean_20_cast_fp16")]; + tensor sub_13_cast_fp16 = sub(x = linear_30_cast_fp16, y = reduce_mean_20_cast_fp16)[name = tensor("sub_13_cast_fp16")]; + tensor square_10_cast_fp16 = square(x = sub_13_cast_fp16)[name = tensor("square_10_cast_fp16")]; + tensor reduce_mean_21_axes_0 = const()[name = tensor("reduce_mean_21_axes_0"), val = tensor([-1])]; + tensor reduce_mean_21_keep_dims_0 = const()[name = tensor("reduce_mean_21_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_21_cast_fp16 = reduce_mean(axes = reduce_mean_21_axes_0, keep_dims = reduce_mean_21_keep_dims_0, x = square_10_cast_fp16)[name = tensor("reduce_mean_21_cast_fp16")]; + tensor real_div_3_to_fp16 = const()[name = tensor("real_div_3_to_fp16"), val = tensor(0x1.008p+0)]; + tensor mul_3_cast_fp16 = mul(x = reduce_mean_21_cast_fp16, y = real_div_3_to_fp16)[name = tensor("mul_3_cast_fp16")]; + tensor var_503_to_fp16 = const()[name = tensor("op_503_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_21_cast_fp16 = add(x = mul_3_cast_fp16, y = var_503_to_fp16)[name = tensor("var_21_cast_fp16")]; + tensor var_506_epsilon_0 = const()[name = tensor("op_506_epsilon_0"), val = tensor(0x1.197998p-40)]; + tensor var_506_cast_fp16 = rsqrt(epsilon = var_506_epsilon_0, x = var_21_cast_fp16)[name = tensor("op_506_cast_fp16")]; + tensor var_507_cast_fp16 = mul(x = const_5_to_fp16, y = var_506_cast_fp16)[name = tensor("op_507_cast_fp16")]; + tensor var_508_cast_fp16 = mul(x = linear_30_cast_fp16, y = var_507_cast_fp16)[name = tensor("op_508_cast_fp16")]; + tensor var_520_cast_fp16 = add(x = var_471_cast_fp16, y = var_508_cast_fp16)[name = tensor("op_520_cast_fp16")]; + tensor _inversed_t_combined_3_y_0_to_fp16 = const()[name = tensor("_inversed_t_combined_3_y_0_to_fp16"), val = tensor(0x1p-1)]; + tensor _inversed_t_combined_3_cast_fp16 = mul(x = var_520_cast_fp16, y = _inversed_t_combined_3_y_0_to_fp16)[name = tensor("_inversed_t_combined_3_cast_fp16")]; + tensor input_81_cast_fp16 = add(x = _inversed_t_combined_3_cast_fp16, y = linear_5_cast_fp16)[name = tensor("input_81_cast_fp16")]; + tensor input_83_cast_fp16 = silu(x = input_81_cast_fp16)[name = tensor("input_83_cast_fp16")]; + tensor linear_32_cast_fp16 = linear(bias = flow_net_res_blocks_0_adaLN_modulation_1_bias_to_fp16, weight = flow_net_res_blocks_0_adaLN_modulation_1_weight_to_fp16, x = input_83_cast_fp16)[name = tensor("linear_32_cast_fp16")]; + tensor var_535_split_sizes_0 = const()[name = tensor("op_535_split_sizes_0"), val = tensor([512, 512, 512])]; + tensor var_535_axis_0 = const()[name = tensor("op_535_axis_0"), val = tensor(-1)]; + tensor var_535_cast_fp16_0, tensor var_535_cast_fp16_1, tensor var_535_cast_fp16_2 = split(axis = var_535_axis_0, split_sizes = var_535_split_sizes_0, x = linear_32_cast_fp16)[name = tensor("op_535_cast_fp16")]; + tensor mean_15_axes_0 = const()[name = tensor("mean_15_axes_0"), val = tensor([-1])]; + tensor mean_15_keep_dims_0 = const()[name = tensor("mean_15_keep_dims_0"), val = tensor(true)]; + tensor mean_15_cast_fp16 = reduce_mean(axes = mean_15_axes_0, keep_dims = mean_15_keep_dims_0, x = linear_26_cast_fp16)[name = tensor("mean_15_cast_fp16")]; + tensor sub_15_cast_fp16 = sub(x = linear_26_cast_fp16, y = mean_15_cast_fp16)[name = tensor("sub_15_cast_fp16")]; + tensor square_11_cast_fp16 = square(x = sub_15_cast_fp16)[name = tensor("square_11_cast_fp16")]; + tensor reduce_mean_23_axes_0 = const()[name = tensor("reduce_mean_23_axes_0"), val = tensor([-1])]; + tensor reduce_mean_23_keep_dims_0 = const()[name = tensor("reduce_mean_23_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_23_cast_fp16 = reduce_mean(axes = reduce_mean_23_axes_0, keep_dims = reduce_mean_23_keep_dims_0, x = square_11_cast_fp16)[name = tensor("reduce_mean_23_cast_fp16")]; + tensor var_545_to_fp16 = const()[name = tensor("op_545_to_fp16"), val = tensor(0x1.1p-20)]; + tensor var_546_cast_fp16 = add(x = reduce_mean_23_cast_fp16, y = var_545_to_fp16)[name = tensor("op_546_cast_fp16")]; + tensor var_547_cast_fp16 = sqrt(x = var_546_cast_fp16)[name = tensor("op_547_cast_fp16")]; + tensor x_51_cast_fp16 = real_div(x = sub_15_cast_fp16, y = var_547_cast_fp16)[name = tensor("x_51_cast_fp16")]; + tensor var_549_cast_fp16 = mul(x = x_51_cast_fp16, y = flow_net_res_blocks_0_in_ln_weight_to_fp16)[name = tensor("op_549_cast_fp16")]; + tensor x_53_cast_fp16 = add(x = var_549_cast_fp16, y = flow_net_res_blocks_0_in_ln_bias_to_fp16)[name = tensor("x_53_cast_fp16")]; + tensor var_551_promoted_to_fp16 = const()[name = tensor("op_551_promoted_to_fp16"), val = tensor(0x1p+0)]; + tensor var_552_cast_fp16 = add(x = var_535_cast_fp16_1, y = var_551_promoted_to_fp16)[name = tensor("op_552_cast_fp16")]; + tensor var_553_cast_fp16 = mul(x = x_53_cast_fp16, y = var_552_cast_fp16)[name = tensor("op_553_cast_fp16")]; + tensor input_85_cast_fp16 = add(x = var_553_cast_fp16, y = var_535_cast_fp16_0)[name = tensor("input_85_cast_fp16")]; + tensor linear_33_cast_fp16 = linear(bias = flow_net_res_blocks_0_mlp_0_bias_to_fp16, weight = flow_net_res_blocks_0_mlp_0_weight_to_fp16, x = input_85_cast_fp16)[name = tensor("linear_33_cast_fp16")]; + tensor input_89_cast_fp16 = silu(x = linear_33_cast_fp16)[name = tensor("input_89_cast_fp16")]; + tensor linear_34_cast_fp16 = linear(bias = flow_net_res_blocks_0_mlp_2_bias_to_fp16, weight = flow_net_res_blocks_0_mlp_2_weight_to_fp16, x = input_89_cast_fp16)[name = tensor("linear_34_cast_fp16")]; + tensor var_564_cast_fp16 = mul(x = var_535_cast_fp16_2, y = linear_34_cast_fp16)[name = tensor("op_564_cast_fp16")]; + tensor x_55_cast_fp16 = add(x = linear_26_cast_fp16, y = var_564_cast_fp16)[name = tensor("x_55_cast_fp16")]; + tensor linear_35_cast_fp16 = linear(bias = flow_net_res_blocks_1_adaLN_modulation_1_bias_to_fp16, weight = flow_net_res_blocks_1_adaLN_modulation_1_weight_to_fp16, x = input_83_cast_fp16)[name = tensor("linear_35_cast_fp16")]; + tensor var_574_split_sizes_0 = const()[name = tensor("op_574_split_sizes_0"), val = tensor([512, 512, 512])]; + tensor var_574_axis_0 = const()[name = tensor("op_574_axis_0"), val = tensor(-1)]; + tensor var_574_cast_fp16_0, tensor var_574_cast_fp16_1, tensor var_574_cast_fp16_2 = split(axis = var_574_axis_0, split_sizes = var_574_split_sizes_0, x = linear_35_cast_fp16)[name = tensor("op_574_cast_fp16")]; + tensor mean_17_axes_0 = const()[name = tensor("mean_17_axes_0"), val = tensor([-1])]; + tensor mean_17_keep_dims_0 = const()[name = tensor("mean_17_keep_dims_0"), val = tensor(true)]; + tensor mean_17_cast_fp16 = reduce_mean(axes = mean_17_axes_0, keep_dims = mean_17_keep_dims_0, x = x_55_cast_fp16)[name = tensor("mean_17_cast_fp16")]; + tensor sub_16_cast_fp16 = sub(x = x_55_cast_fp16, y = mean_17_cast_fp16)[name = tensor("sub_16_cast_fp16")]; + tensor square_12_cast_fp16 = square(x = sub_16_cast_fp16)[name = tensor("square_12_cast_fp16")]; + tensor reduce_mean_25_axes_0 = const()[name = tensor("reduce_mean_25_axes_0"), val = tensor([-1])]; + tensor reduce_mean_25_keep_dims_0 = const()[name = tensor("reduce_mean_25_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_25_cast_fp16 = reduce_mean(axes = reduce_mean_25_axes_0, keep_dims = reduce_mean_25_keep_dims_0, x = square_12_cast_fp16)[name = tensor("reduce_mean_25_cast_fp16")]; + tensor var_584_to_fp16 = const()[name = tensor("op_584_to_fp16"), val = tensor(0x1.1p-20)]; + tensor var_585_cast_fp16 = add(x = reduce_mean_25_cast_fp16, y = var_584_to_fp16)[name = tensor("op_585_cast_fp16")]; + tensor var_586_cast_fp16 = sqrt(x = var_585_cast_fp16)[name = tensor("op_586_cast_fp16")]; + tensor x_57_cast_fp16 = real_div(x = sub_16_cast_fp16, y = var_586_cast_fp16)[name = tensor("x_57_cast_fp16")]; + tensor var_588_cast_fp16 = mul(x = x_57_cast_fp16, y = flow_net_res_blocks_1_in_ln_weight_to_fp16)[name = tensor("op_588_cast_fp16")]; + tensor x_59_cast_fp16 = add(x = var_588_cast_fp16, y = flow_net_res_blocks_1_in_ln_bias_to_fp16)[name = tensor("x_59_cast_fp16")]; + tensor var_590_promoted_to_fp16 = const()[name = tensor("op_590_promoted_to_fp16"), val = tensor(0x1p+0)]; + tensor var_591_cast_fp16 = add(x = var_574_cast_fp16_1, y = var_590_promoted_to_fp16)[name = tensor("op_591_cast_fp16")]; + tensor var_592_cast_fp16 = mul(x = x_59_cast_fp16, y = var_591_cast_fp16)[name = tensor("op_592_cast_fp16")]; + tensor input_93_cast_fp16 = add(x = var_592_cast_fp16, y = var_574_cast_fp16_0)[name = tensor("input_93_cast_fp16")]; + tensor linear_36_cast_fp16 = linear(bias = flow_net_res_blocks_1_mlp_0_bias_to_fp16, weight = flow_net_res_blocks_1_mlp_0_weight_to_fp16, x = input_93_cast_fp16)[name = tensor("linear_36_cast_fp16")]; + tensor input_97_cast_fp16 = silu(x = linear_36_cast_fp16)[name = tensor("input_97_cast_fp16")]; + tensor linear_37_cast_fp16 = linear(bias = flow_net_res_blocks_1_mlp_2_bias_to_fp16, weight = flow_net_res_blocks_1_mlp_2_weight_to_fp16, x = input_97_cast_fp16)[name = tensor("linear_37_cast_fp16")]; + tensor var_603_cast_fp16 = mul(x = var_574_cast_fp16_2, y = linear_37_cast_fp16)[name = tensor("op_603_cast_fp16")]; + tensor x_61_cast_fp16 = add(x = x_55_cast_fp16, y = var_603_cast_fp16)[name = tensor("x_61_cast_fp16")]; + tensor linear_38_cast_fp16 = linear(bias = flow_net_res_blocks_2_adaLN_modulation_1_bias_to_fp16, weight = flow_net_res_blocks_2_adaLN_modulation_1_weight_to_fp16, x = input_83_cast_fp16)[name = tensor("linear_38_cast_fp16")]; + tensor var_613_split_sizes_0 = const()[name = tensor("op_613_split_sizes_0"), val = tensor([512, 512, 512])]; + tensor var_613_axis_0 = const()[name = tensor("op_613_axis_0"), val = tensor(-1)]; + tensor var_613_cast_fp16_0, tensor var_613_cast_fp16_1, tensor var_613_cast_fp16_2 = split(axis = var_613_axis_0, split_sizes = var_613_split_sizes_0, x = linear_38_cast_fp16)[name = tensor("op_613_cast_fp16")]; + tensor mean_19_axes_0 = const()[name = tensor("mean_19_axes_0"), val = tensor([-1])]; + tensor mean_19_keep_dims_0 = const()[name = tensor("mean_19_keep_dims_0"), val = tensor(true)]; + tensor mean_19_cast_fp16 = reduce_mean(axes = mean_19_axes_0, keep_dims = mean_19_keep_dims_0, x = x_61_cast_fp16)[name = tensor("mean_19_cast_fp16")]; + tensor sub_17_cast_fp16 = sub(x = x_61_cast_fp16, y = mean_19_cast_fp16)[name = tensor("sub_17_cast_fp16")]; + tensor square_13_cast_fp16 = square(x = sub_17_cast_fp16)[name = tensor("square_13_cast_fp16")]; + tensor reduce_mean_27_axes_0 = const()[name = tensor("reduce_mean_27_axes_0"), val = tensor([-1])]; + tensor reduce_mean_27_keep_dims_0 = const()[name = tensor("reduce_mean_27_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_27_cast_fp16 = reduce_mean(axes = reduce_mean_27_axes_0, keep_dims = reduce_mean_27_keep_dims_0, x = square_13_cast_fp16)[name = tensor("reduce_mean_27_cast_fp16")]; + tensor var_623_to_fp16 = const()[name = tensor("op_623_to_fp16"), val = tensor(0x1.1p-20)]; + tensor var_624_cast_fp16 = add(x = reduce_mean_27_cast_fp16, y = var_623_to_fp16)[name = tensor("op_624_cast_fp16")]; + tensor var_625_cast_fp16 = sqrt(x = var_624_cast_fp16)[name = tensor("op_625_cast_fp16")]; + tensor x_63_cast_fp16 = real_div(x = sub_17_cast_fp16, y = var_625_cast_fp16)[name = tensor("x_63_cast_fp16")]; + tensor var_627_cast_fp16 = mul(x = x_63_cast_fp16, y = flow_net_res_blocks_2_in_ln_weight_to_fp16)[name = tensor("op_627_cast_fp16")]; + tensor x_65_cast_fp16 = add(x = var_627_cast_fp16, y = flow_net_res_blocks_2_in_ln_bias_to_fp16)[name = tensor("x_65_cast_fp16")]; + tensor var_629_promoted_to_fp16 = const()[name = tensor("op_629_promoted_to_fp16"), val = tensor(0x1p+0)]; + tensor var_630_cast_fp16 = add(x = var_613_cast_fp16_1, y = var_629_promoted_to_fp16)[name = tensor("op_630_cast_fp16")]; + tensor var_631_cast_fp16 = mul(x = x_65_cast_fp16, y = var_630_cast_fp16)[name = tensor("op_631_cast_fp16")]; + tensor input_101_cast_fp16 = add(x = var_631_cast_fp16, y = var_613_cast_fp16_0)[name = tensor("input_101_cast_fp16")]; + tensor linear_39_cast_fp16 = linear(bias = flow_net_res_blocks_2_mlp_0_bias_to_fp16, weight = flow_net_res_blocks_2_mlp_0_weight_to_fp16, x = input_101_cast_fp16)[name = tensor("linear_39_cast_fp16")]; + tensor input_105_cast_fp16 = silu(x = linear_39_cast_fp16)[name = tensor("input_105_cast_fp16")]; + tensor linear_40_cast_fp16 = linear(bias = flow_net_res_blocks_2_mlp_2_bias_to_fp16, weight = flow_net_res_blocks_2_mlp_2_weight_to_fp16, x = input_105_cast_fp16)[name = tensor("linear_40_cast_fp16")]; + tensor var_642_cast_fp16 = mul(x = var_613_cast_fp16_2, y = linear_40_cast_fp16)[name = tensor("op_642_cast_fp16")]; + tensor x_67_cast_fp16 = add(x = x_61_cast_fp16, y = var_642_cast_fp16)[name = tensor("x_67_cast_fp16")]; + tensor linear_41_cast_fp16 = linear(bias = flow_net_res_blocks_3_adaLN_modulation_1_bias_to_fp16, weight = flow_net_res_blocks_3_adaLN_modulation_1_weight_to_fp16, x = input_83_cast_fp16)[name = tensor("linear_41_cast_fp16")]; + tensor var_652_split_sizes_0 = const()[name = tensor("op_652_split_sizes_0"), val = tensor([512, 512, 512])]; + tensor var_652_axis_0 = const()[name = tensor("op_652_axis_0"), val = tensor(-1)]; + tensor var_652_cast_fp16_0, tensor var_652_cast_fp16_1, tensor var_652_cast_fp16_2 = split(axis = var_652_axis_0, split_sizes = var_652_split_sizes_0, x = linear_41_cast_fp16)[name = tensor("op_652_cast_fp16")]; + tensor mean_21_axes_0 = const()[name = tensor("mean_21_axes_0"), val = tensor([-1])]; + tensor mean_21_keep_dims_0 = const()[name = tensor("mean_21_keep_dims_0"), val = tensor(true)]; + tensor mean_21_cast_fp16 = reduce_mean(axes = mean_21_axes_0, keep_dims = mean_21_keep_dims_0, x = x_67_cast_fp16)[name = tensor("mean_21_cast_fp16")]; + tensor sub_18_cast_fp16 = sub(x = x_67_cast_fp16, y = mean_21_cast_fp16)[name = tensor("sub_18_cast_fp16")]; + tensor square_14_cast_fp16 = square(x = sub_18_cast_fp16)[name = tensor("square_14_cast_fp16")]; + tensor reduce_mean_29_axes_0 = const()[name = tensor("reduce_mean_29_axes_0"), val = tensor([-1])]; + tensor reduce_mean_29_keep_dims_0 = const()[name = tensor("reduce_mean_29_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_29_cast_fp16 = reduce_mean(axes = reduce_mean_29_axes_0, keep_dims = reduce_mean_29_keep_dims_0, x = square_14_cast_fp16)[name = tensor("reduce_mean_29_cast_fp16")]; + tensor var_662_to_fp16 = const()[name = tensor("op_662_to_fp16"), val = tensor(0x1.1p-20)]; + tensor var_663_cast_fp16 = add(x = reduce_mean_29_cast_fp16, y = var_662_to_fp16)[name = tensor("op_663_cast_fp16")]; + tensor var_664_cast_fp16 = sqrt(x = var_663_cast_fp16)[name = tensor("op_664_cast_fp16")]; + tensor x_69_cast_fp16 = real_div(x = sub_18_cast_fp16, y = var_664_cast_fp16)[name = tensor("x_69_cast_fp16")]; + tensor var_666_cast_fp16 = mul(x = x_69_cast_fp16, y = flow_net_res_blocks_3_in_ln_weight_to_fp16)[name = tensor("op_666_cast_fp16")]; + tensor x_71_cast_fp16 = add(x = var_666_cast_fp16, y = flow_net_res_blocks_3_in_ln_bias_to_fp16)[name = tensor("x_71_cast_fp16")]; + tensor var_668_promoted_to_fp16 = const()[name = tensor("op_668_promoted_to_fp16"), val = tensor(0x1p+0)]; + tensor var_669_cast_fp16 = add(x = var_652_cast_fp16_1, y = var_668_promoted_to_fp16)[name = tensor("op_669_cast_fp16")]; + tensor var_670_cast_fp16 = mul(x = x_71_cast_fp16, y = var_669_cast_fp16)[name = tensor("op_670_cast_fp16")]; + tensor input_109_cast_fp16 = add(x = var_670_cast_fp16, y = var_652_cast_fp16_0)[name = tensor("input_109_cast_fp16")]; + tensor linear_42_cast_fp16 = linear(bias = flow_net_res_blocks_3_mlp_0_bias_to_fp16, weight = flow_net_res_blocks_3_mlp_0_weight_to_fp16, x = input_109_cast_fp16)[name = tensor("linear_42_cast_fp16")]; + tensor input_113_cast_fp16 = silu(x = linear_42_cast_fp16)[name = tensor("input_113_cast_fp16")]; + tensor linear_43_cast_fp16 = linear(bias = flow_net_res_blocks_3_mlp_2_bias_to_fp16, weight = flow_net_res_blocks_3_mlp_2_weight_to_fp16, x = input_113_cast_fp16)[name = tensor("linear_43_cast_fp16")]; + tensor var_681_cast_fp16 = mul(x = var_652_cast_fp16_2, y = linear_43_cast_fp16)[name = tensor("op_681_cast_fp16")]; + tensor x_73_cast_fp16 = add(x = x_67_cast_fp16, y = var_681_cast_fp16)[name = tensor("x_73_cast_fp16")]; + tensor linear_44_cast_fp16 = linear(bias = flow_net_res_blocks_4_adaLN_modulation_1_bias_to_fp16, weight = flow_net_res_blocks_4_adaLN_modulation_1_weight_to_fp16, x = input_83_cast_fp16)[name = tensor("linear_44_cast_fp16")]; + tensor var_691_split_sizes_0 = const()[name = tensor("op_691_split_sizes_0"), val = tensor([512, 512, 512])]; + tensor var_691_axis_0 = const()[name = tensor("op_691_axis_0"), val = tensor(-1)]; + tensor var_691_cast_fp16_0, tensor var_691_cast_fp16_1, tensor var_691_cast_fp16_2 = split(axis = var_691_axis_0, split_sizes = var_691_split_sizes_0, x = linear_44_cast_fp16)[name = tensor("op_691_cast_fp16")]; + tensor mean_23_axes_0 = const()[name = tensor("mean_23_axes_0"), val = tensor([-1])]; + tensor mean_23_keep_dims_0 = const()[name = tensor("mean_23_keep_dims_0"), val = tensor(true)]; + tensor mean_23_cast_fp16 = reduce_mean(axes = mean_23_axes_0, keep_dims = mean_23_keep_dims_0, x = x_73_cast_fp16)[name = tensor("mean_23_cast_fp16")]; + tensor sub_19_cast_fp16 = sub(x = x_73_cast_fp16, y = mean_23_cast_fp16)[name = tensor("sub_19_cast_fp16")]; + tensor square_15_cast_fp16 = square(x = sub_19_cast_fp16)[name = tensor("square_15_cast_fp16")]; + tensor reduce_mean_31_axes_0 = const()[name = tensor("reduce_mean_31_axes_0"), val = tensor([-1])]; + tensor reduce_mean_31_keep_dims_0 = const()[name = tensor("reduce_mean_31_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_31_cast_fp16 = reduce_mean(axes = reduce_mean_31_axes_0, keep_dims = reduce_mean_31_keep_dims_0, x = square_15_cast_fp16)[name = tensor("reduce_mean_31_cast_fp16")]; + tensor var_701_to_fp16 = const()[name = tensor("op_701_to_fp16"), val = tensor(0x1.1p-20)]; + tensor var_702_cast_fp16 = add(x = reduce_mean_31_cast_fp16, y = var_701_to_fp16)[name = tensor("op_702_cast_fp16")]; + tensor var_703_cast_fp16 = sqrt(x = var_702_cast_fp16)[name = tensor("op_703_cast_fp16")]; + tensor x_75_cast_fp16 = real_div(x = sub_19_cast_fp16, y = var_703_cast_fp16)[name = tensor("x_75_cast_fp16")]; + tensor var_705_cast_fp16 = mul(x = x_75_cast_fp16, y = flow_net_res_blocks_4_in_ln_weight_to_fp16)[name = tensor("op_705_cast_fp16")]; + tensor x_77_cast_fp16 = add(x = var_705_cast_fp16, y = flow_net_res_blocks_4_in_ln_bias_to_fp16)[name = tensor("x_77_cast_fp16")]; + tensor var_707_promoted_to_fp16 = const()[name = tensor("op_707_promoted_to_fp16"), val = tensor(0x1p+0)]; + tensor var_708_cast_fp16 = add(x = var_691_cast_fp16_1, y = var_707_promoted_to_fp16)[name = tensor("op_708_cast_fp16")]; + tensor var_709_cast_fp16 = mul(x = x_77_cast_fp16, y = var_708_cast_fp16)[name = tensor("op_709_cast_fp16")]; + tensor input_117_cast_fp16 = add(x = var_709_cast_fp16, y = var_691_cast_fp16_0)[name = tensor("input_117_cast_fp16")]; + tensor linear_45_cast_fp16 = linear(bias = flow_net_res_blocks_4_mlp_0_bias_to_fp16, weight = flow_net_res_blocks_4_mlp_0_weight_to_fp16, x = input_117_cast_fp16)[name = tensor("linear_45_cast_fp16")]; + tensor input_121_cast_fp16 = silu(x = linear_45_cast_fp16)[name = tensor("input_121_cast_fp16")]; + tensor linear_46_cast_fp16 = linear(bias = flow_net_res_blocks_4_mlp_2_bias_to_fp16, weight = flow_net_res_blocks_4_mlp_2_weight_to_fp16, x = input_121_cast_fp16)[name = tensor("linear_46_cast_fp16")]; + tensor var_720_cast_fp16 = mul(x = var_691_cast_fp16_2, y = linear_46_cast_fp16)[name = tensor("op_720_cast_fp16")]; + tensor x_79_cast_fp16 = add(x = x_73_cast_fp16, y = var_720_cast_fp16)[name = tensor("x_79_cast_fp16")]; + tensor linear_47_cast_fp16 = linear(bias = flow_net_res_blocks_5_adaLN_modulation_1_bias_to_fp16, weight = flow_net_res_blocks_5_adaLN_modulation_1_weight_to_fp16, x = input_83_cast_fp16)[name = tensor("linear_47_cast_fp16")]; + tensor var_730_split_sizes_0 = const()[name = tensor("op_730_split_sizes_0"), val = tensor([512, 512, 512])]; + tensor var_730_axis_0 = const()[name = tensor("op_730_axis_0"), val = tensor(-1)]; + tensor var_730_cast_fp16_0, tensor var_730_cast_fp16_1, tensor var_730_cast_fp16_2 = split(axis = var_730_axis_0, split_sizes = var_730_split_sizes_0, x = linear_47_cast_fp16)[name = tensor("op_730_cast_fp16")]; + tensor mean_25_axes_0 = const()[name = tensor("mean_25_axes_0"), val = tensor([-1])]; + tensor mean_25_keep_dims_0 = const()[name = tensor("mean_25_keep_dims_0"), val = tensor(true)]; + tensor mean_25_cast_fp16 = reduce_mean(axes = mean_25_axes_0, keep_dims = mean_25_keep_dims_0, x = x_79_cast_fp16)[name = tensor("mean_25_cast_fp16")]; + tensor sub_20_cast_fp16 = sub(x = x_79_cast_fp16, y = mean_25_cast_fp16)[name = tensor("sub_20_cast_fp16")]; + tensor square_16_cast_fp16 = square(x = sub_20_cast_fp16)[name = tensor("square_16_cast_fp16")]; + tensor reduce_mean_33_axes_0 = const()[name = tensor("reduce_mean_33_axes_0"), val = tensor([-1])]; + tensor reduce_mean_33_keep_dims_0 = const()[name = tensor("reduce_mean_33_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_33_cast_fp16 = reduce_mean(axes = reduce_mean_33_axes_0, keep_dims = reduce_mean_33_keep_dims_0, x = square_16_cast_fp16)[name = tensor("reduce_mean_33_cast_fp16")]; + tensor var_740_to_fp16 = const()[name = tensor("op_740_to_fp16"), val = tensor(0x1.1p-20)]; + tensor var_741_cast_fp16 = add(x = reduce_mean_33_cast_fp16, y = var_740_to_fp16)[name = tensor("op_741_cast_fp16")]; + tensor var_742_cast_fp16 = sqrt(x = var_741_cast_fp16)[name = tensor("op_742_cast_fp16")]; + tensor x_81_cast_fp16 = real_div(x = sub_20_cast_fp16, y = var_742_cast_fp16)[name = tensor("x_81_cast_fp16")]; + tensor var_744_cast_fp16 = mul(x = x_81_cast_fp16, y = flow_net_res_blocks_5_in_ln_weight_to_fp16)[name = tensor("op_744_cast_fp16")]; + tensor x_83_cast_fp16 = add(x = var_744_cast_fp16, y = flow_net_res_blocks_5_in_ln_bias_to_fp16)[name = tensor("x_83_cast_fp16")]; + tensor var_746_promoted_to_fp16 = const()[name = tensor("op_746_promoted_to_fp16"), val = tensor(0x1p+0)]; + tensor var_747_cast_fp16 = add(x = var_730_cast_fp16_1, y = var_746_promoted_to_fp16)[name = tensor("op_747_cast_fp16")]; + tensor var_748_cast_fp16 = mul(x = x_83_cast_fp16, y = var_747_cast_fp16)[name = tensor("op_748_cast_fp16")]; + tensor input_125_cast_fp16 = add(x = var_748_cast_fp16, y = var_730_cast_fp16_0)[name = tensor("input_125_cast_fp16")]; + tensor linear_48_cast_fp16 = linear(bias = flow_net_res_blocks_5_mlp_0_bias_to_fp16, weight = flow_net_res_blocks_5_mlp_0_weight_to_fp16, x = input_125_cast_fp16)[name = tensor("linear_48_cast_fp16")]; + tensor input_129_cast_fp16 = silu(x = linear_48_cast_fp16)[name = tensor("input_129_cast_fp16")]; + tensor linear_49_cast_fp16 = linear(bias = flow_net_res_blocks_5_mlp_2_bias_to_fp16, weight = flow_net_res_blocks_5_mlp_2_weight_to_fp16, x = input_129_cast_fp16)[name = tensor("linear_49_cast_fp16")]; + tensor var_759_cast_fp16 = mul(x = var_730_cast_fp16_2, y = linear_49_cast_fp16)[name = tensor("op_759_cast_fp16")]; + tensor x_85_cast_fp16 = add(x = x_79_cast_fp16, y = var_759_cast_fp16)[name = tensor("x_85_cast_fp16")]; + tensor linear_50_cast_fp16 = linear(bias = flow_net_final_layer_adaLN_modulation_1_bias_to_fp16, weight = flow_net_final_layer_adaLN_modulation_1_weight_to_fp16, x = input_83_cast_fp16)[name = tensor("linear_50_cast_fp16")]; + tensor var_768_split_sizes_0 = const()[name = tensor("op_768_split_sizes_0"), val = tensor([512, 512])]; + tensor var_768_axis_0 = const()[name = tensor("op_768_axis_0"), val = tensor(-1)]; + tensor var_768_cast_fp16_0, tensor var_768_cast_fp16_1 = split(axis = var_768_axis_0, split_sizes = var_768_split_sizes_0, x = linear_50_cast_fp16)[name = tensor("op_768_cast_fp16")]; + tensor mean_27_axes_0 = const()[name = tensor("mean_27_axes_0"), val = tensor([-1])]; + tensor mean_27_keep_dims_0 = const()[name = tensor("mean_27_keep_dims_0"), val = tensor(true)]; + tensor mean_27_cast_fp16 = reduce_mean(axes = mean_27_axes_0, keep_dims = mean_27_keep_dims_0, x = x_85_cast_fp16)[name = tensor("mean_27_cast_fp16")]; + tensor sub_21_cast_fp16 = sub(x = x_85_cast_fp16, y = mean_27_cast_fp16)[name = tensor("sub_21_cast_fp16")]; + tensor square_17_cast_fp16 = square(x = sub_21_cast_fp16)[name = tensor("square_17_cast_fp16")]; + tensor reduce_mean_35_axes_0 = const()[name = tensor("reduce_mean_35_axes_0"), val = tensor([-1])]; + tensor reduce_mean_35_keep_dims_0 = const()[name = tensor("reduce_mean_35_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_35_cast_fp16 = reduce_mean(axes = reduce_mean_35_axes_0, keep_dims = reduce_mean_35_keep_dims_0, x = square_17_cast_fp16)[name = tensor("reduce_mean_35_cast_fp16")]; + tensor var_775_to_fp16 = const()[name = tensor("op_775_to_fp16"), val = tensor(0x1.1p-20)]; + tensor var_776_cast_fp16 = add(x = reduce_mean_35_cast_fp16, y = var_775_to_fp16)[name = tensor("op_776_cast_fp16")]; + tensor var_777_cast_fp16 = sqrt(x = var_776_cast_fp16)[name = tensor("op_777_cast_fp16")]; + tensor x_87_cast_fp16 = real_div(x = sub_21_cast_fp16, y = var_777_cast_fp16)[name = tensor("x_87_cast_fp16")]; + tensor var_779_promoted_to_fp16 = const()[name = tensor("op_779_promoted_to_fp16"), val = tensor(0x1p+0)]; + tensor var_780_cast_fp16 = add(x = var_768_cast_fp16_1, y = var_779_promoted_to_fp16)[name = tensor("op_780_cast_fp16")]; + tensor var_781_cast_fp16 = mul(x = x_87_cast_fp16, y = var_780_cast_fp16)[name = tensor("op_781_cast_fp16")]; + tensor input_133_cast_fp16 = add(x = var_781_cast_fp16, y = var_768_cast_fp16_0)[name = tensor("input_133_cast_fp16")]; + tensor linear_51_cast_fp16 = linear(bias = flow_net_final_layer_linear_bias_to_fp16, weight = flow_net_final_layer_linear_weight_to_fp16, x = input_133_cast_fp16)[name = tensor("linear_51_cast_fp16")]; + tensor var_792_to_fp16 = const()[name = tensor("op_792_to_fp16"), val = tensor(0x1p-3)]; + tensor var_793_cast_fp16 = mul(x = linear_51_cast_fp16, y = var_792_to_fp16)[name = tensor("op_793_cast_fp16")]; + tensor input_135_cast_fp16 = add(x = input_67_cast_fp16, y = var_793_cast_fp16)[name = tensor("input_135_cast_fp16")]; + tensor linear_52_cast_fp16 = linear(bias = flow_net_input_proj_bias_to_fp16, weight = flow_net_input_proj_weight_to_fp16, x = input_135_cast_fp16)[name = tensor("linear_52_cast_fp16")]; + tensor input_139_to_fp16 = const()[name = tensor("input_139_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(18999552)))]; + tensor input_141_cast_fp16 = silu(x = input_139_to_fp16)[name = tensor("input_141_cast_fp16")]; + tensor linear_54_cast_fp16 = linear(bias = flow_net_time_embed_0_mlp_2_bias_to_fp16, weight = flow_net_time_embed_0_mlp_2_weight_to_fp16, x = input_141_cast_fp16)[name = tensor("linear_54_cast_fp16")]; + tensor reduce_mean_36_axes_0 = const()[name = tensor("reduce_mean_36_axes_0"), val = tensor([-1])]; + tensor reduce_mean_36_keep_dims_0 = const()[name = tensor("reduce_mean_36_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_36_cast_fp16 = reduce_mean(axes = reduce_mean_36_axes_0, keep_dims = reduce_mean_36_keep_dims_0, x = linear_54_cast_fp16)[name = tensor("reduce_mean_36_cast_fp16")]; + tensor sub_22_cast_fp16 = sub(x = linear_54_cast_fp16, y = reduce_mean_36_cast_fp16)[name = tensor("sub_22_cast_fp16")]; + tensor square_18_cast_fp16 = square(x = sub_22_cast_fp16)[name = tensor("square_18_cast_fp16")]; + tensor reduce_mean_37_axes_0 = const()[name = tensor("reduce_mean_37_axes_0"), val = tensor([-1])]; + tensor reduce_mean_37_keep_dims_0 = const()[name = tensor("reduce_mean_37_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_37_cast_fp16 = reduce_mean(axes = reduce_mean_37_axes_0, keep_dims = reduce_mean_37_keep_dims_0, x = square_18_cast_fp16)[name = tensor("reduce_mean_37_cast_fp16")]; + tensor real_div_4_to_fp16 = const()[name = tensor("real_div_4_to_fp16"), val = tensor(0x1.008p+0)]; + tensor mul_4_cast_fp16 = mul(x = reduce_mean_37_cast_fp16, y = real_div_4_to_fp16)[name = tensor("mul_4_cast_fp16")]; + tensor var_859_to_fp16 = const()[name = tensor("op_859_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_37_cast_fp16 = add(x = mul_4_cast_fp16, y = var_859_to_fp16)[name = tensor("var_37_cast_fp16")]; + tensor var_862_epsilon_0 = const()[name = tensor("op_862_epsilon_0"), val = tensor(0x1.197998p-40)]; + tensor var_862_cast_fp16 = rsqrt(epsilon = var_862_epsilon_0, x = var_37_cast_fp16)[name = tensor("op_862_cast_fp16")]; + tensor var_863_cast_fp16 = mul(x = const_3_to_fp16, y = var_862_cast_fp16)[name = tensor("op_863_cast_fp16")]; + tensor var_864_cast_fp16 = mul(x = linear_54_cast_fp16, y = var_863_cast_fp16)[name = tensor("op_864_cast_fp16")]; + tensor input_145_to_fp16 = const()[name = tensor("input_145_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(19000640)))]; + tensor input_147_cast_fp16 = silu(x = input_145_to_fp16)[name = tensor("input_147_cast_fp16")]; + tensor linear_56_cast_fp16 = linear(bias = flow_net_time_embed_1_mlp_2_bias_to_fp16, weight = flow_net_time_embed_1_mlp_2_weight_to_fp16, x = input_147_cast_fp16)[name = tensor("linear_56_cast_fp16")]; + tensor reduce_mean_38_axes_0 = const()[name = tensor("reduce_mean_38_axes_0"), val = tensor([-1])]; + tensor reduce_mean_38_keep_dims_0 = const()[name = tensor("reduce_mean_38_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_38_cast_fp16 = reduce_mean(axes = reduce_mean_38_axes_0, keep_dims = reduce_mean_38_keep_dims_0, x = linear_56_cast_fp16)[name = tensor("reduce_mean_38_cast_fp16")]; + tensor sub_24_cast_fp16 = sub(x = linear_56_cast_fp16, y = reduce_mean_38_cast_fp16)[name = tensor("sub_24_cast_fp16")]; + tensor square_19_cast_fp16 = square(x = sub_24_cast_fp16)[name = tensor("square_19_cast_fp16")]; + tensor reduce_mean_39_axes_0 = const()[name = tensor("reduce_mean_39_axes_0"), val = tensor([-1])]; + tensor reduce_mean_39_keep_dims_0 = const()[name = tensor("reduce_mean_39_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_39_cast_fp16 = reduce_mean(axes = reduce_mean_39_axes_0, keep_dims = reduce_mean_39_keep_dims_0, x = square_19_cast_fp16)[name = tensor("reduce_mean_39_cast_fp16")]; + tensor real_div_5_to_fp16 = const()[name = tensor("real_div_5_to_fp16"), val = tensor(0x1.008p+0)]; + tensor mul_5_cast_fp16 = mul(x = reduce_mean_39_cast_fp16, y = real_div_5_to_fp16)[name = tensor("mul_5_cast_fp16")]; + tensor var_896_to_fp16 = const()[name = tensor("op_896_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_39_cast_fp16 = add(x = mul_5_cast_fp16, y = var_896_to_fp16)[name = tensor("var_39_cast_fp16")]; + tensor var_899_epsilon_0 = const()[name = tensor("op_899_epsilon_0"), val = tensor(0x1.197998p-40)]; + tensor var_899_cast_fp16 = rsqrt(epsilon = var_899_epsilon_0, x = var_39_cast_fp16)[name = tensor("op_899_cast_fp16")]; + tensor var_900_cast_fp16 = mul(x = const_5_to_fp16, y = var_899_cast_fp16)[name = tensor("op_900_cast_fp16")]; + tensor var_901_cast_fp16 = mul(x = linear_56_cast_fp16, y = var_900_cast_fp16)[name = tensor("op_901_cast_fp16")]; + tensor var_913_cast_fp16 = add(x = var_864_cast_fp16, y = var_901_cast_fp16)[name = tensor("op_913_cast_fp16")]; + tensor _inversed_t_combined_5_y_0_to_fp16 = const()[name = tensor("_inversed_t_combined_5_y_0_to_fp16"), val = tensor(0x1p-1)]; + tensor _inversed_t_combined_5_cast_fp16 = mul(x = var_913_cast_fp16, y = _inversed_t_combined_5_y_0_to_fp16)[name = tensor("_inversed_t_combined_5_cast_fp16")]; + tensor input_149_cast_fp16 = add(x = _inversed_t_combined_5_cast_fp16, y = linear_5_cast_fp16)[name = tensor("input_149_cast_fp16")]; + tensor input_151_cast_fp16 = silu(x = input_149_cast_fp16)[name = tensor("input_151_cast_fp16")]; + tensor linear_58_cast_fp16 = linear(bias = flow_net_res_blocks_0_adaLN_modulation_1_bias_to_fp16, weight = flow_net_res_blocks_0_adaLN_modulation_1_weight_to_fp16, x = input_151_cast_fp16)[name = tensor("linear_58_cast_fp16")]; + tensor var_928_split_sizes_0 = const()[name = tensor("op_928_split_sizes_0"), val = tensor([512, 512, 512])]; + tensor var_928_axis_0 = const()[name = tensor("op_928_axis_0"), val = tensor(-1)]; + tensor var_928_cast_fp16_0, tensor var_928_cast_fp16_1, tensor var_928_cast_fp16_2 = split(axis = var_928_axis_0, split_sizes = var_928_split_sizes_0, x = linear_58_cast_fp16)[name = tensor("op_928_cast_fp16")]; + tensor mean_29_axes_0 = const()[name = tensor("mean_29_axes_0"), val = tensor([-1])]; + tensor mean_29_keep_dims_0 = const()[name = tensor("mean_29_keep_dims_0"), val = tensor(true)]; + tensor mean_29_cast_fp16 = reduce_mean(axes = mean_29_axes_0, keep_dims = mean_29_keep_dims_0, x = linear_52_cast_fp16)[name = tensor("mean_29_cast_fp16")]; + tensor sub_26_cast_fp16 = sub(x = linear_52_cast_fp16, y = mean_29_cast_fp16)[name = tensor("sub_26_cast_fp16")]; + tensor square_20_cast_fp16 = square(x = sub_26_cast_fp16)[name = tensor("square_20_cast_fp16")]; + tensor reduce_mean_41_axes_0 = const()[name = tensor("reduce_mean_41_axes_0"), val = tensor([-1])]; + tensor reduce_mean_41_keep_dims_0 = const()[name = tensor("reduce_mean_41_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_41_cast_fp16 = reduce_mean(axes = reduce_mean_41_axes_0, keep_dims = reduce_mean_41_keep_dims_0, x = square_20_cast_fp16)[name = tensor("reduce_mean_41_cast_fp16")]; + tensor var_938_to_fp16 = const()[name = tensor("op_938_to_fp16"), val = tensor(0x1.1p-20)]; + tensor var_939_cast_fp16 = add(x = reduce_mean_41_cast_fp16, y = var_938_to_fp16)[name = tensor("op_939_cast_fp16")]; + tensor var_940_cast_fp16 = sqrt(x = var_939_cast_fp16)[name = tensor("op_940_cast_fp16")]; + tensor x_95_cast_fp16 = real_div(x = sub_26_cast_fp16, y = var_940_cast_fp16)[name = tensor("x_95_cast_fp16")]; + tensor var_942_cast_fp16 = mul(x = x_95_cast_fp16, y = flow_net_res_blocks_0_in_ln_weight_to_fp16)[name = tensor("op_942_cast_fp16")]; + tensor x_97_cast_fp16 = add(x = var_942_cast_fp16, y = flow_net_res_blocks_0_in_ln_bias_to_fp16)[name = tensor("x_97_cast_fp16")]; + tensor var_944_promoted_to_fp16 = const()[name = tensor("op_944_promoted_to_fp16"), val = tensor(0x1p+0)]; + tensor var_945_cast_fp16 = add(x = var_928_cast_fp16_1, y = var_944_promoted_to_fp16)[name = tensor("op_945_cast_fp16")]; + tensor var_946_cast_fp16 = mul(x = x_97_cast_fp16, y = var_945_cast_fp16)[name = tensor("op_946_cast_fp16")]; + tensor input_153_cast_fp16 = add(x = var_946_cast_fp16, y = var_928_cast_fp16_0)[name = tensor("input_153_cast_fp16")]; + tensor linear_59_cast_fp16 = linear(bias = flow_net_res_blocks_0_mlp_0_bias_to_fp16, weight = flow_net_res_blocks_0_mlp_0_weight_to_fp16, x = input_153_cast_fp16)[name = tensor("linear_59_cast_fp16")]; + tensor input_157_cast_fp16 = silu(x = linear_59_cast_fp16)[name = tensor("input_157_cast_fp16")]; + tensor linear_60_cast_fp16 = linear(bias = flow_net_res_blocks_0_mlp_2_bias_to_fp16, weight = flow_net_res_blocks_0_mlp_2_weight_to_fp16, x = input_157_cast_fp16)[name = tensor("linear_60_cast_fp16")]; + tensor var_957_cast_fp16 = mul(x = var_928_cast_fp16_2, y = linear_60_cast_fp16)[name = tensor("op_957_cast_fp16")]; + tensor x_99_cast_fp16 = add(x = linear_52_cast_fp16, y = var_957_cast_fp16)[name = tensor("x_99_cast_fp16")]; + tensor linear_61_cast_fp16 = linear(bias = flow_net_res_blocks_1_adaLN_modulation_1_bias_to_fp16, weight = flow_net_res_blocks_1_adaLN_modulation_1_weight_to_fp16, x = input_151_cast_fp16)[name = tensor("linear_61_cast_fp16")]; + tensor var_967_split_sizes_0 = const()[name = tensor("op_967_split_sizes_0"), val = tensor([512, 512, 512])]; + tensor var_967_axis_0 = const()[name = tensor("op_967_axis_0"), val = tensor(-1)]; + tensor var_967_cast_fp16_0, tensor var_967_cast_fp16_1, tensor var_967_cast_fp16_2 = split(axis = var_967_axis_0, split_sizes = var_967_split_sizes_0, x = linear_61_cast_fp16)[name = tensor("op_967_cast_fp16")]; + tensor mean_31_axes_0 = const()[name = tensor("mean_31_axes_0"), val = tensor([-1])]; + tensor mean_31_keep_dims_0 = const()[name = tensor("mean_31_keep_dims_0"), val = tensor(true)]; + tensor mean_31_cast_fp16 = reduce_mean(axes = mean_31_axes_0, keep_dims = mean_31_keep_dims_0, x = x_99_cast_fp16)[name = tensor("mean_31_cast_fp16")]; + tensor sub_27_cast_fp16 = sub(x = x_99_cast_fp16, y = mean_31_cast_fp16)[name = tensor("sub_27_cast_fp16")]; + tensor square_21_cast_fp16 = square(x = sub_27_cast_fp16)[name = tensor("square_21_cast_fp16")]; + tensor reduce_mean_43_axes_0 = const()[name = tensor("reduce_mean_43_axes_0"), val = tensor([-1])]; + tensor reduce_mean_43_keep_dims_0 = const()[name = tensor("reduce_mean_43_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_43_cast_fp16 = reduce_mean(axes = reduce_mean_43_axes_0, keep_dims = reduce_mean_43_keep_dims_0, x = square_21_cast_fp16)[name = tensor("reduce_mean_43_cast_fp16")]; + tensor var_977_to_fp16 = const()[name = tensor("op_977_to_fp16"), val = tensor(0x1.1p-20)]; + tensor var_978_cast_fp16 = add(x = reduce_mean_43_cast_fp16, y = var_977_to_fp16)[name = tensor("op_978_cast_fp16")]; + tensor var_979_cast_fp16 = sqrt(x = var_978_cast_fp16)[name = tensor("op_979_cast_fp16")]; + tensor x_101_cast_fp16 = real_div(x = sub_27_cast_fp16, y = var_979_cast_fp16)[name = tensor("x_101_cast_fp16")]; + tensor var_981_cast_fp16 = mul(x = x_101_cast_fp16, y = flow_net_res_blocks_1_in_ln_weight_to_fp16)[name = tensor("op_981_cast_fp16")]; + tensor x_103_cast_fp16 = add(x = var_981_cast_fp16, y = flow_net_res_blocks_1_in_ln_bias_to_fp16)[name = tensor("x_103_cast_fp16")]; + tensor var_983_promoted_to_fp16 = const()[name = tensor("op_983_promoted_to_fp16"), val = tensor(0x1p+0)]; + tensor var_984_cast_fp16 = add(x = var_967_cast_fp16_1, y = var_983_promoted_to_fp16)[name = tensor("op_984_cast_fp16")]; + tensor var_985_cast_fp16 = mul(x = x_103_cast_fp16, y = var_984_cast_fp16)[name = tensor("op_985_cast_fp16")]; + tensor input_161_cast_fp16 = add(x = var_985_cast_fp16, y = var_967_cast_fp16_0)[name = tensor("input_161_cast_fp16")]; + tensor linear_62_cast_fp16 = linear(bias = flow_net_res_blocks_1_mlp_0_bias_to_fp16, weight = flow_net_res_blocks_1_mlp_0_weight_to_fp16, x = input_161_cast_fp16)[name = tensor("linear_62_cast_fp16")]; + tensor input_165_cast_fp16 = silu(x = linear_62_cast_fp16)[name = tensor("input_165_cast_fp16")]; + tensor linear_63_cast_fp16 = linear(bias = flow_net_res_blocks_1_mlp_2_bias_to_fp16, weight = flow_net_res_blocks_1_mlp_2_weight_to_fp16, x = input_165_cast_fp16)[name = tensor("linear_63_cast_fp16")]; + tensor var_996_cast_fp16 = mul(x = var_967_cast_fp16_2, y = linear_63_cast_fp16)[name = tensor("op_996_cast_fp16")]; + tensor x_105_cast_fp16 = add(x = x_99_cast_fp16, y = var_996_cast_fp16)[name = tensor("x_105_cast_fp16")]; + tensor linear_64_cast_fp16 = linear(bias = flow_net_res_blocks_2_adaLN_modulation_1_bias_to_fp16, weight = flow_net_res_blocks_2_adaLN_modulation_1_weight_to_fp16, x = input_151_cast_fp16)[name = tensor("linear_64_cast_fp16")]; + tensor var_1006_split_sizes_0 = const()[name = tensor("op_1006_split_sizes_0"), val = tensor([512, 512, 512])]; + tensor var_1006_axis_0 = const()[name = tensor("op_1006_axis_0"), val = tensor(-1)]; + tensor var_1006_cast_fp16_0, tensor var_1006_cast_fp16_1, tensor var_1006_cast_fp16_2 = split(axis = var_1006_axis_0, split_sizes = var_1006_split_sizes_0, x = linear_64_cast_fp16)[name = tensor("op_1006_cast_fp16")]; + tensor mean_33_axes_0 = const()[name = tensor("mean_33_axes_0"), val = tensor([-1])]; + tensor mean_33_keep_dims_0 = const()[name = tensor("mean_33_keep_dims_0"), val = tensor(true)]; + tensor mean_33_cast_fp16 = reduce_mean(axes = mean_33_axes_0, keep_dims = mean_33_keep_dims_0, x = x_105_cast_fp16)[name = tensor("mean_33_cast_fp16")]; + tensor sub_28_cast_fp16 = sub(x = x_105_cast_fp16, y = mean_33_cast_fp16)[name = tensor("sub_28_cast_fp16")]; + tensor square_22_cast_fp16 = square(x = sub_28_cast_fp16)[name = tensor("square_22_cast_fp16")]; + tensor reduce_mean_45_axes_0 = const()[name = tensor("reduce_mean_45_axes_0"), val = tensor([-1])]; + tensor reduce_mean_45_keep_dims_0 = const()[name = tensor("reduce_mean_45_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_45_cast_fp16 = reduce_mean(axes = reduce_mean_45_axes_0, keep_dims = reduce_mean_45_keep_dims_0, x = square_22_cast_fp16)[name = tensor("reduce_mean_45_cast_fp16")]; + tensor var_1016_to_fp16 = const()[name = tensor("op_1016_to_fp16"), val = tensor(0x1.1p-20)]; + tensor var_1017_cast_fp16 = add(x = reduce_mean_45_cast_fp16, y = var_1016_to_fp16)[name = tensor("op_1017_cast_fp16")]; + tensor var_1018_cast_fp16 = sqrt(x = var_1017_cast_fp16)[name = tensor("op_1018_cast_fp16")]; + tensor x_107_cast_fp16 = real_div(x = sub_28_cast_fp16, y = var_1018_cast_fp16)[name = tensor("x_107_cast_fp16")]; + tensor var_1020_cast_fp16 = mul(x = x_107_cast_fp16, y = flow_net_res_blocks_2_in_ln_weight_to_fp16)[name = tensor("op_1020_cast_fp16")]; + tensor x_109_cast_fp16 = add(x = var_1020_cast_fp16, y = flow_net_res_blocks_2_in_ln_bias_to_fp16)[name = tensor("x_109_cast_fp16")]; + tensor var_1022_promoted_to_fp16 = const()[name = tensor("op_1022_promoted_to_fp16"), val = tensor(0x1p+0)]; + tensor var_1023_cast_fp16 = add(x = var_1006_cast_fp16_1, y = var_1022_promoted_to_fp16)[name = tensor("op_1023_cast_fp16")]; + tensor var_1024_cast_fp16 = mul(x = x_109_cast_fp16, y = var_1023_cast_fp16)[name = tensor("op_1024_cast_fp16")]; + tensor input_169_cast_fp16 = add(x = var_1024_cast_fp16, y = var_1006_cast_fp16_0)[name = tensor("input_169_cast_fp16")]; + tensor linear_65_cast_fp16 = linear(bias = flow_net_res_blocks_2_mlp_0_bias_to_fp16, weight = flow_net_res_blocks_2_mlp_0_weight_to_fp16, x = input_169_cast_fp16)[name = tensor("linear_65_cast_fp16")]; + tensor input_173_cast_fp16 = silu(x = linear_65_cast_fp16)[name = tensor("input_173_cast_fp16")]; + tensor linear_66_cast_fp16 = linear(bias = flow_net_res_blocks_2_mlp_2_bias_to_fp16, weight = flow_net_res_blocks_2_mlp_2_weight_to_fp16, x = input_173_cast_fp16)[name = tensor("linear_66_cast_fp16")]; + tensor var_1035_cast_fp16 = mul(x = var_1006_cast_fp16_2, y = linear_66_cast_fp16)[name = tensor("op_1035_cast_fp16")]; + tensor x_111_cast_fp16 = add(x = x_105_cast_fp16, y = var_1035_cast_fp16)[name = tensor("x_111_cast_fp16")]; + tensor linear_67_cast_fp16 = linear(bias = flow_net_res_blocks_3_adaLN_modulation_1_bias_to_fp16, weight = flow_net_res_blocks_3_adaLN_modulation_1_weight_to_fp16, x = input_151_cast_fp16)[name = tensor("linear_67_cast_fp16")]; + tensor var_1045_split_sizes_0 = const()[name = tensor("op_1045_split_sizes_0"), val = tensor([512, 512, 512])]; + tensor var_1045_axis_0 = const()[name = tensor("op_1045_axis_0"), val = tensor(-1)]; + tensor var_1045_cast_fp16_0, tensor var_1045_cast_fp16_1, tensor var_1045_cast_fp16_2 = split(axis = var_1045_axis_0, split_sizes = var_1045_split_sizes_0, x = linear_67_cast_fp16)[name = tensor("op_1045_cast_fp16")]; + tensor mean_35_axes_0 = const()[name = tensor("mean_35_axes_0"), val = tensor([-1])]; + tensor mean_35_keep_dims_0 = const()[name = tensor("mean_35_keep_dims_0"), val = tensor(true)]; + tensor mean_35_cast_fp16 = reduce_mean(axes = mean_35_axes_0, keep_dims = mean_35_keep_dims_0, x = x_111_cast_fp16)[name = tensor("mean_35_cast_fp16")]; + tensor sub_29_cast_fp16 = sub(x = x_111_cast_fp16, y = mean_35_cast_fp16)[name = tensor("sub_29_cast_fp16")]; + tensor square_23_cast_fp16 = square(x = sub_29_cast_fp16)[name = tensor("square_23_cast_fp16")]; + tensor reduce_mean_47_axes_0 = const()[name = tensor("reduce_mean_47_axes_0"), val = tensor([-1])]; + tensor reduce_mean_47_keep_dims_0 = const()[name = tensor("reduce_mean_47_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_47_cast_fp16 = reduce_mean(axes = reduce_mean_47_axes_0, keep_dims = reduce_mean_47_keep_dims_0, x = square_23_cast_fp16)[name = tensor("reduce_mean_47_cast_fp16")]; + tensor var_1055_to_fp16 = const()[name = tensor("op_1055_to_fp16"), val = tensor(0x1.1p-20)]; + tensor var_1056_cast_fp16 = add(x = reduce_mean_47_cast_fp16, y = var_1055_to_fp16)[name = tensor("op_1056_cast_fp16")]; + tensor var_1057_cast_fp16 = sqrt(x = var_1056_cast_fp16)[name = tensor("op_1057_cast_fp16")]; + tensor x_113_cast_fp16 = real_div(x = sub_29_cast_fp16, y = var_1057_cast_fp16)[name = tensor("x_113_cast_fp16")]; + tensor var_1059_cast_fp16 = mul(x = x_113_cast_fp16, y = flow_net_res_blocks_3_in_ln_weight_to_fp16)[name = tensor("op_1059_cast_fp16")]; + tensor x_115_cast_fp16 = add(x = var_1059_cast_fp16, y = flow_net_res_blocks_3_in_ln_bias_to_fp16)[name = tensor("x_115_cast_fp16")]; + tensor var_1061_promoted_to_fp16 = const()[name = tensor("op_1061_promoted_to_fp16"), val = tensor(0x1p+0)]; + tensor var_1062_cast_fp16 = add(x = var_1045_cast_fp16_1, y = var_1061_promoted_to_fp16)[name = tensor("op_1062_cast_fp16")]; + tensor var_1063_cast_fp16 = mul(x = x_115_cast_fp16, y = var_1062_cast_fp16)[name = tensor("op_1063_cast_fp16")]; + tensor input_177_cast_fp16 = add(x = var_1063_cast_fp16, y = var_1045_cast_fp16_0)[name = tensor("input_177_cast_fp16")]; + tensor linear_68_cast_fp16 = linear(bias = flow_net_res_blocks_3_mlp_0_bias_to_fp16, weight = flow_net_res_blocks_3_mlp_0_weight_to_fp16, x = input_177_cast_fp16)[name = tensor("linear_68_cast_fp16")]; + tensor input_181_cast_fp16 = silu(x = linear_68_cast_fp16)[name = tensor("input_181_cast_fp16")]; + tensor linear_69_cast_fp16 = linear(bias = flow_net_res_blocks_3_mlp_2_bias_to_fp16, weight = flow_net_res_blocks_3_mlp_2_weight_to_fp16, x = input_181_cast_fp16)[name = tensor("linear_69_cast_fp16")]; + tensor var_1074_cast_fp16 = mul(x = var_1045_cast_fp16_2, y = linear_69_cast_fp16)[name = tensor("op_1074_cast_fp16")]; + tensor x_117_cast_fp16 = add(x = x_111_cast_fp16, y = var_1074_cast_fp16)[name = tensor("x_117_cast_fp16")]; + tensor linear_70_cast_fp16 = linear(bias = flow_net_res_blocks_4_adaLN_modulation_1_bias_to_fp16, weight = flow_net_res_blocks_4_adaLN_modulation_1_weight_to_fp16, x = input_151_cast_fp16)[name = tensor("linear_70_cast_fp16")]; + tensor var_1084_split_sizes_0 = const()[name = tensor("op_1084_split_sizes_0"), val = tensor([512, 512, 512])]; + tensor var_1084_axis_0 = const()[name = tensor("op_1084_axis_0"), val = tensor(-1)]; + tensor var_1084_cast_fp16_0, tensor var_1084_cast_fp16_1, tensor var_1084_cast_fp16_2 = split(axis = var_1084_axis_0, split_sizes = var_1084_split_sizes_0, x = linear_70_cast_fp16)[name = tensor("op_1084_cast_fp16")]; + tensor mean_37_axes_0 = const()[name = tensor("mean_37_axes_0"), val = tensor([-1])]; + tensor mean_37_keep_dims_0 = const()[name = tensor("mean_37_keep_dims_0"), val = tensor(true)]; + tensor mean_37_cast_fp16 = reduce_mean(axes = mean_37_axes_0, keep_dims = mean_37_keep_dims_0, x = x_117_cast_fp16)[name = tensor("mean_37_cast_fp16")]; + tensor sub_30_cast_fp16 = sub(x = x_117_cast_fp16, y = mean_37_cast_fp16)[name = tensor("sub_30_cast_fp16")]; + tensor square_24_cast_fp16 = square(x = sub_30_cast_fp16)[name = tensor("square_24_cast_fp16")]; + tensor reduce_mean_49_axes_0 = const()[name = tensor("reduce_mean_49_axes_0"), val = tensor([-1])]; + tensor reduce_mean_49_keep_dims_0 = const()[name = tensor("reduce_mean_49_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_49_cast_fp16 = reduce_mean(axes = reduce_mean_49_axes_0, keep_dims = reduce_mean_49_keep_dims_0, x = square_24_cast_fp16)[name = tensor("reduce_mean_49_cast_fp16")]; + tensor var_1094_to_fp16 = const()[name = tensor("op_1094_to_fp16"), val = tensor(0x1.1p-20)]; + tensor var_1095_cast_fp16 = add(x = reduce_mean_49_cast_fp16, y = var_1094_to_fp16)[name = tensor("op_1095_cast_fp16")]; + tensor var_1096_cast_fp16 = sqrt(x = var_1095_cast_fp16)[name = tensor("op_1096_cast_fp16")]; + tensor x_119_cast_fp16 = real_div(x = sub_30_cast_fp16, y = var_1096_cast_fp16)[name = tensor("x_119_cast_fp16")]; + tensor var_1098_cast_fp16 = mul(x = x_119_cast_fp16, y = flow_net_res_blocks_4_in_ln_weight_to_fp16)[name = tensor("op_1098_cast_fp16")]; + tensor x_121_cast_fp16 = add(x = var_1098_cast_fp16, y = flow_net_res_blocks_4_in_ln_bias_to_fp16)[name = tensor("x_121_cast_fp16")]; + tensor var_1100_promoted_to_fp16 = const()[name = tensor("op_1100_promoted_to_fp16"), val = tensor(0x1p+0)]; + tensor var_1101_cast_fp16 = add(x = var_1084_cast_fp16_1, y = var_1100_promoted_to_fp16)[name = tensor("op_1101_cast_fp16")]; + tensor var_1102_cast_fp16 = mul(x = x_121_cast_fp16, y = var_1101_cast_fp16)[name = tensor("op_1102_cast_fp16")]; + tensor input_185_cast_fp16 = add(x = var_1102_cast_fp16, y = var_1084_cast_fp16_0)[name = tensor("input_185_cast_fp16")]; + tensor linear_71_cast_fp16 = linear(bias = flow_net_res_blocks_4_mlp_0_bias_to_fp16, weight = flow_net_res_blocks_4_mlp_0_weight_to_fp16, x = input_185_cast_fp16)[name = tensor("linear_71_cast_fp16")]; + tensor input_189_cast_fp16 = silu(x = linear_71_cast_fp16)[name = tensor("input_189_cast_fp16")]; + tensor linear_72_cast_fp16 = linear(bias = flow_net_res_blocks_4_mlp_2_bias_to_fp16, weight = flow_net_res_blocks_4_mlp_2_weight_to_fp16, x = input_189_cast_fp16)[name = tensor("linear_72_cast_fp16")]; + tensor var_1113_cast_fp16 = mul(x = var_1084_cast_fp16_2, y = linear_72_cast_fp16)[name = tensor("op_1113_cast_fp16")]; + tensor x_123_cast_fp16 = add(x = x_117_cast_fp16, y = var_1113_cast_fp16)[name = tensor("x_123_cast_fp16")]; + tensor linear_73_cast_fp16 = linear(bias = flow_net_res_blocks_5_adaLN_modulation_1_bias_to_fp16, weight = flow_net_res_blocks_5_adaLN_modulation_1_weight_to_fp16, x = input_151_cast_fp16)[name = tensor("linear_73_cast_fp16")]; + tensor var_1123_split_sizes_0 = const()[name = tensor("op_1123_split_sizes_0"), val = tensor([512, 512, 512])]; + tensor var_1123_axis_0 = const()[name = tensor("op_1123_axis_0"), val = tensor(-1)]; + tensor var_1123_cast_fp16_0, tensor var_1123_cast_fp16_1, tensor var_1123_cast_fp16_2 = split(axis = var_1123_axis_0, split_sizes = var_1123_split_sizes_0, x = linear_73_cast_fp16)[name = tensor("op_1123_cast_fp16")]; + tensor mean_39_axes_0 = const()[name = tensor("mean_39_axes_0"), val = tensor([-1])]; + tensor mean_39_keep_dims_0 = const()[name = tensor("mean_39_keep_dims_0"), val = tensor(true)]; + tensor mean_39_cast_fp16 = reduce_mean(axes = mean_39_axes_0, keep_dims = mean_39_keep_dims_0, x = x_123_cast_fp16)[name = tensor("mean_39_cast_fp16")]; + tensor sub_31_cast_fp16 = sub(x = x_123_cast_fp16, y = mean_39_cast_fp16)[name = tensor("sub_31_cast_fp16")]; + tensor square_25_cast_fp16 = square(x = sub_31_cast_fp16)[name = tensor("square_25_cast_fp16")]; + tensor reduce_mean_51_axes_0 = const()[name = tensor("reduce_mean_51_axes_0"), val = tensor([-1])]; + tensor reduce_mean_51_keep_dims_0 = const()[name = tensor("reduce_mean_51_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_51_cast_fp16 = reduce_mean(axes = reduce_mean_51_axes_0, keep_dims = reduce_mean_51_keep_dims_0, x = square_25_cast_fp16)[name = tensor("reduce_mean_51_cast_fp16")]; + tensor var_1133_to_fp16 = const()[name = tensor("op_1133_to_fp16"), val = tensor(0x1.1p-20)]; + tensor var_1134_cast_fp16 = add(x = reduce_mean_51_cast_fp16, y = var_1133_to_fp16)[name = tensor("op_1134_cast_fp16")]; + tensor var_1135_cast_fp16 = sqrt(x = var_1134_cast_fp16)[name = tensor("op_1135_cast_fp16")]; + tensor x_125_cast_fp16 = real_div(x = sub_31_cast_fp16, y = var_1135_cast_fp16)[name = tensor("x_125_cast_fp16")]; + tensor var_1137_cast_fp16 = mul(x = x_125_cast_fp16, y = flow_net_res_blocks_5_in_ln_weight_to_fp16)[name = tensor("op_1137_cast_fp16")]; + tensor x_127_cast_fp16 = add(x = var_1137_cast_fp16, y = flow_net_res_blocks_5_in_ln_bias_to_fp16)[name = tensor("x_127_cast_fp16")]; + tensor var_1139_promoted_to_fp16 = const()[name = tensor("op_1139_promoted_to_fp16"), val = tensor(0x1p+0)]; + tensor var_1140_cast_fp16 = add(x = var_1123_cast_fp16_1, y = var_1139_promoted_to_fp16)[name = tensor("op_1140_cast_fp16")]; + tensor var_1141_cast_fp16 = mul(x = x_127_cast_fp16, y = var_1140_cast_fp16)[name = tensor("op_1141_cast_fp16")]; + tensor input_193_cast_fp16 = add(x = var_1141_cast_fp16, y = var_1123_cast_fp16_0)[name = tensor("input_193_cast_fp16")]; + tensor linear_74_cast_fp16 = linear(bias = flow_net_res_blocks_5_mlp_0_bias_to_fp16, weight = flow_net_res_blocks_5_mlp_0_weight_to_fp16, x = input_193_cast_fp16)[name = tensor("linear_74_cast_fp16")]; + tensor input_197_cast_fp16 = silu(x = linear_74_cast_fp16)[name = tensor("input_197_cast_fp16")]; + tensor linear_75_cast_fp16 = linear(bias = flow_net_res_blocks_5_mlp_2_bias_to_fp16, weight = flow_net_res_blocks_5_mlp_2_weight_to_fp16, x = input_197_cast_fp16)[name = tensor("linear_75_cast_fp16")]; + tensor var_1152_cast_fp16 = mul(x = var_1123_cast_fp16_2, y = linear_75_cast_fp16)[name = tensor("op_1152_cast_fp16")]; + tensor x_129_cast_fp16 = add(x = x_123_cast_fp16, y = var_1152_cast_fp16)[name = tensor("x_129_cast_fp16")]; + tensor linear_76_cast_fp16 = linear(bias = flow_net_final_layer_adaLN_modulation_1_bias_to_fp16, weight = flow_net_final_layer_adaLN_modulation_1_weight_to_fp16, x = input_151_cast_fp16)[name = tensor("linear_76_cast_fp16")]; + tensor var_1161_split_sizes_0 = const()[name = tensor("op_1161_split_sizes_0"), val = tensor([512, 512])]; + tensor var_1161_axis_0 = const()[name = tensor("op_1161_axis_0"), val = tensor(-1)]; + tensor var_1161_cast_fp16_0, tensor var_1161_cast_fp16_1 = split(axis = var_1161_axis_0, split_sizes = var_1161_split_sizes_0, x = linear_76_cast_fp16)[name = tensor("op_1161_cast_fp16")]; + tensor mean_41_axes_0 = const()[name = tensor("mean_41_axes_0"), val = tensor([-1])]; + tensor mean_41_keep_dims_0 = const()[name = tensor("mean_41_keep_dims_0"), val = tensor(true)]; + tensor mean_41_cast_fp16 = reduce_mean(axes = mean_41_axes_0, keep_dims = mean_41_keep_dims_0, x = x_129_cast_fp16)[name = tensor("mean_41_cast_fp16")]; + tensor sub_32_cast_fp16 = sub(x = x_129_cast_fp16, y = mean_41_cast_fp16)[name = tensor("sub_32_cast_fp16")]; + tensor square_26_cast_fp16 = square(x = sub_32_cast_fp16)[name = tensor("square_26_cast_fp16")]; + tensor reduce_mean_53_axes_0 = const()[name = tensor("reduce_mean_53_axes_0"), val = tensor([-1])]; + tensor reduce_mean_53_keep_dims_0 = const()[name = tensor("reduce_mean_53_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_53_cast_fp16 = reduce_mean(axes = reduce_mean_53_axes_0, keep_dims = reduce_mean_53_keep_dims_0, x = square_26_cast_fp16)[name = tensor("reduce_mean_53_cast_fp16")]; + tensor var_1168_to_fp16 = const()[name = tensor("op_1168_to_fp16"), val = tensor(0x1.1p-20)]; + tensor var_1169_cast_fp16 = add(x = reduce_mean_53_cast_fp16, y = var_1168_to_fp16)[name = tensor("op_1169_cast_fp16")]; + tensor var_1170_cast_fp16 = sqrt(x = var_1169_cast_fp16)[name = tensor("op_1170_cast_fp16")]; + tensor x_131_cast_fp16 = real_div(x = sub_32_cast_fp16, y = var_1170_cast_fp16)[name = tensor("x_131_cast_fp16")]; + tensor var_1172_promoted_to_fp16 = const()[name = tensor("op_1172_promoted_to_fp16"), val = tensor(0x1p+0)]; + tensor var_1173_cast_fp16 = add(x = var_1161_cast_fp16_1, y = var_1172_promoted_to_fp16)[name = tensor("op_1173_cast_fp16")]; + tensor var_1174_cast_fp16 = mul(x = x_131_cast_fp16, y = var_1173_cast_fp16)[name = tensor("op_1174_cast_fp16")]; + tensor input_201_cast_fp16 = add(x = var_1174_cast_fp16, y = var_1161_cast_fp16_0)[name = tensor("input_201_cast_fp16")]; + tensor linear_77_cast_fp16 = linear(bias = flow_net_final_layer_linear_bias_to_fp16, weight = flow_net_final_layer_linear_weight_to_fp16, x = input_201_cast_fp16)[name = tensor("linear_77_cast_fp16")]; + tensor var_1185_to_fp16 = const()[name = tensor("op_1185_to_fp16"), val = tensor(0x1p-3)]; + tensor var_1186_cast_fp16 = mul(x = linear_77_cast_fp16, y = var_1185_to_fp16)[name = tensor("op_1186_cast_fp16")]; + tensor input_203_cast_fp16 = add(x = input_135_cast_fp16, y = var_1186_cast_fp16)[name = tensor("input_203_cast_fp16")]; + tensor linear_78_cast_fp16 = linear(bias = flow_net_input_proj_bias_to_fp16, weight = flow_net_input_proj_weight_to_fp16, x = input_203_cast_fp16)[name = tensor("linear_78_cast_fp16")]; + tensor input_207_to_fp16 = const()[name = tensor("input_207_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(19001728)))]; + tensor input_209_cast_fp16 = silu(x = input_207_to_fp16)[name = tensor("input_209_cast_fp16")]; + tensor linear_80_cast_fp16 = linear(bias = flow_net_time_embed_0_mlp_2_bias_to_fp16, weight = flow_net_time_embed_0_mlp_2_weight_to_fp16, x = input_209_cast_fp16)[name = tensor("linear_80_cast_fp16")]; + tensor reduce_mean_54_axes_0 = const()[name = tensor("reduce_mean_54_axes_0"), val = tensor([-1])]; + tensor reduce_mean_54_keep_dims_0 = const()[name = tensor("reduce_mean_54_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_54_cast_fp16 = reduce_mean(axes = reduce_mean_54_axes_0, keep_dims = reduce_mean_54_keep_dims_0, x = linear_80_cast_fp16)[name = tensor("reduce_mean_54_cast_fp16")]; + tensor sub_33_cast_fp16 = sub(x = linear_80_cast_fp16, y = reduce_mean_54_cast_fp16)[name = tensor("sub_33_cast_fp16")]; + tensor square_27_cast_fp16 = square(x = sub_33_cast_fp16)[name = tensor("square_27_cast_fp16")]; + tensor reduce_mean_55_axes_0 = const()[name = tensor("reduce_mean_55_axes_0"), val = tensor([-1])]; + tensor reduce_mean_55_keep_dims_0 = const()[name = tensor("reduce_mean_55_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_55_cast_fp16 = reduce_mean(axes = reduce_mean_55_axes_0, keep_dims = reduce_mean_55_keep_dims_0, x = square_27_cast_fp16)[name = tensor("reduce_mean_55_cast_fp16")]; + tensor real_div_6_to_fp16 = const()[name = tensor("real_div_6_to_fp16"), val = tensor(0x1.008p+0)]; + tensor mul_6_cast_fp16 = mul(x = reduce_mean_55_cast_fp16, y = real_div_6_to_fp16)[name = tensor("mul_6_cast_fp16")]; + tensor var_1252_to_fp16 = const()[name = tensor("op_1252_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_55_cast_fp16 = add(x = mul_6_cast_fp16, y = var_1252_to_fp16)[name = tensor("var_55_cast_fp16")]; + tensor var_1255_epsilon_0 = const()[name = tensor("op_1255_epsilon_0"), val = tensor(0x1.197998p-40)]; + tensor var_1255_cast_fp16 = rsqrt(epsilon = var_1255_epsilon_0, x = var_55_cast_fp16)[name = tensor("op_1255_cast_fp16")]; + tensor var_1256_cast_fp16 = mul(x = const_3_to_fp16, y = var_1255_cast_fp16)[name = tensor("op_1256_cast_fp16")]; + tensor var_1257_cast_fp16 = mul(x = linear_80_cast_fp16, y = var_1256_cast_fp16)[name = tensor("op_1257_cast_fp16")]; + tensor input_213_to_fp16 = const()[name = tensor("input_213_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(19002816)))]; + tensor input_215_cast_fp16 = silu(x = input_213_to_fp16)[name = tensor("input_215_cast_fp16")]; + tensor linear_82_cast_fp16 = linear(bias = flow_net_time_embed_1_mlp_2_bias_to_fp16, weight = flow_net_time_embed_1_mlp_2_weight_to_fp16, x = input_215_cast_fp16)[name = tensor("linear_82_cast_fp16")]; + tensor reduce_mean_56_axes_0 = const()[name = tensor("reduce_mean_56_axes_0"), val = tensor([-1])]; + tensor reduce_mean_56_keep_dims_0 = const()[name = tensor("reduce_mean_56_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_56_cast_fp16 = reduce_mean(axes = reduce_mean_56_axes_0, keep_dims = reduce_mean_56_keep_dims_0, x = linear_82_cast_fp16)[name = tensor("reduce_mean_56_cast_fp16")]; + tensor sub_35_cast_fp16 = sub(x = linear_82_cast_fp16, y = reduce_mean_56_cast_fp16)[name = tensor("sub_35_cast_fp16")]; + tensor square_28_cast_fp16 = square(x = sub_35_cast_fp16)[name = tensor("square_28_cast_fp16")]; + tensor reduce_mean_57_axes_0 = const()[name = tensor("reduce_mean_57_axes_0"), val = tensor([-1])]; + tensor reduce_mean_57_keep_dims_0 = const()[name = tensor("reduce_mean_57_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_57_cast_fp16 = reduce_mean(axes = reduce_mean_57_axes_0, keep_dims = reduce_mean_57_keep_dims_0, x = square_28_cast_fp16)[name = tensor("reduce_mean_57_cast_fp16")]; + tensor real_div_7_to_fp16 = const()[name = tensor("real_div_7_to_fp16"), val = tensor(0x1.008p+0)]; + tensor mul_7_cast_fp16 = mul(x = reduce_mean_57_cast_fp16, y = real_div_7_to_fp16)[name = tensor("mul_7_cast_fp16")]; + tensor var_1289_to_fp16 = const()[name = tensor("op_1289_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_57_cast_fp16 = add(x = mul_7_cast_fp16, y = var_1289_to_fp16)[name = tensor("var_57_cast_fp16")]; + tensor var_1292_epsilon_0 = const()[name = tensor("op_1292_epsilon_0"), val = tensor(0x1.197998p-40)]; + tensor var_1292_cast_fp16 = rsqrt(epsilon = var_1292_epsilon_0, x = var_57_cast_fp16)[name = tensor("op_1292_cast_fp16")]; + tensor var_1293_cast_fp16 = mul(x = const_5_to_fp16, y = var_1292_cast_fp16)[name = tensor("op_1293_cast_fp16")]; + tensor var_1294_cast_fp16 = mul(x = linear_82_cast_fp16, y = var_1293_cast_fp16)[name = tensor("op_1294_cast_fp16")]; + tensor var_1306_cast_fp16 = add(x = var_1257_cast_fp16, y = var_1294_cast_fp16)[name = tensor("op_1306_cast_fp16")]; + tensor _inversed_t_combined_7_y_0_to_fp16 = const()[name = tensor("_inversed_t_combined_7_y_0_to_fp16"), val = tensor(0x1p-1)]; + tensor _inversed_t_combined_7_cast_fp16 = mul(x = var_1306_cast_fp16, y = _inversed_t_combined_7_y_0_to_fp16)[name = tensor("_inversed_t_combined_7_cast_fp16")]; + tensor input_217_cast_fp16 = add(x = _inversed_t_combined_7_cast_fp16, y = linear_5_cast_fp16)[name = tensor("input_217_cast_fp16")]; + tensor input_219_cast_fp16 = silu(x = input_217_cast_fp16)[name = tensor("input_219_cast_fp16")]; + tensor linear_84_cast_fp16 = linear(bias = flow_net_res_blocks_0_adaLN_modulation_1_bias_to_fp16, weight = flow_net_res_blocks_0_adaLN_modulation_1_weight_to_fp16, x = input_219_cast_fp16)[name = tensor("linear_84_cast_fp16")]; + tensor var_1321_split_sizes_0 = const()[name = tensor("op_1321_split_sizes_0"), val = tensor([512, 512, 512])]; + tensor var_1321_axis_0 = const()[name = tensor("op_1321_axis_0"), val = tensor(-1)]; + tensor var_1321_cast_fp16_0, tensor var_1321_cast_fp16_1, tensor var_1321_cast_fp16_2 = split(axis = var_1321_axis_0, split_sizes = var_1321_split_sizes_0, x = linear_84_cast_fp16)[name = tensor("op_1321_cast_fp16")]; + tensor mean_43_axes_0 = const()[name = tensor("mean_43_axes_0"), val = tensor([-1])]; + tensor mean_43_keep_dims_0 = const()[name = tensor("mean_43_keep_dims_0"), val = tensor(true)]; + tensor mean_43_cast_fp16 = reduce_mean(axes = mean_43_axes_0, keep_dims = mean_43_keep_dims_0, x = linear_78_cast_fp16)[name = tensor("mean_43_cast_fp16")]; + tensor sub_37_cast_fp16 = sub(x = linear_78_cast_fp16, y = mean_43_cast_fp16)[name = tensor("sub_37_cast_fp16")]; + tensor square_29_cast_fp16 = square(x = sub_37_cast_fp16)[name = tensor("square_29_cast_fp16")]; + tensor reduce_mean_59_axes_0 = const()[name = tensor("reduce_mean_59_axes_0"), val = tensor([-1])]; + tensor reduce_mean_59_keep_dims_0 = const()[name = tensor("reduce_mean_59_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_59_cast_fp16 = reduce_mean(axes = reduce_mean_59_axes_0, keep_dims = reduce_mean_59_keep_dims_0, x = square_29_cast_fp16)[name = tensor("reduce_mean_59_cast_fp16")]; + tensor var_1331_to_fp16 = const()[name = tensor("op_1331_to_fp16"), val = tensor(0x1.1p-20)]; + tensor var_1332_cast_fp16 = add(x = reduce_mean_59_cast_fp16, y = var_1331_to_fp16)[name = tensor("op_1332_cast_fp16")]; + tensor var_1333_cast_fp16 = sqrt(x = var_1332_cast_fp16)[name = tensor("op_1333_cast_fp16")]; + tensor x_139_cast_fp16 = real_div(x = sub_37_cast_fp16, y = var_1333_cast_fp16)[name = tensor("x_139_cast_fp16")]; + tensor var_1335_cast_fp16 = mul(x = x_139_cast_fp16, y = flow_net_res_blocks_0_in_ln_weight_to_fp16)[name = tensor("op_1335_cast_fp16")]; + tensor x_141_cast_fp16 = add(x = var_1335_cast_fp16, y = flow_net_res_blocks_0_in_ln_bias_to_fp16)[name = tensor("x_141_cast_fp16")]; + tensor var_1337_promoted_to_fp16 = const()[name = tensor("op_1337_promoted_to_fp16"), val = tensor(0x1p+0)]; + tensor var_1338_cast_fp16 = add(x = var_1321_cast_fp16_1, y = var_1337_promoted_to_fp16)[name = tensor("op_1338_cast_fp16")]; + tensor var_1339_cast_fp16 = mul(x = x_141_cast_fp16, y = var_1338_cast_fp16)[name = tensor("op_1339_cast_fp16")]; + tensor input_221_cast_fp16 = add(x = var_1339_cast_fp16, y = var_1321_cast_fp16_0)[name = tensor("input_221_cast_fp16")]; + tensor linear_85_cast_fp16 = linear(bias = flow_net_res_blocks_0_mlp_0_bias_to_fp16, weight = flow_net_res_blocks_0_mlp_0_weight_to_fp16, x = input_221_cast_fp16)[name = tensor("linear_85_cast_fp16")]; + tensor input_225_cast_fp16 = silu(x = linear_85_cast_fp16)[name = tensor("input_225_cast_fp16")]; + tensor linear_86_cast_fp16 = linear(bias = flow_net_res_blocks_0_mlp_2_bias_to_fp16, weight = flow_net_res_blocks_0_mlp_2_weight_to_fp16, x = input_225_cast_fp16)[name = tensor("linear_86_cast_fp16")]; + tensor var_1350_cast_fp16 = mul(x = var_1321_cast_fp16_2, y = linear_86_cast_fp16)[name = tensor("op_1350_cast_fp16")]; + tensor x_143_cast_fp16 = add(x = linear_78_cast_fp16, y = var_1350_cast_fp16)[name = tensor("x_143_cast_fp16")]; + tensor linear_87_cast_fp16 = linear(bias = flow_net_res_blocks_1_adaLN_modulation_1_bias_to_fp16, weight = flow_net_res_blocks_1_adaLN_modulation_1_weight_to_fp16, x = input_219_cast_fp16)[name = tensor("linear_87_cast_fp16")]; + tensor var_1360_split_sizes_0 = const()[name = tensor("op_1360_split_sizes_0"), val = tensor([512, 512, 512])]; + tensor var_1360_axis_0 = const()[name = tensor("op_1360_axis_0"), val = tensor(-1)]; + tensor var_1360_cast_fp16_0, tensor var_1360_cast_fp16_1, tensor var_1360_cast_fp16_2 = split(axis = var_1360_axis_0, split_sizes = var_1360_split_sizes_0, x = linear_87_cast_fp16)[name = tensor("op_1360_cast_fp16")]; + tensor mean_45_axes_0 = const()[name = tensor("mean_45_axes_0"), val = tensor([-1])]; + tensor mean_45_keep_dims_0 = const()[name = tensor("mean_45_keep_dims_0"), val = tensor(true)]; + tensor mean_45_cast_fp16 = reduce_mean(axes = mean_45_axes_0, keep_dims = mean_45_keep_dims_0, x = x_143_cast_fp16)[name = tensor("mean_45_cast_fp16")]; + tensor sub_38_cast_fp16 = sub(x = x_143_cast_fp16, y = mean_45_cast_fp16)[name = tensor("sub_38_cast_fp16")]; + tensor square_30_cast_fp16 = square(x = sub_38_cast_fp16)[name = tensor("square_30_cast_fp16")]; + tensor reduce_mean_61_axes_0 = const()[name = tensor("reduce_mean_61_axes_0"), val = tensor([-1])]; + tensor reduce_mean_61_keep_dims_0 = const()[name = tensor("reduce_mean_61_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_61_cast_fp16 = reduce_mean(axes = reduce_mean_61_axes_0, keep_dims = reduce_mean_61_keep_dims_0, x = square_30_cast_fp16)[name = tensor("reduce_mean_61_cast_fp16")]; + tensor var_1370_to_fp16 = const()[name = tensor("op_1370_to_fp16"), val = tensor(0x1.1p-20)]; + tensor var_1371_cast_fp16 = add(x = reduce_mean_61_cast_fp16, y = var_1370_to_fp16)[name = tensor("op_1371_cast_fp16")]; + tensor var_1372_cast_fp16 = sqrt(x = var_1371_cast_fp16)[name = tensor("op_1372_cast_fp16")]; + tensor x_145_cast_fp16 = real_div(x = sub_38_cast_fp16, y = var_1372_cast_fp16)[name = tensor("x_145_cast_fp16")]; + tensor var_1374_cast_fp16 = mul(x = x_145_cast_fp16, y = flow_net_res_blocks_1_in_ln_weight_to_fp16)[name = tensor("op_1374_cast_fp16")]; + tensor x_147_cast_fp16 = add(x = var_1374_cast_fp16, y = flow_net_res_blocks_1_in_ln_bias_to_fp16)[name = tensor("x_147_cast_fp16")]; + tensor var_1376_promoted_to_fp16 = const()[name = tensor("op_1376_promoted_to_fp16"), val = tensor(0x1p+0)]; + tensor var_1377_cast_fp16 = add(x = var_1360_cast_fp16_1, y = var_1376_promoted_to_fp16)[name = tensor("op_1377_cast_fp16")]; + tensor var_1378_cast_fp16 = mul(x = x_147_cast_fp16, y = var_1377_cast_fp16)[name = tensor("op_1378_cast_fp16")]; + tensor input_229_cast_fp16 = add(x = var_1378_cast_fp16, y = var_1360_cast_fp16_0)[name = tensor("input_229_cast_fp16")]; + tensor linear_88_cast_fp16 = linear(bias = flow_net_res_blocks_1_mlp_0_bias_to_fp16, weight = flow_net_res_blocks_1_mlp_0_weight_to_fp16, x = input_229_cast_fp16)[name = tensor("linear_88_cast_fp16")]; + tensor input_233_cast_fp16 = silu(x = linear_88_cast_fp16)[name = tensor("input_233_cast_fp16")]; + tensor linear_89_cast_fp16 = linear(bias = flow_net_res_blocks_1_mlp_2_bias_to_fp16, weight = flow_net_res_blocks_1_mlp_2_weight_to_fp16, x = input_233_cast_fp16)[name = tensor("linear_89_cast_fp16")]; + tensor var_1389_cast_fp16 = mul(x = var_1360_cast_fp16_2, y = linear_89_cast_fp16)[name = tensor("op_1389_cast_fp16")]; + tensor x_149_cast_fp16 = add(x = x_143_cast_fp16, y = var_1389_cast_fp16)[name = tensor("x_149_cast_fp16")]; + tensor linear_90_cast_fp16 = linear(bias = flow_net_res_blocks_2_adaLN_modulation_1_bias_to_fp16, weight = flow_net_res_blocks_2_adaLN_modulation_1_weight_to_fp16, x = input_219_cast_fp16)[name = tensor("linear_90_cast_fp16")]; + tensor var_1399_split_sizes_0 = const()[name = tensor("op_1399_split_sizes_0"), val = tensor([512, 512, 512])]; + tensor var_1399_axis_0 = const()[name = tensor("op_1399_axis_0"), val = tensor(-1)]; + tensor var_1399_cast_fp16_0, tensor var_1399_cast_fp16_1, tensor var_1399_cast_fp16_2 = split(axis = var_1399_axis_0, split_sizes = var_1399_split_sizes_0, x = linear_90_cast_fp16)[name = tensor("op_1399_cast_fp16")]; + tensor mean_47_axes_0 = const()[name = tensor("mean_47_axes_0"), val = tensor([-1])]; + tensor mean_47_keep_dims_0 = const()[name = tensor("mean_47_keep_dims_0"), val = tensor(true)]; + tensor mean_47_cast_fp16 = reduce_mean(axes = mean_47_axes_0, keep_dims = mean_47_keep_dims_0, x = x_149_cast_fp16)[name = tensor("mean_47_cast_fp16")]; + tensor sub_39_cast_fp16 = sub(x = x_149_cast_fp16, y = mean_47_cast_fp16)[name = tensor("sub_39_cast_fp16")]; + tensor square_31_cast_fp16 = square(x = sub_39_cast_fp16)[name = tensor("square_31_cast_fp16")]; + tensor reduce_mean_63_axes_0 = const()[name = tensor("reduce_mean_63_axes_0"), val = tensor([-1])]; + tensor reduce_mean_63_keep_dims_0 = const()[name = tensor("reduce_mean_63_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_63_cast_fp16 = reduce_mean(axes = reduce_mean_63_axes_0, keep_dims = reduce_mean_63_keep_dims_0, x = square_31_cast_fp16)[name = tensor("reduce_mean_63_cast_fp16")]; + tensor var_1409_to_fp16 = const()[name = tensor("op_1409_to_fp16"), val = tensor(0x1.1p-20)]; + tensor var_1410_cast_fp16 = add(x = reduce_mean_63_cast_fp16, y = var_1409_to_fp16)[name = tensor("op_1410_cast_fp16")]; + tensor var_1411_cast_fp16 = sqrt(x = var_1410_cast_fp16)[name = tensor("op_1411_cast_fp16")]; + tensor x_151_cast_fp16 = real_div(x = sub_39_cast_fp16, y = var_1411_cast_fp16)[name = tensor("x_151_cast_fp16")]; + tensor var_1413_cast_fp16 = mul(x = x_151_cast_fp16, y = flow_net_res_blocks_2_in_ln_weight_to_fp16)[name = tensor("op_1413_cast_fp16")]; + tensor x_153_cast_fp16 = add(x = var_1413_cast_fp16, y = flow_net_res_blocks_2_in_ln_bias_to_fp16)[name = tensor("x_153_cast_fp16")]; + tensor var_1415_promoted_to_fp16 = const()[name = tensor("op_1415_promoted_to_fp16"), val = tensor(0x1p+0)]; + tensor var_1416_cast_fp16 = add(x = var_1399_cast_fp16_1, y = var_1415_promoted_to_fp16)[name = tensor("op_1416_cast_fp16")]; + tensor var_1417_cast_fp16 = mul(x = x_153_cast_fp16, y = var_1416_cast_fp16)[name = tensor("op_1417_cast_fp16")]; + tensor input_237_cast_fp16 = add(x = var_1417_cast_fp16, y = var_1399_cast_fp16_0)[name = tensor("input_237_cast_fp16")]; + tensor linear_91_cast_fp16 = linear(bias = flow_net_res_blocks_2_mlp_0_bias_to_fp16, weight = flow_net_res_blocks_2_mlp_0_weight_to_fp16, x = input_237_cast_fp16)[name = tensor("linear_91_cast_fp16")]; + tensor input_241_cast_fp16 = silu(x = linear_91_cast_fp16)[name = tensor("input_241_cast_fp16")]; + tensor linear_92_cast_fp16 = linear(bias = flow_net_res_blocks_2_mlp_2_bias_to_fp16, weight = flow_net_res_blocks_2_mlp_2_weight_to_fp16, x = input_241_cast_fp16)[name = tensor("linear_92_cast_fp16")]; + tensor var_1428_cast_fp16 = mul(x = var_1399_cast_fp16_2, y = linear_92_cast_fp16)[name = tensor("op_1428_cast_fp16")]; + tensor x_155_cast_fp16 = add(x = x_149_cast_fp16, y = var_1428_cast_fp16)[name = tensor("x_155_cast_fp16")]; + tensor linear_93_cast_fp16 = linear(bias = flow_net_res_blocks_3_adaLN_modulation_1_bias_to_fp16, weight = flow_net_res_blocks_3_adaLN_modulation_1_weight_to_fp16, x = input_219_cast_fp16)[name = tensor("linear_93_cast_fp16")]; + tensor var_1438_split_sizes_0 = const()[name = tensor("op_1438_split_sizes_0"), val = tensor([512, 512, 512])]; + tensor var_1438_axis_0 = const()[name = tensor("op_1438_axis_0"), val = tensor(-1)]; + tensor var_1438_cast_fp16_0, tensor var_1438_cast_fp16_1, tensor var_1438_cast_fp16_2 = split(axis = var_1438_axis_0, split_sizes = var_1438_split_sizes_0, x = linear_93_cast_fp16)[name = tensor("op_1438_cast_fp16")]; + tensor mean_49_axes_0 = const()[name = tensor("mean_49_axes_0"), val = tensor([-1])]; + tensor mean_49_keep_dims_0 = const()[name = tensor("mean_49_keep_dims_0"), val = tensor(true)]; + tensor mean_49_cast_fp16 = reduce_mean(axes = mean_49_axes_0, keep_dims = mean_49_keep_dims_0, x = x_155_cast_fp16)[name = tensor("mean_49_cast_fp16")]; + tensor sub_40_cast_fp16 = sub(x = x_155_cast_fp16, y = mean_49_cast_fp16)[name = tensor("sub_40_cast_fp16")]; + tensor square_32_cast_fp16 = square(x = sub_40_cast_fp16)[name = tensor("square_32_cast_fp16")]; + tensor reduce_mean_65_axes_0 = const()[name = tensor("reduce_mean_65_axes_0"), val = tensor([-1])]; + tensor reduce_mean_65_keep_dims_0 = const()[name = tensor("reduce_mean_65_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_65_cast_fp16 = reduce_mean(axes = reduce_mean_65_axes_0, keep_dims = reduce_mean_65_keep_dims_0, x = square_32_cast_fp16)[name = tensor("reduce_mean_65_cast_fp16")]; + tensor var_1448_to_fp16 = const()[name = tensor("op_1448_to_fp16"), val = tensor(0x1.1p-20)]; + tensor var_1449_cast_fp16 = add(x = reduce_mean_65_cast_fp16, y = var_1448_to_fp16)[name = tensor("op_1449_cast_fp16")]; + tensor var_1450_cast_fp16 = sqrt(x = var_1449_cast_fp16)[name = tensor("op_1450_cast_fp16")]; + tensor x_157_cast_fp16 = real_div(x = sub_40_cast_fp16, y = var_1450_cast_fp16)[name = tensor("x_157_cast_fp16")]; + tensor var_1452_cast_fp16 = mul(x = x_157_cast_fp16, y = flow_net_res_blocks_3_in_ln_weight_to_fp16)[name = tensor("op_1452_cast_fp16")]; + tensor x_159_cast_fp16 = add(x = var_1452_cast_fp16, y = flow_net_res_blocks_3_in_ln_bias_to_fp16)[name = tensor("x_159_cast_fp16")]; + tensor var_1454_promoted_to_fp16 = const()[name = tensor("op_1454_promoted_to_fp16"), val = tensor(0x1p+0)]; + tensor var_1455_cast_fp16 = add(x = var_1438_cast_fp16_1, y = var_1454_promoted_to_fp16)[name = tensor("op_1455_cast_fp16")]; + tensor var_1456_cast_fp16 = mul(x = x_159_cast_fp16, y = var_1455_cast_fp16)[name = tensor("op_1456_cast_fp16")]; + tensor input_245_cast_fp16 = add(x = var_1456_cast_fp16, y = var_1438_cast_fp16_0)[name = tensor("input_245_cast_fp16")]; + tensor linear_94_cast_fp16 = linear(bias = flow_net_res_blocks_3_mlp_0_bias_to_fp16, weight = flow_net_res_blocks_3_mlp_0_weight_to_fp16, x = input_245_cast_fp16)[name = tensor("linear_94_cast_fp16")]; + tensor input_249_cast_fp16 = silu(x = linear_94_cast_fp16)[name = tensor("input_249_cast_fp16")]; + tensor linear_95_cast_fp16 = linear(bias = flow_net_res_blocks_3_mlp_2_bias_to_fp16, weight = flow_net_res_blocks_3_mlp_2_weight_to_fp16, x = input_249_cast_fp16)[name = tensor("linear_95_cast_fp16")]; + tensor var_1467_cast_fp16 = mul(x = var_1438_cast_fp16_2, y = linear_95_cast_fp16)[name = tensor("op_1467_cast_fp16")]; + tensor x_161_cast_fp16 = add(x = x_155_cast_fp16, y = var_1467_cast_fp16)[name = tensor("x_161_cast_fp16")]; + tensor linear_96_cast_fp16 = linear(bias = flow_net_res_blocks_4_adaLN_modulation_1_bias_to_fp16, weight = flow_net_res_blocks_4_adaLN_modulation_1_weight_to_fp16, x = input_219_cast_fp16)[name = tensor("linear_96_cast_fp16")]; + tensor var_1477_split_sizes_0 = const()[name = tensor("op_1477_split_sizes_0"), val = tensor([512, 512, 512])]; + tensor var_1477_axis_0 = const()[name = tensor("op_1477_axis_0"), val = tensor(-1)]; + tensor var_1477_cast_fp16_0, tensor var_1477_cast_fp16_1, tensor var_1477_cast_fp16_2 = split(axis = var_1477_axis_0, split_sizes = var_1477_split_sizes_0, x = linear_96_cast_fp16)[name = tensor("op_1477_cast_fp16")]; + tensor mean_51_axes_0 = const()[name = tensor("mean_51_axes_0"), val = tensor([-1])]; + tensor mean_51_keep_dims_0 = const()[name = tensor("mean_51_keep_dims_0"), val = tensor(true)]; + tensor mean_51_cast_fp16 = reduce_mean(axes = mean_51_axes_0, keep_dims = mean_51_keep_dims_0, x = x_161_cast_fp16)[name = tensor("mean_51_cast_fp16")]; + tensor sub_41_cast_fp16 = sub(x = x_161_cast_fp16, y = mean_51_cast_fp16)[name = tensor("sub_41_cast_fp16")]; + tensor square_33_cast_fp16 = square(x = sub_41_cast_fp16)[name = tensor("square_33_cast_fp16")]; + tensor reduce_mean_67_axes_0 = const()[name = tensor("reduce_mean_67_axes_0"), val = tensor([-1])]; + tensor reduce_mean_67_keep_dims_0 = const()[name = tensor("reduce_mean_67_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_67_cast_fp16 = reduce_mean(axes = reduce_mean_67_axes_0, keep_dims = reduce_mean_67_keep_dims_0, x = square_33_cast_fp16)[name = tensor("reduce_mean_67_cast_fp16")]; + tensor var_1487_to_fp16 = const()[name = tensor("op_1487_to_fp16"), val = tensor(0x1.1p-20)]; + tensor var_1488_cast_fp16 = add(x = reduce_mean_67_cast_fp16, y = var_1487_to_fp16)[name = tensor("op_1488_cast_fp16")]; + tensor var_1489_cast_fp16 = sqrt(x = var_1488_cast_fp16)[name = tensor("op_1489_cast_fp16")]; + tensor x_163_cast_fp16 = real_div(x = sub_41_cast_fp16, y = var_1489_cast_fp16)[name = tensor("x_163_cast_fp16")]; + tensor var_1491_cast_fp16 = mul(x = x_163_cast_fp16, y = flow_net_res_blocks_4_in_ln_weight_to_fp16)[name = tensor("op_1491_cast_fp16")]; + tensor x_165_cast_fp16 = add(x = var_1491_cast_fp16, y = flow_net_res_blocks_4_in_ln_bias_to_fp16)[name = tensor("x_165_cast_fp16")]; + tensor var_1493_promoted_to_fp16 = const()[name = tensor("op_1493_promoted_to_fp16"), val = tensor(0x1p+0)]; + tensor var_1494_cast_fp16 = add(x = var_1477_cast_fp16_1, y = var_1493_promoted_to_fp16)[name = tensor("op_1494_cast_fp16")]; + tensor var_1495_cast_fp16 = mul(x = x_165_cast_fp16, y = var_1494_cast_fp16)[name = tensor("op_1495_cast_fp16")]; + tensor input_253_cast_fp16 = add(x = var_1495_cast_fp16, y = var_1477_cast_fp16_0)[name = tensor("input_253_cast_fp16")]; + tensor linear_97_cast_fp16 = linear(bias = flow_net_res_blocks_4_mlp_0_bias_to_fp16, weight = flow_net_res_blocks_4_mlp_0_weight_to_fp16, x = input_253_cast_fp16)[name = tensor("linear_97_cast_fp16")]; + tensor input_257_cast_fp16 = silu(x = linear_97_cast_fp16)[name = tensor("input_257_cast_fp16")]; + tensor linear_98_cast_fp16 = linear(bias = flow_net_res_blocks_4_mlp_2_bias_to_fp16, weight = flow_net_res_blocks_4_mlp_2_weight_to_fp16, x = input_257_cast_fp16)[name = tensor("linear_98_cast_fp16")]; + tensor var_1506_cast_fp16 = mul(x = var_1477_cast_fp16_2, y = linear_98_cast_fp16)[name = tensor("op_1506_cast_fp16")]; + tensor x_167_cast_fp16 = add(x = x_161_cast_fp16, y = var_1506_cast_fp16)[name = tensor("x_167_cast_fp16")]; + tensor linear_99_cast_fp16 = linear(bias = flow_net_res_blocks_5_adaLN_modulation_1_bias_to_fp16, weight = flow_net_res_blocks_5_adaLN_modulation_1_weight_to_fp16, x = input_219_cast_fp16)[name = tensor("linear_99_cast_fp16")]; + tensor var_1516_split_sizes_0 = const()[name = tensor("op_1516_split_sizes_0"), val = tensor([512, 512, 512])]; + tensor var_1516_axis_0 = const()[name = tensor("op_1516_axis_0"), val = tensor(-1)]; + tensor var_1516_cast_fp16_0, tensor var_1516_cast_fp16_1, tensor var_1516_cast_fp16_2 = split(axis = var_1516_axis_0, split_sizes = var_1516_split_sizes_0, x = linear_99_cast_fp16)[name = tensor("op_1516_cast_fp16")]; + tensor mean_53_axes_0 = const()[name = tensor("mean_53_axes_0"), val = tensor([-1])]; + tensor mean_53_keep_dims_0 = const()[name = tensor("mean_53_keep_dims_0"), val = tensor(true)]; + tensor mean_53_cast_fp16 = reduce_mean(axes = mean_53_axes_0, keep_dims = mean_53_keep_dims_0, x = x_167_cast_fp16)[name = tensor("mean_53_cast_fp16")]; + tensor sub_42_cast_fp16 = sub(x = x_167_cast_fp16, y = mean_53_cast_fp16)[name = tensor("sub_42_cast_fp16")]; + tensor square_34_cast_fp16 = square(x = sub_42_cast_fp16)[name = tensor("square_34_cast_fp16")]; + tensor reduce_mean_69_axes_0 = const()[name = tensor("reduce_mean_69_axes_0"), val = tensor([-1])]; + tensor reduce_mean_69_keep_dims_0 = const()[name = tensor("reduce_mean_69_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_69_cast_fp16 = reduce_mean(axes = reduce_mean_69_axes_0, keep_dims = reduce_mean_69_keep_dims_0, x = square_34_cast_fp16)[name = tensor("reduce_mean_69_cast_fp16")]; + tensor var_1526_to_fp16 = const()[name = tensor("op_1526_to_fp16"), val = tensor(0x1.1p-20)]; + tensor var_1527_cast_fp16 = add(x = reduce_mean_69_cast_fp16, y = var_1526_to_fp16)[name = tensor("op_1527_cast_fp16")]; + tensor var_1528_cast_fp16 = sqrt(x = var_1527_cast_fp16)[name = tensor("op_1528_cast_fp16")]; + tensor x_169_cast_fp16 = real_div(x = sub_42_cast_fp16, y = var_1528_cast_fp16)[name = tensor("x_169_cast_fp16")]; + tensor var_1530_cast_fp16 = mul(x = x_169_cast_fp16, y = flow_net_res_blocks_5_in_ln_weight_to_fp16)[name = tensor("op_1530_cast_fp16")]; + tensor x_171_cast_fp16 = add(x = var_1530_cast_fp16, y = flow_net_res_blocks_5_in_ln_bias_to_fp16)[name = tensor("x_171_cast_fp16")]; + tensor var_1532_promoted_to_fp16 = const()[name = tensor("op_1532_promoted_to_fp16"), val = tensor(0x1p+0)]; + tensor var_1533_cast_fp16 = add(x = var_1516_cast_fp16_1, y = var_1532_promoted_to_fp16)[name = tensor("op_1533_cast_fp16")]; + tensor var_1534_cast_fp16 = mul(x = x_171_cast_fp16, y = var_1533_cast_fp16)[name = tensor("op_1534_cast_fp16")]; + tensor input_261_cast_fp16 = add(x = var_1534_cast_fp16, y = var_1516_cast_fp16_0)[name = tensor("input_261_cast_fp16")]; + tensor linear_100_cast_fp16 = linear(bias = flow_net_res_blocks_5_mlp_0_bias_to_fp16, weight = flow_net_res_blocks_5_mlp_0_weight_to_fp16, x = input_261_cast_fp16)[name = tensor("linear_100_cast_fp16")]; + tensor input_265_cast_fp16 = silu(x = linear_100_cast_fp16)[name = tensor("input_265_cast_fp16")]; + tensor linear_101_cast_fp16 = linear(bias = flow_net_res_blocks_5_mlp_2_bias_to_fp16, weight = flow_net_res_blocks_5_mlp_2_weight_to_fp16, x = input_265_cast_fp16)[name = tensor("linear_101_cast_fp16")]; + tensor var_1545_cast_fp16 = mul(x = var_1516_cast_fp16_2, y = linear_101_cast_fp16)[name = tensor("op_1545_cast_fp16")]; + tensor x_173_cast_fp16 = add(x = x_167_cast_fp16, y = var_1545_cast_fp16)[name = tensor("x_173_cast_fp16")]; + tensor linear_102_cast_fp16 = linear(bias = flow_net_final_layer_adaLN_modulation_1_bias_to_fp16, weight = flow_net_final_layer_adaLN_modulation_1_weight_to_fp16, x = input_219_cast_fp16)[name = tensor("linear_102_cast_fp16")]; + tensor var_1554_split_sizes_0 = const()[name = tensor("op_1554_split_sizes_0"), val = tensor([512, 512])]; + tensor var_1554_axis_0 = const()[name = tensor("op_1554_axis_0"), val = tensor(-1)]; + tensor var_1554_cast_fp16_0, tensor var_1554_cast_fp16_1 = split(axis = var_1554_axis_0, split_sizes = var_1554_split_sizes_0, x = linear_102_cast_fp16)[name = tensor("op_1554_cast_fp16")]; + tensor mean_55_axes_0 = const()[name = tensor("mean_55_axes_0"), val = tensor([-1])]; + tensor mean_55_keep_dims_0 = const()[name = tensor("mean_55_keep_dims_0"), val = tensor(true)]; + tensor mean_55_cast_fp16 = reduce_mean(axes = mean_55_axes_0, keep_dims = mean_55_keep_dims_0, x = x_173_cast_fp16)[name = tensor("mean_55_cast_fp16")]; + tensor sub_43_cast_fp16 = sub(x = x_173_cast_fp16, y = mean_55_cast_fp16)[name = tensor("sub_43_cast_fp16")]; + tensor square_35_cast_fp16 = square(x = sub_43_cast_fp16)[name = tensor("square_35_cast_fp16")]; + tensor reduce_mean_71_axes_0 = const()[name = tensor("reduce_mean_71_axes_0"), val = tensor([-1])]; + tensor reduce_mean_71_keep_dims_0 = const()[name = tensor("reduce_mean_71_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_71_cast_fp16 = reduce_mean(axes = reduce_mean_71_axes_0, keep_dims = reduce_mean_71_keep_dims_0, x = square_35_cast_fp16)[name = tensor("reduce_mean_71_cast_fp16")]; + tensor var_1561_to_fp16 = const()[name = tensor("op_1561_to_fp16"), val = tensor(0x1.1p-20)]; + tensor var_1562_cast_fp16 = add(x = reduce_mean_71_cast_fp16, y = var_1561_to_fp16)[name = tensor("op_1562_cast_fp16")]; + tensor var_1563_cast_fp16 = sqrt(x = var_1562_cast_fp16)[name = tensor("op_1563_cast_fp16")]; + tensor x_175_cast_fp16 = real_div(x = sub_43_cast_fp16, y = var_1563_cast_fp16)[name = tensor("x_175_cast_fp16")]; + tensor var_1565_promoted_to_fp16 = const()[name = tensor("op_1565_promoted_to_fp16"), val = tensor(0x1p+0)]; + tensor var_1566_cast_fp16 = add(x = var_1554_cast_fp16_1, y = var_1565_promoted_to_fp16)[name = tensor("op_1566_cast_fp16")]; + tensor var_1567_cast_fp16 = mul(x = x_175_cast_fp16, y = var_1566_cast_fp16)[name = tensor("op_1567_cast_fp16")]; + tensor input_269_cast_fp16 = add(x = var_1567_cast_fp16, y = var_1554_cast_fp16_0)[name = tensor("input_269_cast_fp16")]; + tensor linear_103_cast_fp16 = linear(bias = flow_net_final_layer_linear_bias_to_fp16, weight = flow_net_final_layer_linear_weight_to_fp16, x = input_269_cast_fp16)[name = tensor("linear_103_cast_fp16")]; + tensor var_1578_to_fp16 = const()[name = tensor("op_1578_to_fp16"), val = tensor(0x1p-3)]; + tensor var_1579_cast_fp16 = mul(x = linear_103_cast_fp16, y = var_1578_to_fp16)[name = tensor("op_1579_cast_fp16")]; + tensor input_271_cast_fp16 = add(x = input_203_cast_fp16, y = var_1579_cast_fp16)[name = tensor("input_271_cast_fp16")]; + tensor linear_104_cast_fp16 = linear(bias = flow_net_input_proj_bias_to_fp16, weight = flow_net_input_proj_weight_to_fp16, x = input_271_cast_fp16)[name = tensor("linear_104_cast_fp16")]; + tensor input_275_to_fp16 = const()[name = tensor("input_275_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(19003904)))]; + tensor input_277_cast_fp16 = silu(x = input_275_to_fp16)[name = tensor("input_277_cast_fp16")]; + tensor linear_106_cast_fp16 = linear(bias = flow_net_time_embed_0_mlp_2_bias_to_fp16, weight = flow_net_time_embed_0_mlp_2_weight_to_fp16, x = input_277_cast_fp16)[name = tensor("linear_106_cast_fp16")]; + tensor reduce_mean_72_axes_0 = const()[name = tensor("reduce_mean_72_axes_0"), val = tensor([-1])]; + tensor reduce_mean_72_keep_dims_0 = const()[name = tensor("reduce_mean_72_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_72_cast_fp16 = reduce_mean(axes = reduce_mean_72_axes_0, keep_dims = reduce_mean_72_keep_dims_0, x = linear_106_cast_fp16)[name = tensor("reduce_mean_72_cast_fp16")]; + tensor sub_44_cast_fp16 = sub(x = linear_106_cast_fp16, y = reduce_mean_72_cast_fp16)[name = tensor("sub_44_cast_fp16")]; + tensor square_36_cast_fp16 = square(x = sub_44_cast_fp16)[name = tensor("square_36_cast_fp16")]; + tensor reduce_mean_73_axes_0 = const()[name = tensor("reduce_mean_73_axes_0"), val = tensor([-1])]; + tensor reduce_mean_73_keep_dims_0 = const()[name = tensor("reduce_mean_73_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_73_cast_fp16 = reduce_mean(axes = reduce_mean_73_axes_0, keep_dims = reduce_mean_73_keep_dims_0, x = square_36_cast_fp16)[name = tensor("reduce_mean_73_cast_fp16")]; + tensor real_div_8_to_fp16 = const()[name = tensor("real_div_8_to_fp16"), val = tensor(0x1.008p+0)]; + tensor mul_8_cast_fp16 = mul(x = reduce_mean_73_cast_fp16, y = real_div_8_to_fp16)[name = tensor("mul_8_cast_fp16")]; + tensor var_1645_to_fp16 = const()[name = tensor("op_1645_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_73_cast_fp16_0 = add(x = mul_8_cast_fp16, y = var_1645_to_fp16)[name = tensor("var_73_cast_fp16")]; + tensor var_1648_epsilon_0 = const()[name = tensor("op_1648_epsilon_0"), val = tensor(0x1.197998p-40)]; + tensor var_1648_cast_fp16 = rsqrt(epsilon = var_1648_epsilon_0, x = var_73_cast_fp16_0)[name = tensor("op_1648_cast_fp16")]; + tensor var_1649_cast_fp16 = mul(x = const_3_to_fp16, y = var_1648_cast_fp16)[name = tensor("op_1649_cast_fp16")]; + tensor var_1650_cast_fp16 = mul(x = linear_106_cast_fp16, y = var_1649_cast_fp16)[name = tensor("op_1650_cast_fp16")]; + tensor input_281_to_fp16 = const()[name = tensor("input_281_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(19004992)))]; + tensor input_283_cast_fp16 = silu(x = input_281_to_fp16)[name = tensor("input_283_cast_fp16")]; + tensor linear_108_cast_fp16 = linear(bias = flow_net_time_embed_1_mlp_2_bias_to_fp16, weight = flow_net_time_embed_1_mlp_2_weight_to_fp16, x = input_283_cast_fp16)[name = tensor("linear_108_cast_fp16")]; + tensor reduce_mean_74_axes_0 = const()[name = tensor("reduce_mean_74_axes_0"), val = tensor([-1])]; + tensor reduce_mean_74_keep_dims_0 = const()[name = tensor("reduce_mean_74_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_74_cast_fp16 = reduce_mean(axes = reduce_mean_74_axes_0, keep_dims = reduce_mean_74_keep_dims_0, x = linear_108_cast_fp16)[name = tensor("reduce_mean_74_cast_fp16")]; + tensor sub_46_cast_fp16 = sub(x = linear_108_cast_fp16, y = reduce_mean_74_cast_fp16)[name = tensor("sub_46_cast_fp16")]; + tensor square_37_cast_fp16 = square(x = sub_46_cast_fp16)[name = tensor("square_37_cast_fp16")]; + tensor reduce_mean_75_axes_0 = const()[name = tensor("reduce_mean_75_axes_0"), val = tensor([-1])]; + tensor reduce_mean_75_keep_dims_0 = const()[name = tensor("reduce_mean_75_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_75_cast_fp16 = reduce_mean(axes = reduce_mean_75_axes_0, keep_dims = reduce_mean_75_keep_dims_0, x = square_37_cast_fp16)[name = tensor("reduce_mean_75_cast_fp16")]; + tensor real_div_9_to_fp16 = const()[name = tensor("real_div_9_to_fp16"), val = tensor(0x1.008p+0)]; + tensor mul_9_cast_fp16 = mul(x = reduce_mean_75_cast_fp16, y = real_div_9_to_fp16)[name = tensor("mul_9_cast_fp16")]; + tensor var_1682_to_fp16 = const()[name = tensor("op_1682_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_75_cast_fp16_0 = add(x = mul_9_cast_fp16, y = var_1682_to_fp16)[name = tensor("var_75_cast_fp16")]; + tensor var_1685_epsilon_0 = const()[name = tensor("op_1685_epsilon_0"), val = tensor(0x1.197998p-40)]; + tensor var_1685_cast_fp16 = rsqrt(epsilon = var_1685_epsilon_0, x = var_75_cast_fp16_0)[name = tensor("op_1685_cast_fp16")]; + tensor var_1686_cast_fp16 = mul(x = const_5_to_fp16, y = var_1685_cast_fp16)[name = tensor("op_1686_cast_fp16")]; + tensor var_1687_cast_fp16 = mul(x = linear_108_cast_fp16, y = var_1686_cast_fp16)[name = tensor("op_1687_cast_fp16")]; + tensor var_1699_cast_fp16 = add(x = var_1650_cast_fp16, y = var_1687_cast_fp16)[name = tensor("op_1699_cast_fp16")]; + tensor _inversed_t_combined_9_y_0_to_fp16 = const()[name = tensor("_inversed_t_combined_9_y_0_to_fp16"), val = tensor(0x1p-1)]; + tensor _inversed_t_combined_9_cast_fp16 = mul(x = var_1699_cast_fp16, y = _inversed_t_combined_9_y_0_to_fp16)[name = tensor("_inversed_t_combined_9_cast_fp16")]; + tensor input_285_cast_fp16 = add(x = _inversed_t_combined_9_cast_fp16, y = linear_5_cast_fp16)[name = tensor("input_285_cast_fp16")]; + tensor input_287_cast_fp16 = silu(x = input_285_cast_fp16)[name = tensor("input_287_cast_fp16")]; + tensor linear_110_cast_fp16 = linear(bias = flow_net_res_blocks_0_adaLN_modulation_1_bias_to_fp16, weight = flow_net_res_blocks_0_adaLN_modulation_1_weight_to_fp16, x = input_287_cast_fp16)[name = tensor("linear_110_cast_fp16")]; + tensor var_1714_split_sizes_0 = const()[name = tensor("op_1714_split_sizes_0"), val = tensor([512, 512, 512])]; + tensor var_1714_axis_0 = const()[name = tensor("op_1714_axis_0"), val = tensor(-1)]; + tensor var_1714_cast_fp16_0, tensor var_1714_cast_fp16_1, tensor var_1714_cast_fp16_2 = split(axis = var_1714_axis_0, split_sizes = var_1714_split_sizes_0, x = linear_110_cast_fp16)[name = tensor("op_1714_cast_fp16")]; + tensor mean_57_axes_0 = const()[name = tensor("mean_57_axes_0"), val = tensor([-1])]; + tensor mean_57_keep_dims_0 = const()[name = tensor("mean_57_keep_dims_0"), val = tensor(true)]; + tensor mean_57_cast_fp16 = reduce_mean(axes = mean_57_axes_0, keep_dims = mean_57_keep_dims_0, x = linear_104_cast_fp16)[name = tensor("mean_57_cast_fp16")]; + tensor sub_48_cast_fp16 = sub(x = linear_104_cast_fp16, y = mean_57_cast_fp16)[name = tensor("sub_48_cast_fp16")]; + tensor square_38_cast_fp16 = square(x = sub_48_cast_fp16)[name = tensor("square_38_cast_fp16")]; + tensor reduce_mean_77_axes_0 = const()[name = tensor("reduce_mean_77_axes_0"), val = tensor([-1])]; + tensor reduce_mean_77_keep_dims_0 = const()[name = tensor("reduce_mean_77_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_77_cast_fp16 = reduce_mean(axes = reduce_mean_77_axes_0, keep_dims = reduce_mean_77_keep_dims_0, x = square_38_cast_fp16)[name = tensor("reduce_mean_77_cast_fp16")]; + tensor var_1724_to_fp16 = const()[name = tensor("op_1724_to_fp16"), val = tensor(0x1.1p-20)]; + tensor var_1725_cast_fp16 = add(x = reduce_mean_77_cast_fp16, y = var_1724_to_fp16)[name = tensor("op_1725_cast_fp16")]; + tensor var_1726_cast_fp16 = sqrt(x = var_1725_cast_fp16)[name = tensor("op_1726_cast_fp16")]; + tensor x_183_cast_fp16 = real_div(x = sub_48_cast_fp16, y = var_1726_cast_fp16)[name = tensor("x_183_cast_fp16")]; + tensor var_1728_cast_fp16 = mul(x = x_183_cast_fp16, y = flow_net_res_blocks_0_in_ln_weight_to_fp16)[name = tensor("op_1728_cast_fp16")]; + tensor x_185_cast_fp16 = add(x = var_1728_cast_fp16, y = flow_net_res_blocks_0_in_ln_bias_to_fp16)[name = tensor("x_185_cast_fp16")]; + tensor var_1730_promoted_to_fp16 = const()[name = tensor("op_1730_promoted_to_fp16"), val = tensor(0x1p+0)]; + tensor var_1731_cast_fp16 = add(x = var_1714_cast_fp16_1, y = var_1730_promoted_to_fp16)[name = tensor("op_1731_cast_fp16")]; + tensor var_1732_cast_fp16 = mul(x = x_185_cast_fp16, y = var_1731_cast_fp16)[name = tensor("op_1732_cast_fp16")]; + tensor input_289_cast_fp16 = add(x = var_1732_cast_fp16, y = var_1714_cast_fp16_0)[name = tensor("input_289_cast_fp16")]; + tensor linear_111_cast_fp16 = linear(bias = flow_net_res_blocks_0_mlp_0_bias_to_fp16, weight = flow_net_res_blocks_0_mlp_0_weight_to_fp16, x = input_289_cast_fp16)[name = tensor("linear_111_cast_fp16")]; + tensor input_293_cast_fp16 = silu(x = linear_111_cast_fp16)[name = tensor("input_293_cast_fp16")]; + tensor linear_112_cast_fp16 = linear(bias = flow_net_res_blocks_0_mlp_2_bias_to_fp16, weight = flow_net_res_blocks_0_mlp_2_weight_to_fp16, x = input_293_cast_fp16)[name = tensor("linear_112_cast_fp16")]; + tensor var_1743_cast_fp16 = mul(x = var_1714_cast_fp16_2, y = linear_112_cast_fp16)[name = tensor("op_1743_cast_fp16")]; + tensor x_187_cast_fp16 = add(x = linear_104_cast_fp16, y = var_1743_cast_fp16)[name = tensor("x_187_cast_fp16")]; + tensor linear_113_cast_fp16 = linear(bias = flow_net_res_blocks_1_adaLN_modulation_1_bias_to_fp16, weight = flow_net_res_blocks_1_adaLN_modulation_1_weight_to_fp16, x = input_287_cast_fp16)[name = tensor("linear_113_cast_fp16")]; + tensor var_1753_split_sizes_0 = const()[name = tensor("op_1753_split_sizes_0"), val = tensor([512, 512, 512])]; + tensor var_1753_axis_0 = const()[name = tensor("op_1753_axis_0"), val = tensor(-1)]; + tensor var_1753_cast_fp16_0, tensor var_1753_cast_fp16_1, tensor var_1753_cast_fp16_2 = split(axis = var_1753_axis_0, split_sizes = var_1753_split_sizes_0, x = linear_113_cast_fp16)[name = tensor("op_1753_cast_fp16")]; + tensor mean_59_axes_0 = const()[name = tensor("mean_59_axes_0"), val = tensor([-1])]; + tensor mean_59_keep_dims_0 = const()[name = tensor("mean_59_keep_dims_0"), val = tensor(true)]; + tensor mean_59_cast_fp16 = reduce_mean(axes = mean_59_axes_0, keep_dims = mean_59_keep_dims_0, x = x_187_cast_fp16)[name = tensor("mean_59_cast_fp16")]; + tensor sub_49_cast_fp16 = sub(x = x_187_cast_fp16, y = mean_59_cast_fp16)[name = tensor("sub_49_cast_fp16")]; + tensor square_39_cast_fp16 = square(x = sub_49_cast_fp16)[name = tensor("square_39_cast_fp16")]; + tensor reduce_mean_79_axes_0 = const()[name = tensor("reduce_mean_79_axes_0"), val = tensor([-1])]; + tensor reduce_mean_79_keep_dims_0 = const()[name = tensor("reduce_mean_79_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_79_cast_fp16 = reduce_mean(axes = reduce_mean_79_axes_0, keep_dims = reduce_mean_79_keep_dims_0, x = square_39_cast_fp16)[name = tensor("reduce_mean_79_cast_fp16")]; + tensor var_1763_to_fp16 = const()[name = tensor("op_1763_to_fp16"), val = tensor(0x1.1p-20)]; + tensor var_1764_cast_fp16 = add(x = reduce_mean_79_cast_fp16, y = var_1763_to_fp16)[name = tensor("op_1764_cast_fp16")]; + tensor var_1765_cast_fp16 = sqrt(x = var_1764_cast_fp16)[name = tensor("op_1765_cast_fp16")]; + tensor x_189_cast_fp16 = real_div(x = sub_49_cast_fp16, y = var_1765_cast_fp16)[name = tensor("x_189_cast_fp16")]; + tensor var_1767_cast_fp16 = mul(x = x_189_cast_fp16, y = flow_net_res_blocks_1_in_ln_weight_to_fp16)[name = tensor("op_1767_cast_fp16")]; + tensor x_191_cast_fp16 = add(x = var_1767_cast_fp16, y = flow_net_res_blocks_1_in_ln_bias_to_fp16)[name = tensor("x_191_cast_fp16")]; + tensor var_1769_promoted_to_fp16 = const()[name = tensor("op_1769_promoted_to_fp16"), val = tensor(0x1p+0)]; + tensor var_1770_cast_fp16 = add(x = var_1753_cast_fp16_1, y = var_1769_promoted_to_fp16)[name = tensor("op_1770_cast_fp16")]; + tensor var_1771_cast_fp16 = mul(x = x_191_cast_fp16, y = var_1770_cast_fp16)[name = tensor("op_1771_cast_fp16")]; + tensor input_297_cast_fp16 = add(x = var_1771_cast_fp16, y = var_1753_cast_fp16_0)[name = tensor("input_297_cast_fp16")]; + tensor linear_114_cast_fp16 = linear(bias = flow_net_res_blocks_1_mlp_0_bias_to_fp16, weight = flow_net_res_blocks_1_mlp_0_weight_to_fp16, x = input_297_cast_fp16)[name = tensor("linear_114_cast_fp16")]; + tensor input_301_cast_fp16 = silu(x = linear_114_cast_fp16)[name = tensor("input_301_cast_fp16")]; + tensor linear_115_cast_fp16 = linear(bias = flow_net_res_blocks_1_mlp_2_bias_to_fp16, weight = flow_net_res_blocks_1_mlp_2_weight_to_fp16, x = input_301_cast_fp16)[name = tensor("linear_115_cast_fp16")]; + tensor var_1782_cast_fp16 = mul(x = var_1753_cast_fp16_2, y = linear_115_cast_fp16)[name = tensor("op_1782_cast_fp16")]; + tensor x_193_cast_fp16 = add(x = x_187_cast_fp16, y = var_1782_cast_fp16)[name = tensor("x_193_cast_fp16")]; + tensor linear_116_cast_fp16 = linear(bias = flow_net_res_blocks_2_adaLN_modulation_1_bias_to_fp16, weight = flow_net_res_blocks_2_adaLN_modulation_1_weight_to_fp16, x = input_287_cast_fp16)[name = tensor("linear_116_cast_fp16")]; + tensor var_1792_split_sizes_0 = const()[name = tensor("op_1792_split_sizes_0"), val = tensor([512, 512, 512])]; + tensor var_1792_axis_0 = const()[name = tensor("op_1792_axis_0"), val = tensor(-1)]; + tensor var_1792_cast_fp16_0, tensor var_1792_cast_fp16_1, tensor var_1792_cast_fp16_2 = split(axis = var_1792_axis_0, split_sizes = var_1792_split_sizes_0, x = linear_116_cast_fp16)[name = tensor("op_1792_cast_fp16")]; + tensor mean_61_axes_0 = const()[name = tensor("mean_61_axes_0"), val = tensor([-1])]; + tensor mean_61_keep_dims_0 = const()[name = tensor("mean_61_keep_dims_0"), val = tensor(true)]; + tensor mean_61_cast_fp16 = reduce_mean(axes = mean_61_axes_0, keep_dims = mean_61_keep_dims_0, x = x_193_cast_fp16)[name = tensor("mean_61_cast_fp16")]; + tensor sub_50_cast_fp16 = sub(x = x_193_cast_fp16, y = mean_61_cast_fp16)[name = tensor("sub_50_cast_fp16")]; + tensor square_40_cast_fp16 = square(x = sub_50_cast_fp16)[name = tensor("square_40_cast_fp16")]; + tensor reduce_mean_81_axes_0 = const()[name = tensor("reduce_mean_81_axes_0"), val = tensor([-1])]; + tensor reduce_mean_81_keep_dims_0 = const()[name = tensor("reduce_mean_81_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_81_cast_fp16 = reduce_mean(axes = reduce_mean_81_axes_0, keep_dims = reduce_mean_81_keep_dims_0, x = square_40_cast_fp16)[name = tensor("reduce_mean_81_cast_fp16")]; + tensor var_1802_to_fp16 = const()[name = tensor("op_1802_to_fp16"), val = tensor(0x1.1p-20)]; + tensor var_1803_cast_fp16 = add(x = reduce_mean_81_cast_fp16, y = var_1802_to_fp16)[name = tensor("op_1803_cast_fp16")]; + tensor var_1804_cast_fp16 = sqrt(x = var_1803_cast_fp16)[name = tensor("op_1804_cast_fp16")]; + tensor x_195_cast_fp16 = real_div(x = sub_50_cast_fp16, y = var_1804_cast_fp16)[name = tensor("x_195_cast_fp16")]; + tensor var_1806_cast_fp16 = mul(x = x_195_cast_fp16, y = flow_net_res_blocks_2_in_ln_weight_to_fp16)[name = tensor("op_1806_cast_fp16")]; + tensor x_197_cast_fp16 = add(x = var_1806_cast_fp16, y = flow_net_res_blocks_2_in_ln_bias_to_fp16)[name = tensor("x_197_cast_fp16")]; + tensor var_1808_promoted_to_fp16 = const()[name = tensor("op_1808_promoted_to_fp16"), val = tensor(0x1p+0)]; + tensor var_1809_cast_fp16 = add(x = var_1792_cast_fp16_1, y = var_1808_promoted_to_fp16)[name = tensor("op_1809_cast_fp16")]; + tensor var_1810_cast_fp16 = mul(x = x_197_cast_fp16, y = var_1809_cast_fp16)[name = tensor("op_1810_cast_fp16")]; + tensor input_305_cast_fp16 = add(x = var_1810_cast_fp16, y = var_1792_cast_fp16_0)[name = tensor("input_305_cast_fp16")]; + tensor linear_117_cast_fp16 = linear(bias = flow_net_res_blocks_2_mlp_0_bias_to_fp16, weight = flow_net_res_blocks_2_mlp_0_weight_to_fp16, x = input_305_cast_fp16)[name = tensor("linear_117_cast_fp16")]; + tensor input_309_cast_fp16 = silu(x = linear_117_cast_fp16)[name = tensor("input_309_cast_fp16")]; + tensor linear_118_cast_fp16 = linear(bias = flow_net_res_blocks_2_mlp_2_bias_to_fp16, weight = flow_net_res_blocks_2_mlp_2_weight_to_fp16, x = input_309_cast_fp16)[name = tensor("linear_118_cast_fp16")]; + tensor var_1821_cast_fp16 = mul(x = var_1792_cast_fp16_2, y = linear_118_cast_fp16)[name = tensor("op_1821_cast_fp16")]; + tensor x_199_cast_fp16 = add(x = x_193_cast_fp16, y = var_1821_cast_fp16)[name = tensor("x_199_cast_fp16")]; + tensor linear_119_cast_fp16 = linear(bias = flow_net_res_blocks_3_adaLN_modulation_1_bias_to_fp16, weight = flow_net_res_blocks_3_adaLN_modulation_1_weight_to_fp16, x = input_287_cast_fp16)[name = tensor("linear_119_cast_fp16")]; + tensor var_1831_split_sizes_0 = const()[name = tensor("op_1831_split_sizes_0"), val = tensor([512, 512, 512])]; + tensor var_1831_axis_0 = const()[name = tensor("op_1831_axis_0"), val = tensor(-1)]; + tensor var_1831_cast_fp16_0, tensor var_1831_cast_fp16_1, tensor var_1831_cast_fp16_2 = split(axis = var_1831_axis_0, split_sizes = var_1831_split_sizes_0, x = linear_119_cast_fp16)[name = tensor("op_1831_cast_fp16")]; + tensor mean_63_axes_0 = const()[name = tensor("mean_63_axes_0"), val = tensor([-1])]; + tensor mean_63_keep_dims_0 = const()[name = tensor("mean_63_keep_dims_0"), val = tensor(true)]; + tensor mean_63_cast_fp16 = reduce_mean(axes = mean_63_axes_0, keep_dims = mean_63_keep_dims_0, x = x_199_cast_fp16)[name = tensor("mean_63_cast_fp16")]; + tensor sub_51_cast_fp16 = sub(x = x_199_cast_fp16, y = mean_63_cast_fp16)[name = tensor("sub_51_cast_fp16")]; + tensor square_41_cast_fp16 = square(x = sub_51_cast_fp16)[name = tensor("square_41_cast_fp16")]; + tensor reduce_mean_83_axes_0 = const()[name = tensor("reduce_mean_83_axes_0"), val = tensor([-1])]; + tensor reduce_mean_83_keep_dims_0 = const()[name = tensor("reduce_mean_83_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_83_cast_fp16 = reduce_mean(axes = reduce_mean_83_axes_0, keep_dims = reduce_mean_83_keep_dims_0, x = square_41_cast_fp16)[name = tensor("reduce_mean_83_cast_fp16")]; + tensor var_1841_to_fp16 = const()[name = tensor("op_1841_to_fp16"), val = tensor(0x1.1p-20)]; + tensor var_1842_cast_fp16 = add(x = reduce_mean_83_cast_fp16, y = var_1841_to_fp16)[name = tensor("op_1842_cast_fp16")]; + tensor var_1843_cast_fp16 = sqrt(x = var_1842_cast_fp16)[name = tensor("op_1843_cast_fp16")]; + tensor x_201_cast_fp16 = real_div(x = sub_51_cast_fp16, y = var_1843_cast_fp16)[name = tensor("x_201_cast_fp16")]; + tensor var_1845_cast_fp16 = mul(x = x_201_cast_fp16, y = flow_net_res_blocks_3_in_ln_weight_to_fp16)[name = tensor("op_1845_cast_fp16")]; + tensor x_203_cast_fp16 = add(x = var_1845_cast_fp16, y = flow_net_res_blocks_3_in_ln_bias_to_fp16)[name = tensor("x_203_cast_fp16")]; + tensor var_1847_promoted_to_fp16 = const()[name = tensor("op_1847_promoted_to_fp16"), val = tensor(0x1p+0)]; + tensor var_1848_cast_fp16 = add(x = var_1831_cast_fp16_1, y = var_1847_promoted_to_fp16)[name = tensor("op_1848_cast_fp16")]; + tensor var_1849_cast_fp16 = mul(x = x_203_cast_fp16, y = var_1848_cast_fp16)[name = tensor("op_1849_cast_fp16")]; + tensor input_313_cast_fp16 = add(x = var_1849_cast_fp16, y = var_1831_cast_fp16_0)[name = tensor("input_313_cast_fp16")]; + tensor linear_120_cast_fp16 = linear(bias = flow_net_res_blocks_3_mlp_0_bias_to_fp16, weight = flow_net_res_blocks_3_mlp_0_weight_to_fp16, x = input_313_cast_fp16)[name = tensor("linear_120_cast_fp16")]; + tensor input_317_cast_fp16 = silu(x = linear_120_cast_fp16)[name = tensor("input_317_cast_fp16")]; + tensor linear_121_cast_fp16 = linear(bias = flow_net_res_blocks_3_mlp_2_bias_to_fp16, weight = flow_net_res_blocks_3_mlp_2_weight_to_fp16, x = input_317_cast_fp16)[name = tensor("linear_121_cast_fp16")]; + tensor var_1860_cast_fp16 = mul(x = var_1831_cast_fp16_2, y = linear_121_cast_fp16)[name = tensor("op_1860_cast_fp16")]; + tensor x_205_cast_fp16 = add(x = x_199_cast_fp16, y = var_1860_cast_fp16)[name = tensor("x_205_cast_fp16")]; + tensor linear_122_cast_fp16 = linear(bias = flow_net_res_blocks_4_adaLN_modulation_1_bias_to_fp16, weight = flow_net_res_blocks_4_adaLN_modulation_1_weight_to_fp16, x = input_287_cast_fp16)[name = tensor("linear_122_cast_fp16")]; + tensor var_1870_split_sizes_0 = const()[name = tensor("op_1870_split_sizes_0"), val = tensor([512, 512, 512])]; + tensor var_1870_axis_0 = const()[name = tensor("op_1870_axis_0"), val = tensor(-1)]; + tensor var_1870_cast_fp16_0, tensor var_1870_cast_fp16_1, tensor var_1870_cast_fp16_2 = split(axis = var_1870_axis_0, split_sizes = var_1870_split_sizes_0, x = linear_122_cast_fp16)[name = tensor("op_1870_cast_fp16")]; + tensor mean_65_axes_0 = const()[name = tensor("mean_65_axes_0"), val = tensor([-1])]; + tensor mean_65_keep_dims_0 = const()[name = tensor("mean_65_keep_dims_0"), val = tensor(true)]; + tensor mean_65_cast_fp16 = reduce_mean(axes = mean_65_axes_0, keep_dims = mean_65_keep_dims_0, x = x_205_cast_fp16)[name = tensor("mean_65_cast_fp16")]; + tensor sub_52_cast_fp16 = sub(x = x_205_cast_fp16, y = mean_65_cast_fp16)[name = tensor("sub_52_cast_fp16")]; + tensor square_42_cast_fp16 = square(x = sub_52_cast_fp16)[name = tensor("square_42_cast_fp16")]; + tensor reduce_mean_85_axes_0 = const()[name = tensor("reduce_mean_85_axes_0"), val = tensor([-1])]; + tensor reduce_mean_85_keep_dims_0 = const()[name = tensor("reduce_mean_85_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_85_cast_fp16 = reduce_mean(axes = reduce_mean_85_axes_0, keep_dims = reduce_mean_85_keep_dims_0, x = square_42_cast_fp16)[name = tensor("reduce_mean_85_cast_fp16")]; + tensor var_1880_to_fp16 = const()[name = tensor("op_1880_to_fp16"), val = tensor(0x1.1p-20)]; + tensor var_1881_cast_fp16 = add(x = reduce_mean_85_cast_fp16, y = var_1880_to_fp16)[name = tensor("op_1881_cast_fp16")]; + tensor var_1882_cast_fp16 = sqrt(x = var_1881_cast_fp16)[name = tensor("op_1882_cast_fp16")]; + tensor x_207_cast_fp16 = real_div(x = sub_52_cast_fp16, y = var_1882_cast_fp16)[name = tensor("x_207_cast_fp16")]; + tensor var_1884_cast_fp16 = mul(x = x_207_cast_fp16, y = flow_net_res_blocks_4_in_ln_weight_to_fp16)[name = tensor("op_1884_cast_fp16")]; + tensor x_209_cast_fp16 = add(x = var_1884_cast_fp16, y = flow_net_res_blocks_4_in_ln_bias_to_fp16)[name = tensor("x_209_cast_fp16")]; + tensor var_1886_promoted_to_fp16 = const()[name = tensor("op_1886_promoted_to_fp16"), val = tensor(0x1p+0)]; + tensor var_1887_cast_fp16 = add(x = var_1870_cast_fp16_1, y = var_1886_promoted_to_fp16)[name = tensor("op_1887_cast_fp16")]; + tensor var_1888_cast_fp16 = mul(x = x_209_cast_fp16, y = var_1887_cast_fp16)[name = tensor("op_1888_cast_fp16")]; + tensor input_321_cast_fp16 = add(x = var_1888_cast_fp16, y = var_1870_cast_fp16_0)[name = tensor("input_321_cast_fp16")]; + tensor linear_123_cast_fp16 = linear(bias = flow_net_res_blocks_4_mlp_0_bias_to_fp16, weight = flow_net_res_blocks_4_mlp_0_weight_to_fp16, x = input_321_cast_fp16)[name = tensor("linear_123_cast_fp16")]; + tensor input_325_cast_fp16 = silu(x = linear_123_cast_fp16)[name = tensor("input_325_cast_fp16")]; + tensor linear_124_cast_fp16 = linear(bias = flow_net_res_blocks_4_mlp_2_bias_to_fp16, weight = flow_net_res_blocks_4_mlp_2_weight_to_fp16, x = input_325_cast_fp16)[name = tensor("linear_124_cast_fp16")]; + tensor var_1899_cast_fp16 = mul(x = var_1870_cast_fp16_2, y = linear_124_cast_fp16)[name = tensor("op_1899_cast_fp16")]; + tensor x_211_cast_fp16 = add(x = x_205_cast_fp16, y = var_1899_cast_fp16)[name = tensor("x_211_cast_fp16")]; + tensor linear_125_cast_fp16 = linear(bias = flow_net_res_blocks_5_adaLN_modulation_1_bias_to_fp16, weight = flow_net_res_blocks_5_adaLN_modulation_1_weight_to_fp16, x = input_287_cast_fp16)[name = tensor("linear_125_cast_fp16")]; + tensor var_1909_split_sizes_0 = const()[name = tensor("op_1909_split_sizes_0"), val = tensor([512, 512, 512])]; + tensor var_1909_axis_0 = const()[name = tensor("op_1909_axis_0"), val = tensor(-1)]; + tensor var_1909_cast_fp16_0, tensor var_1909_cast_fp16_1, tensor var_1909_cast_fp16_2 = split(axis = var_1909_axis_0, split_sizes = var_1909_split_sizes_0, x = linear_125_cast_fp16)[name = tensor("op_1909_cast_fp16")]; + tensor mean_67_axes_0 = const()[name = tensor("mean_67_axes_0"), val = tensor([-1])]; + tensor mean_67_keep_dims_0 = const()[name = tensor("mean_67_keep_dims_0"), val = tensor(true)]; + tensor mean_67_cast_fp16 = reduce_mean(axes = mean_67_axes_0, keep_dims = mean_67_keep_dims_0, x = x_211_cast_fp16)[name = tensor("mean_67_cast_fp16")]; + tensor sub_53_cast_fp16 = sub(x = x_211_cast_fp16, y = mean_67_cast_fp16)[name = tensor("sub_53_cast_fp16")]; + tensor square_43_cast_fp16 = square(x = sub_53_cast_fp16)[name = tensor("square_43_cast_fp16")]; + tensor reduce_mean_87_axes_0 = const()[name = tensor("reduce_mean_87_axes_0"), val = tensor([-1])]; + tensor reduce_mean_87_keep_dims_0 = const()[name = tensor("reduce_mean_87_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_87_cast_fp16 = reduce_mean(axes = reduce_mean_87_axes_0, keep_dims = reduce_mean_87_keep_dims_0, x = square_43_cast_fp16)[name = tensor("reduce_mean_87_cast_fp16")]; + tensor var_1919_to_fp16 = const()[name = tensor("op_1919_to_fp16"), val = tensor(0x1.1p-20)]; + tensor var_1920_cast_fp16 = add(x = reduce_mean_87_cast_fp16, y = var_1919_to_fp16)[name = tensor("op_1920_cast_fp16")]; + tensor var_1921_cast_fp16 = sqrt(x = var_1920_cast_fp16)[name = tensor("op_1921_cast_fp16")]; + tensor x_213_cast_fp16 = real_div(x = sub_53_cast_fp16, y = var_1921_cast_fp16)[name = tensor("x_213_cast_fp16")]; + tensor var_1923_cast_fp16 = mul(x = x_213_cast_fp16, y = flow_net_res_blocks_5_in_ln_weight_to_fp16)[name = tensor("op_1923_cast_fp16")]; + tensor x_215_cast_fp16 = add(x = var_1923_cast_fp16, y = flow_net_res_blocks_5_in_ln_bias_to_fp16)[name = tensor("x_215_cast_fp16")]; + tensor var_1925_promoted_to_fp16 = const()[name = tensor("op_1925_promoted_to_fp16"), val = tensor(0x1p+0)]; + tensor var_1926_cast_fp16 = add(x = var_1909_cast_fp16_1, y = var_1925_promoted_to_fp16)[name = tensor("op_1926_cast_fp16")]; + tensor var_1927_cast_fp16 = mul(x = x_215_cast_fp16, y = var_1926_cast_fp16)[name = tensor("op_1927_cast_fp16")]; + tensor input_329_cast_fp16 = add(x = var_1927_cast_fp16, y = var_1909_cast_fp16_0)[name = tensor("input_329_cast_fp16")]; + tensor linear_126_cast_fp16 = linear(bias = flow_net_res_blocks_5_mlp_0_bias_to_fp16, weight = flow_net_res_blocks_5_mlp_0_weight_to_fp16, x = input_329_cast_fp16)[name = tensor("linear_126_cast_fp16")]; + tensor input_333_cast_fp16 = silu(x = linear_126_cast_fp16)[name = tensor("input_333_cast_fp16")]; + tensor linear_127_cast_fp16 = linear(bias = flow_net_res_blocks_5_mlp_2_bias_to_fp16, weight = flow_net_res_blocks_5_mlp_2_weight_to_fp16, x = input_333_cast_fp16)[name = tensor("linear_127_cast_fp16")]; + tensor var_1938_cast_fp16 = mul(x = var_1909_cast_fp16_2, y = linear_127_cast_fp16)[name = tensor("op_1938_cast_fp16")]; + tensor x_217_cast_fp16 = add(x = x_211_cast_fp16, y = var_1938_cast_fp16)[name = tensor("x_217_cast_fp16")]; + tensor linear_128_cast_fp16 = linear(bias = flow_net_final_layer_adaLN_modulation_1_bias_to_fp16, weight = flow_net_final_layer_adaLN_modulation_1_weight_to_fp16, x = input_287_cast_fp16)[name = tensor("linear_128_cast_fp16")]; + tensor var_1947_split_sizes_0 = const()[name = tensor("op_1947_split_sizes_0"), val = tensor([512, 512])]; + tensor var_1947_axis_0 = const()[name = tensor("op_1947_axis_0"), val = tensor(-1)]; + tensor var_1947_cast_fp16_0, tensor var_1947_cast_fp16_1 = split(axis = var_1947_axis_0, split_sizes = var_1947_split_sizes_0, x = linear_128_cast_fp16)[name = tensor("op_1947_cast_fp16")]; + tensor mean_69_axes_0 = const()[name = tensor("mean_69_axes_0"), val = tensor([-1])]; + tensor mean_69_keep_dims_0 = const()[name = tensor("mean_69_keep_dims_0"), val = tensor(true)]; + tensor mean_69_cast_fp16 = reduce_mean(axes = mean_69_axes_0, keep_dims = mean_69_keep_dims_0, x = x_217_cast_fp16)[name = tensor("mean_69_cast_fp16")]; + tensor sub_54_cast_fp16 = sub(x = x_217_cast_fp16, y = mean_69_cast_fp16)[name = tensor("sub_54_cast_fp16")]; + tensor square_44_cast_fp16 = square(x = sub_54_cast_fp16)[name = tensor("square_44_cast_fp16")]; + tensor reduce_mean_89_axes_0 = const()[name = tensor("reduce_mean_89_axes_0"), val = tensor([-1])]; + tensor reduce_mean_89_keep_dims_0 = const()[name = tensor("reduce_mean_89_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_89_cast_fp16 = reduce_mean(axes = reduce_mean_89_axes_0, keep_dims = reduce_mean_89_keep_dims_0, x = square_44_cast_fp16)[name = tensor("reduce_mean_89_cast_fp16")]; + tensor var_1954_to_fp16 = const()[name = tensor("op_1954_to_fp16"), val = tensor(0x1.1p-20)]; + tensor var_1955_cast_fp16 = add(x = reduce_mean_89_cast_fp16, y = var_1954_to_fp16)[name = tensor("op_1955_cast_fp16")]; + tensor var_1956_cast_fp16 = sqrt(x = var_1955_cast_fp16)[name = tensor("op_1956_cast_fp16")]; + tensor x_219_cast_fp16 = real_div(x = sub_54_cast_fp16, y = var_1956_cast_fp16)[name = tensor("x_219_cast_fp16")]; + tensor var_1958_promoted_to_fp16 = const()[name = tensor("op_1958_promoted_to_fp16"), val = tensor(0x1p+0)]; + tensor var_1959_cast_fp16 = add(x = var_1947_cast_fp16_1, y = var_1958_promoted_to_fp16)[name = tensor("op_1959_cast_fp16")]; + tensor var_1960_cast_fp16 = mul(x = x_219_cast_fp16, y = var_1959_cast_fp16)[name = tensor("op_1960_cast_fp16")]; + tensor input_337_cast_fp16 = add(x = var_1960_cast_fp16, y = var_1947_cast_fp16_0)[name = tensor("input_337_cast_fp16")]; + tensor linear_129_cast_fp16 = linear(bias = flow_net_final_layer_linear_bias_to_fp16, weight = flow_net_final_layer_linear_weight_to_fp16, x = input_337_cast_fp16)[name = tensor("linear_129_cast_fp16")]; + tensor var_1971_to_fp16 = const()[name = tensor("op_1971_to_fp16"), val = tensor(0x1p-3)]; + tensor var_1972_cast_fp16 = mul(x = linear_129_cast_fp16, y = var_1971_to_fp16)[name = tensor("op_1972_cast_fp16")]; + tensor input_339_cast_fp16 = add(x = input_271_cast_fp16, y = var_1972_cast_fp16)[name = tensor("input_339_cast_fp16")]; + tensor linear_130_cast_fp16 = linear(bias = flow_net_input_proj_bias_to_fp16, weight = flow_net_input_proj_weight_to_fp16, x = input_339_cast_fp16)[name = tensor("linear_130_cast_fp16")]; + tensor input_343_to_fp16 = const()[name = tensor("input_343_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(19006080)))]; + tensor input_345_cast_fp16 = silu(x = input_343_to_fp16)[name = tensor("input_345_cast_fp16")]; + tensor linear_132_cast_fp16 = linear(bias = flow_net_time_embed_0_mlp_2_bias_to_fp16, weight = flow_net_time_embed_0_mlp_2_weight_to_fp16, x = input_345_cast_fp16)[name = tensor("linear_132_cast_fp16")]; + tensor reduce_mean_90_axes_0 = const()[name = tensor("reduce_mean_90_axes_0"), val = tensor([-1])]; + tensor reduce_mean_90_keep_dims_0 = const()[name = tensor("reduce_mean_90_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_90_cast_fp16 = reduce_mean(axes = reduce_mean_90_axes_0, keep_dims = reduce_mean_90_keep_dims_0, x = linear_132_cast_fp16)[name = tensor("reduce_mean_90_cast_fp16")]; + tensor sub_55_cast_fp16 = sub(x = linear_132_cast_fp16, y = reduce_mean_90_cast_fp16)[name = tensor("sub_55_cast_fp16")]; + tensor square_45_cast_fp16 = square(x = sub_55_cast_fp16)[name = tensor("square_45_cast_fp16")]; + tensor reduce_mean_91_axes_0 = const()[name = tensor("reduce_mean_91_axes_0"), val = tensor([-1])]; + tensor reduce_mean_91_keep_dims_0 = const()[name = tensor("reduce_mean_91_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_91_cast_fp16 = reduce_mean(axes = reduce_mean_91_axes_0, keep_dims = reduce_mean_91_keep_dims_0, x = square_45_cast_fp16)[name = tensor("reduce_mean_91_cast_fp16")]; + tensor real_div_10_to_fp16 = const()[name = tensor("real_div_10_to_fp16"), val = tensor(0x1.008p+0)]; + tensor mul_10_cast_fp16 = mul(x = reduce_mean_91_cast_fp16, y = real_div_10_to_fp16)[name = tensor("mul_10_cast_fp16")]; + tensor var_2038_to_fp16 = const()[name = tensor("op_2038_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_91_cast_fp16 = add(x = mul_10_cast_fp16, y = var_2038_to_fp16)[name = tensor("var_91_cast_fp16")]; + tensor var_2041_epsilon_0 = const()[name = tensor("op_2041_epsilon_0"), val = tensor(0x1.197998p-40)]; + tensor var_2041_cast_fp16 = rsqrt(epsilon = var_2041_epsilon_0, x = var_91_cast_fp16)[name = tensor("op_2041_cast_fp16")]; + tensor var_2042_cast_fp16 = mul(x = const_3_to_fp16, y = var_2041_cast_fp16)[name = tensor("op_2042_cast_fp16")]; + tensor var_2043_cast_fp16 = mul(x = linear_132_cast_fp16, y = var_2042_cast_fp16)[name = tensor("op_2043_cast_fp16")]; + tensor input_349_to_fp16 = const()[name = tensor("input_349_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(19007168)))]; + tensor input_351_cast_fp16 = silu(x = input_349_to_fp16)[name = tensor("input_351_cast_fp16")]; + tensor linear_134_cast_fp16 = linear(bias = flow_net_time_embed_1_mlp_2_bias_to_fp16, weight = flow_net_time_embed_1_mlp_2_weight_to_fp16, x = input_351_cast_fp16)[name = tensor("linear_134_cast_fp16")]; + tensor reduce_mean_92_axes_0 = const()[name = tensor("reduce_mean_92_axes_0"), val = tensor([-1])]; + tensor reduce_mean_92_keep_dims_0 = const()[name = tensor("reduce_mean_92_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_92_cast_fp16 = reduce_mean(axes = reduce_mean_92_axes_0, keep_dims = reduce_mean_92_keep_dims_0, x = linear_134_cast_fp16)[name = tensor("reduce_mean_92_cast_fp16")]; + tensor sub_57_cast_fp16 = sub(x = linear_134_cast_fp16, y = reduce_mean_92_cast_fp16)[name = tensor("sub_57_cast_fp16")]; + tensor square_46_cast_fp16 = square(x = sub_57_cast_fp16)[name = tensor("square_46_cast_fp16")]; + tensor reduce_mean_93_axes_0 = const()[name = tensor("reduce_mean_93_axes_0"), val = tensor([-1])]; + tensor reduce_mean_93_keep_dims_0 = const()[name = tensor("reduce_mean_93_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_93_cast_fp16 = reduce_mean(axes = reduce_mean_93_axes_0, keep_dims = reduce_mean_93_keep_dims_0, x = square_46_cast_fp16)[name = tensor("reduce_mean_93_cast_fp16")]; + tensor real_div_11_to_fp16 = const()[name = tensor("real_div_11_to_fp16"), val = tensor(0x1.008p+0)]; + tensor mul_11_cast_fp16 = mul(x = reduce_mean_93_cast_fp16, y = real_div_11_to_fp16)[name = tensor("mul_11_cast_fp16")]; + tensor var_2075_to_fp16 = const()[name = tensor("op_2075_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_93_cast_fp16 = add(x = mul_11_cast_fp16, y = var_2075_to_fp16)[name = tensor("var_93_cast_fp16")]; + tensor var_2078_epsilon_0 = const()[name = tensor("op_2078_epsilon_0"), val = tensor(0x1.197998p-40)]; + tensor var_2078_cast_fp16 = rsqrt(epsilon = var_2078_epsilon_0, x = var_93_cast_fp16)[name = tensor("op_2078_cast_fp16")]; + tensor var_2079_cast_fp16 = mul(x = const_5_to_fp16, y = var_2078_cast_fp16)[name = tensor("op_2079_cast_fp16")]; + tensor var_2080_cast_fp16 = mul(x = linear_134_cast_fp16, y = var_2079_cast_fp16)[name = tensor("op_2080_cast_fp16")]; + tensor var_2092_cast_fp16 = add(x = var_2043_cast_fp16, y = var_2080_cast_fp16)[name = tensor("op_2092_cast_fp16")]; + tensor _inversed_t_combined_11_y_0_to_fp16 = const()[name = tensor("_inversed_t_combined_11_y_0_to_fp16"), val = tensor(0x1p-1)]; + tensor _inversed_t_combined_11_cast_fp16 = mul(x = var_2092_cast_fp16, y = _inversed_t_combined_11_y_0_to_fp16)[name = tensor("_inversed_t_combined_11_cast_fp16")]; + tensor input_353_cast_fp16 = add(x = _inversed_t_combined_11_cast_fp16, y = linear_5_cast_fp16)[name = tensor("input_353_cast_fp16")]; + tensor input_355_cast_fp16 = silu(x = input_353_cast_fp16)[name = tensor("input_355_cast_fp16")]; + tensor linear_136_cast_fp16 = linear(bias = flow_net_res_blocks_0_adaLN_modulation_1_bias_to_fp16, weight = flow_net_res_blocks_0_adaLN_modulation_1_weight_to_fp16, x = input_355_cast_fp16)[name = tensor("linear_136_cast_fp16")]; + tensor var_2107_split_sizes_0 = const()[name = tensor("op_2107_split_sizes_0"), val = tensor([512, 512, 512])]; + tensor var_2107_axis_0 = const()[name = tensor("op_2107_axis_0"), val = tensor(-1)]; + tensor var_2107_cast_fp16_0, tensor var_2107_cast_fp16_1, tensor var_2107_cast_fp16_2 = split(axis = var_2107_axis_0, split_sizes = var_2107_split_sizes_0, x = linear_136_cast_fp16)[name = tensor("op_2107_cast_fp16")]; + tensor mean_71_axes_0 = const()[name = tensor("mean_71_axes_0"), val = tensor([-1])]; + tensor mean_71_keep_dims_0 = const()[name = tensor("mean_71_keep_dims_0"), val = tensor(true)]; + tensor mean_71_cast_fp16 = reduce_mean(axes = mean_71_axes_0, keep_dims = mean_71_keep_dims_0, x = linear_130_cast_fp16)[name = tensor("mean_71_cast_fp16")]; + tensor sub_59_cast_fp16 = sub(x = linear_130_cast_fp16, y = mean_71_cast_fp16)[name = tensor("sub_59_cast_fp16")]; + tensor square_47_cast_fp16 = square(x = sub_59_cast_fp16)[name = tensor("square_47_cast_fp16")]; + tensor reduce_mean_95_axes_0 = const()[name = tensor("reduce_mean_95_axes_0"), val = tensor([-1])]; + tensor reduce_mean_95_keep_dims_0 = const()[name = tensor("reduce_mean_95_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_95_cast_fp16 = reduce_mean(axes = reduce_mean_95_axes_0, keep_dims = reduce_mean_95_keep_dims_0, x = square_47_cast_fp16)[name = tensor("reduce_mean_95_cast_fp16")]; + tensor var_2117_to_fp16 = const()[name = tensor("op_2117_to_fp16"), val = tensor(0x1.1p-20)]; + tensor var_2118_cast_fp16 = add(x = reduce_mean_95_cast_fp16, y = var_2117_to_fp16)[name = tensor("op_2118_cast_fp16")]; + tensor var_2119_cast_fp16 = sqrt(x = var_2118_cast_fp16)[name = tensor("op_2119_cast_fp16")]; + tensor x_227_cast_fp16 = real_div(x = sub_59_cast_fp16, y = var_2119_cast_fp16)[name = tensor("x_227_cast_fp16")]; + tensor var_2121_cast_fp16 = mul(x = x_227_cast_fp16, y = flow_net_res_blocks_0_in_ln_weight_to_fp16)[name = tensor("op_2121_cast_fp16")]; + tensor x_229_cast_fp16 = add(x = var_2121_cast_fp16, y = flow_net_res_blocks_0_in_ln_bias_to_fp16)[name = tensor("x_229_cast_fp16")]; + tensor var_2123_promoted_to_fp16 = const()[name = tensor("op_2123_promoted_to_fp16"), val = tensor(0x1p+0)]; + tensor var_2124_cast_fp16 = add(x = var_2107_cast_fp16_1, y = var_2123_promoted_to_fp16)[name = tensor("op_2124_cast_fp16")]; + tensor var_2125_cast_fp16 = mul(x = x_229_cast_fp16, y = var_2124_cast_fp16)[name = tensor("op_2125_cast_fp16")]; + tensor input_357_cast_fp16 = add(x = var_2125_cast_fp16, y = var_2107_cast_fp16_0)[name = tensor("input_357_cast_fp16")]; + tensor linear_137_cast_fp16 = linear(bias = flow_net_res_blocks_0_mlp_0_bias_to_fp16, weight = flow_net_res_blocks_0_mlp_0_weight_to_fp16, x = input_357_cast_fp16)[name = tensor("linear_137_cast_fp16")]; + tensor input_361_cast_fp16 = silu(x = linear_137_cast_fp16)[name = tensor("input_361_cast_fp16")]; + tensor linear_138_cast_fp16 = linear(bias = flow_net_res_blocks_0_mlp_2_bias_to_fp16, weight = flow_net_res_blocks_0_mlp_2_weight_to_fp16, x = input_361_cast_fp16)[name = tensor("linear_138_cast_fp16")]; + tensor var_2136_cast_fp16 = mul(x = var_2107_cast_fp16_2, y = linear_138_cast_fp16)[name = tensor("op_2136_cast_fp16")]; + tensor x_231_cast_fp16 = add(x = linear_130_cast_fp16, y = var_2136_cast_fp16)[name = tensor("x_231_cast_fp16")]; + tensor linear_139_cast_fp16 = linear(bias = flow_net_res_blocks_1_adaLN_modulation_1_bias_to_fp16, weight = flow_net_res_blocks_1_adaLN_modulation_1_weight_to_fp16, x = input_355_cast_fp16)[name = tensor("linear_139_cast_fp16")]; + tensor var_2146_split_sizes_0 = const()[name = tensor("op_2146_split_sizes_0"), val = tensor([512, 512, 512])]; + tensor var_2146_axis_0 = const()[name = tensor("op_2146_axis_0"), val = tensor(-1)]; + tensor var_2146_cast_fp16_0, tensor var_2146_cast_fp16_1, tensor var_2146_cast_fp16_2 = split(axis = var_2146_axis_0, split_sizes = var_2146_split_sizes_0, x = linear_139_cast_fp16)[name = tensor("op_2146_cast_fp16")]; + tensor mean_73_axes_0 = const()[name = tensor("mean_73_axes_0"), val = tensor([-1])]; + tensor mean_73_keep_dims_0 = const()[name = tensor("mean_73_keep_dims_0"), val = tensor(true)]; + tensor mean_73_cast_fp16 = reduce_mean(axes = mean_73_axes_0, keep_dims = mean_73_keep_dims_0, x = x_231_cast_fp16)[name = tensor("mean_73_cast_fp16")]; + tensor sub_60_cast_fp16 = sub(x = x_231_cast_fp16, y = mean_73_cast_fp16)[name = tensor("sub_60_cast_fp16")]; + tensor square_48_cast_fp16 = square(x = sub_60_cast_fp16)[name = tensor("square_48_cast_fp16")]; + tensor reduce_mean_97_axes_0 = const()[name = tensor("reduce_mean_97_axes_0"), val = tensor([-1])]; + tensor reduce_mean_97_keep_dims_0 = const()[name = tensor("reduce_mean_97_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_97_cast_fp16 = reduce_mean(axes = reduce_mean_97_axes_0, keep_dims = reduce_mean_97_keep_dims_0, x = square_48_cast_fp16)[name = tensor("reduce_mean_97_cast_fp16")]; + tensor var_2156_to_fp16 = const()[name = tensor("op_2156_to_fp16"), val = tensor(0x1.1p-20)]; + tensor var_2157_cast_fp16 = add(x = reduce_mean_97_cast_fp16, y = var_2156_to_fp16)[name = tensor("op_2157_cast_fp16")]; + tensor var_2158_cast_fp16 = sqrt(x = var_2157_cast_fp16)[name = tensor("op_2158_cast_fp16")]; + tensor x_233_cast_fp16 = real_div(x = sub_60_cast_fp16, y = var_2158_cast_fp16)[name = tensor("x_233_cast_fp16")]; + tensor var_2160_cast_fp16 = mul(x = x_233_cast_fp16, y = flow_net_res_blocks_1_in_ln_weight_to_fp16)[name = tensor("op_2160_cast_fp16")]; + tensor x_235_cast_fp16 = add(x = var_2160_cast_fp16, y = flow_net_res_blocks_1_in_ln_bias_to_fp16)[name = tensor("x_235_cast_fp16")]; + tensor var_2162_promoted_to_fp16 = const()[name = tensor("op_2162_promoted_to_fp16"), val = tensor(0x1p+0)]; + tensor var_2163_cast_fp16 = add(x = var_2146_cast_fp16_1, y = var_2162_promoted_to_fp16)[name = tensor("op_2163_cast_fp16")]; + tensor var_2164_cast_fp16 = mul(x = x_235_cast_fp16, y = var_2163_cast_fp16)[name = tensor("op_2164_cast_fp16")]; + tensor input_365_cast_fp16 = add(x = var_2164_cast_fp16, y = var_2146_cast_fp16_0)[name = tensor("input_365_cast_fp16")]; + tensor linear_140_cast_fp16 = linear(bias = flow_net_res_blocks_1_mlp_0_bias_to_fp16, weight = flow_net_res_blocks_1_mlp_0_weight_to_fp16, x = input_365_cast_fp16)[name = tensor("linear_140_cast_fp16")]; + tensor input_369_cast_fp16 = silu(x = linear_140_cast_fp16)[name = tensor("input_369_cast_fp16")]; + tensor linear_141_cast_fp16 = linear(bias = flow_net_res_blocks_1_mlp_2_bias_to_fp16, weight = flow_net_res_blocks_1_mlp_2_weight_to_fp16, x = input_369_cast_fp16)[name = tensor("linear_141_cast_fp16")]; + tensor var_2175_cast_fp16 = mul(x = var_2146_cast_fp16_2, y = linear_141_cast_fp16)[name = tensor("op_2175_cast_fp16")]; + tensor x_237_cast_fp16 = add(x = x_231_cast_fp16, y = var_2175_cast_fp16)[name = tensor("x_237_cast_fp16")]; + tensor linear_142_cast_fp16 = linear(bias = flow_net_res_blocks_2_adaLN_modulation_1_bias_to_fp16, weight = flow_net_res_blocks_2_adaLN_modulation_1_weight_to_fp16, x = input_355_cast_fp16)[name = tensor("linear_142_cast_fp16")]; + tensor var_2185_split_sizes_0 = const()[name = tensor("op_2185_split_sizes_0"), val = tensor([512, 512, 512])]; + tensor var_2185_axis_0 = const()[name = tensor("op_2185_axis_0"), val = tensor(-1)]; + tensor var_2185_cast_fp16_0, tensor var_2185_cast_fp16_1, tensor var_2185_cast_fp16_2 = split(axis = var_2185_axis_0, split_sizes = var_2185_split_sizes_0, x = linear_142_cast_fp16)[name = tensor("op_2185_cast_fp16")]; + tensor mean_75_axes_0 = const()[name = tensor("mean_75_axes_0"), val = tensor([-1])]; + tensor mean_75_keep_dims_0 = const()[name = tensor("mean_75_keep_dims_0"), val = tensor(true)]; + tensor mean_75_cast_fp16 = reduce_mean(axes = mean_75_axes_0, keep_dims = mean_75_keep_dims_0, x = x_237_cast_fp16)[name = tensor("mean_75_cast_fp16")]; + tensor sub_61_cast_fp16 = sub(x = x_237_cast_fp16, y = mean_75_cast_fp16)[name = tensor("sub_61_cast_fp16")]; + tensor square_49_cast_fp16 = square(x = sub_61_cast_fp16)[name = tensor("square_49_cast_fp16")]; + tensor reduce_mean_99_axes_0 = const()[name = tensor("reduce_mean_99_axes_0"), val = tensor([-1])]; + tensor reduce_mean_99_keep_dims_0 = const()[name = tensor("reduce_mean_99_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_99_cast_fp16 = reduce_mean(axes = reduce_mean_99_axes_0, keep_dims = reduce_mean_99_keep_dims_0, x = square_49_cast_fp16)[name = tensor("reduce_mean_99_cast_fp16")]; + tensor var_2195_to_fp16 = const()[name = tensor("op_2195_to_fp16"), val = tensor(0x1.1p-20)]; + tensor var_2196_cast_fp16 = add(x = reduce_mean_99_cast_fp16, y = var_2195_to_fp16)[name = tensor("op_2196_cast_fp16")]; + tensor var_2197_cast_fp16 = sqrt(x = var_2196_cast_fp16)[name = tensor("op_2197_cast_fp16")]; + tensor x_239_cast_fp16 = real_div(x = sub_61_cast_fp16, y = var_2197_cast_fp16)[name = tensor("x_239_cast_fp16")]; + tensor var_2199_cast_fp16 = mul(x = x_239_cast_fp16, y = flow_net_res_blocks_2_in_ln_weight_to_fp16)[name = tensor("op_2199_cast_fp16")]; + tensor x_241_cast_fp16 = add(x = var_2199_cast_fp16, y = flow_net_res_blocks_2_in_ln_bias_to_fp16)[name = tensor("x_241_cast_fp16")]; + tensor var_2201_promoted_to_fp16 = const()[name = tensor("op_2201_promoted_to_fp16"), val = tensor(0x1p+0)]; + tensor var_2202_cast_fp16 = add(x = var_2185_cast_fp16_1, y = var_2201_promoted_to_fp16)[name = tensor("op_2202_cast_fp16")]; + tensor var_2203_cast_fp16 = mul(x = x_241_cast_fp16, y = var_2202_cast_fp16)[name = tensor("op_2203_cast_fp16")]; + tensor input_373_cast_fp16 = add(x = var_2203_cast_fp16, y = var_2185_cast_fp16_0)[name = tensor("input_373_cast_fp16")]; + tensor linear_143_cast_fp16 = linear(bias = flow_net_res_blocks_2_mlp_0_bias_to_fp16, weight = flow_net_res_blocks_2_mlp_0_weight_to_fp16, x = input_373_cast_fp16)[name = tensor("linear_143_cast_fp16")]; + tensor input_377_cast_fp16 = silu(x = linear_143_cast_fp16)[name = tensor("input_377_cast_fp16")]; + tensor linear_144_cast_fp16 = linear(bias = flow_net_res_blocks_2_mlp_2_bias_to_fp16, weight = flow_net_res_blocks_2_mlp_2_weight_to_fp16, x = input_377_cast_fp16)[name = tensor("linear_144_cast_fp16")]; + tensor var_2214_cast_fp16 = mul(x = var_2185_cast_fp16_2, y = linear_144_cast_fp16)[name = tensor("op_2214_cast_fp16")]; + tensor x_243_cast_fp16 = add(x = x_237_cast_fp16, y = var_2214_cast_fp16)[name = tensor("x_243_cast_fp16")]; + tensor linear_145_cast_fp16 = linear(bias = flow_net_res_blocks_3_adaLN_modulation_1_bias_to_fp16, weight = flow_net_res_blocks_3_adaLN_modulation_1_weight_to_fp16, x = input_355_cast_fp16)[name = tensor("linear_145_cast_fp16")]; + tensor var_2224_split_sizes_0 = const()[name = tensor("op_2224_split_sizes_0"), val = tensor([512, 512, 512])]; + tensor var_2224_axis_0 = const()[name = tensor("op_2224_axis_0"), val = tensor(-1)]; + tensor var_2224_cast_fp16_0, tensor var_2224_cast_fp16_1, tensor var_2224_cast_fp16_2 = split(axis = var_2224_axis_0, split_sizes = var_2224_split_sizes_0, x = linear_145_cast_fp16)[name = tensor("op_2224_cast_fp16")]; + tensor mean_77_axes_0 = const()[name = tensor("mean_77_axes_0"), val = tensor([-1])]; + tensor mean_77_keep_dims_0 = const()[name = tensor("mean_77_keep_dims_0"), val = tensor(true)]; + tensor mean_77_cast_fp16 = reduce_mean(axes = mean_77_axes_0, keep_dims = mean_77_keep_dims_0, x = x_243_cast_fp16)[name = tensor("mean_77_cast_fp16")]; + tensor sub_62_cast_fp16 = sub(x = x_243_cast_fp16, y = mean_77_cast_fp16)[name = tensor("sub_62_cast_fp16")]; + tensor square_50_cast_fp16 = square(x = sub_62_cast_fp16)[name = tensor("square_50_cast_fp16")]; + tensor reduce_mean_101_axes_0 = const()[name = tensor("reduce_mean_101_axes_0"), val = tensor([-1])]; + tensor reduce_mean_101_keep_dims_0 = const()[name = tensor("reduce_mean_101_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_101_cast_fp16 = reduce_mean(axes = reduce_mean_101_axes_0, keep_dims = reduce_mean_101_keep_dims_0, x = square_50_cast_fp16)[name = tensor("reduce_mean_101_cast_fp16")]; + tensor var_2234_to_fp16 = const()[name = tensor("op_2234_to_fp16"), val = tensor(0x1.1p-20)]; + tensor var_2235_cast_fp16 = add(x = reduce_mean_101_cast_fp16, y = var_2234_to_fp16)[name = tensor("op_2235_cast_fp16")]; + tensor var_2236_cast_fp16 = sqrt(x = var_2235_cast_fp16)[name = tensor("op_2236_cast_fp16")]; + tensor x_245_cast_fp16 = real_div(x = sub_62_cast_fp16, y = var_2236_cast_fp16)[name = tensor("x_245_cast_fp16")]; + tensor var_2238_cast_fp16 = mul(x = x_245_cast_fp16, y = flow_net_res_blocks_3_in_ln_weight_to_fp16)[name = tensor("op_2238_cast_fp16")]; + tensor x_247_cast_fp16 = add(x = var_2238_cast_fp16, y = flow_net_res_blocks_3_in_ln_bias_to_fp16)[name = tensor("x_247_cast_fp16")]; + tensor var_2240_promoted_to_fp16 = const()[name = tensor("op_2240_promoted_to_fp16"), val = tensor(0x1p+0)]; + tensor var_2241_cast_fp16 = add(x = var_2224_cast_fp16_1, y = var_2240_promoted_to_fp16)[name = tensor("op_2241_cast_fp16")]; + tensor var_2242_cast_fp16 = mul(x = x_247_cast_fp16, y = var_2241_cast_fp16)[name = tensor("op_2242_cast_fp16")]; + tensor input_381_cast_fp16 = add(x = var_2242_cast_fp16, y = var_2224_cast_fp16_0)[name = tensor("input_381_cast_fp16")]; + tensor linear_146_cast_fp16 = linear(bias = flow_net_res_blocks_3_mlp_0_bias_to_fp16, weight = flow_net_res_blocks_3_mlp_0_weight_to_fp16, x = input_381_cast_fp16)[name = tensor("linear_146_cast_fp16")]; + tensor input_385_cast_fp16 = silu(x = linear_146_cast_fp16)[name = tensor("input_385_cast_fp16")]; + tensor linear_147_cast_fp16 = linear(bias = flow_net_res_blocks_3_mlp_2_bias_to_fp16, weight = flow_net_res_blocks_3_mlp_2_weight_to_fp16, x = input_385_cast_fp16)[name = tensor("linear_147_cast_fp16")]; + tensor var_2253_cast_fp16 = mul(x = var_2224_cast_fp16_2, y = linear_147_cast_fp16)[name = tensor("op_2253_cast_fp16")]; + tensor x_249_cast_fp16 = add(x = x_243_cast_fp16, y = var_2253_cast_fp16)[name = tensor("x_249_cast_fp16")]; + tensor linear_148_cast_fp16 = linear(bias = flow_net_res_blocks_4_adaLN_modulation_1_bias_to_fp16, weight = flow_net_res_blocks_4_adaLN_modulation_1_weight_to_fp16, x = input_355_cast_fp16)[name = tensor("linear_148_cast_fp16")]; + tensor var_2263_split_sizes_0 = const()[name = tensor("op_2263_split_sizes_0"), val = tensor([512, 512, 512])]; + tensor var_2263_axis_0 = const()[name = tensor("op_2263_axis_0"), val = tensor(-1)]; + tensor var_2263_cast_fp16_0, tensor var_2263_cast_fp16_1, tensor var_2263_cast_fp16_2 = split(axis = var_2263_axis_0, split_sizes = var_2263_split_sizes_0, x = linear_148_cast_fp16)[name = tensor("op_2263_cast_fp16")]; + tensor mean_79_axes_0 = const()[name = tensor("mean_79_axes_0"), val = tensor([-1])]; + tensor mean_79_keep_dims_0 = const()[name = tensor("mean_79_keep_dims_0"), val = tensor(true)]; + tensor mean_79_cast_fp16 = reduce_mean(axes = mean_79_axes_0, keep_dims = mean_79_keep_dims_0, x = x_249_cast_fp16)[name = tensor("mean_79_cast_fp16")]; + tensor sub_63_cast_fp16 = sub(x = x_249_cast_fp16, y = mean_79_cast_fp16)[name = tensor("sub_63_cast_fp16")]; + tensor square_51_cast_fp16 = square(x = sub_63_cast_fp16)[name = tensor("square_51_cast_fp16")]; + tensor reduce_mean_103_axes_0 = const()[name = tensor("reduce_mean_103_axes_0"), val = tensor([-1])]; + tensor reduce_mean_103_keep_dims_0 = const()[name = tensor("reduce_mean_103_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_103_cast_fp16 = reduce_mean(axes = reduce_mean_103_axes_0, keep_dims = reduce_mean_103_keep_dims_0, x = square_51_cast_fp16)[name = tensor("reduce_mean_103_cast_fp16")]; + tensor var_2273_to_fp16 = const()[name = tensor("op_2273_to_fp16"), val = tensor(0x1.1p-20)]; + tensor var_2274_cast_fp16 = add(x = reduce_mean_103_cast_fp16, y = var_2273_to_fp16)[name = tensor("op_2274_cast_fp16")]; + tensor var_2275_cast_fp16 = sqrt(x = var_2274_cast_fp16)[name = tensor("op_2275_cast_fp16")]; + tensor x_251_cast_fp16 = real_div(x = sub_63_cast_fp16, y = var_2275_cast_fp16)[name = tensor("x_251_cast_fp16")]; + tensor var_2277_cast_fp16 = mul(x = x_251_cast_fp16, y = flow_net_res_blocks_4_in_ln_weight_to_fp16)[name = tensor("op_2277_cast_fp16")]; + tensor x_253_cast_fp16 = add(x = var_2277_cast_fp16, y = flow_net_res_blocks_4_in_ln_bias_to_fp16)[name = tensor("x_253_cast_fp16")]; + tensor var_2279_promoted_to_fp16 = const()[name = tensor("op_2279_promoted_to_fp16"), val = tensor(0x1p+0)]; + tensor var_2280_cast_fp16 = add(x = var_2263_cast_fp16_1, y = var_2279_promoted_to_fp16)[name = tensor("op_2280_cast_fp16")]; + tensor var_2281_cast_fp16 = mul(x = x_253_cast_fp16, y = var_2280_cast_fp16)[name = tensor("op_2281_cast_fp16")]; + tensor input_389_cast_fp16 = add(x = var_2281_cast_fp16, y = var_2263_cast_fp16_0)[name = tensor("input_389_cast_fp16")]; + tensor linear_149_cast_fp16 = linear(bias = flow_net_res_blocks_4_mlp_0_bias_to_fp16, weight = flow_net_res_blocks_4_mlp_0_weight_to_fp16, x = input_389_cast_fp16)[name = tensor("linear_149_cast_fp16")]; + tensor input_393_cast_fp16 = silu(x = linear_149_cast_fp16)[name = tensor("input_393_cast_fp16")]; + tensor linear_150_cast_fp16 = linear(bias = flow_net_res_blocks_4_mlp_2_bias_to_fp16, weight = flow_net_res_blocks_4_mlp_2_weight_to_fp16, x = input_393_cast_fp16)[name = tensor("linear_150_cast_fp16")]; + tensor var_2292_cast_fp16 = mul(x = var_2263_cast_fp16_2, y = linear_150_cast_fp16)[name = tensor("op_2292_cast_fp16")]; + tensor x_255_cast_fp16 = add(x = x_249_cast_fp16, y = var_2292_cast_fp16)[name = tensor("x_255_cast_fp16")]; + tensor linear_151_cast_fp16 = linear(bias = flow_net_res_blocks_5_adaLN_modulation_1_bias_to_fp16, weight = flow_net_res_blocks_5_adaLN_modulation_1_weight_to_fp16, x = input_355_cast_fp16)[name = tensor("linear_151_cast_fp16")]; + tensor var_2302_split_sizes_0 = const()[name = tensor("op_2302_split_sizes_0"), val = tensor([512, 512, 512])]; + tensor var_2302_axis_0 = const()[name = tensor("op_2302_axis_0"), val = tensor(-1)]; + tensor var_2302_cast_fp16_0, tensor var_2302_cast_fp16_1, tensor var_2302_cast_fp16_2 = split(axis = var_2302_axis_0, split_sizes = var_2302_split_sizes_0, x = linear_151_cast_fp16)[name = tensor("op_2302_cast_fp16")]; + tensor mean_81_axes_0 = const()[name = tensor("mean_81_axes_0"), val = tensor([-1])]; + tensor mean_81_keep_dims_0 = const()[name = tensor("mean_81_keep_dims_0"), val = tensor(true)]; + tensor mean_81_cast_fp16 = reduce_mean(axes = mean_81_axes_0, keep_dims = mean_81_keep_dims_0, x = x_255_cast_fp16)[name = tensor("mean_81_cast_fp16")]; + tensor sub_64_cast_fp16 = sub(x = x_255_cast_fp16, y = mean_81_cast_fp16)[name = tensor("sub_64_cast_fp16")]; + tensor square_52_cast_fp16 = square(x = sub_64_cast_fp16)[name = tensor("square_52_cast_fp16")]; + tensor reduce_mean_105_axes_0 = const()[name = tensor("reduce_mean_105_axes_0"), val = tensor([-1])]; + tensor reduce_mean_105_keep_dims_0 = const()[name = tensor("reduce_mean_105_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_105_cast_fp16 = reduce_mean(axes = reduce_mean_105_axes_0, keep_dims = reduce_mean_105_keep_dims_0, x = square_52_cast_fp16)[name = tensor("reduce_mean_105_cast_fp16")]; + tensor var_2312_to_fp16 = const()[name = tensor("op_2312_to_fp16"), val = tensor(0x1.1p-20)]; + tensor var_2313_cast_fp16 = add(x = reduce_mean_105_cast_fp16, y = var_2312_to_fp16)[name = tensor("op_2313_cast_fp16")]; + tensor var_2314_cast_fp16 = sqrt(x = var_2313_cast_fp16)[name = tensor("op_2314_cast_fp16")]; + tensor x_257_cast_fp16 = real_div(x = sub_64_cast_fp16, y = var_2314_cast_fp16)[name = tensor("x_257_cast_fp16")]; + tensor var_2316_cast_fp16 = mul(x = x_257_cast_fp16, y = flow_net_res_blocks_5_in_ln_weight_to_fp16)[name = tensor("op_2316_cast_fp16")]; + tensor x_259_cast_fp16 = add(x = var_2316_cast_fp16, y = flow_net_res_blocks_5_in_ln_bias_to_fp16)[name = tensor("x_259_cast_fp16")]; + tensor var_2318_promoted_to_fp16 = const()[name = tensor("op_2318_promoted_to_fp16"), val = tensor(0x1p+0)]; + tensor var_2319_cast_fp16 = add(x = var_2302_cast_fp16_1, y = var_2318_promoted_to_fp16)[name = tensor("op_2319_cast_fp16")]; + tensor var_2320_cast_fp16 = mul(x = x_259_cast_fp16, y = var_2319_cast_fp16)[name = tensor("op_2320_cast_fp16")]; + tensor input_397_cast_fp16 = add(x = var_2320_cast_fp16, y = var_2302_cast_fp16_0)[name = tensor("input_397_cast_fp16")]; + tensor linear_152_cast_fp16 = linear(bias = flow_net_res_blocks_5_mlp_0_bias_to_fp16, weight = flow_net_res_blocks_5_mlp_0_weight_to_fp16, x = input_397_cast_fp16)[name = tensor("linear_152_cast_fp16")]; + tensor input_401_cast_fp16 = silu(x = linear_152_cast_fp16)[name = tensor("input_401_cast_fp16")]; + tensor linear_153_cast_fp16 = linear(bias = flow_net_res_blocks_5_mlp_2_bias_to_fp16, weight = flow_net_res_blocks_5_mlp_2_weight_to_fp16, x = input_401_cast_fp16)[name = tensor("linear_153_cast_fp16")]; + tensor var_2331_cast_fp16 = mul(x = var_2302_cast_fp16_2, y = linear_153_cast_fp16)[name = tensor("op_2331_cast_fp16")]; + tensor x_261_cast_fp16 = add(x = x_255_cast_fp16, y = var_2331_cast_fp16)[name = tensor("x_261_cast_fp16")]; + tensor linear_154_cast_fp16 = linear(bias = flow_net_final_layer_adaLN_modulation_1_bias_to_fp16, weight = flow_net_final_layer_adaLN_modulation_1_weight_to_fp16, x = input_355_cast_fp16)[name = tensor("linear_154_cast_fp16")]; + tensor var_2340_split_sizes_0 = const()[name = tensor("op_2340_split_sizes_0"), val = tensor([512, 512])]; + tensor var_2340_axis_0 = const()[name = tensor("op_2340_axis_0"), val = tensor(-1)]; + tensor var_2340_cast_fp16_0, tensor var_2340_cast_fp16_1 = split(axis = var_2340_axis_0, split_sizes = var_2340_split_sizes_0, x = linear_154_cast_fp16)[name = tensor("op_2340_cast_fp16")]; + tensor mean_83_axes_0 = const()[name = tensor("mean_83_axes_0"), val = tensor([-1])]; + tensor mean_83_keep_dims_0 = const()[name = tensor("mean_83_keep_dims_0"), val = tensor(true)]; + tensor mean_83_cast_fp16 = reduce_mean(axes = mean_83_axes_0, keep_dims = mean_83_keep_dims_0, x = x_261_cast_fp16)[name = tensor("mean_83_cast_fp16")]; + tensor sub_65_cast_fp16 = sub(x = x_261_cast_fp16, y = mean_83_cast_fp16)[name = tensor("sub_65_cast_fp16")]; + tensor square_53_cast_fp16 = square(x = sub_65_cast_fp16)[name = tensor("square_53_cast_fp16")]; + tensor reduce_mean_107_axes_0 = const()[name = tensor("reduce_mean_107_axes_0"), val = tensor([-1])]; + tensor reduce_mean_107_keep_dims_0 = const()[name = tensor("reduce_mean_107_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_107_cast_fp16 = reduce_mean(axes = reduce_mean_107_axes_0, keep_dims = reduce_mean_107_keep_dims_0, x = square_53_cast_fp16)[name = tensor("reduce_mean_107_cast_fp16")]; + tensor var_2347_to_fp16 = const()[name = tensor("op_2347_to_fp16"), val = tensor(0x1.1p-20)]; + tensor var_2348_cast_fp16 = add(x = reduce_mean_107_cast_fp16, y = var_2347_to_fp16)[name = tensor("op_2348_cast_fp16")]; + tensor var_2349_cast_fp16 = sqrt(x = var_2348_cast_fp16)[name = tensor("op_2349_cast_fp16")]; + tensor x_263_cast_fp16 = real_div(x = sub_65_cast_fp16, y = var_2349_cast_fp16)[name = tensor("x_263_cast_fp16")]; + tensor var_2351_promoted_to_fp16 = const()[name = tensor("op_2351_promoted_to_fp16"), val = tensor(0x1p+0)]; + tensor var_2352_cast_fp16 = add(x = var_2340_cast_fp16_1, y = var_2351_promoted_to_fp16)[name = tensor("op_2352_cast_fp16")]; + tensor var_2353_cast_fp16 = mul(x = x_263_cast_fp16, y = var_2352_cast_fp16)[name = tensor("op_2353_cast_fp16")]; + tensor input_405_cast_fp16 = add(x = var_2353_cast_fp16, y = var_2340_cast_fp16_0)[name = tensor("input_405_cast_fp16")]; + tensor linear_155_cast_fp16 = linear(bias = flow_net_final_layer_linear_bias_to_fp16, weight = flow_net_final_layer_linear_weight_to_fp16, x = input_405_cast_fp16)[name = tensor("linear_155_cast_fp16")]; + tensor var_2364_to_fp16 = const()[name = tensor("op_2364_to_fp16"), val = tensor(0x1p-3)]; + tensor var_2365_cast_fp16 = mul(x = linear_155_cast_fp16, y = var_2364_to_fp16)[name = tensor("op_2365_cast_fp16")]; + tensor input_407_cast_fp16 = add(x = input_339_cast_fp16, y = var_2365_cast_fp16)[name = tensor("input_407_cast_fp16")]; + tensor linear_156_cast_fp16 = linear(bias = flow_net_input_proj_bias_to_fp16, weight = flow_net_input_proj_weight_to_fp16, x = input_407_cast_fp16)[name = tensor("linear_156_cast_fp16")]; + tensor input_411_to_fp16 = const()[name = tensor("input_411_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(19008256)))]; + tensor input_413_cast_fp16 = silu(x = input_411_to_fp16)[name = tensor("input_413_cast_fp16")]; + tensor linear_158_cast_fp16 = linear(bias = flow_net_time_embed_0_mlp_2_bias_to_fp16, weight = flow_net_time_embed_0_mlp_2_weight_to_fp16, x = input_413_cast_fp16)[name = tensor("linear_158_cast_fp16")]; + tensor reduce_mean_108_axes_0 = const()[name = tensor("reduce_mean_108_axes_0"), val = tensor([-1])]; + tensor reduce_mean_108_keep_dims_0 = const()[name = tensor("reduce_mean_108_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_108_cast_fp16 = reduce_mean(axes = reduce_mean_108_axes_0, keep_dims = reduce_mean_108_keep_dims_0, x = linear_158_cast_fp16)[name = tensor("reduce_mean_108_cast_fp16")]; + tensor sub_66_cast_fp16 = sub(x = linear_158_cast_fp16, y = reduce_mean_108_cast_fp16)[name = tensor("sub_66_cast_fp16")]; + tensor square_54_cast_fp16 = square(x = sub_66_cast_fp16)[name = tensor("square_54_cast_fp16")]; + tensor reduce_mean_109_axes_0 = const()[name = tensor("reduce_mean_109_axes_0"), val = tensor([-1])]; + tensor reduce_mean_109_keep_dims_0 = const()[name = tensor("reduce_mean_109_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_109_cast_fp16 = reduce_mean(axes = reduce_mean_109_axes_0, keep_dims = reduce_mean_109_keep_dims_0, x = square_54_cast_fp16)[name = tensor("reduce_mean_109_cast_fp16")]; + tensor real_div_12_to_fp16 = const()[name = tensor("real_div_12_to_fp16"), val = tensor(0x1.008p+0)]; + tensor mul_12_cast_fp16 = mul(x = reduce_mean_109_cast_fp16, y = real_div_12_to_fp16)[name = tensor("mul_12_cast_fp16")]; + tensor var_2431_to_fp16 = const()[name = tensor("op_2431_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_109_cast_fp16 = add(x = mul_12_cast_fp16, y = var_2431_to_fp16)[name = tensor("var_109_cast_fp16")]; + tensor var_2434_epsilon_0 = const()[name = tensor("op_2434_epsilon_0"), val = tensor(0x1.197998p-40)]; + tensor var_2434_cast_fp16 = rsqrt(epsilon = var_2434_epsilon_0, x = var_109_cast_fp16)[name = tensor("op_2434_cast_fp16")]; + tensor var_2435_cast_fp16 = mul(x = const_3_to_fp16, y = var_2434_cast_fp16)[name = tensor("op_2435_cast_fp16")]; + tensor var_2436_cast_fp16 = mul(x = linear_158_cast_fp16, y = var_2435_cast_fp16)[name = tensor("op_2436_cast_fp16")]; + tensor input_417_to_fp16 = const()[name = tensor("input_417_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(19009344)))]; + tensor input_419_cast_fp16 = silu(x = input_417_to_fp16)[name = tensor("input_419_cast_fp16")]; + tensor linear_160_cast_fp16 = linear(bias = flow_net_time_embed_1_mlp_2_bias_to_fp16, weight = flow_net_time_embed_1_mlp_2_weight_to_fp16, x = input_419_cast_fp16)[name = tensor("linear_160_cast_fp16")]; + tensor reduce_mean_110_axes_0 = const()[name = tensor("reduce_mean_110_axes_0"), val = tensor([-1])]; + tensor reduce_mean_110_keep_dims_0 = const()[name = tensor("reduce_mean_110_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_110_cast_fp16 = reduce_mean(axes = reduce_mean_110_axes_0, keep_dims = reduce_mean_110_keep_dims_0, x = linear_160_cast_fp16)[name = tensor("reduce_mean_110_cast_fp16")]; + tensor sub_68_cast_fp16 = sub(x = linear_160_cast_fp16, y = reduce_mean_110_cast_fp16)[name = tensor("sub_68_cast_fp16")]; + tensor square_55_cast_fp16 = square(x = sub_68_cast_fp16)[name = tensor("square_55_cast_fp16")]; + tensor reduce_mean_111_axes_0 = const()[name = tensor("reduce_mean_111_axes_0"), val = tensor([-1])]; + tensor reduce_mean_111_keep_dims_0 = const()[name = tensor("reduce_mean_111_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_111_cast_fp16 = reduce_mean(axes = reduce_mean_111_axes_0, keep_dims = reduce_mean_111_keep_dims_0, x = square_55_cast_fp16)[name = tensor("reduce_mean_111_cast_fp16")]; + tensor real_div_13_to_fp16 = const()[name = tensor("real_div_13_to_fp16"), val = tensor(0x1.008p+0)]; + tensor mul_13_cast_fp16 = mul(x = reduce_mean_111_cast_fp16, y = real_div_13_to_fp16)[name = tensor("mul_13_cast_fp16")]; + tensor var_2468_to_fp16 = const()[name = tensor("op_2468_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_111_cast_fp16 = add(x = mul_13_cast_fp16, y = var_2468_to_fp16)[name = tensor("var_111_cast_fp16")]; + tensor var_2471_epsilon_0 = const()[name = tensor("op_2471_epsilon_0"), val = tensor(0x1.197998p-40)]; + tensor var_2471_cast_fp16 = rsqrt(epsilon = var_2471_epsilon_0, x = var_111_cast_fp16)[name = tensor("op_2471_cast_fp16")]; + tensor var_2472_cast_fp16 = mul(x = const_5_to_fp16, y = var_2471_cast_fp16)[name = tensor("op_2472_cast_fp16")]; + tensor var_2473_cast_fp16 = mul(x = linear_160_cast_fp16, y = var_2472_cast_fp16)[name = tensor("op_2473_cast_fp16")]; + tensor var_2485_cast_fp16 = add(x = var_2436_cast_fp16, y = var_2473_cast_fp16)[name = tensor("op_2485_cast_fp16")]; + tensor _inversed_t_combined_13_y_0_to_fp16 = const()[name = tensor("_inversed_t_combined_13_y_0_to_fp16"), val = tensor(0x1p-1)]; + tensor _inversed_t_combined_13_cast_fp16 = mul(x = var_2485_cast_fp16, y = _inversed_t_combined_13_y_0_to_fp16)[name = tensor("_inversed_t_combined_13_cast_fp16")]; + tensor input_421_cast_fp16 = add(x = _inversed_t_combined_13_cast_fp16, y = linear_5_cast_fp16)[name = tensor("input_421_cast_fp16")]; + tensor input_423_cast_fp16 = silu(x = input_421_cast_fp16)[name = tensor("input_423_cast_fp16")]; + tensor linear_162_cast_fp16 = linear(bias = flow_net_res_blocks_0_adaLN_modulation_1_bias_to_fp16, weight = flow_net_res_blocks_0_adaLN_modulation_1_weight_to_fp16, x = input_423_cast_fp16)[name = tensor("linear_162_cast_fp16")]; + tensor var_2500_split_sizes_0 = const()[name = tensor("op_2500_split_sizes_0"), val = tensor([512, 512, 512])]; + tensor var_2500_axis_0 = const()[name = tensor("op_2500_axis_0"), val = tensor(-1)]; + tensor var_2500_cast_fp16_0, tensor var_2500_cast_fp16_1, tensor var_2500_cast_fp16_2 = split(axis = var_2500_axis_0, split_sizes = var_2500_split_sizes_0, x = linear_162_cast_fp16)[name = tensor("op_2500_cast_fp16")]; + tensor mean_85_axes_0 = const()[name = tensor("mean_85_axes_0"), val = tensor([-1])]; + tensor mean_85_keep_dims_0 = const()[name = tensor("mean_85_keep_dims_0"), val = tensor(true)]; + tensor mean_85_cast_fp16 = reduce_mean(axes = mean_85_axes_0, keep_dims = mean_85_keep_dims_0, x = linear_156_cast_fp16)[name = tensor("mean_85_cast_fp16")]; + tensor sub_70_cast_fp16 = sub(x = linear_156_cast_fp16, y = mean_85_cast_fp16)[name = tensor("sub_70_cast_fp16")]; + tensor square_56_cast_fp16 = square(x = sub_70_cast_fp16)[name = tensor("square_56_cast_fp16")]; + tensor reduce_mean_113_axes_0 = const()[name = tensor("reduce_mean_113_axes_0"), val = tensor([-1])]; + tensor reduce_mean_113_keep_dims_0 = const()[name = tensor("reduce_mean_113_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_113_cast_fp16 = reduce_mean(axes = reduce_mean_113_axes_0, keep_dims = reduce_mean_113_keep_dims_0, x = square_56_cast_fp16)[name = tensor("reduce_mean_113_cast_fp16")]; + tensor var_2510_to_fp16 = const()[name = tensor("op_2510_to_fp16"), val = tensor(0x1.1p-20)]; + tensor var_2511_cast_fp16 = add(x = reduce_mean_113_cast_fp16, y = var_2510_to_fp16)[name = tensor("op_2511_cast_fp16")]; + tensor var_2512_cast_fp16 = sqrt(x = var_2511_cast_fp16)[name = tensor("op_2512_cast_fp16")]; + tensor x_271_cast_fp16 = real_div(x = sub_70_cast_fp16, y = var_2512_cast_fp16)[name = tensor("x_271_cast_fp16")]; + tensor var_2514_cast_fp16 = mul(x = x_271_cast_fp16, y = flow_net_res_blocks_0_in_ln_weight_to_fp16)[name = tensor("op_2514_cast_fp16")]; + tensor x_273_cast_fp16 = add(x = var_2514_cast_fp16, y = flow_net_res_blocks_0_in_ln_bias_to_fp16)[name = tensor("x_273_cast_fp16")]; + tensor var_2516_promoted_to_fp16 = const()[name = tensor("op_2516_promoted_to_fp16"), val = tensor(0x1p+0)]; + tensor var_2517_cast_fp16 = add(x = var_2500_cast_fp16_1, y = var_2516_promoted_to_fp16)[name = tensor("op_2517_cast_fp16")]; + tensor var_2518_cast_fp16 = mul(x = x_273_cast_fp16, y = var_2517_cast_fp16)[name = tensor("op_2518_cast_fp16")]; + tensor input_425_cast_fp16 = add(x = var_2518_cast_fp16, y = var_2500_cast_fp16_0)[name = tensor("input_425_cast_fp16")]; + tensor linear_163_cast_fp16 = linear(bias = flow_net_res_blocks_0_mlp_0_bias_to_fp16, weight = flow_net_res_blocks_0_mlp_0_weight_to_fp16, x = input_425_cast_fp16)[name = tensor("linear_163_cast_fp16")]; + tensor input_429_cast_fp16 = silu(x = linear_163_cast_fp16)[name = tensor("input_429_cast_fp16")]; + tensor linear_164_cast_fp16 = linear(bias = flow_net_res_blocks_0_mlp_2_bias_to_fp16, weight = flow_net_res_blocks_0_mlp_2_weight_to_fp16, x = input_429_cast_fp16)[name = tensor("linear_164_cast_fp16")]; + tensor var_2529_cast_fp16 = mul(x = var_2500_cast_fp16_2, y = linear_164_cast_fp16)[name = tensor("op_2529_cast_fp16")]; + tensor x_275_cast_fp16 = add(x = linear_156_cast_fp16, y = var_2529_cast_fp16)[name = tensor("x_275_cast_fp16")]; + tensor linear_165_cast_fp16 = linear(bias = flow_net_res_blocks_1_adaLN_modulation_1_bias_to_fp16, weight = flow_net_res_blocks_1_adaLN_modulation_1_weight_to_fp16, x = input_423_cast_fp16)[name = tensor("linear_165_cast_fp16")]; + tensor var_2539_split_sizes_0 = const()[name = tensor("op_2539_split_sizes_0"), val = tensor([512, 512, 512])]; + tensor var_2539_axis_0 = const()[name = tensor("op_2539_axis_0"), val = tensor(-1)]; + tensor var_2539_cast_fp16_0, tensor var_2539_cast_fp16_1, tensor var_2539_cast_fp16_2 = split(axis = var_2539_axis_0, split_sizes = var_2539_split_sizes_0, x = linear_165_cast_fp16)[name = tensor("op_2539_cast_fp16")]; + tensor mean_87_axes_0 = const()[name = tensor("mean_87_axes_0"), val = tensor([-1])]; + tensor mean_87_keep_dims_0 = const()[name = tensor("mean_87_keep_dims_0"), val = tensor(true)]; + tensor mean_87_cast_fp16 = reduce_mean(axes = mean_87_axes_0, keep_dims = mean_87_keep_dims_0, x = x_275_cast_fp16)[name = tensor("mean_87_cast_fp16")]; + tensor sub_71_cast_fp16 = sub(x = x_275_cast_fp16, y = mean_87_cast_fp16)[name = tensor("sub_71_cast_fp16")]; + tensor square_57_cast_fp16 = square(x = sub_71_cast_fp16)[name = tensor("square_57_cast_fp16")]; + tensor reduce_mean_115_axes_0 = const()[name = tensor("reduce_mean_115_axes_0"), val = tensor([-1])]; + tensor reduce_mean_115_keep_dims_0 = const()[name = tensor("reduce_mean_115_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_115_cast_fp16 = reduce_mean(axes = reduce_mean_115_axes_0, keep_dims = reduce_mean_115_keep_dims_0, x = square_57_cast_fp16)[name = tensor("reduce_mean_115_cast_fp16")]; + tensor var_2549_to_fp16 = const()[name = tensor("op_2549_to_fp16"), val = tensor(0x1.1p-20)]; + tensor var_2550_cast_fp16 = add(x = reduce_mean_115_cast_fp16, y = var_2549_to_fp16)[name = tensor("op_2550_cast_fp16")]; + tensor var_2551_cast_fp16 = sqrt(x = var_2550_cast_fp16)[name = tensor("op_2551_cast_fp16")]; + tensor x_277_cast_fp16 = real_div(x = sub_71_cast_fp16, y = var_2551_cast_fp16)[name = tensor("x_277_cast_fp16")]; + tensor var_2553_cast_fp16 = mul(x = x_277_cast_fp16, y = flow_net_res_blocks_1_in_ln_weight_to_fp16)[name = tensor("op_2553_cast_fp16")]; + tensor x_279_cast_fp16 = add(x = var_2553_cast_fp16, y = flow_net_res_blocks_1_in_ln_bias_to_fp16)[name = tensor("x_279_cast_fp16")]; + tensor var_2555_promoted_to_fp16 = const()[name = tensor("op_2555_promoted_to_fp16"), val = tensor(0x1p+0)]; + tensor var_2556_cast_fp16 = add(x = var_2539_cast_fp16_1, y = var_2555_promoted_to_fp16)[name = tensor("op_2556_cast_fp16")]; + tensor var_2557_cast_fp16 = mul(x = x_279_cast_fp16, y = var_2556_cast_fp16)[name = tensor("op_2557_cast_fp16")]; + tensor input_433_cast_fp16 = add(x = var_2557_cast_fp16, y = var_2539_cast_fp16_0)[name = tensor("input_433_cast_fp16")]; + tensor linear_166_cast_fp16 = linear(bias = flow_net_res_blocks_1_mlp_0_bias_to_fp16, weight = flow_net_res_blocks_1_mlp_0_weight_to_fp16, x = input_433_cast_fp16)[name = tensor("linear_166_cast_fp16")]; + tensor input_437_cast_fp16 = silu(x = linear_166_cast_fp16)[name = tensor("input_437_cast_fp16")]; + tensor linear_167_cast_fp16 = linear(bias = flow_net_res_blocks_1_mlp_2_bias_to_fp16, weight = flow_net_res_blocks_1_mlp_2_weight_to_fp16, x = input_437_cast_fp16)[name = tensor("linear_167_cast_fp16")]; + tensor var_2568_cast_fp16 = mul(x = var_2539_cast_fp16_2, y = linear_167_cast_fp16)[name = tensor("op_2568_cast_fp16")]; + tensor x_281_cast_fp16 = add(x = x_275_cast_fp16, y = var_2568_cast_fp16)[name = tensor("x_281_cast_fp16")]; + tensor linear_168_cast_fp16 = linear(bias = flow_net_res_blocks_2_adaLN_modulation_1_bias_to_fp16, weight = flow_net_res_blocks_2_adaLN_modulation_1_weight_to_fp16, x = input_423_cast_fp16)[name = tensor("linear_168_cast_fp16")]; + tensor var_2578_split_sizes_0 = const()[name = tensor("op_2578_split_sizes_0"), val = tensor([512, 512, 512])]; + tensor var_2578_axis_0 = const()[name = tensor("op_2578_axis_0"), val = tensor(-1)]; + tensor var_2578_cast_fp16_0, tensor var_2578_cast_fp16_1, tensor var_2578_cast_fp16_2 = split(axis = var_2578_axis_0, split_sizes = var_2578_split_sizes_0, x = linear_168_cast_fp16)[name = tensor("op_2578_cast_fp16")]; + tensor mean_89_axes_0 = const()[name = tensor("mean_89_axes_0"), val = tensor([-1])]; + tensor mean_89_keep_dims_0 = const()[name = tensor("mean_89_keep_dims_0"), val = tensor(true)]; + tensor mean_89_cast_fp16 = reduce_mean(axes = mean_89_axes_0, keep_dims = mean_89_keep_dims_0, x = x_281_cast_fp16)[name = tensor("mean_89_cast_fp16")]; + tensor sub_72_cast_fp16 = sub(x = x_281_cast_fp16, y = mean_89_cast_fp16)[name = tensor("sub_72_cast_fp16")]; + tensor square_58_cast_fp16 = square(x = sub_72_cast_fp16)[name = tensor("square_58_cast_fp16")]; + tensor reduce_mean_117_axes_0 = const()[name = tensor("reduce_mean_117_axes_0"), val = tensor([-1])]; + tensor reduce_mean_117_keep_dims_0 = const()[name = tensor("reduce_mean_117_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_117_cast_fp16 = reduce_mean(axes = reduce_mean_117_axes_0, keep_dims = reduce_mean_117_keep_dims_0, x = square_58_cast_fp16)[name = tensor("reduce_mean_117_cast_fp16")]; + tensor var_2588_to_fp16 = const()[name = tensor("op_2588_to_fp16"), val = tensor(0x1.1p-20)]; + tensor var_2589_cast_fp16 = add(x = reduce_mean_117_cast_fp16, y = var_2588_to_fp16)[name = tensor("op_2589_cast_fp16")]; + tensor var_2590_cast_fp16 = sqrt(x = var_2589_cast_fp16)[name = tensor("op_2590_cast_fp16")]; + tensor x_283_cast_fp16 = real_div(x = sub_72_cast_fp16, y = var_2590_cast_fp16)[name = tensor("x_283_cast_fp16")]; + tensor var_2592_cast_fp16 = mul(x = x_283_cast_fp16, y = flow_net_res_blocks_2_in_ln_weight_to_fp16)[name = tensor("op_2592_cast_fp16")]; + tensor x_285_cast_fp16 = add(x = var_2592_cast_fp16, y = flow_net_res_blocks_2_in_ln_bias_to_fp16)[name = tensor("x_285_cast_fp16")]; + tensor var_2594_promoted_to_fp16 = const()[name = tensor("op_2594_promoted_to_fp16"), val = tensor(0x1p+0)]; + tensor var_2595_cast_fp16 = add(x = var_2578_cast_fp16_1, y = var_2594_promoted_to_fp16)[name = tensor("op_2595_cast_fp16")]; + tensor var_2596_cast_fp16 = mul(x = x_285_cast_fp16, y = var_2595_cast_fp16)[name = tensor("op_2596_cast_fp16")]; + tensor input_441_cast_fp16 = add(x = var_2596_cast_fp16, y = var_2578_cast_fp16_0)[name = tensor("input_441_cast_fp16")]; + tensor linear_169_cast_fp16 = linear(bias = flow_net_res_blocks_2_mlp_0_bias_to_fp16, weight = flow_net_res_blocks_2_mlp_0_weight_to_fp16, x = input_441_cast_fp16)[name = tensor("linear_169_cast_fp16")]; + tensor input_445_cast_fp16 = silu(x = linear_169_cast_fp16)[name = tensor("input_445_cast_fp16")]; + tensor linear_170_cast_fp16 = linear(bias = flow_net_res_blocks_2_mlp_2_bias_to_fp16, weight = flow_net_res_blocks_2_mlp_2_weight_to_fp16, x = input_445_cast_fp16)[name = tensor("linear_170_cast_fp16")]; + tensor var_2607_cast_fp16 = mul(x = var_2578_cast_fp16_2, y = linear_170_cast_fp16)[name = tensor("op_2607_cast_fp16")]; + tensor x_287_cast_fp16 = add(x = x_281_cast_fp16, y = var_2607_cast_fp16)[name = tensor("x_287_cast_fp16")]; + tensor linear_171_cast_fp16 = linear(bias = flow_net_res_blocks_3_adaLN_modulation_1_bias_to_fp16, weight = flow_net_res_blocks_3_adaLN_modulation_1_weight_to_fp16, x = input_423_cast_fp16)[name = tensor("linear_171_cast_fp16")]; + tensor var_2617_split_sizes_0 = const()[name = tensor("op_2617_split_sizes_0"), val = tensor([512, 512, 512])]; + tensor var_2617_axis_0 = const()[name = tensor("op_2617_axis_0"), val = tensor(-1)]; + tensor var_2617_cast_fp16_0, tensor var_2617_cast_fp16_1, tensor var_2617_cast_fp16_2 = split(axis = var_2617_axis_0, split_sizes = var_2617_split_sizes_0, x = linear_171_cast_fp16)[name = tensor("op_2617_cast_fp16")]; + tensor mean_91_axes_0 = const()[name = tensor("mean_91_axes_0"), val = tensor([-1])]; + tensor mean_91_keep_dims_0 = const()[name = tensor("mean_91_keep_dims_0"), val = tensor(true)]; + tensor mean_91_cast_fp16 = reduce_mean(axes = mean_91_axes_0, keep_dims = mean_91_keep_dims_0, x = x_287_cast_fp16)[name = tensor("mean_91_cast_fp16")]; + tensor sub_73_cast_fp16 = sub(x = x_287_cast_fp16, y = mean_91_cast_fp16)[name = tensor("sub_73_cast_fp16")]; + tensor square_59_cast_fp16 = square(x = sub_73_cast_fp16)[name = tensor("square_59_cast_fp16")]; + tensor reduce_mean_119_axes_0 = const()[name = tensor("reduce_mean_119_axes_0"), val = tensor([-1])]; + tensor reduce_mean_119_keep_dims_0 = const()[name = tensor("reduce_mean_119_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_119_cast_fp16 = reduce_mean(axes = reduce_mean_119_axes_0, keep_dims = reduce_mean_119_keep_dims_0, x = square_59_cast_fp16)[name = tensor("reduce_mean_119_cast_fp16")]; + tensor var_2627_to_fp16 = const()[name = tensor("op_2627_to_fp16"), val = tensor(0x1.1p-20)]; + tensor var_2628_cast_fp16 = add(x = reduce_mean_119_cast_fp16, y = var_2627_to_fp16)[name = tensor("op_2628_cast_fp16")]; + tensor var_2629_cast_fp16 = sqrt(x = var_2628_cast_fp16)[name = tensor("op_2629_cast_fp16")]; + tensor x_289_cast_fp16 = real_div(x = sub_73_cast_fp16, y = var_2629_cast_fp16)[name = tensor("x_289_cast_fp16")]; + tensor var_2631_cast_fp16 = mul(x = x_289_cast_fp16, y = flow_net_res_blocks_3_in_ln_weight_to_fp16)[name = tensor("op_2631_cast_fp16")]; + tensor x_291_cast_fp16 = add(x = var_2631_cast_fp16, y = flow_net_res_blocks_3_in_ln_bias_to_fp16)[name = tensor("x_291_cast_fp16")]; + tensor var_2633_promoted_to_fp16 = const()[name = tensor("op_2633_promoted_to_fp16"), val = tensor(0x1p+0)]; + tensor var_2634_cast_fp16 = add(x = var_2617_cast_fp16_1, y = var_2633_promoted_to_fp16)[name = tensor("op_2634_cast_fp16")]; + tensor var_2635_cast_fp16 = mul(x = x_291_cast_fp16, y = var_2634_cast_fp16)[name = tensor("op_2635_cast_fp16")]; + tensor input_449_cast_fp16 = add(x = var_2635_cast_fp16, y = var_2617_cast_fp16_0)[name = tensor("input_449_cast_fp16")]; + tensor linear_172_cast_fp16 = linear(bias = flow_net_res_blocks_3_mlp_0_bias_to_fp16, weight = flow_net_res_blocks_3_mlp_0_weight_to_fp16, x = input_449_cast_fp16)[name = tensor("linear_172_cast_fp16")]; + tensor input_453_cast_fp16 = silu(x = linear_172_cast_fp16)[name = tensor("input_453_cast_fp16")]; + tensor linear_173_cast_fp16 = linear(bias = flow_net_res_blocks_3_mlp_2_bias_to_fp16, weight = flow_net_res_blocks_3_mlp_2_weight_to_fp16, x = input_453_cast_fp16)[name = tensor("linear_173_cast_fp16")]; + tensor var_2646_cast_fp16 = mul(x = var_2617_cast_fp16_2, y = linear_173_cast_fp16)[name = tensor("op_2646_cast_fp16")]; + tensor x_293_cast_fp16 = add(x = x_287_cast_fp16, y = var_2646_cast_fp16)[name = tensor("x_293_cast_fp16")]; + tensor linear_174_cast_fp16 = linear(bias = flow_net_res_blocks_4_adaLN_modulation_1_bias_to_fp16, weight = flow_net_res_blocks_4_adaLN_modulation_1_weight_to_fp16, x = input_423_cast_fp16)[name = tensor("linear_174_cast_fp16")]; + tensor var_2656_split_sizes_0 = const()[name = tensor("op_2656_split_sizes_0"), val = tensor([512, 512, 512])]; + tensor var_2656_axis_0 = const()[name = tensor("op_2656_axis_0"), val = tensor(-1)]; + tensor var_2656_cast_fp16_0, tensor var_2656_cast_fp16_1, tensor var_2656_cast_fp16_2 = split(axis = var_2656_axis_0, split_sizes = var_2656_split_sizes_0, x = linear_174_cast_fp16)[name = tensor("op_2656_cast_fp16")]; + tensor mean_93_axes_0 = const()[name = tensor("mean_93_axes_0"), val = tensor([-1])]; + tensor mean_93_keep_dims_0 = const()[name = tensor("mean_93_keep_dims_0"), val = tensor(true)]; + tensor mean_93_cast_fp16 = reduce_mean(axes = mean_93_axes_0, keep_dims = mean_93_keep_dims_0, x = x_293_cast_fp16)[name = tensor("mean_93_cast_fp16")]; + tensor sub_74_cast_fp16 = sub(x = x_293_cast_fp16, y = mean_93_cast_fp16)[name = tensor("sub_74_cast_fp16")]; + tensor square_60_cast_fp16 = square(x = sub_74_cast_fp16)[name = tensor("square_60_cast_fp16")]; + tensor reduce_mean_121_axes_0 = const()[name = tensor("reduce_mean_121_axes_0"), val = tensor([-1])]; + tensor reduce_mean_121_keep_dims_0 = const()[name = tensor("reduce_mean_121_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_121_cast_fp16 = reduce_mean(axes = reduce_mean_121_axes_0, keep_dims = reduce_mean_121_keep_dims_0, x = square_60_cast_fp16)[name = tensor("reduce_mean_121_cast_fp16")]; + tensor var_2666_to_fp16 = const()[name = tensor("op_2666_to_fp16"), val = tensor(0x1.1p-20)]; + tensor var_2667_cast_fp16 = add(x = reduce_mean_121_cast_fp16, y = var_2666_to_fp16)[name = tensor("op_2667_cast_fp16")]; + tensor var_2668_cast_fp16 = sqrt(x = var_2667_cast_fp16)[name = tensor("op_2668_cast_fp16")]; + tensor x_295_cast_fp16 = real_div(x = sub_74_cast_fp16, y = var_2668_cast_fp16)[name = tensor("x_295_cast_fp16")]; + tensor var_2670_cast_fp16 = mul(x = x_295_cast_fp16, y = flow_net_res_blocks_4_in_ln_weight_to_fp16)[name = tensor("op_2670_cast_fp16")]; + tensor x_297_cast_fp16 = add(x = var_2670_cast_fp16, y = flow_net_res_blocks_4_in_ln_bias_to_fp16)[name = tensor("x_297_cast_fp16")]; + tensor var_2672_promoted_to_fp16 = const()[name = tensor("op_2672_promoted_to_fp16"), val = tensor(0x1p+0)]; + tensor var_2673_cast_fp16 = add(x = var_2656_cast_fp16_1, y = var_2672_promoted_to_fp16)[name = tensor("op_2673_cast_fp16")]; + tensor var_2674_cast_fp16 = mul(x = x_297_cast_fp16, y = var_2673_cast_fp16)[name = tensor("op_2674_cast_fp16")]; + tensor input_457_cast_fp16 = add(x = var_2674_cast_fp16, y = var_2656_cast_fp16_0)[name = tensor("input_457_cast_fp16")]; + tensor linear_175_cast_fp16 = linear(bias = flow_net_res_blocks_4_mlp_0_bias_to_fp16, weight = flow_net_res_blocks_4_mlp_0_weight_to_fp16, x = input_457_cast_fp16)[name = tensor("linear_175_cast_fp16")]; + tensor input_461_cast_fp16 = silu(x = linear_175_cast_fp16)[name = tensor("input_461_cast_fp16")]; + tensor linear_176_cast_fp16 = linear(bias = flow_net_res_blocks_4_mlp_2_bias_to_fp16, weight = flow_net_res_blocks_4_mlp_2_weight_to_fp16, x = input_461_cast_fp16)[name = tensor("linear_176_cast_fp16")]; + tensor var_2685_cast_fp16 = mul(x = var_2656_cast_fp16_2, y = linear_176_cast_fp16)[name = tensor("op_2685_cast_fp16")]; + tensor x_299_cast_fp16 = add(x = x_293_cast_fp16, y = var_2685_cast_fp16)[name = tensor("x_299_cast_fp16")]; + tensor linear_177_cast_fp16 = linear(bias = flow_net_res_blocks_5_adaLN_modulation_1_bias_to_fp16, weight = flow_net_res_blocks_5_adaLN_modulation_1_weight_to_fp16, x = input_423_cast_fp16)[name = tensor("linear_177_cast_fp16")]; + tensor var_2695_split_sizes_0 = const()[name = tensor("op_2695_split_sizes_0"), val = tensor([512, 512, 512])]; + tensor var_2695_axis_0 = const()[name = tensor("op_2695_axis_0"), val = tensor(-1)]; + tensor var_2695_cast_fp16_0, tensor var_2695_cast_fp16_1, tensor var_2695_cast_fp16_2 = split(axis = var_2695_axis_0, split_sizes = var_2695_split_sizes_0, x = linear_177_cast_fp16)[name = tensor("op_2695_cast_fp16")]; + tensor mean_95_axes_0 = const()[name = tensor("mean_95_axes_0"), val = tensor([-1])]; + tensor mean_95_keep_dims_0 = const()[name = tensor("mean_95_keep_dims_0"), val = tensor(true)]; + tensor mean_95_cast_fp16 = reduce_mean(axes = mean_95_axes_0, keep_dims = mean_95_keep_dims_0, x = x_299_cast_fp16)[name = tensor("mean_95_cast_fp16")]; + tensor sub_75_cast_fp16 = sub(x = x_299_cast_fp16, y = mean_95_cast_fp16)[name = tensor("sub_75_cast_fp16")]; + tensor square_61_cast_fp16 = square(x = sub_75_cast_fp16)[name = tensor("square_61_cast_fp16")]; + tensor reduce_mean_123_axes_0 = const()[name = tensor("reduce_mean_123_axes_0"), val = tensor([-1])]; + tensor reduce_mean_123_keep_dims_0 = const()[name = tensor("reduce_mean_123_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_123_cast_fp16 = reduce_mean(axes = reduce_mean_123_axes_0, keep_dims = reduce_mean_123_keep_dims_0, x = square_61_cast_fp16)[name = tensor("reduce_mean_123_cast_fp16")]; + tensor var_2705_to_fp16 = const()[name = tensor("op_2705_to_fp16"), val = tensor(0x1.1p-20)]; + tensor var_2706_cast_fp16 = add(x = reduce_mean_123_cast_fp16, y = var_2705_to_fp16)[name = tensor("op_2706_cast_fp16")]; + tensor var_2707_cast_fp16 = sqrt(x = var_2706_cast_fp16)[name = tensor("op_2707_cast_fp16")]; + tensor x_301_cast_fp16 = real_div(x = sub_75_cast_fp16, y = var_2707_cast_fp16)[name = tensor("x_301_cast_fp16")]; + tensor var_2709_cast_fp16 = mul(x = x_301_cast_fp16, y = flow_net_res_blocks_5_in_ln_weight_to_fp16)[name = tensor("op_2709_cast_fp16")]; + tensor x_303_cast_fp16 = add(x = var_2709_cast_fp16, y = flow_net_res_blocks_5_in_ln_bias_to_fp16)[name = tensor("x_303_cast_fp16")]; + tensor var_2711_promoted_to_fp16 = const()[name = tensor("op_2711_promoted_to_fp16"), val = tensor(0x1p+0)]; + tensor var_2712_cast_fp16 = add(x = var_2695_cast_fp16_1, y = var_2711_promoted_to_fp16)[name = tensor("op_2712_cast_fp16")]; + tensor var_2713_cast_fp16 = mul(x = x_303_cast_fp16, y = var_2712_cast_fp16)[name = tensor("op_2713_cast_fp16")]; + tensor input_465_cast_fp16 = add(x = var_2713_cast_fp16, y = var_2695_cast_fp16_0)[name = tensor("input_465_cast_fp16")]; + tensor linear_178_cast_fp16 = linear(bias = flow_net_res_blocks_5_mlp_0_bias_to_fp16, weight = flow_net_res_blocks_5_mlp_0_weight_to_fp16, x = input_465_cast_fp16)[name = tensor("linear_178_cast_fp16")]; + tensor input_469_cast_fp16 = silu(x = linear_178_cast_fp16)[name = tensor("input_469_cast_fp16")]; + tensor linear_179_cast_fp16 = linear(bias = flow_net_res_blocks_5_mlp_2_bias_to_fp16, weight = flow_net_res_blocks_5_mlp_2_weight_to_fp16, x = input_469_cast_fp16)[name = tensor("linear_179_cast_fp16")]; + tensor var_2724_cast_fp16 = mul(x = var_2695_cast_fp16_2, y = linear_179_cast_fp16)[name = tensor("op_2724_cast_fp16")]; + tensor x_305_cast_fp16 = add(x = x_299_cast_fp16, y = var_2724_cast_fp16)[name = tensor("x_305_cast_fp16")]; + tensor linear_180_cast_fp16 = linear(bias = flow_net_final_layer_adaLN_modulation_1_bias_to_fp16, weight = flow_net_final_layer_adaLN_modulation_1_weight_to_fp16, x = input_423_cast_fp16)[name = tensor("linear_180_cast_fp16")]; + tensor var_2733_split_sizes_0 = const()[name = tensor("op_2733_split_sizes_0"), val = tensor([512, 512])]; + tensor var_2733_axis_0 = const()[name = tensor("op_2733_axis_0"), val = tensor(-1)]; + tensor var_2733_cast_fp16_0, tensor var_2733_cast_fp16_1 = split(axis = var_2733_axis_0, split_sizes = var_2733_split_sizes_0, x = linear_180_cast_fp16)[name = tensor("op_2733_cast_fp16")]; + tensor mean_97_axes_0 = const()[name = tensor("mean_97_axes_0"), val = tensor([-1])]; + tensor mean_97_keep_dims_0 = const()[name = tensor("mean_97_keep_dims_0"), val = tensor(true)]; + tensor mean_97_cast_fp16 = reduce_mean(axes = mean_97_axes_0, keep_dims = mean_97_keep_dims_0, x = x_305_cast_fp16)[name = tensor("mean_97_cast_fp16")]; + tensor sub_76_cast_fp16 = sub(x = x_305_cast_fp16, y = mean_97_cast_fp16)[name = tensor("sub_76_cast_fp16")]; + tensor square_62_cast_fp16 = square(x = sub_76_cast_fp16)[name = tensor("square_62_cast_fp16")]; + tensor reduce_mean_125_axes_0 = const()[name = tensor("reduce_mean_125_axes_0"), val = tensor([-1])]; + tensor reduce_mean_125_keep_dims_0 = const()[name = tensor("reduce_mean_125_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_125_cast_fp16 = reduce_mean(axes = reduce_mean_125_axes_0, keep_dims = reduce_mean_125_keep_dims_0, x = square_62_cast_fp16)[name = tensor("reduce_mean_125_cast_fp16")]; + tensor var_2740_to_fp16 = const()[name = tensor("op_2740_to_fp16"), val = tensor(0x1.1p-20)]; + tensor var_2741_cast_fp16 = add(x = reduce_mean_125_cast_fp16, y = var_2740_to_fp16)[name = tensor("op_2741_cast_fp16")]; + tensor var_2742_cast_fp16 = sqrt(x = var_2741_cast_fp16)[name = tensor("op_2742_cast_fp16")]; + tensor x_307_cast_fp16 = real_div(x = sub_76_cast_fp16, y = var_2742_cast_fp16)[name = tensor("x_307_cast_fp16")]; + tensor var_2744_promoted_to_fp16 = const()[name = tensor("op_2744_promoted_to_fp16"), val = tensor(0x1p+0)]; + tensor var_2745_cast_fp16 = add(x = var_2733_cast_fp16_1, y = var_2744_promoted_to_fp16)[name = tensor("op_2745_cast_fp16")]; + tensor var_2746_cast_fp16 = mul(x = x_307_cast_fp16, y = var_2745_cast_fp16)[name = tensor("op_2746_cast_fp16")]; + tensor input_473_cast_fp16 = add(x = var_2746_cast_fp16, y = var_2733_cast_fp16_0)[name = tensor("input_473_cast_fp16")]; + tensor linear_181_cast_fp16 = linear(bias = flow_net_final_layer_linear_bias_to_fp16, weight = flow_net_final_layer_linear_weight_to_fp16, x = input_473_cast_fp16)[name = tensor("linear_181_cast_fp16")]; + tensor var_2757_to_fp16 = const()[name = tensor("op_2757_to_fp16"), val = tensor(0x1p-3)]; + tensor var_2758_cast_fp16 = mul(x = linear_181_cast_fp16, y = var_2757_to_fp16)[name = tensor("op_2758_cast_fp16")]; + tensor input_475_cast_fp16 = add(x = input_407_cast_fp16, y = var_2758_cast_fp16)[name = tensor("input_475_cast_fp16")]; + tensor linear_182_cast_fp16 = linear(bias = flow_net_input_proj_bias_to_fp16, weight = flow_net_input_proj_weight_to_fp16, x = input_475_cast_fp16)[name = tensor("linear_182_cast_fp16")]; + tensor input_479_to_fp16 = const()[name = tensor("input_479_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(19010432)))]; + tensor input_481_cast_fp16 = silu(x = input_479_to_fp16)[name = tensor("input_481_cast_fp16")]; + tensor linear_184_cast_fp16 = linear(bias = flow_net_time_embed_0_mlp_2_bias_to_fp16, weight = flow_net_time_embed_0_mlp_2_weight_to_fp16, x = input_481_cast_fp16)[name = tensor("linear_184_cast_fp16")]; + tensor reduce_mean_126_axes_0 = const()[name = tensor("reduce_mean_126_axes_0"), val = tensor([-1])]; + tensor reduce_mean_126_keep_dims_0 = const()[name = tensor("reduce_mean_126_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_126_cast_fp16 = reduce_mean(axes = reduce_mean_126_axes_0, keep_dims = reduce_mean_126_keep_dims_0, x = linear_184_cast_fp16)[name = tensor("reduce_mean_126_cast_fp16")]; + tensor sub_77_cast_fp16 = sub(x = linear_184_cast_fp16, y = reduce_mean_126_cast_fp16)[name = tensor("sub_77_cast_fp16")]; + tensor square_63_cast_fp16 = square(x = sub_77_cast_fp16)[name = tensor("square_63_cast_fp16")]; + tensor reduce_mean_127_axes_0 = const()[name = tensor("reduce_mean_127_axes_0"), val = tensor([-1])]; + tensor reduce_mean_127_keep_dims_0 = const()[name = tensor("reduce_mean_127_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_127_cast_fp16 = reduce_mean(axes = reduce_mean_127_axes_0, keep_dims = reduce_mean_127_keep_dims_0, x = square_63_cast_fp16)[name = tensor("reduce_mean_127_cast_fp16")]; + tensor real_div_14_to_fp16 = const()[name = tensor("real_div_14_to_fp16"), val = tensor(0x1.008p+0)]; + tensor mul_14_cast_fp16 = mul(x = reduce_mean_127_cast_fp16, y = real_div_14_to_fp16)[name = tensor("mul_14_cast_fp16")]; + tensor var_2824_to_fp16 = const()[name = tensor("op_2824_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_127_cast_fp16_0 = add(x = mul_14_cast_fp16, y = var_2824_to_fp16)[name = tensor("var_127_cast_fp16")]; + tensor var_2827_epsilon_0 = const()[name = tensor("op_2827_epsilon_0"), val = tensor(0x1.197998p-40)]; + tensor var_2827_cast_fp16 = rsqrt(epsilon = var_2827_epsilon_0, x = var_127_cast_fp16_0)[name = tensor("op_2827_cast_fp16")]; + tensor var_2828_cast_fp16 = mul(x = const_3_to_fp16, y = var_2827_cast_fp16)[name = tensor("op_2828_cast_fp16")]; + tensor var_2829_cast_fp16 = mul(x = linear_184_cast_fp16, y = var_2828_cast_fp16)[name = tensor("op_2829_cast_fp16")]; + tensor input_485_to_fp16 = const()[name = tensor("input_485_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(19011520)))]; + tensor input_487_cast_fp16 = silu(x = input_485_to_fp16)[name = tensor("input_487_cast_fp16")]; + tensor linear_186_cast_fp16 = linear(bias = flow_net_time_embed_1_mlp_2_bias_to_fp16, weight = flow_net_time_embed_1_mlp_2_weight_to_fp16, x = input_487_cast_fp16)[name = tensor("linear_186_cast_fp16")]; + tensor reduce_mean_128_axes_0 = const()[name = tensor("reduce_mean_128_axes_0"), val = tensor([-1])]; + tensor reduce_mean_128_keep_dims_0 = const()[name = tensor("reduce_mean_128_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_128_cast_fp16 = reduce_mean(axes = reduce_mean_128_axes_0, keep_dims = reduce_mean_128_keep_dims_0, x = linear_186_cast_fp16)[name = tensor("reduce_mean_128_cast_fp16")]; + tensor sub_79_cast_fp16 = sub(x = linear_186_cast_fp16, y = reduce_mean_128_cast_fp16)[name = tensor("sub_79_cast_fp16")]; + tensor square_64_cast_fp16 = square(x = sub_79_cast_fp16)[name = tensor("square_64_cast_fp16")]; + tensor reduce_mean_129_axes_0 = const()[name = tensor("reduce_mean_129_axes_0"), val = tensor([-1])]; + tensor reduce_mean_129_keep_dims_0 = const()[name = tensor("reduce_mean_129_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_129_cast_fp16 = reduce_mean(axes = reduce_mean_129_axes_0, keep_dims = reduce_mean_129_keep_dims_0, x = square_64_cast_fp16)[name = tensor("reduce_mean_129_cast_fp16")]; + tensor real_div_15_to_fp16 = const()[name = tensor("real_div_15_to_fp16"), val = tensor(0x1.008p+0)]; + tensor mul_15_cast_fp16 = mul(x = reduce_mean_129_cast_fp16, y = real_div_15_to_fp16)[name = tensor("mul_15_cast_fp16")]; + tensor var_2851_to_fp16 = const()[name = tensor("op_2851_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_129_cast_fp16 = add(x = mul_15_cast_fp16, y = var_2851_to_fp16)[name = tensor("var_129_cast_fp16")]; + tensor var_2854_epsilon_0 = const()[name = tensor("op_2854_epsilon_0"), val = tensor(0x1.197998p-40)]; + tensor var_2854_cast_fp16 = rsqrt(epsilon = var_2854_epsilon_0, x = var_129_cast_fp16)[name = tensor("op_2854_cast_fp16")]; + tensor var_2855_cast_fp16 = mul(x = const_5_to_fp16, y = var_2854_cast_fp16)[name = tensor("op_2855_cast_fp16")]; + tensor var_2856_cast_fp16 = mul(x = linear_186_cast_fp16, y = var_2855_cast_fp16)[name = tensor("op_2856_cast_fp16")]; + tensor var_2858_cast_fp16 = add(x = var_2829_cast_fp16, y = var_2856_cast_fp16)[name = tensor("op_2858_cast_fp16")]; + tensor _inversed_t_combined_y_0_to_fp16 = const()[name = tensor("_inversed_t_combined_y_0_to_fp16"), val = tensor(0x1p-1)]; + tensor _inversed_t_combined_cast_fp16 = mul(x = var_2858_cast_fp16, y = _inversed_t_combined_y_0_to_fp16)[name = tensor("_inversed_t_combined_cast_fp16")]; + tensor input_489_cast_fp16 = add(x = _inversed_t_combined_cast_fp16, y = linear_5_cast_fp16)[name = tensor("input_489_cast_fp16")]; + tensor input_491_cast_fp16 = silu(x = input_489_cast_fp16)[name = tensor("input_491_cast_fp16")]; + tensor linear_188_cast_fp16 = linear(bias = flow_net_res_blocks_0_adaLN_modulation_1_bias_to_fp16, weight = flow_net_res_blocks_0_adaLN_modulation_1_weight_to_fp16, x = input_491_cast_fp16)[name = tensor("linear_188_cast_fp16")]; + tensor var_2873_split_sizes_0 = const()[name = tensor("op_2873_split_sizes_0"), val = tensor([512, 512, 512])]; + tensor var_2873_axis_0 = const()[name = tensor("op_2873_axis_0"), val = tensor(-1)]; + tensor var_2873_cast_fp16_0, tensor var_2873_cast_fp16_1, tensor var_2873_cast_fp16_2 = split(axis = var_2873_axis_0, split_sizes = var_2873_split_sizes_0, x = linear_188_cast_fp16)[name = tensor("op_2873_cast_fp16")]; + tensor mean_99_axes_0 = const()[name = tensor("mean_99_axes_0"), val = tensor([-1])]; + tensor mean_99_keep_dims_0 = const()[name = tensor("mean_99_keep_dims_0"), val = tensor(true)]; + tensor mean_99_cast_fp16 = reduce_mean(axes = mean_99_axes_0, keep_dims = mean_99_keep_dims_0, x = linear_182_cast_fp16)[name = tensor("mean_99_cast_fp16")]; + tensor sub_81_cast_fp16 = sub(x = linear_182_cast_fp16, y = mean_99_cast_fp16)[name = tensor("sub_81_cast_fp16")]; + tensor square_65_cast_fp16 = square(x = sub_81_cast_fp16)[name = tensor("square_65_cast_fp16")]; + tensor reduce_mean_131_axes_0 = const()[name = tensor("reduce_mean_131_axes_0"), val = tensor([-1])]; + tensor reduce_mean_131_keep_dims_0 = const()[name = tensor("reduce_mean_131_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_131_cast_fp16 = reduce_mean(axes = reduce_mean_131_axes_0, keep_dims = reduce_mean_131_keep_dims_0, x = square_65_cast_fp16)[name = tensor("reduce_mean_131_cast_fp16")]; + tensor var_2883_to_fp16 = const()[name = tensor("op_2883_to_fp16"), val = tensor(0x1.1p-20)]; + tensor var_2884_cast_fp16 = add(x = reduce_mean_131_cast_fp16, y = var_2883_to_fp16)[name = tensor("op_2884_cast_fp16")]; + tensor var_2885_cast_fp16 = sqrt(x = var_2884_cast_fp16)[name = tensor("op_2885_cast_fp16")]; + tensor x_315_cast_fp16 = real_div(x = sub_81_cast_fp16, y = var_2885_cast_fp16)[name = tensor("x_315_cast_fp16")]; + tensor var_2887_cast_fp16 = mul(x = x_315_cast_fp16, y = flow_net_res_blocks_0_in_ln_weight_to_fp16)[name = tensor("op_2887_cast_fp16")]; + tensor x_317_cast_fp16 = add(x = var_2887_cast_fp16, y = flow_net_res_blocks_0_in_ln_bias_to_fp16)[name = tensor("x_317_cast_fp16")]; + tensor var_2889_promoted_to_fp16 = const()[name = tensor("op_2889_promoted_to_fp16"), val = tensor(0x1p+0)]; + tensor var_2890_cast_fp16 = add(x = var_2873_cast_fp16_1, y = var_2889_promoted_to_fp16)[name = tensor("op_2890_cast_fp16")]; + tensor var_2891_cast_fp16 = mul(x = x_317_cast_fp16, y = var_2890_cast_fp16)[name = tensor("op_2891_cast_fp16")]; + tensor input_493_cast_fp16 = add(x = var_2891_cast_fp16, y = var_2873_cast_fp16_0)[name = tensor("input_493_cast_fp16")]; + tensor linear_189_cast_fp16 = linear(bias = flow_net_res_blocks_0_mlp_0_bias_to_fp16, weight = flow_net_res_blocks_0_mlp_0_weight_to_fp16, x = input_493_cast_fp16)[name = tensor("linear_189_cast_fp16")]; + tensor input_497_cast_fp16 = silu(x = linear_189_cast_fp16)[name = tensor("input_497_cast_fp16")]; + tensor linear_190_cast_fp16 = linear(bias = flow_net_res_blocks_0_mlp_2_bias_to_fp16, weight = flow_net_res_blocks_0_mlp_2_weight_to_fp16, x = input_497_cast_fp16)[name = tensor("linear_190_cast_fp16")]; + tensor var_2902_cast_fp16 = mul(x = var_2873_cast_fp16_2, y = linear_190_cast_fp16)[name = tensor("op_2902_cast_fp16")]; + tensor x_319_cast_fp16 = add(x = linear_182_cast_fp16, y = var_2902_cast_fp16)[name = tensor("x_319_cast_fp16")]; + tensor linear_191_cast_fp16 = linear(bias = flow_net_res_blocks_1_adaLN_modulation_1_bias_to_fp16, weight = flow_net_res_blocks_1_adaLN_modulation_1_weight_to_fp16, x = input_491_cast_fp16)[name = tensor("linear_191_cast_fp16")]; + tensor var_2912_split_sizes_0 = const()[name = tensor("op_2912_split_sizes_0"), val = tensor([512, 512, 512])]; + tensor var_2912_axis_0 = const()[name = tensor("op_2912_axis_0"), val = tensor(-1)]; + tensor var_2912_cast_fp16_0, tensor var_2912_cast_fp16_1, tensor var_2912_cast_fp16_2 = split(axis = var_2912_axis_0, split_sizes = var_2912_split_sizes_0, x = linear_191_cast_fp16)[name = tensor("op_2912_cast_fp16")]; + tensor mean_101_axes_0 = const()[name = tensor("mean_101_axes_0"), val = tensor([-1])]; + tensor mean_101_keep_dims_0 = const()[name = tensor("mean_101_keep_dims_0"), val = tensor(true)]; + tensor mean_101_cast_fp16 = reduce_mean(axes = mean_101_axes_0, keep_dims = mean_101_keep_dims_0, x = x_319_cast_fp16)[name = tensor("mean_101_cast_fp16")]; + tensor sub_82_cast_fp16 = sub(x = x_319_cast_fp16, y = mean_101_cast_fp16)[name = tensor("sub_82_cast_fp16")]; + tensor square_66_cast_fp16 = square(x = sub_82_cast_fp16)[name = tensor("square_66_cast_fp16")]; + tensor reduce_mean_133_axes_0 = const()[name = tensor("reduce_mean_133_axes_0"), val = tensor([-1])]; + tensor reduce_mean_133_keep_dims_0 = const()[name = tensor("reduce_mean_133_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_133_cast_fp16 = reduce_mean(axes = reduce_mean_133_axes_0, keep_dims = reduce_mean_133_keep_dims_0, x = square_66_cast_fp16)[name = tensor("reduce_mean_133_cast_fp16")]; + tensor var_2922_to_fp16 = const()[name = tensor("op_2922_to_fp16"), val = tensor(0x1.1p-20)]; + tensor var_2923_cast_fp16 = add(x = reduce_mean_133_cast_fp16, y = var_2922_to_fp16)[name = tensor("op_2923_cast_fp16")]; + tensor var_2924_cast_fp16 = sqrt(x = var_2923_cast_fp16)[name = tensor("op_2924_cast_fp16")]; + tensor x_321_cast_fp16 = real_div(x = sub_82_cast_fp16, y = var_2924_cast_fp16)[name = tensor("x_321_cast_fp16")]; + tensor var_2926_cast_fp16 = mul(x = x_321_cast_fp16, y = flow_net_res_blocks_1_in_ln_weight_to_fp16)[name = tensor("op_2926_cast_fp16")]; + tensor x_323_cast_fp16 = add(x = var_2926_cast_fp16, y = flow_net_res_blocks_1_in_ln_bias_to_fp16)[name = tensor("x_323_cast_fp16")]; + tensor var_2928_promoted_to_fp16 = const()[name = tensor("op_2928_promoted_to_fp16"), val = tensor(0x1p+0)]; + tensor var_2929_cast_fp16 = add(x = var_2912_cast_fp16_1, y = var_2928_promoted_to_fp16)[name = tensor("op_2929_cast_fp16")]; + tensor var_2930_cast_fp16 = mul(x = x_323_cast_fp16, y = var_2929_cast_fp16)[name = tensor("op_2930_cast_fp16")]; + tensor input_501_cast_fp16 = add(x = var_2930_cast_fp16, y = var_2912_cast_fp16_0)[name = tensor("input_501_cast_fp16")]; + tensor linear_192_cast_fp16 = linear(bias = flow_net_res_blocks_1_mlp_0_bias_to_fp16, weight = flow_net_res_blocks_1_mlp_0_weight_to_fp16, x = input_501_cast_fp16)[name = tensor("linear_192_cast_fp16")]; + tensor input_505_cast_fp16 = silu(x = linear_192_cast_fp16)[name = tensor("input_505_cast_fp16")]; + tensor linear_193_cast_fp16 = linear(bias = flow_net_res_blocks_1_mlp_2_bias_to_fp16, weight = flow_net_res_blocks_1_mlp_2_weight_to_fp16, x = input_505_cast_fp16)[name = tensor("linear_193_cast_fp16")]; + tensor var_2941_cast_fp16 = mul(x = var_2912_cast_fp16_2, y = linear_193_cast_fp16)[name = tensor("op_2941_cast_fp16")]; + tensor x_325_cast_fp16 = add(x = x_319_cast_fp16, y = var_2941_cast_fp16)[name = tensor("x_325_cast_fp16")]; + tensor linear_194_cast_fp16 = linear(bias = flow_net_res_blocks_2_adaLN_modulation_1_bias_to_fp16, weight = flow_net_res_blocks_2_adaLN_modulation_1_weight_to_fp16, x = input_491_cast_fp16)[name = tensor("linear_194_cast_fp16")]; + tensor var_2951_split_sizes_0 = const()[name = tensor("op_2951_split_sizes_0"), val = tensor([512, 512, 512])]; + tensor var_2951_axis_0 = const()[name = tensor("op_2951_axis_0"), val = tensor(-1)]; + tensor var_2951_cast_fp16_0, tensor var_2951_cast_fp16_1, tensor var_2951_cast_fp16_2 = split(axis = var_2951_axis_0, split_sizes = var_2951_split_sizes_0, x = linear_194_cast_fp16)[name = tensor("op_2951_cast_fp16")]; + tensor mean_103_axes_0 = const()[name = tensor("mean_103_axes_0"), val = tensor([-1])]; + tensor mean_103_keep_dims_0 = const()[name = tensor("mean_103_keep_dims_0"), val = tensor(true)]; + tensor mean_103_cast_fp16 = reduce_mean(axes = mean_103_axes_0, keep_dims = mean_103_keep_dims_0, x = x_325_cast_fp16)[name = tensor("mean_103_cast_fp16")]; + tensor sub_83_cast_fp16 = sub(x = x_325_cast_fp16, y = mean_103_cast_fp16)[name = tensor("sub_83_cast_fp16")]; + tensor square_67_cast_fp16 = square(x = sub_83_cast_fp16)[name = tensor("square_67_cast_fp16")]; + tensor reduce_mean_135_axes_0 = const()[name = tensor("reduce_mean_135_axes_0"), val = tensor([-1])]; + tensor reduce_mean_135_keep_dims_0 = const()[name = tensor("reduce_mean_135_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_135_cast_fp16 = reduce_mean(axes = reduce_mean_135_axes_0, keep_dims = reduce_mean_135_keep_dims_0, x = square_67_cast_fp16)[name = tensor("reduce_mean_135_cast_fp16")]; + tensor var_2961_to_fp16 = const()[name = tensor("op_2961_to_fp16"), val = tensor(0x1.1p-20)]; + tensor var_2962_cast_fp16 = add(x = reduce_mean_135_cast_fp16, y = var_2961_to_fp16)[name = tensor("op_2962_cast_fp16")]; + tensor var_2963_cast_fp16 = sqrt(x = var_2962_cast_fp16)[name = tensor("op_2963_cast_fp16")]; + tensor x_327_cast_fp16 = real_div(x = sub_83_cast_fp16, y = var_2963_cast_fp16)[name = tensor("x_327_cast_fp16")]; + tensor var_2965_cast_fp16 = mul(x = x_327_cast_fp16, y = flow_net_res_blocks_2_in_ln_weight_to_fp16)[name = tensor("op_2965_cast_fp16")]; + tensor x_329_cast_fp16 = add(x = var_2965_cast_fp16, y = flow_net_res_blocks_2_in_ln_bias_to_fp16)[name = tensor("x_329_cast_fp16")]; + tensor var_2967_promoted_to_fp16 = const()[name = tensor("op_2967_promoted_to_fp16"), val = tensor(0x1p+0)]; + tensor var_2968_cast_fp16 = add(x = var_2951_cast_fp16_1, y = var_2967_promoted_to_fp16)[name = tensor("op_2968_cast_fp16")]; + tensor var_2969_cast_fp16 = mul(x = x_329_cast_fp16, y = var_2968_cast_fp16)[name = tensor("op_2969_cast_fp16")]; + tensor input_509_cast_fp16 = add(x = var_2969_cast_fp16, y = var_2951_cast_fp16_0)[name = tensor("input_509_cast_fp16")]; + tensor linear_195_cast_fp16 = linear(bias = flow_net_res_blocks_2_mlp_0_bias_to_fp16, weight = flow_net_res_blocks_2_mlp_0_weight_to_fp16, x = input_509_cast_fp16)[name = tensor("linear_195_cast_fp16")]; + tensor input_513_cast_fp16 = silu(x = linear_195_cast_fp16)[name = tensor("input_513_cast_fp16")]; + tensor linear_196_cast_fp16 = linear(bias = flow_net_res_blocks_2_mlp_2_bias_to_fp16, weight = flow_net_res_blocks_2_mlp_2_weight_to_fp16, x = input_513_cast_fp16)[name = tensor("linear_196_cast_fp16")]; + tensor var_2980_cast_fp16 = mul(x = var_2951_cast_fp16_2, y = linear_196_cast_fp16)[name = tensor("op_2980_cast_fp16")]; + tensor x_331_cast_fp16 = add(x = x_325_cast_fp16, y = var_2980_cast_fp16)[name = tensor("x_331_cast_fp16")]; + tensor linear_197_cast_fp16 = linear(bias = flow_net_res_blocks_3_adaLN_modulation_1_bias_to_fp16, weight = flow_net_res_blocks_3_adaLN_modulation_1_weight_to_fp16, x = input_491_cast_fp16)[name = tensor("linear_197_cast_fp16")]; + tensor var_2990_split_sizes_0 = const()[name = tensor("op_2990_split_sizes_0"), val = tensor([512, 512, 512])]; + tensor var_2990_axis_0 = const()[name = tensor("op_2990_axis_0"), val = tensor(-1)]; + tensor var_2990_cast_fp16_0, tensor var_2990_cast_fp16_1, tensor var_2990_cast_fp16_2 = split(axis = var_2990_axis_0, split_sizes = var_2990_split_sizes_0, x = linear_197_cast_fp16)[name = tensor("op_2990_cast_fp16")]; + tensor mean_105_axes_0 = const()[name = tensor("mean_105_axes_0"), val = tensor([-1])]; + tensor mean_105_keep_dims_0 = const()[name = tensor("mean_105_keep_dims_0"), val = tensor(true)]; + tensor mean_105_cast_fp16 = reduce_mean(axes = mean_105_axes_0, keep_dims = mean_105_keep_dims_0, x = x_331_cast_fp16)[name = tensor("mean_105_cast_fp16")]; + tensor sub_84_cast_fp16 = sub(x = x_331_cast_fp16, y = mean_105_cast_fp16)[name = tensor("sub_84_cast_fp16")]; + tensor square_68_cast_fp16 = square(x = sub_84_cast_fp16)[name = tensor("square_68_cast_fp16")]; + tensor reduce_mean_137_axes_0 = const()[name = tensor("reduce_mean_137_axes_0"), val = tensor([-1])]; + tensor reduce_mean_137_keep_dims_0 = const()[name = tensor("reduce_mean_137_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_137_cast_fp16 = reduce_mean(axes = reduce_mean_137_axes_0, keep_dims = reduce_mean_137_keep_dims_0, x = square_68_cast_fp16)[name = tensor("reduce_mean_137_cast_fp16")]; + tensor var_3000_to_fp16 = const()[name = tensor("op_3000_to_fp16"), val = tensor(0x1.1p-20)]; + tensor var_3001_cast_fp16 = add(x = reduce_mean_137_cast_fp16, y = var_3000_to_fp16)[name = tensor("op_3001_cast_fp16")]; + tensor var_3002_cast_fp16 = sqrt(x = var_3001_cast_fp16)[name = tensor("op_3002_cast_fp16")]; + tensor x_333_cast_fp16 = real_div(x = sub_84_cast_fp16, y = var_3002_cast_fp16)[name = tensor("x_333_cast_fp16")]; + tensor var_3004_cast_fp16 = mul(x = x_333_cast_fp16, y = flow_net_res_blocks_3_in_ln_weight_to_fp16)[name = tensor("op_3004_cast_fp16")]; + tensor x_335_cast_fp16 = add(x = var_3004_cast_fp16, y = flow_net_res_blocks_3_in_ln_bias_to_fp16)[name = tensor("x_335_cast_fp16")]; + tensor var_3006_promoted_to_fp16 = const()[name = tensor("op_3006_promoted_to_fp16"), val = tensor(0x1p+0)]; + tensor var_3007_cast_fp16 = add(x = var_2990_cast_fp16_1, y = var_3006_promoted_to_fp16)[name = tensor("op_3007_cast_fp16")]; + tensor var_3008_cast_fp16 = mul(x = x_335_cast_fp16, y = var_3007_cast_fp16)[name = tensor("op_3008_cast_fp16")]; + tensor input_517_cast_fp16 = add(x = var_3008_cast_fp16, y = var_2990_cast_fp16_0)[name = tensor("input_517_cast_fp16")]; + tensor linear_198_cast_fp16 = linear(bias = flow_net_res_blocks_3_mlp_0_bias_to_fp16, weight = flow_net_res_blocks_3_mlp_0_weight_to_fp16, x = input_517_cast_fp16)[name = tensor("linear_198_cast_fp16")]; + tensor input_521_cast_fp16 = silu(x = linear_198_cast_fp16)[name = tensor("input_521_cast_fp16")]; + tensor linear_199_cast_fp16 = linear(bias = flow_net_res_blocks_3_mlp_2_bias_to_fp16, weight = flow_net_res_blocks_3_mlp_2_weight_to_fp16, x = input_521_cast_fp16)[name = tensor("linear_199_cast_fp16")]; + tensor var_3019_cast_fp16 = mul(x = var_2990_cast_fp16_2, y = linear_199_cast_fp16)[name = tensor("op_3019_cast_fp16")]; + tensor x_337_cast_fp16 = add(x = x_331_cast_fp16, y = var_3019_cast_fp16)[name = tensor("x_337_cast_fp16")]; + tensor linear_200_cast_fp16 = linear(bias = flow_net_res_blocks_4_adaLN_modulation_1_bias_to_fp16, weight = flow_net_res_blocks_4_adaLN_modulation_1_weight_to_fp16, x = input_491_cast_fp16)[name = tensor("linear_200_cast_fp16")]; + tensor var_3029_split_sizes_0 = const()[name = tensor("op_3029_split_sizes_0"), val = tensor([512, 512, 512])]; + tensor var_3029_axis_0 = const()[name = tensor("op_3029_axis_0"), val = tensor(-1)]; + tensor var_3029_cast_fp16_0, tensor var_3029_cast_fp16_1, tensor var_3029_cast_fp16_2 = split(axis = var_3029_axis_0, split_sizes = var_3029_split_sizes_0, x = linear_200_cast_fp16)[name = tensor("op_3029_cast_fp16")]; + tensor mean_107_axes_0 = const()[name = tensor("mean_107_axes_0"), val = tensor([-1])]; + tensor mean_107_keep_dims_0 = const()[name = tensor("mean_107_keep_dims_0"), val = tensor(true)]; + tensor mean_107_cast_fp16 = reduce_mean(axes = mean_107_axes_0, keep_dims = mean_107_keep_dims_0, x = x_337_cast_fp16)[name = tensor("mean_107_cast_fp16")]; + tensor sub_85_cast_fp16 = sub(x = x_337_cast_fp16, y = mean_107_cast_fp16)[name = tensor("sub_85_cast_fp16")]; + tensor square_69_cast_fp16 = square(x = sub_85_cast_fp16)[name = tensor("square_69_cast_fp16")]; + tensor reduce_mean_139_axes_0 = const()[name = tensor("reduce_mean_139_axes_0"), val = tensor([-1])]; + tensor reduce_mean_139_keep_dims_0 = const()[name = tensor("reduce_mean_139_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_139_cast_fp16 = reduce_mean(axes = reduce_mean_139_axes_0, keep_dims = reduce_mean_139_keep_dims_0, x = square_69_cast_fp16)[name = tensor("reduce_mean_139_cast_fp16")]; + tensor var_3039_to_fp16 = const()[name = tensor("op_3039_to_fp16"), val = tensor(0x1.1p-20)]; + tensor var_3040_cast_fp16 = add(x = reduce_mean_139_cast_fp16, y = var_3039_to_fp16)[name = tensor("op_3040_cast_fp16")]; + tensor var_3041_cast_fp16 = sqrt(x = var_3040_cast_fp16)[name = tensor("op_3041_cast_fp16")]; + tensor x_339_cast_fp16 = real_div(x = sub_85_cast_fp16, y = var_3041_cast_fp16)[name = tensor("x_339_cast_fp16")]; + tensor var_3043_cast_fp16 = mul(x = x_339_cast_fp16, y = flow_net_res_blocks_4_in_ln_weight_to_fp16)[name = tensor("op_3043_cast_fp16")]; + tensor x_341_cast_fp16 = add(x = var_3043_cast_fp16, y = flow_net_res_blocks_4_in_ln_bias_to_fp16)[name = tensor("x_341_cast_fp16")]; + tensor var_3045_promoted_to_fp16 = const()[name = tensor("op_3045_promoted_to_fp16"), val = tensor(0x1p+0)]; + tensor var_3046_cast_fp16 = add(x = var_3029_cast_fp16_1, y = var_3045_promoted_to_fp16)[name = tensor("op_3046_cast_fp16")]; + tensor var_3047_cast_fp16 = mul(x = x_341_cast_fp16, y = var_3046_cast_fp16)[name = tensor("op_3047_cast_fp16")]; + tensor input_525_cast_fp16 = add(x = var_3047_cast_fp16, y = var_3029_cast_fp16_0)[name = tensor("input_525_cast_fp16")]; + tensor linear_201_cast_fp16 = linear(bias = flow_net_res_blocks_4_mlp_0_bias_to_fp16, weight = flow_net_res_blocks_4_mlp_0_weight_to_fp16, x = input_525_cast_fp16)[name = tensor("linear_201_cast_fp16")]; + tensor input_529_cast_fp16 = silu(x = linear_201_cast_fp16)[name = tensor("input_529_cast_fp16")]; + tensor linear_202_cast_fp16 = linear(bias = flow_net_res_blocks_4_mlp_2_bias_to_fp16, weight = flow_net_res_blocks_4_mlp_2_weight_to_fp16, x = input_529_cast_fp16)[name = tensor("linear_202_cast_fp16")]; + tensor var_3058_cast_fp16 = mul(x = var_3029_cast_fp16_2, y = linear_202_cast_fp16)[name = tensor("op_3058_cast_fp16")]; + tensor x_343_cast_fp16 = add(x = x_337_cast_fp16, y = var_3058_cast_fp16)[name = tensor("x_343_cast_fp16")]; + tensor linear_203_cast_fp16 = linear(bias = flow_net_res_blocks_5_adaLN_modulation_1_bias_to_fp16, weight = flow_net_res_blocks_5_adaLN_modulation_1_weight_to_fp16, x = input_491_cast_fp16)[name = tensor("linear_203_cast_fp16")]; + tensor var_3068_split_sizes_0 = const()[name = tensor("op_3068_split_sizes_0"), val = tensor([512, 512, 512])]; + tensor var_3068_axis_0 = const()[name = tensor("op_3068_axis_0"), val = tensor(-1)]; + tensor var_3068_cast_fp16_0, tensor var_3068_cast_fp16_1, tensor var_3068_cast_fp16_2 = split(axis = var_3068_axis_0, split_sizes = var_3068_split_sizes_0, x = linear_203_cast_fp16)[name = tensor("op_3068_cast_fp16")]; + tensor mean_109_axes_0 = const()[name = tensor("mean_109_axes_0"), val = tensor([-1])]; + tensor mean_109_keep_dims_0 = const()[name = tensor("mean_109_keep_dims_0"), val = tensor(true)]; + tensor mean_109_cast_fp16 = reduce_mean(axes = mean_109_axes_0, keep_dims = mean_109_keep_dims_0, x = x_343_cast_fp16)[name = tensor("mean_109_cast_fp16")]; + tensor sub_86_cast_fp16 = sub(x = x_343_cast_fp16, y = mean_109_cast_fp16)[name = tensor("sub_86_cast_fp16")]; + tensor square_70_cast_fp16 = square(x = sub_86_cast_fp16)[name = tensor("square_70_cast_fp16")]; + tensor reduce_mean_141_axes_0 = const()[name = tensor("reduce_mean_141_axes_0"), val = tensor([-1])]; + tensor reduce_mean_141_keep_dims_0 = const()[name = tensor("reduce_mean_141_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_141_cast_fp16 = reduce_mean(axes = reduce_mean_141_axes_0, keep_dims = reduce_mean_141_keep_dims_0, x = square_70_cast_fp16)[name = tensor("reduce_mean_141_cast_fp16")]; + tensor var_3078_to_fp16 = const()[name = tensor("op_3078_to_fp16"), val = tensor(0x1.1p-20)]; + tensor var_3079_cast_fp16 = add(x = reduce_mean_141_cast_fp16, y = var_3078_to_fp16)[name = tensor("op_3079_cast_fp16")]; + tensor var_3080_cast_fp16 = sqrt(x = var_3079_cast_fp16)[name = tensor("op_3080_cast_fp16")]; + tensor x_345_cast_fp16 = real_div(x = sub_86_cast_fp16, y = var_3080_cast_fp16)[name = tensor("x_345_cast_fp16")]; + tensor var_3082_cast_fp16 = mul(x = x_345_cast_fp16, y = flow_net_res_blocks_5_in_ln_weight_to_fp16)[name = tensor("op_3082_cast_fp16")]; + tensor x_347_cast_fp16 = add(x = var_3082_cast_fp16, y = flow_net_res_blocks_5_in_ln_bias_to_fp16)[name = tensor("x_347_cast_fp16")]; + tensor var_3084_promoted_to_fp16 = const()[name = tensor("op_3084_promoted_to_fp16"), val = tensor(0x1p+0)]; + tensor var_3085_cast_fp16 = add(x = var_3068_cast_fp16_1, y = var_3084_promoted_to_fp16)[name = tensor("op_3085_cast_fp16")]; + tensor var_3086_cast_fp16 = mul(x = x_347_cast_fp16, y = var_3085_cast_fp16)[name = tensor("op_3086_cast_fp16")]; + tensor input_533_cast_fp16 = add(x = var_3086_cast_fp16, y = var_3068_cast_fp16_0)[name = tensor("input_533_cast_fp16")]; + tensor linear_204_cast_fp16 = linear(bias = flow_net_res_blocks_5_mlp_0_bias_to_fp16, weight = flow_net_res_blocks_5_mlp_0_weight_to_fp16, x = input_533_cast_fp16)[name = tensor("linear_204_cast_fp16")]; + tensor input_537_cast_fp16 = silu(x = linear_204_cast_fp16)[name = tensor("input_537_cast_fp16")]; + tensor linear_205_cast_fp16 = linear(bias = flow_net_res_blocks_5_mlp_2_bias_to_fp16, weight = flow_net_res_blocks_5_mlp_2_weight_to_fp16, x = input_537_cast_fp16)[name = tensor("linear_205_cast_fp16")]; + tensor var_3097_cast_fp16 = mul(x = var_3068_cast_fp16_2, y = linear_205_cast_fp16)[name = tensor("op_3097_cast_fp16")]; + tensor x_349_cast_fp16 = add(x = x_343_cast_fp16, y = var_3097_cast_fp16)[name = tensor("x_349_cast_fp16")]; + tensor linear_206_cast_fp16 = linear(bias = flow_net_final_layer_adaLN_modulation_1_bias_to_fp16, weight = flow_net_final_layer_adaLN_modulation_1_weight_to_fp16, x = input_491_cast_fp16)[name = tensor("linear_206_cast_fp16")]; + tensor var_3106_split_sizes_0 = const()[name = tensor("op_3106_split_sizes_0"), val = tensor([512, 512])]; + tensor var_3106_axis_0 = const()[name = tensor("op_3106_axis_0"), val = tensor(-1)]; + tensor var_3106_cast_fp16_0, tensor var_3106_cast_fp16_1 = split(axis = var_3106_axis_0, split_sizes = var_3106_split_sizes_0, x = linear_206_cast_fp16)[name = tensor("op_3106_cast_fp16")]; + tensor mean_axes_0 = const()[name = tensor("mean_axes_0"), val = tensor([-1])]; + tensor mean_keep_dims_0 = const()[name = tensor("mean_keep_dims_0"), val = tensor(true)]; + tensor mean_cast_fp16 = reduce_mean(axes = mean_axes_0, keep_dims = mean_keep_dims_0, x = x_349_cast_fp16)[name = tensor("mean_cast_fp16")]; + tensor sub_87_cast_fp16 = sub(x = x_349_cast_fp16, y = mean_cast_fp16)[name = tensor("sub_87_cast_fp16")]; + tensor square_71_cast_fp16 = square(x = sub_87_cast_fp16)[name = tensor("square_71_cast_fp16")]; + tensor reduce_mean_143_axes_0 = const()[name = tensor("reduce_mean_143_axes_0"), val = tensor([-1])]; + tensor reduce_mean_143_keep_dims_0 = const()[name = tensor("reduce_mean_143_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_143_cast_fp16 = reduce_mean(axes = reduce_mean_143_axes_0, keep_dims = reduce_mean_143_keep_dims_0, x = square_71_cast_fp16)[name = tensor("reduce_mean_143_cast_fp16")]; + tensor var_3113_to_fp16 = const()[name = tensor("op_3113_to_fp16"), val = tensor(0x1.1p-20)]; + tensor var_3114_cast_fp16 = add(x = reduce_mean_143_cast_fp16, y = var_3113_to_fp16)[name = tensor("op_3114_cast_fp16")]; + tensor var_3115_cast_fp16 = sqrt(x = var_3114_cast_fp16)[name = tensor("op_3115_cast_fp16")]; + tensor x_cast_fp16 = real_div(x = sub_87_cast_fp16, y = var_3115_cast_fp16)[name = tensor("x_cast_fp16")]; + tensor var_3117_promoted_to_fp16 = const()[name = tensor("op_3117_promoted_to_fp16"), val = tensor(0x1p+0)]; + tensor var_3118_cast_fp16 = add(x = var_3106_cast_fp16_1, y = var_3117_promoted_to_fp16)[name = tensor("op_3118_cast_fp16")]; + tensor var_3119_cast_fp16 = mul(x = x_cast_fp16, y = var_3118_cast_fp16)[name = tensor("op_3119_cast_fp16")]; + tensor input_cast_fp16 = add(x = var_3119_cast_fp16, y = var_3106_cast_fp16_0)[name = tensor("input_cast_fp16")]; + tensor linear_207_cast_fp16 = linear(bias = flow_net_final_layer_linear_bias_to_fp16, weight = flow_net_final_layer_linear_weight_to_fp16, x = input_cast_fp16)[name = tensor("linear_207_cast_fp16")]; + tensor var_3124_to_fp16 = const()[name = tensor("op_3124_to_fp16"), val = tensor(0x1p-3)]; + tensor var_3125_cast_fp16 = mul(x = linear_207_cast_fp16, y = var_3124_to_fp16)[name = tensor("op_3125_cast_fp16")]; + tensor var_3127_cast_fp16 = add(x = input_475_cast_fp16, y = var_3125_cast_fp16)[name = tensor("op_3127_cast_fp16")]; + tensor var_3127_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_3127_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor latent_final = cast(dtype = var_3127_cast_fp16_to_fp32_dtype_0, x = var_3127_cast_fp16)[name = tensor("cast_144")]; + } -> (latent_final); +} \ No newline at end of file diff --git a/v2.1/german/flow_decoder_fused.mlmodelc/weights/weight.bin b/v2.1/german/flow_decoder_fused.mlmodelc/weights/weight.bin new file mode 100644 index 0000000000000000000000000000000000000000..3bf21f2fed8a3fdb7abedacc3e096c89597b91cd --- /dev/null +++ b/v2.1/german/flow_decoder_fused.mlmodelc/weights/weight.bin @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:6ee76653a292aaee52385edda87a2e5da948e1fbbb676ebf06e8d253f6582b03 +size 19012608 diff --git a/v2.1/german/flow_decoder_fused.mlpackage/Data/com.apple.CoreML/model.mlmodel b/v2.1/german/flow_decoder_fused.mlpackage/Data/com.apple.CoreML/model.mlmodel new file mode 100644 index 0000000000000000000000000000000000000000..63129de9afcdaa0047aa1893a89b5f29750c7f02 --- /dev/null +++ b/v2.1/german/flow_decoder_fused.mlpackage/Data/com.apple.CoreML/model.mlmodel @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:663a7b46b5f556cb138769e55c64dbcd5972e3ce521554725fd44f729e5160d2 +size 301733 diff --git a/v2.1/german/flow_decoder_fused.mlpackage/Data/com.apple.CoreML/weights/weight.bin b/v2.1/german/flow_decoder_fused.mlpackage/Data/com.apple.CoreML/weights/weight.bin new file mode 100644 index 0000000000000000000000000000000000000000..3bf21f2fed8a3fdb7abedacc3e096c89597b91cd --- /dev/null +++ b/v2.1/german/flow_decoder_fused.mlpackage/Data/com.apple.CoreML/weights/weight.bin @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:6ee76653a292aaee52385edda87a2e5da948e1fbbb676ebf06e8d253f6582b03 +size 19012608 diff --git a/v2.1/german/flow_decoder_fused.mlpackage/Manifest.json b/v2.1/german/flow_decoder_fused.mlpackage/Manifest.json new file mode 100644 index 0000000000000000000000000000000000000000..0375770037859cf8480ef2583745a8ee9d4c5c18 --- /dev/null +++ b/v2.1/german/flow_decoder_fused.mlpackage/Manifest.json @@ -0,0 +1,18 @@ +{ + "fileFormatVersion": "1.0.0", + "itemInfoEntries": { + "B4543635-3336-4A07-9A2D-E8B034AAE329": { + "author": "com.apple.CoreML", + "description": "CoreML Model Weights", + "name": "weights", + "path": "com.apple.CoreML/weights" + }, + "E1EBA6D7-07B4-489A-9FDC-E21D38C8941A": { + "author": "com.apple.CoreML", + "description": "CoreML Model Specification", + "name": "model.mlmodel", + "path": "com.apple.CoreML/model.mlmodel" + } + }, + "rootModelIdentifier": "E1EBA6D7-07B4-489A-9FDC-E21D38C8941A" +} diff --git a/v2.1/german/flowlm_step.mlmodelc/analytics/coremldata.bin b/v2.1/german/flowlm_step.mlmodelc/analytics/coremldata.bin new file mode 100644 index 0000000000000000000000000000000000000000..5917c78d15849749f3e055a09a2ef2b7284c23d1 --- /dev/null +++ b/v2.1/german/flowlm_step.mlmodelc/analytics/coremldata.bin @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:7f183c1be7fe10900e60f98b28fec4da6cf9c8b8eaf20b3ac6f0306fff6473e9 +size 243 diff --git a/v2.1/german/flowlm_step.mlmodelc/coremldata.bin b/v2.1/german/flowlm_step.mlmodelc/coremldata.bin new file mode 100644 index 0000000000000000000000000000000000000000..6054bec6807714ca59ab042de4f457ee68966144 --- /dev/null +++ b/v2.1/german/flowlm_step.mlmodelc/coremldata.bin @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:ea9b92f59f67021b9bb4dd6549df1ea6eb88806a0bd767218d85bb1738ff9465 +size 1191 diff --git a/v2.1/german/flowlm_step.mlmodelc/model.mil b/v2.1/german/flowlm_step.mlmodelc/model.mil new file mode 100644 index 0000000000000000000000000000000000000000..1884341832428143a40ec1b5e377a0722cb42bd1 --- /dev/null +++ b/v2.1/german/flowlm_step.mlmodelc/model.mil @@ -0,0 +1,1359 @@ +program(1.0) +[buildInfo = dict, tensor>({{"coremlc-component-MIL", "3520.4.1"}, {"coremlc-version", "3520.5.1"}, {"coremltools-component-torch", "2.12.0"}, {"coremltools-source-dialect", "TorchScript"}, {"coremltools-version", "9.0"}})] +{ + func main(tensor bos_emb, tensor cache0, tensor cache1, tensor cache2, tensor cache3, tensor cache4, tensor cache5, tensor position0, tensor position1, tensor position2, tensor position3, tensor position4, tensor position5, tensor sequence) { + tensor sequence_to_fp16_dtype_0 = const()[name = tensor("sequence_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor sequence_to_fp16 = cast(dtype = sequence_to_fp16_dtype_0, x = sequence)[name = tensor("cast_141")]; + tensor var_54_cast_fp16 = not_equal(x = sequence_to_fp16, y = sequence_to_fp16)[name = tensor("op_54_cast_fp16")]; + tensor expand_dims_0_axes_0 = const()[name = tensor("expand_dims_0_axes_0"), val = tensor([0, 1])]; + tensor bos_emb_to_fp16_dtype_0 = const()[name = tensor("bos_emb_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor bos_emb_to_fp16 = cast(dtype = bos_emb_to_fp16_dtype_0, x = bos_emb)[name = tensor("cast_140")]; + tensor expand_dims_0_cast_fp16 = expand_dims(axes = expand_dims_0_axes_0, x = bos_emb_to_fp16)[name = tensor("expand_dims_0_cast_fp16")]; + tensor input_1_cast_fp16 = select(a = expand_dims_0_cast_fp16, b = sequence_to_fp16, cond = var_54_cast_fp16)[name = tensor("input_1_cast_fp16")]; + tensor input_linear_weight_to_fp16 = const()[name = tensor("input_linear_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(64)))]; + tensor linear_0_bias_0_to_fp16 = const()[name = tensor("linear_0_bias_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(65664)))]; + tensor linear_0_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = input_linear_weight_to_fp16, x = input_1_cast_fp16)[name = tensor("linear_0_cast_fp16")]; + tensor input_5_axes_0 = const()[name = tensor("input_5_axes_0"), val = tensor([-1])]; + tensor norm0_1_weight_to_fp16 = const()[name = tensor("norm0_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(67776)))]; + tensor norm0_1_bias_to_fp16 = const()[name = tensor("norm0_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(69888)))]; + tensor var_60_to_fp16 = const()[name = tensor("op_60_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_5_cast_fp16 = layer_norm(axes = input_5_axes_0, beta = norm0_1_bias_to_fp16, epsilon = var_60_to_fp16, gamma = norm0_1_weight_to_fp16, x = linear_0_cast_fp16)[name = tensor("input_5_cast_fp16")]; + tensor attn0_in_proj_weight_to_fp16 = const()[name = tensor("attn0_in_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(72000)))]; + tensor linear_1_bias_0_to_fp16 = const()[name = tensor("linear_1_bias_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(6363520)))]; + tensor linear_1_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = attn0_in_proj_weight_to_fp16, x = input_5_cast_fp16)[name = tensor("linear_1_cast_fp16")]; + tensor var_74 = const()[name = tensor("op_74"), val = tensor([1, 1, 3, 16, 64])]; + tensor qkv_1_cast_fp16 = reshape(shape = var_74, x = linear_1_cast_fp16)[name = tensor("qkv_1_cast_fp16")]; + tensor q_1_begin_0 = const()[name = tensor("q_1_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor q_1_end_0 = const()[name = tensor("q_1_end_0"), val = tensor([1, 1, 1, 16, 64])]; + tensor q_1_end_mask_0 = const()[name = tensor("q_1_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor q_1_squeeze_mask_0 = const()[name = tensor("q_1_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor q_1_cast_fp16 = slice_by_index(begin = q_1_begin_0, end = q_1_end_0, end_mask = q_1_end_mask_0, squeeze_mask = q_1_squeeze_mask_0, x = qkv_1_cast_fp16)[name = tensor("q_1_cast_fp16")]; + tensor k_1_begin_0 = const()[name = tensor("k_1_begin_0"), val = tensor([0, 0, 1, 0, 0])]; + tensor k_1_end_0 = const()[name = tensor("k_1_end_0"), val = tensor([1, 1, 2, 16, 64])]; + tensor k_1_end_mask_0 = const()[name = tensor("k_1_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor k_1_squeeze_mask_0 = const()[name = tensor("k_1_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor k_1_cast_fp16 = slice_by_index(begin = k_1_begin_0, end = k_1_end_0, end_mask = k_1_end_mask_0, squeeze_mask = k_1_squeeze_mask_0, x = qkv_1_cast_fp16)[name = tensor("k_1_cast_fp16")]; + tensor v_1_begin_0 = const()[name = tensor("v_1_begin_0"), val = tensor([0, 0, 2, 0, 0])]; + tensor v_1_end_0 = const()[name = tensor("v_1_end_0"), val = tensor([1, 1, 3, 16, 64])]; + tensor v_1_end_mask_0 = const()[name = tensor("v_1_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor v_1_squeeze_mask_0 = const()[name = tensor("v_1_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor v_1_cast_fp16 = slice_by_index(begin = v_1_begin_0, end = v_1_end_0, end_mask = v_1_end_mask_0, squeeze_mask = v_1_squeeze_mask_0, x = qkv_1_cast_fp16)[name = tensor("v_1_cast_fp16")]; + tensor var_140 = const()[name = tensor("op_140"), val = tensor([1, 1, 1, 1])]; + tensor position0_to_fp16_dtype_0 = const()[name = tensor("position0_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor position0_to_fp16 = cast(dtype = position0_to_fp16_dtype_0, x = position0)[name = tensor("cast_139")]; + tensor ts_5_cast_fp16 = reshape(shape = var_140, x = position0_to_fp16)[name = tensor("ts_5_cast_fp16")]; + tensor var_147 = const()[name = tensor("op_147"), val = tensor([1, 1, 16, 32, 2])]; + tensor q_complex_1_cast_fp16 = reshape(shape = var_147, x = q_1_cast_fp16)[name = tensor("q_complex_1_cast_fp16")]; + tensor var_154 = const()[name = tensor("op_154"), val = tensor([1, 1, 16, 32, 2])]; + tensor k_complex_1_cast_fp16 = reshape(shape = var_154, x = k_1_cast_fp16)[name = tensor("k_complex_1_cast_fp16")]; + tensor var_158_begin_0 = const()[name = tensor("op_158_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_158_end_0 = const()[name = tensor("op_158_end_0"), val = tensor([1, 1, 16, 32, 1])]; + tensor var_158_end_mask_0 = const()[name = tensor("op_158_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_158_squeeze_mask_0 = const()[name = tensor("op_158_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_158_cast_fp16 = slice_by_index(begin = var_158_begin_0, end = var_158_end_0, end_mask = var_158_end_mask_0, squeeze_mask = var_158_squeeze_mask_0, x = q_complex_1_cast_fp16)[name = tensor("op_158_cast_fp16")]; + tensor var_166_begin_0 = const()[name = tensor("op_166_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_166_end_0 = const()[name = tensor("op_166_end_0"), val = tensor([1, 1, 16, 32, 2])]; + tensor var_166_end_mask_0 = const()[name = tensor("op_166_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_166_squeeze_mask_0 = const()[name = tensor("op_166_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_166_cast_fp16 = slice_by_index(begin = var_166_begin_0, end = var_166_end_0, end_mask = var_166_end_mask_0, squeeze_mask = var_166_squeeze_mask_0, x = q_complex_1_cast_fp16)[name = tensor("op_166_cast_fp16")]; + tensor var_174_begin_0 = const()[name = tensor("op_174_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_174_end_0 = const()[name = tensor("op_174_end_0"), val = tensor([1, 1, 16, 32, 1])]; + tensor var_174_end_mask_0 = const()[name = tensor("op_174_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_174_squeeze_mask_0 = const()[name = tensor("op_174_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_174_cast_fp16 = slice_by_index(begin = var_174_begin_0, end = var_174_end_0, end_mask = var_174_end_mask_0, squeeze_mask = var_174_squeeze_mask_0, x = k_complex_1_cast_fp16)[name = tensor("op_174_cast_fp16")]; + tensor var_182_begin_0 = const()[name = tensor("op_182_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_182_end_0 = const()[name = tensor("op_182_end_0"), val = tensor([1, 1, 16, 32, 2])]; + tensor var_182_end_mask_0 = const()[name = tensor("op_182_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_182_squeeze_mask_0 = const()[name = tensor("op_182_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_182_cast_fp16 = slice_by_index(begin = var_182_begin_0, end = var_182_end_0, end_mask = var_182_end_mask_0, squeeze_mask = var_182_squeeze_mask_0, x = k_complex_1_cast_fp16)[name = tensor("op_182_cast_fp16")]; + tensor freqs_1_to_fp16 = const()[name = tensor("freqs_1_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(6369728)))]; + tensor var_188_cast_fp16 = mul(x = freqs_1_to_fp16, y = ts_5_cast_fp16)[name = tensor("op_188_cast_fp16")]; + tensor rotr_1_cast_fp16 = cos(x = var_188_cast_fp16)[name = tensor("rotr_1_cast_fp16")]; + tensor roti_1_cast_fp16 = sin(x = var_188_cast_fp16)[name = tensor("roti_1_cast_fp16")]; + tensor var_192_cast_fp16 = mul(x = var_158_cast_fp16, y = rotr_1_cast_fp16)[name = tensor("op_192_cast_fp16")]; + tensor var_193_cast_fp16 = mul(x = var_166_cast_fp16, y = roti_1_cast_fp16)[name = tensor("op_193_cast_fp16")]; + tensor qor_1_cast_fp16 = sub(x = var_192_cast_fp16, y = var_193_cast_fp16)[name = tensor("qor_1_cast_fp16")]; + tensor var_196_cast_fp16 = mul(x = var_158_cast_fp16, y = roti_1_cast_fp16)[name = tensor("op_196_cast_fp16")]; + tensor var_197_cast_fp16 = mul(x = var_166_cast_fp16, y = rotr_1_cast_fp16)[name = tensor("op_197_cast_fp16")]; + tensor qoi_1_cast_fp16 = add(x = var_196_cast_fp16, y = var_197_cast_fp16)[name = tensor("qoi_1_cast_fp16")]; + tensor var_200_cast_fp16 = mul(x = var_174_cast_fp16, y = rotr_1_cast_fp16)[name = tensor("op_200_cast_fp16")]; + tensor var_201_cast_fp16 = mul(x = var_182_cast_fp16, y = roti_1_cast_fp16)[name = tensor("op_201_cast_fp16")]; + tensor kor_1_cast_fp16 = sub(x = var_200_cast_fp16, y = var_201_cast_fp16)[name = tensor("kor_1_cast_fp16")]; + tensor var_204_cast_fp16 = mul(x = var_174_cast_fp16, y = roti_1_cast_fp16)[name = tensor("op_204_cast_fp16")]; + tensor var_205_cast_fp16 = mul(x = var_182_cast_fp16, y = rotr_1_cast_fp16)[name = tensor("op_205_cast_fp16")]; + tensor koi_1_cast_fp16 = add(x = var_204_cast_fp16, y = var_205_cast_fp16)[name = tensor("koi_1_cast_fp16")]; + tensor qo_1_axis_0 = const()[name = tensor("qo_1_axis_0"), val = tensor(-1)]; + tensor qo_1_cast_fp16 = stack(axis = qo_1_axis_0, values = (qor_1_cast_fp16, qoi_1_cast_fp16))[name = tensor("qo_1_cast_fp16")]; + tensor ko_1_axis_0 = const()[name = tensor("ko_1_axis_0"), val = tensor(-1)]; + tensor ko_1_cast_fp16 = stack(axis = ko_1_axis_0, values = (kor_1_cast_fp16, koi_1_cast_fp16))[name = tensor("ko_1_cast_fp16")]; + tensor var_238 = const()[name = tensor("op_238"), val = tensor([1, 1, 16, 64])]; + tensor q_3_cast_fp16 = reshape(shape = var_238, x = qo_1_cast_fp16)[name = tensor("q_3_cast_fp16")]; + tensor var_244 = const()[name = tensor("op_244"), val = tensor([1, 1, 16, 64])]; + tensor k_3_cast_fp16 = reshape(shape = var_244, x = ko_1_cast_fp16)[name = tensor("k_3_cast_fp16")]; + tensor _inversed_269_y_0_to_fp16 = const()[name = tensor("_inversed_269_y_0_to_fp16"), val = tensor(0x1p-9)]; + tensor _inversed_269_cast_fp16 = mul(x = ts_5_cast_fp16, y = _inversed_269_y_0_to_fp16)[name = tensor("_inversed_269_cast_fp16")]; + tensor var_270_cast_fp16 = floor(x = _inversed_269_cast_fp16)[name = tensor("op_270_cast_fp16")]; + tensor var_271_to_fp16 = const()[name = tensor("op_271_to_fp16"), val = tensor(0x1p+9)]; + tensor var_272_cast_fp16 = mul(x = var_270_cast_fp16, y = var_271_to_fp16)[name = tensor("op_272_cast_fp16")]; + tensor write_indices_float_3_cast_fp16 = sub(x = ts_5_cast_fp16, y = var_272_cast_fp16)[name = tensor("write_indices_float_3_cast_fp16")]; + tensor var_279_dtype_0 = const()[name = tensor("op_279_dtype_0"), val = tensor("int32")]; + tensor write_indices_1_reps_0 = const()[name = tensor("write_indices_1_reps_0"), val = tensor([1, 1, 16, 64])]; + tensor write_indices_float_3_cast_fp16_to_int32 = cast(dtype = var_279_dtype_0, x = write_indices_float_3_cast_fp16)[name = tensor("cast_138")]; + tensor write_indices_1 = tile(reps = write_indices_1_reps_0, x = write_indices_float_3_cast_fp16_to_int32)[name = tensor("write_indices_1")]; + tensor var_289_begin_0 = const()[name = tensor("op_289_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_289_end_0 = const()[name = tensor("op_289_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor var_289_end_mask_0 = const()[name = tensor("op_289_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_289_squeeze_mask_0 = const()[name = tensor("op_289_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor cache0_to_fp16_dtype_0 = const()[name = tensor("cache0_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor cache0_to_fp16 = cast(dtype = cache0_to_fp16_dtype_0, x = cache0)[name = tensor("cast_137")]; + tensor var_289_cast_fp16 = slice_by_index(begin = var_289_begin_0, end = var_289_end_0, end_mask = var_289_end_mask_0, squeeze_mask = var_289_squeeze_mask_0, x = cache0_to_fp16)[name = tensor("op_289_cast_fp16")]; + tensor var_291_axis_0 = const()[name = tensor("op_291_axis_0"), val = tensor(1)]; + tensor var_291_mode_0 = const()[name = tensor("op_291_mode_0"), val = tensor("update")]; + tensor var_291_validate_indices_0 = const()[name = tensor("op_291_validate_indices_0"), val = tensor(false)]; + tensor var_291_cast_fp16 = scatter_along_axis(axis = var_291_axis_0, data = var_289_cast_fp16, indices = write_indices_1, mode = var_291_mode_0, updates = k_3_cast_fp16, validate_indices = var_291_validate_indices_0)[name = tensor("op_291_cast_fp16")]; + tensor concat_2 = const()[name = tensor("concat_2"), val = tensor([0, 0, 0, 0, 0])]; + tensor concat_3 = const()[name = tensor("concat_3"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_1_internal_tensor_assign_1_stride_0 = const()[name = tensor("new_cache_1_internal_tensor_assign_1_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_1_internal_tensor_assign_1_begin_mask_0 = const()[name = tensor("new_cache_1_internal_tensor_assign_1_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_1_internal_tensor_assign_1_end_mask_0 = const()[name = tensor("new_cache_1_internal_tensor_assign_1_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_1_internal_tensor_assign_1_squeeze_mask_0 = const()[name = tensor("new_cache_1_internal_tensor_assign_1_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_12 = const()[name = tensor("shape_12"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_0 = const()[name = tensor("reduce_prod_0"), val = tensor(1048576)]; + tensor range_1d_0_start_0 = const()[name = tensor("range_1d_0_start_0"), val = tensor(0)]; + tensor range_1d_0_step_0 = const()[name = tensor("range_1d_0_step_0"), val = tensor(1)]; + tensor range_1d_0 = range_1d(end = reduce_prod_0, start = range_1d_0_start_0, step = range_1d_0_step_0)[name = tensor("range_1d_0")]; + tensor reshape_0 = reshape(shape = shape_12, x = range_1d_0)[name = tensor("reshape_0")]; + tensor slice_by_index_0 = slice_by_index(begin = concat_2, begin_mask = new_cache_1_internal_tensor_assign_1_begin_mask_0, end = concat_3, end_mask = new_cache_1_internal_tensor_assign_1_end_mask_0, squeeze_mask = new_cache_1_internal_tensor_assign_1_squeeze_mask_0, stride = new_cache_1_internal_tensor_assign_1_stride_0, x = reshape_0)[name = tensor("slice_by_index_0")]; + tensor reshape_1_shape_0 = const()[name = tensor("reshape_1_shape_0"), val = tensor([-1])]; + tensor reshape_1 = reshape(shape = reshape_1_shape_0, x = slice_by_index_0)[name = tensor("reshape_1")]; + tensor reshape_2_shape_0 = const()[name = tensor("reshape_2_shape_0"), val = tensor([-1])]; + tensor reshape_2_cast_fp16 = reshape(shape = reshape_2_shape_0, x = var_291_cast_fp16)[name = tensor("reshape_2_cast_fp16")]; + tensor reshape_3_shape_0 = const()[name = tensor("reshape_3_shape_0"), val = tensor([-1])]; + tensor reshape_3_cast_fp16 = reshape(shape = reshape_3_shape_0, x = cache0_to_fp16)[name = tensor("reshape_3_cast_fp16")]; + tensor scatter_0_mode_0 = const()[name = tensor("scatter_0_mode_0"), val = tensor("update")]; + tensor scatter_0_axis_0 = const()[name = tensor("scatter_0_axis_0"), val = tensor(0)]; + tensor scatter_0_validate_indices_0 = const()[name = tensor("scatter_0_validate_indices_0"), val = tensor(false)]; + tensor scatter_0_cast_fp16 = scatter(axis = scatter_0_axis_0, data = reshape_3_cast_fp16, indices = reshape_1, mode = scatter_0_mode_0, updates = reshape_2_cast_fp16, validate_indices = scatter_0_validate_indices_0)[name = tensor("scatter_0_cast_fp16")]; + tensor reshape_4_cast_fp16 = reshape(shape = shape_12, x = scatter_0_cast_fp16)[name = tensor("reshape_4_cast_fp16")]; + tensor var_299_begin_0 = const()[name = tensor("op_299_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor var_299_end_0 = const()[name = tensor("op_299_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor var_299_end_mask_0 = const()[name = tensor("op_299_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_299_squeeze_mask_0 = const()[name = tensor("op_299_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor var_299_cast_fp16 = slice_by_index(begin = var_299_begin_0, end = var_299_end_0, end_mask = var_299_end_mask_0, squeeze_mask = var_299_squeeze_mask_0, x = reshape_4_cast_fp16)[name = tensor("op_299_cast_fp16")]; + tensor var_301_axis_0 = const()[name = tensor("op_301_axis_0"), val = tensor(1)]; + tensor var_301_mode_0 = const()[name = tensor("op_301_mode_0"), val = tensor("update")]; + tensor var_301_validate_indices_0 = const()[name = tensor("op_301_validate_indices_0"), val = tensor(false)]; + tensor var_301_cast_fp16 = scatter_along_axis(axis = var_301_axis_0, data = var_299_cast_fp16, indices = write_indices_1, mode = var_301_mode_0, updates = v_1_cast_fp16, validate_indices = var_301_validate_indices_0)[name = tensor("op_301_cast_fp16")]; + tensor concat_4 = const()[name = tensor("concat_4"), val = tensor([1, 0, 0, 0, 0])]; + tensor concat_5 = const()[name = tensor("concat_5"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_1_internal_tensor_assign_2_stride_0 = const()[name = tensor("new_cache_1_internal_tensor_assign_2_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_1_internal_tensor_assign_2_begin_mask_0 = const()[name = tensor("new_cache_1_internal_tensor_assign_2_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_1_internal_tensor_assign_2_end_mask_0 = const()[name = tensor("new_cache_1_internal_tensor_assign_2_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_1_internal_tensor_assign_2_squeeze_mask_0 = const()[name = tensor("new_cache_1_internal_tensor_assign_2_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_13 = const()[name = tensor("shape_13"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_1 = const()[name = tensor("reduce_prod_1"), val = tensor(1048576)]; + tensor range_1d_1_start_0 = const()[name = tensor("range_1d_1_start_0"), val = tensor(0)]; + tensor range_1d_1_step_0 = const()[name = tensor("range_1d_1_step_0"), val = tensor(1)]; + tensor range_1d_1 = range_1d(end = reduce_prod_1, start = range_1d_1_start_0, step = range_1d_1_step_0)[name = tensor("range_1d_1")]; + tensor reshape_5 = reshape(shape = shape_13, x = range_1d_1)[name = tensor("reshape_5")]; + tensor slice_by_index_1 = slice_by_index(begin = concat_4, begin_mask = new_cache_1_internal_tensor_assign_2_begin_mask_0, end = concat_5, end_mask = new_cache_1_internal_tensor_assign_2_end_mask_0, squeeze_mask = new_cache_1_internal_tensor_assign_2_squeeze_mask_0, stride = new_cache_1_internal_tensor_assign_2_stride_0, x = reshape_5)[name = tensor("slice_by_index_1")]; + tensor reshape_6_shape_0 = const()[name = tensor("reshape_6_shape_0"), val = tensor([-1])]; + tensor reshape_6 = reshape(shape = reshape_6_shape_0, x = slice_by_index_1)[name = tensor("reshape_6")]; + tensor reshape_7_shape_0 = const()[name = tensor("reshape_7_shape_0"), val = tensor([-1])]; + tensor reshape_7_cast_fp16 = reshape(shape = reshape_7_shape_0, x = var_301_cast_fp16)[name = tensor("reshape_7_cast_fp16")]; + tensor reshape_8_shape_0 = const()[name = tensor("reshape_8_shape_0"), val = tensor([-1])]; + tensor reshape_8_cast_fp16 = reshape(shape = reshape_8_shape_0, x = reshape_4_cast_fp16)[name = tensor("reshape_8_cast_fp16")]; + tensor scatter_1_mode_0 = const()[name = tensor("scatter_1_mode_0"), val = tensor("update")]; + tensor scatter_1_axis_0 = const()[name = tensor("scatter_1_axis_0"), val = tensor(0)]; + tensor scatter_1_validate_indices_0 = const()[name = tensor("scatter_1_validate_indices_0"), val = tensor(false)]; + tensor scatter_1_cast_fp16 = scatter(axis = scatter_1_axis_0, data = reshape_8_cast_fp16, indices = reshape_6, mode = scatter_1_mode_0, updates = reshape_7_cast_fp16, validate_indices = scatter_1_validate_indices_0)[name = tensor("scatter_1_cast_fp16")]; + tensor reshape_9_cast_fp16 = reshape(shape = shape_13, x = scatter_1_cast_fp16)[name = tensor("reshape_9_cast_fp16")]; + tensor reshape_9_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("reshape_9_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor keys_1_begin_0 = const()[name = tensor("keys_1_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor keys_1_end_0 = const()[name = tensor("keys_1_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor keys_1_end_mask_0 = const()[name = tensor("keys_1_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor keys_1_squeeze_mask_0 = const()[name = tensor("keys_1_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor keys_1_cast_fp16 = slice_by_index(begin = keys_1_begin_0, end = keys_1_end_0, end_mask = keys_1_end_mask_0, squeeze_mask = keys_1_squeeze_mask_0, x = reshape_9_cast_fp16)[name = tensor("keys_1_cast_fp16")]; + tensor values_1_begin_0 = const()[name = tensor("values_1_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor values_1_end_0 = const()[name = tensor("values_1_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor values_1_end_mask_0 = const()[name = tensor("values_1_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor values_1_squeeze_mask_0 = const()[name = tensor("values_1_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor values_1_cast_fp16 = slice_by_index(begin = values_1_begin_0, end = values_1_end_0, end_mask = values_1_end_mask_0, squeeze_mask = values_1_squeeze_mask_0, x = reshape_9_cast_fp16)[name = tensor("values_1_cast_fp16")]; + tensor var_313_cast_fp16 = not_equal(x = keys_1_cast_fp16, y = keys_1_cast_fp16)[name = tensor("op_313_cast_fp16")]; + tensor var_319_to_fp16 = const()[name = tensor("op_319_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(6369856)))]; + tensor keys_3_cast_fp16 = select(a = var_319_to_fp16, b = keys_1_cast_fp16, cond = var_313_cast_fp16)[name = tensor("keys_3_cast_fp16")]; + tensor var_321_cast_fp16 = not_equal(x = values_1_cast_fp16, y = values_1_cast_fp16)[name = tensor("op_321_cast_fp16")]; + tensor values_3_cast_fp16 = select(a = var_319_to_fp16, b = values_1_cast_fp16, cond = var_321_cast_fp16)[name = tensor("values_3_cast_fp16")]; + tensor var_345 = const()[name = tensor("op_345"), val = tensor([0, 2, 1, 3])]; + tensor var_361 = const()[name = tensor("op_361"), val = tensor([1, 1, 1])]; + tensor var_362_cast_fp16 = reshape(shape = var_361, x = position0_to_fp16)[name = tensor("op_362_cast_fp16")]; + tensor var_382_to_fp16 = const()[name = tensor("op_382_to_fp16"), val = tensor(0x1p+0)]; + tensor valid_len_1_cast_fp16 = add(x = var_362_cast_fp16, y = var_382_to_fp16)[name = tensor("valid_len_1_cast_fp16")]; + tensor k_positions_1_promoted_to_fp16 = const()[name = tensor("k_positions_1_promoted_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(7418496)))]; + tensor valid_mask_1_cast_fp16 = less(x = k_positions_1_promoted_to_fp16, y = valid_len_1_cast_fp16)[name = tensor("valid_mask_1_cast_fp16")]; + tensor causal_mask_1_cast_fp16 = less_equal(x = k_positions_1_promoted_to_fp16, y = var_362_cast_fp16)[name = tensor("causal_mask_1_cast_fp16")]; + tensor attn_mask_1 = logical_and(x = valid_mask_1_cast_fp16, y = causal_mask_1_cast_fp16)[name = tensor("attn_mask_1")]; + tensor attn_mask_3_axes_0 = const()[name = tensor("attn_mask_3_axes_0"), val = tensor([1])]; + tensor attn_mask_3 = expand_dims(axes = attn_mask_3_axes_0, x = attn_mask_1)[name = tensor("attn_mask_3")]; + tensor var_400_transpose_x_0 = const()[name = tensor("op_400_transpose_x_0"), val = tensor(false)]; + tensor var_400_transpose_y_0 = const()[name = tensor("op_400_transpose_y_0"), val = tensor(false)]; + tensor transpose_18_perm_0 = const()[name = tensor("transpose_18_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_19_perm_0 = const()[name = tensor("transpose_19_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_19 = transpose(perm = transpose_19_perm_0, x = keys_3_cast_fp16)[name = tensor("transpose_51")]; + tensor transpose_18 = transpose(perm = transpose_18_perm_0, x = q_3_cast_fp16)[name = tensor("transpose_52")]; + tensor var_400_cast_fp16 = matmul(transpose_x = var_400_transpose_x_0, transpose_y = var_400_transpose_y_0, x = transpose_18, y = transpose_19)[name = tensor("op_400_cast_fp16")]; + tensor var_394_to_fp16 = const()[name = tensor("op_394_to_fp16"), val = tensor([0x1p-3])]; + tensor attn_weights_1_cast_fp16 = mul(x = var_400_cast_fp16, y = var_394_to_fp16)[name = tensor("attn_weights_1_cast_fp16")]; + tensor var_402 = logical_not(x = attn_mask_3)[name = tensor("op_402")]; + tensor var_403_to_fp16 = const()[name = tensor("op_403_to_fp16"), val = tensor(-inf)]; + tensor attn_weights_3_cast_fp16 = select(a = var_403_to_fp16, b = attn_weights_1_cast_fp16, cond = var_402)[name = tensor("attn_weights_3_cast_fp16")]; + tensor var_405 = const()[name = tensor("op_405"), val = tensor(-1)]; + tensor attn_weights_5_cast_fp16 = softmax(axis = var_405, x = attn_weights_3_cast_fp16)[name = tensor("attn_weights_5_cast_fp16")]; + tensor attn_output_1_transpose_x_0 = const()[name = tensor("attn_output_1_transpose_x_0"), val = tensor(false)]; + tensor attn_output_1_transpose_y_0 = const()[name = tensor("attn_output_1_transpose_y_0"), val = tensor(false)]; + tensor values_5_cast_fp16 = transpose(perm = var_345, x = values_3_cast_fp16)[name = tensor("transpose_53")]; + tensor attn_output_1_cast_fp16 = matmul(transpose_x = attn_output_1_transpose_x_0, transpose_y = attn_output_1_transpose_y_0, x = attn_weights_5_cast_fp16, y = values_5_cast_fp16)[name = tensor("attn_output_1_cast_fp16")]; + tensor var_413 = const()[name = tensor("op_413"), val = tensor([0, 2, 1, 3])]; + tensor var_418 = const()[name = tensor("op_418"), val = tensor([1, 1, 1024])]; + tensor var_414_cast_fp16 = transpose(perm = var_413, x = attn_output_1_cast_fp16)[name = tensor("transpose_50")]; + tensor input_7_cast_fp16 = reshape(shape = var_418, x = var_414_cast_fp16)[name = tensor("input_7_cast_fp16")]; + tensor attn0_out_proj_weight_to_fp16 = const()[name = tensor("attn0_out_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(7419584)))]; + tensor linear_2_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = attn0_out_proj_weight_to_fp16, x = input_7_cast_fp16)[name = tensor("linear_2_cast_fp16")]; + tensor var_424_to_fp16 = const()[name = tensor("op_424_to_fp16"), val = tensor(0x1p+0)]; + tensor var_425_cast_fp16 = add(x = position0_to_fp16, y = var_424_to_fp16)[name = tensor("op_425_cast_fp16")]; + tensor var_425_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_425_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor input_9_cast_fp16 = add(x = linear_0_cast_fp16, y = linear_2_cast_fp16)[name = tensor("input_9_cast_fp16")]; + tensor input_11_axes_0 = const()[name = tensor("input_11_axes_0"), val = tensor([-1])]; + tensor norm0_2_weight_to_fp16 = const()[name = tensor("norm0_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9516800)))]; + tensor norm0_2_bias_to_fp16 = const()[name = tensor("norm0_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9518912)))]; + tensor var_429_to_fp16 = const()[name = tensor("op_429_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_11_cast_fp16 = layer_norm(axes = input_11_axes_0, beta = norm0_2_bias_to_fp16, epsilon = var_429_to_fp16, gamma = norm0_2_weight_to_fp16, x = input_9_cast_fp16)[name = tensor("input_11_cast_fp16")]; + tensor linear0_1_weight_to_fp16 = const()[name = tensor("linear0_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9521024)))]; + tensor linear_3_bias_0_to_fp16 = const()[name = tensor("linear_3_bias_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(17909696)))]; + tensor linear_3_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = linear0_1_weight_to_fp16, x = input_11_cast_fp16)[name = tensor("linear_3_cast_fp16")]; + tensor input_13_mode_0 = const()[name = tensor("input_13_mode_0"), val = tensor("EXACT")]; + tensor input_13_cast_fp16 = gelu(mode = input_13_mode_0, x = linear_3_cast_fp16)[name = tensor("input_13_cast_fp16")]; + tensor linear0_2_weight_to_fp16 = const()[name = tensor("linear0_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(17917952)))]; + tensor linear_4_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = linear0_2_weight_to_fp16, x = input_13_cast_fp16)[name = tensor("linear_4_cast_fp16")]; + tensor input_15_cast_fp16 = add(x = input_9_cast_fp16, y = linear_4_cast_fp16)[name = tensor("input_15_cast_fp16")]; + tensor input_17_axes_0 = const()[name = tensor("input_17_axes_0"), val = tensor([-1])]; + tensor norm1_1_weight_to_fp16 = const()[name = tensor("norm1_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(26306624)))]; + tensor norm1_1_bias_to_fp16 = const()[name = tensor("norm1_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(26308736)))]; + tensor var_446_to_fp16 = const()[name = tensor("op_446_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_17_cast_fp16 = layer_norm(axes = input_17_axes_0, beta = norm1_1_bias_to_fp16, epsilon = var_446_to_fp16, gamma = norm1_1_weight_to_fp16, x = input_15_cast_fp16)[name = tensor("input_17_cast_fp16")]; + tensor attn1_in_proj_weight_to_fp16 = const()[name = tensor("attn1_in_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(26310848)))]; + tensor linear_5_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = attn1_in_proj_weight_to_fp16, x = input_17_cast_fp16)[name = tensor("linear_5_cast_fp16")]; + tensor var_460 = const()[name = tensor("op_460"), val = tensor([1, 1, 3, 16, 64])]; + tensor qkv_3_cast_fp16 = reshape(shape = var_460, x = linear_5_cast_fp16)[name = tensor("qkv_3_cast_fp16")]; + tensor q_7_begin_0 = const()[name = tensor("q_7_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor q_7_end_0 = const()[name = tensor("q_7_end_0"), val = tensor([1, 1, 1, 16, 64])]; + tensor q_7_end_mask_0 = const()[name = tensor("q_7_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor q_7_squeeze_mask_0 = const()[name = tensor("q_7_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor q_7_cast_fp16 = slice_by_index(begin = q_7_begin_0, end = q_7_end_0, end_mask = q_7_end_mask_0, squeeze_mask = q_7_squeeze_mask_0, x = qkv_3_cast_fp16)[name = tensor("q_7_cast_fp16")]; + tensor k_5_begin_0 = const()[name = tensor("k_5_begin_0"), val = tensor([0, 0, 1, 0, 0])]; + tensor k_5_end_0 = const()[name = tensor("k_5_end_0"), val = tensor([1, 1, 2, 16, 64])]; + tensor k_5_end_mask_0 = const()[name = tensor("k_5_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor k_5_squeeze_mask_0 = const()[name = tensor("k_5_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor k_5_cast_fp16 = slice_by_index(begin = k_5_begin_0, end = k_5_end_0, end_mask = k_5_end_mask_0, squeeze_mask = k_5_squeeze_mask_0, x = qkv_3_cast_fp16)[name = tensor("k_5_cast_fp16")]; + tensor v_3_begin_0 = const()[name = tensor("v_3_begin_0"), val = tensor([0, 0, 2, 0, 0])]; + tensor v_3_end_0 = const()[name = tensor("v_3_end_0"), val = tensor([1, 1, 3, 16, 64])]; + tensor v_3_end_mask_0 = const()[name = tensor("v_3_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor v_3_squeeze_mask_0 = const()[name = tensor("v_3_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor v_3_cast_fp16 = slice_by_index(begin = v_3_begin_0, end = v_3_end_0, end_mask = v_3_end_mask_0, squeeze_mask = v_3_squeeze_mask_0, x = qkv_3_cast_fp16)[name = tensor("v_3_cast_fp16")]; + tensor var_526 = const()[name = tensor("op_526"), val = tensor([1, 1, 1, 1])]; + tensor position1_to_fp16_dtype_0 = const()[name = tensor("position1_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor position1_to_fp16 = cast(dtype = position1_to_fp16_dtype_0, x = position1)[name = tensor("cast_134")]; + tensor ts_11_cast_fp16 = reshape(shape = var_526, x = position1_to_fp16)[name = tensor("ts_11_cast_fp16")]; + tensor var_533 = const()[name = tensor("op_533"), val = tensor([1, 1, 16, 32, 2])]; + tensor q_complex_3_cast_fp16 = reshape(shape = var_533, x = q_7_cast_fp16)[name = tensor("q_complex_3_cast_fp16")]; + tensor var_540 = const()[name = tensor("op_540"), val = tensor([1, 1, 16, 32, 2])]; + tensor k_complex_3_cast_fp16 = reshape(shape = var_540, x = k_5_cast_fp16)[name = tensor("k_complex_3_cast_fp16")]; + tensor var_544_begin_0 = const()[name = tensor("op_544_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_544_end_0 = const()[name = tensor("op_544_end_0"), val = tensor([1, 1, 16, 32, 1])]; + tensor var_544_end_mask_0 = const()[name = tensor("op_544_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_544_squeeze_mask_0 = const()[name = tensor("op_544_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_544_cast_fp16 = slice_by_index(begin = var_544_begin_0, end = var_544_end_0, end_mask = var_544_end_mask_0, squeeze_mask = var_544_squeeze_mask_0, x = q_complex_3_cast_fp16)[name = tensor("op_544_cast_fp16")]; + tensor var_552_begin_0 = const()[name = tensor("op_552_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_552_end_0 = const()[name = tensor("op_552_end_0"), val = tensor([1, 1, 16, 32, 2])]; + tensor var_552_end_mask_0 = const()[name = tensor("op_552_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_552_squeeze_mask_0 = const()[name = tensor("op_552_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_552_cast_fp16 = slice_by_index(begin = var_552_begin_0, end = var_552_end_0, end_mask = var_552_end_mask_0, squeeze_mask = var_552_squeeze_mask_0, x = q_complex_3_cast_fp16)[name = tensor("op_552_cast_fp16")]; + tensor var_560_begin_0 = const()[name = tensor("op_560_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_560_end_0 = const()[name = tensor("op_560_end_0"), val = tensor([1, 1, 16, 32, 1])]; + tensor var_560_end_mask_0 = const()[name = tensor("op_560_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_560_squeeze_mask_0 = const()[name = tensor("op_560_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_560_cast_fp16 = slice_by_index(begin = var_560_begin_0, end = var_560_end_0, end_mask = var_560_end_mask_0, squeeze_mask = var_560_squeeze_mask_0, x = k_complex_3_cast_fp16)[name = tensor("op_560_cast_fp16")]; + tensor var_568_begin_0 = const()[name = tensor("op_568_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_568_end_0 = const()[name = tensor("op_568_end_0"), val = tensor([1, 1, 16, 32, 2])]; + tensor var_568_end_mask_0 = const()[name = tensor("op_568_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_568_squeeze_mask_0 = const()[name = tensor("op_568_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_568_cast_fp16 = slice_by_index(begin = var_568_begin_0, end = var_568_end_0, end_mask = var_568_end_mask_0, squeeze_mask = var_568_squeeze_mask_0, x = k_complex_3_cast_fp16)[name = tensor("op_568_cast_fp16")]; + tensor freqs_3_to_fp16 = const()[name = tensor("freqs_3_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(32602368)))]; + tensor var_574_cast_fp16 = mul(x = freqs_3_to_fp16, y = ts_11_cast_fp16)[name = tensor("op_574_cast_fp16")]; + tensor rotr_3_cast_fp16 = cos(x = var_574_cast_fp16)[name = tensor("rotr_3_cast_fp16")]; + tensor roti_3_cast_fp16 = sin(x = var_574_cast_fp16)[name = tensor("roti_3_cast_fp16")]; + tensor var_578_cast_fp16 = mul(x = var_544_cast_fp16, y = rotr_3_cast_fp16)[name = tensor("op_578_cast_fp16")]; + tensor var_579_cast_fp16 = mul(x = var_552_cast_fp16, y = roti_3_cast_fp16)[name = tensor("op_579_cast_fp16")]; + tensor qor_5_cast_fp16 = sub(x = var_578_cast_fp16, y = var_579_cast_fp16)[name = tensor("qor_5_cast_fp16")]; + tensor var_582_cast_fp16 = mul(x = var_544_cast_fp16, y = roti_3_cast_fp16)[name = tensor("op_582_cast_fp16")]; + tensor var_583_cast_fp16 = mul(x = var_552_cast_fp16, y = rotr_3_cast_fp16)[name = tensor("op_583_cast_fp16")]; + tensor qoi_5_cast_fp16 = add(x = var_582_cast_fp16, y = var_583_cast_fp16)[name = tensor("qoi_5_cast_fp16")]; + tensor var_586_cast_fp16 = mul(x = var_560_cast_fp16, y = rotr_3_cast_fp16)[name = tensor("op_586_cast_fp16")]; + tensor var_587_cast_fp16 = mul(x = var_568_cast_fp16, y = roti_3_cast_fp16)[name = tensor("op_587_cast_fp16")]; + tensor kor_5_cast_fp16 = sub(x = var_586_cast_fp16, y = var_587_cast_fp16)[name = tensor("kor_5_cast_fp16")]; + tensor var_590_cast_fp16 = mul(x = var_560_cast_fp16, y = roti_3_cast_fp16)[name = tensor("op_590_cast_fp16")]; + tensor var_591_cast_fp16 = mul(x = var_568_cast_fp16, y = rotr_3_cast_fp16)[name = tensor("op_591_cast_fp16")]; + tensor koi_5_cast_fp16 = add(x = var_590_cast_fp16, y = var_591_cast_fp16)[name = tensor("koi_5_cast_fp16")]; + tensor qo_3_axis_0 = const()[name = tensor("qo_3_axis_0"), val = tensor(-1)]; + tensor qo_3_cast_fp16 = stack(axis = qo_3_axis_0, values = (qor_5_cast_fp16, qoi_5_cast_fp16))[name = tensor("qo_3_cast_fp16")]; + tensor ko_3_axis_0 = const()[name = tensor("ko_3_axis_0"), val = tensor(-1)]; + tensor ko_3_cast_fp16 = stack(axis = ko_3_axis_0, values = (kor_5_cast_fp16, koi_5_cast_fp16))[name = tensor("ko_3_cast_fp16")]; + tensor var_624 = const()[name = tensor("op_624"), val = tensor([1, 1, 16, 64])]; + tensor q_9_cast_fp16 = reshape(shape = var_624, x = qo_3_cast_fp16)[name = tensor("q_9_cast_fp16")]; + tensor var_630 = const()[name = tensor("op_630"), val = tensor([1, 1, 16, 64])]; + tensor k_7_cast_fp16 = reshape(shape = var_630, x = ko_3_cast_fp16)[name = tensor("k_7_cast_fp16")]; + tensor _inversed_655_y_0_to_fp16 = const()[name = tensor("_inversed_655_y_0_to_fp16"), val = tensor(0x1p-9)]; + tensor _inversed_655_cast_fp16 = mul(x = ts_11_cast_fp16, y = _inversed_655_y_0_to_fp16)[name = tensor("_inversed_655_cast_fp16")]; + tensor var_656_cast_fp16 = floor(x = _inversed_655_cast_fp16)[name = tensor("op_656_cast_fp16")]; + tensor var_657_to_fp16 = const()[name = tensor("op_657_to_fp16"), val = tensor(0x1p+9)]; + tensor var_658_cast_fp16 = mul(x = var_656_cast_fp16, y = var_657_to_fp16)[name = tensor("op_658_cast_fp16")]; + tensor write_indices_float_7_cast_fp16 = sub(x = ts_11_cast_fp16, y = var_658_cast_fp16)[name = tensor("write_indices_float_7_cast_fp16")]; + tensor var_665_dtype_0 = const()[name = tensor("op_665_dtype_0"), val = tensor("int32")]; + tensor write_indices_3_reps_0 = const()[name = tensor("write_indices_3_reps_0"), val = tensor([1, 1, 16, 64])]; + tensor write_indices_float_7_cast_fp16_to_int32 = cast(dtype = var_665_dtype_0, x = write_indices_float_7_cast_fp16)[name = tensor("cast_133")]; + tensor write_indices_3 = tile(reps = write_indices_3_reps_0, x = write_indices_float_7_cast_fp16_to_int32)[name = tensor("write_indices_3")]; + tensor var_675_begin_0 = const()[name = tensor("op_675_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_675_end_0 = const()[name = tensor("op_675_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor var_675_end_mask_0 = const()[name = tensor("op_675_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_675_squeeze_mask_0 = const()[name = tensor("op_675_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor cache1_to_fp16_dtype_0 = const()[name = tensor("cache1_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor cache1_to_fp16 = cast(dtype = cache1_to_fp16_dtype_0, x = cache1)[name = tensor("cast_132")]; + tensor var_675_cast_fp16 = slice_by_index(begin = var_675_begin_0, end = var_675_end_0, end_mask = var_675_end_mask_0, squeeze_mask = var_675_squeeze_mask_0, x = cache1_to_fp16)[name = tensor("op_675_cast_fp16")]; + tensor var_677_axis_0 = const()[name = tensor("op_677_axis_0"), val = tensor(1)]; + tensor var_677_mode_0 = const()[name = tensor("op_677_mode_0"), val = tensor("update")]; + tensor var_677_validate_indices_0 = const()[name = tensor("op_677_validate_indices_0"), val = tensor(false)]; + tensor var_677_cast_fp16 = scatter_along_axis(axis = var_677_axis_0, data = var_675_cast_fp16, indices = write_indices_3, mode = var_677_mode_0, updates = k_7_cast_fp16, validate_indices = var_677_validate_indices_0)[name = tensor("op_677_cast_fp16")]; + tensor concat_9 = const()[name = tensor("concat_9"), val = tensor([0, 0, 0, 0, 0])]; + tensor concat_10 = const()[name = tensor("concat_10"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_3_internal_tensor_assign_1_stride_0 = const()[name = tensor("new_cache_3_internal_tensor_assign_1_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_3_internal_tensor_assign_1_begin_mask_0 = const()[name = tensor("new_cache_3_internal_tensor_assign_1_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_3_internal_tensor_assign_1_end_mask_0 = const()[name = tensor("new_cache_3_internal_tensor_assign_1_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_3_internal_tensor_assign_1_squeeze_mask_0 = const()[name = tensor("new_cache_3_internal_tensor_assign_1_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_14 = const()[name = tensor("shape_14"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_2 = const()[name = tensor("reduce_prod_2"), val = tensor(1048576)]; + tensor range_1d_2_start_0 = const()[name = tensor("range_1d_2_start_0"), val = tensor(0)]; + tensor range_1d_2_step_0 = const()[name = tensor("range_1d_2_step_0"), val = tensor(1)]; + tensor range_1d_2 = range_1d(end = reduce_prod_2, start = range_1d_2_start_0, step = range_1d_2_step_0)[name = tensor("range_1d_2")]; + tensor reshape_10 = reshape(shape = shape_14, x = range_1d_2)[name = tensor("reshape_10")]; + tensor slice_by_index_2 = slice_by_index(begin = concat_9, begin_mask = new_cache_3_internal_tensor_assign_1_begin_mask_0, end = concat_10, end_mask = new_cache_3_internal_tensor_assign_1_end_mask_0, squeeze_mask = new_cache_3_internal_tensor_assign_1_squeeze_mask_0, stride = new_cache_3_internal_tensor_assign_1_stride_0, x = reshape_10)[name = tensor("slice_by_index_2")]; + tensor reshape_11_shape_0 = const()[name = tensor("reshape_11_shape_0"), val = tensor([-1])]; + tensor reshape_11 = reshape(shape = reshape_11_shape_0, x = slice_by_index_2)[name = tensor("reshape_11")]; + tensor reshape_12_shape_0 = const()[name = tensor("reshape_12_shape_0"), val = tensor([-1])]; + tensor reshape_12_cast_fp16 = reshape(shape = reshape_12_shape_0, x = var_677_cast_fp16)[name = tensor("reshape_12_cast_fp16")]; + tensor reshape_13_shape_0 = const()[name = tensor("reshape_13_shape_0"), val = tensor([-1])]; + tensor reshape_13_cast_fp16 = reshape(shape = reshape_13_shape_0, x = cache1_to_fp16)[name = tensor("reshape_13_cast_fp16")]; + tensor scatter_2_mode_0 = const()[name = tensor("scatter_2_mode_0"), val = tensor("update")]; + tensor scatter_2_axis_0 = const()[name = tensor("scatter_2_axis_0"), val = tensor(0)]; + tensor scatter_2_validate_indices_0 = const()[name = tensor("scatter_2_validate_indices_0"), val = tensor(false)]; + tensor scatter_2_cast_fp16 = scatter(axis = scatter_2_axis_0, data = reshape_13_cast_fp16, indices = reshape_11, mode = scatter_2_mode_0, updates = reshape_12_cast_fp16, validate_indices = scatter_2_validate_indices_0)[name = tensor("scatter_2_cast_fp16")]; + tensor reshape_14_cast_fp16 = reshape(shape = shape_14, x = scatter_2_cast_fp16)[name = tensor("reshape_14_cast_fp16")]; + tensor var_685_begin_0 = const()[name = tensor("op_685_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor var_685_end_0 = const()[name = tensor("op_685_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor var_685_end_mask_0 = const()[name = tensor("op_685_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_685_squeeze_mask_0 = const()[name = tensor("op_685_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor var_685_cast_fp16 = slice_by_index(begin = var_685_begin_0, end = var_685_end_0, end_mask = var_685_end_mask_0, squeeze_mask = var_685_squeeze_mask_0, x = reshape_14_cast_fp16)[name = tensor("op_685_cast_fp16")]; + tensor var_687_axis_0 = const()[name = tensor("op_687_axis_0"), val = tensor(1)]; + tensor var_687_mode_0 = const()[name = tensor("op_687_mode_0"), val = tensor("update")]; + tensor var_687_validate_indices_0 = const()[name = tensor("op_687_validate_indices_0"), val = tensor(false)]; + tensor var_687_cast_fp16 = scatter_along_axis(axis = var_687_axis_0, data = var_685_cast_fp16, indices = write_indices_3, mode = var_687_mode_0, updates = v_3_cast_fp16, validate_indices = var_687_validate_indices_0)[name = tensor("op_687_cast_fp16")]; + tensor concat_11 = const()[name = tensor("concat_11"), val = tensor([1, 0, 0, 0, 0])]; + tensor concat_12 = const()[name = tensor("concat_12"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_3_internal_tensor_assign_2_stride_0 = const()[name = tensor("new_cache_3_internal_tensor_assign_2_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_3_internal_tensor_assign_2_begin_mask_0 = const()[name = tensor("new_cache_3_internal_tensor_assign_2_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_3_internal_tensor_assign_2_end_mask_0 = const()[name = tensor("new_cache_3_internal_tensor_assign_2_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_3_internal_tensor_assign_2_squeeze_mask_0 = const()[name = tensor("new_cache_3_internal_tensor_assign_2_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_15 = const()[name = tensor("shape_15"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_3 = const()[name = tensor("reduce_prod_3"), val = tensor(1048576)]; + tensor range_1d_3_start_0 = const()[name = tensor("range_1d_3_start_0"), val = tensor(0)]; + tensor range_1d_3_step_0 = const()[name = tensor("range_1d_3_step_0"), val = tensor(1)]; + tensor range_1d_3 = range_1d(end = reduce_prod_3, start = range_1d_3_start_0, step = range_1d_3_step_0)[name = tensor("range_1d_3")]; + tensor reshape_15 = reshape(shape = shape_15, x = range_1d_3)[name = tensor("reshape_15")]; + tensor slice_by_index_3 = slice_by_index(begin = concat_11, begin_mask = new_cache_3_internal_tensor_assign_2_begin_mask_0, end = concat_12, end_mask = new_cache_3_internal_tensor_assign_2_end_mask_0, squeeze_mask = new_cache_3_internal_tensor_assign_2_squeeze_mask_0, stride = new_cache_3_internal_tensor_assign_2_stride_0, x = reshape_15)[name = tensor("slice_by_index_3")]; + tensor reshape_16_shape_0 = const()[name = tensor("reshape_16_shape_0"), val = tensor([-1])]; + tensor reshape_16 = reshape(shape = reshape_16_shape_0, x = slice_by_index_3)[name = tensor("reshape_16")]; + tensor reshape_17_shape_0 = const()[name = tensor("reshape_17_shape_0"), val = tensor([-1])]; + tensor reshape_17_cast_fp16 = reshape(shape = reshape_17_shape_0, x = var_687_cast_fp16)[name = tensor("reshape_17_cast_fp16")]; + tensor reshape_18_shape_0 = const()[name = tensor("reshape_18_shape_0"), val = tensor([-1])]; + tensor reshape_18_cast_fp16 = reshape(shape = reshape_18_shape_0, x = reshape_14_cast_fp16)[name = tensor("reshape_18_cast_fp16")]; + tensor scatter_3_mode_0 = const()[name = tensor("scatter_3_mode_0"), val = tensor("update")]; + tensor scatter_3_axis_0 = const()[name = tensor("scatter_3_axis_0"), val = tensor(0)]; + tensor scatter_3_validate_indices_0 = const()[name = tensor("scatter_3_validate_indices_0"), val = tensor(false)]; + tensor scatter_3_cast_fp16 = scatter(axis = scatter_3_axis_0, data = reshape_18_cast_fp16, indices = reshape_16, mode = scatter_3_mode_0, updates = reshape_17_cast_fp16, validate_indices = scatter_3_validate_indices_0)[name = tensor("scatter_3_cast_fp16")]; + tensor reshape_19_cast_fp16 = reshape(shape = shape_15, x = scatter_3_cast_fp16)[name = tensor("reshape_19_cast_fp16")]; + tensor reshape_19_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("reshape_19_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor keys_7_begin_0 = const()[name = tensor("keys_7_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor keys_7_end_0 = const()[name = tensor("keys_7_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor keys_7_end_mask_0 = const()[name = tensor("keys_7_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor keys_7_squeeze_mask_0 = const()[name = tensor("keys_7_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor keys_7_cast_fp16 = slice_by_index(begin = keys_7_begin_0, end = keys_7_end_0, end_mask = keys_7_end_mask_0, squeeze_mask = keys_7_squeeze_mask_0, x = reshape_19_cast_fp16)[name = tensor("keys_7_cast_fp16")]; + tensor values_7_begin_0 = const()[name = tensor("values_7_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor values_7_end_0 = const()[name = tensor("values_7_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor values_7_end_mask_0 = const()[name = tensor("values_7_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor values_7_squeeze_mask_0 = const()[name = tensor("values_7_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor values_7_cast_fp16 = slice_by_index(begin = values_7_begin_0, end = values_7_end_0, end_mask = values_7_end_mask_0, squeeze_mask = values_7_squeeze_mask_0, x = reshape_19_cast_fp16)[name = tensor("values_7_cast_fp16")]; + tensor var_699_cast_fp16 = not_equal(x = keys_7_cast_fp16, y = keys_7_cast_fp16)[name = tensor("op_699_cast_fp16")]; + tensor keys_9_cast_fp16 = select(a = var_319_to_fp16, b = keys_7_cast_fp16, cond = var_699_cast_fp16)[name = tensor("keys_9_cast_fp16")]; + tensor var_707_cast_fp16 = not_equal(x = values_7_cast_fp16, y = values_7_cast_fp16)[name = tensor("op_707_cast_fp16")]; + tensor values_9_cast_fp16 = select(a = var_319_to_fp16, b = values_7_cast_fp16, cond = var_707_cast_fp16)[name = tensor("values_9_cast_fp16")]; + tensor var_731 = const()[name = tensor("op_731"), val = tensor([0, 2, 1, 3])]; + tensor var_747 = const()[name = tensor("op_747"), val = tensor([1, 1, 1])]; + tensor var_748_cast_fp16 = reshape(shape = var_747, x = position1_to_fp16)[name = tensor("op_748_cast_fp16")]; + tensor var_768_to_fp16 = const()[name = tensor("op_768_to_fp16"), val = tensor(0x1p+0)]; + tensor valid_len_3_cast_fp16 = add(x = var_748_cast_fp16, y = var_768_to_fp16)[name = tensor("valid_len_3_cast_fp16")]; + tensor valid_mask_3_cast_fp16 = less(x = k_positions_1_promoted_to_fp16, y = valid_len_3_cast_fp16)[name = tensor("valid_mask_3_cast_fp16")]; + tensor causal_mask_3_cast_fp16 = less_equal(x = k_positions_1_promoted_to_fp16, y = var_748_cast_fp16)[name = tensor("causal_mask_3_cast_fp16")]; + tensor attn_mask_5 = logical_and(x = valid_mask_3_cast_fp16, y = causal_mask_3_cast_fp16)[name = tensor("attn_mask_5")]; + tensor attn_mask_7_axes_0 = const()[name = tensor("attn_mask_7_axes_0"), val = tensor([1])]; + tensor attn_mask_7 = expand_dims(axes = attn_mask_7_axes_0, x = attn_mask_5)[name = tensor("attn_mask_7")]; + tensor var_786_transpose_x_0 = const()[name = tensor("op_786_transpose_x_0"), val = tensor(false)]; + tensor var_786_transpose_y_0 = const()[name = tensor("op_786_transpose_y_0"), val = tensor(false)]; + tensor transpose_20_perm_0 = const()[name = tensor("transpose_20_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_21_perm_0 = const()[name = tensor("transpose_21_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_21 = transpose(perm = transpose_21_perm_0, x = keys_9_cast_fp16)[name = tensor("transpose_47")]; + tensor transpose_20 = transpose(perm = transpose_20_perm_0, x = q_9_cast_fp16)[name = tensor("transpose_48")]; + tensor var_786_cast_fp16 = matmul(transpose_x = var_786_transpose_x_0, transpose_y = var_786_transpose_y_0, x = transpose_20, y = transpose_21)[name = tensor("op_786_cast_fp16")]; + tensor var_780_to_fp16 = const()[name = tensor("op_780_to_fp16"), val = tensor([0x1p-3])]; + tensor attn_weights_7_cast_fp16 = mul(x = var_786_cast_fp16, y = var_780_to_fp16)[name = tensor("attn_weights_7_cast_fp16")]; + tensor var_788 = logical_not(x = attn_mask_7)[name = tensor("op_788")]; + tensor var_789_to_fp16 = const()[name = tensor("op_789_to_fp16"), val = tensor(-inf)]; + tensor attn_weights_9_cast_fp16 = select(a = var_789_to_fp16, b = attn_weights_7_cast_fp16, cond = var_788)[name = tensor("attn_weights_9_cast_fp16")]; + tensor var_791 = const()[name = tensor("op_791"), val = tensor(-1)]; + tensor attn_weights_11_cast_fp16 = softmax(axis = var_791, x = attn_weights_9_cast_fp16)[name = tensor("attn_weights_11_cast_fp16")]; + tensor attn_output_3_transpose_x_0 = const()[name = tensor("attn_output_3_transpose_x_0"), val = tensor(false)]; + tensor attn_output_3_transpose_y_0 = const()[name = tensor("attn_output_3_transpose_y_0"), val = tensor(false)]; + tensor values_11_cast_fp16 = transpose(perm = var_731, x = values_9_cast_fp16)[name = tensor("transpose_49")]; + tensor attn_output_3_cast_fp16 = matmul(transpose_x = attn_output_3_transpose_x_0, transpose_y = attn_output_3_transpose_y_0, x = attn_weights_11_cast_fp16, y = values_11_cast_fp16)[name = tensor("attn_output_3_cast_fp16")]; + tensor var_799 = const()[name = tensor("op_799"), val = tensor([0, 2, 1, 3])]; + tensor var_804 = const()[name = tensor("op_804"), val = tensor([1, 1, 1024])]; + tensor var_800_cast_fp16 = transpose(perm = var_799, x = attn_output_3_cast_fp16)[name = tensor("transpose_46")]; + tensor input_19_cast_fp16 = reshape(shape = var_804, x = var_800_cast_fp16)[name = tensor("input_19_cast_fp16")]; + tensor attn1_out_proj_weight_to_fp16 = const()[name = tensor("attn1_out_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(32602496)))]; + tensor linear_6_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = attn1_out_proj_weight_to_fp16, x = input_19_cast_fp16)[name = tensor("linear_6_cast_fp16")]; + tensor var_810_to_fp16 = const()[name = tensor("op_810_to_fp16"), val = tensor(0x1p+0)]; + tensor var_811_cast_fp16 = add(x = position1_to_fp16, y = var_810_to_fp16)[name = tensor("op_811_cast_fp16")]; + tensor var_811_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_811_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor input_21_cast_fp16 = add(x = input_15_cast_fp16, y = linear_6_cast_fp16)[name = tensor("input_21_cast_fp16")]; + tensor input_23_axes_0 = const()[name = tensor("input_23_axes_0"), val = tensor([-1])]; + tensor norm1_2_weight_to_fp16 = const()[name = tensor("norm1_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(34699712)))]; + tensor norm1_2_bias_to_fp16 = const()[name = tensor("norm1_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(34701824)))]; + tensor var_815_to_fp16 = const()[name = tensor("op_815_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_23_cast_fp16 = layer_norm(axes = input_23_axes_0, beta = norm1_2_bias_to_fp16, epsilon = var_815_to_fp16, gamma = norm1_2_weight_to_fp16, x = input_21_cast_fp16)[name = tensor("input_23_cast_fp16")]; + tensor linear1_1_weight_to_fp16 = const()[name = tensor("linear1_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(34703936)))]; + tensor linear_7_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = linear1_1_weight_to_fp16, x = input_23_cast_fp16)[name = tensor("linear_7_cast_fp16")]; + tensor input_25_mode_0 = const()[name = tensor("input_25_mode_0"), val = tensor("EXACT")]; + tensor input_25_cast_fp16 = gelu(mode = input_25_mode_0, x = linear_7_cast_fp16)[name = tensor("input_25_cast_fp16")]; + tensor linear1_2_weight_to_fp16 = const()[name = tensor("linear1_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(43092608)))]; + tensor linear_8_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = linear1_2_weight_to_fp16, x = input_25_cast_fp16)[name = tensor("linear_8_cast_fp16")]; + tensor input_27_cast_fp16 = add(x = input_21_cast_fp16, y = linear_8_cast_fp16)[name = tensor("input_27_cast_fp16")]; + tensor input_29_axes_0 = const()[name = tensor("input_29_axes_0"), val = tensor([-1])]; + tensor norm2_1_weight_to_fp16 = const()[name = tensor("norm2_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(51481280)))]; + tensor norm2_1_bias_to_fp16 = const()[name = tensor("norm2_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(51483392)))]; + tensor var_832_to_fp16 = const()[name = tensor("op_832_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_29_cast_fp16 = layer_norm(axes = input_29_axes_0, beta = norm2_1_bias_to_fp16, epsilon = var_832_to_fp16, gamma = norm2_1_weight_to_fp16, x = input_27_cast_fp16)[name = tensor("input_29_cast_fp16")]; + tensor attn2_in_proj_weight_to_fp16 = const()[name = tensor("attn2_in_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(51485504)))]; + tensor linear_9_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = attn2_in_proj_weight_to_fp16, x = input_29_cast_fp16)[name = tensor("linear_9_cast_fp16")]; + tensor var_846 = const()[name = tensor("op_846"), val = tensor([1, 1, 3, 16, 64])]; + tensor qkv_5_cast_fp16 = reshape(shape = var_846, x = linear_9_cast_fp16)[name = tensor("qkv_5_cast_fp16")]; + tensor q_13_begin_0 = const()[name = tensor("q_13_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor q_13_end_0 = const()[name = tensor("q_13_end_0"), val = tensor([1, 1, 1, 16, 64])]; + tensor q_13_end_mask_0 = const()[name = tensor("q_13_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor q_13_squeeze_mask_0 = const()[name = tensor("q_13_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor q_13_cast_fp16 = slice_by_index(begin = q_13_begin_0, end = q_13_end_0, end_mask = q_13_end_mask_0, squeeze_mask = q_13_squeeze_mask_0, x = qkv_5_cast_fp16)[name = tensor("q_13_cast_fp16")]; + tensor k_9_begin_0 = const()[name = tensor("k_9_begin_0"), val = tensor([0, 0, 1, 0, 0])]; + tensor k_9_end_0 = const()[name = tensor("k_9_end_0"), val = tensor([1, 1, 2, 16, 64])]; + tensor k_9_end_mask_0 = const()[name = tensor("k_9_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor k_9_squeeze_mask_0 = const()[name = tensor("k_9_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor k_9_cast_fp16 = slice_by_index(begin = k_9_begin_0, end = k_9_end_0, end_mask = k_9_end_mask_0, squeeze_mask = k_9_squeeze_mask_0, x = qkv_5_cast_fp16)[name = tensor("k_9_cast_fp16")]; + tensor v_5_begin_0 = const()[name = tensor("v_5_begin_0"), val = tensor([0, 0, 2, 0, 0])]; + tensor v_5_end_0 = const()[name = tensor("v_5_end_0"), val = tensor([1, 1, 3, 16, 64])]; + tensor v_5_end_mask_0 = const()[name = tensor("v_5_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor v_5_squeeze_mask_0 = const()[name = tensor("v_5_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor v_5_cast_fp16 = slice_by_index(begin = v_5_begin_0, end = v_5_end_0, end_mask = v_5_end_mask_0, squeeze_mask = v_5_squeeze_mask_0, x = qkv_5_cast_fp16)[name = tensor("v_5_cast_fp16")]; + tensor var_912 = const()[name = tensor("op_912"), val = tensor([1, 1, 1, 1])]; + tensor position2_to_fp16_dtype_0 = const()[name = tensor("position2_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor position2_to_fp16 = cast(dtype = position2_to_fp16_dtype_0, x = position2)[name = tensor("cast_129")]; + tensor ts_17_cast_fp16 = reshape(shape = var_912, x = position2_to_fp16)[name = tensor("ts_17_cast_fp16")]; + tensor var_919 = const()[name = tensor("op_919"), val = tensor([1, 1, 16, 32, 2])]; + tensor q_complex_5_cast_fp16 = reshape(shape = var_919, x = q_13_cast_fp16)[name = tensor("q_complex_5_cast_fp16")]; + tensor var_926 = const()[name = tensor("op_926"), val = tensor([1, 1, 16, 32, 2])]; + tensor k_complex_5_cast_fp16 = reshape(shape = var_926, x = k_9_cast_fp16)[name = tensor("k_complex_5_cast_fp16")]; + tensor var_930_begin_0 = const()[name = tensor("op_930_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_930_end_0 = const()[name = tensor("op_930_end_0"), val = tensor([1, 1, 16, 32, 1])]; + tensor var_930_end_mask_0 = const()[name = tensor("op_930_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_930_squeeze_mask_0 = const()[name = tensor("op_930_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_930_cast_fp16 = slice_by_index(begin = var_930_begin_0, end = var_930_end_0, end_mask = var_930_end_mask_0, squeeze_mask = var_930_squeeze_mask_0, x = q_complex_5_cast_fp16)[name = tensor("op_930_cast_fp16")]; + tensor var_938_begin_0 = const()[name = tensor("op_938_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_938_end_0 = const()[name = tensor("op_938_end_0"), val = tensor([1, 1, 16, 32, 2])]; + tensor var_938_end_mask_0 = const()[name = tensor("op_938_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_938_squeeze_mask_0 = const()[name = tensor("op_938_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_938_cast_fp16 = slice_by_index(begin = var_938_begin_0, end = var_938_end_0, end_mask = var_938_end_mask_0, squeeze_mask = var_938_squeeze_mask_0, x = q_complex_5_cast_fp16)[name = tensor("op_938_cast_fp16")]; + tensor var_946_begin_0 = const()[name = tensor("op_946_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_946_end_0 = const()[name = tensor("op_946_end_0"), val = tensor([1, 1, 16, 32, 1])]; + tensor var_946_end_mask_0 = const()[name = tensor("op_946_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_946_squeeze_mask_0 = const()[name = tensor("op_946_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_946_cast_fp16 = slice_by_index(begin = var_946_begin_0, end = var_946_end_0, end_mask = var_946_end_mask_0, squeeze_mask = var_946_squeeze_mask_0, x = k_complex_5_cast_fp16)[name = tensor("op_946_cast_fp16")]; + tensor var_954_begin_0 = const()[name = tensor("op_954_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_954_end_0 = const()[name = tensor("op_954_end_0"), val = tensor([1, 1, 16, 32, 2])]; + tensor var_954_end_mask_0 = const()[name = tensor("op_954_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_954_squeeze_mask_0 = const()[name = tensor("op_954_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_954_cast_fp16 = slice_by_index(begin = var_954_begin_0, end = var_954_end_0, end_mask = var_954_end_mask_0, squeeze_mask = var_954_squeeze_mask_0, x = k_complex_5_cast_fp16)[name = tensor("op_954_cast_fp16")]; + tensor freqs_5_to_fp16 = const()[name = tensor("freqs_5_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(57777024)))]; + tensor var_960_cast_fp16 = mul(x = freqs_5_to_fp16, y = ts_17_cast_fp16)[name = tensor("op_960_cast_fp16")]; + tensor rotr_5_cast_fp16 = cos(x = var_960_cast_fp16)[name = tensor("rotr_5_cast_fp16")]; + tensor roti_5_cast_fp16 = sin(x = var_960_cast_fp16)[name = tensor("roti_5_cast_fp16")]; + tensor var_964_cast_fp16 = mul(x = var_930_cast_fp16, y = rotr_5_cast_fp16)[name = tensor("op_964_cast_fp16")]; + tensor var_965_cast_fp16 = mul(x = var_938_cast_fp16, y = roti_5_cast_fp16)[name = tensor("op_965_cast_fp16")]; + tensor qor_9_cast_fp16 = sub(x = var_964_cast_fp16, y = var_965_cast_fp16)[name = tensor("qor_9_cast_fp16")]; + tensor var_968_cast_fp16 = mul(x = var_930_cast_fp16, y = roti_5_cast_fp16)[name = tensor("op_968_cast_fp16")]; + tensor var_969_cast_fp16 = mul(x = var_938_cast_fp16, y = rotr_5_cast_fp16)[name = tensor("op_969_cast_fp16")]; + tensor qoi_9_cast_fp16 = add(x = var_968_cast_fp16, y = var_969_cast_fp16)[name = tensor("qoi_9_cast_fp16")]; + tensor var_972_cast_fp16 = mul(x = var_946_cast_fp16, y = rotr_5_cast_fp16)[name = tensor("op_972_cast_fp16")]; + tensor var_973_cast_fp16 = mul(x = var_954_cast_fp16, y = roti_5_cast_fp16)[name = tensor("op_973_cast_fp16")]; + tensor kor_9_cast_fp16 = sub(x = var_972_cast_fp16, y = var_973_cast_fp16)[name = tensor("kor_9_cast_fp16")]; + tensor var_976_cast_fp16 = mul(x = var_946_cast_fp16, y = roti_5_cast_fp16)[name = tensor("op_976_cast_fp16")]; + tensor var_977_cast_fp16 = mul(x = var_954_cast_fp16, y = rotr_5_cast_fp16)[name = tensor("op_977_cast_fp16")]; + tensor koi_9_cast_fp16 = add(x = var_976_cast_fp16, y = var_977_cast_fp16)[name = tensor("koi_9_cast_fp16")]; + tensor qo_5_axis_0 = const()[name = tensor("qo_5_axis_0"), val = tensor(-1)]; + tensor qo_5_cast_fp16 = stack(axis = qo_5_axis_0, values = (qor_9_cast_fp16, qoi_9_cast_fp16))[name = tensor("qo_5_cast_fp16")]; + tensor ko_5_axis_0 = const()[name = tensor("ko_5_axis_0"), val = tensor(-1)]; + tensor ko_5_cast_fp16 = stack(axis = ko_5_axis_0, values = (kor_9_cast_fp16, koi_9_cast_fp16))[name = tensor("ko_5_cast_fp16")]; + tensor var_1010 = const()[name = tensor("op_1010"), val = tensor([1, 1, 16, 64])]; + tensor q_15_cast_fp16 = reshape(shape = var_1010, x = qo_5_cast_fp16)[name = tensor("q_15_cast_fp16")]; + tensor var_1016 = const()[name = tensor("op_1016"), val = tensor([1, 1, 16, 64])]; + tensor k_11_cast_fp16 = reshape(shape = var_1016, x = ko_5_cast_fp16)[name = tensor("k_11_cast_fp16")]; + tensor _inversed_1041_y_0_to_fp16 = const()[name = tensor("_inversed_1041_y_0_to_fp16"), val = tensor(0x1p-9)]; + tensor _inversed_1041_cast_fp16 = mul(x = ts_17_cast_fp16, y = _inversed_1041_y_0_to_fp16)[name = tensor("_inversed_1041_cast_fp16")]; + tensor var_1042_cast_fp16 = floor(x = _inversed_1041_cast_fp16)[name = tensor("op_1042_cast_fp16")]; + tensor var_1043_to_fp16 = const()[name = tensor("op_1043_to_fp16"), val = tensor(0x1p+9)]; + tensor var_1044_cast_fp16 = mul(x = var_1042_cast_fp16, y = var_1043_to_fp16)[name = tensor("op_1044_cast_fp16")]; + tensor write_indices_float_11_cast_fp16 = sub(x = ts_17_cast_fp16, y = var_1044_cast_fp16)[name = tensor("write_indices_float_11_cast_fp16")]; + tensor var_1051_dtype_0 = const()[name = tensor("op_1051_dtype_0"), val = tensor("int32")]; + tensor write_indices_5_reps_0 = const()[name = tensor("write_indices_5_reps_0"), val = tensor([1, 1, 16, 64])]; + tensor write_indices_float_11_cast_fp16_to_int32 = cast(dtype = var_1051_dtype_0, x = write_indices_float_11_cast_fp16)[name = tensor("cast_128")]; + tensor write_indices_5 = tile(reps = write_indices_5_reps_0, x = write_indices_float_11_cast_fp16_to_int32)[name = tensor("write_indices_5")]; + tensor var_1061_begin_0 = const()[name = tensor("op_1061_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_1061_end_0 = const()[name = tensor("op_1061_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor var_1061_end_mask_0 = const()[name = tensor("op_1061_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_1061_squeeze_mask_0 = const()[name = tensor("op_1061_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor cache2_to_fp16_dtype_0 = const()[name = tensor("cache2_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor cache2_to_fp16 = cast(dtype = cache2_to_fp16_dtype_0, x = cache2)[name = tensor("cast_127")]; + tensor var_1061_cast_fp16 = slice_by_index(begin = var_1061_begin_0, end = var_1061_end_0, end_mask = var_1061_end_mask_0, squeeze_mask = var_1061_squeeze_mask_0, x = cache2_to_fp16)[name = tensor("op_1061_cast_fp16")]; + tensor var_1063_axis_0 = const()[name = tensor("op_1063_axis_0"), val = tensor(1)]; + tensor var_1063_mode_0 = const()[name = tensor("op_1063_mode_0"), val = tensor("update")]; + tensor var_1063_validate_indices_0 = const()[name = tensor("op_1063_validate_indices_0"), val = tensor(false)]; + tensor var_1063_cast_fp16 = scatter_along_axis(axis = var_1063_axis_0, data = var_1061_cast_fp16, indices = write_indices_5, mode = var_1063_mode_0, updates = k_11_cast_fp16, validate_indices = var_1063_validate_indices_0)[name = tensor("op_1063_cast_fp16")]; + tensor concat_16 = const()[name = tensor("concat_16"), val = tensor([0, 0, 0, 0, 0])]; + tensor concat_17 = const()[name = tensor("concat_17"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_5_internal_tensor_assign_1_stride_0 = const()[name = tensor("new_cache_5_internal_tensor_assign_1_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_5_internal_tensor_assign_1_begin_mask_0 = const()[name = tensor("new_cache_5_internal_tensor_assign_1_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_5_internal_tensor_assign_1_end_mask_0 = const()[name = tensor("new_cache_5_internal_tensor_assign_1_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_5_internal_tensor_assign_1_squeeze_mask_0 = const()[name = tensor("new_cache_5_internal_tensor_assign_1_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_16 = const()[name = tensor("shape_16"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_4 = const()[name = tensor("reduce_prod_4"), val = tensor(1048576)]; + tensor range_1d_4_start_0 = const()[name = tensor("range_1d_4_start_0"), val = tensor(0)]; + tensor range_1d_4_step_0 = const()[name = tensor("range_1d_4_step_0"), val = tensor(1)]; + tensor range_1d_4 = range_1d(end = reduce_prod_4, start = range_1d_4_start_0, step = range_1d_4_step_0)[name = tensor("range_1d_4")]; + tensor reshape_20 = reshape(shape = shape_16, x = range_1d_4)[name = tensor("reshape_20")]; + tensor slice_by_index_4 = slice_by_index(begin = concat_16, begin_mask = new_cache_5_internal_tensor_assign_1_begin_mask_0, end = concat_17, end_mask = new_cache_5_internal_tensor_assign_1_end_mask_0, squeeze_mask = new_cache_5_internal_tensor_assign_1_squeeze_mask_0, stride = new_cache_5_internal_tensor_assign_1_stride_0, x = reshape_20)[name = tensor("slice_by_index_4")]; + tensor reshape_21_shape_0 = const()[name = tensor("reshape_21_shape_0"), val = tensor([-1])]; + tensor reshape_21 = reshape(shape = reshape_21_shape_0, x = slice_by_index_4)[name = tensor("reshape_21")]; + tensor reshape_22_shape_0 = const()[name = tensor("reshape_22_shape_0"), val = tensor([-1])]; + tensor reshape_22_cast_fp16 = reshape(shape = reshape_22_shape_0, x = var_1063_cast_fp16)[name = tensor("reshape_22_cast_fp16")]; + tensor reshape_23_shape_0 = const()[name = tensor("reshape_23_shape_0"), val = tensor([-1])]; + tensor reshape_23_cast_fp16 = reshape(shape = reshape_23_shape_0, x = cache2_to_fp16)[name = tensor("reshape_23_cast_fp16")]; + tensor scatter_4_mode_0 = const()[name = tensor("scatter_4_mode_0"), val = tensor("update")]; + tensor scatter_4_axis_0 = const()[name = tensor("scatter_4_axis_0"), val = tensor(0)]; + tensor scatter_4_validate_indices_0 = const()[name = tensor("scatter_4_validate_indices_0"), val = tensor(false)]; + tensor scatter_4_cast_fp16 = scatter(axis = scatter_4_axis_0, data = reshape_23_cast_fp16, indices = reshape_21, mode = scatter_4_mode_0, updates = reshape_22_cast_fp16, validate_indices = scatter_4_validate_indices_0)[name = tensor("scatter_4_cast_fp16")]; + tensor reshape_24_cast_fp16 = reshape(shape = shape_16, x = scatter_4_cast_fp16)[name = tensor("reshape_24_cast_fp16")]; + tensor var_1071_begin_0 = const()[name = tensor("op_1071_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor var_1071_end_0 = const()[name = tensor("op_1071_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor var_1071_end_mask_0 = const()[name = tensor("op_1071_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_1071_squeeze_mask_0 = const()[name = tensor("op_1071_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor var_1071_cast_fp16 = slice_by_index(begin = var_1071_begin_0, end = var_1071_end_0, end_mask = var_1071_end_mask_0, squeeze_mask = var_1071_squeeze_mask_0, x = reshape_24_cast_fp16)[name = tensor("op_1071_cast_fp16")]; + tensor var_1073_axis_0 = const()[name = tensor("op_1073_axis_0"), val = tensor(1)]; + tensor var_1073_mode_0 = const()[name = tensor("op_1073_mode_0"), val = tensor("update")]; + tensor var_1073_validate_indices_0 = const()[name = tensor("op_1073_validate_indices_0"), val = tensor(false)]; + tensor var_1073_cast_fp16 = scatter_along_axis(axis = var_1073_axis_0, data = var_1071_cast_fp16, indices = write_indices_5, mode = var_1073_mode_0, updates = v_5_cast_fp16, validate_indices = var_1073_validate_indices_0)[name = tensor("op_1073_cast_fp16")]; + tensor concat_18 = const()[name = tensor("concat_18"), val = tensor([1, 0, 0, 0, 0])]; + tensor concat_19 = const()[name = tensor("concat_19"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_5_internal_tensor_assign_2_stride_0 = const()[name = tensor("new_cache_5_internal_tensor_assign_2_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_5_internal_tensor_assign_2_begin_mask_0 = const()[name = tensor("new_cache_5_internal_tensor_assign_2_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_5_internal_tensor_assign_2_end_mask_0 = const()[name = tensor("new_cache_5_internal_tensor_assign_2_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_5_internal_tensor_assign_2_squeeze_mask_0 = const()[name = tensor("new_cache_5_internal_tensor_assign_2_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_17 = const()[name = tensor("shape_17"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_5 = const()[name = tensor("reduce_prod_5"), val = tensor(1048576)]; + tensor range_1d_5_start_0 = const()[name = tensor("range_1d_5_start_0"), val = tensor(0)]; + tensor range_1d_5_step_0 = const()[name = tensor("range_1d_5_step_0"), val = tensor(1)]; + tensor range_1d_5 = range_1d(end = reduce_prod_5, start = range_1d_5_start_0, step = range_1d_5_step_0)[name = tensor("range_1d_5")]; + tensor reshape_25 = reshape(shape = shape_17, x = range_1d_5)[name = tensor("reshape_25")]; + tensor slice_by_index_5 = slice_by_index(begin = concat_18, begin_mask = new_cache_5_internal_tensor_assign_2_begin_mask_0, end = concat_19, end_mask = new_cache_5_internal_tensor_assign_2_end_mask_0, squeeze_mask = new_cache_5_internal_tensor_assign_2_squeeze_mask_0, stride = new_cache_5_internal_tensor_assign_2_stride_0, x = reshape_25)[name = tensor("slice_by_index_5")]; + tensor reshape_26_shape_0 = const()[name = tensor("reshape_26_shape_0"), val = tensor([-1])]; + tensor reshape_26 = reshape(shape = reshape_26_shape_0, x = slice_by_index_5)[name = tensor("reshape_26")]; + tensor reshape_27_shape_0 = const()[name = tensor("reshape_27_shape_0"), val = tensor([-1])]; + tensor reshape_27_cast_fp16 = reshape(shape = reshape_27_shape_0, x = var_1073_cast_fp16)[name = tensor("reshape_27_cast_fp16")]; + tensor reshape_28_shape_0 = const()[name = tensor("reshape_28_shape_0"), val = tensor([-1])]; + tensor reshape_28_cast_fp16 = reshape(shape = reshape_28_shape_0, x = reshape_24_cast_fp16)[name = tensor("reshape_28_cast_fp16")]; + tensor scatter_5_mode_0 = const()[name = tensor("scatter_5_mode_0"), val = tensor("update")]; + tensor scatter_5_axis_0 = const()[name = tensor("scatter_5_axis_0"), val = tensor(0)]; + tensor scatter_5_validate_indices_0 = const()[name = tensor("scatter_5_validate_indices_0"), val = tensor(false)]; + tensor scatter_5_cast_fp16 = scatter(axis = scatter_5_axis_0, data = reshape_28_cast_fp16, indices = reshape_26, mode = scatter_5_mode_0, updates = reshape_27_cast_fp16, validate_indices = scatter_5_validate_indices_0)[name = tensor("scatter_5_cast_fp16")]; + tensor reshape_29_cast_fp16 = reshape(shape = shape_17, x = scatter_5_cast_fp16)[name = tensor("reshape_29_cast_fp16")]; + tensor reshape_29_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("reshape_29_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor keys_13_begin_0 = const()[name = tensor("keys_13_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor keys_13_end_0 = const()[name = tensor("keys_13_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor keys_13_end_mask_0 = const()[name = tensor("keys_13_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor keys_13_squeeze_mask_0 = const()[name = tensor("keys_13_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor keys_13_cast_fp16 = slice_by_index(begin = keys_13_begin_0, end = keys_13_end_0, end_mask = keys_13_end_mask_0, squeeze_mask = keys_13_squeeze_mask_0, x = reshape_29_cast_fp16)[name = tensor("keys_13_cast_fp16")]; + tensor values_13_begin_0 = const()[name = tensor("values_13_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor values_13_end_0 = const()[name = tensor("values_13_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor values_13_end_mask_0 = const()[name = tensor("values_13_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor values_13_squeeze_mask_0 = const()[name = tensor("values_13_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor values_13_cast_fp16 = slice_by_index(begin = values_13_begin_0, end = values_13_end_0, end_mask = values_13_end_mask_0, squeeze_mask = values_13_squeeze_mask_0, x = reshape_29_cast_fp16)[name = tensor("values_13_cast_fp16")]; + tensor var_1085_cast_fp16 = not_equal(x = keys_13_cast_fp16, y = keys_13_cast_fp16)[name = tensor("op_1085_cast_fp16")]; + tensor keys_15_cast_fp16 = select(a = var_319_to_fp16, b = keys_13_cast_fp16, cond = var_1085_cast_fp16)[name = tensor("keys_15_cast_fp16")]; + tensor var_1093_cast_fp16 = not_equal(x = values_13_cast_fp16, y = values_13_cast_fp16)[name = tensor("op_1093_cast_fp16")]; + tensor values_15_cast_fp16 = select(a = var_319_to_fp16, b = values_13_cast_fp16, cond = var_1093_cast_fp16)[name = tensor("values_15_cast_fp16")]; + tensor var_1117 = const()[name = tensor("op_1117"), val = tensor([0, 2, 1, 3])]; + tensor var_1133 = const()[name = tensor("op_1133"), val = tensor([1, 1, 1])]; + tensor var_1134_cast_fp16 = reshape(shape = var_1133, x = position2_to_fp16)[name = tensor("op_1134_cast_fp16")]; + tensor var_1154_to_fp16 = const()[name = tensor("op_1154_to_fp16"), val = tensor(0x1p+0)]; + tensor valid_len_5_cast_fp16 = add(x = var_1134_cast_fp16, y = var_1154_to_fp16)[name = tensor("valid_len_5_cast_fp16")]; + tensor valid_mask_5_cast_fp16 = less(x = k_positions_1_promoted_to_fp16, y = valid_len_5_cast_fp16)[name = tensor("valid_mask_5_cast_fp16")]; + tensor causal_mask_5_cast_fp16 = less_equal(x = k_positions_1_promoted_to_fp16, y = var_1134_cast_fp16)[name = tensor("causal_mask_5_cast_fp16")]; + tensor attn_mask_9 = logical_and(x = valid_mask_5_cast_fp16, y = causal_mask_5_cast_fp16)[name = tensor("attn_mask_9")]; + tensor attn_mask_11_axes_0 = const()[name = tensor("attn_mask_11_axes_0"), val = tensor([1])]; + tensor attn_mask_11 = expand_dims(axes = attn_mask_11_axes_0, x = attn_mask_9)[name = tensor("attn_mask_11")]; + tensor var_1172_transpose_x_0 = const()[name = tensor("op_1172_transpose_x_0"), val = tensor(false)]; + tensor var_1172_transpose_y_0 = const()[name = tensor("op_1172_transpose_y_0"), val = tensor(false)]; + tensor transpose_22_perm_0 = const()[name = tensor("transpose_22_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_23_perm_0 = const()[name = tensor("transpose_23_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_23 = transpose(perm = transpose_23_perm_0, x = keys_15_cast_fp16)[name = tensor("transpose_43")]; + tensor transpose_22 = transpose(perm = transpose_22_perm_0, x = q_15_cast_fp16)[name = tensor("transpose_44")]; + tensor var_1172_cast_fp16 = matmul(transpose_x = var_1172_transpose_x_0, transpose_y = var_1172_transpose_y_0, x = transpose_22, y = transpose_23)[name = tensor("op_1172_cast_fp16")]; + tensor var_1166_to_fp16 = const()[name = tensor("op_1166_to_fp16"), val = tensor([0x1p-3])]; + tensor attn_weights_13_cast_fp16 = mul(x = var_1172_cast_fp16, y = var_1166_to_fp16)[name = tensor("attn_weights_13_cast_fp16")]; + tensor var_1174 = logical_not(x = attn_mask_11)[name = tensor("op_1174")]; + tensor var_1175_to_fp16 = const()[name = tensor("op_1175_to_fp16"), val = tensor(-inf)]; + tensor attn_weights_15_cast_fp16 = select(a = var_1175_to_fp16, b = attn_weights_13_cast_fp16, cond = var_1174)[name = tensor("attn_weights_15_cast_fp16")]; + tensor var_1177 = const()[name = tensor("op_1177"), val = tensor(-1)]; + tensor attn_weights_17_cast_fp16 = softmax(axis = var_1177, x = attn_weights_15_cast_fp16)[name = tensor("attn_weights_17_cast_fp16")]; + tensor attn_output_5_transpose_x_0 = const()[name = tensor("attn_output_5_transpose_x_0"), val = tensor(false)]; + tensor attn_output_5_transpose_y_0 = const()[name = tensor("attn_output_5_transpose_y_0"), val = tensor(false)]; + tensor values_17_cast_fp16 = transpose(perm = var_1117, x = values_15_cast_fp16)[name = tensor("transpose_45")]; + tensor attn_output_5_cast_fp16 = matmul(transpose_x = attn_output_5_transpose_x_0, transpose_y = attn_output_5_transpose_y_0, x = attn_weights_17_cast_fp16, y = values_17_cast_fp16)[name = tensor("attn_output_5_cast_fp16")]; + tensor var_1185 = const()[name = tensor("op_1185"), val = tensor([0, 2, 1, 3])]; + tensor var_1190 = const()[name = tensor("op_1190"), val = tensor([1, 1, 1024])]; + tensor var_1186_cast_fp16 = transpose(perm = var_1185, x = attn_output_5_cast_fp16)[name = tensor("transpose_42")]; + tensor input_31_cast_fp16 = reshape(shape = var_1190, x = var_1186_cast_fp16)[name = tensor("input_31_cast_fp16")]; + tensor attn2_out_proj_weight_to_fp16 = const()[name = tensor("attn2_out_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(57777152)))]; + tensor linear_10_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = attn2_out_proj_weight_to_fp16, x = input_31_cast_fp16)[name = tensor("linear_10_cast_fp16")]; + tensor var_1196_to_fp16 = const()[name = tensor("op_1196_to_fp16"), val = tensor(0x1p+0)]; + tensor var_1197_cast_fp16 = add(x = position2_to_fp16, y = var_1196_to_fp16)[name = tensor("op_1197_cast_fp16")]; + tensor var_1197_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_1197_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor input_33_cast_fp16 = add(x = input_27_cast_fp16, y = linear_10_cast_fp16)[name = tensor("input_33_cast_fp16")]; + tensor input_35_axes_0 = const()[name = tensor("input_35_axes_0"), val = tensor([-1])]; + tensor norm2_2_weight_to_fp16 = const()[name = tensor("norm2_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(59874368)))]; + tensor norm2_2_bias_to_fp16 = const()[name = tensor("norm2_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(59876480)))]; + tensor var_1201_to_fp16 = const()[name = tensor("op_1201_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_35_cast_fp16 = layer_norm(axes = input_35_axes_0, beta = norm2_2_bias_to_fp16, epsilon = var_1201_to_fp16, gamma = norm2_2_weight_to_fp16, x = input_33_cast_fp16)[name = tensor("input_35_cast_fp16")]; + tensor linear2_1_weight_to_fp16 = const()[name = tensor("linear2_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(59878592)))]; + tensor linear_11_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = linear2_1_weight_to_fp16, x = input_35_cast_fp16)[name = tensor("linear_11_cast_fp16")]; + tensor input_37_mode_0 = const()[name = tensor("input_37_mode_0"), val = tensor("EXACT")]; + tensor input_37_cast_fp16 = gelu(mode = input_37_mode_0, x = linear_11_cast_fp16)[name = tensor("input_37_cast_fp16")]; + tensor linear2_2_weight_to_fp16 = const()[name = tensor("linear2_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(68267264)))]; + tensor linear_12_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = linear2_2_weight_to_fp16, x = input_37_cast_fp16)[name = tensor("linear_12_cast_fp16")]; + tensor input_39_cast_fp16 = add(x = input_33_cast_fp16, y = linear_12_cast_fp16)[name = tensor("input_39_cast_fp16")]; + tensor input_41_axes_0 = const()[name = tensor("input_41_axes_0"), val = tensor([-1])]; + tensor norm3_1_weight_to_fp16 = const()[name = tensor("norm3_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(76655936)))]; + tensor norm3_1_bias_to_fp16 = const()[name = tensor("norm3_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(76658048)))]; + tensor var_1218_to_fp16 = const()[name = tensor("op_1218_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_41_cast_fp16 = layer_norm(axes = input_41_axes_0, beta = norm3_1_bias_to_fp16, epsilon = var_1218_to_fp16, gamma = norm3_1_weight_to_fp16, x = input_39_cast_fp16)[name = tensor("input_41_cast_fp16")]; + tensor attn3_in_proj_weight_to_fp16 = const()[name = tensor("attn3_in_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(76660160)))]; + tensor linear_13_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = attn3_in_proj_weight_to_fp16, x = input_41_cast_fp16)[name = tensor("linear_13_cast_fp16")]; + tensor var_1232 = const()[name = tensor("op_1232"), val = tensor([1, 1, 3, 16, 64])]; + tensor qkv_7_cast_fp16 = reshape(shape = var_1232, x = linear_13_cast_fp16)[name = tensor("qkv_7_cast_fp16")]; + tensor q_19_begin_0 = const()[name = tensor("q_19_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor q_19_end_0 = const()[name = tensor("q_19_end_0"), val = tensor([1, 1, 1, 16, 64])]; + tensor q_19_end_mask_0 = const()[name = tensor("q_19_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor q_19_squeeze_mask_0 = const()[name = tensor("q_19_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor q_19_cast_fp16 = slice_by_index(begin = q_19_begin_0, end = q_19_end_0, end_mask = q_19_end_mask_0, squeeze_mask = q_19_squeeze_mask_0, x = qkv_7_cast_fp16)[name = tensor("q_19_cast_fp16")]; + tensor k_13_begin_0 = const()[name = tensor("k_13_begin_0"), val = tensor([0, 0, 1, 0, 0])]; + tensor k_13_end_0 = const()[name = tensor("k_13_end_0"), val = tensor([1, 1, 2, 16, 64])]; + tensor k_13_end_mask_0 = const()[name = tensor("k_13_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor k_13_squeeze_mask_0 = const()[name = tensor("k_13_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor k_13_cast_fp16 = slice_by_index(begin = k_13_begin_0, end = k_13_end_0, end_mask = k_13_end_mask_0, squeeze_mask = k_13_squeeze_mask_0, x = qkv_7_cast_fp16)[name = tensor("k_13_cast_fp16")]; + tensor v_7_begin_0 = const()[name = tensor("v_7_begin_0"), val = tensor([0, 0, 2, 0, 0])]; + tensor v_7_end_0 = const()[name = tensor("v_7_end_0"), val = tensor([1, 1, 3, 16, 64])]; + tensor v_7_end_mask_0 = const()[name = tensor("v_7_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor v_7_squeeze_mask_0 = const()[name = tensor("v_7_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor v_7_cast_fp16 = slice_by_index(begin = v_7_begin_0, end = v_7_end_0, end_mask = v_7_end_mask_0, squeeze_mask = v_7_squeeze_mask_0, x = qkv_7_cast_fp16)[name = tensor("v_7_cast_fp16")]; + tensor var_1298 = const()[name = tensor("op_1298"), val = tensor([1, 1, 1, 1])]; + tensor position3_to_fp16_dtype_0 = const()[name = tensor("position3_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor position3_to_fp16 = cast(dtype = position3_to_fp16_dtype_0, x = position3)[name = tensor("cast_124")]; + tensor ts_23_cast_fp16 = reshape(shape = var_1298, x = position3_to_fp16)[name = tensor("ts_23_cast_fp16")]; + tensor var_1305 = const()[name = tensor("op_1305"), val = tensor([1, 1, 16, 32, 2])]; + tensor q_complex_7_cast_fp16 = reshape(shape = var_1305, x = q_19_cast_fp16)[name = tensor("q_complex_7_cast_fp16")]; + tensor var_1312 = const()[name = tensor("op_1312"), val = tensor([1, 1, 16, 32, 2])]; + tensor k_complex_7_cast_fp16 = reshape(shape = var_1312, x = k_13_cast_fp16)[name = tensor("k_complex_7_cast_fp16")]; + tensor var_1316_begin_0 = const()[name = tensor("op_1316_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_1316_end_0 = const()[name = tensor("op_1316_end_0"), val = tensor([1, 1, 16, 32, 1])]; + tensor var_1316_end_mask_0 = const()[name = tensor("op_1316_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_1316_squeeze_mask_0 = const()[name = tensor("op_1316_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_1316_cast_fp16 = slice_by_index(begin = var_1316_begin_0, end = var_1316_end_0, end_mask = var_1316_end_mask_0, squeeze_mask = var_1316_squeeze_mask_0, x = q_complex_7_cast_fp16)[name = tensor("op_1316_cast_fp16")]; + tensor var_1324_begin_0 = const()[name = tensor("op_1324_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_1324_end_0 = const()[name = tensor("op_1324_end_0"), val = tensor([1, 1, 16, 32, 2])]; + tensor var_1324_end_mask_0 = const()[name = tensor("op_1324_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_1324_squeeze_mask_0 = const()[name = tensor("op_1324_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_1324_cast_fp16 = slice_by_index(begin = var_1324_begin_0, end = var_1324_end_0, end_mask = var_1324_end_mask_0, squeeze_mask = var_1324_squeeze_mask_0, x = q_complex_7_cast_fp16)[name = tensor("op_1324_cast_fp16")]; + tensor var_1332_begin_0 = const()[name = tensor("op_1332_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_1332_end_0 = const()[name = tensor("op_1332_end_0"), val = tensor([1, 1, 16, 32, 1])]; + tensor var_1332_end_mask_0 = const()[name = tensor("op_1332_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_1332_squeeze_mask_0 = const()[name = tensor("op_1332_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_1332_cast_fp16 = slice_by_index(begin = var_1332_begin_0, end = var_1332_end_0, end_mask = var_1332_end_mask_0, squeeze_mask = var_1332_squeeze_mask_0, x = k_complex_7_cast_fp16)[name = tensor("op_1332_cast_fp16")]; + tensor var_1340_begin_0 = const()[name = tensor("op_1340_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_1340_end_0 = const()[name = tensor("op_1340_end_0"), val = tensor([1, 1, 16, 32, 2])]; + tensor var_1340_end_mask_0 = const()[name = tensor("op_1340_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_1340_squeeze_mask_0 = const()[name = tensor("op_1340_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_1340_cast_fp16 = slice_by_index(begin = var_1340_begin_0, end = var_1340_end_0, end_mask = var_1340_end_mask_0, squeeze_mask = var_1340_squeeze_mask_0, x = k_complex_7_cast_fp16)[name = tensor("op_1340_cast_fp16")]; + tensor freqs_7_to_fp16 = const()[name = tensor("freqs_7_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(82951680)))]; + tensor var_1346_cast_fp16 = mul(x = freqs_7_to_fp16, y = ts_23_cast_fp16)[name = tensor("op_1346_cast_fp16")]; + tensor rotr_7_cast_fp16 = cos(x = var_1346_cast_fp16)[name = tensor("rotr_7_cast_fp16")]; + tensor roti_7_cast_fp16 = sin(x = var_1346_cast_fp16)[name = tensor("roti_7_cast_fp16")]; + tensor var_1350_cast_fp16 = mul(x = var_1316_cast_fp16, y = rotr_7_cast_fp16)[name = tensor("op_1350_cast_fp16")]; + tensor var_1351_cast_fp16 = mul(x = var_1324_cast_fp16, y = roti_7_cast_fp16)[name = tensor("op_1351_cast_fp16")]; + tensor qor_13_cast_fp16 = sub(x = var_1350_cast_fp16, y = var_1351_cast_fp16)[name = tensor("qor_13_cast_fp16")]; + tensor var_1354_cast_fp16 = mul(x = var_1316_cast_fp16, y = roti_7_cast_fp16)[name = tensor("op_1354_cast_fp16")]; + tensor var_1355_cast_fp16 = mul(x = var_1324_cast_fp16, y = rotr_7_cast_fp16)[name = tensor("op_1355_cast_fp16")]; + tensor qoi_13_cast_fp16 = add(x = var_1354_cast_fp16, y = var_1355_cast_fp16)[name = tensor("qoi_13_cast_fp16")]; + tensor var_1358_cast_fp16 = mul(x = var_1332_cast_fp16, y = rotr_7_cast_fp16)[name = tensor("op_1358_cast_fp16")]; + tensor var_1359_cast_fp16 = mul(x = var_1340_cast_fp16, y = roti_7_cast_fp16)[name = tensor("op_1359_cast_fp16")]; + tensor kor_13_cast_fp16 = sub(x = var_1358_cast_fp16, y = var_1359_cast_fp16)[name = tensor("kor_13_cast_fp16")]; + tensor var_1362_cast_fp16 = mul(x = var_1332_cast_fp16, y = roti_7_cast_fp16)[name = tensor("op_1362_cast_fp16")]; + tensor var_1363_cast_fp16 = mul(x = var_1340_cast_fp16, y = rotr_7_cast_fp16)[name = tensor("op_1363_cast_fp16")]; + tensor koi_13_cast_fp16 = add(x = var_1362_cast_fp16, y = var_1363_cast_fp16)[name = tensor("koi_13_cast_fp16")]; + tensor qo_7_axis_0 = const()[name = tensor("qo_7_axis_0"), val = tensor(-1)]; + tensor qo_7_cast_fp16 = stack(axis = qo_7_axis_0, values = (qor_13_cast_fp16, qoi_13_cast_fp16))[name = tensor("qo_7_cast_fp16")]; + tensor ko_7_axis_0 = const()[name = tensor("ko_7_axis_0"), val = tensor(-1)]; + tensor ko_7_cast_fp16 = stack(axis = ko_7_axis_0, values = (kor_13_cast_fp16, koi_13_cast_fp16))[name = tensor("ko_7_cast_fp16")]; + tensor var_1396 = const()[name = tensor("op_1396"), val = tensor([1, 1, 16, 64])]; + tensor q_21_cast_fp16 = reshape(shape = var_1396, x = qo_7_cast_fp16)[name = tensor("q_21_cast_fp16")]; + tensor var_1402 = const()[name = tensor("op_1402"), val = tensor([1, 1, 16, 64])]; + tensor k_15_cast_fp16 = reshape(shape = var_1402, x = ko_7_cast_fp16)[name = tensor("k_15_cast_fp16")]; + tensor _inversed_1427_y_0_to_fp16 = const()[name = tensor("_inversed_1427_y_0_to_fp16"), val = tensor(0x1p-9)]; + tensor _inversed_1427_cast_fp16 = mul(x = ts_23_cast_fp16, y = _inversed_1427_y_0_to_fp16)[name = tensor("_inversed_1427_cast_fp16")]; + tensor var_1428_cast_fp16 = floor(x = _inversed_1427_cast_fp16)[name = tensor("op_1428_cast_fp16")]; + tensor var_1429_to_fp16 = const()[name = tensor("op_1429_to_fp16"), val = tensor(0x1p+9)]; + tensor var_1430_cast_fp16 = mul(x = var_1428_cast_fp16, y = var_1429_to_fp16)[name = tensor("op_1430_cast_fp16")]; + tensor write_indices_float_15_cast_fp16 = sub(x = ts_23_cast_fp16, y = var_1430_cast_fp16)[name = tensor("write_indices_float_15_cast_fp16")]; + tensor var_1437_dtype_0 = const()[name = tensor("op_1437_dtype_0"), val = tensor("int32")]; + tensor write_indices_7_reps_0 = const()[name = tensor("write_indices_7_reps_0"), val = tensor([1, 1, 16, 64])]; + tensor write_indices_float_15_cast_fp16_to_int32 = cast(dtype = var_1437_dtype_0, x = write_indices_float_15_cast_fp16)[name = tensor("cast_123")]; + tensor write_indices_7 = tile(reps = write_indices_7_reps_0, x = write_indices_float_15_cast_fp16_to_int32)[name = tensor("write_indices_7")]; + tensor var_1447_begin_0 = const()[name = tensor("op_1447_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_1447_end_0 = const()[name = tensor("op_1447_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor var_1447_end_mask_0 = const()[name = tensor("op_1447_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_1447_squeeze_mask_0 = const()[name = tensor("op_1447_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor cache3_to_fp16_dtype_0 = const()[name = tensor("cache3_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor cache3_to_fp16 = cast(dtype = cache3_to_fp16_dtype_0, x = cache3)[name = tensor("cast_122")]; + tensor var_1447_cast_fp16 = slice_by_index(begin = var_1447_begin_0, end = var_1447_end_0, end_mask = var_1447_end_mask_0, squeeze_mask = var_1447_squeeze_mask_0, x = cache3_to_fp16)[name = tensor("op_1447_cast_fp16")]; + tensor var_1449_axis_0 = const()[name = tensor("op_1449_axis_0"), val = tensor(1)]; + tensor var_1449_mode_0 = const()[name = tensor("op_1449_mode_0"), val = tensor("update")]; + tensor var_1449_validate_indices_0 = const()[name = tensor("op_1449_validate_indices_0"), val = tensor(false)]; + tensor var_1449_cast_fp16 = scatter_along_axis(axis = var_1449_axis_0, data = var_1447_cast_fp16, indices = write_indices_7, mode = var_1449_mode_0, updates = k_15_cast_fp16, validate_indices = var_1449_validate_indices_0)[name = tensor("op_1449_cast_fp16")]; + tensor concat_23 = const()[name = tensor("concat_23"), val = tensor([0, 0, 0, 0, 0])]; + tensor concat_24 = const()[name = tensor("concat_24"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_7_internal_tensor_assign_1_stride_0 = const()[name = tensor("new_cache_7_internal_tensor_assign_1_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_7_internal_tensor_assign_1_begin_mask_0 = const()[name = tensor("new_cache_7_internal_tensor_assign_1_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_7_internal_tensor_assign_1_end_mask_0 = const()[name = tensor("new_cache_7_internal_tensor_assign_1_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_7_internal_tensor_assign_1_squeeze_mask_0 = const()[name = tensor("new_cache_7_internal_tensor_assign_1_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_18 = const()[name = tensor("shape_18"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_6 = const()[name = tensor("reduce_prod_6"), val = tensor(1048576)]; + tensor range_1d_6_start_0 = const()[name = tensor("range_1d_6_start_0"), val = tensor(0)]; + tensor range_1d_6_step_0 = const()[name = tensor("range_1d_6_step_0"), val = tensor(1)]; + tensor range_1d_6 = range_1d(end = reduce_prod_6, start = range_1d_6_start_0, step = range_1d_6_step_0)[name = tensor("range_1d_6")]; + tensor reshape_30 = reshape(shape = shape_18, x = range_1d_6)[name = tensor("reshape_30")]; + tensor slice_by_index_6 = slice_by_index(begin = concat_23, begin_mask = new_cache_7_internal_tensor_assign_1_begin_mask_0, end = concat_24, end_mask = new_cache_7_internal_tensor_assign_1_end_mask_0, squeeze_mask = new_cache_7_internal_tensor_assign_1_squeeze_mask_0, stride = new_cache_7_internal_tensor_assign_1_stride_0, x = reshape_30)[name = tensor("slice_by_index_6")]; + tensor reshape_31_shape_0 = const()[name = tensor("reshape_31_shape_0"), val = tensor([-1])]; + tensor reshape_31 = reshape(shape = reshape_31_shape_0, x = slice_by_index_6)[name = tensor("reshape_31")]; + tensor reshape_32_shape_0 = const()[name = tensor("reshape_32_shape_0"), val = tensor([-1])]; + tensor reshape_32_cast_fp16 = reshape(shape = reshape_32_shape_0, x = var_1449_cast_fp16)[name = tensor("reshape_32_cast_fp16")]; + tensor reshape_33_shape_0 = const()[name = tensor("reshape_33_shape_0"), val = tensor([-1])]; + tensor reshape_33_cast_fp16 = reshape(shape = reshape_33_shape_0, x = cache3_to_fp16)[name = tensor("reshape_33_cast_fp16")]; + tensor scatter_6_mode_0 = const()[name = tensor("scatter_6_mode_0"), val = tensor("update")]; + tensor scatter_6_axis_0 = const()[name = tensor("scatter_6_axis_0"), val = tensor(0)]; + tensor scatter_6_validate_indices_0 = const()[name = tensor("scatter_6_validate_indices_0"), val = tensor(false)]; + tensor scatter_6_cast_fp16 = scatter(axis = scatter_6_axis_0, data = reshape_33_cast_fp16, indices = reshape_31, mode = scatter_6_mode_0, updates = reshape_32_cast_fp16, validate_indices = scatter_6_validate_indices_0)[name = tensor("scatter_6_cast_fp16")]; + tensor reshape_34_cast_fp16 = reshape(shape = shape_18, x = scatter_6_cast_fp16)[name = tensor("reshape_34_cast_fp16")]; + tensor var_1457_begin_0 = const()[name = tensor("op_1457_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor var_1457_end_0 = const()[name = tensor("op_1457_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor var_1457_end_mask_0 = const()[name = tensor("op_1457_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_1457_squeeze_mask_0 = const()[name = tensor("op_1457_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor var_1457_cast_fp16 = slice_by_index(begin = var_1457_begin_0, end = var_1457_end_0, end_mask = var_1457_end_mask_0, squeeze_mask = var_1457_squeeze_mask_0, x = reshape_34_cast_fp16)[name = tensor("op_1457_cast_fp16")]; + tensor var_1459_axis_0 = const()[name = tensor("op_1459_axis_0"), val = tensor(1)]; + tensor var_1459_mode_0 = const()[name = tensor("op_1459_mode_0"), val = tensor("update")]; + tensor var_1459_validate_indices_0 = const()[name = tensor("op_1459_validate_indices_0"), val = tensor(false)]; + tensor var_1459_cast_fp16 = scatter_along_axis(axis = var_1459_axis_0, data = var_1457_cast_fp16, indices = write_indices_7, mode = var_1459_mode_0, updates = v_7_cast_fp16, validate_indices = var_1459_validate_indices_0)[name = tensor("op_1459_cast_fp16")]; + tensor concat_25 = const()[name = tensor("concat_25"), val = tensor([1, 0, 0, 0, 0])]; + tensor concat_26 = const()[name = tensor("concat_26"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_7_internal_tensor_assign_2_stride_0 = const()[name = tensor("new_cache_7_internal_tensor_assign_2_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_7_internal_tensor_assign_2_begin_mask_0 = const()[name = tensor("new_cache_7_internal_tensor_assign_2_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_7_internal_tensor_assign_2_end_mask_0 = const()[name = tensor("new_cache_7_internal_tensor_assign_2_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_7_internal_tensor_assign_2_squeeze_mask_0 = const()[name = tensor("new_cache_7_internal_tensor_assign_2_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_19 = const()[name = tensor("shape_19"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_7 = const()[name = tensor("reduce_prod_7"), val = tensor(1048576)]; + tensor range_1d_7_start_0 = const()[name = tensor("range_1d_7_start_0"), val = tensor(0)]; + tensor range_1d_7_step_0 = const()[name = tensor("range_1d_7_step_0"), val = tensor(1)]; + tensor range_1d_7 = range_1d(end = reduce_prod_7, start = range_1d_7_start_0, step = range_1d_7_step_0)[name = tensor("range_1d_7")]; + tensor reshape_35 = reshape(shape = shape_19, x = range_1d_7)[name = tensor("reshape_35")]; + tensor slice_by_index_7 = slice_by_index(begin = concat_25, begin_mask = new_cache_7_internal_tensor_assign_2_begin_mask_0, end = concat_26, end_mask = new_cache_7_internal_tensor_assign_2_end_mask_0, squeeze_mask = new_cache_7_internal_tensor_assign_2_squeeze_mask_0, stride = new_cache_7_internal_tensor_assign_2_stride_0, x = reshape_35)[name = tensor("slice_by_index_7")]; + tensor reshape_36_shape_0 = const()[name = tensor("reshape_36_shape_0"), val = tensor([-1])]; + tensor reshape_36 = reshape(shape = reshape_36_shape_0, x = slice_by_index_7)[name = tensor("reshape_36")]; + tensor reshape_37_shape_0 = const()[name = tensor("reshape_37_shape_0"), val = tensor([-1])]; + tensor reshape_37_cast_fp16 = reshape(shape = reshape_37_shape_0, x = var_1459_cast_fp16)[name = tensor("reshape_37_cast_fp16")]; + tensor reshape_38_shape_0 = const()[name = tensor("reshape_38_shape_0"), val = tensor([-1])]; + tensor reshape_38_cast_fp16 = reshape(shape = reshape_38_shape_0, x = reshape_34_cast_fp16)[name = tensor("reshape_38_cast_fp16")]; + tensor scatter_7_mode_0 = const()[name = tensor("scatter_7_mode_0"), val = tensor("update")]; + tensor scatter_7_axis_0 = const()[name = tensor("scatter_7_axis_0"), val = tensor(0)]; + tensor scatter_7_validate_indices_0 = const()[name = tensor("scatter_7_validate_indices_0"), val = tensor(false)]; + tensor scatter_7_cast_fp16 = scatter(axis = scatter_7_axis_0, data = reshape_38_cast_fp16, indices = reshape_36, mode = scatter_7_mode_0, updates = reshape_37_cast_fp16, validate_indices = scatter_7_validate_indices_0)[name = tensor("scatter_7_cast_fp16")]; + tensor reshape_39_cast_fp16 = reshape(shape = shape_19, x = scatter_7_cast_fp16)[name = tensor("reshape_39_cast_fp16")]; + tensor reshape_39_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("reshape_39_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor keys_19_begin_0 = const()[name = tensor("keys_19_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor keys_19_end_0 = const()[name = tensor("keys_19_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor keys_19_end_mask_0 = const()[name = tensor("keys_19_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor keys_19_squeeze_mask_0 = const()[name = tensor("keys_19_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor keys_19_cast_fp16 = slice_by_index(begin = keys_19_begin_0, end = keys_19_end_0, end_mask = keys_19_end_mask_0, squeeze_mask = keys_19_squeeze_mask_0, x = reshape_39_cast_fp16)[name = tensor("keys_19_cast_fp16")]; + tensor values_19_begin_0 = const()[name = tensor("values_19_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor values_19_end_0 = const()[name = tensor("values_19_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor values_19_end_mask_0 = const()[name = tensor("values_19_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor values_19_squeeze_mask_0 = const()[name = tensor("values_19_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor values_19_cast_fp16 = slice_by_index(begin = values_19_begin_0, end = values_19_end_0, end_mask = values_19_end_mask_0, squeeze_mask = values_19_squeeze_mask_0, x = reshape_39_cast_fp16)[name = tensor("values_19_cast_fp16")]; + tensor var_1471_cast_fp16 = not_equal(x = keys_19_cast_fp16, y = keys_19_cast_fp16)[name = tensor("op_1471_cast_fp16")]; + tensor keys_21_cast_fp16 = select(a = var_319_to_fp16, b = keys_19_cast_fp16, cond = var_1471_cast_fp16)[name = tensor("keys_21_cast_fp16")]; + tensor var_1479_cast_fp16 = not_equal(x = values_19_cast_fp16, y = values_19_cast_fp16)[name = tensor("op_1479_cast_fp16")]; + tensor values_21_cast_fp16 = select(a = var_319_to_fp16, b = values_19_cast_fp16, cond = var_1479_cast_fp16)[name = tensor("values_21_cast_fp16")]; + tensor var_1503 = const()[name = tensor("op_1503"), val = tensor([0, 2, 1, 3])]; + tensor var_1519 = const()[name = tensor("op_1519"), val = tensor([1, 1, 1])]; + tensor var_1520_cast_fp16 = reshape(shape = var_1519, x = position3_to_fp16)[name = tensor("op_1520_cast_fp16")]; + tensor var_1540_to_fp16 = const()[name = tensor("op_1540_to_fp16"), val = tensor(0x1p+0)]; + tensor valid_len_7_cast_fp16 = add(x = var_1520_cast_fp16, y = var_1540_to_fp16)[name = tensor("valid_len_7_cast_fp16")]; + tensor valid_mask_7_cast_fp16 = less(x = k_positions_1_promoted_to_fp16, y = valid_len_7_cast_fp16)[name = tensor("valid_mask_7_cast_fp16")]; + tensor causal_mask_7_cast_fp16 = less_equal(x = k_positions_1_promoted_to_fp16, y = var_1520_cast_fp16)[name = tensor("causal_mask_7_cast_fp16")]; + tensor attn_mask_13 = logical_and(x = valid_mask_7_cast_fp16, y = causal_mask_7_cast_fp16)[name = tensor("attn_mask_13")]; + tensor attn_mask_15_axes_0 = const()[name = tensor("attn_mask_15_axes_0"), val = tensor([1])]; + tensor attn_mask_15 = expand_dims(axes = attn_mask_15_axes_0, x = attn_mask_13)[name = tensor("attn_mask_15")]; + tensor var_1558_transpose_x_0 = const()[name = tensor("op_1558_transpose_x_0"), val = tensor(false)]; + tensor var_1558_transpose_y_0 = const()[name = tensor("op_1558_transpose_y_0"), val = tensor(false)]; + tensor transpose_24_perm_0 = const()[name = tensor("transpose_24_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_25_perm_0 = const()[name = tensor("transpose_25_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_25 = transpose(perm = transpose_25_perm_0, x = keys_21_cast_fp16)[name = tensor("transpose_39")]; + tensor transpose_24 = transpose(perm = transpose_24_perm_0, x = q_21_cast_fp16)[name = tensor("transpose_40")]; + tensor var_1558_cast_fp16 = matmul(transpose_x = var_1558_transpose_x_0, transpose_y = var_1558_transpose_y_0, x = transpose_24, y = transpose_25)[name = tensor("op_1558_cast_fp16")]; + tensor var_1552_to_fp16 = const()[name = tensor("op_1552_to_fp16"), val = tensor([0x1p-3])]; + tensor attn_weights_19_cast_fp16 = mul(x = var_1558_cast_fp16, y = var_1552_to_fp16)[name = tensor("attn_weights_19_cast_fp16")]; + tensor var_1560 = logical_not(x = attn_mask_15)[name = tensor("op_1560")]; + tensor var_1561_to_fp16 = const()[name = tensor("op_1561_to_fp16"), val = tensor(-inf)]; + tensor attn_weights_21_cast_fp16 = select(a = var_1561_to_fp16, b = attn_weights_19_cast_fp16, cond = var_1560)[name = tensor("attn_weights_21_cast_fp16")]; + tensor var_1563 = const()[name = tensor("op_1563"), val = tensor(-1)]; + tensor attn_weights_23_cast_fp16 = softmax(axis = var_1563, x = attn_weights_21_cast_fp16)[name = tensor("attn_weights_23_cast_fp16")]; + tensor attn_output_7_transpose_x_0 = const()[name = tensor("attn_output_7_transpose_x_0"), val = tensor(false)]; + tensor attn_output_7_transpose_y_0 = const()[name = tensor("attn_output_7_transpose_y_0"), val = tensor(false)]; + tensor values_23_cast_fp16 = transpose(perm = var_1503, x = values_21_cast_fp16)[name = tensor("transpose_41")]; + tensor attn_output_7_cast_fp16 = matmul(transpose_x = attn_output_7_transpose_x_0, transpose_y = attn_output_7_transpose_y_0, x = attn_weights_23_cast_fp16, y = values_23_cast_fp16)[name = tensor("attn_output_7_cast_fp16")]; + tensor var_1571 = const()[name = tensor("op_1571"), val = tensor([0, 2, 1, 3])]; + tensor var_1576 = const()[name = tensor("op_1576"), val = tensor([1, 1, 1024])]; + tensor var_1572_cast_fp16 = transpose(perm = var_1571, x = attn_output_7_cast_fp16)[name = tensor("transpose_38")]; + tensor input_43_cast_fp16 = reshape(shape = var_1576, x = var_1572_cast_fp16)[name = tensor("input_43_cast_fp16")]; + tensor attn3_out_proj_weight_to_fp16 = const()[name = tensor("attn3_out_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(82951808)))]; + tensor linear_14_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = attn3_out_proj_weight_to_fp16, x = input_43_cast_fp16)[name = tensor("linear_14_cast_fp16")]; + tensor var_1582_to_fp16 = const()[name = tensor("op_1582_to_fp16"), val = tensor(0x1p+0)]; + tensor var_1583_cast_fp16 = add(x = position3_to_fp16, y = var_1582_to_fp16)[name = tensor("op_1583_cast_fp16")]; + tensor var_1583_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_1583_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor input_45_cast_fp16 = add(x = input_39_cast_fp16, y = linear_14_cast_fp16)[name = tensor("input_45_cast_fp16")]; + tensor input_47_axes_0 = const()[name = tensor("input_47_axes_0"), val = tensor([-1])]; + tensor norm3_2_weight_to_fp16 = const()[name = tensor("norm3_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(85049024)))]; + tensor norm3_2_bias_to_fp16 = const()[name = tensor("norm3_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(85051136)))]; + tensor var_1587_to_fp16 = const()[name = tensor("op_1587_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_47_cast_fp16 = layer_norm(axes = input_47_axes_0, beta = norm3_2_bias_to_fp16, epsilon = var_1587_to_fp16, gamma = norm3_2_weight_to_fp16, x = input_45_cast_fp16)[name = tensor("input_47_cast_fp16")]; + tensor linear3_1_weight_to_fp16 = const()[name = tensor("linear3_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(85053248)))]; + tensor linear_15_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = linear3_1_weight_to_fp16, x = input_47_cast_fp16)[name = tensor("linear_15_cast_fp16")]; + tensor input_49_mode_0 = const()[name = tensor("input_49_mode_0"), val = tensor("EXACT")]; + tensor input_49_cast_fp16 = gelu(mode = input_49_mode_0, x = linear_15_cast_fp16)[name = tensor("input_49_cast_fp16")]; + tensor linear3_2_weight_to_fp16 = const()[name = tensor("linear3_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(93441920)))]; + tensor linear_16_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = linear3_2_weight_to_fp16, x = input_49_cast_fp16)[name = tensor("linear_16_cast_fp16")]; + tensor input_51_cast_fp16 = add(x = input_45_cast_fp16, y = linear_16_cast_fp16)[name = tensor("input_51_cast_fp16")]; + tensor input_53_axes_0 = const()[name = tensor("input_53_axes_0"), val = tensor([-1])]; + tensor norm4_1_weight_to_fp16 = const()[name = tensor("norm4_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(101830592)))]; + tensor norm4_1_bias_to_fp16 = const()[name = tensor("norm4_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(101832704)))]; + tensor var_1604_to_fp16 = const()[name = tensor("op_1604_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_53_cast_fp16 = layer_norm(axes = input_53_axes_0, beta = norm4_1_bias_to_fp16, epsilon = var_1604_to_fp16, gamma = norm4_1_weight_to_fp16, x = input_51_cast_fp16)[name = tensor("input_53_cast_fp16")]; + tensor attn4_in_proj_weight_to_fp16 = const()[name = tensor("attn4_in_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(101834816)))]; + tensor linear_17_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = attn4_in_proj_weight_to_fp16, x = input_53_cast_fp16)[name = tensor("linear_17_cast_fp16")]; + tensor var_1618 = const()[name = tensor("op_1618"), val = tensor([1, 1, 3, 16, 64])]; + tensor qkv_9_cast_fp16 = reshape(shape = var_1618, x = linear_17_cast_fp16)[name = tensor("qkv_9_cast_fp16")]; + tensor q_25_begin_0 = const()[name = tensor("q_25_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor q_25_end_0 = const()[name = tensor("q_25_end_0"), val = tensor([1, 1, 1, 16, 64])]; + tensor q_25_end_mask_0 = const()[name = tensor("q_25_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor q_25_squeeze_mask_0 = const()[name = tensor("q_25_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor q_25_cast_fp16 = slice_by_index(begin = q_25_begin_0, end = q_25_end_0, end_mask = q_25_end_mask_0, squeeze_mask = q_25_squeeze_mask_0, x = qkv_9_cast_fp16)[name = tensor("q_25_cast_fp16")]; + tensor k_17_begin_0 = const()[name = tensor("k_17_begin_0"), val = tensor([0, 0, 1, 0, 0])]; + tensor k_17_end_0 = const()[name = tensor("k_17_end_0"), val = tensor([1, 1, 2, 16, 64])]; + tensor k_17_end_mask_0 = const()[name = tensor("k_17_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor k_17_squeeze_mask_0 = const()[name = tensor("k_17_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor k_17_cast_fp16 = slice_by_index(begin = k_17_begin_0, end = k_17_end_0, end_mask = k_17_end_mask_0, squeeze_mask = k_17_squeeze_mask_0, x = qkv_9_cast_fp16)[name = tensor("k_17_cast_fp16")]; + tensor v_9_begin_0 = const()[name = tensor("v_9_begin_0"), val = tensor([0, 0, 2, 0, 0])]; + tensor v_9_end_0 = const()[name = tensor("v_9_end_0"), val = tensor([1, 1, 3, 16, 64])]; + tensor v_9_end_mask_0 = const()[name = tensor("v_9_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor v_9_squeeze_mask_0 = const()[name = tensor("v_9_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor v_9_cast_fp16 = slice_by_index(begin = v_9_begin_0, end = v_9_end_0, end_mask = v_9_end_mask_0, squeeze_mask = v_9_squeeze_mask_0, x = qkv_9_cast_fp16)[name = tensor("v_9_cast_fp16")]; + tensor var_1684 = const()[name = tensor("op_1684"), val = tensor([1, 1, 1, 1])]; + tensor position4_to_fp16_dtype_0 = const()[name = tensor("position4_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor position4_to_fp16 = cast(dtype = position4_to_fp16_dtype_0, x = position4)[name = tensor("cast_119")]; + tensor ts_29_cast_fp16 = reshape(shape = var_1684, x = position4_to_fp16)[name = tensor("ts_29_cast_fp16")]; + tensor var_1691 = const()[name = tensor("op_1691"), val = tensor([1, 1, 16, 32, 2])]; + tensor q_complex_9_cast_fp16 = reshape(shape = var_1691, x = q_25_cast_fp16)[name = tensor("q_complex_9_cast_fp16")]; + tensor var_1698 = const()[name = tensor("op_1698"), val = tensor([1, 1, 16, 32, 2])]; + tensor k_complex_9_cast_fp16 = reshape(shape = var_1698, x = k_17_cast_fp16)[name = tensor("k_complex_9_cast_fp16")]; + tensor var_1702_begin_0 = const()[name = tensor("op_1702_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_1702_end_0 = const()[name = tensor("op_1702_end_0"), val = tensor([1, 1, 16, 32, 1])]; + tensor var_1702_end_mask_0 = const()[name = tensor("op_1702_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_1702_squeeze_mask_0 = const()[name = tensor("op_1702_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_1702_cast_fp16 = slice_by_index(begin = var_1702_begin_0, end = var_1702_end_0, end_mask = var_1702_end_mask_0, squeeze_mask = var_1702_squeeze_mask_0, x = q_complex_9_cast_fp16)[name = tensor("op_1702_cast_fp16")]; + tensor var_1710_begin_0 = const()[name = tensor("op_1710_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_1710_end_0 = const()[name = tensor("op_1710_end_0"), val = tensor([1, 1, 16, 32, 2])]; + tensor var_1710_end_mask_0 = const()[name = tensor("op_1710_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_1710_squeeze_mask_0 = const()[name = tensor("op_1710_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_1710_cast_fp16 = slice_by_index(begin = var_1710_begin_0, end = var_1710_end_0, end_mask = var_1710_end_mask_0, squeeze_mask = var_1710_squeeze_mask_0, x = q_complex_9_cast_fp16)[name = tensor("op_1710_cast_fp16")]; + tensor var_1718_begin_0 = const()[name = tensor("op_1718_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_1718_end_0 = const()[name = tensor("op_1718_end_0"), val = tensor([1, 1, 16, 32, 1])]; + tensor var_1718_end_mask_0 = const()[name = tensor("op_1718_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_1718_squeeze_mask_0 = const()[name = tensor("op_1718_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_1718_cast_fp16 = slice_by_index(begin = var_1718_begin_0, end = var_1718_end_0, end_mask = var_1718_end_mask_0, squeeze_mask = var_1718_squeeze_mask_0, x = k_complex_9_cast_fp16)[name = tensor("op_1718_cast_fp16")]; + tensor var_1726_begin_0 = const()[name = tensor("op_1726_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_1726_end_0 = const()[name = tensor("op_1726_end_0"), val = tensor([1, 1, 16, 32, 2])]; + tensor var_1726_end_mask_0 = const()[name = tensor("op_1726_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_1726_squeeze_mask_0 = const()[name = tensor("op_1726_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_1726_cast_fp16 = slice_by_index(begin = var_1726_begin_0, end = var_1726_end_0, end_mask = var_1726_end_mask_0, squeeze_mask = var_1726_squeeze_mask_0, x = k_complex_9_cast_fp16)[name = tensor("op_1726_cast_fp16")]; + tensor freqs_9_to_fp16 = const()[name = tensor("freqs_9_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(108126336)))]; + tensor var_1732_cast_fp16 = mul(x = freqs_9_to_fp16, y = ts_29_cast_fp16)[name = tensor("op_1732_cast_fp16")]; + tensor rotr_9_cast_fp16 = cos(x = var_1732_cast_fp16)[name = tensor("rotr_9_cast_fp16")]; + tensor roti_9_cast_fp16 = sin(x = var_1732_cast_fp16)[name = tensor("roti_9_cast_fp16")]; + tensor var_1736_cast_fp16 = mul(x = var_1702_cast_fp16, y = rotr_9_cast_fp16)[name = tensor("op_1736_cast_fp16")]; + tensor var_1737_cast_fp16 = mul(x = var_1710_cast_fp16, y = roti_9_cast_fp16)[name = tensor("op_1737_cast_fp16")]; + tensor qor_17_cast_fp16 = sub(x = var_1736_cast_fp16, y = var_1737_cast_fp16)[name = tensor("qor_17_cast_fp16")]; + tensor var_1740_cast_fp16 = mul(x = var_1702_cast_fp16, y = roti_9_cast_fp16)[name = tensor("op_1740_cast_fp16")]; + tensor var_1741_cast_fp16 = mul(x = var_1710_cast_fp16, y = rotr_9_cast_fp16)[name = tensor("op_1741_cast_fp16")]; + tensor qoi_17_cast_fp16 = add(x = var_1740_cast_fp16, y = var_1741_cast_fp16)[name = tensor("qoi_17_cast_fp16")]; + tensor var_1744_cast_fp16 = mul(x = var_1718_cast_fp16, y = rotr_9_cast_fp16)[name = tensor("op_1744_cast_fp16")]; + tensor var_1745_cast_fp16 = mul(x = var_1726_cast_fp16, y = roti_9_cast_fp16)[name = tensor("op_1745_cast_fp16")]; + tensor kor_17_cast_fp16 = sub(x = var_1744_cast_fp16, y = var_1745_cast_fp16)[name = tensor("kor_17_cast_fp16")]; + tensor var_1748_cast_fp16 = mul(x = var_1718_cast_fp16, y = roti_9_cast_fp16)[name = tensor("op_1748_cast_fp16")]; + tensor var_1749_cast_fp16 = mul(x = var_1726_cast_fp16, y = rotr_9_cast_fp16)[name = tensor("op_1749_cast_fp16")]; + tensor koi_17_cast_fp16 = add(x = var_1748_cast_fp16, y = var_1749_cast_fp16)[name = tensor("koi_17_cast_fp16")]; + tensor qo_9_axis_0 = const()[name = tensor("qo_9_axis_0"), val = tensor(-1)]; + tensor qo_9_cast_fp16 = stack(axis = qo_9_axis_0, values = (qor_17_cast_fp16, qoi_17_cast_fp16))[name = tensor("qo_9_cast_fp16")]; + tensor ko_9_axis_0 = const()[name = tensor("ko_9_axis_0"), val = tensor(-1)]; + tensor ko_9_cast_fp16 = stack(axis = ko_9_axis_0, values = (kor_17_cast_fp16, koi_17_cast_fp16))[name = tensor("ko_9_cast_fp16")]; + tensor var_1782 = const()[name = tensor("op_1782"), val = tensor([1, 1, 16, 64])]; + tensor q_27_cast_fp16 = reshape(shape = var_1782, x = qo_9_cast_fp16)[name = tensor("q_27_cast_fp16")]; + tensor var_1788 = const()[name = tensor("op_1788"), val = tensor([1, 1, 16, 64])]; + tensor k_19_cast_fp16 = reshape(shape = var_1788, x = ko_9_cast_fp16)[name = tensor("k_19_cast_fp16")]; + tensor _inversed_1813_y_0_to_fp16 = const()[name = tensor("_inversed_1813_y_0_to_fp16"), val = tensor(0x1p-9)]; + tensor _inversed_1813_cast_fp16 = mul(x = ts_29_cast_fp16, y = _inversed_1813_y_0_to_fp16)[name = tensor("_inversed_1813_cast_fp16")]; + tensor var_1814_cast_fp16 = floor(x = _inversed_1813_cast_fp16)[name = tensor("op_1814_cast_fp16")]; + tensor var_1815_to_fp16 = const()[name = tensor("op_1815_to_fp16"), val = tensor(0x1p+9)]; + tensor var_1816_cast_fp16 = mul(x = var_1814_cast_fp16, y = var_1815_to_fp16)[name = tensor("op_1816_cast_fp16")]; + tensor write_indices_float_19_cast_fp16 = sub(x = ts_29_cast_fp16, y = var_1816_cast_fp16)[name = tensor("write_indices_float_19_cast_fp16")]; + tensor var_1823_dtype_0 = const()[name = tensor("op_1823_dtype_0"), val = tensor("int32")]; + tensor write_indices_9_reps_0 = const()[name = tensor("write_indices_9_reps_0"), val = tensor([1, 1, 16, 64])]; + tensor write_indices_float_19_cast_fp16_to_int32 = cast(dtype = var_1823_dtype_0, x = write_indices_float_19_cast_fp16)[name = tensor("cast_118")]; + tensor write_indices_9 = tile(reps = write_indices_9_reps_0, x = write_indices_float_19_cast_fp16_to_int32)[name = tensor("write_indices_9")]; + tensor var_1833_begin_0 = const()[name = tensor("op_1833_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_1833_end_0 = const()[name = tensor("op_1833_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor var_1833_end_mask_0 = const()[name = tensor("op_1833_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_1833_squeeze_mask_0 = const()[name = tensor("op_1833_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor cache4_to_fp16_dtype_0 = const()[name = tensor("cache4_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor cache4_to_fp16 = cast(dtype = cache4_to_fp16_dtype_0, x = cache4)[name = tensor("cast_117")]; + tensor var_1833_cast_fp16 = slice_by_index(begin = var_1833_begin_0, end = var_1833_end_0, end_mask = var_1833_end_mask_0, squeeze_mask = var_1833_squeeze_mask_0, x = cache4_to_fp16)[name = tensor("op_1833_cast_fp16")]; + tensor var_1835_axis_0 = const()[name = tensor("op_1835_axis_0"), val = tensor(1)]; + tensor var_1835_mode_0 = const()[name = tensor("op_1835_mode_0"), val = tensor("update")]; + tensor var_1835_validate_indices_0 = const()[name = tensor("op_1835_validate_indices_0"), val = tensor(false)]; + tensor var_1835_cast_fp16 = scatter_along_axis(axis = var_1835_axis_0, data = var_1833_cast_fp16, indices = write_indices_9, mode = var_1835_mode_0, updates = k_19_cast_fp16, validate_indices = var_1835_validate_indices_0)[name = tensor("op_1835_cast_fp16")]; + tensor concat_30 = const()[name = tensor("concat_30"), val = tensor([0, 0, 0, 0, 0])]; + tensor concat_31 = const()[name = tensor("concat_31"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_9_internal_tensor_assign_1_stride_0 = const()[name = tensor("new_cache_9_internal_tensor_assign_1_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_9_internal_tensor_assign_1_begin_mask_0 = const()[name = tensor("new_cache_9_internal_tensor_assign_1_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_9_internal_tensor_assign_1_end_mask_0 = const()[name = tensor("new_cache_9_internal_tensor_assign_1_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_9_internal_tensor_assign_1_squeeze_mask_0 = const()[name = tensor("new_cache_9_internal_tensor_assign_1_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_20 = const()[name = tensor("shape_20"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_8 = const()[name = tensor("reduce_prod_8"), val = tensor(1048576)]; + tensor range_1d_8_start_0 = const()[name = tensor("range_1d_8_start_0"), val = tensor(0)]; + tensor range_1d_8_step_0 = const()[name = tensor("range_1d_8_step_0"), val = tensor(1)]; + tensor range_1d_8 = range_1d(end = reduce_prod_8, start = range_1d_8_start_0, step = range_1d_8_step_0)[name = tensor("range_1d_8")]; + tensor reshape_40 = reshape(shape = shape_20, x = range_1d_8)[name = tensor("reshape_40")]; + tensor slice_by_index_8 = slice_by_index(begin = concat_30, begin_mask = new_cache_9_internal_tensor_assign_1_begin_mask_0, end = concat_31, end_mask = new_cache_9_internal_tensor_assign_1_end_mask_0, squeeze_mask = new_cache_9_internal_tensor_assign_1_squeeze_mask_0, stride = new_cache_9_internal_tensor_assign_1_stride_0, x = reshape_40)[name = tensor("slice_by_index_8")]; + tensor reshape_41_shape_0 = const()[name = tensor("reshape_41_shape_0"), val = tensor([-1])]; + tensor reshape_41 = reshape(shape = reshape_41_shape_0, x = slice_by_index_8)[name = tensor("reshape_41")]; + tensor reshape_42_shape_0 = const()[name = tensor("reshape_42_shape_0"), val = tensor([-1])]; + tensor reshape_42_cast_fp16 = reshape(shape = reshape_42_shape_0, x = var_1835_cast_fp16)[name = tensor("reshape_42_cast_fp16")]; + tensor reshape_43_shape_0 = const()[name = tensor("reshape_43_shape_0"), val = tensor([-1])]; + tensor reshape_43_cast_fp16 = reshape(shape = reshape_43_shape_0, x = cache4_to_fp16)[name = tensor("reshape_43_cast_fp16")]; + tensor scatter_8_mode_0 = const()[name = tensor("scatter_8_mode_0"), val = tensor("update")]; + tensor scatter_8_axis_0 = const()[name = tensor("scatter_8_axis_0"), val = tensor(0)]; + tensor scatter_8_validate_indices_0 = const()[name = tensor("scatter_8_validate_indices_0"), val = tensor(false)]; + tensor scatter_8_cast_fp16 = scatter(axis = scatter_8_axis_0, data = reshape_43_cast_fp16, indices = reshape_41, mode = scatter_8_mode_0, updates = reshape_42_cast_fp16, validate_indices = scatter_8_validate_indices_0)[name = tensor("scatter_8_cast_fp16")]; + tensor reshape_44_cast_fp16 = reshape(shape = shape_20, x = scatter_8_cast_fp16)[name = tensor("reshape_44_cast_fp16")]; + tensor var_1843_begin_0 = const()[name = tensor("op_1843_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor var_1843_end_0 = const()[name = tensor("op_1843_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor var_1843_end_mask_0 = const()[name = tensor("op_1843_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_1843_squeeze_mask_0 = const()[name = tensor("op_1843_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor var_1843_cast_fp16 = slice_by_index(begin = var_1843_begin_0, end = var_1843_end_0, end_mask = var_1843_end_mask_0, squeeze_mask = var_1843_squeeze_mask_0, x = reshape_44_cast_fp16)[name = tensor("op_1843_cast_fp16")]; + tensor var_1845_axis_0 = const()[name = tensor("op_1845_axis_0"), val = tensor(1)]; + tensor var_1845_mode_0 = const()[name = tensor("op_1845_mode_0"), val = tensor("update")]; + tensor var_1845_validate_indices_0 = const()[name = tensor("op_1845_validate_indices_0"), val = tensor(false)]; + tensor var_1845_cast_fp16 = scatter_along_axis(axis = var_1845_axis_0, data = var_1843_cast_fp16, indices = write_indices_9, mode = var_1845_mode_0, updates = v_9_cast_fp16, validate_indices = var_1845_validate_indices_0)[name = tensor("op_1845_cast_fp16")]; + tensor concat_32 = const()[name = tensor("concat_32"), val = tensor([1, 0, 0, 0, 0])]; + tensor concat_33 = const()[name = tensor("concat_33"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_9_internal_tensor_assign_2_stride_0 = const()[name = tensor("new_cache_9_internal_tensor_assign_2_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_9_internal_tensor_assign_2_begin_mask_0 = const()[name = tensor("new_cache_9_internal_tensor_assign_2_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_9_internal_tensor_assign_2_end_mask_0 = const()[name = tensor("new_cache_9_internal_tensor_assign_2_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_9_internal_tensor_assign_2_squeeze_mask_0 = const()[name = tensor("new_cache_9_internal_tensor_assign_2_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_21 = const()[name = tensor("shape_21"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_9 = const()[name = tensor("reduce_prod_9"), val = tensor(1048576)]; + tensor range_1d_9_start_0 = const()[name = tensor("range_1d_9_start_0"), val = tensor(0)]; + tensor range_1d_9_step_0 = const()[name = tensor("range_1d_9_step_0"), val = tensor(1)]; + tensor range_1d_9 = range_1d(end = reduce_prod_9, start = range_1d_9_start_0, step = range_1d_9_step_0)[name = tensor("range_1d_9")]; + tensor reshape_45 = reshape(shape = shape_21, x = range_1d_9)[name = tensor("reshape_45")]; + tensor slice_by_index_9 = slice_by_index(begin = concat_32, begin_mask = new_cache_9_internal_tensor_assign_2_begin_mask_0, end = concat_33, end_mask = new_cache_9_internal_tensor_assign_2_end_mask_0, squeeze_mask = new_cache_9_internal_tensor_assign_2_squeeze_mask_0, stride = new_cache_9_internal_tensor_assign_2_stride_0, x = reshape_45)[name = tensor("slice_by_index_9")]; + tensor reshape_46_shape_0 = const()[name = tensor("reshape_46_shape_0"), val = tensor([-1])]; + tensor reshape_46 = reshape(shape = reshape_46_shape_0, x = slice_by_index_9)[name = tensor("reshape_46")]; + tensor reshape_47_shape_0 = const()[name = tensor("reshape_47_shape_0"), val = tensor([-1])]; + tensor reshape_47_cast_fp16 = reshape(shape = reshape_47_shape_0, x = var_1845_cast_fp16)[name = tensor("reshape_47_cast_fp16")]; + tensor reshape_48_shape_0 = const()[name = tensor("reshape_48_shape_0"), val = tensor([-1])]; + tensor reshape_48_cast_fp16 = reshape(shape = reshape_48_shape_0, x = reshape_44_cast_fp16)[name = tensor("reshape_48_cast_fp16")]; + tensor scatter_9_mode_0 = const()[name = tensor("scatter_9_mode_0"), val = tensor("update")]; + tensor scatter_9_axis_0 = const()[name = tensor("scatter_9_axis_0"), val = tensor(0)]; + tensor scatter_9_validate_indices_0 = const()[name = tensor("scatter_9_validate_indices_0"), val = tensor(false)]; + tensor scatter_9_cast_fp16 = scatter(axis = scatter_9_axis_0, data = reshape_48_cast_fp16, indices = reshape_46, mode = scatter_9_mode_0, updates = reshape_47_cast_fp16, validate_indices = scatter_9_validate_indices_0)[name = tensor("scatter_9_cast_fp16")]; + tensor reshape_49_cast_fp16 = reshape(shape = shape_21, x = scatter_9_cast_fp16)[name = tensor("reshape_49_cast_fp16")]; + tensor reshape_49_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("reshape_49_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor keys_25_begin_0 = const()[name = tensor("keys_25_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor keys_25_end_0 = const()[name = tensor("keys_25_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor keys_25_end_mask_0 = const()[name = tensor("keys_25_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor keys_25_squeeze_mask_0 = const()[name = tensor("keys_25_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor keys_25_cast_fp16 = slice_by_index(begin = keys_25_begin_0, end = keys_25_end_0, end_mask = keys_25_end_mask_0, squeeze_mask = keys_25_squeeze_mask_0, x = reshape_49_cast_fp16)[name = tensor("keys_25_cast_fp16")]; + tensor values_25_begin_0 = const()[name = tensor("values_25_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor values_25_end_0 = const()[name = tensor("values_25_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor values_25_end_mask_0 = const()[name = tensor("values_25_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor values_25_squeeze_mask_0 = const()[name = tensor("values_25_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor values_25_cast_fp16 = slice_by_index(begin = values_25_begin_0, end = values_25_end_0, end_mask = values_25_end_mask_0, squeeze_mask = values_25_squeeze_mask_0, x = reshape_49_cast_fp16)[name = tensor("values_25_cast_fp16")]; + tensor var_1857_cast_fp16 = not_equal(x = keys_25_cast_fp16, y = keys_25_cast_fp16)[name = tensor("op_1857_cast_fp16")]; + tensor keys_27_cast_fp16 = select(a = var_319_to_fp16, b = keys_25_cast_fp16, cond = var_1857_cast_fp16)[name = tensor("keys_27_cast_fp16")]; + tensor var_1865_cast_fp16 = not_equal(x = values_25_cast_fp16, y = values_25_cast_fp16)[name = tensor("op_1865_cast_fp16")]; + tensor values_27_cast_fp16 = select(a = var_319_to_fp16, b = values_25_cast_fp16, cond = var_1865_cast_fp16)[name = tensor("values_27_cast_fp16")]; + tensor var_1889 = const()[name = tensor("op_1889"), val = tensor([0, 2, 1, 3])]; + tensor var_1905 = const()[name = tensor("op_1905"), val = tensor([1, 1, 1])]; + tensor var_1906_cast_fp16 = reshape(shape = var_1905, x = position4_to_fp16)[name = tensor("op_1906_cast_fp16")]; + tensor var_1926_to_fp16 = const()[name = tensor("op_1926_to_fp16"), val = tensor(0x1p+0)]; + tensor valid_len_9_cast_fp16 = add(x = var_1906_cast_fp16, y = var_1926_to_fp16)[name = tensor("valid_len_9_cast_fp16")]; + tensor valid_mask_9_cast_fp16 = less(x = k_positions_1_promoted_to_fp16, y = valid_len_9_cast_fp16)[name = tensor("valid_mask_9_cast_fp16")]; + tensor causal_mask_9_cast_fp16 = less_equal(x = k_positions_1_promoted_to_fp16, y = var_1906_cast_fp16)[name = tensor("causal_mask_9_cast_fp16")]; + tensor attn_mask_17 = logical_and(x = valid_mask_9_cast_fp16, y = causal_mask_9_cast_fp16)[name = tensor("attn_mask_17")]; + tensor attn_mask_19_axes_0 = const()[name = tensor("attn_mask_19_axes_0"), val = tensor([1])]; + tensor attn_mask_19 = expand_dims(axes = attn_mask_19_axes_0, x = attn_mask_17)[name = tensor("attn_mask_19")]; + tensor var_1944_transpose_x_0 = const()[name = tensor("op_1944_transpose_x_0"), val = tensor(false)]; + tensor var_1944_transpose_y_0 = const()[name = tensor("op_1944_transpose_y_0"), val = tensor(false)]; + tensor transpose_26_perm_0 = const()[name = tensor("transpose_26_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_27_perm_0 = const()[name = tensor("transpose_27_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_27 = transpose(perm = transpose_27_perm_0, x = keys_27_cast_fp16)[name = tensor("transpose_35")]; + tensor transpose_26 = transpose(perm = transpose_26_perm_0, x = q_27_cast_fp16)[name = tensor("transpose_36")]; + tensor var_1944_cast_fp16 = matmul(transpose_x = var_1944_transpose_x_0, transpose_y = var_1944_transpose_y_0, x = transpose_26, y = transpose_27)[name = tensor("op_1944_cast_fp16")]; + tensor var_1938_to_fp16 = const()[name = tensor("op_1938_to_fp16"), val = tensor([0x1p-3])]; + tensor attn_weights_25_cast_fp16 = mul(x = var_1944_cast_fp16, y = var_1938_to_fp16)[name = tensor("attn_weights_25_cast_fp16")]; + tensor var_1946 = logical_not(x = attn_mask_19)[name = tensor("op_1946")]; + tensor var_1947_to_fp16 = const()[name = tensor("op_1947_to_fp16"), val = tensor(-inf)]; + tensor attn_weights_27_cast_fp16 = select(a = var_1947_to_fp16, b = attn_weights_25_cast_fp16, cond = var_1946)[name = tensor("attn_weights_27_cast_fp16")]; + tensor var_1949 = const()[name = tensor("op_1949"), val = tensor(-1)]; + tensor attn_weights_29_cast_fp16 = softmax(axis = var_1949, x = attn_weights_27_cast_fp16)[name = tensor("attn_weights_29_cast_fp16")]; + tensor attn_output_9_transpose_x_0 = const()[name = tensor("attn_output_9_transpose_x_0"), val = tensor(false)]; + tensor attn_output_9_transpose_y_0 = const()[name = tensor("attn_output_9_transpose_y_0"), val = tensor(false)]; + tensor values_29_cast_fp16 = transpose(perm = var_1889, x = values_27_cast_fp16)[name = tensor("transpose_37")]; + tensor attn_output_9_cast_fp16 = matmul(transpose_x = attn_output_9_transpose_x_0, transpose_y = attn_output_9_transpose_y_0, x = attn_weights_29_cast_fp16, y = values_29_cast_fp16)[name = tensor("attn_output_9_cast_fp16")]; + tensor var_1957 = const()[name = tensor("op_1957"), val = tensor([0, 2, 1, 3])]; + tensor var_1962 = const()[name = tensor("op_1962"), val = tensor([1, 1, 1024])]; + tensor var_1958_cast_fp16 = transpose(perm = var_1957, x = attn_output_9_cast_fp16)[name = tensor("transpose_34")]; + tensor input_55_cast_fp16 = reshape(shape = var_1962, x = var_1958_cast_fp16)[name = tensor("input_55_cast_fp16")]; + tensor attn4_out_proj_weight_to_fp16 = const()[name = tensor("attn4_out_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(108126464)))]; + tensor linear_18_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = attn4_out_proj_weight_to_fp16, x = input_55_cast_fp16)[name = tensor("linear_18_cast_fp16")]; + tensor var_1968_to_fp16 = const()[name = tensor("op_1968_to_fp16"), val = tensor(0x1p+0)]; + tensor var_1969_cast_fp16 = add(x = position4_to_fp16, y = var_1968_to_fp16)[name = tensor("op_1969_cast_fp16")]; + tensor var_1969_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_1969_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor input_57_cast_fp16 = add(x = input_51_cast_fp16, y = linear_18_cast_fp16)[name = tensor("input_57_cast_fp16")]; + tensor input_59_axes_0 = const()[name = tensor("input_59_axes_0"), val = tensor([-1])]; + tensor norm4_2_weight_to_fp16 = const()[name = tensor("norm4_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(110223680)))]; + tensor norm4_2_bias_to_fp16 = const()[name = tensor("norm4_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(110225792)))]; + tensor var_1973_to_fp16 = const()[name = tensor("op_1973_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_59_cast_fp16 = layer_norm(axes = input_59_axes_0, beta = norm4_2_bias_to_fp16, epsilon = var_1973_to_fp16, gamma = norm4_2_weight_to_fp16, x = input_57_cast_fp16)[name = tensor("input_59_cast_fp16")]; + tensor linear4_1_weight_to_fp16 = const()[name = tensor("linear4_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(110227904)))]; + tensor linear_19_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = linear4_1_weight_to_fp16, x = input_59_cast_fp16)[name = tensor("linear_19_cast_fp16")]; + tensor input_61_mode_0 = const()[name = tensor("input_61_mode_0"), val = tensor("EXACT")]; + tensor input_61_cast_fp16 = gelu(mode = input_61_mode_0, x = linear_19_cast_fp16)[name = tensor("input_61_cast_fp16")]; + tensor linear4_2_weight_to_fp16 = const()[name = tensor("linear4_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(118616576)))]; + tensor linear_20_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = linear4_2_weight_to_fp16, x = input_61_cast_fp16)[name = tensor("linear_20_cast_fp16")]; + tensor input_63_cast_fp16 = add(x = input_57_cast_fp16, y = linear_20_cast_fp16)[name = tensor("input_63_cast_fp16")]; + tensor input_65_axes_0 = const()[name = tensor("input_65_axes_0"), val = tensor([-1])]; + tensor norm5_1_weight_to_fp16 = const()[name = tensor("norm5_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(127005248)))]; + tensor norm5_1_bias_to_fp16 = const()[name = tensor("norm5_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(127007360)))]; + tensor var_1990_to_fp16 = const()[name = tensor("op_1990_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_65_cast_fp16 = layer_norm(axes = input_65_axes_0, beta = norm5_1_bias_to_fp16, epsilon = var_1990_to_fp16, gamma = norm5_1_weight_to_fp16, x = input_63_cast_fp16)[name = tensor("input_65_cast_fp16")]; + tensor attn5_in_proj_weight_to_fp16 = const()[name = tensor("attn5_in_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(127009472)))]; + tensor linear_21_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = attn5_in_proj_weight_to_fp16, x = input_65_cast_fp16)[name = tensor("linear_21_cast_fp16")]; + tensor var_2004 = const()[name = tensor("op_2004"), val = tensor([1, 1, 3, 16, 64])]; + tensor qkv_cast_fp16 = reshape(shape = var_2004, x = linear_21_cast_fp16)[name = tensor("qkv_cast_fp16")]; + tensor q_31_begin_0 = const()[name = tensor("q_31_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor q_31_end_0 = const()[name = tensor("q_31_end_0"), val = tensor([1, 1, 1, 16, 64])]; + tensor q_31_end_mask_0 = const()[name = tensor("q_31_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor q_31_squeeze_mask_0 = const()[name = tensor("q_31_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor q_31_cast_fp16 = slice_by_index(begin = q_31_begin_0, end = q_31_end_0, end_mask = q_31_end_mask_0, squeeze_mask = q_31_squeeze_mask_0, x = qkv_cast_fp16)[name = tensor("q_31_cast_fp16")]; + tensor k_21_begin_0 = const()[name = tensor("k_21_begin_0"), val = tensor([0, 0, 1, 0, 0])]; + tensor k_21_end_0 = const()[name = tensor("k_21_end_0"), val = tensor([1, 1, 2, 16, 64])]; + tensor k_21_end_mask_0 = const()[name = tensor("k_21_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor k_21_squeeze_mask_0 = const()[name = tensor("k_21_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor k_21_cast_fp16 = slice_by_index(begin = k_21_begin_0, end = k_21_end_0, end_mask = k_21_end_mask_0, squeeze_mask = k_21_squeeze_mask_0, x = qkv_cast_fp16)[name = tensor("k_21_cast_fp16")]; + tensor v_begin_0 = const()[name = tensor("v_begin_0"), val = tensor([0, 0, 2, 0, 0])]; + tensor v_end_0 = const()[name = tensor("v_end_0"), val = tensor([1, 1, 3, 16, 64])]; + tensor v_end_mask_0 = const()[name = tensor("v_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor v_squeeze_mask_0 = const()[name = tensor("v_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor v_cast_fp16 = slice_by_index(begin = v_begin_0, end = v_end_0, end_mask = v_end_mask_0, squeeze_mask = v_squeeze_mask_0, x = qkv_cast_fp16)[name = tensor("v_cast_fp16")]; + tensor var_2070 = const()[name = tensor("op_2070"), val = tensor([1, 1, 1, 1])]; + tensor position5_to_fp16_dtype_0 = const()[name = tensor("position5_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor position5_to_fp16 = cast(dtype = position5_to_fp16_dtype_0, x = position5)[name = tensor("cast_114")]; + tensor ts_cast_fp16 = reshape(shape = var_2070, x = position5_to_fp16)[name = tensor("ts_cast_fp16")]; + tensor var_2077 = const()[name = tensor("op_2077"), val = tensor([1, 1, 16, 32, 2])]; + tensor q_complex_cast_fp16 = reshape(shape = var_2077, x = q_31_cast_fp16)[name = tensor("q_complex_cast_fp16")]; + tensor var_2084 = const()[name = tensor("op_2084"), val = tensor([1, 1, 16, 32, 2])]; + tensor k_complex_cast_fp16 = reshape(shape = var_2084, x = k_21_cast_fp16)[name = tensor("k_complex_cast_fp16")]; + tensor var_2088_begin_0 = const()[name = tensor("op_2088_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_2088_end_0 = const()[name = tensor("op_2088_end_0"), val = tensor([1, 1, 16, 32, 1])]; + tensor var_2088_end_mask_0 = const()[name = tensor("op_2088_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_2088_squeeze_mask_0 = const()[name = tensor("op_2088_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_2088_cast_fp16 = slice_by_index(begin = var_2088_begin_0, end = var_2088_end_0, end_mask = var_2088_end_mask_0, squeeze_mask = var_2088_squeeze_mask_0, x = q_complex_cast_fp16)[name = tensor("op_2088_cast_fp16")]; + tensor var_2096_begin_0 = const()[name = tensor("op_2096_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_2096_end_0 = const()[name = tensor("op_2096_end_0"), val = tensor([1, 1, 16, 32, 2])]; + tensor var_2096_end_mask_0 = const()[name = tensor("op_2096_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_2096_squeeze_mask_0 = const()[name = tensor("op_2096_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_2096_cast_fp16 = slice_by_index(begin = var_2096_begin_0, end = var_2096_end_0, end_mask = var_2096_end_mask_0, squeeze_mask = var_2096_squeeze_mask_0, x = q_complex_cast_fp16)[name = tensor("op_2096_cast_fp16")]; + tensor var_2104_begin_0 = const()[name = tensor("op_2104_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_2104_end_0 = const()[name = tensor("op_2104_end_0"), val = tensor([1, 1, 16, 32, 1])]; + tensor var_2104_end_mask_0 = const()[name = tensor("op_2104_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_2104_squeeze_mask_0 = const()[name = tensor("op_2104_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_2104_cast_fp16 = slice_by_index(begin = var_2104_begin_0, end = var_2104_end_0, end_mask = var_2104_end_mask_0, squeeze_mask = var_2104_squeeze_mask_0, x = k_complex_cast_fp16)[name = tensor("op_2104_cast_fp16")]; + tensor var_2112_begin_0 = const()[name = tensor("op_2112_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_2112_end_0 = const()[name = tensor("op_2112_end_0"), val = tensor([1, 1, 16, 32, 2])]; + tensor var_2112_end_mask_0 = const()[name = tensor("op_2112_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_2112_squeeze_mask_0 = const()[name = tensor("op_2112_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_2112_cast_fp16 = slice_by_index(begin = var_2112_begin_0, end = var_2112_end_0, end_mask = var_2112_end_mask_0, squeeze_mask = var_2112_squeeze_mask_0, x = k_complex_cast_fp16)[name = tensor("op_2112_cast_fp16")]; + tensor freqs_to_fp16 = const()[name = tensor("freqs_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(133300992)))]; + tensor var_2118_cast_fp16 = mul(x = freqs_to_fp16, y = ts_cast_fp16)[name = tensor("op_2118_cast_fp16")]; + tensor rotr_cast_fp16 = cos(x = var_2118_cast_fp16)[name = tensor("rotr_cast_fp16")]; + tensor roti_cast_fp16 = sin(x = var_2118_cast_fp16)[name = tensor("roti_cast_fp16")]; + tensor var_2122_cast_fp16 = mul(x = var_2088_cast_fp16, y = rotr_cast_fp16)[name = tensor("op_2122_cast_fp16")]; + tensor var_2123_cast_fp16 = mul(x = var_2096_cast_fp16, y = roti_cast_fp16)[name = tensor("op_2123_cast_fp16")]; + tensor qor_21_cast_fp16 = sub(x = var_2122_cast_fp16, y = var_2123_cast_fp16)[name = tensor("qor_21_cast_fp16")]; + tensor var_2126_cast_fp16 = mul(x = var_2088_cast_fp16, y = roti_cast_fp16)[name = tensor("op_2126_cast_fp16")]; + tensor var_2127_cast_fp16 = mul(x = var_2096_cast_fp16, y = rotr_cast_fp16)[name = tensor("op_2127_cast_fp16")]; + tensor qoi_21_cast_fp16 = add(x = var_2126_cast_fp16, y = var_2127_cast_fp16)[name = tensor("qoi_21_cast_fp16")]; + tensor var_2130_cast_fp16 = mul(x = var_2104_cast_fp16, y = rotr_cast_fp16)[name = tensor("op_2130_cast_fp16")]; + tensor var_2131_cast_fp16 = mul(x = var_2112_cast_fp16, y = roti_cast_fp16)[name = tensor("op_2131_cast_fp16")]; + tensor kor_21_cast_fp16 = sub(x = var_2130_cast_fp16, y = var_2131_cast_fp16)[name = tensor("kor_21_cast_fp16")]; + tensor var_2134_cast_fp16 = mul(x = var_2104_cast_fp16, y = roti_cast_fp16)[name = tensor("op_2134_cast_fp16")]; + tensor var_2135_cast_fp16 = mul(x = var_2112_cast_fp16, y = rotr_cast_fp16)[name = tensor("op_2135_cast_fp16")]; + tensor koi_21_cast_fp16 = add(x = var_2134_cast_fp16, y = var_2135_cast_fp16)[name = tensor("koi_21_cast_fp16")]; + tensor qo_axis_0 = const()[name = tensor("qo_axis_0"), val = tensor(-1)]; + tensor qo_cast_fp16 = stack(axis = qo_axis_0, values = (qor_21_cast_fp16, qoi_21_cast_fp16))[name = tensor("qo_cast_fp16")]; + tensor ko_axis_0 = const()[name = tensor("ko_axis_0"), val = tensor(-1)]; + tensor ko_cast_fp16 = stack(axis = ko_axis_0, values = (kor_21_cast_fp16, koi_21_cast_fp16))[name = tensor("ko_cast_fp16")]; + tensor var_2168 = const()[name = tensor("op_2168"), val = tensor([1, 1, 16, 64])]; + tensor q_33_cast_fp16 = reshape(shape = var_2168, x = qo_cast_fp16)[name = tensor("q_33_cast_fp16")]; + tensor var_2174 = const()[name = tensor("op_2174"), val = tensor([1, 1, 16, 64])]; + tensor k_cast_fp16 = reshape(shape = var_2174, x = ko_cast_fp16)[name = tensor("k_cast_fp16")]; + tensor _inversed_2199_y_0_to_fp16 = const()[name = tensor("_inversed_2199_y_0_to_fp16"), val = tensor(0x1p-9)]; + tensor _inversed_2199_cast_fp16 = mul(x = ts_cast_fp16, y = _inversed_2199_y_0_to_fp16)[name = tensor("_inversed_2199_cast_fp16")]; + tensor var_2200_cast_fp16 = floor(x = _inversed_2199_cast_fp16)[name = tensor("op_2200_cast_fp16")]; + tensor var_2201_to_fp16 = const()[name = tensor("op_2201_to_fp16"), val = tensor(0x1p+9)]; + tensor var_2202_cast_fp16 = mul(x = var_2200_cast_fp16, y = var_2201_to_fp16)[name = tensor("op_2202_cast_fp16")]; + tensor write_indices_float_cast_fp16 = sub(x = ts_cast_fp16, y = var_2202_cast_fp16)[name = tensor("write_indices_float_cast_fp16")]; + tensor var_2209_dtype_0 = const()[name = tensor("op_2209_dtype_0"), val = tensor("int32")]; + tensor write_indices_reps_0 = const()[name = tensor("write_indices_reps_0"), val = tensor([1, 1, 16, 64])]; + tensor write_indices_float_cast_fp16_to_int32 = cast(dtype = var_2209_dtype_0, x = write_indices_float_cast_fp16)[name = tensor("cast_113")]; + tensor write_indices = tile(reps = write_indices_reps_0, x = write_indices_float_cast_fp16_to_int32)[name = tensor("write_indices")]; + tensor var_2219_begin_0 = const()[name = tensor("op_2219_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_2219_end_0 = const()[name = tensor("op_2219_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor var_2219_end_mask_0 = const()[name = tensor("op_2219_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_2219_squeeze_mask_0 = const()[name = tensor("op_2219_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor cache5_to_fp16_dtype_0 = const()[name = tensor("cache5_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor cache5_to_fp16 = cast(dtype = cache5_to_fp16_dtype_0, x = cache5)[name = tensor("cast_112")]; + tensor var_2219_cast_fp16 = slice_by_index(begin = var_2219_begin_0, end = var_2219_end_0, end_mask = var_2219_end_mask_0, squeeze_mask = var_2219_squeeze_mask_0, x = cache5_to_fp16)[name = tensor("op_2219_cast_fp16")]; + tensor var_2221_axis_0 = const()[name = tensor("op_2221_axis_0"), val = tensor(1)]; + tensor var_2221_mode_0 = const()[name = tensor("op_2221_mode_0"), val = tensor("update")]; + tensor var_2221_validate_indices_0 = const()[name = tensor("op_2221_validate_indices_0"), val = tensor(false)]; + tensor var_2221_cast_fp16 = scatter_along_axis(axis = var_2221_axis_0, data = var_2219_cast_fp16, indices = write_indices, mode = var_2221_mode_0, updates = k_cast_fp16, validate_indices = var_2221_validate_indices_0)[name = tensor("op_2221_cast_fp16")]; + tensor concat_37 = const()[name = tensor("concat_37"), val = tensor([0, 0, 0, 0, 0])]; + tensor concat_38 = const()[name = tensor("concat_38"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_internal_tensor_assign_1_stride_0 = const()[name = tensor("new_cache_internal_tensor_assign_1_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_internal_tensor_assign_1_begin_mask_0 = const()[name = tensor("new_cache_internal_tensor_assign_1_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_internal_tensor_assign_1_end_mask_0 = const()[name = tensor("new_cache_internal_tensor_assign_1_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_internal_tensor_assign_1_squeeze_mask_0 = const()[name = tensor("new_cache_internal_tensor_assign_1_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_22 = const()[name = tensor("shape_22"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_10 = const()[name = tensor("reduce_prod_10"), val = tensor(1048576)]; + tensor range_1d_10_start_0 = const()[name = tensor("range_1d_10_start_0"), val = tensor(0)]; + tensor range_1d_10_step_0 = const()[name = tensor("range_1d_10_step_0"), val = tensor(1)]; + tensor range_1d_10 = range_1d(end = reduce_prod_10, start = range_1d_10_start_0, step = range_1d_10_step_0)[name = tensor("range_1d_10")]; + tensor reshape_50 = reshape(shape = shape_22, x = range_1d_10)[name = tensor("reshape_50")]; + tensor slice_by_index_10 = slice_by_index(begin = concat_37, begin_mask = new_cache_internal_tensor_assign_1_begin_mask_0, end = concat_38, end_mask = new_cache_internal_tensor_assign_1_end_mask_0, squeeze_mask = new_cache_internal_tensor_assign_1_squeeze_mask_0, stride = new_cache_internal_tensor_assign_1_stride_0, x = reshape_50)[name = tensor("slice_by_index_10")]; + tensor reshape_51_shape_0 = const()[name = tensor("reshape_51_shape_0"), val = tensor([-1])]; + tensor reshape_51 = reshape(shape = reshape_51_shape_0, x = slice_by_index_10)[name = tensor("reshape_51")]; + tensor reshape_52_shape_0 = const()[name = tensor("reshape_52_shape_0"), val = tensor([-1])]; + tensor reshape_52_cast_fp16 = reshape(shape = reshape_52_shape_0, x = var_2221_cast_fp16)[name = tensor("reshape_52_cast_fp16")]; + tensor reshape_53_shape_0 = const()[name = tensor("reshape_53_shape_0"), val = tensor([-1])]; + tensor reshape_53_cast_fp16 = reshape(shape = reshape_53_shape_0, x = cache5_to_fp16)[name = tensor("reshape_53_cast_fp16")]; + tensor scatter_10_mode_0 = const()[name = tensor("scatter_10_mode_0"), val = tensor("update")]; + tensor scatter_10_axis_0 = const()[name = tensor("scatter_10_axis_0"), val = tensor(0)]; + tensor scatter_10_validate_indices_0 = const()[name = tensor("scatter_10_validate_indices_0"), val = tensor(false)]; + tensor scatter_10_cast_fp16 = scatter(axis = scatter_10_axis_0, data = reshape_53_cast_fp16, indices = reshape_51, mode = scatter_10_mode_0, updates = reshape_52_cast_fp16, validate_indices = scatter_10_validate_indices_0)[name = tensor("scatter_10_cast_fp16")]; + tensor reshape_54_cast_fp16 = reshape(shape = shape_22, x = scatter_10_cast_fp16)[name = tensor("reshape_54_cast_fp16")]; + tensor var_2229_begin_0 = const()[name = tensor("op_2229_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor var_2229_end_0 = const()[name = tensor("op_2229_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor var_2229_end_mask_0 = const()[name = tensor("op_2229_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_2229_squeeze_mask_0 = const()[name = tensor("op_2229_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor var_2229_cast_fp16 = slice_by_index(begin = var_2229_begin_0, end = var_2229_end_0, end_mask = var_2229_end_mask_0, squeeze_mask = var_2229_squeeze_mask_0, x = reshape_54_cast_fp16)[name = tensor("op_2229_cast_fp16")]; + tensor var_2231_axis_0 = const()[name = tensor("op_2231_axis_0"), val = tensor(1)]; + tensor var_2231_mode_0 = const()[name = tensor("op_2231_mode_0"), val = tensor("update")]; + tensor var_2231_validate_indices_0 = const()[name = tensor("op_2231_validate_indices_0"), val = tensor(false)]; + tensor var_2231_cast_fp16 = scatter_along_axis(axis = var_2231_axis_0, data = var_2229_cast_fp16, indices = write_indices, mode = var_2231_mode_0, updates = v_cast_fp16, validate_indices = var_2231_validate_indices_0)[name = tensor("op_2231_cast_fp16")]; + tensor concat_39 = const()[name = tensor("concat_39"), val = tensor([1, 0, 0, 0, 0])]; + tensor concat_40 = const()[name = tensor("concat_40"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_internal_tensor_assign_2_stride_0 = const()[name = tensor("new_cache_internal_tensor_assign_2_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_internal_tensor_assign_2_begin_mask_0 = const()[name = tensor("new_cache_internal_tensor_assign_2_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_internal_tensor_assign_2_end_mask_0 = const()[name = tensor("new_cache_internal_tensor_assign_2_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_internal_tensor_assign_2_squeeze_mask_0 = const()[name = tensor("new_cache_internal_tensor_assign_2_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_23 = const()[name = tensor("shape_23"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_11 = const()[name = tensor("reduce_prod_11"), val = tensor(1048576)]; + tensor range_1d_11_start_0 = const()[name = tensor("range_1d_11_start_0"), val = tensor(0)]; + tensor range_1d_11_step_0 = const()[name = tensor("range_1d_11_step_0"), val = tensor(1)]; + tensor range_1d_11 = range_1d(end = reduce_prod_11, start = range_1d_11_start_0, step = range_1d_11_step_0)[name = tensor("range_1d_11")]; + tensor reshape_55 = reshape(shape = shape_23, x = range_1d_11)[name = tensor("reshape_55")]; + tensor slice_by_index_11 = slice_by_index(begin = concat_39, begin_mask = new_cache_internal_tensor_assign_2_begin_mask_0, end = concat_40, end_mask = new_cache_internal_tensor_assign_2_end_mask_0, squeeze_mask = new_cache_internal_tensor_assign_2_squeeze_mask_0, stride = new_cache_internal_tensor_assign_2_stride_0, x = reshape_55)[name = tensor("slice_by_index_11")]; + tensor reshape_56_shape_0 = const()[name = tensor("reshape_56_shape_0"), val = tensor([-1])]; + tensor reshape_56 = reshape(shape = reshape_56_shape_0, x = slice_by_index_11)[name = tensor("reshape_56")]; + tensor reshape_57_shape_0 = const()[name = tensor("reshape_57_shape_0"), val = tensor([-1])]; + tensor reshape_57_cast_fp16 = reshape(shape = reshape_57_shape_0, x = var_2231_cast_fp16)[name = tensor("reshape_57_cast_fp16")]; + tensor reshape_58_shape_0 = const()[name = tensor("reshape_58_shape_0"), val = tensor([-1])]; + tensor reshape_58_cast_fp16 = reshape(shape = reshape_58_shape_0, x = reshape_54_cast_fp16)[name = tensor("reshape_58_cast_fp16")]; + tensor scatter_11_mode_0 = const()[name = tensor("scatter_11_mode_0"), val = tensor("update")]; + tensor scatter_11_axis_0 = const()[name = tensor("scatter_11_axis_0"), val = tensor(0)]; + tensor scatter_11_validate_indices_0 = const()[name = tensor("scatter_11_validate_indices_0"), val = tensor(false)]; + tensor scatter_11_cast_fp16 = scatter(axis = scatter_11_axis_0, data = reshape_58_cast_fp16, indices = reshape_56, mode = scatter_11_mode_0, updates = reshape_57_cast_fp16, validate_indices = scatter_11_validate_indices_0)[name = tensor("scatter_11_cast_fp16")]; + tensor reshape_59_cast_fp16 = reshape(shape = shape_23, x = scatter_11_cast_fp16)[name = tensor("reshape_59_cast_fp16")]; + tensor reshape_59_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("reshape_59_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor keys_31_begin_0 = const()[name = tensor("keys_31_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor keys_31_end_0 = const()[name = tensor("keys_31_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor keys_31_end_mask_0 = const()[name = tensor("keys_31_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor keys_31_squeeze_mask_0 = const()[name = tensor("keys_31_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor keys_31_cast_fp16 = slice_by_index(begin = keys_31_begin_0, end = keys_31_end_0, end_mask = keys_31_end_mask_0, squeeze_mask = keys_31_squeeze_mask_0, x = reshape_59_cast_fp16)[name = tensor("keys_31_cast_fp16")]; + tensor values_31_begin_0 = const()[name = tensor("values_31_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor values_31_end_0 = const()[name = tensor("values_31_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor values_31_end_mask_0 = const()[name = tensor("values_31_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor values_31_squeeze_mask_0 = const()[name = tensor("values_31_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor values_31_cast_fp16 = slice_by_index(begin = values_31_begin_0, end = values_31_end_0, end_mask = values_31_end_mask_0, squeeze_mask = values_31_squeeze_mask_0, x = reshape_59_cast_fp16)[name = tensor("values_31_cast_fp16")]; + tensor var_2243_cast_fp16 = not_equal(x = keys_31_cast_fp16, y = keys_31_cast_fp16)[name = tensor("op_2243_cast_fp16")]; + tensor keys_33_cast_fp16 = select(a = var_319_to_fp16, b = keys_31_cast_fp16, cond = var_2243_cast_fp16)[name = tensor("keys_33_cast_fp16")]; + tensor var_2251_cast_fp16 = not_equal(x = values_31_cast_fp16, y = values_31_cast_fp16)[name = tensor("op_2251_cast_fp16")]; + tensor values_33_cast_fp16 = select(a = var_319_to_fp16, b = values_31_cast_fp16, cond = var_2251_cast_fp16)[name = tensor("values_33_cast_fp16")]; + tensor var_2275 = const()[name = tensor("op_2275"), val = tensor([0, 2, 1, 3])]; + tensor var_2291 = const()[name = tensor("op_2291"), val = tensor([1, 1, 1])]; + tensor var_2292_cast_fp16 = reshape(shape = var_2291, x = position5_to_fp16)[name = tensor("op_2292_cast_fp16")]; + tensor var_2312_to_fp16 = const()[name = tensor("op_2312_to_fp16"), val = tensor(0x1p+0)]; + tensor valid_len_cast_fp16 = add(x = var_2292_cast_fp16, y = var_2312_to_fp16)[name = tensor("valid_len_cast_fp16")]; + tensor valid_mask_cast_fp16 = less(x = k_positions_1_promoted_to_fp16, y = valid_len_cast_fp16)[name = tensor("valid_mask_cast_fp16")]; + tensor causal_mask_cast_fp16 = less_equal(x = k_positions_1_promoted_to_fp16, y = var_2292_cast_fp16)[name = tensor("causal_mask_cast_fp16")]; + tensor attn_mask_21 = logical_and(x = valid_mask_cast_fp16, y = causal_mask_cast_fp16)[name = tensor("attn_mask_21")]; + tensor attn_mask_axes_0 = const()[name = tensor("attn_mask_axes_0"), val = tensor([1])]; + tensor attn_mask = expand_dims(axes = attn_mask_axes_0, x = attn_mask_21)[name = tensor("attn_mask")]; + tensor var_2330_transpose_x_0 = const()[name = tensor("op_2330_transpose_x_0"), val = tensor(false)]; + tensor var_2330_transpose_y_0 = const()[name = tensor("op_2330_transpose_y_0"), val = tensor(false)]; + tensor transpose_28_perm_0 = const()[name = tensor("transpose_28_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_29_perm_0 = const()[name = tensor("transpose_29_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_29 = transpose(perm = transpose_29_perm_0, x = keys_33_cast_fp16)[name = tensor("transpose_31")]; + tensor transpose_28 = transpose(perm = transpose_28_perm_0, x = q_33_cast_fp16)[name = tensor("transpose_32")]; + tensor var_2330_cast_fp16 = matmul(transpose_x = var_2330_transpose_x_0, transpose_y = var_2330_transpose_y_0, x = transpose_28, y = transpose_29)[name = tensor("op_2330_cast_fp16")]; + tensor var_2324_to_fp16 = const()[name = tensor("op_2324_to_fp16"), val = tensor([0x1p-3])]; + tensor attn_weights_31_cast_fp16 = mul(x = var_2330_cast_fp16, y = var_2324_to_fp16)[name = tensor("attn_weights_31_cast_fp16")]; + tensor var_2332 = logical_not(x = attn_mask)[name = tensor("op_2332")]; + tensor var_2333_to_fp16 = const()[name = tensor("op_2333_to_fp16"), val = tensor(-inf)]; + tensor attn_weights_33_cast_fp16 = select(a = var_2333_to_fp16, b = attn_weights_31_cast_fp16, cond = var_2332)[name = tensor("attn_weights_33_cast_fp16")]; + tensor var_2335 = const()[name = tensor("op_2335"), val = tensor(-1)]; + tensor attn_weights_cast_fp16 = softmax(axis = var_2335, x = attn_weights_33_cast_fp16)[name = tensor("attn_weights_cast_fp16")]; + tensor attn_output_transpose_x_0 = const()[name = tensor("attn_output_transpose_x_0"), val = tensor(false)]; + tensor attn_output_transpose_y_0 = const()[name = tensor("attn_output_transpose_y_0"), val = tensor(false)]; + tensor values_cast_fp16 = transpose(perm = var_2275, x = values_33_cast_fp16)[name = tensor("transpose_33")]; + tensor attn_output_cast_fp16 = matmul(transpose_x = attn_output_transpose_x_0, transpose_y = attn_output_transpose_y_0, x = attn_weights_cast_fp16, y = values_cast_fp16)[name = tensor("attn_output_cast_fp16")]; + tensor var_2343 = const()[name = tensor("op_2343"), val = tensor([0, 2, 1, 3])]; + tensor var_2348 = const()[name = tensor("op_2348"), val = tensor([1, 1, 1024])]; + tensor var_2344_cast_fp16 = transpose(perm = var_2343, x = attn_output_cast_fp16)[name = tensor("transpose_30")]; + tensor input_67_cast_fp16 = reshape(shape = var_2348, x = var_2344_cast_fp16)[name = tensor("input_67_cast_fp16")]; + tensor attn5_out_proj_weight_to_fp16 = const()[name = tensor("attn5_out_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(133301120)))]; + tensor linear_22_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = attn5_out_proj_weight_to_fp16, x = input_67_cast_fp16)[name = tensor("linear_22_cast_fp16")]; + tensor var_2354_to_fp16 = const()[name = tensor("op_2354_to_fp16"), val = tensor(0x1p+0)]; + tensor var_2355_cast_fp16 = add(x = position5_to_fp16, y = var_2354_to_fp16)[name = tensor("op_2355_cast_fp16")]; + tensor var_2355_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_2355_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor input_69_cast_fp16 = add(x = input_63_cast_fp16, y = linear_22_cast_fp16)[name = tensor("input_69_cast_fp16")]; + tensor input_71_axes_0 = const()[name = tensor("input_71_axes_0"), val = tensor([-1])]; + tensor norm5_2_weight_to_fp16 = const()[name = tensor("norm5_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(135398336)))]; + tensor norm5_2_bias_to_fp16 = const()[name = tensor("norm5_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(135400448)))]; + tensor var_2359_to_fp16 = const()[name = tensor("op_2359_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_71_cast_fp16 = layer_norm(axes = input_71_axes_0, beta = norm5_2_bias_to_fp16, epsilon = var_2359_to_fp16, gamma = norm5_2_weight_to_fp16, x = input_69_cast_fp16)[name = tensor("input_71_cast_fp16")]; + tensor linear5_1_weight_to_fp16 = const()[name = tensor("linear5_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(135402560)))]; + tensor linear_23_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = linear5_1_weight_to_fp16, x = input_71_cast_fp16)[name = tensor("linear_23_cast_fp16")]; + tensor input_73_mode_0 = const()[name = tensor("input_73_mode_0"), val = tensor("EXACT")]; + tensor input_73_cast_fp16 = gelu(mode = input_73_mode_0, x = linear_23_cast_fp16)[name = tensor("input_73_cast_fp16")]; + tensor linear5_2_weight_to_fp16 = const()[name = tensor("linear5_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(143791232)))]; + tensor linear_24_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = linear5_2_weight_to_fp16, x = input_73_cast_fp16)[name = tensor("linear_24_cast_fp16")]; + tensor input_75_cast_fp16 = add(x = input_69_cast_fp16, y = linear_24_cast_fp16)[name = tensor("input_75_cast_fp16")]; + tensor input_axes_0 = const()[name = tensor("input_axes_0"), val = tensor([-1])]; + tensor out_norm_weight_to_fp16 = const()[name = tensor("out_norm_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(152179904)))]; + tensor out_norm_bias_to_fp16 = const()[name = tensor("out_norm_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(152182016)))]; + tensor var_2376_to_fp16 = const()[name = tensor("op_2376_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_cast_fp16 = layer_norm(axes = input_axes_0, beta = out_norm_bias_to_fp16, epsilon = var_2376_to_fp16, gamma = out_norm_weight_to_fp16, x = input_75_cast_fp16)[name = tensor("input_cast_fp16")]; + tensor input_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("input_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor out_eos_weight_to_fp16 = const()[name = tensor("out_eos_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(152184128)))]; + tensor out_eos_bias_to_fp16 = const()[name = tensor("out_eos_bias_to_fp16"), val = tensor([-0x1.3ep-3])]; + tensor linear_25_cast_fp16 = linear(bias = out_eos_bias_to_fp16, weight = out_eos_weight_to_fp16, x = input_cast_fp16)[name = tensor("linear_25_cast_fp16")]; + tensor linear_25_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("linear_25_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor var_2384 = cast(dtype = linear_25_cast_fp16_to_fp32_dtype_0, x = linear_25_cast_fp16)[name = tensor("cast_108")]; + tensor input = cast(dtype = input_cast_fp16_to_fp32_dtype_0, x = input_cast_fp16)[name = tensor("cast_109")]; + tensor var_2355 = cast(dtype = var_2355_cast_fp16_to_fp32_dtype_0, x = var_2355_cast_fp16)[name = tensor("cast_110")]; + tensor new_cache_internal_tensor_assign_2 = cast(dtype = reshape_59_cast_fp16_to_fp32_dtype_0, x = reshape_59_cast_fp16)[name = tensor("cast_111")]; + tensor var_1969 = cast(dtype = var_1969_cast_fp16_to_fp32_dtype_0, x = var_1969_cast_fp16)[name = tensor("cast_115")]; + tensor new_cache_9_internal_tensor_assign_2 = cast(dtype = reshape_49_cast_fp16_to_fp32_dtype_0, x = reshape_49_cast_fp16)[name = tensor("cast_116")]; + tensor var_1583 = cast(dtype = var_1583_cast_fp16_to_fp32_dtype_0, x = var_1583_cast_fp16)[name = tensor("cast_120")]; + tensor new_cache_7_internal_tensor_assign_2 = cast(dtype = reshape_39_cast_fp16_to_fp32_dtype_0, x = reshape_39_cast_fp16)[name = tensor("cast_121")]; + tensor var_1197 = cast(dtype = var_1197_cast_fp16_to_fp32_dtype_0, x = var_1197_cast_fp16)[name = tensor("cast_125")]; + tensor new_cache_5_internal_tensor_assign_2 = cast(dtype = reshape_29_cast_fp16_to_fp32_dtype_0, x = reshape_29_cast_fp16)[name = tensor("cast_126")]; + tensor var_811 = cast(dtype = var_811_cast_fp16_to_fp32_dtype_0, x = var_811_cast_fp16)[name = tensor("cast_130")]; + tensor new_cache_3_internal_tensor_assign_2 = cast(dtype = reshape_19_cast_fp16_to_fp32_dtype_0, x = reshape_19_cast_fp16)[name = tensor("cast_131")]; + tensor var_425 = cast(dtype = var_425_cast_fp16_to_fp32_dtype_0, x = var_425_cast_fp16)[name = tensor("cast_135")]; + tensor new_cache_1_internal_tensor_assign_2 = cast(dtype = reshape_9_cast_fp16_to_fp32_dtype_0, x = reshape_9_cast_fp16)[name = tensor("cast_136")]; + } -> (input, var_2384, new_cache_1_internal_tensor_assign_2, var_425, new_cache_3_internal_tensor_assign_2, var_811, new_cache_5_internal_tensor_assign_2, var_1197, new_cache_7_internal_tensor_assign_2, var_1583, new_cache_9_internal_tensor_assign_2, var_1969, new_cache_internal_tensor_assign_2, var_2355); +} \ No newline at end of file diff --git a/v2.1/german/flowlm_step.mlmodelc/weights/weight.bin b/v2.1/german/flowlm_step.mlmodelc/weights/weight.bin new file mode 100644 index 0000000000000000000000000000000000000000..715d73b8b954166b22f25da6746a8484e9f1430b --- /dev/null +++ b/v2.1/german/flowlm_step.mlmodelc/weights/weight.bin @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:871e7e2668c661e77a9255b1fb7f89eba0a60181b6cf04df8a42e793ef7f80c5 +size 152186240 diff --git a/v2.1/german/flowlm_step.mlpackage/Data/com.apple.CoreML/model.mlmodel b/v2.1/german/flowlm_step.mlpackage/Data/com.apple.CoreML/model.mlmodel new file mode 100644 index 0000000000000000000000000000000000000000..1692a20c82b71e9d8c523cbc1ed02a5067c47e7d --- /dev/null +++ b/v2.1/german/flowlm_step.mlpackage/Data/com.apple.CoreML/model.mlmodel @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:206c2fd7a0ca0fdfe6dc5bc9a1149a50c14fa97efd569004de0a9575a1d5df6d +size 207618 diff --git a/v2.1/german/flowlm_step.mlpackage/Data/com.apple.CoreML/weights/weight.bin b/v2.1/german/flowlm_step.mlpackage/Data/com.apple.CoreML/weights/weight.bin new file mode 100644 index 0000000000000000000000000000000000000000..715d73b8b954166b22f25da6746a8484e9f1430b --- /dev/null +++ b/v2.1/german/flowlm_step.mlpackage/Data/com.apple.CoreML/weights/weight.bin @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:871e7e2668c661e77a9255b1fb7f89eba0a60181b6cf04df8a42e793ef7f80c5 +size 152186240 diff --git a/v2.1/german/flowlm_step.mlpackage/Manifest.json b/v2.1/german/flowlm_step.mlpackage/Manifest.json new file mode 100644 index 0000000000000000000000000000000000000000..70b12dcec55048b62f519f36f284fa2e0005c49d --- /dev/null +++ b/v2.1/german/flowlm_step.mlpackage/Manifest.json @@ -0,0 +1,18 @@ +{ + "fileFormatVersion": "1.0.0", + "itemInfoEntries": { + "6CB2175A-E4F6-418F-9445-5A3A275D69C5": { + "author": "com.apple.CoreML", + "description": "CoreML Model Specification", + "name": "model.mlmodel", + "path": "com.apple.CoreML/model.mlmodel" + }, + "BCE97DC5-7C1F-4C4A-8DC3-917096087887": { + "author": "com.apple.CoreML", + "description": "CoreML Model Weights", + "name": "weights", + "path": "com.apple.CoreML/weights" + } + }, + "rootModelIdentifier": "6CB2175A-E4F6-418F-9445-5A3A275D69C5" +} diff --git a/v2.1/german/flowlm_stepv2.mlmodelc/analytics/coremldata.bin b/v2.1/german/flowlm_stepv2.mlmodelc/analytics/coremldata.bin new file mode 100644 index 0000000000000000000000000000000000000000..4e73c22e6fa8ca2fec3407014480503305f78da8 --- /dev/null +++ b/v2.1/german/flowlm_stepv2.mlmodelc/analytics/coremldata.bin @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:d1e11e84fd25af0c128df2e80eefd43586d0e81742e88cc5249c66cef7d87467 +size 243 diff --git a/v2.1/german/flowlm_stepv2.mlmodelc/coremldata.bin b/v2.1/german/flowlm_stepv2.mlmodelc/coremldata.bin new file mode 100644 index 0000000000000000000000000000000000000000..b9a24a8c4f016d7e80426c1c877dff5ac3ba3597 --- /dev/null +++ b/v2.1/german/flowlm_stepv2.mlmodelc/coremldata.bin @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:77dc3ac8949371f41fa4be5037195d1543adc3b75af750e92ce58aa2fe8485dd +size 1190 diff --git a/v2.1/german/flowlm_stepv2.mlmodelc/metadata.json b/v2.1/german/flowlm_stepv2.mlmodelc/metadata.json new file mode 100644 index 0000000000000000000000000000000000000000..f2969c83e1cf7a6025c56eb3502b7ec456a89630 --- /dev/null +++ b/v2.1/german/flowlm_stepv2.mlmodelc/metadata.json @@ -0,0 +1,348 @@ +[ + { + "metadataOutputVersion" : "3.0", + "storagePrecision" : "Mixed (Float16, Int8)", + "outputSchema" : [ + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1 × 1 × 1024)", + "shortDescription" : "", + "shape" : "[1, 1, 1024]", + "name" : "input", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1 × 1 × 1)", + "shortDescription" : "", + "shape" : "[1, 1, 1]", + "name" : "var_2582", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 2 × 1 × 512 × 16 × 64)", + "shortDescription" : "", + "shape" : "[2, 1, 512, 16, 64]", + "name" : "new_cache_1_internal_tensor_assign_2", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1)", + "shortDescription" : "", + "shape" : "[1]", + "name" : "var_458", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 2 × 1 × 512 × 16 × 64)", + "shortDescription" : "", + "shape" : "[2, 1, 512, 16, 64]", + "name" : "new_cache_3_internal_tensor_assign_2", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1)", + "shortDescription" : "", + "shape" : "[1]", + "name" : "var_877", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 2 × 1 × 512 × 16 × 64)", + "shortDescription" : "", + "shape" : "[2, 1, 512, 16, 64]", + "name" : "new_cache_5_internal_tensor_assign_2", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1)", + "shortDescription" : "", + "shape" : "[1]", + "name" : "var_1296", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 2 × 1 × 512 × 16 × 64)", + "shortDescription" : "", + "shape" : "[2, 1, 512, 16, 64]", + "name" : "new_cache_7_internal_tensor_assign_2", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1)", + "shortDescription" : "", + "shape" : "[1]", + "name" : "var_1715", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 2 × 1 × 512 × 16 × 64)", + "shortDescription" : "", + "shape" : "[2, 1, 512, 16, 64]", + "name" : "new_cache_9_internal_tensor_assign_2", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1)", + "shortDescription" : "", + "shape" : "[1]", + "name" : "var_2134", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 2 × 1 × 512 × 16 × 64)", + "shortDescription" : "", + "shape" : "[2, 1, 512, 16, 64]", + "name" : "new_cache_internal_tensor_assign_2", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1)", + "shortDescription" : "", + "shape" : "[1]", + "name" : "var_2553", + "type" : "MultiArray" + } + ], + "modelParameters" : [ + + ], + "specificationVersion" : 8, + "mlProgramOperationTypeHistogram" : { + "Ios17.cos" : 6, + "Range1d" : 12, + "Ios17.notEqual" : 13, + "Ios17.floor" : 6, + "Ios17.reshape" : 108, + "Ios17.logicalAnd" : 6, + "Ios16.softmax" : 6, + "Ios17.scatter" : 12, + "Ios17.matmul" : 12, + "Ios17.sin" : 6, + "Ios17.transpose" : 24, + "Select" : 19, + "Ios17.expandDims" : 7, + "Ios17.add" : 36, + "Tile" : 6, + "Ios17.sliceByIndex" : 78, + "Ios17.lessEqual" : 6, + "Ios17.scatterAlongAxis" : 12, + "Ios17.logicalNot" : 6, + "Ios17.layerNorm" : 13, + "Ios16.constexprAffineDequantize" : 24, + "Ios17.cast" : 34, + "Ios17.sub" : 18, + "Ios17.less" : 6, + "Ios16.gelu" : 6, + "Ios17.linear" : 26, + "Stack" : 12, + "Ios17.mul" : 72 + }, + "computePrecision" : "Mixed (Float16, Float32, Int32)", + "isUpdatable" : "0", + "stateSchema" : [ + + ], + "availability" : { + "macOS" : "14.0", + "tvOS" : "17.0", + "visionOS" : "1.0", + "watchOS" : "10.0", + "iOS" : "17.0", + "macCatalyst" : "17.0" + }, + "modelType" : { + "name" : "MLModelType_mlProgram" + }, + "userDefinedMetadata" : { + "com.github.apple.coremltools.conversion_date" : "2026-04-27", + "com.github.apple.coremltools.source" : "torch==2.9.1", + "com.github.apple.coremltools.version" : "9.0", + "com.github.apple.coremltools.source_dialect" : "TorchScript" + }, + "inputSchema" : [ + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1 × 1 × 32)", + "shortDescription" : "", + "shape" : "[1, 1, 32]", + "name" : "sequence", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 32)", + "shortDescription" : "", + "shape" : "[32]", + "name" : "bos_emb", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 2 × 1 × 512 × 16 × 64)", + "shortDescription" : "", + "shape" : "[2, 1, 512, 16, 64]", + "name" : "cache0", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1)", + "shortDescription" : "", + "shape" : "[1]", + "name" : "position0", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 2 × 1 × 512 × 16 × 64)", + "shortDescription" : "", + "shape" : "[2, 1, 512, 16, 64]", + "name" : "cache1", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1)", + "shortDescription" : "", + "shape" : "[1]", + "name" : "position1", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 2 × 1 × 512 × 16 × 64)", + "shortDescription" : "", + "shape" : "[2, 1, 512, 16, 64]", + "name" : "cache2", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1)", + "shortDescription" : "", + "shape" : "[1]", + "name" : "position2", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 2 × 1 × 512 × 16 × 64)", + "shortDescription" : "", + "shape" : "[2, 1, 512, 16, 64]", + "name" : "cache3", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1)", + "shortDescription" : "", + "shape" : "[1]", + "name" : "position3", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 2 × 1 × 512 × 16 × 64)", + "shortDescription" : "", + "shape" : "[2, 1, 512, 16, 64]", + "name" : "cache4", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1)", + "shortDescription" : "", + "shape" : "[1]", + "name" : "position4", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 2 × 1 × 512 × 16 × 64)", + "shortDescription" : "", + "shape" : "[2, 1, 512, 16, 64]", + "name" : "cache5", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1)", + "shortDescription" : "", + "shape" : "[1]", + "name" : "position5", + "type" : "MultiArray" + } + ], + "generatedClassName" : "flowlm_stepv2", + "method" : "predict" + } +] \ No newline at end of file diff --git a/v2.1/german/flowlm_stepv2.mlmodelc/model.mil b/v2.1/german/flowlm_stepv2.mlmodelc/model.mil new file mode 100644 index 0000000000000000000000000000000000000000..611248bd4e1807461954fe936b4b87f70ddb2725 --- /dev/null +++ b/v2.1/german/flowlm_stepv2.mlmodelc/model.mil @@ -0,0 +1,1359 @@ +program(1.0) +[buildInfo = dict, tensor>({{"coremlc-component-MIL", "3520.4.1"}, {"coremlc-version", "3520.5.1"}, {"coremltools-component-torch", "2.9.1"}, {"coremltools-source-dialect", "TorchScript"}, {"coremltools-version", "9.0"}})] +{ + func main(tensor bos_emb, tensor cache0, tensor cache1, tensor cache2, tensor cache3, tensor cache4, tensor cache5, tensor position0, tensor position1, tensor position2, tensor position3, tensor position4, tensor position5, tensor sequence) { + tensor attn0_in_proj_weight_cast_fp16 = constexpr_affine_dequantize()[axis = tensor(0), name = tensor("attn0_in_proj_weight_cast_fp16"), quantized_data = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(64))), scale = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(3148992))), zero_point = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(3145856)))]; + tensor attn0_out_proj_weight_cast_fp16 = constexpr_affine_dequantize()[axis = tensor(0), name = tensor("attn0_out_proj_weight_cast_fp16"), quantized_data = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(3155200))), scale = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(4204928))), zero_point = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(4203840)))]; + tensor linear0_1_weight_cast_fp16 = constexpr_affine_dequantize()[axis = tensor(0), name = tensor("linear0_1_weight_cast_fp16"), quantized_data = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(4207040))), scale = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8405568))), zero_point = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8401408)))]; + tensor linear0_2_weight_cast_fp16 = constexpr_affine_dequantize()[axis = tensor(0), name = tensor("linear0_2_weight_cast_fp16"), quantized_data = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8413824))), scale = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(12608192))), zero_point = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(4203840)))]; + tensor attn1_in_proj_weight_cast_fp16 = constexpr_affine_dequantize()[axis = tensor(0), name = tensor("attn1_in_proj_weight_cast_fp16"), quantized_data = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(12610304))), scale = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15756096))), zero_point = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(3145856)))]; + tensor attn1_out_proj_weight_cast_fp16 = constexpr_affine_dequantize()[axis = tensor(0), name = tensor("attn1_out_proj_weight_cast_fp16"), quantized_data = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15762304))), scale = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(16810944))), zero_point = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(4203840)))]; + tensor linear1_1_weight_cast_fp16 = constexpr_affine_dequantize()[axis = tensor(0), name = tensor("linear1_1_weight_cast_fp16"), quantized_data = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(16813056))), scale = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21007424))), zero_point = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8401408)))]; + tensor linear1_2_weight_cast_fp16 = constexpr_affine_dequantize()[axis = tensor(0), name = tensor("linear1_2_weight_cast_fp16"), quantized_data = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21015680))), scale = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(25210048))), zero_point = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(4203840)))]; + tensor attn2_in_proj_weight_cast_fp16 = constexpr_affine_dequantize()[axis = tensor(0), name = tensor("attn2_in_proj_weight_cast_fp16"), quantized_data = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(25212160))), scale = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28357952))), zero_point = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(3145856)))]; + tensor attn2_out_proj_weight_cast_fp16 = constexpr_affine_dequantize()[axis = tensor(0), name = tensor("attn2_out_proj_weight_cast_fp16"), quantized_data = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28364160))), scale = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(29412800))), zero_point = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(4203840)))]; + tensor linear2_1_weight_cast_fp16 = constexpr_affine_dequantize()[axis = tensor(0), name = tensor("linear2_1_weight_cast_fp16"), quantized_data = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(29414912))), scale = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(33609280))), zero_point = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8401408)))]; + tensor linear2_2_weight_cast_fp16 = constexpr_affine_dequantize()[axis = tensor(0), name = tensor("linear2_2_weight_cast_fp16"), quantized_data = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(33617536))), scale = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(37811904))), zero_point = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(4203840)))]; + tensor attn3_in_proj_weight_cast_fp16 = constexpr_affine_dequantize()[axis = tensor(0), name = tensor("attn3_in_proj_weight_cast_fp16"), quantized_data = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(37814016))), scale = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(40959808))), zero_point = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(3145856)))]; + tensor attn3_out_proj_weight_cast_fp16 = constexpr_affine_dequantize()[axis = tensor(0), name = tensor("attn3_out_proj_weight_cast_fp16"), quantized_data = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(40966016))), scale = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(42014656))), zero_point = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(4203840)))]; + tensor linear3_1_weight_cast_fp16 = constexpr_affine_dequantize()[axis = tensor(0), name = tensor("linear3_1_weight_cast_fp16"), quantized_data = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(42016768))), scale = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(46211136))), zero_point = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8401408)))]; + tensor linear3_2_weight_cast_fp16 = constexpr_affine_dequantize()[axis = tensor(0), name = tensor("linear3_2_weight_cast_fp16"), quantized_data = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(46219392))), scale = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(50413760))), zero_point = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(4203840)))]; + tensor attn4_in_proj_weight_cast_fp16 = constexpr_affine_dequantize()[axis = tensor(0), name = tensor("attn4_in_proj_weight_cast_fp16"), quantized_data = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(50415872))), scale = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(53561664))), zero_point = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(3145856)))]; + tensor attn4_out_proj_weight_cast_fp16 = constexpr_affine_dequantize()[axis = tensor(0), name = tensor("attn4_out_proj_weight_cast_fp16"), quantized_data = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(53567872))), scale = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(54616512))), zero_point = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(4203840)))]; + tensor linear4_1_weight_cast_fp16 = constexpr_affine_dequantize()[axis = tensor(0), name = tensor("linear4_1_weight_cast_fp16"), quantized_data = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(54618624))), scale = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(58812992))), zero_point = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8401408)))]; + tensor linear4_2_weight_cast_fp16 = constexpr_affine_dequantize()[axis = tensor(0), name = tensor("linear4_2_weight_cast_fp16"), quantized_data = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(58821248))), scale = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(63015616))), zero_point = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(4203840)))]; + tensor attn5_in_proj_weight_cast_fp16 = constexpr_affine_dequantize()[axis = tensor(0), name = tensor("attn5_in_proj_weight_cast_fp16"), quantized_data = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(63017728))), scale = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(66163520))), zero_point = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(3145856)))]; + tensor attn5_out_proj_weight_cast_fp16 = constexpr_affine_dequantize()[axis = tensor(0), name = tensor("attn5_out_proj_weight_cast_fp16"), quantized_data = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(66169728))), scale = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(67218368))), zero_point = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(4203840)))]; + tensor linear5_1_weight_cast_fp16 = constexpr_affine_dequantize()[axis = tensor(0), name = tensor("linear5_1_weight_cast_fp16"), quantized_data = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(67220480))), scale = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(71414848))), zero_point = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8401408)))]; + tensor linear5_2_weight_cast_fp16 = constexpr_affine_dequantize()[axis = tensor(0), name = tensor("linear5_2_weight_cast_fp16"), quantized_data = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(71423104))), scale = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(75617472))), zero_point = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(4203840)))]; + tensor sequence_to_fp16_dtype_0 = const()[name = tensor("sequence_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor sequence_to_fp16 = cast(dtype = sequence_to_fp16_dtype_0, x = sequence)[name = tensor("cast_141")]; + tensor var_54_cast_fp16 = not_equal(x = sequence_to_fp16, y = sequence_to_fp16)[name = tensor("op_54_cast_fp16")]; + tensor expand_dims_0_axes_0 = const()[name = tensor("expand_dims_0_axes_0"), val = tensor([0, 1])]; + tensor bos_emb_to_fp16_dtype_0 = const()[name = tensor("bos_emb_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor bos_emb_to_fp16 = cast(dtype = bos_emb_to_fp16_dtype_0, x = bos_emb)[name = tensor("cast_140")]; + tensor expand_dims_0_cast_fp16 = expand_dims(axes = expand_dims_0_axes_0, x = bos_emb_to_fp16)[name = tensor("expand_dims_0_cast_fp16")]; + tensor input_1_cast_fp16 = select(a = expand_dims_0_cast_fp16, b = sequence_to_fp16, cond = var_54_cast_fp16)[name = tensor("input_1_cast_fp16")]; + tensor input_linear_weight_to_fp16 = const()[name = tensor("input_linear_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(75619584)))]; + tensor linear_0_bias_0_to_fp16 = const()[name = tensor("linear_0_bias_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(75685184)))]; + tensor linear_0_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = input_linear_weight_to_fp16, x = input_1_cast_fp16)[name = tensor("linear_0_cast_fp16")]; + tensor x_1_axes_0 = const()[name = tensor("x_1_axes_0"), val = tensor([-1])]; + tensor norm0_1_weight_to_fp16 = const()[name = tensor("norm0_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(75687296)))]; + tensor norm0_1_bias_to_fp16 = const()[name = tensor("norm0_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(75689408)))]; + tensor var_60_to_fp16 = const()[name = tensor("op_60_to_fp16"), val = tensor(0x1.5p-17)]; + tensor x_1_cast_fp16 = layer_norm(axes = x_1_axes_0, beta = norm0_1_bias_to_fp16, epsilon = var_60_to_fp16, gamma = norm0_1_weight_to_fp16, x = linear_0_cast_fp16)[name = tensor("x_1_cast_fp16")]; + tensor linear_1_bias_0_to_fp16 = const()[name = tensor("linear_1_bias_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(75691520)))]; + tensor linear_1_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = attn0_in_proj_weight_cast_fp16, x = x_1_cast_fp16)[name = tensor("linear_1_cast_fp16")]; + tensor var_96 = const()[name = tensor("op_96"), val = tensor([1, 1, 3, 16, 64])]; + tensor qkv_1_cast_fp16 = reshape(shape = var_96, x = linear_1_cast_fp16)[name = tensor("qkv_1_cast_fp16")]; + tensor q_1_begin_0 = const()[name = tensor("q_1_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor q_1_end_0 = const()[name = tensor("q_1_end_0"), val = tensor([1, 1, 1, 16, 64])]; + tensor q_1_end_mask_0 = const()[name = tensor("q_1_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor q_1_squeeze_mask_0 = const()[name = tensor("q_1_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor q_1_cast_fp16 = slice_by_index(begin = q_1_begin_0, end = q_1_end_0, end_mask = q_1_end_mask_0, squeeze_mask = q_1_squeeze_mask_0, x = qkv_1_cast_fp16)[name = tensor("q_1_cast_fp16")]; + tensor k_1_begin_0 = const()[name = tensor("k_1_begin_0"), val = tensor([0, 0, 1, 0, 0])]; + tensor k_1_end_0 = const()[name = tensor("k_1_end_0"), val = tensor([1, 1, 2, 16, 64])]; + tensor k_1_end_mask_0 = const()[name = tensor("k_1_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor k_1_squeeze_mask_0 = const()[name = tensor("k_1_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor k_1_cast_fp16 = slice_by_index(begin = k_1_begin_0, end = k_1_end_0, end_mask = k_1_end_mask_0, squeeze_mask = k_1_squeeze_mask_0, x = qkv_1_cast_fp16)[name = tensor("k_1_cast_fp16")]; + tensor v_1_begin_0 = const()[name = tensor("v_1_begin_0"), val = tensor([0, 0, 2, 0, 0])]; + tensor v_1_end_0 = const()[name = tensor("v_1_end_0"), val = tensor([1, 1, 3, 16, 64])]; + tensor v_1_end_mask_0 = const()[name = tensor("v_1_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor v_1_squeeze_mask_0 = const()[name = tensor("v_1_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor v_1_cast_fp16 = slice_by_index(begin = v_1_begin_0, end = v_1_end_0, end_mask = v_1_end_mask_0, squeeze_mask = v_1_squeeze_mask_0, x = qkv_1_cast_fp16)[name = tensor("v_1_cast_fp16")]; + tensor var_200 = const()[name = tensor("op_200"), val = tensor([1, 1, 1, 1])]; + tensor position0_to_fp16_dtype_0 = const()[name = tensor("position0_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor position0_to_fp16 = cast(dtype = position0_to_fp16_dtype_0, x = position0)[name = tensor("cast_139")]; + tensor ts_5_cast_fp16 = reshape(shape = var_200, x = position0_to_fp16)[name = tensor("ts_5_cast_fp16")]; + tensor var_204 = const()[name = tensor("op_204"), val = tensor([1, 1, 16, 32, 2])]; + tensor q_complex_1_cast_fp16 = reshape(shape = var_204, x = q_1_cast_fp16)[name = tensor("q_complex_1_cast_fp16")]; + tensor var_208 = const()[name = tensor("op_208"), val = tensor([1, 1, 16, 32, 2])]; + tensor k_complex_1_cast_fp16 = reshape(shape = var_208, x = k_1_cast_fp16)[name = tensor("k_complex_1_cast_fp16")]; + tensor var_212_begin_0 = const()[name = tensor("op_212_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_212_end_0 = const()[name = tensor("op_212_end_0"), val = tensor([1, 1, 16, 32, 1])]; + tensor var_212_end_mask_0 = const()[name = tensor("op_212_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_212_squeeze_mask_0 = const()[name = tensor("op_212_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_212_cast_fp16 = slice_by_index(begin = var_212_begin_0, end = var_212_end_0, end_mask = var_212_end_mask_0, squeeze_mask = var_212_squeeze_mask_0, x = q_complex_1_cast_fp16)[name = tensor("op_212_cast_fp16")]; + tensor var_220_begin_0 = const()[name = tensor("op_220_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_220_end_0 = const()[name = tensor("op_220_end_0"), val = tensor([1, 1, 16, 32, 2])]; + tensor var_220_end_mask_0 = const()[name = tensor("op_220_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_220_squeeze_mask_0 = const()[name = tensor("op_220_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_220_cast_fp16 = slice_by_index(begin = var_220_begin_0, end = var_220_end_0, end_mask = var_220_end_mask_0, squeeze_mask = var_220_squeeze_mask_0, x = q_complex_1_cast_fp16)[name = tensor("op_220_cast_fp16")]; + tensor var_228_begin_0 = const()[name = tensor("op_228_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_228_end_0 = const()[name = tensor("op_228_end_0"), val = tensor([1, 1, 16, 32, 1])]; + tensor var_228_end_mask_0 = const()[name = tensor("op_228_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_228_squeeze_mask_0 = const()[name = tensor("op_228_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_228_cast_fp16 = slice_by_index(begin = var_228_begin_0, end = var_228_end_0, end_mask = var_228_end_mask_0, squeeze_mask = var_228_squeeze_mask_0, x = k_complex_1_cast_fp16)[name = tensor("op_228_cast_fp16")]; + tensor var_236_begin_0 = const()[name = tensor("op_236_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_236_end_0 = const()[name = tensor("op_236_end_0"), val = tensor([1, 1, 16, 32, 2])]; + tensor var_236_end_mask_0 = const()[name = tensor("op_236_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_236_squeeze_mask_0 = const()[name = tensor("op_236_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_236_cast_fp16 = slice_by_index(begin = var_236_begin_0, end = var_236_end_0, end_mask = var_236_end_mask_0, squeeze_mask = var_236_squeeze_mask_0, x = k_complex_1_cast_fp16)[name = tensor("op_236_cast_fp16")]; + tensor freqs_1_to_fp16 = const()[name = tensor("freqs_1_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(75697728)))]; + tensor var_242_cast_fp16 = mul(x = freqs_1_to_fp16, y = ts_5_cast_fp16)[name = tensor("op_242_cast_fp16")]; + tensor rotr_1_cast_fp16 = cos(x = var_242_cast_fp16)[name = tensor("rotr_1_cast_fp16")]; + tensor roti_1_cast_fp16 = sin(x = var_242_cast_fp16)[name = tensor("roti_1_cast_fp16")]; + tensor var_246_cast_fp16 = mul(x = var_212_cast_fp16, y = rotr_1_cast_fp16)[name = tensor("op_246_cast_fp16")]; + tensor var_247_cast_fp16 = mul(x = var_220_cast_fp16, y = roti_1_cast_fp16)[name = tensor("op_247_cast_fp16")]; + tensor qor_1_cast_fp16 = sub(x = var_246_cast_fp16, y = var_247_cast_fp16)[name = tensor("qor_1_cast_fp16")]; + tensor var_250_cast_fp16 = mul(x = var_212_cast_fp16, y = roti_1_cast_fp16)[name = tensor("op_250_cast_fp16")]; + tensor var_251_cast_fp16 = mul(x = var_220_cast_fp16, y = rotr_1_cast_fp16)[name = tensor("op_251_cast_fp16")]; + tensor qoi_1_cast_fp16 = add(x = var_250_cast_fp16, y = var_251_cast_fp16)[name = tensor("qoi_1_cast_fp16")]; + tensor var_254_cast_fp16 = mul(x = var_228_cast_fp16, y = rotr_1_cast_fp16)[name = tensor("op_254_cast_fp16")]; + tensor var_255_cast_fp16 = mul(x = var_236_cast_fp16, y = roti_1_cast_fp16)[name = tensor("op_255_cast_fp16")]; + tensor kor_1_cast_fp16 = sub(x = var_254_cast_fp16, y = var_255_cast_fp16)[name = tensor("kor_1_cast_fp16")]; + tensor var_258_cast_fp16 = mul(x = var_228_cast_fp16, y = roti_1_cast_fp16)[name = tensor("op_258_cast_fp16")]; + tensor var_259_cast_fp16 = mul(x = var_236_cast_fp16, y = rotr_1_cast_fp16)[name = tensor("op_259_cast_fp16")]; + tensor koi_1_cast_fp16 = add(x = var_258_cast_fp16, y = var_259_cast_fp16)[name = tensor("koi_1_cast_fp16")]; + tensor qo_1_axis_0 = const()[name = tensor("qo_1_axis_0"), val = tensor(-1)]; + tensor qo_1_cast_fp16 = stack(axis = qo_1_axis_0, values = (qor_1_cast_fp16, qoi_1_cast_fp16))[name = tensor("qo_1_cast_fp16")]; + tensor ko_1_axis_0 = const()[name = tensor("ko_1_axis_0"), val = tensor(-1)]; + tensor ko_1_cast_fp16 = stack(axis = ko_1_axis_0, values = (kor_1_cast_fp16, koi_1_cast_fp16))[name = tensor("ko_1_cast_fp16")]; + tensor var_288 = const()[name = tensor("op_288"), val = tensor([1, 1, 16, 64])]; + tensor q_3_cast_fp16 = reshape(shape = var_288, x = qo_1_cast_fp16)[name = tensor("q_3_cast_fp16")]; + tensor var_290 = const()[name = tensor("op_290"), val = tensor([1, 1, 16, 64])]; + tensor k_3_cast_fp16 = reshape(shape = var_290, x = ko_1_cast_fp16)[name = tensor("k_3_cast_fp16")]; + tensor _inversed_312_y_0_to_fp16 = const()[name = tensor("_inversed_312_y_0_to_fp16"), val = tensor(0x1p-9)]; + tensor _inversed_312_cast_fp16 = mul(x = ts_5_cast_fp16, y = _inversed_312_y_0_to_fp16)[name = tensor("_inversed_312_cast_fp16")]; + tensor var_313_cast_fp16 = floor(x = _inversed_312_cast_fp16)[name = tensor("op_313_cast_fp16")]; + tensor var_314_to_fp16 = const()[name = tensor("op_314_to_fp16"), val = tensor(0x1p+9)]; + tensor var_315_cast_fp16 = mul(x = var_313_cast_fp16, y = var_314_to_fp16)[name = tensor("op_315_cast_fp16")]; + tensor write_indices_float_3_cast_fp16 = sub(x = ts_5_cast_fp16, y = var_315_cast_fp16)[name = tensor("write_indices_float_3_cast_fp16")]; + tensor var_322_dtype_0 = const()[name = tensor("op_322_dtype_0"), val = tensor("int32")]; + tensor write_indices_1_reps_0 = const()[name = tensor("write_indices_1_reps_0"), val = tensor([1, 1, 16, 64])]; + tensor write_indices_float_3_cast_fp16_to_int32 = cast(dtype = var_322_dtype_0, x = write_indices_float_3_cast_fp16)[name = tensor("cast_138")]; + tensor write_indices_1 = tile(reps = write_indices_1_reps_0, x = write_indices_float_3_cast_fp16_to_int32)[name = tensor("write_indices_1")]; + tensor var_330_begin_0 = const()[name = tensor("op_330_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_330_end_0 = const()[name = tensor("op_330_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor var_330_end_mask_0 = const()[name = tensor("op_330_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_330_squeeze_mask_0 = const()[name = tensor("op_330_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor cache0_to_fp16_dtype_0 = const()[name = tensor("cache0_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor cache0_to_fp16 = cast(dtype = cache0_to_fp16_dtype_0, x = cache0)[name = tensor("cast_137")]; + tensor var_330_cast_fp16 = slice_by_index(begin = var_330_begin_0, end = var_330_end_0, end_mask = var_330_end_mask_0, squeeze_mask = var_330_squeeze_mask_0, x = cache0_to_fp16)[name = tensor("op_330_cast_fp16")]; + tensor var_332_axis_0 = const()[name = tensor("op_332_axis_0"), val = tensor(1)]; + tensor var_332_mode_0 = const()[name = tensor("op_332_mode_0"), val = tensor("update")]; + tensor var_332_validate_indices_0 = const()[name = tensor("op_332_validate_indices_0"), val = tensor(false)]; + tensor var_332_cast_fp16 = scatter_along_axis(axis = var_332_axis_0, data = var_330_cast_fp16, indices = write_indices_1, mode = var_332_mode_0, updates = k_3_cast_fp16, validate_indices = var_332_validate_indices_0)[name = tensor("op_332_cast_fp16")]; + tensor concat_2 = const()[name = tensor("concat_2"), val = tensor([0, 0, 0, 0, 0])]; + tensor concat_3 = const()[name = tensor("concat_3"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_1_internal_tensor_assign_1_stride_0 = const()[name = tensor("new_cache_1_internal_tensor_assign_1_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_1_internal_tensor_assign_1_begin_mask_0 = const()[name = tensor("new_cache_1_internal_tensor_assign_1_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_1_internal_tensor_assign_1_end_mask_0 = const()[name = tensor("new_cache_1_internal_tensor_assign_1_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_1_internal_tensor_assign_1_squeeze_mask_0 = const()[name = tensor("new_cache_1_internal_tensor_assign_1_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_12 = const()[name = tensor("shape_12"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_0 = const()[name = tensor("reduce_prod_0"), val = tensor(1048576)]; + tensor range_1d_0_start_0 = const()[name = tensor("range_1d_0_start_0"), val = tensor(0)]; + tensor range_1d_0_step_0 = const()[name = tensor("range_1d_0_step_0"), val = tensor(1)]; + tensor range_1d_0 = range_1d(end = reduce_prod_0, start = range_1d_0_start_0, step = range_1d_0_step_0)[name = tensor("range_1d_0")]; + tensor reshape_0 = reshape(shape = shape_12, x = range_1d_0)[name = tensor("reshape_0")]; + tensor slice_by_index_0 = slice_by_index(begin = concat_2, begin_mask = new_cache_1_internal_tensor_assign_1_begin_mask_0, end = concat_3, end_mask = new_cache_1_internal_tensor_assign_1_end_mask_0, squeeze_mask = new_cache_1_internal_tensor_assign_1_squeeze_mask_0, stride = new_cache_1_internal_tensor_assign_1_stride_0, x = reshape_0)[name = tensor("slice_by_index_0")]; + tensor reshape_1_shape_0 = const()[name = tensor("reshape_1_shape_0"), val = tensor([-1])]; + tensor reshape_1 = reshape(shape = reshape_1_shape_0, x = slice_by_index_0)[name = tensor("reshape_1")]; + tensor reshape_2_shape_0 = const()[name = tensor("reshape_2_shape_0"), val = tensor([-1])]; + tensor reshape_2_cast_fp16 = reshape(shape = reshape_2_shape_0, x = var_332_cast_fp16)[name = tensor("reshape_2_cast_fp16")]; + tensor reshape_3_shape_0 = const()[name = tensor("reshape_3_shape_0"), val = tensor([-1])]; + tensor reshape_3_cast_fp16 = reshape(shape = reshape_3_shape_0, x = cache0_to_fp16)[name = tensor("reshape_3_cast_fp16")]; + tensor scatter_0_mode_0 = const()[name = tensor("scatter_0_mode_0"), val = tensor("update")]; + tensor scatter_0_axis_0 = const()[name = tensor("scatter_0_axis_0"), val = tensor(0)]; + tensor scatter_0_validate_indices_0 = const()[name = tensor("scatter_0_validate_indices_0"), val = tensor(false)]; + tensor scatter_0_cast_fp16 = scatter(axis = scatter_0_axis_0, data = reshape_3_cast_fp16, indices = reshape_1, mode = scatter_0_mode_0, updates = reshape_2_cast_fp16, validate_indices = scatter_0_validate_indices_0)[name = tensor("scatter_0_cast_fp16")]; + tensor reshape_4_cast_fp16 = reshape(shape = shape_12, x = scatter_0_cast_fp16)[name = tensor("reshape_4_cast_fp16")]; + tensor var_340_begin_0 = const()[name = tensor("op_340_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor var_340_end_0 = const()[name = tensor("op_340_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor var_340_end_mask_0 = const()[name = tensor("op_340_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_340_squeeze_mask_0 = const()[name = tensor("op_340_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor var_340_cast_fp16 = slice_by_index(begin = var_340_begin_0, end = var_340_end_0, end_mask = var_340_end_mask_0, squeeze_mask = var_340_squeeze_mask_0, x = reshape_4_cast_fp16)[name = tensor("op_340_cast_fp16")]; + tensor var_342_axis_0 = const()[name = tensor("op_342_axis_0"), val = tensor(1)]; + tensor var_342_mode_0 = const()[name = tensor("op_342_mode_0"), val = tensor("update")]; + tensor var_342_validate_indices_0 = const()[name = tensor("op_342_validate_indices_0"), val = tensor(false)]; + tensor var_342_cast_fp16 = scatter_along_axis(axis = var_342_axis_0, data = var_340_cast_fp16, indices = write_indices_1, mode = var_342_mode_0, updates = v_1_cast_fp16, validate_indices = var_342_validate_indices_0)[name = tensor("op_342_cast_fp16")]; + tensor concat_4 = const()[name = tensor("concat_4"), val = tensor([1, 0, 0, 0, 0])]; + tensor concat_5 = const()[name = tensor("concat_5"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_1_internal_tensor_assign_2_stride_0 = const()[name = tensor("new_cache_1_internal_tensor_assign_2_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_1_internal_tensor_assign_2_begin_mask_0 = const()[name = tensor("new_cache_1_internal_tensor_assign_2_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_1_internal_tensor_assign_2_end_mask_0 = const()[name = tensor("new_cache_1_internal_tensor_assign_2_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_1_internal_tensor_assign_2_squeeze_mask_0 = const()[name = tensor("new_cache_1_internal_tensor_assign_2_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_13 = const()[name = tensor("shape_13"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_1 = const()[name = tensor("reduce_prod_1"), val = tensor(1048576)]; + tensor range_1d_1_start_0 = const()[name = tensor("range_1d_1_start_0"), val = tensor(0)]; + tensor range_1d_1_step_0 = const()[name = tensor("range_1d_1_step_0"), val = tensor(1)]; + tensor range_1d_1 = range_1d(end = reduce_prod_1, start = range_1d_1_start_0, step = range_1d_1_step_0)[name = tensor("range_1d_1")]; + tensor reshape_5 = reshape(shape = shape_13, x = range_1d_1)[name = tensor("reshape_5")]; + tensor slice_by_index_1 = slice_by_index(begin = concat_4, begin_mask = new_cache_1_internal_tensor_assign_2_begin_mask_0, end = concat_5, end_mask = new_cache_1_internal_tensor_assign_2_end_mask_0, squeeze_mask = new_cache_1_internal_tensor_assign_2_squeeze_mask_0, stride = new_cache_1_internal_tensor_assign_2_stride_0, x = reshape_5)[name = tensor("slice_by_index_1")]; + tensor reshape_6_shape_0 = const()[name = tensor("reshape_6_shape_0"), val = tensor([-1])]; + tensor reshape_6 = reshape(shape = reshape_6_shape_0, x = slice_by_index_1)[name = tensor("reshape_6")]; + tensor reshape_7_shape_0 = const()[name = tensor("reshape_7_shape_0"), val = tensor([-1])]; + tensor reshape_7_cast_fp16 = reshape(shape = reshape_7_shape_0, x = var_342_cast_fp16)[name = tensor("reshape_7_cast_fp16")]; + tensor reshape_8_shape_0 = const()[name = tensor("reshape_8_shape_0"), val = tensor([-1])]; + tensor reshape_8_cast_fp16 = reshape(shape = reshape_8_shape_0, x = reshape_4_cast_fp16)[name = tensor("reshape_8_cast_fp16")]; + tensor scatter_1_mode_0 = const()[name = tensor("scatter_1_mode_0"), val = tensor("update")]; + tensor scatter_1_axis_0 = const()[name = tensor("scatter_1_axis_0"), val = tensor(0)]; + tensor scatter_1_validate_indices_0 = const()[name = tensor("scatter_1_validate_indices_0"), val = tensor(false)]; + tensor scatter_1_cast_fp16 = scatter(axis = scatter_1_axis_0, data = reshape_8_cast_fp16, indices = reshape_6, mode = scatter_1_mode_0, updates = reshape_7_cast_fp16, validate_indices = scatter_1_validate_indices_0)[name = tensor("scatter_1_cast_fp16")]; + tensor reshape_9_cast_fp16 = reshape(shape = shape_13, x = scatter_1_cast_fp16)[name = tensor("reshape_9_cast_fp16")]; + tensor reshape_9_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("reshape_9_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor keys_1_begin_0 = const()[name = tensor("keys_1_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor keys_1_end_0 = const()[name = tensor("keys_1_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor keys_1_end_mask_0 = const()[name = tensor("keys_1_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor keys_1_squeeze_mask_0 = const()[name = tensor("keys_1_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor keys_1_cast_fp16 = slice_by_index(begin = keys_1_begin_0, end = keys_1_end_0, end_mask = keys_1_end_mask_0, squeeze_mask = keys_1_squeeze_mask_0, x = reshape_9_cast_fp16)[name = tensor("keys_1_cast_fp16")]; + tensor values_1_begin_0 = const()[name = tensor("values_1_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor values_1_end_0 = const()[name = tensor("values_1_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor values_1_end_mask_0 = const()[name = tensor("values_1_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor values_1_squeeze_mask_0 = const()[name = tensor("values_1_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor values_1_cast_fp16 = slice_by_index(begin = values_1_begin_0, end = values_1_end_0, end_mask = values_1_end_mask_0, squeeze_mask = values_1_squeeze_mask_0, x = reshape_9_cast_fp16)[name = tensor("values_1_cast_fp16")]; + tensor var_354_cast_fp16 = not_equal(x = keys_1_cast_fp16, y = keys_1_cast_fp16)[name = tensor("op_354_cast_fp16")]; + tensor var_360_to_fp16 = const()[name = tensor("op_360_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(75697856)))]; + tensor keys_3_cast_fp16 = select(a = var_360_to_fp16, b = keys_1_cast_fp16, cond = var_354_cast_fp16)[name = tensor("keys_3_cast_fp16")]; + tensor var_362_cast_fp16 = not_equal(x = values_1_cast_fp16, y = values_1_cast_fp16)[name = tensor("op_362_cast_fp16")]; + tensor values_3_cast_fp16 = select(a = var_360_to_fp16, b = values_1_cast_fp16, cond = var_362_cast_fp16)[name = tensor("values_3_cast_fp16")]; + tensor var_386 = const()[name = tensor("op_386"), val = tensor([0, 2, 1, 3])]; + tensor var_399 = const()[name = tensor("op_399"), val = tensor([1, 1, 1])]; + tensor var_400_cast_fp16 = reshape(shape = var_399, x = position0_to_fp16)[name = tensor("op_400_cast_fp16")]; + tensor var_417_to_fp16 = const()[name = tensor("op_417_to_fp16"), val = tensor(0x1p+0)]; + tensor valid_len_1_cast_fp16 = add(x = var_400_cast_fp16, y = var_417_to_fp16)[name = tensor("valid_len_1_cast_fp16")]; + tensor k_positions_1_promoted_to_fp16 = const()[name = tensor("k_positions_1_promoted_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(76746496)))]; + tensor valid_mask_1_cast_fp16 = less(x = k_positions_1_promoted_to_fp16, y = valid_len_1_cast_fp16)[name = tensor("valid_mask_1_cast_fp16")]; + tensor causal_mask_1_cast_fp16 = less_equal(x = k_positions_1_promoted_to_fp16, y = var_400_cast_fp16)[name = tensor("causal_mask_1_cast_fp16")]; + tensor attn_mask_1 = logical_and(x = valid_mask_1_cast_fp16, y = causal_mask_1_cast_fp16)[name = tensor("attn_mask_1")]; + tensor attn_mask_3_axes_0 = const()[name = tensor("attn_mask_3_axes_0"), val = tensor([1])]; + tensor attn_mask_3 = expand_dims(axes = attn_mask_3_axes_0, x = attn_mask_1)[name = tensor("attn_mask_3")]; + tensor var_435_transpose_x_0 = const()[name = tensor("op_435_transpose_x_0"), val = tensor(false)]; + tensor var_435_transpose_y_0 = const()[name = tensor("op_435_transpose_y_0"), val = tensor(false)]; + tensor transpose_18_perm_0 = const()[name = tensor("transpose_18_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_19_perm_0 = const()[name = tensor("transpose_19_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_19 = transpose(perm = transpose_19_perm_0, x = keys_3_cast_fp16)[name = tensor("transpose_51")]; + tensor transpose_18 = transpose(perm = transpose_18_perm_0, x = q_3_cast_fp16)[name = tensor("transpose_52")]; + tensor var_435_cast_fp16 = matmul(transpose_x = var_435_transpose_x_0, transpose_y = var_435_transpose_y_0, x = transpose_18, y = transpose_19)[name = tensor("op_435_cast_fp16")]; + tensor var_429_to_fp16 = const()[name = tensor("op_429_to_fp16"), val = tensor([0x1p-3])]; + tensor attn_weights_1_cast_fp16 = mul(x = var_435_cast_fp16, y = var_429_to_fp16)[name = tensor("attn_weights_1_cast_fp16")]; + tensor var_437 = logical_not(x = attn_mask_3)[name = tensor("op_437")]; + tensor var_438_to_fp16 = const()[name = tensor("op_438_to_fp16"), val = tensor(-inf)]; + tensor attn_weights_3_cast_fp16 = select(a = var_438_to_fp16, b = attn_weights_1_cast_fp16, cond = var_437)[name = tensor("attn_weights_3_cast_fp16")]; + tensor var_440 = const()[name = tensor("op_440"), val = tensor(-1)]; + tensor attn_weights_5_cast_fp16 = softmax(axis = var_440, x = attn_weights_3_cast_fp16)[name = tensor("attn_weights_5_cast_fp16")]; + tensor attn_output_1_transpose_x_0 = const()[name = tensor("attn_output_1_transpose_x_0"), val = tensor(false)]; + tensor attn_output_1_transpose_y_0 = const()[name = tensor("attn_output_1_transpose_y_0"), val = tensor(false)]; + tensor values_5_cast_fp16 = transpose(perm = var_386, x = values_3_cast_fp16)[name = tensor("transpose_53")]; + tensor attn_output_1_cast_fp16 = matmul(transpose_x = attn_output_1_transpose_x_0, transpose_y = attn_output_1_transpose_y_0, x = attn_weights_5_cast_fp16, y = values_5_cast_fp16)[name = tensor("attn_output_1_cast_fp16")]; + tensor var_448 = const()[name = tensor("op_448"), val = tensor([0, 2, 1, 3])]; + tensor var_451 = const()[name = tensor("op_451"), val = tensor([1, 1, 1024])]; + tensor var_449_cast_fp16 = transpose(perm = var_448, x = attn_output_1_cast_fp16)[name = tensor("transpose_50")]; + tensor input_5_cast_fp16 = reshape(shape = var_451, x = var_449_cast_fp16)[name = tensor("input_5_cast_fp16")]; + tensor linear_2_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = attn0_out_proj_weight_cast_fp16, x = input_5_cast_fp16)[name = tensor("linear_2_cast_fp16")]; + tensor var_457_to_fp16 = const()[name = tensor("op_457_to_fp16"), val = tensor(0x1p+0)]; + tensor var_458_cast_fp16 = add(x = position0_to_fp16, y = var_457_to_fp16)[name = tensor("op_458_cast_fp16")]; + tensor var_458_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_458_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor input_7_cast_fp16 = add(x = linear_0_cast_fp16, y = linear_2_cast_fp16)[name = tensor("input_7_cast_fp16")]; + tensor input_9_axes_0 = const()[name = tensor("input_9_axes_0"), val = tensor([-1])]; + tensor norm0_2_weight_to_fp16 = const()[name = tensor("norm0_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(76747584)))]; + tensor norm0_2_bias_to_fp16 = const()[name = tensor("norm0_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(76749696)))]; + tensor var_462_to_fp16 = const()[name = tensor("op_462_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_9_cast_fp16 = layer_norm(axes = input_9_axes_0, beta = norm0_2_bias_to_fp16, epsilon = var_462_to_fp16, gamma = norm0_2_weight_to_fp16, x = input_7_cast_fp16)[name = tensor("input_9_cast_fp16")]; + tensor linear_3_bias_0_to_fp16 = const()[name = tensor("linear_3_bias_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(76751808)))]; + tensor linear_3_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = linear0_1_weight_cast_fp16, x = input_9_cast_fp16)[name = tensor("linear_3_cast_fp16")]; + tensor input_11_mode_0 = const()[name = tensor("input_11_mode_0"), val = tensor("EXACT")]; + tensor input_11_cast_fp16 = gelu(mode = input_11_mode_0, x = linear_3_cast_fp16)[name = tensor("input_11_cast_fp16")]; + tensor linear_4_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = linear0_2_weight_cast_fp16, x = input_11_cast_fp16)[name = tensor("linear_4_cast_fp16")]; + tensor input_13_cast_fp16 = add(x = input_7_cast_fp16, y = linear_4_cast_fp16)[name = tensor("input_13_cast_fp16")]; + tensor x_3_axes_0 = const()[name = tensor("x_3_axes_0"), val = tensor([-1])]; + tensor norm1_1_weight_to_fp16 = const()[name = tensor("norm1_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(76760064)))]; + tensor norm1_1_bias_to_fp16 = const()[name = tensor("norm1_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(76762176)))]; + tensor var_479_to_fp16 = const()[name = tensor("op_479_to_fp16"), val = tensor(0x1.5p-17)]; + tensor x_3_cast_fp16 = layer_norm(axes = x_3_axes_0, beta = norm1_1_bias_to_fp16, epsilon = var_479_to_fp16, gamma = norm1_1_weight_to_fp16, x = input_13_cast_fp16)[name = tensor("x_3_cast_fp16")]; + tensor linear_5_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = attn1_in_proj_weight_cast_fp16, x = x_3_cast_fp16)[name = tensor("linear_5_cast_fp16")]; + tensor var_515 = const()[name = tensor("op_515"), val = tensor([1, 1, 3, 16, 64])]; + tensor qkv_3_cast_fp16 = reshape(shape = var_515, x = linear_5_cast_fp16)[name = tensor("qkv_3_cast_fp16")]; + tensor q_7_begin_0 = const()[name = tensor("q_7_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor q_7_end_0 = const()[name = tensor("q_7_end_0"), val = tensor([1, 1, 1, 16, 64])]; + tensor q_7_end_mask_0 = const()[name = tensor("q_7_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor q_7_squeeze_mask_0 = const()[name = tensor("q_7_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor q_7_cast_fp16 = slice_by_index(begin = q_7_begin_0, end = q_7_end_0, end_mask = q_7_end_mask_0, squeeze_mask = q_7_squeeze_mask_0, x = qkv_3_cast_fp16)[name = tensor("q_7_cast_fp16")]; + tensor k_5_begin_0 = const()[name = tensor("k_5_begin_0"), val = tensor([0, 0, 1, 0, 0])]; + tensor k_5_end_0 = const()[name = tensor("k_5_end_0"), val = tensor([1, 1, 2, 16, 64])]; + tensor k_5_end_mask_0 = const()[name = tensor("k_5_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor k_5_squeeze_mask_0 = const()[name = tensor("k_5_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor k_5_cast_fp16 = slice_by_index(begin = k_5_begin_0, end = k_5_end_0, end_mask = k_5_end_mask_0, squeeze_mask = k_5_squeeze_mask_0, x = qkv_3_cast_fp16)[name = tensor("k_5_cast_fp16")]; + tensor v_3_begin_0 = const()[name = tensor("v_3_begin_0"), val = tensor([0, 0, 2, 0, 0])]; + tensor v_3_end_0 = const()[name = tensor("v_3_end_0"), val = tensor([1, 1, 3, 16, 64])]; + tensor v_3_end_mask_0 = const()[name = tensor("v_3_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor v_3_squeeze_mask_0 = const()[name = tensor("v_3_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor v_3_cast_fp16 = slice_by_index(begin = v_3_begin_0, end = v_3_end_0, end_mask = v_3_end_mask_0, squeeze_mask = v_3_squeeze_mask_0, x = qkv_3_cast_fp16)[name = tensor("v_3_cast_fp16")]; + tensor var_619 = const()[name = tensor("op_619"), val = tensor([1, 1, 1, 1])]; + tensor position1_to_fp16_dtype_0 = const()[name = tensor("position1_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor position1_to_fp16 = cast(dtype = position1_to_fp16_dtype_0, x = position1)[name = tensor("cast_134")]; + tensor ts_11_cast_fp16 = reshape(shape = var_619, x = position1_to_fp16)[name = tensor("ts_11_cast_fp16")]; + tensor var_623 = const()[name = tensor("op_623"), val = tensor([1, 1, 16, 32, 2])]; + tensor q_complex_3_cast_fp16 = reshape(shape = var_623, x = q_7_cast_fp16)[name = tensor("q_complex_3_cast_fp16")]; + tensor var_627 = const()[name = tensor("op_627"), val = tensor([1, 1, 16, 32, 2])]; + tensor k_complex_3_cast_fp16 = reshape(shape = var_627, x = k_5_cast_fp16)[name = tensor("k_complex_3_cast_fp16")]; + tensor var_631_begin_0 = const()[name = tensor("op_631_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_631_end_0 = const()[name = tensor("op_631_end_0"), val = tensor([1, 1, 16, 32, 1])]; + tensor var_631_end_mask_0 = const()[name = tensor("op_631_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_631_squeeze_mask_0 = const()[name = tensor("op_631_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_631_cast_fp16 = slice_by_index(begin = var_631_begin_0, end = var_631_end_0, end_mask = var_631_end_mask_0, squeeze_mask = var_631_squeeze_mask_0, x = q_complex_3_cast_fp16)[name = tensor("op_631_cast_fp16")]; + tensor var_639_begin_0 = const()[name = tensor("op_639_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_639_end_0 = const()[name = tensor("op_639_end_0"), val = tensor([1, 1, 16, 32, 2])]; + tensor var_639_end_mask_0 = const()[name = tensor("op_639_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_639_squeeze_mask_0 = const()[name = tensor("op_639_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_639_cast_fp16 = slice_by_index(begin = var_639_begin_0, end = var_639_end_0, end_mask = var_639_end_mask_0, squeeze_mask = var_639_squeeze_mask_0, x = q_complex_3_cast_fp16)[name = tensor("op_639_cast_fp16")]; + tensor var_647_begin_0 = const()[name = tensor("op_647_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_647_end_0 = const()[name = tensor("op_647_end_0"), val = tensor([1, 1, 16, 32, 1])]; + tensor var_647_end_mask_0 = const()[name = tensor("op_647_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_647_squeeze_mask_0 = const()[name = tensor("op_647_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_647_cast_fp16 = slice_by_index(begin = var_647_begin_0, end = var_647_end_0, end_mask = var_647_end_mask_0, squeeze_mask = var_647_squeeze_mask_0, x = k_complex_3_cast_fp16)[name = tensor("op_647_cast_fp16")]; + tensor var_655_begin_0 = const()[name = tensor("op_655_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_655_end_0 = const()[name = tensor("op_655_end_0"), val = tensor([1, 1, 16, 32, 2])]; + tensor var_655_end_mask_0 = const()[name = tensor("op_655_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_655_squeeze_mask_0 = const()[name = tensor("op_655_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_655_cast_fp16 = slice_by_index(begin = var_655_begin_0, end = var_655_end_0, end_mask = var_655_end_mask_0, squeeze_mask = var_655_squeeze_mask_0, x = k_complex_3_cast_fp16)[name = tensor("op_655_cast_fp16")]; + tensor freqs_3_to_fp16 = const()[name = tensor("freqs_3_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(76764288)))]; + tensor var_661_cast_fp16 = mul(x = freqs_3_to_fp16, y = ts_11_cast_fp16)[name = tensor("op_661_cast_fp16")]; + tensor rotr_3_cast_fp16 = cos(x = var_661_cast_fp16)[name = tensor("rotr_3_cast_fp16")]; + tensor roti_3_cast_fp16 = sin(x = var_661_cast_fp16)[name = tensor("roti_3_cast_fp16")]; + tensor var_665_cast_fp16 = mul(x = var_631_cast_fp16, y = rotr_3_cast_fp16)[name = tensor("op_665_cast_fp16")]; + tensor var_666_cast_fp16 = mul(x = var_639_cast_fp16, y = roti_3_cast_fp16)[name = tensor("op_666_cast_fp16")]; + tensor qor_5_cast_fp16 = sub(x = var_665_cast_fp16, y = var_666_cast_fp16)[name = tensor("qor_5_cast_fp16")]; + tensor var_669_cast_fp16 = mul(x = var_631_cast_fp16, y = roti_3_cast_fp16)[name = tensor("op_669_cast_fp16")]; + tensor var_670_cast_fp16 = mul(x = var_639_cast_fp16, y = rotr_3_cast_fp16)[name = tensor("op_670_cast_fp16")]; + tensor qoi_5_cast_fp16 = add(x = var_669_cast_fp16, y = var_670_cast_fp16)[name = tensor("qoi_5_cast_fp16")]; + tensor var_673_cast_fp16 = mul(x = var_647_cast_fp16, y = rotr_3_cast_fp16)[name = tensor("op_673_cast_fp16")]; + tensor var_674_cast_fp16 = mul(x = var_655_cast_fp16, y = roti_3_cast_fp16)[name = tensor("op_674_cast_fp16")]; + tensor kor_5_cast_fp16 = sub(x = var_673_cast_fp16, y = var_674_cast_fp16)[name = tensor("kor_5_cast_fp16")]; + tensor var_677_cast_fp16 = mul(x = var_647_cast_fp16, y = roti_3_cast_fp16)[name = tensor("op_677_cast_fp16")]; + tensor var_678_cast_fp16 = mul(x = var_655_cast_fp16, y = rotr_3_cast_fp16)[name = tensor("op_678_cast_fp16")]; + tensor koi_5_cast_fp16 = add(x = var_677_cast_fp16, y = var_678_cast_fp16)[name = tensor("koi_5_cast_fp16")]; + tensor qo_3_axis_0 = const()[name = tensor("qo_3_axis_0"), val = tensor(-1)]; + tensor qo_3_cast_fp16 = stack(axis = qo_3_axis_0, values = (qor_5_cast_fp16, qoi_5_cast_fp16))[name = tensor("qo_3_cast_fp16")]; + tensor ko_3_axis_0 = const()[name = tensor("ko_3_axis_0"), val = tensor(-1)]; + tensor ko_3_cast_fp16 = stack(axis = ko_3_axis_0, values = (kor_5_cast_fp16, koi_5_cast_fp16))[name = tensor("ko_3_cast_fp16")]; + tensor var_707 = const()[name = tensor("op_707"), val = tensor([1, 1, 16, 64])]; + tensor q_9_cast_fp16 = reshape(shape = var_707, x = qo_3_cast_fp16)[name = tensor("q_9_cast_fp16")]; + tensor var_709 = const()[name = tensor("op_709"), val = tensor([1, 1, 16, 64])]; + tensor k_7_cast_fp16 = reshape(shape = var_709, x = ko_3_cast_fp16)[name = tensor("k_7_cast_fp16")]; + tensor _inversed_731_y_0_to_fp16 = const()[name = tensor("_inversed_731_y_0_to_fp16"), val = tensor(0x1p-9)]; + tensor _inversed_731_cast_fp16 = mul(x = ts_11_cast_fp16, y = _inversed_731_y_0_to_fp16)[name = tensor("_inversed_731_cast_fp16")]; + tensor var_732_cast_fp16 = floor(x = _inversed_731_cast_fp16)[name = tensor("op_732_cast_fp16")]; + tensor var_733_to_fp16 = const()[name = tensor("op_733_to_fp16"), val = tensor(0x1p+9)]; + tensor var_734_cast_fp16 = mul(x = var_732_cast_fp16, y = var_733_to_fp16)[name = tensor("op_734_cast_fp16")]; + tensor write_indices_float_7_cast_fp16 = sub(x = ts_11_cast_fp16, y = var_734_cast_fp16)[name = tensor("write_indices_float_7_cast_fp16")]; + tensor var_741_dtype_0 = const()[name = tensor("op_741_dtype_0"), val = tensor("int32")]; + tensor write_indices_3_reps_0 = const()[name = tensor("write_indices_3_reps_0"), val = tensor([1, 1, 16, 64])]; + tensor write_indices_float_7_cast_fp16_to_int32 = cast(dtype = var_741_dtype_0, x = write_indices_float_7_cast_fp16)[name = tensor("cast_133")]; + tensor write_indices_3 = tile(reps = write_indices_3_reps_0, x = write_indices_float_7_cast_fp16_to_int32)[name = tensor("write_indices_3")]; + tensor var_749_begin_0 = const()[name = tensor("op_749_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_749_end_0 = const()[name = tensor("op_749_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor var_749_end_mask_0 = const()[name = tensor("op_749_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_749_squeeze_mask_0 = const()[name = tensor("op_749_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor cache1_to_fp16_dtype_0 = const()[name = tensor("cache1_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor cache1_to_fp16 = cast(dtype = cache1_to_fp16_dtype_0, x = cache1)[name = tensor("cast_132")]; + tensor var_749_cast_fp16 = slice_by_index(begin = var_749_begin_0, end = var_749_end_0, end_mask = var_749_end_mask_0, squeeze_mask = var_749_squeeze_mask_0, x = cache1_to_fp16)[name = tensor("op_749_cast_fp16")]; + tensor var_751_axis_0 = const()[name = tensor("op_751_axis_0"), val = tensor(1)]; + tensor var_751_mode_0 = const()[name = tensor("op_751_mode_0"), val = tensor("update")]; + tensor var_751_validate_indices_0 = const()[name = tensor("op_751_validate_indices_0"), val = tensor(false)]; + tensor var_751_cast_fp16 = scatter_along_axis(axis = var_751_axis_0, data = var_749_cast_fp16, indices = write_indices_3, mode = var_751_mode_0, updates = k_7_cast_fp16, validate_indices = var_751_validate_indices_0)[name = tensor("op_751_cast_fp16")]; + tensor concat_9 = const()[name = tensor("concat_9"), val = tensor([0, 0, 0, 0, 0])]; + tensor concat_10 = const()[name = tensor("concat_10"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_3_internal_tensor_assign_1_stride_0 = const()[name = tensor("new_cache_3_internal_tensor_assign_1_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_3_internal_tensor_assign_1_begin_mask_0 = const()[name = tensor("new_cache_3_internal_tensor_assign_1_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_3_internal_tensor_assign_1_end_mask_0 = const()[name = tensor("new_cache_3_internal_tensor_assign_1_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_3_internal_tensor_assign_1_squeeze_mask_0 = const()[name = tensor("new_cache_3_internal_tensor_assign_1_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_14 = const()[name = tensor("shape_14"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_2 = const()[name = tensor("reduce_prod_2"), val = tensor(1048576)]; + tensor range_1d_2_start_0 = const()[name = tensor("range_1d_2_start_0"), val = tensor(0)]; + tensor range_1d_2_step_0 = const()[name = tensor("range_1d_2_step_0"), val = tensor(1)]; + tensor range_1d_2 = range_1d(end = reduce_prod_2, start = range_1d_2_start_0, step = range_1d_2_step_0)[name = tensor("range_1d_2")]; + tensor reshape_10 = reshape(shape = shape_14, x = range_1d_2)[name = tensor("reshape_10")]; + tensor slice_by_index_2 = slice_by_index(begin = concat_9, begin_mask = new_cache_3_internal_tensor_assign_1_begin_mask_0, end = concat_10, end_mask = new_cache_3_internal_tensor_assign_1_end_mask_0, squeeze_mask = new_cache_3_internal_tensor_assign_1_squeeze_mask_0, stride = new_cache_3_internal_tensor_assign_1_stride_0, x = reshape_10)[name = tensor("slice_by_index_2")]; + tensor reshape_11_shape_0 = const()[name = tensor("reshape_11_shape_0"), val = tensor([-1])]; + tensor reshape_11 = reshape(shape = reshape_11_shape_0, x = slice_by_index_2)[name = tensor("reshape_11")]; + tensor reshape_12_shape_0 = const()[name = tensor("reshape_12_shape_0"), val = tensor([-1])]; + tensor reshape_12_cast_fp16 = reshape(shape = reshape_12_shape_0, x = var_751_cast_fp16)[name = tensor("reshape_12_cast_fp16")]; + tensor reshape_13_shape_0 = const()[name = tensor("reshape_13_shape_0"), val = tensor([-1])]; + tensor reshape_13_cast_fp16 = reshape(shape = reshape_13_shape_0, x = cache1_to_fp16)[name = tensor("reshape_13_cast_fp16")]; + tensor scatter_2_mode_0 = const()[name = tensor("scatter_2_mode_0"), val = tensor("update")]; + tensor scatter_2_axis_0 = const()[name = tensor("scatter_2_axis_0"), val = tensor(0)]; + tensor scatter_2_validate_indices_0 = const()[name = tensor("scatter_2_validate_indices_0"), val = tensor(false)]; + tensor scatter_2_cast_fp16 = scatter(axis = scatter_2_axis_0, data = reshape_13_cast_fp16, indices = reshape_11, mode = scatter_2_mode_0, updates = reshape_12_cast_fp16, validate_indices = scatter_2_validate_indices_0)[name = tensor("scatter_2_cast_fp16")]; + tensor reshape_14_cast_fp16 = reshape(shape = shape_14, x = scatter_2_cast_fp16)[name = tensor("reshape_14_cast_fp16")]; + tensor var_759_begin_0 = const()[name = tensor("op_759_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor var_759_end_0 = const()[name = tensor("op_759_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor var_759_end_mask_0 = const()[name = tensor("op_759_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_759_squeeze_mask_0 = const()[name = tensor("op_759_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor var_759_cast_fp16 = slice_by_index(begin = var_759_begin_0, end = var_759_end_0, end_mask = var_759_end_mask_0, squeeze_mask = var_759_squeeze_mask_0, x = reshape_14_cast_fp16)[name = tensor("op_759_cast_fp16")]; + tensor var_761_axis_0 = const()[name = tensor("op_761_axis_0"), val = tensor(1)]; + tensor var_761_mode_0 = const()[name = tensor("op_761_mode_0"), val = tensor("update")]; + tensor var_761_validate_indices_0 = const()[name = tensor("op_761_validate_indices_0"), val = tensor(false)]; + tensor var_761_cast_fp16 = scatter_along_axis(axis = var_761_axis_0, data = var_759_cast_fp16, indices = write_indices_3, mode = var_761_mode_0, updates = v_3_cast_fp16, validate_indices = var_761_validate_indices_0)[name = tensor("op_761_cast_fp16")]; + tensor concat_11 = const()[name = tensor("concat_11"), val = tensor([1, 0, 0, 0, 0])]; + tensor concat_12 = const()[name = tensor("concat_12"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_3_internal_tensor_assign_2_stride_0 = const()[name = tensor("new_cache_3_internal_tensor_assign_2_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_3_internal_tensor_assign_2_begin_mask_0 = const()[name = tensor("new_cache_3_internal_tensor_assign_2_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_3_internal_tensor_assign_2_end_mask_0 = const()[name = tensor("new_cache_3_internal_tensor_assign_2_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_3_internal_tensor_assign_2_squeeze_mask_0 = const()[name = tensor("new_cache_3_internal_tensor_assign_2_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_15 = const()[name = tensor("shape_15"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_3 = const()[name = tensor("reduce_prod_3"), val = tensor(1048576)]; + tensor range_1d_3_start_0 = const()[name = tensor("range_1d_3_start_0"), val = tensor(0)]; + tensor range_1d_3_step_0 = const()[name = tensor("range_1d_3_step_0"), val = tensor(1)]; + tensor range_1d_3 = range_1d(end = reduce_prod_3, start = range_1d_3_start_0, step = range_1d_3_step_0)[name = tensor("range_1d_3")]; + tensor reshape_15 = reshape(shape = shape_15, x = range_1d_3)[name = tensor("reshape_15")]; + tensor slice_by_index_3 = slice_by_index(begin = concat_11, begin_mask = new_cache_3_internal_tensor_assign_2_begin_mask_0, end = concat_12, end_mask = new_cache_3_internal_tensor_assign_2_end_mask_0, squeeze_mask = new_cache_3_internal_tensor_assign_2_squeeze_mask_0, stride = new_cache_3_internal_tensor_assign_2_stride_0, x = reshape_15)[name = tensor("slice_by_index_3")]; + tensor reshape_16_shape_0 = const()[name = tensor("reshape_16_shape_0"), val = tensor([-1])]; + tensor reshape_16 = reshape(shape = reshape_16_shape_0, x = slice_by_index_3)[name = tensor("reshape_16")]; + tensor reshape_17_shape_0 = const()[name = tensor("reshape_17_shape_0"), val = tensor([-1])]; + tensor reshape_17_cast_fp16 = reshape(shape = reshape_17_shape_0, x = var_761_cast_fp16)[name = tensor("reshape_17_cast_fp16")]; + tensor reshape_18_shape_0 = const()[name = tensor("reshape_18_shape_0"), val = tensor([-1])]; + tensor reshape_18_cast_fp16 = reshape(shape = reshape_18_shape_0, x = reshape_14_cast_fp16)[name = tensor("reshape_18_cast_fp16")]; + tensor scatter_3_mode_0 = const()[name = tensor("scatter_3_mode_0"), val = tensor("update")]; + tensor scatter_3_axis_0 = const()[name = tensor("scatter_3_axis_0"), val = tensor(0)]; + tensor scatter_3_validate_indices_0 = const()[name = tensor("scatter_3_validate_indices_0"), val = tensor(false)]; + tensor scatter_3_cast_fp16 = scatter(axis = scatter_3_axis_0, data = reshape_18_cast_fp16, indices = reshape_16, mode = scatter_3_mode_0, updates = reshape_17_cast_fp16, validate_indices = scatter_3_validate_indices_0)[name = tensor("scatter_3_cast_fp16")]; + tensor reshape_19_cast_fp16 = reshape(shape = shape_15, x = scatter_3_cast_fp16)[name = tensor("reshape_19_cast_fp16")]; + tensor reshape_19_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("reshape_19_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor keys_7_begin_0 = const()[name = tensor("keys_7_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor keys_7_end_0 = const()[name = tensor("keys_7_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor keys_7_end_mask_0 = const()[name = tensor("keys_7_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor keys_7_squeeze_mask_0 = const()[name = tensor("keys_7_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor keys_7_cast_fp16 = slice_by_index(begin = keys_7_begin_0, end = keys_7_end_0, end_mask = keys_7_end_mask_0, squeeze_mask = keys_7_squeeze_mask_0, x = reshape_19_cast_fp16)[name = tensor("keys_7_cast_fp16")]; + tensor values_7_begin_0 = const()[name = tensor("values_7_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor values_7_end_0 = const()[name = tensor("values_7_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor values_7_end_mask_0 = const()[name = tensor("values_7_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor values_7_squeeze_mask_0 = const()[name = tensor("values_7_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor values_7_cast_fp16 = slice_by_index(begin = values_7_begin_0, end = values_7_end_0, end_mask = values_7_end_mask_0, squeeze_mask = values_7_squeeze_mask_0, x = reshape_19_cast_fp16)[name = tensor("values_7_cast_fp16")]; + tensor var_773_cast_fp16 = not_equal(x = keys_7_cast_fp16, y = keys_7_cast_fp16)[name = tensor("op_773_cast_fp16")]; + tensor keys_9_cast_fp16 = select(a = var_360_to_fp16, b = keys_7_cast_fp16, cond = var_773_cast_fp16)[name = tensor("keys_9_cast_fp16")]; + tensor var_781_cast_fp16 = not_equal(x = values_7_cast_fp16, y = values_7_cast_fp16)[name = tensor("op_781_cast_fp16")]; + tensor values_9_cast_fp16 = select(a = var_360_to_fp16, b = values_7_cast_fp16, cond = var_781_cast_fp16)[name = tensor("values_9_cast_fp16")]; + tensor var_805 = const()[name = tensor("op_805"), val = tensor([0, 2, 1, 3])]; + tensor var_818 = const()[name = tensor("op_818"), val = tensor([1, 1, 1])]; + tensor var_819_cast_fp16 = reshape(shape = var_818, x = position1_to_fp16)[name = tensor("op_819_cast_fp16")]; + tensor var_836_to_fp16 = const()[name = tensor("op_836_to_fp16"), val = tensor(0x1p+0)]; + tensor valid_len_3_cast_fp16 = add(x = var_819_cast_fp16, y = var_836_to_fp16)[name = tensor("valid_len_3_cast_fp16")]; + tensor valid_mask_3_cast_fp16 = less(x = k_positions_1_promoted_to_fp16, y = valid_len_3_cast_fp16)[name = tensor("valid_mask_3_cast_fp16")]; + tensor causal_mask_3_cast_fp16 = less_equal(x = k_positions_1_promoted_to_fp16, y = var_819_cast_fp16)[name = tensor("causal_mask_3_cast_fp16")]; + tensor attn_mask_5 = logical_and(x = valid_mask_3_cast_fp16, y = causal_mask_3_cast_fp16)[name = tensor("attn_mask_5")]; + tensor attn_mask_7_axes_0 = const()[name = tensor("attn_mask_7_axes_0"), val = tensor([1])]; + tensor attn_mask_7 = expand_dims(axes = attn_mask_7_axes_0, x = attn_mask_5)[name = tensor("attn_mask_7")]; + tensor var_854_transpose_x_0 = const()[name = tensor("op_854_transpose_x_0"), val = tensor(false)]; + tensor var_854_transpose_y_0 = const()[name = tensor("op_854_transpose_y_0"), val = tensor(false)]; + tensor transpose_20_perm_0 = const()[name = tensor("transpose_20_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_21_perm_0 = const()[name = tensor("transpose_21_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_21 = transpose(perm = transpose_21_perm_0, x = keys_9_cast_fp16)[name = tensor("transpose_47")]; + tensor transpose_20 = transpose(perm = transpose_20_perm_0, x = q_9_cast_fp16)[name = tensor("transpose_48")]; + tensor var_854_cast_fp16 = matmul(transpose_x = var_854_transpose_x_0, transpose_y = var_854_transpose_y_0, x = transpose_20, y = transpose_21)[name = tensor("op_854_cast_fp16")]; + tensor var_848_to_fp16 = const()[name = tensor("op_848_to_fp16"), val = tensor([0x1p-3])]; + tensor attn_weights_7_cast_fp16 = mul(x = var_854_cast_fp16, y = var_848_to_fp16)[name = tensor("attn_weights_7_cast_fp16")]; + tensor var_856 = logical_not(x = attn_mask_7)[name = tensor("op_856")]; + tensor var_857_to_fp16 = const()[name = tensor("op_857_to_fp16"), val = tensor(-inf)]; + tensor attn_weights_9_cast_fp16 = select(a = var_857_to_fp16, b = attn_weights_7_cast_fp16, cond = var_856)[name = tensor("attn_weights_9_cast_fp16")]; + tensor var_859 = const()[name = tensor("op_859"), val = tensor(-1)]; + tensor attn_weights_11_cast_fp16 = softmax(axis = var_859, x = attn_weights_9_cast_fp16)[name = tensor("attn_weights_11_cast_fp16")]; + tensor attn_output_3_transpose_x_0 = const()[name = tensor("attn_output_3_transpose_x_0"), val = tensor(false)]; + tensor attn_output_3_transpose_y_0 = const()[name = tensor("attn_output_3_transpose_y_0"), val = tensor(false)]; + tensor values_11_cast_fp16 = transpose(perm = var_805, x = values_9_cast_fp16)[name = tensor("transpose_49")]; + tensor attn_output_3_cast_fp16 = matmul(transpose_x = attn_output_3_transpose_x_0, transpose_y = attn_output_3_transpose_y_0, x = attn_weights_11_cast_fp16, y = values_11_cast_fp16)[name = tensor("attn_output_3_cast_fp16")]; + tensor var_867 = const()[name = tensor("op_867"), val = tensor([0, 2, 1, 3])]; + tensor var_870 = const()[name = tensor("op_870"), val = tensor([1, 1, 1024])]; + tensor var_868_cast_fp16 = transpose(perm = var_867, x = attn_output_3_cast_fp16)[name = tensor("transpose_46")]; + tensor input_15_cast_fp16 = reshape(shape = var_870, x = var_868_cast_fp16)[name = tensor("input_15_cast_fp16")]; + tensor linear_6_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = attn1_out_proj_weight_cast_fp16, x = input_15_cast_fp16)[name = tensor("linear_6_cast_fp16")]; + tensor var_876_to_fp16 = const()[name = tensor("op_876_to_fp16"), val = tensor(0x1p+0)]; + tensor var_877_cast_fp16 = add(x = position1_to_fp16, y = var_876_to_fp16)[name = tensor("op_877_cast_fp16")]; + tensor var_877_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_877_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor input_17_cast_fp16 = add(x = input_13_cast_fp16, y = linear_6_cast_fp16)[name = tensor("input_17_cast_fp16")]; + tensor input_19_axes_0 = const()[name = tensor("input_19_axes_0"), val = tensor([-1])]; + tensor norm1_2_weight_to_fp16 = const()[name = tensor("norm1_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(76764416)))]; + tensor norm1_2_bias_to_fp16 = const()[name = tensor("norm1_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(76766528)))]; + tensor var_881_to_fp16 = const()[name = tensor("op_881_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_19_cast_fp16 = layer_norm(axes = input_19_axes_0, beta = norm1_2_bias_to_fp16, epsilon = var_881_to_fp16, gamma = norm1_2_weight_to_fp16, x = input_17_cast_fp16)[name = tensor("input_19_cast_fp16")]; + tensor linear_7_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = linear1_1_weight_cast_fp16, x = input_19_cast_fp16)[name = tensor("linear_7_cast_fp16")]; + tensor input_21_mode_0 = const()[name = tensor("input_21_mode_0"), val = tensor("EXACT")]; + tensor input_21_cast_fp16 = gelu(mode = input_21_mode_0, x = linear_7_cast_fp16)[name = tensor("input_21_cast_fp16")]; + tensor linear_8_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = linear1_2_weight_cast_fp16, x = input_21_cast_fp16)[name = tensor("linear_8_cast_fp16")]; + tensor input_23_cast_fp16 = add(x = input_17_cast_fp16, y = linear_8_cast_fp16)[name = tensor("input_23_cast_fp16")]; + tensor x_5_axes_0 = const()[name = tensor("x_5_axes_0"), val = tensor([-1])]; + tensor norm2_1_weight_to_fp16 = const()[name = tensor("norm2_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(76768640)))]; + tensor norm2_1_bias_to_fp16 = const()[name = tensor("norm2_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(76770752)))]; + tensor var_898_to_fp16 = const()[name = tensor("op_898_to_fp16"), val = tensor(0x1.5p-17)]; + tensor x_5_cast_fp16 = layer_norm(axes = x_5_axes_0, beta = norm2_1_bias_to_fp16, epsilon = var_898_to_fp16, gamma = norm2_1_weight_to_fp16, x = input_23_cast_fp16)[name = tensor("x_5_cast_fp16")]; + tensor linear_9_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = attn2_in_proj_weight_cast_fp16, x = x_5_cast_fp16)[name = tensor("linear_9_cast_fp16")]; + tensor var_934 = const()[name = tensor("op_934"), val = tensor([1, 1, 3, 16, 64])]; + tensor qkv_5_cast_fp16 = reshape(shape = var_934, x = linear_9_cast_fp16)[name = tensor("qkv_5_cast_fp16")]; + tensor q_13_begin_0 = const()[name = tensor("q_13_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor q_13_end_0 = const()[name = tensor("q_13_end_0"), val = tensor([1, 1, 1, 16, 64])]; + tensor q_13_end_mask_0 = const()[name = tensor("q_13_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor q_13_squeeze_mask_0 = const()[name = tensor("q_13_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor q_13_cast_fp16 = slice_by_index(begin = q_13_begin_0, end = q_13_end_0, end_mask = q_13_end_mask_0, squeeze_mask = q_13_squeeze_mask_0, x = qkv_5_cast_fp16)[name = tensor("q_13_cast_fp16")]; + tensor k_9_begin_0 = const()[name = tensor("k_9_begin_0"), val = tensor([0, 0, 1, 0, 0])]; + tensor k_9_end_0 = const()[name = tensor("k_9_end_0"), val = tensor([1, 1, 2, 16, 64])]; + tensor k_9_end_mask_0 = const()[name = tensor("k_9_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor k_9_squeeze_mask_0 = const()[name = tensor("k_9_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor k_9_cast_fp16 = slice_by_index(begin = k_9_begin_0, end = k_9_end_0, end_mask = k_9_end_mask_0, squeeze_mask = k_9_squeeze_mask_0, x = qkv_5_cast_fp16)[name = tensor("k_9_cast_fp16")]; + tensor v_5_begin_0 = const()[name = tensor("v_5_begin_0"), val = tensor([0, 0, 2, 0, 0])]; + tensor v_5_end_0 = const()[name = tensor("v_5_end_0"), val = tensor([1, 1, 3, 16, 64])]; + tensor v_5_end_mask_0 = const()[name = tensor("v_5_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor v_5_squeeze_mask_0 = const()[name = tensor("v_5_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor v_5_cast_fp16 = slice_by_index(begin = v_5_begin_0, end = v_5_end_0, end_mask = v_5_end_mask_0, squeeze_mask = v_5_squeeze_mask_0, x = qkv_5_cast_fp16)[name = tensor("v_5_cast_fp16")]; + tensor var_1038 = const()[name = tensor("op_1038"), val = tensor([1, 1, 1, 1])]; + tensor position2_to_fp16_dtype_0 = const()[name = tensor("position2_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor position2_to_fp16 = cast(dtype = position2_to_fp16_dtype_0, x = position2)[name = tensor("cast_129")]; + tensor ts_17_cast_fp16 = reshape(shape = var_1038, x = position2_to_fp16)[name = tensor("ts_17_cast_fp16")]; + tensor var_1042 = const()[name = tensor("op_1042"), val = tensor([1, 1, 16, 32, 2])]; + tensor q_complex_5_cast_fp16 = reshape(shape = var_1042, x = q_13_cast_fp16)[name = tensor("q_complex_5_cast_fp16")]; + tensor var_1046 = const()[name = tensor("op_1046"), val = tensor([1, 1, 16, 32, 2])]; + tensor k_complex_5_cast_fp16 = reshape(shape = var_1046, x = k_9_cast_fp16)[name = tensor("k_complex_5_cast_fp16")]; + tensor var_1050_begin_0 = const()[name = tensor("op_1050_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_1050_end_0 = const()[name = tensor("op_1050_end_0"), val = tensor([1, 1, 16, 32, 1])]; + tensor var_1050_end_mask_0 = const()[name = tensor("op_1050_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_1050_squeeze_mask_0 = const()[name = tensor("op_1050_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_1050_cast_fp16 = slice_by_index(begin = var_1050_begin_0, end = var_1050_end_0, end_mask = var_1050_end_mask_0, squeeze_mask = var_1050_squeeze_mask_0, x = q_complex_5_cast_fp16)[name = tensor("op_1050_cast_fp16")]; + tensor var_1058_begin_0 = const()[name = tensor("op_1058_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_1058_end_0 = const()[name = tensor("op_1058_end_0"), val = tensor([1, 1, 16, 32, 2])]; + tensor var_1058_end_mask_0 = const()[name = tensor("op_1058_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_1058_squeeze_mask_0 = const()[name = tensor("op_1058_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_1058_cast_fp16 = slice_by_index(begin = var_1058_begin_0, end = var_1058_end_0, end_mask = var_1058_end_mask_0, squeeze_mask = var_1058_squeeze_mask_0, x = q_complex_5_cast_fp16)[name = tensor("op_1058_cast_fp16")]; + tensor var_1066_begin_0 = const()[name = tensor("op_1066_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_1066_end_0 = const()[name = tensor("op_1066_end_0"), val = tensor([1, 1, 16, 32, 1])]; + tensor var_1066_end_mask_0 = const()[name = tensor("op_1066_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_1066_squeeze_mask_0 = const()[name = tensor("op_1066_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_1066_cast_fp16 = slice_by_index(begin = var_1066_begin_0, end = var_1066_end_0, end_mask = var_1066_end_mask_0, squeeze_mask = var_1066_squeeze_mask_0, x = k_complex_5_cast_fp16)[name = tensor("op_1066_cast_fp16")]; + tensor var_1074_begin_0 = const()[name = tensor("op_1074_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_1074_end_0 = const()[name = tensor("op_1074_end_0"), val = tensor([1, 1, 16, 32, 2])]; + tensor var_1074_end_mask_0 = const()[name = tensor("op_1074_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_1074_squeeze_mask_0 = const()[name = tensor("op_1074_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_1074_cast_fp16 = slice_by_index(begin = var_1074_begin_0, end = var_1074_end_0, end_mask = var_1074_end_mask_0, squeeze_mask = var_1074_squeeze_mask_0, x = k_complex_5_cast_fp16)[name = tensor("op_1074_cast_fp16")]; + tensor freqs_5_to_fp16 = const()[name = tensor("freqs_5_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(76772864)))]; + tensor var_1080_cast_fp16 = mul(x = freqs_5_to_fp16, y = ts_17_cast_fp16)[name = tensor("op_1080_cast_fp16")]; + tensor rotr_5_cast_fp16 = cos(x = var_1080_cast_fp16)[name = tensor("rotr_5_cast_fp16")]; + tensor roti_5_cast_fp16 = sin(x = var_1080_cast_fp16)[name = tensor("roti_5_cast_fp16")]; + tensor var_1084_cast_fp16 = mul(x = var_1050_cast_fp16, y = rotr_5_cast_fp16)[name = tensor("op_1084_cast_fp16")]; + tensor var_1085_cast_fp16 = mul(x = var_1058_cast_fp16, y = roti_5_cast_fp16)[name = tensor("op_1085_cast_fp16")]; + tensor qor_9_cast_fp16 = sub(x = var_1084_cast_fp16, y = var_1085_cast_fp16)[name = tensor("qor_9_cast_fp16")]; + tensor var_1088_cast_fp16 = mul(x = var_1050_cast_fp16, y = roti_5_cast_fp16)[name = tensor("op_1088_cast_fp16")]; + tensor var_1089_cast_fp16 = mul(x = var_1058_cast_fp16, y = rotr_5_cast_fp16)[name = tensor("op_1089_cast_fp16")]; + tensor qoi_9_cast_fp16 = add(x = var_1088_cast_fp16, y = var_1089_cast_fp16)[name = tensor("qoi_9_cast_fp16")]; + tensor var_1092_cast_fp16 = mul(x = var_1066_cast_fp16, y = rotr_5_cast_fp16)[name = tensor("op_1092_cast_fp16")]; + tensor var_1093_cast_fp16 = mul(x = var_1074_cast_fp16, y = roti_5_cast_fp16)[name = tensor("op_1093_cast_fp16")]; + tensor kor_9_cast_fp16 = sub(x = var_1092_cast_fp16, y = var_1093_cast_fp16)[name = tensor("kor_9_cast_fp16")]; + tensor var_1096_cast_fp16 = mul(x = var_1066_cast_fp16, y = roti_5_cast_fp16)[name = tensor("op_1096_cast_fp16")]; + tensor var_1097_cast_fp16 = mul(x = var_1074_cast_fp16, y = rotr_5_cast_fp16)[name = tensor("op_1097_cast_fp16")]; + tensor koi_9_cast_fp16 = add(x = var_1096_cast_fp16, y = var_1097_cast_fp16)[name = tensor("koi_9_cast_fp16")]; + tensor qo_5_axis_0 = const()[name = tensor("qo_5_axis_0"), val = tensor(-1)]; + tensor qo_5_cast_fp16 = stack(axis = qo_5_axis_0, values = (qor_9_cast_fp16, qoi_9_cast_fp16))[name = tensor("qo_5_cast_fp16")]; + tensor ko_5_axis_0 = const()[name = tensor("ko_5_axis_0"), val = tensor(-1)]; + tensor ko_5_cast_fp16 = stack(axis = ko_5_axis_0, values = (kor_9_cast_fp16, koi_9_cast_fp16))[name = tensor("ko_5_cast_fp16")]; + tensor var_1126 = const()[name = tensor("op_1126"), val = tensor([1, 1, 16, 64])]; + tensor q_15_cast_fp16 = reshape(shape = var_1126, x = qo_5_cast_fp16)[name = tensor("q_15_cast_fp16")]; + tensor var_1128 = const()[name = tensor("op_1128"), val = tensor([1, 1, 16, 64])]; + tensor k_11_cast_fp16 = reshape(shape = var_1128, x = ko_5_cast_fp16)[name = tensor("k_11_cast_fp16")]; + tensor _inversed_1150_y_0_to_fp16 = const()[name = tensor("_inversed_1150_y_0_to_fp16"), val = tensor(0x1p-9)]; + tensor _inversed_1150_cast_fp16 = mul(x = ts_17_cast_fp16, y = _inversed_1150_y_0_to_fp16)[name = tensor("_inversed_1150_cast_fp16")]; + tensor var_1151_cast_fp16 = floor(x = _inversed_1150_cast_fp16)[name = tensor("op_1151_cast_fp16")]; + tensor var_1152_to_fp16 = const()[name = tensor("op_1152_to_fp16"), val = tensor(0x1p+9)]; + tensor var_1153_cast_fp16 = mul(x = var_1151_cast_fp16, y = var_1152_to_fp16)[name = tensor("op_1153_cast_fp16")]; + tensor write_indices_float_11_cast_fp16 = sub(x = ts_17_cast_fp16, y = var_1153_cast_fp16)[name = tensor("write_indices_float_11_cast_fp16")]; + tensor var_1160_dtype_0 = const()[name = tensor("op_1160_dtype_0"), val = tensor("int32")]; + tensor write_indices_5_reps_0 = const()[name = tensor("write_indices_5_reps_0"), val = tensor([1, 1, 16, 64])]; + tensor write_indices_float_11_cast_fp16_to_int32 = cast(dtype = var_1160_dtype_0, x = write_indices_float_11_cast_fp16)[name = tensor("cast_128")]; + tensor write_indices_5 = tile(reps = write_indices_5_reps_0, x = write_indices_float_11_cast_fp16_to_int32)[name = tensor("write_indices_5")]; + tensor var_1168_begin_0 = const()[name = tensor("op_1168_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_1168_end_0 = const()[name = tensor("op_1168_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor var_1168_end_mask_0 = const()[name = tensor("op_1168_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_1168_squeeze_mask_0 = const()[name = tensor("op_1168_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor cache2_to_fp16_dtype_0 = const()[name = tensor("cache2_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor cache2_to_fp16 = cast(dtype = cache2_to_fp16_dtype_0, x = cache2)[name = tensor("cast_127")]; + tensor var_1168_cast_fp16 = slice_by_index(begin = var_1168_begin_0, end = var_1168_end_0, end_mask = var_1168_end_mask_0, squeeze_mask = var_1168_squeeze_mask_0, x = cache2_to_fp16)[name = tensor("op_1168_cast_fp16")]; + tensor var_1170_axis_0 = const()[name = tensor("op_1170_axis_0"), val = tensor(1)]; + tensor var_1170_mode_0 = const()[name = tensor("op_1170_mode_0"), val = tensor("update")]; + tensor var_1170_validate_indices_0 = const()[name = tensor("op_1170_validate_indices_0"), val = tensor(false)]; + tensor var_1170_cast_fp16 = scatter_along_axis(axis = var_1170_axis_0, data = var_1168_cast_fp16, indices = write_indices_5, mode = var_1170_mode_0, updates = k_11_cast_fp16, validate_indices = var_1170_validate_indices_0)[name = tensor("op_1170_cast_fp16")]; + tensor concat_16 = const()[name = tensor("concat_16"), val = tensor([0, 0, 0, 0, 0])]; + tensor concat_17 = const()[name = tensor("concat_17"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_5_internal_tensor_assign_1_stride_0 = const()[name = tensor("new_cache_5_internal_tensor_assign_1_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_5_internal_tensor_assign_1_begin_mask_0 = const()[name = tensor("new_cache_5_internal_tensor_assign_1_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_5_internal_tensor_assign_1_end_mask_0 = const()[name = tensor("new_cache_5_internal_tensor_assign_1_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_5_internal_tensor_assign_1_squeeze_mask_0 = const()[name = tensor("new_cache_5_internal_tensor_assign_1_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_16 = const()[name = tensor("shape_16"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_4 = const()[name = tensor("reduce_prod_4"), val = tensor(1048576)]; + tensor range_1d_4_start_0 = const()[name = tensor("range_1d_4_start_0"), val = tensor(0)]; + tensor range_1d_4_step_0 = const()[name = tensor("range_1d_4_step_0"), val = tensor(1)]; + tensor range_1d_4 = range_1d(end = reduce_prod_4, start = range_1d_4_start_0, step = range_1d_4_step_0)[name = tensor("range_1d_4")]; + tensor reshape_20 = reshape(shape = shape_16, x = range_1d_4)[name = tensor("reshape_20")]; + tensor slice_by_index_4 = slice_by_index(begin = concat_16, begin_mask = new_cache_5_internal_tensor_assign_1_begin_mask_0, end = concat_17, end_mask = new_cache_5_internal_tensor_assign_1_end_mask_0, squeeze_mask = new_cache_5_internal_tensor_assign_1_squeeze_mask_0, stride = new_cache_5_internal_tensor_assign_1_stride_0, x = reshape_20)[name = tensor("slice_by_index_4")]; + tensor reshape_21_shape_0 = const()[name = tensor("reshape_21_shape_0"), val = tensor([-1])]; + tensor reshape_21 = reshape(shape = reshape_21_shape_0, x = slice_by_index_4)[name = tensor("reshape_21")]; + tensor reshape_22_shape_0 = const()[name = tensor("reshape_22_shape_0"), val = tensor([-1])]; + tensor reshape_22_cast_fp16 = reshape(shape = reshape_22_shape_0, x = var_1170_cast_fp16)[name = tensor("reshape_22_cast_fp16")]; + tensor reshape_23_shape_0 = const()[name = tensor("reshape_23_shape_0"), val = tensor([-1])]; + tensor reshape_23_cast_fp16 = reshape(shape = reshape_23_shape_0, x = cache2_to_fp16)[name = tensor("reshape_23_cast_fp16")]; + tensor scatter_4_mode_0 = const()[name = tensor("scatter_4_mode_0"), val = tensor("update")]; + tensor scatter_4_axis_0 = const()[name = tensor("scatter_4_axis_0"), val = tensor(0)]; + tensor scatter_4_validate_indices_0 = const()[name = tensor("scatter_4_validate_indices_0"), val = tensor(false)]; + tensor scatter_4_cast_fp16 = scatter(axis = scatter_4_axis_0, data = reshape_23_cast_fp16, indices = reshape_21, mode = scatter_4_mode_0, updates = reshape_22_cast_fp16, validate_indices = scatter_4_validate_indices_0)[name = tensor("scatter_4_cast_fp16")]; + tensor reshape_24_cast_fp16 = reshape(shape = shape_16, x = scatter_4_cast_fp16)[name = tensor("reshape_24_cast_fp16")]; + tensor var_1178_begin_0 = const()[name = tensor("op_1178_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor var_1178_end_0 = const()[name = tensor("op_1178_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor var_1178_end_mask_0 = const()[name = tensor("op_1178_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_1178_squeeze_mask_0 = const()[name = tensor("op_1178_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor var_1178_cast_fp16 = slice_by_index(begin = var_1178_begin_0, end = var_1178_end_0, end_mask = var_1178_end_mask_0, squeeze_mask = var_1178_squeeze_mask_0, x = reshape_24_cast_fp16)[name = tensor("op_1178_cast_fp16")]; + tensor var_1180_axis_0 = const()[name = tensor("op_1180_axis_0"), val = tensor(1)]; + tensor var_1180_mode_0 = const()[name = tensor("op_1180_mode_0"), val = tensor("update")]; + tensor var_1180_validate_indices_0 = const()[name = tensor("op_1180_validate_indices_0"), val = tensor(false)]; + tensor var_1180_cast_fp16 = scatter_along_axis(axis = var_1180_axis_0, data = var_1178_cast_fp16, indices = write_indices_5, mode = var_1180_mode_0, updates = v_5_cast_fp16, validate_indices = var_1180_validate_indices_0)[name = tensor("op_1180_cast_fp16")]; + tensor concat_18 = const()[name = tensor("concat_18"), val = tensor([1, 0, 0, 0, 0])]; + tensor concat_19 = const()[name = tensor("concat_19"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_5_internal_tensor_assign_2_stride_0 = const()[name = tensor("new_cache_5_internal_tensor_assign_2_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_5_internal_tensor_assign_2_begin_mask_0 = const()[name = tensor("new_cache_5_internal_tensor_assign_2_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_5_internal_tensor_assign_2_end_mask_0 = const()[name = tensor("new_cache_5_internal_tensor_assign_2_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_5_internal_tensor_assign_2_squeeze_mask_0 = const()[name = tensor("new_cache_5_internal_tensor_assign_2_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_17 = const()[name = tensor("shape_17"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_5 = const()[name = tensor("reduce_prod_5"), val = tensor(1048576)]; + tensor range_1d_5_start_0 = const()[name = tensor("range_1d_5_start_0"), val = tensor(0)]; + tensor range_1d_5_step_0 = const()[name = tensor("range_1d_5_step_0"), val = tensor(1)]; + tensor range_1d_5 = range_1d(end = reduce_prod_5, start = range_1d_5_start_0, step = range_1d_5_step_0)[name = tensor("range_1d_5")]; + tensor reshape_25 = reshape(shape = shape_17, x = range_1d_5)[name = tensor("reshape_25")]; + tensor slice_by_index_5 = slice_by_index(begin = concat_18, begin_mask = new_cache_5_internal_tensor_assign_2_begin_mask_0, end = concat_19, end_mask = new_cache_5_internal_tensor_assign_2_end_mask_0, squeeze_mask = new_cache_5_internal_tensor_assign_2_squeeze_mask_0, stride = new_cache_5_internal_tensor_assign_2_stride_0, x = reshape_25)[name = tensor("slice_by_index_5")]; + tensor reshape_26_shape_0 = const()[name = tensor("reshape_26_shape_0"), val = tensor([-1])]; + tensor reshape_26 = reshape(shape = reshape_26_shape_0, x = slice_by_index_5)[name = tensor("reshape_26")]; + tensor reshape_27_shape_0 = const()[name = tensor("reshape_27_shape_0"), val = tensor([-1])]; + tensor reshape_27_cast_fp16 = reshape(shape = reshape_27_shape_0, x = var_1180_cast_fp16)[name = tensor("reshape_27_cast_fp16")]; + tensor reshape_28_shape_0 = const()[name = tensor("reshape_28_shape_0"), val = tensor([-1])]; + tensor reshape_28_cast_fp16 = reshape(shape = reshape_28_shape_0, x = reshape_24_cast_fp16)[name = tensor("reshape_28_cast_fp16")]; + tensor scatter_5_mode_0 = const()[name = tensor("scatter_5_mode_0"), val = tensor("update")]; + tensor scatter_5_axis_0 = const()[name = tensor("scatter_5_axis_0"), val = tensor(0)]; + tensor scatter_5_validate_indices_0 = const()[name = tensor("scatter_5_validate_indices_0"), val = tensor(false)]; + tensor scatter_5_cast_fp16 = scatter(axis = scatter_5_axis_0, data = reshape_28_cast_fp16, indices = reshape_26, mode = scatter_5_mode_0, updates = reshape_27_cast_fp16, validate_indices = scatter_5_validate_indices_0)[name = tensor("scatter_5_cast_fp16")]; + tensor reshape_29_cast_fp16 = reshape(shape = shape_17, x = scatter_5_cast_fp16)[name = tensor("reshape_29_cast_fp16")]; + tensor reshape_29_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("reshape_29_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor keys_13_begin_0 = const()[name = tensor("keys_13_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor keys_13_end_0 = const()[name = tensor("keys_13_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor keys_13_end_mask_0 = const()[name = tensor("keys_13_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor keys_13_squeeze_mask_0 = const()[name = tensor("keys_13_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor keys_13_cast_fp16 = slice_by_index(begin = keys_13_begin_0, end = keys_13_end_0, end_mask = keys_13_end_mask_0, squeeze_mask = keys_13_squeeze_mask_0, x = reshape_29_cast_fp16)[name = tensor("keys_13_cast_fp16")]; + tensor values_13_begin_0 = const()[name = tensor("values_13_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor values_13_end_0 = const()[name = tensor("values_13_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor values_13_end_mask_0 = const()[name = tensor("values_13_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor values_13_squeeze_mask_0 = const()[name = tensor("values_13_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor values_13_cast_fp16 = slice_by_index(begin = values_13_begin_0, end = values_13_end_0, end_mask = values_13_end_mask_0, squeeze_mask = values_13_squeeze_mask_0, x = reshape_29_cast_fp16)[name = tensor("values_13_cast_fp16")]; + tensor var_1192_cast_fp16 = not_equal(x = keys_13_cast_fp16, y = keys_13_cast_fp16)[name = tensor("op_1192_cast_fp16")]; + tensor keys_15_cast_fp16 = select(a = var_360_to_fp16, b = keys_13_cast_fp16, cond = var_1192_cast_fp16)[name = tensor("keys_15_cast_fp16")]; + tensor var_1200_cast_fp16 = not_equal(x = values_13_cast_fp16, y = values_13_cast_fp16)[name = tensor("op_1200_cast_fp16")]; + tensor values_15_cast_fp16 = select(a = var_360_to_fp16, b = values_13_cast_fp16, cond = var_1200_cast_fp16)[name = tensor("values_15_cast_fp16")]; + tensor var_1224 = const()[name = tensor("op_1224"), val = tensor([0, 2, 1, 3])]; + tensor var_1237 = const()[name = tensor("op_1237"), val = tensor([1, 1, 1])]; + tensor var_1238_cast_fp16 = reshape(shape = var_1237, x = position2_to_fp16)[name = tensor("op_1238_cast_fp16")]; + tensor var_1255_to_fp16 = const()[name = tensor("op_1255_to_fp16"), val = tensor(0x1p+0)]; + tensor valid_len_5_cast_fp16 = add(x = var_1238_cast_fp16, y = var_1255_to_fp16)[name = tensor("valid_len_5_cast_fp16")]; + tensor valid_mask_5_cast_fp16 = less(x = k_positions_1_promoted_to_fp16, y = valid_len_5_cast_fp16)[name = tensor("valid_mask_5_cast_fp16")]; + tensor causal_mask_5_cast_fp16 = less_equal(x = k_positions_1_promoted_to_fp16, y = var_1238_cast_fp16)[name = tensor("causal_mask_5_cast_fp16")]; + tensor attn_mask_9 = logical_and(x = valid_mask_5_cast_fp16, y = causal_mask_5_cast_fp16)[name = tensor("attn_mask_9")]; + tensor attn_mask_11_axes_0 = const()[name = tensor("attn_mask_11_axes_0"), val = tensor([1])]; + tensor attn_mask_11 = expand_dims(axes = attn_mask_11_axes_0, x = attn_mask_9)[name = tensor("attn_mask_11")]; + tensor var_1273_transpose_x_0 = const()[name = tensor("op_1273_transpose_x_0"), val = tensor(false)]; + tensor var_1273_transpose_y_0 = const()[name = tensor("op_1273_transpose_y_0"), val = tensor(false)]; + tensor transpose_22_perm_0 = const()[name = tensor("transpose_22_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_23_perm_0 = const()[name = tensor("transpose_23_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_23 = transpose(perm = transpose_23_perm_0, x = keys_15_cast_fp16)[name = tensor("transpose_43")]; + tensor transpose_22 = transpose(perm = transpose_22_perm_0, x = q_15_cast_fp16)[name = tensor("transpose_44")]; + tensor var_1273_cast_fp16 = matmul(transpose_x = var_1273_transpose_x_0, transpose_y = var_1273_transpose_y_0, x = transpose_22, y = transpose_23)[name = tensor("op_1273_cast_fp16")]; + tensor var_1267_to_fp16 = const()[name = tensor("op_1267_to_fp16"), val = tensor([0x1p-3])]; + tensor attn_weights_13_cast_fp16 = mul(x = var_1273_cast_fp16, y = var_1267_to_fp16)[name = tensor("attn_weights_13_cast_fp16")]; + tensor var_1275 = logical_not(x = attn_mask_11)[name = tensor("op_1275")]; + tensor var_1276_to_fp16 = const()[name = tensor("op_1276_to_fp16"), val = tensor(-inf)]; + tensor attn_weights_15_cast_fp16 = select(a = var_1276_to_fp16, b = attn_weights_13_cast_fp16, cond = var_1275)[name = tensor("attn_weights_15_cast_fp16")]; + tensor var_1278 = const()[name = tensor("op_1278"), val = tensor(-1)]; + tensor attn_weights_17_cast_fp16 = softmax(axis = var_1278, x = attn_weights_15_cast_fp16)[name = tensor("attn_weights_17_cast_fp16")]; + tensor attn_output_5_transpose_x_0 = const()[name = tensor("attn_output_5_transpose_x_0"), val = tensor(false)]; + tensor attn_output_5_transpose_y_0 = const()[name = tensor("attn_output_5_transpose_y_0"), val = tensor(false)]; + tensor values_17_cast_fp16 = transpose(perm = var_1224, x = values_15_cast_fp16)[name = tensor("transpose_45")]; + tensor attn_output_5_cast_fp16 = matmul(transpose_x = attn_output_5_transpose_x_0, transpose_y = attn_output_5_transpose_y_0, x = attn_weights_17_cast_fp16, y = values_17_cast_fp16)[name = tensor("attn_output_5_cast_fp16")]; + tensor var_1286 = const()[name = tensor("op_1286"), val = tensor([0, 2, 1, 3])]; + tensor var_1289 = const()[name = tensor("op_1289"), val = tensor([1, 1, 1024])]; + tensor var_1287_cast_fp16 = transpose(perm = var_1286, x = attn_output_5_cast_fp16)[name = tensor("transpose_42")]; + tensor input_25_cast_fp16 = reshape(shape = var_1289, x = var_1287_cast_fp16)[name = tensor("input_25_cast_fp16")]; + tensor linear_10_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = attn2_out_proj_weight_cast_fp16, x = input_25_cast_fp16)[name = tensor("linear_10_cast_fp16")]; + tensor var_1295_to_fp16 = const()[name = tensor("op_1295_to_fp16"), val = tensor(0x1p+0)]; + tensor var_1296_cast_fp16 = add(x = position2_to_fp16, y = var_1295_to_fp16)[name = tensor("op_1296_cast_fp16")]; + tensor var_1296_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_1296_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor input_27_cast_fp16 = add(x = input_23_cast_fp16, y = linear_10_cast_fp16)[name = tensor("input_27_cast_fp16")]; + tensor input_29_axes_0 = const()[name = tensor("input_29_axes_0"), val = tensor([-1])]; + tensor norm2_2_weight_to_fp16 = const()[name = tensor("norm2_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(76772992)))]; + tensor norm2_2_bias_to_fp16 = const()[name = tensor("norm2_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(76775104)))]; + tensor var_1300_to_fp16 = const()[name = tensor("op_1300_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_29_cast_fp16 = layer_norm(axes = input_29_axes_0, beta = norm2_2_bias_to_fp16, epsilon = var_1300_to_fp16, gamma = norm2_2_weight_to_fp16, x = input_27_cast_fp16)[name = tensor("input_29_cast_fp16")]; + tensor linear_11_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = linear2_1_weight_cast_fp16, x = input_29_cast_fp16)[name = tensor("linear_11_cast_fp16")]; + tensor input_31_mode_0 = const()[name = tensor("input_31_mode_0"), val = tensor("EXACT")]; + tensor input_31_cast_fp16 = gelu(mode = input_31_mode_0, x = linear_11_cast_fp16)[name = tensor("input_31_cast_fp16")]; + tensor linear_12_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = linear2_2_weight_cast_fp16, x = input_31_cast_fp16)[name = tensor("linear_12_cast_fp16")]; + tensor input_33_cast_fp16 = add(x = input_27_cast_fp16, y = linear_12_cast_fp16)[name = tensor("input_33_cast_fp16")]; + tensor x_7_axes_0 = const()[name = tensor("x_7_axes_0"), val = tensor([-1])]; + tensor norm3_1_weight_to_fp16 = const()[name = tensor("norm3_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(76777216)))]; + tensor norm3_1_bias_to_fp16 = const()[name = tensor("norm3_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(76779328)))]; + tensor var_1317_to_fp16 = const()[name = tensor("op_1317_to_fp16"), val = tensor(0x1.5p-17)]; + tensor x_7_cast_fp16 = layer_norm(axes = x_7_axes_0, beta = norm3_1_bias_to_fp16, epsilon = var_1317_to_fp16, gamma = norm3_1_weight_to_fp16, x = input_33_cast_fp16)[name = tensor("x_7_cast_fp16")]; + tensor linear_13_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = attn3_in_proj_weight_cast_fp16, x = x_7_cast_fp16)[name = tensor("linear_13_cast_fp16")]; + tensor var_1353 = const()[name = tensor("op_1353"), val = tensor([1, 1, 3, 16, 64])]; + tensor qkv_7_cast_fp16 = reshape(shape = var_1353, x = linear_13_cast_fp16)[name = tensor("qkv_7_cast_fp16")]; + tensor q_19_begin_0 = const()[name = tensor("q_19_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor q_19_end_0 = const()[name = tensor("q_19_end_0"), val = tensor([1, 1, 1, 16, 64])]; + tensor q_19_end_mask_0 = const()[name = tensor("q_19_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor q_19_squeeze_mask_0 = const()[name = tensor("q_19_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor q_19_cast_fp16 = slice_by_index(begin = q_19_begin_0, end = q_19_end_0, end_mask = q_19_end_mask_0, squeeze_mask = q_19_squeeze_mask_0, x = qkv_7_cast_fp16)[name = tensor("q_19_cast_fp16")]; + tensor k_13_begin_0 = const()[name = tensor("k_13_begin_0"), val = tensor([0, 0, 1, 0, 0])]; + tensor k_13_end_0 = const()[name = tensor("k_13_end_0"), val = tensor([1, 1, 2, 16, 64])]; + tensor k_13_end_mask_0 = const()[name = tensor("k_13_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor k_13_squeeze_mask_0 = const()[name = tensor("k_13_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor k_13_cast_fp16 = slice_by_index(begin = k_13_begin_0, end = k_13_end_0, end_mask = k_13_end_mask_0, squeeze_mask = k_13_squeeze_mask_0, x = qkv_7_cast_fp16)[name = tensor("k_13_cast_fp16")]; + tensor v_7_begin_0 = const()[name = tensor("v_7_begin_0"), val = tensor([0, 0, 2, 0, 0])]; + tensor v_7_end_0 = const()[name = tensor("v_7_end_0"), val = tensor([1, 1, 3, 16, 64])]; + tensor v_7_end_mask_0 = const()[name = tensor("v_7_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor v_7_squeeze_mask_0 = const()[name = tensor("v_7_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor v_7_cast_fp16 = slice_by_index(begin = v_7_begin_0, end = v_7_end_0, end_mask = v_7_end_mask_0, squeeze_mask = v_7_squeeze_mask_0, x = qkv_7_cast_fp16)[name = tensor("v_7_cast_fp16")]; + tensor var_1457 = const()[name = tensor("op_1457"), val = tensor([1, 1, 1, 1])]; + tensor position3_to_fp16_dtype_0 = const()[name = tensor("position3_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor position3_to_fp16 = cast(dtype = position3_to_fp16_dtype_0, x = position3)[name = tensor("cast_124")]; + tensor ts_23_cast_fp16 = reshape(shape = var_1457, x = position3_to_fp16)[name = tensor("ts_23_cast_fp16")]; + tensor var_1461 = const()[name = tensor("op_1461"), val = tensor([1, 1, 16, 32, 2])]; + tensor q_complex_7_cast_fp16 = reshape(shape = var_1461, x = q_19_cast_fp16)[name = tensor("q_complex_7_cast_fp16")]; + tensor var_1465 = const()[name = tensor("op_1465"), val = tensor([1, 1, 16, 32, 2])]; + tensor k_complex_7_cast_fp16 = reshape(shape = var_1465, x = k_13_cast_fp16)[name = tensor("k_complex_7_cast_fp16")]; + tensor var_1469_begin_0 = const()[name = tensor("op_1469_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_1469_end_0 = const()[name = tensor("op_1469_end_0"), val = tensor([1, 1, 16, 32, 1])]; + tensor var_1469_end_mask_0 = const()[name = tensor("op_1469_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_1469_squeeze_mask_0 = const()[name = tensor("op_1469_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_1469_cast_fp16 = slice_by_index(begin = var_1469_begin_0, end = var_1469_end_0, end_mask = var_1469_end_mask_0, squeeze_mask = var_1469_squeeze_mask_0, x = q_complex_7_cast_fp16)[name = tensor("op_1469_cast_fp16")]; + tensor var_1477_begin_0 = const()[name = tensor("op_1477_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_1477_end_0 = const()[name = tensor("op_1477_end_0"), val = tensor([1, 1, 16, 32, 2])]; + tensor var_1477_end_mask_0 = const()[name = tensor("op_1477_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_1477_squeeze_mask_0 = const()[name = tensor("op_1477_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_1477_cast_fp16 = slice_by_index(begin = var_1477_begin_0, end = var_1477_end_0, end_mask = var_1477_end_mask_0, squeeze_mask = var_1477_squeeze_mask_0, x = q_complex_7_cast_fp16)[name = tensor("op_1477_cast_fp16")]; + tensor var_1485_begin_0 = const()[name = tensor("op_1485_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_1485_end_0 = const()[name = tensor("op_1485_end_0"), val = tensor([1, 1, 16, 32, 1])]; + tensor var_1485_end_mask_0 = const()[name = tensor("op_1485_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_1485_squeeze_mask_0 = const()[name = tensor("op_1485_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_1485_cast_fp16 = slice_by_index(begin = var_1485_begin_0, end = var_1485_end_0, end_mask = var_1485_end_mask_0, squeeze_mask = var_1485_squeeze_mask_0, x = k_complex_7_cast_fp16)[name = tensor("op_1485_cast_fp16")]; + tensor var_1493_begin_0 = const()[name = tensor("op_1493_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_1493_end_0 = const()[name = tensor("op_1493_end_0"), val = tensor([1, 1, 16, 32, 2])]; + tensor var_1493_end_mask_0 = const()[name = tensor("op_1493_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_1493_squeeze_mask_0 = const()[name = tensor("op_1493_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_1493_cast_fp16 = slice_by_index(begin = var_1493_begin_0, end = var_1493_end_0, end_mask = var_1493_end_mask_0, squeeze_mask = var_1493_squeeze_mask_0, x = k_complex_7_cast_fp16)[name = tensor("op_1493_cast_fp16")]; + tensor freqs_7_to_fp16 = const()[name = tensor("freqs_7_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(76781440)))]; + tensor var_1499_cast_fp16 = mul(x = freqs_7_to_fp16, y = ts_23_cast_fp16)[name = tensor("op_1499_cast_fp16")]; + tensor rotr_7_cast_fp16 = cos(x = var_1499_cast_fp16)[name = tensor("rotr_7_cast_fp16")]; + tensor roti_7_cast_fp16 = sin(x = var_1499_cast_fp16)[name = tensor("roti_7_cast_fp16")]; + tensor var_1503_cast_fp16 = mul(x = var_1469_cast_fp16, y = rotr_7_cast_fp16)[name = tensor("op_1503_cast_fp16")]; + tensor var_1504_cast_fp16 = mul(x = var_1477_cast_fp16, y = roti_7_cast_fp16)[name = tensor("op_1504_cast_fp16")]; + tensor qor_13_cast_fp16 = sub(x = var_1503_cast_fp16, y = var_1504_cast_fp16)[name = tensor("qor_13_cast_fp16")]; + tensor var_1507_cast_fp16 = mul(x = var_1469_cast_fp16, y = roti_7_cast_fp16)[name = tensor("op_1507_cast_fp16")]; + tensor var_1508_cast_fp16 = mul(x = var_1477_cast_fp16, y = rotr_7_cast_fp16)[name = tensor("op_1508_cast_fp16")]; + tensor qoi_13_cast_fp16 = add(x = var_1507_cast_fp16, y = var_1508_cast_fp16)[name = tensor("qoi_13_cast_fp16")]; + tensor var_1511_cast_fp16 = mul(x = var_1485_cast_fp16, y = rotr_7_cast_fp16)[name = tensor("op_1511_cast_fp16")]; + tensor var_1512_cast_fp16 = mul(x = var_1493_cast_fp16, y = roti_7_cast_fp16)[name = tensor("op_1512_cast_fp16")]; + tensor kor_13_cast_fp16 = sub(x = var_1511_cast_fp16, y = var_1512_cast_fp16)[name = tensor("kor_13_cast_fp16")]; + tensor var_1515_cast_fp16 = mul(x = var_1485_cast_fp16, y = roti_7_cast_fp16)[name = tensor("op_1515_cast_fp16")]; + tensor var_1516_cast_fp16 = mul(x = var_1493_cast_fp16, y = rotr_7_cast_fp16)[name = tensor("op_1516_cast_fp16")]; + tensor koi_13_cast_fp16 = add(x = var_1515_cast_fp16, y = var_1516_cast_fp16)[name = tensor("koi_13_cast_fp16")]; + tensor qo_7_axis_0 = const()[name = tensor("qo_7_axis_0"), val = tensor(-1)]; + tensor qo_7_cast_fp16 = stack(axis = qo_7_axis_0, values = (qor_13_cast_fp16, qoi_13_cast_fp16))[name = tensor("qo_7_cast_fp16")]; + tensor ko_7_axis_0 = const()[name = tensor("ko_7_axis_0"), val = tensor(-1)]; + tensor ko_7_cast_fp16 = stack(axis = ko_7_axis_0, values = (kor_13_cast_fp16, koi_13_cast_fp16))[name = tensor("ko_7_cast_fp16")]; + tensor var_1545 = const()[name = tensor("op_1545"), val = tensor([1, 1, 16, 64])]; + tensor q_21_cast_fp16 = reshape(shape = var_1545, x = qo_7_cast_fp16)[name = tensor("q_21_cast_fp16")]; + tensor var_1547 = const()[name = tensor("op_1547"), val = tensor([1, 1, 16, 64])]; + tensor k_15_cast_fp16 = reshape(shape = var_1547, x = ko_7_cast_fp16)[name = tensor("k_15_cast_fp16")]; + tensor _inversed_1569_y_0_to_fp16 = const()[name = tensor("_inversed_1569_y_0_to_fp16"), val = tensor(0x1p-9)]; + tensor _inversed_1569_cast_fp16 = mul(x = ts_23_cast_fp16, y = _inversed_1569_y_0_to_fp16)[name = tensor("_inversed_1569_cast_fp16")]; + tensor var_1570_cast_fp16 = floor(x = _inversed_1569_cast_fp16)[name = tensor("op_1570_cast_fp16")]; + tensor var_1571_to_fp16 = const()[name = tensor("op_1571_to_fp16"), val = tensor(0x1p+9)]; + tensor var_1572_cast_fp16 = mul(x = var_1570_cast_fp16, y = var_1571_to_fp16)[name = tensor("op_1572_cast_fp16")]; + tensor write_indices_float_15_cast_fp16 = sub(x = ts_23_cast_fp16, y = var_1572_cast_fp16)[name = tensor("write_indices_float_15_cast_fp16")]; + tensor var_1579_dtype_0 = const()[name = tensor("op_1579_dtype_0"), val = tensor("int32")]; + tensor write_indices_7_reps_0 = const()[name = tensor("write_indices_7_reps_0"), val = tensor([1, 1, 16, 64])]; + tensor write_indices_float_15_cast_fp16_to_int32 = cast(dtype = var_1579_dtype_0, x = write_indices_float_15_cast_fp16)[name = tensor("cast_123")]; + tensor write_indices_7 = tile(reps = write_indices_7_reps_0, x = write_indices_float_15_cast_fp16_to_int32)[name = tensor("write_indices_7")]; + tensor var_1587_begin_0 = const()[name = tensor("op_1587_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_1587_end_0 = const()[name = tensor("op_1587_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor var_1587_end_mask_0 = const()[name = tensor("op_1587_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_1587_squeeze_mask_0 = const()[name = tensor("op_1587_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor cache3_to_fp16_dtype_0 = const()[name = tensor("cache3_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor cache3_to_fp16 = cast(dtype = cache3_to_fp16_dtype_0, x = cache3)[name = tensor("cast_122")]; + tensor var_1587_cast_fp16 = slice_by_index(begin = var_1587_begin_0, end = var_1587_end_0, end_mask = var_1587_end_mask_0, squeeze_mask = var_1587_squeeze_mask_0, x = cache3_to_fp16)[name = tensor("op_1587_cast_fp16")]; + tensor var_1589_axis_0 = const()[name = tensor("op_1589_axis_0"), val = tensor(1)]; + tensor var_1589_mode_0 = const()[name = tensor("op_1589_mode_0"), val = tensor("update")]; + tensor var_1589_validate_indices_0 = const()[name = tensor("op_1589_validate_indices_0"), val = tensor(false)]; + tensor var_1589_cast_fp16 = scatter_along_axis(axis = var_1589_axis_0, data = var_1587_cast_fp16, indices = write_indices_7, mode = var_1589_mode_0, updates = k_15_cast_fp16, validate_indices = var_1589_validate_indices_0)[name = tensor("op_1589_cast_fp16")]; + tensor concat_23 = const()[name = tensor("concat_23"), val = tensor([0, 0, 0, 0, 0])]; + tensor concat_24 = const()[name = tensor("concat_24"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_7_internal_tensor_assign_1_stride_0 = const()[name = tensor("new_cache_7_internal_tensor_assign_1_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_7_internal_tensor_assign_1_begin_mask_0 = const()[name = tensor("new_cache_7_internal_tensor_assign_1_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_7_internal_tensor_assign_1_end_mask_0 = const()[name = tensor("new_cache_7_internal_tensor_assign_1_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_7_internal_tensor_assign_1_squeeze_mask_0 = const()[name = tensor("new_cache_7_internal_tensor_assign_1_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_18 = const()[name = tensor("shape_18"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_6 = const()[name = tensor("reduce_prod_6"), val = tensor(1048576)]; + tensor range_1d_6_start_0 = const()[name = tensor("range_1d_6_start_0"), val = tensor(0)]; + tensor range_1d_6_step_0 = const()[name = tensor("range_1d_6_step_0"), val = tensor(1)]; + tensor range_1d_6 = range_1d(end = reduce_prod_6, start = range_1d_6_start_0, step = range_1d_6_step_0)[name = tensor("range_1d_6")]; + tensor reshape_30 = reshape(shape = shape_18, x = range_1d_6)[name = tensor("reshape_30")]; + tensor slice_by_index_6 = slice_by_index(begin = concat_23, begin_mask = new_cache_7_internal_tensor_assign_1_begin_mask_0, end = concat_24, end_mask = new_cache_7_internal_tensor_assign_1_end_mask_0, squeeze_mask = new_cache_7_internal_tensor_assign_1_squeeze_mask_0, stride = new_cache_7_internal_tensor_assign_1_stride_0, x = reshape_30)[name = tensor("slice_by_index_6")]; + tensor reshape_31_shape_0 = const()[name = tensor("reshape_31_shape_0"), val = tensor([-1])]; + tensor reshape_31 = reshape(shape = reshape_31_shape_0, x = slice_by_index_6)[name = tensor("reshape_31")]; + tensor reshape_32_shape_0 = const()[name = tensor("reshape_32_shape_0"), val = tensor([-1])]; + tensor reshape_32_cast_fp16 = reshape(shape = reshape_32_shape_0, x = var_1589_cast_fp16)[name = tensor("reshape_32_cast_fp16")]; + tensor reshape_33_shape_0 = const()[name = tensor("reshape_33_shape_0"), val = tensor([-1])]; + tensor reshape_33_cast_fp16 = reshape(shape = reshape_33_shape_0, x = cache3_to_fp16)[name = tensor("reshape_33_cast_fp16")]; + tensor scatter_6_mode_0 = const()[name = tensor("scatter_6_mode_0"), val = tensor("update")]; + tensor scatter_6_axis_0 = const()[name = tensor("scatter_6_axis_0"), val = tensor(0)]; + tensor scatter_6_validate_indices_0 = const()[name = tensor("scatter_6_validate_indices_0"), val = tensor(false)]; + tensor scatter_6_cast_fp16 = scatter(axis = scatter_6_axis_0, data = reshape_33_cast_fp16, indices = reshape_31, mode = scatter_6_mode_0, updates = reshape_32_cast_fp16, validate_indices = scatter_6_validate_indices_0)[name = tensor("scatter_6_cast_fp16")]; + tensor reshape_34_cast_fp16 = reshape(shape = shape_18, x = scatter_6_cast_fp16)[name = tensor("reshape_34_cast_fp16")]; + tensor var_1597_begin_0 = const()[name = tensor("op_1597_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor var_1597_end_0 = const()[name = tensor("op_1597_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor var_1597_end_mask_0 = const()[name = tensor("op_1597_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_1597_squeeze_mask_0 = const()[name = tensor("op_1597_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor var_1597_cast_fp16 = slice_by_index(begin = var_1597_begin_0, end = var_1597_end_0, end_mask = var_1597_end_mask_0, squeeze_mask = var_1597_squeeze_mask_0, x = reshape_34_cast_fp16)[name = tensor("op_1597_cast_fp16")]; + tensor var_1599_axis_0 = const()[name = tensor("op_1599_axis_0"), val = tensor(1)]; + tensor var_1599_mode_0 = const()[name = tensor("op_1599_mode_0"), val = tensor("update")]; + tensor var_1599_validate_indices_0 = const()[name = tensor("op_1599_validate_indices_0"), val = tensor(false)]; + tensor var_1599_cast_fp16 = scatter_along_axis(axis = var_1599_axis_0, data = var_1597_cast_fp16, indices = write_indices_7, mode = var_1599_mode_0, updates = v_7_cast_fp16, validate_indices = var_1599_validate_indices_0)[name = tensor("op_1599_cast_fp16")]; + tensor concat_25 = const()[name = tensor("concat_25"), val = tensor([1, 0, 0, 0, 0])]; + tensor concat_26 = const()[name = tensor("concat_26"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_7_internal_tensor_assign_2_stride_0 = const()[name = tensor("new_cache_7_internal_tensor_assign_2_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_7_internal_tensor_assign_2_begin_mask_0 = const()[name = tensor("new_cache_7_internal_tensor_assign_2_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_7_internal_tensor_assign_2_end_mask_0 = const()[name = tensor("new_cache_7_internal_tensor_assign_2_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_7_internal_tensor_assign_2_squeeze_mask_0 = const()[name = tensor("new_cache_7_internal_tensor_assign_2_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_19 = const()[name = tensor("shape_19"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_7 = const()[name = tensor("reduce_prod_7"), val = tensor(1048576)]; + tensor range_1d_7_start_0 = const()[name = tensor("range_1d_7_start_0"), val = tensor(0)]; + tensor range_1d_7_step_0 = const()[name = tensor("range_1d_7_step_0"), val = tensor(1)]; + tensor range_1d_7 = range_1d(end = reduce_prod_7, start = range_1d_7_start_0, step = range_1d_7_step_0)[name = tensor("range_1d_7")]; + tensor reshape_35 = reshape(shape = shape_19, x = range_1d_7)[name = tensor("reshape_35")]; + tensor slice_by_index_7 = slice_by_index(begin = concat_25, begin_mask = new_cache_7_internal_tensor_assign_2_begin_mask_0, end = concat_26, end_mask = new_cache_7_internal_tensor_assign_2_end_mask_0, squeeze_mask = new_cache_7_internal_tensor_assign_2_squeeze_mask_0, stride = new_cache_7_internal_tensor_assign_2_stride_0, x = reshape_35)[name = tensor("slice_by_index_7")]; + tensor reshape_36_shape_0 = const()[name = tensor("reshape_36_shape_0"), val = tensor([-1])]; + tensor reshape_36 = reshape(shape = reshape_36_shape_0, x = slice_by_index_7)[name = tensor("reshape_36")]; + tensor reshape_37_shape_0 = const()[name = tensor("reshape_37_shape_0"), val = tensor([-1])]; + tensor reshape_37_cast_fp16 = reshape(shape = reshape_37_shape_0, x = var_1599_cast_fp16)[name = tensor("reshape_37_cast_fp16")]; + tensor reshape_38_shape_0 = const()[name = tensor("reshape_38_shape_0"), val = tensor([-1])]; + tensor reshape_38_cast_fp16 = reshape(shape = reshape_38_shape_0, x = reshape_34_cast_fp16)[name = tensor("reshape_38_cast_fp16")]; + tensor scatter_7_mode_0 = const()[name = tensor("scatter_7_mode_0"), val = tensor("update")]; + tensor scatter_7_axis_0 = const()[name = tensor("scatter_7_axis_0"), val = tensor(0)]; + tensor scatter_7_validate_indices_0 = const()[name = tensor("scatter_7_validate_indices_0"), val = tensor(false)]; + tensor scatter_7_cast_fp16 = scatter(axis = scatter_7_axis_0, data = reshape_38_cast_fp16, indices = reshape_36, mode = scatter_7_mode_0, updates = reshape_37_cast_fp16, validate_indices = scatter_7_validate_indices_0)[name = tensor("scatter_7_cast_fp16")]; + tensor reshape_39_cast_fp16 = reshape(shape = shape_19, x = scatter_7_cast_fp16)[name = tensor("reshape_39_cast_fp16")]; + tensor reshape_39_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("reshape_39_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor keys_19_begin_0 = const()[name = tensor("keys_19_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor keys_19_end_0 = const()[name = tensor("keys_19_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor keys_19_end_mask_0 = const()[name = tensor("keys_19_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor keys_19_squeeze_mask_0 = const()[name = tensor("keys_19_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor keys_19_cast_fp16 = slice_by_index(begin = keys_19_begin_0, end = keys_19_end_0, end_mask = keys_19_end_mask_0, squeeze_mask = keys_19_squeeze_mask_0, x = reshape_39_cast_fp16)[name = tensor("keys_19_cast_fp16")]; + tensor values_19_begin_0 = const()[name = tensor("values_19_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor values_19_end_0 = const()[name = tensor("values_19_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor values_19_end_mask_0 = const()[name = tensor("values_19_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor values_19_squeeze_mask_0 = const()[name = tensor("values_19_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor values_19_cast_fp16 = slice_by_index(begin = values_19_begin_0, end = values_19_end_0, end_mask = values_19_end_mask_0, squeeze_mask = values_19_squeeze_mask_0, x = reshape_39_cast_fp16)[name = tensor("values_19_cast_fp16")]; + tensor var_1611_cast_fp16 = not_equal(x = keys_19_cast_fp16, y = keys_19_cast_fp16)[name = tensor("op_1611_cast_fp16")]; + tensor keys_21_cast_fp16 = select(a = var_360_to_fp16, b = keys_19_cast_fp16, cond = var_1611_cast_fp16)[name = tensor("keys_21_cast_fp16")]; + tensor var_1619_cast_fp16 = not_equal(x = values_19_cast_fp16, y = values_19_cast_fp16)[name = tensor("op_1619_cast_fp16")]; + tensor values_21_cast_fp16 = select(a = var_360_to_fp16, b = values_19_cast_fp16, cond = var_1619_cast_fp16)[name = tensor("values_21_cast_fp16")]; + tensor var_1643 = const()[name = tensor("op_1643"), val = tensor([0, 2, 1, 3])]; + tensor var_1656 = const()[name = tensor("op_1656"), val = tensor([1, 1, 1])]; + tensor var_1657_cast_fp16 = reshape(shape = var_1656, x = position3_to_fp16)[name = tensor("op_1657_cast_fp16")]; + tensor var_1674_to_fp16 = const()[name = tensor("op_1674_to_fp16"), val = tensor(0x1p+0)]; + tensor valid_len_7_cast_fp16 = add(x = var_1657_cast_fp16, y = var_1674_to_fp16)[name = tensor("valid_len_7_cast_fp16")]; + tensor valid_mask_7_cast_fp16 = less(x = k_positions_1_promoted_to_fp16, y = valid_len_7_cast_fp16)[name = tensor("valid_mask_7_cast_fp16")]; + tensor causal_mask_7_cast_fp16 = less_equal(x = k_positions_1_promoted_to_fp16, y = var_1657_cast_fp16)[name = tensor("causal_mask_7_cast_fp16")]; + tensor attn_mask_13 = logical_and(x = valid_mask_7_cast_fp16, y = causal_mask_7_cast_fp16)[name = tensor("attn_mask_13")]; + tensor attn_mask_15_axes_0 = const()[name = tensor("attn_mask_15_axes_0"), val = tensor([1])]; + tensor attn_mask_15 = expand_dims(axes = attn_mask_15_axes_0, x = attn_mask_13)[name = tensor("attn_mask_15")]; + tensor var_1692_transpose_x_0 = const()[name = tensor("op_1692_transpose_x_0"), val = tensor(false)]; + tensor var_1692_transpose_y_0 = const()[name = tensor("op_1692_transpose_y_0"), val = tensor(false)]; + tensor transpose_24_perm_0 = const()[name = tensor("transpose_24_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_25_perm_0 = const()[name = tensor("transpose_25_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_25 = transpose(perm = transpose_25_perm_0, x = keys_21_cast_fp16)[name = tensor("transpose_39")]; + tensor transpose_24 = transpose(perm = transpose_24_perm_0, x = q_21_cast_fp16)[name = tensor("transpose_40")]; + tensor var_1692_cast_fp16 = matmul(transpose_x = var_1692_transpose_x_0, transpose_y = var_1692_transpose_y_0, x = transpose_24, y = transpose_25)[name = tensor("op_1692_cast_fp16")]; + tensor var_1686_to_fp16 = const()[name = tensor("op_1686_to_fp16"), val = tensor([0x1p-3])]; + tensor attn_weights_19_cast_fp16 = mul(x = var_1692_cast_fp16, y = var_1686_to_fp16)[name = tensor("attn_weights_19_cast_fp16")]; + tensor var_1694 = logical_not(x = attn_mask_15)[name = tensor("op_1694")]; + tensor var_1695_to_fp16 = const()[name = tensor("op_1695_to_fp16"), val = tensor(-inf)]; + tensor attn_weights_21_cast_fp16 = select(a = var_1695_to_fp16, b = attn_weights_19_cast_fp16, cond = var_1694)[name = tensor("attn_weights_21_cast_fp16")]; + tensor var_1697 = const()[name = tensor("op_1697"), val = tensor(-1)]; + tensor attn_weights_23_cast_fp16 = softmax(axis = var_1697, x = attn_weights_21_cast_fp16)[name = tensor("attn_weights_23_cast_fp16")]; + tensor attn_output_7_transpose_x_0 = const()[name = tensor("attn_output_7_transpose_x_0"), val = tensor(false)]; + tensor attn_output_7_transpose_y_0 = const()[name = tensor("attn_output_7_transpose_y_0"), val = tensor(false)]; + tensor values_23_cast_fp16 = transpose(perm = var_1643, x = values_21_cast_fp16)[name = tensor("transpose_41")]; + tensor attn_output_7_cast_fp16 = matmul(transpose_x = attn_output_7_transpose_x_0, transpose_y = attn_output_7_transpose_y_0, x = attn_weights_23_cast_fp16, y = values_23_cast_fp16)[name = tensor("attn_output_7_cast_fp16")]; + tensor var_1705 = const()[name = tensor("op_1705"), val = tensor([0, 2, 1, 3])]; + tensor var_1708 = const()[name = tensor("op_1708"), val = tensor([1, 1, 1024])]; + tensor var_1706_cast_fp16 = transpose(perm = var_1705, x = attn_output_7_cast_fp16)[name = tensor("transpose_38")]; + tensor input_35_cast_fp16 = reshape(shape = var_1708, x = var_1706_cast_fp16)[name = tensor("input_35_cast_fp16")]; + tensor linear_14_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = attn3_out_proj_weight_cast_fp16, x = input_35_cast_fp16)[name = tensor("linear_14_cast_fp16")]; + tensor var_1714_to_fp16 = const()[name = tensor("op_1714_to_fp16"), val = tensor(0x1p+0)]; + tensor var_1715_cast_fp16 = add(x = position3_to_fp16, y = var_1714_to_fp16)[name = tensor("op_1715_cast_fp16")]; + tensor var_1715_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_1715_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor input_37_cast_fp16 = add(x = input_33_cast_fp16, y = linear_14_cast_fp16)[name = tensor("input_37_cast_fp16")]; + tensor input_39_axes_0 = const()[name = tensor("input_39_axes_0"), val = tensor([-1])]; + tensor norm3_2_weight_to_fp16 = const()[name = tensor("norm3_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(76781568)))]; + tensor norm3_2_bias_to_fp16 = const()[name = tensor("norm3_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(76783680)))]; + tensor var_1719_to_fp16 = const()[name = tensor("op_1719_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_39_cast_fp16 = layer_norm(axes = input_39_axes_0, beta = norm3_2_bias_to_fp16, epsilon = var_1719_to_fp16, gamma = norm3_2_weight_to_fp16, x = input_37_cast_fp16)[name = tensor("input_39_cast_fp16")]; + tensor linear_15_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = linear3_1_weight_cast_fp16, x = input_39_cast_fp16)[name = tensor("linear_15_cast_fp16")]; + tensor input_41_mode_0 = const()[name = tensor("input_41_mode_0"), val = tensor("EXACT")]; + tensor input_41_cast_fp16 = gelu(mode = input_41_mode_0, x = linear_15_cast_fp16)[name = tensor("input_41_cast_fp16")]; + tensor linear_16_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = linear3_2_weight_cast_fp16, x = input_41_cast_fp16)[name = tensor("linear_16_cast_fp16")]; + tensor input_43_cast_fp16 = add(x = input_37_cast_fp16, y = linear_16_cast_fp16)[name = tensor("input_43_cast_fp16")]; + tensor x_9_axes_0 = const()[name = tensor("x_9_axes_0"), val = tensor([-1])]; + tensor norm4_1_weight_to_fp16 = const()[name = tensor("norm4_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(76785792)))]; + tensor norm4_1_bias_to_fp16 = const()[name = tensor("norm4_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(76787904)))]; + tensor var_1736_to_fp16 = const()[name = tensor("op_1736_to_fp16"), val = tensor(0x1.5p-17)]; + tensor x_9_cast_fp16 = layer_norm(axes = x_9_axes_0, beta = norm4_1_bias_to_fp16, epsilon = var_1736_to_fp16, gamma = norm4_1_weight_to_fp16, x = input_43_cast_fp16)[name = tensor("x_9_cast_fp16")]; + tensor linear_17_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = attn4_in_proj_weight_cast_fp16, x = x_9_cast_fp16)[name = tensor("linear_17_cast_fp16")]; + tensor var_1772 = const()[name = tensor("op_1772"), val = tensor([1, 1, 3, 16, 64])]; + tensor qkv_9_cast_fp16 = reshape(shape = var_1772, x = linear_17_cast_fp16)[name = tensor("qkv_9_cast_fp16")]; + tensor q_25_begin_0 = const()[name = tensor("q_25_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor q_25_end_0 = const()[name = tensor("q_25_end_0"), val = tensor([1, 1, 1, 16, 64])]; + tensor q_25_end_mask_0 = const()[name = tensor("q_25_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor q_25_squeeze_mask_0 = const()[name = tensor("q_25_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor q_25_cast_fp16 = slice_by_index(begin = q_25_begin_0, end = q_25_end_0, end_mask = q_25_end_mask_0, squeeze_mask = q_25_squeeze_mask_0, x = qkv_9_cast_fp16)[name = tensor("q_25_cast_fp16")]; + tensor k_17_begin_0 = const()[name = tensor("k_17_begin_0"), val = tensor([0, 0, 1, 0, 0])]; + tensor k_17_end_0 = const()[name = tensor("k_17_end_0"), val = tensor([1, 1, 2, 16, 64])]; + tensor k_17_end_mask_0 = const()[name = tensor("k_17_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor k_17_squeeze_mask_0 = const()[name = tensor("k_17_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor k_17_cast_fp16 = slice_by_index(begin = k_17_begin_0, end = k_17_end_0, end_mask = k_17_end_mask_0, squeeze_mask = k_17_squeeze_mask_0, x = qkv_9_cast_fp16)[name = tensor("k_17_cast_fp16")]; + tensor v_9_begin_0 = const()[name = tensor("v_9_begin_0"), val = tensor([0, 0, 2, 0, 0])]; + tensor v_9_end_0 = const()[name = tensor("v_9_end_0"), val = tensor([1, 1, 3, 16, 64])]; + tensor v_9_end_mask_0 = const()[name = tensor("v_9_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor v_9_squeeze_mask_0 = const()[name = tensor("v_9_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor v_9_cast_fp16 = slice_by_index(begin = v_9_begin_0, end = v_9_end_0, end_mask = v_9_end_mask_0, squeeze_mask = v_9_squeeze_mask_0, x = qkv_9_cast_fp16)[name = tensor("v_9_cast_fp16")]; + tensor var_1876 = const()[name = tensor("op_1876"), val = tensor([1, 1, 1, 1])]; + tensor position4_to_fp16_dtype_0 = const()[name = tensor("position4_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor position4_to_fp16 = cast(dtype = position4_to_fp16_dtype_0, x = position4)[name = tensor("cast_119")]; + tensor ts_29_cast_fp16 = reshape(shape = var_1876, x = position4_to_fp16)[name = tensor("ts_29_cast_fp16")]; + tensor var_1880 = const()[name = tensor("op_1880"), val = tensor([1, 1, 16, 32, 2])]; + tensor q_complex_9_cast_fp16 = reshape(shape = var_1880, x = q_25_cast_fp16)[name = tensor("q_complex_9_cast_fp16")]; + tensor var_1884 = const()[name = tensor("op_1884"), val = tensor([1, 1, 16, 32, 2])]; + tensor k_complex_9_cast_fp16 = reshape(shape = var_1884, x = k_17_cast_fp16)[name = tensor("k_complex_9_cast_fp16")]; + tensor var_1888_begin_0 = const()[name = tensor("op_1888_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_1888_end_0 = const()[name = tensor("op_1888_end_0"), val = tensor([1, 1, 16, 32, 1])]; + tensor var_1888_end_mask_0 = const()[name = tensor("op_1888_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_1888_squeeze_mask_0 = const()[name = tensor("op_1888_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_1888_cast_fp16 = slice_by_index(begin = var_1888_begin_0, end = var_1888_end_0, end_mask = var_1888_end_mask_0, squeeze_mask = var_1888_squeeze_mask_0, x = q_complex_9_cast_fp16)[name = tensor("op_1888_cast_fp16")]; + tensor var_1896_begin_0 = const()[name = tensor("op_1896_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_1896_end_0 = const()[name = tensor("op_1896_end_0"), val = tensor([1, 1, 16, 32, 2])]; + tensor var_1896_end_mask_0 = const()[name = tensor("op_1896_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_1896_squeeze_mask_0 = const()[name = tensor("op_1896_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_1896_cast_fp16 = slice_by_index(begin = var_1896_begin_0, end = var_1896_end_0, end_mask = var_1896_end_mask_0, squeeze_mask = var_1896_squeeze_mask_0, x = q_complex_9_cast_fp16)[name = tensor("op_1896_cast_fp16")]; + tensor var_1904_begin_0 = const()[name = tensor("op_1904_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_1904_end_0 = const()[name = tensor("op_1904_end_0"), val = tensor([1, 1, 16, 32, 1])]; + tensor var_1904_end_mask_0 = const()[name = tensor("op_1904_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_1904_squeeze_mask_0 = const()[name = tensor("op_1904_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_1904_cast_fp16 = slice_by_index(begin = var_1904_begin_0, end = var_1904_end_0, end_mask = var_1904_end_mask_0, squeeze_mask = var_1904_squeeze_mask_0, x = k_complex_9_cast_fp16)[name = tensor("op_1904_cast_fp16")]; + tensor var_1912_begin_0 = const()[name = tensor("op_1912_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_1912_end_0 = const()[name = tensor("op_1912_end_0"), val = tensor([1, 1, 16, 32, 2])]; + tensor var_1912_end_mask_0 = const()[name = tensor("op_1912_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_1912_squeeze_mask_0 = const()[name = tensor("op_1912_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_1912_cast_fp16 = slice_by_index(begin = var_1912_begin_0, end = var_1912_end_0, end_mask = var_1912_end_mask_0, squeeze_mask = var_1912_squeeze_mask_0, x = k_complex_9_cast_fp16)[name = tensor("op_1912_cast_fp16")]; + tensor freqs_9_to_fp16 = const()[name = tensor("freqs_9_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(76790016)))]; + tensor var_1918_cast_fp16 = mul(x = freqs_9_to_fp16, y = ts_29_cast_fp16)[name = tensor("op_1918_cast_fp16")]; + tensor rotr_9_cast_fp16 = cos(x = var_1918_cast_fp16)[name = tensor("rotr_9_cast_fp16")]; + tensor roti_9_cast_fp16 = sin(x = var_1918_cast_fp16)[name = tensor("roti_9_cast_fp16")]; + tensor var_1922_cast_fp16 = mul(x = var_1888_cast_fp16, y = rotr_9_cast_fp16)[name = tensor("op_1922_cast_fp16")]; + tensor var_1923_cast_fp16 = mul(x = var_1896_cast_fp16, y = roti_9_cast_fp16)[name = tensor("op_1923_cast_fp16")]; + tensor qor_17_cast_fp16 = sub(x = var_1922_cast_fp16, y = var_1923_cast_fp16)[name = tensor("qor_17_cast_fp16")]; + tensor var_1926_cast_fp16 = mul(x = var_1888_cast_fp16, y = roti_9_cast_fp16)[name = tensor("op_1926_cast_fp16")]; + tensor var_1927_cast_fp16 = mul(x = var_1896_cast_fp16, y = rotr_9_cast_fp16)[name = tensor("op_1927_cast_fp16")]; + tensor qoi_17_cast_fp16 = add(x = var_1926_cast_fp16, y = var_1927_cast_fp16)[name = tensor("qoi_17_cast_fp16")]; + tensor var_1930_cast_fp16 = mul(x = var_1904_cast_fp16, y = rotr_9_cast_fp16)[name = tensor("op_1930_cast_fp16")]; + tensor var_1931_cast_fp16 = mul(x = var_1912_cast_fp16, y = roti_9_cast_fp16)[name = tensor("op_1931_cast_fp16")]; + tensor kor_17_cast_fp16 = sub(x = var_1930_cast_fp16, y = var_1931_cast_fp16)[name = tensor("kor_17_cast_fp16")]; + tensor var_1934_cast_fp16 = mul(x = var_1904_cast_fp16, y = roti_9_cast_fp16)[name = tensor("op_1934_cast_fp16")]; + tensor var_1935_cast_fp16 = mul(x = var_1912_cast_fp16, y = rotr_9_cast_fp16)[name = tensor("op_1935_cast_fp16")]; + tensor koi_17_cast_fp16 = add(x = var_1934_cast_fp16, y = var_1935_cast_fp16)[name = tensor("koi_17_cast_fp16")]; + tensor qo_9_axis_0 = const()[name = tensor("qo_9_axis_0"), val = tensor(-1)]; + tensor qo_9_cast_fp16 = stack(axis = qo_9_axis_0, values = (qor_17_cast_fp16, qoi_17_cast_fp16))[name = tensor("qo_9_cast_fp16")]; + tensor ko_9_axis_0 = const()[name = tensor("ko_9_axis_0"), val = tensor(-1)]; + tensor ko_9_cast_fp16 = stack(axis = ko_9_axis_0, values = (kor_17_cast_fp16, koi_17_cast_fp16))[name = tensor("ko_9_cast_fp16")]; + tensor var_1964 = const()[name = tensor("op_1964"), val = tensor([1, 1, 16, 64])]; + tensor q_27_cast_fp16 = reshape(shape = var_1964, x = qo_9_cast_fp16)[name = tensor("q_27_cast_fp16")]; + tensor var_1966 = const()[name = tensor("op_1966"), val = tensor([1, 1, 16, 64])]; + tensor k_19_cast_fp16 = reshape(shape = var_1966, x = ko_9_cast_fp16)[name = tensor("k_19_cast_fp16")]; + tensor _inversed_1988_y_0_to_fp16 = const()[name = tensor("_inversed_1988_y_0_to_fp16"), val = tensor(0x1p-9)]; + tensor _inversed_1988_cast_fp16 = mul(x = ts_29_cast_fp16, y = _inversed_1988_y_0_to_fp16)[name = tensor("_inversed_1988_cast_fp16")]; + tensor var_1989_cast_fp16 = floor(x = _inversed_1988_cast_fp16)[name = tensor("op_1989_cast_fp16")]; + tensor var_1990_to_fp16 = const()[name = tensor("op_1990_to_fp16"), val = tensor(0x1p+9)]; + tensor var_1991_cast_fp16 = mul(x = var_1989_cast_fp16, y = var_1990_to_fp16)[name = tensor("op_1991_cast_fp16")]; + tensor write_indices_float_19_cast_fp16 = sub(x = ts_29_cast_fp16, y = var_1991_cast_fp16)[name = tensor("write_indices_float_19_cast_fp16")]; + tensor var_1998_dtype_0 = const()[name = tensor("op_1998_dtype_0"), val = tensor("int32")]; + tensor write_indices_9_reps_0 = const()[name = tensor("write_indices_9_reps_0"), val = tensor([1, 1, 16, 64])]; + tensor write_indices_float_19_cast_fp16_to_int32 = cast(dtype = var_1998_dtype_0, x = write_indices_float_19_cast_fp16)[name = tensor("cast_118")]; + tensor write_indices_9 = tile(reps = write_indices_9_reps_0, x = write_indices_float_19_cast_fp16_to_int32)[name = tensor("write_indices_9")]; + tensor var_2006_begin_0 = const()[name = tensor("op_2006_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_2006_end_0 = const()[name = tensor("op_2006_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor var_2006_end_mask_0 = const()[name = tensor("op_2006_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_2006_squeeze_mask_0 = const()[name = tensor("op_2006_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor cache4_to_fp16_dtype_0 = const()[name = tensor("cache4_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor cache4_to_fp16 = cast(dtype = cache4_to_fp16_dtype_0, x = cache4)[name = tensor("cast_117")]; + tensor var_2006_cast_fp16 = slice_by_index(begin = var_2006_begin_0, end = var_2006_end_0, end_mask = var_2006_end_mask_0, squeeze_mask = var_2006_squeeze_mask_0, x = cache4_to_fp16)[name = tensor("op_2006_cast_fp16")]; + tensor var_2008_axis_0 = const()[name = tensor("op_2008_axis_0"), val = tensor(1)]; + tensor var_2008_mode_0 = const()[name = tensor("op_2008_mode_0"), val = tensor("update")]; + tensor var_2008_validate_indices_0 = const()[name = tensor("op_2008_validate_indices_0"), val = tensor(false)]; + tensor var_2008_cast_fp16 = scatter_along_axis(axis = var_2008_axis_0, data = var_2006_cast_fp16, indices = write_indices_9, mode = var_2008_mode_0, updates = k_19_cast_fp16, validate_indices = var_2008_validate_indices_0)[name = tensor("op_2008_cast_fp16")]; + tensor concat_30 = const()[name = tensor("concat_30"), val = tensor([0, 0, 0, 0, 0])]; + tensor concat_31 = const()[name = tensor("concat_31"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_9_internal_tensor_assign_1_stride_0 = const()[name = tensor("new_cache_9_internal_tensor_assign_1_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_9_internal_tensor_assign_1_begin_mask_0 = const()[name = tensor("new_cache_9_internal_tensor_assign_1_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_9_internal_tensor_assign_1_end_mask_0 = const()[name = tensor("new_cache_9_internal_tensor_assign_1_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_9_internal_tensor_assign_1_squeeze_mask_0 = const()[name = tensor("new_cache_9_internal_tensor_assign_1_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_20 = const()[name = tensor("shape_20"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_8 = const()[name = tensor("reduce_prod_8"), val = tensor(1048576)]; + tensor range_1d_8_start_0 = const()[name = tensor("range_1d_8_start_0"), val = tensor(0)]; + tensor range_1d_8_step_0 = const()[name = tensor("range_1d_8_step_0"), val = tensor(1)]; + tensor range_1d_8 = range_1d(end = reduce_prod_8, start = range_1d_8_start_0, step = range_1d_8_step_0)[name = tensor("range_1d_8")]; + tensor reshape_40 = reshape(shape = shape_20, x = range_1d_8)[name = tensor("reshape_40")]; + tensor slice_by_index_8 = slice_by_index(begin = concat_30, begin_mask = new_cache_9_internal_tensor_assign_1_begin_mask_0, end = concat_31, end_mask = new_cache_9_internal_tensor_assign_1_end_mask_0, squeeze_mask = new_cache_9_internal_tensor_assign_1_squeeze_mask_0, stride = new_cache_9_internal_tensor_assign_1_stride_0, x = reshape_40)[name = tensor("slice_by_index_8")]; + tensor reshape_41_shape_0 = const()[name = tensor("reshape_41_shape_0"), val = tensor([-1])]; + tensor reshape_41 = reshape(shape = reshape_41_shape_0, x = slice_by_index_8)[name = tensor("reshape_41")]; + tensor reshape_42_shape_0 = const()[name = tensor("reshape_42_shape_0"), val = tensor([-1])]; + tensor reshape_42_cast_fp16 = reshape(shape = reshape_42_shape_0, x = var_2008_cast_fp16)[name = tensor("reshape_42_cast_fp16")]; + tensor reshape_43_shape_0 = const()[name = tensor("reshape_43_shape_0"), val = tensor([-1])]; + tensor reshape_43_cast_fp16 = reshape(shape = reshape_43_shape_0, x = cache4_to_fp16)[name = tensor("reshape_43_cast_fp16")]; + tensor scatter_8_mode_0 = const()[name = tensor("scatter_8_mode_0"), val = tensor("update")]; + tensor scatter_8_axis_0 = const()[name = tensor("scatter_8_axis_0"), val = tensor(0)]; + tensor scatter_8_validate_indices_0 = const()[name = tensor("scatter_8_validate_indices_0"), val = tensor(false)]; + tensor scatter_8_cast_fp16 = scatter(axis = scatter_8_axis_0, data = reshape_43_cast_fp16, indices = reshape_41, mode = scatter_8_mode_0, updates = reshape_42_cast_fp16, validate_indices = scatter_8_validate_indices_0)[name = tensor("scatter_8_cast_fp16")]; + tensor reshape_44_cast_fp16 = reshape(shape = shape_20, x = scatter_8_cast_fp16)[name = tensor("reshape_44_cast_fp16")]; + tensor var_2016_begin_0 = const()[name = tensor("op_2016_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor var_2016_end_0 = const()[name = tensor("op_2016_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor var_2016_end_mask_0 = const()[name = tensor("op_2016_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_2016_squeeze_mask_0 = const()[name = tensor("op_2016_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor var_2016_cast_fp16 = slice_by_index(begin = var_2016_begin_0, end = var_2016_end_0, end_mask = var_2016_end_mask_0, squeeze_mask = var_2016_squeeze_mask_0, x = reshape_44_cast_fp16)[name = tensor("op_2016_cast_fp16")]; + tensor var_2018_axis_0 = const()[name = tensor("op_2018_axis_0"), val = tensor(1)]; + tensor var_2018_mode_0 = const()[name = tensor("op_2018_mode_0"), val = tensor("update")]; + tensor var_2018_validate_indices_0 = const()[name = tensor("op_2018_validate_indices_0"), val = tensor(false)]; + tensor var_2018_cast_fp16 = scatter_along_axis(axis = var_2018_axis_0, data = var_2016_cast_fp16, indices = write_indices_9, mode = var_2018_mode_0, updates = v_9_cast_fp16, validate_indices = var_2018_validate_indices_0)[name = tensor("op_2018_cast_fp16")]; + tensor concat_32 = const()[name = tensor("concat_32"), val = tensor([1, 0, 0, 0, 0])]; + tensor concat_33 = const()[name = tensor("concat_33"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_9_internal_tensor_assign_2_stride_0 = const()[name = tensor("new_cache_9_internal_tensor_assign_2_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_9_internal_tensor_assign_2_begin_mask_0 = const()[name = tensor("new_cache_9_internal_tensor_assign_2_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_9_internal_tensor_assign_2_end_mask_0 = const()[name = tensor("new_cache_9_internal_tensor_assign_2_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_9_internal_tensor_assign_2_squeeze_mask_0 = const()[name = tensor("new_cache_9_internal_tensor_assign_2_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_21 = const()[name = tensor("shape_21"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_9 = const()[name = tensor("reduce_prod_9"), val = tensor(1048576)]; + tensor range_1d_9_start_0 = const()[name = tensor("range_1d_9_start_0"), val = tensor(0)]; + tensor range_1d_9_step_0 = const()[name = tensor("range_1d_9_step_0"), val = tensor(1)]; + tensor range_1d_9 = range_1d(end = reduce_prod_9, start = range_1d_9_start_0, step = range_1d_9_step_0)[name = tensor("range_1d_9")]; + tensor reshape_45 = reshape(shape = shape_21, x = range_1d_9)[name = tensor("reshape_45")]; + tensor slice_by_index_9 = slice_by_index(begin = concat_32, begin_mask = new_cache_9_internal_tensor_assign_2_begin_mask_0, end = concat_33, end_mask = new_cache_9_internal_tensor_assign_2_end_mask_0, squeeze_mask = new_cache_9_internal_tensor_assign_2_squeeze_mask_0, stride = new_cache_9_internal_tensor_assign_2_stride_0, x = reshape_45)[name = tensor("slice_by_index_9")]; + tensor reshape_46_shape_0 = const()[name = tensor("reshape_46_shape_0"), val = tensor([-1])]; + tensor reshape_46 = reshape(shape = reshape_46_shape_0, x = slice_by_index_9)[name = tensor("reshape_46")]; + tensor reshape_47_shape_0 = const()[name = tensor("reshape_47_shape_0"), val = tensor([-1])]; + tensor reshape_47_cast_fp16 = reshape(shape = reshape_47_shape_0, x = var_2018_cast_fp16)[name = tensor("reshape_47_cast_fp16")]; + tensor reshape_48_shape_0 = const()[name = tensor("reshape_48_shape_0"), val = tensor([-1])]; + tensor reshape_48_cast_fp16 = reshape(shape = reshape_48_shape_0, x = reshape_44_cast_fp16)[name = tensor("reshape_48_cast_fp16")]; + tensor scatter_9_mode_0 = const()[name = tensor("scatter_9_mode_0"), val = tensor("update")]; + tensor scatter_9_axis_0 = const()[name = tensor("scatter_9_axis_0"), val = tensor(0)]; + tensor scatter_9_validate_indices_0 = const()[name = tensor("scatter_9_validate_indices_0"), val = tensor(false)]; + tensor scatter_9_cast_fp16 = scatter(axis = scatter_9_axis_0, data = reshape_48_cast_fp16, indices = reshape_46, mode = scatter_9_mode_0, updates = reshape_47_cast_fp16, validate_indices = scatter_9_validate_indices_0)[name = tensor("scatter_9_cast_fp16")]; + tensor reshape_49_cast_fp16 = reshape(shape = shape_21, x = scatter_9_cast_fp16)[name = tensor("reshape_49_cast_fp16")]; + tensor reshape_49_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("reshape_49_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor keys_25_begin_0 = const()[name = tensor("keys_25_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor keys_25_end_0 = const()[name = tensor("keys_25_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor keys_25_end_mask_0 = const()[name = tensor("keys_25_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor keys_25_squeeze_mask_0 = const()[name = tensor("keys_25_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor keys_25_cast_fp16 = slice_by_index(begin = keys_25_begin_0, end = keys_25_end_0, end_mask = keys_25_end_mask_0, squeeze_mask = keys_25_squeeze_mask_0, x = reshape_49_cast_fp16)[name = tensor("keys_25_cast_fp16")]; + tensor values_25_begin_0 = const()[name = tensor("values_25_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor values_25_end_0 = const()[name = tensor("values_25_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor values_25_end_mask_0 = const()[name = tensor("values_25_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor values_25_squeeze_mask_0 = const()[name = tensor("values_25_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor values_25_cast_fp16 = slice_by_index(begin = values_25_begin_0, end = values_25_end_0, end_mask = values_25_end_mask_0, squeeze_mask = values_25_squeeze_mask_0, x = reshape_49_cast_fp16)[name = tensor("values_25_cast_fp16")]; + tensor var_2030_cast_fp16 = not_equal(x = keys_25_cast_fp16, y = keys_25_cast_fp16)[name = tensor("op_2030_cast_fp16")]; + tensor keys_27_cast_fp16 = select(a = var_360_to_fp16, b = keys_25_cast_fp16, cond = var_2030_cast_fp16)[name = tensor("keys_27_cast_fp16")]; + tensor var_2038_cast_fp16 = not_equal(x = values_25_cast_fp16, y = values_25_cast_fp16)[name = tensor("op_2038_cast_fp16")]; + tensor values_27_cast_fp16 = select(a = var_360_to_fp16, b = values_25_cast_fp16, cond = var_2038_cast_fp16)[name = tensor("values_27_cast_fp16")]; + tensor var_2062 = const()[name = tensor("op_2062"), val = tensor([0, 2, 1, 3])]; + tensor var_2075 = const()[name = tensor("op_2075"), val = tensor([1, 1, 1])]; + tensor var_2076_cast_fp16 = reshape(shape = var_2075, x = position4_to_fp16)[name = tensor("op_2076_cast_fp16")]; + tensor var_2093_to_fp16 = const()[name = tensor("op_2093_to_fp16"), val = tensor(0x1p+0)]; + tensor valid_len_9_cast_fp16 = add(x = var_2076_cast_fp16, y = var_2093_to_fp16)[name = tensor("valid_len_9_cast_fp16")]; + tensor valid_mask_9_cast_fp16 = less(x = k_positions_1_promoted_to_fp16, y = valid_len_9_cast_fp16)[name = tensor("valid_mask_9_cast_fp16")]; + tensor causal_mask_9_cast_fp16 = less_equal(x = k_positions_1_promoted_to_fp16, y = var_2076_cast_fp16)[name = tensor("causal_mask_9_cast_fp16")]; + tensor attn_mask_17 = logical_and(x = valid_mask_9_cast_fp16, y = causal_mask_9_cast_fp16)[name = tensor("attn_mask_17")]; + tensor attn_mask_19_axes_0 = const()[name = tensor("attn_mask_19_axes_0"), val = tensor([1])]; + tensor attn_mask_19 = expand_dims(axes = attn_mask_19_axes_0, x = attn_mask_17)[name = tensor("attn_mask_19")]; + tensor var_2111_transpose_x_0 = const()[name = tensor("op_2111_transpose_x_0"), val = tensor(false)]; + tensor var_2111_transpose_y_0 = const()[name = tensor("op_2111_transpose_y_0"), val = tensor(false)]; + tensor transpose_26_perm_0 = const()[name = tensor("transpose_26_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_27_perm_0 = const()[name = tensor("transpose_27_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_27 = transpose(perm = transpose_27_perm_0, x = keys_27_cast_fp16)[name = tensor("transpose_35")]; + tensor transpose_26 = transpose(perm = transpose_26_perm_0, x = q_27_cast_fp16)[name = tensor("transpose_36")]; + tensor var_2111_cast_fp16 = matmul(transpose_x = var_2111_transpose_x_0, transpose_y = var_2111_transpose_y_0, x = transpose_26, y = transpose_27)[name = tensor("op_2111_cast_fp16")]; + tensor var_2105_to_fp16 = const()[name = tensor("op_2105_to_fp16"), val = tensor([0x1p-3])]; + tensor attn_weights_25_cast_fp16 = mul(x = var_2111_cast_fp16, y = var_2105_to_fp16)[name = tensor("attn_weights_25_cast_fp16")]; + tensor var_2113 = logical_not(x = attn_mask_19)[name = tensor("op_2113")]; + tensor var_2114_to_fp16 = const()[name = tensor("op_2114_to_fp16"), val = tensor(-inf)]; + tensor attn_weights_27_cast_fp16 = select(a = var_2114_to_fp16, b = attn_weights_25_cast_fp16, cond = var_2113)[name = tensor("attn_weights_27_cast_fp16")]; + tensor var_2116 = const()[name = tensor("op_2116"), val = tensor(-1)]; + tensor attn_weights_29_cast_fp16 = softmax(axis = var_2116, x = attn_weights_27_cast_fp16)[name = tensor("attn_weights_29_cast_fp16")]; + tensor attn_output_9_transpose_x_0 = const()[name = tensor("attn_output_9_transpose_x_0"), val = tensor(false)]; + tensor attn_output_9_transpose_y_0 = const()[name = tensor("attn_output_9_transpose_y_0"), val = tensor(false)]; + tensor values_29_cast_fp16 = transpose(perm = var_2062, x = values_27_cast_fp16)[name = tensor("transpose_37")]; + tensor attn_output_9_cast_fp16 = matmul(transpose_x = attn_output_9_transpose_x_0, transpose_y = attn_output_9_transpose_y_0, x = attn_weights_29_cast_fp16, y = values_29_cast_fp16)[name = tensor("attn_output_9_cast_fp16")]; + tensor var_2124 = const()[name = tensor("op_2124"), val = tensor([0, 2, 1, 3])]; + tensor var_2127 = const()[name = tensor("op_2127"), val = tensor([1, 1, 1024])]; + tensor var_2125_cast_fp16 = transpose(perm = var_2124, x = attn_output_9_cast_fp16)[name = tensor("transpose_34")]; + tensor input_45_cast_fp16 = reshape(shape = var_2127, x = var_2125_cast_fp16)[name = tensor("input_45_cast_fp16")]; + tensor linear_18_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = attn4_out_proj_weight_cast_fp16, x = input_45_cast_fp16)[name = tensor("linear_18_cast_fp16")]; + tensor var_2133_to_fp16 = const()[name = tensor("op_2133_to_fp16"), val = tensor(0x1p+0)]; + tensor var_2134_cast_fp16 = add(x = position4_to_fp16, y = var_2133_to_fp16)[name = tensor("op_2134_cast_fp16")]; + tensor var_2134_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_2134_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor input_47_cast_fp16 = add(x = input_43_cast_fp16, y = linear_18_cast_fp16)[name = tensor("input_47_cast_fp16")]; + tensor input_49_axes_0 = const()[name = tensor("input_49_axes_0"), val = tensor([-1])]; + tensor norm4_2_weight_to_fp16 = const()[name = tensor("norm4_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(76790144)))]; + tensor norm4_2_bias_to_fp16 = const()[name = tensor("norm4_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(76792256)))]; + tensor var_2138_to_fp16 = const()[name = tensor("op_2138_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_49_cast_fp16 = layer_norm(axes = input_49_axes_0, beta = norm4_2_bias_to_fp16, epsilon = var_2138_to_fp16, gamma = norm4_2_weight_to_fp16, x = input_47_cast_fp16)[name = tensor("input_49_cast_fp16")]; + tensor linear_19_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = linear4_1_weight_cast_fp16, x = input_49_cast_fp16)[name = tensor("linear_19_cast_fp16")]; + tensor input_51_mode_0 = const()[name = tensor("input_51_mode_0"), val = tensor("EXACT")]; + tensor input_51_cast_fp16 = gelu(mode = input_51_mode_0, x = linear_19_cast_fp16)[name = tensor("input_51_cast_fp16")]; + tensor linear_20_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = linear4_2_weight_cast_fp16, x = input_51_cast_fp16)[name = tensor("linear_20_cast_fp16")]; + tensor input_53_cast_fp16 = add(x = input_47_cast_fp16, y = linear_20_cast_fp16)[name = tensor("input_53_cast_fp16")]; + tensor x_axes_0 = const()[name = tensor("x_axes_0"), val = tensor([-1])]; + tensor norm5_1_weight_to_fp16 = const()[name = tensor("norm5_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(76794368)))]; + tensor norm5_1_bias_to_fp16 = const()[name = tensor("norm5_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(76796480)))]; + tensor var_2155_to_fp16 = const()[name = tensor("op_2155_to_fp16"), val = tensor(0x1.5p-17)]; + tensor x_cast_fp16 = layer_norm(axes = x_axes_0, beta = norm5_1_bias_to_fp16, epsilon = var_2155_to_fp16, gamma = norm5_1_weight_to_fp16, x = input_53_cast_fp16)[name = tensor("x_cast_fp16")]; + tensor linear_21_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = attn5_in_proj_weight_cast_fp16, x = x_cast_fp16)[name = tensor("linear_21_cast_fp16")]; + tensor var_2191 = const()[name = tensor("op_2191"), val = tensor([1, 1, 3, 16, 64])]; + tensor qkv_cast_fp16 = reshape(shape = var_2191, x = linear_21_cast_fp16)[name = tensor("qkv_cast_fp16")]; + tensor q_31_begin_0 = const()[name = tensor("q_31_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor q_31_end_0 = const()[name = tensor("q_31_end_0"), val = tensor([1, 1, 1, 16, 64])]; + tensor q_31_end_mask_0 = const()[name = tensor("q_31_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor q_31_squeeze_mask_0 = const()[name = tensor("q_31_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor q_31_cast_fp16 = slice_by_index(begin = q_31_begin_0, end = q_31_end_0, end_mask = q_31_end_mask_0, squeeze_mask = q_31_squeeze_mask_0, x = qkv_cast_fp16)[name = tensor("q_31_cast_fp16")]; + tensor k_21_begin_0 = const()[name = tensor("k_21_begin_0"), val = tensor([0, 0, 1, 0, 0])]; + tensor k_21_end_0 = const()[name = tensor("k_21_end_0"), val = tensor([1, 1, 2, 16, 64])]; + tensor k_21_end_mask_0 = const()[name = tensor("k_21_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor k_21_squeeze_mask_0 = const()[name = tensor("k_21_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor k_21_cast_fp16 = slice_by_index(begin = k_21_begin_0, end = k_21_end_0, end_mask = k_21_end_mask_0, squeeze_mask = k_21_squeeze_mask_0, x = qkv_cast_fp16)[name = tensor("k_21_cast_fp16")]; + tensor v_begin_0 = const()[name = tensor("v_begin_0"), val = tensor([0, 0, 2, 0, 0])]; + tensor v_end_0 = const()[name = tensor("v_end_0"), val = tensor([1, 1, 3, 16, 64])]; + tensor v_end_mask_0 = const()[name = tensor("v_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor v_squeeze_mask_0 = const()[name = tensor("v_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor v_cast_fp16 = slice_by_index(begin = v_begin_0, end = v_end_0, end_mask = v_end_mask_0, squeeze_mask = v_squeeze_mask_0, x = qkv_cast_fp16)[name = tensor("v_cast_fp16")]; + tensor var_2295 = const()[name = tensor("op_2295"), val = tensor([1, 1, 1, 1])]; + tensor position5_to_fp16_dtype_0 = const()[name = tensor("position5_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor position5_to_fp16 = cast(dtype = position5_to_fp16_dtype_0, x = position5)[name = tensor("cast_114")]; + tensor ts_cast_fp16 = reshape(shape = var_2295, x = position5_to_fp16)[name = tensor("ts_cast_fp16")]; + tensor var_2299 = const()[name = tensor("op_2299"), val = tensor([1, 1, 16, 32, 2])]; + tensor q_complex_cast_fp16 = reshape(shape = var_2299, x = q_31_cast_fp16)[name = tensor("q_complex_cast_fp16")]; + tensor var_2303 = const()[name = tensor("op_2303"), val = tensor([1, 1, 16, 32, 2])]; + tensor k_complex_cast_fp16 = reshape(shape = var_2303, x = k_21_cast_fp16)[name = tensor("k_complex_cast_fp16")]; + tensor var_2307_begin_0 = const()[name = tensor("op_2307_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_2307_end_0 = const()[name = tensor("op_2307_end_0"), val = tensor([1, 1, 16, 32, 1])]; + tensor var_2307_end_mask_0 = const()[name = tensor("op_2307_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_2307_squeeze_mask_0 = const()[name = tensor("op_2307_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_2307_cast_fp16 = slice_by_index(begin = var_2307_begin_0, end = var_2307_end_0, end_mask = var_2307_end_mask_0, squeeze_mask = var_2307_squeeze_mask_0, x = q_complex_cast_fp16)[name = tensor("op_2307_cast_fp16")]; + tensor var_2315_begin_0 = const()[name = tensor("op_2315_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_2315_end_0 = const()[name = tensor("op_2315_end_0"), val = tensor([1, 1, 16, 32, 2])]; + tensor var_2315_end_mask_0 = const()[name = tensor("op_2315_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_2315_squeeze_mask_0 = const()[name = tensor("op_2315_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_2315_cast_fp16 = slice_by_index(begin = var_2315_begin_0, end = var_2315_end_0, end_mask = var_2315_end_mask_0, squeeze_mask = var_2315_squeeze_mask_0, x = q_complex_cast_fp16)[name = tensor("op_2315_cast_fp16")]; + tensor var_2323_begin_0 = const()[name = tensor("op_2323_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_2323_end_0 = const()[name = tensor("op_2323_end_0"), val = tensor([1, 1, 16, 32, 1])]; + tensor var_2323_end_mask_0 = const()[name = tensor("op_2323_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_2323_squeeze_mask_0 = const()[name = tensor("op_2323_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_2323_cast_fp16 = slice_by_index(begin = var_2323_begin_0, end = var_2323_end_0, end_mask = var_2323_end_mask_0, squeeze_mask = var_2323_squeeze_mask_0, x = k_complex_cast_fp16)[name = tensor("op_2323_cast_fp16")]; + tensor var_2331_begin_0 = const()[name = tensor("op_2331_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_2331_end_0 = const()[name = tensor("op_2331_end_0"), val = tensor([1, 1, 16, 32, 2])]; + tensor var_2331_end_mask_0 = const()[name = tensor("op_2331_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_2331_squeeze_mask_0 = const()[name = tensor("op_2331_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_2331_cast_fp16 = slice_by_index(begin = var_2331_begin_0, end = var_2331_end_0, end_mask = var_2331_end_mask_0, squeeze_mask = var_2331_squeeze_mask_0, x = k_complex_cast_fp16)[name = tensor("op_2331_cast_fp16")]; + tensor freqs_to_fp16 = const()[name = tensor("freqs_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(76798592)))]; + tensor var_2337_cast_fp16 = mul(x = freqs_to_fp16, y = ts_cast_fp16)[name = tensor("op_2337_cast_fp16")]; + tensor rotr_cast_fp16 = cos(x = var_2337_cast_fp16)[name = tensor("rotr_cast_fp16")]; + tensor roti_cast_fp16 = sin(x = var_2337_cast_fp16)[name = tensor("roti_cast_fp16")]; + tensor var_2341_cast_fp16 = mul(x = var_2307_cast_fp16, y = rotr_cast_fp16)[name = tensor("op_2341_cast_fp16")]; + tensor var_2342_cast_fp16 = mul(x = var_2315_cast_fp16, y = roti_cast_fp16)[name = tensor("op_2342_cast_fp16")]; + tensor qor_21_cast_fp16 = sub(x = var_2341_cast_fp16, y = var_2342_cast_fp16)[name = tensor("qor_21_cast_fp16")]; + tensor var_2345_cast_fp16 = mul(x = var_2307_cast_fp16, y = roti_cast_fp16)[name = tensor("op_2345_cast_fp16")]; + tensor var_2346_cast_fp16 = mul(x = var_2315_cast_fp16, y = rotr_cast_fp16)[name = tensor("op_2346_cast_fp16")]; + tensor qoi_21_cast_fp16 = add(x = var_2345_cast_fp16, y = var_2346_cast_fp16)[name = tensor("qoi_21_cast_fp16")]; + tensor var_2349_cast_fp16 = mul(x = var_2323_cast_fp16, y = rotr_cast_fp16)[name = tensor("op_2349_cast_fp16")]; + tensor var_2350_cast_fp16 = mul(x = var_2331_cast_fp16, y = roti_cast_fp16)[name = tensor("op_2350_cast_fp16")]; + tensor kor_21_cast_fp16 = sub(x = var_2349_cast_fp16, y = var_2350_cast_fp16)[name = tensor("kor_21_cast_fp16")]; + tensor var_2353_cast_fp16 = mul(x = var_2323_cast_fp16, y = roti_cast_fp16)[name = tensor("op_2353_cast_fp16")]; + tensor var_2354_cast_fp16 = mul(x = var_2331_cast_fp16, y = rotr_cast_fp16)[name = tensor("op_2354_cast_fp16")]; + tensor koi_21_cast_fp16 = add(x = var_2353_cast_fp16, y = var_2354_cast_fp16)[name = tensor("koi_21_cast_fp16")]; + tensor qo_axis_0 = const()[name = tensor("qo_axis_0"), val = tensor(-1)]; + tensor qo_cast_fp16 = stack(axis = qo_axis_0, values = (qor_21_cast_fp16, qoi_21_cast_fp16))[name = tensor("qo_cast_fp16")]; + tensor ko_axis_0 = const()[name = tensor("ko_axis_0"), val = tensor(-1)]; + tensor ko_cast_fp16 = stack(axis = ko_axis_0, values = (kor_21_cast_fp16, koi_21_cast_fp16))[name = tensor("ko_cast_fp16")]; + tensor var_2383 = const()[name = tensor("op_2383"), val = tensor([1, 1, 16, 64])]; + tensor q_33_cast_fp16 = reshape(shape = var_2383, x = qo_cast_fp16)[name = tensor("q_33_cast_fp16")]; + tensor var_2385 = const()[name = tensor("op_2385"), val = tensor([1, 1, 16, 64])]; + tensor k_cast_fp16 = reshape(shape = var_2385, x = ko_cast_fp16)[name = tensor("k_cast_fp16")]; + tensor _inversed_2407_y_0_to_fp16 = const()[name = tensor("_inversed_2407_y_0_to_fp16"), val = tensor(0x1p-9)]; + tensor _inversed_2407_cast_fp16 = mul(x = ts_cast_fp16, y = _inversed_2407_y_0_to_fp16)[name = tensor("_inversed_2407_cast_fp16")]; + tensor var_2408_cast_fp16 = floor(x = _inversed_2407_cast_fp16)[name = tensor("op_2408_cast_fp16")]; + tensor var_2409_to_fp16 = const()[name = tensor("op_2409_to_fp16"), val = tensor(0x1p+9)]; + tensor var_2410_cast_fp16 = mul(x = var_2408_cast_fp16, y = var_2409_to_fp16)[name = tensor("op_2410_cast_fp16")]; + tensor write_indices_float_cast_fp16 = sub(x = ts_cast_fp16, y = var_2410_cast_fp16)[name = tensor("write_indices_float_cast_fp16")]; + tensor var_2417_dtype_0 = const()[name = tensor("op_2417_dtype_0"), val = tensor("int32")]; + tensor write_indices_reps_0 = const()[name = tensor("write_indices_reps_0"), val = tensor([1, 1, 16, 64])]; + tensor write_indices_float_cast_fp16_to_int32 = cast(dtype = var_2417_dtype_0, x = write_indices_float_cast_fp16)[name = tensor("cast_113")]; + tensor write_indices = tile(reps = write_indices_reps_0, x = write_indices_float_cast_fp16_to_int32)[name = tensor("write_indices")]; + tensor var_2425_begin_0 = const()[name = tensor("op_2425_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_2425_end_0 = const()[name = tensor("op_2425_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor var_2425_end_mask_0 = const()[name = tensor("op_2425_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_2425_squeeze_mask_0 = const()[name = tensor("op_2425_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor cache5_to_fp16_dtype_0 = const()[name = tensor("cache5_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor cache5_to_fp16 = cast(dtype = cache5_to_fp16_dtype_0, x = cache5)[name = tensor("cast_112")]; + tensor var_2425_cast_fp16 = slice_by_index(begin = var_2425_begin_0, end = var_2425_end_0, end_mask = var_2425_end_mask_0, squeeze_mask = var_2425_squeeze_mask_0, x = cache5_to_fp16)[name = tensor("op_2425_cast_fp16")]; + tensor var_2427_axis_0 = const()[name = tensor("op_2427_axis_0"), val = tensor(1)]; + tensor var_2427_mode_0 = const()[name = tensor("op_2427_mode_0"), val = tensor("update")]; + tensor var_2427_validate_indices_0 = const()[name = tensor("op_2427_validate_indices_0"), val = tensor(false)]; + tensor var_2427_cast_fp16 = scatter_along_axis(axis = var_2427_axis_0, data = var_2425_cast_fp16, indices = write_indices, mode = var_2427_mode_0, updates = k_cast_fp16, validate_indices = var_2427_validate_indices_0)[name = tensor("op_2427_cast_fp16")]; + tensor concat_37 = const()[name = tensor("concat_37"), val = tensor([0, 0, 0, 0, 0])]; + tensor concat_38 = const()[name = tensor("concat_38"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_internal_tensor_assign_1_stride_0 = const()[name = tensor("new_cache_internal_tensor_assign_1_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_internal_tensor_assign_1_begin_mask_0 = const()[name = tensor("new_cache_internal_tensor_assign_1_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_internal_tensor_assign_1_end_mask_0 = const()[name = tensor("new_cache_internal_tensor_assign_1_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_internal_tensor_assign_1_squeeze_mask_0 = const()[name = tensor("new_cache_internal_tensor_assign_1_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_22 = const()[name = tensor("shape_22"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_10 = const()[name = tensor("reduce_prod_10"), val = tensor(1048576)]; + tensor range_1d_10_start_0 = const()[name = tensor("range_1d_10_start_0"), val = tensor(0)]; + tensor range_1d_10_step_0 = const()[name = tensor("range_1d_10_step_0"), val = tensor(1)]; + tensor range_1d_10 = range_1d(end = reduce_prod_10, start = range_1d_10_start_0, step = range_1d_10_step_0)[name = tensor("range_1d_10")]; + tensor reshape_50 = reshape(shape = shape_22, x = range_1d_10)[name = tensor("reshape_50")]; + tensor slice_by_index_10 = slice_by_index(begin = concat_37, begin_mask = new_cache_internal_tensor_assign_1_begin_mask_0, end = concat_38, end_mask = new_cache_internal_tensor_assign_1_end_mask_0, squeeze_mask = new_cache_internal_tensor_assign_1_squeeze_mask_0, stride = new_cache_internal_tensor_assign_1_stride_0, x = reshape_50)[name = tensor("slice_by_index_10")]; + tensor reshape_51_shape_0 = const()[name = tensor("reshape_51_shape_0"), val = tensor([-1])]; + tensor reshape_51 = reshape(shape = reshape_51_shape_0, x = slice_by_index_10)[name = tensor("reshape_51")]; + tensor reshape_52_shape_0 = const()[name = tensor("reshape_52_shape_0"), val = tensor([-1])]; + tensor reshape_52_cast_fp16 = reshape(shape = reshape_52_shape_0, x = var_2427_cast_fp16)[name = tensor("reshape_52_cast_fp16")]; + tensor reshape_53_shape_0 = const()[name = tensor("reshape_53_shape_0"), val = tensor([-1])]; + tensor reshape_53_cast_fp16 = reshape(shape = reshape_53_shape_0, x = cache5_to_fp16)[name = tensor("reshape_53_cast_fp16")]; + tensor scatter_10_mode_0 = const()[name = tensor("scatter_10_mode_0"), val = tensor("update")]; + tensor scatter_10_axis_0 = const()[name = tensor("scatter_10_axis_0"), val = tensor(0)]; + tensor scatter_10_validate_indices_0 = const()[name = tensor("scatter_10_validate_indices_0"), val = tensor(false)]; + tensor scatter_10_cast_fp16 = scatter(axis = scatter_10_axis_0, data = reshape_53_cast_fp16, indices = reshape_51, mode = scatter_10_mode_0, updates = reshape_52_cast_fp16, validate_indices = scatter_10_validate_indices_0)[name = tensor("scatter_10_cast_fp16")]; + tensor reshape_54_cast_fp16 = reshape(shape = shape_22, x = scatter_10_cast_fp16)[name = tensor("reshape_54_cast_fp16")]; + tensor var_2435_begin_0 = const()[name = tensor("op_2435_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor var_2435_end_0 = const()[name = tensor("op_2435_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor var_2435_end_mask_0 = const()[name = tensor("op_2435_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_2435_squeeze_mask_0 = const()[name = tensor("op_2435_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor var_2435_cast_fp16 = slice_by_index(begin = var_2435_begin_0, end = var_2435_end_0, end_mask = var_2435_end_mask_0, squeeze_mask = var_2435_squeeze_mask_0, x = reshape_54_cast_fp16)[name = tensor("op_2435_cast_fp16")]; + tensor var_2437_axis_0 = const()[name = tensor("op_2437_axis_0"), val = tensor(1)]; + tensor var_2437_mode_0 = const()[name = tensor("op_2437_mode_0"), val = tensor("update")]; + tensor var_2437_validate_indices_0 = const()[name = tensor("op_2437_validate_indices_0"), val = tensor(false)]; + tensor var_2437_cast_fp16 = scatter_along_axis(axis = var_2437_axis_0, data = var_2435_cast_fp16, indices = write_indices, mode = var_2437_mode_0, updates = v_cast_fp16, validate_indices = var_2437_validate_indices_0)[name = tensor("op_2437_cast_fp16")]; + tensor concat_39 = const()[name = tensor("concat_39"), val = tensor([1, 0, 0, 0, 0])]; + tensor concat_40 = const()[name = tensor("concat_40"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_internal_tensor_assign_2_stride_0 = const()[name = tensor("new_cache_internal_tensor_assign_2_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_internal_tensor_assign_2_begin_mask_0 = const()[name = tensor("new_cache_internal_tensor_assign_2_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_internal_tensor_assign_2_end_mask_0 = const()[name = tensor("new_cache_internal_tensor_assign_2_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_internal_tensor_assign_2_squeeze_mask_0 = const()[name = tensor("new_cache_internal_tensor_assign_2_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_23 = const()[name = tensor("shape_23"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_11 = const()[name = tensor("reduce_prod_11"), val = tensor(1048576)]; + tensor range_1d_11_start_0 = const()[name = tensor("range_1d_11_start_0"), val = tensor(0)]; + tensor range_1d_11_step_0 = const()[name = tensor("range_1d_11_step_0"), val = tensor(1)]; + tensor range_1d_11 = range_1d(end = reduce_prod_11, start = range_1d_11_start_0, step = range_1d_11_step_0)[name = tensor("range_1d_11")]; + tensor reshape_55 = reshape(shape = shape_23, x = range_1d_11)[name = tensor("reshape_55")]; + tensor slice_by_index_11 = slice_by_index(begin = concat_39, begin_mask = new_cache_internal_tensor_assign_2_begin_mask_0, end = concat_40, end_mask = new_cache_internal_tensor_assign_2_end_mask_0, squeeze_mask = new_cache_internal_tensor_assign_2_squeeze_mask_0, stride = new_cache_internal_tensor_assign_2_stride_0, x = reshape_55)[name = tensor("slice_by_index_11")]; + tensor reshape_56_shape_0 = const()[name = tensor("reshape_56_shape_0"), val = tensor([-1])]; + tensor reshape_56 = reshape(shape = reshape_56_shape_0, x = slice_by_index_11)[name = tensor("reshape_56")]; + tensor reshape_57_shape_0 = const()[name = tensor("reshape_57_shape_0"), val = tensor([-1])]; + tensor reshape_57_cast_fp16 = reshape(shape = reshape_57_shape_0, x = var_2437_cast_fp16)[name = tensor("reshape_57_cast_fp16")]; + tensor reshape_58_shape_0 = const()[name = tensor("reshape_58_shape_0"), val = tensor([-1])]; + tensor reshape_58_cast_fp16 = reshape(shape = reshape_58_shape_0, x = reshape_54_cast_fp16)[name = tensor("reshape_58_cast_fp16")]; + tensor scatter_11_mode_0 = const()[name = tensor("scatter_11_mode_0"), val = tensor("update")]; + tensor scatter_11_axis_0 = const()[name = tensor("scatter_11_axis_0"), val = tensor(0)]; + tensor scatter_11_validate_indices_0 = const()[name = tensor("scatter_11_validate_indices_0"), val = tensor(false)]; + tensor scatter_11_cast_fp16 = scatter(axis = scatter_11_axis_0, data = reshape_58_cast_fp16, indices = reshape_56, mode = scatter_11_mode_0, updates = reshape_57_cast_fp16, validate_indices = scatter_11_validate_indices_0)[name = tensor("scatter_11_cast_fp16")]; + tensor reshape_59_cast_fp16 = reshape(shape = shape_23, x = scatter_11_cast_fp16)[name = tensor("reshape_59_cast_fp16")]; + tensor reshape_59_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("reshape_59_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor keys_31_begin_0 = const()[name = tensor("keys_31_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor keys_31_end_0 = const()[name = tensor("keys_31_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor keys_31_end_mask_0 = const()[name = tensor("keys_31_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor keys_31_squeeze_mask_0 = const()[name = tensor("keys_31_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor keys_31_cast_fp16 = slice_by_index(begin = keys_31_begin_0, end = keys_31_end_0, end_mask = keys_31_end_mask_0, squeeze_mask = keys_31_squeeze_mask_0, x = reshape_59_cast_fp16)[name = tensor("keys_31_cast_fp16")]; + tensor values_31_begin_0 = const()[name = tensor("values_31_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor values_31_end_0 = const()[name = tensor("values_31_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor values_31_end_mask_0 = const()[name = tensor("values_31_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor values_31_squeeze_mask_0 = const()[name = tensor("values_31_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor values_31_cast_fp16 = slice_by_index(begin = values_31_begin_0, end = values_31_end_0, end_mask = values_31_end_mask_0, squeeze_mask = values_31_squeeze_mask_0, x = reshape_59_cast_fp16)[name = tensor("values_31_cast_fp16")]; + tensor var_2449_cast_fp16 = not_equal(x = keys_31_cast_fp16, y = keys_31_cast_fp16)[name = tensor("op_2449_cast_fp16")]; + tensor keys_33_cast_fp16 = select(a = var_360_to_fp16, b = keys_31_cast_fp16, cond = var_2449_cast_fp16)[name = tensor("keys_33_cast_fp16")]; + tensor var_2457_cast_fp16 = not_equal(x = values_31_cast_fp16, y = values_31_cast_fp16)[name = tensor("op_2457_cast_fp16")]; + tensor values_33_cast_fp16 = select(a = var_360_to_fp16, b = values_31_cast_fp16, cond = var_2457_cast_fp16)[name = tensor("values_33_cast_fp16")]; + tensor var_2481 = const()[name = tensor("op_2481"), val = tensor([0, 2, 1, 3])]; + tensor var_2494 = const()[name = tensor("op_2494"), val = tensor([1, 1, 1])]; + tensor var_2495_cast_fp16 = reshape(shape = var_2494, x = position5_to_fp16)[name = tensor("op_2495_cast_fp16")]; + tensor var_2512_to_fp16 = const()[name = tensor("op_2512_to_fp16"), val = tensor(0x1p+0)]; + tensor valid_len_cast_fp16 = add(x = var_2495_cast_fp16, y = var_2512_to_fp16)[name = tensor("valid_len_cast_fp16")]; + tensor valid_mask_cast_fp16 = less(x = k_positions_1_promoted_to_fp16, y = valid_len_cast_fp16)[name = tensor("valid_mask_cast_fp16")]; + tensor causal_mask_cast_fp16 = less_equal(x = k_positions_1_promoted_to_fp16, y = var_2495_cast_fp16)[name = tensor("causal_mask_cast_fp16")]; + tensor attn_mask_21 = logical_and(x = valid_mask_cast_fp16, y = causal_mask_cast_fp16)[name = tensor("attn_mask_21")]; + tensor attn_mask_axes_0 = const()[name = tensor("attn_mask_axes_0"), val = tensor([1])]; + tensor attn_mask = expand_dims(axes = attn_mask_axes_0, x = attn_mask_21)[name = tensor("attn_mask")]; + tensor var_2530_transpose_x_0 = const()[name = tensor("op_2530_transpose_x_0"), val = tensor(false)]; + tensor var_2530_transpose_y_0 = const()[name = tensor("op_2530_transpose_y_0"), val = tensor(false)]; + tensor transpose_28_perm_0 = const()[name = tensor("transpose_28_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_29_perm_0 = const()[name = tensor("transpose_29_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_29 = transpose(perm = transpose_29_perm_0, x = keys_33_cast_fp16)[name = tensor("transpose_31")]; + tensor transpose_28 = transpose(perm = transpose_28_perm_0, x = q_33_cast_fp16)[name = tensor("transpose_32")]; + tensor var_2530_cast_fp16 = matmul(transpose_x = var_2530_transpose_x_0, transpose_y = var_2530_transpose_y_0, x = transpose_28, y = transpose_29)[name = tensor("op_2530_cast_fp16")]; + tensor var_2524_to_fp16 = const()[name = tensor("op_2524_to_fp16"), val = tensor([0x1p-3])]; + tensor attn_weights_31_cast_fp16 = mul(x = var_2530_cast_fp16, y = var_2524_to_fp16)[name = tensor("attn_weights_31_cast_fp16")]; + tensor var_2532 = logical_not(x = attn_mask)[name = tensor("op_2532")]; + tensor var_2533_to_fp16 = const()[name = tensor("op_2533_to_fp16"), val = tensor(-inf)]; + tensor attn_weights_33_cast_fp16 = select(a = var_2533_to_fp16, b = attn_weights_31_cast_fp16, cond = var_2532)[name = tensor("attn_weights_33_cast_fp16")]; + tensor var_2535 = const()[name = tensor("op_2535"), val = tensor(-1)]; + tensor attn_weights_cast_fp16 = softmax(axis = var_2535, x = attn_weights_33_cast_fp16)[name = tensor("attn_weights_cast_fp16")]; + tensor attn_output_transpose_x_0 = const()[name = tensor("attn_output_transpose_x_0"), val = tensor(false)]; + tensor attn_output_transpose_y_0 = const()[name = tensor("attn_output_transpose_y_0"), val = tensor(false)]; + tensor values_cast_fp16 = transpose(perm = var_2481, x = values_33_cast_fp16)[name = tensor("transpose_33")]; + tensor attn_output_cast_fp16 = matmul(transpose_x = attn_output_transpose_x_0, transpose_y = attn_output_transpose_y_0, x = attn_weights_cast_fp16, y = values_cast_fp16)[name = tensor("attn_output_cast_fp16")]; + tensor var_2543 = const()[name = tensor("op_2543"), val = tensor([0, 2, 1, 3])]; + tensor var_2546 = const()[name = tensor("op_2546"), val = tensor([1, 1, 1024])]; + tensor var_2544_cast_fp16 = transpose(perm = var_2543, x = attn_output_cast_fp16)[name = tensor("transpose_30")]; + tensor input_55_cast_fp16 = reshape(shape = var_2546, x = var_2544_cast_fp16)[name = tensor("input_55_cast_fp16")]; + tensor linear_22_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = attn5_out_proj_weight_cast_fp16, x = input_55_cast_fp16)[name = tensor("linear_22_cast_fp16")]; + tensor var_2552_to_fp16 = const()[name = tensor("op_2552_to_fp16"), val = tensor(0x1p+0)]; + tensor var_2553_cast_fp16 = add(x = position5_to_fp16, y = var_2552_to_fp16)[name = tensor("op_2553_cast_fp16")]; + tensor var_2553_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_2553_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor input_57_cast_fp16 = add(x = input_53_cast_fp16, y = linear_22_cast_fp16)[name = tensor("input_57_cast_fp16")]; + tensor input_59_axes_0 = const()[name = tensor("input_59_axes_0"), val = tensor([-1])]; + tensor norm5_2_weight_to_fp16 = const()[name = tensor("norm5_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(76798720)))]; + tensor norm5_2_bias_to_fp16 = const()[name = tensor("norm5_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(76800832)))]; + tensor var_2557_to_fp16 = const()[name = tensor("op_2557_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_59_cast_fp16 = layer_norm(axes = input_59_axes_0, beta = norm5_2_bias_to_fp16, epsilon = var_2557_to_fp16, gamma = norm5_2_weight_to_fp16, x = input_57_cast_fp16)[name = tensor("input_59_cast_fp16")]; + tensor linear_23_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = linear5_1_weight_cast_fp16, x = input_59_cast_fp16)[name = tensor("linear_23_cast_fp16")]; + tensor input_61_mode_0 = const()[name = tensor("input_61_mode_0"), val = tensor("EXACT")]; + tensor input_61_cast_fp16 = gelu(mode = input_61_mode_0, x = linear_23_cast_fp16)[name = tensor("input_61_cast_fp16")]; + tensor linear_24_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = linear5_2_weight_cast_fp16, x = input_61_cast_fp16)[name = tensor("linear_24_cast_fp16")]; + tensor input_63_cast_fp16 = add(x = input_57_cast_fp16, y = linear_24_cast_fp16)[name = tensor("input_63_cast_fp16")]; + tensor input_axes_0 = const()[name = tensor("input_axes_0"), val = tensor([-1])]; + tensor out_norm_weight_to_fp16 = const()[name = tensor("out_norm_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(76802944)))]; + tensor out_norm_bias_to_fp16 = const()[name = tensor("out_norm_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(76805056)))]; + tensor var_2574_to_fp16 = const()[name = tensor("op_2574_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_cast_fp16 = layer_norm(axes = input_axes_0, beta = out_norm_bias_to_fp16, epsilon = var_2574_to_fp16, gamma = out_norm_weight_to_fp16, x = input_63_cast_fp16)[name = tensor("input_cast_fp16")]; + tensor input_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("input_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor out_eos_weight_to_fp16 = const()[name = tensor("out_eos_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(76807168)))]; + tensor out_eos_bias_to_fp16 = const()[name = tensor("out_eos_bias_to_fp16"), val = tensor([-0x1.3ep-3])]; + tensor linear_25_cast_fp16 = linear(bias = out_eos_bias_to_fp16, weight = out_eos_weight_to_fp16, x = input_cast_fp16)[name = tensor("linear_25_cast_fp16")]; + tensor linear_25_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("linear_25_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor var_2582 = cast(dtype = linear_25_cast_fp16_to_fp32_dtype_0, x = linear_25_cast_fp16)[name = tensor("cast_108")]; + tensor input = cast(dtype = input_cast_fp16_to_fp32_dtype_0, x = input_cast_fp16)[name = tensor("cast_109")]; + tensor var_2553 = cast(dtype = var_2553_cast_fp16_to_fp32_dtype_0, x = var_2553_cast_fp16)[name = tensor("cast_110")]; + tensor new_cache_internal_tensor_assign_2 = cast(dtype = reshape_59_cast_fp16_to_fp32_dtype_0, x = reshape_59_cast_fp16)[name = tensor("cast_111")]; + tensor var_2134 = cast(dtype = var_2134_cast_fp16_to_fp32_dtype_0, x = var_2134_cast_fp16)[name = tensor("cast_115")]; + tensor new_cache_9_internal_tensor_assign_2 = cast(dtype = reshape_49_cast_fp16_to_fp32_dtype_0, x = reshape_49_cast_fp16)[name = tensor("cast_116")]; + tensor var_1715 = cast(dtype = var_1715_cast_fp16_to_fp32_dtype_0, x = var_1715_cast_fp16)[name = tensor("cast_120")]; + tensor new_cache_7_internal_tensor_assign_2 = cast(dtype = reshape_39_cast_fp16_to_fp32_dtype_0, x = reshape_39_cast_fp16)[name = tensor("cast_121")]; + tensor var_1296 = cast(dtype = var_1296_cast_fp16_to_fp32_dtype_0, x = var_1296_cast_fp16)[name = tensor("cast_125")]; + tensor new_cache_5_internal_tensor_assign_2 = cast(dtype = reshape_29_cast_fp16_to_fp32_dtype_0, x = reshape_29_cast_fp16)[name = tensor("cast_126")]; + tensor var_877 = cast(dtype = var_877_cast_fp16_to_fp32_dtype_0, x = var_877_cast_fp16)[name = tensor("cast_130")]; + tensor new_cache_3_internal_tensor_assign_2 = cast(dtype = reshape_19_cast_fp16_to_fp32_dtype_0, x = reshape_19_cast_fp16)[name = tensor("cast_131")]; + tensor var_458 = cast(dtype = var_458_cast_fp16_to_fp32_dtype_0, x = var_458_cast_fp16)[name = tensor("cast_135")]; + tensor new_cache_1_internal_tensor_assign_2 = cast(dtype = reshape_9_cast_fp16_to_fp32_dtype_0, x = reshape_9_cast_fp16)[name = tensor("cast_136")]; + } -> (input, var_2582, new_cache_1_internal_tensor_assign_2, var_458, new_cache_3_internal_tensor_assign_2, var_877, new_cache_5_internal_tensor_assign_2, var_1296, new_cache_7_internal_tensor_assign_2, var_1715, new_cache_9_internal_tensor_assign_2, var_2134, new_cache_internal_tensor_assign_2, var_2553); +} \ No newline at end of file diff --git a/v2.1/german/flowlm_stepv2.mlmodelc/weights/weight.bin b/v2.1/german/flowlm_stepv2.mlmodelc/weights/weight.bin new file mode 100644 index 0000000000000000000000000000000000000000..4997c7acb14d8a29cd40b89f154fe1d13ed22c8e --- /dev/null +++ b/v2.1/german/flowlm_stepv2.mlmodelc/weights/weight.bin @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:f67bbab3508d3e5fb2b1537a715e6c9f6dc1045043b272dcd1cbedbc688ee72d +size 76809280 diff --git a/v2.1/german/flowlm_stepv2.mlpackage/Data/com.apple.CoreML/model.mlmodel b/v2.1/german/flowlm_stepv2.mlpackage/Data/com.apple.CoreML/model.mlmodel new file mode 100644 index 0000000000000000000000000000000000000000..87927abdaba18b4570ccb8ff54ff300b993edb0c --- /dev/null +++ b/v2.1/german/flowlm_stepv2.mlpackage/Data/com.apple.CoreML/model.mlmodel @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:17b71431390bd43b0085a74fa1e9c836eb4d7ab3eda220dd59f38eb68e9aaa47 +size 212305 diff --git a/v2.1/german/flowlm_stepv2.mlpackage/Data/com.apple.CoreML/weights/weight.bin b/v2.1/german/flowlm_stepv2.mlpackage/Data/com.apple.CoreML/weights/weight.bin new file mode 100644 index 0000000000000000000000000000000000000000..4997c7acb14d8a29cd40b89f154fe1d13ed22c8e --- /dev/null +++ b/v2.1/german/flowlm_stepv2.mlpackage/Data/com.apple.CoreML/weights/weight.bin @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:f67bbab3508d3e5fb2b1537a715e6c9f6dc1045043b272dcd1cbedbc688ee72d +size 76809280 diff --git a/v2.1/german/flowlm_stepv2.mlpackage/Manifest.json b/v2.1/german/flowlm_stepv2.mlpackage/Manifest.json new file mode 100644 index 0000000000000000000000000000000000000000..375fe1a296c396fe9ccda569e338ddd52a3d1769 --- /dev/null +++ b/v2.1/german/flowlm_stepv2.mlpackage/Manifest.json @@ -0,0 +1,18 @@ +{ + "fileFormatVersion": "1.0.0", + "itemInfoEntries": { + "70E8EEA8-3327-49AB-B440-1FA147DA383F": { + "author": "com.apple.CoreML", + "description": "CoreML Model Weights", + "name": "weights", + "path": "com.apple.CoreML/weights" + }, + "9E81B35A-9338-440B-B812-A27BE63EC4A9": { + "author": "com.apple.CoreML", + "description": "CoreML Model Specification", + "name": "model.mlmodel", + "path": "com.apple.CoreML/model.mlmodel" + } + }, + "rootModelIdentifier": "9E81B35A-9338-440B-B812-A27BE63EC4A9" +} diff --git a/v2.1/german/manifest.json b/v2.1/german/manifest.json new file mode 100644 index 0000000000000000000000000000000000000000..fc7da98920126d1261a9c47517b1303db1599db7 --- /dev/null +++ b/v2.1/german/manifest.json @@ -0,0 +1,27 @@ +{ + "version": "2.1", + "base": "v2", + "kind": "optimization-reconvert", + "self_contained": true, + "note": "Same weights as v2. Re-converted for speed: fused flow decoder (100% ANE), one-shot cond prefill, fp16 flowlm. NOT a finetune.", + "new": [ + "flow_decoder_fused", + "cond_prefill", + "flowlm_step" + ], + "copied_from_v2": [ + "flowlm_stepv2.mlpackage", + "flowlm_stepv2.mlmodelc", + "mimi_decoder.mlpackage", + "mimi_decoder.mlmodelc", + "constants", + "constants_bin" + ], + "compute_units": { + "flow_decoder_fused": "all (ANE)", + "cond_prefill": "all (GPU)", + "flowlm_step": "all (GPU, fp16)", + "flowlm_stepv2": "cpuAndGpu (GPU, int8 \u2014 fastest flowlm)", + "mimi_decoder": "cpuOnly" + } +} diff --git a/v2.1/german/mimi_decoder.mlmodelc/analytics/coremldata.bin b/v2.1/german/mimi_decoder.mlmodelc/analytics/coremldata.bin new file mode 100644 index 0000000000000000000000000000000000000000..0b70c68114524d9c935536cf6128d1a85ca48150 --- /dev/null +++ b/v2.1/german/mimi_decoder.mlmodelc/analytics/coremldata.bin @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:0f7c42f7b8c9c6e53e1ced55eebc608f339bf1b2e724ff2302150667fd412ba2 +size 243 diff --git a/v2.1/german/mimi_decoder.mlmodelc/coremldata.bin b/v2.1/german/mimi_decoder.mlmodelc/coremldata.bin new file mode 100644 index 0000000000000000000000000000000000000000..76546b8da4a2a86e85df21fa619ee0286400f2f7 --- /dev/null +++ b/v2.1/german/mimi_decoder.mlmodelc/coremldata.bin @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:2fb9d7ab308ebf86a14e5cf3841c34a9c6e5716d2be56301f35609261ce54f2e +size 1777 diff --git a/v2.1/german/mimi_decoder.mlmodelc/metadata.json b/v2.1/german/mimi_decoder.mlmodelc/metadata.json new file mode 100644 index 0000000000000000000000000000000000000000..f66f9e7ff5be2ffbc44c8300f3501f67a77afaf6 --- /dev/null +++ b/v2.1/german/mimi_decoder.mlmodelc/metadata.json @@ -0,0 +1,573 @@ +[ + { + "metadataOutputVersion" : "3.0", + "storagePrecision" : "Float32", + "outputSchema" : [ + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1 × 1 × 1920)", + "shortDescription" : "", + "shape" : "[1, 1, 1920]", + "name" : "var_724", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1 × 512 × 16)", + "shortDescription" : "", + "shape" : "[1, 512, 16]", + "name" : "var_77", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 2 × 1 × 256 × 8 × 64)", + "shortDescription" : "", + "shape" : "[2, 1, 256, 8, 64]", + "name" : "var_210", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1)", + "shortDescription" : "", + "shape" : "[1]", + "name" : "var_741", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 2 × 1 × 256 × 8 × 64)", + "shortDescription" : "", + "shape" : "[2, 1, 256, 8, 64]", + "name" : "var_400", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1)", + "shortDescription" : "", + "shape" : "[1]", + "name" : "var_744", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1 × 512 × 6)", + "shortDescription" : "", + "shape" : "[1, 512, 6]", + "name" : "var_542", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1)", + "shortDescription" : "", + "shape" : "[1]", + "name" : "conv0_first", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1 × 256 × 6)", + "shortDescription" : "", + "shape" : "[1, 256, 6]", + "name" : "var_565", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1 × 256 × 2)", + "shortDescription" : "", + "shape" : "[1, 256, 2]", + "name" : "var_585", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1)", + "shortDescription" : "", + "shape" : "[1]", + "name" : "res0_conv0_first", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1 × 128 × 0)", + "shortDescription" : "", + "shape" : "[1, 128, 0]", + "name" : "res0_conv1_prev", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1)", + "shortDescription" : "", + "shape" : "[1]", + "name" : "res0_conv1_first", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1 × 128 × 5)", + "shortDescription" : "", + "shape" : "[1, 128, 5]", + "name" : "var_621", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1 × 128 × 2)", + "shortDescription" : "", + "shape" : "[1, 128, 2]", + "name" : "var_641", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1)", + "shortDescription" : "", + "shape" : "[1]", + "name" : "res1_conv0_first", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1 × 64 × 0)", + "shortDescription" : "", + "shape" : "[1, 64, 0]", + "name" : "res1_conv1_prev", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1)", + "shortDescription" : "", + "shape" : "[1]", + "name" : "res1_conv1_first", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1 × 64 × 4)", + "shortDescription" : "", + "shape" : "[1, 64, 4]", + "name" : "var_677", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1 × 64 × 2)", + "shortDescription" : "", + "shape" : "[1, 64, 2]", + "name" : "var_697", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1)", + "shortDescription" : "", + "shape" : "[1]", + "name" : "res2_conv0_first", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1 × 32 × 0)", + "shortDescription" : "", + "shape" : "[1, 32, 0]", + "name" : "res2_conv1_prev", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1)", + "shortDescription" : "", + "shape" : "[1]", + "name" : "res2_conv1_first", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1 × 64 × 2)", + "shortDescription" : "", + "shape" : "[1, 64, 2]", + "name" : "var_725", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1)", + "shortDescription" : "", + "shape" : "[1]", + "name" : "conv_final_first", + "type" : "MultiArray" + } + ], + "modelParameters" : [ + + ], + "specificationVersion" : 8, + "mlProgramOperationTypeHistogram" : { + "Ios17.notEqual" : 4, + "Tile" : 2, + "Ios17.lessEqual" : 2, + "Ios17.layerNorm" : 4, + "Stack" : 6, + "Ios17.greaterEqual" : 4, + "Ios17.scatterAlongAxis" : 4, + "Ios17.sub" : 17, + "Identity" : 11, + "Ios17.elu" : 10, + "Ios17.logicalNot" : 2, + "Ios17.sliceByIndex" : 35, + "Ios17.concat" : 9, + "Ios17.convTranspose" : 4, + "Ios17.floorDiv" : 4, + "Ios17.expandDims" : 7, + "Ios17.less" : 2, + "Ios17.mul" : 29, + "Ios17.transpose" : 10, + "Ios17.matmul" : 4, + "Ios16.gelu" : 2, + "Ios17.conv" : 9, + "Ios17.reshape" : 22, + "Split" : 2, + "Ios17.squeeze" : 6, + "Ios17.cast" : 6, + "Ios17.sin" : 2, + "Ios17.linear" : 8, + "Select" : 6, + "Ios16.softmax" : 2, + "Ios17.add" : 28, + "Ios17.cos" : 2, + "Ios17.logicalAnd" : 6 + }, + "computePrecision" : "Mixed (Float32, Int32)", + "isUpdatable" : "0", + "stateSchema" : [ + + ], + "availability" : { + "macOS" : "14.0", + "tvOS" : "17.0", + "visionOS" : "1.0", + "watchOS" : "10.0", + "iOS" : "17.0", + "macCatalyst" : "17.0" + }, + "modelType" : { + "name" : "MLModelType_mlProgram" + }, + "userDefinedMetadata" : { + "com.github.apple.coremltools.conversion_date" : "2026-04-24", + "com.github.apple.coremltools.source" : "torch==2.9.1", + "com.github.apple.coremltools.version" : "9.0", + "com.github.apple.coremltools.source_dialect" : "TorchScript" + }, + "inputSchema" : [ + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1 × 32)", + "shortDescription" : "", + "shape" : "[1, 32]", + "name" : "latent", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1 × 512 × 16)", + "shortDescription" : "", + "shape" : "[1, 512, 16]", + "name" : "upsample_partial", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 2 × 1 × 256 × 8 × 64)", + "shortDescription" : "", + "shape" : "[2, 1, 256, 8, 64]", + "name" : "attn0_cache", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1)", + "shortDescription" : "", + "shape" : "[1]", + "name" : "attn0_offset", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 2 × 1 × 256 × 8 × 64)", + "shortDescription" : "", + "shape" : "[2, 1, 256, 8, 64]", + "name" : "attn1_cache", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1)", + "shortDescription" : "", + "shape" : "[1]", + "name" : "attn1_offset", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1 × 512 × 6)", + "shortDescription" : "", + "shape" : "[1, 512, 6]", + "name" : "conv0_prev", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1)", + "shortDescription" : "", + "shape" : "[1]", + "name" : "conv0_first", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1 × 256 × 6)", + "shortDescription" : "", + "shape" : "[1, 256, 6]", + "name" : "convtr0_partial", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1 × 256 × 2)", + "shortDescription" : "", + "shape" : "[1, 256, 2]", + "name" : "res0_conv0_prev", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1)", + "shortDescription" : "", + "shape" : "[1]", + "name" : "res0_conv0_first", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1 × 128 × 0)", + "shortDescription" : "", + "shape" : "[1, 128, 0]", + "name" : "res0_conv1_prev", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1)", + "shortDescription" : "", + "shape" : "[1]", + "name" : "res0_conv1_first", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1 × 128 × 5)", + "shortDescription" : "", + "shape" : "[1, 128, 5]", + "name" : "convtr1_partial", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1 × 128 × 2)", + "shortDescription" : "", + "shape" : "[1, 128, 2]", + "name" : "res1_conv0_prev", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1)", + "shortDescription" : "", + "shape" : "[1]", + "name" : "res1_conv0_first", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1 × 64 × 0)", + "shortDescription" : "", + "shape" : "[1, 64, 0]", + "name" : "res1_conv1_prev", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1)", + "shortDescription" : "", + "shape" : "[1]", + "name" : "res1_conv1_first", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1 × 64 × 4)", + "shortDescription" : "", + "shape" : "[1, 64, 4]", + "name" : "convtr2_partial", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1 × 64 × 2)", + "shortDescription" : "", + "shape" : "[1, 64, 2]", + "name" : "res2_conv0_prev", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1)", + "shortDescription" : "", + "shape" : "[1]", + "name" : "res2_conv0_first", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1 × 32 × 0)", + "shortDescription" : "", + "shape" : "[1, 32, 0]", + "name" : "res2_conv1_prev", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1)", + "shortDescription" : "", + "shape" : "[1]", + "name" : "res2_conv1_first", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1 × 64 × 2)", + "shortDescription" : "", + "shape" : "[1, 64, 2]", + "name" : "conv_final_prev", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1)", + "shortDescription" : "", + "shape" : "[1]", + "name" : "conv_final_first", + "type" : "MultiArray" + } + ], + "generatedClassName" : "mimi_decoder", + "method" : "predict" + } +] \ No newline at end of file diff --git a/v2.1/german/mimi_decoder.mlmodelc/model.mil b/v2.1/german/mimi_decoder.mlmodelc/model.mil new file mode 100644 index 0000000000000000000000000000000000000000..85cbf2d02c036f2f26d39a5240feb64c94bd10c9 --- /dev/null +++ b/v2.1/german/mimi_decoder.mlmodelc/model.mil @@ -0,0 +1,646 @@ +program(1.0) +[buildInfo = dict, tensor>({{"coremlc-component-MIL", "3520.4.1"}, {"coremlc-version", "3520.5.1"}, {"coremltools-component-torch", "2.9.1"}, {"coremltools-source-dialect", "TorchScript"}, {"coremltools-version", "9.0"}})] +{ + func main(tensor attn0_cache, tensor attn0_offset, tensor attn1_cache, tensor attn1_offset, tensor conv0_first, tensor conv0_prev, tensor conv_final_first, tensor conv_final_prev, tensor convtr0_partial, tensor convtr1_partial, tensor convtr2_partial, tensor latent, tensor res0_conv0_first, tensor res0_conv0_prev, tensor res0_conv1_first, tensor res0_conv1_prev, tensor res1_conv0_first, tensor res1_conv0_prev, tensor res1_conv1_first, tensor res1_conv1_prev, tensor res2_conv0_first, tensor res2_conv0_prev, tensor res2_conv1_first, tensor res2_conv1_prev, tensor upsample_partial) { + tensor emb_mean = const()[name = tensor("emb_mean"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(64)))]; + tensor emb_std = const()[name = tensor("emb_std"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(256)))]; + tensor mimi_quantizer_output_proj_weight = const()[name = tensor("mimi_quantizer_output_proj_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(448)))]; + tensor mimi_upsample_convtr_convtr_weight = const()[name = tensor("mimi_upsample_convtr_convtr_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(66048)))]; + tensor mimi_decoder_transformer_transformer_layers_0_norm1_bias = const()[name = tensor("mimi_decoder_transformer_transformer_layers_0_norm1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(131648)))]; + tensor mimi_decoder_transformer_transformer_layers_0_norm1_weight = const()[name = tensor("mimi_decoder_transformer_transformer_layers_0_norm1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(133760)))]; + tensor mimi_decoder_transformer_transformer_layers_0_self_attn_in_proj_weight = const()[name = tensor("mimi_decoder_transformer_transformer_layers_0_self_attn_in_proj_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(135872)))]; + tensor mimi_decoder_transformer_transformer_layers_0_self_attn_out_proj_weight = const()[name = tensor("mimi_decoder_transformer_transformer_layers_0_self_attn_out_proj_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(3281664)))]; + tensor mimi_decoder_transformer_transformer_layers_0_layer_scale_1_scale = const()[name = tensor("mimi_decoder_transformer_transformer_layers_0_layer_scale_1_scale"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(4330304)))]; + tensor mimi_decoder_transformer_transformer_layers_0_norm2_bias = const()[name = tensor("mimi_decoder_transformer_transformer_layers_0_norm2_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(4332416)))]; + tensor mimi_decoder_transformer_transformer_layers_0_norm2_weight = const()[name = tensor("mimi_decoder_transformer_transformer_layers_0_norm2_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(4334528)))]; + tensor mimi_decoder_transformer_transformer_layers_0_linear1_weight = const()[name = tensor("mimi_decoder_transformer_transformer_layers_0_linear1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(4336640)))]; + tensor mimi_decoder_transformer_transformer_layers_0_linear2_weight = const()[name = tensor("mimi_decoder_transformer_transformer_layers_0_linear2_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8531008)))]; + tensor mimi_decoder_transformer_transformer_layers_0_layer_scale_2_scale = const()[name = tensor("mimi_decoder_transformer_transformer_layers_0_layer_scale_2_scale"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(12725376)))]; + tensor mimi_decoder_transformer_transformer_layers_1_norm1_bias = const()[name = tensor("mimi_decoder_transformer_transformer_layers_1_norm1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(12727488)))]; + tensor mimi_decoder_transformer_transformer_layers_1_norm1_weight = const()[name = tensor("mimi_decoder_transformer_transformer_layers_1_norm1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(12729600)))]; + tensor mimi_decoder_transformer_transformer_layers_1_self_attn_in_proj_weight = const()[name = tensor("mimi_decoder_transformer_transformer_layers_1_self_attn_in_proj_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(12731712)))]; + tensor mimi_decoder_transformer_transformer_layers_1_self_attn_out_proj_weight = const()[name = tensor("mimi_decoder_transformer_transformer_layers_1_self_attn_out_proj_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15877504)))]; + tensor mimi_decoder_transformer_transformer_layers_1_layer_scale_1_scale = const()[name = tensor("mimi_decoder_transformer_transformer_layers_1_layer_scale_1_scale"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(16926144)))]; + tensor mimi_decoder_transformer_transformer_layers_1_norm2_bias = const()[name = tensor("mimi_decoder_transformer_transformer_layers_1_norm2_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(16928256)))]; + tensor mimi_decoder_transformer_transformer_layers_1_norm2_weight = const()[name = tensor("mimi_decoder_transformer_transformer_layers_1_norm2_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(16930368)))]; + tensor mimi_decoder_transformer_transformer_layers_1_linear1_weight = const()[name = tensor("mimi_decoder_transformer_transformer_layers_1_linear1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(16932480)))]; + tensor mimi_decoder_transformer_transformer_layers_1_linear2_weight = const()[name = tensor("mimi_decoder_transformer_transformer_layers_1_linear2_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21126848)))]; + tensor mimi_decoder_transformer_transformer_layers_1_layer_scale_2_scale = const()[name = tensor("mimi_decoder_transformer_transformer_layers_1_layer_scale_2_scale"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(25321216)))]; + tensor mimi_decoder_model_0_conv_bias = const()[name = tensor("mimi_decoder_model_0_conv_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(25323328)))]; + tensor mimi_decoder_model_0_conv_weight = const()[name = tensor("mimi_decoder_model_0_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(25325440)))]; + tensor mimi_decoder_model_2_convtr_bias = const()[name = tensor("mimi_decoder_model_2_convtr_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(32665536)))]; + tensor mimi_decoder_model_2_convtr_weight = const()[name = tensor("mimi_decoder_model_2_convtr_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(32666624)))]; + tensor mimi_decoder_model_3_block_1_conv_bias = const()[name = tensor("mimi_decoder_model_3_block_1_conv_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(38958144)))]; + tensor mimi_decoder_model_3_block_1_conv_weight = const()[name = tensor("mimi_decoder_model_3_block_1_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(38958720)))]; + tensor mimi_decoder_model_3_block_3_conv_bias = const()[name = tensor("mimi_decoder_model_3_block_3_conv_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(39352000)))]; + tensor mimi_decoder_model_3_block_3_conv_weight = const()[name = tensor("mimi_decoder_model_3_block_3_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(39353088)))]; + tensor mimi_decoder_model_5_convtr_bias = const()[name = tensor("mimi_decoder_model_5_convtr_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(39484224)))]; + tensor mimi_decoder_model_5_convtr_weight = const()[name = tensor("mimi_decoder_model_5_convtr_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(39484800)))]; + tensor mimi_decoder_model_6_block_1_conv_bias = const()[name = tensor("mimi_decoder_model_6_block_1_conv_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(40795584)))]; + tensor mimi_decoder_model_6_block_1_conv_weight = const()[name = tensor("mimi_decoder_model_6_block_1_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(40795904)))]; + tensor mimi_decoder_model_6_block_3_conv_bias = const()[name = tensor("mimi_decoder_model_6_block_3_conv_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(40894272)))]; + tensor mimi_decoder_model_6_block_3_conv_weight = const()[name = tensor("mimi_decoder_model_6_block_3_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(40894848)))]; + tensor mimi_decoder_model_8_convtr_bias = const()[name = tensor("mimi_decoder_model_8_convtr_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(40927680)))]; + tensor mimi_decoder_model_8_convtr_weight = const()[name = tensor("mimi_decoder_model_8_convtr_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(40928000)))]; + tensor mimi_decoder_model_9_block_1_conv_bias = const()[name = tensor("mimi_decoder_model_9_block_1_conv_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(41190208)))]; + tensor mimi_decoder_model_9_block_1_conv_weight = const()[name = tensor("mimi_decoder_model_9_block_1_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(41190400)))]; + tensor mimi_decoder_model_9_block_3_conv_bias = const()[name = tensor("mimi_decoder_model_9_block_3_conv_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(41215040)))]; + tensor mimi_decoder_model_9_block_3_conv_weight = const()[name = tensor("mimi_decoder_model_9_block_3_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(41215360)))]; + tensor mimi_decoder_model_11_conv_bias = const()[name = tensor("mimi_decoder_model_11_conv_bias"), val = tensor([-0x1.8p-13])]; + tensor mimi_decoder_model_11_conv_weight = const()[name = tensor("mimi_decoder_model_11_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(41223616)))]; + tensor var_38 = mul(x = latent, y = emb_std)[name = tensor("op_38")]; + tensor denorm = add(x = var_38, y = emb_mean)[name = tensor("denorm")]; + tensor input_1_axes_0 = const()[name = tensor("input_1_axes_0"), val = tensor([-1])]; + tensor input_1 = expand_dims(axes = input_1_axes_0, x = denorm)[name = tensor("input_1")]; + tensor x_1_pad_type_0 = const()[name = tensor("x_1_pad_type_0"), val = tensor("valid")]; + tensor x_1_strides_0 = const()[name = tensor("x_1_strides_0"), val = tensor([1])]; + tensor x_1_pad_0 = const()[name = tensor("x_1_pad_0"), val = tensor([0, 0])]; + tensor x_1_dilations_0 = const()[name = tensor("x_1_dilations_0"), val = tensor([1])]; + tensor x_1_groups_0 = const()[name = tensor("x_1_groups_0"), val = tensor(1)]; + tensor x_1 = conv(dilations = x_1_dilations_0, groups = x_1_groups_0, pad = x_1_pad_0, pad_type = x_1_pad_type_0, strides = x_1_strides_0, weight = mimi_quantizer_output_proj_weight, x = input_1)[name = tensor("x_1")]; + tensor var_62 = const()[name = tensor("op_62"), val = tensor(-1)]; + tensor y_1_pad_type_0 = const()[name = tensor("y_1_pad_type_0"), val = tensor("valid")]; + tensor y_1_strides_0 = const()[name = tensor("y_1_strides_0"), val = tensor([16])]; + tensor y_1_groups_0 = const()[name = tensor("y_1_groups_0"), val = tensor(512)]; + tensor y_1_pad_0 = const()[name = tensor("y_1_pad_0"), val = tensor([0, 0])]; + tensor y_1_dilations_0 = const()[name = tensor("y_1_dilations_0"), val = tensor([1])]; + tensor y_1_has_output_shape_output_shape_0 = const()[name = tensor("y_1_has_output_shape_output_shape_0"), val = tensor([1, 512, 32])]; + tensor y_1_has_output_shape = conv_transpose(dilations = y_1_dilations_0, groups = y_1_groups_0, output_shape = y_1_has_output_shape_output_shape_0, pad = y_1_pad_0, pad_type = y_1_pad_type_0, strides = y_1_strides_0, weight = mimi_upsample_convtr_convtr_weight, x = x_1)[name = tensor("y_1_has_output_shape")]; + tensor var_72_begin_0 = const()[name = tensor("op_72_begin_0"), val = tensor([0, 0, 0])]; + tensor var_72_end_0 = const()[name = tensor("op_72_end_0"), val = tensor([1, 512, 16])]; + tensor var_72_end_mask_0 = const()[name = tensor("op_72_end_mask_0"), val = tensor([true, true, false])]; + tensor var_72 = slice_by_index(begin = var_72_begin_0, end = var_72_end_0, end_mask = var_72_end_mask_0, x = y_1_has_output_shape)[name = tensor("op_72")]; + tensor var_73 = add(x = var_72, y = upsample_partial)[name = tensor("op_73")]; + tensor var_74_begin_0 = const()[name = tensor("op_74_begin_0"), val = tensor([0, 0, 16])]; + tensor var_74_end_0 = const()[name = tensor("op_74_end_0"), val = tensor([1, 512, 32])]; + tensor var_74_end_mask_0 = const()[name = tensor("op_74_end_mask_0"), val = tensor([true, true, true])]; + tensor var_74 = slice_by_index(begin = var_74_begin_0, end = var_74_end_0, end_mask = var_74_end_mask_0, x = y_1_has_output_shape)[name = tensor("op_74")]; + tensor y_3_interleave_0 = const()[name = tensor("y_3_interleave_0"), val = tensor(false)]; + tensor y_3 = concat(axis = var_62, interleave = y_3_interleave_0, values = (var_73, var_74))[name = tensor("y_3")]; + tensor var_77_begin_0 = const()[name = tensor("op_77_begin_0"), val = tensor([0, 0, 16])]; + tensor var_77_end_0 = const()[name = tensor("op_77_end_0"), val = tensor([1, 512, 32])]; + tensor var_77_end_mask_0 = const()[name = tensor("op_77_end_mask_0"), val = tensor([true, true, true])]; + tensor var_77 = slice_by_index(begin = var_77_begin_0, end = var_77_end_0, end_mask = var_77_end_mask_0, x = y_3)[name = tensor("op_77")]; + tensor x_3_begin_0 = const()[name = tensor("x_3_begin_0"), val = tensor([0, 0, 0])]; + tensor x_3_end_0 = const()[name = tensor("x_3_end_0"), val = tensor([1, 512, 16])]; + tensor x_3_end_mask_0 = const()[name = tensor("x_3_end_mask_0"), val = tensor([true, true, false])]; + tensor x_3 = slice_by_index(begin = x_3_begin_0, end = x_3_end_0, end_mask = x_3_end_mask_0, x = y_3)[name = tensor("x_3")]; + tensor var_86 = const()[name = tensor("op_86"), val = tensor(0)]; + tensor var_91 = const()[name = tensor("op_91"), val = tensor(-1)]; + tensor var_100 = const()[name = tensor("op_100"), val = tensor(-0x1.ff933cp+127)]; + tensor var_102 = const()[name = tensor("op_102"), val = tensor(0x1.4f8b58p-17)]; + tensor input_3_perm_0 = const()[name = tensor("input_3_perm_0"), val = tensor([0, 2, 1])]; + tensor query_1_axes_0 = const()[name = tensor("query_1_axes_0"), val = tensor([-1])]; + tensor input_3 = transpose(perm = input_3_perm_0, x = x_3)[name = tensor("transpose_19")]; + tensor query_1 = layer_norm(axes = query_1_axes_0, beta = mimi_decoder_transformer_transformer_layers_0_norm1_bias, epsilon = var_102, gamma = mimi_decoder_transformer_transformer_layers_0_norm1_weight, x = input_3)[name = tensor("query_1")]; + tensor linear_0_bias_0 = const()[name = tensor("linear_0_bias_0"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(41224448)))]; + tensor projected_1 = linear(bias = linear_0_bias_0, weight = mimi_decoder_transformer_transformer_layers_0_self_attn_in_proj_weight, x = query_1)[name = tensor("linear_0")]; + tensor var_130 = const()[name = tensor("op_130"), val = tensor([1, 16, 3, 8, 64])]; + tensor packed_1 = reshape(shape = var_130, x = projected_1)[name = tensor("packed_1")]; + tensor var_132_split_sizes_0 = const()[name = tensor("op_132_split_sizes_0"), val = tensor([1, 1, 1])]; + tensor var_132_axis_0 = const()[name = tensor("op_132_axis_0"), val = tensor(2)]; + tensor var_132_0, tensor var_132_1, tensor var_132_2 = split(axis = var_132_axis_0, split_sizes = var_132_split_sizes_0, x = packed_1)[name = tensor("op_132")]; + tensor squeeze_0_axes_0 = const()[name = tensor("squeeze_0_axes_0"), val = tensor([2])]; + tensor squeeze_0 = squeeze(axes = squeeze_0_axes_0, x = var_132_0)[name = tensor("squeeze_0")]; + tensor squeeze_1_axes_0 = const()[name = tensor("squeeze_1_axes_0"), val = tensor([2])]; + tensor squeeze_1 = squeeze(axes = squeeze_1_axes_0, x = var_132_1)[name = tensor("squeeze_1")]; + tensor squeeze_2_axes_0 = const()[name = tensor("squeeze_2_axes_0"), val = tensor([2])]; + tensor squeeze_2 = squeeze(axes = squeeze_2_axes_0, x = var_132_2)[name = tensor("squeeze_2")]; + tensor offset_3_begin_0 = const()[name = tensor("offset_3_begin_0"), val = tensor([0])]; + tensor offset_3_end_0 = const()[name = tensor("offset_3_end_0"), val = tensor([1])]; + tensor offset_3_end_mask_0 = const()[name = tensor("offset_3_end_mask_0"), val = tensor([false])]; + tensor offset_3_squeeze_mask_0 = const()[name = tensor("offset_3_squeeze_mask_0"), val = tensor([true])]; + tensor offset_3 = slice_by_index(begin = offset_3_begin_0, end = offset_3_end_0, end_mask = offset_3_end_mask_0, squeeze_mask = offset_3_squeeze_mask_0, x = attn0_offset)[name = tensor("offset_3")]; + tensor freqs_1 = const()[name = tensor("freqs_1"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(41230656)))]; + tensor ts_1_promoted = const()[name = tensor("ts_1_promoted"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(41230848)))]; + tensor ts_3 = add(x = ts_1_promoted, y = offset_3)[name = tensor("ts_3")]; + tensor var_148 = const()[name = tensor("op_148"), val = tensor([-1, 1, 1])]; + tensor ts_5 = reshape(shape = var_148, x = ts_3)[name = tensor("ts_5")]; + tensor var_150 = const()[name = tensor("op_150"), val = tensor([1, 16, 8, 32, 2])]; + tensor q_3 = reshape(shape = var_150, x = squeeze_0)[name = tensor("q_3")]; + tensor var_152 = const()[name = tensor("op_152"), val = tensor([1, 16, 8, 32, 2])]; + tensor k_3 = reshape(shape = var_152, x = squeeze_1)[name = tensor("k_3")]; + tensor var_154_begin_0 = const()[name = tensor("op_154_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_154_end_0 = const()[name = tensor("op_154_end_0"), val = tensor([1, 16, 8, 32, 1])]; + tensor var_154_end_mask_0 = const()[name = tensor("op_154_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_154_squeeze_mask_0 = const()[name = tensor("op_154_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_154 = slice_by_index(begin = var_154_begin_0, end = var_154_end_0, end_mask = var_154_end_mask_0, squeeze_mask = var_154_squeeze_mask_0, x = q_3)[name = tensor("op_154")]; + tensor var_156_begin_0 = const()[name = tensor("op_156_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_156_end_0 = const()[name = tensor("op_156_end_0"), val = tensor([1, 16, 8, 32, 2])]; + tensor var_156_end_mask_0 = const()[name = tensor("op_156_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_156_squeeze_mask_0 = const()[name = tensor("op_156_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_156 = slice_by_index(begin = var_156_begin_0, end = var_156_end_0, end_mask = var_156_end_mask_0, squeeze_mask = var_156_squeeze_mask_0, x = q_3)[name = tensor("op_156")]; + tensor var_158_begin_0 = const()[name = tensor("op_158_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_158_end_0 = const()[name = tensor("op_158_end_0"), val = tensor([1, 16, 8, 32, 1])]; + tensor var_158_end_mask_0 = const()[name = tensor("op_158_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_158_squeeze_mask_0 = const()[name = tensor("op_158_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_158 = slice_by_index(begin = var_158_begin_0, end = var_158_end_0, end_mask = var_158_end_mask_0, squeeze_mask = var_158_squeeze_mask_0, x = k_3)[name = tensor("op_158")]; + tensor var_160_begin_0 = const()[name = tensor("op_160_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_160_end_0 = const()[name = tensor("op_160_end_0"), val = tensor([1, 16, 8, 32, 2])]; + tensor var_160_end_mask_0 = const()[name = tensor("op_160_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_160_squeeze_mask_0 = const()[name = tensor("op_160_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_160 = slice_by_index(begin = var_160_begin_0, end = var_160_end_0, end_mask = var_160_end_mask_0, squeeze_mask = var_160_squeeze_mask_0, x = k_3)[name = tensor("op_160")]; + tensor var_162 = mul(x = freqs_1, y = ts_5)[name = tensor("op_162")]; + tensor rotr_1 = cos(x = var_162)[name = tensor("rotr_1")]; + tensor roti_1 = sin(x = var_162)[name = tensor("roti_1")]; + tensor var_166 = mul(x = var_154, y = rotr_1)[name = tensor("op_166")]; + tensor var_167 = mul(x = var_156, y = roti_1)[name = tensor("op_167")]; + tensor qor_1 = sub(x = var_166, y = var_167)[name = tensor("qor_1")]; + tensor var_169 = mul(x = var_154, y = roti_1)[name = tensor("op_169")]; + tensor var_170 = mul(x = var_156, y = rotr_1)[name = tensor("op_170")]; + tensor qoi_1 = add(x = var_169, y = var_170)[name = tensor("qoi_1")]; + tensor var_172 = mul(x = var_158, y = rotr_1)[name = tensor("op_172")]; + tensor var_173 = mul(x = var_160, y = roti_1)[name = tensor("op_173")]; + tensor kor_1 = sub(x = var_172, y = var_173)[name = tensor("kor_1")]; + tensor var_175 = mul(x = var_158, y = roti_1)[name = tensor("op_175")]; + tensor var_176 = mul(x = var_160, y = rotr_1)[name = tensor("op_176")]; + tensor koi_1 = add(x = var_175, y = var_176)[name = tensor("koi_1")]; + tensor qo_1_axis_0 = const()[name = tensor("qo_1_axis_0"), val = tensor(-1)]; + tensor qo_1 = stack(axis = qo_1_axis_0, values = (qor_1, qoi_1))[name = tensor("qo_1")]; + tensor ko_1_axis_0 = const()[name = tensor("ko_1_axis_0"), val = tensor(-1)]; + tensor ko_1 = stack(axis = ko_1_axis_0, values = (kor_1, koi_1))[name = tensor("ko_1")]; + tensor var_186 = const()[name = tensor("op_186"), val = tensor([1, 16, 8, 64])]; + tensor q_5 = reshape(shape = var_186, x = qo_1)[name = tensor("q_5")]; + tensor var_188 = const()[name = tensor("op_188"), val = tensor([1, 16, 8, 64])]; + tensor k_5 = reshape(shape = var_188, x = ko_1)[name = tensor("k_5")]; + tensor capacity_1 = const()[name = tensor("capacity_1"), val = tensor([256])]; + tensor var_193_dtype_0 = const()[name = tensor("op_193_dtype_0"), val = tensor("int32")]; + tensor var_194 = const()[name = tensor("op_194"), val = tensor([1, 1])]; + tensor var_193 = cast(dtype = var_193_dtype_0, x = attn0_offset)[name = tensor("cast_49")]; + tensor write_base_1 = reshape(shape = var_194, x = var_193)[name = tensor("write_base_1")]; + tensor write_range_1 = const()[name = tensor("write_range_1"), val = tensor([[0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15]])]; + tensor abs_idx_1 = add(x = write_base_1, y = write_range_1)[name = tensor("abs_idx_1")]; + tensor wrapped_1_div = floor_div(x = abs_idx_1, y = capacity_1)[name = tensor("wrapped_1_div")]; + tensor wrapped_1_div_scaled = mul(x = wrapped_1_div, y = capacity_1)[name = tensor("wrapped_1_div_scaled")]; + tensor wrapped_1 = sub(x = abs_idx_1, y = wrapped_1_div_scaled)[name = tensor("wrapped_1")]; + tensor var_201 = const()[name = tensor("op_201"), val = tensor([1, 16, 1, 1])]; + tensor var_202 = reshape(shape = var_201, x = wrapped_1)[name = tensor("op_202")]; + tensor write_indexes_1_reps_0 = const()[name = tensor("write_indexes_1_reps_0"), val = tensor([1, 1, 8, 64])]; + tensor write_indexes_1 = tile(reps = write_indexes_1_reps_0, x = var_202)[name = tensor("write_indexes_1")]; + tensor var_205_begin_0 = const()[name = tensor("op_205_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_205_end_0 = const()[name = tensor("op_205_end_0"), val = tensor([1, 1, 256, 8, 64])]; + tensor var_205_end_mask_0 = const()[name = tensor("op_205_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_205_squeeze_mask_0 = const()[name = tensor("op_205_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor var_205 = slice_by_index(begin = var_205_begin_0, end = var_205_end_0, end_mask = var_205_end_mask_0, squeeze_mask = var_205_squeeze_mask_0, x = attn0_cache)[name = tensor("op_205")]; + tensor new_k_cache_1_axis_0 = const()[name = tensor("new_k_cache_1_axis_0"), val = tensor(1)]; + tensor new_k_cache_1_mode_0 = const()[name = tensor("new_k_cache_1_mode_0"), val = tensor("update")]; + tensor new_k_cache_1_validate_indices_0 = const()[name = tensor("new_k_cache_1_validate_indices_0"), val = tensor(false)]; + tensor new_k_cache_1 = scatter_along_axis(axis = new_k_cache_1_axis_0, data = var_205, indices = write_indexes_1, mode = new_k_cache_1_mode_0, updates = k_5, validate_indices = new_k_cache_1_validate_indices_0)[name = tensor("new_k_cache_1")]; + tensor var_207_begin_0 = const()[name = tensor("op_207_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor var_207_end_0 = const()[name = tensor("op_207_end_0"), val = tensor([2, 1, 256, 8, 64])]; + tensor var_207_end_mask_0 = const()[name = tensor("op_207_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_207_squeeze_mask_0 = const()[name = tensor("op_207_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor var_207 = slice_by_index(begin = var_207_begin_0, end = var_207_end_0, end_mask = var_207_end_mask_0, squeeze_mask = var_207_squeeze_mask_0, x = attn0_cache)[name = tensor("op_207")]; + tensor new_v_cache_1_axis_0 = const()[name = tensor("new_v_cache_1_axis_0"), val = tensor(1)]; + tensor new_v_cache_1_mode_0 = const()[name = tensor("new_v_cache_1_mode_0"), val = tensor("update")]; + tensor new_v_cache_1_validate_indices_0 = const()[name = tensor("new_v_cache_1_validate_indices_0"), val = tensor(false)]; + tensor new_v_cache_1 = scatter_along_axis(axis = new_v_cache_1_axis_0, data = var_207, indices = write_indexes_1, mode = new_v_cache_1_mode_0, updates = squeeze_2, validate_indices = new_v_cache_1_validate_indices_0)[name = tensor("new_v_cache_1")]; + tensor var_210_axis_0 = const()[name = tensor("op_210_axis_0"), val = tensor(0)]; + tensor var_210 = stack(axis = var_210_axis_0, values = (new_k_cache_1, new_v_cache_1))[name = tensor("op_210")]; + tensor var_211 = not_equal(x = new_k_cache_1, y = new_k_cache_1)[name = tensor("op_211")]; + tensor var_212 = const()[name = tensor("op_212"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(41230976)))]; + tensor new_k_cache_3 = select(a = var_212, b = new_k_cache_1, cond = var_211)[name = tensor("new_k_cache_3")]; + tensor var_214 = not_equal(x = new_v_cache_1, y = new_v_cache_1)[name = tensor("op_214")]; + tensor new_v_cache_3 = select(a = var_212, b = new_v_cache_1, cond = var_214)[name = tensor("new_v_cache_3")]; + tensor var_219 = const()[name = tensor("op_219"), val = tensor([0, 2, 1, 3])]; + tensor var_221 = const()[name = tensor("op_221"), val = tensor([1, 1])]; + tensor var_222 = reshape(shape = var_221, x = attn0_offset)[name = tensor("op_222")]; + tensor var_224_promoted = const()[name = tensor("op_224_promoted"), val = tensor([0x1.ep+3])]; + tensor var_225 = add(x = var_222, y = var_224_promoted)[name = tensor("op_225")]; + tensor last_pos_1_dtype_0 = const()[name = tensor("last_pos_1_dtype_0"), val = tensor("int32")]; + tensor slot_idx_1 = const()[name = tensor("slot_idx_1"), val = tensor([[0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19, 20, 21, 22, 23, 24, 25, 26, 27, 28, 29, 30, 31, 32, 33, 34, 35, 36, 37, 38, 39, 40, 41, 42, 43, 44, 45, 46, 47, 48, 49, 50, 51, 52, 53, 54, 55, 56, 57, 58, 59, 60, 61, 62, 63, 64, 65, 66, 67, 68, 69, 70, 71, 72, 73, 74, 75, 76, 77, 78, 79, 80, 81, 82, 83, 84, 85, 86, 87, 88, 89, 90, 91, 92, 93, 94, 95, 96, 97, 98, 99, 100, 101, 102, 103, 104, 105, 106, 107, 108, 109, 110, 111, 112, 113, 114, 115, 116, 117, 118, 119, 120, 121, 122, 123, 124, 125, 126, 127, 128, 129, 130, 131, 132, 133, 134, 135, 136, 137, 138, 139, 140, 141, 142, 143, 144, 145, 146, 147, 148, 149, 150, 151, 152, 153, 154, 155, 156, 157, 158, 159, 160, 161, 162, 163, 164, 165, 166, 167, 168, 169, 170, 171, 172, 173, 174, 175, 176, 177, 178, 179, 180, 181, 182, 183, 184, 185, 186, 187, 188, 189, 190, 191, 192, 193, 194, 195, 196, 197, 198, 199, 200, 201, 202, 203, 204, 205, 206, 207, 208, 209, 210, 211, 212, 213, 214, 215, 216, 217, 218, 219, 220, 221, 222, 223, 224, 225, 226, 227, 228, 229, 230, 231, 232, 233, 234, 235, 236, 237, 238, 239, 240, 241, 242, 243, 244, 245, 246, 247, 248, 249, 250, 251, 252, 253, 254, 255]])]; + tensor last_pos_1 = cast(dtype = last_pos_1_dtype_0, x = var_225)[name = tensor("cast_48")]; + tensor diff_1 = sub(x = last_pos_1, y = slot_idx_1)[name = tensor("diff_1")]; + tensor var_231_div = floor_div(x = diff_1, y = capacity_1)[name = tensor("op_231_div")]; + tensor var_231_div_scaled = mul(x = var_231_div, y = capacity_1)[name = tensor("op_231_div_scaled")]; + tensor var_231 = sub(x = diff_1, y = var_231_div_scaled)[name = tensor("op_231")]; + tensor pos_k_1 = sub(x = last_pos_1, y = var_231)[name = tensor("pos_k_1")]; + tensor var_237_promoted = const()[name = tensor("op_237_promoted"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(41755328)))]; + tensor pos_q_1 = add(x = var_222, y = var_237_promoted)[name = tensor("pos_q_1")]; + tensor var_241_axes_0 = const()[name = tensor("op_241_axes_0"), val = tensor([2])]; + tensor var_241 = expand_dims(axes = var_241_axes_0, x = pos_q_1)[name = tensor("op_241")]; + tensor var_243_axes_0 = const()[name = tensor("op_243_axes_0"), val = tensor([1])]; + tensor var_243 = expand_dims(axes = var_243_axes_0, x = pos_k_1)[name = tensor("op_243")]; + tensor var_244_promoted_dtype_0 = const()[name = tensor("op_244_promoted_dtype_0"), val = tensor("fp32")]; + tensor var_244_promoted = cast(dtype = var_244_promoted_dtype_0, x = var_243)[name = tensor("cast_47")]; + tensor delta_1 = sub(x = var_241, y = var_244_promoted)[name = tensor("delta_1")]; + tensor valid_1 = greater_equal(x = var_243, y = var_86)[name = tensor("valid_1")]; + tensor var_253 = const()[name = tensor("op_253"), val = tensor([1, 1, 1])]; + tensor var_254 = reshape(shape = var_253, x = attn0_offset)[name = tensor("op_254")]; + tensor var_256_promoted = const()[name = tensor("op_256_promoted"), val = tensor([0x1.ep+3])]; + tensor var_257 = add(x = var_254, y = var_256_promoted)[name = tensor("op_257")]; + tensor var_258 = less_equal(x = var_244_promoted, y = var_257)[name = tensor("op_258")]; + tensor valid_3 = logical_and(x = valid_1, y = var_258)[name = tensor("valid_3")]; + tensor var_86_promoted = const()[name = tensor("op_86_promoted"), val = tensor(0x0p+0)]; + tensor var_260 = greater_equal(x = delta_1, y = var_86_promoted)[name = tensor("op_260")]; + tensor attn_mask_1 = logical_and(x = valid_3, y = var_260)[name = tensor("attn_mask_1")]; + tensor var_98_promoted = const()[name = tensor("op_98_promoted"), val = tensor(0x1.f4p+7)]; + tensor var_262 = less(x = delta_1, y = var_98_promoted)[name = tensor("op_262")]; + tensor attn_mask_3 = logical_and(x = attn_mask_1, y = var_262)[name = tensor("attn_mask_3")]; + tensor attn_mask_5_axes_0 = const()[name = tensor("attn_mask_5_axes_0"), val = tensor([1])]; + tensor attn_mask_5 = expand_dims(axes = attn_mask_5_axes_0, x = attn_mask_3)[name = tensor("attn_mask_5")]; + tensor var_267_transpose_x_0 = const()[name = tensor("op_267_transpose_x_0"), val = tensor(false)]; + tensor var_267_transpose_y_0 = const()[name = tensor("op_267_transpose_y_0"), val = tensor(false)]; + tensor transpose_6_perm_0 = const()[name = tensor("transpose_6_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_7_perm_0 = const()[name = tensor("transpose_7_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_7 = transpose(perm = transpose_7_perm_0, x = new_k_cache_3)[name = tensor("transpose_16")]; + tensor transpose_6 = transpose(perm = transpose_6_perm_0, x = q_5)[name = tensor("transpose_17")]; + tensor var_267 = matmul(transpose_x = var_267_transpose_x_0, transpose_y = var_267_transpose_y_0, x = transpose_6, y = transpose_7)[name = tensor("op_267")]; + tensor var_268 = const()[name = tensor("op_268"), val = tensor(0x1p-3)]; + tensor attn_1 = mul(x = var_267, y = var_268)[name = tensor("attn_1")]; + tensor var_270 = logical_not(x = attn_mask_5)[name = tensor("op_270")]; + tensor attn_3 = select(a = var_100, b = attn_1, cond = var_270)[name = tensor("attn_3")]; + tensor attn_5 = softmax(axis = var_91, x = attn_3)[name = tensor("attn_5")]; + tensor x_5_transpose_x_0 = const()[name = tensor("x_5_transpose_x_0"), val = tensor(false)]; + tensor x_5_transpose_y_0 = const()[name = tensor("x_5_transpose_y_0"), val = tensor(false)]; + tensor v_attn_1 = transpose(perm = var_219, x = new_v_cache_3)[name = tensor("transpose_18")]; + tensor x_5 = matmul(transpose_x = x_5_transpose_x_0, transpose_y = x_5_transpose_y_0, x = attn_5, y = v_attn_1)[name = tensor("x_5")]; + tensor var_274_perm_0 = const()[name = tensor("op_274_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_275 = const()[name = tensor("op_275"), val = tensor([1, 16, 512])]; + tensor var_274 = transpose(perm = var_274_perm_0, x = x_5)[name = tensor("transpose_15")]; + tensor input_5 = reshape(shape = var_275, x = var_274)[name = tensor("input_5")]; + tensor linear_1_bias_0 = const()[name = tensor("linear_1_bias_0"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(41755456)))]; + tensor x_7 = linear(bias = linear_1_bias_0, weight = mimi_decoder_transformer_transformer_layers_0_self_attn_out_proj_weight, x = input_5)[name = tensor("linear_1")]; + tensor var_284 = mul(x = mimi_decoder_transformer_transformer_layers_0_layer_scale_1_scale, y = x_7)[name = tensor("op_284")]; + tensor input_7 = add(x = input_3, y = var_284)[name = tensor("input_7")]; + tensor input_9_axes_0 = const()[name = tensor("input_9_axes_0"), val = tensor([-1])]; + tensor input_9 = layer_norm(axes = input_9_axes_0, beta = mimi_decoder_transformer_transformer_layers_0_norm2_bias, epsilon = var_102, gamma = mimi_decoder_transformer_transformer_layers_0_norm2_weight, x = input_7)[name = tensor("input_9")]; + tensor linear_2_bias_0 = const()[name = tensor("linear_2_bias_0"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(41757568)))]; + tensor var_291 = linear(bias = linear_2_bias_0, weight = mimi_decoder_transformer_transformer_layers_0_linear1_weight, x = input_9)[name = tensor("linear_2")]; + tensor input_11_mode_0 = const()[name = tensor("input_11_mode_0"), val = tensor("EXACT")]; + tensor input_11 = gelu(mode = input_11_mode_0, x = var_291)[name = tensor("input_11")]; + tensor x_9 = linear(bias = linear_1_bias_0, weight = mimi_decoder_transformer_transformer_layers_0_linear2_weight, x = input_11)[name = tensor("linear_3")]; + tensor var_297 = mul(x = mimi_decoder_transformer_transformer_layers_0_layer_scale_2_scale, y = x_9)[name = tensor("op_297")]; + tensor input_13 = add(x = input_7, y = var_297)[name = tensor("input_13")]; + tensor query_axes_0 = const()[name = tensor("query_axes_0"), val = tensor([-1])]; + tensor query = layer_norm(axes = query_axes_0, beta = mimi_decoder_transformer_transformer_layers_1_norm1_bias, epsilon = var_102, gamma = mimi_decoder_transformer_transformer_layers_1_norm1_weight, x = input_13)[name = tensor("query")]; + tensor projected = linear(bias = linear_0_bias_0, weight = mimi_decoder_transformer_transformer_layers_1_self_attn_in_proj_weight, x = query)[name = tensor("linear_4")]; + tensor var_320 = const()[name = tensor("op_320"), val = tensor([1, 16, 3, 8, 64])]; + tensor packed = reshape(shape = var_320, x = projected)[name = tensor("packed")]; + tensor var_322_split_sizes_0 = const()[name = tensor("op_322_split_sizes_0"), val = tensor([1, 1, 1])]; + tensor var_322_axis_0 = const()[name = tensor("op_322_axis_0"), val = tensor(2)]; + tensor var_322_0, tensor var_322_1, tensor var_322_2 = split(axis = var_322_axis_0, split_sizes = var_322_split_sizes_0, x = packed)[name = tensor("op_322")]; + tensor squeeze_3_axes_0 = const()[name = tensor("squeeze_3_axes_0"), val = tensor([2])]; + tensor squeeze_3 = squeeze(axes = squeeze_3_axes_0, x = var_322_0)[name = tensor("squeeze_3")]; + tensor squeeze_4_axes_0 = const()[name = tensor("squeeze_4_axes_0"), val = tensor([2])]; + tensor squeeze_4 = squeeze(axes = squeeze_4_axes_0, x = var_322_1)[name = tensor("squeeze_4")]; + tensor squeeze_5_axes_0 = const()[name = tensor("squeeze_5_axes_0"), val = tensor([2])]; + tensor squeeze_5 = squeeze(axes = squeeze_5_axes_0, x = var_322_2)[name = tensor("squeeze_5")]; + tensor offset_begin_0 = const()[name = tensor("offset_begin_0"), val = tensor([0])]; + tensor offset_end_0 = const()[name = tensor("offset_end_0"), val = tensor([1])]; + tensor offset_end_mask_0 = const()[name = tensor("offset_end_mask_0"), val = tensor([false])]; + tensor offset_squeeze_mask_0 = const()[name = tensor("offset_squeeze_mask_0"), val = tensor([true])]; + tensor offset = slice_by_index(begin = offset_begin_0, end = offset_end_0, end_mask = offset_end_mask_0, squeeze_mask = offset_squeeze_mask_0, x = attn1_offset)[name = tensor("offset")]; + tensor freqs = const()[name = tensor("freqs"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(41765824)))]; + tensor ts_7_promoted = const()[name = tensor("ts_7_promoted"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(41766016)))]; + tensor ts_9 = add(x = ts_7_promoted, y = offset)[name = tensor("ts_9")]; + tensor var_338 = const()[name = tensor("op_338"), val = tensor([-1, 1, 1])]; + tensor ts = reshape(shape = var_338, x = ts_9)[name = tensor("ts")]; + tensor var_340 = const()[name = tensor("op_340"), val = tensor([1, 16, 8, 32, 2])]; + tensor q_9 = reshape(shape = var_340, x = squeeze_3)[name = tensor("q_9")]; + tensor var_342 = const()[name = tensor("op_342"), val = tensor([1, 16, 8, 32, 2])]; + tensor k_9 = reshape(shape = var_342, x = squeeze_4)[name = tensor("k_9")]; + tensor var_344_begin_0 = const()[name = tensor("op_344_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_344_end_0 = const()[name = tensor("op_344_end_0"), val = tensor([1, 16, 8, 32, 1])]; + tensor var_344_end_mask_0 = const()[name = tensor("op_344_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_344_squeeze_mask_0 = const()[name = tensor("op_344_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_344 = slice_by_index(begin = var_344_begin_0, end = var_344_end_0, end_mask = var_344_end_mask_0, squeeze_mask = var_344_squeeze_mask_0, x = q_9)[name = tensor("op_344")]; + tensor var_346_begin_0 = const()[name = tensor("op_346_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_346_end_0 = const()[name = tensor("op_346_end_0"), val = tensor([1, 16, 8, 32, 2])]; + tensor var_346_end_mask_0 = const()[name = tensor("op_346_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_346_squeeze_mask_0 = const()[name = tensor("op_346_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_346 = slice_by_index(begin = var_346_begin_0, end = var_346_end_0, end_mask = var_346_end_mask_0, squeeze_mask = var_346_squeeze_mask_0, x = q_9)[name = tensor("op_346")]; + tensor var_348_begin_0 = const()[name = tensor("op_348_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_348_end_0 = const()[name = tensor("op_348_end_0"), val = tensor([1, 16, 8, 32, 1])]; + tensor var_348_end_mask_0 = const()[name = tensor("op_348_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_348_squeeze_mask_0 = const()[name = tensor("op_348_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_348 = slice_by_index(begin = var_348_begin_0, end = var_348_end_0, end_mask = var_348_end_mask_0, squeeze_mask = var_348_squeeze_mask_0, x = k_9)[name = tensor("op_348")]; + tensor var_350_begin_0 = const()[name = tensor("op_350_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_350_end_0 = const()[name = tensor("op_350_end_0"), val = tensor([1, 16, 8, 32, 2])]; + tensor var_350_end_mask_0 = const()[name = tensor("op_350_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_350_squeeze_mask_0 = const()[name = tensor("op_350_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_350 = slice_by_index(begin = var_350_begin_0, end = var_350_end_0, end_mask = var_350_end_mask_0, squeeze_mask = var_350_squeeze_mask_0, x = k_9)[name = tensor("op_350")]; + tensor var_352 = mul(x = freqs, y = ts)[name = tensor("op_352")]; + tensor rotr = cos(x = var_352)[name = tensor("rotr")]; + tensor roti = sin(x = var_352)[name = tensor("roti")]; + tensor var_356 = mul(x = var_344, y = rotr)[name = tensor("op_356")]; + tensor var_357 = mul(x = var_346, y = roti)[name = tensor("op_357")]; + tensor qor_5 = sub(x = var_356, y = var_357)[name = tensor("qor_5")]; + tensor var_359 = mul(x = var_344, y = roti)[name = tensor("op_359")]; + tensor var_360 = mul(x = var_346, y = rotr)[name = tensor("op_360")]; + tensor qoi_5 = add(x = var_359, y = var_360)[name = tensor("qoi_5")]; + tensor var_362 = mul(x = var_348, y = rotr)[name = tensor("op_362")]; + tensor var_363 = mul(x = var_350, y = roti)[name = tensor("op_363")]; + tensor kor_5 = sub(x = var_362, y = var_363)[name = tensor("kor_5")]; + tensor var_365 = mul(x = var_348, y = roti)[name = tensor("op_365")]; + tensor var_366 = mul(x = var_350, y = rotr)[name = tensor("op_366")]; + tensor koi_5 = add(x = var_365, y = var_366)[name = tensor("koi_5")]; + tensor qo_axis_0 = const()[name = tensor("qo_axis_0"), val = tensor(-1)]; + tensor qo = stack(axis = qo_axis_0, values = (qor_5, qoi_5))[name = tensor("qo")]; + tensor ko_axis_0 = const()[name = tensor("ko_axis_0"), val = tensor(-1)]; + tensor ko = stack(axis = ko_axis_0, values = (kor_5, koi_5))[name = tensor("ko")]; + tensor var_376 = const()[name = tensor("op_376"), val = tensor([1, 16, 8, 64])]; + tensor q = reshape(shape = var_376, x = qo)[name = tensor("q")]; + tensor var_378 = const()[name = tensor("op_378"), val = tensor([1, 16, 8, 64])]; + tensor k = reshape(shape = var_378, x = ko)[name = tensor("k")]; + tensor capacity = const()[name = tensor("capacity"), val = tensor([256])]; + tensor var_383_dtype_0 = const()[name = tensor("op_383_dtype_0"), val = tensor("int32")]; + tensor var_384 = const()[name = tensor("op_384"), val = tensor([1, 1])]; + tensor var_383 = cast(dtype = var_383_dtype_0, x = attn1_offset)[name = tensor("cast_46")]; + tensor write_base = reshape(shape = var_384, x = var_383)[name = tensor("write_base")]; + tensor write_range = const()[name = tensor("write_range"), val = tensor([[0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15]])]; + tensor abs_idx = add(x = write_base, y = write_range)[name = tensor("abs_idx")]; + tensor wrapped_div = floor_div(x = abs_idx, y = capacity)[name = tensor("wrapped_div")]; + tensor wrapped_div_scaled = mul(x = wrapped_div, y = capacity)[name = tensor("wrapped_div_scaled")]; + tensor wrapped = sub(x = abs_idx, y = wrapped_div_scaled)[name = tensor("wrapped")]; + tensor var_391 = const()[name = tensor("op_391"), val = tensor([1, 16, 1, 1])]; + tensor var_392 = reshape(shape = var_391, x = wrapped)[name = tensor("op_392")]; + tensor write_indexes_reps_0 = const()[name = tensor("write_indexes_reps_0"), val = tensor([1, 1, 8, 64])]; + tensor write_indexes = tile(reps = write_indexes_reps_0, x = var_392)[name = tensor("write_indexes")]; + tensor var_395_begin_0 = const()[name = tensor("op_395_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_395_end_0 = const()[name = tensor("op_395_end_0"), val = tensor([1, 1, 256, 8, 64])]; + tensor var_395_end_mask_0 = const()[name = tensor("op_395_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_395_squeeze_mask_0 = const()[name = tensor("op_395_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor var_395 = slice_by_index(begin = var_395_begin_0, end = var_395_end_0, end_mask = var_395_end_mask_0, squeeze_mask = var_395_squeeze_mask_0, x = attn1_cache)[name = tensor("op_395")]; + tensor new_k_cache_5_axis_0 = const()[name = tensor("new_k_cache_5_axis_0"), val = tensor(1)]; + tensor new_k_cache_5_mode_0 = const()[name = tensor("new_k_cache_5_mode_0"), val = tensor("update")]; + tensor new_k_cache_5_validate_indices_0 = const()[name = tensor("new_k_cache_5_validate_indices_0"), val = tensor(false)]; + tensor new_k_cache_5 = scatter_along_axis(axis = new_k_cache_5_axis_0, data = var_395, indices = write_indexes, mode = new_k_cache_5_mode_0, updates = k, validate_indices = new_k_cache_5_validate_indices_0)[name = tensor("new_k_cache_5")]; + tensor var_397_begin_0 = const()[name = tensor("op_397_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor var_397_end_0 = const()[name = tensor("op_397_end_0"), val = tensor([2, 1, 256, 8, 64])]; + tensor var_397_end_mask_0 = const()[name = tensor("op_397_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_397_squeeze_mask_0 = const()[name = tensor("op_397_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor var_397 = slice_by_index(begin = var_397_begin_0, end = var_397_end_0, end_mask = var_397_end_mask_0, squeeze_mask = var_397_squeeze_mask_0, x = attn1_cache)[name = tensor("op_397")]; + tensor new_v_cache_5_axis_0 = const()[name = tensor("new_v_cache_5_axis_0"), val = tensor(1)]; + tensor new_v_cache_5_mode_0 = const()[name = tensor("new_v_cache_5_mode_0"), val = tensor("update")]; + tensor new_v_cache_5_validate_indices_0 = const()[name = tensor("new_v_cache_5_validate_indices_0"), val = tensor(false)]; + tensor new_v_cache_5 = scatter_along_axis(axis = new_v_cache_5_axis_0, data = var_397, indices = write_indexes, mode = new_v_cache_5_mode_0, updates = squeeze_5, validate_indices = new_v_cache_5_validate_indices_0)[name = tensor("new_v_cache_5")]; + tensor var_400_axis_0 = const()[name = tensor("op_400_axis_0"), val = tensor(0)]; + tensor var_400 = stack(axis = var_400_axis_0, values = (new_k_cache_5, new_v_cache_5))[name = tensor("op_400")]; + tensor var_401 = not_equal(x = new_k_cache_5, y = new_k_cache_5)[name = tensor("op_401")]; + tensor new_k_cache = select(a = var_212, b = new_k_cache_5, cond = var_401)[name = tensor("new_k_cache")]; + tensor var_404 = not_equal(x = new_v_cache_5, y = new_v_cache_5)[name = tensor("op_404")]; + tensor new_v_cache = select(a = var_212, b = new_v_cache_5, cond = var_404)[name = tensor("new_v_cache")]; + tensor var_409 = const()[name = tensor("op_409"), val = tensor([0, 2, 1, 3])]; + tensor var_411 = const()[name = tensor("op_411"), val = tensor([1, 1])]; + tensor var_412 = reshape(shape = var_411, x = attn1_offset)[name = tensor("op_412")]; + tensor var_414_promoted = const()[name = tensor("op_414_promoted"), val = tensor([0x1.ep+3])]; + tensor var_415 = add(x = var_412, y = var_414_promoted)[name = tensor("op_415")]; + tensor last_pos_dtype_0 = const()[name = tensor("last_pos_dtype_0"), val = tensor("int32")]; + tensor last_pos = cast(dtype = last_pos_dtype_0, x = var_415)[name = tensor("cast_45")]; + tensor diff = sub(x = last_pos, y = slot_idx_1)[name = tensor("diff")]; + tensor var_421_div = floor_div(x = diff, y = capacity)[name = tensor("op_421_div")]; + tensor var_421_div_scaled = mul(x = var_421_div, y = capacity)[name = tensor("op_421_div_scaled")]; + tensor var_421 = sub(x = diff, y = var_421_div_scaled)[name = tensor("op_421")]; + tensor pos_k = sub(x = last_pos, y = var_421)[name = tensor("pos_k")]; + tensor var_427_promoted = const()[name = tensor("op_427_promoted"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(41766144)))]; + tensor pos_q = add(x = var_412, y = var_427_promoted)[name = tensor("pos_q")]; + tensor var_431_axes_0 = const()[name = tensor("op_431_axes_0"), val = tensor([2])]; + tensor var_431 = expand_dims(axes = var_431_axes_0, x = pos_q)[name = tensor("op_431")]; + tensor var_433_axes_0 = const()[name = tensor("op_433_axes_0"), val = tensor([1])]; + tensor var_433 = expand_dims(axes = var_433_axes_0, x = pos_k)[name = tensor("op_433")]; + tensor var_434_promoted_dtype_0 = const()[name = tensor("op_434_promoted_dtype_0"), val = tensor("fp32")]; + tensor var_434_promoted = cast(dtype = var_434_promoted_dtype_0, x = var_433)[name = tensor("cast_44")]; + tensor delta = sub(x = var_431, y = var_434_promoted)[name = tensor("delta")]; + tensor valid_5 = greater_equal(x = var_433, y = var_86)[name = tensor("valid_5")]; + tensor var_443 = const()[name = tensor("op_443"), val = tensor([1, 1, 1])]; + tensor var_444 = reshape(shape = var_443, x = attn1_offset)[name = tensor("op_444")]; + tensor var_446_promoted = const()[name = tensor("op_446_promoted"), val = tensor([0x1.ep+3])]; + tensor var_447 = add(x = var_444, y = var_446_promoted)[name = tensor("op_447")]; + tensor var_448 = less_equal(x = var_434_promoted, y = var_447)[name = tensor("op_448")]; + tensor valid = logical_and(x = valid_5, y = var_448)[name = tensor("valid")]; + tensor var_86_promoted_1 = const()[name = tensor("op_86_promoted_1"), val = tensor(0x0p+0)]; + tensor var_450 = greater_equal(x = delta, y = var_86_promoted_1)[name = tensor("op_450")]; + tensor attn_mask_7 = logical_and(x = valid, y = var_450)[name = tensor("attn_mask_7")]; + tensor var_98_promoted_1 = const()[name = tensor("op_98_promoted_1"), val = tensor(0x1.f4p+7)]; + tensor var_452 = less(x = delta, y = var_98_promoted_1)[name = tensor("op_452")]; + tensor attn_mask_9 = logical_and(x = attn_mask_7, y = var_452)[name = tensor("attn_mask_9")]; + tensor attn_mask_axes_0 = const()[name = tensor("attn_mask_axes_0"), val = tensor([1])]; + tensor attn_mask = expand_dims(axes = attn_mask_axes_0, x = attn_mask_9)[name = tensor("attn_mask")]; + tensor var_457_transpose_x_0 = const()[name = tensor("op_457_transpose_x_0"), val = tensor(false)]; + tensor var_457_transpose_y_0 = const()[name = tensor("op_457_transpose_y_0"), val = tensor(false)]; + tensor transpose_8_perm_0 = const()[name = tensor("transpose_8_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_9_perm_0 = const()[name = tensor("transpose_9_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_9 = transpose(perm = transpose_9_perm_0, x = new_k_cache)[name = tensor("transpose_12")]; + tensor transpose_8 = transpose(perm = transpose_8_perm_0, x = q)[name = tensor("transpose_13")]; + tensor var_457 = matmul(transpose_x = var_457_transpose_x_0, transpose_y = var_457_transpose_y_0, x = transpose_8, y = transpose_9)[name = tensor("op_457")]; + tensor var_458 = const()[name = tensor("op_458"), val = tensor(0x1p-3)]; + tensor attn_7 = mul(x = var_457, y = var_458)[name = tensor("attn_7")]; + tensor var_460 = logical_not(x = attn_mask)[name = tensor("op_460")]; + tensor attn_9 = select(a = var_100, b = attn_7, cond = var_460)[name = tensor("attn_9")]; + tensor attn = softmax(axis = var_91, x = attn_9)[name = tensor("attn")]; + tensor x_11_transpose_x_0 = const()[name = tensor("x_11_transpose_x_0"), val = tensor(false)]; + tensor x_11_transpose_y_0 = const()[name = tensor("x_11_transpose_y_0"), val = tensor(false)]; + tensor v_attn = transpose(perm = var_409, x = new_v_cache)[name = tensor("transpose_14")]; + tensor x_11 = matmul(transpose_x = x_11_transpose_x_0, transpose_y = x_11_transpose_y_0, x = attn, y = v_attn)[name = tensor("x_11")]; + tensor var_464_perm_0 = const()[name = tensor("op_464_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_465 = const()[name = tensor("op_465"), val = tensor([1, 16, 512])]; + tensor var_464 = transpose(perm = var_464_perm_0, x = x_11)[name = tensor("transpose_11")]; + tensor input_15 = reshape(shape = var_465, x = var_464)[name = tensor("input_15")]; + tensor x_13 = linear(bias = linear_1_bias_0, weight = mimi_decoder_transformer_transformer_layers_1_self_attn_out_proj_weight, x = input_15)[name = tensor("linear_5")]; + tensor var_474 = mul(x = mimi_decoder_transformer_transformer_layers_1_layer_scale_1_scale, y = x_13)[name = tensor("op_474")]; + tensor input_17 = add(x = input_13, y = var_474)[name = tensor("input_17")]; + tensor input_19_axes_0 = const()[name = tensor("input_19_axes_0"), val = tensor([-1])]; + tensor input_19 = layer_norm(axes = input_19_axes_0, beta = mimi_decoder_transformer_transformer_layers_1_norm2_bias, epsilon = var_102, gamma = mimi_decoder_transformer_transformer_layers_1_norm2_weight, x = input_17)[name = tensor("input_19")]; + tensor var_481 = linear(bias = linear_2_bias_0, weight = mimi_decoder_transformer_transformer_layers_1_linear1_weight, x = input_19)[name = tensor("linear_6")]; + tensor input_21_mode_0 = const()[name = tensor("input_21_mode_0"), val = tensor("EXACT")]; + tensor input_21 = gelu(mode = input_21_mode_0, x = var_481)[name = tensor("input_21")]; + tensor x_15 = linear(bias = linear_1_bias_0, weight = mimi_decoder_transformer_transformer_layers_1_linear2_weight, x = input_21)[name = tensor("linear_7")]; + tensor var_487 = mul(x = mimi_decoder_transformer_transformer_layers_1_layer_scale_2_scale, y = x_15)[name = tensor("op_487")]; + tensor z = add(x = input_17, y = var_487)[name = tensor("z")]; + tensor x_17_perm_0 = const()[name = tensor("x_17_perm_0"), val = tensor([0, 2, 1])]; + tensor var_507 = const()[name = tensor("op_507"), val = tensor(0x1p+0)]; + tensor var_508 = const()[name = tensor("op_508"), val = tensor(-1)]; + tensor input_23_interleave_0 = const()[name = tensor("input_23_interleave_0"), val = tensor(false)]; + tensor x_17 = transpose(perm = x_17_perm_0, x = z)[name = tensor("transpose_10")]; + tensor input_23 = concat(axis = var_508, interleave = input_23_interleave_0, values = (conv0_prev, x_17))[name = tensor("input_23")]; + tensor input_25_pad_type_0 = const()[name = tensor("input_25_pad_type_0"), val = tensor("valid")]; + tensor input_25_strides_0 = const()[name = tensor("input_25_strides_0"), val = tensor([1])]; + tensor input_25_pad_0 = const()[name = tensor("input_25_pad_0"), val = tensor([0, 0])]; + tensor input_25_dilations_0 = const()[name = tensor("input_25_dilations_0"), val = tensor([1])]; + tensor input_25_groups_0 = const()[name = tensor("input_25_groups_0"), val = tensor(1)]; + tensor input_25 = conv(bias = mimi_decoder_model_0_conv_bias, dilations = input_25_dilations_0, groups = input_25_groups_0, pad = input_25_pad_0, pad_type = input_25_pad_type_0, strides = input_25_strides_0, weight = mimi_decoder_model_0_conv_weight, x = input_23)[name = tensor("input_25")]; + tensor var_542_begin_0 = const()[name = tensor("op_542_begin_0"), val = tensor([0, 0, 16])]; + tensor var_542_end_0 = const()[name = tensor("op_542_end_0"), val = tensor([1, 512, 22])]; + tensor var_542_end_mask_0 = const()[name = tensor("op_542_end_mask_0"), val = tensor([true, true, true])]; + tensor var_542 = slice_by_index(begin = var_542_begin_0, end = var_542_end_0, end_mask = var_542_end_mask_0, x = input_23)[name = tensor("op_542")]; + tensor input_27 = elu(alpha = var_507, x = input_25)[name = tensor("input_27")]; + tensor y_5_pad_type_0 = const()[name = tensor("y_5_pad_type_0"), val = tensor("valid")]; + tensor y_5_strides_0 = const()[name = tensor("y_5_strides_0"), val = tensor([6])]; + tensor y_5_pad_0 = const()[name = tensor("y_5_pad_0"), val = tensor([0, 0])]; + tensor y_5_dilations_0 = const()[name = tensor("y_5_dilations_0"), val = tensor([1])]; + tensor y_5_groups_0 = const()[name = tensor("y_5_groups_0"), val = tensor(1)]; + tensor y_5_has_output_shape_output_shape_0 = const()[name = tensor("y_5_has_output_shape_output_shape_0"), val = tensor([1, 256, 102])]; + tensor y_5_has_output_shape = conv_transpose(bias = mimi_decoder_model_2_convtr_bias, dilations = y_5_dilations_0, groups = y_5_groups_0, output_shape = y_5_has_output_shape_output_shape_0, pad = y_5_pad_0, pad_type = y_5_pad_type_0, strides = y_5_strides_0, weight = mimi_decoder_model_2_convtr_weight, x = input_27)[name = tensor("y_5_has_output_shape")]; + tensor var_557_begin_0 = const()[name = tensor("op_557_begin_0"), val = tensor([0, 0, 0])]; + tensor var_557_end_0 = const()[name = tensor("op_557_end_0"), val = tensor([1, 256, 6])]; + tensor var_557_end_mask_0 = const()[name = tensor("op_557_end_mask_0"), val = tensor([true, true, false])]; + tensor var_557 = slice_by_index(begin = var_557_begin_0, end = var_557_end_0, end_mask = var_557_end_mask_0, x = y_5_has_output_shape)[name = tensor("op_557")]; + tensor var_558 = add(x = var_557, y = convtr0_partial)[name = tensor("op_558")]; + tensor var_559_begin_0 = const()[name = tensor("op_559_begin_0"), val = tensor([0, 0, 6])]; + tensor var_559_end_0 = const()[name = tensor("op_559_end_0"), val = tensor([1, 256, 102])]; + tensor var_559_end_mask_0 = const()[name = tensor("op_559_end_mask_0"), val = tensor([true, true, true])]; + tensor var_559 = slice_by_index(begin = var_559_begin_0, end = var_559_end_0, end_mask = var_559_end_mask_0, x = y_5_has_output_shape)[name = tensor("op_559")]; + tensor y_7_interleave_0 = const()[name = tensor("y_7_interleave_0"), val = tensor(false)]; + tensor y_7 = concat(axis = var_508, interleave = y_7_interleave_0, values = (var_558, var_559))[name = tensor("y_7")]; + tensor new_partial_1_begin_0 = const()[name = tensor("new_partial_1_begin_0"), val = tensor([0, 0, 96])]; + tensor new_partial_1_end_0 = const()[name = tensor("new_partial_1_end_0"), val = tensor([1, 256, 102])]; + tensor new_partial_1_end_mask_0 = const()[name = tensor("new_partial_1_end_mask_0"), val = tensor([true, true, true])]; + tensor new_partial_1 = slice_by_index(begin = new_partial_1_begin_0, end = new_partial_1_end_0, end_mask = new_partial_1_end_mask_0, x = y_7)[name = tensor("new_partial_1")]; + tensor var_564 = const()[name = tensor("op_564"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(41766272)))]; + tensor var_565 = sub(x = new_partial_1, y = var_564)[name = tensor("op_565")]; + tensor input_29_begin_0 = const()[name = tensor("input_29_begin_0"), val = tensor([0, 0, 0])]; + tensor input_29_end_0 = const()[name = tensor("input_29_end_0"), val = tensor([1, 256, 96])]; + tensor input_29_end_mask_0 = const()[name = tensor("input_29_end_mask_0"), val = tensor([true, true, false])]; + tensor input_29 = slice_by_index(begin = input_29_begin_0, end = input_29_end_0, end_mask = input_29_end_mask_0, x = y_7)[name = tensor("input_29")]; + tensor x_19 = elu(alpha = var_507, x = input_29)[name = tensor("x_19")]; + tensor input_31_interleave_0 = const()[name = tensor("input_31_interleave_0"), val = tensor(false)]; + tensor input_31 = concat(axis = var_508, interleave = input_31_interleave_0, values = (res0_conv0_prev, x_19))[name = tensor("input_31")]; + tensor input_33_pad_type_0 = const()[name = tensor("input_33_pad_type_0"), val = tensor("valid")]; + tensor input_33_strides_0 = const()[name = tensor("input_33_strides_0"), val = tensor([1])]; + tensor input_33_pad_0 = const()[name = tensor("input_33_pad_0"), val = tensor([0, 0])]; + tensor input_33_dilations_0 = const()[name = tensor("input_33_dilations_0"), val = tensor([1])]; + tensor input_33_groups_0 = const()[name = tensor("input_33_groups_0"), val = tensor(1)]; + tensor input_33 = conv(bias = mimi_decoder_model_3_block_1_conv_bias, dilations = input_33_dilations_0, groups = input_33_groups_0, pad = input_33_pad_0, pad_type = input_33_pad_type_0, strides = input_33_strides_0, weight = mimi_decoder_model_3_block_1_conv_weight, x = input_31)[name = tensor("input_33")]; + tensor var_585_begin_0 = const()[name = tensor("op_585_begin_0"), val = tensor([0, 0, 96])]; + tensor var_585_end_0 = const()[name = tensor("op_585_end_0"), val = tensor([1, 256, 98])]; + tensor var_585_end_mask_0 = const()[name = tensor("op_585_end_mask_0"), val = tensor([true, true, true])]; + tensor var_585 = slice_by_index(begin = var_585_begin_0, end = var_585_end_0, end_mask = var_585_end_mask_0, x = input_31)[name = tensor("op_585")]; + tensor x_21 = elu(alpha = var_507, x = input_33)[name = tensor("x_21")]; + tensor v_5_pad_type_0 = const()[name = tensor("v_5_pad_type_0"), val = tensor("valid")]; + tensor v_5_strides_0 = const()[name = tensor("v_5_strides_0"), val = tensor([1])]; + tensor v_5_pad_0 = const()[name = tensor("v_5_pad_0"), val = tensor([0, 0])]; + tensor v_5_dilations_0 = const()[name = tensor("v_5_dilations_0"), val = tensor([1])]; + tensor v_5_groups_0 = const()[name = tensor("v_5_groups_0"), val = tensor(1)]; + tensor v_5 = conv(bias = mimi_decoder_model_3_block_3_conv_bias, dilations = v_5_dilations_0, groups = v_5_groups_0, pad = v_5_pad_0, pad_type = v_5_pad_type_0, strides = v_5_strides_0, weight = mimi_decoder_model_3_block_3_conv_weight, x = x_21)[name = tensor("v_5")]; + tensor input_35 = add(x = input_29, y = v_5)[name = tensor("input_35")]; + tensor input_37 = elu(alpha = var_507, x = input_35)[name = tensor("input_37")]; + tensor y_9_pad_type_0 = const()[name = tensor("y_9_pad_type_0"), val = tensor("valid")]; + tensor y_9_strides_0 = const()[name = tensor("y_9_strides_0"), val = tensor([5])]; + tensor y_9_pad_0 = const()[name = tensor("y_9_pad_0"), val = tensor([0, 0])]; + tensor y_9_dilations_0 = const()[name = tensor("y_9_dilations_0"), val = tensor([1])]; + tensor y_9_groups_0 = const()[name = tensor("y_9_groups_0"), val = tensor(1)]; + tensor y_9_has_output_shape_output_shape_0 = const()[name = tensor("y_9_has_output_shape_output_shape_0"), val = tensor([1, 128, 485])]; + tensor y_9_has_output_shape = conv_transpose(bias = mimi_decoder_model_5_convtr_bias, dilations = y_9_dilations_0, groups = y_9_groups_0, output_shape = y_9_has_output_shape_output_shape_0, pad = y_9_pad_0, pad_type = y_9_pad_type_0, strides = y_9_strides_0, weight = mimi_decoder_model_5_convtr_weight, x = input_37)[name = tensor("y_9_has_output_shape")]; + tensor var_613_begin_0 = const()[name = tensor("op_613_begin_0"), val = tensor([0, 0, 0])]; + tensor var_613_end_0 = const()[name = tensor("op_613_end_0"), val = tensor([1, 128, 5])]; + tensor var_613_end_mask_0 = const()[name = tensor("op_613_end_mask_0"), val = tensor([true, true, false])]; + tensor var_613 = slice_by_index(begin = var_613_begin_0, end = var_613_end_0, end_mask = var_613_end_mask_0, x = y_9_has_output_shape)[name = tensor("op_613")]; + tensor var_614 = add(x = var_613, y = convtr1_partial)[name = tensor("op_614")]; + tensor var_615_begin_0 = const()[name = tensor("op_615_begin_0"), val = tensor([0, 0, 5])]; + tensor var_615_end_0 = const()[name = tensor("op_615_end_0"), val = tensor([1, 128, 485])]; + tensor var_615_end_mask_0 = const()[name = tensor("op_615_end_mask_0"), val = tensor([true, true, true])]; + tensor var_615 = slice_by_index(begin = var_615_begin_0, end = var_615_end_0, end_mask = var_615_end_mask_0, x = y_9_has_output_shape)[name = tensor("op_615")]; + tensor y_11_interleave_0 = const()[name = tensor("y_11_interleave_0"), val = tensor(false)]; + tensor y_11 = concat(axis = var_508, interleave = y_11_interleave_0, values = (var_614, var_615))[name = tensor("y_11")]; + tensor new_partial_3_begin_0 = const()[name = tensor("new_partial_3_begin_0"), val = tensor([0, 0, 480])]; + tensor new_partial_3_end_0 = const()[name = tensor("new_partial_3_end_0"), val = tensor([1, 128, 485])]; + tensor new_partial_3_end_mask_0 = const()[name = tensor("new_partial_3_end_mask_0"), val = tensor([true, true, true])]; + tensor new_partial_3 = slice_by_index(begin = new_partial_3_begin_0, end = new_partial_3_end_0, end_mask = new_partial_3_end_mask_0, x = y_11)[name = tensor("new_partial_3")]; + tensor var_620 = const()[name = tensor("op_620"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(41767360)))]; + tensor var_621 = sub(x = new_partial_3, y = var_620)[name = tensor("op_621")]; + tensor input_39_begin_0 = const()[name = tensor("input_39_begin_0"), val = tensor([0, 0, 0])]; + tensor input_39_end_0 = const()[name = tensor("input_39_end_0"), val = tensor([1, 128, 480])]; + tensor input_39_end_mask_0 = const()[name = tensor("input_39_end_mask_0"), val = tensor([true, true, false])]; + tensor input_39 = slice_by_index(begin = input_39_begin_0, end = input_39_end_0, end_mask = input_39_end_mask_0, x = y_11)[name = tensor("input_39")]; + tensor x_23 = elu(alpha = var_507, x = input_39)[name = tensor("x_23")]; + tensor input_41_interleave_0 = const()[name = tensor("input_41_interleave_0"), val = tensor(false)]; + tensor input_41 = concat(axis = var_508, interleave = input_41_interleave_0, values = (res1_conv0_prev, x_23))[name = tensor("input_41")]; + tensor input_43_pad_type_0 = const()[name = tensor("input_43_pad_type_0"), val = tensor("valid")]; + tensor input_43_strides_0 = const()[name = tensor("input_43_strides_0"), val = tensor([1])]; + tensor input_43_pad_0 = const()[name = tensor("input_43_pad_0"), val = tensor([0, 0])]; + tensor input_43_dilations_0 = const()[name = tensor("input_43_dilations_0"), val = tensor([1])]; + tensor input_43_groups_0 = const()[name = tensor("input_43_groups_0"), val = tensor(1)]; + tensor input_43 = conv(bias = mimi_decoder_model_6_block_1_conv_bias, dilations = input_43_dilations_0, groups = input_43_groups_0, pad = input_43_pad_0, pad_type = input_43_pad_type_0, strides = input_43_strides_0, weight = mimi_decoder_model_6_block_1_conv_weight, x = input_41)[name = tensor("input_43")]; + tensor var_641_begin_0 = const()[name = tensor("op_641_begin_0"), val = tensor([0, 0, 480])]; + tensor var_641_end_0 = const()[name = tensor("op_641_end_0"), val = tensor([1, 128, 482])]; + tensor var_641_end_mask_0 = const()[name = tensor("op_641_end_mask_0"), val = tensor([true, true, true])]; + tensor var_641 = slice_by_index(begin = var_641_begin_0, end = var_641_end_0, end_mask = var_641_end_mask_0, x = input_41)[name = tensor("op_641")]; + tensor x_25 = elu(alpha = var_507, x = input_43)[name = tensor("x_25")]; + tensor v_7_pad_type_0 = const()[name = tensor("v_7_pad_type_0"), val = tensor("valid")]; + tensor v_7_strides_0 = const()[name = tensor("v_7_strides_0"), val = tensor([1])]; + tensor v_7_pad_0 = const()[name = tensor("v_7_pad_0"), val = tensor([0, 0])]; + tensor v_7_dilations_0 = const()[name = tensor("v_7_dilations_0"), val = tensor([1])]; + tensor v_7_groups_0 = const()[name = tensor("v_7_groups_0"), val = tensor(1)]; + tensor v_7 = conv(bias = mimi_decoder_model_6_block_3_conv_bias, dilations = v_7_dilations_0, groups = v_7_groups_0, pad = v_7_pad_0, pad_type = v_7_pad_type_0, strides = v_7_strides_0, weight = mimi_decoder_model_6_block_3_conv_weight, x = x_25)[name = tensor("v_7")]; + tensor input_45 = add(x = input_39, y = v_7)[name = tensor("input_45")]; + tensor input_47 = elu(alpha = var_507, x = input_45)[name = tensor("input_47")]; + tensor y_13_pad_type_0 = const()[name = tensor("y_13_pad_type_0"), val = tensor("valid")]; + tensor y_13_strides_0 = const()[name = tensor("y_13_strides_0"), val = tensor([4])]; + tensor y_13_pad_0 = const()[name = tensor("y_13_pad_0"), val = tensor([0, 0])]; + tensor y_13_dilations_0 = const()[name = tensor("y_13_dilations_0"), val = tensor([1])]; + tensor y_13_groups_0 = const()[name = tensor("y_13_groups_0"), val = tensor(1)]; + tensor y_13_has_output_shape_output_shape_0 = const()[name = tensor("y_13_has_output_shape_output_shape_0"), val = tensor([1, 64, 1924])]; + tensor y_13_has_output_shape = conv_transpose(bias = mimi_decoder_model_8_convtr_bias, dilations = y_13_dilations_0, groups = y_13_groups_0, output_shape = y_13_has_output_shape_output_shape_0, pad = y_13_pad_0, pad_type = y_13_pad_type_0, strides = y_13_strides_0, weight = mimi_decoder_model_8_convtr_weight, x = input_47)[name = tensor("y_13_has_output_shape")]; + tensor var_669_begin_0 = const()[name = tensor("op_669_begin_0"), val = tensor([0, 0, 0])]; + tensor var_669_end_0 = const()[name = tensor("op_669_end_0"), val = tensor([1, 64, 4])]; + tensor var_669_end_mask_0 = const()[name = tensor("op_669_end_mask_0"), val = tensor([true, true, false])]; + tensor var_669 = slice_by_index(begin = var_669_begin_0, end = var_669_end_0, end_mask = var_669_end_mask_0, x = y_13_has_output_shape)[name = tensor("op_669")]; + tensor var_670 = add(x = var_669, y = convtr2_partial)[name = tensor("op_670")]; + tensor var_671_begin_0 = const()[name = tensor("op_671_begin_0"), val = tensor([0, 0, 4])]; + tensor var_671_end_0 = const()[name = tensor("op_671_end_0"), val = tensor([1, 64, 1924])]; + tensor var_671_end_mask_0 = const()[name = tensor("op_671_end_mask_0"), val = tensor([true, true, true])]; + tensor var_671 = slice_by_index(begin = var_671_begin_0, end = var_671_end_0, end_mask = var_671_end_mask_0, x = y_13_has_output_shape)[name = tensor("op_671")]; + tensor y_interleave_0 = const()[name = tensor("y_interleave_0"), val = tensor(false)]; + tensor y = concat(axis = var_508, interleave = y_interleave_0, values = (var_670, var_671))[name = tensor("y")]; + tensor new_partial_begin_0 = const()[name = tensor("new_partial_begin_0"), val = tensor([0, 0, 1920])]; + tensor new_partial_end_0 = const()[name = tensor("new_partial_end_0"), val = tensor([1, 64, 1924])]; + tensor new_partial_end_mask_0 = const()[name = tensor("new_partial_end_mask_0"), val = tensor([true, true, true])]; + tensor new_partial = slice_by_index(begin = new_partial_begin_0, end = new_partial_end_0, end_mask = new_partial_end_mask_0, x = y)[name = tensor("new_partial")]; + tensor var_676 = const()[name = tensor("op_676"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(41767936)))]; + tensor var_677 = sub(x = new_partial, y = var_676)[name = tensor("op_677")]; + tensor input_49_begin_0 = const()[name = tensor("input_49_begin_0"), val = tensor([0, 0, 0])]; + tensor input_49_end_0 = const()[name = tensor("input_49_end_0"), val = tensor([1, 64, 1920])]; + tensor input_49_end_mask_0 = const()[name = tensor("input_49_end_mask_0"), val = tensor([true, true, false])]; + tensor input_49 = slice_by_index(begin = input_49_begin_0, end = input_49_end_0, end_mask = input_49_end_mask_0, x = y)[name = tensor("input_49")]; + tensor x_27 = elu(alpha = var_507, x = input_49)[name = tensor("x_27")]; + tensor input_51_interleave_0 = const()[name = tensor("input_51_interleave_0"), val = tensor(false)]; + tensor input_51 = concat(axis = var_508, interleave = input_51_interleave_0, values = (res2_conv0_prev, x_27))[name = tensor("input_51")]; + tensor input_53_pad_type_0 = const()[name = tensor("input_53_pad_type_0"), val = tensor("valid")]; + tensor input_53_strides_0 = const()[name = tensor("input_53_strides_0"), val = tensor([1])]; + tensor input_53_pad_0 = const()[name = tensor("input_53_pad_0"), val = tensor([0, 0])]; + tensor input_53_dilations_0 = const()[name = tensor("input_53_dilations_0"), val = tensor([1])]; + tensor input_53_groups_0 = const()[name = tensor("input_53_groups_0"), val = tensor(1)]; + tensor input_53 = conv(bias = mimi_decoder_model_9_block_1_conv_bias, dilations = input_53_dilations_0, groups = input_53_groups_0, pad = input_53_pad_0, pad_type = input_53_pad_type_0, strides = input_53_strides_0, weight = mimi_decoder_model_9_block_1_conv_weight, x = input_51)[name = tensor("input_53")]; + tensor var_697_begin_0 = const()[name = tensor("op_697_begin_0"), val = tensor([0, 0, 1920])]; + tensor var_697_end_0 = const()[name = tensor("op_697_end_0"), val = tensor([1, 64, 1922])]; + tensor var_697_end_mask_0 = const()[name = tensor("op_697_end_mask_0"), val = tensor([true, true, true])]; + tensor var_697 = slice_by_index(begin = var_697_begin_0, end = var_697_end_0, end_mask = var_697_end_mask_0, x = input_51)[name = tensor("op_697")]; + tensor x_29 = elu(alpha = var_507, x = input_53)[name = tensor("x_29")]; + tensor v_pad_type_0 = const()[name = tensor("v_pad_type_0"), val = tensor("valid")]; + tensor v_strides_0 = const()[name = tensor("v_strides_0"), val = tensor([1])]; + tensor v_pad_0 = const()[name = tensor("v_pad_0"), val = tensor([0, 0])]; + tensor v_dilations_0 = const()[name = tensor("v_dilations_0"), val = tensor([1])]; + tensor v_groups_0 = const()[name = tensor("v_groups_0"), val = tensor(1)]; + tensor v = conv(bias = mimi_decoder_model_9_block_3_conv_bias, dilations = v_dilations_0, groups = v_groups_0, pad = v_pad_0, pad_type = v_pad_type_0, strides = v_strides_0, weight = mimi_decoder_model_9_block_3_conv_weight, x = x_29)[name = tensor("v")]; + tensor input_55 = add(x = input_49, y = v)[name = tensor("input_55")]; + tensor x = elu(alpha = var_507, x = input_55)[name = tensor("x")]; + tensor input_interleave_0 = const()[name = tensor("input_interleave_0"), val = tensor(false)]; + tensor input = concat(axis = var_508, interleave = input_interleave_0, values = (conv_final_prev, x))[name = tensor("input")]; + tensor var_724_pad_type_0 = const()[name = tensor("op_724_pad_type_0"), val = tensor("valid")]; + tensor var_724_strides_0 = const()[name = tensor("op_724_strides_0"), val = tensor([1])]; + tensor var_724_pad_0 = const()[name = tensor("op_724_pad_0"), val = tensor([0, 0])]; + tensor var_724_dilations_0 = const()[name = tensor("op_724_dilations_0"), val = tensor([1])]; + tensor var_724_groups_0 = const()[name = tensor("op_724_groups_0"), val = tensor(1)]; + tensor var_724 = conv(bias = mimi_decoder_model_11_conv_bias, dilations = var_724_dilations_0, groups = var_724_groups_0, pad = var_724_pad_0, pad_type = var_724_pad_type_0, strides = var_724_strides_0, weight = mimi_decoder_model_11_conv_weight, x = input)[name = tensor("op_724")]; + tensor var_725_begin_0 = const()[name = tensor("op_725_begin_0"), val = tensor([0, 0, 1920])]; + tensor var_725_end_0 = const()[name = tensor("op_725_end_0"), val = tensor([1, 64, 1922])]; + tensor var_725_end_mask_0 = const()[name = tensor("op_725_end_mask_0"), val = tensor([true, true, true])]; + tensor var_725 = slice_by_index(begin = var_725_begin_0, end = var_725_end_0, end_mask = var_725_end_mask_0, x = input)[name = tensor("op_725")]; + tensor var_740_promoted = const()[name = tensor("op_740_promoted"), val = tensor(0x1p+4)]; + tensor var_741 = add(x = attn0_offset, y = var_740_promoted)[name = tensor("op_741")]; + tensor var_743_promoted = const()[name = tensor("op_743_promoted"), val = tensor(0x1p+4)]; + tensor var_744 = add(x = attn1_offset, y = var_743_promoted)[name = tensor("op_744")]; + tensor conv0_first_tmp = identity(x = conv0_first)[name = tensor("conv0_first_tmp")]; + tensor res0_conv0_first_tmp = identity(x = res0_conv0_first)[name = tensor("res0_conv0_first_tmp")]; + tensor res0_conv1_prev_tmp = identity(x = res0_conv1_prev)[name = tensor("res0_conv1_prev_tmp")]; + tensor res0_conv1_first_tmp = identity(x = res0_conv1_first)[name = tensor("res0_conv1_first_tmp")]; + tensor res1_conv0_first_tmp = identity(x = res1_conv0_first)[name = tensor("res1_conv0_first_tmp")]; + tensor res1_conv1_prev_tmp = identity(x = res1_conv1_prev)[name = tensor("res1_conv1_prev_tmp")]; + tensor res1_conv1_first_tmp = identity(x = res1_conv1_first)[name = tensor("res1_conv1_first_tmp")]; + tensor res2_conv0_first_tmp = identity(x = res2_conv0_first)[name = tensor("res2_conv0_first_tmp")]; + tensor res2_conv1_prev_tmp = identity(x = res2_conv1_prev)[name = tensor("res2_conv1_prev_tmp")]; + tensor res2_conv1_first_tmp = identity(x = res2_conv1_first)[name = tensor("res2_conv1_first_tmp")]; + tensor conv_final_first_tmp = identity(x = conv_final_first)[name = tensor("conv_final_first_tmp")]; + } -> (var_724, var_77, var_210, var_741, var_400, var_744, var_542, conv0_first, var_565, var_585, res0_conv0_first, res0_conv1_prev, res0_conv1_first, var_621, var_641, res1_conv0_first, res1_conv1_prev, res1_conv1_first, var_677, var_697, res2_conv0_first, res2_conv1_prev, res2_conv1_first, var_725, conv_final_first); +} \ No newline at end of file diff --git a/v2.1/german/mimi_decoder.mlmodelc/weights/weight.bin b/v2.1/german/mimi_decoder.mlmodelc/weights/weight.bin new file mode 100644 index 0000000000000000000000000000000000000000..0bf03171362a32f81af7fa74d4b45c0b6c262a82 --- /dev/null +++ b/v2.1/german/mimi_decoder.mlmodelc/weights/weight.bin @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:fc64ac703e883ade4ec052f3c14cc2a4489161bcabfdb24c6496e420eb3965b3 +size 41768256 diff --git a/v2.1/german/mimi_decoder.mlpackage/Data/com.apple.CoreML/model.mlmodel b/v2.1/german/mimi_decoder.mlpackage/Data/com.apple.CoreML/model.mlmodel new file mode 100644 index 0000000000000000000000000000000000000000..f63ada87718675ea20ae364844c945d5ae430572 --- /dev/null +++ b/v2.1/german/mimi_decoder.mlpackage/Data/com.apple.CoreML/model.mlmodel @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:70e82d98f99ea3438f824b56281c8e6e159a24022c9e47f7c0c6fb19e74ac5e5 +size 93270 diff --git a/v2.1/german/mimi_decoder.mlpackage/Data/com.apple.CoreML/weights/weight.bin b/v2.1/german/mimi_decoder.mlpackage/Data/com.apple.CoreML/weights/weight.bin new file mode 100644 index 0000000000000000000000000000000000000000..0bf03171362a32f81af7fa74d4b45c0b6c262a82 --- /dev/null +++ b/v2.1/german/mimi_decoder.mlpackage/Data/com.apple.CoreML/weights/weight.bin @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:fc64ac703e883ade4ec052f3c14cc2a4489161bcabfdb24c6496e420eb3965b3 +size 41768256 diff --git a/v2.1/german/mimi_decoder.mlpackage/Manifest.json b/v2.1/german/mimi_decoder.mlpackage/Manifest.json new file mode 100644 index 0000000000000000000000000000000000000000..3b19c46d4f85c8989bdcafc2547296ccd39de73b --- /dev/null +++ b/v2.1/german/mimi_decoder.mlpackage/Manifest.json @@ -0,0 +1,18 @@ +{ + "fileFormatVersion": "1.0.0", + "itemInfoEntries": { + "384424C2-95AF-4B75-88DD-07A1D0F432E1": { + "author": "com.apple.CoreML", + "description": "CoreML Model Weights", + "name": "weights", + "path": "com.apple.CoreML/weights" + }, + "9AB0C58F-D2D1-4B44-9E5D-D336F6CF7D82": { + "author": "com.apple.CoreML", + "description": "CoreML Model Specification", + "name": "model.mlmodel", + "path": "com.apple.CoreML/model.mlmodel" + } + }, + "rootModelIdentifier": "9AB0C58F-D2D1-4B44-9E5D-D336F6CF7D82" +} diff --git a/v2.1/german_24l/cond_prefill.mlmodelc/analytics/coremldata.bin b/v2.1/german_24l/cond_prefill.mlmodelc/analytics/coremldata.bin new file mode 100644 index 0000000000000000000000000000000000000000..d3b280f1968a0a85a8a5773cbb383ebac7a080ba --- /dev/null +++ b/v2.1/german_24l/cond_prefill.mlmodelc/analytics/coremldata.bin @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:42b6b657c67ca9c8ef8743eca2df2ee4f7dcefc3b0b5610589ccb9d7b2a42c81 +size 243 diff --git a/v2.1/german_24l/cond_prefill.mlmodelc/coremldata.bin b/v2.1/german_24l/cond_prefill.mlmodelc/coremldata.bin new file mode 100644 index 0000000000000000000000000000000000000000..1785fb31531e3bce1dc283e9568e181375d86086 --- /dev/null +++ b/v2.1/german_24l/cond_prefill.mlmodelc/coremldata.bin @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:6c02efb89c164fdef27a53f29dea316d25d1ae4ae798741528257942b0e7ad35 +size 3519 diff --git a/v2.1/german_24l/cond_prefill.mlmodelc/model.mil b/v2.1/german_24l/cond_prefill.mlmodelc/model.mil new file mode 100644 index 0000000000000000000000000000000000000000..7028cd94630c33bb5143ac20b6b86ca016eb1d25 --- /dev/null +++ b/v2.1/german_24l/cond_prefill.mlmodelc/model.mil @@ -0,0 +1,5431 @@ +program(1.0) +[buildInfo = dict, tensor>({{"coremlc-component-MIL", "3520.4.1"}, {"coremlc-version", "3520.5.1"}, {"coremltools-component-torch", "2.12.0"}, {"coremltools-source-dialect", "TorchScript"}, {"coremltools-version", "9.0"}})] +{ + func main(tensor cache0, tensor cache1, tensor cache10, tensor cache11, tensor cache12, tensor cache13, tensor cache14, tensor cache15, tensor cache16, tensor cache17, tensor cache18, tensor cache19, tensor cache2, tensor cache20, tensor cache21, tensor cache22, tensor cache23, tensor cache3, tensor cache4, tensor cache5, tensor cache6, tensor cache7, tensor cache8, tensor cache9, tensor conditioning, tensor position0, tensor position1, tensor position10, tensor position11, tensor position12, tensor position13, tensor position14, tensor position15, tensor position16, tensor position17, tensor position18, tensor position19, tensor position2, tensor position20, tensor position21, tensor position22, tensor position23, tensor position3, tensor position4, tensor position5, tensor position6, tensor position7, tensor position8, tensor position9, tensor valid_len) { + tensor input_3_axes_0 = const()[name = tensor("input_3_axes_0"), val = tensor([-1])]; + tensor conditioning_to_fp16_dtype_0 = const()[name = tensor("conditioning_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor norm0_1_weight_to_fp16 = const()[name = tensor("norm0_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(64)))]; + tensor norm0_1_bias_to_fp16 = const()[name = tensor("norm0_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(2176)))]; + tensor var_192_to_fp16 = const()[name = tensor("op_192_to_fp16"), val = tensor(0x1.5p-17)]; + tensor conditioning_to_fp16 = cast(dtype = conditioning_to_fp16_dtype_0, x = conditioning)[name = tensor("cast_663")]; + tensor input_3_cast_fp16 = layer_norm(axes = input_3_axes_0, beta = norm0_1_bias_to_fp16, epsilon = var_192_to_fp16, gamma = norm0_1_weight_to_fp16, x = conditioning_to_fp16)[name = tensor("input_3_cast_fp16")]; + tensor attn0_in_proj_weight_to_fp16 = const()[name = tensor("attn0_in_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(4288)))]; + tensor linear_0_bias_0_to_fp16 = const()[name = tensor("linear_0_bias_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(6295808)))]; + tensor linear_0_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = attn0_in_proj_weight_to_fp16, x = input_3_cast_fp16)[name = tensor("linear_0_cast_fp16")]; + tensor var_206 = const()[name = tensor("op_206"), val = tensor([1, 256, 3, 16, 64])]; + tensor qkv_1_cast_fp16 = reshape(shape = var_206, x = linear_0_cast_fp16)[name = tensor("qkv_1_cast_fp16")]; + tensor q_1_begin_0 = const()[name = tensor("q_1_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor q_1_end_0 = const()[name = tensor("q_1_end_0"), val = tensor([1, 256, 1, 16, 64])]; + tensor q_1_end_mask_0 = const()[name = tensor("q_1_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor q_1_squeeze_mask_0 = const()[name = tensor("q_1_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor q_1_cast_fp16 = slice_by_index(begin = q_1_begin_0, end = q_1_end_0, end_mask = q_1_end_mask_0, squeeze_mask = q_1_squeeze_mask_0, x = qkv_1_cast_fp16)[name = tensor("q_1_cast_fp16")]; + tensor k_1_begin_0 = const()[name = tensor("k_1_begin_0"), val = tensor([0, 0, 1, 0, 0])]; + tensor k_1_end_0 = const()[name = tensor("k_1_end_0"), val = tensor([1, 256, 2, 16, 64])]; + tensor k_1_end_mask_0 = const()[name = tensor("k_1_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor k_1_squeeze_mask_0 = const()[name = tensor("k_1_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor k_1_cast_fp16 = slice_by_index(begin = k_1_begin_0, end = k_1_end_0, end_mask = k_1_end_mask_0, squeeze_mask = k_1_squeeze_mask_0, x = qkv_1_cast_fp16)[name = tensor("k_1_cast_fp16")]; + tensor v_1_begin_0 = const()[name = tensor("v_1_begin_0"), val = tensor([0, 0, 2, 0, 0])]; + tensor v_1_end_0 = const()[name = tensor("v_1_end_0"), val = tensor([1, 256, 3, 16, 64])]; + tensor v_1_end_mask_0 = const()[name = tensor("v_1_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor v_1_squeeze_mask_0 = const()[name = tensor("v_1_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor v_1_cast_fp16 = slice_by_index(begin = v_1_begin_0, end = v_1_end_0, end_mask = v_1_end_mask_0, squeeze_mask = v_1_squeeze_mask_0, x = qkv_1_cast_fp16)[name = tensor("v_1_cast_fp16")]; + tensor var_264 = const()[name = tensor("op_264"), val = tensor([1, 1])]; + tensor position0_to_fp16_dtype_0 = const()[name = tensor("position0_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor position0_to_fp16 = cast(dtype = position0_to_fp16_dtype_0, x = position0)[name = tensor("cast_662")]; + tensor var_265_cast_fp16 = reshape(shape = var_264, x = position0_to_fp16)[name = tensor("op_265_cast_fp16")]; + tensor ts_1_promoted_to_fp16 = const()[name = tensor("ts_1_promoted_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(6302016)))]; + tensor ts_3_cast_fp16 = add(x = ts_1_promoted_to_fp16, y = var_265_cast_fp16)[name = tensor("ts_3_cast_fp16")]; + tensor var_272 = const()[name = tensor("op_272"), val = tensor([1, 256, 1, 1])]; + tensor ts_5_cast_fp16 = reshape(shape = var_272, x = ts_3_cast_fp16)[name = tensor("ts_5_cast_fp16")]; + tensor var_279 = const()[name = tensor("op_279"), val = tensor([1, 256, 16, 32, 2])]; + tensor q_complex_1_cast_fp16 = reshape(shape = var_279, x = q_1_cast_fp16)[name = tensor("q_complex_1_cast_fp16")]; + tensor var_286 = const()[name = tensor("op_286"), val = tensor([1, 256, 16, 32, 2])]; + tensor k_complex_1_cast_fp16 = reshape(shape = var_286, x = k_1_cast_fp16)[name = tensor("k_complex_1_cast_fp16")]; + tensor var_290_begin_0 = const()[name = tensor("op_290_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_290_end_0 = const()[name = tensor("op_290_end_0"), val = tensor([1, 256, 16, 32, 1])]; + tensor var_290_end_mask_0 = const()[name = tensor("op_290_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_290_squeeze_mask_0 = const()[name = tensor("op_290_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_290_cast_fp16 = slice_by_index(begin = var_290_begin_0, end = var_290_end_0, end_mask = var_290_end_mask_0, squeeze_mask = var_290_squeeze_mask_0, x = q_complex_1_cast_fp16)[name = tensor("op_290_cast_fp16")]; + tensor var_298_begin_0 = const()[name = tensor("op_298_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_298_end_0 = const()[name = tensor("op_298_end_0"), val = tensor([1, 256, 16, 32, 2])]; + tensor var_298_end_mask_0 = const()[name = tensor("op_298_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_298_squeeze_mask_0 = const()[name = tensor("op_298_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_298_cast_fp16 = slice_by_index(begin = var_298_begin_0, end = var_298_end_0, end_mask = var_298_end_mask_0, squeeze_mask = var_298_squeeze_mask_0, x = q_complex_1_cast_fp16)[name = tensor("op_298_cast_fp16")]; + tensor var_306_begin_0 = const()[name = tensor("op_306_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_306_end_0 = const()[name = tensor("op_306_end_0"), val = tensor([1, 256, 16, 32, 1])]; + tensor var_306_end_mask_0 = const()[name = tensor("op_306_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_306_squeeze_mask_0 = const()[name = tensor("op_306_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_306_cast_fp16 = slice_by_index(begin = var_306_begin_0, end = var_306_end_0, end_mask = var_306_end_mask_0, squeeze_mask = var_306_squeeze_mask_0, x = k_complex_1_cast_fp16)[name = tensor("op_306_cast_fp16")]; + tensor var_314_begin_0 = const()[name = tensor("op_314_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_314_end_0 = const()[name = tensor("op_314_end_0"), val = tensor([1, 256, 16, 32, 2])]; + tensor var_314_end_mask_0 = const()[name = tensor("op_314_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_314_squeeze_mask_0 = const()[name = tensor("op_314_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_314_cast_fp16 = slice_by_index(begin = var_314_begin_0, end = var_314_end_0, end_mask = var_314_end_mask_0, squeeze_mask = var_314_squeeze_mask_0, x = k_complex_1_cast_fp16)[name = tensor("op_314_cast_fp16")]; + tensor freqs_1_to_fp16 = const()[name = tensor("freqs_1_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(6302592)))]; + tensor var_320_cast_fp16 = mul(x = freqs_1_to_fp16, y = ts_5_cast_fp16)[name = tensor("op_320_cast_fp16")]; + tensor rotr_1_cast_fp16 = cos(x = var_320_cast_fp16)[name = tensor("rotr_1_cast_fp16")]; + tensor roti_1_cast_fp16 = sin(x = var_320_cast_fp16)[name = tensor("roti_1_cast_fp16")]; + tensor var_324_cast_fp16 = mul(x = var_290_cast_fp16, y = rotr_1_cast_fp16)[name = tensor("op_324_cast_fp16")]; + tensor var_325_cast_fp16 = mul(x = var_298_cast_fp16, y = roti_1_cast_fp16)[name = tensor("op_325_cast_fp16")]; + tensor qor_1_cast_fp16 = sub(x = var_324_cast_fp16, y = var_325_cast_fp16)[name = tensor("qor_1_cast_fp16")]; + tensor var_328_cast_fp16 = mul(x = var_290_cast_fp16, y = roti_1_cast_fp16)[name = tensor("op_328_cast_fp16")]; + tensor var_329_cast_fp16 = mul(x = var_298_cast_fp16, y = rotr_1_cast_fp16)[name = tensor("op_329_cast_fp16")]; + tensor qoi_1_cast_fp16 = add(x = var_328_cast_fp16, y = var_329_cast_fp16)[name = tensor("qoi_1_cast_fp16")]; + tensor var_332_cast_fp16 = mul(x = var_306_cast_fp16, y = rotr_1_cast_fp16)[name = tensor("op_332_cast_fp16")]; + tensor var_333_cast_fp16 = mul(x = var_314_cast_fp16, y = roti_1_cast_fp16)[name = tensor("op_333_cast_fp16")]; + tensor kor_1_cast_fp16 = sub(x = var_332_cast_fp16, y = var_333_cast_fp16)[name = tensor("kor_1_cast_fp16")]; + tensor var_336_cast_fp16 = mul(x = var_306_cast_fp16, y = roti_1_cast_fp16)[name = tensor("op_336_cast_fp16")]; + tensor var_337_cast_fp16 = mul(x = var_314_cast_fp16, y = rotr_1_cast_fp16)[name = tensor("op_337_cast_fp16")]; + tensor koi_1_cast_fp16 = add(x = var_336_cast_fp16, y = var_337_cast_fp16)[name = tensor("koi_1_cast_fp16")]; + tensor qo_1_axis_0 = const()[name = tensor("qo_1_axis_0"), val = tensor(-1)]; + tensor qo_1_cast_fp16 = stack(axis = qo_1_axis_0, values = (qor_1_cast_fp16, qoi_1_cast_fp16))[name = tensor("qo_1_cast_fp16")]; + tensor ko_1_axis_0 = const()[name = tensor("ko_1_axis_0"), val = tensor(-1)]; + tensor ko_1_cast_fp16 = stack(axis = ko_1_axis_0, values = (kor_1_cast_fp16, koi_1_cast_fp16))[name = tensor("ko_1_cast_fp16")]; + tensor var_370 = const()[name = tensor("op_370"), val = tensor([1, 256, 16, 64])]; + tensor q_3_cast_fp16 = reshape(shape = var_370, x = qo_1_cast_fp16)[name = tensor("q_3_cast_fp16")]; + tensor var_376 = const()[name = tensor("op_376"), val = tensor([1, 256, 16, 64])]; + tensor k_3_cast_fp16 = reshape(shape = var_376, x = ko_1_cast_fp16)[name = tensor("k_3_cast_fp16")]; + tensor var_393_promoted_to_fp16 = const()[name = tensor("op_393_promoted_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(6302720)))]; + tensor write_indices_float_1_cast_fp16 = add(x = var_265_cast_fp16, y = var_393_promoted_to_fp16)[name = tensor("write_indices_float_1_cast_fp16")]; + tensor _inversed_397_y_0_to_fp16 = const()[name = tensor("_inversed_397_y_0_to_fp16"), val = tensor(0x1p-9)]; + tensor _inversed_397_cast_fp16 = mul(x = write_indices_float_1_cast_fp16, y = _inversed_397_y_0_to_fp16)[name = tensor("_inversed_397_cast_fp16")]; + tensor var_398_cast_fp16 = floor(x = _inversed_397_cast_fp16)[name = tensor("op_398_cast_fp16")]; + tensor var_399_to_fp16 = const()[name = tensor("op_399_to_fp16"), val = tensor(0x1p+9)]; + tensor var_400_cast_fp16 = mul(x = var_398_cast_fp16, y = var_399_to_fp16)[name = tensor("op_400_cast_fp16")]; + tensor write_indices_float_3_cast_fp16 = sub(x = write_indices_float_1_cast_fp16, y = var_400_cast_fp16)[name = tensor("write_indices_float_3_cast_fp16")]; + tensor var_409 = const()[name = tensor("op_409"), val = tensor([1, 1])]; + tensor valid_len_to_fp16_dtype_0 = const()[name = tensor("valid_len_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor valid_len_to_fp16 = cast(dtype = valid_len_to_fp16_dtype_0, x = valid_len)[name = tensor("cast_661")]; + tensor var_410_cast_fp16 = reshape(shape = var_409, x = valid_len_to_fp16)[name = tensor("op_410_cast_fp16")]; + tensor token_valid_1_cast_fp16 = less(x = var_393_promoted_to_fp16, y = var_410_cast_fp16)[name = tensor("token_valid_1_cast_fp16")]; + tensor var_418_value_0_to_fp16 = const()[name = tensor("op_418_value_0_to_fp16"), val = tensor(0x1.ffp+8)]; + tensor var_418_cast_fp16 = fill_like(ref_tensor = write_indices_float_3_cast_fp16, value = var_418_value_0_to_fp16)[name = tensor("op_418_cast_fp16")]; + tensor write_indices_float_5_cast_fp16 = select(a = write_indices_float_3_cast_fp16, b = var_418_cast_fp16, cond = token_valid_1_cast_fp16)[name = tensor("write_indices_float_5_cast_fp16")]; + tensor var_424_dtype_0 = const()[name = tensor("op_424_dtype_0"), val = tensor("int32")]; + tensor var_429 = const()[name = tensor("op_429"), val = tensor([1, 256, 1, 1])]; + tensor write_indices_float_5_cast_fp16_to_int32 = cast(dtype = var_424_dtype_0, x = write_indices_float_5_cast_fp16)[name = tensor("cast_660")]; + tensor var_430 = reshape(shape = var_429, x = write_indices_float_5_cast_fp16_to_int32)[name = tensor("op_430")]; + tensor write_indices_1_reps_0 = const()[name = tensor("write_indices_1_reps_0"), val = tensor([1, 1, 16, 64])]; + tensor write_indices_1 = tile(reps = write_indices_1_reps_0, x = var_430)[name = tensor("write_indices_1")]; + tensor var_440_begin_0 = const()[name = tensor("op_440_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_440_end_0 = const()[name = tensor("op_440_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor var_440_end_mask_0 = const()[name = tensor("op_440_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_440_squeeze_mask_0 = const()[name = tensor("op_440_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor cache0_to_fp16_dtype_0 = const()[name = tensor("cache0_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor cache0_to_fp16 = cast(dtype = cache0_to_fp16_dtype_0, x = cache0)[name = tensor("cast_659")]; + tensor var_440_cast_fp16 = slice_by_index(begin = var_440_begin_0, end = var_440_end_0, end_mask = var_440_end_mask_0, squeeze_mask = var_440_squeeze_mask_0, x = cache0_to_fp16)[name = tensor("op_440_cast_fp16")]; + tensor var_442_axis_0 = const()[name = tensor("op_442_axis_0"), val = tensor(1)]; + tensor var_442_mode_0 = const()[name = tensor("op_442_mode_0"), val = tensor("update")]; + tensor var_442_validate_indices_0 = const()[name = tensor("op_442_validate_indices_0"), val = tensor(false)]; + tensor var_442_cast_fp16 = scatter_along_axis(axis = var_442_axis_0, data = var_440_cast_fp16, indices = write_indices_1, mode = var_442_mode_0, updates = k_3_cast_fp16, validate_indices = var_442_validate_indices_0)[name = tensor("op_442_cast_fp16")]; + tensor concat_2 = const()[name = tensor("concat_2"), val = tensor([0, 0, 0, 0, 0])]; + tensor concat_3 = const()[name = tensor("concat_3"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_1_internal_tensor_assign_1_stride_0 = const()[name = tensor("new_cache_1_internal_tensor_assign_1_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_1_internal_tensor_assign_1_begin_mask_0 = const()[name = tensor("new_cache_1_internal_tensor_assign_1_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_1_internal_tensor_assign_1_end_mask_0 = const()[name = tensor("new_cache_1_internal_tensor_assign_1_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_1_internal_tensor_assign_1_squeeze_mask_0 = const()[name = tensor("new_cache_1_internal_tensor_assign_1_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_46 = const()[name = tensor("shape_46"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_0 = const()[name = tensor("reduce_prod_0"), val = tensor(1048576)]; + tensor range_1d_0_start_0 = const()[name = tensor("range_1d_0_start_0"), val = tensor(0)]; + tensor range_1d_0_step_0 = const()[name = tensor("range_1d_0_step_0"), val = tensor(1)]; + tensor range_1d_0 = range_1d(end = reduce_prod_0, start = range_1d_0_start_0, step = range_1d_0_step_0)[name = tensor("range_1d_0")]; + tensor reshape_0 = reshape(shape = shape_46, x = range_1d_0)[name = tensor("reshape_0")]; + tensor slice_by_index_0 = slice_by_index(begin = concat_2, begin_mask = new_cache_1_internal_tensor_assign_1_begin_mask_0, end = concat_3, end_mask = new_cache_1_internal_tensor_assign_1_end_mask_0, squeeze_mask = new_cache_1_internal_tensor_assign_1_squeeze_mask_0, stride = new_cache_1_internal_tensor_assign_1_stride_0, x = reshape_0)[name = tensor("slice_by_index_0")]; + tensor reshape_1_shape_0 = const()[name = tensor("reshape_1_shape_0"), val = tensor([-1])]; + tensor reshape_1 = reshape(shape = reshape_1_shape_0, x = slice_by_index_0)[name = tensor("reshape_1")]; + tensor reshape_2_shape_0 = const()[name = tensor("reshape_2_shape_0"), val = tensor([-1])]; + tensor reshape_2_cast_fp16 = reshape(shape = reshape_2_shape_0, x = var_442_cast_fp16)[name = tensor("reshape_2_cast_fp16")]; + tensor reshape_3_shape_0 = const()[name = tensor("reshape_3_shape_0"), val = tensor([-1])]; + tensor reshape_3_cast_fp16 = reshape(shape = reshape_3_shape_0, x = cache0_to_fp16)[name = tensor("reshape_3_cast_fp16")]; + tensor scatter_0_mode_0 = const()[name = tensor("scatter_0_mode_0"), val = tensor("update")]; + tensor scatter_0_axis_0 = const()[name = tensor("scatter_0_axis_0"), val = tensor(0)]; + tensor scatter_0_validate_indices_0 = const()[name = tensor("scatter_0_validate_indices_0"), val = tensor(false)]; + tensor scatter_0_cast_fp16 = scatter(axis = scatter_0_axis_0, data = reshape_3_cast_fp16, indices = reshape_1, mode = scatter_0_mode_0, updates = reshape_2_cast_fp16, validate_indices = scatter_0_validate_indices_0)[name = tensor("scatter_0_cast_fp16")]; + tensor reshape_4_cast_fp16 = reshape(shape = shape_46, x = scatter_0_cast_fp16)[name = tensor("reshape_4_cast_fp16")]; + tensor var_450_begin_0 = const()[name = tensor("op_450_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor var_450_end_0 = const()[name = tensor("op_450_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor var_450_end_mask_0 = const()[name = tensor("op_450_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_450_squeeze_mask_0 = const()[name = tensor("op_450_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor var_450_cast_fp16 = slice_by_index(begin = var_450_begin_0, end = var_450_end_0, end_mask = var_450_end_mask_0, squeeze_mask = var_450_squeeze_mask_0, x = reshape_4_cast_fp16)[name = tensor("op_450_cast_fp16")]; + tensor var_452_axis_0 = const()[name = tensor("op_452_axis_0"), val = tensor(1)]; + tensor var_452_mode_0 = const()[name = tensor("op_452_mode_0"), val = tensor("update")]; + tensor var_452_validate_indices_0 = const()[name = tensor("op_452_validate_indices_0"), val = tensor(false)]; + tensor var_452_cast_fp16 = scatter_along_axis(axis = var_452_axis_0, data = var_450_cast_fp16, indices = write_indices_1, mode = var_452_mode_0, updates = v_1_cast_fp16, validate_indices = var_452_validate_indices_0)[name = tensor("op_452_cast_fp16")]; + tensor concat_4 = const()[name = tensor("concat_4"), val = tensor([1, 0, 0, 0, 0])]; + tensor concat_5 = const()[name = tensor("concat_5"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_1_internal_tensor_assign_2_stride_0 = const()[name = tensor("new_cache_1_internal_tensor_assign_2_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_1_internal_tensor_assign_2_begin_mask_0 = const()[name = tensor("new_cache_1_internal_tensor_assign_2_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_1_internal_tensor_assign_2_end_mask_0 = const()[name = tensor("new_cache_1_internal_tensor_assign_2_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_1_internal_tensor_assign_2_squeeze_mask_0 = const()[name = tensor("new_cache_1_internal_tensor_assign_2_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_47 = const()[name = tensor("shape_47"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_1 = const()[name = tensor("reduce_prod_1"), val = tensor(1048576)]; + tensor range_1d_1_start_0 = const()[name = tensor("range_1d_1_start_0"), val = tensor(0)]; + tensor range_1d_1_step_0 = const()[name = tensor("range_1d_1_step_0"), val = tensor(1)]; + tensor range_1d_1 = range_1d(end = reduce_prod_1, start = range_1d_1_start_0, step = range_1d_1_step_0)[name = tensor("range_1d_1")]; + tensor reshape_5 = reshape(shape = shape_47, x = range_1d_1)[name = tensor("reshape_5")]; + tensor slice_by_index_1 = slice_by_index(begin = concat_4, begin_mask = new_cache_1_internal_tensor_assign_2_begin_mask_0, end = concat_5, end_mask = new_cache_1_internal_tensor_assign_2_end_mask_0, squeeze_mask = new_cache_1_internal_tensor_assign_2_squeeze_mask_0, stride = new_cache_1_internal_tensor_assign_2_stride_0, x = reshape_5)[name = tensor("slice_by_index_1")]; + tensor reshape_6_shape_0 = const()[name = tensor("reshape_6_shape_0"), val = tensor([-1])]; + tensor reshape_6 = reshape(shape = reshape_6_shape_0, x = slice_by_index_1)[name = tensor("reshape_6")]; + tensor reshape_7_shape_0 = const()[name = tensor("reshape_7_shape_0"), val = tensor([-1])]; + tensor reshape_7_cast_fp16 = reshape(shape = reshape_7_shape_0, x = var_452_cast_fp16)[name = tensor("reshape_7_cast_fp16")]; + tensor reshape_8_shape_0 = const()[name = tensor("reshape_8_shape_0"), val = tensor([-1])]; + tensor reshape_8_cast_fp16 = reshape(shape = reshape_8_shape_0, x = reshape_4_cast_fp16)[name = tensor("reshape_8_cast_fp16")]; + tensor scatter_1_mode_0 = const()[name = tensor("scatter_1_mode_0"), val = tensor("update")]; + tensor scatter_1_axis_0 = const()[name = tensor("scatter_1_axis_0"), val = tensor(0)]; + tensor scatter_1_validate_indices_0 = const()[name = tensor("scatter_1_validate_indices_0"), val = tensor(false)]; + tensor scatter_1_cast_fp16 = scatter(axis = scatter_1_axis_0, data = reshape_8_cast_fp16, indices = reshape_6, mode = scatter_1_mode_0, updates = reshape_7_cast_fp16, validate_indices = scatter_1_validate_indices_0)[name = tensor("scatter_1_cast_fp16")]; + tensor reshape_9_cast_fp16 = reshape(shape = shape_47, x = scatter_1_cast_fp16)[name = tensor("reshape_9_cast_fp16")]; + tensor reshape_9_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("reshape_9_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor keys_1_begin_0 = const()[name = tensor("keys_1_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor keys_1_end_0 = const()[name = tensor("keys_1_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor keys_1_end_mask_0 = const()[name = tensor("keys_1_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor keys_1_squeeze_mask_0 = const()[name = tensor("keys_1_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor keys_1_cast_fp16 = slice_by_index(begin = keys_1_begin_0, end = keys_1_end_0, end_mask = keys_1_end_mask_0, squeeze_mask = keys_1_squeeze_mask_0, x = reshape_9_cast_fp16)[name = tensor("keys_1_cast_fp16")]; + tensor values_1_begin_0 = const()[name = tensor("values_1_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor values_1_end_0 = const()[name = tensor("values_1_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor values_1_end_mask_0 = const()[name = tensor("values_1_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor values_1_squeeze_mask_0 = const()[name = tensor("values_1_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor values_1_cast_fp16 = slice_by_index(begin = values_1_begin_0, end = values_1_end_0, end_mask = values_1_end_mask_0, squeeze_mask = values_1_squeeze_mask_0, x = reshape_9_cast_fp16)[name = tensor("values_1_cast_fp16")]; + tensor var_464_cast_fp16 = not_equal(x = keys_1_cast_fp16, y = keys_1_cast_fp16)[name = tensor("op_464_cast_fp16")]; + tensor var_470_to_fp16 = const()[name = tensor("op_470_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(6303296)))]; + tensor keys_3_cast_fp16 = select(a = var_470_to_fp16, b = keys_1_cast_fp16, cond = var_464_cast_fp16)[name = tensor("keys_3_cast_fp16")]; + tensor var_472_cast_fp16 = not_equal(x = values_1_cast_fp16, y = values_1_cast_fp16)[name = tensor("op_472_cast_fp16")]; + tensor values_3_cast_fp16 = select(a = var_470_to_fp16, b = values_1_cast_fp16, cond = var_472_cast_fp16)[name = tensor("values_3_cast_fp16")]; + tensor var_496 = const()[name = tensor("op_496"), val = tensor([0, 2, 1, 3])]; + tensor var_512 = const()[name = tensor("op_512"), val = tensor([1, 1, 1])]; + tensor var_513_cast_fp16 = reshape(shape = var_512, x = position0_to_fp16)[name = tensor("op_513_cast_fp16")]; + tensor q_offsets_1_promoted_to_fp16 = const()[name = tensor("q_offsets_1_promoted_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(7351936)))]; + tensor q_positions_1_cast_fp16 = add(x = var_513_cast_fp16, y = q_offsets_1_promoted_to_fp16)[name = tensor("q_positions_1_cast_fp16")]; + tensor var_535 = const()[name = tensor("op_535"), val = tensor([1, 1, 1])]; + tensor var_536_cast_fp16 = reshape(shape = var_535, x = valid_len_to_fp16)[name = tensor("op_536_cast_fp16")]; + tensor valid_end_1_cast_fp16 = add(x = var_513_cast_fp16, y = var_536_cast_fp16)[name = tensor("valid_end_1_cast_fp16")]; + tensor k_positions_1_promoted_to_fp16 = const()[name = tensor("k_positions_1_promoted_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(7352512)))]; + tensor valid_mask_1_cast_fp16 = less(x = k_positions_1_promoted_to_fp16, y = valid_end_1_cast_fp16)[name = tensor("valid_mask_1_cast_fp16")]; + tensor causal_mask_1_cast_fp16 = less_equal(x = k_positions_1_promoted_to_fp16, y = q_positions_1_cast_fp16)[name = tensor("causal_mask_1_cast_fp16")]; + tensor var_541 = logical_and(x = valid_mask_1_cast_fp16, y = causal_mask_1_cast_fp16)[name = tensor("op_541")]; + tensor attn_mask_1_axes_0 = const()[name = tensor("attn_mask_1_axes_0"), val = tensor([1])]; + tensor attn_mask_1 = expand_dims(axes = attn_mask_1_axes_0, x = var_541)[name = tensor("attn_mask_1")]; + tensor var_555_transpose_x_0 = const()[name = tensor("op_555_transpose_x_0"), val = tensor(false)]; + tensor var_555_transpose_y_0 = const()[name = tensor("op_555_transpose_y_0"), val = tensor(false)]; + tensor transpose_69_perm_0 = const()[name = tensor("transpose_69_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_70_perm_0 = const()[name = tensor("transpose_70_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_70 = transpose(perm = transpose_70_perm_0, x = keys_3_cast_fp16)[name = tensor("transpose_204")]; + tensor transpose_69 = transpose(perm = transpose_69_perm_0, x = q_3_cast_fp16)[name = tensor("transpose_205")]; + tensor var_555_cast_fp16 = matmul(transpose_x = var_555_transpose_x_0, transpose_y = var_555_transpose_y_0, x = transpose_69, y = transpose_70)[name = tensor("op_555_cast_fp16")]; + tensor var_549_to_fp16 = const()[name = tensor("op_549_to_fp16"), val = tensor([0x1p-3])]; + tensor attn_weights_1_cast_fp16 = mul(x = var_555_cast_fp16, y = var_549_to_fp16)[name = tensor("attn_weights_1_cast_fp16")]; + tensor var_557 = logical_not(x = attn_mask_1)[name = tensor("op_557")]; + tensor var_558_to_fp16 = const()[name = tensor("op_558_to_fp16"), val = tensor(-inf)]; + tensor attn_weights_3_cast_fp16 = select(a = var_558_to_fp16, b = attn_weights_1_cast_fp16, cond = var_557)[name = tensor("attn_weights_3_cast_fp16")]; + tensor var_560 = const()[name = tensor("op_560"), val = tensor(-1)]; + tensor attn_weights_5_cast_fp16 = softmax(axis = var_560, x = attn_weights_3_cast_fp16)[name = tensor("attn_weights_5_cast_fp16")]; + tensor attn_output_1_transpose_x_0 = const()[name = tensor("attn_output_1_transpose_x_0"), val = tensor(false)]; + tensor attn_output_1_transpose_y_0 = const()[name = tensor("attn_output_1_transpose_y_0"), val = tensor(false)]; + tensor values_5_cast_fp16 = transpose(perm = var_496, x = values_3_cast_fp16)[name = tensor("transpose_206")]; + tensor attn_output_1_cast_fp16 = matmul(transpose_x = attn_output_1_transpose_x_0, transpose_y = attn_output_1_transpose_y_0, x = attn_weights_5_cast_fp16, y = values_5_cast_fp16)[name = tensor("attn_output_1_cast_fp16")]; + tensor var_568 = const()[name = tensor("op_568"), val = tensor([0, 2, 1, 3])]; + tensor var_573 = const()[name = tensor("op_573"), val = tensor([1, 256, 1024])]; + tensor var_569_cast_fp16 = transpose(perm = var_568, x = attn_output_1_cast_fp16)[name = tensor("transpose_203")]; + tensor input_5_cast_fp16 = reshape(shape = var_573, x = var_569_cast_fp16)[name = tensor("input_5_cast_fp16")]; + tensor attn0_out_proj_weight_to_fp16 = const()[name = tensor("attn0_out_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(7353600)))]; + tensor linear_1_bias_0_to_fp16 = const()[name = tensor("linear_1_bias_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9450816)))]; + tensor linear_1_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = attn0_out_proj_weight_to_fp16, x = input_5_cast_fp16)[name = tensor("linear_1_cast_fp16")]; + tensor var_582_cast_fp16 = add(x = position0_to_fp16, y = valid_len_to_fp16)[name = tensor("op_582_cast_fp16")]; + tensor var_582_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_582_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor input_7_cast_fp16 = add(x = conditioning_to_fp16, y = linear_1_cast_fp16)[name = tensor("input_7_cast_fp16")]; + tensor input_9_axes_0 = const()[name = tensor("input_9_axes_0"), val = tensor([-1])]; + tensor norm0_2_weight_to_fp16 = const()[name = tensor("norm0_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9452928)))]; + tensor norm0_2_bias_to_fp16 = const()[name = tensor("norm0_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9455040)))]; + tensor var_586_to_fp16 = const()[name = tensor("op_586_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_9_cast_fp16 = layer_norm(axes = input_9_axes_0, beta = norm0_2_bias_to_fp16, epsilon = var_586_to_fp16, gamma = norm0_2_weight_to_fp16, x = input_7_cast_fp16)[name = tensor("input_9_cast_fp16")]; + tensor linear0_1_weight_to_fp16 = const()[name = tensor("linear0_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9457152)))]; + tensor linear_2_bias_0_to_fp16 = const()[name = tensor("linear_2_bias_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(17845824)))]; + tensor linear_2_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = linear0_1_weight_to_fp16, x = input_9_cast_fp16)[name = tensor("linear_2_cast_fp16")]; + tensor input_11_mode_0 = const()[name = tensor("input_11_mode_0"), val = tensor("EXACT")]; + tensor input_11_cast_fp16 = gelu(mode = input_11_mode_0, x = linear_2_cast_fp16)[name = tensor("input_11_cast_fp16")]; + tensor linear0_2_weight_to_fp16 = const()[name = tensor("linear0_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(17854080)))]; + tensor linear_3_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = linear0_2_weight_to_fp16, x = input_11_cast_fp16)[name = tensor("linear_3_cast_fp16")]; + tensor input_13_cast_fp16 = add(x = input_7_cast_fp16, y = linear_3_cast_fp16)[name = tensor("input_13_cast_fp16")]; + tensor input_15_axes_0 = const()[name = tensor("input_15_axes_0"), val = tensor([-1])]; + tensor norm1_1_weight_to_fp16 = const()[name = tensor("norm1_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(26242752)))]; + tensor norm1_1_bias_to_fp16 = const()[name = tensor("norm1_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(26244864)))]; + tensor var_603_to_fp16 = const()[name = tensor("op_603_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_15_cast_fp16 = layer_norm(axes = input_15_axes_0, beta = norm1_1_bias_to_fp16, epsilon = var_603_to_fp16, gamma = norm1_1_weight_to_fp16, x = input_13_cast_fp16)[name = tensor("input_15_cast_fp16")]; + tensor attn1_in_proj_weight_to_fp16 = const()[name = tensor("attn1_in_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(26246976)))]; + tensor linear_4_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = attn1_in_proj_weight_to_fp16, x = input_15_cast_fp16)[name = tensor("linear_4_cast_fp16")]; + tensor var_617 = const()[name = tensor("op_617"), val = tensor([1, 256, 3, 16, 64])]; + tensor qkv_3_cast_fp16 = reshape(shape = var_617, x = linear_4_cast_fp16)[name = tensor("qkv_3_cast_fp16")]; + tensor q_7_begin_0 = const()[name = tensor("q_7_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor q_7_end_0 = const()[name = tensor("q_7_end_0"), val = tensor([1, 256, 1, 16, 64])]; + tensor q_7_end_mask_0 = const()[name = tensor("q_7_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor q_7_squeeze_mask_0 = const()[name = tensor("q_7_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor q_7_cast_fp16 = slice_by_index(begin = q_7_begin_0, end = q_7_end_0, end_mask = q_7_end_mask_0, squeeze_mask = q_7_squeeze_mask_0, x = qkv_3_cast_fp16)[name = tensor("q_7_cast_fp16")]; + tensor k_5_begin_0 = const()[name = tensor("k_5_begin_0"), val = tensor([0, 0, 1, 0, 0])]; + tensor k_5_end_0 = const()[name = tensor("k_5_end_0"), val = tensor([1, 256, 2, 16, 64])]; + tensor k_5_end_mask_0 = const()[name = tensor("k_5_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor k_5_squeeze_mask_0 = const()[name = tensor("k_5_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor k_5_cast_fp16 = slice_by_index(begin = k_5_begin_0, end = k_5_end_0, end_mask = k_5_end_mask_0, squeeze_mask = k_5_squeeze_mask_0, x = qkv_3_cast_fp16)[name = tensor("k_5_cast_fp16")]; + tensor v_3_begin_0 = const()[name = tensor("v_3_begin_0"), val = tensor([0, 0, 2, 0, 0])]; + tensor v_3_end_0 = const()[name = tensor("v_3_end_0"), val = tensor([1, 256, 3, 16, 64])]; + tensor v_3_end_mask_0 = const()[name = tensor("v_3_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor v_3_squeeze_mask_0 = const()[name = tensor("v_3_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor v_3_cast_fp16 = slice_by_index(begin = v_3_begin_0, end = v_3_end_0, end_mask = v_3_end_mask_0, squeeze_mask = v_3_squeeze_mask_0, x = qkv_3_cast_fp16)[name = tensor("v_3_cast_fp16")]; + tensor var_675 = const()[name = tensor("op_675"), val = tensor([1, 1])]; + tensor position1_to_fp16_dtype_0 = const()[name = tensor("position1_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor position1_to_fp16 = cast(dtype = position1_to_fp16_dtype_0, x = position1)[name = tensor("cast_656")]; + tensor var_676_cast_fp16 = reshape(shape = var_675, x = position1_to_fp16)[name = tensor("op_676_cast_fp16")]; + tensor ts_9_cast_fp16 = add(x = ts_1_promoted_to_fp16, y = var_676_cast_fp16)[name = tensor("ts_9_cast_fp16")]; + tensor var_683 = const()[name = tensor("op_683"), val = tensor([1, 256, 1, 1])]; + tensor ts_11_cast_fp16 = reshape(shape = var_683, x = ts_9_cast_fp16)[name = tensor("ts_11_cast_fp16")]; + tensor var_690 = const()[name = tensor("op_690"), val = tensor([1, 256, 16, 32, 2])]; + tensor q_complex_3_cast_fp16 = reshape(shape = var_690, x = q_7_cast_fp16)[name = tensor("q_complex_3_cast_fp16")]; + tensor var_697 = const()[name = tensor("op_697"), val = tensor([1, 256, 16, 32, 2])]; + tensor k_complex_3_cast_fp16 = reshape(shape = var_697, x = k_5_cast_fp16)[name = tensor("k_complex_3_cast_fp16")]; + tensor var_701_begin_0 = const()[name = tensor("op_701_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_701_end_0 = const()[name = tensor("op_701_end_0"), val = tensor([1, 256, 16, 32, 1])]; + tensor var_701_end_mask_0 = const()[name = tensor("op_701_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_701_squeeze_mask_0 = const()[name = tensor("op_701_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_701_cast_fp16 = slice_by_index(begin = var_701_begin_0, end = var_701_end_0, end_mask = var_701_end_mask_0, squeeze_mask = var_701_squeeze_mask_0, x = q_complex_3_cast_fp16)[name = tensor("op_701_cast_fp16")]; + tensor var_709_begin_0 = const()[name = tensor("op_709_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_709_end_0 = const()[name = tensor("op_709_end_0"), val = tensor([1, 256, 16, 32, 2])]; + tensor var_709_end_mask_0 = const()[name = tensor("op_709_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_709_squeeze_mask_0 = const()[name = tensor("op_709_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_709_cast_fp16 = slice_by_index(begin = var_709_begin_0, end = var_709_end_0, end_mask = var_709_end_mask_0, squeeze_mask = var_709_squeeze_mask_0, x = q_complex_3_cast_fp16)[name = tensor("op_709_cast_fp16")]; + tensor var_717_begin_0 = const()[name = tensor("op_717_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_717_end_0 = const()[name = tensor("op_717_end_0"), val = tensor([1, 256, 16, 32, 1])]; + tensor var_717_end_mask_0 = const()[name = tensor("op_717_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_717_squeeze_mask_0 = const()[name = tensor("op_717_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_717_cast_fp16 = slice_by_index(begin = var_717_begin_0, end = var_717_end_0, end_mask = var_717_end_mask_0, squeeze_mask = var_717_squeeze_mask_0, x = k_complex_3_cast_fp16)[name = tensor("op_717_cast_fp16")]; + tensor var_725_begin_0 = const()[name = tensor("op_725_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_725_end_0 = const()[name = tensor("op_725_end_0"), val = tensor([1, 256, 16, 32, 2])]; + tensor var_725_end_mask_0 = const()[name = tensor("op_725_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_725_squeeze_mask_0 = const()[name = tensor("op_725_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_725_cast_fp16 = slice_by_index(begin = var_725_begin_0, end = var_725_end_0, end_mask = var_725_end_mask_0, squeeze_mask = var_725_squeeze_mask_0, x = k_complex_3_cast_fp16)[name = tensor("op_725_cast_fp16")]; + tensor freqs_3_to_fp16 = const()[name = tensor("freqs_3_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(32538496)))]; + tensor var_731_cast_fp16 = mul(x = freqs_3_to_fp16, y = ts_11_cast_fp16)[name = tensor("op_731_cast_fp16")]; + tensor rotr_3_cast_fp16 = cos(x = var_731_cast_fp16)[name = tensor("rotr_3_cast_fp16")]; + tensor roti_3_cast_fp16 = sin(x = var_731_cast_fp16)[name = tensor("roti_3_cast_fp16")]; + tensor var_735_cast_fp16 = mul(x = var_701_cast_fp16, y = rotr_3_cast_fp16)[name = tensor("op_735_cast_fp16")]; + tensor var_736_cast_fp16 = mul(x = var_709_cast_fp16, y = roti_3_cast_fp16)[name = tensor("op_736_cast_fp16")]; + tensor qor_5_cast_fp16 = sub(x = var_735_cast_fp16, y = var_736_cast_fp16)[name = tensor("qor_5_cast_fp16")]; + tensor var_739_cast_fp16 = mul(x = var_701_cast_fp16, y = roti_3_cast_fp16)[name = tensor("op_739_cast_fp16")]; + tensor var_740_cast_fp16 = mul(x = var_709_cast_fp16, y = rotr_3_cast_fp16)[name = tensor("op_740_cast_fp16")]; + tensor qoi_5_cast_fp16 = add(x = var_739_cast_fp16, y = var_740_cast_fp16)[name = tensor("qoi_5_cast_fp16")]; + tensor var_743_cast_fp16 = mul(x = var_717_cast_fp16, y = rotr_3_cast_fp16)[name = tensor("op_743_cast_fp16")]; + tensor var_744_cast_fp16 = mul(x = var_725_cast_fp16, y = roti_3_cast_fp16)[name = tensor("op_744_cast_fp16")]; + tensor kor_5_cast_fp16 = sub(x = var_743_cast_fp16, y = var_744_cast_fp16)[name = tensor("kor_5_cast_fp16")]; + tensor var_747_cast_fp16 = mul(x = var_717_cast_fp16, y = roti_3_cast_fp16)[name = tensor("op_747_cast_fp16")]; + tensor var_748_cast_fp16 = mul(x = var_725_cast_fp16, y = rotr_3_cast_fp16)[name = tensor("op_748_cast_fp16")]; + tensor koi_5_cast_fp16 = add(x = var_747_cast_fp16, y = var_748_cast_fp16)[name = tensor("koi_5_cast_fp16")]; + tensor qo_3_axis_0 = const()[name = tensor("qo_3_axis_0"), val = tensor(-1)]; + tensor qo_3_cast_fp16 = stack(axis = qo_3_axis_0, values = (qor_5_cast_fp16, qoi_5_cast_fp16))[name = tensor("qo_3_cast_fp16")]; + tensor ko_3_axis_0 = const()[name = tensor("ko_3_axis_0"), val = tensor(-1)]; + tensor ko_3_cast_fp16 = stack(axis = ko_3_axis_0, values = (kor_5_cast_fp16, koi_5_cast_fp16))[name = tensor("ko_3_cast_fp16")]; + tensor var_781 = const()[name = tensor("op_781"), val = tensor([1, 256, 16, 64])]; + tensor q_9_cast_fp16 = reshape(shape = var_781, x = qo_3_cast_fp16)[name = tensor("q_9_cast_fp16")]; + tensor var_787 = const()[name = tensor("op_787"), val = tensor([1, 256, 16, 64])]; + tensor k_7_cast_fp16 = reshape(shape = var_787, x = ko_3_cast_fp16)[name = tensor("k_7_cast_fp16")]; + tensor write_indices_float_7_cast_fp16 = add(x = var_676_cast_fp16, y = var_393_promoted_to_fp16)[name = tensor("write_indices_float_7_cast_fp16")]; + tensor _inversed_808_y_0_to_fp16 = const()[name = tensor("_inversed_808_y_0_to_fp16"), val = tensor(0x1p-9)]; + tensor _inversed_808_cast_fp16 = mul(x = write_indices_float_7_cast_fp16, y = _inversed_808_y_0_to_fp16)[name = tensor("_inversed_808_cast_fp16")]; + tensor var_809_cast_fp16 = floor(x = _inversed_808_cast_fp16)[name = tensor("op_809_cast_fp16")]; + tensor var_810_to_fp16 = const()[name = tensor("op_810_to_fp16"), val = tensor(0x1p+9)]; + tensor var_811_cast_fp16 = mul(x = var_809_cast_fp16, y = var_810_to_fp16)[name = tensor("op_811_cast_fp16")]; + tensor write_indices_float_9_cast_fp16 = sub(x = write_indices_float_7_cast_fp16, y = var_811_cast_fp16)[name = tensor("write_indices_float_9_cast_fp16")]; + tensor var_829_value_0_to_fp16 = const()[name = tensor("op_829_value_0_to_fp16"), val = tensor(0x1.ffp+8)]; + tensor var_829_cast_fp16 = fill_like(ref_tensor = write_indices_float_9_cast_fp16, value = var_829_value_0_to_fp16)[name = tensor("op_829_cast_fp16")]; + tensor write_indices_float_11_cast_fp16 = select(a = write_indices_float_9_cast_fp16, b = var_829_cast_fp16, cond = token_valid_1_cast_fp16)[name = tensor("write_indices_float_11_cast_fp16")]; + tensor var_835_dtype_0 = const()[name = tensor("op_835_dtype_0"), val = tensor("int32")]; + tensor var_840 = const()[name = tensor("op_840"), val = tensor([1, 256, 1, 1])]; + tensor write_indices_float_11_cast_fp16_to_int32 = cast(dtype = var_835_dtype_0, x = write_indices_float_11_cast_fp16)[name = tensor("cast_655")]; + tensor var_841 = reshape(shape = var_840, x = write_indices_float_11_cast_fp16_to_int32)[name = tensor("op_841")]; + tensor write_indices_3_reps_0 = const()[name = tensor("write_indices_3_reps_0"), val = tensor([1, 1, 16, 64])]; + tensor write_indices_3 = tile(reps = write_indices_3_reps_0, x = var_841)[name = tensor("write_indices_3")]; + tensor var_851_begin_0 = const()[name = tensor("op_851_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_851_end_0 = const()[name = tensor("op_851_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor var_851_end_mask_0 = const()[name = tensor("op_851_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_851_squeeze_mask_0 = const()[name = tensor("op_851_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor cache1_to_fp16_dtype_0 = const()[name = tensor("cache1_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor cache1_to_fp16 = cast(dtype = cache1_to_fp16_dtype_0, x = cache1)[name = tensor("cast_654")]; + tensor var_851_cast_fp16 = slice_by_index(begin = var_851_begin_0, end = var_851_end_0, end_mask = var_851_end_mask_0, squeeze_mask = var_851_squeeze_mask_0, x = cache1_to_fp16)[name = tensor("op_851_cast_fp16")]; + tensor var_853_axis_0 = const()[name = tensor("op_853_axis_0"), val = tensor(1)]; + tensor var_853_mode_0 = const()[name = tensor("op_853_mode_0"), val = tensor("update")]; + tensor var_853_validate_indices_0 = const()[name = tensor("op_853_validate_indices_0"), val = tensor(false)]; + tensor var_853_cast_fp16 = scatter_along_axis(axis = var_853_axis_0, data = var_851_cast_fp16, indices = write_indices_3, mode = var_853_mode_0, updates = k_7_cast_fp16, validate_indices = var_853_validate_indices_0)[name = tensor("op_853_cast_fp16")]; + tensor concat_10 = const()[name = tensor("concat_10"), val = tensor([0, 0, 0, 0, 0])]; + tensor concat_11 = const()[name = tensor("concat_11"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_3_internal_tensor_assign_1_stride_0 = const()[name = tensor("new_cache_3_internal_tensor_assign_1_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_3_internal_tensor_assign_1_begin_mask_0 = const()[name = tensor("new_cache_3_internal_tensor_assign_1_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_3_internal_tensor_assign_1_end_mask_0 = const()[name = tensor("new_cache_3_internal_tensor_assign_1_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_3_internal_tensor_assign_1_squeeze_mask_0 = const()[name = tensor("new_cache_3_internal_tensor_assign_1_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_48 = const()[name = tensor("shape_48"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_2 = const()[name = tensor("reduce_prod_2"), val = tensor(1048576)]; + tensor range_1d_2_start_0 = const()[name = tensor("range_1d_2_start_0"), val = tensor(0)]; + tensor range_1d_2_step_0 = const()[name = tensor("range_1d_2_step_0"), val = tensor(1)]; + tensor range_1d_2 = range_1d(end = reduce_prod_2, start = range_1d_2_start_0, step = range_1d_2_step_0)[name = tensor("range_1d_2")]; + tensor reshape_10 = reshape(shape = shape_48, x = range_1d_2)[name = tensor("reshape_10")]; + tensor slice_by_index_2 = slice_by_index(begin = concat_10, begin_mask = new_cache_3_internal_tensor_assign_1_begin_mask_0, end = concat_11, end_mask = new_cache_3_internal_tensor_assign_1_end_mask_0, squeeze_mask = new_cache_3_internal_tensor_assign_1_squeeze_mask_0, stride = new_cache_3_internal_tensor_assign_1_stride_0, x = reshape_10)[name = tensor("slice_by_index_2")]; + tensor reshape_11_shape_0 = const()[name = tensor("reshape_11_shape_0"), val = tensor([-1])]; + tensor reshape_11 = reshape(shape = reshape_11_shape_0, x = slice_by_index_2)[name = tensor("reshape_11")]; + tensor reshape_12_shape_0 = const()[name = tensor("reshape_12_shape_0"), val = tensor([-1])]; + tensor reshape_12_cast_fp16 = reshape(shape = reshape_12_shape_0, x = var_853_cast_fp16)[name = tensor("reshape_12_cast_fp16")]; + tensor reshape_13_shape_0 = const()[name = tensor("reshape_13_shape_0"), val = tensor([-1])]; + tensor reshape_13_cast_fp16 = reshape(shape = reshape_13_shape_0, x = cache1_to_fp16)[name = tensor("reshape_13_cast_fp16")]; + tensor scatter_2_mode_0 = const()[name = tensor("scatter_2_mode_0"), val = tensor("update")]; + tensor scatter_2_axis_0 = const()[name = tensor("scatter_2_axis_0"), val = tensor(0)]; + tensor scatter_2_validate_indices_0 = const()[name = tensor("scatter_2_validate_indices_0"), val = tensor(false)]; + tensor scatter_2_cast_fp16 = scatter(axis = scatter_2_axis_0, data = reshape_13_cast_fp16, indices = reshape_11, mode = scatter_2_mode_0, updates = reshape_12_cast_fp16, validate_indices = scatter_2_validate_indices_0)[name = tensor("scatter_2_cast_fp16")]; + tensor reshape_14_cast_fp16 = reshape(shape = shape_48, x = scatter_2_cast_fp16)[name = tensor("reshape_14_cast_fp16")]; + tensor var_861_begin_0 = const()[name = tensor("op_861_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor var_861_end_0 = const()[name = tensor("op_861_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor var_861_end_mask_0 = const()[name = tensor("op_861_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_861_squeeze_mask_0 = const()[name = tensor("op_861_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor var_861_cast_fp16 = slice_by_index(begin = var_861_begin_0, end = var_861_end_0, end_mask = var_861_end_mask_0, squeeze_mask = var_861_squeeze_mask_0, x = reshape_14_cast_fp16)[name = tensor("op_861_cast_fp16")]; + tensor var_863_axis_0 = const()[name = tensor("op_863_axis_0"), val = tensor(1)]; + tensor var_863_mode_0 = const()[name = tensor("op_863_mode_0"), val = tensor("update")]; + tensor var_863_validate_indices_0 = const()[name = tensor("op_863_validate_indices_0"), val = tensor(false)]; + tensor var_863_cast_fp16 = scatter_along_axis(axis = var_863_axis_0, data = var_861_cast_fp16, indices = write_indices_3, mode = var_863_mode_0, updates = v_3_cast_fp16, validate_indices = var_863_validate_indices_0)[name = tensor("op_863_cast_fp16")]; + tensor concat_12 = const()[name = tensor("concat_12"), val = tensor([1, 0, 0, 0, 0])]; + tensor concat_13 = const()[name = tensor("concat_13"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_3_internal_tensor_assign_2_stride_0 = const()[name = tensor("new_cache_3_internal_tensor_assign_2_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_3_internal_tensor_assign_2_begin_mask_0 = const()[name = tensor("new_cache_3_internal_tensor_assign_2_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_3_internal_tensor_assign_2_end_mask_0 = const()[name = tensor("new_cache_3_internal_tensor_assign_2_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_3_internal_tensor_assign_2_squeeze_mask_0 = const()[name = tensor("new_cache_3_internal_tensor_assign_2_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_49 = const()[name = tensor("shape_49"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_3 = const()[name = tensor("reduce_prod_3"), val = tensor(1048576)]; + tensor range_1d_3_start_0 = const()[name = tensor("range_1d_3_start_0"), val = tensor(0)]; + tensor range_1d_3_step_0 = const()[name = tensor("range_1d_3_step_0"), val = tensor(1)]; + tensor range_1d_3 = range_1d(end = reduce_prod_3, start = range_1d_3_start_0, step = range_1d_3_step_0)[name = tensor("range_1d_3")]; + tensor reshape_15 = reshape(shape = shape_49, x = range_1d_3)[name = tensor("reshape_15")]; + tensor slice_by_index_3 = slice_by_index(begin = concat_12, begin_mask = new_cache_3_internal_tensor_assign_2_begin_mask_0, end = concat_13, end_mask = new_cache_3_internal_tensor_assign_2_end_mask_0, squeeze_mask = new_cache_3_internal_tensor_assign_2_squeeze_mask_0, stride = new_cache_3_internal_tensor_assign_2_stride_0, x = reshape_15)[name = tensor("slice_by_index_3")]; + tensor reshape_16_shape_0 = const()[name = tensor("reshape_16_shape_0"), val = tensor([-1])]; + tensor reshape_16 = reshape(shape = reshape_16_shape_0, x = slice_by_index_3)[name = tensor("reshape_16")]; + tensor reshape_17_shape_0 = const()[name = tensor("reshape_17_shape_0"), val = tensor([-1])]; + tensor reshape_17_cast_fp16 = reshape(shape = reshape_17_shape_0, x = var_863_cast_fp16)[name = tensor("reshape_17_cast_fp16")]; + tensor reshape_18_shape_0 = const()[name = tensor("reshape_18_shape_0"), val = tensor([-1])]; + tensor reshape_18_cast_fp16 = reshape(shape = reshape_18_shape_0, x = reshape_14_cast_fp16)[name = tensor("reshape_18_cast_fp16")]; + tensor scatter_3_mode_0 = const()[name = tensor("scatter_3_mode_0"), val = tensor("update")]; + tensor scatter_3_axis_0 = const()[name = tensor("scatter_3_axis_0"), val = tensor(0)]; + tensor scatter_3_validate_indices_0 = const()[name = tensor("scatter_3_validate_indices_0"), val = tensor(false)]; + tensor scatter_3_cast_fp16 = scatter(axis = scatter_3_axis_0, data = reshape_18_cast_fp16, indices = reshape_16, mode = scatter_3_mode_0, updates = reshape_17_cast_fp16, validate_indices = scatter_3_validate_indices_0)[name = tensor("scatter_3_cast_fp16")]; + tensor reshape_19_cast_fp16 = reshape(shape = shape_49, x = scatter_3_cast_fp16)[name = tensor("reshape_19_cast_fp16")]; + tensor reshape_19_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("reshape_19_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor keys_7_begin_0 = const()[name = tensor("keys_7_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor keys_7_end_0 = const()[name = tensor("keys_7_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor keys_7_end_mask_0 = const()[name = tensor("keys_7_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor keys_7_squeeze_mask_0 = const()[name = tensor("keys_7_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor keys_7_cast_fp16 = slice_by_index(begin = keys_7_begin_0, end = keys_7_end_0, end_mask = keys_7_end_mask_0, squeeze_mask = keys_7_squeeze_mask_0, x = reshape_19_cast_fp16)[name = tensor("keys_7_cast_fp16")]; + tensor values_7_begin_0 = const()[name = tensor("values_7_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor values_7_end_0 = const()[name = tensor("values_7_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor values_7_end_mask_0 = const()[name = tensor("values_7_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor values_7_squeeze_mask_0 = const()[name = tensor("values_7_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor values_7_cast_fp16 = slice_by_index(begin = values_7_begin_0, end = values_7_end_0, end_mask = values_7_end_mask_0, squeeze_mask = values_7_squeeze_mask_0, x = reshape_19_cast_fp16)[name = tensor("values_7_cast_fp16")]; + tensor var_875_cast_fp16 = not_equal(x = keys_7_cast_fp16, y = keys_7_cast_fp16)[name = tensor("op_875_cast_fp16")]; + tensor keys_9_cast_fp16 = select(a = var_470_to_fp16, b = keys_7_cast_fp16, cond = var_875_cast_fp16)[name = tensor("keys_9_cast_fp16")]; + tensor var_883_cast_fp16 = not_equal(x = values_7_cast_fp16, y = values_7_cast_fp16)[name = tensor("op_883_cast_fp16")]; + tensor values_9_cast_fp16 = select(a = var_470_to_fp16, b = values_7_cast_fp16, cond = var_883_cast_fp16)[name = tensor("values_9_cast_fp16")]; + tensor var_907 = const()[name = tensor("op_907"), val = tensor([0, 2, 1, 3])]; + tensor var_923 = const()[name = tensor("op_923"), val = tensor([1, 1, 1])]; + tensor var_924_cast_fp16 = reshape(shape = var_923, x = position1_to_fp16)[name = tensor("op_924_cast_fp16")]; + tensor q_positions_3_cast_fp16 = add(x = var_924_cast_fp16, y = q_offsets_1_promoted_to_fp16)[name = tensor("q_positions_3_cast_fp16")]; + tensor valid_end_3_cast_fp16 = add(x = var_924_cast_fp16, y = var_536_cast_fp16)[name = tensor("valid_end_3_cast_fp16")]; + tensor valid_mask_3_cast_fp16 = less(x = k_positions_1_promoted_to_fp16, y = valid_end_3_cast_fp16)[name = tensor("valid_mask_3_cast_fp16")]; + tensor causal_mask_3_cast_fp16 = less_equal(x = k_positions_1_promoted_to_fp16, y = q_positions_3_cast_fp16)[name = tensor("causal_mask_3_cast_fp16")]; + tensor var_952 = logical_and(x = valid_mask_3_cast_fp16, y = causal_mask_3_cast_fp16)[name = tensor("op_952")]; + tensor attn_mask_3_axes_0 = const()[name = tensor("attn_mask_3_axes_0"), val = tensor([1])]; + tensor attn_mask_3 = expand_dims(axes = attn_mask_3_axes_0, x = var_952)[name = tensor("attn_mask_3")]; + tensor var_966_transpose_x_0 = const()[name = tensor("op_966_transpose_x_0"), val = tensor(false)]; + tensor var_966_transpose_y_0 = const()[name = tensor("op_966_transpose_y_0"), val = tensor(false)]; + tensor transpose_71_perm_0 = const()[name = tensor("transpose_71_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_72_perm_0 = const()[name = tensor("transpose_72_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_72 = transpose(perm = transpose_72_perm_0, x = keys_9_cast_fp16)[name = tensor("transpose_200")]; + tensor transpose_71 = transpose(perm = transpose_71_perm_0, x = q_9_cast_fp16)[name = tensor("transpose_201")]; + tensor var_966_cast_fp16 = matmul(transpose_x = var_966_transpose_x_0, transpose_y = var_966_transpose_y_0, x = transpose_71, y = transpose_72)[name = tensor("op_966_cast_fp16")]; + tensor var_960_to_fp16 = const()[name = tensor("op_960_to_fp16"), val = tensor([0x1p-3])]; + tensor attn_weights_7_cast_fp16 = mul(x = var_966_cast_fp16, y = var_960_to_fp16)[name = tensor("attn_weights_7_cast_fp16")]; + tensor var_968 = logical_not(x = attn_mask_3)[name = tensor("op_968")]; + tensor var_969_to_fp16 = const()[name = tensor("op_969_to_fp16"), val = tensor(-inf)]; + tensor attn_weights_9_cast_fp16 = select(a = var_969_to_fp16, b = attn_weights_7_cast_fp16, cond = var_968)[name = tensor("attn_weights_9_cast_fp16")]; + tensor var_971 = const()[name = tensor("op_971"), val = tensor(-1)]; + tensor attn_weights_11_cast_fp16 = softmax(axis = var_971, x = attn_weights_9_cast_fp16)[name = tensor("attn_weights_11_cast_fp16")]; + tensor attn_output_3_transpose_x_0 = const()[name = tensor("attn_output_3_transpose_x_0"), val = tensor(false)]; + tensor attn_output_3_transpose_y_0 = const()[name = tensor("attn_output_3_transpose_y_0"), val = tensor(false)]; + tensor values_11_cast_fp16 = transpose(perm = var_907, x = values_9_cast_fp16)[name = tensor("transpose_202")]; + tensor attn_output_3_cast_fp16 = matmul(transpose_x = attn_output_3_transpose_x_0, transpose_y = attn_output_3_transpose_y_0, x = attn_weights_11_cast_fp16, y = values_11_cast_fp16)[name = tensor("attn_output_3_cast_fp16")]; + tensor var_979 = const()[name = tensor("op_979"), val = tensor([0, 2, 1, 3])]; + tensor var_984 = const()[name = tensor("op_984"), val = tensor([1, 256, 1024])]; + tensor var_980_cast_fp16 = transpose(perm = var_979, x = attn_output_3_cast_fp16)[name = tensor("transpose_199")]; + tensor input_17_cast_fp16 = reshape(shape = var_984, x = var_980_cast_fp16)[name = tensor("input_17_cast_fp16")]; + tensor attn1_out_proj_weight_to_fp16 = const()[name = tensor("attn1_out_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(32538624)))]; + tensor linear_5_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = attn1_out_proj_weight_to_fp16, x = input_17_cast_fp16)[name = tensor("linear_5_cast_fp16")]; + tensor var_993_cast_fp16 = add(x = position1_to_fp16, y = valid_len_to_fp16)[name = tensor("op_993_cast_fp16")]; + tensor var_993_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_993_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor input_19_cast_fp16 = add(x = input_13_cast_fp16, y = linear_5_cast_fp16)[name = tensor("input_19_cast_fp16")]; + tensor input_21_axes_0 = const()[name = tensor("input_21_axes_0"), val = tensor([-1])]; + tensor norm1_2_weight_to_fp16 = const()[name = tensor("norm1_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(34635840)))]; + tensor norm1_2_bias_to_fp16 = const()[name = tensor("norm1_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(34637952)))]; + tensor var_997_to_fp16 = const()[name = tensor("op_997_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_21_cast_fp16 = layer_norm(axes = input_21_axes_0, beta = norm1_2_bias_to_fp16, epsilon = var_997_to_fp16, gamma = norm1_2_weight_to_fp16, x = input_19_cast_fp16)[name = tensor("input_21_cast_fp16")]; + tensor linear1_1_weight_to_fp16 = const()[name = tensor("linear1_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(34640064)))]; + tensor linear_6_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = linear1_1_weight_to_fp16, x = input_21_cast_fp16)[name = tensor("linear_6_cast_fp16")]; + tensor input_23_mode_0 = const()[name = tensor("input_23_mode_0"), val = tensor("EXACT")]; + tensor input_23_cast_fp16 = gelu(mode = input_23_mode_0, x = linear_6_cast_fp16)[name = tensor("input_23_cast_fp16")]; + tensor linear1_2_weight_to_fp16 = const()[name = tensor("linear1_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(43028736)))]; + tensor linear_7_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = linear1_2_weight_to_fp16, x = input_23_cast_fp16)[name = tensor("linear_7_cast_fp16")]; + tensor input_25_cast_fp16 = add(x = input_19_cast_fp16, y = linear_7_cast_fp16)[name = tensor("input_25_cast_fp16")]; + tensor input_27_axes_0 = const()[name = tensor("input_27_axes_0"), val = tensor([-1])]; + tensor norm2_1_weight_to_fp16 = const()[name = tensor("norm2_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(51417408)))]; + tensor norm2_1_bias_to_fp16 = const()[name = tensor("norm2_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(51419520)))]; + tensor var_1014_to_fp16 = const()[name = tensor("op_1014_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_27_cast_fp16 = layer_norm(axes = input_27_axes_0, beta = norm2_1_bias_to_fp16, epsilon = var_1014_to_fp16, gamma = norm2_1_weight_to_fp16, x = input_25_cast_fp16)[name = tensor("input_27_cast_fp16")]; + tensor attn2_in_proj_weight_to_fp16 = const()[name = tensor("attn2_in_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(51421632)))]; + tensor linear_8_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = attn2_in_proj_weight_to_fp16, x = input_27_cast_fp16)[name = tensor("linear_8_cast_fp16")]; + tensor var_1028 = const()[name = tensor("op_1028"), val = tensor([1, 256, 3, 16, 64])]; + tensor qkv_5_cast_fp16 = reshape(shape = var_1028, x = linear_8_cast_fp16)[name = tensor("qkv_5_cast_fp16")]; + tensor q_13_begin_0 = const()[name = tensor("q_13_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor q_13_end_0 = const()[name = tensor("q_13_end_0"), val = tensor([1, 256, 1, 16, 64])]; + tensor q_13_end_mask_0 = const()[name = tensor("q_13_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor q_13_squeeze_mask_0 = const()[name = tensor("q_13_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor q_13_cast_fp16 = slice_by_index(begin = q_13_begin_0, end = q_13_end_0, end_mask = q_13_end_mask_0, squeeze_mask = q_13_squeeze_mask_0, x = qkv_5_cast_fp16)[name = tensor("q_13_cast_fp16")]; + tensor k_9_begin_0 = const()[name = tensor("k_9_begin_0"), val = tensor([0, 0, 1, 0, 0])]; + tensor k_9_end_0 = const()[name = tensor("k_9_end_0"), val = tensor([1, 256, 2, 16, 64])]; + tensor k_9_end_mask_0 = const()[name = tensor("k_9_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor k_9_squeeze_mask_0 = const()[name = tensor("k_9_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor k_9_cast_fp16 = slice_by_index(begin = k_9_begin_0, end = k_9_end_0, end_mask = k_9_end_mask_0, squeeze_mask = k_9_squeeze_mask_0, x = qkv_5_cast_fp16)[name = tensor("k_9_cast_fp16")]; + tensor v_5_begin_0 = const()[name = tensor("v_5_begin_0"), val = tensor([0, 0, 2, 0, 0])]; + tensor v_5_end_0 = const()[name = tensor("v_5_end_0"), val = tensor([1, 256, 3, 16, 64])]; + tensor v_5_end_mask_0 = const()[name = tensor("v_5_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor v_5_squeeze_mask_0 = const()[name = tensor("v_5_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor v_5_cast_fp16 = slice_by_index(begin = v_5_begin_0, end = v_5_end_0, end_mask = v_5_end_mask_0, squeeze_mask = v_5_squeeze_mask_0, x = qkv_5_cast_fp16)[name = tensor("v_5_cast_fp16")]; + tensor var_1086 = const()[name = tensor("op_1086"), val = tensor([1, 1])]; + tensor position2_to_fp16_dtype_0 = const()[name = tensor("position2_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor position2_to_fp16 = cast(dtype = position2_to_fp16_dtype_0, x = position2)[name = tensor("cast_651")]; + tensor var_1087_cast_fp16 = reshape(shape = var_1086, x = position2_to_fp16)[name = tensor("op_1087_cast_fp16")]; + tensor ts_15_cast_fp16 = add(x = ts_1_promoted_to_fp16, y = var_1087_cast_fp16)[name = tensor("ts_15_cast_fp16")]; + tensor var_1094 = const()[name = tensor("op_1094"), val = tensor([1, 256, 1, 1])]; + tensor ts_17_cast_fp16 = reshape(shape = var_1094, x = ts_15_cast_fp16)[name = tensor("ts_17_cast_fp16")]; + tensor var_1101 = const()[name = tensor("op_1101"), val = tensor([1, 256, 16, 32, 2])]; + tensor q_complex_5_cast_fp16 = reshape(shape = var_1101, x = q_13_cast_fp16)[name = tensor("q_complex_5_cast_fp16")]; + tensor var_1108 = const()[name = tensor("op_1108"), val = tensor([1, 256, 16, 32, 2])]; + tensor k_complex_5_cast_fp16 = reshape(shape = var_1108, x = k_9_cast_fp16)[name = tensor("k_complex_5_cast_fp16")]; + tensor var_1112_begin_0 = const()[name = tensor("op_1112_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_1112_end_0 = const()[name = tensor("op_1112_end_0"), val = tensor([1, 256, 16, 32, 1])]; + tensor var_1112_end_mask_0 = const()[name = tensor("op_1112_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_1112_squeeze_mask_0 = const()[name = tensor("op_1112_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_1112_cast_fp16 = slice_by_index(begin = var_1112_begin_0, end = var_1112_end_0, end_mask = var_1112_end_mask_0, squeeze_mask = var_1112_squeeze_mask_0, x = q_complex_5_cast_fp16)[name = tensor("op_1112_cast_fp16")]; + tensor var_1120_begin_0 = const()[name = tensor("op_1120_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_1120_end_0 = const()[name = tensor("op_1120_end_0"), val = tensor([1, 256, 16, 32, 2])]; + tensor var_1120_end_mask_0 = const()[name = tensor("op_1120_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_1120_squeeze_mask_0 = const()[name = tensor("op_1120_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_1120_cast_fp16 = slice_by_index(begin = var_1120_begin_0, end = var_1120_end_0, end_mask = var_1120_end_mask_0, squeeze_mask = var_1120_squeeze_mask_0, x = q_complex_5_cast_fp16)[name = tensor("op_1120_cast_fp16")]; + tensor var_1128_begin_0 = const()[name = tensor("op_1128_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_1128_end_0 = const()[name = tensor("op_1128_end_0"), val = tensor([1, 256, 16, 32, 1])]; + tensor var_1128_end_mask_0 = const()[name = tensor("op_1128_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_1128_squeeze_mask_0 = const()[name = tensor("op_1128_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_1128_cast_fp16 = slice_by_index(begin = var_1128_begin_0, end = var_1128_end_0, end_mask = var_1128_end_mask_0, squeeze_mask = var_1128_squeeze_mask_0, x = k_complex_5_cast_fp16)[name = tensor("op_1128_cast_fp16")]; + tensor var_1136_begin_0 = const()[name = tensor("op_1136_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_1136_end_0 = const()[name = tensor("op_1136_end_0"), val = tensor([1, 256, 16, 32, 2])]; + tensor var_1136_end_mask_0 = const()[name = tensor("op_1136_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_1136_squeeze_mask_0 = const()[name = tensor("op_1136_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_1136_cast_fp16 = slice_by_index(begin = var_1136_begin_0, end = var_1136_end_0, end_mask = var_1136_end_mask_0, squeeze_mask = var_1136_squeeze_mask_0, x = k_complex_5_cast_fp16)[name = tensor("op_1136_cast_fp16")]; + tensor freqs_5_to_fp16 = const()[name = tensor("freqs_5_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(57713152)))]; + tensor var_1142_cast_fp16 = mul(x = freqs_5_to_fp16, y = ts_17_cast_fp16)[name = tensor("op_1142_cast_fp16")]; + tensor rotr_5_cast_fp16 = cos(x = var_1142_cast_fp16)[name = tensor("rotr_5_cast_fp16")]; + tensor roti_5_cast_fp16 = sin(x = var_1142_cast_fp16)[name = tensor("roti_5_cast_fp16")]; + tensor var_1146_cast_fp16 = mul(x = var_1112_cast_fp16, y = rotr_5_cast_fp16)[name = tensor("op_1146_cast_fp16")]; + tensor var_1147_cast_fp16 = mul(x = var_1120_cast_fp16, y = roti_5_cast_fp16)[name = tensor("op_1147_cast_fp16")]; + tensor qor_9_cast_fp16 = sub(x = var_1146_cast_fp16, y = var_1147_cast_fp16)[name = tensor("qor_9_cast_fp16")]; + tensor var_1150_cast_fp16 = mul(x = var_1112_cast_fp16, y = roti_5_cast_fp16)[name = tensor("op_1150_cast_fp16")]; + tensor var_1151_cast_fp16 = mul(x = var_1120_cast_fp16, y = rotr_5_cast_fp16)[name = tensor("op_1151_cast_fp16")]; + tensor qoi_9_cast_fp16 = add(x = var_1150_cast_fp16, y = var_1151_cast_fp16)[name = tensor("qoi_9_cast_fp16")]; + tensor var_1154_cast_fp16 = mul(x = var_1128_cast_fp16, y = rotr_5_cast_fp16)[name = tensor("op_1154_cast_fp16")]; + tensor var_1155_cast_fp16 = mul(x = var_1136_cast_fp16, y = roti_5_cast_fp16)[name = tensor("op_1155_cast_fp16")]; + tensor kor_9_cast_fp16 = sub(x = var_1154_cast_fp16, y = var_1155_cast_fp16)[name = tensor("kor_9_cast_fp16")]; + tensor var_1158_cast_fp16 = mul(x = var_1128_cast_fp16, y = roti_5_cast_fp16)[name = tensor("op_1158_cast_fp16")]; + tensor var_1159_cast_fp16 = mul(x = var_1136_cast_fp16, y = rotr_5_cast_fp16)[name = tensor("op_1159_cast_fp16")]; + tensor koi_9_cast_fp16 = add(x = var_1158_cast_fp16, y = var_1159_cast_fp16)[name = tensor("koi_9_cast_fp16")]; + tensor qo_5_axis_0 = const()[name = tensor("qo_5_axis_0"), val = tensor(-1)]; + tensor qo_5_cast_fp16 = stack(axis = qo_5_axis_0, values = (qor_9_cast_fp16, qoi_9_cast_fp16))[name = tensor("qo_5_cast_fp16")]; + tensor ko_5_axis_0 = const()[name = tensor("ko_5_axis_0"), val = tensor(-1)]; + tensor ko_5_cast_fp16 = stack(axis = ko_5_axis_0, values = (kor_9_cast_fp16, koi_9_cast_fp16))[name = tensor("ko_5_cast_fp16")]; + tensor var_1192 = const()[name = tensor("op_1192"), val = tensor([1, 256, 16, 64])]; + tensor q_15_cast_fp16 = reshape(shape = var_1192, x = qo_5_cast_fp16)[name = tensor("q_15_cast_fp16")]; + tensor var_1198 = const()[name = tensor("op_1198"), val = tensor([1, 256, 16, 64])]; + tensor k_11_cast_fp16 = reshape(shape = var_1198, x = ko_5_cast_fp16)[name = tensor("k_11_cast_fp16")]; + tensor write_indices_float_13_cast_fp16 = add(x = var_1087_cast_fp16, y = var_393_promoted_to_fp16)[name = tensor("write_indices_float_13_cast_fp16")]; + tensor _inversed_1219_y_0_to_fp16 = const()[name = tensor("_inversed_1219_y_0_to_fp16"), val = tensor(0x1p-9)]; + tensor _inversed_1219_cast_fp16 = mul(x = write_indices_float_13_cast_fp16, y = _inversed_1219_y_0_to_fp16)[name = tensor("_inversed_1219_cast_fp16")]; + tensor var_1220_cast_fp16 = floor(x = _inversed_1219_cast_fp16)[name = tensor("op_1220_cast_fp16")]; + tensor var_1221_to_fp16 = const()[name = tensor("op_1221_to_fp16"), val = tensor(0x1p+9)]; + tensor var_1222_cast_fp16 = mul(x = var_1220_cast_fp16, y = var_1221_to_fp16)[name = tensor("op_1222_cast_fp16")]; + tensor write_indices_float_15_cast_fp16 = sub(x = write_indices_float_13_cast_fp16, y = var_1222_cast_fp16)[name = tensor("write_indices_float_15_cast_fp16")]; + tensor var_1240_value_0_to_fp16 = const()[name = tensor("op_1240_value_0_to_fp16"), val = tensor(0x1.ffp+8)]; + tensor var_1240_cast_fp16 = fill_like(ref_tensor = write_indices_float_15_cast_fp16, value = var_1240_value_0_to_fp16)[name = tensor("op_1240_cast_fp16")]; + tensor write_indices_float_17_cast_fp16 = select(a = write_indices_float_15_cast_fp16, b = var_1240_cast_fp16, cond = token_valid_1_cast_fp16)[name = tensor("write_indices_float_17_cast_fp16")]; + tensor var_1246_dtype_0 = const()[name = tensor("op_1246_dtype_0"), val = tensor("int32")]; + tensor var_1251 = const()[name = tensor("op_1251"), val = tensor([1, 256, 1, 1])]; + tensor write_indices_float_17_cast_fp16_to_int32 = cast(dtype = var_1246_dtype_0, x = write_indices_float_17_cast_fp16)[name = tensor("cast_650")]; + tensor var_1252 = reshape(shape = var_1251, x = write_indices_float_17_cast_fp16_to_int32)[name = tensor("op_1252")]; + tensor write_indices_5_reps_0 = const()[name = tensor("write_indices_5_reps_0"), val = tensor([1, 1, 16, 64])]; + tensor write_indices_5 = tile(reps = write_indices_5_reps_0, x = var_1252)[name = tensor("write_indices_5")]; + tensor var_1262_begin_0 = const()[name = tensor("op_1262_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_1262_end_0 = const()[name = tensor("op_1262_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor var_1262_end_mask_0 = const()[name = tensor("op_1262_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_1262_squeeze_mask_0 = const()[name = tensor("op_1262_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor cache2_to_fp16_dtype_0 = const()[name = tensor("cache2_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor cache2_to_fp16 = cast(dtype = cache2_to_fp16_dtype_0, x = cache2)[name = tensor("cast_649")]; + tensor var_1262_cast_fp16 = slice_by_index(begin = var_1262_begin_0, end = var_1262_end_0, end_mask = var_1262_end_mask_0, squeeze_mask = var_1262_squeeze_mask_0, x = cache2_to_fp16)[name = tensor("op_1262_cast_fp16")]; + tensor var_1264_axis_0 = const()[name = tensor("op_1264_axis_0"), val = tensor(1)]; + tensor var_1264_mode_0 = const()[name = tensor("op_1264_mode_0"), val = tensor("update")]; + tensor var_1264_validate_indices_0 = const()[name = tensor("op_1264_validate_indices_0"), val = tensor(false)]; + tensor var_1264_cast_fp16 = scatter_along_axis(axis = var_1264_axis_0, data = var_1262_cast_fp16, indices = write_indices_5, mode = var_1264_mode_0, updates = k_11_cast_fp16, validate_indices = var_1264_validate_indices_0)[name = tensor("op_1264_cast_fp16")]; + tensor concat_18 = const()[name = tensor("concat_18"), val = tensor([0, 0, 0, 0, 0])]; + tensor concat_19 = const()[name = tensor("concat_19"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_5_internal_tensor_assign_1_stride_0 = const()[name = tensor("new_cache_5_internal_tensor_assign_1_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_5_internal_tensor_assign_1_begin_mask_0 = const()[name = tensor("new_cache_5_internal_tensor_assign_1_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_5_internal_tensor_assign_1_end_mask_0 = const()[name = tensor("new_cache_5_internal_tensor_assign_1_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_5_internal_tensor_assign_1_squeeze_mask_0 = const()[name = tensor("new_cache_5_internal_tensor_assign_1_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_50 = const()[name = tensor("shape_50"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_4 = const()[name = tensor("reduce_prod_4"), val = tensor(1048576)]; + tensor range_1d_4_start_0 = const()[name = tensor("range_1d_4_start_0"), val = tensor(0)]; + tensor range_1d_4_step_0 = const()[name = tensor("range_1d_4_step_0"), val = tensor(1)]; + tensor range_1d_4 = range_1d(end = reduce_prod_4, start = range_1d_4_start_0, step = range_1d_4_step_0)[name = tensor("range_1d_4")]; + tensor reshape_20 = reshape(shape = shape_50, x = range_1d_4)[name = tensor("reshape_20")]; + tensor slice_by_index_4 = slice_by_index(begin = concat_18, begin_mask = new_cache_5_internal_tensor_assign_1_begin_mask_0, end = concat_19, end_mask = new_cache_5_internal_tensor_assign_1_end_mask_0, squeeze_mask = new_cache_5_internal_tensor_assign_1_squeeze_mask_0, stride = new_cache_5_internal_tensor_assign_1_stride_0, x = reshape_20)[name = tensor("slice_by_index_4")]; + tensor reshape_21_shape_0 = const()[name = tensor("reshape_21_shape_0"), val = tensor([-1])]; + tensor reshape_21 = reshape(shape = reshape_21_shape_0, x = slice_by_index_4)[name = tensor("reshape_21")]; + tensor reshape_22_shape_0 = const()[name = tensor("reshape_22_shape_0"), val = tensor([-1])]; + tensor reshape_22_cast_fp16 = reshape(shape = reshape_22_shape_0, x = var_1264_cast_fp16)[name = tensor("reshape_22_cast_fp16")]; + tensor reshape_23_shape_0 = const()[name = tensor("reshape_23_shape_0"), val = tensor([-1])]; + tensor reshape_23_cast_fp16 = reshape(shape = reshape_23_shape_0, x = cache2_to_fp16)[name = tensor("reshape_23_cast_fp16")]; + tensor scatter_4_mode_0 = const()[name = tensor("scatter_4_mode_0"), val = tensor("update")]; + tensor scatter_4_axis_0 = const()[name = tensor("scatter_4_axis_0"), val = tensor(0)]; + tensor scatter_4_validate_indices_0 = const()[name = tensor("scatter_4_validate_indices_0"), val = tensor(false)]; + tensor scatter_4_cast_fp16 = scatter(axis = scatter_4_axis_0, data = reshape_23_cast_fp16, indices = reshape_21, mode = scatter_4_mode_0, updates = reshape_22_cast_fp16, validate_indices = scatter_4_validate_indices_0)[name = tensor("scatter_4_cast_fp16")]; + tensor reshape_24_cast_fp16 = reshape(shape = shape_50, x = scatter_4_cast_fp16)[name = tensor("reshape_24_cast_fp16")]; + tensor var_1272_begin_0 = const()[name = tensor("op_1272_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor var_1272_end_0 = const()[name = tensor("op_1272_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor var_1272_end_mask_0 = const()[name = tensor("op_1272_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_1272_squeeze_mask_0 = const()[name = tensor("op_1272_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor var_1272_cast_fp16 = slice_by_index(begin = var_1272_begin_0, end = var_1272_end_0, end_mask = var_1272_end_mask_0, squeeze_mask = var_1272_squeeze_mask_0, x = reshape_24_cast_fp16)[name = tensor("op_1272_cast_fp16")]; + tensor var_1274_axis_0 = const()[name = tensor("op_1274_axis_0"), val = tensor(1)]; + tensor var_1274_mode_0 = const()[name = tensor("op_1274_mode_0"), val = tensor("update")]; + tensor var_1274_validate_indices_0 = const()[name = tensor("op_1274_validate_indices_0"), val = tensor(false)]; + tensor var_1274_cast_fp16 = scatter_along_axis(axis = var_1274_axis_0, data = var_1272_cast_fp16, indices = write_indices_5, mode = var_1274_mode_0, updates = v_5_cast_fp16, validate_indices = var_1274_validate_indices_0)[name = tensor("op_1274_cast_fp16")]; + tensor concat_20 = const()[name = tensor("concat_20"), val = tensor([1, 0, 0, 0, 0])]; + tensor concat_21 = const()[name = tensor("concat_21"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_5_internal_tensor_assign_2_stride_0 = const()[name = tensor("new_cache_5_internal_tensor_assign_2_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_5_internal_tensor_assign_2_begin_mask_0 = const()[name = tensor("new_cache_5_internal_tensor_assign_2_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_5_internal_tensor_assign_2_end_mask_0 = const()[name = tensor("new_cache_5_internal_tensor_assign_2_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_5_internal_tensor_assign_2_squeeze_mask_0 = const()[name = tensor("new_cache_5_internal_tensor_assign_2_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_51 = const()[name = tensor("shape_51"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_5 = const()[name = tensor("reduce_prod_5"), val = tensor(1048576)]; + tensor range_1d_5_start_0 = const()[name = tensor("range_1d_5_start_0"), val = tensor(0)]; + tensor range_1d_5_step_0 = const()[name = tensor("range_1d_5_step_0"), val = tensor(1)]; + tensor range_1d_5 = range_1d(end = reduce_prod_5, start = range_1d_5_start_0, step = range_1d_5_step_0)[name = tensor("range_1d_5")]; + tensor reshape_25 = reshape(shape = shape_51, x = range_1d_5)[name = tensor("reshape_25")]; + tensor slice_by_index_5 = slice_by_index(begin = concat_20, begin_mask = new_cache_5_internal_tensor_assign_2_begin_mask_0, end = concat_21, end_mask = new_cache_5_internal_tensor_assign_2_end_mask_0, squeeze_mask = new_cache_5_internal_tensor_assign_2_squeeze_mask_0, stride = new_cache_5_internal_tensor_assign_2_stride_0, x = reshape_25)[name = tensor("slice_by_index_5")]; + tensor reshape_26_shape_0 = const()[name = tensor("reshape_26_shape_0"), val = tensor([-1])]; + tensor reshape_26 = reshape(shape = reshape_26_shape_0, x = slice_by_index_5)[name = tensor("reshape_26")]; + tensor reshape_27_shape_0 = const()[name = tensor("reshape_27_shape_0"), val = tensor([-1])]; + tensor reshape_27_cast_fp16 = reshape(shape = reshape_27_shape_0, x = var_1274_cast_fp16)[name = tensor("reshape_27_cast_fp16")]; + tensor reshape_28_shape_0 = const()[name = tensor("reshape_28_shape_0"), val = tensor([-1])]; + tensor reshape_28_cast_fp16 = reshape(shape = reshape_28_shape_0, x = reshape_24_cast_fp16)[name = tensor("reshape_28_cast_fp16")]; + tensor scatter_5_mode_0 = const()[name = tensor("scatter_5_mode_0"), val = tensor("update")]; + tensor scatter_5_axis_0 = const()[name = tensor("scatter_5_axis_0"), val = tensor(0)]; + tensor scatter_5_validate_indices_0 = const()[name = tensor("scatter_5_validate_indices_0"), val = tensor(false)]; + tensor scatter_5_cast_fp16 = scatter(axis = scatter_5_axis_0, data = reshape_28_cast_fp16, indices = reshape_26, mode = scatter_5_mode_0, updates = reshape_27_cast_fp16, validate_indices = scatter_5_validate_indices_0)[name = tensor("scatter_5_cast_fp16")]; + tensor reshape_29_cast_fp16 = reshape(shape = shape_51, x = scatter_5_cast_fp16)[name = tensor("reshape_29_cast_fp16")]; + tensor reshape_29_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("reshape_29_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor keys_13_begin_0 = const()[name = tensor("keys_13_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor keys_13_end_0 = const()[name = tensor("keys_13_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor keys_13_end_mask_0 = const()[name = tensor("keys_13_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor keys_13_squeeze_mask_0 = const()[name = tensor("keys_13_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor keys_13_cast_fp16 = slice_by_index(begin = keys_13_begin_0, end = keys_13_end_0, end_mask = keys_13_end_mask_0, squeeze_mask = keys_13_squeeze_mask_0, x = reshape_29_cast_fp16)[name = tensor("keys_13_cast_fp16")]; + tensor values_13_begin_0 = const()[name = tensor("values_13_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor values_13_end_0 = const()[name = tensor("values_13_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor values_13_end_mask_0 = const()[name = tensor("values_13_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor values_13_squeeze_mask_0 = const()[name = tensor("values_13_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor values_13_cast_fp16 = slice_by_index(begin = values_13_begin_0, end = values_13_end_0, end_mask = values_13_end_mask_0, squeeze_mask = values_13_squeeze_mask_0, x = reshape_29_cast_fp16)[name = tensor("values_13_cast_fp16")]; + tensor var_1286_cast_fp16 = not_equal(x = keys_13_cast_fp16, y = keys_13_cast_fp16)[name = tensor("op_1286_cast_fp16")]; + tensor keys_15_cast_fp16 = select(a = var_470_to_fp16, b = keys_13_cast_fp16, cond = var_1286_cast_fp16)[name = tensor("keys_15_cast_fp16")]; + tensor var_1294_cast_fp16 = not_equal(x = values_13_cast_fp16, y = values_13_cast_fp16)[name = tensor("op_1294_cast_fp16")]; + tensor values_15_cast_fp16 = select(a = var_470_to_fp16, b = values_13_cast_fp16, cond = var_1294_cast_fp16)[name = tensor("values_15_cast_fp16")]; + tensor var_1318 = const()[name = tensor("op_1318"), val = tensor([0, 2, 1, 3])]; + tensor var_1334 = const()[name = tensor("op_1334"), val = tensor([1, 1, 1])]; + tensor var_1335_cast_fp16 = reshape(shape = var_1334, x = position2_to_fp16)[name = tensor("op_1335_cast_fp16")]; + tensor q_positions_5_cast_fp16 = add(x = var_1335_cast_fp16, y = q_offsets_1_promoted_to_fp16)[name = tensor("q_positions_5_cast_fp16")]; + tensor valid_end_5_cast_fp16 = add(x = var_1335_cast_fp16, y = var_536_cast_fp16)[name = tensor("valid_end_5_cast_fp16")]; + tensor valid_mask_5_cast_fp16 = less(x = k_positions_1_promoted_to_fp16, y = valid_end_5_cast_fp16)[name = tensor("valid_mask_5_cast_fp16")]; + tensor causal_mask_5_cast_fp16 = less_equal(x = k_positions_1_promoted_to_fp16, y = q_positions_5_cast_fp16)[name = tensor("causal_mask_5_cast_fp16")]; + tensor var_1363 = logical_and(x = valid_mask_5_cast_fp16, y = causal_mask_5_cast_fp16)[name = tensor("op_1363")]; + tensor attn_mask_5_axes_0 = const()[name = tensor("attn_mask_5_axes_0"), val = tensor([1])]; + tensor attn_mask_5 = expand_dims(axes = attn_mask_5_axes_0, x = var_1363)[name = tensor("attn_mask_5")]; + tensor var_1377_transpose_x_0 = const()[name = tensor("op_1377_transpose_x_0"), val = tensor(false)]; + tensor var_1377_transpose_y_0 = const()[name = tensor("op_1377_transpose_y_0"), val = tensor(false)]; + tensor transpose_73_perm_0 = const()[name = tensor("transpose_73_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_74_perm_0 = const()[name = tensor("transpose_74_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_74 = transpose(perm = transpose_74_perm_0, x = keys_15_cast_fp16)[name = tensor("transpose_196")]; + tensor transpose_73 = transpose(perm = transpose_73_perm_0, x = q_15_cast_fp16)[name = tensor("transpose_197")]; + tensor var_1377_cast_fp16 = matmul(transpose_x = var_1377_transpose_x_0, transpose_y = var_1377_transpose_y_0, x = transpose_73, y = transpose_74)[name = tensor("op_1377_cast_fp16")]; + tensor var_1371_to_fp16 = const()[name = tensor("op_1371_to_fp16"), val = tensor([0x1p-3])]; + tensor attn_weights_13_cast_fp16 = mul(x = var_1377_cast_fp16, y = var_1371_to_fp16)[name = tensor("attn_weights_13_cast_fp16")]; + tensor var_1379 = logical_not(x = attn_mask_5)[name = tensor("op_1379")]; + tensor var_1380_to_fp16 = const()[name = tensor("op_1380_to_fp16"), val = tensor(-inf)]; + tensor attn_weights_15_cast_fp16 = select(a = var_1380_to_fp16, b = attn_weights_13_cast_fp16, cond = var_1379)[name = tensor("attn_weights_15_cast_fp16")]; + tensor var_1382 = const()[name = tensor("op_1382"), val = tensor(-1)]; + tensor attn_weights_17_cast_fp16 = softmax(axis = var_1382, x = attn_weights_15_cast_fp16)[name = tensor("attn_weights_17_cast_fp16")]; + tensor attn_output_5_transpose_x_0 = const()[name = tensor("attn_output_5_transpose_x_0"), val = tensor(false)]; + tensor attn_output_5_transpose_y_0 = const()[name = tensor("attn_output_5_transpose_y_0"), val = tensor(false)]; + tensor values_17_cast_fp16 = transpose(perm = var_1318, x = values_15_cast_fp16)[name = tensor("transpose_198")]; + tensor attn_output_5_cast_fp16 = matmul(transpose_x = attn_output_5_transpose_x_0, transpose_y = attn_output_5_transpose_y_0, x = attn_weights_17_cast_fp16, y = values_17_cast_fp16)[name = tensor("attn_output_5_cast_fp16")]; + tensor var_1390 = const()[name = tensor("op_1390"), val = tensor([0, 2, 1, 3])]; + tensor var_1395 = const()[name = tensor("op_1395"), val = tensor([1, 256, 1024])]; + tensor var_1391_cast_fp16 = transpose(perm = var_1390, x = attn_output_5_cast_fp16)[name = tensor("transpose_195")]; + tensor input_29_cast_fp16 = reshape(shape = var_1395, x = var_1391_cast_fp16)[name = tensor("input_29_cast_fp16")]; + tensor attn2_out_proj_weight_to_fp16 = const()[name = tensor("attn2_out_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(57713280)))]; + tensor linear_9_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = attn2_out_proj_weight_to_fp16, x = input_29_cast_fp16)[name = tensor("linear_9_cast_fp16")]; + tensor var_1404_cast_fp16 = add(x = position2_to_fp16, y = valid_len_to_fp16)[name = tensor("op_1404_cast_fp16")]; + tensor var_1404_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_1404_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor input_31_cast_fp16 = add(x = input_25_cast_fp16, y = linear_9_cast_fp16)[name = tensor("input_31_cast_fp16")]; + tensor input_33_axes_0 = const()[name = tensor("input_33_axes_0"), val = tensor([-1])]; + tensor norm2_2_weight_to_fp16 = const()[name = tensor("norm2_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(59810496)))]; + tensor norm2_2_bias_to_fp16 = const()[name = tensor("norm2_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(59812608)))]; + tensor var_1408_to_fp16 = const()[name = tensor("op_1408_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_33_cast_fp16 = layer_norm(axes = input_33_axes_0, beta = norm2_2_bias_to_fp16, epsilon = var_1408_to_fp16, gamma = norm2_2_weight_to_fp16, x = input_31_cast_fp16)[name = tensor("input_33_cast_fp16")]; + tensor linear2_1_weight_to_fp16 = const()[name = tensor("linear2_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(59814720)))]; + tensor linear_10_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = linear2_1_weight_to_fp16, x = input_33_cast_fp16)[name = tensor("linear_10_cast_fp16")]; + tensor input_35_mode_0 = const()[name = tensor("input_35_mode_0"), val = tensor("EXACT")]; + tensor input_35_cast_fp16 = gelu(mode = input_35_mode_0, x = linear_10_cast_fp16)[name = tensor("input_35_cast_fp16")]; + tensor linear2_2_weight_to_fp16 = const()[name = tensor("linear2_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(68203392)))]; + tensor linear_11_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = linear2_2_weight_to_fp16, x = input_35_cast_fp16)[name = tensor("linear_11_cast_fp16")]; + tensor input_37_cast_fp16 = add(x = input_31_cast_fp16, y = linear_11_cast_fp16)[name = tensor("input_37_cast_fp16")]; + tensor input_39_axes_0 = const()[name = tensor("input_39_axes_0"), val = tensor([-1])]; + tensor norm3_1_weight_to_fp16 = const()[name = tensor("norm3_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(76592064)))]; + tensor norm3_1_bias_to_fp16 = const()[name = tensor("norm3_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(76594176)))]; + tensor var_1425_to_fp16 = const()[name = tensor("op_1425_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_39_cast_fp16 = layer_norm(axes = input_39_axes_0, beta = norm3_1_bias_to_fp16, epsilon = var_1425_to_fp16, gamma = norm3_1_weight_to_fp16, x = input_37_cast_fp16)[name = tensor("input_39_cast_fp16")]; + tensor attn3_in_proj_weight_to_fp16 = const()[name = tensor("attn3_in_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(76596288)))]; + tensor linear_12_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = attn3_in_proj_weight_to_fp16, x = input_39_cast_fp16)[name = tensor("linear_12_cast_fp16")]; + tensor var_1439 = const()[name = tensor("op_1439"), val = tensor([1, 256, 3, 16, 64])]; + tensor qkv_7_cast_fp16 = reshape(shape = var_1439, x = linear_12_cast_fp16)[name = tensor("qkv_7_cast_fp16")]; + tensor q_19_begin_0 = const()[name = tensor("q_19_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor q_19_end_0 = const()[name = tensor("q_19_end_0"), val = tensor([1, 256, 1, 16, 64])]; + tensor q_19_end_mask_0 = const()[name = tensor("q_19_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor q_19_squeeze_mask_0 = const()[name = tensor("q_19_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor q_19_cast_fp16 = slice_by_index(begin = q_19_begin_0, end = q_19_end_0, end_mask = q_19_end_mask_0, squeeze_mask = q_19_squeeze_mask_0, x = qkv_7_cast_fp16)[name = tensor("q_19_cast_fp16")]; + tensor k_13_begin_0 = const()[name = tensor("k_13_begin_0"), val = tensor([0, 0, 1, 0, 0])]; + tensor k_13_end_0 = const()[name = tensor("k_13_end_0"), val = tensor([1, 256, 2, 16, 64])]; + tensor k_13_end_mask_0 = const()[name = tensor("k_13_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor k_13_squeeze_mask_0 = const()[name = tensor("k_13_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor k_13_cast_fp16 = slice_by_index(begin = k_13_begin_0, end = k_13_end_0, end_mask = k_13_end_mask_0, squeeze_mask = k_13_squeeze_mask_0, x = qkv_7_cast_fp16)[name = tensor("k_13_cast_fp16")]; + tensor v_7_begin_0 = const()[name = tensor("v_7_begin_0"), val = tensor([0, 0, 2, 0, 0])]; + tensor v_7_end_0 = const()[name = tensor("v_7_end_0"), val = tensor([1, 256, 3, 16, 64])]; + tensor v_7_end_mask_0 = const()[name = tensor("v_7_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor v_7_squeeze_mask_0 = const()[name = tensor("v_7_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor v_7_cast_fp16 = slice_by_index(begin = v_7_begin_0, end = v_7_end_0, end_mask = v_7_end_mask_0, squeeze_mask = v_7_squeeze_mask_0, x = qkv_7_cast_fp16)[name = tensor("v_7_cast_fp16")]; + tensor var_1497 = const()[name = tensor("op_1497"), val = tensor([1, 1])]; + tensor position3_to_fp16_dtype_0 = const()[name = tensor("position3_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor position3_to_fp16 = cast(dtype = position3_to_fp16_dtype_0, x = position3)[name = tensor("cast_646")]; + tensor var_1498_cast_fp16 = reshape(shape = var_1497, x = position3_to_fp16)[name = tensor("op_1498_cast_fp16")]; + tensor ts_21_cast_fp16 = add(x = ts_1_promoted_to_fp16, y = var_1498_cast_fp16)[name = tensor("ts_21_cast_fp16")]; + tensor var_1505 = const()[name = tensor("op_1505"), val = tensor([1, 256, 1, 1])]; + tensor ts_23_cast_fp16 = reshape(shape = var_1505, x = ts_21_cast_fp16)[name = tensor("ts_23_cast_fp16")]; + tensor var_1512 = const()[name = tensor("op_1512"), val = tensor([1, 256, 16, 32, 2])]; + tensor q_complex_7_cast_fp16 = reshape(shape = var_1512, x = q_19_cast_fp16)[name = tensor("q_complex_7_cast_fp16")]; + tensor var_1519 = const()[name = tensor("op_1519"), val = tensor([1, 256, 16, 32, 2])]; + tensor k_complex_7_cast_fp16 = reshape(shape = var_1519, x = k_13_cast_fp16)[name = tensor("k_complex_7_cast_fp16")]; + tensor var_1523_begin_0 = const()[name = tensor("op_1523_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_1523_end_0 = const()[name = tensor("op_1523_end_0"), val = tensor([1, 256, 16, 32, 1])]; + tensor var_1523_end_mask_0 = const()[name = tensor("op_1523_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_1523_squeeze_mask_0 = const()[name = tensor("op_1523_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_1523_cast_fp16 = slice_by_index(begin = var_1523_begin_0, end = var_1523_end_0, end_mask = var_1523_end_mask_0, squeeze_mask = var_1523_squeeze_mask_0, x = q_complex_7_cast_fp16)[name = tensor("op_1523_cast_fp16")]; + tensor var_1531_begin_0 = const()[name = tensor("op_1531_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_1531_end_0 = const()[name = tensor("op_1531_end_0"), val = tensor([1, 256, 16, 32, 2])]; + tensor var_1531_end_mask_0 = const()[name = tensor("op_1531_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_1531_squeeze_mask_0 = const()[name = tensor("op_1531_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_1531_cast_fp16 = slice_by_index(begin = var_1531_begin_0, end = var_1531_end_0, end_mask = var_1531_end_mask_0, squeeze_mask = var_1531_squeeze_mask_0, x = q_complex_7_cast_fp16)[name = tensor("op_1531_cast_fp16")]; + tensor var_1539_begin_0 = const()[name = tensor("op_1539_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_1539_end_0 = const()[name = tensor("op_1539_end_0"), val = tensor([1, 256, 16, 32, 1])]; + tensor var_1539_end_mask_0 = const()[name = tensor("op_1539_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_1539_squeeze_mask_0 = const()[name = tensor("op_1539_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_1539_cast_fp16 = slice_by_index(begin = var_1539_begin_0, end = var_1539_end_0, end_mask = var_1539_end_mask_0, squeeze_mask = var_1539_squeeze_mask_0, x = k_complex_7_cast_fp16)[name = tensor("op_1539_cast_fp16")]; + tensor var_1547_begin_0 = const()[name = tensor("op_1547_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_1547_end_0 = const()[name = tensor("op_1547_end_0"), val = tensor([1, 256, 16, 32, 2])]; + tensor var_1547_end_mask_0 = const()[name = tensor("op_1547_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_1547_squeeze_mask_0 = const()[name = tensor("op_1547_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_1547_cast_fp16 = slice_by_index(begin = var_1547_begin_0, end = var_1547_end_0, end_mask = var_1547_end_mask_0, squeeze_mask = var_1547_squeeze_mask_0, x = k_complex_7_cast_fp16)[name = tensor("op_1547_cast_fp16")]; + tensor freqs_7_to_fp16 = const()[name = tensor("freqs_7_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(82887808)))]; + tensor var_1553_cast_fp16 = mul(x = freqs_7_to_fp16, y = ts_23_cast_fp16)[name = tensor("op_1553_cast_fp16")]; + tensor rotr_7_cast_fp16 = cos(x = var_1553_cast_fp16)[name = tensor("rotr_7_cast_fp16")]; + tensor roti_7_cast_fp16 = sin(x = var_1553_cast_fp16)[name = tensor("roti_7_cast_fp16")]; + tensor var_1557_cast_fp16 = mul(x = var_1523_cast_fp16, y = rotr_7_cast_fp16)[name = tensor("op_1557_cast_fp16")]; + tensor var_1558_cast_fp16 = mul(x = var_1531_cast_fp16, y = roti_7_cast_fp16)[name = tensor("op_1558_cast_fp16")]; + tensor qor_13_cast_fp16 = sub(x = var_1557_cast_fp16, y = var_1558_cast_fp16)[name = tensor("qor_13_cast_fp16")]; + tensor var_1561_cast_fp16 = mul(x = var_1523_cast_fp16, y = roti_7_cast_fp16)[name = tensor("op_1561_cast_fp16")]; + tensor var_1562_cast_fp16 = mul(x = var_1531_cast_fp16, y = rotr_7_cast_fp16)[name = tensor("op_1562_cast_fp16")]; + tensor qoi_13_cast_fp16 = add(x = var_1561_cast_fp16, y = var_1562_cast_fp16)[name = tensor("qoi_13_cast_fp16")]; + tensor var_1565_cast_fp16 = mul(x = var_1539_cast_fp16, y = rotr_7_cast_fp16)[name = tensor("op_1565_cast_fp16")]; + tensor var_1566_cast_fp16 = mul(x = var_1547_cast_fp16, y = roti_7_cast_fp16)[name = tensor("op_1566_cast_fp16")]; + tensor kor_13_cast_fp16 = sub(x = var_1565_cast_fp16, y = var_1566_cast_fp16)[name = tensor("kor_13_cast_fp16")]; + tensor var_1569_cast_fp16 = mul(x = var_1539_cast_fp16, y = roti_7_cast_fp16)[name = tensor("op_1569_cast_fp16")]; + tensor var_1570_cast_fp16 = mul(x = var_1547_cast_fp16, y = rotr_7_cast_fp16)[name = tensor("op_1570_cast_fp16")]; + tensor koi_13_cast_fp16 = add(x = var_1569_cast_fp16, y = var_1570_cast_fp16)[name = tensor("koi_13_cast_fp16")]; + tensor qo_7_axis_0 = const()[name = tensor("qo_7_axis_0"), val = tensor(-1)]; + tensor qo_7_cast_fp16 = stack(axis = qo_7_axis_0, values = (qor_13_cast_fp16, qoi_13_cast_fp16))[name = tensor("qo_7_cast_fp16")]; + tensor ko_7_axis_0 = const()[name = tensor("ko_7_axis_0"), val = tensor(-1)]; + tensor ko_7_cast_fp16 = stack(axis = ko_7_axis_0, values = (kor_13_cast_fp16, koi_13_cast_fp16))[name = tensor("ko_7_cast_fp16")]; + tensor var_1603 = const()[name = tensor("op_1603"), val = tensor([1, 256, 16, 64])]; + tensor q_21_cast_fp16 = reshape(shape = var_1603, x = qo_7_cast_fp16)[name = tensor("q_21_cast_fp16")]; + tensor var_1609 = const()[name = tensor("op_1609"), val = tensor([1, 256, 16, 64])]; + tensor k_15_cast_fp16 = reshape(shape = var_1609, x = ko_7_cast_fp16)[name = tensor("k_15_cast_fp16")]; + tensor write_indices_float_19_cast_fp16 = add(x = var_1498_cast_fp16, y = var_393_promoted_to_fp16)[name = tensor("write_indices_float_19_cast_fp16")]; + tensor _inversed_1630_y_0_to_fp16 = const()[name = tensor("_inversed_1630_y_0_to_fp16"), val = tensor(0x1p-9)]; + tensor _inversed_1630_cast_fp16 = mul(x = write_indices_float_19_cast_fp16, y = _inversed_1630_y_0_to_fp16)[name = tensor("_inversed_1630_cast_fp16")]; + tensor var_1631_cast_fp16 = floor(x = _inversed_1630_cast_fp16)[name = tensor("op_1631_cast_fp16")]; + tensor var_1632_to_fp16 = const()[name = tensor("op_1632_to_fp16"), val = tensor(0x1p+9)]; + tensor var_1633_cast_fp16 = mul(x = var_1631_cast_fp16, y = var_1632_to_fp16)[name = tensor("op_1633_cast_fp16")]; + tensor write_indices_float_21_cast_fp16 = sub(x = write_indices_float_19_cast_fp16, y = var_1633_cast_fp16)[name = tensor("write_indices_float_21_cast_fp16")]; + tensor var_1651_value_0_to_fp16 = const()[name = tensor("op_1651_value_0_to_fp16"), val = tensor(0x1.ffp+8)]; + tensor var_1651_cast_fp16 = fill_like(ref_tensor = write_indices_float_21_cast_fp16, value = var_1651_value_0_to_fp16)[name = tensor("op_1651_cast_fp16")]; + tensor write_indices_float_23_cast_fp16 = select(a = write_indices_float_21_cast_fp16, b = var_1651_cast_fp16, cond = token_valid_1_cast_fp16)[name = tensor("write_indices_float_23_cast_fp16")]; + tensor var_1657_dtype_0 = const()[name = tensor("op_1657_dtype_0"), val = tensor("int32")]; + tensor var_1662 = const()[name = tensor("op_1662"), val = tensor([1, 256, 1, 1])]; + tensor write_indices_float_23_cast_fp16_to_int32 = cast(dtype = var_1657_dtype_0, x = write_indices_float_23_cast_fp16)[name = tensor("cast_645")]; + tensor var_1663 = reshape(shape = var_1662, x = write_indices_float_23_cast_fp16_to_int32)[name = tensor("op_1663")]; + tensor write_indices_7_reps_0 = const()[name = tensor("write_indices_7_reps_0"), val = tensor([1, 1, 16, 64])]; + tensor write_indices_7 = tile(reps = write_indices_7_reps_0, x = var_1663)[name = tensor("write_indices_7")]; + tensor var_1673_begin_0 = const()[name = tensor("op_1673_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_1673_end_0 = const()[name = tensor("op_1673_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor var_1673_end_mask_0 = const()[name = tensor("op_1673_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_1673_squeeze_mask_0 = const()[name = tensor("op_1673_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor cache3_to_fp16_dtype_0 = const()[name = tensor("cache3_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor cache3_to_fp16 = cast(dtype = cache3_to_fp16_dtype_0, x = cache3)[name = tensor("cast_644")]; + tensor var_1673_cast_fp16 = slice_by_index(begin = var_1673_begin_0, end = var_1673_end_0, end_mask = var_1673_end_mask_0, squeeze_mask = var_1673_squeeze_mask_0, x = cache3_to_fp16)[name = tensor("op_1673_cast_fp16")]; + tensor var_1675_axis_0 = const()[name = tensor("op_1675_axis_0"), val = tensor(1)]; + tensor var_1675_mode_0 = const()[name = tensor("op_1675_mode_0"), val = tensor("update")]; + tensor var_1675_validate_indices_0 = const()[name = tensor("op_1675_validate_indices_0"), val = tensor(false)]; + tensor var_1675_cast_fp16 = scatter_along_axis(axis = var_1675_axis_0, data = var_1673_cast_fp16, indices = write_indices_7, mode = var_1675_mode_0, updates = k_15_cast_fp16, validate_indices = var_1675_validate_indices_0)[name = tensor("op_1675_cast_fp16")]; + tensor concat_26 = const()[name = tensor("concat_26"), val = tensor([0, 0, 0, 0, 0])]; + tensor concat_27 = const()[name = tensor("concat_27"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_7_internal_tensor_assign_1_stride_0 = const()[name = tensor("new_cache_7_internal_tensor_assign_1_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_7_internal_tensor_assign_1_begin_mask_0 = const()[name = tensor("new_cache_7_internal_tensor_assign_1_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_7_internal_tensor_assign_1_end_mask_0 = const()[name = tensor("new_cache_7_internal_tensor_assign_1_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_7_internal_tensor_assign_1_squeeze_mask_0 = const()[name = tensor("new_cache_7_internal_tensor_assign_1_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_52 = const()[name = tensor("shape_52"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_6 = const()[name = tensor("reduce_prod_6"), val = tensor(1048576)]; + tensor range_1d_6_start_0 = const()[name = tensor("range_1d_6_start_0"), val = tensor(0)]; + tensor range_1d_6_step_0 = const()[name = tensor("range_1d_6_step_0"), val = tensor(1)]; + tensor range_1d_6 = range_1d(end = reduce_prod_6, start = range_1d_6_start_0, step = range_1d_6_step_0)[name = tensor("range_1d_6")]; + tensor reshape_30 = reshape(shape = shape_52, x = range_1d_6)[name = tensor("reshape_30")]; + tensor slice_by_index_6 = slice_by_index(begin = concat_26, begin_mask = new_cache_7_internal_tensor_assign_1_begin_mask_0, end = concat_27, end_mask = new_cache_7_internal_tensor_assign_1_end_mask_0, squeeze_mask = new_cache_7_internal_tensor_assign_1_squeeze_mask_0, stride = new_cache_7_internal_tensor_assign_1_stride_0, x = reshape_30)[name = tensor("slice_by_index_6")]; + tensor reshape_31_shape_0 = const()[name = tensor("reshape_31_shape_0"), val = tensor([-1])]; + tensor reshape_31 = reshape(shape = reshape_31_shape_0, x = slice_by_index_6)[name = tensor("reshape_31")]; + tensor reshape_32_shape_0 = const()[name = tensor("reshape_32_shape_0"), val = tensor([-1])]; + tensor reshape_32_cast_fp16 = reshape(shape = reshape_32_shape_0, x = var_1675_cast_fp16)[name = tensor("reshape_32_cast_fp16")]; + tensor reshape_33_shape_0 = const()[name = tensor("reshape_33_shape_0"), val = tensor([-1])]; + tensor reshape_33_cast_fp16 = reshape(shape = reshape_33_shape_0, x = cache3_to_fp16)[name = tensor("reshape_33_cast_fp16")]; + tensor scatter_6_mode_0 = const()[name = tensor("scatter_6_mode_0"), val = tensor("update")]; + tensor scatter_6_axis_0 = const()[name = tensor("scatter_6_axis_0"), val = tensor(0)]; + tensor scatter_6_validate_indices_0 = const()[name = tensor("scatter_6_validate_indices_0"), val = tensor(false)]; + tensor scatter_6_cast_fp16 = scatter(axis = scatter_6_axis_0, data = reshape_33_cast_fp16, indices = reshape_31, mode = scatter_6_mode_0, updates = reshape_32_cast_fp16, validate_indices = scatter_6_validate_indices_0)[name = tensor("scatter_6_cast_fp16")]; + tensor reshape_34_cast_fp16 = reshape(shape = shape_52, x = scatter_6_cast_fp16)[name = tensor("reshape_34_cast_fp16")]; + tensor var_1683_begin_0 = const()[name = tensor("op_1683_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor var_1683_end_0 = const()[name = tensor("op_1683_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor var_1683_end_mask_0 = const()[name = tensor("op_1683_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_1683_squeeze_mask_0 = const()[name = tensor("op_1683_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor var_1683_cast_fp16 = slice_by_index(begin = var_1683_begin_0, end = var_1683_end_0, end_mask = var_1683_end_mask_0, squeeze_mask = var_1683_squeeze_mask_0, x = reshape_34_cast_fp16)[name = tensor("op_1683_cast_fp16")]; + tensor var_1685_axis_0 = const()[name = tensor("op_1685_axis_0"), val = tensor(1)]; + tensor var_1685_mode_0 = const()[name = tensor("op_1685_mode_0"), val = tensor("update")]; + tensor var_1685_validate_indices_0 = const()[name = tensor("op_1685_validate_indices_0"), val = tensor(false)]; + tensor var_1685_cast_fp16 = scatter_along_axis(axis = var_1685_axis_0, data = var_1683_cast_fp16, indices = write_indices_7, mode = var_1685_mode_0, updates = v_7_cast_fp16, validate_indices = var_1685_validate_indices_0)[name = tensor("op_1685_cast_fp16")]; + tensor concat_28 = const()[name = tensor("concat_28"), val = tensor([1, 0, 0, 0, 0])]; + tensor concat_29 = const()[name = tensor("concat_29"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_7_internal_tensor_assign_2_stride_0 = const()[name = tensor("new_cache_7_internal_tensor_assign_2_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_7_internal_tensor_assign_2_begin_mask_0 = const()[name = tensor("new_cache_7_internal_tensor_assign_2_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_7_internal_tensor_assign_2_end_mask_0 = const()[name = tensor("new_cache_7_internal_tensor_assign_2_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_7_internal_tensor_assign_2_squeeze_mask_0 = const()[name = tensor("new_cache_7_internal_tensor_assign_2_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_53 = const()[name = tensor("shape_53"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_7 = const()[name = tensor("reduce_prod_7"), val = tensor(1048576)]; + tensor range_1d_7_start_0 = const()[name = tensor("range_1d_7_start_0"), val = tensor(0)]; + tensor range_1d_7_step_0 = const()[name = tensor("range_1d_7_step_0"), val = tensor(1)]; + tensor range_1d_7 = range_1d(end = reduce_prod_7, start = range_1d_7_start_0, step = range_1d_7_step_0)[name = tensor("range_1d_7")]; + tensor reshape_35 = reshape(shape = shape_53, x = range_1d_7)[name = tensor("reshape_35")]; + tensor slice_by_index_7 = slice_by_index(begin = concat_28, begin_mask = new_cache_7_internal_tensor_assign_2_begin_mask_0, end = concat_29, end_mask = new_cache_7_internal_tensor_assign_2_end_mask_0, squeeze_mask = new_cache_7_internal_tensor_assign_2_squeeze_mask_0, stride = new_cache_7_internal_tensor_assign_2_stride_0, x = reshape_35)[name = tensor("slice_by_index_7")]; + tensor reshape_36_shape_0 = const()[name = tensor("reshape_36_shape_0"), val = tensor([-1])]; + tensor reshape_36 = reshape(shape = reshape_36_shape_0, x = slice_by_index_7)[name = tensor("reshape_36")]; + tensor reshape_37_shape_0 = const()[name = tensor("reshape_37_shape_0"), val = tensor([-1])]; + tensor reshape_37_cast_fp16 = reshape(shape = reshape_37_shape_0, x = var_1685_cast_fp16)[name = tensor("reshape_37_cast_fp16")]; + tensor reshape_38_shape_0 = const()[name = tensor("reshape_38_shape_0"), val = tensor([-1])]; + tensor reshape_38_cast_fp16 = reshape(shape = reshape_38_shape_0, x = reshape_34_cast_fp16)[name = tensor("reshape_38_cast_fp16")]; + tensor scatter_7_mode_0 = const()[name = tensor("scatter_7_mode_0"), val = tensor("update")]; + tensor scatter_7_axis_0 = const()[name = tensor("scatter_7_axis_0"), val = tensor(0)]; + tensor scatter_7_validate_indices_0 = const()[name = tensor("scatter_7_validate_indices_0"), val = tensor(false)]; + tensor scatter_7_cast_fp16 = scatter(axis = scatter_7_axis_0, data = reshape_38_cast_fp16, indices = reshape_36, mode = scatter_7_mode_0, updates = reshape_37_cast_fp16, validate_indices = scatter_7_validate_indices_0)[name = tensor("scatter_7_cast_fp16")]; + tensor reshape_39_cast_fp16 = reshape(shape = shape_53, x = scatter_7_cast_fp16)[name = tensor("reshape_39_cast_fp16")]; + tensor reshape_39_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("reshape_39_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor keys_19_begin_0 = const()[name = tensor("keys_19_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor keys_19_end_0 = const()[name = tensor("keys_19_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor keys_19_end_mask_0 = const()[name = tensor("keys_19_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor keys_19_squeeze_mask_0 = const()[name = tensor("keys_19_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor keys_19_cast_fp16 = slice_by_index(begin = keys_19_begin_0, end = keys_19_end_0, end_mask = keys_19_end_mask_0, squeeze_mask = keys_19_squeeze_mask_0, x = reshape_39_cast_fp16)[name = tensor("keys_19_cast_fp16")]; + tensor values_19_begin_0 = const()[name = tensor("values_19_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor values_19_end_0 = const()[name = tensor("values_19_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor values_19_end_mask_0 = const()[name = tensor("values_19_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor values_19_squeeze_mask_0 = const()[name = tensor("values_19_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor values_19_cast_fp16 = slice_by_index(begin = values_19_begin_0, end = values_19_end_0, end_mask = values_19_end_mask_0, squeeze_mask = values_19_squeeze_mask_0, x = reshape_39_cast_fp16)[name = tensor("values_19_cast_fp16")]; + tensor var_1697_cast_fp16 = not_equal(x = keys_19_cast_fp16, y = keys_19_cast_fp16)[name = tensor("op_1697_cast_fp16")]; + tensor keys_21_cast_fp16 = select(a = var_470_to_fp16, b = keys_19_cast_fp16, cond = var_1697_cast_fp16)[name = tensor("keys_21_cast_fp16")]; + tensor var_1705_cast_fp16 = not_equal(x = values_19_cast_fp16, y = values_19_cast_fp16)[name = tensor("op_1705_cast_fp16")]; + tensor values_21_cast_fp16 = select(a = var_470_to_fp16, b = values_19_cast_fp16, cond = var_1705_cast_fp16)[name = tensor("values_21_cast_fp16")]; + tensor var_1729 = const()[name = tensor("op_1729"), val = tensor([0, 2, 1, 3])]; + tensor var_1745 = const()[name = tensor("op_1745"), val = tensor([1, 1, 1])]; + tensor var_1746_cast_fp16 = reshape(shape = var_1745, x = position3_to_fp16)[name = tensor("op_1746_cast_fp16")]; + tensor q_positions_7_cast_fp16 = add(x = var_1746_cast_fp16, y = q_offsets_1_promoted_to_fp16)[name = tensor("q_positions_7_cast_fp16")]; + tensor valid_end_7_cast_fp16 = add(x = var_1746_cast_fp16, y = var_536_cast_fp16)[name = tensor("valid_end_7_cast_fp16")]; + tensor valid_mask_7_cast_fp16 = less(x = k_positions_1_promoted_to_fp16, y = valid_end_7_cast_fp16)[name = tensor("valid_mask_7_cast_fp16")]; + tensor causal_mask_7_cast_fp16 = less_equal(x = k_positions_1_promoted_to_fp16, y = q_positions_7_cast_fp16)[name = tensor("causal_mask_7_cast_fp16")]; + tensor var_1774 = logical_and(x = valid_mask_7_cast_fp16, y = causal_mask_7_cast_fp16)[name = tensor("op_1774")]; + tensor attn_mask_7_axes_0 = const()[name = tensor("attn_mask_7_axes_0"), val = tensor([1])]; + tensor attn_mask_7 = expand_dims(axes = attn_mask_7_axes_0, x = var_1774)[name = tensor("attn_mask_7")]; + tensor var_1788_transpose_x_0 = const()[name = tensor("op_1788_transpose_x_0"), val = tensor(false)]; + tensor var_1788_transpose_y_0 = const()[name = tensor("op_1788_transpose_y_0"), val = tensor(false)]; + tensor transpose_75_perm_0 = const()[name = tensor("transpose_75_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_76_perm_0 = const()[name = tensor("transpose_76_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_76 = transpose(perm = transpose_76_perm_0, x = keys_21_cast_fp16)[name = tensor("transpose_192")]; + tensor transpose_75 = transpose(perm = transpose_75_perm_0, x = q_21_cast_fp16)[name = tensor("transpose_193")]; + tensor var_1788_cast_fp16 = matmul(transpose_x = var_1788_transpose_x_0, transpose_y = var_1788_transpose_y_0, x = transpose_75, y = transpose_76)[name = tensor("op_1788_cast_fp16")]; + tensor var_1782_to_fp16 = const()[name = tensor("op_1782_to_fp16"), val = tensor([0x1p-3])]; + tensor attn_weights_19_cast_fp16 = mul(x = var_1788_cast_fp16, y = var_1782_to_fp16)[name = tensor("attn_weights_19_cast_fp16")]; + tensor var_1790 = logical_not(x = attn_mask_7)[name = tensor("op_1790")]; + tensor var_1791_to_fp16 = const()[name = tensor("op_1791_to_fp16"), val = tensor(-inf)]; + tensor attn_weights_21_cast_fp16 = select(a = var_1791_to_fp16, b = attn_weights_19_cast_fp16, cond = var_1790)[name = tensor("attn_weights_21_cast_fp16")]; + tensor var_1793 = const()[name = tensor("op_1793"), val = tensor(-1)]; + tensor attn_weights_23_cast_fp16 = softmax(axis = var_1793, x = attn_weights_21_cast_fp16)[name = tensor("attn_weights_23_cast_fp16")]; + tensor attn_output_7_transpose_x_0 = const()[name = tensor("attn_output_7_transpose_x_0"), val = tensor(false)]; + tensor attn_output_7_transpose_y_0 = const()[name = tensor("attn_output_7_transpose_y_0"), val = tensor(false)]; + tensor values_23_cast_fp16 = transpose(perm = var_1729, x = values_21_cast_fp16)[name = tensor("transpose_194")]; + tensor attn_output_7_cast_fp16 = matmul(transpose_x = attn_output_7_transpose_x_0, transpose_y = attn_output_7_transpose_y_0, x = attn_weights_23_cast_fp16, y = values_23_cast_fp16)[name = tensor("attn_output_7_cast_fp16")]; + tensor var_1801 = const()[name = tensor("op_1801"), val = tensor([0, 2, 1, 3])]; + tensor var_1806 = const()[name = tensor("op_1806"), val = tensor([1, 256, 1024])]; + tensor var_1802_cast_fp16 = transpose(perm = var_1801, x = attn_output_7_cast_fp16)[name = tensor("transpose_191")]; + tensor input_41_cast_fp16 = reshape(shape = var_1806, x = var_1802_cast_fp16)[name = tensor("input_41_cast_fp16")]; + tensor attn3_out_proj_weight_to_fp16 = const()[name = tensor("attn3_out_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(82887936)))]; + tensor linear_13_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = attn3_out_proj_weight_to_fp16, x = input_41_cast_fp16)[name = tensor("linear_13_cast_fp16")]; + tensor var_1815_cast_fp16 = add(x = position3_to_fp16, y = valid_len_to_fp16)[name = tensor("op_1815_cast_fp16")]; + tensor var_1815_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_1815_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor input_43_cast_fp16 = add(x = input_37_cast_fp16, y = linear_13_cast_fp16)[name = tensor("input_43_cast_fp16")]; + tensor input_45_axes_0 = const()[name = tensor("input_45_axes_0"), val = tensor([-1])]; + tensor norm3_2_weight_to_fp16 = const()[name = tensor("norm3_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(84985152)))]; + tensor norm3_2_bias_to_fp16 = const()[name = tensor("norm3_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(84987264)))]; + tensor var_1819_to_fp16 = const()[name = tensor("op_1819_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_45_cast_fp16 = layer_norm(axes = input_45_axes_0, beta = norm3_2_bias_to_fp16, epsilon = var_1819_to_fp16, gamma = norm3_2_weight_to_fp16, x = input_43_cast_fp16)[name = tensor("input_45_cast_fp16")]; + tensor linear3_1_weight_to_fp16 = const()[name = tensor("linear3_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(84989376)))]; + tensor linear_14_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = linear3_1_weight_to_fp16, x = input_45_cast_fp16)[name = tensor("linear_14_cast_fp16")]; + tensor input_47_mode_0 = const()[name = tensor("input_47_mode_0"), val = tensor("EXACT")]; + tensor input_47_cast_fp16 = gelu(mode = input_47_mode_0, x = linear_14_cast_fp16)[name = tensor("input_47_cast_fp16")]; + tensor linear3_2_weight_to_fp16 = const()[name = tensor("linear3_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(93378048)))]; + tensor linear_15_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = linear3_2_weight_to_fp16, x = input_47_cast_fp16)[name = tensor("linear_15_cast_fp16")]; + tensor input_49_cast_fp16 = add(x = input_43_cast_fp16, y = linear_15_cast_fp16)[name = tensor("input_49_cast_fp16")]; + tensor input_51_axes_0 = const()[name = tensor("input_51_axes_0"), val = tensor([-1])]; + tensor norm4_1_weight_to_fp16 = const()[name = tensor("norm4_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(101766720)))]; + tensor norm4_1_bias_to_fp16 = const()[name = tensor("norm4_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(101768832)))]; + tensor var_1836_to_fp16 = const()[name = tensor("op_1836_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_51_cast_fp16 = layer_norm(axes = input_51_axes_0, beta = norm4_1_bias_to_fp16, epsilon = var_1836_to_fp16, gamma = norm4_1_weight_to_fp16, x = input_49_cast_fp16)[name = tensor("input_51_cast_fp16")]; + tensor attn4_in_proj_weight_to_fp16 = const()[name = tensor("attn4_in_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(101770944)))]; + tensor linear_16_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = attn4_in_proj_weight_to_fp16, x = input_51_cast_fp16)[name = tensor("linear_16_cast_fp16")]; + tensor var_1850 = const()[name = tensor("op_1850"), val = tensor([1, 256, 3, 16, 64])]; + tensor qkv_9_cast_fp16 = reshape(shape = var_1850, x = linear_16_cast_fp16)[name = tensor("qkv_9_cast_fp16")]; + tensor q_25_begin_0 = const()[name = tensor("q_25_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor q_25_end_0 = const()[name = tensor("q_25_end_0"), val = tensor([1, 256, 1, 16, 64])]; + tensor q_25_end_mask_0 = const()[name = tensor("q_25_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor q_25_squeeze_mask_0 = const()[name = tensor("q_25_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor q_25_cast_fp16 = slice_by_index(begin = q_25_begin_0, end = q_25_end_0, end_mask = q_25_end_mask_0, squeeze_mask = q_25_squeeze_mask_0, x = qkv_9_cast_fp16)[name = tensor("q_25_cast_fp16")]; + tensor k_17_begin_0 = const()[name = tensor("k_17_begin_0"), val = tensor([0, 0, 1, 0, 0])]; + tensor k_17_end_0 = const()[name = tensor("k_17_end_0"), val = tensor([1, 256, 2, 16, 64])]; + tensor k_17_end_mask_0 = const()[name = tensor("k_17_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor k_17_squeeze_mask_0 = const()[name = tensor("k_17_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor k_17_cast_fp16 = slice_by_index(begin = k_17_begin_0, end = k_17_end_0, end_mask = k_17_end_mask_0, squeeze_mask = k_17_squeeze_mask_0, x = qkv_9_cast_fp16)[name = tensor("k_17_cast_fp16")]; + tensor v_9_begin_0 = const()[name = tensor("v_9_begin_0"), val = tensor([0, 0, 2, 0, 0])]; + tensor v_9_end_0 = const()[name = tensor("v_9_end_0"), val = tensor([1, 256, 3, 16, 64])]; + tensor v_9_end_mask_0 = const()[name = tensor("v_9_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor v_9_squeeze_mask_0 = const()[name = tensor("v_9_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor v_9_cast_fp16 = slice_by_index(begin = v_9_begin_0, end = v_9_end_0, end_mask = v_9_end_mask_0, squeeze_mask = v_9_squeeze_mask_0, x = qkv_9_cast_fp16)[name = tensor("v_9_cast_fp16")]; + tensor var_1908 = const()[name = tensor("op_1908"), val = tensor([1, 1])]; + tensor position4_to_fp16_dtype_0 = const()[name = tensor("position4_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor position4_to_fp16 = cast(dtype = position4_to_fp16_dtype_0, x = position4)[name = tensor("cast_641")]; + tensor var_1909_cast_fp16 = reshape(shape = var_1908, x = position4_to_fp16)[name = tensor("op_1909_cast_fp16")]; + tensor ts_27_cast_fp16 = add(x = ts_1_promoted_to_fp16, y = var_1909_cast_fp16)[name = tensor("ts_27_cast_fp16")]; + tensor var_1916 = const()[name = tensor("op_1916"), val = tensor([1, 256, 1, 1])]; + tensor ts_29_cast_fp16 = reshape(shape = var_1916, x = ts_27_cast_fp16)[name = tensor("ts_29_cast_fp16")]; + tensor var_1923 = const()[name = tensor("op_1923"), val = tensor([1, 256, 16, 32, 2])]; + tensor q_complex_9_cast_fp16 = reshape(shape = var_1923, x = q_25_cast_fp16)[name = tensor("q_complex_9_cast_fp16")]; + tensor var_1930 = const()[name = tensor("op_1930"), val = tensor([1, 256, 16, 32, 2])]; + tensor k_complex_9_cast_fp16 = reshape(shape = var_1930, x = k_17_cast_fp16)[name = tensor("k_complex_9_cast_fp16")]; + tensor var_1934_begin_0 = const()[name = tensor("op_1934_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_1934_end_0 = const()[name = tensor("op_1934_end_0"), val = tensor([1, 256, 16, 32, 1])]; + tensor var_1934_end_mask_0 = const()[name = tensor("op_1934_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_1934_squeeze_mask_0 = const()[name = tensor("op_1934_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_1934_cast_fp16 = slice_by_index(begin = var_1934_begin_0, end = var_1934_end_0, end_mask = var_1934_end_mask_0, squeeze_mask = var_1934_squeeze_mask_0, x = q_complex_9_cast_fp16)[name = tensor("op_1934_cast_fp16")]; + tensor var_1942_begin_0 = const()[name = tensor("op_1942_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_1942_end_0 = const()[name = tensor("op_1942_end_0"), val = tensor([1, 256, 16, 32, 2])]; + tensor var_1942_end_mask_0 = const()[name = tensor("op_1942_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_1942_squeeze_mask_0 = const()[name = tensor("op_1942_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_1942_cast_fp16 = slice_by_index(begin = var_1942_begin_0, end = var_1942_end_0, end_mask = var_1942_end_mask_0, squeeze_mask = var_1942_squeeze_mask_0, x = q_complex_9_cast_fp16)[name = tensor("op_1942_cast_fp16")]; + tensor var_1950_begin_0 = const()[name = tensor("op_1950_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_1950_end_0 = const()[name = tensor("op_1950_end_0"), val = tensor([1, 256, 16, 32, 1])]; + tensor var_1950_end_mask_0 = const()[name = tensor("op_1950_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_1950_squeeze_mask_0 = const()[name = tensor("op_1950_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_1950_cast_fp16 = slice_by_index(begin = var_1950_begin_0, end = var_1950_end_0, end_mask = var_1950_end_mask_0, squeeze_mask = var_1950_squeeze_mask_0, x = k_complex_9_cast_fp16)[name = tensor("op_1950_cast_fp16")]; + tensor var_1958_begin_0 = const()[name = tensor("op_1958_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_1958_end_0 = const()[name = tensor("op_1958_end_0"), val = tensor([1, 256, 16, 32, 2])]; + tensor var_1958_end_mask_0 = const()[name = tensor("op_1958_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_1958_squeeze_mask_0 = const()[name = tensor("op_1958_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_1958_cast_fp16 = slice_by_index(begin = var_1958_begin_0, end = var_1958_end_0, end_mask = var_1958_end_mask_0, squeeze_mask = var_1958_squeeze_mask_0, x = k_complex_9_cast_fp16)[name = tensor("op_1958_cast_fp16")]; + tensor freqs_9_to_fp16 = const()[name = tensor("freqs_9_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(108062464)))]; + tensor var_1964_cast_fp16 = mul(x = freqs_9_to_fp16, y = ts_29_cast_fp16)[name = tensor("op_1964_cast_fp16")]; + tensor rotr_9_cast_fp16 = cos(x = var_1964_cast_fp16)[name = tensor("rotr_9_cast_fp16")]; + tensor roti_9_cast_fp16 = sin(x = var_1964_cast_fp16)[name = tensor("roti_9_cast_fp16")]; + tensor var_1968_cast_fp16 = mul(x = var_1934_cast_fp16, y = rotr_9_cast_fp16)[name = tensor("op_1968_cast_fp16")]; + tensor var_1969_cast_fp16 = mul(x = var_1942_cast_fp16, y = roti_9_cast_fp16)[name = tensor("op_1969_cast_fp16")]; + tensor qor_17_cast_fp16 = sub(x = var_1968_cast_fp16, y = var_1969_cast_fp16)[name = tensor("qor_17_cast_fp16")]; + tensor var_1972_cast_fp16 = mul(x = var_1934_cast_fp16, y = roti_9_cast_fp16)[name = tensor("op_1972_cast_fp16")]; + tensor var_1973_cast_fp16 = mul(x = var_1942_cast_fp16, y = rotr_9_cast_fp16)[name = tensor("op_1973_cast_fp16")]; + tensor qoi_17_cast_fp16 = add(x = var_1972_cast_fp16, y = var_1973_cast_fp16)[name = tensor("qoi_17_cast_fp16")]; + tensor var_1976_cast_fp16 = mul(x = var_1950_cast_fp16, y = rotr_9_cast_fp16)[name = tensor("op_1976_cast_fp16")]; + tensor var_1977_cast_fp16 = mul(x = var_1958_cast_fp16, y = roti_9_cast_fp16)[name = tensor("op_1977_cast_fp16")]; + tensor kor_17_cast_fp16 = sub(x = var_1976_cast_fp16, y = var_1977_cast_fp16)[name = tensor("kor_17_cast_fp16")]; + tensor var_1980_cast_fp16 = mul(x = var_1950_cast_fp16, y = roti_9_cast_fp16)[name = tensor("op_1980_cast_fp16")]; + tensor var_1981_cast_fp16 = mul(x = var_1958_cast_fp16, y = rotr_9_cast_fp16)[name = tensor("op_1981_cast_fp16")]; + tensor koi_17_cast_fp16 = add(x = var_1980_cast_fp16, y = var_1981_cast_fp16)[name = tensor("koi_17_cast_fp16")]; + tensor qo_9_axis_0 = const()[name = tensor("qo_9_axis_0"), val = tensor(-1)]; + tensor qo_9_cast_fp16 = stack(axis = qo_9_axis_0, values = (qor_17_cast_fp16, qoi_17_cast_fp16))[name = tensor("qo_9_cast_fp16")]; + tensor ko_9_axis_0 = const()[name = tensor("ko_9_axis_0"), val = tensor(-1)]; + tensor ko_9_cast_fp16 = stack(axis = ko_9_axis_0, values = (kor_17_cast_fp16, koi_17_cast_fp16))[name = tensor("ko_9_cast_fp16")]; + tensor var_2014 = const()[name = tensor("op_2014"), val = tensor([1, 256, 16, 64])]; + tensor q_27_cast_fp16 = reshape(shape = var_2014, x = qo_9_cast_fp16)[name = tensor("q_27_cast_fp16")]; + tensor var_2020 = const()[name = tensor("op_2020"), val = tensor([1, 256, 16, 64])]; + tensor k_19_cast_fp16 = reshape(shape = var_2020, x = ko_9_cast_fp16)[name = tensor("k_19_cast_fp16")]; + tensor write_indices_float_25_cast_fp16 = add(x = var_1909_cast_fp16, y = var_393_promoted_to_fp16)[name = tensor("write_indices_float_25_cast_fp16")]; + tensor _inversed_2041_y_0_to_fp16 = const()[name = tensor("_inversed_2041_y_0_to_fp16"), val = tensor(0x1p-9)]; + tensor _inversed_2041_cast_fp16 = mul(x = write_indices_float_25_cast_fp16, y = _inversed_2041_y_0_to_fp16)[name = tensor("_inversed_2041_cast_fp16")]; + tensor var_2042_cast_fp16 = floor(x = _inversed_2041_cast_fp16)[name = tensor("op_2042_cast_fp16")]; + tensor var_2043_to_fp16 = const()[name = tensor("op_2043_to_fp16"), val = tensor(0x1p+9)]; + tensor var_2044_cast_fp16 = mul(x = var_2042_cast_fp16, y = var_2043_to_fp16)[name = tensor("op_2044_cast_fp16")]; + tensor write_indices_float_27_cast_fp16 = sub(x = write_indices_float_25_cast_fp16, y = var_2044_cast_fp16)[name = tensor("write_indices_float_27_cast_fp16")]; + tensor var_2062_value_0_to_fp16 = const()[name = tensor("op_2062_value_0_to_fp16"), val = tensor(0x1.ffp+8)]; + tensor var_2062_cast_fp16 = fill_like(ref_tensor = write_indices_float_27_cast_fp16, value = var_2062_value_0_to_fp16)[name = tensor("op_2062_cast_fp16")]; + tensor write_indices_float_29_cast_fp16 = select(a = write_indices_float_27_cast_fp16, b = var_2062_cast_fp16, cond = token_valid_1_cast_fp16)[name = tensor("write_indices_float_29_cast_fp16")]; + tensor var_2068_dtype_0 = const()[name = tensor("op_2068_dtype_0"), val = tensor("int32")]; + tensor var_2073 = const()[name = tensor("op_2073"), val = tensor([1, 256, 1, 1])]; + tensor write_indices_float_29_cast_fp16_to_int32 = cast(dtype = var_2068_dtype_0, x = write_indices_float_29_cast_fp16)[name = tensor("cast_640")]; + tensor var_2074 = reshape(shape = var_2073, x = write_indices_float_29_cast_fp16_to_int32)[name = tensor("op_2074")]; + tensor write_indices_9_reps_0 = const()[name = tensor("write_indices_9_reps_0"), val = tensor([1, 1, 16, 64])]; + tensor write_indices_9 = tile(reps = write_indices_9_reps_0, x = var_2074)[name = tensor("write_indices_9")]; + tensor var_2084_begin_0 = const()[name = tensor("op_2084_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_2084_end_0 = const()[name = tensor("op_2084_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor var_2084_end_mask_0 = const()[name = tensor("op_2084_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_2084_squeeze_mask_0 = const()[name = tensor("op_2084_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor cache4_to_fp16_dtype_0 = const()[name = tensor("cache4_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor cache4_to_fp16 = cast(dtype = cache4_to_fp16_dtype_0, x = cache4)[name = tensor("cast_639")]; + tensor var_2084_cast_fp16 = slice_by_index(begin = var_2084_begin_0, end = var_2084_end_0, end_mask = var_2084_end_mask_0, squeeze_mask = var_2084_squeeze_mask_0, x = cache4_to_fp16)[name = tensor("op_2084_cast_fp16")]; + tensor var_2086_axis_0 = const()[name = tensor("op_2086_axis_0"), val = tensor(1)]; + tensor var_2086_mode_0 = const()[name = tensor("op_2086_mode_0"), val = tensor("update")]; + tensor var_2086_validate_indices_0 = const()[name = tensor("op_2086_validate_indices_0"), val = tensor(false)]; + tensor var_2086_cast_fp16 = scatter_along_axis(axis = var_2086_axis_0, data = var_2084_cast_fp16, indices = write_indices_9, mode = var_2086_mode_0, updates = k_19_cast_fp16, validate_indices = var_2086_validate_indices_0)[name = tensor("op_2086_cast_fp16")]; + tensor concat_34 = const()[name = tensor("concat_34"), val = tensor([0, 0, 0, 0, 0])]; + tensor concat_35 = const()[name = tensor("concat_35"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_9_internal_tensor_assign_1_stride_0 = const()[name = tensor("new_cache_9_internal_tensor_assign_1_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_9_internal_tensor_assign_1_begin_mask_0 = const()[name = tensor("new_cache_9_internal_tensor_assign_1_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_9_internal_tensor_assign_1_end_mask_0 = const()[name = tensor("new_cache_9_internal_tensor_assign_1_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_9_internal_tensor_assign_1_squeeze_mask_0 = const()[name = tensor("new_cache_9_internal_tensor_assign_1_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_54 = const()[name = tensor("shape_54"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_8 = const()[name = tensor("reduce_prod_8"), val = tensor(1048576)]; + tensor range_1d_8_start_0 = const()[name = tensor("range_1d_8_start_0"), val = tensor(0)]; + tensor range_1d_8_step_0 = const()[name = tensor("range_1d_8_step_0"), val = tensor(1)]; + tensor range_1d_8 = range_1d(end = reduce_prod_8, start = range_1d_8_start_0, step = range_1d_8_step_0)[name = tensor("range_1d_8")]; + tensor reshape_40 = reshape(shape = shape_54, x = range_1d_8)[name = tensor("reshape_40")]; + tensor slice_by_index_8 = slice_by_index(begin = concat_34, begin_mask = new_cache_9_internal_tensor_assign_1_begin_mask_0, end = concat_35, end_mask = new_cache_9_internal_tensor_assign_1_end_mask_0, squeeze_mask = new_cache_9_internal_tensor_assign_1_squeeze_mask_0, stride = new_cache_9_internal_tensor_assign_1_stride_0, x = reshape_40)[name = tensor("slice_by_index_8")]; + tensor reshape_41_shape_0 = const()[name = tensor("reshape_41_shape_0"), val = tensor([-1])]; + tensor reshape_41 = reshape(shape = reshape_41_shape_0, x = slice_by_index_8)[name = tensor("reshape_41")]; + tensor reshape_42_shape_0 = const()[name = tensor("reshape_42_shape_0"), val = tensor([-1])]; + tensor reshape_42_cast_fp16 = reshape(shape = reshape_42_shape_0, x = var_2086_cast_fp16)[name = tensor("reshape_42_cast_fp16")]; + tensor reshape_43_shape_0 = const()[name = tensor("reshape_43_shape_0"), val = tensor([-1])]; + tensor reshape_43_cast_fp16 = reshape(shape = reshape_43_shape_0, x = cache4_to_fp16)[name = tensor("reshape_43_cast_fp16")]; + tensor scatter_8_mode_0 = const()[name = tensor("scatter_8_mode_0"), val = tensor("update")]; + tensor scatter_8_axis_0 = const()[name = tensor("scatter_8_axis_0"), val = tensor(0)]; + tensor scatter_8_validate_indices_0 = const()[name = tensor("scatter_8_validate_indices_0"), val = tensor(false)]; + tensor scatter_8_cast_fp16 = scatter(axis = scatter_8_axis_0, data = reshape_43_cast_fp16, indices = reshape_41, mode = scatter_8_mode_0, updates = reshape_42_cast_fp16, validate_indices = scatter_8_validate_indices_0)[name = tensor("scatter_8_cast_fp16")]; + tensor reshape_44_cast_fp16 = reshape(shape = shape_54, x = scatter_8_cast_fp16)[name = tensor("reshape_44_cast_fp16")]; + tensor var_2094_begin_0 = const()[name = tensor("op_2094_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor var_2094_end_0 = const()[name = tensor("op_2094_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor var_2094_end_mask_0 = const()[name = tensor("op_2094_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_2094_squeeze_mask_0 = const()[name = tensor("op_2094_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor var_2094_cast_fp16 = slice_by_index(begin = var_2094_begin_0, end = var_2094_end_0, end_mask = var_2094_end_mask_0, squeeze_mask = var_2094_squeeze_mask_0, x = reshape_44_cast_fp16)[name = tensor("op_2094_cast_fp16")]; + tensor var_2096_axis_0 = const()[name = tensor("op_2096_axis_0"), val = tensor(1)]; + tensor var_2096_mode_0 = const()[name = tensor("op_2096_mode_0"), val = tensor("update")]; + tensor var_2096_validate_indices_0 = const()[name = tensor("op_2096_validate_indices_0"), val = tensor(false)]; + tensor var_2096_cast_fp16 = scatter_along_axis(axis = var_2096_axis_0, data = var_2094_cast_fp16, indices = write_indices_9, mode = var_2096_mode_0, updates = v_9_cast_fp16, validate_indices = var_2096_validate_indices_0)[name = tensor("op_2096_cast_fp16")]; + tensor concat_36 = const()[name = tensor("concat_36"), val = tensor([1, 0, 0, 0, 0])]; + tensor concat_37 = const()[name = tensor("concat_37"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_9_internal_tensor_assign_2_stride_0 = const()[name = tensor("new_cache_9_internal_tensor_assign_2_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_9_internal_tensor_assign_2_begin_mask_0 = const()[name = tensor("new_cache_9_internal_tensor_assign_2_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_9_internal_tensor_assign_2_end_mask_0 = const()[name = tensor("new_cache_9_internal_tensor_assign_2_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_9_internal_tensor_assign_2_squeeze_mask_0 = const()[name = tensor("new_cache_9_internal_tensor_assign_2_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_55 = const()[name = tensor("shape_55"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_9 = const()[name = tensor("reduce_prod_9"), val = tensor(1048576)]; + tensor range_1d_9_start_0 = const()[name = tensor("range_1d_9_start_0"), val = tensor(0)]; + tensor range_1d_9_step_0 = const()[name = tensor("range_1d_9_step_0"), val = tensor(1)]; + tensor range_1d_9 = range_1d(end = reduce_prod_9, start = range_1d_9_start_0, step = range_1d_9_step_0)[name = tensor("range_1d_9")]; + tensor reshape_45 = reshape(shape = shape_55, x = range_1d_9)[name = tensor("reshape_45")]; + tensor slice_by_index_9 = slice_by_index(begin = concat_36, begin_mask = new_cache_9_internal_tensor_assign_2_begin_mask_0, end = concat_37, end_mask = new_cache_9_internal_tensor_assign_2_end_mask_0, squeeze_mask = new_cache_9_internal_tensor_assign_2_squeeze_mask_0, stride = new_cache_9_internal_tensor_assign_2_stride_0, x = reshape_45)[name = tensor("slice_by_index_9")]; + tensor reshape_46_shape_0 = const()[name = tensor("reshape_46_shape_0"), val = tensor([-1])]; + tensor reshape_46 = reshape(shape = reshape_46_shape_0, x = slice_by_index_9)[name = tensor("reshape_46")]; + tensor reshape_47_shape_0 = const()[name = tensor("reshape_47_shape_0"), val = tensor([-1])]; + tensor reshape_47_cast_fp16 = reshape(shape = reshape_47_shape_0, x = var_2096_cast_fp16)[name = tensor("reshape_47_cast_fp16")]; + tensor reshape_48_shape_0 = const()[name = tensor("reshape_48_shape_0"), val = tensor([-1])]; + tensor reshape_48_cast_fp16 = reshape(shape = reshape_48_shape_0, x = reshape_44_cast_fp16)[name = tensor("reshape_48_cast_fp16")]; + tensor scatter_9_mode_0 = const()[name = tensor("scatter_9_mode_0"), val = tensor("update")]; + tensor scatter_9_axis_0 = const()[name = tensor("scatter_9_axis_0"), val = tensor(0)]; + tensor scatter_9_validate_indices_0 = const()[name = tensor("scatter_9_validate_indices_0"), val = tensor(false)]; + tensor scatter_9_cast_fp16 = scatter(axis = scatter_9_axis_0, data = reshape_48_cast_fp16, indices = reshape_46, mode = scatter_9_mode_0, updates = reshape_47_cast_fp16, validate_indices = scatter_9_validate_indices_0)[name = tensor("scatter_9_cast_fp16")]; + tensor reshape_49_cast_fp16 = reshape(shape = shape_55, x = scatter_9_cast_fp16)[name = tensor("reshape_49_cast_fp16")]; + tensor reshape_49_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("reshape_49_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor keys_25_begin_0 = const()[name = tensor("keys_25_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor keys_25_end_0 = const()[name = tensor("keys_25_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor keys_25_end_mask_0 = const()[name = tensor("keys_25_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor keys_25_squeeze_mask_0 = const()[name = tensor("keys_25_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor keys_25_cast_fp16 = slice_by_index(begin = keys_25_begin_0, end = keys_25_end_0, end_mask = keys_25_end_mask_0, squeeze_mask = keys_25_squeeze_mask_0, x = reshape_49_cast_fp16)[name = tensor("keys_25_cast_fp16")]; + tensor values_25_begin_0 = const()[name = tensor("values_25_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor values_25_end_0 = const()[name = tensor("values_25_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor values_25_end_mask_0 = const()[name = tensor("values_25_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor values_25_squeeze_mask_0 = const()[name = tensor("values_25_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor values_25_cast_fp16 = slice_by_index(begin = values_25_begin_0, end = values_25_end_0, end_mask = values_25_end_mask_0, squeeze_mask = values_25_squeeze_mask_0, x = reshape_49_cast_fp16)[name = tensor("values_25_cast_fp16")]; + tensor var_2108_cast_fp16 = not_equal(x = keys_25_cast_fp16, y = keys_25_cast_fp16)[name = tensor("op_2108_cast_fp16")]; + tensor keys_27_cast_fp16 = select(a = var_470_to_fp16, b = keys_25_cast_fp16, cond = var_2108_cast_fp16)[name = tensor("keys_27_cast_fp16")]; + tensor var_2116_cast_fp16 = not_equal(x = values_25_cast_fp16, y = values_25_cast_fp16)[name = tensor("op_2116_cast_fp16")]; + tensor values_27_cast_fp16 = select(a = var_470_to_fp16, b = values_25_cast_fp16, cond = var_2116_cast_fp16)[name = tensor("values_27_cast_fp16")]; + tensor var_2140 = const()[name = tensor("op_2140"), val = tensor([0, 2, 1, 3])]; + tensor var_2156 = const()[name = tensor("op_2156"), val = tensor([1, 1, 1])]; + tensor var_2157_cast_fp16 = reshape(shape = var_2156, x = position4_to_fp16)[name = tensor("op_2157_cast_fp16")]; + tensor q_positions_9_cast_fp16 = add(x = var_2157_cast_fp16, y = q_offsets_1_promoted_to_fp16)[name = tensor("q_positions_9_cast_fp16")]; + tensor valid_end_9_cast_fp16 = add(x = var_2157_cast_fp16, y = var_536_cast_fp16)[name = tensor("valid_end_9_cast_fp16")]; + tensor valid_mask_9_cast_fp16 = less(x = k_positions_1_promoted_to_fp16, y = valid_end_9_cast_fp16)[name = tensor("valid_mask_9_cast_fp16")]; + tensor causal_mask_9_cast_fp16 = less_equal(x = k_positions_1_promoted_to_fp16, y = q_positions_9_cast_fp16)[name = tensor("causal_mask_9_cast_fp16")]; + tensor var_2185 = logical_and(x = valid_mask_9_cast_fp16, y = causal_mask_9_cast_fp16)[name = tensor("op_2185")]; + tensor attn_mask_9_axes_0 = const()[name = tensor("attn_mask_9_axes_0"), val = tensor([1])]; + tensor attn_mask_9 = expand_dims(axes = attn_mask_9_axes_0, x = var_2185)[name = tensor("attn_mask_9")]; + tensor var_2199_transpose_x_0 = const()[name = tensor("op_2199_transpose_x_0"), val = tensor(false)]; + tensor var_2199_transpose_y_0 = const()[name = tensor("op_2199_transpose_y_0"), val = tensor(false)]; + tensor transpose_77_perm_0 = const()[name = tensor("transpose_77_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_78_perm_0 = const()[name = tensor("transpose_78_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_78 = transpose(perm = transpose_78_perm_0, x = keys_27_cast_fp16)[name = tensor("transpose_188")]; + tensor transpose_77 = transpose(perm = transpose_77_perm_0, x = q_27_cast_fp16)[name = tensor("transpose_189")]; + tensor var_2199_cast_fp16 = matmul(transpose_x = var_2199_transpose_x_0, transpose_y = var_2199_transpose_y_0, x = transpose_77, y = transpose_78)[name = tensor("op_2199_cast_fp16")]; + tensor var_2193_to_fp16 = const()[name = tensor("op_2193_to_fp16"), val = tensor([0x1p-3])]; + tensor attn_weights_25_cast_fp16 = mul(x = var_2199_cast_fp16, y = var_2193_to_fp16)[name = tensor("attn_weights_25_cast_fp16")]; + tensor var_2201 = logical_not(x = attn_mask_9)[name = tensor("op_2201")]; + tensor var_2202_to_fp16 = const()[name = tensor("op_2202_to_fp16"), val = tensor(-inf)]; + tensor attn_weights_27_cast_fp16 = select(a = var_2202_to_fp16, b = attn_weights_25_cast_fp16, cond = var_2201)[name = tensor("attn_weights_27_cast_fp16")]; + tensor var_2204 = const()[name = tensor("op_2204"), val = tensor(-1)]; + tensor attn_weights_29_cast_fp16 = softmax(axis = var_2204, x = attn_weights_27_cast_fp16)[name = tensor("attn_weights_29_cast_fp16")]; + tensor attn_output_9_transpose_x_0 = const()[name = tensor("attn_output_9_transpose_x_0"), val = tensor(false)]; + tensor attn_output_9_transpose_y_0 = const()[name = tensor("attn_output_9_transpose_y_0"), val = tensor(false)]; + tensor values_29_cast_fp16 = transpose(perm = var_2140, x = values_27_cast_fp16)[name = tensor("transpose_190")]; + tensor attn_output_9_cast_fp16 = matmul(transpose_x = attn_output_9_transpose_x_0, transpose_y = attn_output_9_transpose_y_0, x = attn_weights_29_cast_fp16, y = values_29_cast_fp16)[name = tensor("attn_output_9_cast_fp16")]; + tensor var_2212 = const()[name = tensor("op_2212"), val = tensor([0, 2, 1, 3])]; + tensor var_2217 = const()[name = tensor("op_2217"), val = tensor([1, 256, 1024])]; + tensor var_2213_cast_fp16 = transpose(perm = var_2212, x = attn_output_9_cast_fp16)[name = tensor("transpose_187")]; + tensor input_53_cast_fp16 = reshape(shape = var_2217, x = var_2213_cast_fp16)[name = tensor("input_53_cast_fp16")]; + tensor attn4_out_proj_weight_to_fp16 = const()[name = tensor("attn4_out_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(108062592)))]; + tensor linear_17_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = attn4_out_proj_weight_to_fp16, x = input_53_cast_fp16)[name = tensor("linear_17_cast_fp16")]; + tensor var_2226_cast_fp16 = add(x = position4_to_fp16, y = valid_len_to_fp16)[name = tensor("op_2226_cast_fp16")]; + tensor var_2226_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_2226_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor input_55_cast_fp16 = add(x = input_49_cast_fp16, y = linear_17_cast_fp16)[name = tensor("input_55_cast_fp16")]; + tensor input_57_axes_0 = const()[name = tensor("input_57_axes_0"), val = tensor([-1])]; + tensor norm4_2_weight_to_fp16 = const()[name = tensor("norm4_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(110159808)))]; + tensor norm4_2_bias_to_fp16 = const()[name = tensor("norm4_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(110161920)))]; + tensor var_2230_to_fp16 = const()[name = tensor("op_2230_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_57_cast_fp16 = layer_norm(axes = input_57_axes_0, beta = norm4_2_bias_to_fp16, epsilon = var_2230_to_fp16, gamma = norm4_2_weight_to_fp16, x = input_55_cast_fp16)[name = tensor("input_57_cast_fp16")]; + tensor linear4_1_weight_to_fp16 = const()[name = tensor("linear4_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(110164032)))]; + tensor linear_18_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = linear4_1_weight_to_fp16, x = input_57_cast_fp16)[name = tensor("linear_18_cast_fp16")]; + tensor input_59_mode_0 = const()[name = tensor("input_59_mode_0"), val = tensor("EXACT")]; + tensor input_59_cast_fp16 = gelu(mode = input_59_mode_0, x = linear_18_cast_fp16)[name = tensor("input_59_cast_fp16")]; + tensor linear4_2_weight_to_fp16 = const()[name = tensor("linear4_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(118552704)))]; + tensor linear_19_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = linear4_2_weight_to_fp16, x = input_59_cast_fp16)[name = tensor("linear_19_cast_fp16")]; + tensor input_61_cast_fp16 = add(x = input_55_cast_fp16, y = linear_19_cast_fp16)[name = tensor("input_61_cast_fp16")]; + tensor input_63_axes_0 = const()[name = tensor("input_63_axes_0"), val = tensor([-1])]; + tensor norm5_1_weight_to_fp16 = const()[name = tensor("norm5_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(126941376)))]; + tensor norm5_1_bias_to_fp16 = const()[name = tensor("norm5_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(126943488)))]; + tensor var_2247_to_fp16 = const()[name = tensor("op_2247_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_63_cast_fp16 = layer_norm(axes = input_63_axes_0, beta = norm5_1_bias_to_fp16, epsilon = var_2247_to_fp16, gamma = norm5_1_weight_to_fp16, x = input_61_cast_fp16)[name = tensor("input_63_cast_fp16")]; + tensor attn5_in_proj_weight_to_fp16 = const()[name = tensor("attn5_in_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(126945600)))]; + tensor linear_20_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = attn5_in_proj_weight_to_fp16, x = input_63_cast_fp16)[name = tensor("linear_20_cast_fp16")]; + tensor var_2261 = const()[name = tensor("op_2261"), val = tensor([1, 256, 3, 16, 64])]; + tensor qkv_11_cast_fp16 = reshape(shape = var_2261, x = linear_20_cast_fp16)[name = tensor("qkv_11_cast_fp16")]; + tensor q_31_begin_0 = const()[name = tensor("q_31_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor q_31_end_0 = const()[name = tensor("q_31_end_0"), val = tensor([1, 256, 1, 16, 64])]; + tensor q_31_end_mask_0 = const()[name = tensor("q_31_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor q_31_squeeze_mask_0 = const()[name = tensor("q_31_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor q_31_cast_fp16 = slice_by_index(begin = q_31_begin_0, end = q_31_end_0, end_mask = q_31_end_mask_0, squeeze_mask = q_31_squeeze_mask_0, x = qkv_11_cast_fp16)[name = tensor("q_31_cast_fp16")]; + tensor k_21_begin_0 = const()[name = tensor("k_21_begin_0"), val = tensor([0, 0, 1, 0, 0])]; + tensor k_21_end_0 = const()[name = tensor("k_21_end_0"), val = tensor([1, 256, 2, 16, 64])]; + tensor k_21_end_mask_0 = const()[name = tensor("k_21_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor k_21_squeeze_mask_0 = const()[name = tensor("k_21_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor k_21_cast_fp16 = slice_by_index(begin = k_21_begin_0, end = k_21_end_0, end_mask = k_21_end_mask_0, squeeze_mask = k_21_squeeze_mask_0, x = qkv_11_cast_fp16)[name = tensor("k_21_cast_fp16")]; + tensor v_11_begin_0 = const()[name = tensor("v_11_begin_0"), val = tensor([0, 0, 2, 0, 0])]; + tensor v_11_end_0 = const()[name = tensor("v_11_end_0"), val = tensor([1, 256, 3, 16, 64])]; + tensor v_11_end_mask_0 = const()[name = tensor("v_11_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor v_11_squeeze_mask_0 = const()[name = tensor("v_11_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor v_11_cast_fp16 = slice_by_index(begin = v_11_begin_0, end = v_11_end_0, end_mask = v_11_end_mask_0, squeeze_mask = v_11_squeeze_mask_0, x = qkv_11_cast_fp16)[name = tensor("v_11_cast_fp16")]; + tensor var_2319 = const()[name = tensor("op_2319"), val = tensor([1, 1])]; + tensor position5_to_fp16_dtype_0 = const()[name = tensor("position5_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor position5_to_fp16 = cast(dtype = position5_to_fp16_dtype_0, x = position5)[name = tensor("cast_636")]; + tensor var_2320_cast_fp16 = reshape(shape = var_2319, x = position5_to_fp16)[name = tensor("op_2320_cast_fp16")]; + tensor ts_33_cast_fp16 = add(x = ts_1_promoted_to_fp16, y = var_2320_cast_fp16)[name = tensor("ts_33_cast_fp16")]; + tensor var_2327 = const()[name = tensor("op_2327"), val = tensor([1, 256, 1, 1])]; + tensor ts_35_cast_fp16 = reshape(shape = var_2327, x = ts_33_cast_fp16)[name = tensor("ts_35_cast_fp16")]; + tensor var_2334 = const()[name = tensor("op_2334"), val = tensor([1, 256, 16, 32, 2])]; + tensor q_complex_11_cast_fp16 = reshape(shape = var_2334, x = q_31_cast_fp16)[name = tensor("q_complex_11_cast_fp16")]; + tensor var_2341 = const()[name = tensor("op_2341"), val = tensor([1, 256, 16, 32, 2])]; + tensor k_complex_11_cast_fp16 = reshape(shape = var_2341, x = k_21_cast_fp16)[name = tensor("k_complex_11_cast_fp16")]; + tensor var_2345_begin_0 = const()[name = tensor("op_2345_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_2345_end_0 = const()[name = tensor("op_2345_end_0"), val = tensor([1, 256, 16, 32, 1])]; + tensor var_2345_end_mask_0 = const()[name = tensor("op_2345_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_2345_squeeze_mask_0 = const()[name = tensor("op_2345_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_2345_cast_fp16 = slice_by_index(begin = var_2345_begin_0, end = var_2345_end_0, end_mask = var_2345_end_mask_0, squeeze_mask = var_2345_squeeze_mask_0, x = q_complex_11_cast_fp16)[name = tensor("op_2345_cast_fp16")]; + tensor var_2353_begin_0 = const()[name = tensor("op_2353_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_2353_end_0 = const()[name = tensor("op_2353_end_0"), val = tensor([1, 256, 16, 32, 2])]; + tensor var_2353_end_mask_0 = const()[name = tensor("op_2353_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_2353_squeeze_mask_0 = const()[name = tensor("op_2353_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_2353_cast_fp16 = slice_by_index(begin = var_2353_begin_0, end = var_2353_end_0, end_mask = var_2353_end_mask_0, squeeze_mask = var_2353_squeeze_mask_0, x = q_complex_11_cast_fp16)[name = tensor("op_2353_cast_fp16")]; + tensor var_2361_begin_0 = const()[name = tensor("op_2361_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_2361_end_0 = const()[name = tensor("op_2361_end_0"), val = tensor([1, 256, 16, 32, 1])]; + tensor var_2361_end_mask_0 = const()[name = tensor("op_2361_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_2361_squeeze_mask_0 = const()[name = tensor("op_2361_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_2361_cast_fp16 = slice_by_index(begin = var_2361_begin_0, end = var_2361_end_0, end_mask = var_2361_end_mask_0, squeeze_mask = var_2361_squeeze_mask_0, x = k_complex_11_cast_fp16)[name = tensor("op_2361_cast_fp16")]; + tensor var_2369_begin_0 = const()[name = tensor("op_2369_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_2369_end_0 = const()[name = tensor("op_2369_end_0"), val = tensor([1, 256, 16, 32, 2])]; + tensor var_2369_end_mask_0 = const()[name = tensor("op_2369_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_2369_squeeze_mask_0 = const()[name = tensor("op_2369_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_2369_cast_fp16 = slice_by_index(begin = var_2369_begin_0, end = var_2369_end_0, end_mask = var_2369_end_mask_0, squeeze_mask = var_2369_squeeze_mask_0, x = k_complex_11_cast_fp16)[name = tensor("op_2369_cast_fp16")]; + tensor freqs_11_to_fp16 = const()[name = tensor("freqs_11_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(133237120)))]; + tensor var_2375_cast_fp16 = mul(x = freqs_11_to_fp16, y = ts_35_cast_fp16)[name = tensor("op_2375_cast_fp16")]; + tensor rotr_11_cast_fp16 = cos(x = var_2375_cast_fp16)[name = tensor("rotr_11_cast_fp16")]; + tensor roti_11_cast_fp16 = sin(x = var_2375_cast_fp16)[name = tensor("roti_11_cast_fp16")]; + tensor var_2379_cast_fp16 = mul(x = var_2345_cast_fp16, y = rotr_11_cast_fp16)[name = tensor("op_2379_cast_fp16")]; + tensor var_2380_cast_fp16 = mul(x = var_2353_cast_fp16, y = roti_11_cast_fp16)[name = tensor("op_2380_cast_fp16")]; + tensor qor_21_cast_fp16 = sub(x = var_2379_cast_fp16, y = var_2380_cast_fp16)[name = tensor("qor_21_cast_fp16")]; + tensor var_2383_cast_fp16 = mul(x = var_2345_cast_fp16, y = roti_11_cast_fp16)[name = tensor("op_2383_cast_fp16")]; + tensor var_2384_cast_fp16 = mul(x = var_2353_cast_fp16, y = rotr_11_cast_fp16)[name = tensor("op_2384_cast_fp16")]; + tensor qoi_21_cast_fp16 = add(x = var_2383_cast_fp16, y = var_2384_cast_fp16)[name = tensor("qoi_21_cast_fp16")]; + tensor var_2387_cast_fp16 = mul(x = var_2361_cast_fp16, y = rotr_11_cast_fp16)[name = tensor("op_2387_cast_fp16")]; + tensor var_2388_cast_fp16 = mul(x = var_2369_cast_fp16, y = roti_11_cast_fp16)[name = tensor("op_2388_cast_fp16")]; + tensor kor_21_cast_fp16 = sub(x = var_2387_cast_fp16, y = var_2388_cast_fp16)[name = tensor("kor_21_cast_fp16")]; + tensor var_2391_cast_fp16 = mul(x = var_2361_cast_fp16, y = roti_11_cast_fp16)[name = tensor("op_2391_cast_fp16")]; + tensor var_2392_cast_fp16 = mul(x = var_2369_cast_fp16, y = rotr_11_cast_fp16)[name = tensor("op_2392_cast_fp16")]; + tensor koi_21_cast_fp16 = add(x = var_2391_cast_fp16, y = var_2392_cast_fp16)[name = tensor("koi_21_cast_fp16")]; + tensor qo_11_axis_0 = const()[name = tensor("qo_11_axis_0"), val = tensor(-1)]; + tensor qo_11_cast_fp16 = stack(axis = qo_11_axis_0, values = (qor_21_cast_fp16, qoi_21_cast_fp16))[name = tensor("qo_11_cast_fp16")]; + tensor ko_11_axis_0 = const()[name = tensor("ko_11_axis_0"), val = tensor(-1)]; + tensor ko_11_cast_fp16 = stack(axis = ko_11_axis_0, values = (kor_21_cast_fp16, koi_21_cast_fp16))[name = tensor("ko_11_cast_fp16")]; + tensor var_2425 = const()[name = tensor("op_2425"), val = tensor([1, 256, 16, 64])]; + tensor q_33_cast_fp16 = reshape(shape = var_2425, x = qo_11_cast_fp16)[name = tensor("q_33_cast_fp16")]; + tensor var_2431 = const()[name = tensor("op_2431"), val = tensor([1, 256, 16, 64])]; + tensor k_23_cast_fp16 = reshape(shape = var_2431, x = ko_11_cast_fp16)[name = tensor("k_23_cast_fp16")]; + tensor write_indices_float_31_cast_fp16 = add(x = var_2320_cast_fp16, y = var_393_promoted_to_fp16)[name = tensor("write_indices_float_31_cast_fp16")]; + tensor _inversed_2452_y_0_to_fp16 = const()[name = tensor("_inversed_2452_y_0_to_fp16"), val = tensor(0x1p-9)]; + tensor _inversed_2452_cast_fp16 = mul(x = write_indices_float_31_cast_fp16, y = _inversed_2452_y_0_to_fp16)[name = tensor("_inversed_2452_cast_fp16")]; + tensor var_2453_cast_fp16 = floor(x = _inversed_2452_cast_fp16)[name = tensor("op_2453_cast_fp16")]; + tensor var_2454_to_fp16 = const()[name = tensor("op_2454_to_fp16"), val = tensor(0x1p+9)]; + tensor var_2455_cast_fp16 = mul(x = var_2453_cast_fp16, y = var_2454_to_fp16)[name = tensor("op_2455_cast_fp16")]; + tensor write_indices_float_33_cast_fp16 = sub(x = write_indices_float_31_cast_fp16, y = var_2455_cast_fp16)[name = tensor("write_indices_float_33_cast_fp16")]; + tensor var_2473_value_0_to_fp16 = const()[name = tensor("op_2473_value_0_to_fp16"), val = tensor(0x1.ffp+8)]; + tensor var_2473_cast_fp16 = fill_like(ref_tensor = write_indices_float_33_cast_fp16, value = var_2473_value_0_to_fp16)[name = tensor("op_2473_cast_fp16")]; + tensor write_indices_float_35_cast_fp16 = select(a = write_indices_float_33_cast_fp16, b = var_2473_cast_fp16, cond = token_valid_1_cast_fp16)[name = tensor("write_indices_float_35_cast_fp16")]; + tensor var_2479_dtype_0 = const()[name = tensor("op_2479_dtype_0"), val = tensor("int32")]; + tensor var_2484 = const()[name = tensor("op_2484"), val = tensor([1, 256, 1, 1])]; + tensor write_indices_float_35_cast_fp16_to_int32 = cast(dtype = var_2479_dtype_0, x = write_indices_float_35_cast_fp16)[name = tensor("cast_635")]; + tensor var_2485 = reshape(shape = var_2484, x = write_indices_float_35_cast_fp16_to_int32)[name = tensor("op_2485")]; + tensor write_indices_11_reps_0 = const()[name = tensor("write_indices_11_reps_0"), val = tensor([1, 1, 16, 64])]; + tensor write_indices_11 = tile(reps = write_indices_11_reps_0, x = var_2485)[name = tensor("write_indices_11")]; + tensor var_2495_begin_0 = const()[name = tensor("op_2495_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_2495_end_0 = const()[name = tensor("op_2495_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor var_2495_end_mask_0 = const()[name = tensor("op_2495_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_2495_squeeze_mask_0 = const()[name = tensor("op_2495_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor cache5_to_fp16_dtype_0 = const()[name = tensor("cache5_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor cache5_to_fp16 = cast(dtype = cache5_to_fp16_dtype_0, x = cache5)[name = tensor("cast_634")]; + tensor var_2495_cast_fp16 = slice_by_index(begin = var_2495_begin_0, end = var_2495_end_0, end_mask = var_2495_end_mask_0, squeeze_mask = var_2495_squeeze_mask_0, x = cache5_to_fp16)[name = tensor("op_2495_cast_fp16")]; + tensor var_2497_axis_0 = const()[name = tensor("op_2497_axis_0"), val = tensor(1)]; + tensor var_2497_mode_0 = const()[name = tensor("op_2497_mode_0"), val = tensor("update")]; + tensor var_2497_validate_indices_0 = const()[name = tensor("op_2497_validate_indices_0"), val = tensor(false)]; + tensor var_2497_cast_fp16 = scatter_along_axis(axis = var_2497_axis_0, data = var_2495_cast_fp16, indices = write_indices_11, mode = var_2497_mode_0, updates = k_23_cast_fp16, validate_indices = var_2497_validate_indices_0)[name = tensor("op_2497_cast_fp16")]; + tensor concat_42 = const()[name = tensor("concat_42"), val = tensor([0, 0, 0, 0, 0])]; + tensor concat_43 = const()[name = tensor("concat_43"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_11_internal_tensor_assign_1_stride_0 = const()[name = tensor("new_cache_11_internal_tensor_assign_1_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_11_internal_tensor_assign_1_begin_mask_0 = const()[name = tensor("new_cache_11_internal_tensor_assign_1_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_11_internal_tensor_assign_1_end_mask_0 = const()[name = tensor("new_cache_11_internal_tensor_assign_1_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_11_internal_tensor_assign_1_squeeze_mask_0 = const()[name = tensor("new_cache_11_internal_tensor_assign_1_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_56 = const()[name = tensor("shape_56"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_10 = const()[name = tensor("reduce_prod_10"), val = tensor(1048576)]; + tensor range_1d_10_start_0 = const()[name = tensor("range_1d_10_start_0"), val = tensor(0)]; + tensor range_1d_10_step_0 = const()[name = tensor("range_1d_10_step_0"), val = tensor(1)]; + tensor range_1d_10 = range_1d(end = reduce_prod_10, start = range_1d_10_start_0, step = range_1d_10_step_0)[name = tensor("range_1d_10")]; + tensor reshape_50 = reshape(shape = shape_56, x = range_1d_10)[name = tensor("reshape_50")]; + tensor slice_by_index_10 = slice_by_index(begin = concat_42, begin_mask = new_cache_11_internal_tensor_assign_1_begin_mask_0, end = concat_43, end_mask = new_cache_11_internal_tensor_assign_1_end_mask_0, squeeze_mask = new_cache_11_internal_tensor_assign_1_squeeze_mask_0, stride = new_cache_11_internal_tensor_assign_1_stride_0, x = reshape_50)[name = tensor("slice_by_index_10")]; + tensor reshape_51_shape_0 = const()[name = tensor("reshape_51_shape_0"), val = tensor([-1])]; + tensor reshape_51 = reshape(shape = reshape_51_shape_0, x = slice_by_index_10)[name = tensor("reshape_51")]; + tensor reshape_52_shape_0 = const()[name = tensor("reshape_52_shape_0"), val = tensor([-1])]; + tensor reshape_52_cast_fp16 = reshape(shape = reshape_52_shape_0, x = var_2497_cast_fp16)[name = tensor("reshape_52_cast_fp16")]; + tensor reshape_53_shape_0 = const()[name = tensor("reshape_53_shape_0"), val = tensor([-1])]; + tensor reshape_53_cast_fp16 = reshape(shape = reshape_53_shape_0, x = cache5_to_fp16)[name = tensor("reshape_53_cast_fp16")]; + tensor scatter_10_mode_0 = const()[name = tensor("scatter_10_mode_0"), val = tensor("update")]; + tensor scatter_10_axis_0 = const()[name = tensor("scatter_10_axis_0"), val = tensor(0)]; + tensor scatter_10_validate_indices_0 = const()[name = tensor("scatter_10_validate_indices_0"), val = tensor(false)]; + tensor scatter_10_cast_fp16 = scatter(axis = scatter_10_axis_0, data = reshape_53_cast_fp16, indices = reshape_51, mode = scatter_10_mode_0, updates = reshape_52_cast_fp16, validate_indices = scatter_10_validate_indices_0)[name = tensor("scatter_10_cast_fp16")]; + tensor reshape_54_cast_fp16 = reshape(shape = shape_56, x = scatter_10_cast_fp16)[name = tensor("reshape_54_cast_fp16")]; + tensor var_2505_begin_0 = const()[name = tensor("op_2505_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor var_2505_end_0 = const()[name = tensor("op_2505_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor var_2505_end_mask_0 = const()[name = tensor("op_2505_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_2505_squeeze_mask_0 = const()[name = tensor("op_2505_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor var_2505_cast_fp16 = slice_by_index(begin = var_2505_begin_0, end = var_2505_end_0, end_mask = var_2505_end_mask_0, squeeze_mask = var_2505_squeeze_mask_0, x = reshape_54_cast_fp16)[name = tensor("op_2505_cast_fp16")]; + tensor var_2507_axis_0 = const()[name = tensor("op_2507_axis_0"), val = tensor(1)]; + tensor var_2507_mode_0 = const()[name = tensor("op_2507_mode_0"), val = tensor("update")]; + tensor var_2507_validate_indices_0 = const()[name = tensor("op_2507_validate_indices_0"), val = tensor(false)]; + tensor var_2507_cast_fp16 = scatter_along_axis(axis = var_2507_axis_0, data = var_2505_cast_fp16, indices = write_indices_11, mode = var_2507_mode_0, updates = v_11_cast_fp16, validate_indices = var_2507_validate_indices_0)[name = tensor("op_2507_cast_fp16")]; + tensor concat_44 = const()[name = tensor("concat_44"), val = tensor([1, 0, 0, 0, 0])]; + tensor concat_45 = const()[name = tensor("concat_45"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_11_internal_tensor_assign_2_stride_0 = const()[name = tensor("new_cache_11_internal_tensor_assign_2_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_11_internal_tensor_assign_2_begin_mask_0 = const()[name = tensor("new_cache_11_internal_tensor_assign_2_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_11_internal_tensor_assign_2_end_mask_0 = const()[name = tensor("new_cache_11_internal_tensor_assign_2_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_11_internal_tensor_assign_2_squeeze_mask_0 = const()[name = tensor("new_cache_11_internal_tensor_assign_2_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_57 = const()[name = tensor("shape_57"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_11 = const()[name = tensor("reduce_prod_11"), val = tensor(1048576)]; + tensor range_1d_11_start_0 = const()[name = tensor("range_1d_11_start_0"), val = tensor(0)]; + tensor range_1d_11_step_0 = const()[name = tensor("range_1d_11_step_0"), val = tensor(1)]; + tensor range_1d_11 = range_1d(end = reduce_prod_11, start = range_1d_11_start_0, step = range_1d_11_step_0)[name = tensor("range_1d_11")]; + tensor reshape_55 = reshape(shape = shape_57, x = range_1d_11)[name = tensor("reshape_55")]; + tensor slice_by_index_11 = slice_by_index(begin = concat_44, begin_mask = new_cache_11_internal_tensor_assign_2_begin_mask_0, end = concat_45, end_mask = new_cache_11_internal_tensor_assign_2_end_mask_0, squeeze_mask = new_cache_11_internal_tensor_assign_2_squeeze_mask_0, stride = new_cache_11_internal_tensor_assign_2_stride_0, x = reshape_55)[name = tensor("slice_by_index_11")]; + tensor reshape_56_shape_0 = const()[name = tensor("reshape_56_shape_0"), val = tensor([-1])]; + tensor reshape_56 = reshape(shape = reshape_56_shape_0, x = slice_by_index_11)[name = tensor("reshape_56")]; + tensor reshape_57_shape_0 = const()[name = tensor("reshape_57_shape_0"), val = tensor([-1])]; + tensor reshape_57_cast_fp16 = reshape(shape = reshape_57_shape_0, x = var_2507_cast_fp16)[name = tensor("reshape_57_cast_fp16")]; + tensor reshape_58_shape_0 = const()[name = tensor("reshape_58_shape_0"), val = tensor([-1])]; + tensor reshape_58_cast_fp16 = reshape(shape = reshape_58_shape_0, x = reshape_54_cast_fp16)[name = tensor("reshape_58_cast_fp16")]; + tensor scatter_11_mode_0 = const()[name = tensor("scatter_11_mode_0"), val = tensor("update")]; + tensor scatter_11_axis_0 = const()[name = tensor("scatter_11_axis_0"), val = tensor(0)]; + tensor scatter_11_validate_indices_0 = const()[name = tensor("scatter_11_validate_indices_0"), val = tensor(false)]; + tensor scatter_11_cast_fp16 = scatter(axis = scatter_11_axis_0, data = reshape_58_cast_fp16, indices = reshape_56, mode = scatter_11_mode_0, updates = reshape_57_cast_fp16, validate_indices = scatter_11_validate_indices_0)[name = tensor("scatter_11_cast_fp16")]; + tensor reshape_59_cast_fp16 = reshape(shape = shape_57, x = scatter_11_cast_fp16)[name = tensor("reshape_59_cast_fp16")]; + tensor reshape_59_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("reshape_59_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor keys_31_begin_0 = const()[name = tensor("keys_31_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor keys_31_end_0 = const()[name = tensor("keys_31_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor keys_31_end_mask_0 = const()[name = tensor("keys_31_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor keys_31_squeeze_mask_0 = const()[name = tensor("keys_31_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor keys_31_cast_fp16 = slice_by_index(begin = keys_31_begin_0, end = keys_31_end_0, end_mask = keys_31_end_mask_0, squeeze_mask = keys_31_squeeze_mask_0, x = reshape_59_cast_fp16)[name = tensor("keys_31_cast_fp16")]; + tensor values_31_begin_0 = const()[name = tensor("values_31_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor values_31_end_0 = const()[name = tensor("values_31_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor values_31_end_mask_0 = const()[name = tensor("values_31_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor values_31_squeeze_mask_0 = const()[name = tensor("values_31_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor values_31_cast_fp16 = slice_by_index(begin = values_31_begin_0, end = values_31_end_0, end_mask = values_31_end_mask_0, squeeze_mask = values_31_squeeze_mask_0, x = reshape_59_cast_fp16)[name = tensor("values_31_cast_fp16")]; + tensor var_2519_cast_fp16 = not_equal(x = keys_31_cast_fp16, y = keys_31_cast_fp16)[name = tensor("op_2519_cast_fp16")]; + tensor keys_33_cast_fp16 = select(a = var_470_to_fp16, b = keys_31_cast_fp16, cond = var_2519_cast_fp16)[name = tensor("keys_33_cast_fp16")]; + tensor var_2527_cast_fp16 = not_equal(x = values_31_cast_fp16, y = values_31_cast_fp16)[name = tensor("op_2527_cast_fp16")]; + tensor values_33_cast_fp16 = select(a = var_470_to_fp16, b = values_31_cast_fp16, cond = var_2527_cast_fp16)[name = tensor("values_33_cast_fp16")]; + tensor var_2551 = const()[name = tensor("op_2551"), val = tensor([0, 2, 1, 3])]; + tensor var_2567 = const()[name = tensor("op_2567"), val = tensor([1, 1, 1])]; + tensor var_2568_cast_fp16 = reshape(shape = var_2567, x = position5_to_fp16)[name = tensor("op_2568_cast_fp16")]; + tensor q_positions_11_cast_fp16 = add(x = var_2568_cast_fp16, y = q_offsets_1_promoted_to_fp16)[name = tensor("q_positions_11_cast_fp16")]; + tensor valid_end_11_cast_fp16 = add(x = var_2568_cast_fp16, y = var_536_cast_fp16)[name = tensor("valid_end_11_cast_fp16")]; + tensor valid_mask_11_cast_fp16 = less(x = k_positions_1_promoted_to_fp16, y = valid_end_11_cast_fp16)[name = tensor("valid_mask_11_cast_fp16")]; + tensor causal_mask_11_cast_fp16 = less_equal(x = k_positions_1_promoted_to_fp16, y = q_positions_11_cast_fp16)[name = tensor("causal_mask_11_cast_fp16")]; + tensor var_2596 = logical_and(x = valid_mask_11_cast_fp16, y = causal_mask_11_cast_fp16)[name = tensor("op_2596")]; + tensor attn_mask_11_axes_0 = const()[name = tensor("attn_mask_11_axes_0"), val = tensor([1])]; + tensor attn_mask_11 = expand_dims(axes = attn_mask_11_axes_0, x = var_2596)[name = tensor("attn_mask_11")]; + tensor var_2610_transpose_x_0 = const()[name = tensor("op_2610_transpose_x_0"), val = tensor(false)]; + tensor var_2610_transpose_y_0 = const()[name = tensor("op_2610_transpose_y_0"), val = tensor(false)]; + tensor transpose_79_perm_0 = const()[name = tensor("transpose_79_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_80_perm_0 = const()[name = tensor("transpose_80_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_80 = transpose(perm = transpose_80_perm_0, x = keys_33_cast_fp16)[name = tensor("transpose_184")]; + tensor transpose_79 = transpose(perm = transpose_79_perm_0, x = q_33_cast_fp16)[name = tensor("transpose_185")]; + tensor var_2610_cast_fp16 = matmul(transpose_x = var_2610_transpose_x_0, transpose_y = var_2610_transpose_y_0, x = transpose_79, y = transpose_80)[name = tensor("op_2610_cast_fp16")]; + tensor var_2604_to_fp16 = const()[name = tensor("op_2604_to_fp16"), val = tensor([0x1p-3])]; + tensor attn_weights_31_cast_fp16 = mul(x = var_2610_cast_fp16, y = var_2604_to_fp16)[name = tensor("attn_weights_31_cast_fp16")]; + tensor var_2612 = logical_not(x = attn_mask_11)[name = tensor("op_2612")]; + tensor var_2613_to_fp16 = const()[name = tensor("op_2613_to_fp16"), val = tensor(-inf)]; + tensor attn_weights_33_cast_fp16 = select(a = var_2613_to_fp16, b = attn_weights_31_cast_fp16, cond = var_2612)[name = tensor("attn_weights_33_cast_fp16")]; + tensor var_2615 = const()[name = tensor("op_2615"), val = tensor(-1)]; + tensor attn_weights_35_cast_fp16 = softmax(axis = var_2615, x = attn_weights_33_cast_fp16)[name = tensor("attn_weights_35_cast_fp16")]; + tensor attn_output_11_transpose_x_0 = const()[name = tensor("attn_output_11_transpose_x_0"), val = tensor(false)]; + tensor attn_output_11_transpose_y_0 = const()[name = tensor("attn_output_11_transpose_y_0"), val = tensor(false)]; + tensor values_35_cast_fp16 = transpose(perm = var_2551, x = values_33_cast_fp16)[name = tensor("transpose_186")]; + tensor attn_output_11_cast_fp16 = matmul(transpose_x = attn_output_11_transpose_x_0, transpose_y = attn_output_11_transpose_y_0, x = attn_weights_35_cast_fp16, y = values_35_cast_fp16)[name = tensor("attn_output_11_cast_fp16")]; + tensor var_2623 = const()[name = tensor("op_2623"), val = tensor([0, 2, 1, 3])]; + tensor var_2628 = const()[name = tensor("op_2628"), val = tensor([1, 256, 1024])]; + tensor var_2624_cast_fp16 = transpose(perm = var_2623, x = attn_output_11_cast_fp16)[name = tensor("transpose_183")]; + tensor input_65_cast_fp16 = reshape(shape = var_2628, x = var_2624_cast_fp16)[name = tensor("input_65_cast_fp16")]; + tensor attn5_out_proj_weight_to_fp16 = const()[name = tensor("attn5_out_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(133237248)))]; + tensor linear_21_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = attn5_out_proj_weight_to_fp16, x = input_65_cast_fp16)[name = tensor("linear_21_cast_fp16")]; + tensor var_2637_cast_fp16 = add(x = position5_to_fp16, y = valid_len_to_fp16)[name = tensor("op_2637_cast_fp16")]; + tensor var_2637_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_2637_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor input_67_cast_fp16 = add(x = input_61_cast_fp16, y = linear_21_cast_fp16)[name = tensor("input_67_cast_fp16")]; + tensor input_69_axes_0 = const()[name = tensor("input_69_axes_0"), val = tensor([-1])]; + tensor norm5_2_weight_to_fp16 = const()[name = tensor("norm5_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(135334464)))]; + tensor norm5_2_bias_to_fp16 = const()[name = tensor("norm5_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(135336576)))]; + tensor var_2641_to_fp16 = const()[name = tensor("op_2641_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_69_cast_fp16 = layer_norm(axes = input_69_axes_0, beta = norm5_2_bias_to_fp16, epsilon = var_2641_to_fp16, gamma = norm5_2_weight_to_fp16, x = input_67_cast_fp16)[name = tensor("input_69_cast_fp16")]; + tensor linear5_1_weight_to_fp16 = const()[name = tensor("linear5_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(135338688)))]; + tensor linear_22_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = linear5_1_weight_to_fp16, x = input_69_cast_fp16)[name = tensor("linear_22_cast_fp16")]; + tensor input_71_mode_0 = const()[name = tensor("input_71_mode_0"), val = tensor("EXACT")]; + tensor input_71_cast_fp16 = gelu(mode = input_71_mode_0, x = linear_22_cast_fp16)[name = tensor("input_71_cast_fp16")]; + tensor linear5_2_weight_to_fp16 = const()[name = tensor("linear5_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(143727360)))]; + tensor linear_23_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = linear5_2_weight_to_fp16, x = input_71_cast_fp16)[name = tensor("linear_23_cast_fp16")]; + tensor input_73_cast_fp16 = add(x = input_67_cast_fp16, y = linear_23_cast_fp16)[name = tensor("input_73_cast_fp16")]; + tensor input_75_axes_0 = const()[name = tensor("input_75_axes_0"), val = tensor([-1])]; + tensor norm6_1_weight_to_fp16 = const()[name = tensor("norm6_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(152116032)))]; + tensor norm6_1_bias_to_fp16 = const()[name = tensor("norm6_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(152118144)))]; + tensor var_2658_to_fp16 = const()[name = tensor("op_2658_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_75_cast_fp16 = layer_norm(axes = input_75_axes_0, beta = norm6_1_bias_to_fp16, epsilon = var_2658_to_fp16, gamma = norm6_1_weight_to_fp16, x = input_73_cast_fp16)[name = tensor("input_75_cast_fp16")]; + tensor attn6_in_proj_weight_to_fp16 = const()[name = tensor("attn6_in_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(152120256)))]; + tensor linear_24_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = attn6_in_proj_weight_to_fp16, x = input_75_cast_fp16)[name = tensor("linear_24_cast_fp16")]; + tensor var_2672 = const()[name = tensor("op_2672"), val = tensor([1, 256, 3, 16, 64])]; + tensor qkv_13_cast_fp16 = reshape(shape = var_2672, x = linear_24_cast_fp16)[name = tensor("qkv_13_cast_fp16")]; + tensor q_37_begin_0 = const()[name = tensor("q_37_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor q_37_end_0 = const()[name = tensor("q_37_end_0"), val = tensor([1, 256, 1, 16, 64])]; + tensor q_37_end_mask_0 = const()[name = tensor("q_37_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor q_37_squeeze_mask_0 = const()[name = tensor("q_37_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor q_37_cast_fp16 = slice_by_index(begin = q_37_begin_0, end = q_37_end_0, end_mask = q_37_end_mask_0, squeeze_mask = q_37_squeeze_mask_0, x = qkv_13_cast_fp16)[name = tensor("q_37_cast_fp16")]; + tensor k_25_begin_0 = const()[name = tensor("k_25_begin_0"), val = tensor([0, 0, 1, 0, 0])]; + tensor k_25_end_0 = const()[name = tensor("k_25_end_0"), val = tensor([1, 256, 2, 16, 64])]; + tensor k_25_end_mask_0 = const()[name = tensor("k_25_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor k_25_squeeze_mask_0 = const()[name = tensor("k_25_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor k_25_cast_fp16 = slice_by_index(begin = k_25_begin_0, end = k_25_end_0, end_mask = k_25_end_mask_0, squeeze_mask = k_25_squeeze_mask_0, x = qkv_13_cast_fp16)[name = tensor("k_25_cast_fp16")]; + tensor v_13_begin_0 = const()[name = tensor("v_13_begin_0"), val = tensor([0, 0, 2, 0, 0])]; + tensor v_13_end_0 = const()[name = tensor("v_13_end_0"), val = tensor([1, 256, 3, 16, 64])]; + tensor v_13_end_mask_0 = const()[name = tensor("v_13_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor v_13_squeeze_mask_0 = const()[name = tensor("v_13_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor v_13_cast_fp16 = slice_by_index(begin = v_13_begin_0, end = v_13_end_0, end_mask = v_13_end_mask_0, squeeze_mask = v_13_squeeze_mask_0, x = qkv_13_cast_fp16)[name = tensor("v_13_cast_fp16")]; + tensor var_2730 = const()[name = tensor("op_2730"), val = tensor([1, 1])]; + tensor position6_to_fp16_dtype_0 = const()[name = tensor("position6_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor position6_to_fp16 = cast(dtype = position6_to_fp16_dtype_0, x = position6)[name = tensor("cast_631")]; + tensor var_2731_cast_fp16 = reshape(shape = var_2730, x = position6_to_fp16)[name = tensor("op_2731_cast_fp16")]; + tensor ts_39_cast_fp16 = add(x = ts_1_promoted_to_fp16, y = var_2731_cast_fp16)[name = tensor("ts_39_cast_fp16")]; + tensor var_2738 = const()[name = tensor("op_2738"), val = tensor([1, 256, 1, 1])]; + tensor ts_41_cast_fp16 = reshape(shape = var_2738, x = ts_39_cast_fp16)[name = tensor("ts_41_cast_fp16")]; + tensor var_2745 = const()[name = tensor("op_2745"), val = tensor([1, 256, 16, 32, 2])]; + tensor q_complex_13_cast_fp16 = reshape(shape = var_2745, x = q_37_cast_fp16)[name = tensor("q_complex_13_cast_fp16")]; + tensor var_2752 = const()[name = tensor("op_2752"), val = tensor([1, 256, 16, 32, 2])]; + tensor k_complex_13_cast_fp16 = reshape(shape = var_2752, x = k_25_cast_fp16)[name = tensor("k_complex_13_cast_fp16")]; + tensor var_2756_begin_0 = const()[name = tensor("op_2756_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_2756_end_0 = const()[name = tensor("op_2756_end_0"), val = tensor([1, 256, 16, 32, 1])]; + tensor var_2756_end_mask_0 = const()[name = tensor("op_2756_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_2756_squeeze_mask_0 = const()[name = tensor("op_2756_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_2756_cast_fp16 = slice_by_index(begin = var_2756_begin_0, end = var_2756_end_0, end_mask = var_2756_end_mask_0, squeeze_mask = var_2756_squeeze_mask_0, x = q_complex_13_cast_fp16)[name = tensor("op_2756_cast_fp16")]; + tensor var_2764_begin_0 = const()[name = tensor("op_2764_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_2764_end_0 = const()[name = tensor("op_2764_end_0"), val = tensor([1, 256, 16, 32, 2])]; + tensor var_2764_end_mask_0 = const()[name = tensor("op_2764_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_2764_squeeze_mask_0 = const()[name = tensor("op_2764_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_2764_cast_fp16 = slice_by_index(begin = var_2764_begin_0, end = var_2764_end_0, end_mask = var_2764_end_mask_0, squeeze_mask = var_2764_squeeze_mask_0, x = q_complex_13_cast_fp16)[name = tensor("op_2764_cast_fp16")]; + tensor var_2772_begin_0 = const()[name = tensor("op_2772_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_2772_end_0 = const()[name = tensor("op_2772_end_0"), val = tensor([1, 256, 16, 32, 1])]; + tensor var_2772_end_mask_0 = const()[name = tensor("op_2772_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_2772_squeeze_mask_0 = const()[name = tensor("op_2772_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_2772_cast_fp16 = slice_by_index(begin = var_2772_begin_0, end = var_2772_end_0, end_mask = var_2772_end_mask_0, squeeze_mask = var_2772_squeeze_mask_0, x = k_complex_13_cast_fp16)[name = tensor("op_2772_cast_fp16")]; + tensor var_2780_begin_0 = const()[name = tensor("op_2780_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_2780_end_0 = const()[name = tensor("op_2780_end_0"), val = tensor([1, 256, 16, 32, 2])]; + tensor var_2780_end_mask_0 = const()[name = tensor("op_2780_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_2780_squeeze_mask_0 = const()[name = tensor("op_2780_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_2780_cast_fp16 = slice_by_index(begin = var_2780_begin_0, end = var_2780_end_0, end_mask = var_2780_end_mask_0, squeeze_mask = var_2780_squeeze_mask_0, x = k_complex_13_cast_fp16)[name = tensor("op_2780_cast_fp16")]; + tensor freqs_13_to_fp16 = const()[name = tensor("freqs_13_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(158411776)))]; + tensor var_2786_cast_fp16 = mul(x = freqs_13_to_fp16, y = ts_41_cast_fp16)[name = tensor("op_2786_cast_fp16")]; + tensor rotr_13_cast_fp16 = cos(x = var_2786_cast_fp16)[name = tensor("rotr_13_cast_fp16")]; + tensor roti_13_cast_fp16 = sin(x = var_2786_cast_fp16)[name = tensor("roti_13_cast_fp16")]; + tensor var_2790_cast_fp16 = mul(x = var_2756_cast_fp16, y = rotr_13_cast_fp16)[name = tensor("op_2790_cast_fp16")]; + tensor var_2791_cast_fp16 = mul(x = var_2764_cast_fp16, y = roti_13_cast_fp16)[name = tensor("op_2791_cast_fp16")]; + tensor qor_25_cast_fp16 = sub(x = var_2790_cast_fp16, y = var_2791_cast_fp16)[name = tensor("qor_25_cast_fp16")]; + tensor var_2794_cast_fp16 = mul(x = var_2756_cast_fp16, y = roti_13_cast_fp16)[name = tensor("op_2794_cast_fp16")]; + tensor var_2795_cast_fp16 = mul(x = var_2764_cast_fp16, y = rotr_13_cast_fp16)[name = tensor("op_2795_cast_fp16")]; + tensor qoi_25_cast_fp16 = add(x = var_2794_cast_fp16, y = var_2795_cast_fp16)[name = tensor("qoi_25_cast_fp16")]; + tensor var_2798_cast_fp16 = mul(x = var_2772_cast_fp16, y = rotr_13_cast_fp16)[name = tensor("op_2798_cast_fp16")]; + tensor var_2799_cast_fp16 = mul(x = var_2780_cast_fp16, y = roti_13_cast_fp16)[name = tensor("op_2799_cast_fp16")]; + tensor kor_25_cast_fp16 = sub(x = var_2798_cast_fp16, y = var_2799_cast_fp16)[name = tensor("kor_25_cast_fp16")]; + tensor var_2802_cast_fp16 = mul(x = var_2772_cast_fp16, y = roti_13_cast_fp16)[name = tensor("op_2802_cast_fp16")]; + tensor var_2803_cast_fp16 = mul(x = var_2780_cast_fp16, y = rotr_13_cast_fp16)[name = tensor("op_2803_cast_fp16")]; + tensor koi_25_cast_fp16 = add(x = var_2802_cast_fp16, y = var_2803_cast_fp16)[name = tensor("koi_25_cast_fp16")]; + tensor qo_13_axis_0 = const()[name = tensor("qo_13_axis_0"), val = tensor(-1)]; + tensor qo_13_cast_fp16 = stack(axis = qo_13_axis_0, values = (qor_25_cast_fp16, qoi_25_cast_fp16))[name = tensor("qo_13_cast_fp16")]; + tensor ko_13_axis_0 = const()[name = tensor("ko_13_axis_0"), val = tensor(-1)]; + tensor ko_13_cast_fp16 = stack(axis = ko_13_axis_0, values = (kor_25_cast_fp16, koi_25_cast_fp16))[name = tensor("ko_13_cast_fp16")]; + tensor var_2836 = const()[name = tensor("op_2836"), val = tensor([1, 256, 16, 64])]; + tensor q_39_cast_fp16 = reshape(shape = var_2836, x = qo_13_cast_fp16)[name = tensor("q_39_cast_fp16")]; + tensor var_2842 = const()[name = tensor("op_2842"), val = tensor([1, 256, 16, 64])]; + tensor k_27_cast_fp16 = reshape(shape = var_2842, x = ko_13_cast_fp16)[name = tensor("k_27_cast_fp16")]; + tensor write_indices_float_37_cast_fp16 = add(x = var_2731_cast_fp16, y = var_393_promoted_to_fp16)[name = tensor("write_indices_float_37_cast_fp16")]; + tensor _inversed_2863_y_0_to_fp16 = const()[name = tensor("_inversed_2863_y_0_to_fp16"), val = tensor(0x1p-9)]; + tensor _inversed_2863_cast_fp16 = mul(x = write_indices_float_37_cast_fp16, y = _inversed_2863_y_0_to_fp16)[name = tensor("_inversed_2863_cast_fp16")]; + tensor var_2864_cast_fp16 = floor(x = _inversed_2863_cast_fp16)[name = tensor("op_2864_cast_fp16")]; + tensor var_2865_to_fp16 = const()[name = tensor("op_2865_to_fp16"), val = tensor(0x1p+9)]; + tensor var_2866_cast_fp16 = mul(x = var_2864_cast_fp16, y = var_2865_to_fp16)[name = tensor("op_2866_cast_fp16")]; + tensor write_indices_float_39_cast_fp16 = sub(x = write_indices_float_37_cast_fp16, y = var_2866_cast_fp16)[name = tensor("write_indices_float_39_cast_fp16")]; + tensor var_2884_value_0_to_fp16 = const()[name = tensor("op_2884_value_0_to_fp16"), val = tensor(0x1.ffp+8)]; + tensor var_2884_cast_fp16 = fill_like(ref_tensor = write_indices_float_39_cast_fp16, value = var_2884_value_0_to_fp16)[name = tensor("op_2884_cast_fp16")]; + tensor write_indices_float_41_cast_fp16 = select(a = write_indices_float_39_cast_fp16, b = var_2884_cast_fp16, cond = token_valid_1_cast_fp16)[name = tensor("write_indices_float_41_cast_fp16")]; + tensor var_2890_dtype_0 = const()[name = tensor("op_2890_dtype_0"), val = tensor("int32")]; + tensor var_2895 = const()[name = tensor("op_2895"), val = tensor([1, 256, 1, 1])]; + tensor write_indices_float_41_cast_fp16_to_int32 = cast(dtype = var_2890_dtype_0, x = write_indices_float_41_cast_fp16)[name = tensor("cast_630")]; + tensor var_2896 = reshape(shape = var_2895, x = write_indices_float_41_cast_fp16_to_int32)[name = tensor("op_2896")]; + tensor write_indices_13_reps_0 = const()[name = tensor("write_indices_13_reps_0"), val = tensor([1, 1, 16, 64])]; + tensor write_indices_13 = tile(reps = write_indices_13_reps_0, x = var_2896)[name = tensor("write_indices_13")]; + tensor var_2906_begin_0 = const()[name = tensor("op_2906_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_2906_end_0 = const()[name = tensor("op_2906_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor var_2906_end_mask_0 = const()[name = tensor("op_2906_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_2906_squeeze_mask_0 = const()[name = tensor("op_2906_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor cache6_to_fp16_dtype_0 = const()[name = tensor("cache6_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor cache6_to_fp16 = cast(dtype = cache6_to_fp16_dtype_0, x = cache6)[name = tensor("cast_629")]; + tensor var_2906_cast_fp16 = slice_by_index(begin = var_2906_begin_0, end = var_2906_end_0, end_mask = var_2906_end_mask_0, squeeze_mask = var_2906_squeeze_mask_0, x = cache6_to_fp16)[name = tensor("op_2906_cast_fp16")]; + tensor var_2908_axis_0 = const()[name = tensor("op_2908_axis_0"), val = tensor(1)]; + tensor var_2908_mode_0 = const()[name = tensor("op_2908_mode_0"), val = tensor("update")]; + tensor var_2908_validate_indices_0 = const()[name = tensor("op_2908_validate_indices_0"), val = tensor(false)]; + tensor var_2908_cast_fp16 = scatter_along_axis(axis = var_2908_axis_0, data = var_2906_cast_fp16, indices = write_indices_13, mode = var_2908_mode_0, updates = k_27_cast_fp16, validate_indices = var_2908_validate_indices_0)[name = tensor("op_2908_cast_fp16")]; + tensor concat_50 = const()[name = tensor("concat_50"), val = tensor([0, 0, 0, 0, 0])]; + tensor concat_51 = const()[name = tensor("concat_51"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_13_internal_tensor_assign_1_stride_0 = const()[name = tensor("new_cache_13_internal_tensor_assign_1_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_13_internal_tensor_assign_1_begin_mask_0 = const()[name = tensor("new_cache_13_internal_tensor_assign_1_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_13_internal_tensor_assign_1_end_mask_0 = const()[name = tensor("new_cache_13_internal_tensor_assign_1_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_13_internal_tensor_assign_1_squeeze_mask_0 = const()[name = tensor("new_cache_13_internal_tensor_assign_1_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_58 = const()[name = tensor("shape_58"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_12 = const()[name = tensor("reduce_prod_12"), val = tensor(1048576)]; + tensor range_1d_12_start_0 = const()[name = tensor("range_1d_12_start_0"), val = tensor(0)]; + tensor range_1d_12_step_0 = const()[name = tensor("range_1d_12_step_0"), val = tensor(1)]; + tensor range_1d_12 = range_1d(end = reduce_prod_12, start = range_1d_12_start_0, step = range_1d_12_step_0)[name = tensor("range_1d_12")]; + tensor reshape_60 = reshape(shape = shape_58, x = range_1d_12)[name = tensor("reshape_60")]; + tensor slice_by_index_12 = slice_by_index(begin = concat_50, begin_mask = new_cache_13_internal_tensor_assign_1_begin_mask_0, end = concat_51, end_mask = new_cache_13_internal_tensor_assign_1_end_mask_0, squeeze_mask = new_cache_13_internal_tensor_assign_1_squeeze_mask_0, stride = new_cache_13_internal_tensor_assign_1_stride_0, x = reshape_60)[name = tensor("slice_by_index_12")]; + tensor reshape_61_shape_0 = const()[name = tensor("reshape_61_shape_0"), val = tensor([-1])]; + tensor reshape_61 = reshape(shape = reshape_61_shape_0, x = slice_by_index_12)[name = tensor("reshape_61")]; + tensor reshape_62_shape_0 = const()[name = tensor("reshape_62_shape_0"), val = tensor([-1])]; + tensor reshape_62_cast_fp16 = reshape(shape = reshape_62_shape_0, x = var_2908_cast_fp16)[name = tensor("reshape_62_cast_fp16")]; + tensor reshape_63_shape_0 = const()[name = tensor("reshape_63_shape_0"), val = tensor([-1])]; + tensor reshape_63_cast_fp16 = reshape(shape = reshape_63_shape_0, x = cache6_to_fp16)[name = tensor("reshape_63_cast_fp16")]; + tensor scatter_12_mode_0 = const()[name = tensor("scatter_12_mode_0"), val = tensor("update")]; + tensor scatter_12_axis_0 = const()[name = tensor("scatter_12_axis_0"), val = tensor(0)]; + tensor scatter_12_validate_indices_0 = const()[name = tensor("scatter_12_validate_indices_0"), val = tensor(false)]; + tensor scatter_12_cast_fp16 = scatter(axis = scatter_12_axis_0, data = reshape_63_cast_fp16, indices = reshape_61, mode = scatter_12_mode_0, updates = reshape_62_cast_fp16, validate_indices = scatter_12_validate_indices_0)[name = tensor("scatter_12_cast_fp16")]; + tensor reshape_64_cast_fp16 = reshape(shape = shape_58, x = scatter_12_cast_fp16)[name = tensor("reshape_64_cast_fp16")]; + tensor var_2916_begin_0 = const()[name = tensor("op_2916_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor var_2916_end_0 = const()[name = tensor("op_2916_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor var_2916_end_mask_0 = const()[name = tensor("op_2916_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_2916_squeeze_mask_0 = const()[name = tensor("op_2916_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor var_2916_cast_fp16 = slice_by_index(begin = var_2916_begin_0, end = var_2916_end_0, end_mask = var_2916_end_mask_0, squeeze_mask = var_2916_squeeze_mask_0, x = reshape_64_cast_fp16)[name = tensor("op_2916_cast_fp16")]; + tensor var_2918_axis_0 = const()[name = tensor("op_2918_axis_0"), val = tensor(1)]; + tensor var_2918_mode_0 = const()[name = tensor("op_2918_mode_0"), val = tensor("update")]; + tensor var_2918_validate_indices_0 = const()[name = tensor("op_2918_validate_indices_0"), val = tensor(false)]; + tensor var_2918_cast_fp16 = scatter_along_axis(axis = var_2918_axis_0, data = var_2916_cast_fp16, indices = write_indices_13, mode = var_2918_mode_0, updates = v_13_cast_fp16, validate_indices = var_2918_validate_indices_0)[name = tensor("op_2918_cast_fp16")]; + tensor concat_52 = const()[name = tensor("concat_52"), val = tensor([1, 0, 0, 0, 0])]; + tensor concat_53 = const()[name = tensor("concat_53"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_13_internal_tensor_assign_2_stride_0 = const()[name = tensor("new_cache_13_internal_tensor_assign_2_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_13_internal_tensor_assign_2_begin_mask_0 = const()[name = tensor("new_cache_13_internal_tensor_assign_2_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_13_internal_tensor_assign_2_end_mask_0 = const()[name = tensor("new_cache_13_internal_tensor_assign_2_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_13_internal_tensor_assign_2_squeeze_mask_0 = const()[name = tensor("new_cache_13_internal_tensor_assign_2_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_59 = const()[name = tensor("shape_59"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_13 = const()[name = tensor("reduce_prod_13"), val = tensor(1048576)]; + tensor range_1d_13_start_0 = const()[name = tensor("range_1d_13_start_0"), val = tensor(0)]; + tensor range_1d_13_step_0 = const()[name = tensor("range_1d_13_step_0"), val = tensor(1)]; + tensor range_1d_13 = range_1d(end = reduce_prod_13, start = range_1d_13_start_0, step = range_1d_13_step_0)[name = tensor("range_1d_13")]; + tensor reshape_65 = reshape(shape = shape_59, x = range_1d_13)[name = tensor("reshape_65")]; + tensor slice_by_index_13 = slice_by_index(begin = concat_52, begin_mask = new_cache_13_internal_tensor_assign_2_begin_mask_0, end = concat_53, end_mask = new_cache_13_internal_tensor_assign_2_end_mask_0, squeeze_mask = new_cache_13_internal_tensor_assign_2_squeeze_mask_0, stride = new_cache_13_internal_tensor_assign_2_stride_0, x = reshape_65)[name = tensor("slice_by_index_13")]; + tensor reshape_66_shape_0 = const()[name = tensor("reshape_66_shape_0"), val = tensor([-1])]; + tensor reshape_66 = reshape(shape = reshape_66_shape_0, x = slice_by_index_13)[name = tensor("reshape_66")]; + tensor reshape_67_shape_0 = const()[name = tensor("reshape_67_shape_0"), val = tensor([-1])]; + tensor reshape_67_cast_fp16 = reshape(shape = reshape_67_shape_0, x = var_2918_cast_fp16)[name = tensor("reshape_67_cast_fp16")]; + tensor reshape_68_shape_0 = const()[name = tensor("reshape_68_shape_0"), val = tensor([-1])]; + tensor reshape_68_cast_fp16 = reshape(shape = reshape_68_shape_0, x = reshape_64_cast_fp16)[name = tensor("reshape_68_cast_fp16")]; + tensor scatter_13_mode_0 = const()[name = tensor("scatter_13_mode_0"), val = tensor("update")]; + tensor scatter_13_axis_0 = const()[name = tensor("scatter_13_axis_0"), val = tensor(0)]; + tensor scatter_13_validate_indices_0 = const()[name = tensor("scatter_13_validate_indices_0"), val = tensor(false)]; + tensor scatter_13_cast_fp16 = scatter(axis = scatter_13_axis_0, data = reshape_68_cast_fp16, indices = reshape_66, mode = scatter_13_mode_0, updates = reshape_67_cast_fp16, validate_indices = scatter_13_validate_indices_0)[name = tensor("scatter_13_cast_fp16")]; + tensor reshape_69_cast_fp16 = reshape(shape = shape_59, x = scatter_13_cast_fp16)[name = tensor("reshape_69_cast_fp16")]; + tensor reshape_69_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("reshape_69_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor keys_37_begin_0 = const()[name = tensor("keys_37_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor keys_37_end_0 = const()[name = tensor("keys_37_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor keys_37_end_mask_0 = const()[name = tensor("keys_37_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor keys_37_squeeze_mask_0 = const()[name = tensor("keys_37_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor keys_37_cast_fp16 = slice_by_index(begin = keys_37_begin_0, end = keys_37_end_0, end_mask = keys_37_end_mask_0, squeeze_mask = keys_37_squeeze_mask_0, x = reshape_69_cast_fp16)[name = tensor("keys_37_cast_fp16")]; + tensor values_37_begin_0 = const()[name = tensor("values_37_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor values_37_end_0 = const()[name = tensor("values_37_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor values_37_end_mask_0 = const()[name = tensor("values_37_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor values_37_squeeze_mask_0 = const()[name = tensor("values_37_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor values_37_cast_fp16 = slice_by_index(begin = values_37_begin_0, end = values_37_end_0, end_mask = values_37_end_mask_0, squeeze_mask = values_37_squeeze_mask_0, x = reshape_69_cast_fp16)[name = tensor("values_37_cast_fp16")]; + tensor var_2930_cast_fp16 = not_equal(x = keys_37_cast_fp16, y = keys_37_cast_fp16)[name = tensor("op_2930_cast_fp16")]; + tensor keys_39_cast_fp16 = select(a = var_470_to_fp16, b = keys_37_cast_fp16, cond = var_2930_cast_fp16)[name = tensor("keys_39_cast_fp16")]; + tensor var_2938_cast_fp16 = not_equal(x = values_37_cast_fp16, y = values_37_cast_fp16)[name = tensor("op_2938_cast_fp16")]; + tensor values_39_cast_fp16 = select(a = var_470_to_fp16, b = values_37_cast_fp16, cond = var_2938_cast_fp16)[name = tensor("values_39_cast_fp16")]; + tensor var_2962 = const()[name = tensor("op_2962"), val = tensor([0, 2, 1, 3])]; + tensor var_2978 = const()[name = tensor("op_2978"), val = tensor([1, 1, 1])]; + tensor var_2979_cast_fp16 = reshape(shape = var_2978, x = position6_to_fp16)[name = tensor("op_2979_cast_fp16")]; + tensor q_positions_13_cast_fp16 = add(x = var_2979_cast_fp16, y = q_offsets_1_promoted_to_fp16)[name = tensor("q_positions_13_cast_fp16")]; + tensor valid_end_13_cast_fp16 = add(x = var_2979_cast_fp16, y = var_536_cast_fp16)[name = tensor("valid_end_13_cast_fp16")]; + tensor valid_mask_13_cast_fp16 = less(x = k_positions_1_promoted_to_fp16, y = valid_end_13_cast_fp16)[name = tensor("valid_mask_13_cast_fp16")]; + tensor causal_mask_13_cast_fp16 = less_equal(x = k_positions_1_promoted_to_fp16, y = q_positions_13_cast_fp16)[name = tensor("causal_mask_13_cast_fp16")]; + tensor var_3007 = logical_and(x = valid_mask_13_cast_fp16, y = causal_mask_13_cast_fp16)[name = tensor("op_3007")]; + tensor attn_mask_13_axes_0 = const()[name = tensor("attn_mask_13_axes_0"), val = tensor([1])]; + tensor attn_mask_13 = expand_dims(axes = attn_mask_13_axes_0, x = var_3007)[name = tensor("attn_mask_13")]; + tensor var_3021_transpose_x_0 = const()[name = tensor("op_3021_transpose_x_0"), val = tensor(false)]; + tensor var_3021_transpose_y_0 = const()[name = tensor("op_3021_transpose_y_0"), val = tensor(false)]; + tensor transpose_81_perm_0 = const()[name = tensor("transpose_81_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_82_perm_0 = const()[name = tensor("transpose_82_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_82 = transpose(perm = transpose_82_perm_0, x = keys_39_cast_fp16)[name = tensor("transpose_180")]; + tensor transpose_81 = transpose(perm = transpose_81_perm_0, x = q_39_cast_fp16)[name = tensor("transpose_181")]; + tensor var_3021_cast_fp16 = matmul(transpose_x = var_3021_transpose_x_0, transpose_y = var_3021_transpose_y_0, x = transpose_81, y = transpose_82)[name = tensor("op_3021_cast_fp16")]; + tensor var_3015_to_fp16 = const()[name = tensor("op_3015_to_fp16"), val = tensor([0x1p-3])]; + tensor attn_weights_37_cast_fp16 = mul(x = var_3021_cast_fp16, y = var_3015_to_fp16)[name = tensor("attn_weights_37_cast_fp16")]; + tensor var_3023 = logical_not(x = attn_mask_13)[name = tensor("op_3023")]; + tensor var_3024_to_fp16 = const()[name = tensor("op_3024_to_fp16"), val = tensor(-inf)]; + tensor attn_weights_39_cast_fp16 = select(a = var_3024_to_fp16, b = attn_weights_37_cast_fp16, cond = var_3023)[name = tensor("attn_weights_39_cast_fp16")]; + tensor var_3026 = const()[name = tensor("op_3026"), val = tensor(-1)]; + tensor attn_weights_41_cast_fp16 = softmax(axis = var_3026, x = attn_weights_39_cast_fp16)[name = tensor("attn_weights_41_cast_fp16")]; + tensor attn_output_13_transpose_x_0 = const()[name = tensor("attn_output_13_transpose_x_0"), val = tensor(false)]; + tensor attn_output_13_transpose_y_0 = const()[name = tensor("attn_output_13_transpose_y_0"), val = tensor(false)]; + tensor values_41_cast_fp16 = transpose(perm = var_2962, x = values_39_cast_fp16)[name = tensor("transpose_182")]; + tensor attn_output_13_cast_fp16 = matmul(transpose_x = attn_output_13_transpose_x_0, transpose_y = attn_output_13_transpose_y_0, x = attn_weights_41_cast_fp16, y = values_41_cast_fp16)[name = tensor("attn_output_13_cast_fp16")]; + tensor var_3034 = const()[name = tensor("op_3034"), val = tensor([0, 2, 1, 3])]; + tensor var_3039 = const()[name = tensor("op_3039"), val = tensor([1, 256, 1024])]; + tensor var_3035_cast_fp16 = transpose(perm = var_3034, x = attn_output_13_cast_fp16)[name = tensor("transpose_179")]; + tensor input_77_cast_fp16 = reshape(shape = var_3039, x = var_3035_cast_fp16)[name = tensor("input_77_cast_fp16")]; + tensor attn6_out_proj_weight_to_fp16 = const()[name = tensor("attn6_out_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(158411904)))]; + tensor linear_25_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = attn6_out_proj_weight_to_fp16, x = input_77_cast_fp16)[name = tensor("linear_25_cast_fp16")]; + tensor var_3048_cast_fp16 = add(x = position6_to_fp16, y = valid_len_to_fp16)[name = tensor("op_3048_cast_fp16")]; + tensor var_3048_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_3048_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor input_79_cast_fp16 = add(x = input_73_cast_fp16, y = linear_25_cast_fp16)[name = tensor("input_79_cast_fp16")]; + tensor input_81_axes_0 = const()[name = tensor("input_81_axes_0"), val = tensor([-1])]; + tensor norm6_2_weight_to_fp16 = const()[name = tensor("norm6_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(160509120)))]; + tensor norm6_2_bias_to_fp16 = const()[name = tensor("norm6_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(160511232)))]; + tensor var_3052_to_fp16 = const()[name = tensor("op_3052_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_81_cast_fp16 = layer_norm(axes = input_81_axes_0, beta = norm6_2_bias_to_fp16, epsilon = var_3052_to_fp16, gamma = norm6_2_weight_to_fp16, x = input_79_cast_fp16)[name = tensor("input_81_cast_fp16")]; + tensor linear6_1_weight_to_fp16 = const()[name = tensor("linear6_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(160513344)))]; + tensor linear_26_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = linear6_1_weight_to_fp16, x = input_81_cast_fp16)[name = tensor("linear_26_cast_fp16")]; + tensor input_83_mode_0 = const()[name = tensor("input_83_mode_0"), val = tensor("EXACT")]; + tensor input_83_cast_fp16 = gelu(mode = input_83_mode_0, x = linear_26_cast_fp16)[name = tensor("input_83_cast_fp16")]; + tensor linear6_2_weight_to_fp16 = const()[name = tensor("linear6_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(168902016)))]; + tensor linear_27_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = linear6_2_weight_to_fp16, x = input_83_cast_fp16)[name = tensor("linear_27_cast_fp16")]; + tensor input_85_cast_fp16 = add(x = input_79_cast_fp16, y = linear_27_cast_fp16)[name = tensor("input_85_cast_fp16")]; + tensor input_87_axes_0 = const()[name = tensor("input_87_axes_0"), val = tensor([-1])]; + tensor norm7_1_weight_to_fp16 = const()[name = tensor("norm7_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(177290688)))]; + tensor norm7_1_bias_to_fp16 = const()[name = tensor("norm7_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(177292800)))]; + tensor var_3069_to_fp16 = const()[name = tensor("op_3069_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_87_cast_fp16 = layer_norm(axes = input_87_axes_0, beta = norm7_1_bias_to_fp16, epsilon = var_3069_to_fp16, gamma = norm7_1_weight_to_fp16, x = input_85_cast_fp16)[name = tensor("input_87_cast_fp16")]; + tensor attn7_in_proj_weight_to_fp16 = const()[name = tensor("attn7_in_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(177294912)))]; + tensor linear_28_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = attn7_in_proj_weight_to_fp16, x = input_87_cast_fp16)[name = tensor("linear_28_cast_fp16")]; + tensor var_3083 = const()[name = tensor("op_3083"), val = tensor([1, 256, 3, 16, 64])]; + tensor qkv_15_cast_fp16 = reshape(shape = var_3083, x = linear_28_cast_fp16)[name = tensor("qkv_15_cast_fp16")]; + tensor q_43_begin_0 = const()[name = tensor("q_43_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor q_43_end_0 = const()[name = tensor("q_43_end_0"), val = tensor([1, 256, 1, 16, 64])]; + tensor q_43_end_mask_0 = const()[name = tensor("q_43_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor q_43_squeeze_mask_0 = const()[name = tensor("q_43_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor q_43_cast_fp16 = slice_by_index(begin = q_43_begin_0, end = q_43_end_0, end_mask = q_43_end_mask_0, squeeze_mask = q_43_squeeze_mask_0, x = qkv_15_cast_fp16)[name = tensor("q_43_cast_fp16")]; + tensor k_29_begin_0 = const()[name = tensor("k_29_begin_0"), val = tensor([0, 0, 1, 0, 0])]; + tensor k_29_end_0 = const()[name = tensor("k_29_end_0"), val = tensor([1, 256, 2, 16, 64])]; + tensor k_29_end_mask_0 = const()[name = tensor("k_29_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor k_29_squeeze_mask_0 = const()[name = tensor("k_29_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor k_29_cast_fp16 = slice_by_index(begin = k_29_begin_0, end = k_29_end_0, end_mask = k_29_end_mask_0, squeeze_mask = k_29_squeeze_mask_0, x = qkv_15_cast_fp16)[name = tensor("k_29_cast_fp16")]; + tensor v_15_begin_0 = const()[name = tensor("v_15_begin_0"), val = tensor([0, 0, 2, 0, 0])]; + tensor v_15_end_0 = const()[name = tensor("v_15_end_0"), val = tensor([1, 256, 3, 16, 64])]; + tensor v_15_end_mask_0 = const()[name = tensor("v_15_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor v_15_squeeze_mask_0 = const()[name = tensor("v_15_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor v_15_cast_fp16 = slice_by_index(begin = v_15_begin_0, end = v_15_end_0, end_mask = v_15_end_mask_0, squeeze_mask = v_15_squeeze_mask_0, x = qkv_15_cast_fp16)[name = tensor("v_15_cast_fp16")]; + tensor var_3141 = const()[name = tensor("op_3141"), val = tensor([1, 1])]; + tensor position7_to_fp16_dtype_0 = const()[name = tensor("position7_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor position7_to_fp16 = cast(dtype = position7_to_fp16_dtype_0, x = position7)[name = tensor("cast_626")]; + tensor var_3142_cast_fp16 = reshape(shape = var_3141, x = position7_to_fp16)[name = tensor("op_3142_cast_fp16")]; + tensor ts_45_cast_fp16 = add(x = ts_1_promoted_to_fp16, y = var_3142_cast_fp16)[name = tensor("ts_45_cast_fp16")]; + tensor var_3149 = const()[name = tensor("op_3149"), val = tensor([1, 256, 1, 1])]; + tensor ts_47_cast_fp16 = reshape(shape = var_3149, x = ts_45_cast_fp16)[name = tensor("ts_47_cast_fp16")]; + tensor var_3156 = const()[name = tensor("op_3156"), val = tensor([1, 256, 16, 32, 2])]; + tensor q_complex_15_cast_fp16 = reshape(shape = var_3156, x = q_43_cast_fp16)[name = tensor("q_complex_15_cast_fp16")]; + tensor var_3163 = const()[name = tensor("op_3163"), val = tensor([1, 256, 16, 32, 2])]; + tensor k_complex_15_cast_fp16 = reshape(shape = var_3163, x = k_29_cast_fp16)[name = tensor("k_complex_15_cast_fp16")]; + tensor var_3167_begin_0 = const()[name = tensor("op_3167_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_3167_end_0 = const()[name = tensor("op_3167_end_0"), val = tensor([1, 256, 16, 32, 1])]; + tensor var_3167_end_mask_0 = const()[name = tensor("op_3167_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_3167_squeeze_mask_0 = const()[name = tensor("op_3167_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_3167_cast_fp16 = slice_by_index(begin = var_3167_begin_0, end = var_3167_end_0, end_mask = var_3167_end_mask_0, squeeze_mask = var_3167_squeeze_mask_0, x = q_complex_15_cast_fp16)[name = tensor("op_3167_cast_fp16")]; + tensor var_3175_begin_0 = const()[name = tensor("op_3175_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_3175_end_0 = const()[name = tensor("op_3175_end_0"), val = tensor([1, 256, 16, 32, 2])]; + tensor var_3175_end_mask_0 = const()[name = tensor("op_3175_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_3175_squeeze_mask_0 = const()[name = tensor("op_3175_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_3175_cast_fp16 = slice_by_index(begin = var_3175_begin_0, end = var_3175_end_0, end_mask = var_3175_end_mask_0, squeeze_mask = var_3175_squeeze_mask_0, x = q_complex_15_cast_fp16)[name = tensor("op_3175_cast_fp16")]; + tensor var_3183_begin_0 = const()[name = tensor("op_3183_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_3183_end_0 = const()[name = tensor("op_3183_end_0"), val = tensor([1, 256, 16, 32, 1])]; + tensor var_3183_end_mask_0 = const()[name = tensor("op_3183_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_3183_squeeze_mask_0 = const()[name = tensor("op_3183_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_3183_cast_fp16 = slice_by_index(begin = var_3183_begin_0, end = var_3183_end_0, end_mask = var_3183_end_mask_0, squeeze_mask = var_3183_squeeze_mask_0, x = k_complex_15_cast_fp16)[name = tensor("op_3183_cast_fp16")]; + tensor var_3191_begin_0 = const()[name = tensor("op_3191_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_3191_end_0 = const()[name = tensor("op_3191_end_0"), val = tensor([1, 256, 16, 32, 2])]; + tensor var_3191_end_mask_0 = const()[name = tensor("op_3191_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_3191_squeeze_mask_0 = const()[name = tensor("op_3191_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_3191_cast_fp16 = slice_by_index(begin = var_3191_begin_0, end = var_3191_end_0, end_mask = var_3191_end_mask_0, squeeze_mask = var_3191_squeeze_mask_0, x = k_complex_15_cast_fp16)[name = tensor("op_3191_cast_fp16")]; + tensor freqs_15_to_fp16 = const()[name = tensor("freqs_15_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(183586432)))]; + tensor var_3197_cast_fp16 = mul(x = freqs_15_to_fp16, y = ts_47_cast_fp16)[name = tensor("op_3197_cast_fp16")]; + tensor rotr_15_cast_fp16 = cos(x = var_3197_cast_fp16)[name = tensor("rotr_15_cast_fp16")]; + tensor roti_15_cast_fp16 = sin(x = var_3197_cast_fp16)[name = tensor("roti_15_cast_fp16")]; + tensor var_3201_cast_fp16 = mul(x = var_3167_cast_fp16, y = rotr_15_cast_fp16)[name = tensor("op_3201_cast_fp16")]; + tensor var_3202_cast_fp16 = mul(x = var_3175_cast_fp16, y = roti_15_cast_fp16)[name = tensor("op_3202_cast_fp16")]; + tensor qor_29_cast_fp16 = sub(x = var_3201_cast_fp16, y = var_3202_cast_fp16)[name = tensor("qor_29_cast_fp16")]; + tensor var_3205_cast_fp16 = mul(x = var_3167_cast_fp16, y = roti_15_cast_fp16)[name = tensor("op_3205_cast_fp16")]; + tensor var_3206_cast_fp16 = mul(x = var_3175_cast_fp16, y = rotr_15_cast_fp16)[name = tensor("op_3206_cast_fp16")]; + tensor qoi_29_cast_fp16 = add(x = var_3205_cast_fp16, y = var_3206_cast_fp16)[name = tensor("qoi_29_cast_fp16")]; + tensor var_3209_cast_fp16 = mul(x = var_3183_cast_fp16, y = rotr_15_cast_fp16)[name = tensor("op_3209_cast_fp16")]; + tensor var_3210_cast_fp16 = mul(x = var_3191_cast_fp16, y = roti_15_cast_fp16)[name = tensor("op_3210_cast_fp16")]; + tensor kor_29_cast_fp16 = sub(x = var_3209_cast_fp16, y = var_3210_cast_fp16)[name = tensor("kor_29_cast_fp16")]; + tensor var_3213_cast_fp16 = mul(x = var_3183_cast_fp16, y = roti_15_cast_fp16)[name = tensor("op_3213_cast_fp16")]; + tensor var_3214_cast_fp16 = mul(x = var_3191_cast_fp16, y = rotr_15_cast_fp16)[name = tensor("op_3214_cast_fp16")]; + tensor koi_29_cast_fp16 = add(x = var_3213_cast_fp16, y = var_3214_cast_fp16)[name = tensor("koi_29_cast_fp16")]; + tensor qo_15_axis_0 = const()[name = tensor("qo_15_axis_0"), val = tensor(-1)]; + tensor qo_15_cast_fp16 = stack(axis = qo_15_axis_0, values = (qor_29_cast_fp16, qoi_29_cast_fp16))[name = tensor("qo_15_cast_fp16")]; + tensor ko_15_axis_0 = const()[name = tensor("ko_15_axis_0"), val = tensor(-1)]; + tensor ko_15_cast_fp16 = stack(axis = ko_15_axis_0, values = (kor_29_cast_fp16, koi_29_cast_fp16))[name = tensor("ko_15_cast_fp16")]; + tensor var_3247 = const()[name = tensor("op_3247"), val = tensor([1, 256, 16, 64])]; + tensor q_45_cast_fp16 = reshape(shape = var_3247, x = qo_15_cast_fp16)[name = tensor("q_45_cast_fp16")]; + tensor var_3253 = const()[name = tensor("op_3253"), val = tensor([1, 256, 16, 64])]; + tensor k_31_cast_fp16 = reshape(shape = var_3253, x = ko_15_cast_fp16)[name = tensor("k_31_cast_fp16")]; + tensor write_indices_float_43_cast_fp16 = add(x = var_3142_cast_fp16, y = var_393_promoted_to_fp16)[name = tensor("write_indices_float_43_cast_fp16")]; + tensor _inversed_3274_y_0_to_fp16 = const()[name = tensor("_inversed_3274_y_0_to_fp16"), val = tensor(0x1p-9)]; + tensor _inversed_3274_cast_fp16 = mul(x = write_indices_float_43_cast_fp16, y = _inversed_3274_y_0_to_fp16)[name = tensor("_inversed_3274_cast_fp16")]; + tensor var_3275_cast_fp16 = floor(x = _inversed_3274_cast_fp16)[name = tensor("op_3275_cast_fp16")]; + tensor var_3276_to_fp16 = const()[name = tensor("op_3276_to_fp16"), val = tensor(0x1p+9)]; + tensor var_3277_cast_fp16 = mul(x = var_3275_cast_fp16, y = var_3276_to_fp16)[name = tensor("op_3277_cast_fp16")]; + tensor write_indices_float_45_cast_fp16 = sub(x = write_indices_float_43_cast_fp16, y = var_3277_cast_fp16)[name = tensor("write_indices_float_45_cast_fp16")]; + tensor var_3295_value_0_to_fp16 = const()[name = tensor("op_3295_value_0_to_fp16"), val = tensor(0x1.ffp+8)]; + tensor var_3295_cast_fp16 = fill_like(ref_tensor = write_indices_float_45_cast_fp16, value = var_3295_value_0_to_fp16)[name = tensor("op_3295_cast_fp16")]; + tensor write_indices_float_47_cast_fp16 = select(a = write_indices_float_45_cast_fp16, b = var_3295_cast_fp16, cond = token_valid_1_cast_fp16)[name = tensor("write_indices_float_47_cast_fp16")]; + tensor var_3301_dtype_0 = const()[name = tensor("op_3301_dtype_0"), val = tensor("int32")]; + tensor var_3306 = const()[name = tensor("op_3306"), val = tensor([1, 256, 1, 1])]; + tensor write_indices_float_47_cast_fp16_to_int32 = cast(dtype = var_3301_dtype_0, x = write_indices_float_47_cast_fp16)[name = tensor("cast_625")]; + tensor var_3307 = reshape(shape = var_3306, x = write_indices_float_47_cast_fp16_to_int32)[name = tensor("op_3307")]; + tensor write_indices_15_reps_0 = const()[name = tensor("write_indices_15_reps_0"), val = tensor([1, 1, 16, 64])]; + tensor write_indices_15 = tile(reps = write_indices_15_reps_0, x = var_3307)[name = tensor("write_indices_15")]; + tensor var_3317_begin_0 = const()[name = tensor("op_3317_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_3317_end_0 = const()[name = tensor("op_3317_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor var_3317_end_mask_0 = const()[name = tensor("op_3317_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_3317_squeeze_mask_0 = const()[name = tensor("op_3317_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor cache7_to_fp16_dtype_0 = const()[name = tensor("cache7_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor cache7_to_fp16 = cast(dtype = cache7_to_fp16_dtype_0, x = cache7)[name = tensor("cast_624")]; + tensor var_3317_cast_fp16 = slice_by_index(begin = var_3317_begin_0, end = var_3317_end_0, end_mask = var_3317_end_mask_0, squeeze_mask = var_3317_squeeze_mask_0, x = cache7_to_fp16)[name = tensor("op_3317_cast_fp16")]; + tensor var_3319_axis_0 = const()[name = tensor("op_3319_axis_0"), val = tensor(1)]; + tensor var_3319_mode_0 = const()[name = tensor("op_3319_mode_0"), val = tensor("update")]; + tensor var_3319_validate_indices_0 = const()[name = tensor("op_3319_validate_indices_0"), val = tensor(false)]; + tensor var_3319_cast_fp16 = scatter_along_axis(axis = var_3319_axis_0, data = var_3317_cast_fp16, indices = write_indices_15, mode = var_3319_mode_0, updates = k_31_cast_fp16, validate_indices = var_3319_validate_indices_0)[name = tensor("op_3319_cast_fp16")]; + tensor concat_58 = const()[name = tensor("concat_58"), val = tensor([0, 0, 0, 0, 0])]; + tensor concat_59 = const()[name = tensor("concat_59"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_15_internal_tensor_assign_1_stride_0 = const()[name = tensor("new_cache_15_internal_tensor_assign_1_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_15_internal_tensor_assign_1_begin_mask_0 = const()[name = tensor("new_cache_15_internal_tensor_assign_1_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_15_internal_tensor_assign_1_end_mask_0 = const()[name = tensor("new_cache_15_internal_tensor_assign_1_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_15_internal_tensor_assign_1_squeeze_mask_0 = const()[name = tensor("new_cache_15_internal_tensor_assign_1_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_60 = const()[name = tensor("shape_60"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_14 = const()[name = tensor("reduce_prod_14"), val = tensor(1048576)]; + tensor range_1d_14_start_0 = const()[name = tensor("range_1d_14_start_0"), val = tensor(0)]; + tensor range_1d_14_step_0 = const()[name = tensor("range_1d_14_step_0"), val = tensor(1)]; + tensor range_1d_14 = range_1d(end = reduce_prod_14, start = range_1d_14_start_0, step = range_1d_14_step_0)[name = tensor("range_1d_14")]; + tensor reshape_70 = reshape(shape = shape_60, x = range_1d_14)[name = tensor("reshape_70")]; + tensor slice_by_index_14 = slice_by_index(begin = concat_58, begin_mask = new_cache_15_internal_tensor_assign_1_begin_mask_0, end = concat_59, end_mask = new_cache_15_internal_tensor_assign_1_end_mask_0, squeeze_mask = new_cache_15_internal_tensor_assign_1_squeeze_mask_0, stride = new_cache_15_internal_tensor_assign_1_stride_0, x = reshape_70)[name = tensor("slice_by_index_14")]; + tensor reshape_71_shape_0 = const()[name = tensor("reshape_71_shape_0"), val = tensor([-1])]; + tensor reshape_71 = reshape(shape = reshape_71_shape_0, x = slice_by_index_14)[name = tensor("reshape_71")]; + tensor reshape_72_shape_0 = const()[name = tensor("reshape_72_shape_0"), val = tensor([-1])]; + tensor reshape_72_cast_fp16 = reshape(shape = reshape_72_shape_0, x = var_3319_cast_fp16)[name = tensor("reshape_72_cast_fp16")]; + tensor reshape_73_shape_0 = const()[name = tensor("reshape_73_shape_0"), val = tensor([-1])]; + tensor reshape_73_cast_fp16 = reshape(shape = reshape_73_shape_0, x = cache7_to_fp16)[name = tensor("reshape_73_cast_fp16")]; + tensor scatter_14_mode_0 = const()[name = tensor("scatter_14_mode_0"), val = tensor("update")]; + tensor scatter_14_axis_0 = const()[name = tensor("scatter_14_axis_0"), val = tensor(0)]; + tensor scatter_14_validate_indices_0 = const()[name = tensor("scatter_14_validate_indices_0"), val = tensor(false)]; + tensor scatter_14_cast_fp16 = scatter(axis = scatter_14_axis_0, data = reshape_73_cast_fp16, indices = reshape_71, mode = scatter_14_mode_0, updates = reshape_72_cast_fp16, validate_indices = scatter_14_validate_indices_0)[name = tensor("scatter_14_cast_fp16")]; + tensor reshape_74_cast_fp16 = reshape(shape = shape_60, x = scatter_14_cast_fp16)[name = tensor("reshape_74_cast_fp16")]; + tensor var_3327_begin_0 = const()[name = tensor("op_3327_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor var_3327_end_0 = const()[name = tensor("op_3327_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor var_3327_end_mask_0 = const()[name = tensor("op_3327_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_3327_squeeze_mask_0 = const()[name = tensor("op_3327_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor var_3327_cast_fp16 = slice_by_index(begin = var_3327_begin_0, end = var_3327_end_0, end_mask = var_3327_end_mask_0, squeeze_mask = var_3327_squeeze_mask_0, x = reshape_74_cast_fp16)[name = tensor("op_3327_cast_fp16")]; + tensor var_3329_axis_0 = const()[name = tensor("op_3329_axis_0"), val = tensor(1)]; + tensor var_3329_mode_0 = const()[name = tensor("op_3329_mode_0"), val = tensor("update")]; + tensor var_3329_validate_indices_0 = const()[name = tensor("op_3329_validate_indices_0"), val = tensor(false)]; + tensor var_3329_cast_fp16 = scatter_along_axis(axis = var_3329_axis_0, data = var_3327_cast_fp16, indices = write_indices_15, mode = var_3329_mode_0, updates = v_15_cast_fp16, validate_indices = var_3329_validate_indices_0)[name = tensor("op_3329_cast_fp16")]; + tensor concat_60 = const()[name = tensor("concat_60"), val = tensor([1, 0, 0, 0, 0])]; + tensor concat_61 = const()[name = tensor("concat_61"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_15_internal_tensor_assign_2_stride_0 = const()[name = tensor("new_cache_15_internal_tensor_assign_2_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_15_internal_tensor_assign_2_begin_mask_0 = const()[name = tensor("new_cache_15_internal_tensor_assign_2_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_15_internal_tensor_assign_2_end_mask_0 = const()[name = tensor("new_cache_15_internal_tensor_assign_2_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_15_internal_tensor_assign_2_squeeze_mask_0 = const()[name = tensor("new_cache_15_internal_tensor_assign_2_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_61 = const()[name = tensor("shape_61"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_15 = const()[name = tensor("reduce_prod_15"), val = tensor(1048576)]; + tensor range_1d_15_start_0 = const()[name = tensor("range_1d_15_start_0"), val = tensor(0)]; + tensor range_1d_15_step_0 = const()[name = tensor("range_1d_15_step_0"), val = tensor(1)]; + tensor range_1d_15 = range_1d(end = reduce_prod_15, start = range_1d_15_start_0, step = range_1d_15_step_0)[name = tensor("range_1d_15")]; + tensor reshape_75 = reshape(shape = shape_61, x = range_1d_15)[name = tensor("reshape_75")]; + tensor slice_by_index_15 = slice_by_index(begin = concat_60, begin_mask = new_cache_15_internal_tensor_assign_2_begin_mask_0, end = concat_61, end_mask = new_cache_15_internal_tensor_assign_2_end_mask_0, squeeze_mask = new_cache_15_internal_tensor_assign_2_squeeze_mask_0, stride = new_cache_15_internal_tensor_assign_2_stride_0, x = reshape_75)[name = tensor("slice_by_index_15")]; + tensor reshape_76_shape_0 = const()[name = tensor("reshape_76_shape_0"), val = tensor([-1])]; + tensor reshape_76 = reshape(shape = reshape_76_shape_0, x = slice_by_index_15)[name = tensor("reshape_76")]; + tensor reshape_77_shape_0 = const()[name = tensor("reshape_77_shape_0"), val = tensor([-1])]; + tensor reshape_77_cast_fp16 = reshape(shape = reshape_77_shape_0, x = var_3329_cast_fp16)[name = tensor("reshape_77_cast_fp16")]; + tensor reshape_78_shape_0 = const()[name = tensor("reshape_78_shape_0"), val = tensor([-1])]; + tensor reshape_78_cast_fp16 = reshape(shape = reshape_78_shape_0, x = reshape_74_cast_fp16)[name = tensor("reshape_78_cast_fp16")]; + tensor scatter_15_mode_0 = const()[name = tensor("scatter_15_mode_0"), val = tensor("update")]; + tensor scatter_15_axis_0 = const()[name = tensor("scatter_15_axis_0"), val = tensor(0)]; + tensor scatter_15_validate_indices_0 = const()[name = tensor("scatter_15_validate_indices_0"), val = tensor(false)]; + tensor scatter_15_cast_fp16 = scatter(axis = scatter_15_axis_0, data = reshape_78_cast_fp16, indices = reshape_76, mode = scatter_15_mode_0, updates = reshape_77_cast_fp16, validate_indices = scatter_15_validate_indices_0)[name = tensor("scatter_15_cast_fp16")]; + tensor reshape_79_cast_fp16 = reshape(shape = shape_61, x = scatter_15_cast_fp16)[name = tensor("reshape_79_cast_fp16")]; + tensor reshape_79_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("reshape_79_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor keys_43_begin_0 = const()[name = tensor("keys_43_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor keys_43_end_0 = const()[name = tensor("keys_43_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor keys_43_end_mask_0 = const()[name = tensor("keys_43_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor keys_43_squeeze_mask_0 = const()[name = tensor("keys_43_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor keys_43_cast_fp16 = slice_by_index(begin = keys_43_begin_0, end = keys_43_end_0, end_mask = keys_43_end_mask_0, squeeze_mask = keys_43_squeeze_mask_0, x = reshape_79_cast_fp16)[name = tensor("keys_43_cast_fp16")]; + tensor values_43_begin_0 = const()[name = tensor("values_43_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor values_43_end_0 = const()[name = tensor("values_43_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor values_43_end_mask_0 = const()[name = tensor("values_43_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor values_43_squeeze_mask_0 = const()[name = tensor("values_43_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor values_43_cast_fp16 = slice_by_index(begin = values_43_begin_0, end = values_43_end_0, end_mask = values_43_end_mask_0, squeeze_mask = values_43_squeeze_mask_0, x = reshape_79_cast_fp16)[name = tensor("values_43_cast_fp16")]; + tensor var_3341_cast_fp16 = not_equal(x = keys_43_cast_fp16, y = keys_43_cast_fp16)[name = tensor("op_3341_cast_fp16")]; + tensor keys_45_cast_fp16 = select(a = var_470_to_fp16, b = keys_43_cast_fp16, cond = var_3341_cast_fp16)[name = tensor("keys_45_cast_fp16")]; + tensor var_3349_cast_fp16 = not_equal(x = values_43_cast_fp16, y = values_43_cast_fp16)[name = tensor("op_3349_cast_fp16")]; + tensor values_45_cast_fp16 = select(a = var_470_to_fp16, b = values_43_cast_fp16, cond = var_3349_cast_fp16)[name = tensor("values_45_cast_fp16")]; + tensor var_3373 = const()[name = tensor("op_3373"), val = tensor([0, 2, 1, 3])]; + tensor var_3389 = const()[name = tensor("op_3389"), val = tensor([1, 1, 1])]; + tensor var_3390_cast_fp16 = reshape(shape = var_3389, x = position7_to_fp16)[name = tensor("op_3390_cast_fp16")]; + tensor q_positions_15_cast_fp16 = add(x = var_3390_cast_fp16, y = q_offsets_1_promoted_to_fp16)[name = tensor("q_positions_15_cast_fp16")]; + tensor valid_end_15_cast_fp16 = add(x = var_3390_cast_fp16, y = var_536_cast_fp16)[name = tensor("valid_end_15_cast_fp16")]; + tensor valid_mask_15_cast_fp16 = less(x = k_positions_1_promoted_to_fp16, y = valid_end_15_cast_fp16)[name = tensor("valid_mask_15_cast_fp16")]; + tensor causal_mask_15_cast_fp16 = less_equal(x = k_positions_1_promoted_to_fp16, y = q_positions_15_cast_fp16)[name = tensor("causal_mask_15_cast_fp16")]; + tensor var_3418 = logical_and(x = valid_mask_15_cast_fp16, y = causal_mask_15_cast_fp16)[name = tensor("op_3418")]; + tensor attn_mask_15_axes_0 = const()[name = tensor("attn_mask_15_axes_0"), val = tensor([1])]; + tensor attn_mask_15 = expand_dims(axes = attn_mask_15_axes_0, x = var_3418)[name = tensor("attn_mask_15")]; + tensor var_3432_transpose_x_0 = const()[name = tensor("op_3432_transpose_x_0"), val = tensor(false)]; + tensor var_3432_transpose_y_0 = const()[name = tensor("op_3432_transpose_y_0"), val = tensor(false)]; + tensor transpose_83_perm_0 = const()[name = tensor("transpose_83_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_84_perm_0 = const()[name = tensor("transpose_84_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_84 = transpose(perm = transpose_84_perm_0, x = keys_45_cast_fp16)[name = tensor("transpose_176")]; + tensor transpose_83 = transpose(perm = transpose_83_perm_0, x = q_45_cast_fp16)[name = tensor("transpose_177")]; + tensor var_3432_cast_fp16 = matmul(transpose_x = var_3432_transpose_x_0, transpose_y = var_3432_transpose_y_0, x = transpose_83, y = transpose_84)[name = tensor("op_3432_cast_fp16")]; + tensor var_3426_to_fp16 = const()[name = tensor("op_3426_to_fp16"), val = tensor([0x1p-3])]; + tensor attn_weights_43_cast_fp16 = mul(x = var_3432_cast_fp16, y = var_3426_to_fp16)[name = tensor("attn_weights_43_cast_fp16")]; + tensor var_3434 = logical_not(x = attn_mask_15)[name = tensor("op_3434")]; + tensor var_3435_to_fp16 = const()[name = tensor("op_3435_to_fp16"), val = tensor(-inf)]; + tensor attn_weights_45_cast_fp16 = select(a = var_3435_to_fp16, b = attn_weights_43_cast_fp16, cond = var_3434)[name = tensor("attn_weights_45_cast_fp16")]; + tensor var_3437 = const()[name = tensor("op_3437"), val = tensor(-1)]; + tensor attn_weights_47_cast_fp16 = softmax(axis = var_3437, x = attn_weights_45_cast_fp16)[name = tensor("attn_weights_47_cast_fp16")]; + tensor attn_output_15_transpose_x_0 = const()[name = tensor("attn_output_15_transpose_x_0"), val = tensor(false)]; + tensor attn_output_15_transpose_y_0 = const()[name = tensor("attn_output_15_transpose_y_0"), val = tensor(false)]; + tensor values_47_cast_fp16 = transpose(perm = var_3373, x = values_45_cast_fp16)[name = tensor("transpose_178")]; + tensor attn_output_15_cast_fp16 = matmul(transpose_x = attn_output_15_transpose_x_0, transpose_y = attn_output_15_transpose_y_0, x = attn_weights_47_cast_fp16, y = values_47_cast_fp16)[name = tensor("attn_output_15_cast_fp16")]; + tensor var_3445 = const()[name = tensor("op_3445"), val = tensor([0, 2, 1, 3])]; + tensor var_3450 = const()[name = tensor("op_3450"), val = tensor([1, 256, 1024])]; + tensor var_3446_cast_fp16 = transpose(perm = var_3445, x = attn_output_15_cast_fp16)[name = tensor("transpose_175")]; + tensor input_89_cast_fp16 = reshape(shape = var_3450, x = var_3446_cast_fp16)[name = tensor("input_89_cast_fp16")]; + tensor attn7_out_proj_weight_to_fp16 = const()[name = tensor("attn7_out_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(183586560)))]; + tensor linear_29_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = attn7_out_proj_weight_to_fp16, x = input_89_cast_fp16)[name = tensor("linear_29_cast_fp16")]; + tensor var_3459_cast_fp16 = add(x = position7_to_fp16, y = valid_len_to_fp16)[name = tensor("op_3459_cast_fp16")]; + tensor var_3459_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_3459_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor input_91_cast_fp16 = add(x = input_85_cast_fp16, y = linear_29_cast_fp16)[name = tensor("input_91_cast_fp16")]; + tensor input_93_axes_0 = const()[name = tensor("input_93_axes_0"), val = tensor([-1])]; + tensor norm7_2_weight_to_fp16 = const()[name = tensor("norm7_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(185683776)))]; + tensor norm7_2_bias_to_fp16 = const()[name = tensor("norm7_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(185685888)))]; + tensor var_3463_to_fp16 = const()[name = tensor("op_3463_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_93_cast_fp16 = layer_norm(axes = input_93_axes_0, beta = norm7_2_bias_to_fp16, epsilon = var_3463_to_fp16, gamma = norm7_2_weight_to_fp16, x = input_91_cast_fp16)[name = tensor("input_93_cast_fp16")]; + tensor linear7_1_weight_to_fp16 = const()[name = tensor("linear7_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(185688000)))]; + tensor linear_30_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = linear7_1_weight_to_fp16, x = input_93_cast_fp16)[name = tensor("linear_30_cast_fp16")]; + tensor input_95_mode_0 = const()[name = tensor("input_95_mode_0"), val = tensor("EXACT")]; + tensor input_95_cast_fp16 = gelu(mode = input_95_mode_0, x = linear_30_cast_fp16)[name = tensor("input_95_cast_fp16")]; + tensor linear7_2_weight_to_fp16 = const()[name = tensor("linear7_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(194076672)))]; + tensor linear_31_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = linear7_2_weight_to_fp16, x = input_95_cast_fp16)[name = tensor("linear_31_cast_fp16")]; + tensor input_97_cast_fp16 = add(x = input_91_cast_fp16, y = linear_31_cast_fp16)[name = tensor("input_97_cast_fp16")]; + tensor input_99_axes_0 = const()[name = tensor("input_99_axes_0"), val = tensor([-1])]; + tensor norm8_1_weight_to_fp16 = const()[name = tensor("norm8_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(202465344)))]; + tensor norm8_1_bias_to_fp16 = const()[name = tensor("norm8_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(202467456)))]; + tensor var_3480_to_fp16 = const()[name = tensor("op_3480_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_99_cast_fp16 = layer_norm(axes = input_99_axes_0, beta = norm8_1_bias_to_fp16, epsilon = var_3480_to_fp16, gamma = norm8_1_weight_to_fp16, x = input_97_cast_fp16)[name = tensor("input_99_cast_fp16")]; + tensor attn8_in_proj_weight_to_fp16 = const()[name = tensor("attn8_in_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(202469568)))]; + tensor linear_32_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = attn8_in_proj_weight_to_fp16, x = input_99_cast_fp16)[name = tensor("linear_32_cast_fp16")]; + tensor var_3494 = const()[name = tensor("op_3494"), val = tensor([1, 256, 3, 16, 64])]; + tensor qkv_17_cast_fp16 = reshape(shape = var_3494, x = linear_32_cast_fp16)[name = tensor("qkv_17_cast_fp16")]; + tensor q_49_begin_0 = const()[name = tensor("q_49_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor q_49_end_0 = const()[name = tensor("q_49_end_0"), val = tensor([1, 256, 1, 16, 64])]; + tensor q_49_end_mask_0 = const()[name = tensor("q_49_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor q_49_squeeze_mask_0 = const()[name = tensor("q_49_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor q_49_cast_fp16 = slice_by_index(begin = q_49_begin_0, end = q_49_end_0, end_mask = q_49_end_mask_0, squeeze_mask = q_49_squeeze_mask_0, x = qkv_17_cast_fp16)[name = tensor("q_49_cast_fp16")]; + tensor k_33_begin_0 = const()[name = tensor("k_33_begin_0"), val = tensor([0, 0, 1, 0, 0])]; + tensor k_33_end_0 = const()[name = tensor("k_33_end_0"), val = tensor([1, 256, 2, 16, 64])]; + tensor k_33_end_mask_0 = const()[name = tensor("k_33_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor k_33_squeeze_mask_0 = const()[name = tensor("k_33_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor k_33_cast_fp16 = slice_by_index(begin = k_33_begin_0, end = k_33_end_0, end_mask = k_33_end_mask_0, squeeze_mask = k_33_squeeze_mask_0, x = qkv_17_cast_fp16)[name = tensor("k_33_cast_fp16")]; + tensor v_17_begin_0 = const()[name = tensor("v_17_begin_0"), val = tensor([0, 0, 2, 0, 0])]; + tensor v_17_end_0 = const()[name = tensor("v_17_end_0"), val = tensor([1, 256, 3, 16, 64])]; + tensor v_17_end_mask_0 = const()[name = tensor("v_17_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor v_17_squeeze_mask_0 = const()[name = tensor("v_17_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor v_17_cast_fp16 = slice_by_index(begin = v_17_begin_0, end = v_17_end_0, end_mask = v_17_end_mask_0, squeeze_mask = v_17_squeeze_mask_0, x = qkv_17_cast_fp16)[name = tensor("v_17_cast_fp16")]; + tensor var_3552 = const()[name = tensor("op_3552"), val = tensor([1, 1])]; + tensor position8_to_fp16_dtype_0 = const()[name = tensor("position8_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor position8_to_fp16 = cast(dtype = position8_to_fp16_dtype_0, x = position8)[name = tensor("cast_621")]; + tensor var_3553_cast_fp16 = reshape(shape = var_3552, x = position8_to_fp16)[name = tensor("op_3553_cast_fp16")]; + tensor ts_51_cast_fp16 = add(x = ts_1_promoted_to_fp16, y = var_3553_cast_fp16)[name = tensor("ts_51_cast_fp16")]; + tensor var_3560 = const()[name = tensor("op_3560"), val = tensor([1, 256, 1, 1])]; + tensor ts_53_cast_fp16 = reshape(shape = var_3560, x = ts_51_cast_fp16)[name = tensor("ts_53_cast_fp16")]; + tensor var_3567 = const()[name = tensor("op_3567"), val = tensor([1, 256, 16, 32, 2])]; + tensor q_complex_17_cast_fp16 = reshape(shape = var_3567, x = q_49_cast_fp16)[name = tensor("q_complex_17_cast_fp16")]; + tensor var_3574 = const()[name = tensor("op_3574"), val = tensor([1, 256, 16, 32, 2])]; + tensor k_complex_17_cast_fp16 = reshape(shape = var_3574, x = k_33_cast_fp16)[name = tensor("k_complex_17_cast_fp16")]; + tensor var_3578_begin_0 = const()[name = tensor("op_3578_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_3578_end_0 = const()[name = tensor("op_3578_end_0"), val = tensor([1, 256, 16, 32, 1])]; + tensor var_3578_end_mask_0 = const()[name = tensor("op_3578_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_3578_squeeze_mask_0 = const()[name = tensor("op_3578_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_3578_cast_fp16 = slice_by_index(begin = var_3578_begin_0, end = var_3578_end_0, end_mask = var_3578_end_mask_0, squeeze_mask = var_3578_squeeze_mask_0, x = q_complex_17_cast_fp16)[name = tensor("op_3578_cast_fp16")]; + tensor var_3586_begin_0 = const()[name = tensor("op_3586_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_3586_end_0 = const()[name = tensor("op_3586_end_0"), val = tensor([1, 256, 16, 32, 2])]; + tensor var_3586_end_mask_0 = const()[name = tensor("op_3586_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_3586_squeeze_mask_0 = const()[name = tensor("op_3586_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_3586_cast_fp16 = slice_by_index(begin = var_3586_begin_0, end = var_3586_end_0, end_mask = var_3586_end_mask_0, squeeze_mask = var_3586_squeeze_mask_0, x = q_complex_17_cast_fp16)[name = tensor("op_3586_cast_fp16")]; + tensor var_3594_begin_0 = const()[name = tensor("op_3594_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_3594_end_0 = const()[name = tensor("op_3594_end_0"), val = tensor([1, 256, 16, 32, 1])]; + tensor var_3594_end_mask_0 = const()[name = tensor("op_3594_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_3594_squeeze_mask_0 = const()[name = tensor("op_3594_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_3594_cast_fp16 = slice_by_index(begin = var_3594_begin_0, end = var_3594_end_0, end_mask = var_3594_end_mask_0, squeeze_mask = var_3594_squeeze_mask_0, x = k_complex_17_cast_fp16)[name = tensor("op_3594_cast_fp16")]; + tensor var_3602_begin_0 = const()[name = tensor("op_3602_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_3602_end_0 = const()[name = tensor("op_3602_end_0"), val = tensor([1, 256, 16, 32, 2])]; + tensor var_3602_end_mask_0 = const()[name = tensor("op_3602_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_3602_squeeze_mask_0 = const()[name = tensor("op_3602_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_3602_cast_fp16 = slice_by_index(begin = var_3602_begin_0, end = var_3602_end_0, end_mask = var_3602_end_mask_0, squeeze_mask = var_3602_squeeze_mask_0, x = k_complex_17_cast_fp16)[name = tensor("op_3602_cast_fp16")]; + tensor freqs_17_to_fp16 = const()[name = tensor("freqs_17_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(208761088)))]; + tensor var_3608_cast_fp16 = mul(x = freqs_17_to_fp16, y = ts_53_cast_fp16)[name = tensor("op_3608_cast_fp16")]; + tensor rotr_17_cast_fp16 = cos(x = var_3608_cast_fp16)[name = tensor("rotr_17_cast_fp16")]; + tensor roti_17_cast_fp16 = sin(x = var_3608_cast_fp16)[name = tensor("roti_17_cast_fp16")]; + tensor var_3612_cast_fp16 = mul(x = var_3578_cast_fp16, y = rotr_17_cast_fp16)[name = tensor("op_3612_cast_fp16")]; + tensor var_3613_cast_fp16 = mul(x = var_3586_cast_fp16, y = roti_17_cast_fp16)[name = tensor("op_3613_cast_fp16")]; + tensor qor_33_cast_fp16 = sub(x = var_3612_cast_fp16, y = var_3613_cast_fp16)[name = tensor("qor_33_cast_fp16")]; + tensor var_3616_cast_fp16 = mul(x = var_3578_cast_fp16, y = roti_17_cast_fp16)[name = tensor("op_3616_cast_fp16")]; + tensor var_3617_cast_fp16 = mul(x = var_3586_cast_fp16, y = rotr_17_cast_fp16)[name = tensor("op_3617_cast_fp16")]; + tensor qoi_33_cast_fp16 = add(x = var_3616_cast_fp16, y = var_3617_cast_fp16)[name = tensor("qoi_33_cast_fp16")]; + tensor var_3620_cast_fp16 = mul(x = var_3594_cast_fp16, y = rotr_17_cast_fp16)[name = tensor("op_3620_cast_fp16")]; + tensor var_3621_cast_fp16 = mul(x = var_3602_cast_fp16, y = roti_17_cast_fp16)[name = tensor("op_3621_cast_fp16")]; + tensor kor_33_cast_fp16 = sub(x = var_3620_cast_fp16, y = var_3621_cast_fp16)[name = tensor("kor_33_cast_fp16")]; + tensor var_3624_cast_fp16 = mul(x = var_3594_cast_fp16, y = roti_17_cast_fp16)[name = tensor("op_3624_cast_fp16")]; + tensor var_3625_cast_fp16 = mul(x = var_3602_cast_fp16, y = rotr_17_cast_fp16)[name = tensor("op_3625_cast_fp16")]; + tensor koi_33_cast_fp16 = add(x = var_3624_cast_fp16, y = var_3625_cast_fp16)[name = tensor("koi_33_cast_fp16")]; + tensor qo_17_axis_0 = const()[name = tensor("qo_17_axis_0"), val = tensor(-1)]; + tensor qo_17_cast_fp16 = stack(axis = qo_17_axis_0, values = (qor_33_cast_fp16, qoi_33_cast_fp16))[name = tensor("qo_17_cast_fp16")]; + tensor ko_17_axis_0 = const()[name = tensor("ko_17_axis_0"), val = tensor(-1)]; + tensor ko_17_cast_fp16 = stack(axis = ko_17_axis_0, values = (kor_33_cast_fp16, koi_33_cast_fp16))[name = tensor("ko_17_cast_fp16")]; + tensor var_3658 = const()[name = tensor("op_3658"), val = tensor([1, 256, 16, 64])]; + tensor q_51_cast_fp16 = reshape(shape = var_3658, x = qo_17_cast_fp16)[name = tensor("q_51_cast_fp16")]; + tensor var_3664 = const()[name = tensor("op_3664"), val = tensor([1, 256, 16, 64])]; + tensor k_35_cast_fp16 = reshape(shape = var_3664, x = ko_17_cast_fp16)[name = tensor("k_35_cast_fp16")]; + tensor write_indices_float_49_cast_fp16 = add(x = var_3553_cast_fp16, y = var_393_promoted_to_fp16)[name = tensor("write_indices_float_49_cast_fp16")]; + tensor _inversed_3685_y_0_to_fp16 = const()[name = tensor("_inversed_3685_y_0_to_fp16"), val = tensor(0x1p-9)]; + tensor _inversed_3685_cast_fp16 = mul(x = write_indices_float_49_cast_fp16, y = _inversed_3685_y_0_to_fp16)[name = tensor("_inversed_3685_cast_fp16")]; + tensor var_3686_cast_fp16 = floor(x = _inversed_3685_cast_fp16)[name = tensor("op_3686_cast_fp16")]; + tensor var_3687_to_fp16 = const()[name = tensor("op_3687_to_fp16"), val = tensor(0x1p+9)]; + tensor var_3688_cast_fp16 = mul(x = var_3686_cast_fp16, y = var_3687_to_fp16)[name = tensor("op_3688_cast_fp16")]; + tensor write_indices_float_51_cast_fp16 = sub(x = write_indices_float_49_cast_fp16, y = var_3688_cast_fp16)[name = tensor("write_indices_float_51_cast_fp16")]; + tensor var_3706_value_0_to_fp16 = const()[name = tensor("op_3706_value_0_to_fp16"), val = tensor(0x1.ffp+8)]; + tensor var_3706_cast_fp16 = fill_like(ref_tensor = write_indices_float_51_cast_fp16, value = var_3706_value_0_to_fp16)[name = tensor("op_3706_cast_fp16")]; + tensor write_indices_float_53_cast_fp16 = select(a = write_indices_float_51_cast_fp16, b = var_3706_cast_fp16, cond = token_valid_1_cast_fp16)[name = tensor("write_indices_float_53_cast_fp16")]; + tensor var_3712_dtype_0 = const()[name = tensor("op_3712_dtype_0"), val = tensor("int32")]; + tensor var_3717 = const()[name = tensor("op_3717"), val = tensor([1, 256, 1, 1])]; + tensor write_indices_float_53_cast_fp16_to_int32 = cast(dtype = var_3712_dtype_0, x = write_indices_float_53_cast_fp16)[name = tensor("cast_620")]; + tensor var_3718 = reshape(shape = var_3717, x = write_indices_float_53_cast_fp16_to_int32)[name = tensor("op_3718")]; + tensor write_indices_17_reps_0 = const()[name = tensor("write_indices_17_reps_0"), val = tensor([1, 1, 16, 64])]; + tensor write_indices_17 = tile(reps = write_indices_17_reps_0, x = var_3718)[name = tensor("write_indices_17")]; + tensor var_3728_begin_0 = const()[name = tensor("op_3728_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_3728_end_0 = const()[name = tensor("op_3728_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor var_3728_end_mask_0 = const()[name = tensor("op_3728_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_3728_squeeze_mask_0 = const()[name = tensor("op_3728_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor cache8_to_fp16_dtype_0 = const()[name = tensor("cache8_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor cache8_to_fp16 = cast(dtype = cache8_to_fp16_dtype_0, x = cache8)[name = tensor("cast_619")]; + tensor var_3728_cast_fp16 = slice_by_index(begin = var_3728_begin_0, end = var_3728_end_0, end_mask = var_3728_end_mask_0, squeeze_mask = var_3728_squeeze_mask_0, x = cache8_to_fp16)[name = tensor("op_3728_cast_fp16")]; + tensor var_3730_axis_0 = const()[name = tensor("op_3730_axis_0"), val = tensor(1)]; + tensor var_3730_mode_0 = const()[name = tensor("op_3730_mode_0"), val = tensor("update")]; + tensor var_3730_validate_indices_0 = const()[name = tensor("op_3730_validate_indices_0"), val = tensor(false)]; + tensor var_3730_cast_fp16 = scatter_along_axis(axis = var_3730_axis_0, data = var_3728_cast_fp16, indices = write_indices_17, mode = var_3730_mode_0, updates = k_35_cast_fp16, validate_indices = var_3730_validate_indices_0)[name = tensor("op_3730_cast_fp16")]; + tensor concat_66 = const()[name = tensor("concat_66"), val = tensor([0, 0, 0, 0, 0])]; + tensor concat_67 = const()[name = tensor("concat_67"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_17_internal_tensor_assign_1_stride_0 = const()[name = tensor("new_cache_17_internal_tensor_assign_1_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_17_internal_tensor_assign_1_begin_mask_0 = const()[name = tensor("new_cache_17_internal_tensor_assign_1_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_17_internal_tensor_assign_1_end_mask_0 = const()[name = tensor("new_cache_17_internal_tensor_assign_1_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_17_internal_tensor_assign_1_squeeze_mask_0 = const()[name = tensor("new_cache_17_internal_tensor_assign_1_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_62 = const()[name = tensor("shape_62"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_16 = const()[name = tensor("reduce_prod_16"), val = tensor(1048576)]; + tensor range_1d_16_start_0 = const()[name = tensor("range_1d_16_start_0"), val = tensor(0)]; + tensor range_1d_16_step_0 = const()[name = tensor("range_1d_16_step_0"), val = tensor(1)]; + tensor range_1d_16 = range_1d(end = reduce_prod_16, start = range_1d_16_start_0, step = range_1d_16_step_0)[name = tensor("range_1d_16")]; + tensor reshape_80 = reshape(shape = shape_62, x = range_1d_16)[name = tensor("reshape_80")]; + tensor slice_by_index_16 = slice_by_index(begin = concat_66, begin_mask = new_cache_17_internal_tensor_assign_1_begin_mask_0, end = concat_67, end_mask = new_cache_17_internal_tensor_assign_1_end_mask_0, squeeze_mask = new_cache_17_internal_tensor_assign_1_squeeze_mask_0, stride = new_cache_17_internal_tensor_assign_1_stride_0, x = reshape_80)[name = tensor("slice_by_index_16")]; + tensor reshape_81_shape_0 = const()[name = tensor("reshape_81_shape_0"), val = tensor([-1])]; + tensor reshape_81 = reshape(shape = reshape_81_shape_0, x = slice_by_index_16)[name = tensor("reshape_81")]; + tensor reshape_82_shape_0 = const()[name = tensor("reshape_82_shape_0"), val = tensor([-1])]; + tensor reshape_82_cast_fp16 = reshape(shape = reshape_82_shape_0, x = var_3730_cast_fp16)[name = tensor("reshape_82_cast_fp16")]; + tensor reshape_83_shape_0 = const()[name = tensor("reshape_83_shape_0"), val = tensor([-1])]; + tensor reshape_83_cast_fp16 = reshape(shape = reshape_83_shape_0, x = cache8_to_fp16)[name = tensor("reshape_83_cast_fp16")]; + tensor scatter_16_mode_0 = const()[name = tensor("scatter_16_mode_0"), val = tensor("update")]; + tensor scatter_16_axis_0 = const()[name = tensor("scatter_16_axis_0"), val = tensor(0)]; + tensor scatter_16_validate_indices_0 = const()[name = tensor("scatter_16_validate_indices_0"), val = tensor(false)]; + tensor scatter_16_cast_fp16 = scatter(axis = scatter_16_axis_0, data = reshape_83_cast_fp16, indices = reshape_81, mode = scatter_16_mode_0, updates = reshape_82_cast_fp16, validate_indices = scatter_16_validate_indices_0)[name = tensor("scatter_16_cast_fp16")]; + tensor reshape_84_cast_fp16 = reshape(shape = shape_62, x = scatter_16_cast_fp16)[name = tensor("reshape_84_cast_fp16")]; + tensor var_3738_begin_0 = const()[name = tensor("op_3738_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor var_3738_end_0 = const()[name = tensor("op_3738_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor var_3738_end_mask_0 = const()[name = tensor("op_3738_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_3738_squeeze_mask_0 = const()[name = tensor("op_3738_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor var_3738_cast_fp16 = slice_by_index(begin = var_3738_begin_0, end = var_3738_end_0, end_mask = var_3738_end_mask_0, squeeze_mask = var_3738_squeeze_mask_0, x = reshape_84_cast_fp16)[name = tensor("op_3738_cast_fp16")]; + tensor var_3740_axis_0 = const()[name = tensor("op_3740_axis_0"), val = tensor(1)]; + tensor var_3740_mode_0 = const()[name = tensor("op_3740_mode_0"), val = tensor("update")]; + tensor var_3740_validate_indices_0 = const()[name = tensor("op_3740_validate_indices_0"), val = tensor(false)]; + tensor var_3740_cast_fp16 = scatter_along_axis(axis = var_3740_axis_0, data = var_3738_cast_fp16, indices = write_indices_17, mode = var_3740_mode_0, updates = v_17_cast_fp16, validate_indices = var_3740_validate_indices_0)[name = tensor("op_3740_cast_fp16")]; + tensor concat_68 = const()[name = tensor("concat_68"), val = tensor([1, 0, 0, 0, 0])]; + tensor concat_69 = const()[name = tensor("concat_69"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_17_internal_tensor_assign_2_stride_0 = const()[name = tensor("new_cache_17_internal_tensor_assign_2_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_17_internal_tensor_assign_2_begin_mask_0 = const()[name = tensor("new_cache_17_internal_tensor_assign_2_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_17_internal_tensor_assign_2_end_mask_0 = const()[name = tensor("new_cache_17_internal_tensor_assign_2_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_17_internal_tensor_assign_2_squeeze_mask_0 = const()[name = tensor("new_cache_17_internal_tensor_assign_2_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_63 = const()[name = tensor("shape_63"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_17 = const()[name = tensor("reduce_prod_17"), val = tensor(1048576)]; + tensor range_1d_17_start_0 = const()[name = tensor("range_1d_17_start_0"), val = tensor(0)]; + tensor range_1d_17_step_0 = const()[name = tensor("range_1d_17_step_0"), val = tensor(1)]; + tensor range_1d_17 = range_1d(end = reduce_prod_17, start = range_1d_17_start_0, step = range_1d_17_step_0)[name = tensor("range_1d_17")]; + tensor reshape_85 = reshape(shape = shape_63, x = range_1d_17)[name = tensor("reshape_85")]; + tensor slice_by_index_17 = slice_by_index(begin = concat_68, begin_mask = new_cache_17_internal_tensor_assign_2_begin_mask_0, end = concat_69, end_mask = new_cache_17_internal_tensor_assign_2_end_mask_0, squeeze_mask = new_cache_17_internal_tensor_assign_2_squeeze_mask_0, stride = new_cache_17_internal_tensor_assign_2_stride_0, x = reshape_85)[name = tensor("slice_by_index_17")]; + tensor reshape_86_shape_0 = const()[name = tensor("reshape_86_shape_0"), val = tensor([-1])]; + tensor reshape_86 = reshape(shape = reshape_86_shape_0, x = slice_by_index_17)[name = tensor("reshape_86")]; + tensor reshape_87_shape_0 = const()[name = tensor("reshape_87_shape_0"), val = tensor([-1])]; + tensor reshape_87_cast_fp16 = reshape(shape = reshape_87_shape_0, x = var_3740_cast_fp16)[name = tensor("reshape_87_cast_fp16")]; + tensor reshape_88_shape_0 = const()[name = tensor("reshape_88_shape_0"), val = tensor([-1])]; + tensor reshape_88_cast_fp16 = reshape(shape = reshape_88_shape_0, x = reshape_84_cast_fp16)[name = tensor("reshape_88_cast_fp16")]; + tensor scatter_17_mode_0 = const()[name = tensor("scatter_17_mode_0"), val = tensor("update")]; + tensor scatter_17_axis_0 = const()[name = tensor("scatter_17_axis_0"), val = tensor(0)]; + tensor scatter_17_validate_indices_0 = const()[name = tensor("scatter_17_validate_indices_0"), val = tensor(false)]; + tensor scatter_17_cast_fp16 = scatter(axis = scatter_17_axis_0, data = reshape_88_cast_fp16, indices = reshape_86, mode = scatter_17_mode_0, updates = reshape_87_cast_fp16, validate_indices = scatter_17_validate_indices_0)[name = tensor("scatter_17_cast_fp16")]; + tensor reshape_89_cast_fp16 = reshape(shape = shape_63, x = scatter_17_cast_fp16)[name = tensor("reshape_89_cast_fp16")]; + tensor reshape_89_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("reshape_89_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor keys_49_begin_0 = const()[name = tensor("keys_49_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor keys_49_end_0 = const()[name = tensor("keys_49_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor keys_49_end_mask_0 = const()[name = tensor("keys_49_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor keys_49_squeeze_mask_0 = const()[name = tensor("keys_49_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor keys_49_cast_fp16 = slice_by_index(begin = keys_49_begin_0, end = keys_49_end_0, end_mask = keys_49_end_mask_0, squeeze_mask = keys_49_squeeze_mask_0, x = reshape_89_cast_fp16)[name = tensor("keys_49_cast_fp16")]; + tensor values_49_begin_0 = const()[name = tensor("values_49_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor values_49_end_0 = const()[name = tensor("values_49_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor values_49_end_mask_0 = const()[name = tensor("values_49_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor values_49_squeeze_mask_0 = const()[name = tensor("values_49_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor values_49_cast_fp16 = slice_by_index(begin = values_49_begin_0, end = values_49_end_0, end_mask = values_49_end_mask_0, squeeze_mask = values_49_squeeze_mask_0, x = reshape_89_cast_fp16)[name = tensor("values_49_cast_fp16")]; + tensor var_3752_cast_fp16 = not_equal(x = keys_49_cast_fp16, y = keys_49_cast_fp16)[name = tensor("op_3752_cast_fp16")]; + tensor keys_51_cast_fp16 = select(a = var_470_to_fp16, b = keys_49_cast_fp16, cond = var_3752_cast_fp16)[name = tensor("keys_51_cast_fp16")]; + tensor var_3760_cast_fp16 = not_equal(x = values_49_cast_fp16, y = values_49_cast_fp16)[name = tensor("op_3760_cast_fp16")]; + tensor values_51_cast_fp16 = select(a = var_470_to_fp16, b = values_49_cast_fp16, cond = var_3760_cast_fp16)[name = tensor("values_51_cast_fp16")]; + tensor var_3784 = const()[name = tensor("op_3784"), val = tensor([0, 2, 1, 3])]; + tensor var_3800 = const()[name = tensor("op_3800"), val = tensor([1, 1, 1])]; + tensor var_3801_cast_fp16 = reshape(shape = var_3800, x = position8_to_fp16)[name = tensor("op_3801_cast_fp16")]; + tensor q_positions_17_cast_fp16 = add(x = var_3801_cast_fp16, y = q_offsets_1_promoted_to_fp16)[name = tensor("q_positions_17_cast_fp16")]; + tensor valid_end_17_cast_fp16 = add(x = var_3801_cast_fp16, y = var_536_cast_fp16)[name = tensor("valid_end_17_cast_fp16")]; + tensor valid_mask_17_cast_fp16 = less(x = k_positions_1_promoted_to_fp16, y = valid_end_17_cast_fp16)[name = tensor("valid_mask_17_cast_fp16")]; + tensor causal_mask_17_cast_fp16 = less_equal(x = k_positions_1_promoted_to_fp16, y = q_positions_17_cast_fp16)[name = tensor("causal_mask_17_cast_fp16")]; + tensor var_3829 = logical_and(x = valid_mask_17_cast_fp16, y = causal_mask_17_cast_fp16)[name = tensor("op_3829")]; + tensor attn_mask_17_axes_0 = const()[name = tensor("attn_mask_17_axes_0"), val = tensor([1])]; + tensor attn_mask_17 = expand_dims(axes = attn_mask_17_axes_0, x = var_3829)[name = tensor("attn_mask_17")]; + tensor var_3843_transpose_x_0 = const()[name = tensor("op_3843_transpose_x_0"), val = tensor(false)]; + tensor var_3843_transpose_y_0 = const()[name = tensor("op_3843_transpose_y_0"), val = tensor(false)]; + tensor transpose_85_perm_0 = const()[name = tensor("transpose_85_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_86_perm_0 = const()[name = tensor("transpose_86_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_86 = transpose(perm = transpose_86_perm_0, x = keys_51_cast_fp16)[name = tensor("transpose_172")]; + tensor transpose_85 = transpose(perm = transpose_85_perm_0, x = q_51_cast_fp16)[name = tensor("transpose_173")]; + tensor var_3843_cast_fp16 = matmul(transpose_x = var_3843_transpose_x_0, transpose_y = var_3843_transpose_y_0, x = transpose_85, y = transpose_86)[name = tensor("op_3843_cast_fp16")]; + tensor var_3837_to_fp16 = const()[name = tensor("op_3837_to_fp16"), val = tensor([0x1p-3])]; + tensor attn_weights_49_cast_fp16 = mul(x = var_3843_cast_fp16, y = var_3837_to_fp16)[name = tensor("attn_weights_49_cast_fp16")]; + tensor var_3845 = logical_not(x = attn_mask_17)[name = tensor("op_3845")]; + tensor var_3846_to_fp16 = const()[name = tensor("op_3846_to_fp16"), val = tensor(-inf)]; + tensor attn_weights_51_cast_fp16 = select(a = var_3846_to_fp16, b = attn_weights_49_cast_fp16, cond = var_3845)[name = tensor("attn_weights_51_cast_fp16")]; + tensor var_3848 = const()[name = tensor("op_3848"), val = tensor(-1)]; + tensor attn_weights_53_cast_fp16 = softmax(axis = var_3848, x = attn_weights_51_cast_fp16)[name = tensor("attn_weights_53_cast_fp16")]; + tensor attn_output_17_transpose_x_0 = const()[name = tensor("attn_output_17_transpose_x_0"), val = tensor(false)]; + tensor attn_output_17_transpose_y_0 = const()[name = tensor("attn_output_17_transpose_y_0"), val = tensor(false)]; + tensor values_53_cast_fp16 = transpose(perm = var_3784, x = values_51_cast_fp16)[name = tensor("transpose_174")]; + tensor attn_output_17_cast_fp16 = matmul(transpose_x = attn_output_17_transpose_x_0, transpose_y = attn_output_17_transpose_y_0, x = attn_weights_53_cast_fp16, y = values_53_cast_fp16)[name = tensor("attn_output_17_cast_fp16")]; + tensor var_3856 = const()[name = tensor("op_3856"), val = tensor([0, 2, 1, 3])]; + tensor var_3861 = const()[name = tensor("op_3861"), val = tensor([1, 256, 1024])]; + tensor var_3857_cast_fp16 = transpose(perm = var_3856, x = attn_output_17_cast_fp16)[name = tensor("transpose_171")]; + tensor input_101_cast_fp16 = reshape(shape = var_3861, x = var_3857_cast_fp16)[name = tensor("input_101_cast_fp16")]; + tensor attn8_out_proj_weight_to_fp16 = const()[name = tensor("attn8_out_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(208761216)))]; + tensor linear_33_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = attn8_out_proj_weight_to_fp16, x = input_101_cast_fp16)[name = tensor("linear_33_cast_fp16")]; + tensor var_3870_cast_fp16 = add(x = position8_to_fp16, y = valid_len_to_fp16)[name = tensor("op_3870_cast_fp16")]; + tensor var_3870_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_3870_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor input_103_cast_fp16 = add(x = input_97_cast_fp16, y = linear_33_cast_fp16)[name = tensor("input_103_cast_fp16")]; + tensor input_105_axes_0 = const()[name = tensor("input_105_axes_0"), val = tensor([-1])]; + tensor norm8_2_weight_to_fp16 = const()[name = tensor("norm8_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(210858432)))]; + tensor norm8_2_bias_to_fp16 = const()[name = tensor("norm8_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(210860544)))]; + tensor var_3874_to_fp16 = const()[name = tensor("op_3874_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_105_cast_fp16 = layer_norm(axes = input_105_axes_0, beta = norm8_2_bias_to_fp16, epsilon = var_3874_to_fp16, gamma = norm8_2_weight_to_fp16, x = input_103_cast_fp16)[name = tensor("input_105_cast_fp16")]; + tensor linear8_1_weight_to_fp16 = const()[name = tensor("linear8_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(210862656)))]; + tensor linear_34_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = linear8_1_weight_to_fp16, x = input_105_cast_fp16)[name = tensor("linear_34_cast_fp16")]; + tensor input_107_mode_0 = const()[name = tensor("input_107_mode_0"), val = tensor("EXACT")]; + tensor input_107_cast_fp16 = gelu(mode = input_107_mode_0, x = linear_34_cast_fp16)[name = tensor("input_107_cast_fp16")]; + tensor linear8_2_weight_to_fp16 = const()[name = tensor("linear8_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(219251328)))]; + tensor linear_35_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = linear8_2_weight_to_fp16, x = input_107_cast_fp16)[name = tensor("linear_35_cast_fp16")]; + tensor input_109_cast_fp16 = add(x = input_103_cast_fp16, y = linear_35_cast_fp16)[name = tensor("input_109_cast_fp16")]; + tensor input_111_axes_0 = const()[name = tensor("input_111_axes_0"), val = tensor([-1])]; + tensor norm9_1_weight_to_fp16 = const()[name = tensor("norm9_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(227640000)))]; + tensor norm9_1_bias_to_fp16 = const()[name = tensor("norm9_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(227642112)))]; + tensor var_3891_to_fp16 = const()[name = tensor("op_3891_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_111_cast_fp16 = layer_norm(axes = input_111_axes_0, beta = norm9_1_bias_to_fp16, epsilon = var_3891_to_fp16, gamma = norm9_1_weight_to_fp16, x = input_109_cast_fp16)[name = tensor("input_111_cast_fp16")]; + tensor attn9_in_proj_weight_to_fp16 = const()[name = tensor("attn9_in_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(227644224)))]; + tensor linear_36_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = attn9_in_proj_weight_to_fp16, x = input_111_cast_fp16)[name = tensor("linear_36_cast_fp16")]; + tensor var_3905 = const()[name = tensor("op_3905"), val = tensor([1, 256, 3, 16, 64])]; + tensor qkv_19_cast_fp16 = reshape(shape = var_3905, x = linear_36_cast_fp16)[name = tensor("qkv_19_cast_fp16")]; + tensor q_55_begin_0 = const()[name = tensor("q_55_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor q_55_end_0 = const()[name = tensor("q_55_end_0"), val = tensor([1, 256, 1, 16, 64])]; + tensor q_55_end_mask_0 = const()[name = tensor("q_55_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor q_55_squeeze_mask_0 = const()[name = tensor("q_55_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor q_55_cast_fp16 = slice_by_index(begin = q_55_begin_0, end = q_55_end_0, end_mask = q_55_end_mask_0, squeeze_mask = q_55_squeeze_mask_0, x = qkv_19_cast_fp16)[name = tensor("q_55_cast_fp16")]; + tensor k_37_begin_0 = const()[name = tensor("k_37_begin_0"), val = tensor([0, 0, 1, 0, 0])]; + tensor k_37_end_0 = const()[name = tensor("k_37_end_0"), val = tensor([1, 256, 2, 16, 64])]; + tensor k_37_end_mask_0 = const()[name = tensor("k_37_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor k_37_squeeze_mask_0 = const()[name = tensor("k_37_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor k_37_cast_fp16 = slice_by_index(begin = k_37_begin_0, end = k_37_end_0, end_mask = k_37_end_mask_0, squeeze_mask = k_37_squeeze_mask_0, x = qkv_19_cast_fp16)[name = tensor("k_37_cast_fp16")]; + tensor v_19_begin_0 = const()[name = tensor("v_19_begin_0"), val = tensor([0, 0, 2, 0, 0])]; + tensor v_19_end_0 = const()[name = tensor("v_19_end_0"), val = tensor([1, 256, 3, 16, 64])]; + tensor v_19_end_mask_0 = const()[name = tensor("v_19_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor v_19_squeeze_mask_0 = const()[name = tensor("v_19_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor v_19_cast_fp16 = slice_by_index(begin = v_19_begin_0, end = v_19_end_0, end_mask = v_19_end_mask_0, squeeze_mask = v_19_squeeze_mask_0, x = qkv_19_cast_fp16)[name = tensor("v_19_cast_fp16")]; + tensor var_3963 = const()[name = tensor("op_3963"), val = tensor([1, 1])]; + tensor position9_to_fp16_dtype_0 = const()[name = tensor("position9_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor position9_to_fp16 = cast(dtype = position9_to_fp16_dtype_0, x = position9)[name = tensor("cast_616")]; + tensor var_3964_cast_fp16 = reshape(shape = var_3963, x = position9_to_fp16)[name = tensor("op_3964_cast_fp16")]; + tensor ts_57_cast_fp16 = add(x = ts_1_promoted_to_fp16, y = var_3964_cast_fp16)[name = tensor("ts_57_cast_fp16")]; + tensor var_3971 = const()[name = tensor("op_3971"), val = tensor([1, 256, 1, 1])]; + tensor ts_59_cast_fp16 = reshape(shape = var_3971, x = ts_57_cast_fp16)[name = tensor("ts_59_cast_fp16")]; + tensor var_3978 = const()[name = tensor("op_3978"), val = tensor([1, 256, 16, 32, 2])]; + tensor q_complex_19_cast_fp16 = reshape(shape = var_3978, x = q_55_cast_fp16)[name = tensor("q_complex_19_cast_fp16")]; + tensor var_3985 = const()[name = tensor("op_3985"), val = tensor([1, 256, 16, 32, 2])]; + tensor k_complex_19_cast_fp16 = reshape(shape = var_3985, x = k_37_cast_fp16)[name = tensor("k_complex_19_cast_fp16")]; + tensor var_3989_begin_0 = const()[name = tensor("op_3989_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_3989_end_0 = const()[name = tensor("op_3989_end_0"), val = tensor([1, 256, 16, 32, 1])]; + tensor var_3989_end_mask_0 = const()[name = tensor("op_3989_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_3989_squeeze_mask_0 = const()[name = tensor("op_3989_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_3989_cast_fp16 = slice_by_index(begin = var_3989_begin_0, end = var_3989_end_0, end_mask = var_3989_end_mask_0, squeeze_mask = var_3989_squeeze_mask_0, x = q_complex_19_cast_fp16)[name = tensor("op_3989_cast_fp16")]; + tensor var_3997_begin_0 = const()[name = tensor("op_3997_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_3997_end_0 = const()[name = tensor("op_3997_end_0"), val = tensor([1, 256, 16, 32, 2])]; + tensor var_3997_end_mask_0 = const()[name = tensor("op_3997_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_3997_squeeze_mask_0 = const()[name = tensor("op_3997_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_3997_cast_fp16 = slice_by_index(begin = var_3997_begin_0, end = var_3997_end_0, end_mask = var_3997_end_mask_0, squeeze_mask = var_3997_squeeze_mask_0, x = q_complex_19_cast_fp16)[name = tensor("op_3997_cast_fp16")]; + tensor var_4005_begin_0 = const()[name = tensor("op_4005_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_4005_end_0 = const()[name = tensor("op_4005_end_0"), val = tensor([1, 256, 16, 32, 1])]; + tensor var_4005_end_mask_0 = const()[name = tensor("op_4005_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_4005_squeeze_mask_0 = const()[name = tensor("op_4005_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_4005_cast_fp16 = slice_by_index(begin = var_4005_begin_0, end = var_4005_end_0, end_mask = var_4005_end_mask_0, squeeze_mask = var_4005_squeeze_mask_0, x = k_complex_19_cast_fp16)[name = tensor("op_4005_cast_fp16")]; + tensor var_4013_begin_0 = const()[name = tensor("op_4013_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_4013_end_0 = const()[name = tensor("op_4013_end_0"), val = tensor([1, 256, 16, 32, 2])]; + tensor var_4013_end_mask_0 = const()[name = tensor("op_4013_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_4013_squeeze_mask_0 = const()[name = tensor("op_4013_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_4013_cast_fp16 = slice_by_index(begin = var_4013_begin_0, end = var_4013_end_0, end_mask = var_4013_end_mask_0, squeeze_mask = var_4013_squeeze_mask_0, x = k_complex_19_cast_fp16)[name = tensor("op_4013_cast_fp16")]; + tensor freqs_19_to_fp16 = const()[name = tensor("freqs_19_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(233935744)))]; + tensor var_4019_cast_fp16 = mul(x = freqs_19_to_fp16, y = ts_59_cast_fp16)[name = tensor("op_4019_cast_fp16")]; + tensor rotr_19_cast_fp16 = cos(x = var_4019_cast_fp16)[name = tensor("rotr_19_cast_fp16")]; + tensor roti_19_cast_fp16 = sin(x = var_4019_cast_fp16)[name = tensor("roti_19_cast_fp16")]; + tensor var_4023_cast_fp16 = mul(x = var_3989_cast_fp16, y = rotr_19_cast_fp16)[name = tensor("op_4023_cast_fp16")]; + tensor var_4024_cast_fp16 = mul(x = var_3997_cast_fp16, y = roti_19_cast_fp16)[name = tensor("op_4024_cast_fp16")]; + tensor qor_37_cast_fp16 = sub(x = var_4023_cast_fp16, y = var_4024_cast_fp16)[name = tensor("qor_37_cast_fp16")]; + tensor var_4027_cast_fp16 = mul(x = var_3989_cast_fp16, y = roti_19_cast_fp16)[name = tensor("op_4027_cast_fp16")]; + tensor var_4028_cast_fp16 = mul(x = var_3997_cast_fp16, y = rotr_19_cast_fp16)[name = tensor("op_4028_cast_fp16")]; + tensor qoi_37_cast_fp16 = add(x = var_4027_cast_fp16, y = var_4028_cast_fp16)[name = tensor("qoi_37_cast_fp16")]; + tensor var_4031_cast_fp16 = mul(x = var_4005_cast_fp16, y = rotr_19_cast_fp16)[name = tensor("op_4031_cast_fp16")]; + tensor var_4032_cast_fp16 = mul(x = var_4013_cast_fp16, y = roti_19_cast_fp16)[name = tensor("op_4032_cast_fp16")]; + tensor kor_37_cast_fp16 = sub(x = var_4031_cast_fp16, y = var_4032_cast_fp16)[name = tensor("kor_37_cast_fp16")]; + tensor var_4035_cast_fp16 = mul(x = var_4005_cast_fp16, y = roti_19_cast_fp16)[name = tensor("op_4035_cast_fp16")]; + tensor var_4036_cast_fp16 = mul(x = var_4013_cast_fp16, y = rotr_19_cast_fp16)[name = tensor("op_4036_cast_fp16")]; + tensor koi_37_cast_fp16 = add(x = var_4035_cast_fp16, y = var_4036_cast_fp16)[name = tensor("koi_37_cast_fp16")]; + tensor qo_19_axis_0 = const()[name = tensor("qo_19_axis_0"), val = tensor(-1)]; + tensor qo_19_cast_fp16 = stack(axis = qo_19_axis_0, values = (qor_37_cast_fp16, qoi_37_cast_fp16))[name = tensor("qo_19_cast_fp16")]; + tensor ko_19_axis_0 = const()[name = tensor("ko_19_axis_0"), val = tensor(-1)]; + tensor ko_19_cast_fp16 = stack(axis = ko_19_axis_0, values = (kor_37_cast_fp16, koi_37_cast_fp16))[name = tensor("ko_19_cast_fp16")]; + tensor var_4069 = const()[name = tensor("op_4069"), val = tensor([1, 256, 16, 64])]; + tensor q_57_cast_fp16 = reshape(shape = var_4069, x = qo_19_cast_fp16)[name = tensor("q_57_cast_fp16")]; + tensor var_4075 = const()[name = tensor("op_4075"), val = tensor([1, 256, 16, 64])]; + tensor k_39_cast_fp16 = reshape(shape = var_4075, x = ko_19_cast_fp16)[name = tensor("k_39_cast_fp16")]; + tensor write_indices_float_55_cast_fp16 = add(x = var_3964_cast_fp16, y = var_393_promoted_to_fp16)[name = tensor("write_indices_float_55_cast_fp16")]; + tensor _inversed_4096_y_0_to_fp16 = const()[name = tensor("_inversed_4096_y_0_to_fp16"), val = tensor(0x1p-9)]; + tensor _inversed_4096_cast_fp16 = mul(x = write_indices_float_55_cast_fp16, y = _inversed_4096_y_0_to_fp16)[name = tensor("_inversed_4096_cast_fp16")]; + tensor var_4097_cast_fp16 = floor(x = _inversed_4096_cast_fp16)[name = tensor("op_4097_cast_fp16")]; + tensor var_4098_to_fp16 = const()[name = tensor("op_4098_to_fp16"), val = tensor(0x1p+9)]; + tensor var_4099_cast_fp16 = mul(x = var_4097_cast_fp16, y = var_4098_to_fp16)[name = tensor("op_4099_cast_fp16")]; + tensor write_indices_float_57_cast_fp16 = sub(x = write_indices_float_55_cast_fp16, y = var_4099_cast_fp16)[name = tensor("write_indices_float_57_cast_fp16")]; + tensor var_4117_value_0_to_fp16 = const()[name = tensor("op_4117_value_0_to_fp16"), val = tensor(0x1.ffp+8)]; + tensor var_4117_cast_fp16 = fill_like(ref_tensor = write_indices_float_57_cast_fp16, value = var_4117_value_0_to_fp16)[name = tensor("op_4117_cast_fp16")]; + tensor write_indices_float_59_cast_fp16 = select(a = write_indices_float_57_cast_fp16, b = var_4117_cast_fp16, cond = token_valid_1_cast_fp16)[name = tensor("write_indices_float_59_cast_fp16")]; + tensor var_4123_dtype_0 = const()[name = tensor("op_4123_dtype_0"), val = tensor("int32")]; + tensor var_4128 = const()[name = tensor("op_4128"), val = tensor([1, 256, 1, 1])]; + tensor write_indices_float_59_cast_fp16_to_int32 = cast(dtype = var_4123_dtype_0, x = write_indices_float_59_cast_fp16)[name = tensor("cast_615")]; + tensor var_4129 = reshape(shape = var_4128, x = write_indices_float_59_cast_fp16_to_int32)[name = tensor("op_4129")]; + tensor write_indices_19_reps_0 = const()[name = tensor("write_indices_19_reps_0"), val = tensor([1, 1, 16, 64])]; + tensor write_indices_19 = tile(reps = write_indices_19_reps_0, x = var_4129)[name = tensor("write_indices_19")]; + tensor var_4139_begin_0 = const()[name = tensor("op_4139_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_4139_end_0 = const()[name = tensor("op_4139_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor var_4139_end_mask_0 = const()[name = tensor("op_4139_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_4139_squeeze_mask_0 = const()[name = tensor("op_4139_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor cache9_to_fp16_dtype_0 = const()[name = tensor("cache9_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor cache9_to_fp16 = cast(dtype = cache9_to_fp16_dtype_0, x = cache9)[name = tensor("cast_614")]; + tensor var_4139_cast_fp16 = slice_by_index(begin = var_4139_begin_0, end = var_4139_end_0, end_mask = var_4139_end_mask_0, squeeze_mask = var_4139_squeeze_mask_0, x = cache9_to_fp16)[name = tensor("op_4139_cast_fp16")]; + tensor var_4141_axis_0 = const()[name = tensor("op_4141_axis_0"), val = tensor(1)]; + tensor var_4141_mode_0 = const()[name = tensor("op_4141_mode_0"), val = tensor("update")]; + tensor var_4141_validate_indices_0 = const()[name = tensor("op_4141_validate_indices_0"), val = tensor(false)]; + tensor var_4141_cast_fp16 = scatter_along_axis(axis = var_4141_axis_0, data = var_4139_cast_fp16, indices = write_indices_19, mode = var_4141_mode_0, updates = k_39_cast_fp16, validate_indices = var_4141_validate_indices_0)[name = tensor("op_4141_cast_fp16")]; + tensor concat_74 = const()[name = tensor("concat_74"), val = tensor([0, 0, 0, 0, 0])]; + tensor concat_75 = const()[name = tensor("concat_75"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_19_internal_tensor_assign_1_stride_0 = const()[name = tensor("new_cache_19_internal_tensor_assign_1_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_19_internal_tensor_assign_1_begin_mask_0 = const()[name = tensor("new_cache_19_internal_tensor_assign_1_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_19_internal_tensor_assign_1_end_mask_0 = const()[name = tensor("new_cache_19_internal_tensor_assign_1_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_19_internal_tensor_assign_1_squeeze_mask_0 = const()[name = tensor("new_cache_19_internal_tensor_assign_1_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_64 = const()[name = tensor("shape_64"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_18 = const()[name = tensor("reduce_prod_18"), val = tensor(1048576)]; + tensor range_1d_18_start_0 = const()[name = tensor("range_1d_18_start_0"), val = tensor(0)]; + tensor range_1d_18_step_0 = const()[name = tensor("range_1d_18_step_0"), val = tensor(1)]; + tensor range_1d_18 = range_1d(end = reduce_prod_18, start = range_1d_18_start_0, step = range_1d_18_step_0)[name = tensor("range_1d_18")]; + tensor reshape_90 = reshape(shape = shape_64, x = range_1d_18)[name = tensor("reshape_90")]; + tensor slice_by_index_18 = slice_by_index(begin = concat_74, begin_mask = new_cache_19_internal_tensor_assign_1_begin_mask_0, end = concat_75, end_mask = new_cache_19_internal_tensor_assign_1_end_mask_0, squeeze_mask = new_cache_19_internal_tensor_assign_1_squeeze_mask_0, stride = new_cache_19_internal_tensor_assign_1_stride_0, x = reshape_90)[name = tensor("slice_by_index_18")]; + tensor reshape_91_shape_0 = const()[name = tensor("reshape_91_shape_0"), val = tensor([-1])]; + tensor reshape_91 = reshape(shape = reshape_91_shape_0, x = slice_by_index_18)[name = tensor("reshape_91")]; + tensor reshape_92_shape_0 = const()[name = tensor("reshape_92_shape_0"), val = tensor([-1])]; + tensor reshape_92_cast_fp16 = reshape(shape = reshape_92_shape_0, x = var_4141_cast_fp16)[name = tensor("reshape_92_cast_fp16")]; + tensor reshape_93_shape_0 = const()[name = tensor("reshape_93_shape_0"), val = tensor([-1])]; + tensor reshape_93_cast_fp16 = reshape(shape = reshape_93_shape_0, x = cache9_to_fp16)[name = tensor("reshape_93_cast_fp16")]; + tensor scatter_18_mode_0 = const()[name = tensor("scatter_18_mode_0"), val = tensor("update")]; + tensor scatter_18_axis_0 = const()[name = tensor("scatter_18_axis_0"), val = tensor(0)]; + tensor scatter_18_validate_indices_0 = const()[name = tensor("scatter_18_validate_indices_0"), val = tensor(false)]; + tensor scatter_18_cast_fp16 = scatter(axis = scatter_18_axis_0, data = reshape_93_cast_fp16, indices = reshape_91, mode = scatter_18_mode_0, updates = reshape_92_cast_fp16, validate_indices = scatter_18_validate_indices_0)[name = tensor("scatter_18_cast_fp16")]; + tensor reshape_94_cast_fp16 = reshape(shape = shape_64, x = scatter_18_cast_fp16)[name = tensor("reshape_94_cast_fp16")]; + tensor var_4149_begin_0 = const()[name = tensor("op_4149_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor var_4149_end_0 = const()[name = tensor("op_4149_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor var_4149_end_mask_0 = const()[name = tensor("op_4149_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_4149_squeeze_mask_0 = const()[name = tensor("op_4149_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor var_4149_cast_fp16 = slice_by_index(begin = var_4149_begin_0, end = var_4149_end_0, end_mask = var_4149_end_mask_0, squeeze_mask = var_4149_squeeze_mask_0, x = reshape_94_cast_fp16)[name = tensor("op_4149_cast_fp16")]; + tensor var_4151_axis_0 = const()[name = tensor("op_4151_axis_0"), val = tensor(1)]; + tensor var_4151_mode_0 = const()[name = tensor("op_4151_mode_0"), val = tensor("update")]; + tensor var_4151_validate_indices_0 = const()[name = tensor("op_4151_validate_indices_0"), val = tensor(false)]; + tensor var_4151_cast_fp16 = scatter_along_axis(axis = var_4151_axis_0, data = var_4149_cast_fp16, indices = write_indices_19, mode = var_4151_mode_0, updates = v_19_cast_fp16, validate_indices = var_4151_validate_indices_0)[name = tensor("op_4151_cast_fp16")]; + tensor concat_76 = const()[name = tensor("concat_76"), val = tensor([1, 0, 0, 0, 0])]; + tensor concat_77 = const()[name = tensor("concat_77"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_19_internal_tensor_assign_2_stride_0 = const()[name = tensor("new_cache_19_internal_tensor_assign_2_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_19_internal_tensor_assign_2_begin_mask_0 = const()[name = tensor("new_cache_19_internal_tensor_assign_2_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_19_internal_tensor_assign_2_end_mask_0 = const()[name = tensor("new_cache_19_internal_tensor_assign_2_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_19_internal_tensor_assign_2_squeeze_mask_0 = const()[name = tensor("new_cache_19_internal_tensor_assign_2_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_65 = const()[name = tensor("shape_65"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_19 = const()[name = tensor("reduce_prod_19"), val = tensor(1048576)]; + tensor range_1d_19_start_0 = const()[name = tensor("range_1d_19_start_0"), val = tensor(0)]; + tensor range_1d_19_step_0 = const()[name = tensor("range_1d_19_step_0"), val = tensor(1)]; + tensor range_1d_19 = range_1d(end = reduce_prod_19, start = range_1d_19_start_0, step = range_1d_19_step_0)[name = tensor("range_1d_19")]; + tensor reshape_95 = reshape(shape = shape_65, x = range_1d_19)[name = tensor("reshape_95")]; + tensor slice_by_index_19 = slice_by_index(begin = concat_76, begin_mask = new_cache_19_internal_tensor_assign_2_begin_mask_0, end = concat_77, end_mask = new_cache_19_internal_tensor_assign_2_end_mask_0, squeeze_mask = new_cache_19_internal_tensor_assign_2_squeeze_mask_0, stride = new_cache_19_internal_tensor_assign_2_stride_0, x = reshape_95)[name = tensor("slice_by_index_19")]; + tensor reshape_96_shape_0 = const()[name = tensor("reshape_96_shape_0"), val = tensor([-1])]; + tensor reshape_96 = reshape(shape = reshape_96_shape_0, x = slice_by_index_19)[name = tensor("reshape_96")]; + tensor reshape_97_shape_0 = const()[name = tensor("reshape_97_shape_0"), val = tensor([-1])]; + tensor reshape_97_cast_fp16 = reshape(shape = reshape_97_shape_0, x = var_4151_cast_fp16)[name = tensor("reshape_97_cast_fp16")]; + tensor reshape_98_shape_0 = const()[name = tensor("reshape_98_shape_0"), val = tensor([-1])]; + tensor reshape_98_cast_fp16 = reshape(shape = reshape_98_shape_0, x = reshape_94_cast_fp16)[name = tensor("reshape_98_cast_fp16")]; + tensor scatter_19_mode_0 = const()[name = tensor("scatter_19_mode_0"), val = tensor("update")]; + tensor scatter_19_axis_0 = const()[name = tensor("scatter_19_axis_0"), val = tensor(0)]; + tensor scatter_19_validate_indices_0 = const()[name = tensor("scatter_19_validate_indices_0"), val = tensor(false)]; + tensor scatter_19_cast_fp16 = scatter(axis = scatter_19_axis_0, data = reshape_98_cast_fp16, indices = reshape_96, mode = scatter_19_mode_0, updates = reshape_97_cast_fp16, validate_indices = scatter_19_validate_indices_0)[name = tensor("scatter_19_cast_fp16")]; + tensor reshape_99_cast_fp16 = reshape(shape = shape_65, x = scatter_19_cast_fp16)[name = tensor("reshape_99_cast_fp16")]; + tensor reshape_99_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("reshape_99_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor keys_55_begin_0 = const()[name = tensor("keys_55_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor keys_55_end_0 = const()[name = tensor("keys_55_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor keys_55_end_mask_0 = const()[name = tensor("keys_55_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor keys_55_squeeze_mask_0 = const()[name = tensor("keys_55_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor keys_55_cast_fp16 = slice_by_index(begin = keys_55_begin_0, end = keys_55_end_0, end_mask = keys_55_end_mask_0, squeeze_mask = keys_55_squeeze_mask_0, x = reshape_99_cast_fp16)[name = tensor("keys_55_cast_fp16")]; + tensor values_55_begin_0 = const()[name = tensor("values_55_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor values_55_end_0 = const()[name = tensor("values_55_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor values_55_end_mask_0 = const()[name = tensor("values_55_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor values_55_squeeze_mask_0 = const()[name = tensor("values_55_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor values_55_cast_fp16 = slice_by_index(begin = values_55_begin_0, end = values_55_end_0, end_mask = values_55_end_mask_0, squeeze_mask = values_55_squeeze_mask_0, x = reshape_99_cast_fp16)[name = tensor("values_55_cast_fp16")]; + tensor var_4163_cast_fp16 = not_equal(x = keys_55_cast_fp16, y = keys_55_cast_fp16)[name = tensor("op_4163_cast_fp16")]; + tensor keys_57_cast_fp16 = select(a = var_470_to_fp16, b = keys_55_cast_fp16, cond = var_4163_cast_fp16)[name = tensor("keys_57_cast_fp16")]; + tensor var_4171_cast_fp16 = not_equal(x = values_55_cast_fp16, y = values_55_cast_fp16)[name = tensor("op_4171_cast_fp16")]; + tensor values_57_cast_fp16 = select(a = var_470_to_fp16, b = values_55_cast_fp16, cond = var_4171_cast_fp16)[name = tensor("values_57_cast_fp16")]; + tensor var_4195 = const()[name = tensor("op_4195"), val = tensor([0, 2, 1, 3])]; + tensor var_4211 = const()[name = tensor("op_4211"), val = tensor([1, 1, 1])]; + tensor var_4212_cast_fp16 = reshape(shape = var_4211, x = position9_to_fp16)[name = tensor("op_4212_cast_fp16")]; + tensor q_positions_19_cast_fp16 = add(x = var_4212_cast_fp16, y = q_offsets_1_promoted_to_fp16)[name = tensor("q_positions_19_cast_fp16")]; + tensor valid_end_19_cast_fp16 = add(x = var_4212_cast_fp16, y = var_536_cast_fp16)[name = tensor("valid_end_19_cast_fp16")]; + tensor valid_mask_19_cast_fp16 = less(x = k_positions_1_promoted_to_fp16, y = valid_end_19_cast_fp16)[name = tensor("valid_mask_19_cast_fp16")]; + tensor causal_mask_19_cast_fp16 = less_equal(x = k_positions_1_promoted_to_fp16, y = q_positions_19_cast_fp16)[name = tensor("causal_mask_19_cast_fp16")]; + tensor var_4240 = logical_and(x = valid_mask_19_cast_fp16, y = causal_mask_19_cast_fp16)[name = tensor("op_4240")]; + tensor attn_mask_19_axes_0 = const()[name = tensor("attn_mask_19_axes_0"), val = tensor([1])]; + tensor attn_mask_19 = expand_dims(axes = attn_mask_19_axes_0, x = var_4240)[name = tensor("attn_mask_19")]; + tensor var_4254_transpose_x_0 = const()[name = tensor("op_4254_transpose_x_0"), val = tensor(false)]; + tensor var_4254_transpose_y_0 = const()[name = tensor("op_4254_transpose_y_0"), val = tensor(false)]; + tensor transpose_87_perm_0 = const()[name = tensor("transpose_87_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_88_perm_0 = const()[name = tensor("transpose_88_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_88 = transpose(perm = transpose_88_perm_0, x = keys_57_cast_fp16)[name = tensor("transpose_168")]; + tensor transpose_87 = transpose(perm = transpose_87_perm_0, x = q_57_cast_fp16)[name = tensor("transpose_169")]; + tensor var_4254_cast_fp16 = matmul(transpose_x = var_4254_transpose_x_0, transpose_y = var_4254_transpose_y_0, x = transpose_87, y = transpose_88)[name = tensor("op_4254_cast_fp16")]; + tensor var_4248_to_fp16 = const()[name = tensor("op_4248_to_fp16"), val = tensor([0x1p-3])]; + tensor attn_weights_55_cast_fp16 = mul(x = var_4254_cast_fp16, y = var_4248_to_fp16)[name = tensor("attn_weights_55_cast_fp16")]; + tensor var_4256 = logical_not(x = attn_mask_19)[name = tensor("op_4256")]; + tensor var_4257_to_fp16 = const()[name = tensor("op_4257_to_fp16"), val = tensor(-inf)]; + tensor attn_weights_57_cast_fp16 = select(a = var_4257_to_fp16, b = attn_weights_55_cast_fp16, cond = var_4256)[name = tensor("attn_weights_57_cast_fp16")]; + tensor var_4259 = const()[name = tensor("op_4259"), val = tensor(-1)]; + tensor attn_weights_59_cast_fp16 = softmax(axis = var_4259, x = attn_weights_57_cast_fp16)[name = tensor("attn_weights_59_cast_fp16")]; + tensor attn_output_19_transpose_x_0 = const()[name = tensor("attn_output_19_transpose_x_0"), val = tensor(false)]; + tensor attn_output_19_transpose_y_0 = const()[name = tensor("attn_output_19_transpose_y_0"), val = tensor(false)]; + tensor values_59_cast_fp16 = transpose(perm = var_4195, x = values_57_cast_fp16)[name = tensor("transpose_170")]; + tensor attn_output_19_cast_fp16 = matmul(transpose_x = attn_output_19_transpose_x_0, transpose_y = attn_output_19_transpose_y_0, x = attn_weights_59_cast_fp16, y = values_59_cast_fp16)[name = tensor("attn_output_19_cast_fp16")]; + tensor var_4267 = const()[name = tensor("op_4267"), val = tensor([0, 2, 1, 3])]; + tensor var_4272 = const()[name = tensor("op_4272"), val = tensor([1, 256, 1024])]; + tensor var_4268_cast_fp16 = transpose(perm = var_4267, x = attn_output_19_cast_fp16)[name = tensor("transpose_167")]; + tensor input_113_cast_fp16 = reshape(shape = var_4272, x = var_4268_cast_fp16)[name = tensor("input_113_cast_fp16")]; + tensor attn9_out_proj_weight_to_fp16 = const()[name = tensor("attn9_out_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(233935872)))]; + tensor linear_37_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = attn9_out_proj_weight_to_fp16, x = input_113_cast_fp16)[name = tensor("linear_37_cast_fp16")]; + tensor var_4281_cast_fp16 = add(x = position9_to_fp16, y = valid_len_to_fp16)[name = tensor("op_4281_cast_fp16")]; + tensor var_4281_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_4281_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor input_115_cast_fp16 = add(x = input_109_cast_fp16, y = linear_37_cast_fp16)[name = tensor("input_115_cast_fp16")]; + tensor input_117_axes_0 = const()[name = tensor("input_117_axes_0"), val = tensor([-1])]; + tensor norm9_2_weight_to_fp16 = const()[name = tensor("norm9_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(236033088)))]; + tensor norm9_2_bias_to_fp16 = const()[name = tensor("norm9_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(236035200)))]; + tensor var_4285_to_fp16 = const()[name = tensor("op_4285_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_117_cast_fp16 = layer_norm(axes = input_117_axes_0, beta = norm9_2_bias_to_fp16, epsilon = var_4285_to_fp16, gamma = norm9_2_weight_to_fp16, x = input_115_cast_fp16)[name = tensor("input_117_cast_fp16")]; + tensor linear9_1_weight_to_fp16 = const()[name = tensor("linear9_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(236037312)))]; + tensor linear_38_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = linear9_1_weight_to_fp16, x = input_117_cast_fp16)[name = tensor("linear_38_cast_fp16")]; + tensor input_119_mode_0 = const()[name = tensor("input_119_mode_0"), val = tensor("EXACT")]; + tensor input_119_cast_fp16 = gelu(mode = input_119_mode_0, x = linear_38_cast_fp16)[name = tensor("input_119_cast_fp16")]; + tensor linear9_2_weight_to_fp16 = const()[name = tensor("linear9_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(244425984)))]; + tensor linear_39_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = linear9_2_weight_to_fp16, x = input_119_cast_fp16)[name = tensor("linear_39_cast_fp16")]; + tensor input_121_cast_fp16 = add(x = input_115_cast_fp16, y = linear_39_cast_fp16)[name = tensor("input_121_cast_fp16")]; + tensor input_123_axes_0 = const()[name = tensor("input_123_axes_0"), val = tensor([-1])]; + tensor norm10_1_weight_to_fp16 = const()[name = tensor("norm10_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(252814656)))]; + tensor norm10_1_bias_to_fp16 = const()[name = tensor("norm10_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(252816768)))]; + tensor var_4302_to_fp16 = const()[name = tensor("op_4302_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_123_cast_fp16 = layer_norm(axes = input_123_axes_0, beta = norm10_1_bias_to_fp16, epsilon = var_4302_to_fp16, gamma = norm10_1_weight_to_fp16, x = input_121_cast_fp16)[name = tensor("input_123_cast_fp16")]; + tensor attn10_in_proj_weight_to_fp16 = const()[name = tensor("attn10_in_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(252818880)))]; + tensor linear_40_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = attn10_in_proj_weight_to_fp16, x = input_123_cast_fp16)[name = tensor("linear_40_cast_fp16")]; + tensor var_4316 = const()[name = tensor("op_4316"), val = tensor([1, 256, 3, 16, 64])]; + tensor qkv_21_cast_fp16 = reshape(shape = var_4316, x = linear_40_cast_fp16)[name = tensor("qkv_21_cast_fp16")]; + tensor q_61_begin_0 = const()[name = tensor("q_61_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor q_61_end_0 = const()[name = tensor("q_61_end_0"), val = tensor([1, 256, 1, 16, 64])]; + tensor q_61_end_mask_0 = const()[name = tensor("q_61_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor q_61_squeeze_mask_0 = const()[name = tensor("q_61_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor q_61_cast_fp16 = slice_by_index(begin = q_61_begin_0, end = q_61_end_0, end_mask = q_61_end_mask_0, squeeze_mask = q_61_squeeze_mask_0, x = qkv_21_cast_fp16)[name = tensor("q_61_cast_fp16")]; + tensor k_41_begin_0 = const()[name = tensor("k_41_begin_0"), val = tensor([0, 0, 1, 0, 0])]; + tensor k_41_end_0 = const()[name = tensor("k_41_end_0"), val = tensor([1, 256, 2, 16, 64])]; + tensor k_41_end_mask_0 = const()[name = tensor("k_41_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor k_41_squeeze_mask_0 = const()[name = tensor("k_41_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor k_41_cast_fp16 = slice_by_index(begin = k_41_begin_0, end = k_41_end_0, end_mask = k_41_end_mask_0, squeeze_mask = k_41_squeeze_mask_0, x = qkv_21_cast_fp16)[name = tensor("k_41_cast_fp16")]; + tensor v_21_begin_0 = const()[name = tensor("v_21_begin_0"), val = tensor([0, 0, 2, 0, 0])]; + tensor v_21_end_0 = const()[name = tensor("v_21_end_0"), val = tensor([1, 256, 3, 16, 64])]; + tensor v_21_end_mask_0 = const()[name = tensor("v_21_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor v_21_squeeze_mask_0 = const()[name = tensor("v_21_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor v_21_cast_fp16 = slice_by_index(begin = v_21_begin_0, end = v_21_end_0, end_mask = v_21_end_mask_0, squeeze_mask = v_21_squeeze_mask_0, x = qkv_21_cast_fp16)[name = tensor("v_21_cast_fp16")]; + tensor var_4374 = const()[name = tensor("op_4374"), val = tensor([1, 1])]; + tensor position10_to_fp16_dtype_0 = const()[name = tensor("position10_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor position10_to_fp16 = cast(dtype = position10_to_fp16_dtype_0, x = position10)[name = tensor("cast_611")]; + tensor var_4375_cast_fp16 = reshape(shape = var_4374, x = position10_to_fp16)[name = tensor("op_4375_cast_fp16")]; + tensor ts_63_cast_fp16 = add(x = ts_1_promoted_to_fp16, y = var_4375_cast_fp16)[name = tensor("ts_63_cast_fp16")]; + tensor var_4382 = const()[name = tensor("op_4382"), val = tensor([1, 256, 1, 1])]; + tensor ts_65_cast_fp16 = reshape(shape = var_4382, x = ts_63_cast_fp16)[name = tensor("ts_65_cast_fp16")]; + tensor var_4389 = const()[name = tensor("op_4389"), val = tensor([1, 256, 16, 32, 2])]; + tensor q_complex_21_cast_fp16 = reshape(shape = var_4389, x = q_61_cast_fp16)[name = tensor("q_complex_21_cast_fp16")]; + tensor var_4396 = const()[name = tensor("op_4396"), val = tensor([1, 256, 16, 32, 2])]; + tensor k_complex_21_cast_fp16 = reshape(shape = var_4396, x = k_41_cast_fp16)[name = tensor("k_complex_21_cast_fp16")]; + tensor var_4400_begin_0 = const()[name = tensor("op_4400_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_4400_end_0 = const()[name = tensor("op_4400_end_0"), val = tensor([1, 256, 16, 32, 1])]; + tensor var_4400_end_mask_0 = const()[name = tensor("op_4400_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_4400_squeeze_mask_0 = const()[name = tensor("op_4400_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_4400_cast_fp16 = slice_by_index(begin = var_4400_begin_0, end = var_4400_end_0, end_mask = var_4400_end_mask_0, squeeze_mask = var_4400_squeeze_mask_0, x = q_complex_21_cast_fp16)[name = tensor("op_4400_cast_fp16")]; + tensor var_4408_begin_0 = const()[name = tensor("op_4408_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_4408_end_0 = const()[name = tensor("op_4408_end_0"), val = tensor([1, 256, 16, 32, 2])]; + tensor var_4408_end_mask_0 = const()[name = tensor("op_4408_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_4408_squeeze_mask_0 = const()[name = tensor("op_4408_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_4408_cast_fp16 = slice_by_index(begin = var_4408_begin_0, end = var_4408_end_0, end_mask = var_4408_end_mask_0, squeeze_mask = var_4408_squeeze_mask_0, x = q_complex_21_cast_fp16)[name = tensor("op_4408_cast_fp16")]; + tensor var_4416_begin_0 = const()[name = tensor("op_4416_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_4416_end_0 = const()[name = tensor("op_4416_end_0"), val = tensor([1, 256, 16, 32, 1])]; + tensor var_4416_end_mask_0 = const()[name = tensor("op_4416_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_4416_squeeze_mask_0 = const()[name = tensor("op_4416_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_4416_cast_fp16 = slice_by_index(begin = var_4416_begin_0, end = var_4416_end_0, end_mask = var_4416_end_mask_0, squeeze_mask = var_4416_squeeze_mask_0, x = k_complex_21_cast_fp16)[name = tensor("op_4416_cast_fp16")]; + tensor var_4424_begin_0 = const()[name = tensor("op_4424_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_4424_end_0 = const()[name = tensor("op_4424_end_0"), val = tensor([1, 256, 16, 32, 2])]; + tensor var_4424_end_mask_0 = const()[name = tensor("op_4424_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_4424_squeeze_mask_0 = const()[name = tensor("op_4424_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_4424_cast_fp16 = slice_by_index(begin = var_4424_begin_0, end = var_4424_end_0, end_mask = var_4424_end_mask_0, squeeze_mask = var_4424_squeeze_mask_0, x = k_complex_21_cast_fp16)[name = tensor("op_4424_cast_fp16")]; + tensor freqs_21_to_fp16 = const()[name = tensor("freqs_21_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(259110400)))]; + tensor var_4430_cast_fp16 = mul(x = freqs_21_to_fp16, y = ts_65_cast_fp16)[name = tensor("op_4430_cast_fp16")]; + tensor rotr_21_cast_fp16 = cos(x = var_4430_cast_fp16)[name = tensor("rotr_21_cast_fp16")]; + tensor roti_21_cast_fp16 = sin(x = var_4430_cast_fp16)[name = tensor("roti_21_cast_fp16")]; + tensor var_4434_cast_fp16 = mul(x = var_4400_cast_fp16, y = rotr_21_cast_fp16)[name = tensor("op_4434_cast_fp16")]; + tensor var_4435_cast_fp16 = mul(x = var_4408_cast_fp16, y = roti_21_cast_fp16)[name = tensor("op_4435_cast_fp16")]; + tensor qor_41_cast_fp16 = sub(x = var_4434_cast_fp16, y = var_4435_cast_fp16)[name = tensor("qor_41_cast_fp16")]; + tensor var_4438_cast_fp16 = mul(x = var_4400_cast_fp16, y = roti_21_cast_fp16)[name = tensor("op_4438_cast_fp16")]; + tensor var_4439_cast_fp16 = mul(x = var_4408_cast_fp16, y = rotr_21_cast_fp16)[name = tensor("op_4439_cast_fp16")]; + tensor qoi_41_cast_fp16 = add(x = var_4438_cast_fp16, y = var_4439_cast_fp16)[name = tensor("qoi_41_cast_fp16")]; + tensor var_4442_cast_fp16 = mul(x = var_4416_cast_fp16, y = rotr_21_cast_fp16)[name = tensor("op_4442_cast_fp16")]; + tensor var_4443_cast_fp16 = mul(x = var_4424_cast_fp16, y = roti_21_cast_fp16)[name = tensor("op_4443_cast_fp16")]; + tensor kor_41_cast_fp16 = sub(x = var_4442_cast_fp16, y = var_4443_cast_fp16)[name = tensor("kor_41_cast_fp16")]; + tensor var_4446_cast_fp16 = mul(x = var_4416_cast_fp16, y = roti_21_cast_fp16)[name = tensor("op_4446_cast_fp16")]; + tensor var_4447_cast_fp16 = mul(x = var_4424_cast_fp16, y = rotr_21_cast_fp16)[name = tensor("op_4447_cast_fp16")]; + tensor koi_41_cast_fp16 = add(x = var_4446_cast_fp16, y = var_4447_cast_fp16)[name = tensor("koi_41_cast_fp16")]; + tensor qo_21_axis_0 = const()[name = tensor("qo_21_axis_0"), val = tensor(-1)]; + tensor qo_21_cast_fp16 = stack(axis = qo_21_axis_0, values = (qor_41_cast_fp16, qoi_41_cast_fp16))[name = tensor("qo_21_cast_fp16")]; + tensor ko_21_axis_0 = const()[name = tensor("ko_21_axis_0"), val = tensor(-1)]; + tensor ko_21_cast_fp16 = stack(axis = ko_21_axis_0, values = (kor_41_cast_fp16, koi_41_cast_fp16))[name = tensor("ko_21_cast_fp16")]; + tensor var_4480 = const()[name = tensor("op_4480"), val = tensor([1, 256, 16, 64])]; + tensor q_63_cast_fp16 = reshape(shape = var_4480, x = qo_21_cast_fp16)[name = tensor("q_63_cast_fp16")]; + tensor var_4486 = const()[name = tensor("op_4486"), val = tensor([1, 256, 16, 64])]; + tensor k_43_cast_fp16 = reshape(shape = var_4486, x = ko_21_cast_fp16)[name = tensor("k_43_cast_fp16")]; + tensor write_indices_float_61_cast_fp16 = add(x = var_4375_cast_fp16, y = var_393_promoted_to_fp16)[name = tensor("write_indices_float_61_cast_fp16")]; + tensor _inversed_4507_y_0_to_fp16 = const()[name = tensor("_inversed_4507_y_0_to_fp16"), val = tensor(0x1p-9)]; + tensor _inversed_4507_cast_fp16 = mul(x = write_indices_float_61_cast_fp16, y = _inversed_4507_y_0_to_fp16)[name = tensor("_inversed_4507_cast_fp16")]; + tensor var_4508_cast_fp16 = floor(x = _inversed_4507_cast_fp16)[name = tensor("op_4508_cast_fp16")]; + tensor var_4509_to_fp16 = const()[name = tensor("op_4509_to_fp16"), val = tensor(0x1p+9)]; + tensor var_4510_cast_fp16 = mul(x = var_4508_cast_fp16, y = var_4509_to_fp16)[name = tensor("op_4510_cast_fp16")]; + tensor write_indices_float_63_cast_fp16 = sub(x = write_indices_float_61_cast_fp16, y = var_4510_cast_fp16)[name = tensor("write_indices_float_63_cast_fp16")]; + tensor var_4528_value_0_to_fp16 = const()[name = tensor("op_4528_value_0_to_fp16"), val = tensor(0x1.ffp+8)]; + tensor var_4528_cast_fp16 = fill_like(ref_tensor = write_indices_float_63_cast_fp16, value = var_4528_value_0_to_fp16)[name = tensor("op_4528_cast_fp16")]; + tensor write_indices_float_65_cast_fp16 = select(a = write_indices_float_63_cast_fp16, b = var_4528_cast_fp16, cond = token_valid_1_cast_fp16)[name = tensor("write_indices_float_65_cast_fp16")]; + tensor var_4534_dtype_0 = const()[name = tensor("op_4534_dtype_0"), val = tensor("int32")]; + tensor var_4539 = const()[name = tensor("op_4539"), val = tensor([1, 256, 1, 1])]; + tensor write_indices_float_65_cast_fp16_to_int32 = cast(dtype = var_4534_dtype_0, x = write_indices_float_65_cast_fp16)[name = tensor("cast_610")]; + tensor var_4540 = reshape(shape = var_4539, x = write_indices_float_65_cast_fp16_to_int32)[name = tensor("op_4540")]; + tensor write_indices_21_reps_0 = const()[name = tensor("write_indices_21_reps_0"), val = tensor([1, 1, 16, 64])]; + tensor write_indices_21 = tile(reps = write_indices_21_reps_0, x = var_4540)[name = tensor("write_indices_21")]; + tensor var_4550_begin_0 = const()[name = tensor("op_4550_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_4550_end_0 = const()[name = tensor("op_4550_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor var_4550_end_mask_0 = const()[name = tensor("op_4550_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_4550_squeeze_mask_0 = const()[name = tensor("op_4550_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor cache10_to_fp16_dtype_0 = const()[name = tensor("cache10_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor cache10_to_fp16 = cast(dtype = cache10_to_fp16_dtype_0, x = cache10)[name = tensor("cast_609")]; + tensor var_4550_cast_fp16 = slice_by_index(begin = var_4550_begin_0, end = var_4550_end_0, end_mask = var_4550_end_mask_0, squeeze_mask = var_4550_squeeze_mask_0, x = cache10_to_fp16)[name = tensor("op_4550_cast_fp16")]; + tensor var_4552_axis_0 = const()[name = tensor("op_4552_axis_0"), val = tensor(1)]; + tensor var_4552_mode_0 = const()[name = tensor("op_4552_mode_0"), val = tensor("update")]; + tensor var_4552_validate_indices_0 = const()[name = tensor("op_4552_validate_indices_0"), val = tensor(false)]; + tensor var_4552_cast_fp16 = scatter_along_axis(axis = var_4552_axis_0, data = var_4550_cast_fp16, indices = write_indices_21, mode = var_4552_mode_0, updates = k_43_cast_fp16, validate_indices = var_4552_validate_indices_0)[name = tensor("op_4552_cast_fp16")]; + tensor concat_82 = const()[name = tensor("concat_82"), val = tensor([0, 0, 0, 0, 0])]; + tensor concat_83 = const()[name = tensor("concat_83"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_21_internal_tensor_assign_1_stride_0 = const()[name = tensor("new_cache_21_internal_tensor_assign_1_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_21_internal_tensor_assign_1_begin_mask_0 = const()[name = tensor("new_cache_21_internal_tensor_assign_1_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_21_internal_tensor_assign_1_end_mask_0 = const()[name = tensor("new_cache_21_internal_tensor_assign_1_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_21_internal_tensor_assign_1_squeeze_mask_0 = const()[name = tensor("new_cache_21_internal_tensor_assign_1_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_66 = const()[name = tensor("shape_66"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_20 = const()[name = tensor("reduce_prod_20"), val = tensor(1048576)]; + tensor range_1d_20_start_0 = const()[name = tensor("range_1d_20_start_0"), val = tensor(0)]; + tensor range_1d_20_step_0 = const()[name = tensor("range_1d_20_step_0"), val = tensor(1)]; + tensor range_1d_20 = range_1d(end = reduce_prod_20, start = range_1d_20_start_0, step = range_1d_20_step_0)[name = tensor("range_1d_20")]; + tensor reshape_100 = reshape(shape = shape_66, x = range_1d_20)[name = tensor("reshape_100")]; + tensor slice_by_index_20 = slice_by_index(begin = concat_82, begin_mask = new_cache_21_internal_tensor_assign_1_begin_mask_0, end = concat_83, end_mask = new_cache_21_internal_tensor_assign_1_end_mask_0, squeeze_mask = new_cache_21_internal_tensor_assign_1_squeeze_mask_0, stride = new_cache_21_internal_tensor_assign_1_stride_0, x = reshape_100)[name = tensor("slice_by_index_20")]; + tensor reshape_101_shape_0 = const()[name = tensor("reshape_101_shape_0"), val = tensor([-1])]; + tensor reshape_101 = reshape(shape = reshape_101_shape_0, x = slice_by_index_20)[name = tensor("reshape_101")]; + tensor reshape_102_shape_0 = const()[name = tensor("reshape_102_shape_0"), val = tensor([-1])]; + tensor reshape_102_cast_fp16 = reshape(shape = reshape_102_shape_0, x = var_4552_cast_fp16)[name = tensor("reshape_102_cast_fp16")]; + tensor reshape_103_shape_0 = const()[name = tensor("reshape_103_shape_0"), val = tensor([-1])]; + tensor reshape_103_cast_fp16 = reshape(shape = reshape_103_shape_0, x = cache10_to_fp16)[name = tensor("reshape_103_cast_fp16")]; + tensor scatter_20_mode_0 = const()[name = tensor("scatter_20_mode_0"), val = tensor("update")]; + tensor scatter_20_axis_0 = const()[name = tensor("scatter_20_axis_0"), val = tensor(0)]; + tensor scatter_20_validate_indices_0 = const()[name = tensor("scatter_20_validate_indices_0"), val = tensor(false)]; + tensor scatter_20_cast_fp16 = scatter(axis = scatter_20_axis_0, data = reshape_103_cast_fp16, indices = reshape_101, mode = scatter_20_mode_0, updates = reshape_102_cast_fp16, validate_indices = scatter_20_validate_indices_0)[name = tensor("scatter_20_cast_fp16")]; + tensor reshape_104_cast_fp16 = reshape(shape = shape_66, x = scatter_20_cast_fp16)[name = tensor("reshape_104_cast_fp16")]; + tensor var_4560_begin_0 = const()[name = tensor("op_4560_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor var_4560_end_0 = const()[name = tensor("op_4560_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor var_4560_end_mask_0 = const()[name = tensor("op_4560_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_4560_squeeze_mask_0 = const()[name = tensor("op_4560_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor var_4560_cast_fp16 = slice_by_index(begin = var_4560_begin_0, end = var_4560_end_0, end_mask = var_4560_end_mask_0, squeeze_mask = var_4560_squeeze_mask_0, x = reshape_104_cast_fp16)[name = tensor("op_4560_cast_fp16")]; + tensor var_4562_axis_0 = const()[name = tensor("op_4562_axis_0"), val = tensor(1)]; + tensor var_4562_mode_0 = const()[name = tensor("op_4562_mode_0"), val = tensor("update")]; + tensor var_4562_validate_indices_0 = const()[name = tensor("op_4562_validate_indices_0"), val = tensor(false)]; + tensor var_4562_cast_fp16 = scatter_along_axis(axis = var_4562_axis_0, data = var_4560_cast_fp16, indices = write_indices_21, mode = var_4562_mode_0, updates = v_21_cast_fp16, validate_indices = var_4562_validate_indices_0)[name = tensor("op_4562_cast_fp16")]; + tensor concat_84 = const()[name = tensor("concat_84"), val = tensor([1, 0, 0, 0, 0])]; + tensor concat_85 = const()[name = tensor("concat_85"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_21_internal_tensor_assign_2_stride_0 = const()[name = tensor("new_cache_21_internal_tensor_assign_2_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_21_internal_tensor_assign_2_begin_mask_0 = const()[name = tensor("new_cache_21_internal_tensor_assign_2_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_21_internal_tensor_assign_2_end_mask_0 = const()[name = tensor("new_cache_21_internal_tensor_assign_2_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_21_internal_tensor_assign_2_squeeze_mask_0 = const()[name = tensor("new_cache_21_internal_tensor_assign_2_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_67 = const()[name = tensor("shape_67"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_21 = const()[name = tensor("reduce_prod_21"), val = tensor(1048576)]; + tensor range_1d_21_start_0 = const()[name = tensor("range_1d_21_start_0"), val = tensor(0)]; + tensor range_1d_21_step_0 = const()[name = tensor("range_1d_21_step_0"), val = tensor(1)]; + tensor range_1d_21 = range_1d(end = reduce_prod_21, start = range_1d_21_start_0, step = range_1d_21_step_0)[name = tensor("range_1d_21")]; + tensor reshape_105 = reshape(shape = shape_67, x = range_1d_21)[name = tensor("reshape_105")]; + tensor slice_by_index_21 = slice_by_index(begin = concat_84, begin_mask = new_cache_21_internal_tensor_assign_2_begin_mask_0, end = concat_85, end_mask = new_cache_21_internal_tensor_assign_2_end_mask_0, squeeze_mask = new_cache_21_internal_tensor_assign_2_squeeze_mask_0, stride = new_cache_21_internal_tensor_assign_2_stride_0, x = reshape_105)[name = tensor("slice_by_index_21")]; + tensor reshape_106_shape_0 = const()[name = tensor("reshape_106_shape_0"), val = tensor([-1])]; + tensor reshape_106 = reshape(shape = reshape_106_shape_0, x = slice_by_index_21)[name = tensor("reshape_106")]; + tensor reshape_107_shape_0 = const()[name = tensor("reshape_107_shape_0"), val = tensor([-1])]; + tensor reshape_107_cast_fp16 = reshape(shape = reshape_107_shape_0, x = var_4562_cast_fp16)[name = tensor("reshape_107_cast_fp16")]; + tensor reshape_108_shape_0 = const()[name = tensor("reshape_108_shape_0"), val = tensor([-1])]; + tensor reshape_108_cast_fp16 = reshape(shape = reshape_108_shape_0, x = reshape_104_cast_fp16)[name = tensor("reshape_108_cast_fp16")]; + tensor scatter_21_mode_0 = const()[name = tensor("scatter_21_mode_0"), val = tensor("update")]; + tensor scatter_21_axis_0 = const()[name = tensor("scatter_21_axis_0"), val = tensor(0)]; + tensor scatter_21_validate_indices_0 = const()[name = tensor("scatter_21_validate_indices_0"), val = tensor(false)]; + tensor scatter_21_cast_fp16 = scatter(axis = scatter_21_axis_0, data = reshape_108_cast_fp16, indices = reshape_106, mode = scatter_21_mode_0, updates = reshape_107_cast_fp16, validate_indices = scatter_21_validate_indices_0)[name = tensor("scatter_21_cast_fp16")]; + tensor reshape_109_cast_fp16 = reshape(shape = shape_67, x = scatter_21_cast_fp16)[name = tensor("reshape_109_cast_fp16")]; + tensor reshape_109_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("reshape_109_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor keys_61_begin_0 = const()[name = tensor("keys_61_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor keys_61_end_0 = const()[name = tensor("keys_61_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor keys_61_end_mask_0 = const()[name = tensor("keys_61_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor keys_61_squeeze_mask_0 = const()[name = tensor("keys_61_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor keys_61_cast_fp16 = slice_by_index(begin = keys_61_begin_0, end = keys_61_end_0, end_mask = keys_61_end_mask_0, squeeze_mask = keys_61_squeeze_mask_0, x = reshape_109_cast_fp16)[name = tensor("keys_61_cast_fp16")]; + tensor values_61_begin_0 = const()[name = tensor("values_61_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor values_61_end_0 = const()[name = tensor("values_61_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor values_61_end_mask_0 = const()[name = tensor("values_61_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor values_61_squeeze_mask_0 = const()[name = tensor("values_61_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor values_61_cast_fp16 = slice_by_index(begin = values_61_begin_0, end = values_61_end_0, end_mask = values_61_end_mask_0, squeeze_mask = values_61_squeeze_mask_0, x = reshape_109_cast_fp16)[name = tensor("values_61_cast_fp16")]; + tensor var_4574_cast_fp16 = not_equal(x = keys_61_cast_fp16, y = keys_61_cast_fp16)[name = tensor("op_4574_cast_fp16")]; + tensor keys_63_cast_fp16 = select(a = var_470_to_fp16, b = keys_61_cast_fp16, cond = var_4574_cast_fp16)[name = tensor("keys_63_cast_fp16")]; + tensor var_4582_cast_fp16 = not_equal(x = values_61_cast_fp16, y = values_61_cast_fp16)[name = tensor("op_4582_cast_fp16")]; + tensor values_63_cast_fp16 = select(a = var_470_to_fp16, b = values_61_cast_fp16, cond = var_4582_cast_fp16)[name = tensor("values_63_cast_fp16")]; + tensor var_4606 = const()[name = tensor("op_4606"), val = tensor([0, 2, 1, 3])]; + tensor var_4622 = const()[name = tensor("op_4622"), val = tensor([1, 1, 1])]; + tensor var_4623_cast_fp16 = reshape(shape = var_4622, x = position10_to_fp16)[name = tensor("op_4623_cast_fp16")]; + tensor q_positions_21_cast_fp16 = add(x = var_4623_cast_fp16, y = q_offsets_1_promoted_to_fp16)[name = tensor("q_positions_21_cast_fp16")]; + tensor valid_end_21_cast_fp16 = add(x = var_4623_cast_fp16, y = var_536_cast_fp16)[name = tensor("valid_end_21_cast_fp16")]; + tensor valid_mask_21_cast_fp16 = less(x = k_positions_1_promoted_to_fp16, y = valid_end_21_cast_fp16)[name = tensor("valid_mask_21_cast_fp16")]; + tensor causal_mask_21_cast_fp16 = less_equal(x = k_positions_1_promoted_to_fp16, y = q_positions_21_cast_fp16)[name = tensor("causal_mask_21_cast_fp16")]; + tensor var_4651 = logical_and(x = valid_mask_21_cast_fp16, y = causal_mask_21_cast_fp16)[name = tensor("op_4651")]; + tensor attn_mask_21_axes_0 = const()[name = tensor("attn_mask_21_axes_0"), val = tensor([1])]; + tensor attn_mask_21 = expand_dims(axes = attn_mask_21_axes_0, x = var_4651)[name = tensor("attn_mask_21")]; + tensor var_4665_transpose_x_0 = const()[name = tensor("op_4665_transpose_x_0"), val = tensor(false)]; + tensor var_4665_transpose_y_0 = const()[name = tensor("op_4665_transpose_y_0"), val = tensor(false)]; + tensor transpose_89_perm_0 = const()[name = tensor("transpose_89_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_90_perm_0 = const()[name = tensor("transpose_90_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_90 = transpose(perm = transpose_90_perm_0, x = keys_63_cast_fp16)[name = tensor("transpose_164")]; + tensor transpose_89 = transpose(perm = transpose_89_perm_0, x = q_63_cast_fp16)[name = tensor("transpose_165")]; + tensor var_4665_cast_fp16 = matmul(transpose_x = var_4665_transpose_x_0, transpose_y = var_4665_transpose_y_0, x = transpose_89, y = transpose_90)[name = tensor("op_4665_cast_fp16")]; + tensor var_4659_to_fp16 = const()[name = tensor("op_4659_to_fp16"), val = tensor([0x1p-3])]; + tensor attn_weights_61_cast_fp16 = mul(x = var_4665_cast_fp16, y = var_4659_to_fp16)[name = tensor("attn_weights_61_cast_fp16")]; + tensor var_4667 = logical_not(x = attn_mask_21)[name = tensor("op_4667")]; + tensor var_4668_to_fp16 = const()[name = tensor("op_4668_to_fp16"), val = tensor(-inf)]; + tensor attn_weights_63_cast_fp16 = select(a = var_4668_to_fp16, b = attn_weights_61_cast_fp16, cond = var_4667)[name = tensor("attn_weights_63_cast_fp16")]; + tensor var_4670 = const()[name = tensor("op_4670"), val = tensor(-1)]; + tensor attn_weights_65_cast_fp16 = softmax(axis = var_4670, x = attn_weights_63_cast_fp16)[name = tensor("attn_weights_65_cast_fp16")]; + tensor attn_output_21_transpose_x_0 = const()[name = tensor("attn_output_21_transpose_x_0"), val = tensor(false)]; + tensor attn_output_21_transpose_y_0 = const()[name = tensor("attn_output_21_transpose_y_0"), val = tensor(false)]; + tensor values_65_cast_fp16 = transpose(perm = var_4606, x = values_63_cast_fp16)[name = tensor("transpose_166")]; + tensor attn_output_21_cast_fp16 = matmul(transpose_x = attn_output_21_transpose_x_0, transpose_y = attn_output_21_transpose_y_0, x = attn_weights_65_cast_fp16, y = values_65_cast_fp16)[name = tensor("attn_output_21_cast_fp16")]; + tensor var_4678 = const()[name = tensor("op_4678"), val = tensor([0, 2, 1, 3])]; + tensor var_4683 = const()[name = tensor("op_4683"), val = tensor([1, 256, 1024])]; + tensor var_4679_cast_fp16 = transpose(perm = var_4678, x = attn_output_21_cast_fp16)[name = tensor("transpose_163")]; + tensor input_125_cast_fp16 = reshape(shape = var_4683, x = var_4679_cast_fp16)[name = tensor("input_125_cast_fp16")]; + tensor attn10_out_proj_weight_to_fp16 = const()[name = tensor("attn10_out_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(259110528)))]; + tensor linear_41_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = attn10_out_proj_weight_to_fp16, x = input_125_cast_fp16)[name = tensor("linear_41_cast_fp16")]; + tensor var_4692_cast_fp16 = add(x = position10_to_fp16, y = valid_len_to_fp16)[name = tensor("op_4692_cast_fp16")]; + tensor var_4692_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_4692_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor input_127_cast_fp16 = add(x = input_121_cast_fp16, y = linear_41_cast_fp16)[name = tensor("input_127_cast_fp16")]; + tensor input_129_axes_0 = const()[name = tensor("input_129_axes_0"), val = tensor([-1])]; + tensor norm10_2_weight_to_fp16 = const()[name = tensor("norm10_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(261207744)))]; + tensor norm10_2_bias_to_fp16 = const()[name = tensor("norm10_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(261209856)))]; + tensor var_4696_to_fp16 = const()[name = tensor("op_4696_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_129_cast_fp16 = layer_norm(axes = input_129_axes_0, beta = norm10_2_bias_to_fp16, epsilon = var_4696_to_fp16, gamma = norm10_2_weight_to_fp16, x = input_127_cast_fp16)[name = tensor("input_129_cast_fp16")]; + tensor linear10_1_weight_to_fp16 = const()[name = tensor("linear10_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(261211968)))]; + tensor linear_42_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = linear10_1_weight_to_fp16, x = input_129_cast_fp16)[name = tensor("linear_42_cast_fp16")]; + tensor input_131_mode_0 = const()[name = tensor("input_131_mode_0"), val = tensor("EXACT")]; + tensor input_131_cast_fp16 = gelu(mode = input_131_mode_0, x = linear_42_cast_fp16)[name = tensor("input_131_cast_fp16")]; + tensor linear10_2_weight_to_fp16 = const()[name = tensor("linear10_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(269600640)))]; + tensor linear_43_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = linear10_2_weight_to_fp16, x = input_131_cast_fp16)[name = tensor("linear_43_cast_fp16")]; + tensor input_133_cast_fp16 = add(x = input_127_cast_fp16, y = linear_43_cast_fp16)[name = tensor("input_133_cast_fp16")]; + tensor input_135_axes_0 = const()[name = tensor("input_135_axes_0"), val = tensor([-1])]; + tensor norm11_1_weight_to_fp16 = const()[name = tensor("norm11_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(277989312)))]; + tensor norm11_1_bias_to_fp16 = const()[name = tensor("norm11_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(277991424)))]; + tensor var_4713_to_fp16 = const()[name = tensor("op_4713_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_135_cast_fp16 = layer_norm(axes = input_135_axes_0, beta = norm11_1_bias_to_fp16, epsilon = var_4713_to_fp16, gamma = norm11_1_weight_to_fp16, x = input_133_cast_fp16)[name = tensor("input_135_cast_fp16")]; + tensor attn11_in_proj_weight_to_fp16 = const()[name = tensor("attn11_in_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(277993536)))]; + tensor linear_44_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = attn11_in_proj_weight_to_fp16, x = input_135_cast_fp16)[name = tensor("linear_44_cast_fp16")]; + tensor var_4727 = const()[name = tensor("op_4727"), val = tensor([1, 256, 3, 16, 64])]; + tensor qkv_23_cast_fp16 = reshape(shape = var_4727, x = linear_44_cast_fp16)[name = tensor("qkv_23_cast_fp16")]; + tensor q_67_begin_0 = const()[name = tensor("q_67_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor q_67_end_0 = const()[name = tensor("q_67_end_0"), val = tensor([1, 256, 1, 16, 64])]; + tensor q_67_end_mask_0 = const()[name = tensor("q_67_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor q_67_squeeze_mask_0 = const()[name = tensor("q_67_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor q_67_cast_fp16 = slice_by_index(begin = q_67_begin_0, end = q_67_end_0, end_mask = q_67_end_mask_0, squeeze_mask = q_67_squeeze_mask_0, x = qkv_23_cast_fp16)[name = tensor("q_67_cast_fp16")]; + tensor k_45_begin_0 = const()[name = tensor("k_45_begin_0"), val = tensor([0, 0, 1, 0, 0])]; + tensor k_45_end_0 = const()[name = tensor("k_45_end_0"), val = tensor([1, 256, 2, 16, 64])]; + tensor k_45_end_mask_0 = const()[name = tensor("k_45_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor k_45_squeeze_mask_0 = const()[name = tensor("k_45_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor k_45_cast_fp16 = slice_by_index(begin = k_45_begin_0, end = k_45_end_0, end_mask = k_45_end_mask_0, squeeze_mask = k_45_squeeze_mask_0, x = qkv_23_cast_fp16)[name = tensor("k_45_cast_fp16")]; + tensor v_23_begin_0 = const()[name = tensor("v_23_begin_0"), val = tensor([0, 0, 2, 0, 0])]; + tensor v_23_end_0 = const()[name = tensor("v_23_end_0"), val = tensor([1, 256, 3, 16, 64])]; + tensor v_23_end_mask_0 = const()[name = tensor("v_23_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor v_23_squeeze_mask_0 = const()[name = tensor("v_23_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor v_23_cast_fp16 = slice_by_index(begin = v_23_begin_0, end = v_23_end_0, end_mask = v_23_end_mask_0, squeeze_mask = v_23_squeeze_mask_0, x = qkv_23_cast_fp16)[name = tensor("v_23_cast_fp16")]; + tensor var_4785 = const()[name = tensor("op_4785"), val = tensor([1, 1])]; + tensor position11_to_fp16_dtype_0 = const()[name = tensor("position11_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor position11_to_fp16 = cast(dtype = position11_to_fp16_dtype_0, x = position11)[name = tensor("cast_606")]; + tensor var_4786_cast_fp16 = reshape(shape = var_4785, x = position11_to_fp16)[name = tensor("op_4786_cast_fp16")]; + tensor ts_69_cast_fp16 = add(x = ts_1_promoted_to_fp16, y = var_4786_cast_fp16)[name = tensor("ts_69_cast_fp16")]; + tensor var_4793 = const()[name = tensor("op_4793"), val = tensor([1, 256, 1, 1])]; + tensor ts_71_cast_fp16 = reshape(shape = var_4793, x = ts_69_cast_fp16)[name = tensor("ts_71_cast_fp16")]; + tensor var_4800 = const()[name = tensor("op_4800"), val = tensor([1, 256, 16, 32, 2])]; + tensor q_complex_23_cast_fp16 = reshape(shape = var_4800, x = q_67_cast_fp16)[name = tensor("q_complex_23_cast_fp16")]; + tensor var_4807 = const()[name = tensor("op_4807"), val = tensor([1, 256, 16, 32, 2])]; + tensor k_complex_23_cast_fp16 = reshape(shape = var_4807, x = k_45_cast_fp16)[name = tensor("k_complex_23_cast_fp16")]; + tensor var_4811_begin_0 = const()[name = tensor("op_4811_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_4811_end_0 = const()[name = tensor("op_4811_end_0"), val = tensor([1, 256, 16, 32, 1])]; + tensor var_4811_end_mask_0 = const()[name = tensor("op_4811_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_4811_squeeze_mask_0 = const()[name = tensor("op_4811_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_4811_cast_fp16 = slice_by_index(begin = var_4811_begin_0, end = var_4811_end_0, end_mask = var_4811_end_mask_0, squeeze_mask = var_4811_squeeze_mask_0, x = q_complex_23_cast_fp16)[name = tensor("op_4811_cast_fp16")]; + tensor var_4819_begin_0 = const()[name = tensor("op_4819_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_4819_end_0 = const()[name = tensor("op_4819_end_0"), val = tensor([1, 256, 16, 32, 2])]; + tensor var_4819_end_mask_0 = const()[name = tensor("op_4819_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_4819_squeeze_mask_0 = const()[name = tensor("op_4819_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_4819_cast_fp16 = slice_by_index(begin = var_4819_begin_0, end = var_4819_end_0, end_mask = var_4819_end_mask_0, squeeze_mask = var_4819_squeeze_mask_0, x = q_complex_23_cast_fp16)[name = tensor("op_4819_cast_fp16")]; + tensor var_4827_begin_0 = const()[name = tensor("op_4827_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_4827_end_0 = const()[name = tensor("op_4827_end_0"), val = tensor([1, 256, 16, 32, 1])]; + tensor var_4827_end_mask_0 = const()[name = tensor("op_4827_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_4827_squeeze_mask_0 = const()[name = tensor("op_4827_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_4827_cast_fp16 = slice_by_index(begin = var_4827_begin_0, end = var_4827_end_0, end_mask = var_4827_end_mask_0, squeeze_mask = var_4827_squeeze_mask_0, x = k_complex_23_cast_fp16)[name = tensor("op_4827_cast_fp16")]; + tensor var_4835_begin_0 = const()[name = tensor("op_4835_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_4835_end_0 = const()[name = tensor("op_4835_end_0"), val = tensor([1, 256, 16, 32, 2])]; + tensor var_4835_end_mask_0 = const()[name = tensor("op_4835_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_4835_squeeze_mask_0 = const()[name = tensor("op_4835_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_4835_cast_fp16 = slice_by_index(begin = var_4835_begin_0, end = var_4835_end_0, end_mask = var_4835_end_mask_0, squeeze_mask = var_4835_squeeze_mask_0, x = k_complex_23_cast_fp16)[name = tensor("op_4835_cast_fp16")]; + tensor freqs_23_to_fp16 = const()[name = tensor("freqs_23_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(284285056)))]; + tensor var_4841_cast_fp16 = mul(x = freqs_23_to_fp16, y = ts_71_cast_fp16)[name = tensor("op_4841_cast_fp16")]; + tensor rotr_23_cast_fp16 = cos(x = var_4841_cast_fp16)[name = tensor("rotr_23_cast_fp16")]; + tensor roti_23_cast_fp16 = sin(x = var_4841_cast_fp16)[name = tensor("roti_23_cast_fp16")]; + tensor var_4845_cast_fp16 = mul(x = var_4811_cast_fp16, y = rotr_23_cast_fp16)[name = tensor("op_4845_cast_fp16")]; + tensor var_4846_cast_fp16 = mul(x = var_4819_cast_fp16, y = roti_23_cast_fp16)[name = tensor("op_4846_cast_fp16")]; + tensor qor_45_cast_fp16 = sub(x = var_4845_cast_fp16, y = var_4846_cast_fp16)[name = tensor("qor_45_cast_fp16")]; + tensor var_4849_cast_fp16 = mul(x = var_4811_cast_fp16, y = roti_23_cast_fp16)[name = tensor("op_4849_cast_fp16")]; + tensor var_4850_cast_fp16 = mul(x = var_4819_cast_fp16, y = rotr_23_cast_fp16)[name = tensor("op_4850_cast_fp16")]; + tensor qoi_45_cast_fp16 = add(x = var_4849_cast_fp16, y = var_4850_cast_fp16)[name = tensor("qoi_45_cast_fp16")]; + tensor var_4853_cast_fp16 = mul(x = var_4827_cast_fp16, y = rotr_23_cast_fp16)[name = tensor("op_4853_cast_fp16")]; + tensor var_4854_cast_fp16 = mul(x = var_4835_cast_fp16, y = roti_23_cast_fp16)[name = tensor("op_4854_cast_fp16")]; + tensor kor_45_cast_fp16 = sub(x = var_4853_cast_fp16, y = var_4854_cast_fp16)[name = tensor("kor_45_cast_fp16")]; + tensor var_4857_cast_fp16 = mul(x = var_4827_cast_fp16, y = roti_23_cast_fp16)[name = tensor("op_4857_cast_fp16")]; + tensor var_4858_cast_fp16 = mul(x = var_4835_cast_fp16, y = rotr_23_cast_fp16)[name = tensor("op_4858_cast_fp16")]; + tensor koi_45_cast_fp16 = add(x = var_4857_cast_fp16, y = var_4858_cast_fp16)[name = tensor("koi_45_cast_fp16")]; + tensor qo_23_axis_0 = const()[name = tensor("qo_23_axis_0"), val = tensor(-1)]; + tensor qo_23_cast_fp16 = stack(axis = qo_23_axis_0, values = (qor_45_cast_fp16, qoi_45_cast_fp16))[name = tensor("qo_23_cast_fp16")]; + tensor ko_23_axis_0 = const()[name = tensor("ko_23_axis_0"), val = tensor(-1)]; + tensor ko_23_cast_fp16 = stack(axis = ko_23_axis_0, values = (kor_45_cast_fp16, koi_45_cast_fp16))[name = tensor("ko_23_cast_fp16")]; + tensor var_4891 = const()[name = tensor("op_4891"), val = tensor([1, 256, 16, 64])]; + tensor q_69_cast_fp16 = reshape(shape = var_4891, x = qo_23_cast_fp16)[name = tensor("q_69_cast_fp16")]; + tensor var_4897 = const()[name = tensor("op_4897"), val = tensor([1, 256, 16, 64])]; + tensor k_47_cast_fp16 = reshape(shape = var_4897, x = ko_23_cast_fp16)[name = tensor("k_47_cast_fp16")]; + tensor write_indices_float_67_cast_fp16 = add(x = var_4786_cast_fp16, y = var_393_promoted_to_fp16)[name = tensor("write_indices_float_67_cast_fp16")]; + tensor _inversed_4918_y_0_to_fp16 = const()[name = tensor("_inversed_4918_y_0_to_fp16"), val = tensor(0x1p-9)]; + tensor _inversed_4918_cast_fp16 = mul(x = write_indices_float_67_cast_fp16, y = _inversed_4918_y_0_to_fp16)[name = tensor("_inversed_4918_cast_fp16")]; + tensor var_4919_cast_fp16 = floor(x = _inversed_4918_cast_fp16)[name = tensor("op_4919_cast_fp16")]; + tensor var_4920_to_fp16 = const()[name = tensor("op_4920_to_fp16"), val = tensor(0x1p+9)]; + tensor var_4921_cast_fp16 = mul(x = var_4919_cast_fp16, y = var_4920_to_fp16)[name = tensor("op_4921_cast_fp16")]; + tensor write_indices_float_69_cast_fp16 = sub(x = write_indices_float_67_cast_fp16, y = var_4921_cast_fp16)[name = tensor("write_indices_float_69_cast_fp16")]; + tensor var_4939_value_0_to_fp16 = const()[name = tensor("op_4939_value_0_to_fp16"), val = tensor(0x1.ffp+8)]; + tensor var_4939_cast_fp16 = fill_like(ref_tensor = write_indices_float_69_cast_fp16, value = var_4939_value_0_to_fp16)[name = tensor("op_4939_cast_fp16")]; + tensor write_indices_float_71_cast_fp16 = select(a = write_indices_float_69_cast_fp16, b = var_4939_cast_fp16, cond = token_valid_1_cast_fp16)[name = tensor("write_indices_float_71_cast_fp16")]; + tensor var_4945_dtype_0 = const()[name = tensor("op_4945_dtype_0"), val = tensor("int32")]; + tensor var_4950 = const()[name = tensor("op_4950"), val = tensor([1, 256, 1, 1])]; + tensor write_indices_float_71_cast_fp16_to_int32 = cast(dtype = var_4945_dtype_0, x = write_indices_float_71_cast_fp16)[name = tensor("cast_605")]; + tensor var_4951 = reshape(shape = var_4950, x = write_indices_float_71_cast_fp16_to_int32)[name = tensor("op_4951")]; + tensor write_indices_23_reps_0 = const()[name = tensor("write_indices_23_reps_0"), val = tensor([1, 1, 16, 64])]; + tensor write_indices_23 = tile(reps = write_indices_23_reps_0, x = var_4951)[name = tensor("write_indices_23")]; + tensor var_4961_begin_0 = const()[name = tensor("op_4961_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_4961_end_0 = const()[name = tensor("op_4961_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor var_4961_end_mask_0 = const()[name = tensor("op_4961_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_4961_squeeze_mask_0 = const()[name = tensor("op_4961_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor cache11_to_fp16_dtype_0 = const()[name = tensor("cache11_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor cache11_to_fp16 = cast(dtype = cache11_to_fp16_dtype_0, x = cache11)[name = tensor("cast_604")]; + tensor var_4961_cast_fp16 = slice_by_index(begin = var_4961_begin_0, end = var_4961_end_0, end_mask = var_4961_end_mask_0, squeeze_mask = var_4961_squeeze_mask_0, x = cache11_to_fp16)[name = tensor("op_4961_cast_fp16")]; + tensor var_4963_axis_0 = const()[name = tensor("op_4963_axis_0"), val = tensor(1)]; + tensor var_4963_mode_0 = const()[name = tensor("op_4963_mode_0"), val = tensor("update")]; + tensor var_4963_validate_indices_0 = const()[name = tensor("op_4963_validate_indices_0"), val = tensor(false)]; + tensor var_4963_cast_fp16 = scatter_along_axis(axis = var_4963_axis_0, data = var_4961_cast_fp16, indices = write_indices_23, mode = var_4963_mode_0, updates = k_47_cast_fp16, validate_indices = var_4963_validate_indices_0)[name = tensor("op_4963_cast_fp16")]; + tensor concat_90 = const()[name = tensor("concat_90"), val = tensor([0, 0, 0, 0, 0])]; + tensor concat_91 = const()[name = tensor("concat_91"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_23_internal_tensor_assign_1_stride_0 = const()[name = tensor("new_cache_23_internal_tensor_assign_1_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_23_internal_tensor_assign_1_begin_mask_0 = const()[name = tensor("new_cache_23_internal_tensor_assign_1_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_23_internal_tensor_assign_1_end_mask_0 = const()[name = tensor("new_cache_23_internal_tensor_assign_1_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_23_internal_tensor_assign_1_squeeze_mask_0 = const()[name = tensor("new_cache_23_internal_tensor_assign_1_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_68 = const()[name = tensor("shape_68"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_22 = const()[name = tensor("reduce_prod_22"), val = tensor(1048576)]; + tensor range_1d_22_start_0 = const()[name = tensor("range_1d_22_start_0"), val = tensor(0)]; + tensor range_1d_22_step_0 = const()[name = tensor("range_1d_22_step_0"), val = tensor(1)]; + tensor range_1d_22 = range_1d(end = reduce_prod_22, start = range_1d_22_start_0, step = range_1d_22_step_0)[name = tensor("range_1d_22")]; + tensor reshape_110 = reshape(shape = shape_68, x = range_1d_22)[name = tensor("reshape_110")]; + tensor slice_by_index_22 = slice_by_index(begin = concat_90, begin_mask = new_cache_23_internal_tensor_assign_1_begin_mask_0, end = concat_91, end_mask = new_cache_23_internal_tensor_assign_1_end_mask_0, squeeze_mask = new_cache_23_internal_tensor_assign_1_squeeze_mask_0, stride = new_cache_23_internal_tensor_assign_1_stride_0, x = reshape_110)[name = tensor("slice_by_index_22")]; + tensor reshape_111_shape_0 = const()[name = tensor("reshape_111_shape_0"), val = tensor([-1])]; + tensor reshape_111 = reshape(shape = reshape_111_shape_0, x = slice_by_index_22)[name = tensor("reshape_111")]; + tensor reshape_112_shape_0 = const()[name = tensor("reshape_112_shape_0"), val = tensor([-1])]; + tensor reshape_112_cast_fp16 = reshape(shape = reshape_112_shape_0, x = var_4963_cast_fp16)[name = tensor("reshape_112_cast_fp16")]; + tensor reshape_113_shape_0 = const()[name = tensor("reshape_113_shape_0"), val = tensor([-1])]; + tensor reshape_113_cast_fp16 = reshape(shape = reshape_113_shape_0, x = cache11_to_fp16)[name = tensor("reshape_113_cast_fp16")]; + tensor scatter_22_mode_0 = const()[name = tensor("scatter_22_mode_0"), val = tensor("update")]; + tensor scatter_22_axis_0 = const()[name = tensor("scatter_22_axis_0"), val = tensor(0)]; + tensor scatter_22_validate_indices_0 = const()[name = tensor("scatter_22_validate_indices_0"), val = tensor(false)]; + tensor scatter_22_cast_fp16 = scatter(axis = scatter_22_axis_0, data = reshape_113_cast_fp16, indices = reshape_111, mode = scatter_22_mode_0, updates = reshape_112_cast_fp16, validate_indices = scatter_22_validate_indices_0)[name = tensor("scatter_22_cast_fp16")]; + tensor reshape_114_cast_fp16 = reshape(shape = shape_68, x = scatter_22_cast_fp16)[name = tensor("reshape_114_cast_fp16")]; + tensor var_4971_begin_0 = const()[name = tensor("op_4971_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor var_4971_end_0 = const()[name = tensor("op_4971_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor var_4971_end_mask_0 = const()[name = tensor("op_4971_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_4971_squeeze_mask_0 = const()[name = tensor("op_4971_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor var_4971_cast_fp16 = slice_by_index(begin = var_4971_begin_0, end = var_4971_end_0, end_mask = var_4971_end_mask_0, squeeze_mask = var_4971_squeeze_mask_0, x = reshape_114_cast_fp16)[name = tensor("op_4971_cast_fp16")]; + tensor var_4973_axis_0 = const()[name = tensor("op_4973_axis_0"), val = tensor(1)]; + tensor var_4973_mode_0 = const()[name = tensor("op_4973_mode_0"), val = tensor("update")]; + tensor var_4973_validate_indices_0 = const()[name = tensor("op_4973_validate_indices_0"), val = tensor(false)]; + tensor var_4973_cast_fp16 = scatter_along_axis(axis = var_4973_axis_0, data = var_4971_cast_fp16, indices = write_indices_23, mode = var_4973_mode_0, updates = v_23_cast_fp16, validate_indices = var_4973_validate_indices_0)[name = tensor("op_4973_cast_fp16")]; + tensor concat_92 = const()[name = tensor("concat_92"), val = tensor([1, 0, 0, 0, 0])]; + tensor concat_93 = const()[name = tensor("concat_93"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_23_internal_tensor_assign_2_stride_0 = const()[name = tensor("new_cache_23_internal_tensor_assign_2_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_23_internal_tensor_assign_2_begin_mask_0 = const()[name = tensor("new_cache_23_internal_tensor_assign_2_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_23_internal_tensor_assign_2_end_mask_0 = const()[name = tensor("new_cache_23_internal_tensor_assign_2_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_23_internal_tensor_assign_2_squeeze_mask_0 = const()[name = tensor("new_cache_23_internal_tensor_assign_2_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_69 = const()[name = tensor("shape_69"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_23 = const()[name = tensor("reduce_prod_23"), val = tensor(1048576)]; + tensor range_1d_23_start_0 = const()[name = tensor("range_1d_23_start_0"), val = tensor(0)]; + tensor range_1d_23_step_0 = const()[name = tensor("range_1d_23_step_0"), val = tensor(1)]; + tensor range_1d_23 = range_1d(end = reduce_prod_23, start = range_1d_23_start_0, step = range_1d_23_step_0)[name = tensor("range_1d_23")]; + tensor reshape_115 = reshape(shape = shape_69, x = range_1d_23)[name = tensor("reshape_115")]; + tensor slice_by_index_23 = slice_by_index(begin = concat_92, begin_mask = new_cache_23_internal_tensor_assign_2_begin_mask_0, end = concat_93, end_mask = new_cache_23_internal_tensor_assign_2_end_mask_0, squeeze_mask = new_cache_23_internal_tensor_assign_2_squeeze_mask_0, stride = new_cache_23_internal_tensor_assign_2_stride_0, x = reshape_115)[name = tensor("slice_by_index_23")]; + tensor reshape_116_shape_0 = const()[name = tensor("reshape_116_shape_0"), val = tensor([-1])]; + tensor reshape_116 = reshape(shape = reshape_116_shape_0, x = slice_by_index_23)[name = tensor("reshape_116")]; + tensor reshape_117_shape_0 = const()[name = tensor("reshape_117_shape_0"), val = tensor([-1])]; + tensor reshape_117_cast_fp16 = reshape(shape = reshape_117_shape_0, x = var_4973_cast_fp16)[name = tensor("reshape_117_cast_fp16")]; + tensor reshape_118_shape_0 = const()[name = tensor("reshape_118_shape_0"), val = tensor([-1])]; + tensor reshape_118_cast_fp16 = reshape(shape = reshape_118_shape_0, x = reshape_114_cast_fp16)[name = tensor("reshape_118_cast_fp16")]; + tensor scatter_23_mode_0 = const()[name = tensor("scatter_23_mode_0"), val = tensor("update")]; + tensor scatter_23_axis_0 = const()[name = tensor("scatter_23_axis_0"), val = tensor(0)]; + tensor scatter_23_validate_indices_0 = const()[name = tensor("scatter_23_validate_indices_0"), val = tensor(false)]; + tensor scatter_23_cast_fp16 = scatter(axis = scatter_23_axis_0, data = reshape_118_cast_fp16, indices = reshape_116, mode = scatter_23_mode_0, updates = reshape_117_cast_fp16, validate_indices = scatter_23_validate_indices_0)[name = tensor("scatter_23_cast_fp16")]; + tensor reshape_119_cast_fp16 = reshape(shape = shape_69, x = scatter_23_cast_fp16)[name = tensor("reshape_119_cast_fp16")]; + tensor reshape_119_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("reshape_119_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor keys_67_begin_0 = const()[name = tensor("keys_67_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor keys_67_end_0 = const()[name = tensor("keys_67_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor keys_67_end_mask_0 = const()[name = tensor("keys_67_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor keys_67_squeeze_mask_0 = const()[name = tensor("keys_67_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor keys_67_cast_fp16 = slice_by_index(begin = keys_67_begin_0, end = keys_67_end_0, end_mask = keys_67_end_mask_0, squeeze_mask = keys_67_squeeze_mask_0, x = reshape_119_cast_fp16)[name = tensor("keys_67_cast_fp16")]; + tensor values_67_begin_0 = const()[name = tensor("values_67_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor values_67_end_0 = const()[name = tensor("values_67_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor values_67_end_mask_0 = const()[name = tensor("values_67_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor values_67_squeeze_mask_0 = const()[name = tensor("values_67_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor values_67_cast_fp16 = slice_by_index(begin = values_67_begin_0, end = values_67_end_0, end_mask = values_67_end_mask_0, squeeze_mask = values_67_squeeze_mask_0, x = reshape_119_cast_fp16)[name = tensor("values_67_cast_fp16")]; + tensor var_4985_cast_fp16 = not_equal(x = keys_67_cast_fp16, y = keys_67_cast_fp16)[name = tensor("op_4985_cast_fp16")]; + tensor keys_69_cast_fp16 = select(a = var_470_to_fp16, b = keys_67_cast_fp16, cond = var_4985_cast_fp16)[name = tensor("keys_69_cast_fp16")]; + tensor var_4993_cast_fp16 = not_equal(x = values_67_cast_fp16, y = values_67_cast_fp16)[name = tensor("op_4993_cast_fp16")]; + tensor values_69_cast_fp16 = select(a = var_470_to_fp16, b = values_67_cast_fp16, cond = var_4993_cast_fp16)[name = tensor("values_69_cast_fp16")]; + tensor var_5017 = const()[name = tensor("op_5017"), val = tensor([0, 2, 1, 3])]; + tensor var_5033 = const()[name = tensor("op_5033"), val = tensor([1, 1, 1])]; + tensor var_5034_cast_fp16 = reshape(shape = var_5033, x = position11_to_fp16)[name = tensor("op_5034_cast_fp16")]; + tensor q_positions_23_cast_fp16 = add(x = var_5034_cast_fp16, y = q_offsets_1_promoted_to_fp16)[name = tensor("q_positions_23_cast_fp16")]; + tensor valid_end_23_cast_fp16 = add(x = var_5034_cast_fp16, y = var_536_cast_fp16)[name = tensor("valid_end_23_cast_fp16")]; + tensor valid_mask_23_cast_fp16 = less(x = k_positions_1_promoted_to_fp16, y = valid_end_23_cast_fp16)[name = tensor("valid_mask_23_cast_fp16")]; + tensor causal_mask_23_cast_fp16 = less_equal(x = k_positions_1_promoted_to_fp16, y = q_positions_23_cast_fp16)[name = tensor("causal_mask_23_cast_fp16")]; + tensor var_5062 = logical_and(x = valid_mask_23_cast_fp16, y = causal_mask_23_cast_fp16)[name = tensor("op_5062")]; + tensor attn_mask_23_axes_0 = const()[name = tensor("attn_mask_23_axes_0"), val = tensor([1])]; + tensor attn_mask_23 = expand_dims(axes = attn_mask_23_axes_0, x = var_5062)[name = tensor("attn_mask_23")]; + tensor var_5076_transpose_x_0 = const()[name = tensor("op_5076_transpose_x_0"), val = tensor(false)]; + tensor var_5076_transpose_y_0 = const()[name = tensor("op_5076_transpose_y_0"), val = tensor(false)]; + tensor transpose_91_perm_0 = const()[name = tensor("transpose_91_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_92_perm_0 = const()[name = tensor("transpose_92_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_92 = transpose(perm = transpose_92_perm_0, x = keys_69_cast_fp16)[name = tensor("transpose_160")]; + tensor transpose_91 = transpose(perm = transpose_91_perm_0, x = q_69_cast_fp16)[name = tensor("transpose_161")]; + tensor var_5076_cast_fp16 = matmul(transpose_x = var_5076_transpose_x_0, transpose_y = var_5076_transpose_y_0, x = transpose_91, y = transpose_92)[name = tensor("op_5076_cast_fp16")]; + tensor var_5070_to_fp16 = const()[name = tensor("op_5070_to_fp16"), val = tensor([0x1p-3])]; + tensor attn_weights_67_cast_fp16 = mul(x = var_5076_cast_fp16, y = var_5070_to_fp16)[name = tensor("attn_weights_67_cast_fp16")]; + tensor var_5078 = logical_not(x = attn_mask_23)[name = tensor("op_5078")]; + tensor var_5079_to_fp16 = const()[name = tensor("op_5079_to_fp16"), val = tensor(-inf)]; + tensor attn_weights_69_cast_fp16 = select(a = var_5079_to_fp16, b = attn_weights_67_cast_fp16, cond = var_5078)[name = tensor("attn_weights_69_cast_fp16")]; + tensor var_5081 = const()[name = tensor("op_5081"), val = tensor(-1)]; + tensor attn_weights_71_cast_fp16 = softmax(axis = var_5081, x = attn_weights_69_cast_fp16)[name = tensor("attn_weights_71_cast_fp16")]; + tensor attn_output_23_transpose_x_0 = const()[name = tensor("attn_output_23_transpose_x_0"), val = tensor(false)]; + tensor attn_output_23_transpose_y_0 = const()[name = tensor("attn_output_23_transpose_y_0"), val = tensor(false)]; + tensor values_71_cast_fp16 = transpose(perm = var_5017, x = values_69_cast_fp16)[name = tensor("transpose_162")]; + tensor attn_output_23_cast_fp16 = matmul(transpose_x = attn_output_23_transpose_x_0, transpose_y = attn_output_23_transpose_y_0, x = attn_weights_71_cast_fp16, y = values_71_cast_fp16)[name = tensor("attn_output_23_cast_fp16")]; + tensor var_5089 = const()[name = tensor("op_5089"), val = tensor([0, 2, 1, 3])]; + tensor var_5094 = const()[name = tensor("op_5094"), val = tensor([1, 256, 1024])]; + tensor var_5090_cast_fp16 = transpose(perm = var_5089, x = attn_output_23_cast_fp16)[name = tensor("transpose_159")]; + tensor input_137_cast_fp16 = reshape(shape = var_5094, x = var_5090_cast_fp16)[name = tensor("input_137_cast_fp16")]; + tensor attn11_out_proj_weight_to_fp16 = const()[name = tensor("attn11_out_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(284285184)))]; + tensor linear_45_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = attn11_out_proj_weight_to_fp16, x = input_137_cast_fp16)[name = tensor("linear_45_cast_fp16")]; + tensor var_5103_cast_fp16 = add(x = position11_to_fp16, y = valid_len_to_fp16)[name = tensor("op_5103_cast_fp16")]; + tensor var_5103_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_5103_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor input_139_cast_fp16 = add(x = input_133_cast_fp16, y = linear_45_cast_fp16)[name = tensor("input_139_cast_fp16")]; + tensor input_141_axes_0 = const()[name = tensor("input_141_axes_0"), val = tensor([-1])]; + tensor norm11_2_weight_to_fp16 = const()[name = tensor("norm11_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(286382400)))]; + tensor norm11_2_bias_to_fp16 = const()[name = tensor("norm11_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(286384512)))]; + tensor var_5107_to_fp16 = const()[name = tensor("op_5107_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_141_cast_fp16 = layer_norm(axes = input_141_axes_0, beta = norm11_2_bias_to_fp16, epsilon = var_5107_to_fp16, gamma = norm11_2_weight_to_fp16, x = input_139_cast_fp16)[name = tensor("input_141_cast_fp16")]; + tensor linear11_1_weight_to_fp16 = const()[name = tensor("linear11_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(286386624)))]; + tensor linear_46_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = linear11_1_weight_to_fp16, x = input_141_cast_fp16)[name = tensor("linear_46_cast_fp16")]; + tensor input_143_mode_0 = const()[name = tensor("input_143_mode_0"), val = tensor("EXACT")]; + tensor input_143_cast_fp16 = gelu(mode = input_143_mode_0, x = linear_46_cast_fp16)[name = tensor("input_143_cast_fp16")]; + tensor linear11_2_weight_to_fp16 = const()[name = tensor("linear11_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(294775296)))]; + tensor linear_47_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = linear11_2_weight_to_fp16, x = input_143_cast_fp16)[name = tensor("linear_47_cast_fp16")]; + tensor input_145_cast_fp16 = add(x = input_139_cast_fp16, y = linear_47_cast_fp16)[name = tensor("input_145_cast_fp16")]; + tensor input_147_axes_0 = const()[name = tensor("input_147_axes_0"), val = tensor([-1])]; + tensor norm12_1_weight_to_fp16 = const()[name = tensor("norm12_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(303163968)))]; + tensor norm12_1_bias_to_fp16 = const()[name = tensor("norm12_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(303166080)))]; + tensor var_5124_to_fp16 = const()[name = tensor("op_5124_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_147_cast_fp16 = layer_norm(axes = input_147_axes_0, beta = norm12_1_bias_to_fp16, epsilon = var_5124_to_fp16, gamma = norm12_1_weight_to_fp16, x = input_145_cast_fp16)[name = tensor("input_147_cast_fp16")]; + tensor attn12_in_proj_weight_to_fp16 = const()[name = tensor("attn12_in_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(303168192)))]; + tensor linear_48_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = attn12_in_proj_weight_to_fp16, x = input_147_cast_fp16)[name = tensor("linear_48_cast_fp16")]; + tensor var_5138 = const()[name = tensor("op_5138"), val = tensor([1, 256, 3, 16, 64])]; + tensor qkv_25_cast_fp16 = reshape(shape = var_5138, x = linear_48_cast_fp16)[name = tensor("qkv_25_cast_fp16")]; + tensor q_73_begin_0 = const()[name = tensor("q_73_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor q_73_end_0 = const()[name = tensor("q_73_end_0"), val = tensor([1, 256, 1, 16, 64])]; + tensor q_73_end_mask_0 = const()[name = tensor("q_73_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor q_73_squeeze_mask_0 = const()[name = tensor("q_73_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor q_73_cast_fp16 = slice_by_index(begin = q_73_begin_0, end = q_73_end_0, end_mask = q_73_end_mask_0, squeeze_mask = q_73_squeeze_mask_0, x = qkv_25_cast_fp16)[name = tensor("q_73_cast_fp16")]; + tensor k_49_begin_0 = const()[name = tensor("k_49_begin_0"), val = tensor([0, 0, 1, 0, 0])]; + tensor k_49_end_0 = const()[name = tensor("k_49_end_0"), val = tensor([1, 256, 2, 16, 64])]; + tensor k_49_end_mask_0 = const()[name = tensor("k_49_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor k_49_squeeze_mask_0 = const()[name = tensor("k_49_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor k_49_cast_fp16 = slice_by_index(begin = k_49_begin_0, end = k_49_end_0, end_mask = k_49_end_mask_0, squeeze_mask = k_49_squeeze_mask_0, x = qkv_25_cast_fp16)[name = tensor("k_49_cast_fp16")]; + tensor v_25_begin_0 = const()[name = tensor("v_25_begin_0"), val = tensor([0, 0, 2, 0, 0])]; + tensor v_25_end_0 = const()[name = tensor("v_25_end_0"), val = tensor([1, 256, 3, 16, 64])]; + tensor v_25_end_mask_0 = const()[name = tensor("v_25_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor v_25_squeeze_mask_0 = const()[name = tensor("v_25_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor v_25_cast_fp16 = slice_by_index(begin = v_25_begin_0, end = v_25_end_0, end_mask = v_25_end_mask_0, squeeze_mask = v_25_squeeze_mask_0, x = qkv_25_cast_fp16)[name = tensor("v_25_cast_fp16")]; + tensor var_5196 = const()[name = tensor("op_5196"), val = tensor([1, 1])]; + tensor position12_to_fp16_dtype_0 = const()[name = tensor("position12_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor position12_to_fp16 = cast(dtype = position12_to_fp16_dtype_0, x = position12)[name = tensor("cast_601")]; + tensor var_5197_cast_fp16 = reshape(shape = var_5196, x = position12_to_fp16)[name = tensor("op_5197_cast_fp16")]; + tensor ts_75_cast_fp16 = add(x = ts_1_promoted_to_fp16, y = var_5197_cast_fp16)[name = tensor("ts_75_cast_fp16")]; + tensor var_5204 = const()[name = tensor("op_5204"), val = tensor([1, 256, 1, 1])]; + tensor ts_77_cast_fp16 = reshape(shape = var_5204, x = ts_75_cast_fp16)[name = tensor("ts_77_cast_fp16")]; + tensor var_5211 = const()[name = tensor("op_5211"), val = tensor([1, 256, 16, 32, 2])]; + tensor q_complex_25_cast_fp16 = reshape(shape = var_5211, x = q_73_cast_fp16)[name = tensor("q_complex_25_cast_fp16")]; + tensor var_5218 = const()[name = tensor("op_5218"), val = tensor([1, 256, 16, 32, 2])]; + tensor k_complex_25_cast_fp16 = reshape(shape = var_5218, x = k_49_cast_fp16)[name = tensor("k_complex_25_cast_fp16")]; + tensor var_5222_begin_0 = const()[name = tensor("op_5222_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_5222_end_0 = const()[name = tensor("op_5222_end_0"), val = tensor([1, 256, 16, 32, 1])]; + tensor var_5222_end_mask_0 = const()[name = tensor("op_5222_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_5222_squeeze_mask_0 = const()[name = tensor("op_5222_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_5222_cast_fp16 = slice_by_index(begin = var_5222_begin_0, end = var_5222_end_0, end_mask = var_5222_end_mask_0, squeeze_mask = var_5222_squeeze_mask_0, x = q_complex_25_cast_fp16)[name = tensor("op_5222_cast_fp16")]; + tensor var_5230_begin_0 = const()[name = tensor("op_5230_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_5230_end_0 = const()[name = tensor("op_5230_end_0"), val = tensor([1, 256, 16, 32, 2])]; + tensor var_5230_end_mask_0 = const()[name = tensor("op_5230_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_5230_squeeze_mask_0 = const()[name = tensor("op_5230_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_5230_cast_fp16 = slice_by_index(begin = var_5230_begin_0, end = var_5230_end_0, end_mask = var_5230_end_mask_0, squeeze_mask = var_5230_squeeze_mask_0, x = q_complex_25_cast_fp16)[name = tensor("op_5230_cast_fp16")]; + tensor var_5238_begin_0 = const()[name = tensor("op_5238_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_5238_end_0 = const()[name = tensor("op_5238_end_0"), val = tensor([1, 256, 16, 32, 1])]; + tensor var_5238_end_mask_0 = const()[name = tensor("op_5238_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_5238_squeeze_mask_0 = const()[name = tensor("op_5238_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_5238_cast_fp16 = slice_by_index(begin = var_5238_begin_0, end = var_5238_end_0, end_mask = var_5238_end_mask_0, squeeze_mask = var_5238_squeeze_mask_0, x = k_complex_25_cast_fp16)[name = tensor("op_5238_cast_fp16")]; + tensor var_5246_begin_0 = const()[name = tensor("op_5246_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_5246_end_0 = const()[name = tensor("op_5246_end_0"), val = tensor([1, 256, 16, 32, 2])]; + tensor var_5246_end_mask_0 = const()[name = tensor("op_5246_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_5246_squeeze_mask_0 = const()[name = tensor("op_5246_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_5246_cast_fp16 = slice_by_index(begin = var_5246_begin_0, end = var_5246_end_0, end_mask = var_5246_end_mask_0, squeeze_mask = var_5246_squeeze_mask_0, x = k_complex_25_cast_fp16)[name = tensor("op_5246_cast_fp16")]; + tensor freqs_25_to_fp16 = const()[name = tensor("freqs_25_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(309459712)))]; + tensor var_5252_cast_fp16 = mul(x = freqs_25_to_fp16, y = ts_77_cast_fp16)[name = tensor("op_5252_cast_fp16")]; + tensor rotr_25_cast_fp16 = cos(x = var_5252_cast_fp16)[name = tensor("rotr_25_cast_fp16")]; + tensor roti_25_cast_fp16 = sin(x = var_5252_cast_fp16)[name = tensor("roti_25_cast_fp16")]; + tensor var_5256_cast_fp16 = mul(x = var_5222_cast_fp16, y = rotr_25_cast_fp16)[name = tensor("op_5256_cast_fp16")]; + tensor var_5257_cast_fp16 = mul(x = var_5230_cast_fp16, y = roti_25_cast_fp16)[name = tensor("op_5257_cast_fp16")]; + tensor qor_49_cast_fp16 = sub(x = var_5256_cast_fp16, y = var_5257_cast_fp16)[name = tensor("qor_49_cast_fp16")]; + tensor var_5260_cast_fp16 = mul(x = var_5222_cast_fp16, y = roti_25_cast_fp16)[name = tensor("op_5260_cast_fp16")]; + tensor var_5261_cast_fp16 = mul(x = var_5230_cast_fp16, y = rotr_25_cast_fp16)[name = tensor("op_5261_cast_fp16")]; + tensor qoi_49_cast_fp16 = add(x = var_5260_cast_fp16, y = var_5261_cast_fp16)[name = tensor("qoi_49_cast_fp16")]; + tensor var_5264_cast_fp16 = mul(x = var_5238_cast_fp16, y = rotr_25_cast_fp16)[name = tensor("op_5264_cast_fp16")]; + tensor var_5265_cast_fp16 = mul(x = var_5246_cast_fp16, y = roti_25_cast_fp16)[name = tensor("op_5265_cast_fp16")]; + tensor kor_49_cast_fp16 = sub(x = var_5264_cast_fp16, y = var_5265_cast_fp16)[name = tensor("kor_49_cast_fp16")]; + tensor var_5268_cast_fp16 = mul(x = var_5238_cast_fp16, y = roti_25_cast_fp16)[name = tensor("op_5268_cast_fp16")]; + tensor var_5269_cast_fp16 = mul(x = var_5246_cast_fp16, y = rotr_25_cast_fp16)[name = tensor("op_5269_cast_fp16")]; + tensor koi_49_cast_fp16 = add(x = var_5268_cast_fp16, y = var_5269_cast_fp16)[name = tensor("koi_49_cast_fp16")]; + tensor qo_25_axis_0 = const()[name = tensor("qo_25_axis_0"), val = tensor(-1)]; + tensor qo_25_cast_fp16 = stack(axis = qo_25_axis_0, values = (qor_49_cast_fp16, qoi_49_cast_fp16))[name = tensor("qo_25_cast_fp16")]; + tensor ko_25_axis_0 = const()[name = tensor("ko_25_axis_0"), val = tensor(-1)]; + tensor ko_25_cast_fp16 = stack(axis = ko_25_axis_0, values = (kor_49_cast_fp16, koi_49_cast_fp16))[name = tensor("ko_25_cast_fp16")]; + tensor var_5302 = const()[name = tensor("op_5302"), val = tensor([1, 256, 16, 64])]; + tensor q_75_cast_fp16 = reshape(shape = var_5302, x = qo_25_cast_fp16)[name = tensor("q_75_cast_fp16")]; + tensor var_5308 = const()[name = tensor("op_5308"), val = tensor([1, 256, 16, 64])]; + tensor k_51_cast_fp16 = reshape(shape = var_5308, x = ko_25_cast_fp16)[name = tensor("k_51_cast_fp16")]; + tensor write_indices_float_73_cast_fp16 = add(x = var_5197_cast_fp16, y = var_393_promoted_to_fp16)[name = tensor("write_indices_float_73_cast_fp16")]; + tensor _inversed_5329_y_0_to_fp16 = const()[name = tensor("_inversed_5329_y_0_to_fp16"), val = tensor(0x1p-9)]; + tensor _inversed_5329_cast_fp16 = mul(x = write_indices_float_73_cast_fp16, y = _inversed_5329_y_0_to_fp16)[name = tensor("_inversed_5329_cast_fp16")]; + tensor var_5330_cast_fp16 = floor(x = _inversed_5329_cast_fp16)[name = tensor("op_5330_cast_fp16")]; + tensor var_5331_to_fp16 = const()[name = tensor("op_5331_to_fp16"), val = tensor(0x1p+9)]; + tensor var_5332_cast_fp16 = mul(x = var_5330_cast_fp16, y = var_5331_to_fp16)[name = tensor("op_5332_cast_fp16")]; + tensor write_indices_float_75_cast_fp16 = sub(x = write_indices_float_73_cast_fp16, y = var_5332_cast_fp16)[name = tensor("write_indices_float_75_cast_fp16")]; + tensor var_5350_value_0_to_fp16 = const()[name = tensor("op_5350_value_0_to_fp16"), val = tensor(0x1.ffp+8)]; + tensor var_5350_cast_fp16 = fill_like(ref_tensor = write_indices_float_75_cast_fp16, value = var_5350_value_0_to_fp16)[name = tensor("op_5350_cast_fp16")]; + tensor write_indices_float_77_cast_fp16 = select(a = write_indices_float_75_cast_fp16, b = var_5350_cast_fp16, cond = token_valid_1_cast_fp16)[name = tensor("write_indices_float_77_cast_fp16")]; + tensor var_5356_dtype_0 = const()[name = tensor("op_5356_dtype_0"), val = tensor("int32")]; + tensor var_5361 = const()[name = tensor("op_5361"), val = tensor([1, 256, 1, 1])]; + tensor write_indices_float_77_cast_fp16_to_int32 = cast(dtype = var_5356_dtype_0, x = write_indices_float_77_cast_fp16)[name = tensor("cast_600")]; + tensor var_5362 = reshape(shape = var_5361, x = write_indices_float_77_cast_fp16_to_int32)[name = tensor("op_5362")]; + tensor write_indices_25_reps_0 = const()[name = tensor("write_indices_25_reps_0"), val = tensor([1, 1, 16, 64])]; + tensor write_indices_25 = tile(reps = write_indices_25_reps_0, x = var_5362)[name = tensor("write_indices_25")]; + tensor var_5372_begin_0 = const()[name = tensor("op_5372_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_5372_end_0 = const()[name = tensor("op_5372_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor var_5372_end_mask_0 = const()[name = tensor("op_5372_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_5372_squeeze_mask_0 = const()[name = tensor("op_5372_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor cache12_to_fp16_dtype_0 = const()[name = tensor("cache12_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor cache12_to_fp16 = cast(dtype = cache12_to_fp16_dtype_0, x = cache12)[name = tensor("cast_599")]; + tensor var_5372_cast_fp16 = slice_by_index(begin = var_5372_begin_0, end = var_5372_end_0, end_mask = var_5372_end_mask_0, squeeze_mask = var_5372_squeeze_mask_0, x = cache12_to_fp16)[name = tensor("op_5372_cast_fp16")]; + tensor var_5374_axis_0 = const()[name = tensor("op_5374_axis_0"), val = tensor(1)]; + tensor var_5374_mode_0 = const()[name = tensor("op_5374_mode_0"), val = tensor("update")]; + tensor var_5374_validate_indices_0 = const()[name = tensor("op_5374_validate_indices_0"), val = tensor(false)]; + tensor var_5374_cast_fp16 = scatter_along_axis(axis = var_5374_axis_0, data = var_5372_cast_fp16, indices = write_indices_25, mode = var_5374_mode_0, updates = k_51_cast_fp16, validate_indices = var_5374_validate_indices_0)[name = tensor("op_5374_cast_fp16")]; + tensor concat_98 = const()[name = tensor("concat_98"), val = tensor([0, 0, 0, 0, 0])]; + tensor concat_99 = const()[name = tensor("concat_99"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_25_internal_tensor_assign_1_stride_0 = const()[name = tensor("new_cache_25_internal_tensor_assign_1_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_25_internal_tensor_assign_1_begin_mask_0 = const()[name = tensor("new_cache_25_internal_tensor_assign_1_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_25_internal_tensor_assign_1_end_mask_0 = const()[name = tensor("new_cache_25_internal_tensor_assign_1_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_25_internal_tensor_assign_1_squeeze_mask_0 = const()[name = tensor("new_cache_25_internal_tensor_assign_1_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_70 = const()[name = tensor("shape_70"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_24 = const()[name = tensor("reduce_prod_24"), val = tensor(1048576)]; + tensor range_1d_24_start_0 = const()[name = tensor("range_1d_24_start_0"), val = tensor(0)]; + tensor range_1d_24_step_0 = const()[name = tensor("range_1d_24_step_0"), val = tensor(1)]; + tensor range_1d_24 = range_1d(end = reduce_prod_24, start = range_1d_24_start_0, step = range_1d_24_step_0)[name = tensor("range_1d_24")]; + tensor reshape_120 = reshape(shape = shape_70, x = range_1d_24)[name = tensor("reshape_120")]; + tensor slice_by_index_24 = slice_by_index(begin = concat_98, begin_mask = new_cache_25_internal_tensor_assign_1_begin_mask_0, end = concat_99, end_mask = new_cache_25_internal_tensor_assign_1_end_mask_0, squeeze_mask = new_cache_25_internal_tensor_assign_1_squeeze_mask_0, stride = new_cache_25_internal_tensor_assign_1_stride_0, x = reshape_120)[name = tensor("slice_by_index_24")]; + tensor reshape_121_shape_0 = const()[name = tensor("reshape_121_shape_0"), val = tensor([-1])]; + tensor reshape_121 = reshape(shape = reshape_121_shape_0, x = slice_by_index_24)[name = tensor("reshape_121")]; + tensor reshape_122_shape_0 = const()[name = tensor("reshape_122_shape_0"), val = tensor([-1])]; + tensor reshape_122_cast_fp16 = reshape(shape = reshape_122_shape_0, x = var_5374_cast_fp16)[name = tensor("reshape_122_cast_fp16")]; + tensor reshape_123_shape_0 = const()[name = tensor("reshape_123_shape_0"), val = tensor([-1])]; + tensor reshape_123_cast_fp16 = reshape(shape = reshape_123_shape_0, x = cache12_to_fp16)[name = tensor("reshape_123_cast_fp16")]; + tensor scatter_24_mode_0 = const()[name = tensor("scatter_24_mode_0"), val = tensor("update")]; + tensor scatter_24_axis_0 = const()[name = tensor("scatter_24_axis_0"), val = tensor(0)]; + tensor scatter_24_validate_indices_0 = const()[name = tensor("scatter_24_validate_indices_0"), val = tensor(false)]; + tensor scatter_24_cast_fp16 = scatter(axis = scatter_24_axis_0, data = reshape_123_cast_fp16, indices = reshape_121, mode = scatter_24_mode_0, updates = reshape_122_cast_fp16, validate_indices = scatter_24_validate_indices_0)[name = tensor("scatter_24_cast_fp16")]; + tensor reshape_124_cast_fp16 = reshape(shape = shape_70, x = scatter_24_cast_fp16)[name = tensor("reshape_124_cast_fp16")]; + tensor var_5382_begin_0 = const()[name = tensor("op_5382_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor var_5382_end_0 = const()[name = tensor("op_5382_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor var_5382_end_mask_0 = const()[name = tensor("op_5382_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_5382_squeeze_mask_0 = const()[name = tensor("op_5382_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor var_5382_cast_fp16 = slice_by_index(begin = var_5382_begin_0, end = var_5382_end_0, end_mask = var_5382_end_mask_0, squeeze_mask = var_5382_squeeze_mask_0, x = reshape_124_cast_fp16)[name = tensor("op_5382_cast_fp16")]; + tensor var_5384_axis_0 = const()[name = tensor("op_5384_axis_0"), val = tensor(1)]; + tensor var_5384_mode_0 = const()[name = tensor("op_5384_mode_0"), val = tensor("update")]; + tensor var_5384_validate_indices_0 = const()[name = tensor("op_5384_validate_indices_0"), val = tensor(false)]; + tensor var_5384_cast_fp16 = scatter_along_axis(axis = var_5384_axis_0, data = var_5382_cast_fp16, indices = write_indices_25, mode = var_5384_mode_0, updates = v_25_cast_fp16, validate_indices = var_5384_validate_indices_0)[name = tensor("op_5384_cast_fp16")]; + tensor concat_100 = const()[name = tensor("concat_100"), val = tensor([1, 0, 0, 0, 0])]; + tensor concat_101 = const()[name = tensor("concat_101"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_25_internal_tensor_assign_2_stride_0 = const()[name = tensor("new_cache_25_internal_tensor_assign_2_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_25_internal_tensor_assign_2_begin_mask_0 = const()[name = tensor("new_cache_25_internal_tensor_assign_2_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_25_internal_tensor_assign_2_end_mask_0 = const()[name = tensor("new_cache_25_internal_tensor_assign_2_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_25_internal_tensor_assign_2_squeeze_mask_0 = const()[name = tensor("new_cache_25_internal_tensor_assign_2_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_71 = const()[name = tensor("shape_71"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_25 = const()[name = tensor("reduce_prod_25"), val = tensor(1048576)]; + tensor range_1d_25_start_0 = const()[name = tensor("range_1d_25_start_0"), val = tensor(0)]; + tensor range_1d_25_step_0 = const()[name = tensor("range_1d_25_step_0"), val = tensor(1)]; + tensor range_1d_25 = range_1d(end = reduce_prod_25, start = range_1d_25_start_0, step = range_1d_25_step_0)[name = tensor("range_1d_25")]; + tensor reshape_125 = reshape(shape = shape_71, x = range_1d_25)[name = tensor("reshape_125")]; + tensor slice_by_index_25 = slice_by_index(begin = concat_100, begin_mask = new_cache_25_internal_tensor_assign_2_begin_mask_0, end = concat_101, end_mask = new_cache_25_internal_tensor_assign_2_end_mask_0, squeeze_mask = new_cache_25_internal_tensor_assign_2_squeeze_mask_0, stride = new_cache_25_internal_tensor_assign_2_stride_0, x = reshape_125)[name = tensor("slice_by_index_25")]; + tensor reshape_126_shape_0 = const()[name = tensor("reshape_126_shape_0"), val = tensor([-1])]; + tensor reshape_126 = reshape(shape = reshape_126_shape_0, x = slice_by_index_25)[name = tensor("reshape_126")]; + tensor reshape_127_shape_0 = const()[name = tensor("reshape_127_shape_0"), val = tensor([-1])]; + tensor reshape_127_cast_fp16 = reshape(shape = reshape_127_shape_0, x = var_5384_cast_fp16)[name = tensor("reshape_127_cast_fp16")]; + tensor reshape_128_shape_0 = const()[name = tensor("reshape_128_shape_0"), val = tensor([-1])]; + tensor reshape_128_cast_fp16 = reshape(shape = reshape_128_shape_0, x = reshape_124_cast_fp16)[name = tensor("reshape_128_cast_fp16")]; + tensor scatter_25_mode_0 = const()[name = tensor("scatter_25_mode_0"), val = tensor("update")]; + tensor scatter_25_axis_0 = const()[name = tensor("scatter_25_axis_0"), val = tensor(0)]; + tensor scatter_25_validate_indices_0 = const()[name = tensor("scatter_25_validate_indices_0"), val = tensor(false)]; + tensor scatter_25_cast_fp16 = scatter(axis = scatter_25_axis_0, data = reshape_128_cast_fp16, indices = reshape_126, mode = scatter_25_mode_0, updates = reshape_127_cast_fp16, validate_indices = scatter_25_validate_indices_0)[name = tensor("scatter_25_cast_fp16")]; + tensor reshape_129_cast_fp16 = reshape(shape = shape_71, x = scatter_25_cast_fp16)[name = tensor("reshape_129_cast_fp16")]; + tensor reshape_129_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("reshape_129_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor keys_73_begin_0 = const()[name = tensor("keys_73_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor keys_73_end_0 = const()[name = tensor("keys_73_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor keys_73_end_mask_0 = const()[name = tensor("keys_73_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor keys_73_squeeze_mask_0 = const()[name = tensor("keys_73_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor keys_73_cast_fp16 = slice_by_index(begin = keys_73_begin_0, end = keys_73_end_0, end_mask = keys_73_end_mask_0, squeeze_mask = keys_73_squeeze_mask_0, x = reshape_129_cast_fp16)[name = tensor("keys_73_cast_fp16")]; + tensor values_73_begin_0 = const()[name = tensor("values_73_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor values_73_end_0 = const()[name = tensor("values_73_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor values_73_end_mask_0 = const()[name = tensor("values_73_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor values_73_squeeze_mask_0 = const()[name = tensor("values_73_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor values_73_cast_fp16 = slice_by_index(begin = values_73_begin_0, end = values_73_end_0, end_mask = values_73_end_mask_0, squeeze_mask = values_73_squeeze_mask_0, x = reshape_129_cast_fp16)[name = tensor("values_73_cast_fp16")]; + tensor var_5396_cast_fp16 = not_equal(x = keys_73_cast_fp16, y = keys_73_cast_fp16)[name = tensor("op_5396_cast_fp16")]; + tensor keys_75_cast_fp16 = select(a = var_470_to_fp16, b = keys_73_cast_fp16, cond = var_5396_cast_fp16)[name = tensor("keys_75_cast_fp16")]; + tensor var_5404_cast_fp16 = not_equal(x = values_73_cast_fp16, y = values_73_cast_fp16)[name = tensor("op_5404_cast_fp16")]; + tensor values_75_cast_fp16 = select(a = var_470_to_fp16, b = values_73_cast_fp16, cond = var_5404_cast_fp16)[name = tensor("values_75_cast_fp16")]; + tensor var_5428 = const()[name = tensor("op_5428"), val = tensor([0, 2, 1, 3])]; + tensor var_5444 = const()[name = tensor("op_5444"), val = tensor([1, 1, 1])]; + tensor var_5445_cast_fp16 = reshape(shape = var_5444, x = position12_to_fp16)[name = tensor("op_5445_cast_fp16")]; + tensor q_positions_25_cast_fp16 = add(x = var_5445_cast_fp16, y = q_offsets_1_promoted_to_fp16)[name = tensor("q_positions_25_cast_fp16")]; + tensor valid_end_25_cast_fp16 = add(x = var_5445_cast_fp16, y = var_536_cast_fp16)[name = tensor("valid_end_25_cast_fp16")]; + tensor valid_mask_25_cast_fp16 = less(x = k_positions_1_promoted_to_fp16, y = valid_end_25_cast_fp16)[name = tensor("valid_mask_25_cast_fp16")]; + tensor causal_mask_25_cast_fp16 = less_equal(x = k_positions_1_promoted_to_fp16, y = q_positions_25_cast_fp16)[name = tensor("causal_mask_25_cast_fp16")]; + tensor var_5473 = logical_and(x = valid_mask_25_cast_fp16, y = causal_mask_25_cast_fp16)[name = tensor("op_5473")]; + tensor attn_mask_25_axes_0 = const()[name = tensor("attn_mask_25_axes_0"), val = tensor([1])]; + tensor attn_mask_25 = expand_dims(axes = attn_mask_25_axes_0, x = var_5473)[name = tensor("attn_mask_25")]; + tensor var_5487_transpose_x_0 = const()[name = tensor("op_5487_transpose_x_0"), val = tensor(false)]; + tensor var_5487_transpose_y_0 = const()[name = tensor("op_5487_transpose_y_0"), val = tensor(false)]; + tensor transpose_93_perm_0 = const()[name = tensor("transpose_93_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_94_perm_0 = const()[name = tensor("transpose_94_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_94 = transpose(perm = transpose_94_perm_0, x = keys_75_cast_fp16)[name = tensor("transpose_156")]; + tensor transpose_93 = transpose(perm = transpose_93_perm_0, x = q_75_cast_fp16)[name = tensor("transpose_157")]; + tensor var_5487_cast_fp16 = matmul(transpose_x = var_5487_transpose_x_0, transpose_y = var_5487_transpose_y_0, x = transpose_93, y = transpose_94)[name = tensor("op_5487_cast_fp16")]; + tensor var_5481_to_fp16 = const()[name = tensor("op_5481_to_fp16"), val = tensor([0x1p-3])]; + tensor attn_weights_73_cast_fp16 = mul(x = var_5487_cast_fp16, y = var_5481_to_fp16)[name = tensor("attn_weights_73_cast_fp16")]; + tensor var_5489 = logical_not(x = attn_mask_25)[name = tensor("op_5489")]; + tensor var_5490_to_fp16 = const()[name = tensor("op_5490_to_fp16"), val = tensor(-inf)]; + tensor attn_weights_75_cast_fp16 = select(a = var_5490_to_fp16, b = attn_weights_73_cast_fp16, cond = var_5489)[name = tensor("attn_weights_75_cast_fp16")]; + tensor var_5492 = const()[name = tensor("op_5492"), val = tensor(-1)]; + tensor attn_weights_77_cast_fp16 = softmax(axis = var_5492, x = attn_weights_75_cast_fp16)[name = tensor("attn_weights_77_cast_fp16")]; + tensor attn_output_25_transpose_x_0 = const()[name = tensor("attn_output_25_transpose_x_0"), val = tensor(false)]; + tensor attn_output_25_transpose_y_0 = const()[name = tensor("attn_output_25_transpose_y_0"), val = tensor(false)]; + tensor values_77_cast_fp16 = transpose(perm = var_5428, x = values_75_cast_fp16)[name = tensor("transpose_158")]; + tensor attn_output_25_cast_fp16 = matmul(transpose_x = attn_output_25_transpose_x_0, transpose_y = attn_output_25_transpose_y_0, x = attn_weights_77_cast_fp16, y = values_77_cast_fp16)[name = tensor("attn_output_25_cast_fp16")]; + tensor var_5500 = const()[name = tensor("op_5500"), val = tensor([0, 2, 1, 3])]; + tensor var_5505 = const()[name = tensor("op_5505"), val = tensor([1, 256, 1024])]; + tensor var_5501_cast_fp16 = transpose(perm = var_5500, x = attn_output_25_cast_fp16)[name = tensor("transpose_155")]; + tensor input_149_cast_fp16 = reshape(shape = var_5505, x = var_5501_cast_fp16)[name = tensor("input_149_cast_fp16")]; + tensor attn12_out_proj_weight_to_fp16 = const()[name = tensor("attn12_out_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(309459840)))]; + tensor linear_49_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = attn12_out_proj_weight_to_fp16, x = input_149_cast_fp16)[name = tensor("linear_49_cast_fp16")]; + tensor var_5514_cast_fp16 = add(x = position12_to_fp16, y = valid_len_to_fp16)[name = tensor("op_5514_cast_fp16")]; + tensor var_5514_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_5514_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor input_151_cast_fp16 = add(x = input_145_cast_fp16, y = linear_49_cast_fp16)[name = tensor("input_151_cast_fp16")]; + tensor input_153_axes_0 = const()[name = tensor("input_153_axes_0"), val = tensor([-1])]; + tensor norm12_2_weight_to_fp16 = const()[name = tensor("norm12_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(311557056)))]; + tensor norm12_2_bias_to_fp16 = const()[name = tensor("norm12_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(311559168)))]; + tensor var_5518_to_fp16 = const()[name = tensor("op_5518_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_153_cast_fp16 = layer_norm(axes = input_153_axes_0, beta = norm12_2_bias_to_fp16, epsilon = var_5518_to_fp16, gamma = norm12_2_weight_to_fp16, x = input_151_cast_fp16)[name = tensor("input_153_cast_fp16")]; + tensor linear12_1_weight_to_fp16 = const()[name = tensor("linear12_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(311561280)))]; + tensor linear_50_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = linear12_1_weight_to_fp16, x = input_153_cast_fp16)[name = tensor("linear_50_cast_fp16")]; + tensor input_155_mode_0 = const()[name = tensor("input_155_mode_0"), val = tensor("EXACT")]; + tensor input_155_cast_fp16 = gelu(mode = input_155_mode_0, x = linear_50_cast_fp16)[name = tensor("input_155_cast_fp16")]; + tensor linear12_2_weight_to_fp16 = const()[name = tensor("linear12_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(319949952)))]; + tensor linear_51_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = linear12_2_weight_to_fp16, x = input_155_cast_fp16)[name = tensor("linear_51_cast_fp16")]; + tensor input_157_cast_fp16 = add(x = input_151_cast_fp16, y = linear_51_cast_fp16)[name = tensor("input_157_cast_fp16")]; + tensor input_159_axes_0 = const()[name = tensor("input_159_axes_0"), val = tensor([-1])]; + tensor norm13_1_weight_to_fp16 = const()[name = tensor("norm13_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(328338624)))]; + tensor norm13_1_bias_to_fp16 = const()[name = tensor("norm13_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(328340736)))]; + tensor var_5535_to_fp16 = const()[name = tensor("op_5535_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_159_cast_fp16 = layer_norm(axes = input_159_axes_0, beta = norm13_1_bias_to_fp16, epsilon = var_5535_to_fp16, gamma = norm13_1_weight_to_fp16, x = input_157_cast_fp16)[name = tensor("input_159_cast_fp16")]; + tensor attn13_in_proj_weight_to_fp16 = const()[name = tensor("attn13_in_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(328342848)))]; + tensor linear_52_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = attn13_in_proj_weight_to_fp16, x = input_159_cast_fp16)[name = tensor("linear_52_cast_fp16")]; + tensor var_5549 = const()[name = tensor("op_5549"), val = tensor([1, 256, 3, 16, 64])]; + tensor qkv_27_cast_fp16 = reshape(shape = var_5549, x = linear_52_cast_fp16)[name = tensor("qkv_27_cast_fp16")]; + tensor q_79_begin_0 = const()[name = tensor("q_79_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor q_79_end_0 = const()[name = tensor("q_79_end_0"), val = tensor([1, 256, 1, 16, 64])]; + tensor q_79_end_mask_0 = const()[name = tensor("q_79_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor q_79_squeeze_mask_0 = const()[name = tensor("q_79_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor q_79_cast_fp16 = slice_by_index(begin = q_79_begin_0, end = q_79_end_0, end_mask = q_79_end_mask_0, squeeze_mask = q_79_squeeze_mask_0, x = qkv_27_cast_fp16)[name = tensor("q_79_cast_fp16")]; + tensor k_53_begin_0 = const()[name = tensor("k_53_begin_0"), val = tensor([0, 0, 1, 0, 0])]; + tensor k_53_end_0 = const()[name = tensor("k_53_end_0"), val = tensor([1, 256, 2, 16, 64])]; + tensor k_53_end_mask_0 = const()[name = tensor("k_53_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor k_53_squeeze_mask_0 = const()[name = tensor("k_53_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor k_53_cast_fp16 = slice_by_index(begin = k_53_begin_0, end = k_53_end_0, end_mask = k_53_end_mask_0, squeeze_mask = k_53_squeeze_mask_0, x = qkv_27_cast_fp16)[name = tensor("k_53_cast_fp16")]; + tensor v_27_begin_0 = const()[name = tensor("v_27_begin_0"), val = tensor([0, 0, 2, 0, 0])]; + tensor v_27_end_0 = const()[name = tensor("v_27_end_0"), val = tensor([1, 256, 3, 16, 64])]; + tensor v_27_end_mask_0 = const()[name = tensor("v_27_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor v_27_squeeze_mask_0 = const()[name = tensor("v_27_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor v_27_cast_fp16 = slice_by_index(begin = v_27_begin_0, end = v_27_end_0, end_mask = v_27_end_mask_0, squeeze_mask = v_27_squeeze_mask_0, x = qkv_27_cast_fp16)[name = tensor("v_27_cast_fp16")]; + tensor var_5607 = const()[name = tensor("op_5607"), val = tensor([1, 1])]; + tensor position13_to_fp16_dtype_0 = const()[name = tensor("position13_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor position13_to_fp16 = cast(dtype = position13_to_fp16_dtype_0, x = position13)[name = tensor("cast_596")]; + tensor var_5608_cast_fp16 = reshape(shape = var_5607, x = position13_to_fp16)[name = tensor("op_5608_cast_fp16")]; + tensor ts_81_cast_fp16 = add(x = ts_1_promoted_to_fp16, y = var_5608_cast_fp16)[name = tensor("ts_81_cast_fp16")]; + tensor var_5615 = const()[name = tensor("op_5615"), val = tensor([1, 256, 1, 1])]; + tensor ts_83_cast_fp16 = reshape(shape = var_5615, x = ts_81_cast_fp16)[name = tensor("ts_83_cast_fp16")]; + tensor var_5622 = const()[name = tensor("op_5622"), val = tensor([1, 256, 16, 32, 2])]; + tensor q_complex_27_cast_fp16 = reshape(shape = var_5622, x = q_79_cast_fp16)[name = tensor("q_complex_27_cast_fp16")]; + tensor var_5629 = const()[name = tensor("op_5629"), val = tensor([1, 256, 16, 32, 2])]; + tensor k_complex_27_cast_fp16 = reshape(shape = var_5629, x = k_53_cast_fp16)[name = tensor("k_complex_27_cast_fp16")]; + tensor var_5633_begin_0 = const()[name = tensor("op_5633_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_5633_end_0 = const()[name = tensor("op_5633_end_0"), val = tensor([1, 256, 16, 32, 1])]; + tensor var_5633_end_mask_0 = const()[name = tensor("op_5633_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_5633_squeeze_mask_0 = const()[name = tensor("op_5633_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_5633_cast_fp16 = slice_by_index(begin = var_5633_begin_0, end = var_5633_end_0, end_mask = var_5633_end_mask_0, squeeze_mask = var_5633_squeeze_mask_0, x = q_complex_27_cast_fp16)[name = tensor("op_5633_cast_fp16")]; + tensor var_5641_begin_0 = const()[name = tensor("op_5641_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_5641_end_0 = const()[name = tensor("op_5641_end_0"), val = tensor([1, 256, 16, 32, 2])]; + tensor var_5641_end_mask_0 = const()[name = tensor("op_5641_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_5641_squeeze_mask_0 = const()[name = tensor("op_5641_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_5641_cast_fp16 = slice_by_index(begin = var_5641_begin_0, end = var_5641_end_0, end_mask = var_5641_end_mask_0, squeeze_mask = var_5641_squeeze_mask_0, x = q_complex_27_cast_fp16)[name = tensor("op_5641_cast_fp16")]; + tensor var_5649_begin_0 = const()[name = tensor("op_5649_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_5649_end_0 = const()[name = tensor("op_5649_end_0"), val = tensor([1, 256, 16, 32, 1])]; + tensor var_5649_end_mask_0 = const()[name = tensor("op_5649_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_5649_squeeze_mask_0 = const()[name = tensor("op_5649_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_5649_cast_fp16 = slice_by_index(begin = var_5649_begin_0, end = var_5649_end_0, end_mask = var_5649_end_mask_0, squeeze_mask = var_5649_squeeze_mask_0, x = k_complex_27_cast_fp16)[name = tensor("op_5649_cast_fp16")]; + tensor var_5657_begin_0 = const()[name = tensor("op_5657_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_5657_end_0 = const()[name = tensor("op_5657_end_0"), val = tensor([1, 256, 16, 32, 2])]; + tensor var_5657_end_mask_0 = const()[name = tensor("op_5657_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_5657_squeeze_mask_0 = const()[name = tensor("op_5657_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_5657_cast_fp16 = slice_by_index(begin = var_5657_begin_0, end = var_5657_end_0, end_mask = var_5657_end_mask_0, squeeze_mask = var_5657_squeeze_mask_0, x = k_complex_27_cast_fp16)[name = tensor("op_5657_cast_fp16")]; + tensor freqs_27_to_fp16 = const()[name = tensor("freqs_27_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(334634368)))]; + tensor var_5663_cast_fp16 = mul(x = freqs_27_to_fp16, y = ts_83_cast_fp16)[name = tensor("op_5663_cast_fp16")]; + tensor rotr_27_cast_fp16 = cos(x = var_5663_cast_fp16)[name = tensor("rotr_27_cast_fp16")]; + tensor roti_27_cast_fp16 = sin(x = var_5663_cast_fp16)[name = tensor("roti_27_cast_fp16")]; + tensor var_5667_cast_fp16 = mul(x = var_5633_cast_fp16, y = rotr_27_cast_fp16)[name = tensor("op_5667_cast_fp16")]; + tensor var_5668_cast_fp16 = mul(x = var_5641_cast_fp16, y = roti_27_cast_fp16)[name = tensor("op_5668_cast_fp16")]; + tensor qor_53_cast_fp16 = sub(x = var_5667_cast_fp16, y = var_5668_cast_fp16)[name = tensor("qor_53_cast_fp16")]; + tensor var_5671_cast_fp16 = mul(x = var_5633_cast_fp16, y = roti_27_cast_fp16)[name = tensor("op_5671_cast_fp16")]; + tensor var_5672_cast_fp16 = mul(x = var_5641_cast_fp16, y = rotr_27_cast_fp16)[name = tensor("op_5672_cast_fp16")]; + tensor qoi_53_cast_fp16 = add(x = var_5671_cast_fp16, y = var_5672_cast_fp16)[name = tensor("qoi_53_cast_fp16")]; + tensor var_5675_cast_fp16 = mul(x = var_5649_cast_fp16, y = rotr_27_cast_fp16)[name = tensor("op_5675_cast_fp16")]; + tensor var_5676_cast_fp16 = mul(x = var_5657_cast_fp16, y = roti_27_cast_fp16)[name = tensor("op_5676_cast_fp16")]; + tensor kor_53_cast_fp16 = sub(x = var_5675_cast_fp16, y = var_5676_cast_fp16)[name = tensor("kor_53_cast_fp16")]; + tensor var_5679_cast_fp16 = mul(x = var_5649_cast_fp16, y = roti_27_cast_fp16)[name = tensor("op_5679_cast_fp16")]; + tensor var_5680_cast_fp16 = mul(x = var_5657_cast_fp16, y = rotr_27_cast_fp16)[name = tensor("op_5680_cast_fp16")]; + tensor koi_53_cast_fp16 = add(x = var_5679_cast_fp16, y = var_5680_cast_fp16)[name = tensor("koi_53_cast_fp16")]; + tensor qo_27_axis_0 = const()[name = tensor("qo_27_axis_0"), val = tensor(-1)]; + tensor qo_27_cast_fp16 = stack(axis = qo_27_axis_0, values = (qor_53_cast_fp16, qoi_53_cast_fp16))[name = tensor("qo_27_cast_fp16")]; + tensor ko_27_axis_0 = const()[name = tensor("ko_27_axis_0"), val = tensor(-1)]; + tensor ko_27_cast_fp16 = stack(axis = ko_27_axis_0, values = (kor_53_cast_fp16, koi_53_cast_fp16))[name = tensor("ko_27_cast_fp16")]; + tensor var_5713 = const()[name = tensor("op_5713"), val = tensor([1, 256, 16, 64])]; + tensor q_81_cast_fp16 = reshape(shape = var_5713, x = qo_27_cast_fp16)[name = tensor("q_81_cast_fp16")]; + tensor var_5719 = const()[name = tensor("op_5719"), val = tensor([1, 256, 16, 64])]; + tensor k_55_cast_fp16 = reshape(shape = var_5719, x = ko_27_cast_fp16)[name = tensor("k_55_cast_fp16")]; + tensor write_indices_float_79_cast_fp16 = add(x = var_5608_cast_fp16, y = var_393_promoted_to_fp16)[name = tensor("write_indices_float_79_cast_fp16")]; + tensor _inversed_5740_y_0_to_fp16 = const()[name = tensor("_inversed_5740_y_0_to_fp16"), val = tensor(0x1p-9)]; + tensor _inversed_5740_cast_fp16 = mul(x = write_indices_float_79_cast_fp16, y = _inversed_5740_y_0_to_fp16)[name = tensor("_inversed_5740_cast_fp16")]; + tensor var_5741_cast_fp16 = floor(x = _inversed_5740_cast_fp16)[name = tensor("op_5741_cast_fp16")]; + tensor var_5742_to_fp16 = const()[name = tensor("op_5742_to_fp16"), val = tensor(0x1p+9)]; + tensor var_5743_cast_fp16 = mul(x = var_5741_cast_fp16, y = var_5742_to_fp16)[name = tensor("op_5743_cast_fp16")]; + tensor write_indices_float_81_cast_fp16 = sub(x = write_indices_float_79_cast_fp16, y = var_5743_cast_fp16)[name = tensor("write_indices_float_81_cast_fp16")]; + tensor var_5761_value_0_to_fp16 = const()[name = tensor("op_5761_value_0_to_fp16"), val = tensor(0x1.ffp+8)]; + tensor var_5761_cast_fp16 = fill_like(ref_tensor = write_indices_float_81_cast_fp16, value = var_5761_value_0_to_fp16)[name = tensor("op_5761_cast_fp16")]; + tensor write_indices_float_83_cast_fp16 = select(a = write_indices_float_81_cast_fp16, b = var_5761_cast_fp16, cond = token_valid_1_cast_fp16)[name = tensor("write_indices_float_83_cast_fp16")]; + tensor var_5767_dtype_0 = const()[name = tensor("op_5767_dtype_0"), val = tensor("int32")]; + tensor var_5772 = const()[name = tensor("op_5772"), val = tensor([1, 256, 1, 1])]; + tensor write_indices_float_83_cast_fp16_to_int32 = cast(dtype = var_5767_dtype_0, x = write_indices_float_83_cast_fp16)[name = tensor("cast_595")]; + tensor var_5773 = reshape(shape = var_5772, x = write_indices_float_83_cast_fp16_to_int32)[name = tensor("op_5773")]; + tensor write_indices_27_reps_0 = const()[name = tensor("write_indices_27_reps_0"), val = tensor([1, 1, 16, 64])]; + tensor write_indices_27 = tile(reps = write_indices_27_reps_0, x = var_5773)[name = tensor("write_indices_27")]; + tensor var_5783_begin_0 = const()[name = tensor("op_5783_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_5783_end_0 = const()[name = tensor("op_5783_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor var_5783_end_mask_0 = const()[name = tensor("op_5783_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_5783_squeeze_mask_0 = const()[name = tensor("op_5783_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor cache13_to_fp16_dtype_0 = const()[name = tensor("cache13_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor cache13_to_fp16 = cast(dtype = cache13_to_fp16_dtype_0, x = cache13)[name = tensor("cast_594")]; + tensor var_5783_cast_fp16 = slice_by_index(begin = var_5783_begin_0, end = var_5783_end_0, end_mask = var_5783_end_mask_0, squeeze_mask = var_5783_squeeze_mask_0, x = cache13_to_fp16)[name = tensor("op_5783_cast_fp16")]; + tensor var_5785_axis_0 = const()[name = tensor("op_5785_axis_0"), val = tensor(1)]; + tensor var_5785_mode_0 = const()[name = tensor("op_5785_mode_0"), val = tensor("update")]; + tensor var_5785_validate_indices_0 = const()[name = tensor("op_5785_validate_indices_0"), val = tensor(false)]; + tensor var_5785_cast_fp16 = scatter_along_axis(axis = var_5785_axis_0, data = var_5783_cast_fp16, indices = write_indices_27, mode = var_5785_mode_0, updates = k_55_cast_fp16, validate_indices = var_5785_validate_indices_0)[name = tensor("op_5785_cast_fp16")]; + tensor concat_106 = const()[name = tensor("concat_106"), val = tensor([0, 0, 0, 0, 0])]; + tensor concat_107 = const()[name = tensor("concat_107"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_27_internal_tensor_assign_1_stride_0 = const()[name = tensor("new_cache_27_internal_tensor_assign_1_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_27_internal_tensor_assign_1_begin_mask_0 = const()[name = tensor("new_cache_27_internal_tensor_assign_1_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_27_internal_tensor_assign_1_end_mask_0 = const()[name = tensor("new_cache_27_internal_tensor_assign_1_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_27_internal_tensor_assign_1_squeeze_mask_0 = const()[name = tensor("new_cache_27_internal_tensor_assign_1_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_72 = const()[name = tensor("shape_72"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_26 = const()[name = tensor("reduce_prod_26"), val = tensor(1048576)]; + tensor range_1d_26_start_0 = const()[name = tensor("range_1d_26_start_0"), val = tensor(0)]; + tensor range_1d_26_step_0 = const()[name = tensor("range_1d_26_step_0"), val = tensor(1)]; + tensor range_1d_26 = range_1d(end = reduce_prod_26, start = range_1d_26_start_0, step = range_1d_26_step_0)[name = tensor("range_1d_26")]; + tensor reshape_130 = reshape(shape = shape_72, x = range_1d_26)[name = tensor("reshape_130")]; + tensor slice_by_index_26 = slice_by_index(begin = concat_106, begin_mask = new_cache_27_internal_tensor_assign_1_begin_mask_0, end = concat_107, end_mask = new_cache_27_internal_tensor_assign_1_end_mask_0, squeeze_mask = new_cache_27_internal_tensor_assign_1_squeeze_mask_0, stride = new_cache_27_internal_tensor_assign_1_stride_0, x = reshape_130)[name = tensor("slice_by_index_26")]; + tensor reshape_131_shape_0 = const()[name = tensor("reshape_131_shape_0"), val = tensor([-1])]; + tensor reshape_131 = reshape(shape = reshape_131_shape_0, x = slice_by_index_26)[name = tensor("reshape_131")]; + tensor reshape_132_shape_0 = const()[name = tensor("reshape_132_shape_0"), val = tensor([-1])]; + tensor reshape_132_cast_fp16 = reshape(shape = reshape_132_shape_0, x = var_5785_cast_fp16)[name = tensor("reshape_132_cast_fp16")]; + tensor reshape_133_shape_0 = const()[name = tensor("reshape_133_shape_0"), val = tensor([-1])]; + tensor reshape_133_cast_fp16 = reshape(shape = reshape_133_shape_0, x = cache13_to_fp16)[name = tensor("reshape_133_cast_fp16")]; + tensor scatter_26_mode_0 = const()[name = tensor("scatter_26_mode_0"), val = tensor("update")]; + tensor scatter_26_axis_0 = const()[name = tensor("scatter_26_axis_0"), val = tensor(0)]; + tensor scatter_26_validate_indices_0 = const()[name = tensor("scatter_26_validate_indices_0"), val = tensor(false)]; + tensor scatter_26_cast_fp16 = scatter(axis = scatter_26_axis_0, data = reshape_133_cast_fp16, indices = reshape_131, mode = scatter_26_mode_0, updates = reshape_132_cast_fp16, validate_indices = scatter_26_validate_indices_0)[name = tensor("scatter_26_cast_fp16")]; + tensor reshape_134_cast_fp16 = reshape(shape = shape_72, x = scatter_26_cast_fp16)[name = tensor("reshape_134_cast_fp16")]; + tensor var_5793_begin_0 = const()[name = tensor("op_5793_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor var_5793_end_0 = const()[name = tensor("op_5793_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor var_5793_end_mask_0 = const()[name = tensor("op_5793_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_5793_squeeze_mask_0 = const()[name = tensor("op_5793_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor var_5793_cast_fp16 = slice_by_index(begin = var_5793_begin_0, end = var_5793_end_0, end_mask = var_5793_end_mask_0, squeeze_mask = var_5793_squeeze_mask_0, x = reshape_134_cast_fp16)[name = tensor("op_5793_cast_fp16")]; + tensor var_5795_axis_0 = const()[name = tensor("op_5795_axis_0"), val = tensor(1)]; + tensor var_5795_mode_0 = const()[name = tensor("op_5795_mode_0"), val = tensor("update")]; + tensor var_5795_validate_indices_0 = const()[name = tensor("op_5795_validate_indices_0"), val = tensor(false)]; + tensor var_5795_cast_fp16 = scatter_along_axis(axis = var_5795_axis_0, data = var_5793_cast_fp16, indices = write_indices_27, mode = var_5795_mode_0, updates = v_27_cast_fp16, validate_indices = var_5795_validate_indices_0)[name = tensor("op_5795_cast_fp16")]; + tensor concat_108 = const()[name = tensor("concat_108"), val = tensor([1, 0, 0, 0, 0])]; + tensor concat_109 = const()[name = tensor("concat_109"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_27_internal_tensor_assign_2_stride_0 = const()[name = tensor("new_cache_27_internal_tensor_assign_2_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_27_internal_tensor_assign_2_begin_mask_0 = const()[name = tensor("new_cache_27_internal_tensor_assign_2_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_27_internal_tensor_assign_2_end_mask_0 = const()[name = tensor("new_cache_27_internal_tensor_assign_2_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_27_internal_tensor_assign_2_squeeze_mask_0 = const()[name = tensor("new_cache_27_internal_tensor_assign_2_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_73 = const()[name = tensor("shape_73"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_27 = const()[name = tensor("reduce_prod_27"), val = tensor(1048576)]; + tensor range_1d_27_start_0 = const()[name = tensor("range_1d_27_start_0"), val = tensor(0)]; + tensor range_1d_27_step_0 = const()[name = tensor("range_1d_27_step_0"), val = tensor(1)]; + tensor range_1d_27 = range_1d(end = reduce_prod_27, start = range_1d_27_start_0, step = range_1d_27_step_0)[name = tensor("range_1d_27")]; + tensor reshape_135 = reshape(shape = shape_73, x = range_1d_27)[name = tensor("reshape_135")]; + tensor slice_by_index_27 = slice_by_index(begin = concat_108, begin_mask = new_cache_27_internal_tensor_assign_2_begin_mask_0, end = concat_109, end_mask = new_cache_27_internal_tensor_assign_2_end_mask_0, squeeze_mask = new_cache_27_internal_tensor_assign_2_squeeze_mask_0, stride = new_cache_27_internal_tensor_assign_2_stride_0, x = reshape_135)[name = tensor("slice_by_index_27")]; + tensor reshape_136_shape_0 = const()[name = tensor("reshape_136_shape_0"), val = tensor([-1])]; + tensor reshape_136 = reshape(shape = reshape_136_shape_0, x = slice_by_index_27)[name = tensor("reshape_136")]; + tensor reshape_137_shape_0 = const()[name = tensor("reshape_137_shape_0"), val = tensor([-1])]; + tensor reshape_137_cast_fp16 = reshape(shape = reshape_137_shape_0, x = var_5795_cast_fp16)[name = tensor("reshape_137_cast_fp16")]; + tensor reshape_138_shape_0 = const()[name = tensor("reshape_138_shape_0"), val = tensor([-1])]; + tensor reshape_138_cast_fp16 = reshape(shape = reshape_138_shape_0, x = reshape_134_cast_fp16)[name = tensor("reshape_138_cast_fp16")]; + tensor scatter_27_mode_0 = const()[name = tensor("scatter_27_mode_0"), val = tensor("update")]; + tensor scatter_27_axis_0 = const()[name = tensor("scatter_27_axis_0"), val = tensor(0)]; + tensor scatter_27_validate_indices_0 = const()[name = tensor("scatter_27_validate_indices_0"), val = tensor(false)]; + tensor scatter_27_cast_fp16 = scatter(axis = scatter_27_axis_0, data = reshape_138_cast_fp16, indices = reshape_136, mode = scatter_27_mode_0, updates = reshape_137_cast_fp16, validate_indices = scatter_27_validate_indices_0)[name = tensor("scatter_27_cast_fp16")]; + tensor reshape_139_cast_fp16 = reshape(shape = shape_73, x = scatter_27_cast_fp16)[name = tensor("reshape_139_cast_fp16")]; + tensor reshape_139_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("reshape_139_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor keys_79_begin_0 = const()[name = tensor("keys_79_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor keys_79_end_0 = const()[name = tensor("keys_79_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor keys_79_end_mask_0 = const()[name = tensor("keys_79_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor keys_79_squeeze_mask_0 = const()[name = tensor("keys_79_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor keys_79_cast_fp16 = slice_by_index(begin = keys_79_begin_0, end = keys_79_end_0, end_mask = keys_79_end_mask_0, squeeze_mask = keys_79_squeeze_mask_0, x = reshape_139_cast_fp16)[name = tensor("keys_79_cast_fp16")]; + tensor values_79_begin_0 = const()[name = tensor("values_79_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor values_79_end_0 = const()[name = tensor("values_79_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor values_79_end_mask_0 = const()[name = tensor("values_79_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor values_79_squeeze_mask_0 = const()[name = tensor("values_79_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor values_79_cast_fp16 = slice_by_index(begin = values_79_begin_0, end = values_79_end_0, end_mask = values_79_end_mask_0, squeeze_mask = values_79_squeeze_mask_0, x = reshape_139_cast_fp16)[name = tensor("values_79_cast_fp16")]; + tensor var_5807_cast_fp16 = not_equal(x = keys_79_cast_fp16, y = keys_79_cast_fp16)[name = tensor("op_5807_cast_fp16")]; + tensor keys_81_cast_fp16 = select(a = var_470_to_fp16, b = keys_79_cast_fp16, cond = var_5807_cast_fp16)[name = tensor("keys_81_cast_fp16")]; + tensor var_5815_cast_fp16 = not_equal(x = values_79_cast_fp16, y = values_79_cast_fp16)[name = tensor("op_5815_cast_fp16")]; + tensor values_81_cast_fp16 = select(a = var_470_to_fp16, b = values_79_cast_fp16, cond = var_5815_cast_fp16)[name = tensor("values_81_cast_fp16")]; + tensor var_5839 = const()[name = tensor("op_5839"), val = tensor([0, 2, 1, 3])]; + tensor var_5855 = const()[name = tensor("op_5855"), val = tensor([1, 1, 1])]; + tensor var_5856_cast_fp16 = reshape(shape = var_5855, x = position13_to_fp16)[name = tensor("op_5856_cast_fp16")]; + tensor q_positions_27_cast_fp16 = add(x = var_5856_cast_fp16, y = q_offsets_1_promoted_to_fp16)[name = tensor("q_positions_27_cast_fp16")]; + tensor valid_end_27_cast_fp16 = add(x = var_5856_cast_fp16, y = var_536_cast_fp16)[name = tensor("valid_end_27_cast_fp16")]; + tensor valid_mask_27_cast_fp16 = less(x = k_positions_1_promoted_to_fp16, y = valid_end_27_cast_fp16)[name = tensor("valid_mask_27_cast_fp16")]; + tensor causal_mask_27_cast_fp16 = less_equal(x = k_positions_1_promoted_to_fp16, y = q_positions_27_cast_fp16)[name = tensor("causal_mask_27_cast_fp16")]; + tensor var_5884 = logical_and(x = valid_mask_27_cast_fp16, y = causal_mask_27_cast_fp16)[name = tensor("op_5884")]; + tensor attn_mask_27_axes_0 = const()[name = tensor("attn_mask_27_axes_0"), val = tensor([1])]; + tensor attn_mask_27 = expand_dims(axes = attn_mask_27_axes_0, x = var_5884)[name = tensor("attn_mask_27")]; + tensor var_5898_transpose_x_0 = const()[name = tensor("op_5898_transpose_x_0"), val = tensor(false)]; + tensor var_5898_transpose_y_0 = const()[name = tensor("op_5898_transpose_y_0"), val = tensor(false)]; + tensor transpose_95_perm_0 = const()[name = tensor("transpose_95_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_96_perm_0 = const()[name = tensor("transpose_96_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_96 = transpose(perm = transpose_96_perm_0, x = keys_81_cast_fp16)[name = tensor("transpose_152")]; + tensor transpose_95 = transpose(perm = transpose_95_perm_0, x = q_81_cast_fp16)[name = tensor("transpose_153")]; + tensor var_5898_cast_fp16 = matmul(transpose_x = var_5898_transpose_x_0, transpose_y = var_5898_transpose_y_0, x = transpose_95, y = transpose_96)[name = tensor("op_5898_cast_fp16")]; + tensor var_5892_to_fp16 = const()[name = tensor("op_5892_to_fp16"), val = tensor([0x1p-3])]; + tensor attn_weights_79_cast_fp16 = mul(x = var_5898_cast_fp16, y = var_5892_to_fp16)[name = tensor("attn_weights_79_cast_fp16")]; + tensor var_5900 = logical_not(x = attn_mask_27)[name = tensor("op_5900")]; + tensor var_5901_to_fp16 = const()[name = tensor("op_5901_to_fp16"), val = tensor(-inf)]; + tensor attn_weights_81_cast_fp16 = select(a = var_5901_to_fp16, b = attn_weights_79_cast_fp16, cond = var_5900)[name = tensor("attn_weights_81_cast_fp16")]; + tensor var_5903 = const()[name = tensor("op_5903"), val = tensor(-1)]; + tensor attn_weights_83_cast_fp16 = softmax(axis = var_5903, x = attn_weights_81_cast_fp16)[name = tensor("attn_weights_83_cast_fp16")]; + tensor attn_output_27_transpose_x_0 = const()[name = tensor("attn_output_27_transpose_x_0"), val = tensor(false)]; + tensor attn_output_27_transpose_y_0 = const()[name = tensor("attn_output_27_transpose_y_0"), val = tensor(false)]; + tensor values_83_cast_fp16 = transpose(perm = var_5839, x = values_81_cast_fp16)[name = tensor("transpose_154")]; + tensor attn_output_27_cast_fp16 = matmul(transpose_x = attn_output_27_transpose_x_0, transpose_y = attn_output_27_transpose_y_0, x = attn_weights_83_cast_fp16, y = values_83_cast_fp16)[name = tensor("attn_output_27_cast_fp16")]; + tensor var_5911 = const()[name = tensor("op_5911"), val = tensor([0, 2, 1, 3])]; + tensor var_5916 = const()[name = tensor("op_5916"), val = tensor([1, 256, 1024])]; + tensor var_5912_cast_fp16 = transpose(perm = var_5911, x = attn_output_27_cast_fp16)[name = tensor("transpose_151")]; + tensor input_161_cast_fp16 = reshape(shape = var_5916, x = var_5912_cast_fp16)[name = tensor("input_161_cast_fp16")]; + tensor attn13_out_proj_weight_to_fp16 = const()[name = tensor("attn13_out_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(334634496)))]; + tensor linear_53_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = attn13_out_proj_weight_to_fp16, x = input_161_cast_fp16)[name = tensor("linear_53_cast_fp16")]; + tensor var_5925_cast_fp16 = add(x = position13_to_fp16, y = valid_len_to_fp16)[name = tensor("op_5925_cast_fp16")]; + tensor var_5925_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_5925_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor input_163_cast_fp16 = add(x = input_157_cast_fp16, y = linear_53_cast_fp16)[name = tensor("input_163_cast_fp16")]; + tensor input_165_axes_0 = const()[name = tensor("input_165_axes_0"), val = tensor([-1])]; + tensor norm13_2_weight_to_fp16 = const()[name = tensor("norm13_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(336731712)))]; + tensor norm13_2_bias_to_fp16 = const()[name = tensor("norm13_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(336733824)))]; + tensor var_5929_to_fp16 = const()[name = tensor("op_5929_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_165_cast_fp16 = layer_norm(axes = input_165_axes_0, beta = norm13_2_bias_to_fp16, epsilon = var_5929_to_fp16, gamma = norm13_2_weight_to_fp16, x = input_163_cast_fp16)[name = tensor("input_165_cast_fp16")]; + tensor linear13_1_weight_to_fp16 = const()[name = tensor("linear13_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(336735936)))]; + tensor linear_54_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = linear13_1_weight_to_fp16, x = input_165_cast_fp16)[name = tensor("linear_54_cast_fp16")]; + tensor input_167_mode_0 = const()[name = tensor("input_167_mode_0"), val = tensor("EXACT")]; + tensor input_167_cast_fp16 = gelu(mode = input_167_mode_0, x = linear_54_cast_fp16)[name = tensor("input_167_cast_fp16")]; + tensor linear13_2_weight_to_fp16 = const()[name = tensor("linear13_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(345124608)))]; + tensor linear_55_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = linear13_2_weight_to_fp16, x = input_167_cast_fp16)[name = tensor("linear_55_cast_fp16")]; + tensor input_169_cast_fp16 = add(x = input_163_cast_fp16, y = linear_55_cast_fp16)[name = tensor("input_169_cast_fp16")]; + tensor input_171_axes_0 = const()[name = tensor("input_171_axes_0"), val = tensor([-1])]; + tensor norm14_1_weight_to_fp16 = const()[name = tensor("norm14_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(353513280)))]; + tensor norm14_1_bias_to_fp16 = const()[name = tensor("norm14_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(353515392)))]; + tensor var_5946_to_fp16 = const()[name = tensor("op_5946_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_171_cast_fp16 = layer_norm(axes = input_171_axes_0, beta = norm14_1_bias_to_fp16, epsilon = var_5946_to_fp16, gamma = norm14_1_weight_to_fp16, x = input_169_cast_fp16)[name = tensor("input_171_cast_fp16")]; + tensor attn14_in_proj_weight_to_fp16 = const()[name = tensor("attn14_in_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(353517504)))]; + tensor linear_56_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = attn14_in_proj_weight_to_fp16, x = input_171_cast_fp16)[name = tensor("linear_56_cast_fp16")]; + tensor var_5960 = const()[name = tensor("op_5960"), val = tensor([1, 256, 3, 16, 64])]; + tensor qkv_29_cast_fp16 = reshape(shape = var_5960, x = linear_56_cast_fp16)[name = tensor("qkv_29_cast_fp16")]; + tensor q_85_begin_0 = const()[name = tensor("q_85_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor q_85_end_0 = const()[name = tensor("q_85_end_0"), val = tensor([1, 256, 1, 16, 64])]; + tensor q_85_end_mask_0 = const()[name = tensor("q_85_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor q_85_squeeze_mask_0 = const()[name = tensor("q_85_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor q_85_cast_fp16 = slice_by_index(begin = q_85_begin_0, end = q_85_end_0, end_mask = q_85_end_mask_0, squeeze_mask = q_85_squeeze_mask_0, x = qkv_29_cast_fp16)[name = tensor("q_85_cast_fp16")]; + tensor k_57_begin_0 = const()[name = tensor("k_57_begin_0"), val = tensor([0, 0, 1, 0, 0])]; + tensor k_57_end_0 = const()[name = tensor("k_57_end_0"), val = tensor([1, 256, 2, 16, 64])]; + tensor k_57_end_mask_0 = const()[name = tensor("k_57_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor k_57_squeeze_mask_0 = const()[name = tensor("k_57_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor k_57_cast_fp16 = slice_by_index(begin = k_57_begin_0, end = k_57_end_0, end_mask = k_57_end_mask_0, squeeze_mask = k_57_squeeze_mask_0, x = qkv_29_cast_fp16)[name = tensor("k_57_cast_fp16")]; + tensor v_29_begin_0 = const()[name = tensor("v_29_begin_0"), val = tensor([0, 0, 2, 0, 0])]; + tensor v_29_end_0 = const()[name = tensor("v_29_end_0"), val = tensor([1, 256, 3, 16, 64])]; + tensor v_29_end_mask_0 = const()[name = tensor("v_29_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor v_29_squeeze_mask_0 = const()[name = tensor("v_29_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor v_29_cast_fp16 = slice_by_index(begin = v_29_begin_0, end = v_29_end_0, end_mask = v_29_end_mask_0, squeeze_mask = v_29_squeeze_mask_0, x = qkv_29_cast_fp16)[name = tensor("v_29_cast_fp16")]; + tensor var_6018 = const()[name = tensor("op_6018"), val = tensor([1, 1])]; + tensor position14_to_fp16_dtype_0 = const()[name = tensor("position14_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor position14_to_fp16 = cast(dtype = position14_to_fp16_dtype_0, x = position14)[name = tensor("cast_591")]; + tensor var_6019_cast_fp16 = reshape(shape = var_6018, x = position14_to_fp16)[name = tensor("op_6019_cast_fp16")]; + tensor ts_87_cast_fp16 = add(x = ts_1_promoted_to_fp16, y = var_6019_cast_fp16)[name = tensor("ts_87_cast_fp16")]; + tensor var_6026 = const()[name = tensor("op_6026"), val = tensor([1, 256, 1, 1])]; + tensor ts_89_cast_fp16 = reshape(shape = var_6026, x = ts_87_cast_fp16)[name = tensor("ts_89_cast_fp16")]; + tensor var_6033 = const()[name = tensor("op_6033"), val = tensor([1, 256, 16, 32, 2])]; + tensor q_complex_29_cast_fp16 = reshape(shape = var_6033, x = q_85_cast_fp16)[name = tensor("q_complex_29_cast_fp16")]; + tensor var_6040 = const()[name = tensor("op_6040"), val = tensor([1, 256, 16, 32, 2])]; + tensor k_complex_29_cast_fp16 = reshape(shape = var_6040, x = k_57_cast_fp16)[name = tensor("k_complex_29_cast_fp16")]; + tensor var_6044_begin_0 = const()[name = tensor("op_6044_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_6044_end_0 = const()[name = tensor("op_6044_end_0"), val = tensor([1, 256, 16, 32, 1])]; + tensor var_6044_end_mask_0 = const()[name = tensor("op_6044_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_6044_squeeze_mask_0 = const()[name = tensor("op_6044_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_6044_cast_fp16 = slice_by_index(begin = var_6044_begin_0, end = var_6044_end_0, end_mask = var_6044_end_mask_0, squeeze_mask = var_6044_squeeze_mask_0, x = q_complex_29_cast_fp16)[name = tensor("op_6044_cast_fp16")]; + tensor var_6052_begin_0 = const()[name = tensor("op_6052_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_6052_end_0 = const()[name = tensor("op_6052_end_0"), val = tensor([1, 256, 16, 32, 2])]; + tensor var_6052_end_mask_0 = const()[name = tensor("op_6052_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_6052_squeeze_mask_0 = const()[name = tensor("op_6052_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_6052_cast_fp16 = slice_by_index(begin = var_6052_begin_0, end = var_6052_end_0, end_mask = var_6052_end_mask_0, squeeze_mask = var_6052_squeeze_mask_0, x = q_complex_29_cast_fp16)[name = tensor("op_6052_cast_fp16")]; + tensor var_6060_begin_0 = const()[name = tensor("op_6060_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_6060_end_0 = const()[name = tensor("op_6060_end_0"), val = tensor([1, 256, 16, 32, 1])]; + tensor var_6060_end_mask_0 = const()[name = tensor("op_6060_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_6060_squeeze_mask_0 = const()[name = tensor("op_6060_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_6060_cast_fp16 = slice_by_index(begin = var_6060_begin_0, end = var_6060_end_0, end_mask = var_6060_end_mask_0, squeeze_mask = var_6060_squeeze_mask_0, x = k_complex_29_cast_fp16)[name = tensor("op_6060_cast_fp16")]; + tensor var_6068_begin_0 = const()[name = tensor("op_6068_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_6068_end_0 = const()[name = tensor("op_6068_end_0"), val = tensor([1, 256, 16, 32, 2])]; + tensor var_6068_end_mask_0 = const()[name = tensor("op_6068_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_6068_squeeze_mask_0 = const()[name = tensor("op_6068_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_6068_cast_fp16 = slice_by_index(begin = var_6068_begin_0, end = var_6068_end_0, end_mask = var_6068_end_mask_0, squeeze_mask = var_6068_squeeze_mask_0, x = k_complex_29_cast_fp16)[name = tensor("op_6068_cast_fp16")]; + tensor freqs_29_to_fp16 = const()[name = tensor("freqs_29_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(359809024)))]; + tensor var_6074_cast_fp16 = mul(x = freqs_29_to_fp16, y = ts_89_cast_fp16)[name = tensor("op_6074_cast_fp16")]; + tensor rotr_29_cast_fp16 = cos(x = var_6074_cast_fp16)[name = tensor("rotr_29_cast_fp16")]; + tensor roti_29_cast_fp16 = sin(x = var_6074_cast_fp16)[name = tensor("roti_29_cast_fp16")]; + tensor var_6078_cast_fp16 = mul(x = var_6044_cast_fp16, y = rotr_29_cast_fp16)[name = tensor("op_6078_cast_fp16")]; + tensor var_6079_cast_fp16 = mul(x = var_6052_cast_fp16, y = roti_29_cast_fp16)[name = tensor("op_6079_cast_fp16")]; + tensor qor_57_cast_fp16 = sub(x = var_6078_cast_fp16, y = var_6079_cast_fp16)[name = tensor("qor_57_cast_fp16")]; + tensor var_6082_cast_fp16 = mul(x = var_6044_cast_fp16, y = roti_29_cast_fp16)[name = tensor("op_6082_cast_fp16")]; + tensor var_6083_cast_fp16 = mul(x = var_6052_cast_fp16, y = rotr_29_cast_fp16)[name = tensor("op_6083_cast_fp16")]; + tensor qoi_57_cast_fp16 = add(x = var_6082_cast_fp16, y = var_6083_cast_fp16)[name = tensor("qoi_57_cast_fp16")]; + tensor var_6086_cast_fp16 = mul(x = var_6060_cast_fp16, y = rotr_29_cast_fp16)[name = tensor("op_6086_cast_fp16")]; + tensor var_6087_cast_fp16 = mul(x = var_6068_cast_fp16, y = roti_29_cast_fp16)[name = tensor("op_6087_cast_fp16")]; + tensor kor_57_cast_fp16 = sub(x = var_6086_cast_fp16, y = var_6087_cast_fp16)[name = tensor("kor_57_cast_fp16")]; + tensor var_6090_cast_fp16 = mul(x = var_6060_cast_fp16, y = roti_29_cast_fp16)[name = tensor("op_6090_cast_fp16")]; + tensor var_6091_cast_fp16 = mul(x = var_6068_cast_fp16, y = rotr_29_cast_fp16)[name = tensor("op_6091_cast_fp16")]; + tensor koi_57_cast_fp16 = add(x = var_6090_cast_fp16, y = var_6091_cast_fp16)[name = tensor("koi_57_cast_fp16")]; + tensor qo_29_axis_0 = const()[name = tensor("qo_29_axis_0"), val = tensor(-1)]; + tensor qo_29_cast_fp16 = stack(axis = qo_29_axis_0, values = (qor_57_cast_fp16, qoi_57_cast_fp16))[name = tensor("qo_29_cast_fp16")]; + tensor ko_29_axis_0 = const()[name = tensor("ko_29_axis_0"), val = tensor(-1)]; + tensor ko_29_cast_fp16 = stack(axis = ko_29_axis_0, values = (kor_57_cast_fp16, koi_57_cast_fp16))[name = tensor("ko_29_cast_fp16")]; + tensor var_6124 = const()[name = tensor("op_6124"), val = tensor([1, 256, 16, 64])]; + tensor q_87_cast_fp16 = reshape(shape = var_6124, x = qo_29_cast_fp16)[name = tensor("q_87_cast_fp16")]; + tensor var_6130 = const()[name = tensor("op_6130"), val = tensor([1, 256, 16, 64])]; + tensor k_59_cast_fp16 = reshape(shape = var_6130, x = ko_29_cast_fp16)[name = tensor("k_59_cast_fp16")]; + tensor write_indices_float_85_cast_fp16 = add(x = var_6019_cast_fp16, y = var_393_promoted_to_fp16)[name = tensor("write_indices_float_85_cast_fp16")]; + tensor _inversed_6151_y_0_to_fp16 = const()[name = tensor("_inversed_6151_y_0_to_fp16"), val = tensor(0x1p-9)]; + tensor _inversed_6151_cast_fp16 = mul(x = write_indices_float_85_cast_fp16, y = _inversed_6151_y_0_to_fp16)[name = tensor("_inversed_6151_cast_fp16")]; + tensor var_6152_cast_fp16 = floor(x = _inversed_6151_cast_fp16)[name = tensor("op_6152_cast_fp16")]; + tensor var_6153_to_fp16 = const()[name = tensor("op_6153_to_fp16"), val = tensor(0x1p+9)]; + tensor var_6154_cast_fp16 = mul(x = var_6152_cast_fp16, y = var_6153_to_fp16)[name = tensor("op_6154_cast_fp16")]; + tensor write_indices_float_87_cast_fp16 = sub(x = write_indices_float_85_cast_fp16, y = var_6154_cast_fp16)[name = tensor("write_indices_float_87_cast_fp16")]; + tensor var_6172_value_0_to_fp16 = const()[name = tensor("op_6172_value_0_to_fp16"), val = tensor(0x1.ffp+8)]; + tensor var_6172_cast_fp16 = fill_like(ref_tensor = write_indices_float_87_cast_fp16, value = var_6172_value_0_to_fp16)[name = tensor("op_6172_cast_fp16")]; + tensor write_indices_float_89_cast_fp16 = select(a = write_indices_float_87_cast_fp16, b = var_6172_cast_fp16, cond = token_valid_1_cast_fp16)[name = tensor("write_indices_float_89_cast_fp16")]; + tensor var_6178_dtype_0 = const()[name = tensor("op_6178_dtype_0"), val = tensor("int32")]; + tensor var_6183 = const()[name = tensor("op_6183"), val = tensor([1, 256, 1, 1])]; + tensor write_indices_float_89_cast_fp16_to_int32 = cast(dtype = var_6178_dtype_0, x = write_indices_float_89_cast_fp16)[name = tensor("cast_590")]; + tensor var_6184 = reshape(shape = var_6183, x = write_indices_float_89_cast_fp16_to_int32)[name = tensor("op_6184")]; + tensor write_indices_29_reps_0 = const()[name = tensor("write_indices_29_reps_0"), val = tensor([1, 1, 16, 64])]; + tensor write_indices_29 = tile(reps = write_indices_29_reps_0, x = var_6184)[name = tensor("write_indices_29")]; + tensor var_6194_begin_0 = const()[name = tensor("op_6194_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_6194_end_0 = const()[name = tensor("op_6194_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor var_6194_end_mask_0 = const()[name = tensor("op_6194_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_6194_squeeze_mask_0 = const()[name = tensor("op_6194_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor cache14_to_fp16_dtype_0 = const()[name = tensor("cache14_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor cache14_to_fp16 = cast(dtype = cache14_to_fp16_dtype_0, x = cache14)[name = tensor("cast_589")]; + tensor var_6194_cast_fp16 = slice_by_index(begin = var_6194_begin_0, end = var_6194_end_0, end_mask = var_6194_end_mask_0, squeeze_mask = var_6194_squeeze_mask_0, x = cache14_to_fp16)[name = tensor("op_6194_cast_fp16")]; + tensor var_6196_axis_0 = const()[name = tensor("op_6196_axis_0"), val = tensor(1)]; + tensor var_6196_mode_0 = const()[name = tensor("op_6196_mode_0"), val = tensor("update")]; + tensor var_6196_validate_indices_0 = const()[name = tensor("op_6196_validate_indices_0"), val = tensor(false)]; + tensor var_6196_cast_fp16 = scatter_along_axis(axis = var_6196_axis_0, data = var_6194_cast_fp16, indices = write_indices_29, mode = var_6196_mode_0, updates = k_59_cast_fp16, validate_indices = var_6196_validate_indices_0)[name = tensor("op_6196_cast_fp16")]; + tensor concat_114 = const()[name = tensor("concat_114"), val = tensor([0, 0, 0, 0, 0])]; + tensor concat_115 = const()[name = tensor("concat_115"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_29_internal_tensor_assign_1_stride_0 = const()[name = tensor("new_cache_29_internal_tensor_assign_1_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_29_internal_tensor_assign_1_begin_mask_0 = const()[name = tensor("new_cache_29_internal_tensor_assign_1_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_29_internal_tensor_assign_1_end_mask_0 = const()[name = tensor("new_cache_29_internal_tensor_assign_1_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_29_internal_tensor_assign_1_squeeze_mask_0 = const()[name = tensor("new_cache_29_internal_tensor_assign_1_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_74 = const()[name = tensor("shape_74"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_28 = const()[name = tensor("reduce_prod_28"), val = tensor(1048576)]; + tensor range_1d_28_start_0 = const()[name = tensor("range_1d_28_start_0"), val = tensor(0)]; + tensor range_1d_28_step_0 = const()[name = tensor("range_1d_28_step_0"), val = tensor(1)]; + tensor range_1d_28 = range_1d(end = reduce_prod_28, start = range_1d_28_start_0, step = range_1d_28_step_0)[name = tensor("range_1d_28")]; + tensor reshape_140 = reshape(shape = shape_74, x = range_1d_28)[name = tensor("reshape_140")]; + tensor slice_by_index_28 = slice_by_index(begin = concat_114, begin_mask = new_cache_29_internal_tensor_assign_1_begin_mask_0, end = concat_115, end_mask = new_cache_29_internal_tensor_assign_1_end_mask_0, squeeze_mask = new_cache_29_internal_tensor_assign_1_squeeze_mask_0, stride = new_cache_29_internal_tensor_assign_1_stride_0, x = reshape_140)[name = tensor("slice_by_index_28")]; + tensor reshape_141_shape_0 = const()[name = tensor("reshape_141_shape_0"), val = tensor([-1])]; + tensor reshape_141 = reshape(shape = reshape_141_shape_0, x = slice_by_index_28)[name = tensor("reshape_141")]; + tensor reshape_142_shape_0 = const()[name = tensor("reshape_142_shape_0"), val = tensor([-1])]; + tensor reshape_142_cast_fp16 = reshape(shape = reshape_142_shape_0, x = var_6196_cast_fp16)[name = tensor("reshape_142_cast_fp16")]; + tensor reshape_143_shape_0 = const()[name = tensor("reshape_143_shape_0"), val = tensor([-1])]; + tensor reshape_143_cast_fp16 = reshape(shape = reshape_143_shape_0, x = cache14_to_fp16)[name = tensor("reshape_143_cast_fp16")]; + tensor scatter_28_mode_0 = const()[name = tensor("scatter_28_mode_0"), val = tensor("update")]; + tensor scatter_28_axis_0 = const()[name = tensor("scatter_28_axis_0"), val = tensor(0)]; + tensor scatter_28_validate_indices_0 = const()[name = tensor("scatter_28_validate_indices_0"), val = tensor(false)]; + tensor scatter_28_cast_fp16 = scatter(axis = scatter_28_axis_0, data = reshape_143_cast_fp16, indices = reshape_141, mode = scatter_28_mode_0, updates = reshape_142_cast_fp16, validate_indices = scatter_28_validate_indices_0)[name = tensor("scatter_28_cast_fp16")]; + tensor reshape_144_cast_fp16 = reshape(shape = shape_74, x = scatter_28_cast_fp16)[name = tensor("reshape_144_cast_fp16")]; + tensor var_6204_begin_0 = const()[name = tensor("op_6204_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor var_6204_end_0 = const()[name = tensor("op_6204_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor var_6204_end_mask_0 = const()[name = tensor("op_6204_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_6204_squeeze_mask_0 = const()[name = tensor("op_6204_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor var_6204_cast_fp16 = slice_by_index(begin = var_6204_begin_0, end = var_6204_end_0, end_mask = var_6204_end_mask_0, squeeze_mask = var_6204_squeeze_mask_0, x = reshape_144_cast_fp16)[name = tensor("op_6204_cast_fp16")]; + tensor var_6206_axis_0 = const()[name = tensor("op_6206_axis_0"), val = tensor(1)]; + tensor var_6206_mode_0 = const()[name = tensor("op_6206_mode_0"), val = tensor("update")]; + tensor var_6206_validate_indices_0 = const()[name = tensor("op_6206_validate_indices_0"), val = tensor(false)]; + tensor var_6206_cast_fp16 = scatter_along_axis(axis = var_6206_axis_0, data = var_6204_cast_fp16, indices = write_indices_29, mode = var_6206_mode_0, updates = v_29_cast_fp16, validate_indices = var_6206_validate_indices_0)[name = tensor("op_6206_cast_fp16")]; + tensor concat_116 = const()[name = tensor("concat_116"), val = tensor([1, 0, 0, 0, 0])]; + tensor concat_117 = const()[name = tensor("concat_117"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_29_internal_tensor_assign_2_stride_0 = const()[name = tensor("new_cache_29_internal_tensor_assign_2_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_29_internal_tensor_assign_2_begin_mask_0 = const()[name = tensor("new_cache_29_internal_tensor_assign_2_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_29_internal_tensor_assign_2_end_mask_0 = const()[name = tensor("new_cache_29_internal_tensor_assign_2_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_29_internal_tensor_assign_2_squeeze_mask_0 = const()[name = tensor("new_cache_29_internal_tensor_assign_2_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_75 = const()[name = tensor("shape_75"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_29 = const()[name = tensor("reduce_prod_29"), val = tensor(1048576)]; + tensor range_1d_29_start_0 = const()[name = tensor("range_1d_29_start_0"), val = tensor(0)]; + tensor range_1d_29_step_0 = const()[name = tensor("range_1d_29_step_0"), val = tensor(1)]; + tensor range_1d_29 = range_1d(end = reduce_prod_29, start = range_1d_29_start_0, step = range_1d_29_step_0)[name = tensor("range_1d_29")]; + tensor reshape_145 = reshape(shape = shape_75, x = range_1d_29)[name = tensor("reshape_145")]; + tensor slice_by_index_29 = slice_by_index(begin = concat_116, begin_mask = new_cache_29_internal_tensor_assign_2_begin_mask_0, end = concat_117, end_mask = new_cache_29_internal_tensor_assign_2_end_mask_0, squeeze_mask = new_cache_29_internal_tensor_assign_2_squeeze_mask_0, stride = new_cache_29_internal_tensor_assign_2_stride_0, x = reshape_145)[name = tensor("slice_by_index_29")]; + tensor reshape_146_shape_0 = const()[name = tensor("reshape_146_shape_0"), val = tensor([-1])]; + tensor reshape_146 = reshape(shape = reshape_146_shape_0, x = slice_by_index_29)[name = tensor("reshape_146")]; + tensor reshape_147_shape_0 = const()[name = tensor("reshape_147_shape_0"), val = tensor([-1])]; + tensor reshape_147_cast_fp16 = reshape(shape = reshape_147_shape_0, x = var_6206_cast_fp16)[name = tensor("reshape_147_cast_fp16")]; + tensor reshape_148_shape_0 = const()[name = tensor("reshape_148_shape_0"), val = tensor([-1])]; + tensor reshape_148_cast_fp16 = reshape(shape = reshape_148_shape_0, x = reshape_144_cast_fp16)[name = tensor("reshape_148_cast_fp16")]; + tensor scatter_29_mode_0 = const()[name = tensor("scatter_29_mode_0"), val = tensor("update")]; + tensor scatter_29_axis_0 = const()[name = tensor("scatter_29_axis_0"), val = tensor(0)]; + tensor scatter_29_validate_indices_0 = const()[name = tensor("scatter_29_validate_indices_0"), val = tensor(false)]; + tensor scatter_29_cast_fp16 = scatter(axis = scatter_29_axis_0, data = reshape_148_cast_fp16, indices = reshape_146, mode = scatter_29_mode_0, updates = reshape_147_cast_fp16, validate_indices = scatter_29_validate_indices_0)[name = tensor("scatter_29_cast_fp16")]; + tensor reshape_149_cast_fp16 = reshape(shape = shape_75, x = scatter_29_cast_fp16)[name = tensor("reshape_149_cast_fp16")]; + tensor reshape_149_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("reshape_149_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor keys_85_begin_0 = const()[name = tensor("keys_85_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor keys_85_end_0 = const()[name = tensor("keys_85_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor keys_85_end_mask_0 = const()[name = tensor("keys_85_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor keys_85_squeeze_mask_0 = const()[name = tensor("keys_85_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor keys_85_cast_fp16 = slice_by_index(begin = keys_85_begin_0, end = keys_85_end_0, end_mask = keys_85_end_mask_0, squeeze_mask = keys_85_squeeze_mask_0, x = reshape_149_cast_fp16)[name = tensor("keys_85_cast_fp16")]; + tensor values_85_begin_0 = const()[name = tensor("values_85_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor values_85_end_0 = const()[name = tensor("values_85_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor values_85_end_mask_0 = const()[name = tensor("values_85_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor values_85_squeeze_mask_0 = const()[name = tensor("values_85_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor values_85_cast_fp16 = slice_by_index(begin = values_85_begin_0, end = values_85_end_0, end_mask = values_85_end_mask_0, squeeze_mask = values_85_squeeze_mask_0, x = reshape_149_cast_fp16)[name = tensor("values_85_cast_fp16")]; + tensor var_6218_cast_fp16 = not_equal(x = keys_85_cast_fp16, y = keys_85_cast_fp16)[name = tensor("op_6218_cast_fp16")]; + tensor keys_87_cast_fp16 = select(a = var_470_to_fp16, b = keys_85_cast_fp16, cond = var_6218_cast_fp16)[name = tensor("keys_87_cast_fp16")]; + tensor var_6226_cast_fp16 = not_equal(x = values_85_cast_fp16, y = values_85_cast_fp16)[name = tensor("op_6226_cast_fp16")]; + tensor values_87_cast_fp16 = select(a = var_470_to_fp16, b = values_85_cast_fp16, cond = var_6226_cast_fp16)[name = tensor("values_87_cast_fp16")]; + tensor var_6250 = const()[name = tensor("op_6250"), val = tensor([0, 2, 1, 3])]; + tensor var_6266 = const()[name = tensor("op_6266"), val = tensor([1, 1, 1])]; + tensor var_6267_cast_fp16 = reshape(shape = var_6266, x = position14_to_fp16)[name = tensor("op_6267_cast_fp16")]; + tensor q_positions_29_cast_fp16 = add(x = var_6267_cast_fp16, y = q_offsets_1_promoted_to_fp16)[name = tensor("q_positions_29_cast_fp16")]; + tensor valid_end_29_cast_fp16 = add(x = var_6267_cast_fp16, y = var_536_cast_fp16)[name = tensor("valid_end_29_cast_fp16")]; + tensor valid_mask_29_cast_fp16 = less(x = k_positions_1_promoted_to_fp16, y = valid_end_29_cast_fp16)[name = tensor("valid_mask_29_cast_fp16")]; + tensor causal_mask_29_cast_fp16 = less_equal(x = k_positions_1_promoted_to_fp16, y = q_positions_29_cast_fp16)[name = tensor("causal_mask_29_cast_fp16")]; + tensor var_6295 = logical_and(x = valid_mask_29_cast_fp16, y = causal_mask_29_cast_fp16)[name = tensor("op_6295")]; + tensor attn_mask_29_axes_0 = const()[name = tensor("attn_mask_29_axes_0"), val = tensor([1])]; + tensor attn_mask_29 = expand_dims(axes = attn_mask_29_axes_0, x = var_6295)[name = tensor("attn_mask_29")]; + tensor var_6309_transpose_x_0 = const()[name = tensor("op_6309_transpose_x_0"), val = tensor(false)]; + tensor var_6309_transpose_y_0 = const()[name = tensor("op_6309_transpose_y_0"), val = tensor(false)]; + tensor transpose_97_perm_0 = const()[name = tensor("transpose_97_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_98_perm_0 = const()[name = tensor("transpose_98_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_98 = transpose(perm = transpose_98_perm_0, x = keys_87_cast_fp16)[name = tensor("transpose_148")]; + tensor transpose_97 = transpose(perm = transpose_97_perm_0, x = q_87_cast_fp16)[name = tensor("transpose_149")]; + tensor var_6309_cast_fp16 = matmul(transpose_x = var_6309_transpose_x_0, transpose_y = var_6309_transpose_y_0, x = transpose_97, y = transpose_98)[name = tensor("op_6309_cast_fp16")]; + tensor var_6303_to_fp16 = const()[name = tensor("op_6303_to_fp16"), val = tensor([0x1p-3])]; + tensor attn_weights_85_cast_fp16 = mul(x = var_6309_cast_fp16, y = var_6303_to_fp16)[name = tensor("attn_weights_85_cast_fp16")]; + tensor var_6311 = logical_not(x = attn_mask_29)[name = tensor("op_6311")]; + tensor var_6312_to_fp16 = const()[name = tensor("op_6312_to_fp16"), val = tensor(-inf)]; + tensor attn_weights_87_cast_fp16 = select(a = var_6312_to_fp16, b = attn_weights_85_cast_fp16, cond = var_6311)[name = tensor("attn_weights_87_cast_fp16")]; + tensor var_6314 = const()[name = tensor("op_6314"), val = tensor(-1)]; + tensor attn_weights_89_cast_fp16 = softmax(axis = var_6314, x = attn_weights_87_cast_fp16)[name = tensor("attn_weights_89_cast_fp16")]; + tensor attn_output_29_transpose_x_0 = const()[name = tensor("attn_output_29_transpose_x_0"), val = tensor(false)]; + tensor attn_output_29_transpose_y_0 = const()[name = tensor("attn_output_29_transpose_y_0"), val = tensor(false)]; + tensor values_89_cast_fp16 = transpose(perm = var_6250, x = values_87_cast_fp16)[name = tensor("transpose_150")]; + tensor attn_output_29_cast_fp16 = matmul(transpose_x = attn_output_29_transpose_x_0, transpose_y = attn_output_29_transpose_y_0, x = attn_weights_89_cast_fp16, y = values_89_cast_fp16)[name = tensor("attn_output_29_cast_fp16")]; + tensor var_6322 = const()[name = tensor("op_6322"), val = tensor([0, 2, 1, 3])]; + tensor var_6327 = const()[name = tensor("op_6327"), val = tensor([1, 256, 1024])]; + tensor var_6323_cast_fp16 = transpose(perm = var_6322, x = attn_output_29_cast_fp16)[name = tensor("transpose_147")]; + tensor input_173_cast_fp16 = reshape(shape = var_6327, x = var_6323_cast_fp16)[name = tensor("input_173_cast_fp16")]; + tensor attn14_out_proj_weight_to_fp16 = const()[name = tensor("attn14_out_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(359809152)))]; + tensor linear_57_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = attn14_out_proj_weight_to_fp16, x = input_173_cast_fp16)[name = tensor("linear_57_cast_fp16")]; + tensor var_6336_cast_fp16 = add(x = position14_to_fp16, y = valid_len_to_fp16)[name = tensor("op_6336_cast_fp16")]; + tensor var_6336_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_6336_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor input_175_cast_fp16 = add(x = input_169_cast_fp16, y = linear_57_cast_fp16)[name = tensor("input_175_cast_fp16")]; + tensor input_177_axes_0 = const()[name = tensor("input_177_axes_0"), val = tensor([-1])]; + tensor norm14_2_weight_to_fp16 = const()[name = tensor("norm14_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(361906368)))]; + tensor norm14_2_bias_to_fp16 = const()[name = tensor("norm14_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(361908480)))]; + tensor var_6340_to_fp16 = const()[name = tensor("op_6340_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_177_cast_fp16 = layer_norm(axes = input_177_axes_0, beta = norm14_2_bias_to_fp16, epsilon = var_6340_to_fp16, gamma = norm14_2_weight_to_fp16, x = input_175_cast_fp16)[name = tensor("input_177_cast_fp16")]; + tensor linear14_1_weight_to_fp16 = const()[name = tensor("linear14_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(361910592)))]; + tensor linear_58_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = linear14_1_weight_to_fp16, x = input_177_cast_fp16)[name = tensor("linear_58_cast_fp16")]; + tensor input_179_mode_0 = const()[name = tensor("input_179_mode_0"), val = tensor("EXACT")]; + tensor input_179_cast_fp16 = gelu(mode = input_179_mode_0, x = linear_58_cast_fp16)[name = tensor("input_179_cast_fp16")]; + tensor linear14_2_weight_to_fp16 = const()[name = tensor("linear14_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(370299264)))]; + tensor linear_59_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = linear14_2_weight_to_fp16, x = input_179_cast_fp16)[name = tensor("linear_59_cast_fp16")]; + tensor input_181_cast_fp16 = add(x = input_175_cast_fp16, y = linear_59_cast_fp16)[name = tensor("input_181_cast_fp16")]; + tensor input_183_axes_0 = const()[name = tensor("input_183_axes_0"), val = tensor([-1])]; + tensor norm15_1_weight_to_fp16 = const()[name = tensor("norm15_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(378687936)))]; + tensor norm15_1_bias_to_fp16 = const()[name = tensor("norm15_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(378690048)))]; + tensor var_6357_to_fp16 = const()[name = tensor("op_6357_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_183_cast_fp16 = layer_norm(axes = input_183_axes_0, beta = norm15_1_bias_to_fp16, epsilon = var_6357_to_fp16, gamma = norm15_1_weight_to_fp16, x = input_181_cast_fp16)[name = tensor("input_183_cast_fp16")]; + tensor attn15_in_proj_weight_to_fp16 = const()[name = tensor("attn15_in_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(378692160)))]; + tensor linear_60_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = attn15_in_proj_weight_to_fp16, x = input_183_cast_fp16)[name = tensor("linear_60_cast_fp16")]; + tensor var_6371 = const()[name = tensor("op_6371"), val = tensor([1, 256, 3, 16, 64])]; + tensor qkv_31_cast_fp16 = reshape(shape = var_6371, x = linear_60_cast_fp16)[name = tensor("qkv_31_cast_fp16")]; + tensor q_91_begin_0 = const()[name = tensor("q_91_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor q_91_end_0 = const()[name = tensor("q_91_end_0"), val = tensor([1, 256, 1, 16, 64])]; + tensor q_91_end_mask_0 = const()[name = tensor("q_91_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor q_91_squeeze_mask_0 = const()[name = tensor("q_91_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor q_91_cast_fp16 = slice_by_index(begin = q_91_begin_0, end = q_91_end_0, end_mask = q_91_end_mask_0, squeeze_mask = q_91_squeeze_mask_0, x = qkv_31_cast_fp16)[name = tensor("q_91_cast_fp16")]; + tensor k_61_begin_0 = const()[name = tensor("k_61_begin_0"), val = tensor([0, 0, 1, 0, 0])]; + tensor k_61_end_0 = const()[name = tensor("k_61_end_0"), val = tensor([1, 256, 2, 16, 64])]; + tensor k_61_end_mask_0 = const()[name = tensor("k_61_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor k_61_squeeze_mask_0 = const()[name = tensor("k_61_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor k_61_cast_fp16 = slice_by_index(begin = k_61_begin_0, end = k_61_end_0, end_mask = k_61_end_mask_0, squeeze_mask = k_61_squeeze_mask_0, x = qkv_31_cast_fp16)[name = tensor("k_61_cast_fp16")]; + tensor v_31_begin_0 = const()[name = tensor("v_31_begin_0"), val = tensor([0, 0, 2, 0, 0])]; + tensor v_31_end_0 = const()[name = tensor("v_31_end_0"), val = tensor([1, 256, 3, 16, 64])]; + tensor v_31_end_mask_0 = const()[name = tensor("v_31_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor v_31_squeeze_mask_0 = const()[name = tensor("v_31_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor v_31_cast_fp16 = slice_by_index(begin = v_31_begin_0, end = v_31_end_0, end_mask = v_31_end_mask_0, squeeze_mask = v_31_squeeze_mask_0, x = qkv_31_cast_fp16)[name = tensor("v_31_cast_fp16")]; + tensor var_6429 = const()[name = tensor("op_6429"), val = tensor([1, 1])]; + tensor position15_to_fp16_dtype_0 = const()[name = tensor("position15_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor position15_to_fp16 = cast(dtype = position15_to_fp16_dtype_0, x = position15)[name = tensor("cast_586")]; + tensor var_6430_cast_fp16 = reshape(shape = var_6429, x = position15_to_fp16)[name = tensor("op_6430_cast_fp16")]; + tensor ts_93_cast_fp16 = add(x = ts_1_promoted_to_fp16, y = var_6430_cast_fp16)[name = tensor("ts_93_cast_fp16")]; + tensor var_6437 = const()[name = tensor("op_6437"), val = tensor([1, 256, 1, 1])]; + tensor ts_95_cast_fp16 = reshape(shape = var_6437, x = ts_93_cast_fp16)[name = tensor("ts_95_cast_fp16")]; + tensor var_6444 = const()[name = tensor("op_6444"), val = tensor([1, 256, 16, 32, 2])]; + tensor q_complex_31_cast_fp16 = reshape(shape = var_6444, x = q_91_cast_fp16)[name = tensor("q_complex_31_cast_fp16")]; + tensor var_6451 = const()[name = tensor("op_6451"), val = tensor([1, 256, 16, 32, 2])]; + tensor k_complex_31_cast_fp16 = reshape(shape = var_6451, x = k_61_cast_fp16)[name = tensor("k_complex_31_cast_fp16")]; + tensor var_6455_begin_0 = const()[name = tensor("op_6455_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_6455_end_0 = const()[name = tensor("op_6455_end_0"), val = tensor([1, 256, 16, 32, 1])]; + tensor var_6455_end_mask_0 = const()[name = tensor("op_6455_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_6455_squeeze_mask_0 = const()[name = tensor("op_6455_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_6455_cast_fp16 = slice_by_index(begin = var_6455_begin_0, end = var_6455_end_0, end_mask = var_6455_end_mask_0, squeeze_mask = var_6455_squeeze_mask_0, x = q_complex_31_cast_fp16)[name = tensor("op_6455_cast_fp16")]; + tensor var_6463_begin_0 = const()[name = tensor("op_6463_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_6463_end_0 = const()[name = tensor("op_6463_end_0"), val = tensor([1, 256, 16, 32, 2])]; + tensor var_6463_end_mask_0 = const()[name = tensor("op_6463_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_6463_squeeze_mask_0 = const()[name = tensor("op_6463_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_6463_cast_fp16 = slice_by_index(begin = var_6463_begin_0, end = var_6463_end_0, end_mask = var_6463_end_mask_0, squeeze_mask = var_6463_squeeze_mask_0, x = q_complex_31_cast_fp16)[name = tensor("op_6463_cast_fp16")]; + tensor var_6471_begin_0 = const()[name = tensor("op_6471_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_6471_end_0 = const()[name = tensor("op_6471_end_0"), val = tensor([1, 256, 16, 32, 1])]; + tensor var_6471_end_mask_0 = const()[name = tensor("op_6471_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_6471_squeeze_mask_0 = const()[name = tensor("op_6471_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_6471_cast_fp16 = slice_by_index(begin = var_6471_begin_0, end = var_6471_end_0, end_mask = var_6471_end_mask_0, squeeze_mask = var_6471_squeeze_mask_0, x = k_complex_31_cast_fp16)[name = tensor("op_6471_cast_fp16")]; + tensor var_6479_begin_0 = const()[name = tensor("op_6479_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_6479_end_0 = const()[name = tensor("op_6479_end_0"), val = tensor([1, 256, 16, 32, 2])]; + tensor var_6479_end_mask_0 = const()[name = tensor("op_6479_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_6479_squeeze_mask_0 = const()[name = tensor("op_6479_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_6479_cast_fp16 = slice_by_index(begin = var_6479_begin_0, end = var_6479_end_0, end_mask = var_6479_end_mask_0, squeeze_mask = var_6479_squeeze_mask_0, x = k_complex_31_cast_fp16)[name = tensor("op_6479_cast_fp16")]; + tensor freqs_31_to_fp16 = const()[name = tensor("freqs_31_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(384983680)))]; + tensor var_6485_cast_fp16 = mul(x = freqs_31_to_fp16, y = ts_95_cast_fp16)[name = tensor("op_6485_cast_fp16")]; + tensor rotr_31_cast_fp16 = cos(x = var_6485_cast_fp16)[name = tensor("rotr_31_cast_fp16")]; + tensor roti_31_cast_fp16 = sin(x = var_6485_cast_fp16)[name = tensor("roti_31_cast_fp16")]; + tensor var_6489_cast_fp16 = mul(x = var_6455_cast_fp16, y = rotr_31_cast_fp16)[name = tensor("op_6489_cast_fp16")]; + tensor var_6490_cast_fp16 = mul(x = var_6463_cast_fp16, y = roti_31_cast_fp16)[name = tensor("op_6490_cast_fp16")]; + tensor qor_61_cast_fp16 = sub(x = var_6489_cast_fp16, y = var_6490_cast_fp16)[name = tensor("qor_61_cast_fp16")]; + tensor var_6493_cast_fp16 = mul(x = var_6455_cast_fp16, y = roti_31_cast_fp16)[name = tensor("op_6493_cast_fp16")]; + tensor var_6494_cast_fp16 = mul(x = var_6463_cast_fp16, y = rotr_31_cast_fp16)[name = tensor("op_6494_cast_fp16")]; + tensor qoi_61_cast_fp16 = add(x = var_6493_cast_fp16, y = var_6494_cast_fp16)[name = tensor("qoi_61_cast_fp16")]; + tensor var_6497_cast_fp16 = mul(x = var_6471_cast_fp16, y = rotr_31_cast_fp16)[name = tensor("op_6497_cast_fp16")]; + tensor var_6498_cast_fp16 = mul(x = var_6479_cast_fp16, y = roti_31_cast_fp16)[name = tensor("op_6498_cast_fp16")]; + tensor kor_61_cast_fp16 = sub(x = var_6497_cast_fp16, y = var_6498_cast_fp16)[name = tensor("kor_61_cast_fp16")]; + tensor var_6501_cast_fp16 = mul(x = var_6471_cast_fp16, y = roti_31_cast_fp16)[name = tensor("op_6501_cast_fp16")]; + tensor var_6502_cast_fp16 = mul(x = var_6479_cast_fp16, y = rotr_31_cast_fp16)[name = tensor("op_6502_cast_fp16")]; + tensor koi_61_cast_fp16 = add(x = var_6501_cast_fp16, y = var_6502_cast_fp16)[name = tensor("koi_61_cast_fp16")]; + tensor qo_31_axis_0 = const()[name = tensor("qo_31_axis_0"), val = tensor(-1)]; + tensor qo_31_cast_fp16 = stack(axis = qo_31_axis_0, values = (qor_61_cast_fp16, qoi_61_cast_fp16))[name = tensor("qo_31_cast_fp16")]; + tensor ko_31_axis_0 = const()[name = tensor("ko_31_axis_0"), val = tensor(-1)]; + tensor ko_31_cast_fp16 = stack(axis = ko_31_axis_0, values = (kor_61_cast_fp16, koi_61_cast_fp16))[name = tensor("ko_31_cast_fp16")]; + tensor var_6535 = const()[name = tensor("op_6535"), val = tensor([1, 256, 16, 64])]; + tensor q_93_cast_fp16 = reshape(shape = var_6535, x = qo_31_cast_fp16)[name = tensor("q_93_cast_fp16")]; + tensor var_6541 = const()[name = tensor("op_6541"), val = tensor([1, 256, 16, 64])]; + tensor k_63_cast_fp16 = reshape(shape = var_6541, x = ko_31_cast_fp16)[name = tensor("k_63_cast_fp16")]; + tensor write_indices_float_91_cast_fp16 = add(x = var_6430_cast_fp16, y = var_393_promoted_to_fp16)[name = tensor("write_indices_float_91_cast_fp16")]; + tensor _inversed_6562_y_0_to_fp16 = const()[name = tensor("_inversed_6562_y_0_to_fp16"), val = tensor(0x1p-9)]; + tensor _inversed_6562_cast_fp16 = mul(x = write_indices_float_91_cast_fp16, y = _inversed_6562_y_0_to_fp16)[name = tensor("_inversed_6562_cast_fp16")]; + tensor var_6563_cast_fp16 = floor(x = _inversed_6562_cast_fp16)[name = tensor("op_6563_cast_fp16")]; + tensor var_6564_to_fp16 = const()[name = tensor("op_6564_to_fp16"), val = tensor(0x1p+9)]; + tensor var_6565_cast_fp16 = mul(x = var_6563_cast_fp16, y = var_6564_to_fp16)[name = tensor("op_6565_cast_fp16")]; + tensor write_indices_float_93_cast_fp16 = sub(x = write_indices_float_91_cast_fp16, y = var_6565_cast_fp16)[name = tensor("write_indices_float_93_cast_fp16")]; + tensor var_6583_value_0_to_fp16 = const()[name = tensor("op_6583_value_0_to_fp16"), val = tensor(0x1.ffp+8)]; + tensor var_6583_cast_fp16 = fill_like(ref_tensor = write_indices_float_93_cast_fp16, value = var_6583_value_0_to_fp16)[name = tensor("op_6583_cast_fp16")]; + tensor write_indices_float_95_cast_fp16 = select(a = write_indices_float_93_cast_fp16, b = var_6583_cast_fp16, cond = token_valid_1_cast_fp16)[name = tensor("write_indices_float_95_cast_fp16")]; + tensor var_6589_dtype_0 = const()[name = tensor("op_6589_dtype_0"), val = tensor("int32")]; + tensor var_6594 = const()[name = tensor("op_6594"), val = tensor([1, 256, 1, 1])]; + tensor write_indices_float_95_cast_fp16_to_int32 = cast(dtype = var_6589_dtype_0, x = write_indices_float_95_cast_fp16)[name = tensor("cast_585")]; + tensor var_6595 = reshape(shape = var_6594, x = write_indices_float_95_cast_fp16_to_int32)[name = tensor("op_6595")]; + tensor write_indices_31_reps_0 = const()[name = tensor("write_indices_31_reps_0"), val = tensor([1, 1, 16, 64])]; + tensor write_indices_31 = tile(reps = write_indices_31_reps_0, x = var_6595)[name = tensor("write_indices_31")]; + tensor var_6605_begin_0 = const()[name = tensor("op_6605_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_6605_end_0 = const()[name = tensor("op_6605_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor var_6605_end_mask_0 = const()[name = tensor("op_6605_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_6605_squeeze_mask_0 = const()[name = tensor("op_6605_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor cache15_to_fp16_dtype_0 = const()[name = tensor("cache15_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor cache15_to_fp16 = cast(dtype = cache15_to_fp16_dtype_0, x = cache15)[name = tensor("cast_584")]; + tensor var_6605_cast_fp16 = slice_by_index(begin = var_6605_begin_0, end = var_6605_end_0, end_mask = var_6605_end_mask_0, squeeze_mask = var_6605_squeeze_mask_0, x = cache15_to_fp16)[name = tensor("op_6605_cast_fp16")]; + tensor var_6607_axis_0 = const()[name = tensor("op_6607_axis_0"), val = tensor(1)]; + tensor var_6607_mode_0 = const()[name = tensor("op_6607_mode_0"), val = tensor("update")]; + tensor var_6607_validate_indices_0 = const()[name = tensor("op_6607_validate_indices_0"), val = tensor(false)]; + tensor var_6607_cast_fp16 = scatter_along_axis(axis = var_6607_axis_0, data = var_6605_cast_fp16, indices = write_indices_31, mode = var_6607_mode_0, updates = k_63_cast_fp16, validate_indices = var_6607_validate_indices_0)[name = tensor("op_6607_cast_fp16")]; + tensor concat_122 = const()[name = tensor("concat_122"), val = tensor([0, 0, 0, 0, 0])]; + tensor concat_123 = const()[name = tensor("concat_123"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_31_internal_tensor_assign_1_stride_0 = const()[name = tensor("new_cache_31_internal_tensor_assign_1_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_31_internal_tensor_assign_1_begin_mask_0 = const()[name = tensor("new_cache_31_internal_tensor_assign_1_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_31_internal_tensor_assign_1_end_mask_0 = const()[name = tensor("new_cache_31_internal_tensor_assign_1_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_31_internal_tensor_assign_1_squeeze_mask_0 = const()[name = tensor("new_cache_31_internal_tensor_assign_1_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_76 = const()[name = tensor("shape_76"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_30 = const()[name = tensor("reduce_prod_30"), val = tensor(1048576)]; + tensor range_1d_30_start_0 = const()[name = tensor("range_1d_30_start_0"), val = tensor(0)]; + tensor range_1d_30_step_0 = const()[name = tensor("range_1d_30_step_0"), val = tensor(1)]; + tensor range_1d_30 = range_1d(end = reduce_prod_30, start = range_1d_30_start_0, step = range_1d_30_step_0)[name = tensor("range_1d_30")]; + tensor reshape_150 = reshape(shape = shape_76, x = range_1d_30)[name = tensor("reshape_150")]; + tensor slice_by_index_30 = slice_by_index(begin = concat_122, begin_mask = new_cache_31_internal_tensor_assign_1_begin_mask_0, end = concat_123, end_mask = new_cache_31_internal_tensor_assign_1_end_mask_0, squeeze_mask = new_cache_31_internal_tensor_assign_1_squeeze_mask_0, stride = new_cache_31_internal_tensor_assign_1_stride_0, x = reshape_150)[name = tensor("slice_by_index_30")]; + tensor reshape_151_shape_0 = const()[name = tensor("reshape_151_shape_0"), val = tensor([-1])]; + tensor reshape_151 = reshape(shape = reshape_151_shape_0, x = slice_by_index_30)[name = tensor("reshape_151")]; + tensor reshape_152_shape_0 = const()[name = tensor("reshape_152_shape_0"), val = tensor([-1])]; + tensor reshape_152_cast_fp16 = reshape(shape = reshape_152_shape_0, x = var_6607_cast_fp16)[name = tensor("reshape_152_cast_fp16")]; + tensor reshape_153_shape_0 = const()[name = tensor("reshape_153_shape_0"), val = tensor([-1])]; + tensor reshape_153_cast_fp16 = reshape(shape = reshape_153_shape_0, x = cache15_to_fp16)[name = tensor("reshape_153_cast_fp16")]; + tensor scatter_30_mode_0 = const()[name = tensor("scatter_30_mode_0"), val = tensor("update")]; + tensor scatter_30_axis_0 = const()[name = tensor("scatter_30_axis_0"), val = tensor(0)]; + tensor scatter_30_validate_indices_0 = const()[name = tensor("scatter_30_validate_indices_0"), val = tensor(false)]; + tensor scatter_30_cast_fp16 = scatter(axis = scatter_30_axis_0, data = reshape_153_cast_fp16, indices = reshape_151, mode = scatter_30_mode_0, updates = reshape_152_cast_fp16, validate_indices = scatter_30_validate_indices_0)[name = tensor("scatter_30_cast_fp16")]; + tensor reshape_154_cast_fp16 = reshape(shape = shape_76, x = scatter_30_cast_fp16)[name = tensor("reshape_154_cast_fp16")]; + tensor var_6615_begin_0 = const()[name = tensor("op_6615_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor var_6615_end_0 = const()[name = tensor("op_6615_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor var_6615_end_mask_0 = const()[name = tensor("op_6615_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_6615_squeeze_mask_0 = const()[name = tensor("op_6615_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor var_6615_cast_fp16 = slice_by_index(begin = var_6615_begin_0, end = var_6615_end_0, end_mask = var_6615_end_mask_0, squeeze_mask = var_6615_squeeze_mask_0, x = reshape_154_cast_fp16)[name = tensor("op_6615_cast_fp16")]; + tensor var_6617_axis_0 = const()[name = tensor("op_6617_axis_0"), val = tensor(1)]; + tensor var_6617_mode_0 = const()[name = tensor("op_6617_mode_0"), val = tensor("update")]; + tensor var_6617_validate_indices_0 = const()[name = tensor("op_6617_validate_indices_0"), val = tensor(false)]; + tensor var_6617_cast_fp16 = scatter_along_axis(axis = var_6617_axis_0, data = var_6615_cast_fp16, indices = write_indices_31, mode = var_6617_mode_0, updates = v_31_cast_fp16, validate_indices = var_6617_validate_indices_0)[name = tensor("op_6617_cast_fp16")]; + tensor concat_124 = const()[name = tensor("concat_124"), val = tensor([1, 0, 0, 0, 0])]; + tensor concat_125 = const()[name = tensor("concat_125"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_31_internal_tensor_assign_2_stride_0 = const()[name = tensor("new_cache_31_internal_tensor_assign_2_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_31_internal_tensor_assign_2_begin_mask_0 = const()[name = tensor("new_cache_31_internal_tensor_assign_2_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_31_internal_tensor_assign_2_end_mask_0 = const()[name = tensor("new_cache_31_internal_tensor_assign_2_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_31_internal_tensor_assign_2_squeeze_mask_0 = const()[name = tensor("new_cache_31_internal_tensor_assign_2_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_77 = const()[name = tensor("shape_77"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_31 = const()[name = tensor("reduce_prod_31"), val = tensor(1048576)]; + tensor range_1d_31_start_0 = const()[name = tensor("range_1d_31_start_0"), val = tensor(0)]; + tensor range_1d_31_step_0 = const()[name = tensor("range_1d_31_step_0"), val = tensor(1)]; + tensor range_1d_31 = range_1d(end = reduce_prod_31, start = range_1d_31_start_0, step = range_1d_31_step_0)[name = tensor("range_1d_31")]; + tensor reshape_155 = reshape(shape = shape_77, x = range_1d_31)[name = tensor("reshape_155")]; + tensor slice_by_index_31 = slice_by_index(begin = concat_124, begin_mask = new_cache_31_internal_tensor_assign_2_begin_mask_0, end = concat_125, end_mask = new_cache_31_internal_tensor_assign_2_end_mask_0, squeeze_mask = new_cache_31_internal_tensor_assign_2_squeeze_mask_0, stride = new_cache_31_internal_tensor_assign_2_stride_0, x = reshape_155)[name = tensor("slice_by_index_31")]; + tensor reshape_156_shape_0 = const()[name = tensor("reshape_156_shape_0"), val = tensor([-1])]; + tensor reshape_156 = reshape(shape = reshape_156_shape_0, x = slice_by_index_31)[name = tensor("reshape_156")]; + tensor reshape_157_shape_0 = const()[name = tensor("reshape_157_shape_0"), val = tensor([-1])]; + tensor reshape_157_cast_fp16 = reshape(shape = reshape_157_shape_0, x = var_6617_cast_fp16)[name = tensor("reshape_157_cast_fp16")]; + tensor reshape_158_shape_0 = const()[name = tensor("reshape_158_shape_0"), val = tensor([-1])]; + tensor reshape_158_cast_fp16 = reshape(shape = reshape_158_shape_0, x = reshape_154_cast_fp16)[name = tensor("reshape_158_cast_fp16")]; + tensor scatter_31_mode_0 = const()[name = tensor("scatter_31_mode_0"), val = tensor("update")]; + tensor scatter_31_axis_0 = const()[name = tensor("scatter_31_axis_0"), val = tensor(0)]; + tensor scatter_31_validate_indices_0 = const()[name = tensor("scatter_31_validate_indices_0"), val = tensor(false)]; + tensor scatter_31_cast_fp16 = scatter(axis = scatter_31_axis_0, data = reshape_158_cast_fp16, indices = reshape_156, mode = scatter_31_mode_0, updates = reshape_157_cast_fp16, validate_indices = scatter_31_validate_indices_0)[name = tensor("scatter_31_cast_fp16")]; + tensor reshape_159_cast_fp16 = reshape(shape = shape_77, x = scatter_31_cast_fp16)[name = tensor("reshape_159_cast_fp16")]; + tensor reshape_159_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("reshape_159_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor keys_91_begin_0 = const()[name = tensor("keys_91_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor keys_91_end_0 = const()[name = tensor("keys_91_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor keys_91_end_mask_0 = const()[name = tensor("keys_91_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor keys_91_squeeze_mask_0 = const()[name = tensor("keys_91_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor keys_91_cast_fp16 = slice_by_index(begin = keys_91_begin_0, end = keys_91_end_0, end_mask = keys_91_end_mask_0, squeeze_mask = keys_91_squeeze_mask_0, x = reshape_159_cast_fp16)[name = tensor("keys_91_cast_fp16")]; + tensor values_91_begin_0 = const()[name = tensor("values_91_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor values_91_end_0 = const()[name = tensor("values_91_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor values_91_end_mask_0 = const()[name = tensor("values_91_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor values_91_squeeze_mask_0 = const()[name = tensor("values_91_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor values_91_cast_fp16 = slice_by_index(begin = values_91_begin_0, end = values_91_end_0, end_mask = values_91_end_mask_0, squeeze_mask = values_91_squeeze_mask_0, x = reshape_159_cast_fp16)[name = tensor("values_91_cast_fp16")]; + tensor var_6629_cast_fp16 = not_equal(x = keys_91_cast_fp16, y = keys_91_cast_fp16)[name = tensor("op_6629_cast_fp16")]; + tensor keys_93_cast_fp16 = select(a = var_470_to_fp16, b = keys_91_cast_fp16, cond = var_6629_cast_fp16)[name = tensor("keys_93_cast_fp16")]; + tensor var_6637_cast_fp16 = not_equal(x = values_91_cast_fp16, y = values_91_cast_fp16)[name = tensor("op_6637_cast_fp16")]; + tensor values_93_cast_fp16 = select(a = var_470_to_fp16, b = values_91_cast_fp16, cond = var_6637_cast_fp16)[name = tensor("values_93_cast_fp16")]; + tensor var_6661 = const()[name = tensor("op_6661"), val = tensor([0, 2, 1, 3])]; + tensor var_6677 = const()[name = tensor("op_6677"), val = tensor([1, 1, 1])]; + tensor var_6678_cast_fp16 = reshape(shape = var_6677, x = position15_to_fp16)[name = tensor("op_6678_cast_fp16")]; + tensor q_positions_31_cast_fp16 = add(x = var_6678_cast_fp16, y = q_offsets_1_promoted_to_fp16)[name = tensor("q_positions_31_cast_fp16")]; + tensor valid_end_31_cast_fp16 = add(x = var_6678_cast_fp16, y = var_536_cast_fp16)[name = tensor("valid_end_31_cast_fp16")]; + tensor valid_mask_31_cast_fp16 = less(x = k_positions_1_promoted_to_fp16, y = valid_end_31_cast_fp16)[name = tensor("valid_mask_31_cast_fp16")]; + tensor causal_mask_31_cast_fp16 = less_equal(x = k_positions_1_promoted_to_fp16, y = q_positions_31_cast_fp16)[name = tensor("causal_mask_31_cast_fp16")]; + tensor var_6706 = logical_and(x = valid_mask_31_cast_fp16, y = causal_mask_31_cast_fp16)[name = tensor("op_6706")]; + tensor attn_mask_31_axes_0 = const()[name = tensor("attn_mask_31_axes_0"), val = tensor([1])]; + tensor attn_mask_31 = expand_dims(axes = attn_mask_31_axes_0, x = var_6706)[name = tensor("attn_mask_31")]; + tensor var_6720_transpose_x_0 = const()[name = tensor("op_6720_transpose_x_0"), val = tensor(false)]; + tensor var_6720_transpose_y_0 = const()[name = tensor("op_6720_transpose_y_0"), val = tensor(false)]; + tensor transpose_99_perm_0 = const()[name = tensor("transpose_99_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_100_perm_0 = const()[name = tensor("transpose_100_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_100 = transpose(perm = transpose_100_perm_0, x = keys_93_cast_fp16)[name = tensor("transpose_144")]; + tensor transpose_99 = transpose(perm = transpose_99_perm_0, x = q_93_cast_fp16)[name = tensor("transpose_145")]; + tensor var_6720_cast_fp16 = matmul(transpose_x = var_6720_transpose_x_0, transpose_y = var_6720_transpose_y_0, x = transpose_99, y = transpose_100)[name = tensor("op_6720_cast_fp16")]; + tensor var_6714_to_fp16 = const()[name = tensor("op_6714_to_fp16"), val = tensor([0x1p-3])]; + tensor attn_weights_91_cast_fp16 = mul(x = var_6720_cast_fp16, y = var_6714_to_fp16)[name = tensor("attn_weights_91_cast_fp16")]; + tensor var_6722 = logical_not(x = attn_mask_31)[name = tensor("op_6722")]; + tensor var_6723_to_fp16 = const()[name = tensor("op_6723_to_fp16"), val = tensor(-inf)]; + tensor attn_weights_93_cast_fp16 = select(a = var_6723_to_fp16, b = attn_weights_91_cast_fp16, cond = var_6722)[name = tensor("attn_weights_93_cast_fp16")]; + tensor var_6725 = const()[name = tensor("op_6725"), val = tensor(-1)]; + tensor attn_weights_95_cast_fp16 = softmax(axis = var_6725, x = attn_weights_93_cast_fp16)[name = tensor("attn_weights_95_cast_fp16")]; + tensor attn_output_31_transpose_x_0 = const()[name = tensor("attn_output_31_transpose_x_0"), val = tensor(false)]; + tensor attn_output_31_transpose_y_0 = const()[name = tensor("attn_output_31_transpose_y_0"), val = tensor(false)]; + tensor values_95_cast_fp16 = transpose(perm = var_6661, x = values_93_cast_fp16)[name = tensor("transpose_146")]; + tensor attn_output_31_cast_fp16 = matmul(transpose_x = attn_output_31_transpose_x_0, transpose_y = attn_output_31_transpose_y_0, x = attn_weights_95_cast_fp16, y = values_95_cast_fp16)[name = tensor("attn_output_31_cast_fp16")]; + tensor var_6733 = const()[name = tensor("op_6733"), val = tensor([0, 2, 1, 3])]; + tensor var_6738 = const()[name = tensor("op_6738"), val = tensor([1, 256, 1024])]; + tensor var_6734_cast_fp16 = transpose(perm = var_6733, x = attn_output_31_cast_fp16)[name = tensor("transpose_143")]; + tensor input_185_cast_fp16 = reshape(shape = var_6738, x = var_6734_cast_fp16)[name = tensor("input_185_cast_fp16")]; + tensor attn15_out_proj_weight_to_fp16 = const()[name = tensor("attn15_out_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(384983808)))]; + tensor linear_61_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = attn15_out_proj_weight_to_fp16, x = input_185_cast_fp16)[name = tensor("linear_61_cast_fp16")]; + tensor var_6747_cast_fp16 = add(x = position15_to_fp16, y = valid_len_to_fp16)[name = tensor("op_6747_cast_fp16")]; + tensor var_6747_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_6747_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor input_187_cast_fp16 = add(x = input_181_cast_fp16, y = linear_61_cast_fp16)[name = tensor("input_187_cast_fp16")]; + tensor input_189_axes_0 = const()[name = tensor("input_189_axes_0"), val = tensor([-1])]; + tensor norm15_2_weight_to_fp16 = const()[name = tensor("norm15_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(387081024)))]; + tensor norm15_2_bias_to_fp16 = const()[name = tensor("norm15_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(387083136)))]; + tensor var_6751_to_fp16 = const()[name = tensor("op_6751_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_189_cast_fp16 = layer_norm(axes = input_189_axes_0, beta = norm15_2_bias_to_fp16, epsilon = var_6751_to_fp16, gamma = norm15_2_weight_to_fp16, x = input_187_cast_fp16)[name = tensor("input_189_cast_fp16")]; + tensor linear15_1_weight_to_fp16 = const()[name = tensor("linear15_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(387085248)))]; + tensor linear_62_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = linear15_1_weight_to_fp16, x = input_189_cast_fp16)[name = tensor("linear_62_cast_fp16")]; + tensor input_191_mode_0 = const()[name = tensor("input_191_mode_0"), val = tensor("EXACT")]; + tensor input_191_cast_fp16 = gelu(mode = input_191_mode_0, x = linear_62_cast_fp16)[name = tensor("input_191_cast_fp16")]; + tensor linear15_2_weight_to_fp16 = const()[name = tensor("linear15_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(395473920)))]; + tensor linear_63_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = linear15_2_weight_to_fp16, x = input_191_cast_fp16)[name = tensor("linear_63_cast_fp16")]; + tensor input_193_cast_fp16 = add(x = input_187_cast_fp16, y = linear_63_cast_fp16)[name = tensor("input_193_cast_fp16")]; + tensor input_195_axes_0 = const()[name = tensor("input_195_axes_0"), val = tensor([-1])]; + tensor norm16_1_weight_to_fp16 = const()[name = tensor("norm16_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(403862592)))]; + tensor norm16_1_bias_to_fp16 = const()[name = tensor("norm16_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(403864704)))]; + tensor var_6768_to_fp16 = const()[name = tensor("op_6768_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_195_cast_fp16 = layer_norm(axes = input_195_axes_0, beta = norm16_1_bias_to_fp16, epsilon = var_6768_to_fp16, gamma = norm16_1_weight_to_fp16, x = input_193_cast_fp16)[name = tensor("input_195_cast_fp16")]; + tensor attn16_in_proj_weight_to_fp16 = const()[name = tensor("attn16_in_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(403866816)))]; + tensor linear_64_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = attn16_in_proj_weight_to_fp16, x = input_195_cast_fp16)[name = tensor("linear_64_cast_fp16")]; + tensor var_6782 = const()[name = tensor("op_6782"), val = tensor([1, 256, 3, 16, 64])]; + tensor qkv_33_cast_fp16 = reshape(shape = var_6782, x = linear_64_cast_fp16)[name = tensor("qkv_33_cast_fp16")]; + tensor q_97_begin_0 = const()[name = tensor("q_97_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor q_97_end_0 = const()[name = tensor("q_97_end_0"), val = tensor([1, 256, 1, 16, 64])]; + tensor q_97_end_mask_0 = const()[name = tensor("q_97_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor q_97_squeeze_mask_0 = const()[name = tensor("q_97_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor q_97_cast_fp16 = slice_by_index(begin = q_97_begin_0, end = q_97_end_0, end_mask = q_97_end_mask_0, squeeze_mask = q_97_squeeze_mask_0, x = qkv_33_cast_fp16)[name = tensor("q_97_cast_fp16")]; + tensor k_65_begin_0 = const()[name = tensor("k_65_begin_0"), val = tensor([0, 0, 1, 0, 0])]; + tensor k_65_end_0 = const()[name = tensor("k_65_end_0"), val = tensor([1, 256, 2, 16, 64])]; + tensor k_65_end_mask_0 = const()[name = tensor("k_65_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor k_65_squeeze_mask_0 = const()[name = tensor("k_65_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor k_65_cast_fp16 = slice_by_index(begin = k_65_begin_0, end = k_65_end_0, end_mask = k_65_end_mask_0, squeeze_mask = k_65_squeeze_mask_0, x = qkv_33_cast_fp16)[name = tensor("k_65_cast_fp16")]; + tensor v_33_begin_0 = const()[name = tensor("v_33_begin_0"), val = tensor([0, 0, 2, 0, 0])]; + tensor v_33_end_0 = const()[name = tensor("v_33_end_0"), val = tensor([1, 256, 3, 16, 64])]; + tensor v_33_end_mask_0 = const()[name = tensor("v_33_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor v_33_squeeze_mask_0 = const()[name = tensor("v_33_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor v_33_cast_fp16 = slice_by_index(begin = v_33_begin_0, end = v_33_end_0, end_mask = v_33_end_mask_0, squeeze_mask = v_33_squeeze_mask_0, x = qkv_33_cast_fp16)[name = tensor("v_33_cast_fp16")]; + tensor var_6840 = const()[name = tensor("op_6840"), val = tensor([1, 1])]; + tensor position16_to_fp16_dtype_0 = const()[name = tensor("position16_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor position16_to_fp16 = cast(dtype = position16_to_fp16_dtype_0, x = position16)[name = tensor("cast_581")]; + tensor var_6841_cast_fp16 = reshape(shape = var_6840, x = position16_to_fp16)[name = tensor("op_6841_cast_fp16")]; + tensor ts_99_cast_fp16 = add(x = ts_1_promoted_to_fp16, y = var_6841_cast_fp16)[name = tensor("ts_99_cast_fp16")]; + tensor var_6848 = const()[name = tensor("op_6848"), val = tensor([1, 256, 1, 1])]; + tensor ts_101_cast_fp16 = reshape(shape = var_6848, x = ts_99_cast_fp16)[name = tensor("ts_101_cast_fp16")]; + tensor var_6855 = const()[name = tensor("op_6855"), val = tensor([1, 256, 16, 32, 2])]; + tensor q_complex_33_cast_fp16 = reshape(shape = var_6855, x = q_97_cast_fp16)[name = tensor("q_complex_33_cast_fp16")]; + tensor var_6862 = const()[name = tensor("op_6862"), val = tensor([1, 256, 16, 32, 2])]; + tensor k_complex_33_cast_fp16 = reshape(shape = var_6862, x = k_65_cast_fp16)[name = tensor("k_complex_33_cast_fp16")]; + tensor var_6866_begin_0 = const()[name = tensor("op_6866_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_6866_end_0 = const()[name = tensor("op_6866_end_0"), val = tensor([1, 256, 16, 32, 1])]; + tensor var_6866_end_mask_0 = const()[name = tensor("op_6866_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_6866_squeeze_mask_0 = const()[name = tensor("op_6866_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_6866_cast_fp16 = slice_by_index(begin = var_6866_begin_0, end = var_6866_end_0, end_mask = var_6866_end_mask_0, squeeze_mask = var_6866_squeeze_mask_0, x = q_complex_33_cast_fp16)[name = tensor("op_6866_cast_fp16")]; + tensor var_6874_begin_0 = const()[name = tensor("op_6874_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_6874_end_0 = const()[name = tensor("op_6874_end_0"), val = tensor([1, 256, 16, 32, 2])]; + tensor var_6874_end_mask_0 = const()[name = tensor("op_6874_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_6874_squeeze_mask_0 = const()[name = tensor("op_6874_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_6874_cast_fp16 = slice_by_index(begin = var_6874_begin_0, end = var_6874_end_0, end_mask = var_6874_end_mask_0, squeeze_mask = var_6874_squeeze_mask_0, x = q_complex_33_cast_fp16)[name = tensor("op_6874_cast_fp16")]; + tensor var_6882_begin_0 = const()[name = tensor("op_6882_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_6882_end_0 = const()[name = tensor("op_6882_end_0"), val = tensor([1, 256, 16, 32, 1])]; + tensor var_6882_end_mask_0 = const()[name = tensor("op_6882_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_6882_squeeze_mask_0 = const()[name = tensor("op_6882_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_6882_cast_fp16 = slice_by_index(begin = var_6882_begin_0, end = var_6882_end_0, end_mask = var_6882_end_mask_0, squeeze_mask = var_6882_squeeze_mask_0, x = k_complex_33_cast_fp16)[name = tensor("op_6882_cast_fp16")]; + tensor var_6890_begin_0 = const()[name = tensor("op_6890_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_6890_end_0 = const()[name = tensor("op_6890_end_0"), val = tensor([1, 256, 16, 32, 2])]; + tensor var_6890_end_mask_0 = const()[name = tensor("op_6890_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_6890_squeeze_mask_0 = const()[name = tensor("op_6890_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_6890_cast_fp16 = slice_by_index(begin = var_6890_begin_0, end = var_6890_end_0, end_mask = var_6890_end_mask_0, squeeze_mask = var_6890_squeeze_mask_0, x = k_complex_33_cast_fp16)[name = tensor("op_6890_cast_fp16")]; + tensor freqs_33_to_fp16 = const()[name = tensor("freqs_33_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(410158336)))]; + tensor var_6896_cast_fp16 = mul(x = freqs_33_to_fp16, y = ts_101_cast_fp16)[name = tensor("op_6896_cast_fp16")]; + tensor rotr_33_cast_fp16 = cos(x = var_6896_cast_fp16)[name = tensor("rotr_33_cast_fp16")]; + tensor roti_33_cast_fp16 = sin(x = var_6896_cast_fp16)[name = tensor("roti_33_cast_fp16")]; + tensor var_6900_cast_fp16 = mul(x = var_6866_cast_fp16, y = rotr_33_cast_fp16)[name = tensor("op_6900_cast_fp16")]; + tensor var_6901_cast_fp16 = mul(x = var_6874_cast_fp16, y = roti_33_cast_fp16)[name = tensor("op_6901_cast_fp16")]; + tensor qor_65_cast_fp16 = sub(x = var_6900_cast_fp16, y = var_6901_cast_fp16)[name = tensor("qor_65_cast_fp16")]; + tensor var_6904_cast_fp16 = mul(x = var_6866_cast_fp16, y = roti_33_cast_fp16)[name = tensor("op_6904_cast_fp16")]; + tensor var_6905_cast_fp16 = mul(x = var_6874_cast_fp16, y = rotr_33_cast_fp16)[name = tensor("op_6905_cast_fp16")]; + tensor qoi_65_cast_fp16 = add(x = var_6904_cast_fp16, y = var_6905_cast_fp16)[name = tensor("qoi_65_cast_fp16")]; + tensor var_6908_cast_fp16 = mul(x = var_6882_cast_fp16, y = rotr_33_cast_fp16)[name = tensor("op_6908_cast_fp16")]; + tensor var_6909_cast_fp16 = mul(x = var_6890_cast_fp16, y = roti_33_cast_fp16)[name = tensor("op_6909_cast_fp16")]; + tensor kor_65_cast_fp16 = sub(x = var_6908_cast_fp16, y = var_6909_cast_fp16)[name = tensor("kor_65_cast_fp16")]; + tensor var_6912_cast_fp16 = mul(x = var_6882_cast_fp16, y = roti_33_cast_fp16)[name = tensor("op_6912_cast_fp16")]; + tensor var_6913_cast_fp16 = mul(x = var_6890_cast_fp16, y = rotr_33_cast_fp16)[name = tensor("op_6913_cast_fp16")]; + tensor koi_65_cast_fp16 = add(x = var_6912_cast_fp16, y = var_6913_cast_fp16)[name = tensor("koi_65_cast_fp16")]; + tensor qo_33_axis_0 = const()[name = tensor("qo_33_axis_0"), val = tensor(-1)]; + tensor qo_33_cast_fp16 = stack(axis = qo_33_axis_0, values = (qor_65_cast_fp16, qoi_65_cast_fp16))[name = tensor("qo_33_cast_fp16")]; + tensor ko_33_axis_0 = const()[name = tensor("ko_33_axis_0"), val = tensor(-1)]; + tensor ko_33_cast_fp16 = stack(axis = ko_33_axis_0, values = (kor_65_cast_fp16, koi_65_cast_fp16))[name = tensor("ko_33_cast_fp16")]; + tensor var_6946 = const()[name = tensor("op_6946"), val = tensor([1, 256, 16, 64])]; + tensor q_99_cast_fp16 = reshape(shape = var_6946, x = qo_33_cast_fp16)[name = tensor("q_99_cast_fp16")]; + tensor var_6952 = const()[name = tensor("op_6952"), val = tensor([1, 256, 16, 64])]; + tensor k_67_cast_fp16 = reshape(shape = var_6952, x = ko_33_cast_fp16)[name = tensor("k_67_cast_fp16")]; + tensor write_indices_float_97_cast_fp16 = add(x = var_6841_cast_fp16, y = var_393_promoted_to_fp16)[name = tensor("write_indices_float_97_cast_fp16")]; + tensor _inversed_6973_y_0_to_fp16 = const()[name = tensor("_inversed_6973_y_0_to_fp16"), val = tensor(0x1p-9)]; + tensor _inversed_6973_cast_fp16 = mul(x = write_indices_float_97_cast_fp16, y = _inversed_6973_y_0_to_fp16)[name = tensor("_inversed_6973_cast_fp16")]; + tensor var_6974_cast_fp16 = floor(x = _inversed_6973_cast_fp16)[name = tensor("op_6974_cast_fp16")]; + tensor var_6975_to_fp16 = const()[name = tensor("op_6975_to_fp16"), val = tensor(0x1p+9)]; + tensor var_6976_cast_fp16 = mul(x = var_6974_cast_fp16, y = var_6975_to_fp16)[name = tensor("op_6976_cast_fp16")]; + tensor write_indices_float_99_cast_fp16 = sub(x = write_indices_float_97_cast_fp16, y = var_6976_cast_fp16)[name = tensor("write_indices_float_99_cast_fp16")]; + tensor var_6994_value_0_to_fp16 = const()[name = tensor("op_6994_value_0_to_fp16"), val = tensor(0x1.ffp+8)]; + tensor var_6994_cast_fp16 = fill_like(ref_tensor = write_indices_float_99_cast_fp16, value = var_6994_value_0_to_fp16)[name = tensor("op_6994_cast_fp16")]; + tensor write_indices_float_101_cast_fp16 = select(a = write_indices_float_99_cast_fp16, b = var_6994_cast_fp16, cond = token_valid_1_cast_fp16)[name = tensor("write_indices_float_101_cast_fp16")]; + tensor var_7000_dtype_0 = const()[name = tensor("op_7000_dtype_0"), val = tensor("int32")]; + tensor var_7005 = const()[name = tensor("op_7005"), val = tensor([1, 256, 1, 1])]; + tensor write_indices_float_101_cast_fp16_to_int32 = cast(dtype = var_7000_dtype_0, x = write_indices_float_101_cast_fp16)[name = tensor("cast_580")]; + tensor var_7006 = reshape(shape = var_7005, x = write_indices_float_101_cast_fp16_to_int32)[name = tensor("op_7006")]; + tensor write_indices_33_reps_0 = const()[name = tensor("write_indices_33_reps_0"), val = tensor([1, 1, 16, 64])]; + tensor write_indices_33 = tile(reps = write_indices_33_reps_0, x = var_7006)[name = tensor("write_indices_33")]; + tensor var_7016_begin_0 = const()[name = tensor("op_7016_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_7016_end_0 = const()[name = tensor("op_7016_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor var_7016_end_mask_0 = const()[name = tensor("op_7016_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_7016_squeeze_mask_0 = const()[name = tensor("op_7016_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor cache16_to_fp16_dtype_0 = const()[name = tensor("cache16_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor cache16_to_fp16 = cast(dtype = cache16_to_fp16_dtype_0, x = cache16)[name = tensor("cast_579")]; + tensor var_7016_cast_fp16 = slice_by_index(begin = var_7016_begin_0, end = var_7016_end_0, end_mask = var_7016_end_mask_0, squeeze_mask = var_7016_squeeze_mask_0, x = cache16_to_fp16)[name = tensor("op_7016_cast_fp16")]; + tensor var_7018_axis_0 = const()[name = tensor("op_7018_axis_0"), val = tensor(1)]; + tensor var_7018_mode_0 = const()[name = tensor("op_7018_mode_0"), val = tensor("update")]; + tensor var_7018_validate_indices_0 = const()[name = tensor("op_7018_validate_indices_0"), val = tensor(false)]; + tensor var_7018_cast_fp16 = scatter_along_axis(axis = var_7018_axis_0, data = var_7016_cast_fp16, indices = write_indices_33, mode = var_7018_mode_0, updates = k_67_cast_fp16, validate_indices = var_7018_validate_indices_0)[name = tensor("op_7018_cast_fp16")]; + tensor concat_130 = const()[name = tensor("concat_130"), val = tensor([0, 0, 0, 0, 0])]; + tensor concat_131 = const()[name = tensor("concat_131"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_33_internal_tensor_assign_1_stride_0 = const()[name = tensor("new_cache_33_internal_tensor_assign_1_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_33_internal_tensor_assign_1_begin_mask_0 = const()[name = tensor("new_cache_33_internal_tensor_assign_1_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_33_internal_tensor_assign_1_end_mask_0 = const()[name = tensor("new_cache_33_internal_tensor_assign_1_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_33_internal_tensor_assign_1_squeeze_mask_0 = const()[name = tensor("new_cache_33_internal_tensor_assign_1_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_78 = const()[name = tensor("shape_78"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_32 = const()[name = tensor("reduce_prod_32"), val = tensor(1048576)]; + tensor range_1d_32_start_0 = const()[name = tensor("range_1d_32_start_0"), val = tensor(0)]; + tensor range_1d_32_step_0 = const()[name = tensor("range_1d_32_step_0"), val = tensor(1)]; + tensor range_1d_32 = range_1d(end = reduce_prod_32, start = range_1d_32_start_0, step = range_1d_32_step_0)[name = tensor("range_1d_32")]; + tensor reshape_160 = reshape(shape = shape_78, x = range_1d_32)[name = tensor("reshape_160")]; + tensor slice_by_index_32 = slice_by_index(begin = concat_130, begin_mask = new_cache_33_internal_tensor_assign_1_begin_mask_0, end = concat_131, end_mask = new_cache_33_internal_tensor_assign_1_end_mask_0, squeeze_mask = new_cache_33_internal_tensor_assign_1_squeeze_mask_0, stride = new_cache_33_internal_tensor_assign_1_stride_0, x = reshape_160)[name = tensor("slice_by_index_32")]; + tensor reshape_161_shape_0 = const()[name = tensor("reshape_161_shape_0"), val = tensor([-1])]; + tensor reshape_161 = reshape(shape = reshape_161_shape_0, x = slice_by_index_32)[name = tensor("reshape_161")]; + tensor reshape_162_shape_0 = const()[name = tensor("reshape_162_shape_0"), val = tensor([-1])]; + tensor reshape_162_cast_fp16 = reshape(shape = reshape_162_shape_0, x = var_7018_cast_fp16)[name = tensor("reshape_162_cast_fp16")]; + tensor reshape_163_shape_0 = const()[name = tensor("reshape_163_shape_0"), val = tensor([-1])]; + tensor reshape_163_cast_fp16 = reshape(shape = reshape_163_shape_0, x = cache16_to_fp16)[name = tensor("reshape_163_cast_fp16")]; + tensor scatter_32_mode_0 = const()[name = tensor("scatter_32_mode_0"), val = tensor("update")]; + tensor scatter_32_axis_0 = const()[name = tensor("scatter_32_axis_0"), val = tensor(0)]; + tensor scatter_32_validate_indices_0 = const()[name = tensor("scatter_32_validate_indices_0"), val = tensor(false)]; + tensor scatter_32_cast_fp16 = scatter(axis = scatter_32_axis_0, data = reshape_163_cast_fp16, indices = reshape_161, mode = scatter_32_mode_0, updates = reshape_162_cast_fp16, validate_indices = scatter_32_validate_indices_0)[name = tensor("scatter_32_cast_fp16")]; + tensor reshape_164_cast_fp16 = reshape(shape = shape_78, x = scatter_32_cast_fp16)[name = tensor("reshape_164_cast_fp16")]; + tensor var_7026_begin_0 = const()[name = tensor("op_7026_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor var_7026_end_0 = const()[name = tensor("op_7026_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor var_7026_end_mask_0 = const()[name = tensor("op_7026_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_7026_squeeze_mask_0 = const()[name = tensor("op_7026_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor var_7026_cast_fp16 = slice_by_index(begin = var_7026_begin_0, end = var_7026_end_0, end_mask = var_7026_end_mask_0, squeeze_mask = var_7026_squeeze_mask_0, x = reshape_164_cast_fp16)[name = tensor("op_7026_cast_fp16")]; + tensor var_7028_axis_0 = const()[name = tensor("op_7028_axis_0"), val = tensor(1)]; + tensor var_7028_mode_0 = const()[name = tensor("op_7028_mode_0"), val = tensor("update")]; + tensor var_7028_validate_indices_0 = const()[name = tensor("op_7028_validate_indices_0"), val = tensor(false)]; + tensor var_7028_cast_fp16 = scatter_along_axis(axis = var_7028_axis_0, data = var_7026_cast_fp16, indices = write_indices_33, mode = var_7028_mode_0, updates = v_33_cast_fp16, validate_indices = var_7028_validate_indices_0)[name = tensor("op_7028_cast_fp16")]; + tensor concat_132 = const()[name = tensor("concat_132"), val = tensor([1, 0, 0, 0, 0])]; + tensor concat_133 = const()[name = tensor("concat_133"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_33_internal_tensor_assign_2_stride_0 = const()[name = tensor("new_cache_33_internal_tensor_assign_2_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_33_internal_tensor_assign_2_begin_mask_0 = const()[name = tensor("new_cache_33_internal_tensor_assign_2_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_33_internal_tensor_assign_2_end_mask_0 = const()[name = tensor("new_cache_33_internal_tensor_assign_2_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_33_internal_tensor_assign_2_squeeze_mask_0 = const()[name = tensor("new_cache_33_internal_tensor_assign_2_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_79 = const()[name = tensor("shape_79"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_33 = const()[name = tensor("reduce_prod_33"), val = tensor(1048576)]; + tensor range_1d_33_start_0 = const()[name = tensor("range_1d_33_start_0"), val = tensor(0)]; + tensor range_1d_33_step_0 = const()[name = tensor("range_1d_33_step_0"), val = tensor(1)]; + tensor range_1d_33 = range_1d(end = reduce_prod_33, start = range_1d_33_start_0, step = range_1d_33_step_0)[name = tensor("range_1d_33")]; + tensor reshape_165 = reshape(shape = shape_79, x = range_1d_33)[name = tensor("reshape_165")]; + tensor slice_by_index_33 = slice_by_index(begin = concat_132, begin_mask = new_cache_33_internal_tensor_assign_2_begin_mask_0, end = concat_133, end_mask = new_cache_33_internal_tensor_assign_2_end_mask_0, squeeze_mask = new_cache_33_internal_tensor_assign_2_squeeze_mask_0, stride = new_cache_33_internal_tensor_assign_2_stride_0, x = reshape_165)[name = tensor("slice_by_index_33")]; + tensor reshape_166_shape_0 = const()[name = tensor("reshape_166_shape_0"), val = tensor([-1])]; + tensor reshape_166 = reshape(shape = reshape_166_shape_0, x = slice_by_index_33)[name = tensor("reshape_166")]; + tensor reshape_167_shape_0 = const()[name = tensor("reshape_167_shape_0"), val = tensor([-1])]; + tensor reshape_167_cast_fp16 = reshape(shape = reshape_167_shape_0, x = var_7028_cast_fp16)[name = tensor("reshape_167_cast_fp16")]; + tensor reshape_168_shape_0 = const()[name = tensor("reshape_168_shape_0"), val = tensor([-1])]; + tensor reshape_168_cast_fp16 = reshape(shape = reshape_168_shape_0, x = reshape_164_cast_fp16)[name = tensor("reshape_168_cast_fp16")]; + tensor scatter_33_mode_0 = const()[name = tensor("scatter_33_mode_0"), val = tensor("update")]; + tensor scatter_33_axis_0 = const()[name = tensor("scatter_33_axis_0"), val = tensor(0)]; + tensor scatter_33_validate_indices_0 = const()[name = tensor("scatter_33_validate_indices_0"), val = tensor(false)]; + tensor scatter_33_cast_fp16 = scatter(axis = scatter_33_axis_0, data = reshape_168_cast_fp16, indices = reshape_166, mode = scatter_33_mode_0, updates = reshape_167_cast_fp16, validate_indices = scatter_33_validate_indices_0)[name = tensor("scatter_33_cast_fp16")]; + tensor reshape_169_cast_fp16 = reshape(shape = shape_79, x = scatter_33_cast_fp16)[name = tensor("reshape_169_cast_fp16")]; + tensor reshape_169_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("reshape_169_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor keys_97_begin_0 = const()[name = tensor("keys_97_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor keys_97_end_0 = const()[name = tensor("keys_97_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor keys_97_end_mask_0 = const()[name = tensor("keys_97_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor keys_97_squeeze_mask_0 = const()[name = tensor("keys_97_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor keys_97_cast_fp16 = slice_by_index(begin = keys_97_begin_0, end = keys_97_end_0, end_mask = keys_97_end_mask_0, squeeze_mask = keys_97_squeeze_mask_0, x = reshape_169_cast_fp16)[name = tensor("keys_97_cast_fp16")]; + tensor values_97_begin_0 = const()[name = tensor("values_97_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor values_97_end_0 = const()[name = tensor("values_97_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor values_97_end_mask_0 = const()[name = tensor("values_97_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor values_97_squeeze_mask_0 = const()[name = tensor("values_97_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor values_97_cast_fp16 = slice_by_index(begin = values_97_begin_0, end = values_97_end_0, end_mask = values_97_end_mask_0, squeeze_mask = values_97_squeeze_mask_0, x = reshape_169_cast_fp16)[name = tensor("values_97_cast_fp16")]; + tensor var_7040_cast_fp16 = not_equal(x = keys_97_cast_fp16, y = keys_97_cast_fp16)[name = tensor("op_7040_cast_fp16")]; + tensor keys_99_cast_fp16 = select(a = var_470_to_fp16, b = keys_97_cast_fp16, cond = var_7040_cast_fp16)[name = tensor("keys_99_cast_fp16")]; + tensor var_7048_cast_fp16 = not_equal(x = values_97_cast_fp16, y = values_97_cast_fp16)[name = tensor("op_7048_cast_fp16")]; + tensor values_99_cast_fp16 = select(a = var_470_to_fp16, b = values_97_cast_fp16, cond = var_7048_cast_fp16)[name = tensor("values_99_cast_fp16")]; + tensor var_7072 = const()[name = tensor("op_7072"), val = tensor([0, 2, 1, 3])]; + tensor var_7088 = const()[name = tensor("op_7088"), val = tensor([1, 1, 1])]; + tensor var_7089_cast_fp16 = reshape(shape = var_7088, x = position16_to_fp16)[name = tensor("op_7089_cast_fp16")]; + tensor q_positions_33_cast_fp16 = add(x = var_7089_cast_fp16, y = q_offsets_1_promoted_to_fp16)[name = tensor("q_positions_33_cast_fp16")]; + tensor valid_end_33_cast_fp16 = add(x = var_7089_cast_fp16, y = var_536_cast_fp16)[name = tensor("valid_end_33_cast_fp16")]; + tensor valid_mask_33_cast_fp16 = less(x = k_positions_1_promoted_to_fp16, y = valid_end_33_cast_fp16)[name = tensor("valid_mask_33_cast_fp16")]; + tensor causal_mask_33_cast_fp16 = less_equal(x = k_positions_1_promoted_to_fp16, y = q_positions_33_cast_fp16)[name = tensor("causal_mask_33_cast_fp16")]; + tensor var_7117 = logical_and(x = valid_mask_33_cast_fp16, y = causal_mask_33_cast_fp16)[name = tensor("op_7117")]; + tensor attn_mask_33_axes_0 = const()[name = tensor("attn_mask_33_axes_0"), val = tensor([1])]; + tensor attn_mask_33 = expand_dims(axes = attn_mask_33_axes_0, x = var_7117)[name = tensor("attn_mask_33")]; + tensor var_7131_transpose_x_0 = const()[name = tensor("op_7131_transpose_x_0"), val = tensor(false)]; + tensor var_7131_transpose_y_0 = const()[name = tensor("op_7131_transpose_y_0"), val = tensor(false)]; + tensor transpose_101_perm_0 = const()[name = tensor("transpose_101_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_102_perm_0 = const()[name = tensor("transpose_102_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_102 = transpose(perm = transpose_102_perm_0, x = keys_99_cast_fp16)[name = tensor("transpose_140")]; + tensor transpose_101 = transpose(perm = transpose_101_perm_0, x = q_99_cast_fp16)[name = tensor("transpose_141")]; + tensor var_7131_cast_fp16 = matmul(transpose_x = var_7131_transpose_x_0, transpose_y = var_7131_transpose_y_0, x = transpose_101, y = transpose_102)[name = tensor("op_7131_cast_fp16")]; + tensor var_7125_to_fp16 = const()[name = tensor("op_7125_to_fp16"), val = tensor([0x1p-3])]; + tensor attn_weights_97_cast_fp16 = mul(x = var_7131_cast_fp16, y = var_7125_to_fp16)[name = tensor("attn_weights_97_cast_fp16")]; + tensor var_7133 = logical_not(x = attn_mask_33)[name = tensor("op_7133")]; + tensor var_7134_to_fp16 = const()[name = tensor("op_7134_to_fp16"), val = tensor(-inf)]; + tensor attn_weights_99_cast_fp16 = select(a = var_7134_to_fp16, b = attn_weights_97_cast_fp16, cond = var_7133)[name = tensor("attn_weights_99_cast_fp16")]; + tensor var_7136 = const()[name = tensor("op_7136"), val = tensor(-1)]; + tensor attn_weights_101_cast_fp16 = softmax(axis = var_7136, x = attn_weights_99_cast_fp16)[name = tensor("attn_weights_101_cast_fp16")]; + tensor attn_output_33_transpose_x_0 = const()[name = tensor("attn_output_33_transpose_x_0"), val = tensor(false)]; + tensor attn_output_33_transpose_y_0 = const()[name = tensor("attn_output_33_transpose_y_0"), val = tensor(false)]; + tensor values_101_cast_fp16 = transpose(perm = var_7072, x = values_99_cast_fp16)[name = tensor("transpose_142")]; + tensor attn_output_33_cast_fp16 = matmul(transpose_x = attn_output_33_transpose_x_0, transpose_y = attn_output_33_transpose_y_0, x = attn_weights_101_cast_fp16, y = values_101_cast_fp16)[name = tensor("attn_output_33_cast_fp16")]; + tensor var_7144 = const()[name = tensor("op_7144"), val = tensor([0, 2, 1, 3])]; + tensor var_7149 = const()[name = tensor("op_7149"), val = tensor([1, 256, 1024])]; + tensor var_7145_cast_fp16 = transpose(perm = var_7144, x = attn_output_33_cast_fp16)[name = tensor("transpose_139")]; + tensor input_197_cast_fp16 = reshape(shape = var_7149, x = var_7145_cast_fp16)[name = tensor("input_197_cast_fp16")]; + tensor attn16_out_proj_weight_to_fp16 = const()[name = tensor("attn16_out_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(410158464)))]; + tensor linear_65_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = attn16_out_proj_weight_to_fp16, x = input_197_cast_fp16)[name = tensor("linear_65_cast_fp16")]; + tensor var_7158_cast_fp16 = add(x = position16_to_fp16, y = valid_len_to_fp16)[name = tensor("op_7158_cast_fp16")]; + tensor var_7158_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_7158_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor input_199_cast_fp16 = add(x = input_193_cast_fp16, y = linear_65_cast_fp16)[name = tensor("input_199_cast_fp16")]; + tensor input_201_axes_0 = const()[name = tensor("input_201_axes_0"), val = tensor([-1])]; + tensor norm16_2_weight_to_fp16 = const()[name = tensor("norm16_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(412255680)))]; + tensor norm16_2_bias_to_fp16 = const()[name = tensor("norm16_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(412257792)))]; + tensor var_7162_to_fp16 = const()[name = tensor("op_7162_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_201_cast_fp16 = layer_norm(axes = input_201_axes_0, beta = norm16_2_bias_to_fp16, epsilon = var_7162_to_fp16, gamma = norm16_2_weight_to_fp16, x = input_199_cast_fp16)[name = tensor("input_201_cast_fp16")]; + tensor linear16_1_weight_to_fp16 = const()[name = tensor("linear16_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(412259904)))]; + tensor linear_66_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = linear16_1_weight_to_fp16, x = input_201_cast_fp16)[name = tensor("linear_66_cast_fp16")]; + tensor input_203_mode_0 = const()[name = tensor("input_203_mode_0"), val = tensor("EXACT")]; + tensor input_203_cast_fp16 = gelu(mode = input_203_mode_0, x = linear_66_cast_fp16)[name = tensor("input_203_cast_fp16")]; + tensor linear16_2_weight_to_fp16 = const()[name = tensor("linear16_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(420648576)))]; + tensor linear_67_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = linear16_2_weight_to_fp16, x = input_203_cast_fp16)[name = tensor("linear_67_cast_fp16")]; + tensor input_205_cast_fp16 = add(x = input_199_cast_fp16, y = linear_67_cast_fp16)[name = tensor("input_205_cast_fp16")]; + tensor input_207_axes_0 = const()[name = tensor("input_207_axes_0"), val = tensor([-1])]; + tensor norm17_1_weight_to_fp16 = const()[name = tensor("norm17_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(429037248)))]; + tensor norm17_1_bias_to_fp16 = const()[name = tensor("norm17_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(429039360)))]; + tensor var_7179_to_fp16 = const()[name = tensor("op_7179_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_207_cast_fp16 = layer_norm(axes = input_207_axes_0, beta = norm17_1_bias_to_fp16, epsilon = var_7179_to_fp16, gamma = norm17_1_weight_to_fp16, x = input_205_cast_fp16)[name = tensor("input_207_cast_fp16")]; + tensor attn17_in_proj_weight_to_fp16 = const()[name = tensor("attn17_in_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(429041472)))]; + tensor linear_68_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = attn17_in_proj_weight_to_fp16, x = input_207_cast_fp16)[name = tensor("linear_68_cast_fp16")]; + tensor var_7193 = const()[name = tensor("op_7193"), val = tensor([1, 256, 3, 16, 64])]; + tensor qkv_35_cast_fp16 = reshape(shape = var_7193, x = linear_68_cast_fp16)[name = tensor("qkv_35_cast_fp16")]; + tensor q_103_begin_0 = const()[name = tensor("q_103_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor q_103_end_0 = const()[name = tensor("q_103_end_0"), val = tensor([1, 256, 1, 16, 64])]; + tensor q_103_end_mask_0 = const()[name = tensor("q_103_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor q_103_squeeze_mask_0 = const()[name = tensor("q_103_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor q_103_cast_fp16 = slice_by_index(begin = q_103_begin_0, end = q_103_end_0, end_mask = q_103_end_mask_0, squeeze_mask = q_103_squeeze_mask_0, x = qkv_35_cast_fp16)[name = tensor("q_103_cast_fp16")]; + tensor k_69_begin_0 = const()[name = tensor("k_69_begin_0"), val = tensor([0, 0, 1, 0, 0])]; + tensor k_69_end_0 = const()[name = tensor("k_69_end_0"), val = tensor([1, 256, 2, 16, 64])]; + tensor k_69_end_mask_0 = const()[name = tensor("k_69_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor k_69_squeeze_mask_0 = const()[name = tensor("k_69_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor k_69_cast_fp16 = slice_by_index(begin = k_69_begin_0, end = k_69_end_0, end_mask = k_69_end_mask_0, squeeze_mask = k_69_squeeze_mask_0, x = qkv_35_cast_fp16)[name = tensor("k_69_cast_fp16")]; + tensor v_35_begin_0 = const()[name = tensor("v_35_begin_0"), val = tensor([0, 0, 2, 0, 0])]; + tensor v_35_end_0 = const()[name = tensor("v_35_end_0"), val = tensor([1, 256, 3, 16, 64])]; + tensor v_35_end_mask_0 = const()[name = tensor("v_35_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor v_35_squeeze_mask_0 = const()[name = tensor("v_35_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor v_35_cast_fp16 = slice_by_index(begin = v_35_begin_0, end = v_35_end_0, end_mask = v_35_end_mask_0, squeeze_mask = v_35_squeeze_mask_0, x = qkv_35_cast_fp16)[name = tensor("v_35_cast_fp16")]; + tensor var_7251 = const()[name = tensor("op_7251"), val = tensor([1, 1])]; + tensor position17_to_fp16_dtype_0 = const()[name = tensor("position17_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor position17_to_fp16 = cast(dtype = position17_to_fp16_dtype_0, x = position17)[name = tensor("cast_576")]; + tensor var_7252_cast_fp16 = reshape(shape = var_7251, x = position17_to_fp16)[name = tensor("op_7252_cast_fp16")]; + tensor ts_105_cast_fp16 = add(x = ts_1_promoted_to_fp16, y = var_7252_cast_fp16)[name = tensor("ts_105_cast_fp16")]; + tensor var_7259 = const()[name = tensor("op_7259"), val = tensor([1, 256, 1, 1])]; + tensor ts_107_cast_fp16 = reshape(shape = var_7259, x = ts_105_cast_fp16)[name = tensor("ts_107_cast_fp16")]; + tensor var_7266 = const()[name = tensor("op_7266"), val = tensor([1, 256, 16, 32, 2])]; + tensor q_complex_35_cast_fp16 = reshape(shape = var_7266, x = q_103_cast_fp16)[name = tensor("q_complex_35_cast_fp16")]; + tensor var_7273 = const()[name = tensor("op_7273"), val = tensor([1, 256, 16, 32, 2])]; + tensor k_complex_35_cast_fp16 = reshape(shape = var_7273, x = k_69_cast_fp16)[name = tensor("k_complex_35_cast_fp16")]; + tensor var_7277_begin_0 = const()[name = tensor("op_7277_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_7277_end_0 = const()[name = tensor("op_7277_end_0"), val = tensor([1, 256, 16, 32, 1])]; + tensor var_7277_end_mask_0 = const()[name = tensor("op_7277_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_7277_squeeze_mask_0 = const()[name = tensor("op_7277_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_7277_cast_fp16 = slice_by_index(begin = var_7277_begin_0, end = var_7277_end_0, end_mask = var_7277_end_mask_0, squeeze_mask = var_7277_squeeze_mask_0, x = q_complex_35_cast_fp16)[name = tensor("op_7277_cast_fp16")]; + tensor var_7285_begin_0 = const()[name = tensor("op_7285_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_7285_end_0 = const()[name = tensor("op_7285_end_0"), val = tensor([1, 256, 16, 32, 2])]; + tensor var_7285_end_mask_0 = const()[name = tensor("op_7285_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_7285_squeeze_mask_0 = const()[name = tensor("op_7285_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_7285_cast_fp16 = slice_by_index(begin = var_7285_begin_0, end = var_7285_end_0, end_mask = var_7285_end_mask_0, squeeze_mask = var_7285_squeeze_mask_0, x = q_complex_35_cast_fp16)[name = tensor("op_7285_cast_fp16")]; + tensor var_7293_begin_0 = const()[name = tensor("op_7293_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_7293_end_0 = const()[name = tensor("op_7293_end_0"), val = tensor([1, 256, 16, 32, 1])]; + tensor var_7293_end_mask_0 = const()[name = tensor("op_7293_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_7293_squeeze_mask_0 = const()[name = tensor("op_7293_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_7293_cast_fp16 = slice_by_index(begin = var_7293_begin_0, end = var_7293_end_0, end_mask = var_7293_end_mask_0, squeeze_mask = var_7293_squeeze_mask_0, x = k_complex_35_cast_fp16)[name = tensor("op_7293_cast_fp16")]; + tensor var_7301_begin_0 = const()[name = tensor("op_7301_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_7301_end_0 = const()[name = tensor("op_7301_end_0"), val = tensor([1, 256, 16, 32, 2])]; + tensor var_7301_end_mask_0 = const()[name = tensor("op_7301_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_7301_squeeze_mask_0 = const()[name = tensor("op_7301_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_7301_cast_fp16 = slice_by_index(begin = var_7301_begin_0, end = var_7301_end_0, end_mask = var_7301_end_mask_0, squeeze_mask = var_7301_squeeze_mask_0, x = k_complex_35_cast_fp16)[name = tensor("op_7301_cast_fp16")]; + tensor freqs_35_to_fp16 = const()[name = tensor("freqs_35_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(435332992)))]; + tensor var_7307_cast_fp16 = mul(x = freqs_35_to_fp16, y = ts_107_cast_fp16)[name = tensor("op_7307_cast_fp16")]; + tensor rotr_35_cast_fp16 = cos(x = var_7307_cast_fp16)[name = tensor("rotr_35_cast_fp16")]; + tensor roti_35_cast_fp16 = sin(x = var_7307_cast_fp16)[name = tensor("roti_35_cast_fp16")]; + tensor var_7311_cast_fp16 = mul(x = var_7277_cast_fp16, y = rotr_35_cast_fp16)[name = tensor("op_7311_cast_fp16")]; + tensor var_7312_cast_fp16 = mul(x = var_7285_cast_fp16, y = roti_35_cast_fp16)[name = tensor("op_7312_cast_fp16")]; + tensor qor_69_cast_fp16 = sub(x = var_7311_cast_fp16, y = var_7312_cast_fp16)[name = tensor("qor_69_cast_fp16")]; + tensor var_7315_cast_fp16 = mul(x = var_7277_cast_fp16, y = roti_35_cast_fp16)[name = tensor("op_7315_cast_fp16")]; + tensor var_7316_cast_fp16 = mul(x = var_7285_cast_fp16, y = rotr_35_cast_fp16)[name = tensor("op_7316_cast_fp16")]; + tensor qoi_69_cast_fp16 = add(x = var_7315_cast_fp16, y = var_7316_cast_fp16)[name = tensor("qoi_69_cast_fp16")]; + tensor var_7319_cast_fp16 = mul(x = var_7293_cast_fp16, y = rotr_35_cast_fp16)[name = tensor("op_7319_cast_fp16")]; + tensor var_7320_cast_fp16 = mul(x = var_7301_cast_fp16, y = roti_35_cast_fp16)[name = tensor("op_7320_cast_fp16")]; + tensor kor_69_cast_fp16 = sub(x = var_7319_cast_fp16, y = var_7320_cast_fp16)[name = tensor("kor_69_cast_fp16")]; + tensor var_7323_cast_fp16 = mul(x = var_7293_cast_fp16, y = roti_35_cast_fp16)[name = tensor("op_7323_cast_fp16")]; + tensor var_7324_cast_fp16 = mul(x = var_7301_cast_fp16, y = rotr_35_cast_fp16)[name = tensor("op_7324_cast_fp16")]; + tensor koi_69_cast_fp16 = add(x = var_7323_cast_fp16, y = var_7324_cast_fp16)[name = tensor("koi_69_cast_fp16")]; + tensor qo_35_axis_0 = const()[name = tensor("qo_35_axis_0"), val = tensor(-1)]; + tensor qo_35_cast_fp16 = stack(axis = qo_35_axis_0, values = (qor_69_cast_fp16, qoi_69_cast_fp16))[name = tensor("qo_35_cast_fp16")]; + tensor ko_35_axis_0 = const()[name = tensor("ko_35_axis_0"), val = tensor(-1)]; + tensor ko_35_cast_fp16 = stack(axis = ko_35_axis_0, values = (kor_69_cast_fp16, koi_69_cast_fp16))[name = tensor("ko_35_cast_fp16")]; + tensor var_7357 = const()[name = tensor("op_7357"), val = tensor([1, 256, 16, 64])]; + tensor q_105_cast_fp16 = reshape(shape = var_7357, x = qo_35_cast_fp16)[name = tensor("q_105_cast_fp16")]; + tensor var_7363 = const()[name = tensor("op_7363"), val = tensor([1, 256, 16, 64])]; + tensor k_71_cast_fp16 = reshape(shape = var_7363, x = ko_35_cast_fp16)[name = tensor("k_71_cast_fp16")]; + tensor write_indices_float_103_cast_fp16 = add(x = var_7252_cast_fp16, y = var_393_promoted_to_fp16)[name = tensor("write_indices_float_103_cast_fp16")]; + tensor _inversed_7384_y_0_to_fp16 = const()[name = tensor("_inversed_7384_y_0_to_fp16"), val = tensor(0x1p-9)]; + tensor _inversed_7384_cast_fp16 = mul(x = write_indices_float_103_cast_fp16, y = _inversed_7384_y_0_to_fp16)[name = tensor("_inversed_7384_cast_fp16")]; + tensor var_7385_cast_fp16 = floor(x = _inversed_7384_cast_fp16)[name = tensor("op_7385_cast_fp16")]; + tensor var_7386_to_fp16 = const()[name = tensor("op_7386_to_fp16"), val = tensor(0x1p+9)]; + tensor var_7387_cast_fp16 = mul(x = var_7385_cast_fp16, y = var_7386_to_fp16)[name = tensor("op_7387_cast_fp16")]; + tensor write_indices_float_105_cast_fp16 = sub(x = write_indices_float_103_cast_fp16, y = var_7387_cast_fp16)[name = tensor("write_indices_float_105_cast_fp16")]; + tensor var_7405_value_0_to_fp16 = const()[name = tensor("op_7405_value_0_to_fp16"), val = tensor(0x1.ffp+8)]; + tensor var_7405_cast_fp16 = fill_like(ref_tensor = write_indices_float_105_cast_fp16, value = var_7405_value_0_to_fp16)[name = tensor("op_7405_cast_fp16")]; + tensor write_indices_float_107_cast_fp16 = select(a = write_indices_float_105_cast_fp16, b = var_7405_cast_fp16, cond = token_valid_1_cast_fp16)[name = tensor("write_indices_float_107_cast_fp16")]; + tensor var_7411_dtype_0 = const()[name = tensor("op_7411_dtype_0"), val = tensor("int32")]; + tensor var_7416 = const()[name = tensor("op_7416"), val = tensor([1, 256, 1, 1])]; + tensor write_indices_float_107_cast_fp16_to_int32 = cast(dtype = var_7411_dtype_0, x = write_indices_float_107_cast_fp16)[name = tensor("cast_575")]; + tensor var_7417 = reshape(shape = var_7416, x = write_indices_float_107_cast_fp16_to_int32)[name = tensor("op_7417")]; + tensor write_indices_35_reps_0 = const()[name = tensor("write_indices_35_reps_0"), val = tensor([1, 1, 16, 64])]; + tensor write_indices_35 = tile(reps = write_indices_35_reps_0, x = var_7417)[name = tensor("write_indices_35")]; + tensor var_7427_begin_0 = const()[name = tensor("op_7427_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_7427_end_0 = const()[name = tensor("op_7427_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor var_7427_end_mask_0 = const()[name = tensor("op_7427_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_7427_squeeze_mask_0 = const()[name = tensor("op_7427_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor cache17_to_fp16_dtype_0 = const()[name = tensor("cache17_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor cache17_to_fp16 = cast(dtype = cache17_to_fp16_dtype_0, x = cache17)[name = tensor("cast_574")]; + tensor var_7427_cast_fp16 = slice_by_index(begin = var_7427_begin_0, end = var_7427_end_0, end_mask = var_7427_end_mask_0, squeeze_mask = var_7427_squeeze_mask_0, x = cache17_to_fp16)[name = tensor("op_7427_cast_fp16")]; + tensor var_7429_axis_0 = const()[name = tensor("op_7429_axis_0"), val = tensor(1)]; + tensor var_7429_mode_0 = const()[name = tensor("op_7429_mode_0"), val = tensor("update")]; + tensor var_7429_validate_indices_0 = const()[name = tensor("op_7429_validate_indices_0"), val = tensor(false)]; + tensor var_7429_cast_fp16 = scatter_along_axis(axis = var_7429_axis_0, data = var_7427_cast_fp16, indices = write_indices_35, mode = var_7429_mode_0, updates = k_71_cast_fp16, validate_indices = var_7429_validate_indices_0)[name = tensor("op_7429_cast_fp16")]; + tensor concat_138 = const()[name = tensor("concat_138"), val = tensor([0, 0, 0, 0, 0])]; + tensor concat_139 = const()[name = tensor("concat_139"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_35_internal_tensor_assign_1_stride_0 = const()[name = tensor("new_cache_35_internal_tensor_assign_1_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_35_internal_tensor_assign_1_begin_mask_0 = const()[name = tensor("new_cache_35_internal_tensor_assign_1_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_35_internal_tensor_assign_1_end_mask_0 = const()[name = tensor("new_cache_35_internal_tensor_assign_1_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_35_internal_tensor_assign_1_squeeze_mask_0 = const()[name = tensor("new_cache_35_internal_tensor_assign_1_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_80 = const()[name = tensor("shape_80"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_34 = const()[name = tensor("reduce_prod_34"), val = tensor(1048576)]; + tensor range_1d_34_start_0 = const()[name = tensor("range_1d_34_start_0"), val = tensor(0)]; + tensor range_1d_34_step_0 = const()[name = tensor("range_1d_34_step_0"), val = tensor(1)]; + tensor range_1d_34 = range_1d(end = reduce_prod_34, start = range_1d_34_start_0, step = range_1d_34_step_0)[name = tensor("range_1d_34")]; + tensor reshape_170 = reshape(shape = shape_80, x = range_1d_34)[name = tensor("reshape_170")]; + tensor slice_by_index_34 = slice_by_index(begin = concat_138, begin_mask = new_cache_35_internal_tensor_assign_1_begin_mask_0, end = concat_139, end_mask = new_cache_35_internal_tensor_assign_1_end_mask_0, squeeze_mask = new_cache_35_internal_tensor_assign_1_squeeze_mask_0, stride = new_cache_35_internal_tensor_assign_1_stride_0, x = reshape_170)[name = tensor("slice_by_index_34")]; + tensor reshape_171_shape_0 = const()[name = tensor("reshape_171_shape_0"), val = tensor([-1])]; + tensor reshape_171 = reshape(shape = reshape_171_shape_0, x = slice_by_index_34)[name = tensor("reshape_171")]; + tensor reshape_172_shape_0 = const()[name = tensor("reshape_172_shape_0"), val = tensor([-1])]; + tensor reshape_172_cast_fp16 = reshape(shape = reshape_172_shape_0, x = var_7429_cast_fp16)[name = tensor("reshape_172_cast_fp16")]; + tensor reshape_173_shape_0 = const()[name = tensor("reshape_173_shape_0"), val = tensor([-1])]; + tensor reshape_173_cast_fp16 = reshape(shape = reshape_173_shape_0, x = cache17_to_fp16)[name = tensor("reshape_173_cast_fp16")]; + tensor scatter_34_mode_0 = const()[name = tensor("scatter_34_mode_0"), val = tensor("update")]; + tensor scatter_34_axis_0 = const()[name = tensor("scatter_34_axis_0"), val = tensor(0)]; + tensor scatter_34_validate_indices_0 = const()[name = tensor("scatter_34_validate_indices_0"), val = tensor(false)]; + tensor scatter_34_cast_fp16 = scatter(axis = scatter_34_axis_0, data = reshape_173_cast_fp16, indices = reshape_171, mode = scatter_34_mode_0, updates = reshape_172_cast_fp16, validate_indices = scatter_34_validate_indices_0)[name = tensor("scatter_34_cast_fp16")]; + tensor reshape_174_cast_fp16 = reshape(shape = shape_80, x = scatter_34_cast_fp16)[name = tensor("reshape_174_cast_fp16")]; + tensor var_7437_begin_0 = const()[name = tensor("op_7437_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor var_7437_end_0 = const()[name = tensor("op_7437_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor var_7437_end_mask_0 = const()[name = tensor("op_7437_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_7437_squeeze_mask_0 = const()[name = tensor("op_7437_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor var_7437_cast_fp16 = slice_by_index(begin = var_7437_begin_0, end = var_7437_end_0, end_mask = var_7437_end_mask_0, squeeze_mask = var_7437_squeeze_mask_0, x = reshape_174_cast_fp16)[name = tensor("op_7437_cast_fp16")]; + tensor var_7439_axis_0 = const()[name = tensor("op_7439_axis_0"), val = tensor(1)]; + tensor var_7439_mode_0 = const()[name = tensor("op_7439_mode_0"), val = tensor("update")]; + tensor var_7439_validate_indices_0 = const()[name = tensor("op_7439_validate_indices_0"), val = tensor(false)]; + tensor var_7439_cast_fp16 = scatter_along_axis(axis = var_7439_axis_0, data = var_7437_cast_fp16, indices = write_indices_35, mode = var_7439_mode_0, updates = v_35_cast_fp16, validate_indices = var_7439_validate_indices_0)[name = tensor("op_7439_cast_fp16")]; + tensor concat_140 = const()[name = tensor("concat_140"), val = tensor([1, 0, 0, 0, 0])]; + tensor concat_141 = const()[name = tensor("concat_141"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_35_internal_tensor_assign_2_stride_0 = const()[name = tensor("new_cache_35_internal_tensor_assign_2_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_35_internal_tensor_assign_2_begin_mask_0 = const()[name = tensor("new_cache_35_internal_tensor_assign_2_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_35_internal_tensor_assign_2_end_mask_0 = const()[name = tensor("new_cache_35_internal_tensor_assign_2_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_35_internal_tensor_assign_2_squeeze_mask_0 = const()[name = tensor("new_cache_35_internal_tensor_assign_2_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_81 = const()[name = tensor("shape_81"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_35 = const()[name = tensor("reduce_prod_35"), val = tensor(1048576)]; + tensor range_1d_35_start_0 = const()[name = tensor("range_1d_35_start_0"), val = tensor(0)]; + tensor range_1d_35_step_0 = const()[name = tensor("range_1d_35_step_0"), val = tensor(1)]; + tensor range_1d_35 = range_1d(end = reduce_prod_35, start = range_1d_35_start_0, step = range_1d_35_step_0)[name = tensor("range_1d_35")]; + tensor reshape_175 = reshape(shape = shape_81, x = range_1d_35)[name = tensor("reshape_175")]; + tensor slice_by_index_35 = slice_by_index(begin = concat_140, begin_mask = new_cache_35_internal_tensor_assign_2_begin_mask_0, end = concat_141, end_mask = new_cache_35_internal_tensor_assign_2_end_mask_0, squeeze_mask = new_cache_35_internal_tensor_assign_2_squeeze_mask_0, stride = new_cache_35_internal_tensor_assign_2_stride_0, x = reshape_175)[name = tensor("slice_by_index_35")]; + tensor reshape_176_shape_0 = const()[name = tensor("reshape_176_shape_0"), val = tensor([-1])]; + tensor reshape_176 = reshape(shape = reshape_176_shape_0, x = slice_by_index_35)[name = tensor("reshape_176")]; + tensor reshape_177_shape_0 = const()[name = tensor("reshape_177_shape_0"), val = tensor([-1])]; + tensor reshape_177_cast_fp16 = reshape(shape = reshape_177_shape_0, x = var_7439_cast_fp16)[name = tensor("reshape_177_cast_fp16")]; + tensor reshape_178_shape_0 = const()[name = tensor("reshape_178_shape_0"), val = tensor([-1])]; + tensor reshape_178_cast_fp16 = reshape(shape = reshape_178_shape_0, x = reshape_174_cast_fp16)[name = tensor("reshape_178_cast_fp16")]; + tensor scatter_35_mode_0 = const()[name = tensor("scatter_35_mode_0"), val = tensor("update")]; + tensor scatter_35_axis_0 = const()[name = tensor("scatter_35_axis_0"), val = tensor(0)]; + tensor scatter_35_validate_indices_0 = const()[name = tensor("scatter_35_validate_indices_0"), val = tensor(false)]; + tensor scatter_35_cast_fp16 = scatter(axis = scatter_35_axis_0, data = reshape_178_cast_fp16, indices = reshape_176, mode = scatter_35_mode_0, updates = reshape_177_cast_fp16, validate_indices = scatter_35_validate_indices_0)[name = tensor("scatter_35_cast_fp16")]; + tensor reshape_179_cast_fp16 = reshape(shape = shape_81, x = scatter_35_cast_fp16)[name = tensor("reshape_179_cast_fp16")]; + tensor reshape_179_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("reshape_179_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor keys_103_begin_0 = const()[name = tensor("keys_103_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor keys_103_end_0 = const()[name = tensor("keys_103_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor keys_103_end_mask_0 = const()[name = tensor("keys_103_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor keys_103_squeeze_mask_0 = const()[name = tensor("keys_103_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor keys_103_cast_fp16 = slice_by_index(begin = keys_103_begin_0, end = keys_103_end_0, end_mask = keys_103_end_mask_0, squeeze_mask = keys_103_squeeze_mask_0, x = reshape_179_cast_fp16)[name = tensor("keys_103_cast_fp16")]; + tensor values_103_begin_0 = const()[name = tensor("values_103_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor values_103_end_0 = const()[name = tensor("values_103_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor values_103_end_mask_0 = const()[name = tensor("values_103_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor values_103_squeeze_mask_0 = const()[name = tensor("values_103_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor values_103_cast_fp16 = slice_by_index(begin = values_103_begin_0, end = values_103_end_0, end_mask = values_103_end_mask_0, squeeze_mask = values_103_squeeze_mask_0, x = reshape_179_cast_fp16)[name = tensor("values_103_cast_fp16")]; + tensor var_7451_cast_fp16 = not_equal(x = keys_103_cast_fp16, y = keys_103_cast_fp16)[name = tensor("op_7451_cast_fp16")]; + tensor keys_105_cast_fp16 = select(a = var_470_to_fp16, b = keys_103_cast_fp16, cond = var_7451_cast_fp16)[name = tensor("keys_105_cast_fp16")]; + tensor var_7459_cast_fp16 = not_equal(x = values_103_cast_fp16, y = values_103_cast_fp16)[name = tensor("op_7459_cast_fp16")]; + tensor values_105_cast_fp16 = select(a = var_470_to_fp16, b = values_103_cast_fp16, cond = var_7459_cast_fp16)[name = tensor("values_105_cast_fp16")]; + tensor var_7483 = const()[name = tensor("op_7483"), val = tensor([0, 2, 1, 3])]; + tensor var_7499 = const()[name = tensor("op_7499"), val = tensor([1, 1, 1])]; + tensor var_7500_cast_fp16 = reshape(shape = var_7499, x = position17_to_fp16)[name = tensor("op_7500_cast_fp16")]; + tensor q_positions_35_cast_fp16 = add(x = var_7500_cast_fp16, y = q_offsets_1_promoted_to_fp16)[name = tensor("q_positions_35_cast_fp16")]; + tensor valid_end_35_cast_fp16 = add(x = var_7500_cast_fp16, y = var_536_cast_fp16)[name = tensor("valid_end_35_cast_fp16")]; + tensor valid_mask_35_cast_fp16 = less(x = k_positions_1_promoted_to_fp16, y = valid_end_35_cast_fp16)[name = tensor("valid_mask_35_cast_fp16")]; + tensor causal_mask_35_cast_fp16 = less_equal(x = k_positions_1_promoted_to_fp16, y = q_positions_35_cast_fp16)[name = tensor("causal_mask_35_cast_fp16")]; + tensor var_7528 = logical_and(x = valid_mask_35_cast_fp16, y = causal_mask_35_cast_fp16)[name = tensor("op_7528")]; + tensor attn_mask_35_axes_0 = const()[name = tensor("attn_mask_35_axes_0"), val = tensor([1])]; + tensor attn_mask_35 = expand_dims(axes = attn_mask_35_axes_0, x = var_7528)[name = tensor("attn_mask_35")]; + tensor var_7542_transpose_x_0 = const()[name = tensor("op_7542_transpose_x_0"), val = tensor(false)]; + tensor var_7542_transpose_y_0 = const()[name = tensor("op_7542_transpose_y_0"), val = tensor(false)]; + tensor transpose_103_perm_0 = const()[name = tensor("transpose_103_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_104_perm_0 = const()[name = tensor("transpose_104_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_104 = transpose(perm = transpose_104_perm_0, x = keys_105_cast_fp16)[name = tensor("transpose_136")]; + tensor transpose_103 = transpose(perm = transpose_103_perm_0, x = q_105_cast_fp16)[name = tensor("transpose_137")]; + tensor var_7542_cast_fp16 = matmul(transpose_x = var_7542_transpose_x_0, transpose_y = var_7542_transpose_y_0, x = transpose_103, y = transpose_104)[name = tensor("op_7542_cast_fp16")]; + tensor var_7536_to_fp16 = const()[name = tensor("op_7536_to_fp16"), val = tensor([0x1p-3])]; + tensor attn_weights_103_cast_fp16 = mul(x = var_7542_cast_fp16, y = var_7536_to_fp16)[name = tensor("attn_weights_103_cast_fp16")]; + tensor var_7544 = logical_not(x = attn_mask_35)[name = tensor("op_7544")]; + tensor var_7545_to_fp16 = const()[name = tensor("op_7545_to_fp16"), val = tensor(-inf)]; + tensor attn_weights_105_cast_fp16 = select(a = var_7545_to_fp16, b = attn_weights_103_cast_fp16, cond = var_7544)[name = tensor("attn_weights_105_cast_fp16")]; + tensor var_7547 = const()[name = tensor("op_7547"), val = tensor(-1)]; + tensor attn_weights_107_cast_fp16 = softmax(axis = var_7547, x = attn_weights_105_cast_fp16)[name = tensor("attn_weights_107_cast_fp16")]; + tensor attn_output_35_transpose_x_0 = const()[name = tensor("attn_output_35_transpose_x_0"), val = tensor(false)]; + tensor attn_output_35_transpose_y_0 = const()[name = tensor("attn_output_35_transpose_y_0"), val = tensor(false)]; + tensor values_107_cast_fp16 = transpose(perm = var_7483, x = values_105_cast_fp16)[name = tensor("transpose_138")]; + tensor attn_output_35_cast_fp16 = matmul(transpose_x = attn_output_35_transpose_x_0, transpose_y = attn_output_35_transpose_y_0, x = attn_weights_107_cast_fp16, y = values_107_cast_fp16)[name = tensor("attn_output_35_cast_fp16")]; + tensor var_7555 = const()[name = tensor("op_7555"), val = tensor([0, 2, 1, 3])]; + tensor var_7560 = const()[name = tensor("op_7560"), val = tensor([1, 256, 1024])]; + tensor var_7556_cast_fp16 = transpose(perm = var_7555, x = attn_output_35_cast_fp16)[name = tensor("transpose_135")]; + tensor input_209_cast_fp16 = reshape(shape = var_7560, x = var_7556_cast_fp16)[name = tensor("input_209_cast_fp16")]; + tensor attn17_out_proj_weight_to_fp16 = const()[name = tensor("attn17_out_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(435333120)))]; + tensor linear_69_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = attn17_out_proj_weight_to_fp16, x = input_209_cast_fp16)[name = tensor("linear_69_cast_fp16")]; + tensor var_7569_cast_fp16 = add(x = position17_to_fp16, y = valid_len_to_fp16)[name = tensor("op_7569_cast_fp16")]; + tensor var_7569_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_7569_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor input_211_cast_fp16 = add(x = input_205_cast_fp16, y = linear_69_cast_fp16)[name = tensor("input_211_cast_fp16")]; + tensor input_213_axes_0 = const()[name = tensor("input_213_axes_0"), val = tensor([-1])]; + tensor norm17_2_weight_to_fp16 = const()[name = tensor("norm17_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(437430336)))]; + tensor norm17_2_bias_to_fp16 = const()[name = tensor("norm17_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(437432448)))]; + tensor var_7573_to_fp16 = const()[name = tensor("op_7573_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_213_cast_fp16 = layer_norm(axes = input_213_axes_0, beta = norm17_2_bias_to_fp16, epsilon = var_7573_to_fp16, gamma = norm17_2_weight_to_fp16, x = input_211_cast_fp16)[name = tensor("input_213_cast_fp16")]; + tensor linear17_1_weight_to_fp16 = const()[name = tensor("linear17_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(437434560)))]; + tensor linear_70_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = linear17_1_weight_to_fp16, x = input_213_cast_fp16)[name = tensor("linear_70_cast_fp16")]; + tensor input_215_mode_0 = const()[name = tensor("input_215_mode_0"), val = tensor("EXACT")]; + tensor input_215_cast_fp16 = gelu(mode = input_215_mode_0, x = linear_70_cast_fp16)[name = tensor("input_215_cast_fp16")]; + tensor linear17_2_weight_to_fp16 = const()[name = tensor("linear17_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(445823232)))]; + tensor linear_71_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = linear17_2_weight_to_fp16, x = input_215_cast_fp16)[name = tensor("linear_71_cast_fp16")]; + tensor input_217_cast_fp16 = add(x = input_211_cast_fp16, y = linear_71_cast_fp16)[name = tensor("input_217_cast_fp16")]; + tensor input_219_axes_0 = const()[name = tensor("input_219_axes_0"), val = tensor([-1])]; + tensor norm18_1_weight_to_fp16 = const()[name = tensor("norm18_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(454211904)))]; + tensor norm18_1_bias_to_fp16 = const()[name = tensor("norm18_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(454214016)))]; + tensor var_7590_to_fp16 = const()[name = tensor("op_7590_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_219_cast_fp16 = layer_norm(axes = input_219_axes_0, beta = norm18_1_bias_to_fp16, epsilon = var_7590_to_fp16, gamma = norm18_1_weight_to_fp16, x = input_217_cast_fp16)[name = tensor("input_219_cast_fp16")]; + tensor attn18_in_proj_weight_to_fp16 = const()[name = tensor("attn18_in_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(454216128)))]; + tensor linear_72_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = attn18_in_proj_weight_to_fp16, x = input_219_cast_fp16)[name = tensor("linear_72_cast_fp16")]; + tensor var_7604 = const()[name = tensor("op_7604"), val = tensor([1, 256, 3, 16, 64])]; + tensor qkv_37_cast_fp16 = reshape(shape = var_7604, x = linear_72_cast_fp16)[name = tensor("qkv_37_cast_fp16")]; + tensor q_109_begin_0 = const()[name = tensor("q_109_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor q_109_end_0 = const()[name = tensor("q_109_end_0"), val = tensor([1, 256, 1, 16, 64])]; + tensor q_109_end_mask_0 = const()[name = tensor("q_109_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor q_109_squeeze_mask_0 = const()[name = tensor("q_109_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor q_109_cast_fp16 = slice_by_index(begin = q_109_begin_0, end = q_109_end_0, end_mask = q_109_end_mask_0, squeeze_mask = q_109_squeeze_mask_0, x = qkv_37_cast_fp16)[name = tensor("q_109_cast_fp16")]; + tensor k_73_begin_0 = const()[name = tensor("k_73_begin_0"), val = tensor([0, 0, 1, 0, 0])]; + tensor k_73_end_0 = const()[name = tensor("k_73_end_0"), val = tensor([1, 256, 2, 16, 64])]; + tensor k_73_end_mask_0 = const()[name = tensor("k_73_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor k_73_squeeze_mask_0 = const()[name = tensor("k_73_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor k_73_cast_fp16 = slice_by_index(begin = k_73_begin_0, end = k_73_end_0, end_mask = k_73_end_mask_0, squeeze_mask = k_73_squeeze_mask_0, x = qkv_37_cast_fp16)[name = tensor("k_73_cast_fp16")]; + tensor v_37_begin_0 = const()[name = tensor("v_37_begin_0"), val = tensor([0, 0, 2, 0, 0])]; + tensor v_37_end_0 = const()[name = tensor("v_37_end_0"), val = tensor([1, 256, 3, 16, 64])]; + tensor v_37_end_mask_0 = const()[name = tensor("v_37_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor v_37_squeeze_mask_0 = const()[name = tensor("v_37_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor v_37_cast_fp16 = slice_by_index(begin = v_37_begin_0, end = v_37_end_0, end_mask = v_37_end_mask_0, squeeze_mask = v_37_squeeze_mask_0, x = qkv_37_cast_fp16)[name = tensor("v_37_cast_fp16")]; + tensor var_7662 = const()[name = tensor("op_7662"), val = tensor([1, 1])]; + tensor position18_to_fp16_dtype_0 = const()[name = tensor("position18_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor position18_to_fp16 = cast(dtype = position18_to_fp16_dtype_0, x = position18)[name = tensor("cast_571")]; + tensor var_7663_cast_fp16 = reshape(shape = var_7662, x = position18_to_fp16)[name = tensor("op_7663_cast_fp16")]; + tensor ts_111_cast_fp16 = add(x = ts_1_promoted_to_fp16, y = var_7663_cast_fp16)[name = tensor("ts_111_cast_fp16")]; + tensor var_7670 = const()[name = tensor("op_7670"), val = tensor([1, 256, 1, 1])]; + tensor ts_113_cast_fp16 = reshape(shape = var_7670, x = ts_111_cast_fp16)[name = tensor("ts_113_cast_fp16")]; + tensor var_7677 = const()[name = tensor("op_7677"), val = tensor([1, 256, 16, 32, 2])]; + tensor q_complex_37_cast_fp16 = reshape(shape = var_7677, x = q_109_cast_fp16)[name = tensor("q_complex_37_cast_fp16")]; + tensor var_7684 = const()[name = tensor("op_7684"), val = tensor([1, 256, 16, 32, 2])]; + tensor k_complex_37_cast_fp16 = reshape(shape = var_7684, x = k_73_cast_fp16)[name = tensor("k_complex_37_cast_fp16")]; + tensor var_7688_begin_0 = const()[name = tensor("op_7688_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_7688_end_0 = const()[name = tensor("op_7688_end_0"), val = tensor([1, 256, 16, 32, 1])]; + tensor var_7688_end_mask_0 = const()[name = tensor("op_7688_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_7688_squeeze_mask_0 = const()[name = tensor("op_7688_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_7688_cast_fp16 = slice_by_index(begin = var_7688_begin_0, end = var_7688_end_0, end_mask = var_7688_end_mask_0, squeeze_mask = var_7688_squeeze_mask_0, x = q_complex_37_cast_fp16)[name = tensor("op_7688_cast_fp16")]; + tensor var_7696_begin_0 = const()[name = tensor("op_7696_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_7696_end_0 = const()[name = tensor("op_7696_end_0"), val = tensor([1, 256, 16, 32, 2])]; + tensor var_7696_end_mask_0 = const()[name = tensor("op_7696_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_7696_squeeze_mask_0 = const()[name = tensor("op_7696_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_7696_cast_fp16 = slice_by_index(begin = var_7696_begin_0, end = var_7696_end_0, end_mask = var_7696_end_mask_0, squeeze_mask = var_7696_squeeze_mask_0, x = q_complex_37_cast_fp16)[name = tensor("op_7696_cast_fp16")]; + tensor var_7704_begin_0 = const()[name = tensor("op_7704_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_7704_end_0 = const()[name = tensor("op_7704_end_0"), val = tensor([1, 256, 16, 32, 1])]; + tensor var_7704_end_mask_0 = const()[name = tensor("op_7704_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_7704_squeeze_mask_0 = const()[name = tensor("op_7704_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_7704_cast_fp16 = slice_by_index(begin = var_7704_begin_0, end = var_7704_end_0, end_mask = var_7704_end_mask_0, squeeze_mask = var_7704_squeeze_mask_0, x = k_complex_37_cast_fp16)[name = tensor("op_7704_cast_fp16")]; + tensor var_7712_begin_0 = const()[name = tensor("op_7712_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_7712_end_0 = const()[name = tensor("op_7712_end_0"), val = tensor([1, 256, 16, 32, 2])]; + tensor var_7712_end_mask_0 = const()[name = tensor("op_7712_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_7712_squeeze_mask_0 = const()[name = tensor("op_7712_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_7712_cast_fp16 = slice_by_index(begin = var_7712_begin_0, end = var_7712_end_0, end_mask = var_7712_end_mask_0, squeeze_mask = var_7712_squeeze_mask_0, x = k_complex_37_cast_fp16)[name = tensor("op_7712_cast_fp16")]; + tensor freqs_37_to_fp16 = const()[name = tensor("freqs_37_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(460507648)))]; + tensor var_7718_cast_fp16 = mul(x = freqs_37_to_fp16, y = ts_113_cast_fp16)[name = tensor("op_7718_cast_fp16")]; + tensor rotr_37_cast_fp16 = cos(x = var_7718_cast_fp16)[name = tensor("rotr_37_cast_fp16")]; + tensor roti_37_cast_fp16 = sin(x = var_7718_cast_fp16)[name = tensor("roti_37_cast_fp16")]; + tensor var_7722_cast_fp16 = mul(x = var_7688_cast_fp16, y = rotr_37_cast_fp16)[name = tensor("op_7722_cast_fp16")]; + tensor var_7723_cast_fp16 = mul(x = var_7696_cast_fp16, y = roti_37_cast_fp16)[name = tensor("op_7723_cast_fp16")]; + tensor qor_73_cast_fp16 = sub(x = var_7722_cast_fp16, y = var_7723_cast_fp16)[name = tensor("qor_73_cast_fp16")]; + tensor var_7726_cast_fp16 = mul(x = var_7688_cast_fp16, y = roti_37_cast_fp16)[name = tensor("op_7726_cast_fp16")]; + tensor var_7727_cast_fp16 = mul(x = var_7696_cast_fp16, y = rotr_37_cast_fp16)[name = tensor("op_7727_cast_fp16")]; + tensor qoi_73_cast_fp16 = add(x = var_7726_cast_fp16, y = var_7727_cast_fp16)[name = tensor("qoi_73_cast_fp16")]; + tensor var_7730_cast_fp16 = mul(x = var_7704_cast_fp16, y = rotr_37_cast_fp16)[name = tensor("op_7730_cast_fp16")]; + tensor var_7731_cast_fp16 = mul(x = var_7712_cast_fp16, y = roti_37_cast_fp16)[name = tensor("op_7731_cast_fp16")]; + tensor kor_73_cast_fp16 = sub(x = var_7730_cast_fp16, y = var_7731_cast_fp16)[name = tensor("kor_73_cast_fp16")]; + tensor var_7734_cast_fp16 = mul(x = var_7704_cast_fp16, y = roti_37_cast_fp16)[name = tensor("op_7734_cast_fp16")]; + tensor var_7735_cast_fp16 = mul(x = var_7712_cast_fp16, y = rotr_37_cast_fp16)[name = tensor("op_7735_cast_fp16")]; + tensor koi_73_cast_fp16 = add(x = var_7734_cast_fp16, y = var_7735_cast_fp16)[name = tensor("koi_73_cast_fp16")]; + tensor qo_37_axis_0 = const()[name = tensor("qo_37_axis_0"), val = tensor(-1)]; + tensor qo_37_cast_fp16 = stack(axis = qo_37_axis_0, values = (qor_73_cast_fp16, qoi_73_cast_fp16))[name = tensor("qo_37_cast_fp16")]; + tensor ko_37_axis_0 = const()[name = tensor("ko_37_axis_0"), val = tensor(-1)]; + tensor ko_37_cast_fp16 = stack(axis = ko_37_axis_0, values = (kor_73_cast_fp16, koi_73_cast_fp16))[name = tensor("ko_37_cast_fp16")]; + tensor var_7768 = const()[name = tensor("op_7768"), val = tensor([1, 256, 16, 64])]; + tensor q_111_cast_fp16 = reshape(shape = var_7768, x = qo_37_cast_fp16)[name = tensor("q_111_cast_fp16")]; + tensor var_7774 = const()[name = tensor("op_7774"), val = tensor([1, 256, 16, 64])]; + tensor k_75_cast_fp16 = reshape(shape = var_7774, x = ko_37_cast_fp16)[name = tensor("k_75_cast_fp16")]; + tensor write_indices_float_109_cast_fp16 = add(x = var_7663_cast_fp16, y = var_393_promoted_to_fp16)[name = tensor("write_indices_float_109_cast_fp16")]; + tensor _inversed_7795_y_0_to_fp16 = const()[name = tensor("_inversed_7795_y_0_to_fp16"), val = tensor(0x1p-9)]; + tensor _inversed_7795_cast_fp16 = mul(x = write_indices_float_109_cast_fp16, y = _inversed_7795_y_0_to_fp16)[name = tensor("_inversed_7795_cast_fp16")]; + tensor var_7796_cast_fp16 = floor(x = _inversed_7795_cast_fp16)[name = tensor("op_7796_cast_fp16")]; + tensor var_7797_to_fp16 = const()[name = tensor("op_7797_to_fp16"), val = tensor(0x1p+9)]; + tensor var_7798_cast_fp16 = mul(x = var_7796_cast_fp16, y = var_7797_to_fp16)[name = tensor("op_7798_cast_fp16")]; + tensor write_indices_float_111_cast_fp16 = sub(x = write_indices_float_109_cast_fp16, y = var_7798_cast_fp16)[name = tensor("write_indices_float_111_cast_fp16")]; + tensor var_7816_value_0_to_fp16 = const()[name = tensor("op_7816_value_0_to_fp16"), val = tensor(0x1.ffp+8)]; + tensor var_7816_cast_fp16 = fill_like(ref_tensor = write_indices_float_111_cast_fp16, value = var_7816_value_0_to_fp16)[name = tensor("op_7816_cast_fp16")]; + tensor write_indices_float_113_cast_fp16 = select(a = write_indices_float_111_cast_fp16, b = var_7816_cast_fp16, cond = token_valid_1_cast_fp16)[name = tensor("write_indices_float_113_cast_fp16")]; + tensor var_7822_dtype_0 = const()[name = tensor("op_7822_dtype_0"), val = tensor("int32")]; + tensor var_7827 = const()[name = tensor("op_7827"), val = tensor([1, 256, 1, 1])]; + tensor write_indices_float_113_cast_fp16_to_int32 = cast(dtype = var_7822_dtype_0, x = write_indices_float_113_cast_fp16)[name = tensor("cast_570")]; + tensor var_7828 = reshape(shape = var_7827, x = write_indices_float_113_cast_fp16_to_int32)[name = tensor("op_7828")]; + tensor write_indices_37_reps_0 = const()[name = tensor("write_indices_37_reps_0"), val = tensor([1, 1, 16, 64])]; + tensor write_indices_37 = tile(reps = write_indices_37_reps_0, x = var_7828)[name = tensor("write_indices_37")]; + tensor var_7838_begin_0 = const()[name = tensor("op_7838_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_7838_end_0 = const()[name = tensor("op_7838_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor var_7838_end_mask_0 = const()[name = tensor("op_7838_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_7838_squeeze_mask_0 = const()[name = tensor("op_7838_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor cache18_to_fp16_dtype_0 = const()[name = tensor("cache18_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor cache18_to_fp16 = cast(dtype = cache18_to_fp16_dtype_0, x = cache18)[name = tensor("cast_569")]; + tensor var_7838_cast_fp16 = slice_by_index(begin = var_7838_begin_0, end = var_7838_end_0, end_mask = var_7838_end_mask_0, squeeze_mask = var_7838_squeeze_mask_0, x = cache18_to_fp16)[name = tensor("op_7838_cast_fp16")]; + tensor var_7840_axis_0 = const()[name = tensor("op_7840_axis_0"), val = tensor(1)]; + tensor var_7840_mode_0 = const()[name = tensor("op_7840_mode_0"), val = tensor("update")]; + tensor var_7840_validate_indices_0 = const()[name = tensor("op_7840_validate_indices_0"), val = tensor(false)]; + tensor var_7840_cast_fp16 = scatter_along_axis(axis = var_7840_axis_0, data = var_7838_cast_fp16, indices = write_indices_37, mode = var_7840_mode_0, updates = k_75_cast_fp16, validate_indices = var_7840_validate_indices_0)[name = tensor("op_7840_cast_fp16")]; + tensor concat_146 = const()[name = tensor("concat_146"), val = tensor([0, 0, 0, 0, 0])]; + tensor concat_147 = const()[name = tensor("concat_147"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_37_internal_tensor_assign_1_stride_0 = const()[name = tensor("new_cache_37_internal_tensor_assign_1_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_37_internal_tensor_assign_1_begin_mask_0 = const()[name = tensor("new_cache_37_internal_tensor_assign_1_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_37_internal_tensor_assign_1_end_mask_0 = const()[name = tensor("new_cache_37_internal_tensor_assign_1_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_37_internal_tensor_assign_1_squeeze_mask_0 = const()[name = tensor("new_cache_37_internal_tensor_assign_1_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_82 = const()[name = tensor("shape_82"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_36 = const()[name = tensor("reduce_prod_36"), val = tensor(1048576)]; + tensor range_1d_36_start_0 = const()[name = tensor("range_1d_36_start_0"), val = tensor(0)]; + tensor range_1d_36_step_0 = const()[name = tensor("range_1d_36_step_0"), val = tensor(1)]; + tensor range_1d_36 = range_1d(end = reduce_prod_36, start = range_1d_36_start_0, step = range_1d_36_step_0)[name = tensor("range_1d_36")]; + tensor reshape_180 = reshape(shape = shape_82, x = range_1d_36)[name = tensor("reshape_180")]; + tensor slice_by_index_36 = slice_by_index(begin = concat_146, begin_mask = new_cache_37_internal_tensor_assign_1_begin_mask_0, end = concat_147, end_mask = new_cache_37_internal_tensor_assign_1_end_mask_0, squeeze_mask = new_cache_37_internal_tensor_assign_1_squeeze_mask_0, stride = new_cache_37_internal_tensor_assign_1_stride_0, x = reshape_180)[name = tensor("slice_by_index_36")]; + tensor reshape_181_shape_0 = const()[name = tensor("reshape_181_shape_0"), val = tensor([-1])]; + tensor reshape_181 = reshape(shape = reshape_181_shape_0, x = slice_by_index_36)[name = tensor("reshape_181")]; + tensor reshape_182_shape_0 = const()[name = tensor("reshape_182_shape_0"), val = tensor([-1])]; + tensor reshape_182_cast_fp16 = reshape(shape = reshape_182_shape_0, x = var_7840_cast_fp16)[name = tensor("reshape_182_cast_fp16")]; + tensor reshape_183_shape_0 = const()[name = tensor("reshape_183_shape_0"), val = tensor([-1])]; + tensor reshape_183_cast_fp16 = reshape(shape = reshape_183_shape_0, x = cache18_to_fp16)[name = tensor("reshape_183_cast_fp16")]; + tensor scatter_36_mode_0 = const()[name = tensor("scatter_36_mode_0"), val = tensor("update")]; + tensor scatter_36_axis_0 = const()[name = tensor("scatter_36_axis_0"), val = tensor(0)]; + tensor scatter_36_validate_indices_0 = const()[name = tensor("scatter_36_validate_indices_0"), val = tensor(false)]; + tensor scatter_36_cast_fp16 = scatter(axis = scatter_36_axis_0, data = reshape_183_cast_fp16, indices = reshape_181, mode = scatter_36_mode_0, updates = reshape_182_cast_fp16, validate_indices = scatter_36_validate_indices_0)[name = tensor("scatter_36_cast_fp16")]; + tensor reshape_184_cast_fp16 = reshape(shape = shape_82, x = scatter_36_cast_fp16)[name = tensor("reshape_184_cast_fp16")]; + tensor var_7848_begin_0 = const()[name = tensor("op_7848_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor var_7848_end_0 = const()[name = tensor("op_7848_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor var_7848_end_mask_0 = const()[name = tensor("op_7848_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_7848_squeeze_mask_0 = const()[name = tensor("op_7848_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor var_7848_cast_fp16 = slice_by_index(begin = var_7848_begin_0, end = var_7848_end_0, end_mask = var_7848_end_mask_0, squeeze_mask = var_7848_squeeze_mask_0, x = reshape_184_cast_fp16)[name = tensor("op_7848_cast_fp16")]; + tensor var_7850_axis_0 = const()[name = tensor("op_7850_axis_0"), val = tensor(1)]; + tensor var_7850_mode_0 = const()[name = tensor("op_7850_mode_0"), val = tensor("update")]; + tensor var_7850_validate_indices_0 = const()[name = tensor("op_7850_validate_indices_0"), val = tensor(false)]; + tensor var_7850_cast_fp16 = scatter_along_axis(axis = var_7850_axis_0, data = var_7848_cast_fp16, indices = write_indices_37, mode = var_7850_mode_0, updates = v_37_cast_fp16, validate_indices = var_7850_validate_indices_0)[name = tensor("op_7850_cast_fp16")]; + tensor concat_148 = const()[name = tensor("concat_148"), val = tensor([1, 0, 0, 0, 0])]; + tensor concat_149 = const()[name = tensor("concat_149"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_37_internal_tensor_assign_2_stride_0 = const()[name = tensor("new_cache_37_internal_tensor_assign_2_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_37_internal_tensor_assign_2_begin_mask_0 = const()[name = tensor("new_cache_37_internal_tensor_assign_2_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_37_internal_tensor_assign_2_end_mask_0 = const()[name = tensor("new_cache_37_internal_tensor_assign_2_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_37_internal_tensor_assign_2_squeeze_mask_0 = const()[name = tensor("new_cache_37_internal_tensor_assign_2_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_83 = const()[name = tensor("shape_83"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_37 = const()[name = tensor("reduce_prod_37"), val = tensor(1048576)]; + tensor range_1d_37_start_0 = const()[name = tensor("range_1d_37_start_0"), val = tensor(0)]; + tensor range_1d_37_step_0 = const()[name = tensor("range_1d_37_step_0"), val = tensor(1)]; + tensor range_1d_37 = range_1d(end = reduce_prod_37, start = range_1d_37_start_0, step = range_1d_37_step_0)[name = tensor("range_1d_37")]; + tensor reshape_185 = reshape(shape = shape_83, x = range_1d_37)[name = tensor("reshape_185")]; + tensor slice_by_index_37 = slice_by_index(begin = concat_148, begin_mask = new_cache_37_internal_tensor_assign_2_begin_mask_0, end = concat_149, end_mask = new_cache_37_internal_tensor_assign_2_end_mask_0, squeeze_mask = new_cache_37_internal_tensor_assign_2_squeeze_mask_0, stride = new_cache_37_internal_tensor_assign_2_stride_0, x = reshape_185)[name = tensor("slice_by_index_37")]; + tensor reshape_186_shape_0 = const()[name = tensor("reshape_186_shape_0"), val = tensor([-1])]; + tensor reshape_186 = reshape(shape = reshape_186_shape_0, x = slice_by_index_37)[name = tensor("reshape_186")]; + tensor reshape_187_shape_0 = const()[name = tensor("reshape_187_shape_0"), val = tensor([-1])]; + tensor reshape_187_cast_fp16 = reshape(shape = reshape_187_shape_0, x = var_7850_cast_fp16)[name = tensor("reshape_187_cast_fp16")]; + tensor reshape_188_shape_0 = const()[name = tensor("reshape_188_shape_0"), val = tensor([-1])]; + tensor reshape_188_cast_fp16 = reshape(shape = reshape_188_shape_0, x = reshape_184_cast_fp16)[name = tensor("reshape_188_cast_fp16")]; + tensor scatter_37_mode_0 = const()[name = tensor("scatter_37_mode_0"), val = tensor("update")]; + tensor scatter_37_axis_0 = const()[name = tensor("scatter_37_axis_0"), val = tensor(0)]; + tensor scatter_37_validate_indices_0 = const()[name = tensor("scatter_37_validate_indices_0"), val = tensor(false)]; + tensor scatter_37_cast_fp16 = scatter(axis = scatter_37_axis_0, data = reshape_188_cast_fp16, indices = reshape_186, mode = scatter_37_mode_0, updates = reshape_187_cast_fp16, validate_indices = scatter_37_validate_indices_0)[name = tensor("scatter_37_cast_fp16")]; + tensor reshape_189_cast_fp16 = reshape(shape = shape_83, x = scatter_37_cast_fp16)[name = tensor("reshape_189_cast_fp16")]; + tensor reshape_189_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("reshape_189_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor keys_109_begin_0 = const()[name = tensor("keys_109_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor keys_109_end_0 = const()[name = tensor("keys_109_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor keys_109_end_mask_0 = const()[name = tensor("keys_109_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor keys_109_squeeze_mask_0 = const()[name = tensor("keys_109_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor keys_109_cast_fp16 = slice_by_index(begin = keys_109_begin_0, end = keys_109_end_0, end_mask = keys_109_end_mask_0, squeeze_mask = keys_109_squeeze_mask_0, x = reshape_189_cast_fp16)[name = tensor("keys_109_cast_fp16")]; + tensor values_109_begin_0 = const()[name = tensor("values_109_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor values_109_end_0 = const()[name = tensor("values_109_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor values_109_end_mask_0 = const()[name = tensor("values_109_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor values_109_squeeze_mask_0 = const()[name = tensor("values_109_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor values_109_cast_fp16 = slice_by_index(begin = values_109_begin_0, end = values_109_end_0, end_mask = values_109_end_mask_0, squeeze_mask = values_109_squeeze_mask_0, x = reshape_189_cast_fp16)[name = tensor("values_109_cast_fp16")]; + tensor var_7862_cast_fp16 = not_equal(x = keys_109_cast_fp16, y = keys_109_cast_fp16)[name = tensor("op_7862_cast_fp16")]; + tensor keys_111_cast_fp16 = select(a = var_470_to_fp16, b = keys_109_cast_fp16, cond = var_7862_cast_fp16)[name = tensor("keys_111_cast_fp16")]; + tensor var_7870_cast_fp16 = not_equal(x = values_109_cast_fp16, y = values_109_cast_fp16)[name = tensor("op_7870_cast_fp16")]; + tensor values_111_cast_fp16 = select(a = var_470_to_fp16, b = values_109_cast_fp16, cond = var_7870_cast_fp16)[name = tensor("values_111_cast_fp16")]; + tensor var_7894 = const()[name = tensor("op_7894"), val = tensor([0, 2, 1, 3])]; + tensor var_7910 = const()[name = tensor("op_7910"), val = tensor([1, 1, 1])]; + tensor var_7911_cast_fp16 = reshape(shape = var_7910, x = position18_to_fp16)[name = tensor("op_7911_cast_fp16")]; + tensor q_positions_37_cast_fp16 = add(x = var_7911_cast_fp16, y = q_offsets_1_promoted_to_fp16)[name = tensor("q_positions_37_cast_fp16")]; + tensor valid_end_37_cast_fp16 = add(x = var_7911_cast_fp16, y = var_536_cast_fp16)[name = tensor("valid_end_37_cast_fp16")]; + tensor valid_mask_37_cast_fp16 = less(x = k_positions_1_promoted_to_fp16, y = valid_end_37_cast_fp16)[name = tensor("valid_mask_37_cast_fp16")]; + tensor causal_mask_37_cast_fp16 = less_equal(x = k_positions_1_promoted_to_fp16, y = q_positions_37_cast_fp16)[name = tensor("causal_mask_37_cast_fp16")]; + tensor var_7939 = logical_and(x = valid_mask_37_cast_fp16, y = causal_mask_37_cast_fp16)[name = tensor("op_7939")]; + tensor attn_mask_37_axes_0 = const()[name = tensor("attn_mask_37_axes_0"), val = tensor([1])]; + tensor attn_mask_37 = expand_dims(axes = attn_mask_37_axes_0, x = var_7939)[name = tensor("attn_mask_37")]; + tensor var_7953_transpose_x_0 = const()[name = tensor("op_7953_transpose_x_0"), val = tensor(false)]; + tensor var_7953_transpose_y_0 = const()[name = tensor("op_7953_transpose_y_0"), val = tensor(false)]; + tensor transpose_105_perm_0 = const()[name = tensor("transpose_105_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_106_perm_0 = const()[name = tensor("transpose_106_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_106 = transpose(perm = transpose_106_perm_0, x = keys_111_cast_fp16)[name = tensor("transpose_132")]; + tensor transpose_105 = transpose(perm = transpose_105_perm_0, x = q_111_cast_fp16)[name = tensor("transpose_133")]; + tensor var_7953_cast_fp16 = matmul(transpose_x = var_7953_transpose_x_0, transpose_y = var_7953_transpose_y_0, x = transpose_105, y = transpose_106)[name = tensor("op_7953_cast_fp16")]; + tensor var_7947_to_fp16 = const()[name = tensor("op_7947_to_fp16"), val = tensor([0x1p-3])]; + tensor attn_weights_109_cast_fp16 = mul(x = var_7953_cast_fp16, y = var_7947_to_fp16)[name = tensor("attn_weights_109_cast_fp16")]; + tensor var_7955 = logical_not(x = attn_mask_37)[name = tensor("op_7955")]; + tensor var_7956_to_fp16 = const()[name = tensor("op_7956_to_fp16"), val = tensor(-inf)]; + tensor attn_weights_111_cast_fp16 = select(a = var_7956_to_fp16, b = attn_weights_109_cast_fp16, cond = var_7955)[name = tensor("attn_weights_111_cast_fp16")]; + tensor var_7958 = const()[name = tensor("op_7958"), val = tensor(-1)]; + tensor attn_weights_113_cast_fp16 = softmax(axis = var_7958, x = attn_weights_111_cast_fp16)[name = tensor("attn_weights_113_cast_fp16")]; + tensor attn_output_37_transpose_x_0 = const()[name = tensor("attn_output_37_transpose_x_0"), val = tensor(false)]; + tensor attn_output_37_transpose_y_0 = const()[name = tensor("attn_output_37_transpose_y_0"), val = tensor(false)]; + tensor values_113_cast_fp16 = transpose(perm = var_7894, x = values_111_cast_fp16)[name = tensor("transpose_134")]; + tensor attn_output_37_cast_fp16 = matmul(transpose_x = attn_output_37_transpose_x_0, transpose_y = attn_output_37_transpose_y_0, x = attn_weights_113_cast_fp16, y = values_113_cast_fp16)[name = tensor("attn_output_37_cast_fp16")]; + tensor var_7966 = const()[name = tensor("op_7966"), val = tensor([0, 2, 1, 3])]; + tensor var_7971 = const()[name = tensor("op_7971"), val = tensor([1, 256, 1024])]; + tensor var_7967_cast_fp16 = transpose(perm = var_7966, x = attn_output_37_cast_fp16)[name = tensor("transpose_131")]; + tensor input_221_cast_fp16 = reshape(shape = var_7971, x = var_7967_cast_fp16)[name = tensor("input_221_cast_fp16")]; + tensor attn18_out_proj_weight_to_fp16 = const()[name = tensor("attn18_out_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(460507776)))]; + tensor linear_73_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = attn18_out_proj_weight_to_fp16, x = input_221_cast_fp16)[name = tensor("linear_73_cast_fp16")]; + tensor var_7980_cast_fp16 = add(x = position18_to_fp16, y = valid_len_to_fp16)[name = tensor("op_7980_cast_fp16")]; + tensor var_7980_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_7980_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor input_223_cast_fp16 = add(x = input_217_cast_fp16, y = linear_73_cast_fp16)[name = tensor("input_223_cast_fp16")]; + tensor input_225_axes_0 = const()[name = tensor("input_225_axes_0"), val = tensor([-1])]; + tensor norm18_2_weight_to_fp16 = const()[name = tensor("norm18_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(462604992)))]; + tensor norm18_2_bias_to_fp16 = const()[name = tensor("norm18_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(462607104)))]; + tensor var_7984_to_fp16 = const()[name = tensor("op_7984_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_225_cast_fp16 = layer_norm(axes = input_225_axes_0, beta = norm18_2_bias_to_fp16, epsilon = var_7984_to_fp16, gamma = norm18_2_weight_to_fp16, x = input_223_cast_fp16)[name = tensor("input_225_cast_fp16")]; + tensor linear18_1_weight_to_fp16 = const()[name = tensor("linear18_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(462609216)))]; + tensor linear_74_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = linear18_1_weight_to_fp16, x = input_225_cast_fp16)[name = tensor("linear_74_cast_fp16")]; + tensor input_227_mode_0 = const()[name = tensor("input_227_mode_0"), val = tensor("EXACT")]; + tensor input_227_cast_fp16 = gelu(mode = input_227_mode_0, x = linear_74_cast_fp16)[name = tensor("input_227_cast_fp16")]; + tensor linear18_2_weight_to_fp16 = const()[name = tensor("linear18_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(470997888)))]; + tensor linear_75_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = linear18_2_weight_to_fp16, x = input_227_cast_fp16)[name = tensor("linear_75_cast_fp16")]; + tensor input_229_cast_fp16 = add(x = input_223_cast_fp16, y = linear_75_cast_fp16)[name = tensor("input_229_cast_fp16")]; + tensor input_231_axes_0 = const()[name = tensor("input_231_axes_0"), val = tensor([-1])]; + tensor norm19_1_weight_to_fp16 = const()[name = tensor("norm19_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(479386560)))]; + tensor norm19_1_bias_to_fp16 = const()[name = tensor("norm19_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(479388672)))]; + tensor var_8001_to_fp16 = const()[name = tensor("op_8001_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_231_cast_fp16 = layer_norm(axes = input_231_axes_0, beta = norm19_1_bias_to_fp16, epsilon = var_8001_to_fp16, gamma = norm19_1_weight_to_fp16, x = input_229_cast_fp16)[name = tensor("input_231_cast_fp16")]; + tensor attn19_in_proj_weight_to_fp16 = const()[name = tensor("attn19_in_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(479390784)))]; + tensor linear_76_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = attn19_in_proj_weight_to_fp16, x = input_231_cast_fp16)[name = tensor("linear_76_cast_fp16")]; + tensor var_8015 = const()[name = tensor("op_8015"), val = tensor([1, 256, 3, 16, 64])]; + tensor qkv_39_cast_fp16 = reshape(shape = var_8015, x = linear_76_cast_fp16)[name = tensor("qkv_39_cast_fp16")]; + tensor q_115_begin_0 = const()[name = tensor("q_115_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor q_115_end_0 = const()[name = tensor("q_115_end_0"), val = tensor([1, 256, 1, 16, 64])]; + tensor q_115_end_mask_0 = const()[name = tensor("q_115_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor q_115_squeeze_mask_0 = const()[name = tensor("q_115_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor q_115_cast_fp16 = slice_by_index(begin = q_115_begin_0, end = q_115_end_0, end_mask = q_115_end_mask_0, squeeze_mask = q_115_squeeze_mask_0, x = qkv_39_cast_fp16)[name = tensor("q_115_cast_fp16")]; + tensor k_77_begin_0 = const()[name = tensor("k_77_begin_0"), val = tensor([0, 0, 1, 0, 0])]; + tensor k_77_end_0 = const()[name = tensor("k_77_end_0"), val = tensor([1, 256, 2, 16, 64])]; + tensor k_77_end_mask_0 = const()[name = tensor("k_77_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor k_77_squeeze_mask_0 = const()[name = tensor("k_77_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor k_77_cast_fp16 = slice_by_index(begin = k_77_begin_0, end = k_77_end_0, end_mask = k_77_end_mask_0, squeeze_mask = k_77_squeeze_mask_0, x = qkv_39_cast_fp16)[name = tensor("k_77_cast_fp16")]; + tensor v_39_begin_0 = const()[name = tensor("v_39_begin_0"), val = tensor([0, 0, 2, 0, 0])]; + tensor v_39_end_0 = const()[name = tensor("v_39_end_0"), val = tensor([1, 256, 3, 16, 64])]; + tensor v_39_end_mask_0 = const()[name = tensor("v_39_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor v_39_squeeze_mask_0 = const()[name = tensor("v_39_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor v_39_cast_fp16 = slice_by_index(begin = v_39_begin_0, end = v_39_end_0, end_mask = v_39_end_mask_0, squeeze_mask = v_39_squeeze_mask_0, x = qkv_39_cast_fp16)[name = tensor("v_39_cast_fp16")]; + tensor var_8073 = const()[name = tensor("op_8073"), val = tensor([1, 1])]; + tensor position19_to_fp16_dtype_0 = const()[name = tensor("position19_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor position19_to_fp16 = cast(dtype = position19_to_fp16_dtype_0, x = position19)[name = tensor("cast_566")]; + tensor var_8074_cast_fp16 = reshape(shape = var_8073, x = position19_to_fp16)[name = tensor("op_8074_cast_fp16")]; + tensor ts_117_cast_fp16 = add(x = ts_1_promoted_to_fp16, y = var_8074_cast_fp16)[name = tensor("ts_117_cast_fp16")]; + tensor var_8081 = const()[name = tensor("op_8081"), val = tensor([1, 256, 1, 1])]; + tensor ts_119_cast_fp16 = reshape(shape = var_8081, x = ts_117_cast_fp16)[name = tensor("ts_119_cast_fp16")]; + tensor var_8088 = const()[name = tensor("op_8088"), val = tensor([1, 256, 16, 32, 2])]; + tensor q_complex_39_cast_fp16 = reshape(shape = var_8088, x = q_115_cast_fp16)[name = tensor("q_complex_39_cast_fp16")]; + tensor var_8095 = const()[name = tensor("op_8095"), val = tensor([1, 256, 16, 32, 2])]; + tensor k_complex_39_cast_fp16 = reshape(shape = var_8095, x = k_77_cast_fp16)[name = tensor("k_complex_39_cast_fp16")]; + tensor var_8099_begin_0 = const()[name = tensor("op_8099_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_8099_end_0 = const()[name = tensor("op_8099_end_0"), val = tensor([1, 256, 16, 32, 1])]; + tensor var_8099_end_mask_0 = const()[name = tensor("op_8099_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_8099_squeeze_mask_0 = const()[name = tensor("op_8099_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_8099_cast_fp16 = slice_by_index(begin = var_8099_begin_0, end = var_8099_end_0, end_mask = var_8099_end_mask_0, squeeze_mask = var_8099_squeeze_mask_0, x = q_complex_39_cast_fp16)[name = tensor("op_8099_cast_fp16")]; + tensor var_8107_begin_0 = const()[name = tensor("op_8107_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_8107_end_0 = const()[name = tensor("op_8107_end_0"), val = tensor([1, 256, 16, 32, 2])]; + tensor var_8107_end_mask_0 = const()[name = tensor("op_8107_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_8107_squeeze_mask_0 = const()[name = tensor("op_8107_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_8107_cast_fp16 = slice_by_index(begin = var_8107_begin_0, end = var_8107_end_0, end_mask = var_8107_end_mask_0, squeeze_mask = var_8107_squeeze_mask_0, x = q_complex_39_cast_fp16)[name = tensor("op_8107_cast_fp16")]; + tensor var_8115_begin_0 = const()[name = tensor("op_8115_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_8115_end_0 = const()[name = tensor("op_8115_end_0"), val = tensor([1, 256, 16, 32, 1])]; + tensor var_8115_end_mask_0 = const()[name = tensor("op_8115_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_8115_squeeze_mask_0 = const()[name = tensor("op_8115_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_8115_cast_fp16 = slice_by_index(begin = var_8115_begin_0, end = var_8115_end_0, end_mask = var_8115_end_mask_0, squeeze_mask = var_8115_squeeze_mask_0, x = k_complex_39_cast_fp16)[name = tensor("op_8115_cast_fp16")]; + tensor var_8123_begin_0 = const()[name = tensor("op_8123_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_8123_end_0 = const()[name = tensor("op_8123_end_0"), val = tensor([1, 256, 16, 32, 2])]; + tensor var_8123_end_mask_0 = const()[name = tensor("op_8123_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_8123_squeeze_mask_0 = const()[name = tensor("op_8123_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_8123_cast_fp16 = slice_by_index(begin = var_8123_begin_0, end = var_8123_end_0, end_mask = var_8123_end_mask_0, squeeze_mask = var_8123_squeeze_mask_0, x = k_complex_39_cast_fp16)[name = tensor("op_8123_cast_fp16")]; + tensor freqs_39_to_fp16 = const()[name = tensor("freqs_39_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(485682304)))]; + tensor var_8129_cast_fp16 = mul(x = freqs_39_to_fp16, y = ts_119_cast_fp16)[name = tensor("op_8129_cast_fp16")]; + tensor rotr_39_cast_fp16 = cos(x = var_8129_cast_fp16)[name = tensor("rotr_39_cast_fp16")]; + tensor roti_39_cast_fp16 = sin(x = var_8129_cast_fp16)[name = tensor("roti_39_cast_fp16")]; + tensor var_8133_cast_fp16 = mul(x = var_8099_cast_fp16, y = rotr_39_cast_fp16)[name = tensor("op_8133_cast_fp16")]; + tensor var_8134_cast_fp16 = mul(x = var_8107_cast_fp16, y = roti_39_cast_fp16)[name = tensor("op_8134_cast_fp16")]; + tensor qor_77_cast_fp16 = sub(x = var_8133_cast_fp16, y = var_8134_cast_fp16)[name = tensor("qor_77_cast_fp16")]; + tensor var_8137_cast_fp16 = mul(x = var_8099_cast_fp16, y = roti_39_cast_fp16)[name = tensor("op_8137_cast_fp16")]; + tensor var_8138_cast_fp16 = mul(x = var_8107_cast_fp16, y = rotr_39_cast_fp16)[name = tensor("op_8138_cast_fp16")]; + tensor qoi_77_cast_fp16 = add(x = var_8137_cast_fp16, y = var_8138_cast_fp16)[name = tensor("qoi_77_cast_fp16")]; + tensor var_8141_cast_fp16 = mul(x = var_8115_cast_fp16, y = rotr_39_cast_fp16)[name = tensor("op_8141_cast_fp16")]; + tensor var_8142_cast_fp16 = mul(x = var_8123_cast_fp16, y = roti_39_cast_fp16)[name = tensor("op_8142_cast_fp16")]; + tensor kor_77_cast_fp16 = sub(x = var_8141_cast_fp16, y = var_8142_cast_fp16)[name = tensor("kor_77_cast_fp16")]; + tensor var_8145_cast_fp16 = mul(x = var_8115_cast_fp16, y = roti_39_cast_fp16)[name = tensor("op_8145_cast_fp16")]; + tensor var_8146_cast_fp16 = mul(x = var_8123_cast_fp16, y = rotr_39_cast_fp16)[name = tensor("op_8146_cast_fp16")]; + tensor koi_77_cast_fp16 = add(x = var_8145_cast_fp16, y = var_8146_cast_fp16)[name = tensor("koi_77_cast_fp16")]; + tensor qo_39_axis_0 = const()[name = tensor("qo_39_axis_0"), val = tensor(-1)]; + tensor qo_39_cast_fp16 = stack(axis = qo_39_axis_0, values = (qor_77_cast_fp16, qoi_77_cast_fp16))[name = tensor("qo_39_cast_fp16")]; + tensor ko_39_axis_0 = const()[name = tensor("ko_39_axis_0"), val = tensor(-1)]; + tensor ko_39_cast_fp16 = stack(axis = ko_39_axis_0, values = (kor_77_cast_fp16, koi_77_cast_fp16))[name = tensor("ko_39_cast_fp16")]; + tensor var_8179 = const()[name = tensor("op_8179"), val = tensor([1, 256, 16, 64])]; + tensor q_117_cast_fp16 = reshape(shape = var_8179, x = qo_39_cast_fp16)[name = tensor("q_117_cast_fp16")]; + tensor var_8185 = const()[name = tensor("op_8185"), val = tensor([1, 256, 16, 64])]; + tensor k_79_cast_fp16 = reshape(shape = var_8185, x = ko_39_cast_fp16)[name = tensor("k_79_cast_fp16")]; + tensor write_indices_float_115_cast_fp16 = add(x = var_8074_cast_fp16, y = var_393_promoted_to_fp16)[name = tensor("write_indices_float_115_cast_fp16")]; + tensor _inversed_8206_y_0_to_fp16 = const()[name = tensor("_inversed_8206_y_0_to_fp16"), val = tensor(0x1p-9)]; + tensor _inversed_8206_cast_fp16 = mul(x = write_indices_float_115_cast_fp16, y = _inversed_8206_y_0_to_fp16)[name = tensor("_inversed_8206_cast_fp16")]; + tensor var_8207_cast_fp16 = floor(x = _inversed_8206_cast_fp16)[name = tensor("op_8207_cast_fp16")]; + tensor var_8208_to_fp16 = const()[name = tensor("op_8208_to_fp16"), val = tensor(0x1p+9)]; + tensor var_8209_cast_fp16 = mul(x = var_8207_cast_fp16, y = var_8208_to_fp16)[name = tensor("op_8209_cast_fp16")]; + tensor write_indices_float_117_cast_fp16 = sub(x = write_indices_float_115_cast_fp16, y = var_8209_cast_fp16)[name = tensor("write_indices_float_117_cast_fp16")]; + tensor var_8227_value_0_to_fp16 = const()[name = tensor("op_8227_value_0_to_fp16"), val = tensor(0x1.ffp+8)]; + tensor var_8227_cast_fp16 = fill_like(ref_tensor = write_indices_float_117_cast_fp16, value = var_8227_value_0_to_fp16)[name = tensor("op_8227_cast_fp16")]; + tensor write_indices_float_119_cast_fp16 = select(a = write_indices_float_117_cast_fp16, b = var_8227_cast_fp16, cond = token_valid_1_cast_fp16)[name = tensor("write_indices_float_119_cast_fp16")]; + tensor var_8233_dtype_0 = const()[name = tensor("op_8233_dtype_0"), val = tensor("int32")]; + tensor var_8238 = const()[name = tensor("op_8238"), val = tensor([1, 256, 1, 1])]; + tensor write_indices_float_119_cast_fp16_to_int32 = cast(dtype = var_8233_dtype_0, x = write_indices_float_119_cast_fp16)[name = tensor("cast_565")]; + tensor var_8239 = reshape(shape = var_8238, x = write_indices_float_119_cast_fp16_to_int32)[name = tensor("op_8239")]; + tensor write_indices_39_reps_0 = const()[name = tensor("write_indices_39_reps_0"), val = tensor([1, 1, 16, 64])]; + tensor write_indices_39 = tile(reps = write_indices_39_reps_0, x = var_8239)[name = tensor("write_indices_39")]; + tensor var_8249_begin_0 = const()[name = tensor("op_8249_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_8249_end_0 = const()[name = tensor("op_8249_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor var_8249_end_mask_0 = const()[name = tensor("op_8249_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_8249_squeeze_mask_0 = const()[name = tensor("op_8249_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor cache19_to_fp16_dtype_0 = const()[name = tensor("cache19_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor cache19_to_fp16 = cast(dtype = cache19_to_fp16_dtype_0, x = cache19)[name = tensor("cast_564")]; + tensor var_8249_cast_fp16 = slice_by_index(begin = var_8249_begin_0, end = var_8249_end_0, end_mask = var_8249_end_mask_0, squeeze_mask = var_8249_squeeze_mask_0, x = cache19_to_fp16)[name = tensor("op_8249_cast_fp16")]; + tensor var_8251_axis_0 = const()[name = tensor("op_8251_axis_0"), val = tensor(1)]; + tensor var_8251_mode_0 = const()[name = tensor("op_8251_mode_0"), val = tensor("update")]; + tensor var_8251_validate_indices_0 = const()[name = tensor("op_8251_validate_indices_0"), val = tensor(false)]; + tensor var_8251_cast_fp16 = scatter_along_axis(axis = var_8251_axis_0, data = var_8249_cast_fp16, indices = write_indices_39, mode = var_8251_mode_0, updates = k_79_cast_fp16, validate_indices = var_8251_validate_indices_0)[name = tensor("op_8251_cast_fp16")]; + tensor concat_154 = const()[name = tensor("concat_154"), val = tensor([0, 0, 0, 0, 0])]; + tensor concat_155 = const()[name = tensor("concat_155"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_39_internal_tensor_assign_1_stride_0 = const()[name = tensor("new_cache_39_internal_tensor_assign_1_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_39_internal_tensor_assign_1_begin_mask_0 = const()[name = tensor("new_cache_39_internal_tensor_assign_1_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_39_internal_tensor_assign_1_end_mask_0 = const()[name = tensor("new_cache_39_internal_tensor_assign_1_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_39_internal_tensor_assign_1_squeeze_mask_0 = const()[name = tensor("new_cache_39_internal_tensor_assign_1_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_84 = const()[name = tensor("shape_84"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_38 = const()[name = tensor("reduce_prod_38"), val = tensor(1048576)]; + tensor range_1d_38_start_0 = const()[name = tensor("range_1d_38_start_0"), val = tensor(0)]; + tensor range_1d_38_step_0 = const()[name = tensor("range_1d_38_step_0"), val = tensor(1)]; + tensor range_1d_38 = range_1d(end = reduce_prod_38, start = range_1d_38_start_0, step = range_1d_38_step_0)[name = tensor("range_1d_38")]; + tensor reshape_190 = reshape(shape = shape_84, x = range_1d_38)[name = tensor("reshape_190")]; + tensor slice_by_index_38 = slice_by_index(begin = concat_154, begin_mask = new_cache_39_internal_tensor_assign_1_begin_mask_0, end = concat_155, end_mask = new_cache_39_internal_tensor_assign_1_end_mask_0, squeeze_mask = new_cache_39_internal_tensor_assign_1_squeeze_mask_0, stride = new_cache_39_internal_tensor_assign_1_stride_0, x = reshape_190)[name = tensor("slice_by_index_38")]; + tensor reshape_191_shape_0 = const()[name = tensor("reshape_191_shape_0"), val = tensor([-1])]; + tensor reshape_191 = reshape(shape = reshape_191_shape_0, x = slice_by_index_38)[name = tensor("reshape_191")]; + tensor reshape_192_shape_0 = const()[name = tensor("reshape_192_shape_0"), val = tensor([-1])]; + tensor reshape_192_cast_fp16 = reshape(shape = reshape_192_shape_0, x = var_8251_cast_fp16)[name = tensor("reshape_192_cast_fp16")]; + tensor reshape_193_shape_0 = const()[name = tensor("reshape_193_shape_0"), val = tensor([-1])]; + tensor reshape_193_cast_fp16 = reshape(shape = reshape_193_shape_0, x = cache19_to_fp16)[name = tensor("reshape_193_cast_fp16")]; + tensor scatter_38_mode_0 = const()[name = tensor("scatter_38_mode_0"), val = tensor("update")]; + tensor scatter_38_axis_0 = const()[name = tensor("scatter_38_axis_0"), val = tensor(0)]; + tensor scatter_38_validate_indices_0 = const()[name = tensor("scatter_38_validate_indices_0"), val = tensor(false)]; + tensor scatter_38_cast_fp16 = scatter(axis = scatter_38_axis_0, data = reshape_193_cast_fp16, indices = reshape_191, mode = scatter_38_mode_0, updates = reshape_192_cast_fp16, validate_indices = scatter_38_validate_indices_0)[name = tensor("scatter_38_cast_fp16")]; + tensor reshape_194_cast_fp16 = reshape(shape = shape_84, x = scatter_38_cast_fp16)[name = tensor("reshape_194_cast_fp16")]; + tensor var_8259_begin_0 = const()[name = tensor("op_8259_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor var_8259_end_0 = const()[name = tensor("op_8259_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor var_8259_end_mask_0 = const()[name = tensor("op_8259_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_8259_squeeze_mask_0 = const()[name = tensor("op_8259_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor var_8259_cast_fp16 = slice_by_index(begin = var_8259_begin_0, end = var_8259_end_0, end_mask = var_8259_end_mask_0, squeeze_mask = var_8259_squeeze_mask_0, x = reshape_194_cast_fp16)[name = tensor("op_8259_cast_fp16")]; + tensor var_8261_axis_0 = const()[name = tensor("op_8261_axis_0"), val = tensor(1)]; + tensor var_8261_mode_0 = const()[name = tensor("op_8261_mode_0"), val = tensor("update")]; + tensor var_8261_validate_indices_0 = const()[name = tensor("op_8261_validate_indices_0"), val = tensor(false)]; + tensor var_8261_cast_fp16 = scatter_along_axis(axis = var_8261_axis_0, data = var_8259_cast_fp16, indices = write_indices_39, mode = var_8261_mode_0, updates = v_39_cast_fp16, validate_indices = var_8261_validate_indices_0)[name = tensor("op_8261_cast_fp16")]; + tensor concat_156 = const()[name = tensor("concat_156"), val = tensor([1, 0, 0, 0, 0])]; + tensor concat_157 = const()[name = tensor("concat_157"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_39_internal_tensor_assign_2_stride_0 = const()[name = tensor("new_cache_39_internal_tensor_assign_2_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_39_internal_tensor_assign_2_begin_mask_0 = const()[name = tensor("new_cache_39_internal_tensor_assign_2_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_39_internal_tensor_assign_2_end_mask_0 = const()[name = tensor("new_cache_39_internal_tensor_assign_2_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_39_internal_tensor_assign_2_squeeze_mask_0 = const()[name = tensor("new_cache_39_internal_tensor_assign_2_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_85 = const()[name = tensor("shape_85"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_39 = const()[name = tensor("reduce_prod_39"), val = tensor(1048576)]; + tensor range_1d_39_start_0 = const()[name = tensor("range_1d_39_start_0"), val = tensor(0)]; + tensor range_1d_39_step_0 = const()[name = tensor("range_1d_39_step_0"), val = tensor(1)]; + tensor range_1d_39 = range_1d(end = reduce_prod_39, start = range_1d_39_start_0, step = range_1d_39_step_0)[name = tensor("range_1d_39")]; + tensor reshape_195 = reshape(shape = shape_85, x = range_1d_39)[name = tensor("reshape_195")]; + tensor slice_by_index_39 = slice_by_index(begin = concat_156, begin_mask = new_cache_39_internal_tensor_assign_2_begin_mask_0, end = concat_157, end_mask = new_cache_39_internal_tensor_assign_2_end_mask_0, squeeze_mask = new_cache_39_internal_tensor_assign_2_squeeze_mask_0, stride = new_cache_39_internal_tensor_assign_2_stride_0, x = reshape_195)[name = tensor("slice_by_index_39")]; + tensor reshape_196_shape_0 = const()[name = tensor("reshape_196_shape_0"), val = tensor([-1])]; + tensor reshape_196 = reshape(shape = reshape_196_shape_0, x = slice_by_index_39)[name = tensor("reshape_196")]; + tensor reshape_197_shape_0 = const()[name = tensor("reshape_197_shape_0"), val = tensor([-1])]; + tensor reshape_197_cast_fp16 = reshape(shape = reshape_197_shape_0, x = var_8261_cast_fp16)[name = tensor("reshape_197_cast_fp16")]; + tensor reshape_198_shape_0 = const()[name = tensor("reshape_198_shape_0"), val = tensor([-1])]; + tensor reshape_198_cast_fp16 = reshape(shape = reshape_198_shape_0, x = reshape_194_cast_fp16)[name = tensor("reshape_198_cast_fp16")]; + tensor scatter_39_mode_0 = const()[name = tensor("scatter_39_mode_0"), val = tensor("update")]; + tensor scatter_39_axis_0 = const()[name = tensor("scatter_39_axis_0"), val = tensor(0)]; + tensor scatter_39_validate_indices_0 = const()[name = tensor("scatter_39_validate_indices_0"), val = tensor(false)]; + tensor scatter_39_cast_fp16 = scatter(axis = scatter_39_axis_0, data = reshape_198_cast_fp16, indices = reshape_196, mode = scatter_39_mode_0, updates = reshape_197_cast_fp16, validate_indices = scatter_39_validate_indices_0)[name = tensor("scatter_39_cast_fp16")]; + tensor reshape_199_cast_fp16 = reshape(shape = shape_85, x = scatter_39_cast_fp16)[name = tensor("reshape_199_cast_fp16")]; + tensor reshape_199_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("reshape_199_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor keys_115_begin_0 = const()[name = tensor("keys_115_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor keys_115_end_0 = const()[name = tensor("keys_115_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor keys_115_end_mask_0 = const()[name = tensor("keys_115_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor keys_115_squeeze_mask_0 = const()[name = tensor("keys_115_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor keys_115_cast_fp16 = slice_by_index(begin = keys_115_begin_0, end = keys_115_end_0, end_mask = keys_115_end_mask_0, squeeze_mask = keys_115_squeeze_mask_0, x = reshape_199_cast_fp16)[name = tensor("keys_115_cast_fp16")]; + tensor values_115_begin_0 = const()[name = tensor("values_115_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor values_115_end_0 = const()[name = tensor("values_115_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor values_115_end_mask_0 = const()[name = tensor("values_115_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor values_115_squeeze_mask_0 = const()[name = tensor("values_115_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor values_115_cast_fp16 = slice_by_index(begin = values_115_begin_0, end = values_115_end_0, end_mask = values_115_end_mask_0, squeeze_mask = values_115_squeeze_mask_0, x = reshape_199_cast_fp16)[name = tensor("values_115_cast_fp16")]; + tensor var_8273_cast_fp16 = not_equal(x = keys_115_cast_fp16, y = keys_115_cast_fp16)[name = tensor("op_8273_cast_fp16")]; + tensor keys_117_cast_fp16 = select(a = var_470_to_fp16, b = keys_115_cast_fp16, cond = var_8273_cast_fp16)[name = tensor("keys_117_cast_fp16")]; + tensor var_8281_cast_fp16 = not_equal(x = values_115_cast_fp16, y = values_115_cast_fp16)[name = tensor("op_8281_cast_fp16")]; + tensor values_117_cast_fp16 = select(a = var_470_to_fp16, b = values_115_cast_fp16, cond = var_8281_cast_fp16)[name = tensor("values_117_cast_fp16")]; + tensor var_8305 = const()[name = tensor("op_8305"), val = tensor([0, 2, 1, 3])]; + tensor var_8321 = const()[name = tensor("op_8321"), val = tensor([1, 1, 1])]; + tensor var_8322_cast_fp16 = reshape(shape = var_8321, x = position19_to_fp16)[name = tensor("op_8322_cast_fp16")]; + tensor q_positions_39_cast_fp16 = add(x = var_8322_cast_fp16, y = q_offsets_1_promoted_to_fp16)[name = tensor("q_positions_39_cast_fp16")]; + tensor valid_end_39_cast_fp16 = add(x = var_8322_cast_fp16, y = var_536_cast_fp16)[name = tensor("valid_end_39_cast_fp16")]; + tensor valid_mask_39_cast_fp16 = less(x = k_positions_1_promoted_to_fp16, y = valid_end_39_cast_fp16)[name = tensor("valid_mask_39_cast_fp16")]; + tensor causal_mask_39_cast_fp16 = less_equal(x = k_positions_1_promoted_to_fp16, y = q_positions_39_cast_fp16)[name = tensor("causal_mask_39_cast_fp16")]; + tensor var_8350 = logical_and(x = valid_mask_39_cast_fp16, y = causal_mask_39_cast_fp16)[name = tensor("op_8350")]; + tensor attn_mask_39_axes_0 = const()[name = tensor("attn_mask_39_axes_0"), val = tensor([1])]; + tensor attn_mask_39 = expand_dims(axes = attn_mask_39_axes_0, x = var_8350)[name = tensor("attn_mask_39")]; + tensor var_8364_transpose_x_0 = const()[name = tensor("op_8364_transpose_x_0"), val = tensor(false)]; + tensor var_8364_transpose_y_0 = const()[name = tensor("op_8364_transpose_y_0"), val = tensor(false)]; + tensor transpose_107_perm_0 = const()[name = tensor("transpose_107_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_108_perm_0 = const()[name = tensor("transpose_108_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_108 = transpose(perm = transpose_108_perm_0, x = keys_117_cast_fp16)[name = tensor("transpose_128")]; + tensor transpose_107 = transpose(perm = transpose_107_perm_0, x = q_117_cast_fp16)[name = tensor("transpose_129")]; + tensor var_8364_cast_fp16 = matmul(transpose_x = var_8364_transpose_x_0, transpose_y = var_8364_transpose_y_0, x = transpose_107, y = transpose_108)[name = tensor("op_8364_cast_fp16")]; + tensor var_8358_to_fp16 = const()[name = tensor("op_8358_to_fp16"), val = tensor([0x1p-3])]; + tensor attn_weights_115_cast_fp16 = mul(x = var_8364_cast_fp16, y = var_8358_to_fp16)[name = tensor("attn_weights_115_cast_fp16")]; + tensor var_8366 = logical_not(x = attn_mask_39)[name = tensor("op_8366")]; + tensor var_8367_to_fp16 = const()[name = tensor("op_8367_to_fp16"), val = tensor(-inf)]; + tensor attn_weights_117_cast_fp16 = select(a = var_8367_to_fp16, b = attn_weights_115_cast_fp16, cond = var_8366)[name = tensor("attn_weights_117_cast_fp16")]; + tensor var_8369 = const()[name = tensor("op_8369"), val = tensor(-1)]; + tensor attn_weights_119_cast_fp16 = softmax(axis = var_8369, x = attn_weights_117_cast_fp16)[name = tensor("attn_weights_119_cast_fp16")]; + tensor attn_output_39_transpose_x_0 = const()[name = tensor("attn_output_39_transpose_x_0"), val = tensor(false)]; + tensor attn_output_39_transpose_y_0 = const()[name = tensor("attn_output_39_transpose_y_0"), val = tensor(false)]; + tensor values_119_cast_fp16 = transpose(perm = var_8305, x = values_117_cast_fp16)[name = tensor("transpose_130")]; + tensor attn_output_39_cast_fp16 = matmul(transpose_x = attn_output_39_transpose_x_0, transpose_y = attn_output_39_transpose_y_0, x = attn_weights_119_cast_fp16, y = values_119_cast_fp16)[name = tensor("attn_output_39_cast_fp16")]; + tensor var_8377 = const()[name = tensor("op_8377"), val = tensor([0, 2, 1, 3])]; + tensor var_8382 = const()[name = tensor("op_8382"), val = tensor([1, 256, 1024])]; + tensor var_8378_cast_fp16 = transpose(perm = var_8377, x = attn_output_39_cast_fp16)[name = tensor("transpose_127")]; + tensor input_233_cast_fp16 = reshape(shape = var_8382, x = var_8378_cast_fp16)[name = tensor("input_233_cast_fp16")]; + tensor attn19_out_proj_weight_to_fp16 = const()[name = tensor("attn19_out_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(485682432)))]; + tensor linear_77_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = attn19_out_proj_weight_to_fp16, x = input_233_cast_fp16)[name = tensor("linear_77_cast_fp16")]; + tensor var_8391_cast_fp16 = add(x = position19_to_fp16, y = valid_len_to_fp16)[name = tensor("op_8391_cast_fp16")]; + tensor var_8391_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_8391_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor input_235_cast_fp16 = add(x = input_229_cast_fp16, y = linear_77_cast_fp16)[name = tensor("input_235_cast_fp16")]; + tensor input_237_axes_0 = const()[name = tensor("input_237_axes_0"), val = tensor([-1])]; + tensor norm19_2_weight_to_fp16 = const()[name = tensor("norm19_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(487779648)))]; + tensor norm19_2_bias_to_fp16 = const()[name = tensor("norm19_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(487781760)))]; + tensor var_8395_to_fp16 = const()[name = tensor("op_8395_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_237_cast_fp16 = layer_norm(axes = input_237_axes_0, beta = norm19_2_bias_to_fp16, epsilon = var_8395_to_fp16, gamma = norm19_2_weight_to_fp16, x = input_235_cast_fp16)[name = tensor("input_237_cast_fp16")]; + tensor linear19_1_weight_to_fp16 = const()[name = tensor("linear19_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(487783872)))]; + tensor linear_78_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = linear19_1_weight_to_fp16, x = input_237_cast_fp16)[name = tensor("linear_78_cast_fp16")]; + tensor input_239_mode_0 = const()[name = tensor("input_239_mode_0"), val = tensor("EXACT")]; + tensor input_239_cast_fp16 = gelu(mode = input_239_mode_0, x = linear_78_cast_fp16)[name = tensor("input_239_cast_fp16")]; + tensor linear19_2_weight_to_fp16 = const()[name = tensor("linear19_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(496172544)))]; + tensor linear_79_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = linear19_2_weight_to_fp16, x = input_239_cast_fp16)[name = tensor("linear_79_cast_fp16")]; + tensor input_241_cast_fp16 = add(x = input_235_cast_fp16, y = linear_79_cast_fp16)[name = tensor("input_241_cast_fp16")]; + tensor input_243_axes_0 = const()[name = tensor("input_243_axes_0"), val = tensor([-1])]; + tensor norm20_1_weight_to_fp16 = const()[name = tensor("norm20_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(504561216)))]; + tensor norm20_1_bias_to_fp16 = const()[name = tensor("norm20_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(504563328)))]; + tensor var_8412_to_fp16 = const()[name = tensor("op_8412_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_243_cast_fp16 = layer_norm(axes = input_243_axes_0, beta = norm20_1_bias_to_fp16, epsilon = var_8412_to_fp16, gamma = norm20_1_weight_to_fp16, x = input_241_cast_fp16)[name = tensor("input_243_cast_fp16")]; + tensor attn20_in_proj_weight_to_fp16 = const()[name = tensor("attn20_in_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(504565440)))]; + tensor linear_80_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = attn20_in_proj_weight_to_fp16, x = input_243_cast_fp16)[name = tensor("linear_80_cast_fp16")]; + tensor var_8426 = const()[name = tensor("op_8426"), val = tensor([1, 256, 3, 16, 64])]; + tensor qkv_41_cast_fp16 = reshape(shape = var_8426, x = linear_80_cast_fp16)[name = tensor("qkv_41_cast_fp16")]; + tensor q_121_begin_0 = const()[name = tensor("q_121_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor q_121_end_0 = const()[name = tensor("q_121_end_0"), val = tensor([1, 256, 1, 16, 64])]; + tensor q_121_end_mask_0 = const()[name = tensor("q_121_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor q_121_squeeze_mask_0 = const()[name = tensor("q_121_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor q_121_cast_fp16 = slice_by_index(begin = q_121_begin_0, end = q_121_end_0, end_mask = q_121_end_mask_0, squeeze_mask = q_121_squeeze_mask_0, x = qkv_41_cast_fp16)[name = tensor("q_121_cast_fp16")]; + tensor k_81_begin_0 = const()[name = tensor("k_81_begin_0"), val = tensor([0, 0, 1, 0, 0])]; + tensor k_81_end_0 = const()[name = tensor("k_81_end_0"), val = tensor([1, 256, 2, 16, 64])]; + tensor k_81_end_mask_0 = const()[name = tensor("k_81_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor k_81_squeeze_mask_0 = const()[name = tensor("k_81_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor k_81_cast_fp16 = slice_by_index(begin = k_81_begin_0, end = k_81_end_0, end_mask = k_81_end_mask_0, squeeze_mask = k_81_squeeze_mask_0, x = qkv_41_cast_fp16)[name = tensor("k_81_cast_fp16")]; + tensor v_41_begin_0 = const()[name = tensor("v_41_begin_0"), val = tensor([0, 0, 2, 0, 0])]; + tensor v_41_end_0 = const()[name = tensor("v_41_end_0"), val = tensor([1, 256, 3, 16, 64])]; + tensor v_41_end_mask_0 = const()[name = tensor("v_41_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor v_41_squeeze_mask_0 = const()[name = tensor("v_41_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor v_41_cast_fp16 = slice_by_index(begin = v_41_begin_0, end = v_41_end_0, end_mask = v_41_end_mask_0, squeeze_mask = v_41_squeeze_mask_0, x = qkv_41_cast_fp16)[name = tensor("v_41_cast_fp16")]; + tensor var_8484 = const()[name = tensor("op_8484"), val = tensor([1, 1])]; + tensor position20_to_fp16_dtype_0 = const()[name = tensor("position20_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor position20_to_fp16 = cast(dtype = position20_to_fp16_dtype_0, x = position20)[name = tensor("cast_561")]; + tensor var_8485_cast_fp16 = reshape(shape = var_8484, x = position20_to_fp16)[name = tensor("op_8485_cast_fp16")]; + tensor ts_123_cast_fp16 = add(x = ts_1_promoted_to_fp16, y = var_8485_cast_fp16)[name = tensor("ts_123_cast_fp16")]; + tensor var_8492 = const()[name = tensor("op_8492"), val = tensor([1, 256, 1, 1])]; + tensor ts_125_cast_fp16 = reshape(shape = var_8492, x = ts_123_cast_fp16)[name = tensor("ts_125_cast_fp16")]; + tensor var_8499 = const()[name = tensor("op_8499"), val = tensor([1, 256, 16, 32, 2])]; + tensor q_complex_41_cast_fp16 = reshape(shape = var_8499, x = q_121_cast_fp16)[name = tensor("q_complex_41_cast_fp16")]; + tensor var_8506 = const()[name = tensor("op_8506"), val = tensor([1, 256, 16, 32, 2])]; + tensor k_complex_41_cast_fp16 = reshape(shape = var_8506, x = k_81_cast_fp16)[name = tensor("k_complex_41_cast_fp16")]; + tensor var_8510_begin_0 = const()[name = tensor("op_8510_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_8510_end_0 = const()[name = tensor("op_8510_end_0"), val = tensor([1, 256, 16, 32, 1])]; + tensor var_8510_end_mask_0 = const()[name = tensor("op_8510_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_8510_squeeze_mask_0 = const()[name = tensor("op_8510_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_8510_cast_fp16 = slice_by_index(begin = var_8510_begin_0, end = var_8510_end_0, end_mask = var_8510_end_mask_0, squeeze_mask = var_8510_squeeze_mask_0, x = q_complex_41_cast_fp16)[name = tensor("op_8510_cast_fp16")]; + tensor var_8518_begin_0 = const()[name = tensor("op_8518_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_8518_end_0 = const()[name = tensor("op_8518_end_0"), val = tensor([1, 256, 16, 32, 2])]; + tensor var_8518_end_mask_0 = const()[name = tensor("op_8518_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_8518_squeeze_mask_0 = const()[name = tensor("op_8518_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_8518_cast_fp16 = slice_by_index(begin = var_8518_begin_0, end = var_8518_end_0, end_mask = var_8518_end_mask_0, squeeze_mask = var_8518_squeeze_mask_0, x = q_complex_41_cast_fp16)[name = tensor("op_8518_cast_fp16")]; + tensor var_8526_begin_0 = const()[name = tensor("op_8526_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_8526_end_0 = const()[name = tensor("op_8526_end_0"), val = tensor([1, 256, 16, 32, 1])]; + tensor var_8526_end_mask_0 = const()[name = tensor("op_8526_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_8526_squeeze_mask_0 = const()[name = tensor("op_8526_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_8526_cast_fp16 = slice_by_index(begin = var_8526_begin_0, end = var_8526_end_0, end_mask = var_8526_end_mask_0, squeeze_mask = var_8526_squeeze_mask_0, x = k_complex_41_cast_fp16)[name = tensor("op_8526_cast_fp16")]; + tensor var_8534_begin_0 = const()[name = tensor("op_8534_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_8534_end_0 = const()[name = tensor("op_8534_end_0"), val = tensor([1, 256, 16, 32, 2])]; + tensor var_8534_end_mask_0 = const()[name = tensor("op_8534_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_8534_squeeze_mask_0 = const()[name = tensor("op_8534_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_8534_cast_fp16 = slice_by_index(begin = var_8534_begin_0, end = var_8534_end_0, end_mask = var_8534_end_mask_0, squeeze_mask = var_8534_squeeze_mask_0, x = k_complex_41_cast_fp16)[name = tensor("op_8534_cast_fp16")]; + tensor freqs_41_to_fp16 = const()[name = tensor("freqs_41_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(510856960)))]; + tensor var_8540_cast_fp16 = mul(x = freqs_41_to_fp16, y = ts_125_cast_fp16)[name = tensor("op_8540_cast_fp16")]; + tensor rotr_41_cast_fp16 = cos(x = var_8540_cast_fp16)[name = tensor("rotr_41_cast_fp16")]; + tensor roti_41_cast_fp16 = sin(x = var_8540_cast_fp16)[name = tensor("roti_41_cast_fp16")]; + tensor var_8544_cast_fp16 = mul(x = var_8510_cast_fp16, y = rotr_41_cast_fp16)[name = tensor("op_8544_cast_fp16")]; + tensor var_8545_cast_fp16 = mul(x = var_8518_cast_fp16, y = roti_41_cast_fp16)[name = tensor("op_8545_cast_fp16")]; + tensor qor_81_cast_fp16 = sub(x = var_8544_cast_fp16, y = var_8545_cast_fp16)[name = tensor("qor_81_cast_fp16")]; + tensor var_8548_cast_fp16 = mul(x = var_8510_cast_fp16, y = roti_41_cast_fp16)[name = tensor("op_8548_cast_fp16")]; + tensor var_8549_cast_fp16 = mul(x = var_8518_cast_fp16, y = rotr_41_cast_fp16)[name = tensor("op_8549_cast_fp16")]; + tensor qoi_81_cast_fp16 = add(x = var_8548_cast_fp16, y = var_8549_cast_fp16)[name = tensor("qoi_81_cast_fp16")]; + tensor var_8552_cast_fp16 = mul(x = var_8526_cast_fp16, y = rotr_41_cast_fp16)[name = tensor("op_8552_cast_fp16")]; + tensor var_8553_cast_fp16 = mul(x = var_8534_cast_fp16, y = roti_41_cast_fp16)[name = tensor("op_8553_cast_fp16")]; + tensor kor_81_cast_fp16 = sub(x = var_8552_cast_fp16, y = var_8553_cast_fp16)[name = tensor("kor_81_cast_fp16")]; + tensor var_8556_cast_fp16 = mul(x = var_8526_cast_fp16, y = roti_41_cast_fp16)[name = tensor("op_8556_cast_fp16")]; + tensor var_8557_cast_fp16 = mul(x = var_8534_cast_fp16, y = rotr_41_cast_fp16)[name = tensor("op_8557_cast_fp16")]; + tensor koi_81_cast_fp16 = add(x = var_8556_cast_fp16, y = var_8557_cast_fp16)[name = tensor("koi_81_cast_fp16")]; + tensor qo_41_axis_0 = const()[name = tensor("qo_41_axis_0"), val = tensor(-1)]; + tensor qo_41_cast_fp16 = stack(axis = qo_41_axis_0, values = (qor_81_cast_fp16, qoi_81_cast_fp16))[name = tensor("qo_41_cast_fp16")]; + tensor ko_41_axis_0 = const()[name = tensor("ko_41_axis_0"), val = tensor(-1)]; + tensor ko_41_cast_fp16 = stack(axis = ko_41_axis_0, values = (kor_81_cast_fp16, koi_81_cast_fp16))[name = tensor("ko_41_cast_fp16")]; + tensor var_8590 = const()[name = tensor("op_8590"), val = tensor([1, 256, 16, 64])]; + tensor q_123_cast_fp16 = reshape(shape = var_8590, x = qo_41_cast_fp16)[name = tensor("q_123_cast_fp16")]; + tensor var_8596 = const()[name = tensor("op_8596"), val = tensor([1, 256, 16, 64])]; + tensor k_83_cast_fp16 = reshape(shape = var_8596, x = ko_41_cast_fp16)[name = tensor("k_83_cast_fp16")]; + tensor write_indices_float_121_cast_fp16 = add(x = var_8485_cast_fp16, y = var_393_promoted_to_fp16)[name = tensor("write_indices_float_121_cast_fp16")]; + tensor _inversed_8617_y_0_to_fp16 = const()[name = tensor("_inversed_8617_y_0_to_fp16"), val = tensor(0x1p-9)]; + tensor _inversed_8617_cast_fp16 = mul(x = write_indices_float_121_cast_fp16, y = _inversed_8617_y_0_to_fp16)[name = tensor("_inversed_8617_cast_fp16")]; + tensor var_8618_cast_fp16 = floor(x = _inversed_8617_cast_fp16)[name = tensor("op_8618_cast_fp16")]; + tensor var_8619_to_fp16 = const()[name = tensor("op_8619_to_fp16"), val = tensor(0x1p+9)]; + tensor var_8620_cast_fp16 = mul(x = var_8618_cast_fp16, y = var_8619_to_fp16)[name = tensor("op_8620_cast_fp16")]; + tensor write_indices_float_123_cast_fp16 = sub(x = write_indices_float_121_cast_fp16, y = var_8620_cast_fp16)[name = tensor("write_indices_float_123_cast_fp16")]; + tensor var_8638_value_0_to_fp16 = const()[name = tensor("op_8638_value_0_to_fp16"), val = tensor(0x1.ffp+8)]; + tensor var_8638_cast_fp16 = fill_like(ref_tensor = write_indices_float_123_cast_fp16, value = var_8638_value_0_to_fp16)[name = tensor("op_8638_cast_fp16")]; + tensor write_indices_float_125_cast_fp16 = select(a = write_indices_float_123_cast_fp16, b = var_8638_cast_fp16, cond = token_valid_1_cast_fp16)[name = tensor("write_indices_float_125_cast_fp16")]; + tensor var_8644_dtype_0 = const()[name = tensor("op_8644_dtype_0"), val = tensor("int32")]; + tensor var_8649 = const()[name = tensor("op_8649"), val = tensor([1, 256, 1, 1])]; + tensor write_indices_float_125_cast_fp16_to_int32 = cast(dtype = var_8644_dtype_0, x = write_indices_float_125_cast_fp16)[name = tensor("cast_560")]; + tensor var_8650 = reshape(shape = var_8649, x = write_indices_float_125_cast_fp16_to_int32)[name = tensor("op_8650")]; + tensor write_indices_41_reps_0 = const()[name = tensor("write_indices_41_reps_0"), val = tensor([1, 1, 16, 64])]; + tensor write_indices_41 = tile(reps = write_indices_41_reps_0, x = var_8650)[name = tensor("write_indices_41")]; + tensor var_8660_begin_0 = const()[name = tensor("op_8660_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_8660_end_0 = const()[name = tensor("op_8660_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor var_8660_end_mask_0 = const()[name = tensor("op_8660_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_8660_squeeze_mask_0 = const()[name = tensor("op_8660_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor cache20_to_fp16_dtype_0 = const()[name = tensor("cache20_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor cache20_to_fp16 = cast(dtype = cache20_to_fp16_dtype_0, x = cache20)[name = tensor("cast_559")]; + tensor var_8660_cast_fp16 = slice_by_index(begin = var_8660_begin_0, end = var_8660_end_0, end_mask = var_8660_end_mask_0, squeeze_mask = var_8660_squeeze_mask_0, x = cache20_to_fp16)[name = tensor("op_8660_cast_fp16")]; + tensor var_8662_axis_0 = const()[name = tensor("op_8662_axis_0"), val = tensor(1)]; + tensor var_8662_mode_0 = const()[name = tensor("op_8662_mode_0"), val = tensor("update")]; + tensor var_8662_validate_indices_0 = const()[name = tensor("op_8662_validate_indices_0"), val = tensor(false)]; + tensor var_8662_cast_fp16 = scatter_along_axis(axis = var_8662_axis_0, data = var_8660_cast_fp16, indices = write_indices_41, mode = var_8662_mode_0, updates = k_83_cast_fp16, validate_indices = var_8662_validate_indices_0)[name = tensor("op_8662_cast_fp16")]; + tensor concat_162 = const()[name = tensor("concat_162"), val = tensor([0, 0, 0, 0, 0])]; + tensor concat_163 = const()[name = tensor("concat_163"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_41_internal_tensor_assign_1_stride_0 = const()[name = tensor("new_cache_41_internal_tensor_assign_1_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_41_internal_tensor_assign_1_begin_mask_0 = const()[name = tensor("new_cache_41_internal_tensor_assign_1_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_41_internal_tensor_assign_1_end_mask_0 = const()[name = tensor("new_cache_41_internal_tensor_assign_1_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_41_internal_tensor_assign_1_squeeze_mask_0 = const()[name = tensor("new_cache_41_internal_tensor_assign_1_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_86 = const()[name = tensor("shape_86"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_40 = const()[name = tensor("reduce_prod_40"), val = tensor(1048576)]; + tensor range_1d_40_start_0 = const()[name = tensor("range_1d_40_start_0"), val = tensor(0)]; + tensor range_1d_40_step_0 = const()[name = tensor("range_1d_40_step_0"), val = tensor(1)]; + tensor range_1d_40 = range_1d(end = reduce_prod_40, start = range_1d_40_start_0, step = range_1d_40_step_0)[name = tensor("range_1d_40")]; + tensor reshape_200 = reshape(shape = shape_86, x = range_1d_40)[name = tensor("reshape_200")]; + tensor slice_by_index_40 = slice_by_index(begin = concat_162, begin_mask = new_cache_41_internal_tensor_assign_1_begin_mask_0, end = concat_163, end_mask = new_cache_41_internal_tensor_assign_1_end_mask_0, squeeze_mask = new_cache_41_internal_tensor_assign_1_squeeze_mask_0, stride = new_cache_41_internal_tensor_assign_1_stride_0, x = reshape_200)[name = tensor("slice_by_index_40")]; + tensor reshape_201_shape_0 = const()[name = tensor("reshape_201_shape_0"), val = tensor([-1])]; + tensor reshape_201 = reshape(shape = reshape_201_shape_0, x = slice_by_index_40)[name = tensor("reshape_201")]; + tensor reshape_202_shape_0 = const()[name = tensor("reshape_202_shape_0"), val = tensor([-1])]; + tensor reshape_202_cast_fp16 = reshape(shape = reshape_202_shape_0, x = var_8662_cast_fp16)[name = tensor("reshape_202_cast_fp16")]; + tensor reshape_203_shape_0 = const()[name = tensor("reshape_203_shape_0"), val = tensor([-1])]; + tensor reshape_203_cast_fp16 = reshape(shape = reshape_203_shape_0, x = cache20_to_fp16)[name = tensor("reshape_203_cast_fp16")]; + tensor scatter_40_mode_0 = const()[name = tensor("scatter_40_mode_0"), val = tensor("update")]; + tensor scatter_40_axis_0 = const()[name = tensor("scatter_40_axis_0"), val = tensor(0)]; + tensor scatter_40_validate_indices_0 = const()[name = tensor("scatter_40_validate_indices_0"), val = tensor(false)]; + tensor scatter_40_cast_fp16 = scatter(axis = scatter_40_axis_0, data = reshape_203_cast_fp16, indices = reshape_201, mode = scatter_40_mode_0, updates = reshape_202_cast_fp16, validate_indices = scatter_40_validate_indices_0)[name = tensor("scatter_40_cast_fp16")]; + tensor reshape_204_cast_fp16 = reshape(shape = shape_86, x = scatter_40_cast_fp16)[name = tensor("reshape_204_cast_fp16")]; + tensor var_8670_begin_0 = const()[name = tensor("op_8670_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor var_8670_end_0 = const()[name = tensor("op_8670_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor var_8670_end_mask_0 = const()[name = tensor("op_8670_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_8670_squeeze_mask_0 = const()[name = tensor("op_8670_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor var_8670_cast_fp16 = slice_by_index(begin = var_8670_begin_0, end = var_8670_end_0, end_mask = var_8670_end_mask_0, squeeze_mask = var_8670_squeeze_mask_0, x = reshape_204_cast_fp16)[name = tensor("op_8670_cast_fp16")]; + tensor var_8672_axis_0 = const()[name = tensor("op_8672_axis_0"), val = tensor(1)]; + tensor var_8672_mode_0 = const()[name = tensor("op_8672_mode_0"), val = tensor("update")]; + tensor var_8672_validate_indices_0 = const()[name = tensor("op_8672_validate_indices_0"), val = tensor(false)]; + tensor var_8672_cast_fp16 = scatter_along_axis(axis = var_8672_axis_0, data = var_8670_cast_fp16, indices = write_indices_41, mode = var_8672_mode_0, updates = v_41_cast_fp16, validate_indices = var_8672_validate_indices_0)[name = tensor("op_8672_cast_fp16")]; + tensor concat_164 = const()[name = tensor("concat_164"), val = tensor([1, 0, 0, 0, 0])]; + tensor concat_165 = const()[name = tensor("concat_165"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_41_internal_tensor_assign_2_stride_0 = const()[name = tensor("new_cache_41_internal_tensor_assign_2_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_41_internal_tensor_assign_2_begin_mask_0 = const()[name = tensor("new_cache_41_internal_tensor_assign_2_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_41_internal_tensor_assign_2_end_mask_0 = const()[name = tensor("new_cache_41_internal_tensor_assign_2_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_41_internal_tensor_assign_2_squeeze_mask_0 = const()[name = tensor("new_cache_41_internal_tensor_assign_2_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_87 = const()[name = tensor("shape_87"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_41 = const()[name = tensor("reduce_prod_41"), val = tensor(1048576)]; + tensor range_1d_41_start_0 = const()[name = tensor("range_1d_41_start_0"), val = tensor(0)]; + tensor range_1d_41_step_0 = const()[name = tensor("range_1d_41_step_0"), val = tensor(1)]; + tensor range_1d_41 = range_1d(end = reduce_prod_41, start = range_1d_41_start_0, step = range_1d_41_step_0)[name = tensor("range_1d_41")]; + tensor reshape_205 = reshape(shape = shape_87, x = range_1d_41)[name = tensor("reshape_205")]; + tensor slice_by_index_41 = slice_by_index(begin = concat_164, begin_mask = new_cache_41_internal_tensor_assign_2_begin_mask_0, end = concat_165, end_mask = new_cache_41_internal_tensor_assign_2_end_mask_0, squeeze_mask = new_cache_41_internal_tensor_assign_2_squeeze_mask_0, stride = new_cache_41_internal_tensor_assign_2_stride_0, x = reshape_205)[name = tensor("slice_by_index_41")]; + tensor reshape_206_shape_0 = const()[name = tensor("reshape_206_shape_0"), val = tensor([-1])]; + tensor reshape_206 = reshape(shape = reshape_206_shape_0, x = slice_by_index_41)[name = tensor("reshape_206")]; + tensor reshape_207_shape_0 = const()[name = tensor("reshape_207_shape_0"), val = tensor([-1])]; + tensor reshape_207_cast_fp16 = reshape(shape = reshape_207_shape_0, x = var_8672_cast_fp16)[name = tensor("reshape_207_cast_fp16")]; + tensor reshape_208_shape_0 = const()[name = tensor("reshape_208_shape_0"), val = tensor([-1])]; + tensor reshape_208_cast_fp16 = reshape(shape = reshape_208_shape_0, x = reshape_204_cast_fp16)[name = tensor("reshape_208_cast_fp16")]; + tensor scatter_41_mode_0 = const()[name = tensor("scatter_41_mode_0"), val = tensor("update")]; + tensor scatter_41_axis_0 = const()[name = tensor("scatter_41_axis_0"), val = tensor(0)]; + tensor scatter_41_validate_indices_0 = const()[name = tensor("scatter_41_validate_indices_0"), val = tensor(false)]; + tensor scatter_41_cast_fp16 = scatter(axis = scatter_41_axis_0, data = reshape_208_cast_fp16, indices = reshape_206, mode = scatter_41_mode_0, updates = reshape_207_cast_fp16, validate_indices = scatter_41_validate_indices_0)[name = tensor("scatter_41_cast_fp16")]; + tensor reshape_209_cast_fp16 = reshape(shape = shape_87, x = scatter_41_cast_fp16)[name = tensor("reshape_209_cast_fp16")]; + tensor reshape_209_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("reshape_209_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor keys_121_begin_0 = const()[name = tensor("keys_121_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor keys_121_end_0 = const()[name = tensor("keys_121_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor keys_121_end_mask_0 = const()[name = tensor("keys_121_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor keys_121_squeeze_mask_0 = const()[name = tensor("keys_121_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor keys_121_cast_fp16 = slice_by_index(begin = keys_121_begin_0, end = keys_121_end_0, end_mask = keys_121_end_mask_0, squeeze_mask = keys_121_squeeze_mask_0, x = reshape_209_cast_fp16)[name = tensor("keys_121_cast_fp16")]; + tensor values_121_begin_0 = const()[name = tensor("values_121_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor values_121_end_0 = const()[name = tensor("values_121_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor values_121_end_mask_0 = const()[name = tensor("values_121_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor values_121_squeeze_mask_0 = const()[name = tensor("values_121_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor values_121_cast_fp16 = slice_by_index(begin = values_121_begin_0, end = values_121_end_0, end_mask = values_121_end_mask_0, squeeze_mask = values_121_squeeze_mask_0, x = reshape_209_cast_fp16)[name = tensor("values_121_cast_fp16")]; + tensor var_8684_cast_fp16 = not_equal(x = keys_121_cast_fp16, y = keys_121_cast_fp16)[name = tensor("op_8684_cast_fp16")]; + tensor keys_123_cast_fp16 = select(a = var_470_to_fp16, b = keys_121_cast_fp16, cond = var_8684_cast_fp16)[name = tensor("keys_123_cast_fp16")]; + tensor var_8692_cast_fp16 = not_equal(x = values_121_cast_fp16, y = values_121_cast_fp16)[name = tensor("op_8692_cast_fp16")]; + tensor values_123_cast_fp16 = select(a = var_470_to_fp16, b = values_121_cast_fp16, cond = var_8692_cast_fp16)[name = tensor("values_123_cast_fp16")]; + tensor var_8716 = const()[name = tensor("op_8716"), val = tensor([0, 2, 1, 3])]; + tensor var_8732 = const()[name = tensor("op_8732"), val = tensor([1, 1, 1])]; + tensor var_8733_cast_fp16 = reshape(shape = var_8732, x = position20_to_fp16)[name = tensor("op_8733_cast_fp16")]; + tensor q_positions_41_cast_fp16 = add(x = var_8733_cast_fp16, y = q_offsets_1_promoted_to_fp16)[name = tensor("q_positions_41_cast_fp16")]; + tensor valid_end_41_cast_fp16 = add(x = var_8733_cast_fp16, y = var_536_cast_fp16)[name = tensor("valid_end_41_cast_fp16")]; + tensor valid_mask_41_cast_fp16 = less(x = k_positions_1_promoted_to_fp16, y = valid_end_41_cast_fp16)[name = tensor("valid_mask_41_cast_fp16")]; + tensor causal_mask_41_cast_fp16 = less_equal(x = k_positions_1_promoted_to_fp16, y = q_positions_41_cast_fp16)[name = tensor("causal_mask_41_cast_fp16")]; + tensor var_8761 = logical_and(x = valid_mask_41_cast_fp16, y = causal_mask_41_cast_fp16)[name = tensor("op_8761")]; + tensor attn_mask_41_axes_0 = const()[name = tensor("attn_mask_41_axes_0"), val = tensor([1])]; + tensor attn_mask_41 = expand_dims(axes = attn_mask_41_axes_0, x = var_8761)[name = tensor("attn_mask_41")]; + tensor var_8775_transpose_x_0 = const()[name = tensor("op_8775_transpose_x_0"), val = tensor(false)]; + tensor var_8775_transpose_y_0 = const()[name = tensor("op_8775_transpose_y_0"), val = tensor(false)]; + tensor transpose_109_perm_0 = const()[name = tensor("transpose_109_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_110_perm_0 = const()[name = tensor("transpose_110_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_110 = transpose(perm = transpose_110_perm_0, x = keys_123_cast_fp16)[name = tensor("transpose_124")]; + tensor transpose_109 = transpose(perm = transpose_109_perm_0, x = q_123_cast_fp16)[name = tensor("transpose_125")]; + tensor var_8775_cast_fp16 = matmul(transpose_x = var_8775_transpose_x_0, transpose_y = var_8775_transpose_y_0, x = transpose_109, y = transpose_110)[name = tensor("op_8775_cast_fp16")]; + tensor var_8769_to_fp16 = const()[name = tensor("op_8769_to_fp16"), val = tensor([0x1p-3])]; + tensor attn_weights_121_cast_fp16 = mul(x = var_8775_cast_fp16, y = var_8769_to_fp16)[name = tensor("attn_weights_121_cast_fp16")]; + tensor var_8777 = logical_not(x = attn_mask_41)[name = tensor("op_8777")]; + tensor var_8778_to_fp16 = const()[name = tensor("op_8778_to_fp16"), val = tensor(-inf)]; + tensor attn_weights_123_cast_fp16 = select(a = var_8778_to_fp16, b = attn_weights_121_cast_fp16, cond = var_8777)[name = tensor("attn_weights_123_cast_fp16")]; + tensor var_8780 = const()[name = tensor("op_8780"), val = tensor(-1)]; + tensor attn_weights_125_cast_fp16 = softmax(axis = var_8780, x = attn_weights_123_cast_fp16)[name = tensor("attn_weights_125_cast_fp16")]; + tensor attn_output_41_transpose_x_0 = const()[name = tensor("attn_output_41_transpose_x_0"), val = tensor(false)]; + tensor attn_output_41_transpose_y_0 = const()[name = tensor("attn_output_41_transpose_y_0"), val = tensor(false)]; + tensor values_125_cast_fp16 = transpose(perm = var_8716, x = values_123_cast_fp16)[name = tensor("transpose_126")]; + tensor attn_output_41_cast_fp16 = matmul(transpose_x = attn_output_41_transpose_x_0, transpose_y = attn_output_41_transpose_y_0, x = attn_weights_125_cast_fp16, y = values_125_cast_fp16)[name = tensor("attn_output_41_cast_fp16")]; + tensor var_8788 = const()[name = tensor("op_8788"), val = tensor([0, 2, 1, 3])]; + tensor var_8793 = const()[name = tensor("op_8793"), val = tensor([1, 256, 1024])]; + tensor var_8789_cast_fp16 = transpose(perm = var_8788, x = attn_output_41_cast_fp16)[name = tensor("transpose_123")]; + tensor input_245_cast_fp16 = reshape(shape = var_8793, x = var_8789_cast_fp16)[name = tensor("input_245_cast_fp16")]; + tensor attn20_out_proj_weight_to_fp16 = const()[name = tensor("attn20_out_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(510857088)))]; + tensor linear_81_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = attn20_out_proj_weight_to_fp16, x = input_245_cast_fp16)[name = tensor("linear_81_cast_fp16")]; + tensor var_8802_cast_fp16 = add(x = position20_to_fp16, y = valid_len_to_fp16)[name = tensor("op_8802_cast_fp16")]; + tensor var_8802_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_8802_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor input_247_cast_fp16 = add(x = input_241_cast_fp16, y = linear_81_cast_fp16)[name = tensor("input_247_cast_fp16")]; + tensor input_249_axes_0 = const()[name = tensor("input_249_axes_0"), val = tensor([-1])]; + tensor norm20_2_weight_to_fp16 = const()[name = tensor("norm20_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(512954304)))]; + tensor norm20_2_bias_to_fp16 = const()[name = tensor("norm20_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(512956416)))]; + tensor var_8806_to_fp16 = const()[name = tensor("op_8806_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_249_cast_fp16 = layer_norm(axes = input_249_axes_0, beta = norm20_2_bias_to_fp16, epsilon = var_8806_to_fp16, gamma = norm20_2_weight_to_fp16, x = input_247_cast_fp16)[name = tensor("input_249_cast_fp16")]; + tensor linear20_1_weight_to_fp16 = const()[name = tensor("linear20_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(512958528)))]; + tensor linear_82_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = linear20_1_weight_to_fp16, x = input_249_cast_fp16)[name = tensor("linear_82_cast_fp16")]; + tensor input_251_mode_0 = const()[name = tensor("input_251_mode_0"), val = tensor("EXACT")]; + tensor input_251_cast_fp16 = gelu(mode = input_251_mode_0, x = linear_82_cast_fp16)[name = tensor("input_251_cast_fp16")]; + tensor linear20_2_weight_to_fp16 = const()[name = tensor("linear20_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(521347200)))]; + tensor linear_83_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = linear20_2_weight_to_fp16, x = input_251_cast_fp16)[name = tensor("linear_83_cast_fp16")]; + tensor input_253_cast_fp16 = add(x = input_247_cast_fp16, y = linear_83_cast_fp16)[name = tensor("input_253_cast_fp16")]; + tensor input_255_axes_0 = const()[name = tensor("input_255_axes_0"), val = tensor([-1])]; + tensor norm21_1_weight_to_fp16 = const()[name = tensor("norm21_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(529735872)))]; + tensor norm21_1_bias_to_fp16 = const()[name = tensor("norm21_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(529737984)))]; + tensor var_8823_to_fp16 = const()[name = tensor("op_8823_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_255_cast_fp16 = layer_norm(axes = input_255_axes_0, beta = norm21_1_bias_to_fp16, epsilon = var_8823_to_fp16, gamma = norm21_1_weight_to_fp16, x = input_253_cast_fp16)[name = tensor("input_255_cast_fp16")]; + tensor attn21_in_proj_weight_to_fp16 = const()[name = tensor("attn21_in_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(529740096)))]; + tensor linear_84_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = attn21_in_proj_weight_to_fp16, x = input_255_cast_fp16)[name = tensor("linear_84_cast_fp16")]; + tensor var_8837 = const()[name = tensor("op_8837"), val = tensor([1, 256, 3, 16, 64])]; + tensor qkv_43_cast_fp16 = reshape(shape = var_8837, x = linear_84_cast_fp16)[name = tensor("qkv_43_cast_fp16")]; + tensor q_127_begin_0 = const()[name = tensor("q_127_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor q_127_end_0 = const()[name = tensor("q_127_end_0"), val = tensor([1, 256, 1, 16, 64])]; + tensor q_127_end_mask_0 = const()[name = tensor("q_127_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor q_127_squeeze_mask_0 = const()[name = tensor("q_127_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor q_127_cast_fp16 = slice_by_index(begin = q_127_begin_0, end = q_127_end_0, end_mask = q_127_end_mask_0, squeeze_mask = q_127_squeeze_mask_0, x = qkv_43_cast_fp16)[name = tensor("q_127_cast_fp16")]; + tensor k_85_begin_0 = const()[name = tensor("k_85_begin_0"), val = tensor([0, 0, 1, 0, 0])]; + tensor k_85_end_0 = const()[name = tensor("k_85_end_0"), val = tensor([1, 256, 2, 16, 64])]; + tensor k_85_end_mask_0 = const()[name = tensor("k_85_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor k_85_squeeze_mask_0 = const()[name = tensor("k_85_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor k_85_cast_fp16 = slice_by_index(begin = k_85_begin_0, end = k_85_end_0, end_mask = k_85_end_mask_0, squeeze_mask = k_85_squeeze_mask_0, x = qkv_43_cast_fp16)[name = tensor("k_85_cast_fp16")]; + tensor v_43_begin_0 = const()[name = tensor("v_43_begin_0"), val = tensor([0, 0, 2, 0, 0])]; + tensor v_43_end_0 = const()[name = tensor("v_43_end_0"), val = tensor([1, 256, 3, 16, 64])]; + tensor v_43_end_mask_0 = const()[name = tensor("v_43_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor v_43_squeeze_mask_0 = const()[name = tensor("v_43_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor v_43_cast_fp16 = slice_by_index(begin = v_43_begin_0, end = v_43_end_0, end_mask = v_43_end_mask_0, squeeze_mask = v_43_squeeze_mask_0, x = qkv_43_cast_fp16)[name = tensor("v_43_cast_fp16")]; + tensor var_8895 = const()[name = tensor("op_8895"), val = tensor([1, 1])]; + tensor position21_to_fp16_dtype_0 = const()[name = tensor("position21_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor position21_to_fp16 = cast(dtype = position21_to_fp16_dtype_0, x = position21)[name = tensor("cast_556")]; + tensor var_8896_cast_fp16 = reshape(shape = var_8895, x = position21_to_fp16)[name = tensor("op_8896_cast_fp16")]; + tensor ts_129_cast_fp16 = add(x = ts_1_promoted_to_fp16, y = var_8896_cast_fp16)[name = tensor("ts_129_cast_fp16")]; + tensor var_8903 = const()[name = tensor("op_8903"), val = tensor([1, 256, 1, 1])]; + tensor ts_131_cast_fp16 = reshape(shape = var_8903, x = ts_129_cast_fp16)[name = tensor("ts_131_cast_fp16")]; + tensor var_8910 = const()[name = tensor("op_8910"), val = tensor([1, 256, 16, 32, 2])]; + tensor q_complex_43_cast_fp16 = reshape(shape = var_8910, x = q_127_cast_fp16)[name = tensor("q_complex_43_cast_fp16")]; + tensor var_8917 = const()[name = tensor("op_8917"), val = tensor([1, 256, 16, 32, 2])]; + tensor k_complex_43_cast_fp16 = reshape(shape = var_8917, x = k_85_cast_fp16)[name = tensor("k_complex_43_cast_fp16")]; + tensor var_8921_begin_0 = const()[name = tensor("op_8921_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_8921_end_0 = const()[name = tensor("op_8921_end_0"), val = tensor([1, 256, 16, 32, 1])]; + tensor var_8921_end_mask_0 = const()[name = tensor("op_8921_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_8921_squeeze_mask_0 = const()[name = tensor("op_8921_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_8921_cast_fp16 = slice_by_index(begin = var_8921_begin_0, end = var_8921_end_0, end_mask = var_8921_end_mask_0, squeeze_mask = var_8921_squeeze_mask_0, x = q_complex_43_cast_fp16)[name = tensor("op_8921_cast_fp16")]; + tensor var_8929_begin_0 = const()[name = tensor("op_8929_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_8929_end_0 = const()[name = tensor("op_8929_end_0"), val = tensor([1, 256, 16, 32, 2])]; + tensor var_8929_end_mask_0 = const()[name = tensor("op_8929_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_8929_squeeze_mask_0 = const()[name = tensor("op_8929_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_8929_cast_fp16 = slice_by_index(begin = var_8929_begin_0, end = var_8929_end_0, end_mask = var_8929_end_mask_0, squeeze_mask = var_8929_squeeze_mask_0, x = q_complex_43_cast_fp16)[name = tensor("op_8929_cast_fp16")]; + tensor var_8937_begin_0 = const()[name = tensor("op_8937_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_8937_end_0 = const()[name = tensor("op_8937_end_0"), val = tensor([1, 256, 16, 32, 1])]; + tensor var_8937_end_mask_0 = const()[name = tensor("op_8937_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_8937_squeeze_mask_0 = const()[name = tensor("op_8937_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_8937_cast_fp16 = slice_by_index(begin = var_8937_begin_0, end = var_8937_end_0, end_mask = var_8937_end_mask_0, squeeze_mask = var_8937_squeeze_mask_0, x = k_complex_43_cast_fp16)[name = tensor("op_8937_cast_fp16")]; + tensor var_8945_begin_0 = const()[name = tensor("op_8945_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_8945_end_0 = const()[name = tensor("op_8945_end_0"), val = tensor([1, 256, 16, 32, 2])]; + tensor var_8945_end_mask_0 = const()[name = tensor("op_8945_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_8945_squeeze_mask_0 = const()[name = tensor("op_8945_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_8945_cast_fp16 = slice_by_index(begin = var_8945_begin_0, end = var_8945_end_0, end_mask = var_8945_end_mask_0, squeeze_mask = var_8945_squeeze_mask_0, x = k_complex_43_cast_fp16)[name = tensor("op_8945_cast_fp16")]; + tensor freqs_43_to_fp16 = const()[name = tensor("freqs_43_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(536031616)))]; + tensor var_8951_cast_fp16 = mul(x = freqs_43_to_fp16, y = ts_131_cast_fp16)[name = tensor("op_8951_cast_fp16")]; + tensor rotr_43_cast_fp16 = cos(x = var_8951_cast_fp16)[name = tensor("rotr_43_cast_fp16")]; + tensor roti_43_cast_fp16 = sin(x = var_8951_cast_fp16)[name = tensor("roti_43_cast_fp16")]; + tensor var_8955_cast_fp16 = mul(x = var_8921_cast_fp16, y = rotr_43_cast_fp16)[name = tensor("op_8955_cast_fp16")]; + tensor var_8956_cast_fp16 = mul(x = var_8929_cast_fp16, y = roti_43_cast_fp16)[name = tensor("op_8956_cast_fp16")]; + tensor qor_85_cast_fp16 = sub(x = var_8955_cast_fp16, y = var_8956_cast_fp16)[name = tensor("qor_85_cast_fp16")]; + tensor var_8959_cast_fp16 = mul(x = var_8921_cast_fp16, y = roti_43_cast_fp16)[name = tensor("op_8959_cast_fp16")]; + tensor var_8960_cast_fp16 = mul(x = var_8929_cast_fp16, y = rotr_43_cast_fp16)[name = tensor("op_8960_cast_fp16")]; + tensor qoi_85_cast_fp16 = add(x = var_8959_cast_fp16, y = var_8960_cast_fp16)[name = tensor("qoi_85_cast_fp16")]; + tensor var_8963_cast_fp16 = mul(x = var_8937_cast_fp16, y = rotr_43_cast_fp16)[name = tensor("op_8963_cast_fp16")]; + tensor var_8964_cast_fp16 = mul(x = var_8945_cast_fp16, y = roti_43_cast_fp16)[name = tensor("op_8964_cast_fp16")]; + tensor kor_85_cast_fp16 = sub(x = var_8963_cast_fp16, y = var_8964_cast_fp16)[name = tensor("kor_85_cast_fp16")]; + tensor var_8967_cast_fp16 = mul(x = var_8937_cast_fp16, y = roti_43_cast_fp16)[name = tensor("op_8967_cast_fp16")]; + tensor var_8968_cast_fp16 = mul(x = var_8945_cast_fp16, y = rotr_43_cast_fp16)[name = tensor("op_8968_cast_fp16")]; + tensor koi_85_cast_fp16 = add(x = var_8967_cast_fp16, y = var_8968_cast_fp16)[name = tensor("koi_85_cast_fp16")]; + tensor qo_43_axis_0 = const()[name = tensor("qo_43_axis_0"), val = tensor(-1)]; + tensor qo_43_cast_fp16 = stack(axis = qo_43_axis_0, values = (qor_85_cast_fp16, qoi_85_cast_fp16))[name = tensor("qo_43_cast_fp16")]; + tensor ko_43_axis_0 = const()[name = tensor("ko_43_axis_0"), val = tensor(-1)]; + tensor ko_43_cast_fp16 = stack(axis = ko_43_axis_0, values = (kor_85_cast_fp16, koi_85_cast_fp16))[name = tensor("ko_43_cast_fp16")]; + tensor var_9001 = const()[name = tensor("op_9001"), val = tensor([1, 256, 16, 64])]; + tensor q_129_cast_fp16 = reshape(shape = var_9001, x = qo_43_cast_fp16)[name = tensor("q_129_cast_fp16")]; + tensor var_9007 = const()[name = tensor("op_9007"), val = tensor([1, 256, 16, 64])]; + tensor k_87_cast_fp16 = reshape(shape = var_9007, x = ko_43_cast_fp16)[name = tensor("k_87_cast_fp16")]; + tensor write_indices_float_127_cast_fp16 = add(x = var_8896_cast_fp16, y = var_393_promoted_to_fp16)[name = tensor("write_indices_float_127_cast_fp16")]; + tensor _inversed_9028_y_0_to_fp16 = const()[name = tensor("_inversed_9028_y_0_to_fp16"), val = tensor(0x1p-9)]; + tensor _inversed_9028_cast_fp16 = mul(x = write_indices_float_127_cast_fp16, y = _inversed_9028_y_0_to_fp16)[name = tensor("_inversed_9028_cast_fp16")]; + tensor var_9029_cast_fp16 = floor(x = _inversed_9028_cast_fp16)[name = tensor("op_9029_cast_fp16")]; + tensor var_9030_to_fp16 = const()[name = tensor("op_9030_to_fp16"), val = tensor(0x1p+9)]; + tensor var_9031_cast_fp16 = mul(x = var_9029_cast_fp16, y = var_9030_to_fp16)[name = tensor("op_9031_cast_fp16")]; + tensor write_indices_float_129_cast_fp16 = sub(x = write_indices_float_127_cast_fp16, y = var_9031_cast_fp16)[name = tensor("write_indices_float_129_cast_fp16")]; + tensor var_9049_value_0_to_fp16 = const()[name = tensor("op_9049_value_0_to_fp16"), val = tensor(0x1.ffp+8)]; + tensor var_9049_cast_fp16 = fill_like(ref_tensor = write_indices_float_129_cast_fp16, value = var_9049_value_0_to_fp16)[name = tensor("op_9049_cast_fp16")]; + tensor write_indices_float_131_cast_fp16 = select(a = write_indices_float_129_cast_fp16, b = var_9049_cast_fp16, cond = token_valid_1_cast_fp16)[name = tensor("write_indices_float_131_cast_fp16")]; + tensor var_9055_dtype_0 = const()[name = tensor("op_9055_dtype_0"), val = tensor("int32")]; + tensor var_9060 = const()[name = tensor("op_9060"), val = tensor([1, 256, 1, 1])]; + tensor write_indices_float_131_cast_fp16_to_int32 = cast(dtype = var_9055_dtype_0, x = write_indices_float_131_cast_fp16)[name = tensor("cast_555")]; + tensor var_9061 = reshape(shape = var_9060, x = write_indices_float_131_cast_fp16_to_int32)[name = tensor("op_9061")]; + tensor write_indices_43_reps_0 = const()[name = tensor("write_indices_43_reps_0"), val = tensor([1, 1, 16, 64])]; + tensor write_indices_43 = tile(reps = write_indices_43_reps_0, x = var_9061)[name = tensor("write_indices_43")]; + tensor var_9071_begin_0 = const()[name = tensor("op_9071_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_9071_end_0 = const()[name = tensor("op_9071_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor var_9071_end_mask_0 = const()[name = tensor("op_9071_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_9071_squeeze_mask_0 = const()[name = tensor("op_9071_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor cache21_to_fp16_dtype_0 = const()[name = tensor("cache21_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor cache21_to_fp16 = cast(dtype = cache21_to_fp16_dtype_0, x = cache21)[name = tensor("cast_554")]; + tensor var_9071_cast_fp16 = slice_by_index(begin = var_9071_begin_0, end = var_9071_end_0, end_mask = var_9071_end_mask_0, squeeze_mask = var_9071_squeeze_mask_0, x = cache21_to_fp16)[name = tensor("op_9071_cast_fp16")]; + tensor var_9073_axis_0 = const()[name = tensor("op_9073_axis_0"), val = tensor(1)]; + tensor var_9073_mode_0 = const()[name = tensor("op_9073_mode_0"), val = tensor("update")]; + tensor var_9073_validate_indices_0 = const()[name = tensor("op_9073_validate_indices_0"), val = tensor(false)]; + tensor var_9073_cast_fp16 = scatter_along_axis(axis = var_9073_axis_0, data = var_9071_cast_fp16, indices = write_indices_43, mode = var_9073_mode_0, updates = k_87_cast_fp16, validate_indices = var_9073_validate_indices_0)[name = tensor("op_9073_cast_fp16")]; + tensor concat_170 = const()[name = tensor("concat_170"), val = tensor([0, 0, 0, 0, 0])]; + tensor concat_171 = const()[name = tensor("concat_171"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_43_internal_tensor_assign_1_stride_0 = const()[name = tensor("new_cache_43_internal_tensor_assign_1_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_43_internal_tensor_assign_1_begin_mask_0 = const()[name = tensor("new_cache_43_internal_tensor_assign_1_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_43_internal_tensor_assign_1_end_mask_0 = const()[name = tensor("new_cache_43_internal_tensor_assign_1_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_43_internal_tensor_assign_1_squeeze_mask_0 = const()[name = tensor("new_cache_43_internal_tensor_assign_1_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_88 = const()[name = tensor("shape_88"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_42 = const()[name = tensor("reduce_prod_42"), val = tensor(1048576)]; + tensor range_1d_42_start_0 = const()[name = tensor("range_1d_42_start_0"), val = tensor(0)]; + tensor range_1d_42_step_0 = const()[name = tensor("range_1d_42_step_0"), val = tensor(1)]; + tensor range_1d_42 = range_1d(end = reduce_prod_42, start = range_1d_42_start_0, step = range_1d_42_step_0)[name = tensor("range_1d_42")]; + tensor reshape_210 = reshape(shape = shape_88, x = range_1d_42)[name = tensor("reshape_210")]; + tensor slice_by_index_42 = slice_by_index(begin = concat_170, begin_mask = new_cache_43_internal_tensor_assign_1_begin_mask_0, end = concat_171, end_mask = new_cache_43_internal_tensor_assign_1_end_mask_0, squeeze_mask = new_cache_43_internal_tensor_assign_1_squeeze_mask_0, stride = new_cache_43_internal_tensor_assign_1_stride_0, x = reshape_210)[name = tensor("slice_by_index_42")]; + tensor reshape_211_shape_0 = const()[name = tensor("reshape_211_shape_0"), val = tensor([-1])]; + tensor reshape_211 = reshape(shape = reshape_211_shape_0, x = slice_by_index_42)[name = tensor("reshape_211")]; + tensor reshape_212_shape_0 = const()[name = tensor("reshape_212_shape_0"), val = tensor([-1])]; + tensor reshape_212_cast_fp16 = reshape(shape = reshape_212_shape_0, x = var_9073_cast_fp16)[name = tensor("reshape_212_cast_fp16")]; + tensor reshape_213_shape_0 = const()[name = tensor("reshape_213_shape_0"), val = tensor([-1])]; + tensor reshape_213_cast_fp16 = reshape(shape = reshape_213_shape_0, x = cache21_to_fp16)[name = tensor("reshape_213_cast_fp16")]; + tensor scatter_42_mode_0 = const()[name = tensor("scatter_42_mode_0"), val = tensor("update")]; + tensor scatter_42_axis_0 = const()[name = tensor("scatter_42_axis_0"), val = tensor(0)]; + tensor scatter_42_validate_indices_0 = const()[name = tensor("scatter_42_validate_indices_0"), val = tensor(false)]; + tensor scatter_42_cast_fp16 = scatter(axis = scatter_42_axis_0, data = reshape_213_cast_fp16, indices = reshape_211, mode = scatter_42_mode_0, updates = reshape_212_cast_fp16, validate_indices = scatter_42_validate_indices_0)[name = tensor("scatter_42_cast_fp16")]; + tensor reshape_214_cast_fp16 = reshape(shape = shape_88, x = scatter_42_cast_fp16)[name = tensor("reshape_214_cast_fp16")]; + tensor var_9081_begin_0 = const()[name = tensor("op_9081_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor var_9081_end_0 = const()[name = tensor("op_9081_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor var_9081_end_mask_0 = const()[name = tensor("op_9081_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_9081_squeeze_mask_0 = const()[name = tensor("op_9081_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor var_9081_cast_fp16 = slice_by_index(begin = var_9081_begin_0, end = var_9081_end_0, end_mask = var_9081_end_mask_0, squeeze_mask = var_9081_squeeze_mask_0, x = reshape_214_cast_fp16)[name = tensor("op_9081_cast_fp16")]; + tensor var_9083_axis_0 = const()[name = tensor("op_9083_axis_0"), val = tensor(1)]; + tensor var_9083_mode_0 = const()[name = tensor("op_9083_mode_0"), val = tensor("update")]; + tensor var_9083_validate_indices_0 = const()[name = tensor("op_9083_validate_indices_0"), val = tensor(false)]; + tensor var_9083_cast_fp16 = scatter_along_axis(axis = var_9083_axis_0, data = var_9081_cast_fp16, indices = write_indices_43, mode = var_9083_mode_0, updates = v_43_cast_fp16, validate_indices = var_9083_validate_indices_0)[name = tensor("op_9083_cast_fp16")]; + tensor concat_172 = const()[name = tensor("concat_172"), val = tensor([1, 0, 0, 0, 0])]; + tensor concat_173 = const()[name = tensor("concat_173"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_43_internal_tensor_assign_2_stride_0 = const()[name = tensor("new_cache_43_internal_tensor_assign_2_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_43_internal_tensor_assign_2_begin_mask_0 = const()[name = tensor("new_cache_43_internal_tensor_assign_2_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_43_internal_tensor_assign_2_end_mask_0 = const()[name = tensor("new_cache_43_internal_tensor_assign_2_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_43_internal_tensor_assign_2_squeeze_mask_0 = const()[name = tensor("new_cache_43_internal_tensor_assign_2_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_89 = const()[name = tensor("shape_89"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_43 = const()[name = tensor("reduce_prod_43"), val = tensor(1048576)]; + tensor range_1d_43_start_0 = const()[name = tensor("range_1d_43_start_0"), val = tensor(0)]; + tensor range_1d_43_step_0 = const()[name = tensor("range_1d_43_step_0"), val = tensor(1)]; + tensor range_1d_43 = range_1d(end = reduce_prod_43, start = range_1d_43_start_0, step = range_1d_43_step_0)[name = tensor("range_1d_43")]; + tensor reshape_215 = reshape(shape = shape_89, x = range_1d_43)[name = tensor("reshape_215")]; + tensor slice_by_index_43 = slice_by_index(begin = concat_172, begin_mask = new_cache_43_internal_tensor_assign_2_begin_mask_0, end = concat_173, end_mask = new_cache_43_internal_tensor_assign_2_end_mask_0, squeeze_mask = new_cache_43_internal_tensor_assign_2_squeeze_mask_0, stride = new_cache_43_internal_tensor_assign_2_stride_0, x = reshape_215)[name = tensor("slice_by_index_43")]; + tensor reshape_216_shape_0 = const()[name = tensor("reshape_216_shape_0"), val = tensor([-1])]; + tensor reshape_216 = reshape(shape = reshape_216_shape_0, x = slice_by_index_43)[name = tensor("reshape_216")]; + tensor reshape_217_shape_0 = const()[name = tensor("reshape_217_shape_0"), val = tensor([-1])]; + tensor reshape_217_cast_fp16 = reshape(shape = reshape_217_shape_0, x = var_9083_cast_fp16)[name = tensor("reshape_217_cast_fp16")]; + tensor reshape_218_shape_0 = const()[name = tensor("reshape_218_shape_0"), val = tensor([-1])]; + tensor reshape_218_cast_fp16 = reshape(shape = reshape_218_shape_0, x = reshape_214_cast_fp16)[name = tensor("reshape_218_cast_fp16")]; + tensor scatter_43_mode_0 = const()[name = tensor("scatter_43_mode_0"), val = tensor("update")]; + tensor scatter_43_axis_0 = const()[name = tensor("scatter_43_axis_0"), val = tensor(0)]; + tensor scatter_43_validate_indices_0 = const()[name = tensor("scatter_43_validate_indices_0"), val = tensor(false)]; + tensor scatter_43_cast_fp16 = scatter(axis = scatter_43_axis_0, data = reshape_218_cast_fp16, indices = reshape_216, mode = scatter_43_mode_0, updates = reshape_217_cast_fp16, validate_indices = scatter_43_validate_indices_0)[name = tensor("scatter_43_cast_fp16")]; + tensor reshape_219_cast_fp16 = reshape(shape = shape_89, x = scatter_43_cast_fp16)[name = tensor("reshape_219_cast_fp16")]; + tensor reshape_219_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("reshape_219_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor keys_127_begin_0 = const()[name = tensor("keys_127_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor keys_127_end_0 = const()[name = tensor("keys_127_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor keys_127_end_mask_0 = const()[name = tensor("keys_127_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor keys_127_squeeze_mask_0 = const()[name = tensor("keys_127_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor keys_127_cast_fp16 = slice_by_index(begin = keys_127_begin_0, end = keys_127_end_0, end_mask = keys_127_end_mask_0, squeeze_mask = keys_127_squeeze_mask_0, x = reshape_219_cast_fp16)[name = tensor("keys_127_cast_fp16")]; + tensor values_127_begin_0 = const()[name = tensor("values_127_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor values_127_end_0 = const()[name = tensor("values_127_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor values_127_end_mask_0 = const()[name = tensor("values_127_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor values_127_squeeze_mask_0 = const()[name = tensor("values_127_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor values_127_cast_fp16 = slice_by_index(begin = values_127_begin_0, end = values_127_end_0, end_mask = values_127_end_mask_0, squeeze_mask = values_127_squeeze_mask_0, x = reshape_219_cast_fp16)[name = tensor("values_127_cast_fp16")]; + tensor var_9095_cast_fp16 = not_equal(x = keys_127_cast_fp16, y = keys_127_cast_fp16)[name = tensor("op_9095_cast_fp16")]; + tensor keys_129_cast_fp16 = select(a = var_470_to_fp16, b = keys_127_cast_fp16, cond = var_9095_cast_fp16)[name = tensor("keys_129_cast_fp16")]; + tensor var_9103_cast_fp16 = not_equal(x = values_127_cast_fp16, y = values_127_cast_fp16)[name = tensor("op_9103_cast_fp16")]; + tensor values_129_cast_fp16 = select(a = var_470_to_fp16, b = values_127_cast_fp16, cond = var_9103_cast_fp16)[name = tensor("values_129_cast_fp16")]; + tensor var_9127 = const()[name = tensor("op_9127"), val = tensor([0, 2, 1, 3])]; + tensor var_9143 = const()[name = tensor("op_9143"), val = tensor([1, 1, 1])]; + tensor var_9144_cast_fp16 = reshape(shape = var_9143, x = position21_to_fp16)[name = tensor("op_9144_cast_fp16")]; + tensor q_positions_43_cast_fp16 = add(x = var_9144_cast_fp16, y = q_offsets_1_promoted_to_fp16)[name = tensor("q_positions_43_cast_fp16")]; + tensor valid_end_43_cast_fp16 = add(x = var_9144_cast_fp16, y = var_536_cast_fp16)[name = tensor("valid_end_43_cast_fp16")]; + tensor valid_mask_43_cast_fp16 = less(x = k_positions_1_promoted_to_fp16, y = valid_end_43_cast_fp16)[name = tensor("valid_mask_43_cast_fp16")]; + tensor causal_mask_43_cast_fp16 = less_equal(x = k_positions_1_promoted_to_fp16, y = q_positions_43_cast_fp16)[name = tensor("causal_mask_43_cast_fp16")]; + tensor var_9172 = logical_and(x = valid_mask_43_cast_fp16, y = causal_mask_43_cast_fp16)[name = tensor("op_9172")]; + tensor attn_mask_43_axes_0 = const()[name = tensor("attn_mask_43_axes_0"), val = tensor([1])]; + tensor attn_mask_43 = expand_dims(axes = attn_mask_43_axes_0, x = var_9172)[name = tensor("attn_mask_43")]; + tensor var_9186_transpose_x_0 = const()[name = tensor("op_9186_transpose_x_0"), val = tensor(false)]; + tensor var_9186_transpose_y_0 = const()[name = tensor("op_9186_transpose_y_0"), val = tensor(false)]; + tensor transpose_111_perm_0 = const()[name = tensor("transpose_111_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_112_perm_0 = const()[name = tensor("transpose_112_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_112 = transpose(perm = transpose_112_perm_0, x = keys_129_cast_fp16)[name = tensor("transpose_120")]; + tensor transpose_111 = transpose(perm = transpose_111_perm_0, x = q_129_cast_fp16)[name = tensor("transpose_121")]; + tensor var_9186_cast_fp16 = matmul(transpose_x = var_9186_transpose_x_0, transpose_y = var_9186_transpose_y_0, x = transpose_111, y = transpose_112)[name = tensor("op_9186_cast_fp16")]; + tensor var_9180_to_fp16 = const()[name = tensor("op_9180_to_fp16"), val = tensor([0x1p-3])]; + tensor attn_weights_127_cast_fp16 = mul(x = var_9186_cast_fp16, y = var_9180_to_fp16)[name = tensor("attn_weights_127_cast_fp16")]; + tensor var_9188 = logical_not(x = attn_mask_43)[name = tensor("op_9188")]; + tensor var_9189_to_fp16 = const()[name = tensor("op_9189_to_fp16"), val = tensor(-inf)]; + tensor attn_weights_129_cast_fp16 = select(a = var_9189_to_fp16, b = attn_weights_127_cast_fp16, cond = var_9188)[name = tensor("attn_weights_129_cast_fp16")]; + tensor var_9191 = const()[name = tensor("op_9191"), val = tensor(-1)]; + tensor attn_weights_131_cast_fp16 = softmax(axis = var_9191, x = attn_weights_129_cast_fp16)[name = tensor("attn_weights_131_cast_fp16")]; + tensor attn_output_43_transpose_x_0 = const()[name = tensor("attn_output_43_transpose_x_0"), val = tensor(false)]; + tensor attn_output_43_transpose_y_0 = const()[name = tensor("attn_output_43_transpose_y_0"), val = tensor(false)]; + tensor values_131_cast_fp16 = transpose(perm = var_9127, x = values_129_cast_fp16)[name = tensor("transpose_122")]; + tensor attn_output_43_cast_fp16 = matmul(transpose_x = attn_output_43_transpose_x_0, transpose_y = attn_output_43_transpose_y_0, x = attn_weights_131_cast_fp16, y = values_131_cast_fp16)[name = tensor("attn_output_43_cast_fp16")]; + tensor var_9199 = const()[name = tensor("op_9199"), val = tensor([0, 2, 1, 3])]; + tensor var_9204 = const()[name = tensor("op_9204"), val = tensor([1, 256, 1024])]; + tensor var_9200_cast_fp16 = transpose(perm = var_9199, x = attn_output_43_cast_fp16)[name = tensor("transpose_119")]; + tensor input_257_cast_fp16 = reshape(shape = var_9204, x = var_9200_cast_fp16)[name = tensor("input_257_cast_fp16")]; + tensor attn21_out_proj_weight_to_fp16 = const()[name = tensor("attn21_out_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(536031744)))]; + tensor linear_85_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = attn21_out_proj_weight_to_fp16, x = input_257_cast_fp16)[name = tensor("linear_85_cast_fp16")]; + tensor var_9213_cast_fp16 = add(x = position21_to_fp16, y = valid_len_to_fp16)[name = tensor("op_9213_cast_fp16")]; + tensor var_9213_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_9213_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor input_259_cast_fp16 = add(x = input_253_cast_fp16, y = linear_85_cast_fp16)[name = tensor("input_259_cast_fp16")]; + tensor input_261_axes_0 = const()[name = tensor("input_261_axes_0"), val = tensor([-1])]; + tensor norm21_2_weight_to_fp16 = const()[name = tensor("norm21_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(538128960)))]; + tensor norm21_2_bias_to_fp16 = const()[name = tensor("norm21_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(538131072)))]; + tensor var_9217_to_fp16 = const()[name = tensor("op_9217_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_261_cast_fp16 = layer_norm(axes = input_261_axes_0, beta = norm21_2_bias_to_fp16, epsilon = var_9217_to_fp16, gamma = norm21_2_weight_to_fp16, x = input_259_cast_fp16)[name = tensor("input_261_cast_fp16")]; + tensor linear21_1_weight_to_fp16 = const()[name = tensor("linear21_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(538133184)))]; + tensor linear_86_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = linear21_1_weight_to_fp16, x = input_261_cast_fp16)[name = tensor("linear_86_cast_fp16")]; + tensor input_263_mode_0 = const()[name = tensor("input_263_mode_0"), val = tensor("EXACT")]; + tensor input_263_cast_fp16 = gelu(mode = input_263_mode_0, x = linear_86_cast_fp16)[name = tensor("input_263_cast_fp16")]; + tensor linear21_2_weight_to_fp16 = const()[name = tensor("linear21_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(546521856)))]; + tensor linear_87_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = linear21_2_weight_to_fp16, x = input_263_cast_fp16)[name = tensor("linear_87_cast_fp16")]; + tensor input_265_cast_fp16 = add(x = input_259_cast_fp16, y = linear_87_cast_fp16)[name = tensor("input_265_cast_fp16")]; + tensor input_267_axes_0 = const()[name = tensor("input_267_axes_0"), val = tensor([-1])]; + tensor norm22_1_weight_to_fp16 = const()[name = tensor("norm22_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(554910528)))]; + tensor norm22_1_bias_to_fp16 = const()[name = tensor("norm22_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(554912640)))]; + tensor var_9234_to_fp16 = const()[name = tensor("op_9234_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_267_cast_fp16 = layer_norm(axes = input_267_axes_0, beta = norm22_1_bias_to_fp16, epsilon = var_9234_to_fp16, gamma = norm22_1_weight_to_fp16, x = input_265_cast_fp16)[name = tensor("input_267_cast_fp16")]; + tensor attn22_in_proj_weight_to_fp16 = const()[name = tensor("attn22_in_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(554914752)))]; + tensor linear_88_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = attn22_in_proj_weight_to_fp16, x = input_267_cast_fp16)[name = tensor("linear_88_cast_fp16")]; + tensor var_9248 = const()[name = tensor("op_9248"), val = tensor([1, 256, 3, 16, 64])]; + tensor qkv_45_cast_fp16 = reshape(shape = var_9248, x = linear_88_cast_fp16)[name = tensor("qkv_45_cast_fp16")]; + tensor q_133_begin_0 = const()[name = tensor("q_133_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor q_133_end_0 = const()[name = tensor("q_133_end_0"), val = tensor([1, 256, 1, 16, 64])]; + tensor q_133_end_mask_0 = const()[name = tensor("q_133_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor q_133_squeeze_mask_0 = const()[name = tensor("q_133_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor q_133_cast_fp16 = slice_by_index(begin = q_133_begin_0, end = q_133_end_0, end_mask = q_133_end_mask_0, squeeze_mask = q_133_squeeze_mask_0, x = qkv_45_cast_fp16)[name = tensor("q_133_cast_fp16")]; + tensor k_89_begin_0 = const()[name = tensor("k_89_begin_0"), val = tensor([0, 0, 1, 0, 0])]; + tensor k_89_end_0 = const()[name = tensor("k_89_end_0"), val = tensor([1, 256, 2, 16, 64])]; + tensor k_89_end_mask_0 = const()[name = tensor("k_89_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor k_89_squeeze_mask_0 = const()[name = tensor("k_89_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor k_89_cast_fp16 = slice_by_index(begin = k_89_begin_0, end = k_89_end_0, end_mask = k_89_end_mask_0, squeeze_mask = k_89_squeeze_mask_0, x = qkv_45_cast_fp16)[name = tensor("k_89_cast_fp16")]; + tensor v_45_begin_0 = const()[name = tensor("v_45_begin_0"), val = tensor([0, 0, 2, 0, 0])]; + tensor v_45_end_0 = const()[name = tensor("v_45_end_0"), val = tensor([1, 256, 3, 16, 64])]; + tensor v_45_end_mask_0 = const()[name = tensor("v_45_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor v_45_squeeze_mask_0 = const()[name = tensor("v_45_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor v_45_cast_fp16 = slice_by_index(begin = v_45_begin_0, end = v_45_end_0, end_mask = v_45_end_mask_0, squeeze_mask = v_45_squeeze_mask_0, x = qkv_45_cast_fp16)[name = tensor("v_45_cast_fp16")]; + tensor var_9306 = const()[name = tensor("op_9306"), val = tensor([1, 1])]; + tensor position22_to_fp16_dtype_0 = const()[name = tensor("position22_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor position22_to_fp16 = cast(dtype = position22_to_fp16_dtype_0, x = position22)[name = tensor("cast_551")]; + tensor var_9307_cast_fp16 = reshape(shape = var_9306, x = position22_to_fp16)[name = tensor("op_9307_cast_fp16")]; + tensor ts_135_cast_fp16 = add(x = ts_1_promoted_to_fp16, y = var_9307_cast_fp16)[name = tensor("ts_135_cast_fp16")]; + tensor var_9314 = const()[name = tensor("op_9314"), val = tensor([1, 256, 1, 1])]; + tensor ts_137_cast_fp16 = reshape(shape = var_9314, x = ts_135_cast_fp16)[name = tensor("ts_137_cast_fp16")]; + tensor var_9321 = const()[name = tensor("op_9321"), val = tensor([1, 256, 16, 32, 2])]; + tensor q_complex_45_cast_fp16 = reshape(shape = var_9321, x = q_133_cast_fp16)[name = tensor("q_complex_45_cast_fp16")]; + tensor var_9328 = const()[name = tensor("op_9328"), val = tensor([1, 256, 16, 32, 2])]; + tensor k_complex_45_cast_fp16 = reshape(shape = var_9328, x = k_89_cast_fp16)[name = tensor("k_complex_45_cast_fp16")]; + tensor var_9332_begin_0 = const()[name = tensor("op_9332_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_9332_end_0 = const()[name = tensor("op_9332_end_0"), val = tensor([1, 256, 16, 32, 1])]; + tensor var_9332_end_mask_0 = const()[name = tensor("op_9332_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_9332_squeeze_mask_0 = const()[name = tensor("op_9332_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_9332_cast_fp16 = slice_by_index(begin = var_9332_begin_0, end = var_9332_end_0, end_mask = var_9332_end_mask_0, squeeze_mask = var_9332_squeeze_mask_0, x = q_complex_45_cast_fp16)[name = tensor("op_9332_cast_fp16")]; + tensor var_9340_begin_0 = const()[name = tensor("op_9340_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_9340_end_0 = const()[name = tensor("op_9340_end_0"), val = tensor([1, 256, 16, 32, 2])]; + tensor var_9340_end_mask_0 = const()[name = tensor("op_9340_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_9340_squeeze_mask_0 = const()[name = tensor("op_9340_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_9340_cast_fp16 = slice_by_index(begin = var_9340_begin_0, end = var_9340_end_0, end_mask = var_9340_end_mask_0, squeeze_mask = var_9340_squeeze_mask_0, x = q_complex_45_cast_fp16)[name = tensor("op_9340_cast_fp16")]; + tensor var_9348_begin_0 = const()[name = tensor("op_9348_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_9348_end_0 = const()[name = tensor("op_9348_end_0"), val = tensor([1, 256, 16, 32, 1])]; + tensor var_9348_end_mask_0 = const()[name = tensor("op_9348_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_9348_squeeze_mask_0 = const()[name = tensor("op_9348_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_9348_cast_fp16 = slice_by_index(begin = var_9348_begin_0, end = var_9348_end_0, end_mask = var_9348_end_mask_0, squeeze_mask = var_9348_squeeze_mask_0, x = k_complex_45_cast_fp16)[name = tensor("op_9348_cast_fp16")]; + tensor var_9356_begin_0 = const()[name = tensor("op_9356_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_9356_end_0 = const()[name = tensor("op_9356_end_0"), val = tensor([1, 256, 16, 32, 2])]; + tensor var_9356_end_mask_0 = const()[name = tensor("op_9356_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_9356_squeeze_mask_0 = const()[name = tensor("op_9356_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_9356_cast_fp16 = slice_by_index(begin = var_9356_begin_0, end = var_9356_end_0, end_mask = var_9356_end_mask_0, squeeze_mask = var_9356_squeeze_mask_0, x = k_complex_45_cast_fp16)[name = tensor("op_9356_cast_fp16")]; + tensor freqs_45_to_fp16 = const()[name = tensor("freqs_45_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(561206272)))]; + tensor var_9362_cast_fp16 = mul(x = freqs_45_to_fp16, y = ts_137_cast_fp16)[name = tensor("op_9362_cast_fp16")]; + tensor rotr_45_cast_fp16 = cos(x = var_9362_cast_fp16)[name = tensor("rotr_45_cast_fp16")]; + tensor roti_45_cast_fp16 = sin(x = var_9362_cast_fp16)[name = tensor("roti_45_cast_fp16")]; + tensor var_9366_cast_fp16 = mul(x = var_9332_cast_fp16, y = rotr_45_cast_fp16)[name = tensor("op_9366_cast_fp16")]; + tensor var_9367_cast_fp16 = mul(x = var_9340_cast_fp16, y = roti_45_cast_fp16)[name = tensor("op_9367_cast_fp16")]; + tensor qor_89_cast_fp16 = sub(x = var_9366_cast_fp16, y = var_9367_cast_fp16)[name = tensor("qor_89_cast_fp16")]; + tensor var_9370_cast_fp16 = mul(x = var_9332_cast_fp16, y = roti_45_cast_fp16)[name = tensor("op_9370_cast_fp16")]; + tensor var_9371_cast_fp16 = mul(x = var_9340_cast_fp16, y = rotr_45_cast_fp16)[name = tensor("op_9371_cast_fp16")]; + tensor qoi_89_cast_fp16 = add(x = var_9370_cast_fp16, y = var_9371_cast_fp16)[name = tensor("qoi_89_cast_fp16")]; + tensor var_9374_cast_fp16 = mul(x = var_9348_cast_fp16, y = rotr_45_cast_fp16)[name = tensor("op_9374_cast_fp16")]; + tensor var_9375_cast_fp16 = mul(x = var_9356_cast_fp16, y = roti_45_cast_fp16)[name = tensor("op_9375_cast_fp16")]; + tensor kor_89_cast_fp16 = sub(x = var_9374_cast_fp16, y = var_9375_cast_fp16)[name = tensor("kor_89_cast_fp16")]; + tensor var_9378_cast_fp16 = mul(x = var_9348_cast_fp16, y = roti_45_cast_fp16)[name = tensor("op_9378_cast_fp16")]; + tensor var_9379_cast_fp16 = mul(x = var_9356_cast_fp16, y = rotr_45_cast_fp16)[name = tensor("op_9379_cast_fp16")]; + tensor koi_89_cast_fp16 = add(x = var_9378_cast_fp16, y = var_9379_cast_fp16)[name = tensor("koi_89_cast_fp16")]; + tensor qo_45_axis_0 = const()[name = tensor("qo_45_axis_0"), val = tensor(-1)]; + tensor qo_45_cast_fp16 = stack(axis = qo_45_axis_0, values = (qor_89_cast_fp16, qoi_89_cast_fp16))[name = tensor("qo_45_cast_fp16")]; + tensor ko_45_axis_0 = const()[name = tensor("ko_45_axis_0"), val = tensor(-1)]; + tensor ko_45_cast_fp16 = stack(axis = ko_45_axis_0, values = (kor_89_cast_fp16, koi_89_cast_fp16))[name = tensor("ko_45_cast_fp16")]; + tensor var_9412 = const()[name = tensor("op_9412"), val = tensor([1, 256, 16, 64])]; + tensor q_135_cast_fp16 = reshape(shape = var_9412, x = qo_45_cast_fp16)[name = tensor("q_135_cast_fp16")]; + tensor var_9418 = const()[name = tensor("op_9418"), val = tensor([1, 256, 16, 64])]; + tensor k_91_cast_fp16 = reshape(shape = var_9418, x = ko_45_cast_fp16)[name = tensor("k_91_cast_fp16")]; + tensor write_indices_float_133_cast_fp16 = add(x = var_9307_cast_fp16, y = var_393_promoted_to_fp16)[name = tensor("write_indices_float_133_cast_fp16")]; + tensor _inversed_9439_y_0_to_fp16 = const()[name = tensor("_inversed_9439_y_0_to_fp16"), val = tensor(0x1p-9)]; + tensor _inversed_9439_cast_fp16 = mul(x = write_indices_float_133_cast_fp16, y = _inversed_9439_y_0_to_fp16)[name = tensor("_inversed_9439_cast_fp16")]; + tensor var_9440_cast_fp16 = floor(x = _inversed_9439_cast_fp16)[name = tensor("op_9440_cast_fp16")]; + tensor var_9441_to_fp16 = const()[name = tensor("op_9441_to_fp16"), val = tensor(0x1p+9)]; + tensor var_9442_cast_fp16 = mul(x = var_9440_cast_fp16, y = var_9441_to_fp16)[name = tensor("op_9442_cast_fp16")]; + tensor write_indices_float_135_cast_fp16 = sub(x = write_indices_float_133_cast_fp16, y = var_9442_cast_fp16)[name = tensor("write_indices_float_135_cast_fp16")]; + tensor var_9460_value_0_to_fp16 = const()[name = tensor("op_9460_value_0_to_fp16"), val = tensor(0x1.ffp+8)]; + tensor var_9460_cast_fp16 = fill_like(ref_tensor = write_indices_float_135_cast_fp16, value = var_9460_value_0_to_fp16)[name = tensor("op_9460_cast_fp16")]; + tensor write_indices_float_137_cast_fp16 = select(a = write_indices_float_135_cast_fp16, b = var_9460_cast_fp16, cond = token_valid_1_cast_fp16)[name = tensor("write_indices_float_137_cast_fp16")]; + tensor var_9466_dtype_0 = const()[name = tensor("op_9466_dtype_0"), val = tensor("int32")]; + tensor var_9471 = const()[name = tensor("op_9471"), val = tensor([1, 256, 1, 1])]; + tensor write_indices_float_137_cast_fp16_to_int32 = cast(dtype = var_9466_dtype_0, x = write_indices_float_137_cast_fp16)[name = tensor("cast_550")]; + tensor var_9472 = reshape(shape = var_9471, x = write_indices_float_137_cast_fp16_to_int32)[name = tensor("op_9472")]; + tensor write_indices_45_reps_0 = const()[name = tensor("write_indices_45_reps_0"), val = tensor([1, 1, 16, 64])]; + tensor write_indices_45 = tile(reps = write_indices_45_reps_0, x = var_9472)[name = tensor("write_indices_45")]; + tensor var_9482_begin_0 = const()[name = tensor("op_9482_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_9482_end_0 = const()[name = tensor("op_9482_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor var_9482_end_mask_0 = const()[name = tensor("op_9482_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_9482_squeeze_mask_0 = const()[name = tensor("op_9482_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor cache22_to_fp16_dtype_0 = const()[name = tensor("cache22_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor cache22_to_fp16 = cast(dtype = cache22_to_fp16_dtype_0, x = cache22)[name = tensor("cast_549")]; + tensor var_9482_cast_fp16 = slice_by_index(begin = var_9482_begin_0, end = var_9482_end_0, end_mask = var_9482_end_mask_0, squeeze_mask = var_9482_squeeze_mask_0, x = cache22_to_fp16)[name = tensor("op_9482_cast_fp16")]; + tensor var_9484_axis_0 = const()[name = tensor("op_9484_axis_0"), val = tensor(1)]; + tensor var_9484_mode_0 = const()[name = tensor("op_9484_mode_0"), val = tensor("update")]; + tensor var_9484_validate_indices_0 = const()[name = tensor("op_9484_validate_indices_0"), val = tensor(false)]; + tensor var_9484_cast_fp16 = scatter_along_axis(axis = var_9484_axis_0, data = var_9482_cast_fp16, indices = write_indices_45, mode = var_9484_mode_0, updates = k_91_cast_fp16, validate_indices = var_9484_validate_indices_0)[name = tensor("op_9484_cast_fp16")]; + tensor concat_178 = const()[name = tensor("concat_178"), val = tensor([0, 0, 0, 0, 0])]; + tensor concat_179 = const()[name = tensor("concat_179"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_45_internal_tensor_assign_1_stride_0 = const()[name = tensor("new_cache_45_internal_tensor_assign_1_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_45_internal_tensor_assign_1_begin_mask_0 = const()[name = tensor("new_cache_45_internal_tensor_assign_1_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_45_internal_tensor_assign_1_end_mask_0 = const()[name = tensor("new_cache_45_internal_tensor_assign_1_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_45_internal_tensor_assign_1_squeeze_mask_0 = const()[name = tensor("new_cache_45_internal_tensor_assign_1_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_90 = const()[name = tensor("shape_90"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_44 = const()[name = tensor("reduce_prod_44"), val = tensor(1048576)]; + tensor range_1d_44_start_0 = const()[name = tensor("range_1d_44_start_0"), val = tensor(0)]; + tensor range_1d_44_step_0 = const()[name = tensor("range_1d_44_step_0"), val = tensor(1)]; + tensor range_1d_44 = range_1d(end = reduce_prod_44, start = range_1d_44_start_0, step = range_1d_44_step_0)[name = tensor("range_1d_44")]; + tensor reshape_220 = reshape(shape = shape_90, x = range_1d_44)[name = tensor("reshape_220")]; + tensor slice_by_index_44 = slice_by_index(begin = concat_178, begin_mask = new_cache_45_internal_tensor_assign_1_begin_mask_0, end = concat_179, end_mask = new_cache_45_internal_tensor_assign_1_end_mask_0, squeeze_mask = new_cache_45_internal_tensor_assign_1_squeeze_mask_0, stride = new_cache_45_internal_tensor_assign_1_stride_0, x = reshape_220)[name = tensor("slice_by_index_44")]; + tensor reshape_221_shape_0 = const()[name = tensor("reshape_221_shape_0"), val = tensor([-1])]; + tensor reshape_221 = reshape(shape = reshape_221_shape_0, x = slice_by_index_44)[name = tensor("reshape_221")]; + tensor reshape_222_shape_0 = const()[name = tensor("reshape_222_shape_0"), val = tensor([-1])]; + tensor reshape_222_cast_fp16 = reshape(shape = reshape_222_shape_0, x = var_9484_cast_fp16)[name = tensor("reshape_222_cast_fp16")]; + tensor reshape_223_shape_0 = const()[name = tensor("reshape_223_shape_0"), val = tensor([-1])]; + tensor reshape_223_cast_fp16 = reshape(shape = reshape_223_shape_0, x = cache22_to_fp16)[name = tensor("reshape_223_cast_fp16")]; + tensor scatter_44_mode_0 = const()[name = tensor("scatter_44_mode_0"), val = tensor("update")]; + tensor scatter_44_axis_0 = const()[name = tensor("scatter_44_axis_0"), val = tensor(0)]; + tensor scatter_44_validate_indices_0 = const()[name = tensor("scatter_44_validate_indices_0"), val = tensor(false)]; + tensor scatter_44_cast_fp16 = scatter(axis = scatter_44_axis_0, data = reshape_223_cast_fp16, indices = reshape_221, mode = scatter_44_mode_0, updates = reshape_222_cast_fp16, validate_indices = scatter_44_validate_indices_0)[name = tensor("scatter_44_cast_fp16")]; + tensor reshape_224_cast_fp16 = reshape(shape = shape_90, x = scatter_44_cast_fp16)[name = tensor("reshape_224_cast_fp16")]; + tensor var_9492_begin_0 = const()[name = tensor("op_9492_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor var_9492_end_0 = const()[name = tensor("op_9492_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor var_9492_end_mask_0 = const()[name = tensor("op_9492_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_9492_squeeze_mask_0 = const()[name = tensor("op_9492_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor var_9492_cast_fp16 = slice_by_index(begin = var_9492_begin_0, end = var_9492_end_0, end_mask = var_9492_end_mask_0, squeeze_mask = var_9492_squeeze_mask_0, x = reshape_224_cast_fp16)[name = tensor("op_9492_cast_fp16")]; + tensor var_9494_axis_0 = const()[name = tensor("op_9494_axis_0"), val = tensor(1)]; + tensor var_9494_mode_0 = const()[name = tensor("op_9494_mode_0"), val = tensor("update")]; + tensor var_9494_validate_indices_0 = const()[name = tensor("op_9494_validate_indices_0"), val = tensor(false)]; + tensor var_9494_cast_fp16 = scatter_along_axis(axis = var_9494_axis_0, data = var_9492_cast_fp16, indices = write_indices_45, mode = var_9494_mode_0, updates = v_45_cast_fp16, validate_indices = var_9494_validate_indices_0)[name = tensor("op_9494_cast_fp16")]; + tensor concat_180 = const()[name = tensor("concat_180"), val = tensor([1, 0, 0, 0, 0])]; + tensor concat_181 = const()[name = tensor("concat_181"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_45_internal_tensor_assign_2_stride_0 = const()[name = tensor("new_cache_45_internal_tensor_assign_2_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_45_internal_tensor_assign_2_begin_mask_0 = const()[name = tensor("new_cache_45_internal_tensor_assign_2_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_45_internal_tensor_assign_2_end_mask_0 = const()[name = tensor("new_cache_45_internal_tensor_assign_2_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_45_internal_tensor_assign_2_squeeze_mask_0 = const()[name = tensor("new_cache_45_internal_tensor_assign_2_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_91 = const()[name = tensor("shape_91"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_45 = const()[name = tensor("reduce_prod_45"), val = tensor(1048576)]; + tensor range_1d_45_start_0 = const()[name = tensor("range_1d_45_start_0"), val = tensor(0)]; + tensor range_1d_45_step_0 = const()[name = tensor("range_1d_45_step_0"), val = tensor(1)]; + tensor range_1d_45 = range_1d(end = reduce_prod_45, start = range_1d_45_start_0, step = range_1d_45_step_0)[name = tensor("range_1d_45")]; + tensor reshape_225 = reshape(shape = shape_91, x = range_1d_45)[name = tensor("reshape_225")]; + tensor slice_by_index_45 = slice_by_index(begin = concat_180, begin_mask = new_cache_45_internal_tensor_assign_2_begin_mask_0, end = concat_181, end_mask = new_cache_45_internal_tensor_assign_2_end_mask_0, squeeze_mask = new_cache_45_internal_tensor_assign_2_squeeze_mask_0, stride = new_cache_45_internal_tensor_assign_2_stride_0, x = reshape_225)[name = tensor("slice_by_index_45")]; + tensor reshape_226_shape_0 = const()[name = tensor("reshape_226_shape_0"), val = tensor([-1])]; + tensor reshape_226 = reshape(shape = reshape_226_shape_0, x = slice_by_index_45)[name = tensor("reshape_226")]; + tensor reshape_227_shape_0 = const()[name = tensor("reshape_227_shape_0"), val = tensor([-1])]; + tensor reshape_227_cast_fp16 = reshape(shape = reshape_227_shape_0, x = var_9494_cast_fp16)[name = tensor("reshape_227_cast_fp16")]; + tensor reshape_228_shape_0 = const()[name = tensor("reshape_228_shape_0"), val = tensor([-1])]; + tensor reshape_228_cast_fp16 = reshape(shape = reshape_228_shape_0, x = reshape_224_cast_fp16)[name = tensor("reshape_228_cast_fp16")]; + tensor scatter_45_mode_0 = const()[name = tensor("scatter_45_mode_0"), val = tensor("update")]; + tensor scatter_45_axis_0 = const()[name = tensor("scatter_45_axis_0"), val = tensor(0)]; + tensor scatter_45_validate_indices_0 = const()[name = tensor("scatter_45_validate_indices_0"), val = tensor(false)]; + tensor scatter_45_cast_fp16 = scatter(axis = scatter_45_axis_0, data = reshape_228_cast_fp16, indices = reshape_226, mode = scatter_45_mode_0, updates = reshape_227_cast_fp16, validate_indices = scatter_45_validate_indices_0)[name = tensor("scatter_45_cast_fp16")]; + tensor reshape_229_cast_fp16 = reshape(shape = shape_91, x = scatter_45_cast_fp16)[name = tensor("reshape_229_cast_fp16")]; + tensor reshape_229_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("reshape_229_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor keys_133_begin_0 = const()[name = tensor("keys_133_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor keys_133_end_0 = const()[name = tensor("keys_133_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor keys_133_end_mask_0 = const()[name = tensor("keys_133_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor keys_133_squeeze_mask_0 = const()[name = tensor("keys_133_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor keys_133_cast_fp16 = slice_by_index(begin = keys_133_begin_0, end = keys_133_end_0, end_mask = keys_133_end_mask_0, squeeze_mask = keys_133_squeeze_mask_0, x = reshape_229_cast_fp16)[name = tensor("keys_133_cast_fp16")]; + tensor values_133_begin_0 = const()[name = tensor("values_133_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor values_133_end_0 = const()[name = tensor("values_133_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor values_133_end_mask_0 = const()[name = tensor("values_133_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor values_133_squeeze_mask_0 = const()[name = tensor("values_133_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor values_133_cast_fp16 = slice_by_index(begin = values_133_begin_0, end = values_133_end_0, end_mask = values_133_end_mask_0, squeeze_mask = values_133_squeeze_mask_0, x = reshape_229_cast_fp16)[name = tensor("values_133_cast_fp16")]; + tensor var_9506_cast_fp16 = not_equal(x = keys_133_cast_fp16, y = keys_133_cast_fp16)[name = tensor("op_9506_cast_fp16")]; + tensor keys_135_cast_fp16 = select(a = var_470_to_fp16, b = keys_133_cast_fp16, cond = var_9506_cast_fp16)[name = tensor("keys_135_cast_fp16")]; + tensor var_9514_cast_fp16 = not_equal(x = values_133_cast_fp16, y = values_133_cast_fp16)[name = tensor("op_9514_cast_fp16")]; + tensor values_135_cast_fp16 = select(a = var_470_to_fp16, b = values_133_cast_fp16, cond = var_9514_cast_fp16)[name = tensor("values_135_cast_fp16")]; + tensor var_9538 = const()[name = tensor("op_9538"), val = tensor([0, 2, 1, 3])]; + tensor var_9554 = const()[name = tensor("op_9554"), val = tensor([1, 1, 1])]; + tensor var_9555_cast_fp16 = reshape(shape = var_9554, x = position22_to_fp16)[name = tensor("op_9555_cast_fp16")]; + tensor q_positions_45_cast_fp16 = add(x = var_9555_cast_fp16, y = q_offsets_1_promoted_to_fp16)[name = tensor("q_positions_45_cast_fp16")]; + tensor valid_end_45_cast_fp16 = add(x = var_9555_cast_fp16, y = var_536_cast_fp16)[name = tensor("valid_end_45_cast_fp16")]; + tensor valid_mask_45_cast_fp16 = less(x = k_positions_1_promoted_to_fp16, y = valid_end_45_cast_fp16)[name = tensor("valid_mask_45_cast_fp16")]; + tensor causal_mask_45_cast_fp16 = less_equal(x = k_positions_1_promoted_to_fp16, y = q_positions_45_cast_fp16)[name = tensor("causal_mask_45_cast_fp16")]; + tensor var_9583 = logical_and(x = valid_mask_45_cast_fp16, y = causal_mask_45_cast_fp16)[name = tensor("op_9583")]; + tensor attn_mask_45_axes_0 = const()[name = tensor("attn_mask_45_axes_0"), val = tensor([1])]; + tensor attn_mask_45 = expand_dims(axes = attn_mask_45_axes_0, x = var_9583)[name = tensor("attn_mask_45")]; + tensor var_9597_transpose_x_0 = const()[name = tensor("op_9597_transpose_x_0"), val = tensor(false)]; + tensor var_9597_transpose_y_0 = const()[name = tensor("op_9597_transpose_y_0"), val = tensor(false)]; + tensor transpose_113_perm_0 = const()[name = tensor("transpose_113_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_114_perm_0 = const()[name = tensor("transpose_114_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_114 = transpose(perm = transpose_114_perm_0, x = keys_135_cast_fp16)[name = tensor("transpose_116")]; + tensor transpose_113 = transpose(perm = transpose_113_perm_0, x = q_135_cast_fp16)[name = tensor("transpose_117")]; + tensor var_9597_cast_fp16 = matmul(transpose_x = var_9597_transpose_x_0, transpose_y = var_9597_transpose_y_0, x = transpose_113, y = transpose_114)[name = tensor("op_9597_cast_fp16")]; + tensor var_9591_to_fp16 = const()[name = tensor("op_9591_to_fp16"), val = tensor([0x1p-3])]; + tensor attn_weights_133_cast_fp16 = mul(x = var_9597_cast_fp16, y = var_9591_to_fp16)[name = tensor("attn_weights_133_cast_fp16")]; + tensor var_9599 = logical_not(x = attn_mask_45)[name = tensor("op_9599")]; + tensor var_9600_to_fp16 = const()[name = tensor("op_9600_to_fp16"), val = tensor(-inf)]; + tensor attn_weights_135_cast_fp16 = select(a = var_9600_to_fp16, b = attn_weights_133_cast_fp16, cond = var_9599)[name = tensor("attn_weights_135_cast_fp16")]; + tensor var_9602 = const()[name = tensor("op_9602"), val = tensor(-1)]; + tensor attn_weights_137_cast_fp16 = softmax(axis = var_9602, x = attn_weights_135_cast_fp16)[name = tensor("attn_weights_137_cast_fp16")]; + tensor attn_output_45_transpose_x_0 = const()[name = tensor("attn_output_45_transpose_x_0"), val = tensor(false)]; + tensor attn_output_45_transpose_y_0 = const()[name = tensor("attn_output_45_transpose_y_0"), val = tensor(false)]; + tensor values_137_cast_fp16 = transpose(perm = var_9538, x = values_135_cast_fp16)[name = tensor("transpose_118")]; + tensor attn_output_45_cast_fp16 = matmul(transpose_x = attn_output_45_transpose_x_0, transpose_y = attn_output_45_transpose_y_0, x = attn_weights_137_cast_fp16, y = values_137_cast_fp16)[name = tensor("attn_output_45_cast_fp16")]; + tensor var_9610 = const()[name = tensor("op_9610"), val = tensor([0, 2, 1, 3])]; + tensor var_9615 = const()[name = tensor("op_9615"), val = tensor([1, 256, 1024])]; + tensor var_9611_cast_fp16 = transpose(perm = var_9610, x = attn_output_45_cast_fp16)[name = tensor("transpose_115")]; + tensor input_269_cast_fp16 = reshape(shape = var_9615, x = var_9611_cast_fp16)[name = tensor("input_269_cast_fp16")]; + tensor attn22_out_proj_weight_to_fp16 = const()[name = tensor("attn22_out_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(561206400)))]; + tensor linear_89_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = attn22_out_proj_weight_to_fp16, x = input_269_cast_fp16)[name = tensor("linear_89_cast_fp16")]; + tensor var_9624_cast_fp16 = add(x = position22_to_fp16, y = valid_len_to_fp16)[name = tensor("op_9624_cast_fp16")]; + tensor var_9624_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_9624_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor input_271_cast_fp16 = add(x = input_265_cast_fp16, y = linear_89_cast_fp16)[name = tensor("input_271_cast_fp16")]; + tensor input_273_axes_0 = const()[name = tensor("input_273_axes_0"), val = tensor([-1])]; + tensor norm22_2_weight_to_fp16 = const()[name = tensor("norm22_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(563303616)))]; + tensor norm22_2_bias_to_fp16 = const()[name = tensor("norm22_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(563305728)))]; + tensor var_9628_to_fp16 = const()[name = tensor("op_9628_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_273_cast_fp16 = layer_norm(axes = input_273_axes_0, beta = norm22_2_bias_to_fp16, epsilon = var_9628_to_fp16, gamma = norm22_2_weight_to_fp16, x = input_271_cast_fp16)[name = tensor("input_273_cast_fp16")]; + tensor linear22_1_weight_to_fp16 = const()[name = tensor("linear22_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(563307840)))]; + tensor linear_90_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = linear22_1_weight_to_fp16, x = input_273_cast_fp16)[name = tensor("linear_90_cast_fp16")]; + tensor input_275_mode_0 = const()[name = tensor("input_275_mode_0"), val = tensor("EXACT")]; + tensor input_275_cast_fp16 = gelu(mode = input_275_mode_0, x = linear_90_cast_fp16)[name = tensor("input_275_cast_fp16")]; + tensor linear22_2_weight_to_fp16 = const()[name = tensor("linear22_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(571696512)))]; + tensor linear_91_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = linear22_2_weight_to_fp16, x = input_275_cast_fp16)[name = tensor("linear_91_cast_fp16")]; + tensor input_277_cast_fp16 = add(x = input_271_cast_fp16, y = linear_91_cast_fp16)[name = tensor("input_277_cast_fp16")]; + tensor input_279_axes_0 = const()[name = tensor("input_279_axes_0"), val = tensor([-1])]; + tensor norm23_1_weight_to_fp16 = const()[name = tensor("norm23_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(580085184)))]; + tensor norm23_1_bias_to_fp16 = const()[name = tensor("norm23_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(580087296)))]; + tensor var_9645_to_fp16 = const()[name = tensor("op_9645_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_279_cast_fp16 = layer_norm(axes = input_279_axes_0, beta = norm23_1_bias_to_fp16, epsilon = var_9645_to_fp16, gamma = norm23_1_weight_to_fp16, x = input_277_cast_fp16)[name = tensor("input_279_cast_fp16")]; + tensor attn23_in_proj_weight_to_fp16 = const()[name = tensor("attn23_in_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(580089408)))]; + tensor linear_92_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = attn23_in_proj_weight_to_fp16, x = input_279_cast_fp16)[name = tensor("linear_92_cast_fp16")]; + tensor var_9659 = const()[name = tensor("op_9659"), val = tensor([1, 256, 3, 16, 64])]; + tensor qkv_cast_fp16 = reshape(shape = var_9659, x = linear_92_cast_fp16)[name = tensor("qkv_cast_fp16")]; + tensor k_93_begin_0 = const()[name = tensor("k_93_begin_0"), val = tensor([0, 0, 1, 0, 0])]; + tensor k_93_end_0 = const()[name = tensor("k_93_end_0"), val = tensor([1, 256, 2, 16, 64])]; + tensor k_93_end_mask_0 = const()[name = tensor("k_93_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor k_93_squeeze_mask_0 = const()[name = tensor("k_93_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor k_93_cast_fp16 = slice_by_index(begin = k_93_begin_0, end = k_93_end_0, end_mask = k_93_end_mask_0, squeeze_mask = k_93_squeeze_mask_0, x = qkv_cast_fp16)[name = tensor("k_93_cast_fp16")]; + tensor v_begin_0 = const()[name = tensor("v_begin_0"), val = tensor([0, 0, 2, 0, 0])]; + tensor v_end_0 = const()[name = tensor("v_end_0"), val = tensor([1, 256, 3, 16, 64])]; + tensor v_end_mask_0 = const()[name = tensor("v_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor v_squeeze_mask_0 = const()[name = tensor("v_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor v_cast_fp16 = slice_by_index(begin = v_begin_0, end = v_end_0, end_mask = v_end_mask_0, squeeze_mask = v_squeeze_mask_0, x = qkv_cast_fp16)[name = tensor("v_cast_fp16")]; + tensor var_9704 = const()[name = tensor("op_9704"), val = tensor([1, 1])]; + tensor position23_to_fp16_dtype_0 = const()[name = tensor("position23_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor position23_to_fp16 = cast(dtype = position23_to_fp16_dtype_0, x = position23)[name = tensor("cast_546")]; + tensor var_9705_cast_fp16 = reshape(shape = var_9704, x = position23_to_fp16)[name = tensor("op_9705_cast_fp16")]; + tensor ts_141_cast_fp16 = add(x = ts_1_promoted_to_fp16, y = var_9705_cast_fp16)[name = tensor("ts_141_cast_fp16")]; + tensor var_9712 = const()[name = tensor("op_9712"), val = tensor([1, 256, 1, 1])]; + tensor ts_cast_fp16 = reshape(shape = var_9712, x = ts_141_cast_fp16)[name = tensor("ts_cast_fp16")]; + tensor var_9719 = const()[name = tensor("op_9719"), val = tensor([1, 256, 16, 32, 2])]; + tensor k_complex_cast_fp16 = reshape(shape = var_9719, x = k_93_cast_fp16)[name = tensor("k_complex_cast_fp16")]; + tensor var_9723_begin_0 = const()[name = tensor("op_9723_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_9723_end_0 = const()[name = tensor("op_9723_end_0"), val = tensor([1, 256, 16, 32, 1])]; + tensor var_9723_end_mask_0 = const()[name = tensor("op_9723_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_9723_squeeze_mask_0 = const()[name = tensor("op_9723_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_9723_cast_fp16 = slice_by_index(begin = var_9723_begin_0, end = var_9723_end_0, end_mask = var_9723_end_mask_0, squeeze_mask = var_9723_squeeze_mask_0, x = k_complex_cast_fp16)[name = tensor("op_9723_cast_fp16")]; + tensor var_9731_begin_0 = const()[name = tensor("op_9731_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_9731_end_0 = const()[name = tensor("op_9731_end_0"), val = tensor([1, 256, 16, 32, 2])]; + tensor var_9731_end_mask_0 = const()[name = tensor("op_9731_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_9731_squeeze_mask_0 = const()[name = tensor("op_9731_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_9731_cast_fp16 = slice_by_index(begin = var_9731_begin_0, end = var_9731_end_0, end_mask = var_9731_end_mask_0, squeeze_mask = var_9731_squeeze_mask_0, x = k_complex_cast_fp16)[name = tensor("op_9731_cast_fp16")]; + tensor freqs_to_fp16 = const()[name = tensor("freqs_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(586380928)))]; + tensor var_9737_cast_fp16 = mul(x = freqs_to_fp16, y = ts_cast_fp16)[name = tensor("op_9737_cast_fp16")]; + tensor rotr_cast_fp16 = cos(x = var_9737_cast_fp16)[name = tensor("rotr_cast_fp16")]; + tensor roti_cast_fp16 = sin(x = var_9737_cast_fp16)[name = tensor("roti_cast_fp16")]; + tensor var_9741_cast_fp16 = mul(x = var_9723_cast_fp16, y = rotr_cast_fp16)[name = tensor("op_9741_cast_fp16")]; + tensor var_9742_cast_fp16 = mul(x = var_9731_cast_fp16, y = roti_cast_fp16)[name = tensor("op_9742_cast_fp16")]; + tensor kor_93_cast_fp16 = sub(x = var_9741_cast_fp16, y = var_9742_cast_fp16)[name = tensor("kor_93_cast_fp16")]; + tensor var_9745_cast_fp16 = mul(x = var_9723_cast_fp16, y = roti_cast_fp16)[name = tensor("op_9745_cast_fp16")]; + tensor var_9746_cast_fp16 = mul(x = var_9731_cast_fp16, y = rotr_cast_fp16)[name = tensor("op_9746_cast_fp16")]; + tensor koi_93_cast_fp16 = add(x = var_9745_cast_fp16, y = var_9746_cast_fp16)[name = tensor("koi_93_cast_fp16")]; + tensor ko_axis_0 = const()[name = tensor("ko_axis_0"), val = tensor(-1)]; + tensor ko_cast_fp16 = stack(axis = ko_axis_0, values = (kor_93_cast_fp16, koi_93_cast_fp16))[name = tensor("ko_cast_fp16")]; + tensor var_9766 = const()[name = tensor("op_9766"), val = tensor([1, 256, 16, 64])]; + tensor k_cast_fp16 = reshape(shape = var_9766, x = ko_cast_fp16)[name = tensor("k_cast_fp16")]; + tensor write_indices_float_139_cast_fp16 = add(x = var_9705_cast_fp16, y = var_393_promoted_to_fp16)[name = tensor("write_indices_float_139_cast_fp16")]; + tensor _inversed_9787_y_0_to_fp16 = const()[name = tensor("_inversed_9787_y_0_to_fp16"), val = tensor(0x1p-9)]; + tensor _inversed_9787_cast_fp16 = mul(x = write_indices_float_139_cast_fp16, y = _inversed_9787_y_0_to_fp16)[name = tensor("_inversed_9787_cast_fp16")]; + tensor var_9788_cast_fp16 = floor(x = _inversed_9787_cast_fp16)[name = tensor("op_9788_cast_fp16")]; + tensor var_9789_to_fp16 = const()[name = tensor("op_9789_to_fp16"), val = tensor(0x1p+9)]; + tensor var_9790_cast_fp16 = mul(x = var_9788_cast_fp16, y = var_9789_to_fp16)[name = tensor("op_9790_cast_fp16")]; + tensor write_indices_float_141_cast_fp16 = sub(x = write_indices_float_139_cast_fp16, y = var_9790_cast_fp16)[name = tensor("write_indices_float_141_cast_fp16")]; + tensor var_9808_value_0_to_fp16 = const()[name = tensor("op_9808_value_0_to_fp16"), val = tensor(0x1.ffp+8)]; + tensor var_9808_cast_fp16 = fill_like(ref_tensor = write_indices_float_141_cast_fp16, value = var_9808_value_0_to_fp16)[name = tensor("op_9808_cast_fp16")]; + tensor write_indices_float_cast_fp16 = select(a = write_indices_float_141_cast_fp16, b = var_9808_cast_fp16, cond = token_valid_1_cast_fp16)[name = tensor("write_indices_float_cast_fp16")]; + tensor var_9814_dtype_0 = const()[name = tensor("op_9814_dtype_0"), val = tensor("int32")]; + tensor var_9819 = const()[name = tensor("op_9819"), val = tensor([1, 256, 1, 1])]; + tensor write_indices_float_cast_fp16_to_int32 = cast(dtype = var_9814_dtype_0, x = write_indices_float_cast_fp16)[name = tensor("cast_545")]; + tensor var_9820 = reshape(shape = var_9819, x = write_indices_float_cast_fp16_to_int32)[name = tensor("op_9820")]; + tensor write_indices_reps_0 = const()[name = tensor("write_indices_reps_0"), val = tensor([1, 1, 16, 64])]; + tensor write_indices = tile(reps = write_indices_reps_0, x = var_9820)[name = tensor("write_indices")]; + tensor var_9830_begin_0 = const()[name = tensor("op_9830_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_9830_end_0 = const()[name = tensor("op_9830_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor var_9830_end_mask_0 = const()[name = tensor("op_9830_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_9830_squeeze_mask_0 = const()[name = tensor("op_9830_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor cache23_to_fp16_dtype_0 = const()[name = tensor("cache23_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor cache23_to_fp16 = cast(dtype = cache23_to_fp16_dtype_0, x = cache23)[name = tensor("cast_544")]; + tensor var_9830_cast_fp16 = slice_by_index(begin = var_9830_begin_0, end = var_9830_end_0, end_mask = var_9830_end_mask_0, squeeze_mask = var_9830_squeeze_mask_0, x = cache23_to_fp16)[name = tensor("op_9830_cast_fp16")]; + tensor var_9832_axis_0 = const()[name = tensor("op_9832_axis_0"), val = tensor(1)]; + tensor var_9832_mode_0 = const()[name = tensor("op_9832_mode_0"), val = tensor("update")]; + tensor var_9832_validate_indices_0 = const()[name = tensor("op_9832_validate_indices_0"), val = tensor(false)]; + tensor var_9832_cast_fp16 = scatter_along_axis(axis = var_9832_axis_0, data = var_9830_cast_fp16, indices = write_indices, mode = var_9832_mode_0, updates = k_cast_fp16, validate_indices = var_9832_validate_indices_0)[name = tensor("op_9832_cast_fp16")]; + tensor concat_186 = const()[name = tensor("concat_186"), val = tensor([0, 0, 0, 0, 0])]; + tensor concat_187 = const()[name = tensor("concat_187"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_internal_tensor_assign_1_stride_0 = const()[name = tensor("new_cache_internal_tensor_assign_1_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_internal_tensor_assign_1_begin_mask_0 = const()[name = tensor("new_cache_internal_tensor_assign_1_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_internal_tensor_assign_1_end_mask_0 = const()[name = tensor("new_cache_internal_tensor_assign_1_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_internal_tensor_assign_1_squeeze_mask_0 = const()[name = tensor("new_cache_internal_tensor_assign_1_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_92 = const()[name = tensor("shape_92"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_46 = const()[name = tensor("reduce_prod_46"), val = tensor(1048576)]; + tensor range_1d_46_start_0 = const()[name = tensor("range_1d_46_start_0"), val = tensor(0)]; + tensor range_1d_46_step_0 = const()[name = tensor("range_1d_46_step_0"), val = tensor(1)]; + tensor range_1d_46 = range_1d(end = reduce_prod_46, start = range_1d_46_start_0, step = range_1d_46_step_0)[name = tensor("range_1d_46")]; + tensor reshape_230 = reshape(shape = shape_92, x = range_1d_46)[name = tensor("reshape_230")]; + tensor slice_by_index_46 = slice_by_index(begin = concat_186, begin_mask = new_cache_internal_tensor_assign_1_begin_mask_0, end = concat_187, end_mask = new_cache_internal_tensor_assign_1_end_mask_0, squeeze_mask = new_cache_internal_tensor_assign_1_squeeze_mask_0, stride = new_cache_internal_tensor_assign_1_stride_0, x = reshape_230)[name = tensor("slice_by_index_46")]; + tensor reshape_231_shape_0 = const()[name = tensor("reshape_231_shape_0"), val = tensor([-1])]; + tensor reshape_231 = reshape(shape = reshape_231_shape_0, x = slice_by_index_46)[name = tensor("reshape_231")]; + tensor reshape_232_shape_0 = const()[name = tensor("reshape_232_shape_0"), val = tensor([-1])]; + tensor reshape_232_cast_fp16 = reshape(shape = reshape_232_shape_0, x = var_9832_cast_fp16)[name = tensor("reshape_232_cast_fp16")]; + tensor reshape_233_shape_0 = const()[name = tensor("reshape_233_shape_0"), val = tensor([-1])]; + tensor reshape_233_cast_fp16 = reshape(shape = reshape_233_shape_0, x = cache23_to_fp16)[name = tensor("reshape_233_cast_fp16")]; + tensor scatter_46_mode_0 = const()[name = tensor("scatter_46_mode_0"), val = tensor("update")]; + tensor scatter_46_axis_0 = const()[name = tensor("scatter_46_axis_0"), val = tensor(0)]; + tensor scatter_46_validate_indices_0 = const()[name = tensor("scatter_46_validate_indices_0"), val = tensor(false)]; + tensor scatter_46_cast_fp16 = scatter(axis = scatter_46_axis_0, data = reshape_233_cast_fp16, indices = reshape_231, mode = scatter_46_mode_0, updates = reshape_232_cast_fp16, validate_indices = scatter_46_validate_indices_0)[name = tensor("scatter_46_cast_fp16")]; + tensor reshape_234_cast_fp16 = reshape(shape = shape_92, x = scatter_46_cast_fp16)[name = tensor("reshape_234_cast_fp16")]; + tensor var_9840_begin_0 = const()[name = tensor("op_9840_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor var_9840_end_0 = const()[name = tensor("op_9840_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor var_9840_end_mask_0 = const()[name = tensor("op_9840_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_9840_squeeze_mask_0 = const()[name = tensor("op_9840_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor var_9840_cast_fp16 = slice_by_index(begin = var_9840_begin_0, end = var_9840_end_0, end_mask = var_9840_end_mask_0, squeeze_mask = var_9840_squeeze_mask_0, x = reshape_234_cast_fp16)[name = tensor("op_9840_cast_fp16")]; + tensor var_9842_axis_0 = const()[name = tensor("op_9842_axis_0"), val = tensor(1)]; + tensor var_9842_mode_0 = const()[name = tensor("op_9842_mode_0"), val = tensor("update")]; + tensor var_9842_validate_indices_0 = const()[name = tensor("op_9842_validate_indices_0"), val = tensor(false)]; + tensor var_9842_cast_fp16 = scatter_along_axis(axis = var_9842_axis_0, data = var_9840_cast_fp16, indices = write_indices, mode = var_9842_mode_0, updates = v_cast_fp16, validate_indices = var_9842_validate_indices_0)[name = tensor("op_9842_cast_fp16")]; + tensor concat_188 = const()[name = tensor("concat_188"), val = tensor([1, 0, 0, 0, 0])]; + tensor concat_189 = const()[name = tensor("concat_189"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_internal_tensor_assign_2_stride_0 = const()[name = tensor("new_cache_internal_tensor_assign_2_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_internal_tensor_assign_2_begin_mask_0 = const()[name = tensor("new_cache_internal_tensor_assign_2_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_internal_tensor_assign_2_end_mask_0 = const()[name = tensor("new_cache_internal_tensor_assign_2_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_internal_tensor_assign_2_squeeze_mask_0 = const()[name = tensor("new_cache_internal_tensor_assign_2_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_93 = const()[name = tensor("shape_93"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_47 = const()[name = tensor("reduce_prod_47"), val = tensor(1048576)]; + tensor range_1d_47_start_0 = const()[name = tensor("range_1d_47_start_0"), val = tensor(0)]; + tensor range_1d_47_step_0 = const()[name = tensor("range_1d_47_step_0"), val = tensor(1)]; + tensor range_1d_47 = range_1d(end = reduce_prod_47, start = range_1d_47_start_0, step = range_1d_47_step_0)[name = tensor("range_1d_47")]; + tensor reshape_235 = reshape(shape = shape_93, x = range_1d_47)[name = tensor("reshape_235")]; + tensor slice_by_index_47 = slice_by_index(begin = concat_188, begin_mask = new_cache_internal_tensor_assign_2_begin_mask_0, end = concat_189, end_mask = new_cache_internal_tensor_assign_2_end_mask_0, squeeze_mask = new_cache_internal_tensor_assign_2_squeeze_mask_0, stride = new_cache_internal_tensor_assign_2_stride_0, x = reshape_235)[name = tensor("slice_by_index_47")]; + tensor reshape_236_shape_0 = const()[name = tensor("reshape_236_shape_0"), val = tensor([-1])]; + tensor reshape_236 = reshape(shape = reshape_236_shape_0, x = slice_by_index_47)[name = tensor("reshape_236")]; + tensor reshape_237_shape_0 = const()[name = tensor("reshape_237_shape_0"), val = tensor([-1])]; + tensor reshape_237_cast_fp16 = reshape(shape = reshape_237_shape_0, x = var_9842_cast_fp16)[name = tensor("reshape_237_cast_fp16")]; + tensor reshape_238_shape_0 = const()[name = tensor("reshape_238_shape_0"), val = tensor([-1])]; + tensor reshape_238_cast_fp16 = reshape(shape = reshape_238_shape_0, x = reshape_234_cast_fp16)[name = tensor("reshape_238_cast_fp16")]; + tensor scatter_47_mode_0 = const()[name = tensor("scatter_47_mode_0"), val = tensor("update")]; + tensor scatter_47_axis_0 = const()[name = tensor("scatter_47_axis_0"), val = tensor(0)]; + tensor scatter_47_validate_indices_0 = const()[name = tensor("scatter_47_validate_indices_0"), val = tensor(false)]; + tensor scatter_47_cast_fp16 = scatter(axis = scatter_47_axis_0, data = reshape_238_cast_fp16, indices = reshape_236, mode = scatter_47_mode_0, updates = reshape_237_cast_fp16, validate_indices = scatter_47_validate_indices_0)[name = tensor("scatter_47_cast_fp16")]; + tensor reshape_239_cast_fp16 = reshape(shape = shape_93, x = scatter_47_cast_fp16)[name = tensor("reshape_239_cast_fp16")]; + tensor reshape_239_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("reshape_239_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor var_9852_cast_fp16 = add(x = position23_to_fp16, y = valid_len_to_fp16)[name = tensor("op_9852_cast_fp16")]; + tensor var_9852_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_9852_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor var_9852 = cast(dtype = var_9852_cast_fp16_to_fp32_dtype_0, x = var_9852_cast_fp16)[name = tensor("cast_542")]; + tensor new_cache_internal_tensor_assign_2 = cast(dtype = reshape_239_cast_fp16_to_fp32_dtype_0, x = reshape_239_cast_fp16)[name = tensor("cast_543")]; + tensor var_9624 = cast(dtype = var_9624_cast_fp16_to_fp32_dtype_0, x = var_9624_cast_fp16)[name = tensor("cast_547")]; + tensor new_cache_45_internal_tensor_assign_2 = cast(dtype = reshape_229_cast_fp16_to_fp32_dtype_0, x = reshape_229_cast_fp16)[name = tensor("cast_548")]; + tensor var_9213 = cast(dtype = var_9213_cast_fp16_to_fp32_dtype_0, x = var_9213_cast_fp16)[name = tensor("cast_552")]; + tensor new_cache_43_internal_tensor_assign_2 = cast(dtype = reshape_219_cast_fp16_to_fp32_dtype_0, x = reshape_219_cast_fp16)[name = tensor("cast_553")]; + tensor var_8802 = cast(dtype = var_8802_cast_fp16_to_fp32_dtype_0, x = var_8802_cast_fp16)[name = tensor("cast_557")]; + tensor new_cache_41_internal_tensor_assign_2 = cast(dtype = reshape_209_cast_fp16_to_fp32_dtype_0, x = reshape_209_cast_fp16)[name = tensor("cast_558")]; + tensor var_8391 = cast(dtype = var_8391_cast_fp16_to_fp32_dtype_0, x = var_8391_cast_fp16)[name = tensor("cast_562")]; + tensor new_cache_39_internal_tensor_assign_2 = cast(dtype = reshape_199_cast_fp16_to_fp32_dtype_0, x = reshape_199_cast_fp16)[name = tensor("cast_563")]; + tensor var_7980 = cast(dtype = var_7980_cast_fp16_to_fp32_dtype_0, x = var_7980_cast_fp16)[name = tensor("cast_567")]; + tensor new_cache_37_internal_tensor_assign_2 = cast(dtype = reshape_189_cast_fp16_to_fp32_dtype_0, x = reshape_189_cast_fp16)[name = tensor("cast_568")]; + tensor var_7569 = cast(dtype = var_7569_cast_fp16_to_fp32_dtype_0, x = var_7569_cast_fp16)[name = tensor("cast_572")]; + tensor new_cache_35_internal_tensor_assign_2 = cast(dtype = reshape_179_cast_fp16_to_fp32_dtype_0, x = reshape_179_cast_fp16)[name = tensor("cast_573")]; + tensor var_7158 = cast(dtype = var_7158_cast_fp16_to_fp32_dtype_0, x = var_7158_cast_fp16)[name = tensor("cast_577")]; + tensor new_cache_33_internal_tensor_assign_2 = cast(dtype = reshape_169_cast_fp16_to_fp32_dtype_0, x = reshape_169_cast_fp16)[name = tensor("cast_578")]; + tensor var_6747 = cast(dtype = var_6747_cast_fp16_to_fp32_dtype_0, x = var_6747_cast_fp16)[name = tensor("cast_582")]; + tensor new_cache_31_internal_tensor_assign_2 = cast(dtype = reshape_159_cast_fp16_to_fp32_dtype_0, x = reshape_159_cast_fp16)[name = tensor("cast_583")]; + tensor var_6336 = cast(dtype = var_6336_cast_fp16_to_fp32_dtype_0, x = var_6336_cast_fp16)[name = tensor("cast_587")]; + tensor new_cache_29_internal_tensor_assign_2 = cast(dtype = reshape_149_cast_fp16_to_fp32_dtype_0, x = reshape_149_cast_fp16)[name = tensor("cast_588")]; + tensor var_5925 = cast(dtype = var_5925_cast_fp16_to_fp32_dtype_0, x = var_5925_cast_fp16)[name = tensor("cast_592")]; + tensor new_cache_27_internal_tensor_assign_2 = cast(dtype = reshape_139_cast_fp16_to_fp32_dtype_0, x = reshape_139_cast_fp16)[name = tensor("cast_593")]; + tensor var_5514 = cast(dtype = var_5514_cast_fp16_to_fp32_dtype_0, x = var_5514_cast_fp16)[name = tensor("cast_597")]; + tensor new_cache_25_internal_tensor_assign_2 = cast(dtype = reshape_129_cast_fp16_to_fp32_dtype_0, x = reshape_129_cast_fp16)[name = tensor("cast_598")]; + tensor var_5103 = cast(dtype = var_5103_cast_fp16_to_fp32_dtype_0, x = var_5103_cast_fp16)[name = tensor("cast_602")]; + tensor new_cache_23_internal_tensor_assign_2 = cast(dtype = reshape_119_cast_fp16_to_fp32_dtype_0, x = reshape_119_cast_fp16)[name = tensor("cast_603")]; + tensor var_4692 = cast(dtype = var_4692_cast_fp16_to_fp32_dtype_0, x = var_4692_cast_fp16)[name = tensor("cast_607")]; + tensor new_cache_21_internal_tensor_assign_2 = cast(dtype = reshape_109_cast_fp16_to_fp32_dtype_0, x = reshape_109_cast_fp16)[name = tensor("cast_608")]; + tensor var_4281 = cast(dtype = var_4281_cast_fp16_to_fp32_dtype_0, x = var_4281_cast_fp16)[name = tensor("cast_612")]; + tensor new_cache_19_internal_tensor_assign_2 = cast(dtype = reshape_99_cast_fp16_to_fp32_dtype_0, x = reshape_99_cast_fp16)[name = tensor("cast_613")]; + tensor var_3870 = cast(dtype = var_3870_cast_fp16_to_fp32_dtype_0, x = var_3870_cast_fp16)[name = tensor("cast_617")]; + tensor new_cache_17_internal_tensor_assign_2 = cast(dtype = reshape_89_cast_fp16_to_fp32_dtype_0, x = reshape_89_cast_fp16)[name = tensor("cast_618")]; + tensor var_3459 = cast(dtype = var_3459_cast_fp16_to_fp32_dtype_0, x = var_3459_cast_fp16)[name = tensor("cast_622")]; + tensor new_cache_15_internal_tensor_assign_2 = cast(dtype = reshape_79_cast_fp16_to_fp32_dtype_0, x = reshape_79_cast_fp16)[name = tensor("cast_623")]; + tensor var_3048 = cast(dtype = var_3048_cast_fp16_to_fp32_dtype_0, x = var_3048_cast_fp16)[name = tensor("cast_627")]; + tensor new_cache_13_internal_tensor_assign_2 = cast(dtype = reshape_69_cast_fp16_to_fp32_dtype_0, x = reshape_69_cast_fp16)[name = tensor("cast_628")]; + tensor var_2637 = cast(dtype = var_2637_cast_fp16_to_fp32_dtype_0, x = var_2637_cast_fp16)[name = tensor("cast_632")]; + tensor new_cache_11_internal_tensor_assign_2 = cast(dtype = reshape_59_cast_fp16_to_fp32_dtype_0, x = reshape_59_cast_fp16)[name = tensor("cast_633")]; + tensor var_2226 = cast(dtype = var_2226_cast_fp16_to_fp32_dtype_0, x = var_2226_cast_fp16)[name = tensor("cast_637")]; + tensor new_cache_9_internal_tensor_assign_2 = cast(dtype = reshape_49_cast_fp16_to_fp32_dtype_0, x = reshape_49_cast_fp16)[name = tensor("cast_638")]; + tensor var_1815 = cast(dtype = var_1815_cast_fp16_to_fp32_dtype_0, x = var_1815_cast_fp16)[name = tensor("cast_642")]; + tensor new_cache_7_internal_tensor_assign_2 = cast(dtype = reshape_39_cast_fp16_to_fp32_dtype_0, x = reshape_39_cast_fp16)[name = tensor("cast_643")]; + tensor var_1404 = cast(dtype = var_1404_cast_fp16_to_fp32_dtype_0, x = var_1404_cast_fp16)[name = tensor("cast_647")]; + tensor new_cache_5_internal_tensor_assign_2 = cast(dtype = reshape_29_cast_fp16_to_fp32_dtype_0, x = reshape_29_cast_fp16)[name = tensor("cast_648")]; + tensor var_993 = cast(dtype = var_993_cast_fp16_to_fp32_dtype_0, x = var_993_cast_fp16)[name = tensor("cast_652")]; + tensor new_cache_3_internal_tensor_assign_2 = cast(dtype = reshape_19_cast_fp16_to_fp32_dtype_0, x = reshape_19_cast_fp16)[name = tensor("cast_653")]; + tensor var_582 = cast(dtype = var_582_cast_fp16_to_fp32_dtype_0, x = var_582_cast_fp16)[name = tensor("cast_657")]; + tensor new_cache_1_internal_tensor_assign_2 = cast(dtype = reshape_9_cast_fp16_to_fp32_dtype_0, x = reshape_9_cast_fp16)[name = tensor("cast_658")]; + } -> (new_cache_1_internal_tensor_assign_2, var_582, new_cache_3_internal_tensor_assign_2, var_993, new_cache_5_internal_tensor_assign_2, var_1404, new_cache_7_internal_tensor_assign_2, var_1815, new_cache_9_internal_tensor_assign_2, var_2226, new_cache_11_internal_tensor_assign_2, var_2637, new_cache_13_internal_tensor_assign_2, var_3048, new_cache_15_internal_tensor_assign_2, var_3459, new_cache_17_internal_tensor_assign_2, var_3870, new_cache_19_internal_tensor_assign_2, var_4281, new_cache_21_internal_tensor_assign_2, var_4692, new_cache_23_internal_tensor_assign_2, var_5103, new_cache_25_internal_tensor_assign_2, var_5514, new_cache_27_internal_tensor_assign_2, var_5925, new_cache_29_internal_tensor_assign_2, var_6336, new_cache_31_internal_tensor_assign_2, var_6747, new_cache_33_internal_tensor_assign_2, var_7158, new_cache_35_internal_tensor_assign_2, var_7569, new_cache_37_internal_tensor_assign_2, var_7980, new_cache_39_internal_tensor_assign_2, var_8391, new_cache_41_internal_tensor_assign_2, var_8802, new_cache_43_internal_tensor_assign_2, var_9213, new_cache_45_internal_tensor_assign_2, var_9624, new_cache_internal_tensor_assign_2, var_9852); +} \ No newline at end of file diff --git a/v2.1/german_24l/cond_prefill.mlmodelc/weights/weight.bin b/v2.1/german_24l/cond_prefill.mlmodelc/weights/weight.bin new file mode 100644 index 0000000000000000000000000000000000000000..8e9e217a411e1cf6fbf540b021d4967a58f214b6 --- /dev/null +++ b/v2.1/german_24l/cond_prefill.mlmodelc/weights/weight.bin @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:9025814420033df3b8205d51b69e147f8ce30dffbd95ae0d7b9c4d917b057fd4 +size 586381056 diff --git a/v2.1/german_24l/cond_prefill.mlpackage/Data/com.apple.CoreML/model.mlmodel b/v2.1/german_24l/cond_prefill.mlpackage/Data/com.apple.CoreML/model.mlmodel new file mode 100644 index 0000000000000000000000000000000000000000..883715913c70c5603236e2fab0c69586590b26d8 --- /dev/null +++ b/v2.1/german_24l/cond_prefill.mlpackage/Data/com.apple.CoreML/model.mlmodel @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:762966e4782a10f92961128bfcde3cd7874859d181cbb8df614685e8d508ab5d +size 842494 diff --git a/v2.1/german_24l/cond_prefill.mlpackage/Data/com.apple.CoreML/weights/weight.bin b/v2.1/german_24l/cond_prefill.mlpackage/Data/com.apple.CoreML/weights/weight.bin new file mode 100644 index 0000000000000000000000000000000000000000..8e9e217a411e1cf6fbf540b021d4967a58f214b6 --- /dev/null +++ b/v2.1/german_24l/cond_prefill.mlpackage/Data/com.apple.CoreML/weights/weight.bin @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:9025814420033df3b8205d51b69e147f8ce30dffbd95ae0d7b9c4d917b057fd4 +size 586381056 diff --git a/v2.1/german_24l/cond_prefill.mlpackage/Manifest.json b/v2.1/german_24l/cond_prefill.mlpackage/Manifest.json new file mode 100644 index 0000000000000000000000000000000000000000..0d131a0c0a0f017296ff2d6552866be8dceea2af --- /dev/null +++ b/v2.1/german_24l/cond_prefill.mlpackage/Manifest.json @@ -0,0 +1,18 @@ +{ + "fileFormatVersion": "1.0.0", + "itemInfoEntries": { + "14FC4189-7E6C-4426-B289-52C8536BB020": { + "author": "com.apple.CoreML", + "description": "CoreML Model Specification", + "name": "model.mlmodel", + "path": "com.apple.CoreML/model.mlmodel" + }, + "1FE2E759-A95C-4241-B9B3-B875916B364E": { + "author": "com.apple.CoreML", + "description": "CoreML Model Weights", + "name": "weights", + "path": "com.apple.CoreML/weights" + } + }, + "rootModelIdentifier": "14FC4189-7E6C-4426-B289-52C8536BB020" +} diff --git a/v2.1/german_24l/constants/bos_emb.npy b/v2.1/german_24l/constants/bos_emb.npy new file mode 100644 index 0000000000000000000000000000000000000000..fedc81f58e80f1639ccc197433ccefd26692755e --- /dev/null +++ b/v2.1/german_24l/constants/bos_emb.npy @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:183bf02b93710b760b46cb447e46a1de38cfbc8205f8fd2aecca4920d0d2c96d +size 256 diff --git a/v2.1/german_24l/constants/emb_mean.npy b/v2.1/german_24l/constants/emb_mean.npy new file mode 100644 index 0000000000000000000000000000000000000000..0466a09c1264dc319a854a8ccc370ca251302e20 --- /dev/null +++ b/v2.1/german_24l/constants/emb_mean.npy @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:00204e3732d8ade46adeba6f24922470c703638da7dc3fb4788f95ea012873a0 +size 256 diff --git a/v2.1/german_24l/constants/emb_std.npy b/v2.1/german_24l/constants/emb_std.npy new file mode 100644 index 0000000000000000000000000000000000000000..afdbc50ed0cfe53be8b317c2f24f6a674ddab536 --- /dev/null +++ b/v2.1/german_24l/constants/emb_std.npy @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:11ae1576585fff8884e2c1f1d12427479594e634ddc6a8c7209e8bf44020b5a5 +size 256 diff --git a/v2.1/german_24l/constants/mimi_init_state.npz b/v2.1/german_24l/constants/mimi_init_state.npz new file mode 100644 index 0000000000000000000000000000000000000000..3b52f383a86ef0439f9993506e4529c4c1234828 --- /dev/null +++ b/v2.1/german_24l/constants/mimi_init_state.npz @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:515588e10e9c14a6b009f1cfda9fcd592542b1640ad856d5fc69e50024702a39 +size 1083108 diff --git a/v2.1/german_24l/constants/quantizer_weight.npy b/v2.1/german_24l/constants/quantizer_weight.npy new file mode 100644 index 0000000000000000000000000000000000000000..2d53e5ced6278307fbff45e258705b0439fa985c --- /dev/null +++ b/v2.1/german_24l/constants/quantizer_weight.npy @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:b9686cb8cb675e4c94515f5bee5268858e8a455fbc5934aaec4677be73fa97a1 +size 65664 diff --git a/v2.1/german_24l/constants/text_embed_table.npy b/v2.1/german_24l/constants/text_embed_table.npy new file mode 100644 index 0000000000000000000000000000000000000000..df71d39236f6122b024dd018eb76db426118fb2b --- /dev/null +++ b/v2.1/german_24l/constants/text_embed_table.npy @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:953c340d549f9535f25a2735302f624e236a4eab66a674d83519394f3988198a +size 16388224 diff --git a/v2.1/german_24l/constants_bin/alba.safetensors b/v2.1/german_24l/constants_bin/alba.safetensors new file mode 100644 index 0000000000000000000000000000000000000000..1d530324c3bc20c1fb27a36b915283f76caae16e --- /dev/null +++ b/v2.1/german_24l/constants_bin/alba.safetensors @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:58273ae4d4aca6ea4ac50a7603040acd269fb9f200711ed4ef156a342dd4a6d1 +size 24777760 diff --git a/v2.1/german_24l/constants_bin/anna.safetensors b/v2.1/german_24l/constants_bin/anna.safetensors new file mode 100644 index 0000000000000000000000000000000000000000..8be5cfe7a20e24392f882a508788ac69de802dae --- /dev/null +++ b/v2.1/german_24l/constants_bin/anna.safetensors @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:1df6f5b97f8ce1d26a46ac0ddfe9c3d77f41f760b2fb7fa78a14844dc2c43924 +size 31265824 diff --git a/v2.1/german_24l/constants_bin/azelma.safetensors b/v2.1/german_24l/constants_bin/azelma.safetensors new file mode 100644 index 0000000000000000000000000000000000000000..9933c845634d598a2302dfad834395ee44eb427b --- /dev/null +++ b/v2.1/german_24l/constants_bin/azelma.safetensors @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:b1eb7710f9424c36048cbbf9229c69de5035995355366bc2ca09fe07d5262aca +size 31855648 diff --git a/v2.1/german_24l/constants_bin/bill_boerst.safetensors b/v2.1/german_24l/constants_bin/bill_boerst.safetensors new file mode 100644 index 0000000000000000000000000000000000000000..f40504dd18f28917c88543e7bfce87015637aa85 --- /dev/null +++ b/v2.1/german_24l/constants_bin/bill_boerst.safetensors @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:0ef62df88a468eea1d4cee8c746bde87ff4141000bf2475ad1dca23d135a4f77 +size 26940448 diff --git a/v2.1/german_24l/constants_bin/bos_before_voice.bin b/v2.1/german_24l/constants_bin/bos_before_voice.bin new file mode 100644 index 0000000000000000000000000000000000000000..dd78964ca9d31ee69de05b1268cd2242f1370ac6 --- /dev/null +++ b/v2.1/german_24l/constants_bin/bos_before_voice.bin @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:841b647bca5ba1cab895e3212343d968df6ea0dde54bfd2d4403b9e5d9012b60 +size 4096 diff --git a/v2.1/german_24l/constants_bin/bos_emb.bin b/v2.1/german_24l/constants_bin/bos_emb.bin new file mode 100644 index 0000000000000000000000000000000000000000..458b9902032c7edc4cfda6e5b2c34f0ed6f66a2b --- /dev/null +++ b/v2.1/german_24l/constants_bin/bos_emb.bin @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:0ad5279a1664228d96e2f063f9babb27fad13c0b8e73d003e02275ad93181173 +size 128 diff --git a/v2.1/german_24l/constants_bin/caro_davy.safetensors b/v2.1/german_24l/constants_bin/caro_davy.safetensors new file mode 100644 index 0000000000000000000000000000000000000000..21553b5c74996470d4ba532c0752f15cfca45b00 --- /dev/null +++ b/v2.1/german_24l/constants_bin/caro_davy.safetensors @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:6ab1847ec2da8cb30c2509a793debab79ebd98b741dad927061674a72b43fb50 +size 21042200 diff --git a/v2.1/german_24l/constants_bin/charles.safetensors b/v2.1/german_24l/constants_bin/charles.safetensors new file mode 100644 index 0000000000000000000000000000000000000000..ff990c53175d7291b08933bced2bea5c0b41531d --- /dev/null +++ b/v2.1/german_24l/constants_bin/charles.safetensors @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:1dab8552536b517b7378927fd7c955461e986b311f867678705752bda4019f68 +size 24777760 diff --git a/v2.1/german_24l/constants_bin/cosette.safetensors b/v2.1/german_24l/constants_bin/cosette.safetensors new file mode 100644 index 0000000000000000000000000000000000000000..d30f4b02e95c9973b017a16626befc9908d24fdd --- /dev/null +++ b/v2.1/german_24l/constants_bin/cosette.safetensors @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:71658d75661ebf5a4fee70c83dfe2ee69e97ef229b03438117a210637da0ccc1 +size 24777760 diff --git a/v2.1/german_24l/constants_bin/eponine.safetensors b/v2.1/german_24l/constants_bin/eponine.safetensors new file mode 100644 index 0000000000000000000000000000000000000000..8dc4408db122057b13d9cad6c64eb881fe59d6f5 --- /dev/null +++ b/v2.1/german_24l/constants_bin/eponine.safetensors @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:d06ac765ad96de2b79342a7f09952b004fb0ea3b4c621ee6103416a8943ee416 +size 27726880 diff --git a/v2.1/german_24l/constants_bin/estelle.safetensors b/v2.1/german_24l/constants_bin/estelle.safetensors new file mode 100644 index 0000000000000000000000000000000000000000..d33728d174de33ea9e9fa8edd614e7f6da6d38aa --- /dev/null +++ b/v2.1/german_24l/constants_bin/estelle.safetensors @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:b167b134fbf8d908c069170c7940d0c21d2e5b2a546926e73c71a303b1740911 +size 33035296 diff --git a/v2.1/german_24l/constants_bin/eve.safetensors b/v2.1/german_24l/constants_bin/eve.safetensors new file mode 100644 index 0000000000000000000000000000000000000000..6793343abaaf8c282c54ea1db1e266a8a5225289 --- /dev/null +++ b/v2.1/german_24l/constants_bin/eve.safetensors @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:72d34ccc7c4f124d25cde1b4331def40e8cf50fa1a5c8a7331671364f94f0be4 +size 26154016 diff --git a/v2.1/german_24l/constants_bin/fantine.safetensors b/v2.1/german_24l/constants_bin/fantine.safetensors new file mode 100644 index 0000000000000000000000000000000000000000..c70d55de32c230add94d9655cf2b73f98dc66e9d --- /dev/null +++ b/v2.1/german_24l/constants_bin/fantine.safetensors @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:a214c3e2ecca568a9160af75134ab848474ce1b154288b148ea8c7bf6fbd212b +size 26154016 diff --git a/v2.1/german_24l/constants_bin/george.safetensors b/v2.1/german_24l/constants_bin/george.safetensors new file mode 100644 index 0000000000000000000000000000000000000000..0799a3971f4dedfc6d6e5bdf1679b76d89507b47 --- /dev/null +++ b/v2.1/german_24l/constants_bin/george.safetensors @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:5a72e2af9f88663d8d0c5dc91bab4c4f639daa0ed0236145b234ca6fcad25d0e +size 24974368 diff --git a/v2.1/german_24l/constants_bin/giovanni.safetensors b/v2.1/german_24l/constants_bin/giovanni.safetensors new file mode 100644 index 0000000000000000000000000000000000000000..09bca0bbdd22c6d43544adec0272f4bfe5007bda --- /dev/null +++ b/v2.1/german_24l/constants_bin/giovanni.safetensors @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:9cc77677c5c5c47e61cd8e74c3c70584ff8519a3e6ebb53f9b41b0aae2c03be7 +size 18486272 diff --git a/v2.1/german_24l/constants_bin/jane.safetensors b/v2.1/german_24l/constants_bin/jane.safetensors new file mode 100644 index 0000000000000000000000000000000000000000..5d50bd242d3bf76eb4d12a182793078175bee153 --- /dev/null +++ b/v2.1/german_24l/constants_bin/jane.safetensors @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:dce4c2f55920c15025751b4a8e8e5ee988e9c14a4ffa7617d15d5fa98dbe87a6 +size 29496352 diff --git a/v2.1/german_24l/constants_bin/javert.safetensors b/v2.1/german_24l/constants_bin/javert.safetensors new file mode 100644 index 0000000000000000000000000000000000000000..7f35cb204c6e4c5ef7dcfedabaf83a6041836152 --- /dev/null +++ b/v2.1/german_24l/constants_bin/javert.safetensors @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:d18798a0d30d1f09f8b219dba8bc344f45e4eebc074f0b8ed371267e4f7d7380 +size 24777760 diff --git a/v2.1/german_24l/constants_bin/jean.safetensors b/v2.1/german_24l/constants_bin/jean.safetensors new file mode 100644 index 0000000000000000000000000000000000000000..44b212ff8073edd2e53c836b6347abefa0742d84 --- /dev/null +++ b/v2.1/german_24l/constants_bin/jean.safetensors @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:dac764701c0e449119d7eb25ccacb6557df63133e53acba2fded552782746dae +size 24777760 diff --git a/v2.1/german_24l/constants_bin/juergen.safetensors b/v2.1/german_24l/constants_bin/juergen.safetensors new file mode 100644 index 0000000000000000000000000000000000000000..140bdb4f221b8617532dfb646e4475aed29b20b6 --- /dev/null +++ b/v2.1/german_24l/constants_bin/juergen.safetensors @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:e18630b75ebdbff7930c161a66d60351a2010a18875c0c06fa9c547013e6913f +size 24974368 diff --git a/v2.1/german_24l/constants_bin/lola.safetensors b/v2.1/german_24l/constants_bin/lola.safetensors new file mode 100644 index 0000000000000000000000000000000000000000..2318f95f260cd1e814003e5f5f1cd84830c28fc3 --- /dev/null +++ b/v2.1/german_24l/constants_bin/lola.safetensors @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:adebe1f4c560b87111c6b5b3521913741055265314bb5c264f93a15362cfa3e3 +size 23794712 diff --git a/v2.1/german_24l/constants_bin/marius.safetensors b/v2.1/german_24l/constants_bin/marius.safetensors new file mode 100644 index 0000000000000000000000000000000000000000..0c4dc9e810864eb414f2871ae1ec26564198e925 --- /dev/null +++ b/v2.1/german_24l/constants_bin/marius.safetensors @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:5f1eac8916357a16367008e50469cbf9dc7dc49afe4fb63391200e6373609d75 +size 24777760 diff --git a/v2.1/german_24l/constants_bin/mary.safetensors b/v2.1/german_24l/constants_bin/mary.safetensors new file mode 100644 index 0000000000000000000000000000000000000000..e2164fbab392601534c35a249118a99982076efa --- /dev/null +++ b/v2.1/german_24l/constants_bin/mary.safetensors @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:ab44629656234eac2c15dacbd418a29e935019616cb748f7cdaebc84b6224c0c +size 24777760 diff --git a/v2.1/german_24l/constants_bin/michael.safetensors b/v2.1/german_24l/constants_bin/michael.safetensors new file mode 100644 index 0000000000000000000000000000000000000000..c93ad5defd341e2f33661ccf7e763b0ae321ee43 --- /dev/null +++ b/v2.1/german_24l/constants_bin/michael.safetensors @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:800c1877cae884a22a04caf5576ef63aab740c0cc92d3b5e44934d9fa0f66c97 +size 29103136 diff --git a/v2.1/german_24l/constants_bin/paul.safetensors b/v2.1/german_24l/constants_bin/paul.safetensors new file mode 100644 index 0000000000000000000000000000000000000000..835dc2f96f18ce802401a7c2282f308b979b6245 --- /dev/null +++ b/v2.1/german_24l/constants_bin/paul.safetensors @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:a3d6b6d28ec24efa60566833321d4b4113bf38a1fa1843aee5f2273e1290cfa6 +size 27923488 diff --git a/v2.1/german_24l/constants_bin/peter_yearsley.safetensors b/v2.1/german_24l/constants_bin/peter_yearsley.safetensors new file mode 100644 index 0000000000000000000000000000000000000000..aee014ab173ca32c0eb94cb5b48f44f30656072f --- /dev/null +++ b/v2.1/german_24l/constants_bin/peter_yearsley.safetensors @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:823f15e725490b610e8bd018b82b2e243fc6e6b2096f7e8aeb56e382279a803a +size 14947320 diff --git a/v2.1/german_24l/constants_bin/rafael.safetensors b/v2.1/german_24l/constants_bin/rafael.safetensors new file mode 100644 index 0000000000000000000000000000000000000000..01ef991af12179d0cedb4da93e37179e0d054f6b --- /dev/null +++ b/v2.1/german_24l/constants_bin/rafael.safetensors @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:efb1b3be69cb079e450ce3181f6e96b413d048b146fc664abd67ba67803f4713 +size 24777760 diff --git a/v2.1/german_24l/constants_bin/stuart_bell.safetensors b/v2.1/german_24l/constants_bin/stuart_bell.safetensors new file mode 100644 index 0000000000000000000000000000000000000000..35b31aac96f7da7754fe4f1b3b5d5338445ed39a --- /dev/null +++ b/v2.1/german_24l/constants_bin/stuart_bell.safetensors @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:f7e0d664e5894d42fb6d16f648810f876169ba08e5c711759ef694f35b2dd1b1 +size 21042200 diff --git a/v2.1/german_24l/constants_bin/text_embed_table.bin b/v2.1/german_24l/constants_bin/text_embed_table.bin new file mode 100644 index 0000000000000000000000000000000000000000..dbf679c240d4613df768552709ebd04ce22cf08d --- /dev/null +++ b/v2.1/german_24l/constants_bin/text_embed_table.bin @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:c3ae7d828116c94c4ca29f360067d62c79df7a6ad4556351613e6f6563872eab +size 16388096 diff --git a/v2.1/german_24l/constants_bin/tokenizer.model b/v2.1/german_24l/constants_bin/tokenizer.model new file mode 100644 index 0000000000000000000000000000000000000000..40e0004b17234ee1071f728c7780d7edcc176bdd --- /dev/null +++ b/v2.1/german_24l/constants_bin/tokenizer.model @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:389079b9c67cd17ff1c565d53bc695df390b4ca8abf2300947e659585edd9d2a +size 59837 diff --git a/v2.1/german_24l/constants_bin/vera.safetensors b/v2.1/german_24l/constants_bin/vera.safetensors new file mode 100644 index 0000000000000000000000000000000000000000..3e2560723c289ab25e7e1583c4f133c47d389949 --- /dev/null +++ b/v2.1/german_24l/constants_bin/vera.safetensors @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:7a9f60ee102d7e7934c910ef979f23946c373b05ce0e7f0107b6f27e8d000514 +size 26940448 diff --git a/v2.1/german_24l/flow_decoder_fused.mlmodelc/analytics/coremldata.bin b/v2.1/german_24l/flow_decoder_fused.mlmodelc/analytics/coremldata.bin new file mode 100644 index 0000000000000000000000000000000000000000..c76e3a20200850ea6aea86fd94ac8a67ec2b2a21 --- /dev/null +++ b/v2.1/german_24l/flow_decoder_fused.mlmodelc/analytics/coremldata.bin @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:0b2e7bca72ebfafd2797268227a40a0adc76601f39b643bc7525d38f34e9c2aa +size 243 diff --git a/v2.1/german_24l/flow_decoder_fused.mlmodelc/coremldata.bin b/v2.1/german_24l/flow_decoder_fused.mlmodelc/coremldata.bin new file mode 100644 index 0000000000000000000000000000000000000000..c30cb5a5266993877ea3c9790430c8761986c919 --- /dev/null +++ b/v2.1/german_24l/flow_decoder_fused.mlmodelc/coremldata.bin @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:ee0dd65c27b5c92c6ded2da7b2a3a0cca403b74539f023533711de758947bb47 +size 413 diff --git a/v2.1/german_24l/flow_decoder_fused.mlmodelc/model.mil b/v2.1/german_24l/flow_decoder_fused.mlmodelc/model.mil new file mode 100644 index 0000000000000000000000000000000000000000..c64a1f34555e7bbf9c8d4502d29f9ef52dd6b0ec --- /dev/null +++ b/v2.1/german_24l/flow_decoder_fused.mlmodelc/model.mil @@ -0,0 +1,1915 @@ +program(1.0) +[buildInfo = dict, tensor>({{"coremlc-component-MIL", "3520.4.1"}, {"coremlc-version", "3520.5.1"}, {"coremltools-component-torch", "2.12.0"}, {"coremltools-source-dialect", "TorchScript"}, {"coremltools-version", "9.0"}})] +{ + func main(tensor latent_init, tensor transformer_out) { + tensor latent_init_to_fp16_dtype_0 = const()[name = tensor("latent_init_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor flow_net_input_proj_weight_to_fp16 = const()[name = tensor("flow_net_input_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(64)))]; + tensor flow_net_input_proj_bias_to_fp16 = const()[name = tensor("flow_net_input_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(32896)))]; + tensor latent_init_to_fp16 = cast(dtype = latent_init_to_fp16_dtype_0, x = latent_init)[name = tensor("cast_146")]; + tensor linear_0_cast_fp16 = linear(bias = flow_net_input_proj_bias_to_fp16, weight = flow_net_input_proj_weight_to_fp16, x = latent_init_to_fp16)[name = tensor("linear_0_cast_fp16")]; + tensor input_3_to_fp16 = const()[name = tensor("input_3_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(33984)))]; + tensor input_5_cast_fp16 = silu(x = input_3_to_fp16)[name = tensor("input_5_cast_fp16")]; + tensor flow_net_time_embed_0_mlp_2_weight_to_fp16 = const()[name = tensor("flow_net_time_embed_0_mlp_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(35072)))]; + tensor flow_net_time_embed_0_mlp_2_bias_to_fp16 = const()[name = tensor("flow_net_time_embed_0_mlp_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(559424)))]; + tensor linear_2_cast_fp16 = linear(bias = flow_net_time_embed_0_mlp_2_bias_to_fp16, weight = flow_net_time_embed_0_mlp_2_weight_to_fp16, x = input_5_cast_fp16)[name = tensor("linear_2_cast_fp16")]; + tensor reduce_mean_0_axes_0 = const()[name = tensor("reduce_mean_0_axes_0"), val = tensor([-1])]; + tensor reduce_mean_0_keep_dims_0 = const()[name = tensor("reduce_mean_0_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_0_cast_fp16 = reduce_mean(axes = reduce_mean_0_axes_0, keep_dims = reduce_mean_0_keep_dims_0, x = linear_2_cast_fp16)[name = tensor("reduce_mean_0_cast_fp16")]; + tensor sub_0_cast_fp16 = sub(x = linear_2_cast_fp16, y = reduce_mean_0_cast_fp16)[name = tensor("sub_0_cast_fp16")]; + tensor square_0_cast_fp16 = square(x = sub_0_cast_fp16)[name = tensor("square_0_cast_fp16")]; + tensor reduce_mean_1_axes_0 = const()[name = tensor("reduce_mean_1_axes_0"), val = tensor([-1])]; + tensor reduce_mean_1_keep_dims_0 = const()[name = tensor("reduce_mean_1_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_1_cast_fp16 = reduce_mean(axes = reduce_mean_1_axes_0, keep_dims = reduce_mean_1_keep_dims_0, x = square_0_cast_fp16)[name = tensor("reduce_mean_1_cast_fp16")]; + tensor real_div_0_to_fp16 = const()[name = tensor("real_div_0_to_fp16"), val = tensor(0x1.008p+0)]; + tensor mul_0_cast_fp16 = mul(x = reduce_mean_1_cast_fp16, y = real_div_0_to_fp16)[name = tensor("mul_0_cast_fp16")]; + tensor var_70_to_fp16 = const()[name = tensor("op_70_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_1_cast_fp16 = add(x = mul_0_cast_fp16, y = var_70_to_fp16)[name = tensor("var_1_cast_fp16")]; + tensor var_73_epsilon_0 = const()[name = tensor("op_73_epsilon_0"), val = tensor(0x1.197998p-40)]; + tensor var_73_cast_fp16 = rsqrt(epsilon = var_73_epsilon_0, x = var_1_cast_fp16)[name = tensor("op_73_cast_fp16")]; + tensor const_3_to_fp16 = const()[name = tensor("const_3_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(560512)))]; + tensor var_74_cast_fp16 = mul(x = const_3_to_fp16, y = var_73_cast_fp16)[name = tensor("op_74_cast_fp16")]; + tensor var_75_cast_fp16 = mul(x = linear_2_cast_fp16, y = var_74_cast_fp16)[name = tensor("op_75_cast_fp16")]; + tensor input_9_to_fp16 = const()[name = tensor("input_9_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(561600)))]; + tensor input_11_cast_fp16 = silu(x = input_9_to_fp16)[name = tensor("input_11_cast_fp16")]; + tensor flow_net_time_embed_1_mlp_2_weight_to_fp16 = const()[name = tensor("flow_net_time_embed_1_mlp_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(562688)))]; + tensor flow_net_time_embed_1_mlp_2_bias_to_fp16 = const()[name = tensor("flow_net_time_embed_1_mlp_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1087040)))]; + tensor linear_4_cast_fp16 = linear(bias = flow_net_time_embed_1_mlp_2_bias_to_fp16, weight = flow_net_time_embed_1_mlp_2_weight_to_fp16, x = input_11_cast_fp16)[name = tensor("linear_4_cast_fp16")]; + tensor reduce_mean_2_axes_0 = const()[name = tensor("reduce_mean_2_axes_0"), val = tensor([-1])]; + tensor reduce_mean_2_keep_dims_0 = const()[name = tensor("reduce_mean_2_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_2_cast_fp16 = reduce_mean(axes = reduce_mean_2_axes_0, keep_dims = reduce_mean_2_keep_dims_0, x = linear_4_cast_fp16)[name = tensor("reduce_mean_2_cast_fp16")]; + tensor sub_2_cast_fp16 = sub(x = linear_4_cast_fp16, y = reduce_mean_2_cast_fp16)[name = tensor("sub_2_cast_fp16")]; + tensor square_1_cast_fp16 = square(x = sub_2_cast_fp16)[name = tensor("square_1_cast_fp16")]; + tensor reduce_mean_3_axes_0 = const()[name = tensor("reduce_mean_3_axes_0"), val = tensor([-1])]; + tensor reduce_mean_3_keep_dims_0 = const()[name = tensor("reduce_mean_3_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_3_cast_fp16 = reduce_mean(axes = reduce_mean_3_axes_0, keep_dims = reduce_mean_3_keep_dims_0, x = square_1_cast_fp16)[name = tensor("reduce_mean_3_cast_fp16")]; + tensor real_div_1_to_fp16 = const()[name = tensor("real_div_1_to_fp16"), val = tensor(0x1.008p+0)]; + tensor mul_1_cast_fp16 = mul(x = reduce_mean_3_cast_fp16, y = real_div_1_to_fp16)[name = tensor("mul_1_cast_fp16")]; + tensor var_110_to_fp16 = const()[name = tensor("op_110_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_3_cast_fp16 = add(x = mul_1_cast_fp16, y = var_110_to_fp16)[name = tensor("var_3_cast_fp16")]; + tensor var_113_epsilon_0 = const()[name = tensor("op_113_epsilon_0"), val = tensor(0x1.197998p-40)]; + tensor var_113_cast_fp16 = rsqrt(epsilon = var_113_epsilon_0, x = var_3_cast_fp16)[name = tensor("op_113_cast_fp16")]; + tensor const_5_to_fp16 = const()[name = tensor("const_5_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1088128)))]; + tensor var_114_cast_fp16 = mul(x = const_5_to_fp16, y = var_113_cast_fp16)[name = tensor("op_114_cast_fp16")]; + tensor var_115_cast_fp16 = mul(x = linear_4_cast_fp16, y = var_114_cast_fp16)[name = tensor("op_115_cast_fp16")]; + tensor var_127_cast_fp16 = add(x = var_75_cast_fp16, y = var_115_cast_fp16)[name = tensor("op_127_cast_fp16")]; + tensor _inversed_t_combined_1_y_0_to_fp16 = const()[name = tensor("_inversed_t_combined_1_y_0_to_fp16"), val = tensor(0x1p-1)]; + tensor _inversed_t_combined_1_cast_fp16 = mul(x = var_127_cast_fp16, y = _inversed_t_combined_1_y_0_to_fp16)[name = tensor("_inversed_t_combined_1_cast_fp16")]; + tensor transformer_out_to_fp16_dtype_0 = const()[name = tensor("transformer_out_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor flow_net_cond_embed_weight_to_fp16 = const()[name = tensor("flow_net_cond_embed_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1089216)))]; + tensor flow_net_cond_embed_bias_to_fp16 = const()[name = tensor("flow_net_cond_embed_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(2137856)))]; + tensor transformer_out_to_fp16 = cast(dtype = transformer_out_to_fp16_dtype_0, x = transformer_out)[name = tensor("cast_145")]; + tensor linear_5_cast_fp16 = linear(bias = flow_net_cond_embed_bias_to_fp16, weight = flow_net_cond_embed_weight_to_fp16, x = transformer_out_to_fp16)[name = tensor("linear_5_cast_fp16")]; + tensor input_13_cast_fp16 = add(x = _inversed_t_combined_1_cast_fp16, y = linear_5_cast_fp16)[name = tensor("input_13_cast_fp16")]; + tensor input_15_cast_fp16 = silu(x = input_13_cast_fp16)[name = tensor("input_15_cast_fp16")]; + tensor flow_net_res_blocks_0_adaLN_modulation_1_weight_to_fp16 = const()[name = tensor("flow_net_res_blocks_0_adaLN_modulation_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(2138944)))]; + tensor flow_net_res_blocks_0_adaLN_modulation_1_bias_to_fp16 = const()[name = tensor("flow_net_res_blocks_0_adaLN_modulation_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(3711872)))]; + tensor linear_6_cast_fp16 = linear(bias = flow_net_res_blocks_0_adaLN_modulation_1_bias_to_fp16, weight = flow_net_res_blocks_0_adaLN_modulation_1_weight_to_fp16, x = input_15_cast_fp16)[name = tensor("linear_6_cast_fp16")]; + tensor var_142_split_sizes_0 = const()[name = tensor("op_142_split_sizes_0"), val = tensor([512, 512, 512])]; + tensor var_142_axis_0 = const()[name = tensor("op_142_axis_0"), val = tensor(-1)]; + tensor var_142_cast_fp16_0, tensor var_142_cast_fp16_1, tensor var_142_cast_fp16_2 = split(axis = var_142_axis_0, split_sizes = var_142_split_sizes_0, x = linear_6_cast_fp16)[name = tensor("op_142_cast_fp16")]; + tensor mean_1_axes_0 = const()[name = tensor("mean_1_axes_0"), val = tensor([-1])]; + tensor mean_1_keep_dims_0 = const()[name = tensor("mean_1_keep_dims_0"), val = tensor(true)]; + tensor mean_1_cast_fp16 = reduce_mean(axes = mean_1_axes_0, keep_dims = mean_1_keep_dims_0, x = linear_0_cast_fp16)[name = tensor("mean_1_cast_fp16")]; + tensor sub_4_cast_fp16 = sub(x = linear_0_cast_fp16, y = mean_1_cast_fp16)[name = tensor("sub_4_cast_fp16")]; + tensor square_2_cast_fp16 = square(x = sub_4_cast_fp16)[name = tensor("square_2_cast_fp16")]; + tensor reduce_mean_5_axes_0 = const()[name = tensor("reduce_mean_5_axes_0"), val = tensor([-1])]; + tensor reduce_mean_5_keep_dims_0 = const()[name = tensor("reduce_mean_5_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_5_cast_fp16 = reduce_mean(axes = reduce_mean_5_axes_0, keep_dims = reduce_mean_5_keep_dims_0, x = square_2_cast_fp16)[name = tensor("reduce_mean_5_cast_fp16")]; + tensor var_152_to_fp16 = const()[name = tensor("op_152_to_fp16"), val = tensor(0x1.1p-20)]; + tensor var_153_cast_fp16 = add(x = reduce_mean_5_cast_fp16, y = var_152_to_fp16)[name = tensor("op_153_cast_fp16")]; + tensor var_154_cast_fp16 = sqrt(x = var_153_cast_fp16)[name = tensor("op_154_cast_fp16")]; + tensor x_7_cast_fp16 = real_div(x = sub_4_cast_fp16, y = var_154_cast_fp16)[name = tensor("x_7_cast_fp16")]; + tensor flow_net_res_blocks_0_in_ln_weight_to_fp16 = const()[name = tensor("flow_net_res_blocks_0_in_ln_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(3715008)))]; + tensor var_156_cast_fp16 = mul(x = x_7_cast_fp16, y = flow_net_res_blocks_0_in_ln_weight_to_fp16)[name = tensor("op_156_cast_fp16")]; + tensor flow_net_res_blocks_0_in_ln_bias_to_fp16 = const()[name = tensor("flow_net_res_blocks_0_in_ln_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(3716096)))]; + tensor x_9_cast_fp16 = add(x = var_156_cast_fp16, y = flow_net_res_blocks_0_in_ln_bias_to_fp16)[name = tensor("x_9_cast_fp16")]; + tensor var_158_promoted_to_fp16 = const()[name = tensor("op_158_promoted_to_fp16"), val = tensor(0x1p+0)]; + tensor var_159_cast_fp16 = add(x = var_142_cast_fp16_1, y = var_158_promoted_to_fp16)[name = tensor("op_159_cast_fp16")]; + tensor var_160_cast_fp16 = mul(x = x_9_cast_fp16, y = var_159_cast_fp16)[name = tensor("op_160_cast_fp16")]; + tensor input_17_cast_fp16 = add(x = var_160_cast_fp16, y = var_142_cast_fp16_0)[name = tensor("input_17_cast_fp16")]; + tensor flow_net_res_blocks_0_mlp_0_weight_to_fp16 = const()[name = tensor("flow_net_res_blocks_0_mlp_0_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(3717184)))]; + tensor flow_net_res_blocks_0_mlp_0_bias_to_fp16 = const()[name = tensor("flow_net_res_blocks_0_mlp_0_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(4241536)))]; + tensor linear_7_cast_fp16 = linear(bias = flow_net_res_blocks_0_mlp_0_bias_to_fp16, weight = flow_net_res_blocks_0_mlp_0_weight_to_fp16, x = input_17_cast_fp16)[name = tensor("linear_7_cast_fp16")]; + tensor input_21_cast_fp16 = silu(x = linear_7_cast_fp16)[name = tensor("input_21_cast_fp16")]; + tensor flow_net_res_blocks_0_mlp_2_weight_to_fp16 = const()[name = tensor("flow_net_res_blocks_0_mlp_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(4242624)))]; + tensor flow_net_res_blocks_0_mlp_2_bias_to_fp16 = const()[name = tensor("flow_net_res_blocks_0_mlp_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(4766976)))]; + tensor linear_8_cast_fp16 = linear(bias = flow_net_res_blocks_0_mlp_2_bias_to_fp16, weight = flow_net_res_blocks_0_mlp_2_weight_to_fp16, x = input_21_cast_fp16)[name = tensor("linear_8_cast_fp16")]; + tensor var_171_cast_fp16 = mul(x = var_142_cast_fp16_2, y = linear_8_cast_fp16)[name = tensor("op_171_cast_fp16")]; + tensor x_11_cast_fp16 = add(x = linear_0_cast_fp16, y = var_171_cast_fp16)[name = tensor("x_11_cast_fp16")]; + tensor flow_net_res_blocks_1_adaLN_modulation_1_weight_to_fp16 = const()[name = tensor("flow_net_res_blocks_1_adaLN_modulation_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(4768064)))]; + tensor flow_net_res_blocks_1_adaLN_modulation_1_bias_to_fp16 = const()[name = tensor("flow_net_res_blocks_1_adaLN_modulation_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(6340992)))]; + tensor linear_9_cast_fp16 = linear(bias = flow_net_res_blocks_1_adaLN_modulation_1_bias_to_fp16, weight = flow_net_res_blocks_1_adaLN_modulation_1_weight_to_fp16, x = input_15_cast_fp16)[name = tensor("linear_9_cast_fp16")]; + tensor var_181_split_sizes_0 = const()[name = tensor("op_181_split_sizes_0"), val = tensor([512, 512, 512])]; + tensor var_181_axis_0 = const()[name = tensor("op_181_axis_0"), val = tensor(-1)]; + tensor var_181_cast_fp16_0, tensor var_181_cast_fp16_1, tensor var_181_cast_fp16_2 = split(axis = var_181_axis_0, split_sizes = var_181_split_sizes_0, x = linear_9_cast_fp16)[name = tensor("op_181_cast_fp16")]; + tensor mean_3_axes_0 = const()[name = tensor("mean_3_axes_0"), val = tensor([-1])]; + tensor mean_3_keep_dims_0 = const()[name = tensor("mean_3_keep_dims_0"), val = tensor(true)]; + tensor mean_3_cast_fp16 = reduce_mean(axes = mean_3_axes_0, keep_dims = mean_3_keep_dims_0, x = x_11_cast_fp16)[name = tensor("mean_3_cast_fp16")]; + tensor sub_5_cast_fp16 = sub(x = x_11_cast_fp16, y = mean_3_cast_fp16)[name = tensor("sub_5_cast_fp16")]; + tensor square_3_cast_fp16 = square(x = sub_5_cast_fp16)[name = tensor("square_3_cast_fp16")]; + tensor reduce_mean_7_axes_0 = const()[name = tensor("reduce_mean_7_axes_0"), val = tensor([-1])]; + tensor reduce_mean_7_keep_dims_0 = const()[name = tensor("reduce_mean_7_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_7_cast_fp16 = reduce_mean(axes = reduce_mean_7_axes_0, keep_dims = reduce_mean_7_keep_dims_0, x = square_3_cast_fp16)[name = tensor("reduce_mean_7_cast_fp16")]; + tensor var_191_to_fp16 = const()[name = tensor("op_191_to_fp16"), val = tensor(0x1.1p-20)]; + tensor var_192_cast_fp16 = add(x = reduce_mean_7_cast_fp16, y = var_191_to_fp16)[name = tensor("op_192_cast_fp16")]; + tensor var_193_cast_fp16 = sqrt(x = var_192_cast_fp16)[name = tensor("op_193_cast_fp16")]; + tensor x_13_cast_fp16 = real_div(x = sub_5_cast_fp16, y = var_193_cast_fp16)[name = tensor("x_13_cast_fp16")]; + tensor flow_net_res_blocks_1_in_ln_weight_to_fp16 = const()[name = tensor("flow_net_res_blocks_1_in_ln_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(6344128)))]; + tensor var_195_cast_fp16 = mul(x = x_13_cast_fp16, y = flow_net_res_blocks_1_in_ln_weight_to_fp16)[name = tensor("op_195_cast_fp16")]; + tensor flow_net_res_blocks_1_in_ln_bias_to_fp16 = const()[name = tensor("flow_net_res_blocks_1_in_ln_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(6345216)))]; + tensor x_15_cast_fp16 = add(x = var_195_cast_fp16, y = flow_net_res_blocks_1_in_ln_bias_to_fp16)[name = tensor("x_15_cast_fp16")]; + tensor var_197_promoted_to_fp16 = const()[name = tensor("op_197_promoted_to_fp16"), val = tensor(0x1p+0)]; + tensor var_198_cast_fp16 = add(x = var_181_cast_fp16_1, y = var_197_promoted_to_fp16)[name = tensor("op_198_cast_fp16")]; + tensor var_199_cast_fp16 = mul(x = x_15_cast_fp16, y = var_198_cast_fp16)[name = tensor("op_199_cast_fp16")]; + tensor input_25_cast_fp16 = add(x = var_199_cast_fp16, y = var_181_cast_fp16_0)[name = tensor("input_25_cast_fp16")]; + tensor flow_net_res_blocks_1_mlp_0_weight_to_fp16 = const()[name = tensor("flow_net_res_blocks_1_mlp_0_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(6346304)))]; + tensor flow_net_res_blocks_1_mlp_0_bias_to_fp16 = const()[name = tensor("flow_net_res_blocks_1_mlp_0_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(6870656)))]; + tensor linear_10_cast_fp16 = linear(bias = flow_net_res_blocks_1_mlp_0_bias_to_fp16, weight = flow_net_res_blocks_1_mlp_0_weight_to_fp16, x = input_25_cast_fp16)[name = tensor("linear_10_cast_fp16")]; + tensor input_29_cast_fp16 = silu(x = linear_10_cast_fp16)[name = tensor("input_29_cast_fp16")]; + tensor flow_net_res_blocks_1_mlp_2_weight_to_fp16 = const()[name = tensor("flow_net_res_blocks_1_mlp_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(6871744)))]; + tensor flow_net_res_blocks_1_mlp_2_bias_to_fp16 = const()[name = tensor("flow_net_res_blocks_1_mlp_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(7396096)))]; + tensor linear_11_cast_fp16 = linear(bias = flow_net_res_blocks_1_mlp_2_bias_to_fp16, weight = flow_net_res_blocks_1_mlp_2_weight_to_fp16, x = input_29_cast_fp16)[name = tensor("linear_11_cast_fp16")]; + tensor var_210_cast_fp16 = mul(x = var_181_cast_fp16_2, y = linear_11_cast_fp16)[name = tensor("op_210_cast_fp16")]; + tensor x_17_cast_fp16 = add(x = x_11_cast_fp16, y = var_210_cast_fp16)[name = tensor("x_17_cast_fp16")]; + tensor flow_net_res_blocks_2_adaLN_modulation_1_weight_to_fp16 = const()[name = tensor("flow_net_res_blocks_2_adaLN_modulation_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(7397184)))]; + tensor flow_net_res_blocks_2_adaLN_modulation_1_bias_to_fp16 = const()[name = tensor("flow_net_res_blocks_2_adaLN_modulation_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8970112)))]; + tensor linear_12_cast_fp16 = linear(bias = flow_net_res_blocks_2_adaLN_modulation_1_bias_to_fp16, weight = flow_net_res_blocks_2_adaLN_modulation_1_weight_to_fp16, x = input_15_cast_fp16)[name = tensor("linear_12_cast_fp16")]; + tensor var_220_split_sizes_0 = const()[name = tensor("op_220_split_sizes_0"), val = tensor([512, 512, 512])]; + tensor var_220_axis_0 = const()[name = tensor("op_220_axis_0"), val = tensor(-1)]; + tensor var_220_cast_fp16_0, tensor var_220_cast_fp16_1, tensor var_220_cast_fp16_2 = split(axis = var_220_axis_0, split_sizes = var_220_split_sizes_0, x = linear_12_cast_fp16)[name = tensor("op_220_cast_fp16")]; + tensor mean_5_axes_0 = const()[name = tensor("mean_5_axes_0"), val = tensor([-1])]; + tensor mean_5_keep_dims_0 = const()[name = tensor("mean_5_keep_dims_0"), val = tensor(true)]; + tensor mean_5_cast_fp16 = reduce_mean(axes = mean_5_axes_0, keep_dims = mean_5_keep_dims_0, x = x_17_cast_fp16)[name = tensor("mean_5_cast_fp16")]; + tensor sub_6_cast_fp16 = sub(x = x_17_cast_fp16, y = mean_5_cast_fp16)[name = tensor("sub_6_cast_fp16")]; + tensor square_4_cast_fp16 = square(x = sub_6_cast_fp16)[name = tensor("square_4_cast_fp16")]; + tensor reduce_mean_9_axes_0 = const()[name = tensor("reduce_mean_9_axes_0"), val = tensor([-1])]; + tensor reduce_mean_9_keep_dims_0 = const()[name = tensor("reduce_mean_9_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_9_cast_fp16 = reduce_mean(axes = reduce_mean_9_axes_0, keep_dims = reduce_mean_9_keep_dims_0, x = square_4_cast_fp16)[name = tensor("reduce_mean_9_cast_fp16")]; + tensor var_230_to_fp16 = const()[name = tensor("op_230_to_fp16"), val = tensor(0x1.1p-20)]; + tensor var_231_cast_fp16 = add(x = reduce_mean_9_cast_fp16, y = var_230_to_fp16)[name = tensor("op_231_cast_fp16")]; + tensor var_232_cast_fp16 = sqrt(x = var_231_cast_fp16)[name = tensor("op_232_cast_fp16")]; + tensor x_19_cast_fp16 = real_div(x = sub_6_cast_fp16, y = var_232_cast_fp16)[name = tensor("x_19_cast_fp16")]; + tensor flow_net_res_blocks_2_in_ln_weight_to_fp16 = const()[name = tensor("flow_net_res_blocks_2_in_ln_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8973248)))]; + tensor var_234_cast_fp16 = mul(x = x_19_cast_fp16, y = flow_net_res_blocks_2_in_ln_weight_to_fp16)[name = tensor("op_234_cast_fp16")]; + tensor flow_net_res_blocks_2_in_ln_bias_to_fp16 = const()[name = tensor("flow_net_res_blocks_2_in_ln_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8974336)))]; + tensor x_21_cast_fp16 = add(x = var_234_cast_fp16, y = flow_net_res_blocks_2_in_ln_bias_to_fp16)[name = tensor("x_21_cast_fp16")]; + tensor var_236_promoted_to_fp16 = const()[name = tensor("op_236_promoted_to_fp16"), val = tensor(0x1p+0)]; + tensor var_237_cast_fp16 = add(x = var_220_cast_fp16_1, y = var_236_promoted_to_fp16)[name = tensor("op_237_cast_fp16")]; + tensor var_238_cast_fp16 = mul(x = x_21_cast_fp16, y = var_237_cast_fp16)[name = tensor("op_238_cast_fp16")]; + tensor input_33_cast_fp16 = add(x = var_238_cast_fp16, y = var_220_cast_fp16_0)[name = tensor("input_33_cast_fp16")]; + tensor flow_net_res_blocks_2_mlp_0_weight_to_fp16 = const()[name = tensor("flow_net_res_blocks_2_mlp_0_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8975424)))]; + tensor flow_net_res_blocks_2_mlp_0_bias_to_fp16 = const()[name = tensor("flow_net_res_blocks_2_mlp_0_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9499776)))]; + tensor linear_13_cast_fp16 = linear(bias = flow_net_res_blocks_2_mlp_0_bias_to_fp16, weight = flow_net_res_blocks_2_mlp_0_weight_to_fp16, x = input_33_cast_fp16)[name = tensor("linear_13_cast_fp16")]; + tensor input_37_cast_fp16 = silu(x = linear_13_cast_fp16)[name = tensor("input_37_cast_fp16")]; + tensor flow_net_res_blocks_2_mlp_2_weight_to_fp16 = const()[name = tensor("flow_net_res_blocks_2_mlp_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9500864)))]; + tensor flow_net_res_blocks_2_mlp_2_bias_to_fp16 = const()[name = tensor("flow_net_res_blocks_2_mlp_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(10025216)))]; + tensor linear_14_cast_fp16 = linear(bias = flow_net_res_blocks_2_mlp_2_bias_to_fp16, weight = flow_net_res_blocks_2_mlp_2_weight_to_fp16, x = input_37_cast_fp16)[name = tensor("linear_14_cast_fp16")]; + tensor var_249_cast_fp16 = mul(x = var_220_cast_fp16_2, y = linear_14_cast_fp16)[name = tensor("op_249_cast_fp16")]; + tensor x_23_cast_fp16 = add(x = x_17_cast_fp16, y = var_249_cast_fp16)[name = tensor("x_23_cast_fp16")]; + tensor flow_net_res_blocks_3_adaLN_modulation_1_weight_to_fp16 = const()[name = tensor("flow_net_res_blocks_3_adaLN_modulation_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(10026304)))]; + tensor flow_net_res_blocks_3_adaLN_modulation_1_bias_to_fp16 = const()[name = tensor("flow_net_res_blocks_3_adaLN_modulation_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(11599232)))]; + tensor linear_15_cast_fp16 = linear(bias = flow_net_res_blocks_3_adaLN_modulation_1_bias_to_fp16, weight = flow_net_res_blocks_3_adaLN_modulation_1_weight_to_fp16, x = input_15_cast_fp16)[name = tensor("linear_15_cast_fp16")]; + tensor var_259_split_sizes_0 = const()[name = tensor("op_259_split_sizes_0"), val = tensor([512, 512, 512])]; + tensor var_259_axis_0 = const()[name = tensor("op_259_axis_0"), val = tensor(-1)]; + tensor var_259_cast_fp16_0, tensor var_259_cast_fp16_1, tensor var_259_cast_fp16_2 = split(axis = var_259_axis_0, split_sizes = var_259_split_sizes_0, x = linear_15_cast_fp16)[name = tensor("op_259_cast_fp16")]; + tensor mean_7_axes_0 = const()[name = tensor("mean_7_axes_0"), val = tensor([-1])]; + tensor mean_7_keep_dims_0 = const()[name = tensor("mean_7_keep_dims_0"), val = tensor(true)]; + tensor mean_7_cast_fp16 = reduce_mean(axes = mean_7_axes_0, keep_dims = mean_7_keep_dims_0, x = x_23_cast_fp16)[name = tensor("mean_7_cast_fp16")]; + tensor sub_7_cast_fp16 = sub(x = x_23_cast_fp16, y = mean_7_cast_fp16)[name = tensor("sub_7_cast_fp16")]; + tensor square_5_cast_fp16 = square(x = sub_7_cast_fp16)[name = tensor("square_5_cast_fp16")]; + tensor reduce_mean_11_axes_0 = const()[name = tensor("reduce_mean_11_axes_0"), val = tensor([-1])]; + tensor reduce_mean_11_keep_dims_0 = const()[name = tensor("reduce_mean_11_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_11_cast_fp16 = reduce_mean(axes = reduce_mean_11_axes_0, keep_dims = reduce_mean_11_keep_dims_0, x = square_5_cast_fp16)[name = tensor("reduce_mean_11_cast_fp16")]; + tensor var_269_to_fp16 = const()[name = tensor("op_269_to_fp16"), val = tensor(0x1.1p-20)]; + tensor var_270_cast_fp16 = add(x = reduce_mean_11_cast_fp16, y = var_269_to_fp16)[name = tensor("op_270_cast_fp16")]; + tensor var_271_cast_fp16 = sqrt(x = var_270_cast_fp16)[name = tensor("op_271_cast_fp16")]; + tensor x_25_cast_fp16 = real_div(x = sub_7_cast_fp16, y = var_271_cast_fp16)[name = tensor("x_25_cast_fp16")]; + tensor flow_net_res_blocks_3_in_ln_weight_to_fp16 = const()[name = tensor("flow_net_res_blocks_3_in_ln_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(11602368)))]; + tensor var_273_cast_fp16 = mul(x = x_25_cast_fp16, y = flow_net_res_blocks_3_in_ln_weight_to_fp16)[name = tensor("op_273_cast_fp16")]; + tensor flow_net_res_blocks_3_in_ln_bias_to_fp16 = const()[name = tensor("flow_net_res_blocks_3_in_ln_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(11603456)))]; + tensor x_27_cast_fp16 = add(x = var_273_cast_fp16, y = flow_net_res_blocks_3_in_ln_bias_to_fp16)[name = tensor("x_27_cast_fp16")]; + tensor var_275_promoted_to_fp16 = const()[name = tensor("op_275_promoted_to_fp16"), val = tensor(0x1p+0)]; + tensor var_276_cast_fp16 = add(x = var_259_cast_fp16_1, y = var_275_promoted_to_fp16)[name = tensor("op_276_cast_fp16")]; + tensor var_277_cast_fp16 = mul(x = x_27_cast_fp16, y = var_276_cast_fp16)[name = tensor("op_277_cast_fp16")]; + tensor input_41_cast_fp16 = add(x = var_277_cast_fp16, y = var_259_cast_fp16_0)[name = tensor("input_41_cast_fp16")]; + tensor flow_net_res_blocks_3_mlp_0_weight_to_fp16 = const()[name = tensor("flow_net_res_blocks_3_mlp_0_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(11604544)))]; + tensor flow_net_res_blocks_3_mlp_0_bias_to_fp16 = const()[name = tensor("flow_net_res_blocks_3_mlp_0_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(12128896)))]; + tensor linear_16_cast_fp16 = linear(bias = flow_net_res_blocks_3_mlp_0_bias_to_fp16, weight = flow_net_res_blocks_3_mlp_0_weight_to_fp16, x = input_41_cast_fp16)[name = tensor("linear_16_cast_fp16")]; + tensor input_45_cast_fp16 = silu(x = linear_16_cast_fp16)[name = tensor("input_45_cast_fp16")]; + tensor flow_net_res_blocks_3_mlp_2_weight_to_fp16 = const()[name = tensor("flow_net_res_blocks_3_mlp_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(12129984)))]; + tensor flow_net_res_blocks_3_mlp_2_bias_to_fp16 = const()[name = tensor("flow_net_res_blocks_3_mlp_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(12654336)))]; + tensor linear_17_cast_fp16 = linear(bias = flow_net_res_blocks_3_mlp_2_bias_to_fp16, weight = flow_net_res_blocks_3_mlp_2_weight_to_fp16, x = input_45_cast_fp16)[name = tensor("linear_17_cast_fp16")]; + tensor var_288_cast_fp16 = mul(x = var_259_cast_fp16_2, y = linear_17_cast_fp16)[name = tensor("op_288_cast_fp16")]; + tensor x_29_cast_fp16 = add(x = x_23_cast_fp16, y = var_288_cast_fp16)[name = tensor("x_29_cast_fp16")]; + tensor flow_net_res_blocks_4_adaLN_modulation_1_weight_to_fp16 = const()[name = tensor("flow_net_res_blocks_4_adaLN_modulation_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(12655424)))]; + tensor flow_net_res_blocks_4_adaLN_modulation_1_bias_to_fp16 = const()[name = tensor("flow_net_res_blocks_4_adaLN_modulation_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(14228352)))]; + tensor linear_18_cast_fp16 = linear(bias = flow_net_res_blocks_4_adaLN_modulation_1_bias_to_fp16, weight = flow_net_res_blocks_4_adaLN_modulation_1_weight_to_fp16, x = input_15_cast_fp16)[name = tensor("linear_18_cast_fp16")]; + tensor var_298_split_sizes_0 = const()[name = tensor("op_298_split_sizes_0"), val = tensor([512, 512, 512])]; + tensor var_298_axis_0 = const()[name = tensor("op_298_axis_0"), val = tensor(-1)]; + tensor var_298_cast_fp16_0, tensor var_298_cast_fp16_1, tensor var_298_cast_fp16_2 = split(axis = var_298_axis_0, split_sizes = var_298_split_sizes_0, x = linear_18_cast_fp16)[name = tensor("op_298_cast_fp16")]; + tensor mean_9_axes_0 = const()[name = tensor("mean_9_axes_0"), val = tensor([-1])]; + tensor mean_9_keep_dims_0 = const()[name = tensor("mean_9_keep_dims_0"), val = tensor(true)]; + tensor mean_9_cast_fp16 = reduce_mean(axes = mean_9_axes_0, keep_dims = mean_9_keep_dims_0, x = x_29_cast_fp16)[name = tensor("mean_9_cast_fp16")]; + tensor sub_8_cast_fp16 = sub(x = x_29_cast_fp16, y = mean_9_cast_fp16)[name = tensor("sub_8_cast_fp16")]; + tensor square_6_cast_fp16 = square(x = sub_8_cast_fp16)[name = tensor("square_6_cast_fp16")]; + tensor reduce_mean_13_axes_0 = const()[name = tensor("reduce_mean_13_axes_0"), val = tensor([-1])]; + tensor reduce_mean_13_keep_dims_0 = const()[name = tensor("reduce_mean_13_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_13_cast_fp16 = reduce_mean(axes = reduce_mean_13_axes_0, keep_dims = reduce_mean_13_keep_dims_0, x = square_6_cast_fp16)[name = tensor("reduce_mean_13_cast_fp16")]; + tensor var_308_to_fp16 = const()[name = tensor("op_308_to_fp16"), val = tensor(0x1.1p-20)]; + tensor var_309_cast_fp16 = add(x = reduce_mean_13_cast_fp16, y = var_308_to_fp16)[name = tensor("op_309_cast_fp16")]; + tensor var_310_cast_fp16 = sqrt(x = var_309_cast_fp16)[name = tensor("op_310_cast_fp16")]; + tensor x_31_cast_fp16 = real_div(x = sub_8_cast_fp16, y = var_310_cast_fp16)[name = tensor("x_31_cast_fp16")]; + tensor flow_net_res_blocks_4_in_ln_weight_to_fp16 = const()[name = tensor("flow_net_res_blocks_4_in_ln_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(14231488)))]; + tensor var_312_cast_fp16 = mul(x = x_31_cast_fp16, y = flow_net_res_blocks_4_in_ln_weight_to_fp16)[name = tensor("op_312_cast_fp16")]; + tensor flow_net_res_blocks_4_in_ln_bias_to_fp16 = const()[name = tensor("flow_net_res_blocks_4_in_ln_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(14232576)))]; + tensor x_33_cast_fp16 = add(x = var_312_cast_fp16, y = flow_net_res_blocks_4_in_ln_bias_to_fp16)[name = tensor("x_33_cast_fp16")]; + tensor var_314_promoted_to_fp16 = const()[name = tensor("op_314_promoted_to_fp16"), val = tensor(0x1p+0)]; + tensor var_315_cast_fp16 = add(x = var_298_cast_fp16_1, y = var_314_promoted_to_fp16)[name = tensor("op_315_cast_fp16")]; + tensor var_316_cast_fp16 = mul(x = x_33_cast_fp16, y = var_315_cast_fp16)[name = tensor("op_316_cast_fp16")]; + tensor input_49_cast_fp16 = add(x = var_316_cast_fp16, y = var_298_cast_fp16_0)[name = tensor("input_49_cast_fp16")]; + tensor flow_net_res_blocks_4_mlp_0_weight_to_fp16 = const()[name = tensor("flow_net_res_blocks_4_mlp_0_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(14233664)))]; + tensor flow_net_res_blocks_4_mlp_0_bias_to_fp16 = const()[name = tensor("flow_net_res_blocks_4_mlp_0_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(14758016)))]; + tensor linear_19_cast_fp16 = linear(bias = flow_net_res_blocks_4_mlp_0_bias_to_fp16, weight = flow_net_res_blocks_4_mlp_0_weight_to_fp16, x = input_49_cast_fp16)[name = tensor("linear_19_cast_fp16")]; + tensor input_53_cast_fp16 = silu(x = linear_19_cast_fp16)[name = tensor("input_53_cast_fp16")]; + tensor flow_net_res_blocks_4_mlp_2_weight_to_fp16 = const()[name = tensor("flow_net_res_blocks_4_mlp_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(14759104)))]; + tensor flow_net_res_blocks_4_mlp_2_bias_to_fp16 = const()[name = tensor("flow_net_res_blocks_4_mlp_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15283456)))]; + tensor linear_20_cast_fp16 = linear(bias = flow_net_res_blocks_4_mlp_2_bias_to_fp16, weight = flow_net_res_blocks_4_mlp_2_weight_to_fp16, x = input_53_cast_fp16)[name = tensor("linear_20_cast_fp16")]; + tensor var_327_cast_fp16 = mul(x = var_298_cast_fp16_2, y = linear_20_cast_fp16)[name = tensor("op_327_cast_fp16")]; + tensor x_35_cast_fp16 = add(x = x_29_cast_fp16, y = var_327_cast_fp16)[name = tensor("x_35_cast_fp16")]; + tensor flow_net_res_blocks_5_adaLN_modulation_1_weight_to_fp16 = const()[name = tensor("flow_net_res_blocks_5_adaLN_modulation_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15284544)))]; + tensor flow_net_res_blocks_5_adaLN_modulation_1_bias_to_fp16 = const()[name = tensor("flow_net_res_blocks_5_adaLN_modulation_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(16857472)))]; + tensor linear_21_cast_fp16 = linear(bias = flow_net_res_blocks_5_adaLN_modulation_1_bias_to_fp16, weight = flow_net_res_blocks_5_adaLN_modulation_1_weight_to_fp16, x = input_15_cast_fp16)[name = tensor("linear_21_cast_fp16")]; + tensor var_337_split_sizes_0 = const()[name = tensor("op_337_split_sizes_0"), val = tensor([512, 512, 512])]; + tensor var_337_axis_0 = const()[name = tensor("op_337_axis_0"), val = tensor(-1)]; + tensor var_337_cast_fp16_0, tensor var_337_cast_fp16_1, tensor var_337_cast_fp16_2 = split(axis = var_337_axis_0, split_sizes = var_337_split_sizes_0, x = linear_21_cast_fp16)[name = tensor("op_337_cast_fp16")]; + tensor mean_11_axes_0 = const()[name = tensor("mean_11_axes_0"), val = tensor([-1])]; + tensor mean_11_keep_dims_0 = const()[name = tensor("mean_11_keep_dims_0"), val = tensor(true)]; + tensor mean_11_cast_fp16 = reduce_mean(axes = mean_11_axes_0, keep_dims = mean_11_keep_dims_0, x = x_35_cast_fp16)[name = tensor("mean_11_cast_fp16")]; + tensor sub_9_cast_fp16 = sub(x = x_35_cast_fp16, y = mean_11_cast_fp16)[name = tensor("sub_9_cast_fp16")]; + tensor square_7_cast_fp16 = square(x = sub_9_cast_fp16)[name = tensor("square_7_cast_fp16")]; + tensor reduce_mean_15_axes_0 = const()[name = tensor("reduce_mean_15_axes_0"), val = tensor([-1])]; + tensor reduce_mean_15_keep_dims_0 = const()[name = tensor("reduce_mean_15_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_15_cast_fp16 = reduce_mean(axes = reduce_mean_15_axes_0, keep_dims = reduce_mean_15_keep_dims_0, x = square_7_cast_fp16)[name = tensor("reduce_mean_15_cast_fp16")]; + tensor var_347_to_fp16 = const()[name = tensor("op_347_to_fp16"), val = tensor(0x1.1p-20)]; + tensor var_348_cast_fp16 = add(x = reduce_mean_15_cast_fp16, y = var_347_to_fp16)[name = tensor("op_348_cast_fp16")]; + tensor var_349_cast_fp16 = sqrt(x = var_348_cast_fp16)[name = tensor("op_349_cast_fp16")]; + tensor x_37_cast_fp16 = real_div(x = sub_9_cast_fp16, y = var_349_cast_fp16)[name = tensor("x_37_cast_fp16")]; + tensor flow_net_res_blocks_5_in_ln_weight_to_fp16 = const()[name = tensor("flow_net_res_blocks_5_in_ln_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(16860608)))]; + tensor var_351_cast_fp16 = mul(x = x_37_cast_fp16, y = flow_net_res_blocks_5_in_ln_weight_to_fp16)[name = tensor("op_351_cast_fp16")]; + tensor flow_net_res_blocks_5_in_ln_bias_to_fp16 = const()[name = tensor("flow_net_res_blocks_5_in_ln_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(16861696)))]; + tensor x_39_cast_fp16 = add(x = var_351_cast_fp16, y = flow_net_res_blocks_5_in_ln_bias_to_fp16)[name = tensor("x_39_cast_fp16")]; + tensor var_353_promoted_to_fp16 = const()[name = tensor("op_353_promoted_to_fp16"), val = tensor(0x1p+0)]; + tensor var_354_cast_fp16 = add(x = var_337_cast_fp16_1, y = var_353_promoted_to_fp16)[name = tensor("op_354_cast_fp16")]; + tensor var_355_cast_fp16 = mul(x = x_39_cast_fp16, y = var_354_cast_fp16)[name = tensor("op_355_cast_fp16")]; + tensor input_57_cast_fp16 = add(x = var_355_cast_fp16, y = var_337_cast_fp16_0)[name = tensor("input_57_cast_fp16")]; + tensor flow_net_res_blocks_5_mlp_0_weight_to_fp16 = const()[name = tensor("flow_net_res_blocks_5_mlp_0_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(16862784)))]; + tensor flow_net_res_blocks_5_mlp_0_bias_to_fp16 = const()[name = tensor("flow_net_res_blocks_5_mlp_0_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(17387136)))]; + tensor linear_22_cast_fp16 = linear(bias = flow_net_res_blocks_5_mlp_0_bias_to_fp16, weight = flow_net_res_blocks_5_mlp_0_weight_to_fp16, x = input_57_cast_fp16)[name = tensor("linear_22_cast_fp16")]; + tensor input_61_cast_fp16 = silu(x = linear_22_cast_fp16)[name = tensor("input_61_cast_fp16")]; + tensor flow_net_res_blocks_5_mlp_2_weight_to_fp16 = const()[name = tensor("flow_net_res_blocks_5_mlp_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(17388224)))]; + tensor flow_net_res_blocks_5_mlp_2_bias_to_fp16 = const()[name = tensor("flow_net_res_blocks_5_mlp_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(17912576)))]; + tensor linear_23_cast_fp16 = linear(bias = flow_net_res_blocks_5_mlp_2_bias_to_fp16, weight = flow_net_res_blocks_5_mlp_2_weight_to_fp16, x = input_61_cast_fp16)[name = tensor("linear_23_cast_fp16")]; + tensor var_366_cast_fp16 = mul(x = var_337_cast_fp16_2, y = linear_23_cast_fp16)[name = tensor("op_366_cast_fp16")]; + tensor x_41_cast_fp16 = add(x = x_35_cast_fp16, y = var_366_cast_fp16)[name = tensor("x_41_cast_fp16")]; + tensor flow_net_final_layer_adaLN_modulation_1_weight_to_fp16 = const()[name = tensor("flow_net_final_layer_adaLN_modulation_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(17913664)))]; + tensor flow_net_final_layer_adaLN_modulation_1_bias_to_fp16 = const()[name = tensor("flow_net_final_layer_adaLN_modulation_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(18962304)))]; + tensor linear_24_cast_fp16 = linear(bias = flow_net_final_layer_adaLN_modulation_1_bias_to_fp16, weight = flow_net_final_layer_adaLN_modulation_1_weight_to_fp16, x = input_15_cast_fp16)[name = tensor("linear_24_cast_fp16")]; + tensor var_375_split_sizes_0 = const()[name = tensor("op_375_split_sizes_0"), val = tensor([512, 512])]; + tensor var_375_axis_0 = const()[name = tensor("op_375_axis_0"), val = tensor(-1)]; + tensor var_375_cast_fp16_0, tensor var_375_cast_fp16_1 = split(axis = var_375_axis_0, split_sizes = var_375_split_sizes_0, x = linear_24_cast_fp16)[name = tensor("op_375_cast_fp16")]; + tensor mean_13_axes_0 = const()[name = tensor("mean_13_axes_0"), val = tensor([-1])]; + tensor mean_13_keep_dims_0 = const()[name = tensor("mean_13_keep_dims_0"), val = tensor(true)]; + tensor mean_13_cast_fp16 = reduce_mean(axes = mean_13_axes_0, keep_dims = mean_13_keep_dims_0, x = x_41_cast_fp16)[name = tensor("mean_13_cast_fp16")]; + tensor sub_10_cast_fp16 = sub(x = x_41_cast_fp16, y = mean_13_cast_fp16)[name = tensor("sub_10_cast_fp16")]; + tensor square_8_cast_fp16 = square(x = sub_10_cast_fp16)[name = tensor("square_8_cast_fp16")]; + tensor reduce_mean_17_axes_0 = const()[name = tensor("reduce_mean_17_axes_0"), val = tensor([-1])]; + tensor reduce_mean_17_keep_dims_0 = const()[name = tensor("reduce_mean_17_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_17_cast_fp16 = reduce_mean(axes = reduce_mean_17_axes_0, keep_dims = reduce_mean_17_keep_dims_0, x = square_8_cast_fp16)[name = tensor("reduce_mean_17_cast_fp16")]; + tensor var_382_to_fp16 = const()[name = tensor("op_382_to_fp16"), val = tensor(0x1.1p-20)]; + tensor var_383_cast_fp16 = add(x = reduce_mean_17_cast_fp16, y = var_382_to_fp16)[name = tensor("op_383_cast_fp16")]; + tensor var_384_cast_fp16 = sqrt(x = var_383_cast_fp16)[name = tensor("op_384_cast_fp16")]; + tensor x_43_cast_fp16 = real_div(x = sub_10_cast_fp16, y = var_384_cast_fp16)[name = tensor("x_43_cast_fp16")]; + tensor var_386_promoted_to_fp16 = const()[name = tensor("op_386_promoted_to_fp16"), val = tensor(0x1p+0)]; + tensor var_387_cast_fp16 = add(x = var_375_cast_fp16_1, y = var_386_promoted_to_fp16)[name = tensor("op_387_cast_fp16")]; + tensor var_388_cast_fp16 = mul(x = x_43_cast_fp16, y = var_387_cast_fp16)[name = tensor("op_388_cast_fp16")]; + tensor input_65_cast_fp16 = add(x = var_388_cast_fp16, y = var_375_cast_fp16_0)[name = tensor("input_65_cast_fp16")]; + tensor flow_net_final_layer_linear_weight_to_fp16 = const()[name = tensor("flow_net_final_layer_linear_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(18964416)))]; + tensor flow_net_final_layer_linear_bias_to_fp16 = const()[name = tensor("flow_net_final_layer_linear_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(18997248)))]; + tensor linear_25_cast_fp16 = linear(bias = flow_net_final_layer_linear_bias_to_fp16, weight = flow_net_final_layer_linear_weight_to_fp16, x = input_65_cast_fp16)[name = tensor("linear_25_cast_fp16")]; + tensor var_399_to_fp16 = const()[name = tensor("op_399_to_fp16"), val = tensor(0x1p-3)]; + tensor var_400_cast_fp16 = mul(x = linear_25_cast_fp16, y = var_399_to_fp16)[name = tensor("op_400_cast_fp16")]; + tensor input_67_cast_fp16 = add(x = latent_init_to_fp16, y = var_400_cast_fp16)[name = tensor("input_67_cast_fp16")]; + tensor linear_26_cast_fp16 = linear(bias = flow_net_input_proj_bias_to_fp16, weight = flow_net_input_proj_weight_to_fp16, x = input_67_cast_fp16)[name = tensor("linear_26_cast_fp16")]; + tensor input_71_to_fp16 = const()[name = tensor("input_71_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(18997376)))]; + tensor input_73_cast_fp16 = silu(x = input_71_to_fp16)[name = tensor("input_73_cast_fp16")]; + tensor linear_28_cast_fp16 = linear(bias = flow_net_time_embed_0_mlp_2_bias_to_fp16, weight = flow_net_time_embed_0_mlp_2_weight_to_fp16, x = input_73_cast_fp16)[name = tensor("linear_28_cast_fp16")]; + tensor reduce_mean_18_axes_0 = const()[name = tensor("reduce_mean_18_axes_0"), val = tensor([-1])]; + tensor reduce_mean_18_keep_dims_0 = const()[name = tensor("reduce_mean_18_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_18_cast_fp16 = reduce_mean(axes = reduce_mean_18_axes_0, keep_dims = reduce_mean_18_keep_dims_0, x = linear_28_cast_fp16)[name = tensor("reduce_mean_18_cast_fp16")]; + tensor sub_11_cast_fp16 = sub(x = linear_28_cast_fp16, y = reduce_mean_18_cast_fp16)[name = tensor("sub_11_cast_fp16")]; + tensor square_9_cast_fp16 = square(x = sub_11_cast_fp16)[name = tensor("square_9_cast_fp16")]; + tensor reduce_mean_19_axes_0 = const()[name = tensor("reduce_mean_19_axes_0"), val = tensor([-1])]; + tensor reduce_mean_19_keep_dims_0 = const()[name = tensor("reduce_mean_19_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_19_cast_fp16 = reduce_mean(axes = reduce_mean_19_axes_0, keep_dims = reduce_mean_19_keep_dims_0, x = square_9_cast_fp16)[name = tensor("reduce_mean_19_cast_fp16")]; + tensor real_div_2_to_fp16 = const()[name = tensor("real_div_2_to_fp16"), val = tensor(0x1.008p+0)]; + tensor mul_2_cast_fp16 = mul(x = reduce_mean_19_cast_fp16, y = real_div_2_to_fp16)[name = tensor("mul_2_cast_fp16")]; + tensor var_466_to_fp16 = const()[name = tensor("op_466_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_19_cast_fp16 = add(x = mul_2_cast_fp16, y = var_466_to_fp16)[name = tensor("var_19_cast_fp16")]; + tensor var_469_epsilon_0 = const()[name = tensor("op_469_epsilon_0"), val = tensor(0x1.197998p-40)]; + tensor var_469_cast_fp16 = rsqrt(epsilon = var_469_epsilon_0, x = var_19_cast_fp16)[name = tensor("op_469_cast_fp16")]; + tensor var_470_cast_fp16 = mul(x = const_3_to_fp16, y = var_469_cast_fp16)[name = tensor("op_470_cast_fp16")]; + tensor var_471_cast_fp16 = mul(x = linear_28_cast_fp16, y = var_470_cast_fp16)[name = tensor("op_471_cast_fp16")]; + tensor input_77_to_fp16 = const()[name = tensor("input_77_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(18998464)))]; + tensor input_79_cast_fp16 = silu(x = input_77_to_fp16)[name = tensor("input_79_cast_fp16")]; + tensor linear_30_cast_fp16 = linear(bias = flow_net_time_embed_1_mlp_2_bias_to_fp16, weight = flow_net_time_embed_1_mlp_2_weight_to_fp16, x = input_79_cast_fp16)[name = tensor("linear_30_cast_fp16")]; + tensor reduce_mean_20_axes_0 = const()[name = tensor("reduce_mean_20_axes_0"), val = tensor([-1])]; + tensor reduce_mean_20_keep_dims_0 = const()[name = tensor("reduce_mean_20_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_20_cast_fp16 = reduce_mean(axes = reduce_mean_20_axes_0, keep_dims = reduce_mean_20_keep_dims_0, x = linear_30_cast_fp16)[name = tensor("reduce_mean_20_cast_fp16")]; + tensor sub_13_cast_fp16 = sub(x = linear_30_cast_fp16, y = reduce_mean_20_cast_fp16)[name = tensor("sub_13_cast_fp16")]; + tensor square_10_cast_fp16 = square(x = sub_13_cast_fp16)[name = tensor("square_10_cast_fp16")]; + tensor reduce_mean_21_axes_0 = const()[name = tensor("reduce_mean_21_axes_0"), val = tensor([-1])]; + tensor reduce_mean_21_keep_dims_0 = const()[name = tensor("reduce_mean_21_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_21_cast_fp16 = reduce_mean(axes = reduce_mean_21_axes_0, keep_dims = reduce_mean_21_keep_dims_0, x = square_10_cast_fp16)[name = tensor("reduce_mean_21_cast_fp16")]; + tensor real_div_3_to_fp16 = const()[name = tensor("real_div_3_to_fp16"), val = tensor(0x1.008p+0)]; + tensor mul_3_cast_fp16 = mul(x = reduce_mean_21_cast_fp16, y = real_div_3_to_fp16)[name = tensor("mul_3_cast_fp16")]; + tensor var_503_to_fp16 = const()[name = tensor("op_503_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_21_cast_fp16 = add(x = mul_3_cast_fp16, y = var_503_to_fp16)[name = tensor("var_21_cast_fp16")]; + tensor var_506_epsilon_0 = const()[name = tensor("op_506_epsilon_0"), val = tensor(0x1.197998p-40)]; + tensor var_506_cast_fp16 = rsqrt(epsilon = var_506_epsilon_0, x = var_21_cast_fp16)[name = tensor("op_506_cast_fp16")]; + tensor var_507_cast_fp16 = mul(x = const_5_to_fp16, y = var_506_cast_fp16)[name = tensor("op_507_cast_fp16")]; + tensor var_508_cast_fp16 = mul(x = linear_30_cast_fp16, y = var_507_cast_fp16)[name = tensor("op_508_cast_fp16")]; + tensor var_520_cast_fp16 = add(x = var_471_cast_fp16, y = var_508_cast_fp16)[name = tensor("op_520_cast_fp16")]; + tensor _inversed_t_combined_3_y_0_to_fp16 = const()[name = tensor("_inversed_t_combined_3_y_0_to_fp16"), val = tensor(0x1p-1)]; + tensor _inversed_t_combined_3_cast_fp16 = mul(x = var_520_cast_fp16, y = _inversed_t_combined_3_y_0_to_fp16)[name = tensor("_inversed_t_combined_3_cast_fp16")]; + tensor input_81_cast_fp16 = add(x = _inversed_t_combined_3_cast_fp16, y = linear_5_cast_fp16)[name = tensor("input_81_cast_fp16")]; + tensor input_83_cast_fp16 = silu(x = input_81_cast_fp16)[name = tensor("input_83_cast_fp16")]; + tensor linear_32_cast_fp16 = linear(bias = flow_net_res_blocks_0_adaLN_modulation_1_bias_to_fp16, weight = flow_net_res_blocks_0_adaLN_modulation_1_weight_to_fp16, x = input_83_cast_fp16)[name = tensor("linear_32_cast_fp16")]; + tensor var_535_split_sizes_0 = const()[name = tensor("op_535_split_sizes_0"), val = tensor([512, 512, 512])]; + tensor var_535_axis_0 = const()[name = tensor("op_535_axis_0"), val = tensor(-1)]; + tensor var_535_cast_fp16_0, tensor var_535_cast_fp16_1, tensor var_535_cast_fp16_2 = split(axis = var_535_axis_0, split_sizes = var_535_split_sizes_0, x = linear_32_cast_fp16)[name = tensor("op_535_cast_fp16")]; + tensor mean_15_axes_0 = const()[name = tensor("mean_15_axes_0"), val = tensor([-1])]; + tensor mean_15_keep_dims_0 = const()[name = tensor("mean_15_keep_dims_0"), val = tensor(true)]; + tensor mean_15_cast_fp16 = reduce_mean(axes = mean_15_axes_0, keep_dims = mean_15_keep_dims_0, x = linear_26_cast_fp16)[name = tensor("mean_15_cast_fp16")]; + tensor sub_15_cast_fp16 = sub(x = linear_26_cast_fp16, y = mean_15_cast_fp16)[name = tensor("sub_15_cast_fp16")]; + tensor square_11_cast_fp16 = square(x = sub_15_cast_fp16)[name = tensor("square_11_cast_fp16")]; + tensor reduce_mean_23_axes_0 = const()[name = tensor("reduce_mean_23_axes_0"), val = tensor([-1])]; + tensor reduce_mean_23_keep_dims_0 = const()[name = tensor("reduce_mean_23_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_23_cast_fp16 = reduce_mean(axes = reduce_mean_23_axes_0, keep_dims = reduce_mean_23_keep_dims_0, x = square_11_cast_fp16)[name = tensor("reduce_mean_23_cast_fp16")]; + tensor var_545_to_fp16 = const()[name = tensor("op_545_to_fp16"), val = tensor(0x1.1p-20)]; + tensor var_546_cast_fp16 = add(x = reduce_mean_23_cast_fp16, y = var_545_to_fp16)[name = tensor("op_546_cast_fp16")]; + tensor var_547_cast_fp16 = sqrt(x = var_546_cast_fp16)[name = tensor("op_547_cast_fp16")]; + tensor x_51_cast_fp16 = real_div(x = sub_15_cast_fp16, y = var_547_cast_fp16)[name = tensor("x_51_cast_fp16")]; + tensor var_549_cast_fp16 = mul(x = x_51_cast_fp16, y = flow_net_res_blocks_0_in_ln_weight_to_fp16)[name = tensor("op_549_cast_fp16")]; + tensor x_53_cast_fp16 = add(x = var_549_cast_fp16, y = flow_net_res_blocks_0_in_ln_bias_to_fp16)[name = tensor("x_53_cast_fp16")]; + tensor var_551_promoted_to_fp16 = const()[name = tensor("op_551_promoted_to_fp16"), val = tensor(0x1p+0)]; + tensor var_552_cast_fp16 = add(x = var_535_cast_fp16_1, y = var_551_promoted_to_fp16)[name = tensor("op_552_cast_fp16")]; + tensor var_553_cast_fp16 = mul(x = x_53_cast_fp16, y = var_552_cast_fp16)[name = tensor("op_553_cast_fp16")]; + tensor input_85_cast_fp16 = add(x = var_553_cast_fp16, y = var_535_cast_fp16_0)[name = tensor("input_85_cast_fp16")]; + tensor linear_33_cast_fp16 = linear(bias = flow_net_res_blocks_0_mlp_0_bias_to_fp16, weight = flow_net_res_blocks_0_mlp_0_weight_to_fp16, x = input_85_cast_fp16)[name = tensor("linear_33_cast_fp16")]; + tensor input_89_cast_fp16 = silu(x = linear_33_cast_fp16)[name = tensor("input_89_cast_fp16")]; + tensor linear_34_cast_fp16 = linear(bias = flow_net_res_blocks_0_mlp_2_bias_to_fp16, weight = flow_net_res_blocks_0_mlp_2_weight_to_fp16, x = input_89_cast_fp16)[name = tensor("linear_34_cast_fp16")]; + tensor var_564_cast_fp16 = mul(x = var_535_cast_fp16_2, y = linear_34_cast_fp16)[name = tensor("op_564_cast_fp16")]; + tensor x_55_cast_fp16 = add(x = linear_26_cast_fp16, y = var_564_cast_fp16)[name = tensor("x_55_cast_fp16")]; + tensor linear_35_cast_fp16 = linear(bias = flow_net_res_blocks_1_adaLN_modulation_1_bias_to_fp16, weight = flow_net_res_blocks_1_adaLN_modulation_1_weight_to_fp16, x = input_83_cast_fp16)[name = tensor("linear_35_cast_fp16")]; + tensor var_574_split_sizes_0 = const()[name = tensor("op_574_split_sizes_0"), val = tensor([512, 512, 512])]; + tensor var_574_axis_0 = const()[name = tensor("op_574_axis_0"), val = tensor(-1)]; + tensor var_574_cast_fp16_0, tensor var_574_cast_fp16_1, tensor var_574_cast_fp16_2 = split(axis = var_574_axis_0, split_sizes = var_574_split_sizes_0, x = linear_35_cast_fp16)[name = tensor("op_574_cast_fp16")]; + tensor mean_17_axes_0 = const()[name = tensor("mean_17_axes_0"), val = tensor([-1])]; + tensor mean_17_keep_dims_0 = const()[name = tensor("mean_17_keep_dims_0"), val = tensor(true)]; + tensor mean_17_cast_fp16 = reduce_mean(axes = mean_17_axes_0, keep_dims = mean_17_keep_dims_0, x = x_55_cast_fp16)[name = tensor("mean_17_cast_fp16")]; + tensor sub_16_cast_fp16 = sub(x = x_55_cast_fp16, y = mean_17_cast_fp16)[name = tensor("sub_16_cast_fp16")]; + tensor square_12_cast_fp16 = square(x = sub_16_cast_fp16)[name = tensor("square_12_cast_fp16")]; + tensor reduce_mean_25_axes_0 = const()[name = tensor("reduce_mean_25_axes_0"), val = tensor([-1])]; + tensor reduce_mean_25_keep_dims_0 = const()[name = tensor("reduce_mean_25_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_25_cast_fp16 = reduce_mean(axes = reduce_mean_25_axes_0, keep_dims = reduce_mean_25_keep_dims_0, x = square_12_cast_fp16)[name = tensor("reduce_mean_25_cast_fp16")]; + tensor var_584_to_fp16 = const()[name = tensor("op_584_to_fp16"), val = tensor(0x1.1p-20)]; + tensor var_585_cast_fp16 = add(x = reduce_mean_25_cast_fp16, y = var_584_to_fp16)[name = tensor("op_585_cast_fp16")]; + tensor var_586_cast_fp16 = sqrt(x = var_585_cast_fp16)[name = tensor("op_586_cast_fp16")]; + tensor x_57_cast_fp16 = real_div(x = sub_16_cast_fp16, y = var_586_cast_fp16)[name = tensor("x_57_cast_fp16")]; + tensor var_588_cast_fp16 = mul(x = x_57_cast_fp16, y = flow_net_res_blocks_1_in_ln_weight_to_fp16)[name = tensor("op_588_cast_fp16")]; + tensor x_59_cast_fp16 = add(x = var_588_cast_fp16, y = flow_net_res_blocks_1_in_ln_bias_to_fp16)[name = tensor("x_59_cast_fp16")]; + tensor var_590_promoted_to_fp16 = const()[name = tensor("op_590_promoted_to_fp16"), val = tensor(0x1p+0)]; + tensor var_591_cast_fp16 = add(x = var_574_cast_fp16_1, y = var_590_promoted_to_fp16)[name = tensor("op_591_cast_fp16")]; + tensor var_592_cast_fp16 = mul(x = x_59_cast_fp16, y = var_591_cast_fp16)[name = tensor("op_592_cast_fp16")]; + tensor input_93_cast_fp16 = add(x = var_592_cast_fp16, y = var_574_cast_fp16_0)[name = tensor("input_93_cast_fp16")]; + tensor linear_36_cast_fp16 = linear(bias = flow_net_res_blocks_1_mlp_0_bias_to_fp16, weight = flow_net_res_blocks_1_mlp_0_weight_to_fp16, x = input_93_cast_fp16)[name = tensor("linear_36_cast_fp16")]; + tensor input_97_cast_fp16 = silu(x = linear_36_cast_fp16)[name = tensor("input_97_cast_fp16")]; + tensor linear_37_cast_fp16 = linear(bias = flow_net_res_blocks_1_mlp_2_bias_to_fp16, weight = flow_net_res_blocks_1_mlp_2_weight_to_fp16, x = input_97_cast_fp16)[name = tensor("linear_37_cast_fp16")]; + tensor var_603_cast_fp16 = mul(x = var_574_cast_fp16_2, y = linear_37_cast_fp16)[name = tensor("op_603_cast_fp16")]; + tensor x_61_cast_fp16 = add(x = x_55_cast_fp16, y = var_603_cast_fp16)[name = tensor("x_61_cast_fp16")]; + tensor linear_38_cast_fp16 = linear(bias = flow_net_res_blocks_2_adaLN_modulation_1_bias_to_fp16, weight = flow_net_res_blocks_2_adaLN_modulation_1_weight_to_fp16, x = input_83_cast_fp16)[name = tensor("linear_38_cast_fp16")]; + tensor var_613_split_sizes_0 = const()[name = tensor("op_613_split_sizes_0"), val = tensor([512, 512, 512])]; + tensor var_613_axis_0 = const()[name = tensor("op_613_axis_0"), val = tensor(-1)]; + tensor var_613_cast_fp16_0, tensor var_613_cast_fp16_1, tensor var_613_cast_fp16_2 = split(axis = var_613_axis_0, split_sizes = var_613_split_sizes_0, x = linear_38_cast_fp16)[name = tensor("op_613_cast_fp16")]; + tensor mean_19_axes_0 = const()[name = tensor("mean_19_axes_0"), val = tensor([-1])]; + tensor mean_19_keep_dims_0 = const()[name = tensor("mean_19_keep_dims_0"), val = tensor(true)]; + tensor mean_19_cast_fp16 = reduce_mean(axes = mean_19_axes_0, keep_dims = mean_19_keep_dims_0, x = x_61_cast_fp16)[name = tensor("mean_19_cast_fp16")]; + tensor sub_17_cast_fp16 = sub(x = x_61_cast_fp16, y = mean_19_cast_fp16)[name = tensor("sub_17_cast_fp16")]; + tensor square_13_cast_fp16 = square(x = sub_17_cast_fp16)[name = tensor("square_13_cast_fp16")]; + tensor reduce_mean_27_axes_0 = const()[name = tensor("reduce_mean_27_axes_0"), val = tensor([-1])]; + tensor reduce_mean_27_keep_dims_0 = const()[name = tensor("reduce_mean_27_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_27_cast_fp16 = reduce_mean(axes = reduce_mean_27_axes_0, keep_dims = reduce_mean_27_keep_dims_0, x = square_13_cast_fp16)[name = tensor("reduce_mean_27_cast_fp16")]; + tensor var_623_to_fp16 = const()[name = tensor("op_623_to_fp16"), val = tensor(0x1.1p-20)]; + tensor var_624_cast_fp16 = add(x = reduce_mean_27_cast_fp16, y = var_623_to_fp16)[name = tensor("op_624_cast_fp16")]; + tensor var_625_cast_fp16 = sqrt(x = var_624_cast_fp16)[name = tensor("op_625_cast_fp16")]; + tensor x_63_cast_fp16 = real_div(x = sub_17_cast_fp16, y = var_625_cast_fp16)[name = tensor("x_63_cast_fp16")]; + tensor var_627_cast_fp16 = mul(x = x_63_cast_fp16, y = flow_net_res_blocks_2_in_ln_weight_to_fp16)[name = tensor("op_627_cast_fp16")]; + tensor x_65_cast_fp16 = add(x = var_627_cast_fp16, y = flow_net_res_blocks_2_in_ln_bias_to_fp16)[name = tensor("x_65_cast_fp16")]; + tensor var_629_promoted_to_fp16 = const()[name = tensor("op_629_promoted_to_fp16"), val = tensor(0x1p+0)]; + tensor var_630_cast_fp16 = add(x = var_613_cast_fp16_1, y = var_629_promoted_to_fp16)[name = tensor("op_630_cast_fp16")]; + tensor var_631_cast_fp16 = mul(x = x_65_cast_fp16, y = var_630_cast_fp16)[name = tensor("op_631_cast_fp16")]; + tensor input_101_cast_fp16 = add(x = var_631_cast_fp16, y = var_613_cast_fp16_0)[name = tensor("input_101_cast_fp16")]; + tensor linear_39_cast_fp16 = linear(bias = flow_net_res_blocks_2_mlp_0_bias_to_fp16, weight = flow_net_res_blocks_2_mlp_0_weight_to_fp16, x = input_101_cast_fp16)[name = tensor("linear_39_cast_fp16")]; + tensor input_105_cast_fp16 = silu(x = linear_39_cast_fp16)[name = tensor("input_105_cast_fp16")]; + tensor linear_40_cast_fp16 = linear(bias = flow_net_res_blocks_2_mlp_2_bias_to_fp16, weight = flow_net_res_blocks_2_mlp_2_weight_to_fp16, x = input_105_cast_fp16)[name = tensor("linear_40_cast_fp16")]; + tensor var_642_cast_fp16 = mul(x = var_613_cast_fp16_2, y = linear_40_cast_fp16)[name = tensor("op_642_cast_fp16")]; + tensor x_67_cast_fp16 = add(x = x_61_cast_fp16, y = var_642_cast_fp16)[name = tensor("x_67_cast_fp16")]; + tensor linear_41_cast_fp16 = linear(bias = flow_net_res_blocks_3_adaLN_modulation_1_bias_to_fp16, weight = flow_net_res_blocks_3_adaLN_modulation_1_weight_to_fp16, x = input_83_cast_fp16)[name = tensor("linear_41_cast_fp16")]; + tensor var_652_split_sizes_0 = const()[name = tensor("op_652_split_sizes_0"), val = tensor([512, 512, 512])]; + tensor var_652_axis_0 = const()[name = tensor("op_652_axis_0"), val = tensor(-1)]; + tensor var_652_cast_fp16_0, tensor var_652_cast_fp16_1, tensor var_652_cast_fp16_2 = split(axis = var_652_axis_0, split_sizes = var_652_split_sizes_0, x = linear_41_cast_fp16)[name = tensor("op_652_cast_fp16")]; + tensor mean_21_axes_0 = const()[name = tensor("mean_21_axes_0"), val = tensor([-1])]; + tensor mean_21_keep_dims_0 = const()[name = tensor("mean_21_keep_dims_0"), val = tensor(true)]; + tensor mean_21_cast_fp16 = reduce_mean(axes = mean_21_axes_0, keep_dims = mean_21_keep_dims_0, x = x_67_cast_fp16)[name = tensor("mean_21_cast_fp16")]; + tensor sub_18_cast_fp16 = sub(x = x_67_cast_fp16, y = mean_21_cast_fp16)[name = tensor("sub_18_cast_fp16")]; + tensor square_14_cast_fp16 = square(x = sub_18_cast_fp16)[name = tensor("square_14_cast_fp16")]; + tensor reduce_mean_29_axes_0 = const()[name = tensor("reduce_mean_29_axes_0"), val = tensor([-1])]; + tensor reduce_mean_29_keep_dims_0 = const()[name = tensor("reduce_mean_29_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_29_cast_fp16 = reduce_mean(axes = reduce_mean_29_axes_0, keep_dims = reduce_mean_29_keep_dims_0, x = square_14_cast_fp16)[name = tensor("reduce_mean_29_cast_fp16")]; + tensor var_662_to_fp16 = const()[name = tensor("op_662_to_fp16"), val = tensor(0x1.1p-20)]; + tensor var_663_cast_fp16 = add(x = reduce_mean_29_cast_fp16, y = var_662_to_fp16)[name = tensor("op_663_cast_fp16")]; + tensor var_664_cast_fp16 = sqrt(x = var_663_cast_fp16)[name = tensor("op_664_cast_fp16")]; + tensor x_69_cast_fp16 = real_div(x = sub_18_cast_fp16, y = var_664_cast_fp16)[name = tensor("x_69_cast_fp16")]; + tensor var_666_cast_fp16 = mul(x = x_69_cast_fp16, y = flow_net_res_blocks_3_in_ln_weight_to_fp16)[name = tensor("op_666_cast_fp16")]; + tensor x_71_cast_fp16 = add(x = var_666_cast_fp16, y = flow_net_res_blocks_3_in_ln_bias_to_fp16)[name = tensor("x_71_cast_fp16")]; + tensor var_668_promoted_to_fp16 = const()[name = tensor("op_668_promoted_to_fp16"), val = tensor(0x1p+0)]; + tensor var_669_cast_fp16 = add(x = var_652_cast_fp16_1, y = var_668_promoted_to_fp16)[name = tensor("op_669_cast_fp16")]; + tensor var_670_cast_fp16 = mul(x = x_71_cast_fp16, y = var_669_cast_fp16)[name = tensor("op_670_cast_fp16")]; + tensor input_109_cast_fp16 = add(x = var_670_cast_fp16, y = var_652_cast_fp16_0)[name = tensor("input_109_cast_fp16")]; + tensor linear_42_cast_fp16 = linear(bias = flow_net_res_blocks_3_mlp_0_bias_to_fp16, weight = flow_net_res_blocks_3_mlp_0_weight_to_fp16, x = input_109_cast_fp16)[name = tensor("linear_42_cast_fp16")]; + tensor input_113_cast_fp16 = silu(x = linear_42_cast_fp16)[name = tensor("input_113_cast_fp16")]; + tensor linear_43_cast_fp16 = linear(bias = flow_net_res_blocks_3_mlp_2_bias_to_fp16, weight = flow_net_res_blocks_3_mlp_2_weight_to_fp16, x = input_113_cast_fp16)[name = tensor("linear_43_cast_fp16")]; + tensor var_681_cast_fp16 = mul(x = var_652_cast_fp16_2, y = linear_43_cast_fp16)[name = tensor("op_681_cast_fp16")]; + tensor x_73_cast_fp16 = add(x = x_67_cast_fp16, y = var_681_cast_fp16)[name = tensor("x_73_cast_fp16")]; + tensor linear_44_cast_fp16 = linear(bias = flow_net_res_blocks_4_adaLN_modulation_1_bias_to_fp16, weight = flow_net_res_blocks_4_adaLN_modulation_1_weight_to_fp16, x = input_83_cast_fp16)[name = tensor("linear_44_cast_fp16")]; + tensor var_691_split_sizes_0 = const()[name = tensor("op_691_split_sizes_0"), val = tensor([512, 512, 512])]; + tensor var_691_axis_0 = const()[name = tensor("op_691_axis_0"), val = tensor(-1)]; + tensor var_691_cast_fp16_0, tensor var_691_cast_fp16_1, tensor var_691_cast_fp16_2 = split(axis = var_691_axis_0, split_sizes = var_691_split_sizes_0, x = linear_44_cast_fp16)[name = tensor("op_691_cast_fp16")]; + tensor mean_23_axes_0 = const()[name = tensor("mean_23_axes_0"), val = tensor([-1])]; + tensor mean_23_keep_dims_0 = const()[name = tensor("mean_23_keep_dims_0"), val = tensor(true)]; + tensor mean_23_cast_fp16 = reduce_mean(axes = mean_23_axes_0, keep_dims = mean_23_keep_dims_0, x = x_73_cast_fp16)[name = tensor("mean_23_cast_fp16")]; + tensor sub_19_cast_fp16 = sub(x = x_73_cast_fp16, y = mean_23_cast_fp16)[name = tensor("sub_19_cast_fp16")]; + tensor square_15_cast_fp16 = square(x = sub_19_cast_fp16)[name = tensor("square_15_cast_fp16")]; + tensor reduce_mean_31_axes_0 = const()[name = tensor("reduce_mean_31_axes_0"), val = tensor([-1])]; + tensor reduce_mean_31_keep_dims_0 = const()[name = tensor("reduce_mean_31_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_31_cast_fp16 = reduce_mean(axes = reduce_mean_31_axes_0, keep_dims = reduce_mean_31_keep_dims_0, x = square_15_cast_fp16)[name = tensor("reduce_mean_31_cast_fp16")]; + tensor var_701_to_fp16 = const()[name = tensor("op_701_to_fp16"), val = tensor(0x1.1p-20)]; + tensor var_702_cast_fp16 = add(x = reduce_mean_31_cast_fp16, y = var_701_to_fp16)[name = tensor("op_702_cast_fp16")]; + tensor var_703_cast_fp16 = sqrt(x = var_702_cast_fp16)[name = tensor("op_703_cast_fp16")]; + tensor x_75_cast_fp16 = real_div(x = sub_19_cast_fp16, y = var_703_cast_fp16)[name = tensor("x_75_cast_fp16")]; + tensor var_705_cast_fp16 = mul(x = x_75_cast_fp16, y = flow_net_res_blocks_4_in_ln_weight_to_fp16)[name = tensor("op_705_cast_fp16")]; + tensor x_77_cast_fp16 = add(x = var_705_cast_fp16, y = flow_net_res_blocks_4_in_ln_bias_to_fp16)[name = tensor("x_77_cast_fp16")]; + tensor var_707_promoted_to_fp16 = const()[name = tensor("op_707_promoted_to_fp16"), val = tensor(0x1p+0)]; + tensor var_708_cast_fp16 = add(x = var_691_cast_fp16_1, y = var_707_promoted_to_fp16)[name = tensor("op_708_cast_fp16")]; + tensor var_709_cast_fp16 = mul(x = x_77_cast_fp16, y = var_708_cast_fp16)[name = tensor("op_709_cast_fp16")]; + tensor input_117_cast_fp16 = add(x = var_709_cast_fp16, y = var_691_cast_fp16_0)[name = tensor("input_117_cast_fp16")]; + tensor linear_45_cast_fp16 = linear(bias = flow_net_res_blocks_4_mlp_0_bias_to_fp16, weight = flow_net_res_blocks_4_mlp_0_weight_to_fp16, x = input_117_cast_fp16)[name = tensor("linear_45_cast_fp16")]; + tensor input_121_cast_fp16 = silu(x = linear_45_cast_fp16)[name = tensor("input_121_cast_fp16")]; + tensor linear_46_cast_fp16 = linear(bias = flow_net_res_blocks_4_mlp_2_bias_to_fp16, weight = flow_net_res_blocks_4_mlp_2_weight_to_fp16, x = input_121_cast_fp16)[name = tensor("linear_46_cast_fp16")]; + tensor var_720_cast_fp16 = mul(x = var_691_cast_fp16_2, y = linear_46_cast_fp16)[name = tensor("op_720_cast_fp16")]; + tensor x_79_cast_fp16 = add(x = x_73_cast_fp16, y = var_720_cast_fp16)[name = tensor("x_79_cast_fp16")]; + tensor linear_47_cast_fp16 = linear(bias = flow_net_res_blocks_5_adaLN_modulation_1_bias_to_fp16, weight = flow_net_res_blocks_5_adaLN_modulation_1_weight_to_fp16, x = input_83_cast_fp16)[name = tensor("linear_47_cast_fp16")]; + tensor var_730_split_sizes_0 = const()[name = tensor("op_730_split_sizes_0"), val = tensor([512, 512, 512])]; + tensor var_730_axis_0 = const()[name = tensor("op_730_axis_0"), val = tensor(-1)]; + tensor var_730_cast_fp16_0, tensor var_730_cast_fp16_1, tensor var_730_cast_fp16_2 = split(axis = var_730_axis_0, split_sizes = var_730_split_sizes_0, x = linear_47_cast_fp16)[name = tensor("op_730_cast_fp16")]; + tensor mean_25_axes_0 = const()[name = tensor("mean_25_axes_0"), val = tensor([-1])]; + tensor mean_25_keep_dims_0 = const()[name = tensor("mean_25_keep_dims_0"), val = tensor(true)]; + tensor mean_25_cast_fp16 = reduce_mean(axes = mean_25_axes_0, keep_dims = mean_25_keep_dims_0, x = x_79_cast_fp16)[name = tensor("mean_25_cast_fp16")]; + tensor sub_20_cast_fp16 = sub(x = x_79_cast_fp16, y = mean_25_cast_fp16)[name = tensor("sub_20_cast_fp16")]; + tensor square_16_cast_fp16 = square(x = sub_20_cast_fp16)[name = tensor("square_16_cast_fp16")]; + tensor reduce_mean_33_axes_0 = const()[name = tensor("reduce_mean_33_axes_0"), val = tensor([-1])]; + tensor reduce_mean_33_keep_dims_0 = const()[name = tensor("reduce_mean_33_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_33_cast_fp16 = reduce_mean(axes = reduce_mean_33_axes_0, keep_dims = reduce_mean_33_keep_dims_0, x = square_16_cast_fp16)[name = tensor("reduce_mean_33_cast_fp16")]; + tensor var_740_to_fp16 = const()[name = tensor("op_740_to_fp16"), val = tensor(0x1.1p-20)]; + tensor var_741_cast_fp16 = add(x = reduce_mean_33_cast_fp16, y = var_740_to_fp16)[name = tensor("op_741_cast_fp16")]; + tensor var_742_cast_fp16 = sqrt(x = var_741_cast_fp16)[name = tensor("op_742_cast_fp16")]; + tensor x_81_cast_fp16 = real_div(x = sub_20_cast_fp16, y = var_742_cast_fp16)[name = tensor("x_81_cast_fp16")]; + tensor var_744_cast_fp16 = mul(x = x_81_cast_fp16, y = flow_net_res_blocks_5_in_ln_weight_to_fp16)[name = tensor("op_744_cast_fp16")]; + tensor x_83_cast_fp16 = add(x = var_744_cast_fp16, y = flow_net_res_blocks_5_in_ln_bias_to_fp16)[name = tensor("x_83_cast_fp16")]; + tensor var_746_promoted_to_fp16 = const()[name = tensor("op_746_promoted_to_fp16"), val = tensor(0x1p+0)]; + tensor var_747_cast_fp16 = add(x = var_730_cast_fp16_1, y = var_746_promoted_to_fp16)[name = tensor("op_747_cast_fp16")]; + tensor var_748_cast_fp16 = mul(x = x_83_cast_fp16, y = var_747_cast_fp16)[name = tensor("op_748_cast_fp16")]; + tensor input_125_cast_fp16 = add(x = var_748_cast_fp16, y = var_730_cast_fp16_0)[name = tensor("input_125_cast_fp16")]; + tensor linear_48_cast_fp16 = linear(bias = flow_net_res_blocks_5_mlp_0_bias_to_fp16, weight = flow_net_res_blocks_5_mlp_0_weight_to_fp16, x = input_125_cast_fp16)[name = tensor("linear_48_cast_fp16")]; + tensor input_129_cast_fp16 = silu(x = linear_48_cast_fp16)[name = tensor("input_129_cast_fp16")]; + tensor linear_49_cast_fp16 = linear(bias = flow_net_res_blocks_5_mlp_2_bias_to_fp16, weight = flow_net_res_blocks_5_mlp_2_weight_to_fp16, x = input_129_cast_fp16)[name = tensor("linear_49_cast_fp16")]; + tensor var_759_cast_fp16 = mul(x = var_730_cast_fp16_2, y = linear_49_cast_fp16)[name = tensor("op_759_cast_fp16")]; + tensor x_85_cast_fp16 = add(x = x_79_cast_fp16, y = var_759_cast_fp16)[name = tensor("x_85_cast_fp16")]; + tensor linear_50_cast_fp16 = linear(bias = flow_net_final_layer_adaLN_modulation_1_bias_to_fp16, weight = flow_net_final_layer_adaLN_modulation_1_weight_to_fp16, x = input_83_cast_fp16)[name = tensor("linear_50_cast_fp16")]; + tensor var_768_split_sizes_0 = const()[name = tensor("op_768_split_sizes_0"), val = tensor([512, 512])]; + tensor var_768_axis_0 = const()[name = tensor("op_768_axis_0"), val = tensor(-1)]; + tensor var_768_cast_fp16_0, tensor var_768_cast_fp16_1 = split(axis = var_768_axis_0, split_sizes = var_768_split_sizes_0, x = linear_50_cast_fp16)[name = tensor("op_768_cast_fp16")]; + tensor mean_27_axes_0 = const()[name = tensor("mean_27_axes_0"), val = tensor([-1])]; + tensor mean_27_keep_dims_0 = const()[name = tensor("mean_27_keep_dims_0"), val = tensor(true)]; + tensor mean_27_cast_fp16 = reduce_mean(axes = mean_27_axes_0, keep_dims = mean_27_keep_dims_0, x = x_85_cast_fp16)[name = tensor("mean_27_cast_fp16")]; + tensor sub_21_cast_fp16 = sub(x = x_85_cast_fp16, y = mean_27_cast_fp16)[name = tensor("sub_21_cast_fp16")]; + tensor square_17_cast_fp16 = square(x = sub_21_cast_fp16)[name = tensor("square_17_cast_fp16")]; + tensor reduce_mean_35_axes_0 = const()[name = tensor("reduce_mean_35_axes_0"), val = tensor([-1])]; + tensor reduce_mean_35_keep_dims_0 = const()[name = tensor("reduce_mean_35_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_35_cast_fp16 = reduce_mean(axes = reduce_mean_35_axes_0, keep_dims = reduce_mean_35_keep_dims_0, x = square_17_cast_fp16)[name = tensor("reduce_mean_35_cast_fp16")]; + tensor var_775_to_fp16 = const()[name = tensor("op_775_to_fp16"), val = tensor(0x1.1p-20)]; + tensor var_776_cast_fp16 = add(x = reduce_mean_35_cast_fp16, y = var_775_to_fp16)[name = tensor("op_776_cast_fp16")]; + tensor var_777_cast_fp16 = sqrt(x = var_776_cast_fp16)[name = tensor("op_777_cast_fp16")]; + tensor x_87_cast_fp16 = real_div(x = sub_21_cast_fp16, y = var_777_cast_fp16)[name = tensor("x_87_cast_fp16")]; + tensor var_779_promoted_to_fp16 = const()[name = tensor("op_779_promoted_to_fp16"), val = tensor(0x1p+0)]; + tensor var_780_cast_fp16 = add(x = var_768_cast_fp16_1, y = var_779_promoted_to_fp16)[name = tensor("op_780_cast_fp16")]; + tensor var_781_cast_fp16 = mul(x = x_87_cast_fp16, y = var_780_cast_fp16)[name = tensor("op_781_cast_fp16")]; + tensor input_133_cast_fp16 = add(x = var_781_cast_fp16, y = var_768_cast_fp16_0)[name = tensor("input_133_cast_fp16")]; + tensor linear_51_cast_fp16 = linear(bias = flow_net_final_layer_linear_bias_to_fp16, weight = flow_net_final_layer_linear_weight_to_fp16, x = input_133_cast_fp16)[name = tensor("linear_51_cast_fp16")]; + tensor var_792_to_fp16 = const()[name = tensor("op_792_to_fp16"), val = tensor(0x1p-3)]; + tensor var_793_cast_fp16 = mul(x = linear_51_cast_fp16, y = var_792_to_fp16)[name = tensor("op_793_cast_fp16")]; + tensor input_135_cast_fp16 = add(x = input_67_cast_fp16, y = var_793_cast_fp16)[name = tensor("input_135_cast_fp16")]; + tensor linear_52_cast_fp16 = linear(bias = flow_net_input_proj_bias_to_fp16, weight = flow_net_input_proj_weight_to_fp16, x = input_135_cast_fp16)[name = tensor("linear_52_cast_fp16")]; + tensor input_139_to_fp16 = const()[name = tensor("input_139_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(18999552)))]; + tensor input_141_cast_fp16 = silu(x = input_139_to_fp16)[name = tensor("input_141_cast_fp16")]; + tensor linear_54_cast_fp16 = linear(bias = flow_net_time_embed_0_mlp_2_bias_to_fp16, weight = flow_net_time_embed_0_mlp_2_weight_to_fp16, x = input_141_cast_fp16)[name = tensor("linear_54_cast_fp16")]; + tensor reduce_mean_36_axes_0 = const()[name = tensor("reduce_mean_36_axes_0"), val = tensor([-1])]; + tensor reduce_mean_36_keep_dims_0 = const()[name = tensor("reduce_mean_36_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_36_cast_fp16 = reduce_mean(axes = reduce_mean_36_axes_0, keep_dims = reduce_mean_36_keep_dims_0, x = linear_54_cast_fp16)[name = tensor("reduce_mean_36_cast_fp16")]; + tensor sub_22_cast_fp16 = sub(x = linear_54_cast_fp16, y = reduce_mean_36_cast_fp16)[name = tensor("sub_22_cast_fp16")]; + tensor square_18_cast_fp16 = square(x = sub_22_cast_fp16)[name = tensor("square_18_cast_fp16")]; + tensor reduce_mean_37_axes_0 = const()[name = tensor("reduce_mean_37_axes_0"), val = tensor([-1])]; + tensor reduce_mean_37_keep_dims_0 = const()[name = tensor("reduce_mean_37_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_37_cast_fp16 = reduce_mean(axes = reduce_mean_37_axes_0, keep_dims = reduce_mean_37_keep_dims_0, x = square_18_cast_fp16)[name = tensor("reduce_mean_37_cast_fp16")]; + tensor real_div_4_to_fp16 = const()[name = tensor("real_div_4_to_fp16"), val = tensor(0x1.008p+0)]; + tensor mul_4_cast_fp16 = mul(x = reduce_mean_37_cast_fp16, y = real_div_4_to_fp16)[name = tensor("mul_4_cast_fp16")]; + tensor var_859_to_fp16 = const()[name = tensor("op_859_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_37_cast_fp16 = add(x = mul_4_cast_fp16, y = var_859_to_fp16)[name = tensor("var_37_cast_fp16")]; + tensor var_862_epsilon_0 = const()[name = tensor("op_862_epsilon_0"), val = tensor(0x1.197998p-40)]; + tensor var_862_cast_fp16 = rsqrt(epsilon = var_862_epsilon_0, x = var_37_cast_fp16)[name = tensor("op_862_cast_fp16")]; + tensor var_863_cast_fp16 = mul(x = const_3_to_fp16, y = var_862_cast_fp16)[name = tensor("op_863_cast_fp16")]; + tensor var_864_cast_fp16 = mul(x = linear_54_cast_fp16, y = var_863_cast_fp16)[name = tensor("op_864_cast_fp16")]; + tensor input_145_to_fp16 = const()[name = tensor("input_145_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(19000640)))]; + tensor input_147_cast_fp16 = silu(x = input_145_to_fp16)[name = tensor("input_147_cast_fp16")]; + tensor linear_56_cast_fp16 = linear(bias = flow_net_time_embed_1_mlp_2_bias_to_fp16, weight = flow_net_time_embed_1_mlp_2_weight_to_fp16, x = input_147_cast_fp16)[name = tensor("linear_56_cast_fp16")]; + tensor reduce_mean_38_axes_0 = const()[name = tensor("reduce_mean_38_axes_0"), val = tensor([-1])]; + tensor reduce_mean_38_keep_dims_0 = const()[name = tensor("reduce_mean_38_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_38_cast_fp16 = reduce_mean(axes = reduce_mean_38_axes_0, keep_dims = reduce_mean_38_keep_dims_0, x = linear_56_cast_fp16)[name = tensor("reduce_mean_38_cast_fp16")]; + tensor sub_24_cast_fp16 = sub(x = linear_56_cast_fp16, y = reduce_mean_38_cast_fp16)[name = tensor("sub_24_cast_fp16")]; + tensor square_19_cast_fp16 = square(x = sub_24_cast_fp16)[name = tensor("square_19_cast_fp16")]; + tensor reduce_mean_39_axes_0 = const()[name = tensor("reduce_mean_39_axes_0"), val = tensor([-1])]; + tensor reduce_mean_39_keep_dims_0 = const()[name = tensor("reduce_mean_39_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_39_cast_fp16 = reduce_mean(axes = reduce_mean_39_axes_0, keep_dims = reduce_mean_39_keep_dims_0, x = square_19_cast_fp16)[name = tensor("reduce_mean_39_cast_fp16")]; + tensor real_div_5_to_fp16 = const()[name = tensor("real_div_5_to_fp16"), val = tensor(0x1.008p+0)]; + tensor mul_5_cast_fp16 = mul(x = reduce_mean_39_cast_fp16, y = real_div_5_to_fp16)[name = tensor("mul_5_cast_fp16")]; + tensor var_896_to_fp16 = const()[name = tensor("op_896_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_39_cast_fp16 = add(x = mul_5_cast_fp16, y = var_896_to_fp16)[name = tensor("var_39_cast_fp16")]; + tensor var_899_epsilon_0 = const()[name = tensor("op_899_epsilon_0"), val = tensor(0x1.197998p-40)]; + tensor var_899_cast_fp16 = rsqrt(epsilon = var_899_epsilon_0, x = var_39_cast_fp16)[name = tensor("op_899_cast_fp16")]; + tensor var_900_cast_fp16 = mul(x = const_5_to_fp16, y = var_899_cast_fp16)[name = tensor("op_900_cast_fp16")]; + tensor var_901_cast_fp16 = mul(x = linear_56_cast_fp16, y = var_900_cast_fp16)[name = tensor("op_901_cast_fp16")]; + tensor var_913_cast_fp16 = add(x = var_864_cast_fp16, y = var_901_cast_fp16)[name = tensor("op_913_cast_fp16")]; + tensor _inversed_t_combined_5_y_0_to_fp16 = const()[name = tensor("_inversed_t_combined_5_y_0_to_fp16"), val = tensor(0x1p-1)]; + tensor _inversed_t_combined_5_cast_fp16 = mul(x = var_913_cast_fp16, y = _inversed_t_combined_5_y_0_to_fp16)[name = tensor("_inversed_t_combined_5_cast_fp16")]; + tensor input_149_cast_fp16 = add(x = _inversed_t_combined_5_cast_fp16, y = linear_5_cast_fp16)[name = tensor("input_149_cast_fp16")]; + tensor input_151_cast_fp16 = silu(x = input_149_cast_fp16)[name = tensor("input_151_cast_fp16")]; + tensor linear_58_cast_fp16 = linear(bias = flow_net_res_blocks_0_adaLN_modulation_1_bias_to_fp16, weight = flow_net_res_blocks_0_adaLN_modulation_1_weight_to_fp16, x = input_151_cast_fp16)[name = tensor("linear_58_cast_fp16")]; + tensor var_928_split_sizes_0 = const()[name = tensor("op_928_split_sizes_0"), val = tensor([512, 512, 512])]; + tensor var_928_axis_0 = const()[name = tensor("op_928_axis_0"), val = tensor(-1)]; + tensor var_928_cast_fp16_0, tensor var_928_cast_fp16_1, tensor var_928_cast_fp16_2 = split(axis = var_928_axis_0, split_sizes = var_928_split_sizes_0, x = linear_58_cast_fp16)[name = tensor("op_928_cast_fp16")]; + tensor mean_29_axes_0 = const()[name = tensor("mean_29_axes_0"), val = tensor([-1])]; + tensor mean_29_keep_dims_0 = const()[name = tensor("mean_29_keep_dims_0"), val = tensor(true)]; + tensor mean_29_cast_fp16 = reduce_mean(axes = mean_29_axes_0, keep_dims = mean_29_keep_dims_0, x = linear_52_cast_fp16)[name = tensor("mean_29_cast_fp16")]; + tensor sub_26_cast_fp16 = sub(x = linear_52_cast_fp16, y = mean_29_cast_fp16)[name = tensor("sub_26_cast_fp16")]; + tensor square_20_cast_fp16 = square(x = sub_26_cast_fp16)[name = tensor("square_20_cast_fp16")]; + tensor reduce_mean_41_axes_0 = const()[name = tensor("reduce_mean_41_axes_0"), val = tensor([-1])]; + tensor reduce_mean_41_keep_dims_0 = const()[name = tensor("reduce_mean_41_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_41_cast_fp16 = reduce_mean(axes = reduce_mean_41_axes_0, keep_dims = reduce_mean_41_keep_dims_0, x = square_20_cast_fp16)[name = tensor("reduce_mean_41_cast_fp16")]; + tensor var_938_to_fp16 = const()[name = tensor("op_938_to_fp16"), val = tensor(0x1.1p-20)]; + tensor var_939_cast_fp16 = add(x = reduce_mean_41_cast_fp16, y = var_938_to_fp16)[name = tensor("op_939_cast_fp16")]; + tensor var_940_cast_fp16 = sqrt(x = var_939_cast_fp16)[name = tensor("op_940_cast_fp16")]; + tensor x_95_cast_fp16 = real_div(x = sub_26_cast_fp16, y = var_940_cast_fp16)[name = tensor("x_95_cast_fp16")]; + tensor var_942_cast_fp16 = mul(x = x_95_cast_fp16, y = flow_net_res_blocks_0_in_ln_weight_to_fp16)[name = tensor("op_942_cast_fp16")]; + tensor x_97_cast_fp16 = add(x = var_942_cast_fp16, y = flow_net_res_blocks_0_in_ln_bias_to_fp16)[name = tensor("x_97_cast_fp16")]; + tensor var_944_promoted_to_fp16 = const()[name = tensor("op_944_promoted_to_fp16"), val = tensor(0x1p+0)]; + tensor var_945_cast_fp16 = add(x = var_928_cast_fp16_1, y = var_944_promoted_to_fp16)[name = tensor("op_945_cast_fp16")]; + tensor var_946_cast_fp16 = mul(x = x_97_cast_fp16, y = var_945_cast_fp16)[name = tensor("op_946_cast_fp16")]; + tensor input_153_cast_fp16 = add(x = var_946_cast_fp16, y = var_928_cast_fp16_0)[name = tensor("input_153_cast_fp16")]; + tensor linear_59_cast_fp16 = linear(bias = flow_net_res_blocks_0_mlp_0_bias_to_fp16, weight = flow_net_res_blocks_0_mlp_0_weight_to_fp16, x = input_153_cast_fp16)[name = tensor("linear_59_cast_fp16")]; + tensor input_157_cast_fp16 = silu(x = linear_59_cast_fp16)[name = tensor("input_157_cast_fp16")]; + tensor linear_60_cast_fp16 = linear(bias = flow_net_res_blocks_0_mlp_2_bias_to_fp16, weight = flow_net_res_blocks_0_mlp_2_weight_to_fp16, x = input_157_cast_fp16)[name = tensor("linear_60_cast_fp16")]; + tensor var_957_cast_fp16 = mul(x = var_928_cast_fp16_2, y = linear_60_cast_fp16)[name = tensor("op_957_cast_fp16")]; + tensor x_99_cast_fp16 = add(x = linear_52_cast_fp16, y = var_957_cast_fp16)[name = tensor("x_99_cast_fp16")]; + tensor linear_61_cast_fp16 = linear(bias = flow_net_res_blocks_1_adaLN_modulation_1_bias_to_fp16, weight = flow_net_res_blocks_1_adaLN_modulation_1_weight_to_fp16, x = input_151_cast_fp16)[name = tensor("linear_61_cast_fp16")]; + tensor var_967_split_sizes_0 = const()[name = tensor("op_967_split_sizes_0"), val = tensor([512, 512, 512])]; + tensor var_967_axis_0 = const()[name = tensor("op_967_axis_0"), val = tensor(-1)]; + tensor var_967_cast_fp16_0, tensor var_967_cast_fp16_1, tensor var_967_cast_fp16_2 = split(axis = var_967_axis_0, split_sizes = var_967_split_sizes_0, x = linear_61_cast_fp16)[name = tensor("op_967_cast_fp16")]; + tensor mean_31_axes_0 = const()[name = tensor("mean_31_axes_0"), val = tensor([-1])]; + tensor mean_31_keep_dims_0 = const()[name = tensor("mean_31_keep_dims_0"), val = tensor(true)]; + tensor mean_31_cast_fp16 = reduce_mean(axes = mean_31_axes_0, keep_dims = mean_31_keep_dims_0, x = x_99_cast_fp16)[name = tensor("mean_31_cast_fp16")]; + tensor sub_27_cast_fp16 = sub(x = x_99_cast_fp16, y = mean_31_cast_fp16)[name = tensor("sub_27_cast_fp16")]; + tensor square_21_cast_fp16 = square(x = sub_27_cast_fp16)[name = tensor("square_21_cast_fp16")]; + tensor reduce_mean_43_axes_0 = const()[name = tensor("reduce_mean_43_axes_0"), val = tensor([-1])]; + tensor reduce_mean_43_keep_dims_0 = const()[name = tensor("reduce_mean_43_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_43_cast_fp16 = reduce_mean(axes = reduce_mean_43_axes_0, keep_dims = reduce_mean_43_keep_dims_0, x = square_21_cast_fp16)[name = tensor("reduce_mean_43_cast_fp16")]; + tensor var_977_to_fp16 = const()[name = tensor("op_977_to_fp16"), val = tensor(0x1.1p-20)]; + tensor var_978_cast_fp16 = add(x = reduce_mean_43_cast_fp16, y = var_977_to_fp16)[name = tensor("op_978_cast_fp16")]; + tensor var_979_cast_fp16 = sqrt(x = var_978_cast_fp16)[name = tensor("op_979_cast_fp16")]; + tensor x_101_cast_fp16 = real_div(x = sub_27_cast_fp16, y = var_979_cast_fp16)[name = tensor("x_101_cast_fp16")]; + tensor var_981_cast_fp16 = mul(x = x_101_cast_fp16, y = flow_net_res_blocks_1_in_ln_weight_to_fp16)[name = tensor("op_981_cast_fp16")]; + tensor x_103_cast_fp16 = add(x = var_981_cast_fp16, y = flow_net_res_blocks_1_in_ln_bias_to_fp16)[name = tensor("x_103_cast_fp16")]; + tensor var_983_promoted_to_fp16 = const()[name = tensor("op_983_promoted_to_fp16"), val = tensor(0x1p+0)]; + tensor var_984_cast_fp16 = add(x = var_967_cast_fp16_1, y = var_983_promoted_to_fp16)[name = tensor("op_984_cast_fp16")]; + tensor var_985_cast_fp16 = mul(x = x_103_cast_fp16, y = var_984_cast_fp16)[name = tensor("op_985_cast_fp16")]; + tensor input_161_cast_fp16 = add(x = var_985_cast_fp16, y = var_967_cast_fp16_0)[name = tensor("input_161_cast_fp16")]; + tensor linear_62_cast_fp16 = linear(bias = flow_net_res_blocks_1_mlp_0_bias_to_fp16, weight = flow_net_res_blocks_1_mlp_0_weight_to_fp16, x = input_161_cast_fp16)[name = tensor("linear_62_cast_fp16")]; + tensor input_165_cast_fp16 = silu(x = linear_62_cast_fp16)[name = tensor("input_165_cast_fp16")]; + tensor linear_63_cast_fp16 = linear(bias = flow_net_res_blocks_1_mlp_2_bias_to_fp16, weight = flow_net_res_blocks_1_mlp_2_weight_to_fp16, x = input_165_cast_fp16)[name = tensor("linear_63_cast_fp16")]; + tensor var_996_cast_fp16 = mul(x = var_967_cast_fp16_2, y = linear_63_cast_fp16)[name = tensor("op_996_cast_fp16")]; + tensor x_105_cast_fp16 = add(x = x_99_cast_fp16, y = var_996_cast_fp16)[name = tensor("x_105_cast_fp16")]; + tensor linear_64_cast_fp16 = linear(bias = flow_net_res_blocks_2_adaLN_modulation_1_bias_to_fp16, weight = flow_net_res_blocks_2_adaLN_modulation_1_weight_to_fp16, x = input_151_cast_fp16)[name = tensor("linear_64_cast_fp16")]; + tensor var_1006_split_sizes_0 = const()[name = tensor("op_1006_split_sizes_0"), val = tensor([512, 512, 512])]; + tensor var_1006_axis_0 = const()[name = tensor("op_1006_axis_0"), val = tensor(-1)]; + tensor var_1006_cast_fp16_0, tensor var_1006_cast_fp16_1, tensor var_1006_cast_fp16_2 = split(axis = var_1006_axis_0, split_sizes = var_1006_split_sizes_0, x = linear_64_cast_fp16)[name = tensor("op_1006_cast_fp16")]; + tensor mean_33_axes_0 = const()[name = tensor("mean_33_axes_0"), val = tensor([-1])]; + tensor mean_33_keep_dims_0 = const()[name = tensor("mean_33_keep_dims_0"), val = tensor(true)]; + tensor mean_33_cast_fp16 = reduce_mean(axes = mean_33_axes_0, keep_dims = mean_33_keep_dims_0, x = x_105_cast_fp16)[name = tensor("mean_33_cast_fp16")]; + tensor sub_28_cast_fp16 = sub(x = x_105_cast_fp16, y = mean_33_cast_fp16)[name = tensor("sub_28_cast_fp16")]; + tensor square_22_cast_fp16 = square(x = sub_28_cast_fp16)[name = tensor("square_22_cast_fp16")]; + tensor reduce_mean_45_axes_0 = const()[name = tensor("reduce_mean_45_axes_0"), val = tensor([-1])]; + tensor reduce_mean_45_keep_dims_0 = const()[name = tensor("reduce_mean_45_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_45_cast_fp16 = reduce_mean(axes = reduce_mean_45_axes_0, keep_dims = reduce_mean_45_keep_dims_0, x = square_22_cast_fp16)[name = tensor("reduce_mean_45_cast_fp16")]; + tensor var_1016_to_fp16 = const()[name = tensor("op_1016_to_fp16"), val = tensor(0x1.1p-20)]; + tensor var_1017_cast_fp16 = add(x = reduce_mean_45_cast_fp16, y = var_1016_to_fp16)[name = tensor("op_1017_cast_fp16")]; + tensor var_1018_cast_fp16 = sqrt(x = var_1017_cast_fp16)[name = tensor("op_1018_cast_fp16")]; + tensor x_107_cast_fp16 = real_div(x = sub_28_cast_fp16, y = var_1018_cast_fp16)[name = tensor("x_107_cast_fp16")]; + tensor var_1020_cast_fp16 = mul(x = x_107_cast_fp16, y = flow_net_res_blocks_2_in_ln_weight_to_fp16)[name = tensor("op_1020_cast_fp16")]; + tensor x_109_cast_fp16 = add(x = var_1020_cast_fp16, y = flow_net_res_blocks_2_in_ln_bias_to_fp16)[name = tensor("x_109_cast_fp16")]; + tensor var_1022_promoted_to_fp16 = const()[name = tensor("op_1022_promoted_to_fp16"), val = tensor(0x1p+0)]; + tensor var_1023_cast_fp16 = add(x = var_1006_cast_fp16_1, y = var_1022_promoted_to_fp16)[name = tensor("op_1023_cast_fp16")]; + tensor var_1024_cast_fp16 = mul(x = x_109_cast_fp16, y = var_1023_cast_fp16)[name = tensor("op_1024_cast_fp16")]; + tensor input_169_cast_fp16 = add(x = var_1024_cast_fp16, y = var_1006_cast_fp16_0)[name = tensor("input_169_cast_fp16")]; + tensor linear_65_cast_fp16 = linear(bias = flow_net_res_blocks_2_mlp_0_bias_to_fp16, weight = flow_net_res_blocks_2_mlp_0_weight_to_fp16, x = input_169_cast_fp16)[name = tensor("linear_65_cast_fp16")]; + tensor input_173_cast_fp16 = silu(x = linear_65_cast_fp16)[name = tensor("input_173_cast_fp16")]; + tensor linear_66_cast_fp16 = linear(bias = flow_net_res_blocks_2_mlp_2_bias_to_fp16, weight = flow_net_res_blocks_2_mlp_2_weight_to_fp16, x = input_173_cast_fp16)[name = tensor("linear_66_cast_fp16")]; + tensor var_1035_cast_fp16 = mul(x = var_1006_cast_fp16_2, y = linear_66_cast_fp16)[name = tensor("op_1035_cast_fp16")]; + tensor x_111_cast_fp16 = add(x = x_105_cast_fp16, y = var_1035_cast_fp16)[name = tensor("x_111_cast_fp16")]; + tensor linear_67_cast_fp16 = linear(bias = flow_net_res_blocks_3_adaLN_modulation_1_bias_to_fp16, weight = flow_net_res_blocks_3_adaLN_modulation_1_weight_to_fp16, x = input_151_cast_fp16)[name = tensor("linear_67_cast_fp16")]; + tensor var_1045_split_sizes_0 = const()[name = tensor("op_1045_split_sizes_0"), val = tensor([512, 512, 512])]; + tensor var_1045_axis_0 = const()[name = tensor("op_1045_axis_0"), val = tensor(-1)]; + tensor var_1045_cast_fp16_0, tensor var_1045_cast_fp16_1, tensor var_1045_cast_fp16_2 = split(axis = var_1045_axis_0, split_sizes = var_1045_split_sizes_0, x = linear_67_cast_fp16)[name = tensor("op_1045_cast_fp16")]; + tensor mean_35_axes_0 = const()[name = tensor("mean_35_axes_0"), val = tensor([-1])]; + tensor mean_35_keep_dims_0 = const()[name = tensor("mean_35_keep_dims_0"), val = tensor(true)]; + tensor mean_35_cast_fp16 = reduce_mean(axes = mean_35_axes_0, keep_dims = mean_35_keep_dims_0, x = x_111_cast_fp16)[name = tensor("mean_35_cast_fp16")]; + tensor sub_29_cast_fp16 = sub(x = x_111_cast_fp16, y = mean_35_cast_fp16)[name = tensor("sub_29_cast_fp16")]; + tensor square_23_cast_fp16 = square(x = sub_29_cast_fp16)[name = tensor("square_23_cast_fp16")]; + tensor reduce_mean_47_axes_0 = const()[name = tensor("reduce_mean_47_axes_0"), val = tensor([-1])]; + tensor reduce_mean_47_keep_dims_0 = const()[name = tensor("reduce_mean_47_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_47_cast_fp16 = reduce_mean(axes = reduce_mean_47_axes_0, keep_dims = reduce_mean_47_keep_dims_0, x = square_23_cast_fp16)[name = tensor("reduce_mean_47_cast_fp16")]; + tensor var_1055_to_fp16 = const()[name = tensor("op_1055_to_fp16"), val = tensor(0x1.1p-20)]; + tensor var_1056_cast_fp16 = add(x = reduce_mean_47_cast_fp16, y = var_1055_to_fp16)[name = tensor("op_1056_cast_fp16")]; + tensor var_1057_cast_fp16 = sqrt(x = var_1056_cast_fp16)[name = tensor("op_1057_cast_fp16")]; + tensor x_113_cast_fp16 = real_div(x = sub_29_cast_fp16, y = var_1057_cast_fp16)[name = tensor("x_113_cast_fp16")]; + tensor var_1059_cast_fp16 = mul(x = x_113_cast_fp16, y = flow_net_res_blocks_3_in_ln_weight_to_fp16)[name = tensor("op_1059_cast_fp16")]; + tensor x_115_cast_fp16 = add(x = var_1059_cast_fp16, y = flow_net_res_blocks_3_in_ln_bias_to_fp16)[name = tensor("x_115_cast_fp16")]; + tensor var_1061_promoted_to_fp16 = const()[name = tensor("op_1061_promoted_to_fp16"), val = tensor(0x1p+0)]; + tensor var_1062_cast_fp16 = add(x = var_1045_cast_fp16_1, y = var_1061_promoted_to_fp16)[name = tensor("op_1062_cast_fp16")]; + tensor var_1063_cast_fp16 = mul(x = x_115_cast_fp16, y = var_1062_cast_fp16)[name = tensor("op_1063_cast_fp16")]; + tensor input_177_cast_fp16 = add(x = var_1063_cast_fp16, y = var_1045_cast_fp16_0)[name = tensor("input_177_cast_fp16")]; + tensor linear_68_cast_fp16 = linear(bias = flow_net_res_blocks_3_mlp_0_bias_to_fp16, weight = flow_net_res_blocks_3_mlp_0_weight_to_fp16, x = input_177_cast_fp16)[name = tensor("linear_68_cast_fp16")]; + tensor input_181_cast_fp16 = silu(x = linear_68_cast_fp16)[name = tensor("input_181_cast_fp16")]; + tensor linear_69_cast_fp16 = linear(bias = flow_net_res_blocks_3_mlp_2_bias_to_fp16, weight = flow_net_res_blocks_3_mlp_2_weight_to_fp16, x = input_181_cast_fp16)[name = tensor("linear_69_cast_fp16")]; + tensor var_1074_cast_fp16 = mul(x = var_1045_cast_fp16_2, y = linear_69_cast_fp16)[name = tensor("op_1074_cast_fp16")]; + tensor x_117_cast_fp16 = add(x = x_111_cast_fp16, y = var_1074_cast_fp16)[name = tensor("x_117_cast_fp16")]; + tensor linear_70_cast_fp16 = linear(bias = flow_net_res_blocks_4_adaLN_modulation_1_bias_to_fp16, weight = flow_net_res_blocks_4_adaLN_modulation_1_weight_to_fp16, x = input_151_cast_fp16)[name = tensor("linear_70_cast_fp16")]; + tensor var_1084_split_sizes_0 = const()[name = tensor("op_1084_split_sizes_0"), val = tensor([512, 512, 512])]; + tensor var_1084_axis_0 = const()[name = tensor("op_1084_axis_0"), val = tensor(-1)]; + tensor var_1084_cast_fp16_0, tensor var_1084_cast_fp16_1, tensor var_1084_cast_fp16_2 = split(axis = var_1084_axis_0, split_sizes = var_1084_split_sizes_0, x = linear_70_cast_fp16)[name = tensor("op_1084_cast_fp16")]; + tensor mean_37_axes_0 = const()[name = tensor("mean_37_axes_0"), val = tensor([-1])]; + tensor mean_37_keep_dims_0 = const()[name = tensor("mean_37_keep_dims_0"), val = tensor(true)]; + tensor mean_37_cast_fp16 = reduce_mean(axes = mean_37_axes_0, keep_dims = mean_37_keep_dims_0, x = x_117_cast_fp16)[name = tensor("mean_37_cast_fp16")]; + tensor sub_30_cast_fp16 = sub(x = x_117_cast_fp16, y = mean_37_cast_fp16)[name = tensor("sub_30_cast_fp16")]; + tensor square_24_cast_fp16 = square(x = sub_30_cast_fp16)[name = tensor("square_24_cast_fp16")]; + tensor reduce_mean_49_axes_0 = const()[name = tensor("reduce_mean_49_axes_0"), val = tensor([-1])]; + tensor reduce_mean_49_keep_dims_0 = const()[name = tensor("reduce_mean_49_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_49_cast_fp16 = reduce_mean(axes = reduce_mean_49_axes_0, keep_dims = reduce_mean_49_keep_dims_0, x = square_24_cast_fp16)[name = tensor("reduce_mean_49_cast_fp16")]; + tensor var_1094_to_fp16 = const()[name = tensor("op_1094_to_fp16"), val = tensor(0x1.1p-20)]; + tensor var_1095_cast_fp16 = add(x = reduce_mean_49_cast_fp16, y = var_1094_to_fp16)[name = tensor("op_1095_cast_fp16")]; + tensor var_1096_cast_fp16 = sqrt(x = var_1095_cast_fp16)[name = tensor("op_1096_cast_fp16")]; + tensor x_119_cast_fp16 = real_div(x = sub_30_cast_fp16, y = var_1096_cast_fp16)[name = tensor("x_119_cast_fp16")]; + tensor var_1098_cast_fp16 = mul(x = x_119_cast_fp16, y = flow_net_res_blocks_4_in_ln_weight_to_fp16)[name = tensor("op_1098_cast_fp16")]; + tensor x_121_cast_fp16 = add(x = var_1098_cast_fp16, y = flow_net_res_blocks_4_in_ln_bias_to_fp16)[name = tensor("x_121_cast_fp16")]; + tensor var_1100_promoted_to_fp16 = const()[name = tensor("op_1100_promoted_to_fp16"), val = tensor(0x1p+0)]; + tensor var_1101_cast_fp16 = add(x = var_1084_cast_fp16_1, y = var_1100_promoted_to_fp16)[name = tensor("op_1101_cast_fp16")]; + tensor var_1102_cast_fp16 = mul(x = x_121_cast_fp16, y = var_1101_cast_fp16)[name = tensor("op_1102_cast_fp16")]; + tensor input_185_cast_fp16 = add(x = var_1102_cast_fp16, y = var_1084_cast_fp16_0)[name = tensor("input_185_cast_fp16")]; + tensor linear_71_cast_fp16 = linear(bias = flow_net_res_blocks_4_mlp_0_bias_to_fp16, weight = flow_net_res_blocks_4_mlp_0_weight_to_fp16, x = input_185_cast_fp16)[name = tensor("linear_71_cast_fp16")]; + tensor input_189_cast_fp16 = silu(x = linear_71_cast_fp16)[name = tensor("input_189_cast_fp16")]; + tensor linear_72_cast_fp16 = linear(bias = flow_net_res_blocks_4_mlp_2_bias_to_fp16, weight = flow_net_res_blocks_4_mlp_2_weight_to_fp16, x = input_189_cast_fp16)[name = tensor("linear_72_cast_fp16")]; + tensor var_1113_cast_fp16 = mul(x = var_1084_cast_fp16_2, y = linear_72_cast_fp16)[name = tensor("op_1113_cast_fp16")]; + tensor x_123_cast_fp16 = add(x = x_117_cast_fp16, y = var_1113_cast_fp16)[name = tensor("x_123_cast_fp16")]; + tensor linear_73_cast_fp16 = linear(bias = flow_net_res_blocks_5_adaLN_modulation_1_bias_to_fp16, weight = flow_net_res_blocks_5_adaLN_modulation_1_weight_to_fp16, x = input_151_cast_fp16)[name = tensor("linear_73_cast_fp16")]; + tensor var_1123_split_sizes_0 = const()[name = tensor("op_1123_split_sizes_0"), val = tensor([512, 512, 512])]; + tensor var_1123_axis_0 = const()[name = tensor("op_1123_axis_0"), val = tensor(-1)]; + tensor var_1123_cast_fp16_0, tensor var_1123_cast_fp16_1, tensor var_1123_cast_fp16_2 = split(axis = var_1123_axis_0, split_sizes = var_1123_split_sizes_0, x = linear_73_cast_fp16)[name = tensor("op_1123_cast_fp16")]; + tensor mean_39_axes_0 = const()[name = tensor("mean_39_axes_0"), val = tensor([-1])]; + tensor mean_39_keep_dims_0 = const()[name = tensor("mean_39_keep_dims_0"), val = tensor(true)]; + tensor mean_39_cast_fp16 = reduce_mean(axes = mean_39_axes_0, keep_dims = mean_39_keep_dims_0, x = x_123_cast_fp16)[name = tensor("mean_39_cast_fp16")]; + tensor sub_31_cast_fp16 = sub(x = x_123_cast_fp16, y = mean_39_cast_fp16)[name = tensor("sub_31_cast_fp16")]; + tensor square_25_cast_fp16 = square(x = sub_31_cast_fp16)[name = tensor("square_25_cast_fp16")]; + tensor reduce_mean_51_axes_0 = const()[name = tensor("reduce_mean_51_axes_0"), val = tensor([-1])]; + tensor reduce_mean_51_keep_dims_0 = const()[name = tensor("reduce_mean_51_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_51_cast_fp16 = reduce_mean(axes = reduce_mean_51_axes_0, keep_dims = reduce_mean_51_keep_dims_0, x = square_25_cast_fp16)[name = tensor("reduce_mean_51_cast_fp16")]; + tensor var_1133_to_fp16 = const()[name = tensor("op_1133_to_fp16"), val = tensor(0x1.1p-20)]; + tensor var_1134_cast_fp16 = add(x = reduce_mean_51_cast_fp16, y = var_1133_to_fp16)[name = tensor("op_1134_cast_fp16")]; + tensor var_1135_cast_fp16 = sqrt(x = var_1134_cast_fp16)[name = tensor("op_1135_cast_fp16")]; + tensor x_125_cast_fp16 = real_div(x = sub_31_cast_fp16, y = var_1135_cast_fp16)[name = tensor("x_125_cast_fp16")]; + tensor var_1137_cast_fp16 = mul(x = x_125_cast_fp16, y = flow_net_res_blocks_5_in_ln_weight_to_fp16)[name = tensor("op_1137_cast_fp16")]; + tensor x_127_cast_fp16 = add(x = var_1137_cast_fp16, y = flow_net_res_blocks_5_in_ln_bias_to_fp16)[name = tensor("x_127_cast_fp16")]; + tensor var_1139_promoted_to_fp16 = const()[name = tensor("op_1139_promoted_to_fp16"), val = tensor(0x1p+0)]; + tensor var_1140_cast_fp16 = add(x = var_1123_cast_fp16_1, y = var_1139_promoted_to_fp16)[name = tensor("op_1140_cast_fp16")]; + tensor var_1141_cast_fp16 = mul(x = x_127_cast_fp16, y = var_1140_cast_fp16)[name = tensor("op_1141_cast_fp16")]; + tensor input_193_cast_fp16 = add(x = var_1141_cast_fp16, y = var_1123_cast_fp16_0)[name = tensor("input_193_cast_fp16")]; + tensor linear_74_cast_fp16 = linear(bias = flow_net_res_blocks_5_mlp_0_bias_to_fp16, weight = flow_net_res_blocks_5_mlp_0_weight_to_fp16, x = input_193_cast_fp16)[name = tensor("linear_74_cast_fp16")]; + tensor input_197_cast_fp16 = silu(x = linear_74_cast_fp16)[name = tensor("input_197_cast_fp16")]; + tensor linear_75_cast_fp16 = linear(bias = flow_net_res_blocks_5_mlp_2_bias_to_fp16, weight = flow_net_res_blocks_5_mlp_2_weight_to_fp16, x = input_197_cast_fp16)[name = tensor("linear_75_cast_fp16")]; + tensor var_1152_cast_fp16 = mul(x = var_1123_cast_fp16_2, y = linear_75_cast_fp16)[name = tensor("op_1152_cast_fp16")]; + tensor x_129_cast_fp16 = add(x = x_123_cast_fp16, y = var_1152_cast_fp16)[name = tensor("x_129_cast_fp16")]; + tensor linear_76_cast_fp16 = linear(bias = flow_net_final_layer_adaLN_modulation_1_bias_to_fp16, weight = flow_net_final_layer_adaLN_modulation_1_weight_to_fp16, x = input_151_cast_fp16)[name = tensor("linear_76_cast_fp16")]; + tensor var_1161_split_sizes_0 = const()[name = tensor("op_1161_split_sizes_0"), val = tensor([512, 512])]; + tensor var_1161_axis_0 = const()[name = tensor("op_1161_axis_0"), val = tensor(-1)]; + tensor var_1161_cast_fp16_0, tensor var_1161_cast_fp16_1 = split(axis = var_1161_axis_0, split_sizes = var_1161_split_sizes_0, x = linear_76_cast_fp16)[name = tensor("op_1161_cast_fp16")]; + tensor mean_41_axes_0 = const()[name = tensor("mean_41_axes_0"), val = tensor([-1])]; + tensor mean_41_keep_dims_0 = const()[name = tensor("mean_41_keep_dims_0"), val = tensor(true)]; + tensor mean_41_cast_fp16 = reduce_mean(axes = mean_41_axes_0, keep_dims = mean_41_keep_dims_0, x = x_129_cast_fp16)[name = tensor("mean_41_cast_fp16")]; + tensor sub_32_cast_fp16 = sub(x = x_129_cast_fp16, y = mean_41_cast_fp16)[name = tensor("sub_32_cast_fp16")]; + tensor square_26_cast_fp16 = square(x = sub_32_cast_fp16)[name = tensor("square_26_cast_fp16")]; + tensor reduce_mean_53_axes_0 = const()[name = tensor("reduce_mean_53_axes_0"), val = tensor([-1])]; + tensor reduce_mean_53_keep_dims_0 = const()[name = tensor("reduce_mean_53_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_53_cast_fp16 = reduce_mean(axes = reduce_mean_53_axes_0, keep_dims = reduce_mean_53_keep_dims_0, x = square_26_cast_fp16)[name = tensor("reduce_mean_53_cast_fp16")]; + tensor var_1168_to_fp16 = const()[name = tensor("op_1168_to_fp16"), val = tensor(0x1.1p-20)]; + tensor var_1169_cast_fp16 = add(x = reduce_mean_53_cast_fp16, y = var_1168_to_fp16)[name = tensor("op_1169_cast_fp16")]; + tensor var_1170_cast_fp16 = sqrt(x = var_1169_cast_fp16)[name = tensor("op_1170_cast_fp16")]; + tensor x_131_cast_fp16 = real_div(x = sub_32_cast_fp16, y = var_1170_cast_fp16)[name = tensor("x_131_cast_fp16")]; + tensor var_1172_promoted_to_fp16 = const()[name = tensor("op_1172_promoted_to_fp16"), val = tensor(0x1p+0)]; + tensor var_1173_cast_fp16 = add(x = var_1161_cast_fp16_1, y = var_1172_promoted_to_fp16)[name = tensor("op_1173_cast_fp16")]; + tensor var_1174_cast_fp16 = mul(x = x_131_cast_fp16, y = var_1173_cast_fp16)[name = tensor("op_1174_cast_fp16")]; + tensor input_201_cast_fp16 = add(x = var_1174_cast_fp16, y = var_1161_cast_fp16_0)[name = tensor("input_201_cast_fp16")]; + tensor linear_77_cast_fp16 = linear(bias = flow_net_final_layer_linear_bias_to_fp16, weight = flow_net_final_layer_linear_weight_to_fp16, x = input_201_cast_fp16)[name = tensor("linear_77_cast_fp16")]; + tensor var_1185_to_fp16 = const()[name = tensor("op_1185_to_fp16"), val = tensor(0x1p-3)]; + tensor var_1186_cast_fp16 = mul(x = linear_77_cast_fp16, y = var_1185_to_fp16)[name = tensor("op_1186_cast_fp16")]; + tensor input_203_cast_fp16 = add(x = input_135_cast_fp16, y = var_1186_cast_fp16)[name = tensor("input_203_cast_fp16")]; + tensor linear_78_cast_fp16 = linear(bias = flow_net_input_proj_bias_to_fp16, weight = flow_net_input_proj_weight_to_fp16, x = input_203_cast_fp16)[name = tensor("linear_78_cast_fp16")]; + tensor input_207_to_fp16 = const()[name = tensor("input_207_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(19001728)))]; + tensor input_209_cast_fp16 = silu(x = input_207_to_fp16)[name = tensor("input_209_cast_fp16")]; + tensor linear_80_cast_fp16 = linear(bias = flow_net_time_embed_0_mlp_2_bias_to_fp16, weight = flow_net_time_embed_0_mlp_2_weight_to_fp16, x = input_209_cast_fp16)[name = tensor("linear_80_cast_fp16")]; + tensor reduce_mean_54_axes_0 = const()[name = tensor("reduce_mean_54_axes_0"), val = tensor([-1])]; + tensor reduce_mean_54_keep_dims_0 = const()[name = tensor("reduce_mean_54_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_54_cast_fp16 = reduce_mean(axes = reduce_mean_54_axes_0, keep_dims = reduce_mean_54_keep_dims_0, x = linear_80_cast_fp16)[name = tensor("reduce_mean_54_cast_fp16")]; + tensor sub_33_cast_fp16 = sub(x = linear_80_cast_fp16, y = reduce_mean_54_cast_fp16)[name = tensor("sub_33_cast_fp16")]; + tensor square_27_cast_fp16 = square(x = sub_33_cast_fp16)[name = tensor("square_27_cast_fp16")]; + tensor reduce_mean_55_axes_0 = const()[name = tensor("reduce_mean_55_axes_0"), val = tensor([-1])]; + tensor reduce_mean_55_keep_dims_0 = const()[name = tensor("reduce_mean_55_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_55_cast_fp16 = reduce_mean(axes = reduce_mean_55_axes_0, keep_dims = reduce_mean_55_keep_dims_0, x = square_27_cast_fp16)[name = tensor("reduce_mean_55_cast_fp16")]; + tensor real_div_6_to_fp16 = const()[name = tensor("real_div_6_to_fp16"), val = tensor(0x1.008p+0)]; + tensor mul_6_cast_fp16 = mul(x = reduce_mean_55_cast_fp16, y = real_div_6_to_fp16)[name = tensor("mul_6_cast_fp16")]; + tensor var_1252_to_fp16 = const()[name = tensor("op_1252_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_55_cast_fp16 = add(x = mul_6_cast_fp16, y = var_1252_to_fp16)[name = tensor("var_55_cast_fp16")]; + tensor var_1255_epsilon_0 = const()[name = tensor("op_1255_epsilon_0"), val = tensor(0x1.197998p-40)]; + tensor var_1255_cast_fp16 = rsqrt(epsilon = var_1255_epsilon_0, x = var_55_cast_fp16)[name = tensor("op_1255_cast_fp16")]; + tensor var_1256_cast_fp16 = mul(x = const_3_to_fp16, y = var_1255_cast_fp16)[name = tensor("op_1256_cast_fp16")]; + tensor var_1257_cast_fp16 = mul(x = linear_80_cast_fp16, y = var_1256_cast_fp16)[name = tensor("op_1257_cast_fp16")]; + tensor input_213_to_fp16 = const()[name = tensor("input_213_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(19002816)))]; + tensor input_215_cast_fp16 = silu(x = input_213_to_fp16)[name = tensor("input_215_cast_fp16")]; + tensor linear_82_cast_fp16 = linear(bias = flow_net_time_embed_1_mlp_2_bias_to_fp16, weight = flow_net_time_embed_1_mlp_2_weight_to_fp16, x = input_215_cast_fp16)[name = tensor("linear_82_cast_fp16")]; + tensor reduce_mean_56_axes_0 = const()[name = tensor("reduce_mean_56_axes_0"), val = tensor([-1])]; + tensor reduce_mean_56_keep_dims_0 = const()[name = tensor("reduce_mean_56_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_56_cast_fp16 = reduce_mean(axes = reduce_mean_56_axes_0, keep_dims = reduce_mean_56_keep_dims_0, x = linear_82_cast_fp16)[name = tensor("reduce_mean_56_cast_fp16")]; + tensor sub_35_cast_fp16 = sub(x = linear_82_cast_fp16, y = reduce_mean_56_cast_fp16)[name = tensor("sub_35_cast_fp16")]; + tensor square_28_cast_fp16 = square(x = sub_35_cast_fp16)[name = tensor("square_28_cast_fp16")]; + tensor reduce_mean_57_axes_0 = const()[name = tensor("reduce_mean_57_axes_0"), val = tensor([-1])]; + tensor reduce_mean_57_keep_dims_0 = const()[name = tensor("reduce_mean_57_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_57_cast_fp16 = reduce_mean(axes = reduce_mean_57_axes_0, keep_dims = reduce_mean_57_keep_dims_0, x = square_28_cast_fp16)[name = tensor("reduce_mean_57_cast_fp16")]; + tensor real_div_7_to_fp16 = const()[name = tensor("real_div_7_to_fp16"), val = tensor(0x1.008p+0)]; + tensor mul_7_cast_fp16 = mul(x = reduce_mean_57_cast_fp16, y = real_div_7_to_fp16)[name = tensor("mul_7_cast_fp16")]; + tensor var_1289_to_fp16 = const()[name = tensor("op_1289_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_57_cast_fp16 = add(x = mul_7_cast_fp16, y = var_1289_to_fp16)[name = tensor("var_57_cast_fp16")]; + tensor var_1292_epsilon_0 = const()[name = tensor("op_1292_epsilon_0"), val = tensor(0x1.197998p-40)]; + tensor var_1292_cast_fp16 = rsqrt(epsilon = var_1292_epsilon_0, x = var_57_cast_fp16)[name = tensor("op_1292_cast_fp16")]; + tensor var_1293_cast_fp16 = mul(x = const_5_to_fp16, y = var_1292_cast_fp16)[name = tensor("op_1293_cast_fp16")]; + tensor var_1294_cast_fp16 = mul(x = linear_82_cast_fp16, y = var_1293_cast_fp16)[name = tensor("op_1294_cast_fp16")]; + tensor var_1306_cast_fp16 = add(x = var_1257_cast_fp16, y = var_1294_cast_fp16)[name = tensor("op_1306_cast_fp16")]; + tensor _inversed_t_combined_7_y_0_to_fp16 = const()[name = tensor("_inversed_t_combined_7_y_0_to_fp16"), val = tensor(0x1p-1)]; + tensor _inversed_t_combined_7_cast_fp16 = mul(x = var_1306_cast_fp16, y = _inversed_t_combined_7_y_0_to_fp16)[name = tensor("_inversed_t_combined_7_cast_fp16")]; + tensor input_217_cast_fp16 = add(x = _inversed_t_combined_7_cast_fp16, y = linear_5_cast_fp16)[name = tensor("input_217_cast_fp16")]; + tensor input_219_cast_fp16 = silu(x = input_217_cast_fp16)[name = tensor("input_219_cast_fp16")]; + tensor linear_84_cast_fp16 = linear(bias = flow_net_res_blocks_0_adaLN_modulation_1_bias_to_fp16, weight = flow_net_res_blocks_0_adaLN_modulation_1_weight_to_fp16, x = input_219_cast_fp16)[name = tensor("linear_84_cast_fp16")]; + tensor var_1321_split_sizes_0 = const()[name = tensor("op_1321_split_sizes_0"), val = tensor([512, 512, 512])]; + tensor var_1321_axis_0 = const()[name = tensor("op_1321_axis_0"), val = tensor(-1)]; + tensor var_1321_cast_fp16_0, tensor var_1321_cast_fp16_1, tensor var_1321_cast_fp16_2 = split(axis = var_1321_axis_0, split_sizes = var_1321_split_sizes_0, x = linear_84_cast_fp16)[name = tensor("op_1321_cast_fp16")]; + tensor mean_43_axes_0 = const()[name = tensor("mean_43_axes_0"), val = tensor([-1])]; + tensor mean_43_keep_dims_0 = const()[name = tensor("mean_43_keep_dims_0"), val = tensor(true)]; + tensor mean_43_cast_fp16 = reduce_mean(axes = mean_43_axes_0, keep_dims = mean_43_keep_dims_0, x = linear_78_cast_fp16)[name = tensor("mean_43_cast_fp16")]; + tensor sub_37_cast_fp16 = sub(x = linear_78_cast_fp16, y = mean_43_cast_fp16)[name = tensor("sub_37_cast_fp16")]; + tensor square_29_cast_fp16 = square(x = sub_37_cast_fp16)[name = tensor("square_29_cast_fp16")]; + tensor reduce_mean_59_axes_0 = const()[name = tensor("reduce_mean_59_axes_0"), val = tensor([-1])]; + tensor reduce_mean_59_keep_dims_0 = const()[name = tensor("reduce_mean_59_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_59_cast_fp16 = reduce_mean(axes = reduce_mean_59_axes_0, keep_dims = reduce_mean_59_keep_dims_0, x = square_29_cast_fp16)[name = tensor("reduce_mean_59_cast_fp16")]; + tensor var_1331_to_fp16 = const()[name = tensor("op_1331_to_fp16"), val = tensor(0x1.1p-20)]; + tensor var_1332_cast_fp16 = add(x = reduce_mean_59_cast_fp16, y = var_1331_to_fp16)[name = tensor("op_1332_cast_fp16")]; + tensor var_1333_cast_fp16 = sqrt(x = var_1332_cast_fp16)[name = tensor("op_1333_cast_fp16")]; + tensor x_139_cast_fp16 = real_div(x = sub_37_cast_fp16, y = var_1333_cast_fp16)[name = tensor("x_139_cast_fp16")]; + tensor var_1335_cast_fp16 = mul(x = x_139_cast_fp16, y = flow_net_res_blocks_0_in_ln_weight_to_fp16)[name = tensor("op_1335_cast_fp16")]; + tensor x_141_cast_fp16 = add(x = var_1335_cast_fp16, y = flow_net_res_blocks_0_in_ln_bias_to_fp16)[name = tensor("x_141_cast_fp16")]; + tensor var_1337_promoted_to_fp16 = const()[name = tensor("op_1337_promoted_to_fp16"), val = tensor(0x1p+0)]; + tensor var_1338_cast_fp16 = add(x = var_1321_cast_fp16_1, y = var_1337_promoted_to_fp16)[name = tensor("op_1338_cast_fp16")]; + tensor var_1339_cast_fp16 = mul(x = x_141_cast_fp16, y = var_1338_cast_fp16)[name = tensor("op_1339_cast_fp16")]; + tensor input_221_cast_fp16 = add(x = var_1339_cast_fp16, y = var_1321_cast_fp16_0)[name = tensor("input_221_cast_fp16")]; + tensor linear_85_cast_fp16 = linear(bias = flow_net_res_blocks_0_mlp_0_bias_to_fp16, weight = flow_net_res_blocks_0_mlp_0_weight_to_fp16, x = input_221_cast_fp16)[name = tensor("linear_85_cast_fp16")]; + tensor input_225_cast_fp16 = silu(x = linear_85_cast_fp16)[name = tensor("input_225_cast_fp16")]; + tensor linear_86_cast_fp16 = linear(bias = flow_net_res_blocks_0_mlp_2_bias_to_fp16, weight = flow_net_res_blocks_0_mlp_2_weight_to_fp16, x = input_225_cast_fp16)[name = tensor("linear_86_cast_fp16")]; + tensor var_1350_cast_fp16 = mul(x = var_1321_cast_fp16_2, y = linear_86_cast_fp16)[name = tensor("op_1350_cast_fp16")]; + tensor x_143_cast_fp16 = add(x = linear_78_cast_fp16, y = var_1350_cast_fp16)[name = tensor("x_143_cast_fp16")]; + tensor linear_87_cast_fp16 = linear(bias = flow_net_res_blocks_1_adaLN_modulation_1_bias_to_fp16, weight = flow_net_res_blocks_1_adaLN_modulation_1_weight_to_fp16, x = input_219_cast_fp16)[name = tensor("linear_87_cast_fp16")]; + tensor var_1360_split_sizes_0 = const()[name = tensor("op_1360_split_sizes_0"), val = tensor([512, 512, 512])]; + tensor var_1360_axis_0 = const()[name = tensor("op_1360_axis_0"), val = tensor(-1)]; + tensor var_1360_cast_fp16_0, tensor var_1360_cast_fp16_1, tensor var_1360_cast_fp16_2 = split(axis = var_1360_axis_0, split_sizes = var_1360_split_sizes_0, x = linear_87_cast_fp16)[name = tensor("op_1360_cast_fp16")]; + tensor mean_45_axes_0 = const()[name = tensor("mean_45_axes_0"), val = tensor([-1])]; + tensor mean_45_keep_dims_0 = const()[name = tensor("mean_45_keep_dims_0"), val = tensor(true)]; + tensor mean_45_cast_fp16 = reduce_mean(axes = mean_45_axes_0, keep_dims = mean_45_keep_dims_0, x = x_143_cast_fp16)[name = tensor("mean_45_cast_fp16")]; + tensor sub_38_cast_fp16 = sub(x = x_143_cast_fp16, y = mean_45_cast_fp16)[name = tensor("sub_38_cast_fp16")]; + tensor square_30_cast_fp16 = square(x = sub_38_cast_fp16)[name = tensor("square_30_cast_fp16")]; + tensor reduce_mean_61_axes_0 = const()[name = tensor("reduce_mean_61_axes_0"), val = tensor([-1])]; + tensor reduce_mean_61_keep_dims_0 = const()[name = tensor("reduce_mean_61_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_61_cast_fp16 = reduce_mean(axes = reduce_mean_61_axes_0, keep_dims = reduce_mean_61_keep_dims_0, x = square_30_cast_fp16)[name = tensor("reduce_mean_61_cast_fp16")]; + tensor var_1370_to_fp16 = const()[name = tensor("op_1370_to_fp16"), val = tensor(0x1.1p-20)]; + tensor var_1371_cast_fp16 = add(x = reduce_mean_61_cast_fp16, y = var_1370_to_fp16)[name = tensor("op_1371_cast_fp16")]; + tensor var_1372_cast_fp16 = sqrt(x = var_1371_cast_fp16)[name = tensor("op_1372_cast_fp16")]; + tensor x_145_cast_fp16 = real_div(x = sub_38_cast_fp16, y = var_1372_cast_fp16)[name = tensor("x_145_cast_fp16")]; + tensor var_1374_cast_fp16 = mul(x = x_145_cast_fp16, y = flow_net_res_blocks_1_in_ln_weight_to_fp16)[name = tensor("op_1374_cast_fp16")]; + tensor x_147_cast_fp16 = add(x = var_1374_cast_fp16, y = flow_net_res_blocks_1_in_ln_bias_to_fp16)[name = tensor("x_147_cast_fp16")]; + tensor var_1376_promoted_to_fp16 = const()[name = tensor("op_1376_promoted_to_fp16"), val = tensor(0x1p+0)]; + tensor var_1377_cast_fp16 = add(x = var_1360_cast_fp16_1, y = var_1376_promoted_to_fp16)[name = tensor("op_1377_cast_fp16")]; + tensor var_1378_cast_fp16 = mul(x = x_147_cast_fp16, y = var_1377_cast_fp16)[name = tensor("op_1378_cast_fp16")]; + tensor input_229_cast_fp16 = add(x = var_1378_cast_fp16, y = var_1360_cast_fp16_0)[name = tensor("input_229_cast_fp16")]; + tensor linear_88_cast_fp16 = linear(bias = flow_net_res_blocks_1_mlp_0_bias_to_fp16, weight = flow_net_res_blocks_1_mlp_0_weight_to_fp16, x = input_229_cast_fp16)[name = tensor("linear_88_cast_fp16")]; + tensor input_233_cast_fp16 = silu(x = linear_88_cast_fp16)[name = tensor("input_233_cast_fp16")]; + tensor linear_89_cast_fp16 = linear(bias = flow_net_res_blocks_1_mlp_2_bias_to_fp16, weight = flow_net_res_blocks_1_mlp_2_weight_to_fp16, x = input_233_cast_fp16)[name = tensor("linear_89_cast_fp16")]; + tensor var_1389_cast_fp16 = mul(x = var_1360_cast_fp16_2, y = linear_89_cast_fp16)[name = tensor("op_1389_cast_fp16")]; + tensor x_149_cast_fp16 = add(x = x_143_cast_fp16, y = var_1389_cast_fp16)[name = tensor("x_149_cast_fp16")]; + tensor linear_90_cast_fp16 = linear(bias = flow_net_res_blocks_2_adaLN_modulation_1_bias_to_fp16, weight = flow_net_res_blocks_2_adaLN_modulation_1_weight_to_fp16, x = input_219_cast_fp16)[name = tensor("linear_90_cast_fp16")]; + tensor var_1399_split_sizes_0 = const()[name = tensor("op_1399_split_sizes_0"), val = tensor([512, 512, 512])]; + tensor var_1399_axis_0 = const()[name = tensor("op_1399_axis_0"), val = tensor(-1)]; + tensor var_1399_cast_fp16_0, tensor var_1399_cast_fp16_1, tensor var_1399_cast_fp16_2 = split(axis = var_1399_axis_0, split_sizes = var_1399_split_sizes_0, x = linear_90_cast_fp16)[name = tensor("op_1399_cast_fp16")]; + tensor mean_47_axes_0 = const()[name = tensor("mean_47_axes_0"), val = tensor([-1])]; + tensor mean_47_keep_dims_0 = const()[name = tensor("mean_47_keep_dims_0"), val = tensor(true)]; + tensor mean_47_cast_fp16 = reduce_mean(axes = mean_47_axes_0, keep_dims = mean_47_keep_dims_0, x = x_149_cast_fp16)[name = tensor("mean_47_cast_fp16")]; + tensor sub_39_cast_fp16 = sub(x = x_149_cast_fp16, y = mean_47_cast_fp16)[name = tensor("sub_39_cast_fp16")]; + tensor square_31_cast_fp16 = square(x = sub_39_cast_fp16)[name = tensor("square_31_cast_fp16")]; + tensor reduce_mean_63_axes_0 = const()[name = tensor("reduce_mean_63_axes_0"), val = tensor([-1])]; + tensor reduce_mean_63_keep_dims_0 = const()[name = tensor("reduce_mean_63_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_63_cast_fp16 = reduce_mean(axes = reduce_mean_63_axes_0, keep_dims = reduce_mean_63_keep_dims_0, x = square_31_cast_fp16)[name = tensor("reduce_mean_63_cast_fp16")]; + tensor var_1409_to_fp16 = const()[name = tensor("op_1409_to_fp16"), val = tensor(0x1.1p-20)]; + tensor var_1410_cast_fp16 = add(x = reduce_mean_63_cast_fp16, y = var_1409_to_fp16)[name = tensor("op_1410_cast_fp16")]; + tensor var_1411_cast_fp16 = sqrt(x = var_1410_cast_fp16)[name = tensor("op_1411_cast_fp16")]; + tensor x_151_cast_fp16 = real_div(x = sub_39_cast_fp16, y = var_1411_cast_fp16)[name = tensor("x_151_cast_fp16")]; + tensor var_1413_cast_fp16 = mul(x = x_151_cast_fp16, y = flow_net_res_blocks_2_in_ln_weight_to_fp16)[name = tensor("op_1413_cast_fp16")]; + tensor x_153_cast_fp16 = add(x = var_1413_cast_fp16, y = flow_net_res_blocks_2_in_ln_bias_to_fp16)[name = tensor("x_153_cast_fp16")]; + tensor var_1415_promoted_to_fp16 = const()[name = tensor("op_1415_promoted_to_fp16"), val = tensor(0x1p+0)]; + tensor var_1416_cast_fp16 = add(x = var_1399_cast_fp16_1, y = var_1415_promoted_to_fp16)[name = tensor("op_1416_cast_fp16")]; + tensor var_1417_cast_fp16 = mul(x = x_153_cast_fp16, y = var_1416_cast_fp16)[name = tensor("op_1417_cast_fp16")]; + tensor input_237_cast_fp16 = add(x = var_1417_cast_fp16, y = var_1399_cast_fp16_0)[name = tensor("input_237_cast_fp16")]; + tensor linear_91_cast_fp16 = linear(bias = flow_net_res_blocks_2_mlp_0_bias_to_fp16, weight = flow_net_res_blocks_2_mlp_0_weight_to_fp16, x = input_237_cast_fp16)[name = tensor("linear_91_cast_fp16")]; + tensor input_241_cast_fp16 = silu(x = linear_91_cast_fp16)[name = tensor("input_241_cast_fp16")]; + tensor linear_92_cast_fp16 = linear(bias = flow_net_res_blocks_2_mlp_2_bias_to_fp16, weight = flow_net_res_blocks_2_mlp_2_weight_to_fp16, x = input_241_cast_fp16)[name = tensor("linear_92_cast_fp16")]; + tensor var_1428_cast_fp16 = mul(x = var_1399_cast_fp16_2, y = linear_92_cast_fp16)[name = tensor("op_1428_cast_fp16")]; + tensor x_155_cast_fp16 = add(x = x_149_cast_fp16, y = var_1428_cast_fp16)[name = tensor("x_155_cast_fp16")]; + tensor linear_93_cast_fp16 = linear(bias = flow_net_res_blocks_3_adaLN_modulation_1_bias_to_fp16, weight = flow_net_res_blocks_3_adaLN_modulation_1_weight_to_fp16, x = input_219_cast_fp16)[name = tensor("linear_93_cast_fp16")]; + tensor var_1438_split_sizes_0 = const()[name = tensor("op_1438_split_sizes_0"), val = tensor([512, 512, 512])]; + tensor var_1438_axis_0 = const()[name = tensor("op_1438_axis_0"), val = tensor(-1)]; + tensor var_1438_cast_fp16_0, tensor var_1438_cast_fp16_1, tensor var_1438_cast_fp16_2 = split(axis = var_1438_axis_0, split_sizes = var_1438_split_sizes_0, x = linear_93_cast_fp16)[name = tensor("op_1438_cast_fp16")]; + tensor mean_49_axes_0 = const()[name = tensor("mean_49_axes_0"), val = tensor([-1])]; + tensor mean_49_keep_dims_0 = const()[name = tensor("mean_49_keep_dims_0"), val = tensor(true)]; + tensor mean_49_cast_fp16 = reduce_mean(axes = mean_49_axes_0, keep_dims = mean_49_keep_dims_0, x = x_155_cast_fp16)[name = tensor("mean_49_cast_fp16")]; + tensor sub_40_cast_fp16 = sub(x = x_155_cast_fp16, y = mean_49_cast_fp16)[name = tensor("sub_40_cast_fp16")]; + tensor square_32_cast_fp16 = square(x = sub_40_cast_fp16)[name = tensor("square_32_cast_fp16")]; + tensor reduce_mean_65_axes_0 = const()[name = tensor("reduce_mean_65_axes_0"), val = tensor([-1])]; + tensor reduce_mean_65_keep_dims_0 = const()[name = tensor("reduce_mean_65_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_65_cast_fp16 = reduce_mean(axes = reduce_mean_65_axes_0, keep_dims = reduce_mean_65_keep_dims_0, x = square_32_cast_fp16)[name = tensor("reduce_mean_65_cast_fp16")]; + tensor var_1448_to_fp16 = const()[name = tensor("op_1448_to_fp16"), val = tensor(0x1.1p-20)]; + tensor var_1449_cast_fp16 = add(x = reduce_mean_65_cast_fp16, y = var_1448_to_fp16)[name = tensor("op_1449_cast_fp16")]; + tensor var_1450_cast_fp16 = sqrt(x = var_1449_cast_fp16)[name = tensor("op_1450_cast_fp16")]; + tensor x_157_cast_fp16 = real_div(x = sub_40_cast_fp16, y = var_1450_cast_fp16)[name = tensor("x_157_cast_fp16")]; + tensor var_1452_cast_fp16 = mul(x = x_157_cast_fp16, y = flow_net_res_blocks_3_in_ln_weight_to_fp16)[name = tensor("op_1452_cast_fp16")]; + tensor x_159_cast_fp16 = add(x = var_1452_cast_fp16, y = flow_net_res_blocks_3_in_ln_bias_to_fp16)[name = tensor("x_159_cast_fp16")]; + tensor var_1454_promoted_to_fp16 = const()[name = tensor("op_1454_promoted_to_fp16"), val = tensor(0x1p+0)]; + tensor var_1455_cast_fp16 = add(x = var_1438_cast_fp16_1, y = var_1454_promoted_to_fp16)[name = tensor("op_1455_cast_fp16")]; + tensor var_1456_cast_fp16 = mul(x = x_159_cast_fp16, y = var_1455_cast_fp16)[name = tensor("op_1456_cast_fp16")]; + tensor input_245_cast_fp16 = add(x = var_1456_cast_fp16, y = var_1438_cast_fp16_0)[name = tensor("input_245_cast_fp16")]; + tensor linear_94_cast_fp16 = linear(bias = flow_net_res_blocks_3_mlp_0_bias_to_fp16, weight = flow_net_res_blocks_3_mlp_0_weight_to_fp16, x = input_245_cast_fp16)[name = tensor("linear_94_cast_fp16")]; + tensor input_249_cast_fp16 = silu(x = linear_94_cast_fp16)[name = tensor("input_249_cast_fp16")]; + tensor linear_95_cast_fp16 = linear(bias = flow_net_res_blocks_3_mlp_2_bias_to_fp16, weight = flow_net_res_blocks_3_mlp_2_weight_to_fp16, x = input_249_cast_fp16)[name = tensor("linear_95_cast_fp16")]; + tensor var_1467_cast_fp16 = mul(x = var_1438_cast_fp16_2, y = linear_95_cast_fp16)[name = tensor("op_1467_cast_fp16")]; + tensor x_161_cast_fp16 = add(x = x_155_cast_fp16, y = var_1467_cast_fp16)[name = tensor("x_161_cast_fp16")]; + tensor linear_96_cast_fp16 = linear(bias = flow_net_res_blocks_4_adaLN_modulation_1_bias_to_fp16, weight = flow_net_res_blocks_4_adaLN_modulation_1_weight_to_fp16, x = input_219_cast_fp16)[name = tensor("linear_96_cast_fp16")]; + tensor var_1477_split_sizes_0 = const()[name = tensor("op_1477_split_sizes_0"), val = tensor([512, 512, 512])]; + tensor var_1477_axis_0 = const()[name = tensor("op_1477_axis_0"), val = tensor(-1)]; + tensor var_1477_cast_fp16_0, tensor var_1477_cast_fp16_1, tensor var_1477_cast_fp16_2 = split(axis = var_1477_axis_0, split_sizes = var_1477_split_sizes_0, x = linear_96_cast_fp16)[name = tensor("op_1477_cast_fp16")]; + tensor mean_51_axes_0 = const()[name = tensor("mean_51_axes_0"), val = tensor([-1])]; + tensor mean_51_keep_dims_0 = const()[name = tensor("mean_51_keep_dims_0"), val = tensor(true)]; + tensor mean_51_cast_fp16 = reduce_mean(axes = mean_51_axes_0, keep_dims = mean_51_keep_dims_0, x = x_161_cast_fp16)[name = tensor("mean_51_cast_fp16")]; + tensor sub_41_cast_fp16 = sub(x = x_161_cast_fp16, y = mean_51_cast_fp16)[name = tensor("sub_41_cast_fp16")]; + tensor square_33_cast_fp16 = square(x = sub_41_cast_fp16)[name = tensor("square_33_cast_fp16")]; + tensor reduce_mean_67_axes_0 = const()[name = tensor("reduce_mean_67_axes_0"), val = tensor([-1])]; + tensor reduce_mean_67_keep_dims_0 = const()[name = tensor("reduce_mean_67_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_67_cast_fp16 = reduce_mean(axes = reduce_mean_67_axes_0, keep_dims = reduce_mean_67_keep_dims_0, x = square_33_cast_fp16)[name = tensor("reduce_mean_67_cast_fp16")]; + tensor var_1487_to_fp16 = const()[name = tensor("op_1487_to_fp16"), val = tensor(0x1.1p-20)]; + tensor var_1488_cast_fp16 = add(x = reduce_mean_67_cast_fp16, y = var_1487_to_fp16)[name = tensor("op_1488_cast_fp16")]; + tensor var_1489_cast_fp16 = sqrt(x = var_1488_cast_fp16)[name = tensor("op_1489_cast_fp16")]; + tensor x_163_cast_fp16 = real_div(x = sub_41_cast_fp16, y = var_1489_cast_fp16)[name = tensor("x_163_cast_fp16")]; + tensor var_1491_cast_fp16 = mul(x = x_163_cast_fp16, y = flow_net_res_blocks_4_in_ln_weight_to_fp16)[name = tensor("op_1491_cast_fp16")]; + tensor x_165_cast_fp16 = add(x = var_1491_cast_fp16, y = flow_net_res_blocks_4_in_ln_bias_to_fp16)[name = tensor("x_165_cast_fp16")]; + tensor var_1493_promoted_to_fp16 = const()[name = tensor("op_1493_promoted_to_fp16"), val = tensor(0x1p+0)]; + tensor var_1494_cast_fp16 = add(x = var_1477_cast_fp16_1, y = var_1493_promoted_to_fp16)[name = tensor("op_1494_cast_fp16")]; + tensor var_1495_cast_fp16 = mul(x = x_165_cast_fp16, y = var_1494_cast_fp16)[name = tensor("op_1495_cast_fp16")]; + tensor input_253_cast_fp16 = add(x = var_1495_cast_fp16, y = var_1477_cast_fp16_0)[name = tensor("input_253_cast_fp16")]; + tensor linear_97_cast_fp16 = linear(bias = flow_net_res_blocks_4_mlp_0_bias_to_fp16, weight = flow_net_res_blocks_4_mlp_0_weight_to_fp16, x = input_253_cast_fp16)[name = tensor("linear_97_cast_fp16")]; + tensor input_257_cast_fp16 = silu(x = linear_97_cast_fp16)[name = tensor("input_257_cast_fp16")]; + tensor linear_98_cast_fp16 = linear(bias = flow_net_res_blocks_4_mlp_2_bias_to_fp16, weight = flow_net_res_blocks_4_mlp_2_weight_to_fp16, x = input_257_cast_fp16)[name = tensor("linear_98_cast_fp16")]; + tensor var_1506_cast_fp16 = mul(x = var_1477_cast_fp16_2, y = linear_98_cast_fp16)[name = tensor("op_1506_cast_fp16")]; + tensor x_167_cast_fp16 = add(x = x_161_cast_fp16, y = var_1506_cast_fp16)[name = tensor("x_167_cast_fp16")]; + tensor linear_99_cast_fp16 = linear(bias = flow_net_res_blocks_5_adaLN_modulation_1_bias_to_fp16, weight = flow_net_res_blocks_5_adaLN_modulation_1_weight_to_fp16, x = input_219_cast_fp16)[name = tensor("linear_99_cast_fp16")]; + tensor var_1516_split_sizes_0 = const()[name = tensor("op_1516_split_sizes_0"), val = tensor([512, 512, 512])]; + tensor var_1516_axis_0 = const()[name = tensor("op_1516_axis_0"), val = tensor(-1)]; + tensor var_1516_cast_fp16_0, tensor var_1516_cast_fp16_1, tensor var_1516_cast_fp16_2 = split(axis = var_1516_axis_0, split_sizes = var_1516_split_sizes_0, x = linear_99_cast_fp16)[name = tensor("op_1516_cast_fp16")]; + tensor mean_53_axes_0 = const()[name = tensor("mean_53_axes_0"), val = tensor([-1])]; + tensor mean_53_keep_dims_0 = const()[name = tensor("mean_53_keep_dims_0"), val = tensor(true)]; + tensor mean_53_cast_fp16 = reduce_mean(axes = mean_53_axes_0, keep_dims = mean_53_keep_dims_0, x = x_167_cast_fp16)[name = tensor("mean_53_cast_fp16")]; + tensor sub_42_cast_fp16 = sub(x = x_167_cast_fp16, y = mean_53_cast_fp16)[name = tensor("sub_42_cast_fp16")]; + tensor square_34_cast_fp16 = square(x = sub_42_cast_fp16)[name = tensor("square_34_cast_fp16")]; + tensor reduce_mean_69_axes_0 = const()[name = tensor("reduce_mean_69_axes_0"), val = tensor([-1])]; + tensor reduce_mean_69_keep_dims_0 = const()[name = tensor("reduce_mean_69_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_69_cast_fp16 = reduce_mean(axes = reduce_mean_69_axes_0, keep_dims = reduce_mean_69_keep_dims_0, x = square_34_cast_fp16)[name = tensor("reduce_mean_69_cast_fp16")]; + tensor var_1526_to_fp16 = const()[name = tensor("op_1526_to_fp16"), val = tensor(0x1.1p-20)]; + tensor var_1527_cast_fp16 = add(x = reduce_mean_69_cast_fp16, y = var_1526_to_fp16)[name = tensor("op_1527_cast_fp16")]; + tensor var_1528_cast_fp16 = sqrt(x = var_1527_cast_fp16)[name = tensor("op_1528_cast_fp16")]; + tensor x_169_cast_fp16 = real_div(x = sub_42_cast_fp16, y = var_1528_cast_fp16)[name = tensor("x_169_cast_fp16")]; + tensor var_1530_cast_fp16 = mul(x = x_169_cast_fp16, y = flow_net_res_blocks_5_in_ln_weight_to_fp16)[name = tensor("op_1530_cast_fp16")]; + tensor x_171_cast_fp16 = add(x = var_1530_cast_fp16, y = flow_net_res_blocks_5_in_ln_bias_to_fp16)[name = tensor("x_171_cast_fp16")]; + tensor var_1532_promoted_to_fp16 = const()[name = tensor("op_1532_promoted_to_fp16"), val = tensor(0x1p+0)]; + tensor var_1533_cast_fp16 = add(x = var_1516_cast_fp16_1, y = var_1532_promoted_to_fp16)[name = tensor("op_1533_cast_fp16")]; + tensor var_1534_cast_fp16 = mul(x = x_171_cast_fp16, y = var_1533_cast_fp16)[name = tensor("op_1534_cast_fp16")]; + tensor input_261_cast_fp16 = add(x = var_1534_cast_fp16, y = var_1516_cast_fp16_0)[name = tensor("input_261_cast_fp16")]; + tensor linear_100_cast_fp16 = linear(bias = flow_net_res_blocks_5_mlp_0_bias_to_fp16, weight = flow_net_res_blocks_5_mlp_0_weight_to_fp16, x = input_261_cast_fp16)[name = tensor("linear_100_cast_fp16")]; + tensor input_265_cast_fp16 = silu(x = linear_100_cast_fp16)[name = tensor("input_265_cast_fp16")]; + tensor linear_101_cast_fp16 = linear(bias = flow_net_res_blocks_5_mlp_2_bias_to_fp16, weight = flow_net_res_blocks_5_mlp_2_weight_to_fp16, x = input_265_cast_fp16)[name = tensor("linear_101_cast_fp16")]; + tensor var_1545_cast_fp16 = mul(x = var_1516_cast_fp16_2, y = linear_101_cast_fp16)[name = tensor("op_1545_cast_fp16")]; + tensor x_173_cast_fp16 = add(x = x_167_cast_fp16, y = var_1545_cast_fp16)[name = tensor("x_173_cast_fp16")]; + tensor linear_102_cast_fp16 = linear(bias = flow_net_final_layer_adaLN_modulation_1_bias_to_fp16, weight = flow_net_final_layer_adaLN_modulation_1_weight_to_fp16, x = input_219_cast_fp16)[name = tensor("linear_102_cast_fp16")]; + tensor var_1554_split_sizes_0 = const()[name = tensor("op_1554_split_sizes_0"), val = tensor([512, 512])]; + tensor var_1554_axis_0 = const()[name = tensor("op_1554_axis_0"), val = tensor(-1)]; + tensor var_1554_cast_fp16_0, tensor var_1554_cast_fp16_1 = split(axis = var_1554_axis_0, split_sizes = var_1554_split_sizes_0, x = linear_102_cast_fp16)[name = tensor("op_1554_cast_fp16")]; + tensor mean_55_axes_0 = const()[name = tensor("mean_55_axes_0"), val = tensor([-1])]; + tensor mean_55_keep_dims_0 = const()[name = tensor("mean_55_keep_dims_0"), val = tensor(true)]; + tensor mean_55_cast_fp16 = reduce_mean(axes = mean_55_axes_0, keep_dims = mean_55_keep_dims_0, x = x_173_cast_fp16)[name = tensor("mean_55_cast_fp16")]; + tensor sub_43_cast_fp16 = sub(x = x_173_cast_fp16, y = mean_55_cast_fp16)[name = tensor("sub_43_cast_fp16")]; + tensor square_35_cast_fp16 = square(x = sub_43_cast_fp16)[name = tensor("square_35_cast_fp16")]; + tensor reduce_mean_71_axes_0 = const()[name = tensor("reduce_mean_71_axes_0"), val = tensor([-1])]; + tensor reduce_mean_71_keep_dims_0 = const()[name = tensor("reduce_mean_71_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_71_cast_fp16 = reduce_mean(axes = reduce_mean_71_axes_0, keep_dims = reduce_mean_71_keep_dims_0, x = square_35_cast_fp16)[name = tensor("reduce_mean_71_cast_fp16")]; + tensor var_1561_to_fp16 = const()[name = tensor("op_1561_to_fp16"), val = tensor(0x1.1p-20)]; + tensor var_1562_cast_fp16 = add(x = reduce_mean_71_cast_fp16, y = var_1561_to_fp16)[name = tensor("op_1562_cast_fp16")]; + tensor var_1563_cast_fp16 = sqrt(x = var_1562_cast_fp16)[name = tensor("op_1563_cast_fp16")]; + tensor x_175_cast_fp16 = real_div(x = sub_43_cast_fp16, y = var_1563_cast_fp16)[name = tensor("x_175_cast_fp16")]; + tensor var_1565_promoted_to_fp16 = const()[name = tensor("op_1565_promoted_to_fp16"), val = tensor(0x1p+0)]; + tensor var_1566_cast_fp16 = add(x = var_1554_cast_fp16_1, y = var_1565_promoted_to_fp16)[name = tensor("op_1566_cast_fp16")]; + tensor var_1567_cast_fp16 = mul(x = x_175_cast_fp16, y = var_1566_cast_fp16)[name = tensor("op_1567_cast_fp16")]; + tensor input_269_cast_fp16 = add(x = var_1567_cast_fp16, y = var_1554_cast_fp16_0)[name = tensor("input_269_cast_fp16")]; + tensor linear_103_cast_fp16 = linear(bias = flow_net_final_layer_linear_bias_to_fp16, weight = flow_net_final_layer_linear_weight_to_fp16, x = input_269_cast_fp16)[name = tensor("linear_103_cast_fp16")]; + tensor var_1578_to_fp16 = const()[name = tensor("op_1578_to_fp16"), val = tensor(0x1p-3)]; + tensor var_1579_cast_fp16 = mul(x = linear_103_cast_fp16, y = var_1578_to_fp16)[name = tensor("op_1579_cast_fp16")]; + tensor input_271_cast_fp16 = add(x = input_203_cast_fp16, y = var_1579_cast_fp16)[name = tensor("input_271_cast_fp16")]; + tensor linear_104_cast_fp16 = linear(bias = flow_net_input_proj_bias_to_fp16, weight = flow_net_input_proj_weight_to_fp16, x = input_271_cast_fp16)[name = tensor("linear_104_cast_fp16")]; + tensor input_275_to_fp16 = const()[name = tensor("input_275_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(19003904)))]; + tensor input_277_cast_fp16 = silu(x = input_275_to_fp16)[name = tensor("input_277_cast_fp16")]; + tensor linear_106_cast_fp16 = linear(bias = flow_net_time_embed_0_mlp_2_bias_to_fp16, weight = flow_net_time_embed_0_mlp_2_weight_to_fp16, x = input_277_cast_fp16)[name = tensor("linear_106_cast_fp16")]; + tensor reduce_mean_72_axes_0 = const()[name = tensor("reduce_mean_72_axes_0"), val = tensor([-1])]; + tensor reduce_mean_72_keep_dims_0 = const()[name = tensor("reduce_mean_72_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_72_cast_fp16 = reduce_mean(axes = reduce_mean_72_axes_0, keep_dims = reduce_mean_72_keep_dims_0, x = linear_106_cast_fp16)[name = tensor("reduce_mean_72_cast_fp16")]; + tensor sub_44_cast_fp16 = sub(x = linear_106_cast_fp16, y = reduce_mean_72_cast_fp16)[name = tensor("sub_44_cast_fp16")]; + tensor square_36_cast_fp16 = square(x = sub_44_cast_fp16)[name = tensor("square_36_cast_fp16")]; + tensor reduce_mean_73_axes_0 = const()[name = tensor("reduce_mean_73_axes_0"), val = tensor([-1])]; + tensor reduce_mean_73_keep_dims_0 = const()[name = tensor("reduce_mean_73_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_73_cast_fp16 = reduce_mean(axes = reduce_mean_73_axes_0, keep_dims = reduce_mean_73_keep_dims_0, x = square_36_cast_fp16)[name = tensor("reduce_mean_73_cast_fp16")]; + tensor real_div_8_to_fp16 = const()[name = tensor("real_div_8_to_fp16"), val = tensor(0x1.008p+0)]; + tensor mul_8_cast_fp16 = mul(x = reduce_mean_73_cast_fp16, y = real_div_8_to_fp16)[name = tensor("mul_8_cast_fp16")]; + tensor var_1645_to_fp16 = const()[name = tensor("op_1645_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_73_cast_fp16_0 = add(x = mul_8_cast_fp16, y = var_1645_to_fp16)[name = tensor("var_73_cast_fp16")]; + tensor var_1648_epsilon_0 = const()[name = tensor("op_1648_epsilon_0"), val = tensor(0x1.197998p-40)]; + tensor var_1648_cast_fp16 = rsqrt(epsilon = var_1648_epsilon_0, x = var_73_cast_fp16_0)[name = tensor("op_1648_cast_fp16")]; + tensor var_1649_cast_fp16 = mul(x = const_3_to_fp16, y = var_1648_cast_fp16)[name = tensor("op_1649_cast_fp16")]; + tensor var_1650_cast_fp16 = mul(x = linear_106_cast_fp16, y = var_1649_cast_fp16)[name = tensor("op_1650_cast_fp16")]; + tensor input_281_to_fp16 = const()[name = tensor("input_281_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(19004992)))]; + tensor input_283_cast_fp16 = silu(x = input_281_to_fp16)[name = tensor("input_283_cast_fp16")]; + tensor linear_108_cast_fp16 = linear(bias = flow_net_time_embed_1_mlp_2_bias_to_fp16, weight = flow_net_time_embed_1_mlp_2_weight_to_fp16, x = input_283_cast_fp16)[name = tensor("linear_108_cast_fp16")]; + tensor reduce_mean_74_axes_0 = const()[name = tensor("reduce_mean_74_axes_0"), val = tensor([-1])]; + tensor reduce_mean_74_keep_dims_0 = const()[name = tensor("reduce_mean_74_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_74_cast_fp16 = reduce_mean(axes = reduce_mean_74_axes_0, keep_dims = reduce_mean_74_keep_dims_0, x = linear_108_cast_fp16)[name = tensor("reduce_mean_74_cast_fp16")]; + tensor sub_46_cast_fp16 = sub(x = linear_108_cast_fp16, y = reduce_mean_74_cast_fp16)[name = tensor("sub_46_cast_fp16")]; + tensor square_37_cast_fp16 = square(x = sub_46_cast_fp16)[name = tensor("square_37_cast_fp16")]; + tensor reduce_mean_75_axes_0 = const()[name = tensor("reduce_mean_75_axes_0"), val = tensor([-1])]; + tensor reduce_mean_75_keep_dims_0 = const()[name = tensor("reduce_mean_75_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_75_cast_fp16 = reduce_mean(axes = reduce_mean_75_axes_0, keep_dims = reduce_mean_75_keep_dims_0, x = square_37_cast_fp16)[name = tensor("reduce_mean_75_cast_fp16")]; + tensor real_div_9_to_fp16 = const()[name = tensor("real_div_9_to_fp16"), val = tensor(0x1.008p+0)]; + tensor mul_9_cast_fp16 = mul(x = reduce_mean_75_cast_fp16, y = real_div_9_to_fp16)[name = tensor("mul_9_cast_fp16")]; + tensor var_1682_to_fp16 = const()[name = tensor("op_1682_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_75_cast_fp16_0 = add(x = mul_9_cast_fp16, y = var_1682_to_fp16)[name = tensor("var_75_cast_fp16")]; + tensor var_1685_epsilon_0 = const()[name = tensor("op_1685_epsilon_0"), val = tensor(0x1.197998p-40)]; + tensor var_1685_cast_fp16 = rsqrt(epsilon = var_1685_epsilon_0, x = var_75_cast_fp16_0)[name = tensor("op_1685_cast_fp16")]; + tensor var_1686_cast_fp16 = mul(x = const_5_to_fp16, y = var_1685_cast_fp16)[name = tensor("op_1686_cast_fp16")]; + tensor var_1687_cast_fp16 = mul(x = linear_108_cast_fp16, y = var_1686_cast_fp16)[name = tensor("op_1687_cast_fp16")]; + tensor var_1699_cast_fp16 = add(x = var_1650_cast_fp16, y = var_1687_cast_fp16)[name = tensor("op_1699_cast_fp16")]; + tensor _inversed_t_combined_9_y_0_to_fp16 = const()[name = tensor("_inversed_t_combined_9_y_0_to_fp16"), val = tensor(0x1p-1)]; + tensor _inversed_t_combined_9_cast_fp16 = mul(x = var_1699_cast_fp16, y = _inversed_t_combined_9_y_0_to_fp16)[name = tensor("_inversed_t_combined_9_cast_fp16")]; + tensor input_285_cast_fp16 = add(x = _inversed_t_combined_9_cast_fp16, y = linear_5_cast_fp16)[name = tensor("input_285_cast_fp16")]; + tensor input_287_cast_fp16 = silu(x = input_285_cast_fp16)[name = tensor("input_287_cast_fp16")]; + tensor linear_110_cast_fp16 = linear(bias = flow_net_res_blocks_0_adaLN_modulation_1_bias_to_fp16, weight = flow_net_res_blocks_0_adaLN_modulation_1_weight_to_fp16, x = input_287_cast_fp16)[name = tensor("linear_110_cast_fp16")]; + tensor var_1714_split_sizes_0 = const()[name = tensor("op_1714_split_sizes_0"), val = tensor([512, 512, 512])]; + tensor var_1714_axis_0 = const()[name = tensor("op_1714_axis_0"), val = tensor(-1)]; + tensor var_1714_cast_fp16_0, tensor var_1714_cast_fp16_1, tensor var_1714_cast_fp16_2 = split(axis = var_1714_axis_0, split_sizes = var_1714_split_sizes_0, x = linear_110_cast_fp16)[name = tensor("op_1714_cast_fp16")]; + tensor mean_57_axes_0 = const()[name = tensor("mean_57_axes_0"), val = tensor([-1])]; + tensor mean_57_keep_dims_0 = const()[name = tensor("mean_57_keep_dims_0"), val = tensor(true)]; + tensor mean_57_cast_fp16 = reduce_mean(axes = mean_57_axes_0, keep_dims = mean_57_keep_dims_0, x = linear_104_cast_fp16)[name = tensor("mean_57_cast_fp16")]; + tensor sub_48_cast_fp16 = sub(x = linear_104_cast_fp16, y = mean_57_cast_fp16)[name = tensor("sub_48_cast_fp16")]; + tensor square_38_cast_fp16 = square(x = sub_48_cast_fp16)[name = tensor("square_38_cast_fp16")]; + tensor reduce_mean_77_axes_0 = const()[name = tensor("reduce_mean_77_axes_0"), val = tensor([-1])]; + tensor reduce_mean_77_keep_dims_0 = const()[name = tensor("reduce_mean_77_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_77_cast_fp16 = reduce_mean(axes = reduce_mean_77_axes_0, keep_dims = reduce_mean_77_keep_dims_0, x = square_38_cast_fp16)[name = tensor("reduce_mean_77_cast_fp16")]; + tensor var_1724_to_fp16 = const()[name = tensor("op_1724_to_fp16"), val = tensor(0x1.1p-20)]; + tensor var_1725_cast_fp16 = add(x = reduce_mean_77_cast_fp16, y = var_1724_to_fp16)[name = tensor("op_1725_cast_fp16")]; + tensor var_1726_cast_fp16 = sqrt(x = var_1725_cast_fp16)[name = tensor("op_1726_cast_fp16")]; + tensor x_183_cast_fp16 = real_div(x = sub_48_cast_fp16, y = var_1726_cast_fp16)[name = tensor("x_183_cast_fp16")]; + tensor var_1728_cast_fp16 = mul(x = x_183_cast_fp16, y = flow_net_res_blocks_0_in_ln_weight_to_fp16)[name = tensor("op_1728_cast_fp16")]; + tensor x_185_cast_fp16 = add(x = var_1728_cast_fp16, y = flow_net_res_blocks_0_in_ln_bias_to_fp16)[name = tensor("x_185_cast_fp16")]; + tensor var_1730_promoted_to_fp16 = const()[name = tensor("op_1730_promoted_to_fp16"), val = tensor(0x1p+0)]; + tensor var_1731_cast_fp16 = add(x = var_1714_cast_fp16_1, y = var_1730_promoted_to_fp16)[name = tensor("op_1731_cast_fp16")]; + tensor var_1732_cast_fp16 = mul(x = x_185_cast_fp16, y = var_1731_cast_fp16)[name = tensor("op_1732_cast_fp16")]; + tensor input_289_cast_fp16 = add(x = var_1732_cast_fp16, y = var_1714_cast_fp16_0)[name = tensor("input_289_cast_fp16")]; + tensor linear_111_cast_fp16 = linear(bias = flow_net_res_blocks_0_mlp_0_bias_to_fp16, weight = flow_net_res_blocks_0_mlp_0_weight_to_fp16, x = input_289_cast_fp16)[name = tensor("linear_111_cast_fp16")]; + tensor input_293_cast_fp16 = silu(x = linear_111_cast_fp16)[name = tensor("input_293_cast_fp16")]; + tensor linear_112_cast_fp16 = linear(bias = flow_net_res_blocks_0_mlp_2_bias_to_fp16, weight = flow_net_res_blocks_0_mlp_2_weight_to_fp16, x = input_293_cast_fp16)[name = tensor("linear_112_cast_fp16")]; + tensor var_1743_cast_fp16 = mul(x = var_1714_cast_fp16_2, y = linear_112_cast_fp16)[name = tensor("op_1743_cast_fp16")]; + tensor x_187_cast_fp16 = add(x = linear_104_cast_fp16, y = var_1743_cast_fp16)[name = tensor("x_187_cast_fp16")]; + tensor linear_113_cast_fp16 = linear(bias = flow_net_res_blocks_1_adaLN_modulation_1_bias_to_fp16, weight = flow_net_res_blocks_1_adaLN_modulation_1_weight_to_fp16, x = input_287_cast_fp16)[name = tensor("linear_113_cast_fp16")]; + tensor var_1753_split_sizes_0 = const()[name = tensor("op_1753_split_sizes_0"), val = tensor([512, 512, 512])]; + tensor var_1753_axis_0 = const()[name = tensor("op_1753_axis_0"), val = tensor(-1)]; + tensor var_1753_cast_fp16_0, tensor var_1753_cast_fp16_1, tensor var_1753_cast_fp16_2 = split(axis = var_1753_axis_0, split_sizes = var_1753_split_sizes_0, x = linear_113_cast_fp16)[name = tensor("op_1753_cast_fp16")]; + tensor mean_59_axes_0 = const()[name = tensor("mean_59_axes_0"), val = tensor([-1])]; + tensor mean_59_keep_dims_0 = const()[name = tensor("mean_59_keep_dims_0"), val = tensor(true)]; + tensor mean_59_cast_fp16 = reduce_mean(axes = mean_59_axes_0, keep_dims = mean_59_keep_dims_0, x = x_187_cast_fp16)[name = tensor("mean_59_cast_fp16")]; + tensor sub_49_cast_fp16 = sub(x = x_187_cast_fp16, y = mean_59_cast_fp16)[name = tensor("sub_49_cast_fp16")]; + tensor square_39_cast_fp16 = square(x = sub_49_cast_fp16)[name = tensor("square_39_cast_fp16")]; + tensor reduce_mean_79_axes_0 = const()[name = tensor("reduce_mean_79_axes_0"), val = tensor([-1])]; + tensor reduce_mean_79_keep_dims_0 = const()[name = tensor("reduce_mean_79_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_79_cast_fp16 = reduce_mean(axes = reduce_mean_79_axes_0, keep_dims = reduce_mean_79_keep_dims_0, x = square_39_cast_fp16)[name = tensor("reduce_mean_79_cast_fp16")]; + tensor var_1763_to_fp16 = const()[name = tensor("op_1763_to_fp16"), val = tensor(0x1.1p-20)]; + tensor var_1764_cast_fp16 = add(x = reduce_mean_79_cast_fp16, y = var_1763_to_fp16)[name = tensor("op_1764_cast_fp16")]; + tensor var_1765_cast_fp16 = sqrt(x = var_1764_cast_fp16)[name = tensor("op_1765_cast_fp16")]; + tensor x_189_cast_fp16 = real_div(x = sub_49_cast_fp16, y = var_1765_cast_fp16)[name = tensor("x_189_cast_fp16")]; + tensor var_1767_cast_fp16 = mul(x = x_189_cast_fp16, y = flow_net_res_blocks_1_in_ln_weight_to_fp16)[name = tensor("op_1767_cast_fp16")]; + tensor x_191_cast_fp16 = add(x = var_1767_cast_fp16, y = flow_net_res_blocks_1_in_ln_bias_to_fp16)[name = tensor("x_191_cast_fp16")]; + tensor var_1769_promoted_to_fp16 = const()[name = tensor("op_1769_promoted_to_fp16"), val = tensor(0x1p+0)]; + tensor var_1770_cast_fp16 = add(x = var_1753_cast_fp16_1, y = var_1769_promoted_to_fp16)[name = tensor("op_1770_cast_fp16")]; + tensor var_1771_cast_fp16 = mul(x = x_191_cast_fp16, y = var_1770_cast_fp16)[name = tensor("op_1771_cast_fp16")]; + tensor input_297_cast_fp16 = add(x = var_1771_cast_fp16, y = var_1753_cast_fp16_0)[name = tensor("input_297_cast_fp16")]; + tensor linear_114_cast_fp16 = linear(bias = flow_net_res_blocks_1_mlp_0_bias_to_fp16, weight = flow_net_res_blocks_1_mlp_0_weight_to_fp16, x = input_297_cast_fp16)[name = tensor("linear_114_cast_fp16")]; + tensor input_301_cast_fp16 = silu(x = linear_114_cast_fp16)[name = tensor("input_301_cast_fp16")]; + tensor linear_115_cast_fp16 = linear(bias = flow_net_res_blocks_1_mlp_2_bias_to_fp16, weight = flow_net_res_blocks_1_mlp_2_weight_to_fp16, x = input_301_cast_fp16)[name = tensor("linear_115_cast_fp16")]; + tensor var_1782_cast_fp16 = mul(x = var_1753_cast_fp16_2, y = linear_115_cast_fp16)[name = tensor("op_1782_cast_fp16")]; + tensor x_193_cast_fp16 = add(x = x_187_cast_fp16, y = var_1782_cast_fp16)[name = tensor("x_193_cast_fp16")]; + tensor linear_116_cast_fp16 = linear(bias = flow_net_res_blocks_2_adaLN_modulation_1_bias_to_fp16, weight = flow_net_res_blocks_2_adaLN_modulation_1_weight_to_fp16, x = input_287_cast_fp16)[name = tensor("linear_116_cast_fp16")]; + tensor var_1792_split_sizes_0 = const()[name = tensor("op_1792_split_sizes_0"), val = tensor([512, 512, 512])]; + tensor var_1792_axis_0 = const()[name = tensor("op_1792_axis_0"), val = tensor(-1)]; + tensor var_1792_cast_fp16_0, tensor var_1792_cast_fp16_1, tensor var_1792_cast_fp16_2 = split(axis = var_1792_axis_0, split_sizes = var_1792_split_sizes_0, x = linear_116_cast_fp16)[name = tensor("op_1792_cast_fp16")]; + tensor mean_61_axes_0 = const()[name = tensor("mean_61_axes_0"), val = tensor([-1])]; + tensor mean_61_keep_dims_0 = const()[name = tensor("mean_61_keep_dims_0"), val = tensor(true)]; + tensor mean_61_cast_fp16 = reduce_mean(axes = mean_61_axes_0, keep_dims = mean_61_keep_dims_0, x = x_193_cast_fp16)[name = tensor("mean_61_cast_fp16")]; + tensor sub_50_cast_fp16 = sub(x = x_193_cast_fp16, y = mean_61_cast_fp16)[name = tensor("sub_50_cast_fp16")]; + tensor square_40_cast_fp16 = square(x = sub_50_cast_fp16)[name = tensor("square_40_cast_fp16")]; + tensor reduce_mean_81_axes_0 = const()[name = tensor("reduce_mean_81_axes_0"), val = tensor([-1])]; + tensor reduce_mean_81_keep_dims_0 = const()[name = tensor("reduce_mean_81_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_81_cast_fp16 = reduce_mean(axes = reduce_mean_81_axes_0, keep_dims = reduce_mean_81_keep_dims_0, x = square_40_cast_fp16)[name = tensor("reduce_mean_81_cast_fp16")]; + tensor var_1802_to_fp16 = const()[name = tensor("op_1802_to_fp16"), val = tensor(0x1.1p-20)]; + tensor var_1803_cast_fp16 = add(x = reduce_mean_81_cast_fp16, y = var_1802_to_fp16)[name = tensor("op_1803_cast_fp16")]; + tensor var_1804_cast_fp16 = sqrt(x = var_1803_cast_fp16)[name = tensor("op_1804_cast_fp16")]; + tensor x_195_cast_fp16 = real_div(x = sub_50_cast_fp16, y = var_1804_cast_fp16)[name = tensor("x_195_cast_fp16")]; + tensor var_1806_cast_fp16 = mul(x = x_195_cast_fp16, y = flow_net_res_blocks_2_in_ln_weight_to_fp16)[name = tensor("op_1806_cast_fp16")]; + tensor x_197_cast_fp16 = add(x = var_1806_cast_fp16, y = flow_net_res_blocks_2_in_ln_bias_to_fp16)[name = tensor("x_197_cast_fp16")]; + tensor var_1808_promoted_to_fp16 = const()[name = tensor("op_1808_promoted_to_fp16"), val = tensor(0x1p+0)]; + tensor var_1809_cast_fp16 = add(x = var_1792_cast_fp16_1, y = var_1808_promoted_to_fp16)[name = tensor("op_1809_cast_fp16")]; + tensor var_1810_cast_fp16 = mul(x = x_197_cast_fp16, y = var_1809_cast_fp16)[name = tensor("op_1810_cast_fp16")]; + tensor input_305_cast_fp16 = add(x = var_1810_cast_fp16, y = var_1792_cast_fp16_0)[name = tensor("input_305_cast_fp16")]; + tensor linear_117_cast_fp16 = linear(bias = flow_net_res_blocks_2_mlp_0_bias_to_fp16, weight = flow_net_res_blocks_2_mlp_0_weight_to_fp16, x = input_305_cast_fp16)[name = tensor("linear_117_cast_fp16")]; + tensor input_309_cast_fp16 = silu(x = linear_117_cast_fp16)[name = tensor("input_309_cast_fp16")]; + tensor linear_118_cast_fp16 = linear(bias = flow_net_res_blocks_2_mlp_2_bias_to_fp16, weight = flow_net_res_blocks_2_mlp_2_weight_to_fp16, x = input_309_cast_fp16)[name = tensor("linear_118_cast_fp16")]; + tensor var_1821_cast_fp16 = mul(x = var_1792_cast_fp16_2, y = linear_118_cast_fp16)[name = tensor("op_1821_cast_fp16")]; + tensor x_199_cast_fp16 = add(x = x_193_cast_fp16, y = var_1821_cast_fp16)[name = tensor("x_199_cast_fp16")]; + tensor linear_119_cast_fp16 = linear(bias = flow_net_res_blocks_3_adaLN_modulation_1_bias_to_fp16, weight = flow_net_res_blocks_3_adaLN_modulation_1_weight_to_fp16, x = input_287_cast_fp16)[name = tensor("linear_119_cast_fp16")]; + tensor var_1831_split_sizes_0 = const()[name = tensor("op_1831_split_sizes_0"), val = tensor([512, 512, 512])]; + tensor var_1831_axis_0 = const()[name = tensor("op_1831_axis_0"), val = tensor(-1)]; + tensor var_1831_cast_fp16_0, tensor var_1831_cast_fp16_1, tensor var_1831_cast_fp16_2 = split(axis = var_1831_axis_0, split_sizes = var_1831_split_sizes_0, x = linear_119_cast_fp16)[name = tensor("op_1831_cast_fp16")]; + tensor mean_63_axes_0 = const()[name = tensor("mean_63_axes_0"), val = tensor([-1])]; + tensor mean_63_keep_dims_0 = const()[name = tensor("mean_63_keep_dims_0"), val = tensor(true)]; + tensor mean_63_cast_fp16 = reduce_mean(axes = mean_63_axes_0, keep_dims = mean_63_keep_dims_0, x = x_199_cast_fp16)[name = tensor("mean_63_cast_fp16")]; + tensor sub_51_cast_fp16 = sub(x = x_199_cast_fp16, y = mean_63_cast_fp16)[name = tensor("sub_51_cast_fp16")]; + tensor square_41_cast_fp16 = square(x = sub_51_cast_fp16)[name = tensor("square_41_cast_fp16")]; + tensor reduce_mean_83_axes_0 = const()[name = tensor("reduce_mean_83_axes_0"), val = tensor([-1])]; + tensor reduce_mean_83_keep_dims_0 = const()[name = tensor("reduce_mean_83_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_83_cast_fp16 = reduce_mean(axes = reduce_mean_83_axes_0, keep_dims = reduce_mean_83_keep_dims_0, x = square_41_cast_fp16)[name = tensor("reduce_mean_83_cast_fp16")]; + tensor var_1841_to_fp16 = const()[name = tensor("op_1841_to_fp16"), val = tensor(0x1.1p-20)]; + tensor var_1842_cast_fp16 = add(x = reduce_mean_83_cast_fp16, y = var_1841_to_fp16)[name = tensor("op_1842_cast_fp16")]; + tensor var_1843_cast_fp16 = sqrt(x = var_1842_cast_fp16)[name = tensor("op_1843_cast_fp16")]; + tensor x_201_cast_fp16 = real_div(x = sub_51_cast_fp16, y = var_1843_cast_fp16)[name = tensor("x_201_cast_fp16")]; + tensor var_1845_cast_fp16 = mul(x = x_201_cast_fp16, y = flow_net_res_blocks_3_in_ln_weight_to_fp16)[name = tensor("op_1845_cast_fp16")]; + tensor x_203_cast_fp16 = add(x = var_1845_cast_fp16, y = flow_net_res_blocks_3_in_ln_bias_to_fp16)[name = tensor("x_203_cast_fp16")]; + tensor var_1847_promoted_to_fp16 = const()[name = tensor("op_1847_promoted_to_fp16"), val = tensor(0x1p+0)]; + tensor var_1848_cast_fp16 = add(x = var_1831_cast_fp16_1, y = var_1847_promoted_to_fp16)[name = tensor("op_1848_cast_fp16")]; + tensor var_1849_cast_fp16 = mul(x = x_203_cast_fp16, y = var_1848_cast_fp16)[name = tensor("op_1849_cast_fp16")]; + tensor input_313_cast_fp16 = add(x = var_1849_cast_fp16, y = var_1831_cast_fp16_0)[name = tensor("input_313_cast_fp16")]; + tensor linear_120_cast_fp16 = linear(bias = flow_net_res_blocks_3_mlp_0_bias_to_fp16, weight = flow_net_res_blocks_3_mlp_0_weight_to_fp16, x = input_313_cast_fp16)[name = tensor("linear_120_cast_fp16")]; + tensor input_317_cast_fp16 = silu(x = linear_120_cast_fp16)[name = tensor("input_317_cast_fp16")]; + tensor linear_121_cast_fp16 = linear(bias = flow_net_res_blocks_3_mlp_2_bias_to_fp16, weight = flow_net_res_blocks_3_mlp_2_weight_to_fp16, x = input_317_cast_fp16)[name = tensor("linear_121_cast_fp16")]; + tensor var_1860_cast_fp16 = mul(x = var_1831_cast_fp16_2, y = linear_121_cast_fp16)[name = tensor("op_1860_cast_fp16")]; + tensor x_205_cast_fp16 = add(x = x_199_cast_fp16, y = var_1860_cast_fp16)[name = tensor("x_205_cast_fp16")]; + tensor linear_122_cast_fp16 = linear(bias = flow_net_res_blocks_4_adaLN_modulation_1_bias_to_fp16, weight = flow_net_res_blocks_4_adaLN_modulation_1_weight_to_fp16, x = input_287_cast_fp16)[name = tensor("linear_122_cast_fp16")]; + tensor var_1870_split_sizes_0 = const()[name = tensor("op_1870_split_sizes_0"), val = tensor([512, 512, 512])]; + tensor var_1870_axis_0 = const()[name = tensor("op_1870_axis_0"), val = tensor(-1)]; + tensor var_1870_cast_fp16_0, tensor var_1870_cast_fp16_1, tensor var_1870_cast_fp16_2 = split(axis = var_1870_axis_0, split_sizes = var_1870_split_sizes_0, x = linear_122_cast_fp16)[name = tensor("op_1870_cast_fp16")]; + tensor mean_65_axes_0 = const()[name = tensor("mean_65_axes_0"), val = tensor([-1])]; + tensor mean_65_keep_dims_0 = const()[name = tensor("mean_65_keep_dims_0"), val = tensor(true)]; + tensor mean_65_cast_fp16 = reduce_mean(axes = mean_65_axes_0, keep_dims = mean_65_keep_dims_0, x = x_205_cast_fp16)[name = tensor("mean_65_cast_fp16")]; + tensor sub_52_cast_fp16 = sub(x = x_205_cast_fp16, y = mean_65_cast_fp16)[name = tensor("sub_52_cast_fp16")]; + tensor square_42_cast_fp16 = square(x = sub_52_cast_fp16)[name = tensor("square_42_cast_fp16")]; + tensor reduce_mean_85_axes_0 = const()[name = tensor("reduce_mean_85_axes_0"), val = tensor([-1])]; + tensor reduce_mean_85_keep_dims_0 = const()[name = tensor("reduce_mean_85_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_85_cast_fp16 = reduce_mean(axes = reduce_mean_85_axes_0, keep_dims = reduce_mean_85_keep_dims_0, x = square_42_cast_fp16)[name = tensor("reduce_mean_85_cast_fp16")]; + tensor var_1880_to_fp16 = const()[name = tensor("op_1880_to_fp16"), val = tensor(0x1.1p-20)]; + tensor var_1881_cast_fp16 = add(x = reduce_mean_85_cast_fp16, y = var_1880_to_fp16)[name = tensor("op_1881_cast_fp16")]; + tensor var_1882_cast_fp16 = sqrt(x = var_1881_cast_fp16)[name = tensor("op_1882_cast_fp16")]; + tensor x_207_cast_fp16 = real_div(x = sub_52_cast_fp16, y = var_1882_cast_fp16)[name = tensor("x_207_cast_fp16")]; + tensor var_1884_cast_fp16 = mul(x = x_207_cast_fp16, y = flow_net_res_blocks_4_in_ln_weight_to_fp16)[name = tensor("op_1884_cast_fp16")]; + tensor x_209_cast_fp16 = add(x = var_1884_cast_fp16, y = flow_net_res_blocks_4_in_ln_bias_to_fp16)[name = tensor("x_209_cast_fp16")]; + tensor var_1886_promoted_to_fp16 = const()[name = tensor("op_1886_promoted_to_fp16"), val = tensor(0x1p+0)]; + tensor var_1887_cast_fp16 = add(x = var_1870_cast_fp16_1, y = var_1886_promoted_to_fp16)[name = tensor("op_1887_cast_fp16")]; + tensor var_1888_cast_fp16 = mul(x = x_209_cast_fp16, y = var_1887_cast_fp16)[name = tensor("op_1888_cast_fp16")]; + tensor input_321_cast_fp16 = add(x = var_1888_cast_fp16, y = var_1870_cast_fp16_0)[name = tensor("input_321_cast_fp16")]; + tensor linear_123_cast_fp16 = linear(bias = flow_net_res_blocks_4_mlp_0_bias_to_fp16, weight = flow_net_res_blocks_4_mlp_0_weight_to_fp16, x = input_321_cast_fp16)[name = tensor("linear_123_cast_fp16")]; + tensor input_325_cast_fp16 = silu(x = linear_123_cast_fp16)[name = tensor("input_325_cast_fp16")]; + tensor linear_124_cast_fp16 = linear(bias = flow_net_res_blocks_4_mlp_2_bias_to_fp16, weight = flow_net_res_blocks_4_mlp_2_weight_to_fp16, x = input_325_cast_fp16)[name = tensor("linear_124_cast_fp16")]; + tensor var_1899_cast_fp16 = mul(x = var_1870_cast_fp16_2, y = linear_124_cast_fp16)[name = tensor("op_1899_cast_fp16")]; + tensor x_211_cast_fp16 = add(x = x_205_cast_fp16, y = var_1899_cast_fp16)[name = tensor("x_211_cast_fp16")]; + tensor linear_125_cast_fp16 = linear(bias = flow_net_res_blocks_5_adaLN_modulation_1_bias_to_fp16, weight = flow_net_res_blocks_5_adaLN_modulation_1_weight_to_fp16, x = input_287_cast_fp16)[name = tensor("linear_125_cast_fp16")]; + tensor var_1909_split_sizes_0 = const()[name = tensor("op_1909_split_sizes_0"), val = tensor([512, 512, 512])]; + tensor var_1909_axis_0 = const()[name = tensor("op_1909_axis_0"), val = tensor(-1)]; + tensor var_1909_cast_fp16_0, tensor var_1909_cast_fp16_1, tensor var_1909_cast_fp16_2 = split(axis = var_1909_axis_0, split_sizes = var_1909_split_sizes_0, x = linear_125_cast_fp16)[name = tensor("op_1909_cast_fp16")]; + tensor mean_67_axes_0 = const()[name = tensor("mean_67_axes_0"), val = tensor([-1])]; + tensor mean_67_keep_dims_0 = const()[name = tensor("mean_67_keep_dims_0"), val = tensor(true)]; + tensor mean_67_cast_fp16 = reduce_mean(axes = mean_67_axes_0, keep_dims = mean_67_keep_dims_0, x = x_211_cast_fp16)[name = tensor("mean_67_cast_fp16")]; + tensor sub_53_cast_fp16 = sub(x = x_211_cast_fp16, y = mean_67_cast_fp16)[name = tensor("sub_53_cast_fp16")]; + tensor square_43_cast_fp16 = square(x = sub_53_cast_fp16)[name = tensor("square_43_cast_fp16")]; + tensor reduce_mean_87_axes_0 = const()[name = tensor("reduce_mean_87_axes_0"), val = tensor([-1])]; + tensor reduce_mean_87_keep_dims_0 = const()[name = tensor("reduce_mean_87_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_87_cast_fp16 = reduce_mean(axes = reduce_mean_87_axes_0, keep_dims = reduce_mean_87_keep_dims_0, x = square_43_cast_fp16)[name = tensor("reduce_mean_87_cast_fp16")]; + tensor var_1919_to_fp16 = const()[name = tensor("op_1919_to_fp16"), val = tensor(0x1.1p-20)]; + tensor var_1920_cast_fp16 = add(x = reduce_mean_87_cast_fp16, y = var_1919_to_fp16)[name = tensor("op_1920_cast_fp16")]; + tensor var_1921_cast_fp16 = sqrt(x = var_1920_cast_fp16)[name = tensor("op_1921_cast_fp16")]; + tensor x_213_cast_fp16 = real_div(x = sub_53_cast_fp16, y = var_1921_cast_fp16)[name = tensor("x_213_cast_fp16")]; + tensor var_1923_cast_fp16 = mul(x = x_213_cast_fp16, y = flow_net_res_blocks_5_in_ln_weight_to_fp16)[name = tensor("op_1923_cast_fp16")]; + tensor x_215_cast_fp16 = add(x = var_1923_cast_fp16, y = flow_net_res_blocks_5_in_ln_bias_to_fp16)[name = tensor("x_215_cast_fp16")]; + tensor var_1925_promoted_to_fp16 = const()[name = tensor("op_1925_promoted_to_fp16"), val = tensor(0x1p+0)]; + tensor var_1926_cast_fp16 = add(x = var_1909_cast_fp16_1, y = var_1925_promoted_to_fp16)[name = tensor("op_1926_cast_fp16")]; + tensor var_1927_cast_fp16 = mul(x = x_215_cast_fp16, y = var_1926_cast_fp16)[name = tensor("op_1927_cast_fp16")]; + tensor input_329_cast_fp16 = add(x = var_1927_cast_fp16, y = var_1909_cast_fp16_0)[name = tensor("input_329_cast_fp16")]; + tensor linear_126_cast_fp16 = linear(bias = flow_net_res_blocks_5_mlp_0_bias_to_fp16, weight = flow_net_res_blocks_5_mlp_0_weight_to_fp16, x = input_329_cast_fp16)[name = tensor("linear_126_cast_fp16")]; + tensor input_333_cast_fp16 = silu(x = linear_126_cast_fp16)[name = tensor("input_333_cast_fp16")]; + tensor linear_127_cast_fp16 = linear(bias = flow_net_res_blocks_5_mlp_2_bias_to_fp16, weight = flow_net_res_blocks_5_mlp_2_weight_to_fp16, x = input_333_cast_fp16)[name = tensor("linear_127_cast_fp16")]; + tensor var_1938_cast_fp16 = mul(x = var_1909_cast_fp16_2, y = linear_127_cast_fp16)[name = tensor("op_1938_cast_fp16")]; + tensor x_217_cast_fp16 = add(x = x_211_cast_fp16, y = var_1938_cast_fp16)[name = tensor("x_217_cast_fp16")]; + tensor linear_128_cast_fp16 = linear(bias = flow_net_final_layer_adaLN_modulation_1_bias_to_fp16, weight = flow_net_final_layer_adaLN_modulation_1_weight_to_fp16, x = input_287_cast_fp16)[name = tensor("linear_128_cast_fp16")]; + tensor var_1947_split_sizes_0 = const()[name = tensor("op_1947_split_sizes_0"), val = tensor([512, 512])]; + tensor var_1947_axis_0 = const()[name = tensor("op_1947_axis_0"), val = tensor(-1)]; + tensor var_1947_cast_fp16_0, tensor var_1947_cast_fp16_1 = split(axis = var_1947_axis_0, split_sizes = var_1947_split_sizes_0, x = linear_128_cast_fp16)[name = tensor("op_1947_cast_fp16")]; + tensor mean_69_axes_0 = const()[name = tensor("mean_69_axes_0"), val = tensor([-1])]; + tensor mean_69_keep_dims_0 = const()[name = tensor("mean_69_keep_dims_0"), val = tensor(true)]; + tensor mean_69_cast_fp16 = reduce_mean(axes = mean_69_axes_0, keep_dims = mean_69_keep_dims_0, x = x_217_cast_fp16)[name = tensor("mean_69_cast_fp16")]; + tensor sub_54_cast_fp16 = sub(x = x_217_cast_fp16, y = mean_69_cast_fp16)[name = tensor("sub_54_cast_fp16")]; + tensor square_44_cast_fp16 = square(x = sub_54_cast_fp16)[name = tensor("square_44_cast_fp16")]; + tensor reduce_mean_89_axes_0 = const()[name = tensor("reduce_mean_89_axes_0"), val = tensor([-1])]; + tensor reduce_mean_89_keep_dims_0 = const()[name = tensor("reduce_mean_89_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_89_cast_fp16 = reduce_mean(axes = reduce_mean_89_axes_0, keep_dims = reduce_mean_89_keep_dims_0, x = square_44_cast_fp16)[name = tensor("reduce_mean_89_cast_fp16")]; + tensor var_1954_to_fp16 = const()[name = tensor("op_1954_to_fp16"), val = tensor(0x1.1p-20)]; + tensor var_1955_cast_fp16 = add(x = reduce_mean_89_cast_fp16, y = var_1954_to_fp16)[name = tensor("op_1955_cast_fp16")]; + tensor var_1956_cast_fp16 = sqrt(x = var_1955_cast_fp16)[name = tensor("op_1956_cast_fp16")]; + tensor x_219_cast_fp16 = real_div(x = sub_54_cast_fp16, y = var_1956_cast_fp16)[name = tensor("x_219_cast_fp16")]; + tensor var_1958_promoted_to_fp16 = const()[name = tensor("op_1958_promoted_to_fp16"), val = tensor(0x1p+0)]; + tensor var_1959_cast_fp16 = add(x = var_1947_cast_fp16_1, y = var_1958_promoted_to_fp16)[name = tensor("op_1959_cast_fp16")]; + tensor var_1960_cast_fp16 = mul(x = x_219_cast_fp16, y = var_1959_cast_fp16)[name = tensor("op_1960_cast_fp16")]; + tensor input_337_cast_fp16 = add(x = var_1960_cast_fp16, y = var_1947_cast_fp16_0)[name = tensor("input_337_cast_fp16")]; + tensor linear_129_cast_fp16 = linear(bias = flow_net_final_layer_linear_bias_to_fp16, weight = flow_net_final_layer_linear_weight_to_fp16, x = input_337_cast_fp16)[name = tensor("linear_129_cast_fp16")]; + tensor var_1971_to_fp16 = const()[name = tensor("op_1971_to_fp16"), val = tensor(0x1p-3)]; + tensor var_1972_cast_fp16 = mul(x = linear_129_cast_fp16, y = var_1971_to_fp16)[name = tensor("op_1972_cast_fp16")]; + tensor input_339_cast_fp16 = add(x = input_271_cast_fp16, y = var_1972_cast_fp16)[name = tensor("input_339_cast_fp16")]; + tensor linear_130_cast_fp16 = linear(bias = flow_net_input_proj_bias_to_fp16, weight = flow_net_input_proj_weight_to_fp16, x = input_339_cast_fp16)[name = tensor("linear_130_cast_fp16")]; + tensor input_343_to_fp16 = const()[name = tensor("input_343_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(19006080)))]; + tensor input_345_cast_fp16 = silu(x = input_343_to_fp16)[name = tensor("input_345_cast_fp16")]; + tensor linear_132_cast_fp16 = linear(bias = flow_net_time_embed_0_mlp_2_bias_to_fp16, weight = flow_net_time_embed_0_mlp_2_weight_to_fp16, x = input_345_cast_fp16)[name = tensor("linear_132_cast_fp16")]; + tensor reduce_mean_90_axes_0 = const()[name = tensor("reduce_mean_90_axes_0"), val = tensor([-1])]; + tensor reduce_mean_90_keep_dims_0 = const()[name = tensor("reduce_mean_90_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_90_cast_fp16 = reduce_mean(axes = reduce_mean_90_axes_0, keep_dims = reduce_mean_90_keep_dims_0, x = linear_132_cast_fp16)[name = tensor("reduce_mean_90_cast_fp16")]; + tensor sub_55_cast_fp16 = sub(x = linear_132_cast_fp16, y = reduce_mean_90_cast_fp16)[name = tensor("sub_55_cast_fp16")]; + tensor square_45_cast_fp16 = square(x = sub_55_cast_fp16)[name = tensor("square_45_cast_fp16")]; + tensor reduce_mean_91_axes_0 = const()[name = tensor("reduce_mean_91_axes_0"), val = tensor([-1])]; + tensor reduce_mean_91_keep_dims_0 = const()[name = tensor("reduce_mean_91_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_91_cast_fp16 = reduce_mean(axes = reduce_mean_91_axes_0, keep_dims = reduce_mean_91_keep_dims_0, x = square_45_cast_fp16)[name = tensor("reduce_mean_91_cast_fp16")]; + tensor real_div_10_to_fp16 = const()[name = tensor("real_div_10_to_fp16"), val = tensor(0x1.008p+0)]; + tensor mul_10_cast_fp16 = mul(x = reduce_mean_91_cast_fp16, y = real_div_10_to_fp16)[name = tensor("mul_10_cast_fp16")]; + tensor var_2038_to_fp16 = const()[name = tensor("op_2038_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_91_cast_fp16 = add(x = mul_10_cast_fp16, y = var_2038_to_fp16)[name = tensor("var_91_cast_fp16")]; + tensor var_2041_epsilon_0 = const()[name = tensor("op_2041_epsilon_0"), val = tensor(0x1.197998p-40)]; + tensor var_2041_cast_fp16 = rsqrt(epsilon = var_2041_epsilon_0, x = var_91_cast_fp16)[name = tensor("op_2041_cast_fp16")]; + tensor var_2042_cast_fp16 = mul(x = const_3_to_fp16, y = var_2041_cast_fp16)[name = tensor("op_2042_cast_fp16")]; + tensor var_2043_cast_fp16 = mul(x = linear_132_cast_fp16, y = var_2042_cast_fp16)[name = tensor("op_2043_cast_fp16")]; + tensor input_349_to_fp16 = const()[name = tensor("input_349_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(19007168)))]; + tensor input_351_cast_fp16 = silu(x = input_349_to_fp16)[name = tensor("input_351_cast_fp16")]; + tensor linear_134_cast_fp16 = linear(bias = flow_net_time_embed_1_mlp_2_bias_to_fp16, weight = flow_net_time_embed_1_mlp_2_weight_to_fp16, x = input_351_cast_fp16)[name = tensor("linear_134_cast_fp16")]; + tensor reduce_mean_92_axes_0 = const()[name = tensor("reduce_mean_92_axes_0"), val = tensor([-1])]; + tensor reduce_mean_92_keep_dims_0 = const()[name = tensor("reduce_mean_92_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_92_cast_fp16 = reduce_mean(axes = reduce_mean_92_axes_0, keep_dims = reduce_mean_92_keep_dims_0, x = linear_134_cast_fp16)[name = tensor("reduce_mean_92_cast_fp16")]; + tensor sub_57_cast_fp16 = sub(x = linear_134_cast_fp16, y = reduce_mean_92_cast_fp16)[name = tensor("sub_57_cast_fp16")]; + tensor square_46_cast_fp16 = square(x = sub_57_cast_fp16)[name = tensor("square_46_cast_fp16")]; + tensor reduce_mean_93_axes_0 = const()[name = tensor("reduce_mean_93_axes_0"), val = tensor([-1])]; + tensor reduce_mean_93_keep_dims_0 = const()[name = tensor("reduce_mean_93_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_93_cast_fp16 = reduce_mean(axes = reduce_mean_93_axes_0, keep_dims = reduce_mean_93_keep_dims_0, x = square_46_cast_fp16)[name = tensor("reduce_mean_93_cast_fp16")]; + tensor real_div_11_to_fp16 = const()[name = tensor("real_div_11_to_fp16"), val = tensor(0x1.008p+0)]; + tensor mul_11_cast_fp16 = mul(x = reduce_mean_93_cast_fp16, y = real_div_11_to_fp16)[name = tensor("mul_11_cast_fp16")]; + tensor var_2075_to_fp16 = const()[name = tensor("op_2075_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_93_cast_fp16 = add(x = mul_11_cast_fp16, y = var_2075_to_fp16)[name = tensor("var_93_cast_fp16")]; + tensor var_2078_epsilon_0 = const()[name = tensor("op_2078_epsilon_0"), val = tensor(0x1.197998p-40)]; + tensor var_2078_cast_fp16 = rsqrt(epsilon = var_2078_epsilon_0, x = var_93_cast_fp16)[name = tensor("op_2078_cast_fp16")]; + tensor var_2079_cast_fp16 = mul(x = const_5_to_fp16, y = var_2078_cast_fp16)[name = tensor("op_2079_cast_fp16")]; + tensor var_2080_cast_fp16 = mul(x = linear_134_cast_fp16, y = var_2079_cast_fp16)[name = tensor("op_2080_cast_fp16")]; + tensor var_2092_cast_fp16 = add(x = var_2043_cast_fp16, y = var_2080_cast_fp16)[name = tensor("op_2092_cast_fp16")]; + tensor _inversed_t_combined_11_y_0_to_fp16 = const()[name = tensor("_inversed_t_combined_11_y_0_to_fp16"), val = tensor(0x1p-1)]; + tensor _inversed_t_combined_11_cast_fp16 = mul(x = var_2092_cast_fp16, y = _inversed_t_combined_11_y_0_to_fp16)[name = tensor("_inversed_t_combined_11_cast_fp16")]; + tensor input_353_cast_fp16 = add(x = _inversed_t_combined_11_cast_fp16, y = linear_5_cast_fp16)[name = tensor("input_353_cast_fp16")]; + tensor input_355_cast_fp16 = silu(x = input_353_cast_fp16)[name = tensor("input_355_cast_fp16")]; + tensor linear_136_cast_fp16 = linear(bias = flow_net_res_blocks_0_adaLN_modulation_1_bias_to_fp16, weight = flow_net_res_blocks_0_adaLN_modulation_1_weight_to_fp16, x = input_355_cast_fp16)[name = tensor("linear_136_cast_fp16")]; + tensor var_2107_split_sizes_0 = const()[name = tensor("op_2107_split_sizes_0"), val = tensor([512, 512, 512])]; + tensor var_2107_axis_0 = const()[name = tensor("op_2107_axis_0"), val = tensor(-1)]; + tensor var_2107_cast_fp16_0, tensor var_2107_cast_fp16_1, tensor var_2107_cast_fp16_2 = split(axis = var_2107_axis_0, split_sizes = var_2107_split_sizes_0, x = linear_136_cast_fp16)[name = tensor("op_2107_cast_fp16")]; + tensor mean_71_axes_0 = const()[name = tensor("mean_71_axes_0"), val = tensor([-1])]; + tensor mean_71_keep_dims_0 = const()[name = tensor("mean_71_keep_dims_0"), val = tensor(true)]; + tensor mean_71_cast_fp16 = reduce_mean(axes = mean_71_axes_0, keep_dims = mean_71_keep_dims_0, x = linear_130_cast_fp16)[name = tensor("mean_71_cast_fp16")]; + tensor sub_59_cast_fp16 = sub(x = linear_130_cast_fp16, y = mean_71_cast_fp16)[name = tensor("sub_59_cast_fp16")]; + tensor square_47_cast_fp16 = square(x = sub_59_cast_fp16)[name = tensor("square_47_cast_fp16")]; + tensor reduce_mean_95_axes_0 = const()[name = tensor("reduce_mean_95_axes_0"), val = tensor([-1])]; + tensor reduce_mean_95_keep_dims_0 = const()[name = tensor("reduce_mean_95_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_95_cast_fp16 = reduce_mean(axes = reduce_mean_95_axes_0, keep_dims = reduce_mean_95_keep_dims_0, x = square_47_cast_fp16)[name = tensor("reduce_mean_95_cast_fp16")]; + tensor var_2117_to_fp16 = const()[name = tensor("op_2117_to_fp16"), val = tensor(0x1.1p-20)]; + tensor var_2118_cast_fp16 = add(x = reduce_mean_95_cast_fp16, y = var_2117_to_fp16)[name = tensor("op_2118_cast_fp16")]; + tensor var_2119_cast_fp16 = sqrt(x = var_2118_cast_fp16)[name = tensor("op_2119_cast_fp16")]; + tensor x_227_cast_fp16 = real_div(x = sub_59_cast_fp16, y = var_2119_cast_fp16)[name = tensor("x_227_cast_fp16")]; + tensor var_2121_cast_fp16 = mul(x = x_227_cast_fp16, y = flow_net_res_blocks_0_in_ln_weight_to_fp16)[name = tensor("op_2121_cast_fp16")]; + tensor x_229_cast_fp16 = add(x = var_2121_cast_fp16, y = flow_net_res_blocks_0_in_ln_bias_to_fp16)[name = tensor("x_229_cast_fp16")]; + tensor var_2123_promoted_to_fp16 = const()[name = tensor("op_2123_promoted_to_fp16"), val = tensor(0x1p+0)]; + tensor var_2124_cast_fp16 = add(x = var_2107_cast_fp16_1, y = var_2123_promoted_to_fp16)[name = tensor("op_2124_cast_fp16")]; + tensor var_2125_cast_fp16 = mul(x = x_229_cast_fp16, y = var_2124_cast_fp16)[name = tensor("op_2125_cast_fp16")]; + tensor input_357_cast_fp16 = add(x = var_2125_cast_fp16, y = var_2107_cast_fp16_0)[name = tensor("input_357_cast_fp16")]; + tensor linear_137_cast_fp16 = linear(bias = flow_net_res_blocks_0_mlp_0_bias_to_fp16, weight = flow_net_res_blocks_0_mlp_0_weight_to_fp16, x = input_357_cast_fp16)[name = tensor("linear_137_cast_fp16")]; + tensor input_361_cast_fp16 = silu(x = linear_137_cast_fp16)[name = tensor("input_361_cast_fp16")]; + tensor linear_138_cast_fp16 = linear(bias = flow_net_res_blocks_0_mlp_2_bias_to_fp16, weight = flow_net_res_blocks_0_mlp_2_weight_to_fp16, x = input_361_cast_fp16)[name = tensor("linear_138_cast_fp16")]; + tensor var_2136_cast_fp16 = mul(x = var_2107_cast_fp16_2, y = linear_138_cast_fp16)[name = tensor("op_2136_cast_fp16")]; + tensor x_231_cast_fp16 = add(x = linear_130_cast_fp16, y = var_2136_cast_fp16)[name = tensor("x_231_cast_fp16")]; + tensor linear_139_cast_fp16 = linear(bias = flow_net_res_blocks_1_adaLN_modulation_1_bias_to_fp16, weight = flow_net_res_blocks_1_adaLN_modulation_1_weight_to_fp16, x = input_355_cast_fp16)[name = tensor("linear_139_cast_fp16")]; + tensor var_2146_split_sizes_0 = const()[name = tensor("op_2146_split_sizes_0"), val = tensor([512, 512, 512])]; + tensor var_2146_axis_0 = const()[name = tensor("op_2146_axis_0"), val = tensor(-1)]; + tensor var_2146_cast_fp16_0, tensor var_2146_cast_fp16_1, tensor var_2146_cast_fp16_2 = split(axis = var_2146_axis_0, split_sizes = var_2146_split_sizes_0, x = linear_139_cast_fp16)[name = tensor("op_2146_cast_fp16")]; + tensor mean_73_axes_0 = const()[name = tensor("mean_73_axes_0"), val = tensor([-1])]; + tensor mean_73_keep_dims_0 = const()[name = tensor("mean_73_keep_dims_0"), val = tensor(true)]; + tensor mean_73_cast_fp16 = reduce_mean(axes = mean_73_axes_0, keep_dims = mean_73_keep_dims_0, x = x_231_cast_fp16)[name = tensor("mean_73_cast_fp16")]; + tensor sub_60_cast_fp16 = sub(x = x_231_cast_fp16, y = mean_73_cast_fp16)[name = tensor("sub_60_cast_fp16")]; + tensor square_48_cast_fp16 = square(x = sub_60_cast_fp16)[name = tensor("square_48_cast_fp16")]; + tensor reduce_mean_97_axes_0 = const()[name = tensor("reduce_mean_97_axes_0"), val = tensor([-1])]; + tensor reduce_mean_97_keep_dims_0 = const()[name = tensor("reduce_mean_97_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_97_cast_fp16 = reduce_mean(axes = reduce_mean_97_axes_0, keep_dims = reduce_mean_97_keep_dims_0, x = square_48_cast_fp16)[name = tensor("reduce_mean_97_cast_fp16")]; + tensor var_2156_to_fp16 = const()[name = tensor("op_2156_to_fp16"), val = tensor(0x1.1p-20)]; + tensor var_2157_cast_fp16 = add(x = reduce_mean_97_cast_fp16, y = var_2156_to_fp16)[name = tensor("op_2157_cast_fp16")]; + tensor var_2158_cast_fp16 = sqrt(x = var_2157_cast_fp16)[name = tensor("op_2158_cast_fp16")]; + tensor x_233_cast_fp16 = real_div(x = sub_60_cast_fp16, y = var_2158_cast_fp16)[name = tensor("x_233_cast_fp16")]; + tensor var_2160_cast_fp16 = mul(x = x_233_cast_fp16, y = flow_net_res_blocks_1_in_ln_weight_to_fp16)[name = tensor("op_2160_cast_fp16")]; + tensor x_235_cast_fp16 = add(x = var_2160_cast_fp16, y = flow_net_res_blocks_1_in_ln_bias_to_fp16)[name = tensor("x_235_cast_fp16")]; + tensor var_2162_promoted_to_fp16 = const()[name = tensor("op_2162_promoted_to_fp16"), val = tensor(0x1p+0)]; + tensor var_2163_cast_fp16 = add(x = var_2146_cast_fp16_1, y = var_2162_promoted_to_fp16)[name = tensor("op_2163_cast_fp16")]; + tensor var_2164_cast_fp16 = mul(x = x_235_cast_fp16, y = var_2163_cast_fp16)[name = tensor("op_2164_cast_fp16")]; + tensor input_365_cast_fp16 = add(x = var_2164_cast_fp16, y = var_2146_cast_fp16_0)[name = tensor("input_365_cast_fp16")]; + tensor linear_140_cast_fp16 = linear(bias = flow_net_res_blocks_1_mlp_0_bias_to_fp16, weight = flow_net_res_blocks_1_mlp_0_weight_to_fp16, x = input_365_cast_fp16)[name = tensor("linear_140_cast_fp16")]; + tensor input_369_cast_fp16 = silu(x = linear_140_cast_fp16)[name = tensor("input_369_cast_fp16")]; + tensor linear_141_cast_fp16 = linear(bias = flow_net_res_blocks_1_mlp_2_bias_to_fp16, weight = flow_net_res_blocks_1_mlp_2_weight_to_fp16, x = input_369_cast_fp16)[name = tensor("linear_141_cast_fp16")]; + tensor var_2175_cast_fp16 = mul(x = var_2146_cast_fp16_2, y = linear_141_cast_fp16)[name = tensor("op_2175_cast_fp16")]; + tensor x_237_cast_fp16 = add(x = x_231_cast_fp16, y = var_2175_cast_fp16)[name = tensor("x_237_cast_fp16")]; + tensor linear_142_cast_fp16 = linear(bias = flow_net_res_blocks_2_adaLN_modulation_1_bias_to_fp16, weight = flow_net_res_blocks_2_adaLN_modulation_1_weight_to_fp16, x = input_355_cast_fp16)[name = tensor("linear_142_cast_fp16")]; + tensor var_2185_split_sizes_0 = const()[name = tensor("op_2185_split_sizes_0"), val = tensor([512, 512, 512])]; + tensor var_2185_axis_0 = const()[name = tensor("op_2185_axis_0"), val = tensor(-1)]; + tensor var_2185_cast_fp16_0, tensor var_2185_cast_fp16_1, tensor var_2185_cast_fp16_2 = split(axis = var_2185_axis_0, split_sizes = var_2185_split_sizes_0, x = linear_142_cast_fp16)[name = tensor("op_2185_cast_fp16")]; + tensor mean_75_axes_0 = const()[name = tensor("mean_75_axes_0"), val = tensor([-1])]; + tensor mean_75_keep_dims_0 = const()[name = tensor("mean_75_keep_dims_0"), val = tensor(true)]; + tensor mean_75_cast_fp16 = reduce_mean(axes = mean_75_axes_0, keep_dims = mean_75_keep_dims_0, x = x_237_cast_fp16)[name = tensor("mean_75_cast_fp16")]; + tensor sub_61_cast_fp16 = sub(x = x_237_cast_fp16, y = mean_75_cast_fp16)[name = tensor("sub_61_cast_fp16")]; + tensor square_49_cast_fp16 = square(x = sub_61_cast_fp16)[name = tensor("square_49_cast_fp16")]; + tensor reduce_mean_99_axes_0 = const()[name = tensor("reduce_mean_99_axes_0"), val = tensor([-1])]; + tensor reduce_mean_99_keep_dims_0 = const()[name = tensor("reduce_mean_99_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_99_cast_fp16 = reduce_mean(axes = reduce_mean_99_axes_0, keep_dims = reduce_mean_99_keep_dims_0, x = square_49_cast_fp16)[name = tensor("reduce_mean_99_cast_fp16")]; + tensor var_2195_to_fp16 = const()[name = tensor("op_2195_to_fp16"), val = tensor(0x1.1p-20)]; + tensor var_2196_cast_fp16 = add(x = reduce_mean_99_cast_fp16, y = var_2195_to_fp16)[name = tensor("op_2196_cast_fp16")]; + tensor var_2197_cast_fp16 = sqrt(x = var_2196_cast_fp16)[name = tensor("op_2197_cast_fp16")]; + tensor x_239_cast_fp16 = real_div(x = sub_61_cast_fp16, y = var_2197_cast_fp16)[name = tensor("x_239_cast_fp16")]; + tensor var_2199_cast_fp16 = mul(x = x_239_cast_fp16, y = flow_net_res_blocks_2_in_ln_weight_to_fp16)[name = tensor("op_2199_cast_fp16")]; + tensor x_241_cast_fp16 = add(x = var_2199_cast_fp16, y = flow_net_res_blocks_2_in_ln_bias_to_fp16)[name = tensor("x_241_cast_fp16")]; + tensor var_2201_promoted_to_fp16 = const()[name = tensor("op_2201_promoted_to_fp16"), val = tensor(0x1p+0)]; + tensor var_2202_cast_fp16 = add(x = var_2185_cast_fp16_1, y = var_2201_promoted_to_fp16)[name = tensor("op_2202_cast_fp16")]; + tensor var_2203_cast_fp16 = mul(x = x_241_cast_fp16, y = var_2202_cast_fp16)[name = tensor("op_2203_cast_fp16")]; + tensor input_373_cast_fp16 = add(x = var_2203_cast_fp16, y = var_2185_cast_fp16_0)[name = tensor("input_373_cast_fp16")]; + tensor linear_143_cast_fp16 = linear(bias = flow_net_res_blocks_2_mlp_0_bias_to_fp16, weight = flow_net_res_blocks_2_mlp_0_weight_to_fp16, x = input_373_cast_fp16)[name = tensor("linear_143_cast_fp16")]; + tensor input_377_cast_fp16 = silu(x = linear_143_cast_fp16)[name = tensor("input_377_cast_fp16")]; + tensor linear_144_cast_fp16 = linear(bias = flow_net_res_blocks_2_mlp_2_bias_to_fp16, weight = flow_net_res_blocks_2_mlp_2_weight_to_fp16, x = input_377_cast_fp16)[name = tensor("linear_144_cast_fp16")]; + tensor var_2214_cast_fp16 = mul(x = var_2185_cast_fp16_2, y = linear_144_cast_fp16)[name = tensor("op_2214_cast_fp16")]; + tensor x_243_cast_fp16 = add(x = x_237_cast_fp16, y = var_2214_cast_fp16)[name = tensor("x_243_cast_fp16")]; + tensor linear_145_cast_fp16 = linear(bias = flow_net_res_blocks_3_adaLN_modulation_1_bias_to_fp16, weight = flow_net_res_blocks_3_adaLN_modulation_1_weight_to_fp16, x = input_355_cast_fp16)[name = tensor("linear_145_cast_fp16")]; + tensor var_2224_split_sizes_0 = const()[name = tensor("op_2224_split_sizes_0"), val = tensor([512, 512, 512])]; + tensor var_2224_axis_0 = const()[name = tensor("op_2224_axis_0"), val = tensor(-1)]; + tensor var_2224_cast_fp16_0, tensor var_2224_cast_fp16_1, tensor var_2224_cast_fp16_2 = split(axis = var_2224_axis_0, split_sizes = var_2224_split_sizes_0, x = linear_145_cast_fp16)[name = tensor("op_2224_cast_fp16")]; + tensor mean_77_axes_0 = const()[name = tensor("mean_77_axes_0"), val = tensor([-1])]; + tensor mean_77_keep_dims_0 = const()[name = tensor("mean_77_keep_dims_0"), val = tensor(true)]; + tensor mean_77_cast_fp16 = reduce_mean(axes = mean_77_axes_0, keep_dims = mean_77_keep_dims_0, x = x_243_cast_fp16)[name = tensor("mean_77_cast_fp16")]; + tensor sub_62_cast_fp16 = sub(x = x_243_cast_fp16, y = mean_77_cast_fp16)[name = tensor("sub_62_cast_fp16")]; + tensor square_50_cast_fp16 = square(x = sub_62_cast_fp16)[name = tensor("square_50_cast_fp16")]; + tensor reduce_mean_101_axes_0 = const()[name = tensor("reduce_mean_101_axes_0"), val = tensor([-1])]; + tensor reduce_mean_101_keep_dims_0 = const()[name = tensor("reduce_mean_101_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_101_cast_fp16 = reduce_mean(axes = reduce_mean_101_axes_0, keep_dims = reduce_mean_101_keep_dims_0, x = square_50_cast_fp16)[name = tensor("reduce_mean_101_cast_fp16")]; + tensor var_2234_to_fp16 = const()[name = tensor("op_2234_to_fp16"), val = tensor(0x1.1p-20)]; + tensor var_2235_cast_fp16 = add(x = reduce_mean_101_cast_fp16, y = var_2234_to_fp16)[name = tensor("op_2235_cast_fp16")]; + tensor var_2236_cast_fp16 = sqrt(x = var_2235_cast_fp16)[name = tensor("op_2236_cast_fp16")]; + tensor x_245_cast_fp16 = real_div(x = sub_62_cast_fp16, y = var_2236_cast_fp16)[name = tensor("x_245_cast_fp16")]; + tensor var_2238_cast_fp16 = mul(x = x_245_cast_fp16, y = flow_net_res_blocks_3_in_ln_weight_to_fp16)[name = tensor("op_2238_cast_fp16")]; + tensor x_247_cast_fp16 = add(x = var_2238_cast_fp16, y = flow_net_res_blocks_3_in_ln_bias_to_fp16)[name = tensor("x_247_cast_fp16")]; + tensor var_2240_promoted_to_fp16 = const()[name = tensor("op_2240_promoted_to_fp16"), val = tensor(0x1p+0)]; + tensor var_2241_cast_fp16 = add(x = var_2224_cast_fp16_1, y = var_2240_promoted_to_fp16)[name = tensor("op_2241_cast_fp16")]; + tensor var_2242_cast_fp16 = mul(x = x_247_cast_fp16, y = var_2241_cast_fp16)[name = tensor("op_2242_cast_fp16")]; + tensor input_381_cast_fp16 = add(x = var_2242_cast_fp16, y = var_2224_cast_fp16_0)[name = tensor("input_381_cast_fp16")]; + tensor linear_146_cast_fp16 = linear(bias = flow_net_res_blocks_3_mlp_0_bias_to_fp16, weight = flow_net_res_blocks_3_mlp_0_weight_to_fp16, x = input_381_cast_fp16)[name = tensor("linear_146_cast_fp16")]; + tensor input_385_cast_fp16 = silu(x = linear_146_cast_fp16)[name = tensor("input_385_cast_fp16")]; + tensor linear_147_cast_fp16 = linear(bias = flow_net_res_blocks_3_mlp_2_bias_to_fp16, weight = flow_net_res_blocks_3_mlp_2_weight_to_fp16, x = input_385_cast_fp16)[name = tensor("linear_147_cast_fp16")]; + tensor var_2253_cast_fp16 = mul(x = var_2224_cast_fp16_2, y = linear_147_cast_fp16)[name = tensor("op_2253_cast_fp16")]; + tensor x_249_cast_fp16 = add(x = x_243_cast_fp16, y = var_2253_cast_fp16)[name = tensor("x_249_cast_fp16")]; + tensor linear_148_cast_fp16 = linear(bias = flow_net_res_blocks_4_adaLN_modulation_1_bias_to_fp16, weight = flow_net_res_blocks_4_adaLN_modulation_1_weight_to_fp16, x = input_355_cast_fp16)[name = tensor("linear_148_cast_fp16")]; + tensor var_2263_split_sizes_0 = const()[name = tensor("op_2263_split_sizes_0"), val = tensor([512, 512, 512])]; + tensor var_2263_axis_0 = const()[name = tensor("op_2263_axis_0"), val = tensor(-1)]; + tensor var_2263_cast_fp16_0, tensor var_2263_cast_fp16_1, tensor var_2263_cast_fp16_2 = split(axis = var_2263_axis_0, split_sizes = var_2263_split_sizes_0, x = linear_148_cast_fp16)[name = tensor("op_2263_cast_fp16")]; + tensor mean_79_axes_0 = const()[name = tensor("mean_79_axes_0"), val = tensor([-1])]; + tensor mean_79_keep_dims_0 = const()[name = tensor("mean_79_keep_dims_0"), val = tensor(true)]; + tensor mean_79_cast_fp16 = reduce_mean(axes = mean_79_axes_0, keep_dims = mean_79_keep_dims_0, x = x_249_cast_fp16)[name = tensor("mean_79_cast_fp16")]; + tensor sub_63_cast_fp16 = sub(x = x_249_cast_fp16, y = mean_79_cast_fp16)[name = tensor("sub_63_cast_fp16")]; + tensor square_51_cast_fp16 = square(x = sub_63_cast_fp16)[name = tensor("square_51_cast_fp16")]; + tensor reduce_mean_103_axes_0 = const()[name = tensor("reduce_mean_103_axes_0"), val = tensor([-1])]; + tensor reduce_mean_103_keep_dims_0 = const()[name = tensor("reduce_mean_103_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_103_cast_fp16 = reduce_mean(axes = reduce_mean_103_axes_0, keep_dims = reduce_mean_103_keep_dims_0, x = square_51_cast_fp16)[name = tensor("reduce_mean_103_cast_fp16")]; + tensor var_2273_to_fp16 = const()[name = tensor("op_2273_to_fp16"), val = tensor(0x1.1p-20)]; + tensor var_2274_cast_fp16 = add(x = reduce_mean_103_cast_fp16, y = var_2273_to_fp16)[name = tensor("op_2274_cast_fp16")]; + tensor var_2275_cast_fp16 = sqrt(x = var_2274_cast_fp16)[name = tensor("op_2275_cast_fp16")]; + tensor x_251_cast_fp16 = real_div(x = sub_63_cast_fp16, y = var_2275_cast_fp16)[name = tensor("x_251_cast_fp16")]; + tensor var_2277_cast_fp16 = mul(x = x_251_cast_fp16, y = flow_net_res_blocks_4_in_ln_weight_to_fp16)[name = tensor("op_2277_cast_fp16")]; + tensor x_253_cast_fp16 = add(x = var_2277_cast_fp16, y = flow_net_res_blocks_4_in_ln_bias_to_fp16)[name = tensor("x_253_cast_fp16")]; + tensor var_2279_promoted_to_fp16 = const()[name = tensor("op_2279_promoted_to_fp16"), val = tensor(0x1p+0)]; + tensor var_2280_cast_fp16 = add(x = var_2263_cast_fp16_1, y = var_2279_promoted_to_fp16)[name = tensor("op_2280_cast_fp16")]; + tensor var_2281_cast_fp16 = mul(x = x_253_cast_fp16, y = var_2280_cast_fp16)[name = tensor("op_2281_cast_fp16")]; + tensor input_389_cast_fp16 = add(x = var_2281_cast_fp16, y = var_2263_cast_fp16_0)[name = tensor("input_389_cast_fp16")]; + tensor linear_149_cast_fp16 = linear(bias = flow_net_res_blocks_4_mlp_0_bias_to_fp16, weight = flow_net_res_blocks_4_mlp_0_weight_to_fp16, x = input_389_cast_fp16)[name = tensor("linear_149_cast_fp16")]; + tensor input_393_cast_fp16 = silu(x = linear_149_cast_fp16)[name = tensor("input_393_cast_fp16")]; + tensor linear_150_cast_fp16 = linear(bias = flow_net_res_blocks_4_mlp_2_bias_to_fp16, weight = flow_net_res_blocks_4_mlp_2_weight_to_fp16, x = input_393_cast_fp16)[name = tensor("linear_150_cast_fp16")]; + tensor var_2292_cast_fp16 = mul(x = var_2263_cast_fp16_2, y = linear_150_cast_fp16)[name = tensor("op_2292_cast_fp16")]; + tensor x_255_cast_fp16 = add(x = x_249_cast_fp16, y = var_2292_cast_fp16)[name = tensor("x_255_cast_fp16")]; + tensor linear_151_cast_fp16 = linear(bias = flow_net_res_blocks_5_adaLN_modulation_1_bias_to_fp16, weight = flow_net_res_blocks_5_adaLN_modulation_1_weight_to_fp16, x = input_355_cast_fp16)[name = tensor("linear_151_cast_fp16")]; + tensor var_2302_split_sizes_0 = const()[name = tensor("op_2302_split_sizes_0"), val = tensor([512, 512, 512])]; + tensor var_2302_axis_0 = const()[name = tensor("op_2302_axis_0"), val = tensor(-1)]; + tensor var_2302_cast_fp16_0, tensor var_2302_cast_fp16_1, tensor var_2302_cast_fp16_2 = split(axis = var_2302_axis_0, split_sizes = var_2302_split_sizes_0, x = linear_151_cast_fp16)[name = tensor("op_2302_cast_fp16")]; + tensor mean_81_axes_0 = const()[name = tensor("mean_81_axes_0"), val = tensor([-1])]; + tensor mean_81_keep_dims_0 = const()[name = tensor("mean_81_keep_dims_0"), val = tensor(true)]; + tensor mean_81_cast_fp16 = reduce_mean(axes = mean_81_axes_0, keep_dims = mean_81_keep_dims_0, x = x_255_cast_fp16)[name = tensor("mean_81_cast_fp16")]; + tensor sub_64_cast_fp16 = sub(x = x_255_cast_fp16, y = mean_81_cast_fp16)[name = tensor("sub_64_cast_fp16")]; + tensor square_52_cast_fp16 = square(x = sub_64_cast_fp16)[name = tensor("square_52_cast_fp16")]; + tensor reduce_mean_105_axes_0 = const()[name = tensor("reduce_mean_105_axes_0"), val = tensor([-1])]; + tensor reduce_mean_105_keep_dims_0 = const()[name = tensor("reduce_mean_105_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_105_cast_fp16 = reduce_mean(axes = reduce_mean_105_axes_0, keep_dims = reduce_mean_105_keep_dims_0, x = square_52_cast_fp16)[name = tensor("reduce_mean_105_cast_fp16")]; + tensor var_2312_to_fp16 = const()[name = tensor("op_2312_to_fp16"), val = tensor(0x1.1p-20)]; + tensor var_2313_cast_fp16 = add(x = reduce_mean_105_cast_fp16, y = var_2312_to_fp16)[name = tensor("op_2313_cast_fp16")]; + tensor var_2314_cast_fp16 = sqrt(x = var_2313_cast_fp16)[name = tensor("op_2314_cast_fp16")]; + tensor x_257_cast_fp16 = real_div(x = sub_64_cast_fp16, y = var_2314_cast_fp16)[name = tensor("x_257_cast_fp16")]; + tensor var_2316_cast_fp16 = mul(x = x_257_cast_fp16, y = flow_net_res_blocks_5_in_ln_weight_to_fp16)[name = tensor("op_2316_cast_fp16")]; + tensor x_259_cast_fp16 = add(x = var_2316_cast_fp16, y = flow_net_res_blocks_5_in_ln_bias_to_fp16)[name = tensor("x_259_cast_fp16")]; + tensor var_2318_promoted_to_fp16 = const()[name = tensor("op_2318_promoted_to_fp16"), val = tensor(0x1p+0)]; + tensor var_2319_cast_fp16 = add(x = var_2302_cast_fp16_1, y = var_2318_promoted_to_fp16)[name = tensor("op_2319_cast_fp16")]; + tensor var_2320_cast_fp16 = mul(x = x_259_cast_fp16, y = var_2319_cast_fp16)[name = tensor("op_2320_cast_fp16")]; + tensor input_397_cast_fp16 = add(x = var_2320_cast_fp16, y = var_2302_cast_fp16_0)[name = tensor("input_397_cast_fp16")]; + tensor linear_152_cast_fp16 = linear(bias = flow_net_res_blocks_5_mlp_0_bias_to_fp16, weight = flow_net_res_blocks_5_mlp_0_weight_to_fp16, x = input_397_cast_fp16)[name = tensor("linear_152_cast_fp16")]; + tensor input_401_cast_fp16 = silu(x = linear_152_cast_fp16)[name = tensor("input_401_cast_fp16")]; + tensor linear_153_cast_fp16 = linear(bias = flow_net_res_blocks_5_mlp_2_bias_to_fp16, weight = flow_net_res_blocks_5_mlp_2_weight_to_fp16, x = input_401_cast_fp16)[name = tensor("linear_153_cast_fp16")]; + tensor var_2331_cast_fp16 = mul(x = var_2302_cast_fp16_2, y = linear_153_cast_fp16)[name = tensor("op_2331_cast_fp16")]; + tensor x_261_cast_fp16 = add(x = x_255_cast_fp16, y = var_2331_cast_fp16)[name = tensor("x_261_cast_fp16")]; + tensor linear_154_cast_fp16 = linear(bias = flow_net_final_layer_adaLN_modulation_1_bias_to_fp16, weight = flow_net_final_layer_adaLN_modulation_1_weight_to_fp16, x = input_355_cast_fp16)[name = tensor("linear_154_cast_fp16")]; + tensor var_2340_split_sizes_0 = const()[name = tensor("op_2340_split_sizes_0"), val = tensor([512, 512])]; + tensor var_2340_axis_0 = const()[name = tensor("op_2340_axis_0"), val = tensor(-1)]; + tensor var_2340_cast_fp16_0, tensor var_2340_cast_fp16_1 = split(axis = var_2340_axis_0, split_sizes = var_2340_split_sizes_0, x = linear_154_cast_fp16)[name = tensor("op_2340_cast_fp16")]; + tensor mean_83_axes_0 = const()[name = tensor("mean_83_axes_0"), val = tensor([-1])]; + tensor mean_83_keep_dims_0 = const()[name = tensor("mean_83_keep_dims_0"), val = tensor(true)]; + tensor mean_83_cast_fp16 = reduce_mean(axes = mean_83_axes_0, keep_dims = mean_83_keep_dims_0, x = x_261_cast_fp16)[name = tensor("mean_83_cast_fp16")]; + tensor sub_65_cast_fp16 = sub(x = x_261_cast_fp16, y = mean_83_cast_fp16)[name = tensor("sub_65_cast_fp16")]; + tensor square_53_cast_fp16 = square(x = sub_65_cast_fp16)[name = tensor("square_53_cast_fp16")]; + tensor reduce_mean_107_axes_0 = const()[name = tensor("reduce_mean_107_axes_0"), val = tensor([-1])]; + tensor reduce_mean_107_keep_dims_0 = const()[name = tensor("reduce_mean_107_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_107_cast_fp16 = reduce_mean(axes = reduce_mean_107_axes_0, keep_dims = reduce_mean_107_keep_dims_0, x = square_53_cast_fp16)[name = tensor("reduce_mean_107_cast_fp16")]; + tensor var_2347_to_fp16 = const()[name = tensor("op_2347_to_fp16"), val = tensor(0x1.1p-20)]; + tensor var_2348_cast_fp16 = add(x = reduce_mean_107_cast_fp16, y = var_2347_to_fp16)[name = tensor("op_2348_cast_fp16")]; + tensor var_2349_cast_fp16 = sqrt(x = var_2348_cast_fp16)[name = tensor("op_2349_cast_fp16")]; + tensor x_263_cast_fp16 = real_div(x = sub_65_cast_fp16, y = var_2349_cast_fp16)[name = tensor("x_263_cast_fp16")]; + tensor var_2351_promoted_to_fp16 = const()[name = tensor("op_2351_promoted_to_fp16"), val = tensor(0x1p+0)]; + tensor var_2352_cast_fp16 = add(x = var_2340_cast_fp16_1, y = var_2351_promoted_to_fp16)[name = tensor("op_2352_cast_fp16")]; + tensor var_2353_cast_fp16 = mul(x = x_263_cast_fp16, y = var_2352_cast_fp16)[name = tensor("op_2353_cast_fp16")]; + tensor input_405_cast_fp16 = add(x = var_2353_cast_fp16, y = var_2340_cast_fp16_0)[name = tensor("input_405_cast_fp16")]; + tensor linear_155_cast_fp16 = linear(bias = flow_net_final_layer_linear_bias_to_fp16, weight = flow_net_final_layer_linear_weight_to_fp16, x = input_405_cast_fp16)[name = tensor("linear_155_cast_fp16")]; + tensor var_2364_to_fp16 = const()[name = tensor("op_2364_to_fp16"), val = tensor(0x1p-3)]; + tensor var_2365_cast_fp16 = mul(x = linear_155_cast_fp16, y = var_2364_to_fp16)[name = tensor("op_2365_cast_fp16")]; + tensor input_407_cast_fp16 = add(x = input_339_cast_fp16, y = var_2365_cast_fp16)[name = tensor("input_407_cast_fp16")]; + tensor linear_156_cast_fp16 = linear(bias = flow_net_input_proj_bias_to_fp16, weight = flow_net_input_proj_weight_to_fp16, x = input_407_cast_fp16)[name = tensor("linear_156_cast_fp16")]; + tensor input_411_to_fp16 = const()[name = tensor("input_411_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(19008256)))]; + tensor input_413_cast_fp16 = silu(x = input_411_to_fp16)[name = tensor("input_413_cast_fp16")]; + tensor linear_158_cast_fp16 = linear(bias = flow_net_time_embed_0_mlp_2_bias_to_fp16, weight = flow_net_time_embed_0_mlp_2_weight_to_fp16, x = input_413_cast_fp16)[name = tensor("linear_158_cast_fp16")]; + tensor reduce_mean_108_axes_0 = const()[name = tensor("reduce_mean_108_axes_0"), val = tensor([-1])]; + tensor reduce_mean_108_keep_dims_0 = const()[name = tensor("reduce_mean_108_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_108_cast_fp16 = reduce_mean(axes = reduce_mean_108_axes_0, keep_dims = reduce_mean_108_keep_dims_0, x = linear_158_cast_fp16)[name = tensor("reduce_mean_108_cast_fp16")]; + tensor sub_66_cast_fp16 = sub(x = linear_158_cast_fp16, y = reduce_mean_108_cast_fp16)[name = tensor("sub_66_cast_fp16")]; + tensor square_54_cast_fp16 = square(x = sub_66_cast_fp16)[name = tensor("square_54_cast_fp16")]; + tensor reduce_mean_109_axes_0 = const()[name = tensor("reduce_mean_109_axes_0"), val = tensor([-1])]; + tensor reduce_mean_109_keep_dims_0 = const()[name = tensor("reduce_mean_109_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_109_cast_fp16 = reduce_mean(axes = reduce_mean_109_axes_0, keep_dims = reduce_mean_109_keep_dims_0, x = square_54_cast_fp16)[name = tensor("reduce_mean_109_cast_fp16")]; + tensor real_div_12_to_fp16 = const()[name = tensor("real_div_12_to_fp16"), val = tensor(0x1.008p+0)]; + tensor mul_12_cast_fp16 = mul(x = reduce_mean_109_cast_fp16, y = real_div_12_to_fp16)[name = tensor("mul_12_cast_fp16")]; + tensor var_2431_to_fp16 = const()[name = tensor("op_2431_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_109_cast_fp16 = add(x = mul_12_cast_fp16, y = var_2431_to_fp16)[name = tensor("var_109_cast_fp16")]; + tensor var_2434_epsilon_0 = const()[name = tensor("op_2434_epsilon_0"), val = tensor(0x1.197998p-40)]; + tensor var_2434_cast_fp16 = rsqrt(epsilon = var_2434_epsilon_0, x = var_109_cast_fp16)[name = tensor("op_2434_cast_fp16")]; + tensor var_2435_cast_fp16 = mul(x = const_3_to_fp16, y = var_2434_cast_fp16)[name = tensor("op_2435_cast_fp16")]; + tensor var_2436_cast_fp16 = mul(x = linear_158_cast_fp16, y = var_2435_cast_fp16)[name = tensor("op_2436_cast_fp16")]; + tensor input_417_to_fp16 = const()[name = tensor("input_417_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(19009344)))]; + tensor input_419_cast_fp16 = silu(x = input_417_to_fp16)[name = tensor("input_419_cast_fp16")]; + tensor linear_160_cast_fp16 = linear(bias = flow_net_time_embed_1_mlp_2_bias_to_fp16, weight = flow_net_time_embed_1_mlp_2_weight_to_fp16, x = input_419_cast_fp16)[name = tensor("linear_160_cast_fp16")]; + tensor reduce_mean_110_axes_0 = const()[name = tensor("reduce_mean_110_axes_0"), val = tensor([-1])]; + tensor reduce_mean_110_keep_dims_0 = const()[name = tensor("reduce_mean_110_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_110_cast_fp16 = reduce_mean(axes = reduce_mean_110_axes_0, keep_dims = reduce_mean_110_keep_dims_0, x = linear_160_cast_fp16)[name = tensor("reduce_mean_110_cast_fp16")]; + tensor sub_68_cast_fp16 = sub(x = linear_160_cast_fp16, y = reduce_mean_110_cast_fp16)[name = tensor("sub_68_cast_fp16")]; + tensor square_55_cast_fp16 = square(x = sub_68_cast_fp16)[name = tensor("square_55_cast_fp16")]; + tensor reduce_mean_111_axes_0 = const()[name = tensor("reduce_mean_111_axes_0"), val = tensor([-1])]; + tensor reduce_mean_111_keep_dims_0 = const()[name = tensor("reduce_mean_111_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_111_cast_fp16 = reduce_mean(axes = reduce_mean_111_axes_0, keep_dims = reduce_mean_111_keep_dims_0, x = square_55_cast_fp16)[name = tensor("reduce_mean_111_cast_fp16")]; + tensor real_div_13_to_fp16 = const()[name = tensor("real_div_13_to_fp16"), val = tensor(0x1.008p+0)]; + tensor mul_13_cast_fp16 = mul(x = reduce_mean_111_cast_fp16, y = real_div_13_to_fp16)[name = tensor("mul_13_cast_fp16")]; + tensor var_2468_to_fp16 = const()[name = tensor("op_2468_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_111_cast_fp16 = add(x = mul_13_cast_fp16, y = var_2468_to_fp16)[name = tensor("var_111_cast_fp16")]; + tensor var_2471_epsilon_0 = const()[name = tensor("op_2471_epsilon_0"), val = tensor(0x1.197998p-40)]; + tensor var_2471_cast_fp16 = rsqrt(epsilon = var_2471_epsilon_0, x = var_111_cast_fp16)[name = tensor("op_2471_cast_fp16")]; + tensor var_2472_cast_fp16 = mul(x = const_5_to_fp16, y = var_2471_cast_fp16)[name = tensor("op_2472_cast_fp16")]; + tensor var_2473_cast_fp16 = mul(x = linear_160_cast_fp16, y = var_2472_cast_fp16)[name = tensor("op_2473_cast_fp16")]; + tensor var_2485_cast_fp16 = add(x = var_2436_cast_fp16, y = var_2473_cast_fp16)[name = tensor("op_2485_cast_fp16")]; + tensor _inversed_t_combined_13_y_0_to_fp16 = const()[name = tensor("_inversed_t_combined_13_y_0_to_fp16"), val = tensor(0x1p-1)]; + tensor _inversed_t_combined_13_cast_fp16 = mul(x = var_2485_cast_fp16, y = _inversed_t_combined_13_y_0_to_fp16)[name = tensor("_inversed_t_combined_13_cast_fp16")]; + tensor input_421_cast_fp16 = add(x = _inversed_t_combined_13_cast_fp16, y = linear_5_cast_fp16)[name = tensor("input_421_cast_fp16")]; + tensor input_423_cast_fp16 = silu(x = input_421_cast_fp16)[name = tensor("input_423_cast_fp16")]; + tensor linear_162_cast_fp16 = linear(bias = flow_net_res_blocks_0_adaLN_modulation_1_bias_to_fp16, weight = flow_net_res_blocks_0_adaLN_modulation_1_weight_to_fp16, x = input_423_cast_fp16)[name = tensor("linear_162_cast_fp16")]; + tensor var_2500_split_sizes_0 = const()[name = tensor("op_2500_split_sizes_0"), val = tensor([512, 512, 512])]; + tensor var_2500_axis_0 = const()[name = tensor("op_2500_axis_0"), val = tensor(-1)]; + tensor var_2500_cast_fp16_0, tensor var_2500_cast_fp16_1, tensor var_2500_cast_fp16_2 = split(axis = var_2500_axis_0, split_sizes = var_2500_split_sizes_0, x = linear_162_cast_fp16)[name = tensor("op_2500_cast_fp16")]; + tensor mean_85_axes_0 = const()[name = tensor("mean_85_axes_0"), val = tensor([-1])]; + tensor mean_85_keep_dims_0 = const()[name = tensor("mean_85_keep_dims_0"), val = tensor(true)]; + tensor mean_85_cast_fp16 = reduce_mean(axes = mean_85_axes_0, keep_dims = mean_85_keep_dims_0, x = linear_156_cast_fp16)[name = tensor("mean_85_cast_fp16")]; + tensor sub_70_cast_fp16 = sub(x = linear_156_cast_fp16, y = mean_85_cast_fp16)[name = tensor("sub_70_cast_fp16")]; + tensor square_56_cast_fp16 = square(x = sub_70_cast_fp16)[name = tensor("square_56_cast_fp16")]; + tensor reduce_mean_113_axes_0 = const()[name = tensor("reduce_mean_113_axes_0"), val = tensor([-1])]; + tensor reduce_mean_113_keep_dims_0 = const()[name = tensor("reduce_mean_113_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_113_cast_fp16 = reduce_mean(axes = reduce_mean_113_axes_0, keep_dims = reduce_mean_113_keep_dims_0, x = square_56_cast_fp16)[name = tensor("reduce_mean_113_cast_fp16")]; + tensor var_2510_to_fp16 = const()[name = tensor("op_2510_to_fp16"), val = tensor(0x1.1p-20)]; + tensor var_2511_cast_fp16 = add(x = reduce_mean_113_cast_fp16, y = var_2510_to_fp16)[name = tensor("op_2511_cast_fp16")]; + tensor var_2512_cast_fp16 = sqrt(x = var_2511_cast_fp16)[name = tensor("op_2512_cast_fp16")]; + tensor x_271_cast_fp16 = real_div(x = sub_70_cast_fp16, y = var_2512_cast_fp16)[name = tensor("x_271_cast_fp16")]; + tensor var_2514_cast_fp16 = mul(x = x_271_cast_fp16, y = flow_net_res_blocks_0_in_ln_weight_to_fp16)[name = tensor("op_2514_cast_fp16")]; + tensor x_273_cast_fp16 = add(x = var_2514_cast_fp16, y = flow_net_res_blocks_0_in_ln_bias_to_fp16)[name = tensor("x_273_cast_fp16")]; + tensor var_2516_promoted_to_fp16 = const()[name = tensor("op_2516_promoted_to_fp16"), val = tensor(0x1p+0)]; + tensor var_2517_cast_fp16 = add(x = var_2500_cast_fp16_1, y = var_2516_promoted_to_fp16)[name = tensor("op_2517_cast_fp16")]; + tensor var_2518_cast_fp16 = mul(x = x_273_cast_fp16, y = var_2517_cast_fp16)[name = tensor("op_2518_cast_fp16")]; + tensor input_425_cast_fp16 = add(x = var_2518_cast_fp16, y = var_2500_cast_fp16_0)[name = tensor("input_425_cast_fp16")]; + tensor linear_163_cast_fp16 = linear(bias = flow_net_res_blocks_0_mlp_0_bias_to_fp16, weight = flow_net_res_blocks_0_mlp_0_weight_to_fp16, x = input_425_cast_fp16)[name = tensor("linear_163_cast_fp16")]; + tensor input_429_cast_fp16 = silu(x = linear_163_cast_fp16)[name = tensor("input_429_cast_fp16")]; + tensor linear_164_cast_fp16 = linear(bias = flow_net_res_blocks_0_mlp_2_bias_to_fp16, weight = flow_net_res_blocks_0_mlp_2_weight_to_fp16, x = input_429_cast_fp16)[name = tensor("linear_164_cast_fp16")]; + tensor var_2529_cast_fp16 = mul(x = var_2500_cast_fp16_2, y = linear_164_cast_fp16)[name = tensor("op_2529_cast_fp16")]; + tensor x_275_cast_fp16 = add(x = linear_156_cast_fp16, y = var_2529_cast_fp16)[name = tensor("x_275_cast_fp16")]; + tensor linear_165_cast_fp16 = linear(bias = flow_net_res_blocks_1_adaLN_modulation_1_bias_to_fp16, weight = flow_net_res_blocks_1_adaLN_modulation_1_weight_to_fp16, x = input_423_cast_fp16)[name = tensor("linear_165_cast_fp16")]; + tensor var_2539_split_sizes_0 = const()[name = tensor("op_2539_split_sizes_0"), val = tensor([512, 512, 512])]; + tensor var_2539_axis_0 = const()[name = tensor("op_2539_axis_0"), val = tensor(-1)]; + tensor var_2539_cast_fp16_0, tensor var_2539_cast_fp16_1, tensor var_2539_cast_fp16_2 = split(axis = var_2539_axis_0, split_sizes = var_2539_split_sizes_0, x = linear_165_cast_fp16)[name = tensor("op_2539_cast_fp16")]; + tensor mean_87_axes_0 = const()[name = tensor("mean_87_axes_0"), val = tensor([-1])]; + tensor mean_87_keep_dims_0 = const()[name = tensor("mean_87_keep_dims_0"), val = tensor(true)]; + tensor mean_87_cast_fp16 = reduce_mean(axes = mean_87_axes_0, keep_dims = mean_87_keep_dims_0, x = x_275_cast_fp16)[name = tensor("mean_87_cast_fp16")]; + tensor sub_71_cast_fp16 = sub(x = x_275_cast_fp16, y = mean_87_cast_fp16)[name = tensor("sub_71_cast_fp16")]; + tensor square_57_cast_fp16 = square(x = sub_71_cast_fp16)[name = tensor("square_57_cast_fp16")]; + tensor reduce_mean_115_axes_0 = const()[name = tensor("reduce_mean_115_axes_0"), val = tensor([-1])]; + tensor reduce_mean_115_keep_dims_0 = const()[name = tensor("reduce_mean_115_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_115_cast_fp16 = reduce_mean(axes = reduce_mean_115_axes_0, keep_dims = reduce_mean_115_keep_dims_0, x = square_57_cast_fp16)[name = tensor("reduce_mean_115_cast_fp16")]; + tensor var_2549_to_fp16 = const()[name = tensor("op_2549_to_fp16"), val = tensor(0x1.1p-20)]; + tensor var_2550_cast_fp16 = add(x = reduce_mean_115_cast_fp16, y = var_2549_to_fp16)[name = tensor("op_2550_cast_fp16")]; + tensor var_2551_cast_fp16 = sqrt(x = var_2550_cast_fp16)[name = tensor("op_2551_cast_fp16")]; + tensor x_277_cast_fp16 = real_div(x = sub_71_cast_fp16, y = var_2551_cast_fp16)[name = tensor("x_277_cast_fp16")]; + tensor var_2553_cast_fp16 = mul(x = x_277_cast_fp16, y = flow_net_res_blocks_1_in_ln_weight_to_fp16)[name = tensor("op_2553_cast_fp16")]; + tensor x_279_cast_fp16 = add(x = var_2553_cast_fp16, y = flow_net_res_blocks_1_in_ln_bias_to_fp16)[name = tensor("x_279_cast_fp16")]; + tensor var_2555_promoted_to_fp16 = const()[name = tensor("op_2555_promoted_to_fp16"), val = tensor(0x1p+0)]; + tensor var_2556_cast_fp16 = add(x = var_2539_cast_fp16_1, y = var_2555_promoted_to_fp16)[name = tensor("op_2556_cast_fp16")]; + tensor var_2557_cast_fp16 = mul(x = x_279_cast_fp16, y = var_2556_cast_fp16)[name = tensor("op_2557_cast_fp16")]; + tensor input_433_cast_fp16 = add(x = var_2557_cast_fp16, y = var_2539_cast_fp16_0)[name = tensor("input_433_cast_fp16")]; + tensor linear_166_cast_fp16 = linear(bias = flow_net_res_blocks_1_mlp_0_bias_to_fp16, weight = flow_net_res_blocks_1_mlp_0_weight_to_fp16, x = input_433_cast_fp16)[name = tensor("linear_166_cast_fp16")]; + tensor input_437_cast_fp16 = silu(x = linear_166_cast_fp16)[name = tensor("input_437_cast_fp16")]; + tensor linear_167_cast_fp16 = linear(bias = flow_net_res_blocks_1_mlp_2_bias_to_fp16, weight = flow_net_res_blocks_1_mlp_2_weight_to_fp16, x = input_437_cast_fp16)[name = tensor("linear_167_cast_fp16")]; + tensor var_2568_cast_fp16 = mul(x = var_2539_cast_fp16_2, y = linear_167_cast_fp16)[name = tensor("op_2568_cast_fp16")]; + tensor x_281_cast_fp16 = add(x = x_275_cast_fp16, y = var_2568_cast_fp16)[name = tensor("x_281_cast_fp16")]; + tensor linear_168_cast_fp16 = linear(bias = flow_net_res_blocks_2_adaLN_modulation_1_bias_to_fp16, weight = flow_net_res_blocks_2_adaLN_modulation_1_weight_to_fp16, x = input_423_cast_fp16)[name = tensor("linear_168_cast_fp16")]; + tensor var_2578_split_sizes_0 = const()[name = tensor("op_2578_split_sizes_0"), val = tensor([512, 512, 512])]; + tensor var_2578_axis_0 = const()[name = tensor("op_2578_axis_0"), val = tensor(-1)]; + tensor var_2578_cast_fp16_0, tensor var_2578_cast_fp16_1, tensor var_2578_cast_fp16_2 = split(axis = var_2578_axis_0, split_sizes = var_2578_split_sizes_0, x = linear_168_cast_fp16)[name = tensor("op_2578_cast_fp16")]; + tensor mean_89_axes_0 = const()[name = tensor("mean_89_axes_0"), val = tensor([-1])]; + tensor mean_89_keep_dims_0 = const()[name = tensor("mean_89_keep_dims_0"), val = tensor(true)]; + tensor mean_89_cast_fp16 = reduce_mean(axes = mean_89_axes_0, keep_dims = mean_89_keep_dims_0, x = x_281_cast_fp16)[name = tensor("mean_89_cast_fp16")]; + tensor sub_72_cast_fp16 = sub(x = x_281_cast_fp16, y = mean_89_cast_fp16)[name = tensor("sub_72_cast_fp16")]; + tensor square_58_cast_fp16 = square(x = sub_72_cast_fp16)[name = tensor("square_58_cast_fp16")]; + tensor reduce_mean_117_axes_0 = const()[name = tensor("reduce_mean_117_axes_0"), val = tensor([-1])]; + tensor reduce_mean_117_keep_dims_0 = const()[name = tensor("reduce_mean_117_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_117_cast_fp16 = reduce_mean(axes = reduce_mean_117_axes_0, keep_dims = reduce_mean_117_keep_dims_0, x = square_58_cast_fp16)[name = tensor("reduce_mean_117_cast_fp16")]; + tensor var_2588_to_fp16 = const()[name = tensor("op_2588_to_fp16"), val = tensor(0x1.1p-20)]; + tensor var_2589_cast_fp16 = add(x = reduce_mean_117_cast_fp16, y = var_2588_to_fp16)[name = tensor("op_2589_cast_fp16")]; + tensor var_2590_cast_fp16 = sqrt(x = var_2589_cast_fp16)[name = tensor("op_2590_cast_fp16")]; + tensor x_283_cast_fp16 = real_div(x = sub_72_cast_fp16, y = var_2590_cast_fp16)[name = tensor("x_283_cast_fp16")]; + tensor var_2592_cast_fp16 = mul(x = x_283_cast_fp16, y = flow_net_res_blocks_2_in_ln_weight_to_fp16)[name = tensor("op_2592_cast_fp16")]; + tensor x_285_cast_fp16 = add(x = var_2592_cast_fp16, y = flow_net_res_blocks_2_in_ln_bias_to_fp16)[name = tensor("x_285_cast_fp16")]; + tensor var_2594_promoted_to_fp16 = const()[name = tensor("op_2594_promoted_to_fp16"), val = tensor(0x1p+0)]; + tensor var_2595_cast_fp16 = add(x = var_2578_cast_fp16_1, y = var_2594_promoted_to_fp16)[name = tensor("op_2595_cast_fp16")]; + tensor var_2596_cast_fp16 = mul(x = x_285_cast_fp16, y = var_2595_cast_fp16)[name = tensor("op_2596_cast_fp16")]; + tensor input_441_cast_fp16 = add(x = var_2596_cast_fp16, y = var_2578_cast_fp16_0)[name = tensor("input_441_cast_fp16")]; + tensor linear_169_cast_fp16 = linear(bias = flow_net_res_blocks_2_mlp_0_bias_to_fp16, weight = flow_net_res_blocks_2_mlp_0_weight_to_fp16, x = input_441_cast_fp16)[name = tensor("linear_169_cast_fp16")]; + tensor input_445_cast_fp16 = silu(x = linear_169_cast_fp16)[name = tensor("input_445_cast_fp16")]; + tensor linear_170_cast_fp16 = linear(bias = flow_net_res_blocks_2_mlp_2_bias_to_fp16, weight = flow_net_res_blocks_2_mlp_2_weight_to_fp16, x = input_445_cast_fp16)[name = tensor("linear_170_cast_fp16")]; + tensor var_2607_cast_fp16 = mul(x = var_2578_cast_fp16_2, y = linear_170_cast_fp16)[name = tensor("op_2607_cast_fp16")]; + tensor x_287_cast_fp16 = add(x = x_281_cast_fp16, y = var_2607_cast_fp16)[name = tensor("x_287_cast_fp16")]; + tensor linear_171_cast_fp16 = linear(bias = flow_net_res_blocks_3_adaLN_modulation_1_bias_to_fp16, weight = flow_net_res_blocks_3_adaLN_modulation_1_weight_to_fp16, x = input_423_cast_fp16)[name = tensor("linear_171_cast_fp16")]; + tensor var_2617_split_sizes_0 = const()[name = tensor("op_2617_split_sizes_0"), val = tensor([512, 512, 512])]; + tensor var_2617_axis_0 = const()[name = tensor("op_2617_axis_0"), val = tensor(-1)]; + tensor var_2617_cast_fp16_0, tensor var_2617_cast_fp16_1, tensor var_2617_cast_fp16_2 = split(axis = var_2617_axis_0, split_sizes = var_2617_split_sizes_0, x = linear_171_cast_fp16)[name = tensor("op_2617_cast_fp16")]; + tensor mean_91_axes_0 = const()[name = tensor("mean_91_axes_0"), val = tensor([-1])]; + tensor mean_91_keep_dims_0 = const()[name = tensor("mean_91_keep_dims_0"), val = tensor(true)]; + tensor mean_91_cast_fp16 = reduce_mean(axes = mean_91_axes_0, keep_dims = mean_91_keep_dims_0, x = x_287_cast_fp16)[name = tensor("mean_91_cast_fp16")]; + tensor sub_73_cast_fp16 = sub(x = x_287_cast_fp16, y = mean_91_cast_fp16)[name = tensor("sub_73_cast_fp16")]; + tensor square_59_cast_fp16 = square(x = sub_73_cast_fp16)[name = tensor("square_59_cast_fp16")]; + tensor reduce_mean_119_axes_0 = const()[name = tensor("reduce_mean_119_axes_0"), val = tensor([-1])]; + tensor reduce_mean_119_keep_dims_0 = const()[name = tensor("reduce_mean_119_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_119_cast_fp16 = reduce_mean(axes = reduce_mean_119_axes_0, keep_dims = reduce_mean_119_keep_dims_0, x = square_59_cast_fp16)[name = tensor("reduce_mean_119_cast_fp16")]; + tensor var_2627_to_fp16 = const()[name = tensor("op_2627_to_fp16"), val = tensor(0x1.1p-20)]; + tensor var_2628_cast_fp16 = add(x = reduce_mean_119_cast_fp16, y = var_2627_to_fp16)[name = tensor("op_2628_cast_fp16")]; + tensor var_2629_cast_fp16 = sqrt(x = var_2628_cast_fp16)[name = tensor("op_2629_cast_fp16")]; + tensor x_289_cast_fp16 = real_div(x = sub_73_cast_fp16, y = var_2629_cast_fp16)[name = tensor("x_289_cast_fp16")]; + tensor var_2631_cast_fp16 = mul(x = x_289_cast_fp16, y = flow_net_res_blocks_3_in_ln_weight_to_fp16)[name = tensor("op_2631_cast_fp16")]; + tensor x_291_cast_fp16 = add(x = var_2631_cast_fp16, y = flow_net_res_blocks_3_in_ln_bias_to_fp16)[name = tensor("x_291_cast_fp16")]; + tensor var_2633_promoted_to_fp16 = const()[name = tensor("op_2633_promoted_to_fp16"), val = tensor(0x1p+0)]; + tensor var_2634_cast_fp16 = add(x = var_2617_cast_fp16_1, y = var_2633_promoted_to_fp16)[name = tensor("op_2634_cast_fp16")]; + tensor var_2635_cast_fp16 = mul(x = x_291_cast_fp16, y = var_2634_cast_fp16)[name = tensor("op_2635_cast_fp16")]; + tensor input_449_cast_fp16 = add(x = var_2635_cast_fp16, y = var_2617_cast_fp16_0)[name = tensor("input_449_cast_fp16")]; + tensor linear_172_cast_fp16 = linear(bias = flow_net_res_blocks_3_mlp_0_bias_to_fp16, weight = flow_net_res_blocks_3_mlp_0_weight_to_fp16, x = input_449_cast_fp16)[name = tensor("linear_172_cast_fp16")]; + tensor input_453_cast_fp16 = silu(x = linear_172_cast_fp16)[name = tensor("input_453_cast_fp16")]; + tensor linear_173_cast_fp16 = linear(bias = flow_net_res_blocks_3_mlp_2_bias_to_fp16, weight = flow_net_res_blocks_3_mlp_2_weight_to_fp16, x = input_453_cast_fp16)[name = tensor("linear_173_cast_fp16")]; + tensor var_2646_cast_fp16 = mul(x = var_2617_cast_fp16_2, y = linear_173_cast_fp16)[name = tensor("op_2646_cast_fp16")]; + tensor x_293_cast_fp16 = add(x = x_287_cast_fp16, y = var_2646_cast_fp16)[name = tensor("x_293_cast_fp16")]; + tensor linear_174_cast_fp16 = linear(bias = flow_net_res_blocks_4_adaLN_modulation_1_bias_to_fp16, weight = flow_net_res_blocks_4_adaLN_modulation_1_weight_to_fp16, x = input_423_cast_fp16)[name = tensor("linear_174_cast_fp16")]; + tensor var_2656_split_sizes_0 = const()[name = tensor("op_2656_split_sizes_0"), val = tensor([512, 512, 512])]; + tensor var_2656_axis_0 = const()[name = tensor("op_2656_axis_0"), val = tensor(-1)]; + tensor var_2656_cast_fp16_0, tensor var_2656_cast_fp16_1, tensor var_2656_cast_fp16_2 = split(axis = var_2656_axis_0, split_sizes = var_2656_split_sizes_0, x = linear_174_cast_fp16)[name = tensor("op_2656_cast_fp16")]; + tensor mean_93_axes_0 = const()[name = tensor("mean_93_axes_0"), val = tensor([-1])]; + tensor mean_93_keep_dims_0 = const()[name = tensor("mean_93_keep_dims_0"), val = tensor(true)]; + tensor mean_93_cast_fp16 = reduce_mean(axes = mean_93_axes_0, keep_dims = mean_93_keep_dims_0, x = x_293_cast_fp16)[name = tensor("mean_93_cast_fp16")]; + tensor sub_74_cast_fp16 = sub(x = x_293_cast_fp16, y = mean_93_cast_fp16)[name = tensor("sub_74_cast_fp16")]; + tensor square_60_cast_fp16 = square(x = sub_74_cast_fp16)[name = tensor("square_60_cast_fp16")]; + tensor reduce_mean_121_axes_0 = const()[name = tensor("reduce_mean_121_axes_0"), val = tensor([-1])]; + tensor reduce_mean_121_keep_dims_0 = const()[name = tensor("reduce_mean_121_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_121_cast_fp16 = reduce_mean(axes = reduce_mean_121_axes_0, keep_dims = reduce_mean_121_keep_dims_0, x = square_60_cast_fp16)[name = tensor("reduce_mean_121_cast_fp16")]; + tensor var_2666_to_fp16 = const()[name = tensor("op_2666_to_fp16"), val = tensor(0x1.1p-20)]; + tensor var_2667_cast_fp16 = add(x = reduce_mean_121_cast_fp16, y = var_2666_to_fp16)[name = tensor("op_2667_cast_fp16")]; + tensor var_2668_cast_fp16 = sqrt(x = var_2667_cast_fp16)[name = tensor("op_2668_cast_fp16")]; + tensor x_295_cast_fp16 = real_div(x = sub_74_cast_fp16, y = var_2668_cast_fp16)[name = tensor("x_295_cast_fp16")]; + tensor var_2670_cast_fp16 = mul(x = x_295_cast_fp16, y = flow_net_res_blocks_4_in_ln_weight_to_fp16)[name = tensor("op_2670_cast_fp16")]; + tensor x_297_cast_fp16 = add(x = var_2670_cast_fp16, y = flow_net_res_blocks_4_in_ln_bias_to_fp16)[name = tensor("x_297_cast_fp16")]; + tensor var_2672_promoted_to_fp16 = const()[name = tensor("op_2672_promoted_to_fp16"), val = tensor(0x1p+0)]; + tensor var_2673_cast_fp16 = add(x = var_2656_cast_fp16_1, y = var_2672_promoted_to_fp16)[name = tensor("op_2673_cast_fp16")]; + tensor var_2674_cast_fp16 = mul(x = x_297_cast_fp16, y = var_2673_cast_fp16)[name = tensor("op_2674_cast_fp16")]; + tensor input_457_cast_fp16 = add(x = var_2674_cast_fp16, y = var_2656_cast_fp16_0)[name = tensor("input_457_cast_fp16")]; + tensor linear_175_cast_fp16 = linear(bias = flow_net_res_blocks_4_mlp_0_bias_to_fp16, weight = flow_net_res_blocks_4_mlp_0_weight_to_fp16, x = input_457_cast_fp16)[name = tensor("linear_175_cast_fp16")]; + tensor input_461_cast_fp16 = silu(x = linear_175_cast_fp16)[name = tensor("input_461_cast_fp16")]; + tensor linear_176_cast_fp16 = linear(bias = flow_net_res_blocks_4_mlp_2_bias_to_fp16, weight = flow_net_res_blocks_4_mlp_2_weight_to_fp16, x = input_461_cast_fp16)[name = tensor("linear_176_cast_fp16")]; + tensor var_2685_cast_fp16 = mul(x = var_2656_cast_fp16_2, y = linear_176_cast_fp16)[name = tensor("op_2685_cast_fp16")]; + tensor x_299_cast_fp16 = add(x = x_293_cast_fp16, y = var_2685_cast_fp16)[name = tensor("x_299_cast_fp16")]; + tensor linear_177_cast_fp16 = linear(bias = flow_net_res_blocks_5_adaLN_modulation_1_bias_to_fp16, weight = flow_net_res_blocks_5_adaLN_modulation_1_weight_to_fp16, x = input_423_cast_fp16)[name = tensor("linear_177_cast_fp16")]; + tensor var_2695_split_sizes_0 = const()[name = tensor("op_2695_split_sizes_0"), val = tensor([512, 512, 512])]; + tensor var_2695_axis_0 = const()[name = tensor("op_2695_axis_0"), val = tensor(-1)]; + tensor var_2695_cast_fp16_0, tensor var_2695_cast_fp16_1, tensor var_2695_cast_fp16_2 = split(axis = var_2695_axis_0, split_sizes = var_2695_split_sizes_0, x = linear_177_cast_fp16)[name = tensor("op_2695_cast_fp16")]; + tensor mean_95_axes_0 = const()[name = tensor("mean_95_axes_0"), val = tensor([-1])]; + tensor mean_95_keep_dims_0 = const()[name = tensor("mean_95_keep_dims_0"), val = tensor(true)]; + tensor mean_95_cast_fp16 = reduce_mean(axes = mean_95_axes_0, keep_dims = mean_95_keep_dims_0, x = x_299_cast_fp16)[name = tensor("mean_95_cast_fp16")]; + tensor sub_75_cast_fp16 = sub(x = x_299_cast_fp16, y = mean_95_cast_fp16)[name = tensor("sub_75_cast_fp16")]; + tensor square_61_cast_fp16 = square(x = sub_75_cast_fp16)[name = tensor("square_61_cast_fp16")]; + tensor reduce_mean_123_axes_0 = const()[name = tensor("reduce_mean_123_axes_0"), val = tensor([-1])]; + tensor reduce_mean_123_keep_dims_0 = const()[name = tensor("reduce_mean_123_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_123_cast_fp16 = reduce_mean(axes = reduce_mean_123_axes_0, keep_dims = reduce_mean_123_keep_dims_0, x = square_61_cast_fp16)[name = tensor("reduce_mean_123_cast_fp16")]; + tensor var_2705_to_fp16 = const()[name = tensor("op_2705_to_fp16"), val = tensor(0x1.1p-20)]; + tensor var_2706_cast_fp16 = add(x = reduce_mean_123_cast_fp16, y = var_2705_to_fp16)[name = tensor("op_2706_cast_fp16")]; + tensor var_2707_cast_fp16 = sqrt(x = var_2706_cast_fp16)[name = tensor("op_2707_cast_fp16")]; + tensor x_301_cast_fp16 = real_div(x = sub_75_cast_fp16, y = var_2707_cast_fp16)[name = tensor("x_301_cast_fp16")]; + tensor var_2709_cast_fp16 = mul(x = x_301_cast_fp16, y = flow_net_res_blocks_5_in_ln_weight_to_fp16)[name = tensor("op_2709_cast_fp16")]; + tensor x_303_cast_fp16 = add(x = var_2709_cast_fp16, y = flow_net_res_blocks_5_in_ln_bias_to_fp16)[name = tensor("x_303_cast_fp16")]; + tensor var_2711_promoted_to_fp16 = const()[name = tensor("op_2711_promoted_to_fp16"), val = tensor(0x1p+0)]; + tensor var_2712_cast_fp16 = add(x = var_2695_cast_fp16_1, y = var_2711_promoted_to_fp16)[name = tensor("op_2712_cast_fp16")]; + tensor var_2713_cast_fp16 = mul(x = x_303_cast_fp16, y = var_2712_cast_fp16)[name = tensor("op_2713_cast_fp16")]; + tensor input_465_cast_fp16 = add(x = var_2713_cast_fp16, y = var_2695_cast_fp16_0)[name = tensor("input_465_cast_fp16")]; + tensor linear_178_cast_fp16 = linear(bias = flow_net_res_blocks_5_mlp_0_bias_to_fp16, weight = flow_net_res_blocks_5_mlp_0_weight_to_fp16, x = input_465_cast_fp16)[name = tensor("linear_178_cast_fp16")]; + tensor input_469_cast_fp16 = silu(x = linear_178_cast_fp16)[name = tensor("input_469_cast_fp16")]; + tensor linear_179_cast_fp16 = linear(bias = flow_net_res_blocks_5_mlp_2_bias_to_fp16, weight = flow_net_res_blocks_5_mlp_2_weight_to_fp16, x = input_469_cast_fp16)[name = tensor("linear_179_cast_fp16")]; + tensor var_2724_cast_fp16 = mul(x = var_2695_cast_fp16_2, y = linear_179_cast_fp16)[name = tensor("op_2724_cast_fp16")]; + tensor x_305_cast_fp16 = add(x = x_299_cast_fp16, y = var_2724_cast_fp16)[name = tensor("x_305_cast_fp16")]; + tensor linear_180_cast_fp16 = linear(bias = flow_net_final_layer_adaLN_modulation_1_bias_to_fp16, weight = flow_net_final_layer_adaLN_modulation_1_weight_to_fp16, x = input_423_cast_fp16)[name = tensor("linear_180_cast_fp16")]; + tensor var_2733_split_sizes_0 = const()[name = tensor("op_2733_split_sizes_0"), val = tensor([512, 512])]; + tensor var_2733_axis_0 = const()[name = tensor("op_2733_axis_0"), val = tensor(-1)]; + tensor var_2733_cast_fp16_0, tensor var_2733_cast_fp16_1 = split(axis = var_2733_axis_0, split_sizes = var_2733_split_sizes_0, x = linear_180_cast_fp16)[name = tensor("op_2733_cast_fp16")]; + tensor mean_97_axes_0 = const()[name = tensor("mean_97_axes_0"), val = tensor([-1])]; + tensor mean_97_keep_dims_0 = const()[name = tensor("mean_97_keep_dims_0"), val = tensor(true)]; + tensor mean_97_cast_fp16 = reduce_mean(axes = mean_97_axes_0, keep_dims = mean_97_keep_dims_0, x = x_305_cast_fp16)[name = tensor("mean_97_cast_fp16")]; + tensor sub_76_cast_fp16 = sub(x = x_305_cast_fp16, y = mean_97_cast_fp16)[name = tensor("sub_76_cast_fp16")]; + tensor square_62_cast_fp16 = square(x = sub_76_cast_fp16)[name = tensor("square_62_cast_fp16")]; + tensor reduce_mean_125_axes_0 = const()[name = tensor("reduce_mean_125_axes_0"), val = tensor([-1])]; + tensor reduce_mean_125_keep_dims_0 = const()[name = tensor("reduce_mean_125_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_125_cast_fp16 = reduce_mean(axes = reduce_mean_125_axes_0, keep_dims = reduce_mean_125_keep_dims_0, x = square_62_cast_fp16)[name = tensor("reduce_mean_125_cast_fp16")]; + tensor var_2740_to_fp16 = const()[name = tensor("op_2740_to_fp16"), val = tensor(0x1.1p-20)]; + tensor var_2741_cast_fp16 = add(x = reduce_mean_125_cast_fp16, y = var_2740_to_fp16)[name = tensor("op_2741_cast_fp16")]; + tensor var_2742_cast_fp16 = sqrt(x = var_2741_cast_fp16)[name = tensor("op_2742_cast_fp16")]; + tensor x_307_cast_fp16 = real_div(x = sub_76_cast_fp16, y = var_2742_cast_fp16)[name = tensor("x_307_cast_fp16")]; + tensor var_2744_promoted_to_fp16 = const()[name = tensor("op_2744_promoted_to_fp16"), val = tensor(0x1p+0)]; + tensor var_2745_cast_fp16 = add(x = var_2733_cast_fp16_1, y = var_2744_promoted_to_fp16)[name = tensor("op_2745_cast_fp16")]; + tensor var_2746_cast_fp16 = mul(x = x_307_cast_fp16, y = var_2745_cast_fp16)[name = tensor("op_2746_cast_fp16")]; + tensor input_473_cast_fp16 = add(x = var_2746_cast_fp16, y = var_2733_cast_fp16_0)[name = tensor("input_473_cast_fp16")]; + tensor linear_181_cast_fp16 = linear(bias = flow_net_final_layer_linear_bias_to_fp16, weight = flow_net_final_layer_linear_weight_to_fp16, x = input_473_cast_fp16)[name = tensor("linear_181_cast_fp16")]; + tensor var_2757_to_fp16 = const()[name = tensor("op_2757_to_fp16"), val = tensor(0x1p-3)]; + tensor var_2758_cast_fp16 = mul(x = linear_181_cast_fp16, y = var_2757_to_fp16)[name = tensor("op_2758_cast_fp16")]; + tensor input_475_cast_fp16 = add(x = input_407_cast_fp16, y = var_2758_cast_fp16)[name = tensor("input_475_cast_fp16")]; + tensor linear_182_cast_fp16 = linear(bias = flow_net_input_proj_bias_to_fp16, weight = flow_net_input_proj_weight_to_fp16, x = input_475_cast_fp16)[name = tensor("linear_182_cast_fp16")]; + tensor input_479_to_fp16 = const()[name = tensor("input_479_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(19010432)))]; + tensor input_481_cast_fp16 = silu(x = input_479_to_fp16)[name = tensor("input_481_cast_fp16")]; + tensor linear_184_cast_fp16 = linear(bias = flow_net_time_embed_0_mlp_2_bias_to_fp16, weight = flow_net_time_embed_0_mlp_2_weight_to_fp16, x = input_481_cast_fp16)[name = tensor("linear_184_cast_fp16")]; + tensor reduce_mean_126_axes_0 = const()[name = tensor("reduce_mean_126_axes_0"), val = tensor([-1])]; + tensor reduce_mean_126_keep_dims_0 = const()[name = tensor("reduce_mean_126_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_126_cast_fp16 = reduce_mean(axes = reduce_mean_126_axes_0, keep_dims = reduce_mean_126_keep_dims_0, x = linear_184_cast_fp16)[name = tensor("reduce_mean_126_cast_fp16")]; + tensor sub_77_cast_fp16 = sub(x = linear_184_cast_fp16, y = reduce_mean_126_cast_fp16)[name = tensor("sub_77_cast_fp16")]; + tensor square_63_cast_fp16 = square(x = sub_77_cast_fp16)[name = tensor("square_63_cast_fp16")]; + tensor reduce_mean_127_axes_0 = const()[name = tensor("reduce_mean_127_axes_0"), val = tensor([-1])]; + tensor reduce_mean_127_keep_dims_0 = const()[name = tensor("reduce_mean_127_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_127_cast_fp16 = reduce_mean(axes = reduce_mean_127_axes_0, keep_dims = reduce_mean_127_keep_dims_0, x = square_63_cast_fp16)[name = tensor("reduce_mean_127_cast_fp16")]; + tensor real_div_14_to_fp16 = const()[name = tensor("real_div_14_to_fp16"), val = tensor(0x1.008p+0)]; + tensor mul_14_cast_fp16 = mul(x = reduce_mean_127_cast_fp16, y = real_div_14_to_fp16)[name = tensor("mul_14_cast_fp16")]; + tensor var_2824_to_fp16 = const()[name = tensor("op_2824_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_127_cast_fp16_0 = add(x = mul_14_cast_fp16, y = var_2824_to_fp16)[name = tensor("var_127_cast_fp16")]; + tensor var_2827_epsilon_0 = const()[name = tensor("op_2827_epsilon_0"), val = tensor(0x1.197998p-40)]; + tensor var_2827_cast_fp16 = rsqrt(epsilon = var_2827_epsilon_0, x = var_127_cast_fp16_0)[name = tensor("op_2827_cast_fp16")]; + tensor var_2828_cast_fp16 = mul(x = const_3_to_fp16, y = var_2827_cast_fp16)[name = tensor("op_2828_cast_fp16")]; + tensor var_2829_cast_fp16 = mul(x = linear_184_cast_fp16, y = var_2828_cast_fp16)[name = tensor("op_2829_cast_fp16")]; + tensor input_485_to_fp16 = const()[name = tensor("input_485_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(19011520)))]; + tensor input_487_cast_fp16 = silu(x = input_485_to_fp16)[name = tensor("input_487_cast_fp16")]; + tensor linear_186_cast_fp16 = linear(bias = flow_net_time_embed_1_mlp_2_bias_to_fp16, weight = flow_net_time_embed_1_mlp_2_weight_to_fp16, x = input_487_cast_fp16)[name = tensor("linear_186_cast_fp16")]; + tensor reduce_mean_128_axes_0 = const()[name = tensor("reduce_mean_128_axes_0"), val = tensor([-1])]; + tensor reduce_mean_128_keep_dims_0 = const()[name = tensor("reduce_mean_128_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_128_cast_fp16 = reduce_mean(axes = reduce_mean_128_axes_0, keep_dims = reduce_mean_128_keep_dims_0, x = linear_186_cast_fp16)[name = tensor("reduce_mean_128_cast_fp16")]; + tensor sub_79_cast_fp16 = sub(x = linear_186_cast_fp16, y = reduce_mean_128_cast_fp16)[name = tensor("sub_79_cast_fp16")]; + tensor square_64_cast_fp16 = square(x = sub_79_cast_fp16)[name = tensor("square_64_cast_fp16")]; + tensor reduce_mean_129_axes_0 = const()[name = tensor("reduce_mean_129_axes_0"), val = tensor([-1])]; + tensor reduce_mean_129_keep_dims_0 = const()[name = tensor("reduce_mean_129_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_129_cast_fp16 = reduce_mean(axes = reduce_mean_129_axes_0, keep_dims = reduce_mean_129_keep_dims_0, x = square_64_cast_fp16)[name = tensor("reduce_mean_129_cast_fp16")]; + tensor real_div_15_to_fp16 = const()[name = tensor("real_div_15_to_fp16"), val = tensor(0x1.008p+0)]; + tensor mul_15_cast_fp16 = mul(x = reduce_mean_129_cast_fp16, y = real_div_15_to_fp16)[name = tensor("mul_15_cast_fp16")]; + tensor var_2851_to_fp16 = const()[name = tensor("op_2851_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_129_cast_fp16 = add(x = mul_15_cast_fp16, y = var_2851_to_fp16)[name = tensor("var_129_cast_fp16")]; + tensor var_2854_epsilon_0 = const()[name = tensor("op_2854_epsilon_0"), val = tensor(0x1.197998p-40)]; + tensor var_2854_cast_fp16 = rsqrt(epsilon = var_2854_epsilon_0, x = var_129_cast_fp16)[name = tensor("op_2854_cast_fp16")]; + tensor var_2855_cast_fp16 = mul(x = const_5_to_fp16, y = var_2854_cast_fp16)[name = tensor("op_2855_cast_fp16")]; + tensor var_2856_cast_fp16 = mul(x = linear_186_cast_fp16, y = var_2855_cast_fp16)[name = tensor("op_2856_cast_fp16")]; + tensor var_2858_cast_fp16 = add(x = var_2829_cast_fp16, y = var_2856_cast_fp16)[name = tensor("op_2858_cast_fp16")]; + tensor _inversed_t_combined_y_0_to_fp16 = const()[name = tensor("_inversed_t_combined_y_0_to_fp16"), val = tensor(0x1p-1)]; + tensor _inversed_t_combined_cast_fp16 = mul(x = var_2858_cast_fp16, y = _inversed_t_combined_y_0_to_fp16)[name = tensor("_inversed_t_combined_cast_fp16")]; + tensor input_489_cast_fp16 = add(x = _inversed_t_combined_cast_fp16, y = linear_5_cast_fp16)[name = tensor("input_489_cast_fp16")]; + tensor input_491_cast_fp16 = silu(x = input_489_cast_fp16)[name = tensor("input_491_cast_fp16")]; + tensor linear_188_cast_fp16 = linear(bias = flow_net_res_blocks_0_adaLN_modulation_1_bias_to_fp16, weight = flow_net_res_blocks_0_adaLN_modulation_1_weight_to_fp16, x = input_491_cast_fp16)[name = tensor("linear_188_cast_fp16")]; + tensor var_2873_split_sizes_0 = const()[name = tensor("op_2873_split_sizes_0"), val = tensor([512, 512, 512])]; + tensor var_2873_axis_0 = const()[name = tensor("op_2873_axis_0"), val = tensor(-1)]; + tensor var_2873_cast_fp16_0, tensor var_2873_cast_fp16_1, tensor var_2873_cast_fp16_2 = split(axis = var_2873_axis_0, split_sizes = var_2873_split_sizes_0, x = linear_188_cast_fp16)[name = tensor("op_2873_cast_fp16")]; + tensor mean_99_axes_0 = const()[name = tensor("mean_99_axes_0"), val = tensor([-1])]; + tensor mean_99_keep_dims_0 = const()[name = tensor("mean_99_keep_dims_0"), val = tensor(true)]; + tensor mean_99_cast_fp16 = reduce_mean(axes = mean_99_axes_0, keep_dims = mean_99_keep_dims_0, x = linear_182_cast_fp16)[name = tensor("mean_99_cast_fp16")]; + tensor sub_81_cast_fp16 = sub(x = linear_182_cast_fp16, y = mean_99_cast_fp16)[name = tensor("sub_81_cast_fp16")]; + tensor square_65_cast_fp16 = square(x = sub_81_cast_fp16)[name = tensor("square_65_cast_fp16")]; + tensor reduce_mean_131_axes_0 = const()[name = tensor("reduce_mean_131_axes_0"), val = tensor([-1])]; + tensor reduce_mean_131_keep_dims_0 = const()[name = tensor("reduce_mean_131_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_131_cast_fp16 = reduce_mean(axes = reduce_mean_131_axes_0, keep_dims = reduce_mean_131_keep_dims_0, x = square_65_cast_fp16)[name = tensor("reduce_mean_131_cast_fp16")]; + tensor var_2883_to_fp16 = const()[name = tensor("op_2883_to_fp16"), val = tensor(0x1.1p-20)]; + tensor var_2884_cast_fp16 = add(x = reduce_mean_131_cast_fp16, y = var_2883_to_fp16)[name = tensor("op_2884_cast_fp16")]; + tensor var_2885_cast_fp16 = sqrt(x = var_2884_cast_fp16)[name = tensor("op_2885_cast_fp16")]; + tensor x_315_cast_fp16 = real_div(x = sub_81_cast_fp16, y = var_2885_cast_fp16)[name = tensor("x_315_cast_fp16")]; + tensor var_2887_cast_fp16 = mul(x = x_315_cast_fp16, y = flow_net_res_blocks_0_in_ln_weight_to_fp16)[name = tensor("op_2887_cast_fp16")]; + tensor x_317_cast_fp16 = add(x = var_2887_cast_fp16, y = flow_net_res_blocks_0_in_ln_bias_to_fp16)[name = tensor("x_317_cast_fp16")]; + tensor var_2889_promoted_to_fp16 = const()[name = tensor("op_2889_promoted_to_fp16"), val = tensor(0x1p+0)]; + tensor var_2890_cast_fp16 = add(x = var_2873_cast_fp16_1, y = var_2889_promoted_to_fp16)[name = tensor("op_2890_cast_fp16")]; + tensor var_2891_cast_fp16 = mul(x = x_317_cast_fp16, y = var_2890_cast_fp16)[name = tensor("op_2891_cast_fp16")]; + tensor input_493_cast_fp16 = add(x = var_2891_cast_fp16, y = var_2873_cast_fp16_0)[name = tensor("input_493_cast_fp16")]; + tensor linear_189_cast_fp16 = linear(bias = flow_net_res_blocks_0_mlp_0_bias_to_fp16, weight = flow_net_res_blocks_0_mlp_0_weight_to_fp16, x = input_493_cast_fp16)[name = tensor("linear_189_cast_fp16")]; + tensor input_497_cast_fp16 = silu(x = linear_189_cast_fp16)[name = tensor("input_497_cast_fp16")]; + tensor linear_190_cast_fp16 = linear(bias = flow_net_res_blocks_0_mlp_2_bias_to_fp16, weight = flow_net_res_blocks_0_mlp_2_weight_to_fp16, x = input_497_cast_fp16)[name = tensor("linear_190_cast_fp16")]; + tensor var_2902_cast_fp16 = mul(x = var_2873_cast_fp16_2, y = linear_190_cast_fp16)[name = tensor("op_2902_cast_fp16")]; + tensor x_319_cast_fp16 = add(x = linear_182_cast_fp16, y = var_2902_cast_fp16)[name = tensor("x_319_cast_fp16")]; + tensor linear_191_cast_fp16 = linear(bias = flow_net_res_blocks_1_adaLN_modulation_1_bias_to_fp16, weight = flow_net_res_blocks_1_adaLN_modulation_1_weight_to_fp16, x = input_491_cast_fp16)[name = tensor("linear_191_cast_fp16")]; + tensor var_2912_split_sizes_0 = const()[name = tensor("op_2912_split_sizes_0"), val = tensor([512, 512, 512])]; + tensor var_2912_axis_0 = const()[name = tensor("op_2912_axis_0"), val = tensor(-1)]; + tensor var_2912_cast_fp16_0, tensor var_2912_cast_fp16_1, tensor var_2912_cast_fp16_2 = split(axis = var_2912_axis_0, split_sizes = var_2912_split_sizes_0, x = linear_191_cast_fp16)[name = tensor("op_2912_cast_fp16")]; + tensor mean_101_axes_0 = const()[name = tensor("mean_101_axes_0"), val = tensor([-1])]; + tensor mean_101_keep_dims_0 = const()[name = tensor("mean_101_keep_dims_0"), val = tensor(true)]; + tensor mean_101_cast_fp16 = reduce_mean(axes = mean_101_axes_0, keep_dims = mean_101_keep_dims_0, x = x_319_cast_fp16)[name = tensor("mean_101_cast_fp16")]; + tensor sub_82_cast_fp16 = sub(x = x_319_cast_fp16, y = mean_101_cast_fp16)[name = tensor("sub_82_cast_fp16")]; + tensor square_66_cast_fp16 = square(x = sub_82_cast_fp16)[name = tensor("square_66_cast_fp16")]; + tensor reduce_mean_133_axes_0 = const()[name = tensor("reduce_mean_133_axes_0"), val = tensor([-1])]; + tensor reduce_mean_133_keep_dims_0 = const()[name = tensor("reduce_mean_133_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_133_cast_fp16 = reduce_mean(axes = reduce_mean_133_axes_0, keep_dims = reduce_mean_133_keep_dims_0, x = square_66_cast_fp16)[name = tensor("reduce_mean_133_cast_fp16")]; + tensor var_2922_to_fp16 = const()[name = tensor("op_2922_to_fp16"), val = tensor(0x1.1p-20)]; + tensor var_2923_cast_fp16 = add(x = reduce_mean_133_cast_fp16, y = var_2922_to_fp16)[name = tensor("op_2923_cast_fp16")]; + tensor var_2924_cast_fp16 = sqrt(x = var_2923_cast_fp16)[name = tensor("op_2924_cast_fp16")]; + tensor x_321_cast_fp16 = real_div(x = sub_82_cast_fp16, y = var_2924_cast_fp16)[name = tensor("x_321_cast_fp16")]; + tensor var_2926_cast_fp16 = mul(x = x_321_cast_fp16, y = flow_net_res_blocks_1_in_ln_weight_to_fp16)[name = tensor("op_2926_cast_fp16")]; + tensor x_323_cast_fp16 = add(x = var_2926_cast_fp16, y = flow_net_res_blocks_1_in_ln_bias_to_fp16)[name = tensor("x_323_cast_fp16")]; + tensor var_2928_promoted_to_fp16 = const()[name = tensor("op_2928_promoted_to_fp16"), val = tensor(0x1p+0)]; + tensor var_2929_cast_fp16 = add(x = var_2912_cast_fp16_1, y = var_2928_promoted_to_fp16)[name = tensor("op_2929_cast_fp16")]; + tensor var_2930_cast_fp16 = mul(x = x_323_cast_fp16, y = var_2929_cast_fp16)[name = tensor("op_2930_cast_fp16")]; + tensor input_501_cast_fp16 = add(x = var_2930_cast_fp16, y = var_2912_cast_fp16_0)[name = tensor("input_501_cast_fp16")]; + tensor linear_192_cast_fp16 = linear(bias = flow_net_res_blocks_1_mlp_0_bias_to_fp16, weight = flow_net_res_blocks_1_mlp_0_weight_to_fp16, x = input_501_cast_fp16)[name = tensor("linear_192_cast_fp16")]; + tensor input_505_cast_fp16 = silu(x = linear_192_cast_fp16)[name = tensor("input_505_cast_fp16")]; + tensor linear_193_cast_fp16 = linear(bias = flow_net_res_blocks_1_mlp_2_bias_to_fp16, weight = flow_net_res_blocks_1_mlp_2_weight_to_fp16, x = input_505_cast_fp16)[name = tensor("linear_193_cast_fp16")]; + tensor var_2941_cast_fp16 = mul(x = var_2912_cast_fp16_2, y = linear_193_cast_fp16)[name = tensor("op_2941_cast_fp16")]; + tensor x_325_cast_fp16 = add(x = x_319_cast_fp16, y = var_2941_cast_fp16)[name = tensor("x_325_cast_fp16")]; + tensor linear_194_cast_fp16 = linear(bias = flow_net_res_blocks_2_adaLN_modulation_1_bias_to_fp16, weight = flow_net_res_blocks_2_adaLN_modulation_1_weight_to_fp16, x = input_491_cast_fp16)[name = tensor("linear_194_cast_fp16")]; + tensor var_2951_split_sizes_0 = const()[name = tensor("op_2951_split_sizes_0"), val = tensor([512, 512, 512])]; + tensor var_2951_axis_0 = const()[name = tensor("op_2951_axis_0"), val = tensor(-1)]; + tensor var_2951_cast_fp16_0, tensor var_2951_cast_fp16_1, tensor var_2951_cast_fp16_2 = split(axis = var_2951_axis_0, split_sizes = var_2951_split_sizes_0, x = linear_194_cast_fp16)[name = tensor("op_2951_cast_fp16")]; + tensor mean_103_axes_0 = const()[name = tensor("mean_103_axes_0"), val = tensor([-1])]; + tensor mean_103_keep_dims_0 = const()[name = tensor("mean_103_keep_dims_0"), val = tensor(true)]; + tensor mean_103_cast_fp16 = reduce_mean(axes = mean_103_axes_0, keep_dims = mean_103_keep_dims_0, x = x_325_cast_fp16)[name = tensor("mean_103_cast_fp16")]; + tensor sub_83_cast_fp16 = sub(x = x_325_cast_fp16, y = mean_103_cast_fp16)[name = tensor("sub_83_cast_fp16")]; + tensor square_67_cast_fp16 = square(x = sub_83_cast_fp16)[name = tensor("square_67_cast_fp16")]; + tensor reduce_mean_135_axes_0 = const()[name = tensor("reduce_mean_135_axes_0"), val = tensor([-1])]; + tensor reduce_mean_135_keep_dims_0 = const()[name = tensor("reduce_mean_135_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_135_cast_fp16 = reduce_mean(axes = reduce_mean_135_axes_0, keep_dims = reduce_mean_135_keep_dims_0, x = square_67_cast_fp16)[name = tensor("reduce_mean_135_cast_fp16")]; + tensor var_2961_to_fp16 = const()[name = tensor("op_2961_to_fp16"), val = tensor(0x1.1p-20)]; + tensor var_2962_cast_fp16 = add(x = reduce_mean_135_cast_fp16, y = var_2961_to_fp16)[name = tensor("op_2962_cast_fp16")]; + tensor var_2963_cast_fp16 = sqrt(x = var_2962_cast_fp16)[name = tensor("op_2963_cast_fp16")]; + tensor x_327_cast_fp16 = real_div(x = sub_83_cast_fp16, y = var_2963_cast_fp16)[name = tensor("x_327_cast_fp16")]; + tensor var_2965_cast_fp16 = mul(x = x_327_cast_fp16, y = flow_net_res_blocks_2_in_ln_weight_to_fp16)[name = tensor("op_2965_cast_fp16")]; + tensor x_329_cast_fp16 = add(x = var_2965_cast_fp16, y = flow_net_res_blocks_2_in_ln_bias_to_fp16)[name = tensor("x_329_cast_fp16")]; + tensor var_2967_promoted_to_fp16 = const()[name = tensor("op_2967_promoted_to_fp16"), val = tensor(0x1p+0)]; + tensor var_2968_cast_fp16 = add(x = var_2951_cast_fp16_1, y = var_2967_promoted_to_fp16)[name = tensor("op_2968_cast_fp16")]; + tensor var_2969_cast_fp16 = mul(x = x_329_cast_fp16, y = var_2968_cast_fp16)[name = tensor("op_2969_cast_fp16")]; + tensor input_509_cast_fp16 = add(x = var_2969_cast_fp16, y = var_2951_cast_fp16_0)[name = tensor("input_509_cast_fp16")]; + tensor linear_195_cast_fp16 = linear(bias = flow_net_res_blocks_2_mlp_0_bias_to_fp16, weight = flow_net_res_blocks_2_mlp_0_weight_to_fp16, x = input_509_cast_fp16)[name = tensor("linear_195_cast_fp16")]; + tensor input_513_cast_fp16 = silu(x = linear_195_cast_fp16)[name = tensor("input_513_cast_fp16")]; + tensor linear_196_cast_fp16 = linear(bias = flow_net_res_blocks_2_mlp_2_bias_to_fp16, weight = flow_net_res_blocks_2_mlp_2_weight_to_fp16, x = input_513_cast_fp16)[name = tensor("linear_196_cast_fp16")]; + tensor var_2980_cast_fp16 = mul(x = var_2951_cast_fp16_2, y = linear_196_cast_fp16)[name = tensor("op_2980_cast_fp16")]; + tensor x_331_cast_fp16 = add(x = x_325_cast_fp16, y = var_2980_cast_fp16)[name = tensor("x_331_cast_fp16")]; + tensor linear_197_cast_fp16 = linear(bias = flow_net_res_blocks_3_adaLN_modulation_1_bias_to_fp16, weight = flow_net_res_blocks_3_adaLN_modulation_1_weight_to_fp16, x = input_491_cast_fp16)[name = tensor("linear_197_cast_fp16")]; + tensor var_2990_split_sizes_0 = const()[name = tensor("op_2990_split_sizes_0"), val = tensor([512, 512, 512])]; + tensor var_2990_axis_0 = const()[name = tensor("op_2990_axis_0"), val = tensor(-1)]; + tensor var_2990_cast_fp16_0, tensor var_2990_cast_fp16_1, tensor var_2990_cast_fp16_2 = split(axis = var_2990_axis_0, split_sizes = var_2990_split_sizes_0, x = linear_197_cast_fp16)[name = tensor("op_2990_cast_fp16")]; + tensor mean_105_axes_0 = const()[name = tensor("mean_105_axes_0"), val = tensor([-1])]; + tensor mean_105_keep_dims_0 = const()[name = tensor("mean_105_keep_dims_0"), val = tensor(true)]; + tensor mean_105_cast_fp16 = reduce_mean(axes = mean_105_axes_0, keep_dims = mean_105_keep_dims_0, x = x_331_cast_fp16)[name = tensor("mean_105_cast_fp16")]; + tensor sub_84_cast_fp16 = sub(x = x_331_cast_fp16, y = mean_105_cast_fp16)[name = tensor("sub_84_cast_fp16")]; + tensor square_68_cast_fp16 = square(x = sub_84_cast_fp16)[name = tensor("square_68_cast_fp16")]; + tensor reduce_mean_137_axes_0 = const()[name = tensor("reduce_mean_137_axes_0"), val = tensor([-1])]; + tensor reduce_mean_137_keep_dims_0 = const()[name = tensor("reduce_mean_137_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_137_cast_fp16 = reduce_mean(axes = reduce_mean_137_axes_0, keep_dims = reduce_mean_137_keep_dims_0, x = square_68_cast_fp16)[name = tensor("reduce_mean_137_cast_fp16")]; + tensor var_3000_to_fp16 = const()[name = tensor("op_3000_to_fp16"), val = tensor(0x1.1p-20)]; + tensor var_3001_cast_fp16 = add(x = reduce_mean_137_cast_fp16, y = var_3000_to_fp16)[name = tensor("op_3001_cast_fp16")]; + tensor var_3002_cast_fp16 = sqrt(x = var_3001_cast_fp16)[name = tensor("op_3002_cast_fp16")]; + tensor x_333_cast_fp16 = real_div(x = sub_84_cast_fp16, y = var_3002_cast_fp16)[name = tensor("x_333_cast_fp16")]; + tensor var_3004_cast_fp16 = mul(x = x_333_cast_fp16, y = flow_net_res_blocks_3_in_ln_weight_to_fp16)[name = tensor("op_3004_cast_fp16")]; + tensor x_335_cast_fp16 = add(x = var_3004_cast_fp16, y = flow_net_res_blocks_3_in_ln_bias_to_fp16)[name = tensor("x_335_cast_fp16")]; + tensor var_3006_promoted_to_fp16 = const()[name = tensor("op_3006_promoted_to_fp16"), val = tensor(0x1p+0)]; + tensor var_3007_cast_fp16 = add(x = var_2990_cast_fp16_1, y = var_3006_promoted_to_fp16)[name = tensor("op_3007_cast_fp16")]; + tensor var_3008_cast_fp16 = mul(x = x_335_cast_fp16, y = var_3007_cast_fp16)[name = tensor("op_3008_cast_fp16")]; + tensor input_517_cast_fp16 = add(x = var_3008_cast_fp16, y = var_2990_cast_fp16_0)[name = tensor("input_517_cast_fp16")]; + tensor linear_198_cast_fp16 = linear(bias = flow_net_res_blocks_3_mlp_0_bias_to_fp16, weight = flow_net_res_blocks_3_mlp_0_weight_to_fp16, x = input_517_cast_fp16)[name = tensor("linear_198_cast_fp16")]; + tensor input_521_cast_fp16 = silu(x = linear_198_cast_fp16)[name = tensor("input_521_cast_fp16")]; + tensor linear_199_cast_fp16 = linear(bias = flow_net_res_blocks_3_mlp_2_bias_to_fp16, weight = flow_net_res_blocks_3_mlp_2_weight_to_fp16, x = input_521_cast_fp16)[name = tensor("linear_199_cast_fp16")]; + tensor var_3019_cast_fp16 = mul(x = var_2990_cast_fp16_2, y = linear_199_cast_fp16)[name = tensor("op_3019_cast_fp16")]; + tensor x_337_cast_fp16 = add(x = x_331_cast_fp16, y = var_3019_cast_fp16)[name = tensor("x_337_cast_fp16")]; + tensor linear_200_cast_fp16 = linear(bias = flow_net_res_blocks_4_adaLN_modulation_1_bias_to_fp16, weight = flow_net_res_blocks_4_adaLN_modulation_1_weight_to_fp16, x = input_491_cast_fp16)[name = tensor("linear_200_cast_fp16")]; + tensor var_3029_split_sizes_0 = const()[name = tensor("op_3029_split_sizes_0"), val = tensor([512, 512, 512])]; + tensor var_3029_axis_0 = const()[name = tensor("op_3029_axis_0"), val = tensor(-1)]; + tensor var_3029_cast_fp16_0, tensor var_3029_cast_fp16_1, tensor var_3029_cast_fp16_2 = split(axis = var_3029_axis_0, split_sizes = var_3029_split_sizes_0, x = linear_200_cast_fp16)[name = tensor("op_3029_cast_fp16")]; + tensor mean_107_axes_0 = const()[name = tensor("mean_107_axes_0"), val = tensor([-1])]; + tensor mean_107_keep_dims_0 = const()[name = tensor("mean_107_keep_dims_0"), val = tensor(true)]; + tensor mean_107_cast_fp16 = reduce_mean(axes = mean_107_axes_0, keep_dims = mean_107_keep_dims_0, x = x_337_cast_fp16)[name = tensor("mean_107_cast_fp16")]; + tensor sub_85_cast_fp16 = sub(x = x_337_cast_fp16, y = mean_107_cast_fp16)[name = tensor("sub_85_cast_fp16")]; + tensor square_69_cast_fp16 = square(x = sub_85_cast_fp16)[name = tensor("square_69_cast_fp16")]; + tensor reduce_mean_139_axes_0 = const()[name = tensor("reduce_mean_139_axes_0"), val = tensor([-1])]; + tensor reduce_mean_139_keep_dims_0 = const()[name = tensor("reduce_mean_139_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_139_cast_fp16 = reduce_mean(axes = reduce_mean_139_axes_0, keep_dims = reduce_mean_139_keep_dims_0, x = square_69_cast_fp16)[name = tensor("reduce_mean_139_cast_fp16")]; + tensor var_3039_to_fp16 = const()[name = tensor("op_3039_to_fp16"), val = tensor(0x1.1p-20)]; + tensor var_3040_cast_fp16 = add(x = reduce_mean_139_cast_fp16, y = var_3039_to_fp16)[name = tensor("op_3040_cast_fp16")]; + tensor var_3041_cast_fp16 = sqrt(x = var_3040_cast_fp16)[name = tensor("op_3041_cast_fp16")]; + tensor x_339_cast_fp16 = real_div(x = sub_85_cast_fp16, y = var_3041_cast_fp16)[name = tensor("x_339_cast_fp16")]; + tensor var_3043_cast_fp16 = mul(x = x_339_cast_fp16, y = flow_net_res_blocks_4_in_ln_weight_to_fp16)[name = tensor("op_3043_cast_fp16")]; + tensor x_341_cast_fp16 = add(x = var_3043_cast_fp16, y = flow_net_res_blocks_4_in_ln_bias_to_fp16)[name = tensor("x_341_cast_fp16")]; + tensor var_3045_promoted_to_fp16 = const()[name = tensor("op_3045_promoted_to_fp16"), val = tensor(0x1p+0)]; + tensor var_3046_cast_fp16 = add(x = var_3029_cast_fp16_1, y = var_3045_promoted_to_fp16)[name = tensor("op_3046_cast_fp16")]; + tensor var_3047_cast_fp16 = mul(x = x_341_cast_fp16, y = var_3046_cast_fp16)[name = tensor("op_3047_cast_fp16")]; + tensor input_525_cast_fp16 = add(x = var_3047_cast_fp16, y = var_3029_cast_fp16_0)[name = tensor("input_525_cast_fp16")]; + tensor linear_201_cast_fp16 = linear(bias = flow_net_res_blocks_4_mlp_0_bias_to_fp16, weight = flow_net_res_blocks_4_mlp_0_weight_to_fp16, x = input_525_cast_fp16)[name = tensor("linear_201_cast_fp16")]; + tensor input_529_cast_fp16 = silu(x = linear_201_cast_fp16)[name = tensor("input_529_cast_fp16")]; + tensor linear_202_cast_fp16 = linear(bias = flow_net_res_blocks_4_mlp_2_bias_to_fp16, weight = flow_net_res_blocks_4_mlp_2_weight_to_fp16, x = input_529_cast_fp16)[name = tensor("linear_202_cast_fp16")]; + tensor var_3058_cast_fp16 = mul(x = var_3029_cast_fp16_2, y = linear_202_cast_fp16)[name = tensor("op_3058_cast_fp16")]; + tensor x_343_cast_fp16 = add(x = x_337_cast_fp16, y = var_3058_cast_fp16)[name = tensor("x_343_cast_fp16")]; + tensor linear_203_cast_fp16 = linear(bias = flow_net_res_blocks_5_adaLN_modulation_1_bias_to_fp16, weight = flow_net_res_blocks_5_adaLN_modulation_1_weight_to_fp16, x = input_491_cast_fp16)[name = tensor("linear_203_cast_fp16")]; + tensor var_3068_split_sizes_0 = const()[name = tensor("op_3068_split_sizes_0"), val = tensor([512, 512, 512])]; + tensor var_3068_axis_0 = const()[name = tensor("op_3068_axis_0"), val = tensor(-1)]; + tensor var_3068_cast_fp16_0, tensor var_3068_cast_fp16_1, tensor var_3068_cast_fp16_2 = split(axis = var_3068_axis_0, split_sizes = var_3068_split_sizes_0, x = linear_203_cast_fp16)[name = tensor("op_3068_cast_fp16")]; + tensor mean_109_axes_0 = const()[name = tensor("mean_109_axes_0"), val = tensor([-1])]; + tensor mean_109_keep_dims_0 = const()[name = tensor("mean_109_keep_dims_0"), val = tensor(true)]; + tensor mean_109_cast_fp16 = reduce_mean(axes = mean_109_axes_0, keep_dims = mean_109_keep_dims_0, x = x_343_cast_fp16)[name = tensor("mean_109_cast_fp16")]; + tensor sub_86_cast_fp16 = sub(x = x_343_cast_fp16, y = mean_109_cast_fp16)[name = tensor("sub_86_cast_fp16")]; + tensor square_70_cast_fp16 = square(x = sub_86_cast_fp16)[name = tensor("square_70_cast_fp16")]; + tensor reduce_mean_141_axes_0 = const()[name = tensor("reduce_mean_141_axes_0"), val = tensor([-1])]; + tensor reduce_mean_141_keep_dims_0 = const()[name = tensor("reduce_mean_141_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_141_cast_fp16 = reduce_mean(axes = reduce_mean_141_axes_0, keep_dims = reduce_mean_141_keep_dims_0, x = square_70_cast_fp16)[name = tensor("reduce_mean_141_cast_fp16")]; + tensor var_3078_to_fp16 = const()[name = tensor("op_3078_to_fp16"), val = tensor(0x1.1p-20)]; + tensor var_3079_cast_fp16 = add(x = reduce_mean_141_cast_fp16, y = var_3078_to_fp16)[name = tensor("op_3079_cast_fp16")]; + tensor var_3080_cast_fp16 = sqrt(x = var_3079_cast_fp16)[name = tensor("op_3080_cast_fp16")]; + tensor x_345_cast_fp16 = real_div(x = sub_86_cast_fp16, y = var_3080_cast_fp16)[name = tensor("x_345_cast_fp16")]; + tensor var_3082_cast_fp16 = mul(x = x_345_cast_fp16, y = flow_net_res_blocks_5_in_ln_weight_to_fp16)[name = tensor("op_3082_cast_fp16")]; + tensor x_347_cast_fp16 = add(x = var_3082_cast_fp16, y = flow_net_res_blocks_5_in_ln_bias_to_fp16)[name = tensor("x_347_cast_fp16")]; + tensor var_3084_promoted_to_fp16 = const()[name = tensor("op_3084_promoted_to_fp16"), val = tensor(0x1p+0)]; + tensor var_3085_cast_fp16 = add(x = var_3068_cast_fp16_1, y = var_3084_promoted_to_fp16)[name = tensor("op_3085_cast_fp16")]; + tensor var_3086_cast_fp16 = mul(x = x_347_cast_fp16, y = var_3085_cast_fp16)[name = tensor("op_3086_cast_fp16")]; + tensor input_533_cast_fp16 = add(x = var_3086_cast_fp16, y = var_3068_cast_fp16_0)[name = tensor("input_533_cast_fp16")]; + tensor linear_204_cast_fp16 = linear(bias = flow_net_res_blocks_5_mlp_0_bias_to_fp16, weight = flow_net_res_blocks_5_mlp_0_weight_to_fp16, x = input_533_cast_fp16)[name = tensor("linear_204_cast_fp16")]; + tensor input_537_cast_fp16 = silu(x = linear_204_cast_fp16)[name = tensor("input_537_cast_fp16")]; + tensor linear_205_cast_fp16 = linear(bias = flow_net_res_blocks_5_mlp_2_bias_to_fp16, weight = flow_net_res_blocks_5_mlp_2_weight_to_fp16, x = input_537_cast_fp16)[name = tensor("linear_205_cast_fp16")]; + tensor var_3097_cast_fp16 = mul(x = var_3068_cast_fp16_2, y = linear_205_cast_fp16)[name = tensor("op_3097_cast_fp16")]; + tensor x_349_cast_fp16 = add(x = x_343_cast_fp16, y = var_3097_cast_fp16)[name = tensor("x_349_cast_fp16")]; + tensor linear_206_cast_fp16 = linear(bias = flow_net_final_layer_adaLN_modulation_1_bias_to_fp16, weight = flow_net_final_layer_adaLN_modulation_1_weight_to_fp16, x = input_491_cast_fp16)[name = tensor("linear_206_cast_fp16")]; + tensor var_3106_split_sizes_0 = const()[name = tensor("op_3106_split_sizes_0"), val = tensor([512, 512])]; + tensor var_3106_axis_0 = const()[name = tensor("op_3106_axis_0"), val = tensor(-1)]; + tensor var_3106_cast_fp16_0, tensor var_3106_cast_fp16_1 = split(axis = var_3106_axis_0, split_sizes = var_3106_split_sizes_0, x = linear_206_cast_fp16)[name = tensor("op_3106_cast_fp16")]; + tensor mean_axes_0 = const()[name = tensor("mean_axes_0"), val = tensor([-1])]; + tensor mean_keep_dims_0 = const()[name = tensor("mean_keep_dims_0"), val = tensor(true)]; + tensor mean_cast_fp16 = reduce_mean(axes = mean_axes_0, keep_dims = mean_keep_dims_0, x = x_349_cast_fp16)[name = tensor("mean_cast_fp16")]; + tensor sub_87_cast_fp16 = sub(x = x_349_cast_fp16, y = mean_cast_fp16)[name = tensor("sub_87_cast_fp16")]; + tensor square_71_cast_fp16 = square(x = sub_87_cast_fp16)[name = tensor("square_71_cast_fp16")]; + tensor reduce_mean_143_axes_0 = const()[name = tensor("reduce_mean_143_axes_0"), val = tensor([-1])]; + tensor reduce_mean_143_keep_dims_0 = const()[name = tensor("reduce_mean_143_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_143_cast_fp16 = reduce_mean(axes = reduce_mean_143_axes_0, keep_dims = reduce_mean_143_keep_dims_0, x = square_71_cast_fp16)[name = tensor("reduce_mean_143_cast_fp16")]; + tensor var_3113_to_fp16 = const()[name = tensor("op_3113_to_fp16"), val = tensor(0x1.1p-20)]; + tensor var_3114_cast_fp16 = add(x = reduce_mean_143_cast_fp16, y = var_3113_to_fp16)[name = tensor("op_3114_cast_fp16")]; + tensor var_3115_cast_fp16 = sqrt(x = var_3114_cast_fp16)[name = tensor("op_3115_cast_fp16")]; + tensor x_cast_fp16 = real_div(x = sub_87_cast_fp16, y = var_3115_cast_fp16)[name = tensor("x_cast_fp16")]; + tensor var_3117_promoted_to_fp16 = const()[name = tensor("op_3117_promoted_to_fp16"), val = tensor(0x1p+0)]; + tensor var_3118_cast_fp16 = add(x = var_3106_cast_fp16_1, y = var_3117_promoted_to_fp16)[name = tensor("op_3118_cast_fp16")]; + tensor var_3119_cast_fp16 = mul(x = x_cast_fp16, y = var_3118_cast_fp16)[name = tensor("op_3119_cast_fp16")]; + tensor input_cast_fp16 = add(x = var_3119_cast_fp16, y = var_3106_cast_fp16_0)[name = tensor("input_cast_fp16")]; + tensor linear_207_cast_fp16 = linear(bias = flow_net_final_layer_linear_bias_to_fp16, weight = flow_net_final_layer_linear_weight_to_fp16, x = input_cast_fp16)[name = tensor("linear_207_cast_fp16")]; + tensor var_3124_to_fp16 = const()[name = tensor("op_3124_to_fp16"), val = tensor(0x1p-3)]; + tensor var_3125_cast_fp16 = mul(x = linear_207_cast_fp16, y = var_3124_to_fp16)[name = tensor("op_3125_cast_fp16")]; + tensor var_3127_cast_fp16 = add(x = input_475_cast_fp16, y = var_3125_cast_fp16)[name = tensor("op_3127_cast_fp16")]; + tensor var_3127_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_3127_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor latent_final = cast(dtype = var_3127_cast_fp16_to_fp32_dtype_0, x = var_3127_cast_fp16)[name = tensor("cast_144")]; + } -> (latent_final); +} \ No newline at end of file diff --git a/v2.1/german_24l/flow_decoder_fused.mlmodelc/weights/weight.bin b/v2.1/german_24l/flow_decoder_fused.mlmodelc/weights/weight.bin new file mode 100644 index 0000000000000000000000000000000000000000..3bf21f2fed8a3fdb7abedacc3e096c89597b91cd --- /dev/null +++ b/v2.1/german_24l/flow_decoder_fused.mlmodelc/weights/weight.bin @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:6ee76653a292aaee52385edda87a2e5da948e1fbbb676ebf06e8d253f6582b03 +size 19012608 diff --git a/v2.1/german_24l/flow_decoder_fused.mlpackage/Data/com.apple.CoreML/model.mlmodel b/v2.1/german_24l/flow_decoder_fused.mlpackage/Data/com.apple.CoreML/model.mlmodel new file mode 100644 index 0000000000000000000000000000000000000000..b713225be2bb4b666ad0ae642d962d0c4fe8f7a0 --- /dev/null +++ b/v2.1/german_24l/flow_decoder_fused.mlpackage/Data/com.apple.CoreML/model.mlmodel @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:769f124d1cc051148970e7cf259a1e9ac6065e11ccacce1971241990fc224f7b +size 301733 diff --git a/v2.1/german_24l/flow_decoder_fused.mlpackage/Data/com.apple.CoreML/weights/weight.bin b/v2.1/german_24l/flow_decoder_fused.mlpackage/Data/com.apple.CoreML/weights/weight.bin new file mode 100644 index 0000000000000000000000000000000000000000..3bf21f2fed8a3fdb7abedacc3e096c89597b91cd --- /dev/null +++ b/v2.1/german_24l/flow_decoder_fused.mlpackage/Data/com.apple.CoreML/weights/weight.bin @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:6ee76653a292aaee52385edda87a2e5da948e1fbbb676ebf06e8d253f6582b03 +size 19012608 diff --git a/v2.1/german_24l/flow_decoder_fused.mlpackage/Manifest.json b/v2.1/german_24l/flow_decoder_fused.mlpackage/Manifest.json new file mode 100644 index 0000000000000000000000000000000000000000..a59614b2541c0873b5c9e02676a44c8268a446aa --- /dev/null +++ b/v2.1/german_24l/flow_decoder_fused.mlpackage/Manifest.json @@ -0,0 +1,18 @@ +{ + "fileFormatVersion": "1.0.0", + "itemInfoEntries": { + "6148FF90-6235-4DD0-9F77-4625A590468D": { + "author": "com.apple.CoreML", + "description": "CoreML Model Weights", + "name": "weights", + "path": "com.apple.CoreML/weights" + }, + "EF5949F6-A979-43AE-BAC0-B755622A10D2": { + "author": "com.apple.CoreML", + "description": "CoreML Model Specification", + "name": "model.mlmodel", + "path": "com.apple.CoreML/model.mlmodel" + } + }, + "rootModelIdentifier": "EF5949F6-A979-43AE-BAC0-B755622A10D2" +} diff --git a/v2.1/german_24l/flowlm_step.mlmodelc/analytics/coremldata.bin b/v2.1/german_24l/flowlm_step.mlmodelc/analytics/coremldata.bin new file mode 100644 index 0000000000000000000000000000000000000000..c5946217b7ba099f8c73e881126c492fb543d719 --- /dev/null +++ b/v2.1/german_24l/flowlm_step.mlmodelc/analytics/coremldata.bin @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:62aa4609f9c87c4a70769d04b2ec8faf34e48c4712205561178c9c6b6ad0a5c3 +size 243 diff --git a/v2.1/german_24l/flowlm_step.mlmodelc/coremldata.bin b/v2.1/german_24l/flowlm_step.mlmodelc/coremldata.bin new file mode 100644 index 0000000000000000000000000000000000000000..038a3dcda6213d061bb1635e99151b96308b4ec7 --- /dev/null +++ b/v2.1/german_24l/flowlm_step.mlmodelc/coremldata.bin @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:698b266d26fbbf9dc967cbfd2b94ae1466679cb7a3bb64c3150c9368acc7ba22 +size 3559 diff --git a/v2.1/german_24l/flowlm_step.mlmodelc/model.mil b/v2.1/german_24l/flowlm_step.mlmodelc/model.mil new file mode 100644 index 0000000000000000000000000000000000000000..f4deb57a696051a81d5878c658d52db6bf7db9b7 --- /dev/null +++ b/v2.1/german_24l/flowlm_step.mlmodelc/model.mil @@ -0,0 +1,5337 @@ +program(1.0) +[buildInfo = dict, tensor>({{"coremlc-component-MIL", "3520.4.1"}, {"coremlc-version", "3520.5.1"}, {"coremltools-component-torch", "2.12.0"}, {"coremltools-source-dialect", "TorchScript"}, {"coremltools-version", "9.0"}})] +{ + func main(tensor bos_emb, tensor cache0, tensor cache1, tensor cache10, tensor cache11, tensor cache12, tensor cache13, tensor cache14, tensor cache15, tensor cache16, tensor cache17, tensor cache18, tensor cache19, tensor cache2, tensor cache20, tensor cache21, tensor cache22, tensor cache23, tensor cache3, tensor cache4, tensor cache5, tensor cache6, tensor cache7, tensor cache8, tensor cache9, tensor position0, tensor position1, tensor position10, tensor position11, tensor position12, tensor position13, tensor position14, tensor position15, tensor position16, tensor position17, tensor position18, tensor position19, tensor position2, tensor position20, tensor position21, tensor position22, tensor position23, tensor position3, tensor position4, tensor position5, tensor position6, tensor position7, tensor position8, tensor position9, tensor sequence) { + tensor sequence_to_fp16_dtype_0 = const()[name = tensor("sequence_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor sequence_to_fp16 = cast(dtype = sequence_to_fp16_dtype_0, x = sequence)[name = tensor("cast_555")]; + tensor var_198_cast_fp16 = not_equal(x = sequence_to_fp16, y = sequence_to_fp16)[name = tensor("op_198_cast_fp16")]; + tensor expand_dims_0_axes_0 = const()[name = tensor("expand_dims_0_axes_0"), val = tensor([0, 1])]; + tensor bos_emb_to_fp16_dtype_0 = const()[name = tensor("bos_emb_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor bos_emb_to_fp16 = cast(dtype = bos_emb_to_fp16_dtype_0, x = bos_emb)[name = tensor("cast_554")]; + tensor expand_dims_0_cast_fp16 = expand_dims(axes = expand_dims_0_axes_0, x = bos_emb_to_fp16)[name = tensor("expand_dims_0_cast_fp16")]; + tensor input_1_cast_fp16 = select(a = expand_dims_0_cast_fp16, b = sequence_to_fp16, cond = var_198_cast_fp16)[name = tensor("input_1_cast_fp16")]; + tensor input_linear_weight_to_fp16 = const()[name = tensor("input_linear_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(64)))]; + tensor linear_0_bias_0_to_fp16 = const()[name = tensor("linear_0_bias_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(65664)))]; + tensor linear_0_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = input_linear_weight_to_fp16, x = input_1_cast_fp16)[name = tensor("linear_0_cast_fp16")]; + tensor input_5_axes_0 = const()[name = tensor("input_5_axes_0"), val = tensor([-1])]; + tensor norm0_1_weight_to_fp16 = const()[name = tensor("norm0_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(67776)))]; + tensor norm0_1_bias_to_fp16 = const()[name = tensor("norm0_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(69888)))]; + tensor var_204_to_fp16 = const()[name = tensor("op_204_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_5_cast_fp16 = layer_norm(axes = input_5_axes_0, beta = norm0_1_bias_to_fp16, epsilon = var_204_to_fp16, gamma = norm0_1_weight_to_fp16, x = linear_0_cast_fp16)[name = tensor("input_5_cast_fp16")]; + tensor attn0_in_proj_weight_to_fp16 = const()[name = tensor("attn0_in_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(72000)))]; + tensor linear_1_bias_0_to_fp16 = const()[name = tensor("linear_1_bias_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(6363520)))]; + tensor linear_1_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = attn0_in_proj_weight_to_fp16, x = input_5_cast_fp16)[name = tensor("linear_1_cast_fp16")]; + tensor var_218 = const()[name = tensor("op_218"), val = tensor([1, 1, 3, 16, 64])]; + tensor qkv_1_cast_fp16 = reshape(shape = var_218, x = linear_1_cast_fp16)[name = tensor("qkv_1_cast_fp16")]; + tensor q_1_begin_0 = const()[name = tensor("q_1_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor q_1_end_0 = const()[name = tensor("q_1_end_0"), val = tensor([1, 1, 1, 16, 64])]; + tensor q_1_end_mask_0 = const()[name = tensor("q_1_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor q_1_squeeze_mask_0 = const()[name = tensor("q_1_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor q_1_cast_fp16 = slice_by_index(begin = q_1_begin_0, end = q_1_end_0, end_mask = q_1_end_mask_0, squeeze_mask = q_1_squeeze_mask_0, x = qkv_1_cast_fp16)[name = tensor("q_1_cast_fp16")]; + tensor k_1_begin_0 = const()[name = tensor("k_1_begin_0"), val = tensor([0, 0, 1, 0, 0])]; + tensor k_1_end_0 = const()[name = tensor("k_1_end_0"), val = tensor([1, 1, 2, 16, 64])]; + tensor k_1_end_mask_0 = const()[name = tensor("k_1_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor k_1_squeeze_mask_0 = const()[name = tensor("k_1_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor k_1_cast_fp16 = slice_by_index(begin = k_1_begin_0, end = k_1_end_0, end_mask = k_1_end_mask_0, squeeze_mask = k_1_squeeze_mask_0, x = qkv_1_cast_fp16)[name = tensor("k_1_cast_fp16")]; + tensor v_1_begin_0 = const()[name = tensor("v_1_begin_0"), val = tensor([0, 0, 2, 0, 0])]; + tensor v_1_end_0 = const()[name = tensor("v_1_end_0"), val = tensor([1, 1, 3, 16, 64])]; + tensor v_1_end_mask_0 = const()[name = tensor("v_1_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor v_1_squeeze_mask_0 = const()[name = tensor("v_1_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor v_1_cast_fp16 = slice_by_index(begin = v_1_begin_0, end = v_1_end_0, end_mask = v_1_end_mask_0, squeeze_mask = v_1_squeeze_mask_0, x = qkv_1_cast_fp16)[name = tensor("v_1_cast_fp16")]; + tensor var_284 = const()[name = tensor("op_284"), val = tensor([1, 1, 1, 1])]; + tensor position0_to_fp16_dtype_0 = const()[name = tensor("position0_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor position0_to_fp16 = cast(dtype = position0_to_fp16_dtype_0, x = position0)[name = tensor("cast_553")]; + tensor ts_5_cast_fp16 = reshape(shape = var_284, x = position0_to_fp16)[name = tensor("ts_5_cast_fp16")]; + tensor var_291 = const()[name = tensor("op_291"), val = tensor([1, 1, 16, 32, 2])]; + tensor q_complex_1_cast_fp16 = reshape(shape = var_291, x = q_1_cast_fp16)[name = tensor("q_complex_1_cast_fp16")]; + tensor var_298 = const()[name = tensor("op_298"), val = tensor([1, 1, 16, 32, 2])]; + tensor k_complex_1_cast_fp16 = reshape(shape = var_298, x = k_1_cast_fp16)[name = tensor("k_complex_1_cast_fp16")]; + tensor var_302_begin_0 = const()[name = tensor("op_302_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_302_end_0 = const()[name = tensor("op_302_end_0"), val = tensor([1, 1, 16, 32, 1])]; + tensor var_302_end_mask_0 = const()[name = tensor("op_302_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_302_squeeze_mask_0 = const()[name = tensor("op_302_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_302_cast_fp16 = slice_by_index(begin = var_302_begin_0, end = var_302_end_0, end_mask = var_302_end_mask_0, squeeze_mask = var_302_squeeze_mask_0, x = q_complex_1_cast_fp16)[name = tensor("op_302_cast_fp16")]; + tensor var_310_begin_0 = const()[name = tensor("op_310_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_310_end_0 = const()[name = tensor("op_310_end_0"), val = tensor([1, 1, 16, 32, 2])]; + tensor var_310_end_mask_0 = const()[name = tensor("op_310_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_310_squeeze_mask_0 = const()[name = tensor("op_310_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_310_cast_fp16 = slice_by_index(begin = var_310_begin_0, end = var_310_end_0, end_mask = var_310_end_mask_0, squeeze_mask = var_310_squeeze_mask_0, x = q_complex_1_cast_fp16)[name = tensor("op_310_cast_fp16")]; + tensor var_318_begin_0 = const()[name = tensor("op_318_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_318_end_0 = const()[name = tensor("op_318_end_0"), val = tensor([1, 1, 16, 32, 1])]; + tensor var_318_end_mask_0 = const()[name = tensor("op_318_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_318_squeeze_mask_0 = const()[name = tensor("op_318_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_318_cast_fp16 = slice_by_index(begin = var_318_begin_0, end = var_318_end_0, end_mask = var_318_end_mask_0, squeeze_mask = var_318_squeeze_mask_0, x = k_complex_1_cast_fp16)[name = tensor("op_318_cast_fp16")]; + tensor var_326_begin_0 = const()[name = tensor("op_326_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_326_end_0 = const()[name = tensor("op_326_end_0"), val = tensor([1, 1, 16, 32, 2])]; + tensor var_326_end_mask_0 = const()[name = tensor("op_326_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_326_squeeze_mask_0 = const()[name = tensor("op_326_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_326_cast_fp16 = slice_by_index(begin = var_326_begin_0, end = var_326_end_0, end_mask = var_326_end_mask_0, squeeze_mask = var_326_squeeze_mask_0, x = k_complex_1_cast_fp16)[name = tensor("op_326_cast_fp16")]; + tensor freqs_1_to_fp16 = const()[name = tensor("freqs_1_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(6369728)))]; + tensor var_332_cast_fp16 = mul(x = freqs_1_to_fp16, y = ts_5_cast_fp16)[name = tensor("op_332_cast_fp16")]; + tensor rotr_1_cast_fp16 = cos(x = var_332_cast_fp16)[name = tensor("rotr_1_cast_fp16")]; + tensor roti_1_cast_fp16 = sin(x = var_332_cast_fp16)[name = tensor("roti_1_cast_fp16")]; + tensor var_336_cast_fp16 = mul(x = var_302_cast_fp16, y = rotr_1_cast_fp16)[name = tensor("op_336_cast_fp16")]; + tensor var_337_cast_fp16 = mul(x = var_310_cast_fp16, y = roti_1_cast_fp16)[name = tensor("op_337_cast_fp16")]; + tensor qor_1_cast_fp16 = sub(x = var_336_cast_fp16, y = var_337_cast_fp16)[name = tensor("qor_1_cast_fp16")]; + tensor var_340_cast_fp16 = mul(x = var_302_cast_fp16, y = roti_1_cast_fp16)[name = tensor("op_340_cast_fp16")]; + tensor var_341_cast_fp16 = mul(x = var_310_cast_fp16, y = rotr_1_cast_fp16)[name = tensor("op_341_cast_fp16")]; + tensor qoi_1_cast_fp16 = add(x = var_340_cast_fp16, y = var_341_cast_fp16)[name = tensor("qoi_1_cast_fp16")]; + tensor var_344_cast_fp16 = mul(x = var_318_cast_fp16, y = rotr_1_cast_fp16)[name = tensor("op_344_cast_fp16")]; + tensor var_345_cast_fp16 = mul(x = var_326_cast_fp16, y = roti_1_cast_fp16)[name = tensor("op_345_cast_fp16")]; + tensor kor_1_cast_fp16 = sub(x = var_344_cast_fp16, y = var_345_cast_fp16)[name = tensor("kor_1_cast_fp16")]; + tensor var_348_cast_fp16 = mul(x = var_318_cast_fp16, y = roti_1_cast_fp16)[name = tensor("op_348_cast_fp16")]; + tensor var_349_cast_fp16 = mul(x = var_326_cast_fp16, y = rotr_1_cast_fp16)[name = tensor("op_349_cast_fp16")]; + tensor koi_1_cast_fp16 = add(x = var_348_cast_fp16, y = var_349_cast_fp16)[name = tensor("koi_1_cast_fp16")]; + tensor qo_1_axis_0 = const()[name = tensor("qo_1_axis_0"), val = tensor(-1)]; + tensor qo_1_cast_fp16 = stack(axis = qo_1_axis_0, values = (qor_1_cast_fp16, qoi_1_cast_fp16))[name = tensor("qo_1_cast_fp16")]; + tensor ko_1_axis_0 = const()[name = tensor("ko_1_axis_0"), val = tensor(-1)]; + tensor ko_1_cast_fp16 = stack(axis = ko_1_axis_0, values = (kor_1_cast_fp16, koi_1_cast_fp16))[name = tensor("ko_1_cast_fp16")]; + tensor var_382 = const()[name = tensor("op_382"), val = tensor([1, 1, 16, 64])]; + tensor q_3_cast_fp16 = reshape(shape = var_382, x = qo_1_cast_fp16)[name = tensor("q_3_cast_fp16")]; + tensor var_388 = const()[name = tensor("op_388"), val = tensor([1, 1, 16, 64])]; + tensor k_3_cast_fp16 = reshape(shape = var_388, x = ko_1_cast_fp16)[name = tensor("k_3_cast_fp16")]; + tensor _inversed_413_y_0_to_fp16 = const()[name = tensor("_inversed_413_y_0_to_fp16"), val = tensor(0x1p-9)]; + tensor _inversed_413_cast_fp16 = mul(x = ts_5_cast_fp16, y = _inversed_413_y_0_to_fp16)[name = tensor("_inversed_413_cast_fp16")]; + tensor var_414_cast_fp16 = floor(x = _inversed_413_cast_fp16)[name = tensor("op_414_cast_fp16")]; + tensor var_415_to_fp16 = const()[name = tensor("op_415_to_fp16"), val = tensor(0x1p+9)]; + tensor var_416_cast_fp16 = mul(x = var_414_cast_fp16, y = var_415_to_fp16)[name = tensor("op_416_cast_fp16")]; + tensor write_indices_float_3_cast_fp16 = sub(x = ts_5_cast_fp16, y = var_416_cast_fp16)[name = tensor("write_indices_float_3_cast_fp16")]; + tensor var_423_dtype_0 = const()[name = tensor("op_423_dtype_0"), val = tensor("int32")]; + tensor write_indices_1_reps_0 = const()[name = tensor("write_indices_1_reps_0"), val = tensor([1, 1, 16, 64])]; + tensor write_indices_float_3_cast_fp16_to_int32 = cast(dtype = var_423_dtype_0, x = write_indices_float_3_cast_fp16)[name = tensor("cast_552")]; + tensor write_indices_1 = tile(reps = write_indices_1_reps_0, x = write_indices_float_3_cast_fp16_to_int32)[name = tensor("write_indices_1")]; + tensor var_433_begin_0 = const()[name = tensor("op_433_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_433_end_0 = const()[name = tensor("op_433_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor var_433_end_mask_0 = const()[name = tensor("op_433_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_433_squeeze_mask_0 = const()[name = tensor("op_433_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor cache0_to_fp16_dtype_0 = const()[name = tensor("cache0_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor cache0_to_fp16 = cast(dtype = cache0_to_fp16_dtype_0, x = cache0)[name = tensor("cast_551")]; + tensor var_433_cast_fp16 = slice_by_index(begin = var_433_begin_0, end = var_433_end_0, end_mask = var_433_end_mask_0, squeeze_mask = var_433_squeeze_mask_0, x = cache0_to_fp16)[name = tensor("op_433_cast_fp16")]; + tensor var_435_axis_0 = const()[name = tensor("op_435_axis_0"), val = tensor(1)]; + tensor var_435_mode_0 = const()[name = tensor("op_435_mode_0"), val = tensor("update")]; + tensor var_435_validate_indices_0 = const()[name = tensor("op_435_validate_indices_0"), val = tensor(false)]; + tensor var_435_cast_fp16 = scatter_along_axis(axis = var_435_axis_0, data = var_433_cast_fp16, indices = write_indices_1, mode = var_435_mode_0, updates = k_3_cast_fp16, validate_indices = var_435_validate_indices_0)[name = tensor("op_435_cast_fp16")]; + tensor concat_2 = const()[name = tensor("concat_2"), val = tensor([0, 0, 0, 0, 0])]; + tensor concat_3 = const()[name = tensor("concat_3"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_1_internal_tensor_assign_1_stride_0 = const()[name = tensor("new_cache_1_internal_tensor_assign_1_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_1_internal_tensor_assign_1_begin_mask_0 = const()[name = tensor("new_cache_1_internal_tensor_assign_1_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_1_internal_tensor_assign_1_end_mask_0 = const()[name = tensor("new_cache_1_internal_tensor_assign_1_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_1_internal_tensor_assign_1_squeeze_mask_0 = const()[name = tensor("new_cache_1_internal_tensor_assign_1_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_48 = const()[name = tensor("shape_48"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_0 = const()[name = tensor("reduce_prod_0"), val = tensor(1048576)]; + tensor range_1d_0_start_0 = const()[name = tensor("range_1d_0_start_0"), val = tensor(0)]; + tensor range_1d_0_step_0 = const()[name = tensor("range_1d_0_step_0"), val = tensor(1)]; + tensor range_1d_0 = range_1d(end = reduce_prod_0, start = range_1d_0_start_0, step = range_1d_0_step_0)[name = tensor("range_1d_0")]; + tensor reshape_0 = reshape(shape = shape_48, x = range_1d_0)[name = tensor("reshape_0")]; + tensor slice_by_index_0 = slice_by_index(begin = concat_2, begin_mask = new_cache_1_internal_tensor_assign_1_begin_mask_0, end = concat_3, end_mask = new_cache_1_internal_tensor_assign_1_end_mask_0, squeeze_mask = new_cache_1_internal_tensor_assign_1_squeeze_mask_0, stride = new_cache_1_internal_tensor_assign_1_stride_0, x = reshape_0)[name = tensor("slice_by_index_0")]; + tensor reshape_1_shape_0 = const()[name = tensor("reshape_1_shape_0"), val = tensor([-1])]; + tensor reshape_1 = reshape(shape = reshape_1_shape_0, x = slice_by_index_0)[name = tensor("reshape_1")]; + tensor reshape_2_shape_0 = const()[name = tensor("reshape_2_shape_0"), val = tensor([-1])]; + tensor reshape_2_cast_fp16 = reshape(shape = reshape_2_shape_0, x = var_435_cast_fp16)[name = tensor("reshape_2_cast_fp16")]; + tensor reshape_3_shape_0 = const()[name = tensor("reshape_3_shape_0"), val = tensor([-1])]; + tensor reshape_3_cast_fp16 = reshape(shape = reshape_3_shape_0, x = cache0_to_fp16)[name = tensor("reshape_3_cast_fp16")]; + tensor scatter_0_mode_0 = const()[name = tensor("scatter_0_mode_0"), val = tensor("update")]; + tensor scatter_0_axis_0 = const()[name = tensor("scatter_0_axis_0"), val = tensor(0)]; + tensor scatter_0_validate_indices_0 = const()[name = tensor("scatter_0_validate_indices_0"), val = tensor(false)]; + tensor scatter_0_cast_fp16 = scatter(axis = scatter_0_axis_0, data = reshape_3_cast_fp16, indices = reshape_1, mode = scatter_0_mode_0, updates = reshape_2_cast_fp16, validate_indices = scatter_0_validate_indices_0)[name = tensor("scatter_0_cast_fp16")]; + tensor reshape_4_cast_fp16 = reshape(shape = shape_48, x = scatter_0_cast_fp16)[name = tensor("reshape_4_cast_fp16")]; + tensor var_443_begin_0 = const()[name = tensor("op_443_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor var_443_end_0 = const()[name = tensor("op_443_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor var_443_end_mask_0 = const()[name = tensor("op_443_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_443_squeeze_mask_0 = const()[name = tensor("op_443_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor var_443_cast_fp16 = slice_by_index(begin = var_443_begin_0, end = var_443_end_0, end_mask = var_443_end_mask_0, squeeze_mask = var_443_squeeze_mask_0, x = reshape_4_cast_fp16)[name = tensor("op_443_cast_fp16")]; + tensor var_445_axis_0 = const()[name = tensor("op_445_axis_0"), val = tensor(1)]; + tensor var_445_mode_0 = const()[name = tensor("op_445_mode_0"), val = tensor("update")]; + tensor var_445_validate_indices_0 = const()[name = tensor("op_445_validate_indices_0"), val = tensor(false)]; + tensor var_445_cast_fp16 = scatter_along_axis(axis = var_445_axis_0, data = var_443_cast_fp16, indices = write_indices_1, mode = var_445_mode_0, updates = v_1_cast_fp16, validate_indices = var_445_validate_indices_0)[name = tensor("op_445_cast_fp16")]; + tensor concat_4 = const()[name = tensor("concat_4"), val = tensor([1, 0, 0, 0, 0])]; + tensor concat_5 = const()[name = tensor("concat_5"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_1_internal_tensor_assign_2_stride_0 = const()[name = tensor("new_cache_1_internal_tensor_assign_2_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_1_internal_tensor_assign_2_begin_mask_0 = const()[name = tensor("new_cache_1_internal_tensor_assign_2_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_1_internal_tensor_assign_2_end_mask_0 = const()[name = tensor("new_cache_1_internal_tensor_assign_2_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_1_internal_tensor_assign_2_squeeze_mask_0 = const()[name = tensor("new_cache_1_internal_tensor_assign_2_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_49 = const()[name = tensor("shape_49"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_1 = const()[name = tensor("reduce_prod_1"), val = tensor(1048576)]; + tensor range_1d_1_start_0 = const()[name = tensor("range_1d_1_start_0"), val = tensor(0)]; + tensor range_1d_1_step_0 = const()[name = tensor("range_1d_1_step_0"), val = tensor(1)]; + tensor range_1d_1 = range_1d(end = reduce_prod_1, start = range_1d_1_start_0, step = range_1d_1_step_0)[name = tensor("range_1d_1")]; + tensor reshape_5 = reshape(shape = shape_49, x = range_1d_1)[name = tensor("reshape_5")]; + tensor slice_by_index_1 = slice_by_index(begin = concat_4, begin_mask = new_cache_1_internal_tensor_assign_2_begin_mask_0, end = concat_5, end_mask = new_cache_1_internal_tensor_assign_2_end_mask_0, squeeze_mask = new_cache_1_internal_tensor_assign_2_squeeze_mask_0, stride = new_cache_1_internal_tensor_assign_2_stride_0, x = reshape_5)[name = tensor("slice_by_index_1")]; + tensor reshape_6_shape_0 = const()[name = tensor("reshape_6_shape_0"), val = tensor([-1])]; + tensor reshape_6 = reshape(shape = reshape_6_shape_0, x = slice_by_index_1)[name = tensor("reshape_6")]; + tensor reshape_7_shape_0 = const()[name = tensor("reshape_7_shape_0"), val = tensor([-1])]; + tensor reshape_7_cast_fp16 = reshape(shape = reshape_7_shape_0, x = var_445_cast_fp16)[name = tensor("reshape_7_cast_fp16")]; + tensor reshape_8_shape_0 = const()[name = tensor("reshape_8_shape_0"), val = tensor([-1])]; + tensor reshape_8_cast_fp16 = reshape(shape = reshape_8_shape_0, x = reshape_4_cast_fp16)[name = tensor("reshape_8_cast_fp16")]; + tensor scatter_1_mode_0 = const()[name = tensor("scatter_1_mode_0"), val = tensor("update")]; + tensor scatter_1_axis_0 = const()[name = tensor("scatter_1_axis_0"), val = tensor(0)]; + tensor scatter_1_validate_indices_0 = const()[name = tensor("scatter_1_validate_indices_0"), val = tensor(false)]; + tensor scatter_1_cast_fp16 = scatter(axis = scatter_1_axis_0, data = reshape_8_cast_fp16, indices = reshape_6, mode = scatter_1_mode_0, updates = reshape_7_cast_fp16, validate_indices = scatter_1_validate_indices_0)[name = tensor("scatter_1_cast_fp16")]; + tensor reshape_9_cast_fp16 = reshape(shape = shape_49, x = scatter_1_cast_fp16)[name = tensor("reshape_9_cast_fp16")]; + tensor reshape_9_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("reshape_9_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor keys_1_begin_0 = const()[name = tensor("keys_1_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor keys_1_end_0 = const()[name = tensor("keys_1_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor keys_1_end_mask_0 = const()[name = tensor("keys_1_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor keys_1_squeeze_mask_0 = const()[name = tensor("keys_1_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor keys_1_cast_fp16 = slice_by_index(begin = keys_1_begin_0, end = keys_1_end_0, end_mask = keys_1_end_mask_0, squeeze_mask = keys_1_squeeze_mask_0, x = reshape_9_cast_fp16)[name = tensor("keys_1_cast_fp16")]; + tensor values_1_begin_0 = const()[name = tensor("values_1_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor values_1_end_0 = const()[name = tensor("values_1_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor values_1_end_mask_0 = const()[name = tensor("values_1_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor values_1_squeeze_mask_0 = const()[name = tensor("values_1_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor values_1_cast_fp16 = slice_by_index(begin = values_1_begin_0, end = values_1_end_0, end_mask = values_1_end_mask_0, squeeze_mask = values_1_squeeze_mask_0, x = reshape_9_cast_fp16)[name = tensor("values_1_cast_fp16")]; + tensor var_457_cast_fp16 = not_equal(x = keys_1_cast_fp16, y = keys_1_cast_fp16)[name = tensor("op_457_cast_fp16")]; + tensor var_463_to_fp16 = const()[name = tensor("op_463_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(6369856)))]; + tensor keys_3_cast_fp16 = select(a = var_463_to_fp16, b = keys_1_cast_fp16, cond = var_457_cast_fp16)[name = tensor("keys_3_cast_fp16")]; + tensor var_465_cast_fp16 = not_equal(x = values_1_cast_fp16, y = values_1_cast_fp16)[name = tensor("op_465_cast_fp16")]; + tensor values_3_cast_fp16 = select(a = var_463_to_fp16, b = values_1_cast_fp16, cond = var_465_cast_fp16)[name = tensor("values_3_cast_fp16")]; + tensor var_489 = const()[name = tensor("op_489"), val = tensor([0, 2, 1, 3])]; + tensor var_505 = const()[name = tensor("op_505"), val = tensor([1, 1, 1])]; + tensor var_506_cast_fp16 = reshape(shape = var_505, x = position0_to_fp16)[name = tensor("op_506_cast_fp16")]; + tensor var_526_to_fp16 = const()[name = tensor("op_526_to_fp16"), val = tensor(0x1p+0)]; + tensor valid_len_1_cast_fp16 = add(x = var_506_cast_fp16, y = var_526_to_fp16)[name = tensor("valid_len_1_cast_fp16")]; + tensor k_positions_1_promoted_to_fp16 = const()[name = tensor("k_positions_1_promoted_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(7418496)))]; + tensor valid_mask_1_cast_fp16 = less(x = k_positions_1_promoted_to_fp16, y = valid_len_1_cast_fp16)[name = tensor("valid_mask_1_cast_fp16")]; + tensor causal_mask_1_cast_fp16 = less_equal(x = k_positions_1_promoted_to_fp16, y = var_506_cast_fp16)[name = tensor("causal_mask_1_cast_fp16")]; + tensor attn_mask_1 = logical_and(x = valid_mask_1_cast_fp16, y = causal_mask_1_cast_fp16)[name = tensor("attn_mask_1")]; + tensor attn_mask_3_axes_0 = const()[name = tensor("attn_mask_3_axes_0"), val = tensor([1])]; + tensor attn_mask_3 = expand_dims(axes = attn_mask_3_axes_0, x = attn_mask_1)[name = tensor("attn_mask_3")]; + tensor var_544_transpose_x_0 = const()[name = tensor("op_544_transpose_x_0"), val = tensor(false)]; + tensor var_544_transpose_y_0 = const()[name = tensor("op_544_transpose_y_0"), val = tensor(false)]; + tensor transpose_72_perm_0 = const()[name = tensor("transpose_72_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_73_perm_0 = const()[name = tensor("transpose_73_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_73 = transpose(perm = transpose_73_perm_0, x = keys_3_cast_fp16)[name = tensor("transpose_213")]; + tensor transpose_72 = transpose(perm = transpose_72_perm_0, x = q_3_cast_fp16)[name = tensor("transpose_214")]; + tensor var_544_cast_fp16 = matmul(transpose_x = var_544_transpose_x_0, transpose_y = var_544_transpose_y_0, x = transpose_72, y = transpose_73)[name = tensor("op_544_cast_fp16")]; + tensor var_538_to_fp16 = const()[name = tensor("op_538_to_fp16"), val = tensor([0x1p-3])]; + tensor attn_weights_1_cast_fp16 = mul(x = var_544_cast_fp16, y = var_538_to_fp16)[name = tensor("attn_weights_1_cast_fp16")]; + tensor var_546 = logical_not(x = attn_mask_3)[name = tensor("op_546")]; + tensor var_547_to_fp16 = const()[name = tensor("op_547_to_fp16"), val = tensor(-inf)]; + tensor attn_weights_3_cast_fp16 = select(a = var_547_to_fp16, b = attn_weights_1_cast_fp16, cond = var_546)[name = tensor("attn_weights_3_cast_fp16")]; + tensor var_549 = const()[name = tensor("op_549"), val = tensor(-1)]; + tensor attn_weights_5_cast_fp16 = softmax(axis = var_549, x = attn_weights_3_cast_fp16)[name = tensor("attn_weights_5_cast_fp16")]; + tensor attn_output_1_transpose_x_0 = const()[name = tensor("attn_output_1_transpose_x_0"), val = tensor(false)]; + tensor attn_output_1_transpose_y_0 = const()[name = tensor("attn_output_1_transpose_y_0"), val = tensor(false)]; + tensor values_5_cast_fp16 = transpose(perm = var_489, x = values_3_cast_fp16)[name = tensor("transpose_215")]; + tensor attn_output_1_cast_fp16 = matmul(transpose_x = attn_output_1_transpose_x_0, transpose_y = attn_output_1_transpose_y_0, x = attn_weights_5_cast_fp16, y = values_5_cast_fp16)[name = tensor("attn_output_1_cast_fp16")]; + tensor var_557 = const()[name = tensor("op_557"), val = tensor([0, 2, 1, 3])]; + tensor var_562 = const()[name = tensor("op_562"), val = tensor([1, 1, 1024])]; + tensor var_558_cast_fp16 = transpose(perm = var_557, x = attn_output_1_cast_fp16)[name = tensor("transpose_212")]; + tensor input_7_cast_fp16 = reshape(shape = var_562, x = var_558_cast_fp16)[name = tensor("input_7_cast_fp16")]; + tensor attn0_out_proj_weight_to_fp16 = const()[name = tensor("attn0_out_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(7419584)))]; + tensor linear_2_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = attn0_out_proj_weight_to_fp16, x = input_7_cast_fp16)[name = tensor("linear_2_cast_fp16")]; + tensor var_568_to_fp16 = const()[name = tensor("op_568_to_fp16"), val = tensor(0x1p+0)]; + tensor var_569_cast_fp16 = add(x = position0_to_fp16, y = var_568_to_fp16)[name = tensor("op_569_cast_fp16")]; + tensor var_569_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_569_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor input_9_cast_fp16 = add(x = linear_0_cast_fp16, y = linear_2_cast_fp16)[name = tensor("input_9_cast_fp16")]; + tensor input_11_axes_0 = const()[name = tensor("input_11_axes_0"), val = tensor([-1])]; + tensor norm0_2_weight_to_fp16 = const()[name = tensor("norm0_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9516800)))]; + tensor norm0_2_bias_to_fp16 = const()[name = tensor("norm0_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9518912)))]; + tensor var_573_to_fp16 = const()[name = tensor("op_573_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_11_cast_fp16 = layer_norm(axes = input_11_axes_0, beta = norm0_2_bias_to_fp16, epsilon = var_573_to_fp16, gamma = norm0_2_weight_to_fp16, x = input_9_cast_fp16)[name = tensor("input_11_cast_fp16")]; + tensor linear0_1_weight_to_fp16 = const()[name = tensor("linear0_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9521024)))]; + tensor linear_3_bias_0_to_fp16 = const()[name = tensor("linear_3_bias_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(17909696)))]; + tensor linear_3_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = linear0_1_weight_to_fp16, x = input_11_cast_fp16)[name = tensor("linear_3_cast_fp16")]; + tensor input_13_mode_0 = const()[name = tensor("input_13_mode_0"), val = tensor("EXACT")]; + tensor input_13_cast_fp16 = gelu(mode = input_13_mode_0, x = linear_3_cast_fp16)[name = tensor("input_13_cast_fp16")]; + tensor linear0_2_weight_to_fp16 = const()[name = tensor("linear0_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(17917952)))]; + tensor linear_4_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = linear0_2_weight_to_fp16, x = input_13_cast_fp16)[name = tensor("linear_4_cast_fp16")]; + tensor input_15_cast_fp16 = add(x = input_9_cast_fp16, y = linear_4_cast_fp16)[name = tensor("input_15_cast_fp16")]; + tensor input_17_axes_0 = const()[name = tensor("input_17_axes_0"), val = tensor([-1])]; + tensor norm1_1_weight_to_fp16 = const()[name = tensor("norm1_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(26306624)))]; + tensor norm1_1_bias_to_fp16 = const()[name = tensor("norm1_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(26308736)))]; + tensor var_590_to_fp16 = const()[name = tensor("op_590_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_17_cast_fp16 = layer_norm(axes = input_17_axes_0, beta = norm1_1_bias_to_fp16, epsilon = var_590_to_fp16, gamma = norm1_1_weight_to_fp16, x = input_15_cast_fp16)[name = tensor("input_17_cast_fp16")]; + tensor attn1_in_proj_weight_to_fp16 = const()[name = tensor("attn1_in_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(26310848)))]; + tensor linear_5_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = attn1_in_proj_weight_to_fp16, x = input_17_cast_fp16)[name = tensor("linear_5_cast_fp16")]; + tensor var_604 = const()[name = tensor("op_604"), val = tensor([1, 1, 3, 16, 64])]; + tensor qkv_3_cast_fp16 = reshape(shape = var_604, x = linear_5_cast_fp16)[name = tensor("qkv_3_cast_fp16")]; + tensor q_7_begin_0 = const()[name = tensor("q_7_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor q_7_end_0 = const()[name = tensor("q_7_end_0"), val = tensor([1, 1, 1, 16, 64])]; + tensor q_7_end_mask_0 = const()[name = tensor("q_7_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor q_7_squeeze_mask_0 = const()[name = tensor("q_7_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor q_7_cast_fp16 = slice_by_index(begin = q_7_begin_0, end = q_7_end_0, end_mask = q_7_end_mask_0, squeeze_mask = q_7_squeeze_mask_0, x = qkv_3_cast_fp16)[name = tensor("q_7_cast_fp16")]; + tensor k_5_begin_0 = const()[name = tensor("k_5_begin_0"), val = tensor([0, 0, 1, 0, 0])]; + tensor k_5_end_0 = const()[name = tensor("k_5_end_0"), val = tensor([1, 1, 2, 16, 64])]; + tensor k_5_end_mask_0 = const()[name = tensor("k_5_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor k_5_squeeze_mask_0 = const()[name = tensor("k_5_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor k_5_cast_fp16 = slice_by_index(begin = k_5_begin_0, end = k_5_end_0, end_mask = k_5_end_mask_0, squeeze_mask = k_5_squeeze_mask_0, x = qkv_3_cast_fp16)[name = tensor("k_5_cast_fp16")]; + tensor v_3_begin_0 = const()[name = tensor("v_3_begin_0"), val = tensor([0, 0, 2, 0, 0])]; + tensor v_3_end_0 = const()[name = tensor("v_3_end_0"), val = tensor([1, 1, 3, 16, 64])]; + tensor v_3_end_mask_0 = const()[name = tensor("v_3_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor v_3_squeeze_mask_0 = const()[name = tensor("v_3_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor v_3_cast_fp16 = slice_by_index(begin = v_3_begin_0, end = v_3_end_0, end_mask = v_3_end_mask_0, squeeze_mask = v_3_squeeze_mask_0, x = qkv_3_cast_fp16)[name = tensor("v_3_cast_fp16")]; + tensor var_670 = const()[name = tensor("op_670"), val = tensor([1, 1, 1, 1])]; + tensor position1_to_fp16_dtype_0 = const()[name = tensor("position1_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor position1_to_fp16 = cast(dtype = position1_to_fp16_dtype_0, x = position1)[name = tensor("cast_548")]; + tensor ts_11_cast_fp16 = reshape(shape = var_670, x = position1_to_fp16)[name = tensor("ts_11_cast_fp16")]; + tensor var_677 = const()[name = tensor("op_677"), val = tensor([1, 1, 16, 32, 2])]; + tensor q_complex_3_cast_fp16 = reshape(shape = var_677, x = q_7_cast_fp16)[name = tensor("q_complex_3_cast_fp16")]; + tensor var_684 = const()[name = tensor("op_684"), val = tensor([1, 1, 16, 32, 2])]; + tensor k_complex_3_cast_fp16 = reshape(shape = var_684, x = k_5_cast_fp16)[name = tensor("k_complex_3_cast_fp16")]; + tensor var_688_begin_0 = const()[name = tensor("op_688_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_688_end_0 = const()[name = tensor("op_688_end_0"), val = tensor([1, 1, 16, 32, 1])]; + tensor var_688_end_mask_0 = const()[name = tensor("op_688_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_688_squeeze_mask_0 = const()[name = tensor("op_688_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_688_cast_fp16 = slice_by_index(begin = var_688_begin_0, end = var_688_end_0, end_mask = var_688_end_mask_0, squeeze_mask = var_688_squeeze_mask_0, x = q_complex_3_cast_fp16)[name = tensor("op_688_cast_fp16")]; + tensor var_696_begin_0 = const()[name = tensor("op_696_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_696_end_0 = const()[name = tensor("op_696_end_0"), val = tensor([1, 1, 16, 32, 2])]; + tensor var_696_end_mask_0 = const()[name = tensor("op_696_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_696_squeeze_mask_0 = const()[name = tensor("op_696_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_696_cast_fp16 = slice_by_index(begin = var_696_begin_0, end = var_696_end_0, end_mask = var_696_end_mask_0, squeeze_mask = var_696_squeeze_mask_0, x = q_complex_3_cast_fp16)[name = tensor("op_696_cast_fp16")]; + tensor var_704_begin_0 = const()[name = tensor("op_704_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_704_end_0 = const()[name = tensor("op_704_end_0"), val = tensor([1, 1, 16, 32, 1])]; + tensor var_704_end_mask_0 = const()[name = tensor("op_704_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_704_squeeze_mask_0 = const()[name = tensor("op_704_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_704_cast_fp16 = slice_by_index(begin = var_704_begin_0, end = var_704_end_0, end_mask = var_704_end_mask_0, squeeze_mask = var_704_squeeze_mask_0, x = k_complex_3_cast_fp16)[name = tensor("op_704_cast_fp16")]; + tensor var_712_begin_0 = const()[name = tensor("op_712_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_712_end_0 = const()[name = tensor("op_712_end_0"), val = tensor([1, 1, 16, 32, 2])]; + tensor var_712_end_mask_0 = const()[name = tensor("op_712_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_712_squeeze_mask_0 = const()[name = tensor("op_712_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_712_cast_fp16 = slice_by_index(begin = var_712_begin_0, end = var_712_end_0, end_mask = var_712_end_mask_0, squeeze_mask = var_712_squeeze_mask_0, x = k_complex_3_cast_fp16)[name = tensor("op_712_cast_fp16")]; + tensor freqs_3_to_fp16 = const()[name = tensor("freqs_3_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(32602368)))]; + tensor var_718_cast_fp16 = mul(x = freqs_3_to_fp16, y = ts_11_cast_fp16)[name = tensor("op_718_cast_fp16")]; + tensor rotr_3_cast_fp16 = cos(x = var_718_cast_fp16)[name = tensor("rotr_3_cast_fp16")]; + tensor roti_3_cast_fp16 = sin(x = var_718_cast_fp16)[name = tensor("roti_3_cast_fp16")]; + tensor var_722_cast_fp16 = mul(x = var_688_cast_fp16, y = rotr_3_cast_fp16)[name = tensor("op_722_cast_fp16")]; + tensor var_723_cast_fp16 = mul(x = var_696_cast_fp16, y = roti_3_cast_fp16)[name = tensor("op_723_cast_fp16")]; + tensor qor_5_cast_fp16 = sub(x = var_722_cast_fp16, y = var_723_cast_fp16)[name = tensor("qor_5_cast_fp16")]; + tensor var_726_cast_fp16 = mul(x = var_688_cast_fp16, y = roti_3_cast_fp16)[name = tensor("op_726_cast_fp16")]; + tensor var_727_cast_fp16 = mul(x = var_696_cast_fp16, y = rotr_3_cast_fp16)[name = tensor("op_727_cast_fp16")]; + tensor qoi_5_cast_fp16 = add(x = var_726_cast_fp16, y = var_727_cast_fp16)[name = tensor("qoi_5_cast_fp16")]; + tensor var_730_cast_fp16 = mul(x = var_704_cast_fp16, y = rotr_3_cast_fp16)[name = tensor("op_730_cast_fp16")]; + tensor var_731_cast_fp16 = mul(x = var_712_cast_fp16, y = roti_3_cast_fp16)[name = tensor("op_731_cast_fp16")]; + tensor kor_5_cast_fp16 = sub(x = var_730_cast_fp16, y = var_731_cast_fp16)[name = tensor("kor_5_cast_fp16")]; + tensor var_734_cast_fp16 = mul(x = var_704_cast_fp16, y = roti_3_cast_fp16)[name = tensor("op_734_cast_fp16")]; + tensor var_735_cast_fp16 = mul(x = var_712_cast_fp16, y = rotr_3_cast_fp16)[name = tensor("op_735_cast_fp16")]; + tensor koi_5_cast_fp16 = add(x = var_734_cast_fp16, y = var_735_cast_fp16)[name = tensor("koi_5_cast_fp16")]; + tensor qo_3_axis_0 = const()[name = tensor("qo_3_axis_0"), val = tensor(-1)]; + tensor qo_3_cast_fp16 = stack(axis = qo_3_axis_0, values = (qor_5_cast_fp16, qoi_5_cast_fp16))[name = tensor("qo_3_cast_fp16")]; + tensor ko_3_axis_0 = const()[name = tensor("ko_3_axis_0"), val = tensor(-1)]; + tensor ko_3_cast_fp16 = stack(axis = ko_3_axis_0, values = (kor_5_cast_fp16, koi_5_cast_fp16))[name = tensor("ko_3_cast_fp16")]; + tensor var_768 = const()[name = tensor("op_768"), val = tensor([1, 1, 16, 64])]; + tensor q_9_cast_fp16 = reshape(shape = var_768, x = qo_3_cast_fp16)[name = tensor("q_9_cast_fp16")]; + tensor var_774 = const()[name = tensor("op_774"), val = tensor([1, 1, 16, 64])]; + tensor k_7_cast_fp16 = reshape(shape = var_774, x = ko_3_cast_fp16)[name = tensor("k_7_cast_fp16")]; + tensor _inversed_799_y_0_to_fp16 = const()[name = tensor("_inversed_799_y_0_to_fp16"), val = tensor(0x1p-9)]; + tensor _inversed_799_cast_fp16 = mul(x = ts_11_cast_fp16, y = _inversed_799_y_0_to_fp16)[name = tensor("_inversed_799_cast_fp16")]; + tensor var_800_cast_fp16 = floor(x = _inversed_799_cast_fp16)[name = tensor("op_800_cast_fp16")]; + tensor var_801_to_fp16 = const()[name = tensor("op_801_to_fp16"), val = tensor(0x1p+9)]; + tensor var_802_cast_fp16 = mul(x = var_800_cast_fp16, y = var_801_to_fp16)[name = tensor("op_802_cast_fp16")]; + tensor write_indices_float_7_cast_fp16 = sub(x = ts_11_cast_fp16, y = var_802_cast_fp16)[name = tensor("write_indices_float_7_cast_fp16")]; + tensor var_809_dtype_0 = const()[name = tensor("op_809_dtype_0"), val = tensor("int32")]; + tensor write_indices_3_reps_0 = const()[name = tensor("write_indices_3_reps_0"), val = tensor([1, 1, 16, 64])]; + tensor write_indices_float_7_cast_fp16_to_int32 = cast(dtype = var_809_dtype_0, x = write_indices_float_7_cast_fp16)[name = tensor("cast_547")]; + tensor write_indices_3 = tile(reps = write_indices_3_reps_0, x = write_indices_float_7_cast_fp16_to_int32)[name = tensor("write_indices_3")]; + tensor var_819_begin_0 = const()[name = tensor("op_819_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_819_end_0 = const()[name = tensor("op_819_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor var_819_end_mask_0 = const()[name = tensor("op_819_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_819_squeeze_mask_0 = const()[name = tensor("op_819_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor cache1_to_fp16_dtype_0 = const()[name = tensor("cache1_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor cache1_to_fp16 = cast(dtype = cache1_to_fp16_dtype_0, x = cache1)[name = tensor("cast_546")]; + tensor var_819_cast_fp16 = slice_by_index(begin = var_819_begin_0, end = var_819_end_0, end_mask = var_819_end_mask_0, squeeze_mask = var_819_squeeze_mask_0, x = cache1_to_fp16)[name = tensor("op_819_cast_fp16")]; + tensor var_821_axis_0 = const()[name = tensor("op_821_axis_0"), val = tensor(1)]; + tensor var_821_mode_0 = const()[name = tensor("op_821_mode_0"), val = tensor("update")]; + tensor var_821_validate_indices_0 = const()[name = tensor("op_821_validate_indices_0"), val = tensor(false)]; + tensor var_821_cast_fp16 = scatter_along_axis(axis = var_821_axis_0, data = var_819_cast_fp16, indices = write_indices_3, mode = var_821_mode_0, updates = k_7_cast_fp16, validate_indices = var_821_validate_indices_0)[name = tensor("op_821_cast_fp16")]; + tensor concat_9 = const()[name = tensor("concat_9"), val = tensor([0, 0, 0, 0, 0])]; + tensor concat_10 = const()[name = tensor("concat_10"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_3_internal_tensor_assign_1_stride_0 = const()[name = tensor("new_cache_3_internal_tensor_assign_1_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_3_internal_tensor_assign_1_begin_mask_0 = const()[name = tensor("new_cache_3_internal_tensor_assign_1_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_3_internal_tensor_assign_1_end_mask_0 = const()[name = tensor("new_cache_3_internal_tensor_assign_1_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_3_internal_tensor_assign_1_squeeze_mask_0 = const()[name = tensor("new_cache_3_internal_tensor_assign_1_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_50 = const()[name = tensor("shape_50"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_2 = const()[name = tensor("reduce_prod_2"), val = tensor(1048576)]; + tensor range_1d_2_start_0 = const()[name = tensor("range_1d_2_start_0"), val = tensor(0)]; + tensor range_1d_2_step_0 = const()[name = tensor("range_1d_2_step_0"), val = tensor(1)]; + tensor range_1d_2 = range_1d(end = reduce_prod_2, start = range_1d_2_start_0, step = range_1d_2_step_0)[name = tensor("range_1d_2")]; + tensor reshape_10 = reshape(shape = shape_50, x = range_1d_2)[name = tensor("reshape_10")]; + tensor slice_by_index_2 = slice_by_index(begin = concat_9, begin_mask = new_cache_3_internal_tensor_assign_1_begin_mask_0, end = concat_10, end_mask = new_cache_3_internal_tensor_assign_1_end_mask_0, squeeze_mask = new_cache_3_internal_tensor_assign_1_squeeze_mask_0, stride = new_cache_3_internal_tensor_assign_1_stride_0, x = reshape_10)[name = tensor("slice_by_index_2")]; + tensor reshape_11_shape_0 = const()[name = tensor("reshape_11_shape_0"), val = tensor([-1])]; + tensor reshape_11 = reshape(shape = reshape_11_shape_0, x = slice_by_index_2)[name = tensor("reshape_11")]; + tensor reshape_12_shape_0 = const()[name = tensor("reshape_12_shape_0"), val = tensor([-1])]; + tensor reshape_12_cast_fp16 = reshape(shape = reshape_12_shape_0, x = var_821_cast_fp16)[name = tensor("reshape_12_cast_fp16")]; + tensor reshape_13_shape_0 = const()[name = tensor("reshape_13_shape_0"), val = tensor([-1])]; + tensor reshape_13_cast_fp16 = reshape(shape = reshape_13_shape_0, x = cache1_to_fp16)[name = tensor("reshape_13_cast_fp16")]; + tensor scatter_2_mode_0 = const()[name = tensor("scatter_2_mode_0"), val = tensor("update")]; + tensor scatter_2_axis_0 = const()[name = tensor("scatter_2_axis_0"), val = tensor(0)]; + tensor scatter_2_validate_indices_0 = const()[name = tensor("scatter_2_validate_indices_0"), val = tensor(false)]; + tensor scatter_2_cast_fp16 = scatter(axis = scatter_2_axis_0, data = reshape_13_cast_fp16, indices = reshape_11, mode = scatter_2_mode_0, updates = reshape_12_cast_fp16, validate_indices = scatter_2_validate_indices_0)[name = tensor("scatter_2_cast_fp16")]; + tensor reshape_14_cast_fp16 = reshape(shape = shape_50, x = scatter_2_cast_fp16)[name = tensor("reshape_14_cast_fp16")]; + tensor var_829_begin_0 = const()[name = tensor("op_829_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor var_829_end_0 = const()[name = tensor("op_829_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor var_829_end_mask_0 = const()[name = tensor("op_829_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_829_squeeze_mask_0 = const()[name = tensor("op_829_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor var_829_cast_fp16 = slice_by_index(begin = var_829_begin_0, end = var_829_end_0, end_mask = var_829_end_mask_0, squeeze_mask = var_829_squeeze_mask_0, x = reshape_14_cast_fp16)[name = tensor("op_829_cast_fp16")]; + tensor var_831_axis_0 = const()[name = tensor("op_831_axis_0"), val = tensor(1)]; + tensor var_831_mode_0 = const()[name = tensor("op_831_mode_0"), val = tensor("update")]; + tensor var_831_validate_indices_0 = const()[name = tensor("op_831_validate_indices_0"), val = tensor(false)]; + tensor var_831_cast_fp16 = scatter_along_axis(axis = var_831_axis_0, data = var_829_cast_fp16, indices = write_indices_3, mode = var_831_mode_0, updates = v_3_cast_fp16, validate_indices = var_831_validate_indices_0)[name = tensor("op_831_cast_fp16")]; + tensor concat_11 = const()[name = tensor("concat_11"), val = tensor([1, 0, 0, 0, 0])]; + tensor concat_12 = const()[name = tensor("concat_12"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_3_internal_tensor_assign_2_stride_0 = const()[name = tensor("new_cache_3_internal_tensor_assign_2_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_3_internal_tensor_assign_2_begin_mask_0 = const()[name = tensor("new_cache_3_internal_tensor_assign_2_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_3_internal_tensor_assign_2_end_mask_0 = const()[name = tensor("new_cache_3_internal_tensor_assign_2_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_3_internal_tensor_assign_2_squeeze_mask_0 = const()[name = tensor("new_cache_3_internal_tensor_assign_2_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_51 = const()[name = tensor("shape_51"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_3 = const()[name = tensor("reduce_prod_3"), val = tensor(1048576)]; + tensor range_1d_3_start_0 = const()[name = tensor("range_1d_3_start_0"), val = tensor(0)]; + tensor range_1d_3_step_0 = const()[name = tensor("range_1d_3_step_0"), val = tensor(1)]; + tensor range_1d_3 = range_1d(end = reduce_prod_3, start = range_1d_3_start_0, step = range_1d_3_step_0)[name = tensor("range_1d_3")]; + tensor reshape_15 = reshape(shape = shape_51, x = range_1d_3)[name = tensor("reshape_15")]; + tensor slice_by_index_3 = slice_by_index(begin = concat_11, begin_mask = new_cache_3_internal_tensor_assign_2_begin_mask_0, end = concat_12, end_mask = new_cache_3_internal_tensor_assign_2_end_mask_0, squeeze_mask = new_cache_3_internal_tensor_assign_2_squeeze_mask_0, stride = new_cache_3_internal_tensor_assign_2_stride_0, x = reshape_15)[name = tensor("slice_by_index_3")]; + tensor reshape_16_shape_0 = const()[name = tensor("reshape_16_shape_0"), val = tensor([-1])]; + tensor reshape_16 = reshape(shape = reshape_16_shape_0, x = slice_by_index_3)[name = tensor("reshape_16")]; + tensor reshape_17_shape_0 = const()[name = tensor("reshape_17_shape_0"), val = tensor([-1])]; + tensor reshape_17_cast_fp16 = reshape(shape = reshape_17_shape_0, x = var_831_cast_fp16)[name = tensor("reshape_17_cast_fp16")]; + tensor reshape_18_shape_0 = const()[name = tensor("reshape_18_shape_0"), val = tensor([-1])]; + tensor reshape_18_cast_fp16 = reshape(shape = reshape_18_shape_0, x = reshape_14_cast_fp16)[name = tensor("reshape_18_cast_fp16")]; + tensor scatter_3_mode_0 = const()[name = tensor("scatter_3_mode_0"), val = tensor("update")]; + tensor scatter_3_axis_0 = const()[name = tensor("scatter_3_axis_0"), val = tensor(0)]; + tensor scatter_3_validate_indices_0 = const()[name = tensor("scatter_3_validate_indices_0"), val = tensor(false)]; + tensor scatter_3_cast_fp16 = scatter(axis = scatter_3_axis_0, data = reshape_18_cast_fp16, indices = reshape_16, mode = scatter_3_mode_0, updates = reshape_17_cast_fp16, validate_indices = scatter_3_validate_indices_0)[name = tensor("scatter_3_cast_fp16")]; + tensor reshape_19_cast_fp16 = reshape(shape = shape_51, x = scatter_3_cast_fp16)[name = tensor("reshape_19_cast_fp16")]; + tensor reshape_19_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("reshape_19_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor keys_7_begin_0 = const()[name = tensor("keys_7_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor keys_7_end_0 = const()[name = tensor("keys_7_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor keys_7_end_mask_0 = const()[name = tensor("keys_7_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor keys_7_squeeze_mask_0 = const()[name = tensor("keys_7_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor keys_7_cast_fp16 = slice_by_index(begin = keys_7_begin_0, end = keys_7_end_0, end_mask = keys_7_end_mask_0, squeeze_mask = keys_7_squeeze_mask_0, x = reshape_19_cast_fp16)[name = tensor("keys_7_cast_fp16")]; + tensor values_7_begin_0 = const()[name = tensor("values_7_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor values_7_end_0 = const()[name = tensor("values_7_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor values_7_end_mask_0 = const()[name = tensor("values_7_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor values_7_squeeze_mask_0 = const()[name = tensor("values_7_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor values_7_cast_fp16 = slice_by_index(begin = values_7_begin_0, end = values_7_end_0, end_mask = values_7_end_mask_0, squeeze_mask = values_7_squeeze_mask_0, x = reshape_19_cast_fp16)[name = tensor("values_7_cast_fp16")]; + tensor var_843_cast_fp16 = not_equal(x = keys_7_cast_fp16, y = keys_7_cast_fp16)[name = tensor("op_843_cast_fp16")]; + tensor keys_9_cast_fp16 = select(a = var_463_to_fp16, b = keys_7_cast_fp16, cond = var_843_cast_fp16)[name = tensor("keys_9_cast_fp16")]; + tensor var_851_cast_fp16 = not_equal(x = values_7_cast_fp16, y = values_7_cast_fp16)[name = tensor("op_851_cast_fp16")]; + tensor values_9_cast_fp16 = select(a = var_463_to_fp16, b = values_7_cast_fp16, cond = var_851_cast_fp16)[name = tensor("values_9_cast_fp16")]; + tensor var_875 = const()[name = tensor("op_875"), val = tensor([0, 2, 1, 3])]; + tensor var_891 = const()[name = tensor("op_891"), val = tensor([1, 1, 1])]; + tensor var_892_cast_fp16 = reshape(shape = var_891, x = position1_to_fp16)[name = tensor("op_892_cast_fp16")]; + tensor var_912_to_fp16 = const()[name = tensor("op_912_to_fp16"), val = tensor(0x1p+0)]; + tensor valid_len_3_cast_fp16 = add(x = var_892_cast_fp16, y = var_912_to_fp16)[name = tensor("valid_len_3_cast_fp16")]; + tensor valid_mask_3_cast_fp16 = less(x = k_positions_1_promoted_to_fp16, y = valid_len_3_cast_fp16)[name = tensor("valid_mask_3_cast_fp16")]; + tensor causal_mask_3_cast_fp16 = less_equal(x = k_positions_1_promoted_to_fp16, y = var_892_cast_fp16)[name = tensor("causal_mask_3_cast_fp16")]; + tensor attn_mask_5 = logical_and(x = valid_mask_3_cast_fp16, y = causal_mask_3_cast_fp16)[name = tensor("attn_mask_5")]; + tensor attn_mask_7_axes_0 = const()[name = tensor("attn_mask_7_axes_0"), val = tensor([1])]; + tensor attn_mask_7 = expand_dims(axes = attn_mask_7_axes_0, x = attn_mask_5)[name = tensor("attn_mask_7")]; + tensor var_930_transpose_x_0 = const()[name = tensor("op_930_transpose_x_0"), val = tensor(false)]; + tensor var_930_transpose_y_0 = const()[name = tensor("op_930_transpose_y_0"), val = tensor(false)]; + tensor transpose_74_perm_0 = const()[name = tensor("transpose_74_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_75_perm_0 = const()[name = tensor("transpose_75_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_75 = transpose(perm = transpose_75_perm_0, x = keys_9_cast_fp16)[name = tensor("transpose_209")]; + tensor transpose_74 = transpose(perm = transpose_74_perm_0, x = q_9_cast_fp16)[name = tensor("transpose_210")]; + tensor var_930_cast_fp16 = matmul(transpose_x = var_930_transpose_x_0, transpose_y = var_930_transpose_y_0, x = transpose_74, y = transpose_75)[name = tensor("op_930_cast_fp16")]; + tensor var_924_to_fp16 = const()[name = tensor("op_924_to_fp16"), val = tensor([0x1p-3])]; + tensor attn_weights_7_cast_fp16 = mul(x = var_930_cast_fp16, y = var_924_to_fp16)[name = tensor("attn_weights_7_cast_fp16")]; + tensor var_932 = logical_not(x = attn_mask_7)[name = tensor("op_932")]; + tensor var_933_to_fp16 = const()[name = tensor("op_933_to_fp16"), val = tensor(-inf)]; + tensor attn_weights_9_cast_fp16 = select(a = var_933_to_fp16, b = attn_weights_7_cast_fp16, cond = var_932)[name = tensor("attn_weights_9_cast_fp16")]; + tensor var_935 = const()[name = tensor("op_935"), val = tensor(-1)]; + tensor attn_weights_11_cast_fp16 = softmax(axis = var_935, x = attn_weights_9_cast_fp16)[name = tensor("attn_weights_11_cast_fp16")]; + tensor attn_output_3_transpose_x_0 = const()[name = tensor("attn_output_3_transpose_x_0"), val = tensor(false)]; + tensor attn_output_3_transpose_y_0 = const()[name = tensor("attn_output_3_transpose_y_0"), val = tensor(false)]; + tensor values_11_cast_fp16 = transpose(perm = var_875, x = values_9_cast_fp16)[name = tensor("transpose_211")]; + tensor attn_output_3_cast_fp16 = matmul(transpose_x = attn_output_3_transpose_x_0, transpose_y = attn_output_3_transpose_y_0, x = attn_weights_11_cast_fp16, y = values_11_cast_fp16)[name = tensor("attn_output_3_cast_fp16")]; + tensor var_943 = const()[name = tensor("op_943"), val = tensor([0, 2, 1, 3])]; + tensor var_948 = const()[name = tensor("op_948"), val = tensor([1, 1, 1024])]; + tensor var_944_cast_fp16 = transpose(perm = var_943, x = attn_output_3_cast_fp16)[name = tensor("transpose_208")]; + tensor input_19_cast_fp16 = reshape(shape = var_948, x = var_944_cast_fp16)[name = tensor("input_19_cast_fp16")]; + tensor attn1_out_proj_weight_to_fp16 = const()[name = tensor("attn1_out_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(32602496)))]; + tensor linear_6_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = attn1_out_proj_weight_to_fp16, x = input_19_cast_fp16)[name = tensor("linear_6_cast_fp16")]; + tensor var_954_to_fp16 = const()[name = tensor("op_954_to_fp16"), val = tensor(0x1p+0)]; + tensor var_955_cast_fp16 = add(x = position1_to_fp16, y = var_954_to_fp16)[name = tensor("op_955_cast_fp16")]; + tensor var_955_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_955_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor input_21_cast_fp16 = add(x = input_15_cast_fp16, y = linear_6_cast_fp16)[name = tensor("input_21_cast_fp16")]; + tensor input_23_axes_0 = const()[name = tensor("input_23_axes_0"), val = tensor([-1])]; + tensor norm1_2_weight_to_fp16 = const()[name = tensor("norm1_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(34699712)))]; + tensor norm1_2_bias_to_fp16 = const()[name = tensor("norm1_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(34701824)))]; + tensor var_959_to_fp16 = const()[name = tensor("op_959_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_23_cast_fp16 = layer_norm(axes = input_23_axes_0, beta = norm1_2_bias_to_fp16, epsilon = var_959_to_fp16, gamma = norm1_2_weight_to_fp16, x = input_21_cast_fp16)[name = tensor("input_23_cast_fp16")]; + tensor linear1_1_weight_to_fp16 = const()[name = tensor("linear1_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(34703936)))]; + tensor linear_7_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = linear1_1_weight_to_fp16, x = input_23_cast_fp16)[name = tensor("linear_7_cast_fp16")]; + tensor input_25_mode_0 = const()[name = tensor("input_25_mode_0"), val = tensor("EXACT")]; + tensor input_25_cast_fp16 = gelu(mode = input_25_mode_0, x = linear_7_cast_fp16)[name = tensor("input_25_cast_fp16")]; + tensor linear1_2_weight_to_fp16 = const()[name = tensor("linear1_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(43092608)))]; + tensor linear_8_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = linear1_2_weight_to_fp16, x = input_25_cast_fp16)[name = tensor("linear_8_cast_fp16")]; + tensor input_27_cast_fp16 = add(x = input_21_cast_fp16, y = linear_8_cast_fp16)[name = tensor("input_27_cast_fp16")]; + tensor input_29_axes_0 = const()[name = tensor("input_29_axes_0"), val = tensor([-1])]; + tensor norm2_1_weight_to_fp16 = const()[name = tensor("norm2_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(51481280)))]; + tensor norm2_1_bias_to_fp16 = const()[name = tensor("norm2_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(51483392)))]; + tensor var_976_to_fp16 = const()[name = tensor("op_976_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_29_cast_fp16 = layer_norm(axes = input_29_axes_0, beta = norm2_1_bias_to_fp16, epsilon = var_976_to_fp16, gamma = norm2_1_weight_to_fp16, x = input_27_cast_fp16)[name = tensor("input_29_cast_fp16")]; + tensor attn2_in_proj_weight_to_fp16 = const()[name = tensor("attn2_in_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(51485504)))]; + tensor linear_9_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = attn2_in_proj_weight_to_fp16, x = input_29_cast_fp16)[name = tensor("linear_9_cast_fp16")]; + tensor var_990 = const()[name = tensor("op_990"), val = tensor([1, 1, 3, 16, 64])]; + tensor qkv_5_cast_fp16 = reshape(shape = var_990, x = linear_9_cast_fp16)[name = tensor("qkv_5_cast_fp16")]; + tensor q_13_begin_0 = const()[name = tensor("q_13_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor q_13_end_0 = const()[name = tensor("q_13_end_0"), val = tensor([1, 1, 1, 16, 64])]; + tensor q_13_end_mask_0 = const()[name = tensor("q_13_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor q_13_squeeze_mask_0 = const()[name = tensor("q_13_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor q_13_cast_fp16 = slice_by_index(begin = q_13_begin_0, end = q_13_end_0, end_mask = q_13_end_mask_0, squeeze_mask = q_13_squeeze_mask_0, x = qkv_5_cast_fp16)[name = tensor("q_13_cast_fp16")]; + tensor k_9_begin_0 = const()[name = tensor("k_9_begin_0"), val = tensor([0, 0, 1, 0, 0])]; + tensor k_9_end_0 = const()[name = tensor("k_9_end_0"), val = tensor([1, 1, 2, 16, 64])]; + tensor k_9_end_mask_0 = const()[name = tensor("k_9_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor k_9_squeeze_mask_0 = const()[name = tensor("k_9_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor k_9_cast_fp16 = slice_by_index(begin = k_9_begin_0, end = k_9_end_0, end_mask = k_9_end_mask_0, squeeze_mask = k_9_squeeze_mask_0, x = qkv_5_cast_fp16)[name = tensor("k_9_cast_fp16")]; + tensor v_5_begin_0 = const()[name = tensor("v_5_begin_0"), val = tensor([0, 0, 2, 0, 0])]; + tensor v_5_end_0 = const()[name = tensor("v_5_end_0"), val = tensor([1, 1, 3, 16, 64])]; + tensor v_5_end_mask_0 = const()[name = tensor("v_5_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor v_5_squeeze_mask_0 = const()[name = tensor("v_5_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor v_5_cast_fp16 = slice_by_index(begin = v_5_begin_0, end = v_5_end_0, end_mask = v_5_end_mask_0, squeeze_mask = v_5_squeeze_mask_0, x = qkv_5_cast_fp16)[name = tensor("v_5_cast_fp16")]; + tensor var_1056 = const()[name = tensor("op_1056"), val = tensor([1, 1, 1, 1])]; + tensor position2_to_fp16_dtype_0 = const()[name = tensor("position2_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor position2_to_fp16 = cast(dtype = position2_to_fp16_dtype_0, x = position2)[name = tensor("cast_543")]; + tensor ts_17_cast_fp16 = reshape(shape = var_1056, x = position2_to_fp16)[name = tensor("ts_17_cast_fp16")]; + tensor var_1063 = const()[name = tensor("op_1063"), val = tensor([1, 1, 16, 32, 2])]; + tensor q_complex_5_cast_fp16 = reshape(shape = var_1063, x = q_13_cast_fp16)[name = tensor("q_complex_5_cast_fp16")]; + tensor var_1070 = const()[name = tensor("op_1070"), val = tensor([1, 1, 16, 32, 2])]; + tensor k_complex_5_cast_fp16 = reshape(shape = var_1070, x = k_9_cast_fp16)[name = tensor("k_complex_5_cast_fp16")]; + tensor var_1074_begin_0 = const()[name = tensor("op_1074_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_1074_end_0 = const()[name = tensor("op_1074_end_0"), val = tensor([1, 1, 16, 32, 1])]; + tensor var_1074_end_mask_0 = const()[name = tensor("op_1074_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_1074_squeeze_mask_0 = const()[name = tensor("op_1074_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_1074_cast_fp16 = slice_by_index(begin = var_1074_begin_0, end = var_1074_end_0, end_mask = var_1074_end_mask_0, squeeze_mask = var_1074_squeeze_mask_0, x = q_complex_5_cast_fp16)[name = tensor("op_1074_cast_fp16")]; + tensor var_1082_begin_0 = const()[name = tensor("op_1082_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_1082_end_0 = const()[name = tensor("op_1082_end_0"), val = tensor([1, 1, 16, 32, 2])]; + tensor var_1082_end_mask_0 = const()[name = tensor("op_1082_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_1082_squeeze_mask_0 = const()[name = tensor("op_1082_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_1082_cast_fp16 = slice_by_index(begin = var_1082_begin_0, end = var_1082_end_0, end_mask = var_1082_end_mask_0, squeeze_mask = var_1082_squeeze_mask_0, x = q_complex_5_cast_fp16)[name = tensor("op_1082_cast_fp16")]; + tensor var_1090_begin_0 = const()[name = tensor("op_1090_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_1090_end_0 = const()[name = tensor("op_1090_end_0"), val = tensor([1, 1, 16, 32, 1])]; + tensor var_1090_end_mask_0 = const()[name = tensor("op_1090_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_1090_squeeze_mask_0 = const()[name = tensor("op_1090_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_1090_cast_fp16 = slice_by_index(begin = var_1090_begin_0, end = var_1090_end_0, end_mask = var_1090_end_mask_0, squeeze_mask = var_1090_squeeze_mask_0, x = k_complex_5_cast_fp16)[name = tensor("op_1090_cast_fp16")]; + tensor var_1098_begin_0 = const()[name = tensor("op_1098_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_1098_end_0 = const()[name = tensor("op_1098_end_0"), val = tensor([1, 1, 16, 32, 2])]; + tensor var_1098_end_mask_0 = const()[name = tensor("op_1098_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_1098_squeeze_mask_0 = const()[name = tensor("op_1098_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_1098_cast_fp16 = slice_by_index(begin = var_1098_begin_0, end = var_1098_end_0, end_mask = var_1098_end_mask_0, squeeze_mask = var_1098_squeeze_mask_0, x = k_complex_5_cast_fp16)[name = tensor("op_1098_cast_fp16")]; + tensor freqs_5_to_fp16 = const()[name = tensor("freqs_5_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(57777024)))]; + tensor var_1104_cast_fp16 = mul(x = freqs_5_to_fp16, y = ts_17_cast_fp16)[name = tensor("op_1104_cast_fp16")]; + tensor rotr_5_cast_fp16 = cos(x = var_1104_cast_fp16)[name = tensor("rotr_5_cast_fp16")]; + tensor roti_5_cast_fp16 = sin(x = var_1104_cast_fp16)[name = tensor("roti_5_cast_fp16")]; + tensor var_1108_cast_fp16 = mul(x = var_1074_cast_fp16, y = rotr_5_cast_fp16)[name = tensor("op_1108_cast_fp16")]; + tensor var_1109_cast_fp16 = mul(x = var_1082_cast_fp16, y = roti_5_cast_fp16)[name = tensor("op_1109_cast_fp16")]; + tensor qor_9_cast_fp16 = sub(x = var_1108_cast_fp16, y = var_1109_cast_fp16)[name = tensor("qor_9_cast_fp16")]; + tensor var_1112_cast_fp16 = mul(x = var_1074_cast_fp16, y = roti_5_cast_fp16)[name = tensor("op_1112_cast_fp16")]; + tensor var_1113_cast_fp16 = mul(x = var_1082_cast_fp16, y = rotr_5_cast_fp16)[name = tensor("op_1113_cast_fp16")]; + tensor qoi_9_cast_fp16 = add(x = var_1112_cast_fp16, y = var_1113_cast_fp16)[name = tensor("qoi_9_cast_fp16")]; + tensor var_1116_cast_fp16 = mul(x = var_1090_cast_fp16, y = rotr_5_cast_fp16)[name = tensor("op_1116_cast_fp16")]; + tensor var_1117_cast_fp16 = mul(x = var_1098_cast_fp16, y = roti_5_cast_fp16)[name = tensor("op_1117_cast_fp16")]; + tensor kor_9_cast_fp16 = sub(x = var_1116_cast_fp16, y = var_1117_cast_fp16)[name = tensor("kor_9_cast_fp16")]; + tensor var_1120_cast_fp16 = mul(x = var_1090_cast_fp16, y = roti_5_cast_fp16)[name = tensor("op_1120_cast_fp16")]; + tensor var_1121_cast_fp16 = mul(x = var_1098_cast_fp16, y = rotr_5_cast_fp16)[name = tensor("op_1121_cast_fp16")]; + tensor koi_9_cast_fp16 = add(x = var_1120_cast_fp16, y = var_1121_cast_fp16)[name = tensor("koi_9_cast_fp16")]; + tensor qo_5_axis_0 = const()[name = tensor("qo_5_axis_0"), val = tensor(-1)]; + tensor qo_5_cast_fp16 = stack(axis = qo_5_axis_0, values = (qor_9_cast_fp16, qoi_9_cast_fp16))[name = tensor("qo_5_cast_fp16")]; + tensor ko_5_axis_0 = const()[name = tensor("ko_5_axis_0"), val = tensor(-1)]; + tensor ko_5_cast_fp16 = stack(axis = ko_5_axis_0, values = (kor_9_cast_fp16, koi_9_cast_fp16))[name = tensor("ko_5_cast_fp16")]; + tensor var_1154 = const()[name = tensor("op_1154"), val = tensor([1, 1, 16, 64])]; + tensor q_15_cast_fp16 = reshape(shape = var_1154, x = qo_5_cast_fp16)[name = tensor("q_15_cast_fp16")]; + tensor var_1160 = const()[name = tensor("op_1160"), val = tensor([1, 1, 16, 64])]; + tensor k_11_cast_fp16 = reshape(shape = var_1160, x = ko_5_cast_fp16)[name = tensor("k_11_cast_fp16")]; + tensor _inversed_1185_y_0_to_fp16 = const()[name = tensor("_inversed_1185_y_0_to_fp16"), val = tensor(0x1p-9)]; + tensor _inversed_1185_cast_fp16 = mul(x = ts_17_cast_fp16, y = _inversed_1185_y_0_to_fp16)[name = tensor("_inversed_1185_cast_fp16")]; + tensor var_1186_cast_fp16 = floor(x = _inversed_1185_cast_fp16)[name = tensor("op_1186_cast_fp16")]; + tensor var_1187_to_fp16 = const()[name = tensor("op_1187_to_fp16"), val = tensor(0x1p+9)]; + tensor var_1188_cast_fp16 = mul(x = var_1186_cast_fp16, y = var_1187_to_fp16)[name = tensor("op_1188_cast_fp16")]; + tensor write_indices_float_11_cast_fp16 = sub(x = ts_17_cast_fp16, y = var_1188_cast_fp16)[name = tensor("write_indices_float_11_cast_fp16")]; + tensor var_1195_dtype_0 = const()[name = tensor("op_1195_dtype_0"), val = tensor("int32")]; + tensor write_indices_5_reps_0 = const()[name = tensor("write_indices_5_reps_0"), val = tensor([1, 1, 16, 64])]; + tensor write_indices_float_11_cast_fp16_to_int32 = cast(dtype = var_1195_dtype_0, x = write_indices_float_11_cast_fp16)[name = tensor("cast_542")]; + tensor write_indices_5 = tile(reps = write_indices_5_reps_0, x = write_indices_float_11_cast_fp16_to_int32)[name = tensor("write_indices_5")]; + tensor var_1205_begin_0 = const()[name = tensor("op_1205_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_1205_end_0 = const()[name = tensor("op_1205_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor var_1205_end_mask_0 = const()[name = tensor("op_1205_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_1205_squeeze_mask_0 = const()[name = tensor("op_1205_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor cache2_to_fp16_dtype_0 = const()[name = tensor("cache2_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor cache2_to_fp16 = cast(dtype = cache2_to_fp16_dtype_0, x = cache2)[name = tensor("cast_541")]; + tensor var_1205_cast_fp16 = slice_by_index(begin = var_1205_begin_0, end = var_1205_end_0, end_mask = var_1205_end_mask_0, squeeze_mask = var_1205_squeeze_mask_0, x = cache2_to_fp16)[name = tensor("op_1205_cast_fp16")]; + tensor var_1207_axis_0 = const()[name = tensor("op_1207_axis_0"), val = tensor(1)]; + tensor var_1207_mode_0 = const()[name = tensor("op_1207_mode_0"), val = tensor("update")]; + tensor var_1207_validate_indices_0 = const()[name = tensor("op_1207_validate_indices_0"), val = tensor(false)]; + tensor var_1207_cast_fp16 = scatter_along_axis(axis = var_1207_axis_0, data = var_1205_cast_fp16, indices = write_indices_5, mode = var_1207_mode_0, updates = k_11_cast_fp16, validate_indices = var_1207_validate_indices_0)[name = tensor("op_1207_cast_fp16")]; + tensor concat_16 = const()[name = tensor("concat_16"), val = tensor([0, 0, 0, 0, 0])]; + tensor concat_17 = const()[name = tensor("concat_17"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_5_internal_tensor_assign_1_stride_0 = const()[name = tensor("new_cache_5_internal_tensor_assign_1_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_5_internal_tensor_assign_1_begin_mask_0 = const()[name = tensor("new_cache_5_internal_tensor_assign_1_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_5_internal_tensor_assign_1_end_mask_0 = const()[name = tensor("new_cache_5_internal_tensor_assign_1_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_5_internal_tensor_assign_1_squeeze_mask_0 = const()[name = tensor("new_cache_5_internal_tensor_assign_1_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_52 = const()[name = tensor("shape_52"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_4 = const()[name = tensor("reduce_prod_4"), val = tensor(1048576)]; + tensor range_1d_4_start_0 = const()[name = tensor("range_1d_4_start_0"), val = tensor(0)]; + tensor range_1d_4_step_0 = const()[name = tensor("range_1d_4_step_0"), val = tensor(1)]; + tensor range_1d_4 = range_1d(end = reduce_prod_4, start = range_1d_4_start_0, step = range_1d_4_step_0)[name = tensor("range_1d_4")]; + tensor reshape_20 = reshape(shape = shape_52, x = range_1d_4)[name = tensor("reshape_20")]; + tensor slice_by_index_4 = slice_by_index(begin = concat_16, begin_mask = new_cache_5_internal_tensor_assign_1_begin_mask_0, end = concat_17, end_mask = new_cache_5_internal_tensor_assign_1_end_mask_0, squeeze_mask = new_cache_5_internal_tensor_assign_1_squeeze_mask_0, stride = new_cache_5_internal_tensor_assign_1_stride_0, x = reshape_20)[name = tensor("slice_by_index_4")]; + tensor reshape_21_shape_0 = const()[name = tensor("reshape_21_shape_0"), val = tensor([-1])]; + tensor reshape_21 = reshape(shape = reshape_21_shape_0, x = slice_by_index_4)[name = tensor("reshape_21")]; + tensor reshape_22_shape_0 = const()[name = tensor("reshape_22_shape_0"), val = tensor([-1])]; + tensor reshape_22_cast_fp16 = reshape(shape = reshape_22_shape_0, x = var_1207_cast_fp16)[name = tensor("reshape_22_cast_fp16")]; + tensor reshape_23_shape_0 = const()[name = tensor("reshape_23_shape_0"), val = tensor([-1])]; + tensor reshape_23_cast_fp16 = reshape(shape = reshape_23_shape_0, x = cache2_to_fp16)[name = tensor("reshape_23_cast_fp16")]; + tensor scatter_4_mode_0 = const()[name = tensor("scatter_4_mode_0"), val = tensor("update")]; + tensor scatter_4_axis_0 = const()[name = tensor("scatter_4_axis_0"), val = tensor(0)]; + tensor scatter_4_validate_indices_0 = const()[name = tensor("scatter_4_validate_indices_0"), val = tensor(false)]; + tensor scatter_4_cast_fp16 = scatter(axis = scatter_4_axis_0, data = reshape_23_cast_fp16, indices = reshape_21, mode = scatter_4_mode_0, updates = reshape_22_cast_fp16, validate_indices = scatter_4_validate_indices_0)[name = tensor("scatter_4_cast_fp16")]; + tensor reshape_24_cast_fp16 = reshape(shape = shape_52, x = scatter_4_cast_fp16)[name = tensor("reshape_24_cast_fp16")]; + tensor var_1215_begin_0 = const()[name = tensor("op_1215_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor var_1215_end_0 = const()[name = tensor("op_1215_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor var_1215_end_mask_0 = const()[name = tensor("op_1215_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_1215_squeeze_mask_0 = const()[name = tensor("op_1215_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor var_1215_cast_fp16 = slice_by_index(begin = var_1215_begin_0, end = var_1215_end_0, end_mask = var_1215_end_mask_0, squeeze_mask = var_1215_squeeze_mask_0, x = reshape_24_cast_fp16)[name = tensor("op_1215_cast_fp16")]; + tensor var_1217_axis_0 = const()[name = tensor("op_1217_axis_0"), val = tensor(1)]; + tensor var_1217_mode_0 = const()[name = tensor("op_1217_mode_0"), val = tensor("update")]; + tensor var_1217_validate_indices_0 = const()[name = tensor("op_1217_validate_indices_0"), val = tensor(false)]; + tensor var_1217_cast_fp16 = scatter_along_axis(axis = var_1217_axis_0, data = var_1215_cast_fp16, indices = write_indices_5, mode = var_1217_mode_0, updates = v_5_cast_fp16, validate_indices = var_1217_validate_indices_0)[name = tensor("op_1217_cast_fp16")]; + tensor concat_18 = const()[name = tensor("concat_18"), val = tensor([1, 0, 0, 0, 0])]; + tensor concat_19 = const()[name = tensor("concat_19"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_5_internal_tensor_assign_2_stride_0 = const()[name = tensor("new_cache_5_internal_tensor_assign_2_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_5_internal_tensor_assign_2_begin_mask_0 = const()[name = tensor("new_cache_5_internal_tensor_assign_2_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_5_internal_tensor_assign_2_end_mask_0 = const()[name = tensor("new_cache_5_internal_tensor_assign_2_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_5_internal_tensor_assign_2_squeeze_mask_0 = const()[name = tensor("new_cache_5_internal_tensor_assign_2_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_53 = const()[name = tensor("shape_53"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_5 = const()[name = tensor("reduce_prod_5"), val = tensor(1048576)]; + tensor range_1d_5_start_0 = const()[name = tensor("range_1d_5_start_0"), val = tensor(0)]; + tensor range_1d_5_step_0 = const()[name = tensor("range_1d_5_step_0"), val = tensor(1)]; + tensor range_1d_5 = range_1d(end = reduce_prod_5, start = range_1d_5_start_0, step = range_1d_5_step_0)[name = tensor("range_1d_5")]; + tensor reshape_25 = reshape(shape = shape_53, x = range_1d_5)[name = tensor("reshape_25")]; + tensor slice_by_index_5 = slice_by_index(begin = concat_18, begin_mask = new_cache_5_internal_tensor_assign_2_begin_mask_0, end = concat_19, end_mask = new_cache_5_internal_tensor_assign_2_end_mask_0, squeeze_mask = new_cache_5_internal_tensor_assign_2_squeeze_mask_0, stride = new_cache_5_internal_tensor_assign_2_stride_0, x = reshape_25)[name = tensor("slice_by_index_5")]; + tensor reshape_26_shape_0 = const()[name = tensor("reshape_26_shape_0"), val = tensor([-1])]; + tensor reshape_26 = reshape(shape = reshape_26_shape_0, x = slice_by_index_5)[name = tensor("reshape_26")]; + tensor reshape_27_shape_0 = const()[name = tensor("reshape_27_shape_0"), val = tensor([-1])]; + tensor reshape_27_cast_fp16 = reshape(shape = reshape_27_shape_0, x = var_1217_cast_fp16)[name = tensor("reshape_27_cast_fp16")]; + tensor reshape_28_shape_0 = const()[name = tensor("reshape_28_shape_0"), val = tensor([-1])]; + tensor reshape_28_cast_fp16 = reshape(shape = reshape_28_shape_0, x = reshape_24_cast_fp16)[name = tensor("reshape_28_cast_fp16")]; + tensor scatter_5_mode_0 = const()[name = tensor("scatter_5_mode_0"), val = tensor("update")]; + tensor scatter_5_axis_0 = const()[name = tensor("scatter_5_axis_0"), val = tensor(0)]; + tensor scatter_5_validate_indices_0 = const()[name = tensor("scatter_5_validate_indices_0"), val = tensor(false)]; + tensor scatter_5_cast_fp16 = scatter(axis = scatter_5_axis_0, data = reshape_28_cast_fp16, indices = reshape_26, mode = scatter_5_mode_0, updates = reshape_27_cast_fp16, validate_indices = scatter_5_validate_indices_0)[name = tensor("scatter_5_cast_fp16")]; + tensor reshape_29_cast_fp16 = reshape(shape = shape_53, x = scatter_5_cast_fp16)[name = tensor("reshape_29_cast_fp16")]; + tensor reshape_29_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("reshape_29_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor keys_13_begin_0 = const()[name = tensor("keys_13_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor keys_13_end_0 = const()[name = tensor("keys_13_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor keys_13_end_mask_0 = const()[name = tensor("keys_13_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor keys_13_squeeze_mask_0 = const()[name = tensor("keys_13_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor keys_13_cast_fp16 = slice_by_index(begin = keys_13_begin_0, end = keys_13_end_0, end_mask = keys_13_end_mask_0, squeeze_mask = keys_13_squeeze_mask_0, x = reshape_29_cast_fp16)[name = tensor("keys_13_cast_fp16")]; + tensor values_13_begin_0 = const()[name = tensor("values_13_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor values_13_end_0 = const()[name = tensor("values_13_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor values_13_end_mask_0 = const()[name = tensor("values_13_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor values_13_squeeze_mask_0 = const()[name = tensor("values_13_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor values_13_cast_fp16 = slice_by_index(begin = values_13_begin_0, end = values_13_end_0, end_mask = values_13_end_mask_0, squeeze_mask = values_13_squeeze_mask_0, x = reshape_29_cast_fp16)[name = tensor("values_13_cast_fp16")]; + tensor var_1229_cast_fp16 = not_equal(x = keys_13_cast_fp16, y = keys_13_cast_fp16)[name = tensor("op_1229_cast_fp16")]; + tensor keys_15_cast_fp16 = select(a = var_463_to_fp16, b = keys_13_cast_fp16, cond = var_1229_cast_fp16)[name = tensor("keys_15_cast_fp16")]; + tensor var_1237_cast_fp16 = not_equal(x = values_13_cast_fp16, y = values_13_cast_fp16)[name = tensor("op_1237_cast_fp16")]; + tensor values_15_cast_fp16 = select(a = var_463_to_fp16, b = values_13_cast_fp16, cond = var_1237_cast_fp16)[name = tensor("values_15_cast_fp16")]; + tensor var_1261 = const()[name = tensor("op_1261"), val = tensor([0, 2, 1, 3])]; + tensor var_1277 = const()[name = tensor("op_1277"), val = tensor([1, 1, 1])]; + tensor var_1278_cast_fp16 = reshape(shape = var_1277, x = position2_to_fp16)[name = tensor("op_1278_cast_fp16")]; + tensor var_1298_to_fp16 = const()[name = tensor("op_1298_to_fp16"), val = tensor(0x1p+0)]; + tensor valid_len_5_cast_fp16 = add(x = var_1278_cast_fp16, y = var_1298_to_fp16)[name = tensor("valid_len_5_cast_fp16")]; + tensor valid_mask_5_cast_fp16 = less(x = k_positions_1_promoted_to_fp16, y = valid_len_5_cast_fp16)[name = tensor("valid_mask_5_cast_fp16")]; + tensor causal_mask_5_cast_fp16 = less_equal(x = k_positions_1_promoted_to_fp16, y = var_1278_cast_fp16)[name = tensor("causal_mask_5_cast_fp16")]; + tensor attn_mask_9 = logical_and(x = valid_mask_5_cast_fp16, y = causal_mask_5_cast_fp16)[name = tensor("attn_mask_9")]; + tensor attn_mask_11_axes_0 = const()[name = tensor("attn_mask_11_axes_0"), val = tensor([1])]; + tensor attn_mask_11 = expand_dims(axes = attn_mask_11_axes_0, x = attn_mask_9)[name = tensor("attn_mask_11")]; + tensor var_1316_transpose_x_0 = const()[name = tensor("op_1316_transpose_x_0"), val = tensor(false)]; + tensor var_1316_transpose_y_0 = const()[name = tensor("op_1316_transpose_y_0"), val = tensor(false)]; + tensor transpose_76_perm_0 = const()[name = tensor("transpose_76_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_77_perm_0 = const()[name = tensor("transpose_77_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_77 = transpose(perm = transpose_77_perm_0, x = keys_15_cast_fp16)[name = tensor("transpose_205")]; + tensor transpose_76 = transpose(perm = transpose_76_perm_0, x = q_15_cast_fp16)[name = tensor("transpose_206")]; + tensor var_1316_cast_fp16 = matmul(transpose_x = var_1316_transpose_x_0, transpose_y = var_1316_transpose_y_0, x = transpose_76, y = transpose_77)[name = tensor("op_1316_cast_fp16")]; + tensor var_1310_to_fp16 = const()[name = tensor("op_1310_to_fp16"), val = tensor([0x1p-3])]; + tensor attn_weights_13_cast_fp16 = mul(x = var_1316_cast_fp16, y = var_1310_to_fp16)[name = tensor("attn_weights_13_cast_fp16")]; + tensor var_1318 = logical_not(x = attn_mask_11)[name = tensor("op_1318")]; + tensor var_1319_to_fp16 = const()[name = tensor("op_1319_to_fp16"), val = tensor(-inf)]; + tensor attn_weights_15_cast_fp16 = select(a = var_1319_to_fp16, b = attn_weights_13_cast_fp16, cond = var_1318)[name = tensor("attn_weights_15_cast_fp16")]; + tensor var_1321 = const()[name = tensor("op_1321"), val = tensor(-1)]; + tensor attn_weights_17_cast_fp16 = softmax(axis = var_1321, x = attn_weights_15_cast_fp16)[name = tensor("attn_weights_17_cast_fp16")]; + tensor attn_output_5_transpose_x_0 = const()[name = tensor("attn_output_5_transpose_x_0"), val = tensor(false)]; + tensor attn_output_5_transpose_y_0 = const()[name = tensor("attn_output_5_transpose_y_0"), val = tensor(false)]; + tensor values_17_cast_fp16 = transpose(perm = var_1261, x = values_15_cast_fp16)[name = tensor("transpose_207")]; + tensor attn_output_5_cast_fp16 = matmul(transpose_x = attn_output_5_transpose_x_0, transpose_y = attn_output_5_transpose_y_0, x = attn_weights_17_cast_fp16, y = values_17_cast_fp16)[name = tensor("attn_output_5_cast_fp16")]; + tensor var_1329 = const()[name = tensor("op_1329"), val = tensor([0, 2, 1, 3])]; + tensor var_1334 = const()[name = tensor("op_1334"), val = tensor([1, 1, 1024])]; + tensor var_1330_cast_fp16 = transpose(perm = var_1329, x = attn_output_5_cast_fp16)[name = tensor("transpose_204")]; + tensor input_31_cast_fp16 = reshape(shape = var_1334, x = var_1330_cast_fp16)[name = tensor("input_31_cast_fp16")]; + tensor attn2_out_proj_weight_to_fp16 = const()[name = tensor("attn2_out_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(57777152)))]; + tensor linear_10_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = attn2_out_proj_weight_to_fp16, x = input_31_cast_fp16)[name = tensor("linear_10_cast_fp16")]; + tensor var_1340_to_fp16 = const()[name = tensor("op_1340_to_fp16"), val = tensor(0x1p+0)]; + tensor var_1341_cast_fp16 = add(x = position2_to_fp16, y = var_1340_to_fp16)[name = tensor("op_1341_cast_fp16")]; + tensor var_1341_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_1341_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor input_33_cast_fp16 = add(x = input_27_cast_fp16, y = linear_10_cast_fp16)[name = tensor("input_33_cast_fp16")]; + tensor input_35_axes_0 = const()[name = tensor("input_35_axes_0"), val = tensor([-1])]; + tensor norm2_2_weight_to_fp16 = const()[name = tensor("norm2_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(59874368)))]; + tensor norm2_2_bias_to_fp16 = const()[name = tensor("norm2_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(59876480)))]; + tensor var_1345_to_fp16 = const()[name = tensor("op_1345_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_35_cast_fp16 = layer_norm(axes = input_35_axes_0, beta = norm2_2_bias_to_fp16, epsilon = var_1345_to_fp16, gamma = norm2_2_weight_to_fp16, x = input_33_cast_fp16)[name = tensor("input_35_cast_fp16")]; + tensor linear2_1_weight_to_fp16 = const()[name = tensor("linear2_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(59878592)))]; + tensor linear_11_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = linear2_1_weight_to_fp16, x = input_35_cast_fp16)[name = tensor("linear_11_cast_fp16")]; + tensor input_37_mode_0 = const()[name = tensor("input_37_mode_0"), val = tensor("EXACT")]; + tensor input_37_cast_fp16 = gelu(mode = input_37_mode_0, x = linear_11_cast_fp16)[name = tensor("input_37_cast_fp16")]; + tensor linear2_2_weight_to_fp16 = const()[name = tensor("linear2_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(68267264)))]; + tensor linear_12_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = linear2_2_weight_to_fp16, x = input_37_cast_fp16)[name = tensor("linear_12_cast_fp16")]; + tensor input_39_cast_fp16 = add(x = input_33_cast_fp16, y = linear_12_cast_fp16)[name = tensor("input_39_cast_fp16")]; + tensor input_41_axes_0 = const()[name = tensor("input_41_axes_0"), val = tensor([-1])]; + tensor norm3_1_weight_to_fp16 = const()[name = tensor("norm3_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(76655936)))]; + tensor norm3_1_bias_to_fp16 = const()[name = tensor("norm3_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(76658048)))]; + tensor var_1362_to_fp16 = const()[name = tensor("op_1362_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_41_cast_fp16 = layer_norm(axes = input_41_axes_0, beta = norm3_1_bias_to_fp16, epsilon = var_1362_to_fp16, gamma = norm3_1_weight_to_fp16, x = input_39_cast_fp16)[name = tensor("input_41_cast_fp16")]; + tensor attn3_in_proj_weight_to_fp16 = const()[name = tensor("attn3_in_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(76660160)))]; + tensor linear_13_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = attn3_in_proj_weight_to_fp16, x = input_41_cast_fp16)[name = tensor("linear_13_cast_fp16")]; + tensor var_1376 = const()[name = tensor("op_1376"), val = tensor([1, 1, 3, 16, 64])]; + tensor qkv_7_cast_fp16 = reshape(shape = var_1376, x = linear_13_cast_fp16)[name = tensor("qkv_7_cast_fp16")]; + tensor q_19_begin_0 = const()[name = tensor("q_19_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor q_19_end_0 = const()[name = tensor("q_19_end_0"), val = tensor([1, 1, 1, 16, 64])]; + tensor q_19_end_mask_0 = const()[name = tensor("q_19_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor q_19_squeeze_mask_0 = const()[name = tensor("q_19_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor q_19_cast_fp16 = slice_by_index(begin = q_19_begin_0, end = q_19_end_0, end_mask = q_19_end_mask_0, squeeze_mask = q_19_squeeze_mask_0, x = qkv_7_cast_fp16)[name = tensor("q_19_cast_fp16")]; + tensor k_13_begin_0 = const()[name = tensor("k_13_begin_0"), val = tensor([0, 0, 1, 0, 0])]; + tensor k_13_end_0 = const()[name = tensor("k_13_end_0"), val = tensor([1, 1, 2, 16, 64])]; + tensor k_13_end_mask_0 = const()[name = tensor("k_13_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor k_13_squeeze_mask_0 = const()[name = tensor("k_13_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor k_13_cast_fp16 = slice_by_index(begin = k_13_begin_0, end = k_13_end_0, end_mask = k_13_end_mask_0, squeeze_mask = k_13_squeeze_mask_0, x = qkv_7_cast_fp16)[name = tensor("k_13_cast_fp16")]; + tensor v_7_begin_0 = const()[name = tensor("v_7_begin_0"), val = tensor([0, 0, 2, 0, 0])]; + tensor v_7_end_0 = const()[name = tensor("v_7_end_0"), val = tensor([1, 1, 3, 16, 64])]; + tensor v_7_end_mask_0 = const()[name = tensor("v_7_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor v_7_squeeze_mask_0 = const()[name = tensor("v_7_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor v_7_cast_fp16 = slice_by_index(begin = v_7_begin_0, end = v_7_end_0, end_mask = v_7_end_mask_0, squeeze_mask = v_7_squeeze_mask_0, x = qkv_7_cast_fp16)[name = tensor("v_7_cast_fp16")]; + tensor var_1442 = const()[name = tensor("op_1442"), val = tensor([1, 1, 1, 1])]; + tensor position3_to_fp16_dtype_0 = const()[name = tensor("position3_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor position3_to_fp16 = cast(dtype = position3_to_fp16_dtype_0, x = position3)[name = tensor("cast_538")]; + tensor ts_23_cast_fp16 = reshape(shape = var_1442, x = position3_to_fp16)[name = tensor("ts_23_cast_fp16")]; + tensor var_1449 = const()[name = tensor("op_1449"), val = tensor([1, 1, 16, 32, 2])]; + tensor q_complex_7_cast_fp16 = reshape(shape = var_1449, x = q_19_cast_fp16)[name = tensor("q_complex_7_cast_fp16")]; + tensor var_1456 = const()[name = tensor("op_1456"), val = tensor([1, 1, 16, 32, 2])]; + tensor k_complex_7_cast_fp16 = reshape(shape = var_1456, x = k_13_cast_fp16)[name = tensor("k_complex_7_cast_fp16")]; + tensor var_1460_begin_0 = const()[name = tensor("op_1460_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_1460_end_0 = const()[name = tensor("op_1460_end_0"), val = tensor([1, 1, 16, 32, 1])]; + tensor var_1460_end_mask_0 = const()[name = tensor("op_1460_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_1460_squeeze_mask_0 = const()[name = tensor("op_1460_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_1460_cast_fp16 = slice_by_index(begin = var_1460_begin_0, end = var_1460_end_0, end_mask = var_1460_end_mask_0, squeeze_mask = var_1460_squeeze_mask_0, x = q_complex_7_cast_fp16)[name = tensor("op_1460_cast_fp16")]; + tensor var_1468_begin_0 = const()[name = tensor("op_1468_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_1468_end_0 = const()[name = tensor("op_1468_end_0"), val = tensor([1, 1, 16, 32, 2])]; + tensor var_1468_end_mask_0 = const()[name = tensor("op_1468_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_1468_squeeze_mask_0 = const()[name = tensor("op_1468_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_1468_cast_fp16 = slice_by_index(begin = var_1468_begin_0, end = var_1468_end_0, end_mask = var_1468_end_mask_0, squeeze_mask = var_1468_squeeze_mask_0, x = q_complex_7_cast_fp16)[name = tensor("op_1468_cast_fp16")]; + tensor var_1476_begin_0 = const()[name = tensor("op_1476_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_1476_end_0 = const()[name = tensor("op_1476_end_0"), val = tensor([1, 1, 16, 32, 1])]; + tensor var_1476_end_mask_0 = const()[name = tensor("op_1476_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_1476_squeeze_mask_0 = const()[name = tensor("op_1476_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_1476_cast_fp16 = slice_by_index(begin = var_1476_begin_0, end = var_1476_end_0, end_mask = var_1476_end_mask_0, squeeze_mask = var_1476_squeeze_mask_0, x = k_complex_7_cast_fp16)[name = tensor("op_1476_cast_fp16")]; + tensor var_1484_begin_0 = const()[name = tensor("op_1484_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_1484_end_0 = const()[name = tensor("op_1484_end_0"), val = tensor([1, 1, 16, 32, 2])]; + tensor var_1484_end_mask_0 = const()[name = tensor("op_1484_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_1484_squeeze_mask_0 = const()[name = tensor("op_1484_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_1484_cast_fp16 = slice_by_index(begin = var_1484_begin_0, end = var_1484_end_0, end_mask = var_1484_end_mask_0, squeeze_mask = var_1484_squeeze_mask_0, x = k_complex_7_cast_fp16)[name = tensor("op_1484_cast_fp16")]; + tensor freqs_7_to_fp16 = const()[name = tensor("freqs_7_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(82951680)))]; + tensor var_1490_cast_fp16 = mul(x = freqs_7_to_fp16, y = ts_23_cast_fp16)[name = tensor("op_1490_cast_fp16")]; + tensor rotr_7_cast_fp16 = cos(x = var_1490_cast_fp16)[name = tensor("rotr_7_cast_fp16")]; + tensor roti_7_cast_fp16 = sin(x = var_1490_cast_fp16)[name = tensor("roti_7_cast_fp16")]; + tensor var_1494_cast_fp16 = mul(x = var_1460_cast_fp16, y = rotr_7_cast_fp16)[name = tensor("op_1494_cast_fp16")]; + tensor var_1495_cast_fp16 = mul(x = var_1468_cast_fp16, y = roti_7_cast_fp16)[name = tensor("op_1495_cast_fp16")]; + tensor qor_13_cast_fp16 = sub(x = var_1494_cast_fp16, y = var_1495_cast_fp16)[name = tensor("qor_13_cast_fp16")]; + tensor var_1498_cast_fp16 = mul(x = var_1460_cast_fp16, y = roti_7_cast_fp16)[name = tensor("op_1498_cast_fp16")]; + tensor var_1499_cast_fp16 = mul(x = var_1468_cast_fp16, y = rotr_7_cast_fp16)[name = tensor("op_1499_cast_fp16")]; + tensor qoi_13_cast_fp16 = add(x = var_1498_cast_fp16, y = var_1499_cast_fp16)[name = tensor("qoi_13_cast_fp16")]; + tensor var_1502_cast_fp16 = mul(x = var_1476_cast_fp16, y = rotr_7_cast_fp16)[name = tensor("op_1502_cast_fp16")]; + tensor var_1503_cast_fp16 = mul(x = var_1484_cast_fp16, y = roti_7_cast_fp16)[name = tensor("op_1503_cast_fp16")]; + tensor kor_13_cast_fp16 = sub(x = var_1502_cast_fp16, y = var_1503_cast_fp16)[name = tensor("kor_13_cast_fp16")]; + tensor var_1506_cast_fp16 = mul(x = var_1476_cast_fp16, y = roti_7_cast_fp16)[name = tensor("op_1506_cast_fp16")]; + tensor var_1507_cast_fp16 = mul(x = var_1484_cast_fp16, y = rotr_7_cast_fp16)[name = tensor("op_1507_cast_fp16")]; + tensor koi_13_cast_fp16 = add(x = var_1506_cast_fp16, y = var_1507_cast_fp16)[name = tensor("koi_13_cast_fp16")]; + tensor qo_7_axis_0 = const()[name = tensor("qo_7_axis_0"), val = tensor(-1)]; + tensor qo_7_cast_fp16 = stack(axis = qo_7_axis_0, values = (qor_13_cast_fp16, qoi_13_cast_fp16))[name = tensor("qo_7_cast_fp16")]; + tensor ko_7_axis_0 = const()[name = tensor("ko_7_axis_0"), val = tensor(-1)]; + tensor ko_7_cast_fp16 = stack(axis = ko_7_axis_0, values = (kor_13_cast_fp16, koi_13_cast_fp16))[name = tensor("ko_7_cast_fp16")]; + tensor var_1540 = const()[name = tensor("op_1540"), val = tensor([1, 1, 16, 64])]; + tensor q_21_cast_fp16 = reshape(shape = var_1540, x = qo_7_cast_fp16)[name = tensor("q_21_cast_fp16")]; + tensor var_1546 = const()[name = tensor("op_1546"), val = tensor([1, 1, 16, 64])]; + tensor k_15_cast_fp16 = reshape(shape = var_1546, x = ko_7_cast_fp16)[name = tensor("k_15_cast_fp16")]; + tensor _inversed_1571_y_0_to_fp16 = const()[name = tensor("_inversed_1571_y_0_to_fp16"), val = tensor(0x1p-9)]; + tensor _inversed_1571_cast_fp16 = mul(x = ts_23_cast_fp16, y = _inversed_1571_y_0_to_fp16)[name = tensor("_inversed_1571_cast_fp16")]; + tensor var_1572_cast_fp16 = floor(x = _inversed_1571_cast_fp16)[name = tensor("op_1572_cast_fp16")]; + tensor var_1573_to_fp16 = const()[name = tensor("op_1573_to_fp16"), val = tensor(0x1p+9)]; + tensor var_1574_cast_fp16 = mul(x = var_1572_cast_fp16, y = var_1573_to_fp16)[name = tensor("op_1574_cast_fp16")]; + tensor write_indices_float_15_cast_fp16 = sub(x = ts_23_cast_fp16, y = var_1574_cast_fp16)[name = tensor("write_indices_float_15_cast_fp16")]; + tensor var_1581_dtype_0 = const()[name = tensor("op_1581_dtype_0"), val = tensor("int32")]; + tensor write_indices_7_reps_0 = const()[name = tensor("write_indices_7_reps_0"), val = tensor([1, 1, 16, 64])]; + tensor write_indices_float_15_cast_fp16_to_int32 = cast(dtype = var_1581_dtype_0, x = write_indices_float_15_cast_fp16)[name = tensor("cast_537")]; + tensor write_indices_7 = tile(reps = write_indices_7_reps_0, x = write_indices_float_15_cast_fp16_to_int32)[name = tensor("write_indices_7")]; + tensor var_1591_begin_0 = const()[name = tensor("op_1591_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_1591_end_0 = const()[name = tensor("op_1591_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor var_1591_end_mask_0 = const()[name = tensor("op_1591_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_1591_squeeze_mask_0 = const()[name = tensor("op_1591_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor cache3_to_fp16_dtype_0 = const()[name = tensor("cache3_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor cache3_to_fp16 = cast(dtype = cache3_to_fp16_dtype_0, x = cache3)[name = tensor("cast_536")]; + tensor var_1591_cast_fp16 = slice_by_index(begin = var_1591_begin_0, end = var_1591_end_0, end_mask = var_1591_end_mask_0, squeeze_mask = var_1591_squeeze_mask_0, x = cache3_to_fp16)[name = tensor("op_1591_cast_fp16")]; + tensor var_1593_axis_0 = const()[name = tensor("op_1593_axis_0"), val = tensor(1)]; + tensor var_1593_mode_0 = const()[name = tensor("op_1593_mode_0"), val = tensor("update")]; + tensor var_1593_validate_indices_0 = const()[name = tensor("op_1593_validate_indices_0"), val = tensor(false)]; + tensor var_1593_cast_fp16 = scatter_along_axis(axis = var_1593_axis_0, data = var_1591_cast_fp16, indices = write_indices_7, mode = var_1593_mode_0, updates = k_15_cast_fp16, validate_indices = var_1593_validate_indices_0)[name = tensor("op_1593_cast_fp16")]; + tensor concat_23 = const()[name = tensor("concat_23"), val = tensor([0, 0, 0, 0, 0])]; + tensor concat_24 = const()[name = tensor("concat_24"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_7_internal_tensor_assign_1_stride_0 = const()[name = tensor("new_cache_7_internal_tensor_assign_1_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_7_internal_tensor_assign_1_begin_mask_0 = const()[name = tensor("new_cache_7_internal_tensor_assign_1_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_7_internal_tensor_assign_1_end_mask_0 = const()[name = tensor("new_cache_7_internal_tensor_assign_1_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_7_internal_tensor_assign_1_squeeze_mask_0 = const()[name = tensor("new_cache_7_internal_tensor_assign_1_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_54 = const()[name = tensor("shape_54"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_6 = const()[name = tensor("reduce_prod_6"), val = tensor(1048576)]; + tensor range_1d_6_start_0 = const()[name = tensor("range_1d_6_start_0"), val = tensor(0)]; + tensor range_1d_6_step_0 = const()[name = tensor("range_1d_6_step_0"), val = tensor(1)]; + tensor range_1d_6 = range_1d(end = reduce_prod_6, start = range_1d_6_start_0, step = range_1d_6_step_0)[name = tensor("range_1d_6")]; + tensor reshape_30 = reshape(shape = shape_54, x = range_1d_6)[name = tensor("reshape_30")]; + tensor slice_by_index_6 = slice_by_index(begin = concat_23, begin_mask = new_cache_7_internal_tensor_assign_1_begin_mask_0, end = concat_24, end_mask = new_cache_7_internal_tensor_assign_1_end_mask_0, squeeze_mask = new_cache_7_internal_tensor_assign_1_squeeze_mask_0, stride = new_cache_7_internal_tensor_assign_1_stride_0, x = reshape_30)[name = tensor("slice_by_index_6")]; + tensor reshape_31_shape_0 = const()[name = tensor("reshape_31_shape_0"), val = tensor([-1])]; + tensor reshape_31 = reshape(shape = reshape_31_shape_0, x = slice_by_index_6)[name = tensor("reshape_31")]; + tensor reshape_32_shape_0 = const()[name = tensor("reshape_32_shape_0"), val = tensor([-1])]; + tensor reshape_32_cast_fp16 = reshape(shape = reshape_32_shape_0, x = var_1593_cast_fp16)[name = tensor("reshape_32_cast_fp16")]; + tensor reshape_33_shape_0 = const()[name = tensor("reshape_33_shape_0"), val = tensor([-1])]; + tensor reshape_33_cast_fp16 = reshape(shape = reshape_33_shape_0, x = cache3_to_fp16)[name = tensor("reshape_33_cast_fp16")]; + tensor scatter_6_mode_0 = const()[name = tensor("scatter_6_mode_0"), val = tensor("update")]; + tensor scatter_6_axis_0 = const()[name = tensor("scatter_6_axis_0"), val = tensor(0)]; + tensor scatter_6_validate_indices_0 = const()[name = tensor("scatter_6_validate_indices_0"), val = tensor(false)]; + tensor scatter_6_cast_fp16 = scatter(axis = scatter_6_axis_0, data = reshape_33_cast_fp16, indices = reshape_31, mode = scatter_6_mode_0, updates = reshape_32_cast_fp16, validate_indices = scatter_6_validate_indices_0)[name = tensor("scatter_6_cast_fp16")]; + tensor reshape_34_cast_fp16 = reshape(shape = shape_54, x = scatter_6_cast_fp16)[name = tensor("reshape_34_cast_fp16")]; + tensor var_1601_begin_0 = const()[name = tensor("op_1601_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor var_1601_end_0 = const()[name = tensor("op_1601_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor var_1601_end_mask_0 = const()[name = tensor("op_1601_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_1601_squeeze_mask_0 = const()[name = tensor("op_1601_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor var_1601_cast_fp16 = slice_by_index(begin = var_1601_begin_0, end = var_1601_end_0, end_mask = var_1601_end_mask_0, squeeze_mask = var_1601_squeeze_mask_0, x = reshape_34_cast_fp16)[name = tensor("op_1601_cast_fp16")]; + tensor var_1603_axis_0 = const()[name = tensor("op_1603_axis_0"), val = tensor(1)]; + tensor var_1603_mode_0 = const()[name = tensor("op_1603_mode_0"), val = tensor("update")]; + tensor var_1603_validate_indices_0 = const()[name = tensor("op_1603_validate_indices_0"), val = tensor(false)]; + tensor var_1603_cast_fp16 = scatter_along_axis(axis = var_1603_axis_0, data = var_1601_cast_fp16, indices = write_indices_7, mode = var_1603_mode_0, updates = v_7_cast_fp16, validate_indices = var_1603_validate_indices_0)[name = tensor("op_1603_cast_fp16")]; + tensor concat_25 = const()[name = tensor("concat_25"), val = tensor([1, 0, 0, 0, 0])]; + tensor concat_26 = const()[name = tensor("concat_26"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_7_internal_tensor_assign_2_stride_0 = const()[name = tensor("new_cache_7_internal_tensor_assign_2_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_7_internal_tensor_assign_2_begin_mask_0 = const()[name = tensor("new_cache_7_internal_tensor_assign_2_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_7_internal_tensor_assign_2_end_mask_0 = const()[name = tensor("new_cache_7_internal_tensor_assign_2_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_7_internal_tensor_assign_2_squeeze_mask_0 = const()[name = tensor("new_cache_7_internal_tensor_assign_2_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_55 = const()[name = tensor("shape_55"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_7 = const()[name = tensor("reduce_prod_7"), val = tensor(1048576)]; + tensor range_1d_7_start_0 = const()[name = tensor("range_1d_7_start_0"), val = tensor(0)]; + tensor range_1d_7_step_0 = const()[name = tensor("range_1d_7_step_0"), val = tensor(1)]; + tensor range_1d_7 = range_1d(end = reduce_prod_7, start = range_1d_7_start_0, step = range_1d_7_step_0)[name = tensor("range_1d_7")]; + tensor reshape_35 = reshape(shape = shape_55, x = range_1d_7)[name = tensor("reshape_35")]; + tensor slice_by_index_7 = slice_by_index(begin = concat_25, begin_mask = new_cache_7_internal_tensor_assign_2_begin_mask_0, end = concat_26, end_mask = new_cache_7_internal_tensor_assign_2_end_mask_0, squeeze_mask = new_cache_7_internal_tensor_assign_2_squeeze_mask_0, stride = new_cache_7_internal_tensor_assign_2_stride_0, x = reshape_35)[name = tensor("slice_by_index_7")]; + tensor reshape_36_shape_0 = const()[name = tensor("reshape_36_shape_0"), val = tensor([-1])]; + tensor reshape_36 = reshape(shape = reshape_36_shape_0, x = slice_by_index_7)[name = tensor("reshape_36")]; + tensor reshape_37_shape_0 = const()[name = tensor("reshape_37_shape_0"), val = tensor([-1])]; + tensor reshape_37_cast_fp16 = reshape(shape = reshape_37_shape_0, x = var_1603_cast_fp16)[name = tensor("reshape_37_cast_fp16")]; + tensor reshape_38_shape_0 = const()[name = tensor("reshape_38_shape_0"), val = tensor([-1])]; + tensor reshape_38_cast_fp16 = reshape(shape = reshape_38_shape_0, x = reshape_34_cast_fp16)[name = tensor("reshape_38_cast_fp16")]; + tensor scatter_7_mode_0 = const()[name = tensor("scatter_7_mode_0"), val = tensor("update")]; + tensor scatter_7_axis_0 = const()[name = tensor("scatter_7_axis_0"), val = tensor(0)]; + tensor scatter_7_validate_indices_0 = const()[name = tensor("scatter_7_validate_indices_0"), val = tensor(false)]; + tensor scatter_7_cast_fp16 = scatter(axis = scatter_7_axis_0, data = reshape_38_cast_fp16, indices = reshape_36, mode = scatter_7_mode_0, updates = reshape_37_cast_fp16, validate_indices = scatter_7_validate_indices_0)[name = tensor("scatter_7_cast_fp16")]; + tensor reshape_39_cast_fp16 = reshape(shape = shape_55, x = scatter_7_cast_fp16)[name = tensor("reshape_39_cast_fp16")]; + tensor reshape_39_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("reshape_39_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor keys_19_begin_0 = const()[name = tensor("keys_19_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor keys_19_end_0 = const()[name = tensor("keys_19_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor keys_19_end_mask_0 = const()[name = tensor("keys_19_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor keys_19_squeeze_mask_0 = const()[name = tensor("keys_19_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor keys_19_cast_fp16 = slice_by_index(begin = keys_19_begin_0, end = keys_19_end_0, end_mask = keys_19_end_mask_0, squeeze_mask = keys_19_squeeze_mask_0, x = reshape_39_cast_fp16)[name = tensor("keys_19_cast_fp16")]; + tensor values_19_begin_0 = const()[name = tensor("values_19_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor values_19_end_0 = const()[name = tensor("values_19_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor values_19_end_mask_0 = const()[name = tensor("values_19_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor values_19_squeeze_mask_0 = const()[name = tensor("values_19_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor values_19_cast_fp16 = slice_by_index(begin = values_19_begin_0, end = values_19_end_0, end_mask = values_19_end_mask_0, squeeze_mask = values_19_squeeze_mask_0, x = reshape_39_cast_fp16)[name = tensor("values_19_cast_fp16")]; + tensor var_1615_cast_fp16 = not_equal(x = keys_19_cast_fp16, y = keys_19_cast_fp16)[name = tensor("op_1615_cast_fp16")]; + tensor keys_21_cast_fp16 = select(a = var_463_to_fp16, b = keys_19_cast_fp16, cond = var_1615_cast_fp16)[name = tensor("keys_21_cast_fp16")]; + tensor var_1623_cast_fp16 = not_equal(x = values_19_cast_fp16, y = values_19_cast_fp16)[name = tensor("op_1623_cast_fp16")]; + tensor values_21_cast_fp16 = select(a = var_463_to_fp16, b = values_19_cast_fp16, cond = var_1623_cast_fp16)[name = tensor("values_21_cast_fp16")]; + tensor var_1647 = const()[name = tensor("op_1647"), val = tensor([0, 2, 1, 3])]; + tensor var_1663 = const()[name = tensor("op_1663"), val = tensor([1, 1, 1])]; + tensor var_1664_cast_fp16 = reshape(shape = var_1663, x = position3_to_fp16)[name = tensor("op_1664_cast_fp16")]; + tensor var_1684_to_fp16 = const()[name = tensor("op_1684_to_fp16"), val = tensor(0x1p+0)]; + tensor valid_len_7_cast_fp16 = add(x = var_1664_cast_fp16, y = var_1684_to_fp16)[name = tensor("valid_len_7_cast_fp16")]; + tensor valid_mask_7_cast_fp16 = less(x = k_positions_1_promoted_to_fp16, y = valid_len_7_cast_fp16)[name = tensor("valid_mask_7_cast_fp16")]; + tensor causal_mask_7_cast_fp16 = less_equal(x = k_positions_1_promoted_to_fp16, y = var_1664_cast_fp16)[name = tensor("causal_mask_7_cast_fp16")]; + tensor attn_mask_13 = logical_and(x = valid_mask_7_cast_fp16, y = causal_mask_7_cast_fp16)[name = tensor("attn_mask_13")]; + tensor attn_mask_15_axes_0 = const()[name = tensor("attn_mask_15_axes_0"), val = tensor([1])]; + tensor attn_mask_15 = expand_dims(axes = attn_mask_15_axes_0, x = attn_mask_13)[name = tensor("attn_mask_15")]; + tensor var_1702_transpose_x_0 = const()[name = tensor("op_1702_transpose_x_0"), val = tensor(false)]; + tensor var_1702_transpose_y_0 = const()[name = tensor("op_1702_transpose_y_0"), val = tensor(false)]; + tensor transpose_78_perm_0 = const()[name = tensor("transpose_78_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_79_perm_0 = const()[name = tensor("transpose_79_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_79 = transpose(perm = transpose_79_perm_0, x = keys_21_cast_fp16)[name = tensor("transpose_201")]; + tensor transpose_78 = transpose(perm = transpose_78_perm_0, x = q_21_cast_fp16)[name = tensor("transpose_202")]; + tensor var_1702_cast_fp16 = matmul(transpose_x = var_1702_transpose_x_0, transpose_y = var_1702_transpose_y_0, x = transpose_78, y = transpose_79)[name = tensor("op_1702_cast_fp16")]; + tensor var_1696_to_fp16 = const()[name = tensor("op_1696_to_fp16"), val = tensor([0x1p-3])]; + tensor attn_weights_19_cast_fp16 = mul(x = var_1702_cast_fp16, y = var_1696_to_fp16)[name = tensor("attn_weights_19_cast_fp16")]; + tensor var_1704 = logical_not(x = attn_mask_15)[name = tensor("op_1704")]; + tensor var_1705_to_fp16 = const()[name = tensor("op_1705_to_fp16"), val = tensor(-inf)]; + tensor attn_weights_21_cast_fp16 = select(a = var_1705_to_fp16, b = attn_weights_19_cast_fp16, cond = var_1704)[name = tensor("attn_weights_21_cast_fp16")]; + tensor var_1707 = const()[name = tensor("op_1707"), val = tensor(-1)]; + tensor attn_weights_23_cast_fp16 = softmax(axis = var_1707, x = attn_weights_21_cast_fp16)[name = tensor("attn_weights_23_cast_fp16")]; + tensor attn_output_7_transpose_x_0 = const()[name = tensor("attn_output_7_transpose_x_0"), val = tensor(false)]; + tensor attn_output_7_transpose_y_0 = const()[name = tensor("attn_output_7_transpose_y_0"), val = tensor(false)]; + tensor values_23_cast_fp16 = transpose(perm = var_1647, x = values_21_cast_fp16)[name = tensor("transpose_203")]; + tensor attn_output_7_cast_fp16 = matmul(transpose_x = attn_output_7_transpose_x_0, transpose_y = attn_output_7_transpose_y_0, x = attn_weights_23_cast_fp16, y = values_23_cast_fp16)[name = tensor("attn_output_7_cast_fp16")]; + tensor var_1715 = const()[name = tensor("op_1715"), val = tensor([0, 2, 1, 3])]; + tensor var_1720 = const()[name = tensor("op_1720"), val = tensor([1, 1, 1024])]; + tensor var_1716_cast_fp16 = transpose(perm = var_1715, x = attn_output_7_cast_fp16)[name = tensor("transpose_200")]; + tensor input_43_cast_fp16 = reshape(shape = var_1720, x = var_1716_cast_fp16)[name = tensor("input_43_cast_fp16")]; + tensor attn3_out_proj_weight_to_fp16 = const()[name = tensor("attn3_out_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(82951808)))]; + tensor linear_14_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = attn3_out_proj_weight_to_fp16, x = input_43_cast_fp16)[name = tensor("linear_14_cast_fp16")]; + tensor var_1726_to_fp16 = const()[name = tensor("op_1726_to_fp16"), val = tensor(0x1p+0)]; + tensor var_1727_cast_fp16 = add(x = position3_to_fp16, y = var_1726_to_fp16)[name = tensor("op_1727_cast_fp16")]; + tensor var_1727_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_1727_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor input_45_cast_fp16 = add(x = input_39_cast_fp16, y = linear_14_cast_fp16)[name = tensor("input_45_cast_fp16")]; + tensor input_47_axes_0 = const()[name = tensor("input_47_axes_0"), val = tensor([-1])]; + tensor norm3_2_weight_to_fp16 = const()[name = tensor("norm3_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(85049024)))]; + tensor norm3_2_bias_to_fp16 = const()[name = tensor("norm3_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(85051136)))]; + tensor var_1731_to_fp16 = const()[name = tensor("op_1731_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_47_cast_fp16 = layer_norm(axes = input_47_axes_0, beta = norm3_2_bias_to_fp16, epsilon = var_1731_to_fp16, gamma = norm3_2_weight_to_fp16, x = input_45_cast_fp16)[name = tensor("input_47_cast_fp16")]; + tensor linear3_1_weight_to_fp16 = const()[name = tensor("linear3_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(85053248)))]; + tensor linear_15_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = linear3_1_weight_to_fp16, x = input_47_cast_fp16)[name = tensor("linear_15_cast_fp16")]; + tensor input_49_mode_0 = const()[name = tensor("input_49_mode_0"), val = tensor("EXACT")]; + tensor input_49_cast_fp16 = gelu(mode = input_49_mode_0, x = linear_15_cast_fp16)[name = tensor("input_49_cast_fp16")]; + tensor linear3_2_weight_to_fp16 = const()[name = tensor("linear3_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(93441920)))]; + tensor linear_16_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = linear3_2_weight_to_fp16, x = input_49_cast_fp16)[name = tensor("linear_16_cast_fp16")]; + tensor input_51_cast_fp16 = add(x = input_45_cast_fp16, y = linear_16_cast_fp16)[name = tensor("input_51_cast_fp16")]; + tensor input_53_axes_0 = const()[name = tensor("input_53_axes_0"), val = tensor([-1])]; + tensor norm4_1_weight_to_fp16 = const()[name = tensor("norm4_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(101830592)))]; + tensor norm4_1_bias_to_fp16 = const()[name = tensor("norm4_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(101832704)))]; + tensor var_1748_to_fp16 = const()[name = tensor("op_1748_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_53_cast_fp16 = layer_norm(axes = input_53_axes_0, beta = norm4_1_bias_to_fp16, epsilon = var_1748_to_fp16, gamma = norm4_1_weight_to_fp16, x = input_51_cast_fp16)[name = tensor("input_53_cast_fp16")]; + tensor attn4_in_proj_weight_to_fp16 = const()[name = tensor("attn4_in_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(101834816)))]; + tensor linear_17_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = attn4_in_proj_weight_to_fp16, x = input_53_cast_fp16)[name = tensor("linear_17_cast_fp16")]; + tensor var_1762 = const()[name = tensor("op_1762"), val = tensor([1, 1, 3, 16, 64])]; + tensor qkv_9_cast_fp16 = reshape(shape = var_1762, x = linear_17_cast_fp16)[name = tensor("qkv_9_cast_fp16")]; + tensor q_25_begin_0 = const()[name = tensor("q_25_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor q_25_end_0 = const()[name = tensor("q_25_end_0"), val = tensor([1, 1, 1, 16, 64])]; + tensor q_25_end_mask_0 = const()[name = tensor("q_25_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor q_25_squeeze_mask_0 = const()[name = tensor("q_25_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor q_25_cast_fp16 = slice_by_index(begin = q_25_begin_0, end = q_25_end_0, end_mask = q_25_end_mask_0, squeeze_mask = q_25_squeeze_mask_0, x = qkv_9_cast_fp16)[name = tensor("q_25_cast_fp16")]; + tensor k_17_begin_0 = const()[name = tensor("k_17_begin_0"), val = tensor([0, 0, 1, 0, 0])]; + tensor k_17_end_0 = const()[name = tensor("k_17_end_0"), val = tensor([1, 1, 2, 16, 64])]; + tensor k_17_end_mask_0 = const()[name = tensor("k_17_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor k_17_squeeze_mask_0 = const()[name = tensor("k_17_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor k_17_cast_fp16 = slice_by_index(begin = k_17_begin_0, end = k_17_end_0, end_mask = k_17_end_mask_0, squeeze_mask = k_17_squeeze_mask_0, x = qkv_9_cast_fp16)[name = tensor("k_17_cast_fp16")]; + tensor v_9_begin_0 = const()[name = tensor("v_9_begin_0"), val = tensor([0, 0, 2, 0, 0])]; + tensor v_9_end_0 = const()[name = tensor("v_9_end_0"), val = tensor([1, 1, 3, 16, 64])]; + tensor v_9_end_mask_0 = const()[name = tensor("v_9_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor v_9_squeeze_mask_0 = const()[name = tensor("v_9_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor v_9_cast_fp16 = slice_by_index(begin = v_9_begin_0, end = v_9_end_0, end_mask = v_9_end_mask_0, squeeze_mask = v_9_squeeze_mask_0, x = qkv_9_cast_fp16)[name = tensor("v_9_cast_fp16")]; + tensor var_1828 = const()[name = tensor("op_1828"), val = tensor([1, 1, 1, 1])]; + tensor position4_to_fp16_dtype_0 = const()[name = tensor("position4_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor position4_to_fp16 = cast(dtype = position4_to_fp16_dtype_0, x = position4)[name = tensor("cast_533")]; + tensor ts_29_cast_fp16 = reshape(shape = var_1828, x = position4_to_fp16)[name = tensor("ts_29_cast_fp16")]; + tensor var_1835 = const()[name = tensor("op_1835"), val = tensor([1, 1, 16, 32, 2])]; + tensor q_complex_9_cast_fp16 = reshape(shape = var_1835, x = q_25_cast_fp16)[name = tensor("q_complex_9_cast_fp16")]; + tensor var_1842 = const()[name = tensor("op_1842"), val = tensor([1, 1, 16, 32, 2])]; + tensor k_complex_9_cast_fp16 = reshape(shape = var_1842, x = k_17_cast_fp16)[name = tensor("k_complex_9_cast_fp16")]; + tensor var_1846_begin_0 = const()[name = tensor("op_1846_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_1846_end_0 = const()[name = tensor("op_1846_end_0"), val = tensor([1, 1, 16, 32, 1])]; + tensor var_1846_end_mask_0 = const()[name = tensor("op_1846_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_1846_squeeze_mask_0 = const()[name = tensor("op_1846_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_1846_cast_fp16 = slice_by_index(begin = var_1846_begin_0, end = var_1846_end_0, end_mask = var_1846_end_mask_0, squeeze_mask = var_1846_squeeze_mask_0, x = q_complex_9_cast_fp16)[name = tensor("op_1846_cast_fp16")]; + tensor var_1854_begin_0 = const()[name = tensor("op_1854_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_1854_end_0 = const()[name = tensor("op_1854_end_0"), val = tensor([1, 1, 16, 32, 2])]; + tensor var_1854_end_mask_0 = const()[name = tensor("op_1854_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_1854_squeeze_mask_0 = const()[name = tensor("op_1854_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_1854_cast_fp16 = slice_by_index(begin = var_1854_begin_0, end = var_1854_end_0, end_mask = var_1854_end_mask_0, squeeze_mask = var_1854_squeeze_mask_0, x = q_complex_9_cast_fp16)[name = tensor("op_1854_cast_fp16")]; + tensor var_1862_begin_0 = const()[name = tensor("op_1862_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_1862_end_0 = const()[name = tensor("op_1862_end_0"), val = tensor([1, 1, 16, 32, 1])]; + tensor var_1862_end_mask_0 = const()[name = tensor("op_1862_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_1862_squeeze_mask_0 = const()[name = tensor("op_1862_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_1862_cast_fp16 = slice_by_index(begin = var_1862_begin_0, end = var_1862_end_0, end_mask = var_1862_end_mask_0, squeeze_mask = var_1862_squeeze_mask_0, x = k_complex_9_cast_fp16)[name = tensor("op_1862_cast_fp16")]; + tensor var_1870_begin_0 = const()[name = tensor("op_1870_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_1870_end_0 = const()[name = tensor("op_1870_end_0"), val = tensor([1, 1, 16, 32, 2])]; + tensor var_1870_end_mask_0 = const()[name = tensor("op_1870_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_1870_squeeze_mask_0 = const()[name = tensor("op_1870_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_1870_cast_fp16 = slice_by_index(begin = var_1870_begin_0, end = var_1870_end_0, end_mask = var_1870_end_mask_0, squeeze_mask = var_1870_squeeze_mask_0, x = k_complex_9_cast_fp16)[name = tensor("op_1870_cast_fp16")]; + tensor freqs_9_to_fp16 = const()[name = tensor("freqs_9_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(108126336)))]; + tensor var_1876_cast_fp16 = mul(x = freqs_9_to_fp16, y = ts_29_cast_fp16)[name = tensor("op_1876_cast_fp16")]; + tensor rotr_9_cast_fp16 = cos(x = var_1876_cast_fp16)[name = tensor("rotr_9_cast_fp16")]; + tensor roti_9_cast_fp16 = sin(x = var_1876_cast_fp16)[name = tensor("roti_9_cast_fp16")]; + tensor var_1880_cast_fp16 = mul(x = var_1846_cast_fp16, y = rotr_9_cast_fp16)[name = tensor("op_1880_cast_fp16")]; + tensor var_1881_cast_fp16 = mul(x = var_1854_cast_fp16, y = roti_9_cast_fp16)[name = tensor("op_1881_cast_fp16")]; + tensor qor_17_cast_fp16 = sub(x = var_1880_cast_fp16, y = var_1881_cast_fp16)[name = tensor("qor_17_cast_fp16")]; + tensor var_1884_cast_fp16 = mul(x = var_1846_cast_fp16, y = roti_9_cast_fp16)[name = tensor("op_1884_cast_fp16")]; + tensor var_1885_cast_fp16 = mul(x = var_1854_cast_fp16, y = rotr_9_cast_fp16)[name = tensor("op_1885_cast_fp16")]; + tensor qoi_17_cast_fp16 = add(x = var_1884_cast_fp16, y = var_1885_cast_fp16)[name = tensor("qoi_17_cast_fp16")]; + tensor var_1888_cast_fp16 = mul(x = var_1862_cast_fp16, y = rotr_9_cast_fp16)[name = tensor("op_1888_cast_fp16")]; + tensor var_1889_cast_fp16 = mul(x = var_1870_cast_fp16, y = roti_9_cast_fp16)[name = tensor("op_1889_cast_fp16")]; + tensor kor_17_cast_fp16 = sub(x = var_1888_cast_fp16, y = var_1889_cast_fp16)[name = tensor("kor_17_cast_fp16")]; + tensor var_1892_cast_fp16 = mul(x = var_1862_cast_fp16, y = roti_9_cast_fp16)[name = tensor("op_1892_cast_fp16")]; + tensor var_1893_cast_fp16 = mul(x = var_1870_cast_fp16, y = rotr_9_cast_fp16)[name = tensor("op_1893_cast_fp16")]; + tensor koi_17_cast_fp16 = add(x = var_1892_cast_fp16, y = var_1893_cast_fp16)[name = tensor("koi_17_cast_fp16")]; + tensor qo_9_axis_0 = const()[name = tensor("qo_9_axis_0"), val = tensor(-1)]; + tensor qo_9_cast_fp16 = stack(axis = qo_9_axis_0, values = (qor_17_cast_fp16, qoi_17_cast_fp16))[name = tensor("qo_9_cast_fp16")]; + tensor ko_9_axis_0 = const()[name = tensor("ko_9_axis_0"), val = tensor(-1)]; + tensor ko_9_cast_fp16 = stack(axis = ko_9_axis_0, values = (kor_17_cast_fp16, koi_17_cast_fp16))[name = tensor("ko_9_cast_fp16")]; + tensor var_1926 = const()[name = tensor("op_1926"), val = tensor([1, 1, 16, 64])]; + tensor q_27_cast_fp16 = reshape(shape = var_1926, x = qo_9_cast_fp16)[name = tensor("q_27_cast_fp16")]; + tensor var_1932 = const()[name = tensor("op_1932"), val = tensor([1, 1, 16, 64])]; + tensor k_19_cast_fp16 = reshape(shape = var_1932, x = ko_9_cast_fp16)[name = tensor("k_19_cast_fp16")]; + tensor _inversed_1957_y_0_to_fp16 = const()[name = tensor("_inversed_1957_y_0_to_fp16"), val = tensor(0x1p-9)]; + tensor _inversed_1957_cast_fp16 = mul(x = ts_29_cast_fp16, y = _inversed_1957_y_0_to_fp16)[name = tensor("_inversed_1957_cast_fp16")]; + tensor var_1958_cast_fp16 = floor(x = _inversed_1957_cast_fp16)[name = tensor("op_1958_cast_fp16")]; + tensor var_1959_to_fp16 = const()[name = tensor("op_1959_to_fp16"), val = tensor(0x1p+9)]; + tensor var_1960_cast_fp16 = mul(x = var_1958_cast_fp16, y = var_1959_to_fp16)[name = tensor("op_1960_cast_fp16")]; + tensor write_indices_float_19_cast_fp16 = sub(x = ts_29_cast_fp16, y = var_1960_cast_fp16)[name = tensor("write_indices_float_19_cast_fp16")]; + tensor var_1967_dtype_0 = const()[name = tensor("op_1967_dtype_0"), val = tensor("int32")]; + tensor write_indices_9_reps_0 = const()[name = tensor("write_indices_9_reps_0"), val = tensor([1, 1, 16, 64])]; + tensor write_indices_float_19_cast_fp16_to_int32 = cast(dtype = var_1967_dtype_0, x = write_indices_float_19_cast_fp16)[name = tensor("cast_532")]; + tensor write_indices_9 = tile(reps = write_indices_9_reps_0, x = write_indices_float_19_cast_fp16_to_int32)[name = tensor("write_indices_9")]; + tensor var_1977_begin_0 = const()[name = tensor("op_1977_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_1977_end_0 = const()[name = tensor("op_1977_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor var_1977_end_mask_0 = const()[name = tensor("op_1977_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_1977_squeeze_mask_0 = const()[name = tensor("op_1977_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor cache4_to_fp16_dtype_0 = const()[name = tensor("cache4_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor cache4_to_fp16 = cast(dtype = cache4_to_fp16_dtype_0, x = cache4)[name = tensor("cast_531")]; + tensor var_1977_cast_fp16 = slice_by_index(begin = var_1977_begin_0, end = var_1977_end_0, end_mask = var_1977_end_mask_0, squeeze_mask = var_1977_squeeze_mask_0, x = cache4_to_fp16)[name = tensor("op_1977_cast_fp16")]; + tensor var_1979_axis_0 = const()[name = tensor("op_1979_axis_0"), val = tensor(1)]; + tensor var_1979_mode_0 = const()[name = tensor("op_1979_mode_0"), val = tensor("update")]; + tensor var_1979_validate_indices_0 = const()[name = tensor("op_1979_validate_indices_0"), val = tensor(false)]; + tensor var_1979_cast_fp16 = scatter_along_axis(axis = var_1979_axis_0, data = var_1977_cast_fp16, indices = write_indices_9, mode = var_1979_mode_0, updates = k_19_cast_fp16, validate_indices = var_1979_validate_indices_0)[name = tensor("op_1979_cast_fp16")]; + tensor concat_30 = const()[name = tensor("concat_30"), val = tensor([0, 0, 0, 0, 0])]; + tensor concat_31 = const()[name = tensor("concat_31"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_9_internal_tensor_assign_1_stride_0 = const()[name = tensor("new_cache_9_internal_tensor_assign_1_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_9_internal_tensor_assign_1_begin_mask_0 = const()[name = tensor("new_cache_9_internal_tensor_assign_1_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_9_internal_tensor_assign_1_end_mask_0 = const()[name = tensor("new_cache_9_internal_tensor_assign_1_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_9_internal_tensor_assign_1_squeeze_mask_0 = const()[name = tensor("new_cache_9_internal_tensor_assign_1_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_56 = const()[name = tensor("shape_56"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_8 = const()[name = tensor("reduce_prod_8"), val = tensor(1048576)]; + tensor range_1d_8_start_0 = const()[name = tensor("range_1d_8_start_0"), val = tensor(0)]; + tensor range_1d_8_step_0 = const()[name = tensor("range_1d_8_step_0"), val = tensor(1)]; + tensor range_1d_8 = range_1d(end = reduce_prod_8, start = range_1d_8_start_0, step = range_1d_8_step_0)[name = tensor("range_1d_8")]; + tensor reshape_40 = reshape(shape = shape_56, x = range_1d_8)[name = tensor("reshape_40")]; + tensor slice_by_index_8 = slice_by_index(begin = concat_30, begin_mask = new_cache_9_internal_tensor_assign_1_begin_mask_0, end = concat_31, end_mask = new_cache_9_internal_tensor_assign_1_end_mask_0, squeeze_mask = new_cache_9_internal_tensor_assign_1_squeeze_mask_0, stride = new_cache_9_internal_tensor_assign_1_stride_0, x = reshape_40)[name = tensor("slice_by_index_8")]; + tensor reshape_41_shape_0 = const()[name = tensor("reshape_41_shape_0"), val = tensor([-1])]; + tensor reshape_41 = reshape(shape = reshape_41_shape_0, x = slice_by_index_8)[name = tensor("reshape_41")]; + tensor reshape_42_shape_0 = const()[name = tensor("reshape_42_shape_0"), val = tensor([-1])]; + tensor reshape_42_cast_fp16 = reshape(shape = reshape_42_shape_0, x = var_1979_cast_fp16)[name = tensor("reshape_42_cast_fp16")]; + tensor reshape_43_shape_0 = const()[name = tensor("reshape_43_shape_0"), val = tensor([-1])]; + tensor reshape_43_cast_fp16 = reshape(shape = reshape_43_shape_0, x = cache4_to_fp16)[name = tensor("reshape_43_cast_fp16")]; + tensor scatter_8_mode_0 = const()[name = tensor("scatter_8_mode_0"), val = tensor("update")]; + tensor scatter_8_axis_0 = const()[name = tensor("scatter_8_axis_0"), val = tensor(0)]; + tensor scatter_8_validate_indices_0 = const()[name = tensor("scatter_8_validate_indices_0"), val = tensor(false)]; + tensor scatter_8_cast_fp16 = scatter(axis = scatter_8_axis_0, data = reshape_43_cast_fp16, indices = reshape_41, mode = scatter_8_mode_0, updates = reshape_42_cast_fp16, validate_indices = scatter_8_validate_indices_0)[name = tensor("scatter_8_cast_fp16")]; + tensor reshape_44_cast_fp16 = reshape(shape = shape_56, x = scatter_8_cast_fp16)[name = tensor("reshape_44_cast_fp16")]; + tensor var_1987_begin_0 = const()[name = tensor("op_1987_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor var_1987_end_0 = const()[name = tensor("op_1987_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor var_1987_end_mask_0 = const()[name = tensor("op_1987_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_1987_squeeze_mask_0 = const()[name = tensor("op_1987_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor var_1987_cast_fp16 = slice_by_index(begin = var_1987_begin_0, end = var_1987_end_0, end_mask = var_1987_end_mask_0, squeeze_mask = var_1987_squeeze_mask_0, x = reshape_44_cast_fp16)[name = tensor("op_1987_cast_fp16")]; + tensor var_1989_axis_0 = const()[name = tensor("op_1989_axis_0"), val = tensor(1)]; + tensor var_1989_mode_0 = const()[name = tensor("op_1989_mode_0"), val = tensor("update")]; + tensor var_1989_validate_indices_0 = const()[name = tensor("op_1989_validate_indices_0"), val = tensor(false)]; + tensor var_1989_cast_fp16 = scatter_along_axis(axis = var_1989_axis_0, data = var_1987_cast_fp16, indices = write_indices_9, mode = var_1989_mode_0, updates = v_9_cast_fp16, validate_indices = var_1989_validate_indices_0)[name = tensor("op_1989_cast_fp16")]; + tensor concat_32 = const()[name = tensor("concat_32"), val = tensor([1, 0, 0, 0, 0])]; + tensor concat_33 = const()[name = tensor("concat_33"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_9_internal_tensor_assign_2_stride_0 = const()[name = tensor("new_cache_9_internal_tensor_assign_2_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_9_internal_tensor_assign_2_begin_mask_0 = const()[name = tensor("new_cache_9_internal_tensor_assign_2_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_9_internal_tensor_assign_2_end_mask_0 = const()[name = tensor("new_cache_9_internal_tensor_assign_2_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_9_internal_tensor_assign_2_squeeze_mask_0 = const()[name = tensor("new_cache_9_internal_tensor_assign_2_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_57 = const()[name = tensor("shape_57"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_9 = const()[name = tensor("reduce_prod_9"), val = tensor(1048576)]; + tensor range_1d_9_start_0 = const()[name = tensor("range_1d_9_start_0"), val = tensor(0)]; + tensor range_1d_9_step_0 = const()[name = tensor("range_1d_9_step_0"), val = tensor(1)]; + tensor range_1d_9 = range_1d(end = reduce_prod_9, start = range_1d_9_start_0, step = range_1d_9_step_0)[name = tensor("range_1d_9")]; + tensor reshape_45 = reshape(shape = shape_57, x = range_1d_9)[name = tensor("reshape_45")]; + tensor slice_by_index_9 = slice_by_index(begin = concat_32, begin_mask = new_cache_9_internal_tensor_assign_2_begin_mask_0, end = concat_33, end_mask = new_cache_9_internal_tensor_assign_2_end_mask_0, squeeze_mask = new_cache_9_internal_tensor_assign_2_squeeze_mask_0, stride = new_cache_9_internal_tensor_assign_2_stride_0, x = reshape_45)[name = tensor("slice_by_index_9")]; + tensor reshape_46_shape_0 = const()[name = tensor("reshape_46_shape_0"), val = tensor([-1])]; + tensor reshape_46 = reshape(shape = reshape_46_shape_0, x = slice_by_index_9)[name = tensor("reshape_46")]; + tensor reshape_47_shape_0 = const()[name = tensor("reshape_47_shape_0"), val = tensor([-1])]; + tensor reshape_47_cast_fp16 = reshape(shape = reshape_47_shape_0, x = var_1989_cast_fp16)[name = tensor("reshape_47_cast_fp16")]; + tensor reshape_48_shape_0 = const()[name = tensor("reshape_48_shape_0"), val = tensor([-1])]; + tensor reshape_48_cast_fp16 = reshape(shape = reshape_48_shape_0, x = reshape_44_cast_fp16)[name = tensor("reshape_48_cast_fp16")]; + tensor scatter_9_mode_0 = const()[name = tensor("scatter_9_mode_0"), val = tensor("update")]; + tensor scatter_9_axis_0 = const()[name = tensor("scatter_9_axis_0"), val = tensor(0)]; + tensor scatter_9_validate_indices_0 = const()[name = tensor("scatter_9_validate_indices_0"), val = tensor(false)]; + tensor scatter_9_cast_fp16 = scatter(axis = scatter_9_axis_0, data = reshape_48_cast_fp16, indices = reshape_46, mode = scatter_9_mode_0, updates = reshape_47_cast_fp16, validate_indices = scatter_9_validate_indices_0)[name = tensor("scatter_9_cast_fp16")]; + tensor reshape_49_cast_fp16 = reshape(shape = shape_57, x = scatter_9_cast_fp16)[name = tensor("reshape_49_cast_fp16")]; + tensor reshape_49_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("reshape_49_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor keys_25_begin_0 = const()[name = tensor("keys_25_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor keys_25_end_0 = const()[name = tensor("keys_25_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor keys_25_end_mask_0 = const()[name = tensor("keys_25_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor keys_25_squeeze_mask_0 = const()[name = tensor("keys_25_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor keys_25_cast_fp16 = slice_by_index(begin = keys_25_begin_0, end = keys_25_end_0, end_mask = keys_25_end_mask_0, squeeze_mask = keys_25_squeeze_mask_0, x = reshape_49_cast_fp16)[name = tensor("keys_25_cast_fp16")]; + tensor values_25_begin_0 = const()[name = tensor("values_25_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor values_25_end_0 = const()[name = tensor("values_25_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor values_25_end_mask_0 = const()[name = tensor("values_25_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor values_25_squeeze_mask_0 = const()[name = tensor("values_25_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor values_25_cast_fp16 = slice_by_index(begin = values_25_begin_0, end = values_25_end_0, end_mask = values_25_end_mask_0, squeeze_mask = values_25_squeeze_mask_0, x = reshape_49_cast_fp16)[name = tensor("values_25_cast_fp16")]; + tensor var_2001_cast_fp16 = not_equal(x = keys_25_cast_fp16, y = keys_25_cast_fp16)[name = tensor("op_2001_cast_fp16")]; + tensor keys_27_cast_fp16 = select(a = var_463_to_fp16, b = keys_25_cast_fp16, cond = var_2001_cast_fp16)[name = tensor("keys_27_cast_fp16")]; + tensor var_2009_cast_fp16 = not_equal(x = values_25_cast_fp16, y = values_25_cast_fp16)[name = tensor("op_2009_cast_fp16")]; + tensor values_27_cast_fp16 = select(a = var_463_to_fp16, b = values_25_cast_fp16, cond = var_2009_cast_fp16)[name = tensor("values_27_cast_fp16")]; + tensor var_2033 = const()[name = tensor("op_2033"), val = tensor([0, 2, 1, 3])]; + tensor var_2049 = const()[name = tensor("op_2049"), val = tensor([1, 1, 1])]; + tensor var_2050_cast_fp16 = reshape(shape = var_2049, x = position4_to_fp16)[name = tensor("op_2050_cast_fp16")]; + tensor var_2070_to_fp16 = const()[name = tensor("op_2070_to_fp16"), val = tensor(0x1p+0)]; + tensor valid_len_9_cast_fp16 = add(x = var_2050_cast_fp16, y = var_2070_to_fp16)[name = tensor("valid_len_9_cast_fp16")]; + tensor valid_mask_9_cast_fp16 = less(x = k_positions_1_promoted_to_fp16, y = valid_len_9_cast_fp16)[name = tensor("valid_mask_9_cast_fp16")]; + tensor causal_mask_9_cast_fp16 = less_equal(x = k_positions_1_promoted_to_fp16, y = var_2050_cast_fp16)[name = tensor("causal_mask_9_cast_fp16")]; + tensor attn_mask_17 = logical_and(x = valid_mask_9_cast_fp16, y = causal_mask_9_cast_fp16)[name = tensor("attn_mask_17")]; + tensor attn_mask_19_axes_0 = const()[name = tensor("attn_mask_19_axes_0"), val = tensor([1])]; + tensor attn_mask_19 = expand_dims(axes = attn_mask_19_axes_0, x = attn_mask_17)[name = tensor("attn_mask_19")]; + tensor var_2088_transpose_x_0 = const()[name = tensor("op_2088_transpose_x_0"), val = tensor(false)]; + tensor var_2088_transpose_y_0 = const()[name = tensor("op_2088_transpose_y_0"), val = tensor(false)]; + tensor transpose_80_perm_0 = const()[name = tensor("transpose_80_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_81_perm_0 = const()[name = tensor("transpose_81_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_81 = transpose(perm = transpose_81_perm_0, x = keys_27_cast_fp16)[name = tensor("transpose_197")]; + tensor transpose_80 = transpose(perm = transpose_80_perm_0, x = q_27_cast_fp16)[name = tensor("transpose_198")]; + tensor var_2088_cast_fp16 = matmul(transpose_x = var_2088_transpose_x_0, transpose_y = var_2088_transpose_y_0, x = transpose_80, y = transpose_81)[name = tensor("op_2088_cast_fp16")]; + tensor var_2082_to_fp16 = const()[name = tensor("op_2082_to_fp16"), val = tensor([0x1p-3])]; + tensor attn_weights_25_cast_fp16 = mul(x = var_2088_cast_fp16, y = var_2082_to_fp16)[name = tensor("attn_weights_25_cast_fp16")]; + tensor var_2090 = logical_not(x = attn_mask_19)[name = tensor("op_2090")]; + tensor var_2091_to_fp16 = const()[name = tensor("op_2091_to_fp16"), val = tensor(-inf)]; + tensor attn_weights_27_cast_fp16 = select(a = var_2091_to_fp16, b = attn_weights_25_cast_fp16, cond = var_2090)[name = tensor("attn_weights_27_cast_fp16")]; + tensor var_2093 = const()[name = tensor("op_2093"), val = tensor(-1)]; + tensor attn_weights_29_cast_fp16 = softmax(axis = var_2093, x = attn_weights_27_cast_fp16)[name = tensor("attn_weights_29_cast_fp16")]; + tensor attn_output_9_transpose_x_0 = const()[name = tensor("attn_output_9_transpose_x_0"), val = tensor(false)]; + tensor attn_output_9_transpose_y_0 = const()[name = tensor("attn_output_9_transpose_y_0"), val = tensor(false)]; + tensor values_29_cast_fp16 = transpose(perm = var_2033, x = values_27_cast_fp16)[name = tensor("transpose_199")]; + tensor attn_output_9_cast_fp16 = matmul(transpose_x = attn_output_9_transpose_x_0, transpose_y = attn_output_9_transpose_y_0, x = attn_weights_29_cast_fp16, y = values_29_cast_fp16)[name = tensor("attn_output_9_cast_fp16")]; + tensor var_2101 = const()[name = tensor("op_2101"), val = tensor([0, 2, 1, 3])]; + tensor var_2106 = const()[name = tensor("op_2106"), val = tensor([1, 1, 1024])]; + tensor var_2102_cast_fp16 = transpose(perm = var_2101, x = attn_output_9_cast_fp16)[name = tensor("transpose_196")]; + tensor input_55_cast_fp16 = reshape(shape = var_2106, x = var_2102_cast_fp16)[name = tensor("input_55_cast_fp16")]; + tensor attn4_out_proj_weight_to_fp16 = const()[name = tensor("attn4_out_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(108126464)))]; + tensor linear_18_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = attn4_out_proj_weight_to_fp16, x = input_55_cast_fp16)[name = tensor("linear_18_cast_fp16")]; + tensor var_2112_to_fp16 = const()[name = tensor("op_2112_to_fp16"), val = tensor(0x1p+0)]; + tensor var_2113_cast_fp16 = add(x = position4_to_fp16, y = var_2112_to_fp16)[name = tensor("op_2113_cast_fp16")]; + tensor var_2113_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_2113_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor input_57_cast_fp16 = add(x = input_51_cast_fp16, y = linear_18_cast_fp16)[name = tensor("input_57_cast_fp16")]; + tensor input_59_axes_0 = const()[name = tensor("input_59_axes_0"), val = tensor([-1])]; + tensor norm4_2_weight_to_fp16 = const()[name = tensor("norm4_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(110223680)))]; + tensor norm4_2_bias_to_fp16 = const()[name = tensor("norm4_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(110225792)))]; + tensor var_2117_to_fp16 = const()[name = tensor("op_2117_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_59_cast_fp16 = layer_norm(axes = input_59_axes_0, beta = norm4_2_bias_to_fp16, epsilon = var_2117_to_fp16, gamma = norm4_2_weight_to_fp16, x = input_57_cast_fp16)[name = tensor("input_59_cast_fp16")]; + tensor linear4_1_weight_to_fp16 = const()[name = tensor("linear4_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(110227904)))]; + tensor linear_19_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = linear4_1_weight_to_fp16, x = input_59_cast_fp16)[name = tensor("linear_19_cast_fp16")]; + tensor input_61_mode_0 = const()[name = tensor("input_61_mode_0"), val = tensor("EXACT")]; + tensor input_61_cast_fp16 = gelu(mode = input_61_mode_0, x = linear_19_cast_fp16)[name = tensor("input_61_cast_fp16")]; + tensor linear4_2_weight_to_fp16 = const()[name = tensor("linear4_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(118616576)))]; + tensor linear_20_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = linear4_2_weight_to_fp16, x = input_61_cast_fp16)[name = tensor("linear_20_cast_fp16")]; + tensor input_63_cast_fp16 = add(x = input_57_cast_fp16, y = linear_20_cast_fp16)[name = tensor("input_63_cast_fp16")]; + tensor input_65_axes_0 = const()[name = tensor("input_65_axes_0"), val = tensor([-1])]; + tensor norm5_1_weight_to_fp16 = const()[name = tensor("norm5_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(127005248)))]; + tensor norm5_1_bias_to_fp16 = const()[name = tensor("norm5_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(127007360)))]; + tensor var_2134_to_fp16 = const()[name = tensor("op_2134_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_65_cast_fp16 = layer_norm(axes = input_65_axes_0, beta = norm5_1_bias_to_fp16, epsilon = var_2134_to_fp16, gamma = norm5_1_weight_to_fp16, x = input_63_cast_fp16)[name = tensor("input_65_cast_fp16")]; + tensor attn5_in_proj_weight_to_fp16 = const()[name = tensor("attn5_in_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(127009472)))]; + tensor linear_21_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = attn5_in_proj_weight_to_fp16, x = input_65_cast_fp16)[name = tensor("linear_21_cast_fp16")]; + tensor var_2148 = const()[name = tensor("op_2148"), val = tensor([1, 1, 3, 16, 64])]; + tensor qkv_11_cast_fp16 = reshape(shape = var_2148, x = linear_21_cast_fp16)[name = tensor("qkv_11_cast_fp16")]; + tensor q_31_begin_0 = const()[name = tensor("q_31_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor q_31_end_0 = const()[name = tensor("q_31_end_0"), val = tensor([1, 1, 1, 16, 64])]; + tensor q_31_end_mask_0 = const()[name = tensor("q_31_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor q_31_squeeze_mask_0 = const()[name = tensor("q_31_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor q_31_cast_fp16 = slice_by_index(begin = q_31_begin_0, end = q_31_end_0, end_mask = q_31_end_mask_0, squeeze_mask = q_31_squeeze_mask_0, x = qkv_11_cast_fp16)[name = tensor("q_31_cast_fp16")]; + tensor k_21_begin_0 = const()[name = tensor("k_21_begin_0"), val = tensor([0, 0, 1, 0, 0])]; + tensor k_21_end_0 = const()[name = tensor("k_21_end_0"), val = tensor([1, 1, 2, 16, 64])]; + tensor k_21_end_mask_0 = const()[name = tensor("k_21_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor k_21_squeeze_mask_0 = const()[name = tensor("k_21_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor k_21_cast_fp16 = slice_by_index(begin = k_21_begin_0, end = k_21_end_0, end_mask = k_21_end_mask_0, squeeze_mask = k_21_squeeze_mask_0, x = qkv_11_cast_fp16)[name = tensor("k_21_cast_fp16")]; + tensor v_11_begin_0 = const()[name = tensor("v_11_begin_0"), val = tensor([0, 0, 2, 0, 0])]; + tensor v_11_end_0 = const()[name = tensor("v_11_end_0"), val = tensor([1, 1, 3, 16, 64])]; + tensor v_11_end_mask_0 = const()[name = tensor("v_11_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor v_11_squeeze_mask_0 = const()[name = tensor("v_11_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor v_11_cast_fp16 = slice_by_index(begin = v_11_begin_0, end = v_11_end_0, end_mask = v_11_end_mask_0, squeeze_mask = v_11_squeeze_mask_0, x = qkv_11_cast_fp16)[name = tensor("v_11_cast_fp16")]; + tensor var_2214 = const()[name = tensor("op_2214"), val = tensor([1, 1, 1, 1])]; + tensor position5_to_fp16_dtype_0 = const()[name = tensor("position5_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor position5_to_fp16 = cast(dtype = position5_to_fp16_dtype_0, x = position5)[name = tensor("cast_528")]; + tensor ts_35_cast_fp16 = reshape(shape = var_2214, x = position5_to_fp16)[name = tensor("ts_35_cast_fp16")]; + tensor var_2221 = const()[name = tensor("op_2221"), val = tensor([1, 1, 16, 32, 2])]; + tensor q_complex_11_cast_fp16 = reshape(shape = var_2221, x = q_31_cast_fp16)[name = tensor("q_complex_11_cast_fp16")]; + tensor var_2228 = const()[name = tensor("op_2228"), val = tensor([1, 1, 16, 32, 2])]; + tensor k_complex_11_cast_fp16 = reshape(shape = var_2228, x = k_21_cast_fp16)[name = tensor("k_complex_11_cast_fp16")]; + tensor var_2232_begin_0 = const()[name = tensor("op_2232_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_2232_end_0 = const()[name = tensor("op_2232_end_0"), val = tensor([1, 1, 16, 32, 1])]; + tensor var_2232_end_mask_0 = const()[name = tensor("op_2232_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_2232_squeeze_mask_0 = const()[name = tensor("op_2232_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_2232_cast_fp16 = slice_by_index(begin = var_2232_begin_0, end = var_2232_end_0, end_mask = var_2232_end_mask_0, squeeze_mask = var_2232_squeeze_mask_0, x = q_complex_11_cast_fp16)[name = tensor("op_2232_cast_fp16")]; + tensor var_2240_begin_0 = const()[name = tensor("op_2240_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_2240_end_0 = const()[name = tensor("op_2240_end_0"), val = tensor([1, 1, 16, 32, 2])]; + tensor var_2240_end_mask_0 = const()[name = tensor("op_2240_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_2240_squeeze_mask_0 = const()[name = tensor("op_2240_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_2240_cast_fp16 = slice_by_index(begin = var_2240_begin_0, end = var_2240_end_0, end_mask = var_2240_end_mask_0, squeeze_mask = var_2240_squeeze_mask_0, x = q_complex_11_cast_fp16)[name = tensor("op_2240_cast_fp16")]; + tensor var_2248_begin_0 = const()[name = tensor("op_2248_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_2248_end_0 = const()[name = tensor("op_2248_end_0"), val = tensor([1, 1, 16, 32, 1])]; + tensor var_2248_end_mask_0 = const()[name = tensor("op_2248_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_2248_squeeze_mask_0 = const()[name = tensor("op_2248_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_2248_cast_fp16 = slice_by_index(begin = var_2248_begin_0, end = var_2248_end_0, end_mask = var_2248_end_mask_0, squeeze_mask = var_2248_squeeze_mask_0, x = k_complex_11_cast_fp16)[name = tensor("op_2248_cast_fp16")]; + tensor var_2256_begin_0 = const()[name = tensor("op_2256_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_2256_end_0 = const()[name = tensor("op_2256_end_0"), val = tensor([1, 1, 16, 32, 2])]; + tensor var_2256_end_mask_0 = const()[name = tensor("op_2256_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_2256_squeeze_mask_0 = const()[name = tensor("op_2256_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_2256_cast_fp16 = slice_by_index(begin = var_2256_begin_0, end = var_2256_end_0, end_mask = var_2256_end_mask_0, squeeze_mask = var_2256_squeeze_mask_0, x = k_complex_11_cast_fp16)[name = tensor("op_2256_cast_fp16")]; + tensor freqs_11_to_fp16 = const()[name = tensor("freqs_11_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(133300992)))]; + tensor var_2262_cast_fp16 = mul(x = freqs_11_to_fp16, y = ts_35_cast_fp16)[name = tensor("op_2262_cast_fp16")]; + tensor rotr_11_cast_fp16 = cos(x = var_2262_cast_fp16)[name = tensor("rotr_11_cast_fp16")]; + tensor roti_11_cast_fp16 = sin(x = var_2262_cast_fp16)[name = tensor("roti_11_cast_fp16")]; + tensor var_2266_cast_fp16 = mul(x = var_2232_cast_fp16, y = rotr_11_cast_fp16)[name = tensor("op_2266_cast_fp16")]; + tensor var_2267_cast_fp16 = mul(x = var_2240_cast_fp16, y = roti_11_cast_fp16)[name = tensor("op_2267_cast_fp16")]; + tensor qor_21_cast_fp16 = sub(x = var_2266_cast_fp16, y = var_2267_cast_fp16)[name = tensor("qor_21_cast_fp16")]; + tensor var_2270_cast_fp16 = mul(x = var_2232_cast_fp16, y = roti_11_cast_fp16)[name = tensor("op_2270_cast_fp16")]; + tensor var_2271_cast_fp16 = mul(x = var_2240_cast_fp16, y = rotr_11_cast_fp16)[name = tensor("op_2271_cast_fp16")]; + tensor qoi_21_cast_fp16 = add(x = var_2270_cast_fp16, y = var_2271_cast_fp16)[name = tensor("qoi_21_cast_fp16")]; + tensor var_2274_cast_fp16 = mul(x = var_2248_cast_fp16, y = rotr_11_cast_fp16)[name = tensor("op_2274_cast_fp16")]; + tensor var_2275_cast_fp16 = mul(x = var_2256_cast_fp16, y = roti_11_cast_fp16)[name = tensor("op_2275_cast_fp16")]; + tensor kor_21_cast_fp16 = sub(x = var_2274_cast_fp16, y = var_2275_cast_fp16)[name = tensor("kor_21_cast_fp16")]; + tensor var_2278_cast_fp16 = mul(x = var_2248_cast_fp16, y = roti_11_cast_fp16)[name = tensor("op_2278_cast_fp16")]; + tensor var_2279_cast_fp16 = mul(x = var_2256_cast_fp16, y = rotr_11_cast_fp16)[name = tensor("op_2279_cast_fp16")]; + tensor koi_21_cast_fp16 = add(x = var_2278_cast_fp16, y = var_2279_cast_fp16)[name = tensor("koi_21_cast_fp16")]; + tensor qo_11_axis_0 = const()[name = tensor("qo_11_axis_0"), val = tensor(-1)]; + tensor qo_11_cast_fp16 = stack(axis = qo_11_axis_0, values = (qor_21_cast_fp16, qoi_21_cast_fp16))[name = tensor("qo_11_cast_fp16")]; + tensor ko_11_axis_0 = const()[name = tensor("ko_11_axis_0"), val = tensor(-1)]; + tensor ko_11_cast_fp16 = stack(axis = ko_11_axis_0, values = (kor_21_cast_fp16, koi_21_cast_fp16))[name = tensor("ko_11_cast_fp16")]; + tensor var_2312 = const()[name = tensor("op_2312"), val = tensor([1, 1, 16, 64])]; + tensor q_33_cast_fp16 = reshape(shape = var_2312, x = qo_11_cast_fp16)[name = tensor("q_33_cast_fp16")]; + tensor var_2318 = const()[name = tensor("op_2318"), val = tensor([1, 1, 16, 64])]; + tensor k_23_cast_fp16 = reshape(shape = var_2318, x = ko_11_cast_fp16)[name = tensor("k_23_cast_fp16")]; + tensor _inversed_2343_y_0_to_fp16 = const()[name = tensor("_inversed_2343_y_0_to_fp16"), val = tensor(0x1p-9)]; + tensor _inversed_2343_cast_fp16 = mul(x = ts_35_cast_fp16, y = _inversed_2343_y_0_to_fp16)[name = tensor("_inversed_2343_cast_fp16")]; + tensor var_2344_cast_fp16 = floor(x = _inversed_2343_cast_fp16)[name = tensor("op_2344_cast_fp16")]; + tensor var_2345_to_fp16 = const()[name = tensor("op_2345_to_fp16"), val = tensor(0x1p+9)]; + tensor var_2346_cast_fp16 = mul(x = var_2344_cast_fp16, y = var_2345_to_fp16)[name = tensor("op_2346_cast_fp16")]; + tensor write_indices_float_23_cast_fp16 = sub(x = ts_35_cast_fp16, y = var_2346_cast_fp16)[name = tensor("write_indices_float_23_cast_fp16")]; + tensor var_2353_dtype_0 = const()[name = tensor("op_2353_dtype_0"), val = tensor("int32")]; + tensor write_indices_11_reps_0 = const()[name = tensor("write_indices_11_reps_0"), val = tensor([1, 1, 16, 64])]; + tensor write_indices_float_23_cast_fp16_to_int32 = cast(dtype = var_2353_dtype_0, x = write_indices_float_23_cast_fp16)[name = tensor("cast_527")]; + tensor write_indices_11 = tile(reps = write_indices_11_reps_0, x = write_indices_float_23_cast_fp16_to_int32)[name = tensor("write_indices_11")]; + tensor var_2363_begin_0 = const()[name = tensor("op_2363_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_2363_end_0 = const()[name = tensor("op_2363_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor var_2363_end_mask_0 = const()[name = tensor("op_2363_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_2363_squeeze_mask_0 = const()[name = tensor("op_2363_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor cache5_to_fp16_dtype_0 = const()[name = tensor("cache5_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor cache5_to_fp16 = cast(dtype = cache5_to_fp16_dtype_0, x = cache5)[name = tensor("cast_526")]; + tensor var_2363_cast_fp16 = slice_by_index(begin = var_2363_begin_0, end = var_2363_end_0, end_mask = var_2363_end_mask_0, squeeze_mask = var_2363_squeeze_mask_0, x = cache5_to_fp16)[name = tensor("op_2363_cast_fp16")]; + tensor var_2365_axis_0 = const()[name = tensor("op_2365_axis_0"), val = tensor(1)]; + tensor var_2365_mode_0 = const()[name = tensor("op_2365_mode_0"), val = tensor("update")]; + tensor var_2365_validate_indices_0 = const()[name = tensor("op_2365_validate_indices_0"), val = tensor(false)]; + tensor var_2365_cast_fp16 = scatter_along_axis(axis = var_2365_axis_0, data = var_2363_cast_fp16, indices = write_indices_11, mode = var_2365_mode_0, updates = k_23_cast_fp16, validate_indices = var_2365_validate_indices_0)[name = tensor("op_2365_cast_fp16")]; + tensor concat_37 = const()[name = tensor("concat_37"), val = tensor([0, 0, 0, 0, 0])]; + tensor concat_38 = const()[name = tensor("concat_38"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_11_internal_tensor_assign_1_stride_0 = const()[name = tensor("new_cache_11_internal_tensor_assign_1_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_11_internal_tensor_assign_1_begin_mask_0 = const()[name = tensor("new_cache_11_internal_tensor_assign_1_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_11_internal_tensor_assign_1_end_mask_0 = const()[name = tensor("new_cache_11_internal_tensor_assign_1_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_11_internal_tensor_assign_1_squeeze_mask_0 = const()[name = tensor("new_cache_11_internal_tensor_assign_1_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_58 = const()[name = tensor("shape_58"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_10 = const()[name = tensor("reduce_prod_10"), val = tensor(1048576)]; + tensor range_1d_10_start_0 = const()[name = tensor("range_1d_10_start_0"), val = tensor(0)]; + tensor range_1d_10_step_0 = const()[name = tensor("range_1d_10_step_0"), val = tensor(1)]; + tensor range_1d_10 = range_1d(end = reduce_prod_10, start = range_1d_10_start_0, step = range_1d_10_step_0)[name = tensor("range_1d_10")]; + tensor reshape_50 = reshape(shape = shape_58, x = range_1d_10)[name = tensor("reshape_50")]; + tensor slice_by_index_10 = slice_by_index(begin = concat_37, begin_mask = new_cache_11_internal_tensor_assign_1_begin_mask_0, end = concat_38, end_mask = new_cache_11_internal_tensor_assign_1_end_mask_0, squeeze_mask = new_cache_11_internal_tensor_assign_1_squeeze_mask_0, stride = new_cache_11_internal_tensor_assign_1_stride_0, x = reshape_50)[name = tensor("slice_by_index_10")]; + tensor reshape_51_shape_0 = const()[name = tensor("reshape_51_shape_0"), val = tensor([-1])]; + tensor reshape_51 = reshape(shape = reshape_51_shape_0, x = slice_by_index_10)[name = tensor("reshape_51")]; + tensor reshape_52_shape_0 = const()[name = tensor("reshape_52_shape_0"), val = tensor([-1])]; + tensor reshape_52_cast_fp16 = reshape(shape = reshape_52_shape_0, x = var_2365_cast_fp16)[name = tensor("reshape_52_cast_fp16")]; + tensor reshape_53_shape_0 = const()[name = tensor("reshape_53_shape_0"), val = tensor([-1])]; + tensor reshape_53_cast_fp16 = reshape(shape = reshape_53_shape_0, x = cache5_to_fp16)[name = tensor("reshape_53_cast_fp16")]; + tensor scatter_10_mode_0 = const()[name = tensor("scatter_10_mode_0"), val = tensor("update")]; + tensor scatter_10_axis_0 = const()[name = tensor("scatter_10_axis_0"), val = tensor(0)]; + tensor scatter_10_validate_indices_0 = const()[name = tensor("scatter_10_validate_indices_0"), val = tensor(false)]; + tensor scatter_10_cast_fp16 = scatter(axis = scatter_10_axis_0, data = reshape_53_cast_fp16, indices = reshape_51, mode = scatter_10_mode_0, updates = reshape_52_cast_fp16, validate_indices = scatter_10_validate_indices_0)[name = tensor("scatter_10_cast_fp16")]; + tensor reshape_54_cast_fp16 = reshape(shape = shape_58, x = scatter_10_cast_fp16)[name = tensor("reshape_54_cast_fp16")]; + tensor var_2373_begin_0 = const()[name = tensor("op_2373_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor var_2373_end_0 = const()[name = tensor("op_2373_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor var_2373_end_mask_0 = const()[name = tensor("op_2373_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_2373_squeeze_mask_0 = const()[name = tensor("op_2373_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor var_2373_cast_fp16 = slice_by_index(begin = var_2373_begin_0, end = var_2373_end_0, end_mask = var_2373_end_mask_0, squeeze_mask = var_2373_squeeze_mask_0, x = reshape_54_cast_fp16)[name = tensor("op_2373_cast_fp16")]; + tensor var_2375_axis_0 = const()[name = tensor("op_2375_axis_0"), val = tensor(1)]; + tensor var_2375_mode_0 = const()[name = tensor("op_2375_mode_0"), val = tensor("update")]; + tensor var_2375_validate_indices_0 = const()[name = tensor("op_2375_validate_indices_0"), val = tensor(false)]; + tensor var_2375_cast_fp16 = scatter_along_axis(axis = var_2375_axis_0, data = var_2373_cast_fp16, indices = write_indices_11, mode = var_2375_mode_0, updates = v_11_cast_fp16, validate_indices = var_2375_validate_indices_0)[name = tensor("op_2375_cast_fp16")]; + tensor concat_39 = const()[name = tensor("concat_39"), val = tensor([1, 0, 0, 0, 0])]; + tensor concat_40 = const()[name = tensor("concat_40"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_11_internal_tensor_assign_2_stride_0 = const()[name = tensor("new_cache_11_internal_tensor_assign_2_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_11_internal_tensor_assign_2_begin_mask_0 = const()[name = tensor("new_cache_11_internal_tensor_assign_2_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_11_internal_tensor_assign_2_end_mask_0 = const()[name = tensor("new_cache_11_internal_tensor_assign_2_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_11_internal_tensor_assign_2_squeeze_mask_0 = const()[name = tensor("new_cache_11_internal_tensor_assign_2_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_59 = const()[name = tensor("shape_59"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_11 = const()[name = tensor("reduce_prod_11"), val = tensor(1048576)]; + tensor range_1d_11_start_0 = const()[name = tensor("range_1d_11_start_0"), val = tensor(0)]; + tensor range_1d_11_step_0 = const()[name = tensor("range_1d_11_step_0"), val = tensor(1)]; + tensor range_1d_11 = range_1d(end = reduce_prod_11, start = range_1d_11_start_0, step = range_1d_11_step_0)[name = tensor("range_1d_11")]; + tensor reshape_55 = reshape(shape = shape_59, x = range_1d_11)[name = tensor("reshape_55")]; + tensor slice_by_index_11 = slice_by_index(begin = concat_39, begin_mask = new_cache_11_internal_tensor_assign_2_begin_mask_0, end = concat_40, end_mask = new_cache_11_internal_tensor_assign_2_end_mask_0, squeeze_mask = new_cache_11_internal_tensor_assign_2_squeeze_mask_0, stride = new_cache_11_internal_tensor_assign_2_stride_0, x = reshape_55)[name = tensor("slice_by_index_11")]; + tensor reshape_56_shape_0 = const()[name = tensor("reshape_56_shape_0"), val = tensor([-1])]; + tensor reshape_56 = reshape(shape = reshape_56_shape_0, x = slice_by_index_11)[name = tensor("reshape_56")]; + tensor reshape_57_shape_0 = const()[name = tensor("reshape_57_shape_0"), val = tensor([-1])]; + tensor reshape_57_cast_fp16 = reshape(shape = reshape_57_shape_0, x = var_2375_cast_fp16)[name = tensor("reshape_57_cast_fp16")]; + tensor reshape_58_shape_0 = const()[name = tensor("reshape_58_shape_0"), val = tensor([-1])]; + tensor reshape_58_cast_fp16 = reshape(shape = reshape_58_shape_0, x = reshape_54_cast_fp16)[name = tensor("reshape_58_cast_fp16")]; + tensor scatter_11_mode_0 = const()[name = tensor("scatter_11_mode_0"), val = tensor("update")]; + tensor scatter_11_axis_0 = const()[name = tensor("scatter_11_axis_0"), val = tensor(0)]; + tensor scatter_11_validate_indices_0 = const()[name = tensor("scatter_11_validate_indices_0"), val = tensor(false)]; + tensor scatter_11_cast_fp16 = scatter(axis = scatter_11_axis_0, data = reshape_58_cast_fp16, indices = reshape_56, mode = scatter_11_mode_0, updates = reshape_57_cast_fp16, validate_indices = scatter_11_validate_indices_0)[name = tensor("scatter_11_cast_fp16")]; + tensor reshape_59_cast_fp16 = reshape(shape = shape_59, x = scatter_11_cast_fp16)[name = tensor("reshape_59_cast_fp16")]; + tensor reshape_59_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("reshape_59_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor keys_31_begin_0 = const()[name = tensor("keys_31_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor keys_31_end_0 = const()[name = tensor("keys_31_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor keys_31_end_mask_0 = const()[name = tensor("keys_31_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor keys_31_squeeze_mask_0 = const()[name = tensor("keys_31_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor keys_31_cast_fp16 = slice_by_index(begin = keys_31_begin_0, end = keys_31_end_0, end_mask = keys_31_end_mask_0, squeeze_mask = keys_31_squeeze_mask_0, x = reshape_59_cast_fp16)[name = tensor("keys_31_cast_fp16")]; + tensor values_31_begin_0 = const()[name = tensor("values_31_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor values_31_end_0 = const()[name = tensor("values_31_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor values_31_end_mask_0 = const()[name = tensor("values_31_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor values_31_squeeze_mask_0 = const()[name = tensor("values_31_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor values_31_cast_fp16 = slice_by_index(begin = values_31_begin_0, end = values_31_end_0, end_mask = values_31_end_mask_0, squeeze_mask = values_31_squeeze_mask_0, x = reshape_59_cast_fp16)[name = tensor("values_31_cast_fp16")]; + tensor var_2387_cast_fp16 = not_equal(x = keys_31_cast_fp16, y = keys_31_cast_fp16)[name = tensor("op_2387_cast_fp16")]; + tensor keys_33_cast_fp16 = select(a = var_463_to_fp16, b = keys_31_cast_fp16, cond = var_2387_cast_fp16)[name = tensor("keys_33_cast_fp16")]; + tensor var_2395_cast_fp16 = not_equal(x = values_31_cast_fp16, y = values_31_cast_fp16)[name = tensor("op_2395_cast_fp16")]; + tensor values_33_cast_fp16 = select(a = var_463_to_fp16, b = values_31_cast_fp16, cond = var_2395_cast_fp16)[name = tensor("values_33_cast_fp16")]; + tensor var_2419 = const()[name = tensor("op_2419"), val = tensor([0, 2, 1, 3])]; + tensor var_2435 = const()[name = tensor("op_2435"), val = tensor([1, 1, 1])]; + tensor var_2436_cast_fp16 = reshape(shape = var_2435, x = position5_to_fp16)[name = tensor("op_2436_cast_fp16")]; + tensor var_2456_to_fp16 = const()[name = tensor("op_2456_to_fp16"), val = tensor(0x1p+0)]; + tensor valid_len_11_cast_fp16 = add(x = var_2436_cast_fp16, y = var_2456_to_fp16)[name = tensor("valid_len_11_cast_fp16")]; + tensor valid_mask_11_cast_fp16 = less(x = k_positions_1_promoted_to_fp16, y = valid_len_11_cast_fp16)[name = tensor("valid_mask_11_cast_fp16")]; + tensor causal_mask_11_cast_fp16 = less_equal(x = k_positions_1_promoted_to_fp16, y = var_2436_cast_fp16)[name = tensor("causal_mask_11_cast_fp16")]; + tensor attn_mask_21 = logical_and(x = valid_mask_11_cast_fp16, y = causal_mask_11_cast_fp16)[name = tensor("attn_mask_21")]; + tensor attn_mask_23_axes_0 = const()[name = tensor("attn_mask_23_axes_0"), val = tensor([1])]; + tensor attn_mask_23 = expand_dims(axes = attn_mask_23_axes_0, x = attn_mask_21)[name = tensor("attn_mask_23")]; + tensor var_2474_transpose_x_0 = const()[name = tensor("op_2474_transpose_x_0"), val = tensor(false)]; + tensor var_2474_transpose_y_0 = const()[name = tensor("op_2474_transpose_y_0"), val = tensor(false)]; + tensor transpose_82_perm_0 = const()[name = tensor("transpose_82_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_83_perm_0 = const()[name = tensor("transpose_83_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_83 = transpose(perm = transpose_83_perm_0, x = keys_33_cast_fp16)[name = tensor("transpose_193")]; + tensor transpose_82 = transpose(perm = transpose_82_perm_0, x = q_33_cast_fp16)[name = tensor("transpose_194")]; + tensor var_2474_cast_fp16 = matmul(transpose_x = var_2474_transpose_x_0, transpose_y = var_2474_transpose_y_0, x = transpose_82, y = transpose_83)[name = tensor("op_2474_cast_fp16")]; + tensor var_2468_to_fp16 = const()[name = tensor("op_2468_to_fp16"), val = tensor([0x1p-3])]; + tensor attn_weights_31_cast_fp16 = mul(x = var_2474_cast_fp16, y = var_2468_to_fp16)[name = tensor("attn_weights_31_cast_fp16")]; + tensor var_2476 = logical_not(x = attn_mask_23)[name = tensor("op_2476")]; + tensor var_2477_to_fp16 = const()[name = tensor("op_2477_to_fp16"), val = tensor(-inf)]; + tensor attn_weights_33_cast_fp16 = select(a = var_2477_to_fp16, b = attn_weights_31_cast_fp16, cond = var_2476)[name = tensor("attn_weights_33_cast_fp16")]; + tensor var_2479 = const()[name = tensor("op_2479"), val = tensor(-1)]; + tensor attn_weights_35_cast_fp16 = softmax(axis = var_2479, x = attn_weights_33_cast_fp16)[name = tensor("attn_weights_35_cast_fp16")]; + tensor attn_output_11_transpose_x_0 = const()[name = tensor("attn_output_11_transpose_x_0"), val = tensor(false)]; + tensor attn_output_11_transpose_y_0 = const()[name = tensor("attn_output_11_transpose_y_0"), val = tensor(false)]; + tensor values_35_cast_fp16 = transpose(perm = var_2419, x = values_33_cast_fp16)[name = tensor("transpose_195")]; + tensor attn_output_11_cast_fp16 = matmul(transpose_x = attn_output_11_transpose_x_0, transpose_y = attn_output_11_transpose_y_0, x = attn_weights_35_cast_fp16, y = values_35_cast_fp16)[name = tensor("attn_output_11_cast_fp16")]; + tensor var_2487 = const()[name = tensor("op_2487"), val = tensor([0, 2, 1, 3])]; + tensor var_2492 = const()[name = tensor("op_2492"), val = tensor([1, 1, 1024])]; + tensor var_2488_cast_fp16 = transpose(perm = var_2487, x = attn_output_11_cast_fp16)[name = tensor("transpose_192")]; + tensor input_67_cast_fp16 = reshape(shape = var_2492, x = var_2488_cast_fp16)[name = tensor("input_67_cast_fp16")]; + tensor attn5_out_proj_weight_to_fp16 = const()[name = tensor("attn5_out_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(133301120)))]; + tensor linear_22_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = attn5_out_proj_weight_to_fp16, x = input_67_cast_fp16)[name = tensor("linear_22_cast_fp16")]; + tensor var_2498_to_fp16 = const()[name = tensor("op_2498_to_fp16"), val = tensor(0x1p+0)]; + tensor var_2499_cast_fp16 = add(x = position5_to_fp16, y = var_2498_to_fp16)[name = tensor("op_2499_cast_fp16")]; + tensor var_2499_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_2499_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor input_69_cast_fp16 = add(x = input_63_cast_fp16, y = linear_22_cast_fp16)[name = tensor("input_69_cast_fp16")]; + tensor input_71_axes_0 = const()[name = tensor("input_71_axes_0"), val = tensor([-1])]; + tensor norm5_2_weight_to_fp16 = const()[name = tensor("norm5_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(135398336)))]; + tensor norm5_2_bias_to_fp16 = const()[name = tensor("norm5_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(135400448)))]; + tensor var_2503_to_fp16 = const()[name = tensor("op_2503_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_71_cast_fp16 = layer_norm(axes = input_71_axes_0, beta = norm5_2_bias_to_fp16, epsilon = var_2503_to_fp16, gamma = norm5_2_weight_to_fp16, x = input_69_cast_fp16)[name = tensor("input_71_cast_fp16")]; + tensor linear5_1_weight_to_fp16 = const()[name = tensor("linear5_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(135402560)))]; + tensor linear_23_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = linear5_1_weight_to_fp16, x = input_71_cast_fp16)[name = tensor("linear_23_cast_fp16")]; + tensor input_73_mode_0 = const()[name = tensor("input_73_mode_0"), val = tensor("EXACT")]; + tensor input_73_cast_fp16 = gelu(mode = input_73_mode_0, x = linear_23_cast_fp16)[name = tensor("input_73_cast_fp16")]; + tensor linear5_2_weight_to_fp16 = const()[name = tensor("linear5_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(143791232)))]; + tensor linear_24_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = linear5_2_weight_to_fp16, x = input_73_cast_fp16)[name = tensor("linear_24_cast_fp16")]; + tensor input_75_cast_fp16 = add(x = input_69_cast_fp16, y = linear_24_cast_fp16)[name = tensor("input_75_cast_fp16")]; + tensor input_77_axes_0 = const()[name = tensor("input_77_axes_0"), val = tensor([-1])]; + tensor norm6_1_weight_to_fp16 = const()[name = tensor("norm6_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(152179904)))]; + tensor norm6_1_bias_to_fp16 = const()[name = tensor("norm6_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(152182016)))]; + tensor var_2520_to_fp16 = const()[name = tensor("op_2520_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_77_cast_fp16 = layer_norm(axes = input_77_axes_0, beta = norm6_1_bias_to_fp16, epsilon = var_2520_to_fp16, gamma = norm6_1_weight_to_fp16, x = input_75_cast_fp16)[name = tensor("input_77_cast_fp16")]; + tensor attn6_in_proj_weight_to_fp16 = const()[name = tensor("attn6_in_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(152184128)))]; + tensor linear_25_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = attn6_in_proj_weight_to_fp16, x = input_77_cast_fp16)[name = tensor("linear_25_cast_fp16")]; + tensor var_2534 = const()[name = tensor("op_2534"), val = tensor([1, 1, 3, 16, 64])]; + tensor qkv_13_cast_fp16 = reshape(shape = var_2534, x = linear_25_cast_fp16)[name = tensor("qkv_13_cast_fp16")]; + tensor q_37_begin_0 = const()[name = tensor("q_37_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor q_37_end_0 = const()[name = tensor("q_37_end_0"), val = tensor([1, 1, 1, 16, 64])]; + tensor q_37_end_mask_0 = const()[name = tensor("q_37_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor q_37_squeeze_mask_0 = const()[name = tensor("q_37_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor q_37_cast_fp16 = slice_by_index(begin = q_37_begin_0, end = q_37_end_0, end_mask = q_37_end_mask_0, squeeze_mask = q_37_squeeze_mask_0, x = qkv_13_cast_fp16)[name = tensor("q_37_cast_fp16")]; + tensor k_25_begin_0 = const()[name = tensor("k_25_begin_0"), val = tensor([0, 0, 1, 0, 0])]; + tensor k_25_end_0 = const()[name = tensor("k_25_end_0"), val = tensor([1, 1, 2, 16, 64])]; + tensor k_25_end_mask_0 = const()[name = tensor("k_25_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor k_25_squeeze_mask_0 = const()[name = tensor("k_25_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor k_25_cast_fp16 = slice_by_index(begin = k_25_begin_0, end = k_25_end_0, end_mask = k_25_end_mask_0, squeeze_mask = k_25_squeeze_mask_0, x = qkv_13_cast_fp16)[name = tensor("k_25_cast_fp16")]; + tensor v_13_begin_0 = const()[name = tensor("v_13_begin_0"), val = tensor([0, 0, 2, 0, 0])]; + tensor v_13_end_0 = const()[name = tensor("v_13_end_0"), val = tensor([1, 1, 3, 16, 64])]; + tensor v_13_end_mask_0 = const()[name = tensor("v_13_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor v_13_squeeze_mask_0 = const()[name = tensor("v_13_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor v_13_cast_fp16 = slice_by_index(begin = v_13_begin_0, end = v_13_end_0, end_mask = v_13_end_mask_0, squeeze_mask = v_13_squeeze_mask_0, x = qkv_13_cast_fp16)[name = tensor("v_13_cast_fp16")]; + tensor var_2600 = const()[name = tensor("op_2600"), val = tensor([1, 1, 1, 1])]; + tensor position6_to_fp16_dtype_0 = const()[name = tensor("position6_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor position6_to_fp16 = cast(dtype = position6_to_fp16_dtype_0, x = position6)[name = tensor("cast_523")]; + tensor ts_41_cast_fp16 = reshape(shape = var_2600, x = position6_to_fp16)[name = tensor("ts_41_cast_fp16")]; + tensor var_2607 = const()[name = tensor("op_2607"), val = tensor([1, 1, 16, 32, 2])]; + tensor q_complex_13_cast_fp16 = reshape(shape = var_2607, x = q_37_cast_fp16)[name = tensor("q_complex_13_cast_fp16")]; + tensor var_2614 = const()[name = tensor("op_2614"), val = tensor([1, 1, 16, 32, 2])]; + tensor k_complex_13_cast_fp16 = reshape(shape = var_2614, x = k_25_cast_fp16)[name = tensor("k_complex_13_cast_fp16")]; + tensor var_2618_begin_0 = const()[name = tensor("op_2618_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_2618_end_0 = const()[name = tensor("op_2618_end_0"), val = tensor([1, 1, 16, 32, 1])]; + tensor var_2618_end_mask_0 = const()[name = tensor("op_2618_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_2618_squeeze_mask_0 = const()[name = tensor("op_2618_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_2618_cast_fp16 = slice_by_index(begin = var_2618_begin_0, end = var_2618_end_0, end_mask = var_2618_end_mask_0, squeeze_mask = var_2618_squeeze_mask_0, x = q_complex_13_cast_fp16)[name = tensor("op_2618_cast_fp16")]; + tensor var_2626_begin_0 = const()[name = tensor("op_2626_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_2626_end_0 = const()[name = tensor("op_2626_end_0"), val = tensor([1, 1, 16, 32, 2])]; + tensor var_2626_end_mask_0 = const()[name = tensor("op_2626_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_2626_squeeze_mask_0 = const()[name = tensor("op_2626_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_2626_cast_fp16 = slice_by_index(begin = var_2626_begin_0, end = var_2626_end_0, end_mask = var_2626_end_mask_0, squeeze_mask = var_2626_squeeze_mask_0, x = q_complex_13_cast_fp16)[name = tensor("op_2626_cast_fp16")]; + tensor var_2634_begin_0 = const()[name = tensor("op_2634_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_2634_end_0 = const()[name = tensor("op_2634_end_0"), val = tensor([1, 1, 16, 32, 1])]; + tensor var_2634_end_mask_0 = const()[name = tensor("op_2634_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_2634_squeeze_mask_0 = const()[name = tensor("op_2634_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_2634_cast_fp16 = slice_by_index(begin = var_2634_begin_0, end = var_2634_end_0, end_mask = var_2634_end_mask_0, squeeze_mask = var_2634_squeeze_mask_0, x = k_complex_13_cast_fp16)[name = tensor("op_2634_cast_fp16")]; + tensor var_2642_begin_0 = const()[name = tensor("op_2642_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_2642_end_0 = const()[name = tensor("op_2642_end_0"), val = tensor([1, 1, 16, 32, 2])]; + tensor var_2642_end_mask_0 = const()[name = tensor("op_2642_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_2642_squeeze_mask_0 = const()[name = tensor("op_2642_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_2642_cast_fp16 = slice_by_index(begin = var_2642_begin_0, end = var_2642_end_0, end_mask = var_2642_end_mask_0, squeeze_mask = var_2642_squeeze_mask_0, x = k_complex_13_cast_fp16)[name = tensor("op_2642_cast_fp16")]; + tensor freqs_13_to_fp16 = const()[name = tensor("freqs_13_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(158475648)))]; + tensor var_2648_cast_fp16 = mul(x = freqs_13_to_fp16, y = ts_41_cast_fp16)[name = tensor("op_2648_cast_fp16")]; + tensor rotr_13_cast_fp16 = cos(x = var_2648_cast_fp16)[name = tensor("rotr_13_cast_fp16")]; + tensor roti_13_cast_fp16 = sin(x = var_2648_cast_fp16)[name = tensor("roti_13_cast_fp16")]; + tensor var_2652_cast_fp16 = mul(x = var_2618_cast_fp16, y = rotr_13_cast_fp16)[name = tensor("op_2652_cast_fp16")]; + tensor var_2653_cast_fp16 = mul(x = var_2626_cast_fp16, y = roti_13_cast_fp16)[name = tensor("op_2653_cast_fp16")]; + tensor qor_25_cast_fp16 = sub(x = var_2652_cast_fp16, y = var_2653_cast_fp16)[name = tensor("qor_25_cast_fp16")]; + tensor var_2656_cast_fp16 = mul(x = var_2618_cast_fp16, y = roti_13_cast_fp16)[name = tensor("op_2656_cast_fp16")]; + tensor var_2657_cast_fp16 = mul(x = var_2626_cast_fp16, y = rotr_13_cast_fp16)[name = tensor("op_2657_cast_fp16")]; + tensor qoi_25_cast_fp16 = add(x = var_2656_cast_fp16, y = var_2657_cast_fp16)[name = tensor("qoi_25_cast_fp16")]; + tensor var_2660_cast_fp16 = mul(x = var_2634_cast_fp16, y = rotr_13_cast_fp16)[name = tensor("op_2660_cast_fp16")]; + tensor var_2661_cast_fp16 = mul(x = var_2642_cast_fp16, y = roti_13_cast_fp16)[name = tensor("op_2661_cast_fp16")]; + tensor kor_25_cast_fp16 = sub(x = var_2660_cast_fp16, y = var_2661_cast_fp16)[name = tensor("kor_25_cast_fp16")]; + tensor var_2664_cast_fp16 = mul(x = var_2634_cast_fp16, y = roti_13_cast_fp16)[name = tensor("op_2664_cast_fp16")]; + tensor var_2665_cast_fp16 = mul(x = var_2642_cast_fp16, y = rotr_13_cast_fp16)[name = tensor("op_2665_cast_fp16")]; + tensor koi_25_cast_fp16 = add(x = var_2664_cast_fp16, y = var_2665_cast_fp16)[name = tensor("koi_25_cast_fp16")]; + tensor qo_13_axis_0 = const()[name = tensor("qo_13_axis_0"), val = tensor(-1)]; + tensor qo_13_cast_fp16 = stack(axis = qo_13_axis_0, values = (qor_25_cast_fp16, qoi_25_cast_fp16))[name = tensor("qo_13_cast_fp16")]; + tensor ko_13_axis_0 = const()[name = tensor("ko_13_axis_0"), val = tensor(-1)]; + tensor ko_13_cast_fp16 = stack(axis = ko_13_axis_0, values = (kor_25_cast_fp16, koi_25_cast_fp16))[name = tensor("ko_13_cast_fp16")]; + tensor var_2698 = const()[name = tensor("op_2698"), val = tensor([1, 1, 16, 64])]; + tensor q_39_cast_fp16 = reshape(shape = var_2698, x = qo_13_cast_fp16)[name = tensor("q_39_cast_fp16")]; + tensor var_2704 = const()[name = tensor("op_2704"), val = tensor([1, 1, 16, 64])]; + tensor k_27_cast_fp16 = reshape(shape = var_2704, x = ko_13_cast_fp16)[name = tensor("k_27_cast_fp16")]; + tensor _inversed_2729_y_0_to_fp16 = const()[name = tensor("_inversed_2729_y_0_to_fp16"), val = tensor(0x1p-9)]; + tensor _inversed_2729_cast_fp16 = mul(x = ts_41_cast_fp16, y = _inversed_2729_y_0_to_fp16)[name = tensor("_inversed_2729_cast_fp16")]; + tensor var_2730_cast_fp16 = floor(x = _inversed_2729_cast_fp16)[name = tensor("op_2730_cast_fp16")]; + tensor var_2731_to_fp16 = const()[name = tensor("op_2731_to_fp16"), val = tensor(0x1p+9)]; + tensor var_2732_cast_fp16 = mul(x = var_2730_cast_fp16, y = var_2731_to_fp16)[name = tensor("op_2732_cast_fp16")]; + tensor write_indices_float_27_cast_fp16 = sub(x = ts_41_cast_fp16, y = var_2732_cast_fp16)[name = tensor("write_indices_float_27_cast_fp16")]; + tensor var_2739_dtype_0 = const()[name = tensor("op_2739_dtype_0"), val = tensor("int32")]; + tensor write_indices_13_reps_0 = const()[name = tensor("write_indices_13_reps_0"), val = tensor([1, 1, 16, 64])]; + tensor write_indices_float_27_cast_fp16_to_int32 = cast(dtype = var_2739_dtype_0, x = write_indices_float_27_cast_fp16)[name = tensor("cast_522")]; + tensor write_indices_13 = tile(reps = write_indices_13_reps_0, x = write_indices_float_27_cast_fp16_to_int32)[name = tensor("write_indices_13")]; + tensor var_2749_begin_0 = const()[name = tensor("op_2749_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_2749_end_0 = const()[name = tensor("op_2749_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor var_2749_end_mask_0 = const()[name = tensor("op_2749_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_2749_squeeze_mask_0 = const()[name = tensor("op_2749_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor cache6_to_fp16_dtype_0 = const()[name = tensor("cache6_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor cache6_to_fp16 = cast(dtype = cache6_to_fp16_dtype_0, x = cache6)[name = tensor("cast_521")]; + tensor var_2749_cast_fp16 = slice_by_index(begin = var_2749_begin_0, end = var_2749_end_0, end_mask = var_2749_end_mask_0, squeeze_mask = var_2749_squeeze_mask_0, x = cache6_to_fp16)[name = tensor("op_2749_cast_fp16")]; + tensor var_2751_axis_0 = const()[name = tensor("op_2751_axis_0"), val = tensor(1)]; + tensor var_2751_mode_0 = const()[name = tensor("op_2751_mode_0"), val = tensor("update")]; + tensor var_2751_validate_indices_0 = const()[name = tensor("op_2751_validate_indices_0"), val = tensor(false)]; + tensor var_2751_cast_fp16 = scatter_along_axis(axis = var_2751_axis_0, data = var_2749_cast_fp16, indices = write_indices_13, mode = var_2751_mode_0, updates = k_27_cast_fp16, validate_indices = var_2751_validate_indices_0)[name = tensor("op_2751_cast_fp16")]; + tensor concat_44 = const()[name = tensor("concat_44"), val = tensor([0, 0, 0, 0, 0])]; + tensor concat_45 = const()[name = tensor("concat_45"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_13_internal_tensor_assign_1_stride_0 = const()[name = tensor("new_cache_13_internal_tensor_assign_1_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_13_internal_tensor_assign_1_begin_mask_0 = const()[name = tensor("new_cache_13_internal_tensor_assign_1_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_13_internal_tensor_assign_1_end_mask_0 = const()[name = tensor("new_cache_13_internal_tensor_assign_1_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_13_internal_tensor_assign_1_squeeze_mask_0 = const()[name = tensor("new_cache_13_internal_tensor_assign_1_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_60 = const()[name = tensor("shape_60"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_12 = const()[name = tensor("reduce_prod_12"), val = tensor(1048576)]; + tensor range_1d_12_start_0 = const()[name = tensor("range_1d_12_start_0"), val = tensor(0)]; + tensor range_1d_12_step_0 = const()[name = tensor("range_1d_12_step_0"), val = tensor(1)]; + tensor range_1d_12 = range_1d(end = reduce_prod_12, start = range_1d_12_start_0, step = range_1d_12_step_0)[name = tensor("range_1d_12")]; + tensor reshape_60 = reshape(shape = shape_60, x = range_1d_12)[name = tensor("reshape_60")]; + tensor slice_by_index_12 = slice_by_index(begin = concat_44, begin_mask = new_cache_13_internal_tensor_assign_1_begin_mask_0, end = concat_45, end_mask = new_cache_13_internal_tensor_assign_1_end_mask_0, squeeze_mask = new_cache_13_internal_tensor_assign_1_squeeze_mask_0, stride = new_cache_13_internal_tensor_assign_1_stride_0, x = reshape_60)[name = tensor("slice_by_index_12")]; + tensor reshape_61_shape_0 = const()[name = tensor("reshape_61_shape_0"), val = tensor([-1])]; + tensor reshape_61 = reshape(shape = reshape_61_shape_0, x = slice_by_index_12)[name = tensor("reshape_61")]; + tensor reshape_62_shape_0 = const()[name = tensor("reshape_62_shape_0"), val = tensor([-1])]; + tensor reshape_62_cast_fp16 = reshape(shape = reshape_62_shape_0, x = var_2751_cast_fp16)[name = tensor("reshape_62_cast_fp16")]; + tensor reshape_63_shape_0 = const()[name = tensor("reshape_63_shape_0"), val = tensor([-1])]; + tensor reshape_63_cast_fp16 = reshape(shape = reshape_63_shape_0, x = cache6_to_fp16)[name = tensor("reshape_63_cast_fp16")]; + tensor scatter_12_mode_0 = const()[name = tensor("scatter_12_mode_0"), val = tensor("update")]; + tensor scatter_12_axis_0 = const()[name = tensor("scatter_12_axis_0"), val = tensor(0)]; + tensor scatter_12_validate_indices_0 = const()[name = tensor("scatter_12_validate_indices_0"), val = tensor(false)]; + tensor scatter_12_cast_fp16 = scatter(axis = scatter_12_axis_0, data = reshape_63_cast_fp16, indices = reshape_61, mode = scatter_12_mode_0, updates = reshape_62_cast_fp16, validate_indices = scatter_12_validate_indices_0)[name = tensor("scatter_12_cast_fp16")]; + tensor reshape_64_cast_fp16 = reshape(shape = shape_60, x = scatter_12_cast_fp16)[name = tensor("reshape_64_cast_fp16")]; + tensor var_2759_begin_0 = const()[name = tensor("op_2759_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor var_2759_end_0 = const()[name = tensor("op_2759_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor var_2759_end_mask_0 = const()[name = tensor("op_2759_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_2759_squeeze_mask_0 = const()[name = tensor("op_2759_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor var_2759_cast_fp16 = slice_by_index(begin = var_2759_begin_0, end = var_2759_end_0, end_mask = var_2759_end_mask_0, squeeze_mask = var_2759_squeeze_mask_0, x = reshape_64_cast_fp16)[name = tensor("op_2759_cast_fp16")]; + tensor var_2761_axis_0 = const()[name = tensor("op_2761_axis_0"), val = tensor(1)]; + tensor var_2761_mode_0 = const()[name = tensor("op_2761_mode_0"), val = tensor("update")]; + tensor var_2761_validate_indices_0 = const()[name = tensor("op_2761_validate_indices_0"), val = tensor(false)]; + tensor var_2761_cast_fp16 = scatter_along_axis(axis = var_2761_axis_0, data = var_2759_cast_fp16, indices = write_indices_13, mode = var_2761_mode_0, updates = v_13_cast_fp16, validate_indices = var_2761_validate_indices_0)[name = tensor("op_2761_cast_fp16")]; + tensor concat_46 = const()[name = tensor("concat_46"), val = tensor([1, 0, 0, 0, 0])]; + tensor concat_47 = const()[name = tensor("concat_47"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_13_internal_tensor_assign_2_stride_0 = const()[name = tensor("new_cache_13_internal_tensor_assign_2_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_13_internal_tensor_assign_2_begin_mask_0 = const()[name = tensor("new_cache_13_internal_tensor_assign_2_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_13_internal_tensor_assign_2_end_mask_0 = const()[name = tensor("new_cache_13_internal_tensor_assign_2_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_13_internal_tensor_assign_2_squeeze_mask_0 = const()[name = tensor("new_cache_13_internal_tensor_assign_2_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_61 = const()[name = tensor("shape_61"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_13 = const()[name = tensor("reduce_prod_13"), val = tensor(1048576)]; + tensor range_1d_13_start_0 = const()[name = tensor("range_1d_13_start_0"), val = tensor(0)]; + tensor range_1d_13_step_0 = const()[name = tensor("range_1d_13_step_0"), val = tensor(1)]; + tensor range_1d_13 = range_1d(end = reduce_prod_13, start = range_1d_13_start_0, step = range_1d_13_step_0)[name = tensor("range_1d_13")]; + tensor reshape_65 = reshape(shape = shape_61, x = range_1d_13)[name = tensor("reshape_65")]; + tensor slice_by_index_13 = slice_by_index(begin = concat_46, begin_mask = new_cache_13_internal_tensor_assign_2_begin_mask_0, end = concat_47, end_mask = new_cache_13_internal_tensor_assign_2_end_mask_0, squeeze_mask = new_cache_13_internal_tensor_assign_2_squeeze_mask_0, stride = new_cache_13_internal_tensor_assign_2_stride_0, x = reshape_65)[name = tensor("slice_by_index_13")]; + tensor reshape_66_shape_0 = const()[name = tensor("reshape_66_shape_0"), val = tensor([-1])]; + tensor reshape_66 = reshape(shape = reshape_66_shape_0, x = slice_by_index_13)[name = tensor("reshape_66")]; + tensor reshape_67_shape_0 = const()[name = tensor("reshape_67_shape_0"), val = tensor([-1])]; + tensor reshape_67_cast_fp16 = reshape(shape = reshape_67_shape_0, x = var_2761_cast_fp16)[name = tensor("reshape_67_cast_fp16")]; + tensor reshape_68_shape_0 = const()[name = tensor("reshape_68_shape_0"), val = tensor([-1])]; + tensor reshape_68_cast_fp16 = reshape(shape = reshape_68_shape_0, x = reshape_64_cast_fp16)[name = tensor("reshape_68_cast_fp16")]; + tensor scatter_13_mode_0 = const()[name = tensor("scatter_13_mode_0"), val = tensor("update")]; + tensor scatter_13_axis_0 = const()[name = tensor("scatter_13_axis_0"), val = tensor(0)]; + tensor scatter_13_validate_indices_0 = const()[name = tensor("scatter_13_validate_indices_0"), val = tensor(false)]; + tensor scatter_13_cast_fp16 = scatter(axis = scatter_13_axis_0, data = reshape_68_cast_fp16, indices = reshape_66, mode = scatter_13_mode_0, updates = reshape_67_cast_fp16, validate_indices = scatter_13_validate_indices_0)[name = tensor("scatter_13_cast_fp16")]; + tensor reshape_69_cast_fp16 = reshape(shape = shape_61, x = scatter_13_cast_fp16)[name = tensor("reshape_69_cast_fp16")]; + tensor reshape_69_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("reshape_69_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor keys_37_begin_0 = const()[name = tensor("keys_37_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor keys_37_end_0 = const()[name = tensor("keys_37_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor keys_37_end_mask_0 = const()[name = tensor("keys_37_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor keys_37_squeeze_mask_0 = const()[name = tensor("keys_37_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor keys_37_cast_fp16 = slice_by_index(begin = keys_37_begin_0, end = keys_37_end_0, end_mask = keys_37_end_mask_0, squeeze_mask = keys_37_squeeze_mask_0, x = reshape_69_cast_fp16)[name = tensor("keys_37_cast_fp16")]; + tensor values_37_begin_0 = const()[name = tensor("values_37_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor values_37_end_0 = const()[name = tensor("values_37_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor values_37_end_mask_0 = const()[name = tensor("values_37_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor values_37_squeeze_mask_0 = const()[name = tensor("values_37_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor values_37_cast_fp16 = slice_by_index(begin = values_37_begin_0, end = values_37_end_0, end_mask = values_37_end_mask_0, squeeze_mask = values_37_squeeze_mask_0, x = reshape_69_cast_fp16)[name = tensor("values_37_cast_fp16")]; + tensor var_2773_cast_fp16 = not_equal(x = keys_37_cast_fp16, y = keys_37_cast_fp16)[name = tensor("op_2773_cast_fp16")]; + tensor keys_39_cast_fp16 = select(a = var_463_to_fp16, b = keys_37_cast_fp16, cond = var_2773_cast_fp16)[name = tensor("keys_39_cast_fp16")]; + tensor var_2781_cast_fp16 = not_equal(x = values_37_cast_fp16, y = values_37_cast_fp16)[name = tensor("op_2781_cast_fp16")]; + tensor values_39_cast_fp16 = select(a = var_463_to_fp16, b = values_37_cast_fp16, cond = var_2781_cast_fp16)[name = tensor("values_39_cast_fp16")]; + tensor var_2805 = const()[name = tensor("op_2805"), val = tensor([0, 2, 1, 3])]; + tensor var_2821 = const()[name = tensor("op_2821"), val = tensor([1, 1, 1])]; + tensor var_2822_cast_fp16 = reshape(shape = var_2821, x = position6_to_fp16)[name = tensor("op_2822_cast_fp16")]; + tensor var_2842_to_fp16 = const()[name = tensor("op_2842_to_fp16"), val = tensor(0x1p+0)]; + tensor valid_len_13_cast_fp16 = add(x = var_2822_cast_fp16, y = var_2842_to_fp16)[name = tensor("valid_len_13_cast_fp16")]; + tensor valid_mask_13_cast_fp16 = less(x = k_positions_1_promoted_to_fp16, y = valid_len_13_cast_fp16)[name = tensor("valid_mask_13_cast_fp16")]; + tensor causal_mask_13_cast_fp16 = less_equal(x = k_positions_1_promoted_to_fp16, y = var_2822_cast_fp16)[name = tensor("causal_mask_13_cast_fp16")]; + tensor attn_mask_25 = logical_and(x = valid_mask_13_cast_fp16, y = causal_mask_13_cast_fp16)[name = tensor("attn_mask_25")]; + tensor attn_mask_27_axes_0 = const()[name = tensor("attn_mask_27_axes_0"), val = tensor([1])]; + tensor attn_mask_27 = expand_dims(axes = attn_mask_27_axes_0, x = attn_mask_25)[name = tensor("attn_mask_27")]; + tensor var_2860_transpose_x_0 = const()[name = tensor("op_2860_transpose_x_0"), val = tensor(false)]; + tensor var_2860_transpose_y_0 = const()[name = tensor("op_2860_transpose_y_0"), val = tensor(false)]; + tensor transpose_84_perm_0 = const()[name = tensor("transpose_84_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_85_perm_0 = const()[name = tensor("transpose_85_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_85 = transpose(perm = transpose_85_perm_0, x = keys_39_cast_fp16)[name = tensor("transpose_189")]; + tensor transpose_84 = transpose(perm = transpose_84_perm_0, x = q_39_cast_fp16)[name = tensor("transpose_190")]; + tensor var_2860_cast_fp16 = matmul(transpose_x = var_2860_transpose_x_0, transpose_y = var_2860_transpose_y_0, x = transpose_84, y = transpose_85)[name = tensor("op_2860_cast_fp16")]; + tensor var_2854_to_fp16 = const()[name = tensor("op_2854_to_fp16"), val = tensor([0x1p-3])]; + tensor attn_weights_37_cast_fp16 = mul(x = var_2860_cast_fp16, y = var_2854_to_fp16)[name = tensor("attn_weights_37_cast_fp16")]; + tensor var_2862 = logical_not(x = attn_mask_27)[name = tensor("op_2862")]; + tensor var_2863_to_fp16 = const()[name = tensor("op_2863_to_fp16"), val = tensor(-inf)]; + tensor attn_weights_39_cast_fp16 = select(a = var_2863_to_fp16, b = attn_weights_37_cast_fp16, cond = var_2862)[name = tensor("attn_weights_39_cast_fp16")]; + tensor var_2865 = const()[name = tensor("op_2865"), val = tensor(-1)]; + tensor attn_weights_41_cast_fp16 = softmax(axis = var_2865, x = attn_weights_39_cast_fp16)[name = tensor("attn_weights_41_cast_fp16")]; + tensor attn_output_13_transpose_x_0 = const()[name = tensor("attn_output_13_transpose_x_0"), val = tensor(false)]; + tensor attn_output_13_transpose_y_0 = const()[name = tensor("attn_output_13_transpose_y_0"), val = tensor(false)]; + tensor values_41_cast_fp16 = transpose(perm = var_2805, x = values_39_cast_fp16)[name = tensor("transpose_191")]; + tensor attn_output_13_cast_fp16 = matmul(transpose_x = attn_output_13_transpose_x_0, transpose_y = attn_output_13_transpose_y_0, x = attn_weights_41_cast_fp16, y = values_41_cast_fp16)[name = tensor("attn_output_13_cast_fp16")]; + tensor var_2873 = const()[name = tensor("op_2873"), val = tensor([0, 2, 1, 3])]; + tensor var_2878 = const()[name = tensor("op_2878"), val = tensor([1, 1, 1024])]; + tensor var_2874_cast_fp16 = transpose(perm = var_2873, x = attn_output_13_cast_fp16)[name = tensor("transpose_188")]; + tensor input_79_cast_fp16 = reshape(shape = var_2878, x = var_2874_cast_fp16)[name = tensor("input_79_cast_fp16")]; + tensor attn6_out_proj_weight_to_fp16 = const()[name = tensor("attn6_out_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(158475776)))]; + tensor linear_26_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = attn6_out_proj_weight_to_fp16, x = input_79_cast_fp16)[name = tensor("linear_26_cast_fp16")]; + tensor var_2884_to_fp16 = const()[name = tensor("op_2884_to_fp16"), val = tensor(0x1p+0)]; + tensor var_2885_cast_fp16 = add(x = position6_to_fp16, y = var_2884_to_fp16)[name = tensor("op_2885_cast_fp16")]; + tensor var_2885_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_2885_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor input_81_cast_fp16 = add(x = input_75_cast_fp16, y = linear_26_cast_fp16)[name = tensor("input_81_cast_fp16")]; + tensor input_83_axes_0 = const()[name = tensor("input_83_axes_0"), val = tensor([-1])]; + tensor norm6_2_weight_to_fp16 = const()[name = tensor("norm6_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(160572992)))]; + tensor norm6_2_bias_to_fp16 = const()[name = tensor("norm6_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(160575104)))]; + tensor var_2889_to_fp16 = const()[name = tensor("op_2889_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_83_cast_fp16 = layer_norm(axes = input_83_axes_0, beta = norm6_2_bias_to_fp16, epsilon = var_2889_to_fp16, gamma = norm6_2_weight_to_fp16, x = input_81_cast_fp16)[name = tensor("input_83_cast_fp16")]; + tensor linear6_1_weight_to_fp16 = const()[name = tensor("linear6_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(160577216)))]; + tensor linear_27_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = linear6_1_weight_to_fp16, x = input_83_cast_fp16)[name = tensor("linear_27_cast_fp16")]; + tensor input_85_mode_0 = const()[name = tensor("input_85_mode_0"), val = tensor("EXACT")]; + tensor input_85_cast_fp16 = gelu(mode = input_85_mode_0, x = linear_27_cast_fp16)[name = tensor("input_85_cast_fp16")]; + tensor linear6_2_weight_to_fp16 = const()[name = tensor("linear6_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(168965888)))]; + tensor linear_28_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = linear6_2_weight_to_fp16, x = input_85_cast_fp16)[name = tensor("linear_28_cast_fp16")]; + tensor input_87_cast_fp16 = add(x = input_81_cast_fp16, y = linear_28_cast_fp16)[name = tensor("input_87_cast_fp16")]; + tensor input_89_axes_0 = const()[name = tensor("input_89_axes_0"), val = tensor([-1])]; + tensor norm7_1_weight_to_fp16 = const()[name = tensor("norm7_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(177354560)))]; + tensor norm7_1_bias_to_fp16 = const()[name = tensor("norm7_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(177356672)))]; + tensor var_2906_to_fp16 = const()[name = tensor("op_2906_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_89_cast_fp16 = layer_norm(axes = input_89_axes_0, beta = norm7_1_bias_to_fp16, epsilon = var_2906_to_fp16, gamma = norm7_1_weight_to_fp16, x = input_87_cast_fp16)[name = tensor("input_89_cast_fp16")]; + tensor attn7_in_proj_weight_to_fp16 = const()[name = tensor("attn7_in_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(177358784)))]; + tensor linear_29_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = attn7_in_proj_weight_to_fp16, x = input_89_cast_fp16)[name = tensor("linear_29_cast_fp16")]; + tensor var_2920 = const()[name = tensor("op_2920"), val = tensor([1, 1, 3, 16, 64])]; + tensor qkv_15_cast_fp16 = reshape(shape = var_2920, x = linear_29_cast_fp16)[name = tensor("qkv_15_cast_fp16")]; + tensor q_43_begin_0 = const()[name = tensor("q_43_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor q_43_end_0 = const()[name = tensor("q_43_end_0"), val = tensor([1, 1, 1, 16, 64])]; + tensor q_43_end_mask_0 = const()[name = tensor("q_43_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor q_43_squeeze_mask_0 = const()[name = tensor("q_43_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor q_43_cast_fp16 = slice_by_index(begin = q_43_begin_0, end = q_43_end_0, end_mask = q_43_end_mask_0, squeeze_mask = q_43_squeeze_mask_0, x = qkv_15_cast_fp16)[name = tensor("q_43_cast_fp16")]; + tensor k_29_begin_0 = const()[name = tensor("k_29_begin_0"), val = tensor([0, 0, 1, 0, 0])]; + tensor k_29_end_0 = const()[name = tensor("k_29_end_0"), val = tensor([1, 1, 2, 16, 64])]; + tensor k_29_end_mask_0 = const()[name = tensor("k_29_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor k_29_squeeze_mask_0 = const()[name = tensor("k_29_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor k_29_cast_fp16 = slice_by_index(begin = k_29_begin_0, end = k_29_end_0, end_mask = k_29_end_mask_0, squeeze_mask = k_29_squeeze_mask_0, x = qkv_15_cast_fp16)[name = tensor("k_29_cast_fp16")]; + tensor v_15_begin_0 = const()[name = tensor("v_15_begin_0"), val = tensor([0, 0, 2, 0, 0])]; + tensor v_15_end_0 = const()[name = tensor("v_15_end_0"), val = tensor([1, 1, 3, 16, 64])]; + tensor v_15_end_mask_0 = const()[name = tensor("v_15_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor v_15_squeeze_mask_0 = const()[name = tensor("v_15_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor v_15_cast_fp16 = slice_by_index(begin = v_15_begin_0, end = v_15_end_0, end_mask = v_15_end_mask_0, squeeze_mask = v_15_squeeze_mask_0, x = qkv_15_cast_fp16)[name = tensor("v_15_cast_fp16")]; + tensor var_2986 = const()[name = tensor("op_2986"), val = tensor([1, 1, 1, 1])]; + tensor position7_to_fp16_dtype_0 = const()[name = tensor("position7_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor position7_to_fp16 = cast(dtype = position7_to_fp16_dtype_0, x = position7)[name = tensor("cast_518")]; + tensor ts_47_cast_fp16 = reshape(shape = var_2986, x = position7_to_fp16)[name = tensor("ts_47_cast_fp16")]; + tensor var_2993 = const()[name = tensor("op_2993"), val = tensor([1, 1, 16, 32, 2])]; + tensor q_complex_15_cast_fp16 = reshape(shape = var_2993, x = q_43_cast_fp16)[name = tensor("q_complex_15_cast_fp16")]; + tensor var_3000 = const()[name = tensor("op_3000"), val = tensor([1, 1, 16, 32, 2])]; + tensor k_complex_15_cast_fp16 = reshape(shape = var_3000, x = k_29_cast_fp16)[name = tensor("k_complex_15_cast_fp16")]; + tensor var_3004_begin_0 = const()[name = tensor("op_3004_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_3004_end_0 = const()[name = tensor("op_3004_end_0"), val = tensor([1, 1, 16, 32, 1])]; + tensor var_3004_end_mask_0 = const()[name = tensor("op_3004_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_3004_squeeze_mask_0 = const()[name = tensor("op_3004_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_3004_cast_fp16 = slice_by_index(begin = var_3004_begin_0, end = var_3004_end_0, end_mask = var_3004_end_mask_0, squeeze_mask = var_3004_squeeze_mask_0, x = q_complex_15_cast_fp16)[name = tensor("op_3004_cast_fp16")]; + tensor var_3012_begin_0 = const()[name = tensor("op_3012_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_3012_end_0 = const()[name = tensor("op_3012_end_0"), val = tensor([1, 1, 16, 32, 2])]; + tensor var_3012_end_mask_0 = const()[name = tensor("op_3012_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_3012_squeeze_mask_0 = const()[name = tensor("op_3012_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_3012_cast_fp16 = slice_by_index(begin = var_3012_begin_0, end = var_3012_end_0, end_mask = var_3012_end_mask_0, squeeze_mask = var_3012_squeeze_mask_0, x = q_complex_15_cast_fp16)[name = tensor("op_3012_cast_fp16")]; + tensor var_3020_begin_0 = const()[name = tensor("op_3020_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_3020_end_0 = const()[name = tensor("op_3020_end_0"), val = tensor([1, 1, 16, 32, 1])]; + tensor var_3020_end_mask_0 = const()[name = tensor("op_3020_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_3020_squeeze_mask_0 = const()[name = tensor("op_3020_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_3020_cast_fp16 = slice_by_index(begin = var_3020_begin_0, end = var_3020_end_0, end_mask = var_3020_end_mask_0, squeeze_mask = var_3020_squeeze_mask_0, x = k_complex_15_cast_fp16)[name = tensor("op_3020_cast_fp16")]; + tensor var_3028_begin_0 = const()[name = tensor("op_3028_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_3028_end_0 = const()[name = tensor("op_3028_end_0"), val = tensor([1, 1, 16, 32, 2])]; + tensor var_3028_end_mask_0 = const()[name = tensor("op_3028_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_3028_squeeze_mask_0 = const()[name = tensor("op_3028_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_3028_cast_fp16 = slice_by_index(begin = var_3028_begin_0, end = var_3028_end_0, end_mask = var_3028_end_mask_0, squeeze_mask = var_3028_squeeze_mask_0, x = k_complex_15_cast_fp16)[name = tensor("op_3028_cast_fp16")]; + tensor freqs_15_to_fp16 = const()[name = tensor("freqs_15_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(183650304)))]; + tensor var_3034_cast_fp16 = mul(x = freqs_15_to_fp16, y = ts_47_cast_fp16)[name = tensor("op_3034_cast_fp16")]; + tensor rotr_15_cast_fp16 = cos(x = var_3034_cast_fp16)[name = tensor("rotr_15_cast_fp16")]; + tensor roti_15_cast_fp16 = sin(x = var_3034_cast_fp16)[name = tensor("roti_15_cast_fp16")]; + tensor var_3038_cast_fp16 = mul(x = var_3004_cast_fp16, y = rotr_15_cast_fp16)[name = tensor("op_3038_cast_fp16")]; + tensor var_3039_cast_fp16 = mul(x = var_3012_cast_fp16, y = roti_15_cast_fp16)[name = tensor("op_3039_cast_fp16")]; + tensor qor_29_cast_fp16 = sub(x = var_3038_cast_fp16, y = var_3039_cast_fp16)[name = tensor("qor_29_cast_fp16")]; + tensor var_3042_cast_fp16 = mul(x = var_3004_cast_fp16, y = roti_15_cast_fp16)[name = tensor("op_3042_cast_fp16")]; + tensor var_3043_cast_fp16 = mul(x = var_3012_cast_fp16, y = rotr_15_cast_fp16)[name = tensor("op_3043_cast_fp16")]; + tensor qoi_29_cast_fp16 = add(x = var_3042_cast_fp16, y = var_3043_cast_fp16)[name = tensor("qoi_29_cast_fp16")]; + tensor var_3046_cast_fp16 = mul(x = var_3020_cast_fp16, y = rotr_15_cast_fp16)[name = tensor("op_3046_cast_fp16")]; + tensor var_3047_cast_fp16 = mul(x = var_3028_cast_fp16, y = roti_15_cast_fp16)[name = tensor("op_3047_cast_fp16")]; + tensor kor_29_cast_fp16 = sub(x = var_3046_cast_fp16, y = var_3047_cast_fp16)[name = tensor("kor_29_cast_fp16")]; + tensor var_3050_cast_fp16 = mul(x = var_3020_cast_fp16, y = roti_15_cast_fp16)[name = tensor("op_3050_cast_fp16")]; + tensor var_3051_cast_fp16 = mul(x = var_3028_cast_fp16, y = rotr_15_cast_fp16)[name = tensor("op_3051_cast_fp16")]; + tensor koi_29_cast_fp16 = add(x = var_3050_cast_fp16, y = var_3051_cast_fp16)[name = tensor("koi_29_cast_fp16")]; + tensor qo_15_axis_0 = const()[name = tensor("qo_15_axis_0"), val = tensor(-1)]; + tensor qo_15_cast_fp16 = stack(axis = qo_15_axis_0, values = (qor_29_cast_fp16, qoi_29_cast_fp16))[name = tensor("qo_15_cast_fp16")]; + tensor ko_15_axis_0 = const()[name = tensor("ko_15_axis_0"), val = tensor(-1)]; + tensor ko_15_cast_fp16 = stack(axis = ko_15_axis_0, values = (kor_29_cast_fp16, koi_29_cast_fp16))[name = tensor("ko_15_cast_fp16")]; + tensor var_3084 = const()[name = tensor("op_3084"), val = tensor([1, 1, 16, 64])]; + tensor q_45_cast_fp16 = reshape(shape = var_3084, x = qo_15_cast_fp16)[name = tensor("q_45_cast_fp16")]; + tensor var_3090 = const()[name = tensor("op_3090"), val = tensor([1, 1, 16, 64])]; + tensor k_31_cast_fp16 = reshape(shape = var_3090, x = ko_15_cast_fp16)[name = tensor("k_31_cast_fp16")]; + tensor _inversed_3115_y_0_to_fp16 = const()[name = tensor("_inversed_3115_y_0_to_fp16"), val = tensor(0x1p-9)]; + tensor _inversed_3115_cast_fp16 = mul(x = ts_47_cast_fp16, y = _inversed_3115_y_0_to_fp16)[name = tensor("_inversed_3115_cast_fp16")]; + tensor var_3116_cast_fp16 = floor(x = _inversed_3115_cast_fp16)[name = tensor("op_3116_cast_fp16")]; + tensor var_3117_to_fp16 = const()[name = tensor("op_3117_to_fp16"), val = tensor(0x1p+9)]; + tensor var_3118_cast_fp16 = mul(x = var_3116_cast_fp16, y = var_3117_to_fp16)[name = tensor("op_3118_cast_fp16")]; + tensor write_indices_float_31_cast_fp16 = sub(x = ts_47_cast_fp16, y = var_3118_cast_fp16)[name = tensor("write_indices_float_31_cast_fp16")]; + tensor var_3125_dtype_0 = const()[name = tensor("op_3125_dtype_0"), val = tensor("int32")]; + tensor write_indices_15_reps_0 = const()[name = tensor("write_indices_15_reps_0"), val = tensor([1, 1, 16, 64])]; + tensor write_indices_float_31_cast_fp16_to_int32 = cast(dtype = var_3125_dtype_0, x = write_indices_float_31_cast_fp16)[name = tensor("cast_517")]; + tensor write_indices_15 = tile(reps = write_indices_15_reps_0, x = write_indices_float_31_cast_fp16_to_int32)[name = tensor("write_indices_15")]; + tensor var_3135_begin_0 = const()[name = tensor("op_3135_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_3135_end_0 = const()[name = tensor("op_3135_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor var_3135_end_mask_0 = const()[name = tensor("op_3135_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_3135_squeeze_mask_0 = const()[name = tensor("op_3135_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor cache7_to_fp16_dtype_0 = const()[name = tensor("cache7_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor cache7_to_fp16 = cast(dtype = cache7_to_fp16_dtype_0, x = cache7)[name = tensor("cast_516")]; + tensor var_3135_cast_fp16 = slice_by_index(begin = var_3135_begin_0, end = var_3135_end_0, end_mask = var_3135_end_mask_0, squeeze_mask = var_3135_squeeze_mask_0, x = cache7_to_fp16)[name = tensor("op_3135_cast_fp16")]; + tensor var_3137_axis_0 = const()[name = tensor("op_3137_axis_0"), val = tensor(1)]; + tensor var_3137_mode_0 = const()[name = tensor("op_3137_mode_0"), val = tensor("update")]; + tensor var_3137_validate_indices_0 = const()[name = tensor("op_3137_validate_indices_0"), val = tensor(false)]; + tensor var_3137_cast_fp16 = scatter_along_axis(axis = var_3137_axis_0, data = var_3135_cast_fp16, indices = write_indices_15, mode = var_3137_mode_0, updates = k_31_cast_fp16, validate_indices = var_3137_validate_indices_0)[name = tensor("op_3137_cast_fp16")]; + tensor concat_51 = const()[name = tensor("concat_51"), val = tensor([0, 0, 0, 0, 0])]; + tensor concat_52 = const()[name = tensor("concat_52"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_15_internal_tensor_assign_1_stride_0 = const()[name = tensor("new_cache_15_internal_tensor_assign_1_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_15_internal_tensor_assign_1_begin_mask_0 = const()[name = tensor("new_cache_15_internal_tensor_assign_1_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_15_internal_tensor_assign_1_end_mask_0 = const()[name = tensor("new_cache_15_internal_tensor_assign_1_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_15_internal_tensor_assign_1_squeeze_mask_0 = const()[name = tensor("new_cache_15_internal_tensor_assign_1_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_62 = const()[name = tensor("shape_62"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_14 = const()[name = tensor("reduce_prod_14"), val = tensor(1048576)]; + tensor range_1d_14_start_0 = const()[name = tensor("range_1d_14_start_0"), val = tensor(0)]; + tensor range_1d_14_step_0 = const()[name = tensor("range_1d_14_step_0"), val = tensor(1)]; + tensor range_1d_14 = range_1d(end = reduce_prod_14, start = range_1d_14_start_0, step = range_1d_14_step_0)[name = tensor("range_1d_14")]; + tensor reshape_70 = reshape(shape = shape_62, x = range_1d_14)[name = tensor("reshape_70")]; + tensor slice_by_index_14 = slice_by_index(begin = concat_51, begin_mask = new_cache_15_internal_tensor_assign_1_begin_mask_0, end = concat_52, end_mask = new_cache_15_internal_tensor_assign_1_end_mask_0, squeeze_mask = new_cache_15_internal_tensor_assign_1_squeeze_mask_0, stride = new_cache_15_internal_tensor_assign_1_stride_0, x = reshape_70)[name = tensor("slice_by_index_14")]; + tensor reshape_71_shape_0 = const()[name = tensor("reshape_71_shape_0"), val = tensor([-1])]; + tensor reshape_71 = reshape(shape = reshape_71_shape_0, x = slice_by_index_14)[name = tensor("reshape_71")]; + tensor reshape_72_shape_0 = const()[name = tensor("reshape_72_shape_0"), val = tensor([-1])]; + tensor reshape_72_cast_fp16 = reshape(shape = reshape_72_shape_0, x = var_3137_cast_fp16)[name = tensor("reshape_72_cast_fp16")]; + tensor reshape_73_shape_0 = const()[name = tensor("reshape_73_shape_0"), val = tensor([-1])]; + tensor reshape_73_cast_fp16 = reshape(shape = reshape_73_shape_0, x = cache7_to_fp16)[name = tensor("reshape_73_cast_fp16")]; + tensor scatter_14_mode_0 = const()[name = tensor("scatter_14_mode_0"), val = tensor("update")]; + tensor scatter_14_axis_0 = const()[name = tensor("scatter_14_axis_0"), val = tensor(0)]; + tensor scatter_14_validate_indices_0 = const()[name = tensor("scatter_14_validate_indices_0"), val = tensor(false)]; + tensor scatter_14_cast_fp16 = scatter(axis = scatter_14_axis_0, data = reshape_73_cast_fp16, indices = reshape_71, mode = scatter_14_mode_0, updates = reshape_72_cast_fp16, validate_indices = scatter_14_validate_indices_0)[name = tensor("scatter_14_cast_fp16")]; + tensor reshape_74_cast_fp16 = reshape(shape = shape_62, x = scatter_14_cast_fp16)[name = tensor("reshape_74_cast_fp16")]; + tensor var_3145_begin_0 = const()[name = tensor("op_3145_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor var_3145_end_0 = const()[name = tensor("op_3145_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor var_3145_end_mask_0 = const()[name = tensor("op_3145_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_3145_squeeze_mask_0 = const()[name = tensor("op_3145_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor var_3145_cast_fp16 = slice_by_index(begin = var_3145_begin_0, end = var_3145_end_0, end_mask = var_3145_end_mask_0, squeeze_mask = var_3145_squeeze_mask_0, x = reshape_74_cast_fp16)[name = tensor("op_3145_cast_fp16")]; + tensor var_3147_axis_0 = const()[name = tensor("op_3147_axis_0"), val = tensor(1)]; + tensor var_3147_mode_0 = const()[name = tensor("op_3147_mode_0"), val = tensor("update")]; + tensor var_3147_validate_indices_0 = const()[name = tensor("op_3147_validate_indices_0"), val = tensor(false)]; + tensor var_3147_cast_fp16 = scatter_along_axis(axis = var_3147_axis_0, data = var_3145_cast_fp16, indices = write_indices_15, mode = var_3147_mode_0, updates = v_15_cast_fp16, validate_indices = var_3147_validate_indices_0)[name = tensor("op_3147_cast_fp16")]; + tensor concat_53 = const()[name = tensor("concat_53"), val = tensor([1, 0, 0, 0, 0])]; + tensor concat_54 = const()[name = tensor("concat_54"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_15_internal_tensor_assign_2_stride_0 = const()[name = tensor("new_cache_15_internal_tensor_assign_2_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_15_internal_tensor_assign_2_begin_mask_0 = const()[name = tensor("new_cache_15_internal_tensor_assign_2_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_15_internal_tensor_assign_2_end_mask_0 = const()[name = tensor("new_cache_15_internal_tensor_assign_2_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_15_internal_tensor_assign_2_squeeze_mask_0 = const()[name = tensor("new_cache_15_internal_tensor_assign_2_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_63 = const()[name = tensor("shape_63"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_15 = const()[name = tensor("reduce_prod_15"), val = tensor(1048576)]; + tensor range_1d_15_start_0 = const()[name = tensor("range_1d_15_start_0"), val = tensor(0)]; + tensor range_1d_15_step_0 = const()[name = tensor("range_1d_15_step_0"), val = tensor(1)]; + tensor range_1d_15 = range_1d(end = reduce_prod_15, start = range_1d_15_start_0, step = range_1d_15_step_0)[name = tensor("range_1d_15")]; + tensor reshape_75 = reshape(shape = shape_63, x = range_1d_15)[name = tensor("reshape_75")]; + tensor slice_by_index_15 = slice_by_index(begin = concat_53, begin_mask = new_cache_15_internal_tensor_assign_2_begin_mask_0, end = concat_54, end_mask = new_cache_15_internal_tensor_assign_2_end_mask_0, squeeze_mask = new_cache_15_internal_tensor_assign_2_squeeze_mask_0, stride = new_cache_15_internal_tensor_assign_2_stride_0, x = reshape_75)[name = tensor("slice_by_index_15")]; + tensor reshape_76_shape_0 = const()[name = tensor("reshape_76_shape_0"), val = tensor([-1])]; + tensor reshape_76 = reshape(shape = reshape_76_shape_0, x = slice_by_index_15)[name = tensor("reshape_76")]; + tensor reshape_77_shape_0 = const()[name = tensor("reshape_77_shape_0"), val = tensor([-1])]; + tensor reshape_77_cast_fp16 = reshape(shape = reshape_77_shape_0, x = var_3147_cast_fp16)[name = tensor("reshape_77_cast_fp16")]; + tensor reshape_78_shape_0 = const()[name = tensor("reshape_78_shape_0"), val = tensor([-1])]; + tensor reshape_78_cast_fp16 = reshape(shape = reshape_78_shape_0, x = reshape_74_cast_fp16)[name = tensor("reshape_78_cast_fp16")]; + tensor scatter_15_mode_0 = const()[name = tensor("scatter_15_mode_0"), val = tensor("update")]; + tensor scatter_15_axis_0 = const()[name = tensor("scatter_15_axis_0"), val = tensor(0)]; + tensor scatter_15_validate_indices_0 = const()[name = tensor("scatter_15_validate_indices_0"), val = tensor(false)]; + tensor scatter_15_cast_fp16 = scatter(axis = scatter_15_axis_0, data = reshape_78_cast_fp16, indices = reshape_76, mode = scatter_15_mode_0, updates = reshape_77_cast_fp16, validate_indices = scatter_15_validate_indices_0)[name = tensor("scatter_15_cast_fp16")]; + tensor reshape_79_cast_fp16 = reshape(shape = shape_63, x = scatter_15_cast_fp16)[name = tensor("reshape_79_cast_fp16")]; + tensor reshape_79_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("reshape_79_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor keys_43_begin_0 = const()[name = tensor("keys_43_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor keys_43_end_0 = const()[name = tensor("keys_43_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor keys_43_end_mask_0 = const()[name = tensor("keys_43_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor keys_43_squeeze_mask_0 = const()[name = tensor("keys_43_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor keys_43_cast_fp16 = slice_by_index(begin = keys_43_begin_0, end = keys_43_end_0, end_mask = keys_43_end_mask_0, squeeze_mask = keys_43_squeeze_mask_0, x = reshape_79_cast_fp16)[name = tensor("keys_43_cast_fp16")]; + tensor values_43_begin_0 = const()[name = tensor("values_43_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor values_43_end_0 = const()[name = tensor("values_43_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor values_43_end_mask_0 = const()[name = tensor("values_43_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor values_43_squeeze_mask_0 = const()[name = tensor("values_43_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor values_43_cast_fp16 = slice_by_index(begin = values_43_begin_0, end = values_43_end_0, end_mask = values_43_end_mask_0, squeeze_mask = values_43_squeeze_mask_0, x = reshape_79_cast_fp16)[name = tensor("values_43_cast_fp16")]; + tensor var_3159_cast_fp16 = not_equal(x = keys_43_cast_fp16, y = keys_43_cast_fp16)[name = tensor("op_3159_cast_fp16")]; + tensor keys_45_cast_fp16 = select(a = var_463_to_fp16, b = keys_43_cast_fp16, cond = var_3159_cast_fp16)[name = tensor("keys_45_cast_fp16")]; + tensor var_3167_cast_fp16 = not_equal(x = values_43_cast_fp16, y = values_43_cast_fp16)[name = tensor("op_3167_cast_fp16")]; + tensor values_45_cast_fp16 = select(a = var_463_to_fp16, b = values_43_cast_fp16, cond = var_3167_cast_fp16)[name = tensor("values_45_cast_fp16")]; + tensor var_3191 = const()[name = tensor("op_3191"), val = tensor([0, 2, 1, 3])]; + tensor var_3207 = const()[name = tensor("op_3207"), val = tensor([1, 1, 1])]; + tensor var_3208_cast_fp16 = reshape(shape = var_3207, x = position7_to_fp16)[name = tensor("op_3208_cast_fp16")]; + tensor var_3228_to_fp16 = const()[name = tensor("op_3228_to_fp16"), val = tensor(0x1p+0)]; + tensor valid_len_15_cast_fp16 = add(x = var_3208_cast_fp16, y = var_3228_to_fp16)[name = tensor("valid_len_15_cast_fp16")]; + tensor valid_mask_15_cast_fp16 = less(x = k_positions_1_promoted_to_fp16, y = valid_len_15_cast_fp16)[name = tensor("valid_mask_15_cast_fp16")]; + tensor causal_mask_15_cast_fp16 = less_equal(x = k_positions_1_promoted_to_fp16, y = var_3208_cast_fp16)[name = tensor("causal_mask_15_cast_fp16")]; + tensor attn_mask_29 = logical_and(x = valid_mask_15_cast_fp16, y = causal_mask_15_cast_fp16)[name = tensor("attn_mask_29")]; + tensor attn_mask_31_axes_0 = const()[name = tensor("attn_mask_31_axes_0"), val = tensor([1])]; + tensor attn_mask_31 = expand_dims(axes = attn_mask_31_axes_0, x = attn_mask_29)[name = tensor("attn_mask_31")]; + tensor var_3246_transpose_x_0 = const()[name = tensor("op_3246_transpose_x_0"), val = tensor(false)]; + tensor var_3246_transpose_y_0 = const()[name = tensor("op_3246_transpose_y_0"), val = tensor(false)]; + tensor transpose_86_perm_0 = const()[name = tensor("transpose_86_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_87_perm_0 = const()[name = tensor("transpose_87_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_87 = transpose(perm = transpose_87_perm_0, x = keys_45_cast_fp16)[name = tensor("transpose_185")]; + tensor transpose_86 = transpose(perm = transpose_86_perm_0, x = q_45_cast_fp16)[name = tensor("transpose_186")]; + tensor var_3246_cast_fp16 = matmul(transpose_x = var_3246_transpose_x_0, transpose_y = var_3246_transpose_y_0, x = transpose_86, y = transpose_87)[name = tensor("op_3246_cast_fp16")]; + tensor var_3240_to_fp16 = const()[name = tensor("op_3240_to_fp16"), val = tensor([0x1p-3])]; + tensor attn_weights_43_cast_fp16 = mul(x = var_3246_cast_fp16, y = var_3240_to_fp16)[name = tensor("attn_weights_43_cast_fp16")]; + tensor var_3248 = logical_not(x = attn_mask_31)[name = tensor("op_3248")]; + tensor var_3249_to_fp16 = const()[name = tensor("op_3249_to_fp16"), val = tensor(-inf)]; + tensor attn_weights_45_cast_fp16 = select(a = var_3249_to_fp16, b = attn_weights_43_cast_fp16, cond = var_3248)[name = tensor("attn_weights_45_cast_fp16")]; + tensor var_3251 = const()[name = tensor("op_3251"), val = tensor(-1)]; + tensor attn_weights_47_cast_fp16 = softmax(axis = var_3251, x = attn_weights_45_cast_fp16)[name = tensor("attn_weights_47_cast_fp16")]; + tensor attn_output_15_transpose_x_0 = const()[name = tensor("attn_output_15_transpose_x_0"), val = tensor(false)]; + tensor attn_output_15_transpose_y_0 = const()[name = tensor("attn_output_15_transpose_y_0"), val = tensor(false)]; + tensor values_47_cast_fp16 = transpose(perm = var_3191, x = values_45_cast_fp16)[name = tensor("transpose_187")]; + tensor attn_output_15_cast_fp16 = matmul(transpose_x = attn_output_15_transpose_x_0, transpose_y = attn_output_15_transpose_y_0, x = attn_weights_47_cast_fp16, y = values_47_cast_fp16)[name = tensor("attn_output_15_cast_fp16")]; + tensor var_3259 = const()[name = tensor("op_3259"), val = tensor([0, 2, 1, 3])]; + tensor var_3264 = const()[name = tensor("op_3264"), val = tensor([1, 1, 1024])]; + tensor var_3260_cast_fp16 = transpose(perm = var_3259, x = attn_output_15_cast_fp16)[name = tensor("transpose_184")]; + tensor input_91_cast_fp16 = reshape(shape = var_3264, x = var_3260_cast_fp16)[name = tensor("input_91_cast_fp16")]; + tensor attn7_out_proj_weight_to_fp16 = const()[name = tensor("attn7_out_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(183650432)))]; + tensor linear_30_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = attn7_out_proj_weight_to_fp16, x = input_91_cast_fp16)[name = tensor("linear_30_cast_fp16")]; + tensor var_3270_to_fp16 = const()[name = tensor("op_3270_to_fp16"), val = tensor(0x1p+0)]; + tensor var_3271_cast_fp16 = add(x = position7_to_fp16, y = var_3270_to_fp16)[name = tensor("op_3271_cast_fp16")]; + tensor var_3271_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_3271_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor input_93_cast_fp16 = add(x = input_87_cast_fp16, y = linear_30_cast_fp16)[name = tensor("input_93_cast_fp16")]; + tensor input_95_axes_0 = const()[name = tensor("input_95_axes_0"), val = tensor([-1])]; + tensor norm7_2_weight_to_fp16 = const()[name = tensor("norm7_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(185747648)))]; + tensor norm7_2_bias_to_fp16 = const()[name = tensor("norm7_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(185749760)))]; + tensor var_3275_to_fp16 = const()[name = tensor("op_3275_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_95_cast_fp16 = layer_norm(axes = input_95_axes_0, beta = norm7_2_bias_to_fp16, epsilon = var_3275_to_fp16, gamma = norm7_2_weight_to_fp16, x = input_93_cast_fp16)[name = tensor("input_95_cast_fp16")]; + tensor linear7_1_weight_to_fp16 = const()[name = tensor("linear7_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(185751872)))]; + tensor linear_31_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = linear7_1_weight_to_fp16, x = input_95_cast_fp16)[name = tensor("linear_31_cast_fp16")]; + tensor input_97_mode_0 = const()[name = tensor("input_97_mode_0"), val = tensor("EXACT")]; + tensor input_97_cast_fp16 = gelu(mode = input_97_mode_0, x = linear_31_cast_fp16)[name = tensor("input_97_cast_fp16")]; + tensor linear7_2_weight_to_fp16 = const()[name = tensor("linear7_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(194140544)))]; + tensor linear_32_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = linear7_2_weight_to_fp16, x = input_97_cast_fp16)[name = tensor("linear_32_cast_fp16")]; + tensor input_99_cast_fp16 = add(x = input_93_cast_fp16, y = linear_32_cast_fp16)[name = tensor("input_99_cast_fp16")]; + tensor input_101_axes_0 = const()[name = tensor("input_101_axes_0"), val = tensor([-1])]; + tensor norm8_1_weight_to_fp16 = const()[name = tensor("norm8_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(202529216)))]; + tensor norm8_1_bias_to_fp16 = const()[name = tensor("norm8_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(202531328)))]; + tensor var_3292_to_fp16 = const()[name = tensor("op_3292_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_101_cast_fp16 = layer_norm(axes = input_101_axes_0, beta = norm8_1_bias_to_fp16, epsilon = var_3292_to_fp16, gamma = norm8_1_weight_to_fp16, x = input_99_cast_fp16)[name = tensor("input_101_cast_fp16")]; + tensor attn8_in_proj_weight_to_fp16 = const()[name = tensor("attn8_in_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(202533440)))]; + tensor linear_33_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = attn8_in_proj_weight_to_fp16, x = input_101_cast_fp16)[name = tensor("linear_33_cast_fp16")]; + tensor var_3306 = const()[name = tensor("op_3306"), val = tensor([1, 1, 3, 16, 64])]; + tensor qkv_17_cast_fp16 = reshape(shape = var_3306, x = linear_33_cast_fp16)[name = tensor("qkv_17_cast_fp16")]; + tensor q_49_begin_0 = const()[name = tensor("q_49_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor q_49_end_0 = const()[name = tensor("q_49_end_0"), val = tensor([1, 1, 1, 16, 64])]; + tensor q_49_end_mask_0 = const()[name = tensor("q_49_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor q_49_squeeze_mask_0 = const()[name = tensor("q_49_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor q_49_cast_fp16 = slice_by_index(begin = q_49_begin_0, end = q_49_end_0, end_mask = q_49_end_mask_0, squeeze_mask = q_49_squeeze_mask_0, x = qkv_17_cast_fp16)[name = tensor("q_49_cast_fp16")]; + tensor k_33_begin_0 = const()[name = tensor("k_33_begin_0"), val = tensor([0, 0, 1, 0, 0])]; + tensor k_33_end_0 = const()[name = tensor("k_33_end_0"), val = tensor([1, 1, 2, 16, 64])]; + tensor k_33_end_mask_0 = const()[name = tensor("k_33_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor k_33_squeeze_mask_0 = const()[name = tensor("k_33_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor k_33_cast_fp16 = slice_by_index(begin = k_33_begin_0, end = k_33_end_0, end_mask = k_33_end_mask_0, squeeze_mask = k_33_squeeze_mask_0, x = qkv_17_cast_fp16)[name = tensor("k_33_cast_fp16")]; + tensor v_17_begin_0 = const()[name = tensor("v_17_begin_0"), val = tensor([0, 0, 2, 0, 0])]; + tensor v_17_end_0 = const()[name = tensor("v_17_end_0"), val = tensor([1, 1, 3, 16, 64])]; + tensor v_17_end_mask_0 = const()[name = tensor("v_17_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor v_17_squeeze_mask_0 = const()[name = tensor("v_17_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor v_17_cast_fp16 = slice_by_index(begin = v_17_begin_0, end = v_17_end_0, end_mask = v_17_end_mask_0, squeeze_mask = v_17_squeeze_mask_0, x = qkv_17_cast_fp16)[name = tensor("v_17_cast_fp16")]; + tensor var_3372 = const()[name = tensor("op_3372"), val = tensor([1, 1, 1, 1])]; + tensor position8_to_fp16_dtype_0 = const()[name = tensor("position8_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor position8_to_fp16 = cast(dtype = position8_to_fp16_dtype_0, x = position8)[name = tensor("cast_513")]; + tensor ts_53_cast_fp16 = reshape(shape = var_3372, x = position8_to_fp16)[name = tensor("ts_53_cast_fp16")]; + tensor var_3379 = const()[name = tensor("op_3379"), val = tensor([1, 1, 16, 32, 2])]; + tensor q_complex_17_cast_fp16 = reshape(shape = var_3379, x = q_49_cast_fp16)[name = tensor("q_complex_17_cast_fp16")]; + tensor var_3386 = const()[name = tensor("op_3386"), val = tensor([1, 1, 16, 32, 2])]; + tensor k_complex_17_cast_fp16 = reshape(shape = var_3386, x = k_33_cast_fp16)[name = tensor("k_complex_17_cast_fp16")]; + tensor var_3390_begin_0 = const()[name = tensor("op_3390_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_3390_end_0 = const()[name = tensor("op_3390_end_0"), val = tensor([1, 1, 16, 32, 1])]; + tensor var_3390_end_mask_0 = const()[name = tensor("op_3390_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_3390_squeeze_mask_0 = const()[name = tensor("op_3390_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_3390_cast_fp16 = slice_by_index(begin = var_3390_begin_0, end = var_3390_end_0, end_mask = var_3390_end_mask_0, squeeze_mask = var_3390_squeeze_mask_0, x = q_complex_17_cast_fp16)[name = tensor("op_3390_cast_fp16")]; + tensor var_3398_begin_0 = const()[name = tensor("op_3398_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_3398_end_0 = const()[name = tensor("op_3398_end_0"), val = tensor([1, 1, 16, 32, 2])]; + tensor var_3398_end_mask_0 = const()[name = tensor("op_3398_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_3398_squeeze_mask_0 = const()[name = tensor("op_3398_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_3398_cast_fp16 = slice_by_index(begin = var_3398_begin_0, end = var_3398_end_0, end_mask = var_3398_end_mask_0, squeeze_mask = var_3398_squeeze_mask_0, x = q_complex_17_cast_fp16)[name = tensor("op_3398_cast_fp16")]; + tensor var_3406_begin_0 = const()[name = tensor("op_3406_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_3406_end_0 = const()[name = tensor("op_3406_end_0"), val = tensor([1, 1, 16, 32, 1])]; + tensor var_3406_end_mask_0 = const()[name = tensor("op_3406_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_3406_squeeze_mask_0 = const()[name = tensor("op_3406_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_3406_cast_fp16 = slice_by_index(begin = var_3406_begin_0, end = var_3406_end_0, end_mask = var_3406_end_mask_0, squeeze_mask = var_3406_squeeze_mask_0, x = k_complex_17_cast_fp16)[name = tensor("op_3406_cast_fp16")]; + tensor var_3414_begin_0 = const()[name = tensor("op_3414_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_3414_end_0 = const()[name = tensor("op_3414_end_0"), val = tensor([1, 1, 16, 32, 2])]; + tensor var_3414_end_mask_0 = const()[name = tensor("op_3414_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_3414_squeeze_mask_0 = const()[name = tensor("op_3414_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_3414_cast_fp16 = slice_by_index(begin = var_3414_begin_0, end = var_3414_end_0, end_mask = var_3414_end_mask_0, squeeze_mask = var_3414_squeeze_mask_0, x = k_complex_17_cast_fp16)[name = tensor("op_3414_cast_fp16")]; + tensor freqs_17_to_fp16 = const()[name = tensor("freqs_17_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(208824960)))]; + tensor var_3420_cast_fp16 = mul(x = freqs_17_to_fp16, y = ts_53_cast_fp16)[name = tensor("op_3420_cast_fp16")]; + tensor rotr_17_cast_fp16 = cos(x = var_3420_cast_fp16)[name = tensor("rotr_17_cast_fp16")]; + tensor roti_17_cast_fp16 = sin(x = var_3420_cast_fp16)[name = tensor("roti_17_cast_fp16")]; + tensor var_3424_cast_fp16 = mul(x = var_3390_cast_fp16, y = rotr_17_cast_fp16)[name = tensor("op_3424_cast_fp16")]; + tensor var_3425_cast_fp16 = mul(x = var_3398_cast_fp16, y = roti_17_cast_fp16)[name = tensor("op_3425_cast_fp16")]; + tensor qor_33_cast_fp16 = sub(x = var_3424_cast_fp16, y = var_3425_cast_fp16)[name = tensor("qor_33_cast_fp16")]; + tensor var_3428_cast_fp16 = mul(x = var_3390_cast_fp16, y = roti_17_cast_fp16)[name = tensor("op_3428_cast_fp16")]; + tensor var_3429_cast_fp16 = mul(x = var_3398_cast_fp16, y = rotr_17_cast_fp16)[name = tensor("op_3429_cast_fp16")]; + tensor qoi_33_cast_fp16 = add(x = var_3428_cast_fp16, y = var_3429_cast_fp16)[name = tensor("qoi_33_cast_fp16")]; + tensor var_3432_cast_fp16 = mul(x = var_3406_cast_fp16, y = rotr_17_cast_fp16)[name = tensor("op_3432_cast_fp16")]; + tensor var_3433_cast_fp16 = mul(x = var_3414_cast_fp16, y = roti_17_cast_fp16)[name = tensor("op_3433_cast_fp16")]; + tensor kor_33_cast_fp16 = sub(x = var_3432_cast_fp16, y = var_3433_cast_fp16)[name = tensor("kor_33_cast_fp16")]; + tensor var_3436_cast_fp16 = mul(x = var_3406_cast_fp16, y = roti_17_cast_fp16)[name = tensor("op_3436_cast_fp16")]; + tensor var_3437_cast_fp16 = mul(x = var_3414_cast_fp16, y = rotr_17_cast_fp16)[name = tensor("op_3437_cast_fp16")]; + tensor koi_33_cast_fp16 = add(x = var_3436_cast_fp16, y = var_3437_cast_fp16)[name = tensor("koi_33_cast_fp16")]; + tensor qo_17_axis_0 = const()[name = tensor("qo_17_axis_0"), val = tensor(-1)]; + tensor qo_17_cast_fp16 = stack(axis = qo_17_axis_0, values = (qor_33_cast_fp16, qoi_33_cast_fp16))[name = tensor("qo_17_cast_fp16")]; + tensor ko_17_axis_0 = const()[name = tensor("ko_17_axis_0"), val = tensor(-1)]; + tensor ko_17_cast_fp16 = stack(axis = ko_17_axis_0, values = (kor_33_cast_fp16, koi_33_cast_fp16))[name = tensor("ko_17_cast_fp16")]; + tensor var_3470 = const()[name = tensor("op_3470"), val = tensor([1, 1, 16, 64])]; + tensor q_51_cast_fp16 = reshape(shape = var_3470, x = qo_17_cast_fp16)[name = tensor("q_51_cast_fp16")]; + tensor var_3476 = const()[name = tensor("op_3476"), val = tensor([1, 1, 16, 64])]; + tensor k_35_cast_fp16 = reshape(shape = var_3476, x = ko_17_cast_fp16)[name = tensor("k_35_cast_fp16")]; + tensor _inversed_3501_y_0_to_fp16 = const()[name = tensor("_inversed_3501_y_0_to_fp16"), val = tensor(0x1p-9)]; + tensor _inversed_3501_cast_fp16 = mul(x = ts_53_cast_fp16, y = _inversed_3501_y_0_to_fp16)[name = tensor("_inversed_3501_cast_fp16")]; + tensor var_3502_cast_fp16 = floor(x = _inversed_3501_cast_fp16)[name = tensor("op_3502_cast_fp16")]; + tensor var_3503_to_fp16 = const()[name = tensor("op_3503_to_fp16"), val = tensor(0x1p+9)]; + tensor var_3504_cast_fp16 = mul(x = var_3502_cast_fp16, y = var_3503_to_fp16)[name = tensor("op_3504_cast_fp16")]; + tensor write_indices_float_35_cast_fp16 = sub(x = ts_53_cast_fp16, y = var_3504_cast_fp16)[name = tensor("write_indices_float_35_cast_fp16")]; + tensor var_3511_dtype_0 = const()[name = tensor("op_3511_dtype_0"), val = tensor("int32")]; + tensor write_indices_17_reps_0 = const()[name = tensor("write_indices_17_reps_0"), val = tensor([1, 1, 16, 64])]; + tensor write_indices_float_35_cast_fp16_to_int32 = cast(dtype = var_3511_dtype_0, x = write_indices_float_35_cast_fp16)[name = tensor("cast_512")]; + tensor write_indices_17 = tile(reps = write_indices_17_reps_0, x = write_indices_float_35_cast_fp16_to_int32)[name = tensor("write_indices_17")]; + tensor var_3521_begin_0 = const()[name = tensor("op_3521_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_3521_end_0 = const()[name = tensor("op_3521_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor var_3521_end_mask_0 = const()[name = tensor("op_3521_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_3521_squeeze_mask_0 = const()[name = tensor("op_3521_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor cache8_to_fp16_dtype_0 = const()[name = tensor("cache8_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor cache8_to_fp16 = cast(dtype = cache8_to_fp16_dtype_0, x = cache8)[name = tensor("cast_511")]; + tensor var_3521_cast_fp16 = slice_by_index(begin = var_3521_begin_0, end = var_3521_end_0, end_mask = var_3521_end_mask_0, squeeze_mask = var_3521_squeeze_mask_0, x = cache8_to_fp16)[name = tensor("op_3521_cast_fp16")]; + tensor var_3523_axis_0 = const()[name = tensor("op_3523_axis_0"), val = tensor(1)]; + tensor var_3523_mode_0 = const()[name = tensor("op_3523_mode_0"), val = tensor("update")]; + tensor var_3523_validate_indices_0 = const()[name = tensor("op_3523_validate_indices_0"), val = tensor(false)]; + tensor var_3523_cast_fp16 = scatter_along_axis(axis = var_3523_axis_0, data = var_3521_cast_fp16, indices = write_indices_17, mode = var_3523_mode_0, updates = k_35_cast_fp16, validate_indices = var_3523_validate_indices_0)[name = tensor("op_3523_cast_fp16")]; + tensor concat_58 = const()[name = tensor("concat_58"), val = tensor([0, 0, 0, 0, 0])]; + tensor concat_59 = const()[name = tensor("concat_59"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_17_internal_tensor_assign_1_stride_0 = const()[name = tensor("new_cache_17_internal_tensor_assign_1_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_17_internal_tensor_assign_1_begin_mask_0 = const()[name = tensor("new_cache_17_internal_tensor_assign_1_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_17_internal_tensor_assign_1_end_mask_0 = const()[name = tensor("new_cache_17_internal_tensor_assign_1_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_17_internal_tensor_assign_1_squeeze_mask_0 = const()[name = tensor("new_cache_17_internal_tensor_assign_1_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_64 = const()[name = tensor("shape_64"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_16 = const()[name = tensor("reduce_prod_16"), val = tensor(1048576)]; + tensor range_1d_16_start_0 = const()[name = tensor("range_1d_16_start_0"), val = tensor(0)]; + tensor range_1d_16_step_0 = const()[name = tensor("range_1d_16_step_0"), val = tensor(1)]; + tensor range_1d_16 = range_1d(end = reduce_prod_16, start = range_1d_16_start_0, step = range_1d_16_step_0)[name = tensor("range_1d_16")]; + tensor reshape_80 = reshape(shape = shape_64, x = range_1d_16)[name = tensor("reshape_80")]; + tensor slice_by_index_16 = slice_by_index(begin = concat_58, begin_mask = new_cache_17_internal_tensor_assign_1_begin_mask_0, end = concat_59, end_mask = new_cache_17_internal_tensor_assign_1_end_mask_0, squeeze_mask = new_cache_17_internal_tensor_assign_1_squeeze_mask_0, stride = new_cache_17_internal_tensor_assign_1_stride_0, x = reshape_80)[name = tensor("slice_by_index_16")]; + tensor reshape_81_shape_0 = const()[name = tensor("reshape_81_shape_0"), val = tensor([-1])]; + tensor reshape_81 = reshape(shape = reshape_81_shape_0, x = slice_by_index_16)[name = tensor("reshape_81")]; + tensor reshape_82_shape_0 = const()[name = tensor("reshape_82_shape_0"), val = tensor([-1])]; + tensor reshape_82_cast_fp16 = reshape(shape = reshape_82_shape_0, x = var_3523_cast_fp16)[name = tensor("reshape_82_cast_fp16")]; + tensor reshape_83_shape_0 = const()[name = tensor("reshape_83_shape_0"), val = tensor([-1])]; + tensor reshape_83_cast_fp16 = reshape(shape = reshape_83_shape_0, x = cache8_to_fp16)[name = tensor("reshape_83_cast_fp16")]; + tensor scatter_16_mode_0 = const()[name = tensor("scatter_16_mode_0"), val = tensor("update")]; + tensor scatter_16_axis_0 = const()[name = tensor("scatter_16_axis_0"), val = tensor(0)]; + tensor scatter_16_validate_indices_0 = const()[name = tensor("scatter_16_validate_indices_0"), val = tensor(false)]; + tensor scatter_16_cast_fp16 = scatter(axis = scatter_16_axis_0, data = reshape_83_cast_fp16, indices = reshape_81, mode = scatter_16_mode_0, updates = reshape_82_cast_fp16, validate_indices = scatter_16_validate_indices_0)[name = tensor("scatter_16_cast_fp16")]; + tensor reshape_84_cast_fp16 = reshape(shape = shape_64, x = scatter_16_cast_fp16)[name = tensor("reshape_84_cast_fp16")]; + tensor var_3531_begin_0 = const()[name = tensor("op_3531_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor var_3531_end_0 = const()[name = tensor("op_3531_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor var_3531_end_mask_0 = const()[name = tensor("op_3531_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_3531_squeeze_mask_0 = const()[name = tensor("op_3531_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor var_3531_cast_fp16 = slice_by_index(begin = var_3531_begin_0, end = var_3531_end_0, end_mask = var_3531_end_mask_0, squeeze_mask = var_3531_squeeze_mask_0, x = reshape_84_cast_fp16)[name = tensor("op_3531_cast_fp16")]; + tensor var_3533_axis_0 = const()[name = tensor("op_3533_axis_0"), val = tensor(1)]; + tensor var_3533_mode_0 = const()[name = tensor("op_3533_mode_0"), val = tensor("update")]; + tensor var_3533_validate_indices_0 = const()[name = tensor("op_3533_validate_indices_0"), val = tensor(false)]; + tensor var_3533_cast_fp16 = scatter_along_axis(axis = var_3533_axis_0, data = var_3531_cast_fp16, indices = write_indices_17, mode = var_3533_mode_0, updates = v_17_cast_fp16, validate_indices = var_3533_validate_indices_0)[name = tensor("op_3533_cast_fp16")]; + tensor concat_60 = const()[name = tensor("concat_60"), val = tensor([1, 0, 0, 0, 0])]; + tensor concat_61 = const()[name = tensor("concat_61"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_17_internal_tensor_assign_2_stride_0 = const()[name = tensor("new_cache_17_internal_tensor_assign_2_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_17_internal_tensor_assign_2_begin_mask_0 = const()[name = tensor("new_cache_17_internal_tensor_assign_2_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_17_internal_tensor_assign_2_end_mask_0 = const()[name = tensor("new_cache_17_internal_tensor_assign_2_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_17_internal_tensor_assign_2_squeeze_mask_0 = const()[name = tensor("new_cache_17_internal_tensor_assign_2_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_65 = const()[name = tensor("shape_65"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_17 = const()[name = tensor("reduce_prod_17"), val = tensor(1048576)]; + tensor range_1d_17_start_0 = const()[name = tensor("range_1d_17_start_0"), val = tensor(0)]; + tensor range_1d_17_step_0 = const()[name = tensor("range_1d_17_step_0"), val = tensor(1)]; + tensor range_1d_17 = range_1d(end = reduce_prod_17, start = range_1d_17_start_0, step = range_1d_17_step_0)[name = tensor("range_1d_17")]; + tensor reshape_85 = reshape(shape = shape_65, x = range_1d_17)[name = tensor("reshape_85")]; + tensor slice_by_index_17 = slice_by_index(begin = concat_60, begin_mask = new_cache_17_internal_tensor_assign_2_begin_mask_0, end = concat_61, end_mask = new_cache_17_internal_tensor_assign_2_end_mask_0, squeeze_mask = new_cache_17_internal_tensor_assign_2_squeeze_mask_0, stride = new_cache_17_internal_tensor_assign_2_stride_0, x = reshape_85)[name = tensor("slice_by_index_17")]; + tensor reshape_86_shape_0 = const()[name = tensor("reshape_86_shape_0"), val = tensor([-1])]; + tensor reshape_86 = reshape(shape = reshape_86_shape_0, x = slice_by_index_17)[name = tensor("reshape_86")]; + tensor reshape_87_shape_0 = const()[name = tensor("reshape_87_shape_0"), val = tensor([-1])]; + tensor reshape_87_cast_fp16 = reshape(shape = reshape_87_shape_0, x = var_3533_cast_fp16)[name = tensor("reshape_87_cast_fp16")]; + tensor reshape_88_shape_0 = const()[name = tensor("reshape_88_shape_0"), val = tensor([-1])]; + tensor reshape_88_cast_fp16 = reshape(shape = reshape_88_shape_0, x = reshape_84_cast_fp16)[name = tensor("reshape_88_cast_fp16")]; + tensor scatter_17_mode_0 = const()[name = tensor("scatter_17_mode_0"), val = tensor("update")]; + tensor scatter_17_axis_0 = const()[name = tensor("scatter_17_axis_0"), val = tensor(0)]; + tensor scatter_17_validate_indices_0 = const()[name = tensor("scatter_17_validate_indices_0"), val = tensor(false)]; + tensor scatter_17_cast_fp16 = scatter(axis = scatter_17_axis_0, data = reshape_88_cast_fp16, indices = reshape_86, mode = scatter_17_mode_0, updates = reshape_87_cast_fp16, validate_indices = scatter_17_validate_indices_0)[name = tensor("scatter_17_cast_fp16")]; + tensor reshape_89_cast_fp16 = reshape(shape = shape_65, x = scatter_17_cast_fp16)[name = tensor("reshape_89_cast_fp16")]; + tensor reshape_89_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("reshape_89_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor keys_49_begin_0 = const()[name = tensor("keys_49_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor keys_49_end_0 = const()[name = tensor("keys_49_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor keys_49_end_mask_0 = const()[name = tensor("keys_49_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor keys_49_squeeze_mask_0 = const()[name = tensor("keys_49_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor keys_49_cast_fp16 = slice_by_index(begin = keys_49_begin_0, end = keys_49_end_0, end_mask = keys_49_end_mask_0, squeeze_mask = keys_49_squeeze_mask_0, x = reshape_89_cast_fp16)[name = tensor("keys_49_cast_fp16")]; + tensor values_49_begin_0 = const()[name = tensor("values_49_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor values_49_end_0 = const()[name = tensor("values_49_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor values_49_end_mask_0 = const()[name = tensor("values_49_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor values_49_squeeze_mask_0 = const()[name = tensor("values_49_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor values_49_cast_fp16 = slice_by_index(begin = values_49_begin_0, end = values_49_end_0, end_mask = values_49_end_mask_0, squeeze_mask = values_49_squeeze_mask_0, x = reshape_89_cast_fp16)[name = tensor("values_49_cast_fp16")]; + tensor var_3545_cast_fp16 = not_equal(x = keys_49_cast_fp16, y = keys_49_cast_fp16)[name = tensor("op_3545_cast_fp16")]; + tensor keys_51_cast_fp16 = select(a = var_463_to_fp16, b = keys_49_cast_fp16, cond = var_3545_cast_fp16)[name = tensor("keys_51_cast_fp16")]; + tensor var_3553_cast_fp16 = not_equal(x = values_49_cast_fp16, y = values_49_cast_fp16)[name = tensor("op_3553_cast_fp16")]; + tensor values_51_cast_fp16 = select(a = var_463_to_fp16, b = values_49_cast_fp16, cond = var_3553_cast_fp16)[name = tensor("values_51_cast_fp16")]; + tensor var_3577 = const()[name = tensor("op_3577"), val = tensor([0, 2, 1, 3])]; + tensor var_3593 = const()[name = tensor("op_3593"), val = tensor([1, 1, 1])]; + tensor var_3594_cast_fp16 = reshape(shape = var_3593, x = position8_to_fp16)[name = tensor("op_3594_cast_fp16")]; + tensor var_3614_to_fp16 = const()[name = tensor("op_3614_to_fp16"), val = tensor(0x1p+0)]; + tensor valid_len_17_cast_fp16 = add(x = var_3594_cast_fp16, y = var_3614_to_fp16)[name = tensor("valid_len_17_cast_fp16")]; + tensor valid_mask_17_cast_fp16 = less(x = k_positions_1_promoted_to_fp16, y = valid_len_17_cast_fp16)[name = tensor("valid_mask_17_cast_fp16")]; + tensor causal_mask_17_cast_fp16 = less_equal(x = k_positions_1_promoted_to_fp16, y = var_3594_cast_fp16)[name = tensor("causal_mask_17_cast_fp16")]; + tensor attn_mask_33 = logical_and(x = valid_mask_17_cast_fp16, y = causal_mask_17_cast_fp16)[name = tensor("attn_mask_33")]; + tensor attn_mask_35_axes_0 = const()[name = tensor("attn_mask_35_axes_0"), val = tensor([1])]; + tensor attn_mask_35 = expand_dims(axes = attn_mask_35_axes_0, x = attn_mask_33)[name = tensor("attn_mask_35")]; + tensor var_3632_transpose_x_0 = const()[name = tensor("op_3632_transpose_x_0"), val = tensor(false)]; + tensor var_3632_transpose_y_0 = const()[name = tensor("op_3632_transpose_y_0"), val = tensor(false)]; + tensor transpose_88_perm_0 = const()[name = tensor("transpose_88_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_89_perm_0 = const()[name = tensor("transpose_89_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_89 = transpose(perm = transpose_89_perm_0, x = keys_51_cast_fp16)[name = tensor("transpose_181")]; + tensor transpose_88 = transpose(perm = transpose_88_perm_0, x = q_51_cast_fp16)[name = tensor("transpose_182")]; + tensor var_3632_cast_fp16 = matmul(transpose_x = var_3632_transpose_x_0, transpose_y = var_3632_transpose_y_0, x = transpose_88, y = transpose_89)[name = tensor("op_3632_cast_fp16")]; + tensor var_3626_to_fp16 = const()[name = tensor("op_3626_to_fp16"), val = tensor([0x1p-3])]; + tensor attn_weights_49_cast_fp16 = mul(x = var_3632_cast_fp16, y = var_3626_to_fp16)[name = tensor("attn_weights_49_cast_fp16")]; + tensor var_3634 = logical_not(x = attn_mask_35)[name = tensor("op_3634")]; + tensor var_3635_to_fp16 = const()[name = tensor("op_3635_to_fp16"), val = tensor(-inf)]; + tensor attn_weights_51_cast_fp16 = select(a = var_3635_to_fp16, b = attn_weights_49_cast_fp16, cond = var_3634)[name = tensor("attn_weights_51_cast_fp16")]; + tensor var_3637 = const()[name = tensor("op_3637"), val = tensor(-1)]; + tensor attn_weights_53_cast_fp16 = softmax(axis = var_3637, x = attn_weights_51_cast_fp16)[name = tensor("attn_weights_53_cast_fp16")]; + tensor attn_output_17_transpose_x_0 = const()[name = tensor("attn_output_17_transpose_x_0"), val = tensor(false)]; + tensor attn_output_17_transpose_y_0 = const()[name = tensor("attn_output_17_transpose_y_0"), val = tensor(false)]; + tensor values_53_cast_fp16 = transpose(perm = var_3577, x = values_51_cast_fp16)[name = tensor("transpose_183")]; + tensor attn_output_17_cast_fp16 = matmul(transpose_x = attn_output_17_transpose_x_0, transpose_y = attn_output_17_transpose_y_0, x = attn_weights_53_cast_fp16, y = values_53_cast_fp16)[name = tensor("attn_output_17_cast_fp16")]; + tensor var_3645 = const()[name = tensor("op_3645"), val = tensor([0, 2, 1, 3])]; + tensor var_3650 = const()[name = tensor("op_3650"), val = tensor([1, 1, 1024])]; + tensor var_3646_cast_fp16 = transpose(perm = var_3645, x = attn_output_17_cast_fp16)[name = tensor("transpose_180")]; + tensor input_103_cast_fp16 = reshape(shape = var_3650, x = var_3646_cast_fp16)[name = tensor("input_103_cast_fp16")]; + tensor attn8_out_proj_weight_to_fp16 = const()[name = tensor("attn8_out_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(208825088)))]; + tensor linear_34_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = attn8_out_proj_weight_to_fp16, x = input_103_cast_fp16)[name = tensor("linear_34_cast_fp16")]; + tensor var_3656_to_fp16 = const()[name = tensor("op_3656_to_fp16"), val = tensor(0x1p+0)]; + tensor var_3657_cast_fp16 = add(x = position8_to_fp16, y = var_3656_to_fp16)[name = tensor("op_3657_cast_fp16")]; + tensor var_3657_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_3657_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor input_105_cast_fp16 = add(x = input_99_cast_fp16, y = linear_34_cast_fp16)[name = tensor("input_105_cast_fp16")]; + tensor input_107_axes_0 = const()[name = tensor("input_107_axes_0"), val = tensor([-1])]; + tensor norm8_2_weight_to_fp16 = const()[name = tensor("norm8_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(210922304)))]; + tensor norm8_2_bias_to_fp16 = const()[name = tensor("norm8_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(210924416)))]; + tensor var_3661_to_fp16 = const()[name = tensor("op_3661_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_107_cast_fp16 = layer_norm(axes = input_107_axes_0, beta = norm8_2_bias_to_fp16, epsilon = var_3661_to_fp16, gamma = norm8_2_weight_to_fp16, x = input_105_cast_fp16)[name = tensor("input_107_cast_fp16")]; + tensor linear8_1_weight_to_fp16 = const()[name = tensor("linear8_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(210926528)))]; + tensor linear_35_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = linear8_1_weight_to_fp16, x = input_107_cast_fp16)[name = tensor("linear_35_cast_fp16")]; + tensor input_109_mode_0 = const()[name = tensor("input_109_mode_0"), val = tensor("EXACT")]; + tensor input_109_cast_fp16 = gelu(mode = input_109_mode_0, x = linear_35_cast_fp16)[name = tensor("input_109_cast_fp16")]; + tensor linear8_2_weight_to_fp16 = const()[name = tensor("linear8_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(219315200)))]; + tensor linear_36_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = linear8_2_weight_to_fp16, x = input_109_cast_fp16)[name = tensor("linear_36_cast_fp16")]; + tensor input_111_cast_fp16 = add(x = input_105_cast_fp16, y = linear_36_cast_fp16)[name = tensor("input_111_cast_fp16")]; + tensor input_113_axes_0 = const()[name = tensor("input_113_axes_0"), val = tensor([-1])]; + tensor norm9_1_weight_to_fp16 = const()[name = tensor("norm9_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(227703872)))]; + tensor norm9_1_bias_to_fp16 = const()[name = tensor("norm9_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(227705984)))]; + tensor var_3678_to_fp16 = const()[name = tensor("op_3678_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_113_cast_fp16 = layer_norm(axes = input_113_axes_0, beta = norm9_1_bias_to_fp16, epsilon = var_3678_to_fp16, gamma = norm9_1_weight_to_fp16, x = input_111_cast_fp16)[name = tensor("input_113_cast_fp16")]; + tensor attn9_in_proj_weight_to_fp16 = const()[name = tensor("attn9_in_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(227708096)))]; + tensor linear_37_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = attn9_in_proj_weight_to_fp16, x = input_113_cast_fp16)[name = tensor("linear_37_cast_fp16")]; + tensor var_3692 = const()[name = tensor("op_3692"), val = tensor([1, 1, 3, 16, 64])]; + tensor qkv_19_cast_fp16 = reshape(shape = var_3692, x = linear_37_cast_fp16)[name = tensor("qkv_19_cast_fp16")]; + tensor q_55_begin_0 = const()[name = tensor("q_55_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor q_55_end_0 = const()[name = tensor("q_55_end_0"), val = tensor([1, 1, 1, 16, 64])]; + tensor q_55_end_mask_0 = const()[name = tensor("q_55_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor q_55_squeeze_mask_0 = const()[name = tensor("q_55_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor q_55_cast_fp16 = slice_by_index(begin = q_55_begin_0, end = q_55_end_0, end_mask = q_55_end_mask_0, squeeze_mask = q_55_squeeze_mask_0, x = qkv_19_cast_fp16)[name = tensor("q_55_cast_fp16")]; + tensor k_37_begin_0 = const()[name = tensor("k_37_begin_0"), val = tensor([0, 0, 1, 0, 0])]; + tensor k_37_end_0 = const()[name = tensor("k_37_end_0"), val = tensor([1, 1, 2, 16, 64])]; + tensor k_37_end_mask_0 = const()[name = tensor("k_37_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor k_37_squeeze_mask_0 = const()[name = tensor("k_37_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor k_37_cast_fp16 = slice_by_index(begin = k_37_begin_0, end = k_37_end_0, end_mask = k_37_end_mask_0, squeeze_mask = k_37_squeeze_mask_0, x = qkv_19_cast_fp16)[name = tensor("k_37_cast_fp16")]; + tensor v_19_begin_0 = const()[name = tensor("v_19_begin_0"), val = tensor([0, 0, 2, 0, 0])]; + tensor v_19_end_0 = const()[name = tensor("v_19_end_0"), val = tensor([1, 1, 3, 16, 64])]; + tensor v_19_end_mask_0 = const()[name = tensor("v_19_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor v_19_squeeze_mask_0 = const()[name = tensor("v_19_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor v_19_cast_fp16 = slice_by_index(begin = v_19_begin_0, end = v_19_end_0, end_mask = v_19_end_mask_0, squeeze_mask = v_19_squeeze_mask_0, x = qkv_19_cast_fp16)[name = tensor("v_19_cast_fp16")]; + tensor var_3758 = const()[name = tensor("op_3758"), val = tensor([1, 1, 1, 1])]; + tensor position9_to_fp16_dtype_0 = const()[name = tensor("position9_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor position9_to_fp16 = cast(dtype = position9_to_fp16_dtype_0, x = position9)[name = tensor("cast_508")]; + tensor ts_59_cast_fp16 = reshape(shape = var_3758, x = position9_to_fp16)[name = tensor("ts_59_cast_fp16")]; + tensor var_3765 = const()[name = tensor("op_3765"), val = tensor([1, 1, 16, 32, 2])]; + tensor q_complex_19_cast_fp16 = reshape(shape = var_3765, x = q_55_cast_fp16)[name = tensor("q_complex_19_cast_fp16")]; + tensor var_3772 = const()[name = tensor("op_3772"), val = tensor([1, 1, 16, 32, 2])]; + tensor k_complex_19_cast_fp16 = reshape(shape = var_3772, x = k_37_cast_fp16)[name = tensor("k_complex_19_cast_fp16")]; + tensor var_3776_begin_0 = const()[name = tensor("op_3776_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_3776_end_0 = const()[name = tensor("op_3776_end_0"), val = tensor([1, 1, 16, 32, 1])]; + tensor var_3776_end_mask_0 = const()[name = tensor("op_3776_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_3776_squeeze_mask_0 = const()[name = tensor("op_3776_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_3776_cast_fp16 = slice_by_index(begin = var_3776_begin_0, end = var_3776_end_0, end_mask = var_3776_end_mask_0, squeeze_mask = var_3776_squeeze_mask_0, x = q_complex_19_cast_fp16)[name = tensor("op_3776_cast_fp16")]; + tensor var_3784_begin_0 = const()[name = tensor("op_3784_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_3784_end_0 = const()[name = tensor("op_3784_end_0"), val = tensor([1, 1, 16, 32, 2])]; + tensor var_3784_end_mask_0 = const()[name = tensor("op_3784_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_3784_squeeze_mask_0 = const()[name = tensor("op_3784_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_3784_cast_fp16 = slice_by_index(begin = var_3784_begin_0, end = var_3784_end_0, end_mask = var_3784_end_mask_0, squeeze_mask = var_3784_squeeze_mask_0, x = q_complex_19_cast_fp16)[name = tensor("op_3784_cast_fp16")]; + tensor var_3792_begin_0 = const()[name = tensor("op_3792_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_3792_end_0 = const()[name = tensor("op_3792_end_0"), val = tensor([1, 1, 16, 32, 1])]; + tensor var_3792_end_mask_0 = const()[name = tensor("op_3792_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_3792_squeeze_mask_0 = const()[name = tensor("op_3792_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_3792_cast_fp16 = slice_by_index(begin = var_3792_begin_0, end = var_3792_end_0, end_mask = var_3792_end_mask_0, squeeze_mask = var_3792_squeeze_mask_0, x = k_complex_19_cast_fp16)[name = tensor("op_3792_cast_fp16")]; + tensor var_3800_begin_0 = const()[name = tensor("op_3800_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_3800_end_0 = const()[name = tensor("op_3800_end_0"), val = tensor([1, 1, 16, 32, 2])]; + tensor var_3800_end_mask_0 = const()[name = tensor("op_3800_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_3800_squeeze_mask_0 = const()[name = tensor("op_3800_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_3800_cast_fp16 = slice_by_index(begin = var_3800_begin_0, end = var_3800_end_0, end_mask = var_3800_end_mask_0, squeeze_mask = var_3800_squeeze_mask_0, x = k_complex_19_cast_fp16)[name = tensor("op_3800_cast_fp16")]; + tensor freqs_19_to_fp16 = const()[name = tensor("freqs_19_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(233999616)))]; + tensor var_3806_cast_fp16 = mul(x = freqs_19_to_fp16, y = ts_59_cast_fp16)[name = tensor("op_3806_cast_fp16")]; + tensor rotr_19_cast_fp16 = cos(x = var_3806_cast_fp16)[name = tensor("rotr_19_cast_fp16")]; + tensor roti_19_cast_fp16 = sin(x = var_3806_cast_fp16)[name = tensor("roti_19_cast_fp16")]; + tensor var_3810_cast_fp16 = mul(x = var_3776_cast_fp16, y = rotr_19_cast_fp16)[name = tensor("op_3810_cast_fp16")]; + tensor var_3811_cast_fp16 = mul(x = var_3784_cast_fp16, y = roti_19_cast_fp16)[name = tensor("op_3811_cast_fp16")]; + tensor qor_37_cast_fp16 = sub(x = var_3810_cast_fp16, y = var_3811_cast_fp16)[name = tensor("qor_37_cast_fp16")]; + tensor var_3814_cast_fp16 = mul(x = var_3776_cast_fp16, y = roti_19_cast_fp16)[name = tensor("op_3814_cast_fp16")]; + tensor var_3815_cast_fp16 = mul(x = var_3784_cast_fp16, y = rotr_19_cast_fp16)[name = tensor("op_3815_cast_fp16")]; + tensor qoi_37_cast_fp16 = add(x = var_3814_cast_fp16, y = var_3815_cast_fp16)[name = tensor("qoi_37_cast_fp16")]; + tensor var_3818_cast_fp16 = mul(x = var_3792_cast_fp16, y = rotr_19_cast_fp16)[name = tensor("op_3818_cast_fp16")]; + tensor var_3819_cast_fp16 = mul(x = var_3800_cast_fp16, y = roti_19_cast_fp16)[name = tensor("op_3819_cast_fp16")]; + tensor kor_37_cast_fp16 = sub(x = var_3818_cast_fp16, y = var_3819_cast_fp16)[name = tensor("kor_37_cast_fp16")]; + tensor var_3822_cast_fp16 = mul(x = var_3792_cast_fp16, y = roti_19_cast_fp16)[name = tensor("op_3822_cast_fp16")]; + tensor var_3823_cast_fp16 = mul(x = var_3800_cast_fp16, y = rotr_19_cast_fp16)[name = tensor("op_3823_cast_fp16")]; + tensor koi_37_cast_fp16 = add(x = var_3822_cast_fp16, y = var_3823_cast_fp16)[name = tensor("koi_37_cast_fp16")]; + tensor qo_19_axis_0 = const()[name = tensor("qo_19_axis_0"), val = tensor(-1)]; + tensor qo_19_cast_fp16 = stack(axis = qo_19_axis_0, values = (qor_37_cast_fp16, qoi_37_cast_fp16))[name = tensor("qo_19_cast_fp16")]; + tensor ko_19_axis_0 = const()[name = tensor("ko_19_axis_0"), val = tensor(-1)]; + tensor ko_19_cast_fp16 = stack(axis = ko_19_axis_0, values = (kor_37_cast_fp16, koi_37_cast_fp16))[name = tensor("ko_19_cast_fp16")]; + tensor var_3856 = const()[name = tensor("op_3856"), val = tensor([1, 1, 16, 64])]; + tensor q_57_cast_fp16 = reshape(shape = var_3856, x = qo_19_cast_fp16)[name = tensor("q_57_cast_fp16")]; + tensor var_3862 = const()[name = tensor("op_3862"), val = tensor([1, 1, 16, 64])]; + tensor k_39_cast_fp16 = reshape(shape = var_3862, x = ko_19_cast_fp16)[name = tensor("k_39_cast_fp16")]; + tensor _inversed_3887_y_0_to_fp16 = const()[name = tensor("_inversed_3887_y_0_to_fp16"), val = tensor(0x1p-9)]; + tensor _inversed_3887_cast_fp16 = mul(x = ts_59_cast_fp16, y = _inversed_3887_y_0_to_fp16)[name = tensor("_inversed_3887_cast_fp16")]; + tensor var_3888_cast_fp16 = floor(x = _inversed_3887_cast_fp16)[name = tensor("op_3888_cast_fp16")]; + tensor var_3889_to_fp16 = const()[name = tensor("op_3889_to_fp16"), val = tensor(0x1p+9)]; + tensor var_3890_cast_fp16 = mul(x = var_3888_cast_fp16, y = var_3889_to_fp16)[name = tensor("op_3890_cast_fp16")]; + tensor write_indices_float_39_cast_fp16 = sub(x = ts_59_cast_fp16, y = var_3890_cast_fp16)[name = tensor("write_indices_float_39_cast_fp16")]; + tensor var_3897_dtype_0 = const()[name = tensor("op_3897_dtype_0"), val = tensor("int32")]; + tensor write_indices_19_reps_0 = const()[name = tensor("write_indices_19_reps_0"), val = tensor([1, 1, 16, 64])]; + tensor write_indices_float_39_cast_fp16_to_int32 = cast(dtype = var_3897_dtype_0, x = write_indices_float_39_cast_fp16)[name = tensor("cast_507")]; + tensor write_indices_19 = tile(reps = write_indices_19_reps_0, x = write_indices_float_39_cast_fp16_to_int32)[name = tensor("write_indices_19")]; + tensor var_3907_begin_0 = const()[name = tensor("op_3907_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_3907_end_0 = const()[name = tensor("op_3907_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor var_3907_end_mask_0 = const()[name = tensor("op_3907_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_3907_squeeze_mask_0 = const()[name = tensor("op_3907_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor cache9_to_fp16_dtype_0 = const()[name = tensor("cache9_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor cache9_to_fp16 = cast(dtype = cache9_to_fp16_dtype_0, x = cache9)[name = tensor("cast_506")]; + tensor var_3907_cast_fp16 = slice_by_index(begin = var_3907_begin_0, end = var_3907_end_0, end_mask = var_3907_end_mask_0, squeeze_mask = var_3907_squeeze_mask_0, x = cache9_to_fp16)[name = tensor("op_3907_cast_fp16")]; + tensor var_3909_axis_0 = const()[name = tensor("op_3909_axis_0"), val = tensor(1)]; + tensor var_3909_mode_0 = const()[name = tensor("op_3909_mode_0"), val = tensor("update")]; + tensor var_3909_validate_indices_0 = const()[name = tensor("op_3909_validate_indices_0"), val = tensor(false)]; + tensor var_3909_cast_fp16 = scatter_along_axis(axis = var_3909_axis_0, data = var_3907_cast_fp16, indices = write_indices_19, mode = var_3909_mode_0, updates = k_39_cast_fp16, validate_indices = var_3909_validate_indices_0)[name = tensor("op_3909_cast_fp16")]; + tensor concat_65 = const()[name = tensor("concat_65"), val = tensor([0, 0, 0, 0, 0])]; + tensor concat_66 = const()[name = tensor("concat_66"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_19_internal_tensor_assign_1_stride_0 = const()[name = tensor("new_cache_19_internal_tensor_assign_1_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_19_internal_tensor_assign_1_begin_mask_0 = const()[name = tensor("new_cache_19_internal_tensor_assign_1_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_19_internal_tensor_assign_1_end_mask_0 = const()[name = tensor("new_cache_19_internal_tensor_assign_1_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_19_internal_tensor_assign_1_squeeze_mask_0 = const()[name = tensor("new_cache_19_internal_tensor_assign_1_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_66 = const()[name = tensor("shape_66"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_18 = const()[name = tensor("reduce_prod_18"), val = tensor(1048576)]; + tensor range_1d_18_start_0 = const()[name = tensor("range_1d_18_start_0"), val = tensor(0)]; + tensor range_1d_18_step_0 = const()[name = tensor("range_1d_18_step_0"), val = tensor(1)]; + tensor range_1d_18 = range_1d(end = reduce_prod_18, start = range_1d_18_start_0, step = range_1d_18_step_0)[name = tensor("range_1d_18")]; + tensor reshape_90 = reshape(shape = shape_66, x = range_1d_18)[name = tensor("reshape_90")]; + tensor slice_by_index_18 = slice_by_index(begin = concat_65, begin_mask = new_cache_19_internal_tensor_assign_1_begin_mask_0, end = concat_66, end_mask = new_cache_19_internal_tensor_assign_1_end_mask_0, squeeze_mask = new_cache_19_internal_tensor_assign_1_squeeze_mask_0, stride = new_cache_19_internal_tensor_assign_1_stride_0, x = reshape_90)[name = tensor("slice_by_index_18")]; + tensor reshape_91_shape_0 = const()[name = tensor("reshape_91_shape_0"), val = tensor([-1])]; + tensor reshape_91 = reshape(shape = reshape_91_shape_0, x = slice_by_index_18)[name = tensor("reshape_91")]; + tensor reshape_92_shape_0 = const()[name = tensor("reshape_92_shape_0"), val = tensor([-1])]; + tensor reshape_92_cast_fp16 = reshape(shape = reshape_92_shape_0, x = var_3909_cast_fp16)[name = tensor("reshape_92_cast_fp16")]; + tensor reshape_93_shape_0 = const()[name = tensor("reshape_93_shape_0"), val = tensor([-1])]; + tensor reshape_93_cast_fp16 = reshape(shape = reshape_93_shape_0, x = cache9_to_fp16)[name = tensor("reshape_93_cast_fp16")]; + tensor scatter_18_mode_0 = const()[name = tensor("scatter_18_mode_0"), val = tensor("update")]; + tensor scatter_18_axis_0 = const()[name = tensor("scatter_18_axis_0"), val = tensor(0)]; + tensor scatter_18_validate_indices_0 = const()[name = tensor("scatter_18_validate_indices_0"), val = tensor(false)]; + tensor scatter_18_cast_fp16 = scatter(axis = scatter_18_axis_0, data = reshape_93_cast_fp16, indices = reshape_91, mode = scatter_18_mode_0, updates = reshape_92_cast_fp16, validate_indices = scatter_18_validate_indices_0)[name = tensor("scatter_18_cast_fp16")]; + tensor reshape_94_cast_fp16 = reshape(shape = shape_66, x = scatter_18_cast_fp16)[name = tensor("reshape_94_cast_fp16")]; + tensor var_3917_begin_0 = const()[name = tensor("op_3917_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor var_3917_end_0 = const()[name = tensor("op_3917_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor var_3917_end_mask_0 = const()[name = tensor("op_3917_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_3917_squeeze_mask_0 = const()[name = tensor("op_3917_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor var_3917_cast_fp16 = slice_by_index(begin = var_3917_begin_0, end = var_3917_end_0, end_mask = var_3917_end_mask_0, squeeze_mask = var_3917_squeeze_mask_0, x = reshape_94_cast_fp16)[name = tensor("op_3917_cast_fp16")]; + tensor var_3919_axis_0 = const()[name = tensor("op_3919_axis_0"), val = tensor(1)]; + tensor var_3919_mode_0 = const()[name = tensor("op_3919_mode_0"), val = tensor("update")]; + tensor var_3919_validate_indices_0 = const()[name = tensor("op_3919_validate_indices_0"), val = tensor(false)]; + tensor var_3919_cast_fp16 = scatter_along_axis(axis = var_3919_axis_0, data = var_3917_cast_fp16, indices = write_indices_19, mode = var_3919_mode_0, updates = v_19_cast_fp16, validate_indices = var_3919_validate_indices_0)[name = tensor("op_3919_cast_fp16")]; + tensor concat_67 = const()[name = tensor("concat_67"), val = tensor([1, 0, 0, 0, 0])]; + tensor concat_68 = const()[name = tensor("concat_68"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_19_internal_tensor_assign_2_stride_0 = const()[name = tensor("new_cache_19_internal_tensor_assign_2_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_19_internal_tensor_assign_2_begin_mask_0 = const()[name = tensor("new_cache_19_internal_tensor_assign_2_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_19_internal_tensor_assign_2_end_mask_0 = const()[name = tensor("new_cache_19_internal_tensor_assign_2_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_19_internal_tensor_assign_2_squeeze_mask_0 = const()[name = tensor("new_cache_19_internal_tensor_assign_2_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_67 = const()[name = tensor("shape_67"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_19 = const()[name = tensor("reduce_prod_19"), val = tensor(1048576)]; + tensor range_1d_19_start_0 = const()[name = tensor("range_1d_19_start_0"), val = tensor(0)]; + tensor range_1d_19_step_0 = const()[name = tensor("range_1d_19_step_0"), val = tensor(1)]; + tensor range_1d_19 = range_1d(end = reduce_prod_19, start = range_1d_19_start_0, step = range_1d_19_step_0)[name = tensor("range_1d_19")]; + tensor reshape_95 = reshape(shape = shape_67, x = range_1d_19)[name = tensor("reshape_95")]; + tensor slice_by_index_19 = slice_by_index(begin = concat_67, begin_mask = new_cache_19_internal_tensor_assign_2_begin_mask_0, end = concat_68, end_mask = new_cache_19_internal_tensor_assign_2_end_mask_0, squeeze_mask = new_cache_19_internal_tensor_assign_2_squeeze_mask_0, stride = new_cache_19_internal_tensor_assign_2_stride_0, x = reshape_95)[name = tensor("slice_by_index_19")]; + tensor reshape_96_shape_0 = const()[name = tensor("reshape_96_shape_0"), val = tensor([-1])]; + tensor reshape_96 = reshape(shape = reshape_96_shape_0, x = slice_by_index_19)[name = tensor("reshape_96")]; + tensor reshape_97_shape_0 = const()[name = tensor("reshape_97_shape_0"), val = tensor([-1])]; + tensor reshape_97_cast_fp16 = reshape(shape = reshape_97_shape_0, x = var_3919_cast_fp16)[name = tensor("reshape_97_cast_fp16")]; + tensor reshape_98_shape_0 = const()[name = tensor("reshape_98_shape_0"), val = tensor([-1])]; + tensor reshape_98_cast_fp16 = reshape(shape = reshape_98_shape_0, x = reshape_94_cast_fp16)[name = tensor("reshape_98_cast_fp16")]; + tensor scatter_19_mode_0 = const()[name = tensor("scatter_19_mode_0"), val = tensor("update")]; + tensor scatter_19_axis_0 = const()[name = tensor("scatter_19_axis_0"), val = tensor(0)]; + tensor scatter_19_validate_indices_0 = const()[name = tensor("scatter_19_validate_indices_0"), val = tensor(false)]; + tensor scatter_19_cast_fp16 = scatter(axis = scatter_19_axis_0, data = reshape_98_cast_fp16, indices = reshape_96, mode = scatter_19_mode_0, updates = reshape_97_cast_fp16, validate_indices = scatter_19_validate_indices_0)[name = tensor("scatter_19_cast_fp16")]; + tensor reshape_99_cast_fp16 = reshape(shape = shape_67, x = scatter_19_cast_fp16)[name = tensor("reshape_99_cast_fp16")]; + tensor reshape_99_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("reshape_99_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor keys_55_begin_0 = const()[name = tensor("keys_55_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor keys_55_end_0 = const()[name = tensor("keys_55_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor keys_55_end_mask_0 = const()[name = tensor("keys_55_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor keys_55_squeeze_mask_0 = const()[name = tensor("keys_55_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor keys_55_cast_fp16 = slice_by_index(begin = keys_55_begin_0, end = keys_55_end_0, end_mask = keys_55_end_mask_0, squeeze_mask = keys_55_squeeze_mask_0, x = reshape_99_cast_fp16)[name = tensor("keys_55_cast_fp16")]; + tensor values_55_begin_0 = const()[name = tensor("values_55_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor values_55_end_0 = const()[name = tensor("values_55_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor values_55_end_mask_0 = const()[name = tensor("values_55_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor values_55_squeeze_mask_0 = const()[name = tensor("values_55_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor values_55_cast_fp16 = slice_by_index(begin = values_55_begin_0, end = values_55_end_0, end_mask = values_55_end_mask_0, squeeze_mask = values_55_squeeze_mask_0, x = reshape_99_cast_fp16)[name = tensor("values_55_cast_fp16")]; + tensor var_3931_cast_fp16 = not_equal(x = keys_55_cast_fp16, y = keys_55_cast_fp16)[name = tensor("op_3931_cast_fp16")]; + tensor keys_57_cast_fp16 = select(a = var_463_to_fp16, b = keys_55_cast_fp16, cond = var_3931_cast_fp16)[name = tensor("keys_57_cast_fp16")]; + tensor var_3939_cast_fp16 = not_equal(x = values_55_cast_fp16, y = values_55_cast_fp16)[name = tensor("op_3939_cast_fp16")]; + tensor values_57_cast_fp16 = select(a = var_463_to_fp16, b = values_55_cast_fp16, cond = var_3939_cast_fp16)[name = tensor("values_57_cast_fp16")]; + tensor var_3963 = const()[name = tensor("op_3963"), val = tensor([0, 2, 1, 3])]; + tensor var_3979 = const()[name = tensor("op_3979"), val = tensor([1, 1, 1])]; + tensor var_3980_cast_fp16 = reshape(shape = var_3979, x = position9_to_fp16)[name = tensor("op_3980_cast_fp16")]; + tensor var_4000_to_fp16 = const()[name = tensor("op_4000_to_fp16"), val = tensor(0x1p+0)]; + tensor valid_len_19_cast_fp16 = add(x = var_3980_cast_fp16, y = var_4000_to_fp16)[name = tensor("valid_len_19_cast_fp16")]; + tensor valid_mask_19_cast_fp16 = less(x = k_positions_1_promoted_to_fp16, y = valid_len_19_cast_fp16)[name = tensor("valid_mask_19_cast_fp16")]; + tensor causal_mask_19_cast_fp16 = less_equal(x = k_positions_1_promoted_to_fp16, y = var_3980_cast_fp16)[name = tensor("causal_mask_19_cast_fp16")]; + tensor attn_mask_37 = logical_and(x = valid_mask_19_cast_fp16, y = causal_mask_19_cast_fp16)[name = tensor("attn_mask_37")]; + tensor attn_mask_39_axes_0 = const()[name = tensor("attn_mask_39_axes_0"), val = tensor([1])]; + tensor attn_mask_39 = expand_dims(axes = attn_mask_39_axes_0, x = attn_mask_37)[name = tensor("attn_mask_39")]; + tensor var_4018_transpose_x_0 = const()[name = tensor("op_4018_transpose_x_0"), val = tensor(false)]; + tensor var_4018_transpose_y_0 = const()[name = tensor("op_4018_transpose_y_0"), val = tensor(false)]; + tensor transpose_90_perm_0 = const()[name = tensor("transpose_90_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_91_perm_0 = const()[name = tensor("transpose_91_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_91 = transpose(perm = transpose_91_perm_0, x = keys_57_cast_fp16)[name = tensor("transpose_177")]; + tensor transpose_90 = transpose(perm = transpose_90_perm_0, x = q_57_cast_fp16)[name = tensor("transpose_178")]; + tensor var_4018_cast_fp16 = matmul(transpose_x = var_4018_transpose_x_0, transpose_y = var_4018_transpose_y_0, x = transpose_90, y = transpose_91)[name = tensor("op_4018_cast_fp16")]; + tensor var_4012_to_fp16 = const()[name = tensor("op_4012_to_fp16"), val = tensor([0x1p-3])]; + tensor attn_weights_55_cast_fp16 = mul(x = var_4018_cast_fp16, y = var_4012_to_fp16)[name = tensor("attn_weights_55_cast_fp16")]; + tensor var_4020 = logical_not(x = attn_mask_39)[name = tensor("op_4020")]; + tensor var_4021_to_fp16 = const()[name = tensor("op_4021_to_fp16"), val = tensor(-inf)]; + tensor attn_weights_57_cast_fp16 = select(a = var_4021_to_fp16, b = attn_weights_55_cast_fp16, cond = var_4020)[name = tensor("attn_weights_57_cast_fp16")]; + tensor var_4023 = const()[name = tensor("op_4023"), val = tensor(-1)]; + tensor attn_weights_59_cast_fp16 = softmax(axis = var_4023, x = attn_weights_57_cast_fp16)[name = tensor("attn_weights_59_cast_fp16")]; + tensor attn_output_19_transpose_x_0 = const()[name = tensor("attn_output_19_transpose_x_0"), val = tensor(false)]; + tensor attn_output_19_transpose_y_0 = const()[name = tensor("attn_output_19_transpose_y_0"), val = tensor(false)]; + tensor values_59_cast_fp16 = transpose(perm = var_3963, x = values_57_cast_fp16)[name = tensor("transpose_179")]; + tensor attn_output_19_cast_fp16 = matmul(transpose_x = attn_output_19_transpose_x_0, transpose_y = attn_output_19_transpose_y_0, x = attn_weights_59_cast_fp16, y = values_59_cast_fp16)[name = tensor("attn_output_19_cast_fp16")]; + tensor var_4031 = const()[name = tensor("op_4031"), val = tensor([0, 2, 1, 3])]; + tensor var_4036 = const()[name = tensor("op_4036"), val = tensor([1, 1, 1024])]; + tensor var_4032_cast_fp16 = transpose(perm = var_4031, x = attn_output_19_cast_fp16)[name = tensor("transpose_176")]; + tensor input_115_cast_fp16 = reshape(shape = var_4036, x = var_4032_cast_fp16)[name = tensor("input_115_cast_fp16")]; + tensor attn9_out_proj_weight_to_fp16 = const()[name = tensor("attn9_out_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(233999744)))]; + tensor linear_38_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = attn9_out_proj_weight_to_fp16, x = input_115_cast_fp16)[name = tensor("linear_38_cast_fp16")]; + tensor var_4042_to_fp16 = const()[name = tensor("op_4042_to_fp16"), val = tensor(0x1p+0)]; + tensor var_4043_cast_fp16 = add(x = position9_to_fp16, y = var_4042_to_fp16)[name = tensor("op_4043_cast_fp16")]; + tensor var_4043_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_4043_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor input_117_cast_fp16 = add(x = input_111_cast_fp16, y = linear_38_cast_fp16)[name = tensor("input_117_cast_fp16")]; + tensor input_119_axes_0 = const()[name = tensor("input_119_axes_0"), val = tensor([-1])]; + tensor norm9_2_weight_to_fp16 = const()[name = tensor("norm9_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(236096960)))]; + tensor norm9_2_bias_to_fp16 = const()[name = tensor("norm9_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(236099072)))]; + tensor var_4047_to_fp16 = const()[name = tensor("op_4047_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_119_cast_fp16 = layer_norm(axes = input_119_axes_0, beta = norm9_2_bias_to_fp16, epsilon = var_4047_to_fp16, gamma = norm9_2_weight_to_fp16, x = input_117_cast_fp16)[name = tensor("input_119_cast_fp16")]; + tensor linear9_1_weight_to_fp16 = const()[name = tensor("linear9_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(236101184)))]; + tensor linear_39_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = linear9_1_weight_to_fp16, x = input_119_cast_fp16)[name = tensor("linear_39_cast_fp16")]; + tensor input_121_mode_0 = const()[name = tensor("input_121_mode_0"), val = tensor("EXACT")]; + tensor input_121_cast_fp16 = gelu(mode = input_121_mode_0, x = linear_39_cast_fp16)[name = tensor("input_121_cast_fp16")]; + tensor linear9_2_weight_to_fp16 = const()[name = tensor("linear9_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(244489856)))]; + tensor linear_40_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = linear9_2_weight_to_fp16, x = input_121_cast_fp16)[name = tensor("linear_40_cast_fp16")]; + tensor input_123_cast_fp16 = add(x = input_117_cast_fp16, y = linear_40_cast_fp16)[name = tensor("input_123_cast_fp16")]; + tensor input_125_axes_0 = const()[name = tensor("input_125_axes_0"), val = tensor([-1])]; + tensor norm10_1_weight_to_fp16 = const()[name = tensor("norm10_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(252878528)))]; + tensor norm10_1_bias_to_fp16 = const()[name = tensor("norm10_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(252880640)))]; + tensor var_4064_to_fp16 = const()[name = tensor("op_4064_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_125_cast_fp16 = layer_norm(axes = input_125_axes_0, beta = norm10_1_bias_to_fp16, epsilon = var_4064_to_fp16, gamma = norm10_1_weight_to_fp16, x = input_123_cast_fp16)[name = tensor("input_125_cast_fp16")]; + tensor attn10_in_proj_weight_to_fp16 = const()[name = tensor("attn10_in_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(252882752)))]; + tensor linear_41_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = attn10_in_proj_weight_to_fp16, x = input_125_cast_fp16)[name = tensor("linear_41_cast_fp16")]; + tensor var_4078 = const()[name = tensor("op_4078"), val = tensor([1, 1, 3, 16, 64])]; + tensor qkv_21_cast_fp16 = reshape(shape = var_4078, x = linear_41_cast_fp16)[name = tensor("qkv_21_cast_fp16")]; + tensor q_61_begin_0 = const()[name = tensor("q_61_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor q_61_end_0 = const()[name = tensor("q_61_end_0"), val = tensor([1, 1, 1, 16, 64])]; + tensor q_61_end_mask_0 = const()[name = tensor("q_61_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor q_61_squeeze_mask_0 = const()[name = tensor("q_61_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor q_61_cast_fp16 = slice_by_index(begin = q_61_begin_0, end = q_61_end_0, end_mask = q_61_end_mask_0, squeeze_mask = q_61_squeeze_mask_0, x = qkv_21_cast_fp16)[name = tensor("q_61_cast_fp16")]; + tensor k_41_begin_0 = const()[name = tensor("k_41_begin_0"), val = tensor([0, 0, 1, 0, 0])]; + tensor k_41_end_0 = const()[name = tensor("k_41_end_0"), val = tensor([1, 1, 2, 16, 64])]; + tensor k_41_end_mask_0 = const()[name = tensor("k_41_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor k_41_squeeze_mask_0 = const()[name = tensor("k_41_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor k_41_cast_fp16 = slice_by_index(begin = k_41_begin_0, end = k_41_end_0, end_mask = k_41_end_mask_0, squeeze_mask = k_41_squeeze_mask_0, x = qkv_21_cast_fp16)[name = tensor("k_41_cast_fp16")]; + tensor v_21_begin_0 = const()[name = tensor("v_21_begin_0"), val = tensor([0, 0, 2, 0, 0])]; + tensor v_21_end_0 = const()[name = tensor("v_21_end_0"), val = tensor([1, 1, 3, 16, 64])]; + tensor v_21_end_mask_0 = const()[name = tensor("v_21_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor v_21_squeeze_mask_0 = const()[name = tensor("v_21_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor v_21_cast_fp16 = slice_by_index(begin = v_21_begin_0, end = v_21_end_0, end_mask = v_21_end_mask_0, squeeze_mask = v_21_squeeze_mask_0, x = qkv_21_cast_fp16)[name = tensor("v_21_cast_fp16")]; + tensor var_4144 = const()[name = tensor("op_4144"), val = tensor([1, 1, 1, 1])]; + tensor position10_to_fp16_dtype_0 = const()[name = tensor("position10_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor position10_to_fp16 = cast(dtype = position10_to_fp16_dtype_0, x = position10)[name = tensor("cast_503")]; + tensor ts_65_cast_fp16 = reshape(shape = var_4144, x = position10_to_fp16)[name = tensor("ts_65_cast_fp16")]; + tensor var_4151 = const()[name = tensor("op_4151"), val = tensor([1, 1, 16, 32, 2])]; + tensor q_complex_21_cast_fp16 = reshape(shape = var_4151, x = q_61_cast_fp16)[name = tensor("q_complex_21_cast_fp16")]; + tensor var_4158 = const()[name = tensor("op_4158"), val = tensor([1, 1, 16, 32, 2])]; + tensor k_complex_21_cast_fp16 = reshape(shape = var_4158, x = k_41_cast_fp16)[name = tensor("k_complex_21_cast_fp16")]; + tensor var_4162_begin_0 = const()[name = tensor("op_4162_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_4162_end_0 = const()[name = tensor("op_4162_end_0"), val = tensor([1, 1, 16, 32, 1])]; + tensor var_4162_end_mask_0 = const()[name = tensor("op_4162_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_4162_squeeze_mask_0 = const()[name = tensor("op_4162_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_4162_cast_fp16 = slice_by_index(begin = var_4162_begin_0, end = var_4162_end_0, end_mask = var_4162_end_mask_0, squeeze_mask = var_4162_squeeze_mask_0, x = q_complex_21_cast_fp16)[name = tensor("op_4162_cast_fp16")]; + tensor var_4170_begin_0 = const()[name = tensor("op_4170_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_4170_end_0 = const()[name = tensor("op_4170_end_0"), val = tensor([1, 1, 16, 32, 2])]; + tensor var_4170_end_mask_0 = const()[name = tensor("op_4170_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_4170_squeeze_mask_0 = const()[name = tensor("op_4170_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_4170_cast_fp16 = slice_by_index(begin = var_4170_begin_0, end = var_4170_end_0, end_mask = var_4170_end_mask_0, squeeze_mask = var_4170_squeeze_mask_0, x = q_complex_21_cast_fp16)[name = tensor("op_4170_cast_fp16")]; + tensor var_4178_begin_0 = const()[name = tensor("op_4178_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_4178_end_0 = const()[name = tensor("op_4178_end_0"), val = tensor([1, 1, 16, 32, 1])]; + tensor var_4178_end_mask_0 = const()[name = tensor("op_4178_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_4178_squeeze_mask_0 = const()[name = tensor("op_4178_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_4178_cast_fp16 = slice_by_index(begin = var_4178_begin_0, end = var_4178_end_0, end_mask = var_4178_end_mask_0, squeeze_mask = var_4178_squeeze_mask_0, x = k_complex_21_cast_fp16)[name = tensor("op_4178_cast_fp16")]; + tensor var_4186_begin_0 = const()[name = tensor("op_4186_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_4186_end_0 = const()[name = tensor("op_4186_end_0"), val = tensor([1, 1, 16, 32, 2])]; + tensor var_4186_end_mask_0 = const()[name = tensor("op_4186_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_4186_squeeze_mask_0 = const()[name = tensor("op_4186_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_4186_cast_fp16 = slice_by_index(begin = var_4186_begin_0, end = var_4186_end_0, end_mask = var_4186_end_mask_0, squeeze_mask = var_4186_squeeze_mask_0, x = k_complex_21_cast_fp16)[name = tensor("op_4186_cast_fp16")]; + tensor freqs_21_to_fp16 = const()[name = tensor("freqs_21_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(259174272)))]; + tensor var_4192_cast_fp16 = mul(x = freqs_21_to_fp16, y = ts_65_cast_fp16)[name = tensor("op_4192_cast_fp16")]; + tensor rotr_21_cast_fp16 = cos(x = var_4192_cast_fp16)[name = tensor("rotr_21_cast_fp16")]; + tensor roti_21_cast_fp16 = sin(x = var_4192_cast_fp16)[name = tensor("roti_21_cast_fp16")]; + tensor var_4196_cast_fp16 = mul(x = var_4162_cast_fp16, y = rotr_21_cast_fp16)[name = tensor("op_4196_cast_fp16")]; + tensor var_4197_cast_fp16 = mul(x = var_4170_cast_fp16, y = roti_21_cast_fp16)[name = tensor("op_4197_cast_fp16")]; + tensor qor_41_cast_fp16 = sub(x = var_4196_cast_fp16, y = var_4197_cast_fp16)[name = tensor("qor_41_cast_fp16")]; + tensor var_4200_cast_fp16 = mul(x = var_4162_cast_fp16, y = roti_21_cast_fp16)[name = tensor("op_4200_cast_fp16")]; + tensor var_4201_cast_fp16 = mul(x = var_4170_cast_fp16, y = rotr_21_cast_fp16)[name = tensor("op_4201_cast_fp16")]; + tensor qoi_41_cast_fp16 = add(x = var_4200_cast_fp16, y = var_4201_cast_fp16)[name = tensor("qoi_41_cast_fp16")]; + tensor var_4204_cast_fp16 = mul(x = var_4178_cast_fp16, y = rotr_21_cast_fp16)[name = tensor("op_4204_cast_fp16")]; + tensor var_4205_cast_fp16 = mul(x = var_4186_cast_fp16, y = roti_21_cast_fp16)[name = tensor("op_4205_cast_fp16")]; + tensor kor_41_cast_fp16 = sub(x = var_4204_cast_fp16, y = var_4205_cast_fp16)[name = tensor("kor_41_cast_fp16")]; + tensor var_4208_cast_fp16 = mul(x = var_4178_cast_fp16, y = roti_21_cast_fp16)[name = tensor("op_4208_cast_fp16")]; + tensor var_4209_cast_fp16 = mul(x = var_4186_cast_fp16, y = rotr_21_cast_fp16)[name = tensor("op_4209_cast_fp16")]; + tensor koi_41_cast_fp16 = add(x = var_4208_cast_fp16, y = var_4209_cast_fp16)[name = tensor("koi_41_cast_fp16")]; + tensor qo_21_axis_0 = const()[name = tensor("qo_21_axis_0"), val = tensor(-1)]; + tensor qo_21_cast_fp16 = stack(axis = qo_21_axis_0, values = (qor_41_cast_fp16, qoi_41_cast_fp16))[name = tensor("qo_21_cast_fp16")]; + tensor ko_21_axis_0 = const()[name = tensor("ko_21_axis_0"), val = tensor(-1)]; + tensor ko_21_cast_fp16 = stack(axis = ko_21_axis_0, values = (kor_41_cast_fp16, koi_41_cast_fp16))[name = tensor("ko_21_cast_fp16")]; + tensor var_4242 = const()[name = tensor("op_4242"), val = tensor([1, 1, 16, 64])]; + tensor q_63_cast_fp16 = reshape(shape = var_4242, x = qo_21_cast_fp16)[name = tensor("q_63_cast_fp16")]; + tensor var_4248 = const()[name = tensor("op_4248"), val = tensor([1, 1, 16, 64])]; + tensor k_43_cast_fp16 = reshape(shape = var_4248, x = ko_21_cast_fp16)[name = tensor("k_43_cast_fp16")]; + tensor _inversed_4273_y_0_to_fp16 = const()[name = tensor("_inversed_4273_y_0_to_fp16"), val = tensor(0x1p-9)]; + tensor _inversed_4273_cast_fp16 = mul(x = ts_65_cast_fp16, y = _inversed_4273_y_0_to_fp16)[name = tensor("_inversed_4273_cast_fp16")]; + tensor var_4274_cast_fp16 = floor(x = _inversed_4273_cast_fp16)[name = tensor("op_4274_cast_fp16")]; + tensor var_4275_to_fp16 = const()[name = tensor("op_4275_to_fp16"), val = tensor(0x1p+9)]; + tensor var_4276_cast_fp16 = mul(x = var_4274_cast_fp16, y = var_4275_to_fp16)[name = tensor("op_4276_cast_fp16")]; + tensor write_indices_float_43_cast_fp16 = sub(x = ts_65_cast_fp16, y = var_4276_cast_fp16)[name = tensor("write_indices_float_43_cast_fp16")]; + tensor var_4283_dtype_0 = const()[name = tensor("op_4283_dtype_0"), val = tensor("int32")]; + tensor write_indices_21_reps_0 = const()[name = tensor("write_indices_21_reps_0"), val = tensor([1, 1, 16, 64])]; + tensor write_indices_float_43_cast_fp16_to_int32 = cast(dtype = var_4283_dtype_0, x = write_indices_float_43_cast_fp16)[name = tensor("cast_502")]; + tensor write_indices_21 = tile(reps = write_indices_21_reps_0, x = write_indices_float_43_cast_fp16_to_int32)[name = tensor("write_indices_21")]; + tensor var_4293_begin_0 = const()[name = tensor("op_4293_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_4293_end_0 = const()[name = tensor("op_4293_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor var_4293_end_mask_0 = const()[name = tensor("op_4293_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_4293_squeeze_mask_0 = const()[name = tensor("op_4293_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor cache10_to_fp16_dtype_0 = const()[name = tensor("cache10_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor cache10_to_fp16 = cast(dtype = cache10_to_fp16_dtype_0, x = cache10)[name = tensor("cast_501")]; + tensor var_4293_cast_fp16 = slice_by_index(begin = var_4293_begin_0, end = var_4293_end_0, end_mask = var_4293_end_mask_0, squeeze_mask = var_4293_squeeze_mask_0, x = cache10_to_fp16)[name = tensor("op_4293_cast_fp16")]; + tensor var_4295_axis_0 = const()[name = tensor("op_4295_axis_0"), val = tensor(1)]; + tensor var_4295_mode_0 = const()[name = tensor("op_4295_mode_0"), val = tensor("update")]; + tensor var_4295_validate_indices_0 = const()[name = tensor("op_4295_validate_indices_0"), val = tensor(false)]; + tensor var_4295_cast_fp16 = scatter_along_axis(axis = var_4295_axis_0, data = var_4293_cast_fp16, indices = write_indices_21, mode = var_4295_mode_0, updates = k_43_cast_fp16, validate_indices = var_4295_validate_indices_0)[name = tensor("op_4295_cast_fp16")]; + tensor concat_72 = const()[name = tensor("concat_72"), val = tensor([0, 0, 0, 0, 0])]; + tensor concat_73 = const()[name = tensor("concat_73"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_21_internal_tensor_assign_1_stride_0 = const()[name = tensor("new_cache_21_internal_tensor_assign_1_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_21_internal_tensor_assign_1_begin_mask_0 = const()[name = tensor("new_cache_21_internal_tensor_assign_1_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_21_internal_tensor_assign_1_end_mask_0 = const()[name = tensor("new_cache_21_internal_tensor_assign_1_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_21_internal_tensor_assign_1_squeeze_mask_0 = const()[name = tensor("new_cache_21_internal_tensor_assign_1_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_68 = const()[name = tensor("shape_68"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_20 = const()[name = tensor("reduce_prod_20"), val = tensor(1048576)]; + tensor range_1d_20_start_0 = const()[name = tensor("range_1d_20_start_0"), val = tensor(0)]; + tensor range_1d_20_step_0 = const()[name = tensor("range_1d_20_step_0"), val = tensor(1)]; + tensor range_1d_20 = range_1d(end = reduce_prod_20, start = range_1d_20_start_0, step = range_1d_20_step_0)[name = tensor("range_1d_20")]; + tensor reshape_100 = reshape(shape = shape_68, x = range_1d_20)[name = tensor("reshape_100")]; + tensor slice_by_index_20 = slice_by_index(begin = concat_72, begin_mask = new_cache_21_internal_tensor_assign_1_begin_mask_0, end = concat_73, end_mask = new_cache_21_internal_tensor_assign_1_end_mask_0, squeeze_mask = new_cache_21_internal_tensor_assign_1_squeeze_mask_0, stride = new_cache_21_internal_tensor_assign_1_stride_0, x = reshape_100)[name = tensor("slice_by_index_20")]; + tensor reshape_101_shape_0 = const()[name = tensor("reshape_101_shape_0"), val = tensor([-1])]; + tensor reshape_101 = reshape(shape = reshape_101_shape_0, x = slice_by_index_20)[name = tensor("reshape_101")]; + tensor reshape_102_shape_0 = const()[name = tensor("reshape_102_shape_0"), val = tensor([-1])]; + tensor reshape_102_cast_fp16 = reshape(shape = reshape_102_shape_0, x = var_4295_cast_fp16)[name = tensor("reshape_102_cast_fp16")]; + tensor reshape_103_shape_0 = const()[name = tensor("reshape_103_shape_0"), val = tensor([-1])]; + tensor reshape_103_cast_fp16 = reshape(shape = reshape_103_shape_0, x = cache10_to_fp16)[name = tensor("reshape_103_cast_fp16")]; + tensor scatter_20_mode_0 = const()[name = tensor("scatter_20_mode_0"), val = tensor("update")]; + tensor scatter_20_axis_0 = const()[name = tensor("scatter_20_axis_0"), val = tensor(0)]; + tensor scatter_20_validate_indices_0 = const()[name = tensor("scatter_20_validate_indices_0"), val = tensor(false)]; + tensor scatter_20_cast_fp16 = scatter(axis = scatter_20_axis_0, data = reshape_103_cast_fp16, indices = reshape_101, mode = scatter_20_mode_0, updates = reshape_102_cast_fp16, validate_indices = scatter_20_validate_indices_0)[name = tensor("scatter_20_cast_fp16")]; + tensor reshape_104_cast_fp16 = reshape(shape = shape_68, x = scatter_20_cast_fp16)[name = tensor("reshape_104_cast_fp16")]; + tensor var_4303_begin_0 = const()[name = tensor("op_4303_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor var_4303_end_0 = const()[name = tensor("op_4303_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor var_4303_end_mask_0 = const()[name = tensor("op_4303_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_4303_squeeze_mask_0 = const()[name = tensor("op_4303_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor var_4303_cast_fp16 = slice_by_index(begin = var_4303_begin_0, end = var_4303_end_0, end_mask = var_4303_end_mask_0, squeeze_mask = var_4303_squeeze_mask_0, x = reshape_104_cast_fp16)[name = tensor("op_4303_cast_fp16")]; + tensor var_4305_axis_0 = const()[name = tensor("op_4305_axis_0"), val = tensor(1)]; + tensor var_4305_mode_0 = const()[name = tensor("op_4305_mode_0"), val = tensor("update")]; + tensor var_4305_validate_indices_0 = const()[name = tensor("op_4305_validate_indices_0"), val = tensor(false)]; + tensor var_4305_cast_fp16 = scatter_along_axis(axis = var_4305_axis_0, data = var_4303_cast_fp16, indices = write_indices_21, mode = var_4305_mode_0, updates = v_21_cast_fp16, validate_indices = var_4305_validate_indices_0)[name = tensor("op_4305_cast_fp16")]; + tensor concat_74 = const()[name = tensor("concat_74"), val = tensor([1, 0, 0, 0, 0])]; + tensor concat_75 = const()[name = tensor("concat_75"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_21_internal_tensor_assign_2_stride_0 = const()[name = tensor("new_cache_21_internal_tensor_assign_2_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_21_internal_tensor_assign_2_begin_mask_0 = const()[name = tensor("new_cache_21_internal_tensor_assign_2_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_21_internal_tensor_assign_2_end_mask_0 = const()[name = tensor("new_cache_21_internal_tensor_assign_2_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_21_internal_tensor_assign_2_squeeze_mask_0 = const()[name = tensor("new_cache_21_internal_tensor_assign_2_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_69 = const()[name = tensor("shape_69"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_21 = const()[name = tensor("reduce_prod_21"), val = tensor(1048576)]; + tensor range_1d_21_start_0 = const()[name = tensor("range_1d_21_start_0"), val = tensor(0)]; + tensor range_1d_21_step_0 = const()[name = tensor("range_1d_21_step_0"), val = tensor(1)]; + tensor range_1d_21 = range_1d(end = reduce_prod_21, start = range_1d_21_start_0, step = range_1d_21_step_0)[name = tensor("range_1d_21")]; + tensor reshape_105 = reshape(shape = shape_69, x = range_1d_21)[name = tensor("reshape_105")]; + tensor slice_by_index_21 = slice_by_index(begin = concat_74, begin_mask = new_cache_21_internal_tensor_assign_2_begin_mask_0, end = concat_75, end_mask = new_cache_21_internal_tensor_assign_2_end_mask_0, squeeze_mask = new_cache_21_internal_tensor_assign_2_squeeze_mask_0, stride = new_cache_21_internal_tensor_assign_2_stride_0, x = reshape_105)[name = tensor("slice_by_index_21")]; + tensor reshape_106_shape_0 = const()[name = tensor("reshape_106_shape_0"), val = tensor([-1])]; + tensor reshape_106 = reshape(shape = reshape_106_shape_0, x = slice_by_index_21)[name = tensor("reshape_106")]; + tensor reshape_107_shape_0 = const()[name = tensor("reshape_107_shape_0"), val = tensor([-1])]; + tensor reshape_107_cast_fp16 = reshape(shape = reshape_107_shape_0, x = var_4305_cast_fp16)[name = tensor("reshape_107_cast_fp16")]; + tensor reshape_108_shape_0 = const()[name = tensor("reshape_108_shape_0"), val = tensor([-1])]; + tensor reshape_108_cast_fp16 = reshape(shape = reshape_108_shape_0, x = reshape_104_cast_fp16)[name = tensor("reshape_108_cast_fp16")]; + tensor scatter_21_mode_0 = const()[name = tensor("scatter_21_mode_0"), val = tensor("update")]; + tensor scatter_21_axis_0 = const()[name = tensor("scatter_21_axis_0"), val = tensor(0)]; + tensor scatter_21_validate_indices_0 = const()[name = tensor("scatter_21_validate_indices_0"), val = tensor(false)]; + tensor scatter_21_cast_fp16 = scatter(axis = scatter_21_axis_0, data = reshape_108_cast_fp16, indices = reshape_106, mode = scatter_21_mode_0, updates = reshape_107_cast_fp16, validate_indices = scatter_21_validate_indices_0)[name = tensor("scatter_21_cast_fp16")]; + tensor reshape_109_cast_fp16 = reshape(shape = shape_69, x = scatter_21_cast_fp16)[name = tensor("reshape_109_cast_fp16")]; + tensor reshape_109_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("reshape_109_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor keys_61_begin_0 = const()[name = tensor("keys_61_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor keys_61_end_0 = const()[name = tensor("keys_61_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor keys_61_end_mask_0 = const()[name = tensor("keys_61_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor keys_61_squeeze_mask_0 = const()[name = tensor("keys_61_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor keys_61_cast_fp16 = slice_by_index(begin = keys_61_begin_0, end = keys_61_end_0, end_mask = keys_61_end_mask_0, squeeze_mask = keys_61_squeeze_mask_0, x = reshape_109_cast_fp16)[name = tensor("keys_61_cast_fp16")]; + tensor values_61_begin_0 = const()[name = tensor("values_61_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor values_61_end_0 = const()[name = tensor("values_61_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor values_61_end_mask_0 = const()[name = tensor("values_61_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor values_61_squeeze_mask_0 = const()[name = tensor("values_61_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor values_61_cast_fp16 = slice_by_index(begin = values_61_begin_0, end = values_61_end_0, end_mask = values_61_end_mask_0, squeeze_mask = values_61_squeeze_mask_0, x = reshape_109_cast_fp16)[name = tensor("values_61_cast_fp16")]; + tensor var_4317_cast_fp16 = not_equal(x = keys_61_cast_fp16, y = keys_61_cast_fp16)[name = tensor("op_4317_cast_fp16")]; + tensor keys_63_cast_fp16 = select(a = var_463_to_fp16, b = keys_61_cast_fp16, cond = var_4317_cast_fp16)[name = tensor("keys_63_cast_fp16")]; + tensor var_4325_cast_fp16 = not_equal(x = values_61_cast_fp16, y = values_61_cast_fp16)[name = tensor("op_4325_cast_fp16")]; + tensor values_63_cast_fp16 = select(a = var_463_to_fp16, b = values_61_cast_fp16, cond = var_4325_cast_fp16)[name = tensor("values_63_cast_fp16")]; + tensor var_4349 = const()[name = tensor("op_4349"), val = tensor([0, 2, 1, 3])]; + tensor var_4365 = const()[name = tensor("op_4365"), val = tensor([1, 1, 1])]; + tensor var_4366_cast_fp16 = reshape(shape = var_4365, x = position10_to_fp16)[name = tensor("op_4366_cast_fp16")]; + tensor var_4386_to_fp16 = const()[name = tensor("op_4386_to_fp16"), val = tensor(0x1p+0)]; + tensor valid_len_21_cast_fp16 = add(x = var_4366_cast_fp16, y = var_4386_to_fp16)[name = tensor("valid_len_21_cast_fp16")]; + tensor valid_mask_21_cast_fp16 = less(x = k_positions_1_promoted_to_fp16, y = valid_len_21_cast_fp16)[name = tensor("valid_mask_21_cast_fp16")]; + tensor causal_mask_21_cast_fp16 = less_equal(x = k_positions_1_promoted_to_fp16, y = var_4366_cast_fp16)[name = tensor("causal_mask_21_cast_fp16")]; + tensor attn_mask_41 = logical_and(x = valid_mask_21_cast_fp16, y = causal_mask_21_cast_fp16)[name = tensor("attn_mask_41")]; + tensor attn_mask_43_axes_0 = const()[name = tensor("attn_mask_43_axes_0"), val = tensor([1])]; + tensor attn_mask_43 = expand_dims(axes = attn_mask_43_axes_0, x = attn_mask_41)[name = tensor("attn_mask_43")]; + tensor var_4404_transpose_x_0 = const()[name = tensor("op_4404_transpose_x_0"), val = tensor(false)]; + tensor var_4404_transpose_y_0 = const()[name = tensor("op_4404_transpose_y_0"), val = tensor(false)]; + tensor transpose_92_perm_0 = const()[name = tensor("transpose_92_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_93_perm_0 = const()[name = tensor("transpose_93_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_93 = transpose(perm = transpose_93_perm_0, x = keys_63_cast_fp16)[name = tensor("transpose_173")]; + tensor transpose_92 = transpose(perm = transpose_92_perm_0, x = q_63_cast_fp16)[name = tensor("transpose_174")]; + tensor var_4404_cast_fp16 = matmul(transpose_x = var_4404_transpose_x_0, transpose_y = var_4404_transpose_y_0, x = transpose_92, y = transpose_93)[name = tensor("op_4404_cast_fp16")]; + tensor var_4398_to_fp16 = const()[name = tensor("op_4398_to_fp16"), val = tensor([0x1p-3])]; + tensor attn_weights_61_cast_fp16 = mul(x = var_4404_cast_fp16, y = var_4398_to_fp16)[name = tensor("attn_weights_61_cast_fp16")]; + tensor var_4406 = logical_not(x = attn_mask_43)[name = tensor("op_4406")]; + tensor var_4407_to_fp16 = const()[name = tensor("op_4407_to_fp16"), val = tensor(-inf)]; + tensor attn_weights_63_cast_fp16 = select(a = var_4407_to_fp16, b = attn_weights_61_cast_fp16, cond = var_4406)[name = tensor("attn_weights_63_cast_fp16")]; + tensor var_4409 = const()[name = tensor("op_4409"), val = tensor(-1)]; + tensor attn_weights_65_cast_fp16 = softmax(axis = var_4409, x = attn_weights_63_cast_fp16)[name = tensor("attn_weights_65_cast_fp16")]; + tensor attn_output_21_transpose_x_0 = const()[name = tensor("attn_output_21_transpose_x_0"), val = tensor(false)]; + tensor attn_output_21_transpose_y_0 = const()[name = tensor("attn_output_21_transpose_y_0"), val = tensor(false)]; + tensor values_65_cast_fp16 = transpose(perm = var_4349, x = values_63_cast_fp16)[name = tensor("transpose_175")]; + tensor attn_output_21_cast_fp16 = matmul(transpose_x = attn_output_21_transpose_x_0, transpose_y = attn_output_21_transpose_y_0, x = attn_weights_65_cast_fp16, y = values_65_cast_fp16)[name = tensor("attn_output_21_cast_fp16")]; + tensor var_4417 = const()[name = tensor("op_4417"), val = tensor([0, 2, 1, 3])]; + tensor var_4422 = const()[name = tensor("op_4422"), val = tensor([1, 1, 1024])]; + tensor var_4418_cast_fp16 = transpose(perm = var_4417, x = attn_output_21_cast_fp16)[name = tensor("transpose_172")]; + tensor input_127_cast_fp16 = reshape(shape = var_4422, x = var_4418_cast_fp16)[name = tensor("input_127_cast_fp16")]; + tensor attn10_out_proj_weight_to_fp16 = const()[name = tensor("attn10_out_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(259174400)))]; + tensor linear_42_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = attn10_out_proj_weight_to_fp16, x = input_127_cast_fp16)[name = tensor("linear_42_cast_fp16")]; + tensor var_4428_to_fp16 = const()[name = tensor("op_4428_to_fp16"), val = tensor(0x1p+0)]; + tensor var_4429_cast_fp16 = add(x = position10_to_fp16, y = var_4428_to_fp16)[name = tensor("op_4429_cast_fp16")]; + tensor var_4429_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_4429_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor input_129_cast_fp16 = add(x = input_123_cast_fp16, y = linear_42_cast_fp16)[name = tensor("input_129_cast_fp16")]; + tensor input_131_axes_0 = const()[name = tensor("input_131_axes_0"), val = tensor([-1])]; + tensor norm10_2_weight_to_fp16 = const()[name = tensor("norm10_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(261271616)))]; + tensor norm10_2_bias_to_fp16 = const()[name = tensor("norm10_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(261273728)))]; + tensor var_4433_to_fp16 = const()[name = tensor("op_4433_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_131_cast_fp16 = layer_norm(axes = input_131_axes_0, beta = norm10_2_bias_to_fp16, epsilon = var_4433_to_fp16, gamma = norm10_2_weight_to_fp16, x = input_129_cast_fp16)[name = tensor("input_131_cast_fp16")]; + tensor linear10_1_weight_to_fp16 = const()[name = tensor("linear10_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(261275840)))]; + tensor linear_43_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = linear10_1_weight_to_fp16, x = input_131_cast_fp16)[name = tensor("linear_43_cast_fp16")]; + tensor input_133_mode_0 = const()[name = tensor("input_133_mode_0"), val = tensor("EXACT")]; + tensor input_133_cast_fp16 = gelu(mode = input_133_mode_0, x = linear_43_cast_fp16)[name = tensor("input_133_cast_fp16")]; + tensor linear10_2_weight_to_fp16 = const()[name = tensor("linear10_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(269664512)))]; + tensor linear_44_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = linear10_2_weight_to_fp16, x = input_133_cast_fp16)[name = tensor("linear_44_cast_fp16")]; + tensor input_135_cast_fp16 = add(x = input_129_cast_fp16, y = linear_44_cast_fp16)[name = tensor("input_135_cast_fp16")]; + tensor input_137_axes_0 = const()[name = tensor("input_137_axes_0"), val = tensor([-1])]; + tensor norm11_1_weight_to_fp16 = const()[name = tensor("norm11_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(278053184)))]; + tensor norm11_1_bias_to_fp16 = const()[name = tensor("norm11_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(278055296)))]; + tensor var_4450_to_fp16 = const()[name = tensor("op_4450_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_137_cast_fp16 = layer_norm(axes = input_137_axes_0, beta = norm11_1_bias_to_fp16, epsilon = var_4450_to_fp16, gamma = norm11_1_weight_to_fp16, x = input_135_cast_fp16)[name = tensor("input_137_cast_fp16")]; + tensor attn11_in_proj_weight_to_fp16 = const()[name = tensor("attn11_in_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(278057408)))]; + tensor linear_45_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = attn11_in_proj_weight_to_fp16, x = input_137_cast_fp16)[name = tensor("linear_45_cast_fp16")]; + tensor var_4464 = const()[name = tensor("op_4464"), val = tensor([1, 1, 3, 16, 64])]; + tensor qkv_23_cast_fp16 = reshape(shape = var_4464, x = linear_45_cast_fp16)[name = tensor("qkv_23_cast_fp16")]; + tensor q_67_begin_0 = const()[name = tensor("q_67_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor q_67_end_0 = const()[name = tensor("q_67_end_0"), val = tensor([1, 1, 1, 16, 64])]; + tensor q_67_end_mask_0 = const()[name = tensor("q_67_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor q_67_squeeze_mask_0 = const()[name = tensor("q_67_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor q_67_cast_fp16 = slice_by_index(begin = q_67_begin_0, end = q_67_end_0, end_mask = q_67_end_mask_0, squeeze_mask = q_67_squeeze_mask_0, x = qkv_23_cast_fp16)[name = tensor("q_67_cast_fp16")]; + tensor k_45_begin_0 = const()[name = tensor("k_45_begin_0"), val = tensor([0, 0, 1, 0, 0])]; + tensor k_45_end_0 = const()[name = tensor("k_45_end_0"), val = tensor([1, 1, 2, 16, 64])]; + tensor k_45_end_mask_0 = const()[name = tensor("k_45_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor k_45_squeeze_mask_0 = const()[name = tensor("k_45_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor k_45_cast_fp16 = slice_by_index(begin = k_45_begin_0, end = k_45_end_0, end_mask = k_45_end_mask_0, squeeze_mask = k_45_squeeze_mask_0, x = qkv_23_cast_fp16)[name = tensor("k_45_cast_fp16")]; + tensor v_23_begin_0 = const()[name = tensor("v_23_begin_0"), val = tensor([0, 0, 2, 0, 0])]; + tensor v_23_end_0 = const()[name = tensor("v_23_end_0"), val = tensor([1, 1, 3, 16, 64])]; + tensor v_23_end_mask_0 = const()[name = tensor("v_23_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor v_23_squeeze_mask_0 = const()[name = tensor("v_23_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor v_23_cast_fp16 = slice_by_index(begin = v_23_begin_0, end = v_23_end_0, end_mask = v_23_end_mask_0, squeeze_mask = v_23_squeeze_mask_0, x = qkv_23_cast_fp16)[name = tensor("v_23_cast_fp16")]; + tensor var_4530 = const()[name = tensor("op_4530"), val = tensor([1, 1, 1, 1])]; + tensor position11_to_fp16_dtype_0 = const()[name = tensor("position11_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor position11_to_fp16 = cast(dtype = position11_to_fp16_dtype_0, x = position11)[name = tensor("cast_498")]; + tensor ts_71_cast_fp16 = reshape(shape = var_4530, x = position11_to_fp16)[name = tensor("ts_71_cast_fp16")]; + tensor var_4537 = const()[name = tensor("op_4537"), val = tensor([1, 1, 16, 32, 2])]; + tensor q_complex_23_cast_fp16 = reshape(shape = var_4537, x = q_67_cast_fp16)[name = tensor("q_complex_23_cast_fp16")]; + tensor var_4544 = const()[name = tensor("op_4544"), val = tensor([1, 1, 16, 32, 2])]; + tensor k_complex_23_cast_fp16 = reshape(shape = var_4544, x = k_45_cast_fp16)[name = tensor("k_complex_23_cast_fp16")]; + tensor var_4548_begin_0 = const()[name = tensor("op_4548_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_4548_end_0 = const()[name = tensor("op_4548_end_0"), val = tensor([1, 1, 16, 32, 1])]; + tensor var_4548_end_mask_0 = const()[name = tensor("op_4548_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_4548_squeeze_mask_0 = const()[name = tensor("op_4548_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_4548_cast_fp16 = slice_by_index(begin = var_4548_begin_0, end = var_4548_end_0, end_mask = var_4548_end_mask_0, squeeze_mask = var_4548_squeeze_mask_0, x = q_complex_23_cast_fp16)[name = tensor("op_4548_cast_fp16")]; + tensor var_4556_begin_0 = const()[name = tensor("op_4556_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_4556_end_0 = const()[name = tensor("op_4556_end_0"), val = tensor([1, 1, 16, 32, 2])]; + tensor var_4556_end_mask_0 = const()[name = tensor("op_4556_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_4556_squeeze_mask_0 = const()[name = tensor("op_4556_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_4556_cast_fp16 = slice_by_index(begin = var_4556_begin_0, end = var_4556_end_0, end_mask = var_4556_end_mask_0, squeeze_mask = var_4556_squeeze_mask_0, x = q_complex_23_cast_fp16)[name = tensor("op_4556_cast_fp16")]; + tensor var_4564_begin_0 = const()[name = tensor("op_4564_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_4564_end_0 = const()[name = tensor("op_4564_end_0"), val = tensor([1, 1, 16, 32, 1])]; + tensor var_4564_end_mask_0 = const()[name = tensor("op_4564_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_4564_squeeze_mask_0 = const()[name = tensor("op_4564_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_4564_cast_fp16 = slice_by_index(begin = var_4564_begin_0, end = var_4564_end_0, end_mask = var_4564_end_mask_0, squeeze_mask = var_4564_squeeze_mask_0, x = k_complex_23_cast_fp16)[name = tensor("op_4564_cast_fp16")]; + tensor var_4572_begin_0 = const()[name = tensor("op_4572_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_4572_end_0 = const()[name = tensor("op_4572_end_0"), val = tensor([1, 1, 16, 32, 2])]; + tensor var_4572_end_mask_0 = const()[name = tensor("op_4572_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_4572_squeeze_mask_0 = const()[name = tensor("op_4572_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_4572_cast_fp16 = slice_by_index(begin = var_4572_begin_0, end = var_4572_end_0, end_mask = var_4572_end_mask_0, squeeze_mask = var_4572_squeeze_mask_0, x = k_complex_23_cast_fp16)[name = tensor("op_4572_cast_fp16")]; + tensor freqs_23_to_fp16 = const()[name = tensor("freqs_23_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(284348928)))]; + tensor var_4578_cast_fp16 = mul(x = freqs_23_to_fp16, y = ts_71_cast_fp16)[name = tensor("op_4578_cast_fp16")]; + tensor rotr_23_cast_fp16 = cos(x = var_4578_cast_fp16)[name = tensor("rotr_23_cast_fp16")]; + tensor roti_23_cast_fp16 = sin(x = var_4578_cast_fp16)[name = tensor("roti_23_cast_fp16")]; + tensor var_4582_cast_fp16 = mul(x = var_4548_cast_fp16, y = rotr_23_cast_fp16)[name = tensor("op_4582_cast_fp16")]; + tensor var_4583_cast_fp16 = mul(x = var_4556_cast_fp16, y = roti_23_cast_fp16)[name = tensor("op_4583_cast_fp16")]; + tensor qor_45_cast_fp16 = sub(x = var_4582_cast_fp16, y = var_4583_cast_fp16)[name = tensor("qor_45_cast_fp16")]; + tensor var_4586_cast_fp16 = mul(x = var_4548_cast_fp16, y = roti_23_cast_fp16)[name = tensor("op_4586_cast_fp16")]; + tensor var_4587_cast_fp16 = mul(x = var_4556_cast_fp16, y = rotr_23_cast_fp16)[name = tensor("op_4587_cast_fp16")]; + tensor qoi_45_cast_fp16 = add(x = var_4586_cast_fp16, y = var_4587_cast_fp16)[name = tensor("qoi_45_cast_fp16")]; + tensor var_4590_cast_fp16 = mul(x = var_4564_cast_fp16, y = rotr_23_cast_fp16)[name = tensor("op_4590_cast_fp16")]; + tensor var_4591_cast_fp16 = mul(x = var_4572_cast_fp16, y = roti_23_cast_fp16)[name = tensor("op_4591_cast_fp16")]; + tensor kor_45_cast_fp16 = sub(x = var_4590_cast_fp16, y = var_4591_cast_fp16)[name = tensor("kor_45_cast_fp16")]; + tensor var_4594_cast_fp16 = mul(x = var_4564_cast_fp16, y = roti_23_cast_fp16)[name = tensor("op_4594_cast_fp16")]; + tensor var_4595_cast_fp16 = mul(x = var_4572_cast_fp16, y = rotr_23_cast_fp16)[name = tensor("op_4595_cast_fp16")]; + tensor koi_45_cast_fp16 = add(x = var_4594_cast_fp16, y = var_4595_cast_fp16)[name = tensor("koi_45_cast_fp16")]; + tensor qo_23_axis_0 = const()[name = tensor("qo_23_axis_0"), val = tensor(-1)]; + tensor qo_23_cast_fp16 = stack(axis = qo_23_axis_0, values = (qor_45_cast_fp16, qoi_45_cast_fp16))[name = tensor("qo_23_cast_fp16")]; + tensor ko_23_axis_0 = const()[name = tensor("ko_23_axis_0"), val = tensor(-1)]; + tensor ko_23_cast_fp16 = stack(axis = ko_23_axis_0, values = (kor_45_cast_fp16, koi_45_cast_fp16))[name = tensor("ko_23_cast_fp16")]; + tensor var_4628 = const()[name = tensor("op_4628"), val = tensor([1, 1, 16, 64])]; + tensor q_69_cast_fp16 = reshape(shape = var_4628, x = qo_23_cast_fp16)[name = tensor("q_69_cast_fp16")]; + tensor var_4634 = const()[name = tensor("op_4634"), val = tensor([1, 1, 16, 64])]; + tensor k_47_cast_fp16 = reshape(shape = var_4634, x = ko_23_cast_fp16)[name = tensor("k_47_cast_fp16")]; + tensor _inversed_4659_y_0_to_fp16 = const()[name = tensor("_inversed_4659_y_0_to_fp16"), val = tensor(0x1p-9)]; + tensor _inversed_4659_cast_fp16 = mul(x = ts_71_cast_fp16, y = _inversed_4659_y_0_to_fp16)[name = tensor("_inversed_4659_cast_fp16")]; + tensor var_4660_cast_fp16 = floor(x = _inversed_4659_cast_fp16)[name = tensor("op_4660_cast_fp16")]; + tensor var_4661_to_fp16 = const()[name = tensor("op_4661_to_fp16"), val = tensor(0x1p+9)]; + tensor var_4662_cast_fp16 = mul(x = var_4660_cast_fp16, y = var_4661_to_fp16)[name = tensor("op_4662_cast_fp16")]; + tensor write_indices_float_47_cast_fp16 = sub(x = ts_71_cast_fp16, y = var_4662_cast_fp16)[name = tensor("write_indices_float_47_cast_fp16")]; + tensor var_4669_dtype_0 = const()[name = tensor("op_4669_dtype_0"), val = tensor("int32")]; + tensor write_indices_23_reps_0 = const()[name = tensor("write_indices_23_reps_0"), val = tensor([1, 1, 16, 64])]; + tensor write_indices_float_47_cast_fp16_to_int32 = cast(dtype = var_4669_dtype_0, x = write_indices_float_47_cast_fp16)[name = tensor("cast_497")]; + tensor write_indices_23 = tile(reps = write_indices_23_reps_0, x = write_indices_float_47_cast_fp16_to_int32)[name = tensor("write_indices_23")]; + tensor var_4679_begin_0 = const()[name = tensor("op_4679_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_4679_end_0 = const()[name = tensor("op_4679_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor var_4679_end_mask_0 = const()[name = tensor("op_4679_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_4679_squeeze_mask_0 = const()[name = tensor("op_4679_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor cache11_to_fp16_dtype_0 = const()[name = tensor("cache11_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor cache11_to_fp16 = cast(dtype = cache11_to_fp16_dtype_0, x = cache11)[name = tensor("cast_496")]; + tensor var_4679_cast_fp16 = slice_by_index(begin = var_4679_begin_0, end = var_4679_end_0, end_mask = var_4679_end_mask_0, squeeze_mask = var_4679_squeeze_mask_0, x = cache11_to_fp16)[name = tensor("op_4679_cast_fp16")]; + tensor var_4681_axis_0 = const()[name = tensor("op_4681_axis_0"), val = tensor(1)]; + tensor var_4681_mode_0 = const()[name = tensor("op_4681_mode_0"), val = tensor("update")]; + tensor var_4681_validate_indices_0 = const()[name = tensor("op_4681_validate_indices_0"), val = tensor(false)]; + tensor var_4681_cast_fp16 = scatter_along_axis(axis = var_4681_axis_0, data = var_4679_cast_fp16, indices = write_indices_23, mode = var_4681_mode_0, updates = k_47_cast_fp16, validate_indices = var_4681_validate_indices_0)[name = tensor("op_4681_cast_fp16")]; + tensor concat_79 = const()[name = tensor("concat_79"), val = tensor([0, 0, 0, 0, 0])]; + tensor concat_80 = const()[name = tensor("concat_80"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_23_internal_tensor_assign_1_stride_0 = const()[name = tensor("new_cache_23_internal_tensor_assign_1_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_23_internal_tensor_assign_1_begin_mask_0 = const()[name = tensor("new_cache_23_internal_tensor_assign_1_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_23_internal_tensor_assign_1_end_mask_0 = const()[name = tensor("new_cache_23_internal_tensor_assign_1_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_23_internal_tensor_assign_1_squeeze_mask_0 = const()[name = tensor("new_cache_23_internal_tensor_assign_1_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_70 = const()[name = tensor("shape_70"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_22 = const()[name = tensor("reduce_prod_22"), val = tensor(1048576)]; + tensor range_1d_22_start_0 = const()[name = tensor("range_1d_22_start_0"), val = tensor(0)]; + tensor range_1d_22_step_0 = const()[name = tensor("range_1d_22_step_0"), val = tensor(1)]; + tensor range_1d_22 = range_1d(end = reduce_prod_22, start = range_1d_22_start_0, step = range_1d_22_step_0)[name = tensor("range_1d_22")]; + tensor reshape_110 = reshape(shape = shape_70, x = range_1d_22)[name = tensor("reshape_110")]; + tensor slice_by_index_22 = slice_by_index(begin = concat_79, begin_mask = new_cache_23_internal_tensor_assign_1_begin_mask_0, end = concat_80, end_mask = new_cache_23_internal_tensor_assign_1_end_mask_0, squeeze_mask = new_cache_23_internal_tensor_assign_1_squeeze_mask_0, stride = new_cache_23_internal_tensor_assign_1_stride_0, x = reshape_110)[name = tensor("slice_by_index_22")]; + tensor reshape_111_shape_0 = const()[name = tensor("reshape_111_shape_0"), val = tensor([-1])]; + tensor reshape_111 = reshape(shape = reshape_111_shape_0, x = slice_by_index_22)[name = tensor("reshape_111")]; + tensor reshape_112_shape_0 = const()[name = tensor("reshape_112_shape_0"), val = tensor([-1])]; + tensor reshape_112_cast_fp16 = reshape(shape = reshape_112_shape_0, x = var_4681_cast_fp16)[name = tensor("reshape_112_cast_fp16")]; + tensor reshape_113_shape_0 = const()[name = tensor("reshape_113_shape_0"), val = tensor([-1])]; + tensor reshape_113_cast_fp16 = reshape(shape = reshape_113_shape_0, x = cache11_to_fp16)[name = tensor("reshape_113_cast_fp16")]; + tensor scatter_22_mode_0 = const()[name = tensor("scatter_22_mode_0"), val = tensor("update")]; + tensor scatter_22_axis_0 = const()[name = tensor("scatter_22_axis_0"), val = tensor(0)]; + tensor scatter_22_validate_indices_0 = const()[name = tensor("scatter_22_validate_indices_0"), val = tensor(false)]; + tensor scatter_22_cast_fp16 = scatter(axis = scatter_22_axis_0, data = reshape_113_cast_fp16, indices = reshape_111, mode = scatter_22_mode_0, updates = reshape_112_cast_fp16, validate_indices = scatter_22_validate_indices_0)[name = tensor("scatter_22_cast_fp16")]; + tensor reshape_114_cast_fp16 = reshape(shape = shape_70, x = scatter_22_cast_fp16)[name = tensor("reshape_114_cast_fp16")]; + tensor var_4689_begin_0 = const()[name = tensor("op_4689_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor var_4689_end_0 = const()[name = tensor("op_4689_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor var_4689_end_mask_0 = const()[name = tensor("op_4689_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_4689_squeeze_mask_0 = const()[name = tensor("op_4689_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor var_4689_cast_fp16 = slice_by_index(begin = var_4689_begin_0, end = var_4689_end_0, end_mask = var_4689_end_mask_0, squeeze_mask = var_4689_squeeze_mask_0, x = reshape_114_cast_fp16)[name = tensor("op_4689_cast_fp16")]; + tensor var_4691_axis_0 = const()[name = tensor("op_4691_axis_0"), val = tensor(1)]; + tensor var_4691_mode_0 = const()[name = tensor("op_4691_mode_0"), val = tensor("update")]; + tensor var_4691_validate_indices_0 = const()[name = tensor("op_4691_validate_indices_0"), val = tensor(false)]; + tensor var_4691_cast_fp16 = scatter_along_axis(axis = var_4691_axis_0, data = var_4689_cast_fp16, indices = write_indices_23, mode = var_4691_mode_0, updates = v_23_cast_fp16, validate_indices = var_4691_validate_indices_0)[name = tensor("op_4691_cast_fp16")]; + tensor concat_81 = const()[name = tensor("concat_81"), val = tensor([1, 0, 0, 0, 0])]; + tensor concat_82 = const()[name = tensor("concat_82"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_23_internal_tensor_assign_2_stride_0 = const()[name = tensor("new_cache_23_internal_tensor_assign_2_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_23_internal_tensor_assign_2_begin_mask_0 = const()[name = tensor("new_cache_23_internal_tensor_assign_2_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_23_internal_tensor_assign_2_end_mask_0 = const()[name = tensor("new_cache_23_internal_tensor_assign_2_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_23_internal_tensor_assign_2_squeeze_mask_0 = const()[name = tensor("new_cache_23_internal_tensor_assign_2_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_71 = const()[name = tensor("shape_71"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_23 = const()[name = tensor("reduce_prod_23"), val = tensor(1048576)]; + tensor range_1d_23_start_0 = const()[name = tensor("range_1d_23_start_0"), val = tensor(0)]; + tensor range_1d_23_step_0 = const()[name = tensor("range_1d_23_step_0"), val = tensor(1)]; + tensor range_1d_23 = range_1d(end = reduce_prod_23, start = range_1d_23_start_0, step = range_1d_23_step_0)[name = tensor("range_1d_23")]; + tensor reshape_115 = reshape(shape = shape_71, x = range_1d_23)[name = tensor("reshape_115")]; + tensor slice_by_index_23 = slice_by_index(begin = concat_81, begin_mask = new_cache_23_internal_tensor_assign_2_begin_mask_0, end = concat_82, end_mask = new_cache_23_internal_tensor_assign_2_end_mask_0, squeeze_mask = new_cache_23_internal_tensor_assign_2_squeeze_mask_0, stride = new_cache_23_internal_tensor_assign_2_stride_0, x = reshape_115)[name = tensor("slice_by_index_23")]; + tensor reshape_116_shape_0 = const()[name = tensor("reshape_116_shape_0"), val = tensor([-1])]; + tensor reshape_116 = reshape(shape = reshape_116_shape_0, x = slice_by_index_23)[name = tensor("reshape_116")]; + tensor reshape_117_shape_0 = const()[name = tensor("reshape_117_shape_0"), val = tensor([-1])]; + tensor reshape_117_cast_fp16 = reshape(shape = reshape_117_shape_0, x = var_4691_cast_fp16)[name = tensor("reshape_117_cast_fp16")]; + tensor reshape_118_shape_0 = const()[name = tensor("reshape_118_shape_0"), val = tensor([-1])]; + tensor reshape_118_cast_fp16 = reshape(shape = reshape_118_shape_0, x = reshape_114_cast_fp16)[name = tensor("reshape_118_cast_fp16")]; + tensor scatter_23_mode_0 = const()[name = tensor("scatter_23_mode_0"), val = tensor("update")]; + tensor scatter_23_axis_0 = const()[name = tensor("scatter_23_axis_0"), val = tensor(0)]; + tensor scatter_23_validate_indices_0 = const()[name = tensor("scatter_23_validate_indices_0"), val = tensor(false)]; + tensor scatter_23_cast_fp16 = scatter(axis = scatter_23_axis_0, data = reshape_118_cast_fp16, indices = reshape_116, mode = scatter_23_mode_0, updates = reshape_117_cast_fp16, validate_indices = scatter_23_validate_indices_0)[name = tensor("scatter_23_cast_fp16")]; + tensor reshape_119_cast_fp16 = reshape(shape = shape_71, x = scatter_23_cast_fp16)[name = tensor("reshape_119_cast_fp16")]; + tensor reshape_119_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("reshape_119_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor keys_67_begin_0 = const()[name = tensor("keys_67_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor keys_67_end_0 = const()[name = tensor("keys_67_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor keys_67_end_mask_0 = const()[name = tensor("keys_67_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor keys_67_squeeze_mask_0 = const()[name = tensor("keys_67_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor keys_67_cast_fp16 = slice_by_index(begin = keys_67_begin_0, end = keys_67_end_0, end_mask = keys_67_end_mask_0, squeeze_mask = keys_67_squeeze_mask_0, x = reshape_119_cast_fp16)[name = tensor("keys_67_cast_fp16")]; + tensor values_67_begin_0 = const()[name = tensor("values_67_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor values_67_end_0 = const()[name = tensor("values_67_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor values_67_end_mask_0 = const()[name = tensor("values_67_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor values_67_squeeze_mask_0 = const()[name = tensor("values_67_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor values_67_cast_fp16 = slice_by_index(begin = values_67_begin_0, end = values_67_end_0, end_mask = values_67_end_mask_0, squeeze_mask = values_67_squeeze_mask_0, x = reshape_119_cast_fp16)[name = tensor("values_67_cast_fp16")]; + tensor var_4703_cast_fp16 = not_equal(x = keys_67_cast_fp16, y = keys_67_cast_fp16)[name = tensor("op_4703_cast_fp16")]; + tensor keys_69_cast_fp16 = select(a = var_463_to_fp16, b = keys_67_cast_fp16, cond = var_4703_cast_fp16)[name = tensor("keys_69_cast_fp16")]; + tensor var_4711_cast_fp16 = not_equal(x = values_67_cast_fp16, y = values_67_cast_fp16)[name = tensor("op_4711_cast_fp16")]; + tensor values_69_cast_fp16 = select(a = var_463_to_fp16, b = values_67_cast_fp16, cond = var_4711_cast_fp16)[name = tensor("values_69_cast_fp16")]; + tensor var_4735 = const()[name = tensor("op_4735"), val = tensor([0, 2, 1, 3])]; + tensor var_4751 = const()[name = tensor("op_4751"), val = tensor([1, 1, 1])]; + tensor var_4752_cast_fp16 = reshape(shape = var_4751, x = position11_to_fp16)[name = tensor("op_4752_cast_fp16")]; + tensor var_4772_to_fp16 = const()[name = tensor("op_4772_to_fp16"), val = tensor(0x1p+0)]; + tensor valid_len_23_cast_fp16 = add(x = var_4752_cast_fp16, y = var_4772_to_fp16)[name = tensor("valid_len_23_cast_fp16")]; + tensor valid_mask_23_cast_fp16 = less(x = k_positions_1_promoted_to_fp16, y = valid_len_23_cast_fp16)[name = tensor("valid_mask_23_cast_fp16")]; + tensor causal_mask_23_cast_fp16 = less_equal(x = k_positions_1_promoted_to_fp16, y = var_4752_cast_fp16)[name = tensor("causal_mask_23_cast_fp16")]; + tensor attn_mask_45 = logical_and(x = valid_mask_23_cast_fp16, y = causal_mask_23_cast_fp16)[name = tensor("attn_mask_45")]; + tensor attn_mask_47_axes_0 = const()[name = tensor("attn_mask_47_axes_0"), val = tensor([1])]; + tensor attn_mask_47 = expand_dims(axes = attn_mask_47_axes_0, x = attn_mask_45)[name = tensor("attn_mask_47")]; + tensor var_4790_transpose_x_0 = const()[name = tensor("op_4790_transpose_x_0"), val = tensor(false)]; + tensor var_4790_transpose_y_0 = const()[name = tensor("op_4790_transpose_y_0"), val = tensor(false)]; + tensor transpose_94_perm_0 = const()[name = tensor("transpose_94_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_95_perm_0 = const()[name = tensor("transpose_95_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_95 = transpose(perm = transpose_95_perm_0, x = keys_69_cast_fp16)[name = tensor("transpose_169")]; + tensor transpose_94 = transpose(perm = transpose_94_perm_0, x = q_69_cast_fp16)[name = tensor("transpose_170")]; + tensor var_4790_cast_fp16 = matmul(transpose_x = var_4790_transpose_x_0, transpose_y = var_4790_transpose_y_0, x = transpose_94, y = transpose_95)[name = tensor("op_4790_cast_fp16")]; + tensor var_4784_to_fp16 = const()[name = tensor("op_4784_to_fp16"), val = tensor([0x1p-3])]; + tensor attn_weights_67_cast_fp16 = mul(x = var_4790_cast_fp16, y = var_4784_to_fp16)[name = tensor("attn_weights_67_cast_fp16")]; + tensor var_4792 = logical_not(x = attn_mask_47)[name = tensor("op_4792")]; + tensor var_4793_to_fp16 = const()[name = tensor("op_4793_to_fp16"), val = tensor(-inf)]; + tensor attn_weights_69_cast_fp16 = select(a = var_4793_to_fp16, b = attn_weights_67_cast_fp16, cond = var_4792)[name = tensor("attn_weights_69_cast_fp16")]; + tensor var_4795 = const()[name = tensor("op_4795"), val = tensor(-1)]; + tensor attn_weights_71_cast_fp16 = softmax(axis = var_4795, x = attn_weights_69_cast_fp16)[name = tensor("attn_weights_71_cast_fp16")]; + tensor attn_output_23_transpose_x_0 = const()[name = tensor("attn_output_23_transpose_x_0"), val = tensor(false)]; + tensor attn_output_23_transpose_y_0 = const()[name = tensor("attn_output_23_transpose_y_0"), val = tensor(false)]; + tensor values_71_cast_fp16 = transpose(perm = var_4735, x = values_69_cast_fp16)[name = tensor("transpose_171")]; + tensor attn_output_23_cast_fp16 = matmul(transpose_x = attn_output_23_transpose_x_0, transpose_y = attn_output_23_transpose_y_0, x = attn_weights_71_cast_fp16, y = values_71_cast_fp16)[name = tensor("attn_output_23_cast_fp16")]; + tensor var_4803 = const()[name = tensor("op_4803"), val = tensor([0, 2, 1, 3])]; + tensor var_4808 = const()[name = tensor("op_4808"), val = tensor([1, 1, 1024])]; + tensor var_4804_cast_fp16 = transpose(perm = var_4803, x = attn_output_23_cast_fp16)[name = tensor("transpose_168")]; + tensor input_139_cast_fp16 = reshape(shape = var_4808, x = var_4804_cast_fp16)[name = tensor("input_139_cast_fp16")]; + tensor attn11_out_proj_weight_to_fp16 = const()[name = tensor("attn11_out_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(284349056)))]; + tensor linear_46_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = attn11_out_proj_weight_to_fp16, x = input_139_cast_fp16)[name = tensor("linear_46_cast_fp16")]; + tensor var_4814_to_fp16 = const()[name = tensor("op_4814_to_fp16"), val = tensor(0x1p+0)]; + tensor var_4815_cast_fp16 = add(x = position11_to_fp16, y = var_4814_to_fp16)[name = tensor("op_4815_cast_fp16")]; + tensor var_4815_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_4815_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor input_141_cast_fp16 = add(x = input_135_cast_fp16, y = linear_46_cast_fp16)[name = tensor("input_141_cast_fp16")]; + tensor input_143_axes_0 = const()[name = tensor("input_143_axes_0"), val = tensor([-1])]; + tensor norm11_2_weight_to_fp16 = const()[name = tensor("norm11_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(286446272)))]; + tensor norm11_2_bias_to_fp16 = const()[name = tensor("norm11_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(286448384)))]; + tensor var_4819_to_fp16 = const()[name = tensor("op_4819_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_143_cast_fp16 = layer_norm(axes = input_143_axes_0, beta = norm11_2_bias_to_fp16, epsilon = var_4819_to_fp16, gamma = norm11_2_weight_to_fp16, x = input_141_cast_fp16)[name = tensor("input_143_cast_fp16")]; + tensor linear11_1_weight_to_fp16 = const()[name = tensor("linear11_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(286450496)))]; + tensor linear_47_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = linear11_1_weight_to_fp16, x = input_143_cast_fp16)[name = tensor("linear_47_cast_fp16")]; + tensor input_145_mode_0 = const()[name = tensor("input_145_mode_0"), val = tensor("EXACT")]; + tensor input_145_cast_fp16 = gelu(mode = input_145_mode_0, x = linear_47_cast_fp16)[name = tensor("input_145_cast_fp16")]; + tensor linear11_2_weight_to_fp16 = const()[name = tensor("linear11_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(294839168)))]; + tensor linear_48_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = linear11_2_weight_to_fp16, x = input_145_cast_fp16)[name = tensor("linear_48_cast_fp16")]; + tensor input_147_cast_fp16 = add(x = input_141_cast_fp16, y = linear_48_cast_fp16)[name = tensor("input_147_cast_fp16")]; + tensor input_149_axes_0 = const()[name = tensor("input_149_axes_0"), val = tensor([-1])]; + tensor norm12_1_weight_to_fp16 = const()[name = tensor("norm12_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(303227840)))]; + tensor norm12_1_bias_to_fp16 = const()[name = tensor("norm12_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(303229952)))]; + tensor var_4836_to_fp16 = const()[name = tensor("op_4836_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_149_cast_fp16 = layer_norm(axes = input_149_axes_0, beta = norm12_1_bias_to_fp16, epsilon = var_4836_to_fp16, gamma = norm12_1_weight_to_fp16, x = input_147_cast_fp16)[name = tensor("input_149_cast_fp16")]; + tensor attn12_in_proj_weight_to_fp16 = const()[name = tensor("attn12_in_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(303232064)))]; + tensor linear_49_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = attn12_in_proj_weight_to_fp16, x = input_149_cast_fp16)[name = tensor("linear_49_cast_fp16")]; + tensor var_4850 = const()[name = tensor("op_4850"), val = tensor([1, 1, 3, 16, 64])]; + tensor qkv_25_cast_fp16 = reshape(shape = var_4850, x = linear_49_cast_fp16)[name = tensor("qkv_25_cast_fp16")]; + tensor q_73_begin_0 = const()[name = tensor("q_73_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor q_73_end_0 = const()[name = tensor("q_73_end_0"), val = tensor([1, 1, 1, 16, 64])]; + tensor q_73_end_mask_0 = const()[name = tensor("q_73_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor q_73_squeeze_mask_0 = const()[name = tensor("q_73_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor q_73_cast_fp16 = slice_by_index(begin = q_73_begin_0, end = q_73_end_0, end_mask = q_73_end_mask_0, squeeze_mask = q_73_squeeze_mask_0, x = qkv_25_cast_fp16)[name = tensor("q_73_cast_fp16")]; + tensor k_49_begin_0 = const()[name = tensor("k_49_begin_0"), val = tensor([0, 0, 1, 0, 0])]; + tensor k_49_end_0 = const()[name = tensor("k_49_end_0"), val = tensor([1, 1, 2, 16, 64])]; + tensor k_49_end_mask_0 = const()[name = tensor("k_49_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor k_49_squeeze_mask_0 = const()[name = tensor("k_49_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor k_49_cast_fp16 = slice_by_index(begin = k_49_begin_0, end = k_49_end_0, end_mask = k_49_end_mask_0, squeeze_mask = k_49_squeeze_mask_0, x = qkv_25_cast_fp16)[name = tensor("k_49_cast_fp16")]; + tensor v_25_begin_0 = const()[name = tensor("v_25_begin_0"), val = tensor([0, 0, 2, 0, 0])]; + tensor v_25_end_0 = const()[name = tensor("v_25_end_0"), val = tensor([1, 1, 3, 16, 64])]; + tensor v_25_end_mask_0 = const()[name = tensor("v_25_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor v_25_squeeze_mask_0 = const()[name = tensor("v_25_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor v_25_cast_fp16 = slice_by_index(begin = v_25_begin_0, end = v_25_end_0, end_mask = v_25_end_mask_0, squeeze_mask = v_25_squeeze_mask_0, x = qkv_25_cast_fp16)[name = tensor("v_25_cast_fp16")]; + tensor var_4916 = const()[name = tensor("op_4916"), val = tensor([1, 1, 1, 1])]; + tensor position12_to_fp16_dtype_0 = const()[name = tensor("position12_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor position12_to_fp16 = cast(dtype = position12_to_fp16_dtype_0, x = position12)[name = tensor("cast_493")]; + tensor ts_77_cast_fp16 = reshape(shape = var_4916, x = position12_to_fp16)[name = tensor("ts_77_cast_fp16")]; + tensor var_4923 = const()[name = tensor("op_4923"), val = tensor([1, 1, 16, 32, 2])]; + tensor q_complex_25_cast_fp16 = reshape(shape = var_4923, x = q_73_cast_fp16)[name = tensor("q_complex_25_cast_fp16")]; + tensor var_4930 = const()[name = tensor("op_4930"), val = tensor([1, 1, 16, 32, 2])]; + tensor k_complex_25_cast_fp16 = reshape(shape = var_4930, x = k_49_cast_fp16)[name = tensor("k_complex_25_cast_fp16")]; + tensor var_4934_begin_0 = const()[name = tensor("op_4934_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_4934_end_0 = const()[name = tensor("op_4934_end_0"), val = tensor([1, 1, 16, 32, 1])]; + tensor var_4934_end_mask_0 = const()[name = tensor("op_4934_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_4934_squeeze_mask_0 = const()[name = tensor("op_4934_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_4934_cast_fp16 = slice_by_index(begin = var_4934_begin_0, end = var_4934_end_0, end_mask = var_4934_end_mask_0, squeeze_mask = var_4934_squeeze_mask_0, x = q_complex_25_cast_fp16)[name = tensor("op_4934_cast_fp16")]; + tensor var_4942_begin_0 = const()[name = tensor("op_4942_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_4942_end_0 = const()[name = tensor("op_4942_end_0"), val = tensor([1, 1, 16, 32, 2])]; + tensor var_4942_end_mask_0 = const()[name = tensor("op_4942_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_4942_squeeze_mask_0 = const()[name = tensor("op_4942_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_4942_cast_fp16 = slice_by_index(begin = var_4942_begin_0, end = var_4942_end_0, end_mask = var_4942_end_mask_0, squeeze_mask = var_4942_squeeze_mask_0, x = q_complex_25_cast_fp16)[name = tensor("op_4942_cast_fp16")]; + tensor var_4950_begin_0 = const()[name = tensor("op_4950_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_4950_end_0 = const()[name = tensor("op_4950_end_0"), val = tensor([1, 1, 16, 32, 1])]; + tensor var_4950_end_mask_0 = const()[name = tensor("op_4950_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_4950_squeeze_mask_0 = const()[name = tensor("op_4950_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_4950_cast_fp16 = slice_by_index(begin = var_4950_begin_0, end = var_4950_end_0, end_mask = var_4950_end_mask_0, squeeze_mask = var_4950_squeeze_mask_0, x = k_complex_25_cast_fp16)[name = tensor("op_4950_cast_fp16")]; + tensor var_4958_begin_0 = const()[name = tensor("op_4958_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_4958_end_0 = const()[name = tensor("op_4958_end_0"), val = tensor([1, 1, 16, 32, 2])]; + tensor var_4958_end_mask_0 = const()[name = tensor("op_4958_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_4958_squeeze_mask_0 = const()[name = tensor("op_4958_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_4958_cast_fp16 = slice_by_index(begin = var_4958_begin_0, end = var_4958_end_0, end_mask = var_4958_end_mask_0, squeeze_mask = var_4958_squeeze_mask_0, x = k_complex_25_cast_fp16)[name = tensor("op_4958_cast_fp16")]; + tensor freqs_25_to_fp16 = const()[name = tensor("freqs_25_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(309523584)))]; + tensor var_4964_cast_fp16 = mul(x = freqs_25_to_fp16, y = ts_77_cast_fp16)[name = tensor("op_4964_cast_fp16")]; + tensor rotr_25_cast_fp16 = cos(x = var_4964_cast_fp16)[name = tensor("rotr_25_cast_fp16")]; + tensor roti_25_cast_fp16 = sin(x = var_4964_cast_fp16)[name = tensor("roti_25_cast_fp16")]; + tensor var_4968_cast_fp16 = mul(x = var_4934_cast_fp16, y = rotr_25_cast_fp16)[name = tensor("op_4968_cast_fp16")]; + tensor var_4969_cast_fp16 = mul(x = var_4942_cast_fp16, y = roti_25_cast_fp16)[name = tensor("op_4969_cast_fp16")]; + tensor qor_49_cast_fp16 = sub(x = var_4968_cast_fp16, y = var_4969_cast_fp16)[name = tensor("qor_49_cast_fp16")]; + tensor var_4972_cast_fp16 = mul(x = var_4934_cast_fp16, y = roti_25_cast_fp16)[name = tensor("op_4972_cast_fp16")]; + tensor var_4973_cast_fp16 = mul(x = var_4942_cast_fp16, y = rotr_25_cast_fp16)[name = tensor("op_4973_cast_fp16")]; + tensor qoi_49_cast_fp16 = add(x = var_4972_cast_fp16, y = var_4973_cast_fp16)[name = tensor("qoi_49_cast_fp16")]; + tensor var_4976_cast_fp16 = mul(x = var_4950_cast_fp16, y = rotr_25_cast_fp16)[name = tensor("op_4976_cast_fp16")]; + tensor var_4977_cast_fp16 = mul(x = var_4958_cast_fp16, y = roti_25_cast_fp16)[name = tensor("op_4977_cast_fp16")]; + tensor kor_49_cast_fp16 = sub(x = var_4976_cast_fp16, y = var_4977_cast_fp16)[name = tensor("kor_49_cast_fp16")]; + tensor var_4980_cast_fp16 = mul(x = var_4950_cast_fp16, y = roti_25_cast_fp16)[name = tensor("op_4980_cast_fp16")]; + tensor var_4981_cast_fp16 = mul(x = var_4958_cast_fp16, y = rotr_25_cast_fp16)[name = tensor("op_4981_cast_fp16")]; + tensor koi_49_cast_fp16 = add(x = var_4980_cast_fp16, y = var_4981_cast_fp16)[name = tensor("koi_49_cast_fp16")]; + tensor qo_25_axis_0 = const()[name = tensor("qo_25_axis_0"), val = tensor(-1)]; + tensor qo_25_cast_fp16 = stack(axis = qo_25_axis_0, values = (qor_49_cast_fp16, qoi_49_cast_fp16))[name = tensor("qo_25_cast_fp16")]; + tensor ko_25_axis_0 = const()[name = tensor("ko_25_axis_0"), val = tensor(-1)]; + tensor ko_25_cast_fp16 = stack(axis = ko_25_axis_0, values = (kor_49_cast_fp16, koi_49_cast_fp16))[name = tensor("ko_25_cast_fp16")]; + tensor var_5014 = const()[name = tensor("op_5014"), val = tensor([1, 1, 16, 64])]; + tensor q_75_cast_fp16 = reshape(shape = var_5014, x = qo_25_cast_fp16)[name = tensor("q_75_cast_fp16")]; + tensor var_5020 = const()[name = tensor("op_5020"), val = tensor([1, 1, 16, 64])]; + tensor k_51_cast_fp16 = reshape(shape = var_5020, x = ko_25_cast_fp16)[name = tensor("k_51_cast_fp16")]; + tensor _inversed_5045_y_0_to_fp16 = const()[name = tensor("_inversed_5045_y_0_to_fp16"), val = tensor(0x1p-9)]; + tensor _inversed_5045_cast_fp16 = mul(x = ts_77_cast_fp16, y = _inversed_5045_y_0_to_fp16)[name = tensor("_inversed_5045_cast_fp16")]; + tensor var_5046_cast_fp16 = floor(x = _inversed_5045_cast_fp16)[name = tensor("op_5046_cast_fp16")]; + tensor var_5047_to_fp16 = const()[name = tensor("op_5047_to_fp16"), val = tensor(0x1p+9)]; + tensor var_5048_cast_fp16 = mul(x = var_5046_cast_fp16, y = var_5047_to_fp16)[name = tensor("op_5048_cast_fp16")]; + tensor write_indices_float_51_cast_fp16 = sub(x = ts_77_cast_fp16, y = var_5048_cast_fp16)[name = tensor("write_indices_float_51_cast_fp16")]; + tensor var_5055_dtype_0 = const()[name = tensor("op_5055_dtype_0"), val = tensor("int32")]; + tensor write_indices_25_reps_0 = const()[name = tensor("write_indices_25_reps_0"), val = tensor([1, 1, 16, 64])]; + tensor write_indices_float_51_cast_fp16_to_int32 = cast(dtype = var_5055_dtype_0, x = write_indices_float_51_cast_fp16)[name = tensor("cast_492")]; + tensor write_indices_25 = tile(reps = write_indices_25_reps_0, x = write_indices_float_51_cast_fp16_to_int32)[name = tensor("write_indices_25")]; + tensor var_5065_begin_0 = const()[name = tensor("op_5065_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_5065_end_0 = const()[name = tensor("op_5065_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor var_5065_end_mask_0 = const()[name = tensor("op_5065_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_5065_squeeze_mask_0 = const()[name = tensor("op_5065_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor cache12_to_fp16_dtype_0 = const()[name = tensor("cache12_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor cache12_to_fp16 = cast(dtype = cache12_to_fp16_dtype_0, x = cache12)[name = tensor("cast_491")]; + tensor var_5065_cast_fp16 = slice_by_index(begin = var_5065_begin_0, end = var_5065_end_0, end_mask = var_5065_end_mask_0, squeeze_mask = var_5065_squeeze_mask_0, x = cache12_to_fp16)[name = tensor("op_5065_cast_fp16")]; + tensor var_5067_axis_0 = const()[name = tensor("op_5067_axis_0"), val = tensor(1)]; + tensor var_5067_mode_0 = const()[name = tensor("op_5067_mode_0"), val = tensor("update")]; + tensor var_5067_validate_indices_0 = const()[name = tensor("op_5067_validate_indices_0"), val = tensor(false)]; + tensor var_5067_cast_fp16 = scatter_along_axis(axis = var_5067_axis_0, data = var_5065_cast_fp16, indices = write_indices_25, mode = var_5067_mode_0, updates = k_51_cast_fp16, validate_indices = var_5067_validate_indices_0)[name = tensor("op_5067_cast_fp16")]; + tensor concat_86 = const()[name = tensor("concat_86"), val = tensor([0, 0, 0, 0, 0])]; + tensor concat_87 = const()[name = tensor("concat_87"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_25_internal_tensor_assign_1_stride_0 = const()[name = tensor("new_cache_25_internal_tensor_assign_1_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_25_internal_tensor_assign_1_begin_mask_0 = const()[name = tensor("new_cache_25_internal_tensor_assign_1_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_25_internal_tensor_assign_1_end_mask_0 = const()[name = tensor("new_cache_25_internal_tensor_assign_1_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_25_internal_tensor_assign_1_squeeze_mask_0 = const()[name = tensor("new_cache_25_internal_tensor_assign_1_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_72 = const()[name = tensor("shape_72"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_24 = const()[name = tensor("reduce_prod_24"), val = tensor(1048576)]; + tensor range_1d_24_start_0 = const()[name = tensor("range_1d_24_start_0"), val = tensor(0)]; + tensor range_1d_24_step_0 = const()[name = tensor("range_1d_24_step_0"), val = tensor(1)]; + tensor range_1d_24 = range_1d(end = reduce_prod_24, start = range_1d_24_start_0, step = range_1d_24_step_0)[name = tensor("range_1d_24")]; + tensor reshape_120 = reshape(shape = shape_72, x = range_1d_24)[name = tensor("reshape_120")]; + tensor slice_by_index_24 = slice_by_index(begin = concat_86, begin_mask = new_cache_25_internal_tensor_assign_1_begin_mask_0, end = concat_87, end_mask = new_cache_25_internal_tensor_assign_1_end_mask_0, squeeze_mask = new_cache_25_internal_tensor_assign_1_squeeze_mask_0, stride = new_cache_25_internal_tensor_assign_1_stride_0, x = reshape_120)[name = tensor("slice_by_index_24")]; + tensor reshape_121_shape_0 = const()[name = tensor("reshape_121_shape_0"), val = tensor([-1])]; + tensor reshape_121 = reshape(shape = reshape_121_shape_0, x = slice_by_index_24)[name = tensor("reshape_121")]; + tensor reshape_122_shape_0 = const()[name = tensor("reshape_122_shape_0"), val = tensor([-1])]; + tensor reshape_122_cast_fp16 = reshape(shape = reshape_122_shape_0, x = var_5067_cast_fp16)[name = tensor("reshape_122_cast_fp16")]; + tensor reshape_123_shape_0 = const()[name = tensor("reshape_123_shape_0"), val = tensor([-1])]; + tensor reshape_123_cast_fp16 = reshape(shape = reshape_123_shape_0, x = cache12_to_fp16)[name = tensor("reshape_123_cast_fp16")]; + tensor scatter_24_mode_0 = const()[name = tensor("scatter_24_mode_0"), val = tensor("update")]; + tensor scatter_24_axis_0 = const()[name = tensor("scatter_24_axis_0"), val = tensor(0)]; + tensor scatter_24_validate_indices_0 = const()[name = tensor("scatter_24_validate_indices_0"), val = tensor(false)]; + tensor scatter_24_cast_fp16 = scatter(axis = scatter_24_axis_0, data = reshape_123_cast_fp16, indices = reshape_121, mode = scatter_24_mode_0, updates = reshape_122_cast_fp16, validate_indices = scatter_24_validate_indices_0)[name = tensor("scatter_24_cast_fp16")]; + tensor reshape_124_cast_fp16 = reshape(shape = shape_72, x = scatter_24_cast_fp16)[name = tensor("reshape_124_cast_fp16")]; + tensor var_5075_begin_0 = const()[name = tensor("op_5075_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor var_5075_end_0 = const()[name = tensor("op_5075_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor var_5075_end_mask_0 = const()[name = tensor("op_5075_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_5075_squeeze_mask_0 = const()[name = tensor("op_5075_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor var_5075_cast_fp16 = slice_by_index(begin = var_5075_begin_0, end = var_5075_end_0, end_mask = var_5075_end_mask_0, squeeze_mask = var_5075_squeeze_mask_0, x = reshape_124_cast_fp16)[name = tensor("op_5075_cast_fp16")]; + tensor var_5077_axis_0 = const()[name = tensor("op_5077_axis_0"), val = tensor(1)]; + tensor var_5077_mode_0 = const()[name = tensor("op_5077_mode_0"), val = tensor("update")]; + tensor var_5077_validate_indices_0 = const()[name = tensor("op_5077_validate_indices_0"), val = tensor(false)]; + tensor var_5077_cast_fp16 = scatter_along_axis(axis = var_5077_axis_0, data = var_5075_cast_fp16, indices = write_indices_25, mode = var_5077_mode_0, updates = v_25_cast_fp16, validate_indices = var_5077_validate_indices_0)[name = tensor("op_5077_cast_fp16")]; + tensor concat_88 = const()[name = tensor("concat_88"), val = tensor([1, 0, 0, 0, 0])]; + tensor concat_89 = const()[name = tensor("concat_89"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_25_internal_tensor_assign_2_stride_0 = const()[name = tensor("new_cache_25_internal_tensor_assign_2_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_25_internal_tensor_assign_2_begin_mask_0 = const()[name = tensor("new_cache_25_internal_tensor_assign_2_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_25_internal_tensor_assign_2_end_mask_0 = const()[name = tensor("new_cache_25_internal_tensor_assign_2_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_25_internal_tensor_assign_2_squeeze_mask_0 = const()[name = tensor("new_cache_25_internal_tensor_assign_2_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_73 = const()[name = tensor("shape_73"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_25 = const()[name = tensor("reduce_prod_25"), val = tensor(1048576)]; + tensor range_1d_25_start_0 = const()[name = tensor("range_1d_25_start_0"), val = tensor(0)]; + tensor range_1d_25_step_0 = const()[name = tensor("range_1d_25_step_0"), val = tensor(1)]; + tensor range_1d_25 = range_1d(end = reduce_prod_25, start = range_1d_25_start_0, step = range_1d_25_step_0)[name = tensor("range_1d_25")]; + tensor reshape_125 = reshape(shape = shape_73, x = range_1d_25)[name = tensor("reshape_125")]; + tensor slice_by_index_25 = slice_by_index(begin = concat_88, begin_mask = new_cache_25_internal_tensor_assign_2_begin_mask_0, end = concat_89, end_mask = new_cache_25_internal_tensor_assign_2_end_mask_0, squeeze_mask = new_cache_25_internal_tensor_assign_2_squeeze_mask_0, stride = new_cache_25_internal_tensor_assign_2_stride_0, x = reshape_125)[name = tensor("slice_by_index_25")]; + tensor reshape_126_shape_0 = const()[name = tensor("reshape_126_shape_0"), val = tensor([-1])]; + tensor reshape_126 = reshape(shape = reshape_126_shape_0, x = slice_by_index_25)[name = tensor("reshape_126")]; + tensor reshape_127_shape_0 = const()[name = tensor("reshape_127_shape_0"), val = tensor([-1])]; + tensor reshape_127_cast_fp16 = reshape(shape = reshape_127_shape_0, x = var_5077_cast_fp16)[name = tensor("reshape_127_cast_fp16")]; + tensor reshape_128_shape_0 = const()[name = tensor("reshape_128_shape_0"), val = tensor([-1])]; + tensor reshape_128_cast_fp16 = reshape(shape = reshape_128_shape_0, x = reshape_124_cast_fp16)[name = tensor("reshape_128_cast_fp16")]; + tensor scatter_25_mode_0 = const()[name = tensor("scatter_25_mode_0"), val = tensor("update")]; + tensor scatter_25_axis_0 = const()[name = tensor("scatter_25_axis_0"), val = tensor(0)]; + tensor scatter_25_validate_indices_0 = const()[name = tensor("scatter_25_validate_indices_0"), val = tensor(false)]; + tensor scatter_25_cast_fp16 = scatter(axis = scatter_25_axis_0, data = reshape_128_cast_fp16, indices = reshape_126, mode = scatter_25_mode_0, updates = reshape_127_cast_fp16, validate_indices = scatter_25_validate_indices_0)[name = tensor("scatter_25_cast_fp16")]; + tensor reshape_129_cast_fp16 = reshape(shape = shape_73, x = scatter_25_cast_fp16)[name = tensor("reshape_129_cast_fp16")]; + tensor reshape_129_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("reshape_129_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor keys_73_begin_0 = const()[name = tensor("keys_73_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor keys_73_end_0 = const()[name = tensor("keys_73_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor keys_73_end_mask_0 = const()[name = tensor("keys_73_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor keys_73_squeeze_mask_0 = const()[name = tensor("keys_73_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor keys_73_cast_fp16 = slice_by_index(begin = keys_73_begin_0, end = keys_73_end_0, end_mask = keys_73_end_mask_0, squeeze_mask = keys_73_squeeze_mask_0, x = reshape_129_cast_fp16)[name = tensor("keys_73_cast_fp16")]; + tensor values_73_begin_0 = const()[name = tensor("values_73_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor values_73_end_0 = const()[name = tensor("values_73_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor values_73_end_mask_0 = const()[name = tensor("values_73_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor values_73_squeeze_mask_0 = const()[name = tensor("values_73_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor values_73_cast_fp16 = slice_by_index(begin = values_73_begin_0, end = values_73_end_0, end_mask = values_73_end_mask_0, squeeze_mask = values_73_squeeze_mask_0, x = reshape_129_cast_fp16)[name = tensor("values_73_cast_fp16")]; + tensor var_5089_cast_fp16 = not_equal(x = keys_73_cast_fp16, y = keys_73_cast_fp16)[name = tensor("op_5089_cast_fp16")]; + tensor keys_75_cast_fp16 = select(a = var_463_to_fp16, b = keys_73_cast_fp16, cond = var_5089_cast_fp16)[name = tensor("keys_75_cast_fp16")]; + tensor var_5097_cast_fp16 = not_equal(x = values_73_cast_fp16, y = values_73_cast_fp16)[name = tensor("op_5097_cast_fp16")]; + tensor values_75_cast_fp16 = select(a = var_463_to_fp16, b = values_73_cast_fp16, cond = var_5097_cast_fp16)[name = tensor("values_75_cast_fp16")]; + tensor var_5121 = const()[name = tensor("op_5121"), val = tensor([0, 2, 1, 3])]; + tensor var_5137 = const()[name = tensor("op_5137"), val = tensor([1, 1, 1])]; + tensor var_5138_cast_fp16 = reshape(shape = var_5137, x = position12_to_fp16)[name = tensor("op_5138_cast_fp16")]; + tensor var_5158_to_fp16 = const()[name = tensor("op_5158_to_fp16"), val = tensor(0x1p+0)]; + tensor valid_len_25_cast_fp16 = add(x = var_5138_cast_fp16, y = var_5158_to_fp16)[name = tensor("valid_len_25_cast_fp16")]; + tensor valid_mask_25_cast_fp16 = less(x = k_positions_1_promoted_to_fp16, y = valid_len_25_cast_fp16)[name = tensor("valid_mask_25_cast_fp16")]; + tensor causal_mask_25_cast_fp16 = less_equal(x = k_positions_1_promoted_to_fp16, y = var_5138_cast_fp16)[name = tensor("causal_mask_25_cast_fp16")]; + tensor attn_mask_49 = logical_and(x = valid_mask_25_cast_fp16, y = causal_mask_25_cast_fp16)[name = tensor("attn_mask_49")]; + tensor attn_mask_51_axes_0 = const()[name = tensor("attn_mask_51_axes_0"), val = tensor([1])]; + tensor attn_mask_51 = expand_dims(axes = attn_mask_51_axes_0, x = attn_mask_49)[name = tensor("attn_mask_51")]; + tensor var_5176_transpose_x_0 = const()[name = tensor("op_5176_transpose_x_0"), val = tensor(false)]; + tensor var_5176_transpose_y_0 = const()[name = tensor("op_5176_transpose_y_0"), val = tensor(false)]; + tensor transpose_96_perm_0 = const()[name = tensor("transpose_96_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_97_perm_0 = const()[name = tensor("transpose_97_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_97 = transpose(perm = transpose_97_perm_0, x = keys_75_cast_fp16)[name = tensor("transpose_165")]; + tensor transpose_96 = transpose(perm = transpose_96_perm_0, x = q_75_cast_fp16)[name = tensor("transpose_166")]; + tensor var_5176_cast_fp16 = matmul(transpose_x = var_5176_transpose_x_0, transpose_y = var_5176_transpose_y_0, x = transpose_96, y = transpose_97)[name = tensor("op_5176_cast_fp16")]; + tensor var_5170_to_fp16 = const()[name = tensor("op_5170_to_fp16"), val = tensor([0x1p-3])]; + tensor attn_weights_73_cast_fp16 = mul(x = var_5176_cast_fp16, y = var_5170_to_fp16)[name = tensor("attn_weights_73_cast_fp16")]; + tensor var_5178 = logical_not(x = attn_mask_51)[name = tensor("op_5178")]; + tensor var_5179_to_fp16 = const()[name = tensor("op_5179_to_fp16"), val = tensor(-inf)]; + tensor attn_weights_75_cast_fp16 = select(a = var_5179_to_fp16, b = attn_weights_73_cast_fp16, cond = var_5178)[name = tensor("attn_weights_75_cast_fp16")]; + tensor var_5181 = const()[name = tensor("op_5181"), val = tensor(-1)]; + tensor attn_weights_77_cast_fp16 = softmax(axis = var_5181, x = attn_weights_75_cast_fp16)[name = tensor("attn_weights_77_cast_fp16")]; + tensor attn_output_25_transpose_x_0 = const()[name = tensor("attn_output_25_transpose_x_0"), val = tensor(false)]; + tensor attn_output_25_transpose_y_0 = const()[name = tensor("attn_output_25_transpose_y_0"), val = tensor(false)]; + tensor values_77_cast_fp16 = transpose(perm = var_5121, x = values_75_cast_fp16)[name = tensor("transpose_167")]; + tensor attn_output_25_cast_fp16 = matmul(transpose_x = attn_output_25_transpose_x_0, transpose_y = attn_output_25_transpose_y_0, x = attn_weights_77_cast_fp16, y = values_77_cast_fp16)[name = tensor("attn_output_25_cast_fp16")]; + tensor var_5189 = const()[name = tensor("op_5189"), val = tensor([0, 2, 1, 3])]; + tensor var_5194 = const()[name = tensor("op_5194"), val = tensor([1, 1, 1024])]; + tensor var_5190_cast_fp16 = transpose(perm = var_5189, x = attn_output_25_cast_fp16)[name = tensor("transpose_164")]; + tensor input_151_cast_fp16 = reshape(shape = var_5194, x = var_5190_cast_fp16)[name = tensor("input_151_cast_fp16")]; + tensor attn12_out_proj_weight_to_fp16 = const()[name = tensor("attn12_out_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(309523712)))]; + tensor linear_50_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = attn12_out_proj_weight_to_fp16, x = input_151_cast_fp16)[name = tensor("linear_50_cast_fp16")]; + tensor var_5200_to_fp16 = const()[name = tensor("op_5200_to_fp16"), val = tensor(0x1p+0)]; + tensor var_5201_cast_fp16 = add(x = position12_to_fp16, y = var_5200_to_fp16)[name = tensor("op_5201_cast_fp16")]; + tensor var_5201_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_5201_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor input_153_cast_fp16 = add(x = input_147_cast_fp16, y = linear_50_cast_fp16)[name = tensor("input_153_cast_fp16")]; + tensor input_155_axes_0 = const()[name = tensor("input_155_axes_0"), val = tensor([-1])]; + tensor norm12_2_weight_to_fp16 = const()[name = tensor("norm12_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(311620928)))]; + tensor norm12_2_bias_to_fp16 = const()[name = tensor("norm12_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(311623040)))]; + tensor var_5205_to_fp16 = const()[name = tensor("op_5205_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_155_cast_fp16 = layer_norm(axes = input_155_axes_0, beta = norm12_2_bias_to_fp16, epsilon = var_5205_to_fp16, gamma = norm12_2_weight_to_fp16, x = input_153_cast_fp16)[name = tensor("input_155_cast_fp16")]; + tensor linear12_1_weight_to_fp16 = const()[name = tensor("linear12_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(311625152)))]; + tensor linear_51_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = linear12_1_weight_to_fp16, x = input_155_cast_fp16)[name = tensor("linear_51_cast_fp16")]; + tensor input_157_mode_0 = const()[name = tensor("input_157_mode_0"), val = tensor("EXACT")]; + tensor input_157_cast_fp16 = gelu(mode = input_157_mode_0, x = linear_51_cast_fp16)[name = tensor("input_157_cast_fp16")]; + tensor linear12_2_weight_to_fp16 = const()[name = tensor("linear12_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(320013824)))]; + tensor linear_52_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = linear12_2_weight_to_fp16, x = input_157_cast_fp16)[name = tensor("linear_52_cast_fp16")]; + tensor input_159_cast_fp16 = add(x = input_153_cast_fp16, y = linear_52_cast_fp16)[name = tensor("input_159_cast_fp16")]; + tensor input_161_axes_0 = const()[name = tensor("input_161_axes_0"), val = tensor([-1])]; + tensor norm13_1_weight_to_fp16 = const()[name = tensor("norm13_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(328402496)))]; + tensor norm13_1_bias_to_fp16 = const()[name = tensor("norm13_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(328404608)))]; + tensor var_5222_to_fp16 = const()[name = tensor("op_5222_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_161_cast_fp16 = layer_norm(axes = input_161_axes_0, beta = norm13_1_bias_to_fp16, epsilon = var_5222_to_fp16, gamma = norm13_1_weight_to_fp16, x = input_159_cast_fp16)[name = tensor("input_161_cast_fp16")]; + tensor attn13_in_proj_weight_to_fp16 = const()[name = tensor("attn13_in_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(328406720)))]; + tensor linear_53_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = attn13_in_proj_weight_to_fp16, x = input_161_cast_fp16)[name = tensor("linear_53_cast_fp16")]; + tensor var_5236 = const()[name = tensor("op_5236"), val = tensor([1, 1, 3, 16, 64])]; + tensor qkv_27_cast_fp16 = reshape(shape = var_5236, x = linear_53_cast_fp16)[name = tensor("qkv_27_cast_fp16")]; + tensor q_79_begin_0 = const()[name = tensor("q_79_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor q_79_end_0 = const()[name = tensor("q_79_end_0"), val = tensor([1, 1, 1, 16, 64])]; + tensor q_79_end_mask_0 = const()[name = tensor("q_79_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor q_79_squeeze_mask_0 = const()[name = tensor("q_79_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor q_79_cast_fp16 = slice_by_index(begin = q_79_begin_0, end = q_79_end_0, end_mask = q_79_end_mask_0, squeeze_mask = q_79_squeeze_mask_0, x = qkv_27_cast_fp16)[name = tensor("q_79_cast_fp16")]; + tensor k_53_begin_0 = const()[name = tensor("k_53_begin_0"), val = tensor([0, 0, 1, 0, 0])]; + tensor k_53_end_0 = const()[name = tensor("k_53_end_0"), val = tensor([1, 1, 2, 16, 64])]; + tensor k_53_end_mask_0 = const()[name = tensor("k_53_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor k_53_squeeze_mask_0 = const()[name = tensor("k_53_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor k_53_cast_fp16 = slice_by_index(begin = k_53_begin_0, end = k_53_end_0, end_mask = k_53_end_mask_0, squeeze_mask = k_53_squeeze_mask_0, x = qkv_27_cast_fp16)[name = tensor("k_53_cast_fp16")]; + tensor v_27_begin_0 = const()[name = tensor("v_27_begin_0"), val = tensor([0, 0, 2, 0, 0])]; + tensor v_27_end_0 = const()[name = tensor("v_27_end_0"), val = tensor([1, 1, 3, 16, 64])]; + tensor v_27_end_mask_0 = const()[name = tensor("v_27_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor v_27_squeeze_mask_0 = const()[name = tensor("v_27_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor v_27_cast_fp16 = slice_by_index(begin = v_27_begin_0, end = v_27_end_0, end_mask = v_27_end_mask_0, squeeze_mask = v_27_squeeze_mask_0, x = qkv_27_cast_fp16)[name = tensor("v_27_cast_fp16")]; + tensor var_5302 = const()[name = tensor("op_5302"), val = tensor([1, 1, 1, 1])]; + tensor position13_to_fp16_dtype_0 = const()[name = tensor("position13_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor position13_to_fp16 = cast(dtype = position13_to_fp16_dtype_0, x = position13)[name = tensor("cast_488")]; + tensor ts_83_cast_fp16 = reshape(shape = var_5302, x = position13_to_fp16)[name = tensor("ts_83_cast_fp16")]; + tensor var_5309 = const()[name = tensor("op_5309"), val = tensor([1, 1, 16, 32, 2])]; + tensor q_complex_27_cast_fp16 = reshape(shape = var_5309, x = q_79_cast_fp16)[name = tensor("q_complex_27_cast_fp16")]; + tensor var_5316 = const()[name = tensor("op_5316"), val = tensor([1, 1, 16, 32, 2])]; + tensor k_complex_27_cast_fp16 = reshape(shape = var_5316, x = k_53_cast_fp16)[name = tensor("k_complex_27_cast_fp16")]; + tensor var_5320_begin_0 = const()[name = tensor("op_5320_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_5320_end_0 = const()[name = tensor("op_5320_end_0"), val = tensor([1, 1, 16, 32, 1])]; + tensor var_5320_end_mask_0 = const()[name = tensor("op_5320_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_5320_squeeze_mask_0 = const()[name = tensor("op_5320_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_5320_cast_fp16 = slice_by_index(begin = var_5320_begin_0, end = var_5320_end_0, end_mask = var_5320_end_mask_0, squeeze_mask = var_5320_squeeze_mask_0, x = q_complex_27_cast_fp16)[name = tensor("op_5320_cast_fp16")]; + tensor var_5328_begin_0 = const()[name = tensor("op_5328_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_5328_end_0 = const()[name = tensor("op_5328_end_0"), val = tensor([1, 1, 16, 32, 2])]; + tensor var_5328_end_mask_0 = const()[name = tensor("op_5328_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_5328_squeeze_mask_0 = const()[name = tensor("op_5328_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_5328_cast_fp16 = slice_by_index(begin = var_5328_begin_0, end = var_5328_end_0, end_mask = var_5328_end_mask_0, squeeze_mask = var_5328_squeeze_mask_0, x = q_complex_27_cast_fp16)[name = tensor("op_5328_cast_fp16")]; + tensor var_5336_begin_0 = const()[name = tensor("op_5336_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_5336_end_0 = const()[name = tensor("op_5336_end_0"), val = tensor([1, 1, 16, 32, 1])]; + tensor var_5336_end_mask_0 = const()[name = tensor("op_5336_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_5336_squeeze_mask_0 = const()[name = tensor("op_5336_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_5336_cast_fp16 = slice_by_index(begin = var_5336_begin_0, end = var_5336_end_0, end_mask = var_5336_end_mask_0, squeeze_mask = var_5336_squeeze_mask_0, x = k_complex_27_cast_fp16)[name = tensor("op_5336_cast_fp16")]; + tensor var_5344_begin_0 = const()[name = tensor("op_5344_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_5344_end_0 = const()[name = tensor("op_5344_end_0"), val = tensor([1, 1, 16, 32, 2])]; + tensor var_5344_end_mask_0 = const()[name = tensor("op_5344_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_5344_squeeze_mask_0 = const()[name = tensor("op_5344_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_5344_cast_fp16 = slice_by_index(begin = var_5344_begin_0, end = var_5344_end_0, end_mask = var_5344_end_mask_0, squeeze_mask = var_5344_squeeze_mask_0, x = k_complex_27_cast_fp16)[name = tensor("op_5344_cast_fp16")]; + tensor freqs_27_to_fp16 = const()[name = tensor("freqs_27_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(334698240)))]; + tensor var_5350_cast_fp16 = mul(x = freqs_27_to_fp16, y = ts_83_cast_fp16)[name = tensor("op_5350_cast_fp16")]; + tensor rotr_27_cast_fp16 = cos(x = var_5350_cast_fp16)[name = tensor("rotr_27_cast_fp16")]; + tensor roti_27_cast_fp16 = sin(x = var_5350_cast_fp16)[name = tensor("roti_27_cast_fp16")]; + tensor var_5354_cast_fp16 = mul(x = var_5320_cast_fp16, y = rotr_27_cast_fp16)[name = tensor("op_5354_cast_fp16")]; + tensor var_5355_cast_fp16 = mul(x = var_5328_cast_fp16, y = roti_27_cast_fp16)[name = tensor("op_5355_cast_fp16")]; + tensor qor_53_cast_fp16 = sub(x = var_5354_cast_fp16, y = var_5355_cast_fp16)[name = tensor("qor_53_cast_fp16")]; + tensor var_5358_cast_fp16 = mul(x = var_5320_cast_fp16, y = roti_27_cast_fp16)[name = tensor("op_5358_cast_fp16")]; + tensor var_5359_cast_fp16 = mul(x = var_5328_cast_fp16, y = rotr_27_cast_fp16)[name = tensor("op_5359_cast_fp16")]; + tensor qoi_53_cast_fp16 = add(x = var_5358_cast_fp16, y = var_5359_cast_fp16)[name = tensor("qoi_53_cast_fp16")]; + tensor var_5362_cast_fp16 = mul(x = var_5336_cast_fp16, y = rotr_27_cast_fp16)[name = tensor("op_5362_cast_fp16")]; + tensor var_5363_cast_fp16 = mul(x = var_5344_cast_fp16, y = roti_27_cast_fp16)[name = tensor("op_5363_cast_fp16")]; + tensor kor_53_cast_fp16 = sub(x = var_5362_cast_fp16, y = var_5363_cast_fp16)[name = tensor("kor_53_cast_fp16")]; + tensor var_5366_cast_fp16 = mul(x = var_5336_cast_fp16, y = roti_27_cast_fp16)[name = tensor("op_5366_cast_fp16")]; + tensor var_5367_cast_fp16 = mul(x = var_5344_cast_fp16, y = rotr_27_cast_fp16)[name = tensor("op_5367_cast_fp16")]; + tensor koi_53_cast_fp16 = add(x = var_5366_cast_fp16, y = var_5367_cast_fp16)[name = tensor("koi_53_cast_fp16")]; + tensor qo_27_axis_0 = const()[name = tensor("qo_27_axis_0"), val = tensor(-1)]; + tensor qo_27_cast_fp16 = stack(axis = qo_27_axis_0, values = (qor_53_cast_fp16, qoi_53_cast_fp16))[name = tensor("qo_27_cast_fp16")]; + tensor ko_27_axis_0 = const()[name = tensor("ko_27_axis_0"), val = tensor(-1)]; + tensor ko_27_cast_fp16 = stack(axis = ko_27_axis_0, values = (kor_53_cast_fp16, koi_53_cast_fp16))[name = tensor("ko_27_cast_fp16")]; + tensor var_5400 = const()[name = tensor("op_5400"), val = tensor([1, 1, 16, 64])]; + tensor q_81_cast_fp16 = reshape(shape = var_5400, x = qo_27_cast_fp16)[name = tensor("q_81_cast_fp16")]; + tensor var_5406 = const()[name = tensor("op_5406"), val = tensor([1, 1, 16, 64])]; + tensor k_55_cast_fp16 = reshape(shape = var_5406, x = ko_27_cast_fp16)[name = tensor("k_55_cast_fp16")]; + tensor _inversed_5431_y_0_to_fp16 = const()[name = tensor("_inversed_5431_y_0_to_fp16"), val = tensor(0x1p-9)]; + tensor _inversed_5431_cast_fp16 = mul(x = ts_83_cast_fp16, y = _inversed_5431_y_0_to_fp16)[name = tensor("_inversed_5431_cast_fp16")]; + tensor var_5432_cast_fp16 = floor(x = _inversed_5431_cast_fp16)[name = tensor("op_5432_cast_fp16")]; + tensor var_5433_to_fp16 = const()[name = tensor("op_5433_to_fp16"), val = tensor(0x1p+9)]; + tensor var_5434_cast_fp16 = mul(x = var_5432_cast_fp16, y = var_5433_to_fp16)[name = tensor("op_5434_cast_fp16")]; + tensor write_indices_float_55_cast_fp16 = sub(x = ts_83_cast_fp16, y = var_5434_cast_fp16)[name = tensor("write_indices_float_55_cast_fp16")]; + tensor var_5441_dtype_0 = const()[name = tensor("op_5441_dtype_0"), val = tensor("int32")]; + tensor write_indices_27_reps_0 = const()[name = tensor("write_indices_27_reps_0"), val = tensor([1, 1, 16, 64])]; + tensor write_indices_float_55_cast_fp16_to_int32 = cast(dtype = var_5441_dtype_0, x = write_indices_float_55_cast_fp16)[name = tensor("cast_487")]; + tensor write_indices_27 = tile(reps = write_indices_27_reps_0, x = write_indices_float_55_cast_fp16_to_int32)[name = tensor("write_indices_27")]; + tensor var_5451_begin_0 = const()[name = tensor("op_5451_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_5451_end_0 = const()[name = tensor("op_5451_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor var_5451_end_mask_0 = const()[name = tensor("op_5451_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_5451_squeeze_mask_0 = const()[name = tensor("op_5451_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor cache13_to_fp16_dtype_0 = const()[name = tensor("cache13_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor cache13_to_fp16 = cast(dtype = cache13_to_fp16_dtype_0, x = cache13)[name = tensor("cast_486")]; + tensor var_5451_cast_fp16 = slice_by_index(begin = var_5451_begin_0, end = var_5451_end_0, end_mask = var_5451_end_mask_0, squeeze_mask = var_5451_squeeze_mask_0, x = cache13_to_fp16)[name = tensor("op_5451_cast_fp16")]; + tensor var_5453_axis_0 = const()[name = tensor("op_5453_axis_0"), val = tensor(1)]; + tensor var_5453_mode_0 = const()[name = tensor("op_5453_mode_0"), val = tensor("update")]; + tensor var_5453_validate_indices_0 = const()[name = tensor("op_5453_validate_indices_0"), val = tensor(false)]; + tensor var_5453_cast_fp16 = scatter_along_axis(axis = var_5453_axis_0, data = var_5451_cast_fp16, indices = write_indices_27, mode = var_5453_mode_0, updates = k_55_cast_fp16, validate_indices = var_5453_validate_indices_0)[name = tensor("op_5453_cast_fp16")]; + tensor concat_93 = const()[name = tensor("concat_93"), val = tensor([0, 0, 0, 0, 0])]; + tensor concat_94 = const()[name = tensor("concat_94"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_27_internal_tensor_assign_1_stride_0 = const()[name = tensor("new_cache_27_internal_tensor_assign_1_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_27_internal_tensor_assign_1_begin_mask_0 = const()[name = tensor("new_cache_27_internal_tensor_assign_1_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_27_internal_tensor_assign_1_end_mask_0 = const()[name = tensor("new_cache_27_internal_tensor_assign_1_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_27_internal_tensor_assign_1_squeeze_mask_0 = const()[name = tensor("new_cache_27_internal_tensor_assign_1_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_74 = const()[name = tensor("shape_74"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_26 = const()[name = tensor("reduce_prod_26"), val = tensor(1048576)]; + tensor range_1d_26_start_0 = const()[name = tensor("range_1d_26_start_0"), val = tensor(0)]; + tensor range_1d_26_step_0 = const()[name = tensor("range_1d_26_step_0"), val = tensor(1)]; + tensor range_1d_26 = range_1d(end = reduce_prod_26, start = range_1d_26_start_0, step = range_1d_26_step_0)[name = tensor("range_1d_26")]; + tensor reshape_130 = reshape(shape = shape_74, x = range_1d_26)[name = tensor("reshape_130")]; + tensor slice_by_index_26 = slice_by_index(begin = concat_93, begin_mask = new_cache_27_internal_tensor_assign_1_begin_mask_0, end = concat_94, end_mask = new_cache_27_internal_tensor_assign_1_end_mask_0, squeeze_mask = new_cache_27_internal_tensor_assign_1_squeeze_mask_0, stride = new_cache_27_internal_tensor_assign_1_stride_0, x = reshape_130)[name = tensor("slice_by_index_26")]; + tensor reshape_131_shape_0 = const()[name = tensor("reshape_131_shape_0"), val = tensor([-1])]; + tensor reshape_131 = reshape(shape = reshape_131_shape_0, x = slice_by_index_26)[name = tensor("reshape_131")]; + tensor reshape_132_shape_0 = const()[name = tensor("reshape_132_shape_0"), val = tensor([-1])]; + tensor reshape_132_cast_fp16 = reshape(shape = reshape_132_shape_0, x = var_5453_cast_fp16)[name = tensor("reshape_132_cast_fp16")]; + tensor reshape_133_shape_0 = const()[name = tensor("reshape_133_shape_0"), val = tensor([-1])]; + tensor reshape_133_cast_fp16 = reshape(shape = reshape_133_shape_0, x = cache13_to_fp16)[name = tensor("reshape_133_cast_fp16")]; + tensor scatter_26_mode_0 = const()[name = tensor("scatter_26_mode_0"), val = tensor("update")]; + tensor scatter_26_axis_0 = const()[name = tensor("scatter_26_axis_0"), val = tensor(0)]; + tensor scatter_26_validate_indices_0 = const()[name = tensor("scatter_26_validate_indices_0"), val = tensor(false)]; + tensor scatter_26_cast_fp16 = scatter(axis = scatter_26_axis_0, data = reshape_133_cast_fp16, indices = reshape_131, mode = scatter_26_mode_0, updates = reshape_132_cast_fp16, validate_indices = scatter_26_validate_indices_0)[name = tensor("scatter_26_cast_fp16")]; + tensor reshape_134_cast_fp16 = reshape(shape = shape_74, x = scatter_26_cast_fp16)[name = tensor("reshape_134_cast_fp16")]; + tensor var_5461_begin_0 = const()[name = tensor("op_5461_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor var_5461_end_0 = const()[name = tensor("op_5461_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor var_5461_end_mask_0 = const()[name = tensor("op_5461_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_5461_squeeze_mask_0 = const()[name = tensor("op_5461_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor var_5461_cast_fp16 = slice_by_index(begin = var_5461_begin_0, end = var_5461_end_0, end_mask = var_5461_end_mask_0, squeeze_mask = var_5461_squeeze_mask_0, x = reshape_134_cast_fp16)[name = tensor("op_5461_cast_fp16")]; + tensor var_5463_axis_0 = const()[name = tensor("op_5463_axis_0"), val = tensor(1)]; + tensor var_5463_mode_0 = const()[name = tensor("op_5463_mode_0"), val = tensor("update")]; + tensor var_5463_validate_indices_0 = const()[name = tensor("op_5463_validate_indices_0"), val = tensor(false)]; + tensor var_5463_cast_fp16 = scatter_along_axis(axis = var_5463_axis_0, data = var_5461_cast_fp16, indices = write_indices_27, mode = var_5463_mode_0, updates = v_27_cast_fp16, validate_indices = var_5463_validate_indices_0)[name = tensor("op_5463_cast_fp16")]; + tensor concat_95 = const()[name = tensor("concat_95"), val = tensor([1, 0, 0, 0, 0])]; + tensor concat_96 = const()[name = tensor("concat_96"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_27_internal_tensor_assign_2_stride_0 = const()[name = tensor("new_cache_27_internal_tensor_assign_2_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_27_internal_tensor_assign_2_begin_mask_0 = const()[name = tensor("new_cache_27_internal_tensor_assign_2_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_27_internal_tensor_assign_2_end_mask_0 = const()[name = tensor("new_cache_27_internal_tensor_assign_2_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_27_internal_tensor_assign_2_squeeze_mask_0 = const()[name = tensor("new_cache_27_internal_tensor_assign_2_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_75 = const()[name = tensor("shape_75"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_27 = const()[name = tensor("reduce_prod_27"), val = tensor(1048576)]; + tensor range_1d_27_start_0 = const()[name = tensor("range_1d_27_start_0"), val = tensor(0)]; + tensor range_1d_27_step_0 = const()[name = tensor("range_1d_27_step_0"), val = tensor(1)]; + tensor range_1d_27 = range_1d(end = reduce_prod_27, start = range_1d_27_start_0, step = range_1d_27_step_0)[name = tensor("range_1d_27")]; + tensor reshape_135 = reshape(shape = shape_75, x = range_1d_27)[name = tensor("reshape_135")]; + tensor slice_by_index_27 = slice_by_index(begin = concat_95, begin_mask = new_cache_27_internal_tensor_assign_2_begin_mask_0, end = concat_96, end_mask = new_cache_27_internal_tensor_assign_2_end_mask_0, squeeze_mask = new_cache_27_internal_tensor_assign_2_squeeze_mask_0, stride = new_cache_27_internal_tensor_assign_2_stride_0, x = reshape_135)[name = tensor("slice_by_index_27")]; + tensor reshape_136_shape_0 = const()[name = tensor("reshape_136_shape_0"), val = tensor([-1])]; + tensor reshape_136 = reshape(shape = reshape_136_shape_0, x = slice_by_index_27)[name = tensor("reshape_136")]; + tensor reshape_137_shape_0 = const()[name = tensor("reshape_137_shape_0"), val = tensor([-1])]; + tensor reshape_137_cast_fp16 = reshape(shape = reshape_137_shape_0, x = var_5463_cast_fp16)[name = tensor("reshape_137_cast_fp16")]; + tensor reshape_138_shape_0 = const()[name = tensor("reshape_138_shape_0"), val = tensor([-1])]; + tensor reshape_138_cast_fp16 = reshape(shape = reshape_138_shape_0, x = reshape_134_cast_fp16)[name = tensor("reshape_138_cast_fp16")]; + tensor scatter_27_mode_0 = const()[name = tensor("scatter_27_mode_0"), val = tensor("update")]; + tensor scatter_27_axis_0 = const()[name = tensor("scatter_27_axis_0"), val = tensor(0)]; + tensor scatter_27_validate_indices_0 = const()[name = tensor("scatter_27_validate_indices_0"), val = tensor(false)]; + tensor scatter_27_cast_fp16 = scatter(axis = scatter_27_axis_0, data = reshape_138_cast_fp16, indices = reshape_136, mode = scatter_27_mode_0, updates = reshape_137_cast_fp16, validate_indices = scatter_27_validate_indices_0)[name = tensor("scatter_27_cast_fp16")]; + tensor reshape_139_cast_fp16 = reshape(shape = shape_75, x = scatter_27_cast_fp16)[name = tensor("reshape_139_cast_fp16")]; + tensor reshape_139_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("reshape_139_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor keys_79_begin_0 = const()[name = tensor("keys_79_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor keys_79_end_0 = const()[name = tensor("keys_79_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor keys_79_end_mask_0 = const()[name = tensor("keys_79_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor keys_79_squeeze_mask_0 = const()[name = tensor("keys_79_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor keys_79_cast_fp16 = slice_by_index(begin = keys_79_begin_0, end = keys_79_end_0, end_mask = keys_79_end_mask_0, squeeze_mask = keys_79_squeeze_mask_0, x = reshape_139_cast_fp16)[name = tensor("keys_79_cast_fp16")]; + tensor values_79_begin_0 = const()[name = tensor("values_79_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor values_79_end_0 = const()[name = tensor("values_79_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor values_79_end_mask_0 = const()[name = tensor("values_79_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor values_79_squeeze_mask_0 = const()[name = tensor("values_79_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor values_79_cast_fp16 = slice_by_index(begin = values_79_begin_0, end = values_79_end_0, end_mask = values_79_end_mask_0, squeeze_mask = values_79_squeeze_mask_0, x = reshape_139_cast_fp16)[name = tensor("values_79_cast_fp16")]; + tensor var_5475_cast_fp16 = not_equal(x = keys_79_cast_fp16, y = keys_79_cast_fp16)[name = tensor("op_5475_cast_fp16")]; + tensor keys_81_cast_fp16 = select(a = var_463_to_fp16, b = keys_79_cast_fp16, cond = var_5475_cast_fp16)[name = tensor("keys_81_cast_fp16")]; + tensor var_5483_cast_fp16 = not_equal(x = values_79_cast_fp16, y = values_79_cast_fp16)[name = tensor("op_5483_cast_fp16")]; + tensor values_81_cast_fp16 = select(a = var_463_to_fp16, b = values_79_cast_fp16, cond = var_5483_cast_fp16)[name = tensor("values_81_cast_fp16")]; + tensor var_5507 = const()[name = tensor("op_5507"), val = tensor([0, 2, 1, 3])]; + tensor var_5523 = const()[name = tensor("op_5523"), val = tensor([1, 1, 1])]; + tensor var_5524_cast_fp16 = reshape(shape = var_5523, x = position13_to_fp16)[name = tensor("op_5524_cast_fp16")]; + tensor var_5544_to_fp16 = const()[name = tensor("op_5544_to_fp16"), val = tensor(0x1p+0)]; + tensor valid_len_27_cast_fp16 = add(x = var_5524_cast_fp16, y = var_5544_to_fp16)[name = tensor("valid_len_27_cast_fp16")]; + tensor valid_mask_27_cast_fp16 = less(x = k_positions_1_promoted_to_fp16, y = valid_len_27_cast_fp16)[name = tensor("valid_mask_27_cast_fp16")]; + tensor causal_mask_27_cast_fp16 = less_equal(x = k_positions_1_promoted_to_fp16, y = var_5524_cast_fp16)[name = tensor("causal_mask_27_cast_fp16")]; + tensor attn_mask_53 = logical_and(x = valid_mask_27_cast_fp16, y = causal_mask_27_cast_fp16)[name = tensor("attn_mask_53")]; + tensor attn_mask_55_axes_0 = const()[name = tensor("attn_mask_55_axes_0"), val = tensor([1])]; + tensor attn_mask_55 = expand_dims(axes = attn_mask_55_axes_0, x = attn_mask_53)[name = tensor("attn_mask_55")]; + tensor var_5562_transpose_x_0 = const()[name = tensor("op_5562_transpose_x_0"), val = tensor(false)]; + tensor var_5562_transpose_y_0 = const()[name = tensor("op_5562_transpose_y_0"), val = tensor(false)]; + tensor transpose_98_perm_0 = const()[name = tensor("transpose_98_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_99_perm_0 = const()[name = tensor("transpose_99_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_99 = transpose(perm = transpose_99_perm_0, x = keys_81_cast_fp16)[name = tensor("transpose_161")]; + tensor transpose_98 = transpose(perm = transpose_98_perm_0, x = q_81_cast_fp16)[name = tensor("transpose_162")]; + tensor var_5562_cast_fp16 = matmul(transpose_x = var_5562_transpose_x_0, transpose_y = var_5562_transpose_y_0, x = transpose_98, y = transpose_99)[name = tensor("op_5562_cast_fp16")]; + tensor var_5556_to_fp16 = const()[name = tensor("op_5556_to_fp16"), val = tensor([0x1p-3])]; + tensor attn_weights_79_cast_fp16 = mul(x = var_5562_cast_fp16, y = var_5556_to_fp16)[name = tensor("attn_weights_79_cast_fp16")]; + tensor var_5564 = logical_not(x = attn_mask_55)[name = tensor("op_5564")]; + tensor var_5565_to_fp16 = const()[name = tensor("op_5565_to_fp16"), val = tensor(-inf)]; + tensor attn_weights_81_cast_fp16 = select(a = var_5565_to_fp16, b = attn_weights_79_cast_fp16, cond = var_5564)[name = tensor("attn_weights_81_cast_fp16")]; + tensor var_5567 = const()[name = tensor("op_5567"), val = tensor(-1)]; + tensor attn_weights_83_cast_fp16 = softmax(axis = var_5567, x = attn_weights_81_cast_fp16)[name = tensor("attn_weights_83_cast_fp16")]; + tensor attn_output_27_transpose_x_0 = const()[name = tensor("attn_output_27_transpose_x_0"), val = tensor(false)]; + tensor attn_output_27_transpose_y_0 = const()[name = tensor("attn_output_27_transpose_y_0"), val = tensor(false)]; + tensor values_83_cast_fp16 = transpose(perm = var_5507, x = values_81_cast_fp16)[name = tensor("transpose_163")]; + tensor attn_output_27_cast_fp16 = matmul(transpose_x = attn_output_27_transpose_x_0, transpose_y = attn_output_27_transpose_y_0, x = attn_weights_83_cast_fp16, y = values_83_cast_fp16)[name = tensor("attn_output_27_cast_fp16")]; + tensor var_5575 = const()[name = tensor("op_5575"), val = tensor([0, 2, 1, 3])]; + tensor var_5580 = const()[name = tensor("op_5580"), val = tensor([1, 1, 1024])]; + tensor var_5576_cast_fp16 = transpose(perm = var_5575, x = attn_output_27_cast_fp16)[name = tensor("transpose_160")]; + tensor input_163_cast_fp16 = reshape(shape = var_5580, x = var_5576_cast_fp16)[name = tensor("input_163_cast_fp16")]; + tensor attn13_out_proj_weight_to_fp16 = const()[name = tensor("attn13_out_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(334698368)))]; + tensor linear_54_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = attn13_out_proj_weight_to_fp16, x = input_163_cast_fp16)[name = tensor("linear_54_cast_fp16")]; + tensor var_5586_to_fp16 = const()[name = tensor("op_5586_to_fp16"), val = tensor(0x1p+0)]; + tensor var_5587_cast_fp16 = add(x = position13_to_fp16, y = var_5586_to_fp16)[name = tensor("op_5587_cast_fp16")]; + tensor var_5587_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_5587_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor input_165_cast_fp16 = add(x = input_159_cast_fp16, y = linear_54_cast_fp16)[name = tensor("input_165_cast_fp16")]; + tensor input_167_axes_0 = const()[name = tensor("input_167_axes_0"), val = tensor([-1])]; + tensor norm13_2_weight_to_fp16 = const()[name = tensor("norm13_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(336795584)))]; + tensor norm13_2_bias_to_fp16 = const()[name = tensor("norm13_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(336797696)))]; + tensor var_5591_to_fp16 = const()[name = tensor("op_5591_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_167_cast_fp16 = layer_norm(axes = input_167_axes_0, beta = norm13_2_bias_to_fp16, epsilon = var_5591_to_fp16, gamma = norm13_2_weight_to_fp16, x = input_165_cast_fp16)[name = tensor("input_167_cast_fp16")]; + tensor linear13_1_weight_to_fp16 = const()[name = tensor("linear13_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(336799808)))]; + tensor linear_55_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = linear13_1_weight_to_fp16, x = input_167_cast_fp16)[name = tensor("linear_55_cast_fp16")]; + tensor input_169_mode_0 = const()[name = tensor("input_169_mode_0"), val = tensor("EXACT")]; + tensor input_169_cast_fp16 = gelu(mode = input_169_mode_0, x = linear_55_cast_fp16)[name = tensor("input_169_cast_fp16")]; + tensor linear13_2_weight_to_fp16 = const()[name = tensor("linear13_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(345188480)))]; + tensor linear_56_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = linear13_2_weight_to_fp16, x = input_169_cast_fp16)[name = tensor("linear_56_cast_fp16")]; + tensor input_171_cast_fp16 = add(x = input_165_cast_fp16, y = linear_56_cast_fp16)[name = tensor("input_171_cast_fp16")]; + tensor input_173_axes_0 = const()[name = tensor("input_173_axes_0"), val = tensor([-1])]; + tensor norm14_1_weight_to_fp16 = const()[name = tensor("norm14_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(353577152)))]; + tensor norm14_1_bias_to_fp16 = const()[name = tensor("norm14_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(353579264)))]; + tensor var_5608_to_fp16 = const()[name = tensor("op_5608_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_173_cast_fp16 = layer_norm(axes = input_173_axes_0, beta = norm14_1_bias_to_fp16, epsilon = var_5608_to_fp16, gamma = norm14_1_weight_to_fp16, x = input_171_cast_fp16)[name = tensor("input_173_cast_fp16")]; + tensor attn14_in_proj_weight_to_fp16 = const()[name = tensor("attn14_in_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(353581376)))]; + tensor linear_57_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = attn14_in_proj_weight_to_fp16, x = input_173_cast_fp16)[name = tensor("linear_57_cast_fp16")]; + tensor var_5622 = const()[name = tensor("op_5622"), val = tensor([1, 1, 3, 16, 64])]; + tensor qkv_29_cast_fp16 = reshape(shape = var_5622, x = linear_57_cast_fp16)[name = tensor("qkv_29_cast_fp16")]; + tensor q_85_begin_0 = const()[name = tensor("q_85_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor q_85_end_0 = const()[name = tensor("q_85_end_0"), val = tensor([1, 1, 1, 16, 64])]; + tensor q_85_end_mask_0 = const()[name = tensor("q_85_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor q_85_squeeze_mask_0 = const()[name = tensor("q_85_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor q_85_cast_fp16 = slice_by_index(begin = q_85_begin_0, end = q_85_end_0, end_mask = q_85_end_mask_0, squeeze_mask = q_85_squeeze_mask_0, x = qkv_29_cast_fp16)[name = tensor("q_85_cast_fp16")]; + tensor k_57_begin_0 = const()[name = tensor("k_57_begin_0"), val = tensor([0, 0, 1, 0, 0])]; + tensor k_57_end_0 = const()[name = tensor("k_57_end_0"), val = tensor([1, 1, 2, 16, 64])]; + tensor k_57_end_mask_0 = const()[name = tensor("k_57_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor k_57_squeeze_mask_0 = const()[name = tensor("k_57_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor k_57_cast_fp16 = slice_by_index(begin = k_57_begin_0, end = k_57_end_0, end_mask = k_57_end_mask_0, squeeze_mask = k_57_squeeze_mask_0, x = qkv_29_cast_fp16)[name = tensor("k_57_cast_fp16")]; + tensor v_29_begin_0 = const()[name = tensor("v_29_begin_0"), val = tensor([0, 0, 2, 0, 0])]; + tensor v_29_end_0 = const()[name = tensor("v_29_end_0"), val = tensor([1, 1, 3, 16, 64])]; + tensor v_29_end_mask_0 = const()[name = tensor("v_29_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor v_29_squeeze_mask_0 = const()[name = tensor("v_29_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor v_29_cast_fp16 = slice_by_index(begin = v_29_begin_0, end = v_29_end_0, end_mask = v_29_end_mask_0, squeeze_mask = v_29_squeeze_mask_0, x = qkv_29_cast_fp16)[name = tensor("v_29_cast_fp16")]; + tensor var_5688 = const()[name = tensor("op_5688"), val = tensor([1, 1, 1, 1])]; + tensor position14_to_fp16_dtype_0 = const()[name = tensor("position14_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor position14_to_fp16 = cast(dtype = position14_to_fp16_dtype_0, x = position14)[name = tensor("cast_483")]; + tensor ts_89_cast_fp16 = reshape(shape = var_5688, x = position14_to_fp16)[name = tensor("ts_89_cast_fp16")]; + tensor var_5695 = const()[name = tensor("op_5695"), val = tensor([1, 1, 16, 32, 2])]; + tensor q_complex_29_cast_fp16 = reshape(shape = var_5695, x = q_85_cast_fp16)[name = tensor("q_complex_29_cast_fp16")]; + tensor var_5702 = const()[name = tensor("op_5702"), val = tensor([1, 1, 16, 32, 2])]; + tensor k_complex_29_cast_fp16 = reshape(shape = var_5702, x = k_57_cast_fp16)[name = tensor("k_complex_29_cast_fp16")]; + tensor var_5706_begin_0 = const()[name = tensor("op_5706_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_5706_end_0 = const()[name = tensor("op_5706_end_0"), val = tensor([1, 1, 16, 32, 1])]; + tensor var_5706_end_mask_0 = const()[name = tensor("op_5706_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_5706_squeeze_mask_0 = const()[name = tensor("op_5706_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_5706_cast_fp16 = slice_by_index(begin = var_5706_begin_0, end = var_5706_end_0, end_mask = var_5706_end_mask_0, squeeze_mask = var_5706_squeeze_mask_0, x = q_complex_29_cast_fp16)[name = tensor("op_5706_cast_fp16")]; + tensor var_5714_begin_0 = const()[name = tensor("op_5714_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_5714_end_0 = const()[name = tensor("op_5714_end_0"), val = tensor([1, 1, 16, 32, 2])]; + tensor var_5714_end_mask_0 = const()[name = tensor("op_5714_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_5714_squeeze_mask_0 = const()[name = tensor("op_5714_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_5714_cast_fp16 = slice_by_index(begin = var_5714_begin_0, end = var_5714_end_0, end_mask = var_5714_end_mask_0, squeeze_mask = var_5714_squeeze_mask_0, x = q_complex_29_cast_fp16)[name = tensor("op_5714_cast_fp16")]; + tensor var_5722_begin_0 = const()[name = tensor("op_5722_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_5722_end_0 = const()[name = tensor("op_5722_end_0"), val = tensor([1, 1, 16, 32, 1])]; + tensor var_5722_end_mask_0 = const()[name = tensor("op_5722_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_5722_squeeze_mask_0 = const()[name = tensor("op_5722_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_5722_cast_fp16 = slice_by_index(begin = var_5722_begin_0, end = var_5722_end_0, end_mask = var_5722_end_mask_0, squeeze_mask = var_5722_squeeze_mask_0, x = k_complex_29_cast_fp16)[name = tensor("op_5722_cast_fp16")]; + tensor var_5730_begin_0 = const()[name = tensor("op_5730_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_5730_end_0 = const()[name = tensor("op_5730_end_0"), val = tensor([1, 1, 16, 32, 2])]; + tensor var_5730_end_mask_0 = const()[name = tensor("op_5730_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_5730_squeeze_mask_0 = const()[name = tensor("op_5730_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_5730_cast_fp16 = slice_by_index(begin = var_5730_begin_0, end = var_5730_end_0, end_mask = var_5730_end_mask_0, squeeze_mask = var_5730_squeeze_mask_0, x = k_complex_29_cast_fp16)[name = tensor("op_5730_cast_fp16")]; + tensor freqs_29_to_fp16 = const()[name = tensor("freqs_29_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(359872896)))]; + tensor var_5736_cast_fp16 = mul(x = freqs_29_to_fp16, y = ts_89_cast_fp16)[name = tensor("op_5736_cast_fp16")]; + tensor rotr_29_cast_fp16 = cos(x = var_5736_cast_fp16)[name = tensor("rotr_29_cast_fp16")]; + tensor roti_29_cast_fp16 = sin(x = var_5736_cast_fp16)[name = tensor("roti_29_cast_fp16")]; + tensor var_5740_cast_fp16 = mul(x = var_5706_cast_fp16, y = rotr_29_cast_fp16)[name = tensor("op_5740_cast_fp16")]; + tensor var_5741_cast_fp16 = mul(x = var_5714_cast_fp16, y = roti_29_cast_fp16)[name = tensor("op_5741_cast_fp16")]; + tensor qor_57_cast_fp16 = sub(x = var_5740_cast_fp16, y = var_5741_cast_fp16)[name = tensor("qor_57_cast_fp16")]; + tensor var_5744_cast_fp16 = mul(x = var_5706_cast_fp16, y = roti_29_cast_fp16)[name = tensor("op_5744_cast_fp16")]; + tensor var_5745_cast_fp16 = mul(x = var_5714_cast_fp16, y = rotr_29_cast_fp16)[name = tensor("op_5745_cast_fp16")]; + tensor qoi_57_cast_fp16 = add(x = var_5744_cast_fp16, y = var_5745_cast_fp16)[name = tensor("qoi_57_cast_fp16")]; + tensor var_5748_cast_fp16 = mul(x = var_5722_cast_fp16, y = rotr_29_cast_fp16)[name = tensor("op_5748_cast_fp16")]; + tensor var_5749_cast_fp16 = mul(x = var_5730_cast_fp16, y = roti_29_cast_fp16)[name = tensor("op_5749_cast_fp16")]; + tensor kor_57_cast_fp16 = sub(x = var_5748_cast_fp16, y = var_5749_cast_fp16)[name = tensor("kor_57_cast_fp16")]; + tensor var_5752_cast_fp16 = mul(x = var_5722_cast_fp16, y = roti_29_cast_fp16)[name = tensor("op_5752_cast_fp16")]; + tensor var_5753_cast_fp16 = mul(x = var_5730_cast_fp16, y = rotr_29_cast_fp16)[name = tensor("op_5753_cast_fp16")]; + tensor koi_57_cast_fp16 = add(x = var_5752_cast_fp16, y = var_5753_cast_fp16)[name = tensor("koi_57_cast_fp16")]; + tensor qo_29_axis_0 = const()[name = tensor("qo_29_axis_0"), val = tensor(-1)]; + tensor qo_29_cast_fp16 = stack(axis = qo_29_axis_0, values = (qor_57_cast_fp16, qoi_57_cast_fp16))[name = tensor("qo_29_cast_fp16")]; + tensor ko_29_axis_0 = const()[name = tensor("ko_29_axis_0"), val = tensor(-1)]; + tensor ko_29_cast_fp16 = stack(axis = ko_29_axis_0, values = (kor_57_cast_fp16, koi_57_cast_fp16))[name = tensor("ko_29_cast_fp16")]; + tensor var_5786 = const()[name = tensor("op_5786"), val = tensor([1, 1, 16, 64])]; + tensor q_87_cast_fp16 = reshape(shape = var_5786, x = qo_29_cast_fp16)[name = tensor("q_87_cast_fp16")]; + tensor var_5792 = const()[name = tensor("op_5792"), val = tensor([1, 1, 16, 64])]; + tensor k_59_cast_fp16 = reshape(shape = var_5792, x = ko_29_cast_fp16)[name = tensor("k_59_cast_fp16")]; + tensor _inversed_5817_y_0_to_fp16 = const()[name = tensor("_inversed_5817_y_0_to_fp16"), val = tensor(0x1p-9)]; + tensor _inversed_5817_cast_fp16 = mul(x = ts_89_cast_fp16, y = _inversed_5817_y_0_to_fp16)[name = tensor("_inversed_5817_cast_fp16")]; + tensor var_5818_cast_fp16 = floor(x = _inversed_5817_cast_fp16)[name = tensor("op_5818_cast_fp16")]; + tensor var_5819_to_fp16 = const()[name = tensor("op_5819_to_fp16"), val = tensor(0x1p+9)]; + tensor var_5820_cast_fp16 = mul(x = var_5818_cast_fp16, y = var_5819_to_fp16)[name = tensor("op_5820_cast_fp16")]; + tensor write_indices_float_59_cast_fp16 = sub(x = ts_89_cast_fp16, y = var_5820_cast_fp16)[name = tensor("write_indices_float_59_cast_fp16")]; + tensor var_5827_dtype_0 = const()[name = tensor("op_5827_dtype_0"), val = tensor("int32")]; + tensor write_indices_29_reps_0 = const()[name = tensor("write_indices_29_reps_0"), val = tensor([1, 1, 16, 64])]; + tensor write_indices_float_59_cast_fp16_to_int32 = cast(dtype = var_5827_dtype_0, x = write_indices_float_59_cast_fp16)[name = tensor("cast_482")]; + tensor write_indices_29 = tile(reps = write_indices_29_reps_0, x = write_indices_float_59_cast_fp16_to_int32)[name = tensor("write_indices_29")]; + tensor var_5837_begin_0 = const()[name = tensor("op_5837_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_5837_end_0 = const()[name = tensor("op_5837_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor var_5837_end_mask_0 = const()[name = tensor("op_5837_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_5837_squeeze_mask_0 = const()[name = tensor("op_5837_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor cache14_to_fp16_dtype_0 = const()[name = tensor("cache14_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor cache14_to_fp16 = cast(dtype = cache14_to_fp16_dtype_0, x = cache14)[name = tensor("cast_481")]; + tensor var_5837_cast_fp16 = slice_by_index(begin = var_5837_begin_0, end = var_5837_end_0, end_mask = var_5837_end_mask_0, squeeze_mask = var_5837_squeeze_mask_0, x = cache14_to_fp16)[name = tensor("op_5837_cast_fp16")]; + tensor var_5839_axis_0 = const()[name = tensor("op_5839_axis_0"), val = tensor(1)]; + tensor var_5839_mode_0 = const()[name = tensor("op_5839_mode_0"), val = tensor("update")]; + tensor var_5839_validate_indices_0 = const()[name = tensor("op_5839_validate_indices_0"), val = tensor(false)]; + tensor var_5839_cast_fp16 = scatter_along_axis(axis = var_5839_axis_0, data = var_5837_cast_fp16, indices = write_indices_29, mode = var_5839_mode_0, updates = k_59_cast_fp16, validate_indices = var_5839_validate_indices_0)[name = tensor("op_5839_cast_fp16")]; + tensor concat_100 = const()[name = tensor("concat_100"), val = tensor([0, 0, 0, 0, 0])]; + tensor concat_101 = const()[name = tensor("concat_101"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_29_internal_tensor_assign_1_stride_0 = const()[name = tensor("new_cache_29_internal_tensor_assign_1_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_29_internal_tensor_assign_1_begin_mask_0 = const()[name = tensor("new_cache_29_internal_tensor_assign_1_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_29_internal_tensor_assign_1_end_mask_0 = const()[name = tensor("new_cache_29_internal_tensor_assign_1_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_29_internal_tensor_assign_1_squeeze_mask_0 = const()[name = tensor("new_cache_29_internal_tensor_assign_1_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_76 = const()[name = tensor("shape_76"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_28 = const()[name = tensor("reduce_prod_28"), val = tensor(1048576)]; + tensor range_1d_28_start_0 = const()[name = tensor("range_1d_28_start_0"), val = tensor(0)]; + tensor range_1d_28_step_0 = const()[name = tensor("range_1d_28_step_0"), val = tensor(1)]; + tensor range_1d_28 = range_1d(end = reduce_prod_28, start = range_1d_28_start_0, step = range_1d_28_step_0)[name = tensor("range_1d_28")]; + tensor reshape_140 = reshape(shape = shape_76, x = range_1d_28)[name = tensor("reshape_140")]; + tensor slice_by_index_28 = slice_by_index(begin = concat_100, begin_mask = new_cache_29_internal_tensor_assign_1_begin_mask_0, end = concat_101, end_mask = new_cache_29_internal_tensor_assign_1_end_mask_0, squeeze_mask = new_cache_29_internal_tensor_assign_1_squeeze_mask_0, stride = new_cache_29_internal_tensor_assign_1_stride_0, x = reshape_140)[name = tensor("slice_by_index_28")]; + tensor reshape_141_shape_0 = const()[name = tensor("reshape_141_shape_0"), val = tensor([-1])]; + tensor reshape_141 = reshape(shape = reshape_141_shape_0, x = slice_by_index_28)[name = tensor("reshape_141")]; + tensor reshape_142_shape_0 = const()[name = tensor("reshape_142_shape_0"), val = tensor([-1])]; + tensor reshape_142_cast_fp16 = reshape(shape = reshape_142_shape_0, x = var_5839_cast_fp16)[name = tensor("reshape_142_cast_fp16")]; + tensor reshape_143_shape_0 = const()[name = tensor("reshape_143_shape_0"), val = tensor([-1])]; + tensor reshape_143_cast_fp16 = reshape(shape = reshape_143_shape_0, x = cache14_to_fp16)[name = tensor("reshape_143_cast_fp16")]; + tensor scatter_28_mode_0 = const()[name = tensor("scatter_28_mode_0"), val = tensor("update")]; + tensor scatter_28_axis_0 = const()[name = tensor("scatter_28_axis_0"), val = tensor(0)]; + tensor scatter_28_validate_indices_0 = const()[name = tensor("scatter_28_validate_indices_0"), val = tensor(false)]; + tensor scatter_28_cast_fp16 = scatter(axis = scatter_28_axis_0, data = reshape_143_cast_fp16, indices = reshape_141, mode = scatter_28_mode_0, updates = reshape_142_cast_fp16, validate_indices = scatter_28_validate_indices_0)[name = tensor("scatter_28_cast_fp16")]; + tensor reshape_144_cast_fp16 = reshape(shape = shape_76, x = scatter_28_cast_fp16)[name = tensor("reshape_144_cast_fp16")]; + tensor var_5847_begin_0 = const()[name = tensor("op_5847_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor var_5847_end_0 = const()[name = tensor("op_5847_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor var_5847_end_mask_0 = const()[name = tensor("op_5847_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_5847_squeeze_mask_0 = const()[name = tensor("op_5847_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor var_5847_cast_fp16 = slice_by_index(begin = var_5847_begin_0, end = var_5847_end_0, end_mask = var_5847_end_mask_0, squeeze_mask = var_5847_squeeze_mask_0, x = reshape_144_cast_fp16)[name = tensor("op_5847_cast_fp16")]; + tensor var_5849_axis_0 = const()[name = tensor("op_5849_axis_0"), val = tensor(1)]; + tensor var_5849_mode_0 = const()[name = tensor("op_5849_mode_0"), val = tensor("update")]; + tensor var_5849_validate_indices_0 = const()[name = tensor("op_5849_validate_indices_0"), val = tensor(false)]; + tensor var_5849_cast_fp16 = scatter_along_axis(axis = var_5849_axis_0, data = var_5847_cast_fp16, indices = write_indices_29, mode = var_5849_mode_0, updates = v_29_cast_fp16, validate_indices = var_5849_validate_indices_0)[name = tensor("op_5849_cast_fp16")]; + tensor concat_102 = const()[name = tensor("concat_102"), val = tensor([1, 0, 0, 0, 0])]; + tensor concat_103 = const()[name = tensor("concat_103"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_29_internal_tensor_assign_2_stride_0 = const()[name = tensor("new_cache_29_internal_tensor_assign_2_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_29_internal_tensor_assign_2_begin_mask_0 = const()[name = tensor("new_cache_29_internal_tensor_assign_2_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_29_internal_tensor_assign_2_end_mask_0 = const()[name = tensor("new_cache_29_internal_tensor_assign_2_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_29_internal_tensor_assign_2_squeeze_mask_0 = const()[name = tensor("new_cache_29_internal_tensor_assign_2_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_77 = const()[name = tensor("shape_77"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_29 = const()[name = tensor("reduce_prod_29"), val = tensor(1048576)]; + tensor range_1d_29_start_0 = const()[name = tensor("range_1d_29_start_0"), val = tensor(0)]; + tensor range_1d_29_step_0 = const()[name = tensor("range_1d_29_step_0"), val = tensor(1)]; + tensor range_1d_29 = range_1d(end = reduce_prod_29, start = range_1d_29_start_0, step = range_1d_29_step_0)[name = tensor("range_1d_29")]; + tensor reshape_145 = reshape(shape = shape_77, x = range_1d_29)[name = tensor("reshape_145")]; + tensor slice_by_index_29 = slice_by_index(begin = concat_102, begin_mask = new_cache_29_internal_tensor_assign_2_begin_mask_0, end = concat_103, end_mask = new_cache_29_internal_tensor_assign_2_end_mask_0, squeeze_mask = new_cache_29_internal_tensor_assign_2_squeeze_mask_0, stride = new_cache_29_internal_tensor_assign_2_stride_0, x = reshape_145)[name = tensor("slice_by_index_29")]; + tensor reshape_146_shape_0 = const()[name = tensor("reshape_146_shape_0"), val = tensor([-1])]; + tensor reshape_146 = reshape(shape = reshape_146_shape_0, x = slice_by_index_29)[name = tensor("reshape_146")]; + tensor reshape_147_shape_0 = const()[name = tensor("reshape_147_shape_0"), val = tensor([-1])]; + tensor reshape_147_cast_fp16 = reshape(shape = reshape_147_shape_0, x = var_5849_cast_fp16)[name = tensor("reshape_147_cast_fp16")]; + tensor reshape_148_shape_0 = const()[name = tensor("reshape_148_shape_0"), val = tensor([-1])]; + tensor reshape_148_cast_fp16 = reshape(shape = reshape_148_shape_0, x = reshape_144_cast_fp16)[name = tensor("reshape_148_cast_fp16")]; + tensor scatter_29_mode_0 = const()[name = tensor("scatter_29_mode_0"), val = tensor("update")]; + tensor scatter_29_axis_0 = const()[name = tensor("scatter_29_axis_0"), val = tensor(0)]; + tensor scatter_29_validate_indices_0 = const()[name = tensor("scatter_29_validate_indices_0"), val = tensor(false)]; + tensor scatter_29_cast_fp16 = scatter(axis = scatter_29_axis_0, data = reshape_148_cast_fp16, indices = reshape_146, mode = scatter_29_mode_0, updates = reshape_147_cast_fp16, validate_indices = scatter_29_validate_indices_0)[name = tensor("scatter_29_cast_fp16")]; + tensor reshape_149_cast_fp16 = reshape(shape = shape_77, x = scatter_29_cast_fp16)[name = tensor("reshape_149_cast_fp16")]; + tensor reshape_149_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("reshape_149_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor keys_85_begin_0 = const()[name = tensor("keys_85_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor keys_85_end_0 = const()[name = tensor("keys_85_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor keys_85_end_mask_0 = const()[name = tensor("keys_85_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor keys_85_squeeze_mask_0 = const()[name = tensor("keys_85_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor keys_85_cast_fp16 = slice_by_index(begin = keys_85_begin_0, end = keys_85_end_0, end_mask = keys_85_end_mask_0, squeeze_mask = keys_85_squeeze_mask_0, x = reshape_149_cast_fp16)[name = tensor("keys_85_cast_fp16")]; + tensor values_85_begin_0 = const()[name = tensor("values_85_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor values_85_end_0 = const()[name = tensor("values_85_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor values_85_end_mask_0 = const()[name = tensor("values_85_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor values_85_squeeze_mask_0 = const()[name = tensor("values_85_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor values_85_cast_fp16 = slice_by_index(begin = values_85_begin_0, end = values_85_end_0, end_mask = values_85_end_mask_0, squeeze_mask = values_85_squeeze_mask_0, x = reshape_149_cast_fp16)[name = tensor("values_85_cast_fp16")]; + tensor var_5861_cast_fp16 = not_equal(x = keys_85_cast_fp16, y = keys_85_cast_fp16)[name = tensor("op_5861_cast_fp16")]; + tensor keys_87_cast_fp16 = select(a = var_463_to_fp16, b = keys_85_cast_fp16, cond = var_5861_cast_fp16)[name = tensor("keys_87_cast_fp16")]; + tensor var_5869_cast_fp16 = not_equal(x = values_85_cast_fp16, y = values_85_cast_fp16)[name = tensor("op_5869_cast_fp16")]; + tensor values_87_cast_fp16 = select(a = var_463_to_fp16, b = values_85_cast_fp16, cond = var_5869_cast_fp16)[name = tensor("values_87_cast_fp16")]; + tensor var_5893 = const()[name = tensor("op_5893"), val = tensor([0, 2, 1, 3])]; + tensor var_5909 = const()[name = tensor("op_5909"), val = tensor([1, 1, 1])]; + tensor var_5910_cast_fp16 = reshape(shape = var_5909, x = position14_to_fp16)[name = tensor("op_5910_cast_fp16")]; + tensor var_5930_to_fp16 = const()[name = tensor("op_5930_to_fp16"), val = tensor(0x1p+0)]; + tensor valid_len_29_cast_fp16 = add(x = var_5910_cast_fp16, y = var_5930_to_fp16)[name = tensor("valid_len_29_cast_fp16")]; + tensor valid_mask_29_cast_fp16 = less(x = k_positions_1_promoted_to_fp16, y = valid_len_29_cast_fp16)[name = tensor("valid_mask_29_cast_fp16")]; + tensor causal_mask_29_cast_fp16 = less_equal(x = k_positions_1_promoted_to_fp16, y = var_5910_cast_fp16)[name = tensor("causal_mask_29_cast_fp16")]; + tensor attn_mask_57 = logical_and(x = valid_mask_29_cast_fp16, y = causal_mask_29_cast_fp16)[name = tensor("attn_mask_57")]; + tensor attn_mask_59_axes_0 = const()[name = tensor("attn_mask_59_axes_0"), val = tensor([1])]; + tensor attn_mask_59 = expand_dims(axes = attn_mask_59_axes_0, x = attn_mask_57)[name = tensor("attn_mask_59")]; + tensor var_5948_transpose_x_0 = const()[name = tensor("op_5948_transpose_x_0"), val = tensor(false)]; + tensor var_5948_transpose_y_0 = const()[name = tensor("op_5948_transpose_y_0"), val = tensor(false)]; + tensor transpose_100_perm_0 = const()[name = tensor("transpose_100_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_101_perm_0 = const()[name = tensor("transpose_101_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_101 = transpose(perm = transpose_101_perm_0, x = keys_87_cast_fp16)[name = tensor("transpose_157")]; + tensor transpose_100 = transpose(perm = transpose_100_perm_0, x = q_87_cast_fp16)[name = tensor("transpose_158")]; + tensor var_5948_cast_fp16 = matmul(transpose_x = var_5948_transpose_x_0, transpose_y = var_5948_transpose_y_0, x = transpose_100, y = transpose_101)[name = tensor("op_5948_cast_fp16")]; + tensor var_5942_to_fp16 = const()[name = tensor("op_5942_to_fp16"), val = tensor([0x1p-3])]; + tensor attn_weights_85_cast_fp16 = mul(x = var_5948_cast_fp16, y = var_5942_to_fp16)[name = tensor("attn_weights_85_cast_fp16")]; + tensor var_5950 = logical_not(x = attn_mask_59)[name = tensor("op_5950")]; + tensor var_5951_to_fp16 = const()[name = tensor("op_5951_to_fp16"), val = tensor(-inf)]; + tensor attn_weights_87_cast_fp16 = select(a = var_5951_to_fp16, b = attn_weights_85_cast_fp16, cond = var_5950)[name = tensor("attn_weights_87_cast_fp16")]; + tensor var_5953 = const()[name = tensor("op_5953"), val = tensor(-1)]; + tensor attn_weights_89_cast_fp16 = softmax(axis = var_5953, x = attn_weights_87_cast_fp16)[name = tensor("attn_weights_89_cast_fp16")]; + tensor attn_output_29_transpose_x_0 = const()[name = tensor("attn_output_29_transpose_x_0"), val = tensor(false)]; + tensor attn_output_29_transpose_y_0 = const()[name = tensor("attn_output_29_transpose_y_0"), val = tensor(false)]; + tensor values_89_cast_fp16 = transpose(perm = var_5893, x = values_87_cast_fp16)[name = tensor("transpose_159")]; + tensor attn_output_29_cast_fp16 = matmul(transpose_x = attn_output_29_transpose_x_0, transpose_y = attn_output_29_transpose_y_0, x = attn_weights_89_cast_fp16, y = values_89_cast_fp16)[name = tensor("attn_output_29_cast_fp16")]; + tensor var_5961 = const()[name = tensor("op_5961"), val = tensor([0, 2, 1, 3])]; + tensor var_5966 = const()[name = tensor("op_5966"), val = tensor([1, 1, 1024])]; + tensor var_5962_cast_fp16 = transpose(perm = var_5961, x = attn_output_29_cast_fp16)[name = tensor("transpose_156")]; + tensor input_175_cast_fp16 = reshape(shape = var_5966, x = var_5962_cast_fp16)[name = tensor("input_175_cast_fp16")]; + tensor attn14_out_proj_weight_to_fp16 = const()[name = tensor("attn14_out_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(359873024)))]; + tensor linear_58_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = attn14_out_proj_weight_to_fp16, x = input_175_cast_fp16)[name = tensor("linear_58_cast_fp16")]; + tensor var_5972_to_fp16 = const()[name = tensor("op_5972_to_fp16"), val = tensor(0x1p+0)]; + tensor var_5973_cast_fp16 = add(x = position14_to_fp16, y = var_5972_to_fp16)[name = tensor("op_5973_cast_fp16")]; + tensor var_5973_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_5973_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor input_177_cast_fp16 = add(x = input_171_cast_fp16, y = linear_58_cast_fp16)[name = tensor("input_177_cast_fp16")]; + tensor input_179_axes_0 = const()[name = tensor("input_179_axes_0"), val = tensor([-1])]; + tensor norm14_2_weight_to_fp16 = const()[name = tensor("norm14_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(361970240)))]; + tensor norm14_2_bias_to_fp16 = const()[name = tensor("norm14_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(361972352)))]; + tensor var_5977_to_fp16 = const()[name = tensor("op_5977_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_179_cast_fp16 = layer_norm(axes = input_179_axes_0, beta = norm14_2_bias_to_fp16, epsilon = var_5977_to_fp16, gamma = norm14_2_weight_to_fp16, x = input_177_cast_fp16)[name = tensor("input_179_cast_fp16")]; + tensor linear14_1_weight_to_fp16 = const()[name = tensor("linear14_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(361974464)))]; + tensor linear_59_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = linear14_1_weight_to_fp16, x = input_179_cast_fp16)[name = tensor("linear_59_cast_fp16")]; + tensor input_181_mode_0 = const()[name = tensor("input_181_mode_0"), val = tensor("EXACT")]; + tensor input_181_cast_fp16 = gelu(mode = input_181_mode_0, x = linear_59_cast_fp16)[name = tensor("input_181_cast_fp16")]; + tensor linear14_2_weight_to_fp16 = const()[name = tensor("linear14_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(370363136)))]; + tensor linear_60_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = linear14_2_weight_to_fp16, x = input_181_cast_fp16)[name = tensor("linear_60_cast_fp16")]; + tensor input_183_cast_fp16 = add(x = input_177_cast_fp16, y = linear_60_cast_fp16)[name = tensor("input_183_cast_fp16")]; + tensor input_185_axes_0 = const()[name = tensor("input_185_axes_0"), val = tensor([-1])]; + tensor norm15_1_weight_to_fp16 = const()[name = tensor("norm15_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(378751808)))]; + tensor norm15_1_bias_to_fp16 = const()[name = tensor("norm15_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(378753920)))]; + tensor var_5994_to_fp16 = const()[name = tensor("op_5994_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_185_cast_fp16 = layer_norm(axes = input_185_axes_0, beta = norm15_1_bias_to_fp16, epsilon = var_5994_to_fp16, gamma = norm15_1_weight_to_fp16, x = input_183_cast_fp16)[name = tensor("input_185_cast_fp16")]; + tensor attn15_in_proj_weight_to_fp16 = const()[name = tensor("attn15_in_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(378756032)))]; + tensor linear_61_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = attn15_in_proj_weight_to_fp16, x = input_185_cast_fp16)[name = tensor("linear_61_cast_fp16")]; + tensor var_6008 = const()[name = tensor("op_6008"), val = tensor([1, 1, 3, 16, 64])]; + tensor qkv_31_cast_fp16 = reshape(shape = var_6008, x = linear_61_cast_fp16)[name = tensor("qkv_31_cast_fp16")]; + tensor q_91_begin_0 = const()[name = tensor("q_91_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor q_91_end_0 = const()[name = tensor("q_91_end_0"), val = tensor([1, 1, 1, 16, 64])]; + tensor q_91_end_mask_0 = const()[name = tensor("q_91_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor q_91_squeeze_mask_0 = const()[name = tensor("q_91_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor q_91_cast_fp16 = slice_by_index(begin = q_91_begin_0, end = q_91_end_0, end_mask = q_91_end_mask_0, squeeze_mask = q_91_squeeze_mask_0, x = qkv_31_cast_fp16)[name = tensor("q_91_cast_fp16")]; + tensor k_61_begin_0 = const()[name = tensor("k_61_begin_0"), val = tensor([0, 0, 1, 0, 0])]; + tensor k_61_end_0 = const()[name = tensor("k_61_end_0"), val = tensor([1, 1, 2, 16, 64])]; + tensor k_61_end_mask_0 = const()[name = tensor("k_61_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor k_61_squeeze_mask_0 = const()[name = tensor("k_61_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor k_61_cast_fp16 = slice_by_index(begin = k_61_begin_0, end = k_61_end_0, end_mask = k_61_end_mask_0, squeeze_mask = k_61_squeeze_mask_0, x = qkv_31_cast_fp16)[name = tensor("k_61_cast_fp16")]; + tensor v_31_begin_0 = const()[name = tensor("v_31_begin_0"), val = tensor([0, 0, 2, 0, 0])]; + tensor v_31_end_0 = const()[name = tensor("v_31_end_0"), val = tensor([1, 1, 3, 16, 64])]; + tensor v_31_end_mask_0 = const()[name = tensor("v_31_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor v_31_squeeze_mask_0 = const()[name = tensor("v_31_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor v_31_cast_fp16 = slice_by_index(begin = v_31_begin_0, end = v_31_end_0, end_mask = v_31_end_mask_0, squeeze_mask = v_31_squeeze_mask_0, x = qkv_31_cast_fp16)[name = tensor("v_31_cast_fp16")]; + tensor var_6074 = const()[name = tensor("op_6074"), val = tensor([1, 1, 1, 1])]; + tensor position15_to_fp16_dtype_0 = const()[name = tensor("position15_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor position15_to_fp16 = cast(dtype = position15_to_fp16_dtype_0, x = position15)[name = tensor("cast_478")]; + tensor ts_95_cast_fp16 = reshape(shape = var_6074, x = position15_to_fp16)[name = tensor("ts_95_cast_fp16")]; + tensor var_6081 = const()[name = tensor("op_6081"), val = tensor([1, 1, 16, 32, 2])]; + tensor q_complex_31_cast_fp16 = reshape(shape = var_6081, x = q_91_cast_fp16)[name = tensor("q_complex_31_cast_fp16")]; + tensor var_6088 = const()[name = tensor("op_6088"), val = tensor([1, 1, 16, 32, 2])]; + tensor k_complex_31_cast_fp16 = reshape(shape = var_6088, x = k_61_cast_fp16)[name = tensor("k_complex_31_cast_fp16")]; + tensor var_6092_begin_0 = const()[name = tensor("op_6092_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_6092_end_0 = const()[name = tensor("op_6092_end_0"), val = tensor([1, 1, 16, 32, 1])]; + tensor var_6092_end_mask_0 = const()[name = tensor("op_6092_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_6092_squeeze_mask_0 = const()[name = tensor("op_6092_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_6092_cast_fp16 = slice_by_index(begin = var_6092_begin_0, end = var_6092_end_0, end_mask = var_6092_end_mask_0, squeeze_mask = var_6092_squeeze_mask_0, x = q_complex_31_cast_fp16)[name = tensor("op_6092_cast_fp16")]; + tensor var_6100_begin_0 = const()[name = tensor("op_6100_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_6100_end_0 = const()[name = tensor("op_6100_end_0"), val = tensor([1, 1, 16, 32, 2])]; + tensor var_6100_end_mask_0 = const()[name = tensor("op_6100_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_6100_squeeze_mask_0 = const()[name = tensor("op_6100_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_6100_cast_fp16 = slice_by_index(begin = var_6100_begin_0, end = var_6100_end_0, end_mask = var_6100_end_mask_0, squeeze_mask = var_6100_squeeze_mask_0, x = q_complex_31_cast_fp16)[name = tensor("op_6100_cast_fp16")]; + tensor var_6108_begin_0 = const()[name = tensor("op_6108_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_6108_end_0 = const()[name = tensor("op_6108_end_0"), val = tensor([1, 1, 16, 32, 1])]; + tensor var_6108_end_mask_0 = const()[name = tensor("op_6108_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_6108_squeeze_mask_0 = const()[name = tensor("op_6108_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_6108_cast_fp16 = slice_by_index(begin = var_6108_begin_0, end = var_6108_end_0, end_mask = var_6108_end_mask_0, squeeze_mask = var_6108_squeeze_mask_0, x = k_complex_31_cast_fp16)[name = tensor("op_6108_cast_fp16")]; + tensor var_6116_begin_0 = const()[name = tensor("op_6116_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_6116_end_0 = const()[name = tensor("op_6116_end_0"), val = tensor([1, 1, 16, 32, 2])]; + tensor var_6116_end_mask_0 = const()[name = tensor("op_6116_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_6116_squeeze_mask_0 = const()[name = tensor("op_6116_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_6116_cast_fp16 = slice_by_index(begin = var_6116_begin_0, end = var_6116_end_0, end_mask = var_6116_end_mask_0, squeeze_mask = var_6116_squeeze_mask_0, x = k_complex_31_cast_fp16)[name = tensor("op_6116_cast_fp16")]; + tensor freqs_31_to_fp16 = const()[name = tensor("freqs_31_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(385047552)))]; + tensor var_6122_cast_fp16 = mul(x = freqs_31_to_fp16, y = ts_95_cast_fp16)[name = tensor("op_6122_cast_fp16")]; + tensor rotr_31_cast_fp16 = cos(x = var_6122_cast_fp16)[name = tensor("rotr_31_cast_fp16")]; + tensor roti_31_cast_fp16 = sin(x = var_6122_cast_fp16)[name = tensor("roti_31_cast_fp16")]; + tensor var_6126_cast_fp16 = mul(x = var_6092_cast_fp16, y = rotr_31_cast_fp16)[name = tensor("op_6126_cast_fp16")]; + tensor var_6127_cast_fp16 = mul(x = var_6100_cast_fp16, y = roti_31_cast_fp16)[name = tensor("op_6127_cast_fp16")]; + tensor qor_61_cast_fp16 = sub(x = var_6126_cast_fp16, y = var_6127_cast_fp16)[name = tensor("qor_61_cast_fp16")]; + tensor var_6130_cast_fp16 = mul(x = var_6092_cast_fp16, y = roti_31_cast_fp16)[name = tensor("op_6130_cast_fp16")]; + tensor var_6131_cast_fp16 = mul(x = var_6100_cast_fp16, y = rotr_31_cast_fp16)[name = tensor("op_6131_cast_fp16")]; + tensor qoi_61_cast_fp16 = add(x = var_6130_cast_fp16, y = var_6131_cast_fp16)[name = tensor("qoi_61_cast_fp16")]; + tensor var_6134_cast_fp16 = mul(x = var_6108_cast_fp16, y = rotr_31_cast_fp16)[name = tensor("op_6134_cast_fp16")]; + tensor var_6135_cast_fp16 = mul(x = var_6116_cast_fp16, y = roti_31_cast_fp16)[name = tensor("op_6135_cast_fp16")]; + tensor kor_61_cast_fp16 = sub(x = var_6134_cast_fp16, y = var_6135_cast_fp16)[name = tensor("kor_61_cast_fp16")]; + tensor var_6138_cast_fp16 = mul(x = var_6108_cast_fp16, y = roti_31_cast_fp16)[name = tensor("op_6138_cast_fp16")]; + tensor var_6139_cast_fp16 = mul(x = var_6116_cast_fp16, y = rotr_31_cast_fp16)[name = tensor("op_6139_cast_fp16")]; + tensor koi_61_cast_fp16 = add(x = var_6138_cast_fp16, y = var_6139_cast_fp16)[name = tensor("koi_61_cast_fp16")]; + tensor qo_31_axis_0 = const()[name = tensor("qo_31_axis_0"), val = tensor(-1)]; + tensor qo_31_cast_fp16 = stack(axis = qo_31_axis_0, values = (qor_61_cast_fp16, qoi_61_cast_fp16))[name = tensor("qo_31_cast_fp16")]; + tensor ko_31_axis_0 = const()[name = tensor("ko_31_axis_0"), val = tensor(-1)]; + tensor ko_31_cast_fp16 = stack(axis = ko_31_axis_0, values = (kor_61_cast_fp16, koi_61_cast_fp16))[name = tensor("ko_31_cast_fp16")]; + tensor var_6172 = const()[name = tensor("op_6172"), val = tensor([1, 1, 16, 64])]; + tensor q_93_cast_fp16 = reshape(shape = var_6172, x = qo_31_cast_fp16)[name = tensor("q_93_cast_fp16")]; + tensor var_6178 = const()[name = tensor("op_6178"), val = tensor([1, 1, 16, 64])]; + tensor k_63_cast_fp16 = reshape(shape = var_6178, x = ko_31_cast_fp16)[name = tensor("k_63_cast_fp16")]; + tensor _inversed_6203_y_0_to_fp16 = const()[name = tensor("_inversed_6203_y_0_to_fp16"), val = tensor(0x1p-9)]; + tensor _inversed_6203_cast_fp16 = mul(x = ts_95_cast_fp16, y = _inversed_6203_y_0_to_fp16)[name = tensor("_inversed_6203_cast_fp16")]; + tensor var_6204_cast_fp16 = floor(x = _inversed_6203_cast_fp16)[name = tensor("op_6204_cast_fp16")]; + tensor var_6205_to_fp16 = const()[name = tensor("op_6205_to_fp16"), val = tensor(0x1p+9)]; + tensor var_6206_cast_fp16 = mul(x = var_6204_cast_fp16, y = var_6205_to_fp16)[name = tensor("op_6206_cast_fp16")]; + tensor write_indices_float_63_cast_fp16 = sub(x = ts_95_cast_fp16, y = var_6206_cast_fp16)[name = tensor("write_indices_float_63_cast_fp16")]; + tensor var_6213_dtype_0 = const()[name = tensor("op_6213_dtype_0"), val = tensor("int32")]; + tensor write_indices_31_reps_0 = const()[name = tensor("write_indices_31_reps_0"), val = tensor([1, 1, 16, 64])]; + tensor write_indices_float_63_cast_fp16_to_int32 = cast(dtype = var_6213_dtype_0, x = write_indices_float_63_cast_fp16)[name = tensor("cast_477")]; + tensor write_indices_31 = tile(reps = write_indices_31_reps_0, x = write_indices_float_63_cast_fp16_to_int32)[name = tensor("write_indices_31")]; + tensor var_6223_begin_0 = const()[name = tensor("op_6223_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_6223_end_0 = const()[name = tensor("op_6223_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor var_6223_end_mask_0 = const()[name = tensor("op_6223_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_6223_squeeze_mask_0 = const()[name = tensor("op_6223_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor cache15_to_fp16_dtype_0 = const()[name = tensor("cache15_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor cache15_to_fp16 = cast(dtype = cache15_to_fp16_dtype_0, x = cache15)[name = tensor("cast_476")]; + tensor var_6223_cast_fp16 = slice_by_index(begin = var_6223_begin_0, end = var_6223_end_0, end_mask = var_6223_end_mask_0, squeeze_mask = var_6223_squeeze_mask_0, x = cache15_to_fp16)[name = tensor("op_6223_cast_fp16")]; + tensor var_6225_axis_0 = const()[name = tensor("op_6225_axis_0"), val = tensor(1)]; + tensor var_6225_mode_0 = const()[name = tensor("op_6225_mode_0"), val = tensor("update")]; + tensor var_6225_validate_indices_0 = const()[name = tensor("op_6225_validate_indices_0"), val = tensor(false)]; + tensor var_6225_cast_fp16 = scatter_along_axis(axis = var_6225_axis_0, data = var_6223_cast_fp16, indices = write_indices_31, mode = var_6225_mode_0, updates = k_63_cast_fp16, validate_indices = var_6225_validate_indices_0)[name = tensor("op_6225_cast_fp16")]; + tensor concat_107 = const()[name = tensor("concat_107"), val = tensor([0, 0, 0, 0, 0])]; + tensor concat_108 = const()[name = tensor("concat_108"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_31_internal_tensor_assign_1_stride_0 = const()[name = tensor("new_cache_31_internal_tensor_assign_1_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_31_internal_tensor_assign_1_begin_mask_0 = const()[name = tensor("new_cache_31_internal_tensor_assign_1_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_31_internal_tensor_assign_1_end_mask_0 = const()[name = tensor("new_cache_31_internal_tensor_assign_1_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_31_internal_tensor_assign_1_squeeze_mask_0 = const()[name = tensor("new_cache_31_internal_tensor_assign_1_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_78 = const()[name = tensor("shape_78"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_30 = const()[name = tensor("reduce_prod_30"), val = tensor(1048576)]; + tensor range_1d_30_start_0 = const()[name = tensor("range_1d_30_start_0"), val = tensor(0)]; + tensor range_1d_30_step_0 = const()[name = tensor("range_1d_30_step_0"), val = tensor(1)]; + tensor range_1d_30 = range_1d(end = reduce_prod_30, start = range_1d_30_start_0, step = range_1d_30_step_0)[name = tensor("range_1d_30")]; + tensor reshape_150 = reshape(shape = shape_78, x = range_1d_30)[name = tensor("reshape_150")]; + tensor slice_by_index_30 = slice_by_index(begin = concat_107, begin_mask = new_cache_31_internal_tensor_assign_1_begin_mask_0, end = concat_108, end_mask = new_cache_31_internal_tensor_assign_1_end_mask_0, squeeze_mask = new_cache_31_internal_tensor_assign_1_squeeze_mask_0, stride = new_cache_31_internal_tensor_assign_1_stride_0, x = reshape_150)[name = tensor("slice_by_index_30")]; + tensor reshape_151_shape_0 = const()[name = tensor("reshape_151_shape_0"), val = tensor([-1])]; + tensor reshape_151 = reshape(shape = reshape_151_shape_0, x = slice_by_index_30)[name = tensor("reshape_151")]; + tensor reshape_152_shape_0 = const()[name = tensor("reshape_152_shape_0"), val = tensor([-1])]; + tensor reshape_152_cast_fp16 = reshape(shape = reshape_152_shape_0, x = var_6225_cast_fp16)[name = tensor("reshape_152_cast_fp16")]; + tensor reshape_153_shape_0 = const()[name = tensor("reshape_153_shape_0"), val = tensor([-1])]; + tensor reshape_153_cast_fp16 = reshape(shape = reshape_153_shape_0, x = cache15_to_fp16)[name = tensor("reshape_153_cast_fp16")]; + tensor scatter_30_mode_0 = const()[name = tensor("scatter_30_mode_0"), val = tensor("update")]; + tensor scatter_30_axis_0 = const()[name = tensor("scatter_30_axis_0"), val = tensor(0)]; + tensor scatter_30_validate_indices_0 = const()[name = tensor("scatter_30_validate_indices_0"), val = tensor(false)]; + tensor scatter_30_cast_fp16 = scatter(axis = scatter_30_axis_0, data = reshape_153_cast_fp16, indices = reshape_151, mode = scatter_30_mode_0, updates = reshape_152_cast_fp16, validate_indices = scatter_30_validate_indices_0)[name = tensor("scatter_30_cast_fp16")]; + tensor reshape_154_cast_fp16 = reshape(shape = shape_78, x = scatter_30_cast_fp16)[name = tensor("reshape_154_cast_fp16")]; + tensor var_6233_begin_0 = const()[name = tensor("op_6233_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor var_6233_end_0 = const()[name = tensor("op_6233_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor var_6233_end_mask_0 = const()[name = tensor("op_6233_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_6233_squeeze_mask_0 = const()[name = tensor("op_6233_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor var_6233_cast_fp16 = slice_by_index(begin = var_6233_begin_0, end = var_6233_end_0, end_mask = var_6233_end_mask_0, squeeze_mask = var_6233_squeeze_mask_0, x = reshape_154_cast_fp16)[name = tensor("op_6233_cast_fp16")]; + tensor var_6235_axis_0 = const()[name = tensor("op_6235_axis_0"), val = tensor(1)]; + tensor var_6235_mode_0 = const()[name = tensor("op_6235_mode_0"), val = tensor("update")]; + tensor var_6235_validate_indices_0 = const()[name = tensor("op_6235_validate_indices_0"), val = tensor(false)]; + tensor var_6235_cast_fp16 = scatter_along_axis(axis = var_6235_axis_0, data = var_6233_cast_fp16, indices = write_indices_31, mode = var_6235_mode_0, updates = v_31_cast_fp16, validate_indices = var_6235_validate_indices_0)[name = tensor("op_6235_cast_fp16")]; + tensor concat_109 = const()[name = tensor("concat_109"), val = tensor([1, 0, 0, 0, 0])]; + tensor concat_110 = const()[name = tensor("concat_110"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_31_internal_tensor_assign_2_stride_0 = const()[name = tensor("new_cache_31_internal_tensor_assign_2_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_31_internal_tensor_assign_2_begin_mask_0 = const()[name = tensor("new_cache_31_internal_tensor_assign_2_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_31_internal_tensor_assign_2_end_mask_0 = const()[name = tensor("new_cache_31_internal_tensor_assign_2_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_31_internal_tensor_assign_2_squeeze_mask_0 = const()[name = tensor("new_cache_31_internal_tensor_assign_2_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_79 = const()[name = tensor("shape_79"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_31 = const()[name = tensor("reduce_prod_31"), val = tensor(1048576)]; + tensor range_1d_31_start_0 = const()[name = tensor("range_1d_31_start_0"), val = tensor(0)]; + tensor range_1d_31_step_0 = const()[name = tensor("range_1d_31_step_0"), val = tensor(1)]; + tensor range_1d_31 = range_1d(end = reduce_prod_31, start = range_1d_31_start_0, step = range_1d_31_step_0)[name = tensor("range_1d_31")]; + tensor reshape_155 = reshape(shape = shape_79, x = range_1d_31)[name = tensor("reshape_155")]; + tensor slice_by_index_31 = slice_by_index(begin = concat_109, begin_mask = new_cache_31_internal_tensor_assign_2_begin_mask_0, end = concat_110, end_mask = new_cache_31_internal_tensor_assign_2_end_mask_0, squeeze_mask = new_cache_31_internal_tensor_assign_2_squeeze_mask_0, stride = new_cache_31_internal_tensor_assign_2_stride_0, x = reshape_155)[name = tensor("slice_by_index_31")]; + tensor reshape_156_shape_0 = const()[name = tensor("reshape_156_shape_0"), val = tensor([-1])]; + tensor reshape_156 = reshape(shape = reshape_156_shape_0, x = slice_by_index_31)[name = tensor("reshape_156")]; + tensor reshape_157_shape_0 = const()[name = tensor("reshape_157_shape_0"), val = tensor([-1])]; + tensor reshape_157_cast_fp16 = reshape(shape = reshape_157_shape_0, x = var_6235_cast_fp16)[name = tensor("reshape_157_cast_fp16")]; + tensor reshape_158_shape_0 = const()[name = tensor("reshape_158_shape_0"), val = tensor([-1])]; + tensor reshape_158_cast_fp16 = reshape(shape = reshape_158_shape_0, x = reshape_154_cast_fp16)[name = tensor("reshape_158_cast_fp16")]; + tensor scatter_31_mode_0 = const()[name = tensor("scatter_31_mode_0"), val = tensor("update")]; + tensor scatter_31_axis_0 = const()[name = tensor("scatter_31_axis_0"), val = tensor(0)]; + tensor scatter_31_validate_indices_0 = const()[name = tensor("scatter_31_validate_indices_0"), val = tensor(false)]; + tensor scatter_31_cast_fp16 = scatter(axis = scatter_31_axis_0, data = reshape_158_cast_fp16, indices = reshape_156, mode = scatter_31_mode_0, updates = reshape_157_cast_fp16, validate_indices = scatter_31_validate_indices_0)[name = tensor("scatter_31_cast_fp16")]; + tensor reshape_159_cast_fp16 = reshape(shape = shape_79, x = scatter_31_cast_fp16)[name = tensor("reshape_159_cast_fp16")]; + tensor reshape_159_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("reshape_159_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor keys_91_begin_0 = const()[name = tensor("keys_91_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor keys_91_end_0 = const()[name = tensor("keys_91_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor keys_91_end_mask_0 = const()[name = tensor("keys_91_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor keys_91_squeeze_mask_0 = const()[name = tensor("keys_91_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor keys_91_cast_fp16 = slice_by_index(begin = keys_91_begin_0, end = keys_91_end_0, end_mask = keys_91_end_mask_0, squeeze_mask = keys_91_squeeze_mask_0, x = reshape_159_cast_fp16)[name = tensor("keys_91_cast_fp16")]; + tensor values_91_begin_0 = const()[name = tensor("values_91_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor values_91_end_0 = const()[name = tensor("values_91_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor values_91_end_mask_0 = const()[name = tensor("values_91_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor values_91_squeeze_mask_0 = const()[name = tensor("values_91_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor values_91_cast_fp16 = slice_by_index(begin = values_91_begin_0, end = values_91_end_0, end_mask = values_91_end_mask_0, squeeze_mask = values_91_squeeze_mask_0, x = reshape_159_cast_fp16)[name = tensor("values_91_cast_fp16")]; + tensor var_6247_cast_fp16 = not_equal(x = keys_91_cast_fp16, y = keys_91_cast_fp16)[name = tensor("op_6247_cast_fp16")]; + tensor keys_93_cast_fp16 = select(a = var_463_to_fp16, b = keys_91_cast_fp16, cond = var_6247_cast_fp16)[name = tensor("keys_93_cast_fp16")]; + tensor var_6255_cast_fp16 = not_equal(x = values_91_cast_fp16, y = values_91_cast_fp16)[name = tensor("op_6255_cast_fp16")]; + tensor values_93_cast_fp16 = select(a = var_463_to_fp16, b = values_91_cast_fp16, cond = var_6255_cast_fp16)[name = tensor("values_93_cast_fp16")]; + tensor var_6279 = const()[name = tensor("op_6279"), val = tensor([0, 2, 1, 3])]; + tensor var_6295 = const()[name = tensor("op_6295"), val = tensor([1, 1, 1])]; + tensor var_6296_cast_fp16 = reshape(shape = var_6295, x = position15_to_fp16)[name = tensor("op_6296_cast_fp16")]; + tensor var_6316_to_fp16 = const()[name = tensor("op_6316_to_fp16"), val = tensor(0x1p+0)]; + tensor valid_len_31_cast_fp16 = add(x = var_6296_cast_fp16, y = var_6316_to_fp16)[name = tensor("valid_len_31_cast_fp16")]; + tensor valid_mask_31_cast_fp16 = less(x = k_positions_1_promoted_to_fp16, y = valid_len_31_cast_fp16)[name = tensor("valid_mask_31_cast_fp16")]; + tensor causal_mask_31_cast_fp16 = less_equal(x = k_positions_1_promoted_to_fp16, y = var_6296_cast_fp16)[name = tensor("causal_mask_31_cast_fp16")]; + tensor attn_mask_61 = logical_and(x = valid_mask_31_cast_fp16, y = causal_mask_31_cast_fp16)[name = tensor("attn_mask_61")]; + tensor attn_mask_63_axes_0 = const()[name = tensor("attn_mask_63_axes_0"), val = tensor([1])]; + tensor attn_mask_63 = expand_dims(axes = attn_mask_63_axes_0, x = attn_mask_61)[name = tensor("attn_mask_63")]; + tensor var_6334_transpose_x_0 = const()[name = tensor("op_6334_transpose_x_0"), val = tensor(false)]; + tensor var_6334_transpose_y_0 = const()[name = tensor("op_6334_transpose_y_0"), val = tensor(false)]; + tensor transpose_102_perm_0 = const()[name = tensor("transpose_102_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_103_perm_0 = const()[name = tensor("transpose_103_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_103 = transpose(perm = transpose_103_perm_0, x = keys_93_cast_fp16)[name = tensor("transpose_153")]; + tensor transpose_102 = transpose(perm = transpose_102_perm_0, x = q_93_cast_fp16)[name = tensor("transpose_154")]; + tensor var_6334_cast_fp16 = matmul(transpose_x = var_6334_transpose_x_0, transpose_y = var_6334_transpose_y_0, x = transpose_102, y = transpose_103)[name = tensor("op_6334_cast_fp16")]; + tensor var_6328_to_fp16 = const()[name = tensor("op_6328_to_fp16"), val = tensor([0x1p-3])]; + tensor attn_weights_91_cast_fp16 = mul(x = var_6334_cast_fp16, y = var_6328_to_fp16)[name = tensor("attn_weights_91_cast_fp16")]; + tensor var_6336 = logical_not(x = attn_mask_63)[name = tensor("op_6336")]; + tensor var_6337_to_fp16 = const()[name = tensor("op_6337_to_fp16"), val = tensor(-inf)]; + tensor attn_weights_93_cast_fp16 = select(a = var_6337_to_fp16, b = attn_weights_91_cast_fp16, cond = var_6336)[name = tensor("attn_weights_93_cast_fp16")]; + tensor var_6339 = const()[name = tensor("op_6339"), val = tensor(-1)]; + tensor attn_weights_95_cast_fp16 = softmax(axis = var_6339, x = attn_weights_93_cast_fp16)[name = tensor("attn_weights_95_cast_fp16")]; + tensor attn_output_31_transpose_x_0 = const()[name = tensor("attn_output_31_transpose_x_0"), val = tensor(false)]; + tensor attn_output_31_transpose_y_0 = const()[name = tensor("attn_output_31_transpose_y_0"), val = tensor(false)]; + tensor values_95_cast_fp16 = transpose(perm = var_6279, x = values_93_cast_fp16)[name = tensor("transpose_155")]; + tensor attn_output_31_cast_fp16 = matmul(transpose_x = attn_output_31_transpose_x_0, transpose_y = attn_output_31_transpose_y_0, x = attn_weights_95_cast_fp16, y = values_95_cast_fp16)[name = tensor("attn_output_31_cast_fp16")]; + tensor var_6347 = const()[name = tensor("op_6347"), val = tensor([0, 2, 1, 3])]; + tensor var_6352 = const()[name = tensor("op_6352"), val = tensor([1, 1, 1024])]; + tensor var_6348_cast_fp16 = transpose(perm = var_6347, x = attn_output_31_cast_fp16)[name = tensor("transpose_152")]; + tensor input_187_cast_fp16 = reshape(shape = var_6352, x = var_6348_cast_fp16)[name = tensor("input_187_cast_fp16")]; + tensor attn15_out_proj_weight_to_fp16 = const()[name = tensor("attn15_out_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(385047680)))]; + tensor linear_62_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = attn15_out_proj_weight_to_fp16, x = input_187_cast_fp16)[name = tensor("linear_62_cast_fp16")]; + tensor var_6358_to_fp16 = const()[name = tensor("op_6358_to_fp16"), val = tensor(0x1p+0)]; + tensor var_6359_cast_fp16 = add(x = position15_to_fp16, y = var_6358_to_fp16)[name = tensor("op_6359_cast_fp16")]; + tensor var_6359_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_6359_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor input_189_cast_fp16 = add(x = input_183_cast_fp16, y = linear_62_cast_fp16)[name = tensor("input_189_cast_fp16")]; + tensor input_191_axes_0 = const()[name = tensor("input_191_axes_0"), val = tensor([-1])]; + tensor norm15_2_weight_to_fp16 = const()[name = tensor("norm15_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(387144896)))]; + tensor norm15_2_bias_to_fp16 = const()[name = tensor("norm15_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(387147008)))]; + tensor var_6363_to_fp16 = const()[name = tensor("op_6363_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_191_cast_fp16 = layer_norm(axes = input_191_axes_0, beta = norm15_2_bias_to_fp16, epsilon = var_6363_to_fp16, gamma = norm15_2_weight_to_fp16, x = input_189_cast_fp16)[name = tensor("input_191_cast_fp16")]; + tensor linear15_1_weight_to_fp16 = const()[name = tensor("linear15_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(387149120)))]; + tensor linear_63_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = linear15_1_weight_to_fp16, x = input_191_cast_fp16)[name = tensor("linear_63_cast_fp16")]; + tensor input_193_mode_0 = const()[name = tensor("input_193_mode_0"), val = tensor("EXACT")]; + tensor input_193_cast_fp16 = gelu(mode = input_193_mode_0, x = linear_63_cast_fp16)[name = tensor("input_193_cast_fp16")]; + tensor linear15_2_weight_to_fp16 = const()[name = tensor("linear15_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(395537792)))]; + tensor linear_64_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = linear15_2_weight_to_fp16, x = input_193_cast_fp16)[name = tensor("linear_64_cast_fp16")]; + tensor input_195_cast_fp16 = add(x = input_189_cast_fp16, y = linear_64_cast_fp16)[name = tensor("input_195_cast_fp16")]; + tensor input_197_axes_0 = const()[name = tensor("input_197_axes_0"), val = tensor([-1])]; + tensor norm16_1_weight_to_fp16 = const()[name = tensor("norm16_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(403926464)))]; + tensor norm16_1_bias_to_fp16 = const()[name = tensor("norm16_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(403928576)))]; + tensor var_6380_to_fp16 = const()[name = tensor("op_6380_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_197_cast_fp16 = layer_norm(axes = input_197_axes_0, beta = norm16_1_bias_to_fp16, epsilon = var_6380_to_fp16, gamma = norm16_1_weight_to_fp16, x = input_195_cast_fp16)[name = tensor("input_197_cast_fp16")]; + tensor attn16_in_proj_weight_to_fp16 = const()[name = tensor("attn16_in_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(403930688)))]; + tensor linear_65_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = attn16_in_proj_weight_to_fp16, x = input_197_cast_fp16)[name = tensor("linear_65_cast_fp16")]; + tensor var_6394 = const()[name = tensor("op_6394"), val = tensor([1, 1, 3, 16, 64])]; + tensor qkv_33_cast_fp16 = reshape(shape = var_6394, x = linear_65_cast_fp16)[name = tensor("qkv_33_cast_fp16")]; + tensor q_97_begin_0 = const()[name = tensor("q_97_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor q_97_end_0 = const()[name = tensor("q_97_end_0"), val = tensor([1, 1, 1, 16, 64])]; + tensor q_97_end_mask_0 = const()[name = tensor("q_97_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor q_97_squeeze_mask_0 = const()[name = tensor("q_97_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor q_97_cast_fp16 = slice_by_index(begin = q_97_begin_0, end = q_97_end_0, end_mask = q_97_end_mask_0, squeeze_mask = q_97_squeeze_mask_0, x = qkv_33_cast_fp16)[name = tensor("q_97_cast_fp16")]; + tensor k_65_begin_0 = const()[name = tensor("k_65_begin_0"), val = tensor([0, 0, 1, 0, 0])]; + tensor k_65_end_0 = const()[name = tensor("k_65_end_0"), val = tensor([1, 1, 2, 16, 64])]; + tensor k_65_end_mask_0 = const()[name = tensor("k_65_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor k_65_squeeze_mask_0 = const()[name = tensor("k_65_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor k_65_cast_fp16 = slice_by_index(begin = k_65_begin_0, end = k_65_end_0, end_mask = k_65_end_mask_0, squeeze_mask = k_65_squeeze_mask_0, x = qkv_33_cast_fp16)[name = tensor("k_65_cast_fp16")]; + tensor v_33_begin_0 = const()[name = tensor("v_33_begin_0"), val = tensor([0, 0, 2, 0, 0])]; + tensor v_33_end_0 = const()[name = tensor("v_33_end_0"), val = tensor([1, 1, 3, 16, 64])]; + tensor v_33_end_mask_0 = const()[name = tensor("v_33_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor v_33_squeeze_mask_0 = const()[name = tensor("v_33_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor v_33_cast_fp16 = slice_by_index(begin = v_33_begin_0, end = v_33_end_0, end_mask = v_33_end_mask_0, squeeze_mask = v_33_squeeze_mask_0, x = qkv_33_cast_fp16)[name = tensor("v_33_cast_fp16")]; + tensor var_6460 = const()[name = tensor("op_6460"), val = tensor([1, 1, 1, 1])]; + tensor position16_to_fp16_dtype_0 = const()[name = tensor("position16_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor position16_to_fp16 = cast(dtype = position16_to_fp16_dtype_0, x = position16)[name = tensor("cast_473")]; + tensor ts_101_cast_fp16 = reshape(shape = var_6460, x = position16_to_fp16)[name = tensor("ts_101_cast_fp16")]; + tensor var_6467 = const()[name = tensor("op_6467"), val = tensor([1, 1, 16, 32, 2])]; + tensor q_complex_33_cast_fp16 = reshape(shape = var_6467, x = q_97_cast_fp16)[name = tensor("q_complex_33_cast_fp16")]; + tensor var_6474 = const()[name = tensor("op_6474"), val = tensor([1, 1, 16, 32, 2])]; + tensor k_complex_33_cast_fp16 = reshape(shape = var_6474, x = k_65_cast_fp16)[name = tensor("k_complex_33_cast_fp16")]; + tensor var_6478_begin_0 = const()[name = tensor("op_6478_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_6478_end_0 = const()[name = tensor("op_6478_end_0"), val = tensor([1, 1, 16, 32, 1])]; + tensor var_6478_end_mask_0 = const()[name = tensor("op_6478_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_6478_squeeze_mask_0 = const()[name = tensor("op_6478_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_6478_cast_fp16 = slice_by_index(begin = var_6478_begin_0, end = var_6478_end_0, end_mask = var_6478_end_mask_0, squeeze_mask = var_6478_squeeze_mask_0, x = q_complex_33_cast_fp16)[name = tensor("op_6478_cast_fp16")]; + tensor var_6486_begin_0 = const()[name = tensor("op_6486_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_6486_end_0 = const()[name = tensor("op_6486_end_0"), val = tensor([1, 1, 16, 32, 2])]; + tensor var_6486_end_mask_0 = const()[name = tensor("op_6486_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_6486_squeeze_mask_0 = const()[name = tensor("op_6486_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_6486_cast_fp16 = slice_by_index(begin = var_6486_begin_0, end = var_6486_end_0, end_mask = var_6486_end_mask_0, squeeze_mask = var_6486_squeeze_mask_0, x = q_complex_33_cast_fp16)[name = tensor("op_6486_cast_fp16")]; + tensor var_6494_begin_0 = const()[name = tensor("op_6494_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_6494_end_0 = const()[name = tensor("op_6494_end_0"), val = tensor([1, 1, 16, 32, 1])]; + tensor var_6494_end_mask_0 = const()[name = tensor("op_6494_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_6494_squeeze_mask_0 = const()[name = tensor("op_6494_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_6494_cast_fp16 = slice_by_index(begin = var_6494_begin_0, end = var_6494_end_0, end_mask = var_6494_end_mask_0, squeeze_mask = var_6494_squeeze_mask_0, x = k_complex_33_cast_fp16)[name = tensor("op_6494_cast_fp16")]; + tensor var_6502_begin_0 = const()[name = tensor("op_6502_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_6502_end_0 = const()[name = tensor("op_6502_end_0"), val = tensor([1, 1, 16, 32, 2])]; + tensor var_6502_end_mask_0 = const()[name = tensor("op_6502_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_6502_squeeze_mask_0 = const()[name = tensor("op_6502_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_6502_cast_fp16 = slice_by_index(begin = var_6502_begin_0, end = var_6502_end_0, end_mask = var_6502_end_mask_0, squeeze_mask = var_6502_squeeze_mask_0, x = k_complex_33_cast_fp16)[name = tensor("op_6502_cast_fp16")]; + tensor freqs_33_to_fp16 = const()[name = tensor("freqs_33_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(410222208)))]; + tensor var_6508_cast_fp16 = mul(x = freqs_33_to_fp16, y = ts_101_cast_fp16)[name = tensor("op_6508_cast_fp16")]; + tensor rotr_33_cast_fp16 = cos(x = var_6508_cast_fp16)[name = tensor("rotr_33_cast_fp16")]; + tensor roti_33_cast_fp16 = sin(x = var_6508_cast_fp16)[name = tensor("roti_33_cast_fp16")]; + tensor var_6512_cast_fp16 = mul(x = var_6478_cast_fp16, y = rotr_33_cast_fp16)[name = tensor("op_6512_cast_fp16")]; + tensor var_6513_cast_fp16 = mul(x = var_6486_cast_fp16, y = roti_33_cast_fp16)[name = tensor("op_6513_cast_fp16")]; + tensor qor_65_cast_fp16 = sub(x = var_6512_cast_fp16, y = var_6513_cast_fp16)[name = tensor("qor_65_cast_fp16")]; + tensor var_6516_cast_fp16 = mul(x = var_6478_cast_fp16, y = roti_33_cast_fp16)[name = tensor("op_6516_cast_fp16")]; + tensor var_6517_cast_fp16 = mul(x = var_6486_cast_fp16, y = rotr_33_cast_fp16)[name = tensor("op_6517_cast_fp16")]; + tensor qoi_65_cast_fp16 = add(x = var_6516_cast_fp16, y = var_6517_cast_fp16)[name = tensor("qoi_65_cast_fp16")]; + tensor var_6520_cast_fp16 = mul(x = var_6494_cast_fp16, y = rotr_33_cast_fp16)[name = tensor("op_6520_cast_fp16")]; + tensor var_6521_cast_fp16 = mul(x = var_6502_cast_fp16, y = roti_33_cast_fp16)[name = tensor("op_6521_cast_fp16")]; + tensor kor_65_cast_fp16 = sub(x = var_6520_cast_fp16, y = var_6521_cast_fp16)[name = tensor("kor_65_cast_fp16")]; + tensor var_6524_cast_fp16 = mul(x = var_6494_cast_fp16, y = roti_33_cast_fp16)[name = tensor("op_6524_cast_fp16")]; + tensor var_6525_cast_fp16 = mul(x = var_6502_cast_fp16, y = rotr_33_cast_fp16)[name = tensor("op_6525_cast_fp16")]; + tensor koi_65_cast_fp16 = add(x = var_6524_cast_fp16, y = var_6525_cast_fp16)[name = tensor("koi_65_cast_fp16")]; + tensor qo_33_axis_0 = const()[name = tensor("qo_33_axis_0"), val = tensor(-1)]; + tensor qo_33_cast_fp16 = stack(axis = qo_33_axis_0, values = (qor_65_cast_fp16, qoi_65_cast_fp16))[name = tensor("qo_33_cast_fp16")]; + tensor ko_33_axis_0 = const()[name = tensor("ko_33_axis_0"), val = tensor(-1)]; + tensor ko_33_cast_fp16 = stack(axis = ko_33_axis_0, values = (kor_65_cast_fp16, koi_65_cast_fp16))[name = tensor("ko_33_cast_fp16")]; + tensor var_6558 = const()[name = tensor("op_6558"), val = tensor([1, 1, 16, 64])]; + tensor q_99_cast_fp16 = reshape(shape = var_6558, x = qo_33_cast_fp16)[name = tensor("q_99_cast_fp16")]; + tensor var_6564 = const()[name = tensor("op_6564"), val = tensor([1, 1, 16, 64])]; + tensor k_67_cast_fp16 = reshape(shape = var_6564, x = ko_33_cast_fp16)[name = tensor("k_67_cast_fp16")]; + tensor _inversed_6589_y_0_to_fp16 = const()[name = tensor("_inversed_6589_y_0_to_fp16"), val = tensor(0x1p-9)]; + tensor _inversed_6589_cast_fp16 = mul(x = ts_101_cast_fp16, y = _inversed_6589_y_0_to_fp16)[name = tensor("_inversed_6589_cast_fp16")]; + tensor var_6590_cast_fp16 = floor(x = _inversed_6589_cast_fp16)[name = tensor("op_6590_cast_fp16")]; + tensor var_6591_to_fp16 = const()[name = tensor("op_6591_to_fp16"), val = tensor(0x1p+9)]; + tensor var_6592_cast_fp16 = mul(x = var_6590_cast_fp16, y = var_6591_to_fp16)[name = tensor("op_6592_cast_fp16")]; + tensor write_indices_float_67_cast_fp16 = sub(x = ts_101_cast_fp16, y = var_6592_cast_fp16)[name = tensor("write_indices_float_67_cast_fp16")]; + tensor var_6599_dtype_0 = const()[name = tensor("op_6599_dtype_0"), val = tensor("int32")]; + tensor write_indices_33_reps_0 = const()[name = tensor("write_indices_33_reps_0"), val = tensor([1, 1, 16, 64])]; + tensor write_indices_float_67_cast_fp16_to_int32 = cast(dtype = var_6599_dtype_0, x = write_indices_float_67_cast_fp16)[name = tensor("cast_472")]; + tensor write_indices_33 = tile(reps = write_indices_33_reps_0, x = write_indices_float_67_cast_fp16_to_int32)[name = tensor("write_indices_33")]; + tensor var_6609_begin_0 = const()[name = tensor("op_6609_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_6609_end_0 = const()[name = tensor("op_6609_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor var_6609_end_mask_0 = const()[name = tensor("op_6609_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_6609_squeeze_mask_0 = const()[name = tensor("op_6609_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor cache16_to_fp16_dtype_0 = const()[name = tensor("cache16_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor cache16_to_fp16 = cast(dtype = cache16_to_fp16_dtype_0, x = cache16)[name = tensor("cast_471")]; + tensor var_6609_cast_fp16 = slice_by_index(begin = var_6609_begin_0, end = var_6609_end_0, end_mask = var_6609_end_mask_0, squeeze_mask = var_6609_squeeze_mask_0, x = cache16_to_fp16)[name = tensor("op_6609_cast_fp16")]; + tensor var_6611_axis_0 = const()[name = tensor("op_6611_axis_0"), val = tensor(1)]; + tensor var_6611_mode_0 = const()[name = tensor("op_6611_mode_0"), val = tensor("update")]; + tensor var_6611_validate_indices_0 = const()[name = tensor("op_6611_validate_indices_0"), val = tensor(false)]; + tensor var_6611_cast_fp16 = scatter_along_axis(axis = var_6611_axis_0, data = var_6609_cast_fp16, indices = write_indices_33, mode = var_6611_mode_0, updates = k_67_cast_fp16, validate_indices = var_6611_validate_indices_0)[name = tensor("op_6611_cast_fp16")]; + tensor concat_114 = const()[name = tensor("concat_114"), val = tensor([0, 0, 0, 0, 0])]; + tensor concat_115 = const()[name = tensor("concat_115"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_33_internal_tensor_assign_1_stride_0 = const()[name = tensor("new_cache_33_internal_tensor_assign_1_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_33_internal_tensor_assign_1_begin_mask_0 = const()[name = tensor("new_cache_33_internal_tensor_assign_1_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_33_internal_tensor_assign_1_end_mask_0 = const()[name = tensor("new_cache_33_internal_tensor_assign_1_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_33_internal_tensor_assign_1_squeeze_mask_0 = const()[name = tensor("new_cache_33_internal_tensor_assign_1_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_80 = const()[name = tensor("shape_80"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_32 = const()[name = tensor("reduce_prod_32"), val = tensor(1048576)]; + tensor range_1d_32_start_0 = const()[name = tensor("range_1d_32_start_0"), val = tensor(0)]; + tensor range_1d_32_step_0 = const()[name = tensor("range_1d_32_step_0"), val = tensor(1)]; + tensor range_1d_32 = range_1d(end = reduce_prod_32, start = range_1d_32_start_0, step = range_1d_32_step_0)[name = tensor("range_1d_32")]; + tensor reshape_160 = reshape(shape = shape_80, x = range_1d_32)[name = tensor("reshape_160")]; + tensor slice_by_index_32 = slice_by_index(begin = concat_114, begin_mask = new_cache_33_internal_tensor_assign_1_begin_mask_0, end = concat_115, end_mask = new_cache_33_internal_tensor_assign_1_end_mask_0, squeeze_mask = new_cache_33_internal_tensor_assign_1_squeeze_mask_0, stride = new_cache_33_internal_tensor_assign_1_stride_0, x = reshape_160)[name = tensor("slice_by_index_32")]; + tensor reshape_161_shape_0 = const()[name = tensor("reshape_161_shape_0"), val = tensor([-1])]; + tensor reshape_161 = reshape(shape = reshape_161_shape_0, x = slice_by_index_32)[name = tensor("reshape_161")]; + tensor reshape_162_shape_0 = const()[name = tensor("reshape_162_shape_0"), val = tensor([-1])]; + tensor reshape_162_cast_fp16 = reshape(shape = reshape_162_shape_0, x = var_6611_cast_fp16)[name = tensor("reshape_162_cast_fp16")]; + tensor reshape_163_shape_0 = const()[name = tensor("reshape_163_shape_0"), val = tensor([-1])]; + tensor reshape_163_cast_fp16 = reshape(shape = reshape_163_shape_0, x = cache16_to_fp16)[name = tensor("reshape_163_cast_fp16")]; + tensor scatter_32_mode_0 = const()[name = tensor("scatter_32_mode_0"), val = tensor("update")]; + tensor scatter_32_axis_0 = const()[name = tensor("scatter_32_axis_0"), val = tensor(0)]; + tensor scatter_32_validate_indices_0 = const()[name = tensor("scatter_32_validate_indices_0"), val = tensor(false)]; + tensor scatter_32_cast_fp16 = scatter(axis = scatter_32_axis_0, data = reshape_163_cast_fp16, indices = reshape_161, mode = scatter_32_mode_0, updates = reshape_162_cast_fp16, validate_indices = scatter_32_validate_indices_0)[name = tensor("scatter_32_cast_fp16")]; + tensor reshape_164_cast_fp16 = reshape(shape = shape_80, x = scatter_32_cast_fp16)[name = tensor("reshape_164_cast_fp16")]; + tensor var_6619_begin_0 = const()[name = tensor("op_6619_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor var_6619_end_0 = const()[name = tensor("op_6619_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor var_6619_end_mask_0 = const()[name = tensor("op_6619_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_6619_squeeze_mask_0 = const()[name = tensor("op_6619_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor var_6619_cast_fp16 = slice_by_index(begin = var_6619_begin_0, end = var_6619_end_0, end_mask = var_6619_end_mask_0, squeeze_mask = var_6619_squeeze_mask_0, x = reshape_164_cast_fp16)[name = tensor("op_6619_cast_fp16")]; + tensor var_6621_axis_0 = const()[name = tensor("op_6621_axis_0"), val = tensor(1)]; + tensor var_6621_mode_0 = const()[name = tensor("op_6621_mode_0"), val = tensor("update")]; + tensor var_6621_validate_indices_0 = const()[name = tensor("op_6621_validate_indices_0"), val = tensor(false)]; + tensor var_6621_cast_fp16 = scatter_along_axis(axis = var_6621_axis_0, data = var_6619_cast_fp16, indices = write_indices_33, mode = var_6621_mode_0, updates = v_33_cast_fp16, validate_indices = var_6621_validate_indices_0)[name = tensor("op_6621_cast_fp16")]; + tensor concat_116 = const()[name = tensor("concat_116"), val = tensor([1, 0, 0, 0, 0])]; + tensor concat_117 = const()[name = tensor("concat_117"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_33_internal_tensor_assign_2_stride_0 = const()[name = tensor("new_cache_33_internal_tensor_assign_2_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_33_internal_tensor_assign_2_begin_mask_0 = const()[name = tensor("new_cache_33_internal_tensor_assign_2_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_33_internal_tensor_assign_2_end_mask_0 = const()[name = tensor("new_cache_33_internal_tensor_assign_2_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_33_internal_tensor_assign_2_squeeze_mask_0 = const()[name = tensor("new_cache_33_internal_tensor_assign_2_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_81 = const()[name = tensor("shape_81"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_33 = const()[name = tensor("reduce_prod_33"), val = tensor(1048576)]; + tensor range_1d_33_start_0 = const()[name = tensor("range_1d_33_start_0"), val = tensor(0)]; + tensor range_1d_33_step_0 = const()[name = tensor("range_1d_33_step_0"), val = tensor(1)]; + tensor range_1d_33 = range_1d(end = reduce_prod_33, start = range_1d_33_start_0, step = range_1d_33_step_0)[name = tensor("range_1d_33")]; + tensor reshape_165 = reshape(shape = shape_81, x = range_1d_33)[name = tensor("reshape_165")]; + tensor slice_by_index_33 = slice_by_index(begin = concat_116, begin_mask = new_cache_33_internal_tensor_assign_2_begin_mask_0, end = concat_117, end_mask = new_cache_33_internal_tensor_assign_2_end_mask_0, squeeze_mask = new_cache_33_internal_tensor_assign_2_squeeze_mask_0, stride = new_cache_33_internal_tensor_assign_2_stride_0, x = reshape_165)[name = tensor("slice_by_index_33")]; + tensor reshape_166_shape_0 = const()[name = tensor("reshape_166_shape_0"), val = tensor([-1])]; + tensor reshape_166 = reshape(shape = reshape_166_shape_0, x = slice_by_index_33)[name = tensor("reshape_166")]; + tensor reshape_167_shape_0 = const()[name = tensor("reshape_167_shape_0"), val = tensor([-1])]; + tensor reshape_167_cast_fp16 = reshape(shape = reshape_167_shape_0, x = var_6621_cast_fp16)[name = tensor("reshape_167_cast_fp16")]; + tensor reshape_168_shape_0 = const()[name = tensor("reshape_168_shape_0"), val = tensor([-1])]; + tensor reshape_168_cast_fp16 = reshape(shape = reshape_168_shape_0, x = reshape_164_cast_fp16)[name = tensor("reshape_168_cast_fp16")]; + tensor scatter_33_mode_0 = const()[name = tensor("scatter_33_mode_0"), val = tensor("update")]; + tensor scatter_33_axis_0 = const()[name = tensor("scatter_33_axis_0"), val = tensor(0)]; + tensor scatter_33_validate_indices_0 = const()[name = tensor("scatter_33_validate_indices_0"), val = tensor(false)]; + tensor scatter_33_cast_fp16 = scatter(axis = scatter_33_axis_0, data = reshape_168_cast_fp16, indices = reshape_166, mode = scatter_33_mode_0, updates = reshape_167_cast_fp16, validate_indices = scatter_33_validate_indices_0)[name = tensor("scatter_33_cast_fp16")]; + tensor reshape_169_cast_fp16 = reshape(shape = shape_81, x = scatter_33_cast_fp16)[name = tensor("reshape_169_cast_fp16")]; + tensor reshape_169_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("reshape_169_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor keys_97_begin_0 = const()[name = tensor("keys_97_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor keys_97_end_0 = const()[name = tensor("keys_97_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor keys_97_end_mask_0 = const()[name = tensor("keys_97_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor keys_97_squeeze_mask_0 = const()[name = tensor("keys_97_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor keys_97_cast_fp16 = slice_by_index(begin = keys_97_begin_0, end = keys_97_end_0, end_mask = keys_97_end_mask_0, squeeze_mask = keys_97_squeeze_mask_0, x = reshape_169_cast_fp16)[name = tensor("keys_97_cast_fp16")]; + tensor values_97_begin_0 = const()[name = tensor("values_97_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor values_97_end_0 = const()[name = tensor("values_97_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor values_97_end_mask_0 = const()[name = tensor("values_97_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor values_97_squeeze_mask_0 = const()[name = tensor("values_97_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor values_97_cast_fp16 = slice_by_index(begin = values_97_begin_0, end = values_97_end_0, end_mask = values_97_end_mask_0, squeeze_mask = values_97_squeeze_mask_0, x = reshape_169_cast_fp16)[name = tensor("values_97_cast_fp16")]; + tensor var_6633_cast_fp16 = not_equal(x = keys_97_cast_fp16, y = keys_97_cast_fp16)[name = tensor("op_6633_cast_fp16")]; + tensor keys_99_cast_fp16 = select(a = var_463_to_fp16, b = keys_97_cast_fp16, cond = var_6633_cast_fp16)[name = tensor("keys_99_cast_fp16")]; + tensor var_6641_cast_fp16 = not_equal(x = values_97_cast_fp16, y = values_97_cast_fp16)[name = tensor("op_6641_cast_fp16")]; + tensor values_99_cast_fp16 = select(a = var_463_to_fp16, b = values_97_cast_fp16, cond = var_6641_cast_fp16)[name = tensor("values_99_cast_fp16")]; + tensor var_6665 = const()[name = tensor("op_6665"), val = tensor([0, 2, 1, 3])]; + tensor var_6681 = const()[name = tensor("op_6681"), val = tensor([1, 1, 1])]; + tensor var_6682_cast_fp16 = reshape(shape = var_6681, x = position16_to_fp16)[name = tensor("op_6682_cast_fp16")]; + tensor var_6702_to_fp16 = const()[name = tensor("op_6702_to_fp16"), val = tensor(0x1p+0)]; + tensor valid_len_33_cast_fp16 = add(x = var_6682_cast_fp16, y = var_6702_to_fp16)[name = tensor("valid_len_33_cast_fp16")]; + tensor valid_mask_33_cast_fp16 = less(x = k_positions_1_promoted_to_fp16, y = valid_len_33_cast_fp16)[name = tensor("valid_mask_33_cast_fp16")]; + tensor causal_mask_33_cast_fp16 = less_equal(x = k_positions_1_promoted_to_fp16, y = var_6682_cast_fp16)[name = tensor("causal_mask_33_cast_fp16")]; + tensor attn_mask_65 = logical_and(x = valid_mask_33_cast_fp16, y = causal_mask_33_cast_fp16)[name = tensor("attn_mask_65")]; + tensor attn_mask_67_axes_0 = const()[name = tensor("attn_mask_67_axes_0"), val = tensor([1])]; + tensor attn_mask_67 = expand_dims(axes = attn_mask_67_axes_0, x = attn_mask_65)[name = tensor("attn_mask_67")]; + tensor var_6720_transpose_x_0 = const()[name = tensor("op_6720_transpose_x_0"), val = tensor(false)]; + tensor var_6720_transpose_y_0 = const()[name = tensor("op_6720_transpose_y_0"), val = tensor(false)]; + tensor transpose_104_perm_0 = const()[name = tensor("transpose_104_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_105_perm_0 = const()[name = tensor("transpose_105_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_105 = transpose(perm = transpose_105_perm_0, x = keys_99_cast_fp16)[name = tensor("transpose_149")]; + tensor transpose_104 = transpose(perm = transpose_104_perm_0, x = q_99_cast_fp16)[name = tensor("transpose_150")]; + tensor var_6720_cast_fp16 = matmul(transpose_x = var_6720_transpose_x_0, transpose_y = var_6720_transpose_y_0, x = transpose_104, y = transpose_105)[name = tensor("op_6720_cast_fp16")]; + tensor var_6714_to_fp16 = const()[name = tensor("op_6714_to_fp16"), val = tensor([0x1p-3])]; + tensor attn_weights_97_cast_fp16 = mul(x = var_6720_cast_fp16, y = var_6714_to_fp16)[name = tensor("attn_weights_97_cast_fp16")]; + tensor var_6722 = logical_not(x = attn_mask_67)[name = tensor("op_6722")]; + tensor var_6723_to_fp16 = const()[name = tensor("op_6723_to_fp16"), val = tensor(-inf)]; + tensor attn_weights_99_cast_fp16 = select(a = var_6723_to_fp16, b = attn_weights_97_cast_fp16, cond = var_6722)[name = tensor("attn_weights_99_cast_fp16")]; + tensor var_6725 = const()[name = tensor("op_6725"), val = tensor(-1)]; + tensor attn_weights_101_cast_fp16 = softmax(axis = var_6725, x = attn_weights_99_cast_fp16)[name = tensor("attn_weights_101_cast_fp16")]; + tensor attn_output_33_transpose_x_0 = const()[name = tensor("attn_output_33_transpose_x_0"), val = tensor(false)]; + tensor attn_output_33_transpose_y_0 = const()[name = tensor("attn_output_33_transpose_y_0"), val = tensor(false)]; + tensor values_101_cast_fp16 = transpose(perm = var_6665, x = values_99_cast_fp16)[name = tensor("transpose_151")]; + tensor attn_output_33_cast_fp16 = matmul(transpose_x = attn_output_33_transpose_x_0, transpose_y = attn_output_33_transpose_y_0, x = attn_weights_101_cast_fp16, y = values_101_cast_fp16)[name = tensor("attn_output_33_cast_fp16")]; + tensor var_6733 = const()[name = tensor("op_6733"), val = tensor([0, 2, 1, 3])]; + tensor var_6738 = const()[name = tensor("op_6738"), val = tensor([1, 1, 1024])]; + tensor var_6734_cast_fp16 = transpose(perm = var_6733, x = attn_output_33_cast_fp16)[name = tensor("transpose_148")]; + tensor input_199_cast_fp16 = reshape(shape = var_6738, x = var_6734_cast_fp16)[name = tensor("input_199_cast_fp16")]; + tensor attn16_out_proj_weight_to_fp16 = const()[name = tensor("attn16_out_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(410222336)))]; + tensor linear_66_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = attn16_out_proj_weight_to_fp16, x = input_199_cast_fp16)[name = tensor("linear_66_cast_fp16")]; + tensor var_6744_to_fp16 = const()[name = tensor("op_6744_to_fp16"), val = tensor(0x1p+0)]; + tensor var_6745_cast_fp16 = add(x = position16_to_fp16, y = var_6744_to_fp16)[name = tensor("op_6745_cast_fp16")]; + tensor var_6745_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_6745_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor input_201_cast_fp16 = add(x = input_195_cast_fp16, y = linear_66_cast_fp16)[name = tensor("input_201_cast_fp16")]; + tensor input_203_axes_0 = const()[name = tensor("input_203_axes_0"), val = tensor([-1])]; + tensor norm16_2_weight_to_fp16 = const()[name = tensor("norm16_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(412319552)))]; + tensor norm16_2_bias_to_fp16 = const()[name = tensor("norm16_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(412321664)))]; + tensor var_6749_to_fp16 = const()[name = tensor("op_6749_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_203_cast_fp16 = layer_norm(axes = input_203_axes_0, beta = norm16_2_bias_to_fp16, epsilon = var_6749_to_fp16, gamma = norm16_2_weight_to_fp16, x = input_201_cast_fp16)[name = tensor("input_203_cast_fp16")]; + tensor linear16_1_weight_to_fp16 = const()[name = tensor("linear16_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(412323776)))]; + tensor linear_67_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = linear16_1_weight_to_fp16, x = input_203_cast_fp16)[name = tensor("linear_67_cast_fp16")]; + tensor input_205_mode_0 = const()[name = tensor("input_205_mode_0"), val = tensor("EXACT")]; + tensor input_205_cast_fp16 = gelu(mode = input_205_mode_0, x = linear_67_cast_fp16)[name = tensor("input_205_cast_fp16")]; + tensor linear16_2_weight_to_fp16 = const()[name = tensor("linear16_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(420712448)))]; + tensor linear_68_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = linear16_2_weight_to_fp16, x = input_205_cast_fp16)[name = tensor("linear_68_cast_fp16")]; + tensor input_207_cast_fp16 = add(x = input_201_cast_fp16, y = linear_68_cast_fp16)[name = tensor("input_207_cast_fp16")]; + tensor input_209_axes_0 = const()[name = tensor("input_209_axes_0"), val = tensor([-1])]; + tensor norm17_1_weight_to_fp16 = const()[name = tensor("norm17_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(429101120)))]; + tensor norm17_1_bias_to_fp16 = const()[name = tensor("norm17_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(429103232)))]; + tensor var_6766_to_fp16 = const()[name = tensor("op_6766_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_209_cast_fp16 = layer_norm(axes = input_209_axes_0, beta = norm17_1_bias_to_fp16, epsilon = var_6766_to_fp16, gamma = norm17_1_weight_to_fp16, x = input_207_cast_fp16)[name = tensor("input_209_cast_fp16")]; + tensor attn17_in_proj_weight_to_fp16 = const()[name = tensor("attn17_in_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(429105344)))]; + tensor linear_69_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = attn17_in_proj_weight_to_fp16, x = input_209_cast_fp16)[name = tensor("linear_69_cast_fp16")]; + tensor var_6780 = const()[name = tensor("op_6780"), val = tensor([1, 1, 3, 16, 64])]; + tensor qkv_35_cast_fp16 = reshape(shape = var_6780, x = linear_69_cast_fp16)[name = tensor("qkv_35_cast_fp16")]; + tensor q_103_begin_0 = const()[name = tensor("q_103_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor q_103_end_0 = const()[name = tensor("q_103_end_0"), val = tensor([1, 1, 1, 16, 64])]; + tensor q_103_end_mask_0 = const()[name = tensor("q_103_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor q_103_squeeze_mask_0 = const()[name = tensor("q_103_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor q_103_cast_fp16 = slice_by_index(begin = q_103_begin_0, end = q_103_end_0, end_mask = q_103_end_mask_0, squeeze_mask = q_103_squeeze_mask_0, x = qkv_35_cast_fp16)[name = tensor("q_103_cast_fp16")]; + tensor k_69_begin_0 = const()[name = tensor("k_69_begin_0"), val = tensor([0, 0, 1, 0, 0])]; + tensor k_69_end_0 = const()[name = tensor("k_69_end_0"), val = tensor([1, 1, 2, 16, 64])]; + tensor k_69_end_mask_0 = const()[name = tensor("k_69_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor k_69_squeeze_mask_0 = const()[name = tensor("k_69_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor k_69_cast_fp16 = slice_by_index(begin = k_69_begin_0, end = k_69_end_0, end_mask = k_69_end_mask_0, squeeze_mask = k_69_squeeze_mask_0, x = qkv_35_cast_fp16)[name = tensor("k_69_cast_fp16")]; + tensor v_35_begin_0 = const()[name = tensor("v_35_begin_0"), val = tensor([0, 0, 2, 0, 0])]; + tensor v_35_end_0 = const()[name = tensor("v_35_end_0"), val = tensor([1, 1, 3, 16, 64])]; + tensor v_35_end_mask_0 = const()[name = tensor("v_35_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor v_35_squeeze_mask_0 = const()[name = tensor("v_35_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor v_35_cast_fp16 = slice_by_index(begin = v_35_begin_0, end = v_35_end_0, end_mask = v_35_end_mask_0, squeeze_mask = v_35_squeeze_mask_0, x = qkv_35_cast_fp16)[name = tensor("v_35_cast_fp16")]; + tensor var_6846 = const()[name = tensor("op_6846"), val = tensor([1, 1, 1, 1])]; + tensor position17_to_fp16_dtype_0 = const()[name = tensor("position17_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor position17_to_fp16 = cast(dtype = position17_to_fp16_dtype_0, x = position17)[name = tensor("cast_468")]; + tensor ts_107_cast_fp16 = reshape(shape = var_6846, x = position17_to_fp16)[name = tensor("ts_107_cast_fp16")]; + tensor var_6853 = const()[name = tensor("op_6853"), val = tensor([1, 1, 16, 32, 2])]; + tensor q_complex_35_cast_fp16 = reshape(shape = var_6853, x = q_103_cast_fp16)[name = tensor("q_complex_35_cast_fp16")]; + tensor var_6860 = const()[name = tensor("op_6860"), val = tensor([1, 1, 16, 32, 2])]; + tensor k_complex_35_cast_fp16 = reshape(shape = var_6860, x = k_69_cast_fp16)[name = tensor("k_complex_35_cast_fp16")]; + tensor var_6864_begin_0 = const()[name = tensor("op_6864_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_6864_end_0 = const()[name = tensor("op_6864_end_0"), val = tensor([1, 1, 16, 32, 1])]; + tensor var_6864_end_mask_0 = const()[name = tensor("op_6864_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_6864_squeeze_mask_0 = const()[name = tensor("op_6864_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_6864_cast_fp16 = slice_by_index(begin = var_6864_begin_0, end = var_6864_end_0, end_mask = var_6864_end_mask_0, squeeze_mask = var_6864_squeeze_mask_0, x = q_complex_35_cast_fp16)[name = tensor("op_6864_cast_fp16")]; + tensor var_6872_begin_0 = const()[name = tensor("op_6872_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_6872_end_0 = const()[name = tensor("op_6872_end_0"), val = tensor([1, 1, 16, 32, 2])]; + tensor var_6872_end_mask_0 = const()[name = tensor("op_6872_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_6872_squeeze_mask_0 = const()[name = tensor("op_6872_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_6872_cast_fp16 = slice_by_index(begin = var_6872_begin_0, end = var_6872_end_0, end_mask = var_6872_end_mask_0, squeeze_mask = var_6872_squeeze_mask_0, x = q_complex_35_cast_fp16)[name = tensor("op_6872_cast_fp16")]; + tensor var_6880_begin_0 = const()[name = tensor("op_6880_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_6880_end_0 = const()[name = tensor("op_6880_end_0"), val = tensor([1, 1, 16, 32, 1])]; + tensor var_6880_end_mask_0 = const()[name = tensor("op_6880_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_6880_squeeze_mask_0 = const()[name = tensor("op_6880_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_6880_cast_fp16 = slice_by_index(begin = var_6880_begin_0, end = var_6880_end_0, end_mask = var_6880_end_mask_0, squeeze_mask = var_6880_squeeze_mask_0, x = k_complex_35_cast_fp16)[name = tensor("op_6880_cast_fp16")]; + tensor var_6888_begin_0 = const()[name = tensor("op_6888_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_6888_end_0 = const()[name = tensor("op_6888_end_0"), val = tensor([1, 1, 16, 32, 2])]; + tensor var_6888_end_mask_0 = const()[name = tensor("op_6888_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_6888_squeeze_mask_0 = const()[name = tensor("op_6888_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_6888_cast_fp16 = slice_by_index(begin = var_6888_begin_0, end = var_6888_end_0, end_mask = var_6888_end_mask_0, squeeze_mask = var_6888_squeeze_mask_0, x = k_complex_35_cast_fp16)[name = tensor("op_6888_cast_fp16")]; + tensor freqs_35_to_fp16 = const()[name = tensor("freqs_35_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(435396864)))]; + tensor var_6894_cast_fp16 = mul(x = freqs_35_to_fp16, y = ts_107_cast_fp16)[name = tensor("op_6894_cast_fp16")]; + tensor rotr_35_cast_fp16 = cos(x = var_6894_cast_fp16)[name = tensor("rotr_35_cast_fp16")]; + tensor roti_35_cast_fp16 = sin(x = var_6894_cast_fp16)[name = tensor("roti_35_cast_fp16")]; + tensor var_6898_cast_fp16 = mul(x = var_6864_cast_fp16, y = rotr_35_cast_fp16)[name = tensor("op_6898_cast_fp16")]; + tensor var_6899_cast_fp16 = mul(x = var_6872_cast_fp16, y = roti_35_cast_fp16)[name = tensor("op_6899_cast_fp16")]; + tensor qor_69_cast_fp16 = sub(x = var_6898_cast_fp16, y = var_6899_cast_fp16)[name = tensor("qor_69_cast_fp16")]; + tensor var_6902_cast_fp16 = mul(x = var_6864_cast_fp16, y = roti_35_cast_fp16)[name = tensor("op_6902_cast_fp16")]; + tensor var_6903_cast_fp16 = mul(x = var_6872_cast_fp16, y = rotr_35_cast_fp16)[name = tensor("op_6903_cast_fp16")]; + tensor qoi_69_cast_fp16 = add(x = var_6902_cast_fp16, y = var_6903_cast_fp16)[name = tensor("qoi_69_cast_fp16")]; + tensor var_6906_cast_fp16 = mul(x = var_6880_cast_fp16, y = rotr_35_cast_fp16)[name = tensor("op_6906_cast_fp16")]; + tensor var_6907_cast_fp16 = mul(x = var_6888_cast_fp16, y = roti_35_cast_fp16)[name = tensor("op_6907_cast_fp16")]; + tensor kor_69_cast_fp16 = sub(x = var_6906_cast_fp16, y = var_6907_cast_fp16)[name = tensor("kor_69_cast_fp16")]; + tensor var_6910_cast_fp16 = mul(x = var_6880_cast_fp16, y = roti_35_cast_fp16)[name = tensor("op_6910_cast_fp16")]; + tensor var_6911_cast_fp16 = mul(x = var_6888_cast_fp16, y = rotr_35_cast_fp16)[name = tensor("op_6911_cast_fp16")]; + tensor koi_69_cast_fp16 = add(x = var_6910_cast_fp16, y = var_6911_cast_fp16)[name = tensor("koi_69_cast_fp16")]; + tensor qo_35_axis_0 = const()[name = tensor("qo_35_axis_0"), val = tensor(-1)]; + tensor qo_35_cast_fp16 = stack(axis = qo_35_axis_0, values = (qor_69_cast_fp16, qoi_69_cast_fp16))[name = tensor("qo_35_cast_fp16")]; + tensor ko_35_axis_0 = const()[name = tensor("ko_35_axis_0"), val = tensor(-1)]; + tensor ko_35_cast_fp16 = stack(axis = ko_35_axis_0, values = (kor_69_cast_fp16, koi_69_cast_fp16))[name = tensor("ko_35_cast_fp16")]; + tensor var_6944 = const()[name = tensor("op_6944"), val = tensor([1, 1, 16, 64])]; + tensor q_105_cast_fp16 = reshape(shape = var_6944, x = qo_35_cast_fp16)[name = tensor("q_105_cast_fp16")]; + tensor var_6950 = const()[name = tensor("op_6950"), val = tensor([1, 1, 16, 64])]; + tensor k_71_cast_fp16 = reshape(shape = var_6950, x = ko_35_cast_fp16)[name = tensor("k_71_cast_fp16")]; + tensor _inversed_6975_y_0_to_fp16 = const()[name = tensor("_inversed_6975_y_0_to_fp16"), val = tensor(0x1p-9)]; + tensor _inversed_6975_cast_fp16 = mul(x = ts_107_cast_fp16, y = _inversed_6975_y_0_to_fp16)[name = tensor("_inversed_6975_cast_fp16")]; + tensor var_6976_cast_fp16 = floor(x = _inversed_6975_cast_fp16)[name = tensor("op_6976_cast_fp16")]; + tensor var_6977_to_fp16 = const()[name = tensor("op_6977_to_fp16"), val = tensor(0x1p+9)]; + tensor var_6978_cast_fp16 = mul(x = var_6976_cast_fp16, y = var_6977_to_fp16)[name = tensor("op_6978_cast_fp16")]; + tensor write_indices_float_71_cast_fp16 = sub(x = ts_107_cast_fp16, y = var_6978_cast_fp16)[name = tensor("write_indices_float_71_cast_fp16")]; + tensor var_6985_dtype_0 = const()[name = tensor("op_6985_dtype_0"), val = tensor("int32")]; + tensor write_indices_35_reps_0 = const()[name = tensor("write_indices_35_reps_0"), val = tensor([1, 1, 16, 64])]; + tensor write_indices_float_71_cast_fp16_to_int32 = cast(dtype = var_6985_dtype_0, x = write_indices_float_71_cast_fp16)[name = tensor("cast_467")]; + tensor write_indices_35 = tile(reps = write_indices_35_reps_0, x = write_indices_float_71_cast_fp16_to_int32)[name = tensor("write_indices_35")]; + tensor var_6995_begin_0 = const()[name = tensor("op_6995_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_6995_end_0 = const()[name = tensor("op_6995_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor var_6995_end_mask_0 = const()[name = tensor("op_6995_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_6995_squeeze_mask_0 = const()[name = tensor("op_6995_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor cache17_to_fp16_dtype_0 = const()[name = tensor("cache17_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor cache17_to_fp16 = cast(dtype = cache17_to_fp16_dtype_0, x = cache17)[name = tensor("cast_466")]; + tensor var_6995_cast_fp16 = slice_by_index(begin = var_6995_begin_0, end = var_6995_end_0, end_mask = var_6995_end_mask_0, squeeze_mask = var_6995_squeeze_mask_0, x = cache17_to_fp16)[name = tensor("op_6995_cast_fp16")]; + tensor var_6997_axis_0 = const()[name = tensor("op_6997_axis_0"), val = tensor(1)]; + tensor var_6997_mode_0 = const()[name = tensor("op_6997_mode_0"), val = tensor("update")]; + tensor var_6997_validate_indices_0 = const()[name = tensor("op_6997_validate_indices_0"), val = tensor(false)]; + tensor var_6997_cast_fp16 = scatter_along_axis(axis = var_6997_axis_0, data = var_6995_cast_fp16, indices = write_indices_35, mode = var_6997_mode_0, updates = k_71_cast_fp16, validate_indices = var_6997_validate_indices_0)[name = tensor("op_6997_cast_fp16")]; + tensor concat_121 = const()[name = tensor("concat_121"), val = tensor([0, 0, 0, 0, 0])]; + tensor concat_122 = const()[name = tensor("concat_122"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_35_internal_tensor_assign_1_stride_0 = const()[name = tensor("new_cache_35_internal_tensor_assign_1_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_35_internal_tensor_assign_1_begin_mask_0 = const()[name = tensor("new_cache_35_internal_tensor_assign_1_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_35_internal_tensor_assign_1_end_mask_0 = const()[name = tensor("new_cache_35_internal_tensor_assign_1_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_35_internal_tensor_assign_1_squeeze_mask_0 = const()[name = tensor("new_cache_35_internal_tensor_assign_1_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_82 = const()[name = tensor("shape_82"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_34 = const()[name = tensor("reduce_prod_34"), val = tensor(1048576)]; + tensor range_1d_34_start_0 = const()[name = tensor("range_1d_34_start_0"), val = tensor(0)]; + tensor range_1d_34_step_0 = const()[name = tensor("range_1d_34_step_0"), val = tensor(1)]; + tensor range_1d_34 = range_1d(end = reduce_prod_34, start = range_1d_34_start_0, step = range_1d_34_step_0)[name = tensor("range_1d_34")]; + tensor reshape_170 = reshape(shape = shape_82, x = range_1d_34)[name = tensor("reshape_170")]; + tensor slice_by_index_34 = slice_by_index(begin = concat_121, begin_mask = new_cache_35_internal_tensor_assign_1_begin_mask_0, end = concat_122, end_mask = new_cache_35_internal_tensor_assign_1_end_mask_0, squeeze_mask = new_cache_35_internal_tensor_assign_1_squeeze_mask_0, stride = new_cache_35_internal_tensor_assign_1_stride_0, x = reshape_170)[name = tensor("slice_by_index_34")]; + tensor reshape_171_shape_0 = const()[name = tensor("reshape_171_shape_0"), val = tensor([-1])]; + tensor reshape_171 = reshape(shape = reshape_171_shape_0, x = slice_by_index_34)[name = tensor("reshape_171")]; + tensor reshape_172_shape_0 = const()[name = tensor("reshape_172_shape_0"), val = tensor([-1])]; + tensor reshape_172_cast_fp16 = reshape(shape = reshape_172_shape_0, x = var_6997_cast_fp16)[name = tensor("reshape_172_cast_fp16")]; + tensor reshape_173_shape_0 = const()[name = tensor("reshape_173_shape_0"), val = tensor([-1])]; + tensor reshape_173_cast_fp16 = reshape(shape = reshape_173_shape_0, x = cache17_to_fp16)[name = tensor("reshape_173_cast_fp16")]; + tensor scatter_34_mode_0 = const()[name = tensor("scatter_34_mode_0"), val = tensor("update")]; + tensor scatter_34_axis_0 = const()[name = tensor("scatter_34_axis_0"), val = tensor(0)]; + tensor scatter_34_validate_indices_0 = const()[name = tensor("scatter_34_validate_indices_0"), val = tensor(false)]; + tensor scatter_34_cast_fp16 = scatter(axis = scatter_34_axis_0, data = reshape_173_cast_fp16, indices = reshape_171, mode = scatter_34_mode_0, updates = reshape_172_cast_fp16, validate_indices = scatter_34_validate_indices_0)[name = tensor("scatter_34_cast_fp16")]; + tensor reshape_174_cast_fp16 = reshape(shape = shape_82, x = scatter_34_cast_fp16)[name = tensor("reshape_174_cast_fp16")]; + tensor var_7005_begin_0 = const()[name = tensor("op_7005_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor var_7005_end_0 = const()[name = tensor("op_7005_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor var_7005_end_mask_0 = const()[name = tensor("op_7005_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_7005_squeeze_mask_0 = const()[name = tensor("op_7005_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor var_7005_cast_fp16 = slice_by_index(begin = var_7005_begin_0, end = var_7005_end_0, end_mask = var_7005_end_mask_0, squeeze_mask = var_7005_squeeze_mask_0, x = reshape_174_cast_fp16)[name = tensor("op_7005_cast_fp16")]; + tensor var_7007_axis_0 = const()[name = tensor("op_7007_axis_0"), val = tensor(1)]; + tensor var_7007_mode_0 = const()[name = tensor("op_7007_mode_0"), val = tensor("update")]; + tensor var_7007_validate_indices_0 = const()[name = tensor("op_7007_validate_indices_0"), val = tensor(false)]; + tensor var_7007_cast_fp16 = scatter_along_axis(axis = var_7007_axis_0, data = var_7005_cast_fp16, indices = write_indices_35, mode = var_7007_mode_0, updates = v_35_cast_fp16, validate_indices = var_7007_validate_indices_0)[name = tensor("op_7007_cast_fp16")]; + tensor concat_123 = const()[name = tensor("concat_123"), val = tensor([1, 0, 0, 0, 0])]; + tensor concat_124 = const()[name = tensor("concat_124"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_35_internal_tensor_assign_2_stride_0 = const()[name = tensor("new_cache_35_internal_tensor_assign_2_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_35_internal_tensor_assign_2_begin_mask_0 = const()[name = tensor("new_cache_35_internal_tensor_assign_2_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_35_internal_tensor_assign_2_end_mask_0 = const()[name = tensor("new_cache_35_internal_tensor_assign_2_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_35_internal_tensor_assign_2_squeeze_mask_0 = const()[name = tensor("new_cache_35_internal_tensor_assign_2_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_83 = const()[name = tensor("shape_83"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_35 = const()[name = tensor("reduce_prod_35"), val = tensor(1048576)]; + tensor range_1d_35_start_0 = const()[name = tensor("range_1d_35_start_0"), val = tensor(0)]; + tensor range_1d_35_step_0 = const()[name = tensor("range_1d_35_step_0"), val = tensor(1)]; + tensor range_1d_35 = range_1d(end = reduce_prod_35, start = range_1d_35_start_0, step = range_1d_35_step_0)[name = tensor("range_1d_35")]; + tensor reshape_175 = reshape(shape = shape_83, x = range_1d_35)[name = tensor("reshape_175")]; + tensor slice_by_index_35 = slice_by_index(begin = concat_123, begin_mask = new_cache_35_internal_tensor_assign_2_begin_mask_0, end = concat_124, end_mask = new_cache_35_internal_tensor_assign_2_end_mask_0, squeeze_mask = new_cache_35_internal_tensor_assign_2_squeeze_mask_0, stride = new_cache_35_internal_tensor_assign_2_stride_0, x = reshape_175)[name = tensor("slice_by_index_35")]; + tensor reshape_176_shape_0 = const()[name = tensor("reshape_176_shape_0"), val = tensor([-1])]; + tensor reshape_176 = reshape(shape = reshape_176_shape_0, x = slice_by_index_35)[name = tensor("reshape_176")]; + tensor reshape_177_shape_0 = const()[name = tensor("reshape_177_shape_0"), val = tensor([-1])]; + tensor reshape_177_cast_fp16 = reshape(shape = reshape_177_shape_0, x = var_7007_cast_fp16)[name = tensor("reshape_177_cast_fp16")]; + tensor reshape_178_shape_0 = const()[name = tensor("reshape_178_shape_0"), val = tensor([-1])]; + tensor reshape_178_cast_fp16 = reshape(shape = reshape_178_shape_0, x = reshape_174_cast_fp16)[name = tensor("reshape_178_cast_fp16")]; + tensor scatter_35_mode_0 = const()[name = tensor("scatter_35_mode_0"), val = tensor("update")]; + tensor scatter_35_axis_0 = const()[name = tensor("scatter_35_axis_0"), val = tensor(0)]; + tensor scatter_35_validate_indices_0 = const()[name = tensor("scatter_35_validate_indices_0"), val = tensor(false)]; + tensor scatter_35_cast_fp16 = scatter(axis = scatter_35_axis_0, data = reshape_178_cast_fp16, indices = reshape_176, mode = scatter_35_mode_0, updates = reshape_177_cast_fp16, validate_indices = scatter_35_validate_indices_0)[name = tensor("scatter_35_cast_fp16")]; + tensor reshape_179_cast_fp16 = reshape(shape = shape_83, x = scatter_35_cast_fp16)[name = tensor("reshape_179_cast_fp16")]; + tensor reshape_179_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("reshape_179_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor keys_103_begin_0 = const()[name = tensor("keys_103_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor keys_103_end_0 = const()[name = tensor("keys_103_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor keys_103_end_mask_0 = const()[name = tensor("keys_103_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor keys_103_squeeze_mask_0 = const()[name = tensor("keys_103_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor keys_103_cast_fp16 = slice_by_index(begin = keys_103_begin_0, end = keys_103_end_0, end_mask = keys_103_end_mask_0, squeeze_mask = keys_103_squeeze_mask_0, x = reshape_179_cast_fp16)[name = tensor("keys_103_cast_fp16")]; + tensor values_103_begin_0 = const()[name = tensor("values_103_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor values_103_end_0 = const()[name = tensor("values_103_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor values_103_end_mask_0 = const()[name = tensor("values_103_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor values_103_squeeze_mask_0 = const()[name = tensor("values_103_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor values_103_cast_fp16 = slice_by_index(begin = values_103_begin_0, end = values_103_end_0, end_mask = values_103_end_mask_0, squeeze_mask = values_103_squeeze_mask_0, x = reshape_179_cast_fp16)[name = tensor("values_103_cast_fp16")]; + tensor var_7019_cast_fp16 = not_equal(x = keys_103_cast_fp16, y = keys_103_cast_fp16)[name = tensor("op_7019_cast_fp16")]; + tensor keys_105_cast_fp16 = select(a = var_463_to_fp16, b = keys_103_cast_fp16, cond = var_7019_cast_fp16)[name = tensor("keys_105_cast_fp16")]; + tensor var_7027_cast_fp16 = not_equal(x = values_103_cast_fp16, y = values_103_cast_fp16)[name = tensor("op_7027_cast_fp16")]; + tensor values_105_cast_fp16 = select(a = var_463_to_fp16, b = values_103_cast_fp16, cond = var_7027_cast_fp16)[name = tensor("values_105_cast_fp16")]; + tensor var_7051 = const()[name = tensor("op_7051"), val = tensor([0, 2, 1, 3])]; + tensor var_7067 = const()[name = tensor("op_7067"), val = tensor([1, 1, 1])]; + tensor var_7068_cast_fp16 = reshape(shape = var_7067, x = position17_to_fp16)[name = tensor("op_7068_cast_fp16")]; + tensor var_7088_to_fp16 = const()[name = tensor("op_7088_to_fp16"), val = tensor(0x1p+0)]; + tensor valid_len_35_cast_fp16 = add(x = var_7068_cast_fp16, y = var_7088_to_fp16)[name = tensor("valid_len_35_cast_fp16")]; + tensor valid_mask_35_cast_fp16 = less(x = k_positions_1_promoted_to_fp16, y = valid_len_35_cast_fp16)[name = tensor("valid_mask_35_cast_fp16")]; + tensor causal_mask_35_cast_fp16 = less_equal(x = k_positions_1_promoted_to_fp16, y = var_7068_cast_fp16)[name = tensor("causal_mask_35_cast_fp16")]; + tensor attn_mask_69 = logical_and(x = valid_mask_35_cast_fp16, y = causal_mask_35_cast_fp16)[name = tensor("attn_mask_69")]; + tensor attn_mask_71_axes_0 = const()[name = tensor("attn_mask_71_axes_0"), val = tensor([1])]; + tensor attn_mask_71 = expand_dims(axes = attn_mask_71_axes_0, x = attn_mask_69)[name = tensor("attn_mask_71")]; + tensor var_7106_transpose_x_0 = const()[name = tensor("op_7106_transpose_x_0"), val = tensor(false)]; + tensor var_7106_transpose_y_0 = const()[name = tensor("op_7106_transpose_y_0"), val = tensor(false)]; + tensor transpose_106_perm_0 = const()[name = tensor("transpose_106_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_107_perm_0 = const()[name = tensor("transpose_107_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_107 = transpose(perm = transpose_107_perm_0, x = keys_105_cast_fp16)[name = tensor("transpose_145")]; + tensor transpose_106 = transpose(perm = transpose_106_perm_0, x = q_105_cast_fp16)[name = tensor("transpose_146")]; + tensor var_7106_cast_fp16 = matmul(transpose_x = var_7106_transpose_x_0, transpose_y = var_7106_transpose_y_0, x = transpose_106, y = transpose_107)[name = tensor("op_7106_cast_fp16")]; + tensor var_7100_to_fp16 = const()[name = tensor("op_7100_to_fp16"), val = tensor([0x1p-3])]; + tensor attn_weights_103_cast_fp16 = mul(x = var_7106_cast_fp16, y = var_7100_to_fp16)[name = tensor("attn_weights_103_cast_fp16")]; + tensor var_7108 = logical_not(x = attn_mask_71)[name = tensor("op_7108")]; + tensor var_7109_to_fp16 = const()[name = tensor("op_7109_to_fp16"), val = tensor(-inf)]; + tensor attn_weights_105_cast_fp16 = select(a = var_7109_to_fp16, b = attn_weights_103_cast_fp16, cond = var_7108)[name = tensor("attn_weights_105_cast_fp16")]; + tensor var_7111 = const()[name = tensor("op_7111"), val = tensor(-1)]; + tensor attn_weights_107_cast_fp16 = softmax(axis = var_7111, x = attn_weights_105_cast_fp16)[name = tensor("attn_weights_107_cast_fp16")]; + tensor attn_output_35_transpose_x_0 = const()[name = tensor("attn_output_35_transpose_x_0"), val = tensor(false)]; + tensor attn_output_35_transpose_y_0 = const()[name = tensor("attn_output_35_transpose_y_0"), val = tensor(false)]; + tensor values_107_cast_fp16 = transpose(perm = var_7051, x = values_105_cast_fp16)[name = tensor("transpose_147")]; + tensor attn_output_35_cast_fp16 = matmul(transpose_x = attn_output_35_transpose_x_0, transpose_y = attn_output_35_transpose_y_0, x = attn_weights_107_cast_fp16, y = values_107_cast_fp16)[name = tensor("attn_output_35_cast_fp16")]; + tensor var_7119 = const()[name = tensor("op_7119"), val = tensor([0, 2, 1, 3])]; + tensor var_7124 = const()[name = tensor("op_7124"), val = tensor([1, 1, 1024])]; + tensor var_7120_cast_fp16 = transpose(perm = var_7119, x = attn_output_35_cast_fp16)[name = tensor("transpose_144")]; + tensor input_211_cast_fp16 = reshape(shape = var_7124, x = var_7120_cast_fp16)[name = tensor("input_211_cast_fp16")]; + tensor attn17_out_proj_weight_to_fp16 = const()[name = tensor("attn17_out_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(435396992)))]; + tensor linear_70_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = attn17_out_proj_weight_to_fp16, x = input_211_cast_fp16)[name = tensor("linear_70_cast_fp16")]; + tensor var_7130_to_fp16 = const()[name = tensor("op_7130_to_fp16"), val = tensor(0x1p+0)]; + tensor var_7131_cast_fp16 = add(x = position17_to_fp16, y = var_7130_to_fp16)[name = tensor("op_7131_cast_fp16")]; + tensor var_7131_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_7131_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor input_213_cast_fp16 = add(x = input_207_cast_fp16, y = linear_70_cast_fp16)[name = tensor("input_213_cast_fp16")]; + tensor input_215_axes_0 = const()[name = tensor("input_215_axes_0"), val = tensor([-1])]; + tensor norm17_2_weight_to_fp16 = const()[name = tensor("norm17_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(437494208)))]; + tensor norm17_2_bias_to_fp16 = const()[name = tensor("norm17_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(437496320)))]; + tensor var_7135_to_fp16 = const()[name = tensor("op_7135_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_215_cast_fp16 = layer_norm(axes = input_215_axes_0, beta = norm17_2_bias_to_fp16, epsilon = var_7135_to_fp16, gamma = norm17_2_weight_to_fp16, x = input_213_cast_fp16)[name = tensor("input_215_cast_fp16")]; + tensor linear17_1_weight_to_fp16 = const()[name = tensor("linear17_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(437498432)))]; + tensor linear_71_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = linear17_1_weight_to_fp16, x = input_215_cast_fp16)[name = tensor("linear_71_cast_fp16")]; + tensor input_217_mode_0 = const()[name = tensor("input_217_mode_0"), val = tensor("EXACT")]; + tensor input_217_cast_fp16 = gelu(mode = input_217_mode_0, x = linear_71_cast_fp16)[name = tensor("input_217_cast_fp16")]; + tensor linear17_2_weight_to_fp16 = const()[name = tensor("linear17_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(445887104)))]; + tensor linear_72_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = linear17_2_weight_to_fp16, x = input_217_cast_fp16)[name = tensor("linear_72_cast_fp16")]; + tensor input_219_cast_fp16 = add(x = input_213_cast_fp16, y = linear_72_cast_fp16)[name = tensor("input_219_cast_fp16")]; + tensor input_221_axes_0 = const()[name = tensor("input_221_axes_0"), val = tensor([-1])]; + tensor norm18_1_weight_to_fp16 = const()[name = tensor("norm18_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(454275776)))]; + tensor norm18_1_bias_to_fp16 = const()[name = tensor("norm18_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(454277888)))]; + tensor var_7152_to_fp16 = const()[name = tensor("op_7152_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_221_cast_fp16 = layer_norm(axes = input_221_axes_0, beta = norm18_1_bias_to_fp16, epsilon = var_7152_to_fp16, gamma = norm18_1_weight_to_fp16, x = input_219_cast_fp16)[name = tensor("input_221_cast_fp16")]; + tensor attn18_in_proj_weight_to_fp16 = const()[name = tensor("attn18_in_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(454280000)))]; + tensor linear_73_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = attn18_in_proj_weight_to_fp16, x = input_221_cast_fp16)[name = tensor("linear_73_cast_fp16")]; + tensor var_7166 = const()[name = tensor("op_7166"), val = tensor([1, 1, 3, 16, 64])]; + tensor qkv_37_cast_fp16 = reshape(shape = var_7166, x = linear_73_cast_fp16)[name = tensor("qkv_37_cast_fp16")]; + tensor q_109_begin_0 = const()[name = tensor("q_109_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor q_109_end_0 = const()[name = tensor("q_109_end_0"), val = tensor([1, 1, 1, 16, 64])]; + tensor q_109_end_mask_0 = const()[name = tensor("q_109_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor q_109_squeeze_mask_0 = const()[name = tensor("q_109_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor q_109_cast_fp16 = slice_by_index(begin = q_109_begin_0, end = q_109_end_0, end_mask = q_109_end_mask_0, squeeze_mask = q_109_squeeze_mask_0, x = qkv_37_cast_fp16)[name = tensor("q_109_cast_fp16")]; + tensor k_73_begin_0 = const()[name = tensor("k_73_begin_0"), val = tensor([0, 0, 1, 0, 0])]; + tensor k_73_end_0 = const()[name = tensor("k_73_end_0"), val = tensor([1, 1, 2, 16, 64])]; + tensor k_73_end_mask_0 = const()[name = tensor("k_73_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor k_73_squeeze_mask_0 = const()[name = tensor("k_73_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor k_73_cast_fp16 = slice_by_index(begin = k_73_begin_0, end = k_73_end_0, end_mask = k_73_end_mask_0, squeeze_mask = k_73_squeeze_mask_0, x = qkv_37_cast_fp16)[name = tensor("k_73_cast_fp16")]; + tensor v_37_begin_0 = const()[name = tensor("v_37_begin_0"), val = tensor([0, 0, 2, 0, 0])]; + tensor v_37_end_0 = const()[name = tensor("v_37_end_0"), val = tensor([1, 1, 3, 16, 64])]; + tensor v_37_end_mask_0 = const()[name = tensor("v_37_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor v_37_squeeze_mask_0 = const()[name = tensor("v_37_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor v_37_cast_fp16 = slice_by_index(begin = v_37_begin_0, end = v_37_end_0, end_mask = v_37_end_mask_0, squeeze_mask = v_37_squeeze_mask_0, x = qkv_37_cast_fp16)[name = tensor("v_37_cast_fp16")]; + tensor var_7232 = const()[name = tensor("op_7232"), val = tensor([1, 1, 1, 1])]; + tensor position18_to_fp16_dtype_0 = const()[name = tensor("position18_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor position18_to_fp16 = cast(dtype = position18_to_fp16_dtype_0, x = position18)[name = tensor("cast_463")]; + tensor ts_113_cast_fp16 = reshape(shape = var_7232, x = position18_to_fp16)[name = tensor("ts_113_cast_fp16")]; + tensor var_7239 = const()[name = tensor("op_7239"), val = tensor([1, 1, 16, 32, 2])]; + tensor q_complex_37_cast_fp16 = reshape(shape = var_7239, x = q_109_cast_fp16)[name = tensor("q_complex_37_cast_fp16")]; + tensor var_7246 = const()[name = tensor("op_7246"), val = tensor([1, 1, 16, 32, 2])]; + tensor k_complex_37_cast_fp16 = reshape(shape = var_7246, x = k_73_cast_fp16)[name = tensor("k_complex_37_cast_fp16")]; + tensor var_7250_begin_0 = const()[name = tensor("op_7250_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_7250_end_0 = const()[name = tensor("op_7250_end_0"), val = tensor([1, 1, 16, 32, 1])]; + tensor var_7250_end_mask_0 = const()[name = tensor("op_7250_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_7250_squeeze_mask_0 = const()[name = tensor("op_7250_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_7250_cast_fp16 = slice_by_index(begin = var_7250_begin_0, end = var_7250_end_0, end_mask = var_7250_end_mask_0, squeeze_mask = var_7250_squeeze_mask_0, x = q_complex_37_cast_fp16)[name = tensor("op_7250_cast_fp16")]; + tensor var_7258_begin_0 = const()[name = tensor("op_7258_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_7258_end_0 = const()[name = tensor("op_7258_end_0"), val = tensor([1, 1, 16, 32, 2])]; + tensor var_7258_end_mask_0 = const()[name = tensor("op_7258_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_7258_squeeze_mask_0 = const()[name = tensor("op_7258_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_7258_cast_fp16 = slice_by_index(begin = var_7258_begin_0, end = var_7258_end_0, end_mask = var_7258_end_mask_0, squeeze_mask = var_7258_squeeze_mask_0, x = q_complex_37_cast_fp16)[name = tensor("op_7258_cast_fp16")]; + tensor var_7266_begin_0 = const()[name = tensor("op_7266_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_7266_end_0 = const()[name = tensor("op_7266_end_0"), val = tensor([1, 1, 16, 32, 1])]; + tensor var_7266_end_mask_0 = const()[name = tensor("op_7266_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_7266_squeeze_mask_0 = const()[name = tensor("op_7266_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_7266_cast_fp16 = slice_by_index(begin = var_7266_begin_0, end = var_7266_end_0, end_mask = var_7266_end_mask_0, squeeze_mask = var_7266_squeeze_mask_0, x = k_complex_37_cast_fp16)[name = tensor("op_7266_cast_fp16")]; + tensor var_7274_begin_0 = const()[name = tensor("op_7274_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_7274_end_0 = const()[name = tensor("op_7274_end_0"), val = tensor([1, 1, 16, 32, 2])]; + tensor var_7274_end_mask_0 = const()[name = tensor("op_7274_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_7274_squeeze_mask_0 = const()[name = tensor("op_7274_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_7274_cast_fp16 = slice_by_index(begin = var_7274_begin_0, end = var_7274_end_0, end_mask = var_7274_end_mask_0, squeeze_mask = var_7274_squeeze_mask_0, x = k_complex_37_cast_fp16)[name = tensor("op_7274_cast_fp16")]; + tensor freqs_37_to_fp16 = const()[name = tensor("freqs_37_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(460571520)))]; + tensor var_7280_cast_fp16 = mul(x = freqs_37_to_fp16, y = ts_113_cast_fp16)[name = tensor("op_7280_cast_fp16")]; + tensor rotr_37_cast_fp16 = cos(x = var_7280_cast_fp16)[name = tensor("rotr_37_cast_fp16")]; + tensor roti_37_cast_fp16 = sin(x = var_7280_cast_fp16)[name = tensor("roti_37_cast_fp16")]; + tensor var_7284_cast_fp16 = mul(x = var_7250_cast_fp16, y = rotr_37_cast_fp16)[name = tensor("op_7284_cast_fp16")]; + tensor var_7285_cast_fp16 = mul(x = var_7258_cast_fp16, y = roti_37_cast_fp16)[name = tensor("op_7285_cast_fp16")]; + tensor qor_73_cast_fp16 = sub(x = var_7284_cast_fp16, y = var_7285_cast_fp16)[name = tensor("qor_73_cast_fp16")]; + tensor var_7288_cast_fp16 = mul(x = var_7250_cast_fp16, y = roti_37_cast_fp16)[name = tensor("op_7288_cast_fp16")]; + tensor var_7289_cast_fp16 = mul(x = var_7258_cast_fp16, y = rotr_37_cast_fp16)[name = tensor("op_7289_cast_fp16")]; + tensor qoi_73_cast_fp16 = add(x = var_7288_cast_fp16, y = var_7289_cast_fp16)[name = tensor("qoi_73_cast_fp16")]; + tensor var_7292_cast_fp16 = mul(x = var_7266_cast_fp16, y = rotr_37_cast_fp16)[name = tensor("op_7292_cast_fp16")]; + tensor var_7293_cast_fp16 = mul(x = var_7274_cast_fp16, y = roti_37_cast_fp16)[name = tensor("op_7293_cast_fp16")]; + tensor kor_73_cast_fp16 = sub(x = var_7292_cast_fp16, y = var_7293_cast_fp16)[name = tensor("kor_73_cast_fp16")]; + tensor var_7296_cast_fp16 = mul(x = var_7266_cast_fp16, y = roti_37_cast_fp16)[name = tensor("op_7296_cast_fp16")]; + tensor var_7297_cast_fp16 = mul(x = var_7274_cast_fp16, y = rotr_37_cast_fp16)[name = tensor("op_7297_cast_fp16")]; + tensor koi_73_cast_fp16 = add(x = var_7296_cast_fp16, y = var_7297_cast_fp16)[name = tensor("koi_73_cast_fp16")]; + tensor qo_37_axis_0 = const()[name = tensor("qo_37_axis_0"), val = tensor(-1)]; + tensor qo_37_cast_fp16 = stack(axis = qo_37_axis_0, values = (qor_73_cast_fp16, qoi_73_cast_fp16))[name = tensor("qo_37_cast_fp16")]; + tensor ko_37_axis_0 = const()[name = tensor("ko_37_axis_0"), val = tensor(-1)]; + tensor ko_37_cast_fp16 = stack(axis = ko_37_axis_0, values = (kor_73_cast_fp16, koi_73_cast_fp16))[name = tensor("ko_37_cast_fp16")]; + tensor var_7330 = const()[name = tensor("op_7330"), val = tensor([1, 1, 16, 64])]; + tensor q_111_cast_fp16 = reshape(shape = var_7330, x = qo_37_cast_fp16)[name = tensor("q_111_cast_fp16")]; + tensor var_7336 = const()[name = tensor("op_7336"), val = tensor([1, 1, 16, 64])]; + tensor k_75_cast_fp16 = reshape(shape = var_7336, x = ko_37_cast_fp16)[name = tensor("k_75_cast_fp16")]; + tensor _inversed_7361_y_0_to_fp16 = const()[name = tensor("_inversed_7361_y_0_to_fp16"), val = tensor(0x1p-9)]; + tensor _inversed_7361_cast_fp16 = mul(x = ts_113_cast_fp16, y = _inversed_7361_y_0_to_fp16)[name = tensor("_inversed_7361_cast_fp16")]; + tensor var_7362_cast_fp16 = floor(x = _inversed_7361_cast_fp16)[name = tensor("op_7362_cast_fp16")]; + tensor var_7363_to_fp16 = const()[name = tensor("op_7363_to_fp16"), val = tensor(0x1p+9)]; + tensor var_7364_cast_fp16 = mul(x = var_7362_cast_fp16, y = var_7363_to_fp16)[name = tensor("op_7364_cast_fp16")]; + tensor write_indices_float_75_cast_fp16 = sub(x = ts_113_cast_fp16, y = var_7364_cast_fp16)[name = tensor("write_indices_float_75_cast_fp16")]; + tensor var_7371_dtype_0 = const()[name = tensor("op_7371_dtype_0"), val = tensor("int32")]; + tensor write_indices_37_reps_0 = const()[name = tensor("write_indices_37_reps_0"), val = tensor([1, 1, 16, 64])]; + tensor write_indices_float_75_cast_fp16_to_int32 = cast(dtype = var_7371_dtype_0, x = write_indices_float_75_cast_fp16)[name = tensor("cast_462")]; + tensor write_indices_37 = tile(reps = write_indices_37_reps_0, x = write_indices_float_75_cast_fp16_to_int32)[name = tensor("write_indices_37")]; + tensor var_7381_begin_0 = const()[name = tensor("op_7381_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_7381_end_0 = const()[name = tensor("op_7381_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor var_7381_end_mask_0 = const()[name = tensor("op_7381_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_7381_squeeze_mask_0 = const()[name = tensor("op_7381_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor cache18_to_fp16_dtype_0 = const()[name = tensor("cache18_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor cache18_to_fp16 = cast(dtype = cache18_to_fp16_dtype_0, x = cache18)[name = tensor("cast_461")]; + tensor var_7381_cast_fp16 = slice_by_index(begin = var_7381_begin_0, end = var_7381_end_0, end_mask = var_7381_end_mask_0, squeeze_mask = var_7381_squeeze_mask_0, x = cache18_to_fp16)[name = tensor("op_7381_cast_fp16")]; + tensor var_7383_axis_0 = const()[name = tensor("op_7383_axis_0"), val = tensor(1)]; + tensor var_7383_mode_0 = const()[name = tensor("op_7383_mode_0"), val = tensor("update")]; + tensor var_7383_validate_indices_0 = const()[name = tensor("op_7383_validate_indices_0"), val = tensor(false)]; + tensor var_7383_cast_fp16 = scatter_along_axis(axis = var_7383_axis_0, data = var_7381_cast_fp16, indices = write_indices_37, mode = var_7383_mode_0, updates = k_75_cast_fp16, validate_indices = var_7383_validate_indices_0)[name = tensor("op_7383_cast_fp16")]; + tensor concat_128 = const()[name = tensor("concat_128"), val = tensor([0, 0, 0, 0, 0])]; + tensor concat_129 = const()[name = tensor("concat_129"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_37_internal_tensor_assign_1_stride_0 = const()[name = tensor("new_cache_37_internal_tensor_assign_1_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_37_internal_tensor_assign_1_begin_mask_0 = const()[name = tensor("new_cache_37_internal_tensor_assign_1_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_37_internal_tensor_assign_1_end_mask_0 = const()[name = tensor("new_cache_37_internal_tensor_assign_1_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_37_internal_tensor_assign_1_squeeze_mask_0 = const()[name = tensor("new_cache_37_internal_tensor_assign_1_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_84 = const()[name = tensor("shape_84"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_36 = const()[name = tensor("reduce_prod_36"), val = tensor(1048576)]; + tensor range_1d_36_start_0 = const()[name = tensor("range_1d_36_start_0"), val = tensor(0)]; + tensor range_1d_36_step_0 = const()[name = tensor("range_1d_36_step_0"), val = tensor(1)]; + tensor range_1d_36 = range_1d(end = reduce_prod_36, start = range_1d_36_start_0, step = range_1d_36_step_0)[name = tensor("range_1d_36")]; + tensor reshape_180 = reshape(shape = shape_84, x = range_1d_36)[name = tensor("reshape_180")]; + tensor slice_by_index_36 = slice_by_index(begin = concat_128, begin_mask = new_cache_37_internal_tensor_assign_1_begin_mask_0, end = concat_129, end_mask = new_cache_37_internal_tensor_assign_1_end_mask_0, squeeze_mask = new_cache_37_internal_tensor_assign_1_squeeze_mask_0, stride = new_cache_37_internal_tensor_assign_1_stride_0, x = reshape_180)[name = tensor("slice_by_index_36")]; + tensor reshape_181_shape_0 = const()[name = tensor("reshape_181_shape_0"), val = tensor([-1])]; + tensor reshape_181 = reshape(shape = reshape_181_shape_0, x = slice_by_index_36)[name = tensor("reshape_181")]; + tensor reshape_182_shape_0 = const()[name = tensor("reshape_182_shape_0"), val = tensor([-1])]; + tensor reshape_182_cast_fp16 = reshape(shape = reshape_182_shape_0, x = var_7383_cast_fp16)[name = tensor("reshape_182_cast_fp16")]; + tensor reshape_183_shape_0 = const()[name = tensor("reshape_183_shape_0"), val = tensor([-1])]; + tensor reshape_183_cast_fp16 = reshape(shape = reshape_183_shape_0, x = cache18_to_fp16)[name = tensor("reshape_183_cast_fp16")]; + tensor scatter_36_mode_0 = const()[name = tensor("scatter_36_mode_0"), val = tensor("update")]; + tensor scatter_36_axis_0 = const()[name = tensor("scatter_36_axis_0"), val = tensor(0)]; + tensor scatter_36_validate_indices_0 = const()[name = tensor("scatter_36_validate_indices_0"), val = tensor(false)]; + tensor scatter_36_cast_fp16 = scatter(axis = scatter_36_axis_0, data = reshape_183_cast_fp16, indices = reshape_181, mode = scatter_36_mode_0, updates = reshape_182_cast_fp16, validate_indices = scatter_36_validate_indices_0)[name = tensor("scatter_36_cast_fp16")]; + tensor reshape_184_cast_fp16 = reshape(shape = shape_84, x = scatter_36_cast_fp16)[name = tensor("reshape_184_cast_fp16")]; + tensor var_7391_begin_0 = const()[name = tensor("op_7391_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor var_7391_end_0 = const()[name = tensor("op_7391_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor var_7391_end_mask_0 = const()[name = tensor("op_7391_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_7391_squeeze_mask_0 = const()[name = tensor("op_7391_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor var_7391_cast_fp16 = slice_by_index(begin = var_7391_begin_0, end = var_7391_end_0, end_mask = var_7391_end_mask_0, squeeze_mask = var_7391_squeeze_mask_0, x = reshape_184_cast_fp16)[name = tensor("op_7391_cast_fp16")]; + tensor var_7393_axis_0 = const()[name = tensor("op_7393_axis_0"), val = tensor(1)]; + tensor var_7393_mode_0 = const()[name = tensor("op_7393_mode_0"), val = tensor("update")]; + tensor var_7393_validate_indices_0 = const()[name = tensor("op_7393_validate_indices_0"), val = tensor(false)]; + tensor var_7393_cast_fp16 = scatter_along_axis(axis = var_7393_axis_0, data = var_7391_cast_fp16, indices = write_indices_37, mode = var_7393_mode_0, updates = v_37_cast_fp16, validate_indices = var_7393_validate_indices_0)[name = tensor("op_7393_cast_fp16")]; + tensor concat_130 = const()[name = tensor("concat_130"), val = tensor([1, 0, 0, 0, 0])]; + tensor concat_131 = const()[name = tensor("concat_131"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_37_internal_tensor_assign_2_stride_0 = const()[name = tensor("new_cache_37_internal_tensor_assign_2_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_37_internal_tensor_assign_2_begin_mask_0 = const()[name = tensor("new_cache_37_internal_tensor_assign_2_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_37_internal_tensor_assign_2_end_mask_0 = const()[name = tensor("new_cache_37_internal_tensor_assign_2_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_37_internal_tensor_assign_2_squeeze_mask_0 = const()[name = tensor("new_cache_37_internal_tensor_assign_2_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_85 = const()[name = tensor("shape_85"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_37 = const()[name = tensor("reduce_prod_37"), val = tensor(1048576)]; + tensor range_1d_37_start_0 = const()[name = tensor("range_1d_37_start_0"), val = tensor(0)]; + tensor range_1d_37_step_0 = const()[name = tensor("range_1d_37_step_0"), val = tensor(1)]; + tensor range_1d_37 = range_1d(end = reduce_prod_37, start = range_1d_37_start_0, step = range_1d_37_step_0)[name = tensor("range_1d_37")]; + tensor reshape_185 = reshape(shape = shape_85, x = range_1d_37)[name = tensor("reshape_185")]; + tensor slice_by_index_37 = slice_by_index(begin = concat_130, begin_mask = new_cache_37_internal_tensor_assign_2_begin_mask_0, end = concat_131, end_mask = new_cache_37_internal_tensor_assign_2_end_mask_0, squeeze_mask = new_cache_37_internal_tensor_assign_2_squeeze_mask_0, stride = new_cache_37_internal_tensor_assign_2_stride_0, x = reshape_185)[name = tensor("slice_by_index_37")]; + tensor reshape_186_shape_0 = const()[name = tensor("reshape_186_shape_0"), val = tensor([-1])]; + tensor reshape_186 = reshape(shape = reshape_186_shape_0, x = slice_by_index_37)[name = tensor("reshape_186")]; + tensor reshape_187_shape_0 = const()[name = tensor("reshape_187_shape_0"), val = tensor([-1])]; + tensor reshape_187_cast_fp16 = reshape(shape = reshape_187_shape_0, x = var_7393_cast_fp16)[name = tensor("reshape_187_cast_fp16")]; + tensor reshape_188_shape_0 = const()[name = tensor("reshape_188_shape_0"), val = tensor([-1])]; + tensor reshape_188_cast_fp16 = reshape(shape = reshape_188_shape_0, x = reshape_184_cast_fp16)[name = tensor("reshape_188_cast_fp16")]; + tensor scatter_37_mode_0 = const()[name = tensor("scatter_37_mode_0"), val = tensor("update")]; + tensor scatter_37_axis_0 = const()[name = tensor("scatter_37_axis_0"), val = tensor(0)]; + tensor scatter_37_validate_indices_0 = const()[name = tensor("scatter_37_validate_indices_0"), val = tensor(false)]; + tensor scatter_37_cast_fp16 = scatter(axis = scatter_37_axis_0, data = reshape_188_cast_fp16, indices = reshape_186, mode = scatter_37_mode_0, updates = reshape_187_cast_fp16, validate_indices = scatter_37_validate_indices_0)[name = tensor("scatter_37_cast_fp16")]; + tensor reshape_189_cast_fp16 = reshape(shape = shape_85, x = scatter_37_cast_fp16)[name = tensor("reshape_189_cast_fp16")]; + tensor reshape_189_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("reshape_189_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor keys_109_begin_0 = const()[name = tensor("keys_109_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor keys_109_end_0 = const()[name = tensor("keys_109_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor keys_109_end_mask_0 = const()[name = tensor("keys_109_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor keys_109_squeeze_mask_0 = const()[name = tensor("keys_109_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor keys_109_cast_fp16 = slice_by_index(begin = keys_109_begin_0, end = keys_109_end_0, end_mask = keys_109_end_mask_0, squeeze_mask = keys_109_squeeze_mask_0, x = reshape_189_cast_fp16)[name = tensor("keys_109_cast_fp16")]; + tensor values_109_begin_0 = const()[name = tensor("values_109_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor values_109_end_0 = const()[name = tensor("values_109_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor values_109_end_mask_0 = const()[name = tensor("values_109_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor values_109_squeeze_mask_0 = const()[name = tensor("values_109_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor values_109_cast_fp16 = slice_by_index(begin = values_109_begin_0, end = values_109_end_0, end_mask = values_109_end_mask_0, squeeze_mask = values_109_squeeze_mask_0, x = reshape_189_cast_fp16)[name = tensor("values_109_cast_fp16")]; + tensor var_7405_cast_fp16 = not_equal(x = keys_109_cast_fp16, y = keys_109_cast_fp16)[name = tensor("op_7405_cast_fp16")]; + tensor keys_111_cast_fp16 = select(a = var_463_to_fp16, b = keys_109_cast_fp16, cond = var_7405_cast_fp16)[name = tensor("keys_111_cast_fp16")]; + tensor var_7413_cast_fp16 = not_equal(x = values_109_cast_fp16, y = values_109_cast_fp16)[name = tensor("op_7413_cast_fp16")]; + tensor values_111_cast_fp16 = select(a = var_463_to_fp16, b = values_109_cast_fp16, cond = var_7413_cast_fp16)[name = tensor("values_111_cast_fp16")]; + tensor var_7437 = const()[name = tensor("op_7437"), val = tensor([0, 2, 1, 3])]; + tensor var_7453 = const()[name = tensor("op_7453"), val = tensor([1, 1, 1])]; + tensor var_7454_cast_fp16 = reshape(shape = var_7453, x = position18_to_fp16)[name = tensor("op_7454_cast_fp16")]; + tensor var_7474_to_fp16 = const()[name = tensor("op_7474_to_fp16"), val = tensor(0x1p+0)]; + tensor valid_len_37_cast_fp16 = add(x = var_7454_cast_fp16, y = var_7474_to_fp16)[name = tensor("valid_len_37_cast_fp16")]; + tensor valid_mask_37_cast_fp16 = less(x = k_positions_1_promoted_to_fp16, y = valid_len_37_cast_fp16)[name = tensor("valid_mask_37_cast_fp16")]; + tensor causal_mask_37_cast_fp16 = less_equal(x = k_positions_1_promoted_to_fp16, y = var_7454_cast_fp16)[name = tensor("causal_mask_37_cast_fp16")]; + tensor attn_mask_73 = logical_and(x = valid_mask_37_cast_fp16, y = causal_mask_37_cast_fp16)[name = tensor("attn_mask_73")]; + tensor attn_mask_75_axes_0 = const()[name = tensor("attn_mask_75_axes_0"), val = tensor([1])]; + tensor attn_mask_75 = expand_dims(axes = attn_mask_75_axes_0, x = attn_mask_73)[name = tensor("attn_mask_75")]; + tensor var_7492_transpose_x_0 = const()[name = tensor("op_7492_transpose_x_0"), val = tensor(false)]; + tensor var_7492_transpose_y_0 = const()[name = tensor("op_7492_transpose_y_0"), val = tensor(false)]; + tensor transpose_108_perm_0 = const()[name = tensor("transpose_108_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_109_perm_0 = const()[name = tensor("transpose_109_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_109 = transpose(perm = transpose_109_perm_0, x = keys_111_cast_fp16)[name = tensor("transpose_141")]; + tensor transpose_108 = transpose(perm = transpose_108_perm_0, x = q_111_cast_fp16)[name = tensor("transpose_142")]; + tensor var_7492_cast_fp16 = matmul(transpose_x = var_7492_transpose_x_0, transpose_y = var_7492_transpose_y_0, x = transpose_108, y = transpose_109)[name = tensor("op_7492_cast_fp16")]; + tensor var_7486_to_fp16 = const()[name = tensor("op_7486_to_fp16"), val = tensor([0x1p-3])]; + tensor attn_weights_109_cast_fp16 = mul(x = var_7492_cast_fp16, y = var_7486_to_fp16)[name = tensor("attn_weights_109_cast_fp16")]; + tensor var_7494 = logical_not(x = attn_mask_75)[name = tensor("op_7494")]; + tensor var_7495_to_fp16 = const()[name = tensor("op_7495_to_fp16"), val = tensor(-inf)]; + tensor attn_weights_111_cast_fp16 = select(a = var_7495_to_fp16, b = attn_weights_109_cast_fp16, cond = var_7494)[name = tensor("attn_weights_111_cast_fp16")]; + tensor var_7497 = const()[name = tensor("op_7497"), val = tensor(-1)]; + tensor attn_weights_113_cast_fp16 = softmax(axis = var_7497, x = attn_weights_111_cast_fp16)[name = tensor("attn_weights_113_cast_fp16")]; + tensor attn_output_37_transpose_x_0 = const()[name = tensor("attn_output_37_transpose_x_0"), val = tensor(false)]; + tensor attn_output_37_transpose_y_0 = const()[name = tensor("attn_output_37_transpose_y_0"), val = tensor(false)]; + tensor values_113_cast_fp16 = transpose(perm = var_7437, x = values_111_cast_fp16)[name = tensor("transpose_143")]; + tensor attn_output_37_cast_fp16 = matmul(transpose_x = attn_output_37_transpose_x_0, transpose_y = attn_output_37_transpose_y_0, x = attn_weights_113_cast_fp16, y = values_113_cast_fp16)[name = tensor("attn_output_37_cast_fp16")]; + tensor var_7505 = const()[name = tensor("op_7505"), val = tensor([0, 2, 1, 3])]; + tensor var_7510 = const()[name = tensor("op_7510"), val = tensor([1, 1, 1024])]; + tensor var_7506_cast_fp16 = transpose(perm = var_7505, x = attn_output_37_cast_fp16)[name = tensor("transpose_140")]; + tensor input_223_cast_fp16 = reshape(shape = var_7510, x = var_7506_cast_fp16)[name = tensor("input_223_cast_fp16")]; + tensor attn18_out_proj_weight_to_fp16 = const()[name = tensor("attn18_out_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(460571648)))]; + tensor linear_74_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = attn18_out_proj_weight_to_fp16, x = input_223_cast_fp16)[name = tensor("linear_74_cast_fp16")]; + tensor var_7516_to_fp16 = const()[name = tensor("op_7516_to_fp16"), val = tensor(0x1p+0)]; + tensor var_7517_cast_fp16 = add(x = position18_to_fp16, y = var_7516_to_fp16)[name = tensor("op_7517_cast_fp16")]; + tensor var_7517_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_7517_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor input_225_cast_fp16 = add(x = input_219_cast_fp16, y = linear_74_cast_fp16)[name = tensor("input_225_cast_fp16")]; + tensor input_227_axes_0 = const()[name = tensor("input_227_axes_0"), val = tensor([-1])]; + tensor norm18_2_weight_to_fp16 = const()[name = tensor("norm18_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(462668864)))]; + tensor norm18_2_bias_to_fp16 = const()[name = tensor("norm18_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(462670976)))]; + tensor var_7521_to_fp16 = const()[name = tensor("op_7521_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_227_cast_fp16 = layer_norm(axes = input_227_axes_0, beta = norm18_2_bias_to_fp16, epsilon = var_7521_to_fp16, gamma = norm18_2_weight_to_fp16, x = input_225_cast_fp16)[name = tensor("input_227_cast_fp16")]; + tensor linear18_1_weight_to_fp16 = const()[name = tensor("linear18_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(462673088)))]; + tensor linear_75_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = linear18_1_weight_to_fp16, x = input_227_cast_fp16)[name = tensor("linear_75_cast_fp16")]; + tensor input_229_mode_0 = const()[name = tensor("input_229_mode_0"), val = tensor("EXACT")]; + tensor input_229_cast_fp16 = gelu(mode = input_229_mode_0, x = linear_75_cast_fp16)[name = tensor("input_229_cast_fp16")]; + tensor linear18_2_weight_to_fp16 = const()[name = tensor("linear18_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(471061760)))]; + tensor linear_76_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = linear18_2_weight_to_fp16, x = input_229_cast_fp16)[name = tensor("linear_76_cast_fp16")]; + tensor input_231_cast_fp16 = add(x = input_225_cast_fp16, y = linear_76_cast_fp16)[name = tensor("input_231_cast_fp16")]; + tensor input_233_axes_0 = const()[name = tensor("input_233_axes_0"), val = tensor([-1])]; + tensor norm19_1_weight_to_fp16 = const()[name = tensor("norm19_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(479450432)))]; + tensor norm19_1_bias_to_fp16 = const()[name = tensor("norm19_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(479452544)))]; + tensor var_7538_to_fp16 = const()[name = tensor("op_7538_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_233_cast_fp16 = layer_norm(axes = input_233_axes_0, beta = norm19_1_bias_to_fp16, epsilon = var_7538_to_fp16, gamma = norm19_1_weight_to_fp16, x = input_231_cast_fp16)[name = tensor("input_233_cast_fp16")]; + tensor attn19_in_proj_weight_to_fp16 = const()[name = tensor("attn19_in_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(479454656)))]; + tensor linear_77_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = attn19_in_proj_weight_to_fp16, x = input_233_cast_fp16)[name = tensor("linear_77_cast_fp16")]; + tensor var_7552 = const()[name = tensor("op_7552"), val = tensor([1, 1, 3, 16, 64])]; + tensor qkv_39_cast_fp16 = reshape(shape = var_7552, x = linear_77_cast_fp16)[name = tensor("qkv_39_cast_fp16")]; + tensor q_115_begin_0 = const()[name = tensor("q_115_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor q_115_end_0 = const()[name = tensor("q_115_end_0"), val = tensor([1, 1, 1, 16, 64])]; + tensor q_115_end_mask_0 = const()[name = tensor("q_115_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor q_115_squeeze_mask_0 = const()[name = tensor("q_115_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor q_115_cast_fp16 = slice_by_index(begin = q_115_begin_0, end = q_115_end_0, end_mask = q_115_end_mask_0, squeeze_mask = q_115_squeeze_mask_0, x = qkv_39_cast_fp16)[name = tensor("q_115_cast_fp16")]; + tensor k_77_begin_0 = const()[name = tensor("k_77_begin_0"), val = tensor([0, 0, 1, 0, 0])]; + tensor k_77_end_0 = const()[name = tensor("k_77_end_0"), val = tensor([1, 1, 2, 16, 64])]; + tensor k_77_end_mask_0 = const()[name = tensor("k_77_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor k_77_squeeze_mask_0 = const()[name = tensor("k_77_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor k_77_cast_fp16 = slice_by_index(begin = k_77_begin_0, end = k_77_end_0, end_mask = k_77_end_mask_0, squeeze_mask = k_77_squeeze_mask_0, x = qkv_39_cast_fp16)[name = tensor("k_77_cast_fp16")]; + tensor v_39_begin_0 = const()[name = tensor("v_39_begin_0"), val = tensor([0, 0, 2, 0, 0])]; + tensor v_39_end_0 = const()[name = tensor("v_39_end_0"), val = tensor([1, 1, 3, 16, 64])]; + tensor v_39_end_mask_0 = const()[name = tensor("v_39_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor v_39_squeeze_mask_0 = const()[name = tensor("v_39_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor v_39_cast_fp16 = slice_by_index(begin = v_39_begin_0, end = v_39_end_0, end_mask = v_39_end_mask_0, squeeze_mask = v_39_squeeze_mask_0, x = qkv_39_cast_fp16)[name = tensor("v_39_cast_fp16")]; + tensor var_7618 = const()[name = tensor("op_7618"), val = tensor([1, 1, 1, 1])]; + tensor position19_to_fp16_dtype_0 = const()[name = tensor("position19_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor position19_to_fp16 = cast(dtype = position19_to_fp16_dtype_0, x = position19)[name = tensor("cast_458")]; + tensor ts_119_cast_fp16 = reshape(shape = var_7618, x = position19_to_fp16)[name = tensor("ts_119_cast_fp16")]; + tensor var_7625 = const()[name = tensor("op_7625"), val = tensor([1, 1, 16, 32, 2])]; + tensor q_complex_39_cast_fp16 = reshape(shape = var_7625, x = q_115_cast_fp16)[name = tensor("q_complex_39_cast_fp16")]; + tensor var_7632 = const()[name = tensor("op_7632"), val = tensor([1, 1, 16, 32, 2])]; + tensor k_complex_39_cast_fp16 = reshape(shape = var_7632, x = k_77_cast_fp16)[name = tensor("k_complex_39_cast_fp16")]; + tensor var_7636_begin_0 = const()[name = tensor("op_7636_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_7636_end_0 = const()[name = tensor("op_7636_end_0"), val = tensor([1, 1, 16, 32, 1])]; + tensor var_7636_end_mask_0 = const()[name = tensor("op_7636_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_7636_squeeze_mask_0 = const()[name = tensor("op_7636_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_7636_cast_fp16 = slice_by_index(begin = var_7636_begin_0, end = var_7636_end_0, end_mask = var_7636_end_mask_0, squeeze_mask = var_7636_squeeze_mask_0, x = q_complex_39_cast_fp16)[name = tensor("op_7636_cast_fp16")]; + tensor var_7644_begin_0 = const()[name = tensor("op_7644_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_7644_end_0 = const()[name = tensor("op_7644_end_0"), val = tensor([1, 1, 16, 32, 2])]; + tensor var_7644_end_mask_0 = const()[name = tensor("op_7644_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_7644_squeeze_mask_0 = const()[name = tensor("op_7644_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_7644_cast_fp16 = slice_by_index(begin = var_7644_begin_0, end = var_7644_end_0, end_mask = var_7644_end_mask_0, squeeze_mask = var_7644_squeeze_mask_0, x = q_complex_39_cast_fp16)[name = tensor("op_7644_cast_fp16")]; + tensor var_7652_begin_0 = const()[name = tensor("op_7652_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_7652_end_0 = const()[name = tensor("op_7652_end_0"), val = tensor([1, 1, 16, 32, 1])]; + tensor var_7652_end_mask_0 = const()[name = tensor("op_7652_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_7652_squeeze_mask_0 = const()[name = tensor("op_7652_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_7652_cast_fp16 = slice_by_index(begin = var_7652_begin_0, end = var_7652_end_0, end_mask = var_7652_end_mask_0, squeeze_mask = var_7652_squeeze_mask_0, x = k_complex_39_cast_fp16)[name = tensor("op_7652_cast_fp16")]; + tensor var_7660_begin_0 = const()[name = tensor("op_7660_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_7660_end_0 = const()[name = tensor("op_7660_end_0"), val = tensor([1, 1, 16, 32, 2])]; + tensor var_7660_end_mask_0 = const()[name = tensor("op_7660_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_7660_squeeze_mask_0 = const()[name = tensor("op_7660_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_7660_cast_fp16 = slice_by_index(begin = var_7660_begin_0, end = var_7660_end_0, end_mask = var_7660_end_mask_0, squeeze_mask = var_7660_squeeze_mask_0, x = k_complex_39_cast_fp16)[name = tensor("op_7660_cast_fp16")]; + tensor freqs_39_to_fp16 = const()[name = tensor("freqs_39_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(485746176)))]; + tensor var_7666_cast_fp16 = mul(x = freqs_39_to_fp16, y = ts_119_cast_fp16)[name = tensor("op_7666_cast_fp16")]; + tensor rotr_39_cast_fp16 = cos(x = var_7666_cast_fp16)[name = tensor("rotr_39_cast_fp16")]; + tensor roti_39_cast_fp16 = sin(x = var_7666_cast_fp16)[name = tensor("roti_39_cast_fp16")]; + tensor var_7670_cast_fp16 = mul(x = var_7636_cast_fp16, y = rotr_39_cast_fp16)[name = tensor("op_7670_cast_fp16")]; + tensor var_7671_cast_fp16 = mul(x = var_7644_cast_fp16, y = roti_39_cast_fp16)[name = tensor("op_7671_cast_fp16")]; + tensor qor_77_cast_fp16 = sub(x = var_7670_cast_fp16, y = var_7671_cast_fp16)[name = tensor("qor_77_cast_fp16")]; + tensor var_7674_cast_fp16 = mul(x = var_7636_cast_fp16, y = roti_39_cast_fp16)[name = tensor("op_7674_cast_fp16")]; + tensor var_7675_cast_fp16 = mul(x = var_7644_cast_fp16, y = rotr_39_cast_fp16)[name = tensor("op_7675_cast_fp16")]; + tensor qoi_77_cast_fp16 = add(x = var_7674_cast_fp16, y = var_7675_cast_fp16)[name = tensor("qoi_77_cast_fp16")]; + tensor var_7678_cast_fp16 = mul(x = var_7652_cast_fp16, y = rotr_39_cast_fp16)[name = tensor("op_7678_cast_fp16")]; + tensor var_7679_cast_fp16 = mul(x = var_7660_cast_fp16, y = roti_39_cast_fp16)[name = tensor("op_7679_cast_fp16")]; + tensor kor_77_cast_fp16 = sub(x = var_7678_cast_fp16, y = var_7679_cast_fp16)[name = tensor("kor_77_cast_fp16")]; + tensor var_7682_cast_fp16 = mul(x = var_7652_cast_fp16, y = roti_39_cast_fp16)[name = tensor("op_7682_cast_fp16")]; + tensor var_7683_cast_fp16 = mul(x = var_7660_cast_fp16, y = rotr_39_cast_fp16)[name = tensor("op_7683_cast_fp16")]; + tensor koi_77_cast_fp16 = add(x = var_7682_cast_fp16, y = var_7683_cast_fp16)[name = tensor("koi_77_cast_fp16")]; + tensor qo_39_axis_0 = const()[name = tensor("qo_39_axis_0"), val = tensor(-1)]; + tensor qo_39_cast_fp16 = stack(axis = qo_39_axis_0, values = (qor_77_cast_fp16, qoi_77_cast_fp16))[name = tensor("qo_39_cast_fp16")]; + tensor ko_39_axis_0 = const()[name = tensor("ko_39_axis_0"), val = tensor(-1)]; + tensor ko_39_cast_fp16 = stack(axis = ko_39_axis_0, values = (kor_77_cast_fp16, koi_77_cast_fp16))[name = tensor("ko_39_cast_fp16")]; + tensor var_7716 = const()[name = tensor("op_7716"), val = tensor([1, 1, 16, 64])]; + tensor q_117_cast_fp16 = reshape(shape = var_7716, x = qo_39_cast_fp16)[name = tensor("q_117_cast_fp16")]; + tensor var_7722 = const()[name = tensor("op_7722"), val = tensor([1, 1, 16, 64])]; + tensor k_79_cast_fp16 = reshape(shape = var_7722, x = ko_39_cast_fp16)[name = tensor("k_79_cast_fp16")]; + tensor _inversed_7747_y_0_to_fp16 = const()[name = tensor("_inversed_7747_y_0_to_fp16"), val = tensor(0x1p-9)]; + tensor _inversed_7747_cast_fp16 = mul(x = ts_119_cast_fp16, y = _inversed_7747_y_0_to_fp16)[name = tensor("_inversed_7747_cast_fp16")]; + tensor var_7748_cast_fp16 = floor(x = _inversed_7747_cast_fp16)[name = tensor("op_7748_cast_fp16")]; + tensor var_7749_to_fp16 = const()[name = tensor("op_7749_to_fp16"), val = tensor(0x1p+9)]; + tensor var_7750_cast_fp16 = mul(x = var_7748_cast_fp16, y = var_7749_to_fp16)[name = tensor("op_7750_cast_fp16")]; + tensor write_indices_float_79_cast_fp16 = sub(x = ts_119_cast_fp16, y = var_7750_cast_fp16)[name = tensor("write_indices_float_79_cast_fp16")]; + tensor var_7757_dtype_0 = const()[name = tensor("op_7757_dtype_0"), val = tensor("int32")]; + tensor write_indices_39_reps_0 = const()[name = tensor("write_indices_39_reps_0"), val = tensor([1, 1, 16, 64])]; + tensor write_indices_float_79_cast_fp16_to_int32 = cast(dtype = var_7757_dtype_0, x = write_indices_float_79_cast_fp16)[name = tensor("cast_457")]; + tensor write_indices_39 = tile(reps = write_indices_39_reps_0, x = write_indices_float_79_cast_fp16_to_int32)[name = tensor("write_indices_39")]; + tensor var_7767_begin_0 = const()[name = tensor("op_7767_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_7767_end_0 = const()[name = tensor("op_7767_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor var_7767_end_mask_0 = const()[name = tensor("op_7767_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_7767_squeeze_mask_0 = const()[name = tensor("op_7767_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor cache19_to_fp16_dtype_0 = const()[name = tensor("cache19_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor cache19_to_fp16 = cast(dtype = cache19_to_fp16_dtype_0, x = cache19)[name = tensor("cast_456")]; + tensor var_7767_cast_fp16 = slice_by_index(begin = var_7767_begin_0, end = var_7767_end_0, end_mask = var_7767_end_mask_0, squeeze_mask = var_7767_squeeze_mask_0, x = cache19_to_fp16)[name = tensor("op_7767_cast_fp16")]; + tensor var_7769_axis_0 = const()[name = tensor("op_7769_axis_0"), val = tensor(1)]; + tensor var_7769_mode_0 = const()[name = tensor("op_7769_mode_0"), val = tensor("update")]; + tensor var_7769_validate_indices_0 = const()[name = tensor("op_7769_validate_indices_0"), val = tensor(false)]; + tensor var_7769_cast_fp16 = scatter_along_axis(axis = var_7769_axis_0, data = var_7767_cast_fp16, indices = write_indices_39, mode = var_7769_mode_0, updates = k_79_cast_fp16, validate_indices = var_7769_validate_indices_0)[name = tensor("op_7769_cast_fp16")]; + tensor concat_135 = const()[name = tensor("concat_135"), val = tensor([0, 0, 0, 0, 0])]; + tensor concat_136 = const()[name = tensor("concat_136"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_39_internal_tensor_assign_1_stride_0 = const()[name = tensor("new_cache_39_internal_tensor_assign_1_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_39_internal_tensor_assign_1_begin_mask_0 = const()[name = tensor("new_cache_39_internal_tensor_assign_1_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_39_internal_tensor_assign_1_end_mask_0 = const()[name = tensor("new_cache_39_internal_tensor_assign_1_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_39_internal_tensor_assign_1_squeeze_mask_0 = const()[name = tensor("new_cache_39_internal_tensor_assign_1_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_86 = const()[name = tensor("shape_86"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_38 = const()[name = tensor("reduce_prod_38"), val = tensor(1048576)]; + tensor range_1d_38_start_0 = const()[name = tensor("range_1d_38_start_0"), val = tensor(0)]; + tensor range_1d_38_step_0 = const()[name = tensor("range_1d_38_step_0"), val = tensor(1)]; + tensor range_1d_38 = range_1d(end = reduce_prod_38, start = range_1d_38_start_0, step = range_1d_38_step_0)[name = tensor("range_1d_38")]; + tensor reshape_190 = reshape(shape = shape_86, x = range_1d_38)[name = tensor("reshape_190")]; + tensor slice_by_index_38 = slice_by_index(begin = concat_135, begin_mask = new_cache_39_internal_tensor_assign_1_begin_mask_0, end = concat_136, end_mask = new_cache_39_internal_tensor_assign_1_end_mask_0, squeeze_mask = new_cache_39_internal_tensor_assign_1_squeeze_mask_0, stride = new_cache_39_internal_tensor_assign_1_stride_0, x = reshape_190)[name = tensor("slice_by_index_38")]; + tensor reshape_191_shape_0 = const()[name = tensor("reshape_191_shape_0"), val = tensor([-1])]; + tensor reshape_191 = reshape(shape = reshape_191_shape_0, x = slice_by_index_38)[name = tensor("reshape_191")]; + tensor reshape_192_shape_0 = const()[name = tensor("reshape_192_shape_0"), val = tensor([-1])]; + tensor reshape_192_cast_fp16 = reshape(shape = reshape_192_shape_0, x = var_7769_cast_fp16)[name = tensor("reshape_192_cast_fp16")]; + tensor reshape_193_shape_0 = const()[name = tensor("reshape_193_shape_0"), val = tensor([-1])]; + tensor reshape_193_cast_fp16 = reshape(shape = reshape_193_shape_0, x = cache19_to_fp16)[name = tensor("reshape_193_cast_fp16")]; + tensor scatter_38_mode_0 = const()[name = tensor("scatter_38_mode_0"), val = tensor("update")]; + tensor scatter_38_axis_0 = const()[name = tensor("scatter_38_axis_0"), val = tensor(0)]; + tensor scatter_38_validate_indices_0 = const()[name = tensor("scatter_38_validate_indices_0"), val = tensor(false)]; + tensor scatter_38_cast_fp16 = scatter(axis = scatter_38_axis_0, data = reshape_193_cast_fp16, indices = reshape_191, mode = scatter_38_mode_0, updates = reshape_192_cast_fp16, validate_indices = scatter_38_validate_indices_0)[name = tensor("scatter_38_cast_fp16")]; + tensor reshape_194_cast_fp16 = reshape(shape = shape_86, x = scatter_38_cast_fp16)[name = tensor("reshape_194_cast_fp16")]; + tensor var_7777_begin_0 = const()[name = tensor("op_7777_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor var_7777_end_0 = const()[name = tensor("op_7777_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor var_7777_end_mask_0 = const()[name = tensor("op_7777_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_7777_squeeze_mask_0 = const()[name = tensor("op_7777_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor var_7777_cast_fp16 = slice_by_index(begin = var_7777_begin_0, end = var_7777_end_0, end_mask = var_7777_end_mask_0, squeeze_mask = var_7777_squeeze_mask_0, x = reshape_194_cast_fp16)[name = tensor("op_7777_cast_fp16")]; + tensor var_7779_axis_0 = const()[name = tensor("op_7779_axis_0"), val = tensor(1)]; + tensor var_7779_mode_0 = const()[name = tensor("op_7779_mode_0"), val = tensor("update")]; + tensor var_7779_validate_indices_0 = const()[name = tensor("op_7779_validate_indices_0"), val = tensor(false)]; + tensor var_7779_cast_fp16 = scatter_along_axis(axis = var_7779_axis_0, data = var_7777_cast_fp16, indices = write_indices_39, mode = var_7779_mode_0, updates = v_39_cast_fp16, validate_indices = var_7779_validate_indices_0)[name = tensor("op_7779_cast_fp16")]; + tensor concat_137 = const()[name = tensor("concat_137"), val = tensor([1, 0, 0, 0, 0])]; + tensor concat_138 = const()[name = tensor("concat_138"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_39_internal_tensor_assign_2_stride_0 = const()[name = tensor("new_cache_39_internal_tensor_assign_2_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_39_internal_tensor_assign_2_begin_mask_0 = const()[name = tensor("new_cache_39_internal_tensor_assign_2_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_39_internal_tensor_assign_2_end_mask_0 = const()[name = tensor("new_cache_39_internal_tensor_assign_2_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_39_internal_tensor_assign_2_squeeze_mask_0 = const()[name = tensor("new_cache_39_internal_tensor_assign_2_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_87 = const()[name = tensor("shape_87"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_39 = const()[name = tensor("reduce_prod_39"), val = tensor(1048576)]; + tensor range_1d_39_start_0 = const()[name = tensor("range_1d_39_start_0"), val = tensor(0)]; + tensor range_1d_39_step_0 = const()[name = tensor("range_1d_39_step_0"), val = tensor(1)]; + tensor range_1d_39 = range_1d(end = reduce_prod_39, start = range_1d_39_start_0, step = range_1d_39_step_0)[name = tensor("range_1d_39")]; + tensor reshape_195 = reshape(shape = shape_87, x = range_1d_39)[name = tensor("reshape_195")]; + tensor slice_by_index_39 = slice_by_index(begin = concat_137, begin_mask = new_cache_39_internal_tensor_assign_2_begin_mask_0, end = concat_138, end_mask = new_cache_39_internal_tensor_assign_2_end_mask_0, squeeze_mask = new_cache_39_internal_tensor_assign_2_squeeze_mask_0, stride = new_cache_39_internal_tensor_assign_2_stride_0, x = reshape_195)[name = tensor("slice_by_index_39")]; + tensor reshape_196_shape_0 = const()[name = tensor("reshape_196_shape_0"), val = tensor([-1])]; + tensor reshape_196 = reshape(shape = reshape_196_shape_0, x = slice_by_index_39)[name = tensor("reshape_196")]; + tensor reshape_197_shape_0 = const()[name = tensor("reshape_197_shape_0"), val = tensor([-1])]; + tensor reshape_197_cast_fp16 = reshape(shape = reshape_197_shape_0, x = var_7779_cast_fp16)[name = tensor("reshape_197_cast_fp16")]; + tensor reshape_198_shape_0 = const()[name = tensor("reshape_198_shape_0"), val = tensor([-1])]; + tensor reshape_198_cast_fp16 = reshape(shape = reshape_198_shape_0, x = reshape_194_cast_fp16)[name = tensor("reshape_198_cast_fp16")]; + tensor scatter_39_mode_0 = const()[name = tensor("scatter_39_mode_0"), val = tensor("update")]; + tensor scatter_39_axis_0 = const()[name = tensor("scatter_39_axis_0"), val = tensor(0)]; + tensor scatter_39_validate_indices_0 = const()[name = tensor("scatter_39_validate_indices_0"), val = tensor(false)]; + tensor scatter_39_cast_fp16 = scatter(axis = scatter_39_axis_0, data = reshape_198_cast_fp16, indices = reshape_196, mode = scatter_39_mode_0, updates = reshape_197_cast_fp16, validate_indices = scatter_39_validate_indices_0)[name = tensor("scatter_39_cast_fp16")]; + tensor reshape_199_cast_fp16 = reshape(shape = shape_87, x = scatter_39_cast_fp16)[name = tensor("reshape_199_cast_fp16")]; + tensor reshape_199_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("reshape_199_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor keys_115_begin_0 = const()[name = tensor("keys_115_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor keys_115_end_0 = const()[name = tensor("keys_115_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor keys_115_end_mask_0 = const()[name = tensor("keys_115_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor keys_115_squeeze_mask_0 = const()[name = tensor("keys_115_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor keys_115_cast_fp16 = slice_by_index(begin = keys_115_begin_0, end = keys_115_end_0, end_mask = keys_115_end_mask_0, squeeze_mask = keys_115_squeeze_mask_0, x = reshape_199_cast_fp16)[name = tensor("keys_115_cast_fp16")]; + tensor values_115_begin_0 = const()[name = tensor("values_115_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor values_115_end_0 = const()[name = tensor("values_115_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor values_115_end_mask_0 = const()[name = tensor("values_115_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor values_115_squeeze_mask_0 = const()[name = tensor("values_115_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor values_115_cast_fp16 = slice_by_index(begin = values_115_begin_0, end = values_115_end_0, end_mask = values_115_end_mask_0, squeeze_mask = values_115_squeeze_mask_0, x = reshape_199_cast_fp16)[name = tensor("values_115_cast_fp16")]; + tensor var_7791_cast_fp16 = not_equal(x = keys_115_cast_fp16, y = keys_115_cast_fp16)[name = tensor("op_7791_cast_fp16")]; + tensor keys_117_cast_fp16 = select(a = var_463_to_fp16, b = keys_115_cast_fp16, cond = var_7791_cast_fp16)[name = tensor("keys_117_cast_fp16")]; + tensor var_7799_cast_fp16 = not_equal(x = values_115_cast_fp16, y = values_115_cast_fp16)[name = tensor("op_7799_cast_fp16")]; + tensor values_117_cast_fp16 = select(a = var_463_to_fp16, b = values_115_cast_fp16, cond = var_7799_cast_fp16)[name = tensor("values_117_cast_fp16")]; + tensor var_7823 = const()[name = tensor("op_7823"), val = tensor([0, 2, 1, 3])]; + tensor var_7839 = const()[name = tensor("op_7839"), val = tensor([1, 1, 1])]; + tensor var_7840_cast_fp16 = reshape(shape = var_7839, x = position19_to_fp16)[name = tensor("op_7840_cast_fp16")]; + tensor var_7860_to_fp16 = const()[name = tensor("op_7860_to_fp16"), val = tensor(0x1p+0)]; + tensor valid_len_39_cast_fp16 = add(x = var_7840_cast_fp16, y = var_7860_to_fp16)[name = tensor("valid_len_39_cast_fp16")]; + tensor valid_mask_39_cast_fp16 = less(x = k_positions_1_promoted_to_fp16, y = valid_len_39_cast_fp16)[name = tensor("valid_mask_39_cast_fp16")]; + tensor causal_mask_39_cast_fp16 = less_equal(x = k_positions_1_promoted_to_fp16, y = var_7840_cast_fp16)[name = tensor("causal_mask_39_cast_fp16")]; + tensor attn_mask_77 = logical_and(x = valid_mask_39_cast_fp16, y = causal_mask_39_cast_fp16)[name = tensor("attn_mask_77")]; + tensor attn_mask_79_axes_0 = const()[name = tensor("attn_mask_79_axes_0"), val = tensor([1])]; + tensor attn_mask_79 = expand_dims(axes = attn_mask_79_axes_0, x = attn_mask_77)[name = tensor("attn_mask_79")]; + tensor var_7878_transpose_x_0 = const()[name = tensor("op_7878_transpose_x_0"), val = tensor(false)]; + tensor var_7878_transpose_y_0 = const()[name = tensor("op_7878_transpose_y_0"), val = tensor(false)]; + tensor transpose_110_perm_0 = const()[name = tensor("transpose_110_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_111_perm_0 = const()[name = tensor("transpose_111_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_111 = transpose(perm = transpose_111_perm_0, x = keys_117_cast_fp16)[name = tensor("transpose_137")]; + tensor transpose_110 = transpose(perm = transpose_110_perm_0, x = q_117_cast_fp16)[name = tensor("transpose_138")]; + tensor var_7878_cast_fp16 = matmul(transpose_x = var_7878_transpose_x_0, transpose_y = var_7878_transpose_y_0, x = transpose_110, y = transpose_111)[name = tensor("op_7878_cast_fp16")]; + tensor var_7872_to_fp16 = const()[name = tensor("op_7872_to_fp16"), val = tensor([0x1p-3])]; + tensor attn_weights_115_cast_fp16 = mul(x = var_7878_cast_fp16, y = var_7872_to_fp16)[name = tensor("attn_weights_115_cast_fp16")]; + tensor var_7880 = logical_not(x = attn_mask_79)[name = tensor("op_7880")]; + tensor var_7881_to_fp16 = const()[name = tensor("op_7881_to_fp16"), val = tensor(-inf)]; + tensor attn_weights_117_cast_fp16 = select(a = var_7881_to_fp16, b = attn_weights_115_cast_fp16, cond = var_7880)[name = tensor("attn_weights_117_cast_fp16")]; + tensor var_7883 = const()[name = tensor("op_7883"), val = tensor(-1)]; + tensor attn_weights_119_cast_fp16 = softmax(axis = var_7883, x = attn_weights_117_cast_fp16)[name = tensor("attn_weights_119_cast_fp16")]; + tensor attn_output_39_transpose_x_0 = const()[name = tensor("attn_output_39_transpose_x_0"), val = tensor(false)]; + tensor attn_output_39_transpose_y_0 = const()[name = tensor("attn_output_39_transpose_y_0"), val = tensor(false)]; + tensor values_119_cast_fp16 = transpose(perm = var_7823, x = values_117_cast_fp16)[name = tensor("transpose_139")]; + tensor attn_output_39_cast_fp16 = matmul(transpose_x = attn_output_39_transpose_x_0, transpose_y = attn_output_39_transpose_y_0, x = attn_weights_119_cast_fp16, y = values_119_cast_fp16)[name = tensor("attn_output_39_cast_fp16")]; + tensor var_7891 = const()[name = tensor("op_7891"), val = tensor([0, 2, 1, 3])]; + tensor var_7896 = const()[name = tensor("op_7896"), val = tensor([1, 1, 1024])]; + tensor var_7892_cast_fp16 = transpose(perm = var_7891, x = attn_output_39_cast_fp16)[name = tensor("transpose_136")]; + tensor input_235_cast_fp16 = reshape(shape = var_7896, x = var_7892_cast_fp16)[name = tensor("input_235_cast_fp16")]; + tensor attn19_out_proj_weight_to_fp16 = const()[name = tensor("attn19_out_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(485746304)))]; + tensor linear_78_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = attn19_out_proj_weight_to_fp16, x = input_235_cast_fp16)[name = tensor("linear_78_cast_fp16")]; + tensor var_7902_to_fp16 = const()[name = tensor("op_7902_to_fp16"), val = tensor(0x1p+0)]; + tensor var_7903_cast_fp16 = add(x = position19_to_fp16, y = var_7902_to_fp16)[name = tensor("op_7903_cast_fp16")]; + tensor var_7903_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_7903_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor input_237_cast_fp16 = add(x = input_231_cast_fp16, y = linear_78_cast_fp16)[name = tensor("input_237_cast_fp16")]; + tensor input_239_axes_0 = const()[name = tensor("input_239_axes_0"), val = tensor([-1])]; + tensor norm19_2_weight_to_fp16 = const()[name = tensor("norm19_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(487843520)))]; + tensor norm19_2_bias_to_fp16 = const()[name = tensor("norm19_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(487845632)))]; + tensor var_7907_to_fp16 = const()[name = tensor("op_7907_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_239_cast_fp16 = layer_norm(axes = input_239_axes_0, beta = norm19_2_bias_to_fp16, epsilon = var_7907_to_fp16, gamma = norm19_2_weight_to_fp16, x = input_237_cast_fp16)[name = tensor("input_239_cast_fp16")]; + tensor linear19_1_weight_to_fp16 = const()[name = tensor("linear19_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(487847744)))]; + tensor linear_79_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = linear19_1_weight_to_fp16, x = input_239_cast_fp16)[name = tensor("linear_79_cast_fp16")]; + tensor input_241_mode_0 = const()[name = tensor("input_241_mode_0"), val = tensor("EXACT")]; + tensor input_241_cast_fp16 = gelu(mode = input_241_mode_0, x = linear_79_cast_fp16)[name = tensor("input_241_cast_fp16")]; + tensor linear19_2_weight_to_fp16 = const()[name = tensor("linear19_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(496236416)))]; + tensor linear_80_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = linear19_2_weight_to_fp16, x = input_241_cast_fp16)[name = tensor("linear_80_cast_fp16")]; + tensor input_243_cast_fp16 = add(x = input_237_cast_fp16, y = linear_80_cast_fp16)[name = tensor("input_243_cast_fp16")]; + tensor input_245_axes_0 = const()[name = tensor("input_245_axes_0"), val = tensor([-1])]; + tensor norm20_1_weight_to_fp16 = const()[name = tensor("norm20_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(504625088)))]; + tensor norm20_1_bias_to_fp16 = const()[name = tensor("norm20_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(504627200)))]; + tensor var_7924_to_fp16 = const()[name = tensor("op_7924_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_245_cast_fp16 = layer_norm(axes = input_245_axes_0, beta = norm20_1_bias_to_fp16, epsilon = var_7924_to_fp16, gamma = norm20_1_weight_to_fp16, x = input_243_cast_fp16)[name = tensor("input_245_cast_fp16")]; + tensor attn20_in_proj_weight_to_fp16 = const()[name = tensor("attn20_in_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(504629312)))]; + tensor linear_81_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = attn20_in_proj_weight_to_fp16, x = input_245_cast_fp16)[name = tensor("linear_81_cast_fp16")]; + tensor var_7938 = const()[name = tensor("op_7938"), val = tensor([1, 1, 3, 16, 64])]; + tensor qkv_41_cast_fp16 = reshape(shape = var_7938, x = linear_81_cast_fp16)[name = tensor("qkv_41_cast_fp16")]; + tensor q_121_begin_0 = const()[name = tensor("q_121_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor q_121_end_0 = const()[name = tensor("q_121_end_0"), val = tensor([1, 1, 1, 16, 64])]; + tensor q_121_end_mask_0 = const()[name = tensor("q_121_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor q_121_squeeze_mask_0 = const()[name = tensor("q_121_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor q_121_cast_fp16 = slice_by_index(begin = q_121_begin_0, end = q_121_end_0, end_mask = q_121_end_mask_0, squeeze_mask = q_121_squeeze_mask_0, x = qkv_41_cast_fp16)[name = tensor("q_121_cast_fp16")]; + tensor k_81_begin_0 = const()[name = tensor("k_81_begin_0"), val = tensor([0, 0, 1, 0, 0])]; + tensor k_81_end_0 = const()[name = tensor("k_81_end_0"), val = tensor([1, 1, 2, 16, 64])]; + tensor k_81_end_mask_0 = const()[name = tensor("k_81_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor k_81_squeeze_mask_0 = const()[name = tensor("k_81_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor k_81_cast_fp16 = slice_by_index(begin = k_81_begin_0, end = k_81_end_0, end_mask = k_81_end_mask_0, squeeze_mask = k_81_squeeze_mask_0, x = qkv_41_cast_fp16)[name = tensor("k_81_cast_fp16")]; + tensor v_41_begin_0 = const()[name = tensor("v_41_begin_0"), val = tensor([0, 0, 2, 0, 0])]; + tensor v_41_end_0 = const()[name = tensor("v_41_end_0"), val = tensor([1, 1, 3, 16, 64])]; + tensor v_41_end_mask_0 = const()[name = tensor("v_41_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor v_41_squeeze_mask_0 = const()[name = tensor("v_41_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor v_41_cast_fp16 = slice_by_index(begin = v_41_begin_0, end = v_41_end_0, end_mask = v_41_end_mask_0, squeeze_mask = v_41_squeeze_mask_0, x = qkv_41_cast_fp16)[name = tensor("v_41_cast_fp16")]; + tensor var_8004 = const()[name = tensor("op_8004"), val = tensor([1, 1, 1, 1])]; + tensor position20_to_fp16_dtype_0 = const()[name = tensor("position20_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor position20_to_fp16 = cast(dtype = position20_to_fp16_dtype_0, x = position20)[name = tensor("cast_453")]; + tensor ts_125_cast_fp16 = reshape(shape = var_8004, x = position20_to_fp16)[name = tensor("ts_125_cast_fp16")]; + tensor var_8011 = const()[name = tensor("op_8011"), val = tensor([1, 1, 16, 32, 2])]; + tensor q_complex_41_cast_fp16 = reshape(shape = var_8011, x = q_121_cast_fp16)[name = tensor("q_complex_41_cast_fp16")]; + tensor var_8018 = const()[name = tensor("op_8018"), val = tensor([1, 1, 16, 32, 2])]; + tensor k_complex_41_cast_fp16 = reshape(shape = var_8018, x = k_81_cast_fp16)[name = tensor("k_complex_41_cast_fp16")]; + tensor var_8022_begin_0 = const()[name = tensor("op_8022_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_8022_end_0 = const()[name = tensor("op_8022_end_0"), val = tensor([1, 1, 16, 32, 1])]; + tensor var_8022_end_mask_0 = const()[name = tensor("op_8022_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_8022_squeeze_mask_0 = const()[name = tensor("op_8022_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_8022_cast_fp16 = slice_by_index(begin = var_8022_begin_0, end = var_8022_end_0, end_mask = var_8022_end_mask_0, squeeze_mask = var_8022_squeeze_mask_0, x = q_complex_41_cast_fp16)[name = tensor("op_8022_cast_fp16")]; + tensor var_8030_begin_0 = const()[name = tensor("op_8030_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_8030_end_0 = const()[name = tensor("op_8030_end_0"), val = tensor([1, 1, 16, 32, 2])]; + tensor var_8030_end_mask_0 = const()[name = tensor("op_8030_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_8030_squeeze_mask_0 = const()[name = tensor("op_8030_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_8030_cast_fp16 = slice_by_index(begin = var_8030_begin_0, end = var_8030_end_0, end_mask = var_8030_end_mask_0, squeeze_mask = var_8030_squeeze_mask_0, x = q_complex_41_cast_fp16)[name = tensor("op_8030_cast_fp16")]; + tensor var_8038_begin_0 = const()[name = tensor("op_8038_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_8038_end_0 = const()[name = tensor("op_8038_end_0"), val = tensor([1, 1, 16, 32, 1])]; + tensor var_8038_end_mask_0 = const()[name = tensor("op_8038_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_8038_squeeze_mask_0 = const()[name = tensor("op_8038_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_8038_cast_fp16 = slice_by_index(begin = var_8038_begin_0, end = var_8038_end_0, end_mask = var_8038_end_mask_0, squeeze_mask = var_8038_squeeze_mask_0, x = k_complex_41_cast_fp16)[name = tensor("op_8038_cast_fp16")]; + tensor var_8046_begin_0 = const()[name = tensor("op_8046_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_8046_end_0 = const()[name = tensor("op_8046_end_0"), val = tensor([1, 1, 16, 32, 2])]; + tensor var_8046_end_mask_0 = const()[name = tensor("op_8046_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_8046_squeeze_mask_0 = const()[name = tensor("op_8046_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_8046_cast_fp16 = slice_by_index(begin = var_8046_begin_0, end = var_8046_end_0, end_mask = var_8046_end_mask_0, squeeze_mask = var_8046_squeeze_mask_0, x = k_complex_41_cast_fp16)[name = tensor("op_8046_cast_fp16")]; + tensor freqs_41_to_fp16 = const()[name = tensor("freqs_41_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(510920832)))]; + tensor var_8052_cast_fp16 = mul(x = freqs_41_to_fp16, y = ts_125_cast_fp16)[name = tensor("op_8052_cast_fp16")]; + tensor rotr_41_cast_fp16 = cos(x = var_8052_cast_fp16)[name = tensor("rotr_41_cast_fp16")]; + tensor roti_41_cast_fp16 = sin(x = var_8052_cast_fp16)[name = tensor("roti_41_cast_fp16")]; + tensor var_8056_cast_fp16 = mul(x = var_8022_cast_fp16, y = rotr_41_cast_fp16)[name = tensor("op_8056_cast_fp16")]; + tensor var_8057_cast_fp16 = mul(x = var_8030_cast_fp16, y = roti_41_cast_fp16)[name = tensor("op_8057_cast_fp16")]; + tensor qor_81_cast_fp16 = sub(x = var_8056_cast_fp16, y = var_8057_cast_fp16)[name = tensor("qor_81_cast_fp16")]; + tensor var_8060_cast_fp16 = mul(x = var_8022_cast_fp16, y = roti_41_cast_fp16)[name = tensor("op_8060_cast_fp16")]; + tensor var_8061_cast_fp16 = mul(x = var_8030_cast_fp16, y = rotr_41_cast_fp16)[name = tensor("op_8061_cast_fp16")]; + tensor qoi_81_cast_fp16 = add(x = var_8060_cast_fp16, y = var_8061_cast_fp16)[name = tensor("qoi_81_cast_fp16")]; + tensor var_8064_cast_fp16 = mul(x = var_8038_cast_fp16, y = rotr_41_cast_fp16)[name = tensor("op_8064_cast_fp16")]; + tensor var_8065_cast_fp16 = mul(x = var_8046_cast_fp16, y = roti_41_cast_fp16)[name = tensor("op_8065_cast_fp16")]; + tensor kor_81_cast_fp16 = sub(x = var_8064_cast_fp16, y = var_8065_cast_fp16)[name = tensor("kor_81_cast_fp16")]; + tensor var_8068_cast_fp16 = mul(x = var_8038_cast_fp16, y = roti_41_cast_fp16)[name = tensor("op_8068_cast_fp16")]; + tensor var_8069_cast_fp16 = mul(x = var_8046_cast_fp16, y = rotr_41_cast_fp16)[name = tensor("op_8069_cast_fp16")]; + tensor koi_81_cast_fp16 = add(x = var_8068_cast_fp16, y = var_8069_cast_fp16)[name = tensor("koi_81_cast_fp16")]; + tensor qo_41_axis_0 = const()[name = tensor("qo_41_axis_0"), val = tensor(-1)]; + tensor qo_41_cast_fp16 = stack(axis = qo_41_axis_0, values = (qor_81_cast_fp16, qoi_81_cast_fp16))[name = tensor("qo_41_cast_fp16")]; + tensor ko_41_axis_0 = const()[name = tensor("ko_41_axis_0"), val = tensor(-1)]; + tensor ko_41_cast_fp16 = stack(axis = ko_41_axis_0, values = (kor_81_cast_fp16, koi_81_cast_fp16))[name = tensor("ko_41_cast_fp16")]; + tensor var_8102 = const()[name = tensor("op_8102"), val = tensor([1, 1, 16, 64])]; + tensor q_123_cast_fp16 = reshape(shape = var_8102, x = qo_41_cast_fp16)[name = tensor("q_123_cast_fp16")]; + tensor var_8108 = const()[name = tensor("op_8108"), val = tensor([1, 1, 16, 64])]; + tensor k_83_cast_fp16 = reshape(shape = var_8108, x = ko_41_cast_fp16)[name = tensor("k_83_cast_fp16")]; + tensor _inversed_8133_y_0_to_fp16 = const()[name = tensor("_inversed_8133_y_0_to_fp16"), val = tensor(0x1p-9)]; + tensor _inversed_8133_cast_fp16 = mul(x = ts_125_cast_fp16, y = _inversed_8133_y_0_to_fp16)[name = tensor("_inversed_8133_cast_fp16")]; + tensor var_8134_cast_fp16 = floor(x = _inversed_8133_cast_fp16)[name = tensor("op_8134_cast_fp16")]; + tensor var_8135_to_fp16 = const()[name = tensor("op_8135_to_fp16"), val = tensor(0x1p+9)]; + tensor var_8136_cast_fp16 = mul(x = var_8134_cast_fp16, y = var_8135_to_fp16)[name = tensor("op_8136_cast_fp16")]; + tensor write_indices_float_83_cast_fp16 = sub(x = ts_125_cast_fp16, y = var_8136_cast_fp16)[name = tensor("write_indices_float_83_cast_fp16")]; + tensor var_8143_dtype_0 = const()[name = tensor("op_8143_dtype_0"), val = tensor("int32")]; + tensor write_indices_41_reps_0 = const()[name = tensor("write_indices_41_reps_0"), val = tensor([1, 1, 16, 64])]; + tensor write_indices_float_83_cast_fp16_to_int32 = cast(dtype = var_8143_dtype_0, x = write_indices_float_83_cast_fp16)[name = tensor("cast_452")]; + tensor write_indices_41 = tile(reps = write_indices_41_reps_0, x = write_indices_float_83_cast_fp16_to_int32)[name = tensor("write_indices_41")]; + tensor var_8153_begin_0 = const()[name = tensor("op_8153_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_8153_end_0 = const()[name = tensor("op_8153_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor var_8153_end_mask_0 = const()[name = tensor("op_8153_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_8153_squeeze_mask_0 = const()[name = tensor("op_8153_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor cache20_to_fp16_dtype_0 = const()[name = tensor("cache20_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor cache20_to_fp16 = cast(dtype = cache20_to_fp16_dtype_0, x = cache20)[name = tensor("cast_451")]; + tensor var_8153_cast_fp16 = slice_by_index(begin = var_8153_begin_0, end = var_8153_end_0, end_mask = var_8153_end_mask_0, squeeze_mask = var_8153_squeeze_mask_0, x = cache20_to_fp16)[name = tensor("op_8153_cast_fp16")]; + tensor var_8155_axis_0 = const()[name = tensor("op_8155_axis_0"), val = tensor(1)]; + tensor var_8155_mode_0 = const()[name = tensor("op_8155_mode_0"), val = tensor("update")]; + tensor var_8155_validate_indices_0 = const()[name = tensor("op_8155_validate_indices_0"), val = tensor(false)]; + tensor var_8155_cast_fp16 = scatter_along_axis(axis = var_8155_axis_0, data = var_8153_cast_fp16, indices = write_indices_41, mode = var_8155_mode_0, updates = k_83_cast_fp16, validate_indices = var_8155_validate_indices_0)[name = tensor("op_8155_cast_fp16")]; + tensor concat_142 = const()[name = tensor("concat_142"), val = tensor([0, 0, 0, 0, 0])]; + tensor concat_143 = const()[name = tensor("concat_143"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_41_internal_tensor_assign_1_stride_0 = const()[name = tensor("new_cache_41_internal_tensor_assign_1_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_41_internal_tensor_assign_1_begin_mask_0 = const()[name = tensor("new_cache_41_internal_tensor_assign_1_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_41_internal_tensor_assign_1_end_mask_0 = const()[name = tensor("new_cache_41_internal_tensor_assign_1_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_41_internal_tensor_assign_1_squeeze_mask_0 = const()[name = tensor("new_cache_41_internal_tensor_assign_1_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_88 = const()[name = tensor("shape_88"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_40 = const()[name = tensor("reduce_prod_40"), val = tensor(1048576)]; + tensor range_1d_40_start_0 = const()[name = tensor("range_1d_40_start_0"), val = tensor(0)]; + tensor range_1d_40_step_0 = const()[name = tensor("range_1d_40_step_0"), val = tensor(1)]; + tensor range_1d_40 = range_1d(end = reduce_prod_40, start = range_1d_40_start_0, step = range_1d_40_step_0)[name = tensor("range_1d_40")]; + tensor reshape_200 = reshape(shape = shape_88, x = range_1d_40)[name = tensor("reshape_200")]; + tensor slice_by_index_40 = slice_by_index(begin = concat_142, begin_mask = new_cache_41_internal_tensor_assign_1_begin_mask_0, end = concat_143, end_mask = new_cache_41_internal_tensor_assign_1_end_mask_0, squeeze_mask = new_cache_41_internal_tensor_assign_1_squeeze_mask_0, stride = new_cache_41_internal_tensor_assign_1_stride_0, x = reshape_200)[name = tensor("slice_by_index_40")]; + tensor reshape_201_shape_0 = const()[name = tensor("reshape_201_shape_0"), val = tensor([-1])]; + tensor reshape_201 = reshape(shape = reshape_201_shape_0, x = slice_by_index_40)[name = tensor("reshape_201")]; + tensor reshape_202_shape_0 = const()[name = tensor("reshape_202_shape_0"), val = tensor([-1])]; + tensor reshape_202_cast_fp16 = reshape(shape = reshape_202_shape_0, x = var_8155_cast_fp16)[name = tensor("reshape_202_cast_fp16")]; + tensor reshape_203_shape_0 = const()[name = tensor("reshape_203_shape_0"), val = tensor([-1])]; + tensor reshape_203_cast_fp16 = reshape(shape = reshape_203_shape_0, x = cache20_to_fp16)[name = tensor("reshape_203_cast_fp16")]; + tensor scatter_40_mode_0 = const()[name = tensor("scatter_40_mode_0"), val = tensor("update")]; + tensor scatter_40_axis_0 = const()[name = tensor("scatter_40_axis_0"), val = tensor(0)]; + tensor scatter_40_validate_indices_0 = const()[name = tensor("scatter_40_validate_indices_0"), val = tensor(false)]; + tensor scatter_40_cast_fp16 = scatter(axis = scatter_40_axis_0, data = reshape_203_cast_fp16, indices = reshape_201, mode = scatter_40_mode_0, updates = reshape_202_cast_fp16, validate_indices = scatter_40_validate_indices_0)[name = tensor("scatter_40_cast_fp16")]; + tensor reshape_204_cast_fp16 = reshape(shape = shape_88, x = scatter_40_cast_fp16)[name = tensor("reshape_204_cast_fp16")]; + tensor var_8163_begin_0 = const()[name = tensor("op_8163_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor var_8163_end_0 = const()[name = tensor("op_8163_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor var_8163_end_mask_0 = const()[name = tensor("op_8163_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_8163_squeeze_mask_0 = const()[name = tensor("op_8163_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor var_8163_cast_fp16 = slice_by_index(begin = var_8163_begin_0, end = var_8163_end_0, end_mask = var_8163_end_mask_0, squeeze_mask = var_8163_squeeze_mask_0, x = reshape_204_cast_fp16)[name = tensor("op_8163_cast_fp16")]; + tensor var_8165_axis_0 = const()[name = tensor("op_8165_axis_0"), val = tensor(1)]; + tensor var_8165_mode_0 = const()[name = tensor("op_8165_mode_0"), val = tensor("update")]; + tensor var_8165_validate_indices_0 = const()[name = tensor("op_8165_validate_indices_0"), val = tensor(false)]; + tensor var_8165_cast_fp16 = scatter_along_axis(axis = var_8165_axis_0, data = var_8163_cast_fp16, indices = write_indices_41, mode = var_8165_mode_0, updates = v_41_cast_fp16, validate_indices = var_8165_validate_indices_0)[name = tensor("op_8165_cast_fp16")]; + tensor concat_144 = const()[name = tensor("concat_144"), val = tensor([1, 0, 0, 0, 0])]; + tensor concat_145 = const()[name = tensor("concat_145"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_41_internal_tensor_assign_2_stride_0 = const()[name = tensor("new_cache_41_internal_tensor_assign_2_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_41_internal_tensor_assign_2_begin_mask_0 = const()[name = tensor("new_cache_41_internal_tensor_assign_2_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_41_internal_tensor_assign_2_end_mask_0 = const()[name = tensor("new_cache_41_internal_tensor_assign_2_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_41_internal_tensor_assign_2_squeeze_mask_0 = const()[name = tensor("new_cache_41_internal_tensor_assign_2_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_89 = const()[name = tensor("shape_89"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_41 = const()[name = tensor("reduce_prod_41"), val = tensor(1048576)]; + tensor range_1d_41_start_0 = const()[name = tensor("range_1d_41_start_0"), val = tensor(0)]; + tensor range_1d_41_step_0 = const()[name = tensor("range_1d_41_step_0"), val = tensor(1)]; + tensor range_1d_41 = range_1d(end = reduce_prod_41, start = range_1d_41_start_0, step = range_1d_41_step_0)[name = tensor("range_1d_41")]; + tensor reshape_205 = reshape(shape = shape_89, x = range_1d_41)[name = tensor("reshape_205")]; + tensor slice_by_index_41 = slice_by_index(begin = concat_144, begin_mask = new_cache_41_internal_tensor_assign_2_begin_mask_0, end = concat_145, end_mask = new_cache_41_internal_tensor_assign_2_end_mask_0, squeeze_mask = new_cache_41_internal_tensor_assign_2_squeeze_mask_0, stride = new_cache_41_internal_tensor_assign_2_stride_0, x = reshape_205)[name = tensor("slice_by_index_41")]; + tensor reshape_206_shape_0 = const()[name = tensor("reshape_206_shape_0"), val = tensor([-1])]; + tensor reshape_206 = reshape(shape = reshape_206_shape_0, x = slice_by_index_41)[name = tensor("reshape_206")]; + tensor reshape_207_shape_0 = const()[name = tensor("reshape_207_shape_0"), val = tensor([-1])]; + tensor reshape_207_cast_fp16 = reshape(shape = reshape_207_shape_0, x = var_8165_cast_fp16)[name = tensor("reshape_207_cast_fp16")]; + tensor reshape_208_shape_0 = const()[name = tensor("reshape_208_shape_0"), val = tensor([-1])]; + tensor reshape_208_cast_fp16 = reshape(shape = reshape_208_shape_0, x = reshape_204_cast_fp16)[name = tensor("reshape_208_cast_fp16")]; + tensor scatter_41_mode_0 = const()[name = tensor("scatter_41_mode_0"), val = tensor("update")]; + tensor scatter_41_axis_0 = const()[name = tensor("scatter_41_axis_0"), val = tensor(0)]; + tensor scatter_41_validate_indices_0 = const()[name = tensor("scatter_41_validate_indices_0"), val = tensor(false)]; + tensor scatter_41_cast_fp16 = scatter(axis = scatter_41_axis_0, data = reshape_208_cast_fp16, indices = reshape_206, mode = scatter_41_mode_0, updates = reshape_207_cast_fp16, validate_indices = scatter_41_validate_indices_0)[name = tensor("scatter_41_cast_fp16")]; + tensor reshape_209_cast_fp16 = reshape(shape = shape_89, x = scatter_41_cast_fp16)[name = tensor("reshape_209_cast_fp16")]; + tensor reshape_209_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("reshape_209_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor keys_121_begin_0 = const()[name = tensor("keys_121_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor keys_121_end_0 = const()[name = tensor("keys_121_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor keys_121_end_mask_0 = const()[name = tensor("keys_121_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor keys_121_squeeze_mask_0 = const()[name = tensor("keys_121_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor keys_121_cast_fp16 = slice_by_index(begin = keys_121_begin_0, end = keys_121_end_0, end_mask = keys_121_end_mask_0, squeeze_mask = keys_121_squeeze_mask_0, x = reshape_209_cast_fp16)[name = tensor("keys_121_cast_fp16")]; + tensor values_121_begin_0 = const()[name = tensor("values_121_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor values_121_end_0 = const()[name = tensor("values_121_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor values_121_end_mask_0 = const()[name = tensor("values_121_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor values_121_squeeze_mask_0 = const()[name = tensor("values_121_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor values_121_cast_fp16 = slice_by_index(begin = values_121_begin_0, end = values_121_end_0, end_mask = values_121_end_mask_0, squeeze_mask = values_121_squeeze_mask_0, x = reshape_209_cast_fp16)[name = tensor("values_121_cast_fp16")]; + tensor var_8177_cast_fp16 = not_equal(x = keys_121_cast_fp16, y = keys_121_cast_fp16)[name = tensor("op_8177_cast_fp16")]; + tensor keys_123_cast_fp16 = select(a = var_463_to_fp16, b = keys_121_cast_fp16, cond = var_8177_cast_fp16)[name = tensor("keys_123_cast_fp16")]; + tensor var_8185_cast_fp16 = not_equal(x = values_121_cast_fp16, y = values_121_cast_fp16)[name = tensor("op_8185_cast_fp16")]; + tensor values_123_cast_fp16 = select(a = var_463_to_fp16, b = values_121_cast_fp16, cond = var_8185_cast_fp16)[name = tensor("values_123_cast_fp16")]; + tensor var_8209 = const()[name = tensor("op_8209"), val = tensor([0, 2, 1, 3])]; + tensor var_8225 = const()[name = tensor("op_8225"), val = tensor([1, 1, 1])]; + tensor var_8226_cast_fp16 = reshape(shape = var_8225, x = position20_to_fp16)[name = tensor("op_8226_cast_fp16")]; + tensor var_8246_to_fp16 = const()[name = tensor("op_8246_to_fp16"), val = tensor(0x1p+0)]; + tensor valid_len_41_cast_fp16 = add(x = var_8226_cast_fp16, y = var_8246_to_fp16)[name = tensor("valid_len_41_cast_fp16")]; + tensor valid_mask_41_cast_fp16 = less(x = k_positions_1_promoted_to_fp16, y = valid_len_41_cast_fp16)[name = tensor("valid_mask_41_cast_fp16")]; + tensor causal_mask_41_cast_fp16 = less_equal(x = k_positions_1_promoted_to_fp16, y = var_8226_cast_fp16)[name = tensor("causal_mask_41_cast_fp16")]; + tensor attn_mask_81 = logical_and(x = valid_mask_41_cast_fp16, y = causal_mask_41_cast_fp16)[name = tensor("attn_mask_81")]; + tensor attn_mask_83_axes_0 = const()[name = tensor("attn_mask_83_axes_0"), val = tensor([1])]; + tensor attn_mask_83 = expand_dims(axes = attn_mask_83_axes_0, x = attn_mask_81)[name = tensor("attn_mask_83")]; + tensor var_8264_transpose_x_0 = const()[name = tensor("op_8264_transpose_x_0"), val = tensor(false)]; + tensor var_8264_transpose_y_0 = const()[name = tensor("op_8264_transpose_y_0"), val = tensor(false)]; + tensor transpose_112_perm_0 = const()[name = tensor("transpose_112_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_113_perm_0 = const()[name = tensor("transpose_113_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_113 = transpose(perm = transpose_113_perm_0, x = keys_123_cast_fp16)[name = tensor("transpose_133")]; + tensor transpose_112 = transpose(perm = transpose_112_perm_0, x = q_123_cast_fp16)[name = tensor("transpose_134")]; + tensor var_8264_cast_fp16 = matmul(transpose_x = var_8264_transpose_x_0, transpose_y = var_8264_transpose_y_0, x = transpose_112, y = transpose_113)[name = tensor("op_8264_cast_fp16")]; + tensor var_8258_to_fp16 = const()[name = tensor("op_8258_to_fp16"), val = tensor([0x1p-3])]; + tensor attn_weights_121_cast_fp16 = mul(x = var_8264_cast_fp16, y = var_8258_to_fp16)[name = tensor("attn_weights_121_cast_fp16")]; + tensor var_8266 = logical_not(x = attn_mask_83)[name = tensor("op_8266")]; + tensor var_8267_to_fp16 = const()[name = tensor("op_8267_to_fp16"), val = tensor(-inf)]; + tensor attn_weights_123_cast_fp16 = select(a = var_8267_to_fp16, b = attn_weights_121_cast_fp16, cond = var_8266)[name = tensor("attn_weights_123_cast_fp16")]; + tensor var_8269 = const()[name = tensor("op_8269"), val = tensor(-1)]; + tensor attn_weights_125_cast_fp16 = softmax(axis = var_8269, x = attn_weights_123_cast_fp16)[name = tensor("attn_weights_125_cast_fp16")]; + tensor attn_output_41_transpose_x_0 = const()[name = tensor("attn_output_41_transpose_x_0"), val = tensor(false)]; + tensor attn_output_41_transpose_y_0 = const()[name = tensor("attn_output_41_transpose_y_0"), val = tensor(false)]; + tensor values_125_cast_fp16 = transpose(perm = var_8209, x = values_123_cast_fp16)[name = tensor("transpose_135")]; + tensor attn_output_41_cast_fp16 = matmul(transpose_x = attn_output_41_transpose_x_0, transpose_y = attn_output_41_transpose_y_0, x = attn_weights_125_cast_fp16, y = values_125_cast_fp16)[name = tensor("attn_output_41_cast_fp16")]; + tensor var_8277 = const()[name = tensor("op_8277"), val = tensor([0, 2, 1, 3])]; + tensor var_8282 = const()[name = tensor("op_8282"), val = tensor([1, 1, 1024])]; + tensor var_8278_cast_fp16 = transpose(perm = var_8277, x = attn_output_41_cast_fp16)[name = tensor("transpose_132")]; + tensor input_247_cast_fp16 = reshape(shape = var_8282, x = var_8278_cast_fp16)[name = tensor("input_247_cast_fp16")]; + tensor attn20_out_proj_weight_to_fp16 = const()[name = tensor("attn20_out_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(510920960)))]; + tensor linear_82_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = attn20_out_proj_weight_to_fp16, x = input_247_cast_fp16)[name = tensor("linear_82_cast_fp16")]; + tensor var_8288_to_fp16 = const()[name = tensor("op_8288_to_fp16"), val = tensor(0x1p+0)]; + tensor var_8289_cast_fp16 = add(x = position20_to_fp16, y = var_8288_to_fp16)[name = tensor("op_8289_cast_fp16")]; + tensor var_8289_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_8289_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor input_249_cast_fp16 = add(x = input_243_cast_fp16, y = linear_82_cast_fp16)[name = tensor("input_249_cast_fp16")]; + tensor input_251_axes_0 = const()[name = tensor("input_251_axes_0"), val = tensor([-1])]; + tensor norm20_2_weight_to_fp16 = const()[name = tensor("norm20_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(513018176)))]; + tensor norm20_2_bias_to_fp16 = const()[name = tensor("norm20_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(513020288)))]; + tensor var_8293_to_fp16 = const()[name = tensor("op_8293_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_251_cast_fp16 = layer_norm(axes = input_251_axes_0, beta = norm20_2_bias_to_fp16, epsilon = var_8293_to_fp16, gamma = norm20_2_weight_to_fp16, x = input_249_cast_fp16)[name = tensor("input_251_cast_fp16")]; + tensor linear20_1_weight_to_fp16 = const()[name = tensor("linear20_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(513022400)))]; + tensor linear_83_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = linear20_1_weight_to_fp16, x = input_251_cast_fp16)[name = tensor("linear_83_cast_fp16")]; + tensor input_253_mode_0 = const()[name = tensor("input_253_mode_0"), val = tensor("EXACT")]; + tensor input_253_cast_fp16 = gelu(mode = input_253_mode_0, x = linear_83_cast_fp16)[name = tensor("input_253_cast_fp16")]; + tensor linear20_2_weight_to_fp16 = const()[name = tensor("linear20_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(521411072)))]; + tensor linear_84_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = linear20_2_weight_to_fp16, x = input_253_cast_fp16)[name = tensor("linear_84_cast_fp16")]; + tensor input_255_cast_fp16 = add(x = input_249_cast_fp16, y = linear_84_cast_fp16)[name = tensor("input_255_cast_fp16")]; + tensor input_257_axes_0 = const()[name = tensor("input_257_axes_0"), val = tensor([-1])]; + tensor norm21_1_weight_to_fp16 = const()[name = tensor("norm21_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(529799744)))]; + tensor norm21_1_bias_to_fp16 = const()[name = tensor("norm21_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(529801856)))]; + tensor var_8310_to_fp16 = const()[name = tensor("op_8310_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_257_cast_fp16 = layer_norm(axes = input_257_axes_0, beta = norm21_1_bias_to_fp16, epsilon = var_8310_to_fp16, gamma = norm21_1_weight_to_fp16, x = input_255_cast_fp16)[name = tensor("input_257_cast_fp16")]; + tensor attn21_in_proj_weight_to_fp16 = const()[name = tensor("attn21_in_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(529803968)))]; + tensor linear_85_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = attn21_in_proj_weight_to_fp16, x = input_257_cast_fp16)[name = tensor("linear_85_cast_fp16")]; + tensor var_8324 = const()[name = tensor("op_8324"), val = tensor([1, 1, 3, 16, 64])]; + tensor qkv_43_cast_fp16 = reshape(shape = var_8324, x = linear_85_cast_fp16)[name = tensor("qkv_43_cast_fp16")]; + tensor q_127_begin_0 = const()[name = tensor("q_127_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor q_127_end_0 = const()[name = tensor("q_127_end_0"), val = tensor([1, 1, 1, 16, 64])]; + tensor q_127_end_mask_0 = const()[name = tensor("q_127_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor q_127_squeeze_mask_0 = const()[name = tensor("q_127_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor q_127_cast_fp16 = slice_by_index(begin = q_127_begin_0, end = q_127_end_0, end_mask = q_127_end_mask_0, squeeze_mask = q_127_squeeze_mask_0, x = qkv_43_cast_fp16)[name = tensor("q_127_cast_fp16")]; + tensor k_85_begin_0 = const()[name = tensor("k_85_begin_0"), val = tensor([0, 0, 1, 0, 0])]; + tensor k_85_end_0 = const()[name = tensor("k_85_end_0"), val = tensor([1, 1, 2, 16, 64])]; + tensor k_85_end_mask_0 = const()[name = tensor("k_85_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor k_85_squeeze_mask_0 = const()[name = tensor("k_85_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor k_85_cast_fp16 = slice_by_index(begin = k_85_begin_0, end = k_85_end_0, end_mask = k_85_end_mask_0, squeeze_mask = k_85_squeeze_mask_0, x = qkv_43_cast_fp16)[name = tensor("k_85_cast_fp16")]; + tensor v_43_begin_0 = const()[name = tensor("v_43_begin_0"), val = tensor([0, 0, 2, 0, 0])]; + tensor v_43_end_0 = const()[name = tensor("v_43_end_0"), val = tensor([1, 1, 3, 16, 64])]; + tensor v_43_end_mask_0 = const()[name = tensor("v_43_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor v_43_squeeze_mask_0 = const()[name = tensor("v_43_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor v_43_cast_fp16 = slice_by_index(begin = v_43_begin_0, end = v_43_end_0, end_mask = v_43_end_mask_0, squeeze_mask = v_43_squeeze_mask_0, x = qkv_43_cast_fp16)[name = tensor("v_43_cast_fp16")]; + tensor var_8390 = const()[name = tensor("op_8390"), val = tensor([1, 1, 1, 1])]; + tensor position21_to_fp16_dtype_0 = const()[name = tensor("position21_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor position21_to_fp16 = cast(dtype = position21_to_fp16_dtype_0, x = position21)[name = tensor("cast_448")]; + tensor ts_131_cast_fp16 = reshape(shape = var_8390, x = position21_to_fp16)[name = tensor("ts_131_cast_fp16")]; + tensor var_8397 = const()[name = tensor("op_8397"), val = tensor([1, 1, 16, 32, 2])]; + tensor q_complex_43_cast_fp16 = reshape(shape = var_8397, x = q_127_cast_fp16)[name = tensor("q_complex_43_cast_fp16")]; + tensor var_8404 = const()[name = tensor("op_8404"), val = tensor([1, 1, 16, 32, 2])]; + tensor k_complex_43_cast_fp16 = reshape(shape = var_8404, x = k_85_cast_fp16)[name = tensor("k_complex_43_cast_fp16")]; + tensor var_8408_begin_0 = const()[name = tensor("op_8408_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_8408_end_0 = const()[name = tensor("op_8408_end_0"), val = tensor([1, 1, 16, 32, 1])]; + tensor var_8408_end_mask_0 = const()[name = tensor("op_8408_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_8408_squeeze_mask_0 = const()[name = tensor("op_8408_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_8408_cast_fp16 = slice_by_index(begin = var_8408_begin_0, end = var_8408_end_0, end_mask = var_8408_end_mask_0, squeeze_mask = var_8408_squeeze_mask_0, x = q_complex_43_cast_fp16)[name = tensor("op_8408_cast_fp16")]; + tensor var_8416_begin_0 = const()[name = tensor("op_8416_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_8416_end_0 = const()[name = tensor("op_8416_end_0"), val = tensor([1, 1, 16, 32, 2])]; + tensor var_8416_end_mask_0 = const()[name = tensor("op_8416_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_8416_squeeze_mask_0 = const()[name = tensor("op_8416_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_8416_cast_fp16 = slice_by_index(begin = var_8416_begin_0, end = var_8416_end_0, end_mask = var_8416_end_mask_0, squeeze_mask = var_8416_squeeze_mask_0, x = q_complex_43_cast_fp16)[name = tensor("op_8416_cast_fp16")]; + tensor var_8424_begin_0 = const()[name = tensor("op_8424_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_8424_end_0 = const()[name = tensor("op_8424_end_0"), val = tensor([1, 1, 16, 32, 1])]; + tensor var_8424_end_mask_0 = const()[name = tensor("op_8424_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_8424_squeeze_mask_0 = const()[name = tensor("op_8424_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_8424_cast_fp16 = slice_by_index(begin = var_8424_begin_0, end = var_8424_end_0, end_mask = var_8424_end_mask_0, squeeze_mask = var_8424_squeeze_mask_0, x = k_complex_43_cast_fp16)[name = tensor("op_8424_cast_fp16")]; + tensor var_8432_begin_0 = const()[name = tensor("op_8432_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_8432_end_0 = const()[name = tensor("op_8432_end_0"), val = tensor([1, 1, 16, 32, 2])]; + tensor var_8432_end_mask_0 = const()[name = tensor("op_8432_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_8432_squeeze_mask_0 = const()[name = tensor("op_8432_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_8432_cast_fp16 = slice_by_index(begin = var_8432_begin_0, end = var_8432_end_0, end_mask = var_8432_end_mask_0, squeeze_mask = var_8432_squeeze_mask_0, x = k_complex_43_cast_fp16)[name = tensor("op_8432_cast_fp16")]; + tensor freqs_43_to_fp16 = const()[name = tensor("freqs_43_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(536095488)))]; + tensor var_8438_cast_fp16 = mul(x = freqs_43_to_fp16, y = ts_131_cast_fp16)[name = tensor("op_8438_cast_fp16")]; + tensor rotr_43_cast_fp16 = cos(x = var_8438_cast_fp16)[name = tensor("rotr_43_cast_fp16")]; + tensor roti_43_cast_fp16 = sin(x = var_8438_cast_fp16)[name = tensor("roti_43_cast_fp16")]; + tensor var_8442_cast_fp16 = mul(x = var_8408_cast_fp16, y = rotr_43_cast_fp16)[name = tensor("op_8442_cast_fp16")]; + tensor var_8443_cast_fp16 = mul(x = var_8416_cast_fp16, y = roti_43_cast_fp16)[name = tensor("op_8443_cast_fp16")]; + tensor qor_85_cast_fp16 = sub(x = var_8442_cast_fp16, y = var_8443_cast_fp16)[name = tensor("qor_85_cast_fp16")]; + tensor var_8446_cast_fp16 = mul(x = var_8408_cast_fp16, y = roti_43_cast_fp16)[name = tensor("op_8446_cast_fp16")]; + tensor var_8447_cast_fp16 = mul(x = var_8416_cast_fp16, y = rotr_43_cast_fp16)[name = tensor("op_8447_cast_fp16")]; + tensor qoi_85_cast_fp16 = add(x = var_8446_cast_fp16, y = var_8447_cast_fp16)[name = tensor("qoi_85_cast_fp16")]; + tensor var_8450_cast_fp16 = mul(x = var_8424_cast_fp16, y = rotr_43_cast_fp16)[name = tensor("op_8450_cast_fp16")]; + tensor var_8451_cast_fp16 = mul(x = var_8432_cast_fp16, y = roti_43_cast_fp16)[name = tensor("op_8451_cast_fp16")]; + tensor kor_85_cast_fp16 = sub(x = var_8450_cast_fp16, y = var_8451_cast_fp16)[name = tensor("kor_85_cast_fp16")]; + tensor var_8454_cast_fp16 = mul(x = var_8424_cast_fp16, y = roti_43_cast_fp16)[name = tensor("op_8454_cast_fp16")]; + tensor var_8455_cast_fp16 = mul(x = var_8432_cast_fp16, y = rotr_43_cast_fp16)[name = tensor("op_8455_cast_fp16")]; + tensor koi_85_cast_fp16 = add(x = var_8454_cast_fp16, y = var_8455_cast_fp16)[name = tensor("koi_85_cast_fp16")]; + tensor qo_43_axis_0 = const()[name = tensor("qo_43_axis_0"), val = tensor(-1)]; + tensor qo_43_cast_fp16 = stack(axis = qo_43_axis_0, values = (qor_85_cast_fp16, qoi_85_cast_fp16))[name = tensor("qo_43_cast_fp16")]; + tensor ko_43_axis_0 = const()[name = tensor("ko_43_axis_0"), val = tensor(-1)]; + tensor ko_43_cast_fp16 = stack(axis = ko_43_axis_0, values = (kor_85_cast_fp16, koi_85_cast_fp16))[name = tensor("ko_43_cast_fp16")]; + tensor var_8488 = const()[name = tensor("op_8488"), val = tensor([1, 1, 16, 64])]; + tensor q_129_cast_fp16 = reshape(shape = var_8488, x = qo_43_cast_fp16)[name = tensor("q_129_cast_fp16")]; + tensor var_8494 = const()[name = tensor("op_8494"), val = tensor([1, 1, 16, 64])]; + tensor k_87_cast_fp16 = reshape(shape = var_8494, x = ko_43_cast_fp16)[name = tensor("k_87_cast_fp16")]; + tensor _inversed_8519_y_0_to_fp16 = const()[name = tensor("_inversed_8519_y_0_to_fp16"), val = tensor(0x1p-9)]; + tensor _inversed_8519_cast_fp16 = mul(x = ts_131_cast_fp16, y = _inversed_8519_y_0_to_fp16)[name = tensor("_inversed_8519_cast_fp16")]; + tensor var_8520_cast_fp16 = floor(x = _inversed_8519_cast_fp16)[name = tensor("op_8520_cast_fp16")]; + tensor var_8521_to_fp16 = const()[name = tensor("op_8521_to_fp16"), val = tensor(0x1p+9)]; + tensor var_8522_cast_fp16 = mul(x = var_8520_cast_fp16, y = var_8521_to_fp16)[name = tensor("op_8522_cast_fp16")]; + tensor write_indices_float_87_cast_fp16 = sub(x = ts_131_cast_fp16, y = var_8522_cast_fp16)[name = tensor("write_indices_float_87_cast_fp16")]; + tensor var_8529_dtype_0 = const()[name = tensor("op_8529_dtype_0"), val = tensor("int32")]; + tensor write_indices_43_reps_0 = const()[name = tensor("write_indices_43_reps_0"), val = tensor([1, 1, 16, 64])]; + tensor write_indices_float_87_cast_fp16_to_int32 = cast(dtype = var_8529_dtype_0, x = write_indices_float_87_cast_fp16)[name = tensor("cast_447")]; + tensor write_indices_43 = tile(reps = write_indices_43_reps_0, x = write_indices_float_87_cast_fp16_to_int32)[name = tensor("write_indices_43")]; + tensor var_8539_begin_0 = const()[name = tensor("op_8539_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_8539_end_0 = const()[name = tensor("op_8539_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor var_8539_end_mask_0 = const()[name = tensor("op_8539_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_8539_squeeze_mask_0 = const()[name = tensor("op_8539_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor cache21_to_fp16_dtype_0 = const()[name = tensor("cache21_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor cache21_to_fp16 = cast(dtype = cache21_to_fp16_dtype_0, x = cache21)[name = tensor("cast_446")]; + tensor var_8539_cast_fp16 = slice_by_index(begin = var_8539_begin_0, end = var_8539_end_0, end_mask = var_8539_end_mask_0, squeeze_mask = var_8539_squeeze_mask_0, x = cache21_to_fp16)[name = tensor("op_8539_cast_fp16")]; + tensor var_8541_axis_0 = const()[name = tensor("op_8541_axis_0"), val = tensor(1)]; + tensor var_8541_mode_0 = const()[name = tensor("op_8541_mode_0"), val = tensor("update")]; + tensor var_8541_validate_indices_0 = const()[name = tensor("op_8541_validate_indices_0"), val = tensor(false)]; + tensor var_8541_cast_fp16 = scatter_along_axis(axis = var_8541_axis_0, data = var_8539_cast_fp16, indices = write_indices_43, mode = var_8541_mode_0, updates = k_87_cast_fp16, validate_indices = var_8541_validate_indices_0)[name = tensor("op_8541_cast_fp16")]; + tensor concat_149 = const()[name = tensor("concat_149"), val = tensor([0, 0, 0, 0, 0])]; + tensor concat_150 = const()[name = tensor("concat_150"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_43_internal_tensor_assign_1_stride_0 = const()[name = tensor("new_cache_43_internal_tensor_assign_1_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_43_internal_tensor_assign_1_begin_mask_0 = const()[name = tensor("new_cache_43_internal_tensor_assign_1_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_43_internal_tensor_assign_1_end_mask_0 = const()[name = tensor("new_cache_43_internal_tensor_assign_1_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_43_internal_tensor_assign_1_squeeze_mask_0 = const()[name = tensor("new_cache_43_internal_tensor_assign_1_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_90 = const()[name = tensor("shape_90"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_42 = const()[name = tensor("reduce_prod_42"), val = tensor(1048576)]; + tensor range_1d_42_start_0 = const()[name = tensor("range_1d_42_start_0"), val = tensor(0)]; + tensor range_1d_42_step_0 = const()[name = tensor("range_1d_42_step_0"), val = tensor(1)]; + tensor range_1d_42 = range_1d(end = reduce_prod_42, start = range_1d_42_start_0, step = range_1d_42_step_0)[name = tensor("range_1d_42")]; + tensor reshape_210 = reshape(shape = shape_90, x = range_1d_42)[name = tensor("reshape_210")]; + tensor slice_by_index_42 = slice_by_index(begin = concat_149, begin_mask = new_cache_43_internal_tensor_assign_1_begin_mask_0, end = concat_150, end_mask = new_cache_43_internal_tensor_assign_1_end_mask_0, squeeze_mask = new_cache_43_internal_tensor_assign_1_squeeze_mask_0, stride = new_cache_43_internal_tensor_assign_1_stride_0, x = reshape_210)[name = tensor("slice_by_index_42")]; + tensor reshape_211_shape_0 = const()[name = tensor("reshape_211_shape_0"), val = tensor([-1])]; + tensor reshape_211 = reshape(shape = reshape_211_shape_0, x = slice_by_index_42)[name = tensor("reshape_211")]; + tensor reshape_212_shape_0 = const()[name = tensor("reshape_212_shape_0"), val = tensor([-1])]; + tensor reshape_212_cast_fp16 = reshape(shape = reshape_212_shape_0, x = var_8541_cast_fp16)[name = tensor("reshape_212_cast_fp16")]; + tensor reshape_213_shape_0 = const()[name = tensor("reshape_213_shape_0"), val = tensor([-1])]; + tensor reshape_213_cast_fp16 = reshape(shape = reshape_213_shape_0, x = cache21_to_fp16)[name = tensor("reshape_213_cast_fp16")]; + tensor scatter_42_mode_0 = const()[name = tensor("scatter_42_mode_0"), val = tensor("update")]; + tensor scatter_42_axis_0 = const()[name = tensor("scatter_42_axis_0"), val = tensor(0)]; + tensor scatter_42_validate_indices_0 = const()[name = tensor("scatter_42_validate_indices_0"), val = tensor(false)]; + tensor scatter_42_cast_fp16 = scatter(axis = scatter_42_axis_0, data = reshape_213_cast_fp16, indices = reshape_211, mode = scatter_42_mode_0, updates = reshape_212_cast_fp16, validate_indices = scatter_42_validate_indices_0)[name = tensor("scatter_42_cast_fp16")]; + tensor reshape_214_cast_fp16 = reshape(shape = shape_90, x = scatter_42_cast_fp16)[name = tensor("reshape_214_cast_fp16")]; + tensor var_8549_begin_0 = const()[name = tensor("op_8549_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor var_8549_end_0 = const()[name = tensor("op_8549_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor var_8549_end_mask_0 = const()[name = tensor("op_8549_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_8549_squeeze_mask_0 = const()[name = tensor("op_8549_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor var_8549_cast_fp16 = slice_by_index(begin = var_8549_begin_0, end = var_8549_end_0, end_mask = var_8549_end_mask_0, squeeze_mask = var_8549_squeeze_mask_0, x = reshape_214_cast_fp16)[name = tensor("op_8549_cast_fp16")]; + tensor var_8551_axis_0 = const()[name = tensor("op_8551_axis_0"), val = tensor(1)]; + tensor var_8551_mode_0 = const()[name = tensor("op_8551_mode_0"), val = tensor("update")]; + tensor var_8551_validate_indices_0 = const()[name = tensor("op_8551_validate_indices_0"), val = tensor(false)]; + tensor var_8551_cast_fp16 = scatter_along_axis(axis = var_8551_axis_0, data = var_8549_cast_fp16, indices = write_indices_43, mode = var_8551_mode_0, updates = v_43_cast_fp16, validate_indices = var_8551_validate_indices_0)[name = tensor("op_8551_cast_fp16")]; + tensor concat_151 = const()[name = tensor("concat_151"), val = tensor([1, 0, 0, 0, 0])]; + tensor concat_152 = const()[name = tensor("concat_152"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_43_internal_tensor_assign_2_stride_0 = const()[name = tensor("new_cache_43_internal_tensor_assign_2_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_43_internal_tensor_assign_2_begin_mask_0 = const()[name = tensor("new_cache_43_internal_tensor_assign_2_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_43_internal_tensor_assign_2_end_mask_0 = const()[name = tensor("new_cache_43_internal_tensor_assign_2_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_43_internal_tensor_assign_2_squeeze_mask_0 = const()[name = tensor("new_cache_43_internal_tensor_assign_2_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_91 = const()[name = tensor("shape_91"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_43 = const()[name = tensor("reduce_prod_43"), val = tensor(1048576)]; + tensor range_1d_43_start_0 = const()[name = tensor("range_1d_43_start_0"), val = tensor(0)]; + tensor range_1d_43_step_0 = const()[name = tensor("range_1d_43_step_0"), val = tensor(1)]; + tensor range_1d_43 = range_1d(end = reduce_prod_43, start = range_1d_43_start_0, step = range_1d_43_step_0)[name = tensor("range_1d_43")]; + tensor reshape_215 = reshape(shape = shape_91, x = range_1d_43)[name = tensor("reshape_215")]; + tensor slice_by_index_43 = slice_by_index(begin = concat_151, begin_mask = new_cache_43_internal_tensor_assign_2_begin_mask_0, end = concat_152, end_mask = new_cache_43_internal_tensor_assign_2_end_mask_0, squeeze_mask = new_cache_43_internal_tensor_assign_2_squeeze_mask_0, stride = new_cache_43_internal_tensor_assign_2_stride_0, x = reshape_215)[name = tensor("slice_by_index_43")]; + tensor reshape_216_shape_0 = const()[name = tensor("reshape_216_shape_0"), val = tensor([-1])]; + tensor reshape_216 = reshape(shape = reshape_216_shape_0, x = slice_by_index_43)[name = tensor("reshape_216")]; + tensor reshape_217_shape_0 = const()[name = tensor("reshape_217_shape_0"), val = tensor([-1])]; + tensor reshape_217_cast_fp16 = reshape(shape = reshape_217_shape_0, x = var_8551_cast_fp16)[name = tensor("reshape_217_cast_fp16")]; + tensor reshape_218_shape_0 = const()[name = tensor("reshape_218_shape_0"), val = tensor([-1])]; + tensor reshape_218_cast_fp16 = reshape(shape = reshape_218_shape_0, x = reshape_214_cast_fp16)[name = tensor("reshape_218_cast_fp16")]; + tensor scatter_43_mode_0 = const()[name = tensor("scatter_43_mode_0"), val = tensor("update")]; + tensor scatter_43_axis_0 = const()[name = tensor("scatter_43_axis_0"), val = tensor(0)]; + tensor scatter_43_validate_indices_0 = const()[name = tensor("scatter_43_validate_indices_0"), val = tensor(false)]; + tensor scatter_43_cast_fp16 = scatter(axis = scatter_43_axis_0, data = reshape_218_cast_fp16, indices = reshape_216, mode = scatter_43_mode_0, updates = reshape_217_cast_fp16, validate_indices = scatter_43_validate_indices_0)[name = tensor("scatter_43_cast_fp16")]; + tensor reshape_219_cast_fp16 = reshape(shape = shape_91, x = scatter_43_cast_fp16)[name = tensor("reshape_219_cast_fp16")]; + tensor reshape_219_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("reshape_219_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor keys_127_begin_0 = const()[name = tensor("keys_127_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor keys_127_end_0 = const()[name = tensor("keys_127_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor keys_127_end_mask_0 = const()[name = tensor("keys_127_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor keys_127_squeeze_mask_0 = const()[name = tensor("keys_127_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor keys_127_cast_fp16 = slice_by_index(begin = keys_127_begin_0, end = keys_127_end_0, end_mask = keys_127_end_mask_0, squeeze_mask = keys_127_squeeze_mask_0, x = reshape_219_cast_fp16)[name = tensor("keys_127_cast_fp16")]; + tensor values_127_begin_0 = const()[name = tensor("values_127_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor values_127_end_0 = const()[name = tensor("values_127_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor values_127_end_mask_0 = const()[name = tensor("values_127_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor values_127_squeeze_mask_0 = const()[name = tensor("values_127_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor values_127_cast_fp16 = slice_by_index(begin = values_127_begin_0, end = values_127_end_0, end_mask = values_127_end_mask_0, squeeze_mask = values_127_squeeze_mask_0, x = reshape_219_cast_fp16)[name = tensor("values_127_cast_fp16")]; + tensor var_8563_cast_fp16 = not_equal(x = keys_127_cast_fp16, y = keys_127_cast_fp16)[name = tensor("op_8563_cast_fp16")]; + tensor keys_129_cast_fp16 = select(a = var_463_to_fp16, b = keys_127_cast_fp16, cond = var_8563_cast_fp16)[name = tensor("keys_129_cast_fp16")]; + tensor var_8571_cast_fp16 = not_equal(x = values_127_cast_fp16, y = values_127_cast_fp16)[name = tensor("op_8571_cast_fp16")]; + tensor values_129_cast_fp16 = select(a = var_463_to_fp16, b = values_127_cast_fp16, cond = var_8571_cast_fp16)[name = tensor("values_129_cast_fp16")]; + tensor var_8595 = const()[name = tensor("op_8595"), val = tensor([0, 2, 1, 3])]; + tensor var_8611 = const()[name = tensor("op_8611"), val = tensor([1, 1, 1])]; + tensor var_8612_cast_fp16 = reshape(shape = var_8611, x = position21_to_fp16)[name = tensor("op_8612_cast_fp16")]; + tensor var_8632_to_fp16 = const()[name = tensor("op_8632_to_fp16"), val = tensor(0x1p+0)]; + tensor valid_len_43_cast_fp16 = add(x = var_8612_cast_fp16, y = var_8632_to_fp16)[name = tensor("valid_len_43_cast_fp16")]; + tensor valid_mask_43_cast_fp16 = less(x = k_positions_1_promoted_to_fp16, y = valid_len_43_cast_fp16)[name = tensor("valid_mask_43_cast_fp16")]; + tensor causal_mask_43_cast_fp16 = less_equal(x = k_positions_1_promoted_to_fp16, y = var_8612_cast_fp16)[name = tensor("causal_mask_43_cast_fp16")]; + tensor attn_mask_85 = logical_and(x = valid_mask_43_cast_fp16, y = causal_mask_43_cast_fp16)[name = tensor("attn_mask_85")]; + tensor attn_mask_87_axes_0 = const()[name = tensor("attn_mask_87_axes_0"), val = tensor([1])]; + tensor attn_mask_87 = expand_dims(axes = attn_mask_87_axes_0, x = attn_mask_85)[name = tensor("attn_mask_87")]; + tensor var_8650_transpose_x_0 = const()[name = tensor("op_8650_transpose_x_0"), val = tensor(false)]; + tensor var_8650_transpose_y_0 = const()[name = tensor("op_8650_transpose_y_0"), val = tensor(false)]; + tensor transpose_114_perm_0 = const()[name = tensor("transpose_114_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_115_perm_0 = const()[name = tensor("transpose_115_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_115 = transpose(perm = transpose_115_perm_0, x = keys_129_cast_fp16)[name = tensor("transpose_129")]; + tensor transpose_114 = transpose(perm = transpose_114_perm_0, x = q_129_cast_fp16)[name = tensor("transpose_130")]; + tensor var_8650_cast_fp16 = matmul(transpose_x = var_8650_transpose_x_0, transpose_y = var_8650_transpose_y_0, x = transpose_114, y = transpose_115)[name = tensor("op_8650_cast_fp16")]; + tensor var_8644_to_fp16 = const()[name = tensor("op_8644_to_fp16"), val = tensor([0x1p-3])]; + tensor attn_weights_127_cast_fp16 = mul(x = var_8650_cast_fp16, y = var_8644_to_fp16)[name = tensor("attn_weights_127_cast_fp16")]; + tensor var_8652 = logical_not(x = attn_mask_87)[name = tensor("op_8652")]; + tensor var_8653_to_fp16 = const()[name = tensor("op_8653_to_fp16"), val = tensor(-inf)]; + tensor attn_weights_129_cast_fp16 = select(a = var_8653_to_fp16, b = attn_weights_127_cast_fp16, cond = var_8652)[name = tensor("attn_weights_129_cast_fp16")]; + tensor var_8655 = const()[name = tensor("op_8655"), val = tensor(-1)]; + tensor attn_weights_131_cast_fp16 = softmax(axis = var_8655, x = attn_weights_129_cast_fp16)[name = tensor("attn_weights_131_cast_fp16")]; + tensor attn_output_43_transpose_x_0 = const()[name = tensor("attn_output_43_transpose_x_0"), val = tensor(false)]; + tensor attn_output_43_transpose_y_0 = const()[name = tensor("attn_output_43_transpose_y_0"), val = tensor(false)]; + tensor values_131_cast_fp16 = transpose(perm = var_8595, x = values_129_cast_fp16)[name = tensor("transpose_131")]; + tensor attn_output_43_cast_fp16 = matmul(transpose_x = attn_output_43_transpose_x_0, transpose_y = attn_output_43_transpose_y_0, x = attn_weights_131_cast_fp16, y = values_131_cast_fp16)[name = tensor("attn_output_43_cast_fp16")]; + tensor var_8663 = const()[name = tensor("op_8663"), val = tensor([0, 2, 1, 3])]; + tensor var_8668 = const()[name = tensor("op_8668"), val = tensor([1, 1, 1024])]; + tensor var_8664_cast_fp16 = transpose(perm = var_8663, x = attn_output_43_cast_fp16)[name = tensor("transpose_128")]; + tensor input_259_cast_fp16 = reshape(shape = var_8668, x = var_8664_cast_fp16)[name = tensor("input_259_cast_fp16")]; + tensor attn21_out_proj_weight_to_fp16 = const()[name = tensor("attn21_out_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(536095616)))]; + tensor linear_86_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = attn21_out_proj_weight_to_fp16, x = input_259_cast_fp16)[name = tensor("linear_86_cast_fp16")]; + tensor var_8674_to_fp16 = const()[name = tensor("op_8674_to_fp16"), val = tensor(0x1p+0)]; + tensor var_8675_cast_fp16 = add(x = position21_to_fp16, y = var_8674_to_fp16)[name = tensor("op_8675_cast_fp16")]; + tensor var_8675_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_8675_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor input_261_cast_fp16 = add(x = input_255_cast_fp16, y = linear_86_cast_fp16)[name = tensor("input_261_cast_fp16")]; + tensor input_263_axes_0 = const()[name = tensor("input_263_axes_0"), val = tensor([-1])]; + tensor norm21_2_weight_to_fp16 = const()[name = tensor("norm21_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(538192832)))]; + tensor norm21_2_bias_to_fp16 = const()[name = tensor("norm21_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(538194944)))]; + tensor var_8679_to_fp16 = const()[name = tensor("op_8679_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_263_cast_fp16 = layer_norm(axes = input_263_axes_0, beta = norm21_2_bias_to_fp16, epsilon = var_8679_to_fp16, gamma = norm21_2_weight_to_fp16, x = input_261_cast_fp16)[name = tensor("input_263_cast_fp16")]; + tensor linear21_1_weight_to_fp16 = const()[name = tensor("linear21_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(538197056)))]; + tensor linear_87_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = linear21_1_weight_to_fp16, x = input_263_cast_fp16)[name = tensor("linear_87_cast_fp16")]; + tensor input_265_mode_0 = const()[name = tensor("input_265_mode_0"), val = tensor("EXACT")]; + tensor input_265_cast_fp16 = gelu(mode = input_265_mode_0, x = linear_87_cast_fp16)[name = tensor("input_265_cast_fp16")]; + tensor linear21_2_weight_to_fp16 = const()[name = tensor("linear21_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(546585728)))]; + tensor linear_88_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = linear21_2_weight_to_fp16, x = input_265_cast_fp16)[name = tensor("linear_88_cast_fp16")]; + tensor input_267_cast_fp16 = add(x = input_261_cast_fp16, y = linear_88_cast_fp16)[name = tensor("input_267_cast_fp16")]; + tensor input_269_axes_0 = const()[name = tensor("input_269_axes_0"), val = tensor([-1])]; + tensor norm22_1_weight_to_fp16 = const()[name = tensor("norm22_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(554974400)))]; + tensor norm22_1_bias_to_fp16 = const()[name = tensor("norm22_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(554976512)))]; + tensor var_8696_to_fp16 = const()[name = tensor("op_8696_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_269_cast_fp16 = layer_norm(axes = input_269_axes_0, beta = norm22_1_bias_to_fp16, epsilon = var_8696_to_fp16, gamma = norm22_1_weight_to_fp16, x = input_267_cast_fp16)[name = tensor("input_269_cast_fp16")]; + tensor attn22_in_proj_weight_to_fp16 = const()[name = tensor("attn22_in_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(554978624)))]; + tensor linear_89_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = attn22_in_proj_weight_to_fp16, x = input_269_cast_fp16)[name = tensor("linear_89_cast_fp16")]; + tensor var_8710 = const()[name = tensor("op_8710"), val = tensor([1, 1, 3, 16, 64])]; + tensor qkv_45_cast_fp16 = reshape(shape = var_8710, x = linear_89_cast_fp16)[name = tensor("qkv_45_cast_fp16")]; + tensor q_133_begin_0 = const()[name = tensor("q_133_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor q_133_end_0 = const()[name = tensor("q_133_end_0"), val = tensor([1, 1, 1, 16, 64])]; + tensor q_133_end_mask_0 = const()[name = tensor("q_133_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor q_133_squeeze_mask_0 = const()[name = tensor("q_133_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor q_133_cast_fp16 = slice_by_index(begin = q_133_begin_0, end = q_133_end_0, end_mask = q_133_end_mask_0, squeeze_mask = q_133_squeeze_mask_0, x = qkv_45_cast_fp16)[name = tensor("q_133_cast_fp16")]; + tensor k_89_begin_0 = const()[name = tensor("k_89_begin_0"), val = tensor([0, 0, 1, 0, 0])]; + tensor k_89_end_0 = const()[name = tensor("k_89_end_0"), val = tensor([1, 1, 2, 16, 64])]; + tensor k_89_end_mask_0 = const()[name = tensor("k_89_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor k_89_squeeze_mask_0 = const()[name = tensor("k_89_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor k_89_cast_fp16 = slice_by_index(begin = k_89_begin_0, end = k_89_end_0, end_mask = k_89_end_mask_0, squeeze_mask = k_89_squeeze_mask_0, x = qkv_45_cast_fp16)[name = tensor("k_89_cast_fp16")]; + tensor v_45_begin_0 = const()[name = tensor("v_45_begin_0"), val = tensor([0, 0, 2, 0, 0])]; + tensor v_45_end_0 = const()[name = tensor("v_45_end_0"), val = tensor([1, 1, 3, 16, 64])]; + tensor v_45_end_mask_0 = const()[name = tensor("v_45_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor v_45_squeeze_mask_0 = const()[name = tensor("v_45_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor v_45_cast_fp16 = slice_by_index(begin = v_45_begin_0, end = v_45_end_0, end_mask = v_45_end_mask_0, squeeze_mask = v_45_squeeze_mask_0, x = qkv_45_cast_fp16)[name = tensor("v_45_cast_fp16")]; + tensor var_8776 = const()[name = tensor("op_8776"), val = tensor([1, 1, 1, 1])]; + tensor position22_to_fp16_dtype_0 = const()[name = tensor("position22_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor position22_to_fp16 = cast(dtype = position22_to_fp16_dtype_0, x = position22)[name = tensor("cast_443")]; + tensor ts_137_cast_fp16 = reshape(shape = var_8776, x = position22_to_fp16)[name = tensor("ts_137_cast_fp16")]; + tensor var_8783 = const()[name = tensor("op_8783"), val = tensor([1, 1, 16, 32, 2])]; + tensor q_complex_45_cast_fp16 = reshape(shape = var_8783, x = q_133_cast_fp16)[name = tensor("q_complex_45_cast_fp16")]; + tensor var_8790 = const()[name = tensor("op_8790"), val = tensor([1, 1, 16, 32, 2])]; + tensor k_complex_45_cast_fp16 = reshape(shape = var_8790, x = k_89_cast_fp16)[name = tensor("k_complex_45_cast_fp16")]; + tensor var_8794_begin_0 = const()[name = tensor("op_8794_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_8794_end_0 = const()[name = tensor("op_8794_end_0"), val = tensor([1, 1, 16, 32, 1])]; + tensor var_8794_end_mask_0 = const()[name = tensor("op_8794_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_8794_squeeze_mask_0 = const()[name = tensor("op_8794_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_8794_cast_fp16 = slice_by_index(begin = var_8794_begin_0, end = var_8794_end_0, end_mask = var_8794_end_mask_0, squeeze_mask = var_8794_squeeze_mask_0, x = q_complex_45_cast_fp16)[name = tensor("op_8794_cast_fp16")]; + tensor var_8802_begin_0 = const()[name = tensor("op_8802_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_8802_end_0 = const()[name = tensor("op_8802_end_0"), val = tensor([1, 1, 16, 32, 2])]; + tensor var_8802_end_mask_0 = const()[name = tensor("op_8802_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_8802_squeeze_mask_0 = const()[name = tensor("op_8802_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_8802_cast_fp16 = slice_by_index(begin = var_8802_begin_0, end = var_8802_end_0, end_mask = var_8802_end_mask_0, squeeze_mask = var_8802_squeeze_mask_0, x = q_complex_45_cast_fp16)[name = tensor("op_8802_cast_fp16")]; + tensor var_8810_begin_0 = const()[name = tensor("op_8810_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_8810_end_0 = const()[name = tensor("op_8810_end_0"), val = tensor([1, 1, 16, 32, 1])]; + tensor var_8810_end_mask_0 = const()[name = tensor("op_8810_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_8810_squeeze_mask_0 = const()[name = tensor("op_8810_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_8810_cast_fp16 = slice_by_index(begin = var_8810_begin_0, end = var_8810_end_0, end_mask = var_8810_end_mask_0, squeeze_mask = var_8810_squeeze_mask_0, x = k_complex_45_cast_fp16)[name = tensor("op_8810_cast_fp16")]; + tensor var_8818_begin_0 = const()[name = tensor("op_8818_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_8818_end_0 = const()[name = tensor("op_8818_end_0"), val = tensor([1, 1, 16, 32, 2])]; + tensor var_8818_end_mask_0 = const()[name = tensor("op_8818_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_8818_squeeze_mask_0 = const()[name = tensor("op_8818_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_8818_cast_fp16 = slice_by_index(begin = var_8818_begin_0, end = var_8818_end_0, end_mask = var_8818_end_mask_0, squeeze_mask = var_8818_squeeze_mask_0, x = k_complex_45_cast_fp16)[name = tensor("op_8818_cast_fp16")]; + tensor freqs_45_to_fp16 = const()[name = tensor("freqs_45_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(561270144)))]; + tensor var_8824_cast_fp16 = mul(x = freqs_45_to_fp16, y = ts_137_cast_fp16)[name = tensor("op_8824_cast_fp16")]; + tensor rotr_45_cast_fp16 = cos(x = var_8824_cast_fp16)[name = tensor("rotr_45_cast_fp16")]; + tensor roti_45_cast_fp16 = sin(x = var_8824_cast_fp16)[name = tensor("roti_45_cast_fp16")]; + tensor var_8828_cast_fp16 = mul(x = var_8794_cast_fp16, y = rotr_45_cast_fp16)[name = tensor("op_8828_cast_fp16")]; + tensor var_8829_cast_fp16 = mul(x = var_8802_cast_fp16, y = roti_45_cast_fp16)[name = tensor("op_8829_cast_fp16")]; + tensor qor_89_cast_fp16 = sub(x = var_8828_cast_fp16, y = var_8829_cast_fp16)[name = tensor("qor_89_cast_fp16")]; + tensor var_8832_cast_fp16 = mul(x = var_8794_cast_fp16, y = roti_45_cast_fp16)[name = tensor("op_8832_cast_fp16")]; + tensor var_8833_cast_fp16 = mul(x = var_8802_cast_fp16, y = rotr_45_cast_fp16)[name = tensor("op_8833_cast_fp16")]; + tensor qoi_89_cast_fp16 = add(x = var_8832_cast_fp16, y = var_8833_cast_fp16)[name = tensor("qoi_89_cast_fp16")]; + tensor var_8836_cast_fp16 = mul(x = var_8810_cast_fp16, y = rotr_45_cast_fp16)[name = tensor("op_8836_cast_fp16")]; + tensor var_8837_cast_fp16 = mul(x = var_8818_cast_fp16, y = roti_45_cast_fp16)[name = tensor("op_8837_cast_fp16")]; + tensor kor_89_cast_fp16 = sub(x = var_8836_cast_fp16, y = var_8837_cast_fp16)[name = tensor("kor_89_cast_fp16")]; + tensor var_8840_cast_fp16 = mul(x = var_8810_cast_fp16, y = roti_45_cast_fp16)[name = tensor("op_8840_cast_fp16")]; + tensor var_8841_cast_fp16 = mul(x = var_8818_cast_fp16, y = rotr_45_cast_fp16)[name = tensor("op_8841_cast_fp16")]; + tensor koi_89_cast_fp16 = add(x = var_8840_cast_fp16, y = var_8841_cast_fp16)[name = tensor("koi_89_cast_fp16")]; + tensor qo_45_axis_0 = const()[name = tensor("qo_45_axis_0"), val = tensor(-1)]; + tensor qo_45_cast_fp16 = stack(axis = qo_45_axis_0, values = (qor_89_cast_fp16, qoi_89_cast_fp16))[name = tensor("qo_45_cast_fp16")]; + tensor ko_45_axis_0 = const()[name = tensor("ko_45_axis_0"), val = tensor(-1)]; + tensor ko_45_cast_fp16 = stack(axis = ko_45_axis_0, values = (kor_89_cast_fp16, koi_89_cast_fp16))[name = tensor("ko_45_cast_fp16")]; + tensor var_8874 = const()[name = tensor("op_8874"), val = tensor([1, 1, 16, 64])]; + tensor q_135_cast_fp16 = reshape(shape = var_8874, x = qo_45_cast_fp16)[name = tensor("q_135_cast_fp16")]; + tensor var_8880 = const()[name = tensor("op_8880"), val = tensor([1, 1, 16, 64])]; + tensor k_91_cast_fp16 = reshape(shape = var_8880, x = ko_45_cast_fp16)[name = tensor("k_91_cast_fp16")]; + tensor _inversed_8905_y_0_to_fp16 = const()[name = tensor("_inversed_8905_y_0_to_fp16"), val = tensor(0x1p-9)]; + tensor _inversed_8905_cast_fp16 = mul(x = ts_137_cast_fp16, y = _inversed_8905_y_0_to_fp16)[name = tensor("_inversed_8905_cast_fp16")]; + tensor var_8906_cast_fp16 = floor(x = _inversed_8905_cast_fp16)[name = tensor("op_8906_cast_fp16")]; + tensor var_8907_to_fp16 = const()[name = tensor("op_8907_to_fp16"), val = tensor(0x1p+9)]; + tensor var_8908_cast_fp16 = mul(x = var_8906_cast_fp16, y = var_8907_to_fp16)[name = tensor("op_8908_cast_fp16")]; + tensor write_indices_float_91_cast_fp16 = sub(x = ts_137_cast_fp16, y = var_8908_cast_fp16)[name = tensor("write_indices_float_91_cast_fp16")]; + tensor var_8915_dtype_0 = const()[name = tensor("op_8915_dtype_0"), val = tensor("int32")]; + tensor write_indices_45_reps_0 = const()[name = tensor("write_indices_45_reps_0"), val = tensor([1, 1, 16, 64])]; + tensor write_indices_float_91_cast_fp16_to_int32 = cast(dtype = var_8915_dtype_0, x = write_indices_float_91_cast_fp16)[name = tensor("cast_442")]; + tensor write_indices_45 = tile(reps = write_indices_45_reps_0, x = write_indices_float_91_cast_fp16_to_int32)[name = tensor("write_indices_45")]; + tensor var_8925_begin_0 = const()[name = tensor("op_8925_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_8925_end_0 = const()[name = tensor("op_8925_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor var_8925_end_mask_0 = const()[name = tensor("op_8925_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_8925_squeeze_mask_0 = const()[name = tensor("op_8925_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor cache22_to_fp16_dtype_0 = const()[name = tensor("cache22_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor cache22_to_fp16 = cast(dtype = cache22_to_fp16_dtype_0, x = cache22)[name = tensor("cast_441")]; + tensor var_8925_cast_fp16 = slice_by_index(begin = var_8925_begin_0, end = var_8925_end_0, end_mask = var_8925_end_mask_0, squeeze_mask = var_8925_squeeze_mask_0, x = cache22_to_fp16)[name = tensor("op_8925_cast_fp16")]; + tensor var_8927_axis_0 = const()[name = tensor("op_8927_axis_0"), val = tensor(1)]; + tensor var_8927_mode_0 = const()[name = tensor("op_8927_mode_0"), val = tensor("update")]; + tensor var_8927_validate_indices_0 = const()[name = tensor("op_8927_validate_indices_0"), val = tensor(false)]; + tensor var_8927_cast_fp16 = scatter_along_axis(axis = var_8927_axis_0, data = var_8925_cast_fp16, indices = write_indices_45, mode = var_8927_mode_0, updates = k_91_cast_fp16, validate_indices = var_8927_validate_indices_0)[name = tensor("op_8927_cast_fp16")]; + tensor concat_156 = const()[name = tensor("concat_156"), val = tensor([0, 0, 0, 0, 0])]; + tensor concat_157 = const()[name = tensor("concat_157"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_45_internal_tensor_assign_1_stride_0 = const()[name = tensor("new_cache_45_internal_tensor_assign_1_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_45_internal_tensor_assign_1_begin_mask_0 = const()[name = tensor("new_cache_45_internal_tensor_assign_1_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_45_internal_tensor_assign_1_end_mask_0 = const()[name = tensor("new_cache_45_internal_tensor_assign_1_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_45_internal_tensor_assign_1_squeeze_mask_0 = const()[name = tensor("new_cache_45_internal_tensor_assign_1_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_92 = const()[name = tensor("shape_92"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_44 = const()[name = tensor("reduce_prod_44"), val = tensor(1048576)]; + tensor range_1d_44_start_0 = const()[name = tensor("range_1d_44_start_0"), val = tensor(0)]; + tensor range_1d_44_step_0 = const()[name = tensor("range_1d_44_step_0"), val = tensor(1)]; + tensor range_1d_44 = range_1d(end = reduce_prod_44, start = range_1d_44_start_0, step = range_1d_44_step_0)[name = tensor("range_1d_44")]; + tensor reshape_220 = reshape(shape = shape_92, x = range_1d_44)[name = tensor("reshape_220")]; + tensor slice_by_index_44 = slice_by_index(begin = concat_156, begin_mask = new_cache_45_internal_tensor_assign_1_begin_mask_0, end = concat_157, end_mask = new_cache_45_internal_tensor_assign_1_end_mask_0, squeeze_mask = new_cache_45_internal_tensor_assign_1_squeeze_mask_0, stride = new_cache_45_internal_tensor_assign_1_stride_0, x = reshape_220)[name = tensor("slice_by_index_44")]; + tensor reshape_221_shape_0 = const()[name = tensor("reshape_221_shape_0"), val = tensor([-1])]; + tensor reshape_221 = reshape(shape = reshape_221_shape_0, x = slice_by_index_44)[name = tensor("reshape_221")]; + tensor reshape_222_shape_0 = const()[name = tensor("reshape_222_shape_0"), val = tensor([-1])]; + tensor reshape_222_cast_fp16 = reshape(shape = reshape_222_shape_0, x = var_8927_cast_fp16)[name = tensor("reshape_222_cast_fp16")]; + tensor reshape_223_shape_0 = const()[name = tensor("reshape_223_shape_0"), val = tensor([-1])]; + tensor reshape_223_cast_fp16 = reshape(shape = reshape_223_shape_0, x = cache22_to_fp16)[name = tensor("reshape_223_cast_fp16")]; + tensor scatter_44_mode_0 = const()[name = tensor("scatter_44_mode_0"), val = tensor("update")]; + tensor scatter_44_axis_0 = const()[name = tensor("scatter_44_axis_0"), val = tensor(0)]; + tensor scatter_44_validate_indices_0 = const()[name = tensor("scatter_44_validate_indices_0"), val = tensor(false)]; + tensor scatter_44_cast_fp16 = scatter(axis = scatter_44_axis_0, data = reshape_223_cast_fp16, indices = reshape_221, mode = scatter_44_mode_0, updates = reshape_222_cast_fp16, validate_indices = scatter_44_validate_indices_0)[name = tensor("scatter_44_cast_fp16")]; + tensor reshape_224_cast_fp16 = reshape(shape = shape_92, x = scatter_44_cast_fp16)[name = tensor("reshape_224_cast_fp16")]; + tensor var_8935_begin_0 = const()[name = tensor("op_8935_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor var_8935_end_0 = const()[name = tensor("op_8935_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor var_8935_end_mask_0 = const()[name = tensor("op_8935_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_8935_squeeze_mask_0 = const()[name = tensor("op_8935_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor var_8935_cast_fp16 = slice_by_index(begin = var_8935_begin_0, end = var_8935_end_0, end_mask = var_8935_end_mask_0, squeeze_mask = var_8935_squeeze_mask_0, x = reshape_224_cast_fp16)[name = tensor("op_8935_cast_fp16")]; + tensor var_8937_axis_0 = const()[name = tensor("op_8937_axis_0"), val = tensor(1)]; + tensor var_8937_mode_0 = const()[name = tensor("op_8937_mode_0"), val = tensor("update")]; + tensor var_8937_validate_indices_0 = const()[name = tensor("op_8937_validate_indices_0"), val = tensor(false)]; + tensor var_8937_cast_fp16 = scatter_along_axis(axis = var_8937_axis_0, data = var_8935_cast_fp16, indices = write_indices_45, mode = var_8937_mode_0, updates = v_45_cast_fp16, validate_indices = var_8937_validate_indices_0)[name = tensor("op_8937_cast_fp16")]; + tensor concat_158 = const()[name = tensor("concat_158"), val = tensor([1, 0, 0, 0, 0])]; + tensor concat_159 = const()[name = tensor("concat_159"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_45_internal_tensor_assign_2_stride_0 = const()[name = tensor("new_cache_45_internal_tensor_assign_2_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_45_internal_tensor_assign_2_begin_mask_0 = const()[name = tensor("new_cache_45_internal_tensor_assign_2_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_45_internal_tensor_assign_2_end_mask_0 = const()[name = tensor("new_cache_45_internal_tensor_assign_2_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_45_internal_tensor_assign_2_squeeze_mask_0 = const()[name = tensor("new_cache_45_internal_tensor_assign_2_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_93 = const()[name = tensor("shape_93"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_45 = const()[name = tensor("reduce_prod_45"), val = tensor(1048576)]; + tensor range_1d_45_start_0 = const()[name = tensor("range_1d_45_start_0"), val = tensor(0)]; + tensor range_1d_45_step_0 = const()[name = tensor("range_1d_45_step_0"), val = tensor(1)]; + tensor range_1d_45 = range_1d(end = reduce_prod_45, start = range_1d_45_start_0, step = range_1d_45_step_0)[name = tensor("range_1d_45")]; + tensor reshape_225 = reshape(shape = shape_93, x = range_1d_45)[name = tensor("reshape_225")]; + tensor slice_by_index_45 = slice_by_index(begin = concat_158, begin_mask = new_cache_45_internal_tensor_assign_2_begin_mask_0, end = concat_159, end_mask = new_cache_45_internal_tensor_assign_2_end_mask_0, squeeze_mask = new_cache_45_internal_tensor_assign_2_squeeze_mask_0, stride = new_cache_45_internal_tensor_assign_2_stride_0, x = reshape_225)[name = tensor("slice_by_index_45")]; + tensor reshape_226_shape_0 = const()[name = tensor("reshape_226_shape_0"), val = tensor([-1])]; + tensor reshape_226 = reshape(shape = reshape_226_shape_0, x = slice_by_index_45)[name = tensor("reshape_226")]; + tensor reshape_227_shape_0 = const()[name = tensor("reshape_227_shape_0"), val = tensor([-1])]; + tensor reshape_227_cast_fp16 = reshape(shape = reshape_227_shape_0, x = var_8937_cast_fp16)[name = tensor("reshape_227_cast_fp16")]; + tensor reshape_228_shape_0 = const()[name = tensor("reshape_228_shape_0"), val = tensor([-1])]; + tensor reshape_228_cast_fp16 = reshape(shape = reshape_228_shape_0, x = reshape_224_cast_fp16)[name = tensor("reshape_228_cast_fp16")]; + tensor scatter_45_mode_0 = const()[name = tensor("scatter_45_mode_0"), val = tensor("update")]; + tensor scatter_45_axis_0 = const()[name = tensor("scatter_45_axis_0"), val = tensor(0)]; + tensor scatter_45_validate_indices_0 = const()[name = tensor("scatter_45_validate_indices_0"), val = tensor(false)]; + tensor scatter_45_cast_fp16 = scatter(axis = scatter_45_axis_0, data = reshape_228_cast_fp16, indices = reshape_226, mode = scatter_45_mode_0, updates = reshape_227_cast_fp16, validate_indices = scatter_45_validate_indices_0)[name = tensor("scatter_45_cast_fp16")]; + tensor reshape_229_cast_fp16 = reshape(shape = shape_93, x = scatter_45_cast_fp16)[name = tensor("reshape_229_cast_fp16")]; + tensor reshape_229_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("reshape_229_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor keys_133_begin_0 = const()[name = tensor("keys_133_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor keys_133_end_0 = const()[name = tensor("keys_133_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor keys_133_end_mask_0 = const()[name = tensor("keys_133_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor keys_133_squeeze_mask_0 = const()[name = tensor("keys_133_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor keys_133_cast_fp16 = slice_by_index(begin = keys_133_begin_0, end = keys_133_end_0, end_mask = keys_133_end_mask_0, squeeze_mask = keys_133_squeeze_mask_0, x = reshape_229_cast_fp16)[name = tensor("keys_133_cast_fp16")]; + tensor values_133_begin_0 = const()[name = tensor("values_133_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor values_133_end_0 = const()[name = tensor("values_133_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor values_133_end_mask_0 = const()[name = tensor("values_133_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor values_133_squeeze_mask_0 = const()[name = tensor("values_133_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor values_133_cast_fp16 = slice_by_index(begin = values_133_begin_0, end = values_133_end_0, end_mask = values_133_end_mask_0, squeeze_mask = values_133_squeeze_mask_0, x = reshape_229_cast_fp16)[name = tensor("values_133_cast_fp16")]; + tensor var_8949_cast_fp16 = not_equal(x = keys_133_cast_fp16, y = keys_133_cast_fp16)[name = tensor("op_8949_cast_fp16")]; + tensor keys_135_cast_fp16 = select(a = var_463_to_fp16, b = keys_133_cast_fp16, cond = var_8949_cast_fp16)[name = tensor("keys_135_cast_fp16")]; + tensor var_8957_cast_fp16 = not_equal(x = values_133_cast_fp16, y = values_133_cast_fp16)[name = tensor("op_8957_cast_fp16")]; + tensor values_135_cast_fp16 = select(a = var_463_to_fp16, b = values_133_cast_fp16, cond = var_8957_cast_fp16)[name = tensor("values_135_cast_fp16")]; + tensor var_8981 = const()[name = tensor("op_8981"), val = tensor([0, 2, 1, 3])]; + tensor var_8997 = const()[name = tensor("op_8997"), val = tensor([1, 1, 1])]; + tensor var_8998_cast_fp16 = reshape(shape = var_8997, x = position22_to_fp16)[name = tensor("op_8998_cast_fp16")]; + tensor var_9018_to_fp16 = const()[name = tensor("op_9018_to_fp16"), val = tensor(0x1p+0)]; + tensor valid_len_45_cast_fp16 = add(x = var_8998_cast_fp16, y = var_9018_to_fp16)[name = tensor("valid_len_45_cast_fp16")]; + tensor valid_mask_45_cast_fp16 = less(x = k_positions_1_promoted_to_fp16, y = valid_len_45_cast_fp16)[name = tensor("valid_mask_45_cast_fp16")]; + tensor causal_mask_45_cast_fp16 = less_equal(x = k_positions_1_promoted_to_fp16, y = var_8998_cast_fp16)[name = tensor("causal_mask_45_cast_fp16")]; + tensor attn_mask_89 = logical_and(x = valid_mask_45_cast_fp16, y = causal_mask_45_cast_fp16)[name = tensor("attn_mask_89")]; + tensor attn_mask_91_axes_0 = const()[name = tensor("attn_mask_91_axes_0"), val = tensor([1])]; + tensor attn_mask_91 = expand_dims(axes = attn_mask_91_axes_0, x = attn_mask_89)[name = tensor("attn_mask_91")]; + tensor var_9036_transpose_x_0 = const()[name = tensor("op_9036_transpose_x_0"), val = tensor(false)]; + tensor var_9036_transpose_y_0 = const()[name = tensor("op_9036_transpose_y_0"), val = tensor(false)]; + tensor transpose_116_perm_0 = const()[name = tensor("transpose_116_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_117_perm_0 = const()[name = tensor("transpose_117_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_117 = transpose(perm = transpose_117_perm_0, x = keys_135_cast_fp16)[name = tensor("transpose_125")]; + tensor transpose_116 = transpose(perm = transpose_116_perm_0, x = q_135_cast_fp16)[name = tensor("transpose_126")]; + tensor var_9036_cast_fp16 = matmul(transpose_x = var_9036_transpose_x_0, transpose_y = var_9036_transpose_y_0, x = transpose_116, y = transpose_117)[name = tensor("op_9036_cast_fp16")]; + tensor var_9030_to_fp16 = const()[name = tensor("op_9030_to_fp16"), val = tensor([0x1p-3])]; + tensor attn_weights_133_cast_fp16 = mul(x = var_9036_cast_fp16, y = var_9030_to_fp16)[name = tensor("attn_weights_133_cast_fp16")]; + tensor var_9038 = logical_not(x = attn_mask_91)[name = tensor("op_9038")]; + tensor var_9039_to_fp16 = const()[name = tensor("op_9039_to_fp16"), val = tensor(-inf)]; + tensor attn_weights_135_cast_fp16 = select(a = var_9039_to_fp16, b = attn_weights_133_cast_fp16, cond = var_9038)[name = tensor("attn_weights_135_cast_fp16")]; + tensor var_9041 = const()[name = tensor("op_9041"), val = tensor(-1)]; + tensor attn_weights_137_cast_fp16 = softmax(axis = var_9041, x = attn_weights_135_cast_fp16)[name = tensor("attn_weights_137_cast_fp16")]; + tensor attn_output_45_transpose_x_0 = const()[name = tensor("attn_output_45_transpose_x_0"), val = tensor(false)]; + tensor attn_output_45_transpose_y_0 = const()[name = tensor("attn_output_45_transpose_y_0"), val = tensor(false)]; + tensor values_137_cast_fp16 = transpose(perm = var_8981, x = values_135_cast_fp16)[name = tensor("transpose_127")]; + tensor attn_output_45_cast_fp16 = matmul(transpose_x = attn_output_45_transpose_x_0, transpose_y = attn_output_45_transpose_y_0, x = attn_weights_137_cast_fp16, y = values_137_cast_fp16)[name = tensor("attn_output_45_cast_fp16")]; + tensor var_9049 = const()[name = tensor("op_9049"), val = tensor([0, 2, 1, 3])]; + tensor var_9054 = const()[name = tensor("op_9054"), val = tensor([1, 1, 1024])]; + tensor var_9050_cast_fp16 = transpose(perm = var_9049, x = attn_output_45_cast_fp16)[name = tensor("transpose_124")]; + tensor input_271_cast_fp16 = reshape(shape = var_9054, x = var_9050_cast_fp16)[name = tensor("input_271_cast_fp16")]; + tensor attn22_out_proj_weight_to_fp16 = const()[name = tensor("attn22_out_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(561270272)))]; + tensor linear_90_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = attn22_out_proj_weight_to_fp16, x = input_271_cast_fp16)[name = tensor("linear_90_cast_fp16")]; + tensor var_9060_to_fp16 = const()[name = tensor("op_9060_to_fp16"), val = tensor(0x1p+0)]; + tensor var_9061_cast_fp16 = add(x = position22_to_fp16, y = var_9060_to_fp16)[name = tensor("op_9061_cast_fp16")]; + tensor var_9061_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_9061_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor input_273_cast_fp16 = add(x = input_267_cast_fp16, y = linear_90_cast_fp16)[name = tensor("input_273_cast_fp16")]; + tensor input_275_axes_0 = const()[name = tensor("input_275_axes_0"), val = tensor([-1])]; + tensor norm22_2_weight_to_fp16 = const()[name = tensor("norm22_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(563367488)))]; + tensor norm22_2_bias_to_fp16 = const()[name = tensor("norm22_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(563369600)))]; + tensor var_9065_to_fp16 = const()[name = tensor("op_9065_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_275_cast_fp16 = layer_norm(axes = input_275_axes_0, beta = norm22_2_bias_to_fp16, epsilon = var_9065_to_fp16, gamma = norm22_2_weight_to_fp16, x = input_273_cast_fp16)[name = tensor("input_275_cast_fp16")]; + tensor linear22_1_weight_to_fp16 = const()[name = tensor("linear22_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(563371712)))]; + tensor linear_91_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = linear22_1_weight_to_fp16, x = input_275_cast_fp16)[name = tensor("linear_91_cast_fp16")]; + tensor input_277_mode_0 = const()[name = tensor("input_277_mode_0"), val = tensor("EXACT")]; + tensor input_277_cast_fp16 = gelu(mode = input_277_mode_0, x = linear_91_cast_fp16)[name = tensor("input_277_cast_fp16")]; + tensor linear22_2_weight_to_fp16 = const()[name = tensor("linear22_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(571760384)))]; + tensor linear_92_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = linear22_2_weight_to_fp16, x = input_277_cast_fp16)[name = tensor("linear_92_cast_fp16")]; + tensor input_279_cast_fp16 = add(x = input_273_cast_fp16, y = linear_92_cast_fp16)[name = tensor("input_279_cast_fp16")]; + tensor input_281_axes_0 = const()[name = tensor("input_281_axes_0"), val = tensor([-1])]; + tensor norm23_1_weight_to_fp16 = const()[name = tensor("norm23_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(580149056)))]; + tensor norm23_1_bias_to_fp16 = const()[name = tensor("norm23_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(580151168)))]; + tensor var_9082_to_fp16 = const()[name = tensor("op_9082_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_281_cast_fp16 = layer_norm(axes = input_281_axes_0, beta = norm23_1_bias_to_fp16, epsilon = var_9082_to_fp16, gamma = norm23_1_weight_to_fp16, x = input_279_cast_fp16)[name = tensor("input_281_cast_fp16")]; + tensor attn23_in_proj_weight_to_fp16 = const()[name = tensor("attn23_in_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(580153280)))]; + tensor linear_93_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = attn23_in_proj_weight_to_fp16, x = input_281_cast_fp16)[name = tensor("linear_93_cast_fp16")]; + tensor var_9096 = const()[name = tensor("op_9096"), val = tensor([1, 1, 3, 16, 64])]; + tensor qkv_cast_fp16 = reshape(shape = var_9096, x = linear_93_cast_fp16)[name = tensor("qkv_cast_fp16")]; + tensor q_139_begin_0 = const()[name = tensor("q_139_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor q_139_end_0 = const()[name = tensor("q_139_end_0"), val = tensor([1, 1, 1, 16, 64])]; + tensor q_139_end_mask_0 = const()[name = tensor("q_139_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor q_139_squeeze_mask_0 = const()[name = tensor("q_139_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor q_139_cast_fp16 = slice_by_index(begin = q_139_begin_0, end = q_139_end_0, end_mask = q_139_end_mask_0, squeeze_mask = q_139_squeeze_mask_0, x = qkv_cast_fp16)[name = tensor("q_139_cast_fp16")]; + tensor k_93_begin_0 = const()[name = tensor("k_93_begin_0"), val = tensor([0, 0, 1, 0, 0])]; + tensor k_93_end_0 = const()[name = tensor("k_93_end_0"), val = tensor([1, 1, 2, 16, 64])]; + tensor k_93_end_mask_0 = const()[name = tensor("k_93_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor k_93_squeeze_mask_0 = const()[name = tensor("k_93_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor k_93_cast_fp16 = slice_by_index(begin = k_93_begin_0, end = k_93_end_0, end_mask = k_93_end_mask_0, squeeze_mask = k_93_squeeze_mask_0, x = qkv_cast_fp16)[name = tensor("k_93_cast_fp16")]; + tensor v_begin_0 = const()[name = tensor("v_begin_0"), val = tensor([0, 0, 2, 0, 0])]; + tensor v_end_0 = const()[name = tensor("v_end_0"), val = tensor([1, 1, 3, 16, 64])]; + tensor v_end_mask_0 = const()[name = tensor("v_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor v_squeeze_mask_0 = const()[name = tensor("v_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor v_cast_fp16 = slice_by_index(begin = v_begin_0, end = v_end_0, end_mask = v_end_mask_0, squeeze_mask = v_squeeze_mask_0, x = qkv_cast_fp16)[name = tensor("v_cast_fp16")]; + tensor var_9162 = const()[name = tensor("op_9162"), val = tensor([1, 1, 1, 1])]; + tensor position23_to_fp16_dtype_0 = const()[name = tensor("position23_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor position23_to_fp16 = cast(dtype = position23_to_fp16_dtype_0, x = position23)[name = tensor("cast_438")]; + tensor ts_cast_fp16 = reshape(shape = var_9162, x = position23_to_fp16)[name = tensor("ts_cast_fp16")]; + tensor var_9169 = const()[name = tensor("op_9169"), val = tensor([1, 1, 16, 32, 2])]; + tensor q_complex_cast_fp16 = reshape(shape = var_9169, x = q_139_cast_fp16)[name = tensor("q_complex_cast_fp16")]; + tensor var_9176 = const()[name = tensor("op_9176"), val = tensor([1, 1, 16, 32, 2])]; + tensor k_complex_cast_fp16 = reshape(shape = var_9176, x = k_93_cast_fp16)[name = tensor("k_complex_cast_fp16")]; + tensor var_9180_begin_0 = const()[name = tensor("op_9180_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_9180_end_0 = const()[name = tensor("op_9180_end_0"), val = tensor([1, 1, 16, 32, 1])]; + tensor var_9180_end_mask_0 = const()[name = tensor("op_9180_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_9180_squeeze_mask_0 = const()[name = tensor("op_9180_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_9180_cast_fp16 = slice_by_index(begin = var_9180_begin_0, end = var_9180_end_0, end_mask = var_9180_end_mask_0, squeeze_mask = var_9180_squeeze_mask_0, x = q_complex_cast_fp16)[name = tensor("op_9180_cast_fp16")]; + tensor var_9188_begin_0 = const()[name = tensor("op_9188_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_9188_end_0 = const()[name = tensor("op_9188_end_0"), val = tensor([1, 1, 16, 32, 2])]; + tensor var_9188_end_mask_0 = const()[name = tensor("op_9188_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_9188_squeeze_mask_0 = const()[name = tensor("op_9188_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_9188_cast_fp16 = slice_by_index(begin = var_9188_begin_0, end = var_9188_end_0, end_mask = var_9188_end_mask_0, squeeze_mask = var_9188_squeeze_mask_0, x = q_complex_cast_fp16)[name = tensor("op_9188_cast_fp16")]; + tensor var_9196_begin_0 = const()[name = tensor("op_9196_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_9196_end_0 = const()[name = tensor("op_9196_end_0"), val = tensor([1, 1, 16, 32, 1])]; + tensor var_9196_end_mask_0 = const()[name = tensor("op_9196_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_9196_squeeze_mask_0 = const()[name = tensor("op_9196_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_9196_cast_fp16 = slice_by_index(begin = var_9196_begin_0, end = var_9196_end_0, end_mask = var_9196_end_mask_0, squeeze_mask = var_9196_squeeze_mask_0, x = k_complex_cast_fp16)[name = tensor("op_9196_cast_fp16")]; + tensor var_9204_begin_0 = const()[name = tensor("op_9204_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_9204_end_0 = const()[name = tensor("op_9204_end_0"), val = tensor([1, 1, 16, 32, 2])]; + tensor var_9204_end_mask_0 = const()[name = tensor("op_9204_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_9204_squeeze_mask_0 = const()[name = tensor("op_9204_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_9204_cast_fp16 = slice_by_index(begin = var_9204_begin_0, end = var_9204_end_0, end_mask = var_9204_end_mask_0, squeeze_mask = var_9204_squeeze_mask_0, x = k_complex_cast_fp16)[name = tensor("op_9204_cast_fp16")]; + tensor freqs_to_fp16 = const()[name = tensor("freqs_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(586444800)))]; + tensor var_9210_cast_fp16 = mul(x = freqs_to_fp16, y = ts_cast_fp16)[name = tensor("op_9210_cast_fp16")]; + tensor rotr_cast_fp16 = cos(x = var_9210_cast_fp16)[name = tensor("rotr_cast_fp16")]; + tensor roti_cast_fp16 = sin(x = var_9210_cast_fp16)[name = tensor("roti_cast_fp16")]; + tensor var_9214_cast_fp16 = mul(x = var_9180_cast_fp16, y = rotr_cast_fp16)[name = tensor("op_9214_cast_fp16")]; + tensor var_9215_cast_fp16 = mul(x = var_9188_cast_fp16, y = roti_cast_fp16)[name = tensor("op_9215_cast_fp16")]; + tensor qor_93_cast_fp16 = sub(x = var_9214_cast_fp16, y = var_9215_cast_fp16)[name = tensor("qor_93_cast_fp16")]; + tensor var_9218_cast_fp16 = mul(x = var_9180_cast_fp16, y = roti_cast_fp16)[name = tensor("op_9218_cast_fp16")]; + tensor var_9219_cast_fp16 = mul(x = var_9188_cast_fp16, y = rotr_cast_fp16)[name = tensor("op_9219_cast_fp16")]; + tensor qoi_93_cast_fp16 = add(x = var_9218_cast_fp16, y = var_9219_cast_fp16)[name = tensor("qoi_93_cast_fp16")]; + tensor var_9222_cast_fp16 = mul(x = var_9196_cast_fp16, y = rotr_cast_fp16)[name = tensor("op_9222_cast_fp16")]; + tensor var_9223_cast_fp16 = mul(x = var_9204_cast_fp16, y = roti_cast_fp16)[name = tensor("op_9223_cast_fp16")]; + tensor kor_93_cast_fp16 = sub(x = var_9222_cast_fp16, y = var_9223_cast_fp16)[name = tensor("kor_93_cast_fp16")]; + tensor var_9226_cast_fp16 = mul(x = var_9196_cast_fp16, y = roti_cast_fp16)[name = tensor("op_9226_cast_fp16")]; + tensor var_9227_cast_fp16 = mul(x = var_9204_cast_fp16, y = rotr_cast_fp16)[name = tensor("op_9227_cast_fp16")]; + tensor koi_93_cast_fp16 = add(x = var_9226_cast_fp16, y = var_9227_cast_fp16)[name = tensor("koi_93_cast_fp16")]; + tensor qo_axis_0 = const()[name = tensor("qo_axis_0"), val = tensor(-1)]; + tensor qo_cast_fp16 = stack(axis = qo_axis_0, values = (qor_93_cast_fp16, qoi_93_cast_fp16))[name = tensor("qo_cast_fp16")]; + tensor ko_axis_0 = const()[name = tensor("ko_axis_0"), val = tensor(-1)]; + tensor ko_cast_fp16 = stack(axis = ko_axis_0, values = (kor_93_cast_fp16, koi_93_cast_fp16))[name = tensor("ko_cast_fp16")]; + tensor var_9260 = const()[name = tensor("op_9260"), val = tensor([1, 1, 16, 64])]; + tensor q_141_cast_fp16 = reshape(shape = var_9260, x = qo_cast_fp16)[name = tensor("q_141_cast_fp16")]; + tensor var_9266 = const()[name = tensor("op_9266"), val = tensor([1, 1, 16, 64])]; + tensor k_cast_fp16 = reshape(shape = var_9266, x = ko_cast_fp16)[name = tensor("k_cast_fp16")]; + tensor _inversed_9291_y_0_to_fp16 = const()[name = tensor("_inversed_9291_y_0_to_fp16"), val = tensor(0x1p-9)]; + tensor _inversed_9291_cast_fp16 = mul(x = ts_cast_fp16, y = _inversed_9291_y_0_to_fp16)[name = tensor("_inversed_9291_cast_fp16")]; + tensor var_9292_cast_fp16 = floor(x = _inversed_9291_cast_fp16)[name = tensor("op_9292_cast_fp16")]; + tensor var_9293_to_fp16 = const()[name = tensor("op_9293_to_fp16"), val = tensor(0x1p+9)]; + tensor var_9294_cast_fp16 = mul(x = var_9292_cast_fp16, y = var_9293_to_fp16)[name = tensor("op_9294_cast_fp16")]; + tensor write_indices_float_cast_fp16 = sub(x = ts_cast_fp16, y = var_9294_cast_fp16)[name = tensor("write_indices_float_cast_fp16")]; + tensor var_9301_dtype_0 = const()[name = tensor("op_9301_dtype_0"), val = tensor("int32")]; + tensor write_indices_reps_0 = const()[name = tensor("write_indices_reps_0"), val = tensor([1, 1, 16, 64])]; + tensor write_indices_float_cast_fp16_to_int32 = cast(dtype = var_9301_dtype_0, x = write_indices_float_cast_fp16)[name = tensor("cast_437")]; + tensor write_indices = tile(reps = write_indices_reps_0, x = write_indices_float_cast_fp16_to_int32)[name = tensor("write_indices")]; + tensor var_9311_begin_0 = const()[name = tensor("op_9311_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_9311_end_0 = const()[name = tensor("op_9311_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor var_9311_end_mask_0 = const()[name = tensor("op_9311_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_9311_squeeze_mask_0 = const()[name = tensor("op_9311_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor cache23_to_fp16_dtype_0 = const()[name = tensor("cache23_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor cache23_to_fp16 = cast(dtype = cache23_to_fp16_dtype_0, x = cache23)[name = tensor("cast_436")]; + tensor var_9311_cast_fp16 = slice_by_index(begin = var_9311_begin_0, end = var_9311_end_0, end_mask = var_9311_end_mask_0, squeeze_mask = var_9311_squeeze_mask_0, x = cache23_to_fp16)[name = tensor("op_9311_cast_fp16")]; + tensor var_9313_axis_0 = const()[name = tensor("op_9313_axis_0"), val = tensor(1)]; + tensor var_9313_mode_0 = const()[name = tensor("op_9313_mode_0"), val = tensor("update")]; + tensor var_9313_validate_indices_0 = const()[name = tensor("op_9313_validate_indices_0"), val = tensor(false)]; + tensor var_9313_cast_fp16 = scatter_along_axis(axis = var_9313_axis_0, data = var_9311_cast_fp16, indices = write_indices, mode = var_9313_mode_0, updates = k_cast_fp16, validate_indices = var_9313_validate_indices_0)[name = tensor("op_9313_cast_fp16")]; + tensor concat_163 = const()[name = tensor("concat_163"), val = tensor([0, 0, 0, 0, 0])]; + tensor concat_164 = const()[name = tensor("concat_164"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_internal_tensor_assign_1_stride_0 = const()[name = tensor("new_cache_internal_tensor_assign_1_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_internal_tensor_assign_1_begin_mask_0 = const()[name = tensor("new_cache_internal_tensor_assign_1_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_internal_tensor_assign_1_end_mask_0 = const()[name = tensor("new_cache_internal_tensor_assign_1_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_internal_tensor_assign_1_squeeze_mask_0 = const()[name = tensor("new_cache_internal_tensor_assign_1_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_94 = const()[name = tensor("shape_94"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_46 = const()[name = tensor("reduce_prod_46"), val = tensor(1048576)]; + tensor range_1d_46_start_0 = const()[name = tensor("range_1d_46_start_0"), val = tensor(0)]; + tensor range_1d_46_step_0 = const()[name = tensor("range_1d_46_step_0"), val = tensor(1)]; + tensor range_1d_46 = range_1d(end = reduce_prod_46, start = range_1d_46_start_0, step = range_1d_46_step_0)[name = tensor("range_1d_46")]; + tensor reshape_230 = reshape(shape = shape_94, x = range_1d_46)[name = tensor("reshape_230")]; + tensor slice_by_index_46 = slice_by_index(begin = concat_163, begin_mask = new_cache_internal_tensor_assign_1_begin_mask_0, end = concat_164, end_mask = new_cache_internal_tensor_assign_1_end_mask_0, squeeze_mask = new_cache_internal_tensor_assign_1_squeeze_mask_0, stride = new_cache_internal_tensor_assign_1_stride_0, x = reshape_230)[name = tensor("slice_by_index_46")]; + tensor reshape_231_shape_0 = const()[name = tensor("reshape_231_shape_0"), val = tensor([-1])]; + tensor reshape_231 = reshape(shape = reshape_231_shape_0, x = slice_by_index_46)[name = tensor("reshape_231")]; + tensor reshape_232_shape_0 = const()[name = tensor("reshape_232_shape_0"), val = tensor([-1])]; + tensor reshape_232_cast_fp16 = reshape(shape = reshape_232_shape_0, x = var_9313_cast_fp16)[name = tensor("reshape_232_cast_fp16")]; + tensor reshape_233_shape_0 = const()[name = tensor("reshape_233_shape_0"), val = tensor([-1])]; + tensor reshape_233_cast_fp16 = reshape(shape = reshape_233_shape_0, x = cache23_to_fp16)[name = tensor("reshape_233_cast_fp16")]; + tensor scatter_46_mode_0 = const()[name = tensor("scatter_46_mode_0"), val = tensor("update")]; + tensor scatter_46_axis_0 = const()[name = tensor("scatter_46_axis_0"), val = tensor(0)]; + tensor scatter_46_validate_indices_0 = const()[name = tensor("scatter_46_validate_indices_0"), val = tensor(false)]; + tensor scatter_46_cast_fp16 = scatter(axis = scatter_46_axis_0, data = reshape_233_cast_fp16, indices = reshape_231, mode = scatter_46_mode_0, updates = reshape_232_cast_fp16, validate_indices = scatter_46_validate_indices_0)[name = tensor("scatter_46_cast_fp16")]; + tensor reshape_234_cast_fp16 = reshape(shape = shape_94, x = scatter_46_cast_fp16)[name = tensor("reshape_234_cast_fp16")]; + tensor var_9321_begin_0 = const()[name = tensor("op_9321_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor var_9321_end_0 = const()[name = tensor("op_9321_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor var_9321_end_mask_0 = const()[name = tensor("op_9321_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_9321_squeeze_mask_0 = const()[name = tensor("op_9321_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor var_9321_cast_fp16 = slice_by_index(begin = var_9321_begin_0, end = var_9321_end_0, end_mask = var_9321_end_mask_0, squeeze_mask = var_9321_squeeze_mask_0, x = reshape_234_cast_fp16)[name = tensor("op_9321_cast_fp16")]; + tensor var_9323_axis_0 = const()[name = tensor("op_9323_axis_0"), val = tensor(1)]; + tensor var_9323_mode_0 = const()[name = tensor("op_9323_mode_0"), val = tensor("update")]; + tensor var_9323_validate_indices_0 = const()[name = tensor("op_9323_validate_indices_0"), val = tensor(false)]; + tensor var_9323_cast_fp16 = scatter_along_axis(axis = var_9323_axis_0, data = var_9321_cast_fp16, indices = write_indices, mode = var_9323_mode_0, updates = v_cast_fp16, validate_indices = var_9323_validate_indices_0)[name = tensor("op_9323_cast_fp16")]; + tensor concat_165 = const()[name = tensor("concat_165"), val = tensor([1, 0, 0, 0, 0])]; + tensor concat_166 = const()[name = tensor("concat_166"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_internal_tensor_assign_2_stride_0 = const()[name = tensor("new_cache_internal_tensor_assign_2_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_internal_tensor_assign_2_begin_mask_0 = const()[name = tensor("new_cache_internal_tensor_assign_2_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_internal_tensor_assign_2_end_mask_0 = const()[name = tensor("new_cache_internal_tensor_assign_2_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_internal_tensor_assign_2_squeeze_mask_0 = const()[name = tensor("new_cache_internal_tensor_assign_2_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_95 = const()[name = tensor("shape_95"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_47 = const()[name = tensor("reduce_prod_47"), val = tensor(1048576)]; + tensor range_1d_47_start_0 = const()[name = tensor("range_1d_47_start_0"), val = tensor(0)]; + tensor range_1d_47_step_0 = const()[name = tensor("range_1d_47_step_0"), val = tensor(1)]; + tensor range_1d_47 = range_1d(end = reduce_prod_47, start = range_1d_47_start_0, step = range_1d_47_step_0)[name = tensor("range_1d_47")]; + tensor reshape_235 = reshape(shape = shape_95, x = range_1d_47)[name = tensor("reshape_235")]; + tensor slice_by_index_47 = slice_by_index(begin = concat_165, begin_mask = new_cache_internal_tensor_assign_2_begin_mask_0, end = concat_166, end_mask = new_cache_internal_tensor_assign_2_end_mask_0, squeeze_mask = new_cache_internal_tensor_assign_2_squeeze_mask_0, stride = new_cache_internal_tensor_assign_2_stride_0, x = reshape_235)[name = tensor("slice_by_index_47")]; + tensor reshape_236_shape_0 = const()[name = tensor("reshape_236_shape_0"), val = tensor([-1])]; + tensor reshape_236 = reshape(shape = reshape_236_shape_0, x = slice_by_index_47)[name = tensor("reshape_236")]; + tensor reshape_237_shape_0 = const()[name = tensor("reshape_237_shape_0"), val = tensor([-1])]; + tensor reshape_237_cast_fp16 = reshape(shape = reshape_237_shape_0, x = var_9323_cast_fp16)[name = tensor("reshape_237_cast_fp16")]; + tensor reshape_238_shape_0 = const()[name = tensor("reshape_238_shape_0"), val = tensor([-1])]; + tensor reshape_238_cast_fp16 = reshape(shape = reshape_238_shape_0, x = reshape_234_cast_fp16)[name = tensor("reshape_238_cast_fp16")]; + tensor scatter_47_mode_0 = const()[name = tensor("scatter_47_mode_0"), val = tensor("update")]; + tensor scatter_47_axis_0 = const()[name = tensor("scatter_47_axis_0"), val = tensor(0)]; + tensor scatter_47_validate_indices_0 = const()[name = tensor("scatter_47_validate_indices_0"), val = tensor(false)]; + tensor scatter_47_cast_fp16 = scatter(axis = scatter_47_axis_0, data = reshape_238_cast_fp16, indices = reshape_236, mode = scatter_47_mode_0, updates = reshape_237_cast_fp16, validate_indices = scatter_47_validate_indices_0)[name = tensor("scatter_47_cast_fp16")]; + tensor reshape_239_cast_fp16 = reshape(shape = shape_95, x = scatter_47_cast_fp16)[name = tensor("reshape_239_cast_fp16")]; + tensor reshape_239_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("reshape_239_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor keys_139_begin_0 = const()[name = tensor("keys_139_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor keys_139_end_0 = const()[name = tensor("keys_139_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor keys_139_end_mask_0 = const()[name = tensor("keys_139_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor keys_139_squeeze_mask_0 = const()[name = tensor("keys_139_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor keys_139_cast_fp16 = slice_by_index(begin = keys_139_begin_0, end = keys_139_end_0, end_mask = keys_139_end_mask_0, squeeze_mask = keys_139_squeeze_mask_0, x = reshape_239_cast_fp16)[name = tensor("keys_139_cast_fp16")]; + tensor values_139_begin_0 = const()[name = tensor("values_139_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor values_139_end_0 = const()[name = tensor("values_139_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor values_139_end_mask_0 = const()[name = tensor("values_139_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor values_139_squeeze_mask_0 = const()[name = tensor("values_139_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor values_139_cast_fp16 = slice_by_index(begin = values_139_begin_0, end = values_139_end_0, end_mask = values_139_end_mask_0, squeeze_mask = values_139_squeeze_mask_0, x = reshape_239_cast_fp16)[name = tensor("values_139_cast_fp16")]; + tensor var_9335_cast_fp16 = not_equal(x = keys_139_cast_fp16, y = keys_139_cast_fp16)[name = tensor("op_9335_cast_fp16")]; + tensor keys_141_cast_fp16 = select(a = var_463_to_fp16, b = keys_139_cast_fp16, cond = var_9335_cast_fp16)[name = tensor("keys_141_cast_fp16")]; + tensor var_9343_cast_fp16 = not_equal(x = values_139_cast_fp16, y = values_139_cast_fp16)[name = tensor("op_9343_cast_fp16")]; + tensor values_141_cast_fp16 = select(a = var_463_to_fp16, b = values_139_cast_fp16, cond = var_9343_cast_fp16)[name = tensor("values_141_cast_fp16")]; + tensor var_9367 = const()[name = tensor("op_9367"), val = tensor([0, 2, 1, 3])]; + tensor var_9383 = const()[name = tensor("op_9383"), val = tensor([1, 1, 1])]; + tensor var_9384_cast_fp16 = reshape(shape = var_9383, x = position23_to_fp16)[name = tensor("op_9384_cast_fp16")]; + tensor var_9404_to_fp16 = const()[name = tensor("op_9404_to_fp16"), val = tensor(0x1p+0)]; + tensor valid_len_cast_fp16 = add(x = var_9384_cast_fp16, y = var_9404_to_fp16)[name = tensor("valid_len_cast_fp16")]; + tensor valid_mask_cast_fp16 = less(x = k_positions_1_promoted_to_fp16, y = valid_len_cast_fp16)[name = tensor("valid_mask_cast_fp16")]; + tensor causal_mask_cast_fp16 = less_equal(x = k_positions_1_promoted_to_fp16, y = var_9384_cast_fp16)[name = tensor("causal_mask_cast_fp16")]; + tensor attn_mask_93 = logical_and(x = valid_mask_cast_fp16, y = causal_mask_cast_fp16)[name = tensor("attn_mask_93")]; + tensor attn_mask_axes_0 = const()[name = tensor("attn_mask_axes_0"), val = tensor([1])]; + tensor attn_mask = expand_dims(axes = attn_mask_axes_0, x = attn_mask_93)[name = tensor("attn_mask")]; + tensor var_9422_transpose_x_0 = const()[name = tensor("op_9422_transpose_x_0"), val = tensor(false)]; + tensor var_9422_transpose_y_0 = const()[name = tensor("op_9422_transpose_y_0"), val = tensor(false)]; + tensor transpose_118_perm_0 = const()[name = tensor("transpose_118_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_119_perm_0 = const()[name = tensor("transpose_119_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_119 = transpose(perm = transpose_119_perm_0, x = keys_141_cast_fp16)[name = tensor("transpose_121")]; + tensor transpose_118 = transpose(perm = transpose_118_perm_0, x = q_141_cast_fp16)[name = tensor("transpose_122")]; + tensor var_9422_cast_fp16 = matmul(transpose_x = var_9422_transpose_x_0, transpose_y = var_9422_transpose_y_0, x = transpose_118, y = transpose_119)[name = tensor("op_9422_cast_fp16")]; + tensor var_9416_to_fp16 = const()[name = tensor("op_9416_to_fp16"), val = tensor([0x1p-3])]; + tensor attn_weights_139_cast_fp16 = mul(x = var_9422_cast_fp16, y = var_9416_to_fp16)[name = tensor("attn_weights_139_cast_fp16")]; + tensor var_9424 = logical_not(x = attn_mask)[name = tensor("op_9424")]; + tensor var_9425_to_fp16 = const()[name = tensor("op_9425_to_fp16"), val = tensor(-inf)]; + tensor attn_weights_141_cast_fp16 = select(a = var_9425_to_fp16, b = attn_weights_139_cast_fp16, cond = var_9424)[name = tensor("attn_weights_141_cast_fp16")]; + tensor var_9427 = const()[name = tensor("op_9427"), val = tensor(-1)]; + tensor attn_weights_cast_fp16 = softmax(axis = var_9427, x = attn_weights_141_cast_fp16)[name = tensor("attn_weights_cast_fp16")]; + tensor attn_output_transpose_x_0 = const()[name = tensor("attn_output_transpose_x_0"), val = tensor(false)]; + tensor attn_output_transpose_y_0 = const()[name = tensor("attn_output_transpose_y_0"), val = tensor(false)]; + tensor values_cast_fp16 = transpose(perm = var_9367, x = values_141_cast_fp16)[name = tensor("transpose_123")]; + tensor attn_output_cast_fp16 = matmul(transpose_x = attn_output_transpose_x_0, transpose_y = attn_output_transpose_y_0, x = attn_weights_cast_fp16, y = values_cast_fp16)[name = tensor("attn_output_cast_fp16")]; + tensor var_9435 = const()[name = tensor("op_9435"), val = tensor([0, 2, 1, 3])]; + tensor var_9440 = const()[name = tensor("op_9440"), val = tensor([1, 1, 1024])]; + tensor var_9436_cast_fp16 = transpose(perm = var_9435, x = attn_output_cast_fp16)[name = tensor("transpose_120")]; + tensor input_283_cast_fp16 = reshape(shape = var_9440, x = var_9436_cast_fp16)[name = tensor("input_283_cast_fp16")]; + tensor attn23_out_proj_weight_to_fp16 = const()[name = tensor("attn23_out_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(586444928)))]; + tensor linear_94_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = attn23_out_proj_weight_to_fp16, x = input_283_cast_fp16)[name = tensor("linear_94_cast_fp16")]; + tensor var_9446_to_fp16 = const()[name = tensor("op_9446_to_fp16"), val = tensor(0x1p+0)]; + tensor var_9447_cast_fp16 = add(x = position23_to_fp16, y = var_9446_to_fp16)[name = tensor("op_9447_cast_fp16")]; + tensor var_9447_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_9447_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor input_285_cast_fp16 = add(x = input_279_cast_fp16, y = linear_94_cast_fp16)[name = tensor("input_285_cast_fp16")]; + tensor input_287_axes_0 = const()[name = tensor("input_287_axes_0"), val = tensor([-1])]; + tensor norm23_2_weight_to_fp16 = const()[name = tensor("norm23_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(588542144)))]; + tensor norm23_2_bias_to_fp16 = const()[name = tensor("norm23_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(588544256)))]; + tensor var_9451_to_fp16 = const()[name = tensor("op_9451_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_287_cast_fp16 = layer_norm(axes = input_287_axes_0, beta = norm23_2_bias_to_fp16, epsilon = var_9451_to_fp16, gamma = norm23_2_weight_to_fp16, x = input_285_cast_fp16)[name = tensor("input_287_cast_fp16")]; + tensor linear23_1_weight_to_fp16 = const()[name = tensor("linear23_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(588546368)))]; + tensor linear_95_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = linear23_1_weight_to_fp16, x = input_287_cast_fp16)[name = tensor("linear_95_cast_fp16")]; + tensor input_289_mode_0 = const()[name = tensor("input_289_mode_0"), val = tensor("EXACT")]; + tensor input_289_cast_fp16 = gelu(mode = input_289_mode_0, x = linear_95_cast_fp16)[name = tensor("input_289_cast_fp16")]; + tensor linear23_2_weight_to_fp16 = const()[name = tensor("linear23_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(596935040)))]; + tensor linear_96_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = linear23_2_weight_to_fp16, x = input_289_cast_fp16)[name = tensor("linear_96_cast_fp16")]; + tensor input_291_cast_fp16 = add(x = input_285_cast_fp16, y = linear_96_cast_fp16)[name = tensor("input_291_cast_fp16")]; + tensor input_axes_0 = const()[name = tensor("input_axes_0"), val = tensor([-1])]; + tensor out_norm_weight_to_fp16 = const()[name = tensor("out_norm_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(605323712)))]; + tensor out_norm_bias_to_fp16 = const()[name = tensor("out_norm_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(605325824)))]; + tensor var_9468_to_fp16 = const()[name = tensor("op_9468_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_cast_fp16 = layer_norm(axes = input_axes_0, beta = out_norm_bias_to_fp16, epsilon = var_9468_to_fp16, gamma = out_norm_weight_to_fp16, x = input_291_cast_fp16)[name = tensor("input_cast_fp16")]; + tensor input_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("input_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor out_eos_weight_to_fp16 = const()[name = tensor("out_eos_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(605327936)))]; + tensor out_eos_bias_to_fp16 = const()[name = tensor("out_eos_bias_to_fp16"), val = tensor([-0x1.3ep-3])]; + tensor linear_97_cast_fp16 = linear(bias = out_eos_bias_to_fp16, weight = out_eos_weight_to_fp16, x = input_cast_fp16)[name = tensor("linear_97_cast_fp16")]; + tensor linear_97_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("linear_97_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor var_9476 = cast(dtype = linear_97_cast_fp16_to_fp32_dtype_0, x = linear_97_cast_fp16)[name = tensor("cast_432")]; + tensor input = cast(dtype = input_cast_fp16_to_fp32_dtype_0, x = input_cast_fp16)[name = tensor("cast_433")]; + tensor var_9447 = cast(dtype = var_9447_cast_fp16_to_fp32_dtype_0, x = var_9447_cast_fp16)[name = tensor("cast_434")]; + tensor new_cache_internal_tensor_assign_2 = cast(dtype = reshape_239_cast_fp16_to_fp32_dtype_0, x = reshape_239_cast_fp16)[name = tensor("cast_435")]; + tensor var_9061 = cast(dtype = var_9061_cast_fp16_to_fp32_dtype_0, x = var_9061_cast_fp16)[name = tensor("cast_439")]; + tensor new_cache_45_internal_tensor_assign_2 = cast(dtype = reshape_229_cast_fp16_to_fp32_dtype_0, x = reshape_229_cast_fp16)[name = tensor("cast_440")]; + tensor var_8675 = cast(dtype = var_8675_cast_fp16_to_fp32_dtype_0, x = var_8675_cast_fp16)[name = tensor("cast_444")]; + tensor new_cache_43_internal_tensor_assign_2 = cast(dtype = reshape_219_cast_fp16_to_fp32_dtype_0, x = reshape_219_cast_fp16)[name = tensor("cast_445")]; + tensor var_8289 = cast(dtype = var_8289_cast_fp16_to_fp32_dtype_0, x = var_8289_cast_fp16)[name = tensor("cast_449")]; + tensor new_cache_41_internal_tensor_assign_2 = cast(dtype = reshape_209_cast_fp16_to_fp32_dtype_0, x = reshape_209_cast_fp16)[name = tensor("cast_450")]; + tensor var_7903 = cast(dtype = var_7903_cast_fp16_to_fp32_dtype_0, x = var_7903_cast_fp16)[name = tensor("cast_454")]; + tensor new_cache_39_internal_tensor_assign_2 = cast(dtype = reshape_199_cast_fp16_to_fp32_dtype_0, x = reshape_199_cast_fp16)[name = tensor("cast_455")]; + tensor var_7517 = cast(dtype = var_7517_cast_fp16_to_fp32_dtype_0, x = var_7517_cast_fp16)[name = tensor("cast_459")]; + tensor new_cache_37_internal_tensor_assign_2 = cast(dtype = reshape_189_cast_fp16_to_fp32_dtype_0, x = reshape_189_cast_fp16)[name = tensor("cast_460")]; + tensor var_7131 = cast(dtype = var_7131_cast_fp16_to_fp32_dtype_0, x = var_7131_cast_fp16)[name = tensor("cast_464")]; + tensor new_cache_35_internal_tensor_assign_2 = cast(dtype = reshape_179_cast_fp16_to_fp32_dtype_0, x = reshape_179_cast_fp16)[name = tensor("cast_465")]; + tensor var_6745 = cast(dtype = var_6745_cast_fp16_to_fp32_dtype_0, x = var_6745_cast_fp16)[name = tensor("cast_469")]; + tensor new_cache_33_internal_tensor_assign_2 = cast(dtype = reshape_169_cast_fp16_to_fp32_dtype_0, x = reshape_169_cast_fp16)[name = tensor("cast_470")]; + tensor var_6359 = cast(dtype = var_6359_cast_fp16_to_fp32_dtype_0, x = var_6359_cast_fp16)[name = tensor("cast_474")]; + tensor new_cache_31_internal_tensor_assign_2 = cast(dtype = reshape_159_cast_fp16_to_fp32_dtype_0, x = reshape_159_cast_fp16)[name = tensor("cast_475")]; + tensor var_5973 = cast(dtype = var_5973_cast_fp16_to_fp32_dtype_0, x = var_5973_cast_fp16)[name = tensor("cast_479")]; + tensor new_cache_29_internal_tensor_assign_2 = cast(dtype = reshape_149_cast_fp16_to_fp32_dtype_0, x = reshape_149_cast_fp16)[name = tensor("cast_480")]; + tensor var_5587 = cast(dtype = var_5587_cast_fp16_to_fp32_dtype_0, x = var_5587_cast_fp16)[name = tensor("cast_484")]; + tensor new_cache_27_internal_tensor_assign_2 = cast(dtype = reshape_139_cast_fp16_to_fp32_dtype_0, x = reshape_139_cast_fp16)[name = tensor("cast_485")]; + tensor var_5201 = cast(dtype = var_5201_cast_fp16_to_fp32_dtype_0, x = var_5201_cast_fp16)[name = tensor("cast_489")]; + tensor new_cache_25_internal_tensor_assign_2 = cast(dtype = reshape_129_cast_fp16_to_fp32_dtype_0, x = reshape_129_cast_fp16)[name = tensor("cast_490")]; + tensor var_4815 = cast(dtype = var_4815_cast_fp16_to_fp32_dtype_0, x = var_4815_cast_fp16)[name = tensor("cast_494")]; + tensor new_cache_23_internal_tensor_assign_2 = cast(dtype = reshape_119_cast_fp16_to_fp32_dtype_0, x = reshape_119_cast_fp16)[name = tensor("cast_495")]; + tensor var_4429 = cast(dtype = var_4429_cast_fp16_to_fp32_dtype_0, x = var_4429_cast_fp16)[name = tensor("cast_499")]; + tensor new_cache_21_internal_tensor_assign_2 = cast(dtype = reshape_109_cast_fp16_to_fp32_dtype_0, x = reshape_109_cast_fp16)[name = tensor("cast_500")]; + tensor var_4043 = cast(dtype = var_4043_cast_fp16_to_fp32_dtype_0, x = var_4043_cast_fp16)[name = tensor("cast_504")]; + tensor new_cache_19_internal_tensor_assign_2 = cast(dtype = reshape_99_cast_fp16_to_fp32_dtype_0, x = reshape_99_cast_fp16)[name = tensor("cast_505")]; + tensor var_3657 = cast(dtype = var_3657_cast_fp16_to_fp32_dtype_0, x = var_3657_cast_fp16)[name = tensor("cast_509")]; + tensor new_cache_17_internal_tensor_assign_2 = cast(dtype = reshape_89_cast_fp16_to_fp32_dtype_0, x = reshape_89_cast_fp16)[name = tensor("cast_510")]; + tensor var_3271 = cast(dtype = var_3271_cast_fp16_to_fp32_dtype_0, x = var_3271_cast_fp16)[name = tensor("cast_514")]; + tensor new_cache_15_internal_tensor_assign_2 = cast(dtype = reshape_79_cast_fp16_to_fp32_dtype_0, x = reshape_79_cast_fp16)[name = tensor("cast_515")]; + tensor var_2885 = cast(dtype = var_2885_cast_fp16_to_fp32_dtype_0, x = var_2885_cast_fp16)[name = tensor("cast_519")]; + tensor new_cache_13_internal_tensor_assign_2 = cast(dtype = reshape_69_cast_fp16_to_fp32_dtype_0, x = reshape_69_cast_fp16)[name = tensor("cast_520")]; + tensor var_2499 = cast(dtype = var_2499_cast_fp16_to_fp32_dtype_0, x = var_2499_cast_fp16)[name = tensor("cast_524")]; + tensor new_cache_11_internal_tensor_assign_2 = cast(dtype = reshape_59_cast_fp16_to_fp32_dtype_0, x = reshape_59_cast_fp16)[name = tensor("cast_525")]; + tensor var_2113 = cast(dtype = var_2113_cast_fp16_to_fp32_dtype_0, x = var_2113_cast_fp16)[name = tensor("cast_529")]; + tensor new_cache_9_internal_tensor_assign_2 = cast(dtype = reshape_49_cast_fp16_to_fp32_dtype_0, x = reshape_49_cast_fp16)[name = tensor("cast_530")]; + tensor var_1727 = cast(dtype = var_1727_cast_fp16_to_fp32_dtype_0, x = var_1727_cast_fp16)[name = tensor("cast_534")]; + tensor new_cache_7_internal_tensor_assign_2 = cast(dtype = reshape_39_cast_fp16_to_fp32_dtype_0, x = reshape_39_cast_fp16)[name = tensor("cast_535")]; + tensor var_1341 = cast(dtype = var_1341_cast_fp16_to_fp32_dtype_0, x = var_1341_cast_fp16)[name = tensor("cast_539")]; + tensor new_cache_5_internal_tensor_assign_2 = cast(dtype = reshape_29_cast_fp16_to_fp32_dtype_0, x = reshape_29_cast_fp16)[name = tensor("cast_540")]; + tensor var_955 = cast(dtype = var_955_cast_fp16_to_fp32_dtype_0, x = var_955_cast_fp16)[name = tensor("cast_544")]; + tensor new_cache_3_internal_tensor_assign_2 = cast(dtype = reshape_19_cast_fp16_to_fp32_dtype_0, x = reshape_19_cast_fp16)[name = tensor("cast_545")]; + tensor var_569 = cast(dtype = var_569_cast_fp16_to_fp32_dtype_0, x = var_569_cast_fp16)[name = tensor("cast_549")]; + tensor new_cache_1_internal_tensor_assign_2 = cast(dtype = reshape_9_cast_fp16_to_fp32_dtype_0, x = reshape_9_cast_fp16)[name = tensor("cast_550")]; + } -> (input, var_9476, new_cache_1_internal_tensor_assign_2, var_569, new_cache_3_internal_tensor_assign_2, var_955, new_cache_5_internal_tensor_assign_2, var_1341, new_cache_7_internal_tensor_assign_2, var_1727, new_cache_9_internal_tensor_assign_2, var_2113, new_cache_11_internal_tensor_assign_2, var_2499, new_cache_13_internal_tensor_assign_2, var_2885, new_cache_15_internal_tensor_assign_2, var_3271, new_cache_17_internal_tensor_assign_2, var_3657, new_cache_19_internal_tensor_assign_2, var_4043, new_cache_21_internal_tensor_assign_2, var_4429, new_cache_23_internal_tensor_assign_2, var_4815, new_cache_25_internal_tensor_assign_2, var_5201, new_cache_27_internal_tensor_assign_2, var_5587, new_cache_29_internal_tensor_assign_2, var_5973, new_cache_31_internal_tensor_assign_2, var_6359, new_cache_33_internal_tensor_assign_2, var_6745, new_cache_35_internal_tensor_assign_2, var_7131, new_cache_37_internal_tensor_assign_2, var_7517, new_cache_39_internal_tensor_assign_2, var_7903, new_cache_41_internal_tensor_assign_2, var_8289, new_cache_43_internal_tensor_assign_2, var_8675, new_cache_45_internal_tensor_assign_2, var_9061, new_cache_internal_tensor_assign_2, var_9447); +} \ No newline at end of file diff --git a/v2.1/german_24l/flowlm_step.mlmodelc/weights/weight.bin b/v2.1/german_24l/flowlm_step.mlmodelc/weights/weight.bin new file mode 100644 index 0000000000000000000000000000000000000000..570efac790af22fd6ac0344b178d1b6fa23cfc57 --- /dev/null +++ b/v2.1/german_24l/flowlm_step.mlmodelc/weights/weight.bin @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:4b61da154a38d5d675b44ad2c7d0587092f2ad343eeca15e85d8df594108db53 +size 605330048 diff --git a/v2.1/german_24l/flowlm_step.mlpackage/Data/com.apple.CoreML/model.mlmodel b/v2.1/german_24l/flowlm_step.mlpackage/Data/com.apple.CoreML/model.mlmodel new file mode 100644 index 0000000000000000000000000000000000000000..484699de0bd8e94c2aba7407520c02b2849ad913 --- /dev/null +++ b/v2.1/german_24l/flowlm_step.mlpackage/Data/com.apple.CoreML/model.mlmodel @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:7ec941a8fed230310364fa21dfa969d7b9bc8209e1fb8b91e7ade5c234b2ff19 +size 825108 diff --git a/v2.1/german_24l/flowlm_step.mlpackage/Data/com.apple.CoreML/weights/weight.bin b/v2.1/german_24l/flowlm_step.mlpackage/Data/com.apple.CoreML/weights/weight.bin new file mode 100644 index 0000000000000000000000000000000000000000..570efac790af22fd6ac0344b178d1b6fa23cfc57 --- /dev/null +++ b/v2.1/german_24l/flowlm_step.mlpackage/Data/com.apple.CoreML/weights/weight.bin @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:4b61da154a38d5d675b44ad2c7d0587092f2ad343eeca15e85d8df594108db53 +size 605330048 diff --git a/v2.1/german_24l/flowlm_step.mlpackage/Manifest.json b/v2.1/german_24l/flowlm_step.mlpackage/Manifest.json new file mode 100644 index 0000000000000000000000000000000000000000..94900ed4dcfa407d571339e998279637bac9f80d --- /dev/null +++ b/v2.1/german_24l/flowlm_step.mlpackage/Manifest.json @@ -0,0 +1,18 @@ +{ + "fileFormatVersion": "1.0.0", + "itemInfoEntries": { + "8CD97442-3981-49B7-A039-DFDEB3A592D7": { + "author": "com.apple.CoreML", + "description": "CoreML Model Weights", + "name": "weights", + "path": "com.apple.CoreML/weights" + }, + "F5C49884-BE55-4555-9CA0-BC821512FC51": { + "author": "com.apple.CoreML", + "description": "CoreML Model Specification", + "name": "model.mlmodel", + "path": "com.apple.CoreML/model.mlmodel" + } + }, + "rootModelIdentifier": "F5C49884-BE55-4555-9CA0-BC821512FC51" +} diff --git a/v2.1/german_24l/flowlm_stepv2.mlmodelc/analytics/coremldata.bin b/v2.1/german_24l/flowlm_stepv2.mlmodelc/analytics/coremldata.bin new file mode 100644 index 0000000000000000000000000000000000000000..efd07b3ae76e57c00cc9764f45785c9b68cb6450 --- /dev/null +++ b/v2.1/german_24l/flowlm_stepv2.mlmodelc/analytics/coremldata.bin @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:67adbead75be4c07caf5958c38f6fa84c90953ad0c63164bbeb61b3e02fa3ff4 +size 243 diff --git a/v2.1/german_24l/flowlm_stepv2.mlmodelc/coremldata.bin b/v2.1/german_24l/flowlm_stepv2.mlmodelc/coremldata.bin new file mode 100644 index 0000000000000000000000000000000000000000..7c014fabfc71894b9a981f302353186c3ded37c8 --- /dev/null +++ b/v2.1/german_24l/flowlm_stepv2.mlmodelc/coremldata.bin @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:ced1a4c3457c5611b38104f42313bd526fef2c53682bd8ad8cf88f0385b7c62e +size 3561 diff --git a/v2.1/german_24l/flowlm_stepv2.mlmodelc/metadata.json b/v2.1/german_24l/flowlm_stepv2.mlmodelc/metadata.json new file mode 100644 index 0000000000000000000000000000000000000000..629da9a69819375413a6c11a7789eb2ca4f04e28 --- /dev/null +++ b/v2.1/german_24l/flowlm_stepv2.mlmodelc/metadata.json @@ -0,0 +1,1068 @@ +[ + { + "metadataOutputVersion" : "3.0", + "storagePrecision" : "Mixed (Float16, Int8)", + "outputSchema" : [ + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1 × 1 × 1024)", + "shortDescription" : "", + "shape" : "[1, 1, 1024]", + "name" : "input", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1 × 1 × 1)", + "shortDescription" : "", + "shape" : "[1, 1, 1]", + "name" : "var_10268", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 2 × 1 × 512 × 16 × 64)", + "shortDescription" : "", + "shape" : "[2, 1, 512, 16, 64]", + "name" : "new_cache_1_internal_tensor_assign_2", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1)", + "shortDescription" : "", + "shape" : "[1]", + "name" : "var_602", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 2 × 1 × 512 × 16 × 64)", + "shortDescription" : "", + "shape" : "[2, 1, 512, 16, 64]", + "name" : "new_cache_3_internal_tensor_assign_2", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1)", + "shortDescription" : "", + "shape" : "[1]", + "name" : "var_1021", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 2 × 1 × 512 × 16 × 64)", + "shortDescription" : "", + "shape" : "[2, 1, 512, 16, 64]", + "name" : "new_cache_5_internal_tensor_assign_2", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1)", + "shortDescription" : "", + "shape" : "[1]", + "name" : "var_1440", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 2 × 1 × 512 × 16 × 64)", + "shortDescription" : "", + "shape" : "[2, 1, 512, 16, 64]", + "name" : "new_cache_7_internal_tensor_assign_2", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1)", + "shortDescription" : "", + "shape" : "[1]", + "name" : "var_1859", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 2 × 1 × 512 × 16 × 64)", + "shortDescription" : "", + "shape" : "[2, 1, 512, 16, 64]", + "name" : "new_cache_9_internal_tensor_assign_2", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1)", + "shortDescription" : "", + "shape" : "[1]", + "name" : "var_2278", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 2 × 1 × 512 × 16 × 64)", + "shortDescription" : "", + "shape" : "[2, 1, 512, 16, 64]", + "name" : "new_cache_11_internal_tensor_assign_2", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1)", + "shortDescription" : "", + "shape" : "[1]", + "name" : "var_2697", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 2 × 1 × 512 × 16 × 64)", + "shortDescription" : "", + "shape" : "[2, 1, 512, 16, 64]", + "name" : "new_cache_13_internal_tensor_assign_2", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1)", + "shortDescription" : "", + "shape" : "[1]", + "name" : "var_3116", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 2 × 1 × 512 × 16 × 64)", + "shortDescription" : "", + "shape" : "[2, 1, 512, 16, 64]", + "name" : "new_cache_15_internal_tensor_assign_2", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1)", + "shortDescription" : "", + "shape" : "[1]", + "name" : "var_3535", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 2 × 1 × 512 × 16 × 64)", + "shortDescription" : "", + "shape" : "[2, 1, 512, 16, 64]", + "name" : "new_cache_17_internal_tensor_assign_2", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1)", + "shortDescription" : "", + "shape" : "[1]", + "name" : "var_3954", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 2 × 1 × 512 × 16 × 64)", + "shortDescription" : "", + "shape" : "[2, 1, 512, 16, 64]", + "name" : "new_cache_19_internal_tensor_assign_2", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1)", + "shortDescription" : "", + "shape" : "[1]", + "name" : "var_4373", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 2 × 1 × 512 × 16 × 64)", + "shortDescription" : "", + "shape" : "[2, 1, 512, 16, 64]", + "name" : "new_cache_21_internal_tensor_assign_2", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1)", + "shortDescription" : "", + "shape" : "[1]", + "name" : "var_4792", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 2 × 1 × 512 × 16 × 64)", + "shortDescription" : "", + "shape" : "[2, 1, 512, 16, 64]", + "name" : "new_cache_23_internal_tensor_assign_2", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1)", + "shortDescription" : "", + "shape" : "[1]", + "name" : "var_5211", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 2 × 1 × 512 × 16 × 64)", + "shortDescription" : "", + "shape" : "[2, 1, 512, 16, 64]", + "name" : "new_cache_25_internal_tensor_assign_2", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1)", + "shortDescription" : "", + "shape" : "[1]", + "name" : "var_5630", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 2 × 1 × 512 × 16 × 64)", + "shortDescription" : "", + "shape" : "[2, 1, 512, 16, 64]", + "name" : "new_cache_27_internal_tensor_assign_2", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1)", + "shortDescription" : "", + "shape" : "[1]", + "name" : "var_6049", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 2 × 1 × 512 × 16 × 64)", + "shortDescription" : "", + "shape" : "[2, 1, 512, 16, 64]", + "name" : "new_cache_29_internal_tensor_assign_2", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1)", + "shortDescription" : "", + "shape" : "[1]", + "name" : "var_6468", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 2 × 1 × 512 × 16 × 64)", + "shortDescription" : "", + "shape" : "[2, 1, 512, 16, 64]", + "name" : "new_cache_31_internal_tensor_assign_2", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1)", + "shortDescription" : "", + "shape" : "[1]", + "name" : "var_6887", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 2 × 1 × 512 × 16 × 64)", + "shortDescription" : "", + "shape" : "[2, 1, 512, 16, 64]", + "name" : "new_cache_33_internal_tensor_assign_2", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1)", + "shortDescription" : "", + "shape" : "[1]", + "name" : "var_7306", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 2 × 1 × 512 × 16 × 64)", + "shortDescription" : "", + "shape" : "[2, 1, 512, 16, 64]", + "name" : "new_cache_35_internal_tensor_assign_2", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1)", + "shortDescription" : "", + "shape" : "[1]", + "name" : "var_7725", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 2 × 1 × 512 × 16 × 64)", + "shortDescription" : "", + "shape" : "[2, 1, 512, 16, 64]", + "name" : "new_cache_37_internal_tensor_assign_2", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1)", + "shortDescription" : "", + "shape" : "[1]", + "name" : "var_8144", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 2 × 1 × 512 × 16 × 64)", + "shortDescription" : "", + "shape" : "[2, 1, 512, 16, 64]", + "name" : "new_cache_39_internal_tensor_assign_2", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1)", + "shortDescription" : "", + "shape" : "[1]", + "name" : "var_8563", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 2 × 1 × 512 × 16 × 64)", + "shortDescription" : "", + "shape" : "[2, 1, 512, 16, 64]", + "name" : "new_cache_41_internal_tensor_assign_2", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1)", + "shortDescription" : "", + "shape" : "[1]", + "name" : "var_8982", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 2 × 1 × 512 × 16 × 64)", + "shortDescription" : "", + "shape" : "[2, 1, 512, 16, 64]", + "name" : "new_cache_43_internal_tensor_assign_2", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1)", + "shortDescription" : "", + "shape" : "[1]", + "name" : "var_9401", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 2 × 1 × 512 × 16 × 64)", + "shortDescription" : "", + "shape" : "[2, 1, 512, 16, 64]", + "name" : "new_cache_45_internal_tensor_assign_2", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1)", + "shortDescription" : "", + "shape" : "[1]", + "name" : "var_9820", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 2 × 1 × 512 × 16 × 64)", + "shortDescription" : "", + "shape" : "[2, 1, 512, 16, 64]", + "name" : "new_cache_internal_tensor_assign_2", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1)", + "shortDescription" : "", + "shape" : "[1]", + "name" : "var_10239", + "type" : "MultiArray" + } + ], + "modelParameters" : [ + + ], + "specificationVersion" : 8, + "mlProgramOperationTypeHistogram" : { + "Ios17.cos" : 24, + "Range1d" : 48, + "Ios17.notEqual" : 49, + "Ios17.floor" : 24, + "Ios17.reshape" : 432, + "Ios17.logicalAnd" : 24, + "Ios16.softmax" : 24, + "Ios17.scatter" : 48, + "Ios17.matmul" : 48, + "Ios17.sin" : 24, + "Ios17.transpose" : 96, + "Select" : 73, + "Ios17.expandDims" : 25, + "Ios17.add" : 144, + "Tile" : 24, + "Ios17.sliceByIndex" : 312, + "Ios17.lessEqual" : 24, + "Ios17.scatterAlongAxis" : 48, + "Ios17.logicalNot" : 24, + "Ios17.layerNorm" : 49, + "Ios16.constexprAffineDequantize" : 96, + "Ios17.cast" : 124, + "Ios17.sub" : 72, + "Ios17.less" : 24, + "Ios16.gelu" : 24, + "Ios17.linear" : 98, + "Stack" : 48, + "Ios17.mul" : 288 + }, + "computePrecision" : "Mixed (Float16, Float32, Int32)", + "isUpdatable" : "0", + "stateSchema" : [ + + ], + "availability" : { + "macOS" : "14.0", + "tvOS" : "17.0", + "visionOS" : "1.0", + "watchOS" : "10.0", + "iOS" : "17.0", + "macCatalyst" : "17.0" + }, + "modelType" : { + "name" : "MLModelType_mlProgram" + }, + "userDefinedMetadata" : { + "com.github.apple.coremltools.conversion_date" : "2026-04-27", + "com.github.apple.coremltools.source" : "torch==2.9.1", + "com.github.apple.coremltools.version" : "9.0", + "com.github.apple.coremltools.source_dialect" : "TorchScript" + }, + "inputSchema" : [ + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1 × 1 × 32)", + "shortDescription" : "", + "shape" : "[1, 1, 32]", + "name" : "sequence", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 32)", + "shortDescription" : "", + "shape" : "[32]", + "name" : "bos_emb", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 2 × 1 × 512 × 16 × 64)", + "shortDescription" : "", + "shape" : "[2, 1, 512, 16, 64]", + "name" : "cache0", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1)", + "shortDescription" : "", + "shape" : "[1]", + "name" : "position0", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 2 × 1 × 512 × 16 × 64)", + "shortDescription" : "", + "shape" : "[2, 1, 512, 16, 64]", + "name" : "cache1", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1)", + "shortDescription" : "", + "shape" : "[1]", + "name" : "position1", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 2 × 1 × 512 × 16 × 64)", + "shortDescription" : "", + "shape" : "[2, 1, 512, 16, 64]", + "name" : "cache2", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1)", + "shortDescription" : "", + "shape" : "[1]", + "name" : "position2", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 2 × 1 × 512 × 16 × 64)", + "shortDescription" : "", + "shape" : "[2, 1, 512, 16, 64]", + "name" : "cache3", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1)", + "shortDescription" : "", + "shape" : "[1]", + "name" : "position3", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 2 × 1 × 512 × 16 × 64)", + "shortDescription" : "", + "shape" : "[2, 1, 512, 16, 64]", + "name" : "cache4", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1)", + "shortDescription" : "", + "shape" : "[1]", + "name" : "position4", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 2 × 1 × 512 × 16 × 64)", + "shortDescription" : "", + "shape" : "[2, 1, 512, 16, 64]", + "name" : "cache5", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1)", + "shortDescription" : "", + "shape" : "[1]", + "name" : "position5", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 2 × 1 × 512 × 16 × 64)", + "shortDescription" : "", + "shape" : "[2, 1, 512, 16, 64]", + "name" : "cache6", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1)", + "shortDescription" : "", + "shape" : "[1]", + "name" : "position6", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 2 × 1 × 512 × 16 × 64)", + "shortDescription" : "", + "shape" : "[2, 1, 512, 16, 64]", + "name" : "cache7", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1)", + "shortDescription" : "", + "shape" : "[1]", + "name" : "position7", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 2 × 1 × 512 × 16 × 64)", + "shortDescription" : "", + "shape" : "[2, 1, 512, 16, 64]", + "name" : "cache8", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1)", + "shortDescription" : "", + "shape" : "[1]", + "name" : "position8", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 2 × 1 × 512 × 16 × 64)", + "shortDescription" : "", + "shape" : "[2, 1, 512, 16, 64]", + "name" : "cache9", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1)", + "shortDescription" : "", + "shape" : "[1]", + "name" : "position9", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 2 × 1 × 512 × 16 × 64)", + "shortDescription" : "", + "shape" : "[2, 1, 512, 16, 64]", + "name" : "cache10", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1)", + "shortDescription" : "", + "shape" : "[1]", + "name" : "position10", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 2 × 1 × 512 × 16 × 64)", + "shortDescription" : "", + "shape" : "[2, 1, 512, 16, 64]", + "name" : "cache11", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1)", + "shortDescription" : "", + "shape" : "[1]", + "name" : "position11", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 2 × 1 × 512 × 16 × 64)", + "shortDescription" : "", + "shape" : "[2, 1, 512, 16, 64]", + "name" : "cache12", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1)", + "shortDescription" : "", + "shape" : "[1]", + "name" : "position12", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 2 × 1 × 512 × 16 × 64)", + "shortDescription" : "", + "shape" : "[2, 1, 512, 16, 64]", + "name" : "cache13", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1)", + "shortDescription" : "", + "shape" : "[1]", + "name" : "position13", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 2 × 1 × 512 × 16 × 64)", + "shortDescription" : "", + "shape" : "[2, 1, 512, 16, 64]", + "name" : "cache14", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1)", + "shortDescription" : "", + "shape" : "[1]", + "name" : "position14", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 2 × 1 × 512 × 16 × 64)", + "shortDescription" : "", + "shape" : "[2, 1, 512, 16, 64]", + "name" : "cache15", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1)", + "shortDescription" : "", + "shape" : "[1]", + "name" : "position15", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 2 × 1 × 512 × 16 × 64)", + "shortDescription" : "", + "shape" : "[2, 1, 512, 16, 64]", + "name" : "cache16", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1)", + "shortDescription" : "", + "shape" : "[1]", + "name" : "position16", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 2 × 1 × 512 × 16 × 64)", + "shortDescription" : "", + "shape" : "[2, 1, 512, 16, 64]", + "name" : "cache17", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1)", + "shortDescription" : "", + "shape" : "[1]", + "name" : "position17", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 2 × 1 × 512 × 16 × 64)", + "shortDescription" : "", + "shape" : "[2, 1, 512, 16, 64]", + "name" : "cache18", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1)", + "shortDescription" : "", + "shape" : "[1]", + "name" : "position18", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 2 × 1 × 512 × 16 × 64)", + "shortDescription" : "", + "shape" : "[2, 1, 512, 16, 64]", + "name" : "cache19", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1)", + "shortDescription" : "", + "shape" : "[1]", + "name" : "position19", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 2 × 1 × 512 × 16 × 64)", + "shortDescription" : "", + "shape" : "[2, 1, 512, 16, 64]", + "name" : "cache20", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1)", + "shortDescription" : "", + "shape" : "[1]", + "name" : "position20", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 2 × 1 × 512 × 16 × 64)", + "shortDescription" : "", + "shape" : "[2, 1, 512, 16, 64]", + "name" : "cache21", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1)", + "shortDescription" : "", + "shape" : "[1]", + "name" : "position21", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 2 × 1 × 512 × 16 × 64)", + "shortDescription" : "", + "shape" : "[2, 1, 512, 16, 64]", + "name" : "cache22", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1)", + "shortDescription" : "", + "shape" : "[1]", + "name" : "position22", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 2 × 1 × 512 × 16 × 64)", + "shortDescription" : "", + "shape" : "[2, 1, 512, 16, 64]", + "name" : "cache23", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1)", + "shortDescription" : "", + "shape" : "[1]", + "name" : "position23", + "type" : "MultiArray" + } + ], + "generatedClassName" : "flowlm_stepv2", + "method" : "predict" + } +] \ No newline at end of file diff --git a/v2.1/german_24l/flowlm_stepv2.mlmodelc/model.mil b/v2.1/german_24l/flowlm_stepv2.mlmodelc/model.mil new file mode 100644 index 0000000000000000000000000000000000000000..3bb20d21231a3f663b9208cf7be7d9c738cefecc --- /dev/null +++ b/v2.1/german_24l/flowlm_stepv2.mlmodelc/model.mil @@ -0,0 +1,5337 @@ +program(1.0) +[buildInfo = dict, tensor>({{"coremlc-component-MIL", "3520.4.1"}, {"coremlc-version", "3520.5.1"}, {"coremltools-component-torch", "2.9.1"}, {"coremltools-source-dialect", "TorchScript"}, {"coremltools-version", "9.0"}})] +{ + func main(tensor bos_emb, tensor cache0, tensor cache1, tensor cache10, tensor cache11, tensor cache12, tensor cache13, tensor cache14, tensor cache15, tensor cache16, tensor cache17, tensor cache18, tensor cache19, tensor cache2, tensor cache20, tensor cache21, tensor cache22, tensor cache23, tensor cache3, tensor cache4, tensor cache5, tensor cache6, tensor cache7, tensor cache8, tensor cache9, tensor position0, tensor position1, tensor position10, tensor position11, tensor position12, tensor position13, tensor position14, tensor position15, tensor position16, tensor position17, tensor position18, tensor position19, tensor position2, tensor position20, tensor position21, tensor position22, tensor position23, tensor position3, tensor position4, tensor position5, tensor position6, tensor position7, tensor position8, tensor position9, tensor sequence) { + tensor attn0_in_proj_weight_cast_fp16 = constexpr_affine_dequantize()[axis = tensor(0), name = tensor("attn0_in_proj_weight_cast_fp16"), quantized_data = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(64))), scale = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(3148992))), zero_point = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(3145856)))]; + tensor attn0_out_proj_weight_cast_fp16 = constexpr_affine_dequantize()[axis = tensor(0), name = tensor("attn0_out_proj_weight_cast_fp16"), quantized_data = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(3155200))), scale = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(4204928))), zero_point = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(4203840)))]; + tensor linear0_1_weight_cast_fp16 = constexpr_affine_dequantize()[axis = tensor(0), name = tensor("linear0_1_weight_cast_fp16"), quantized_data = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(4207040))), scale = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8405568))), zero_point = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8401408)))]; + tensor linear0_2_weight_cast_fp16 = constexpr_affine_dequantize()[axis = tensor(0), name = tensor("linear0_2_weight_cast_fp16"), quantized_data = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8413824))), scale = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(12608192))), zero_point = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(4203840)))]; + tensor attn1_in_proj_weight_cast_fp16 = constexpr_affine_dequantize()[axis = tensor(0), name = tensor("attn1_in_proj_weight_cast_fp16"), quantized_data = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(12610304))), scale = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15756096))), zero_point = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(3145856)))]; + tensor attn1_out_proj_weight_cast_fp16 = constexpr_affine_dequantize()[axis = tensor(0), name = tensor("attn1_out_proj_weight_cast_fp16"), quantized_data = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15762304))), scale = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(16810944))), zero_point = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(4203840)))]; + tensor linear1_1_weight_cast_fp16 = constexpr_affine_dequantize()[axis = tensor(0), name = tensor("linear1_1_weight_cast_fp16"), quantized_data = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(16813056))), scale = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21007424))), zero_point = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8401408)))]; + tensor linear1_2_weight_cast_fp16 = constexpr_affine_dequantize()[axis = tensor(0), name = tensor("linear1_2_weight_cast_fp16"), quantized_data = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21015680))), scale = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(25210048))), zero_point = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(4203840)))]; + tensor attn2_in_proj_weight_cast_fp16 = constexpr_affine_dequantize()[axis = tensor(0), name = tensor("attn2_in_proj_weight_cast_fp16"), quantized_data = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(25212160))), scale = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28357952))), zero_point = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(3145856)))]; + tensor attn2_out_proj_weight_cast_fp16 = constexpr_affine_dequantize()[axis = tensor(0), name = tensor("attn2_out_proj_weight_cast_fp16"), quantized_data = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28364160))), scale = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(29412800))), zero_point = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(4203840)))]; + tensor linear2_1_weight_cast_fp16 = constexpr_affine_dequantize()[axis = tensor(0), name = tensor("linear2_1_weight_cast_fp16"), quantized_data = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(29414912))), scale = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(33609280))), zero_point = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8401408)))]; + tensor linear2_2_weight_cast_fp16 = constexpr_affine_dequantize()[axis = tensor(0), name = tensor("linear2_2_weight_cast_fp16"), quantized_data = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(33617536))), scale = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(37811904))), zero_point = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(4203840)))]; + tensor attn3_in_proj_weight_cast_fp16 = constexpr_affine_dequantize()[axis = tensor(0), name = tensor("attn3_in_proj_weight_cast_fp16"), quantized_data = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(37814016))), scale = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(40959808))), zero_point = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(3145856)))]; + tensor attn3_out_proj_weight_cast_fp16 = constexpr_affine_dequantize()[axis = tensor(0), name = tensor("attn3_out_proj_weight_cast_fp16"), quantized_data = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(40966016))), scale = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(42014656))), zero_point = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(4203840)))]; + tensor linear3_1_weight_cast_fp16 = constexpr_affine_dequantize()[axis = tensor(0), name = tensor("linear3_1_weight_cast_fp16"), quantized_data = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(42016768))), scale = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(46211136))), zero_point = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8401408)))]; + tensor linear3_2_weight_cast_fp16 = constexpr_affine_dequantize()[axis = tensor(0), name = tensor("linear3_2_weight_cast_fp16"), quantized_data = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(46219392))), scale = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(50413760))), zero_point = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(4203840)))]; + tensor attn4_in_proj_weight_cast_fp16 = constexpr_affine_dequantize()[axis = tensor(0), name = tensor("attn4_in_proj_weight_cast_fp16"), quantized_data = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(50415872))), scale = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(53561664))), zero_point = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(3145856)))]; + tensor attn4_out_proj_weight_cast_fp16 = constexpr_affine_dequantize()[axis = tensor(0), name = tensor("attn4_out_proj_weight_cast_fp16"), quantized_data = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(53567872))), scale = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(54616512))), zero_point = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(4203840)))]; + tensor linear4_1_weight_cast_fp16 = constexpr_affine_dequantize()[axis = tensor(0), name = tensor("linear4_1_weight_cast_fp16"), quantized_data = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(54618624))), scale = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(58812992))), zero_point = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8401408)))]; + tensor linear4_2_weight_cast_fp16 = constexpr_affine_dequantize()[axis = tensor(0), name = tensor("linear4_2_weight_cast_fp16"), quantized_data = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(58821248))), scale = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(63015616))), zero_point = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(4203840)))]; + tensor attn5_in_proj_weight_cast_fp16 = constexpr_affine_dequantize()[axis = tensor(0), name = tensor("attn5_in_proj_weight_cast_fp16"), quantized_data = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(63017728))), scale = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(66163520))), zero_point = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(3145856)))]; + tensor attn5_out_proj_weight_cast_fp16 = constexpr_affine_dequantize()[axis = tensor(0), name = tensor("attn5_out_proj_weight_cast_fp16"), quantized_data = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(66169728))), scale = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(67218368))), zero_point = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(4203840)))]; + tensor linear5_1_weight_cast_fp16 = constexpr_affine_dequantize()[axis = tensor(0), name = tensor("linear5_1_weight_cast_fp16"), quantized_data = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(67220480))), scale = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(71414848))), zero_point = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8401408)))]; + tensor linear5_2_weight_cast_fp16 = constexpr_affine_dequantize()[axis = tensor(0), name = tensor("linear5_2_weight_cast_fp16"), quantized_data = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(71423104))), scale = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(75617472))), zero_point = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(4203840)))]; + tensor attn6_in_proj_weight_cast_fp16 = constexpr_affine_dequantize()[axis = tensor(0), name = tensor("attn6_in_proj_weight_cast_fp16"), quantized_data = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(75619584))), scale = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(78765376))), zero_point = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(3145856)))]; + tensor attn6_out_proj_weight_cast_fp16 = constexpr_affine_dequantize()[axis = tensor(0), name = tensor("attn6_out_proj_weight_cast_fp16"), quantized_data = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(78771584))), scale = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(79820224))), zero_point = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(4203840)))]; + tensor linear6_1_weight_cast_fp16 = constexpr_affine_dequantize()[axis = tensor(0), name = tensor("linear6_1_weight_cast_fp16"), quantized_data = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(79822336))), scale = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(84016704))), zero_point = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8401408)))]; + tensor linear6_2_weight_cast_fp16 = constexpr_affine_dequantize()[axis = tensor(0), name = tensor("linear6_2_weight_cast_fp16"), quantized_data = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(84024960))), scale = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(88219328))), zero_point = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(4203840)))]; + tensor attn7_in_proj_weight_cast_fp16 = constexpr_affine_dequantize()[axis = tensor(0), name = tensor("attn7_in_proj_weight_cast_fp16"), quantized_data = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(88221440))), scale = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(91367232))), zero_point = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(3145856)))]; + tensor attn7_out_proj_weight_cast_fp16 = constexpr_affine_dequantize()[axis = tensor(0), name = tensor("attn7_out_proj_weight_cast_fp16"), quantized_data = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(91373440))), scale = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(92422080))), zero_point = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(4203840)))]; + tensor linear7_1_weight_cast_fp16 = constexpr_affine_dequantize()[axis = tensor(0), name = tensor("linear7_1_weight_cast_fp16"), quantized_data = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(92424192))), scale = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(96618560))), zero_point = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8401408)))]; + tensor linear7_2_weight_cast_fp16 = constexpr_affine_dequantize()[axis = tensor(0), name = tensor("linear7_2_weight_cast_fp16"), quantized_data = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(96626816))), scale = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(100821184))), zero_point = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(4203840)))]; + tensor attn8_in_proj_weight_cast_fp16 = constexpr_affine_dequantize()[axis = tensor(0), name = tensor("attn8_in_proj_weight_cast_fp16"), quantized_data = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(100823296))), scale = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(103969088))), zero_point = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(3145856)))]; + tensor attn8_out_proj_weight_cast_fp16 = constexpr_affine_dequantize()[axis = tensor(0), name = tensor("attn8_out_proj_weight_cast_fp16"), quantized_data = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(103975296))), scale = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(105023936))), zero_point = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(4203840)))]; + tensor linear8_1_weight_cast_fp16 = constexpr_affine_dequantize()[axis = tensor(0), name = tensor("linear8_1_weight_cast_fp16"), quantized_data = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(105026048))), scale = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(109220416))), zero_point = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8401408)))]; + tensor linear8_2_weight_cast_fp16 = constexpr_affine_dequantize()[axis = tensor(0), name = tensor("linear8_2_weight_cast_fp16"), quantized_data = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(109228672))), scale = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(113423040))), zero_point = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(4203840)))]; + tensor attn9_in_proj_weight_cast_fp16 = constexpr_affine_dequantize()[axis = tensor(0), name = tensor("attn9_in_proj_weight_cast_fp16"), quantized_data = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(113425152))), scale = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(116570944))), zero_point = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(3145856)))]; + tensor attn9_out_proj_weight_cast_fp16 = constexpr_affine_dequantize()[axis = tensor(0), name = tensor("attn9_out_proj_weight_cast_fp16"), quantized_data = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(116577152))), scale = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(117625792))), zero_point = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(4203840)))]; + tensor linear9_1_weight_cast_fp16 = constexpr_affine_dequantize()[axis = tensor(0), name = tensor("linear9_1_weight_cast_fp16"), quantized_data = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(117627904))), scale = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(121822272))), zero_point = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8401408)))]; + tensor linear9_2_weight_cast_fp16 = constexpr_affine_dequantize()[axis = tensor(0), name = tensor("linear9_2_weight_cast_fp16"), quantized_data = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(121830528))), scale = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(126024896))), zero_point = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(4203840)))]; + tensor attn10_in_proj_weight_cast_fp16 = constexpr_affine_dequantize()[axis = tensor(0), name = tensor("attn10_in_proj_weight_cast_fp16"), quantized_data = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(126027008))), scale = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(129172800))), zero_point = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(3145856)))]; + tensor attn10_out_proj_weight_cast_fp16 = constexpr_affine_dequantize()[axis = tensor(0), name = tensor("attn10_out_proj_weight_cast_fp16"), quantized_data = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(129179008))), scale = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(130227648))), zero_point = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(4203840)))]; + tensor linear10_1_weight_cast_fp16 = constexpr_affine_dequantize()[axis = tensor(0), name = tensor("linear10_1_weight_cast_fp16"), quantized_data = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(130229760))), scale = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(134424128))), zero_point = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8401408)))]; + tensor linear10_2_weight_cast_fp16 = constexpr_affine_dequantize()[axis = tensor(0), name = tensor("linear10_2_weight_cast_fp16"), quantized_data = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(134432384))), scale = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(138626752))), zero_point = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(4203840)))]; + tensor attn11_in_proj_weight_cast_fp16 = constexpr_affine_dequantize()[axis = tensor(0), name = tensor("attn11_in_proj_weight_cast_fp16"), quantized_data = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(138628864))), scale = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(141774656))), zero_point = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(3145856)))]; + tensor attn11_out_proj_weight_cast_fp16 = constexpr_affine_dequantize()[axis = tensor(0), name = tensor("attn11_out_proj_weight_cast_fp16"), quantized_data = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(141780864))), scale = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(142829504))), zero_point = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(4203840)))]; + tensor linear11_1_weight_cast_fp16 = constexpr_affine_dequantize()[axis = tensor(0), name = tensor("linear11_1_weight_cast_fp16"), quantized_data = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(142831616))), scale = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(147025984))), zero_point = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8401408)))]; + tensor linear11_2_weight_cast_fp16 = constexpr_affine_dequantize()[axis = tensor(0), name = tensor("linear11_2_weight_cast_fp16"), quantized_data = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(147034240))), scale = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(151228608))), zero_point = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(4203840)))]; + tensor attn12_in_proj_weight_cast_fp16 = constexpr_affine_dequantize()[axis = tensor(0), name = tensor("attn12_in_proj_weight_cast_fp16"), quantized_data = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(151230720))), scale = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(154376512))), zero_point = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(3145856)))]; + tensor attn12_out_proj_weight_cast_fp16 = constexpr_affine_dequantize()[axis = tensor(0), name = tensor("attn12_out_proj_weight_cast_fp16"), quantized_data = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(154382720))), scale = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(155431360))), zero_point = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(4203840)))]; + tensor linear12_1_weight_cast_fp16 = constexpr_affine_dequantize()[axis = tensor(0), name = tensor("linear12_1_weight_cast_fp16"), quantized_data = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(155433472))), scale = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(159627840))), zero_point = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8401408)))]; + tensor linear12_2_weight_cast_fp16 = constexpr_affine_dequantize()[axis = tensor(0), name = tensor("linear12_2_weight_cast_fp16"), quantized_data = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(159636096))), scale = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(163830464))), zero_point = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(4203840)))]; + tensor attn13_in_proj_weight_cast_fp16 = constexpr_affine_dequantize()[axis = tensor(0), name = tensor("attn13_in_proj_weight_cast_fp16"), quantized_data = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(163832576))), scale = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(166978368))), zero_point = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(3145856)))]; + tensor attn13_out_proj_weight_cast_fp16 = constexpr_affine_dequantize()[axis = tensor(0), name = tensor("attn13_out_proj_weight_cast_fp16"), quantized_data = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(166984576))), scale = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(168033216))), zero_point = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(4203840)))]; + tensor linear13_1_weight_cast_fp16 = constexpr_affine_dequantize()[axis = tensor(0), name = tensor("linear13_1_weight_cast_fp16"), quantized_data = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(168035328))), scale = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(172229696))), zero_point = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8401408)))]; + tensor linear13_2_weight_cast_fp16 = constexpr_affine_dequantize()[axis = tensor(0), name = tensor("linear13_2_weight_cast_fp16"), quantized_data = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(172237952))), scale = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(176432320))), zero_point = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(4203840)))]; + tensor attn14_in_proj_weight_cast_fp16 = constexpr_affine_dequantize()[axis = tensor(0), name = tensor("attn14_in_proj_weight_cast_fp16"), quantized_data = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(176434432))), scale = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(179580224))), zero_point = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(3145856)))]; + tensor attn14_out_proj_weight_cast_fp16 = constexpr_affine_dequantize()[axis = tensor(0), name = tensor("attn14_out_proj_weight_cast_fp16"), quantized_data = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(179586432))), scale = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(180635072))), zero_point = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(4203840)))]; + tensor linear14_1_weight_cast_fp16 = constexpr_affine_dequantize()[axis = tensor(0), name = tensor("linear14_1_weight_cast_fp16"), quantized_data = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(180637184))), scale = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(184831552))), zero_point = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8401408)))]; + tensor linear14_2_weight_cast_fp16 = constexpr_affine_dequantize()[axis = tensor(0), name = tensor("linear14_2_weight_cast_fp16"), quantized_data = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(184839808))), scale = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(189034176))), zero_point = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(4203840)))]; + tensor attn15_in_proj_weight_cast_fp16 = constexpr_affine_dequantize()[axis = tensor(0), name = tensor("attn15_in_proj_weight_cast_fp16"), quantized_data = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(189036288))), scale = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(192182080))), zero_point = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(3145856)))]; + tensor attn15_out_proj_weight_cast_fp16 = constexpr_affine_dequantize()[axis = tensor(0), name = tensor("attn15_out_proj_weight_cast_fp16"), quantized_data = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(192188288))), scale = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(193236928))), zero_point = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(4203840)))]; + tensor linear15_1_weight_cast_fp16 = constexpr_affine_dequantize()[axis = tensor(0), name = tensor("linear15_1_weight_cast_fp16"), quantized_data = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(193239040))), scale = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(197433408))), zero_point = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8401408)))]; + tensor linear15_2_weight_cast_fp16 = constexpr_affine_dequantize()[axis = tensor(0), name = tensor("linear15_2_weight_cast_fp16"), quantized_data = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(197441664))), scale = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(201636032))), zero_point = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(4203840)))]; + tensor attn16_in_proj_weight_cast_fp16 = constexpr_affine_dequantize()[axis = tensor(0), name = tensor("attn16_in_proj_weight_cast_fp16"), quantized_data = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(201638144))), scale = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(204783936))), zero_point = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(3145856)))]; + tensor attn16_out_proj_weight_cast_fp16 = constexpr_affine_dequantize()[axis = tensor(0), name = tensor("attn16_out_proj_weight_cast_fp16"), quantized_data = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(204790144))), scale = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(205838784))), zero_point = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(4203840)))]; + tensor linear16_1_weight_cast_fp16 = constexpr_affine_dequantize()[axis = tensor(0), name = tensor("linear16_1_weight_cast_fp16"), quantized_data = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(205840896))), scale = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(210035264))), zero_point = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8401408)))]; + tensor linear16_2_weight_cast_fp16 = constexpr_affine_dequantize()[axis = tensor(0), name = tensor("linear16_2_weight_cast_fp16"), quantized_data = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(210043520))), scale = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(214237888))), zero_point = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(4203840)))]; + tensor attn17_in_proj_weight_cast_fp16 = constexpr_affine_dequantize()[axis = tensor(0), name = tensor("attn17_in_proj_weight_cast_fp16"), quantized_data = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(214240000))), scale = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(217385792))), zero_point = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(3145856)))]; + tensor attn17_out_proj_weight_cast_fp16 = constexpr_affine_dequantize()[axis = tensor(0), name = tensor("attn17_out_proj_weight_cast_fp16"), quantized_data = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(217392000))), scale = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(218440640))), zero_point = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(4203840)))]; + tensor linear17_1_weight_cast_fp16 = constexpr_affine_dequantize()[axis = tensor(0), name = tensor("linear17_1_weight_cast_fp16"), quantized_data = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(218442752))), scale = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(222637120))), zero_point = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8401408)))]; + tensor linear17_2_weight_cast_fp16 = constexpr_affine_dequantize()[axis = tensor(0), name = tensor("linear17_2_weight_cast_fp16"), quantized_data = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(222645376))), scale = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(226839744))), zero_point = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(4203840)))]; + tensor attn18_in_proj_weight_cast_fp16 = constexpr_affine_dequantize()[axis = tensor(0), name = tensor("attn18_in_proj_weight_cast_fp16"), quantized_data = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(226841856))), scale = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(229987648))), zero_point = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(3145856)))]; + tensor attn18_out_proj_weight_cast_fp16 = constexpr_affine_dequantize()[axis = tensor(0), name = tensor("attn18_out_proj_weight_cast_fp16"), quantized_data = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(229993856))), scale = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(231042496))), zero_point = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(4203840)))]; + tensor linear18_1_weight_cast_fp16 = constexpr_affine_dequantize()[axis = tensor(0), name = tensor("linear18_1_weight_cast_fp16"), quantized_data = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(231044608))), scale = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(235238976))), zero_point = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8401408)))]; + tensor linear18_2_weight_cast_fp16 = constexpr_affine_dequantize()[axis = tensor(0), name = tensor("linear18_2_weight_cast_fp16"), quantized_data = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(235247232))), scale = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(239441600))), zero_point = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(4203840)))]; + tensor attn19_in_proj_weight_cast_fp16 = constexpr_affine_dequantize()[axis = tensor(0), name = tensor("attn19_in_proj_weight_cast_fp16"), quantized_data = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(239443712))), scale = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(242589504))), zero_point = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(3145856)))]; + tensor attn19_out_proj_weight_cast_fp16 = constexpr_affine_dequantize()[axis = tensor(0), name = tensor("attn19_out_proj_weight_cast_fp16"), quantized_data = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(242595712))), scale = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(243644352))), zero_point = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(4203840)))]; + tensor linear19_1_weight_cast_fp16 = constexpr_affine_dequantize()[axis = tensor(0), name = tensor("linear19_1_weight_cast_fp16"), quantized_data = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(243646464))), scale = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(247840832))), zero_point = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8401408)))]; + tensor linear19_2_weight_cast_fp16 = constexpr_affine_dequantize()[axis = tensor(0), name = tensor("linear19_2_weight_cast_fp16"), quantized_data = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(247849088))), scale = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(252043456))), zero_point = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(4203840)))]; + tensor attn20_in_proj_weight_cast_fp16 = constexpr_affine_dequantize()[axis = tensor(0), name = tensor("attn20_in_proj_weight_cast_fp16"), quantized_data = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(252045568))), scale = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(255191360))), zero_point = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(3145856)))]; + tensor attn20_out_proj_weight_cast_fp16 = constexpr_affine_dequantize()[axis = tensor(0), name = tensor("attn20_out_proj_weight_cast_fp16"), quantized_data = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(255197568))), scale = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(256246208))), zero_point = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(4203840)))]; + tensor linear20_1_weight_cast_fp16 = constexpr_affine_dequantize()[axis = tensor(0), name = tensor("linear20_1_weight_cast_fp16"), quantized_data = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(256248320))), scale = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(260442688))), zero_point = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8401408)))]; + tensor linear20_2_weight_cast_fp16 = constexpr_affine_dequantize()[axis = tensor(0), name = tensor("linear20_2_weight_cast_fp16"), quantized_data = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(260450944))), scale = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(264645312))), zero_point = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(4203840)))]; + tensor attn21_in_proj_weight_cast_fp16 = constexpr_affine_dequantize()[axis = tensor(0), name = tensor("attn21_in_proj_weight_cast_fp16"), quantized_data = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(264647424))), scale = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(267793216))), zero_point = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(3145856)))]; + tensor attn21_out_proj_weight_cast_fp16 = constexpr_affine_dequantize()[axis = tensor(0), name = tensor("attn21_out_proj_weight_cast_fp16"), quantized_data = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(267799424))), scale = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(268848064))), zero_point = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(4203840)))]; + tensor linear21_1_weight_cast_fp16 = constexpr_affine_dequantize()[axis = tensor(0), name = tensor("linear21_1_weight_cast_fp16"), quantized_data = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(268850176))), scale = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(273044544))), zero_point = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8401408)))]; + tensor linear21_2_weight_cast_fp16 = constexpr_affine_dequantize()[axis = tensor(0), name = tensor("linear21_2_weight_cast_fp16"), quantized_data = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(273052800))), scale = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(277247168))), zero_point = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(4203840)))]; + tensor attn22_in_proj_weight_cast_fp16 = constexpr_affine_dequantize()[axis = tensor(0), name = tensor("attn22_in_proj_weight_cast_fp16"), quantized_data = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(277249280))), scale = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(280395072))), zero_point = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(3145856)))]; + tensor attn22_out_proj_weight_cast_fp16 = constexpr_affine_dequantize()[axis = tensor(0), name = tensor("attn22_out_proj_weight_cast_fp16"), quantized_data = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(280401280))), scale = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(281449920))), zero_point = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(4203840)))]; + tensor linear22_1_weight_cast_fp16 = constexpr_affine_dequantize()[axis = tensor(0), name = tensor("linear22_1_weight_cast_fp16"), quantized_data = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(281452032))), scale = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(285646400))), zero_point = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8401408)))]; + tensor linear22_2_weight_cast_fp16 = constexpr_affine_dequantize()[axis = tensor(0), name = tensor("linear22_2_weight_cast_fp16"), quantized_data = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(285654656))), scale = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(289849024))), zero_point = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(4203840)))]; + tensor attn23_in_proj_weight_cast_fp16 = constexpr_affine_dequantize()[axis = tensor(0), name = tensor("attn23_in_proj_weight_cast_fp16"), quantized_data = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(289851136))), scale = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(292996928))), zero_point = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(3145856)))]; + tensor attn23_out_proj_weight_cast_fp16 = constexpr_affine_dequantize()[axis = tensor(0), name = tensor("attn23_out_proj_weight_cast_fp16"), quantized_data = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(293003136))), scale = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(294051776))), zero_point = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(4203840)))]; + tensor linear23_1_weight_cast_fp16 = constexpr_affine_dequantize()[axis = tensor(0), name = tensor("linear23_1_weight_cast_fp16"), quantized_data = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(294053888))), scale = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(298248256))), zero_point = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8401408)))]; + tensor linear23_2_weight_cast_fp16 = constexpr_affine_dequantize()[axis = tensor(0), name = tensor("linear23_2_weight_cast_fp16"), quantized_data = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(298256512))), scale = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(302450880))), zero_point = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(4203840)))]; + tensor sequence_to_fp16_dtype_0 = const()[name = tensor("sequence_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor sequence_to_fp16 = cast(dtype = sequence_to_fp16_dtype_0, x = sequence)[name = tensor("cast_555")]; + tensor var_198_cast_fp16 = not_equal(x = sequence_to_fp16, y = sequence_to_fp16)[name = tensor("op_198_cast_fp16")]; + tensor expand_dims_0_axes_0 = const()[name = tensor("expand_dims_0_axes_0"), val = tensor([0, 1])]; + tensor bos_emb_to_fp16_dtype_0 = const()[name = tensor("bos_emb_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor bos_emb_to_fp16 = cast(dtype = bos_emb_to_fp16_dtype_0, x = bos_emb)[name = tensor("cast_554")]; + tensor expand_dims_0_cast_fp16 = expand_dims(axes = expand_dims_0_axes_0, x = bos_emb_to_fp16)[name = tensor("expand_dims_0_cast_fp16")]; + tensor input_1_cast_fp16 = select(a = expand_dims_0_cast_fp16, b = sequence_to_fp16, cond = var_198_cast_fp16)[name = tensor("input_1_cast_fp16")]; + tensor input_linear_weight_to_fp16 = const()[name = tensor("input_linear_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(302452992)))]; + tensor linear_0_bias_0_to_fp16 = const()[name = tensor("linear_0_bias_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(302518592)))]; + tensor linear_0_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = input_linear_weight_to_fp16, x = input_1_cast_fp16)[name = tensor("linear_0_cast_fp16")]; + tensor x_1_axes_0 = const()[name = tensor("x_1_axes_0"), val = tensor([-1])]; + tensor norm0_1_weight_to_fp16 = const()[name = tensor("norm0_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(302520704)))]; + tensor norm0_1_bias_to_fp16 = const()[name = tensor("norm0_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(302522816)))]; + tensor var_204_to_fp16 = const()[name = tensor("op_204_to_fp16"), val = tensor(0x1.5p-17)]; + tensor x_1_cast_fp16 = layer_norm(axes = x_1_axes_0, beta = norm0_1_bias_to_fp16, epsilon = var_204_to_fp16, gamma = norm0_1_weight_to_fp16, x = linear_0_cast_fp16)[name = tensor("x_1_cast_fp16")]; + tensor linear_1_bias_0_to_fp16 = const()[name = tensor("linear_1_bias_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(302524928)))]; + tensor linear_1_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = attn0_in_proj_weight_cast_fp16, x = x_1_cast_fp16)[name = tensor("linear_1_cast_fp16")]; + tensor var_240 = const()[name = tensor("op_240"), val = tensor([1, 1, 3, 16, 64])]; + tensor qkv_1_cast_fp16 = reshape(shape = var_240, x = linear_1_cast_fp16)[name = tensor("qkv_1_cast_fp16")]; + tensor q_1_begin_0 = const()[name = tensor("q_1_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor q_1_end_0 = const()[name = tensor("q_1_end_0"), val = tensor([1, 1, 1, 16, 64])]; + tensor q_1_end_mask_0 = const()[name = tensor("q_1_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor q_1_squeeze_mask_0 = const()[name = tensor("q_1_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor q_1_cast_fp16 = slice_by_index(begin = q_1_begin_0, end = q_1_end_0, end_mask = q_1_end_mask_0, squeeze_mask = q_1_squeeze_mask_0, x = qkv_1_cast_fp16)[name = tensor("q_1_cast_fp16")]; + tensor k_1_begin_0 = const()[name = tensor("k_1_begin_0"), val = tensor([0, 0, 1, 0, 0])]; + tensor k_1_end_0 = const()[name = tensor("k_1_end_0"), val = tensor([1, 1, 2, 16, 64])]; + tensor k_1_end_mask_0 = const()[name = tensor("k_1_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor k_1_squeeze_mask_0 = const()[name = tensor("k_1_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor k_1_cast_fp16 = slice_by_index(begin = k_1_begin_0, end = k_1_end_0, end_mask = k_1_end_mask_0, squeeze_mask = k_1_squeeze_mask_0, x = qkv_1_cast_fp16)[name = tensor("k_1_cast_fp16")]; + tensor v_1_begin_0 = const()[name = tensor("v_1_begin_0"), val = tensor([0, 0, 2, 0, 0])]; + tensor v_1_end_0 = const()[name = tensor("v_1_end_0"), val = tensor([1, 1, 3, 16, 64])]; + tensor v_1_end_mask_0 = const()[name = tensor("v_1_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor v_1_squeeze_mask_0 = const()[name = tensor("v_1_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor v_1_cast_fp16 = slice_by_index(begin = v_1_begin_0, end = v_1_end_0, end_mask = v_1_end_mask_0, squeeze_mask = v_1_squeeze_mask_0, x = qkv_1_cast_fp16)[name = tensor("v_1_cast_fp16")]; + tensor var_344 = const()[name = tensor("op_344"), val = tensor([1, 1, 1, 1])]; + tensor position0_to_fp16_dtype_0 = const()[name = tensor("position0_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor position0_to_fp16 = cast(dtype = position0_to_fp16_dtype_0, x = position0)[name = tensor("cast_553")]; + tensor ts_5_cast_fp16 = reshape(shape = var_344, x = position0_to_fp16)[name = tensor("ts_5_cast_fp16")]; + tensor var_348 = const()[name = tensor("op_348"), val = tensor([1, 1, 16, 32, 2])]; + tensor q_complex_1_cast_fp16 = reshape(shape = var_348, x = q_1_cast_fp16)[name = tensor("q_complex_1_cast_fp16")]; + tensor var_352 = const()[name = tensor("op_352"), val = tensor([1, 1, 16, 32, 2])]; + tensor k_complex_1_cast_fp16 = reshape(shape = var_352, x = k_1_cast_fp16)[name = tensor("k_complex_1_cast_fp16")]; + tensor var_356_begin_0 = const()[name = tensor("op_356_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_356_end_0 = const()[name = tensor("op_356_end_0"), val = tensor([1, 1, 16, 32, 1])]; + tensor var_356_end_mask_0 = const()[name = tensor("op_356_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_356_squeeze_mask_0 = const()[name = tensor("op_356_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_356_cast_fp16 = slice_by_index(begin = var_356_begin_0, end = var_356_end_0, end_mask = var_356_end_mask_0, squeeze_mask = var_356_squeeze_mask_0, x = q_complex_1_cast_fp16)[name = tensor("op_356_cast_fp16")]; + tensor var_364_begin_0 = const()[name = tensor("op_364_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_364_end_0 = const()[name = tensor("op_364_end_0"), val = tensor([1, 1, 16, 32, 2])]; + tensor var_364_end_mask_0 = const()[name = tensor("op_364_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_364_squeeze_mask_0 = const()[name = tensor("op_364_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_364_cast_fp16 = slice_by_index(begin = var_364_begin_0, end = var_364_end_0, end_mask = var_364_end_mask_0, squeeze_mask = var_364_squeeze_mask_0, x = q_complex_1_cast_fp16)[name = tensor("op_364_cast_fp16")]; + tensor var_372_begin_0 = const()[name = tensor("op_372_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_372_end_0 = const()[name = tensor("op_372_end_0"), val = tensor([1, 1, 16, 32, 1])]; + tensor var_372_end_mask_0 = const()[name = tensor("op_372_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_372_squeeze_mask_0 = const()[name = tensor("op_372_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_372_cast_fp16 = slice_by_index(begin = var_372_begin_0, end = var_372_end_0, end_mask = var_372_end_mask_0, squeeze_mask = var_372_squeeze_mask_0, x = k_complex_1_cast_fp16)[name = tensor("op_372_cast_fp16")]; + tensor var_380_begin_0 = const()[name = tensor("op_380_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_380_end_0 = const()[name = tensor("op_380_end_0"), val = tensor([1, 1, 16, 32, 2])]; + tensor var_380_end_mask_0 = const()[name = tensor("op_380_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_380_squeeze_mask_0 = const()[name = tensor("op_380_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_380_cast_fp16 = slice_by_index(begin = var_380_begin_0, end = var_380_end_0, end_mask = var_380_end_mask_0, squeeze_mask = var_380_squeeze_mask_0, x = k_complex_1_cast_fp16)[name = tensor("op_380_cast_fp16")]; + tensor freqs_1_to_fp16 = const()[name = tensor("freqs_1_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(302531136)))]; + tensor var_386_cast_fp16 = mul(x = freqs_1_to_fp16, y = ts_5_cast_fp16)[name = tensor("op_386_cast_fp16")]; + tensor rotr_1_cast_fp16 = cos(x = var_386_cast_fp16)[name = tensor("rotr_1_cast_fp16")]; + tensor roti_1_cast_fp16 = sin(x = var_386_cast_fp16)[name = tensor("roti_1_cast_fp16")]; + tensor var_390_cast_fp16 = mul(x = var_356_cast_fp16, y = rotr_1_cast_fp16)[name = tensor("op_390_cast_fp16")]; + tensor var_391_cast_fp16 = mul(x = var_364_cast_fp16, y = roti_1_cast_fp16)[name = tensor("op_391_cast_fp16")]; + tensor qor_1_cast_fp16 = sub(x = var_390_cast_fp16, y = var_391_cast_fp16)[name = tensor("qor_1_cast_fp16")]; + tensor var_394_cast_fp16 = mul(x = var_356_cast_fp16, y = roti_1_cast_fp16)[name = tensor("op_394_cast_fp16")]; + tensor var_395_cast_fp16 = mul(x = var_364_cast_fp16, y = rotr_1_cast_fp16)[name = tensor("op_395_cast_fp16")]; + tensor qoi_1_cast_fp16 = add(x = var_394_cast_fp16, y = var_395_cast_fp16)[name = tensor("qoi_1_cast_fp16")]; + tensor var_398_cast_fp16 = mul(x = var_372_cast_fp16, y = rotr_1_cast_fp16)[name = tensor("op_398_cast_fp16")]; + tensor var_399_cast_fp16 = mul(x = var_380_cast_fp16, y = roti_1_cast_fp16)[name = tensor("op_399_cast_fp16")]; + tensor kor_1_cast_fp16 = sub(x = var_398_cast_fp16, y = var_399_cast_fp16)[name = tensor("kor_1_cast_fp16")]; + tensor var_402_cast_fp16 = mul(x = var_372_cast_fp16, y = roti_1_cast_fp16)[name = tensor("op_402_cast_fp16")]; + tensor var_403_cast_fp16 = mul(x = var_380_cast_fp16, y = rotr_1_cast_fp16)[name = tensor("op_403_cast_fp16")]; + tensor koi_1_cast_fp16 = add(x = var_402_cast_fp16, y = var_403_cast_fp16)[name = tensor("koi_1_cast_fp16")]; + tensor qo_1_axis_0 = const()[name = tensor("qo_1_axis_0"), val = tensor(-1)]; + tensor qo_1_cast_fp16 = stack(axis = qo_1_axis_0, values = (qor_1_cast_fp16, qoi_1_cast_fp16))[name = tensor("qo_1_cast_fp16")]; + tensor ko_1_axis_0 = const()[name = tensor("ko_1_axis_0"), val = tensor(-1)]; + tensor ko_1_cast_fp16 = stack(axis = ko_1_axis_0, values = (kor_1_cast_fp16, koi_1_cast_fp16))[name = tensor("ko_1_cast_fp16")]; + tensor var_432 = const()[name = tensor("op_432"), val = tensor([1, 1, 16, 64])]; + tensor q_3_cast_fp16 = reshape(shape = var_432, x = qo_1_cast_fp16)[name = tensor("q_3_cast_fp16")]; + tensor var_434 = const()[name = tensor("op_434"), val = tensor([1, 1, 16, 64])]; + tensor k_3_cast_fp16 = reshape(shape = var_434, x = ko_1_cast_fp16)[name = tensor("k_3_cast_fp16")]; + tensor _inversed_456_y_0_to_fp16 = const()[name = tensor("_inversed_456_y_0_to_fp16"), val = tensor(0x1p-9)]; + tensor _inversed_456_cast_fp16 = mul(x = ts_5_cast_fp16, y = _inversed_456_y_0_to_fp16)[name = tensor("_inversed_456_cast_fp16")]; + tensor var_457_cast_fp16 = floor(x = _inversed_456_cast_fp16)[name = tensor("op_457_cast_fp16")]; + tensor var_458_to_fp16 = const()[name = tensor("op_458_to_fp16"), val = tensor(0x1p+9)]; + tensor var_459_cast_fp16 = mul(x = var_457_cast_fp16, y = var_458_to_fp16)[name = tensor("op_459_cast_fp16")]; + tensor write_indices_float_3_cast_fp16 = sub(x = ts_5_cast_fp16, y = var_459_cast_fp16)[name = tensor("write_indices_float_3_cast_fp16")]; + tensor var_466_dtype_0 = const()[name = tensor("op_466_dtype_0"), val = tensor("int32")]; + tensor write_indices_1_reps_0 = const()[name = tensor("write_indices_1_reps_0"), val = tensor([1, 1, 16, 64])]; + tensor write_indices_float_3_cast_fp16_to_int32 = cast(dtype = var_466_dtype_0, x = write_indices_float_3_cast_fp16)[name = tensor("cast_552")]; + tensor write_indices_1 = tile(reps = write_indices_1_reps_0, x = write_indices_float_3_cast_fp16_to_int32)[name = tensor("write_indices_1")]; + tensor var_474_begin_0 = const()[name = tensor("op_474_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_474_end_0 = const()[name = tensor("op_474_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor var_474_end_mask_0 = const()[name = tensor("op_474_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_474_squeeze_mask_0 = const()[name = tensor("op_474_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor cache0_to_fp16_dtype_0 = const()[name = tensor("cache0_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor cache0_to_fp16 = cast(dtype = cache0_to_fp16_dtype_0, x = cache0)[name = tensor("cast_551")]; + tensor var_474_cast_fp16 = slice_by_index(begin = var_474_begin_0, end = var_474_end_0, end_mask = var_474_end_mask_0, squeeze_mask = var_474_squeeze_mask_0, x = cache0_to_fp16)[name = tensor("op_474_cast_fp16")]; + tensor var_476_axis_0 = const()[name = tensor("op_476_axis_0"), val = tensor(1)]; + tensor var_476_mode_0 = const()[name = tensor("op_476_mode_0"), val = tensor("update")]; + tensor var_476_validate_indices_0 = const()[name = tensor("op_476_validate_indices_0"), val = tensor(false)]; + tensor var_476_cast_fp16 = scatter_along_axis(axis = var_476_axis_0, data = var_474_cast_fp16, indices = write_indices_1, mode = var_476_mode_0, updates = k_3_cast_fp16, validate_indices = var_476_validate_indices_0)[name = tensor("op_476_cast_fp16")]; + tensor concat_2 = const()[name = tensor("concat_2"), val = tensor([0, 0, 0, 0, 0])]; + tensor concat_3 = const()[name = tensor("concat_3"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_1_internal_tensor_assign_1_stride_0 = const()[name = tensor("new_cache_1_internal_tensor_assign_1_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_1_internal_tensor_assign_1_begin_mask_0 = const()[name = tensor("new_cache_1_internal_tensor_assign_1_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_1_internal_tensor_assign_1_end_mask_0 = const()[name = tensor("new_cache_1_internal_tensor_assign_1_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_1_internal_tensor_assign_1_squeeze_mask_0 = const()[name = tensor("new_cache_1_internal_tensor_assign_1_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_48 = const()[name = tensor("shape_48"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_0 = const()[name = tensor("reduce_prod_0"), val = tensor(1048576)]; + tensor range_1d_0_start_0 = const()[name = tensor("range_1d_0_start_0"), val = tensor(0)]; + tensor range_1d_0_step_0 = const()[name = tensor("range_1d_0_step_0"), val = tensor(1)]; + tensor range_1d_0 = range_1d(end = reduce_prod_0, start = range_1d_0_start_0, step = range_1d_0_step_0)[name = tensor("range_1d_0")]; + tensor reshape_0 = reshape(shape = shape_48, x = range_1d_0)[name = tensor("reshape_0")]; + tensor slice_by_index_0 = slice_by_index(begin = concat_2, begin_mask = new_cache_1_internal_tensor_assign_1_begin_mask_0, end = concat_3, end_mask = new_cache_1_internal_tensor_assign_1_end_mask_0, squeeze_mask = new_cache_1_internal_tensor_assign_1_squeeze_mask_0, stride = new_cache_1_internal_tensor_assign_1_stride_0, x = reshape_0)[name = tensor("slice_by_index_0")]; + tensor reshape_1_shape_0 = const()[name = tensor("reshape_1_shape_0"), val = tensor([-1])]; + tensor reshape_1 = reshape(shape = reshape_1_shape_0, x = slice_by_index_0)[name = tensor("reshape_1")]; + tensor reshape_2_shape_0 = const()[name = tensor("reshape_2_shape_0"), val = tensor([-1])]; + tensor reshape_2_cast_fp16 = reshape(shape = reshape_2_shape_0, x = var_476_cast_fp16)[name = tensor("reshape_2_cast_fp16")]; + tensor reshape_3_shape_0 = const()[name = tensor("reshape_3_shape_0"), val = tensor([-1])]; + tensor reshape_3_cast_fp16 = reshape(shape = reshape_3_shape_0, x = cache0_to_fp16)[name = tensor("reshape_3_cast_fp16")]; + tensor scatter_0_mode_0 = const()[name = tensor("scatter_0_mode_0"), val = tensor("update")]; + tensor scatter_0_axis_0 = const()[name = tensor("scatter_0_axis_0"), val = tensor(0)]; + tensor scatter_0_validate_indices_0 = const()[name = tensor("scatter_0_validate_indices_0"), val = tensor(false)]; + tensor scatter_0_cast_fp16 = scatter(axis = scatter_0_axis_0, data = reshape_3_cast_fp16, indices = reshape_1, mode = scatter_0_mode_0, updates = reshape_2_cast_fp16, validate_indices = scatter_0_validate_indices_0)[name = tensor("scatter_0_cast_fp16")]; + tensor reshape_4_cast_fp16 = reshape(shape = shape_48, x = scatter_0_cast_fp16)[name = tensor("reshape_4_cast_fp16")]; + tensor var_484_begin_0 = const()[name = tensor("op_484_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor var_484_end_0 = const()[name = tensor("op_484_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor var_484_end_mask_0 = const()[name = tensor("op_484_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_484_squeeze_mask_0 = const()[name = tensor("op_484_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor var_484_cast_fp16 = slice_by_index(begin = var_484_begin_0, end = var_484_end_0, end_mask = var_484_end_mask_0, squeeze_mask = var_484_squeeze_mask_0, x = reshape_4_cast_fp16)[name = tensor("op_484_cast_fp16")]; + tensor var_486_axis_0 = const()[name = tensor("op_486_axis_0"), val = tensor(1)]; + tensor var_486_mode_0 = const()[name = tensor("op_486_mode_0"), val = tensor("update")]; + tensor var_486_validate_indices_0 = const()[name = tensor("op_486_validate_indices_0"), val = tensor(false)]; + tensor var_486_cast_fp16 = scatter_along_axis(axis = var_486_axis_0, data = var_484_cast_fp16, indices = write_indices_1, mode = var_486_mode_0, updates = v_1_cast_fp16, validate_indices = var_486_validate_indices_0)[name = tensor("op_486_cast_fp16")]; + tensor concat_4 = const()[name = tensor("concat_4"), val = tensor([1, 0, 0, 0, 0])]; + tensor concat_5 = const()[name = tensor("concat_5"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_1_internal_tensor_assign_2_stride_0 = const()[name = tensor("new_cache_1_internal_tensor_assign_2_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_1_internal_tensor_assign_2_begin_mask_0 = const()[name = tensor("new_cache_1_internal_tensor_assign_2_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_1_internal_tensor_assign_2_end_mask_0 = const()[name = tensor("new_cache_1_internal_tensor_assign_2_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_1_internal_tensor_assign_2_squeeze_mask_0 = const()[name = tensor("new_cache_1_internal_tensor_assign_2_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_49 = const()[name = tensor("shape_49"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_1 = const()[name = tensor("reduce_prod_1"), val = tensor(1048576)]; + tensor range_1d_1_start_0 = const()[name = tensor("range_1d_1_start_0"), val = tensor(0)]; + tensor range_1d_1_step_0 = const()[name = tensor("range_1d_1_step_0"), val = tensor(1)]; + tensor range_1d_1 = range_1d(end = reduce_prod_1, start = range_1d_1_start_0, step = range_1d_1_step_0)[name = tensor("range_1d_1")]; + tensor reshape_5 = reshape(shape = shape_49, x = range_1d_1)[name = tensor("reshape_5")]; + tensor slice_by_index_1 = slice_by_index(begin = concat_4, begin_mask = new_cache_1_internal_tensor_assign_2_begin_mask_0, end = concat_5, end_mask = new_cache_1_internal_tensor_assign_2_end_mask_0, squeeze_mask = new_cache_1_internal_tensor_assign_2_squeeze_mask_0, stride = new_cache_1_internal_tensor_assign_2_stride_0, x = reshape_5)[name = tensor("slice_by_index_1")]; + tensor reshape_6_shape_0 = const()[name = tensor("reshape_6_shape_0"), val = tensor([-1])]; + tensor reshape_6 = reshape(shape = reshape_6_shape_0, x = slice_by_index_1)[name = tensor("reshape_6")]; + tensor reshape_7_shape_0 = const()[name = tensor("reshape_7_shape_0"), val = tensor([-1])]; + tensor reshape_7_cast_fp16 = reshape(shape = reshape_7_shape_0, x = var_486_cast_fp16)[name = tensor("reshape_7_cast_fp16")]; + tensor reshape_8_shape_0 = const()[name = tensor("reshape_8_shape_0"), val = tensor([-1])]; + tensor reshape_8_cast_fp16 = reshape(shape = reshape_8_shape_0, x = reshape_4_cast_fp16)[name = tensor("reshape_8_cast_fp16")]; + tensor scatter_1_mode_0 = const()[name = tensor("scatter_1_mode_0"), val = tensor("update")]; + tensor scatter_1_axis_0 = const()[name = tensor("scatter_1_axis_0"), val = tensor(0)]; + tensor scatter_1_validate_indices_0 = const()[name = tensor("scatter_1_validate_indices_0"), val = tensor(false)]; + tensor scatter_1_cast_fp16 = scatter(axis = scatter_1_axis_0, data = reshape_8_cast_fp16, indices = reshape_6, mode = scatter_1_mode_0, updates = reshape_7_cast_fp16, validate_indices = scatter_1_validate_indices_0)[name = tensor("scatter_1_cast_fp16")]; + tensor reshape_9_cast_fp16 = reshape(shape = shape_49, x = scatter_1_cast_fp16)[name = tensor("reshape_9_cast_fp16")]; + tensor reshape_9_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("reshape_9_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor keys_1_begin_0 = const()[name = tensor("keys_1_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor keys_1_end_0 = const()[name = tensor("keys_1_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor keys_1_end_mask_0 = const()[name = tensor("keys_1_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor keys_1_squeeze_mask_0 = const()[name = tensor("keys_1_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor keys_1_cast_fp16 = slice_by_index(begin = keys_1_begin_0, end = keys_1_end_0, end_mask = keys_1_end_mask_0, squeeze_mask = keys_1_squeeze_mask_0, x = reshape_9_cast_fp16)[name = tensor("keys_1_cast_fp16")]; + tensor values_1_begin_0 = const()[name = tensor("values_1_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor values_1_end_0 = const()[name = tensor("values_1_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor values_1_end_mask_0 = const()[name = tensor("values_1_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor values_1_squeeze_mask_0 = const()[name = tensor("values_1_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor values_1_cast_fp16 = slice_by_index(begin = values_1_begin_0, end = values_1_end_0, end_mask = values_1_end_mask_0, squeeze_mask = values_1_squeeze_mask_0, x = reshape_9_cast_fp16)[name = tensor("values_1_cast_fp16")]; + tensor var_498_cast_fp16 = not_equal(x = keys_1_cast_fp16, y = keys_1_cast_fp16)[name = tensor("op_498_cast_fp16")]; + tensor var_504_to_fp16 = const()[name = tensor("op_504_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(302531264)))]; + tensor keys_3_cast_fp16 = select(a = var_504_to_fp16, b = keys_1_cast_fp16, cond = var_498_cast_fp16)[name = tensor("keys_3_cast_fp16")]; + tensor var_506_cast_fp16 = not_equal(x = values_1_cast_fp16, y = values_1_cast_fp16)[name = tensor("op_506_cast_fp16")]; + tensor values_3_cast_fp16 = select(a = var_504_to_fp16, b = values_1_cast_fp16, cond = var_506_cast_fp16)[name = tensor("values_3_cast_fp16")]; + tensor var_530 = const()[name = tensor("op_530"), val = tensor([0, 2, 1, 3])]; + tensor var_543 = const()[name = tensor("op_543"), val = tensor([1, 1, 1])]; + tensor var_544_cast_fp16 = reshape(shape = var_543, x = position0_to_fp16)[name = tensor("op_544_cast_fp16")]; + tensor var_561_to_fp16 = const()[name = tensor("op_561_to_fp16"), val = tensor(0x1p+0)]; + tensor valid_len_1_cast_fp16 = add(x = var_544_cast_fp16, y = var_561_to_fp16)[name = tensor("valid_len_1_cast_fp16")]; + tensor k_positions_1_promoted_to_fp16 = const()[name = tensor("k_positions_1_promoted_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(303579904)))]; + tensor valid_mask_1_cast_fp16 = less(x = k_positions_1_promoted_to_fp16, y = valid_len_1_cast_fp16)[name = tensor("valid_mask_1_cast_fp16")]; + tensor causal_mask_1_cast_fp16 = less_equal(x = k_positions_1_promoted_to_fp16, y = var_544_cast_fp16)[name = tensor("causal_mask_1_cast_fp16")]; + tensor attn_mask_1 = logical_and(x = valid_mask_1_cast_fp16, y = causal_mask_1_cast_fp16)[name = tensor("attn_mask_1")]; + tensor attn_mask_3_axes_0 = const()[name = tensor("attn_mask_3_axes_0"), val = tensor([1])]; + tensor attn_mask_3 = expand_dims(axes = attn_mask_3_axes_0, x = attn_mask_1)[name = tensor("attn_mask_3")]; + tensor var_579_transpose_x_0 = const()[name = tensor("op_579_transpose_x_0"), val = tensor(false)]; + tensor var_579_transpose_y_0 = const()[name = tensor("op_579_transpose_y_0"), val = tensor(false)]; + tensor transpose_72_perm_0 = const()[name = tensor("transpose_72_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_73_perm_0 = const()[name = tensor("transpose_73_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_73 = transpose(perm = transpose_73_perm_0, x = keys_3_cast_fp16)[name = tensor("transpose_213")]; + tensor transpose_72 = transpose(perm = transpose_72_perm_0, x = q_3_cast_fp16)[name = tensor("transpose_214")]; + tensor var_579_cast_fp16 = matmul(transpose_x = var_579_transpose_x_0, transpose_y = var_579_transpose_y_0, x = transpose_72, y = transpose_73)[name = tensor("op_579_cast_fp16")]; + tensor var_573_to_fp16 = const()[name = tensor("op_573_to_fp16"), val = tensor([0x1p-3])]; + tensor attn_weights_1_cast_fp16 = mul(x = var_579_cast_fp16, y = var_573_to_fp16)[name = tensor("attn_weights_1_cast_fp16")]; + tensor var_581 = logical_not(x = attn_mask_3)[name = tensor("op_581")]; + tensor var_582_to_fp16 = const()[name = tensor("op_582_to_fp16"), val = tensor(-inf)]; + tensor attn_weights_3_cast_fp16 = select(a = var_582_to_fp16, b = attn_weights_1_cast_fp16, cond = var_581)[name = tensor("attn_weights_3_cast_fp16")]; + tensor var_584 = const()[name = tensor("op_584"), val = tensor(-1)]; + tensor attn_weights_5_cast_fp16 = softmax(axis = var_584, x = attn_weights_3_cast_fp16)[name = tensor("attn_weights_5_cast_fp16")]; + tensor attn_output_1_transpose_x_0 = const()[name = tensor("attn_output_1_transpose_x_0"), val = tensor(false)]; + tensor attn_output_1_transpose_y_0 = const()[name = tensor("attn_output_1_transpose_y_0"), val = tensor(false)]; + tensor values_5_cast_fp16 = transpose(perm = var_530, x = values_3_cast_fp16)[name = tensor("transpose_215")]; + tensor attn_output_1_cast_fp16 = matmul(transpose_x = attn_output_1_transpose_x_0, transpose_y = attn_output_1_transpose_y_0, x = attn_weights_5_cast_fp16, y = values_5_cast_fp16)[name = tensor("attn_output_1_cast_fp16")]; + tensor var_592 = const()[name = tensor("op_592"), val = tensor([0, 2, 1, 3])]; + tensor var_595 = const()[name = tensor("op_595"), val = tensor([1, 1, 1024])]; + tensor var_593_cast_fp16 = transpose(perm = var_592, x = attn_output_1_cast_fp16)[name = tensor("transpose_212")]; + tensor input_5_cast_fp16 = reshape(shape = var_595, x = var_593_cast_fp16)[name = tensor("input_5_cast_fp16")]; + tensor linear_2_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = attn0_out_proj_weight_cast_fp16, x = input_5_cast_fp16)[name = tensor("linear_2_cast_fp16")]; + tensor var_601_to_fp16 = const()[name = tensor("op_601_to_fp16"), val = tensor(0x1p+0)]; + tensor var_602_cast_fp16 = add(x = position0_to_fp16, y = var_601_to_fp16)[name = tensor("op_602_cast_fp16")]; + tensor var_602_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_602_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor input_7_cast_fp16 = add(x = linear_0_cast_fp16, y = linear_2_cast_fp16)[name = tensor("input_7_cast_fp16")]; + tensor input_9_axes_0 = const()[name = tensor("input_9_axes_0"), val = tensor([-1])]; + tensor norm0_2_weight_to_fp16 = const()[name = tensor("norm0_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(303580992)))]; + tensor norm0_2_bias_to_fp16 = const()[name = tensor("norm0_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(303583104)))]; + tensor var_606_to_fp16 = const()[name = tensor("op_606_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_9_cast_fp16 = layer_norm(axes = input_9_axes_0, beta = norm0_2_bias_to_fp16, epsilon = var_606_to_fp16, gamma = norm0_2_weight_to_fp16, x = input_7_cast_fp16)[name = tensor("input_9_cast_fp16")]; + tensor linear_3_bias_0_to_fp16 = const()[name = tensor("linear_3_bias_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(303585216)))]; + tensor linear_3_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = linear0_1_weight_cast_fp16, x = input_9_cast_fp16)[name = tensor("linear_3_cast_fp16")]; + tensor input_11_mode_0 = const()[name = tensor("input_11_mode_0"), val = tensor("EXACT")]; + tensor input_11_cast_fp16 = gelu(mode = input_11_mode_0, x = linear_3_cast_fp16)[name = tensor("input_11_cast_fp16")]; + tensor linear_4_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = linear0_2_weight_cast_fp16, x = input_11_cast_fp16)[name = tensor("linear_4_cast_fp16")]; + tensor input_13_cast_fp16 = add(x = input_7_cast_fp16, y = linear_4_cast_fp16)[name = tensor("input_13_cast_fp16")]; + tensor x_3_axes_0 = const()[name = tensor("x_3_axes_0"), val = tensor([-1])]; + tensor norm1_1_weight_to_fp16 = const()[name = tensor("norm1_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(303593472)))]; + tensor norm1_1_bias_to_fp16 = const()[name = tensor("norm1_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(303595584)))]; + tensor var_623_to_fp16 = const()[name = tensor("op_623_to_fp16"), val = tensor(0x1.5p-17)]; + tensor x_3_cast_fp16 = layer_norm(axes = x_3_axes_0, beta = norm1_1_bias_to_fp16, epsilon = var_623_to_fp16, gamma = norm1_1_weight_to_fp16, x = input_13_cast_fp16)[name = tensor("x_3_cast_fp16")]; + tensor linear_5_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = attn1_in_proj_weight_cast_fp16, x = x_3_cast_fp16)[name = tensor("linear_5_cast_fp16")]; + tensor var_659 = const()[name = tensor("op_659"), val = tensor([1, 1, 3, 16, 64])]; + tensor qkv_3_cast_fp16 = reshape(shape = var_659, x = linear_5_cast_fp16)[name = tensor("qkv_3_cast_fp16")]; + tensor q_7_begin_0 = const()[name = tensor("q_7_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor q_7_end_0 = const()[name = tensor("q_7_end_0"), val = tensor([1, 1, 1, 16, 64])]; + tensor q_7_end_mask_0 = const()[name = tensor("q_7_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor q_7_squeeze_mask_0 = const()[name = tensor("q_7_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor q_7_cast_fp16 = slice_by_index(begin = q_7_begin_0, end = q_7_end_0, end_mask = q_7_end_mask_0, squeeze_mask = q_7_squeeze_mask_0, x = qkv_3_cast_fp16)[name = tensor("q_7_cast_fp16")]; + tensor k_5_begin_0 = const()[name = tensor("k_5_begin_0"), val = tensor([0, 0, 1, 0, 0])]; + tensor k_5_end_0 = const()[name = tensor("k_5_end_0"), val = tensor([1, 1, 2, 16, 64])]; + tensor k_5_end_mask_0 = const()[name = tensor("k_5_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor k_5_squeeze_mask_0 = const()[name = tensor("k_5_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor k_5_cast_fp16 = slice_by_index(begin = k_5_begin_0, end = k_5_end_0, end_mask = k_5_end_mask_0, squeeze_mask = k_5_squeeze_mask_0, x = qkv_3_cast_fp16)[name = tensor("k_5_cast_fp16")]; + tensor v_3_begin_0 = const()[name = tensor("v_3_begin_0"), val = tensor([0, 0, 2, 0, 0])]; + tensor v_3_end_0 = const()[name = tensor("v_3_end_0"), val = tensor([1, 1, 3, 16, 64])]; + tensor v_3_end_mask_0 = const()[name = tensor("v_3_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor v_3_squeeze_mask_0 = const()[name = tensor("v_3_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor v_3_cast_fp16 = slice_by_index(begin = v_3_begin_0, end = v_3_end_0, end_mask = v_3_end_mask_0, squeeze_mask = v_3_squeeze_mask_0, x = qkv_3_cast_fp16)[name = tensor("v_3_cast_fp16")]; + tensor var_763 = const()[name = tensor("op_763"), val = tensor([1, 1, 1, 1])]; + tensor position1_to_fp16_dtype_0 = const()[name = tensor("position1_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor position1_to_fp16 = cast(dtype = position1_to_fp16_dtype_0, x = position1)[name = tensor("cast_548")]; + tensor ts_11_cast_fp16 = reshape(shape = var_763, x = position1_to_fp16)[name = tensor("ts_11_cast_fp16")]; + tensor var_767 = const()[name = tensor("op_767"), val = tensor([1, 1, 16, 32, 2])]; + tensor q_complex_3_cast_fp16 = reshape(shape = var_767, x = q_7_cast_fp16)[name = tensor("q_complex_3_cast_fp16")]; + tensor var_771 = const()[name = tensor("op_771"), val = tensor([1, 1, 16, 32, 2])]; + tensor k_complex_3_cast_fp16 = reshape(shape = var_771, x = k_5_cast_fp16)[name = tensor("k_complex_3_cast_fp16")]; + tensor var_775_begin_0 = const()[name = tensor("op_775_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_775_end_0 = const()[name = tensor("op_775_end_0"), val = tensor([1, 1, 16, 32, 1])]; + tensor var_775_end_mask_0 = const()[name = tensor("op_775_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_775_squeeze_mask_0 = const()[name = tensor("op_775_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_775_cast_fp16 = slice_by_index(begin = var_775_begin_0, end = var_775_end_0, end_mask = var_775_end_mask_0, squeeze_mask = var_775_squeeze_mask_0, x = q_complex_3_cast_fp16)[name = tensor("op_775_cast_fp16")]; + tensor var_783_begin_0 = const()[name = tensor("op_783_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_783_end_0 = const()[name = tensor("op_783_end_0"), val = tensor([1, 1, 16, 32, 2])]; + tensor var_783_end_mask_0 = const()[name = tensor("op_783_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_783_squeeze_mask_0 = const()[name = tensor("op_783_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_783_cast_fp16 = slice_by_index(begin = var_783_begin_0, end = var_783_end_0, end_mask = var_783_end_mask_0, squeeze_mask = var_783_squeeze_mask_0, x = q_complex_3_cast_fp16)[name = tensor("op_783_cast_fp16")]; + tensor var_791_begin_0 = const()[name = tensor("op_791_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_791_end_0 = const()[name = tensor("op_791_end_0"), val = tensor([1, 1, 16, 32, 1])]; + tensor var_791_end_mask_0 = const()[name = tensor("op_791_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_791_squeeze_mask_0 = const()[name = tensor("op_791_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_791_cast_fp16 = slice_by_index(begin = var_791_begin_0, end = var_791_end_0, end_mask = var_791_end_mask_0, squeeze_mask = var_791_squeeze_mask_0, x = k_complex_3_cast_fp16)[name = tensor("op_791_cast_fp16")]; + tensor var_799_begin_0 = const()[name = tensor("op_799_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_799_end_0 = const()[name = tensor("op_799_end_0"), val = tensor([1, 1, 16, 32, 2])]; + tensor var_799_end_mask_0 = const()[name = tensor("op_799_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_799_squeeze_mask_0 = const()[name = tensor("op_799_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_799_cast_fp16 = slice_by_index(begin = var_799_begin_0, end = var_799_end_0, end_mask = var_799_end_mask_0, squeeze_mask = var_799_squeeze_mask_0, x = k_complex_3_cast_fp16)[name = tensor("op_799_cast_fp16")]; + tensor freqs_3_to_fp16 = const()[name = tensor("freqs_3_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(303597696)))]; + tensor var_805_cast_fp16 = mul(x = freqs_3_to_fp16, y = ts_11_cast_fp16)[name = tensor("op_805_cast_fp16")]; + tensor rotr_3_cast_fp16 = cos(x = var_805_cast_fp16)[name = tensor("rotr_3_cast_fp16")]; + tensor roti_3_cast_fp16 = sin(x = var_805_cast_fp16)[name = tensor("roti_3_cast_fp16")]; + tensor var_809_cast_fp16 = mul(x = var_775_cast_fp16, y = rotr_3_cast_fp16)[name = tensor("op_809_cast_fp16")]; + tensor var_810_cast_fp16 = mul(x = var_783_cast_fp16, y = roti_3_cast_fp16)[name = tensor("op_810_cast_fp16")]; + tensor qor_5_cast_fp16 = sub(x = var_809_cast_fp16, y = var_810_cast_fp16)[name = tensor("qor_5_cast_fp16")]; + tensor var_813_cast_fp16 = mul(x = var_775_cast_fp16, y = roti_3_cast_fp16)[name = tensor("op_813_cast_fp16")]; + tensor var_814_cast_fp16 = mul(x = var_783_cast_fp16, y = rotr_3_cast_fp16)[name = tensor("op_814_cast_fp16")]; + tensor qoi_5_cast_fp16 = add(x = var_813_cast_fp16, y = var_814_cast_fp16)[name = tensor("qoi_5_cast_fp16")]; + tensor var_817_cast_fp16 = mul(x = var_791_cast_fp16, y = rotr_3_cast_fp16)[name = tensor("op_817_cast_fp16")]; + tensor var_818_cast_fp16 = mul(x = var_799_cast_fp16, y = roti_3_cast_fp16)[name = tensor("op_818_cast_fp16")]; + tensor kor_5_cast_fp16 = sub(x = var_817_cast_fp16, y = var_818_cast_fp16)[name = tensor("kor_5_cast_fp16")]; + tensor var_821_cast_fp16 = mul(x = var_791_cast_fp16, y = roti_3_cast_fp16)[name = tensor("op_821_cast_fp16")]; + tensor var_822_cast_fp16 = mul(x = var_799_cast_fp16, y = rotr_3_cast_fp16)[name = tensor("op_822_cast_fp16")]; + tensor koi_5_cast_fp16 = add(x = var_821_cast_fp16, y = var_822_cast_fp16)[name = tensor("koi_5_cast_fp16")]; + tensor qo_3_axis_0 = const()[name = tensor("qo_3_axis_0"), val = tensor(-1)]; + tensor qo_3_cast_fp16 = stack(axis = qo_3_axis_0, values = (qor_5_cast_fp16, qoi_5_cast_fp16))[name = tensor("qo_3_cast_fp16")]; + tensor ko_3_axis_0 = const()[name = tensor("ko_3_axis_0"), val = tensor(-1)]; + tensor ko_3_cast_fp16 = stack(axis = ko_3_axis_0, values = (kor_5_cast_fp16, koi_5_cast_fp16))[name = tensor("ko_3_cast_fp16")]; + tensor var_851 = const()[name = tensor("op_851"), val = tensor([1, 1, 16, 64])]; + tensor q_9_cast_fp16 = reshape(shape = var_851, x = qo_3_cast_fp16)[name = tensor("q_9_cast_fp16")]; + tensor var_853 = const()[name = tensor("op_853"), val = tensor([1, 1, 16, 64])]; + tensor k_7_cast_fp16 = reshape(shape = var_853, x = ko_3_cast_fp16)[name = tensor("k_7_cast_fp16")]; + tensor _inversed_875_y_0_to_fp16 = const()[name = tensor("_inversed_875_y_0_to_fp16"), val = tensor(0x1p-9)]; + tensor _inversed_875_cast_fp16 = mul(x = ts_11_cast_fp16, y = _inversed_875_y_0_to_fp16)[name = tensor("_inversed_875_cast_fp16")]; + tensor var_876_cast_fp16 = floor(x = _inversed_875_cast_fp16)[name = tensor("op_876_cast_fp16")]; + tensor var_877_to_fp16 = const()[name = tensor("op_877_to_fp16"), val = tensor(0x1p+9)]; + tensor var_878_cast_fp16 = mul(x = var_876_cast_fp16, y = var_877_to_fp16)[name = tensor("op_878_cast_fp16")]; + tensor write_indices_float_7_cast_fp16 = sub(x = ts_11_cast_fp16, y = var_878_cast_fp16)[name = tensor("write_indices_float_7_cast_fp16")]; + tensor var_885_dtype_0 = const()[name = tensor("op_885_dtype_0"), val = tensor("int32")]; + tensor write_indices_3_reps_0 = const()[name = tensor("write_indices_3_reps_0"), val = tensor([1, 1, 16, 64])]; + tensor write_indices_float_7_cast_fp16_to_int32 = cast(dtype = var_885_dtype_0, x = write_indices_float_7_cast_fp16)[name = tensor("cast_547")]; + tensor write_indices_3 = tile(reps = write_indices_3_reps_0, x = write_indices_float_7_cast_fp16_to_int32)[name = tensor("write_indices_3")]; + tensor var_893_begin_0 = const()[name = tensor("op_893_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_893_end_0 = const()[name = tensor("op_893_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor var_893_end_mask_0 = const()[name = tensor("op_893_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_893_squeeze_mask_0 = const()[name = tensor("op_893_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor cache1_to_fp16_dtype_0 = const()[name = tensor("cache1_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor cache1_to_fp16 = cast(dtype = cache1_to_fp16_dtype_0, x = cache1)[name = tensor("cast_546")]; + tensor var_893_cast_fp16 = slice_by_index(begin = var_893_begin_0, end = var_893_end_0, end_mask = var_893_end_mask_0, squeeze_mask = var_893_squeeze_mask_0, x = cache1_to_fp16)[name = tensor("op_893_cast_fp16")]; + tensor var_895_axis_0 = const()[name = tensor("op_895_axis_0"), val = tensor(1)]; + tensor var_895_mode_0 = const()[name = tensor("op_895_mode_0"), val = tensor("update")]; + tensor var_895_validate_indices_0 = const()[name = tensor("op_895_validate_indices_0"), val = tensor(false)]; + tensor var_895_cast_fp16 = scatter_along_axis(axis = var_895_axis_0, data = var_893_cast_fp16, indices = write_indices_3, mode = var_895_mode_0, updates = k_7_cast_fp16, validate_indices = var_895_validate_indices_0)[name = tensor("op_895_cast_fp16")]; + tensor concat_9 = const()[name = tensor("concat_9"), val = tensor([0, 0, 0, 0, 0])]; + tensor concat_10 = const()[name = tensor("concat_10"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_3_internal_tensor_assign_1_stride_0 = const()[name = tensor("new_cache_3_internal_tensor_assign_1_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_3_internal_tensor_assign_1_begin_mask_0 = const()[name = tensor("new_cache_3_internal_tensor_assign_1_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_3_internal_tensor_assign_1_end_mask_0 = const()[name = tensor("new_cache_3_internal_tensor_assign_1_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_3_internal_tensor_assign_1_squeeze_mask_0 = const()[name = tensor("new_cache_3_internal_tensor_assign_1_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_50 = const()[name = tensor("shape_50"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_2 = const()[name = tensor("reduce_prod_2"), val = tensor(1048576)]; + tensor range_1d_2_start_0 = const()[name = tensor("range_1d_2_start_0"), val = tensor(0)]; + tensor range_1d_2_step_0 = const()[name = tensor("range_1d_2_step_0"), val = tensor(1)]; + tensor range_1d_2 = range_1d(end = reduce_prod_2, start = range_1d_2_start_0, step = range_1d_2_step_0)[name = tensor("range_1d_2")]; + tensor reshape_10 = reshape(shape = shape_50, x = range_1d_2)[name = tensor("reshape_10")]; + tensor slice_by_index_2 = slice_by_index(begin = concat_9, begin_mask = new_cache_3_internal_tensor_assign_1_begin_mask_0, end = concat_10, end_mask = new_cache_3_internal_tensor_assign_1_end_mask_0, squeeze_mask = new_cache_3_internal_tensor_assign_1_squeeze_mask_0, stride = new_cache_3_internal_tensor_assign_1_stride_0, x = reshape_10)[name = tensor("slice_by_index_2")]; + tensor reshape_11_shape_0 = const()[name = tensor("reshape_11_shape_0"), val = tensor([-1])]; + tensor reshape_11 = reshape(shape = reshape_11_shape_0, x = slice_by_index_2)[name = tensor("reshape_11")]; + tensor reshape_12_shape_0 = const()[name = tensor("reshape_12_shape_0"), val = tensor([-1])]; + tensor reshape_12_cast_fp16 = reshape(shape = reshape_12_shape_0, x = var_895_cast_fp16)[name = tensor("reshape_12_cast_fp16")]; + tensor reshape_13_shape_0 = const()[name = tensor("reshape_13_shape_0"), val = tensor([-1])]; + tensor reshape_13_cast_fp16 = reshape(shape = reshape_13_shape_0, x = cache1_to_fp16)[name = tensor("reshape_13_cast_fp16")]; + tensor scatter_2_mode_0 = const()[name = tensor("scatter_2_mode_0"), val = tensor("update")]; + tensor scatter_2_axis_0 = const()[name = tensor("scatter_2_axis_0"), val = tensor(0)]; + tensor scatter_2_validate_indices_0 = const()[name = tensor("scatter_2_validate_indices_0"), val = tensor(false)]; + tensor scatter_2_cast_fp16 = scatter(axis = scatter_2_axis_0, data = reshape_13_cast_fp16, indices = reshape_11, mode = scatter_2_mode_0, updates = reshape_12_cast_fp16, validate_indices = scatter_2_validate_indices_0)[name = tensor("scatter_2_cast_fp16")]; + tensor reshape_14_cast_fp16 = reshape(shape = shape_50, x = scatter_2_cast_fp16)[name = tensor("reshape_14_cast_fp16")]; + tensor var_903_begin_0 = const()[name = tensor("op_903_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor var_903_end_0 = const()[name = tensor("op_903_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor var_903_end_mask_0 = const()[name = tensor("op_903_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_903_squeeze_mask_0 = const()[name = tensor("op_903_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor var_903_cast_fp16 = slice_by_index(begin = var_903_begin_0, end = var_903_end_0, end_mask = var_903_end_mask_0, squeeze_mask = var_903_squeeze_mask_0, x = reshape_14_cast_fp16)[name = tensor("op_903_cast_fp16")]; + tensor var_905_axis_0 = const()[name = tensor("op_905_axis_0"), val = tensor(1)]; + tensor var_905_mode_0 = const()[name = tensor("op_905_mode_0"), val = tensor("update")]; + tensor var_905_validate_indices_0 = const()[name = tensor("op_905_validate_indices_0"), val = tensor(false)]; + tensor var_905_cast_fp16 = scatter_along_axis(axis = var_905_axis_0, data = var_903_cast_fp16, indices = write_indices_3, mode = var_905_mode_0, updates = v_3_cast_fp16, validate_indices = var_905_validate_indices_0)[name = tensor("op_905_cast_fp16")]; + tensor concat_11 = const()[name = tensor("concat_11"), val = tensor([1, 0, 0, 0, 0])]; + tensor concat_12 = const()[name = tensor("concat_12"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_3_internal_tensor_assign_2_stride_0 = const()[name = tensor("new_cache_3_internal_tensor_assign_2_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_3_internal_tensor_assign_2_begin_mask_0 = const()[name = tensor("new_cache_3_internal_tensor_assign_2_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_3_internal_tensor_assign_2_end_mask_0 = const()[name = tensor("new_cache_3_internal_tensor_assign_2_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_3_internal_tensor_assign_2_squeeze_mask_0 = const()[name = tensor("new_cache_3_internal_tensor_assign_2_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_51 = const()[name = tensor("shape_51"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_3 = const()[name = tensor("reduce_prod_3"), val = tensor(1048576)]; + tensor range_1d_3_start_0 = const()[name = tensor("range_1d_3_start_0"), val = tensor(0)]; + tensor range_1d_3_step_0 = const()[name = tensor("range_1d_3_step_0"), val = tensor(1)]; + tensor range_1d_3 = range_1d(end = reduce_prod_3, start = range_1d_3_start_0, step = range_1d_3_step_0)[name = tensor("range_1d_3")]; + tensor reshape_15 = reshape(shape = shape_51, x = range_1d_3)[name = tensor("reshape_15")]; + tensor slice_by_index_3 = slice_by_index(begin = concat_11, begin_mask = new_cache_3_internal_tensor_assign_2_begin_mask_0, end = concat_12, end_mask = new_cache_3_internal_tensor_assign_2_end_mask_0, squeeze_mask = new_cache_3_internal_tensor_assign_2_squeeze_mask_0, stride = new_cache_3_internal_tensor_assign_2_stride_0, x = reshape_15)[name = tensor("slice_by_index_3")]; + tensor reshape_16_shape_0 = const()[name = tensor("reshape_16_shape_0"), val = tensor([-1])]; + tensor reshape_16 = reshape(shape = reshape_16_shape_0, x = slice_by_index_3)[name = tensor("reshape_16")]; + tensor reshape_17_shape_0 = const()[name = tensor("reshape_17_shape_0"), val = tensor([-1])]; + tensor reshape_17_cast_fp16 = reshape(shape = reshape_17_shape_0, x = var_905_cast_fp16)[name = tensor("reshape_17_cast_fp16")]; + tensor reshape_18_shape_0 = const()[name = tensor("reshape_18_shape_0"), val = tensor([-1])]; + tensor reshape_18_cast_fp16 = reshape(shape = reshape_18_shape_0, x = reshape_14_cast_fp16)[name = tensor("reshape_18_cast_fp16")]; + tensor scatter_3_mode_0 = const()[name = tensor("scatter_3_mode_0"), val = tensor("update")]; + tensor scatter_3_axis_0 = const()[name = tensor("scatter_3_axis_0"), val = tensor(0)]; + tensor scatter_3_validate_indices_0 = const()[name = tensor("scatter_3_validate_indices_0"), val = tensor(false)]; + tensor scatter_3_cast_fp16 = scatter(axis = scatter_3_axis_0, data = reshape_18_cast_fp16, indices = reshape_16, mode = scatter_3_mode_0, updates = reshape_17_cast_fp16, validate_indices = scatter_3_validate_indices_0)[name = tensor("scatter_3_cast_fp16")]; + tensor reshape_19_cast_fp16 = reshape(shape = shape_51, x = scatter_3_cast_fp16)[name = tensor("reshape_19_cast_fp16")]; + tensor reshape_19_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("reshape_19_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor keys_7_begin_0 = const()[name = tensor("keys_7_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor keys_7_end_0 = const()[name = tensor("keys_7_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor keys_7_end_mask_0 = const()[name = tensor("keys_7_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor keys_7_squeeze_mask_0 = const()[name = tensor("keys_7_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor keys_7_cast_fp16 = slice_by_index(begin = keys_7_begin_0, end = keys_7_end_0, end_mask = keys_7_end_mask_0, squeeze_mask = keys_7_squeeze_mask_0, x = reshape_19_cast_fp16)[name = tensor("keys_7_cast_fp16")]; + tensor values_7_begin_0 = const()[name = tensor("values_7_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor values_7_end_0 = const()[name = tensor("values_7_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor values_7_end_mask_0 = const()[name = tensor("values_7_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor values_7_squeeze_mask_0 = const()[name = tensor("values_7_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor values_7_cast_fp16 = slice_by_index(begin = values_7_begin_0, end = values_7_end_0, end_mask = values_7_end_mask_0, squeeze_mask = values_7_squeeze_mask_0, x = reshape_19_cast_fp16)[name = tensor("values_7_cast_fp16")]; + tensor var_917_cast_fp16 = not_equal(x = keys_7_cast_fp16, y = keys_7_cast_fp16)[name = tensor("op_917_cast_fp16")]; + tensor keys_9_cast_fp16 = select(a = var_504_to_fp16, b = keys_7_cast_fp16, cond = var_917_cast_fp16)[name = tensor("keys_9_cast_fp16")]; + tensor var_925_cast_fp16 = not_equal(x = values_7_cast_fp16, y = values_7_cast_fp16)[name = tensor("op_925_cast_fp16")]; + tensor values_9_cast_fp16 = select(a = var_504_to_fp16, b = values_7_cast_fp16, cond = var_925_cast_fp16)[name = tensor("values_9_cast_fp16")]; + tensor var_949 = const()[name = tensor("op_949"), val = tensor([0, 2, 1, 3])]; + tensor var_962 = const()[name = tensor("op_962"), val = tensor([1, 1, 1])]; + tensor var_963_cast_fp16 = reshape(shape = var_962, x = position1_to_fp16)[name = tensor("op_963_cast_fp16")]; + tensor var_980_to_fp16 = const()[name = tensor("op_980_to_fp16"), val = tensor(0x1p+0)]; + tensor valid_len_3_cast_fp16 = add(x = var_963_cast_fp16, y = var_980_to_fp16)[name = tensor("valid_len_3_cast_fp16")]; + tensor valid_mask_3_cast_fp16 = less(x = k_positions_1_promoted_to_fp16, y = valid_len_3_cast_fp16)[name = tensor("valid_mask_3_cast_fp16")]; + tensor causal_mask_3_cast_fp16 = less_equal(x = k_positions_1_promoted_to_fp16, y = var_963_cast_fp16)[name = tensor("causal_mask_3_cast_fp16")]; + tensor attn_mask_5 = logical_and(x = valid_mask_3_cast_fp16, y = causal_mask_3_cast_fp16)[name = tensor("attn_mask_5")]; + tensor attn_mask_7_axes_0 = const()[name = tensor("attn_mask_7_axes_0"), val = tensor([1])]; + tensor attn_mask_7 = expand_dims(axes = attn_mask_7_axes_0, x = attn_mask_5)[name = tensor("attn_mask_7")]; + tensor var_998_transpose_x_0 = const()[name = tensor("op_998_transpose_x_0"), val = tensor(false)]; + tensor var_998_transpose_y_0 = const()[name = tensor("op_998_transpose_y_0"), val = tensor(false)]; + tensor transpose_74_perm_0 = const()[name = tensor("transpose_74_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_75_perm_0 = const()[name = tensor("transpose_75_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_75 = transpose(perm = transpose_75_perm_0, x = keys_9_cast_fp16)[name = tensor("transpose_209")]; + tensor transpose_74 = transpose(perm = transpose_74_perm_0, x = q_9_cast_fp16)[name = tensor("transpose_210")]; + tensor var_998_cast_fp16 = matmul(transpose_x = var_998_transpose_x_0, transpose_y = var_998_transpose_y_0, x = transpose_74, y = transpose_75)[name = tensor("op_998_cast_fp16")]; + tensor var_992_to_fp16 = const()[name = tensor("op_992_to_fp16"), val = tensor([0x1p-3])]; + tensor attn_weights_7_cast_fp16 = mul(x = var_998_cast_fp16, y = var_992_to_fp16)[name = tensor("attn_weights_7_cast_fp16")]; + tensor var_1000 = logical_not(x = attn_mask_7)[name = tensor("op_1000")]; + tensor var_1001_to_fp16 = const()[name = tensor("op_1001_to_fp16"), val = tensor(-inf)]; + tensor attn_weights_9_cast_fp16 = select(a = var_1001_to_fp16, b = attn_weights_7_cast_fp16, cond = var_1000)[name = tensor("attn_weights_9_cast_fp16")]; + tensor var_1003 = const()[name = tensor("op_1003"), val = tensor(-1)]; + tensor attn_weights_11_cast_fp16 = softmax(axis = var_1003, x = attn_weights_9_cast_fp16)[name = tensor("attn_weights_11_cast_fp16")]; + tensor attn_output_3_transpose_x_0 = const()[name = tensor("attn_output_3_transpose_x_0"), val = tensor(false)]; + tensor attn_output_3_transpose_y_0 = const()[name = tensor("attn_output_3_transpose_y_0"), val = tensor(false)]; + tensor values_11_cast_fp16 = transpose(perm = var_949, x = values_9_cast_fp16)[name = tensor("transpose_211")]; + tensor attn_output_3_cast_fp16 = matmul(transpose_x = attn_output_3_transpose_x_0, transpose_y = attn_output_3_transpose_y_0, x = attn_weights_11_cast_fp16, y = values_11_cast_fp16)[name = tensor("attn_output_3_cast_fp16")]; + tensor var_1011 = const()[name = tensor("op_1011"), val = tensor([0, 2, 1, 3])]; + tensor var_1014 = const()[name = tensor("op_1014"), val = tensor([1, 1, 1024])]; + tensor var_1012_cast_fp16 = transpose(perm = var_1011, x = attn_output_3_cast_fp16)[name = tensor("transpose_208")]; + tensor input_15_cast_fp16 = reshape(shape = var_1014, x = var_1012_cast_fp16)[name = tensor("input_15_cast_fp16")]; + tensor linear_6_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = attn1_out_proj_weight_cast_fp16, x = input_15_cast_fp16)[name = tensor("linear_6_cast_fp16")]; + tensor var_1020_to_fp16 = const()[name = tensor("op_1020_to_fp16"), val = tensor(0x1p+0)]; + tensor var_1021_cast_fp16 = add(x = position1_to_fp16, y = var_1020_to_fp16)[name = tensor("op_1021_cast_fp16")]; + tensor var_1021_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_1021_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor input_17_cast_fp16 = add(x = input_13_cast_fp16, y = linear_6_cast_fp16)[name = tensor("input_17_cast_fp16")]; + tensor input_19_axes_0 = const()[name = tensor("input_19_axes_0"), val = tensor([-1])]; + tensor norm1_2_weight_to_fp16 = const()[name = tensor("norm1_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(303597824)))]; + tensor norm1_2_bias_to_fp16 = const()[name = tensor("norm1_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(303599936)))]; + tensor var_1025_to_fp16 = const()[name = tensor("op_1025_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_19_cast_fp16 = layer_norm(axes = input_19_axes_0, beta = norm1_2_bias_to_fp16, epsilon = var_1025_to_fp16, gamma = norm1_2_weight_to_fp16, x = input_17_cast_fp16)[name = tensor("input_19_cast_fp16")]; + tensor linear_7_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = linear1_1_weight_cast_fp16, x = input_19_cast_fp16)[name = tensor("linear_7_cast_fp16")]; + tensor input_21_mode_0 = const()[name = tensor("input_21_mode_0"), val = tensor("EXACT")]; + tensor input_21_cast_fp16 = gelu(mode = input_21_mode_0, x = linear_7_cast_fp16)[name = tensor("input_21_cast_fp16")]; + tensor linear_8_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = linear1_2_weight_cast_fp16, x = input_21_cast_fp16)[name = tensor("linear_8_cast_fp16")]; + tensor input_23_cast_fp16 = add(x = input_17_cast_fp16, y = linear_8_cast_fp16)[name = tensor("input_23_cast_fp16")]; + tensor x_5_axes_0 = const()[name = tensor("x_5_axes_0"), val = tensor([-1])]; + tensor norm2_1_weight_to_fp16 = const()[name = tensor("norm2_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(303602048)))]; + tensor norm2_1_bias_to_fp16 = const()[name = tensor("norm2_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(303604160)))]; + tensor var_1042_to_fp16 = const()[name = tensor("op_1042_to_fp16"), val = tensor(0x1.5p-17)]; + tensor x_5_cast_fp16 = layer_norm(axes = x_5_axes_0, beta = norm2_1_bias_to_fp16, epsilon = var_1042_to_fp16, gamma = norm2_1_weight_to_fp16, x = input_23_cast_fp16)[name = tensor("x_5_cast_fp16")]; + tensor linear_9_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = attn2_in_proj_weight_cast_fp16, x = x_5_cast_fp16)[name = tensor("linear_9_cast_fp16")]; + tensor var_1078 = const()[name = tensor("op_1078"), val = tensor([1, 1, 3, 16, 64])]; + tensor qkv_5_cast_fp16 = reshape(shape = var_1078, x = linear_9_cast_fp16)[name = tensor("qkv_5_cast_fp16")]; + tensor q_13_begin_0 = const()[name = tensor("q_13_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor q_13_end_0 = const()[name = tensor("q_13_end_0"), val = tensor([1, 1, 1, 16, 64])]; + tensor q_13_end_mask_0 = const()[name = tensor("q_13_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor q_13_squeeze_mask_0 = const()[name = tensor("q_13_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor q_13_cast_fp16 = slice_by_index(begin = q_13_begin_0, end = q_13_end_0, end_mask = q_13_end_mask_0, squeeze_mask = q_13_squeeze_mask_0, x = qkv_5_cast_fp16)[name = tensor("q_13_cast_fp16")]; + tensor k_9_begin_0 = const()[name = tensor("k_9_begin_0"), val = tensor([0, 0, 1, 0, 0])]; + tensor k_9_end_0 = const()[name = tensor("k_9_end_0"), val = tensor([1, 1, 2, 16, 64])]; + tensor k_9_end_mask_0 = const()[name = tensor("k_9_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor k_9_squeeze_mask_0 = const()[name = tensor("k_9_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor k_9_cast_fp16 = slice_by_index(begin = k_9_begin_0, end = k_9_end_0, end_mask = k_9_end_mask_0, squeeze_mask = k_9_squeeze_mask_0, x = qkv_5_cast_fp16)[name = tensor("k_9_cast_fp16")]; + tensor v_5_begin_0 = const()[name = tensor("v_5_begin_0"), val = tensor([0, 0, 2, 0, 0])]; + tensor v_5_end_0 = const()[name = tensor("v_5_end_0"), val = tensor([1, 1, 3, 16, 64])]; + tensor v_5_end_mask_0 = const()[name = tensor("v_5_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor v_5_squeeze_mask_0 = const()[name = tensor("v_5_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor v_5_cast_fp16 = slice_by_index(begin = v_5_begin_0, end = v_5_end_0, end_mask = v_5_end_mask_0, squeeze_mask = v_5_squeeze_mask_0, x = qkv_5_cast_fp16)[name = tensor("v_5_cast_fp16")]; + tensor var_1182 = const()[name = tensor("op_1182"), val = tensor([1, 1, 1, 1])]; + tensor position2_to_fp16_dtype_0 = const()[name = tensor("position2_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor position2_to_fp16 = cast(dtype = position2_to_fp16_dtype_0, x = position2)[name = tensor("cast_543")]; + tensor ts_17_cast_fp16 = reshape(shape = var_1182, x = position2_to_fp16)[name = tensor("ts_17_cast_fp16")]; + tensor var_1186 = const()[name = tensor("op_1186"), val = tensor([1, 1, 16, 32, 2])]; + tensor q_complex_5_cast_fp16 = reshape(shape = var_1186, x = q_13_cast_fp16)[name = tensor("q_complex_5_cast_fp16")]; + tensor var_1190 = const()[name = tensor("op_1190"), val = tensor([1, 1, 16, 32, 2])]; + tensor k_complex_5_cast_fp16 = reshape(shape = var_1190, x = k_9_cast_fp16)[name = tensor("k_complex_5_cast_fp16")]; + tensor var_1194_begin_0 = const()[name = tensor("op_1194_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_1194_end_0 = const()[name = tensor("op_1194_end_0"), val = tensor([1, 1, 16, 32, 1])]; + tensor var_1194_end_mask_0 = const()[name = tensor("op_1194_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_1194_squeeze_mask_0 = const()[name = tensor("op_1194_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_1194_cast_fp16 = slice_by_index(begin = var_1194_begin_0, end = var_1194_end_0, end_mask = var_1194_end_mask_0, squeeze_mask = var_1194_squeeze_mask_0, x = q_complex_5_cast_fp16)[name = tensor("op_1194_cast_fp16")]; + tensor var_1202_begin_0 = const()[name = tensor("op_1202_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_1202_end_0 = const()[name = tensor("op_1202_end_0"), val = tensor([1, 1, 16, 32, 2])]; + tensor var_1202_end_mask_0 = const()[name = tensor("op_1202_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_1202_squeeze_mask_0 = const()[name = tensor("op_1202_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_1202_cast_fp16 = slice_by_index(begin = var_1202_begin_0, end = var_1202_end_0, end_mask = var_1202_end_mask_0, squeeze_mask = var_1202_squeeze_mask_0, x = q_complex_5_cast_fp16)[name = tensor("op_1202_cast_fp16")]; + tensor var_1210_begin_0 = const()[name = tensor("op_1210_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_1210_end_0 = const()[name = tensor("op_1210_end_0"), val = tensor([1, 1, 16, 32, 1])]; + tensor var_1210_end_mask_0 = const()[name = tensor("op_1210_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_1210_squeeze_mask_0 = const()[name = tensor("op_1210_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_1210_cast_fp16 = slice_by_index(begin = var_1210_begin_0, end = var_1210_end_0, end_mask = var_1210_end_mask_0, squeeze_mask = var_1210_squeeze_mask_0, x = k_complex_5_cast_fp16)[name = tensor("op_1210_cast_fp16")]; + tensor var_1218_begin_0 = const()[name = tensor("op_1218_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_1218_end_0 = const()[name = tensor("op_1218_end_0"), val = tensor([1, 1, 16, 32, 2])]; + tensor var_1218_end_mask_0 = const()[name = tensor("op_1218_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_1218_squeeze_mask_0 = const()[name = tensor("op_1218_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_1218_cast_fp16 = slice_by_index(begin = var_1218_begin_0, end = var_1218_end_0, end_mask = var_1218_end_mask_0, squeeze_mask = var_1218_squeeze_mask_0, x = k_complex_5_cast_fp16)[name = tensor("op_1218_cast_fp16")]; + tensor freqs_5_to_fp16 = const()[name = tensor("freqs_5_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(303606272)))]; + tensor var_1224_cast_fp16 = mul(x = freqs_5_to_fp16, y = ts_17_cast_fp16)[name = tensor("op_1224_cast_fp16")]; + tensor rotr_5_cast_fp16 = cos(x = var_1224_cast_fp16)[name = tensor("rotr_5_cast_fp16")]; + tensor roti_5_cast_fp16 = sin(x = var_1224_cast_fp16)[name = tensor("roti_5_cast_fp16")]; + tensor var_1228_cast_fp16 = mul(x = var_1194_cast_fp16, y = rotr_5_cast_fp16)[name = tensor("op_1228_cast_fp16")]; + tensor var_1229_cast_fp16 = mul(x = var_1202_cast_fp16, y = roti_5_cast_fp16)[name = tensor("op_1229_cast_fp16")]; + tensor qor_9_cast_fp16 = sub(x = var_1228_cast_fp16, y = var_1229_cast_fp16)[name = tensor("qor_9_cast_fp16")]; + tensor var_1232_cast_fp16 = mul(x = var_1194_cast_fp16, y = roti_5_cast_fp16)[name = tensor("op_1232_cast_fp16")]; + tensor var_1233_cast_fp16 = mul(x = var_1202_cast_fp16, y = rotr_5_cast_fp16)[name = tensor("op_1233_cast_fp16")]; + tensor qoi_9_cast_fp16 = add(x = var_1232_cast_fp16, y = var_1233_cast_fp16)[name = tensor("qoi_9_cast_fp16")]; + tensor var_1236_cast_fp16 = mul(x = var_1210_cast_fp16, y = rotr_5_cast_fp16)[name = tensor("op_1236_cast_fp16")]; + tensor var_1237_cast_fp16 = mul(x = var_1218_cast_fp16, y = roti_5_cast_fp16)[name = tensor("op_1237_cast_fp16")]; + tensor kor_9_cast_fp16 = sub(x = var_1236_cast_fp16, y = var_1237_cast_fp16)[name = tensor("kor_9_cast_fp16")]; + tensor var_1240_cast_fp16 = mul(x = var_1210_cast_fp16, y = roti_5_cast_fp16)[name = tensor("op_1240_cast_fp16")]; + tensor var_1241_cast_fp16 = mul(x = var_1218_cast_fp16, y = rotr_5_cast_fp16)[name = tensor("op_1241_cast_fp16")]; + tensor koi_9_cast_fp16 = add(x = var_1240_cast_fp16, y = var_1241_cast_fp16)[name = tensor("koi_9_cast_fp16")]; + tensor qo_5_axis_0 = const()[name = tensor("qo_5_axis_0"), val = tensor(-1)]; + tensor qo_5_cast_fp16 = stack(axis = qo_5_axis_0, values = (qor_9_cast_fp16, qoi_9_cast_fp16))[name = tensor("qo_5_cast_fp16")]; + tensor ko_5_axis_0 = const()[name = tensor("ko_5_axis_0"), val = tensor(-1)]; + tensor ko_5_cast_fp16 = stack(axis = ko_5_axis_0, values = (kor_9_cast_fp16, koi_9_cast_fp16))[name = tensor("ko_5_cast_fp16")]; + tensor var_1270 = const()[name = tensor("op_1270"), val = tensor([1, 1, 16, 64])]; + tensor q_15_cast_fp16 = reshape(shape = var_1270, x = qo_5_cast_fp16)[name = tensor("q_15_cast_fp16")]; + tensor var_1272 = const()[name = tensor("op_1272"), val = tensor([1, 1, 16, 64])]; + tensor k_11_cast_fp16 = reshape(shape = var_1272, x = ko_5_cast_fp16)[name = tensor("k_11_cast_fp16")]; + tensor _inversed_1294_y_0_to_fp16 = const()[name = tensor("_inversed_1294_y_0_to_fp16"), val = tensor(0x1p-9)]; + tensor _inversed_1294_cast_fp16 = mul(x = ts_17_cast_fp16, y = _inversed_1294_y_0_to_fp16)[name = tensor("_inversed_1294_cast_fp16")]; + tensor var_1295_cast_fp16 = floor(x = _inversed_1294_cast_fp16)[name = tensor("op_1295_cast_fp16")]; + tensor var_1296_to_fp16 = const()[name = tensor("op_1296_to_fp16"), val = tensor(0x1p+9)]; + tensor var_1297_cast_fp16 = mul(x = var_1295_cast_fp16, y = var_1296_to_fp16)[name = tensor("op_1297_cast_fp16")]; + tensor write_indices_float_11_cast_fp16 = sub(x = ts_17_cast_fp16, y = var_1297_cast_fp16)[name = tensor("write_indices_float_11_cast_fp16")]; + tensor var_1304_dtype_0 = const()[name = tensor("op_1304_dtype_0"), val = tensor("int32")]; + tensor write_indices_5_reps_0 = const()[name = tensor("write_indices_5_reps_0"), val = tensor([1, 1, 16, 64])]; + tensor write_indices_float_11_cast_fp16_to_int32 = cast(dtype = var_1304_dtype_0, x = write_indices_float_11_cast_fp16)[name = tensor("cast_542")]; + tensor write_indices_5 = tile(reps = write_indices_5_reps_0, x = write_indices_float_11_cast_fp16_to_int32)[name = tensor("write_indices_5")]; + tensor var_1312_begin_0 = const()[name = tensor("op_1312_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_1312_end_0 = const()[name = tensor("op_1312_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor var_1312_end_mask_0 = const()[name = tensor("op_1312_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_1312_squeeze_mask_0 = const()[name = tensor("op_1312_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor cache2_to_fp16_dtype_0 = const()[name = tensor("cache2_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor cache2_to_fp16 = cast(dtype = cache2_to_fp16_dtype_0, x = cache2)[name = tensor("cast_541")]; + tensor var_1312_cast_fp16 = slice_by_index(begin = var_1312_begin_0, end = var_1312_end_0, end_mask = var_1312_end_mask_0, squeeze_mask = var_1312_squeeze_mask_0, x = cache2_to_fp16)[name = tensor("op_1312_cast_fp16")]; + tensor var_1314_axis_0 = const()[name = tensor("op_1314_axis_0"), val = tensor(1)]; + tensor var_1314_mode_0 = const()[name = tensor("op_1314_mode_0"), val = tensor("update")]; + tensor var_1314_validate_indices_0 = const()[name = tensor("op_1314_validate_indices_0"), val = tensor(false)]; + tensor var_1314_cast_fp16 = scatter_along_axis(axis = var_1314_axis_0, data = var_1312_cast_fp16, indices = write_indices_5, mode = var_1314_mode_0, updates = k_11_cast_fp16, validate_indices = var_1314_validate_indices_0)[name = tensor("op_1314_cast_fp16")]; + tensor concat_16 = const()[name = tensor("concat_16"), val = tensor([0, 0, 0, 0, 0])]; + tensor concat_17 = const()[name = tensor("concat_17"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_5_internal_tensor_assign_1_stride_0 = const()[name = tensor("new_cache_5_internal_tensor_assign_1_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_5_internal_tensor_assign_1_begin_mask_0 = const()[name = tensor("new_cache_5_internal_tensor_assign_1_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_5_internal_tensor_assign_1_end_mask_0 = const()[name = tensor("new_cache_5_internal_tensor_assign_1_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_5_internal_tensor_assign_1_squeeze_mask_0 = const()[name = tensor("new_cache_5_internal_tensor_assign_1_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_52 = const()[name = tensor("shape_52"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_4 = const()[name = tensor("reduce_prod_4"), val = tensor(1048576)]; + tensor range_1d_4_start_0 = const()[name = tensor("range_1d_4_start_0"), val = tensor(0)]; + tensor range_1d_4_step_0 = const()[name = tensor("range_1d_4_step_0"), val = tensor(1)]; + tensor range_1d_4 = range_1d(end = reduce_prod_4, start = range_1d_4_start_0, step = range_1d_4_step_0)[name = tensor("range_1d_4")]; + tensor reshape_20 = reshape(shape = shape_52, x = range_1d_4)[name = tensor("reshape_20")]; + tensor slice_by_index_4 = slice_by_index(begin = concat_16, begin_mask = new_cache_5_internal_tensor_assign_1_begin_mask_0, end = concat_17, end_mask = new_cache_5_internal_tensor_assign_1_end_mask_0, squeeze_mask = new_cache_5_internal_tensor_assign_1_squeeze_mask_0, stride = new_cache_5_internal_tensor_assign_1_stride_0, x = reshape_20)[name = tensor("slice_by_index_4")]; + tensor reshape_21_shape_0 = const()[name = tensor("reshape_21_shape_0"), val = tensor([-1])]; + tensor reshape_21 = reshape(shape = reshape_21_shape_0, x = slice_by_index_4)[name = tensor("reshape_21")]; + tensor reshape_22_shape_0 = const()[name = tensor("reshape_22_shape_0"), val = tensor([-1])]; + tensor reshape_22_cast_fp16 = reshape(shape = reshape_22_shape_0, x = var_1314_cast_fp16)[name = tensor("reshape_22_cast_fp16")]; + tensor reshape_23_shape_0 = const()[name = tensor("reshape_23_shape_0"), val = tensor([-1])]; + tensor reshape_23_cast_fp16 = reshape(shape = reshape_23_shape_0, x = cache2_to_fp16)[name = tensor("reshape_23_cast_fp16")]; + tensor scatter_4_mode_0 = const()[name = tensor("scatter_4_mode_0"), val = tensor("update")]; + tensor scatter_4_axis_0 = const()[name = tensor("scatter_4_axis_0"), val = tensor(0)]; + tensor scatter_4_validate_indices_0 = const()[name = tensor("scatter_4_validate_indices_0"), val = tensor(false)]; + tensor scatter_4_cast_fp16 = scatter(axis = scatter_4_axis_0, data = reshape_23_cast_fp16, indices = reshape_21, mode = scatter_4_mode_0, updates = reshape_22_cast_fp16, validate_indices = scatter_4_validate_indices_0)[name = tensor("scatter_4_cast_fp16")]; + tensor reshape_24_cast_fp16 = reshape(shape = shape_52, x = scatter_4_cast_fp16)[name = tensor("reshape_24_cast_fp16")]; + tensor var_1322_begin_0 = const()[name = tensor("op_1322_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor var_1322_end_0 = const()[name = tensor("op_1322_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor var_1322_end_mask_0 = const()[name = tensor("op_1322_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_1322_squeeze_mask_0 = const()[name = tensor("op_1322_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor var_1322_cast_fp16 = slice_by_index(begin = var_1322_begin_0, end = var_1322_end_0, end_mask = var_1322_end_mask_0, squeeze_mask = var_1322_squeeze_mask_0, x = reshape_24_cast_fp16)[name = tensor("op_1322_cast_fp16")]; + tensor var_1324_axis_0 = const()[name = tensor("op_1324_axis_0"), val = tensor(1)]; + tensor var_1324_mode_0 = const()[name = tensor("op_1324_mode_0"), val = tensor("update")]; + tensor var_1324_validate_indices_0 = const()[name = tensor("op_1324_validate_indices_0"), val = tensor(false)]; + tensor var_1324_cast_fp16 = scatter_along_axis(axis = var_1324_axis_0, data = var_1322_cast_fp16, indices = write_indices_5, mode = var_1324_mode_0, updates = v_5_cast_fp16, validate_indices = var_1324_validate_indices_0)[name = tensor("op_1324_cast_fp16")]; + tensor concat_18 = const()[name = tensor("concat_18"), val = tensor([1, 0, 0, 0, 0])]; + tensor concat_19 = const()[name = tensor("concat_19"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_5_internal_tensor_assign_2_stride_0 = const()[name = tensor("new_cache_5_internal_tensor_assign_2_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_5_internal_tensor_assign_2_begin_mask_0 = const()[name = tensor("new_cache_5_internal_tensor_assign_2_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_5_internal_tensor_assign_2_end_mask_0 = const()[name = tensor("new_cache_5_internal_tensor_assign_2_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_5_internal_tensor_assign_2_squeeze_mask_0 = const()[name = tensor("new_cache_5_internal_tensor_assign_2_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_53 = const()[name = tensor("shape_53"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_5 = const()[name = tensor("reduce_prod_5"), val = tensor(1048576)]; + tensor range_1d_5_start_0 = const()[name = tensor("range_1d_5_start_0"), val = tensor(0)]; + tensor range_1d_5_step_0 = const()[name = tensor("range_1d_5_step_0"), val = tensor(1)]; + tensor range_1d_5 = range_1d(end = reduce_prod_5, start = range_1d_5_start_0, step = range_1d_5_step_0)[name = tensor("range_1d_5")]; + tensor reshape_25 = reshape(shape = shape_53, x = range_1d_5)[name = tensor("reshape_25")]; + tensor slice_by_index_5 = slice_by_index(begin = concat_18, begin_mask = new_cache_5_internal_tensor_assign_2_begin_mask_0, end = concat_19, end_mask = new_cache_5_internal_tensor_assign_2_end_mask_0, squeeze_mask = new_cache_5_internal_tensor_assign_2_squeeze_mask_0, stride = new_cache_5_internal_tensor_assign_2_stride_0, x = reshape_25)[name = tensor("slice_by_index_5")]; + tensor reshape_26_shape_0 = const()[name = tensor("reshape_26_shape_0"), val = tensor([-1])]; + tensor reshape_26 = reshape(shape = reshape_26_shape_0, x = slice_by_index_5)[name = tensor("reshape_26")]; + tensor reshape_27_shape_0 = const()[name = tensor("reshape_27_shape_0"), val = tensor([-1])]; + tensor reshape_27_cast_fp16 = reshape(shape = reshape_27_shape_0, x = var_1324_cast_fp16)[name = tensor("reshape_27_cast_fp16")]; + tensor reshape_28_shape_0 = const()[name = tensor("reshape_28_shape_0"), val = tensor([-1])]; + tensor reshape_28_cast_fp16 = reshape(shape = reshape_28_shape_0, x = reshape_24_cast_fp16)[name = tensor("reshape_28_cast_fp16")]; + tensor scatter_5_mode_0 = const()[name = tensor("scatter_5_mode_0"), val = tensor("update")]; + tensor scatter_5_axis_0 = const()[name = tensor("scatter_5_axis_0"), val = tensor(0)]; + tensor scatter_5_validate_indices_0 = const()[name = tensor("scatter_5_validate_indices_0"), val = tensor(false)]; + tensor scatter_5_cast_fp16 = scatter(axis = scatter_5_axis_0, data = reshape_28_cast_fp16, indices = reshape_26, mode = scatter_5_mode_0, updates = reshape_27_cast_fp16, validate_indices = scatter_5_validate_indices_0)[name = tensor("scatter_5_cast_fp16")]; + tensor reshape_29_cast_fp16 = reshape(shape = shape_53, x = scatter_5_cast_fp16)[name = tensor("reshape_29_cast_fp16")]; + tensor reshape_29_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("reshape_29_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor keys_13_begin_0 = const()[name = tensor("keys_13_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor keys_13_end_0 = const()[name = tensor("keys_13_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor keys_13_end_mask_0 = const()[name = tensor("keys_13_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor keys_13_squeeze_mask_0 = const()[name = tensor("keys_13_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor keys_13_cast_fp16 = slice_by_index(begin = keys_13_begin_0, end = keys_13_end_0, end_mask = keys_13_end_mask_0, squeeze_mask = keys_13_squeeze_mask_0, x = reshape_29_cast_fp16)[name = tensor("keys_13_cast_fp16")]; + tensor values_13_begin_0 = const()[name = tensor("values_13_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor values_13_end_0 = const()[name = tensor("values_13_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor values_13_end_mask_0 = const()[name = tensor("values_13_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor values_13_squeeze_mask_0 = const()[name = tensor("values_13_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor values_13_cast_fp16 = slice_by_index(begin = values_13_begin_0, end = values_13_end_0, end_mask = values_13_end_mask_0, squeeze_mask = values_13_squeeze_mask_0, x = reshape_29_cast_fp16)[name = tensor("values_13_cast_fp16")]; + tensor var_1336_cast_fp16 = not_equal(x = keys_13_cast_fp16, y = keys_13_cast_fp16)[name = tensor("op_1336_cast_fp16")]; + tensor keys_15_cast_fp16 = select(a = var_504_to_fp16, b = keys_13_cast_fp16, cond = var_1336_cast_fp16)[name = tensor("keys_15_cast_fp16")]; + tensor var_1344_cast_fp16 = not_equal(x = values_13_cast_fp16, y = values_13_cast_fp16)[name = tensor("op_1344_cast_fp16")]; + tensor values_15_cast_fp16 = select(a = var_504_to_fp16, b = values_13_cast_fp16, cond = var_1344_cast_fp16)[name = tensor("values_15_cast_fp16")]; + tensor var_1368 = const()[name = tensor("op_1368"), val = tensor([0, 2, 1, 3])]; + tensor var_1381 = const()[name = tensor("op_1381"), val = tensor([1, 1, 1])]; + tensor var_1382_cast_fp16 = reshape(shape = var_1381, x = position2_to_fp16)[name = tensor("op_1382_cast_fp16")]; + tensor var_1399_to_fp16 = const()[name = tensor("op_1399_to_fp16"), val = tensor(0x1p+0)]; + tensor valid_len_5_cast_fp16 = add(x = var_1382_cast_fp16, y = var_1399_to_fp16)[name = tensor("valid_len_5_cast_fp16")]; + tensor valid_mask_5_cast_fp16 = less(x = k_positions_1_promoted_to_fp16, y = valid_len_5_cast_fp16)[name = tensor("valid_mask_5_cast_fp16")]; + tensor causal_mask_5_cast_fp16 = less_equal(x = k_positions_1_promoted_to_fp16, y = var_1382_cast_fp16)[name = tensor("causal_mask_5_cast_fp16")]; + tensor attn_mask_9 = logical_and(x = valid_mask_5_cast_fp16, y = causal_mask_5_cast_fp16)[name = tensor("attn_mask_9")]; + tensor attn_mask_11_axes_0 = const()[name = tensor("attn_mask_11_axes_0"), val = tensor([1])]; + tensor attn_mask_11 = expand_dims(axes = attn_mask_11_axes_0, x = attn_mask_9)[name = tensor("attn_mask_11")]; + tensor var_1417_transpose_x_0 = const()[name = tensor("op_1417_transpose_x_0"), val = tensor(false)]; + tensor var_1417_transpose_y_0 = const()[name = tensor("op_1417_transpose_y_0"), val = tensor(false)]; + tensor transpose_76_perm_0 = const()[name = tensor("transpose_76_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_77_perm_0 = const()[name = tensor("transpose_77_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_77 = transpose(perm = transpose_77_perm_0, x = keys_15_cast_fp16)[name = tensor("transpose_205")]; + tensor transpose_76 = transpose(perm = transpose_76_perm_0, x = q_15_cast_fp16)[name = tensor("transpose_206")]; + tensor var_1417_cast_fp16 = matmul(transpose_x = var_1417_transpose_x_0, transpose_y = var_1417_transpose_y_0, x = transpose_76, y = transpose_77)[name = tensor("op_1417_cast_fp16")]; + tensor var_1411_to_fp16 = const()[name = tensor("op_1411_to_fp16"), val = tensor([0x1p-3])]; + tensor attn_weights_13_cast_fp16 = mul(x = var_1417_cast_fp16, y = var_1411_to_fp16)[name = tensor("attn_weights_13_cast_fp16")]; + tensor var_1419 = logical_not(x = attn_mask_11)[name = tensor("op_1419")]; + tensor var_1420_to_fp16 = const()[name = tensor("op_1420_to_fp16"), val = tensor(-inf)]; + tensor attn_weights_15_cast_fp16 = select(a = var_1420_to_fp16, b = attn_weights_13_cast_fp16, cond = var_1419)[name = tensor("attn_weights_15_cast_fp16")]; + tensor var_1422 = const()[name = tensor("op_1422"), val = tensor(-1)]; + tensor attn_weights_17_cast_fp16 = softmax(axis = var_1422, x = attn_weights_15_cast_fp16)[name = tensor("attn_weights_17_cast_fp16")]; + tensor attn_output_5_transpose_x_0 = const()[name = tensor("attn_output_5_transpose_x_0"), val = tensor(false)]; + tensor attn_output_5_transpose_y_0 = const()[name = tensor("attn_output_5_transpose_y_0"), val = tensor(false)]; + tensor values_17_cast_fp16 = transpose(perm = var_1368, x = values_15_cast_fp16)[name = tensor("transpose_207")]; + tensor attn_output_5_cast_fp16 = matmul(transpose_x = attn_output_5_transpose_x_0, transpose_y = attn_output_5_transpose_y_0, x = attn_weights_17_cast_fp16, y = values_17_cast_fp16)[name = tensor("attn_output_5_cast_fp16")]; + tensor var_1430 = const()[name = tensor("op_1430"), val = tensor([0, 2, 1, 3])]; + tensor var_1433 = const()[name = tensor("op_1433"), val = tensor([1, 1, 1024])]; + tensor var_1431_cast_fp16 = transpose(perm = var_1430, x = attn_output_5_cast_fp16)[name = tensor("transpose_204")]; + tensor input_25_cast_fp16 = reshape(shape = var_1433, x = var_1431_cast_fp16)[name = tensor("input_25_cast_fp16")]; + tensor linear_10_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = attn2_out_proj_weight_cast_fp16, x = input_25_cast_fp16)[name = tensor("linear_10_cast_fp16")]; + tensor var_1439_to_fp16 = const()[name = tensor("op_1439_to_fp16"), val = tensor(0x1p+0)]; + tensor var_1440_cast_fp16 = add(x = position2_to_fp16, y = var_1439_to_fp16)[name = tensor("op_1440_cast_fp16")]; + tensor var_1440_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_1440_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor input_27_cast_fp16 = add(x = input_23_cast_fp16, y = linear_10_cast_fp16)[name = tensor("input_27_cast_fp16")]; + tensor input_29_axes_0 = const()[name = tensor("input_29_axes_0"), val = tensor([-1])]; + tensor norm2_2_weight_to_fp16 = const()[name = tensor("norm2_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(303606400)))]; + tensor norm2_2_bias_to_fp16 = const()[name = tensor("norm2_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(303608512)))]; + tensor var_1444_to_fp16 = const()[name = tensor("op_1444_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_29_cast_fp16 = layer_norm(axes = input_29_axes_0, beta = norm2_2_bias_to_fp16, epsilon = var_1444_to_fp16, gamma = norm2_2_weight_to_fp16, x = input_27_cast_fp16)[name = tensor("input_29_cast_fp16")]; + tensor linear_11_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = linear2_1_weight_cast_fp16, x = input_29_cast_fp16)[name = tensor("linear_11_cast_fp16")]; + tensor input_31_mode_0 = const()[name = tensor("input_31_mode_0"), val = tensor("EXACT")]; + tensor input_31_cast_fp16 = gelu(mode = input_31_mode_0, x = linear_11_cast_fp16)[name = tensor("input_31_cast_fp16")]; + tensor linear_12_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = linear2_2_weight_cast_fp16, x = input_31_cast_fp16)[name = tensor("linear_12_cast_fp16")]; + tensor input_33_cast_fp16 = add(x = input_27_cast_fp16, y = linear_12_cast_fp16)[name = tensor("input_33_cast_fp16")]; + tensor x_7_axes_0 = const()[name = tensor("x_7_axes_0"), val = tensor([-1])]; + tensor norm3_1_weight_to_fp16 = const()[name = tensor("norm3_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(303610624)))]; + tensor norm3_1_bias_to_fp16 = const()[name = tensor("norm3_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(303612736)))]; + tensor var_1461_to_fp16 = const()[name = tensor("op_1461_to_fp16"), val = tensor(0x1.5p-17)]; + tensor x_7_cast_fp16 = layer_norm(axes = x_7_axes_0, beta = norm3_1_bias_to_fp16, epsilon = var_1461_to_fp16, gamma = norm3_1_weight_to_fp16, x = input_33_cast_fp16)[name = tensor("x_7_cast_fp16")]; + tensor linear_13_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = attn3_in_proj_weight_cast_fp16, x = x_7_cast_fp16)[name = tensor("linear_13_cast_fp16")]; + tensor var_1497 = const()[name = tensor("op_1497"), val = tensor([1, 1, 3, 16, 64])]; + tensor qkv_7_cast_fp16 = reshape(shape = var_1497, x = linear_13_cast_fp16)[name = tensor("qkv_7_cast_fp16")]; + tensor q_19_begin_0 = const()[name = tensor("q_19_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor q_19_end_0 = const()[name = tensor("q_19_end_0"), val = tensor([1, 1, 1, 16, 64])]; + tensor q_19_end_mask_0 = const()[name = tensor("q_19_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor q_19_squeeze_mask_0 = const()[name = tensor("q_19_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor q_19_cast_fp16 = slice_by_index(begin = q_19_begin_0, end = q_19_end_0, end_mask = q_19_end_mask_0, squeeze_mask = q_19_squeeze_mask_0, x = qkv_7_cast_fp16)[name = tensor("q_19_cast_fp16")]; + tensor k_13_begin_0 = const()[name = tensor("k_13_begin_0"), val = tensor([0, 0, 1, 0, 0])]; + tensor k_13_end_0 = const()[name = tensor("k_13_end_0"), val = tensor([1, 1, 2, 16, 64])]; + tensor k_13_end_mask_0 = const()[name = tensor("k_13_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor k_13_squeeze_mask_0 = const()[name = tensor("k_13_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor k_13_cast_fp16 = slice_by_index(begin = k_13_begin_0, end = k_13_end_0, end_mask = k_13_end_mask_0, squeeze_mask = k_13_squeeze_mask_0, x = qkv_7_cast_fp16)[name = tensor("k_13_cast_fp16")]; + tensor v_7_begin_0 = const()[name = tensor("v_7_begin_0"), val = tensor([0, 0, 2, 0, 0])]; + tensor v_7_end_0 = const()[name = tensor("v_7_end_0"), val = tensor([1, 1, 3, 16, 64])]; + tensor v_7_end_mask_0 = const()[name = tensor("v_7_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor v_7_squeeze_mask_0 = const()[name = tensor("v_7_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor v_7_cast_fp16 = slice_by_index(begin = v_7_begin_0, end = v_7_end_0, end_mask = v_7_end_mask_0, squeeze_mask = v_7_squeeze_mask_0, x = qkv_7_cast_fp16)[name = tensor("v_7_cast_fp16")]; + tensor var_1601 = const()[name = tensor("op_1601"), val = tensor([1, 1, 1, 1])]; + tensor position3_to_fp16_dtype_0 = const()[name = tensor("position3_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor position3_to_fp16 = cast(dtype = position3_to_fp16_dtype_0, x = position3)[name = tensor("cast_538")]; + tensor ts_23_cast_fp16 = reshape(shape = var_1601, x = position3_to_fp16)[name = tensor("ts_23_cast_fp16")]; + tensor var_1605 = const()[name = tensor("op_1605"), val = tensor([1, 1, 16, 32, 2])]; + tensor q_complex_7_cast_fp16 = reshape(shape = var_1605, x = q_19_cast_fp16)[name = tensor("q_complex_7_cast_fp16")]; + tensor var_1609 = const()[name = tensor("op_1609"), val = tensor([1, 1, 16, 32, 2])]; + tensor k_complex_7_cast_fp16 = reshape(shape = var_1609, x = k_13_cast_fp16)[name = tensor("k_complex_7_cast_fp16")]; + tensor var_1613_begin_0 = const()[name = tensor("op_1613_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_1613_end_0 = const()[name = tensor("op_1613_end_0"), val = tensor([1, 1, 16, 32, 1])]; + tensor var_1613_end_mask_0 = const()[name = tensor("op_1613_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_1613_squeeze_mask_0 = const()[name = tensor("op_1613_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_1613_cast_fp16 = slice_by_index(begin = var_1613_begin_0, end = var_1613_end_0, end_mask = var_1613_end_mask_0, squeeze_mask = var_1613_squeeze_mask_0, x = q_complex_7_cast_fp16)[name = tensor("op_1613_cast_fp16")]; + tensor var_1621_begin_0 = const()[name = tensor("op_1621_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_1621_end_0 = const()[name = tensor("op_1621_end_0"), val = tensor([1, 1, 16, 32, 2])]; + tensor var_1621_end_mask_0 = const()[name = tensor("op_1621_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_1621_squeeze_mask_0 = const()[name = tensor("op_1621_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_1621_cast_fp16 = slice_by_index(begin = var_1621_begin_0, end = var_1621_end_0, end_mask = var_1621_end_mask_0, squeeze_mask = var_1621_squeeze_mask_0, x = q_complex_7_cast_fp16)[name = tensor("op_1621_cast_fp16")]; + tensor var_1629_begin_0 = const()[name = tensor("op_1629_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_1629_end_0 = const()[name = tensor("op_1629_end_0"), val = tensor([1, 1, 16, 32, 1])]; + tensor var_1629_end_mask_0 = const()[name = tensor("op_1629_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_1629_squeeze_mask_0 = const()[name = tensor("op_1629_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_1629_cast_fp16 = slice_by_index(begin = var_1629_begin_0, end = var_1629_end_0, end_mask = var_1629_end_mask_0, squeeze_mask = var_1629_squeeze_mask_0, x = k_complex_7_cast_fp16)[name = tensor("op_1629_cast_fp16")]; + tensor var_1637_begin_0 = const()[name = tensor("op_1637_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_1637_end_0 = const()[name = tensor("op_1637_end_0"), val = tensor([1, 1, 16, 32, 2])]; + tensor var_1637_end_mask_0 = const()[name = tensor("op_1637_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_1637_squeeze_mask_0 = const()[name = tensor("op_1637_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_1637_cast_fp16 = slice_by_index(begin = var_1637_begin_0, end = var_1637_end_0, end_mask = var_1637_end_mask_0, squeeze_mask = var_1637_squeeze_mask_0, x = k_complex_7_cast_fp16)[name = tensor("op_1637_cast_fp16")]; + tensor freqs_7_to_fp16 = const()[name = tensor("freqs_7_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(303614848)))]; + tensor var_1643_cast_fp16 = mul(x = freqs_7_to_fp16, y = ts_23_cast_fp16)[name = tensor("op_1643_cast_fp16")]; + tensor rotr_7_cast_fp16 = cos(x = var_1643_cast_fp16)[name = tensor("rotr_7_cast_fp16")]; + tensor roti_7_cast_fp16 = sin(x = var_1643_cast_fp16)[name = tensor("roti_7_cast_fp16")]; + tensor var_1647_cast_fp16 = mul(x = var_1613_cast_fp16, y = rotr_7_cast_fp16)[name = tensor("op_1647_cast_fp16")]; + tensor var_1648_cast_fp16 = mul(x = var_1621_cast_fp16, y = roti_7_cast_fp16)[name = tensor("op_1648_cast_fp16")]; + tensor qor_13_cast_fp16 = sub(x = var_1647_cast_fp16, y = var_1648_cast_fp16)[name = tensor("qor_13_cast_fp16")]; + tensor var_1651_cast_fp16 = mul(x = var_1613_cast_fp16, y = roti_7_cast_fp16)[name = tensor("op_1651_cast_fp16")]; + tensor var_1652_cast_fp16 = mul(x = var_1621_cast_fp16, y = rotr_7_cast_fp16)[name = tensor("op_1652_cast_fp16")]; + tensor qoi_13_cast_fp16 = add(x = var_1651_cast_fp16, y = var_1652_cast_fp16)[name = tensor("qoi_13_cast_fp16")]; + tensor var_1655_cast_fp16 = mul(x = var_1629_cast_fp16, y = rotr_7_cast_fp16)[name = tensor("op_1655_cast_fp16")]; + tensor var_1656_cast_fp16 = mul(x = var_1637_cast_fp16, y = roti_7_cast_fp16)[name = tensor("op_1656_cast_fp16")]; + tensor kor_13_cast_fp16 = sub(x = var_1655_cast_fp16, y = var_1656_cast_fp16)[name = tensor("kor_13_cast_fp16")]; + tensor var_1659_cast_fp16 = mul(x = var_1629_cast_fp16, y = roti_7_cast_fp16)[name = tensor("op_1659_cast_fp16")]; + tensor var_1660_cast_fp16 = mul(x = var_1637_cast_fp16, y = rotr_7_cast_fp16)[name = tensor("op_1660_cast_fp16")]; + tensor koi_13_cast_fp16 = add(x = var_1659_cast_fp16, y = var_1660_cast_fp16)[name = tensor("koi_13_cast_fp16")]; + tensor qo_7_axis_0 = const()[name = tensor("qo_7_axis_0"), val = tensor(-1)]; + tensor qo_7_cast_fp16 = stack(axis = qo_7_axis_0, values = (qor_13_cast_fp16, qoi_13_cast_fp16))[name = tensor("qo_7_cast_fp16")]; + tensor ko_7_axis_0 = const()[name = tensor("ko_7_axis_0"), val = tensor(-1)]; + tensor ko_7_cast_fp16 = stack(axis = ko_7_axis_0, values = (kor_13_cast_fp16, koi_13_cast_fp16))[name = tensor("ko_7_cast_fp16")]; + tensor var_1689 = const()[name = tensor("op_1689"), val = tensor([1, 1, 16, 64])]; + tensor q_21_cast_fp16 = reshape(shape = var_1689, x = qo_7_cast_fp16)[name = tensor("q_21_cast_fp16")]; + tensor var_1691 = const()[name = tensor("op_1691"), val = tensor([1, 1, 16, 64])]; + tensor k_15_cast_fp16 = reshape(shape = var_1691, x = ko_7_cast_fp16)[name = tensor("k_15_cast_fp16")]; + tensor _inversed_1713_y_0_to_fp16 = const()[name = tensor("_inversed_1713_y_0_to_fp16"), val = tensor(0x1p-9)]; + tensor _inversed_1713_cast_fp16 = mul(x = ts_23_cast_fp16, y = _inversed_1713_y_0_to_fp16)[name = tensor("_inversed_1713_cast_fp16")]; + tensor var_1714_cast_fp16 = floor(x = _inversed_1713_cast_fp16)[name = tensor("op_1714_cast_fp16")]; + tensor var_1715_to_fp16 = const()[name = tensor("op_1715_to_fp16"), val = tensor(0x1p+9)]; + tensor var_1716_cast_fp16 = mul(x = var_1714_cast_fp16, y = var_1715_to_fp16)[name = tensor("op_1716_cast_fp16")]; + tensor write_indices_float_15_cast_fp16 = sub(x = ts_23_cast_fp16, y = var_1716_cast_fp16)[name = tensor("write_indices_float_15_cast_fp16")]; + tensor var_1723_dtype_0 = const()[name = tensor("op_1723_dtype_0"), val = tensor("int32")]; + tensor write_indices_7_reps_0 = const()[name = tensor("write_indices_7_reps_0"), val = tensor([1, 1, 16, 64])]; + tensor write_indices_float_15_cast_fp16_to_int32 = cast(dtype = var_1723_dtype_0, x = write_indices_float_15_cast_fp16)[name = tensor("cast_537")]; + tensor write_indices_7 = tile(reps = write_indices_7_reps_0, x = write_indices_float_15_cast_fp16_to_int32)[name = tensor("write_indices_7")]; + tensor var_1731_begin_0 = const()[name = tensor("op_1731_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_1731_end_0 = const()[name = tensor("op_1731_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor var_1731_end_mask_0 = const()[name = tensor("op_1731_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_1731_squeeze_mask_0 = const()[name = tensor("op_1731_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor cache3_to_fp16_dtype_0 = const()[name = tensor("cache3_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor cache3_to_fp16 = cast(dtype = cache3_to_fp16_dtype_0, x = cache3)[name = tensor("cast_536")]; + tensor var_1731_cast_fp16 = slice_by_index(begin = var_1731_begin_0, end = var_1731_end_0, end_mask = var_1731_end_mask_0, squeeze_mask = var_1731_squeeze_mask_0, x = cache3_to_fp16)[name = tensor("op_1731_cast_fp16")]; + tensor var_1733_axis_0 = const()[name = tensor("op_1733_axis_0"), val = tensor(1)]; + tensor var_1733_mode_0 = const()[name = tensor("op_1733_mode_0"), val = tensor("update")]; + tensor var_1733_validate_indices_0 = const()[name = tensor("op_1733_validate_indices_0"), val = tensor(false)]; + tensor var_1733_cast_fp16 = scatter_along_axis(axis = var_1733_axis_0, data = var_1731_cast_fp16, indices = write_indices_7, mode = var_1733_mode_0, updates = k_15_cast_fp16, validate_indices = var_1733_validate_indices_0)[name = tensor("op_1733_cast_fp16")]; + tensor concat_23 = const()[name = tensor("concat_23"), val = tensor([0, 0, 0, 0, 0])]; + tensor concat_24 = const()[name = tensor("concat_24"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_7_internal_tensor_assign_1_stride_0 = const()[name = tensor("new_cache_7_internal_tensor_assign_1_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_7_internal_tensor_assign_1_begin_mask_0 = const()[name = tensor("new_cache_7_internal_tensor_assign_1_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_7_internal_tensor_assign_1_end_mask_0 = const()[name = tensor("new_cache_7_internal_tensor_assign_1_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_7_internal_tensor_assign_1_squeeze_mask_0 = const()[name = tensor("new_cache_7_internal_tensor_assign_1_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_54 = const()[name = tensor("shape_54"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_6 = const()[name = tensor("reduce_prod_6"), val = tensor(1048576)]; + tensor range_1d_6_start_0 = const()[name = tensor("range_1d_6_start_0"), val = tensor(0)]; + tensor range_1d_6_step_0 = const()[name = tensor("range_1d_6_step_0"), val = tensor(1)]; + tensor range_1d_6 = range_1d(end = reduce_prod_6, start = range_1d_6_start_0, step = range_1d_6_step_0)[name = tensor("range_1d_6")]; + tensor reshape_30 = reshape(shape = shape_54, x = range_1d_6)[name = tensor("reshape_30")]; + tensor slice_by_index_6 = slice_by_index(begin = concat_23, begin_mask = new_cache_7_internal_tensor_assign_1_begin_mask_0, end = concat_24, end_mask = new_cache_7_internal_tensor_assign_1_end_mask_0, squeeze_mask = new_cache_7_internal_tensor_assign_1_squeeze_mask_0, stride = new_cache_7_internal_tensor_assign_1_stride_0, x = reshape_30)[name = tensor("slice_by_index_6")]; + tensor reshape_31_shape_0 = const()[name = tensor("reshape_31_shape_0"), val = tensor([-1])]; + tensor reshape_31 = reshape(shape = reshape_31_shape_0, x = slice_by_index_6)[name = tensor("reshape_31")]; + tensor reshape_32_shape_0 = const()[name = tensor("reshape_32_shape_0"), val = tensor([-1])]; + tensor reshape_32_cast_fp16 = reshape(shape = reshape_32_shape_0, x = var_1733_cast_fp16)[name = tensor("reshape_32_cast_fp16")]; + tensor reshape_33_shape_0 = const()[name = tensor("reshape_33_shape_0"), val = tensor([-1])]; + tensor reshape_33_cast_fp16 = reshape(shape = reshape_33_shape_0, x = cache3_to_fp16)[name = tensor("reshape_33_cast_fp16")]; + tensor scatter_6_mode_0 = const()[name = tensor("scatter_6_mode_0"), val = tensor("update")]; + tensor scatter_6_axis_0 = const()[name = tensor("scatter_6_axis_0"), val = tensor(0)]; + tensor scatter_6_validate_indices_0 = const()[name = tensor("scatter_6_validate_indices_0"), val = tensor(false)]; + tensor scatter_6_cast_fp16 = scatter(axis = scatter_6_axis_0, data = reshape_33_cast_fp16, indices = reshape_31, mode = scatter_6_mode_0, updates = reshape_32_cast_fp16, validate_indices = scatter_6_validate_indices_0)[name = tensor("scatter_6_cast_fp16")]; + tensor reshape_34_cast_fp16 = reshape(shape = shape_54, x = scatter_6_cast_fp16)[name = tensor("reshape_34_cast_fp16")]; + tensor var_1741_begin_0 = const()[name = tensor("op_1741_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor var_1741_end_0 = const()[name = tensor("op_1741_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor var_1741_end_mask_0 = const()[name = tensor("op_1741_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_1741_squeeze_mask_0 = const()[name = tensor("op_1741_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor var_1741_cast_fp16 = slice_by_index(begin = var_1741_begin_0, end = var_1741_end_0, end_mask = var_1741_end_mask_0, squeeze_mask = var_1741_squeeze_mask_0, x = reshape_34_cast_fp16)[name = tensor("op_1741_cast_fp16")]; + tensor var_1743_axis_0 = const()[name = tensor("op_1743_axis_0"), val = tensor(1)]; + tensor var_1743_mode_0 = const()[name = tensor("op_1743_mode_0"), val = tensor("update")]; + tensor var_1743_validate_indices_0 = const()[name = tensor("op_1743_validate_indices_0"), val = tensor(false)]; + tensor var_1743_cast_fp16 = scatter_along_axis(axis = var_1743_axis_0, data = var_1741_cast_fp16, indices = write_indices_7, mode = var_1743_mode_0, updates = v_7_cast_fp16, validate_indices = var_1743_validate_indices_0)[name = tensor("op_1743_cast_fp16")]; + tensor concat_25 = const()[name = tensor("concat_25"), val = tensor([1, 0, 0, 0, 0])]; + tensor concat_26 = const()[name = tensor("concat_26"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_7_internal_tensor_assign_2_stride_0 = const()[name = tensor("new_cache_7_internal_tensor_assign_2_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_7_internal_tensor_assign_2_begin_mask_0 = const()[name = tensor("new_cache_7_internal_tensor_assign_2_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_7_internal_tensor_assign_2_end_mask_0 = const()[name = tensor("new_cache_7_internal_tensor_assign_2_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_7_internal_tensor_assign_2_squeeze_mask_0 = const()[name = tensor("new_cache_7_internal_tensor_assign_2_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_55 = const()[name = tensor("shape_55"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_7 = const()[name = tensor("reduce_prod_7"), val = tensor(1048576)]; + tensor range_1d_7_start_0 = const()[name = tensor("range_1d_7_start_0"), val = tensor(0)]; + tensor range_1d_7_step_0 = const()[name = tensor("range_1d_7_step_0"), val = tensor(1)]; + tensor range_1d_7 = range_1d(end = reduce_prod_7, start = range_1d_7_start_0, step = range_1d_7_step_0)[name = tensor("range_1d_7")]; + tensor reshape_35 = reshape(shape = shape_55, x = range_1d_7)[name = tensor("reshape_35")]; + tensor slice_by_index_7 = slice_by_index(begin = concat_25, begin_mask = new_cache_7_internal_tensor_assign_2_begin_mask_0, end = concat_26, end_mask = new_cache_7_internal_tensor_assign_2_end_mask_0, squeeze_mask = new_cache_7_internal_tensor_assign_2_squeeze_mask_0, stride = new_cache_7_internal_tensor_assign_2_stride_0, x = reshape_35)[name = tensor("slice_by_index_7")]; + tensor reshape_36_shape_0 = const()[name = tensor("reshape_36_shape_0"), val = tensor([-1])]; + tensor reshape_36 = reshape(shape = reshape_36_shape_0, x = slice_by_index_7)[name = tensor("reshape_36")]; + tensor reshape_37_shape_0 = const()[name = tensor("reshape_37_shape_0"), val = tensor([-1])]; + tensor reshape_37_cast_fp16 = reshape(shape = reshape_37_shape_0, x = var_1743_cast_fp16)[name = tensor("reshape_37_cast_fp16")]; + tensor reshape_38_shape_0 = const()[name = tensor("reshape_38_shape_0"), val = tensor([-1])]; + tensor reshape_38_cast_fp16 = reshape(shape = reshape_38_shape_0, x = reshape_34_cast_fp16)[name = tensor("reshape_38_cast_fp16")]; + tensor scatter_7_mode_0 = const()[name = tensor("scatter_7_mode_0"), val = tensor("update")]; + tensor scatter_7_axis_0 = const()[name = tensor("scatter_7_axis_0"), val = tensor(0)]; + tensor scatter_7_validate_indices_0 = const()[name = tensor("scatter_7_validate_indices_0"), val = tensor(false)]; + tensor scatter_7_cast_fp16 = scatter(axis = scatter_7_axis_0, data = reshape_38_cast_fp16, indices = reshape_36, mode = scatter_7_mode_0, updates = reshape_37_cast_fp16, validate_indices = scatter_7_validate_indices_0)[name = tensor("scatter_7_cast_fp16")]; + tensor reshape_39_cast_fp16 = reshape(shape = shape_55, x = scatter_7_cast_fp16)[name = tensor("reshape_39_cast_fp16")]; + tensor reshape_39_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("reshape_39_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor keys_19_begin_0 = const()[name = tensor("keys_19_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor keys_19_end_0 = const()[name = tensor("keys_19_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor keys_19_end_mask_0 = const()[name = tensor("keys_19_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor keys_19_squeeze_mask_0 = const()[name = tensor("keys_19_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor keys_19_cast_fp16 = slice_by_index(begin = keys_19_begin_0, end = keys_19_end_0, end_mask = keys_19_end_mask_0, squeeze_mask = keys_19_squeeze_mask_0, x = reshape_39_cast_fp16)[name = tensor("keys_19_cast_fp16")]; + tensor values_19_begin_0 = const()[name = tensor("values_19_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor values_19_end_0 = const()[name = tensor("values_19_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor values_19_end_mask_0 = const()[name = tensor("values_19_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor values_19_squeeze_mask_0 = const()[name = tensor("values_19_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor values_19_cast_fp16 = slice_by_index(begin = values_19_begin_0, end = values_19_end_0, end_mask = values_19_end_mask_0, squeeze_mask = values_19_squeeze_mask_0, x = reshape_39_cast_fp16)[name = tensor("values_19_cast_fp16")]; + tensor var_1755_cast_fp16 = not_equal(x = keys_19_cast_fp16, y = keys_19_cast_fp16)[name = tensor("op_1755_cast_fp16")]; + tensor keys_21_cast_fp16 = select(a = var_504_to_fp16, b = keys_19_cast_fp16, cond = var_1755_cast_fp16)[name = tensor("keys_21_cast_fp16")]; + tensor var_1763_cast_fp16 = not_equal(x = values_19_cast_fp16, y = values_19_cast_fp16)[name = tensor("op_1763_cast_fp16")]; + tensor values_21_cast_fp16 = select(a = var_504_to_fp16, b = values_19_cast_fp16, cond = var_1763_cast_fp16)[name = tensor("values_21_cast_fp16")]; + tensor var_1787 = const()[name = tensor("op_1787"), val = tensor([0, 2, 1, 3])]; + tensor var_1800 = const()[name = tensor("op_1800"), val = tensor([1, 1, 1])]; + tensor var_1801_cast_fp16 = reshape(shape = var_1800, x = position3_to_fp16)[name = tensor("op_1801_cast_fp16")]; + tensor var_1818_to_fp16 = const()[name = tensor("op_1818_to_fp16"), val = tensor(0x1p+0)]; + tensor valid_len_7_cast_fp16 = add(x = var_1801_cast_fp16, y = var_1818_to_fp16)[name = tensor("valid_len_7_cast_fp16")]; + tensor valid_mask_7_cast_fp16 = less(x = k_positions_1_promoted_to_fp16, y = valid_len_7_cast_fp16)[name = tensor("valid_mask_7_cast_fp16")]; + tensor causal_mask_7_cast_fp16 = less_equal(x = k_positions_1_promoted_to_fp16, y = var_1801_cast_fp16)[name = tensor("causal_mask_7_cast_fp16")]; + tensor attn_mask_13 = logical_and(x = valid_mask_7_cast_fp16, y = causal_mask_7_cast_fp16)[name = tensor("attn_mask_13")]; + tensor attn_mask_15_axes_0 = const()[name = tensor("attn_mask_15_axes_0"), val = tensor([1])]; + tensor attn_mask_15 = expand_dims(axes = attn_mask_15_axes_0, x = attn_mask_13)[name = tensor("attn_mask_15")]; + tensor var_1836_transpose_x_0 = const()[name = tensor("op_1836_transpose_x_0"), val = tensor(false)]; + tensor var_1836_transpose_y_0 = const()[name = tensor("op_1836_transpose_y_0"), val = tensor(false)]; + tensor transpose_78_perm_0 = const()[name = tensor("transpose_78_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_79_perm_0 = const()[name = tensor("transpose_79_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_79 = transpose(perm = transpose_79_perm_0, x = keys_21_cast_fp16)[name = tensor("transpose_201")]; + tensor transpose_78 = transpose(perm = transpose_78_perm_0, x = q_21_cast_fp16)[name = tensor("transpose_202")]; + tensor var_1836_cast_fp16 = matmul(transpose_x = var_1836_transpose_x_0, transpose_y = var_1836_transpose_y_0, x = transpose_78, y = transpose_79)[name = tensor("op_1836_cast_fp16")]; + tensor var_1830_to_fp16 = const()[name = tensor("op_1830_to_fp16"), val = tensor([0x1p-3])]; + tensor attn_weights_19_cast_fp16 = mul(x = var_1836_cast_fp16, y = var_1830_to_fp16)[name = tensor("attn_weights_19_cast_fp16")]; + tensor var_1838 = logical_not(x = attn_mask_15)[name = tensor("op_1838")]; + tensor var_1839_to_fp16 = const()[name = tensor("op_1839_to_fp16"), val = tensor(-inf)]; + tensor attn_weights_21_cast_fp16 = select(a = var_1839_to_fp16, b = attn_weights_19_cast_fp16, cond = var_1838)[name = tensor("attn_weights_21_cast_fp16")]; + tensor var_1841 = const()[name = tensor("op_1841"), val = tensor(-1)]; + tensor attn_weights_23_cast_fp16 = softmax(axis = var_1841, x = attn_weights_21_cast_fp16)[name = tensor("attn_weights_23_cast_fp16")]; + tensor attn_output_7_transpose_x_0 = const()[name = tensor("attn_output_7_transpose_x_0"), val = tensor(false)]; + tensor attn_output_7_transpose_y_0 = const()[name = tensor("attn_output_7_transpose_y_0"), val = tensor(false)]; + tensor values_23_cast_fp16 = transpose(perm = var_1787, x = values_21_cast_fp16)[name = tensor("transpose_203")]; + tensor attn_output_7_cast_fp16 = matmul(transpose_x = attn_output_7_transpose_x_0, transpose_y = attn_output_7_transpose_y_0, x = attn_weights_23_cast_fp16, y = values_23_cast_fp16)[name = tensor("attn_output_7_cast_fp16")]; + tensor var_1849 = const()[name = tensor("op_1849"), val = tensor([0, 2, 1, 3])]; + tensor var_1852 = const()[name = tensor("op_1852"), val = tensor([1, 1, 1024])]; + tensor var_1850_cast_fp16 = transpose(perm = var_1849, x = attn_output_7_cast_fp16)[name = tensor("transpose_200")]; + tensor input_35_cast_fp16 = reshape(shape = var_1852, x = var_1850_cast_fp16)[name = tensor("input_35_cast_fp16")]; + tensor linear_14_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = attn3_out_proj_weight_cast_fp16, x = input_35_cast_fp16)[name = tensor("linear_14_cast_fp16")]; + tensor var_1858_to_fp16 = const()[name = tensor("op_1858_to_fp16"), val = tensor(0x1p+0)]; + tensor var_1859_cast_fp16 = add(x = position3_to_fp16, y = var_1858_to_fp16)[name = tensor("op_1859_cast_fp16")]; + tensor var_1859_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_1859_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor input_37_cast_fp16 = add(x = input_33_cast_fp16, y = linear_14_cast_fp16)[name = tensor("input_37_cast_fp16")]; + tensor input_39_axes_0 = const()[name = tensor("input_39_axes_0"), val = tensor([-1])]; + tensor norm3_2_weight_to_fp16 = const()[name = tensor("norm3_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(303614976)))]; + tensor norm3_2_bias_to_fp16 = const()[name = tensor("norm3_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(303617088)))]; + tensor var_1863_to_fp16 = const()[name = tensor("op_1863_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_39_cast_fp16 = layer_norm(axes = input_39_axes_0, beta = norm3_2_bias_to_fp16, epsilon = var_1863_to_fp16, gamma = norm3_2_weight_to_fp16, x = input_37_cast_fp16)[name = tensor("input_39_cast_fp16")]; + tensor linear_15_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = linear3_1_weight_cast_fp16, x = input_39_cast_fp16)[name = tensor("linear_15_cast_fp16")]; + tensor input_41_mode_0 = const()[name = tensor("input_41_mode_0"), val = tensor("EXACT")]; + tensor input_41_cast_fp16 = gelu(mode = input_41_mode_0, x = linear_15_cast_fp16)[name = tensor("input_41_cast_fp16")]; + tensor linear_16_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = linear3_2_weight_cast_fp16, x = input_41_cast_fp16)[name = tensor("linear_16_cast_fp16")]; + tensor input_43_cast_fp16 = add(x = input_37_cast_fp16, y = linear_16_cast_fp16)[name = tensor("input_43_cast_fp16")]; + tensor x_9_axes_0 = const()[name = tensor("x_9_axes_0"), val = tensor([-1])]; + tensor norm4_1_weight_to_fp16 = const()[name = tensor("norm4_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(303619200)))]; + tensor norm4_1_bias_to_fp16 = const()[name = tensor("norm4_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(303621312)))]; + tensor var_1880_to_fp16 = const()[name = tensor("op_1880_to_fp16"), val = tensor(0x1.5p-17)]; + tensor x_9_cast_fp16 = layer_norm(axes = x_9_axes_0, beta = norm4_1_bias_to_fp16, epsilon = var_1880_to_fp16, gamma = norm4_1_weight_to_fp16, x = input_43_cast_fp16)[name = tensor("x_9_cast_fp16")]; + tensor linear_17_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = attn4_in_proj_weight_cast_fp16, x = x_9_cast_fp16)[name = tensor("linear_17_cast_fp16")]; + tensor var_1916 = const()[name = tensor("op_1916"), val = tensor([1, 1, 3, 16, 64])]; + tensor qkv_9_cast_fp16 = reshape(shape = var_1916, x = linear_17_cast_fp16)[name = tensor("qkv_9_cast_fp16")]; + tensor q_25_begin_0 = const()[name = tensor("q_25_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor q_25_end_0 = const()[name = tensor("q_25_end_0"), val = tensor([1, 1, 1, 16, 64])]; + tensor q_25_end_mask_0 = const()[name = tensor("q_25_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor q_25_squeeze_mask_0 = const()[name = tensor("q_25_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor q_25_cast_fp16 = slice_by_index(begin = q_25_begin_0, end = q_25_end_0, end_mask = q_25_end_mask_0, squeeze_mask = q_25_squeeze_mask_0, x = qkv_9_cast_fp16)[name = tensor("q_25_cast_fp16")]; + tensor k_17_begin_0 = const()[name = tensor("k_17_begin_0"), val = tensor([0, 0, 1, 0, 0])]; + tensor k_17_end_0 = const()[name = tensor("k_17_end_0"), val = tensor([1, 1, 2, 16, 64])]; + tensor k_17_end_mask_0 = const()[name = tensor("k_17_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor k_17_squeeze_mask_0 = const()[name = tensor("k_17_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor k_17_cast_fp16 = slice_by_index(begin = k_17_begin_0, end = k_17_end_0, end_mask = k_17_end_mask_0, squeeze_mask = k_17_squeeze_mask_0, x = qkv_9_cast_fp16)[name = tensor("k_17_cast_fp16")]; + tensor v_9_begin_0 = const()[name = tensor("v_9_begin_0"), val = tensor([0, 0, 2, 0, 0])]; + tensor v_9_end_0 = const()[name = tensor("v_9_end_0"), val = tensor([1, 1, 3, 16, 64])]; + tensor v_9_end_mask_0 = const()[name = tensor("v_9_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor v_9_squeeze_mask_0 = const()[name = tensor("v_9_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor v_9_cast_fp16 = slice_by_index(begin = v_9_begin_0, end = v_9_end_0, end_mask = v_9_end_mask_0, squeeze_mask = v_9_squeeze_mask_0, x = qkv_9_cast_fp16)[name = tensor("v_9_cast_fp16")]; + tensor var_2020 = const()[name = tensor("op_2020"), val = tensor([1, 1, 1, 1])]; + tensor position4_to_fp16_dtype_0 = const()[name = tensor("position4_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor position4_to_fp16 = cast(dtype = position4_to_fp16_dtype_0, x = position4)[name = tensor("cast_533")]; + tensor ts_29_cast_fp16 = reshape(shape = var_2020, x = position4_to_fp16)[name = tensor("ts_29_cast_fp16")]; + tensor var_2024 = const()[name = tensor("op_2024"), val = tensor([1, 1, 16, 32, 2])]; + tensor q_complex_9_cast_fp16 = reshape(shape = var_2024, x = q_25_cast_fp16)[name = tensor("q_complex_9_cast_fp16")]; + tensor var_2028 = const()[name = tensor("op_2028"), val = tensor([1, 1, 16, 32, 2])]; + tensor k_complex_9_cast_fp16 = reshape(shape = var_2028, x = k_17_cast_fp16)[name = tensor("k_complex_9_cast_fp16")]; + tensor var_2032_begin_0 = const()[name = tensor("op_2032_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_2032_end_0 = const()[name = tensor("op_2032_end_0"), val = tensor([1, 1, 16, 32, 1])]; + tensor var_2032_end_mask_0 = const()[name = tensor("op_2032_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_2032_squeeze_mask_0 = const()[name = tensor("op_2032_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_2032_cast_fp16 = slice_by_index(begin = var_2032_begin_0, end = var_2032_end_0, end_mask = var_2032_end_mask_0, squeeze_mask = var_2032_squeeze_mask_0, x = q_complex_9_cast_fp16)[name = tensor("op_2032_cast_fp16")]; + tensor var_2040_begin_0 = const()[name = tensor("op_2040_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_2040_end_0 = const()[name = tensor("op_2040_end_0"), val = tensor([1, 1, 16, 32, 2])]; + tensor var_2040_end_mask_0 = const()[name = tensor("op_2040_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_2040_squeeze_mask_0 = const()[name = tensor("op_2040_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_2040_cast_fp16 = slice_by_index(begin = var_2040_begin_0, end = var_2040_end_0, end_mask = var_2040_end_mask_0, squeeze_mask = var_2040_squeeze_mask_0, x = q_complex_9_cast_fp16)[name = tensor("op_2040_cast_fp16")]; + tensor var_2048_begin_0 = const()[name = tensor("op_2048_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_2048_end_0 = const()[name = tensor("op_2048_end_0"), val = tensor([1, 1, 16, 32, 1])]; + tensor var_2048_end_mask_0 = const()[name = tensor("op_2048_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_2048_squeeze_mask_0 = const()[name = tensor("op_2048_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_2048_cast_fp16 = slice_by_index(begin = var_2048_begin_0, end = var_2048_end_0, end_mask = var_2048_end_mask_0, squeeze_mask = var_2048_squeeze_mask_0, x = k_complex_9_cast_fp16)[name = tensor("op_2048_cast_fp16")]; + tensor var_2056_begin_0 = const()[name = tensor("op_2056_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_2056_end_0 = const()[name = tensor("op_2056_end_0"), val = tensor([1, 1, 16, 32, 2])]; + tensor var_2056_end_mask_0 = const()[name = tensor("op_2056_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_2056_squeeze_mask_0 = const()[name = tensor("op_2056_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_2056_cast_fp16 = slice_by_index(begin = var_2056_begin_0, end = var_2056_end_0, end_mask = var_2056_end_mask_0, squeeze_mask = var_2056_squeeze_mask_0, x = k_complex_9_cast_fp16)[name = tensor("op_2056_cast_fp16")]; + tensor freqs_9_to_fp16 = const()[name = tensor("freqs_9_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(303623424)))]; + tensor var_2062_cast_fp16 = mul(x = freqs_9_to_fp16, y = ts_29_cast_fp16)[name = tensor("op_2062_cast_fp16")]; + tensor rotr_9_cast_fp16 = cos(x = var_2062_cast_fp16)[name = tensor("rotr_9_cast_fp16")]; + tensor roti_9_cast_fp16 = sin(x = var_2062_cast_fp16)[name = tensor("roti_9_cast_fp16")]; + tensor var_2066_cast_fp16 = mul(x = var_2032_cast_fp16, y = rotr_9_cast_fp16)[name = tensor("op_2066_cast_fp16")]; + tensor var_2067_cast_fp16 = mul(x = var_2040_cast_fp16, y = roti_9_cast_fp16)[name = tensor("op_2067_cast_fp16")]; + tensor qor_17_cast_fp16 = sub(x = var_2066_cast_fp16, y = var_2067_cast_fp16)[name = tensor("qor_17_cast_fp16")]; + tensor var_2070_cast_fp16 = mul(x = var_2032_cast_fp16, y = roti_9_cast_fp16)[name = tensor("op_2070_cast_fp16")]; + tensor var_2071_cast_fp16 = mul(x = var_2040_cast_fp16, y = rotr_9_cast_fp16)[name = tensor("op_2071_cast_fp16")]; + tensor qoi_17_cast_fp16 = add(x = var_2070_cast_fp16, y = var_2071_cast_fp16)[name = tensor("qoi_17_cast_fp16")]; + tensor var_2074_cast_fp16 = mul(x = var_2048_cast_fp16, y = rotr_9_cast_fp16)[name = tensor("op_2074_cast_fp16")]; + tensor var_2075_cast_fp16 = mul(x = var_2056_cast_fp16, y = roti_9_cast_fp16)[name = tensor("op_2075_cast_fp16")]; + tensor kor_17_cast_fp16 = sub(x = var_2074_cast_fp16, y = var_2075_cast_fp16)[name = tensor("kor_17_cast_fp16")]; + tensor var_2078_cast_fp16 = mul(x = var_2048_cast_fp16, y = roti_9_cast_fp16)[name = tensor("op_2078_cast_fp16")]; + tensor var_2079_cast_fp16 = mul(x = var_2056_cast_fp16, y = rotr_9_cast_fp16)[name = tensor("op_2079_cast_fp16")]; + tensor koi_17_cast_fp16 = add(x = var_2078_cast_fp16, y = var_2079_cast_fp16)[name = tensor("koi_17_cast_fp16")]; + tensor qo_9_axis_0 = const()[name = tensor("qo_9_axis_0"), val = tensor(-1)]; + tensor qo_9_cast_fp16 = stack(axis = qo_9_axis_0, values = (qor_17_cast_fp16, qoi_17_cast_fp16))[name = tensor("qo_9_cast_fp16")]; + tensor ko_9_axis_0 = const()[name = tensor("ko_9_axis_0"), val = tensor(-1)]; + tensor ko_9_cast_fp16 = stack(axis = ko_9_axis_0, values = (kor_17_cast_fp16, koi_17_cast_fp16))[name = tensor("ko_9_cast_fp16")]; + tensor var_2108 = const()[name = tensor("op_2108"), val = tensor([1, 1, 16, 64])]; + tensor q_27_cast_fp16 = reshape(shape = var_2108, x = qo_9_cast_fp16)[name = tensor("q_27_cast_fp16")]; + tensor var_2110 = const()[name = tensor("op_2110"), val = tensor([1, 1, 16, 64])]; + tensor k_19_cast_fp16 = reshape(shape = var_2110, x = ko_9_cast_fp16)[name = tensor("k_19_cast_fp16")]; + tensor _inversed_2132_y_0_to_fp16 = const()[name = tensor("_inversed_2132_y_0_to_fp16"), val = tensor(0x1p-9)]; + tensor _inversed_2132_cast_fp16 = mul(x = ts_29_cast_fp16, y = _inversed_2132_y_0_to_fp16)[name = tensor("_inversed_2132_cast_fp16")]; + tensor var_2133_cast_fp16 = floor(x = _inversed_2132_cast_fp16)[name = tensor("op_2133_cast_fp16")]; + tensor var_2134_to_fp16 = const()[name = tensor("op_2134_to_fp16"), val = tensor(0x1p+9)]; + tensor var_2135_cast_fp16 = mul(x = var_2133_cast_fp16, y = var_2134_to_fp16)[name = tensor("op_2135_cast_fp16")]; + tensor write_indices_float_19_cast_fp16 = sub(x = ts_29_cast_fp16, y = var_2135_cast_fp16)[name = tensor("write_indices_float_19_cast_fp16")]; + tensor var_2142_dtype_0 = const()[name = tensor("op_2142_dtype_0"), val = tensor("int32")]; + tensor write_indices_9_reps_0 = const()[name = tensor("write_indices_9_reps_0"), val = tensor([1, 1, 16, 64])]; + tensor write_indices_float_19_cast_fp16_to_int32 = cast(dtype = var_2142_dtype_0, x = write_indices_float_19_cast_fp16)[name = tensor("cast_532")]; + tensor write_indices_9 = tile(reps = write_indices_9_reps_0, x = write_indices_float_19_cast_fp16_to_int32)[name = tensor("write_indices_9")]; + tensor var_2150_begin_0 = const()[name = tensor("op_2150_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_2150_end_0 = const()[name = tensor("op_2150_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor var_2150_end_mask_0 = const()[name = tensor("op_2150_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_2150_squeeze_mask_0 = const()[name = tensor("op_2150_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor cache4_to_fp16_dtype_0 = const()[name = tensor("cache4_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor cache4_to_fp16 = cast(dtype = cache4_to_fp16_dtype_0, x = cache4)[name = tensor("cast_531")]; + tensor var_2150_cast_fp16 = slice_by_index(begin = var_2150_begin_0, end = var_2150_end_0, end_mask = var_2150_end_mask_0, squeeze_mask = var_2150_squeeze_mask_0, x = cache4_to_fp16)[name = tensor("op_2150_cast_fp16")]; + tensor var_2152_axis_0 = const()[name = tensor("op_2152_axis_0"), val = tensor(1)]; + tensor var_2152_mode_0 = const()[name = tensor("op_2152_mode_0"), val = tensor("update")]; + tensor var_2152_validate_indices_0 = const()[name = tensor("op_2152_validate_indices_0"), val = tensor(false)]; + tensor var_2152_cast_fp16 = scatter_along_axis(axis = var_2152_axis_0, data = var_2150_cast_fp16, indices = write_indices_9, mode = var_2152_mode_0, updates = k_19_cast_fp16, validate_indices = var_2152_validate_indices_0)[name = tensor("op_2152_cast_fp16")]; + tensor concat_30 = const()[name = tensor("concat_30"), val = tensor([0, 0, 0, 0, 0])]; + tensor concat_31 = const()[name = tensor("concat_31"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_9_internal_tensor_assign_1_stride_0 = const()[name = tensor("new_cache_9_internal_tensor_assign_1_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_9_internal_tensor_assign_1_begin_mask_0 = const()[name = tensor("new_cache_9_internal_tensor_assign_1_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_9_internal_tensor_assign_1_end_mask_0 = const()[name = tensor("new_cache_9_internal_tensor_assign_1_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_9_internal_tensor_assign_1_squeeze_mask_0 = const()[name = tensor("new_cache_9_internal_tensor_assign_1_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_56 = const()[name = tensor("shape_56"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_8 = const()[name = tensor("reduce_prod_8"), val = tensor(1048576)]; + tensor range_1d_8_start_0 = const()[name = tensor("range_1d_8_start_0"), val = tensor(0)]; + tensor range_1d_8_step_0 = const()[name = tensor("range_1d_8_step_0"), val = tensor(1)]; + tensor range_1d_8 = range_1d(end = reduce_prod_8, start = range_1d_8_start_0, step = range_1d_8_step_0)[name = tensor("range_1d_8")]; + tensor reshape_40 = reshape(shape = shape_56, x = range_1d_8)[name = tensor("reshape_40")]; + tensor slice_by_index_8 = slice_by_index(begin = concat_30, begin_mask = new_cache_9_internal_tensor_assign_1_begin_mask_0, end = concat_31, end_mask = new_cache_9_internal_tensor_assign_1_end_mask_0, squeeze_mask = new_cache_9_internal_tensor_assign_1_squeeze_mask_0, stride = new_cache_9_internal_tensor_assign_1_stride_0, x = reshape_40)[name = tensor("slice_by_index_8")]; + tensor reshape_41_shape_0 = const()[name = tensor("reshape_41_shape_0"), val = tensor([-1])]; + tensor reshape_41 = reshape(shape = reshape_41_shape_0, x = slice_by_index_8)[name = tensor("reshape_41")]; + tensor reshape_42_shape_0 = const()[name = tensor("reshape_42_shape_0"), val = tensor([-1])]; + tensor reshape_42_cast_fp16 = reshape(shape = reshape_42_shape_0, x = var_2152_cast_fp16)[name = tensor("reshape_42_cast_fp16")]; + tensor reshape_43_shape_0 = const()[name = tensor("reshape_43_shape_0"), val = tensor([-1])]; + tensor reshape_43_cast_fp16 = reshape(shape = reshape_43_shape_0, x = cache4_to_fp16)[name = tensor("reshape_43_cast_fp16")]; + tensor scatter_8_mode_0 = const()[name = tensor("scatter_8_mode_0"), val = tensor("update")]; + tensor scatter_8_axis_0 = const()[name = tensor("scatter_8_axis_0"), val = tensor(0)]; + tensor scatter_8_validate_indices_0 = const()[name = tensor("scatter_8_validate_indices_0"), val = tensor(false)]; + tensor scatter_8_cast_fp16 = scatter(axis = scatter_8_axis_0, data = reshape_43_cast_fp16, indices = reshape_41, mode = scatter_8_mode_0, updates = reshape_42_cast_fp16, validate_indices = scatter_8_validate_indices_0)[name = tensor("scatter_8_cast_fp16")]; + tensor reshape_44_cast_fp16 = reshape(shape = shape_56, x = scatter_8_cast_fp16)[name = tensor("reshape_44_cast_fp16")]; + tensor var_2160_begin_0 = const()[name = tensor("op_2160_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor var_2160_end_0 = const()[name = tensor("op_2160_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor var_2160_end_mask_0 = const()[name = tensor("op_2160_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_2160_squeeze_mask_0 = const()[name = tensor("op_2160_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor var_2160_cast_fp16 = slice_by_index(begin = var_2160_begin_0, end = var_2160_end_0, end_mask = var_2160_end_mask_0, squeeze_mask = var_2160_squeeze_mask_0, x = reshape_44_cast_fp16)[name = tensor("op_2160_cast_fp16")]; + tensor var_2162_axis_0 = const()[name = tensor("op_2162_axis_0"), val = tensor(1)]; + tensor var_2162_mode_0 = const()[name = tensor("op_2162_mode_0"), val = tensor("update")]; + tensor var_2162_validate_indices_0 = const()[name = tensor("op_2162_validate_indices_0"), val = tensor(false)]; + tensor var_2162_cast_fp16 = scatter_along_axis(axis = var_2162_axis_0, data = var_2160_cast_fp16, indices = write_indices_9, mode = var_2162_mode_0, updates = v_9_cast_fp16, validate_indices = var_2162_validate_indices_0)[name = tensor("op_2162_cast_fp16")]; + tensor concat_32 = const()[name = tensor("concat_32"), val = tensor([1, 0, 0, 0, 0])]; + tensor concat_33 = const()[name = tensor("concat_33"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_9_internal_tensor_assign_2_stride_0 = const()[name = tensor("new_cache_9_internal_tensor_assign_2_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_9_internal_tensor_assign_2_begin_mask_0 = const()[name = tensor("new_cache_9_internal_tensor_assign_2_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_9_internal_tensor_assign_2_end_mask_0 = const()[name = tensor("new_cache_9_internal_tensor_assign_2_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_9_internal_tensor_assign_2_squeeze_mask_0 = const()[name = tensor("new_cache_9_internal_tensor_assign_2_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_57 = const()[name = tensor("shape_57"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_9 = const()[name = tensor("reduce_prod_9"), val = tensor(1048576)]; + tensor range_1d_9_start_0 = const()[name = tensor("range_1d_9_start_0"), val = tensor(0)]; + tensor range_1d_9_step_0 = const()[name = tensor("range_1d_9_step_0"), val = tensor(1)]; + tensor range_1d_9 = range_1d(end = reduce_prod_9, start = range_1d_9_start_0, step = range_1d_9_step_0)[name = tensor("range_1d_9")]; + tensor reshape_45 = reshape(shape = shape_57, x = range_1d_9)[name = tensor("reshape_45")]; + tensor slice_by_index_9 = slice_by_index(begin = concat_32, begin_mask = new_cache_9_internal_tensor_assign_2_begin_mask_0, end = concat_33, end_mask = new_cache_9_internal_tensor_assign_2_end_mask_0, squeeze_mask = new_cache_9_internal_tensor_assign_2_squeeze_mask_0, stride = new_cache_9_internal_tensor_assign_2_stride_0, x = reshape_45)[name = tensor("slice_by_index_9")]; + tensor reshape_46_shape_0 = const()[name = tensor("reshape_46_shape_0"), val = tensor([-1])]; + tensor reshape_46 = reshape(shape = reshape_46_shape_0, x = slice_by_index_9)[name = tensor("reshape_46")]; + tensor reshape_47_shape_0 = const()[name = tensor("reshape_47_shape_0"), val = tensor([-1])]; + tensor reshape_47_cast_fp16 = reshape(shape = reshape_47_shape_0, x = var_2162_cast_fp16)[name = tensor("reshape_47_cast_fp16")]; + tensor reshape_48_shape_0 = const()[name = tensor("reshape_48_shape_0"), val = tensor([-1])]; + tensor reshape_48_cast_fp16 = reshape(shape = reshape_48_shape_0, x = reshape_44_cast_fp16)[name = tensor("reshape_48_cast_fp16")]; + tensor scatter_9_mode_0 = const()[name = tensor("scatter_9_mode_0"), val = tensor("update")]; + tensor scatter_9_axis_0 = const()[name = tensor("scatter_9_axis_0"), val = tensor(0)]; + tensor scatter_9_validate_indices_0 = const()[name = tensor("scatter_9_validate_indices_0"), val = tensor(false)]; + tensor scatter_9_cast_fp16 = scatter(axis = scatter_9_axis_0, data = reshape_48_cast_fp16, indices = reshape_46, mode = scatter_9_mode_0, updates = reshape_47_cast_fp16, validate_indices = scatter_9_validate_indices_0)[name = tensor("scatter_9_cast_fp16")]; + tensor reshape_49_cast_fp16 = reshape(shape = shape_57, x = scatter_9_cast_fp16)[name = tensor("reshape_49_cast_fp16")]; + tensor reshape_49_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("reshape_49_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor keys_25_begin_0 = const()[name = tensor("keys_25_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor keys_25_end_0 = const()[name = tensor("keys_25_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor keys_25_end_mask_0 = const()[name = tensor("keys_25_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor keys_25_squeeze_mask_0 = const()[name = tensor("keys_25_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor keys_25_cast_fp16 = slice_by_index(begin = keys_25_begin_0, end = keys_25_end_0, end_mask = keys_25_end_mask_0, squeeze_mask = keys_25_squeeze_mask_0, x = reshape_49_cast_fp16)[name = tensor("keys_25_cast_fp16")]; + tensor values_25_begin_0 = const()[name = tensor("values_25_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor values_25_end_0 = const()[name = tensor("values_25_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor values_25_end_mask_0 = const()[name = tensor("values_25_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor values_25_squeeze_mask_0 = const()[name = tensor("values_25_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor values_25_cast_fp16 = slice_by_index(begin = values_25_begin_0, end = values_25_end_0, end_mask = values_25_end_mask_0, squeeze_mask = values_25_squeeze_mask_0, x = reshape_49_cast_fp16)[name = tensor("values_25_cast_fp16")]; + tensor var_2174_cast_fp16 = not_equal(x = keys_25_cast_fp16, y = keys_25_cast_fp16)[name = tensor("op_2174_cast_fp16")]; + tensor keys_27_cast_fp16 = select(a = var_504_to_fp16, b = keys_25_cast_fp16, cond = var_2174_cast_fp16)[name = tensor("keys_27_cast_fp16")]; + tensor var_2182_cast_fp16 = not_equal(x = values_25_cast_fp16, y = values_25_cast_fp16)[name = tensor("op_2182_cast_fp16")]; + tensor values_27_cast_fp16 = select(a = var_504_to_fp16, b = values_25_cast_fp16, cond = var_2182_cast_fp16)[name = tensor("values_27_cast_fp16")]; + tensor var_2206 = const()[name = tensor("op_2206"), val = tensor([0, 2, 1, 3])]; + tensor var_2219 = const()[name = tensor("op_2219"), val = tensor([1, 1, 1])]; + tensor var_2220_cast_fp16 = reshape(shape = var_2219, x = position4_to_fp16)[name = tensor("op_2220_cast_fp16")]; + tensor var_2237_to_fp16 = const()[name = tensor("op_2237_to_fp16"), val = tensor(0x1p+0)]; + tensor valid_len_9_cast_fp16 = add(x = var_2220_cast_fp16, y = var_2237_to_fp16)[name = tensor("valid_len_9_cast_fp16")]; + tensor valid_mask_9_cast_fp16 = less(x = k_positions_1_promoted_to_fp16, y = valid_len_9_cast_fp16)[name = tensor("valid_mask_9_cast_fp16")]; + tensor causal_mask_9_cast_fp16 = less_equal(x = k_positions_1_promoted_to_fp16, y = var_2220_cast_fp16)[name = tensor("causal_mask_9_cast_fp16")]; + tensor attn_mask_17 = logical_and(x = valid_mask_9_cast_fp16, y = causal_mask_9_cast_fp16)[name = tensor("attn_mask_17")]; + tensor attn_mask_19_axes_0 = const()[name = tensor("attn_mask_19_axes_0"), val = tensor([1])]; + tensor attn_mask_19 = expand_dims(axes = attn_mask_19_axes_0, x = attn_mask_17)[name = tensor("attn_mask_19")]; + tensor var_2255_transpose_x_0 = const()[name = tensor("op_2255_transpose_x_0"), val = tensor(false)]; + tensor var_2255_transpose_y_0 = const()[name = tensor("op_2255_transpose_y_0"), val = tensor(false)]; + tensor transpose_80_perm_0 = const()[name = tensor("transpose_80_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_81_perm_0 = const()[name = tensor("transpose_81_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_81 = transpose(perm = transpose_81_perm_0, x = keys_27_cast_fp16)[name = tensor("transpose_197")]; + tensor transpose_80 = transpose(perm = transpose_80_perm_0, x = q_27_cast_fp16)[name = tensor("transpose_198")]; + tensor var_2255_cast_fp16 = matmul(transpose_x = var_2255_transpose_x_0, transpose_y = var_2255_transpose_y_0, x = transpose_80, y = transpose_81)[name = tensor("op_2255_cast_fp16")]; + tensor var_2249_to_fp16 = const()[name = tensor("op_2249_to_fp16"), val = tensor([0x1p-3])]; + tensor attn_weights_25_cast_fp16 = mul(x = var_2255_cast_fp16, y = var_2249_to_fp16)[name = tensor("attn_weights_25_cast_fp16")]; + tensor var_2257 = logical_not(x = attn_mask_19)[name = tensor("op_2257")]; + tensor var_2258_to_fp16 = const()[name = tensor("op_2258_to_fp16"), val = tensor(-inf)]; + tensor attn_weights_27_cast_fp16 = select(a = var_2258_to_fp16, b = attn_weights_25_cast_fp16, cond = var_2257)[name = tensor("attn_weights_27_cast_fp16")]; + tensor var_2260 = const()[name = tensor("op_2260"), val = tensor(-1)]; + tensor attn_weights_29_cast_fp16 = softmax(axis = var_2260, x = attn_weights_27_cast_fp16)[name = tensor("attn_weights_29_cast_fp16")]; + tensor attn_output_9_transpose_x_0 = const()[name = tensor("attn_output_9_transpose_x_0"), val = tensor(false)]; + tensor attn_output_9_transpose_y_0 = const()[name = tensor("attn_output_9_transpose_y_0"), val = tensor(false)]; + tensor values_29_cast_fp16 = transpose(perm = var_2206, x = values_27_cast_fp16)[name = tensor("transpose_199")]; + tensor attn_output_9_cast_fp16 = matmul(transpose_x = attn_output_9_transpose_x_0, transpose_y = attn_output_9_transpose_y_0, x = attn_weights_29_cast_fp16, y = values_29_cast_fp16)[name = tensor("attn_output_9_cast_fp16")]; + tensor var_2268 = const()[name = tensor("op_2268"), val = tensor([0, 2, 1, 3])]; + tensor var_2271 = const()[name = tensor("op_2271"), val = tensor([1, 1, 1024])]; + tensor var_2269_cast_fp16 = transpose(perm = var_2268, x = attn_output_9_cast_fp16)[name = tensor("transpose_196")]; + tensor input_45_cast_fp16 = reshape(shape = var_2271, x = var_2269_cast_fp16)[name = tensor("input_45_cast_fp16")]; + tensor linear_18_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = attn4_out_proj_weight_cast_fp16, x = input_45_cast_fp16)[name = tensor("linear_18_cast_fp16")]; + tensor var_2277_to_fp16 = const()[name = tensor("op_2277_to_fp16"), val = tensor(0x1p+0)]; + tensor var_2278_cast_fp16 = add(x = position4_to_fp16, y = var_2277_to_fp16)[name = tensor("op_2278_cast_fp16")]; + tensor var_2278_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_2278_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor input_47_cast_fp16 = add(x = input_43_cast_fp16, y = linear_18_cast_fp16)[name = tensor("input_47_cast_fp16")]; + tensor input_49_axes_0 = const()[name = tensor("input_49_axes_0"), val = tensor([-1])]; + tensor norm4_2_weight_to_fp16 = const()[name = tensor("norm4_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(303623552)))]; + tensor norm4_2_bias_to_fp16 = const()[name = tensor("norm4_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(303625664)))]; + tensor var_2282_to_fp16 = const()[name = tensor("op_2282_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_49_cast_fp16 = layer_norm(axes = input_49_axes_0, beta = norm4_2_bias_to_fp16, epsilon = var_2282_to_fp16, gamma = norm4_2_weight_to_fp16, x = input_47_cast_fp16)[name = tensor("input_49_cast_fp16")]; + tensor linear_19_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = linear4_1_weight_cast_fp16, x = input_49_cast_fp16)[name = tensor("linear_19_cast_fp16")]; + tensor input_51_mode_0 = const()[name = tensor("input_51_mode_0"), val = tensor("EXACT")]; + tensor input_51_cast_fp16 = gelu(mode = input_51_mode_0, x = linear_19_cast_fp16)[name = tensor("input_51_cast_fp16")]; + tensor linear_20_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = linear4_2_weight_cast_fp16, x = input_51_cast_fp16)[name = tensor("linear_20_cast_fp16")]; + tensor input_53_cast_fp16 = add(x = input_47_cast_fp16, y = linear_20_cast_fp16)[name = tensor("input_53_cast_fp16")]; + tensor x_11_axes_0 = const()[name = tensor("x_11_axes_0"), val = tensor([-1])]; + tensor norm5_1_weight_to_fp16 = const()[name = tensor("norm5_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(303627776)))]; + tensor norm5_1_bias_to_fp16 = const()[name = tensor("norm5_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(303629888)))]; + tensor var_2299_to_fp16 = const()[name = tensor("op_2299_to_fp16"), val = tensor(0x1.5p-17)]; + tensor x_11_cast_fp16 = layer_norm(axes = x_11_axes_0, beta = norm5_1_bias_to_fp16, epsilon = var_2299_to_fp16, gamma = norm5_1_weight_to_fp16, x = input_53_cast_fp16)[name = tensor("x_11_cast_fp16")]; + tensor linear_21_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = attn5_in_proj_weight_cast_fp16, x = x_11_cast_fp16)[name = tensor("linear_21_cast_fp16")]; + tensor var_2335 = const()[name = tensor("op_2335"), val = tensor([1, 1, 3, 16, 64])]; + tensor qkv_11_cast_fp16 = reshape(shape = var_2335, x = linear_21_cast_fp16)[name = tensor("qkv_11_cast_fp16")]; + tensor q_31_begin_0 = const()[name = tensor("q_31_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor q_31_end_0 = const()[name = tensor("q_31_end_0"), val = tensor([1, 1, 1, 16, 64])]; + tensor q_31_end_mask_0 = const()[name = tensor("q_31_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor q_31_squeeze_mask_0 = const()[name = tensor("q_31_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor q_31_cast_fp16 = slice_by_index(begin = q_31_begin_0, end = q_31_end_0, end_mask = q_31_end_mask_0, squeeze_mask = q_31_squeeze_mask_0, x = qkv_11_cast_fp16)[name = tensor("q_31_cast_fp16")]; + tensor k_21_begin_0 = const()[name = tensor("k_21_begin_0"), val = tensor([0, 0, 1, 0, 0])]; + tensor k_21_end_0 = const()[name = tensor("k_21_end_0"), val = tensor([1, 1, 2, 16, 64])]; + tensor k_21_end_mask_0 = const()[name = tensor("k_21_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor k_21_squeeze_mask_0 = const()[name = tensor("k_21_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor k_21_cast_fp16 = slice_by_index(begin = k_21_begin_0, end = k_21_end_0, end_mask = k_21_end_mask_0, squeeze_mask = k_21_squeeze_mask_0, x = qkv_11_cast_fp16)[name = tensor("k_21_cast_fp16")]; + tensor v_11_begin_0 = const()[name = tensor("v_11_begin_0"), val = tensor([0, 0, 2, 0, 0])]; + tensor v_11_end_0 = const()[name = tensor("v_11_end_0"), val = tensor([1, 1, 3, 16, 64])]; + tensor v_11_end_mask_0 = const()[name = tensor("v_11_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor v_11_squeeze_mask_0 = const()[name = tensor("v_11_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor v_11_cast_fp16 = slice_by_index(begin = v_11_begin_0, end = v_11_end_0, end_mask = v_11_end_mask_0, squeeze_mask = v_11_squeeze_mask_0, x = qkv_11_cast_fp16)[name = tensor("v_11_cast_fp16")]; + tensor var_2439 = const()[name = tensor("op_2439"), val = tensor([1, 1, 1, 1])]; + tensor position5_to_fp16_dtype_0 = const()[name = tensor("position5_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor position5_to_fp16 = cast(dtype = position5_to_fp16_dtype_0, x = position5)[name = tensor("cast_528")]; + tensor ts_35_cast_fp16 = reshape(shape = var_2439, x = position5_to_fp16)[name = tensor("ts_35_cast_fp16")]; + tensor var_2443 = const()[name = tensor("op_2443"), val = tensor([1, 1, 16, 32, 2])]; + tensor q_complex_11_cast_fp16 = reshape(shape = var_2443, x = q_31_cast_fp16)[name = tensor("q_complex_11_cast_fp16")]; + tensor var_2447 = const()[name = tensor("op_2447"), val = tensor([1, 1, 16, 32, 2])]; + tensor k_complex_11_cast_fp16 = reshape(shape = var_2447, x = k_21_cast_fp16)[name = tensor("k_complex_11_cast_fp16")]; + tensor var_2451_begin_0 = const()[name = tensor("op_2451_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_2451_end_0 = const()[name = tensor("op_2451_end_0"), val = tensor([1, 1, 16, 32, 1])]; + tensor var_2451_end_mask_0 = const()[name = tensor("op_2451_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_2451_squeeze_mask_0 = const()[name = tensor("op_2451_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_2451_cast_fp16 = slice_by_index(begin = var_2451_begin_0, end = var_2451_end_0, end_mask = var_2451_end_mask_0, squeeze_mask = var_2451_squeeze_mask_0, x = q_complex_11_cast_fp16)[name = tensor("op_2451_cast_fp16")]; + tensor var_2459_begin_0 = const()[name = tensor("op_2459_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_2459_end_0 = const()[name = tensor("op_2459_end_0"), val = tensor([1, 1, 16, 32, 2])]; + tensor var_2459_end_mask_0 = const()[name = tensor("op_2459_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_2459_squeeze_mask_0 = const()[name = tensor("op_2459_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_2459_cast_fp16 = slice_by_index(begin = var_2459_begin_0, end = var_2459_end_0, end_mask = var_2459_end_mask_0, squeeze_mask = var_2459_squeeze_mask_0, x = q_complex_11_cast_fp16)[name = tensor("op_2459_cast_fp16")]; + tensor var_2467_begin_0 = const()[name = tensor("op_2467_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_2467_end_0 = const()[name = tensor("op_2467_end_0"), val = tensor([1, 1, 16, 32, 1])]; + tensor var_2467_end_mask_0 = const()[name = tensor("op_2467_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_2467_squeeze_mask_0 = const()[name = tensor("op_2467_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_2467_cast_fp16 = slice_by_index(begin = var_2467_begin_0, end = var_2467_end_0, end_mask = var_2467_end_mask_0, squeeze_mask = var_2467_squeeze_mask_0, x = k_complex_11_cast_fp16)[name = tensor("op_2467_cast_fp16")]; + tensor var_2475_begin_0 = const()[name = tensor("op_2475_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_2475_end_0 = const()[name = tensor("op_2475_end_0"), val = tensor([1, 1, 16, 32, 2])]; + tensor var_2475_end_mask_0 = const()[name = tensor("op_2475_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_2475_squeeze_mask_0 = const()[name = tensor("op_2475_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_2475_cast_fp16 = slice_by_index(begin = var_2475_begin_0, end = var_2475_end_0, end_mask = var_2475_end_mask_0, squeeze_mask = var_2475_squeeze_mask_0, x = k_complex_11_cast_fp16)[name = tensor("op_2475_cast_fp16")]; + tensor freqs_11_to_fp16 = const()[name = tensor("freqs_11_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(303632000)))]; + tensor var_2481_cast_fp16 = mul(x = freqs_11_to_fp16, y = ts_35_cast_fp16)[name = tensor("op_2481_cast_fp16")]; + tensor rotr_11_cast_fp16 = cos(x = var_2481_cast_fp16)[name = tensor("rotr_11_cast_fp16")]; + tensor roti_11_cast_fp16 = sin(x = var_2481_cast_fp16)[name = tensor("roti_11_cast_fp16")]; + tensor var_2485_cast_fp16 = mul(x = var_2451_cast_fp16, y = rotr_11_cast_fp16)[name = tensor("op_2485_cast_fp16")]; + tensor var_2486_cast_fp16 = mul(x = var_2459_cast_fp16, y = roti_11_cast_fp16)[name = tensor("op_2486_cast_fp16")]; + tensor qor_21_cast_fp16 = sub(x = var_2485_cast_fp16, y = var_2486_cast_fp16)[name = tensor("qor_21_cast_fp16")]; + tensor var_2489_cast_fp16 = mul(x = var_2451_cast_fp16, y = roti_11_cast_fp16)[name = tensor("op_2489_cast_fp16")]; + tensor var_2490_cast_fp16 = mul(x = var_2459_cast_fp16, y = rotr_11_cast_fp16)[name = tensor("op_2490_cast_fp16")]; + tensor qoi_21_cast_fp16 = add(x = var_2489_cast_fp16, y = var_2490_cast_fp16)[name = tensor("qoi_21_cast_fp16")]; + tensor var_2493_cast_fp16 = mul(x = var_2467_cast_fp16, y = rotr_11_cast_fp16)[name = tensor("op_2493_cast_fp16")]; + tensor var_2494_cast_fp16 = mul(x = var_2475_cast_fp16, y = roti_11_cast_fp16)[name = tensor("op_2494_cast_fp16")]; + tensor kor_21_cast_fp16 = sub(x = var_2493_cast_fp16, y = var_2494_cast_fp16)[name = tensor("kor_21_cast_fp16")]; + tensor var_2497_cast_fp16 = mul(x = var_2467_cast_fp16, y = roti_11_cast_fp16)[name = tensor("op_2497_cast_fp16")]; + tensor var_2498_cast_fp16 = mul(x = var_2475_cast_fp16, y = rotr_11_cast_fp16)[name = tensor("op_2498_cast_fp16")]; + tensor koi_21_cast_fp16 = add(x = var_2497_cast_fp16, y = var_2498_cast_fp16)[name = tensor("koi_21_cast_fp16")]; + tensor qo_11_axis_0 = const()[name = tensor("qo_11_axis_0"), val = tensor(-1)]; + tensor qo_11_cast_fp16 = stack(axis = qo_11_axis_0, values = (qor_21_cast_fp16, qoi_21_cast_fp16))[name = tensor("qo_11_cast_fp16")]; + tensor ko_11_axis_0 = const()[name = tensor("ko_11_axis_0"), val = tensor(-1)]; + tensor ko_11_cast_fp16 = stack(axis = ko_11_axis_0, values = (kor_21_cast_fp16, koi_21_cast_fp16))[name = tensor("ko_11_cast_fp16")]; + tensor var_2527 = const()[name = tensor("op_2527"), val = tensor([1, 1, 16, 64])]; + tensor q_33_cast_fp16 = reshape(shape = var_2527, x = qo_11_cast_fp16)[name = tensor("q_33_cast_fp16")]; + tensor var_2529 = const()[name = tensor("op_2529"), val = tensor([1, 1, 16, 64])]; + tensor k_23_cast_fp16 = reshape(shape = var_2529, x = ko_11_cast_fp16)[name = tensor("k_23_cast_fp16")]; + tensor _inversed_2551_y_0_to_fp16 = const()[name = tensor("_inversed_2551_y_0_to_fp16"), val = tensor(0x1p-9)]; + tensor _inversed_2551_cast_fp16 = mul(x = ts_35_cast_fp16, y = _inversed_2551_y_0_to_fp16)[name = tensor("_inversed_2551_cast_fp16")]; + tensor var_2552_cast_fp16 = floor(x = _inversed_2551_cast_fp16)[name = tensor("op_2552_cast_fp16")]; + tensor var_2553_to_fp16 = const()[name = tensor("op_2553_to_fp16"), val = tensor(0x1p+9)]; + tensor var_2554_cast_fp16 = mul(x = var_2552_cast_fp16, y = var_2553_to_fp16)[name = tensor("op_2554_cast_fp16")]; + tensor write_indices_float_23_cast_fp16 = sub(x = ts_35_cast_fp16, y = var_2554_cast_fp16)[name = tensor("write_indices_float_23_cast_fp16")]; + tensor var_2561_dtype_0 = const()[name = tensor("op_2561_dtype_0"), val = tensor("int32")]; + tensor write_indices_11_reps_0 = const()[name = tensor("write_indices_11_reps_0"), val = tensor([1, 1, 16, 64])]; + tensor write_indices_float_23_cast_fp16_to_int32 = cast(dtype = var_2561_dtype_0, x = write_indices_float_23_cast_fp16)[name = tensor("cast_527")]; + tensor write_indices_11 = tile(reps = write_indices_11_reps_0, x = write_indices_float_23_cast_fp16_to_int32)[name = tensor("write_indices_11")]; + tensor var_2569_begin_0 = const()[name = tensor("op_2569_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_2569_end_0 = const()[name = tensor("op_2569_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor var_2569_end_mask_0 = const()[name = tensor("op_2569_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_2569_squeeze_mask_0 = const()[name = tensor("op_2569_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor cache5_to_fp16_dtype_0 = const()[name = tensor("cache5_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor cache5_to_fp16 = cast(dtype = cache5_to_fp16_dtype_0, x = cache5)[name = tensor("cast_526")]; + tensor var_2569_cast_fp16 = slice_by_index(begin = var_2569_begin_0, end = var_2569_end_0, end_mask = var_2569_end_mask_0, squeeze_mask = var_2569_squeeze_mask_0, x = cache5_to_fp16)[name = tensor("op_2569_cast_fp16")]; + tensor var_2571_axis_0 = const()[name = tensor("op_2571_axis_0"), val = tensor(1)]; + tensor var_2571_mode_0 = const()[name = tensor("op_2571_mode_0"), val = tensor("update")]; + tensor var_2571_validate_indices_0 = const()[name = tensor("op_2571_validate_indices_0"), val = tensor(false)]; + tensor var_2571_cast_fp16 = scatter_along_axis(axis = var_2571_axis_0, data = var_2569_cast_fp16, indices = write_indices_11, mode = var_2571_mode_0, updates = k_23_cast_fp16, validate_indices = var_2571_validate_indices_0)[name = tensor("op_2571_cast_fp16")]; + tensor concat_37 = const()[name = tensor("concat_37"), val = tensor([0, 0, 0, 0, 0])]; + tensor concat_38 = const()[name = tensor("concat_38"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_11_internal_tensor_assign_1_stride_0 = const()[name = tensor("new_cache_11_internal_tensor_assign_1_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_11_internal_tensor_assign_1_begin_mask_0 = const()[name = tensor("new_cache_11_internal_tensor_assign_1_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_11_internal_tensor_assign_1_end_mask_0 = const()[name = tensor("new_cache_11_internal_tensor_assign_1_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_11_internal_tensor_assign_1_squeeze_mask_0 = const()[name = tensor("new_cache_11_internal_tensor_assign_1_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_58 = const()[name = tensor("shape_58"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_10 = const()[name = tensor("reduce_prod_10"), val = tensor(1048576)]; + tensor range_1d_10_start_0 = const()[name = tensor("range_1d_10_start_0"), val = tensor(0)]; + tensor range_1d_10_step_0 = const()[name = tensor("range_1d_10_step_0"), val = tensor(1)]; + tensor range_1d_10 = range_1d(end = reduce_prod_10, start = range_1d_10_start_0, step = range_1d_10_step_0)[name = tensor("range_1d_10")]; + tensor reshape_50 = reshape(shape = shape_58, x = range_1d_10)[name = tensor("reshape_50")]; + tensor slice_by_index_10 = slice_by_index(begin = concat_37, begin_mask = new_cache_11_internal_tensor_assign_1_begin_mask_0, end = concat_38, end_mask = new_cache_11_internal_tensor_assign_1_end_mask_0, squeeze_mask = new_cache_11_internal_tensor_assign_1_squeeze_mask_0, stride = new_cache_11_internal_tensor_assign_1_stride_0, x = reshape_50)[name = tensor("slice_by_index_10")]; + tensor reshape_51_shape_0 = const()[name = tensor("reshape_51_shape_0"), val = tensor([-1])]; + tensor reshape_51 = reshape(shape = reshape_51_shape_0, x = slice_by_index_10)[name = tensor("reshape_51")]; + tensor reshape_52_shape_0 = const()[name = tensor("reshape_52_shape_0"), val = tensor([-1])]; + tensor reshape_52_cast_fp16 = reshape(shape = reshape_52_shape_0, x = var_2571_cast_fp16)[name = tensor("reshape_52_cast_fp16")]; + tensor reshape_53_shape_0 = const()[name = tensor("reshape_53_shape_0"), val = tensor([-1])]; + tensor reshape_53_cast_fp16 = reshape(shape = reshape_53_shape_0, x = cache5_to_fp16)[name = tensor("reshape_53_cast_fp16")]; + tensor scatter_10_mode_0 = const()[name = tensor("scatter_10_mode_0"), val = tensor("update")]; + tensor scatter_10_axis_0 = const()[name = tensor("scatter_10_axis_0"), val = tensor(0)]; + tensor scatter_10_validate_indices_0 = const()[name = tensor("scatter_10_validate_indices_0"), val = tensor(false)]; + tensor scatter_10_cast_fp16 = scatter(axis = scatter_10_axis_0, data = reshape_53_cast_fp16, indices = reshape_51, mode = scatter_10_mode_0, updates = reshape_52_cast_fp16, validate_indices = scatter_10_validate_indices_0)[name = tensor("scatter_10_cast_fp16")]; + tensor reshape_54_cast_fp16 = reshape(shape = shape_58, x = scatter_10_cast_fp16)[name = tensor("reshape_54_cast_fp16")]; + tensor var_2579_begin_0 = const()[name = tensor("op_2579_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor var_2579_end_0 = const()[name = tensor("op_2579_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor var_2579_end_mask_0 = const()[name = tensor("op_2579_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_2579_squeeze_mask_0 = const()[name = tensor("op_2579_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor var_2579_cast_fp16 = slice_by_index(begin = var_2579_begin_0, end = var_2579_end_0, end_mask = var_2579_end_mask_0, squeeze_mask = var_2579_squeeze_mask_0, x = reshape_54_cast_fp16)[name = tensor("op_2579_cast_fp16")]; + tensor var_2581_axis_0 = const()[name = tensor("op_2581_axis_0"), val = tensor(1)]; + tensor var_2581_mode_0 = const()[name = tensor("op_2581_mode_0"), val = tensor("update")]; + tensor var_2581_validate_indices_0 = const()[name = tensor("op_2581_validate_indices_0"), val = tensor(false)]; + tensor var_2581_cast_fp16 = scatter_along_axis(axis = var_2581_axis_0, data = var_2579_cast_fp16, indices = write_indices_11, mode = var_2581_mode_0, updates = v_11_cast_fp16, validate_indices = var_2581_validate_indices_0)[name = tensor("op_2581_cast_fp16")]; + tensor concat_39 = const()[name = tensor("concat_39"), val = tensor([1, 0, 0, 0, 0])]; + tensor concat_40 = const()[name = tensor("concat_40"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_11_internal_tensor_assign_2_stride_0 = const()[name = tensor("new_cache_11_internal_tensor_assign_2_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_11_internal_tensor_assign_2_begin_mask_0 = const()[name = tensor("new_cache_11_internal_tensor_assign_2_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_11_internal_tensor_assign_2_end_mask_0 = const()[name = tensor("new_cache_11_internal_tensor_assign_2_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_11_internal_tensor_assign_2_squeeze_mask_0 = const()[name = tensor("new_cache_11_internal_tensor_assign_2_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_59 = const()[name = tensor("shape_59"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_11 = const()[name = tensor("reduce_prod_11"), val = tensor(1048576)]; + tensor range_1d_11_start_0 = const()[name = tensor("range_1d_11_start_0"), val = tensor(0)]; + tensor range_1d_11_step_0 = const()[name = tensor("range_1d_11_step_0"), val = tensor(1)]; + tensor range_1d_11 = range_1d(end = reduce_prod_11, start = range_1d_11_start_0, step = range_1d_11_step_0)[name = tensor("range_1d_11")]; + tensor reshape_55 = reshape(shape = shape_59, x = range_1d_11)[name = tensor("reshape_55")]; + tensor slice_by_index_11 = slice_by_index(begin = concat_39, begin_mask = new_cache_11_internal_tensor_assign_2_begin_mask_0, end = concat_40, end_mask = new_cache_11_internal_tensor_assign_2_end_mask_0, squeeze_mask = new_cache_11_internal_tensor_assign_2_squeeze_mask_0, stride = new_cache_11_internal_tensor_assign_2_stride_0, x = reshape_55)[name = tensor("slice_by_index_11")]; + tensor reshape_56_shape_0 = const()[name = tensor("reshape_56_shape_0"), val = tensor([-1])]; + tensor reshape_56 = reshape(shape = reshape_56_shape_0, x = slice_by_index_11)[name = tensor("reshape_56")]; + tensor reshape_57_shape_0 = const()[name = tensor("reshape_57_shape_0"), val = tensor([-1])]; + tensor reshape_57_cast_fp16 = reshape(shape = reshape_57_shape_0, x = var_2581_cast_fp16)[name = tensor("reshape_57_cast_fp16")]; + tensor reshape_58_shape_0 = const()[name = tensor("reshape_58_shape_0"), val = tensor([-1])]; + tensor reshape_58_cast_fp16 = reshape(shape = reshape_58_shape_0, x = reshape_54_cast_fp16)[name = tensor("reshape_58_cast_fp16")]; + tensor scatter_11_mode_0 = const()[name = tensor("scatter_11_mode_0"), val = tensor("update")]; + tensor scatter_11_axis_0 = const()[name = tensor("scatter_11_axis_0"), val = tensor(0)]; + tensor scatter_11_validate_indices_0 = const()[name = tensor("scatter_11_validate_indices_0"), val = tensor(false)]; + tensor scatter_11_cast_fp16 = scatter(axis = scatter_11_axis_0, data = reshape_58_cast_fp16, indices = reshape_56, mode = scatter_11_mode_0, updates = reshape_57_cast_fp16, validate_indices = scatter_11_validate_indices_0)[name = tensor("scatter_11_cast_fp16")]; + tensor reshape_59_cast_fp16 = reshape(shape = shape_59, x = scatter_11_cast_fp16)[name = tensor("reshape_59_cast_fp16")]; + tensor reshape_59_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("reshape_59_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor keys_31_begin_0 = const()[name = tensor("keys_31_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor keys_31_end_0 = const()[name = tensor("keys_31_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor keys_31_end_mask_0 = const()[name = tensor("keys_31_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor keys_31_squeeze_mask_0 = const()[name = tensor("keys_31_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor keys_31_cast_fp16 = slice_by_index(begin = keys_31_begin_0, end = keys_31_end_0, end_mask = keys_31_end_mask_0, squeeze_mask = keys_31_squeeze_mask_0, x = reshape_59_cast_fp16)[name = tensor("keys_31_cast_fp16")]; + tensor values_31_begin_0 = const()[name = tensor("values_31_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor values_31_end_0 = const()[name = tensor("values_31_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor values_31_end_mask_0 = const()[name = tensor("values_31_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor values_31_squeeze_mask_0 = const()[name = tensor("values_31_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor values_31_cast_fp16 = slice_by_index(begin = values_31_begin_0, end = values_31_end_0, end_mask = values_31_end_mask_0, squeeze_mask = values_31_squeeze_mask_0, x = reshape_59_cast_fp16)[name = tensor("values_31_cast_fp16")]; + tensor var_2593_cast_fp16 = not_equal(x = keys_31_cast_fp16, y = keys_31_cast_fp16)[name = tensor("op_2593_cast_fp16")]; + tensor keys_33_cast_fp16 = select(a = var_504_to_fp16, b = keys_31_cast_fp16, cond = var_2593_cast_fp16)[name = tensor("keys_33_cast_fp16")]; + tensor var_2601_cast_fp16 = not_equal(x = values_31_cast_fp16, y = values_31_cast_fp16)[name = tensor("op_2601_cast_fp16")]; + tensor values_33_cast_fp16 = select(a = var_504_to_fp16, b = values_31_cast_fp16, cond = var_2601_cast_fp16)[name = tensor("values_33_cast_fp16")]; + tensor var_2625 = const()[name = tensor("op_2625"), val = tensor([0, 2, 1, 3])]; + tensor var_2638 = const()[name = tensor("op_2638"), val = tensor([1, 1, 1])]; + tensor var_2639_cast_fp16 = reshape(shape = var_2638, x = position5_to_fp16)[name = tensor("op_2639_cast_fp16")]; + tensor var_2656_to_fp16 = const()[name = tensor("op_2656_to_fp16"), val = tensor(0x1p+0)]; + tensor valid_len_11_cast_fp16 = add(x = var_2639_cast_fp16, y = var_2656_to_fp16)[name = tensor("valid_len_11_cast_fp16")]; + tensor valid_mask_11_cast_fp16 = less(x = k_positions_1_promoted_to_fp16, y = valid_len_11_cast_fp16)[name = tensor("valid_mask_11_cast_fp16")]; + tensor causal_mask_11_cast_fp16 = less_equal(x = k_positions_1_promoted_to_fp16, y = var_2639_cast_fp16)[name = tensor("causal_mask_11_cast_fp16")]; + tensor attn_mask_21 = logical_and(x = valid_mask_11_cast_fp16, y = causal_mask_11_cast_fp16)[name = tensor("attn_mask_21")]; + tensor attn_mask_23_axes_0 = const()[name = tensor("attn_mask_23_axes_0"), val = tensor([1])]; + tensor attn_mask_23 = expand_dims(axes = attn_mask_23_axes_0, x = attn_mask_21)[name = tensor("attn_mask_23")]; + tensor var_2674_transpose_x_0 = const()[name = tensor("op_2674_transpose_x_0"), val = tensor(false)]; + tensor var_2674_transpose_y_0 = const()[name = tensor("op_2674_transpose_y_0"), val = tensor(false)]; + tensor transpose_82_perm_0 = const()[name = tensor("transpose_82_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_83_perm_0 = const()[name = tensor("transpose_83_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_83 = transpose(perm = transpose_83_perm_0, x = keys_33_cast_fp16)[name = tensor("transpose_193")]; + tensor transpose_82 = transpose(perm = transpose_82_perm_0, x = q_33_cast_fp16)[name = tensor("transpose_194")]; + tensor var_2674_cast_fp16 = matmul(transpose_x = var_2674_transpose_x_0, transpose_y = var_2674_transpose_y_0, x = transpose_82, y = transpose_83)[name = tensor("op_2674_cast_fp16")]; + tensor var_2668_to_fp16 = const()[name = tensor("op_2668_to_fp16"), val = tensor([0x1p-3])]; + tensor attn_weights_31_cast_fp16 = mul(x = var_2674_cast_fp16, y = var_2668_to_fp16)[name = tensor("attn_weights_31_cast_fp16")]; + tensor var_2676 = logical_not(x = attn_mask_23)[name = tensor("op_2676")]; + tensor var_2677_to_fp16 = const()[name = tensor("op_2677_to_fp16"), val = tensor(-inf)]; + tensor attn_weights_33_cast_fp16 = select(a = var_2677_to_fp16, b = attn_weights_31_cast_fp16, cond = var_2676)[name = tensor("attn_weights_33_cast_fp16")]; + tensor var_2679 = const()[name = tensor("op_2679"), val = tensor(-1)]; + tensor attn_weights_35_cast_fp16 = softmax(axis = var_2679, x = attn_weights_33_cast_fp16)[name = tensor("attn_weights_35_cast_fp16")]; + tensor attn_output_11_transpose_x_0 = const()[name = tensor("attn_output_11_transpose_x_0"), val = tensor(false)]; + tensor attn_output_11_transpose_y_0 = const()[name = tensor("attn_output_11_transpose_y_0"), val = tensor(false)]; + tensor values_35_cast_fp16 = transpose(perm = var_2625, x = values_33_cast_fp16)[name = tensor("transpose_195")]; + tensor attn_output_11_cast_fp16 = matmul(transpose_x = attn_output_11_transpose_x_0, transpose_y = attn_output_11_transpose_y_0, x = attn_weights_35_cast_fp16, y = values_35_cast_fp16)[name = tensor("attn_output_11_cast_fp16")]; + tensor var_2687 = const()[name = tensor("op_2687"), val = tensor([0, 2, 1, 3])]; + tensor var_2690 = const()[name = tensor("op_2690"), val = tensor([1, 1, 1024])]; + tensor var_2688_cast_fp16 = transpose(perm = var_2687, x = attn_output_11_cast_fp16)[name = tensor("transpose_192")]; + tensor input_55_cast_fp16 = reshape(shape = var_2690, x = var_2688_cast_fp16)[name = tensor("input_55_cast_fp16")]; + tensor linear_22_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = attn5_out_proj_weight_cast_fp16, x = input_55_cast_fp16)[name = tensor("linear_22_cast_fp16")]; + tensor var_2696_to_fp16 = const()[name = tensor("op_2696_to_fp16"), val = tensor(0x1p+0)]; + tensor var_2697_cast_fp16 = add(x = position5_to_fp16, y = var_2696_to_fp16)[name = tensor("op_2697_cast_fp16")]; + tensor var_2697_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_2697_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor input_57_cast_fp16 = add(x = input_53_cast_fp16, y = linear_22_cast_fp16)[name = tensor("input_57_cast_fp16")]; + tensor input_59_axes_0 = const()[name = tensor("input_59_axes_0"), val = tensor([-1])]; + tensor norm5_2_weight_to_fp16 = const()[name = tensor("norm5_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(303632128)))]; + tensor norm5_2_bias_to_fp16 = const()[name = tensor("norm5_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(303634240)))]; + tensor var_2701_to_fp16 = const()[name = tensor("op_2701_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_59_cast_fp16 = layer_norm(axes = input_59_axes_0, beta = norm5_2_bias_to_fp16, epsilon = var_2701_to_fp16, gamma = norm5_2_weight_to_fp16, x = input_57_cast_fp16)[name = tensor("input_59_cast_fp16")]; + tensor linear_23_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = linear5_1_weight_cast_fp16, x = input_59_cast_fp16)[name = tensor("linear_23_cast_fp16")]; + tensor input_61_mode_0 = const()[name = tensor("input_61_mode_0"), val = tensor("EXACT")]; + tensor input_61_cast_fp16 = gelu(mode = input_61_mode_0, x = linear_23_cast_fp16)[name = tensor("input_61_cast_fp16")]; + tensor linear_24_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = linear5_2_weight_cast_fp16, x = input_61_cast_fp16)[name = tensor("linear_24_cast_fp16")]; + tensor input_63_cast_fp16 = add(x = input_57_cast_fp16, y = linear_24_cast_fp16)[name = tensor("input_63_cast_fp16")]; + tensor x_13_axes_0 = const()[name = tensor("x_13_axes_0"), val = tensor([-1])]; + tensor norm6_1_weight_to_fp16 = const()[name = tensor("norm6_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(303636352)))]; + tensor norm6_1_bias_to_fp16 = const()[name = tensor("norm6_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(303638464)))]; + tensor var_2718_to_fp16 = const()[name = tensor("op_2718_to_fp16"), val = tensor(0x1.5p-17)]; + tensor x_13_cast_fp16 = layer_norm(axes = x_13_axes_0, beta = norm6_1_bias_to_fp16, epsilon = var_2718_to_fp16, gamma = norm6_1_weight_to_fp16, x = input_63_cast_fp16)[name = tensor("x_13_cast_fp16")]; + tensor linear_25_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = attn6_in_proj_weight_cast_fp16, x = x_13_cast_fp16)[name = tensor("linear_25_cast_fp16")]; + tensor var_2754 = const()[name = tensor("op_2754"), val = tensor([1, 1, 3, 16, 64])]; + tensor qkv_13_cast_fp16 = reshape(shape = var_2754, x = linear_25_cast_fp16)[name = tensor("qkv_13_cast_fp16")]; + tensor q_37_begin_0 = const()[name = tensor("q_37_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor q_37_end_0 = const()[name = tensor("q_37_end_0"), val = tensor([1, 1, 1, 16, 64])]; + tensor q_37_end_mask_0 = const()[name = tensor("q_37_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor q_37_squeeze_mask_0 = const()[name = tensor("q_37_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor q_37_cast_fp16 = slice_by_index(begin = q_37_begin_0, end = q_37_end_0, end_mask = q_37_end_mask_0, squeeze_mask = q_37_squeeze_mask_0, x = qkv_13_cast_fp16)[name = tensor("q_37_cast_fp16")]; + tensor k_25_begin_0 = const()[name = tensor("k_25_begin_0"), val = tensor([0, 0, 1, 0, 0])]; + tensor k_25_end_0 = const()[name = tensor("k_25_end_0"), val = tensor([1, 1, 2, 16, 64])]; + tensor k_25_end_mask_0 = const()[name = tensor("k_25_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor k_25_squeeze_mask_0 = const()[name = tensor("k_25_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor k_25_cast_fp16 = slice_by_index(begin = k_25_begin_0, end = k_25_end_0, end_mask = k_25_end_mask_0, squeeze_mask = k_25_squeeze_mask_0, x = qkv_13_cast_fp16)[name = tensor("k_25_cast_fp16")]; + tensor v_13_begin_0 = const()[name = tensor("v_13_begin_0"), val = tensor([0, 0, 2, 0, 0])]; + tensor v_13_end_0 = const()[name = tensor("v_13_end_0"), val = tensor([1, 1, 3, 16, 64])]; + tensor v_13_end_mask_0 = const()[name = tensor("v_13_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor v_13_squeeze_mask_0 = const()[name = tensor("v_13_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor v_13_cast_fp16 = slice_by_index(begin = v_13_begin_0, end = v_13_end_0, end_mask = v_13_end_mask_0, squeeze_mask = v_13_squeeze_mask_0, x = qkv_13_cast_fp16)[name = tensor("v_13_cast_fp16")]; + tensor var_2858 = const()[name = tensor("op_2858"), val = tensor([1, 1, 1, 1])]; + tensor position6_to_fp16_dtype_0 = const()[name = tensor("position6_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor position6_to_fp16 = cast(dtype = position6_to_fp16_dtype_0, x = position6)[name = tensor("cast_523")]; + tensor ts_41_cast_fp16 = reshape(shape = var_2858, x = position6_to_fp16)[name = tensor("ts_41_cast_fp16")]; + tensor var_2862 = const()[name = tensor("op_2862"), val = tensor([1, 1, 16, 32, 2])]; + tensor q_complex_13_cast_fp16 = reshape(shape = var_2862, x = q_37_cast_fp16)[name = tensor("q_complex_13_cast_fp16")]; + tensor var_2866 = const()[name = tensor("op_2866"), val = tensor([1, 1, 16, 32, 2])]; + tensor k_complex_13_cast_fp16 = reshape(shape = var_2866, x = k_25_cast_fp16)[name = tensor("k_complex_13_cast_fp16")]; + tensor var_2870_begin_0 = const()[name = tensor("op_2870_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_2870_end_0 = const()[name = tensor("op_2870_end_0"), val = tensor([1, 1, 16, 32, 1])]; + tensor var_2870_end_mask_0 = const()[name = tensor("op_2870_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_2870_squeeze_mask_0 = const()[name = tensor("op_2870_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_2870_cast_fp16 = slice_by_index(begin = var_2870_begin_0, end = var_2870_end_0, end_mask = var_2870_end_mask_0, squeeze_mask = var_2870_squeeze_mask_0, x = q_complex_13_cast_fp16)[name = tensor("op_2870_cast_fp16")]; + tensor var_2878_begin_0 = const()[name = tensor("op_2878_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_2878_end_0 = const()[name = tensor("op_2878_end_0"), val = tensor([1, 1, 16, 32, 2])]; + tensor var_2878_end_mask_0 = const()[name = tensor("op_2878_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_2878_squeeze_mask_0 = const()[name = tensor("op_2878_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_2878_cast_fp16 = slice_by_index(begin = var_2878_begin_0, end = var_2878_end_0, end_mask = var_2878_end_mask_0, squeeze_mask = var_2878_squeeze_mask_0, x = q_complex_13_cast_fp16)[name = tensor("op_2878_cast_fp16")]; + tensor var_2886_begin_0 = const()[name = tensor("op_2886_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_2886_end_0 = const()[name = tensor("op_2886_end_0"), val = tensor([1, 1, 16, 32, 1])]; + tensor var_2886_end_mask_0 = const()[name = tensor("op_2886_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_2886_squeeze_mask_0 = const()[name = tensor("op_2886_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_2886_cast_fp16 = slice_by_index(begin = var_2886_begin_0, end = var_2886_end_0, end_mask = var_2886_end_mask_0, squeeze_mask = var_2886_squeeze_mask_0, x = k_complex_13_cast_fp16)[name = tensor("op_2886_cast_fp16")]; + tensor var_2894_begin_0 = const()[name = tensor("op_2894_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_2894_end_0 = const()[name = tensor("op_2894_end_0"), val = tensor([1, 1, 16, 32, 2])]; + tensor var_2894_end_mask_0 = const()[name = tensor("op_2894_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_2894_squeeze_mask_0 = const()[name = tensor("op_2894_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_2894_cast_fp16 = slice_by_index(begin = var_2894_begin_0, end = var_2894_end_0, end_mask = var_2894_end_mask_0, squeeze_mask = var_2894_squeeze_mask_0, x = k_complex_13_cast_fp16)[name = tensor("op_2894_cast_fp16")]; + tensor freqs_13_to_fp16 = const()[name = tensor("freqs_13_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(303640576)))]; + tensor var_2900_cast_fp16 = mul(x = freqs_13_to_fp16, y = ts_41_cast_fp16)[name = tensor("op_2900_cast_fp16")]; + tensor rotr_13_cast_fp16 = cos(x = var_2900_cast_fp16)[name = tensor("rotr_13_cast_fp16")]; + tensor roti_13_cast_fp16 = sin(x = var_2900_cast_fp16)[name = tensor("roti_13_cast_fp16")]; + tensor var_2904_cast_fp16 = mul(x = var_2870_cast_fp16, y = rotr_13_cast_fp16)[name = tensor("op_2904_cast_fp16")]; + tensor var_2905_cast_fp16 = mul(x = var_2878_cast_fp16, y = roti_13_cast_fp16)[name = tensor("op_2905_cast_fp16")]; + tensor qor_25_cast_fp16 = sub(x = var_2904_cast_fp16, y = var_2905_cast_fp16)[name = tensor("qor_25_cast_fp16")]; + tensor var_2908_cast_fp16 = mul(x = var_2870_cast_fp16, y = roti_13_cast_fp16)[name = tensor("op_2908_cast_fp16")]; + tensor var_2909_cast_fp16 = mul(x = var_2878_cast_fp16, y = rotr_13_cast_fp16)[name = tensor("op_2909_cast_fp16")]; + tensor qoi_25_cast_fp16 = add(x = var_2908_cast_fp16, y = var_2909_cast_fp16)[name = tensor("qoi_25_cast_fp16")]; + tensor var_2912_cast_fp16 = mul(x = var_2886_cast_fp16, y = rotr_13_cast_fp16)[name = tensor("op_2912_cast_fp16")]; + tensor var_2913_cast_fp16 = mul(x = var_2894_cast_fp16, y = roti_13_cast_fp16)[name = tensor("op_2913_cast_fp16")]; + tensor kor_25_cast_fp16 = sub(x = var_2912_cast_fp16, y = var_2913_cast_fp16)[name = tensor("kor_25_cast_fp16")]; + tensor var_2916_cast_fp16 = mul(x = var_2886_cast_fp16, y = roti_13_cast_fp16)[name = tensor("op_2916_cast_fp16")]; + tensor var_2917_cast_fp16 = mul(x = var_2894_cast_fp16, y = rotr_13_cast_fp16)[name = tensor("op_2917_cast_fp16")]; + tensor koi_25_cast_fp16 = add(x = var_2916_cast_fp16, y = var_2917_cast_fp16)[name = tensor("koi_25_cast_fp16")]; + tensor qo_13_axis_0 = const()[name = tensor("qo_13_axis_0"), val = tensor(-1)]; + tensor qo_13_cast_fp16 = stack(axis = qo_13_axis_0, values = (qor_25_cast_fp16, qoi_25_cast_fp16))[name = tensor("qo_13_cast_fp16")]; + tensor ko_13_axis_0 = const()[name = tensor("ko_13_axis_0"), val = tensor(-1)]; + tensor ko_13_cast_fp16 = stack(axis = ko_13_axis_0, values = (kor_25_cast_fp16, koi_25_cast_fp16))[name = tensor("ko_13_cast_fp16")]; + tensor var_2946 = const()[name = tensor("op_2946"), val = tensor([1, 1, 16, 64])]; + tensor q_39_cast_fp16 = reshape(shape = var_2946, x = qo_13_cast_fp16)[name = tensor("q_39_cast_fp16")]; + tensor var_2948 = const()[name = tensor("op_2948"), val = tensor([1, 1, 16, 64])]; + tensor k_27_cast_fp16 = reshape(shape = var_2948, x = ko_13_cast_fp16)[name = tensor("k_27_cast_fp16")]; + tensor _inversed_2970_y_0_to_fp16 = const()[name = tensor("_inversed_2970_y_0_to_fp16"), val = tensor(0x1p-9)]; + tensor _inversed_2970_cast_fp16 = mul(x = ts_41_cast_fp16, y = _inversed_2970_y_0_to_fp16)[name = tensor("_inversed_2970_cast_fp16")]; + tensor var_2971_cast_fp16 = floor(x = _inversed_2970_cast_fp16)[name = tensor("op_2971_cast_fp16")]; + tensor var_2972_to_fp16 = const()[name = tensor("op_2972_to_fp16"), val = tensor(0x1p+9)]; + tensor var_2973_cast_fp16 = mul(x = var_2971_cast_fp16, y = var_2972_to_fp16)[name = tensor("op_2973_cast_fp16")]; + tensor write_indices_float_27_cast_fp16 = sub(x = ts_41_cast_fp16, y = var_2973_cast_fp16)[name = tensor("write_indices_float_27_cast_fp16")]; + tensor var_2980_dtype_0 = const()[name = tensor("op_2980_dtype_0"), val = tensor("int32")]; + tensor write_indices_13_reps_0 = const()[name = tensor("write_indices_13_reps_0"), val = tensor([1, 1, 16, 64])]; + tensor write_indices_float_27_cast_fp16_to_int32 = cast(dtype = var_2980_dtype_0, x = write_indices_float_27_cast_fp16)[name = tensor("cast_522")]; + tensor write_indices_13 = tile(reps = write_indices_13_reps_0, x = write_indices_float_27_cast_fp16_to_int32)[name = tensor("write_indices_13")]; + tensor var_2988_begin_0 = const()[name = tensor("op_2988_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_2988_end_0 = const()[name = tensor("op_2988_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor var_2988_end_mask_0 = const()[name = tensor("op_2988_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_2988_squeeze_mask_0 = const()[name = tensor("op_2988_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor cache6_to_fp16_dtype_0 = const()[name = tensor("cache6_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor cache6_to_fp16 = cast(dtype = cache6_to_fp16_dtype_0, x = cache6)[name = tensor("cast_521")]; + tensor var_2988_cast_fp16 = slice_by_index(begin = var_2988_begin_0, end = var_2988_end_0, end_mask = var_2988_end_mask_0, squeeze_mask = var_2988_squeeze_mask_0, x = cache6_to_fp16)[name = tensor("op_2988_cast_fp16")]; + tensor var_2990_axis_0 = const()[name = tensor("op_2990_axis_0"), val = tensor(1)]; + tensor var_2990_mode_0 = const()[name = tensor("op_2990_mode_0"), val = tensor("update")]; + tensor var_2990_validate_indices_0 = const()[name = tensor("op_2990_validate_indices_0"), val = tensor(false)]; + tensor var_2990_cast_fp16 = scatter_along_axis(axis = var_2990_axis_0, data = var_2988_cast_fp16, indices = write_indices_13, mode = var_2990_mode_0, updates = k_27_cast_fp16, validate_indices = var_2990_validate_indices_0)[name = tensor("op_2990_cast_fp16")]; + tensor concat_44 = const()[name = tensor("concat_44"), val = tensor([0, 0, 0, 0, 0])]; + tensor concat_45 = const()[name = tensor("concat_45"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_13_internal_tensor_assign_1_stride_0 = const()[name = tensor("new_cache_13_internal_tensor_assign_1_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_13_internal_tensor_assign_1_begin_mask_0 = const()[name = tensor("new_cache_13_internal_tensor_assign_1_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_13_internal_tensor_assign_1_end_mask_0 = const()[name = tensor("new_cache_13_internal_tensor_assign_1_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_13_internal_tensor_assign_1_squeeze_mask_0 = const()[name = tensor("new_cache_13_internal_tensor_assign_1_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_60 = const()[name = tensor("shape_60"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_12 = const()[name = tensor("reduce_prod_12"), val = tensor(1048576)]; + tensor range_1d_12_start_0 = const()[name = tensor("range_1d_12_start_0"), val = tensor(0)]; + tensor range_1d_12_step_0 = const()[name = tensor("range_1d_12_step_0"), val = tensor(1)]; + tensor range_1d_12 = range_1d(end = reduce_prod_12, start = range_1d_12_start_0, step = range_1d_12_step_0)[name = tensor("range_1d_12")]; + tensor reshape_60 = reshape(shape = shape_60, x = range_1d_12)[name = tensor("reshape_60")]; + tensor slice_by_index_12 = slice_by_index(begin = concat_44, begin_mask = new_cache_13_internal_tensor_assign_1_begin_mask_0, end = concat_45, end_mask = new_cache_13_internal_tensor_assign_1_end_mask_0, squeeze_mask = new_cache_13_internal_tensor_assign_1_squeeze_mask_0, stride = new_cache_13_internal_tensor_assign_1_stride_0, x = reshape_60)[name = tensor("slice_by_index_12")]; + tensor reshape_61_shape_0 = const()[name = tensor("reshape_61_shape_0"), val = tensor([-1])]; + tensor reshape_61 = reshape(shape = reshape_61_shape_0, x = slice_by_index_12)[name = tensor("reshape_61")]; + tensor reshape_62_shape_0 = const()[name = tensor("reshape_62_shape_0"), val = tensor([-1])]; + tensor reshape_62_cast_fp16 = reshape(shape = reshape_62_shape_0, x = var_2990_cast_fp16)[name = tensor("reshape_62_cast_fp16")]; + tensor reshape_63_shape_0 = const()[name = tensor("reshape_63_shape_0"), val = tensor([-1])]; + tensor reshape_63_cast_fp16 = reshape(shape = reshape_63_shape_0, x = cache6_to_fp16)[name = tensor("reshape_63_cast_fp16")]; + tensor scatter_12_mode_0 = const()[name = tensor("scatter_12_mode_0"), val = tensor("update")]; + tensor scatter_12_axis_0 = const()[name = tensor("scatter_12_axis_0"), val = tensor(0)]; + tensor scatter_12_validate_indices_0 = const()[name = tensor("scatter_12_validate_indices_0"), val = tensor(false)]; + tensor scatter_12_cast_fp16 = scatter(axis = scatter_12_axis_0, data = reshape_63_cast_fp16, indices = reshape_61, mode = scatter_12_mode_0, updates = reshape_62_cast_fp16, validate_indices = scatter_12_validate_indices_0)[name = tensor("scatter_12_cast_fp16")]; + tensor reshape_64_cast_fp16 = reshape(shape = shape_60, x = scatter_12_cast_fp16)[name = tensor("reshape_64_cast_fp16")]; + tensor var_2998_begin_0 = const()[name = tensor("op_2998_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor var_2998_end_0 = const()[name = tensor("op_2998_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor var_2998_end_mask_0 = const()[name = tensor("op_2998_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_2998_squeeze_mask_0 = const()[name = tensor("op_2998_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor var_2998_cast_fp16 = slice_by_index(begin = var_2998_begin_0, end = var_2998_end_0, end_mask = var_2998_end_mask_0, squeeze_mask = var_2998_squeeze_mask_0, x = reshape_64_cast_fp16)[name = tensor("op_2998_cast_fp16")]; + tensor var_3000_axis_0 = const()[name = tensor("op_3000_axis_0"), val = tensor(1)]; + tensor var_3000_mode_0 = const()[name = tensor("op_3000_mode_0"), val = tensor("update")]; + tensor var_3000_validate_indices_0 = const()[name = tensor("op_3000_validate_indices_0"), val = tensor(false)]; + tensor var_3000_cast_fp16 = scatter_along_axis(axis = var_3000_axis_0, data = var_2998_cast_fp16, indices = write_indices_13, mode = var_3000_mode_0, updates = v_13_cast_fp16, validate_indices = var_3000_validate_indices_0)[name = tensor("op_3000_cast_fp16")]; + tensor concat_46 = const()[name = tensor("concat_46"), val = tensor([1, 0, 0, 0, 0])]; + tensor concat_47 = const()[name = tensor("concat_47"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_13_internal_tensor_assign_2_stride_0 = const()[name = tensor("new_cache_13_internal_tensor_assign_2_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_13_internal_tensor_assign_2_begin_mask_0 = const()[name = tensor("new_cache_13_internal_tensor_assign_2_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_13_internal_tensor_assign_2_end_mask_0 = const()[name = tensor("new_cache_13_internal_tensor_assign_2_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_13_internal_tensor_assign_2_squeeze_mask_0 = const()[name = tensor("new_cache_13_internal_tensor_assign_2_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_61 = const()[name = tensor("shape_61"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_13 = const()[name = tensor("reduce_prod_13"), val = tensor(1048576)]; + tensor range_1d_13_start_0 = const()[name = tensor("range_1d_13_start_0"), val = tensor(0)]; + tensor range_1d_13_step_0 = const()[name = tensor("range_1d_13_step_0"), val = tensor(1)]; + tensor range_1d_13 = range_1d(end = reduce_prod_13, start = range_1d_13_start_0, step = range_1d_13_step_0)[name = tensor("range_1d_13")]; + tensor reshape_65 = reshape(shape = shape_61, x = range_1d_13)[name = tensor("reshape_65")]; + tensor slice_by_index_13 = slice_by_index(begin = concat_46, begin_mask = new_cache_13_internal_tensor_assign_2_begin_mask_0, end = concat_47, end_mask = new_cache_13_internal_tensor_assign_2_end_mask_0, squeeze_mask = new_cache_13_internal_tensor_assign_2_squeeze_mask_0, stride = new_cache_13_internal_tensor_assign_2_stride_0, x = reshape_65)[name = tensor("slice_by_index_13")]; + tensor reshape_66_shape_0 = const()[name = tensor("reshape_66_shape_0"), val = tensor([-1])]; + tensor reshape_66 = reshape(shape = reshape_66_shape_0, x = slice_by_index_13)[name = tensor("reshape_66")]; + tensor reshape_67_shape_0 = const()[name = tensor("reshape_67_shape_0"), val = tensor([-1])]; + tensor reshape_67_cast_fp16 = reshape(shape = reshape_67_shape_0, x = var_3000_cast_fp16)[name = tensor("reshape_67_cast_fp16")]; + tensor reshape_68_shape_0 = const()[name = tensor("reshape_68_shape_0"), val = tensor([-1])]; + tensor reshape_68_cast_fp16 = reshape(shape = reshape_68_shape_0, x = reshape_64_cast_fp16)[name = tensor("reshape_68_cast_fp16")]; + tensor scatter_13_mode_0 = const()[name = tensor("scatter_13_mode_0"), val = tensor("update")]; + tensor scatter_13_axis_0 = const()[name = tensor("scatter_13_axis_0"), val = tensor(0)]; + tensor scatter_13_validate_indices_0 = const()[name = tensor("scatter_13_validate_indices_0"), val = tensor(false)]; + tensor scatter_13_cast_fp16 = scatter(axis = scatter_13_axis_0, data = reshape_68_cast_fp16, indices = reshape_66, mode = scatter_13_mode_0, updates = reshape_67_cast_fp16, validate_indices = scatter_13_validate_indices_0)[name = tensor("scatter_13_cast_fp16")]; + tensor reshape_69_cast_fp16 = reshape(shape = shape_61, x = scatter_13_cast_fp16)[name = tensor("reshape_69_cast_fp16")]; + tensor reshape_69_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("reshape_69_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor keys_37_begin_0 = const()[name = tensor("keys_37_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor keys_37_end_0 = const()[name = tensor("keys_37_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor keys_37_end_mask_0 = const()[name = tensor("keys_37_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor keys_37_squeeze_mask_0 = const()[name = tensor("keys_37_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor keys_37_cast_fp16 = slice_by_index(begin = keys_37_begin_0, end = keys_37_end_0, end_mask = keys_37_end_mask_0, squeeze_mask = keys_37_squeeze_mask_0, x = reshape_69_cast_fp16)[name = tensor("keys_37_cast_fp16")]; + tensor values_37_begin_0 = const()[name = tensor("values_37_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor values_37_end_0 = const()[name = tensor("values_37_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor values_37_end_mask_0 = const()[name = tensor("values_37_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor values_37_squeeze_mask_0 = const()[name = tensor("values_37_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor values_37_cast_fp16 = slice_by_index(begin = values_37_begin_0, end = values_37_end_0, end_mask = values_37_end_mask_0, squeeze_mask = values_37_squeeze_mask_0, x = reshape_69_cast_fp16)[name = tensor("values_37_cast_fp16")]; + tensor var_3012_cast_fp16 = not_equal(x = keys_37_cast_fp16, y = keys_37_cast_fp16)[name = tensor("op_3012_cast_fp16")]; + tensor keys_39_cast_fp16 = select(a = var_504_to_fp16, b = keys_37_cast_fp16, cond = var_3012_cast_fp16)[name = tensor("keys_39_cast_fp16")]; + tensor var_3020_cast_fp16 = not_equal(x = values_37_cast_fp16, y = values_37_cast_fp16)[name = tensor("op_3020_cast_fp16")]; + tensor values_39_cast_fp16 = select(a = var_504_to_fp16, b = values_37_cast_fp16, cond = var_3020_cast_fp16)[name = tensor("values_39_cast_fp16")]; + tensor var_3044 = const()[name = tensor("op_3044"), val = tensor([0, 2, 1, 3])]; + tensor var_3057 = const()[name = tensor("op_3057"), val = tensor([1, 1, 1])]; + tensor var_3058_cast_fp16 = reshape(shape = var_3057, x = position6_to_fp16)[name = tensor("op_3058_cast_fp16")]; + tensor var_3075_to_fp16 = const()[name = tensor("op_3075_to_fp16"), val = tensor(0x1p+0)]; + tensor valid_len_13_cast_fp16 = add(x = var_3058_cast_fp16, y = var_3075_to_fp16)[name = tensor("valid_len_13_cast_fp16")]; + tensor valid_mask_13_cast_fp16 = less(x = k_positions_1_promoted_to_fp16, y = valid_len_13_cast_fp16)[name = tensor("valid_mask_13_cast_fp16")]; + tensor causal_mask_13_cast_fp16 = less_equal(x = k_positions_1_promoted_to_fp16, y = var_3058_cast_fp16)[name = tensor("causal_mask_13_cast_fp16")]; + tensor attn_mask_25 = logical_and(x = valid_mask_13_cast_fp16, y = causal_mask_13_cast_fp16)[name = tensor("attn_mask_25")]; + tensor attn_mask_27_axes_0 = const()[name = tensor("attn_mask_27_axes_0"), val = tensor([1])]; + tensor attn_mask_27 = expand_dims(axes = attn_mask_27_axes_0, x = attn_mask_25)[name = tensor("attn_mask_27")]; + tensor var_3093_transpose_x_0 = const()[name = tensor("op_3093_transpose_x_0"), val = tensor(false)]; + tensor var_3093_transpose_y_0 = const()[name = tensor("op_3093_transpose_y_0"), val = tensor(false)]; + tensor transpose_84_perm_0 = const()[name = tensor("transpose_84_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_85_perm_0 = const()[name = tensor("transpose_85_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_85 = transpose(perm = transpose_85_perm_0, x = keys_39_cast_fp16)[name = tensor("transpose_189")]; + tensor transpose_84 = transpose(perm = transpose_84_perm_0, x = q_39_cast_fp16)[name = tensor("transpose_190")]; + tensor var_3093_cast_fp16 = matmul(transpose_x = var_3093_transpose_x_0, transpose_y = var_3093_transpose_y_0, x = transpose_84, y = transpose_85)[name = tensor("op_3093_cast_fp16")]; + tensor var_3087_to_fp16 = const()[name = tensor("op_3087_to_fp16"), val = tensor([0x1p-3])]; + tensor attn_weights_37_cast_fp16 = mul(x = var_3093_cast_fp16, y = var_3087_to_fp16)[name = tensor("attn_weights_37_cast_fp16")]; + tensor var_3095 = logical_not(x = attn_mask_27)[name = tensor("op_3095")]; + tensor var_3096_to_fp16 = const()[name = tensor("op_3096_to_fp16"), val = tensor(-inf)]; + tensor attn_weights_39_cast_fp16 = select(a = var_3096_to_fp16, b = attn_weights_37_cast_fp16, cond = var_3095)[name = tensor("attn_weights_39_cast_fp16")]; + tensor var_3098 = const()[name = tensor("op_3098"), val = tensor(-1)]; + tensor attn_weights_41_cast_fp16 = softmax(axis = var_3098, x = attn_weights_39_cast_fp16)[name = tensor("attn_weights_41_cast_fp16")]; + tensor attn_output_13_transpose_x_0 = const()[name = tensor("attn_output_13_transpose_x_0"), val = tensor(false)]; + tensor attn_output_13_transpose_y_0 = const()[name = tensor("attn_output_13_transpose_y_0"), val = tensor(false)]; + tensor values_41_cast_fp16 = transpose(perm = var_3044, x = values_39_cast_fp16)[name = tensor("transpose_191")]; + tensor attn_output_13_cast_fp16 = matmul(transpose_x = attn_output_13_transpose_x_0, transpose_y = attn_output_13_transpose_y_0, x = attn_weights_41_cast_fp16, y = values_41_cast_fp16)[name = tensor("attn_output_13_cast_fp16")]; + tensor var_3106 = const()[name = tensor("op_3106"), val = tensor([0, 2, 1, 3])]; + tensor var_3109 = const()[name = tensor("op_3109"), val = tensor([1, 1, 1024])]; + tensor var_3107_cast_fp16 = transpose(perm = var_3106, x = attn_output_13_cast_fp16)[name = tensor("transpose_188")]; + tensor input_65_cast_fp16 = reshape(shape = var_3109, x = var_3107_cast_fp16)[name = tensor("input_65_cast_fp16")]; + tensor linear_26_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = attn6_out_proj_weight_cast_fp16, x = input_65_cast_fp16)[name = tensor("linear_26_cast_fp16")]; + tensor var_3115_to_fp16 = const()[name = tensor("op_3115_to_fp16"), val = tensor(0x1p+0)]; + tensor var_3116_cast_fp16 = add(x = position6_to_fp16, y = var_3115_to_fp16)[name = tensor("op_3116_cast_fp16")]; + tensor var_3116_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_3116_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor input_67_cast_fp16 = add(x = input_63_cast_fp16, y = linear_26_cast_fp16)[name = tensor("input_67_cast_fp16")]; + tensor input_69_axes_0 = const()[name = tensor("input_69_axes_0"), val = tensor([-1])]; + tensor norm6_2_weight_to_fp16 = const()[name = tensor("norm6_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(303640704)))]; + tensor norm6_2_bias_to_fp16 = const()[name = tensor("norm6_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(303642816)))]; + tensor var_3120_to_fp16 = const()[name = tensor("op_3120_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_69_cast_fp16 = layer_norm(axes = input_69_axes_0, beta = norm6_2_bias_to_fp16, epsilon = var_3120_to_fp16, gamma = norm6_2_weight_to_fp16, x = input_67_cast_fp16)[name = tensor("input_69_cast_fp16")]; + tensor linear_27_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = linear6_1_weight_cast_fp16, x = input_69_cast_fp16)[name = tensor("linear_27_cast_fp16")]; + tensor input_71_mode_0 = const()[name = tensor("input_71_mode_0"), val = tensor("EXACT")]; + tensor input_71_cast_fp16 = gelu(mode = input_71_mode_0, x = linear_27_cast_fp16)[name = tensor("input_71_cast_fp16")]; + tensor linear_28_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = linear6_2_weight_cast_fp16, x = input_71_cast_fp16)[name = tensor("linear_28_cast_fp16")]; + tensor input_73_cast_fp16 = add(x = input_67_cast_fp16, y = linear_28_cast_fp16)[name = tensor("input_73_cast_fp16")]; + tensor x_15_axes_0 = const()[name = tensor("x_15_axes_0"), val = tensor([-1])]; + tensor norm7_1_weight_to_fp16 = const()[name = tensor("norm7_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(303644928)))]; + tensor norm7_1_bias_to_fp16 = const()[name = tensor("norm7_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(303647040)))]; + tensor var_3137_to_fp16 = const()[name = tensor("op_3137_to_fp16"), val = tensor(0x1.5p-17)]; + tensor x_15_cast_fp16 = layer_norm(axes = x_15_axes_0, beta = norm7_1_bias_to_fp16, epsilon = var_3137_to_fp16, gamma = norm7_1_weight_to_fp16, x = input_73_cast_fp16)[name = tensor("x_15_cast_fp16")]; + tensor linear_29_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = attn7_in_proj_weight_cast_fp16, x = x_15_cast_fp16)[name = tensor("linear_29_cast_fp16")]; + tensor var_3173 = const()[name = tensor("op_3173"), val = tensor([1, 1, 3, 16, 64])]; + tensor qkv_15_cast_fp16 = reshape(shape = var_3173, x = linear_29_cast_fp16)[name = tensor("qkv_15_cast_fp16")]; + tensor q_43_begin_0 = const()[name = tensor("q_43_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor q_43_end_0 = const()[name = tensor("q_43_end_0"), val = tensor([1, 1, 1, 16, 64])]; + tensor q_43_end_mask_0 = const()[name = tensor("q_43_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor q_43_squeeze_mask_0 = const()[name = tensor("q_43_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor q_43_cast_fp16 = slice_by_index(begin = q_43_begin_0, end = q_43_end_0, end_mask = q_43_end_mask_0, squeeze_mask = q_43_squeeze_mask_0, x = qkv_15_cast_fp16)[name = tensor("q_43_cast_fp16")]; + tensor k_29_begin_0 = const()[name = tensor("k_29_begin_0"), val = tensor([0, 0, 1, 0, 0])]; + tensor k_29_end_0 = const()[name = tensor("k_29_end_0"), val = tensor([1, 1, 2, 16, 64])]; + tensor k_29_end_mask_0 = const()[name = tensor("k_29_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor k_29_squeeze_mask_0 = const()[name = tensor("k_29_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor k_29_cast_fp16 = slice_by_index(begin = k_29_begin_0, end = k_29_end_0, end_mask = k_29_end_mask_0, squeeze_mask = k_29_squeeze_mask_0, x = qkv_15_cast_fp16)[name = tensor("k_29_cast_fp16")]; + tensor v_15_begin_0 = const()[name = tensor("v_15_begin_0"), val = tensor([0, 0, 2, 0, 0])]; + tensor v_15_end_0 = const()[name = tensor("v_15_end_0"), val = tensor([1, 1, 3, 16, 64])]; + tensor v_15_end_mask_0 = const()[name = tensor("v_15_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor v_15_squeeze_mask_0 = const()[name = tensor("v_15_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor v_15_cast_fp16 = slice_by_index(begin = v_15_begin_0, end = v_15_end_0, end_mask = v_15_end_mask_0, squeeze_mask = v_15_squeeze_mask_0, x = qkv_15_cast_fp16)[name = tensor("v_15_cast_fp16")]; + tensor var_3277 = const()[name = tensor("op_3277"), val = tensor([1, 1, 1, 1])]; + tensor position7_to_fp16_dtype_0 = const()[name = tensor("position7_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor position7_to_fp16 = cast(dtype = position7_to_fp16_dtype_0, x = position7)[name = tensor("cast_518")]; + tensor ts_47_cast_fp16 = reshape(shape = var_3277, x = position7_to_fp16)[name = tensor("ts_47_cast_fp16")]; + tensor var_3281 = const()[name = tensor("op_3281"), val = tensor([1, 1, 16, 32, 2])]; + tensor q_complex_15_cast_fp16 = reshape(shape = var_3281, x = q_43_cast_fp16)[name = tensor("q_complex_15_cast_fp16")]; + tensor var_3285 = const()[name = tensor("op_3285"), val = tensor([1, 1, 16, 32, 2])]; + tensor k_complex_15_cast_fp16 = reshape(shape = var_3285, x = k_29_cast_fp16)[name = tensor("k_complex_15_cast_fp16")]; + tensor var_3289_begin_0 = const()[name = tensor("op_3289_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_3289_end_0 = const()[name = tensor("op_3289_end_0"), val = tensor([1, 1, 16, 32, 1])]; + tensor var_3289_end_mask_0 = const()[name = tensor("op_3289_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_3289_squeeze_mask_0 = const()[name = tensor("op_3289_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_3289_cast_fp16 = slice_by_index(begin = var_3289_begin_0, end = var_3289_end_0, end_mask = var_3289_end_mask_0, squeeze_mask = var_3289_squeeze_mask_0, x = q_complex_15_cast_fp16)[name = tensor("op_3289_cast_fp16")]; + tensor var_3297_begin_0 = const()[name = tensor("op_3297_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_3297_end_0 = const()[name = tensor("op_3297_end_0"), val = tensor([1, 1, 16, 32, 2])]; + tensor var_3297_end_mask_0 = const()[name = tensor("op_3297_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_3297_squeeze_mask_0 = const()[name = tensor("op_3297_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_3297_cast_fp16 = slice_by_index(begin = var_3297_begin_0, end = var_3297_end_0, end_mask = var_3297_end_mask_0, squeeze_mask = var_3297_squeeze_mask_0, x = q_complex_15_cast_fp16)[name = tensor("op_3297_cast_fp16")]; + tensor var_3305_begin_0 = const()[name = tensor("op_3305_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_3305_end_0 = const()[name = tensor("op_3305_end_0"), val = tensor([1, 1, 16, 32, 1])]; + tensor var_3305_end_mask_0 = const()[name = tensor("op_3305_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_3305_squeeze_mask_0 = const()[name = tensor("op_3305_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_3305_cast_fp16 = slice_by_index(begin = var_3305_begin_0, end = var_3305_end_0, end_mask = var_3305_end_mask_0, squeeze_mask = var_3305_squeeze_mask_0, x = k_complex_15_cast_fp16)[name = tensor("op_3305_cast_fp16")]; + tensor var_3313_begin_0 = const()[name = tensor("op_3313_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_3313_end_0 = const()[name = tensor("op_3313_end_0"), val = tensor([1, 1, 16, 32, 2])]; + tensor var_3313_end_mask_0 = const()[name = tensor("op_3313_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_3313_squeeze_mask_0 = const()[name = tensor("op_3313_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_3313_cast_fp16 = slice_by_index(begin = var_3313_begin_0, end = var_3313_end_0, end_mask = var_3313_end_mask_0, squeeze_mask = var_3313_squeeze_mask_0, x = k_complex_15_cast_fp16)[name = tensor("op_3313_cast_fp16")]; + tensor freqs_15_to_fp16 = const()[name = tensor("freqs_15_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(303649152)))]; + tensor var_3319_cast_fp16 = mul(x = freqs_15_to_fp16, y = ts_47_cast_fp16)[name = tensor("op_3319_cast_fp16")]; + tensor rotr_15_cast_fp16 = cos(x = var_3319_cast_fp16)[name = tensor("rotr_15_cast_fp16")]; + tensor roti_15_cast_fp16 = sin(x = var_3319_cast_fp16)[name = tensor("roti_15_cast_fp16")]; + tensor var_3323_cast_fp16 = mul(x = var_3289_cast_fp16, y = rotr_15_cast_fp16)[name = tensor("op_3323_cast_fp16")]; + tensor var_3324_cast_fp16 = mul(x = var_3297_cast_fp16, y = roti_15_cast_fp16)[name = tensor("op_3324_cast_fp16")]; + tensor qor_29_cast_fp16 = sub(x = var_3323_cast_fp16, y = var_3324_cast_fp16)[name = tensor("qor_29_cast_fp16")]; + tensor var_3327_cast_fp16 = mul(x = var_3289_cast_fp16, y = roti_15_cast_fp16)[name = tensor("op_3327_cast_fp16")]; + tensor var_3328_cast_fp16 = mul(x = var_3297_cast_fp16, y = rotr_15_cast_fp16)[name = tensor("op_3328_cast_fp16")]; + tensor qoi_29_cast_fp16 = add(x = var_3327_cast_fp16, y = var_3328_cast_fp16)[name = tensor("qoi_29_cast_fp16")]; + tensor var_3331_cast_fp16 = mul(x = var_3305_cast_fp16, y = rotr_15_cast_fp16)[name = tensor("op_3331_cast_fp16")]; + tensor var_3332_cast_fp16 = mul(x = var_3313_cast_fp16, y = roti_15_cast_fp16)[name = tensor("op_3332_cast_fp16")]; + tensor kor_29_cast_fp16 = sub(x = var_3331_cast_fp16, y = var_3332_cast_fp16)[name = tensor("kor_29_cast_fp16")]; + tensor var_3335_cast_fp16 = mul(x = var_3305_cast_fp16, y = roti_15_cast_fp16)[name = tensor("op_3335_cast_fp16")]; + tensor var_3336_cast_fp16 = mul(x = var_3313_cast_fp16, y = rotr_15_cast_fp16)[name = tensor("op_3336_cast_fp16")]; + tensor koi_29_cast_fp16 = add(x = var_3335_cast_fp16, y = var_3336_cast_fp16)[name = tensor("koi_29_cast_fp16")]; + tensor qo_15_axis_0 = const()[name = tensor("qo_15_axis_0"), val = tensor(-1)]; + tensor qo_15_cast_fp16 = stack(axis = qo_15_axis_0, values = (qor_29_cast_fp16, qoi_29_cast_fp16))[name = tensor("qo_15_cast_fp16")]; + tensor ko_15_axis_0 = const()[name = tensor("ko_15_axis_0"), val = tensor(-1)]; + tensor ko_15_cast_fp16 = stack(axis = ko_15_axis_0, values = (kor_29_cast_fp16, koi_29_cast_fp16))[name = tensor("ko_15_cast_fp16")]; + tensor var_3365 = const()[name = tensor("op_3365"), val = tensor([1, 1, 16, 64])]; + tensor q_45_cast_fp16 = reshape(shape = var_3365, x = qo_15_cast_fp16)[name = tensor("q_45_cast_fp16")]; + tensor var_3367 = const()[name = tensor("op_3367"), val = tensor([1, 1, 16, 64])]; + tensor k_31_cast_fp16 = reshape(shape = var_3367, x = ko_15_cast_fp16)[name = tensor("k_31_cast_fp16")]; + tensor _inversed_3389_y_0_to_fp16 = const()[name = tensor("_inversed_3389_y_0_to_fp16"), val = tensor(0x1p-9)]; + tensor _inversed_3389_cast_fp16 = mul(x = ts_47_cast_fp16, y = _inversed_3389_y_0_to_fp16)[name = tensor("_inversed_3389_cast_fp16")]; + tensor var_3390_cast_fp16 = floor(x = _inversed_3389_cast_fp16)[name = tensor("op_3390_cast_fp16")]; + tensor var_3391_to_fp16 = const()[name = tensor("op_3391_to_fp16"), val = tensor(0x1p+9)]; + tensor var_3392_cast_fp16 = mul(x = var_3390_cast_fp16, y = var_3391_to_fp16)[name = tensor("op_3392_cast_fp16")]; + tensor write_indices_float_31_cast_fp16 = sub(x = ts_47_cast_fp16, y = var_3392_cast_fp16)[name = tensor("write_indices_float_31_cast_fp16")]; + tensor var_3399_dtype_0 = const()[name = tensor("op_3399_dtype_0"), val = tensor("int32")]; + tensor write_indices_15_reps_0 = const()[name = tensor("write_indices_15_reps_0"), val = tensor([1, 1, 16, 64])]; + tensor write_indices_float_31_cast_fp16_to_int32 = cast(dtype = var_3399_dtype_0, x = write_indices_float_31_cast_fp16)[name = tensor("cast_517")]; + tensor write_indices_15 = tile(reps = write_indices_15_reps_0, x = write_indices_float_31_cast_fp16_to_int32)[name = tensor("write_indices_15")]; + tensor var_3407_begin_0 = const()[name = tensor("op_3407_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_3407_end_0 = const()[name = tensor("op_3407_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor var_3407_end_mask_0 = const()[name = tensor("op_3407_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_3407_squeeze_mask_0 = const()[name = tensor("op_3407_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor cache7_to_fp16_dtype_0 = const()[name = tensor("cache7_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor cache7_to_fp16 = cast(dtype = cache7_to_fp16_dtype_0, x = cache7)[name = tensor("cast_516")]; + tensor var_3407_cast_fp16 = slice_by_index(begin = var_3407_begin_0, end = var_3407_end_0, end_mask = var_3407_end_mask_0, squeeze_mask = var_3407_squeeze_mask_0, x = cache7_to_fp16)[name = tensor("op_3407_cast_fp16")]; + tensor var_3409_axis_0 = const()[name = tensor("op_3409_axis_0"), val = tensor(1)]; + tensor var_3409_mode_0 = const()[name = tensor("op_3409_mode_0"), val = tensor("update")]; + tensor var_3409_validate_indices_0 = const()[name = tensor("op_3409_validate_indices_0"), val = tensor(false)]; + tensor var_3409_cast_fp16 = scatter_along_axis(axis = var_3409_axis_0, data = var_3407_cast_fp16, indices = write_indices_15, mode = var_3409_mode_0, updates = k_31_cast_fp16, validate_indices = var_3409_validate_indices_0)[name = tensor("op_3409_cast_fp16")]; + tensor concat_51 = const()[name = tensor("concat_51"), val = tensor([0, 0, 0, 0, 0])]; + tensor concat_52 = const()[name = tensor("concat_52"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_15_internal_tensor_assign_1_stride_0 = const()[name = tensor("new_cache_15_internal_tensor_assign_1_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_15_internal_tensor_assign_1_begin_mask_0 = const()[name = tensor("new_cache_15_internal_tensor_assign_1_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_15_internal_tensor_assign_1_end_mask_0 = const()[name = tensor("new_cache_15_internal_tensor_assign_1_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_15_internal_tensor_assign_1_squeeze_mask_0 = const()[name = tensor("new_cache_15_internal_tensor_assign_1_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_62 = const()[name = tensor("shape_62"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_14 = const()[name = tensor("reduce_prod_14"), val = tensor(1048576)]; + tensor range_1d_14_start_0 = const()[name = tensor("range_1d_14_start_0"), val = tensor(0)]; + tensor range_1d_14_step_0 = const()[name = tensor("range_1d_14_step_0"), val = tensor(1)]; + tensor range_1d_14 = range_1d(end = reduce_prod_14, start = range_1d_14_start_0, step = range_1d_14_step_0)[name = tensor("range_1d_14")]; + tensor reshape_70 = reshape(shape = shape_62, x = range_1d_14)[name = tensor("reshape_70")]; + tensor slice_by_index_14 = slice_by_index(begin = concat_51, begin_mask = new_cache_15_internal_tensor_assign_1_begin_mask_0, end = concat_52, end_mask = new_cache_15_internal_tensor_assign_1_end_mask_0, squeeze_mask = new_cache_15_internal_tensor_assign_1_squeeze_mask_0, stride = new_cache_15_internal_tensor_assign_1_stride_0, x = reshape_70)[name = tensor("slice_by_index_14")]; + tensor reshape_71_shape_0 = const()[name = tensor("reshape_71_shape_0"), val = tensor([-1])]; + tensor reshape_71 = reshape(shape = reshape_71_shape_0, x = slice_by_index_14)[name = tensor("reshape_71")]; + tensor reshape_72_shape_0 = const()[name = tensor("reshape_72_shape_0"), val = tensor([-1])]; + tensor reshape_72_cast_fp16 = reshape(shape = reshape_72_shape_0, x = var_3409_cast_fp16)[name = tensor("reshape_72_cast_fp16")]; + tensor reshape_73_shape_0 = const()[name = tensor("reshape_73_shape_0"), val = tensor([-1])]; + tensor reshape_73_cast_fp16 = reshape(shape = reshape_73_shape_0, x = cache7_to_fp16)[name = tensor("reshape_73_cast_fp16")]; + tensor scatter_14_mode_0 = const()[name = tensor("scatter_14_mode_0"), val = tensor("update")]; + tensor scatter_14_axis_0 = const()[name = tensor("scatter_14_axis_0"), val = tensor(0)]; + tensor scatter_14_validate_indices_0 = const()[name = tensor("scatter_14_validate_indices_0"), val = tensor(false)]; + tensor scatter_14_cast_fp16 = scatter(axis = scatter_14_axis_0, data = reshape_73_cast_fp16, indices = reshape_71, mode = scatter_14_mode_0, updates = reshape_72_cast_fp16, validate_indices = scatter_14_validate_indices_0)[name = tensor("scatter_14_cast_fp16")]; + tensor reshape_74_cast_fp16 = reshape(shape = shape_62, x = scatter_14_cast_fp16)[name = tensor("reshape_74_cast_fp16")]; + tensor var_3417_begin_0 = const()[name = tensor("op_3417_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor var_3417_end_0 = const()[name = tensor("op_3417_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor var_3417_end_mask_0 = const()[name = tensor("op_3417_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_3417_squeeze_mask_0 = const()[name = tensor("op_3417_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor var_3417_cast_fp16 = slice_by_index(begin = var_3417_begin_0, end = var_3417_end_0, end_mask = var_3417_end_mask_0, squeeze_mask = var_3417_squeeze_mask_0, x = reshape_74_cast_fp16)[name = tensor("op_3417_cast_fp16")]; + tensor var_3419_axis_0 = const()[name = tensor("op_3419_axis_0"), val = tensor(1)]; + tensor var_3419_mode_0 = const()[name = tensor("op_3419_mode_0"), val = tensor("update")]; + tensor var_3419_validate_indices_0 = const()[name = tensor("op_3419_validate_indices_0"), val = tensor(false)]; + tensor var_3419_cast_fp16 = scatter_along_axis(axis = var_3419_axis_0, data = var_3417_cast_fp16, indices = write_indices_15, mode = var_3419_mode_0, updates = v_15_cast_fp16, validate_indices = var_3419_validate_indices_0)[name = tensor("op_3419_cast_fp16")]; + tensor concat_53 = const()[name = tensor("concat_53"), val = tensor([1, 0, 0, 0, 0])]; + tensor concat_54 = const()[name = tensor("concat_54"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_15_internal_tensor_assign_2_stride_0 = const()[name = tensor("new_cache_15_internal_tensor_assign_2_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_15_internal_tensor_assign_2_begin_mask_0 = const()[name = tensor("new_cache_15_internal_tensor_assign_2_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_15_internal_tensor_assign_2_end_mask_0 = const()[name = tensor("new_cache_15_internal_tensor_assign_2_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_15_internal_tensor_assign_2_squeeze_mask_0 = const()[name = tensor("new_cache_15_internal_tensor_assign_2_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_63 = const()[name = tensor("shape_63"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_15 = const()[name = tensor("reduce_prod_15"), val = tensor(1048576)]; + tensor range_1d_15_start_0 = const()[name = tensor("range_1d_15_start_0"), val = tensor(0)]; + tensor range_1d_15_step_0 = const()[name = tensor("range_1d_15_step_0"), val = tensor(1)]; + tensor range_1d_15 = range_1d(end = reduce_prod_15, start = range_1d_15_start_0, step = range_1d_15_step_0)[name = tensor("range_1d_15")]; + tensor reshape_75 = reshape(shape = shape_63, x = range_1d_15)[name = tensor("reshape_75")]; + tensor slice_by_index_15 = slice_by_index(begin = concat_53, begin_mask = new_cache_15_internal_tensor_assign_2_begin_mask_0, end = concat_54, end_mask = new_cache_15_internal_tensor_assign_2_end_mask_0, squeeze_mask = new_cache_15_internal_tensor_assign_2_squeeze_mask_0, stride = new_cache_15_internal_tensor_assign_2_stride_0, x = reshape_75)[name = tensor("slice_by_index_15")]; + tensor reshape_76_shape_0 = const()[name = tensor("reshape_76_shape_0"), val = tensor([-1])]; + tensor reshape_76 = reshape(shape = reshape_76_shape_0, x = slice_by_index_15)[name = tensor("reshape_76")]; + tensor reshape_77_shape_0 = const()[name = tensor("reshape_77_shape_0"), val = tensor([-1])]; + tensor reshape_77_cast_fp16 = reshape(shape = reshape_77_shape_0, x = var_3419_cast_fp16)[name = tensor("reshape_77_cast_fp16")]; + tensor reshape_78_shape_0 = const()[name = tensor("reshape_78_shape_0"), val = tensor([-1])]; + tensor reshape_78_cast_fp16 = reshape(shape = reshape_78_shape_0, x = reshape_74_cast_fp16)[name = tensor("reshape_78_cast_fp16")]; + tensor scatter_15_mode_0 = const()[name = tensor("scatter_15_mode_0"), val = tensor("update")]; + tensor scatter_15_axis_0 = const()[name = tensor("scatter_15_axis_0"), val = tensor(0)]; + tensor scatter_15_validate_indices_0 = const()[name = tensor("scatter_15_validate_indices_0"), val = tensor(false)]; + tensor scatter_15_cast_fp16 = scatter(axis = scatter_15_axis_0, data = reshape_78_cast_fp16, indices = reshape_76, mode = scatter_15_mode_0, updates = reshape_77_cast_fp16, validate_indices = scatter_15_validate_indices_0)[name = tensor("scatter_15_cast_fp16")]; + tensor reshape_79_cast_fp16 = reshape(shape = shape_63, x = scatter_15_cast_fp16)[name = tensor("reshape_79_cast_fp16")]; + tensor reshape_79_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("reshape_79_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor keys_43_begin_0 = const()[name = tensor("keys_43_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor keys_43_end_0 = const()[name = tensor("keys_43_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor keys_43_end_mask_0 = const()[name = tensor("keys_43_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor keys_43_squeeze_mask_0 = const()[name = tensor("keys_43_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor keys_43_cast_fp16 = slice_by_index(begin = keys_43_begin_0, end = keys_43_end_0, end_mask = keys_43_end_mask_0, squeeze_mask = keys_43_squeeze_mask_0, x = reshape_79_cast_fp16)[name = tensor("keys_43_cast_fp16")]; + tensor values_43_begin_0 = const()[name = tensor("values_43_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor values_43_end_0 = const()[name = tensor("values_43_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor values_43_end_mask_0 = const()[name = tensor("values_43_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor values_43_squeeze_mask_0 = const()[name = tensor("values_43_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor values_43_cast_fp16 = slice_by_index(begin = values_43_begin_0, end = values_43_end_0, end_mask = values_43_end_mask_0, squeeze_mask = values_43_squeeze_mask_0, x = reshape_79_cast_fp16)[name = tensor("values_43_cast_fp16")]; + tensor var_3431_cast_fp16 = not_equal(x = keys_43_cast_fp16, y = keys_43_cast_fp16)[name = tensor("op_3431_cast_fp16")]; + tensor keys_45_cast_fp16 = select(a = var_504_to_fp16, b = keys_43_cast_fp16, cond = var_3431_cast_fp16)[name = tensor("keys_45_cast_fp16")]; + tensor var_3439_cast_fp16 = not_equal(x = values_43_cast_fp16, y = values_43_cast_fp16)[name = tensor("op_3439_cast_fp16")]; + tensor values_45_cast_fp16 = select(a = var_504_to_fp16, b = values_43_cast_fp16, cond = var_3439_cast_fp16)[name = tensor("values_45_cast_fp16")]; + tensor var_3463 = const()[name = tensor("op_3463"), val = tensor([0, 2, 1, 3])]; + tensor var_3476 = const()[name = tensor("op_3476"), val = tensor([1, 1, 1])]; + tensor var_3477_cast_fp16 = reshape(shape = var_3476, x = position7_to_fp16)[name = tensor("op_3477_cast_fp16")]; + tensor var_3494_to_fp16 = const()[name = tensor("op_3494_to_fp16"), val = tensor(0x1p+0)]; + tensor valid_len_15_cast_fp16 = add(x = var_3477_cast_fp16, y = var_3494_to_fp16)[name = tensor("valid_len_15_cast_fp16")]; + tensor valid_mask_15_cast_fp16 = less(x = k_positions_1_promoted_to_fp16, y = valid_len_15_cast_fp16)[name = tensor("valid_mask_15_cast_fp16")]; + tensor causal_mask_15_cast_fp16 = less_equal(x = k_positions_1_promoted_to_fp16, y = var_3477_cast_fp16)[name = tensor("causal_mask_15_cast_fp16")]; + tensor attn_mask_29 = logical_and(x = valid_mask_15_cast_fp16, y = causal_mask_15_cast_fp16)[name = tensor("attn_mask_29")]; + tensor attn_mask_31_axes_0 = const()[name = tensor("attn_mask_31_axes_0"), val = tensor([1])]; + tensor attn_mask_31 = expand_dims(axes = attn_mask_31_axes_0, x = attn_mask_29)[name = tensor("attn_mask_31")]; + tensor var_3512_transpose_x_0 = const()[name = tensor("op_3512_transpose_x_0"), val = tensor(false)]; + tensor var_3512_transpose_y_0 = const()[name = tensor("op_3512_transpose_y_0"), val = tensor(false)]; + tensor transpose_86_perm_0 = const()[name = tensor("transpose_86_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_87_perm_0 = const()[name = tensor("transpose_87_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_87 = transpose(perm = transpose_87_perm_0, x = keys_45_cast_fp16)[name = tensor("transpose_185")]; + tensor transpose_86 = transpose(perm = transpose_86_perm_0, x = q_45_cast_fp16)[name = tensor("transpose_186")]; + tensor var_3512_cast_fp16 = matmul(transpose_x = var_3512_transpose_x_0, transpose_y = var_3512_transpose_y_0, x = transpose_86, y = transpose_87)[name = tensor("op_3512_cast_fp16")]; + tensor var_3506_to_fp16 = const()[name = tensor("op_3506_to_fp16"), val = tensor([0x1p-3])]; + tensor attn_weights_43_cast_fp16 = mul(x = var_3512_cast_fp16, y = var_3506_to_fp16)[name = tensor("attn_weights_43_cast_fp16")]; + tensor var_3514 = logical_not(x = attn_mask_31)[name = tensor("op_3514")]; + tensor var_3515_to_fp16 = const()[name = tensor("op_3515_to_fp16"), val = tensor(-inf)]; + tensor attn_weights_45_cast_fp16 = select(a = var_3515_to_fp16, b = attn_weights_43_cast_fp16, cond = var_3514)[name = tensor("attn_weights_45_cast_fp16")]; + tensor var_3517 = const()[name = tensor("op_3517"), val = tensor(-1)]; + tensor attn_weights_47_cast_fp16 = softmax(axis = var_3517, x = attn_weights_45_cast_fp16)[name = tensor("attn_weights_47_cast_fp16")]; + tensor attn_output_15_transpose_x_0 = const()[name = tensor("attn_output_15_transpose_x_0"), val = tensor(false)]; + tensor attn_output_15_transpose_y_0 = const()[name = tensor("attn_output_15_transpose_y_0"), val = tensor(false)]; + tensor values_47_cast_fp16 = transpose(perm = var_3463, x = values_45_cast_fp16)[name = tensor("transpose_187")]; + tensor attn_output_15_cast_fp16 = matmul(transpose_x = attn_output_15_transpose_x_0, transpose_y = attn_output_15_transpose_y_0, x = attn_weights_47_cast_fp16, y = values_47_cast_fp16)[name = tensor("attn_output_15_cast_fp16")]; + tensor var_3525 = const()[name = tensor("op_3525"), val = tensor([0, 2, 1, 3])]; + tensor var_3528 = const()[name = tensor("op_3528"), val = tensor([1, 1, 1024])]; + tensor var_3526_cast_fp16 = transpose(perm = var_3525, x = attn_output_15_cast_fp16)[name = tensor("transpose_184")]; + tensor input_75_cast_fp16 = reshape(shape = var_3528, x = var_3526_cast_fp16)[name = tensor("input_75_cast_fp16")]; + tensor linear_30_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = attn7_out_proj_weight_cast_fp16, x = input_75_cast_fp16)[name = tensor("linear_30_cast_fp16")]; + tensor var_3534_to_fp16 = const()[name = tensor("op_3534_to_fp16"), val = tensor(0x1p+0)]; + tensor var_3535_cast_fp16 = add(x = position7_to_fp16, y = var_3534_to_fp16)[name = tensor("op_3535_cast_fp16")]; + tensor var_3535_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_3535_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor input_77_cast_fp16 = add(x = input_73_cast_fp16, y = linear_30_cast_fp16)[name = tensor("input_77_cast_fp16")]; + tensor input_79_axes_0 = const()[name = tensor("input_79_axes_0"), val = tensor([-1])]; + tensor norm7_2_weight_to_fp16 = const()[name = tensor("norm7_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(303649280)))]; + tensor norm7_2_bias_to_fp16 = const()[name = tensor("norm7_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(303651392)))]; + tensor var_3539_to_fp16 = const()[name = tensor("op_3539_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_79_cast_fp16 = layer_norm(axes = input_79_axes_0, beta = norm7_2_bias_to_fp16, epsilon = var_3539_to_fp16, gamma = norm7_2_weight_to_fp16, x = input_77_cast_fp16)[name = tensor("input_79_cast_fp16")]; + tensor linear_31_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = linear7_1_weight_cast_fp16, x = input_79_cast_fp16)[name = tensor("linear_31_cast_fp16")]; + tensor input_81_mode_0 = const()[name = tensor("input_81_mode_0"), val = tensor("EXACT")]; + tensor input_81_cast_fp16 = gelu(mode = input_81_mode_0, x = linear_31_cast_fp16)[name = tensor("input_81_cast_fp16")]; + tensor linear_32_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = linear7_2_weight_cast_fp16, x = input_81_cast_fp16)[name = tensor("linear_32_cast_fp16")]; + tensor input_83_cast_fp16 = add(x = input_77_cast_fp16, y = linear_32_cast_fp16)[name = tensor("input_83_cast_fp16")]; + tensor x_17_axes_0 = const()[name = tensor("x_17_axes_0"), val = tensor([-1])]; + tensor norm8_1_weight_to_fp16 = const()[name = tensor("norm8_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(303653504)))]; + tensor norm8_1_bias_to_fp16 = const()[name = tensor("norm8_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(303655616)))]; + tensor var_3556_to_fp16 = const()[name = tensor("op_3556_to_fp16"), val = tensor(0x1.5p-17)]; + tensor x_17_cast_fp16 = layer_norm(axes = x_17_axes_0, beta = norm8_1_bias_to_fp16, epsilon = var_3556_to_fp16, gamma = norm8_1_weight_to_fp16, x = input_83_cast_fp16)[name = tensor("x_17_cast_fp16")]; + tensor linear_33_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = attn8_in_proj_weight_cast_fp16, x = x_17_cast_fp16)[name = tensor("linear_33_cast_fp16")]; + tensor var_3592 = const()[name = tensor("op_3592"), val = tensor([1, 1, 3, 16, 64])]; + tensor qkv_17_cast_fp16 = reshape(shape = var_3592, x = linear_33_cast_fp16)[name = tensor("qkv_17_cast_fp16")]; + tensor q_49_begin_0 = const()[name = tensor("q_49_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor q_49_end_0 = const()[name = tensor("q_49_end_0"), val = tensor([1, 1, 1, 16, 64])]; + tensor q_49_end_mask_0 = const()[name = tensor("q_49_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor q_49_squeeze_mask_0 = const()[name = tensor("q_49_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor q_49_cast_fp16 = slice_by_index(begin = q_49_begin_0, end = q_49_end_0, end_mask = q_49_end_mask_0, squeeze_mask = q_49_squeeze_mask_0, x = qkv_17_cast_fp16)[name = tensor("q_49_cast_fp16")]; + tensor k_33_begin_0 = const()[name = tensor("k_33_begin_0"), val = tensor([0, 0, 1, 0, 0])]; + tensor k_33_end_0 = const()[name = tensor("k_33_end_0"), val = tensor([1, 1, 2, 16, 64])]; + tensor k_33_end_mask_0 = const()[name = tensor("k_33_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor k_33_squeeze_mask_0 = const()[name = tensor("k_33_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor k_33_cast_fp16 = slice_by_index(begin = k_33_begin_0, end = k_33_end_0, end_mask = k_33_end_mask_0, squeeze_mask = k_33_squeeze_mask_0, x = qkv_17_cast_fp16)[name = tensor("k_33_cast_fp16")]; + tensor v_17_begin_0 = const()[name = tensor("v_17_begin_0"), val = tensor([0, 0, 2, 0, 0])]; + tensor v_17_end_0 = const()[name = tensor("v_17_end_0"), val = tensor([1, 1, 3, 16, 64])]; + tensor v_17_end_mask_0 = const()[name = tensor("v_17_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor v_17_squeeze_mask_0 = const()[name = tensor("v_17_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor v_17_cast_fp16 = slice_by_index(begin = v_17_begin_0, end = v_17_end_0, end_mask = v_17_end_mask_0, squeeze_mask = v_17_squeeze_mask_0, x = qkv_17_cast_fp16)[name = tensor("v_17_cast_fp16")]; + tensor var_3696 = const()[name = tensor("op_3696"), val = tensor([1, 1, 1, 1])]; + tensor position8_to_fp16_dtype_0 = const()[name = tensor("position8_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor position8_to_fp16 = cast(dtype = position8_to_fp16_dtype_0, x = position8)[name = tensor("cast_513")]; + tensor ts_53_cast_fp16 = reshape(shape = var_3696, x = position8_to_fp16)[name = tensor("ts_53_cast_fp16")]; + tensor var_3700 = const()[name = tensor("op_3700"), val = tensor([1, 1, 16, 32, 2])]; + tensor q_complex_17_cast_fp16 = reshape(shape = var_3700, x = q_49_cast_fp16)[name = tensor("q_complex_17_cast_fp16")]; + tensor var_3704 = const()[name = tensor("op_3704"), val = tensor([1, 1, 16, 32, 2])]; + tensor k_complex_17_cast_fp16 = reshape(shape = var_3704, x = k_33_cast_fp16)[name = tensor("k_complex_17_cast_fp16")]; + tensor var_3708_begin_0 = const()[name = tensor("op_3708_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_3708_end_0 = const()[name = tensor("op_3708_end_0"), val = tensor([1, 1, 16, 32, 1])]; + tensor var_3708_end_mask_0 = const()[name = tensor("op_3708_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_3708_squeeze_mask_0 = const()[name = tensor("op_3708_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_3708_cast_fp16 = slice_by_index(begin = var_3708_begin_0, end = var_3708_end_0, end_mask = var_3708_end_mask_0, squeeze_mask = var_3708_squeeze_mask_0, x = q_complex_17_cast_fp16)[name = tensor("op_3708_cast_fp16")]; + tensor var_3716_begin_0 = const()[name = tensor("op_3716_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_3716_end_0 = const()[name = tensor("op_3716_end_0"), val = tensor([1, 1, 16, 32, 2])]; + tensor var_3716_end_mask_0 = const()[name = tensor("op_3716_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_3716_squeeze_mask_0 = const()[name = tensor("op_3716_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_3716_cast_fp16 = slice_by_index(begin = var_3716_begin_0, end = var_3716_end_0, end_mask = var_3716_end_mask_0, squeeze_mask = var_3716_squeeze_mask_0, x = q_complex_17_cast_fp16)[name = tensor("op_3716_cast_fp16")]; + tensor var_3724_begin_0 = const()[name = tensor("op_3724_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_3724_end_0 = const()[name = tensor("op_3724_end_0"), val = tensor([1, 1, 16, 32, 1])]; + tensor var_3724_end_mask_0 = const()[name = tensor("op_3724_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_3724_squeeze_mask_0 = const()[name = tensor("op_3724_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_3724_cast_fp16 = slice_by_index(begin = var_3724_begin_0, end = var_3724_end_0, end_mask = var_3724_end_mask_0, squeeze_mask = var_3724_squeeze_mask_0, x = k_complex_17_cast_fp16)[name = tensor("op_3724_cast_fp16")]; + tensor var_3732_begin_0 = const()[name = tensor("op_3732_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_3732_end_0 = const()[name = tensor("op_3732_end_0"), val = tensor([1, 1, 16, 32, 2])]; + tensor var_3732_end_mask_0 = const()[name = tensor("op_3732_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_3732_squeeze_mask_0 = const()[name = tensor("op_3732_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_3732_cast_fp16 = slice_by_index(begin = var_3732_begin_0, end = var_3732_end_0, end_mask = var_3732_end_mask_0, squeeze_mask = var_3732_squeeze_mask_0, x = k_complex_17_cast_fp16)[name = tensor("op_3732_cast_fp16")]; + tensor freqs_17_to_fp16 = const()[name = tensor("freqs_17_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(303657728)))]; + tensor var_3738_cast_fp16 = mul(x = freqs_17_to_fp16, y = ts_53_cast_fp16)[name = tensor("op_3738_cast_fp16")]; + tensor rotr_17_cast_fp16 = cos(x = var_3738_cast_fp16)[name = tensor("rotr_17_cast_fp16")]; + tensor roti_17_cast_fp16 = sin(x = var_3738_cast_fp16)[name = tensor("roti_17_cast_fp16")]; + tensor var_3742_cast_fp16 = mul(x = var_3708_cast_fp16, y = rotr_17_cast_fp16)[name = tensor("op_3742_cast_fp16")]; + tensor var_3743_cast_fp16 = mul(x = var_3716_cast_fp16, y = roti_17_cast_fp16)[name = tensor("op_3743_cast_fp16")]; + tensor qor_33_cast_fp16 = sub(x = var_3742_cast_fp16, y = var_3743_cast_fp16)[name = tensor("qor_33_cast_fp16")]; + tensor var_3746_cast_fp16 = mul(x = var_3708_cast_fp16, y = roti_17_cast_fp16)[name = tensor("op_3746_cast_fp16")]; + tensor var_3747_cast_fp16 = mul(x = var_3716_cast_fp16, y = rotr_17_cast_fp16)[name = tensor("op_3747_cast_fp16")]; + tensor qoi_33_cast_fp16 = add(x = var_3746_cast_fp16, y = var_3747_cast_fp16)[name = tensor("qoi_33_cast_fp16")]; + tensor var_3750_cast_fp16 = mul(x = var_3724_cast_fp16, y = rotr_17_cast_fp16)[name = tensor("op_3750_cast_fp16")]; + tensor var_3751_cast_fp16 = mul(x = var_3732_cast_fp16, y = roti_17_cast_fp16)[name = tensor("op_3751_cast_fp16")]; + tensor kor_33_cast_fp16 = sub(x = var_3750_cast_fp16, y = var_3751_cast_fp16)[name = tensor("kor_33_cast_fp16")]; + tensor var_3754_cast_fp16 = mul(x = var_3724_cast_fp16, y = roti_17_cast_fp16)[name = tensor("op_3754_cast_fp16")]; + tensor var_3755_cast_fp16 = mul(x = var_3732_cast_fp16, y = rotr_17_cast_fp16)[name = tensor("op_3755_cast_fp16")]; + tensor koi_33_cast_fp16 = add(x = var_3754_cast_fp16, y = var_3755_cast_fp16)[name = tensor("koi_33_cast_fp16")]; + tensor qo_17_axis_0 = const()[name = tensor("qo_17_axis_0"), val = tensor(-1)]; + tensor qo_17_cast_fp16 = stack(axis = qo_17_axis_0, values = (qor_33_cast_fp16, qoi_33_cast_fp16))[name = tensor("qo_17_cast_fp16")]; + tensor ko_17_axis_0 = const()[name = tensor("ko_17_axis_0"), val = tensor(-1)]; + tensor ko_17_cast_fp16 = stack(axis = ko_17_axis_0, values = (kor_33_cast_fp16, koi_33_cast_fp16))[name = tensor("ko_17_cast_fp16")]; + tensor var_3784 = const()[name = tensor("op_3784"), val = tensor([1, 1, 16, 64])]; + tensor q_51_cast_fp16 = reshape(shape = var_3784, x = qo_17_cast_fp16)[name = tensor("q_51_cast_fp16")]; + tensor var_3786 = const()[name = tensor("op_3786"), val = tensor([1, 1, 16, 64])]; + tensor k_35_cast_fp16 = reshape(shape = var_3786, x = ko_17_cast_fp16)[name = tensor("k_35_cast_fp16")]; + tensor _inversed_3808_y_0_to_fp16 = const()[name = tensor("_inversed_3808_y_0_to_fp16"), val = tensor(0x1p-9)]; + tensor _inversed_3808_cast_fp16 = mul(x = ts_53_cast_fp16, y = _inversed_3808_y_0_to_fp16)[name = tensor("_inversed_3808_cast_fp16")]; + tensor var_3809_cast_fp16 = floor(x = _inversed_3808_cast_fp16)[name = tensor("op_3809_cast_fp16")]; + tensor var_3810_to_fp16 = const()[name = tensor("op_3810_to_fp16"), val = tensor(0x1p+9)]; + tensor var_3811_cast_fp16 = mul(x = var_3809_cast_fp16, y = var_3810_to_fp16)[name = tensor("op_3811_cast_fp16")]; + tensor write_indices_float_35_cast_fp16 = sub(x = ts_53_cast_fp16, y = var_3811_cast_fp16)[name = tensor("write_indices_float_35_cast_fp16")]; + tensor var_3818_dtype_0 = const()[name = tensor("op_3818_dtype_0"), val = tensor("int32")]; + tensor write_indices_17_reps_0 = const()[name = tensor("write_indices_17_reps_0"), val = tensor([1, 1, 16, 64])]; + tensor write_indices_float_35_cast_fp16_to_int32 = cast(dtype = var_3818_dtype_0, x = write_indices_float_35_cast_fp16)[name = tensor("cast_512")]; + tensor write_indices_17 = tile(reps = write_indices_17_reps_0, x = write_indices_float_35_cast_fp16_to_int32)[name = tensor("write_indices_17")]; + tensor var_3826_begin_0 = const()[name = tensor("op_3826_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_3826_end_0 = const()[name = tensor("op_3826_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor var_3826_end_mask_0 = const()[name = tensor("op_3826_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_3826_squeeze_mask_0 = const()[name = tensor("op_3826_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor cache8_to_fp16_dtype_0 = const()[name = tensor("cache8_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor cache8_to_fp16 = cast(dtype = cache8_to_fp16_dtype_0, x = cache8)[name = tensor("cast_511")]; + tensor var_3826_cast_fp16 = slice_by_index(begin = var_3826_begin_0, end = var_3826_end_0, end_mask = var_3826_end_mask_0, squeeze_mask = var_3826_squeeze_mask_0, x = cache8_to_fp16)[name = tensor("op_3826_cast_fp16")]; + tensor var_3828_axis_0 = const()[name = tensor("op_3828_axis_0"), val = tensor(1)]; + tensor var_3828_mode_0 = const()[name = tensor("op_3828_mode_0"), val = tensor("update")]; + tensor var_3828_validate_indices_0 = const()[name = tensor("op_3828_validate_indices_0"), val = tensor(false)]; + tensor var_3828_cast_fp16 = scatter_along_axis(axis = var_3828_axis_0, data = var_3826_cast_fp16, indices = write_indices_17, mode = var_3828_mode_0, updates = k_35_cast_fp16, validate_indices = var_3828_validate_indices_0)[name = tensor("op_3828_cast_fp16")]; + tensor concat_58 = const()[name = tensor("concat_58"), val = tensor([0, 0, 0, 0, 0])]; + tensor concat_59 = const()[name = tensor("concat_59"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_17_internal_tensor_assign_1_stride_0 = const()[name = tensor("new_cache_17_internal_tensor_assign_1_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_17_internal_tensor_assign_1_begin_mask_0 = const()[name = tensor("new_cache_17_internal_tensor_assign_1_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_17_internal_tensor_assign_1_end_mask_0 = const()[name = tensor("new_cache_17_internal_tensor_assign_1_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_17_internal_tensor_assign_1_squeeze_mask_0 = const()[name = tensor("new_cache_17_internal_tensor_assign_1_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_64 = const()[name = tensor("shape_64"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_16 = const()[name = tensor("reduce_prod_16"), val = tensor(1048576)]; + tensor range_1d_16_start_0 = const()[name = tensor("range_1d_16_start_0"), val = tensor(0)]; + tensor range_1d_16_step_0 = const()[name = tensor("range_1d_16_step_0"), val = tensor(1)]; + tensor range_1d_16 = range_1d(end = reduce_prod_16, start = range_1d_16_start_0, step = range_1d_16_step_0)[name = tensor("range_1d_16")]; + tensor reshape_80 = reshape(shape = shape_64, x = range_1d_16)[name = tensor("reshape_80")]; + tensor slice_by_index_16 = slice_by_index(begin = concat_58, begin_mask = new_cache_17_internal_tensor_assign_1_begin_mask_0, end = concat_59, end_mask = new_cache_17_internal_tensor_assign_1_end_mask_0, squeeze_mask = new_cache_17_internal_tensor_assign_1_squeeze_mask_0, stride = new_cache_17_internal_tensor_assign_1_stride_0, x = reshape_80)[name = tensor("slice_by_index_16")]; + tensor reshape_81_shape_0 = const()[name = tensor("reshape_81_shape_0"), val = tensor([-1])]; + tensor reshape_81 = reshape(shape = reshape_81_shape_0, x = slice_by_index_16)[name = tensor("reshape_81")]; + tensor reshape_82_shape_0 = const()[name = tensor("reshape_82_shape_0"), val = tensor([-1])]; + tensor reshape_82_cast_fp16 = reshape(shape = reshape_82_shape_0, x = var_3828_cast_fp16)[name = tensor("reshape_82_cast_fp16")]; + tensor reshape_83_shape_0 = const()[name = tensor("reshape_83_shape_0"), val = tensor([-1])]; + tensor reshape_83_cast_fp16 = reshape(shape = reshape_83_shape_0, x = cache8_to_fp16)[name = tensor("reshape_83_cast_fp16")]; + tensor scatter_16_mode_0 = const()[name = tensor("scatter_16_mode_0"), val = tensor("update")]; + tensor scatter_16_axis_0 = const()[name = tensor("scatter_16_axis_0"), val = tensor(0)]; + tensor scatter_16_validate_indices_0 = const()[name = tensor("scatter_16_validate_indices_0"), val = tensor(false)]; + tensor scatter_16_cast_fp16 = scatter(axis = scatter_16_axis_0, data = reshape_83_cast_fp16, indices = reshape_81, mode = scatter_16_mode_0, updates = reshape_82_cast_fp16, validate_indices = scatter_16_validate_indices_0)[name = tensor("scatter_16_cast_fp16")]; + tensor reshape_84_cast_fp16 = reshape(shape = shape_64, x = scatter_16_cast_fp16)[name = tensor("reshape_84_cast_fp16")]; + tensor var_3836_begin_0 = const()[name = tensor("op_3836_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor var_3836_end_0 = const()[name = tensor("op_3836_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor var_3836_end_mask_0 = const()[name = tensor("op_3836_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_3836_squeeze_mask_0 = const()[name = tensor("op_3836_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor var_3836_cast_fp16 = slice_by_index(begin = var_3836_begin_0, end = var_3836_end_0, end_mask = var_3836_end_mask_0, squeeze_mask = var_3836_squeeze_mask_0, x = reshape_84_cast_fp16)[name = tensor("op_3836_cast_fp16")]; + tensor var_3838_axis_0 = const()[name = tensor("op_3838_axis_0"), val = tensor(1)]; + tensor var_3838_mode_0 = const()[name = tensor("op_3838_mode_0"), val = tensor("update")]; + tensor var_3838_validate_indices_0 = const()[name = tensor("op_3838_validate_indices_0"), val = tensor(false)]; + tensor var_3838_cast_fp16 = scatter_along_axis(axis = var_3838_axis_0, data = var_3836_cast_fp16, indices = write_indices_17, mode = var_3838_mode_0, updates = v_17_cast_fp16, validate_indices = var_3838_validate_indices_0)[name = tensor("op_3838_cast_fp16")]; + tensor concat_60 = const()[name = tensor("concat_60"), val = tensor([1, 0, 0, 0, 0])]; + tensor concat_61 = const()[name = tensor("concat_61"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_17_internal_tensor_assign_2_stride_0 = const()[name = tensor("new_cache_17_internal_tensor_assign_2_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_17_internal_tensor_assign_2_begin_mask_0 = const()[name = tensor("new_cache_17_internal_tensor_assign_2_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_17_internal_tensor_assign_2_end_mask_0 = const()[name = tensor("new_cache_17_internal_tensor_assign_2_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_17_internal_tensor_assign_2_squeeze_mask_0 = const()[name = tensor("new_cache_17_internal_tensor_assign_2_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_65 = const()[name = tensor("shape_65"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_17 = const()[name = tensor("reduce_prod_17"), val = tensor(1048576)]; + tensor range_1d_17_start_0 = const()[name = tensor("range_1d_17_start_0"), val = tensor(0)]; + tensor range_1d_17_step_0 = const()[name = tensor("range_1d_17_step_0"), val = tensor(1)]; + tensor range_1d_17 = range_1d(end = reduce_prod_17, start = range_1d_17_start_0, step = range_1d_17_step_0)[name = tensor("range_1d_17")]; + tensor reshape_85 = reshape(shape = shape_65, x = range_1d_17)[name = tensor("reshape_85")]; + tensor slice_by_index_17 = slice_by_index(begin = concat_60, begin_mask = new_cache_17_internal_tensor_assign_2_begin_mask_0, end = concat_61, end_mask = new_cache_17_internal_tensor_assign_2_end_mask_0, squeeze_mask = new_cache_17_internal_tensor_assign_2_squeeze_mask_0, stride = new_cache_17_internal_tensor_assign_2_stride_0, x = reshape_85)[name = tensor("slice_by_index_17")]; + tensor reshape_86_shape_0 = const()[name = tensor("reshape_86_shape_0"), val = tensor([-1])]; + tensor reshape_86 = reshape(shape = reshape_86_shape_0, x = slice_by_index_17)[name = tensor("reshape_86")]; + tensor reshape_87_shape_0 = const()[name = tensor("reshape_87_shape_0"), val = tensor([-1])]; + tensor reshape_87_cast_fp16 = reshape(shape = reshape_87_shape_0, x = var_3838_cast_fp16)[name = tensor("reshape_87_cast_fp16")]; + tensor reshape_88_shape_0 = const()[name = tensor("reshape_88_shape_0"), val = tensor([-1])]; + tensor reshape_88_cast_fp16 = reshape(shape = reshape_88_shape_0, x = reshape_84_cast_fp16)[name = tensor("reshape_88_cast_fp16")]; + tensor scatter_17_mode_0 = const()[name = tensor("scatter_17_mode_0"), val = tensor("update")]; + tensor scatter_17_axis_0 = const()[name = tensor("scatter_17_axis_0"), val = tensor(0)]; + tensor scatter_17_validate_indices_0 = const()[name = tensor("scatter_17_validate_indices_0"), val = tensor(false)]; + tensor scatter_17_cast_fp16 = scatter(axis = scatter_17_axis_0, data = reshape_88_cast_fp16, indices = reshape_86, mode = scatter_17_mode_0, updates = reshape_87_cast_fp16, validate_indices = scatter_17_validate_indices_0)[name = tensor("scatter_17_cast_fp16")]; + tensor reshape_89_cast_fp16 = reshape(shape = shape_65, x = scatter_17_cast_fp16)[name = tensor("reshape_89_cast_fp16")]; + tensor reshape_89_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("reshape_89_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor keys_49_begin_0 = const()[name = tensor("keys_49_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor keys_49_end_0 = const()[name = tensor("keys_49_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor keys_49_end_mask_0 = const()[name = tensor("keys_49_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor keys_49_squeeze_mask_0 = const()[name = tensor("keys_49_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor keys_49_cast_fp16 = slice_by_index(begin = keys_49_begin_0, end = keys_49_end_0, end_mask = keys_49_end_mask_0, squeeze_mask = keys_49_squeeze_mask_0, x = reshape_89_cast_fp16)[name = tensor("keys_49_cast_fp16")]; + tensor values_49_begin_0 = const()[name = tensor("values_49_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor values_49_end_0 = const()[name = tensor("values_49_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor values_49_end_mask_0 = const()[name = tensor("values_49_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor values_49_squeeze_mask_0 = const()[name = tensor("values_49_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor values_49_cast_fp16 = slice_by_index(begin = values_49_begin_0, end = values_49_end_0, end_mask = values_49_end_mask_0, squeeze_mask = values_49_squeeze_mask_0, x = reshape_89_cast_fp16)[name = tensor("values_49_cast_fp16")]; + tensor var_3850_cast_fp16 = not_equal(x = keys_49_cast_fp16, y = keys_49_cast_fp16)[name = tensor("op_3850_cast_fp16")]; + tensor keys_51_cast_fp16 = select(a = var_504_to_fp16, b = keys_49_cast_fp16, cond = var_3850_cast_fp16)[name = tensor("keys_51_cast_fp16")]; + tensor var_3858_cast_fp16 = not_equal(x = values_49_cast_fp16, y = values_49_cast_fp16)[name = tensor("op_3858_cast_fp16")]; + tensor values_51_cast_fp16 = select(a = var_504_to_fp16, b = values_49_cast_fp16, cond = var_3858_cast_fp16)[name = tensor("values_51_cast_fp16")]; + tensor var_3882 = const()[name = tensor("op_3882"), val = tensor([0, 2, 1, 3])]; + tensor var_3895 = const()[name = tensor("op_3895"), val = tensor([1, 1, 1])]; + tensor var_3896_cast_fp16 = reshape(shape = var_3895, x = position8_to_fp16)[name = tensor("op_3896_cast_fp16")]; + tensor var_3913_to_fp16 = const()[name = tensor("op_3913_to_fp16"), val = tensor(0x1p+0)]; + tensor valid_len_17_cast_fp16 = add(x = var_3896_cast_fp16, y = var_3913_to_fp16)[name = tensor("valid_len_17_cast_fp16")]; + tensor valid_mask_17_cast_fp16 = less(x = k_positions_1_promoted_to_fp16, y = valid_len_17_cast_fp16)[name = tensor("valid_mask_17_cast_fp16")]; + tensor causal_mask_17_cast_fp16 = less_equal(x = k_positions_1_promoted_to_fp16, y = var_3896_cast_fp16)[name = tensor("causal_mask_17_cast_fp16")]; + tensor attn_mask_33 = logical_and(x = valid_mask_17_cast_fp16, y = causal_mask_17_cast_fp16)[name = tensor("attn_mask_33")]; + tensor attn_mask_35_axes_0 = const()[name = tensor("attn_mask_35_axes_0"), val = tensor([1])]; + tensor attn_mask_35 = expand_dims(axes = attn_mask_35_axes_0, x = attn_mask_33)[name = tensor("attn_mask_35")]; + tensor var_3931_transpose_x_0 = const()[name = tensor("op_3931_transpose_x_0"), val = tensor(false)]; + tensor var_3931_transpose_y_0 = const()[name = tensor("op_3931_transpose_y_0"), val = tensor(false)]; + tensor transpose_88_perm_0 = const()[name = tensor("transpose_88_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_89_perm_0 = const()[name = tensor("transpose_89_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_89 = transpose(perm = transpose_89_perm_0, x = keys_51_cast_fp16)[name = tensor("transpose_181")]; + tensor transpose_88 = transpose(perm = transpose_88_perm_0, x = q_51_cast_fp16)[name = tensor("transpose_182")]; + tensor var_3931_cast_fp16 = matmul(transpose_x = var_3931_transpose_x_0, transpose_y = var_3931_transpose_y_0, x = transpose_88, y = transpose_89)[name = tensor("op_3931_cast_fp16")]; + tensor var_3925_to_fp16 = const()[name = tensor("op_3925_to_fp16"), val = tensor([0x1p-3])]; + tensor attn_weights_49_cast_fp16 = mul(x = var_3931_cast_fp16, y = var_3925_to_fp16)[name = tensor("attn_weights_49_cast_fp16")]; + tensor var_3933 = logical_not(x = attn_mask_35)[name = tensor("op_3933")]; + tensor var_3934_to_fp16 = const()[name = tensor("op_3934_to_fp16"), val = tensor(-inf)]; + tensor attn_weights_51_cast_fp16 = select(a = var_3934_to_fp16, b = attn_weights_49_cast_fp16, cond = var_3933)[name = tensor("attn_weights_51_cast_fp16")]; + tensor var_3936 = const()[name = tensor("op_3936"), val = tensor(-1)]; + tensor attn_weights_53_cast_fp16 = softmax(axis = var_3936, x = attn_weights_51_cast_fp16)[name = tensor("attn_weights_53_cast_fp16")]; + tensor attn_output_17_transpose_x_0 = const()[name = tensor("attn_output_17_transpose_x_0"), val = tensor(false)]; + tensor attn_output_17_transpose_y_0 = const()[name = tensor("attn_output_17_transpose_y_0"), val = tensor(false)]; + tensor values_53_cast_fp16 = transpose(perm = var_3882, x = values_51_cast_fp16)[name = tensor("transpose_183")]; + tensor attn_output_17_cast_fp16 = matmul(transpose_x = attn_output_17_transpose_x_0, transpose_y = attn_output_17_transpose_y_0, x = attn_weights_53_cast_fp16, y = values_53_cast_fp16)[name = tensor("attn_output_17_cast_fp16")]; + tensor var_3944 = const()[name = tensor("op_3944"), val = tensor([0, 2, 1, 3])]; + tensor var_3947 = const()[name = tensor("op_3947"), val = tensor([1, 1, 1024])]; + tensor var_3945_cast_fp16 = transpose(perm = var_3944, x = attn_output_17_cast_fp16)[name = tensor("transpose_180")]; + tensor input_85_cast_fp16 = reshape(shape = var_3947, x = var_3945_cast_fp16)[name = tensor("input_85_cast_fp16")]; + tensor linear_34_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = attn8_out_proj_weight_cast_fp16, x = input_85_cast_fp16)[name = tensor("linear_34_cast_fp16")]; + tensor var_3953_to_fp16 = const()[name = tensor("op_3953_to_fp16"), val = tensor(0x1p+0)]; + tensor var_3954_cast_fp16 = add(x = position8_to_fp16, y = var_3953_to_fp16)[name = tensor("op_3954_cast_fp16")]; + tensor var_3954_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_3954_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor input_87_cast_fp16 = add(x = input_83_cast_fp16, y = linear_34_cast_fp16)[name = tensor("input_87_cast_fp16")]; + tensor input_89_axes_0 = const()[name = tensor("input_89_axes_0"), val = tensor([-1])]; + tensor norm8_2_weight_to_fp16 = const()[name = tensor("norm8_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(303657856)))]; + tensor norm8_2_bias_to_fp16 = const()[name = tensor("norm8_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(303659968)))]; + tensor var_3958_to_fp16 = const()[name = tensor("op_3958_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_89_cast_fp16 = layer_norm(axes = input_89_axes_0, beta = norm8_2_bias_to_fp16, epsilon = var_3958_to_fp16, gamma = norm8_2_weight_to_fp16, x = input_87_cast_fp16)[name = tensor("input_89_cast_fp16")]; + tensor linear_35_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = linear8_1_weight_cast_fp16, x = input_89_cast_fp16)[name = tensor("linear_35_cast_fp16")]; + tensor input_91_mode_0 = const()[name = tensor("input_91_mode_0"), val = tensor("EXACT")]; + tensor input_91_cast_fp16 = gelu(mode = input_91_mode_0, x = linear_35_cast_fp16)[name = tensor("input_91_cast_fp16")]; + tensor linear_36_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = linear8_2_weight_cast_fp16, x = input_91_cast_fp16)[name = tensor("linear_36_cast_fp16")]; + tensor input_93_cast_fp16 = add(x = input_87_cast_fp16, y = linear_36_cast_fp16)[name = tensor("input_93_cast_fp16")]; + tensor x_19_axes_0 = const()[name = tensor("x_19_axes_0"), val = tensor([-1])]; + tensor norm9_1_weight_to_fp16 = const()[name = tensor("norm9_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(303662080)))]; + tensor norm9_1_bias_to_fp16 = const()[name = tensor("norm9_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(303664192)))]; + tensor var_3975_to_fp16 = const()[name = tensor("op_3975_to_fp16"), val = tensor(0x1.5p-17)]; + tensor x_19_cast_fp16 = layer_norm(axes = x_19_axes_0, beta = norm9_1_bias_to_fp16, epsilon = var_3975_to_fp16, gamma = norm9_1_weight_to_fp16, x = input_93_cast_fp16)[name = tensor("x_19_cast_fp16")]; + tensor linear_37_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = attn9_in_proj_weight_cast_fp16, x = x_19_cast_fp16)[name = tensor("linear_37_cast_fp16")]; + tensor var_4011 = const()[name = tensor("op_4011"), val = tensor([1, 1, 3, 16, 64])]; + tensor qkv_19_cast_fp16 = reshape(shape = var_4011, x = linear_37_cast_fp16)[name = tensor("qkv_19_cast_fp16")]; + tensor q_55_begin_0 = const()[name = tensor("q_55_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor q_55_end_0 = const()[name = tensor("q_55_end_0"), val = tensor([1, 1, 1, 16, 64])]; + tensor q_55_end_mask_0 = const()[name = tensor("q_55_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor q_55_squeeze_mask_0 = const()[name = tensor("q_55_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor q_55_cast_fp16 = slice_by_index(begin = q_55_begin_0, end = q_55_end_0, end_mask = q_55_end_mask_0, squeeze_mask = q_55_squeeze_mask_0, x = qkv_19_cast_fp16)[name = tensor("q_55_cast_fp16")]; + tensor k_37_begin_0 = const()[name = tensor("k_37_begin_0"), val = tensor([0, 0, 1, 0, 0])]; + tensor k_37_end_0 = const()[name = tensor("k_37_end_0"), val = tensor([1, 1, 2, 16, 64])]; + tensor k_37_end_mask_0 = const()[name = tensor("k_37_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor k_37_squeeze_mask_0 = const()[name = tensor("k_37_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor k_37_cast_fp16 = slice_by_index(begin = k_37_begin_0, end = k_37_end_0, end_mask = k_37_end_mask_0, squeeze_mask = k_37_squeeze_mask_0, x = qkv_19_cast_fp16)[name = tensor("k_37_cast_fp16")]; + tensor v_19_begin_0 = const()[name = tensor("v_19_begin_0"), val = tensor([0, 0, 2, 0, 0])]; + tensor v_19_end_0 = const()[name = tensor("v_19_end_0"), val = tensor([1, 1, 3, 16, 64])]; + tensor v_19_end_mask_0 = const()[name = tensor("v_19_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor v_19_squeeze_mask_0 = const()[name = tensor("v_19_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor v_19_cast_fp16 = slice_by_index(begin = v_19_begin_0, end = v_19_end_0, end_mask = v_19_end_mask_0, squeeze_mask = v_19_squeeze_mask_0, x = qkv_19_cast_fp16)[name = tensor("v_19_cast_fp16")]; + tensor var_4115 = const()[name = tensor("op_4115"), val = tensor([1, 1, 1, 1])]; + tensor position9_to_fp16_dtype_0 = const()[name = tensor("position9_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor position9_to_fp16 = cast(dtype = position9_to_fp16_dtype_0, x = position9)[name = tensor("cast_508")]; + tensor ts_59_cast_fp16 = reshape(shape = var_4115, x = position9_to_fp16)[name = tensor("ts_59_cast_fp16")]; + tensor var_4119 = const()[name = tensor("op_4119"), val = tensor([1, 1, 16, 32, 2])]; + tensor q_complex_19_cast_fp16 = reshape(shape = var_4119, x = q_55_cast_fp16)[name = tensor("q_complex_19_cast_fp16")]; + tensor var_4123 = const()[name = tensor("op_4123"), val = tensor([1, 1, 16, 32, 2])]; + tensor k_complex_19_cast_fp16 = reshape(shape = var_4123, x = k_37_cast_fp16)[name = tensor("k_complex_19_cast_fp16")]; + tensor var_4127_begin_0 = const()[name = tensor("op_4127_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_4127_end_0 = const()[name = tensor("op_4127_end_0"), val = tensor([1, 1, 16, 32, 1])]; + tensor var_4127_end_mask_0 = const()[name = tensor("op_4127_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_4127_squeeze_mask_0 = const()[name = tensor("op_4127_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_4127_cast_fp16 = slice_by_index(begin = var_4127_begin_0, end = var_4127_end_0, end_mask = var_4127_end_mask_0, squeeze_mask = var_4127_squeeze_mask_0, x = q_complex_19_cast_fp16)[name = tensor("op_4127_cast_fp16")]; + tensor var_4135_begin_0 = const()[name = tensor("op_4135_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_4135_end_0 = const()[name = tensor("op_4135_end_0"), val = tensor([1, 1, 16, 32, 2])]; + tensor var_4135_end_mask_0 = const()[name = tensor("op_4135_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_4135_squeeze_mask_0 = const()[name = tensor("op_4135_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_4135_cast_fp16 = slice_by_index(begin = var_4135_begin_0, end = var_4135_end_0, end_mask = var_4135_end_mask_0, squeeze_mask = var_4135_squeeze_mask_0, x = q_complex_19_cast_fp16)[name = tensor("op_4135_cast_fp16")]; + tensor var_4143_begin_0 = const()[name = tensor("op_4143_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_4143_end_0 = const()[name = tensor("op_4143_end_0"), val = tensor([1, 1, 16, 32, 1])]; + tensor var_4143_end_mask_0 = const()[name = tensor("op_4143_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_4143_squeeze_mask_0 = const()[name = tensor("op_4143_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_4143_cast_fp16 = slice_by_index(begin = var_4143_begin_0, end = var_4143_end_0, end_mask = var_4143_end_mask_0, squeeze_mask = var_4143_squeeze_mask_0, x = k_complex_19_cast_fp16)[name = tensor("op_4143_cast_fp16")]; + tensor var_4151_begin_0 = const()[name = tensor("op_4151_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_4151_end_0 = const()[name = tensor("op_4151_end_0"), val = tensor([1, 1, 16, 32, 2])]; + tensor var_4151_end_mask_0 = const()[name = tensor("op_4151_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_4151_squeeze_mask_0 = const()[name = tensor("op_4151_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_4151_cast_fp16 = slice_by_index(begin = var_4151_begin_0, end = var_4151_end_0, end_mask = var_4151_end_mask_0, squeeze_mask = var_4151_squeeze_mask_0, x = k_complex_19_cast_fp16)[name = tensor("op_4151_cast_fp16")]; + tensor freqs_19_to_fp16 = const()[name = tensor("freqs_19_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(303666304)))]; + tensor var_4157_cast_fp16 = mul(x = freqs_19_to_fp16, y = ts_59_cast_fp16)[name = tensor("op_4157_cast_fp16")]; + tensor rotr_19_cast_fp16 = cos(x = var_4157_cast_fp16)[name = tensor("rotr_19_cast_fp16")]; + tensor roti_19_cast_fp16 = sin(x = var_4157_cast_fp16)[name = tensor("roti_19_cast_fp16")]; + tensor var_4161_cast_fp16 = mul(x = var_4127_cast_fp16, y = rotr_19_cast_fp16)[name = tensor("op_4161_cast_fp16")]; + tensor var_4162_cast_fp16 = mul(x = var_4135_cast_fp16, y = roti_19_cast_fp16)[name = tensor("op_4162_cast_fp16")]; + tensor qor_37_cast_fp16 = sub(x = var_4161_cast_fp16, y = var_4162_cast_fp16)[name = tensor("qor_37_cast_fp16")]; + tensor var_4165_cast_fp16 = mul(x = var_4127_cast_fp16, y = roti_19_cast_fp16)[name = tensor("op_4165_cast_fp16")]; + tensor var_4166_cast_fp16 = mul(x = var_4135_cast_fp16, y = rotr_19_cast_fp16)[name = tensor("op_4166_cast_fp16")]; + tensor qoi_37_cast_fp16 = add(x = var_4165_cast_fp16, y = var_4166_cast_fp16)[name = tensor("qoi_37_cast_fp16")]; + tensor var_4169_cast_fp16 = mul(x = var_4143_cast_fp16, y = rotr_19_cast_fp16)[name = tensor("op_4169_cast_fp16")]; + tensor var_4170_cast_fp16 = mul(x = var_4151_cast_fp16, y = roti_19_cast_fp16)[name = tensor("op_4170_cast_fp16")]; + tensor kor_37_cast_fp16 = sub(x = var_4169_cast_fp16, y = var_4170_cast_fp16)[name = tensor("kor_37_cast_fp16")]; + tensor var_4173_cast_fp16 = mul(x = var_4143_cast_fp16, y = roti_19_cast_fp16)[name = tensor("op_4173_cast_fp16")]; + tensor var_4174_cast_fp16 = mul(x = var_4151_cast_fp16, y = rotr_19_cast_fp16)[name = tensor("op_4174_cast_fp16")]; + tensor koi_37_cast_fp16 = add(x = var_4173_cast_fp16, y = var_4174_cast_fp16)[name = tensor("koi_37_cast_fp16")]; + tensor qo_19_axis_0 = const()[name = tensor("qo_19_axis_0"), val = tensor(-1)]; + tensor qo_19_cast_fp16 = stack(axis = qo_19_axis_0, values = (qor_37_cast_fp16, qoi_37_cast_fp16))[name = tensor("qo_19_cast_fp16")]; + tensor ko_19_axis_0 = const()[name = tensor("ko_19_axis_0"), val = tensor(-1)]; + tensor ko_19_cast_fp16 = stack(axis = ko_19_axis_0, values = (kor_37_cast_fp16, koi_37_cast_fp16))[name = tensor("ko_19_cast_fp16")]; + tensor var_4203 = const()[name = tensor("op_4203"), val = tensor([1, 1, 16, 64])]; + tensor q_57_cast_fp16 = reshape(shape = var_4203, x = qo_19_cast_fp16)[name = tensor("q_57_cast_fp16")]; + tensor var_4205 = const()[name = tensor("op_4205"), val = tensor([1, 1, 16, 64])]; + tensor k_39_cast_fp16 = reshape(shape = var_4205, x = ko_19_cast_fp16)[name = tensor("k_39_cast_fp16")]; + tensor _inversed_4227_y_0_to_fp16 = const()[name = tensor("_inversed_4227_y_0_to_fp16"), val = tensor(0x1p-9)]; + tensor _inversed_4227_cast_fp16 = mul(x = ts_59_cast_fp16, y = _inversed_4227_y_0_to_fp16)[name = tensor("_inversed_4227_cast_fp16")]; + tensor var_4228_cast_fp16 = floor(x = _inversed_4227_cast_fp16)[name = tensor("op_4228_cast_fp16")]; + tensor var_4229_to_fp16 = const()[name = tensor("op_4229_to_fp16"), val = tensor(0x1p+9)]; + tensor var_4230_cast_fp16 = mul(x = var_4228_cast_fp16, y = var_4229_to_fp16)[name = tensor("op_4230_cast_fp16")]; + tensor write_indices_float_39_cast_fp16 = sub(x = ts_59_cast_fp16, y = var_4230_cast_fp16)[name = tensor("write_indices_float_39_cast_fp16")]; + tensor var_4237_dtype_0 = const()[name = tensor("op_4237_dtype_0"), val = tensor("int32")]; + tensor write_indices_19_reps_0 = const()[name = tensor("write_indices_19_reps_0"), val = tensor([1, 1, 16, 64])]; + tensor write_indices_float_39_cast_fp16_to_int32 = cast(dtype = var_4237_dtype_0, x = write_indices_float_39_cast_fp16)[name = tensor("cast_507")]; + tensor write_indices_19 = tile(reps = write_indices_19_reps_0, x = write_indices_float_39_cast_fp16_to_int32)[name = tensor("write_indices_19")]; + tensor var_4245_begin_0 = const()[name = tensor("op_4245_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_4245_end_0 = const()[name = tensor("op_4245_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor var_4245_end_mask_0 = const()[name = tensor("op_4245_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_4245_squeeze_mask_0 = const()[name = tensor("op_4245_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor cache9_to_fp16_dtype_0 = const()[name = tensor("cache9_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor cache9_to_fp16 = cast(dtype = cache9_to_fp16_dtype_0, x = cache9)[name = tensor("cast_506")]; + tensor var_4245_cast_fp16 = slice_by_index(begin = var_4245_begin_0, end = var_4245_end_0, end_mask = var_4245_end_mask_0, squeeze_mask = var_4245_squeeze_mask_0, x = cache9_to_fp16)[name = tensor("op_4245_cast_fp16")]; + tensor var_4247_axis_0 = const()[name = tensor("op_4247_axis_0"), val = tensor(1)]; + tensor var_4247_mode_0 = const()[name = tensor("op_4247_mode_0"), val = tensor("update")]; + tensor var_4247_validate_indices_0 = const()[name = tensor("op_4247_validate_indices_0"), val = tensor(false)]; + tensor var_4247_cast_fp16 = scatter_along_axis(axis = var_4247_axis_0, data = var_4245_cast_fp16, indices = write_indices_19, mode = var_4247_mode_0, updates = k_39_cast_fp16, validate_indices = var_4247_validate_indices_0)[name = tensor("op_4247_cast_fp16")]; + tensor concat_65 = const()[name = tensor("concat_65"), val = tensor([0, 0, 0, 0, 0])]; + tensor concat_66 = const()[name = tensor("concat_66"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_19_internal_tensor_assign_1_stride_0 = const()[name = tensor("new_cache_19_internal_tensor_assign_1_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_19_internal_tensor_assign_1_begin_mask_0 = const()[name = tensor("new_cache_19_internal_tensor_assign_1_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_19_internal_tensor_assign_1_end_mask_0 = const()[name = tensor("new_cache_19_internal_tensor_assign_1_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_19_internal_tensor_assign_1_squeeze_mask_0 = const()[name = tensor("new_cache_19_internal_tensor_assign_1_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_66 = const()[name = tensor("shape_66"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_18 = const()[name = tensor("reduce_prod_18"), val = tensor(1048576)]; + tensor range_1d_18_start_0 = const()[name = tensor("range_1d_18_start_0"), val = tensor(0)]; + tensor range_1d_18_step_0 = const()[name = tensor("range_1d_18_step_0"), val = tensor(1)]; + tensor range_1d_18 = range_1d(end = reduce_prod_18, start = range_1d_18_start_0, step = range_1d_18_step_0)[name = tensor("range_1d_18")]; + tensor reshape_90 = reshape(shape = shape_66, x = range_1d_18)[name = tensor("reshape_90")]; + tensor slice_by_index_18 = slice_by_index(begin = concat_65, begin_mask = new_cache_19_internal_tensor_assign_1_begin_mask_0, end = concat_66, end_mask = new_cache_19_internal_tensor_assign_1_end_mask_0, squeeze_mask = new_cache_19_internal_tensor_assign_1_squeeze_mask_0, stride = new_cache_19_internal_tensor_assign_1_stride_0, x = reshape_90)[name = tensor("slice_by_index_18")]; + tensor reshape_91_shape_0 = const()[name = tensor("reshape_91_shape_0"), val = tensor([-1])]; + tensor reshape_91 = reshape(shape = reshape_91_shape_0, x = slice_by_index_18)[name = tensor("reshape_91")]; + tensor reshape_92_shape_0 = const()[name = tensor("reshape_92_shape_0"), val = tensor([-1])]; + tensor reshape_92_cast_fp16 = reshape(shape = reshape_92_shape_0, x = var_4247_cast_fp16)[name = tensor("reshape_92_cast_fp16")]; + tensor reshape_93_shape_0 = const()[name = tensor("reshape_93_shape_0"), val = tensor([-1])]; + tensor reshape_93_cast_fp16 = reshape(shape = reshape_93_shape_0, x = cache9_to_fp16)[name = tensor("reshape_93_cast_fp16")]; + tensor scatter_18_mode_0 = const()[name = tensor("scatter_18_mode_0"), val = tensor("update")]; + tensor scatter_18_axis_0 = const()[name = tensor("scatter_18_axis_0"), val = tensor(0)]; + tensor scatter_18_validate_indices_0 = const()[name = tensor("scatter_18_validate_indices_0"), val = tensor(false)]; + tensor scatter_18_cast_fp16 = scatter(axis = scatter_18_axis_0, data = reshape_93_cast_fp16, indices = reshape_91, mode = scatter_18_mode_0, updates = reshape_92_cast_fp16, validate_indices = scatter_18_validate_indices_0)[name = tensor("scatter_18_cast_fp16")]; + tensor reshape_94_cast_fp16 = reshape(shape = shape_66, x = scatter_18_cast_fp16)[name = tensor("reshape_94_cast_fp16")]; + tensor var_4255_begin_0 = const()[name = tensor("op_4255_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor var_4255_end_0 = const()[name = tensor("op_4255_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor var_4255_end_mask_0 = const()[name = tensor("op_4255_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_4255_squeeze_mask_0 = const()[name = tensor("op_4255_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor var_4255_cast_fp16 = slice_by_index(begin = var_4255_begin_0, end = var_4255_end_0, end_mask = var_4255_end_mask_0, squeeze_mask = var_4255_squeeze_mask_0, x = reshape_94_cast_fp16)[name = tensor("op_4255_cast_fp16")]; + tensor var_4257_axis_0 = const()[name = tensor("op_4257_axis_0"), val = tensor(1)]; + tensor var_4257_mode_0 = const()[name = tensor("op_4257_mode_0"), val = tensor("update")]; + tensor var_4257_validate_indices_0 = const()[name = tensor("op_4257_validate_indices_0"), val = tensor(false)]; + tensor var_4257_cast_fp16 = scatter_along_axis(axis = var_4257_axis_0, data = var_4255_cast_fp16, indices = write_indices_19, mode = var_4257_mode_0, updates = v_19_cast_fp16, validate_indices = var_4257_validate_indices_0)[name = tensor("op_4257_cast_fp16")]; + tensor concat_67 = const()[name = tensor("concat_67"), val = tensor([1, 0, 0, 0, 0])]; + tensor concat_68 = const()[name = tensor("concat_68"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_19_internal_tensor_assign_2_stride_0 = const()[name = tensor("new_cache_19_internal_tensor_assign_2_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_19_internal_tensor_assign_2_begin_mask_0 = const()[name = tensor("new_cache_19_internal_tensor_assign_2_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_19_internal_tensor_assign_2_end_mask_0 = const()[name = tensor("new_cache_19_internal_tensor_assign_2_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_19_internal_tensor_assign_2_squeeze_mask_0 = const()[name = tensor("new_cache_19_internal_tensor_assign_2_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_67 = const()[name = tensor("shape_67"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_19 = const()[name = tensor("reduce_prod_19"), val = tensor(1048576)]; + tensor range_1d_19_start_0 = const()[name = tensor("range_1d_19_start_0"), val = tensor(0)]; + tensor range_1d_19_step_0 = const()[name = tensor("range_1d_19_step_0"), val = tensor(1)]; + tensor range_1d_19 = range_1d(end = reduce_prod_19, start = range_1d_19_start_0, step = range_1d_19_step_0)[name = tensor("range_1d_19")]; + tensor reshape_95 = reshape(shape = shape_67, x = range_1d_19)[name = tensor("reshape_95")]; + tensor slice_by_index_19 = slice_by_index(begin = concat_67, begin_mask = new_cache_19_internal_tensor_assign_2_begin_mask_0, end = concat_68, end_mask = new_cache_19_internal_tensor_assign_2_end_mask_0, squeeze_mask = new_cache_19_internal_tensor_assign_2_squeeze_mask_0, stride = new_cache_19_internal_tensor_assign_2_stride_0, x = reshape_95)[name = tensor("slice_by_index_19")]; + tensor reshape_96_shape_0 = const()[name = tensor("reshape_96_shape_0"), val = tensor([-1])]; + tensor reshape_96 = reshape(shape = reshape_96_shape_0, x = slice_by_index_19)[name = tensor("reshape_96")]; + tensor reshape_97_shape_0 = const()[name = tensor("reshape_97_shape_0"), val = tensor([-1])]; + tensor reshape_97_cast_fp16 = reshape(shape = reshape_97_shape_0, x = var_4257_cast_fp16)[name = tensor("reshape_97_cast_fp16")]; + tensor reshape_98_shape_0 = const()[name = tensor("reshape_98_shape_0"), val = tensor([-1])]; + tensor reshape_98_cast_fp16 = reshape(shape = reshape_98_shape_0, x = reshape_94_cast_fp16)[name = tensor("reshape_98_cast_fp16")]; + tensor scatter_19_mode_0 = const()[name = tensor("scatter_19_mode_0"), val = tensor("update")]; + tensor scatter_19_axis_0 = const()[name = tensor("scatter_19_axis_0"), val = tensor(0)]; + tensor scatter_19_validate_indices_0 = const()[name = tensor("scatter_19_validate_indices_0"), val = tensor(false)]; + tensor scatter_19_cast_fp16 = scatter(axis = scatter_19_axis_0, data = reshape_98_cast_fp16, indices = reshape_96, mode = scatter_19_mode_0, updates = reshape_97_cast_fp16, validate_indices = scatter_19_validate_indices_0)[name = tensor("scatter_19_cast_fp16")]; + tensor reshape_99_cast_fp16 = reshape(shape = shape_67, x = scatter_19_cast_fp16)[name = tensor("reshape_99_cast_fp16")]; + tensor reshape_99_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("reshape_99_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor keys_55_begin_0 = const()[name = tensor("keys_55_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor keys_55_end_0 = const()[name = tensor("keys_55_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor keys_55_end_mask_0 = const()[name = tensor("keys_55_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor keys_55_squeeze_mask_0 = const()[name = tensor("keys_55_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor keys_55_cast_fp16 = slice_by_index(begin = keys_55_begin_0, end = keys_55_end_0, end_mask = keys_55_end_mask_0, squeeze_mask = keys_55_squeeze_mask_0, x = reshape_99_cast_fp16)[name = tensor("keys_55_cast_fp16")]; + tensor values_55_begin_0 = const()[name = tensor("values_55_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor values_55_end_0 = const()[name = tensor("values_55_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor values_55_end_mask_0 = const()[name = tensor("values_55_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor values_55_squeeze_mask_0 = const()[name = tensor("values_55_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor values_55_cast_fp16 = slice_by_index(begin = values_55_begin_0, end = values_55_end_0, end_mask = values_55_end_mask_0, squeeze_mask = values_55_squeeze_mask_0, x = reshape_99_cast_fp16)[name = tensor("values_55_cast_fp16")]; + tensor var_4269_cast_fp16 = not_equal(x = keys_55_cast_fp16, y = keys_55_cast_fp16)[name = tensor("op_4269_cast_fp16")]; + tensor keys_57_cast_fp16 = select(a = var_504_to_fp16, b = keys_55_cast_fp16, cond = var_4269_cast_fp16)[name = tensor("keys_57_cast_fp16")]; + tensor var_4277_cast_fp16 = not_equal(x = values_55_cast_fp16, y = values_55_cast_fp16)[name = tensor("op_4277_cast_fp16")]; + tensor values_57_cast_fp16 = select(a = var_504_to_fp16, b = values_55_cast_fp16, cond = var_4277_cast_fp16)[name = tensor("values_57_cast_fp16")]; + tensor var_4301 = const()[name = tensor("op_4301"), val = tensor([0, 2, 1, 3])]; + tensor var_4314 = const()[name = tensor("op_4314"), val = tensor([1, 1, 1])]; + tensor var_4315_cast_fp16 = reshape(shape = var_4314, x = position9_to_fp16)[name = tensor("op_4315_cast_fp16")]; + tensor var_4332_to_fp16 = const()[name = tensor("op_4332_to_fp16"), val = tensor(0x1p+0)]; + tensor valid_len_19_cast_fp16 = add(x = var_4315_cast_fp16, y = var_4332_to_fp16)[name = tensor("valid_len_19_cast_fp16")]; + tensor valid_mask_19_cast_fp16 = less(x = k_positions_1_promoted_to_fp16, y = valid_len_19_cast_fp16)[name = tensor("valid_mask_19_cast_fp16")]; + tensor causal_mask_19_cast_fp16 = less_equal(x = k_positions_1_promoted_to_fp16, y = var_4315_cast_fp16)[name = tensor("causal_mask_19_cast_fp16")]; + tensor attn_mask_37 = logical_and(x = valid_mask_19_cast_fp16, y = causal_mask_19_cast_fp16)[name = tensor("attn_mask_37")]; + tensor attn_mask_39_axes_0 = const()[name = tensor("attn_mask_39_axes_0"), val = tensor([1])]; + tensor attn_mask_39 = expand_dims(axes = attn_mask_39_axes_0, x = attn_mask_37)[name = tensor("attn_mask_39")]; + tensor var_4350_transpose_x_0 = const()[name = tensor("op_4350_transpose_x_0"), val = tensor(false)]; + tensor var_4350_transpose_y_0 = const()[name = tensor("op_4350_transpose_y_0"), val = tensor(false)]; + tensor transpose_90_perm_0 = const()[name = tensor("transpose_90_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_91_perm_0 = const()[name = tensor("transpose_91_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_91 = transpose(perm = transpose_91_perm_0, x = keys_57_cast_fp16)[name = tensor("transpose_177")]; + tensor transpose_90 = transpose(perm = transpose_90_perm_0, x = q_57_cast_fp16)[name = tensor("transpose_178")]; + tensor var_4350_cast_fp16 = matmul(transpose_x = var_4350_transpose_x_0, transpose_y = var_4350_transpose_y_0, x = transpose_90, y = transpose_91)[name = tensor("op_4350_cast_fp16")]; + tensor var_4344_to_fp16 = const()[name = tensor("op_4344_to_fp16"), val = tensor([0x1p-3])]; + tensor attn_weights_55_cast_fp16 = mul(x = var_4350_cast_fp16, y = var_4344_to_fp16)[name = tensor("attn_weights_55_cast_fp16")]; + tensor var_4352 = logical_not(x = attn_mask_39)[name = tensor("op_4352")]; + tensor var_4353_to_fp16 = const()[name = tensor("op_4353_to_fp16"), val = tensor(-inf)]; + tensor attn_weights_57_cast_fp16 = select(a = var_4353_to_fp16, b = attn_weights_55_cast_fp16, cond = var_4352)[name = tensor("attn_weights_57_cast_fp16")]; + tensor var_4355 = const()[name = tensor("op_4355"), val = tensor(-1)]; + tensor attn_weights_59_cast_fp16 = softmax(axis = var_4355, x = attn_weights_57_cast_fp16)[name = tensor("attn_weights_59_cast_fp16")]; + tensor attn_output_19_transpose_x_0 = const()[name = tensor("attn_output_19_transpose_x_0"), val = tensor(false)]; + tensor attn_output_19_transpose_y_0 = const()[name = tensor("attn_output_19_transpose_y_0"), val = tensor(false)]; + tensor values_59_cast_fp16 = transpose(perm = var_4301, x = values_57_cast_fp16)[name = tensor("transpose_179")]; + tensor attn_output_19_cast_fp16 = matmul(transpose_x = attn_output_19_transpose_x_0, transpose_y = attn_output_19_transpose_y_0, x = attn_weights_59_cast_fp16, y = values_59_cast_fp16)[name = tensor("attn_output_19_cast_fp16")]; + tensor var_4363 = const()[name = tensor("op_4363"), val = tensor([0, 2, 1, 3])]; + tensor var_4366 = const()[name = tensor("op_4366"), val = tensor([1, 1, 1024])]; + tensor var_4364_cast_fp16 = transpose(perm = var_4363, x = attn_output_19_cast_fp16)[name = tensor("transpose_176")]; + tensor input_95_cast_fp16 = reshape(shape = var_4366, x = var_4364_cast_fp16)[name = tensor("input_95_cast_fp16")]; + tensor linear_38_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = attn9_out_proj_weight_cast_fp16, x = input_95_cast_fp16)[name = tensor("linear_38_cast_fp16")]; + tensor var_4372_to_fp16 = const()[name = tensor("op_4372_to_fp16"), val = tensor(0x1p+0)]; + tensor var_4373_cast_fp16 = add(x = position9_to_fp16, y = var_4372_to_fp16)[name = tensor("op_4373_cast_fp16")]; + tensor var_4373_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_4373_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor input_97_cast_fp16 = add(x = input_93_cast_fp16, y = linear_38_cast_fp16)[name = tensor("input_97_cast_fp16")]; + tensor input_99_axes_0 = const()[name = tensor("input_99_axes_0"), val = tensor([-1])]; + tensor norm9_2_weight_to_fp16 = const()[name = tensor("norm9_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(303666432)))]; + tensor norm9_2_bias_to_fp16 = const()[name = tensor("norm9_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(303668544)))]; + tensor var_4377_to_fp16 = const()[name = tensor("op_4377_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_99_cast_fp16 = layer_norm(axes = input_99_axes_0, beta = norm9_2_bias_to_fp16, epsilon = var_4377_to_fp16, gamma = norm9_2_weight_to_fp16, x = input_97_cast_fp16)[name = tensor("input_99_cast_fp16")]; + tensor linear_39_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = linear9_1_weight_cast_fp16, x = input_99_cast_fp16)[name = tensor("linear_39_cast_fp16")]; + tensor input_101_mode_0 = const()[name = tensor("input_101_mode_0"), val = tensor("EXACT")]; + tensor input_101_cast_fp16 = gelu(mode = input_101_mode_0, x = linear_39_cast_fp16)[name = tensor("input_101_cast_fp16")]; + tensor linear_40_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = linear9_2_weight_cast_fp16, x = input_101_cast_fp16)[name = tensor("linear_40_cast_fp16")]; + tensor input_103_cast_fp16 = add(x = input_97_cast_fp16, y = linear_40_cast_fp16)[name = tensor("input_103_cast_fp16")]; + tensor x_21_axes_0 = const()[name = tensor("x_21_axes_0"), val = tensor([-1])]; + tensor norm10_1_weight_to_fp16 = const()[name = tensor("norm10_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(303670656)))]; + tensor norm10_1_bias_to_fp16 = const()[name = tensor("norm10_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(303672768)))]; + tensor var_4394_to_fp16 = const()[name = tensor("op_4394_to_fp16"), val = tensor(0x1.5p-17)]; + tensor x_21_cast_fp16 = layer_norm(axes = x_21_axes_0, beta = norm10_1_bias_to_fp16, epsilon = var_4394_to_fp16, gamma = norm10_1_weight_to_fp16, x = input_103_cast_fp16)[name = tensor("x_21_cast_fp16")]; + tensor linear_41_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = attn10_in_proj_weight_cast_fp16, x = x_21_cast_fp16)[name = tensor("linear_41_cast_fp16")]; + tensor var_4430 = const()[name = tensor("op_4430"), val = tensor([1, 1, 3, 16, 64])]; + tensor qkv_21_cast_fp16 = reshape(shape = var_4430, x = linear_41_cast_fp16)[name = tensor("qkv_21_cast_fp16")]; + tensor q_61_begin_0 = const()[name = tensor("q_61_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor q_61_end_0 = const()[name = tensor("q_61_end_0"), val = tensor([1, 1, 1, 16, 64])]; + tensor q_61_end_mask_0 = const()[name = tensor("q_61_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor q_61_squeeze_mask_0 = const()[name = tensor("q_61_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor q_61_cast_fp16 = slice_by_index(begin = q_61_begin_0, end = q_61_end_0, end_mask = q_61_end_mask_0, squeeze_mask = q_61_squeeze_mask_0, x = qkv_21_cast_fp16)[name = tensor("q_61_cast_fp16")]; + tensor k_41_begin_0 = const()[name = tensor("k_41_begin_0"), val = tensor([0, 0, 1, 0, 0])]; + tensor k_41_end_0 = const()[name = tensor("k_41_end_0"), val = tensor([1, 1, 2, 16, 64])]; + tensor k_41_end_mask_0 = const()[name = tensor("k_41_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor k_41_squeeze_mask_0 = const()[name = tensor("k_41_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor k_41_cast_fp16 = slice_by_index(begin = k_41_begin_0, end = k_41_end_0, end_mask = k_41_end_mask_0, squeeze_mask = k_41_squeeze_mask_0, x = qkv_21_cast_fp16)[name = tensor("k_41_cast_fp16")]; + tensor v_21_begin_0 = const()[name = tensor("v_21_begin_0"), val = tensor([0, 0, 2, 0, 0])]; + tensor v_21_end_0 = const()[name = tensor("v_21_end_0"), val = tensor([1, 1, 3, 16, 64])]; + tensor v_21_end_mask_0 = const()[name = tensor("v_21_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor v_21_squeeze_mask_0 = const()[name = tensor("v_21_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor v_21_cast_fp16 = slice_by_index(begin = v_21_begin_0, end = v_21_end_0, end_mask = v_21_end_mask_0, squeeze_mask = v_21_squeeze_mask_0, x = qkv_21_cast_fp16)[name = tensor("v_21_cast_fp16")]; + tensor var_4534 = const()[name = tensor("op_4534"), val = tensor([1, 1, 1, 1])]; + tensor position10_to_fp16_dtype_0 = const()[name = tensor("position10_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor position10_to_fp16 = cast(dtype = position10_to_fp16_dtype_0, x = position10)[name = tensor("cast_503")]; + tensor ts_65_cast_fp16 = reshape(shape = var_4534, x = position10_to_fp16)[name = tensor("ts_65_cast_fp16")]; + tensor var_4538 = const()[name = tensor("op_4538"), val = tensor([1, 1, 16, 32, 2])]; + tensor q_complex_21_cast_fp16 = reshape(shape = var_4538, x = q_61_cast_fp16)[name = tensor("q_complex_21_cast_fp16")]; + tensor var_4542 = const()[name = tensor("op_4542"), val = tensor([1, 1, 16, 32, 2])]; + tensor k_complex_21_cast_fp16 = reshape(shape = var_4542, x = k_41_cast_fp16)[name = tensor("k_complex_21_cast_fp16")]; + tensor var_4546_begin_0 = const()[name = tensor("op_4546_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_4546_end_0 = const()[name = tensor("op_4546_end_0"), val = tensor([1, 1, 16, 32, 1])]; + tensor var_4546_end_mask_0 = const()[name = tensor("op_4546_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_4546_squeeze_mask_0 = const()[name = tensor("op_4546_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_4546_cast_fp16 = slice_by_index(begin = var_4546_begin_0, end = var_4546_end_0, end_mask = var_4546_end_mask_0, squeeze_mask = var_4546_squeeze_mask_0, x = q_complex_21_cast_fp16)[name = tensor("op_4546_cast_fp16")]; + tensor var_4554_begin_0 = const()[name = tensor("op_4554_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_4554_end_0 = const()[name = tensor("op_4554_end_0"), val = tensor([1, 1, 16, 32, 2])]; + tensor var_4554_end_mask_0 = const()[name = tensor("op_4554_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_4554_squeeze_mask_0 = const()[name = tensor("op_4554_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_4554_cast_fp16 = slice_by_index(begin = var_4554_begin_0, end = var_4554_end_0, end_mask = var_4554_end_mask_0, squeeze_mask = var_4554_squeeze_mask_0, x = q_complex_21_cast_fp16)[name = tensor("op_4554_cast_fp16")]; + tensor var_4562_begin_0 = const()[name = tensor("op_4562_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_4562_end_0 = const()[name = tensor("op_4562_end_0"), val = tensor([1, 1, 16, 32, 1])]; + tensor var_4562_end_mask_0 = const()[name = tensor("op_4562_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_4562_squeeze_mask_0 = const()[name = tensor("op_4562_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_4562_cast_fp16 = slice_by_index(begin = var_4562_begin_0, end = var_4562_end_0, end_mask = var_4562_end_mask_0, squeeze_mask = var_4562_squeeze_mask_0, x = k_complex_21_cast_fp16)[name = tensor("op_4562_cast_fp16")]; + tensor var_4570_begin_0 = const()[name = tensor("op_4570_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_4570_end_0 = const()[name = tensor("op_4570_end_0"), val = tensor([1, 1, 16, 32, 2])]; + tensor var_4570_end_mask_0 = const()[name = tensor("op_4570_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_4570_squeeze_mask_0 = const()[name = tensor("op_4570_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_4570_cast_fp16 = slice_by_index(begin = var_4570_begin_0, end = var_4570_end_0, end_mask = var_4570_end_mask_0, squeeze_mask = var_4570_squeeze_mask_0, x = k_complex_21_cast_fp16)[name = tensor("op_4570_cast_fp16")]; + tensor freqs_21_to_fp16 = const()[name = tensor("freqs_21_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(303674880)))]; + tensor var_4576_cast_fp16 = mul(x = freqs_21_to_fp16, y = ts_65_cast_fp16)[name = tensor("op_4576_cast_fp16")]; + tensor rotr_21_cast_fp16 = cos(x = var_4576_cast_fp16)[name = tensor("rotr_21_cast_fp16")]; + tensor roti_21_cast_fp16 = sin(x = var_4576_cast_fp16)[name = tensor("roti_21_cast_fp16")]; + tensor var_4580_cast_fp16 = mul(x = var_4546_cast_fp16, y = rotr_21_cast_fp16)[name = tensor("op_4580_cast_fp16")]; + tensor var_4581_cast_fp16 = mul(x = var_4554_cast_fp16, y = roti_21_cast_fp16)[name = tensor("op_4581_cast_fp16")]; + tensor qor_41_cast_fp16 = sub(x = var_4580_cast_fp16, y = var_4581_cast_fp16)[name = tensor("qor_41_cast_fp16")]; + tensor var_4584_cast_fp16 = mul(x = var_4546_cast_fp16, y = roti_21_cast_fp16)[name = tensor("op_4584_cast_fp16")]; + tensor var_4585_cast_fp16 = mul(x = var_4554_cast_fp16, y = rotr_21_cast_fp16)[name = tensor("op_4585_cast_fp16")]; + tensor qoi_41_cast_fp16 = add(x = var_4584_cast_fp16, y = var_4585_cast_fp16)[name = tensor("qoi_41_cast_fp16")]; + tensor var_4588_cast_fp16 = mul(x = var_4562_cast_fp16, y = rotr_21_cast_fp16)[name = tensor("op_4588_cast_fp16")]; + tensor var_4589_cast_fp16 = mul(x = var_4570_cast_fp16, y = roti_21_cast_fp16)[name = tensor("op_4589_cast_fp16")]; + tensor kor_41_cast_fp16 = sub(x = var_4588_cast_fp16, y = var_4589_cast_fp16)[name = tensor("kor_41_cast_fp16")]; + tensor var_4592_cast_fp16 = mul(x = var_4562_cast_fp16, y = roti_21_cast_fp16)[name = tensor("op_4592_cast_fp16")]; + tensor var_4593_cast_fp16 = mul(x = var_4570_cast_fp16, y = rotr_21_cast_fp16)[name = tensor("op_4593_cast_fp16")]; + tensor koi_41_cast_fp16 = add(x = var_4592_cast_fp16, y = var_4593_cast_fp16)[name = tensor("koi_41_cast_fp16")]; + tensor qo_21_axis_0 = const()[name = tensor("qo_21_axis_0"), val = tensor(-1)]; + tensor qo_21_cast_fp16 = stack(axis = qo_21_axis_0, values = (qor_41_cast_fp16, qoi_41_cast_fp16))[name = tensor("qo_21_cast_fp16")]; + tensor ko_21_axis_0 = const()[name = tensor("ko_21_axis_0"), val = tensor(-1)]; + tensor ko_21_cast_fp16 = stack(axis = ko_21_axis_0, values = (kor_41_cast_fp16, koi_41_cast_fp16))[name = tensor("ko_21_cast_fp16")]; + tensor var_4622 = const()[name = tensor("op_4622"), val = tensor([1, 1, 16, 64])]; + tensor q_63_cast_fp16 = reshape(shape = var_4622, x = qo_21_cast_fp16)[name = tensor("q_63_cast_fp16")]; + tensor var_4624 = const()[name = tensor("op_4624"), val = tensor([1, 1, 16, 64])]; + tensor k_43_cast_fp16 = reshape(shape = var_4624, x = ko_21_cast_fp16)[name = tensor("k_43_cast_fp16")]; + tensor _inversed_4646_y_0_to_fp16 = const()[name = tensor("_inversed_4646_y_0_to_fp16"), val = tensor(0x1p-9)]; + tensor _inversed_4646_cast_fp16 = mul(x = ts_65_cast_fp16, y = _inversed_4646_y_0_to_fp16)[name = tensor("_inversed_4646_cast_fp16")]; + tensor var_4647_cast_fp16 = floor(x = _inversed_4646_cast_fp16)[name = tensor("op_4647_cast_fp16")]; + tensor var_4648_to_fp16 = const()[name = tensor("op_4648_to_fp16"), val = tensor(0x1p+9)]; + tensor var_4649_cast_fp16 = mul(x = var_4647_cast_fp16, y = var_4648_to_fp16)[name = tensor("op_4649_cast_fp16")]; + tensor write_indices_float_43_cast_fp16 = sub(x = ts_65_cast_fp16, y = var_4649_cast_fp16)[name = tensor("write_indices_float_43_cast_fp16")]; + tensor var_4656_dtype_0 = const()[name = tensor("op_4656_dtype_0"), val = tensor("int32")]; + tensor write_indices_21_reps_0 = const()[name = tensor("write_indices_21_reps_0"), val = tensor([1, 1, 16, 64])]; + tensor write_indices_float_43_cast_fp16_to_int32 = cast(dtype = var_4656_dtype_0, x = write_indices_float_43_cast_fp16)[name = tensor("cast_502")]; + tensor write_indices_21 = tile(reps = write_indices_21_reps_0, x = write_indices_float_43_cast_fp16_to_int32)[name = tensor("write_indices_21")]; + tensor var_4664_begin_0 = const()[name = tensor("op_4664_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_4664_end_0 = const()[name = tensor("op_4664_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor var_4664_end_mask_0 = const()[name = tensor("op_4664_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_4664_squeeze_mask_0 = const()[name = tensor("op_4664_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor cache10_to_fp16_dtype_0 = const()[name = tensor("cache10_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor cache10_to_fp16 = cast(dtype = cache10_to_fp16_dtype_0, x = cache10)[name = tensor("cast_501")]; + tensor var_4664_cast_fp16 = slice_by_index(begin = var_4664_begin_0, end = var_4664_end_0, end_mask = var_4664_end_mask_0, squeeze_mask = var_4664_squeeze_mask_0, x = cache10_to_fp16)[name = tensor("op_4664_cast_fp16")]; + tensor var_4666_axis_0 = const()[name = tensor("op_4666_axis_0"), val = tensor(1)]; + tensor var_4666_mode_0 = const()[name = tensor("op_4666_mode_0"), val = tensor("update")]; + tensor var_4666_validate_indices_0 = const()[name = tensor("op_4666_validate_indices_0"), val = tensor(false)]; + tensor var_4666_cast_fp16 = scatter_along_axis(axis = var_4666_axis_0, data = var_4664_cast_fp16, indices = write_indices_21, mode = var_4666_mode_0, updates = k_43_cast_fp16, validate_indices = var_4666_validate_indices_0)[name = tensor("op_4666_cast_fp16")]; + tensor concat_72 = const()[name = tensor("concat_72"), val = tensor([0, 0, 0, 0, 0])]; + tensor concat_73 = const()[name = tensor("concat_73"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_21_internal_tensor_assign_1_stride_0 = const()[name = tensor("new_cache_21_internal_tensor_assign_1_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_21_internal_tensor_assign_1_begin_mask_0 = const()[name = tensor("new_cache_21_internal_tensor_assign_1_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_21_internal_tensor_assign_1_end_mask_0 = const()[name = tensor("new_cache_21_internal_tensor_assign_1_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_21_internal_tensor_assign_1_squeeze_mask_0 = const()[name = tensor("new_cache_21_internal_tensor_assign_1_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_68 = const()[name = tensor("shape_68"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_20 = const()[name = tensor("reduce_prod_20"), val = tensor(1048576)]; + tensor range_1d_20_start_0 = const()[name = tensor("range_1d_20_start_0"), val = tensor(0)]; + tensor range_1d_20_step_0 = const()[name = tensor("range_1d_20_step_0"), val = tensor(1)]; + tensor range_1d_20 = range_1d(end = reduce_prod_20, start = range_1d_20_start_0, step = range_1d_20_step_0)[name = tensor("range_1d_20")]; + tensor reshape_100 = reshape(shape = shape_68, x = range_1d_20)[name = tensor("reshape_100")]; + tensor slice_by_index_20 = slice_by_index(begin = concat_72, begin_mask = new_cache_21_internal_tensor_assign_1_begin_mask_0, end = concat_73, end_mask = new_cache_21_internal_tensor_assign_1_end_mask_0, squeeze_mask = new_cache_21_internal_tensor_assign_1_squeeze_mask_0, stride = new_cache_21_internal_tensor_assign_1_stride_0, x = reshape_100)[name = tensor("slice_by_index_20")]; + tensor reshape_101_shape_0 = const()[name = tensor("reshape_101_shape_0"), val = tensor([-1])]; + tensor reshape_101 = reshape(shape = reshape_101_shape_0, x = slice_by_index_20)[name = tensor("reshape_101")]; + tensor reshape_102_shape_0 = const()[name = tensor("reshape_102_shape_0"), val = tensor([-1])]; + tensor reshape_102_cast_fp16 = reshape(shape = reshape_102_shape_0, x = var_4666_cast_fp16)[name = tensor("reshape_102_cast_fp16")]; + tensor reshape_103_shape_0 = const()[name = tensor("reshape_103_shape_0"), val = tensor([-1])]; + tensor reshape_103_cast_fp16 = reshape(shape = reshape_103_shape_0, x = cache10_to_fp16)[name = tensor("reshape_103_cast_fp16")]; + tensor scatter_20_mode_0 = const()[name = tensor("scatter_20_mode_0"), val = tensor("update")]; + tensor scatter_20_axis_0 = const()[name = tensor("scatter_20_axis_0"), val = tensor(0)]; + tensor scatter_20_validate_indices_0 = const()[name = tensor("scatter_20_validate_indices_0"), val = tensor(false)]; + tensor scatter_20_cast_fp16 = scatter(axis = scatter_20_axis_0, data = reshape_103_cast_fp16, indices = reshape_101, mode = scatter_20_mode_0, updates = reshape_102_cast_fp16, validate_indices = scatter_20_validate_indices_0)[name = tensor("scatter_20_cast_fp16")]; + tensor reshape_104_cast_fp16 = reshape(shape = shape_68, x = scatter_20_cast_fp16)[name = tensor("reshape_104_cast_fp16")]; + tensor var_4674_begin_0 = const()[name = tensor("op_4674_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor var_4674_end_0 = const()[name = tensor("op_4674_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor var_4674_end_mask_0 = const()[name = tensor("op_4674_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_4674_squeeze_mask_0 = const()[name = tensor("op_4674_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor var_4674_cast_fp16 = slice_by_index(begin = var_4674_begin_0, end = var_4674_end_0, end_mask = var_4674_end_mask_0, squeeze_mask = var_4674_squeeze_mask_0, x = reshape_104_cast_fp16)[name = tensor("op_4674_cast_fp16")]; + tensor var_4676_axis_0 = const()[name = tensor("op_4676_axis_0"), val = tensor(1)]; + tensor var_4676_mode_0 = const()[name = tensor("op_4676_mode_0"), val = tensor("update")]; + tensor var_4676_validate_indices_0 = const()[name = tensor("op_4676_validate_indices_0"), val = tensor(false)]; + tensor var_4676_cast_fp16 = scatter_along_axis(axis = var_4676_axis_0, data = var_4674_cast_fp16, indices = write_indices_21, mode = var_4676_mode_0, updates = v_21_cast_fp16, validate_indices = var_4676_validate_indices_0)[name = tensor("op_4676_cast_fp16")]; + tensor concat_74 = const()[name = tensor("concat_74"), val = tensor([1, 0, 0, 0, 0])]; + tensor concat_75 = const()[name = tensor("concat_75"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_21_internal_tensor_assign_2_stride_0 = const()[name = tensor("new_cache_21_internal_tensor_assign_2_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_21_internal_tensor_assign_2_begin_mask_0 = const()[name = tensor("new_cache_21_internal_tensor_assign_2_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_21_internal_tensor_assign_2_end_mask_0 = const()[name = tensor("new_cache_21_internal_tensor_assign_2_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_21_internal_tensor_assign_2_squeeze_mask_0 = const()[name = tensor("new_cache_21_internal_tensor_assign_2_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_69 = const()[name = tensor("shape_69"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_21 = const()[name = tensor("reduce_prod_21"), val = tensor(1048576)]; + tensor range_1d_21_start_0 = const()[name = tensor("range_1d_21_start_0"), val = tensor(0)]; + tensor range_1d_21_step_0 = const()[name = tensor("range_1d_21_step_0"), val = tensor(1)]; + tensor range_1d_21 = range_1d(end = reduce_prod_21, start = range_1d_21_start_0, step = range_1d_21_step_0)[name = tensor("range_1d_21")]; + tensor reshape_105 = reshape(shape = shape_69, x = range_1d_21)[name = tensor("reshape_105")]; + tensor slice_by_index_21 = slice_by_index(begin = concat_74, begin_mask = new_cache_21_internal_tensor_assign_2_begin_mask_0, end = concat_75, end_mask = new_cache_21_internal_tensor_assign_2_end_mask_0, squeeze_mask = new_cache_21_internal_tensor_assign_2_squeeze_mask_0, stride = new_cache_21_internal_tensor_assign_2_stride_0, x = reshape_105)[name = tensor("slice_by_index_21")]; + tensor reshape_106_shape_0 = const()[name = tensor("reshape_106_shape_0"), val = tensor([-1])]; + tensor reshape_106 = reshape(shape = reshape_106_shape_0, x = slice_by_index_21)[name = tensor("reshape_106")]; + tensor reshape_107_shape_0 = const()[name = tensor("reshape_107_shape_0"), val = tensor([-1])]; + tensor reshape_107_cast_fp16 = reshape(shape = reshape_107_shape_0, x = var_4676_cast_fp16)[name = tensor("reshape_107_cast_fp16")]; + tensor reshape_108_shape_0 = const()[name = tensor("reshape_108_shape_0"), val = tensor([-1])]; + tensor reshape_108_cast_fp16 = reshape(shape = reshape_108_shape_0, x = reshape_104_cast_fp16)[name = tensor("reshape_108_cast_fp16")]; + tensor scatter_21_mode_0 = const()[name = tensor("scatter_21_mode_0"), val = tensor("update")]; + tensor scatter_21_axis_0 = const()[name = tensor("scatter_21_axis_0"), val = tensor(0)]; + tensor scatter_21_validate_indices_0 = const()[name = tensor("scatter_21_validate_indices_0"), val = tensor(false)]; + tensor scatter_21_cast_fp16 = scatter(axis = scatter_21_axis_0, data = reshape_108_cast_fp16, indices = reshape_106, mode = scatter_21_mode_0, updates = reshape_107_cast_fp16, validate_indices = scatter_21_validate_indices_0)[name = tensor("scatter_21_cast_fp16")]; + tensor reshape_109_cast_fp16 = reshape(shape = shape_69, x = scatter_21_cast_fp16)[name = tensor("reshape_109_cast_fp16")]; + tensor reshape_109_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("reshape_109_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor keys_61_begin_0 = const()[name = tensor("keys_61_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor keys_61_end_0 = const()[name = tensor("keys_61_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor keys_61_end_mask_0 = const()[name = tensor("keys_61_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor keys_61_squeeze_mask_0 = const()[name = tensor("keys_61_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor keys_61_cast_fp16 = slice_by_index(begin = keys_61_begin_0, end = keys_61_end_0, end_mask = keys_61_end_mask_0, squeeze_mask = keys_61_squeeze_mask_0, x = reshape_109_cast_fp16)[name = tensor("keys_61_cast_fp16")]; + tensor values_61_begin_0 = const()[name = tensor("values_61_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor values_61_end_0 = const()[name = tensor("values_61_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor values_61_end_mask_0 = const()[name = tensor("values_61_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor values_61_squeeze_mask_0 = const()[name = tensor("values_61_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor values_61_cast_fp16 = slice_by_index(begin = values_61_begin_0, end = values_61_end_0, end_mask = values_61_end_mask_0, squeeze_mask = values_61_squeeze_mask_0, x = reshape_109_cast_fp16)[name = tensor("values_61_cast_fp16")]; + tensor var_4688_cast_fp16 = not_equal(x = keys_61_cast_fp16, y = keys_61_cast_fp16)[name = tensor("op_4688_cast_fp16")]; + tensor keys_63_cast_fp16 = select(a = var_504_to_fp16, b = keys_61_cast_fp16, cond = var_4688_cast_fp16)[name = tensor("keys_63_cast_fp16")]; + tensor var_4696_cast_fp16 = not_equal(x = values_61_cast_fp16, y = values_61_cast_fp16)[name = tensor("op_4696_cast_fp16")]; + tensor values_63_cast_fp16 = select(a = var_504_to_fp16, b = values_61_cast_fp16, cond = var_4696_cast_fp16)[name = tensor("values_63_cast_fp16")]; + tensor var_4720 = const()[name = tensor("op_4720"), val = tensor([0, 2, 1, 3])]; + tensor var_4733 = const()[name = tensor("op_4733"), val = tensor([1, 1, 1])]; + tensor var_4734_cast_fp16 = reshape(shape = var_4733, x = position10_to_fp16)[name = tensor("op_4734_cast_fp16")]; + tensor var_4751_to_fp16 = const()[name = tensor("op_4751_to_fp16"), val = tensor(0x1p+0)]; + tensor valid_len_21_cast_fp16 = add(x = var_4734_cast_fp16, y = var_4751_to_fp16)[name = tensor("valid_len_21_cast_fp16")]; + tensor valid_mask_21_cast_fp16 = less(x = k_positions_1_promoted_to_fp16, y = valid_len_21_cast_fp16)[name = tensor("valid_mask_21_cast_fp16")]; + tensor causal_mask_21_cast_fp16 = less_equal(x = k_positions_1_promoted_to_fp16, y = var_4734_cast_fp16)[name = tensor("causal_mask_21_cast_fp16")]; + tensor attn_mask_41 = logical_and(x = valid_mask_21_cast_fp16, y = causal_mask_21_cast_fp16)[name = tensor("attn_mask_41")]; + tensor attn_mask_43_axes_0 = const()[name = tensor("attn_mask_43_axes_0"), val = tensor([1])]; + tensor attn_mask_43 = expand_dims(axes = attn_mask_43_axes_0, x = attn_mask_41)[name = tensor("attn_mask_43")]; + tensor var_4769_transpose_x_0 = const()[name = tensor("op_4769_transpose_x_0"), val = tensor(false)]; + tensor var_4769_transpose_y_0 = const()[name = tensor("op_4769_transpose_y_0"), val = tensor(false)]; + tensor transpose_92_perm_0 = const()[name = tensor("transpose_92_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_93_perm_0 = const()[name = tensor("transpose_93_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_93 = transpose(perm = transpose_93_perm_0, x = keys_63_cast_fp16)[name = tensor("transpose_173")]; + tensor transpose_92 = transpose(perm = transpose_92_perm_0, x = q_63_cast_fp16)[name = tensor("transpose_174")]; + tensor var_4769_cast_fp16 = matmul(transpose_x = var_4769_transpose_x_0, transpose_y = var_4769_transpose_y_0, x = transpose_92, y = transpose_93)[name = tensor("op_4769_cast_fp16")]; + tensor var_4763_to_fp16 = const()[name = tensor("op_4763_to_fp16"), val = tensor([0x1p-3])]; + tensor attn_weights_61_cast_fp16 = mul(x = var_4769_cast_fp16, y = var_4763_to_fp16)[name = tensor("attn_weights_61_cast_fp16")]; + tensor var_4771 = logical_not(x = attn_mask_43)[name = tensor("op_4771")]; + tensor var_4772_to_fp16 = const()[name = tensor("op_4772_to_fp16"), val = tensor(-inf)]; + tensor attn_weights_63_cast_fp16 = select(a = var_4772_to_fp16, b = attn_weights_61_cast_fp16, cond = var_4771)[name = tensor("attn_weights_63_cast_fp16")]; + tensor var_4774 = const()[name = tensor("op_4774"), val = tensor(-1)]; + tensor attn_weights_65_cast_fp16 = softmax(axis = var_4774, x = attn_weights_63_cast_fp16)[name = tensor("attn_weights_65_cast_fp16")]; + tensor attn_output_21_transpose_x_0 = const()[name = tensor("attn_output_21_transpose_x_0"), val = tensor(false)]; + tensor attn_output_21_transpose_y_0 = const()[name = tensor("attn_output_21_transpose_y_0"), val = tensor(false)]; + tensor values_65_cast_fp16 = transpose(perm = var_4720, x = values_63_cast_fp16)[name = tensor("transpose_175")]; + tensor attn_output_21_cast_fp16 = matmul(transpose_x = attn_output_21_transpose_x_0, transpose_y = attn_output_21_transpose_y_0, x = attn_weights_65_cast_fp16, y = values_65_cast_fp16)[name = tensor("attn_output_21_cast_fp16")]; + tensor var_4782 = const()[name = tensor("op_4782"), val = tensor([0, 2, 1, 3])]; + tensor var_4785 = const()[name = tensor("op_4785"), val = tensor([1, 1, 1024])]; + tensor var_4783_cast_fp16 = transpose(perm = var_4782, x = attn_output_21_cast_fp16)[name = tensor("transpose_172")]; + tensor input_105_cast_fp16 = reshape(shape = var_4785, x = var_4783_cast_fp16)[name = tensor("input_105_cast_fp16")]; + tensor linear_42_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = attn10_out_proj_weight_cast_fp16, x = input_105_cast_fp16)[name = tensor("linear_42_cast_fp16")]; + tensor var_4791_to_fp16 = const()[name = tensor("op_4791_to_fp16"), val = tensor(0x1p+0)]; + tensor var_4792_cast_fp16 = add(x = position10_to_fp16, y = var_4791_to_fp16)[name = tensor("op_4792_cast_fp16")]; + tensor var_4792_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_4792_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor input_107_cast_fp16 = add(x = input_103_cast_fp16, y = linear_42_cast_fp16)[name = tensor("input_107_cast_fp16")]; + tensor input_109_axes_0 = const()[name = tensor("input_109_axes_0"), val = tensor([-1])]; + tensor norm10_2_weight_to_fp16 = const()[name = tensor("norm10_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(303675008)))]; + tensor norm10_2_bias_to_fp16 = const()[name = tensor("norm10_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(303677120)))]; + tensor var_4796_to_fp16 = const()[name = tensor("op_4796_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_109_cast_fp16 = layer_norm(axes = input_109_axes_0, beta = norm10_2_bias_to_fp16, epsilon = var_4796_to_fp16, gamma = norm10_2_weight_to_fp16, x = input_107_cast_fp16)[name = tensor("input_109_cast_fp16")]; + tensor linear_43_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = linear10_1_weight_cast_fp16, x = input_109_cast_fp16)[name = tensor("linear_43_cast_fp16")]; + tensor input_111_mode_0 = const()[name = tensor("input_111_mode_0"), val = tensor("EXACT")]; + tensor input_111_cast_fp16 = gelu(mode = input_111_mode_0, x = linear_43_cast_fp16)[name = tensor("input_111_cast_fp16")]; + tensor linear_44_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = linear10_2_weight_cast_fp16, x = input_111_cast_fp16)[name = tensor("linear_44_cast_fp16")]; + tensor input_113_cast_fp16 = add(x = input_107_cast_fp16, y = linear_44_cast_fp16)[name = tensor("input_113_cast_fp16")]; + tensor x_23_axes_0 = const()[name = tensor("x_23_axes_0"), val = tensor([-1])]; + tensor norm11_1_weight_to_fp16 = const()[name = tensor("norm11_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(303679232)))]; + tensor norm11_1_bias_to_fp16 = const()[name = tensor("norm11_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(303681344)))]; + tensor var_4813_to_fp16 = const()[name = tensor("op_4813_to_fp16"), val = tensor(0x1.5p-17)]; + tensor x_23_cast_fp16 = layer_norm(axes = x_23_axes_0, beta = norm11_1_bias_to_fp16, epsilon = var_4813_to_fp16, gamma = norm11_1_weight_to_fp16, x = input_113_cast_fp16)[name = tensor("x_23_cast_fp16")]; + tensor linear_45_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = attn11_in_proj_weight_cast_fp16, x = x_23_cast_fp16)[name = tensor("linear_45_cast_fp16")]; + tensor var_4849 = const()[name = tensor("op_4849"), val = tensor([1, 1, 3, 16, 64])]; + tensor qkv_23_cast_fp16 = reshape(shape = var_4849, x = linear_45_cast_fp16)[name = tensor("qkv_23_cast_fp16")]; + tensor q_67_begin_0 = const()[name = tensor("q_67_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor q_67_end_0 = const()[name = tensor("q_67_end_0"), val = tensor([1, 1, 1, 16, 64])]; + tensor q_67_end_mask_0 = const()[name = tensor("q_67_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor q_67_squeeze_mask_0 = const()[name = tensor("q_67_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor q_67_cast_fp16 = slice_by_index(begin = q_67_begin_0, end = q_67_end_0, end_mask = q_67_end_mask_0, squeeze_mask = q_67_squeeze_mask_0, x = qkv_23_cast_fp16)[name = tensor("q_67_cast_fp16")]; + tensor k_45_begin_0 = const()[name = tensor("k_45_begin_0"), val = tensor([0, 0, 1, 0, 0])]; + tensor k_45_end_0 = const()[name = tensor("k_45_end_0"), val = tensor([1, 1, 2, 16, 64])]; + tensor k_45_end_mask_0 = const()[name = tensor("k_45_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor k_45_squeeze_mask_0 = const()[name = tensor("k_45_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor k_45_cast_fp16 = slice_by_index(begin = k_45_begin_0, end = k_45_end_0, end_mask = k_45_end_mask_0, squeeze_mask = k_45_squeeze_mask_0, x = qkv_23_cast_fp16)[name = tensor("k_45_cast_fp16")]; + tensor v_23_begin_0 = const()[name = tensor("v_23_begin_0"), val = tensor([0, 0, 2, 0, 0])]; + tensor v_23_end_0 = const()[name = tensor("v_23_end_0"), val = tensor([1, 1, 3, 16, 64])]; + tensor v_23_end_mask_0 = const()[name = tensor("v_23_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor v_23_squeeze_mask_0 = const()[name = tensor("v_23_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor v_23_cast_fp16 = slice_by_index(begin = v_23_begin_0, end = v_23_end_0, end_mask = v_23_end_mask_0, squeeze_mask = v_23_squeeze_mask_0, x = qkv_23_cast_fp16)[name = tensor("v_23_cast_fp16")]; + tensor var_4953 = const()[name = tensor("op_4953"), val = tensor([1, 1, 1, 1])]; + tensor position11_to_fp16_dtype_0 = const()[name = tensor("position11_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor position11_to_fp16 = cast(dtype = position11_to_fp16_dtype_0, x = position11)[name = tensor("cast_498")]; + tensor ts_71_cast_fp16 = reshape(shape = var_4953, x = position11_to_fp16)[name = tensor("ts_71_cast_fp16")]; + tensor var_4957 = const()[name = tensor("op_4957"), val = tensor([1, 1, 16, 32, 2])]; + tensor q_complex_23_cast_fp16 = reshape(shape = var_4957, x = q_67_cast_fp16)[name = tensor("q_complex_23_cast_fp16")]; + tensor var_4961 = const()[name = tensor("op_4961"), val = tensor([1, 1, 16, 32, 2])]; + tensor k_complex_23_cast_fp16 = reshape(shape = var_4961, x = k_45_cast_fp16)[name = tensor("k_complex_23_cast_fp16")]; + tensor var_4965_begin_0 = const()[name = tensor("op_4965_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_4965_end_0 = const()[name = tensor("op_4965_end_0"), val = tensor([1, 1, 16, 32, 1])]; + tensor var_4965_end_mask_0 = const()[name = tensor("op_4965_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_4965_squeeze_mask_0 = const()[name = tensor("op_4965_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_4965_cast_fp16 = slice_by_index(begin = var_4965_begin_0, end = var_4965_end_0, end_mask = var_4965_end_mask_0, squeeze_mask = var_4965_squeeze_mask_0, x = q_complex_23_cast_fp16)[name = tensor("op_4965_cast_fp16")]; + tensor var_4973_begin_0 = const()[name = tensor("op_4973_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_4973_end_0 = const()[name = tensor("op_4973_end_0"), val = tensor([1, 1, 16, 32, 2])]; + tensor var_4973_end_mask_0 = const()[name = tensor("op_4973_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_4973_squeeze_mask_0 = const()[name = tensor("op_4973_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_4973_cast_fp16 = slice_by_index(begin = var_4973_begin_0, end = var_4973_end_0, end_mask = var_4973_end_mask_0, squeeze_mask = var_4973_squeeze_mask_0, x = q_complex_23_cast_fp16)[name = tensor("op_4973_cast_fp16")]; + tensor var_4981_begin_0 = const()[name = tensor("op_4981_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_4981_end_0 = const()[name = tensor("op_4981_end_0"), val = tensor([1, 1, 16, 32, 1])]; + tensor var_4981_end_mask_0 = const()[name = tensor("op_4981_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_4981_squeeze_mask_0 = const()[name = tensor("op_4981_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_4981_cast_fp16 = slice_by_index(begin = var_4981_begin_0, end = var_4981_end_0, end_mask = var_4981_end_mask_0, squeeze_mask = var_4981_squeeze_mask_0, x = k_complex_23_cast_fp16)[name = tensor("op_4981_cast_fp16")]; + tensor var_4989_begin_0 = const()[name = tensor("op_4989_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_4989_end_0 = const()[name = tensor("op_4989_end_0"), val = tensor([1, 1, 16, 32, 2])]; + tensor var_4989_end_mask_0 = const()[name = tensor("op_4989_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_4989_squeeze_mask_0 = const()[name = tensor("op_4989_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_4989_cast_fp16 = slice_by_index(begin = var_4989_begin_0, end = var_4989_end_0, end_mask = var_4989_end_mask_0, squeeze_mask = var_4989_squeeze_mask_0, x = k_complex_23_cast_fp16)[name = tensor("op_4989_cast_fp16")]; + tensor freqs_23_to_fp16 = const()[name = tensor("freqs_23_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(303683456)))]; + tensor var_4995_cast_fp16 = mul(x = freqs_23_to_fp16, y = ts_71_cast_fp16)[name = tensor("op_4995_cast_fp16")]; + tensor rotr_23_cast_fp16 = cos(x = var_4995_cast_fp16)[name = tensor("rotr_23_cast_fp16")]; + tensor roti_23_cast_fp16 = sin(x = var_4995_cast_fp16)[name = tensor("roti_23_cast_fp16")]; + tensor var_4999_cast_fp16 = mul(x = var_4965_cast_fp16, y = rotr_23_cast_fp16)[name = tensor("op_4999_cast_fp16")]; + tensor var_5000_cast_fp16 = mul(x = var_4973_cast_fp16, y = roti_23_cast_fp16)[name = tensor("op_5000_cast_fp16")]; + tensor qor_45_cast_fp16 = sub(x = var_4999_cast_fp16, y = var_5000_cast_fp16)[name = tensor("qor_45_cast_fp16")]; + tensor var_5003_cast_fp16 = mul(x = var_4965_cast_fp16, y = roti_23_cast_fp16)[name = tensor("op_5003_cast_fp16")]; + tensor var_5004_cast_fp16 = mul(x = var_4973_cast_fp16, y = rotr_23_cast_fp16)[name = tensor("op_5004_cast_fp16")]; + tensor qoi_45_cast_fp16 = add(x = var_5003_cast_fp16, y = var_5004_cast_fp16)[name = tensor("qoi_45_cast_fp16")]; + tensor var_5007_cast_fp16 = mul(x = var_4981_cast_fp16, y = rotr_23_cast_fp16)[name = tensor("op_5007_cast_fp16")]; + tensor var_5008_cast_fp16 = mul(x = var_4989_cast_fp16, y = roti_23_cast_fp16)[name = tensor("op_5008_cast_fp16")]; + tensor kor_45_cast_fp16 = sub(x = var_5007_cast_fp16, y = var_5008_cast_fp16)[name = tensor("kor_45_cast_fp16")]; + tensor var_5011_cast_fp16 = mul(x = var_4981_cast_fp16, y = roti_23_cast_fp16)[name = tensor("op_5011_cast_fp16")]; + tensor var_5012_cast_fp16 = mul(x = var_4989_cast_fp16, y = rotr_23_cast_fp16)[name = tensor("op_5012_cast_fp16")]; + tensor koi_45_cast_fp16 = add(x = var_5011_cast_fp16, y = var_5012_cast_fp16)[name = tensor("koi_45_cast_fp16")]; + tensor qo_23_axis_0 = const()[name = tensor("qo_23_axis_0"), val = tensor(-1)]; + tensor qo_23_cast_fp16 = stack(axis = qo_23_axis_0, values = (qor_45_cast_fp16, qoi_45_cast_fp16))[name = tensor("qo_23_cast_fp16")]; + tensor ko_23_axis_0 = const()[name = tensor("ko_23_axis_0"), val = tensor(-1)]; + tensor ko_23_cast_fp16 = stack(axis = ko_23_axis_0, values = (kor_45_cast_fp16, koi_45_cast_fp16))[name = tensor("ko_23_cast_fp16")]; + tensor var_5041 = const()[name = tensor("op_5041"), val = tensor([1, 1, 16, 64])]; + tensor q_69_cast_fp16 = reshape(shape = var_5041, x = qo_23_cast_fp16)[name = tensor("q_69_cast_fp16")]; + tensor var_5043 = const()[name = tensor("op_5043"), val = tensor([1, 1, 16, 64])]; + tensor k_47_cast_fp16 = reshape(shape = var_5043, x = ko_23_cast_fp16)[name = tensor("k_47_cast_fp16")]; + tensor _inversed_5065_y_0_to_fp16 = const()[name = tensor("_inversed_5065_y_0_to_fp16"), val = tensor(0x1p-9)]; + tensor _inversed_5065_cast_fp16 = mul(x = ts_71_cast_fp16, y = _inversed_5065_y_0_to_fp16)[name = tensor("_inversed_5065_cast_fp16")]; + tensor var_5066_cast_fp16 = floor(x = _inversed_5065_cast_fp16)[name = tensor("op_5066_cast_fp16")]; + tensor var_5067_to_fp16 = const()[name = tensor("op_5067_to_fp16"), val = tensor(0x1p+9)]; + tensor var_5068_cast_fp16 = mul(x = var_5066_cast_fp16, y = var_5067_to_fp16)[name = tensor("op_5068_cast_fp16")]; + tensor write_indices_float_47_cast_fp16 = sub(x = ts_71_cast_fp16, y = var_5068_cast_fp16)[name = tensor("write_indices_float_47_cast_fp16")]; + tensor var_5075_dtype_0 = const()[name = tensor("op_5075_dtype_0"), val = tensor("int32")]; + tensor write_indices_23_reps_0 = const()[name = tensor("write_indices_23_reps_0"), val = tensor([1, 1, 16, 64])]; + tensor write_indices_float_47_cast_fp16_to_int32 = cast(dtype = var_5075_dtype_0, x = write_indices_float_47_cast_fp16)[name = tensor("cast_497")]; + tensor write_indices_23 = tile(reps = write_indices_23_reps_0, x = write_indices_float_47_cast_fp16_to_int32)[name = tensor("write_indices_23")]; + tensor var_5083_begin_0 = const()[name = tensor("op_5083_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_5083_end_0 = const()[name = tensor("op_5083_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor var_5083_end_mask_0 = const()[name = tensor("op_5083_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_5083_squeeze_mask_0 = const()[name = tensor("op_5083_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor cache11_to_fp16_dtype_0 = const()[name = tensor("cache11_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor cache11_to_fp16 = cast(dtype = cache11_to_fp16_dtype_0, x = cache11)[name = tensor("cast_496")]; + tensor var_5083_cast_fp16 = slice_by_index(begin = var_5083_begin_0, end = var_5083_end_0, end_mask = var_5083_end_mask_0, squeeze_mask = var_5083_squeeze_mask_0, x = cache11_to_fp16)[name = tensor("op_5083_cast_fp16")]; + tensor var_5085_axis_0 = const()[name = tensor("op_5085_axis_0"), val = tensor(1)]; + tensor var_5085_mode_0 = const()[name = tensor("op_5085_mode_0"), val = tensor("update")]; + tensor var_5085_validate_indices_0 = const()[name = tensor("op_5085_validate_indices_0"), val = tensor(false)]; + tensor var_5085_cast_fp16 = scatter_along_axis(axis = var_5085_axis_0, data = var_5083_cast_fp16, indices = write_indices_23, mode = var_5085_mode_0, updates = k_47_cast_fp16, validate_indices = var_5085_validate_indices_0)[name = tensor("op_5085_cast_fp16")]; + tensor concat_79 = const()[name = tensor("concat_79"), val = tensor([0, 0, 0, 0, 0])]; + tensor concat_80 = const()[name = tensor("concat_80"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_23_internal_tensor_assign_1_stride_0 = const()[name = tensor("new_cache_23_internal_tensor_assign_1_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_23_internal_tensor_assign_1_begin_mask_0 = const()[name = tensor("new_cache_23_internal_tensor_assign_1_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_23_internal_tensor_assign_1_end_mask_0 = const()[name = tensor("new_cache_23_internal_tensor_assign_1_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_23_internal_tensor_assign_1_squeeze_mask_0 = const()[name = tensor("new_cache_23_internal_tensor_assign_1_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_70 = const()[name = tensor("shape_70"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_22 = const()[name = tensor("reduce_prod_22"), val = tensor(1048576)]; + tensor range_1d_22_start_0 = const()[name = tensor("range_1d_22_start_0"), val = tensor(0)]; + tensor range_1d_22_step_0 = const()[name = tensor("range_1d_22_step_0"), val = tensor(1)]; + tensor range_1d_22 = range_1d(end = reduce_prod_22, start = range_1d_22_start_0, step = range_1d_22_step_0)[name = tensor("range_1d_22")]; + tensor reshape_110 = reshape(shape = shape_70, x = range_1d_22)[name = tensor("reshape_110")]; + tensor slice_by_index_22 = slice_by_index(begin = concat_79, begin_mask = new_cache_23_internal_tensor_assign_1_begin_mask_0, end = concat_80, end_mask = new_cache_23_internal_tensor_assign_1_end_mask_0, squeeze_mask = new_cache_23_internal_tensor_assign_1_squeeze_mask_0, stride = new_cache_23_internal_tensor_assign_1_stride_0, x = reshape_110)[name = tensor("slice_by_index_22")]; + tensor reshape_111_shape_0 = const()[name = tensor("reshape_111_shape_0"), val = tensor([-1])]; + tensor reshape_111 = reshape(shape = reshape_111_shape_0, x = slice_by_index_22)[name = tensor("reshape_111")]; + tensor reshape_112_shape_0 = const()[name = tensor("reshape_112_shape_0"), val = tensor([-1])]; + tensor reshape_112_cast_fp16 = reshape(shape = reshape_112_shape_0, x = var_5085_cast_fp16)[name = tensor("reshape_112_cast_fp16")]; + tensor reshape_113_shape_0 = const()[name = tensor("reshape_113_shape_0"), val = tensor([-1])]; + tensor reshape_113_cast_fp16 = reshape(shape = reshape_113_shape_0, x = cache11_to_fp16)[name = tensor("reshape_113_cast_fp16")]; + tensor scatter_22_mode_0 = const()[name = tensor("scatter_22_mode_0"), val = tensor("update")]; + tensor scatter_22_axis_0 = const()[name = tensor("scatter_22_axis_0"), val = tensor(0)]; + tensor scatter_22_validate_indices_0 = const()[name = tensor("scatter_22_validate_indices_0"), val = tensor(false)]; + tensor scatter_22_cast_fp16 = scatter(axis = scatter_22_axis_0, data = reshape_113_cast_fp16, indices = reshape_111, mode = scatter_22_mode_0, updates = reshape_112_cast_fp16, validate_indices = scatter_22_validate_indices_0)[name = tensor("scatter_22_cast_fp16")]; + tensor reshape_114_cast_fp16 = reshape(shape = shape_70, x = scatter_22_cast_fp16)[name = tensor("reshape_114_cast_fp16")]; + tensor var_5093_begin_0 = const()[name = tensor("op_5093_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor var_5093_end_0 = const()[name = tensor("op_5093_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor var_5093_end_mask_0 = const()[name = tensor("op_5093_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_5093_squeeze_mask_0 = const()[name = tensor("op_5093_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor var_5093_cast_fp16 = slice_by_index(begin = var_5093_begin_0, end = var_5093_end_0, end_mask = var_5093_end_mask_0, squeeze_mask = var_5093_squeeze_mask_0, x = reshape_114_cast_fp16)[name = tensor("op_5093_cast_fp16")]; + tensor var_5095_axis_0 = const()[name = tensor("op_5095_axis_0"), val = tensor(1)]; + tensor var_5095_mode_0 = const()[name = tensor("op_5095_mode_0"), val = tensor("update")]; + tensor var_5095_validate_indices_0 = const()[name = tensor("op_5095_validate_indices_0"), val = tensor(false)]; + tensor var_5095_cast_fp16 = scatter_along_axis(axis = var_5095_axis_0, data = var_5093_cast_fp16, indices = write_indices_23, mode = var_5095_mode_0, updates = v_23_cast_fp16, validate_indices = var_5095_validate_indices_0)[name = tensor("op_5095_cast_fp16")]; + tensor concat_81 = const()[name = tensor("concat_81"), val = tensor([1, 0, 0, 0, 0])]; + tensor concat_82 = const()[name = tensor("concat_82"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_23_internal_tensor_assign_2_stride_0 = const()[name = tensor("new_cache_23_internal_tensor_assign_2_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_23_internal_tensor_assign_2_begin_mask_0 = const()[name = tensor("new_cache_23_internal_tensor_assign_2_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_23_internal_tensor_assign_2_end_mask_0 = const()[name = tensor("new_cache_23_internal_tensor_assign_2_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_23_internal_tensor_assign_2_squeeze_mask_0 = const()[name = tensor("new_cache_23_internal_tensor_assign_2_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_71 = const()[name = tensor("shape_71"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_23 = const()[name = tensor("reduce_prod_23"), val = tensor(1048576)]; + tensor range_1d_23_start_0 = const()[name = tensor("range_1d_23_start_0"), val = tensor(0)]; + tensor range_1d_23_step_0 = const()[name = tensor("range_1d_23_step_0"), val = tensor(1)]; + tensor range_1d_23 = range_1d(end = reduce_prod_23, start = range_1d_23_start_0, step = range_1d_23_step_0)[name = tensor("range_1d_23")]; + tensor reshape_115 = reshape(shape = shape_71, x = range_1d_23)[name = tensor("reshape_115")]; + tensor slice_by_index_23 = slice_by_index(begin = concat_81, begin_mask = new_cache_23_internal_tensor_assign_2_begin_mask_0, end = concat_82, end_mask = new_cache_23_internal_tensor_assign_2_end_mask_0, squeeze_mask = new_cache_23_internal_tensor_assign_2_squeeze_mask_0, stride = new_cache_23_internal_tensor_assign_2_stride_0, x = reshape_115)[name = tensor("slice_by_index_23")]; + tensor reshape_116_shape_0 = const()[name = tensor("reshape_116_shape_0"), val = tensor([-1])]; + tensor reshape_116 = reshape(shape = reshape_116_shape_0, x = slice_by_index_23)[name = tensor("reshape_116")]; + tensor reshape_117_shape_0 = const()[name = tensor("reshape_117_shape_0"), val = tensor([-1])]; + tensor reshape_117_cast_fp16 = reshape(shape = reshape_117_shape_0, x = var_5095_cast_fp16)[name = tensor("reshape_117_cast_fp16")]; + tensor reshape_118_shape_0 = const()[name = tensor("reshape_118_shape_0"), val = tensor([-1])]; + tensor reshape_118_cast_fp16 = reshape(shape = reshape_118_shape_0, x = reshape_114_cast_fp16)[name = tensor("reshape_118_cast_fp16")]; + tensor scatter_23_mode_0 = const()[name = tensor("scatter_23_mode_0"), val = tensor("update")]; + tensor scatter_23_axis_0 = const()[name = tensor("scatter_23_axis_0"), val = tensor(0)]; + tensor scatter_23_validate_indices_0 = const()[name = tensor("scatter_23_validate_indices_0"), val = tensor(false)]; + tensor scatter_23_cast_fp16 = scatter(axis = scatter_23_axis_0, data = reshape_118_cast_fp16, indices = reshape_116, mode = scatter_23_mode_0, updates = reshape_117_cast_fp16, validate_indices = scatter_23_validate_indices_0)[name = tensor("scatter_23_cast_fp16")]; + tensor reshape_119_cast_fp16 = reshape(shape = shape_71, x = scatter_23_cast_fp16)[name = tensor("reshape_119_cast_fp16")]; + tensor reshape_119_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("reshape_119_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor keys_67_begin_0 = const()[name = tensor("keys_67_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor keys_67_end_0 = const()[name = tensor("keys_67_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor keys_67_end_mask_0 = const()[name = tensor("keys_67_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor keys_67_squeeze_mask_0 = const()[name = tensor("keys_67_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor keys_67_cast_fp16 = slice_by_index(begin = keys_67_begin_0, end = keys_67_end_0, end_mask = keys_67_end_mask_0, squeeze_mask = keys_67_squeeze_mask_0, x = reshape_119_cast_fp16)[name = tensor("keys_67_cast_fp16")]; + tensor values_67_begin_0 = const()[name = tensor("values_67_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor values_67_end_0 = const()[name = tensor("values_67_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor values_67_end_mask_0 = const()[name = tensor("values_67_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor values_67_squeeze_mask_0 = const()[name = tensor("values_67_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor values_67_cast_fp16 = slice_by_index(begin = values_67_begin_0, end = values_67_end_0, end_mask = values_67_end_mask_0, squeeze_mask = values_67_squeeze_mask_0, x = reshape_119_cast_fp16)[name = tensor("values_67_cast_fp16")]; + tensor var_5107_cast_fp16 = not_equal(x = keys_67_cast_fp16, y = keys_67_cast_fp16)[name = tensor("op_5107_cast_fp16")]; + tensor keys_69_cast_fp16 = select(a = var_504_to_fp16, b = keys_67_cast_fp16, cond = var_5107_cast_fp16)[name = tensor("keys_69_cast_fp16")]; + tensor var_5115_cast_fp16 = not_equal(x = values_67_cast_fp16, y = values_67_cast_fp16)[name = tensor("op_5115_cast_fp16")]; + tensor values_69_cast_fp16 = select(a = var_504_to_fp16, b = values_67_cast_fp16, cond = var_5115_cast_fp16)[name = tensor("values_69_cast_fp16")]; + tensor var_5139 = const()[name = tensor("op_5139"), val = tensor([0, 2, 1, 3])]; + tensor var_5152 = const()[name = tensor("op_5152"), val = tensor([1, 1, 1])]; + tensor var_5153_cast_fp16 = reshape(shape = var_5152, x = position11_to_fp16)[name = tensor("op_5153_cast_fp16")]; + tensor var_5170_to_fp16 = const()[name = tensor("op_5170_to_fp16"), val = tensor(0x1p+0)]; + tensor valid_len_23_cast_fp16 = add(x = var_5153_cast_fp16, y = var_5170_to_fp16)[name = tensor("valid_len_23_cast_fp16")]; + tensor valid_mask_23_cast_fp16 = less(x = k_positions_1_promoted_to_fp16, y = valid_len_23_cast_fp16)[name = tensor("valid_mask_23_cast_fp16")]; + tensor causal_mask_23_cast_fp16 = less_equal(x = k_positions_1_promoted_to_fp16, y = var_5153_cast_fp16)[name = tensor("causal_mask_23_cast_fp16")]; + tensor attn_mask_45 = logical_and(x = valid_mask_23_cast_fp16, y = causal_mask_23_cast_fp16)[name = tensor("attn_mask_45")]; + tensor attn_mask_47_axes_0 = const()[name = tensor("attn_mask_47_axes_0"), val = tensor([1])]; + tensor attn_mask_47 = expand_dims(axes = attn_mask_47_axes_0, x = attn_mask_45)[name = tensor("attn_mask_47")]; + tensor var_5188_transpose_x_0 = const()[name = tensor("op_5188_transpose_x_0"), val = tensor(false)]; + tensor var_5188_transpose_y_0 = const()[name = tensor("op_5188_transpose_y_0"), val = tensor(false)]; + tensor transpose_94_perm_0 = const()[name = tensor("transpose_94_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_95_perm_0 = const()[name = tensor("transpose_95_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_95 = transpose(perm = transpose_95_perm_0, x = keys_69_cast_fp16)[name = tensor("transpose_169")]; + tensor transpose_94 = transpose(perm = transpose_94_perm_0, x = q_69_cast_fp16)[name = tensor("transpose_170")]; + tensor var_5188_cast_fp16 = matmul(transpose_x = var_5188_transpose_x_0, transpose_y = var_5188_transpose_y_0, x = transpose_94, y = transpose_95)[name = tensor("op_5188_cast_fp16")]; + tensor var_5182_to_fp16 = const()[name = tensor("op_5182_to_fp16"), val = tensor([0x1p-3])]; + tensor attn_weights_67_cast_fp16 = mul(x = var_5188_cast_fp16, y = var_5182_to_fp16)[name = tensor("attn_weights_67_cast_fp16")]; + tensor var_5190 = logical_not(x = attn_mask_47)[name = tensor("op_5190")]; + tensor var_5191_to_fp16 = const()[name = tensor("op_5191_to_fp16"), val = tensor(-inf)]; + tensor attn_weights_69_cast_fp16 = select(a = var_5191_to_fp16, b = attn_weights_67_cast_fp16, cond = var_5190)[name = tensor("attn_weights_69_cast_fp16")]; + tensor var_5193 = const()[name = tensor("op_5193"), val = tensor(-1)]; + tensor attn_weights_71_cast_fp16 = softmax(axis = var_5193, x = attn_weights_69_cast_fp16)[name = tensor("attn_weights_71_cast_fp16")]; + tensor attn_output_23_transpose_x_0 = const()[name = tensor("attn_output_23_transpose_x_0"), val = tensor(false)]; + tensor attn_output_23_transpose_y_0 = const()[name = tensor("attn_output_23_transpose_y_0"), val = tensor(false)]; + tensor values_71_cast_fp16 = transpose(perm = var_5139, x = values_69_cast_fp16)[name = tensor("transpose_171")]; + tensor attn_output_23_cast_fp16 = matmul(transpose_x = attn_output_23_transpose_x_0, transpose_y = attn_output_23_transpose_y_0, x = attn_weights_71_cast_fp16, y = values_71_cast_fp16)[name = tensor("attn_output_23_cast_fp16")]; + tensor var_5201 = const()[name = tensor("op_5201"), val = tensor([0, 2, 1, 3])]; + tensor var_5204 = const()[name = tensor("op_5204"), val = tensor([1, 1, 1024])]; + tensor var_5202_cast_fp16 = transpose(perm = var_5201, x = attn_output_23_cast_fp16)[name = tensor("transpose_168")]; + tensor input_115_cast_fp16 = reshape(shape = var_5204, x = var_5202_cast_fp16)[name = tensor("input_115_cast_fp16")]; + tensor linear_46_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = attn11_out_proj_weight_cast_fp16, x = input_115_cast_fp16)[name = tensor("linear_46_cast_fp16")]; + tensor var_5210_to_fp16 = const()[name = tensor("op_5210_to_fp16"), val = tensor(0x1p+0)]; + tensor var_5211_cast_fp16 = add(x = position11_to_fp16, y = var_5210_to_fp16)[name = tensor("op_5211_cast_fp16")]; + tensor var_5211_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_5211_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor input_117_cast_fp16 = add(x = input_113_cast_fp16, y = linear_46_cast_fp16)[name = tensor("input_117_cast_fp16")]; + tensor input_119_axes_0 = const()[name = tensor("input_119_axes_0"), val = tensor([-1])]; + tensor norm11_2_weight_to_fp16 = const()[name = tensor("norm11_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(303683584)))]; + tensor norm11_2_bias_to_fp16 = const()[name = tensor("norm11_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(303685696)))]; + tensor var_5215_to_fp16 = const()[name = tensor("op_5215_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_119_cast_fp16 = layer_norm(axes = input_119_axes_0, beta = norm11_2_bias_to_fp16, epsilon = var_5215_to_fp16, gamma = norm11_2_weight_to_fp16, x = input_117_cast_fp16)[name = tensor("input_119_cast_fp16")]; + tensor linear_47_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = linear11_1_weight_cast_fp16, x = input_119_cast_fp16)[name = tensor("linear_47_cast_fp16")]; + tensor input_121_mode_0 = const()[name = tensor("input_121_mode_0"), val = tensor("EXACT")]; + tensor input_121_cast_fp16 = gelu(mode = input_121_mode_0, x = linear_47_cast_fp16)[name = tensor("input_121_cast_fp16")]; + tensor linear_48_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = linear11_2_weight_cast_fp16, x = input_121_cast_fp16)[name = tensor("linear_48_cast_fp16")]; + tensor input_123_cast_fp16 = add(x = input_117_cast_fp16, y = linear_48_cast_fp16)[name = tensor("input_123_cast_fp16")]; + tensor x_25_axes_0 = const()[name = tensor("x_25_axes_0"), val = tensor([-1])]; + tensor norm12_1_weight_to_fp16 = const()[name = tensor("norm12_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(303687808)))]; + tensor norm12_1_bias_to_fp16 = const()[name = tensor("norm12_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(303689920)))]; + tensor var_5232_to_fp16 = const()[name = tensor("op_5232_to_fp16"), val = tensor(0x1.5p-17)]; + tensor x_25_cast_fp16 = layer_norm(axes = x_25_axes_0, beta = norm12_1_bias_to_fp16, epsilon = var_5232_to_fp16, gamma = norm12_1_weight_to_fp16, x = input_123_cast_fp16)[name = tensor("x_25_cast_fp16")]; + tensor linear_49_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = attn12_in_proj_weight_cast_fp16, x = x_25_cast_fp16)[name = tensor("linear_49_cast_fp16")]; + tensor var_5268 = const()[name = tensor("op_5268"), val = tensor([1, 1, 3, 16, 64])]; + tensor qkv_25_cast_fp16 = reshape(shape = var_5268, x = linear_49_cast_fp16)[name = tensor("qkv_25_cast_fp16")]; + tensor q_73_begin_0 = const()[name = tensor("q_73_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor q_73_end_0 = const()[name = tensor("q_73_end_0"), val = tensor([1, 1, 1, 16, 64])]; + tensor q_73_end_mask_0 = const()[name = tensor("q_73_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor q_73_squeeze_mask_0 = const()[name = tensor("q_73_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor q_73_cast_fp16 = slice_by_index(begin = q_73_begin_0, end = q_73_end_0, end_mask = q_73_end_mask_0, squeeze_mask = q_73_squeeze_mask_0, x = qkv_25_cast_fp16)[name = tensor("q_73_cast_fp16")]; + tensor k_49_begin_0 = const()[name = tensor("k_49_begin_0"), val = tensor([0, 0, 1, 0, 0])]; + tensor k_49_end_0 = const()[name = tensor("k_49_end_0"), val = tensor([1, 1, 2, 16, 64])]; + tensor k_49_end_mask_0 = const()[name = tensor("k_49_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor k_49_squeeze_mask_0 = const()[name = tensor("k_49_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor k_49_cast_fp16 = slice_by_index(begin = k_49_begin_0, end = k_49_end_0, end_mask = k_49_end_mask_0, squeeze_mask = k_49_squeeze_mask_0, x = qkv_25_cast_fp16)[name = tensor("k_49_cast_fp16")]; + tensor v_25_begin_0 = const()[name = tensor("v_25_begin_0"), val = tensor([0, 0, 2, 0, 0])]; + tensor v_25_end_0 = const()[name = tensor("v_25_end_0"), val = tensor([1, 1, 3, 16, 64])]; + tensor v_25_end_mask_0 = const()[name = tensor("v_25_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor v_25_squeeze_mask_0 = const()[name = tensor("v_25_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor v_25_cast_fp16 = slice_by_index(begin = v_25_begin_0, end = v_25_end_0, end_mask = v_25_end_mask_0, squeeze_mask = v_25_squeeze_mask_0, x = qkv_25_cast_fp16)[name = tensor("v_25_cast_fp16")]; + tensor var_5372 = const()[name = tensor("op_5372"), val = tensor([1, 1, 1, 1])]; + tensor position12_to_fp16_dtype_0 = const()[name = tensor("position12_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor position12_to_fp16 = cast(dtype = position12_to_fp16_dtype_0, x = position12)[name = tensor("cast_493")]; + tensor ts_77_cast_fp16 = reshape(shape = var_5372, x = position12_to_fp16)[name = tensor("ts_77_cast_fp16")]; + tensor var_5376 = const()[name = tensor("op_5376"), val = tensor([1, 1, 16, 32, 2])]; + tensor q_complex_25_cast_fp16 = reshape(shape = var_5376, x = q_73_cast_fp16)[name = tensor("q_complex_25_cast_fp16")]; + tensor var_5380 = const()[name = tensor("op_5380"), val = tensor([1, 1, 16, 32, 2])]; + tensor k_complex_25_cast_fp16 = reshape(shape = var_5380, x = k_49_cast_fp16)[name = tensor("k_complex_25_cast_fp16")]; + tensor var_5384_begin_0 = const()[name = tensor("op_5384_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_5384_end_0 = const()[name = tensor("op_5384_end_0"), val = tensor([1, 1, 16, 32, 1])]; + tensor var_5384_end_mask_0 = const()[name = tensor("op_5384_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_5384_squeeze_mask_0 = const()[name = tensor("op_5384_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_5384_cast_fp16 = slice_by_index(begin = var_5384_begin_0, end = var_5384_end_0, end_mask = var_5384_end_mask_0, squeeze_mask = var_5384_squeeze_mask_0, x = q_complex_25_cast_fp16)[name = tensor("op_5384_cast_fp16")]; + tensor var_5392_begin_0 = const()[name = tensor("op_5392_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_5392_end_0 = const()[name = tensor("op_5392_end_0"), val = tensor([1, 1, 16, 32, 2])]; + tensor var_5392_end_mask_0 = const()[name = tensor("op_5392_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_5392_squeeze_mask_0 = const()[name = tensor("op_5392_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_5392_cast_fp16 = slice_by_index(begin = var_5392_begin_0, end = var_5392_end_0, end_mask = var_5392_end_mask_0, squeeze_mask = var_5392_squeeze_mask_0, x = q_complex_25_cast_fp16)[name = tensor("op_5392_cast_fp16")]; + tensor var_5400_begin_0 = const()[name = tensor("op_5400_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_5400_end_0 = const()[name = tensor("op_5400_end_0"), val = tensor([1, 1, 16, 32, 1])]; + tensor var_5400_end_mask_0 = const()[name = tensor("op_5400_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_5400_squeeze_mask_0 = const()[name = tensor("op_5400_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_5400_cast_fp16 = slice_by_index(begin = var_5400_begin_0, end = var_5400_end_0, end_mask = var_5400_end_mask_0, squeeze_mask = var_5400_squeeze_mask_0, x = k_complex_25_cast_fp16)[name = tensor("op_5400_cast_fp16")]; + tensor var_5408_begin_0 = const()[name = tensor("op_5408_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_5408_end_0 = const()[name = tensor("op_5408_end_0"), val = tensor([1, 1, 16, 32, 2])]; + tensor var_5408_end_mask_0 = const()[name = tensor("op_5408_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_5408_squeeze_mask_0 = const()[name = tensor("op_5408_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_5408_cast_fp16 = slice_by_index(begin = var_5408_begin_0, end = var_5408_end_0, end_mask = var_5408_end_mask_0, squeeze_mask = var_5408_squeeze_mask_0, x = k_complex_25_cast_fp16)[name = tensor("op_5408_cast_fp16")]; + tensor freqs_25_to_fp16 = const()[name = tensor("freqs_25_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(303692032)))]; + tensor var_5414_cast_fp16 = mul(x = freqs_25_to_fp16, y = ts_77_cast_fp16)[name = tensor("op_5414_cast_fp16")]; + tensor rotr_25_cast_fp16 = cos(x = var_5414_cast_fp16)[name = tensor("rotr_25_cast_fp16")]; + tensor roti_25_cast_fp16 = sin(x = var_5414_cast_fp16)[name = tensor("roti_25_cast_fp16")]; + tensor var_5418_cast_fp16 = mul(x = var_5384_cast_fp16, y = rotr_25_cast_fp16)[name = tensor("op_5418_cast_fp16")]; + tensor var_5419_cast_fp16 = mul(x = var_5392_cast_fp16, y = roti_25_cast_fp16)[name = tensor("op_5419_cast_fp16")]; + tensor qor_49_cast_fp16 = sub(x = var_5418_cast_fp16, y = var_5419_cast_fp16)[name = tensor("qor_49_cast_fp16")]; + tensor var_5422_cast_fp16 = mul(x = var_5384_cast_fp16, y = roti_25_cast_fp16)[name = tensor("op_5422_cast_fp16")]; + tensor var_5423_cast_fp16 = mul(x = var_5392_cast_fp16, y = rotr_25_cast_fp16)[name = tensor("op_5423_cast_fp16")]; + tensor qoi_49_cast_fp16 = add(x = var_5422_cast_fp16, y = var_5423_cast_fp16)[name = tensor("qoi_49_cast_fp16")]; + tensor var_5426_cast_fp16 = mul(x = var_5400_cast_fp16, y = rotr_25_cast_fp16)[name = tensor("op_5426_cast_fp16")]; + tensor var_5427_cast_fp16 = mul(x = var_5408_cast_fp16, y = roti_25_cast_fp16)[name = tensor("op_5427_cast_fp16")]; + tensor kor_49_cast_fp16 = sub(x = var_5426_cast_fp16, y = var_5427_cast_fp16)[name = tensor("kor_49_cast_fp16")]; + tensor var_5430_cast_fp16 = mul(x = var_5400_cast_fp16, y = roti_25_cast_fp16)[name = tensor("op_5430_cast_fp16")]; + tensor var_5431_cast_fp16 = mul(x = var_5408_cast_fp16, y = rotr_25_cast_fp16)[name = tensor("op_5431_cast_fp16")]; + tensor koi_49_cast_fp16 = add(x = var_5430_cast_fp16, y = var_5431_cast_fp16)[name = tensor("koi_49_cast_fp16")]; + tensor qo_25_axis_0 = const()[name = tensor("qo_25_axis_0"), val = tensor(-1)]; + tensor qo_25_cast_fp16 = stack(axis = qo_25_axis_0, values = (qor_49_cast_fp16, qoi_49_cast_fp16))[name = tensor("qo_25_cast_fp16")]; + tensor ko_25_axis_0 = const()[name = tensor("ko_25_axis_0"), val = tensor(-1)]; + tensor ko_25_cast_fp16 = stack(axis = ko_25_axis_0, values = (kor_49_cast_fp16, koi_49_cast_fp16))[name = tensor("ko_25_cast_fp16")]; + tensor var_5460 = const()[name = tensor("op_5460"), val = tensor([1, 1, 16, 64])]; + tensor q_75_cast_fp16 = reshape(shape = var_5460, x = qo_25_cast_fp16)[name = tensor("q_75_cast_fp16")]; + tensor var_5462 = const()[name = tensor("op_5462"), val = tensor([1, 1, 16, 64])]; + tensor k_51_cast_fp16 = reshape(shape = var_5462, x = ko_25_cast_fp16)[name = tensor("k_51_cast_fp16")]; + tensor _inversed_5484_y_0_to_fp16 = const()[name = tensor("_inversed_5484_y_0_to_fp16"), val = tensor(0x1p-9)]; + tensor _inversed_5484_cast_fp16 = mul(x = ts_77_cast_fp16, y = _inversed_5484_y_0_to_fp16)[name = tensor("_inversed_5484_cast_fp16")]; + tensor var_5485_cast_fp16 = floor(x = _inversed_5484_cast_fp16)[name = tensor("op_5485_cast_fp16")]; + tensor var_5486_to_fp16 = const()[name = tensor("op_5486_to_fp16"), val = tensor(0x1p+9)]; + tensor var_5487_cast_fp16 = mul(x = var_5485_cast_fp16, y = var_5486_to_fp16)[name = tensor("op_5487_cast_fp16")]; + tensor write_indices_float_51_cast_fp16 = sub(x = ts_77_cast_fp16, y = var_5487_cast_fp16)[name = tensor("write_indices_float_51_cast_fp16")]; + tensor var_5494_dtype_0 = const()[name = tensor("op_5494_dtype_0"), val = tensor("int32")]; + tensor write_indices_25_reps_0 = const()[name = tensor("write_indices_25_reps_0"), val = tensor([1, 1, 16, 64])]; + tensor write_indices_float_51_cast_fp16_to_int32 = cast(dtype = var_5494_dtype_0, x = write_indices_float_51_cast_fp16)[name = tensor("cast_492")]; + tensor write_indices_25 = tile(reps = write_indices_25_reps_0, x = write_indices_float_51_cast_fp16_to_int32)[name = tensor("write_indices_25")]; + tensor var_5502_begin_0 = const()[name = tensor("op_5502_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_5502_end_0 = const()[name = tensor("op_5502_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor var_5502_end_mask_0 = const()[name = tensor("op_5502_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_5502_squeeze_mask_0 = const()[name = tensor("op_5502_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor cache12_to_fp16_dtype_0 = const()[name = tensor("cache12_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor cache12_to_fp16 = cast(dtype = cache12_to_fp16_dtype_0, x = cache12)[name = tensor("cast_491")]; + tensor var_5502_cast_fp16 = slice_by_index(begin = var_5502_begin_0, end = var_5502_end_0, end_mask = var_5502_end_mask_0, squeeze_mask = var_5502_squeeze_mask_0, x = cache12_to_fp16)[name = tensor("op_5502_cast_fp16")]; + tensor var_5504_axis_0 = const()[name = tensor("op_5504_axis_0"), val = tensor(1)]; + tensor var_5504_mode_0 = const()[name = tensor("op_5504_mode_0"), val = tensor("update")]; + tensor var_5504_validate_indices_0 = const()[name = tensor("op_5504_validate_indices_0"), val = tensor(false)]; + tensor var_5504_cast_fp16 = scatter_along_axis(axis = var_5504_axis_0, data = var_5502_cast_fp16, indices = write_indices_25, mode = var_5504_mode_0, updates = k_51_cast_fp16, validate_indices = var_5504_validate_indices_0)[name = tensor("op_5504_cast_fp16")]; + tensor concat_86 = const()[name = tensor("concat_86"), val = tensor([0, 0, 0, 0, 0])]; + tensor concat_87 = const()[name = tensor("concat_87"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_25_internal_tensor_assign_1_stride_0 = const()[name = tensor("new_cache_25_internal_tensor_assign_1_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_25_internal_tensor_assign_1_begin_mask_0 = const()[name = tensor("new_cache_25_internal_tensor_assign_1_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_25_internal_tensor_assign_1_end_mask_0 = const()[name = tensor("new_cache_25_internal_tensor_assign_1_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_25_internal_tensor_assign_1_squeeze_mask_0 = const()[name = tensor("new_cache_25_internal_tensor_assign_1_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_72 = const()[name = tensor("shape_72"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_24 = const()[name = tensor("reduce_prod_24"), val = tensor(1048576)]; + tensor range_1d_24_start_0 = const()[name = tensor("range_1d_24_start_0"), val = tensor(0)]; + tensor range_1d_24_step_0 = const()[name = tensor("range_1d_24_step_0"), val = tensor(1)]; + tensor range_1d_24 = range_1d(end = reduce_prod_24, start = range_1d_24_start_0, step = range_1d_24_step_0)[name = tensor("range_1d_24")]; + tensor reshape_120 = reshape(shape = shape_72, x = range_1d_24)[name = tensor("reshape_120")]; + tensor slice_by_index_24 = slice_by_index(begin = concat_86, begin_mask = new_cache_25_internal_tensor_assign_1_begin_mask_0, end = concat_87, end_mask = new_cache_25_internal_tensor_assign_1_end_mask_0, squeeze_mask = new_cache_25_internal_tensor_assign_1_squeeze_mask_0, stride = new_cache_25_internal_tensor_assign_1_stride_0, x = reshape_120)[name = tensor("slice_by_index_24")]; + tensor reshape_121_shape_0 = const()[name = tensor("reshape_121_shape_0"), val = tensor([-1])]; + tensor reshape_121 = reshape(shape = reshape_121_shape_0, x = slice_by_index_24)[name = tensor("reshape_121")]; + tensor reshape_122_shape_0 = const()[name = tensor("reshape_122_shape_0"), val = tensor([-1])]; + tensor reshape_122_cast_fp16 = reshape(shape = reshape_122_shape_0, x = var_5504_cast_fp16)[name = tensor("reshape_122_cast_fp16")]; + tensor reshape_123_shape_0 = const()[name = tensor("reshape_123_shape_0"), val = tensor([-1])]; + tensor reshape_123_cast_fp16 = reshape(shape = reshape_123_shape_0, x = cache12_to_fp16)[name = tensor("reshape_123_cast_fp16")]; + tensor scatter_24_mode_0 = const()[name = tensor("scatter_24_mode_0"), val = tensor("update")]; + tensor scatter_24_axis_0 = const()[name = tensor("scatter_24_axis_0"), val = tensor(0)]; + tensor scatter_24_validate_indices_0 = const()[name = tensor("scatter_24_validate_indices_0"), val = tensor(false)]; + tensor scatter_24_cast_fp16 = scatter(axis = scatter_24_axis_0, data = reshape_123_cast_fp16, indices = reshape_121, mode = scatter_24_mode_0, updates = reshape_122_cast_fp16, validate_indices = scatter_24_validate_indices_0)[name = tensor("scatter_24_cast_fp16")]; + tensor reshape_124_cast_fp16 = reshape(shape = shape_72, x = scatter_24_cast_fp16)[name = tensor("reshape_124_cast_fp16")]; + tensor var_5512_begin_0 = const()[name = tensor("op_5512_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor var_5512_end_0 = const()[name = tensor("op_5512_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor var_5512_end_mask_0 = const()[name = tensor("op_5512_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_5512_squeeze_mask_0 = const()[name = tensor("op_5512_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor var_5512_cast_fp16 = slice_by_index(begin = var_5512_begin_0, end = var_5512_end_0, end_mask = var_5512_end_mask_0, squeeze_mask = var_5512_squeeze_mask_0, x = reshape_124_cast_fp16)[name = tensor("op_5512_cast_fp16")]; + tensor var_5514_axis_0 = const()[name = tensor("op_5514_axis_0"), val = tensor(1)]; + tensor var_5514_mode_0 = const()[name = tensor("op_5514_mode_0"), val = tensor("update")]; + tensor var_5514_validate_indices_0 = const()[name = tensor("op_5514_validate_indices_0"), val = tensor(false)]; + tensor var_5514_cast_fp16 = scatter_along_axis(axis = var_5514_axis_0, data = var_5512_cast_fp16, indices = write_indices_25, mode = var_5514_mode_0, updates = v_25_cast_fp16, validate_indices = var_5514_validate_indices_0)[name = tensor("op_5514_cast_fp16")]; + tensor concat_88 = const()[name = tensor("concat_88"), val = tensor([1, 0, 0, 0, 0])]; + tensor concat_89 = const()[name = tensor("concat_89"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_25_internal_tensor_assign_2_stride_0 = const()[name = tensor("new_cache_25_internal_tensor_assign_2_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_25_internal_tensor_assign_2_begin_mask_0 = const()[name = tensor("new_cache_25_internal_tensor_assign_2_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_25_internal_tensor_assign_2_end_mask_0 = const()[name = tensor("new_cache_25_internal_tensor_assign_2_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_25_internal_tensor_assign_2_squeeze_mask_0 = const()[name = tensor("new_cache_25_internal_tensor_assign_2_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_73 = const()[name = tensor("shape_73"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_25 = const()[name = tensor("reduce_prod_25"), val = tensor(1048576)]; + tensor range_1d_25_start_0 = const()[name = tensor("range_1d_25_start_0"), val = tensor(0)]; + tensor range_1d_25_step_0 = const()[name = tensor("range_1d_25_step_0"), val = tensor(1)]; + tensor range_1d_25 = range_1d(end = reduce_prod_25, start = range_1d_25_start_0, step = range_1d_25_step_0)[name = tensor("range_1d_25")]; + tensor reshape_125 = reshape(shape = shape_73, x = range_1d_25)[name = tensor("reshape_125")]; + tensor slice_by_index_25 = slice_by_index(begin = concat_88, begin_mask = new_cache_25_internal_tensor_assign_2_begin_mask_0, end = concat_89, end_mask = new_cache_25_internal_tensor_assign_2_end_mask_0, squeeze_mask = new_cache_25_internal_tensor_assign_2_squeeze_mask_0, stride = new_cache_25_internal_tensor_assign_2_stride_0, x = reshape_125)[name = tensor("slice_by_index_25")]; + tensor reshape_126_shape_0 = const()[name = tensor("reshape_126_shape_0"), val = tensor([-1])]; + tensor reshape_126 = reshape(shape = reshape_126_shape_0, x = slice_by_index_25)[name = tensor("reshape_126")]; + tensor reshape_127_shape_0 = const()[name = tensor("reshape_127_shape_0"), val = tensor([-1])]; + tensor reshape_127_cast_fp16 = reshape(shape = reshape_127_shape_0, x = var_5514_cast_fp16)[name = tensor("reshape_127_cast_fp16")]; + tensor reshape_128_shape_0 = const()[name = tensor("reshape_128_shape_0"), val = tensor([-1])]; + tensor reshape_128_cast_fp16 = reshape(shape = reshape_128_shape_0, x = reshape_124_cast_fp16)[name = tensor("reshape_128_cast_fp16")]; + tensor scatter_25_mode_0 = const()[name = tensor("scatter_25_mode_0"), val = tensor("update")]; + tensor scatter_25_axis_0 = const()[name = tensor("scatter_25_axis_0"), val = tensor(0)]; + tensor scatter_25_validate_indices_0 = const()[name = tensor("scatter_25_validate_indices_0"), val = tensor(false)]; + tensor scatter_25_cast_fp16 = scatter(axis = scatter_25_axis_0, data = reshape_128_cast_fp16, indices = reshape_126, mode = scatter_25_mode_0, updates = reshape_127_cast_fp16, validate_indices = scatter_25_validate_indices_0)[name = tensor("scatter_25_cast_fp16")]; + tensor reshape_129_cast_fp16 = reshape(shape = shape_73, x = scatter_25_cast_fp16)[name = tensor("reshape_129_cast_fp16")]; + tensor reshape_129_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("reshape_129_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor keys_73_begin_0 = const()[name = tensor("keys_73_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor keys_73_end_0 = const()[name = tensor("keys_73_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor keys_73_end_mask_0 = const()[name = tensor("keys_73_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor keys_73_squeeze_mask_0 = const()[name = tensor("keys_73_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor keys_73_cast_fp16 = slice_by_index(begin = keys_73_begin_0, end = keys_73_end_0, end_mask = keys_73_end_mask_0, squeeze_mask = keys_73_squeeze_mask_0, x = reshape_129_cast_fp16)[name = tensor("keys_73_cast_fp16")]; + tensor values_73_begin_0 = const()[name = tensor("values_73_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor values_73_end_0 = const()[name = tensor("values_73_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor values_73_end_mask_0 = const()[name = tensor("values_73_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor values_73_squeeze_mask_0 = const()[name = tensor("values_73_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor values_73_cast_fp16 = slice_by_index(begin = values_73_begin_0, end = values_73_end_0, end_mask = values_73_end_mask_0, squeeze_mask = values_73_squeeze_mask_0, x = reshape_129_cast_fp16)[name = tensor("values_73_cast_fp16")]; + tensor var_5526_cast_fp16 = not_equal(x = keys_73_cast_fp16, y = keys_73_cast_fp16)[name = tensor("op_5526_cast_fp16")]; + tensor keys_75_cast_fp16 = select(a = var_504_to_fp16, b = keys_73_cast_fp16, cond = var_5526_cast_fp16)[name = tensor("keys_75_cast_fp16")]; + tensor var_5534_cast_fp16 = not_equal(x = values_73_cast_fp16, y = values_73_cast_fp16)[name = tensor("op_5534_cast_fp16")]; + tensor values_75_cast_fp16 = select(a = var_504_to_fp16, b = values_73_cast_fp16, cond = var_5534_cast_fp16)[name = tensor("values_75_cast_fp16")]; + tensor var_5558 = const()[name = tensor("op_5558"), val = tensor([0, 2, 1, 3])]; + tensor var_5571 = const()[name = tensor("op_5571"), val = tensor([1, 1, 1])]; + tensor var_5572_cast_fp16 = reshape(shape = var_5571, x = position12_to_fp16)[name = tensor("op_5572_cast_fp16")]; + tensor var_5589_to_fp16 = const()[name = tensor("op_5589_to_fp16"), val = tensor(0x1p+0)]; + tensor valid_len_25_cast_fp16 = add(x = var_5572_cast_fp16, y = var_5589_to_fp16)[name = tensor("valid_len_25_cast_fp16")]; + tensor valid_mask_25_cast_fp16 = less(x = k_positions_1_promoted_to_fp16, y = valid_len_25_cast_fp16)[name = tensor("valid_mask_25_cast_fp16")]; + tensor causal_mask_25_cast_fp16 = less_equal(x = k_positions_1_promoted_to_fp16, y = var_5572_cast_fp16)[name = tensor("causal_mask_25_cast_fp16")]; + tensor attn_mask_49 = logical_and(x = valid_mask_25_cast_fp16, y = causal_mask_25_cast_fp16)[name = tensor("attn_mask_49")]; + tensor attn_mask_51_axes_0 = const()[name = tensor("attn_mask_51_axes_0"), val = tensor([1])]; + tensor attn_mask_51 = expand_dims(axes = attn_mask_51_axes_0, x = attn_mask_49)[name = tensor("attn_mask_51")]; + tensor var_5607_transpose_x_0 = const()[name = tensor("op_5607_transpose_x_0"), val = tensor(false)]; + tensor var_5607_transpose_y_0 = const()[name = tensor("op_5607_transpose_y_0"), val = tensor(false)]; + tensor transpose_96_perm_0 = const()[name = tensor("transpose_96_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_97_perm_0 = const()[name = tensor("transpose_97_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_97 = transpose(perm = transpose_97_perm_0, x = keys_75_cast_fp16)[name = tensor("transpose_165")]; + tensor transpose_96 = transpose(perm = transpose_96_perm_0, x = q_75_cast_fp16)[name = tensor("transpose_166")]; + tensor var_5607_cast_fp16 = matmul(transpose_x = var_5607_transpose_x_0, transpose_y = var_5607_transpose_y_0, x = transpose_96, y = transpose_97)[name = tensor("op_5607_cast_fp16")]; + tensor var_5601_to_fp16 = const()[name = tensor("op_5601_to_fp16"), val = tensor([0x1p-3])]; + tensor attn_weights_73_cast_fp16 = mul(x = var_5607_cast_fp16, y = var_5601_to_fp16)[name = tensor("attn_weights_73_cast_fp16")]; + tensor var_5609 = logical_not(x = attn_mask_51)[name = tensor("op_5609")]; + tensor var_5610_to_fp16 = const()[name = tensor("op_5610_to_fp16"), val = tensor(-inf)]; + tensor attn_weights_75_cast_fp16 = select(a = var_5610_to_fp16, b = attn_weights_73_cast_fp16, cond = var_5609)[name = tensor("attn_weights_75_cast_fp16")]; + tensor var_5612 = const()[name = tensor("op_5612"), val = tensor(-1)]; + tensor attn_weights_77_cast_fp16 = softmax(axis = var_5612, x = attn_weights_75_cast_fp16)[name = tensor("attn_weights_77_cast_fp16")]; + tensor attn_output_25_transpose_x_0 = const()[name = tensor("attn_output_25_transpose_x_0"), val = tensor(false)]; + tensor attn_output_25_transpose_y_0 = const()[name = tensor("attn_output_25_transpose_y_0"), val = tensor(false)]; + tensor values_77_cast_fp16 = transpose(perm = var_5558, x = values_75_cast_fp16)[name = tensor("transpose_167")]; + tensor attn_output_25_cast_fp16 = matmul(transpose_x = attn_output_25_transpose_x_0, transpose_y = attn_output_25_transpose_y_0, x = attn_weights_77_cast_fp16, y = values_77_cast_fp16)[name = tensor("attn_output_25_cast_fp16")]; + tensor var_5620 = const()[name = tensor("op_5620"), val = tensor([0, 2, 1, 3])]; + tensor var_5623 = const()[name = tensor("op_5623"), val = tensor([1, 1, 1024])]; + tensor var_5621_cast_fp16 = transpose(perm = var_5620, x = attn_output_25_cast_fp16)[name = tensor("transpose_164")]; + tensor input_125_cast_fp16 = reshape(shape = var_5623, x = var_5621_cast_fp16)[name = tensor("input_125_cast_fp16")]; + tensor linear_50_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = attn12_out_proj_weight_cast_fp16, x = input_125_cast_fp16)[name = tensor("linear_50_cast_fp16")]; + tensor var_5629_to_fp16 = const()[name = tensor("op_5629_to_fp16"), val = tensor(0x1p+0)]; + tensor var_5630_cast_fp16 = add(x = position12_to_fp16, y = var_5629_to_fp16)[name = tensor("op_5630_cast_fp16")]; + tensor var_5630_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_5630_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor input_127_cast_fp16 = add(x = input_123_cast_fp16, y = linear_50_cast_fp16)[name = tensor("input_127_cast_fp16")]; + tensor input_129_axes_0 = const()[name = tensor("input_129_axes_0"), val = tensor([-1])]; + tensor norm12_2_weight_to_fp16 = const()[name = tensor("norm12_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(303692160)))]; + tensor norm12_2_bias_to_fp16 = const()[name = tensor("norm12_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(303694272)))]; + tensor var_5634_to_fp16 = const()[name = tensor("op_5634_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_129_cast_fp16 = layer_norm(axes = input_129_axes_0, beta = norm12_2_bias_to_fp16, epsilon = var_5634_to_fp16, gamma = norm12_2_weight_to_fp16, x = input_127_cast_fp16)[name = tensor("input_129_cast_fp16")]; + tensor linear_51_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = linear12_1_weight_cast_fp16, x = input_129_cast_fp16)[name = tensor("linear_51_cast_fp16")]; + tensor input_131_mode_0 = const()[name = tensor("input_131_mode_0"), val = tensor("EXACT")]; + tensor input_131_cast_fp16 = gelu(mode = input_131_mode_0, x = linear_51_cast_fp16)[name = tensor("input_131_cast_fp16")]; + tensor linear_52_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = linear12_2_weight_cast_fp16, x = input_131_cast_fp16)[name = tensor("linear_52_cast_fp16")]; + tensor input_133_cast_fp16 = add(x = input_127_cast_fp16, y = linear_52_cast_fp16)[name = tensor("input_133_cast_fp16")]; + tensor x_27_axes_0 = const()[name = tensor("x_27_axes_0"), val = tensor([-1])]; + tensor norm13_1_weight_to_fp16 = const()[name = tensor("norm13_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(303696384)))]; + tensor norm13_1_bias_to_fp16 = const()[name = tensor("norm13_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(303698496)))]; + tensor var_5651_to_fp16 = const()[name = tensor("op_5651_to_fp16"), val = tensor(0x1.5p-17)]; + tensor x_27_cast_fp16 = layer_norm(axes = x_27_axes_0, beta = norm13_1_bias_to_fp16, epsilon = var_5651_to_fp16, gamma = norm13_1_weight_to_fp16, x = input_133_cast_fp16)[name = tensor("x_27_cast_fp16")]; + tensor linear_53_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = attn13_in_proj_weight_cast_fp16, x = x_27_cast_fp16)[name = tensor("linear_53_cast_fp16")]; + tensor var_5687 = const()[name = tensor("op_5687"), val = tensor([1, 1, 3, 16, 64])]; + tensor qkv_27_cast_fp16 = reshape(shape = var_5687, x = linear_53_cast_fp16)[name = tensor("qkv_27_cast_fp16")]; + tensor q_79_begin_0 = const()[name = tensor("q_79_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor q_79_end_0 = const()[name = tensor("q_79_end_0"), val = tensor([1, 1, 1, 16, 64])]; + tensor q_79_end_mask_0 = const()[name = tensor("q_79_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor q_79_squeeze_mask_0 = const()[name = tensor("q_79_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor q_79_cast_fp16 = slice_by_index(begin = q_79_begin_0, end = q_79_end_0, end_mask = q_79_end_mask_0, squeeze_mask = q_79_squeeze_mask_0, x = qkv_27_cast_fp16)[name = tensor("q_79_cast_fp16")]; + tensor k_53_begin_0 = const()[name = tensor("k_53_begin_0"), val = tensor([0, 0, 1, 0, 0])]; + tensor k_53_end_0 = const()[name = tensor("k_53_end_0"), val = tensor([1, 1, 2, 16, 64])]; + tensor k_53_end_mask_0 = const()[name = tensor("k_53_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor k_53_squeeze_mask_0 = const()[name = tensor("k_53_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor k_53_cast_fp16 = slice_by_index(begin = k_53_begin_0, end = k_53_end_0, end_mask = k_53_end_mask_0, squeeze_mask = k_53_squeeze_mask_0, x = qkv_27_cast_fp16)[name = tensor("k_53_cast_fp16")]; + tensor v_27_begin_0 = const()[name = tensor("v_27_begin_0"), val = tensor([0, 0, 2, 0, 0])]; + tensor v_27_end_0 = const()[name = tensor("v_27_end_0"), val = tensor([1, 1, 3, 16, 64])]; + tensor v_27_end_mask_0 = const()[name = tensor("v_27_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor v_27_squeeze_mask_0 = const()[name = tensor("v_27_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor v_27_cast_fp16 = slice_by_index(begin = v_27_begin_0, end = v_27_end_0, end_mask = v_27_end_mask_0, squeeze_mask = v_27_squeeze_mask_0, x = qkv_27_cast_fp16)[name = tensor("v_27_cast_fp16")]; + tensor var_5791 = const()[name = tensor("op_5791"), val = tensor([1, 1, 1, 1])]; + tensor position13_to_fp16_dtype_0 = const()[name = tensor("position13_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor position13_to_fp16 = cast(dtype = position13_to_fp16_dtype_0, x = position13)[name = tensor("cast_488")]; + tensor ts_83_cast_fp16 = reshape(shape = var_5791, x = position13_to_fp16)[name = tensor("ts_83_cast_fp16")]; + tensor var_5795 = const()[name = tensor("op_5795"), val = tensor([1, 1, 16, 32, 2])]; + tensor q_complex_27_cast_fp16 = reshape(shape = var_5795, x = q_79_cast_fp16)[name = tensor("q_complex_27_cast_fp16")]; + tensor var_5799 = const()[name = tensor("op_5799"), val = tensor([1, 1, 16, 32, 2])]; + tensor k_complex_27_cast_fp16 = reshape(shape = var_5799, x = k_53_cast_fp16)[name = tensor("k_complex_27_cast_fp16")]; + tensor var_5803_begin_0 = const()[name = tensor("op_5803_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_5803_end_0 = const()[name = tensor("op_5803_end_0"), val = tensor([1, 1, 16, 32, 1])]; + tensor var_5803_end_mask_0 = const()[name = tensor("op_5803_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_5803_squeeze_mask_0 = const()[name = tensor("op_5803_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_5803_cast_fp16 = slice_by_index(begin = var_5803_begin_0, end = var_5803_end_0, end_mask = var_5803_end_mask_0, squeeze_mask = var_5803_squeeze_mask_0, x = q_complex_27_cast_fp16)[name = tensor("op_5803_cast_fp16")]; + tensor var_5811_begin_0 = const()[name = tensor("op_5811_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_5811_end_0 = const()[name = tensor("op_5811_end_0"), val = tensor([1, 1, 16, 32, 2])]; + tensor var_5811_end_mask_0 = const()[name = tensor("op_5811_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_5811_squeeze_mask_0 = const()[name = tensor("op_5811_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_5811_cast_fp16 = slice_by_index(begin = var_5811_begin_0, end = var_5811_end_0, end_mask = var_5811_end_mask_0, squeeze_mask = var_5811_squeeze_mask_0, x = q_complex_27_cast_fp16)[name = tensor("op_5811_cast_fp16")]; + tensor var_5819_begin_0 = const()[name = tensor("op_5819_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_5819_end_0 = const()[name = tensor("op_5819_end_0"), val = tensor([1, 1, 16, 32, 1])]; + tensor var_5819_end_mask_0 = const()[name = tensor("op_5819_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_5819_squeeze_mask_0 = const()[name = tensor("op_5819_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_5819_cast_fp16 = slice_by_index(begin = var_5819_begin_0, end = var_5819_end_0, end_mask = var_5819_end_mask_0, squeeze_mask = var_5819_squeeze_mask_0, x = k_complex_27_cast_fp16)[name = tensor("op_5819_cast_fp16")]; + tensor var_5827_begin_0 = const()[name = tensor("op_5827_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_5827_end_0 = const()[name = tensor("op_5827_end_0"), val = tensor([1, 1, 16, 32, 2])]; + tensor var_5827_end_mask_0 = const()[name = tensor("op_5827_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_5827_squeeze_mask_0 = const()[name = tensor("op_5827_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_5827_cast_fp16 = slice_by_index(begin = var_5827_begin_0, end = var_5827_end_0, end_mask = var_5827_end_mask_0, squeeze_mask = var_5827_squeeze_mask_0, x = k_complex_27_cast_fp16)[name = tensor("op_5827_cast_fp16")]; + tensor freqs_27_to_fp16 = const()[name = tensor("freqs_27_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(303700608)))]; + tensor var_5833_cast_fp16 = mul(x = freqs_27_to_fp16, y = ts_83_cast_fp16)[name = tensor("op_5833_cast_fp16")]; + tensor rotr_27_cast_fp16 = cos(x = var_5833_cast_fp16)[name = tensor("rotr_27_cast_fp16")]; + tensor roti_27_cast_fp16 = sin(x = var_5833_cast_fp16)[name = tensor("roti_27_cast_fp16")]; + tensor var_5837_cast_fp16 = mul(x = var_5803_cast_fp16, y = rotr_27_cast_fp16)[name = tensor("op_5837_cast_fp16")]; + tensor var_5838_cast_fp16 = mul(x = var_5811_cast_fp16, y = roti_27_cast_fp16)[name = tensor("op_5838_cast_fp16")]; + tensor qor_53_cast_fp16 = sub(x = var_5837_cast_fp16, y = var_5838_cast_fp16)[name = tensor("qor_53_cast_fp16")]; + tensor var_5841_cast_fp16 = mul(x = var_5803_cast_fp16, y = roti_27_cast_fp16)[name = tensor("op_5841_cast_fp16")]; + tensor var_5842_cast_fp16 = mul(x = var_5811_cast_fp16, y = rotr_27_cast_fp16)[name = tensor("op_5842_cast_fp16")]; + tensor qoi_53_cast_fp16 = add(x = var_5841_cast_fp16, y = var_5842_cast_fp16)[name = tensor("qoi_53_cast_fp16")]; + tensor var_5845_cast_fp16 = mul(x = var_5819_cast_fp16, y = rotr_27_cast_fp16)[name = tensor("op_5845_cast_fp16")]; + tensor var_5846_cast_fp16 = mul(x = var_5827_cast_fp16, y = roti_27_cast_fp16)[name = tensor("op_5846_cast_fp16")]; + tensor kor_53_cast_fp16 = sub(x = var_5845_cast_fp16, y = var_5846_cast_fp16)[name = tensor("kor_53_cast_fp16")]; + tensor var_5849_cast_fp16 = mul(x = var_5819_cast_fp16, y = roti_27_cast_fp16)[name = tensor("op_5849_cast_fp16")]; + tensor var_5850_cast_fp16 = mul(x = var_5827_cast_fp16, y = rotr_27_cast_fp16)[name = tensor("op_5850_cast_fp16")]; + tensor koi_53_cast_fp16 = add(x = var_5849_cast_fp16, y = var_5850_cast_fp16)[name = tensor("koi_53_cast_fp16")]; + tensor qo_27_axis_0 = const()[name = tensor("qo_27_axis_0"), val = tensor(-1)]; + tensor qo_27_cast_fp16 = stack(axis = qo_27_axis_0, values = (qor_53_cast_fp16, qoi_53_cast_fp16))[name = tensor("qo_27_cast_fp16")]; + tensor ko_27_axis_0 = const()[name = tensor("ko_27_axis_0"), val = tensor(-1)]; + tensor ko_27_cast_fp16 = stack(axis = ko_27_axis_0, values = (kor_53_cast_fp16, koi_53_cast_fp16))[name = tensor("ko_27_cast_fp16")]; + tensor var_5879 = const()[name = tensor("op_5879"), val = tensor([1, 1, 16, 64])]; + tensor q_81_cast_fp16 = reshape(shape = var_5879, x = qo_27_cast_fp16)[name = tensor("q_81_cast_fp16")]; + tensor var_5881 = const()[name = tensor("op_5881"), val = tensor([1, 1, 16, 64])]; + tensor k_55_cast_fp16 = reshape(shape = var_5881, x = ko_27_cast_fp16)[name = tensor("k_55_cast_fp16")]; + tensor _inversed_5903_y_0_to_fp16 = const()[name = tensor("_inversed_5903_y_0_to_fp16"), val = tensor(0x1p-9)]; + tensor _inversed_5903_cast_fp16 = mul(x = ts_83_cast_fp16, y = _inversed_5903_y_0_to_fp16)[name = tensor("_inversed_5903_cast_fp16")]; + tensor var_5904_cast_fp16 = floor(x = _inversed_5903_cast_fp16)[name = tensor("op_5904_cast_fp16")]; + tensor var_5905_to_fp16 = const()[name = tensor("op_5905_to_fp16"), val = tensor(0x1p+9)]; + tensor var_5906_cast_fp16 = mul(x = var_5904_cast_fp16, y = var_5905_to_fp16)[name = tensor("op_5906_cast_fp16")]; + tensor write_indices_float_55_cast_fp16 = sub(x = ts_83_cast_fp16, y = var_5906_cast_fp16)[name = tensor("write_indices_float_55_cast_fp16")]; + tensor var_5913_dtype_0 = const()[name = tensor("op_5913_dtype_0"), val = tensor("int32")]; + tensor write_indices_27_reps_0 = const()[name = tensor("write_indices_27_reps_0"), val = tensor([1, 1, 16, 64])]; + tensor write_indices_float_55_cast_fp16_to_int32 = cast(dtype = var_5913_dtype_0, x = write_indices_float_55_cast_fp16)[name = tensor("cast_487")]; + tensor write_indices_27 = tile(reps = write_indices_27_reps_0, x = write_indices_float_55_cast_fp16_to_int32)[name = tensor("write_indices_27")]; + tensor var_5921_begin_0 = const()[name = tensor("op_5921_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_5921_end_0 = const()[name = tensor("op_5921_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor var_5921_end_mask_0 = const()[name = tensor("op_5921_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_5921_squeeze_mask_0 = const()[name = tensor("op_5921_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor cache13_to_fp16_dtype_0 = const()[name = tensor("cache13_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor cache13_to_fp16 = cast(dtype = cache13_to_fp16_dtype_0, x = cache13)[name = tensor("cast_486")]; + tensor var_5921_cast_fp16 = slice_by_index(begin = var_5921_begin_0, end = var_5921_end_0, end_mask = var_5921_end_mask_0, squeeze_mask = var_5921_squeeze_mask_0, x = cache13_to_fp16)[name = tensor("op_5921_cast_fp16")]; + tensor var_5923_axis_0 = const()[name = tensor("op_5923_axis_0"), val = tensor(1)]; + tensor var_5923_mode_0 = const()[name = tensor("op_5923_mode_0"), val = tensor("update")]; + tensor var_5923_validate_indices_0 = const()[name = tensor("op_5923_validate_indices_0"), val = tensor(false)]; + tensor var_5923_cast_fp16 = scatter_along_axis(axis = var_5923_axis_0, data = var_5921_cast_fp16, indices = write_indices_27, mode = var_5923_mode_0, updates = k_55_cast_fp16, validate_indices = var_5923_validate_indices_0)[name = tensor("op_5923_cast_fp16")]; + tensor concat_93 = const()[name = tensor("concat_93"), val = tensor([0, 0, 0, 0, 0])]; + tensor concat_94 = const()[name = tensor("concat_94"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_27_internal_tensor_assign_1_stride_0 = const()[name = tensor("new_cache_27_internal_tensor_assign_1_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_27_internal_tensor_assign_1_begin_mask_0 = const()[name = tensor("new_cache_27_internal_tensor_assign_1_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_27_internal_tensor_assign_1_end_mask_0 = const()[name = tensor("new_cache_27_internal_tensor_assign_1_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_27_internal_tensor_assign_1_squeeze_mask_0 = const()[name = tensor("new_cache_27_internal_tensor_assign_1_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_74 = const()[name = tensor("shape_74"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_26 = const()[name = tensor("reduce_prod_26"), val = tensor(1048576)]; + tensor range_1d_26_start_0 = const()[name = tensor("range_1d_26_start_0"), val = tensor(0)]; + tensor range_1d_26_step_0 = const()[name = tensor("range_1d_26_step_0"), val = tensor(1)]; + tensor range_1d_26 = range_1d(end = reduce_prod_26, start = range_1d_26_start_0, step = range_1d_26_step_0)[name = tensor("range_1d_26")]; + tensor reshape_130 = reshape(shape = shape_74, x = range_1d_26)[name = tensor("reshape_130")]; + tensor slice_by_index_26 = slice_by_index(begin = concat_93, begin_mask = new_cache_27_internal_tensor_assign_1_begin_mask_0, end = concat_94, end_mask = new_cache_27_internal_tensor_assign_1_end_mask_0, squeeze_mask = new_cache_27_internal_tensor_assign_1_squeeze_mask_0, stride = new_cache_27_internal_tensor_assign_1_stride_0, x = reshape_130)[name = tensor("slice_by_index_26")]; + tensor reshape_131_shape_0 = const()[name = tensor("reshape_131_shape_0"), val = tensor([-1])]; + tensor reshape_131 = reshape(shape = reshape_131_shape_0, x = slice_by_index_26)[name = tensor("reshape_131")]; + tensor reshape_132_shape_0 = const()[name = tensor("reshape_132_shape_0"), val = tensor([-1])]; + tensor reshape_132_cast_fp16 = reshape(shape = reshape_132_shape_0, x = var_5923_cast_fp16)[name = tensor("reshape_132_cast_fp16")]; + tensor reshape_133_shape_0 = const()[name = tensor("reshape_133_shape_0"), val = tensor([-1])]; + tensor reshape_133_cast_fp16 = reshape(shape = reshape_133_shape_0, x = cache13_to_fp16)[name = tensor("reshape_133_cast_fp16")]; + tensor scatter_26_mode_0 = const()[name = tensor("scatter_26_mode_0"), val = tensor("update")]; + tensor scatter_26_axis_0 = const()[name = tensor("scatter_26_axis_0"), val = tensor(0)]; + tensor scatter_26_validate_indices_0 = const()[name = tensor("scatter_26_validate_indices_0"), val = tensor(false)]; + tensor scatter_26_cast_fp16 = scatter(axis = scatter_26_axis_0, data = reshape_133_cast_fp16, indices = reshape_131, mode = scatter_26_mode_0, updates = reshape_132_cast_fp16, validate_indices = scatter_26_validate_indices_0)[name = tensor("scatter_26_cast_fp16")]; + tensor reshape_134_cast_fp16 = reshape(shape = shape_74, x = scatter_26_cast_fp16)[name = tensor("reshape_134_cast_fp16")]; + tensor var_5931_begin_0 = const()[name = tensor("op_5931_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor var_5931_end_0 = const()[name = tensor("op_5931_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor var_5931_end_mask_0 = const()[name = tensor("op_5931_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_5931_squeeze_mask_0 = const()[name = tensor("op_5931_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor var_5931_cast_fp16 = slice_by_index(begin = var_5931_begin_0, end = var_5931_end_0, end_mask = var_5931_end_mask_0, squeeze_mask = var_5931_squeeze_mask_0, x = reshape_134_cast_fp16)[name = tensor("op_5931_cast_fp16")]; + tensor var_5933_axis_0 = const()[name = tensor("op_5933_axis_0"), val = tensor(1)]; + tensor var_5933_mode_0 = const()[name = tensor("op_5933_mode_0"), val = tensor("update")]; + tensor var_5933_validate_indices_0 = const()[name = tensor("op_5933_validate_indices_0"), val = tensor(false)]; + tensor var_5933_cast_fp16 = scatter_along_axis(axis = var_5933_axis_0, data = var_5931_cast_fp16, indices = write_indices_27, mode = var_5933_mode_0, updates = v_27_cast_fp16, validate_indices = var_5933_validate_indices_0)[name = tensor("op_5933_cast_fp16")]; + tensor concat_95 = const()[name = tensor("concat_95"), val = tensor([1, 0, 0, 0, 0])]; + tensor concat_96 = const()[name = tensor("concat_96"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_27_internal_tensor_assign_2_stride_0 = const()[name = tensor("new_cache_27_internal_tensor_assign_2_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_27_internal_tensor_assign_2_begin_mask_0 = const()[name = tensor("new_cache_27_internal_tensor_assign_2_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_27_internal_tensor_assign_2_end_mask_0 = const()[name = tensor("new_cache_27_internal_tensor_assign_2_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_27_internal_tensor_assign_2_squeeze_mask_0 = const()[name = tensor("new_cache_27_internal_tensor_assign_2_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_75 = const()[name = tensor("shape_75"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_27 = const()[name = tensor("reduce_prod_27"), val = tensor(1048576)]; + tensor range_1d_27_start_0 = const()[name = tensor("range_1d_27_start_0"), val = tensor(0)]; + tensor range_1d_27_step_0 = const()[name = tensor("range_1d_27_step_0"), val = tensor(1)]; + tensor range_1d_27 = range_1d(end = reduce_prod_27, start = range_1d_27_start_0, step = range_1d_27_step_0)[name = tensor("range_1d_27")]; + tensor reshape_135 = reshape(shape = shape_75, x = range_1d_27)[name = tensor("reshape_135")]; + tensor slice_by_index_27 = slice_by_index(begin = concat_95, begin_mask = new_cache_27_internal_tensor_assign_2_begin_mask_0, end = concat_96, end_mask = new_cache_27_internal_tensor_assign_2_end_mask_0, squeeze_mask = new_cache_27_internal_tensor_assign_2_squeeze_mask_0, stride = new_cache_27_internal_tensor_assign_2_stride_0, x = reshape_135)[name = tensor("slice_by_index_27")]; + tensor reshape_136_shape_0 = const()[name = tensor("reshape_136_shape_0"), val = tensor([-1])]; + tensor reshape_136 = reshape(shape = reshape_136_shape_0, x = slice_by_index_27)[name = tensor("reshape_136")]; + tensor reshape_137_shape_0 = const()[name = tensor("reshape_137_shape_0"), val = tensor([-1])]; + tensor reshape_137_cast_fp16 = reshape(shape = reshape_137_shape_0, x = var_5933_cast_fp16)[name = tensor("reshape_137_cast_fp16")]; + tensor reshape_138_shape_0 = const()[name = tensor("reshape_138_shape_0"), val = tensor([-1])]; + tensor reshape_138_cast_fp16 = reshape(shape = reshape_138_shape_0, x = reshape_134_cast_fp16)[name = tensor("reshape_138_cast_fp16")]; + tensor scatter_27_mode_0 = const()[name = tensor("scatter_27_mode_0"), val = tensor("update")]; + tensor scatter_27_axis_0 = const()[name = tensor("scatter_27_axis_0"), val = tensor(0)]; + tensor scatter_27_validate_indices_0 = const()[name = tensor("scatter_27_validate_indices_0"), val = tensor(false)]; + tensor scatter_27_cast_fp16 = scatter(axis = scatter_27_axis_0, data = reshape_138_cast_fp16, indices = reshape_136, mode = scatter_27_mode_0, updates = reshape_137_cast_fp16, validate_indices = scatter_27_validate_indices_0)[name = tensor("scatter_27_cast_fp16")]; + tensor reshape_139_cast_fp16 = reshape(shape = shape_75, x = scatter_27_cast_fp16)[name = tensor("reshape_139_cast_fp16")]; + tensor reshape_139_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("reshape_139_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor keys_79_begin_0 = const()[name = tensor("keys_79_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor keys_79_end_0 = const()[name = tensor("keys_79_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor keys_79_end_mask_0 = const()[name = tensor("keys_79_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor keys_79_squeeze_mask_0 = const()[name = tensor("keys_79_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor keys_79_cast_fp16 = slice_by_index(begin = keys_79_begin_0, end = keys_79_end_0, end_mask = keys_79_end_mask_0, squeeze_mask = keys_79_squeeze_mask_0, x = reshape_139_cast_fp16)[name = tensor("keys_79_cast_fp16")]; + tensor values_79_begin_0 = const()[name = tensor("values_79_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor values_79_end_0 = const()[name = tensor("values_79_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor values_79_end_mask_0 = const()[name = tensor("values_79_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor values_79_squeeze_mask_0 = const()[name = tensor("values_79_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor values_79_cast_fp16 = slice_by_index(begin = values_79_begin_0, end = values_79_end_0, end_mask = values_79_end_mask_0, squeeze_mask = values_79_squeeze_mask_0, x = reshape_139_cast_fp16)[name = tensor("values_79_cast_fp16")]; + tensor var_5945_cast_fp16 = not_equal(x = keys_79_cast_fp16, y = keys_79_cast_fp16)[name = tensor("op_5945_cast_fp16")]; + tensor keys_81_cast_fp16 = select(a = var_504_to_fp16, b = keys_79_cast_fp16, cond = var_5945_cast_fp16)[name = tensor("keys_81_cast_fp16")]; + tensor var_5953_cast_fp16 = not_equal(x = values_79_cast_fp16, y = values_79_cast_fp16)[name = tensor("op_5953_cast_fp16")]; + tensor values_81_cast_fp16 = select(a = var_504_to_fp16, b = values_79_cast_fp16, cond = var_5953_cast_fp16)[name = tensor("values_81_cast_fp16")]; + tensor var_5977 = const()[name = tensor("op_5977"), val = tensor([0, 2, 1, 3])]; + tensor var_5990 = const()[name = tensor("op_5990"), val = tensor([1, 1, 1])]; + tensor var_5991_cast_fp16 = reshape(shape = var_5990, x = position13_to_fp16)[name = tensor("op_5991_cast_fp16")]; + tensor var_6008_to_fp16 = const()[name = tensor("op_6008_to_fp16"), val = tensor(0x1p+0)]; + tensor valid_len_27_cast_fp16 = add(x = var_5991_cast_fp16, y = var_6008_to_fp16)[name = tensor("valid_len_27_cast_fp16")]; + tensor valid_mask_27_cast_fp16 = less(x = k_positions_1_promoted_to_fp16, y = valid_len_27_cast_fp16)[name = tensor("valid_mask_27_cast_fp16")]; + tensor causal_mask_27_cast_fp16 = less_equal(x = k_positions_1_promoted_to_fp16, y = var_5991_cast_fp16)[name = tensor("causal_mask_27_cast_fp16")]; + tensor attn_mask_53 = logical_and(x = valid_mask_27_cast_fp16, y = causal_mask_27_cast_fp16)[name = tensor("attn_mask_53")]; + tensor attn_mask_55_axes_0 = const()[name = tensor("attn_mask_55_axes_0"), val = tensor([1])]; + tensor attn_mask_55 = expand_dims(axes = attn_mask_55_axes_0, x = attn_mask_53)[name = tensor("attn_mask_55")]; + tensor var_6026_transpose_x_0 = const()[name = tensor("op_6026_transpose_x_0"), val = tensor(false)]; + tensor var_6026_transpose_y_0 = const()[name = tensor("op_6026_transpose_y_0"), val = tensor(false)]; + tensor transpose_98_perm_0 = const()[name = tensor("transpose_98_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_99_perm_0 = const()[name = tensor("transpose_99_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_99 = transpose(perm = transpose_99_perm_0, x = keys_81_cast_fp16)[name = tensor("transpose_161")]; + tensor transpose_98 = transpose(perm = transpose_98_perm_0, x = q_81_cast_fp16)[name = tensor("transpose_162")]; + tensor var_6026_cast_fp16 = matmul(transpose_x = var_6026_transpose_x_0, transpose_y = var_6026_transpose_y_0, x = transpose_98, y = transpose_99)[name = tensor("op_6026_cast_fp16")]; + tensor var_6020_to_fp16 = const()[name = tensor("op_6020_to_fp16"), val = tensor([0x1p-3])]; + tensor attn_weights_79_cast_fp16 = mul(x = var_6026_cast_fp16, y = var_6020_to_fp16)[name = tensor("attn_weights_79_cast_fp16")]; + tensor var_6028 = logical_not(x = attn_mask_55)[name = tensor("op_6028")]; + tensor var_6029_to_fp16 = const()[name = tensor("op_6029_to_fp16"), val = tensor(-inf)]; + tensor attn_weights_81_cast_fp16 = select(a = var_6029_to_fp16, b = attn_weights_79_cast_fp16, cond = var_6028)[name = tensor("attn_weights_81_cast_fp16")]; + tensor var_6031 = const()[name = tensor("op_6031"), val = tensor(-1)]; + tensor attn_weights_83_cast_fp16 = softmax(axis = var_6031, x = attn_weights_81_cast_fp16)[name = tensor("attn_weights_83_cast_fp16")]; + tensor attn_output_27_transpose_x_0 = const()[name = tensor("attn_output_27_transpose_x_0"), val = tensor(false)]; + tensor attn_output_27_transpose_y_0 = const()[name = tensor("attn_output_27_transpose_y_0"), val = tensor(false)]; + tensor values_83_cast_fp16 = transpose(perm = var_5977, x = values_81_cast_fp16)[name = tensor("transpose_163")]; + tensor attn_output_27_cast_fp16 = matmul(transpose_x = attn_output_27_transpose_x_0, transpose_y = attn_output_27_transpose_y_0, x = attn_weights_83_cast_fp16, y = values_83_cast_fp16)[name = tensor("attn_output_27_cast_fp16")]; + tensor var_6039 = const()[name = tensor("op_6039"), val = tensor([0, 2, 1, 3])]; + tensor var_6042 = const()[name = tensor("op_6042"), val = tensor([1, 1, 1024])]; + tensor var_6040_cast_fp16 = transpose(perm = var_6039, x = attn_output_27_cast_fp16)[name = tensor("transpose_160")]; + tensor input_135_cast_fp16 = reshape(shape = var_6042, x = var_6040_cast_fp16)[name = tensor("input_135_cast_fp16")]; + tensor linear_54_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = attn13_out_proj_weight_cast_fp16, x = input_135_cast_fp16)[name = tensor("linear_54_cast_fp16")]; + tensor var_6048_to_fp16 = const()[name = tensor("op_6048_to_fp16"), val = tensor(0x1p+0)]; + tensor var_6049_cast_fp16 = add(x = position13_to_fp16, y = var_6048_to_fp16)[name = tensor("op_6049_cast_fp16")]; + tensor var_6049_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_6049_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor input_137_cast_fp16 = add(x = input_133_cast_fp16, y = linear_54_cast_fp16)[name = tensor("input_137_cast_fp16")]; + tensor input_139_axes_0 = const()[name = tensor("input_139_axes_0"), val = tensor([-1])]; + tensor norm13_2_weight_to_fp16 = const()[name = tensor("norm13_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(303700736)))]; + tensor norm13_2_bias_to_fp16 = const()[name = tensor("norm13_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(303702848)))]; + tensor var_6053_to_fp16 = const()[name = tensor("op_6053_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_139_cast_fp16 = layer_norm(axes = input_139_axes_0, beta = norm13_2_bias_to_fp16, epsilon = var_6053_to_fp16, gamma = norm13_2_weight_to_fp16, x = input_137_cast_fp16)[name = tensor("input_139_cast_fp16")]; + tensor linear_55_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = linear13_1_weight_cast_fp16, x = input_139_cast_fp16)[name = tensor("linear_55_cast_fp16")]; + tensor input_141_mode_0 = const()[name = tensor("input_141_mode_0"), val = tensor("EXACT")]; + tensor input_141_cast_fp16 = gelu(mode = input_141_mode_0, x = linear_55_cast_fp16)[name = tensor("input_141_cast_fp16")]; + tensor linear_56_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = linear13_2_weight_cast_fp16, x = input_141_cast_fp16)[name = tensor("linear_56_cast_fp16")]; + tensor input_143_cast_fp16 = add(x = input_137_cast_fp16, y = linear_56_cast_fp16)[name = tensor("input_143_cast_fp16")]; + tensor x_29_axes_0 = const()[name = tensor("x_29_axes_0"), val = tensor([-1])]; + tensor norm14_1_weight_to_fp16 = const()[name = tensor("norm14_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(303704960)))]; + tensor norm14_1_bias_to_fp16 = const()[name = tensor("norm14_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(303707072)))]; + tensor var_6070_to_fp16 = const()[name = tensor("op_6070_to_fp16"), val = tensor(0x1.5p-17)]; + tensor x_29_cast_fp16 = layer_norm(axes = x_29_axes_0, beta = norm14_1_bias_to_fp16, epsilon = var_6070_to_fp16, gamma = norm14_1_weight_to_fp16, x = input_143_cast_fp16)[name = tensor("x_29_cast_fp16")]; + tensor linear_57_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = attn14_in_proj_weight_cast_fp16, x = x_29_cast_fp16)[name = tensor("linear_57_cast_fp16")]; + tensor var_6106 = const()[name = tensor("op_6106"), val = tensor([1, 1, 3, 16, 64])]; + tensor qkv_29_cast_fp16 = reshape(shape = var_6106, x = linear_57_cast_fp16)[name = tensor("qkv_29_cast_fp16")]; + tensor q_85_begin_0 = const()[name = tensor("q_85_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor q_85_end_0 = const()[name = tensor("q_85_end_0"), val = tensor([1, 1, 1, 16, 64])]; + tensor q_85_end_mask_0 = const()[name = tensor("q_85_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor q_85_squeeze_mask_0 = const()[name = tensor("q_85_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor q_85_cast_fp16 = slice_by_index(begin = q_85_begin_0, end = q_85_end_0, end_mask = q_85_end_mask_0, squeeze_mask = q_85_squeeze_mask_0, x = qkv_29_cast_fp16)[name = tensor("q_85_cast_fp16")]; + tensor k_57_begin_0 = const()[name = tensor("k_57_begin_0"), val = tensor([0, 0, 1, 0, 0])]; + tensor k_57_end_0 = const()[name = tensor("k_57_end_0"), val = tensor([1, 1, 2, 16, 64])]; + tensor k_57_end_mask_0 = const()[name = tensor("k_57_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor k_57_squeeze_mask_0 = const()[name = tensor("k_57_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor k_57_cast_fp16 = slice_by_index(begin = k_57_begin_0, end = k_57_end_0, end_mask = k_57_end_mask_0, squeeze_mask = k_57_squeeze_mask_0, x = qkv_29_cast_fp16)[name = tensor("k_57_cast_fp16")]; + tensor v_29_begin_0 = const()[name = tensor("v_29_begin_0"), val = tensor([0, 0, 2, 0, 0])]; + tensor v_29_end_0 = const()[name = tensor("v_29_end_0"), val = tensor([1, 1, 3, 16, 64])]; + tensor v_29_end_mask_0 = const()[name = tensor("v_29_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor v_29_squeeze_mask_0 = const()[name = tensor("v_29_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor v_29_cast_fp16 = slice_by_index(begin = v_29_begin_0, end = v_29_end_0, end_mask = v_29_end_mask_0, squeeze_mask = v_29_squeeze_mask_0, x = qkv_29_cast_fp16)[name = tensor("v_29_cast_fp16")]; + tensor var_6210 = const()[name = tensor("op_6210"), val = tensor([1, 1, 1, 1])]; + tensor position14_to_fp16_dtype_0 = const()[name = tensor("position14_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor position14_to_fp16 = cast(dtype = position14_to_fp16_dtype_0, x = position14)[name = tensor("cast_483")]; + tensor ts_89_cast_fp16 = reshape(shape = var_6210, x = position14_to_fp16)[name = tensor("ts_89_cast_fp16")]; + tensor var_6214 = const()[name = tensor("op_6214"), val = tensor([1, 1, 16, 32, 2])]; + tensor q_complex_29_cast_fp16 = reshape(shape = var_6214, x = q_85_cast_fp16)[name = tensor("q_complex_29_cast_fp16")]; + tensor var_6218 = const()[name = tensor("op_6218"), val = tensor([1, 1, 16, 32, 2])]; + tensor k_complex_29_cast_fp16 = reshape(shape = var_6218, x = k_57_cast_fp16)[name = tensor("k_complex_29_cast_fp16")]; + tensor var_6222_begin_0 = const()[name = tensor("op_6222_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_6222_end_0 = const()[name = tensor("op_6222_end_0"), val = tensor([1, 1, 16, 32, 1])]; + tensor var_6222_end_mask_0 = const()[name = tensor("op_6222_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_6222_squeeze_mask_0 = const()[name = tensor("op_6222_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_6222_cast_fp16 = slice_by_index(begin = var_6222_begin_0, end = var_6222_end_0, end_mask = var_6222_end_mask_0, squeeze_mask = var_6222_squeeze_mask_0, x = q_complex_29_cast_fp16)[name = tensor("op_6222_cast_fp16")]; + tensor var_6230_begin_0 = const()[name = tensor("op_6230_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_6230_end_0 = const()[name = tensor("op_6230_end_0"), val = tensor([1, 1, 16, 32, 2])]; + tensor var_6230_end_mask_0 = const()[name = tensor("op_6230_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_6230_squeeze_mask_0 = const()[name = tensor("op_6230_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_6230_cast_fp16 = slice_by_index(begin = var_6230_begin_0, end = var_6230_end_0, end_mask = var_6230_end_mask_0, squeeze_mask = var_6230_squeeze_mask_0, x = q_complex_29_cast_fp16)[name = tensor("op_6230_cast_fp16")]; + tensor var_6238_begin_0 = const()[name = tensor("op_6238_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_6238_end_0 = const()[name = tensor("op_6238_end_0"), val = tensor([1, 1, 16, 32, 1])]; + tensor var_6238_end_mask_0 = const()[name = tensor("op_6238_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_6238_squeeze_mask_0 = const()[name = tensor("op_6238_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_6238_cast_fp16 = slice_by_index(begin = var_6238_begin_0, end = var_6238_end_0, end_mask = var_6238_end_mask_0, squeeze_mask = var_6238_squeeze_mask_0, x = k_complex_29_cast_fp16)[name = tensor("op_6238_cast_fp16")]; + tensor var_6246_begin_0 = const()[name = tensor("op_6246_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_6246_end_0 = const()[name = tensor("op_6246_end_0"), val = tensor([1, 1, 16, 32, 2])]; + tensor var_6246_end_mask_0 = const()[name = tensor("op_6246_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_6246_squeeze_mask_0 = const()[name = tensor("op_6246_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_6246_cast_fp16 = slice_by_index(begin = var_6246_begin_0, end = var_6246_end_0, end_mask = var_6246_end_mask_0, squeeze_mask = var_6246_squeeze_mask_0, x = k_complex_29_cast_fp16)[name = tensor("op_6246_cast_fp16")]; + tensor freqs_29_to_fp16 = const()[name = tensor("freqs_29_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(303709184)))]; + tensor var_6252_cast_fp16 = mul(x = freqs_29_to_fp16, y = ts_89_cast_fp16)[name = tensor("op_6252_cast_fp16")]; + tensor rotr_29_cast_fp16 = cos(x = var_6252_cast_fp16)[name = tensor("rotr_29_cast_fp16")]; + tensor roti_29_cast_fp16 = sin(x = var_6252_cast_fp16)[name = tensor("roti_29_cast_fp16")]; + tensor var_6256_cast_fp16 = mul(x = var_6222_cast_fp16, y = rotr_29_cast_fp16)[name = tensor("op_6256_cast_fp16")]; + tensor var_6257_cast_fp16 = mul(x = var_6230_cast_fp16, y = roti_29_cast_fp16)[name = tensor("op_6257_cast_fp16")]; + tensor qor_57_cast_fp16 = sub(x = var_6256_cast_fp16, y = var_6257_cast_fp16)[name = tensor("qor_57_cast_fp16")]; + tensor var_6260_cast_fp16 = mul(x = var_6222_cast_fp16, y = roti_29_cast_fp16)[name = tensor("op_6260_cast_fp16")]; + tensor var_6261_cast_fp16 = mul(x = var_6230_cast_fp16, y = rotr_29_cast_fp16)[name = tensor("op_6261_cast_fp16")]; + tensor qoi_57_cast_fp16 = add(x = var_6260_cast_fp16, y = var_6261_cast_fp16)[name = tensor("qoi_57_cast_fp16")]; + tensor var_6264_cast_fp16 = mul(x = var_6238_cast_fp16, y = rotr_29_cast_fp16)[name = tensor("op_6264_cast_fp16")]; + tensor var_6265_cast_fp16 = mul(x = var_6246_cast_fp16, y = roti_29_cast_fp16)[name = tensor("op_6265_cast_fp16")]; + tensor kor_57_cast_fp16 = sub(x = var_6264_cast_fp16, y = var_6265_cast_fp16)[name = tensor("kor_57_cast_fp16")]; + tensor var_6268_cast_fp16 = mul(x = var_6238_cast_fp16, y = roti_29_cast_fp16)[name = tensor("op_6268_cast_fp16")]; + tensor var_6269_cast_fp16 = mul(x = var_6246_cast_fp16, y = rotr_29_cast_fp16)[name = tensor("op_6269_cast_fp16")]; + tensor koi_57_cast_fp16 = add(x = var_6268_cast_fp16, y = var_6269_cast_fp16)[name = tensor("koi_57_cast_fp16")]; + tensor qo_29_axis_0 = const()[name = tensor("qo_29_axis_0"), val = tensor(-1)]; + tensor qo_29_cast_fp16 = stack(axis = qo_29_axis_0, values = (qor_57_cast_fp16, qoi_57_cast_fp16))[name = tensor("qo_29_cast_fp16")]; + tensor ko_29_axis_0 = const()[name = tensor("ko_29_axis_0"), val = tensor(-1)]; + tensor ko_29_cast_fp16 = stack(axis = ko_29_axis_0, values = (kor_57_cast_fp16, koi_57_cast_fp16))[name = tensor("ko_29_cast_fp16")]; + tensor var_6298 = const()[name = tensor("op_6298"), val = tensor([1, 1, 16, 64])]; + tensor q_87_cast_fp16 = reshape(shape = var_6298, x = qo_29_cast_fp16)[name = tensor("q_87_cast_fp16")]; + tensor var_6300 = const()[name = tensor("op_6300"), val = tensor([1, 1, 16, 64])]; + tensor k_59_cast_fp16 = reshape(shape = var_6300, x = ko_29_cast_fp16)[name = tensor("k_59_cast_fp16")]; + tensor _inversed_6322_y_0_to_fp16 = const()[name = tensor("_inversed_6322_y_0_to_fp16"), val = tensor(0x1p-9)]; + tensor _inversed_6322_cast_fp16 = mul(x = ts_89_cast_fp16, y = _inversed_6322_y_0_to_fp16)[name = tensor("_inversed_6322_cast_fp16")]; + tensor var_6323_cast_fp16 = floor(x = _inversed_6322_cast_fp16)[name = tensor("op_6323_cast_fp16")]; + tensor var_6324_to_fp16 = const()[name = tensor("op_6324_to_fp16"), val = tensor(0x1p+9)]; + tensor var_6325_cast_fp16 = mul(x = var_6323_cast_fp16, y = var_6324_to_fp16)[name = tensor("op_6325_cast_fp16")]; + tensor write_indices_float_59_cast_fp16 = sub(x = ts_89_cast_fp16, y = var_6325_cast_fp16)[name = tensor("write_indices_float_59_cast_fp16")]; + tensor var_6332_dtype_0 = const()[name = tensor("op_6332_dtype_0"), val = tensor("int32")]; + tensor write_indices_29_reps_0 = const()[name = tensor("write_indices_29_reps_0"), val = tensor([1, 1, 16, 64])]; + tensor write_indices_float_59_cast_fp16_to_int32 = cast(dtype = var_6332_dtype_0, x = write_indices_float_59_cast_fp16)[name = tensor("cast_482")]; + tensor write_indices_29 = tile(reps = write_indices_29_reps_0, x = write_indices_float_59_cast_fp16_to_int32)[name = tensor("write_indices_29")]; + tensor var_6340_begin_0 = const()[name = tensor("op_6340_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_6340_end_0 = const()[name = tensor("op_6340_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor var_6340_end_mask_0 = const()[name = tensor("op_6340_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_6340_squeeze_mask_0 = const()[name = tensor("op_6340_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor cache14_to_fp16_dtype_0 = const()[name = tensor("cache14_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor cache14_to_fp16 = cast(dtype = cache14_to_fp16_dtype_0, x = cache14)[name = tensor("cast_481")]; + tensor var_6340_cast_fp16 = slice_by_index(begin = var_6340_begin_0, end = var_6340_end_0, end_mask = var_6340_end_mask_0, squeeze_mask = var_6340_squeeze_mask_0, x = cache14_to_fp16)[name = tensor("op_6340_cast_fp16")]; + tensor var_6342_axis_0 = const()[name = tensor("op_6342_axis_0"), val = tensor(1)]; + tensor var_6342_mode_0 = const()[name = tensor("op_6342_mode_0"), val = tensor("update")]; + tensor var_6342_validate_indices_0 = const()[name = tensor("op_6342_validate_indices_0"), val = tensor(false)]; + tensor var_6342_cast_fp16 = scatter_along_axis(axis = var_6342_axis_0, data = var_6340_cast_fp16, indices = write_indices_29, mode = var_6342_mode_0, updates = k_59_cast_fp16, validate_indices = var_6342_validate_indices_0)[name = tensor("op_6342_cast_fp16")]; + tensor concat_100 = const()[name = tensor("concat_100"), val = tensor([0, 0, 0, 0, 0])]; + tensor concat_101 = const()[name = tensor("concat_101"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_29_internal_tensor_assign_1_stride_0 = const()[name = tensor("new_cache_29_internal_tensor_assign_1_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_29_internal_tensor_assign_1_begin_mask_0 = const()[name = tensor("new_cache_29_internal_tensor_assign_1_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_29_internal_tensor_assign_1_end_mask_0 = const()[name = tensor("new_cache_29_internal_tensor_assign_1_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_29_internal_tensor_assign_1_squeeze_mask_0 = const()[name = tensor("new_cache_29_internal_tensor_assign_1_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_76 = const()[name = tensor("shape_76"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_28 = const()[name = tensor("reduce_prod_28"), val = tensor(1048576)]; + tensor range_1d_28_start_0 = const()[name = tensor("range_1d_28_start_0"), val = tensor(0)]; + tensor range_1d_28_step_0 = const()[name = tensor("range_1d_28_step_0"), val = tensor(1)]; + tensor range_1d_28 = range_1d(end = reduce_prod_28, start = range_1d_28_start_0, step = range_1d_28_step_0)[name = tensor("range_1d_28")]; + tensor reshape_140 = reshape(shape = shape_76, x = range_1d_28)[name = tensor("reshape_140")]; + tensor slice_by_index_28 = slice_by_index(begin = concat_100, begin_mask = new_cache_29_internal_tensor_assign_1_begin_mask_0, end = concat_101, end_mask = new_cache_29_internal_tensor_assign_1_end_mask_0, squeeze_mask = new_cache_29_internal_tensor_assign_1_squeeze_mask_0, stride = new_cache_29_internal_tensor_assign_1_stride_0, x = reshape_140)[name = tensor("slice_by_index_28")]; + tensor reshape_141_shape_0 = const()[name = tensor("reshape_141_shape_0"), val = tensor([-1])]; + tensor reshape_141 = reshape(shape = reshape_141_shape_0, x = slice_by_index_28)[name = tensor("reshape_141")]; + tensor reshape_142_shape_0 = const()[name = tensor("reshape_142_shape_0"), val = tensor([-1])]; + tensor reshape_142_cast_fp16 = reshape(shape = reshape_142_shape_0, x = var_6342_cast_fp16)[name = tensor("reshape_142_cast_fp16")]; + tensor reshape_143_shape_0 = const()[name = tensor("reshape_143_shape_0"), val = tensor([-1])]; + tensor reshape_143_cast_fp16 = reshape(shape = reshape_143_shape_0, x = cache14_to_fp16)[name = tensor("reshape_143_cast_fp16")]; + tensor scatter_28_mode_0 = const()[name = tensor("scatter_28_mode_0"), val = tensor("update")]; + tensor scatter_28_axis_0 = const()[name = tensor("scatter_28_axis_0"), val = tensor(0)]; + tensor scatter_28_validate_indices_0 = const()[name = tensor("scatter_28_validate_indices_0"), val = tensor(false)]; + tensor scatter_28_cast_fp16 = scatter(axis = scatter_28_axis_0, data = reshape_143_cast_fp16, indices = reshape_141, mode = scatter_28_mode_0, updates = reshape_142_cast_fp16, validate_indices = scatter_28_validate_indices_0)[name = tensor("scatter_28_cast_fp16")]; + tensor reshape_144_cast_fp16 = reshape(shape = shape_76, x = scatter_28_cast_fp16)[name = tensor("reshape_144_cast_fp16")]; + tensor var_6350_begin_0 = const()[name = tensor("op_6350_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor var_6350_end_0 = const()[name = tensor("op_6350_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor var_6350_end_mask_0 = const()[name = tensor("op_6350_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_6350_squeeze_mask_0 = const()[name = tensor("op_6350_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor var_6350_cast_fp16 = slice_by_index(begin = var_6350_begin_0, end = var_6350_end_0, end_mask = var_6350_end_mask_0, squeeze_mask = var_6350_squeeze_mask_0, x = reshape_144_cast_fp16)[name = tensor("op_6350_cast_fp16")]; + tensor var_6352_axis_0 = const()[name = tensor("op_6352_axis_0"), val = tensor(1)]; + tensor var_6352_mode_0 = const()[name = tensor("op_6352_mode_0"), val = tensor("update")]; + tensor var_6352_validate_indices_0 = const()[name = tensor("op_6352_validate_indices_0"), val = tensor(false)]; + tensor var_6352_cast_fp16 = scatter_along_axis(axis = var_6352_axis_0, data = var_6350_cast_fp16, indices = write_indices_29, mode = var_6352_mode_0, updates = v_29_cast_fp16, validate_indices = var_6352_validate_indices_0)[name = tensor("op_6352_cast_fp16")]; + tensor concat_102 = const()[name = tensor("concat_102"), val = tensor([1, 0, 0, 0, 0])]; + tensor concat_103 = const()[name = tensor("concat_103"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_29_internal_tensor_assign_2_stride_0 = const()[name = tensor("new_cache_29_internal_tensor_assign_2_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_29_internal_tensor_assign_2_begin_mask_0 = const()[name = tensor("new_cache_29_internal_tensor_assign_2_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_29_internal_tensor_assign_2_end_mask_0 = const()[name = tensor("new_cache_29_internal_tensor_assign_2_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_29_internal_tensor_assign_2_squeeze_mask_0 = const()[name = tensor("new_cache_29_internal_tensor_assign_2_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_77 = const()[name = tensor("shape_77"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_29 = const()[name = tensor("reduce_prod_29"), val = tensor(1048576)]; + tensor range_1d_29_start_0 = const()[name = tensor("range_1d_29_start_0"), val = tensor(0)]; + tensor range_1d_29_step_0 = const()[name = tensor("range_1d_29_step_0"), val = tensor(1)]; + tensor range_1d_29 = range_1d(end = reduce_prod_29, start = range_1d_29_start_0, step = range_1d_29_step_0)[name = tensor("range_1d_29")]; + tensor reshape_145 = reshape(shape = shape_77, x = range_1d_29)[name = tensor("reshape_145")]; + tensor slice_by_index_29 = slice_by_index(begin = concat_102, begin_mask = new_cache_29_internal_tensor_assign_2_begin_mask_0, end = concat_103, end_mask = new_cache_29_internal_tensor_assign_2_end_mask_0, squeeze_mask = new_cache_29_internal_tensor_assign_2_squeeze_mask_0, stride = new_cache_29_internal_tensor_assign_2_stride_0, x = reshape_145)[name = tensor("slice_by_index_29")]; + tensor reshape_146_shape_0 = const()[name = tensor("reshape_146_shape_0"), val = tensor([-1])]; + tensor reshape_146 = reshape(shape = reshape_146_shape_0, x = slice_by_index_29)[name = tensor("reshape_146")]; + tensor reshape_147_shape_0 = const()[name = tensor("reshape_147_shape_0"), val = tensor([-1])]; + tensor reshape_147_cast_fp16 = reshape(shape = reshape_147_shape_0, x = var_6352_cast_fp16)[name = tensor("reshape_147_cast_fp16")]; + tensor reshape_148_shape_0 = const()[name = tensor("reshape_148_shape_0"), val = tensor([-1])]; + tensor reshape_148_cast_fp16 = reshape(shape = reshape_148_shape_0, x = reshape_144_cast_fp16)[name = tensor("reshape_148_cast_fp16")]; + tensor scatter_29_mode_0 = const()[name = tensor("scatter_29_mode_0"), val = tensor("update")]; + tensor scatter_29_axis_0 = const()[name = tensor("scatter_29_axis_0"), val = tensor(0)]; + tensor scatter_29_validate_indices_0 = const()[name = tensor("scatter_29_validate_indices_0"), val = tensor(false)]; + tensor scatter_29_cast_fp16 = scatter(axis = scatter_29_axis_0, data = reshape_148_cast_fp16, indices = reshape_146, mode = scatter_29_mode_0, updates = reshape_147_cast_fp16, validate_indices = scatter_29_validate_indices_0)[name = tensor("scatter_29_cast_fp16")]; + tensor reshape_149_cast_fp16 = reshape(shape = shape_77, x = scatter_29_cast_fp16)[name = tensor("reshape_149_cast_fp16")]; + tensor reshape_149_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("reshape_149_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor keys_85_begin_0 = const()[name = tensor("keys_85_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor keys_85_end_0 = const()[name = tensor("keys_85_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor keys_85_end_mask_0 = const()[name = tensor("keys_85_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor keys_85_squeeze_mask_0 = const()[name = tensor("keys_85_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor keys_85_cast_fp16 = slice_by_index(begin = keys_85_begin_0, end = keys_85_end_0, end_mask = keys_85_end_mask_0, squeeze_mask = keys_85_squeeze_mask_0, x = reshape_149_cast_fp16)[name = tensor("keys_85_cast_fp16")]; + tensor values_85_begin_0 = const()[name = tensor("values_85_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor values_85_end_0 = const()[name = tensor("values_85_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor values_85_end_mask_0 = const()[name = tensor("values_85_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor values_85_squeeze_mask_0 = const()[name = tensor("values_85_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor values_85_cast_fp16 = slice_by_index(begin = values_85_begin_0, end = values_85_end_0, end_mask = values_85_end_mask_0, squeeze_mask = values_85_squeeze_mask_0, x = reshape_149_cast_fp16)[name = tensor("values_85_cast_fp16")]; + tensor var_6364_cast_fp16 = not_equal(x = keys_85_cast_fp16, y = keys_85_cast_fp16)[name = tensor("op_6364_cast_fp16")]; + tensor keys_87_cast_fp16 = select(a = var_504_to_fp16, b = keys_85_cast_fp16, cond = var_6364_cast_fp16)[name = tensor("keys_87_cast_fp16")]; + tensor var_6372_cast_fp16 = not_equal(x = values_85_cast_fp16, y = values_85_cast_fp16)[name = tensor("op_6372_cast_fp16")]; + tensor values_87_cast_fp16 = select(a = var_504_to_fp16, b = values_85_cast_fp16, cond = var_6372_cast_fp16)[name = tensor("values_87_cast_fp16")]; + tensor var_6396 = const()[name = tensor("op_6396"), val = tensor([0, 2, 1, 3])]; + tensor var_6409 = const()[name = tensor("op_6409"), val = tensor([1, 1, 1])]; + tensor var_6410_cast_fp16 = reshape(shape = var_6409, x = position14_to_fp16)[name = tensor("op_6410_cast_fp16")]; + tensor var_6427_to_fp16 = const()[name = tensor("op_6427_to_fp16"), val = tensor(0x1p+0)]; + tensor valid_len_29_cast_fp16 = add(x = var_6410_cast_fp16, y = var_6427_to_fp16)[name = tensor("valid_len_29_cast_fp16")]; + tensor valid_mask_29_cast_fp16 = less(x = k_positions_1_promoted_to_fp16, y = valid_len_29_cast_fp16)[name = tensor("valid_mask_29_cast_fp16")]; + tensor causal_mask_29_cast_fp16 = less_equal(x = k_positions_1_promoted_to_fp16, y = var_6410_cast_fp16)[name = tensor("causal_mask_29_cast_fp16")]; + tensor attn_mask_57 = logical_and(x = valid_mask_29_cast_fp16, y = causal_mask_29_cast_fp16)[name = tensor("attn_mask_57")]; + tensor attn_mask_59_axes_0 = const()[name = tensor("attn_mask_59_axes_0"), val = tensor([1])]; + tensor attn_mask_59 = expand_dims(axes = attn_mask_59_axes_0, x = attn_mask_57)[name = tensor("attn_mask_59")]; + tensor var_6445_transpose_x_0 = const()[name = tensor("op_6445_transpose_x_0"), val = tensor(false)]; + tensor var_6445_transpose_y_0 = const()[name = tensor("op_6445_transpose_y_0"), val = tensor(false)]; + tensor transpose_100_perm_0 = const()[name = tensor("transpose_100_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_101_perm_0 = const()[name = tensor("transpose_101_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_101 = transpose(perm = transpose_101_perm_0, x = keys_87_cast_fp16)[name = tensor("transpose_157")]; + tensor transpose_100 = transpose(perm = transpose_100_perm_0, x = q_87_cast_fp16)[name = tensor("transpose_158")]; + tensor var_6445_cast_fp16 = matmul(transpose_x = var_6445_transpose_x_0, transpose_y = var_6445_transpose_y_0, x = transpose_100, y = transpose_101)[name = tensor("op_6445_cast_fp16")]; + tensor var_6439_to_fp16 = const()[name = tensor("op_6439_to_fp16"), val = tensor([0x1p-3])]; + tensor attn_weights_85_cast_fp16 = mul(x = var_6445_cast_fp16, y = var_6439_to_fp16)[name = tensor("attn_weights_85_cast_fp16")]; + tensor var_6447 = logical_not(x = attn_mask_59)[name = tensor("op_6447")]; + tensor var_6448_to_fp16 = const()[name = tensor("op_6448_to_fp16"), val = tensor(-inf)]; + tensor attn_weights_87_cast_fp16 = select(a = var_6448_to_fp16, b = attn_weights_85_cast_fp16, cond = var_6447)[name = tensor("attn_weights_87_cast_fp16")]; + tensor var_6450 = const()[name = tensor("op_6450"), val = tensor(-1)]; + tensor attn_weights_89_cast_fp16 = softmax(axis = var_6450, x = attn_weights_87_cast_fp16)[name = tensor("attn_weights_89_cast_fp16")]; + tensor attn_output_29_transpose_x_0 = const()[name = tensor("attn_output_29_transpose_x_0"), val = tensor(false)]; + tensor attn_output_29_transpose_y_0 = const()[name = tensor("attn_output_29_transpose_y_0"), val = tensor(false)]; + tensor values_89_cast_fp16 = transpose(perm = var_6396, x = values_87_cast_fp16)[name = tensor("transpose_159")]; + tensor attn_output_29_cast_fp16 = matmul(transpose_x = attn_output_29_transpose_x_0, transpose_y = attn_output_29_transpose_y_0, x = attn_weights_89_cast_fp16, y = values_89_cast_fp16)[name = tensor("attn_output_29_cast_fp16")]; + tensor var_6458 = const()[name = tensor("op_6458"), val = tensor([0, 2, 1, 3])]; + tensor var_6461 = const()[name = tensor("op_6461"), val = tensor([1, 1, 1024])]; + tensor var_6459_cast_fp16 = transpose(perm = var_6458, x = attn_output_29_cast_fp16)[name = tensor("transpose_156")]; + tensor input_145_cast_fp16 = reshape(shape = var_6461, x = var_6459_cast_fp16)[name = tensor("input_145_cast_fp16")]; + tensor linear_58_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = attn14_out_proj_weight_cast_fp16, x = input_145_cast_fp16)[name = tensor("linear_58_cast_fp16")]; + tensor var_6467_to_fp16 = const()[name = tensor("op_6467_to_fp16"), val = tensor(0x1p+0)]; + tensor var_6468_cast_fp16 = add(x = position14_to_fp16, y = var_6467_to_fp16)[name = tensor("op_6468_cast_fp16")]; + tensor var_6468_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_6468_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor input_147_cast_fp16 = add(x = input_143_cast_fp16, y = linear_58_cast_fp16)[name = tensor("input_147_cast_fp16")]; + tensor input_149_axes_0 = const()[name = tensor("input_149_axes_0"), val = tensor([-1])]; + tensor norm14_2_weight_to_fp16 = const()[name = tensor("norm14_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(303709312)))]; + tensor norm14_2_bias_to_fp16 = const()[name = tensor("norm14_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(303711424)))]; + tensor var_6472_to_fp16 = const()[name = tensor("op_6472_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_149_cast_fp16 = layer_norm(axes = input_149_axes_0, beta = norm14_2_bias_to_fp16, epsilon = var_6472_to_fp16, gamma = norm14_2_weight_to_fp16, x = input_147_cast_fp16)[name = tensor("input_149_cast_fp16")]; + tensor linear_59_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = linear14_1_weight_cast_fp16, x = input_149_cast_fp16)[name = tensor("linear_59_cast_fp16")]; + tensor input_151_mode_0 = const()[name = tensor("input_151_mode_0"), val = tensor("EXACT")]; + tensor input_151_cast_fp16 = gelu(mode = input_151_mode_0, x = linear_59_cast_fp16)[name = tensor("input_151_cast_fp16")]; + tensor linear_60_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = linear14_2_weight_cast_fp16, x = input_151_cast_fp16)[name = tensor("linear_60_cast_fp16")]; + tensor input_153_cast_fp16 = add(x = input_147_cast_fp16, y = linear_60_cast_fp16)[name = tensor("input_153_cast_fp16")]; + tensor x_31_axes_0 = const()[name = tensor("x_31_axes_0"), val = tensor([-1])]; + tensor norm15_1_weight_to_fp16 = const()[name = tensor("norm15_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(303713536)))]; + tensor norm15_1_bias_to_fp16 = const()[name = tensor("norm15_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(303715648)))]; + tensor var_6489_to_fp16 = const()[name = tensor("op_6489_to_fp16"), val = tensor(0x1.5p-17)]; + tensor x_31_cast_fp16 = layer_norm(axes = x_31_axes_0, beta = norm15_1_bias_to_fp16, epsilon = var_6489_to_fp16, gamma = norm15_1_weight_to_fp16, x = input_153_cast_fp16)[name = tensor("x_31_cast_fp16")]; + tensor linear_61_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = attn15_in_proj_weight_cast_fp16, x = x_31_cast_fp16)[name = tensor("linear_61_cast_fp16")]; + tensor var_6525 = const()[name = tensor("op_6525"), val = tensor([1, 1, 3, 16, 64])]; + tensor qkv_31_cast_fp16 = reshape(shape = var_6525, x = linear_61_cast_fp16)[name = tensor("qkv_31_cast_fp16")]; + tensor q_91_begin_0 = const()[name = tensor("q_91_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor q_91_end_0 = const()[name = tensor("q_91_end_0"), val = tensor([1, 1, 1, 16, 64])]; + tensor q_91_end_mask_0 = const()[name = tensor("q_91_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor q_91_squeeze_mask_0 = const()[name = tensor("q_91_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor q_91_cast_fp16 = slice_by_index(begin = q_91_begin_0, end = q_91_end_0, end_mask = q_91_end_mask_0, squeeze_mask = q_91_squeeze_mask_0, x = qkv_31_cast_fp16)[name = tensor("q_91_cast_fp16")]; + tensor k_61_begin_0 = const()[name = tensor("k_61_begin_0"), val = tensor([0, 0, 1, 0, 0])]; + tensor k_61_end_0 = const()[name = tensor("k_61_end_0"), val = tensor([1, 1, 2, 16, 64])]; + tensor k_61_end_mask_0 = const()[name = tensor("k_61_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor k_61_squeeze_mask_0 = const()[name = tensor("k_61_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor k_61_cast_fp16 = slice_by_index(begin = k_61_begin_0, end = k_61_end_0, end_mask = k_61_end_mask_0, squeeze_mask = k_61_squeeze_mask_0, x = qkv_31_cast_fp16)[name = tensor("k_61_cast_fp16")]; + tensor v_31_begin_0 = const()[name = tensor("v_31_begin_0"), val = tensor([0, 0, 2, 0, 0])]; + tensor v_31_end_0 = const()[name = tensor("v_31_end_0"), val = tensor([1, 1, 3, 16, 64])]; + tensor v_31_end_mask_0 = const()[name = tensor("v_31_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor v_31_squeeze_mask_0 = const()[name = tensor("v_31_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor v_31_cast_fp16 = slice_by_index(begin = v_31_begin_0, end = v_31_end_0, end_mask = v_31_end_mask_0, squeeze_mask = v_31_squeeze_mask_0, x = qkv_31_cast_fp16)[name = tensor("v_31_cast_fp16")]; + tensor var_6629 = const()[name = tensor("op_6629"), val = tensor([1, 1, 1, 1])]; + tensor position15_to_fp16_dtype_0 = const()[name = tensor("position15_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor position15_to_fp16 = cast(dtype = position15_to_fp16_dtype_0, x = position15)[name = tensor("cast_478")]; + tensor ts_95_cast_fp16 = reshape(shape = var_6629, x = position15_to_fp16)[name = tensor("ts_95_cast_fp16")]; + tensor var_6633 = const()[name = tensor("op_6633"), val = tensor([1, 1, 16, 32, 2])]; + tensor q_complex_31_cast_fp16 = reshape(shape = var_6633, x = q_91_cast_fp16)[name = tensor("q_complex_31_cast_fp16")]; + tensor var_6637 = const()[name = tensor("op_6637"), val = tensor([1, 1, 16, 32, 2])]; + tensor k_complex_31_cast_fp16 = reshape(shape = var_6637, x = k_61_cast_fp16)[name = tensor("k_complex_31_cast_fp16")]; + tensor var_6641_begin_0 = const()[name = tensor("op_6641_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_6641_end_0 = const()[name = tensor("op_6641_end_0"), val = tensor([1, 1, 16, 32, 1])]; + tensor var_6641_end_mask_0 = const()[name = tensor("op_6641_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_6641_squeeze_mask_0 = const()[name = tensor("op_6641_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_6641_cast_fp16 = slice_by_index(begin = var_6641_begin_0, end = var_6641_end_0, end_mask = var_6641_end_mask_0, squeeze_mask = var_6641_squeeze_mask_0, x = q_complex_31_cast_fp16)[name = tensor("op_6641_cast_fp16")]; + tensor var_6649_begin_0 = const()[name = tensor("op_6649_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_6649_end_0 = const()[name = tensor("op_6649_end_0"), val = tensor([1, 1, 16, 32, 2])]; + tensor var_6649_end_mask_0 = const()[name = tensor("op_6649_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_6649_squeeze_mask_0 = const()[name = tensor("op_6649_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_6649_cast_fp16 = slice_by_index(begin = var_6649_begin_0, end = var_6649_end_0, end_mask = var_6649_end_mask_0, squeeze_mask = var_6649_squeeze_mask_0, x = q_complex_31_cast_fp16)[name = tensor("op_6649_cast_fp16")]; + tensor var_6657_begin_0 = const()[name = tensor("op_6657_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_6657_end_0 = const()[name = tensor("op_6657_end_0"), val = tensor([1, 1, 16, 32, 1])]; + tensor var_6657_end_mask_0 = const()[name = tensor("op_6657_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_6657_squeeze_mask_0 = const()[name = tensor("op_6657_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_6657_cast_fp16 = slice_by_index(begin = var_6657_begin_0, end = var_6657_end_0, end_mask = var_6657_end_mask_0, squeeze_mask = var_6657_squeeze_mask_0, x = k_complex_31_cast_fp16)[name = tensor("op_6657_cast_fp16")]; + tensor var_6665_begin_0 = const()[name = tensor("op_6665_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_6665_end_0 = const()[name = tensor("op_6665_end_0"), val = tensor([1, 1, 16, 32, 2])]; + tensor var_6665_end_mask_0 = const()[name = tensor("op_6665_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_6665_squeeze_mask_0 = const()[name = tensor("op_6665_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_6665_cast_fp16 = slice_by_index(begin = var_6665_begin_0, end = var_6665_end_0, end_mask = var_6665_end_mask_0, squeeze_mask = var_6665_squeeze_mask_0, x = k_complex_31_cast_fp16)[name = tensor("op_6665_cast_fp16")]; + tensor freqs_31_to_fp16 = const()[name = tensor("freqs_31_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(303717760)))]; + tensor var_6671_cast_fp16 = mul(x = freqs_31_to_fp16, y = ts_95_cast_fp16)[name = tensor("op_6671_cast_fp16")]; + tensor rotr_31_cast_fp16 = cos(x = var_6671_cast_fp16)[name = tensor("rotr_31_cast_fp16")]; + tensor roti_31_cast_fp16 = sin(x = var_6671_cast_fp16)[name = tensor("roti_31_cast_fp16")]; + tensor var_6675_cast_fp16 = mul(x = var_6641_cast_fp16, y = rotr_31_cast_fp16)[name = tensor("op_6675_cast_fp16")]; + tensor var_6676_cast_fp16 = mul(x = var_6649_cast_fp16, y = roti_31_cast_fp16)[name = tensor("op_6676_cast_fp16")]; + tensor qor_61_cast_fp16 = sub(x = var_6675_cast_fp16, y = var_6676_cast_fp16)[name = tensor("qor_61_cast_fp16")]; + tensor var_6679_cast_fp16 = mul(x = var_6641_cast_fp16, y = roti_31_cast_fp16)[name = tensor("op_6679_cast_fp16")]; + tensor var_6680_cast_fp16 = mul(x = var_6649_cast_fp16, y = rotr_31_cast_fp16)[name = tensor("op_6680_cast_fp16")]; + tensor qoi_61_cast_fp16 = add(x = var_6679_cast_fp16, y = var_6680_cast_fp16)[name = tensor("qoi_61_cast_fp16")]; + tensor var_6683_cast_fp16 = mul(x = var_6657_cast_fp16, y = rotr_31_cast_fp16)[name = tensor("op_6683_cast_fp16")]; + tensor var_6684_cast_fp16 = mul(x = var_6665_cast_fp16, y = roti_31_cast_fp16)[name = tensor("op_6684_cast_fp16")]; + tensor kor_61_cast_fp16 = sub(x = var_6683_cast_fp16, y = var_6684_cast_fp16)[name = tensor("kor_61_cast_fp16")]; + tensor var_6687_cast_fp16 = mul(x = var_6657_cast_fp16, y = roti_31_cast_fp16)[name = tensor("op_6687_cast_fp16")]; + tensor var_6688_cast_fp16 = mul(x = var_6665_cast_fp16, y = rotr_31_cast_fp16)[name = tensor("op_6688_cast_fp16")]; + tensor koi_61_cast_fp16 = add(x = var_6687_cast_fp16, y = var_6688_cast_fp16)[name = tensor("koi_61_cast_fp16")]; + tensor qo_31_axis_0 = const()[name = tensor("qo_31_axis_0"), val = tensor(-1)]; + tensor qo_31_cast_fp16 = stack(axis = qo_31_axis_0, values = (qor_61_cast_fp16, qoi_61_cast_fp16))[name = tensor("qo_31_cast_fp16")]; + tensor ko_31_axis_0 = const()[name = tensor("ko_31_axis_0"), val = tensor(-1)]; + tensor ko_31_cast_fp16 = stack(axis = ko_31_axis_0, values = (kor_61_cast_fp16, koi_61_cast_fp16))[name = tensor("ko_31_cast_fp16")]; + tensor var_6717 = const()[name = tensor("op_6717"), val = tensor([1, 1, 16, 64])]; + tensor q_93_cast_fp16 = reshape(shape = var_6717, x = qo_31_cast_fp16)[name = tensor("q_93_cast_fp16")]; + tensor var_6719 = const()[name = tensor("op_6719"), val = tensor([1, 1, 16, 64])]; + tensor k_63_cast_fp16 = reshape(shape = var_6719, x = ko_31_cast_fp16)[name = tensor("k_63_cast_fp16")]; + tensor _inversed_6741_y_0_to_fp16 = const()[name = tensor("_inversed_6741_y_0_to_fp16"), val = tensor(0x1p-9)]; + tensor _inversed_6741_cast_fp16 = mul(x = ts_95_cast_fp16, y = _inversed_6741_y_0_to_fp16)[name = tensor("_inversed_6741_cast_fp16")]; + tensor var_6742_cast_fp16 = floor(x = _inversed_6741_cast_fp16)[name = tensor("op_6742_cast_fp16")]; + tensor var_6743_to_fp16 = const()[name = tensor("op_6743_to_fp16"), val = tensor(0x1p+9)]; + tensor var_6744_cast_fp16 = mul(x = var_6742_cast_fp16, y = var_6743_to_fp16)[name = tensor("op_6744_cast_fp16")]; + tensor write_indices_float_63_cast_fp16 = sub(x = ts_95_cast_fp16, y = var_6744_cast_fp16)[name = tensor("write_indices_float_63_cast_fp16")]; + tensor var_6751_dtype_0 = const()[name = tensor("op_6751_dtype_0"), val = tensor("int32")]; + tensor write_indices_31_reps_0 = const()[name = tensor("write_indices_31_reps_0"), val = tensor([1, 1, 16, 64])]; + tensor write_indices_float_63_cast_fp16_to_int32 = cast(dtype = var_6751_dtype_0, x = write_indices_float_63_cast_fp16)[name = tensor("cast_477")]; + tensor write_indices_31 = tile(reps = write_indices_31_reps_0, x = write_indices_float_63_cast_fp16_to_int32)[name = tensor("write_indices_31")]; + tensor var_6759_begin_0 = const()[name = tensor("op_6759_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_6759_end_0 = const()[name = tensor("op_6759_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor var_6759_end_mask_0 = const()[name = tensor("op_6759_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_6759_squeeze_mask_0 = const()[name = tensor("op_6759_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor cache15_to_fp16_dtype_0 = const()[name = tensor("cache15_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor cache15_to_fp16 = cast(dtype = cache15_to_fp16_dtype_0, x = cache15)[name = tensor("cast_476")]; + tensor var_6759_cast_fp16 = slice_by_index(begin = var_6759_begin_0, end = var_6759_end_0, end_mask = var_6759_end_mask_0, squeeze_mask = var_6759_squeeze_mask_0, x = cache15_to_fp16)[name = tensor("op_6759_cast_fp16")]; + tensor var_6761_axis_0 = const()[name = tensor("op_6761_axis_0"), val = tensor(1)]; + tensor var_6761_mode_0 = const()[name = tensor("op_6761_mode_0"), val = tensor("update")]; + tensor var_6761_validate_indices_0 = const()[name = tensor("op_6761_validate_indices_0"), val = tensor(false)]; + tensor var_6761_cast_fp16 = scatter_along_axis(axis = var_6761_axis_0, data = var_6759_cast_fp16, indices = write_indices_31, mode = var_6761_mode_0, updates = k_63_cast_fp16, validate_indices = var_6761_validate_indices_0)[name = tensor("op_6761_cast_fp16")]; + tensor concat_107 = const()[name = tensor("concat_107"), val = tensor([0, 0, 0, 0, 0])]; + tensor concat_108 = const()[name = tensor("concat_108"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_31_internal_tensor_assign_1_stride_0 = const()[name = tensor("new_cache_31_internal_tensor_assign_1_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_31_internal_tensor_assign_1_begin_mask_0 = const()[name = tensor("new_cache_31_internal_tensor_assign_1_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_31_internal_tensor_assign_1_end_mask_0 = const()[name = tensor("new_cache_31_internal_tensor_assign_1_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_31_internal_tensor_assign_1_squeeze_mask_0 = const()[name = tensor("new_cache_31_internal_tensor_assign_1_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_78 = const()[name = tensor("shape_78"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_30 = const()[name = tensor("reduce_prod_30"), val = tensor(1048576)]; + tensor range_1d_30_start_0 = const()[name = tensor("range_1d_30_start_0"), val = tensor(0)]; + tensor range_1d_30_step_0 = const()[name = tensor("range_1d_30_step_0"), val = tensor(1)]; + tensor range_1d_30 = range_1d(end = reduce_prod_30, start = range_1d_30_start_0, step = range_1d_30_step_0)[name = tensor("range_1d_30")]; + tensor reshape_150 = reshape(shape = shape_78, x = range_1d_30)[name = tensor("reshape_150")]; + tensor slice_by_index_30 = slice_by_index(begin = concat_107, begin_mask = new_cache_31_internal_tensor_assign_1_begin_mask_0, end = concat_108, end_mask = new_cache_31_internal_tensor_assign_1_end_mask_0, squeeze_mask = new_cache_31_internal_tensor_assign_1_squeeze_mask_0, stride = new_cache_31_internal_tensor_assign_1_stride_0, x = reshape_150)[name = tensor("slice_by_index_30")]; + tensor reshape_151_shape_0 = const()[name = tensor("reshape_151_shape_0"), val = tensor([-1])]; + tensor reshape_151 = reshape(shape = reshape_151_shape_0, x = slice_by_index_30)[name = tensor("reshape_151")]; + tensor reshape_152_shape_0 = const()[name = tensor("reshape_152_shape_0"), val = tensor([-1])]; + tensor reshape_152_cast_fp16 = reshape(shape = reshape_152_shape_0, x = var_6761_cast_fp16)[name = tensor("reshape_152_cast_fp16")]; + tensor reshape_153_shape_0 = const()[name = tensor("reshape_153_shape_0"), val = tensor([-1])]; + tensor reshape_153_cast_fp16 = reshape(shape = reshape_153_shape_0, x = cache15_to_fp16)[name = tensor("reshape_153_cast_fp16")]; + tensor scatter_30_mode_0 = const()[name = tensor("scatter_30_mode_0"), val = tensor("update")]; + tensor scatter_30_axis_0 = const()[name = tensor("scatter_30_axis_0"), val = tensor(0)]; + tensor scatter_30_validate_indices_0 = const()[name = tensor("scatter_30_validate_indices_0"), val = tensor(false)]; + tensor scatter_30_cast_fp16 = scatter(axis = scatter_30_axis_0, data = reshape_153_cast_fp16, indices = reshape_151, mode = scatter_30_mode_0, updates = reshape_152_cast_fp16, validate_indices = scatter_30_validate_indices_0)[name = tensor("scatter_30_cast_fp16")]; + tensor reshape_154_cast_fp16 = reshape(shape = shape_78, x = scatter_30_cast_fp16)[name = tensor("reshape_154_cast_fp16")]; + tensor var_6769_begin_0 = const()[name = tensor("op_6769_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor var_6769_end_0 = const()[name = tensor("op_6769_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor var_6769_end_mask_0 = const()[name = tensor("op_6769_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_6769_squeeze_mask_0 = const()[name = tensor("op_6769_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor var_6769_cast_fp16 = slice_by_index(begin = var_6769_begin_0, end = var_6769_end_0, end_mask = var_6769_end_mask_0, squeeze_mask = var_6769_squeeze_mask_0, x = reshape_154_cast_fp16)[name = tensor("op_6769_cast_fp16")]; + tensor var_6771_axis_0 = const()[name = tensor("op_6771_axis_0"), val = tensor(1)]; + tensor var_6771_mode_0 = const()[name = tensor("op_6771_mode_0"), val = tensor("update")]; + tensor var_6771_validate_indices_0 = const()[name = tensor("op_6771_validate_indices_0"), val = tensor(false)]; + tensor var_6771_cast_fp16 = scatter_along_axis(axis = var_6771_axis_0, data = var_6769_cast_fp16, indices = write_indices_31, mode = var_6771_mode_0, updates = v_31_cast_fp16, validate_indices = var_6771_validate_indices_0)[name = tensor("op_6771_cast_fp16")]; + tensor concat_109 = const()[name = tensor("concat_109"), val = tensor([1, 0, 0, 0, 0])]; + tensor concat_110 = const()[name = tensor("concat_110"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_31_internal_tensor_assign_2_stride_0 = const()[name = tensor("new_cache_31_internal_tensor_assign_2_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_31_internal_tensor_assign_2_begin_mask_0 = const()[name = tensor("new_cache_31_internal_tensor_assign_2_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_31_internal_tensor_assign_2_end_mask_0 = const()[name = tensor("new_cache_31_internal_tensor_assign_2_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_31_internal_tensor_assign_2_squeeze_mask_0 = const()[name = tensor("new_cache_31_internal_tensor_assign_2_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_79 = const()[name = tensor("shape_79"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_31 = const()[name = tensor("reduce_prod_31"), val = tensor(1048576)]; + tensor range_1d_31_start_0 = const()[name = tensor("range_1d_31_start_0"), val = tensor(0)]; + tensor range_1d_31_step_0 = const()[name = tensor("range_1d_31_step_0"), val = tensor(1)]; + tensor range_1d_31 = range_1d(end = reduce_prod_31, start = range_1d_31_start_0, step = range_1d_31_step_0)[name = tensor("range_1d_31")]; + tensor reshape_155 = reshape(shape = shape_79, x = range_1d_31)[name = tensor("reshape_155")]; + tensor slice_by_index_31 = slice_by_index(begin = concat_109, begin_mask = new_cache_31_internal_tensor_assign_2_begin_mask_0, end = concat_110, end_mask = new_cache_31_internal_tensor_assign_2_end_mask_0, squeeze_mask = new_cache_31_internal_tensor_assign_2_squeeze_mask_0, stride = new_cache_31_internal_tensor_assign_2_stride_0, x = reshape_155)[name = tensor("slice_by_index_31")]; + tensor reshape_156_shape_0 = const()[name = tensor("reshape_156_shape_0"), val = tensor([-1])]; + tensor reshape_156 = reshape(shape = reshape_156_shape_0, x = slice_by_index_31)[name = tensor("reshape_156")]; + tensor reshape_157_shape_0 = const()[name = tensor("reshape_157_shape_0"), val = tensor([-1])]; + tensor reshape_157_cast_fp16 = reshape(shape = reshape_157_shape_0, x = var_6771_cast_fp16)[name = tensor("reshape_157_cast_fp16")]; + tensor reshape_158_shape_0 = const()[name = tensor("reshape_158_shape_0"), val = tensor([-1])]; + tensor reshape_158_cast_fp16 = reshape(shape = reshape_158_shape_0, x = reshape_154_cast_fp16)[name = tensor("reshape_158_cast_fp16")]; + tensor scatter_31_mode_0 = const()[name = tensor("scatter_31_mode_0"), val = tensor("update")]; + tensor scatter_31_axis_0 = const()[name = tensor("scatter_31_axis_0"), val = tensor(0)]; + tensor scatter_31_validate_indices_0 = const()[name = tensor("scatter_31_validate_indices_0"), val = tensor(false)]; + tensor scatter_31_cast_fp16 = scatter(axis = scatter_31_axis_0, data = reshape_158_cast_fp16, indices = reshape_156, mode = scatter_31_mode_0, updates = reshape_157_cast_fp16, validate_indices = scatter_31_validate_indices_0)[name = tensor("scatter_31_cast_fp16")]; + tensor reshape_159_cast_fp16 = reshape(shape = shape_79, x = scatter_31_cast_fp16)[name = tensor("reshape_159_cast_fp16")]; + tensor reshape_159_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("reshape_159_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor keys_91_begin_0 = const()[name = tensor("keys_91_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor keys_91_end_0 = const()[name = tensor("keys_91_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor keys_91_end_mask_0 = const()[name = tensor("keys_91_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor keys_91_squeeze_mask_0 = const()[name = tensor("keys_91_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor keys_91_cast_fp16 = slice_by_index(begin = keys_91_begin_0, end = keys_91_end_0, end_mask = keys_91_end_mask_0, squeeze_mask = keys_91_squeeze_mask_0, x = reshape_159_cast_fp16)[name = tensor("keys_91_cast_fp16")]; + tensor values_91_begin_0 = const()[name = tensor("values_91_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor values_91_end_0 = const()[name = tensor("values_91_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor values_91_end_mask_0 = const()[name = tensor("values_91_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor values_91_squeeze_mask_0 = const()[name = tensor("values_91_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor values_91_cast_fp16 = slice_by_index(begin = values_91_begin_0, end = values_91_end_0, end_mask = values_91_end_mask_0, squeeze_mask = values_91_squeeze_mask_0, x = reshape_159_cast_fp16)[name = tensor("values_91_cast_fp16")]; + tensor var_6783_cast_fp16 = not_equal(x = keys_91_cast_fp16, y = keys_91_cast_fp16)[name = tensor("op_6783_cast_fp16")]; + tensor keys_93_cast_fp16 = select(a = var_504_to_fp16, b = keys_91_cast_fp16, cond = var_6783_cast_fp16)[name = tensor("keys_93_cast_fp16")]; + tensor var_6791_cast_fp16 = not_equal(x = values_91_cast_fp16, y = values_91_cast_fp16)[name = tensor("op_6791_cast_fp16")]; + tensor values_93_cast_fp16 = select(a = var_504_to_fp16, b = values_91_cast_fp16, cond = var_6791_cast_fp16)[name = tensor("values_93_cast_fp16")]; + tensor var_6815 = const()[name = tensor("op_6815"), val = tensor([0, 2, 1, 3])]; + tensor var_6828 = const()[name = tensor("op_6828"), val = tensor([1, 1, 1])]; + tensor var_6829_cast_fp16 = reshape(shape = var_6828, x = position15_to_fp16)[name = tensor("op_6829_cast_fp16")]; + tensor var_6846_to_fp16 = const()[name = tensor("op_6846_to_fp16"), val = tensor(0x1p+0)]; + tensor valid_len_31_cast_fp16 = add(x = var_6829_cast_fp16, y = var_6846_to_fp16)[name = tensor("valid_len_31_cast_fp16")]; + tensor valid_mask_31_cast_fp16 = less(x = k_positions_1_promoted_to_fp16, y = valid_len_31_cast_fp16)[name = tensor("valid_mask_31_cast_fp16")]; + tensor causal_mask_31_cast_fp16 = less_equal(x = k_positions_1_promoted_to_fp16, y = var_6829_cast_fp16)[name = tensor("causal_mask_31_cast_fp16")]; + tensor attn_mask_61 = logical_and(x = valid_mask_31_cast_fp16, y = causal_mask_31_cast_fp16)[name = tensor("attn_mask_61")]; + tensor attn_mask_63_axes_0 = const()[name = tensor("attn_mask_63_axes_0"), val = tensor([1])]; + tensor attn_mask_63 = expand_dims(axes = attn_mask_63_axes_0, x = attn_mask_61)[name = tensor("attn_mask_63")]; + tensor var_6864_transpose_x_0 = const()[name = tensor("op_6864_transpose_x_0"), val = tensor(false)]; + tensor var_6864_transpose_y_0 = const()[name = tensor("op_6864_transpose_y_0"), val = tensor(false)]; + tensor transpose_102_perm_0 = const()[name = tensor("transpose_102_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_103_perm_0 = const()[name = tensor("transpose_103_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_103 = transpose(perm = transpose_103_perm_0, x = keys_93_cast_fp16)[name = tensor("transpose_153")]; + tensor transpose_102 = transpose(perm = transpose_102_perm_0, x = q_93_cast_fp16)[name = tensor("transpose_154")]; + tensor var_6864_cast_fp16 = matmul(transpose_x = var_6864_transpose_x_0, transpose_y = var_6864_transpose_y_0, x = transpose_102, y = transpose_103)[name = tensor("op_6864_cast_fp16")]; + tensor var_6858_to_fp16 = const()[name = tensor("op_6858_to_fp16"), val = tensor([0x1p-3])]; + tensor attn_weights_91_cast_fp16 = mul(x = var_6864_cast_fp16, y = var_6858_to_fp16)[name = tensor("attn_weights_91_cast_fp16")]; + tensor var_6866 = logical_not(x = attn_mask_63)[name = tensor("op_6866")]; + tensor var_6867_to_fp16 = const()[name = tensor("op_6867_to_fp16"), val = tensor(-inf)]; + tensor attn_weights_93_cast_fp16 = select(a = var_6867_to_fp16, b = attn_weights_91_cast_fp16, cond = var_6866)[name = tensor("attn_weights_93_cast_fp16")]; + tensor var_6869 = const()[name = tensor("op_6869"), val = tensor(-1)]; + tensor attn_weights_95_cast_fp16 = softmax(axis = var_6869, x = attn_weights_93_cast_fp16)[name = tensor("attn_weights_95_cast_fp16")]; + tensor attn_output_31_transpose_x_0 = const()[name = tensor("attn_output_31_transpose_x_0"), val = tensor(false)]; + tensor attn_output_31_transpose_y_0 = const()[name = tensor("attn_output_31_transpose_y_0"), val = tensor(false)]; + tensor values_95_cast_fp16 = transpose(perm = var_6815, x = values_93_cast_fp16)[name = tensor("transpose_155")]; + tensor attn_output_31_cast_fp16 = matmul(transpose_x = attn_output_31_transpose_x_0, transpose_y = attn_output_31_transpose_y_0, x = attn_weights_95_cast_fp16, y = values_95_cast_fp16)[name = tensor("attn_output_31_cast_fp16")]; + tensor var_6877 = const()[name = tensor("op_6877"), val = tensor([0, 2, 1, 3])]; + tensor var_6880 = const()[name = tensor("op_6880"), val = tensor([1, 1, 1024])]; + tensor var_6878_cast_fp16 = transpose(perm = var_6877, x = attn_output_31_cast_fp16)[name = tensor("transpose_152")]; + tensor input_155_cast_fp16 = reshape(shape = var_6880, x = var_6878_cast_fp16)[name = tensor("input_155_cast_fp16")]; + tensor linear_62_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = attn15_out_proj_weight_cast_fp16, x = input_155_cast_fp16)[name = tensor("linear_62_cast_fp16")]; + tensor var_6886_to_fp16 = const()[name = tensor("op_6886_to_fp16"), val = tensor(0x1p+0)]; + tensor var_6887_cast_fp16 = add(x = position15_to_fp16, y = var_6886_to_fp16)[name = tensor("op_6887_cast_fp16")]; + tensor var_6887_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_6887_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor input_157_cast_fp16 = add(x = input_153_cast_fp16, y = linear_62_cast_fp16)[name = tensor("input_157_cast_fp16")]; + tensor input_159_axes_0 = const()[name = tensor("input_159_axes_0"), val = tensor([-1])]; + tensor norm15_2_weight_to_fp16 = const()[name = tensor("norm15_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(303717888)))]; + tensor norm15_2_bias_to_fp16 = const()[name = tensor("norm15_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(303720000)))]; + tensor var_6891_to_fp16 = const()[name = tensor("op_6891_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_159_cast_fp16 = layer_norm(axes = input_159_axes_0, beta = norm15_2_bias_to_fp16, epsilon = var_6891_to_fp16, gamma = norm15_2_weight_to_fp16, x = input_157_cast_fp16)[name = tensor("input_159_cast_fp16")]; + tensor linear_63_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = linear15_1_weight_cast_fp16, x = input_159_cast_fp16)[name = tensor("linear_63_cast_fp16")]; + tensor input_161_mode_0 = const()[name = tensor("input_161_mode_0"), val = tensor("EXACT")]; + tensor input_161_cast_fp16 = gelu(mode = input_161_mode_0, x = linear_63_cast_fp16)[name = tensor("input_161_cast_fp16")]; + tensor linear_64_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = linear15_2_weight_cast_fp16, x = input_161_cast_fp16)[name = tensor("linear_64_cast_fp16")]; + tensor input_163_cast_fp16 = add(x = input_157_cast_fp16, y = linear_64_cast_fp16)[name = tensor("input_163_cast_fp16")]; + tensor x_33_axes_0 = const()[name = tensor("x_33_axes_0"), val = tensor([-1])]; + tensor norm16_1_weight_to_fp16 = const()[name = tensor("norm16_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(303722112)))]; + tensor norm16_1_bias_to_fp16 = const()[name = tensor("norm16_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(303724224)))]; + tensor var_6908_to_fp16 = const()[name = tensor("op_6908_to_fp16"), val = tensor(0x1.5p-17)]; + tensor x_33_cast_fp16 = layer_norm(axes = x_33_axes_0, beta = norm16_1_bias_to_fp16, epsilon = var_6908_to_fp16, gamma = norm16_1_weight_to_fp16, x = input_163_cast_fp16)[name = tensor("x_33_cast_fp16")]; + tensor linear_65_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = attn16_in_proj_weight_cast_fp16, x = x_33_cast_fp16)[name = tensor("linear_65_cast_fp16")]; + tensor var_6944 = const()[name = tensor("op_6944"), val = tensor([1, 1, 3, 16, 64])]; + tensor qkv_33_cast_fp16 = reshape(shape = var_6944, x = linear_65_cast_fp16)[name = tensor("qkv_33_cast_fp16")]; + tensor q_97_begin_0 = const()[name = tensor("q_97_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor q_97_end_0 = const()[name = tensor("q_97_end_0"), val = tensor([1, 1, 1, 16, 64])]; + tensor q_97_end_mask_0 = const()[name = tensor("q_97_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor q_97_squeeze_mask_0 = const()[name = tensor("q_97_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor q_97_cast_fp16 = slice_by_index(begin = q_97_begin_0, end = q_97_end_0, end_mask = q_97_end_mask_0, squeeze_mask = q_97_squeeze_mask_0, x = qkv_33_cast_fp16)[name = tensor("q_97_cast_fp16")]; + tensor k_65_begin_0 = const()[name = tensor("k_65_begin_0"), val = tensor([0, 0, 1, 0, 0])]; + tensor k_65_end_0 = const()[name = tensor("k_65_end_0"), val = tensor([1, 1, 2, 16, 64])]; + tensor k_65_end_mask_0 = const()[name = tensor("k_65_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor k_65_squeeze_mask_0 = const()[name = tensor("k_65_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor k_65_cast_fp16 = slice_by_index(begin = k_65_begin_0, end = k_65_end_0, end_mask = k_65_end_mask_0, squeeze_mask = k_65_squeeze_mask_0, x = qkv_33_cast_fp16)[name = tensor("k_65_cast_fp16")]; + tensor v_33_begin_0 = const()[name = tensor("v_33_begin_0"), val = tensor([0, 0, 2, 0, 0])]; + tensor v_33_end_0 = const()[name = tensor("v_33_end_0"), val = tensor([1, 1, 3, 16, 64])]; + tensor v_33_end_mask_0 = const()[name = tensor("v_33_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor v_33_squeeze_mask_0 = const()[name = tensor("v_33_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor v_33_cast_fp16 = slice_by_index(begin = v_33_begin_0, end = v_33_end_0, end_mask = v_33_end_mask_0, squeeze_mask = v_33_squeeze_mask_0, x = qkv_33_cast_fp16)[name = tensor("v_33_cast_fp16")]; + tensor var_7048 = const()[name = tensor("op_7048"), val = tensor([1, 1, 1, 1])]; + tensor position16_to_fp16_dtype_0 = const()[name = tensor("position16_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor position16_to_fp16 = cast(dtype = position16_to_fp16_dtype_0, x = position16)[name = tensor("cast_473")]; + tensor ts_101_cast_fp16 = reshape(shape = var_7048, x = position16_to_fp16)[name = tensor("ts_101_cast_fp16")]; + tensor var_7052 = const()[name = tensor("op_7052"), val = tensor([1, 1, 16, 32, 2])]; + tensor q_complex_33_cast_fp16 = reshape(shape = var_7052, x = q_97_cast_fp16)[name = tensor("q_complex_33_cast_fp16")]; + tensor var_7056 = const()[name = tensor("op_7056"), val = tensor([1, 1, 16, 32, 2])]; + tensor k_complex_33_cast_fp16 = reshape(shape = var_7056, x = k_65_cast_fp16)[name = tensor("k_complex_33_cast_fp16")]; + tensor var_7060_begin_0 = const()[name = tensor("op_7060_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_7060_end_0 = const()[name = tensor("op_7060_end_0"), val = tensor([1, 1, 16, 32, 1])]; + tensor var_7060_end_mask_0 = const()[name = tensor("op_7060_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_7060_squeeze_mask_0 = const()[name = tensor("op_7060_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_7060_cast_fp16 = slice_by_index(begin = var_7060_begin_0, end = var_7060_end_0, end_mask = var_7060_end_mask_0, squeeze_mask = var_7060_squeeze_mask_0, x = q_complex_33_cast_fp16)[name = tensor("op_7060_cast_fp16")]; + tensor var_7068_begin_0 = const()[name = tensor("op_7068_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_7068_end_0 = const()[name = tensor("op_7068_end_0"), val = tensor([1, 1, 16, 32, 2])]; + tensor var_7068_end_mask_0 = const()[name = tensor("op_7068_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_7068_squeeze_mask_0 = const()[name = tensor("op_7068_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_7068_cast_fp16 = slice_by_index(begin = var_7068_begin_0, end = var_7068_end_0, end_mask = var_7068_end_mask_0, squeeze_mask = var_7068_squeeze_mask_0, x = q_complex_33_cast_fp16)[name = tensor("op_7068_cast_fp16")]; + tensor var_7076_begin_0 = const()[name = tensor("op_7076_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_7076_end_0 = const()[name = tensor("op_7076_end_0"), val = tensor([1, 1, 16, 32, 1])]; + tensor var_7076_end_mask_0 = const()[name = tensor("op_7076_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_7076_squeeze_mask_0 = const()[name = tensor("op_7076_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_7076_cast_fp16 = slice_by_index(begin = var_7076_begin_0, end = var_7076_end_0, end_mask = var_7076_end_mask_0, squeeze_mask = var_7076_squeeze_mask_0, x = k_complex_33_cast_fp16)[name = tensor("op_7076_cast_fp16")]; + tensor var_7084_begin_0 = const()[name = tensor("op_7084_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_7084_end_0 = const()[name = tensor("op_7084_end_0"), val = tensor([1, 1, 16, 32, 2])]; + tensor var_7084_end_mask_0 = const()[name = tensor("op_7084_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_7084_squeeze_mask_0 = const()[name = tensor("op_7084_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_7084_cast_fp16 = slice_by_index(begin = var_7084_begin_0, end = var_7084_end_0, end_mask = var_7084_end_mask_0, squeeze_mask = var_7084_squeeze_mask_0, x = k_complex_33_cast_fp16)[name = tensor("op_7084_cast_fp16")]; + tensor freqs_33_to_fp16 = const()[name = tensor("freqs_33_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(303726336)))]; + tensor var_7090_cast_fp16 = mul(x = freqs_33_to_fp16, y = ts_101_cast_fp16)[name = tensor("op_7090_cast_fp16")]; + tensor rotr_33_cast_fp16 = cos(x = var_7090_cast_fp16)[name = tensor("rotr_33_cast_fp16")]; + tensor roti_33_cast_fp16 = sin(x = var_7090_cast_fp16)[name = tensor("roti_33_cast_fp16")]; + tensor var_7094_cast_fp16 = mul(x = var_7060_cast_fp16, y = rotr_33_cast_fp16)[name = tensor("op_7094_cast_fp16")]; + tensor var_7095_cast_fp16 = mul(x = var_7068_cast_fp16, y = roti_33_cast_fp16)[name = tensor("op_7095_cast_fp16")]; + tensor qor_65_cast_fp16 = sub(x = var_7094_cast_fp16, y = var_7095_cast_fp16)[name = tensor("qor_65_cast_fp16")]; + tensor var_7098_cast_fp16 = mul(x = var_7060_cast_fp16, y = roti_33_cast_fp16)[name = tensor("op_7098_cast_fp16")]; + tensor var_7099_cast_fp16 = mul(x = var_7068_cast_fp16, y = rotr_33_cast_fp16)[name = tensor("op_7099_cast_fp16")]; + tensor qoi_65_cast_fp16 = add(x = var_7098_cast_fp16, y = var_7099_cast_fp16)[name = tensor("qoi_65_cast_fp16")]; + tensor var_7102_cast_fp16 = mul(x = var_7076_cast_fp16, y = rotr_33_cast_fp16)[name = tensor("op_7102_cast_fp16")]; + tensor var_7103_cast_fp16 = mul(x = var_7084_cast_fp16, y = roti_33_cast_fp16)[name = tensor("op_7103_cast_fp16")]; + tensor kor_65_cast_fp16 = sub(x = var_7102_cast_fp16, y = var_7103_cast_fp16)[name = tensor("kor_65_cast_fp16")]; + tensor var_7106_cast_fp16 = mul(x = var_7076_cast_fp16, y = roti_33_cast_fp16)[name = tensor("op_7106_cast_fp16")]; + tensor var_7107_cast_fp16 = mul(x = var_7084_cast_fp16, y = rotr_33_cast_fp16)[name = tensor("op_7107_cast_fp16")]; + tensor koi_65_cast_fp16 = add(x = var_7106_cast_fp16, y = var_7107_cast_fp16)[name = tensor("koi_65_cast_fp16")]; + tensor qo_33_axis_0 = const()[name = tensor("qo_33_axis_0"), val = tensor(-1)]; + tensor qo_33_cast_fp16 = stack(axis = qo_33_axis_0, values = (qor_65_cast_fp16, qoi_65_cast_fp16))[name = tensor("qo_33_cast_fp16")]; + tensor ko_33_axis_0 = const()[name = tensor("ko_33_axis_0"), val = tensor(-1)]; + tensor ko_33_cast_fp16 = stack(axis = ko_33_axis_0, values = (kor_65_cast_fp16, koi_65_cast_fp16))[name = tensor("ko_33_cast_fp16")]; + tensor var_7136 = const()[name = tensor("op_7136"), val = tensor([1, 1, 16, 64])]; + tensor q_99_cast_fp16 = reshape(shape = var_7136, x = qo_33_cast_fp16)[name = tensor("q_99_cast_fp16")]; + tensor var_7138 = const()[name = tensor("op_7138"), val = tensor([1, 1, 16, 64])]; + tensor k_67_cast_fp16 = reshape(shape = var_7138, x = ko_33_cast_fp16)[name = tensor("k_67_cast_fp16")]; + tensor _inversed_7160_y_0_to_fp16 = const()[name = tensor("_inversed_7160_y_0_to_fp16"), val = tensor(0x1p-9)]; + tensor _inversed_7160_cast_fp16 = mul(x = ts_101_cast_fp16, y = _inversed_7160_y_0_to_fp16)[name = tensor("_inversed_7160_cast_fp16")]; + tensor var_7161_cast_fp16 = floor(x = _inversed_7160_cast_fp16)[name = tensor("op_7161_cast_fp16")]; + tensor var_7162_to_fp16 = const()[name = tensor("op_7162_to_fp16"), val = tensor(0x1p+9)]; + tensor var_7163_cast_fp16 = mul(x = var_7161_cast_fp16, y = var_7162_to_fp16)[name = tensor("op_7163_cast_fp16")]; + tensor write_indices_float_67_cast_fp16 = sub(x = ts_101_cast_fp16, y = var_7163_cast_fp16)[name = tensor("write_indices_float_67_cast_fp16")]; + tensor var_7170_dtype_0 = const()[name = tensor("op_7170_dtype_0"), val = tensor("int32")]; + tensor write_indices_33_reps_0 = const()[name = tensor("write_indices_33_reps_0"), val = tensor([1, 1, 16, 64])]; + tensor write_indices_float_67_cast_fp16_to_int32 = cast(dtype = var_7170_dtype_0, x = write_indices_float_67_cast_fp16)[name = tensor("cast_472")]; + tensor write_indices_33 = tile(reps = write_indices_33_reps_0, x = write_indices_float_67_cast_fp16_to_int32)[name = tensor("write_indices_33")]; + tensor var_7178_begin_0 = const()[name = tensor("op_7178_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_7178_end_0 = const()[name = tensor("op_7178_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor var_7178_end_mask_0 = const()[name = tensor("op_7178_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_7178_squeeze_mask_0 = const()[name = tensor("op_7178_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor cache16_to_fp16_dtype_0 = const()[name = tensor("cache16_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor cache16_to_fp16 = cast(dtype = cache16_to_fp16_dtype_0, x = cache16)[name = tensor("cast_471")]; + tensor var_7178_cast_fp16 = slice_by_index(begin = var_7178_begin_0, end = var_7178_end_0, end_mask = var_7178_end_mask_0, squeeze_mask = var_7178_squeeze_mask_0, x = cache16_to_fp16)[name = tensor("op_7178_cast_fp16")]; + tensor var_7180_axis_0 = const()[name = tensor("op_7180_axis_0"), val = tensor(1)]; + tensor var_7180_mode_0 = const()[name = tensor("op_7180_mode_0"), val = tensor("update")]; + tensor var_7180_validate_indices_0 = const()[name = tensor("op_7180_validate_indices_0"), val = tensor(false)]; + tensor var_7180_cast_fp16 = scatter_along_axis(axis = var_7180_axis_0, data = var_7178_cast_fp16, indices = write_indices_33, mode = var_7180_mode_0, updates = k_67_cast_fp16, validate_indices = var_7180_validate_indices_0)[name = tensor("op_7180_cast_fp16")]; + tensor concat_114 = const()[name = tensor("concat_114"), val = tensor([0, 0, 0, 0, 0])]; + tensor concat_115 = const()[name = tensor("concat_115"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_33_internal_tensor_assign_1_stride_0 = const()[name = tensor("new_cache_33_internal_tensor_assign_1_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_33_internal_tensor_assign_1_begin_mask_0 = const()[name = tensor("new_cache_33_internal_tensor_assign_1_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_33_internal_tensor_assign_1_end_mask_0 = const()[name = tensor("new_cache_33_internal_tensor_assign_1_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_33_internal_tensor_assign_1_squeeze_mask_0 = const()[name = tensor("new_cache_33_internal_tensor_assign_1_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_80 = const()[name = tensor("shape_80"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_32 = const()[name = tensor("reduce_prod_32"), val = tensor(1048576)]; + tensor range_1d_32_start_0 = const()[name = tensor("range_1d_32_start_0"), val = tensor(0)]; + tensor range_1d_32_step_0 = const()[name = tensor("range_1d_32_step_0"), val = tensor(1)]; + tensor range_1d_32 = range_1d(end = reduce_prod_32, start = range_1d_32_start_0, step = range_1d_32_step_0)[name = tensor("range_1d_32")]; + tensor reshape_160 = reshape(shape = shape_80, x = range_1d_32)[name = tensor("reshape_160")]; + tensor slice_by_index_32 = slice_by_index(begin = concat_114, begin_mask = new_cache_33_internal_tensor_assign_1_begin_mask_0, end = concat_115, end_mask = new_cache_33_internal_tensor_assign_1_end_mask_0, squeeze_mask = new_cache_33_internal_tensor_assign_1_squeeze_mask_0, stride = new_cache_33_internal_tensor_assign_1_stride_0, x = reshape_160)[name = tensor("slice_by_index_32")]; + tensor reshape_161_shape_0 = const()[name = tensor("reshape_161_shape_0"), val = tensor([-1])]; + tensor reshape_161 = reshape(shape = reshape_161_shape_0, x = slice_by_index_32)[name = tensor("reshape_161")]; + tensor reshape_162_shape_0 = const()[name = tensor("reshape_162_shape_0"), val = tensor([-1])]; + tensor reshape_162_cast_fp16 = reshape(shape = reshape_162_shape_0, x = var_7180_cast_fp16)[name = tensor("reshape_162_cast_fp16")]; + tensor reshape_163_shape_0 = const()[name = tensor("reshape_163_shape_0"), val = tensor([-1])]; + tensor reshape_163_cast_fp16 = reshape(shape = reshape_163_shape_0, x = cache16_to_fp16)[name = tensor("reshape_163_cast_fp16")]; + tensor scatter_32_mode_0 = const()[name = tensor("scatter_32_mode_0"), val = tensor("update")]; + tensor scatter_32_axis_0 = const()[name = tensor("scatter_32_axis_0"), val = tensor(0)]; + tensor scatter_32_validate_indices_0 = const()[name = tensor("scatter_32_validate_indices_0"), val = tensor(false)]; + tensor scatter_32_cast_fp16 = scatter(axis = scatter_32_axis_0, data = reshape_163_cast_fp16, indices = reshape_161, mode = scatter_32_mode_0, updates = reshape_162_cast_fp16, validate_indices = scatter_32_validate_indices_0)[name = tensor("scatter_32_cast_fp16")]; + tensor reshape_164_cast_fp16 = reshape(shape = shape_80, x = scatter_32_cast_fp16)[name = tensor("reshape_164_cast_fp16")]; + tensor var_7188_begin_0 = const()[name = tensor("op_7188_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor var_7188_end_0 = const()[name = tensor("op_7188_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor var_7188_end_mask_0 = const()[name = tensor("op_7188_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_7188_squeeze_mask_0 = const()[name = tensor("op_7188_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor var_7188_cast_fp16 = slice_by_index(begin = var_7188_begin_0, end = var_7188_end_0, end_mask = var_7188_end_mask_0, squeeze_mask = var_7188_squeeze_mask_0, x = reshape_164_cast_fp16)[name = tensor("op_7188_cast_fp16")]; + tensor var_7190_axis_0 = const()[name = tensor("op_7190_axis_0"), val = tensor(1)]; + tensor var_7190_mode_0 = const()[name = tensor("op_7190_mode_0"), val = tensor("update")]; + tensor var_7190_validate_indices_0 = const()[name = tensor("op_7190_validate_indices_0"), val = tensor(false)]; + tensor var_7190_cast_fp16 = scatter_along_axis(axis = var_7190_axis_0, data = var_7188_cast_fp16, indices = write_indices_33, mode = var_7190_mode_0, updates = v_33_cast_fp16, validate_indices = var_7190_validate_indices_0)[name = tensor("op_7190_cast_fp16")]; + tensor concat_116 = const()[name = tensor("concat_116"), val = tensor([1, 0, 0, 0, 0])]; + tensor concat_117 = const()[name = tensor("concat_117"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_33_internal_tensor_assign_2_stride_0 = const()[name = tensor("new_cache_33_internal_tensor_assign_2_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_33_internal_tensor_assign_2_begin_mask_0 = const()[name = tensor("new_cache_33_internal_tensor_assign_2_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_33_internal_tensor_assign_2_end_mask_0 = const()[name = tensor("new_cache_33_internal_tensor_assign_2_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_33_internal_tensor_assign_2_squeeze_mask_0 = const()[name = tensor("new_cache_33_internal_tensor_assign_2_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_81 = const()[name = tensor("shape_81"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_33 = const()[name = tensor("reduce_prod_33"), val = tensor(1048576)]; + tensor range_1d_33_start_0 = const()[name = tensor("range_1d_33_start_0"), val = tensor(0)]; + tensor range_1d_33_step_0 = const()[name = tensor("range_1d_33_step_0"), val = tensor(1)]; + tensor range_1d_33 = range_1d(end = reduce_prod_33, start = range_1d_33_start_0, step = range_1d_33_step_0)[name = tensor("range_1d_33")]; + tensor reshape_165 = reshape(shape = shape_81, x = range_1d_33)[name = tensor("reshape_165")]; + tensor slice_by_index_33 = slice_by_index(begin = concat_116, begin_mask = new_cache_33_internal_tensor_assign_2_begin_mask_0, end = concat_117, end_mask = new_cache_33_internal_tensor_assign_2_end_mask_0, squeeze_mask = new_cache_33_internal_tensor_assign_2_squeeze_mask_0, stride = new_cache_33_internal_tensor_assign_2_stride_0, x = reshape_165)[name = tensor("slice_by_index_33")]; + tensor reshape_166_shape_0 = const()[name = tensor("reshape_166_shape_0"), val = tensor([-1])]; + tensor reshape_166 = reshape(shape = reshape_166_shape_0, x = slice_by_index_33)[name = tensor("reshape_166")]; + tensor reshape_167_shape_0 = const()[name = tensor("reshape_167_shape_0"), val = tensor([-1])]; + tensor reshape_167_cast_fp16 = reshape(shape = reshape_167_shape_0, x = var_7190_cast_fp16)[name = tensor("reshape_167_cast_fp16")]; + tensor reshape_168_shape_0 = const()[name = tensor("reshape_168_shape_0"), val = tensor([-1])]; + tensor reshape_168_cast_fp16 = reshape(shape = reshape_168_shape_0, x = reshape_164_cast_fp16)[name = tensor("reshape_168_cast_fp16")]; + tensor scatter_33_mode_0 = const()[name = tensor("scatter_33_mode_0"), val = tensor("update")]; + tensor scatter_33_axis_0 = const()[name = tensor("scatter_33_axis_0"), val = tensor(0)]; + tensor scatter_33_validate_indices_0 = const()[name = tensor("scatter_33_validate_indices_0"), val = tensor(false)]; + tensor scatter_33_cast_fp16 = scatter(axis = scatter_33_axis_0, data = reshape_168_cast_fp16, indices = reshape_166, mode = scatter_33_mode_0, updates = reshape_167_cast_fp16, validate_indices = scatter_33_validate_indices_0)[name = tensor("scatter_33_cast_fp16")]; + tensor reshape_169_cast_fp16 = reshape(shape = shape_81, x = scatter_33_cast_fp16)[name = tensor("reshape_169_cast_fp16")]; + tensor reshape_169_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("reshape_169_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor keys_97_begin_0 = const()[name = tensor("keys_97_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor keys_97_end_0 = const()[name = tensor("keys_97_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor keys_97_end_mask_0 = const()[name = tensor("keys_97_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor keys_97_squeeze_mask_0 = const()[name = tensor("keys_97_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor keys_97_cast_fp16 = slice_by_index(begin = keys_97_begin_0, end = keys_97_end_0, end_mask = keys_97_end_mask_0, squeeze_mask = keys_97_squeeze_mask_0, x = reshape_169_cast_fp16)[name = tensor("keys_97_cast_fp16")]; + tensor values_97_begin_0 = const()[name = tensor("values_97_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor values_97_end_0 = const()[name = tensor("values_97_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor values_97_end_mask_0 = const()[name = tensor("values_97_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor values_97_squeeze_mask_0 = const()[name = tensor("values_97_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor values_97_cast_fp16 = slice_by_index(begin = values_97_begin_0, end = values_97_end_0, end_mask = values_97_end_mask_0, squeeze_mask = values_97_squeeze_mask_0, x = reshape_169_cast_fp16)[name = tensor("values_97_cast_fp16")]; + tensor var_7202_cast_fp16 = not_equal(x = keys_97_cast_fp16, y = keys_97_cast_fp16)[name = tensor("op_7202_cast_fp16")]; + tensor keys_99_cast_fp16 = select(a = var_504_to_fp16, b = keys_97_cast_fp16, cond = var_7202_cast_fp16)[name = tensor("keys_99_cast_fp16")]; + tensor var_7210_cast_fp16 = not_equal(x = values_97_cast_fp16, y = values_97_cast_fp16)[name = tensor("op_7210_cast_fp16")]; + tensor values_99_cast_fp16 = select(a = var_504_to_fp16, b = values_97_cast_fp16, cond = var_7210_cast_fp16)[name = tensor("values_99_cast_fp16")]; + tensor var_7234 = const()[name = tensor("op_7234"), val = tensor([0, 2, 1, 3])]; + tensor var_7247 = const()[name = tensor("op_7247"), val = tensor([1, 1, 1])]; + tensor var_7248_cast_fp16 = reshape(shape = var_7247, x = position16_to_fp16)[name = tensor("op_7248_cast_fp16")]; + tensor var_7265_to_fp16 = const()[name = tensor("op_7265_to_fp16"), val = tensor(0x1p+0)]; + tensor valid_len_33_cast_fp16 = add(x = var_7248_cast_fp16, y = var_7265_to_fp16)[name = tensor("valid_len_33_cast_fp16")]; + tensor valid_mask_33_cast_fp16 = less(x = k_positions_1_promoted_to_fp16, y = valid_len_33_cast_fp16)[name = tensor("valid_mask_33_cast_fp16")]; + tensor causal_mask_33_cast_fp16 = less_equal(x = k_positions_1_promoted_to_fp16, y = var_7248_cast_fp16)[name = tensor("causal_mask_33_cast_fp16")]; + tensor attn_mask_65 = logical_and(x = valid_mask_33_cast_fp16, y = causal_mask_33_cast_fp16)[name = tensor("attn_mask_65")]; + tensor attn_mask_67_axes_0 = const()[name = tensor("attn_mask_67_axes_0"), val = tensor([1])]; + tensor attn_mask_67 = expand_dims(axes = attn_mask_67_axes_0, x = attn_mask_65)[name = tensor("attn_mask_67")]; + tensor var_7283_transpose_x_0 = const()[name = tensor("op_7283_transpose_x_0"), val = tensor(false)]; + tensor var_7283_transpose_y_0 = const()[name = tensor("op_7283_transpose_y_0"), val = tensor(false)]; + tensor transpose_104_perm_0 = const()[name = tensor("transpose_104_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_105_perm_0 = const()[name = tensor("transpose_105_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_105 = transpose(perm = transpose_105_perm_0, x = keys_99_cast_fp16)[name = tensor("transpose_149")]; + tensor transpose_104 = transpose(perm = transpose_104_perm_0, x = q_99_cast_fp16)[name = tensor("transpose_150")]; + tensor var_7283_cast_fp16 = matmul(transpose_x = var_7283_transpose_x_0, transpose_y = var_7283_transpose_y_0, x = transpose_104, y = transpose_105)[name = tensor("op_7283_cast_fp16")]; + tensor var_7277_to_fp16 = const()[name = tensor("op_7277_to_fp16"), val = tensor([0x1p-3])]; + tensor attn_weights_97_cast_fp16 = mul(x = var_7283_cast_fp16, y = var_7277_to_fp16)[name = tensor("attn_weights_97_cast_fp16")]; + tensor var_7285 = logical_not(x = attn_mask_67)[name = tensor("op_7285")]; + tensor var_7286_to_fp16 = const()[name = tensor("op_7286_to_fp16"), val = tensor(-inf)]; + tensor attn_weights_99_cast_fp16 = select(a = var_7286_to_fp16, b = attn_weights_97_cast_fp16, cond = var_7285)[name = tensor("attn_weights_99_cast_fp16")]; + tensor var_7288 = const()[name = tensor("op_7288"), val = tensor(-1)]; + tensor attn_weights_101_cast_fp16 = softmax(axis = var_7288, x = attn_weights_99_cast_fp16)[name = tensor("attn_weights_101_cast_fp16")]; + tensor attn_output_33_transpose_x_0 = const()[name = tensor("attn_output_33_transpose_x_0"), val = tensor(false)]; + tensor attn_output_33_transpose_y_0 = const()[name = tensor("attn_output_33_transpose_y_0"), val = tensor(false)]; + tensor values_101_cast_fp16 = transpose(perm = var_7234, x = values_99_cast_fp16)[name = tensor("transpose_151")]; + tensor attn_output_33_cast_fp16 = matmul(transpose_x = attn_output_33_transpose_x_0, transpose_y = attn_output_33_transpose_y_0, x = attn_weights_101_cast_fp16, y = values_101_cast_fp16)[name = tensor("attn_output_33_cast_fp16")]; + tensor var_7296 = const()[name = tensor("op_7296"), val = tensor([0, 2, 1, 3])]; + tensor var_7299 = const()[name = tensor("op_7299"), val = tensor([1, 1, 1024])]; + tensor var_7297_cast_fp16 = transpose(perm = var_7296, x = attn_output_33_cast_fp16)[name = tensor("transpose_148")]; + tensor input_165_cast_fp16 = reshape(shape = var_7299, x = var_7297_cast_fp16)[name = tensor("input_165_cast_fp16")]; + tensor linear_66_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = attn16_out_proj_weight_cast_fp16, x = input_165_cast_fp16)[name = tensor("linear_66_cast_fp16")]; + tensor var_7305_to_fp16 = const()[name = tensor("op_7305_to_fp16"), val = tensor(0x1p+0)]; + tensor var_7306_cast_fp16 = add(x = position16_to_fp16, y = var_7305_to_fp16)[name = tensor("op_7306_cast_fp16")]; + tensor var_7306_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_7306_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor input_167_cast_fp16 = add(x = input_163_cast_fp16, y = linear_66_cast_fp16)[name = tensor("input_167_cast_fp16")]; + tensor input_169_axes_0 = const()[name = tensor("input_169_axes_0"), val = tensor([-1])]; + tensor norm16_2_weight_to_fp16 = const()[name = tensor("norm16_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(303726464)))]; + tensor norm16_2_bias_to_fp16 = const()[name = tensor("norm16_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(303728576)))]; + tensor var_7310_to_fp16 = const()[name = tensor("op_7310_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_169_cast_fp16 = layer_norm(axes = input_169_axes_0, beta = norm16_2_bias_to_fp16, epsilon = var_7310_to_fp16, gamma = norm16_2_weight_to_fp16, x = input_167_cast_fp16)[name = tensor("input_169_cast_fp16")]; + tensor linear_67_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = linear16_1_weight_cast_fp16, x = input_169_cast_fp16)[name = tensor("linear_67_cast_fp16")]; + tensor input_171_mode_0 = const()[name = tensor("input_171_mode_0"), val = tensor("EXACT")]; + tensor input_171_cast_fp16 = gelu(mode = input_171_mode_0, x = linear_67_cast_fp16)[name = tensor("input_171_cast_fp16")]; + tensor linear_68_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = linear16_2_weight_cast_fp16, x = input_171_cast_fp16)[name = tensor("linear_68_cast_fp16")]; + tensor input_173_cast_fp16 = add(x = input_167_cast_fp16, y = linear_68_cast_fp16)[name = tensor("input_173_cast_fp16")]; + tensor x_35_axes_0 = const()[name = tensor("x_35_axes_0"), val = tensor([-1])]; + tensor norm17_1_weight_to_fp16 = const()[name = tensor("norm17_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(303730688)))]; + tensor norm17_1_bias_to_fp16 = const()[name = tensor("norm17_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(303732800)))]; + tensor var_7327_to_fp16 = const()[name = tensor("op_7327_to_fp16"), val = tensor(0x1.5p-17)]; + tensor x_35_cast_fp16 = layer_norm(axes = x_35_axes_0, beta = norm17_1_bias_to_fp16, epsilon = var_7327_to_fp16, gamma = norm17_1_weight_to_fp16, x = input_173_cast_fp16)[name = tensor("x_35_cast_fp16")]; + tensor linear_69_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = attn17_in_proj_weight_cast_fp16, x = x_35_cast_fp16)[name = tensor("linear_69_cast_fp16")]; + tensor var_7363 = const()[name = tensor("op_7363"), val = tensor([1, 1, 3, 16, 64])]; + tensor qkv_35_cast_fp16 = reshape(shape = var_7363, x = linear_69_cast_fp16)[name = tensor("qkv_35_cast_fp16")]; + tensor q_103_begin_0 = const()[name = tensor("q_103_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor q_103_end_0 = const()[name = tensor("q_103_end_0"), val = tensor([1, 1, 1, 16, 64])]; + tensor q_103_end_mask_0 = const()[name = tensor("q_103_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor q_103_squeeze_mask_0 = const()[name = tensor("q_103_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor q_103_cast_fp16 = slice_by_index(begin = q_103_begin_0, end = q_103_end_0, end_mask = q_103_end_mask_0, squeeze_mask = q_103_squeeze_mask_0, x = qkv_35_cast_fp16)[name = tensor("q_103_cast_fp16")]; + tensor k_69_begin_0 = const()[name = tensor("k_69_begin_0"), val = tensor([0, 0, 1, 0, 0])]; + tensor k_69_end_0 = const()[name = tensor("k_69_end_0"), val = tensor([1, 1, 2, 16, 64])]; + tensor k_69_end_mask_0 = const()[name = tensor("k_69_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor k_69_squeeze_mask_0 = const()[name = tensor("k_69_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor k_69_cast_fp16 = slice_by_index(begin = k_69_begin_0, end = k_69_end_0, end_mask = k_69_end_mask_0, squeeze_mask = k_69_squeeze_mask_0, x = qkv_35_cast_fp16)[name = tensor("k_69_cast_fp16")]; + tensor v_35_begin_0 = const()[name = tensor("v_35_begin_0"), val = tensor([0, 0, 2, 0, 0])]; + tensor v_35_end_0 = const()[name = tensor("v_35_end_0"), val = tensor([1, 1, 3, 16, 64])]; + tensor v_35_end_mask_0 = const()[name = tensor("v_35_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor v_35_squeeze_mask_0 = const()[name = tensor("v_35_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor v_35_cast_fp16 = slice_by_index(begin = v_35_begin_0, end = v_35_end_0, end_mask = v_35_end_mask_0, squeeze_mask = v_35_squeeze_mask_0, x = qkv_35_cast_fp16)[name = tensor("v_35_cast_fp16")]; + tensor var_7467 = const()[name = tensor("op_7467"), val = tensor([1, 1, 1, 1])]; + tensor position17_to_fp16_dtype_0 = const()[name = tensor("position17_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor position17_to_fp16 = cast(dtype = position17_to_fp16_dtype_0, x = position17)[name = tensor("cast_468")]; + tensor ts_107_cast_fp16 = reshape(shape = var_7467, x = position17_to_fp16)[name = tensor("ts_107_cast_fp16")]; + tensor var_7471 = const()[name = tensor("op_7471"), val = tensor([1, 1, 16, 32, 2])]; + tensor q_complex_35_cast_fp16 = reshape(shape = var_7471, x = q_103_cast_fp16)[name = tensor("q_complex_35_cast_fp16")]; + tensor var_7475 = const()[name = tensor("op_7475"), val = tensor([1, 1, 16, 32, 2])]; + tensor k_complex_35_cast_fp16 = reshape(shape = var_7475, x = k_69_cast_fp16)[name = tensor("k_complex_35_cast_fp16")]; + tensor var_7479_begin_0 = const()[name = tensor("op_7479_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_7479_end_0 = const()[name = tensor("op_7479_end_0"), val = tensor([1, 1, 16, 32, 1])]; + tensor var_7479_end_mask_0 = const()[name = tensor("op_7479_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_7479_squeeze_mask_0 = const()[name = tensor("op_7479_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_7479_cast_fp16 = slice_by_index(begin = var_7479_begin_0, end = var_7479_end_0, end_mask = var_7479_end_mask_0, squeeze_mask = var_7479_squeeze_mask_0, x = q_complex_35_cast_fp16)[name = tensor("op_7479_cast_fp16")]; + tensor var_7487_begin_0 = const()[name = tensor("op_7487_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_7487_end_0 = const()[name = tensor("op_7487_end_0"), val = tensor([1, 1, 16, 32, 2])]; + tensor var_7487_end_mask_0 = const()[name = tensor("op_7487_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_7487_squeeze_mask_0 = const()[name = tensor("op_7487_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_7487_cast_fp16 = slice_by_index(begin = var_7487_begin_0, end = var_7487_end_0, end_mask = var_7487_end_mask_0, squeeze_mask = var_7487_squeeze_mask_0, x = q_complex_35_cast_fp16)[name = tensor("op_7487_cast_fp16")]; + tensor var_7495_begin_0 = const()[name = tensor("op_7495_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_7495_end_0 = const()[name = tensor("op_7495_end_0"), val = tensor([1, 1, 16, 32, 1])]; + tensor var_7495_end_mask_0 = const()[name = tensor("op_7495_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_7495_squeeze_mask_0 = const()[name = tensor("op_7495_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_7495_cast_fp16 = slice_by_index(begin = var_7495_begin_0, end = var_7495_end_0, end_mask = var_7495_end_mask_0, squeeze_mask = var_7495_squeeze_mask_0, x = k_complex_35_cast_fp16)[name = tensor("op_7495_cast_fp16")]; + tensor var_7503_begin_0 = const()[name = tensor("op_7503_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_7503_end_0 = const()[name = tensor("op_7503_end_0"), val = tensor([1, 1, 16, 32, 2])]; + tensor var_7503_end_mask_0 = const()[name = tensor("op_7503_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_7503_squeeze_mask_0 = const()[name = tensor("op_7503_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_7503_cast_fp16 = slice_by_index(begin = var_7503_begin_0, end = var_7503_end_0, end_mask = var_7503_end_mask_0, squeeze_mask = var_7503_squeeze_mask_0, x = k_complex_35_cast_fp16)[name = tensor("op_7503_cast_fp16")]; + tensor freqs_35_to_fp16 = const()[name = tensor("freqs_35_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(303734912)))]; + tensor var_7509_cast_fp16 = mul(x = freqs_35_to_fp16, y = ts_107_cast_fp16)[name = tensor("op_7509_cast_fp16")]; + tensor rotr_35_cast_fp16 = cos(x = var_7509_cast_fp16)[name = tensor("rotr_35_cast_fp16")]; + tensor roti_35_cast_fp16 = sin(x = var_7509_cast_fp16)[name = tensor("roti_35_cast_fp16")]; + tensor var_7513_cast_fp16 = mul(x = var_7479_cast_fp16, y = rotr_35_cast_fp16)[name = tensor("op_7513_cast_fp16")]; + tensor var_7514_cast_fp16 = mul(x = var_7487_cast_fp16, y = roti_35_cast_fp16)[name = tensor("op_7514_cast_fp16")]; + tensor qor_69_cast_fp16 = sub(x = var_7513_cast_fp16, y = var_7514_cast_fp16)[name = tensor("qor_69_cast_fp16")]; + tensor var_7517_cast_fp16 = mul(x = var_7479_cast_fp16, y = roti_35_cast_fp16)[name = tensor("op_7517_cast_fp16")]; + tensor var_7518_cast_fp16 = mul(x = var_7487_cast_fp16, y = rotr_35_cast_fp16)[name = tensor("op_7518_cast_fp16")]; + tensor qoi_69_cast_fp16 = add(x = var_7517_cast_fp16, y = var_7518_cast_fp16)[name = tensor("qoi_69_cast_fp16")]; + tensor var_7521_cast_fp16 = mul(x = var_7495_cast_fp16, y = rotr_35_cast_fp16)[name = tensor("op_7521_cast_fp16")]; + tensor var_7522_cast_fp16 = mul(x = var_7503_cast_fp16, y = roti_35_cast_fp16)[name = tensor("op_7522_cast_fp16")]; + tensor kor_69_cast_fp16 = sub(x = var_7521_cast_fp16, y = var_7522_cast_fp16)[name = tensor("kor_69_cast_fp16")]; + tensor var_7525_cast_fp16 = mul(x = var_7495_cast_fp16, y = roti_35_cast_fp16)[name = tensor("op_7525_cast_fp16")]; + tensor var_7526_cast_fp16 = mul(x = var_7503_cast_fp16, y = rotr_35_cast_fp16)[name = tensor("op_7526_cast_fp16")]; + tensor koi_69_cast_fp16 = add(x = var_7525_cast_fp16, y = var_7526_cast_fp16)[name = tensor("koi_69_cast_fp16")]; + tensor qo_35_axis_0 = const()[name = tensor("qo_35_axis_0"), val = tensor(-1)]; + tensor qo_35_cast_fp16 = stack(axis = qo_35_axis_0, values = (qor_69_cast_fp16, qoi_69_cast_fp16))[name = tensor("qo_35_cast_fp16")]; + tensor ko_35_axis_0 = const()[name = tensor("ko_35_axis_0"), val = tensor(-1)]; + tensor ko_35_cast_fp16 = stack(axis = ko_35_axis_0, values = (kor_69_cast_fp16, koi_69_cast_fp16))[name = tensor("ko_35_cast_fp16")]; + tensor var_7555 = const()[name = tensor("op_7555"), val = tensor([1, 1, 16, 64])]; + tensor q_105_cast_fp16 = reshape(shape = var_7555, x = qo_35_cast_fp16)[name = tensor("q_105_cast_fp16")]; + tensor var_7557 = const()[name = tensor("op_7557"), val = tensor([1, 1, 16, 64])]; + tensor k_71_cast_fp16 = reshape(shape = var_7557, x = ko_35_cast_fp16)[name = tensor("k_71_cast_fp16")]; + tensor _inversed_7579_y_0_to_fp16 = const()[name = tensor("_inversed_7579_y_0_to_fp16"), val = tensor(0x1p-9)]; + tensor _inversed_7579_cast_fp16 = mul(x = ts_107_cast_fp16, y = _inversed_7579_y_0_to_fp16)[name = tensor("_inversed_7579_cast_fp16")]; + tensor var_7580_cast_fp16 = floor(x = _inversed_7579_cast_fp16)[name = tensor("op_7580_cast_fp16")]; + tensor var_7581_to_fp16 = const()[name = tensor("op_7581_to_fp16"), val = tensor(0x1p+9)]; + tensor var_7582_cast_fp16 = mul(x = var_7580_cast_fp16, y = var_7581_to_fp16)[name = tensor("op_7582_cast_fp16")]; + tensor write_indices_float_71_cast_fp16 = sub(x = ts_107_cast_fp16, y = var_7582_cast_fp16)[name = tensor("write_indices_float_71_cast_fp16")]; + tensor var_7589_dtype_0 = const()[name = tensor("op_7589_dtype_0"), val = tensor("int32")]; + tensor write_indices_35_reps_0 = const()[name = tensor("write_indices_35_reps_0"), val = tensor([1, 1, 16, 64])]; + tensor write_indices_float_71_cast_fp16_to_int32 = cast(dtype = var_7589_dtype_0, x = write_indices_float_71_cast_fp16)[name = tensor("cast_467")]; + tensor write_indices_35 = tile(reps = write_indices_35_reps_0, x = write_indices_float_71_cast_fp16_to_int32)[name = tensor("write_indices_35")]; + tensor var_7597_begin_0 = const()[name = tensor("op_7597_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_7597_end_0 = const()[name = tensor("op_7597_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor var_7597_end_mask_0 = const()[name = tensor("op_7597_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_7597_squeeze_mask_0 = const()[name = tensor("op_7597_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor cache17_to_fp16_dtype_0 = const()[name = tensor("cache17_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor cache17_to_fp16 = cast(dtype = cache17_to_fp16_dtype_0, x = cache17)[name = tensor("cast_466")]; + tensor var_7597_cast_fp16 = slice_by_index(begin = var_7597_begin_0, end = var_7597_end_0, end_mask = var_7597_end_mask_0, squeeze_mask = var_7597_squeeze_mask_0, x = cache17_to_fp16)[name = tensor("op_7597_cast_fp16")]; + tensor var_7599_axis_0 = const()[name = tensor("op_7599_axis_0"), val = tensor(1)]; + tensor var_7599_mode_0 = const()[name = tensor("op_7599_mode_0"), val = tensor("update")]; + tensor var_7599_validate_indices_0 = const()[name = tensor("op_7599_validate_indices_0"), val = tensor(false)]; + tensor var_7599_cast_fp16 = scatter_along_axis(axis = var_7599_axis_0, data = var_7597_cast_fp16, indices = write_indices_35, mode = var_7599_mode_0, updates = k_71_cast_fp16, validate_indices = var_7599_validate_indices_0)[name = tensor("op_7599_cast_fp16")]; + tensor concat_121 = const()[name = tensor("concat_121"), val = tensor([0, 0, 0, 0, 0])]; + tensor concat_122 = const()[name = tensor("concat_122"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_35_internal_tensor_assign_1_stride_0 = const()[name = tensor("new_cache_35_internal_tensor_assign_1_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_35_internal_tensor_assign_1_begin_mask_0 = const()[name = tensor("new_cache_35_internal_tensor_assign_1_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_35_internal_tensor_assign_1_end_mask_0 = const()[name = tensor("new_cache_35_internal_tensor_assign_1_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_35_internal_tensor_assign_1_squeeze_mask_0 = const()[name = tensor("new_cache_35_internal_tensor_assign_1_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_82 = const()[name = tensor("shape_82"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_34 = const()[name = tensor("reduce_prod_34"), val = tensor(1048576)]; + tensor range_1d_34_start_0 = const()[name = tensor("range_1d_34_start_0"), val = tensor(0)]; + tensor range_1d_34_step_0 = const()[name = tensor("range_1d_34_step_0"), val = tensor(1)]; + tensor range_1d_34 = range_1d(end = reduce_prod_34, start = range_1d_34_start_0, step = range_1d_34_step_0)[name = tensor("range_1d_34")]; + tensor reshape_170 = reshape(shape = shape_82, x = range_1d_34)[name = tensor("reshape_170")]; + tensor slice_by_index_34 = slice_by_index(begin = concat_121, begin_mask = new_cache_35_internal_tensor_assign_1_begin_mask_0, end = concat_122, end_mask = new_cache_35_internal_tensor_assign_1_end_mask_0, squeeze_mask = new_cache_35_internal_tensor_assign_1_squeeze_mask_0, stride = new_cache_35_internal_tensor_assign_1_stride_0, x = reshape_170)[name = tensor("slice_by_index_34")]; + tensor reshape_171_shape_0 = const()[name = tensor("reshape_171_shape_0"), val = tensor([-1])]; + tensor reshape_171 = reshape(shape = reshape_171_shape_0, x = slice_by_index_34)[name = tensor("reshape_171")]; + tensor reshape_172_shape_0 = const()[name = tensor("reshape_172_shape_0"), val = tensor([-1])]; + tensor reshape_172_cast_fp16 = reshape(shape = reshape_172_shape_0, x = var_7599_cast_fp16)[name = tensor("reshape_172_cast_fp16")]; + tensor reshape_173_shape_0 = const()[name = tensor("reshape_173_shape_0"), val = tensor([-1])]; + tensor reshape_173_cast_fp16 = reshape(shape = reshape_173_shape_0, x = cache17_to_fp16)[name = tensor("reshape_173_cast_fp16")]; + tensor scatter_34_mode_0 = const()[name = tensor("scatter_34_mode_0"), val = tensor("update")]; + tensor scatter_34_axis_0 = const()[name = tensor("scatter_34_axis_0"), val = tensor(0)]; + tensor scatter_34_validate_indices_0 = const()[name = tensor("scatter_34_validate_indices_0"), val = tensor(false)]; + tensor scatter_34_cast_fp16 = scatter(axis = scatter_34_axis_0, data = reshape_173_cast_fp16, indices = reshape_171, mode = scatter_34_mode_0, updates = reshape_172_cast_fp16, validate_indices = scatter_34_validate_indices_0)[name = tensor("scatter_34_cast_fp16")]; + tensor reshape_174_cast_fp16 = reshape(shape = shape_82, x = scatter_34_cast_fp16)[name = tensor("reshape_174_cast_fp16")]; + tensor var_7607_begin_0 = const()[name = tensor("op_7607_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor var_7607_end_0 = const()[name = tensor("op_7607_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor var_7607_end_mask_0 = const()[name = tensor("op_7607_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_7607_squeeze_mask_0 = const()[name = tensor("op_7607_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor var_7607_cast_fp16 = slice_by_index(begin = var_7607_begin_0, end = var_7607_end_0, end_mask = var_7607_end_mask_0, squeeze_mask = var_7607_squeeze_mask_0, x = reshape_174_cast_fp16)[name = tensor("op_7607_cast_fp16")]; + tensor var_7609_axis_0 = const()[name = tensor("op_7609_axis_0"), val = tensor(1)]; + tensor var_7609_mode_0 = const()[name = tensor("op_7609_mode_0"), val = tensor("update")]; + tensor var_7609_validate_indices_0 = const()[name = tensor("op_7609_validate_indices_0"), val = tensor(false)]; + tensor var_7609_cast_fp16 = scatter_along_axis(axis = var_7609_axis_0, data = var_7607_cast_fp16, indices = write_indices_35, mode = var_7609_mode_0, updates = v_35_cast_fp16, validate_indices = var_7609_validate_indices_0)[name = tensor("op_7609_cast_fp16")]; + tensor concat_123 = const()[name = tensor("concat_123"), val = tensor([1, 0, 0, 0, 0])]; + tensor concat_124 = const()[name = tensor("concat_124"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_35_internal_tensor_assign_2_stride_0 = const()[name = tensor("new_cache_35_internal_tensor_assign_2_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_35_internal_tensor_assign_2_begin_mask_0 = const()[name = tensor("new_cache_35_internal_tensor_assign_2_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_35_internal_tensor_assign_2_end_mask_0 = const()[name = tensor("new_cache_35_internal_tensor_assign_2_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_35_internal_tensor_assign_2_squeeze_mask_0 = const()[name = tensor("new_cache_35_internal_tensor_assign_2_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_83 = const()[name = tensor("shape_83"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_35 = const()[name = tensor("reduce_prod_35"), val = tensor(1048576)]; + tensor range_1d_35_start_0 = const()[name = tensor("range_1d_35_start_0"), val = tensor(0)]; + tensor range_1d_35_step_0 = const()[name = tensor("range_1d_35_step_0"), val = tensor(1)]; + tensor range_1d_35 = range_1d(end = reduce_prod_35, start = range_1d_35_start_0, step = range_1d_35_step_0)[name = tensor("range_1d_35")]; + tensor reshape_175 = reshape(shape = shape_83, x = range_1d_35)[name = tensor("reshape_175")]; + tensor slice_by_index_35 = slice_by_index(begin = concat_123, begin_mask = new_cache_35_internal_tensor_assign_2_begin_mask_0, end = concat_124, end_mask = new_cache_35_internal_tensor_assign_2_end_mask_0, squeeze_mask = new_cache_35_internal_tensor_assign_2_squeeze_mask_0, stride = new_cache_35_internal_tensor_assign_2_stride_0, x = reshape_175)[name = tensor("slice_by_index_35")]; + tensor reshape_176_shape_0 = const()[name = tensor("reshape_176_shape_0"), val = tensor([-1])]; + tensor reshape_176 = reshape(shape = reshape_176_shape_0, x = slice_by_index_35)[name = tensor("reshape_176")]; + tensor reshape_177_shape_0 = const()[name = tensor("reshape_177_shape_0"), val = tensor([-1])]; + tensor reshape_177_cast_fp16 = reshape(shape = reshape_177_shape_0, x = var_7609_cast_fp16)[name = tensor("reshape_177_cast_fp16")]; + tensor reshape_178_shape_0 = const()[name = tensor("reshape_178_shape_0"), val = tensor([-1])]; + tensor reshape_178_cast_fp16 = reshape(shape = reshape_178_shape_0, x = reshape_174_cast_fp16)[name = tensor("reshape_178_cast_fp16")]; + tensor scatter_35_mode_0 = const()[name = tensor("scatter_35_mode_0"), val = tensor("update")]; + tensor scatter_35_axis_0 = const()[name = tensor("scatter_35_axis_0"), val = tensor(0)]; + tensor scatter_35_validate_indices_0 = const()[name = tensor("scatter_35_validate_indices_0"), val = tensor(false)]; + tensor scatter_35_cast_fp16 = scatter(axis = scatter_35_axis_0, data = reshape_178_cast_fp16, indices = reshape_176, mode = scatter_35_mode_0, updates = reshape_177_cast_fp16, validate_indices = scatter_35_validate_indices_0)[name = tensor("scatter_35_cast_fp16")]; + tensor reshape_179_cast_fp16 = reshape(shape = shape_83, x = scatter_35_cast_fp16)[name = tensor("reshape_179_cast_fp16")]; + tensor reshape_179_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("reshape_179_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor keys_103_begin_0 = const()[name = tensor("keys_103_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor keys_103_end_0 = const()[name = tensor("keys_103_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor keys_103_end_mask_0 = const()[name = tensor("keys_103_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor keys_103_squeeze_mask_0 = const()[name = tensor("keys_103_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor keys_103_cast_fp16 = slice_by_index(begin = keys_103_begin_0, end = keys_103_end_0, end_mask = keys_103_end_mask_0, squeeze_mask = keys_103_squeeze_mask_0, x = reshape_179_cast_fp16)[name = tensor("keys_103_cast_fp16")]; + tensor values_103_begin_0 = const()[name = tensor("values_103_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor values_103_end_0 = const()[name = tensor("values_103_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor values_103_end_mask_0 = const()[name = tensor("values_103_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor values_103_squeeze_mask_0 = const()[name = tensor("values_103_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor values_103_cast_fp16 = slice_by_index(begin = values_103_begin_0, end = values_103_end_0, end_mask = values_103_end_mask_0, squeeze_mask = values_103_squeeze_mask_0, x = reshape_179_cast_fp16)[name = tensor("values_103_cast_fp16")]; + tensor var_7621_cast_fp16 = not_equal(x = keys_103_cast_fp16, y = keys_103_cast_fp16)[name = tensor("op_7621_cast_fp16")]; + tensor keys_105_cast_fp16 = select(a = var_504_to_fp16, b = keys_103_cast_fp16, cond = var_7621_cast_fp16)[name = tensor("keys_105_cast_fp16")]; + tensor var_7629_cast_fp16 = not_equal(x = values_103_cast_fp16, y = values_103_cast_fp16)[name = tensor("op_7629_cast_fp16")]; + tensor values_105_cast_fp16 = select(a = var_504_to_fp16, b = values_103_cast_fp16, cond = var_7629_cast_fp16)[name = tensor("values_105_cast_fp16")]; + tensor var_7653 = const()[name = tensor("op_7653"), val = tensor([0, 2, 1, 3])]; + tensor var_7666 = const()[name = tensor("op_7666"), val = tensor([1, 1, 1])]; + tensor var_7667_cast_fp16 = reshape(shape = var_7666, x = position17_to_fp16)[name = tensor("op_7667_cast_fp16")]; + tensor var_7684_to_fp16 = const()[name = tensor("op_7684_to_fp16"), val = tensor(0x1p+0)]; + tensor valid_len_35_cast_fp16 = add(x = var_7667_cast_fp16, y = var_7684_to_fp16)[name = tensor("valid_len_35_cast_fp16")]; + tensor valid_mask_35_cast_fp16 = less(x = k_positions_1_promoted_to_fp16, y = valid_len_35_cast_fp16)[name = tensor("valid_mask_35_cast_fp16")]; + tensor causal_mask_35_cast_fp16 = less_equal(x = k_positions_1_promoted_to_fp16, y = var_7667_cast_fp16)[name = tensor("causal_mask_35_cast_fp16")]; + tensor attn_mask_69 = logical_and(x = valid_mask_35_cast_fp16, y = causal_mask_35_cast_fp16)[name = tensor("attn_mask_69")]; + tensor attn_mask_71_axes_0 = const()[name = tensor("attn_mask_71_axes_0"), val = tensor([1])]; + tensor attn_mask_71 = expand_dims(axes = attn_mask_71_axes_0, x = attn_mask_69)[name = tensor("attn_mask_71")]; + tensor var_7702_transpose_x_0 = const()[name = tensor("op_7702_transpose_x_0"), val = tensor(false)]; + tensor var_7702_transpose_y_0 = const()[name = tensor("op_7702_transpose_y_0"), val = tensor(false)]; + tensor transpose_106_perm_0 = const()[name = tensor("transpose_106_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_107_perm_0 = const()[name = tensor("transpose_107_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_107 = transpose(perm = transpose_107_perm_0, x = keys_105_cast_fp16)[name = tensor("transpose_145")]; + tensor transpose_106 = transpose(perm = transpose_106_perm_0, x = q_105_cast_fp16)[name = tensor("transpose_146")]; + tensor var_7702_cast_fp16 = matmul(transpose_x = var_7702_transpose_x_0, transpose_y = var_7702_transpose_y_0, x = transpose_106, y = transpose_107)[name = tensor("op_7702_cast_fp16")]; + tensor var_7696_to_fp16 = const()[name = tensor("op_7696_to_fp16"), val = tensor([0x1p-3])]; + tensor attn_weights_103_cast_fp16 = mul(x = var_7702_cast_fp16, y = var_7696_to_fp16)[name = tensor("attn_weights_103_cast_fp16")]; + tensor var_7704 = logical_not(x = attn_mask_71)[name = tensor("op_7704")]; + tensor var_7705_to_fp16 = const()[name = tensor("op_7705_to_fp16"), val = tensor(-inf)]; + tensor attn_weights_105_cast_fp16 = select(a = var_7705_to_fp16, b = attn_weights_103_cast_fp16, cond = var_7704)[name = tensor("attn_weights_105_cast_fp16")]; + tensor var_7707 = const()[name = tensor("op_7707"), val = tensor(-1)]; + tensor attn_weights_107_cast_fp16 = softmax(axis = var_7707, x = attn_weights_105_cast_fp16)[name = tensor("attn_weights_107_cast_fp16")]; + tensor attn_output_35_transpose_x_0 = const()[name = tensor("attn_output_35_transpose_x_0"), val = tensor(false)]; + tensor attn_output_35_transpose_y_0 = const()[name = tensor("attn_output_35_transpose_y_0"), val = tensor(false)]; + tensor values_107_cast_fp16 = transpose(perm = var_7653, x = values_105_cast_fp16)[name = tensor("transpose_147")]; + tensor attn_output_35_cast_fp16 = matmul(transpose_x = attn_output_35_transpose_x_0, transpose_y = attn_output_35_transpose_y_0, x = attn_weights_107_cast_fp16, y = values_107_cast_fp16)[name = tensor("attn_output_35_cast_fp16")]; + tensor var_7715 = const()[name = tensor("op_7715"), val = tensor([0, 2, 1, 3])]; + tensor var_7718 = const()[name = tensor("op_7718"), val = tensor([1, 1, 1024])]; + tensor var_7716_cast_fp16 = transpose(perm = var_7715, x = attn_output_35_cast_fp16)[name = tensor("transpose_144")]; + tensor input_175_cast_fp16 = reshape(shape = var_7718, x = var_7716_cast_fp16)[name = tensor("input_175_cast_fp16")]; + tensor linear_70_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = attn17_out_proj_weight_cast_fp16, x = input_175_cast_fp16)[name = tensor("linear_70_cast_fp16")]; + tensor var_7724_to_fp16 = const()[name = tensor("op_7724_to_fp16"), val = tensor(0x1p+0)]; + tensor var_7725_cast_fp16 = add(x = position17_to_fp16, y = var_7724_to_fp16)[name = tensor("op_7725_cast_fp16")]; + tensor var_7725_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_7725_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor input_177_cast_fp16 = add(x = input_173_cast_fp16, y = linear_70_cast_fp16)[name = tensor("input_177_cast_fp16")]; + tensor input_179_axes_0 = const()[name = tensor("input_179_axes_0"), val = tensor([-1])]; + tensor norm17_2_weight_to_fp16 = const()[name = tensor("norm17_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(303735040)))]; + tensor norm17_2_bias_to_fp16 = const()[name = tensor("norm17_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(303737152)))]; + tensor var_7729_to_fp16 = const()[name = tensor("op_7729_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_179_cast_fp16 = layer_norm(axes = input_179_axes_0, beta = norm17_2_bias_to_fp16, epsilon = var_7729_to_fp16, gamma = norm17_2_weight_to_fp16, x = input_177_cast_fp16)[name = tensor("input_179_cast_fp16")]; + tensor linear_71_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = linear17_1_weight_cast_fp16, x = input_179_cast_fp16)[name = tensor("linear_71_cast_fp16")]; + tensor input_181_mode_0 = const()[name = tensor("input_181_mode_0"), val = tensor("EXACT")]; + tensor input_181_cast_fp16 = gelu(mode = input_181_mode_0, x = linear_71_cast_fp16)[name = tensor("input_181_cast_fp16")]; + tensor linear_72_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = linear17_2_weight_cast_fp16, x = input_181_cast_fp16)[name = tensor("linear_72_cast_fp16")]; + tensor input_183_cast_fp16 = add(x = input_177_cast_fp16, y = linear_72_cast_fp16)[name = tensor("input_183_cast_fp16")]; + tensor x_37_axes_0 = const()[name = tensor("x_37_axes_0"), val = tensor([-1])]; + tensor norm18_1_weight_to_fp16 = const()[name = tensor("norm18_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(303739264)))]; + tensor norm18_1_bias_to_fp16 = const()[name = tensor("norm18_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(303741376)))]; + tensor var_7746_to_fp16 = const()[name = tensor("op_7746_to_fp16"), val = tensor(0x1.5p-17)]; + tensor x_37_cast_fp16 = layer_norm(axes = x_37_axes_0, beta = norm18_1_bias_to_fp16, epsilon = var_7746_to_fp16, gamma = norm18_1_weight_to_fp16, x = input_183_cast_fp16)[name = tensor("x_37_cast_fp16")]; + tensor linear_73_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = attn18_in_proj_weight_cast_fp16, x = x_37_cast_fp16)[name = tensor("linear_73_cast_fp16")]; + tensor var_7782 = const()[name = tensor("op_7782"), val = tensor([1, 1, 3, 16, 64])]; + tensor qkv_37_cast_fp16 = reshape(shape = var_7782, x = linear_73_cast_fp16)[name = tensor("qkv_37_cast_fp16")]; + tensor q_109_begin_0 = const()[name = tensor("q_109_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor q_109_end_0 = const()[name = tensor("q_109_end_0"), val = tensor([1, 1, 1, 16, 64])]; + tensor q_109_end_mask_0 = const()[name = tensor("q_109_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor q_109_squeeze_mask_0 = const()[name = tensor("q_109_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor q_109_cast_fp16 = slice_by_index(begin = q_109_begin_0, end = q_109_end_0, end_mask = q_109_end_mask_0, squeeze_mask = q_109_squeeze_mask_0, x = qkv_37_cast_fp16)[name = tensor("q_109_cast_fp16")]; + tensor k_73_begin_0 = const()[name = tensor("k_73_begin_0"), val = tensor([0, 0, 1, 0, 0])]; + tensor k_73_end_0 = const()[name = tensor("k_73_end_0"), val = tensor([1, 1, 2, 16, 64])]; + tensor k_73_end_mask_0 = const()[name = tensor("k_73_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor k_73_squeeze_mask_0 = const()[name = tensor("k_73_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor k_73_cast_fp16 = slice_by_index(begin = k_73_begin_0, end = k_73_end_0, end_mask = k_73_end_mask_0, squeeze_mask = k_73_squeeze_mask_0, x = qkv_37_cast_fp16)[name = tensor("k_73_cast_fp16")]; + tensor v_37_begin_0 = const()[name = tensor("v_37_begin_0"), val = tensor([0, 0, 2, 0, 0])]; + tensor v_37_end_0 = const()[name = tensor("v_37_end_0"), val = tensor([1, 1, 3, 16, 64])]; + tensor v_37_end_mask_0 = const()[name = tensor("v_37_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor v_37_squeeze_mask_0 = const()[name = tensor("v_37_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor v_37_cast_fp16 = slice_by_index(begin = v_37_begin_0, end = v_37_end_0, end_mask = v_37_end_mask_0, squeeze_mask = v_37_squeeze_mask_0, x = qkv_37_cast_fp16)[name = tensor("v_37_cast_fp16")]; + tensor var_7886 = const()[name = tensor("op_7886"), val = tensor([1, 1, 1, 1])]; + tensor position18_to_fp16_dtype_0 = const()[name = tensor("position18_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor position18_to_fp16 = cast(dtype = position18_to_fp16_dtype_0, x = position18)[name = tensor("cast_463")]; + tensor ts_113_cast_fp16 = reshape(shape = var_7886, x = position18_to_fp16)[name = tensor("ts_113_cast_fp16")]; + tensor var_7890 = const()[name = tensor("op_7890"), val = tensor([1, 1, 16, 32, 2])]; + tensor q_complex_37_cast_fp16 = reshape(shape = var_7890, x = q_109_cast_fp16)[name = tensor("q_complex_37_cast_fp16")]; + tensor var_7894 = const()[name = tensor("op_7894"), val = tensor([1, 1, 16, 32, 2])]; + tensor k_complex_37_cast_fp16 = reshape(shape = var_7894, x = k_73_cast_fp16)[name = tensor("k_complex_37_cast_fp16")]; + tensor var_7898_begin_0 = const()[name = tensor("op_7898_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_7898_end_0 = const()[name = tensor("op_7898_end_0"), val = tensor([1, 1, 16, 32, 1])]; + tensor var_7898_end_mask_0 = const()[name = tensor("op_7898_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_7898_squeeze_mask_0 = const()[name = tensor("op_7898_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_7898_cast_fp16 = slice_by_index(begin = var_7898_begin_0, end = var_7898_end_0, end_mask = var_7898_end_mask_0, squeeze_mask = var_7898_squeeze_mask_0, x = q_complex_37_cast_fp16)[name = tensor("op_7898_cast_fp16")]; + tensor var_7906_begin_0 = const()[name = tensor("op_7906_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_7906_end_0 = const()[name = tensor("op_7906_end_0"), val = tensor([1, 1, 16, 32, 2])]; + tensor var_7906_end_mask_0 = const()[name = tensor("op_7906_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_7906_squeeze_mask_0 = const()[name = tensor("op_7906_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_7906_cast_fp16 = slice_by_index(begin = var_7906_begin_0, end = var_7906_end_0, end_mask = var_7906_end_mask_0, squeeze_mask = var_7906_squeeze_mask_0, x = q_complex_37_cast_fp16)[name = tensor("op_7906_cast_fp16")]; + tensor var_7914_begin_0 = const()[name = tensor("op_7914_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_7914_end_0 = const()[name = tensor("op_7914_end_0"), val = tensor([1, 1, 16, 32, 1])]; + tensor var_7914_end_mask_0 = const()[name = tensor("op_7914_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_7914_squeeze_mask_0 = const()[name = tensor("op_7914_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_7914_cast_fp16 = slice_by_index(begin = var_7914_begin_0, end = var_7914_end_0, end_mask = var_7914_end_mask_0, squeeze_mask = var_7914_squeeze_mask_0, x = k_complex_37_cast_fp16)[name = tensor("op_7914_cast_fp16")]; + tensor var_7922_begin_0 = const()[name = tensor("op_7922_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_7922_end_0 = const()[name = tensor("op_7922_end_0"), val = tensor([1, 1, 16, 32, 2])]; + tensor var_7922_end_mask_0 = const()[name = tensor("op_7922_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_7922_squeeze_mask_0 = const()[name = tensor("op_7922_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_7922_cast_fp16 = slice_by_index(begin = var_7922_begin_0, end = var_7922_end_0, end_mask = var_7922_end_mask_0, squeeze_mask = var_7922_squeeze_mask_0, x = k_complex_37_cast_fp16)[name = tensor("op_7922_cast_fp16")]; + tensor freqs_37_to_fp16 = const()[name = tensor("freqs_37_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(303743488)))]; + tensor var_7928_cast_fp16 = mul(x = freqs_37_to_fp16, y = ts_113_cast_fp16)[name = tensor("op_7928_cast_fp16")]; + tensor rotr_37_cast_fp16 = cos(x = var_7928_cast_fp16)[name = tensor("rotr_37_cast_fp16")]; + tensor roti_37_cast_fp16 = sin(x = var_7928_cast_fp16)[name = tensor("roti_37_cast_fp16")]; + tensor var_7932_cast_fp16 = mul(x = var_7898_cast_fp16, y = rotr_37_cast_fp16)[name = tensor("op_7932_cast_fp16")]; + tensor var_7933_cast_fp16 = mul(x = var_7906_cast_fp16, y = roti_37_cast_fp16)[name = tensor("op_7933_cast_fp16")]; + tensor qor_73_cast_fp16 = sub(x = var_7932_cast_fp16, y = var_7933_cast_fp16)[name = tensor("qor_73_cast_fp16")]; + tensor var_7936_cast_fp16 = mul(x = var_7898_cast_fp16, y = roti_37_cast_fp16)[name = tensor("op_7936_cast_fp16")]; + tensor var_7937_cast_fp16 = mul(x = var_7906_cast_fp16, y = rotr_37_cast_fp16)[name = tensor("op_7937_cast_fp16")]; + tensor qoi_73_cast_fp16 = add(x = var_7936_cast_fp16, y = var_7937_cast_fp16)[name = tensor("qoi_73_cast_fp16")]; + tensor var_7940_cast_fp16 = mul(x = var_7914_cast_fp16, y = rotr_37_cast_fp16)[name = tensor("op_7940_cast_fp16")]; + tensor var_7941_cast_fp16 = mul(x = var_7922_cast_fp16, y = roti_37_cast_fp16)[name = tensor("op_7941_cast_fp16")]; + tensor kor_73_cast_fp16 = sub(x = var_7940_cast_fp16, y = var_7941_cast_fp16)[name = tensor("kor_73_cast_fp16")]; + tensor var_7944_cast_fp16 = mul(x = var_7914_cast_fp16, y = roti_37_cast_fp16)[name = tensor("op_7944_cast_fp16")]; + tensor var_7945_cast_fp16 = mul(x = var_7922_cast_fp16, y = rotr_37_cast_fp16)[name = tensor("op_7945_cast_fp16")]; + tensor koi_73_cast_fp16 = add(x = var_7944_cast_fp16, y = var_7945_cast_fp16)[name = tensor("koi_73_cast_fp16")]; + tensor qo_37_axis_0 = const()[name = tensor("qo_37_axis_0"), val = tensor(-1)]; + tensor qo_37_cast_fp16 = stack(axis = qo_37_axis_0, values = (qor_73_cast_fp16, qoi_73_cast_fp16))[name = tensor("qo_37_cast_fp16")]; + tensor ko_37_axis_0 = const()[name = tensor("ko_37_axis_0"), val = tensor(-1)]; + tensor ko_37_cast_fp16 = stack(axis = ko_37_axis_0, values = (kor_73_cast_fp16, koi_73_cast_fp16))[name = tensor("ko_37_cast_fp16")]; + tensor var_7974 = const()[name = tensor("op_7974"), val = tensor([1, 1, 16, 64])]; + tensor q_111_cast_fp16 = reshape(shape = var_7974, x = qo_37_cast_fp16)[name = tensor("q_111_cast_fp16")]; + tensor var_7976 = const()[name = tensor("op_7976"), val = tensor([1, 1, 16, 64])]; + tensor k_75_cast_fp16 = reshape(shape = var_7976, x = ko_37_cast_fp16)[name = tensor("k_75_cast_fp16")]; + tensor _inversed_7998_y_0_to_fp16 = const()[name = tensor("_inversed_7998_y_0_to_fp16"), val = tensor(0x1p-9)]; + tensor _inversed_7998_cast_fp16 = mul(x = ts_113_cast_fp16, y = _inversed_7998_y_0_to_fp16)[name = tensor("_inversed_7998_cast_fp16")]; + tensor var_7999_cast_fp16 = floor(x = _inversed_7998_cast_fp16)[name = tensor("op_7999_cast_fp16")]; + tensor var_8000_to_fp16 = const()[name = tensor("op_8000_to_fp16"), val = tensor(0x1p+9)]; + tensor var_8001_cast_fp16 = mul(x = var_7999_cast_fp16, y = var_8000_to_fp16)[name = tensor("op_8001_cast_fp16")]; + tensor write_indices_float_75_cast_fp16 = sub(x = ts_113_cast_fp16, y = var_8001_cast_fp16)[name = tensor("write_indices_float_75_cast_fp16")]; + tensor var_8008_dtype_0 = const()[name = tensor("op_8008_dtype_0"), val = tensor("int32")]; + tensor write_indices_37_reps_0 = const()[name = tensor("write_indices_37_reps_0"), val = tensor([1, 1, 16, 64])]; + tensor write_indices_float_75_cast_fp16_to_int32 = cast(dtype = var_8008_dtype_0, x = write_indices_float_75_cast_fp16)[name = tensor("cast_462")]; + tensor write_indices_37 = tile(reps = write_indices_37_reps_0, x = write_indices_float_75_cast_fp16_to_int32)[name = tensor("write_indices_37")]; + tensor var_8016_begin_0 = const()[name = tensor("op_8016_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_8016_end_0 = const()[name = tensor("op_8016_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor var_8016_end_mask_0 = const()[name = tensor("op_8016_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_8016_squeeze_mask_0 = const()[name = tensor("op_8016_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor cache18_to_fp16_dtype_0 = const()[name = tensor("cache18_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor cache18_to_fp16 = cast(dtype = cache18_to_fp16_dtype_0, x = cache18)[name = tensor("cast_461")]; + tensor var_8016_cast_fp16 = slice_by_index(begin = var_8016_begin_0, end = var_8016_end_0, end_mask = var_8016_end_mask_0, squeeze_mask = var_8016_squeeze_mask_0, x = cache18_to_fp16)[name = tensor("op_8016_cast_fp16")]; + tensor var_8018_axis_0 = const()[name = tensor("op_8018_axis_0"), val = tensor(1)]; + tensor var_8018_mode_0 = const()[name = tensor("op_8018_mode_0"), val = tensor("update")]; + tensor var_8018_validate_indices_0 = const()[name = tensor("op_8018_validate_indices_0"), val = tensor(false)]; + tensor var_8018_cast_fp16 = scatter_along_axis(axis = var_8018_axis_0, data = var_8016_cast_fp16, indices = write_indices_37, mode = var_8018_mode_0, updates = k_75_cast_fp16, validate_indices = var_8018_validate_indices_0)[name = tensor("op_8018_cast_fp16")]; + tensor concat_128 = const()[name = tensor("concat_128"), val = tensor([0, 0, 0, 0, 0])]; + tensor concat_129 = const()[name = tensor("concat_129"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_37_internal_tensor_assign_1_stride_0 = const()[name = tensor("new_cache_37_internal_tensor_assign_1_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_37_internal_tensor_assign_1_begin_mask_0 = const()[name = tensor("new_cache_37_internal_tensor_assign_1_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_37_internal_tensor_assign_1_end_mask_0 = const()[name = tensor("new_cache_37_internal_tensor_assign_1_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_37_internal_tensor_assign_1_squeeze_mask_0 = const()[name = tensor("new_cache_37_internal_tensor_assign_1_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_84 = const()[name = tensor("shape_84"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_36 = const()[name = tensor("reduce_prod_36"), val = tensor(1048576)]; + tensor range_1d_36_start_0 = const()[name = tensor("range_1d_36_start_0"), val = tensor(0)]; + tensor range_1d_36_step_0 = const()[name = tensor("range_1d_36_step_0"), val = tensor(1)]; + tensor range_1d_36 = range_1d(end = reduce_prod_36, start = range_1d_36_start_0, step = range_1d_36_step_0)[name = tensor("range_1d_36")]; + tensor reshape_180 = reshape(shape = shape_84, x = range_1d_36)[name = tensor("reshape_180")]; + tensor slice_by_index_36 = slice_by_index(begin = concat_128, begin_mask = new_cache_37_internal_tensor_assign_1_begin_mask_0, end = concat_129, end_mask = new_cache_37_internal_tensor_assign_1_end_mask_0, squeeze_mask = new_cache_37_internal_tensor_assign_1_squeeze_mask_0, stride = new_cache_37_internal_tensor_assign_1_stride_0, x = reshape_180)[name = tensor("slice_by_index_36")]; + tensor reshape_181_shape_0 = const()[name = tensor("reshape_181_shape_0"), val = tensor([-1])]; + tensor reshape_181 = reshape(shape = reshape_181_shape_0, x = slice_by_index_36)[name = tensor("reshape_181")]; + tensor reshape_182_shape_0 = const()[name = tensor("reshape_182_shape_0"), val = tensor([-1])]; + tensor reshape_182_cast_fp16 = reshape(shape = reshape_182_shape_0, x = var_8018_cast_fp16)[name = tensor("reshape_182_cast_fp16")]; + tensor reshape_183_shape_0 = const()[name = tensor("reshape_183_shape_0"), val = tensor([-1])]; + tensor reshape_183_cast_fp16 = reshape(shape = reshape_183_shape_0, x = cache18_to_fp16)[name = tensor("reshape_183_cast_fp16")]; + tensor scatter_36_mode_0 = const()[name = tensor("scatter_36_mode_0"), val = tensor("update")]; + tensor scatter_36_axis_0 = const()[name = tensor("scatter_36_axis_0"), val = tensor(0)]; + tensor scatter_36_validate_indices_0 = const()[name = tensor("scatter_36_validate_indices_0"), val = tensor(false)]; + tensor scatter_36_cast_fp16 = scatter(axis = scatter_36_axis_0, data = reshape_183_cast_fp16, indices = reshape_181, mode = scatter_36_mode_0, updates = reshape_182_cast_fp16, validate_indices = scatter_36_validate_indices_0)[name = tensor("scatter_36_cast_fp16")]; + tensor reshape_184_cast_fp16 = reshape(shape = shape_84, x = scatter_36_cast_fp16)[name = tensor("reshape_184_cast_fp16")]; + tensor var_8026_begin_0 = const()[name = tensor("op_8026_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor var_8026_end_0 = const()[name = tensor("op_8026_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor var_8026_end_mask_0 = const()[name = tensor("op_8026_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_8026_squeeze_mask_0 = const()[name = tensor("op_8026_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor var_8026_cast_fp16 = slice_by_index(begin = var_8026_begin_0, end = var_8026_end_0, end_mask = var_8026_end_mask_0, squeeze_mask = var_8026_squeeze_mask_0, x = reshape_184_cast_fp16)[name = tensor("op_8026_cast_fp16")]; + tensor var_8028_axis_0 = const()[name = tensor("op_8028_axis_0"), val = tensor(1)]; + tensor var_8028_mode_0 = const()[name = tensor("op_8028_mode_0"), val = tensor("update")]; + tensor var_8028_validate_indices_0 = const()[name = tensor("op_8028_validate_indices_0"), val = tensor(false)]; + tensor var_8028_cast_fp16 = scatter_along_axis(axis = var_8028_axis_0, data = var_8026_cast_fp16, indices = write_indices_37, mode = var_8028_mode_0, updates = v_37_cast_fp16, validate_indices = var_8028_validate_indices_0)[name = tensor("op_8028_cast_fp16")]; + tensor concat_130 = const()[name = tensor("concat_130"), val = tensor([1, 0, 0, 0, 0])]; + tensor concat_131 = const()[name = tensor("concat_131"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_37_internal_tensor_assign_2_stride_0 = const()[name = tensor("new_cache_37_internal_tensor_assign_2_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_37_internal_tensor_assign_2_begin_mask_0 = const()[name = tensor("new_cache_37_internal_tensor_assign_2_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_37_internal_tensor_assign_2_end_mask_0 = const()[name = tensor("new_cache_37_internal_tensor_assign_2_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_37_internal_tensor_assign_2_squeeze_mask_0 = const()[name = tensor("new_cache_37_internal_tensor_assign_2_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_85 = const()[name = tensor("shape_85"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_37 = const()[name = tensor("reduce_prod_37"), val = tensor(1048576)]; + tensor range_1d_37_start_0 = const()[name = tensor("range_1d_37_start_0"), val = tensor(0)]; + tensor range_1d_37_step_0 = const()[name = tensor("range_1d_37_step_0"), val = tensor(1)]; + tensor range_1d_37 = range_1d(end = reduce_prod_37, start = range_1d_37_start_0, step = range_1d_37_step_0)[name = tensor("range_1d_37")]; + tensor reshape_185 = reshape(shape = shape_85, x = range_1d_37)[name = tensor("reshape_185")]; + tensor slice_by_index_37 = slice_by_index(begin = concat_130, begin_mask = new_cache_37_internal_tensor_assign_2_begin_mask_0, end = concat_131, end_mask = new_cache_37_internal_tensor_assign_2_end_mask_0, squeeze_mask = new_cache_37_internal_tensor_assign_2_squeeze_mask_0, stride = new_cache_37_internal_tensor_assign_2_stride_0, x = reshape_185)[name = tensor("slice_by_index_37")]; + tensor reshape_186_shape_0 = const()[name = tensor("reshape_186_shape_0"), val = tensor([-1])]; + tensor reshape_186 = reshape(shape = reshape_186_shape_0, x = slice_by_index_37)[name = tensor("reshape_186")]; + tensor reshape_187_shape_0 = const()[name = tensor("reshape_187_shape_0"), val = tensor([-1])]; + tensor reshape_187_cast_fp16 = reshape(shape = reshape_187_shape_0, x = var_8028_cast_fp16)[name = tensor("reshape_187_cast_fp16")]; + tensor reshape_188_shape_0 = const()[name = tensor("reshape_188_shape_0"), val = tensor([-1])]; + tensor reshape_188_cast_fp16 = reshape(shape = reshape_188_shape_0, x = reshape_184_cast_fp16)[name = tensor("reshape_188_cast_fp16")]; + tensor scatter_37_mode_0 = const()[name = tensor("scatter_37_mode_0"), val = tensor("update")]; + tensor scatter_37_axis_0 = const()[name = tensor("scatter_37_axis_0"), val = tensor(0)]; + tensor scatter_37_validate_indices_0 = const()[name = tensor("scatter_37_validate_indices_0"), val = tensor(false)]; + tensor scatter_37_cast_fp16 = scatter(axis = scatter_37_axis_0, data = reshape_188_cast_fp16, indices = reshape_186, mode = scatter_37_mode_0, updates = reshape_187_cast_fp16, validate_indices = scatter_37_validate_indices_0)[name = tensor("scatter_37_cast_fp16")]; + tensor reshape_189_cast_fp16 = reshape(shape = shape_85, x = scatter_37_cast_fp16)[name = tensor("reshape_189_cast_fp16")]; + tensor reshape_189_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("reshape_189_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor keys_109_begin_0 = const()[name = tensor("keys_109_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor keys_109_end_0 = const()[name = tensor("keys_109_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor keys_109_end_mask_0 = const()[name = tensor("keys_109_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor keys_109_squeeze_mask_0 = const()[name = tensor("keys_109_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor keys_109_cast_fp16 = slice_by_index(begin = keys_109_begin_0, end = keys_109_end_0, end_mask = keys_109_end_mask_0, squeeze_mask = keys_109_squeeze_mask_0, x = reshape_189_cast_fp16)[name = tensor("keys_109_cast_fp16")]; + tensor values_109_begin_0 = const()[name = tensor("values_109_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor values_109_end_0 = const()[name = tensor("values_109_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor values_109_end_mask_0 = const()[name = tensor("values_109_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor values_109_squeeze_mask_0 = const()[name = tensor("values_109_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor values_109_cast_fp16 = slice_by_index(begin = values_109_begin_0, end = values_109_end_0, end_mask = values_109_end_mask_0, squeeze_mask = values_109_squeeze_mask_0, x = reshape_189_cast_fp16)[name = tensor("values_109_cast_fp16")]; + tensor var_8040_cast_fp16 = not_equal(x = keys_109_cast_fp16, y = keys_109_cast_fp16)[name = tensor("op_8040_cast_fp16")]; + tensor keys_111_cast_fp16 = select(a = var_504_to_fp16, b = keys_109_cast_fp16, cond = var_8040_cast_fp16)[name = tensor("keys_111_cast_fp16")]; + tensor var_8048_cast_fp16 = not_equal(x = values_109_cast_fp16, y = values_109_cast_fp16)[name = tensor("op_8048_cast_fp16")]; + tensor values_111_cast_fp16 = select(a = var_504_to_fp16, b = values_109_cast_fp16, cond = var_8048_cast_fp16)[name = tensor("values_111_cast_fp16")]; + tensor var_8072 = const()[name = tensor("op_8072"), val = tensor([0, 2, 1, 3])]; + tensor var_8085 = const()[name = tensor("op_8085"), val = tensor([1, 1, 1])]; + tensor var_8086_cast_fp16 = reshape(shape = var_8085, x = position18_to_fp16)[name = tensor("op_8086_cast_fp16")]; + tensor var_8103_to_fp16 = const()[name = tensor("op_8103_to_fp16"), val = tensor(0x1p+0)]; + tensor valid_len_37_cast_fp16 = add(x = var_8086_cast_fp16, y = var_8103_to_fp16)[name = tensor("valid_len_37_cast_fp16")]; + tensor valid_mask_37_cast_fp16 = less(x = k_positions_1_promoted_to_fp16, y = valid_len_37_cast_fp16)[name = tensor("valid_mask_37_cast_fp16")]; + tensor causal_mask_37_cast_fp16 = less_equal(x = k_positions_1_promoted_to_fp16, y = var_8086_cast_fp16)[name = tensor("causal_mask_37_cast_fp16")]; + tensor attn_mask_73 = logical_and(x = valid_mask_37_cast_fp16, y = causal_mask_37_cast_fp16)[name = tensor("attn_mask_73")]; + tensor attn_mask_75_axes_0 = const()[name = tensor("attn_mask_75_axes_0"), val = tensor([1])]; + tensor attn_mask_75 = expand_dims(axes = attn_mask_75_axes_0, x = attn_mask_73)[name = tensor("attn_mask_75")]; + tensor var_8121_transpose_x_0 = const()[name = tensor("op_8121_transpose_x_0"), val = tensor(false)]; + tensor var_8121_transpose_y_0 = const()[name = tensor("op_8121_transpose_y_0"), val = tensor(false)]; + tensor transpose_108_perm_0 = const()[name = tensor("transpose_108_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_109_perm_0 = const()[name = tensor("transpose_109_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_109 = transpose(perm = transpose_109_perm_0, x = keys_111_cast_fp16)[name = tensor("transpose_141")]; + tensor transpose_108 = transpose(perm = transpose_108_perm_0, x = q_111_cast_fp16)[name = tensor("transpose_142")]; + tensor var_8121_cast_fp16 = matmul(transpose_x = var_8121_transpose_x_0, transpose_y = var_8121_transpose_y_0, x = transpose_108, y = transpose_109)[name = tensor("op_8121_cast_fp16")]; + tensor var_8115_to_fp16 = const()[name = tensor("op_8115_to_fp16"), val = tensor([0x1p-3])]; + tensor attn_weights_109_cast_fp16 = mul(x = var_8121_cast_fp16, y = var_8115_to_fp16)[name = tensor("attn_weights_109_cast_fp16")]; + tensor var_8123 = logical_not(x = attn_mask_75)[name = tensor("op_8123")]; + tensor var_8124_to_fp16 = const()[name = tensor("op_8124_to_fp16"), val = tensor(-inf)]; + tensor attn_weights_111_cast_fp16 = select(a = var_8124_to_fp16, b = attn_weights_109_cast_fp16, cond = var_8123)[name = tensor("attn_weights_111_cast_fp16")]; + tensor var_8126 = const()[name = tensor("op_8126"), val = tensor(-1)]; + tensor attn_weights_113_cast_fp16 = softmax(axis = var_8126, x = attn_weights_111_cast_fp16)[name = tensor("attn_weights_113_cast_fp16")]; + tensor attn_output_37_transpose_x_0 = const()[name = tensor("attn_output_37_transpose_x_0"), val = tensor(false)]; + tensor attn_output_37_transpose_y_0 = const()[name = tensor("attn_output_37_transpose_y_0"), val = tensor(false)]; + tensor values_113_cast_fp16 = transpose(perm = var_8072, x = values_111_cast_fp16)[name = tensor("transpose_143")]; + tensor attn_output_37_cast_fp16 = matmul(transpose_x = attn_output_37_transpose_x_0, transpose_y = attn_output_37_transpose_y_0, x = attn_weights_113_cast_fp16, y = values_113_cast_fp16)[name = tensor("attn_output_37_cast_fp16")]; + tensor var_8134 = const()[name = tensor("op_8134"), val = tensor([0, 2, 1, 3])]; + tensor var_8137 = const()[name = tensor("op_8137"), val = tensor([1, 1, 1024])]; + tensor var_8135_cast_fp16 = transpose(perm = var_8134, x = attn_output_37_cast_fp16)[name = tensor("transpose_140")]; + tensor input_185_cast_fp16 = reshape(shape = var_8137, x = var_8135_cast_fp16)[name = tensor("input_185_cast_fp16")]; + tensor linear_74_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = attn18_out_proj_weight_cast_fp16, x = input_185_cast_fp16)[name = tensor("linear_74_cast_fp16")]; + tensor var_8143_to_fp16 = const()[name = tensor("op_8143_to_fp16"), val = tensor(0x1p+0)]; + tensor var_8144_cast_fp16 = add(x = position18_to_fp16, y = var_8143_to_fp16)[name = tensor("op_8144_cast_fp16")]; + tensor var_8144_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_8144_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor input_187_cast_fp16 = add(x = input_183_cast_fp16, y = linear_74_cast_fp16)[name = tensor("input_187_cast_fp16")]; + tensor input_189_axes_0 = const()[name = tensor("input_189_axes_0"), val = tensor([-1])]; + tensor norm18_2_weight_to_fp16 = const()[name = tensor("norm18_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(303743616)))]; + tensor norm18_2_bias_to_fp16 = const()[name = tensor("norm18_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(303745728)))]; + tensor var_8148_to_fp16 = const()[name = tensor("op_8148_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_189_cast_fp16 = layer_norm(axes = input_189_axes_0, beta = norm18_2_bias_to_fp16, epsilon = var_8148_to_fp16, gamma = norm18_2_weight_to_fp16, x = input_187_cast_fp16)[name = tensor("input_189_cast_fp16")]; + tensor linear_75_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = linear18_1_weight_cast_fp16, x = input_189_cast_fp16)[name = tensor("linear_75_cast_fp16")]; + tensor input_191_mode_0 = const()[name = tensor("input_191_mode_0"), val = tensor("EXACT")]; + tensor input_191_cast_fp16 = gelu(mode = input_191_mode_0, x = linear_75_cast_fp16)[name = tensor("input_191_cast_fp16")]; + tensor linear_76_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = linear18_2_weight_cast_fp16, x = input_191_cast_fp16)[name = tensor("linear_76_cast_fp16")]; + tensor input_193_cast_fp16 = add(x = input_187_cast_fp16, y = linear_76_cast_fp16)[name = tensor("input_193_cast_fp16")]; + tensor x_39_axes_0 = const()[name = tensor("x_39_axes_0"), val = tensor([-1])]; + tensor norm19_1_weight_to_fp16 = const()[name = tensor("norm19_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(303747840)))]; + tensor norm19_1_bias_to_fp16 = const()[name = tensor("norm19_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(303749952)))]; + tensor var_8165_to_fp16 = const()[name = tensor("op_8165_to_fp16"), val = tensor(0x1.5p-17)]; + tensor x_39_cast_fp16 = layer_norm(axes = x_39_axes_0, beta = norm19_1_bias_to_fp16, epsilon = var_8165_to_fp16, gamma = norm19_1_weight_to_fp16, x = input_193_cast_fp16)[name = tensor("x_39_cast_fp16")]; + tensor linear_77_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = attn19_in_proj_weight_cast_fp16, x = x_39_cast_fp16)[name = tensor("linear_77_cast_fp16")]; + tensor var_8201 = const()[name = tensor("op_8201"), val = tensor([1, 1, 3, 16, 64])]; + tensor qkv_39_cast_fp16 = reshape(shape = var_8201, x = linear_77_cast_fp16)[name = tensor("qkv_39_cast_fp16")]; + tensor q_115_begin_0 = const()[name = tensor("q_115_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor q_115_end_0 = const()[name = tensor("q_115_end_0"), val = tensor([1, 1, 1, 16, 64])]; + tensor q_115_end_mask_0 = const()[name = tensor("q_115_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor q_115_squeeze_mask_0 = const()[name = tensor("q_115_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor q_115_cast_fp16 = slice_by_index(begin = q_115_begin_0, end = q_115_end_0, end_mask = q_115_end_mask_0, squeeze_mask = q_115_squeeze_mask_0, x = qkv_39_cast_fp16)[name = tensor("q_115_cast_fp16")]; + tensor k_77_begin_0 = const()[name = tensor("k_77_begin_0"), val = tensor([0, 0, 1, 0, 0])]; + tensor k_77_end_0 = const()[name = tensor("k_77_end_0"), val = tensor([1, 1, 2, 16, 64])]; + tensor k_77_end_mask_0 = const()[name = tensor("k_77_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor k_77_squeeze_mask_0 = const()[name = tensor("k_77_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor k_77_cast_fp16 = slice_by_index(begin = k_77_begin_0, end = k_77_end_0, end_mask = k_77_end_mask_0, squeeze_mask = k_77_squeeze_mask_0, x = qkv_39_cast_fp16)[name = tensor("k_77_cast_fp16")]; + tensor v_39_begin_0 = const()[name = tensor("v_39_begin_0"), val = tensor([0, 0, 2, 0, 0])]; + tensor v_39_end_0 = const()[name = tensor("v_39_end_0"), val = tensor([1, 1, 3, 16, 64])]; + tensor v_39_end_mask_0 = const()[name = tensor("v_39_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor v_39_squeeze_mask_0 = const()[name = tensor("v_39_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor v_39_cast_fp16 = slice_by_index(begin = v_39_begin_0, end = v_39_end_0, end_mask = v_39_end_mask_0, squeeze_mask = v_39_squeeze_mask_0, x = qkv_39_cast_fp16)[name = tensor("v_39_cast_fp16")]; + tensor var_8305 = const()[name = tensor("op_8305"), val = tensor([1, 1, 1, 1])]; + tensor position19_to_fp16_dtype_0 = const()[name = tensor("position19_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor position19_to_fp16 = cast(dtype = position19_to_fp16_dtype_0, x = position19)[name = tensor("cast_458")]; + tensor ts_119_cast_fp16 = reshape(shape = var_8305, x = position19_to_fp16)[name = tensor("ts_119_cast_fp16")]; + tensor var_8309 = const()[name = tensor("op_8309"), val = tensor([1, 1, 16, 32, 2])]; + tensor q_complex_39_cast_fp16 = reshape(shape = var_8309, x = q_115_cast_fp16)[name = tensor("q_complex_39_cast_fp16")]; + tensor var_8313 = const()[name = tensor("op_8313"), val = tensor([1, 1, 16, 32, 2])]; + tensor k_complex_39_cast_fp16 = reshape(shape = var_8313, x = k_77_cast_fp16)[name = tensor("k_complex_39_cast_fp16")]; + tensor var_8317_begin_0 = const()[name = tensor("op_8317_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_8317_end_0 = const()[name = tensor("op_8317_end_0"), val = tensor([1, 1, 16, 32, 1])]; + tensor var_8317_end_mask_0 = const()[name = tensor("op_8317_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_8317_squeeze_mask_0 = const()[name = tensor("op_8317_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_8317_cast_fp16 = slice_by_index(begin = var_8317_begin_0, end = var_8317_end_0, end_mask = var_8317_end_mask_0, squeeze_mask = var_8317_squeeze_mask_0, x = q_complex_39_cast_fp16)[name = tensor("op_8317_cast_fp16")]; + tensor var_8325_begin_0 = const()[name = tensor("op_8325_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_8325_end_0 = const()[name = tensor("op_8325_end_0"), val = tensor([1, 1, 16, 32, 2])]; + tensor var_8325_end_mask_0 = const()[name = tensor("op_8325_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_8325_squeeze_mask_0 = const()[name = tensor("op_8325_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_8325_cast_fp16 = slice_by_index(begin = var_8325_begin_0, end = var_8325_end_0, end_mask = var_8325_end_mask_0, squeeze_mask = var_8325_squeeze_mask_0, x = q_complex_39_cast_fp16)[name = tensor("op_8325_cast_fp16")]; + tensor var_8333_begin_0 = const()[name = tensor("op_8333_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_8333_end_0 = const()[name = tensor("op_8333_end_0"), val = tensor([1, 1, 16, 32, 1])]; + tensor var_8333_end_mask_0 = const()[name = tensor("op_8333_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_8333_squeeze_mask_0 = const()[name = tensor("op_8333_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_8333_cast_fp16 = slice_by_index(begin = var_8333_begin_0, end = var_8333_end_0, end_mask = var_8333_end_mask_0, squeeze_mask = var_8333_squeeze_mask_0, x = k_complex_39_cast_fp16)[name = tensor("op_8333_cast_fp16")]; + tensor var_8341_begin_0 = const()[name = tensor("op_8341_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_8341_end_0 = const()[name = tensor("op_8341_end_0"), val = tensor([1, 1, 16, 32, 2])]; + tensor var_8341_end_mask_0 = const()[name = tensor("op_8341_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_8341_squeeze_mask_0 = const()[name = tensor("op_8341_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_8341_cast_fp16 = slice_by_index(begin = var_8341_begin_0, end = var_8341_end_0, end_mask = var_8341_end_mask_0, squeeze_mask = var_8341_squeeze_mask_0, x = k_complex_39_cast_fp16)[name = tensor("op_8341_cast_fp16")]; + tensor freqs_39_to_fp16 = const()[name = tensor("freqs_39_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(303752064)))]; + tensor var_8347_cast_fp16 = mul(x = freqs_39_to_fp16, y = ts_119_cast_fp16)[name = tensor("op_8347_cast_fp16")]; + tensor rotr_39_cast_fp16 = cos(x = var_8347_cast_fp16)[name = tensor("rotr_39_cast_fp16")]; + tensor roti_39_cast_fp16 = sin(x = var_8347_cast_fp16)[name = tensor("roti_39_cast_fp16")]; + tensor var_8351_cast_fp16 = mul(x = var_8317_cast_fp16, y = rotr_39_cast_fp16)[name = tensor("op_8351_cast_fp16")]; + tensor var_8352_cast_fp16 = mul(x = var_8325_cast_fp16, y = roti_39_cast_fp16)[name = tensor("op_8352_cast_fp16")]; + tensor qor_77_cast_fp16 = sub(x = var_8351_cast_fp16, y = var_8352_cast_fp16)[name = tensor("qor_77_cast_fp16")]; + tensor var_8355_cast_fp16 = mul(x = var_8317_cast_fp16, y = roti_39_cast_fp16)[name = tensor("op_8355_cast_fp16")]; + tensor var_8356_cast_fp16 = mul(x = var_8325_cast_fp16, y = rotr_39_cast_fp16)[name = tensor("op_8356_cast_fp16")]; + tensor qoi_77_cast_fp16 = add(x = var_8355_cast_fp16, y = var_8356_cast_fp16)[name = tensor("qoi_77_cast_fp16")]; + tensor var_8359_cast_fp16 = mul(x = var_8333_cast_fp16, y = rotr_39_cast_fp16)[name = tensor("op_8359_cast_fp16")]; + tensor var_8360_cast_fp16 = mul(x = var_8341_cast_fp16, y = roti_39_cast_fp16)[name = tensor("op_8360_cast_fp16")]; + tensor kor_77_cast_fp16 = sub(x = var_8359_cast_fp16, y = var_8360_cast_fp16)[name = tensor("kor_77_cast_fp16")]; + tensor var_8363_cast_fp16 = mul(x = var_8333_cast_fp16, y = roti_39_cast_fp16)[name = tensor("op_8363_cast_fp16")]; + tensor var_8364_cast_fp16 = mul(x = var_8341_cast_fp16, y = rotr_39_cast_fp16)[name = tensor("op_8364_cast_fp16")]; + tensor koi_77_cast_fp16 = add(x = var_8363_cast_fp16, y = var_8364_cast_fp16)[name = tensor("koi_77_cast_fp16")]; + tensor qo_39_axis_0 = const()[name = tensor("qo_39_axis_0"), val = tensor(-1)]; + tensor qo_39_cast_fp16 = stack(axis = qo_39_axis_0, values = (qor_77_cast_fp16, qoi_77_cast_fp16))[name = tensor("qo_39_cast_fp16")]; + tensor ko_39_axis_0 = const()[name = tensor("ko_39_axis_0"), val = tensor(-1)]; + tensor ko_39_cast_fp16 = stack(axis = ko_39_axis_0, values = (kor_77_cast_fp16, koi_77_cast_fp16))[name = tensor("ko_39_cast_fp16")]; + tensor var_8393 = const()[name = tensor("op_8393"), val = tensor([1, 1, 16, 64])]; + tensor q_117_cast_fp16 = reshape(shape = var_8393, x = qo_39_cast_fp16)[name = tensor("q_117_cast_fp16")]; + tensor var_8395 = const()[name = tensor("op_8395"), val = tensor([1, 1, 16, 64])]; + tensor k_79_cast_fp16 = reshape(shape = var_8395, x = ko_39_cast_fp16)[name = tensor("k_79_cast_fp16")]; + tensor _inversed_8417_y_0_to_fp16 = const()[name = tensor("_inversed_8417_y_0_to_fp16"), val = tensor(0x1p-9)]; + tensor _inversed_8417_cast_fp16 = mul(x = ts_119_cast_fp16, y = _inversed_8417_y_0_to_fp16)[name = tensor("_inversed_8417_cast_fp16")]; + tensor var_8418_cast_fp16 = floor(x = _inversed_8417_cast_fp16)[name = tensor("op_8418_cast_fp16")]; + tensor var_8419_to_fp16 = const()[name = tensor("op_8419_to_fp16"), val = tensor(0x1p+9)]; + tensor var_8420_cast_fp16 = mul(x = var_8418_cast_fp16, y = var_8419_to_fp16)[name = tensor("op_8420_cast_fp16")]; + tensor write_indices_float_79_cast_fp16 = sub(x = ts_119_cast_fp16, y = var_8420_cast_fp16)[name = tensor("write_indices_float_79_cast_fp16")]; + tensor var_8427_dtype_0 = const()[name = tensor("op_8427_dtype_0"), val = tensor("int32")]; + tensor write_indices_39_reps_0 = const()[name = tensor("write_indices_39_reps_0"), val = tensor([1, 1, 16, 64])]; + tensor write_indices_float_79_cast_fp16_to_int32 = cast(dtype = var_8427_dtype_0, x = write_indices_float_79_cast_fp16)[name = tensor("cast_457")]; + tensor write_indices_39 = tile(reps = write_indices_39_reps_0, x = write_indices_float_79_cast_fp16_to_int32)[name = tensor("write_indices_39")]; + tensor var_8435_begin_0 = const()[name = tensor("op_8435_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_8435_end_0 = const()[name = tensor("op_8435_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor var_8435_end_mask_0 = const()[name = tensor("op_8435_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_8435_squeeze_mask_0 = const()[name = tensor("op_8435_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor cache19_to_fp16_dtype_0 = const()[name = tensor("cache19_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor cache19_to_fp16 = cast(dtype = cache19_to_fp16_dtype_0, x = cache19)[name = tensor("cast_456")]; + tensor var_8435_cast_fp16 = slice_by_index(begin = var_8435_begin_0, end = var_8435_end_0, end_mask = var_8435_end_mask_0, squeeze_mask = var_8435_squeeze_mask_0, x = cache19_to_fp16)[name = tensor("op_8435_cast_fp16")]; + tensor var_8437_axis_0 = const()[name = tensor("op_8437_axis_0"), val = tensor(1)]; + tensor var_8437_mode_0 = const()[name = tensor("op_8437_mode_0"), val = tensor("update")]; + tensor var_8437_validate_indices_0 = const()[name = tensor("op_8437_validate_indices_0"), val = tensor(false)]; + tensor var_8437_cast_fp16 = scatter_along_axis(axis = var_8437_axis_0, data = var_8435_cast_fp16, indices = write_indices_39, mode = var_8437_mode_0, updates = k_79_cast_fp16, validate_indices = var_8437_validate_indices_0)[name = tensor("op_8437_cast_fp16")]; + tensor concat_135 = const()[name = tensor("concat_135"), val = tensor([0, 0, 0, 0, 0])]; + tensor concat_136 = const()[name = tensor("concat_136"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_39_internal_tensor_assign_1_stride_0 = const()[name = tensor("new_cache_39_internal_tensor_assign_1_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_39_internal_tensor_assign_1_begin_mask_0 = const()[name = tensor("new_cache_39_internal_tensor_assign_1_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_39_internal_tensor_assign_1_end_mask_0 = const()[name = tensor("new_cache_39_internal_tensor_assign_1_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_39_internal_tensor_assign_1_squeeze_mask_0 = const()[name = tensor("new_cache_39_internal_tensor_assign_1_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_86 = const()[name = tensor("shape_86"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_38 = const()[name = tensor("reduce_prod_38"), val = tensor(1048576)]; + tensor range_1d_38_start_0 = const()[name = tensor("range_1d_38_start_0"), val = tensor(0)]; + tensor range_1d_38_step_0 = const()[name = tensor("range_1d_38_step_0"), val = tensor(1)]; + tensor range_1d_38 = range_1d(end = reduce_prod_38, start = range_1d_38_start_0, step = range_1d_38_step_0)[name = tensor("range_1d_38")]; + tensor reshape_190 = reshape(shape = shape_86, x = range_1d_38)[name = tensor("reshape_190")]; + tensor slice_by_index_38 = slice_by_index(begin = concat_135, begin_mask = new_cache_39_internal_tensor_assign_1_begin_mask_0, end = concat_136, end_mask = new_cache_39_internal_tensor_assign_1_end_mask_0, squeeze_mask = new_cache_39_internal_tensor_assign_1_squeeze_mask_0, stride = new_cache_39_internal_tensor_assign_1_stride_0, x = reshape_190)[name = tensor("slice_by_index_38")]; + tensor reshape_191_shape_0 = const()[name = tensor("reshape_191_shape_0"), val = tensor([-1])]; + tensor reshape_191 = reshape(shape = reshape_191_shape_0, x = slice_by_index_38)[name = tensor("reshape_191")]; + tensor reshape_192_shape_0 = const()[name = tensor("reshape_192_shape_0"), val = tensor([-1])]; + tensor reshape_192_cast_fp16 = reshape(shape = reshape_192_shape_0, x = var_8437_cast_fp16)[name = tensor("reshape_192_cast_fp16")]; + tensor reshape_193_shape_0 = const()[name = tensor("reshape_193_shape_0"), val = tensor([-1])]; + tensor reshape_193_cast_fp16 = reshape(shape = reshape_193_shape_0, x = cache19_to_fp16)[name = tensor("reshape_193_cast_fp16")]; + tensor scatter_38_mode_0 = const()[name = tensor("scatter_38_mode_0"), val = tensor("update")]; + tensor scatter_38_axis_0 = const()[name = tensor("scatter_38_axis_0"), val = tensor(0)]; + tensor scatter_38_validate_indices_0 = const()[name = tensor("scatter_38_validate_indices_0"), val = tensor(false)]; + tensor scatter_38_cast_fp16 = scatter(axis = scatter_38_axis_0, data = reshape_193_cast_fp16, indices = reshape_191, mode = scatter_38_mode_0, updates = reshape_192_cast_fp16, validate_indices = scatter_38_validate_indices_0)[name = tensor("scatter_38_cast_fp16")]; + tensor reshape_194_cast_fp16 = reshape(shape = shape_86, x = scatter_38_cast_fp16)[name = tensor("reshape_194_cast_fp16")]; + tensor var_8445_begin_0 = const()[name = tensor("op_8445_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor var_8445_end_0 = const()[name = tensor("op_8445_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor var_8445_end_mask_0 = const()[name = tensor("op_8445_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_8445_squeeze_mask_0 = const()[name = tensor("op_8445_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor var_8445_cast_fp16 = slice_by_index(begin = var_8445_begin_0, end = var_8445_end_0, end_mask = var_8445_end_mask_0, squeeze_mask = var_8445_squeeze_mask_0, x = reshape_194_cast_fp16)[name = tensor("op_8445_cast_fp16")]; + tensor var_8447_axis_0 = const()[name = tensor("op_8447_axis_0"), val = tensor(1)]; + tensor var_8447_mode_0 = const()[name = tensor("op_8447_mode_0"), val = tensor("update")]; + tensor var_8447_validate_indices_0 = const()[name = tensor("op_8447_validate_indices_0"), val = tensor(false)]; + tensor var_8447_cast_fp16 = scatter_along_axis(axis = var_8447_axis_0, data = var_8445_cast_fp16, indices = write_indices_39, mode = var_8447_mode_0, updates = v_39_cast_fp16, validate_indices = var_8447_validate_indices_0)[name = tensor("op_8447_cast_fp16")]; + tensor concat_137 = const()[name = tensor("concat_137"), val = tensor([1, 0, 0, 0, 0])]; + tensor concat_138 = const()[name = tensor("concat_138"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_39_internal_tensor_assign_2_stride_0 = const()[name = tensor("new_cache_39_internal_tensor_assign_2_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_39_internal_tensor_assign_2_begin_mask_0 = const()[name = tensor("new_cache_39_internal_tensor_assign_2_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_39_internal_tensor_assign_2_end_mask_0 = const()[name = tensor("new_cache_39_internal_tensor_assign_2_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_39_internal_tensor_assign_2_squeeze_mask_0 = const()[name = tensor("new_cache_39_internal_tensor_assign_2_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_87 = const()[name = tensor("shape_87"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_39 = const()[name = tensor("reduce_prod_39"), val = tensor(1048576)]; + tensor range_1d_39_start_0 = const()[name = tensor("range_1d_39_start_0"), val = tensor(0)]; + tensor range_1d_39_step_0 = const()[name = tensor("range_1d_39_step_0"), val = tensor(1)]; + tensor range_1d_39 = range_1d(end = reduce_prod_39, start = range_1d_39_start_0, step = range_1d_39_step_0)[name = tensor("range_1d_39")]; + tensor reshape_195 = reshape(shape = shape_87, x = range_1d_39)[name = tensor("reshape_195")]; + tensor slice_by_index_39 = slice_by_index(begin = concat_137, begin_mask = new_cache_39_internal_tensor_assign_2_begin_mask_0, end = concat_138, end_mask = new_cache_39_internal_tensor_assign_2_end_mask_0, squeeze_mask = new_cache_39_internal_tensor_assign_2_squeeze_mask_0, stride = new_cache_39_internal_tensor_assign_2_stride_0, x = reshape_195)[name = tensor("slice_by_index_39")]; + tensor reshape_196_shape_0 = const()[name = tensor("reshape_196_shape_0"), val = tensor([-1])]; + tensor reshape_196 = reshape(shape = reshape_196_shape_0, x = slice_by_index_39)[name = tensor("reshape_196")]; + tensor reshape_197_shape_0 = const()[name = tensor("reshape_197_shape_0"), val = tensor([-1])]; + tensor reshape_197_cast_fp16 = reshape(shape = reshape_197_shape_0, x = var_8447_cast_fp16)[name = tensor("reshape_197_cast_fp16")]; + tensor reshape_198_shape_0 = const()[name = tensor("reshape_198_shape_0"), val = tensor([-1])]; + tensor reshape_198_cast_fp16 = reshape(shape = reshape_198_shape_0, x = reshape_194_cast_fp16)[name = tensor("reshape_198_cast_fp16")]; + tensor scatter_39_mode_0 = const()[name = tensor("scatter_39_mode_0"), val = tensor("update")]; + tensor scatter_39_axis_0 = const()[name = tensor("scatter_39_axis_0"), val = tensor(0)]; + tensor scatter_39_validate_indices_0 = const()[name = tensor("scatter_39_validate_indices_0"), val = tensor(false)]; + tensor scatter_39_cast_fp16 = scatter(axis = scatter_39_axis_0, data = reshape_198_cast_fp16, indices = reshape_196, mode = scatter_39_mode_0, updates = reshape_197_cast_fp16, validate_indices = scatter_39_validate_indices_0)[name = tensor("scatter_39_cast_fp16")]; + tensor reshape_199_cast_fp16 = reshape(shape = shape_87, x = scatter_39_cast_fp16)[name = tensor("reshape_199_cast_fp16")]; + tensor reshape_199_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("reshape_199_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor keys_115_begin_0 = const()[name = tensor("keys_115_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor keys_115_end_0 = const()[name = tensor("keys_115_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor keys_115_end_mask_0 = const()[name = tensor("keys_115_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor keys_115_squeeze_mask_0 = const()[name = tensor("keys_115_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor keys_115_cast_fp16 = slice_by_index(begin = keys_115_begin_0, end = keys_115_end_0, end_mask = keys_115_end_mask_0, squeeze_mask = keys_115_squeeze_mask_0, x = reshape_199_cast_fp16)[name = tensor("keys_115_cast_fp16")]; + tensor values_115_begin_0 = const()[name = tensor("values_115_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor values_115_end_0 = const()[name = tensor("values_115_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor values_115_end_mask_0 = const()[name = tensor("values_115_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor values_115_squeeze_mask_0 = const()[name = tensor("values_115_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor values_115_cast_fp16 = slice_by_index(begin = values_115_begin_0, end = values_115_end_0, end_mask = values_115_end_mask_0, squeeze_mask = values_115_squeeze_mask_0, x = reshape_199_cast_fp16)[name = tensor("values_115_cast_fp16")]; + tensor var_8459_cast_fp16 = not_equal(x = keys_115_cast_fp16, y = keys_115_cast_fp16)[name = tensor("op_8459_cast_fp16")]; + tensor keys_117_cast_fp16 = select(a = var_504_to_fp16, b = keys_115_cast_fp16, cond = var_8459_cast_fp16)[name = tensor("keys_117_cast_fp16")]; + tensor var_8467_cast_fp16 = not_equal(x = values_115_cast_fp16, y = values_115_cast_fp16)[name = tensor("op_8467_cast_fp16")]; + tensor values_117_cast_fp16 = select(a = var_504_to_fp16, b = values_115_cast_fp16, cond = var_8467_cast_fp16)[name = tensor("values_117_cast_fp16")]; + tensor var_8491 = const()[name = tensor("op_8491"), val = tensor([0, 2, 1, 3])]; + tensor var_8504 = const()[name = tensor("op_8504"), val = tensor([1, 1, 1])]; + tensor var_8505_cast_fp16 = reshape(shape = var_8504, x = position19_to_fp16)[name = tensor("op_8505_cast_fp16")]; + tensor var_8522_to_fp16 = const()[name = tensor("op_8522_to_fp16"), val = tensor(0x1p+0)]; + tensor valid_len_39_cast_fp16 = add(x = var_8505_cast_fp16, y = var_8522_to_fp16)[name = tensor("valid_len_39_cast_fp16")]; + tensor valid_mask_39_cast_fp16 = less(x = k_positions_1_promoted_to_fp16, y = valid_len_39_cast_fp16)[name = tensor("valid_mask_39_cast_fp16")]; + tensor causal_mask_39_cast_fp16 = less_equal(x = k_positions_1_promoted_to_fp16, y = var_8505_cast_fp16)[name = tensor("causal_mask_39_cast_fp16")]; + tensor attn_mask_77 = logical_and(x = valid_mask_39_cast_fp16, y = causal_mask_39_cast_fp16)[name = tensor("attn_mask_77")]; + tensor attn_mask_79_axes_0 = const()[name = tensor("attn_mask_79_axes_0"), val = tensor([1])]; + tensor attn_mask_79 = expand_dims(axes = attn_mask_79_axes_0, x = attn_mask_77)[name = tensor("attn_mask_79")]; + tensor var_8540_transpose_x_0 = const()[name = tensor("op_8540_transpose_x_0"), val = tensor(false)]; + tensor var_8540_transpose_y_0 = const()[name = tensor("op_8540_transpose_y_0"), val = tensor(false)]; + tensor transpose_110_perm_0 = const()[name = tensor("transpose_110_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_111_perm_0 = const()[name = tensor("transpose_111_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_111 = transpose(perm = transpose_111_perm_0, x = keys_117_cast_fp16)[name = tensor("transpose_137")]; + tensor transpose_110 = transpose(perm = transpose_110_perm_0, x = q_117_cast_fp16)[name = tensor("transpose_138")]; + tensor var_8540_cast_fp16 = matmul(transpose_x = var_8540_transpose_x_0, transpose_y = var_8540_transpose_y_0, x = transpose_110, y = transpose_111)[name = tensor("op_8540_cast_fp16")]; + tensor var_8534_to_fp16 = const()[name = tensor("op_8534_to_fp16"), val = tensor([0x1p-3])]; + tensor attn_weights_115_cast_fp16 = mul(x = var_8540_cast_fp16, y = var_8534_to_fp16)[name = tensor("attn_weights_115_cast_fp16")]; + tensor var_8542 = logical_not(x = attn_mask_79)[name = tensor("op_8542")]; + tensor var_8543_to_fp16 = const()[name = tensor("op_8543_to_fp16"), val = tensor(-inf)]; + tensor attn_weights_117_cast_fp16 = select(a = var_8543_to_fp16, b = attn_weights_115_cast_fp16, cond = var_8542)[name = tensor("attn_weights_117_cast_fp16")]; + tensor var_8545 = const()[name = tensor("op_8545"), val = tensor(-1)]; + tensor attn_weights_119_cast_fp16 = softmax(axis = var_8545, x = attn_weights_117_cast_fp16)[name = tensor("attn_weights_119_cast_fp16")]; + tensor attn_output_39_transpose_x_0 = const()[name = tensor("attn_output_39_transpose_x_0"), val = tensor(false)]; + tensor attn_output_39_transpose_y_0 = const()[name = tensor("attn_output_39_transpose_y_0"), val = tensor(false)]; + tensor values_119_cast_fp16 = transpose(perm = var_8491, x = values_117_cast_fp16)[name = tensor("transpose_139")]; + tensor attn_output_39_cast_fp16 = matmul(transpose_x = attn_output_39_transpose_x_0, transpose_y = attn_output_39_transpose_y_0, x = attn_weights_119_cast_fp16, y = values_119_cast_fp16)[name = tensor("attn_output_39_cast_fp16")]; + tensor var_8553 = const()[name = tensor("op_8553"), val = tensor([0, 2, 1, 3])]; + tensor var_8556 = const()[name = tensor("op_8556"), val = tensor([1, 1, 1024])]; + tensor var_8554_cast_fp16 = transpose(perm = var_8553, x = attn_output_39_cast_fp16)[name = tensor("transpose_136")]; + tensor input_195_cast_fp16 = reshape(shape = var_8556, x = var_8554_cast_fp16)[name = tensor("input_195_cast_fp16")]; + tensor linear_78_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = attn19_out_proj_weight_cast_fp16, x = input_195_cast_fp16)[name = tensor("linear_78_cast_fp16")]; + tensor var_8562_to_fp16 = const()[name = tensor("op_8562_to_fp16"), val = tensor(0x1p+0)]; + tensor var_8563_cast_fp16 = add(x = position19_to_fp16, y = var_8562_to_fp16)[name = tensor("op_8563_cast_fp16")]; + tensor var_8563_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_8563_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor input_197_cast_fp16 = add(x = input_193_cast_fp16, y = linear_78_cast_fp16)[name = tensor("input_197_cast_fp16")]; + tensor input_199_axes_0 = const()[name = tensor("input_199_axes_0"), val = tensor([-1])]; + tensor norm19_2_weight_to_fp16 = const()[name = tensor("norm19_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(303752192)))]; + tensor norm19_2_bias_to_fp16 = const()[name = tensor("norm19_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(303754304)))]; + tensor var_8567_to_fp16 = const()[name = tensor("op_8567_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_199_cast_fp16 = layer_norm(axes = input_199_axes_0, beta = norm19_2_bias_to_fp16, epsilon = var_8567_to_fp16, gamma = norm19_2_weight_to_fp16, x = input_197_cast_fp16)[name = tensor("input_199_cast_fp16")]; + tensor linear_79_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = linear19_1_weight_cast_fp16, x = input_199_cast_fp16)[name = tensor("linear_79_cast_fp16")]; + tensor input_201_mode_0 = const()[name = tensor("input_201_mode_0"), val = tensor("EXACT")]; + tensor input_201_cast_fp16 = gelu(mode = input_201_mode_0, x = linear_79_cast_fp16)[name = tensor("input_201_cast_fp16")]; + tensor linear_80_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = linear19_2_weight_cast_fp16, x = input_201_cast_fp16)[name = tensor("linear_80_cast_fp16")]; + tensor input_203_cast_fp16 = add(x = input_197_cast_fp16, y = linear_80_cast_fp16)[name = tensor("input_203_cast_fp16")]; + tensor x_41_axes_0 = const()[name = tensor("x_41_axes_0"), val = tensor([-1])]; + tensor norm20_1_weight_to_fp16 = const()[name = tensor("norm20_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(303756416)))]; + tensor norm20_1_bias_to_fp16 = const()[name = tensor("norm20_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(303758528)))]; + tensor var_8584_to_fp16 = const()[name = tensor("op_8584_to_fp16"), val = tensor(0x1.5p-17)]; + tensor x_41_cast_fp16 = layer_norm(axes = x_41_axes_0, beta = norm20_1_bias_to_fp16, epsilon = var_8584_to_fp16, gamma = norm20_1_weight_to_fp16, x = input_203_cast_fp16)[name = tensor("x_41_cast_fp16")]; + tensor linear_81_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = attn20_in_proj_weight_cast_fp16, x = x_41_cast_fp16)[name = tensor("linear_81_cast_fp16")]; + tensor var_8620 = const()[name = tensor("op_8620"), val = tensor([1, 1, 3, 16, 64])]; + tensor qkv_41_cast_fp16 = reshape(shape = var_8620, x = linear_81_cast_fp16)[name = tensor("qkv_41_cast_fp16")]; + tensor q_121_begin_0 = const()[name = tensor("q_121_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor q_121_end_0 = const()[name = tensor("q_121_end_0"), val = tensor([1, 1, 1, 16, 64])]; + tensor q_121_end_mask_0 = const()[name = tensor("q_121_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor q_121_squeeze_mask_0 = const()[name = tensor("q_121_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor q_121_cast_fp16 = slice_by_index(begin = q_121_begin_0, end = q_121_end_0, end_mask = q_121_end_mask_0, squeeze_mask = q_121_squeeze_mask_0, x = qkv_41_cast_fp16)[name = tensor("q_121_cast_fp16")]; + tensor k_81_begin_0 = const()[name = tensor("k_81_begin_0"), val = tensor([0, 0, 1, 0, 0])]; + tensor k_81_end_0 = const()[name = tensor("k_81_end_0"), val = tensor([1, 1, 2, 16, 64])]; + tensor k_81_end_mask_0 = const()[name = tensor("k_81_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor k_81_squeeze_mask_0 = const()[name = tensor("k_81_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor k_81_cast_fp16 = slice_by_index(begin = k_81_begin_0, end = k_81_end_0, end_mask = k_81_end_mask_0, squeeze_mask = k_81_squeeze_mask_0, x = qkv_41_cast_fp16)[name = tensor("k_81_cast_fp16")]; + tensor v_41_begin_0 = const()[name = tensor("v_41_begin_0"), val = tensor([0, 0, 2, 0, 0])]; + tensor v_41_end_0 = const()[name = tensor("v_41_end_0"), val = tensor([1, 1, 3, 16, 64])]; + tensor v_41_end_mask_0 = const()[name = tensor("v_41_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor v_41_squeeze_mask_0 = const()[name = tensor("v_41_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor v_41_cast_fp16 = slice_by_index(begin = v_41_begin_0, end = v_41_end_0, end_mask = v_41_end_mask_0, squeeze_mask = v_41_squeeze_mask_0, x = qkv_41_cast_fp16)[name = tensor("v_41_cast_fp16")]; + tensor var_8724 = const()[name = tensor("op_8724"), val = tensor([1, 1, 1, 1])]; + tensor position20_to_fp16_dtype_0 = const()[name = tensor("position20_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor position20_to_fp16 = cast(dtype = position20_to_fp16_dtype_0, x = position20)[name = tensor("cast_453")]; + tensor ts_125_cast_fp16 = reshape(shape = var_8724, x = position20_to_fp16)[name = tensor("ts_125_cast_fp16")]; + tensor var_8728 = const()[name = tensor("op_8728"), val = tensor([1, 1, 16, 32, 2])]; + tensor q_complex_41_cast_fp16 = reshape(shape = var_8728, x = q_121_cast_fp16)[name = tensor("q_complex_41_cast_fp16")]; + tensor var_8732 = const()[name = tensor("op_8732"), val = tensor([1, 1, 16, 32, 2])]; + tensor k_complex_41_cast_fp16 = reshape(shape = var_8732, x = k_81_cast_fp16)[name = tensor("k_complex_41_cast_fp16")]; + tensor var_8736_begin_0 = const()[name = tensor("op_8736_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_8736_end_0 = const()[name = tensor("op_8736_end_0"), val = tensor([1, 1, 16, 32, 1])]; + tensor var_8736_end_mask_0 = const()[name = tensor("op_8736_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_8736_squeeze_mask_0 = const()[name = tensor("op_8736_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_8736_cast_fp16 = slice_by_index(begin = var_8736_begin_0, end = var_8736_end_0, end_mask = var_8736_end_mask_0, squeeze_mask = var_8736_squeeze_mask_0, x = q_complex_41_cast_fp16)[name = tensor("op_8736_cast_fp16")]; + tensor var_8744_begin_0 = const()[name = tensor("op_8744_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_8744_end_0 = const()[name = tensor("op_8744_end_0"), val = tensor([1, 1, 16, 32, 2])]; + tensor var_8744_end_mask_0 = const()[name = tensor("op_8744_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_8744_squeeze_mask_0 = const()[name = tensor("op_8744_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_8744_cast_fp16 = slice_by_index(begin = var_8744_begin_0, end = var_8744_end_0, end_mask = var_8744_end_mask_0, squeeze_mask = var_8744_squeeze_mask_0, x = q_complex_41_cast_fp16)[name = tensor("op_8744_cast_fp16")]; + tensor var_8752_begin_0 = const()[name = tensor("op_8752_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_8752_end_0 = const()[name = tensor("op_8752_end_0"), val = tensor([1, 1, 16, 32, 1])]; + tensor var_8752_end_mask_0 = const()[name = tensor("op_8752_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_8752_squeeze_mask_0 = const()[name = tensor("op_8752_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_8752_cast_fp16 = slice_by_index(begin = var_8752_begin_0, end = var_8752_end_0, end_mask = var_8752_end_mask_0, squeeze_mask = var_8752_squeeze_mask_0, x = k_complex_41_cast_fp16)[name = tensor("op_8752_cast_fp16")]; + tensor var_8760_begin_0 = const()[name = tensor("op_8760_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_8760_end_0 = const()[name = tensor("op_8760_end_0"), val = tensor([1, 1, 16, 32, 2])]; + tensor var_8760_end_mask_0 = const()[name = tensor("op_8760_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_8760_squeeze_mask_0 = const()[name = tensor("op_8760_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_8760_cast_fp16 = slice_by_index(begin = var_8760_begin_0, end = var_8760_end_0, end_mask = var_8760_end_mask_0, squeeze_mask = var_8760_squeeze_mask_0, x = k_complex_41_cast_fp16)[name = tensor("op_8760_cast_fp16")]; + tensor freqs_41_to_fp16 = const()[name = tensor("freqs_41_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(303760640)))]; + tensor var_8766_cast_fp16 = mul(x = freqs_41_to_fp16, y = ts_125_cast_fp16)[name = tensor("op_8766_cast_fp16")]; + tensor rotr_41_cast_fp16 = cos(x = var_8766_cast_fp16)[name = tensor("rotr_41_cast_fp16")]; + tensor roti_41_cast_fp16 = sin(x = var_8766_cast_fp16)[name = tensor("roti_41_cast_fp16")]; + tensor var_8770_cast_fp16 = mul(x = var_8736_cast_fp16, y = rotr_41_cast_fp16)[name = tensor("op_8770_cast_fp16")]; + tensor var_8771_cast_fp16 = mul(x = var_8744_cast_fp16, y = roti_41_cast_fp16)[name = tensor("op_8771_cast_fp16")]; + tensor qor_81_cast_fp16 = sub(x = var_8770_cast_fp16, y = var_8771_cast_fp16)[name = tensor("qor_81_cast_fp16")]; + tensor var_8774_cast_fp16 = mul(x = var_8736_cast_fp16, y = roti_41_cast_fp16)[name = tensor("op_8774_cast_fp16")]; + tensor var_8775_cast_fp16 = mul(x = var_8744_cast_fp16, y = rotr_41_cast_fp16)[name = tensor("op_8775_cast_fp16")]; + tensor qoi_81_cast_fp16 = add(x = var_8774_cast_fp16, y = var_8775_cast_fp16)[name = tensor("qoi_81_cast_fp16")]; + tensor var_8778_cast_fp16 = mul(x = var_8752_cast_fp16, y = rotr_41_cast_fp16)[name = tensor("op_8778_cast_fp16")]; + tensor var_8779_cast_fp16 = mul(x = var_8760_cast_fp16, y = roti_41_cast_fp16)[name = tensor("op_8779_cast_fp16")]; + tensor kor_81_cast_fp16 = sub(x = var_8778_cast_fp16, y = var_8779_cast_fp16)[name = tensor("kor_81_cast_fp16")]; + tensor var_8782_cast_fp16 = mul(x = var_8752_cast_fp16, y = roti_41_cast_fp16)[name = tensor("op_8782_cast_fp16")]; + tensor var_8783_cast_fp16 = mul(x = var_8760_cast_fp16, y = rotr_41_cast_fp16)[name = tensor("op_8783_cast_fp16")]; + tensor koi_81_cast_fp16 = add(x = var_8782_cast_fp16, y = var_8783_cast_fp16)[name = tensor("koi_81_cast_fp16")]; + tensor qo_41_axis_0 = const()[name = tensor("qo_41_axis_0"), val = tensor(-1)]; + tensor qo_41_cast_fp16 = stack(axis = qo_41_axis_0, values = (qor_81_cast_fp16, qoi_81_cast_fp16))[name = tensor("qo_41_cast_fp16")]; + tensor ko_41_axis_0 = const()[name = tensor("ko_41_axis_0"), val = tensor(-1)]; + tensor ko_41_cast_fp16 = stack(axis = ko_41_axis_0, values = (kor_81_cast_fp16, koi_81_cast_fp16))[name = tensor("ko_41_cast_fp16")]; + tensor var_8812 = const()[name = tensor("op_8812"), val = tensor([1, 1, 16, 64])]; + tensor q_123_cast_fp16 = reshape(shape = var_8812, x = qo_41_cast_fp16)[name = tensor("q_123_cast_fp16")]; + tensor var_8814 = const()[name = tensor("op_8814"), val = tensor([1, 1, 16, 64])]; + tensor k_83_cast_fp16 = reshape(shape = var_8814, x = ko_41_cast_fp16)[name = tensor("k_83_cast_fp16")]; + tensor _inversed_8836_y_0_to_fp16 = const()[name = tensor("_inversed_8836_y_0_to_fp16"), val = tensor(0x1p-9)]; + tensor _inversed_8836_cast_fp16 = mul(x = ts_125_cast_fp16, y = _inversed_8836_y_0_to_fp16)[name = tensor("_inversed_8836_cast_fp16")]; + tensor var_8837_cast_fp16 = floor(x = _inversed_8836_cast_fp16)[name = tensor("op_8837_cast_fp16")]; + tensor var_8838_to_fp16 = const()[name = tensor("op_8838_to_fp16"), val = tensor(0x1p+9)]; + tensor var_8839_cast_fp16 = mul(x = var_8837_cast_fp16, y = var_8838_to_fp16)[name = tensor("op_8839_cast_fp16")]; + tensor write_indices_float_83_cast_fp16 = sub(x = ts_125_cast_fp16, y = var_8839_cast_fp16)[name = tensor("write_indices_float_83_cast_fp16")]; + tensor var_8846_dtype_0 = const()[name = tensor("op_8846_dtype_0"), val = tensor("int32")]; + tensor write_indices_41_reps_0 = const()[name = tensor("write_indices_41_reps_0"), val = tensor([1, 1, 16, 64])]; + tensor write_indices_float_83_cast_fp16_to_int32 = cast(dtype = var_8846_dtype_0, x = write_indices_float_83_cast_fp16)[name = tensor("cast_452")]; + tensor write_indices_41 = tile(reps = write_indices_41_reps_0, x = write_indices_float_83_cast_fp16_to_int32)[name = tensor("write_indices_41")]; + tensor var_8854_begin_0 = const()[name = tensor("op_8854_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_8854_end_0 = const()[name = tensor("op_8854_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor var_8854_end_mask_0 = const()[name = tensor("op_8854_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_8854_squeeze_mask_0 = const()[name = tensor("op_8854_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor cache20_to_fp16_dtype_0 = const()[name = tensor("cache20_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor cache20_to_fp16 = cast(dtype = cache20_to_fp16_dtype_0, x = cache20)[name = tensor("cast_451")]; + tensor var_8854_cast_fp16 = slice_by_index(begin = var_8854_begin_0, end = var_8854_end_0, end_mask = var_8854_end_mask_0, squeeze_mask = var_8854_squeeze_mask_0, x = cache20_to_fp16)[name = tensor("op_8854_cast_fp16")]; + tensor var_8856_axis_0 = const()[name = tensor("op_8856_axis_0"), val = tensor(1)]; + tensor var_8856_mode_0 = const()[name = tensor("op_8856_mode_0"), val = tensor("update")]; + tensor var_8856_validate_indices_0 = const()[name = tensor("op_8856_validate_indices_0"), val = tensor(false)]; + tensor var_8856_cast_fp16 = scatter_along_axis(axis = var_8856_axis_0, data = var_8854_cast_fp16, indices = write_indices_41, mode = var_8856_mode_0, updates = k_83_cast_fp16, validate_indices = var_8856_validate_indices_0)[name = tensor("op_8856_cast_fp16")]; + tensor concat_142 = const()[name = tensor("concat_142"), val = tensor([0, 0, 0, 0, 0])]; + tensor concat_143 = const()[name = tensor("concat_143"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_41_internal_tensor_assign_1_stride_0 = const()[name = tensor("new_cache_41_internal_tensor_assign_1_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_41_internal_tensor_assign_1_begin_mask_0 = const()[name = tensor("new_cache_41_internal_tensor_assign_1_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_41_internal_tensor_assign_1_end_mask_0 = const()[name = tensor("new_cache_41_internal_tensor_assign_1_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_41_internal_tensor_assign_1_squeeze_mask_0 = const()[name = tensor("new_cache_41_internal_tensor_assign_1_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_88 = const()[name = tensor("shape_88"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_40 = const()[name = tensor("reduce_prod_40"), val = tensor(1048576)]; + tensor range_1d_40_start_0 = const()[name = tensor("range_1d_40_start_0"), val = tensor(0)]; + tensor range_1d_40_step_0 = const()[name = tensor("range_1d_40_step_0"), val = tensor(1)]; + tensor range_1d_40 = range_1d(end = reduce_prod_40, start = range_1d_40_start_0, step = range_1d_40_step_0)[name = tensor("range_1d_40")]; + tensor reshape_200 = reshape(shape = shape_88, x = range_1d_40)[name = tensor("reshape_200")]; + tensor slice_by_index_40 = slice_by_index(begin = concat_142, begin_mask = new_cache_41_internal_tensor_assign_1_begin_mask_0, end = concat_143, end_mask = new_cache_41_internal_tensor_assign_1_end_mask_0, squeeze_mask = new_cache_41_internal_tensor_assign_1_squeeze_mask_0, stride = new_cache_41_internal_tensor_assign_1_stride_0, x = reshape_200)[name = tensor("slice_by_index_40")]; + tensor reshape_201_shape_0 = const()[name = tensor("reshape_201_shape_0"), val = tensor([-1])]; + tensor reshape_201 = reshape(shape = reshape_201_shape_0, x = slice_by_index_40)[name = tensor("reshape_201")]; + tensor reshape_202_shape_0 = const()[name = tensor("reshape_202_shape_0"), val = tensor([-1])]; + tensor reshape_202_cast_fp16 = reshape(shape = reshape_202_shape_0, x = var_8856_cast_fp16)[name = tensor("reshape_202_cast_fp16")]; + tensor reshape_203_shape_0 = const()[name = tensor("reshape_203_shape_0"), val = tensor([-1])]; + tensor reshape_203_cast_fp16 = reshape(shape = reshape_203_shape_0, x = cache20_to_fp16)[name = tensor("reshape_203_cast_fp16")]; + tensor scatter_40_mode_0 = const()[name = tensor("scatter_40_mode_0"), val = tensor("update")]; + tensor scatter_40_axis_0 = const()[name = tensor("scatter_40_axis_0"), val = tensor(0)]; + tensor scatter_40_validate_indices_0 = const()[name = tensor("scatter_40_validate_indices_0"), val = tensor(false)]; + tensor scatter_40_cast_fp16 = scatter(axis = scatter_40_axis_0, data = reshape_203_cast_fp16, indices = reshape_201, mode = scatter_40_mode_0, updates = reshape_202_cast_fp16, validate_indices = scatter_40_validate_indices_0)[name = tensor("scatter_40_cast_fp16")]; + tensor reshape_204_cast_fp16 = reshape(shape = shape_88, x = scatter_40_cast_fp16)[name = tensor("reshape_204_cast_fp16")]; + tensor var_8864_begin_0 = const()[name = tensor("op_8864_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor var_8864_end_0 = const()[name = tensor("op_8864_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor var_8864_end_mask_0 = const()[name = tensor("op_8864_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_8864_squeeze_mask_0 = const()[name = tensor("op_8864_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor var_8864_cast_fp16 = slice_by_index(begin = var_8864_begin_0, end = var_8864_end_0, end_mask = var_8864_end_mask_0, squeeze_mask = var_8864_squeeze_mask_0, x = reshape_204_cast_fp16)[name = tensor("op_8864_cast_fp16")]; + tensor var_8866_axis_0 = const()[name = tensor("op_8866_axis_0"), val = tensor(1)]; + tensor var_8866_mode_0 = const()[name = tensor("op_8866_mode_0"), val = tensor("update")]; + tensor var_8866_validate_indices_0 = const()[name = tensor("op_8866_validate_indices_0"), val = tensor(false)]; + tensor var_8866_cast_fp16 = scatter_along_axis(axis = var_8866_axis_0, data = var_8864_cast_fp16, indices = write_indices_41, mode = var_8866_mode_0, updates = v_41_cast_fp16, validate_indices = var_8866_validate_indices_0)[name = tensor("op_8866_cast_fp16")]; + tensor concat_144 = const()[name = tensor("concat_144"), val = tensor([1, 0, 0, 0, 0])]; + tensor concat_145 = const()[name = tensor("concat_145"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_41_internal_tensor_assign_2_stride_0 = const()[name = tensor("new_cache_41_internal_tensor_assign_2_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_41_internal_tensor_assign_2_begin_mask_0 = const()[name = tensor("new_cache_41_internal_tensor_assign_2_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_41_internal_tensor_assign_2_end_mask_0 = const()[name = tensor("new_cache_41_internal_tensor_assign_2_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_41_internal_tensor_assign_2_squeeze_mask_0 = const()[name = tensor("new_cache_41_internal_tensor_assign_2_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_89 = const()[name = tensor("shape_89"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_41 = const()[name = tensor("reduce_prod_41"), val = tensor(1048576)]; + tensor range_1d_41_start_0 = const()[name = tensor("range_1d_41_start_0"), val = tensor(0)]; + tensor range_1d_41_step_0 = const()[name = tensor("range_1d_41_step_0"), val = tensor(1)]; + tensor range_1d_41 = range_1d(end = reduce_prod_41, start = range_1d_41_start_0, step = range_1d_41_step_0)[name = tensor("range_1d_41")]; + tensor reshape_205 = reshape(shape = shape_89, x = range_1d_41)[name = tensor("reshape_205")]; + tensor slice_by_index_41 = slice_by_index(begin = concat_144, begin_mask = new_cache_41_internal_tensor_assign_2_begin_mask_0, end = concat_145, end_mask = new_cache_41_internal_tensor_assign_2_end_mask_0, squeeze_mask = new_cache_41_internal_tensor_assign_2_squeeze_mask_0, stride = new_cache_41_internal_tensor_assign_2_stride_0, x = reshape_205)[name = tensor("slice_by_index_41")]; + tensor reshape_206_shape_0 = const()[name = tensor("reshape_206_shape_0"), val = tensor([-1])]; + tensor reshape_206 = reshape(shape = reshape_206_shape_0, x = slice_by_index_41)[name = tensor("reshape_206")]; + tensor reshape_207_shape_0 = const()[name = tensor("reshape_207_shape_0"), val = tensor([-1])]; + tensor reshape_207_cast_fp16 = reshape(shape = reshape_207_shape_0, x = var_8866_cast_fp16)[name = tensor("reshape_207_cast_fp16")]; + tensor reshape_208_shape_0 = const()[name = tensor("reshape_208_shape_0"), val = tensor([-1])]; + tensor reshape_208_cast_fp16 = reshape(shape = reshape_208_shape_0, x = reshape_204_cast_fp16)[name = tensor("reshape_208_cast_fp16")]; + tensor scatter_41_mode_0 = const()[name = tensor("scatter_41_mode_0"), val = tensor("update")]; + tensor scatter_41_axis_0 = const()[name = tensor("scatter_41_axis_0"), val = tensor(0)]; + tensor scatter_41_validate_indices_0 = const()[name = tensor("scatter_41_validate_indices_0"), val = tensor(false)]; + tensor scatter_41_cast_fp16 = scatter(axis = scatter_41_axis_0, data = reshape_208_cast_fp16, indices = reshape_206, mode = scatter_41_mode_0, updates = reshape_207_cast_fp16, validate_indices = scatter_41_validate_indices_0)[name = tensor("scatter_41_cast_fp16")]; + tensor reshape_209_cast_fp16 = reshape(shape = shape_89, x = scatter_41_cast_fp16)[name = tensor("reshape_209_cast_fp16")]; + tensor reshape_209_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("reshape_209_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor keys_121_begin_0 = const()[name = tensor("keys_121_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor keys_121_end_0 = const()[name = tensor("keys_121_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor keys_121_end_mask_0 = const()[name = tensor("keys_121_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor keys_121_squeeze_mask_0 = const()[name = tensor("keys_121_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor keys_121_cast_fp16 = slice_by_index(begin = keys_121_begin_0, end = keys_121_end_0, end_mask = keys_121_end_mask_0, squeeze_mask = keys_121_squeeze_mask_0, x = reshape_209_cast_fp16)[name = tensor("keys_121_cast_fp16")]; + tensor values_121_begin_0 = const()[name = tensor("values_121_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor values_121_end_0 = const()[name = tensor("values_121_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor values_121_end_mask_0 = const()[name = tensor("values_121_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor values_121_squeeze_mask_0 = const()[name = tensor("values_121_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor values_121_cast_fp16 = slice_by_index(begin = values_121_begin_0, end = values_121_end_0, end_mask = values_121_end_mask_0, squeeze_mask = values_121_squeeze_mask_0, x = reshape_209_cast_fp16)[name = tensor("values_121_cast_fp16")]; + tensor var_8878_cast_fp16 = not_equal(x = keys_121_cast_fp16, y = keys_121_cast_fp16)[name = tensor("op_8878_cast_fp16")]; + tensor keys_123_cast_fp16 = select(a = var_504_to_fp16, b = keys_121_cast_fp16, cond = var_8878_cast_fp16)[name = tensor("keys_123_cast_fp16")]; + tensor var_8886_cast_fp16 = not_equal(x = values_121_cast_fp16, y = values_121_cast_fp16)[name = tensor("op_8886_cast_fp16")]; + tensor values_123_cast_fp16 = select(a = var_504_to_fp16, b = values_121_cast_fp16, cond = var_8886_cast_fp16)[name = tensor("values_123_cast_fp16")]; + tensor var_8910 = const()[name = tensor("op_8910"), val = tensor([0, 2, 1, 3])]; + tensor var_8923 = const()[name = tensor("op_8923"), val = tensor([1, 1, 1])]; + tensor var_8924_cast_fp16 = reshape(shape = var_8923, x = position20_to_fp16)[name = tensor("op_8924_cast_fp16")]; + tensor var_8941_to_fp16 = const()[name = tensor("op_8941_to_fp16"), val = tensor(0x1p+0)]; + tensor valid_len_41_cast_fp16 = add(x = var_8924_cast_fp16, y = var_8941_to_fp16)[name = tensor("valid_len_41_cast_fp16")]; + tensor valid_mask_41_cast_fp16 = less(x = k_positions_1_promoted_to_fp16, y = valid_len_41_cast_fp16)[name = tensor("valid_mask_41_cast_fp16")]; + tensor causal_mask_41_cast_fp16 = less_equal(x = k_positions_1_promoted_to_fp16, y = var_8924_cast_fp16)[name = tensor("causal_mask_41_cast_fp16")]; + tensor attn_mask_81 = logical_and(x = valid_mask_41_cast_fp16, y = causal_mask_41_cast_fp16)[name = tensor("attn_mask_81")]; + tensor attn_mask_83_axes_0 = const()[name = tensor("attn_mask_83_axes_0"), val = tensor([1])]; + tensor attn_mask_83 = expand_dims(axes = attn_mask_83_axes_0, x = attn_mask_81)[name = tensor("attn_mask_83")]; + tensor var_8959_transpose_x_0 = const()[name = tensor("op_8959_transpose_x_0"), val = tensor(false)]; + tensor var_8959_transpose_y_0 = const()[name = tensor("op_8959_transpose_y_0"), val = tensor(false)]; + tensor transpose_112_perm_0 = const()[name = tensor("transpose_112_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_113_perm_0 = const()[name = tensor("transpose_113_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_113 = transpose(perm = transpose_113_perm_0, x = keys_123_cast_fp16)[name = tensor("transpose_133")]; + tensor transpose_112 = transpose(perm = transpose_112_perm_0, x = q_123_cast_fp16)[name = tensor("transpose_134")]; + tensor var_8959_cast_fp16 = matmul(transpose_x = var_8959_transpose_x_0, transpose_y = var_8959_transpose_y_0, x = transpose_112, y = transpose_113)[name = tensor("op_8959_cast_fp16")]; + tensor var_8953_to_fp16 = const()[name = tensor("op_8953_to_fp16"), val = tensor([0x1p-3])]; + tensor attn_weights_121_cast_fp16 = mul(x = var_8959_cast_fp16, y = var_8953_to_fp16)[name = tensor("attn_weights_121_cast_fp16")]; + tensor var_8961 = logical_not(x = attn_mask_83)[name = tensor("op_8961")]; + tensor var_8962_to_fp16 = const()[name = tensor("op_8962_to_fp16"), val = tensor(-inf)]; + tensor attn_weights_123_cast_fp16 = select(a = var_8962_to_fp16, b = attn_weights_121_cast_fp16, cond = var_8961)[name = tensor("attn_weights_123_cast_fp16")]; + tensor var_8964 = const()[name = tensor("op_8964"), val = tensor(-1)]; + tensor attn_weights_125_cast_fp16 = softmax(axis = var_8964, x = attn_weights_123_cast_fp16)[name = tensor("attn_weights_125_cast_fp16")]; + tensor attn_output_41_transpose_x_0 = const()[name = tensor("attn_output_41_transpose_x_0"), val = tensor(false)]; + tensor attn_output_41_transpose_y_0 = const()[name = tensor("attn_output_41_transpose_y_0"), val = tensor(false)]; + tensor values_125_cast_fp16 = transpose(perm = var_8910, x = values_123_cast_fp16)[name = tensor("transpose_135")]; + tensor attn_output_41_cast_fp16 = matmul(transpose_x = attn_output_41_transpose_x_0, transpose_y = attn_output_41_transpose_y_0, x = attn_weights_125_cast_fp16, y = values_125_cast_fp16)[name = tensor("attn_output_41_cast_fp16")]; + tensor var_8972 = const()[name = tensor("op_8972"), val = tensor([0, 2, 1, 3])]; + tensor var_8975 = const()[name = tensor("op_8975"), val = tensor([1, 1, 1024])]; + tensor var_8973_cast_fp16 = transpose(perm = var_8972, x = attn_output_41_cast_fp16)[name = tensor("transpose_132")]; + tensor input_205_cast_fp16 = reshape(shape = var_8975, x = var_8973_cast_fp16)[name = tensor("input_205_cast_fp16")]; + tensor linear_82_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = attn20_out_proj_weight_cast_fp16, x = input_205_cast_fp16)[name = tensor("linear_82_cast_fp16")]; + tensor var_8981_to_fp16 = const()[name = tensor("op_8981_to_fp16"), val = tensor(0x1p+0)]; + tensor var_8982_cast_fp16 = add(x = position20_to_fp16, y = var_8981_to_fp16)[name = tensor("op_8982_cast_fp16")]; + tensor var_8982_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_8982_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor input_207_cast_fp16 = add(x = input_203_cast_fp16, y = linear_82_cast_fp16)[name = tensor("input_207_cast_fp16")]; + tensor input_209_axes_0 = const()[name = tensor("input_209_axes_0"), val = tensor([-1])]; + tensor norm20_2_weight_to_fp16 = const()[name = tensor("norm20_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(303760768)))]; + tensor norm20_2_bias_to_fp16 = const()[name = tensor("norm20_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(303762880)))]; + tensor var_8986_to_fp16 = const()[name = tensor("op_8986_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_209_cast_fp16 = layer_norm(axes = input_209_axes_0, beta = norm20_2_bias_to_fp16, epsilon = var_8986_to_fp16, gamma = norm20_2_weight_to_fp16, x = input_207_cast_fp16)[name = tensor("input_209_cast_fp16")]; + tensor linear_83_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = linear20_1_weight_cast_fp16, x = input_209_cast_fp16)[name = tensor("linear_83_cast_fp16")]; + tensor input_211_mode_0 = const()[name = tensor("input_211_mode_0"), val = tensor("EXACT")]; + tensor input_211_cast_fp16 = gelu(mode = input_211_mode_0, x = linear_83_cast_fp16)[name = tensor("input_211_cast_fp16")]; + tensor linear_84_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = linear20_2_weight_cast_fp16, x = input_211_cast_fp16)[name = tensor("linear_84_cast_fp16")]; + tensor input_213_cast_fp16 = add(x = input_207_cast_fp16, y = linear_84_cast_fp16)[name = tensor("input_213_cast_fp16")]; + tensor x_43_axes_0 = const()[name = tensor("x_43_axes_0"), val = tensor([-1])]; + tensor norm21_1_weight_to_fp16 = const()[name = tensor("norm21_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(303764992)))]; + tensor norm21_1_bias_to_fp16 = const()[name = tensor("norm21_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(303767104)))]; + tensor var_9003_to_fp16 = const()[name = tensor("op_9003_to_fp16"), val = tensor(0x1.5p-17)]; + tensor x_43_cast_fp16 = layer_norm(axes = x_43_axes_0, beta = norm21_1_bias_to_fp16, epsilon = var_9003_to_fp16, gamma = norm21_1_weight_to_fp16, x = input_213_cast_fp16)[name = tensor("x_43_cast_fp16")]; + tensor linear_85_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = attn21_in_proj_weight_cast_fp16, x = x_43_cast_fp16)[name = tensor("linear_85_cast_fp16")]; + tensor var_9039 = const()[name = tensor("op_9039"), val = tensor([1, 1, 3, 16, 64])]; + tensor qkv_43_cast_fp16 = reshape(shape = var_9039, x = linear_85_cast_fp16)[name = tensor("qkv_43_cast_fp16")]; + tensor q_127_begin_0 = const()[name = tensor("q_127_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor q_127_end_0 = const()[name = tensor("q_127_end_0"), val = tensor([1, 1, 1, 16, 64])]; + tensor q_127_end_mask_0 = const()[name = tensor("q_127_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor q_127_squeeze_mask_0 = const()[name = tensor("q_127_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor q_127_cast_fp16 = slice_by_index(begin = q_127_begin_0, end = q_127_end_0, end_mask = q_127_end_mask_0, squeeze_mask = q_127_squeeze_mask_0, x = qkv_43_cast_fp16)[name = tensor("q_127_cast_fp16")]; + tensor k_85_begin_0 = const()[name = tensor("k_85_begin_0"), val = tensor([0, 0, 1, 0, 0])]; + tensor k_85_end_0 = const()[name = tensor("k_85_end_0"), val = tensor([1, 1, 2, 16, 64])]; + tensor k_85_end_mask_0 = const()[name = tensor("k_85_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor k_85_squeeze_mask_0 = const()[name = tensor("k_85_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor k_85_cast_fp16 = slice_by_index(begin = k_85_begin_0, end = k_85_end_0, end_mask = k_85_end_mask_0, squeeze_mask = k_85_squeeze_mask_0, x = qkv_43_cast_fp16)[name = tensor("k_85_cast_fp16")]; + tensor v_43_begin_0 = const()[name = tensor("v_43_begin_0"), val = tensor([0, 0, 2, 0, 0])]; + tensor v_43_end_0 = const()[name = tensor("v_43_end_0"), val = tensor([1, 1, 3, 16, 64])]; + tensor v_43_end_mask_0 = const()[name = tensor("v_43_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor v_43_squeeze_mask_0 = const()[name = tensor("v_43_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor v_43_cast_fp16 = slice_by_index(begin = v_43_begin_0, end = v_43_end_0, end_mask = v_43_end_mask_0, squeeze_mask = v_43_squeeze_mask_0, x = qkv_43_cast_fp16)[name = tensor("v_43_cast_fp16")]; + tensor var_9143 = const()[name = tensor("op_9143"), val = tensor([1, 1, 1, 1])]; + tensor position21_to_fp16_dtype_0 = const()[name = tensor("position21_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor position21_to_fp16 = cast(dtype = position21_to_fp16_dtype_0, x = position21)[name = tensor("cast_448")]; + tensor ts_131_cast_fp16 = reshape(shape = var_9143, x = position21_to_fp16)[name = tensor("ts_131_cast_fp16")]; + tensor var_9147 = const()[name = tensor("op_9147"), val = tensor([1, 1, 16, 32, 2])]; + tensor q_complex_43_cast_fp16 = reshape(shape = var_9147, x = q_127_cast_fp16)[name = tensor("q_complex_43_cast_fp16")]; + tensor var_9151 = const()[name = tensor("op_9151"), val = tensor([1, 1, 16, 32, 2])]; + tensor k_complex_43_cast_fp16 = reshape(shape = var_9151, x = k_85_cast_fp16)[name = tensor("k_complex_43_cast_fp16")]; + tensor var_9155_begin_0 = const()[name = tensor("op_9155_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_9155_end_0 = const()[name = tensor("op_9155_end_0"), val = tensor([1, 1, 16, 32, 1])]; + tensor var_9155_end_mask_0 = const()[name = tensor("op_9155_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_9155_squeeze_mask_0 = const()[name = tensor("op_9155_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_9155_cast_fp16 = slice_by_index(begin = var_9155_begin_0, end = var_9155_end_0, end_mask = var_9155_end_mask_0, squeeze_mask = var_9155_squeeze_mask_0, x = q_complex_43_cast_fp16)[name = tensor("op_9155_cast_fp16")]; + tensor var_9163_begin_0 = const()[name = tensor("op_9163_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_9163_end_0 = const()[name = tensor("op_9163_end_0"), val = tensor([1, 1, 16, 32, 2])]; + tensor var_9163_end_mask_0 = const()[name = tensor("op_9163_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_9163_squeeze_mask_0 = const()[name = tensor("op_9163_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_9163_cast_fp16 = slice_by_index(begin = var_9163_begin_0, end = var_9163_end_0, end_mask = var_9163_end_mask_0, squeeze_mask = var_9163_squeeze_mask_0, x = q_complex_43_cast_fp16)[name = tensor("op_9163_cast_fp16")]; + tensor var_9171_begin_0 = const()[name = tensor("op_9171_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_9171_end_0 = const()[name = tensor("op_9171_end_0"), val = tensor([1, 1, 16, 32, 1])]; + tensor var_9171_end_mask_0 = const()[name = tensor("op_9171_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_9171_squeeze_mask_0 = const()[name = tensor("op_9171_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_9171_cast_fp16 = slice_by_index(begin = var_9171_begin_0, end = var_9171_end_0, end_mask = var_9171_end_mask_0, squeeze_mask = var_9171_squeeze_mask_0, x = k_complex_43_cast_fp16)[name = tensor("op_9171_cast_fp16")]; + tensor var_9179_begin_0 = const()[name = tensor("op_9179_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_9179_end_0 = const()[name = tensor("op_9179_end_0"), val = tensor([1, 1, 16, 32, 2])]; + tensor var_9179_end_mask_0 = const()[name = tensor("op_9179_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_9179_squeeze_mask_0 = const()[name = tensor("op_9179_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_9179_cast_fp16 = slice_by_index(begin = var_9179_begin_0, end = var_9179_end_0, end_mask = var_9179_end_mask_0, squeeze_mask = var_9179_squeeze_mask_0, x = k_complex_43_cast_fp16)[name = tensor("op_9179_cast_fp16")]; + tensor freqs_43_to_fp16 = const()[name = tensor("freqs_43_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(303769216)))]; + tensor var_9185_cast_fp16 = mul(x = freqs_43_to_fp16, y = ts_131_cast_fp16)[name = tensor("op_9185_cast_fp16")]; + tensor rotr_43_cast_fp16 = cos(x = var_9185_cast_fp16)[name = tensor("rotr_43_cast_fp16")]; + tensor roti_43_cast_fp16 = sin(x = var_9185_cast_fp16)[name = tensor("roti_43_cast_fp16")]; + tensor var_9189_cast_fp16 = mul(x = var_9155_cast_fp16, y = rotr_43_cast_fp16)[name = tensor("op_9189_cast_fp16")]; + tensor var_9190_cast_fp16 = mul(x = var_9163_cast_fp16, y = roti_43_cast_fp16)[name = tensor("op_9190_cast_fp16")]; + tensor qor_85_cast_fp16 = sub(x = var_9189_cast_fp16, y = var_9190_cast_fp16)[name = tensor("qor_85_cast_fp16")]; + tensor var_9193_cast_fp16 = mul(x = var_9155_cast_fp16, y = roti_43_cast_fp16)[name = tensor("op_9193_cast_fp16")]; + tensor var_9194_cast_fp16 = mul(x = var_9163_cast_fp16, y = rotr_43_cast_fp16)[name = tensor("op_9194_cast_fp16")]; + tensor qoi_85_cast_fp16 = add(x = var_9193_cast_fp16, y = var_9194_cast_fp16)[name = tensor("qoi_85_cast_fp16")]; + tensor var_9197_cast_fp16 = mul(x = var_9171_cast_fp16, y = rotr_43_cast_fp16)[name = tensor("op_9197_cast_fp16")]; + tensor var_9198_cast_fp16 = mul(x = var_9179_cast_fp16, y = roti_43_cast_fp16)[name = tensor("op_9198_cast_fp16")]; + tensor kor_85_cast_fp16 = sub(x = var_9197_cast_fp16, y = var_9198_cast_fp16)[name = tensor("kor_85_cast_fp16")]; + tensor var_9201_cast_fp16 = mul(x = var_9171_cast_fp16, y = roti_43_cast_fp16)[name = tensor("op_9201_cast_fp16")]; + tensor var_9202_cast_fp16 = mul(x = var_9179_cast_fp16, y = rotr_43_cast_fp16)[name = tensor("op_9202_cast_fp16")]; + tensor koi_85_cast_fp16 = add(x = var_9201_cast_fp16, y = var_9202_cast_fp16)[name = tensor("koi_85_cast_fp16")]; + tensor qo_43_axis_0 = const()[name = tensor("qo_43_axis_0"), val = tensor(-1)]; + tensor qo_43_cast_fp16 = stack(axis = qo_43_axis_0, values = (qor_85_cast_fp16, qoi_85_cast_fp16))[name = tensor("qo_43_cast_fp16")]; + tensor ko_43_axis_0 = const()[name = tensor("ko_43_axis_0"), val = tensor(-1)]; + tensor ko_43_cast_fp16 = stack(axis = ko_43_axis_0, values = (kor_85_cast_fp16, koi_85_cast_fp16))[name = tensor("ko_43_cast_fp16")]; + tensor var_9231 = const()[name = tensor("op_9231"), val = tensor([1, 1, 16, 64])]; + tensor q_129_cast_fp16 = reshape(shape = var_9231, x = qo_43_cast_fp16)[name = tensor("q_129_cast_fp16")]; + tensor var_9233 = const()[name = tensor("op_9233"), val = tensor([1, 1, 16, 64])]; + tensor k_87_cast_fp16 = reshape(shape = var_9233, x = ko_43_cast_fp16)[name = tensor("k_87_cast_fp16")]; + tensor _inversed_9255_y_0_to_fp16 = const()[name = tensor("_inversed_9255_y_0_to_fp16"), val = tensor(0x1p-9)]; + tensor _inversed_9255_cast_fp16 = mul(x = ts_131_cast_fp16, y = _inversed_9255_y_0_to_fp16)[name = tensor("_inversed_9255_cast_fp16")]; + tensor var_9256_cast_fp16 = floor(x = _inversed_9255_cast_fp16)[name = tensor("op_9256_cast_fp16")]; + tensor var_9257_to_fp16 = const()[name = tensor("op_9257_to_fp16"), val = tensor(0x1p+9)]; + tensor var_9258_cast_fp16 = mul(x = var_9256_cast_fp16, y = var_9257_to_fp16)[name = tensor("op_9258_cast_fp16")]; + tensor write_indices_float_87_cast_fp16 = sub(x = ts_131_cast_fp16, y = var_9258_cast_fp16)[name = tensor("write_indices_float_87_cast_fp16")]; + tensor var_9265_dtype_0 = const()[name = tensor("op_9265_dtype_0"), val = tensor("int32")]; + tensor write_indices_43_reps_0 = const()[name = tensor("write_indices_43_reps_0"), val = tensor([1, 1, 16, 64])]; + tensor write_indices_float_87_cast_fp16_to_int32 = cast(dtype = var_9265_dtype_0, x = write_indices_float_87_cast_fp16)[name = tensor("cast_447")]; + tensor write_indices_43 = tile(reps = write_indices_43_reps_0, x = write_indices_float_87_cast_fp16_to_int32)[name = tensor("write_indices_43")]; + tensor var_9273_begin_0 = const()[name = tensor("op_9273_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_9273_end_0 = const()[name = tensor("op_9273_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor var_9273_end_mask_0 = const()[name = tensor("op_9273_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_9273_squeeze_mask_0 = const()[name = tensor("op_9273_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor cache21_to_fp16_dtype_0 = const()[name = tensor("cache21_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor cache21_to_fp16 = cast(dtype = cache21_to_fp16_dtype_0, x = cache21)[name = tensor("cast_446")]; + tensor var_9273_cast_fp16 = slice_by_index(begin = var_9273_begin_0, end = var_9273_end_0, end_mask = var_9273_end_mask_0, squeeze_mask = var_9273_squeeze_mask_0, x = cache21_to_fp16)[name = tensor("op_9273_cast_fp16")]; + tensor var_9275_axis_0 = const()[name = tensor("op_9275_axis_0"), val = tensor(1)]; + tensor var_9275_mode_0 = const()[name = tensor("op_9275_mode_0"), val = tensor("update")]; + tensor var_9275_validate_indices_0 = const()[name = tensor("op_9275_validate_indices_0"), val = tensor(false)]; + tensor var_9275_cast_fp16 = scatter_along_axis(axis = var_9275_axis_0, data = var_9273_cast_fp16, indices = write_indices_43, mode = var_9275_mode_0, updates = k_87_cast_fp16, validate_indices = var_9275_validate_indices_0)[name = tensor("op_9275_cast_fp16")]; + tensor concat_149 = const()[name = tensor("concat_149"), val = tensor([0, 0, 0, 0, 0])]; + tensor concat_150 = const()[name = tensor("concat_150"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_43_internal_tensor_assign_1_stride_0 = const()[name = tensor("new_cache_43_internal_tensor_assign_1_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_43_internal_tensor_assign_1_begin_mask_0 = const()[name = tensor("new_cache_43_internal_tensor_assign_1_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_43_internal_tensor_assign_1_end_mask_0 = const()[name = tensor("new_cache_43_internal_tensor_assign_1_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_43_internal_tensor_assign_1_squeeze_mask_0 = const()[name = tensor("new_cache_43_internal_tensor_assign_1_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_90 = const()[name = tensor("shape_90"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_42 = const()[name = tensor("reduce_prod_42"), val = tensor(1048576)]; + tensor range_1d_42_start_0 = const()[name = tensor("range_1d_42_start_0"), val = tensor(0)]; + tensor range_1d_42_step_0 = const()[name = tensor("range_1d_42_step_0"), val = tensor(1)]; + tensor range_1d_42 = range_1d(end = reduce_prod_42, start = range_1d_42_start_0, step = range_1d_42_step_0)[name = tensor("range_1d_42")]; + tensor reshape_210 = reshape(shape = shape_90, x = range_1d_42)[name = tensor("reshape_210")]; + tensor slice_by_index_42 = slice_by_index(begin = concat_149, begin_mask = new_cache_43_internal_tensor_assign_1_begin_mask_0, end = concat_150, end_mask = new_cache_43_internal_tensor_assign_1_end_mask_0, squeeze_mask = new_cache_43_internal_tensor_assign_1_squeeze_mask_0, stride = new_cache_43_internal_tensor_assign_1_stride_0, x = reshape_210)[name = tensor("slice_by_index_42")]; + tensor reshape_211_shape_0 = const()[name = tensor("reshape_211_shape_0"), val = tensor([-1])]; + tensor reshape_211 = reshape(shape = reshape_211_shape_0, x = slice_by_index_42)[name = tensor("reshape_211")]; + tensor reshape_212_shape_0 = const()[name = tensor("reshape_212_shape_0"), val = tensor([-1])]; + tensor reshape_212_cast_fp16 = reshape(shape = reshape_212_shape_0, x = var_9275_cast_fp16)[name = tensor("reshape_212_cast_fp16")]; + tensor reshape_213_shape_0 = const()[name = tensor("reshape_213_shape_0"), val = tensor([-1])]; + tensor reshape_213_cast_fp16 = reshape(shape = reshape_213_shape_0, x = cache21_to_fp16)[name = tensor("reshape_213_cast_fp16")]; + tensor scatter_42_mode_0 = const()[name = tensor("scatter_42_mode_0"), val = tensor("update")]; + tensor scatter_42_axis_0 = const()[name = tensor("scatter_42_axis_0"), val = tensor(0)]; + tensor scatter_42_validate_indices_0 = const()[name = tensor("scatter_42_validate_indices_0"), val = tensor(false)]; + tensor scatter_42_cast_fp16 = scatter(axis = scatter_42_axis_0, data = reshape_213_cast_fp16, indices = reshape_211, mode = scatter_42_mode_0, updates = reshape_212_cast_fp16, validate_indices = scatter_42_validate_indices_0)[name = tensor("scatter_42_cast_fp16")]; + tensor reshape_214_cast_fp16 = reshape(shape = shape_90, x = scatter_42_cast_fp16)[name = tensor("reshape_214_cast_fp16")]; + tensor var_9283_begin_0 = const()[name = tensor("op_9283_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor var_9283_end_0 = const()[name = tensor("op_9283_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor var_9283_end_mask_0 = const()[name = tensor("op_9283_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_9283_squeeze_mask_0 = const()[name = tensor("op_9283_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor var_9283_cast_fp16 = slice_by_index(begin = var_9283_begin_0, end = var_9283_end_0, end_mask = var_9283_end_mask_0, squeeze_mask = var_9283_squeeze_mask_0, x = reshape_214_cast_fp16)[name = tensor("op_9283_cast_fp16")]; + tensor var_9285_axis_0 = const()[name = tensor("op_9285_axis_0"), val = tensor(1)]; + tensor var_9285_mode_0 = const()[name = tensor("op_9285_mode_0"), val = tensor("update")]; + tensor var_9285_validate_indices_0 = const()[name = tensor("op_9285_validate_indices_0"), val = tensor(false)]; + tensor var_9285_cast_fp16 = scatter_along_axis(axis = var_9285_axis_0, data = var_9283_cast_fp16, indices = write_indices_43, mode = var_9285_mode_0, updates = v_43_cast_fp16, validate_indices = var_9285_validate_indices_0)[name = tensor("op_9285_cast_fp16")]; + tensor concat_151 = const()[name = tensor("concat_151"), val = tensor([1, 0, 0, 0, 0])]; + tensor concat_152 = const()[name = tensor("concat_152"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_43_internal_tensor_assign_2_stride_0 = const()[name = tensor("new_cache_43_internal_tensor_assign_2_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_43_internal_tensor_assign_2_begin_mask_0 = const()[name = tensor("new_cache_43_internal_tensor_assign_2_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_43_internal_tensor_assign_2_end_mask_0 = const()[name = tensor("new_cache_43_internal_tensor_assign_2_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_43_internal_tensor_assign_2_squeeze_mask_0 = const()[name = tensor("new_cache_43_internal_tensor_assign_2_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_91 = const()[name = tensor("shape_91"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_43 = const()[name = tensor("reduce_prod_43"), val = tensor(1048576)]; + tensor range_1d_43_start_0 = const()[name = tensor("range_1d_43_start_0"), val = tensor(0)]; + tensor range_1d_43_step_0 = const()[name = tensor("range_1d_43_step_0"), val = tensor(1)]; + tensor range_1d_43 = range_1d(end = reduce_prod_43, start = range_1d_43_start_0, step = range_1d_43_step_0)[name = tensor("range_1d_43")]; + tensor reshape_215 = reshape(shape = shape_91, x = range_1d_43)[name = tensor("reshape_215")]; + tensor slice_by_index_43 = slice_by_index(begin = concat_151, begin_mask = new_cache_43_internal_tensor_assign_2_begin_mask_0, end = concat_152, end_mask = new_cache_43_internal_tensor_assign_2_end_mask_0, squeeze_mask = new_cache_43_internal_tensor_assign_2_squeeze_mask_0, stride = new_cache_43_internal_tensor_assign_2_stride_0, x = reshape_215)[name = tensor("slice_by_index_43")]; + tensor reshape_216_shape_0 = const()[name = tensor("reshape_216_shape_0"), val = tensor([-1])]; + tensor reshape_216 = reshape(shape = reshape_216_shape_0, x = slice_by_index_43)[name = tensor("reshape_216")]; + tensor reshape_217_shape_0 = const()[name = tensor("reshape_217_shape_0"), val = tensor([-1])]; + tensor reshape_217_cast_fp16 = reshape(shape = reshape_217_shape_0, x = var_9285_cast_fp16)[name = tensor("reshape_217_cast_fp16")]; + tensor reshape_218_shape_0 = const()[name = tensor("reshape_218_shape_0"), val = tensor([-1])]; + tensor reshape_218_cast_fp16 = reshape(shape = reshape_218_shape_0, x = reshape_214_cast_fp16)[name = tensor("reshape_218_cast_fp16")]; + tensor scatter_43_mode_0 = const()[name = tensor("scatter_43_mode_0"), val = tensor("update")]; + tensor scatter_43_axis_0 = const()[name = tensor("scatter_43_axis_0"), val = tensor(0)]; + tensor scatter_43_validate_indices_0 = const()[name = tensor("scatter_43_validate_indices_0"), val = tensor(false)]; + tensor scatter_43_cast_fp16 = scatter(axis = scatter_43_axis_0, data = reshape_218_cast_fp16, indices = reshape_216, mode = scatter_43_mode_0, updates = reshape_217_cast_fp16, validate_indices = scatter_43_validate_indices_0)[name = tensor("scatter_43_cast_fp16")]; + tensor reshape_219_cast_fp16 = reshape(shape = shape_91, x = scatter_43_cast_fp16)[name = tensor("reshape_219_cast_fp16")]; + tensor reshape_219_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("reshape_219_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor keys_127_begin_0 = const()[name = tensor("keys_127_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor keys_127_end_0 = const()[name = tensor("keys_127_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor keys_127_end_mask_0 = const()[name = tensor("keys_127_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor keys_127_squeeze_mask_0 = const()[name = tensor("keys_127_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor keys_127_cast_fp16 = slice_by_index(begin = keys_127_begin_0, end = keys_127_end_0, end_mask = keys_127_end_mask_0, squeeze_mask = keys_127_squeeze_mask_0, x = reshape_219_cast_fp16)[name = tensor("keys_127_cast_fp16")]; + tensor values_127_begin_0 = const()[name = tensor("values_127_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor values_127_end_0 = const()[name = tensor("values_127_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor values_127_end_mask_0 = const()[name = tensor("values_127_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor values_127_squeeze_mask_0 = const()[name = tensor("values_127_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor values_127_cast_fp16 = slice_by_index(begin = values_127_begin_0, end = values_127_end_0, end_mask = values_127_end_mask_0, squeeze_mask = values_127_squeeze_mask_0, x = reshape_219_cast_fp16)[name = tensor("values_127_cast_fp16")]; + tensor var_9297_cast_fp16 = not_equal(x = keys_127_cast_fp16, y = keys_127_cast_fp16)[name = tensor("op_9297_cast_fp16")]; + tensor keys_129_cast_fp16 = select(a = var_504_to_fp16, b = keys_127_cast_fp16, cond = var_9297_cast_fp16)[name = tensor("keys_129_cast_fp16")]; + tensor var_9305_cast_fp16 = not_equal(x = values_127_cast_fp16, y = values_127_cast_fp16)[name = tensor("op_9305_cast_fp16")]; + tensor values_129_cast_fp16 = select(a = var_504_to_fp16, b = values_127_cast_fp16, cond = var_9305_cast_fp16)[name = tensor("values_129_cast_fp16")]; + tensor var_9329 = const()[name = tensor("op_9329"), val = tensor([0, 2, 1, 3])]; + tensor var_9342 = const()[name = tensor("op_9342"), val = tensor([1, 1, 1])]; + tensor var_9343_cast_fp16 = reshape(shape = var_9342, x = position21_to_fp16)[name = tensor("op_9343_cast_fp16")]; + tensor var_9360_to_fp16 = const()[name = tensor("op_9360_to_fp16"), val = tensor(0x1p+0)]; + tensor valid_len_43_cast_fp16 = add(x = var_9343_cast_fp16, y = var_9360_to_fp16)[name = tensor("valid_len_43_cast_fp16")]; + tensor valid_mask_43_cast_fp16 = less(x = k_positions_1_promoted_to_fp16, y = valid_len_43_cast_fp16)[name = tensor("valid_mask_43_cast_fp16")]; + tensor causal_mask_43_cast_fp16 = less_equal(x = k_positions_1_promoted_to_fp16, y = var_9343_cast_fp16)[name = tensor("causal_mask_43_cast_fp16")]; + tensor attn_mask_85 = logical_and(x = valid_mask_43_cast_fp16, y = causal_mask_43_cast_fp16)[name = tensor("attn_mask_85")]; + tensor attn_mask_87_axes_0 = const()[name = tensor("attn_mask_87_axes_0"), val = tensor([1])]; + tensor attn_mask_87 = expand_dims(axes = attn_mask_87_axes_0, x = attn_mask_85)[name = tensor("attn_mask_87")]; + tensor var_9378_transpose_x_0 = const()[name = tensor("op_9378_transpose_x_0"), val = tensor(false)]; + tensor var_9378_transpose_y_0 = const()[name = tensor("op_9378_transpose_y_0"), val = tensor(false)]; + tensor transpose_114_perm_0 = const()[name = tensor("transpose_114_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_115_perm_0 = const()[name = tensor("transpose_115_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_115 = transpose(perm = transpose_115_perm_0, x = keys_129_cast_fp16)[name = tensor("transpose_129")]; + tensor transpose_114 = transpose(perm = transpose_114_perm_0, x = q_129_cast_fp16)[name = tensor("transpose_130")]; + tensor var_9378_cast_fp16 = matmul(transpose_x = var_9378_transpose_x_0, transpose_y = var_9378_transpose_y_0, x = transpose_114, y = transpose_115)[name = tensor("op_9378_cast_fp16")]; + tensor var_9372_to_fp16 = const()[name = tensor("op_9372_to_fp16"), val = tensor([0x1p-3])]; + tensor attn_weights_127_cast_fp16 = mul(x = var_9378_cast_fp16, y = var_9372_to_fp16)[name = tensor("attn_weights_127_cast_fp16")]; + tensor var_9380 = logical_not(x = attn_mask_87)[name = tensor("op_9380")]; + tensor var_9381_to_fp16 = const()[name = tensor("op_9381_to_fp16"), val = tensor(-inf)]; + tensor attn_weights_129_cast_fp16 = select(a = var_9381_to_fp16, b = attn_weights_127_cast_fp16, cond = var_9380)[name = tensor("attn_weights_129_cast_fp16")]; + tensor var_9383 = const()[name = tensor("op_9383"), val = tensor(-1)]; + tensor attn_weights_131_cast_fp16 = softmax(axis = var_9383, x = attn_weights_129_cast_fp16)[name = tensor("attn_weights_131_cast_fp16")]; + tensor attn_output_43_transpose_x_0 = const()[name = tensor("attn_output_43_transpose_x_0"), val = tensor(false)]; + tensor attn_output_43_transpose_y_0 = const()[name = tensor("attn_output_43_transpose_y_0"), val = tensor(false)]; + tensor values_131_cast_fp16 = transpose(perm = var_9329, x = values_129_cast_fp16)[name = tensor("transpose_131")]; + tensor attn_output_43_cast_fp16 = matmul(transpose_x = attn_output_43_transpose_x_0, transpose_y = attn_output_43_transpose_y_0, x = attn_weights_131_cast_fp16, y = values_131_cast_fp16)[name = tensor("attn_output_43_cast_fp16")]; + tensor var_9391 = const()[name = tensor("op_9391"), val = tensor([0, 2, 1, 3])]; + tensor var_9394 = const()[name = tensor("op_9394"), val = tensor([1, 1, 1024])]; + tensor var_9392_cast_fp16 = transpose(perm = var_9391, x = attn_output_43_cast_fp16)[name = tensor("transpose_128")]; + tensor input_215_cast_fp16 = reshape(shape = var_9394, x = var_9392_cast_fp16)[name = tensor("input_215_cast_fp16")]; + tensor linear_86_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = attn21_out_proj_weight_cast_fp16, x = input_215_cast_fp16)[name = tensor("linear_86_cast_fp16")]; + tensor var_9400_to_fp16 = const()[name = tensor("op_9400_to_fp16"), val = tensor(0x1p+0)]; + tensor var_9401_cast_fp16 = add(x = position21_to_fp16, y = var_9400_to_fp16)[name = tensor("op_9401_cast_fp16")]; + tensor var_9401_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_9401_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor input_217_cast_fp16 = add(x = input_213_cast_fp16, y = linear_86_cast_fp16)[name = tensor("input_217_cast_fp16")]; + tensor input_219_axes_0 = const()[name = tensor("input_219_axes_0"), val = tensor([-1])]; + tensor norm21_2_weight_to_fp16 = const()[name = tensor("norm21_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(303769344)))]; + tensor norm21_2_bias_to_fp16 = const()[name = tensor("norm21_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(303771456)))]; + tensor var_9405_to_fp16 = const()[name = tensor("op_9405_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_219_cast_fp16 = layer_norm(axes = input_219_axes_0, beta = norm21_2_bias_to_fp16, epsilon = var_9405_to_fp16, gamma = norm21_2_weight_to_fp16, x = input_217_cast_fp16)[name = tensor("input_219_cast_fp16")]; + tensor linear_87_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = linear21_1_weight_cast_fp16, x = input_219_cast_fp16)[name = tensor("linear_87_cast_fp16")]; + tensor input_221_mode_0 = const()[name = tensor("input_221_mode_0"), val = tensor("EXACT")]; + tensor input_221_cast_fp16 = gelu(mode = input_221_mode_0, x = linear_87_cast_fp16)[name = tensor("input_221_cast_fp16")]; + tensor linear_88_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = linear21_2_weight_cast_fp16, x = input_221_cast_fp16)[name = tensor("linear_88_cast_fp16")]; + tensor input_223_cast_fp16 = add(x = input_217_cast_fp16, y = linear_88_cast_fp16)[name = tensor("input_223_cast_fp16")]; + tensor x_45_axes_0 = const()[name = tensor("x_45_axes_0"), val = tensor([-1])]; + tensor norm22_1_weight_to_fp16 = const()[name = tensor("norm22_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(303773568)))]; + tensor norm22_1_bias_to_fp16 = const()[name = tensor("norm22_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(303775680)))]; + tensor var_9422_to_fp16 = const()[name = tensor("op_9422_to_fp16"), val = tensor(0x1.5p-17)]; + tensor x_45_cast_fp16 = layer_norm(axes = x_45_axes_0, beta = norm22_1_bias_to_fp16, epsilon = var_9422_to_fp16, gamma = norm22_1_weight_to_fp16, x = input_223_cast_fp16)[name = tensor("x_45_cast_fp16")]; + tensor linear_89_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = attn22_in_proj_weight_cast_fp16, x = x_45_cast_fp16)[name = tensor("linear_89_cast_fp16")]; + tensor var_9458 = const()[name = tensor("op_9458"), val = tensor([1, 1, 3, 16, 64])]; + tensor qkv_45_cast_fp16 = reshape(shape = var_9458, x = linear_89_cast_fp16)[name = tensor("qkv_45_cast_fp16")]; + tensor q_133_begin_0 = const()[name = tensor("q_133_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor q_133_end_0 = const()[name = tensor("q_133_end_0"), val = tensor([1, 1, 1, 16, 64])]; + tensor q_133_end_mask_0 = const()[name = tensor("q_133_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor q_133_squeeze_mask_0 = const()[name = tensor("q_133_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor q_133_cast_fp16 = slice_by_index(begin = q_133_begin_0, end = q_133_end_0, end_mask = q_133_end_mask_0, squeeze_mask = q_133_squeeze_mask_0, x = qkv_45_cast_fp16)[name = tensor("q_133_cast_fp16")]; + tensor k_89_begin_0 = const()[name = tensor("k_89_begin_0"), val = tensor([0, 0, 1, 0, 0])]; + tensor k_89_end_0 = const()[name = tensor("k_89_end_0"), val = tensor([1, 1, 2, 16, 64])]; + tensor k_89_end_mask_0 = const()[name = tensor("k_89_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor k_89_squeeze_mask_0 = const()[name = tensor("k_89_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor k_89_cast_fp16 = slice_by_index(begin = k_89_begin_0, end = k_89_end_0, end_mask = k_89_end_mask_0, squeeze_mask = k_89_squeeze_mask_0, x = qkv_45_cast_fp16)[name = tensor("k_89_cast_fp16")]; + tensor v_45_begin_0 = const()[name = tensor("v_45_begin_0"), val = tensor([0, 0, 2, 0, 0])]; + tensor v_45_end_0 = const()[name = tensor("v_45_end_0"), val = tensor([1, 1, 3, 16, 64])]; + tensor v_45_end_mask_0 = const()[name = tensor("v_45_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor v_45_squeeze_mask_0 = const()[name = tensor("v_45_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor v_45_cast_fp16 = slice_by_index(begin = v_45_begin_0, end = v_45_end_0, end_mask = v_45_end_mask_0, squeeze_mask = v_45_squeeze_mask_0, x = qkv_45_cast_fp16)[name = tensor("v_45_cast_fp16")]; + tensor var_9562 = const()[name = tensor("op_9562"), val = tensor([1, 1, 1, 1])]; + tensor position22_to_fp16_dtype_0 = const()[name = tensor("position22_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor position22_to_fp16 = cast(dtype = position22_to_fp16_dtype_0, x = position22)[name = tensor("cast_443")]; + tensor ts_137_cast_fp16 = reshape(shape = var_9562, x = position22_to_fp16)[name = tensor("ts_137_cast_fp16")]; + tensor var_9566 = const()[name = tensor("op_9566"), val = tensor([1, 1, 16, 32, 2])]; + tensor q_complex_45_cast_fp16 = reshape(shape = var_9566, x = q_133_cast_fp16)[name = tensor("q_complex_45_cast_fp16")]; + tensor var_9570 = const()[name = tensor("op_9570"), val = tensor([1, 1, 16, 32, 2])]; + tensor k_complex_45_cast_fp16 = reshape(shape = var_9570, x = k_89_cast_fp16)[name = tensor("k_complex_45_cast_fp16")]; + tensor var_9574_begin_0 = const()[name = tensor("op_9574_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_9574_end_0 = const()[name = tensor("op_9574_end_0"), val = tensor([1, 1, 16, 32, 1])]; + tensor var_9574_end_mask_0 = const()[name = tensor("op_9574_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_9574_squeeze_mask_0 = const()[name = tensor("op_9574_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_9574_cast_fp16 = slice_by_index(begin = var_9574_begin_0, end = var_9574_end_0, end_mask = var_9574_end_mask_0, squeeze_mask = var_9574_squeeze_mask_0, x = q_complex_45_cast_fp16)[name = tensor("op_9574_cast_fp16")]; + tensor var_9582_begin_0 = const()[name = tensor("op_9582_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_9582_end_0 = const()[name = tensor("op_9582_end_0"), val = tensor([1, 1, 16, 32, 2])]; + tensor var_9582_end_mask_0 = const()[name = tensor("op_9582_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_9582_squeeze_mask_0 = const()[name = tensor("op_9582_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_9582_cast_fp16 = slice_by_index(begin = var_9582_begin_0, end = var_9582_end_0, end_mask = var_9582_end_mask_0, squeeze_mask = var_9582_squeeze_mask_0, x = q_complex_45_cast_fp16)[name = tensor("op_9582_cast_fp16")]; + tensor var_9590_begin_0 = const()[name = tensor("op_9590_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_9590_end_0 = const()[name = tensor("op_9590_end_0"), val = tensor([1, 1, 16, 32, 1])]; + tensor var_9590_end_mask_0 = const()[name = tensor("op_9590_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_9590_squeeze_mask_0 = const()[name = tensor("op_9590_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_9590_cast_fp16 = slice_by_index(begin = var_9590_begin_0, end = var_9590_end_0, end_mask = var_9590_end_mask_0, squeeze_mask = var_9590_squeeze_mask_0, x = k_complex_45_cast_fp16)[name = tensor("op_9590_cast_fp16")]; + tensor var_9598_begin_0 = const()[name = tensor("op_9598_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_9598_end_0 = const()[name = tensor("op_9598_end_0"), val = tensor([1, 1, 16, 32, 2])]; + tensor var_9598_end_mask_0 = const()[name = tensor("op_9598_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_9598_squeeze_mask_0 = const()[name = tensor("op_9598_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_9598_cast_fp16 = slice_by_index(begin = var_9598_begin_0, end = var_9598_end_0, end_mask = var_9598_end_mask_0, squeeze_mask = var_9598_squeeze_mask_0, x = k_complex_45_cast_fp16)[name = tensor("op_9598_cast_fp16")]; + tensor freqs_45_to_fp16 = const()[name = tensor("freqs_45_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(303777792)))]; + tensor var_9604_cast_fp16 = mul(x = freqs_45_to_fp16, y = ts_137_cast_fp16)[name = tensor("op_9604_cast_fp16")]; + tensor rotr_45_cast_fp16 = cos(x = var_9604_cast_fp16)[name = tensor("rotr_45_cast_fp16")]; + tensor roti_45_cast_fp16 = sin(x = var_9604_cast_fp16)[name = tensor("roti_45_cast_fp16")]; + tensor var_9608_cast_fp16 = mul(x = var_9574_cast_fp16, y = rotr_45_cast_fp16)[name = tensor("op_9608_cast_fp16")]; + tensor var_9609_cast_fp16 = mul(x = var_9582_cast_fp16, y = roti_45_cast_fp16)[name = tensor("op_9609_cast_fp16")]; + tensor qor_89_cast_fp16 = sub(x = var_9608_cast_fp16, y = var_9609_cast_fp16)[name = tensor("qor_89_cast_fp16")]; + tensor var_9612_cast_fp16 = mul(x = var_9574_cast_fp16, y = roti_45_cast_fp16)[name = tensor("op_9612_cast_fp16")]; + tensor var_9613_cast_fp16 = mul(x = var_9582_cast_fp16, y = rotr_45_cast_fp16)[name = tensor("op_9613_cast_fp16")]; + tensor qoi_89_cast_fp16 = add(x = var_9612_cast_fp16, y = var_9613_cast_fp16)[name = tensor("qoi_89_cast_fp16")]; + tensor var_9616_cast_fp16 = mul(x = var_9590_cast_fp16, y = rotr_45_cast_fp16)[name = tensor("op_9616_cast_fp16")]; + tensor var_9617_cast_fp16 = mul(x = var_9598_cast_fp16, y = roti_45_cast_fp16)[name = tensor("op_9617_cast_fp16")]; + tensor kor_89_cast_fp16 = sub(x = var_9616_cast_fp16, y = var_9617_cast_fp16)[name = tensor("kor_89_cast_fp16")]; + tensor var_9620_cast_fp16 = mul(x = var_9590_cast_fp16, y = roti_45_cast_fp16)[name = tensor("op_9620_cast_fp16")]; + tensor var_9621_cast_fp16 = mul(x = var_9598_cast_fp16, y = rotr_45_cast_fp16)[name = tensor("op_9621_cast_fp16")]; + tensor koi_89_cast_fp16 = add(x = var_9620_cast_fp16, y = var_9621_cast_fp16)[name = tensor("koi_89_cast_fp16")]; + tensor qo_45_axis_0 = const()[name = tensor("qo_45_axis_0"), val = tensor(-1)]; + tensor qo_45_cast_fp16 = stack(axis = qo_45_axis_0, values = (qor_89_cast_fp16, qoi_89_cast_fp16))[name = tensor("qo_45_cast_fp16")]; + tensor ko_45_axis_0 = const()[name = tensor("ko_45_axis_0"), val = tensor(-1)]; + tensor ko_45_cast_fp16 = stack(axis = ko_45_axis_0, values = (kor_89_cast_fp16, koi_89_cast_fp16))[name = tensor("ko_45_cast_fp16")]; + tensor var_9650 = const()[name = tensor("op_9650"), val = tensor([1, 1, 16, 64])]; + tensor q_135_cast_fp16 = reshape(shape = var_9650, x = qo_45_cast_fp16)[name = tensor("q_135_cast_fp16")]; + tensor var_9652 = const()[name = tensor("op_9652"), val = tensor([1, 1, 16, 64])]; + tensor k_91_cast_fp16 = reshape(shape = var_9652, x = ko_45_cast_fp16)[name = tensor("k_91_cast_fp16")]; + tensor _inversed_9674_y_0_to_fp16 = const()[name = tensor("_inversed_9674_y_0_to_fp16"), val = tensor(0x1p-9)]; + tensor _inversed_9674_cast_fp16 = mul(x = ts_137_cast_fp16, y = _inversed_9674_y_0_to_fp16)[name = tensor("_inversed_9674_cast_fp16")]; + tensor var_9675_cast_fp16 = floor(x = _inversed_9674_cast_fp16)[name = tensor("op_9675_cast_fp16")]; + tensor var_9676_to_fp16 = const()[name = tensor("op_9676_to_fp16"), val = tensor(0x1p+9)]; + tensor var_9677_cast_fp16 = mul(x = var_9675_cast_fp16, y = var_9676_to_fp16)[name = tensor("op_9677_cast_fp16")]; + tensor write_indices_float_91_cast_fp16 = sub(x = ts_137_cast_fp16, y = var_9677_cast_fp16)[name = tensor("write_indices_float_91_cast_fp16")]; + tensor var_9684_dtype_0 = const()[name = tensor("op_9684_dtype_0"), val = tensor("int32")]; + tensor write_indices_45_reps_0 = const()[name = tensor("write_indices_45_reps_0"), val = tensor([1, 1, 16, 64])]; + tensor write_indices_float_91_cast_fp16_to_int32 = cast(dtype = var_9684_dtype_0, x = write_indices_float_91_cast_fp16)[name = tensor("cast_442")]; + tensor write_indices_45 = tile(reps = write_indices_45_reps_0, x = write_indices_float_91_cast_fp16_to_int32)[name = tensor("write_indices_45")]; + tensor var_9692_begin_0 = const()[name = tensor("op_9692_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_9692_end_0 = const()[name = tensor("op_9692_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor var_9692_end_mask_0 = const()[name = tensor("op_9692_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_9692_squeeze_mask_0 = const()[name = tensor("op_9692_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor cache22_to_fp16_dtype_0 = const()[name = tensor("cache22_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor cache22_to_fp16 = cast(dtype = cache22_to_fp16_dtype_0, x = cache22)[name = tensor("cast_441")]; + tensor var_9692_cast_fp16 = slice_by_index(begin = var_9692_begin_0, end = var_9692_end_0, end_mask = var_9692_end_mask_0, squeeze_mask = var_9692_squeeze_mask_0, x = cache22_to_fp16)[name = tensor("op_9692_cast_fp16")]; + tensor var_9694_axis_0 = const()[name = tensor("op_9694_axis_0"), val = tensor(1)]; + tensor var_9694_mode_0 = const()[name = tensor("op_9694_mode_0"), val = tensor("update")]; + tensor var_9694_validate_indices_0 = const()[name = tensor("op_9694_validate_indices_0"), val = tensor(false)]; + tensor var_9694_cast_fp16 = scatter_along_axis(axis = var_9694_axis_0, data = var_9692_cast_fp16, indices = write_indices_45, mode = var_9694_mode_0, updates = k_91_cast_fp16, validate_indices = var_9694_validate_indices_0)[name = tensor("op_9694_cast_fp16")]; + tensor concat_156 = const()[name = tensor("concat_156"), val = tensor([0, 0, 0, 0, 0])]; + tensor concat_157 = const()[name = tensor("concat_157"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_45_internal_tensor_assign_1_stride_0 = const()[name = tensor("new_cache_45_internal_tensor_assign_1_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_45_internal_tensor_assign_1_begin_mask_0 = const()[name = tensor("new_cache_45_internal_tensor_assign_1_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_45_internal_tensor_assign_1_end_mask_0 = const()[name = tensor("new_cache_45_internal_tensor_assign_1_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_45_internal_tensor_assign_1_squeeze_mask_0 = const()[name = tensor("new_cache_45_internal_tensor_assign_1_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_92 = const()[name = tensor("shape_92"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_44 = const()[name = tensor("reduce_prod_44"), val = tensor(1048576)]; + tensor range_1d_44_start_0 = const()[name = tensor("range_1d_44_start_0"), val = tensor(0)]; + tensor range_1d_44_step_0 = const()[name = tensor("range_1d_44_step_0"), val = tensor(1)]; + tensor range_1d_44 = range_1d(end = reduce_prod_44, start = range_1d_44_start_0, step = range_1d_44_step_0)[name = tensor("range_1d_44")]; + tensor reshape_220 = reshape(shape = shape_92, x = range_1d_44)[name = tensor("reshape_220")]; + tensor slice_by_index_44 = slice_by_index(begin = concat_156, begin_mask = new_cache_45_internal_tensor_assign_1_begin_mask_0, end = concat_157, end_mask = new_cache_45_internal_tensor_assign_1_end_mask_0, squeeze_mask = new_cache_45_internal_tensor_assign_1_squeeze_mask_0, stride = new_cache_45_internal_tensor_assign_1_stride_0, x = reshape_220)[name = tensor("slice_by_index_44")]; + tensor reshape_221_shape_0 = const()[name = tensor("reshape_221_shape_0"), val = tensor([-1])]; + tensor reshape_221 = reshape(shape = reshape_221_shape_0, x = slice_by_index_44)[name = tensor("reshape_221")]; + tensor reshape_222_shape_0 = const()[name = tensor("reshape_222_shape_0"), val = tensor([-1])]; + tensor reshape_222_cast_fp16 = reshape(shape = reshape_222_shape_0, x = var_9694_cast_fp16)[name = tensor("reshape_222_cast_fp16")]; + tensor reshape_223_shape_0 = const()[name = tensor("reshape_223_shape_0"), val = tensor([-1])]; + tensor reshape_223_cast_fp16 = reshape(shape = reshape_223_shape_0, x = cache22_to_fp16)[name = tensor("reshape_223_cast_fp16")]; + tensor scatter_44_mode_0 = const()[name = tensor("scatter_44_mode_0"), val = tensor("update")]; + tensor scatter_44_axis_0 = const()[name = tensor("scatter_44_axis_0"), val = tensor(0)]; + tensor scatter_44_validate_indices_0 = const()[name = tensor("scatter_44_validate_indices_0"), val = tensor(false)]; + tensor scatter_44_cast_fp16 = scatter(axis = scatter_44_axis_0, data = reshape_223_cast_fp16, indices = reshape_221, mode = scatter_44_mode_0, updates = reshape_222_cast_fp16, validate_indices = scatter_44_validate_indices_0)[name = tensor("scatter_44_cast_fp16")]; + tensor reshape_224_cast_fp16 = reshape(shape = shape_92, x = scatter_44_cast_fp16)[name = tensor("reshape_224_cast_fp16")]; + tensor var_9702_begin_0 = const()[name = tensor("op_9702_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor var_9702_end_0 = const()[name = tensor("op_9702_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor var_9702_end_mask_0 = const()[name = tensor("op_9702_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_9702_squeeze_mask_0 = const()[name = tensor("op_9702_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor var_9702_cast_fp16 = slice_by_index(begin = var_9702_begin_0, end = var_9702_end_0, end_mask = var_9702_end_mask_0, squeeze_mask = var_9702_squeeze_mask_0, x = reshape_224_cast_fp16)[name = tensor("op_9702_cast_fp16")]; + tensor var_9704_axis_0 = const()[name = tensor("op_9704_axis_0"), val = tensor(1)]; + tensor var_9704_mode_0 = const()[name = tensor("op_9704_mode_0"), val = tensor("update")]; + tensor var_9704_validate_indices_0 = const()[name = tensor("op_9704_validate_indices_0"), val = tensor(false)]; + tensor var_9704_cast_fp16 = scatter_along_axis(axis = var_9704_axis_0, data = var_9702_cast_fp16, indices = write_indices_45, mode = var_9704_mode_0, updates = v_45_cast_fp16, validate_indices = var_9704_validate_indices_0)[name = tensor("op_9704_cast_fp16")]; + tensor concat_158 = const()[name = tensor("concat_158"), val = tensor([1, 0, 0, 0, 0])]; + tensor concat_159 = const()[name = tensor("concat_159"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_45_internal_tensor_assign_2_stride_0 = const()[name = tensor("new_cache_45_internal_tensor_assign_2_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_45_internal_tensor_assign_2_begin_mask_0 = const()[name = tensor("new_cache_45_internal_tensor_assign_2_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_45_internal_tensor_assign_2_end_mask_0 = const()[name = tensor("new_cache_45_internal_tensor_assign_2_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_45_internal_tensor_assign_2_squeeze_mask_0 = const()[name = tensor("new_cache_45_internal_tensor_assign_2_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_93 = const()[name = tensor("shape_93"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_45 = const()[name = tensor("reduce_prod_45"), val = tensor(1048576)]; + tensor range_1d_45_start_0 = const()[name = tensor("range_1d_45_start_0"), val = tensor(0)]; + tensor range_1d_45_step_0 = const()[name = tensor("range_1d_45_step_0"), val = tensor(1)]; + tensor range_1d_45 = range_1d(end = reduce_prod_45, start = range_1d_45_start_0, step = range_1d_45_step_0)[name = tensor("range_1d_45")]; + tensor reshape_225 = reshape(shape = shape_93, x = range_1d_45)[name = tensor("reshape_225")]; + tensor slice_by_index_45 = slice_by_index(begin = concat_158, begin_mask = new_cache_45_internal_tensor_assign_2_begin_mask_0, end = concat_159, end_mask = new_cache_45_internal_tensor_assign_2_end_mask_0, squeeze_mask = new_cache_45_internal_tensor_assign_2_squeeze_mask_0, stride = new_cache_45_internal_tensor_assign_2_stride_0, x = reshape_225)[name = tensor("slice_by_index_45")]; + tensor reshape_226_shape_0 = const()[name = tensor("reshape_226_shape_0"), val = tensor([-1])]; + tensor reshape_226 = reshape(shape = reshape_226_shape_0, x = slice_by_index_45)[name = tensor("reshape_226")]; + tensor reshape_227_shape_0 = const()[name = tensor("reshape_227_shape_0"), val = tensor([-1])]; + tensor reshape_227_cast_fp16 = reshape(shape = reshape_227_shape_0, x = var_9704_cast_fp16)[name = tensor("reshape_227_cast_fp16")]; + tensor reshape_228_shape_0 = const()[name = tensor("reshape_228_shape_0"), val = tensor([-1])]; + tensor reshape_228_cast_fp16 = reshape(shape = reshape_228_shape_0, x = reshape_224_cast_fp16)[name = tensor("reshape_228_cast_fp16")]; + tensor scatter_45_mode_0 = const()[name = tensor("scatter_45_mode_0"), val = tensor("update")]; + tensor scatter_45_axis_0 = const()[name = tensor("scatter_45_axis_0"), val = tensor(0)]; + tensor scatter_45_validate_indices_0 = const()[name = tensor("scatter_45_validate_indices_0"), val = tensor(false)]; + tensor scatter_45_cast_fp16 = scatter(axis = scatter_45_axis_0, data = reshape_228_cast_fp16, indices = reshape_226, mode = scatter_45_mode_0, updates = reshape_227_cast_fp16, validate_indices = scatter_45_validate_indices_0)[name = tensor("scatter_45_cast_fp16")]; + tensor reshape_229_cast_fp16 = reshape(shape = shape_93, x = scatter_45_cast_fp16)[name = tensor("reshape_229_cast_fp16")]; + tensor reshape_229_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("reshape_229_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor keys_133_begin_0 = const()[name = tensor("keys_133_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor keys_133_end_0 = const()[name = tensor("keys_133_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor keys_133_end_mask_0 = const()[name = tensor("keys_133_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor keys_133_squeeze_mask_0 = const()[name = tensor("keys_133_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor keys_133_cast_fp16 = slice_by_index(begin = keys_133_begin_0, end = keys_133_end_0, end_mask = keys_133_end_mask_0, squeeze_mask = keys_133_squeeze_mask_0, x = reshape_229_cast_fp16)[name = tensor("keys_133_cast_fp16")]; + tensor values_133_begin_0 = const()[name = tensor("values_133_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor values_133_end_0 = const()[name = tensor("values_133_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor values_133_end_mask_0 = const()[name = tensor("values_133_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor values_133_squeeze_mask_0 = const()[name = tensor("values_133_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor values_133_cast_fp16 = slice_by_index(begin = values_133_begin_0, end = values_133_end_0, end_mask = values_133_end_mask_0, squeeze_mask = values_133_squeeze_mask_0, x = reshape_229_cast_fp16)[name = tensor("values_133_cast_fp16")]; + tensor var_9716_cast_fp16 = not_equal(x = keys_133_cast_fp16, y = keys_133_cast_fp16)[name = tensor("op_9716_cast_fp16")]; + tensor keys_135_cast_fp16 = select(a = var_504_to_fp16, b = keys_133_cast_fp16, cond = var_9716_cast_fp16)[name = tensor("keys_135_cast_fp16")]; + tensor var_9724_cast_fp16 = not_equal(x = values_133_cast_fp16, y = values_133_cast_fp16)[name = tensor("op_9724_cast_fp16")]; + tensor values_135_cast_fp16 = select(a = var_504_to_fp16, b = values_133_cast_fp16, cond = var_9724_cast_fp16)[name = tensor("values_135_cast_fp16")]; + tensor var_9748 = const()[name = tensor("op_9748"), val = tensor([0, 2, 1, 3])]; + tensor var_9761 = const()[name = tensor("op_9761"), val = tensor([1, 1, 1])]; + tensor var_9762_cast_fp16 = reshape(shape = var_9761, x = position22_to_fp16)[name = tensor("op_9762_cast_fp16")]; + tensor var_9779_to_fp16 = const()[name = tensor("op_9779_to_fp16"), val = tensor(0x1p+0)]; + tensor valid_len_45_cast_fp16 = add(x = var_9762_cast_fp16, y = var_9779_to_fp16)[name = tensor("valid_len_45_cast_fp16")]; + tensor valid_mask_45_cast_fp16 = less(x = k_positions_1_promoted_to_fp16, y = valid_len_45_cast_fp16)[name = tensor("valid_mask_45_cast_fp16")]; + tensor causal_mask_45_cast_fp16 = less_equal(x = k_positions_1_promoted_to_fp16, y = var_9762_cast_fp16)[name = tensor("causal_mask_45_cast_fp16")]; + tensor attn_mask_89 = logical_and(x = valid_mask_45_cast_fp16, y = causal_mask_45_cast_fp16)[name = tensor("attn_mask_89")]; + tensor attn_mask_91_axes_0 = const()[name = tensor("attn_mask_91_axes_0"), val = tensor([1])]; + tensor attn_mask_91 = expand_dims(axes = attn_mask_91_axes_0, x = attn_mask_89)[name = tensor("attn_mask_91")]; + tensor var_9797_transpose_x_0 = const()[name = tensor("op_9797_transpose_x_0"), val = tensor(false)]; + tensor var_9797_transpose_y_0 = const()[name = tensor("op_9797_transpose_y_0"), val = tensor(false)]; + tensor transpose_116_perm_0 = const()[name = tensor("transpose_116_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_117_perm_0 = const()[name = tensor("transpose_117_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_117 = transpose(perm = transpose_117_perm_0, x = keys_135_cast_fp16)[name = tensor("transpose_125")]; + tensor transpose_116 = transpose(perm = transpose_116_perm_0, x = q_135_cast_fp16)[name = tensor("transpose_126")]; + tensor var_9797_cast_fp16 = matmul(transpose_x = var_9797_transpose_x_0, transpose_y = var_9797_transpose_y_0, x = transpose_116, y = transpose_117)[name = tensor("op_9797_cast_fp16")]; + tensor var_9791_to_fp16 = const()[name = tensor("op_9791_to_fp16"), val = tensor([0x1p-3])]; + tensor attn_weights_133_cast_fp16 = mul(x = var_9797_cast_fp16, y = var_9791_to_fp16)[name = tensor("attn_weights_133_cast_fp16")]; + tensor var_9799 = logical_not(x = attn_mask_91)[name = tensor("op_9799")]; + tensor var_9800_to_fp16 = const()[name = tensor("op_9800_to_fp16"), val = tensor(-inf)]; + tensor attn_weights_135_cast_fp16 = select(a = var_9800_to_fp16, b = attn_weights_133_cast_fp16, cond = var_9799)[name = tensor("attn_weights_135_cast_fp16")]; + tensor var_9802 = const()[name = tensor("op_9802"), val = tensor(-1)]; + tensor attn_weights_137_cast_fp16 = softmax(axis = var_9802, x = attn_weights_135_cast_fp16)[name = tensor("attn_weights_137_cast_fp16")]; + tensor attn_output_45_transpose_x_0 = const()[name = tensor("attn_output_45_transpose_x_0"), val = tensor(false)]; + tensor attn_output_45_transpose_y_0 = const()[name = tensor("attn_output_45_transpose_y_0"), val = tensor(false)]; + tensor values_137_cast_fp16 = transpose(perm = var_9748, x = values_135_cast_fp16)[name = tensor("transpose_127")]; + tensor attn_output_45_cast_fp16 = matmul(transpose_x = attn_output_45_transpose_x_0, transpose_y = attn_output_45_transpose_y_0, x = attn_weights_137_cast_fp16, y = values_137_cast_fp16)[name = tensor("attn_output_45_cast_fp16")]; + tensor var_9810 = const()[name = tensor("op_9810"), val = tensor([0, 2, 1, 3])]; + tensor var_9813 = const()[name = tensor("op_9813"), val = tensor([1, 1, 1024])]; + tensor var_9811_cast_fp16 = transpose(perm = var_9810, x = attn_output_45_cast_fp16)[name = tensor("transpose_124")]; + tensor input_225_cast_fp16 = reshape(shape = var_9813, x = var_9811_cast_fp16)[name = tensor("input_225_cast_fp16")]; + tensor linear_90_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = attn22_out_proj_weight_cast_fp16, x = input_225_cast_fp16)[name = tensor("linear_90_cast_fp16")]; + tensor var_9819_to_fp16 = const()[name = tensor("op_9819_to_fp16"), val = tensor(0x1p+0)]; + tensor var_9820_cast_fp16 = add(x = position22_to_fp16, y = var_9819_to_fp16)[name = tensor("op_9820_cast_fp16")]; + tensor var_9820_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_9820_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor input_227_cast_fp16 = add(x = input_223_cast_fp16, y = linear_90_cast_fp16)[name = tensor("input_227_cast_fp16")]; + tensor input_229_axes_0 = const()[name = tensor("input_229_axes_0"), val = tensor([-1])]; + tensor norm22_2_weight_to_fp16 = const()[name = tensor("norm22_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(303777920)))]; + tensor norm22_2_bias_to_fp16 = const()[name = tensor("norm22_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(303780032)))]; + tensor var_9824_to_fp16 = const()[name = tensor("op_9824_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_229_cast_fp16 = layer_norm(axes = input_229_axes_0, beta = norm22_2_bias_to_fp16, epsilon = var_9824_to_fp16, gamma = norm22_2_weight_to_fp16, x = input_227_cast_fp16)[name = tensor("input_229_cast_fp16")]; + tensor linear_91_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = linear22_1_weight_cast_fp16, x = input_229_cast_fp16)[name = tensor("linear_91_cast_fp16")]; + tensor input_231_mode_0 = const()[name = tensor("input_231_mode_0"), val = tensor("EXACT")]; + tensor input_231_cast_fp16 = gelu(mode = input_231_mode_0, x = linear_91_cast_fp16)[name = tensor("input_231_cast_fp16")]; + tensor linear_92_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = linear22_2_weight_cast_fp16, x = input_231_cast_fp16)[name = tensor("linear_92_cast_fp16")]; + tensor input_233_cast_fp16 = add(x = input_227_cast_fp16, y = linear_92_cast_fp16)[name = tensor("input_233_cast_fp16")]; + tensor x_axes_0 = const()[name = tensor("x_axes_0"), val = tensor([-1])]; + tensor norm23_1_weight_to_fp16 = const()[name = tensor("norm23_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(303782144)))]; + tensor norm23_1_bias_to_fp16 = const()[name = tensor("norm23_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(303784256)))]; + tensor var_9841_to_fp16 = const()[name = tensor("op_9841_to_fp16"), val = tensor(0x1.5p-17)]; + tensor x_cast_fp16 = layer_norm(axes = x_axes_0, beta = norm23_1_bias_to_fp16, epsilon = var_9841_to_fp16, gamma = norm23_1_weight_to_fp16, x = input_233_cast_fp16)[name = tensor("x_cast_fp16")]; + tensor linear_93_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = attn23_in_proj_weight_cast_fp16, x = x_cast_fp16)[name = tensor("linear_93_cast_fp16")]; + tensor var_9877 = const()[name = tensor("op_9877"), val = tensor([1, 1, 3, 16, 64])]; + tensor qkv_cast_fp16 = reshape(shape = var_9877, x = linear_93_cast_fp16)[name = tensor("qkv_cast_fp16")]; + tensor q_139_begin_0 = const()[name = tensor("q_139_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor q_139_end_0 = const()[name = tensor("q_139_end_0"), val = tensor([1, 1, 1, 16, 64])]; + tensor q_139_end_mask_0 = const()[name = tensor("q_139_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor q_139_squeeze_mask_0 = const()[name = tensor("q_139_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor q_139_cast_fp16 = slice_by_index(begin = q_139_begin_0, end = q_139_end_0, end_mask = q_139_end_mask_0, squeeze_mask = q_139_squeeze_mask_0, x = qkv_cast_fp16)[name = tensor("q_139_cast_fp16")]; + tensor k_93_begin_0 = const()[name = tensor("k_93_begin_0"), val = tensor([0, 0, 1, 0, 0])]; + tensor k_93_end_0 = const()[name = tensor("k_93_end_0"), val = tensor([1, 1, 2, 16, 64])]; + tensor k_93_end_mask_0 = const()[name = tensor("k_93_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor k_93_squeeze_mask_0 = const()[name = tensor("k_93_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor k_93_cast_fp16 = slice_by_index(begin = k_93_begin_0, end = k_93_end_0, end_mask = k_93_end_mask_0, squeeze_mask = k_93_squeeze_mask_0, x = qkv_cast_fp16)[name = tensor("k_93_cast_fp16")]; + tensor v_begin_0 = const()[name = tensor("v_begin_0"), val = tensor([0, 0, 2, 0, 0])]; + tensor v_end_0 = const()[name = tensor("v_end_0"), val = tensor([1, 1, 3, 16, 64])]; + tensor v_end_mask_0 = const()[name = tensor("v_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor v_squeeze_mask_0 = const()[name = tensor("v_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor v_cast_fp16 = slice_by_index(begin = v_begin_0, end = v_end_0, end_mask = v_end_mask_0, squeeze_mask = v_squeeze_mask_0, x = qkv_cast_fp16)[name = tensor("v_cast_fp16")]; + tensor var_9981 = const()[name = tensor("op_9981"), val = tensor([1, 1, 1, 1])]; + tensor position23_to_fp16_dtype_0 = const()[name = tensor("position23_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor position23_to_fp16 = cast(dtype = position23_to_fp16_dtype_0, x = position23)[name = tensor("cast_438")]; + tensor ts_cast_fp16 = reshape(shape = var_9981, x = position23_to_fp16)[name = tensor("ts_cast_fp16")]; + tensor var_9985 = const()[name = tensor("op_9985"), val = tensor([1, 1, 16, 32, 2])]; + tensor q_complex_cast_fp16 = reshape(shape = var_9985, x = q_139_cast_fp16)[name = tensor("q_complex_cast_fp16")]; + tensor var_9989 = const()[name = tensor("op_9989"), val = tensor([1, 1, 16, 32, 2])]; + tensor k_complex_cast_fp16 = reshape(shape = var_9989, x = k_93_cast_fp16)[name = tensor("k_complex_cast_fp16")]; + tensor var_9993_begin_0 = const()[name = tensor("op_9993_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_9993_end_0 = const()[name = tensor("op_9993_end_0"), val = tensor([1, 1, 16, 32, 1])]; + tensor var_9993_end_mask_0 = const()[name = tensor("op_9993_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_9993_squeeze_mask_0 = const()[name = tensor("op_9993_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_9993_cast_fp16 = slice_by_index(begin = var_9993_begin_0, end = var_9993_end_0, end_mask = var_9993_end_mask_0, squeeze_mask = var_9993_squeeze_mask_0, x = q_complex_cast_fp16)[name = tensor("op_9993_cast_fp16")]; + tensor var_10001_begin_0 = const()[name = tensor("op_10001_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_10001_end_0 = const()[name = tensor("op_10001_end_0"), val = tensor([1, 1, 16, 32, 2])]; + tensor var_10001_end_mask_0 = const()[name = tensor("op_10001_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_10001_squeeze_mask_0 = const()[name = tensor("op_10001_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_10001_cast_fp16 = slice_by_index(begin = var_10001_begin_0, end = var_10001_end_0, end_mask = var_10001_end_mask_0, squeeze_mask = var_10001_squeeze_mask_0, x = q_complex_cast_fp16)[name = tensor("op_10001_cast_fp16")]; + tensor var_10009_begin_0 = const()[name = tensor("op_10009_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_10009_end_0 = const()[name = tensor("op_10009_end_0"), val = tensor([1, 1, 16, 32, 1])]; + tensor var_10009_end_mask_0 = const()[name = tensor("op_10009_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_10009_squeeze_mask_0 = const()[name = tensor("op_10009_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_10009_cast_fp16 = slice_by_index(begin = var_10009_begin_0, end = var_10009_end_0, end_mask = var_10009_end_mask_0, squeeze_mask = var_10009_squeeze_mask_0, x = k_complex_cast_fp16)[name = tensor("op_10009_cast_fp16")]; + tensor var_10017_begin_0 = const()[name = tensor("op_10017_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_10017_end_0 = const()[name = tensor("op_10017_end_0"), val = tensor([1, 1, 16, 32, 2])]; + tensor var_10017_end_mask_0 = const()[name = tensor("op_10017_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_10017_squeeze_mask_0 = const()[name = tensor("op_10017_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_10017_cast_fp16 = slice_by_index(begin = var_10017_begin_0, end = var_10017_end_0, end_mask = var_10017_end_mask_0, squeeze_mask = var_10017_squeeze_mask_0, x = k_complex_cast_fp16)[name = tensor("op_10017_cast_fp16")]; + tensor freqs_to_fp16 = const()[name = tensor("freqs_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(303786368)))]; + tensor var_10023_cast_fp16 = mul(x = freqs_to_fp16, y = ts_cast_fp16)[name = tensor("op_10023_cast_fp16")]; + tensor rotr_cast_fp16 = cos(x = var_10023_cast_fp16)[name = tensor("rotr_cast_fp16")]; + tensor roti_cast_fp16 = sin(x = var_10023_cast_fp16)[name = tensor("roti_cast_fp16")]; + tensor var_10027_cast_fp16 = mul(x = var_9993_cast_fp16, y = rotr_cast_fp16)[name = tensor("op_10027_cast_fp16")]; + tensor var_10028_cast_fp16 = mul(x = var_10001_cast_fp16, y = roti_cast_fp16)[name = tensor("op_10028_cast_fp16")]; + tensor qor_93_cast_fp16 = sub(x = var_10027_cast_fp16, y = var_10028_cast_fp16)[name = tensor("qor_93_cast_fp16")]; + tensor var_10031_cast_fp16 = mul(x = var_9993_cast_fp16, y = roti_cast_fp16)[name = tensor("op_10031_cast_fp16")]; + tensor var_10032_cast_fp16 = mul(x = var_10001_cast_fp16, y = rotr_cast_fp16)[name = tensor("op_10032_cast_fp16")]; + tensor qoi_93_cast_fp16 = add(x = var_10031_cast_fp16, y = var_10032_cast_fp16)[name = tensor("qoi_93_cast_fp16")]; + tensor var_10035_cast_fp16 = mul(x = var_10009_cast_fp16, y = rotr_cast_fp16)[name = tensor("op_10035_cast_fp16")]; + tensor var_10036_cast_fp16 = mul(x = var_10017_cast_fp16, y = roti_cast_fp16)[name = tensor("op_10036_cast_fp16")]; + tensor kor_93_cast_fp16 = sub(x = var_10035_cast_fp16, y = var_10036_cast_fp16)[name = tensor("kor_93_cast_fp16")]; + tensor var_10039_cast_fp16 = mul(x = var_10009_cast_fp16, y = roti_cast_fp16)[name = tensor("op_10039_cast_fp16")]; + tensor var_10040_cast_fp16 = mul(x = var_10017_cast_fp16, y = rotr_cast_fp16)[name = tensor("op_10040_cast_fp16")]; + tensor koi_93_cast_fp16 = add(x = var_10039_cast_fp16, y = var_10040_cast_fp16)[name = tensor("koi_93_cast_fp16")]; + tensor qo_axis_0 = const()[name = tensor("qo_axis_0"), val = tensor(-1)]; + tensor qo_cast_fp16 = stack(axis = qo_axis_0, values = (qor_93_cast_fp16, qoi_93_cast_fp16))[name = tensor("qo_cast_fp16")]; + tensor ko_axis_0 = const()[name = tensor("ko_axis_0"), val = tensor(-1)]; + tensor ko_cast_fp16 = stack(axis = ko_axis_0, values = (kor_93_cast_fp16, koi_93_cast_fp16))[name = tensor("ko_cast_fp16")]; + tensor var_10069 = const()[name = tensor("op_10069"), val = tensor([1, 1, 16, 64])]; + tensor q_141_cast_fp16 = reshape(shape = var_10069, x = qo_cast_fp16)[name = tensor("q_141_cast_fp16")]; + tensor var_10071 = const()[name = tensor("op_10071"), val = tensor([1, 1, 16, 64])]; + tensor k_cast_fp16 = reshape(shape = var_10071, x = ko_cast_fp16)[name = tensor("k_cast_fp16")]; + tensor _inversed_10093_y_0_to_fp16 = const()[name = tensor("_inversed_10093_y_0_to_fp16"), val = tensor(0x1p-9)]; + tensor _inversed_10093_cast_fp16 = mul(x = ts_cast_fp16, y = _inversed_10093_y_0_to_fp16)[name = tensor("_inversed_10093_cast_fp16")]; + tensor var_10094_cast_fp16 = floor(x = _inversed_10093_cast_fp16)[name = tensor("op_10094_cast_fp16")]; + tensor var_10095_to_fp16 = const()[name = tensor("op_10095_to_fp16"), val = tensor(0x1p+9)]; + tensor var_10096_cast_fp16 = mul(x = var_10094_cast_fp16, y = var_10095_to_fp16)[name = tensor("op_10096_cast_fp16")]; + tensor write_indices_float_cast_fp16 = sub(x = ts_cast_fp16, y = var_10096_cast_fp16)[name = tensor("write_indices_float_cast_fp16")]; + tensor var_10103_dtype_0 = const()[name = tensor("op_10103_dtype_0"), val = tensor("int32")]; + tensor write_indices_reps_0 = const()[name = tensor("write_indices_reps_0"), val = tensor([1, 1, 16, 64])]; + tensor write_indices_float_cast_fp16_to_int32 = cast(dtype = var_10103_dtype_0, x = write_indices_float_cast_fp16)[name = tensor("cast_437")]; + tensor write_indices = tile(reps = write_indices_reps_0, x = write_indices_float_cast_fp16_to_int32)[name = tensor("write_indices")]; + tensor var_10111_begin_0 = const()[name = tensor("op_10111_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_10111_end_0 = const()[name = tensor("op_10111_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor var_10111_end_mask_0 = const()[name = tensor("op_10111_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_10111_squeeze_mask_0 = const()[name = tensor("op_10111_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor cache23_to_fp16_dtype_0 = const()[name = tensor("cache23_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor cache23_to_fp16 = cast(dtype = cache23_to_fp16_dtype_0, x = cache23)[name = tensor("cast_436")]; + tensor var_10111_cast_fp16 = slice_by_index(begin = var_10111_begin_0, end = var_10111_end_0, end_mask = var_10111_end_mask_0, squeeze_mask = var_10111_squeeze_mask_0, x = cache23_to_fp16)[name = tensor("op_10111_cast_fp16")]; + tensor var_10113_axis_0 = const()[name = tensor("op_10113_axis_0"), val = tensor(1)]; + tensor var_10113_mode_0 = const()[name = tensor("op_10113_mode_0"), val = tensor("update")]; + tensor var_10113_validate_indices_0 = const()[name = tensor("op_10113_validate_indices_0"), val = tensor(false)]; + tensor var_10113_cast_fp16 = scatter_along_axis(axis = var_10113_axis_0, data = var_10111_cast_fp16, indices = write_indices, mode = var_10113_mode_0, updates = k_cast_fp16, validate_indices = var_10113_validate_indices_0)[name = tensor("op_10113_cast_fp16")]; + tensor concat_163 = const()[name = tensor("concat_163"), val = tensor([0, 0, 0, 0, 0])]; + tensor concat_164 = const()[name = tensor("concat_164"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_internal_tensor_assign_1_stride_0 = const()[name = tensor("new_cache_internal_tensor_assign_1_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_internal_tensor_assign_1_begin_mask_0 = const()[name = tensor("new_cache_internal_tensor_assign_1_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_internal_tensor_assign_1_end_mask_0 = const()[name = tensor("new_cache_internal_tensor_assign_1_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_internal_tensor_assign_1_squeeze_mask_0 = const()[name = tensor("new_cache_internal_tensor_assign_1_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_94 = const()[name = tensor("shape_94"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_46 = const()[name = tensor("reduce_prod_46"), val = tensor(1048576)]; + tensor range_1d_46_start_0 = const()[name = tensor("range_1d_46_start_0"), val = tensor(0)]; + tensor range_1d_46_step_0 = const()[name = tensor("range_1d_46_step_0"), val = tensor(1)]; + tensor range_1d_46 = range_1d(end = reduce_prod_46, start = range_1d_46_start_0, step = range_1d_46_step_0)[name = tensor("range_1d_46")]; + tensor reshape_230 = reshape(shape = shape_94, x = range_1d_46)[name = tensor("reshape_230")]; + tensor slice_by_index_46 = slice_by_index(begin = concat_163, begin_mask = new_cache_internal_tensor_assign_1_begin_mask_0, end = concat_164, end_mask = new_cache_internal_tensor_assign_1_end_mask_0, squeeze_mask = new_cache_internal_tensor_assign_1_squeeze_mask_0, stride = new_cache_internal_tensor_assign_1_stride_0, x = reshape_230)[name = tensor("slice_by_index_46")]; + tensor reshape_231_shape_0 = const()[name = tensor("reshape_231_shape_0"), val = tensor([-1])]; + tensor reshape_231 = reshape(shape = reshape_231_shape_0, x = slice_by_index_46)[name = tensor("reshape_231")]; + tensor reshape_232_shape_0 = const()[name = tensor("reshape_232_shape_0"), val = tensor([-1])]; + tensor reshape_232_cast_fp16 = reshape(shape = reshape_232_shape_0, x = var_10113_cast_fp16)[name = tensor("reshape_232_cast_fp16")]; + tensor reshape_233_shape_0 = const()[name = tensor("reshape_233_shape_0"), val = tensor([-1])]; + tensor reshape_233_cast_fp16 = reshape(shape = reshape_233_shape_0, x = cache23_to_fp16)[name = tensor("reshape_233_cast_fp16")]; + tensor scatter_46_mode_0 = const()[name = tensor("scatter_46_mode_0"), val = tensor("update")]; + tensor scatter_46_axis_0 = const()[name = tensor("scatter_46_axis_0"), val = tensor(0)]; + tensor scatter_46_validate_indices_0 = const()[name = tensor("scatter_46_validate_indices_0"), val = tensor(false)]; + tensor scatter_46_cast_fp16 = scatter(axis = scatter_46_axis_0, data = reshape_233_cast_fp16, indices = reshape_231, mode = scatter_46_mode_0, updates = reshape_232_cast_fp16, validate_indices = scatter_46_validate_indices_0)[name = tensor("scatter_46_cast_fp16")]; + tensor reshape_234_cast_fp16 = reshape(shape = shape_94, x = scatter_46_cast_fp16)[name = tensor("reshape_234_cast_fp16")]; + tensor var_10121_begin_0 = const()[name = tensor("op_10121_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor var_10121_end_0 = const()[name = tensor("op_10121_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor var_10121_end_mask_0 = const()[name = tensor("op_10121_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_10121_squeeze_mask_0 = const()[name = tensor("op_10121_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor var_10121_cast_fp16 = slice_by_index(begin = var_10121_begin_0, end = var_10121_end_0, end_mask = var_10121_end_mask_0, squeeze_mask = var_10121_squeeze_mask_0, x = reshape_234_cast_fp16)[name = tensor("op_10121_cast_fp16")]; + tensor var_10123_axis_0 = const()[name = tensor("op_10123_axis_0"), val = tensor(1)]; + tensor var_10123_mode_0 = const()[name = tensor("op_10123_mode_0"), val = tensor("update")]; + tensor var_10123_validate_indices_0 = const()[name = tensor("op_10123_validate_indices_0"), val = tensor(false)]; + tensor var_10123_cast_fp16 = scatter_along_axis(axis = var_10123_axis_0, data = var_10121_cast_fp16, indices = write_indices, mode = var_10123_mode_0, updates = v_cast_fp16, validate_indices = var_10123_validate_indices_0)[name = tensor("op_10123_cast_fp16")]; + tensor concat_165 = const()[name = tensor("concat_165"), val = tensor([1, 0, 0, 0, 0])]; + tensor concat_166 = const()[name = tensor("concat_166"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_internal_tensor_assign_2_stride_0 = const()[name = tensor("new_cache_internal_tensor_assign_2_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_internal_tensor_assign_2_begin_mask_0 = const()[name = tensor("new_cache_internal_tensor_assign_2_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_internal_tensor_assign_2_end_mask_0 = const()[name = tensor("new_cache_internal_tensor_assign_2_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_internal_tensor_assign_2_squeeze_mask_0 = const()[name = tensor("new_cache_internal_tensor_assign_2_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_95 = const()[name = tensor("shape_95"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_47 = const()[name = tensor("reduce_prod_47"), val = tensor(1048576)]; + tensor range_1d_47_start_0 = const()[name = tensor("range_1d_47_start_0"), val = tensor(0)]; + tensor range_1d_47_step_0 = const()[name = tensor("range_1d_47_step_0"), val = tensor(1)]; + tensor range_1d_47 = range_1d(end = reduce_prod_47, start = range_1d_47_start_0, step = range_1d_47_step_0)[name = tensor("range_1d_47")]; + tensor reshape_235 = reshape(shape = shape_95, x = range_1d_47)[name = tensor("reshape_235")]; + tensor slice_by_index_47 = slice_by_index(begin = concat_165, begin_mask = new_cache_internal_tensor_assign_2_begin_mask_0, end = concat_166, end_mask = new_cache_internal_tensor_assign_2_end_mask_0, squeeze_mask = new_cache_internal_tensor_assign_2_squeeze_mask_0, stride = new_cache_internal_tensor_assign_2_stride_0, x = reshape_235)[name = tensor("slice_by_index_47")]; + tensor reshape_236_shape_0 = const()[name = tensor("reshape_236_shape_0"), val = tensor([-1])]; + tensor reshape_236 = reshape(shape = reshape_236_shape_0, x = slice_by_index_47)[name = tensor("reshape_236")]; + tensor reshape_237_shape_0 = const()[name = tensor("reshape_237_shape_0"), val = tensor([-1])]; + tensor reshape_237_cast_fp16 = reshape(shape = reshape_237_shape_0, x = var_10123_cast_fp16)[name = tensor("reshape_237_cast_fp16")]; + tensor reshape_238_shape_0 = const()[name = tensor("reshape_238_shape_0"), val = tensor([-1])]; + tensor reshape_238_cast_fp16 = reshape(shape = reshape_238_shape_0, x = reshape_234_cast_fp16)[name = tensor("reshape_238_cast_fp16")]; + tensor scatter_47_mode_0 = const()[name = tensor("scatter_47_mode_0"), val = tensor("update")]; + tensor scatter_47_axis_0 = const()[name = tensor("scatter_47_axis_0"), val = tensor(0)]; + tensor scatter_47_validate_indices_0 = const()[name = tensor("scatter_47_validate_indices_0"), val = tensor(false)]; + tensor scatter_47_cast_fp16 = scatter(axis = scatter_47_axis_0, data = reshape_238_cast_fp16, indices = reshape_236, mode = scatter_47_mode_0, updates = reshape_237_cast_fp16, validate_indices = scatter_47_validate_indices_0)[name = tensor("scatter_47_cast_fp16")]; + tensor reshape_239_cast_fp16 = reshape(shape = shape_95, x = scatter_47_cast_fp16)[name = tensor("reshape_239_cast_fp16")]; + tensor reshape_239_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("reshape_239_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor keys_139_begin_0 = const()[name = tensor("keys_139_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor keys_139_end_0 = const()[name = tensor("keys_139_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor keys_139_end_mask_0 = const()[name = tensor("keys_139_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor keys_139_squeeze_mask_0 = const()[name = tensor("keys_139_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor keys_139_cast_fp16 = slice_by_index(begin = keys_139_begin_0, end = keys_139_end_0, end_mask = keys_139_end_mask_0, squeeze_mask = keys_139_squeeze_mask_0, x = reshape_239_cast_fp16)[name = tensor("keys_139_cast_fp16")]; + tensor values_139_begin_0 = const()[name = tensor("values_139_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor values_139_end_0 = const()[name = tensor("values_139_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor values_139_end_mask_0 = const()[name = tensor("values_139_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor values_139_squeeze_mask_0 = const()[name = tensor("values_139_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor values_139_cast_fp16 = slice_by_index(begin = values_139_begin_0, end = values_139_end_0, end_mask = values_139_end_mask_0, squeeze_mask = values_139_squeeze_mask_0, x = reshape_239_cast_fp16)[name = tensor("values_139_cast_fp16")]; + tensor var_10135_cast_fp16 = not_equal(x = keys_139_cast_fp16, y = keys_139_cast_fp16)[name = tensor("op_10135_cast_fp16")]; + tensor keys_141_cast_fp16 = select(a = var_504_to_fp16, b = keys_139_cast_fp16, cond = var_10135_cast_fp16)[name = tensor("keys_141_cast_fp16")]; + tensor var_10143_cast_fp16 = not_equal(x = values_139_cast_fp16, y = values_139_cast_fp16)[name = tensor("op_10143_cast_fp16")]; + tensor values_141_cast_fp16 = select(a = var_504_to_fp16, b = values_139_cast_fp16, cond = var_10143_cast_fp16)[name = tensor("values_141_cast_fp16")]; + tensor var_10167 = const()[name = tensor("op_10167"), val = tensor([0, 2, 1, 3])]; + tensor var_10180 = const()[name = tensor("op_10180"), val = tensor([1, 1, 1])]; + tensor var_10181_cast_fp16 = reshape(shape = var_10180, x = position23_to_fp16)[name = tensor("op_10181_cast_fp16")]; + tensor var_10198_to_fp16 = const()[name = tensor("op_10198_to_fp16"), val = tensor(0x1p+0)]; + tensor valid_len_cast_fp16 = add(x = var_10181_cast_fp16, y = var_10198_to_fp16)[name = tensor("valid_len_cast_fp16")]; + tensor valid_mask_cast_fp16 = less(x = k_positions_1_promoted_to_fp16, y = valid_len_cast_fp16)[name = tensor("valid_mask_cast_fp16")]; + tensor causal_mask_cast_fp16 = less_equal(x = k_positions_1_promoted_to_fp16, y = var_10181_cast_fp16)[name = tensor("causal_mask_cast_fp16")]; + tensor attn_mask_93 = logical_and(x = valid_mask_cast_fp16, y = causal_mask_cast_fp16)[name = tensor("attn_mask_93")]; + tensor attn_mask_axes_0 = const()[name = tensor("attn_mask_axes_0"), val = tensor([1])]; + tensor attn_mask = expand_dims(axes = attn_mask_axes_0, x = attn_mask_93)[name = tensor("attn_mask")]; + tensor var_10216_transpose_x_0 = const()[name = tensor("op_10216_transpose_x_0"), val = tensor(false)]; + tensor var_10216_transpose_y_0 = const()[name = tensor("op_10216_transpose_y_0"), val = tensor(false)]; + tensor transpose_118_perm_0 = const()[name = tensor("transpose_118_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_119_perm_0 = const()[name = tensor("transpose_119_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_119 = transpose(perm = transpose_119_perm_0, x = keys_141_cast_fp16)[name = tensor("transpose_121")]; + tensor transpose_118 = transpose(perm = transpose_118_perm_0, x = q_141_cast_fp16)[name = tensor("transpose_122")]; + tensor var_10216_cast_fp16 = matmul(transpose_x = var_10216_transpose_x_0, transpose_y = var_10216_transpose_y_0, x = transpose_118, y = transpose_119)[name = tensor("op_10216_cast_fp16")]; + tensor var_10210_to_fp16 = const()[name = tensor("op_10210_to_fp16"), val = tensor([0x1p-3])]; + tensor attn_weights_139_cast_fp16 = mul(x = var_10216_cast_fp16, y = var_10210_to_fp16)[name = tensor("attn_weights_139_cast_fp16")]; + tensor var_10218 = logical_not(x = attn_mask)[name = tensor("op_10218")]; + tensor var_10219_to_fp16 = const()[name = tensor("op_10219_to_fp16"), val = tensor(-inf)]; + tensor attn_weights_141_cast_fp16 = select(a = var_10219_to_fp16, b = attn_weights_139_cast_fp16, cond = var_10218)[name = tensor("attn_weights_141_cast_fp16")]; + tensor var_10221 = const()[name = tensor("op_10221"), val = tensor(-1)]; + tensor attn_weights_cast_fp16 = softmax(axis = var_10221, x = attn_weights_141_cast_fp16)[name = tensor("attn_weights_cast_fp16")]; + tensor attn_output_transpose_x_0 = const()[name = tensor("attn_output_transpose_x_0"), val = tensor(false)]; + tensor attn_output_transpose_y_0 = const()[name = tensor("attn_output_transpose_y_0"), val = tensor(false)]; + tensor values_cast_fp16 = transpose(perm = var_10167, x = values_141_cast_fp16)[name = tensor("transpose_123")]; + tensor attn_output_cast_fp16 = matmul(transpose_x = attn_output_transpose_x_0, transpose_y = attn_output_transpose_y_0, x = attn_weights_cast_fp16, y = values_cast_fp16)[name = tensor("attn_output_cast_fp16")]; + tensor var_10229 = const()[name = tensor("op_10229"), val = tensor([0, 2, 1, 3])]; + tensor var_10232 = const()[name = tensor("op_10232"), val = tensor([1, 1, 1024])]; + tensor var_10230_cast_fp16 = transpose(perm = var_10229, x = attn_output_cast_fp16)[name = tensor("transpose_120")]; + tensor input_235_cast_fp16 = reshape(shape = var_10232, x = var_10230_cast_fp16)[name = tensor("input_235_cast_fp16")]; + tensor linear_94_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = attn23_out_proj_weight_cast_fp16, x = input_235_cast_fp16)[name = tensor("linear_94_cast_fp16")]; + tensor var_10238_to_fp16 = const()[name = tensor("op_10238_to_fp16"), val = tensor(0x1p+0)]; + tensor var_10239_cast_fp16 = add(x = position23_to_fp16, y = var_10238_to_fp16)[name = tensor("op_10239_cast_fp16")]; + tensor var_10239_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_10239_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor input_237_cast_fp16 = add(x = input_233_cast_fp16, y = linear_94_cast_fp16)[name = tensor("input_237_cast_fp16")]; + tensor input_239_axes_0 = const()[name = tensor("input_239_axes_0"), val = tensor([-1])]; + tensor norm23_2_weight_to_fp16 = const()[name = tensor("norm23_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(303786496)))]; + tensor norm23_2_bias_to_fp16 = const()[name = tensor("norm23_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(303788608)))]; + tensor var_10243_to_fp16 = const()[name = tensor("op_10243_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_239_cast_fp16 = layer_norm(axes = input_239_axes_0, beta = norm23_2_bias_to_fp16, epsilon = var_10243_to_fp16, gamma = norm23_2_weight_to_fp16, x = input_237_cast_fp16)[name = tensor("input_239_cast_fp16")]; + tensor linear_95_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = linear23_1_weight_cast_fp16, x = input_239_cast_fp16)[name = tensor("linear_95_cast_fp16")]; + tensor input_241_mode_0 = const()[name = tensor("input_241_mode_0"), val = tensor("EXACT")]; + tensor input_241_cast_fp16 = gelu(mode = input_241_mode_0, x = linear_95_cast_fp16)[name = tensor("input_241_cast_fp16")]; + tensor linear_96_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = linear23_2_weight_cast_fp16, x = input_241_cast_fp16)[name = tensor("linear_96_cast_fp16")]; + tensor input_243_cast_fp16 = add(x = input_237_cast_fp16, y = linear_96_cast_fp16)[name = tensor("input_243_cast_fp16")]; + tensor input_axes_0 = const()[name = tensor("input_axes_0"), val = tensor([-1])]; + tensor out_norm_weight_to_fp16 = const()[name = tensor("out_norm_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(303790720)))]; + tensor out_norm_bias_to_fp16 = const()[name = tensor("out_norm_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(303792832)))]; + tensor var_10260_to_fp16 = const()[name = tensor("op_10260_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_cast_fp16 = layer_norm(axes = input_axes_0, beta = out_norm_bias_to_fp16, epsilon = var_10260_to_fp16, gamma = out_norm_weight_to_fp16, x = input_243_cast_fp16)[name = tensor("input_cast_fp16")]; + tensor input_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("input_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor out_eos_weight_to_fp16 = const()[name = tensor("out_eos_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(303794944)))]; + tensor out_eos_bias_to_fp16 = const()[name = tensor("out_eos_bias_to_fp16"), val = tensor([-0x1.3ep-3])]; + tensor linear_97_cast_fp16 = linear(bias = out_eos_bias_to_fp16, weight = out_eos_weight_to_fp16, x = input_cast_fp16)[name = tensor("linear_97_cast_fp16")]; + tensor linear_97_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("linear_97_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor var_10268 = cast(dtype = linear_97_cast_fp16_to_fp32_dtype_0, x = linear_97_cast_fp16)[name = tensor("cast_432")]; + tensor input = cast(dtype = input_cast_fp16_to_fp32_dtype_0, x = input_cast_fp16)[name = tensor("cast_433")]; + tensor var_10239 = cast(dtype = var_10239_cast_fp16_to_fp32_dtype_0, x = var_10239_cast_fp16)[name = tensor("cast_434")]; + tensor new_cache_internal_tensor_assign_2 = cast(dtype = reshape_239_cast_fp16_to_fp32_dtype_0, x = reshape_239_cast_fp16)[name = tensor("cast_435")]; + tensor var_9820 = cast(dtype = var_9820_cast_fp16_to_fp32_dtype_0, x = var_9820_cast_fp16)[name = tensor("cast_439")]; + tensor new_cache_45_internal_tensor_assign_2 = cast(dtype = reshape_229_cast_fp16_to_fp32_dtype_0, x = reshape_229_cast_fp16)[name = tensor("cast_440")]; + tensor var_9401 = cast(dtype = var_9401_cast_fp16_to_fp32_dtype_0, x = var_9401_cast_fp16)[name = tensor("cast_444")]; + tensor new_cache_43_internal_tensor_assign_2 = cast(dtype = reshape_219_cast_fp16_to_fp32_dtype_0, x = reshape_219_cast_fp16)[name = tensor("cast_445")]; + tensor var_8982 = cast(dtype = var_8982_cast_fp16_to_fp32_dtype_0, x = var_8982_cast_fp16)[name = tensor("cast_449")]; + tensor new_cache_41_internal_tensor_assign_2 = cast(dtype = reshape_209_cast_fp16_to_fp32_dtype_0, x = reshape_209_cast_fp16)[name = tensor("cast_450")]; + tensor var_8563 = cast(dtype = var_8563_cast_fp16_to_fp32_dtype_0, x = var_8563_cast_fp16)[name = tensor("cast_454")]; + tensor new_cache_39_internal_tensor_assign_2 = cast(dtype = reshape_199_cast_fp16_to_fp32_dtype_0, x = reshape_199_cast_fp16)[name = tensor("cast_455")]; + tensor var_8144 = cast(dtype = var_8144_cast_fp16_to_fp32_dtype_0, x = var_8144_cast_fp16)[name = tensor("cast_459")]; + tensor new_cache_37_internal_tensor_assign_2 = cast(dtype = reshape_189_cast_fp16_to_fp32_dtype_0, x = reshape_189_cast_fp16)[name = tensor("cast_460")]; + tensor var_7725 = cast(dtype = var_7725_cast_fp16_to_fp32_dtype_0, x = var_7725_cast_fp16)[name = tensor("cast_464")]; + tensor new_cache_35_internal_tensor_assign_2 = cast(dtype = reshape_179_cast_fp16_to_fp32_dtype_0, x = reshape_179_cast_fp16)[name = tensor("cast_465")]; + tensor var_7306 = cast(dtype = var_7306_cast_fp16_to_fp32_dtype_0, x = var_7306_cast_fp16)[name = tensor("cast_469")]; + tensor new_cache_33_internal_tensor_assign_2 = cast(dtype = reshape_169_cast_fp16_to_fp32_dtype_0, x = reshape_169_cast_fp16)[name = tensor("cast_470")]; + tensor var_6887 = cast(dtype = var_6887_cast_fp16_to_fp32_dtype_0, x = var_6887_cast_fp16)[name = tensor("cast_474")]; + tensor new_cache_31_internal_tensor_assign_2 = cast(dtype = reshape_159_cast_fp16_to_fp32_dtype_0, x = reshape_159_cast_fp16)[name = tensor("cast_475")]; + tensor var_6468 = cast(dtype = var_6468_cast_fp16_to_fp32_dtype_0, x = var_6468_cast_fp16)[name = tensor("cast_479")]; + tensor new_cache_29_internal_tensor_assign_2 = cast(dtype = reshape_149_cast_fp16_to_fp32_dtype_0, x = reshape_149_cast_fp16)[name = tensor("cast_480")]; + tensor var_6049 = cast(dtype = var_6049_cast_fp16_to_fp32_dtype_0, x = var_6049_cast_fp16)[name = tensor("cast_484")]; + tensor new_cache_27_internal_tensor_assign_2 = cast(dtype = reshape_139_cast_fp16_to_fp32_dtype_0, x = reshape_139_cast_fp16)[name = tensor("cast_485")]; + tensor var_5630 = cast(dtype = var_5630_cast_fp16_to_fp32_dtype_0, x = var_5630_cast_fp16)[name = tensor("cast_489")]; + tensor new_cache_25_internal_tensor_assign_2 = cast(dtype = reshape_129_cast_fp16_to_fp32_dtype_0, x = reshape_129_cast_fp16)[name = tensor("cast_490")]; + tensor var_5211 = cast(dtype = var_5211_cast_fp16_to_fp32_dtype_0, x = var_5211_cast_fp16)[name = tensor("cast_494")]; + tensor new_cache_23_internal_tensor_assign_2 = cast(dtype = reshape_119_cast_fp16_to_fp32_dtype_0, x = reshape_119_cast_fp16)[name = tensor("cast_495")]; + tensor var_4792 = cast(dtype = var_4792_cast_fp16_to_fp32_dtype_0, x = var_4792_cast_fp16)[name = tensor("cast_499")]; + tensor new_cache_21_internal_tensor_assign_2 = cast(dtype = reshape_109_cast_fp16_to_fp32_dtype_0, x = reshape_109_cast_fp16)[name = tensor("cast_500")]; + tensor var_4373 = cast(dtype = var_4373_cast_fp16_to_fp32_dtype_0, x = var_4373_cast_fp16)[name = tensor("cast_504")]; + tensor new_cache_19_internal_tensor_assign_2 = cast(dtype = reshape_99_cast_fp16_to_fp32_dtype_0, x = reshape_99_cast_fp16)[name = tensor("cast_505")]; + tensor var_3954 = cast(dtype = var_3954_cast_fp16_to_fp32_dtype_0, x = var_3954_cast_fp16)[name = tensor("cast_509")]; + tensor new_cache_17_internal_tensor_assign_2 = cast(dtype = reshape_89_cast_fp16_to_fp32_dtype_0, x = reshape_89_cast_fp16)[name = tensor("cast_510")]; + tensor var_3535 = cast(dtype = var_3535_cast_fp16_to_fp32_dtype_0, x = var_3535_cast_fp16)[name = tensor("cast_514")]; + tensor new_cache_15_internal_tensor_assign_2 = cast(dtype = reshape_79_cast_fp16_to_fp32_dtype_0, x = reshape_79_cast_fp16)[name = tensor("cast_515")]; + tensor var_3116 = cast(dtype = var_3116_cast_fp16_to_fp32_dtype_0, x = var_3116_cast_fp16)[name = tensor("cast_519")]; + tensor new_cache_13_internal_tensor_assign_2 = cast(dtype = reshape_69_cast_fp16_to_fp32_dtype_0, x = reshape_69_cast_fp16)[name = tensor("cast_520")]; + tensor var_2697 = cast(dtype = var_2697_cast_fp16_to_fp32_dtype_0, x = var_2697_cast_fp16)[name = tensor("cast_524")]; + tensor new_cache_11_internal_tensor_assign_2 = cast(dtype = reshape_59_cast_fp16_to_fp32_dtype_0, x = reshape_59_cast_fp16)[name = tensor("cast_525")]; + tensor var_2278 = cast(dtype = var_2278_cast_fp16_to_fp32_dtype_0, x = var_2278_cast_fp16)[name = tensor("cast_529")]; + tensor new_cache_9_internal_tensor_assign_2 = cast(dtype = reshape_49_cast_fp16_to_fp32_dtype_0, x = reshape_49_cast_fp16)[name = tensor("cast_530")]; + tensor var_1859 = cast(dtype = var_1859_cast_fp16_to_fp32_dtype_0, x = var_1859_cast_fp16)[name = tensor("cast_534")]; + tensor new_cache_7_internal_tensor_assign_2 = cast(dtype = reshape_39_cast_fp16_to_fp32_dtype_0, x = reshape_39_cast_fp16)[name = tensor("cast_535")]; + tensor var_1440 = cast(dtype = var_1440_cast_fp16_to_fp32_dtype_0, x = var_1440_cast_fp16)[name = tensor("cast_539")]; + tensor new_cache_5_internal_tensor_assign_2 = cast(dtype = reshape_29_cast_fp16_to_fp32_dtype_0, x = reshape_29_cast_fp16)[name = tensor("cast_540")]; + tensor var_1021 = cast(dtype = var_1021_cast_fp16_to_fp32_dtype_0, x = var_1021_cast_fp16)[name = tensor("cast_544")]; + tensor new_cache_3_internal_tensor_assign_2 = cast(dtype = reshape_19_cast_fp16_to_fp32_dtype_0, x = reshape_19_cast_fp16)[name = tensor("cast_545")]; + tensor var_602 = cast(dtype = var_602_cast_fp16_to_fp32_dtype_0, x = var_602_cast_fp16)[name = tensor("cast_549")]; + tensor new_cache_1_internal_tensor_assign_2 = cast(dtype = reshape_9_cast_fp16_to_fp32_dtype_0, x = reshape_9_cast_fp16)[name = tensor("cast_550")]; + } -> (input, var_10268, new_cache_1_internal_tensor_assign_2, var_602, new_cache_3_internal_tensor_assign_2, var_1021, new_cache_5_internal_tensor_assign_2, var_1440, new_cache_7_internal_tensor_assign_2, var_1859, new_cache_9_internal_tensor_assign_2, var_2278, new_cache_11_internal_tensor_assign_2, var_2697, new_cache_13_internal_tensor_assign_2, var_3116, new_cache_15_internal_tensor_assign_2, var_3535, new_cache_17_internal_tensor_assign_2, var_3954, new_cache_19_internal_tensor_assign_2, var_4373, new_cache_21_internal_tensor_assign_2, var_4792, new_cache_23_internal_tensor_assign_2, var_5211, new_cache_25_internal_tensor_assign_2, var_5630, new_cache_27_internal_tensor_assign_2, var_6049, new_cache_29_internal_tensor_assign_2, var_6468, new_cache_31_internal_tensor_assign_2, var_6887, new_cache_33_internal_tensor_assign_2, var_7306, new_cache_35_internal_tensor_assign_2, var_7725, new_cache_37_internal_tensor_assign_2, var_8144, new_cache_39_internal_tensor_assign_2, var_8563, new_cache_41_internal_tensor_assign_2, var_8982, new_cache_43_internal_tensor_assign_2, var_9401, new_cache_45_internal_tensor_assign_2, var_9820, new_cache_internal_tensor_assign_2, var_10239); +} \ No newline at end of file diff --git a/v2.1/german_24l/flowlm_stepv2.mlmodelc/weights/weight.bin b/v2.1/german_24l/flowlm_stepv2.mlmodelc/weights/weight.bin new file mode 100644 index 0000000000000000000000000000000000000000..bbc028e207947a93dce43fc6f726ecf2e46906cd --- /dev/null +++ b/v2.1/german_24l/flowlm_stepv2.mlmodelc/weights/weight.bin @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:c02147cafc4ac29905ef0d55ed77b57e0613cfb33cec69502f55c15a4768a4ce +size 303797056 diff --git a/v2.1/german_24l/flowlm_stepv2.mlpackage/Data/com.apple.CoreML/model.mlmodel b/v2.1/german_24l/flowlm_stepv2.mlpackage/Data/com.apple.CoreML/model.mlmodel new file mode 100644 index 0000000000000000000000000000000000000000..07b44dbe53ef7a3aca48560a271207f00f5db223 --- /dev/null +++ b/v2.1/german_24l/flowlm_stepv2.mlpackage/Data/com.apple.CoreML/model.mlmodel @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:4554810805b9a0524a371c50ea315c11cf5bb35c29aca67322893e12edcf6b0f +size 843748 diff --git a/v2.1/german_24l/flowlm_stepv2.mlpackage/Data/com.apple.CoreML/weights/weight.bin b/v2.1/german_24l/flowlm_stepv2.mlpackage/Data/com.apple.CoreML/weights/weight.bin new file mode 100644 index 0000000000000000000000000000000000000000..bbc028e207947a93dce43fc6f726ecf2e46906cd --- /dev/null +++ b/v2.1/german_24l/flowlm_stepv2.mlpackage/Data/com.apple.CoreML/weights/weight.bin @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:c02147cafc4ac29905ef0d55ed77b57e0613cfb33cec69502f55c15a4768a4ce +size 303797056 diff --git a/v2.1/german_24l/flowlm_stepv2.mlpackage/Manifest.json b/v2.1/german_24l/flowlm_stepv2.mlpackage/Manifest.json new file mode 100644 index 0000000000000000000000000000000000000000..f6bdb7bb4510505417075673820a229e3c9b4613 --- /dev/null +++ b/v2.1/german_24l/flowlm_stepv2.mlpackage/Manifest.json @@ -0,0 +1,18 @@ +{ + "fileFormatVersion": "1.0.0", + "itemInfoEntries": { + "2B60A968-B168-42C7-BABE-7FA9724C0632": { + "author": "com.apple.CoreML", + "description": "CoreML Model Weights", + "name": "weights", + "path": "com.apple.CoreML/weights" + }, + "A7D0F82D-34FF-48DB-BAF0-CEB7A690C3D6": { + "author": "com.apple.CoreML", + "description": "CoreML Model Specification", + "name": "model.mlmodel", + "path": "com.apple.CoreML/model.mlmodel" + } + }, + "rootModelIdentifier": "A7D0F82D-34FF-48DB-BAF0-CEB7A690C3D6" +} diff --git a/v2.1/german_24l/manifest.json b/v2.1/german_24l/manifest.json new file mode 100644 index 0000000000000000000000000000000000000000..fc7da98920126d1261a9c47517b1303db1599db7 --- /dev/null +++ b/v2.1/german_24l/manifest.json @@ -0,0 +1,27 @@ +{ + "version": "2.1", + "base": "v2", + "kind": "optimization-reconvert", + "self_contained": true, + "note": "Same weights as v2. Re-converted for speed: fused flow decoder (100% ANE), one-shot cond prefill, fp16 flowlm. NOT a finetune.", + "new": [ + "flow_decoder_fused", + "cond_prefill", + "flowlm_step" + ], + "copied_from_v2": [ + "flowlm_stepv2.mlpackage", + "flowlm_stepv2.mlmodelc", + "mimi_decoder.mlpackage", + "mimi_decoder.mlmodelc", + "constants", + "constants_bin" + ], + "compute_units": { + "flow_decoder_fused": "all (ANE)", + "cond_prefill": "all (GPU)", + "flowlm_step": "all (GPU, fp16)", + "flowlm_stepv2": "cpuAndGpu (GPU, int8 \u2014 fastest flowlm)", + "mimi_decoder": "cpuOnly" + } +} diff --git a/v2.1/german_24l/mimi_decoder.mlmodelc/analytics/coremldata.bin b/v2.1/german_24l/mimi_decoder.mlmodelc/analytics/coremldata.bin new file mode 100644 index 0000000000000000000000000000000000000000..dea8904b584abd8dbd4c8cced80b283a1ecbb42e --- /dev/null +++ b/v2.1/german_24l/mimi_decoder.mlmodelc/analytics/coremldata.bin @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:9d909b4f10c950569ffac44ec612cd4b8248ae220f79ec8621f06953a7edadb3 +size 243 diff --git a/v2.1/german_24l/mimi_decoder.mlmodelc/coremldata.bin b/v2.1/german_24l/mimi_decoder.mlmodelc/coremldata.bin new file mode 100644 index 0000000000000000000000000000000000000000..76546b8da4a2a86e85df21fa619ee0286400f2f7 --- /dev/null +++ b/v2.1/german_24l/mimi_decoder.mlmodelc/coremldata.bin @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:2fb9d7ab308ebf86a14e5cf3841c34a9c6e5716d2be56301f35609261ce54f2e +size 1777 diff --git a/v2.1/german_24l/mimi_decoder.mlmodelc/metadata.json b/v2.1/german_24l/mimi_decoder.mlmodelc/metadata.json new file mode 100644 index 0000000000000000000000000000000000000000..f66f9e7ff5be2ffbc44c8300f3501f67a77afaf6 --- /dev/null +++ b/v2.1/german_24l/mimi_decoder.mlmodelc/metadata.json @@ -0,0 +1,573 @@ +[ + { + "metadataOutputVersion" : "3.0", + "storagePrecision" : "Float32", + "outputSchema" : [ + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1 × 1 × 1920)", + "shortDescription" : "", + "shape" : "[1, 1, 1920]", + "name" : "var_724", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1 × 512 × 16)", + "shortDescription" : "", + "shape" : "[1, 512, 16]", + "name" : "var_77", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 2 × 1 × 256 × 8 × 64)", + "shortDescription" : "", + "shape" : "[2, 1, 256, 8, 64]", + "name" : "var_210", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1)", + "shortDescription" : "", + "shape" : "[1]", + "name" : "var_741", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 2 × 1 × 256 × 8 × 64)", + "shortDescription" : "", + "shape" : "[2, 1, 256, 8, 64]", + "name" : "var_400", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1)", + "shortDescription" : "", + "shape" : "[1]", + "name" : "var_744", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1 × 512 × 6)", + "shortDescription" : "", + "shape" : "[1, 512, 6]", + "name" : "var_542", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1)", + "shortDescription" : "", + "shape" : "[1]", + "name" : "conv0_first", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1 × 256 × 6)", + "shortDescription" : "", + "shape" : "[1, 256, 6]", + "name" : "var_565", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1 × 256 × 2)", + "shortDescription" : "", + "shape" : "[1, 256, 2]", + "name" : "var_585", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1)", + "shortDescription" : "", + "shape" : "[1]", + "name" : "res0_conv0_first", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1 × 128 × 0)", + "shortDescription" : "", + "shape" : "[1, 128, 0]", + "name" : "res0_conv1_prev", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1)", + "shortDescription" : "", + "shape" : "[1]", + "name" : "res0_conv1_first", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1 × 128 × 5)", + "shortDescription" : "", + "shape" : "[1, 128, 5]", + "name" : "var_621", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1 × 128 × 2)", + "shortDescription" : "", + "shape" : "[1, 128, 2]", + "name" : "var_641", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1)", + "shortDescription" : "", + "shape" : "[1]", + "name" : "res1_conv0_first", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1 × 64 × 0)", + "shortDescription" : "", + "shape" : "[1, 64, 0]", + "name" : "res1_conv1_prev", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1)", + "shortDescription" : "", + "shape" : "[1]", + "name" : "res1_conv1_first", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1 × 64 × 4)", + "shortDescription" : "", + "shape" : "[1, 64, 4]", + "name" : "var_677", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1 × 64 × 2)", + "shortDescription" : "", + "shape" : "[1, 64, 2]", + "name" : "var_697", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1)", + "shortDescription" : "", + "shape" : "[1]", + "name" : "res2_conv0_first", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1 × 32 × 0)", + "shortDescription" : "", + "shape" : "[1, 32, 0]", + "name" : "res2_conv1_prev", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1)", + "shortDescription" : "", + "shape" : "[1]", + "name" : "res2_conv1_first", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1 × 64 × 2)", + "shortDescription" : "", + "shape" : "[1, 64, 2]", + "name" : "var_725", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1)", + "shortDescription" : "", + "shape" : "[1]", + "name" : "conv_final_first", + "type" : "MultiArray" + } + ], + "modelParameters" : [ + + ], + "specificationVersion" : 8, + "mlProgramOperationTypeHistogram" : { + "Ios17.notEqual" : 4, + "Tile" : 2, + "Ios17.lessEqual" : 2, + "Ios17.layerNorm" : 4, + "Stack" : 6, + "Ios17.greaterEqual" : 4, + "Ios17.scatterAlongAxis" : 4, + "Ios17.sub" : 17, + "Identity" : 11, + "Ios17.elu" : 10, + "Ios17.logicalNot" : 2, + "Ios17.sliceByIndex" : 35, + "Ios17.concat" : 9, + "Ios17.convTranspose" : 4, + "Ios17.floorDiv" : 4, + "Ios17.expandDims" : 7, + "Ios17.less" : 2, + "Ios17.mul" : 29, + "Ios17.transpose" : 10, + "Ios17.matmul" : 4, + "Ios16.gelu" : 2, + "Ios17.conv" : 9, + "Ios17.reshape" : 22, + "Split" : 2, + "Ios17.squeeze" : 6, + "Ios17.cast" : 6, + "Ios17.sin" : 2, + "Ios17.linear" : 8, + "Select" : 6, + "Ios16.softmax" : 2, + "Ios17.add" : 28, + "Ios17.cos" : 2, + "Ios17.logicalAnd" : 6 + }, + "computePrecision" : "Mixed (Float32, Int32)", + "isUpdatable" : "0", + "stateSchema" : [ + + ], + "availability" : { + "macOS" : "14.0", + "tvOS" : "17.0", + "visionOS" : "1.0", + "watchOS" : "10.0", + "iOS" : "17.0", + "macCatalyst" : "17.0" + }, + "modelType" : { + "name" : "MLModelType_mlProgram" + }, + "userDefinedMetadata" : { + "com.github.apple.coremltools.conversion_date" : "2026-04-24", + "com.github.apple.coremltools.source" : "torch==2.9.1", + "com.github.apple.coremltools.version" : "9.0", + "com.github.apple.coremltools.source_dialect" : "TorchScript" + }, + "inputSchema" : [ + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1 × 32)", + "shortDescription" : "", + "shape" : "[1, 32]", + "name" : "latent", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1 × 512 × 16)", + "shortDescription" : "", + "shape" : "[1, 512, 16]", + "name" : "upsample_partial", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 2 × 1 × 256 × 8 × 64)", + "shortDescription" : "", + "shape" : "[2, 1, 256, 8, 64]", + "name" : "attn0_cache", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1)", + "shortDescription" : "", + "shape" : "[1]", + "name" : "attn0_offset", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 2 × 1 × 256 × 8 × 64)", + "shortDescription" : "", + "shape" : "[2, 1, 256, 8, 64]", + "name" : "attn1_cache", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1)", + "shortDescription" : "", + "shape" : "[1]", + "name" : "attn1_offset", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1 × 512 × 6)", + "shortDescription" : "", + "shape" : "[1, 512, 6]", + "name" : "conv0_prev", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1)", + "shortDescription" : "", + "shape" : "[1]", + "name" : "conv0_first", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1 × 256 × 6)", + "shortDescription" : "", + "shape" : "[1, 256, 6]", + "name" : "convtr0_partial", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1 × 256 × 2)", + "shortDescription" : "", + "shape" : "[1, 256, 2]", + "name" : "res0_conv0_prev", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1)", + "shortDescription" : "", + "shape" : "[1]", + "name" : "res0_conv0_first", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1 × 128 × 0)", + "shortDescription" : "", + "shape" : "[1, 128, 0]", + "name" : "res0_conv1_prev", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1)", + "shortDescription" : "", + "shape" : "[1]", + "name" : "res0_conv1_first", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1 × 128 × 5)", + "shortDescription" : "", + "shape" : "[1, 128, 5]", + "name" : "convtr1_partial", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1 × 128 × 2)", + "shortDescription" : "", + "shape" : "[1, 128, 2]", + "name" : "res1_conv0_prev", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1)", + "shortDescription" : "", + "shape" : "[1]", + "name" : "res1_conv0_first", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1 × 64 × 0)", + "shortDescription" : "", + "shape" : "[1, 64, 0]", + "name" : "res1_conv1_prev", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1)", + "shortDescription" : "", + "shape" : "[1]", + "name" : "res1_conv1_first", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1 × 64 × 4)", + "shortDescription" : "", + "shape" : "[1, 64, 4]", + "name" : "convtr2_partial", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1 × 64 × 2)", + "shortDescription" : "", + "shape" : "[1, 64, 2]", + "name" : "res2_conv0_prev", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1)", + "shortDescription" : "", + "shape" : "[1]", + "name" : "res2_conv0_first", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1 × 32 × 0)", + "shortDescription" : "", + "shape" : "[1, 32, 0]", + "name" : "res2_conv1_prev", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1)", + "shortDescription" : "", + "shape" : "[1]", + "name" : "res2_conv1_first", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1 × 64 × 2)", + "shortDescription" : "", + "shape" : "[1, 64, 2]", + "name" : "conv_final_prev", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1)", + "shortDescription" : "", + "shape" : "[1]", + "name" : "conv_final_first", + "type" : "MultiArray" + } + ], + "generatedClassName" : "mimi_decoder", + "method" : "predict" + } +] \ No newline at end of file diff --git a/v2.1/german_24l/mimi_decoder.mlmodelc/model.mil b/v2.1/german_24l/mimi_decoder.mlmodelc/model.mil new file mode 100644 index 0000000000000000000000000000000000000000..85cbf2d02c036f2f26d39a5240feb64c94bd10c9 --- /dev/null +++ b/v2.1/german_24l/mimi_decoder.mlmodelc/model.mil @@ -0,0 +1,646 @@ +program(1.0) +[buildInfo = dict, tensor>({{"coremlc-component-MIL", "3520.4.1"}, {"coremlc-version", "3520.5.1"}, {"coremltools-component-torch", "2.9.1"}, {"coremltools-source-dialect", "TorchScript"}, {"coremltools-version", "9.0"}})] +{ + func main(tensor attn0_cache, tensor attn0_offset, tensor attn1_cache, tensor attn1_offset, tensor conv0_first, tensor conv0_prev, tensor conv_final_first, tensor conv_final_prev, tensor convtr0_partial, tensor convtr1_partial, tensor convtr2_partial, tensor latent, tensor res0_conv0_first, tensor res0_conv0_prev, tensor res0_conv1_first, tensor res0_conv1_prev, tensor res1_conv0_first, tensor res1_conv0_prev, tensor res1_conv1_first, tensor res1_conv1_prev, tensor res2_conv0_first, tensor res2_conv0_prev, tensor res2_conv1_first, tensor res2_conv1_prev, tensor upsample_partial) { + tensor emb_mean = const()[name = tensor("emb_mean"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(64)))]; + tensor emb_std = const()[name = tensor("emb_std"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(256)))]; + tensor mimi_quantizer_output_proj_weight = const()[name = tensor("mimi_quantizer_output_proj_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(448)))]; + tensor mimi_upsample_convtr_convtr_weight = const()[name = tensor("mimi_upsample_convtr_convtr_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(66048)))]; + tensor mimi_decoder_transformer_transformer_layers_0_norm1_bias = const()[name = tensor("mimi_decoder_transformer_transformer_layers_0_norm1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(131648)))]; + tensor mimi_decoder_transformer_transformer_layers_0_norm1_weight = const()[name = tensor("mimi_decoder_transformer_transformer_layers_0_norm1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(133760)))]; + tensor mimi_decoder_transformer_transformer_layers_0_self_attn_in_proj_weight = const()[name = tensor("mimi_decoder_transformer_transformer_layers_0_self_attn_in_proj_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(135872)))]; + tensor mimi_decoder_transformer_transformer_layers_0_self_attn_out_proj_weight = const()[name = tensor("mimi_decoder_transformer_transformer_layers_0_self_attn_out_proj_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(3281664)))]; + tensor mimi_decoder_transformer_transformer_layers_0_layer_scale_1_scale = const()[name = tensor("mimi_decoder_transformer_transformer_layers_0_layer_scale_1_scale"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(4330304)))]; + tensor mimi_decoder_transformer_transformer_layers_0_norm2_bias = const()[name = tensor("mimi_decoder_transformer_transformer_layers_0_norm2_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(4332416)))]; + tensor mimi_decoder_transformer_transformer_layers_0_norm2_weight = const()[name = tensor("mimi_decoder_transformer_transformer_layers_0_norm2_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(4334528)))]; + tensor mimi_decoder_transformer_transformer_layers_0_linear1_weight = const()[name = tensor("mimi_decoder_transformer_transformer_layers_0_linear1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(4336640)))]; + tensor mimi_decoder_transformer_transformer_layers_0_linear2_weight = const()[name = tensor("mimi_decoder_transformer_transformer_layers_0_linear2_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8531008)))]; + tensor mimi_decoder_transformer_transformer_layers_0_layer_scale_2_scale = const()[name = tensor("mimi_decoder_transformer_transformer_layers_0_layer_scale_2_scale"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(12725376)))]; + tensor mimi_decoder_transformer_transformer_layers_1_norm1_bias = const()[name = tensor("mimi_decoder_transformer_transformer_layers_1_norm1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(12727488)))]; + tensor mimi_decoder_transformer_transformer_layers_1_norm1_weight = const()[name = tensor("mimi_decoder_transformer_transformer_layers_1_norm1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(12729600)))]; + tensor mimi_decoder_transformer_transformer_layers_1_self_attn_in_proj_weight = const()[name = tensor("mimi_decoder_transformer_transformer_layers_1_self_attn_in_proj_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(12731712)))]; + tensor mimi_decoder_transformer_transformer_layers_1_self_attn_out_proj_weight = const()[name = tensor("mimi_decoder_transformer_transformer_layers_1_self_attn_out_proj_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15877504)))]; + tensor mimi_decoder_transformer_transformer_layers_1_layer_scale_1_scale = const()[name = tensor("mimi_decoder_transformer_transformer_layers_1_layer_scale_1_scale"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(16926144)))]; + tensor mimi_decoder_transformer_transformer_layers_1_norm2_bias = const()[name = tensor("mimi_decoder_transformer_transformer_layers_1_norm2_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(16928256)))]; + tensor mimi_decoder_transformer_transformer_layers_1_norm2_weight = const()[name = tensor("mimi_decoder_transformer_transformer_layers_1_norm2_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(16930368)))]; + tensor mimi_decoder_transformer_transformer_layers_1_linear1_weight = const()[name = tensor("mimi_decoder_transformer_transformer_layers_1_linear1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(16932480)))]; + tensor mimi_decoder_transformer_transformer_layers_1_linear2_weight = const()[name = tensor("mimi_decoder_transformer_transformer_layers_1_linear2_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21126848)))]; + tensor mimi_decoder_transformer_transformer_layers_1_layer_scale_2_scale = const()[name = tensor("mimi_decoder_transformer_transformer_layers_1_layer_scale_2_scale"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(25321216)))]; + tensor mimi_decoder_model_0_conv_bias = const()[name = tensor("mimi_decoder_model_0_conv_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(25323328)))]; + tensor mimi_decoder_model_0_conv_weight = const()[name = tensor("mimi_decoder_model_0_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(25325440)))]; + tensor mimi_decoder_model_2_convtr_bias = const()[name = tensor("mimi_decoder_model_2_convtr_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(32665536)))]; + tensor mimi_decoder_model_2_convtr_weight = const()[name = tensor("mimi_decoder_model_2_convtr_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(32666624)))]; + tensor mimi_decoder_model_3_block_1_conv_bias = const()[name = tensor("mimi_decoder_model_3_block_1_conv_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(38958144)))]; + tensor mimi_decoder_model_3_block_1_conv_weight = const()[name = tensor("mimi_decoder_model_3_block_1_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(38958720)))]; + tensor mimi_decoder_model_3_block_3_conv_bias = const()[name = tensor("mimi_decoder_model_3_block_3_conv_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(39352000)))]; + tensor mimi_decoder_model_3_block_3_conv_weight = const()[name = tensor("mimi_decoder_model_3_block_3_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(39353088)))]; + tensor mimi_decoder_model_5_convtr_bias = const()[name = tensor("mimi_decoder_model_5_convtr_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(39484224)))]; + tensor mimi_decoder_model_5_convtr_weight = const()[name = tensor("mimi_decoder_model_5_convtr_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(39484800)))]; + tensor mimi_decoder_model_6_block_1_conv_bias = const()[name = tensor("mimi_decoder_model_6_block_1_conv_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(40795584)))]; + tensor mimi_decoder_model_6_block_1_conv_weight = const()[name = tensor("mimi_decoder_model_6_block_1_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(40795904)))]; + tensor mimi_decoder_model_6_block_3_conv_bias = const()[name = tensor("mimi_decoder_model_6_block_3_conv_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(40894272)))]; + tensor mimi_decoder_model_6_block_3_conv_weight = const()[name = tensor("mimi_decoder_model_6_block_3_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(40894848)))]; + tensor mimi_decoder_model_8_convtr_bias = const()[name = tensor("mimi_decoder_model_8_convtr_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(40927680)))]; + tensor mimi_decoder_model_8_convtr_weight = const()[name = tensor("mimi_decoder_model_8_convtr_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(40928000)))]; + tensor mimi_decoder_model_9_block_1_conv_bias = const()[name = tensor("mimi_decoder_model_9_block_1_conv_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(41190208)))]; + tensor mimi_decoder_model_9_block_1_conv_weight = const()[name = tensor("mimi_decoder_model_9_block_1_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(41190400)))]; + tensor mimi_decoder_model_9_block_3_conv_bias = const()[name = tensor("mimi_decoder_model_9_block_3_conv_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(41215040)))]; + tensor mimi_decoder_model_9_block_3_conv_weight = const()[name = tensor("mimi_decoder_model_9_block_3_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(41215360)))]; + tensor mimi_decoder_model_11_conv_bias = const()[name = tensor("mimi_decoder_model_11_conv_bias"), val = tensor([-0x1.8p-13])]; + tensor mimi_decoder_model_11_conv_weight = const()[name = tensor("mimi_decoder_model_11_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(41223616)))]; + tensor var_38 = mul(x = latent, y = emb_std)[name = tensor("op_38")]; + tensor denorm = add(x = var_38, y = emb_mean)[name = tensor("denorm")]; + tensor input_1_axes_0 = const()[name = tensor("input_1_axes_0"), val = tensor([-1])]; + tensor input_1 = expand_dims(axes = input_1_axes_0, x = denorm)[name = tensor("input_1")]; + tensor x_1_pad_type_0 = const()[name = tensor("x_1_pad_type_0"), val = tensor("valid")]; + tensor x_1_strides_0 = const()[name = tensor("x_1_strides_0"), val = tensor([1])]; + tensor x_1_pad_0 = const()[name = tensor("x_1_pad_0"), val = tensor([0, 0])]; + tensor x_1_dilations_0 = const()[name = tensor("x_1_dilations_0"), val = tensor([1])]; + tensor x_1_groups_0 = const()[name = tensor("x_1_groups_0"), val = tensor(1)]; + tensor x_1 = conv(dilations = x_1_dilations_0, groups = x_1_groups_0, pad = x_1_pad_0, pad_type = x_1_pad_type_0, strides = x_1_strides_0, weight = mimi_quantizer_output_proj_weight, x = input_1)[name = tensor("x_1")]; + tensor var_62 = const()[name = tensor("op_62"), val = tensor(-1)]; + tensor y_1_pad_type_0 = const()[name = tensor("y_1_pad_type_0"), val = tensor("valid")]; + tensor y_1_strides_0 = const()[name = tensor("y_1_strides_0"), val = tensor([16])]; + tensor y_1_groups_0 = const()[name = tensor("y_1_groups_0"), val = tensor(512)]; + tensor y_1_pad_0 = const()[name = tensor("y_1_pad_0"), val = tensor([0, 0])]; + tensor y_1_dilations_0 = const()[name = tensor("y_1_dilations_0"), val = tensor([1])]; + tensor y_1_has_output_shape_output_shape_0 = const()[name = tensor("y_1_has_output_shape_output_shape_0"), val = tensor([1, 512, 32])]; + tensor y_1_has_output_shape = conv_transpose(dilations = y_1_dilations_0, groups = y_1_groups_0, output_shape = y_1_has_output_shape_output_shape_0, pad = y_1_pad_0, pad_type = y_1_pad_type_0, strides = y_1_strides_0, weight = mimi_upsample_convtr_convtr_weight, x = x_1)[name = tensor("y_1_has_output_shape")]; + tensor var_72_begin_0 = const()[name = tensor("op_72_begin_0"), val = tensor([0, 0, 0])]; + tensor var_72_end_0 = const()[name = tensor("op_72_end_0"), val = tensor([1, 512, 16])]; + tensor var_72_end_mask_0 = const()[name = tensor("op_72_end_mask_0"), val = tensor([true, true, false])]; + tensor var_72 = slice_by_index(begin = var_72_begin_0, end = var_72_end_0, end_mask = var_72_end_mask_0, x = y_1_has_output_shape)[name = tensor("op_72")]; + tensor var_73 = add(x = var_72, y = upsample_partial)[name = tensor("op_73")]; + tensor var_74_begin_0 = const()[name = tensor("op_74_begin_0"), val = tensor([0, 0, 16])]; + tensor var_74_end_0 = const()[name = tensor("op_74_end_0"), val = tensor([1, 512, 32])]; + tensor var_74_end_mask_0 = const()[name = tensor("op_74_end_mask_0"), val = tensor([true, true, true])]; + tensor var_74 = slice_by_index(begin = var_74_begin_0, end = var_74_end_0, end_mask = var_74_end_mask_0, x = y_1_has_output_shape)[name = tensor("op_74")]; + tensor y_3_interleave_0 = const()[name = tensor("y_3_interleave_0"), val = tensor(false)]; + tensor y_3 = concat(axis = var_62, interleave = y_3_interleave_0, values = (var_73, var_74))[name = tensor("y_3")]; + tensor var_77_begin_0 = const()[name = tensor("op_77_begin_0"), val = tensor([0, 0, 16])]; + tensor var_77_end_0 = const()[name = tensor("op_77_end_0"), val = tensor([1, 512, 32])]; + tensor var_77_end_mask_0 = const()[name = tensor("op_77_end_mask_0"), val = tensor([true, true, true])]; + tensor var_77 = slice_by_index(begin = var_77_begin_0, end = var_77_end_0, end_mask = var_77_end_mask_0, x = y_3)[name = tensor("op_77")]; + tensor x_3_begin_0 = const()[name = tensor("x_3_begin_0"), val = tensor([0, 0, 0])]; + tensor x_3_end_0 = const()[name = tensor("x_3_end_0"), val = tensor([1, 512, 16])]; + tensor x_3_end_mask_0 = const()[name = tensor("x_3_end_mask_0"), val = tensor([true, true, false])]; + tensor x_3 = slice_by_index(begin = x_3_begin_0, end = x_3_end_0, end_mask = x_3_end_mask_0, x = y_3)[name = tensor("x_3")]; + tensor var_86 = const()[name = tensor("op_86"), val = tensor(0)]; + tensor var_91 = const()[name = tensor("op_91"), val = tensor(-1)]; + tensor var_100 = const()[name = tensor("op_100"), val = tensor(-0x1.ff933cp+127)]; + tensor var_102 = const()[name = tensor("op_102"), val = tensor(0x1.4f8b58p-17)]; + tensor input_3_perm_0 = const()[name = tensor("input_3_perm_0"), val = tensor([0, 2, 1])]; + tensor query_1_axes_0 = const()[name = tensor("query_1_axes_0"), val = tensor([-1])]; + tensor input_3 = transpose(perm = input_3_perm_0, x = x_3)[name = tensor("transpose_19")]; + tensor query_1 = layer_norm(axes = query_1_axes_0, beta = mimi_decoder_transformer_transformer_layers_0_norm1_bias, epsilon = var_102, gamma = mimi_decoder_transformer_transformer_layers_0_norm1_weight, x = input_3)[name = tensor("query_1")]; + tensor linear_0_bias_0 = const()[name = tensor("linear_0_bias_0"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(41224448)))]; + tensor projected_1 = linear(bias = linear_0_bias_0, weight = mimi_decoder_transformer_transformer_layers_0_self_attn_in_proj_weight, x = query_1)[name = tensor("linear_0")]; + tensor var_130 = const()[name = tensor("op_130"), val = tensor([1, 16, 3, 8, 64])]; + tensor packed_1 = reshape(shape = var_130, x = projected_1)[name = tensor("packed_1")]; + tensor var_132_split_sizes_0 = const()[name = tensor("op_132_split_sizes_0"), val = tensor([1, 1, 1])]; + tensor var_132_axis_0 = const()[name = tensor("op_132_axis_0"), val = tensor(2)]; + tensor var_132_0, tensor var_132_1, tensor var_132_2 = split(axis = var_132_axis_0, split_sizes = var_132_split_sizes_0, x = packed_1)[name = tensor("op_132")]; + tensor squeeze_0_axes_0 = const()[name = tensor("squeeze_0_axes_0"), val = tensor([2])]; + tensor squeeze_0 = squeeze(axes = squeeze_0_axes_0, x = var_132_0)[name = tensor("squeeze_0")]; + tensor squeeze_1_axes_0 = const()[name = tensor("squeeze_1_axes_0"), val = tensor([2])]; + tensor squeeze_1 = squeeze(axes = squeeze_1_axes_0, x = var_132_1)[name = tensor("squeeze_1")]; + tensor squeeze_2_axes_0 = const()[name = tensor("squeeze_2_axes_0"), val = tensor([2])]; + tensor squeeze_2 = squeeze(axes = squeeze_2_axes_0, x = var_132_2)[name = tensor("squeeze_2")]; + tensor offset_3_begin_0 = const()[name = tensor("offset_3_begin_0"), val = tensor([0])]; + tensor offset_3_end_0 = const()[name = tensor("offset_3_end_0"), val = tensor([1])]; + tensor offset_3_end_mask_0 = const()[name = tensor("offset_3_end_mask_0"), val = tensor([false])]; + tensor offset_3_squeeze_mask_0 = const()[name = tensor("offset_3_squeeze_mask_0"), val = tensor([true])]; + tensor offset_3 = slice_by_index(begin = offset_3_begin_0, end = offset_3_end_0, end_mask = offset_3_end_mask_0, squeeze_mask = offset_3_squeeze_mask_0, x = attn0_offset)[name = tensor("offset_3")]; + tensor freqs_1 = const()[name = tensor("freqs_1"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(41230656)))]; + tensor ts_1_promoted = const()[name = tensor("ts_1_promoted"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(41230848)))]; + tensor ts_3 = add(x = ts_1_promoted, y = offset_3)[name = tensor("ts_3")]; + tensor var_148 = const()[name = tensor("op_148"), val = tensor([-1, 1, 1])]; + tensor ts_5 = reshape(shape = var_148, x = ts_3)[name = tensor("ts_5")]; + tensor var_150 = const()[name = tensor("op_150"), val = tensor([1, 16, 8, 32, 2])]; + tensor q_3 = reshape(shape = var_150, x = squeeze_0)[name = tensor("q_3")]; + tensor var_152 = const()[name = tensor("op_152"), val = tensor([1, 16, 8, 32, 2])]; + tensor k_3 = reshape(shape = var_152, x = squeeze_1)[name = tensor("k_3")]; + tensor var_154_begin_0 = const()[name = tensor("op_154_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_154_end_0 = const()[name = tensor("op_154_end_0"), val = tensor([1, 16, 8, 32, 1])]; + tensor var_154_end_mask_0 = const()[name = tensor("op_154_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_154_squeeze_mask_0 = const()[name = tensor("op_154_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_154 = slice_by_index(begin = var_154_begin_0, end = var_154_end_0, end_mask = var_154_end_mask_0, squeeze_mask = var_154_squeeze_mask_0, x = q_3)[name = tensor("op_154")]; + tensor var_156_begin_0 = const()[name = tensor("op_156_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_156_end_0 = const()[name = tensor("op_156_end_0"), val = tensor([1, 16, 8, 32, 2])]; + tensor var_156_end_mask_0 = const()[name = tensor("op_156_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_156_squeeze_mask_0 = const()[name = tensor("op_156_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_156 = slice_by_index(begin = var_156_begin_0, end = var_156_end_0, end_mask = var_156_end_mask_0, squeeze_mask = var_156_squeeze_mask_0, x = q_3)[name = tensor("op_156")]; + tensor var_158_begin_0 = const()[name = tensor("op_158_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_158_end_0 = const()[name = tensor("op_158_end_0"), val = tensor([1, 16, 8, 32, 1])]; + tensor var_158_end_mask_0 = const()[name = tensor("op_158_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_158_squeeze_mask_0 = const()[name = tensor("op_158_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_158 = slice_by_index(begin = var_158_begin_0, end = var_158_end_0, end_mask = var_158_end_mask_0, squeeze_mask = var_158_squeeze_mask_0, x = k_3)[name = tensor("op_158")]; + tensor var_160_begin_0 = const()[name = tensor("op_160_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_160_end_0 = const()[name = tensor("op_160_end_0"), val = tensor([1, 16, 8, 32, 2])]; + tensor var_160_end_mask_0 = const()[name = tensor("op_160_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_160_squeeze_mask_0 = const()[name = tensor("op_160_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_160 = slice_by_index(begin = var_160_begin_0, end = var_160_end_0, end_mask = var_160_end_mask_0, squeeze_mask = var_160_squeeze_mask_0, x = k_3)[name = tensor("op_160")]; + tensor var_162 = mul(x = freqs_1, y = ts_5)[name = tensor("op_162")]; + tensor rotr_1 = cos(x = var_162)[name = tensor("rotr_1")]; + tensor roti_1 = sin(x = var_162)[name = tensor("roti_1")]; + tensor var_166 = mul(x = var_154, y = rotr_1)[name = tensor("op_166")]; + tensor var_167 = mul(x = var_156, y = roti_1)[name = tensor("op_167")]; + tensor qor_1 = sub(x = var_166, y = var_167)[name = tensor("qor_1")]; + tensor var_169 = mul(x = var_154, y = roti_1)[name = tensor("op_169")]; + tensor var_170 = mul(x = var_156, y = rotr_1)[name = tensor("op_170")]; + tensor qoi_1 = add(x = var_169, y = var_170)[name = tensor("qoi_1")]; + tensor var_172 = mul(x = var_158, y = rotr_1)[name = tensor("op_172")]; + tensor var_173 = mul(x = var_160, y = roti_1)[name = tensor("op_173")]; + tensor kor_1 = sub(x = var_172, y = var_173)[name = tensor("kor_1")]; + tensor var_175 = mul(x = var_158, y = roti_1)[name = tensor("op_175")]; + tensor var_176 = mul(x = var_160, y = rotr_1)[name = tensor("op_176")]; + tensor koi_1 = add(x = var_175, y = var_176)[name = tensor("koi_1")]; + tensor qo_1_axis_0 = const()[name = tensor("qo_1_axis_0"), val = tensor(-1)]; + tensor qo_1 = stack(axis = qo_1_axis_0, values = (qor_1, qoi_1))[name = tensor("qo_1")]; + tensor ko_1_axis_0 = const()[name = tensor("ko_1_axis_0"), val = tensor(-1)]; + tensor ko_1 = stack(axis = ko_1_axis_0, values = (kor_1, koi_1))[name = tensor("ko_1")]; + tensor var_186 = const()[name = tensor("op_186"), val = tensor([1, 16, 8, 64])]; + tensor q_5 = reshape(shape = var_186, x = qo_1)[name = tensor("q_5")]; + tensor var_188 = const()[name = tensor("op_188"), val = tensor([1, 16, 8, 64])]; + tensor k_5 = reshape(shape = var_188, x = ko_1)[name = tensor("k_5")]; + tensor capacity_1 = const()[name = tensor("capacity_1"), val = tensor([256])]; + tensor var_193_dtype_0 = const()[name = tensor("op_193_dtype_0"), val = tensor("int32")]; + tensor var_194 = const()[name = tensor("op_194"), val = tensor([1, 1])]; + tensor var_193 = cast(dtype = var_193_dtype_0, x = attn0_offset)[name = tensor("cast_49")]; + tensor write_base_1 = reshape(shape = var_194, x = var_193)[name = tensor("write_base_1")]; + tensor write_range_1 = const()[name = tensor("write_range_1"), val = tensor([[0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15]])]; + tensor abs_idx_1 = add(x = write_base_1, y = write_range_1)[name = tensor("abs_idx_1")]; + tensor wrapped_1_div = floor_div(x = abs_idx_1, y = capacity_1)[name = tensor("wrapped_1_div")]; + tensor wrapped_1_div_scaled = mul(x = wrapped_1_div, y = capacity_1)[name = tensor("wrapped_1_div_scaled")]; + tensor wrapped_1 = sub(x = abs_idx_1, y = wrapped_1_div_scaled)[name = tensor("wrapped_1")]; + tensor var_201 = const()[name = tensor("op_201"), val = tensor([1, 16, 1, 1])]; + tensor var_202 = reshape(shape = var_201, x = wrapped_1)[name = tensor("op_202")]; + tensor write_indexes_1_reps_0 = const()[name = tensor("write_indexes_1_reps_0"), val = tensor([1, 1, 8, 64])]; + tensor write_indexes_1 = tile(reps = write_indexes_1_reps_0, x = var_202)[name = tensor("write_indexes_1")]; + tensor var_205_begin_0 = const()[name = tensor("op_205_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_205_end_0 = const()[name = tensor("op_205_end_0"), val = tensor([1, 1, 256, 8, 64])]; + tensor var_205_end_mask_0 = const()[name = tensor("op_205_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_205_squeeze_mask_0 = const()[name = tensor("op_205_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor var_205 = slice_by_index(begin = var_205_begin_0, end = var_205_end_0, end_mask = var_205_end_mask_0, squeeze_mask = var_205_squeeze_mask_0, x = attn0_cache)[name = tensor("op_205")]; + tensor new_k_cache_1_axis_0 = const()[name = tensor("new_k_cache_1_axis_0"), val = tensor(1)]; + tensor new_k_cache_1_mode_0 = const()[name = tensor("new_k_cache_1_mode_0"), val = tensor("update")]; + tensor new_k_cache_1_validate_indices_0 = const()[name = tensor("new_k_cache_1_validate_indices_0"), val = tensor(false)]; + tensor new_k_cache_1 = scatter_along_axis(axis = new_k_cache_1_axis_0, data = var_205, indices = write_indexes_1, mode = new_k_cache_1_mode_0, updates = k_5, validate_indices = new_k_cache_1_validate_indices_0)[name = tensor("new_k_cache_1")]; + tensor var_207_begin_0 = const()[name = tensor("op_207_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor var_207_end_0 = const()[name = tensor("op_207_end_0"), val = tensor([2, 1, 256, 8, 64])]; + tensor var_207_end_mask_0 = const()[name = tensor("op_207_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_207_squeeze_mask_0 = const()[name = tensor("op_207_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor var_207 = slice_by_index(begin = var_207_begin_0, end = var_207_end_0, end_mask = var_207_end_mask_0, squeeze_mask = var_207_squeeze_mask_0, x = attn0_cache)[name = tensor("op_207")]; + tensor new_v_cache_1_axis_0 = const()[name = tensor("new_v_cache_1_axis_0"), val = tensor(1)]; + tensor new_v_cache_1_mode_0 = const()[name = tensor("new_v_cache_1_mode_0"), val = tensor("update")]; + tensor new_v_cache_1_validate_indices_0 = const()[name = tensor("new_v_cache_1_validate_indices_0"), val = tensor(false)]; + tensor new_v_cache_1 = scatter_along_axis(axis = new_v_cache_1_axis_0, data = var_207, indices = write_indexes_1, mode = new_v_cache_1_mode_0, updates = squeeze_2, validate_indices = new_v_cache_1_validate_indices_0)[name = tensor("new_v_cache_1")]; + tensor var_210_axis_0 = const()[name = tensor("op_210_axis_0"), val = tensor(0)]; + tensor var_210 = stack(axis = var_210_axis_0, values = (new_k_cache_1, new_v_cache_1))[name = tensor("op_210")]; + tensor var_211 = not_equal(x = new_k_cache_1, y = new_k_cache_1)[name = tensor("op_211")]; + tensor var_212 = const()[name = tensor("op_212"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(41230976)))]; + tensor new_k_cache_3 = select(a = var_212, b = new_k_cache_1, cond = var_211)[name = tensor("new_k_cache_3")]; + tensor var_214 = not_equal(x = new_v_cache_1, y = new_v_cache_1)[name = tensor("op_214")]; + tensor new_v_cache_3 = select(a = var_212, b = new_v_cache_1, cond = var_214)[name = tensor("new_v_cache_3")]; + tensor var_219 = const()[name = tensor("op_219"), val = tensor([0, 2, 1, 3])]; + tensor var_221 = const()[name = tensor("op_221"), val = tensor([1, 1])]; + tensor var_222 = reshape(shape = var_221, x = attn0_offset)[name = tensor("op_222")]; + tensor var_224_promoted = const()[name = tensor("op_224_promoted"), val = tensor([0x1.ep+3])]; + tensor var_225 = add(x = var_222, y = var_224_promoted)[name = tensor("op_225")]; + tensor last_pos_1_dtype_0 = const()[name = tensor("last_pos_1_dtype_0"), val = tensor("int32")]; + tensor slot_idx_1 = const()[name = tensor("slot_idx_1"), val = tensor([[0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19, 20, 21, 22, 23, 24, 25, 26, 27, 28, 29, 30, 31, 32, 33, 34, 35, 36, 37, 38, 39, 40, 41, 42, 43, 44, 45, 46, 47, 48, 49, 50, 51, 52, 53, 54, 55, 56, 57, 58, 59, 60, 61, 62, 63, 64, 65, 66, 67, 68, 69, 70, 71, 72, 73, 74, 75, 76, 77, 78, 79, 80, 81, 82, 83, 84, 85, 86, 87, 88, 89, 90, 91, 92, 93, 94, 95, 96, 97, 98, 99, 100, 101, 102, 103, 104, 105, 106, 107, 108, 109, 110, 111, 112, 113, 114, 115, 116, 117, 118, 119, 120, 121, 122, 123, 124, 125, 126, 127, 128, 129, 130, 131, 132, 133, 134, 135, 136, 137, 138, 139, 140, 141, 142, 143, 144, 145, 146, 147, 148, 149, 150, 151, 152, 153, 154, 155, 156, 157, 158, 159, 160, 161, 162, 163, 164, 165, 166, 167, 168, 169, 170, 171, 172, 173, 174, 175, 176, 177, 178, 179, 180, 181, 182, 183, 184, 185, 186, 187, 188, 189, 190, 191, 192, 193, 194, 195, 196, 197, 198, 199, 200, 201, 202, 203, 204, 205, 206, 207, 208, 209, 210, 211, 212, 213, 214, 215, 216, 217, 218, 219, 220, 221, 222, 223, 224, 225, 226, 227, 228, 229, 230, 231, 232, 233, 234, 235, 236, 237, 238, 239, 240, 241, 242, 243, 244, 245, 246, 247, 248, 249, 250, 251, 252, 253, 254, 255]])]; + tensor last_pos_1 = cast(dtype = last_pos_1_dtype_0, x = var_225)[name = tensor("cast_48")]; + tensor diff_1 = sub(x = last_pos_1, y = slot_idx_1)[name = tensor("diff_1")]; + tensor var_231_div = floor_div(x = diff_1, y = capacity_1)[name = tensor("op_231_div")]; + tensor var_231_div_scaled = mul(x = var_231_div, y = capacity_1)[name = tensor("op_231_div_scaled")]; + tensor var_231 = sub(x = diff_1, y = var_231_div_scaled)[name = tensor("op_231")]; + tensor pos_k_1 = sub(x = last_pos_1, y = var_231)[name = tensor("pos_k_1")]; + tensor var_237_promoted = const()[name = tensor("op_237_promoted"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(41755328)))]; + tensor pos_q_1 = add(x = var_222, y = var_237_promoted)[name = tensor("pos_q_1")]; + tensor var_241_axes_0 = const()[name = tensor("op_241_axes_0"), val = tensor([2])]; + tensor var_241 = expand_dims(axes = var_241_axes_0, x = pos_q_1)[name = tensor("op_241")]; + tensor var_243_axes_0 = const()[name = tensor("op_243_axes_0"), val = tensor([1])]; + tensor var_243 = expand_dims(axes = var_243_axes_0, x = pos_k_1)[name = tensor("op_243")]; + tensor var_244_promoted_dtype_0 = const()[name = tensor("op_244_promoted_dtype_0"), val = tensor("fp32")]; + tensor var_244_promoted = cast(dtype = var_244_promoted_dtype_0, x = var_243)[name = tensor("cast_47")]; + tensor delta_1 = sub(x = var_241, y = var_244_promoted)[name = tensor("delta_1")]; + tensor valid_1 = greater_equal(x = var_243, y = var_86)[name = tensor("valid_1")]; + tensor var_253 = const()[name = tensor("op_253"), val = tensor([1, 1, 1])]; + tensor var_254 = reshape(shape = var_253, x = attn0_offset)[name = tensor("op_254")]; + tensor var_256_promoted = const()[name = tensor("op_256_promoted"), val = tensor([0x1.ep+3])]; + tensor var_257 = add(x = var_254, y = var_256_promoted)[name = tensor("op_257")]; + tensor var_258 = less_equal(x = var_244_promoted, y = var_257)[name = tensor("op_258")]; + tensor valid_3 = logical_and(x = valid_1, y = var_258)[name = tensor("valid_3")]; + tensor var_86_promoted = const()[name = tensor("op_86_promoted"), val = tensor(0x0p+0)]; + tensor var_260 = greater_equal(x = delta_1, y = var_86_promoted)[name = tensor("op_260")]; + tensor attn_mask_1 = logical_and(x = valid_3, y = var_260)[name = tensor("attn_mask_1")]; + tensor var_98_promoted = const()[name = tensor("op_98_promoted"), val = tensor(0x1.f4p+7)]; + tensor var_262 = less(x = delta_1, y = var_98_promoted)[name = tensor("op_262")]; + tensor attn_mask_3 = logical_and(x = attn_mask_1, y = var_262)[name = tensor("attn_mask_3")]; + tensor attn_mask_5_axes_0 = const()[name = tensor("attn_mask_5_axes_0"), val = tensor([1])]; + tensor attn_mask_5 = expand_dims(axes = attn_mask_5_axes_0, x = attn_mask_3)[name = tensor("attn_mask_5")]; + tensor var_267_transpose_x_0 = const()[name = tensor("op_267_transpose_x_0"), val = tensor(false)]; + tensor var_267_transpose_y_0 = const()[name = tensor("op_267_transpose_y_0"), val = tensor(false)]; + tensor transpose_6_perm_0 = const()[name = tensor("transpose_6_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_7_perm_0 = const()[name = tensor("transpose_7_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_7 = transpose(perm = transpose_7_perm_0, x = new_k_cache_3)[name = tensor("transpose_16")]; + tensor transpose_6 = transpose(perm = transpose_6_perm_0, x = q_5)[name = tensor("transpose_17")]; + tensor var_267 = matmul(transpose_x = var_267_transpose_x_0, transpose_y = var_267_transpose_y_0, x = transpose_6, y = transpose_7)[name = tensor("op_267")]; + tensor var_268 = const()[name = tensor("op_268"), val = tensor(0x1p-3)]; + tensor attn_1 = mul(x = var_267, y = var_268)[name = tensor("attn_1")]; + tensor var_270 = logical_not(x = attn_mask_5)[name = tensor("op_270")]; + tensor attn_3 = select(a = var_100, b = attn_1, cond = var_270)[name = tensor("attn_3")]; + tensor attn_5 = softmax(axis = var_91, x = attn_3)[name = tensor("attn_5")]; + tensor x_5_transpose_x_0 = const()[name = tensor("x_5_transpose_x_0"), val = tensor(false)]; + tensor x_5_transpose_y_0 = const()[name = tensor("x_5_transpose_y_0"), val = tensor(false)]; + tensor v_attn_1 = transpose(perm = var_219, x = new_v_cache_3)[name = tensor("transpose_18")]; + tensor x_5 = matmul(transpose_x = x_5_transpose_x_0, transpose_y = x_5_transpose_y_0, x = attn_5, y = v_attn_1)[name = tensor("x_5")]; + tensor var_274_perm_0 = const()[name = tensor("op_274_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_275 = const()[name = tensor("op_275"), val = tensor([1, 16, 512])]; + tensor var_274 = transpose(perm = var_274_perm_0, x = x_5)[name = tensor("transpose_15")]; + tensor input_5 = reshape(shape = var_275, x = var_274)[name = tensor("input_5")]; + tensor linear_1_bias_0 = const()[name = tensor("linear_1_bias_0"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(41755456)))]; + tensor x_7 = linear(bias = linear_1_bias_0, weight = mimi_decoder_transformer_transformer_layers_0_self_attn_out_proj_weight, x = input_5)[name = tensor("linear_1")]; + tensor var_284 = mul(x = mimi_decoder_transformer_transformer_layers_0_layer_scale_1_scale, y = x_7)[name = tensor("op_284")]; + tensor input_7 = add(x = input_3, y = var_284)[name = tensor("input_7")]; + tensor input_9_axes_0 = const()[name = tensor("input_9_axes_0"), val = tensor([-1])]; + tensor input_9 = layer_norm(axes = input_9_axes_0, beta = mimi_decoder_transformer_transformer_layers_0_norm2_bias, epsilon = var_102, gamma = mimi_decoder_transformer_transformer_layers_0_norm2_weight, x = input_7)[name = tensor("input_9")]; + tensor linear_2_bias_0 = const()[name = tensor("linear_2_bias_0"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(41757568)))]; + tensor var_291 = linear(bias = linear_2_bias_0, weight = mimi_decoder_transformer_transformer_layers_0_linear1_weight, x = input_9)[name = tensor("linear_2")]; + tensor input_11_mode_0 = const()[name = tensor("input_11_mode_0"), val = tensor("EXACT")]; + tensor input_11 = gelu(mode = input_11_mode_0, x = var_291)[name = tensor("input_11")]; + tensor x_9 = linear(bias = linear_1_bias_0, weight = mimi_decoder_transformer_transformer_layers_0_linear2_weight, x = input_11)[name = tensor("linear_3")]; + tensor var_297 = mul(x = mimi_decoder_transformer_transformer_layers_0_layer_scale_2_scale, y = x_9)[name = tensor("op_297")]; + tensor input_13 = add(x = input_7, y = var_297)[name = tensor("input_13")]; + tensor query_axes_0 = const()[name = tensor("query_axes_0"), val = tensor([-1])]; + tensor query = layer_norm(axes = query_axes_0, beta = mimi_decoder_transformer_transformer_layers_1_norm1_bias, epsilon = var_102, gamma = mimi_decoder_transformer_transformer_layers_1_norm1_weight, x = input_13)[name = tensor("query")]; + tensor projected = linear(bias = linear_0_bias_0, weight = mimi_decoder_transformer_transformer_layers_1_self_attn_in_proj_weight, x = query)[name = tensor("linear_4")]; + tensor var_320 = const()[name = tensor("op_320"), val = tensor([1, 16, 3, 8, 64])]; + tensor packed = reshape(shape = var_320, x = projected)[name = tensor("packed")]; + tensor var_322_split_sizes_0 = const()[name = tensor("op_322_split_sizes_0"), val = tensor([1, 1, 1])]; + tensor var_322_axis_0 = const()[name = tensor("op_322_axis_0"), val = tensor(2)]; + tensor var_322_0, tensor var_322_1, tensor var_322_2 = split(axis = var_322_axis_0, split_sizes = var_322_split_sizes_0, x = packed)[name = tensor("op_322")]; + tensor squeeze_3_axes_0 = const()[name = tensor("squeeze_3_axes_0"), val = tensor([2])]; + tensor squeeze_3 = squeeze(axes = squeeze_3_axes_0, x = var_322_0)[name = tensor("squeeze_3")]; + tensor squeeze_4_axes_0 = const()[name = tensor("squeeze_4_axes_0"), val = tensor([2])]; + tensor squeeze_4 = squeeze(axes = squeeze_4_axes_0, x = var_322_1)[name = tensor("squeeze_4")]; + tensor squeeze_5_axes_0 = const()[name = tensor("squeeze_5_axes_0"), val = tensor([2])]; + tensor squeeze_5 = squeeze(axes = squeeze_5_axes_0, x = var_322_2)[name = tensor("squeeze_5")]; + tensor offset_begin_0 = const()[name = tensor("offset_begin_0"), val = tensor([0])]; + tensor offset_end_0 = const()[name = tensor("offset_end_0"), val = tensor([1])]; + tensor offset_end_mask_0 = const()[name = tensor("offset_end_mask_0"), val = tensor([false])]; + tensor offset_squeeze_mask_0 = const()[name = tensor("offset_squeeze_mask_0"), val = tensor([true])]; + tensor offset = slice_by_index(begin = offset_begin_0, end = offset_end_0, end_mask = offset_end_mask_0, squeeze_mask = offset_squeeze_mask_0, x = attn1_offset)[name = tensor("offset")]; + tensor freqs = const()[name = tensor("freqs"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(41765824)))]; + tensor ts_7_promoted = const()[name = tensor("ts_7_promoted"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(41766016)))]; + tensor ts_9 = add(x = ts_7_promoted, y = offset)[name = tensor("ts_9")]; + tensor var_338 = const()[name = tensor("op_338"), val = tensor([-1, 1, 1])]; + tensor ts = reshape(shape = var_338, x = ts_9)[name = tensor("ts")]; + tensor var_340 = const()[name = tensor("op_340"), val = tensor([1, 16, 8, 32, 2])]; + tensor q_9 = reshape(shape = var_340, x = squeeze_3)[name = tensor("q_9")]; + tensor var_342 = const()[name = tensor("op_342"), val = tensor([1, 16, 8, 32, 2])]; + tensor k_9 = reshape(shape = var_342, x = squeeze_4)[name = tensor("k_9")]; + tensor var_344_begin_0 = const()[name = tensor("op_344_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_344_end_0 = const()[name = tensor("op_344_end_0"), val = tensor([1, 16, 8, 32, 1])]; + tensor var_344_end_mask_0 = const()[name = tensor("op_344_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_344_squeeze_mask_0 = const()[name = tensor("op_344_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_344 = slice_by_index(begin = var_344_begin_0, end = var_344_end_0, end_mask = var_344_end_mask_0, squeeze_mask = var_344_squeeze_mask_0, x = q_9)[name = tensor("op_344")]; + tensor var_346_begin_0 = const()[name = tensor("op_346_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_346_end_0 = const()[name = tensor("op_346_end_0"), val = tensor([1, 16, 8, 32, 2])]; + tensor var_346_end_mask_0 = const()[name = tensor("op_346_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_346_squeeze_mask_0 = const()[name = tensor("op_346_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_346 = slice_by_index(begin = var_346_begin_0, end = var_346_end_0, end_mask = var_346_end_mask_0, squeeze_mask = var_346_squeeze_mask_0, x = q_9)[name = tensor("op_346")]; + tensor var_348_begin_0 = const()[name = tensor("op_348_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_348_end_0 = const()[name = tensor("op_348_end_0"), val = tensor([1, 16, 8, 32, 1])]; + tensor var_348_end_mask_0 = const()[name = tensor("op_348_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_348_squeeze_mask_0 = const()[name = tensor("op_348_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_348 = slice_by_index(begin = var_348_begin_0, end = var_348_end_0, end_mask = var_348_end_mask_0, squeeze_mask = var_348_squeeze_mask_0, x = k_9)[name = tensor("op_348")]; + tensor var_350_begin_0 = const()[name = tensor("op_350_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_350_end_0 = const()[name = tensor("op_350_end_0"), val = tensor([1, 16, 8, 32, 2])]; + tensor var_350_end_mask_0 = const()[name = tensor("op_350_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_350_squeeze_mask_0 = const()[name = tensor("op_350_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_350 = slice_by_index(begin = var_350_begin_0, end = var_350_end_0, end_mask = var_350_end_mask_0, squeeze_mask = var_350_squeeze_mask_0, x = k_9)[name = tensor("op_350")]; + tensor var_352 = mul(x = freqs, y = ts)[name = tensor("op_352")]; + tensor rotr = cos(x = var_352)[name = tensor("rotr")]; + tensor roti = sin(x = var_352)[name = tensor("roti")]; + tensor var_356 = mul(x = var_344, y = rotr)[name = tensor("op_356")]; + tensor var_357 = mul(x = var_346, y = roti)[name = tensor("op_357")]; + tensor qor_5 = sub(x = var_356, y = var_357)[name = tensor("qor_5")]; + tensor var_359 = mul(x = var_344, y = roti)[name = tensor("op_359")]; + tensor var_360 = mul(x = var_346, y = rotr)[name = tensor("op_360")]; + tensor qoi_5 = add(x = var_359, y = var_360)[name = tensor("qoi_5")]; + tensor var_362 = mul(x = var_348, y = rotr)[name = tensor("op_362")]; + tensor var_363 = mul(x = var_350, y = roti)[name = tensor("op_363")]; + tensor kor_5 = sub(x = var_362, y = var_363)[name = tensor("kor_5")]; + tensor var_365 = mul(x = var_348, y = roti)[name = tensor("op_365")]; + tensor var_366 = mul(x = var_350, y = rotr)[name = tensor("op_366")]; + tensor koi_5 = add(x = var_365, y = var_366)[name = tensor("koi_5")]; + tensor qo_axis_0 = const()[name = tensor("qo_axis_0"), val = tensor(-1)]; + tensor qo = stack(axis = qo_axis_0, values = (qor_5, qoi_5))[name = tensor("qo")]; + tensor ko_axis_0 = const()[name = tensor("ko_axis_0"), val = tensor(-1)]; + tensor ko = stack(axis = ko_axis_0, values = (kor_5, koi_5))[name = tensor("ko")]; + tensor var_376 = const()[name = tensor("op_376"), val = tensor([1, 16, 8, 64])]; + tensor q = reshape(shape = var_376, x = qo)[name = tensor("q")]; + tensor var_378 = const()[name = tensor("op_378"), val = tensor([1, 16, 8, 64])]; + tensor k = reshape(shape = var_378, x = ko)[name = tensor("k")]; + tensor capacity = const()[name = tensor("capacity"), val = tensor([256])]; + tensor var_383_dtype_0 = const()[name = tensor("op_383_dtype_0"), val = tensor("int32")]; + tensor var_384 = const()[name = tensor("op_384"), val = tensor([1, 1])]; + tensor var_383 = cast(dtype = var_383_dtype_0, x = attn1_offset)[name = tensor("cast_46")]; + tensor write_base = reshape(shape = var_384, x = var_383)[name = tensor("write_base")]; + tensor write_range = const()[name = tensor("write_range"), val = tensor([[0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15]])]; + tensor abs_idx = add(x = write_base, y = write_range)[name = tensor("abs_idx")]; + tensor wrapped_div = floor_div(x = abs_idx, y = capacity)[name = tensor("wrapped_div")]; + tensor wrapped_div_scaled = mul(x = wrapped_div, y = capacity)[name = tensor("wrapped_div_scaled")]; + tensor wrapped = sub(x = abs_idx, y = wrapped_div_scaled)[name = tensor("wrapped")]; + tensor var_391 = const()[name = tensor("op_391"), val = tensor([1, 16, 1, 1])]; + tensor var_392 = reshape(shape = var_391, x = wrapped)[name = tensor("op_392")]; + tensor write_indexes_reps_0 = const()[name = tensor("write_indexes_reps_0"), val = tensor([1, 1, 8, 64])]; + tensor write_indexes = tile(reps = write_indexes_reps_0, x = var_392)[name = tensor("write_indexes")]; + tensor var_395_begin_0 = const()[name = tensor("op_395_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_395_end_0 = const()[name = tensor("op_395_end_0"), val = tensor([1, 1, 256, 8, 64])]; + tensor var_395_end_mask_0 = const()[name = tensor("op_395_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_395_squeeze_mask_0 = const()[name = tensor("op_395_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor var_395 = slice_by_index(begin = var_395_begin_0, end = var_395_end_0, end_mask = var_395_end_mask_0, squeeze_mask = var_395_squeeze_mask_0, x = attn1_cache)[name = tensor("op_395")]; + tensor new_k_cache_5_axis_0 = const()[name = tensor("new_k_cache_5_axis_0"), val = tensor(1)]; + tensor new_k_cache_5_mode_0 = const()[name = tensor("new_k_cache_5_mode_0"), val = tensor("update")]; + tensor new_k_cache_5_validate_indices_0 = const()[name = tensor("new_k_cache_5_validate_indices_0"), val = tensor(false)]; + tensor new_k_cache_5 = scatter_along_axis(axis = new_k_cache_5_axis_0, data = var_395, indices = write_indexes, mode = new_k_cache_5_mode_0, updates = k, validate_indices = new_k_cache_5_validate_indices_0)[name = tensor("new_k_cache_5")]; + tensor var_397_begin_0 = const()[name = tensor("op_397_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor var_397_end_0 = const()[name = tensor("op_397_end_0"), val = tensor([2, 1, 256, 8, 64])]; + tensor var_397_end_mask_0 = const()[name = tensor("op_397_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_397_squeeze_mask_0 = const()[name = tensor("op_397_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor var_397 = slice_by_index(begin = var_397_begin_0, end = var_397_end_0, end_mask = var_397_end_mask_0, squeeze_mask = var_397_squeeze_mask_0, x = attn1_cache)[name = tensor("op_397")]; + tensor new_v_cache_5_axis_0 = const()[name = tensor("new_v_cache_5_axis_0"), val = tensor(1)]; + tensor new_v_cache_5_mode_0 = const()[name = tensor("new_v_cache_5_mode_0"), val = tensor("update")]; + tensor new_v_cache_5_validate_indices_0 = const()[name = tensor("new_v_cache_5_validate_indices_0"), val = tensor(false)]; + tensor new_v_cache_5 = scatter_along_axis(axis = new_v_cache_5_axis_0, data = var_397, indices = write_indexes, mode = new_v_cache_5_mode_0, updates = squeeze_5, validate_indices = new_v_cache_5_validate_indices_0)[name = tensor("new_v_cache_5")]; + tensor var_400_axis_0 = const()[name = tensor("op_400_axis_0"), val = tensor(0)]; + tensor var_400 = stack(axis = var_400_axis_0, values = (new_k_cache_5, new_v_cache_5))[name = tensor("op_400")]; + tensor var_401 = not_equal(x = new_k_cache_5, y = new_k_cache_5)[name = tensor("op_401")]; + tensor new_k_cache = select(a = var_212, b = new_k_cache_5, cond = var_401)[name = tensor("new_k_cache")]; + tensor var_404 = not_equal(x = new_v_cache_5, y = new_v_cache_5)[name = tensor("op_404")]; + tensor new_v_cache = select(a = var_212, b = new_v_cache_5, cond = var_404)[name = tensor("new_v_cache")]; + tensor var_409 = const()[name = tensor("op_409"), val = tensor([0, 2, 1, 3])]; + tensor var_411 = const()[name = tensor("op_411"), val = tensor([1, 1])]; + tensor var_412 = reshape(shape = var_411, x = attn1_offset)[name = tensor("op_412")]; + tensor var_414_promoted = const()[name = tensor("op_414_promoted"), val = tensor([0x1.ep+3])]; + tensor var_415 = add(x = var_412, y = var_414_promoted)[name = tensor("op_415")]; + tensor last_pos_dtype_0 = const()[name = tensor("last_pos_dtype_0"), val = tensor("int32")]; + tensor last_pos = cast(dtype = last_pos_dtype_0, x = var_415)[name = tensor("cast_45")]; + tensor diff = sub(x = last_pos, y = slot_idx_1)[name = tensor("diff")]; + tensor var_421_div = floor_div(x = diff, y = capacity)[name = tensor("op_421_div")]; + tensor var_421_div_scaled = mul(x = var_421_div, y = capacity)[name = tensor("op_421_div_scaled")]; + tensor var_421 = sub(x = diff, y = var_421_div_scaled)[name = tensor("op_421")]; + tensor pos_k = sub(x = last_pos, y = var_421)[name = tensor("pos_k")]; + tensor var_427_promoted = const()[name = tensor("op_427_promoted"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(41766144)))]; + tensor pos_q = add(x = var_412, y = var_427_promoted)[name = tensor("pos_q")]; + tensor var_431_axes_0 = const()[name = tensor("op_431_axes_0"), val = tensor([2])]; + tensor var_431 = expand_dims(axes = var_431_axes_0, x = pos_q)[name = tensor("op_431")]; + tensor var_433_axes_0 = const()[name = tensor("op_433_axes_0"), val = tensor([1])]; + tensor var_433 = expand_dims(axes = var_433_axes_0, x = pos_k)[name = tensor("op_433")]; + tensor var_434_promoted_dtype_0 = const()[name = tensor("op_434_promoted_dtype_0"), val = tensor("fp32")]; + tensor var_434_promoted = cast(dtype = var_434_promoted_dtype_0, x = var_433)[name = tensor("cast_44")]; + tensor delta = sub(x = var_431, y = var_434_promoted)[name = tensor("delta")]; + tensor valid_5 = greater_equal(x = var_433, y = var_86)[name = tensor("valid_5")]; + tensor var_443 = const()[name = tensor("op_443"), val = tensor([1, 1, 1])]; + tensor var_444 = reshape(shape = var_443, x = attn1_offset)[name = tensor("op_444")]; + tensor var_446_promoted = const()[name = tensor("op_446_promoted"), val = tensor([0x1.ep+3])]; + tensor var_447 = add(x = var_444, y = var_446_promoted)[name = tensor("op_447")]; + tensor var_448 = less_equal(x = var_434_promoted, y = var_447)[name = tensor("op_448")]; + tensor valid = logical_and(x = valid_5, y = var_448)[name = tensor("valid")]; + tensor var_86_promoted_1 = const()[name = tensor("op_86_promoted_1"), val = tensor(0x0p+0)]; + tensor var_450 = greater_equal(x = delta, y = var_86_promoted_1)[name = tensor("op_450")]; + tensor attn_mask_7 = logical_and(x = valid, y = var_450)[name = tensor("attn_mask_7")]; + tensor var_98_promoted_1 = const()[name = tensor("op_98_promoted_1"), val = tensor(0x1.f4p+7)]; + tensor var_452 = less(x = delta, y = var_98_promoted_1)[name = tensor("op_452")]; + tensor attn_mask_9 = logical_and(x = attn_mask_7, y = var_452)[name = tensor("attn_mask_9")]; + tensor attn_mask_axes_0 = const()[name = tensor("attn_mask_axes_0"), val = tensor([1])]; + tensor attn_mask = expand_dims(axes = attn_mask_axes_0, x = attn_mask_9)[name = tensor("attn_mask")]; + tensor var_457_transpose_x_0 = const()[name = tensor("op_457_transpose_x_0"), val = tensor(false)]; + tensor var_457_transpose_y_0 = const()[name = tensor("op_457_transpose_y_0"), val = tensor(false)]; + tensor transpose_8_perm_0 = const()[name = tensor("transpose_8_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_9_perm_0 = const()[name = tensor("transpose_9_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_9 = transpose(perm = transpose_9_perm_0, x = new_k_cache)[name = tensor("transpose_12")]; + tensor transpose_8 = transpose(perm = transpose_8_perm_0, x = q)[name = tensor("transpose_13")]; + tensor var_457 = matmul(transpose_x = var_457_transpose_x_0, transpose_y = var_457_transpose_y_0, x = transpose_8, y = transpose_9)[name = tensor("op_457")]; + tensor var_458 = const()[name = tensor("op_458"), val = tensor(0x1p-3)]; + tensor attn_7 = mul(x = var_457, y = var_458)[name = tensor("attn_7")]; + tensor var_460 = logical_not(x = attn_mask)[name = tensor("op_460")]; + tensor attn_9 = select(a = var_100, b = attn_7, cond = var_460)[name = tensor("attn_9")]; + tensor attn = softmax(axis = var_91, x = attn_9)[name = tensor("attn")]; + tensor x_11_transpose_x_0 = const()[name = tensor("x_11_transpose_x_0"), val = tensor(false)]; + tensor x_11_transpose_y_0 = const()[name = tensor("x_11_transpose_y_0"), val = tensor(false)]; + tensor v_attn = transpose(perm = var_409, x = new_v_cache)[name = tensor("transpose_14")]; + tensor x_11 = matmul(transpose_x = x_11_transpose_x_0, transpose_y = x_11_transpose_y_0, x = attn, y = v_attn)[name = tensor("x_11")]; + tensor var_464_perm_0 = const()[name = tensor("op_464_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_465 = const()[name = tensor("op_465"), val = tensor([1, 16, 512])]; + tensor var_464 = transpose(perm = var_464_perm_0, x = x_11)[name = tensor("transpose_11")]; + tensor input_15 = reshape(shape = var_465, x = var_464)[name = tensor("input_15")]; + tensor x_13 = linear(bias = linear_1_bias_0, weight = mimi_decoder_transformer_transformer_layers_1_self_attn_out_proj_weight, x = input_15)[name = tensor("linear_5")]; + tensor var_474 = mul(x = mimi_decoder_transformer_transformer_layers_1_layer_scale_1_scale, y = x_13)[name = tensor("op_474")]; + tensor input_17 = add(x = input_13, y = var_474)[name = tensor("input_17")]; + tensor input_19_axes_0 = const()[name = tensor("input_19_axes_0"), val = tensor([-1])]; + tensor input_19 = layer_norm(axes = input_19_axes_0, beta = mimi_decoder_transformer_transformer_layers_1_norm2_bias, epsilon = var_102, gamma = mimi_decoder_transformer_transformer_layers_1_norm2_weight, x = input_17)[name = tensor("input_19")]; + tensor var_481 = linear(bias = linear_2_bias_0, weight = mimi_decoder_transformer_transformer_layers_1_linear1_weight, x = input_19)[name = tensor("linear_6")]; + tensor input_21_mode_0 = const()[name = tensor("input_21_mode_0"), val = tensor("EXACT")]; + tensor input_21 = gelu(mode = input_21_mode_0, x = var_481)[name = tensor("input_21")]; + tensor x_15 = linear(bias = linear_1_bias_0, weight = mimi_decoder_transformer_transformer_layers_1_linear2_weight, x = input_21)[name = tensor("linear_7")]; + tensor var_487 = mul(x = mimi_decoder_transformer_transformer_layers_1_layer_scale_2_scale, y = x_15)[name = tensor("op_487")]; + tensor z = add(x = input_17, y = var_487)[name = tensor("z")]; + tensor x_17_perm_0 = const()[name = tensor("x_17_perm_0"), val = tensor([0, 2, 1])]; + tensor var_507 = const()[name = tensor("op_507"), val = tensor(0x1p+0)]; + tensor var_508 = const()[name = tensor("op_508"), val = tensor(-1)]; + tensor input_23_interleave_0 = const()[name = tensor("input_23_interleave_0"), val = tensor(false)]; + tensor x_17 = transpose(perm = x_17_perm_0, x = z)[name = tensor("transpose_10")]; + tensor input_23 = concat(axis = var_508, interleave = input_23_interleave_0, values = (conv0_prev, x_17))[name = tensor("input_23")]; + tensor input_25_pad_type_0 = const()[name = tensor("input_25_pad_type_0"), val = tensor("valid")]; + tensor input_25_strides_0 = const()[name = tensor("input_25_strides_0"), val = tensor([1])]; + tensor input_25_pad_0 = const()[name = tensor("input_25_pad_0"), val = tensor([0, 0])]; + tensor input_25_dilations_0 = const()[name = tensor("input_25_dilations_0"), val = tensor([1])]; + tensor input_25_groups_0 = const()[name = tensor("input_25_groups_0"), val = tensor(1)]; + tensor input_25 = conv(bias = mimi_decoder_model_0_conv_bias, dilations = input_25_dilations_0, groups = input_25_groups_0, pad = input_25_pad_0, pad_type = input_25_pad_type_0, strides = input_25_strides_0, weight = mimi_decoder_model_0_conv_weight, x = input_23)[name = tensor("input_25")]; + tensor var_542_begin_0 = const()[name = tensor("op_542_begin_0"), val = tensor([0, 0, 16])]; + tensor var_542_end_0 = const()[name = tensor("op_542_end_0"), val = tensor([1, 512, 22])]; + tensor var_542_end_mask_0 = const()[name = tensor("op_542_end_mask_0"), val = tensor([true, true, true])]; + tensor var_542 = slice_by_index(begin = var_542_begin_0, end = var_542_end_0, end_mask = var_542_end_mask_0, x = input_23)[name = tensor("op_542")]; + tensor input_27 = elu(alpha = var_507, x = input_25)[name = tensor("input_27")]; + tensor y_5_pad_type_0 = const()[name = tensor("y_5_pad_type_0"), val = tensor("valid")]; + tensor y_5_strides_0 = const()[name = tensor("y_5_strides_0"), val = tensor([6])]; + tensor y_5_pad_0 = const()[name = tensor("y_5_pad_0"), val = tensor([0, 0])]; + tensor y_5_dilations_0 = const()[name = tensor("y_5_dilations_0"), val = tensor([1])]; + tensor y_5_groups_0 = const()[name = tensor("y_5_groups_0"), val = tensor(1)]; + tensor y_5_has_output_shape_output_shape_0 = const()[name = tensor("y_5_has_output_shape_output_shape_0"), val = tensor([1, 256, 102])]; + tensor y_5_has_output_shape = conv_transpose(bias = mimi_decoder_model_2_convtr_bias, dilations = y_5_dilations_0, groups = y_5_groups_0, output_shape = y_5_has_output_shape_output_shape_0, pad = y_5_pad_0, pad_type = y_5_pad_type_0, strides = y_5_strides_0, weight = mimi_decoder_model_2_convtr_weight, x = input_27)[name = tensor("y_5_has_output_shape")]; + tensor var_557_begin_0 = const()[name = tensor("op_557_begin_0"), val = tensor([0, 0, 0])]; + tensor var_557_end_0 = const()[name = tensor("op_557_end_0"), val = tensor([1, 256, 6])]; + tensor var_557_end_mask_0 = const()[name = tensor("op_557_end_mask_0"), val = tensor([true, true, false])]; + tensor var_557 = slice_by_index(begin = var_557_begin_0, end = var_557_end_0, end_mask = var_557_end_mask_0, x = y_5_has_output_shape)[name = tensor("op_557")]; + tensor var_558 = add(x = var_557, y = convtr0_partial)[name = tensor("op_558")]; + tensor var_559_begin_0 = const()[name = tensor("op_559_begin_0"), val = tensor([0, 0, 6])]; + tensor var_559_end_0 = const()[name = tensor("op_559_end_0"), val = tensor([1, 256, 102])]; + tensor var_559_end_mask_0 = const()[name = tensor("op_559_end_mask_0"), val = tensor([true, true, true])]; + tensor var_559 = slice_by_index(begin = var_559_begin_0, end = var_559_end_0, end_mask = var_559_end_mask_0, x = y_5_has_output_shape)[name = tensor("op_559")]; + tensor y_7_interleave_0 = const()[name = tensor("y_7_interleave_0"), val = tensor(false)]; + tensor y_7 = concat(axis = var_508, interleave = y_7_interleave_0, values = (var_558, var_559))[name = tensor("y_7")]; + tensor new_partial_1_begin_0 = const()[name = tensor("new_partial_1_begin_0"), val = tensor([0, 0, 96])]; + tensor new_partial_1_end_0 = const()[name = tensor("new_partial_1_end_0"), val = tensor([1, 256, 102])]; + tensor new_partial_1_end_mask_0 = const()[name = tensor("new_partial_1_end_mask_0"), val = tensor([true, true, true])]; + tensor new_partial_1 = slice_by_index(begin = new_partial_1_begin_0, end = new_partial_1_end_0, end_mask = new_partial_1_end_mask_0, x = y_7)[name = tensor("new_partial_1")]; + tensor var_564 = const()[name = tensor("op_564"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(41766272)))]; + tensor var_565 = sub(x = new_partial_1, y = var_564)[name = tensor("op_565")]; + tensor input_29_begin_0 = const()[name = tensor("input_29_begin_0"), val = tensor([0, 0, 0])]; + tensor input_29_end_0 = const()[name = tensor("input_29_end_0"), val = tensor([1, 256, 96])]; + tensor input_29_end_mask_0 = const()[name = tensor("input_29_end_mask_0"), val = tensor([true, true, false])]; + tensor input_29 = slice_by_index(begin = input_29_begin_0, end = input_29_end_0, end_mask = input_29_end_mask_0, x = y_7)[name = tensor("input_29")]; + tensor x_19 = elu(alpha = var_507, x = input_29)[name = tensor("x_19")]; + tensor input_31_interleave_0 = const()[name = tensor("input_31_interleave_0"), val = tensor(false)]; + tensor input_31 = concat(axis = var_508, interleave = input_31_interleave_0, values = (res0_conv0_prev, x_19))[name = tensor("input_31")]; + tensor input_33_pad_type_0 = const()[name = tensor("input_33_pad_type_0"), val = tensor("valid")]; + tensor input_33_strides_0 = const()[name = tensor("input_33_strides_0"), val = tensor([1])]; + tensor input_33_pad_0 = const()[name = tensor("input_33_pad_0"), val = tensor([0, 0])]; + tensor input_33_dilations_0 = const()[name = tensor("input_33_dilations_0"), val = tensor([1])]; + tensor input_33_groups_0 = const()[name = tensor("input_33_groups_0"), val = tensor(1)]; + tensor input_33 = conv(bias = mimi_decoder_model_3_block_1_conv_bias, dilations = input_33_dilations_0, groups = input_33_groups_0, pad = input_33_pad_0, pad_type = input_33_pad_type_0, strides = input_33_strides_0, weight = mimi_decoder_model_3_block_1_conv_weight, x = input_31)[name = tensor("input_33")]; + tensor var_585_begin_0 = const()[name = tensor("op_585_begin_0"), val = tensor([0, 0, 96])]; + tensor var_585_end_0 = const()[name = tensor("op_585_end_0"), val = tensor([1, 256, 98])]; + tensor var_585_end_mask_0 = const()[name = tensor("op_585_end_mask_0"), val = tensor([true, true, true])]; + tensor var_585 = slice_by_index(begin = var_585_begin_0, end = var_585_end_0, end_mask = var_585_end_mask_0, x = input_31)[name = tensor("op_585")]; + tensor x_21 = elu(alpha = var_507, x = input_33)[name = tensor("x_21")]; + tensor v_5_pad_type_0 = const()[name = tensor("v_5_pad_type_0"), val = tensor("valid")]; + tensor v_5_strides_0 = const()[name = tensor("v_5_strides_0"), val = tensor([1])]; + tensor v_5_pad_0 = const()[name = tensor("v_5_pad_0"), val = tensor([0, 0])]; + tensor v_5_dilations_0 = const()[name = tensor("v_5_dilations_0"), val = tensor([1])]; + tensor v_5_groups_0 = const()[name = tensor("v_5_groups_0"), val = tensor(1)]; + tensor v_5 = conv(bias = mimi_decoder_model_3_block_3_conv_bias, dilations = v_5_dilations_0, groups = v_5_groups_0, pad = v_5_pad_0, pad_type = v_5_pad_type_0, strides = v_5_strides_0, weight = mimi_decoder_model_3_block_3_conv_weight, x = x_21)[name = tensor("v_5")]; + tensor input_35 = add(x = input_29, y = v_5)[name = tensor("input_35")]; + tensor input_37 = elu(alpha = var_507, x = input_35)[name = tensor("input_37")]; + tensor y_9_pad_type_0 = const()[name = tensor("y_9_pad_type_0"), val = tensor("valid")]; + tensor y_9_strides_0 = const()[name = tensor("y_9_strides_0"), val = tensor([5])]; + tensor y_9_pad_0 = const()[name = tensor("y_9_pad_0"), val = tensor([0, 0])]; + tensor y_9_dilations_0 = const()[name = tensor("y_9_dilations_0"), val = tensor([1])]; + tensor y_9_groups_0 = const()[name = tensor("y_9_groups_0"), val = tensor(1)]; + tensor y_9_has_output_shape_output_shape_0 = const()[name = tensor("y_9_has_output_shape_output_shape_0"), val = tensor([1, 128, 485])]; + tensor y_9_has_output_shape = conv_transpose(bias = mimi_decoder_model_5_convtr_bias, dilations = y_9_dilations_0, groups = y_9_groups_0, output_shape = y_9_has_output_shape_output_shape_0, pad = y_9_pad_0, pad_type = y_9_pad_type_0, strides = y_9_strides_0, weight = mimi_decoder_model_5_convtr_weight, x = input_37)[name = tensor("y_9_has_output_shape")]; + tensor var_613_begin_0 = const()[name = tensor("op_613_begin_0"), val = tensor([0, 0, 0])]; + tensor var_613_end_0 = const()[name = tensor("op_613_end_0"), val = tensor([1, 128, 5])]; + tensor var_613_end_mask_0 = const()[name = tensor("op_613_end_mask_0"), val = tensor([true, true, false])]; + tensor var_613 = slice_by_index(begin = var_613_begin_0, end = var_613_end_0, end_mask = var_613_end_mask_0, x = y_9_has_output_shape)[name = tensor("op_613")]; + tensor var_614 = add(x = var_613, y = convtr1_partial)[name = tensor("op_614")]; + tensor var_615_begin_0 = const()[name = tensor("op_615_begin_0"), val = tensor([0, 0, 5])]; + tensor var_615_end_0 = const()[name = tensor("op_615_end_0"), val = tensor([1, 128, 485])]; + tensor var_615_end_mask_0 = const()[name = tensor("op_615_end_mask_0"), val = tensor([true, true, true])]; + tensor var_615 = slice_by_index(begin = var_615_begin_0, end = var_615_end_0, end_mask = var_615_end_mask_0, x = y_9_has_output_shape)[name = tensor("op_615")]; + tensor y_11_interleave_0 = const()[name = tensor("y_11_interleave_0"), val = tensor(false)]; + tensor y_11 = concat(axis = var_508, interleave = y_11_interleave_0, values = (var_614, var_615))[name = tensor("y_11")]; + tensor new_partial_3_begin_0 = const()[name = tensor("new_partial_3_begin_0"), val = tensor([0, 0, 480])]; + tensor new_partial_3_end_0 = const()[name = tensor("new_partial_3_end_0"), val = tensor([1, 128, 485])]; + tensor new_partial_3_end_mask_0 = const()[name = tensor("new_partial_3_end_mask_0"), val = tensor([true, true, true])]; + tensor new_partial_3 = slice_by_index(begin = new_partial_3_begin_0, end = new_partial_3_end_0, end_mask = new_partial_3_end_mask_0, x = y_11)[name = tensor("new_partial_3")]; + tensor var_620 = const()[name = tensor("op_620"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(41767360)))]; + tensor var_621 = sub(x = new_partial_3, y = var_620)[name = tensor("op_621")]; + tensor input_39_begin_0 = const()[name = tensor("input_39_begin_0"), val = tensor([0, 0, 0])]; + tensor input_39_end_0 = const()[name = tensor("input_39_end_0"), val = tensor([1, 128, 480])]; + tensor input_39_end_mask_0 = const()[name = tensor("input_39_end_mask_0"), val = tensor([true, true, false])]; + tensor input_39 = slice_by_index(begin = input_39_begin_0, end = input_39_end_0, end_mask = input_39_end_mask_0, x = y_11)[name = tensor("input_39")]; + tensor x_23 = elu(alpha = var_507, x = input_39)[name = tensor("x_23")]; + tensor input_41_interleave_0 = const()[name = tensor("input_41_interleave_0"), val = tensor(false)]; + tensor input_41 = concat(axis = var_508, interleave = input_41_interleave_0, values = (res1_conv0_prev, x_23))[name = tensor("input_41")]; + tensor input_43_pad_type_0 = const()[name = tensor("input_43_pad_type_0"), val = tensor("valid")]; + tensor input_43_strides_0 = const()[name = tensor("input_43_strides_0"), val = tensor([1])]; + tensor input_43_pad_0 = const()[name = tensor("input_43_pad_0"), val = tensor([0, 0])]; + tensor input_43_dilations_0 = const()[name = tensor("input_43_dilations_0"), val = tensor([1])]; + tensor input_43_groups_0 = const()[name = tensor("input_43_groups_0"), val = tensor(1)]; + tensor input_43 = conv(bias = mimi_decoder_model_6_block_1_conv_bias, dilations = input_43_dilations_0, groups = input_43_groups_0, pad = input_43_pad_0, pad_type = input_43_pad_type_0, strides = input_43_strides_0, weight = mimi_decoder_model_6_block_1_conv_weight, x = input_41)[name = tensor("input_43")]; + tensor var_641_begin_0 = const()[name = tensor("op_641_begin_0"), val = tensor([0, 0, 480])]; + tensor var_641_end_0 = const()[name = tensor("op_641_end_0"), val = tensor([1, 128, 482])]; + tensor var_641_end_mask_0 = const()[name = tensor("op_641_end_mask_0"), val = tensor([true, true, true])]; + tensor var_641 = slice_by_index(begin = var_641_begin_0, end = var_641_end_0, end_mask = var_641_end_mask_0, x = input_41)[name = tensor("op_641")]; + tensor x_25 = elu(alpha = var_507, x = input_43)[name = tensor("x_25")]; + tensor v_7_pad_type_0 = const()[name = tensor("v_7_pad_type_0"), val = tensor("valid")]; + tensor v_7_strides_0 = const()[name = tensor("v_7_strides_0"), val = tensor([1])]; + tensor v_7_pad_0 = const()[name = tensor("v_7_pad_0"), val = tensor([0, 0])]; + tensor v_7_dilations_0 = const()[name = tensor("v_7_dilations_0"), val = tensor([1])]; + tensor v_7_groups_0 = const()[name = tensor("v_7_groups_0"), val = tensor(1)]; + tensor v_7 = conv(bias = mimi_decoder_model_6_block_3_conv_bias, dilations = v_7_dilations_0, groups = v_7_groups_0, pad = v_7_pad_0, pad_type = v_7_pad_type_0, strides = v_7_strides_0, weight = mimi_decoder_model_6_block_3_conv_weight, x = x_25)[name = tensor("v_7")]; + tensor input_45 = add(x = input_39, y = v_7)[name = tensor("input_45")]; + tensor input_47 = elu(alpha = var_507, x = input_45)[name = tensor("input_47")]; + tensor y_13_pad_type_0 = const()[name = tensor("y_13_pad_type_0"), val = tensor("valid")]; + tensor y_13_strides_0 = const()[name = tensor("y_13_strides_0"), val = tensor([4])]; + tensor y_13_pad_0 = const()[name = tensor("y_13_pad_0"), val = tensor([0, 0])]; + tensor y_13_dilations_0 = const()[name = tensor("y_13_dilations_0"), val = tensor([1])]; + tensor y_13_groups_0 = const()[name = tensor("y_13_groups_0"), val = tensor(1)]; + tensor y_13_has_output_shape_output_shape_0 = const()[name = tensor("y_13_has_output_shape_output_shape_0"), val = tensor([1, 64, 1924])]; + tensor y_13_has_output_shape = conv_transpose(bias = mimi_decoder_model_8_convtr_bias, dilations = y_13_dilations_0, groups = y_13_groups_0, output_shape = y_13_has_output_shape_output_shape_0, pad = y_13_pad_0, pad_type = y_13_pad_type_0, strides = y_13_strides_0, weight = mimi_decoder_model_8_convtr_weight, x = input_47)[name = tensor("y_13_has_output_shape")]; + tensor var_669_begin_0 = const()[name = tensor("op_669_begin_0"), val = tensor([0, 0, 0])]; + tensor var_669_end_0 = const()[name = tensor("op_669_end_0"), val = tensor([1, 64, 4])]; + tensor var_669_end_mask_0 = const()[name = tensor("op_669_end_mask_0"), val = tensor([true, true, false])]; + tensor var_669 = slice_by_index(begin = var_669_begin_0, end = var_669_end_0, end_mask = var_669_end_mask_0, x = y_13_has_output_shape)[name = tensor("op_669")]; + tensor var_670 = add(x = var_669, y = convtr2_partial)[name = tensor("op_670")]; + tensor var_671_begin_0 = const()[name = tensor("op_671_begin_0"), val = tensor([0, 0, 4])]; + tensor var_671_end_0 = const()[name = tensor("op_671_end_0"), val = tensor([1, 64, 1924])]; + tensor var_671_end_mask_0 = const()[name = tensor("op_671_end_mask_0"), val = tensor([true, true, true])]; + tensor var_671 = slice_by_index(begin = var_671_begin_0, end = var_671_end_0, end_mask = var_671_end_mask_0, x = y_13_has_output_shape)[name = tensor("op_671")]; + tensor y_interleave_0 = const()[name = tensor("y_interleave_0"), val = tensor(false)]; + tensor y = concat(axis = var_508, interleave = y_interleave_0, values = (var_670, var_671))[name = tensor("y")]; + tensor new_partial_begin_0 = const()[name = tensor("new_partial_begin_0"), val = tensor([0, 0, 1920])]; + tensor new_partial_end_0 = const()[name = tensor("new_partial_end_0"), val = tensor([1, 64, 1924])]; + tensor new_partial_end_mask_0 = const()[name = tensor("new_partial_end_mask_0"), val = tensor([true, true, true])]; + tensor new_partial = slice_by_index(begin = new_partial_begin_0, end = new_partial_end_0, end_mask = new_partial_end_mask_0, x = y)[name = tensor("new_partial")]; + tensor var_676 = const()[name = tensor("op_676"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(41767936)))]; + tensor var_677 = sub(x = new_partial, y = var_676)[name = tensor("op_677")]; + tensor input_49_begin_0 = const()[name = tensor("input_49_begin_0"), val = tensor([0, 0, 0])]; + tensor input_49_end_0 = const()[name = tensor("input_49_end_0"), val = tensor([1, 64, 1920])]; + tensor input_49_end_mask_0 = const()[name = tensor("input_49_end_mask_0"), val = tensor([true, true, false])]; + tensor input_49 = slice_by_index(begin = input_49_begin_0, end = input_49_end_0, end_mask = input_49_end_mask_0, x = y)[name = tensor("input_49")]; + tensor x_27 = elu(alpha = var_507, x = input_49)[name = tensor("x_27")]; + tensor input_51_interleave_0 = const()[name = tensor("input_51_interleave_0"), val = tensor(false)]; + tensor input_51 = concat(axis = var_508, interleave = input_51_interleave_0, values = (res2_conv0_prev, x_27))[name = tensor("input_51")]; + tensor input_53_pad_type_0 = const()[name = tensor("input_53_pad_type_0"), val = tensor("valid")]; + tensor input_53_strides_0 = const()[name = tensor("input_53_strides_0"), val = tensor([1])]; + tensor input_53_pad_0 = const()[name = tensor("input_53_pad_0"), val = tensor([0, 0])]; + tensor input_53_dilations_0 = const()[name = tensor("input_53_dilations_0"), val = tensor([1])]; + tensor input_53_groups_0 = const()[name = tensor("input_53_groups_0"), val = tensor(1)]; + tensor input_53 = conv(bias = mimi_decoder_model_9_block_1_conv_bias, dilations = input_53_dilations_0, groups = input_53_groups_0, pad = input_53_pad_0, pad_type = input_53_pad_type_0, strides = input_53_strides_0, weight = mimi_decoder_model_9_block_1_conv_weight, x = input_51)[name = tensor("input_53")]; + tensor var_697_begin_0 = const()[name = tensor("op_697_begin_0"), val = tensor([0, 0, 1920])]; + tensor var_697_end_0 = const()[name = tensor("op_697_end_0"), val = tensor([1, 64, 1922])]; + tensor var_697_end_mask_0 = const()[name = tensor("op_697_end_mask_0"), val = tensor([true, true, true])]; + tensor var_697 = slice_by_index(begin = var_697_begin_0, end = var_697_end_0, end_mask = var_697_end_mask_0, x = input_51)[name = tensor("op_697")]; + tensor x_29 = elu(alpha = var_507, x = input_53)[name = tensor("x_29")]; + tensor v_pad_type_0 = const()[name = tensor("v_pad_type_0"), val = tensor("valid")]; + tensor v_strides_0 = const()[name = tensor("v_strides_0"), val = tensor([1])]; + tensor v_pad_0 = const()[name = tensor("v_pad_0"), val = tensor([0, 0])]; + tensor v_dilations_0 = const()[name = tensor("v_dilations_0"), val = tensor([1])]; + tensor v_groups_0 = const()[name = tensor("v_groups_0"), val = tensor(1)]; + tensor v = conv(bias = mimi_decoder_model_9_block_3_conv_bias, dilations = v_dilations_0, groups = v_groups_0, pad = v_pad_0, pad_type = v_pad_type_0, strides = v_strides_0, weight = mimi_decoder_model_9_block_3_conv_weight, x = x_29)[name = tensor("v")]; + tensor input_55 = add(x = input_49, y = v)[name = tensor("input_55")]; + tensor x = elu(alpha = var_507, x = input_55)[name = tensor("x")]; + tensor input_interleave_0 = const()[name = tensor("input_interleave_0"), val = tensor(false)]; + tensor input = concat(axis = var_508, interleave = input_interleave_0, values = (conv_final_prev, x))[name = tensor("input")]; + tensor var_724_pad_type_0 = const()[name = tensor("op_724_pad_type_0"), val = tensor("valid")]; + tensor var_724_strides_0 = const()[name = tensor("op_724_strides_0"), val = tensor([1])]; + tensor var_724_pad_0 = const()[name = tensor("op_724_pad_0"), val = tensor([0, 0])]; + tensor var_724_dilations_0 = const()[name = tensor("op_724_dilations_0"), val = tensor([1])]; + tensor var_724_groups_0 = const()[name = tensor("op_724_groups_0"), val = tensor(1)]; + tensor var_724 = conv(bias = mimi_decoder_model_11_conv_bias, dilations = var_724_dilations_0, groups = var_724_groups_0, pad = var_724_pad_0, pad_type = var_724_pad_type_0, strides = var_724_strides_0, weight = mimi_decoder_model_11_conv_weight, x = input)[name = tensor("op_724")]; + tensor var_725_begin_0 = const()[name = tensor("op_725_begin_0"), val = tensor([0, 0, 1920])]; + tensor var_725_end_0 = const()[name = tensor("op_725_end_0"), val = tensor([1, 64, 1922])]; + tensor var_725_end_mask_0 = const()[name = tensor("op_725_end_mask_0"), val = tensor([true, true, true])]; + tensor var_725 = slice_by_index(begin = var_725_begin_0, end = var_725_end_0, end_mask = var_725_end_mask_0, x = input)[name = tensor("op_725")]; + tensor var_740_promoted = const()[name = tensor("op_740_promoted"), val = tensor(0x1p+4)]; + tensor var_741 = add(x = attn0_offset, y = var_740_promoted)[name = tensor("op_741")]; + tensor var_743_promoted = const()[name = tensor("op_743_promoted"), val = tensor(0x1p+4)]; + tensor var_744 = add(x = attn1_offset, y = var_743_promoted)[name = tensor("op_744")]; + tensor conv0_first_tmp = identity(x = conv0_first)[name = tensor("conv0_first_tmp")]; + tensor res0_conv0_first_tmp = identity(x = res0_conv0_first)[name = tensor("res0_conv0_first_tmp")]; + tensor res0_conv1_prev_tmp = identity(x = res0_conv1_prev)[name = tensor("res0_conv1_prev_tmp")]; + tensor res0_conv1_first_tmp = identity(x = res0_conv1_first)[name = tensor("res0_conv1_first_tmp")]; + tensor res1_conv0_first_tmp = identity(x = res1_conv0_first)[name = tensor("res1_conv0_first_tmp")]; + tensor res1_conv1_prev_tmp = identity(x = res1_conv1_prev)[name = tensor("res1_conv1_prev_tmp")]; + tensor res1_conv1_first_tmp = identity(x = res1_conv1_first)[name = tensor("res1_conv1_first_tmp")]; + tensor res2_conv0_first_tmp = identity(x = res2_conv0_first)[name = tensor("res2_conv0_first_tmp")]; + tensor res2_conv1_prev_tmp = identity(x = res2_conv1_prev)[name = tensor("res2_conv1_prev_tmp")]; + tensor res2_conv1_first_tmp = identity(x = res2_conv1_first)[name = tensor("res2_conv1_first_tmp")]; + tensor conv_final_first_tmp = identity(x = conv_final_first)[name = tensor("conv_final_first_tmp")]; + } -> (var_724, var_77, var_210, var_741, var_400, var_744, var_542, conv0_first, var_565, var_585, res0_conv0_first, res0_conv1_prev, res0_conv1_first, var_621, var_641, res1_conv0_first, res1_conv1_prev, res1_conv1_first, var_677, var_697, res2_conv0_first, res2_conv1_prev, res2_conv1_first, var_725, conv_final_first); +} \ No newline at end of file diff --git a/v2.1/german_24l/mimi_decoder.mlmodelc/weights/weight.bin b/v2.1/german_24l/mimi_decoder.mlmodelc/weights/weight.bin new file mode 100644 index 0000000000000000000000000000000000000000..0bf03171362a32f81af7fa74d4b45c0b6c262a82 --- /dev/null +++ b/v2.1/german_24l/mimi_decoder.mlmodelc/weights/weight.bin @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:fc64ac703e883ade4ec052f3c14cc2a4489161bcabfdb24c6496e420eb3965b3 +size 41768256 diff --git a/v2.1/german_24l/mimi_decoder.mlpackage/Data/com.apple.CoreML/model.mlmodel b/v2.1/german_24l/mimi_decoder.mlpackage/Data/com.apple.CoreML/model.mlmodel new file mode 100644 index 0000000000000000000000000000000000000000..b24f8728f4431623117cc30c830b92105973747c --- /dev/null +++ b/v2.1/german_24l/mimi_decoder.mlpackage/Data/com.apple.CoreML/model.mlmodel @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:80ef9ab926a9e3f84726a6763e20cad9b8de95ca12a689fafeb32a754b8643f8 +size 93270 diff --git a/v2.1/german_24l/mimi_decoder.mlpackage/Data/com.apple.CoreML/weights/weight.bin b/v2.1/german_24l/mimi_decoder.mlpackage/Data/com.apple.CoreML/weights/weight.bin new file mode 100644 index 0000000000000000000000000000000000000000..0bf03171362a32f81af7fa74d4b45c0b6c262a82 --- /dev/null +++ b/v2.1/german_24l/mimi_decoder.mlpackage/Data/com.apple.CoreML/weights/weight.bin @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:fc64ac703e883ade4ec052f3c14cc2a4489161bcabfdb24c6496e420eb3965b3 +size 41768256 diff --git a/v2.1/german_24l/mimi_decoder.mlpackage/Manifest.json b/v2.1/german_24l/mimi_decoder.mlpackage/Manifest.json new file mode 100644 index 0000000000000000000000000000000000000000..54c7375ff5b579dec3adc50b4ef045a4ae2d07a4 --- /dev/null +++ b/v2.1/german_24l/mimi_decoder.mlpackage/Manifest.json @@ -0,0 +1,18 @@ +{ + "fileFormatVersion": "1.0.0", + "itemInfoEntries": { + "09EB6E89-4F59-40EE-8293-AA49C9DE537F": { + "author": "com.apple.CoreML", + "description": "CoreML Model Specification", + "name": "model.mlmodel", + "path": "com.apple.CoreML/model.mlmodel" + }, + "A161F999-BA19-464F-89F5-80919FFE2B87": { + "author": "com.apple.CoreML", + "description": "CoreML Model Weights", + "name": "weights", + "path": "com.apple.CoreML/weights" + } + }, + "rootModelIdentifier": "09EB6E89-4F59-40EE-8293-AA49C9DE537F" +} diff --git a/v2.1/italian/cond_prefill.mlmodelc/analytics/coremldata.bin b/v2.1/italian/cond_prefill.mlmodelc/analytics/coremldata.bin new file mode 100644 index 0000000000000000000000000000000000000000..c5b4a45600469433ea6a68c3692fd2fc71830b40 --- /dev/null +++ b/v2.1/italian/cond_prefill.mlmodelc/analytics/coremldata.bin @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:222d1ce60c65fb098049a076f06f9a23f8328886abc1101359a7223449a92a0a +size 243 diff --git a/v2.1/italian/cond_prefill.mlmodelc/coremldata.bin b/v2.1/italian/cond_prefill.mlmodelc/coremldata.bin new file mode 100644 index 0000000000000000000000000000000000000000..fca5c789b2a8e0e6d4d9570854256d19feb5a19b --- /dev/null +++ b/v2.1/italian/cond_prefill.mlmodelc/coremldata.bin @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:6a1e688cb261456d025b8dd2df27c701ff9b5d28f22b8cb5ba45043cf17f0c93 +size 1151 diff --git a/v2.1/italian/cond_prefill.mlmodelc/model.mil b/v2.1/italian/cond_prefill.mlmodelc/model.mil new file mode 100644 index 0000000000000000000000000000000000000000..203b091a2c5043d4b8c07f4402d5a5f0319b3c3f --- /dev/null +++ b/v2.1/italian/cond_prefill.mlmodelc/model.mil @@ -0,0 +1,1309 @@ +program(1.0) +[buildInfo = dict, tensor>({{"coremlc-component-MIL", "3520.4.1"}, {"coremlc-version", "3520.5.1"}, {"coremltools-component-torch", "2.12.0"}, {"coremltools-source-dialect", "TorchScript"}, {"coremltools-version", "9.0"}})] +{ + func main(tensor cache0, tensor cache1, tensor cache2, tensor cache3, tensor cache4, tensor cache5, tensor conditioning, tensor position0, tensor position1, tensor position2, tensor position3, tensor position4, tensor position5, tensor valid_len) { + tensor input_3_axes_0 = const()[name = tensor("input_3_axes_0"), val = tensor([-1])]; + tensor conditioning_to_fp16_dtype_0 = const()[name = tensor("conditioning_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor norm0_1_weight_to_fp16 = const()[name = tensor("norm0_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(64)))]; + tensor norm0_1_bias_to_fp16 = const()[name = tensor("norm0_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(2176)))]; + tensor var_48_to_fp16 = const()[name = tensor("op_48_to_fp16"), val = tensor(0x1.5p-17)]; + tensor conditioning_to_fp16 = cast(dtype = conditioning_to_fp16_dtype_0, x = conditioning)[name = tensor("cast_159")]; + tensor input_3_cast_fp16 = layer_norm(axes = input_3_axes_0, beta = norm0_1_bias_to_fp16, epsilon = var_48_to_fp16, gamma = norm0_1_weight_to_fp16, x = conditioning_to_fp16)[name = tensor("input_3_cast_fp16")]; + tensor attn0_in_proj_weight_to_fp16 = const()[name = tensor("attn0_in_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(4288)))]; + tensor linear_0_bias_0_to_fp16 = const()[name = tensor("linear_0_bias_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(6295808)))]; + tensor linear_0_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = attn0_in_proj_weight_to_fp16, x = input_3_cast_fp16)[name = tensor("linear_0_cast_fp16")]; + tensor var_62 = const()[name = tensor("op_62"), val = tensor([1, 256, 3, 16, 64])]; + tensor qkv_1_cast_fp16 = reshape(shape = var_62, x = linear_0_cast_fp16)[name = tensor("qkv_1_cast_fp16")]; + tensor q_1_begin_0 = const()[name = tensor("q_1_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor q_1_end_0 = const()[name = tensor("q_1_end_0"), val = tensor([1, 256, 1, 16, 64])]; + tensor q_1_end_mask_0 = const()[name = tensor("q_1_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor q_1_squeeze_mask_0 = const()[name = tensor("q_1_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor q_1_cast_fp16 = slice_by_index(begin = q_1_begin_0, end = q_1_end_0, end_mask = q_1_end_mask_0, squeeze_mask = q_1_squeeze_mask_0, x = qkv_1_cast_fp16)[name = tensor("q_1_cast_fp16")]; + tensor k_1_begin_0 = const()[name = tensor("k_1_begin_0"), val = tensor([0, 0, 1, 0, 0])]; + tensor k_1_end_0 = const()[name = tensor("k_1_end_0"), val = tensor([1, 256, 2, 16, 64])]; + tensor k_1_end_mask_0 = const()[name = tensor("k_1_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor k_1_squeeze_mask_0 = const()[name = tensor("k_1_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor k_1_cast_fp16 = slice_by_index(begin = k_1_begin_0, end = k_1_end_0, end_mask = k_1_end_mask_0, squeeze_mask = k_1_squeeze_mask_0, x = qkv_1_cast_fp16)[name = tensor("k_1_cast_fp16")]; + tensor v_1_begin_0 = const()[name = tensor("v_1_begin_0"), val = tensor([0, 0, 2, 0, 0])]; + tensor v_1_end_0 = const()[name = tensor("v_1_end_0"), val = tensor([1, 256, 3, 16, 64])]; + tensor v_1_end_mask_0 = const()[name = tensor("v_1_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor v_1_squeeze_mask_0 = const()[name = tensor("v_1_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor v_1_cast_fp16 = slice_by_index(begin = v_1_begin_0, end = v_1_end_0, end_mask = v_1_end_mask_0, squeeze_mask = v_1_squeeze_mask_0, x = qkv_1_cast_fp16)[name = tensor("v_1_cast_fp16")]; + tensor var_120 = const()[name = tensor("op_120"), val = tensor([1, 1])]; + tensor position0_to_fp16_dtype_0 = const()[name = tensor("position0_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor position0_to_fp16 = cast(dtype = position0_to_fp16_dtype_0, x = position0)[name = tensor("cast_158")]; + tensor var_121_cast_fp16 = reshape(shape = var_120, x = position0_to_fp16)[name = tensor("op_121_cast_fp16")]; + tensor ts_1_promoted_to_fp16 = const()[name = tensor("ts_1_promoted_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(6302016)))]; + tensor ts_3_cast_fp16 = add(x = ts_1_promoted_to_fp16, y = var_121_cast_fp16)[name = tensor("ts_3_cast_fp16")]; + tensor var_128 = const()[name = tensor("op_128"), val = tensor([1, 256, 1, 1])]; + tensor ts_5_cast_fp16 = reshape(shape = var_128, x = ts_3_cast_fp16)[name = tensor("ts_5_cast_fp16")]; + tensor var_135 = const()[name = tensor("op_135"), val = tensor([1, 256, 16, 32, 2])]; + tensor q_complex_1_cast_fp16 = reshape(shape = var_135, x = q_1_cast_fp16)[name = tensor("q_complex_1_cast_fp16")]; + tensor var_142 = const()[name = tensor("op_142"), val = tensor([1, 256, 16, 32, 2])]; + tensor k_complex_1_cast_fp16 = reshape(shape = var_142, x = k_1_cast_fp16)[name = tensor("k_complex_1_cast_fp16")]; + tensor var_146_begin_0 = const()[name = tensor("op_146_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_146_end_0 = const()[name = tensor("op_146_end_0"), val = tensor([1, 256, 16, 32, 1])]; + tensor var_146_end_mask_0 = const()[name = tensor("op_146_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_146_squeeze_mask_0 = const()[name = tensor("op_146_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_146_cast_fp16 = slice_by_index(begin = var_146_begin_0, end = var_146_end_0, end_mask = var_146_end_mask_0, squeeze_mask = var_146_squeeze_mask_0, x = q_complex_1_cast_fp16)[name = tensor("op_146_cast_fp16")]; + tensor var_154_begin_0 = const()[name = tensor("op_154_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_154_end_0 = const()[name = tensor("op_154_end_0"), val = tensor([1, 256, 16, 32, 2])]; + tensor var_154_end_mask_0 = const()[name = tensor("op_154_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_154_squeeze_mask_0 = const()[name = tensor("op_154_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_154_cast_fp16 = slice_by_index(begin = var_154_begin_0, end = var_154_end_0, end_mask = var_154_end_mask_0, squeeze_mask = var_154_squeeze_mask_0, x = q_complex_1_cast_fp16)[name = tensor("op_154_cast_fp16")]; + tensor var_162_begin_0 = const()[name = tensor("op_162_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_162_end_0 = const()[name = tensor("op_162_end_0"), val = tensor([1, 256, 16, 32, 1])]; + tensor var_162_end_mask_0 = const()[name = tensor("op_162_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_162_squeeze_mask_0 = const()[name = tensor("op_162_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_162_cast_fp16 = slice_by_index(begin = var_162_begin_0, end = var_162_end_0, end_mask = var_162_end_mask_0, squeeze_mask = var_162_squeeze_mask_0, x = k_complex_1_cast_fp16)[name = tensor("op_162_cast_fp16")]; + tensor var_170_begin_0 = const()[name = tensor("op_170_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_170_end_0 = const()[name = tensor("op_170_end_0"), val = tensor([1, 256, 16, 32, 2])]; + tensor var_170_end_mask_0 = const()[name = tensor("op_170_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_170_squeeze_mask_0 = const()[name = tensor("op_170_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_170_cast_fp16 = slice_by_index(begin = var_170_begin_0, end = var_170_end_0, end_mask = var_170_end_mask_0, squeeze_mask = var_170_squeeze_mask_0, x = k_complex_1_cast_fp16)[name = tensor("op_170_cast_fp16")]; + tensor freqs_1_to_fp16 = const()[name = tensor("freqs_1_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(6302592)))]; + tensor var_176_cast_fp16 = mul(x = freqs_1_to_fp16, y = ts_5_cast_fp16)[name = tensor("op_176_cast_fp16")]; + tensor rotr_1_cast_fp16 = cos(x = var_176_cast_fp16)[name = tensor("rotr_1_cast_fp16")]; + tensor roti_1_cast_fp16 = sin(x = var_176_cast_fp16)[name = tensor("roti_1_cast_fp16")]; + tensor var_180_cast_fp16 = mul(x = var_146_cast_fp16, y = rotr_1_cast_fp16)[name = tensor("op_180_cast_fp16")]; + tensor var_181_cast_fp16 = mul(x = var_154_cast_fp16, y = roti_1_cast_fp16)[name = tensor("op_181_cast_fp16")]; + tensor qor_1_cast_fp16 = sub(x = var_180_cast_fp16, y = var_181_cast_fp16)[name = tensor("qor_1_cast_fp16")]; + tensor var_184_cast_fp16 = mul(x = var_146_cast_fp16, y = roti_1_cast_fp16)[name = tensor("op_184_cast_fp16")]; + tensor var_185_cast_fp16 = mul(x = var_154_cast_fp16, y = rotr_1_cast_fp16)[name = tensor("op_185_cast_fp16")]; + tensor qoi_1_cast_fp16 = add(x = var_184_cast_fp16, y = var_185_cast_fp16)[name = tensor("qoi_1_cast_fp16")]; + tensor var_188_cast_fp16 = mul(x = var_162_cast_fp16, y = rotr_1_cast_fp16)[name = tensor("op_188_cast_fp16")]; + tensor var_189_cast_fp16 = mul(x = var_170_cast_fp16, y = roti_1_cast_fp16)[name = tensor("op_189_cast_fp16")]; + tensor kor_1_cast_fp16 = sub(x = var_188_cast_fp16, y = var_189_cast_fp16)[name = tensor("kor_1_cast_fp16")]; + tensor var_192_cast_fp16 = mul(x = var_162_cast_fp16, y = roti_1_cast_fp16)[name = tensor("op_192_cast_fp16")]; + tensor var_193_cast_fp16 = mul(x = var_170_cast_fp16, y = rotr_1_cast_fp16)[name = tensor("op_193_cast_fp16")]; + tensor koi_1_cast_fp16 = add(x = var_192_cast_fp16, y = var_193_cast_fp16)[name = tensor("koi_1_cast_fp16")]; + tensor qo_1_axis_0 = const()[name = tensor("qo_1_axis_0"), val = tensor(-1)]; + tensor qo_1_cast_fp16 = stack(axis = qo_1_axis_0, values = (qor_1_cast_fp16, qoi_1_cast_fp16))[name = tensor("qo_1_cast_fp16")]; + tensor ko_1_axis_0 = const()[name = tensor("ko_1_axis_0"), val = tensor(-1)]; + tensor ko_1_cast_fp16 = stack(axis = ko_1_axis_0, values = (kor_1_cast_fp16, koi_1_cast_fp16))[name = tensor("ko_1_cast_fp16")]; + tensor var_226 = const()[name = tensor("op_226"), val = tensor([1, 256, 16, 64])]; + tensor q_3_cast_fp16 = reshape(shape = var_226, x = qo_1_cast_fp16)[name = tensor("q_3_cast_fp16")]; + tensor var_232 = const()[name = tensor("op_232"), val = tensor([1, 256, 16, 64])]; + tensor k_3_cast_fp16 = reshape(shape = var_232, x = ko_1_cast_fp16)[name = tensor("k_3_cast_fp16")]; + tensor var_249_promoted_to_fp16 = const()[name = tensor("op_249_promoted_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(6302720)))]; + tensor write_indices_float_1_cast_fp16 = add(x = var_121_cast_fp16, y = var_249_promoted_to_fp16)[name = tensor("write_indices_float_1_cast_fp16")]; + tensor _inversed_253_y_0_to_fp16 = const()[name = tensor("_inversed_253_y_0_to_fp16"), val = tensor(0x1p-9)]; + tensor _inversed_253_cast_fp16 = mul(x = write_indices_float_1_cast_fp16, y = _inversed_253_y_0_to_fp16)[name = tensor("_inversed_253_cast_fp16")]; + tensor var_254_cast_fp16 = floor(x = _inversed_253_cast_fp16)[name = tensor("op_254_cast_fp16")]; + tensor var_255_to_fp16 = const()[name = tensor("op_255_to_fp16"), val = tensor(0x1p+9)]; + tensor var_256_cast_fp16 = mul(x = var_254_cast_fp16, y = var_255_to_fp16)[name = tensor("op_256_cast_fp16")]; + tensor write_indices_float_3_cast_fp16 = sub(x = write_indices_float_1_cast_fp16, y = var_256_cast_fp16)[name = tensor("write_indices_float_3_cast_fp16")]; + tensor var_265 = const()[name = tensor("op_265"), val = tensor([1, 1])]; + tensor valid_len_to_fp16_dtype_0 = const()[name = tensor("valid_len_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor valid_len_to_fp16 = cast(dtype = valid_len_to_fp16_dtype_0, x = valid_len)[name = tensor("cast_157")]; + tensor var_266_cast_fp16 = reshape(shape = var_265, x = valid_len_to_fp16)[name = tensor("op_266_cast_fp16")]; + tensor token_valid_1_cast_fp16 = less(x = var_249_promoted_to_fp16, y = var_266_cast_fp16)[name = tensor("token_valid_1_cast_fp16")]; + tensor var_274_value_0_to_fp16 = const()[name = tensor("op_274_value_0_to_fp16"), val = tensor(0x1.ffp+8)]; + tensor var_274_cast_fp16 = fill_like(ref_tensor = write_indices_float_3_cast_fp16, value = var_274_value_0_to_fp16)[name = tensor("op_274_cast_fp16")]; + tensor write_indices_float_5_cast_fp16 = select(a = write_indices_float_3_cast_fp16, b = var_274_cast_fp16, cond = token_valid_1_cast_fp16)[name = tensor("write_indices_float_5_cast_fp16")]; + tensor var_280_dtype_0 = const()[name = tensor("op_280_dtype_0"), val = tensor("int32")]; + tensor var_285 = const()[name = tensor("op_285"), val = tensor([1, 256, 1, 1])]; + tensor write_indices_float_5_cast_fp16_to_int32 = cast(dtype = var_280_dtype_0, x = write_indices_float_5_cast_fp16)[name = tensor("cast_156")]; + tensor var_286 = reshape(shape = var_285, x = write_indices_float_5_cast_fp16_to_int32)[name = tensor("op_286")]; + tensor write_indices_1_reps_0 = const()[name = tensor("write_indices_1_reps_0"), val = tensor([1, 1, 16, 64])]; + tensor write_indices_1 = tile(reps = write_indices_1_reps_0, x = var_286)[name = tensor("write_indices_1")]; + tensor var_296_begin_0 = const()[name = tensor("op_296_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_296_end_0 = const()[name = tensor("op_296_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor var_296_end_mask_0 = const()[name = tensor("op_296_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_296_squeeze_mask_0 = const()[name = tensor("op_296_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor cache0_to_fp16_dtype_0 = const()[name = tensor("cache0_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor cache0_to_fp16 = cast(dtype = cache0_to_fp16_dtype_0, x = cache0)[name = tensor("cast_155")]; + tensor var_296_cast_fp16 = slice_by_index(begin = var_296_begin_0, end = var_296_end_0, end_mask = var_296_end_mask_0, squeeze_mask = var_296_squeeze_mask_0, x = cache0_to_fp16)[name = tensor("op_296_cast_fp16")]; + tensor var_298_axis_0 = const()[name = tensor("op_298_axis_0"), val = tensor(1)]; + tensor var_298_mode_0 = const()[name = tensor("op_298_mode_0"), val = tensor("update")]; + tensor var_298_validate_indices_0 = const()[name = tensor("op_298_validate_indices_0"), val = tensor(false)]; + tensor var_298_cast_fp16 = scatter_along_axis(axis = var_298_axis_0, data = var_296_cast_fp16, indices = write_indices_1, mode = var_298_mode_0, updates = k_3_cast_fp16, validate_indices = var_298_validate_indices_0)[name = tensor("op_298_cast_fp16")]; + tensor concat_2 = const()[name = tensor("concat_2"), val = tensor([0, 0, 0, 0, 0])]; + tensor concat_3 = const()[name = tensor("concat_3"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_1_internal_tensor_assign_1_stride_0 = const()[name = tensor("new_cache_1_internal_tensor_assign_1_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_1_internal_tensor_assign_1_begin_mask_0 = const()[name = tensor("new_cache_1_internal_tensor_assign_1_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_1_internal_tensor_assign_1_end_mask_0 = const()[name = tensor("new_cache_1_internal_tensor_assign_1_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_1_internal_tensor_assign_1_squeeze_mask_0 = const()[name = tensor("new_cache_1_internal_tensor_assign_1_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_10 = const()[name = tensor("shape_10"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_0 = const()[name = tensor("reduce_prod_0"), val = tensor(1048576)]; + tensor range_1d_0_start_0 = const()[name = tensor("range_1d_0_start_0"), val = tensor(0)]; + tensor range_1d_0_step_0 = const()[name = tensor("range_1d_0_step_0"), val = tensor(1)]; + tensor range_1d_0 = range_1d(end = reduce_prod_0, start = range_1d_0_start_0, step = range_1d_0_step_0)[name = tensor("range_1d_0")]; + tensor reshape_0 = reshape(shape = shape_10, x = range_1d_0)[name = tensor("reshape_0")]; + tensor slice_by_index_0 = slice_by_index(begin = concat_2, begin_mask = new_cache_1_internal_tensor_assign_1_begin_mask_0, end = concat_3, end_mask = new_cache_1_internal_tensor_assign_1_end_mask_0, squeeze_mask = new_cache_1_internal_tensor_assign_1_squeeze_mask_0, stride = new_cache_1_internal_tensor_assign_1_stride_0, x = reshape_0)[name = tensor("slice_by_index_0")]; + tensor reshape_1_shape_0 = const()[name = tensor("reshape_1_shape_0"), val = tensor([-1])]; + tensor reshape_1 = reshape(shape = reshape_1_shape_0, x = slice_by_index_0)[name = tensor("reshape_1")]; + tensor reshape_2_shape_0 = const()[name = tensor("reshape_2_shape_0"), val = tensor([-1])]; + tensor reshape_2_cast_fp16 = reshape(shape = reshape_2_shape_0, x = var_298_cast_fp16)[name = tensor("reshape_2_cast_fp16")]; + tensor reshape_3_shape_0 = const()[name = tensor("reshape_3_shape_0"), val = tensor([-1])]; + tensor reshape_3_cast_fp16 = reshape(shape = reshape_3_shape_0, x = cache0_to_fp16)[name = tensor("reshape_3_cast_fp16")]; + tensor scatter_0_mode_0 = const()[name = tensor("scatter_0_mode_0"), val = tensor("update")]; + tensor scatter_0_axis_0 = const()[name = tensor("scatter_0_axis_0"), val = tensor(0)]; + tensor scatter_0_validate_indices_0 = const()[name = tensor("scatter_0_validate_indices_0"), val = tensor(false)]; + tensor scatter_0_cast_fp16 = scatter(axis = scatter_0_axis_0, data = reshape_3_cast_fp16, indices = reshape_1, mode = scatter_0_mode_0, updates = reshape_2_cast_fp16, validate_indices = scatter_0_validate_indices_0)[name = tensor("scatter_0_cast_fp16")]; + tensor reshape_4_cast_fp16 = reshape(shape = shape_10, x = scatter_0_cast_fp16)[name = tensor("reshape_4_cast_fp16")]; + tensor var_306_begin_0 = const()[name = tensor("op_306_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor var_306_end_0 = const()[name = tensor("op_306_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor var_306_end_mask_0 = const()[name = tensor("op_306_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_306_squeeze_mask_0 = const()[name = tensor("op_306_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor var_306_cast_fp16 = slice_by_index(begin = var_306_begin_0, end = var_306_end_0, end_mask = var_306_end_mask_0, squeeze_mask = var_306_squeeze_mask_0, x = reshape_4_cast_fp16)[name = tensor("op_306_cast_fp16")]; + tensor var_308_axis_0 = const()[name = tensor("op_308_axis_0"), val = tensor(1)]; + tensor var_308_mode_0 = const()[name = tensor("op_308_mode_0"), val = tensor("update")]; + tensor var_308_validate_indices_0 = const()[name = tensor("op_308_validate_indices_0"), val = tensor(false)]; + tensor var_308_cast_fp16 = scatter_along_axis(axis = var_308_axis_0, data = var_306_cast_fp16, indices = write_indices_1, mode = var_308_mode_0, updates = v_1_cast_fp16, validate_indices = var_308_validate_indices_0)[name = tensor("op_308_cast_fp16")]; + tensor concat_4 = const()[name = tensor("concat_4"), val = tensor([1, 0, 0, 0, 0])]; + tensor concat_5 = const()[name = tensor("concat_5"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_1_internal_tensor_assign_2_stride_0 = const()[name = tensor("new_cache_1_internal_tensor_assign_2_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_1_internal_tensor_assign_2_begin_mask_0 = const()[name = tensor("new_cache_1_internal_tensor_assign_2_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_1_internal_tensor_assign_2_end_mask_0 = const()[name = tensor("new_cache_1_internal_tensor_assign_2_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_1_internal_tensor_assign_2_squeeze_mask_0 = const()[name = tensor("new_cache_1_internal_tensor_assign_2_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_11 = const()[name = tensor("shape_11"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_1 = const()[name = tensor("reduce_prod_1"), val = tensor(1048576)]; + tensor range_1d_1_start_0 = const()[name = tensor("range_1d_1_start_0"), val = tensor(0)]; + tensor range_1d_1_step_0 = const()[name = tensor("range_1d_1_step_0"), val = tensor(1)]; + tensor range_1d_1 = range_1d(end = reduce_prod_1, start = range_1d_1_start_0, step = range_1d_1_step_0)[name = tensor("range_1d_1")]; + tensor reshape_5 = reshape(shape = shape_11, x = range_1d_1)[name = tensor("reshape_5")]; + tensor slice_by_index_1 = slice_by_index(begin = concat_4, begin_mask = new_cache_1_internal_tensor_assign_2_begin_mask_0, end = concat_5, end_mask = new_cache_1_internal_tensor_assign_2_end_mask_0, squeeze_mask = new_cache_1_internal_tensor_assign_2_squeeze_mask_0, stride = new_cache_1_internal_tensor_assign_2_stride_0, x = reshape_5)[name = tensor("slice_by_index_1")]; + tensor reshape_6_shape_0 = const()[name = tensor("reshape_6_shape_0"), val = tensor([-1])]; + tensor reshape_6 = reshape(shape = reshape_6_shape_0, x = slice_by_index_1)[name = tensor("reshape_6")]; + tensor reshape_7_shape_0 = const()[name = tensor("reshape_7_shape_0"), val = tensor([-1])]; + tensor reshape_7_cast_fp16 = reshape(shape = reshape_7_shape_0, x = var_308_cast_fp16)[name = tensor("reshape_7_cast_fp16")]; + tensor reshape_8_shape_0 = const()[name = tensor("reshape_8_shape_0"), val = tensor([-1])]; + tensor reshape_8_cast_fp16 = reshape(shape = reshape_8_shape_0, x = reshape_4_cast_fp16)[name = tensor("reshape_8_cast_fp16")]; + tensor scatter_1_mode_0 = const()[name = tensor("scatter_1_mode_0"), val = tensor("update")]; + tensor scatter_1_axis_0 = const()[name = tensor("scatter_1_axis_0"), val = tensor(0)]; + tensor scatter_1_validate_indices_0 = const()[name = tensor("scatter_1_validate_indices_0"), val = tensor(false)]; + tensor scatter_1_cast_fp16 = scatter(axis = scatter_1_axis_0, data = reshape_8_cast_fp16, indices = reshape_6, mode = scatter_1_mode_0, updates = reshape_7_cast_fp16, validate_indices = scatter_1_validate_indices_0)[name = tensor("scatter_1_cast_fp16")]; + tensor reshape_9_cast_fp16 = reshape(shape = shape_11, x = scatter_1_cast_fp16)[name = tensor("reshape_9_cast_fp16")]; + tensor reshape_9_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("reshape_9_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor keys_1_begin_0 = const()[name = tensor("keys_1_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor keys_1_end_0 = const()[name = tensor("keys_1_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor keys_1_end_mask_0 = const()[name = tensor("keys_1_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor keys_1_squeeze_mask_0 = const()[name = tensor("keys_1_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor keys_1_cast_fp16 = slice_by_index(begin = keys_1_begin_0, end = keys_1_end_0, end_mask = keys_1_end_mask_0, squeeze_mask = keys_1_squeeze_mask_0, x = reshape_9_cast_fp16)[name = tensor("keys_1_cast_fp16")]; + tensor values_1_begin_0 = const()[name = tensor("values_1_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor values_1_end_0 = const()[name = tensor("values_1_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor values_1_end_mask_0 = const()[name = tensor("values_1_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor values_1_squeeze_mask_0 = const()[name = tensor("values_1_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor values_1_cast_fp16 = slice_by_index(begin = values_1_begin_0, end = values_1_end_0, end_mask = values_1_end_mask_0, squeeze_mask = values_1_squeeze_mask_0, x = reshape_9_cast_fp16)[name = tensor("values_1_cast_fp16")]; + tensor var_320_cast_fp16 = not_equal(x = keys_1_cast_fp16, y = keys_1_cast_fp16)[name = tensor("op_320_cast_fp16")]; + tensor var_326_to_fp16 = const()[name = tensor("op_326_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(6303296)))]; + tensor keys_3_cast_fp16 = select(a = var_326_to_fp16, b = keys_1_cast_fp16, cond = var_320_cast_fp16)[name = tensor("keys_3_cast_fp16")]; + tensor var_328_cast_fp16 = not_equal(x = values_1_cast_fp16, y = values_1_cast_fp16)[name = tensor("op_328_cast_fp16")]; + tensor values_3_cast_fp16 = select(a = var_326_to_fp16, b = values_1_cast_fp16, cond = var_328_cast_fp16)[name = tensor("values_3_cast_fp16")]; + tensor var_352 = const()[name = tensor("op_352"), val = tensor([0, 2, 1, 3])]; + tensor var_368 = const()[name = tensor("op_368"), val = tensor([1, 1, 1])]; + tensor var_369_cast_fp16 = reshape(shape = var_368, x = position0_to_fp16)[name = tensor("op_369_cast_fp16")]; + tensor q_offsets_1_promoted_to_fp16 = const()[name = tensor("q_offsets_1_promoted_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(7351936)))]; + tensor q_positions_1_cast_fp16 = add(x = var_369_cast_fp16, y = q_offsets_1_promoted_to_fp16)[name = tensor("q_positions_1_cast_fp16")]; + tensor var_391 = const()[name = tensor("op_391"), val = tensor([1, 1, 1])]; + tensor var_392_cast_fp16 = reshape(shape = var_391, x = valid_len_to_fp16)[name = tensor("op_392_cast_fp16")]; + tensor valid_end_1_cast_fp16 = add(x = var_369_cast_fp16, y = var_392_cast_fp16)[name = tensor("valid_end_1_cast_fp16")]; + tensor k_positions_1_promoted_to_fp16 = const()[name = tensor("k_positions_1_promoted_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(7352512)))]; + tensor valid_mask_1_cast_fp16 = less(x = k_positions_1_promoted_to_fp16, y = valid_end_1_cast_fp16)[name = tensor("valid_mask_1_cast_fp16")]; + tensor causal_mask_1_cast_fp16 = less_equal(x = k_positions_1_promoted_to_fp16, y = q_positions_1_cast_fp16)[name = tensor("causal_mask_1_cast_fp16")]; + tensor var_397 = logical_and(x = valid_mask_1_cast_fp16, y = causal_mask_1_cast_fp16)[name = tensor("op_397")]; + tensor attn_mask_1_axes_0 = const()[name = tensor("attn_mask_1_axes_0"), val = tensor([1])]; + tensor attn_mask_1 = expand_dims(axes = attn_mask_1_axes_0, x = var_397)[name = tensor("attn_mask_1")]; + tensor var_411_transpose_x_0 = const()[name = tensor("op_411_transpose_x_0"), val = tensor(false)]; + tensor var_411_transpose_y_0 = const()[name = tensor("op_411_transpose_y_0"), val = tensor(false)]; + tensor transpose_15_perm_0 = const()[name = tensor("transpose_15_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_16_perm_0 = const()[name = tensor("transpose_16_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_16 = transpose(perm = transpose_16_perm_0, x = keys_3_cast_fp16)[name = tensor("transpose_42")]; + tensor transpose_15 = transpose(perm = transpose_15_perm_0, x = q_3_cast_fp16)[name = tensor("transpose_43")]; + tensor var_411_cast_fp16 = matmul(transpose_x = var_411_transpose_x_0, transpose_y = var_411_transpose_y_0, x = transpose_15, y = transpose_16)[name = tensor("op_411_cast_fp16")]; + tensor var_405_to_fp16 = const()[name = tensor("op_405_to_fp16"), val = tensor([0x1p-3])]; + tensor attn_weights_1_cast_fp16 = mul(x = var_411_cast_fp16, y = var_405_to_fp16)[name = tensor("attn_weights_1_cast_fp16")]; + tensor var_413 = logical_not(x = attn_mask_1)[name = tensor("op_413")]; + tensor var_414_to_fp16 = const()[name = tensor("op_414_to_fp16"), val = tensor(-inf)]; + tensor attn_weights_3_cast_fp16 = select(a = var_414_to_fp16, b = attn_weights_1_cast_fp16, cond = var_413)[name = tensor("attn_weights_3_cast_fp16")]; + tensor var_416 = const()[name = tensor("op_416"), val = tensor(-1)]; + tensor attn_weights_5_cast_fp16 = softmax(axis = var_416, x = attn_weights_3_cast_fp16)[name = tensor("attn_weights_5_cast_fp16")]; + tensor attn_output_1_transpose_x_0 = const()[name = tensor("attn_output_1_transpose_x_0"), val = tensor(false)]; + tensor attn_output_1_transpose_y_0 = const()[name = tensor("attn_output_1_transpose_y_0"), val = tensor(false)]; + tensor values_5_cast_fp16 = transpose(perm = var_352, x = values_3_cast_fp16)[name = tensor("transpose_44")]; + tensor attn_output_1_cast_fp16 = matmul(transpose_x = attn_output_1_transpose_x_0, transpose_y = attn_output_1_transpose_y_0, x = attn_weights_5_cast_fp16, y = values_5_cast_fp16)[name = tensor("attn_output_1_cast_fp16")]; + tensor var_424 = const()[name = tensor("op_424"), val = tensor([0, 2, 1, 3])]; + tensor var_429 = const()[name = tensor("op_429"), val = tensor([1, 256, 1024])]; + tensor var_425_cast_fp16 = transpose(perm = var_424, x = attn_output_1_cast_fp16)[name = tensor("transpose_41")]; + tensor input_5_cast_fp16 = reshape(shape = var_429, x = var_425_cast_fp16)[name = tensor("input_5_cast_fp16")]; + tensor attn0_out_proj_weight_to_fp16 = const()[name = tensor("attn0_out_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(7353600)))]; + tensor linear_1_bias_0_to_fp16 = const()[name = tensor("linear_1_bias_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9450816)))]; + tensor linear_1_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = attn0_out_proj_weight_to_fp16, x = input_5_cast_fp16)[name = tensor("linear_1_cast_fp16")]; + tensor var_438_cast_fp16 = add(x = position0_to_fp16, y = valid_len_to_fp16)[name = tensor("op_438_cast_fp16")]; + tensor var_438_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_438_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor input_7_cast_fp16 = add(x = conditioning_to_fp16, y = linear_1_cast_fp16)[name = tensor("input_7_cast_fp16")]; + tensor input_9_axes_0 = const()[name = tensor("input_9_axes_0"), val = tensor([-1])]; + tensor norm0_2_weight_to_fp16 = const()[name = tensor("norm0_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9452928)))]; + tensor norm0_2_bias_to_fp16 = const()[name = tensor("norm0_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9455040)))]; + tensor var_442_to_fp16 = const()[name = tensor("op_442_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_9_cast_fp16 = layer_norm(axes = input_9_axes_0, beta = norm0_2_bias_to_fp16, epsilon = var_442_to_fp16, gamma = norm0_2_weight_to_fp16, x = input_7_cast_fp16)[name = tensor("input_9_cast_fp16")]; + tensor linear0_1_weight_to_fp16 = const()[name = tensor("linear0_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9457152)))]; + tensor linear_2_bias_0_to_fp16 = const()[name = tensor("linear_2_bias_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(17845824)))]; + tensor linear_2_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = linear0_1_weight_to_fp16, x = input_9_cast_fp16)[name = tensor("linear_2_cast_fp16")]; + tensor input_11_mode_0 = const()[name = tensor("input_11_mode_0"), val = tensor("EXACT")]; + tensor input_11_cast_fp16 = gelu(mode = input_11_mode_0, x = linear_2_cast_fp16)[name = tensor("input_11_cast_fp16")]; + tensor linear0_2_weight_to_fp16 = const()[name = tensor("linear0_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(17854080)))]; + tensor linear_3_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = linear0_2_weight_to_fp16, x = input_11_cast_fp16)[name = tensor("linear_3_cast_fp16")]; + tensor input_13_cast_fp16 = add(x = input_7_cast_fp16, y = linear_3_cast_fp16)[name = tensor("input_13_cast_fp16")]; + tensor input_15_axes_0 = const()[name = tensor("input_15_axes_0"), val = tensor([-1])]; + tensor norm1_1_weight_to_fp16 = const()[name = tensor("norm1_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(26242752)))]; + tensor norm1_1_bias_to_fp16 = const()[name = tensor("norm1_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(26244864)))]; + tensor var_459_to_fp16 = const()[name = tensor("op_459_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_15_cast_fp16 = layer_norm(axes = input_15_axes_0, beta = norm1_1_bias_to_fp16, epsilon = var_459_to_fp16, gamma = norm1_1_weight_to_fp16, x = input_13_cast_fp16)[name = tensor("input_15_cast_fp16")]; + tensor attn1_in_proj_weight_to_fp16 = const()[name = tensor("attn1_in_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(26246976)))]; + tensor linear_4_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = attn1_in_proj_weight_to_fp16, x = input_15_cast_fp16)[name = tensor("linear_4_cast_fp16")]; + tensor var_473 = const()[name = tensor("op_473"), val = tensor([1, 256, 3, 16, 64])]; + tensor qkv_3_cast_fp16 = reshape(shape = var_473, x = linear_4_cast_fp16)[name = tensor("qkv_3_cast_fp16")]; + tensor q_7_begin_0 = const()[name = tensor("q_7_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor q_7_end_0 = const()[name = tensor("q_7_end_0"), val = tensor([1, 256, 1, 16, 64])]; + tensor q_7_end_mask_0 = const()[name = tensor("q_7_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor q_7_squeeze_mask_0 = const()[name = tensor("q_7_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor q_7_cast_fp16 = slice_by_index(begin = q_7_begin_0, end = q_7_end_0, end_mask = q_7_end_mask_0, squeeze_mask = q_7_squeeze_mask_0, x = qkv_3_cast_fp16)[name = tensor("q_7_cast_fp16")]; + tensor k_5_begin_0 = const()[name = tensor("k_5_begin_0"), val = tensor([0, 0, 1, 0, 0])]; + tensor k_5_end_0 = const()[name = tensor("k_5_end_0"), val = tensor([1, 256, 2, 16, 64])]; + tensor k_5_end_mask_0 = const()[name = tensor("k_5_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor k_5_squeeze_mask_0 = const()[name = tensor("k_5_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor k_5_cast_fp16 = slice_by_index(begin = k_5_begin_0, end = k_5_end_0, end_mask = k_5_end_mask_0, squeeze_mask = k_5_squeeze_mask_0, x = qkv_3_cast_fp16)[name = tensor("k_5_cast_fp16")]; + tensor v_3_begin_0 = const()[name = tensor("v_3_begin_0"), val = tensor([0, 0, 2, 0, 0])]; + tensor v_3_end_0 = const()[name = tensor("v_3_end_0"), val = tensor([1, 256, 3, 16, 64])]; + tensor v_3_end_mask_0 = const()[name = tensor("v_3_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor v_3_squeeze_mask_0 = const()[name = tensor("v_3_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor v_3_cast_fp16 = slice_by_index(begin = v_3_begin_0, end = v_3_end_0, end_mask = v_3_end_mask_0, squeeze_mask = v_3_squeeze_mask_0, x = qkv_3_cast_fp16)[name = tensor("v_3_cast_fp16")]; + tensor var_531 = const()[name = tensor("op_531"), val = tensor([1, 1])]; + tensor position1_to_fp16_dtype_0 = const()[name = tensor("position1_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor position1_to_fp16 = cast(dtype = position1_to_fp16_dtype_0, x = position1)[name = tensor("cast_152")]; + tensor var_532_cast_fp16 = reshape(shape = var_531, x = position1_to_fp16)[name = tensor("op_532_cast_fp16")]; + tensor ts_9_cast_fp16 = add(x = ts_1_promoted_to_fp16, y = var_532_cast_fp16)[name = tensor("ts_9_cast_fp16")]; + tensor var_539 = const()[name = tensor("op_539"), val = tensor([1, 256, 1, 1])]; + tensor ts_11_cast_fp16 = reshape(shape = var_539, x = ts_9_cast_fp16)[name = tensor("ts_11_cast_fp16")]; + tensor var_546 = const()[name = tensor("op_546"), val = tensor([1, 256, 16, 32, 2])]; + tensor q_complex_3_cast_fp16 = reshape(shape = var_546, x = q_7_cast_fp16)[name = tensor("q_complex_3_cast_fp16")]; + tensor var_553 = const()[name = tensor("op_553"), val = tensor([1, 256, 16, 32, 2])]; + tensor k_complex_3_cast_fp16 = reshape(shape = var_553, x = k_5_cast_fp16)[name = tensor("k_complex_3_cast_fp16")]; + tensor var_557_begin_0 = const()[name = tensor("op_557_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_557_end_0 = const()[name = tensor("op_557_end_0"), val = tensor([1, 256, 16, 32, 1])]; + tensor var_557_end_mask_0 = const()[name = tensor("op_557_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_557_squeeze_mask_0 = const()[name = tensor("op_557_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_557_cast_fp16 = slice_by_index(begin = var_557_begin_0, end = var_557_end_0, end_mask = var_557_end_mask_0, squeeze_mask = var_557_squeeze_mask_0, x = q_complex_3_cast_fp16)[name = tensor("op_557_cast_fp16")]; + tensor var_565_begin_0 = const()[name = tensor("op_565_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_565_end_0 = const()[name = tensor("op_565_end_0"), val = tensor([1, 256, 16, 32, 2])]; + tensor var_565_end_mask_0 = const()[name = tensor("op_565_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_565_squeeze_mask_0 = const()[name = tensor("op_565_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_565_cast_fp16 = slice_by_index(begin = var_565_begin_0, end = var_565_end_0, end_mask = var_565_end_mask_0, squeeze_mask = var_565_squeeze_mask_0, x = q_complex_3_cast_fp16)[name = tensor("op_565_cast_fp16")]; + tensor var_573_begin_0 = const()[name = tensor("op_573_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_573_end_0 = const()[name = tensor("op_573_end_0"), val = tensor([1, 256, 16, 32, 1])]; + tensor var_573_end_mask_0 = const()[name = tensor("op_573_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_573_squeeze_mask_0 = const()[name = tensor("op_573_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_573_cast_fp16 = slice_by_index(begin = var_573_begin_0, end = var_573_end_0, end_mask = var_573_end_mask_0, squeeze_mask = var_573_squeeze_mask_0, x = k_complex_3_cast_fp16)[name = tensor("op_573_cast_fp16")]; + tensor var_581_begin_0 = const()[name = tensor("op_581_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_581_end_0 = const()[name = tensor("op_581_end_0"), val = tensor([1, 256, 16, 32, 2])]; + tensor var_581_end_mask_0 = const()[name = tensor("op_581_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_581_squeeze_mask_0 = const()[name = tensor("op_581_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_581_cast_fp16 = slice_by_index(begin = var_581_begin_0, end = var_581_end_0, end_mask = var_581_end_mask_0, squeeze_mask = var_581_squeeze_mask_0, x = k_complex_3_cast_fp16)[name = tensor("op_581_cast_fp16")]; + tensor freqs_3_to_fp16 = const()[name = tensor("freqs_3_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(32538496)))]; + tensor var_587_cast_fp16 = mul(x = freqs_3_to_fp16, y = ts_11_cast_fp16)[name = tensor("op_587_cast_fp16")]; + tensor rotr_3_cast_fp16 = cos(x = var_587_cast_fp16)[name = tensor("rotr_3_cast_fp16")]; + tensor roti_3_cast_fp16 = sin(x = var_587_cast_fp16)[name = tensor("roti_3_cast_fp16")]; + tensor var_591_cast_fp16 = mul(x = var_557_cast_fp16, y = rotr_3_cast_fp16)[name = tensor("op_591_cast_fp16")]; + tensor var_592_cast_fp16 = mul(x = var_565_cast_fp16, y = roti_3_cast_fp16)[name = tensor("op_592_cast_fp16")]; + tensor qor_5_cast_fp16 = sub(x = var_591_cast_fp16, y = var_592_cast_fp16)[name = tensor("qor_5_cast_fp16")]; + tensor var_595_cast_fp16 = mul(x = var_557_cast_fp16, y = roti_3_cast_fp16)[name = tensor("op_595_cast_fp16")]; + tensor var_596_cast_fp16 = mul(x = var_565_cast_fp16, y = rotr_3_cast_fp16)[name = tensor("op_596_cast_fp16")]; + tensor qoi_5_cast_fp16 = add(x = var_595_cast_fp16, y = var_596_cast_fp16)[name = tensor("qoi_5_cast_fp16")]; + tensor var_599_cast_fp16 = mul(x = var_573_cast_fp16, y = rotr_3_cast_fp16)[name = tensor("op_599_cast_fp16")]; + tensor var_600_cast_fp16 = mul(x = var_581_cast_fp16, y = roti_3_cast_fp16)[name = tensor("op_600_cast_fp16")]; + tensor kor_5_cast_fp16 = sub(x = var_599_cast_fp16, y = var_600_cast_fp16)[name = tensor("kor_5_cast_fp16")]; + tensor var_603_cast_fp16 = mul(x = var_573_cast_fp16, y = roti_3_cast_fp16)[name = tensor("op_603_cast_fp16")]; + tensor var_604_cast_fp16 = mul(x = var_581_cast_fp16, y = rotr_3_cast_fp16)[name = tensor("op_604_cast_fp16")]; + tensor koi_5_cast_fp16 = add(x = var_603_cast_fp16, y = var_604_cast_fp16)[name = tensor("koi_5_cast_fp16")]; + tensor qo_3_axis_0 = const()[name = tensor("qo_3_axis_0"), val = tensor(-1)]; + tensor qo_3_cast_fp16 = stack(axis = qo_3_axis_0, values = (qor_5_cast_fp16, qoi_5_cast_fp16))[name = tensor("qo_3_cast_fp16")]; + tensor ko_3_axis_0 = const()[name = tensor("ko_3_axis_0"), val = tensor(-1)]; + tensor ko_3_cast_fp16 = stack(axis = ko_3_axis_0, values = (kor_5_cast_fp16, koi_5_cast_fp16))[name = tensor("ko_3_cast_fp16")]; + tensor var_637 = const()[name = tensor("op_637"), val = tensor([1, 256, 16, 64])]; + tensor q_9_cast_fp16 = reshape(shape = var_637, x = qo_3_cast_fp16)[name = tensor("q_9_cast_fp16")]; + tensor var_643 = const()[name = tensor("op_643"), val = tensor([1, 256, 16, 64])]; + tensor k_7_cast_fp16 = reshape(shape = var_643, x = ko_3_cast_fp16)[name = tensor("k_7_cast_fp16")]; + tensor write_indices_float_7_cast_fp16 = add(x = var_532_cast_fp16, y = var_249_promoted_to_fp16)[name = tensor("write_indices_float_7_cast_fp16")]; + tensor _inversed_664_y_0_to_fp16 = const()[name = tensor("_inversed_664_y_0_to_fp16"), val = tensor(0x1p-9)]; + tensor _inversed_664_cast_fp16 = mul(x = write_indices_float_7_cast_fp16, y = _inversed_664_y_0_to_fp16)[name = tensor("_inversed_664_cast_fp16")]; + tensor var_665_cast_fp16 = floor(x = _inversed_664_cast_fp16)[name = tensor("op_665_cast_fp16")]; + tensor var_666_to_fp16 = const()[name = tensor("op_666_to_fp16"), val = tensor(0x1p+9)]; + tensor var_667_cast_fp16 = mul(x = var_665_cast_fp16, y = var_666_to_fp16)[name = tensor("op_667_cast_fp16")]; + tensor write_indices_float_9_cast_fp16 = sub(x = write_indices_float_7_cast_fp16, y = var_667_cast_fp16)[name = tensor("write_indices_float_9_cast_fp16")]; + tensor var_685_value_0_to_fp16 = const()[name = tensor("op_685_value_0_to_fp16"), val = tensor(0x1.ffp+8)]; + tensor var_685_cast_fp16 = fill_like(ref_tensor = write_indices_float_9_cast_fp16, value = var_685_value_0_to_fp16)[name = tensor("op_685_cast_fp16")]; + tensor write_indices_float_11_cast_fp16 = select(a = write_indices_float_9_cast_fp16, b = var_685_cast_fp16, cond = token_valid_1_cast_fp16)[name = tensor("write_indices_float_11_cast_fp16")]; + tensor var_691_dtype_0 = const()[name = tensor("op_691_dtype_0"), val = tensor("int32")]; + tensor var_696 = const()[name = tensor("op_696"), val = tensor([1, 256, 1, 1])]; + tensor write_indices_float_11_cast_fp16_to_int32 = cast(dtype = var_691_dtype_0, x = write_indices_float_11_cast_fp16)[name = tensor("cast_151")]; + tensor var_697 = reshape(shape = var_696, x = write_indices_float_11_cast_fp16_to_int32)[name = tensor("op_697")]; + tensor write_indices_3_reps_0 = const()[name = tensor("write_indices_3_reps_0"), val = tensor([1, 1, 16, 64])]; + tensor write_indices_3 = tile(reps = write_indices_3_reps_0, x = var_697)[name = tensor("write_indices_3")]; + tensor var_707_begin_0 = const()[name = tensor("op_707_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_707_end_0 = const()[name = tensor("op_707_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor var_707_end_mask_0 = const()[name = tensor("op_707_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_707_squeeze_mask_0 = const()[name = tensor("op_707_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor cache1_to_fp16_dtype_0 = const()[name = tensor("cache1_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor cache1_to_fp16 = cast(dtype = cache1_to_fp16_dtype_0, x = cache1)[name = tensor("cast_150")]; + tensor var_707_cast_fp16 = slice_by_index(begin = var_707_begin_0, end = var_707_end_0, end_mask = var_707_end_mask_0, squeeze_mask = var_707_squeeze_mask_0, x = cache1_to_fp16)[name = tensor("op_707_cast_fp16")]; + tensor var_709_axis_0 = const()[name = tensor("op_709_axis_0"), val = tensor(1)]; + tensor var_709_mode_0 = const()[name = tensor("op_709_mode_0"), val = tensor("update")]; + tensor var_709_validate_indices_0 = const()[name = tensor("op_709_validate_indices_0"), val = tensor(false)]; + tensor var_709_cast_fp16 = scatter_along_axis(axis = var_709_axis_0, data = var_707_cast_fp16, indices = write_indices_3, mode = var_709_mode_0, updates = k_7_cast_fp16, validate_indices = var_709_validate_indices_0)[name = tensor("op_709_cast_fp16")]; + tensor concat_10 = const()[name = tensor("concat_10"), val = tensor([0, 0, 0, 0, 0])]; + tensor concat_11 = const()[name = tensor("concat_11"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_3_internal_tensor_assign_1_stride_0 = const()[name = tensor("new_cache_3_internal_tensor_assign_1_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_3_internal_tensor_assign_1_begin_mask_0 = const()[name = tensor("new_cache_3_internal_tensor_assign_1_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_3_internal_tensor_assign_1_end_mask_0 = const()[name = tensor("new_cache_3_internal_tensor_assign_1_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_3_internal_tensor_assign_1_squeeze_mask_0 = const()[name = tensor("new_cache_3_internal_tensor_assign_1_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_12 = const()[name = tensor("shape_12"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_2 = const()[name = tensor("reduce_prod_2"), val = tensor(1048576)]; + tensor range_1d_2_start_0 = const()[name = tensor("range_1d_2_start_0"), val = tensor(0)]; + tensor range_1d_2_step_0 = const()[name = tensor("range_1d_2_step_0"), val = tensor(1)]; + tensor range_1d_2 = range_1d(end = reduce_prod_2, start = range_1d_2_start_0, step = range_1d_2_step_0)[name = tensor("range_1d_2")]; + tensor reshape_10 = reshape(shape = shape_12, x = range_1d_2)[name = tensor("reshape_10")]; + tensor slice_by_index_2 = slice_by_index(begin = concat_10, begin_mask = new_cache_3_internal_tensor_assign_1_begin_mask_0, end = concat_11, end_mask = new_cache_3_internal_tensor_assign_1_end_mask_0, squeeze_mask = new_cache_3_internal_tensor_assign_1_squeeze_mask_0, stride = new_cache_3_internal_tensor_assign_1_stride_0, x = reshape_10)[name = tensor("slice_by_index_2")]; + tensor reshape_11_shape_0 = const()[name = tensor("reshape_11_shape_0"), val = tensor([-1])]; + tensor reshape_11 = reshape(shape = reshape_11_shape_0, x = slice_by_index_2)[name = tensor("reshape_11")]; + tensor reshape_12_shape_0 = const()[name = tensor("reshape_12_shape_0"), val = tensor([-1])]; + tensor reshape_12_cast_fp16 = reshape(shape = reshape_12_shape_0, x = var_709_cast_fp16)[name = tensor("reshape_12_cast_fp16")]; + tensor reshape_13_shape_0 = const()[name = tensor("reshape_13_shape_0"), val = tensor([-1])]; + tensor reshape_13_cast_fp16 = reshape(shape = reshape_13_shape_0, x = cache1_to_fp16)[name = tensor("reshape_13_cast_fp16")]; + tensor scatter_2_mode_0 = const()[name = tensor("scatter_2_mode_0"), val = tensor("update")]; + tensor scatter_2_axis_0 = const()[name = tensor("scatter_2_axis_0"), val = tensor(0)]; + tensor scatter_2_validate_indices_0 = const()[name = tensor("scatter_2_validate_indices_0"), val = tensor(false)]; + tensor scatter_2_cast_fp16 = scatter(axis = scatter_2_axis_0, data = reshape_13_cast_fp16, indices = reshape_11, mode = scatter_2_mode_0, updates = reshape_12_cast_fp16, validate_indices = scatter_2_validate_indices_0)[name = tensor("scatter_2_cast_fp16")]; + tensor reshape_14_cast_fp16 = reshape(shape = shape_12, x = scatter_2_cast_fp16)[name = tensor("reshape_14_cast_fp16")]; + tensor var_717_begin_0 = const()[name = tensor("op_717_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor var_717_end_0 = const()[name = tensor("op_717_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor var_717_end_mask_0 = const()[name = tensor("op_717_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_717_squeeze_mask_0 = const()[name = tensor("op_717_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor var_717_cast_fp16 = slice_by_index(begin = var_717_begin_0, end = var_717_end_0, end_mask = var_717_end_mask_0, squeeze_mask = var_717_squeeze_mask_0, x = reshape_14_cast_fp16)[name = tensor("op_717_cast_fp16")]; + tensor var_719_axis_0 = const()[name = tensor("op_719_axis_0"), val = tensor(1)]; + tensor var_719_mode_0 = const()[name = tensor("op_719_mode_0"), val = tensor("update")]; + tensor var_719_validate_indices_0 = const()[name = tensor("op_719_validate_indices_0"), val = tensor(false)]; + tensor var_719_cast_fp16 = scatter_along_axis(axis = var_719_axis_0, data = var_717_cast_fp16, indices = write_indices_3, mode = var_719_mode_0, updates = v_3_cast_fp16, validate_indices = var_719_validate_indices_0)[name = tensor("op_719_cast_fp16")]; + tensor concat_12 = const()[name = tensor("concat_12"), val = tensor([1, 0, 0, 0, 0])]; + tensor concat_13 = const()[name = tensor("concat_13"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_3_internal_tensor_assign_2_stride_0 = const()[name = tensor("new_cache_3_internal_tensor_assign_2_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_3_internal_tensor_assign_2_begin_mask_0 = const()[name = tensor("new_cache_3_internal_tensor_assign_2_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_3_internal_tensor_assign_2_end_mask_0 = const()[name = tensor("new_cache_3_internal_tensor_assign_2_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_3_internal_tensor_assign_2_squeeze_mask_0 = const()[name = tensor("new_cache_3_internal_tensor_assign_2_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_13 = const()[name = tensor("shape_13"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_3 = const()[name = tensor("reduce_prod_3"), val = tensor(1048576)]; + tensor range_1d_3_start_0 = const()[name = tensor("range_1d_3_start_0"), val = tensor(0)]; + tensor range_1d_3_step_0 = const()[name = tensor("range_1d_3_step_0"), val = tensor(1)]; + tensor range_1d_3 = range_1d(end = reduce_prod_3, start = range_1d_3_start_0, step = range_1d_3_step_0)[name = tensor("range_1d_3")]; + tensor reshape_15 = reshape(shape = shape_13, x = range_1d_3)[name = tensor("reshape_15")]; + tensor slice_by_index_3 = slice_by_index(begin = concat_12, begin_mask = new_cache_3_internal_tensor_assign_2_begin_mask_0, end = concat_13, end_mask = new_cache_3_internal_tensor_assign_2_end_mask_0, squeeze_mask = new_cache_3_internal_tensor_assign_2_squeeze_mask_0, stride = new_cache_3_internal_tensor_assign_2_stride_0, x = reshape_15)[name = tensor("slice_by_index_3")]; + tensor reshape_16_shape_0 = const()[name = tensor("reshape_16_shape_0"), val = tensor([-1])]; + tensor reshape_16 = reshape(shape = reshape_16_shape_0, x = slice_by_index_3)[name = tensor("reshape_16")]; + tensor reshape_17_shape_0 = const()[name = tensor("reshape_17_shape_0"), val = tensor([-1])]; + tensor reshape_17_cast_fp16 = reshape(shape = reshape_17_shape_0, x = var_719_cast_fp16)[name = tensor("reshape_17_cast_fp16")]; + tensor reshape_18_shape_0 = const()[name = tensor("reshape_18_shape_0"), val = tensor([-1])]; + tensor reshape_18_cast_fp16 = reshape(shape = reshape_18_shape_0, x = reshape_14_cast_fp16)[name = tensor("reshape_18_cast_fp16")]; + tensor scatter_3_mode_0 = const()[name = tensor("scatter_3_mode_0"), val = tensor("update")]; + tensor scatter_3_axis_0 = const()[name = tensor("scatter_3_axis_0"), val = tensor(0)]; + tensor scatter_3_validate_indices_0 = const()[name = tensor("scatter_3_validate_indices_0"), val = tensor(false)]; + tensor scatter_3_cast_fp16 = scatter(axis = scatter_3_axis_0, data = reshape_18_cast_fp16, indices = reshape_16, mode = scatter_3_mode_0, updates = reshape_17_cast_fp16, validate_indices = scatter_3_validate_indices_0)[name = tensor("scatter_3_cast_fp16")]; + tensor reshape_19_cast_fp16 = reshape(shape = shape_13, x = scatter_3_cast_fp16)[name = tensor("reshape_19_cast_fp16")]; + tensor reshape_19_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("reshape_19_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor keys_7_begin_0 = const()[name = tensor("keys_7_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor keys_7_end_0 = const()[name = tensor("keys_7_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor keys_7_end_mask_0 = const()[name = tensor("keys_7_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor keys_7_squeeze_mask_0 = const()[name = tensor("keys_7_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor keys_7_cast_fp16 = slice_by_index(begin = keys_7_begin_0, end = keys_7_end_0, end_mask = keys_7_end_mask_0, squeeze_mask = keys_7_squeeze_mask_0, x = reshape_19_cast_fp16)[name = tensor("keys_7_cast_fp16")]; + tensor values_7_begin_0 = const()[name = tensor("values_7_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor values_7_end_0 = const()[name = tensor("values_7_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor values_7_end_mask_0 = const()[name = tensor("values_7_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor values_7_squeeze_mask_0 = const()[name = tensor("values_7_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor values_7_cast_fp16 = slice_by_index(begin = values_7_begin_0, end = values_7_end_0, end_mask = values_7_end_mask_0, squeeze_mask = values_7_squeeze_mask_0, x = reshape_19_cast_fp16)[name = tensor("values_7_cast_fp16")]; + tensor var_731_cast_fp16 = not_equal(x = keys_7_cast_fp16, y = keys_7_cast_fp16)[name = tensor("op_731_cast_fp16")]; + tensor keys_9_cast_fp16 = select(a = var_326_to_fp16, b = keys_7_cast_fp16, cond = var_731_cast_fp16)[name = tensor("keys_9_cast_fp16")]; + tensor var_739_cast_fp16 = not_equal(x = values_7_cast_fp16, y = values_7_cast_fp16)[name = tensor("op_739_cast_fp16")]; + tensor values_9_cast_fp16 = select(a = var_326_to_fp16, b = values_7_cast_fp16, cond = var_739_cast_fp16)[name = tensor("values_9_cast_fp16")]; + tensor var_763 = const()[name = tensor("op_763"), val = tensor([0, 2, 1, 3])]; + tensor var_779 = const()[name = tensor("op_779"), val = tensor([1, 1, 1])]; + tensor var_780_cast_fp16 = reshape(shape = var_779, x = position1_to_fp16)[name = tensor("op_780_cast_fp16")]; + tensor q_positions_3_cast_fp16 = add(x = var_780_cast_fp16, y = q_offsets_1_promoted_to_fp16)[name = tensor("q_positions_3_cast_fp16")]; + tensor valid_end_3_cast_fp16 = add(x = var_780_cast_fp16, y = var_392_cast_fp16)[name = tensor("valid_end_3_cast_fp16")]; + tensor valid_mask_3_cast_fp16 = less(x = k_positions_1_promoted_to_fp16, y = valid_end_3_cast_fp16)[name = tensor("valid_mask_3_cast_fp16")]; + tensor causal_mask_3_cast_fp16 = less_equal(x = k_positions_1_promoted_to_fp16, y = q_positions_3_cast_fp16)[name = tensor("causal_mask_3_cast_fp16")]; + tensor var_808 = logical_and(x = valid_mask_3_cast_fp16, y = causal_mask_3_cast_fp16)[name = tensor("op_808")]; + tensor attn_mask_3_axes_0 = const()[name = tensor("attn_mask_3_axes_0"), val = tensor([1])]; + tensor attn_mask_3 = expand_dims(axes = attn_mask_3_axes_0, x = var_808)[name = tensor("attn_mask_3")]; + tensor var_822_transpose_x_0 = const()[name = tensor("op_822_transpose_x_0"), val = tensor(false)]; + tensor var_822_transpose_y_0 = const()[name = tensor("op_822_transpose_y_0"), val = tensor(false)]; + tensor transpose_17_perm_0 = const()[name = tensor("transpose_17_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_18_perm_0 = const()[name = tensor("transpose_18_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_18 = transpose(perm = transpose_18_perm_0, x = keys_9_cast_fp16)[name = tensor("transpose_38")]; + tensor transpose_17 = transpose(perm = transpose_17_perm_0, x = q_9_cast_fp16)[name = tensor("transpose_39")]; + tensor var_822_cast_fp16 = matmul(transpose_x = var_822_transpose_x_0, transpose_y = var_822_transpose_y_0, x = transpose_17, y = transpose_18)[name = tensor("op_822_cast_fp16")]; + tensor var_816_to_fp16 = const()[name = tensor("op_816_to_fp16"), val = tensor([0x1p-3])]; + tensor attn_weights_7_cast_fp16 = mul(x = var_822_cast_fp16, y = var_816_to_fp16)[name = tensor("attn_weights_7_cast_fp16")]; + tensor var_824 = logical_not(x = attn_mask_3)[name = tensor("op_824")]; + tensor var_825_to_fp16 = const()[name = tensor("op_825_to_fp16"), val = tensor(-inf)]; + tensor attn_weights_9_cast_fp16 = select(a = var_825_to_fp16, b = attn_weights_7_cast_fp16, cond = var_824)[name = tensor("attn_weights_9_cast_fp16")]; + tensor var_827 = const()[name = tensor("op_827"), val = tensor(-1)]; + tensor attn_weights_11_cast_fp16 = softmax(axis = var_827, x = attn_weights_9_cast_fp16)[name = tensor("attn_weights_11_cast_fp16")]; + tensor attn_output_3_transpose_x_0 = const()[name = tensor("attn_output_3_transpose_x_0"), val = tensor(false)]; + tensor attn_output_3_transpose_y_0 = const()[name = tensor("attn_output_3_transpose_y_0"), val = tensor(false)]; + tensor values_11_cast_fp16 = transpose(perm = var_763, x = values_9_cast_fp16)[name = tensor("transpose_40")]; + tensor attn_output_3_cast_fp16 = matmul(transpose_x = attn_output_3_transpose_x_0, transpose_y = attn_output_3_transpose_y_0, x = attn_weights_11_cast_fp16, y = values_11_cast_fp16)[name = tensor("attn_output_3_cast_fp16")]; + tensor var_835 = const()[name = tensor("op_835"), val = tensor([0, 2, 1, 3])]; + tensor var_840 = const()[name = tensor("op_840"), val = tensor([1, 256, 1024])]; + tensor var_836_cast_fp16 = transpose(perm = var_835, x = attn_output_3_cast_fp16)[name = tensor("transpose_37")]; + tensor input_17_cast_fp16 = reshape(shape = var_840, x = var_836_cast_fp16)[name = tensor("input_17_cast_fp16")]; + tensor attn1_out_proj_weight_to_fp16 = const()[name = tensor("attn1_out_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(32538624)))]; + tensor linear_5_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = attn1_out_proj_weight_to_fp16, x = input_17_cast_fp16)[name = tensor("linear_5_cast_fp16")]; + tensor var_849_cast_fp16 = add(x = position1_to_fp16, y = valid_len_to_fp16)[name = tensor("op_849_cast_fp16")]; + tensor var_849_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_849_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor input_19_cast_fp16 = add(x = input_13_cast_fp16, y = linear_5_cast_fp16)[name = tensor("input_19_cast_fp16")]; + tensor input_21_axes_0 = const()[name = tensor("input_21_axes_0"), val = tensor([-1])]; + tensor norm1_2_weight_to_fp16 = const()[name = tensor("norm1_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(34635840)))]; + tensor norm1_2_bias_to_fp16 = const()[name = tensor("norm1_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(34637952)))]; + tensor var_853_to_fp16 = const()[name = tensor("op_853_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_21_cast_fp16 = layer_norm(axes = input_21_axes_0, beta = norm1_2_bias_to_fp16, epsilon = var_853_to_fp16, gamma = norm1_2_weight_to_fp16, x = input_19_cast_fp16)[name = tensor("input_21_cast_fp16")]; + tensor linear1_1_weight_to_fp16 = const()[name = tensor("linear1_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(34640064)))]; + tensor linear_6_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = linear1_1_weight_to_fp16, x = input_21_cast_fp16)[name = tensor("linear_6_cast_fp16")]; + tensor input_23_mode_0 = const()[name = tensor("input_23_mode_0"), val = tensor("EXACT")]; + tensor input_23_cast_fp16 = gelu(mode = input_23_mode_0, x = linear_6_cast_fp16)[name = tensor("input_23_cast_fp16")]; + tensor linear1_2_weight_to_fp16 = const()[name = tensor("linear1_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(43028736)))]; + tensor linear_7_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = linear1_2_weight_to_fp16, x = input_23_cast_fp16)[name = tensor("linear_7_cast_fp16")]; + tensor input_25_cast_fp16 = add(x = input_19_cast_fp16, y = linear_7_cast_fp16)[name = tensor("input_25_cast_fp16")]; + tensor input_27_axes_0 = const()[name = tensor("input_27_axes_0"), val = tensor([-1])]; + tensor norm2_1_weight_to_fp16 = const()[name = tensor("norm2_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(51417408)))]; + tensor norm2_1_bias_to_fp16 = const()[name = tensor("norm2_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(51419520)))]; + tensor var_870_to_fp16 = const()[name = tensor("op_870_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_27_cast_fp16 = layer_norm(axes = input_27_axes_0, beta = norm2_1_bias_to_fp16, epsilon = var_870_to_fp16, gamma = norm2_1_weight_to_fp16, x = input_25_cast_fp16)[name = tensor("input_27_cast_fp16")]; + tensor attn2_in_proj_weight_to_fp16 = const()[name = tensor("attn2_in_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(51421632)))]; + tensor linear_8_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = attn2_in_proj_weight_to_fp16, x = input_27_cast_fp16)[name = tensor("linear_8_cast_fp16")]; + tensor var_884 = const()[name = tensor("op_884"), val = tensor([1, 256, 3, 16, 64])]; + tensor qkv_5_cast_fp16 = reshape(shape = var_884, x = linear_8_cast_fp16)[name = tensor("qkv_5_cast_fp16")]; + tensor q_13_begin_0 = const()[name = tensor("q_13_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor q_13_end_0 = const()[name = tensor("q_13_end_0"), val = tensor([1, 256, 1, 16, 64])]; + tensor q_13_end_mask_0 = const()[name = tensor("q_13_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor q_13_squeeze_mask_0 = const()[name = tensor("q_13_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor q_13_cast_fp16 = slice_by_index(begin = q_13_begin_0, end = q_13_end_0, end_mask = q_13_end_mask_0, squeeze_mask = q_13_squeeze_mask_0, x = qkv_5_cast_fp16)[name = tensor("q_13_cast_fp16")]; + tensor k_9_begin_0 = const()[name = tensor("k_9_begin_0"), val = tensor([0, 0, 1, 0, 0])]; + tensor k_9_end_0 = const()[name = tensor("k_9_end_0"), val = tensor([1, 256, 2, 16, 64])]; + tensor k_9_end_mask_0 = const()[name = tensor("k_9_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor k_9_squeeze_mask_0 = const()[name = tensor("k_9_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor k_9_cast_fp16 = slice_by_index(begin = k_9_begin_0, end = k_9_end_0, end_mask = k_9_end_mask_0, squeeze_mask = k_9_squeeze_mask_0, x = qkv_5_cast_fp16)[name = tensor("k_9_cast_fp16")]; + tensor v_5_begin_0 = const()[name = tensor("v_5_begin_0"), val = tensor([0, 0, 2, 0, 0])]; + tensor v_5_end_0 = const()[name = tensor("v_5_end_0"), val = tensor([1, 256, 3, 16, 64])]; + tensor v_5_end_mask_0 = const()[name = tensor("v_5_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor v_5_squeeze_mask_0 = const()[name = tensor("v_5_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor v_5_cast_fp16 = slice_by_index(begin = v_5_begin_0, end = v_5_end_0, end_mask = v_5_end_mask_0, squeeze_mask = v_5_squeeze_mask_0, x = qkv_5_cast_fp16)[name = tensor("v_5_cast_fp16")]; + tensor var_942 = const()[name = tensor("op_942"), val = tensor([1, 1])]; + tensor position2_to_fp16_dtype_0 = const()[name = tensor("position2_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor position2_to_fp16 = cast(dtype = position2_to_fp16_dtype_0, x = position2)[name = tensor("cast_147")]; + tensor var_943_cast_fp16 = reshape(shape = var_942, x = position2_to_fp16)[name = tensor("op_943_cast_fp16")]; + tensor ts_15_cast_fp16 = add(x = ts_1_promoted_to_fp16, y = var_943_cast_fp16)[name = tensor("ts_15_cast_fp16")]; + tensor var_950 = const()[name = tensor("op_950"), val = tensor([1, 256, 1, 1])]; + tensor ts_17_cast_fp16 = reshape(shape = var_950, x = ts_15_cast_fp16)[name = tensor("ts_17_cast_fp16")]; + tensor var_957 = const()[name = tensor("op_957"), val = tensor([1, 256, 16, 32, 2])]; + tensor q_complex_5_cast_fp16 = reshape(shape = var_957, x = q_13_cast_fp16)[name = tensor("q_complex_5_cast_fp16")]; + tensor var_964 = const()[name = tensor("op_964"), val = tensor([1, 256, 16, 32, 2])]; + tensor k_complex_5_cast_fp16 = reshape(shape = var_964, x = k_9_cast_fp16)[name = tensor("k_complex_5_cast_fp16")]; + tensor var_968_begin_0 = const()[name = tensor("op_968_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_968_end_0 = const()[name = tensor("op_968_end_0"), val = tensor([1, 256, 16, 32, 1])]; + tensor var_968_end_mask_0 = const()[name = tensor("op_968_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_968_squeeze_mask_0 = const()[name = tensor("op_968_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_968_cast_fp16 = slice_by_index(begin = var_968_begin_0, end = var_968_end_0, end_mask = var_968_end_mask_0, squeeze_mask = var_968_squeeze_mask_0, x = q_complex_5_cast_fp16)[name = tensor("op_968_cast_fp16")]; + tensor var_976_begin_0 = const()[name = tensor("op_976_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_976_end_0 = const()[name = tensor("op_976_end_0"), val = tensor([1, 256, 16, 32, 2])]; + tensor var_976_end_mask_0 = const()[name = tensor("op_976_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_976_squeeze_mask_0 = const()[name = tensor("op_976_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_976_cast_fp16 = slice_by_index(begin = var_976_begin_0, end = var_976_end_0, end_mask = var_976_end_mask_0, squeeze_mask = var_976_squeeze_mask_0, x = q_complex_5_cast_fp16)[name = tensor("op_976_cast_fp16")]; + tensor var_984_begin_0 = const()[name = tensor("op_984_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_984_end_0 = const()[name = tensor("op_984_end_0"), val = tensor([1, 256, 16, 32, 1])]; + tensor var_984_end_mask_0 = const()[name = tensor("op_984_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_984_squeeze_mask_0 = const()[name = tensor("op_984_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_984_cast_fp16 = slice_by_index(begin = var_984_begin_0, end = var_984_end_0, end_mask = var_984_end_mask_0, squeeze_mask = var_984_squeeze_mask_0, x = k_complex_5_cast_fp16)[name = tensor("op_984_cast_fp16")]; + tensor var_992_begin_0 = const()[name = tensor("op_992_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_992_end_0 = const()[name = tensor("op_992_end_0"), val = tensor([1, 256, 16, 32, 2])]; + tensor var_992_end_mask_0 = const()[name = tensor("op_992_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_992_squeeze_mask_0 = const()[name = tensor("op_992_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_992_cast_fp16 = slice_by_index(begin = var_992_begin_0, end = var_992_end_0, end_mask = var_992_end_mask_0, squeeze_mask = var_992_squeeze_mask_0, x = k_complex_5_cast_fp16)[name = tensor("op_992_cast_fp16")]; + tensor freqs_5_to_fp16 = const()[name = tensor("freqs_5_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(57713152)))]; + tensor var_998_cast_fp16 = mul(x = freqs_5_to_fp16, y = ts_17_cast_fp16)[name = tensor("op_998_cast_fp16")]; + tensor rotr_5_cast_fp16 = cos(x = var_998_cast_fp16)[name = tensor("rotr_5_cast_fp16")]; + tensor roti_5_cast_fp16 = sin(x = var_998_cast_fp16)[name = tensor("roti_5_cast_fp16")]; + tensor var_1002_cast_fp16 = mul(x = var_968_cast_fp16, y = rotr_5_cast_fp16)[name = tensor("op_1002_cast_fp16")]; + tensor var_1003_cast_fp16 = mul(x = var_976_cast_fp16, y = roti_5_cast_fp16)[name = tensor("op_1003_cast_fp16")]; + tensor qor_9_cast_fp16 = sub(x = var_1002_cast_fp16, y = var_1003_cast_fp16)[name = tensor("qor_9_cast_fp16")]; + tensor var_1006_cast_fp16 = mul(x = var_968_cast_fp16, y = roti_5_cast_fp16)[name = tensor("op_1006_cast_fp16")]; + tensor var_1007_cast_fp16 = mul(x = var_976_cast_fp16, y = rotr_5_cast_fp16)[name = tensor("op_1007_cast_fp16")]; + tensor qoi_9_cast_fp16 = add(x = var_1006_cast_fp16, y = var_1007_cast_fp16)[name = tensor("qoi_9_cast_fp16")]; + tensor var_1010_cast_fp16 = mul(x = var_984_cast_fp16, y = rotr_5_cast_fp16)[name = tensor("op_1010_cast_fp16")]; + tensor var_1011_cast_fp16 = mul(x = var_992_cast_fp16, y = roti_5_cast_fp16)[name = tensor("op_1011_cast_fp16")]; + tensor kor_9_cast_fp16 = sub(x = var_1010_cast_fp16, y = var_1011_cast_fp16)[name = tensor("kor_9_cast_fp16")]; + tensor var_1014_cast_fp16 = mul(x = var_984_cast_fp16, y = roti_5_cast_fp16)[name = tensor("op_1014_cast_fp16")]; + tensor var_1015_cast_fp16 = mul(x = var_992_cast_fp16, y = rotr_5_cast_fp16)[name = tensor("op_1015_cast_fp16")]; + tensor koi_9_cast_fp16 = add(x = var_1014_cast_fp16, y = var_1015_cast_fp16)[name = tensor("koi_9_cast_fp16")]; + tensor qo_5_axis_0 = const()[name = tensor("qo_5_axis_0"), val = tensor(-1)]; + tensor qo_5_cast_fp16 = stack(axis = qo_5_axis_0, values = (qor_9_cast_fp16, qoi_9_cast_fp16))[name = tensor("qo_5_cast_fp16")]; + tensor ko_5_axis_0 = const()[name = tensor("ko_5_axis_0"), val = tensor(-1)]; + tensor ko_5_cast_fp16 = stack(axis = ko_5_axis_0, values = (kor_9_cast_fp16, koi_9_cast_fp16))[name = tensor("ko_5_cast_fp16")]; + tensor var_1048 = const()[name = tensor("op_1048"), val = tensor([1, 256, 16, 64])]; + tensor q_15_cast_fp16 = reshape(shape = var_1048, x = qo_5_cast_fp16)[name = tensor("q_15_cast_fp16")]; + tensor var_1054 = const()[name = tensor("op_1054"), val = tensor([1, 256, 16, 64])]; + tensor k_11_cast_fp16 = reshape(shape = var_1054, x = ko_5_cast_fp16)[name = tensor("k_11_cast_fp16")]; + tensor write_indices_float_13_cast_fp16 = add(x = var_943_cast_fp16, y = var_249_promoted_to_fp16)[name = tensor("write_indices_float_13_cast_fp16")]; + tensor _inversed_1075_y_0_to_fp16 = const()[name = tensor("_inversed_1075_y_0_to_fp16"), val = tensor(0x1p-9)]; + tensor _inversed_1075_cast_fp16 = mul(x = write_indices_float_13_cast_fp16, y = _inversed_1075_y_0_to_fp16)[name = tensor("_inversed_1075_cast_fp16")]; + tensor var_1076_cast_fp16 = floor(x = _inversed_1075_cast_fp16)[name = tensor("op_1076_cast_fp16")]; + tensor var_1077_to_fp16 = const()[name = tensor("op_1077_to_fp16"), val = tensor(0x1p+9)]; + tensor var_1078_cast_fp16 = mul(x = var_1076_cast_fp16, y = var_1077_to_fp16)[name = tensor("op_1078_cast_fp16")]; + tensor write_indices_float_15_cast_fp16 = sub(x = write_indices_float_13_cast_fp16, y = var_1078_cast_fp16)[name = tensor("write_indices_float_15_cast_fp16")]; + tensor var_1096_value_0_to_fp16 = const()[name = tensor("op_1096_value_0_to_fp16"), val = tensor(0x1.ffp+8)]; + tensor var_1096_cast_fp16 = fill_like(ref_tensor = write_indices_float_15_cast_fp16, value = var_1096_value_0_to_fp16)[name = tensor("op_1096_cast_fp16")]; + tensor write_indices_float_17_cast_fp16 = select(a = write_indices_float_15_cast_fp16, b = var_1096_cast_fp16, cond = token_valid_1_cast_fp16)[name = tensor("write_indices_float_17_cast_fp16")]; + tensor var_1102_dtype_0 = const()[name = tensor("op_1102_dtype_0"), val = tensor("int32")]; + tensor var_1107 = const()[name = tensor("op_1107"), val = tensor([1, 256, 1, 1])]; + tensor write_indices_float_17_cast_fp16_to_int32 = cast(dtype = var_1102_dtype_0, x = write_indices_float_17_cast_fp16)[name = tensor("cast_146")]; + tensor var_1108 = reshape(shape = var_1107, x = write_indices_float_17_cast_fp16_to_int32)[name = tensor("op_1108")]; + tensor write_indices_5_reps_0 = const()[name = tensor("write_indices_5_reps_0"), val = tensor([1, 1, 16, 64])]; + tensor write_indices_5 = tile(reps = write_indices_5_reps_0, x = var_1108)[name = tensor("write_indices_5")]; + tensor var_1118_begin_0 = const()[name = tensor("op_1118_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_1118_end_0 = const()[name = tensor("op_1118_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor var_1118_end_mask_0 = const()[name = tensor("op_1118_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_1118_squeeze_mask_0 = const()[name = tensor("op_1118_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor cache2_to_fp16_dtype_0 = const()[name = tensor("cache2_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor cache2_to_fp16 = cast(dtype = cache2_to_fp16_dtype_0, x = cache2)[name = tensor("cast_145")]; + tensor var_1118_cast_fp16 = slice_by_index(begin = var_1118_begin_0, end = var_1118_end_0, end_mask = var_1118_end_mask_0, squeeze_mask = var_1118_squeeze_mask_0, x = cache2_to_fp16)[name = tensor("op_1118_cast_fp16")]; + tensor var_1120_axis_0 = const()[name = tensor("op_1120_axis_0"), val = tensor(1)]; + tensor var_1120_mode_0 = const()[name = tensor("op_1120_mode_0"), val = tensor("update")]; + tensor var_1120_validate_indices_0 = const()[name = tensor("op_1120_validate_indices_0"), val = tensor(false)]; + tensor var_1120_cast_fp16 = scatter_along_axis(axis = var_1120_axis_0, data = var_1118_cast_fp16, indices = write_indices_5, mode = var_1120_mode_0, updates = k_11_cast_fp16, validate_indices = var_1120_validate_indices_0)[name = tensor("op_1120_cast_fp16")]; + tensor concat_18 = const()[name = tensor("concat_18"), val = tensor([0, 0, 0, 0, 0])]; + tensor concat_19 = const()[name = tensor("concat_19"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_5_internal_tensor_assign_1_stride_0 = const()[name = tensor("new_cache_5_internal_tensor_assign_1_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_5_internal_tensor_assign_1_begin_mask_0 = const()[name = tensor("new_cache_5_internal_tensor_assign_1_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_5_internal_tensor_assign_1_end_mask_0 = const()[name = tensor("new_cache_5_internal_tensor_assign_1_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_5_internal_tensor_assign_1_squeeze_mask_0 = const()[name = tensor("new_cache_5_internal_tensor_assign_1_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_14 = const()[name = tensor("shape_14"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_4 = const()[name = tensor("reduce_prod_4"), val = tensor(1048576)]; + tensor range_1d_4_start_0 = const()[name = tensor("range_1d_4_start_0"), val = tensor(0)]; + tensor range_1d_4_step_0 = const()[name = tensor("range_1d_4_step_0"), val = tensor(1)]; + tensor range_1d_4 = range_1d(end = reduce_prod_4, start = range_1d_4_start_0, step = range_1d_4_step_0)[name = tensor("range_1d_4")]; + tensor reshape_20 = reshape(shape = shape_14, x = range_1d_4)[name = tensor("reshape_20")]; + tensor slice_by_index_4 = slice_by_index(begin = concat_18, begin_mask = new_cache_5_internal_tensor_assign_1_begin_mask_0, end = concat_19, end_mask = new_cache_5_internal_tensor_assign_1_end_mask_0, squeeze_mask = new_cache_5_internal_tensor_assign_1_squeeze_mask_0, stride = new_cache_5_internal_tensor_assign_1_stride_0, x = reshape_20)[name = tensor("slice_by_index_4")]; + tensor reshape_21_shape_0 = const()[name = tensor("reshape_21_shape_0"), val = tensor([-1])]; + tensor reshape_21 = reshape(shape = reshape_21_shape_0, x = slice_by_index_4)[name = tensor("reshape_21")]; + tensor reshape_22_shape_0 = const()[name = tensor("reshape_22_shape_0"), val = tensor([-1])]; + tensor reshape_22_cast_fp16 = reshape(shape = reshape_22_shape_0, x = var_1120_cast_fp16)[name = tensor("reshape_22_cast_fp16")]; + tensor reshape_23_shape_0 = const()[name = tensor("reshape_23_shape_0"), val = tensor([-1])]; + tensor reshape_23_cast_fp16 = reshape(shape = reshape_23_shape_0, x = cache2_to_fp16)[name = tensor("reshape_23_cast_fp16")]; + tensor scatter_4_mode_0 = const()[name = tensor("scatter_4_mode_0"), val = tensor("update")]; + tensor scatter_4_axis_0 = const()[name = tensor("scatter_4_axis_0"), val = tensor(0)]; + tensor scatter_4_validate_indices_0 = const()[name = tensor("scatter_4_validate_indices_0"), val = tensor(false)]; + tensor scatter_4_cast_fp16 = scatter(axis = scatter_4_axis_0, data = reshape_23_cast_fp16, indices = reshape_21, mode = scatter_4_mode_0, updates = reshape_22_cast_fp16, validate_indices = scatter_4_validate_indices_0)[name = tensor("scatter_4_cast_fp16")]; + tensor reshape_24_cast_fp16 = reshape(shape = shape_14, x = scatter_4_cast_fp16)[name = tensor("reshape_24_cast_fp16")]; + tensor var_1128_begin_0 = const()[name = tensor("op_1128_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor var_1128_end_0 = const()[name = tensor("op_1128_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor var_1128_end_mask_0 = const()[name = tensor("op_1128_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_1128_squeeze_mask_0 = const()[name = tensor("op_1128_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor var_1128_cast_fp16 = slice_by_index(begin = var_1128_begin_0, end = var_1128_end_0, end_mask = var_1128_end_mask_0, squeeze_mask = var_1128_squeeze_mask_0, x = reshape_24_cast_fp16)[name = tensor("op_1128_cast_fp16")]; + tensor var_1130_axis_0 = const()[name = tensor("op_1130_axis_0"), val = tensor(1)]; + tensor var_1130_mode_0 = const()[name = tensor("op_1130_mode_0"), val = tensor("update")]; + tensor var_1130_validate_indices_0 = const()[name = tensor("op_1130_validate_indices_0"), val = tensor(false)]; + tensor var_1130_cast_fp16 = scatter_along_axis(axis = var_1130_axis_0, data = var_1128_cast_fp16, indices = write_indices_5, mode = var_1130_mode_0, updates = v_5_cast_fp16, validate_indices = var_1130_validate_indices_0)[name = tensor("op_1130_cast_fp16")]; + tensor concat_20 = const()[name = tensor("concat_20"), val = tensor([1, 0, 0, 0, 0])]; + tensor concat_21 = const()[name = tensor("concat_21"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_5_internal_tensor_assign_2_stride_0 = const()[name = tensor("new_cache_5_internal_tensor_assign_2_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_5_internal_tensor_assign_2_begin_mask_0 = const()[name = tensor("new_cache_5_internal_tensor_assign_2_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_5_internal_tensor_assign_2_end_mask_0 = const()[name = tensor("new_cache_5_internal_tensor_assign_2_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_5_internal_tensor_assign_2_squeeze_mask_0 = const()[name = tensor("new_cache_5_internal_tensor_assign_2_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_15 = const()[name = tensor("shape_15"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_5 = const()[name = tensor("reduce_prod_5"), val = tensor(1048576)]; + tensor range_1d_5_start_0 = const()[name = tensor("range_1d_5_start_0"), val = tensor(0)]; + tensor range_1d_5_step_0 = const()[name = tensor("range_1d_5_step_0"), val = tensor(1)]; + tensor range_1d_5 = range_1d(end = reduce_prod_5, start = range_1d_5_start_0, step = range_1d_5_step_0)[name = tensor("range_1d_5")]; + tensor reshape_25 = reshape(shape = shape_15, x = range_1d_5)[name = tensor("reshape_25")]; + tensor slice_by_index_5 = slice_by_index(begin = concat_20, begin_mask = new_cache_5_internal_tensor_assign_2_begin_mask_0, end = concat_21, end_mask = new_cache_5_internal_tensor_assign_2_end_mask_0, squeeze_mask = new_cache_5_internal_tensor_assign_2_squeeze_mask_0, stride = new_cache_5_internal_tensor_assign_2_stride_0, x = reshape_25)[name = tensor("slice_by_index_5")]; + tensor reshape_26_shape_0 = const()[name = tensor("reshape_26_shape_0"), val = tensor([-1])]; + tensor reshape_26 = reshape(shape = reshape_26_shape_0, x = slice_by_index_5)[name = tensor("reshape_26")]; + tensor reshape_27_shape_0 = const()[name = tensor("reshape_27_shape_0"), val = tensor([-1])]; + tensor reshape_27_cast_fp16 = reshape(shape = reshape_27_shape_0, x = var_1130_cast_fp16)[name = tensor("reshape_27_cast_fp16")]; + tensor reshape_28_shape_0 = const()[name = tensor("reshape_28_shape_0"), val = tensor([-1])]; + tensor reshape_28_cast_fp16 = reshape(shape = reshape_28_shape_0, x = reshape_24_cast_fp16)[name = tensor("reshape_28_cast_fp16")]; + tensor scatter_5_mode_0 = const()[name = tensor("scatter_5_mode_0"), val = tensor("update")]; + tensor scatter_5_axis_0 = const()[name = tensor("scatter_5_axis_0"), val = tensor(0)]; + tensor scatter_5_validate_indices_0 = const()[name = tensor("scatter_5_validate_indices_0"), val = tensor(false)]; + tensor scatter_5_cast_fp16 = scatter(axis = scatter_5_axis_0, data = reshape_28_cast_fp16, indices = reshape_26, mode = scatter_5_mode_0, updates = reshape_27_cast_fp16, validate_indices = scatter_5_validate_indices_0)[name = tensor("scatter_5_cast_fp16")]; + tensor reshape_29_cast_fp16 = reshape(shape = shape_15, x = scatter_5_cast_fp16)[name = tensor("reshape_29_cast_fp16")]; + tensor reshape_29_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("reshape_29_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor keys_13_begin_0 = const()[name = tensor("keys_13_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor keys_13_end_0 = const()[name = tensor("keys_13_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor keys_13_end_mask_0 = const()[name = tensor("keys_13_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor keys_13_squeeze_mask_0 = const()[name = tensor("keys_13_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor keys_13_cast_fp16 = slice_by_index(begin = keys_13_begin_0, end = keys_13_end_0, end_mask = keys_13_end_mask_0, squeeze_mask = keys_13_squeeze_mask_0, x = reshape_29_cast_fp16)[name = tensor("keys_13_cast_fp16")]; + tensor values_13_begin_0 = const()[name = tensor("values_13_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor values_13_end_0 = const()[name = tensor("values_13_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor values_13_end_mask_0 = const()[name = tensor("values_13_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor values_13_squeeze_mask_0 = const()[name = tensor("values_13_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor values_13_cast_fp16 = slice_by_index(begin = values_13_begin_0, end = values_13_end_0, end_mask = values_13_end_mask_0, squeeze_mask = values_13_squeeze_mask_0, x = reshape_29_cast_fp16)[name = tensor("values_13_cast_fp16")]; + tensor var_1142_cast_fp16 = not_equal(x = keys_13_cast_fp16, y = keys_13_cast_fp16)[name = tensor("op_1142_cast_fp16")]; + tensor keys_15_cast_fp16 = select(a = var_326_to_fp16, b = keys_13_cast_fp16, cond = var_1142_cast_fp16)[name = tensor("keys_15_cast_fp16")]; + tensor var_1150_cast_fp16 = not_equal(x = values_13_cast_fp16, y = values_13_cast_fp16)[name = tensor("op_1150_cast_fp16")]; + tensor values_15_cast_fp16 = select(a = var_326_to_fp16, b = values_13_cast_fp16, cond = var_1150_cast_fp16)[name = tensor("values_15_cast_fp16")]; + tensor var_1174 = const()[name = tensor("op_1174"), val = tensor([0, 2, 1, 3])]; + tensor var_1190 = const()[name = tensor("op_1190"), val = tensor([1, 1, 1])]; + tensor var_1191_cast_fp16 = reshape(shape = var_1190, x = position2_to_fp16)[name = tensor("op_1191_cast_fp16")]; + tensor q_positions_5_cast_fp16 = add(x = var_1191_cast_fp16, y = q_offsets_1_promoted_to_fp16)[name = tensor("q_positions_5_cast_fp16")]; + tensor valid_end_5_cast_fp16 = add(x = var_1191_cast_fp16, y = var_392_cast_fp16)[name = tensor("valid_end_5_cast_fp16")]; + tensor valid_mask_5_cast_fp16 = less(x = k_positions_1_promoted_to_fp16, y = valid_end_5_cast_fp16)[name = tensor("valid_mask_5_cast_fp16")]; + tensor causal_mask_5_cast_fp16 = less_equal(x = k_positions_1_promoted_to_fp16, y = q_positions_5_cast_fp16)[name = tensor("causal_mask_5_cast_fp16")]; + tensor var_1219 = logical_and(x = valid_mask_5_cast_fp16, y = causal_mask_5_cast_fp16)[name = tensor("op_1219")]; + tensor attn_mask_5_axes_0 = const()[name = tensor("attn_mask_5_axes_0"), val = tensor([1])]; + tensor attn_mask_5 = expand_dims(axes = attn_mask_5_axes_0, x = var_1219)[name = tensor("attn_mask_5")]; + tensor var_1233_transpose_x_0 = const()[name = tensor("op_1233_transpose_x_0"), val = tensor(false)]; + tensor var_1233_transpose_y_0 = const()[name = tensor("op_1233_transpose_y_0"), val = tensor(false)]; + tensor transpose_19_perm_0 = const()[name = tensor("transpose_19_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_20_perm_0 = const()[name = tensor("transpose_20_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_20 = transpose(perm = transpose_20_perm_0, x = keys_15_cast_fp16)[name = tensor("transpose_34")]; + tensor transpose_19 = transpose(perm = transpose_19_perm_0, x = q_15_cast_fp16)[name = tensor("transpose_35")]; + tensor var_1233_cast_fp16 = matmul(transpose_x = var_1233_transpose_x_0, transpose_y = var_1233_transpose_y_0, x = transpose_19, y = transpose_20)[name = tensor("op_1233_cast_fp16")]; + tensor var_1227_to_fp16 = const()[name = tensor("op_1227_to_fp16"), val = tensor([0x1p-3])]; + tensor attn_weights_13_cast_fp16 = mul(x = var_1233_cast_fp16, y = var_1227_to_fp16)[name = tensor("attn_weights_13_cast_fp16")]; + tensor var_1235 = logical_not(x = attn_mask_5)[name = tensor("op_1235")]; + tensor var_1236_to_fp16 = const()[name = tensor("op_1236_to_fp16"), val = tensor(-inf)]; + tensor attn_weights_15_cast_fp16 = select(a = var_1236_to_fp16, b = attn_weights_13_cast_fp16, cond = var_1235)[name = tensor("attn_weights_15_cast_fp16")]; + tensor var_1238 = const()[name = tensor("op_1238"), val = tensor(-1)]; + tensor attn_weights_17_cast_fp16 = softmax(axis = var_1238, x = attn_weights_15_cast_fp16)[name = tensor("attn_weights_17_cast_fp16")]; + tensor attn_output_5_transpose_x_0 = const()[name = tensor("attn_output_5_transpose_x_0"), val = tensor(false)]; + tensor attn_output_5_transpose_y_0 = const()[name = tensor("attn_output_5_transpose_y_0"), val = tensor(false)]; + tensor values_17_cast_fp16 = transpose(perm = var_1174, x = values_15_cast_fp16)[name = tensor("transpose_36")]; + tensor attn_output_5_cast_fp16 = matmul(transpose_x = attn_output_5_transpose_x_0, transpose_y = attn_output_5_transpose_y_0, x = attn_weights_17_cast_fp16, y = values_17_cast_fp16)[name = tensor("attn_output_5_cast_fp16")]; + tensor var_1246 = const()[name = tensor("op_1246"), val = tensor([0, 2, 1, 3])]; + tensor var_1251 = const()[name = tensor("op_1251"), val = tensor([1, 256, 1024])]; + tensor var_1247_cast_fp16 = transpose(perm = var_1246, x = attn_output_5_cast_fp16)[name = tensor("transpose_33")]; + tensor input_29_cast_fp16 = reshape(shape = var_1251, x = var_1247_cast_fp16)[name = tensor("input_29_cast_fp16")]; + tensor attn2_out_proj_weight_to_fp16 = const()[name = tensor("attn2_out_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(57713280)))]; + tensor linear_9_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = attn2_out_proj_weight_to_fp16, x = input_29_cast_fp16)[name = tensor("linear_9_cast_fp16")]; + tensor var_1260_cast_fp16 = add(x = position2_to_fp16, y = valid_len_to_fp16)[name = tensor("op_1260_cast_fp16")]; + tensor var_1260_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_1260_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor input_31_cast_fp16 = add(x = input_25_cast_fp16, y = linear_9_cast_fp16)[name = tensor("input_31_cast_fp16")]; + tensor input_33_axes_0 = const()[name = tensor("input_33_axes_0"), val = tensor([-1])]; + tensor norm2_2_weight_to_fp16 = const()[name = tensor("norm2_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(59810496)))]; + tensor norm2_2_bias_to_fp16 = const()[name = tensor("norm2_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(59812608)))]; + tensor var_1264_to_fp16 = const()[name = tensor("op_1264_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_33_cast_fp16 = layer_norm(axes = input_33_axes_0, beta = norm2_2_bias_to_fp16, epsilon = var_1264_to_fp16, gamma = norm2_2_weight_to_fp16, x = input_31_cast_fp16)[name = tensor("input_33_cast_fp16")]; + tensor linear2_1_weight_to_fp16 = const()[name = tensor("linear2_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(59814720)))]; + tensor linear_10_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = linear2_1_weight_to_fp16, x = input_33_cast_fp16)[name = tensor("linear_10_cast_fp16")]; + tensor input_35_mode_0 = const()[name = tensor("input_35_mode_0"), val = tensor("EXACT")]; + tensor input_35_cast_fp16 = gelu(mode = input_35_mode_0, x = linear_10_cast_fp16)[name = tensor("input_35_cast_fp16")]; + tensor linear2_2_weight_to_fp16 = const()[name = tensor("linear2_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(68203392)))]; + tensor linear_11_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = linear2_2_weight_to_fp16, x = input_35_cast_fp16)[name = tensor("linear_11_cast_fp16")]; + tensor input_37_cast_fp16 = add(x = input_31_cast_fp16, y = linear_11_cast_fp16)[name = tensor("input_37_cast_fp16")]; + tensor input_39_axes_0 = const()[name = tensor("input_39_axes_0"), val = tensor([-1])]; + tensor norm3_1_weight_to_fp16 = const()[name = tensor("norm3_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(76592064)))]; + tensor norm3_1_bias_to_fp16 = const()[name = tensor("norm3_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(76594176)))]; + tensor var_1281_to_fp16 = const()[name = tensor("op_1281_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_39_cast_fp16 = layer_norm(axes = input_39_axes_0, beta = norm3_1_bias_to_fp16, epsilon = var_1281_to_fp16, gamma = norm3_1_weight_to_fp16, x = input_37_cast_fp16)[name = tensor("input_39_cast_fp16")]; + tensor attn3_in_proj_weight_to_fp16 = const()[name = tensor("attn3_in_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(76596288)))]; + tensor linear_12_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = attn3_in_proj_weight_to_fp16, x = input_39_cast_fp16)[name = tensor("linear_12_cast_fp16")]; + tensor var_1295 = const()[name = tensor("op_1295"), val = tensor([1, 256, 3, 16, 64])]; + tensor qkv_7_cast_fp16 = reshape(shape = var_1295, x = linear_12_cast_fp16)[name = tensor("qkv_7_cast_fp16")]; + tensor q_19_begin_0 = const()[name = tensor("q_19_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor q_19_end_0 = const()[name = tensor("q_19_end_0"), val = tensor([1, 256, 1, 16, 64])]; + tensor q_19_end_mask_0 = const()[name = tensor("q_19_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor q_19_squeeze_mask_0 = const()[name = tensor("q_19_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor q_19_cast_fp16 = slice_by_index(begin = q_19_begin_0, end = q_19_end_0, end_mask = q_19_end_mask_0, squeeze_mask = q_19_squeeze_mask_0, x = qkv_7_cast_fp16)[name = tensor("q_19_cast_fp16")]; + tensor k_13_begin_0 = const()[name = tensor("k_13_begin_0"), val = tensor([0, 0, 1, 0, 0])]; + tensor k_13_end_0 = const()[name = tensor("k_13_end_0"), val = tensor([1, 256, 2, 16, 64])]; + tensor k_13_end_mask_0 = const()[name = tensor("k_13_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor k_13_squeeze_mask_0 = const()[name = tensor("k_13_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor k_13_cast_fp16 = slice_by_index(begin = k_13_begin_0, end = k_13_end_0, end_mask = k_13_end_mask_0, squeeze_mask = k_13_squeeze_mask_0, x = qkv_7_cast_fp16)[name = tensor("k_13_cast_fp16")]; + tensor v_7_begin_0 = const()[name = tensor("v_7_begin_0"), val = tensor([0, 0, 2, 0, 0])]; + tensor v_7_end_0 = const()[name = tensor("v_7_end_0"), val = tensor([1, 256, 3, 16, 64])]; + tensor v_7_end_mask_0 = const()[name = tensor("v_7_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor v_7_squeeze_mask_0 = const()[name = tensor("v_7_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor v_7_cast_fp16 = slice_by_index(begin = v_7_begin_0, end = v_7_end_0, end_mask = v_7_end_mask_0, squeeze_mask = v_7_squeeze_mask_0, x = qkv_7_cast_fp16)[name = tensor("v_7_cast_fp16")]; + tensor var_1353 = const()[name = tensor("op_1353"), val = tensor([1, 1])]; + tensor position3_to_fp16_dtype_0 = const()[name = tensor("position3_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor position3_to_fp16 = cast(dtype = position3_to_fp16_dtype_0, x = position3)[name = tensor("cast_142")]; + tensor var_1354_cast_fp16 = reshape(shape = var_1353, x = position3_to_fp16)[name = tensor("op_1354_cast_fp16")]; + tensor ts_21_cast_fp16 = add(x = ts_1_promoted_to_fp16, y = var_1354_cast_fp16)[name = tensor("ts_21_cast_fp16")]; + tensor var_1361 = const()[name = tensor("op_1361"), val = tensor([1, 256, 1, 1])]; + tensor ts_23_cast_fp16 = reshape(shape = var_1361, x = ts_21_cast_fp16)[name = tensor("ts_23_cast_fp16")]; + tensor var_1368 = const()[name = tensor("op_1368"), val = tensor([1, 256, 16, 32, 2])]; + tensor q_complex_7_cast_fp16 = reshape(shape = var_1368, x = q_19_cast_fp16)[name = tensor("q_complex_7_cast_fp16")]; + tensor var_1375 = const()[name = tensor("op_1375"), val = tensor([1, 256, 16, 32, 2])]; + tensor k_complex_7_cast_fp16 = reshape(shape = var_1375, x = k_13_cast_fp16)[name = tensor("k_complex_7_cast_fp16")]; + tensor var_1379_begin_0 = const()[name = tensor("op_1379_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_1379_end_0 = const()[name = tensor("op_1379_end_0"), val = tensor([1, 256, 16, 32, 1])]; + tensor var_1379_end_mask_0 = const()[name = tensor("op_1379_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_1379_squeeze_mask_0 = const()[name = tensor("op_1379_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_1379_cast_fp16 = slice_by_index(begin = var_1379_begin_0, end = var_1379_end_0, end_mask = var_1379_end_mask_0, squeeze_mask = var_1379_squeeze_mask_0, x = q_complex_7_cast_fp16)[name = tensor("op_1379_cast_fp16")]; + tensor var_1387_begin_0 = const()[name = tensor("op_1387_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_1387_end_0 = const()[name = tensor("op_1387_end_0"), val = tensor([1, 256, 16, 32, 2])]; + tensor var_1387_end_mask_0 = const()[name = tensor("op_1387_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_1387_squeeze_mask_0 = const()[name = tensor("op_1387_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_1387_cast_fp16 = slice_by_index(begin = var_1387_begin_0, end = var_1387_end_0, end_mask = var_1387_end_mask_0, squeeze_mask = var_1387_squeeze_mask_0, x = q_complex_7_cast_fp16)[name = tensor("op_1387_cast_fp16")]; + tensor var_1395_begin_0 = const()[name = tensor("op_1395_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_1395_end_0 = const()[name = tensor("op_1395_end_0"), val = tensor([1, 256, 16, 32, 1])]; + tensor var_1395_end_mask_0 = const()[name = tensor("op_1395_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_1395_squeeze_mask_0 = const()[name = tensor("op_1395_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_1395_cast_fp16 = slice_by_index(begin = var_1395_begin_0, end = var_1395_end_0, end_mask = var_1395_end_mask_0, squeeze_mask = var_1395_squeeze_mask_0, x = k_complex_7_cast_fp16)[name = tensor("op_1395_cast_fp16")]; + tensor var_1403_begin_0 = const()[name = tensor("op_1403_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_1403_end_0 = const()[name = tensor("op_1403_end_0"), val = tensor([1, 256, 16, 32, 2])]; + tensor var_1403_end_mask_0 = const()[name = tensor("op_1403_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_1403_squeeze_mask_0 = const()[name = tensor("op_1403_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_1403_cast_fp16 = slice_by_index(begin = var_1403_begin_0, end = var_1403_end_0, end_mask = var_1403_end_mask_0, squeeze_mask = var_1403_squeeze_mask_0, x = k_complex_7_cast_fp16)[name = tensor("op_1403_cast_fp16")]; + tensor freqs_7_to_fp16 = const()[name = tensor("freqs_7_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(82887808)))]; + tensor var_1409_cast_fp16 = mul(x = freqs_7_to_fp16, y = ts_23_cast_fp16)[name = tensor("op_1409_cast_fp16")]; + tensor rotr_7_cast_fp16 = cos(x = var_1409_cast_fp16)[name = tensor("rotr_7_cast_fp16")]; + tensor roti_7_cast_fp16 = sin(x = var_1409_cast_fp16)[name = tensor("roti_7_cast_fp16")]; + tensor var_1413_cast_fp16 = mul(x = var_1379_cast_fp16, y = rotr_7_cast_fp16)[name = tensor("op_1413_cast_fp16")]; + tensor var_1414_cast_fp16 = mul(x = var_1387_cast_fp16, y = roti_7_cast_fp16)[name = tensor("op_1414_cast_fp16")]; + tensor qor_13_cast_fp16 = sub(x = var_1413_cast_fp16, y = var_1414_cast_fp16)[name = tensor("qor_13_cast_fp16")]; + tensor var_1417_cast_fp16 = mul(x = var_1379_cast_fp16, y = roti_7_cast_fp16)[name = tensor("op_1417_cast_fp16")]; + tensor var_1418_cast_fp16 = mul(x = var_1387_cast_fp16, y = rotr_7_cast_fp16)[name = tensor("op_1418_cast_fp16")]; + tensor qoi_13_cast_fp16 = add(x = var_1417_cast_fp16, y = var_1418_cast_fp16)[name = tensor("qoi_13_cast_fp16")]; + tensor var_1421_cast_fp16 = mul(x = var_1395_cast_fp16, y = rotr_7_cast_fp16)[name = tensor("op_1421_cast_fp16")]; + tensor var_1422_cast_fp16 = mul(x = var_1403_cast_fp16, y = roti_7_cast_fp16)[name = tensor("op_1422_cast_fp16")]; + tensor kor_13_cast_fp16 = sub(x = var_1421_cast_fp16, y = var_1422_cast_fp16)[name = tensor("kor_13_cast_fp16")]; + tensor var_1425_cast_fp16 = mul(x = var_1395_cast_fp16, y = roti_7_cast_fp16)[name = tensor("op_1425_cast_fp16")]; + tensor var_1426_cast_fp16 = mul(x = var_1403_cast_fp16, y = rotr_7_cast_fp16)[name = tensor("op_1426_cast_fp16")]; + tensor koi_13_cast_fp16 = add(x = var_1425_cast_fp16, y = var_1426_cast_fp16)[name = tensor("koi_13_cast_fp16")]; + tensor qo_7_axis_0 = const()[name = tensor("qo_7_axis_0"), val = tensor(-1)]; + tensor qo_7_cast_fp16 = stack(axis = qo_7_axis_0, values = (qor_13_cast_fp16, qoi_13_cast_fp16))[name = tensor("qo_7_cast_fp16")]; + tensor ko_7_axis_0 = const()[name = tensor("ko_7_axis_0"), val = tensor(-1)]; + tensor ko_7_cast_fp16 = stack(axis = ko_7_axis_0, values = (kor_13_cast_fp16, koi_13_cast_fp16))[name = tensor("ko_7_cast_fp16")]; + tensor var_1459 = const()[name = tensor("op_1459"), val = tensor([1, 256, 16, 64])]; + tensor q_21_cast_fp16 = reshape(shape = var_1459, x = qo_7_cast_fp16)[name = tensor("q_21_cast_fp16")]; + tensor var_1465 = const()[name = tensor("op_1465"), val = tensor([1, 256, 16, 64])]; + tensor k_15_cast_fp16 = reshape(shape = var_1465, x = ko_7_cast_fp16)[name = tensor("k_15_cast_fp16")]; + tensor write_indices_float_19_cast_fp16 = add(x = var_1354_cast_fp16, y = var_249_promoted_to_fp16)[name = tensor("write_indices_float_19_cast_fp16")]; + tensor _inversed_1486_y_0_to_fp16 = const()[name = tensor("_inversed_1486_y_0_to_fp16"), val = tensor(0x1p-9)]; + tensor _inversed_1486_cast_fp16 = mul(x = write_indices_float_19_cast_fp16, y = _inversed_1486_y_0_to_fp16)[name = tensor("_inversed_1486_cast_fp16")]; + tensor var_1487_cast_fp16 = floor(x = _inversed_1486_cast_fp16)[name = tensor("op_1487_cast_fp16")]; + tensor var_1488_to_fp16 = const()[name = tensor("op_1488_to_fp16"), val = tensor(0x1p+9)]; + tensor var_1489_cast_fp16 = mul(x = var_1487_cast_fp16, y = var_1488_to_fp16)[name = tensor("op_1489_cast_fp16")]; + tensor write_indices_float_21_cast_fp16 = sub(x = write_indices_float_19_cast_fp16, y = var_1489_cast_fp16)[name = tensor("write_indices_float_21_cast_fp16")]; + tensor var_1507_value_0_to_fp16 = const()[name = tensor("op_1507_value_0_to_fp16"), val = tensor(0x1.ffp+8)]; + tensor var_1507_cast_fp16 = fill_like(ref_tensor = write_indices_float_21_cast_fp16, value = var_1507_value_0_to_fp16)[name = tensor("op_1507_cast_fp16")]; + tensor write_indices_float_23_cast_fp16 = select(a = write_indices_float_21_cast_fp16, b = var_1507_cast_fp16, cond = token_valid_1_cast_fp16)[name = tensor("write_indices_float_23_cast_fp16")]; + tensor var_1513_dtype_0 = const()[name = tensor("op_1513_dtype_0"), val = tensor("int32")]; + tensor var_1518 = const()[name = tensor("op_1518"), val = tensor([1, 256, 1, 1])]; + tensor write_indices_float_23_cast_fp16_to_int32 = cast(dtype = var_1513_dtype_0, x = write_indices_float_23_cast_fp16)[name = tensor("cast_141")]; + tensor var_1519 = reshape(shape = var_1518, x = write_indices_float_23_cast_fp16_to_int32)[name = tensor("op_1519")]; + tensor write_indices_7_reps_0 = const()[name = tensor("write_indices_7_reps_0"), val = tensor([1, 1, 16, 64])]; + tensor write_indices_7 = tile(reps = write_indices_7_reps_0, x = var_1519)[name = tensor("write_indices_7")]; + tensor var_1529_begin_0 = const()[name = tensor("op_1529_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_1529_end_0 = const()[name = tensor("op_1529_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor var_1529_end_mask_0 = const()[name = tensor("op_1529_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_1529_squeeze_mask_0 = const()[name = tensor("op_1529_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor cache3_to_fp16_dtype_0 = const()[name = tensor("cache3_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor cache3_to_fp16 = cast(dtype = cache3_to_fp16_dtype_0, x = cache3)[name = tensor("cast_140")]; + tensor var_1529_cast_fp16 = slice_by_index(begin = var_1529_begin_0, end = var_1529_end_0, end_mask = var_1529_end_mask_0, squeeze_mask = var_1529_squeeze_mask_0, x = cache3_to_fp16)[name = tensor("op_1529_cast_fp16")]; + tensor var_1531_axis_0 = const()[name = tensor("op_1531_axis_0"), val = tensor(1)]; + tensor var_1531_mode_0 = const()[name = tensor("op_1531_mode_0"), val = tensor("update")]; + tensor var_1531_validate_indices_0 = const()[name = tensor("op_1531_validate_indices_0"), val = tensor(false)]; + tensor var_1531_cast_fp16 = scatter_along_axis(axis = var_1531_axis_0, data = var_1529_cast_fp16, indices = write_indices_7, mode = var_1531_mode_0, updates = k_15_cast_fp16, validate_indices = var_1531_validate_indices_0)[name = tensor("op_1531_cast_fp16")]; + tensor concat_26 = const()[name = tensor("concat_26"), val = tensor([0, 0, 0, 0, 0])]; + tensor concat_27 = const()[name = tensor("concat_27"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_7_internal_tensor_assign_1_stride_0 = const()[name = tensor("new_cache_7_internal_tensor_assign_1_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_7_internal_tensor_assign_1_begin_mask_0 = const()[name = tensor("new_cache_7_internal_tensor_assign_1_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_7_internal_tensor_assign_1_end_mask_0 = const()[name = tensor("new_cache_7_internal_tensor_assign_1_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_7_internal_tensor_assign_1_squeeze_mask_0 = const()[name = tensor("new_cache_7_internal_tensor_assign_1_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_16 = const()[name = tensor("shape_16"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_6 = const()[name = tensor("reduce_prod_6"), val = tensor(1048576)]; + tensor range_1d_6_start_0 = const()[name = tensor("range_1d_6_start_0"), val = tensor(0)]; + tensor range_1d_6_step_0 = const()[name = tensor("range_1d_6_step_0"), val = tensor(1)]; + tensor range_1d_6 = range_1d(end = reduce_prod_6, start = range_1d_6_start_0, step = range_1d_6_step_0)[name = tensor("range_1d_6")]; + tensor reshape_30 = reshape(shape = shape_16, x = range_1d_6)[name = tensor("reshape_30")]; + tensor slice_by_index_6 = slice_by_index(begin = concat_26, begin_mask = new_cache_7_internal_tensor_assign_1_begin_mask_0, end = concat_27, end_mask = new_cache_7_internal_tensor_assign_1_end_mask_0, squeeze_mask = new_cache_7_internal_tensor_assign_1_squeeze_mask_0, stride = new_cache_7_internal_tensor_assign_1_stride_0, x = reshape_30)[name = tensor("slice_by_index_6")]; + tensor reshape_31_shape_0 = const()[name = tensor("reshape_31_shape_0"), val = tensor([-1])]; + tensor reshape_31 = reshape(shape = reshape_31_shape_0, x = slice_by_index_6)[name = tensor("reshape_31")]; + tensor reshape_32_shape_0 = const()[name = tensor("reshape_32_shape_0"), val = tensor([-1])]; + tensor reshape_32_cast_fp16 = reshape(shape = reshape_32_shape_0, x = var_1531_cast_fp16)[name = tensor("reshape_32_cast_fp16")]; + tensor reshape_33_shape_0 = const()[name = tensor("reshape_33_shape_0"), val = tensor([-1])]; + tensor reshape_33_cast_fp16 = reshape(shape = reshape_33_shape_0, x = cache3_to_fp16)[name = tensor("reshape_33_cast_fp16")]; + tensor scatter_6_mode_0 = const()[name = tensor("scatter_6_mode_0"), val = tensor("update")]; + tensor scatter_6_axis_0 = const()[name = tensor("scatter_6_axis_0"), val = tensor(0)]; + tensor scatter_6_validate_indices_0 = const()[name = tensor("scatter_6_validate_indices_0"), val = tensor(false)]; + tensor scatter_6_cast_fp16 = scatter(axis = scatter_6_axis_0, data = reshape_33_cast_fp16, indices = reshape_31, mode = scatter_6_mode_0, updates = reshape_32_cast_fp16, validate_indices = scatter_6_validate_indices_0)[name = tensor("scatter_6_cast_fp16")]; + tensor reshape_34_cast_fp16 = reshape(shape = shape_16, x = scatter_6_cast_fp16)[name = tensor("reshape_34_cast_fp16")]; + tensor var_1539_begin_0 = const()[name = tensor("op_1539_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor var_1539_end_0 = const()[name = tensor("op_1539_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor var_1539_end_mask_0 = const()[name = tensor("op_1539_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_1539_squeeze_mask_0 = const()[name = tensor("op_1539_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor var_1539_cast_fp16 = slice_by_index(begin = var_1539_begin_0, end = var_1539_end_0, end_mask = var_1539_end_mask_0, squeeze_mask = var_1539_squeeze_mask_0, x = reshape_34_cast_fp16)[name = tensor("op_1539_cast_fp16")]; + tensor var_1541_axis_0 = const()[name = tensor("op_1541_axis_0"), val = tensor(1)]; + tensor var_1541_mode_0 = const()[name = tensor("op_1541_mode_0"), val = tensor("update")]; + tensor var_1541_validate_indices_0 = const()[name = tensor("op_1541_validate_indices_0"), val = tensor(false)]; + tensor var_1541_cast_fp16 = scatter_along_axis(axis = var_1541_axis_0, data = var_1539_cast_fp16, indices = write_indices_7, mode = var_1541_mode_0, updates = v_7_cast_fp16, validate_indices = var_1541_validate_indices_0)[name = tensor("op_1541_cast_fp16")]; + tensor concat_28 = const()[name = tensor("concat_28"), val = tensor([1, 0, 0, 0, 0])]; + tensor concat_29 = const()[name = tensor("concat_29"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_7_internal_tensor_assign_2_stride_0 = const()[name = tensor("new_cache_7_internal_tensor_assign_2_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_7_internal_tensor_assign_2_begin_mask_0 = const()[name = tensor("new_cache_7_internal_tensor_assign_2_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_7_internal_tensor_assign_2_end_mask_0 = const()[name = tensor("new_cache_7_internal_tensor_assign_2_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_7_internal_tensor_assign_2_squeeze_mask_0 = const()[name = tensor("new_cache_7_internal_tensor_assign_2_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_17 = const()[name = tensor("shape_17"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_7 = const()[name = tensor("reduce_prod_7"), val = tensor(1048576)]; + tensor range_1d_7_start_0 = const()[name = tensor("range_1d_7_start_0"), val = tensor(0)]; + tensor range_1d_7_step_0 = const()[name = tensor("range_1d_7_step_0"), val = tensor(1)]; + tensor range_1d_7 = range_1d(end = reduce_prod_7, start = range_1d_7_start_0, step = range_1d_7_step_0)[name = tensor("range_1d_7")]; + tensor reshape_35 = reshape(shape = shape_17, x = range_1d_7)[name = tensor("reshape_35")]; + tensor slice_by_index_7 = slice_by_index(begin = concat_28, begin_mask = new_cache_7_internal_tensor_assign_2_begin_mask_0, end = concat_29, end_mask = new_cache_7_internal_tensor_assign_2_end_mask_0, squeeze_mask = new_cache_7_internal_tensor_assign_2_squeeze_mask_0, stride = new_cache_7_internal_tensor_assign_2_stride_0, x = reshape_35)[name = tensor("slice_by_index_7")]; + tensor reshape_36_shape_0 = const()[name = tensor("reshape_36_shape_0"), val = tensor([-1])]; + tensor reshape_36 = reshape(shape = reshape_36_shape_0, x = slice_by_index_7)[name = tensor("reshape_36")]; + tensor reshape_37_shape_0 = const()[name = tensor("reshape_37_shape_0"), val = tensor([-1])]; + tensor reshape_37_cast_fp16 = reshape(shape = reshape_37_shape_0, x = var_1541_cast_fp16)[name = tensor("reshape_37_cast_fp16")]; + tensor reshape_38_shape_0 = const()[name = tensor("reshape_38_shape_0"), val = tensor([-1])]; + tensor reshape_38_cast_fp16 = reshape(shape = reshape_38_shape_0, x = reshape_34_cast_fp16)[name = tensor("reshape_38_cast_fp16")]; + tensor scatter_7_mode_0 = const()[name = tensor("scatter_7_mode_0"), val = tensor("update")]; + tensor scatter_7_axis_0 = const()[name = tensor("scatter_7_axis_0"), val = tensor(0)]; + tensor scatter_7_validate_indices_0 = const()[name = tensor("scatter_7_validate_indices_0"), val = tensor(false)]; + tensor scatter_7_cast_fp16 = scatter(axis = scatter_7_axis_0, data = reshape_38_cast_fp16, indices = reshape_36, mode = scatter_7_mode_0, updates = reshape_37_cast_fp16, validate_indices = scatter_7_validate_indices_0)[name = tensor("scatter_7_cast_fp16")]; + tensor reshape_39_cast_fp16 = reshape(shape = shape_17, x = scatter_7_cast_fp16)[name = tensor("reshape_39_cast_fp16")]; + tensor reshape_39_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("reshape_39_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor keys_19_begin_0 = const()[name = tensor("keys_19_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor keys_19_end_0 = const()[name = tensor("keys_19_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor keys_19_end_mask_0 = const()[name = tensor("keys_19_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor keys_19_squeeze_mask_0 = const()[name = tensor("keys_19_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor keys_19_cast_fp16 = slice_by_index(begin = keys_19_begin_0, end = keys_19_end_0, end_mask = keys_19_end_mask_0, squeeze_mask = keys_19_squeeze_mask_0, x = reshape_39_cast_fp16)[name = tensor("keys_19_cast_fp16")]; + tensor values_19_begin_0 = const()[name = tensor("values_19_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor values_19_end_0 = const()[name = tensor("values_19_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor values_19_end_mask_0 = const()[name = tensor("values_19_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor values_19_squeeze_mask_0 = const()[name = tensor("values_19_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor values_19_cast_fp16 = slice_by_index(begin = values_19_begin_0, end = values_19_end_0, end_mask = values_19_end_mask_0, squeeze_mask = values_19_squeeze_mask_0, x = reshape_39_cast_fp16)[name = tensor("values_19_cast_fp16")]; + tensor var_1553_cast_fp16 = not_equal(x = keys_19_cast_fp16, y = keys_19_cast_fp16)[name = tensor("op_1553_cast_fp16")]; + tensor keys_21_cast_fp16 = select(a = var_326_to_fp16, b = keys_19_cast_fp16, cond = var_1553_cast_fp16)[name = tensor("keys_21_cast_fp16")]; + tensor var_1561_cast_fp16 = not_equal(x = values_19_cast_fp16, y = values_19_cast_fp16)[name = tensor("op_1561_cast_fp16")]; + tensor values_21_cast_fp16 = select(a = var_326_to_fp16, b = values_19_cast_fp16, cond = var_1561_cast_fp16)[name = tensor("values_21_cast_fp16")]; + tensor var_1585 = const()[name = tensor("op_1585"), val = tensor([0, 2, 1, 3])]; + tensor var_1601 = const()[name = tensor("op_1601"), val = tensor([1, 1, 1])]; + tensor var_1602_cast_fp16 = reshape(shape = var_1601, x = position3_to_fp16)[name = tensor("op_1602_cast_fp16")]; + tensor q_positions_7_cast_fp16 = add(x = var_1602_cast_fp16, y = q_offsets_1_promoted_to_fp16)[name = tensor("q_positions_7_cast_fp16")]; + tensor valid_end_7_cast_fp16 = add(x = var_1602_cast_fp16, y = var_392_cast_fp16)[name = tensor("valid_end_7_cast_fp16")]; + tensor valid_mask_7_cast_fp16 = less(x = k_positions_1_promoted_to_fp16, y = valid_end_7_cast_fp16)[name = tensor("valid_mask_7_cast_fp16")]; + tensor causal_mask_7_cast_fp16 = less_equal(x = k_positions_1_promoted_to_fp16, y = q_positions_7_cast_fp16)[name = tensor("causal_mask_7_cast_fp16")]; + tensor var_1630 = logical_and(x = valid_mask_7_cast_fp16, y = causal_mask_7_cast_fp16)[name = tensor("op_1630")]; + tensor attn_mask_7_axes_0 = const()[name = tensor("attn_mask_7_axes_0"), val = tensor([1])]; + tensor attn_mask_7 = expand_dims(axes = attn_mask_7_axes_0, x = var_1630)[name = tensor("attn_mask_7")]; + tensor var_1644_transpose_x_0 = const()[name = tensor("op_1644_transpose_x_0"), val = tensor(false)]; + tensor var_1644_transpose_y_0 = const()[name = tensor("op_1644_transpose_y_0"), val = tensor(false)]; + tensor transpose_21_perm_0 = const()[name = tensor("transpose_21_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_22_perm_0 = const()[name = tensor("transpose_22_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_22 = transpose(perm = transpose_22_perm_0, x = keys_21_cast_fp16)[name = tensor("transpose_30")]; + tensor transpose_21 = transpose(perm = transpose_21_perm_0, x = q_21_cast_fp16)[name = tensor("transpose_31")]; + tensor var_1644_cast_fp16 = matmul(transpose_x = var_1644_transpose_x_0, transpose_y = var_1644_transpose_y_0, x = transpose_21, y = transpose_22)[name = tensor("op_1644_cast_fp16")]; + tensor var_1638_to_fp16 = const()[name = tensor("op_1638_to_fp16"), val = tensor([0x1p-3])]; + tensor attn_weights_19_cast_fp16 = mul(x = var_1644_cast_fp16, y = var_1638_to_fp16)[name = tensor("attn_weights_19_cast_fp16")]; + tensor var_1646 = logical_not(x = attn_mask_7)[name = tensor("op_1646")]; + tensor var_1647_to_fp16 = const()[name = tensor("op_1647_to_fp16"), val = tensor(-inf)]; + tensor attn_weights_21_cast_fp16 = select(a = var_1647_to_fp16, b = attn_weights_19_cast_fp16, cond = var_1646)[name = tensor("attn_weights_21_cast_fp16")]; + tensor var_1649 = const()[name = tensor("op_1649"), val = tensor(-1)]; + tensor attn_weights_23_cast_fp16 = softmax(axis = var_1649, x = attn_weights_21_cast_fp16)[name = tensor("attn_weights_23_cast_fp16")]; + tensor attn_output_7_transpose_x_0 = const()[name = tensor("attn_output_7_transpose_x_0"), val = tensor(false)]; + tensor attn_output_7_transpose_y_0 = const()[name = tensor("attn_output_7_transpose_y_0"), val = tensor(false)]; + tensor values_23_cast_fp16 = transpose(perm = var_1585, x = values_21_cast_fp16)[name = tensor("transpose_32")]; + tensor attn_output_7_cast_fp16 = matmul(transpose_x = attn_output_7_transpose_x_0, transpose_y = attn_output_7_transpose_y_0, x = attn_weights_23_cast_fp16, y = values_23_cast_fp16)[name = tensor("attn_output_7_cast_fp16")]; + tensor var_1657 = const()[name = tensor("op_1657"), val = tensor([0, 2, 1, 3])]; + tensor var_1662 = const()[name = tensor("op_1662"), val = tensor([1, 256, 1024])]; + tensor var_1658_cast_fp16 = transpose(perm = var_1657, x = attn_output_7_cast_fp16)[name = tensor("transpose_29")]; + tensor input_41_cast_fp16 = reshape(shape = var_1662, x = var_1658_cast_fp16)[name = tensor("input_41_cast_fp16")]; + tensor attn3_out_proj_weight_to_fp16 = const()[name = tensor("attn3_out_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(82887936)))]; + tensor linear_13_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = attn3_out_proj_weight_to_fp16, x = input_41_cast_fp16)[name = tensor("linear_13_cast_fp16")]; + tensor var_1671_cast_fp16 = add(x = position3_to_fp16, y = valid_len_to_fp16)[name = tensor("op_1671_cast_fp16")]; + tensor var_1671_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_1671_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor input_43_cast_fp16 = add(x = input_37_cast_fp16, y = linear_13_cast_fp16)[name = tensor("input_43_cast_fp16")]; + tensor input_45_axes_0 = const()[name = tensor("input_45_axes_0"), val = tensor([-1])]; + tensor norm3_2_weight_to_fp16 = const()[name = tensor("norm3_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(84985152)))]; + tensor norm3_2_bias_to_fp16 = const()[name = tensor("norm3_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(84987264)))]; + tensor var_1675_to_fp16 = const()[name = tensor("op_1675_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_45_cast_fp16 = layer_norm(axes = input_45_axes_0, beta = norm3_2_bias_to_fp16, epsilon = var_1675_to_fp16, gamma = norm3_2_weight_to_fp16, x = input_43_cast_fp16)[name = tensor("input_45_cast_fp16")]; + tensor linear3_1_weight_to_fp16 = const()[name = tensor("linear3_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(84989376)))]; + tensor linear_14_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = linear3_1_weight_to_fp16, x = input_45_cast_fp16)[name = tensor("linear_14_cast_fp16")]; + tensor input_47_mode_0 = const()[name = tensor("input_47_mode_0"), val = tensor("EXACT")]; + tensor input_47_cast_fp16 = gelu(mode = input_47_mode_0, x = linear_14_cast_fp16)[name = tensor("input_47_cast_fp16")]; + tensor linear3_2_weight_to_fp16 = const()[name = tensor("linear3_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(93378048)))]; + tensor linear_15_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = linear3_2_weight_to_fp16, x = input_47_cast_fp16)[name = tensor("linear_15_cast_fp16")]; + tensor input_49_cast_fp16 = add(x = input_43_cast_fp16, y = linear_15_cast_fp16)[name = tensor("input_49_cast_fp16")]; + tensor input_51_axes_0 = const()[name = tensor("input_51_axes_0"), val = tensor([-1])]; + tensor norm4_1_weight_to_fp16 = const()[name = tensor("norm4_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(101766720)))]; + tensor norm4_1_bias_to_fp16 = const()[name = tensor("norm4_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(101768832)))]; + tensor var_1692_to_fp16 = const()[name = tensor("op_1692_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_51_cast_fp16 = layer_norm(axes = input_51_axes_0, beta = norm4_1_bias_to_fp16, epsilon = var_1692_to_fp16, gamma = norm4_1_weight_to_fp16, x = input_49_cast_fp16)[name = tensor("input_51_cast_fp16")]; + tensor attn4_in_proj_weight_to_fp16 = const()[name = tensor("attn4_in_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(101770944)))]; + tensor linear_16_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = attn4_in_proj_weight_to_fp16, x = input_51_cast_fp16)[name = tensor("linear_16_cast_fp16")]; + tensor var_1706 = const()[name = tensor("op_1706"), val = tensor([1, 256, 3, 16, 64])]; + tensor qkv_9_cast_fp16 = reshape(shape = var_1706, x = linear_16_cast_fp16)[name = tensor("qkv_9_cast_fp16")]; + tensor q_25_begin_0 = const()[name = tensor("q_25_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor q_25_end_0 = const()[name = tensor("q_25_end_0"), val = tensor([1, 256, 1, 16, 64])]; + tensor q_25_end_mask_0 = const()[name = tensor("q_25_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor q_25_squeeze_mask_0 = const()[name = tensor("q_25_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor q_25_cast_fp16 = slice_by_index(begin = q_25_begin_0, end = q_25_end_0, end_mask = q_25_end_mask_0, squeeze_mask = q_25_squeeze_mask_0, x = qkv_9_cast_fp16)[name = tensor("q_25_cast_fp16")]; + tensor k_17_begin_0 = const()[name = tensor("k_17_begin_0"), val = tensor([0, 0, 1, 0, 0])]; + tensor k_17_end_0 = const()[name = tensor("k_17_end_0"), val = tensor([1, 256, 2, 16, 64])]; + tensor k_17_end_mask_0 = const()[name = tensor("k_17_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor k_17_squeeze_mask_0 = const()[name = tensor("k_17_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor k_17_cast_fp16 = slice_by_index(begin = k_17_begin_0, end = k_17_end_0, end_mask = k_17_end_mask_0, squeeze_mask = k_17_squeeze_mask_0, x = qkv_9_cast_fp16)[name = tensor("k_17_cast_fp16")]; + tensor v_9_begin_0 = const()[name = tensor("v_9_begin_0"), val = tensor([0, 0, 2, 0, 0])]; + tensor v_9_end_0 = const()[name = tensor("v_9_end_0"), val = tensor([1, 256, 3, 16, 64])]; + tensor v_9_end_mask_0 = const()[name = tensor("v_9_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor v_9_squeeze_mask_0 = const()[name = tensor("v_9_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor v_9_cast_fp16 = slice_by_index(begin = v_9_begin_0, end = v_9_end_0, end_mask = v_9_end_mask_0, squeeze_mask = v_9_squeeze_mask_0, x = qkv_9_cast_fp16)[name = tensor("v_9_cast_fp16")]; + tensor var_1764 = const()[name = tensor("op_1764"), val = tensor([1, 1])]; + tensor position4_to_fp16_dtype_0 = const()[name = tensor("position4_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor position4_to_fp16 = cast(dtype = position4_to_fp16_dtype_0, x = position4)[name = tensor("cast_137")]; + tensor var_1765_cast_fp16 = reshape(shape = var_1764, x = position4_to_fp16)[name = tensor("op_1765_cast_fp16")]; + tensor ts_27_cast_fp16 = add(x = ts_1_promoted_to_fp16, y = var_1765_cast_fp16)[name = tensor("ts_27_cast_fp16")]; + tensor var_1772 = const()[name = tensor("op_1772"), val = tensor([1, 256, 1, 1])]; + tensor ts_29_cast_fp16 = reshape(shape = var_1772, x = ts_27_cast_fp16)[name = tensor("ts_29_cast_fp16")]; + tensor var_1779 = const()[name = tensor("op_1779"), val = tensor([1, 256, 16, 32, 2])]; + tensor q_complex_9_cast_fp16 = reshape(shape = var_1779, x = q_25_cast_fp16)[name = tensor("q_complex_9_cast_fp16")]; + tensor var_1786 = const()[name = tensor("op_1786"), val = tensor([1, 256, 16, 32, 2])]; + tensor k_complex_9_cast_fp16 = reshape(shape = var_1786, x = k_17_cast_fp16)[name = tensor("k_complex_9_cast_fp16")]; + tensor var_1790_begin_0 = const()[name = tensor("op_1790_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_1790_end_0 = const()[name = tensor("op_1790_end_0"), val = tensor([1, 256, 16, 32, 1])]; + tensor var_1790_end_mask_0 = const()[name = tensor("op_1790_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_1790_squeeze_mask_0 = const()[name = tensor("op_1790_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_1790_cast_fp16 = slice_by_index(begin = var_1790_begin_0, end = var_1790_end_0, end_mask = var_1790_end_mask_0, squeeze_mask = var_1790_squeeze_mask_0, x = q_complex_9_cast_fp16)[name = tensor("op_1790_cast_fp16")]; + tensor var_1798_begin_0 = const()[name = tensor("op_1798_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_1798_end_0 = const()[name = tensor("op_1798_end_0"), val = tensor([1, 256, 16, 32, 2])]; + tensor var_1798_end_mask_0 = const()[name = tensor("op_1798_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_1798_squeeze_mask_0 = const()[name = tensor("op_1798_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_1798_cast_fp16 = slice_by_index(begin = var_1798_begin_0, end = var_1798_end_0, end_mask = var_1798_end_mask_0, squeeze_mask = var_1798_squeeze_mask_0, x = q_complex_9_cast_fp16)[name = tensor("op_1798_cast_fp16")]; + tensor var_1806_begin_0 = const()[name = tensor("op_1806_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_1806_end_0 = const()[name = tensor("op_1806_end_0"), val = tensor([1, 256, 16, 32, 1])]; + tensor var_1806_end_mask_0 = const()[name = tensor("op_1806_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_1806_squeeze_mask_0 = const()[name = tensor("op_1806_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_1806_cast_fp16 = slice_by_index(begin = var_1806_begin_0, end = var_1806_end_0, end_mask = var_1806_end_mask_0, squeeze_mask = var_1806_squeeze_mask_0, x = k_complex_9_cast_fp16)[name = tensor("op_1806_cast_fp16")]; + tensor var_1814_begin_0 = const()[name = tensor("op_1814_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_1814_end_0 = const()[name = tensor("op_1814_end_0"), val = tensor([1, 256, 16, 32, 2])]; + tensor var_1814_end_mask_0 = const()[name = tensor("op_1814_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_1814_squeeze_mask_0 = const()[name = tensor("op_1814_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_1814_cast_fp16 = slice_by_index(begin = var_1814_begin_0, end = var_1814_end_0, end_mask = var_1814_end_mask_0, squeeze_mask = var_1814_squeeze_mask_0, x = k_complex_9_cast_fp16)[name = tensor("op_1814_cast_fp16")]; + tensor freqs_9_to_fp16 = const()[name = tensor("freqs_9_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(108062464)))]; + tensor var_1820_cast_fp16 = mul(x = freqs_9_to_fp16, y = ts_29_cast_fp16)[name = tensor("op_1820_cast_fp16")]; + tensor rotr_9_cast_fp16 = cos(x = var_1820_cast_fp16)[name = tensor("rotr_9_cast_fp16")]; + tensor roti_9_cast_fp16 = sin(x = var_1820_cast_fp16)[name = tensor("roti_9_cast_fp16")]; + tensor var_1824_cast_fp16 = mul(x = var_1790_cast_fp16, y = rotr_9_cast_fp16)[name = tensor("op_1824_cast_fp16")]; + tensor var_1825_cast_fp16 = mul(x = var_1798_cast_fp16, y = roti_9_cast_fp16)[name = tensor("op_1825_cast_fp16")]; + tensor qor_17_cast_fp16 = sub(x = var_1824_cast_fp16, y = var_1825_cast_fp16)[name = tensor("qor_17_cast_fp16")]; + tensor var_1828_cast_fp16 = mul(x = var_1790_cast_fp16, y = roti_9_cast_fp16)[name = tensor("op_1828_cast_fp16")]; + tensor var_1829_cast_fp16 = mul(x = var_1798_cast_fp16, y = rotr_9_cast_fp16)[name = tensor("op_1829_cast_fp16")]; + tensor qoi_17_cast_fp16 = add(x = var_1828_cast_fp16, y = var_1829_cast_fp16)[name = tensor("qoi_17_cast_fp16")]; + tensor var_1832_cast_fp16 = mul(x = var_1806_cast_fp16, y = rotr_9_cast_fp16)[name = tensor("op_1832_cast_fp16")]; + tensor var_1833_cast_fp16 = mul(x = var_1814_cast_fp16, y = roti_9_cast_fp16)[name = tensor("op_1833_cast_fp16")]; + tensor kor_17_cast_fp16 = sub(x = var_1832_cast_fp16, y = var_1833_cast_fp16)[name = tensor("kor_17_cast_fp16")]; + tensor var_1836_cast_fp16 = mul(x = var_1806_cast_fp16, y = roti_9_cast_fp16)[name = tensor("op_1836_cast_fp16")]; + tensor var_1837_cast_fp16 = mul(x = var_1814_cast_fp16, y = rotr_9_cast_fp16)[name = tensor("op_1837_cast_fp16")]; + tensor koi_17_cast_fp16 = add(x = var_1836_cast_fp16, y = var_1837_cast_fp16)[name = tensor("koi_17_cast_fp16")]; + tensor qo_9_axis_0 = const()[name = tensor("qo_9_axis_0"), val = tensor(-1)]; + tensor qo_9_cast_fp16 = stack(axis = qo_9_axis_0, values = (qor_17_cast_fp16, qoi_17_cast_fp16))[name = tensor("qo_9_cast_fp16")]; + tensor ko_9_axis_0 = const()[name = tensor("ko_9_axis_0"), val = tensor(-1)]; + tensor ko_9_cast_fp16 = stack(axis = ko_9_axis_0, values = (kor_17_cast_fp16, koi_17_cast_fp16))[name = tensor("ko_9_cast_fp16")]; + tensor var_1870 = const()[name = tensor("op_1870"), val = tensor([1, 256, 16, 64])]; + tensor q_27_cast_fp16 = reshape(shape = var_1870, x = qo_9_cast_fp16)[name = tensor("q_27_cast_fp16")]; + tensor var_1876 = const()[name = tensor("op_1876"), val = tensor([1, 256, 16, 64])]; + tensor k_19_cast_fp16 = reshape(shape = var_1876, x = ko_9_cast_fp16)[name = tensor("k_19_cast_fp16")]; + tensor write_indices_float_25_cast_fp16 = add(x = var_1765_cast_fp16, y = var_249_promoted_to_fp16)[name = tensor("write_indices_float_25_cast_fp16")]; + tensor _inversed_1897_y_0_to_fp16 = const()[name = tensor("_inversed_1897_y_0_to_fp16"), val = tensor(0x1p-9)]; + tensor _inversed_1897_cast_fp16 = mul(x = write_indices_float_25_cast_fp16, y = _inversed_1897_y_0_to_fp16)[name = tensor("_inversed_1897_cast_fp16")]; + tensor var_1898_cast_fp16 = floor(x = _inversed_1897_cast_fp16)[name = tensor("op_1898_cast_fp16")]; + tensor var_1899_to_fp16 = const()[name = tensor("op_1899_to_fp16"), val = tensor(0x1p+9)]; + tensor var_1900_cast_fp16 = mul(x = var_1898_cast_fp16, y = var_1899_to_fp16)[name = tensor("op_1900_cast_fp16")]; + tensor write_indices_float_27_cast_fp16 = sub(x = write_indices_float_25_cast_fp16, y = var_1900_cast_fp16)[name = tensor("write_indices_float_27_cast_fp16")]; + tensor var_1918_value_0_to_fp16 = const()[name = tensor("op_1918_value_0_to_fp16"), val = tensor(0x1.ffp+8)]; + tensor var_1918_cast_fp16 = fill_like(ref_tensor = write_indices_float_27_cast_fp16, value = var_1918_value_0_to_fp16)[name = tensor("op_1918_cast_fp16")]; + tensor write_indices_float_29_cast_fp16 = select(a = write_indices_float_27_cast_fp16, b = var_1918_cast_fp16, cond = token_valid_1_cast_fp16)[name = tensor("write_indices_float_29_cast_fp16")]; + tensor var_1924_dtype_0 = const()[name = tensor("op_1924_dtype_0"), val = tensor("int32")]; + tensor var_1929 = const()[name = tensor("op_1929"), val = tensor([1, 256, 1, 1])]; + tensor write_indices_float_29_cast_fp16_to_int32 = cast(dtype = var_1924_dtype_0, x = write_indices_float_29_cast_fp16)[name = tensor("cast_136")]; + tensor var_1930 = reshape(shape = var_1929, x = write_indices_float_29_cast_fp16_to_int32)[name = tensor("op_1930")]; + tensor write_indices_9_reps_0 = const()[name = tensor("write_indices_9_reps_0"), val = tensor([1, 1, 16, 64])]; + tensor write_indices_9 = tile(reps = write_indices_9_reps_0, x = var_1930)[name = tensor("write_indices_9")]; + tensor var_1940_begin_0 = const()[name = tensor("op_1940_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_1940_end_0 = const()[name = tensor("op_1940_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor var_1940_end_mask_0 = const()[name = tensor("op_1940_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_1940_squeeze_mask_0 = const()[name = tensor("op_1940_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor cache4_to_fp16_dtype_0 = const()[name = tensor("cache4_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor cache4_to_fp16 = cast(dtype = cache4_to_fp16_dtype_0, x = cache4)[name = tensor("cast_135")]; + tensor var_1940_cast_fp16 = slice_by_index(begin = var_1940_begin_0, end = var_1940_end_0, end_mask = var_1940_end_mask_0, squeeze_mask = var_1940_squeeze_mask_0, x = cache4_to_fp16)[name = tensor("op_1940_cast_fp16")]; + tensor var_1942_axis_0 = const()[name = tensor("op_1942_axis_0"), val = tensor(1)]; + tensor var_1942_mode_0 = const()[name = tensor("op_1942_mode_0"), val = tensor("update")]; + tensor var_1942_validate_indices_0 = const()[name = tensor("op_1942_validate_indices_0"), val = tensor(false)]; + tensor var_1942_cast_fp16 = scatter_along_axis(axis = var_1942_axis_0, data = var_1940_cast_fp16, indices = write_indices_9, mode = var_1942_mode_0, updates = k_19_cast_fp16, validate_indices = var_1942_validate_indices_0)[name = tensor("op_1942_cast_fp16")]; + tensor concat_34 = const()[name = tensor("concat_34"), val = tensor([0, 0, 0, 0, 0])]; + tensor concat_35 = const()[name = tensor("concat_35"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_9_internal_tensor_assign_1_stride_0 = const()[name = tensor("new_cache_9_internal_tensor_assign_1_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_9_internal_tensor_assign_1_begin_mask_0 = const()[name = tensor("new_cache_9_internal_tensor_assign_1_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_9_internal_tensor_assign_1_end_mask_0 = const()[name = tensor("new_cache_9_internal_tensor_assign_1_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_9_internal_tensor_assign_1_squeeze_mask_0 = const()[name = tensor("new_cache_9_internal_tensor_assign_1_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_18 = const()[name = tensor("shape_18"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_8 = const()[name = tensor("reduce_prod_8"), val = tensor(1048576)]; + tensor range_1d_8_start_0 = const()[name = tensor("range_1d_8_start_0"), val = tensor(0)]; + tensor range_1d_8_step_0 = const()[name = tensor("range_1d_8_step_0"), val = tensor(1)]; + tensor range_1d_8 = range_1d(end = reduce_prod_8, start = range_1d_8_start_0, step = range_1d_8_step_0)[name = tensor("range_1d_8")]; + tensor reshape_40 = reshape(shape = shape_18, x = range_1d_8)[name = tensor("reshape_40")]; + tensor slice_by_index_8 = slice_by_index(begin = concat_34, begin_mask = new_cache_9_internal_tensor_assign_1_begin_mask_0, end = concat_35, end_mask = new_cache_9_internal_tensor_assign_1_end_mask_0, squeeze_mask = new_cache_9_internal_tensor_assign_1_squeeze_mask_0, stride = new_cache_9_internal_tensor_assign_1_stride_0, x = reshape_40)[name = tensor("slice_by_index_8")]; + tensor reshape_41_shape_0 = const()[name = tensor("reshape_41_shape_0"), val = tensor([-1])]; + tensor reshape_41 = reshape(shape = reshape_41_shape_0, x = slice_by_index_8)[name = tensor("reshape_41")]; + tensor reshape_42_shape_0 = const()[name = tensor("reshape_42_shape_0"), val = tensor([-1])]; + tensor reshape_42_cast_fp16 = reshape(shape = reshape_42_shape_0, x = var_1942_cast_fp16)[name = tensor("reshape_42_cast_fp16")]; + tensor reshape_43_shape_0 = const()[name = tensor("reshape_43_shape_0"), val = tensor([-1])]; + tensor reshape_43_cast_fp16 = reshape(shape = reshape_43_shape_0, x = cache4_to_fp16)[name = tensor("reshape_43_cast_fp16")]; + tensor scatter_8_mode_0 = const()[name = tensor("scatter_8_mode_0"), val = tensor("update")]; + tensor scatter_8_axis_0 = const()[name = tensor("scatter_8_axis_0"), val = tensor(0)]; + tensor scatter_8_validate_indices_0 = const()[name = tensor("scatter_8_validate_indices_0"), val = tensor(false)]; + tensor scatter_8_cast_fp16 = scatter(axis = scatter_8_axis_0, data = reshape_43_cast_fp16, indices = reshape_41, mode = scatter_8_mode_0, updates = reshape_42_cast_fp16, validate_indices = scatter_8_validate_indices_0)[name = tensor("scatter_8_cast_fp16")]; + tensor reshape_44_cast_fp16 = reshape(shape = shape_18, x = scatter_8_cast_fp16)[name = tensor("reshape_44_cast_fp16")]; + tensor var_1950_begin_0 = const()[name = tensor("op_1950_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor var_1950_end_0 = const()[name = tensor("op_1950_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor var_1950_end_mask_0 = const()[name = tensor("op_1950_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_1950_squeeze_mask_0 = const()[name = tensor("op_1950_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor var_1950_cast_fp16 = slice_by_index(begin = var_1950_begin_0, end = var_1950_end_0, end_mask = var_1950_end_mask_0, squeeze_mask = var_1950_squeeze_mask_0, x = reshape_44_cast_fp16)[name = tensor("op_1950_cast_fp16")]; + tensor var_1952_axis_0 = const()[name = tensor("op_1952_axis_0"), val = tensor(1)]; + tensor var_1952_mode_0 = const()[name = tensor("op_1952_mode_0"), val = tensor("update")]; + tensor var_1952_validate_indices_0 = const()[name = tensor("op_1952_validate_indices_0"), val = tensor(false)]; + tensor var_1952_cast_fp16 = scatter_along_axis(axis = var_1952_axis_0, data = var_1950_cast_fp16, indices = write_indices_9, mode = var_1952_mode_0, updates = v_9_cast_fp16, validate_indices = var_1952_validate_indices_0)[name = tensor("op_1952_cast_fp16")]; + tensor concat_36 = const()[name = tensor("concat_36"), val = tensor([1, 0, 0, 0, 0])]; + tensor concat_37 = const()[name = tensor("concat_37"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_9_internal_tensor_assign_2_stride_0 = const()[name = tensor("new_cache_9_internal_tensor_assign_2_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_9_internal_tensor_assign_2_begin_mask_0 = const()[name = tensor("new_cache_9_internal_tensor_assign_2_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_9_internal_tensor_assign_2_end_mask_0 = const()[name = tensor("new_cache_9_internal_tensor_assign_2_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_9_internal_tensor_assign_2_squeeze_mask_0 = const()[name = tensor("new_cache_9_internal_tensor_assign_2_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_19 = const()[name = tensor("shape_19"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_9 = const()[name = tensor("reduce_prod_9"), val = tensor(1048576)]; + tensor range_1d_9_start_0 = const()[name = tensor("range_1d_9_start_0"), val = tensor(0)]; + tensor range_1d_9_step_0 = const()[name = tensor("range_1d_9_step_0"), val = tensor(1)]; + tensor range_1d_9 = range_1d(end = reduce_prod_9, start = range_1d_9_start_0, step = range_1d_9_step_0)[name = tensor("range_1d_9")]; + tensor reshape_45 = reshape(shape = shape_19, x = range_1d_9)[name = tensor("reshape_45")]; + tensor slice_by_index_9 = slice_by_index(begin = concat_36, begin_mask = new_cache_9_internal_tensor_assign_2_begin_mask_0, end = concat_37, end_mask = new_cache_9_internal_tensor_assign_2_end_mask_0, squeeze_mask = new_cache_9_internal_tensor_assign_2_squeeze_mask_0, stride = new_cache_9_internal_tensor_assign_2_stride_0, x = reshape_45)[name = tensor("slice_by_index_9")]; + tensor reshape_46_shape_0 = const()[name = tensor("reshape_46_shape_0"), val = tensor([-1])]; + tensor reshape_46 = reshape(shape = reshape_46_shape_0, x = slice_by_index_9)[name = tensor("reshape_46")]; + tensor reshape_47_shape_0 = const()[name = tensor("reshape_47_shape_0"), val = tensor([-1])]; + tensor reshape_47_cast_fp16 = reshape(shape = reshape_47_shape_0, x = var_1952_cast_fp16)[name = tensor("reshape_47_cast_fp16")]; + tensor reshape_48_shape_0 = const()[name = tensor("reshape_48_shape_0"), val = tensor([-1])]; + tensor reshape_48_cast_fp16 = reshape(shape = reshape_48_shape_0, x = reshape_44_cast_fp16)[name = tensor("reshape_48_cast_fp16")]; + tensor scatter_9_mode_0 = const()[name = tensor("scatter_9_mode_0"), val = tensor("update")]; + tensor scatter_9_axis_0 = const()[name = tensor("scatter_9_axis_0"), val = tensor(0)]; + tensor scatter_9_validate_indices_0 = const()[name = tensor("scatter_9_validate_indices_0"), val = tensor(false)]; + tensor scatter_9_cast_fp16 = scatter(axis = scatter_9_axis_0, data = reshape_48_cast_fp16, indices = reshape_46, mode = scatter_9_mode_0, updates = reshape_47_cast_fp16, validate_indices = scatter_9_validate_indices_0)[name = tensor("scatter_9_cast_fp16")]; + tensor reshape_49_cast_fp16 = reshape(shape = shape_19, x = scatter_9_cast_fp16)[name = tensor("reshape_49_cast_fp16")]; + tensor reshape_49_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("reshape_49_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor keys_25_begin_0 = const()[name = tensor("keys_25_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor keys_25_end_0 = const()[name = tensor("keys_25_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor keys_25_end_mask_0 = const()[name = tensor("keys_25_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor keys_25_squeeze_mask_0 = const()[name = tensor("keys_25_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor keys_25_cast_fp16 = slice_by_index(begin = keys_25_begin_0, end = keys_25_end_0, end_mask = keys_25_end_mask_0, squeeze_mask = keys_25_squeeze_mask_0, x = reshape_49_cast_fp16)[name = tensor("keys_25_cast_fp16")]; + tensor values_25_begin_0 = const()[name = tensor("values_25_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor values_25_end_0 = const()[name = tensor("values_25_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor values_25_end_mask_0 = const()[name = tensor("values_25_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor values_25_squeeze_mask_0 = const()[name = tensor("values_25_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor values_25_cast_fp16 = slice_by_index(begin = values_25_begin_0, end = values_25_end_0, end_mask = values_25_end_mask_0, squeeze_mask = values_25_squeeze_mask_0, x = reshape_49_cast_fp16)[name = tensor("values_25_cast_fp16")]; + tensor var_1964_cast_fp16 = not_equal(x = keys_25_cast_fp16, y = keys_25_cast_fp16)[name = tensor("op_1964_cast_fp16")]; + tensor keys_27_cast_fp16 = select(a = var_326_to_fp16, b = keys_25_cast_fp16, cond = var_1964_cast_fp16)[name = tensor("keys_27_cast_fp16")]; + tensor var_1972_cast_fp16 = not_equal(x = values_25_cast_fp16, y = values_25_cast_fp16)[name = tensor("op_1972_cast_fp16")]; + tensor values_27_cast_fp16 = select(a = var_326_to_fp16, b = values_25_cast_fp16, cond = var_1972_cast_fp16)[name = tensor("values_27_cast_fp16")]; + tensor var_1996 = const()[name = tensor("op_1996"), val = tensor([0, 2, 1, 3])]; + tensor var_2012 = const()[name = tensor("op_2012"), val = tensor([1, 1, 1])]; + tensor var_2013_cast_fp16 = reshape(shape = var_2012, x = position4_to_fp16)[name = tensor("op_2013_cast_fp16")]; + tensor q_positions_9_cast_fp16 = add(x = var_2013_cast_fp16, y = q_offsets_1_promoted_to_fp16)[name = tensor("q_positions_9_cast_fp16")]; + tensor valid_end_9_cast_fp16 = add(x = var_2013_cast_fp16, y = var_392_cast_fp16)[name = tensor("valid_end_9_cast_fp16")]; + tensor valid_mask_9_cast_fp16 = less(x = k_positions_1_promoted_to_fp16, y = valid_end_9_cast_fp16)[name = tensor("valid_mask_9_cast_fp16")]; + tensor causal_mask_9_cast_fp16 = less_equal(x = k_positions_1_promoted_to_fp16, y = q_positions_9_cast_fp16)[name = tensor("causal_mask_9_cast_fp16")]; + tensor var_2041 = logical_and(x = valid_mask_9_cast_fp16, y = causal_mask_9_cast_fp16)[name = tensor("op_2041")]; + tensor attn_mask_9_axes_0 = const()[name = tensor("attn_mask_9_axes_0"), val = tensor([1])]; + tensor attn_mask_9 = expand_dims(axes = attn_mask_9_axes_0, x = var_2041)[name = tensor("attn_mask_9")]; + tensor var_2055_transpose_x_0 = const()[name = tensor("op_2055_transpose_x_0"), val = tensor(false)]; + tensor var_2055_transpose_y_0 = const()[name = tensor("op_2055_transpose_y_0"), val = tensor(false)]; + tensor transpose_23_perm_0 = const()[name = tensor("transpose_23_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_24_perm_0 = const()[name = tensor("transpose_24_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_24 = transpose(perm = transpose_24_perm_0, x = keys_27_cast_fp16)[name = tensor("transpose_26")]; + tensor transpose_23 = transpose(perm = transpose_23_perm_0, x = q_27_cast_fp16)[name = tensor("transpose_27")]; + tensor var_2055_cast_fp16 = matmul(transpose_x = var_2055_transpose_x_0, transpose_y = var_2055_transpose_y_0, x = transpose_23, y = transpose_24)[name = tensor("op_2055_cast_fp16")]; + tensor var_2049_to_fp16 = const()[name = tensor("op_2049_to_fp16"), val = tensor([0x1p-3])]; + tensor attn_weights_25_cast_fp16 = mul(x = var_2055_cast_fp16, y = var_2049_to_fp16)[name = tensor("attn_weights_25_cast_fp16")]; + tensor var_2057 = logical_not(x = attn_mask_9)[name = tensor("op_2057")]; + tensor var_2058_to_fp16 = const()[name = tensor("op_2058_to_fp16"), val = tensor(-inf)]; + tensor attn_weights_27_cast_fp16 = select(a = var_2058_to_fp16, b = attn_weights_25_cast_fp16, cond = var_2057)[name = tensor("attn_weights_27_cast_fp16")]; + tensor var_2060 = const()[name = tensor("op_2060"), val = tensor(-1)]; + tensor attn_weights_29_cast_fp16 = softmax(axis = var_2060, x = attn_weights_27_cast_fp16)[name = tensor("attn_weights_29_cast_fp16")]; + tensor attn_output_9_transpose_x_0 = const()[name = tensor("attn_output_9_transpose_x_0"), val = tensor(false)]; + tensor attn_output_9_transpose_y_0 = const()[name = tensor("attn_output_9_transpose_y_0"), val = tensor(false)]; + tensor values_29_cast_fp16 = transpose(perm = var_1996, x = values_27_cast_fp16)[name = tensor("transpose_28")]; + tensor attn_output_9_cast_fp16 = matmul(transpose_x = attn_output_9_transpose_x_0, transpose_y = attn_output_9_transpose_y_0, x = attn_weights_29_cast_fp16, y = values_29_cast_fp16)[name = tensor("attn_output_9_cast_fp16")]; + tensor var_2068 = const()[name = tensor("op_2068"), val = tensor([0, 2, 1, 3])]; + tensor var_2073 = const()[name = tensor("op_2073"), val = tensor([1, 256, 1024])]; + tensor var_2069_cast_fp16 = transpose(perm = var_2068, x = attn_output_9_cast_fp16)[name = tensor("transpose_25")]; + tensor input_53_cast_fp16 = reshape(shape = var_2073, x = var_2069_cast_fp16)[name = tensor("input_53_cast_fp16")]; + tensor attn4_out_proj_weight_to_fp16 = const()[name = tensor("attn4_out_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(108062592)))]; + tensor linear_17_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = attn4_out_proj_weight_to_fp16, x = input_53_cast_fp16)[name = tensor("linear_17_cast_fp16")]; + tensor var_2082_cast_fp16 = add(x = position4_to_fp16, y = valid_len_to_fp16)[name = tensor("op_2082_cast_fp16")]; + tensor var_2082_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_2082_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor input_55_cast_fp16 = add(x = input_49_cast_fp16, y = linear_17_cast_fp16)[name = tensor("input_55_cast_fp16")]; + tensor input_57_axes_0 = const()[name = tensor("input_57_axes_0"), val = tensor([-1])]; + tensor norm4_2_weight_to_fp16 = const()[name = tensor("norm4_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(110159808)))]; + tensor norm4_2_bias_to_fp16 = const()[name = tensor("norm4_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(110161920)))]; + tensor var_2086_to_fp16 = const()[name = tensor("op_2086_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_57_cast_fp16 = layer_norm(axes = input_57_axes_0, beta = norm4_2_bias_to_fp16, epsilon = var_2086_to_fp16, gamma = norm4_2_weight_to_fp16, x = input_55_cast_fp16)[name = tensor("input_57_cast_fp16")]; + tensor linear4_1_weight_to_fp16 = const()[name = tensor("linear4_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(110164032)))]; + tensor linear_18_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = linear4_1_weight_to_fp16, x = input_57_cast_fp16)[name = tensor("linear_18_cast_fp16")]; + tensor input_59_mode_0 = const()[name = tensor("input_59_mode_0"), val = tensor("EXACT")]; + tensor input_59_cast_fp16 = gelu(mode = input_59_mode_0, x = linear_18_cast_fp16)[name = tensor("input_59_cast_fp16")]; + tensor linear4_2_weight_to_fp16 = const()[name = tensor("linear4_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(118552704)))]; + tensor linear_19_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = linear4_2_weight_to_fp16, x = input_59_cast_fp16)[name = tensor("linear_19_cast_fp16")]; + tensor input_61_cast_fp16 = add(x = input_55_cast_fp16, y = linear_19_cast_fp16)[name = tensor("input_61_cast_fp16")]; + tensor input_63_axes_0 = const()[name = tensor("input_63_axes_0"), val = tensor([-1])]; + tensor norm5_1_weight_to_fp16 = const()[name = tensor("norm5_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(126941376)))]; + tensor norm5_1_bias_to_fp16 = const()[name = tensor("norm5_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(126943488)))]; + tensor var_2103_to_fp16 = const()[name = tensor("op_2103_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_63_cast_fp16 = layer_norm(axes = input_63_axes_0, beta = norm5_1_bias_to_fp16, epsilon = var_2103_to_fp16, gamma = norm5_1_weight_to_fp16, x = input_61_cast_fp16)[name = tensor("input_63_cast_fp16")]; + tensor attn5_in_proj_weight_to_fp16 = const()[name = tensor("attn5_in_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(126945600)))]; + tensor linear_20_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = attn5_in_proj_weight_to_fp16, x = input_63_cast_fp16)[name = tensor("linear_20_cast_fp16")]; + tensor var_2117 = const()[name = tensor("op_2117"), val = tensor([1, 256, 3, 16, 64])]; + tensor qkv_cast_fp16 = reshape(shape = var_2117, x = linear_20_cast_fp16)[name = tensor("qkv_cast_fp16")]; + tensor k_21_begin_0 = const()[name = tensor("k_21_begin_0"), val = tensor([0, 0, 1, 0, 0])]; + tensor k_21_end_0 = const()[name = tensor("k_21_end_0"), val = tensor([1, 256, 2, 16, 64])]; + tensor k_21_end_mask_0 = const()[name = tensor("k_21_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor k_21_squeeze_mask_0 = const()[name = tensor("k_21_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor k_21_cast_fp16 = slice_by_index(begin = k_21_begin_0, end = k_21_end_0, end_mask = k_21_end_mask_0, squeeze_mask = k_21_squeeze_mask_0, x = qkv_cast_fp16)[name = tensor("k_21_cast_fp16")]; + tensor v_begin_0 = const()[name = tensor("v_begin_0"), val = tensor([0, 0, 2, 0, 0])]; + tensor v_end_0 = const()[name = tensor("v_end_0"), val = tensor([1, 256, 3, 16, 64])]; + tensor v_end_mask_0 = const()[name = tensor("v_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor v_squeeze_mask_0 = const()[name = tensor("v_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor v_cast_fp16 = slice_by_index(begin = v_begin_0, end = v_end_0, end_mask = v_end_mask_0, squeeze_mask = v_squeeze_mask_0, x = qkv_cast_fp16)[name = tensor("v_cast_fp16")]; + tensor var_2162 = const()[name = tensor("op_2162"), val = tensor([1, 1])]; + tensor position5_to_fp16_dtype_0 = const()[name = tensor("position5_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor position5_to_fp16 = cast(dtype = position5_to_fp16_dtype_0, x = position5)[name = tensor("cast_132")]; + tensor var_2163_cast_fp16 = reshape(shape = var_2162, x = position5_to_fp16)[name = tensor("op_2163_cast_fp16")]; + tensor ts_33_cast_fp16 = add(x = ts_1_promoted_to_fp16, y = var_2163_cast_fp16)[name = tensor("ts_33_cast_fp16")]; + tensor var_2170 = const()[name = tensor("op_2170"), val = tensor([1, 256, 1, 1])]; + tensor ts_cast_fp16 = reshape(shape = var_2170, x = ts_33_cast_fp16)[name = tensor("ts_cast_fp16")]; + tensor var_2177 = const()[name = tensor("op_2177"), val = tensor([1, 256, 16, 32, 2])]; + tensor k_complex_cast_fp16 = reshape(shape = var_2177, x = k_21_cast_fp16)[name = tensor("k_complex_cast_fp16")]; + tensor var_2181_begin_0 = const()[name = tensor("op_2181_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_2181_end_0 = const()[name = tensor("op_2181_end_0"), val = tensor([1, 256, 16, 32, 1])]; + tensor var_2181_end_mask_0 = const()[name = tensor("op_2181_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_2181_squeeze_mask_0 = const()[name = tensor("op_2181_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_2181_cast_fp16 = slice_by_index(begin = var_2181_begin_0, end = var_2181_end_0, end_mask = var_2181_end_mask_0, squeeze_mask = var_2181_squeeze_mask_0, x = k_complex_cast_fp16)[name = tensor("op_2181_cast_fp16")]; + tensor var_2189_begin_0 = const()[name = tensor("op_2189_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_2189_end_0 = const()[name = tensor("op_2189_end_0"), val = tensor([1, 256, 16, 32, 2])]; + tensor var_2189_end_mask_0 = const()[name = tensor("op_2189_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_2189_squeeze_mask_0 = const()[name = tensor("op_2189_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_2189_cast_fp16 = slice_by_index(begin = var_2189_begin_0, end = var_2189_end_0, end_mask = var_2189_end_mask_0, squeeze_mask = var_2189_squeeze_mask_0, x = k_complex_cast_fp16)[name = tensor("op_2189_cast_fp16")]; + tensor freqs_to_fp16 = const()[name = tensor("freqs_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(133237120)))]; + tensor var_2195_cast_fp16 = mul(x = freqs_to_fp16, y = ts_cast_fp16)[name = tensor("op_2195_cast_fp16")]; + tensor rotr_cast_fp16 = cos(x = var_2195_cast_fp16)[name = tensor("rotr_cast_fp16")]; + tensor roti_cast_fp16 = sin(x = var_2195_cast_fp16)[name = tensor("roti_cast_fp16")]; + tensor var_2199_cast_fp16 = mul(x = var_2181_cast_fp16, y = rotr_cast_fp16)[name = tensor("op_2199_cast_fp16")]; + tensor var_2200_cast_fp16 = mul(x = var_2189_cast_fp16, y = roti_cast_fp16)[name = tensor("op_2200_cast_fp16")]; + tensor kor_21_cast_fp16 = sub(x = var_2199_cast_fp16, y = var_2200_cast_fp16)[name = tensor("kor_21_cast_fp16")]; + tensor var_2203_cast_fp16 = mul(x = var_2181_cast_fp16, y = roti_cast_fp16)[name = tensor("op_2203_cast_fp16")]; + tensor var_2204_cast_fp16 = mul(x = var_2189_cast_fp16, y = rotr_cast_fp16)[name = tensor("op_2204_cast_fp16")]; + tensor koi_21_cast_fp16 = add(x = var_2203_cast_fp16, y = var_2204_cast_fp16)[name = tensor("koi_21_cast_fp16")]; + tensor ko_axis_0 = const()[name = tensor("ko_axis_0"), val = tensor(-1)]; + tensor ko_cast_fp16 = stack(axis = ko_axis_0, values = (kor_21_cast_fp16, koi_21_cast_fp16))[name = tensor("ko_cast_fp16")]; + tensor var_2224 = const()[name = tensor("op_2224"), val = tensor([1, 256, 16, 64])]; + tensor k_cast_fp16 = reshape(shape = var_2224, x = ko_cast_fp16)[name = tensor("k_cast_fp16")]; + tensor write_indices_float_31_cast_fp16 = add(x = var_2163_cast_fp16, y = var_249_promoted_to_fp16)[name = tensor("write_indices_float_31_cast_fp16")]; + tensor _inversed_2245_y_0_to_fp16 = const()[name = tensor("_inversed_2245_y_0_to_fp16"), val = tensor(0x1p-9)]; + tensor _inversed_2245_cast_fp16 = mul(x = write_indices_float_31_cast_fp16, y = _inversed_2245_y_0_to_fp16)[name = tensor("_inversed_2245_cast_fp16")]; + tensor var_2246_cast_fp16 = floor(x = _inversed_2245_cast_fp16)[name = tensor("op_2246_cast_fp16")]; + tensor var_2247_to_fp16 = const()[name = tensor("op_2247_to_fp16"), val = tensor(0x1p+9)]; + tensor var_2248_cast_fp16 = mul(x = var_2246_cast_fp16, y = var_2247_to_fp16)[name = tensor("op_2248_cast_fp16")]; + tensor write_indices_float_33_cast_fp16 = sub(x = write_indices_float_31_cast_fp16, y = var_2248_cast_fp16)[name = tensor("write_indices_float_33_cast_fp16")]; + tensor var_2266_value_0_to_fp16 = const()[name = tensor("op_2266_value_0_to_fp16"), val = tensor(0x1.ffp+8)]; + tensor var_2266_cast_fp16 = fill_like(ref_tensor = write_indices_float_33_cast_fp16, value = var_2266_value_0_to_fp16)[name = tensor("op_2266_cast_fp16")]; + tensor write_indices_float_cast_fp16 = select(a = write_indices_float_33_cast_fp16, b = var_2266_cast_fp16, cond = token_valid_1_cast_fp16)[name = tensor("write_indices_float_cast_fp16")]; + tensor var_2272_dtype_0 = const()[name = tensor("op_2272_dtype_0"), val = tensor("int32")]; + tensor var_2277 = const()[name = tensor("op_2277"), val = tensor([1, 256, 1, 1])]; + tensor write_indices_float_cast_fp16_to_int32 = cast(dtype = var_2272_dtype_0, x = write_indices_float_cast_fp16)[name = tensor("cast_131")]; + tensor var_2278 = reshape(shape = var_2277, x = write_indices_float_cast_fp16_to_int32)[name = tensor("op_2278")]; + tensor write_indices_reps_0 = const()[name = tensor("write_indices_reps_0"), val = tensor([1, 1, 16, 64])]; + tensor write_indices = tile(reps = write_indices_reps_0, x = var_2278)[name = tensor("write_indices")]; + tensor var_2288_begin_0 = const()[name = tensor("op_2288_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_2288_end_0 = const()[name = tensor("op_2288_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor var_2288_end_mask_0 = const()[name = tensor("op_2288_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_2288_squeeze_mask_0 = const()[name = tensor("op_2288_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor cache5_to_fp16_dtype_0 = const()[name = tensor("cache5_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor cache5_to_fp16 = cast(dtype = cache5_to_fp16_dtype_0, x = cache5)[name = tensor("cast_130")]; + tensor var_2288_cast_fp16 = slice_by_index(begin = var_2288_begin_0, end = var_2288_end_0, end_mask = var_2288_end_mask_0, squeeze_mask = var_2288_squeeze_mask_0, x = cache5_to_fp16)[name = tensor("op_2288_cast_fp16")]; + tensor var_2290_axis_0 = const()[name = tensor("op_2290_axis_0"), val = tensor(1)]; + tensor var_2290_mode_0 = const()[name = tensor("op_2290_mode_0"), val = tensor("update")]; + tensor var_2290_validate_indices_0 = const()[name = tensor("op_2290_validate_indices_0"), val = tensor(false)]; + tensor var_2290_cast_fp16 = scatter_along_axis(axis = var_2290_axis_0, data = var_2288_cast_fp16, indices = write_indices, mode = var_2290_mode_0, updates = k_cast_fp16, validate_indices = var_2290_validate_indices_0)[name = tensor("op_2290_cast_fp16")]; + tensor concat_42 = const()[name = tensor("concat_42"), val = tensor([0, 0, 0, 0, 0])]; + tensor concat_43 = const()[name = tensor("concat_43"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_internal_tensor_assign_1_stride_0 = const()[name = tensor("new_cache_internal_tensor_assign_1_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_internal_tensor_assign_1_begin_mask_0 = const()[name = tensor("new_cache_internal_tensor_assign_1_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_internal_tensor_assign_1_end_mask_0 = const()[name = tensor("new_cache_internal_tensor_assign_1_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_internal_tensor_assign_1_squeeze_mask_0 = const()[name = tensor("new_cache_internal_tensor_assign_1_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_20 = const()[name = tensor("shape_20"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_10 = const()[name = tensor("reduce_prod_10"), val = tensor(1048576)]; + tensor range_1d_10_start_0 = const()[name = tensor("range_1d_10_start_0"), val = tensor(0)]; + tensor range_1d_10_step_0 = const()[name = tensor("range_1d_10_step_0"), val = tensor(1)]; + tensor range_1d_10 = range_1d(end = reduce_prod_10, start = range_1d_10_start_0, step = range_1d_10_step_0)[name = tensor("range_1d_10")]; + tensor reshape_50 = reshape(shape = shape_20, x = range_1d_10)[name = tensor("reshape_50")]; + tensor slice_by_index_10 = slice_by_index(begin = concat_42, begin_mask = new_cache_internal_tensor_assign_1_begin_mask_0, end = concat_43, end_mask = new_cache_internal_tensor_assign_1_end_mask_0, squeeze_mask = new_cache_internal_tensor_assign_1_squeeze_mask_0, stride = new_cache_internal_tensor_assign_1_stride_0, x = reshape_50)[name = tensor("slice_by_index_10")]; + tensor reshape_51_shape_0 = const()[name = tensor("reshape_51_shape_0"), val = tensor([-1])]; + tensor reshape_51 = reshape(shape = reshape_51_shape_0, x = slice_by_index_10)[name = tensor("reshape_51")]; + tensor reshape_52_shape_0 = const()[name = tensor("reshape_52_shape_0"), val = tensor([-1])]; + tensor reshape_52_cast_fp16 = reshape(shape = reshape_52_shape_0, x = var_2290_cast_fp16)[name = tensor("reshape_52_cast_fp16")]; + tensor reshape_53_shape_0 = const()[name = tensor("reshape_53_shape_0"), val = tensor([-1])]; + tensor reshape_53_cast_fp16 = reshape(shape = reshape_53_shape_0, x = cache5_to_fp16)[name = tensor("reshape_53_cast_fp16")]; + tensor scatter_10_mode_0 = const()[name = tensor("scatter_10_mode_0"), val = tensor("update")]; + tensor scatter_10_axis_0 = const()[name = tensor("scatter_10_axis_0"), val = tensor(0)]; + tensor scatter_10_validate_indices_0 = const()[name = tensor("scatter_10_validate_indices_0"), val = tensor(false)]; + tensor scatter_10_cast_fp16 = scatter(axis = scatter_10_axis_0, data = reshape_53_cast_fp16, indices = reshape_51, mode = scatter_10_mode_0, updates = reshape_52_cast_fp16, validate_indices = scatter_10_validate_indices_0)[name = tensor("scatter_10_cast_fp16")]; + tensor reshape_54_cast_fp16 = reshape(shape = shape_20, x = scatter_10_cast_fp16)[name = tensor("reshape_54_cast_fp16")]; + tensor var_2298_begin_0 = const()[name = tensor("op_2298_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor var_2298_end_0 = const()[name = tensor("op_2298_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor var_2298_end_mask_0 = const()[name = tensor("op_2298_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_2298_squeeze_mask_0 = const()[name = tensor("op_2298_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor var_2298_cast_fp16 = slice_by_index(begin = var_2298_begin_0, end = var_2298_end_0, end_mask = var_2298_end_mask_0, squeeze_mask = var_2298_squeeze_mask_0, x = reshape_54_cast_fp16)[name = tensor("op_2298_cast_fp16")]; + tensor var_2300_axis_0 = const()[name = tensor("op_2300_axis_0"), val = tensor(1)]; + tensor var_2300_mode_0 = const()[name = tensor("op_2300_mode_0"), val = tensor("update")]; + tensor var_2300_validate_indices_0 = const()[name = tensor("op_2300_validate_indices_0"), val = tensor(false)]; + tensor var_2300_cast_fp16 = scatter_along_axis(axis = var_2300_axis_0, data = var_2298_cast_fp16, indices = write_indices, mode = var_2300_mode_0, updates = v_cast_fp16, validate_indices = var_2300_validate_indices_0)[name = tensor("op_2300_cast_fp16")]; + tensor concat_44 = const()[name = tensor("concat_44"), val = tensor([1, 0, 0, 0, 0])]; + tensor concat_45 = const()[name = tensor("concat_45"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_internal_tensor_assign_2_stride_0 = const()[name = tensor("new_cache_internal_tensor_assign_2_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_internal_tensor_assign_2_begin_mask_0 = const()[name = tensor("new_cache_internal_tensor_assign_2_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_internal_tensor_assign_2_end_mask_0 = const()[name = tensor("new_cache_internal_tensor_assign_2_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_internal_tensor_assign_2_squeeze_mask_0 = const()[name = tensor("new_cache_internal_tensor_assign_2_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_21 = const()[name = tensor("shape_21"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_11 = const()[name = tensor("reduce_prod_11"), val = tensor(1048576)]; + tensor range_1d_11_start_0 = const()[name = tensor("range_1d_11_start_0"), val = tensor(0)]; + tensor range_1d_11_step_0 = const()[name = tensor("range_1d_11_step_0"), val = tensor(1)]; + tensor range_1d_11 = range_1d(end = reduce_prod_11, start = range_1d_11_start_0, step = range_1d_11_step_0)[name = tensor("range_1d_11")]; + tensor reshape_55 = reshape(shape = shape_21, x = range_1d_11)[name = tensor("reshape_55")]; + tensor slice_by_index_11 = slice_by_index(begin = concat_44, begin_mask = new_cache_internal_tensor_assign_2_begin_mask_0, end = concat_45, end_mask = new_cache_internal_tensor_assign_2_end_mask_0, squeeze_mask = new_cache_internal_tensor_assign_2_squeeze_mask_0, stride = new_cache_internal_tensor_assign_2_stride_0, x = reshape_55)[name = tensor("slice_by_index_11")]; + tensor reshape_56_shape_0 = const()[name = tensor("reshape_56_shape_0"), val = tensor([-1])]; + tensor reshape_56 = reshape(shape = reshape_56_shape_0, x = slice_by_index_11)[name = tensor("reshape_56")]; + tensor reshape_57_shape_0 = const()[name = tensor("reshape_57_shape_0"), val = tensor([-1])]; + tensor reshape_57_cast_fp16 = reshape(shape = reshape_57_shape_0, x = var_2300_cast_fp16)[name = tensor("reshape_57_cast_fp16")]; + tensor reshape_58_shape_0 = const()[name = tensor("reshape_58_shape_0"), val = tensor([-1])]; + tensor reshape_58_cast_fp16 = reshape(shape = reshape_58_shape_0, x = reshape_54_cast_fp16)[name = tensor("reshape_58_cast_fp16")]; + tensor scatter_11_mode_0 = const()[name = tensor("scatter_11_mode_0"), val = tensor("update")]; + tensor scatter_11_axis_0 = const()[name = tensor("scatter_11_axis_0"), val = tensor(0)]; + tensor scatter_11_validate_indices_0 = const()[name = tensor("scatter_11_validate_indices_0"), val = tensor(false)]; + tensor scatter_11_cast_fp16 = scatter(axis = scatter_11_axis_0, data = reshape_58_cast_fp16, indices = reshape_56, mode = scatter_11_mode_0, updates = reshape_57_cast_fp16, validate_indices = scatter_11_validate_indices_0)[name = tensor("scatter_11_cast_fp16")]; + tensor reshape_59_cast_fp16 = reshape(shape = shape_21, x = scatter_11_cast_fp16)[name = tensor("reshape_59_cast_fp16")]; + tensor reshape_59_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("reshape_59_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor var_2310_cast_fp16 = add(x = position5_to_fp16, y = valid_len_to_fp16)[name = tensor("op_2310_cast_fp16")]; + tensor var_2310_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_2310_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor var_2310 = cast(dtype = var_2310_cast_fp16_to_fp32_dtype_0, x = var_2310_cast_fp16)[name = tensor("cast_128")]; + tensor new_cache_internal_tensor_assign_2 = cast(dtype = reshape_59_cast_fp16_to_fp32_dtype_0, x = reshape_59_cast_fp16)[name = tensor("cast_129")]; + tensor var_2082 = cast(dtype = var_2082_cast_fp16_to_fp32_dtype_0, x = var_2082_cast_fp16)[name = tensor("cast_133")]; + tensor new_cache_9_internal_tensor_assign_2 = cast(dtype = reshape_49_cast_fp16_to_fp32_dtype_0, x = reshape_49_cast_fp16)[name = tensor("cast_134")]; + tensor var_1671 = cast(dtype = var_1671_cast_fp16_to_fp32_dtype_0, x = var_1671_cast_fp16)[name = tensor("cast_138")]; + tensor new_cache_7_internal_tensor_assign_2 = cast(dtype = reshape_39_cast_fp16_to_fp32_dtype_0, x = reshape_39_cast_fp16)[name = tensor("cast_139")]; + tensor var_1260 = cast(dtype = var_1260_cast_fp16_to_fp32_dtype_0, x = var_1260_cast_fp16)[name = tensor("cast_143")]; + tensor new_cache_5_internal_tensor_assign_2 = cast(dtype = reshape_29_cast_fp16_to_fp32_dtype_0, x = reshape_29_cast_fp16)[name = tensor("cast_144")]; + tensor var_849 = cast(dtype = var_849_cast_fp16_to_fp32_dtype_0, x = var_849_cast_fp16)[name = tensor("cast_148")]; + tensor new_cache_3_internal_tensor_assign_2 = cast(dtype = reshape_19_cast_fp16_to_fp32_dtype_0, x = reshape_19_cast_fp16)[name = tensor("cast_149")]; + tensor var_438 = cast(dtype = var_438_cast_fp16_to_fp32_dtype_0, x = var_438_cast_fp16)[name = tensor("cast_153")]; + tensor new_cache_1_internal_tensor_assign_2 = cast(dtype = reshape_9_cast_fp16_to_fp32_dtype_0, x = reshape_9_cast_fp16)[name = tensor("cast_154")]; + } -> (new_cache_1_internal_tensor_assign_2, var_438, new_cache_3_internal_tensor_assign_2, var_849, new_cache_5_internal_tensor_assign_2, var_1260, new_cache_7_internal_tensor_assign_2, var_1671, new_cache_9_internal_tensor_assign_2, var_2082, new_cache_internal_tensor_assign_2, var_2310); +} \ No newline at end of file diff --git a/v2.1/italian/cond_prefill.mlmodelc/weights/weight.bin b/v2.1/italian/cond_prefill.mlmodelc/weights/weight.bin new file mode 100644 index 0000000000000000000000000000000000000000..e013f7ff31c4c9c32b8497cecb69ed09abfc5db1 --- /dev/null +++ b/v2.1/italian/cond_prefill.mlmodelc/weights/weight.bin @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:078b3872ccc9ea75e0fe47b4529d7a26327bf02011af8c4908ad9a4ac2d0bee4 +size 133237248 diff --git a/v2.1/italian/cond_prefill.mlpackage/Data/com.apple.CoreML/model.mlmodel b/v2.1/italian/cond_prefill.mlpackage/Data/com.apple.CoreML/model.mlmodel new file mode 100644 index 0000000000000000000000000000000000000000..7766742c76c0a3baa632db0526343e458e4a036d --- /dev/null +++ b/v2.1/italian/cond_prefill.mlpackage/Data/com.apple.CoreML/model.mlmodel @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:6e96c61c812ddfd18310fba537a6443190e309a17cb5f80a563568dc4765fd44 +size 200480 diff --git a/v2.1/italian/cond_prefill.mlpackage/Data/com.apple.CoreML/weights/weight.bin b/v2.1/italian/cond_prefill.mlpackage/Data/com.apple.CoreML/weights/weight.bin new file mode 100644 index 0000000000000000000000000000000000000000..e013f7ff31c4c9c32b8497cecb69ed09abfc5db1 --- /dev/null +++ b/v2.1/italian/cond_prefill.mlpackage/Data/com.apple.CoreML/weights/weight.bin @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:078b3872ccc9ea75e0fe47b4529d7a26327bf02011af8c4908ad9a4ac2d0bee4 +size 133237248 diff --git a/v2.1/italian/cond_prefill.mlpackage/Manifest.json b/v2.1/italian/cond_prefill.mlpackage/Manifest.json new file mode 100644 index 0000000000000000000000000000000000000000..0d5f9ac74c4d7f9cfa9262b6a4f3667173ff88bc --- /dev/null +++ b/v2.1/italian/cond_prefill.mlpackage/Manifest.json @@ -0,0 +1,18 @@ +{ + "fileFormatVersion": "1.0.0", + "itemInfoEntries": { + "1B8CA28B-4A36-4210-A45A-967ED745DE1F": { + "author": "com.apple.CoreML", + "description": "CoreML Model Weights", + "name": "weights", + "path": "com.apple.CoreML/weights" + }, + "B895902A-A7AA-40C1-8414-8C05C5B24AC4": { + "author": "com.apple.CoreML", + "description": "CoreML Model Specification", + "name": "model.mlmodel", + "path": "com.apple.CoreML/model.mlmodel" + } + }, + "rootModelIdentifier": "B895902A-A7AA-40C1-8414-8C05C5B24AC4" +} diff --git a/v2.1/italian/constants/bos_emb.npy b/v2.1/italian/constants/bos_emb.npy new file mode 100644 index 0000000000000000000000000000000000000000..d7a7213e5170c021260e3a438ca4d213b8b321cd --- /dev/null +++ b/v2.1/italian/constants/bos_emb.npy @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:a352112cce9f8154ac7f6f6a9fc361ba3e7157261c80b23d49a5715c703f4f58 +size 256 diff --git a/v2.1/italian/constants/emb_mean.npy b/v2.1/italian/constants/emb_mean.npy new file mode 100644 index 0000000000000000000000000000000000000000..caab83ff55325b240ef55fa9cd40b5edf04a0032 --- /dev/null +++ b/v2.1/italian/constants/emb_mean.npy @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:80bbe48ccafadccde914579a7b6607cb9639f76f87ed3b29ae329b2ef3b7d2fd +size 256 diff --git a/v2.1/italian/constants/emb_std.npy b/v2.1/italian/constants/emb_std.npy new file mode 100644 index 0000000000000000000000000000000000000000..722ab578df338c4aea23acf36ce4beccbc7d5956 --- /dev/null +++ b/v2.1/italian/constants/emb_std.npy @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:c3c68664acd8a79629d900c7a7155a1f0813b5118ae51a063b22d64d92ca8f71 +size 256 diff --git a/v2.1/italian/constants/mimi_init_state.npz b/v2.1/italian/constants/mimi_init_state.npz new file mode 100644 index 0000000000000000000000000000000000000000..3b52f383a86ef0439f9993506e4529c4c1234828 --- /dev/null +++ b/v2.1/italian/constants/mimi_init_state.npz @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:515588e10e9c14a6b009f1cfda9fcd592542b1640ad856d5fc69e50024702a39 +size 1083108 diff --git a/v2.1/italian/constants/quantizer_weight.npy b/v2.1/italian/constants/quantizer_weight.npy new file mode 100644 index 0000000000000000000000000000000000000000..f3c0009e88c203a64687e9e0e278ef50f3e50f1e --- /dev/null +++ b/v2.1/italian/constants/quantizer_weight.npy @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:c40a96dbde34920b6d468078ccbf796da40ebb94d259a20bf96b86e8635e6372 +size 65664 diff --git a/v2.1/italian/constants/text_embed_table.npy b/v2.1/italian/constants/text_embed_table.npy new file mode 100644 index 0000000000000000000000000000000000000000..9d100003b2b27b0c6a4359b89d7d8397433f760d --- /dev/null +++ b/v2.1/italian/constants/text_embed_table.npy @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:b094b2e9cb0f35b6f407439fd3ada5bed614b5b032f5a127b7c864ba378887a9 +size 16388224 diff --git a/v2.1/italian/constants_bin/alba.safetensors b/v2.1/italian/constants_bin/alba.safetensors new file mode 100644 index 0000000000000000000000000000000000000000..ad2c30c5d4976641204c2a8f8322a85790752bd5 --- /dev/null +++ b/v2.1/italian/constants_bin/alba.safetensors @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:7d2bd2e5df41e5ede229857bcc2b56d41856725465260c330ba847e8a945442c +size 6194424 diff --git a/v2.1/italian/constants_bin/anna.safetensors b/v2.1/italian/constants_bin/anna.safetensors new file mode 100644 index 0000000000000000000000000000000000000000..812e205820fe416deb9adcb2812765fae434dd87 --- /dev/null +++ b/v2.1/italian/constants_bin/anna.safetensors @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:2d1bb064780b6b3c4a77a336c51335fde6dd07b506ccf329d2d0872ad6327c5f +size 7816440 diff --git a/v2.1/italian/constants_bin/azelma.safetensors b/v2.1/italian/constants_bin/azelma.safetensors new file mode 100644 index 0000000000000000000000000000000000000000..51fe54999befa12e7e3cedad27da1cb0116289a5 --- /dev/null +++ b/v2.1/italian/constants_bin/azelma.safetensors @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:bbd40c3dd7d74434e5d274ab87a1e00ea1c06216a3242eec783a7b6c882cf3ba +size 7963896 diff --git a/v2.1/italian/constants_bin/bill_boerst.safetensors b/v2.1/italian/constants_bin/bill_boerst.safetensors new file mode 100644 index 0000000000000000000000000000000000000000..41d243d010c6ba23574712565f7dfa716e80bddb --- /dev/null +++ b/v2.1/italian/constants_bin/bill_boerst.safetensors @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:fb5fc864bf42564a1362df9e37cc169b2570b40a0c7fc11a024b9009a89f5bc2 +size 6735096 diff --git a/v2.1/italian/constants_bin/bos_before_voice.bin b/v2.1/italian/constants_bin/bos_before_voice.bin new file mode 100644 index 0000000000000000000000000000000000000000..bd84cf74ed85af1a12be3c5b15bdede1d3659d85 --- /dev/null +++ b/v2.1/italian/constants_bin/bos_before_voice.bin @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:89ecd4a43d247428b861b16e21097a8bbb9618765b1e48a638201cb59c499b34 +size 4096 diff --git a/v2.1/italian/constants_bin/bos_emb.bin b/v2.1/italian/constants_bin/bos_emb.bin new file mode 100644 index 0000000000000000000000000000000000000000..b173be4995caef235ddac80714308b24be3114f9 --- /dev/null +++ b/v2.1/italian/constants_bin/bos_emb.bin @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:be5ac18f25321f2432aa1671d0648900e7fb313df7fa0b9929a76ca9c0c4c6f0 +size 128 diff --git a/v2.1/italian/constants_bin/caro_davy.safetensors b/v2.1/italian/constants_bin/caro_davy.safetensors new file mode 100644 index 0000000000000000000000000000000000000000..a32787ea85a1896c664c872cd4307d29db7f38a0 --- /dev/null +++ b/v2.1/italian/constants_bin/caro_davy.safetensors @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:f13fb23a7c105ddeb5908483d6261ded35a2d8b053b97a77af6baf7a6ba4b78a +size 5260536 diff --git a/v2.1/italian/constants_bin/charles.safetensors b/v2.1/italian/constants_bin/charles.safetensors new file mode 100644 index 0000000000000000000000000000000000000000..1f57a37139bb6a2e9c55a84af0877b5e230ac48f --- /dev/null +++ b/v2.1/italian/constants_bin/charles.safetensors @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:1e4b6dafd17e0bb4489590f821f7083b598bd0dbe6d93bcabc8b72bb383881d5 +size 6194424 diff --git a/v2.1/italian/constants_bin/cosette.safetensors b/v2.1/italian/constants_bin/cosette.safetensors new file mode 100644 index 0000000000000000000000000000000000000000..f348173f163690a43899a6d68d59e75ce292a8e4 --- /dev/null +++ b/v2.1/italian/constants_bin/cosette.safetensors @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:b8e2734587ea6a67850076ff2c867186ad4cd7a189e4894ecb609c8b79529002 +size 6194424 diff --git a/v2.1/italian/constants_bin/eponine.safetensors b/v2.1/italian/constants_bin/eponine.safetensors new file mode 100644 index 0000000000000000000000000000000000000000..3fc9c74f11f27cd65fb7c17bf00f233c9719e0c7 --- /dev/null +++ b/v2.1/italian/constants_bin/eponine.safetensors @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:31259726243a19758b6ba2d6dfa80bae3db70eedd96a183f5ce05d496eff4cdb +size 6931704 diff --git a/v2.1/italian/constants_bin/estelle.safetensors b/v2.1/italian/constants_bin/estelle.safetensors new file mode 100644 index 0000000000000000000000000000000000000000..3896793c7b78250005a0c992212e2d32ded54b47 --- /dev/null +++ b/v2.1/italian/constants_bin/estelle.safetensors @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:7c768f5234601c670917be9a8db2fa005f88d0dc97080fb415ea19c2154ddada +size 8258808 diff --git a/v2.1/italian/constants_bin/eve.safetensors b/v2.1/italian/constants_bin/eve.safetensors new file mode 100644 index 0000000000000000000000000000000000000000..b1b32bce3f6e148fcb1aa8d357f9eacb14ba1458 --- /dev/null +++ b/v2.1/italian/constants_bin/eve.safetensors @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:6bd49ded51178b6a388bfad57360d4f3b795f89fa7155923efbd968405f01c9f +size 6538488 diff --git a/v2.1/italian/constants_bin/fantine.safetensors b/v2.1/italian/constants_bin/fantine.safetensors new file mode 100644 index 0000000000000000000000000000000000000000..0d8606a1cfaf0fdfee0ad83c7ab84d972bb8bcb4 --- /dev/null +++ b/v2.1/italian/constants_bin/fantine.safetensors @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:88a60059c7bed630cb77de4787b473da747e20140dfb93a4ade353c176bae7f1 +size 6538488 diff --git a/v2.1/italian/constants_bin/george.safetensors b/v2.1/italian/constants_bin/george.safetensors new file mode 100644 index 0000000000000000000000000000000000000000..a4791d53008bb29d09c569e297c8442beffff905 --- /dev/null +++ b/v2.1/italian/constants_bin/george.safetensors @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:c54fd34b252255f61f007fda923d8cb33b3f960b4806335ace0a74c2227a878d +size 6243576 diff --git a/v2.1/italian/constants_bin/giovanni.safetensors b/v2.1/italian/constants_bin/giovanni.safetensors new file mode 100644 index 0000000000000000000000000000000000000000..9eee162a37a27cf76e7018ebe64684d4e8e35650 --- /dev/null +++ b/v2.1/italian/constants_bin/giovanni.safetensors @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:e2b47d9d63ad18cfb2e98591b714a0464456a739b1b4316b6427fcf9d5ea671f +size 4621552 diff --git a/v2.1/italian/constants_bin/jane.safetensors b/v2.1/italian/constants_bin/jane.safetensors new file mode 100644 index 0000000000000000000000000000000000000000..c15bd599cf64bca6397cbb61be41157597e19dc6 --- /dev/null +++ b/v2.1/italian/constants_bin/jane.safetensors @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:56bcf13e73c2143ab431e59be4d4b67ef0cb086ce63626f290a5edcbd1c3fa7f +size 7374072 diff --git a/v2.1/italian/constants_bin/javert.safetensors b/v2.1/italian/constants_bin/javert.safetensors new file mode 100644 index 0000000000000000000000000000000000000000..739d43ee1e30a49713afd4e3243ae4ad1fd9644d --- /dev/null +++ b/v2.1/italian/constants_bin/javert.safetensors @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:a4f0b3ca1da9b276b0304f67f83342147afd28e823f7dd9775b62cc1c588bba5 +size 6194424 diff --git a/v2.1/italian/constants_bin/jean.safetensors b/v2.1/italian/constants_bin/jean.safetensors new file mode 100644 index 0000000000000000000000000000000000000000..9dd412a33ca9e3dc38e1d1244c54f65f674daa3d --- /dev/null +++ b/v2.1/italian/constants_bin/jean.safetensors @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:53aca4c7e490fcbaef2aec8e071b8ede1f978a7fdc7826118488c97a9dfc26e9 +size 6194424 diff --git a/v2.1/italian/constants_bin/juergen.safetensors b/v2.1/italian/constants_bin/juergen.safetensors new file mode 100644 index 0000000000000000000000000000000000000000..43dd034847e253f98e398490abb45e5299660f53 --- /dev/null +++ b/v2.1/italian/constants_bin/juergen.safetensors @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:ca2423a3a252f07ba86603f912bfac2adeca58d3c4e2e0ca4563add8f840334b +size 6243576 diff --git a/v2.1/italian/constants_bin/lola.safetensors b/v2.1/italian/constants_bin/lola.safetensors new file mode 100644 index 0000000000000000000000000000000000000000..21b28e39be3b8e55c85b4b5c49faf202b0b5e0b3 --- /dev/null +++ b/v2.1/italian/constants_bin/lola.safetensors @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:6c881871e4799a4927f97d1e2c75005f8bd4aebbcc772ab3b33cdb61fa376d49 +size 5948664 diff --git a/v2.1/italian/constants_bin/marius.safetensors b/v2.1/italian/constants_bin/marius.safetensors new file mode 100644 index 0000000000000000000000000000000000000000..684afc647e937754c6956ccb8bde7ec2e8b43fd5 --- /dev/null +++ b/v2.1/italian/constants_bin/marius.safetensors @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:ed70813d5f48d0b37ab671051fc2226588aec0922bf589f8a2f99fb4db6f199f +size 6194424 diff --git a/v2.1/italian/constants_bin/mary.safetensors b/v2.1/italian/constants_bin/mary.safetensors new file mode 100644 index 0000000000000000000000000000000000000000..db2599cb6853efbf9e017e0c61c4cc6f01792b6f --- /dev/null +++ b/v2.1/italian/constants_bin/mary.safetensors @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:8cb071cafd9f2490094715ab3dc7c84a5f10b9c0539ab30526a1cce0416696b4 +size 6194424 diff --git a/v2.1/italian/constants_bin/michael.safetensors b/v2.1/italian/constants_bin/michael.safetensors new file mode 100644 index 0000000000000000000000000000000000000000..3e329d967961db26edbe412055c45c0be8ca6b8b --- /dev/null +++ b/v2.1/italian/constants_bin/michael.safetensors @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:3a52c4eb205affd8da5b8d8d2e384328b258f6652f2ad170311bb28f5cdab714 +size 7275768 diff --git a/v2.1/italian/constants_bin/paul.safetensors b/v2.1/italian/constants_bin/paul.safetensors new file mode 100644 index 0000000000000000000000000000000000000000..0631445cb023323e72f669c8bf4812ae5a839370 --- /dev/null +++ b/v2.1/italian/constants_bin/paul.safetensors @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:c0a1913b9bf224d61d15562c3f1ff778e6ecfb77432c998ce785b3c04a19d9a0 +size 6980856 diff --git a/v2.1/italian/constants_bin/peter_yearsley.safetensors b/v2.1/italian/constants_bin/peter_yearsley.safetensors new file mode 100644 index 0000000000000000000000000000000000000000..6c6801a29c7eaf7abf43a842f7fddde397200cc8 --- /dev/null +++ b/v2.1/italian/constants_bin/peter_yearsley.safetensors @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:74cabecedcaa39bafc75d4b8ccf9435c0bd6da0eeb5417d68bc68fb91c406a81 +size 3736816 diff --git a/v2.1/italian/constants_bin/rafael.safetensors b/v2.1/italian/constants_bin/rafael.safetensors new file mode 100644 index 0000000000000000000000000000000000000000..48b833a0d70dfd265a03c32d0d58082e1ed005fd --- /dev/null +++ b/v2.1/italian/constants_bin/rafael.safetensors @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:9a192c24c3edd9eded30fbf73e5054533fb9682ee0c07067fee86d2c5b4c8b6d +size 6194424 diff --git a/v2.1/italian/constants_bin/stuart_bell.safetensors b/v2.1/italian/constants_bin/stuart_bell.safetensors new file mode 100644 index 0000000000000000000000000000000000000000..f50301f713ff0b7bfc7d8f420ab549a58fc7ada2 --- /dev/null +++ b/v2.1/italian/constants_bin/stuart_bell.safetensors @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:c1438f5e99714fb10ccb1eae17bc00191b076ad753fb85d140575c396e5c1e2e +size 5260536 diff --git a/v2.1/italian/constants_bin/text_embed_table.bin b/v2.1/italian/constants_bin/text_embed_table.bin new file mode 100644 index 0000000000000000000000000000000000000000..15bb8077f511e0893cfb1090046a07c5075c18c9 --- /dev/null +++ b/v2.1/italian/constants_bin/text_embed_table.bin @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:60ad799b054f9e34a27871e872e3819294d3c79e604b66e6ab29f83df73b3803 +size 16388096 diff --git a/v2.1/italian/constants_bin/tokenizer.model b/v2.1/italian/constants_bin/tokenizer.model new file mode 100644 index 0000000000000000000000000000000000000000..c2dab86ea3c7a064b38d657237cec3544371fe8b --- /dev/null +++ b/v2.1/italian/constants_bin/tokenizer.model @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:6583b974a11b90e14d8a4c8e9c43f06c3861b9ede6e5023a4c27ab5a3a7d4c39 +size 60078 diff --git a/v2.1/italian/constants_bin/vera.safetensors b/v2.1/italian/constants_bin/vera.safetensors new file mode 100644 index 0000000000000000000000000000000000000000..4c37f84a39cc2f0f341ef659f48c9b2872184cd8 --- /dev/null +++ b/v2.1/italian/constants_bin/vera.safetensors @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:e5e506fda84961dba96295915a9957c4e0c3db00a49f946cd9ec3f7a546a4583 +size 6735096 diff --git a/v2.1/italian/flow_decoder_fused.mlmodelc/analytics/coremldata.bin b/v2.1/italian/flow_decoder_fused.mlmodelc/analytics/coremldata.bin new file mode 100644 index 0000000000000000000000000000000000000000..928744ed7b15a6133bc5bc68244b26c4c8d72454 --- /dev/null +++ b/v2.1/italian/flow_decoder_fused.mlmodelc/analytics/coremldata.bin @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:e2d914e21c97d5b818bead98509f19b47b6d68fa8b2de77dd1762c9454d18ad8 +size 243 diff --git a/v2.1/italian/flow_decoder_fused.mlmodelc/coremldata.bin b/v2.1/italian/flow_decoder_fused.mlmodelc/coremldata.bin new file mode 100644 index 0000000000000000000000000000000000000000..2193e383f54291453d675f31dd28a51c6a3b1f75 --- /dev/null +++ b/v2.1/italian/flow_decoder_fused.mlmodelc/coremldata.bin @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:b77b677b547d6ce2ee01d58dc20e75cd5a5ecf2ce86b0cb82a32bfc1c6aec496 +size 413 diff --git a/v2.1/italian/flow_decoder_fused.mlmodelc/model.mil b/v2.1/italian/flow_decoder_fused.mlmodelc/model.mil new file mode 100644 index 0000000000000000000000000000000000000000..c64a1f34555e7bbf9c8d4502d29f9ef52dd6b0ec --- /dev/null +++ b/v2.1/italian/flow_decoder_fused.mlmodelc/model.mil @@ -0,0 +1,1915 @@ +program(1.0) +[buildInfo = dict, tensor>({{"coremlc-component-MIL", "3520.4.1"}, {"coremlc-version", "3520.5.1"}, {"coremltools-component-torch", "2.12.0"}, {"coremltools-source-dialect", "TorchScript"}, {"coremltools-version", "9.0"}})] +{ + func main(tensor latent_init, tensor transformer_out) { + tensor latent_init_to_fp16_dtype_0 = const()[name = tensor("latent_init_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor flow_net_input_proj_weight_to_fp16 = const()[name = tensor("flow_net_input_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(64)))]; + tensor flow_net_input_proj_bias_to_fp16 = const()[name = tensor("flow_net_input_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(32896)))]; + tensor latent_init_to_fp16 = cast(dtype = latent_init_to_fp16_dtype_0, x = latent_init)[name = tensor("cast_146")]; + tensor linear_0_cast_fp16 = linear(bias = flow_net_input_proj_bias_to_fp16, weight = flow_net_input_proj_weight_to_fp16, x = latent_init_to_fp16)[name = tensor("linear_0_cast_fp16")]; + tensor input_3_to_fp16 = const()[name = tensor("input_3_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(33984)))]; + tensor input_5_cast_fp16 = silu(x = input_3_to_fp16)[name = tensor("input_5_cast_fp16")]; + tensor flow_net_time_embed_0_mlp_2_weight_to_fp16 = const()[name = tensor("flow_net_time_embed_0_mlp_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(35072)))]; + tensor flow_net_time_embed_0_mlp_2_bias_to_fp16 = const()[name = tensor("flow_net_time_embed_0_mlp_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(559424)))]; + tensor linear_2_cast_fp16 = linear(bias = flow_net_time_embed_0_mlp_2_bias_to_fp16, weight = flow_net_time_embed_0_mlp_2_weight_to_fp16, x = input_5_cast_fp16)[name = tensor("linear_2_cast_fp16")]; + tensor reduce_mean_0_axes_0 = const()[name = tensor("reduce_mean_0_axes_0"), val = tensor([-1])]; + tensor reduce_mean_0_keep_dims_0 = const()[name = tensor("reduce_mean_0_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_0_cast_fp16 = reduce_mean(axes = reduce_mean_0_axes_0, keep_dims = reduce_mean_0_keep_dims_0, x = linear_2_cast_fp16)[name = tensor("reduce_mean_0_cast_fp16")]; + tensor sub_0_cast_fp16 = sub(x = linear_2_cast_fp16, y = reduce_mean_0_cast_fp16)[name = tensor("sub_0_cast_fp16")]; + tensor square_0_cast_fp16 = square(x = sub_0_cast_fp16)[name = tensor("square_0_cast_fp16")]; + tensor reduce_mean_1_axes_0 = const()[name = tensor("reduce_mean_1_axes_0"), val = tensor([-1])]; + tensor reduce_mean_1_keep_dims_0 = const()[name = tensor("reduce_mean_1_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_1_cast_fp16 = reduce_mean(axes = reduce_mean_1_axes_0, keep_dims = reduce_mean_1_keep_dims_0, x = square_0_cast_fp16)[name = tensor("reduce_mean_1_cast_fp16")]; + tensor real_div_0_to_fp16 = const()[name = tensor("real_div_0_to_fp16"), val = tensor(0x1.008p+0)]; + tensor mul_0_cast_fp16 = mul(x = reduce_mean_1_cast_fp16, y = real_div_0_to_fp16)[name = tensor("mul_0_cast_fp16")]; + tensor var_70_to_fp16 = const()[name = tensor("op_70_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_1_cast_fp16 = add(x = mul_0_cast_fp16, y = var_70_to_fp16)[name = tensor("var_1_cast_fp16")]; + tensor var_73_epsilon_0 = const()[name = tensor("op_73_epsilon_0"), val = tensor(0x1.197998p-40)]; + tensor var_73_cast_fp16 = rsqrt(epsilon = var_73_epsilon_0, x = var_1_cast_fp16)[name = tensor("op_73_cast_fp16")]; + tensor const_3_to_fp16 = const()[name = tensor("const_3_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(560512)))]; + tensor var_74_cast_fp16 = mul(x = const_3_to_fp16, y = var_73_cast_fp16)[name = tensor("op_74_cast_fp16")]; + tensor var_75_cast_fp16 = mul(x = linear_2_cast_fp16, y = var_74_cast_fp16)[name = tensor("op_75_cast_fp16")]; + tensor input_9_to_fp16 = const()[name = tensor("input_9_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(561600)))]; + tensor input_11_cast_fp16 = silu(x = input_9_to_fp16)[name = tensor("input_11_cast_fp16")]; + tensor flow_net_time_embed_1_mlp_2_weight_to_fp16 = const()[name = tensor("flow_net_time_embed_1_mlp_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(562688)))]; + tensor flow_net_time_embed_1_mlp_2_bias_to_fp16 = const()[name = tensor("flow_net_time_embed_1_mlp_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1087040)))]; + tensor linear_4_cast_fp16 = linear(bias = flow_net_time_embed_1_mlp_2_bias_to_fp16, weight = flow_net_time_embed_1_mlp_2_weight_to_fp16, x = input_11_cast_fp16)[name = tensor("linear_4_cast_fp16")]; + tensor reduce_mean_2_axes_0 = const()[name = tensor("reduce_mean_2_axes_0"), val = tensor([-1])]; + tensor reduce_mean_2_keep_dims_0 = const()[name = tensor("reduce_mean_2_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_2_cast_fp16 = reduce_mean(axes = reduce_mean_2_axes_0, keep_dims = reduce_mean_2_keep_dims_0, x = linear_4_cast_fp16)[name = tensor("reduce_mean_2_cast_fp16")]; + tensor sub_2_cast_fp16 = sub(x = linear_4_cast_fp16, y = reduce_mean_2_cast_fp16)[name = tensor("sub_2_cast_fp16")]; + tensor square_1_cast_fp16 = square(x = sub_2_cast_fp16)[name = tensor("square_1_cast_fp16")]; + tensor reduce_mean_3_axes_0 = const()[name = tensor("reduce_mean_3_axes_0"), val = tensor([-1])]; + tensor reduce_mean_3_keep_dims_0 = const()[name = tensor("reduce_mean_3_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_3_cast_fp16 = reduce_mean(axes = reduce_mean_3_axes_0, keep_dims = reduce_mean_3_keep_dims_0, x = square_1_cast_fp16)[name = tensor("reduce_mean_3_cast_fp16")]; + tensor real_div_1_to_fp16 = const()[name = tensor("real_div_1_to_fp16"), val = tensor(0x1.008p+0)]; + tensor mul_1_cast_fp16 = mul(x = reduce_mean_3_cast_fp16, y = real_div_1_to_fp16)[name = tensor("mul_1_cast_fp16")]; + tensor var_110_to_fp16 = const()[name = tensor("op_110_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_3_cast_fp16 = add(x = mul_1_cast_fp16, y = var_110_to_fp16)[name = tensor("var_3_cast_fp16")]; + tensor var_113_epsilon_0 = const()[name = tensor("op_113_epsilon_0"), val = tensor(0x1.197998p-40)]; + tensor var_113_cast_fp16 = rsqrt(epsilon = var_113_epsilon_0, x = var_3_cast_fp16)[name = tensor("op_113_cast_fp16")]; + tensor const_5_to_fp16 = const()[name = tensor("const_5_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1088128)))]; + tensor var_114_cast_fp16 = mul(x = const_5_to_fp16, y = var_113_cast_fp16)[name = tensor("op_114_cast_fp16")]; + tensor var_115_cast_fp16 = mul(x = linear_4_cast_fp16, y = var_114_cast_fp16)[name = tensor("op_115_cast_fp16")]; + tensor var_127_cast_fp16 = add(x = var_75_cast_fp16, y = var_115_cast_fp16)[name = tensor("op_127_cast_fp16")]; + tensor _inversed_t_combined_1_y_0_to_fp16 = const()[name = tensor("_inversed_t_combined_1_y_0_to_fp16"), val = tensor(0x1p-1)]; + tensor _inversed_t_combined_1_cast_fp16 = mul(x = var_127_cast_fp16, y = _inversed_t_combined_1_y_0_to_fp16)[name = tensor("_inversed_t_combined_1_cast_fp16")]; + tensor transformer_out_to_fp16_dtype_0 = const()[name = tensor("transformer_out_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor flow_net_cond_embed_weight_to_fp16 = const()[name = tensor("flow_net_cond_embed_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1089216)))]; + tensor flow_net_cond_embed_bias_to_fp16 = const()[name = tensor("flow_net_cond_embed_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(2137856)))]; + tensor transformer_out_to_fp16 = cast(dtype = transformer_out_to_fp16_dtype_0, x = transformer_out)[name = tensor("cast_145")]; + tensor linear_5_cast_fp16 = linear(bias = flow_net_cond_embed_bias_to_fp16, weight = flow_net_cond_embed_weight_to_fp16, x = transformer_out_to_fp16)[name = tensor("linear_5_cast_fp16")]; + tensor input_13_cast_fp16 = add(x = _inversed_t_combined_1_cast_fp16, y = linear_5_cast_fp16)[name = tensor("input_13_cast_fp16")]; + tensor input_15_cast_fp16 = silu(x = input_13_cast_fp16)[name = tensor("input_15_cast_fp16")]; + tensor flow_net_res_blocks_0_adaLN_modulation_1_weight_to_fp16 = const()[name = tensor("flow_net_res_blocks_0_adaLN_modulation_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(2138944)))]; + tensor flow_net_res_blocks_0_adaLN_modulation_1_bias_to_fp16 = const()[name = tensor("flow_net_res_blocks_0_adaLN_modulation_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(3711872)))]; + tensor linear_6_cast_fp16 = linear(bias = flow_net_res_blocks_0_adaLN_modulation_1_bias_to_fp16, weight = flow_net_res_blocks_0_adaLN_modulation_1_weight_to_fp16, x = input_15_cast_fp16)[name = tensor("linear_6_cast_fp16")]; + tensor var_142_split_sizes_0 = const()[name = tensor("op_142_split_sizes_0"), val = tensor([512, 512, 512])]; + tensor var_142_axis_0 = const()[name = tensor("op_142_axis_0"), val = tensor(-1)]; + tensor var_142_cast_fp16_0, tensor var_142_cast_fp16_1, tensor var_142_cast_fp16_2 = split(axis = var_142_axis_0, split_sizes = var_142_split_sizes_0, x = linear_6_cast_fp16)[name = tensor("op_142_cast_fp16")]; + tensor mean_1_axes_0 = const()[name = tensor("mean_1_axes_0"), val = tensor([-1])]; + tensor mean_1_keep_dims_0 = const()[name = tensor("mean_1_keep_dims_0"), val = tensor(true)]; + tensor mean_1_cast_fp16 = reduce_mean(axes = mean_1_axes_0, keep_dims = mean_1_keep_dims_0, x = linear_0_cast_fp16)[name = tensor("mean_1_cast_fp16")]; + tensor sub_4_cast_fp16 = sub(x = linear_0_cast_fp16, y = mean_1_cast_fp16)[name = tensor("sub_4_cast_fp16")]; + tensor square_2_cast_fp16 = square(x = sub_4_cast_fp16)[name = tensor("square_2_cast_fp16")]; + tensor reduce_mean_5_axes_0 = const()[name = tensor("reduce_mean_5_axes_0"), val = tensor([-1])]; + tensor reduce_mean_5_keep_dims_0 = const()[name = tensor("reduce_mean_5_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_5_cast_fp16 = reduce_mean(axes = reduce_mean_5_axes_0, keep_dims = reduce_mean_5_keep_dims_0, x = square_2_cast_fp16)[name = tensor("reduce_mean_5_cast_fp16")]; + tensor var_152_to_fp16 = const()[name = tensor("op_152_to_fp16"), val = tensor(0x1.1p-20)]; + tensor var_153_cast_fp16 = add(x = reduce_mean_5_cast_fp16, y = var_152_to_fp16)[name = tensor("op_153_cast_fp16")]; + tensor var_154_cast_fp16 = sqrt(x = var_153_cast_fp16)[name = tensor("op_154_cast_fp16")]; + tensor x_7_cast_fp16 = real_div(x = sub_4_cast_fp16, y = var_154_cast_fp16)[name = tensor("x_7_cast_fp16")]; + tensor flow_net_res_blocks_0_in_ln_weight_to_fp16 = const()[name = tensor("flow_net_res_blocks_0_in_ln_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(3715008)))]; + tensor var_156_cast_fp16 = mul(x = x_7_cast_fp16, y = flow_net_res_blocks_0_in_ln_weight_to_fp16)[name = tensor("op_156_cast_fp16")]; + tensor flow_net_res_blocks_0_in_ln_bias_to_fp16 = const()[name = tensor("flow_net_res_blocks_0_in_ln_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(3716096)))]; + tensor x_9_cast_fp16 = add(x = var_156_cast_fp16, y = flow_net_res_blocks_0_in_ln_bias_to_fp16)[name = tensor("x_9_cast_fp16")]; + tensor var_158_promoted_to_fp16 = const()[name = tensor("op_158_promoted_to_fp16"), val = tensor(0x1p+0)]; + tensor var_159_cast_fp16 = add(x = var_142_cast_fp16_1, y = var_158_promoted_to_fp16)[name = tensor("op_159_cast_fp16")]; + tensor var_160_cast_fp16 = mul(x = x_9_cast_fp16, y = var_159_cast_fp16)[name = tensor("op_160_cast_fp16")]; + tensor input_17_cast_fp16 = add(x = var_160_cast_fp16, y = var_142_cast_fp16_0)[name = tensor("input_17_cast_fp16")]; + tensor flow_net_res_blocks_0_mlp_0_weight_to_fp16 = const()[name = tensor("flow_net_res_blocks_0_mlp_0_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(3717184)))]; + tensor flow_net_res_blocks_0_mlp_0_bias_to_fp16 = const()[name = tensor("flow_net_res_blocks_0_mlp_0_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(4241536)))]; + tensor linear_7_cast_fp16 = linear(bias = flow_net_res_blocks_0_mlp_0_bias_to_fp16, weight = flow_net_res_blocks_0_mlp_0_weight_to_fp16, x = input_17_cast_fp16)[name = tensor("linear_7_cast_fp16")]; + tensor input_21_cast_fp16 = silu(x = linear_7_cast_fp16)[name = tensor("input_21_cast_fp16")]; + tensor flow_net_res_blocks_0_mlp_2_weight_to_fp16 = const()[name = tensor("flow_net_res_blocks_0_mlp_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(4242624)))]; + tensor flow_net_res_blocks_0_mlp_2_bias_to_fp16 = const()[name = tensor("flow_net_res_blocks_0_mlp_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(4766976)))]; + tensor linear_8_cast_fp16 = linear(bias = flow_net_res_blocks_0_mlp_2_bias_to_fp16, weight = flow_net_res_blocks_0_mlp_2_weight_to_fp16, x = input_21_cast_fp16)[name = tensor("linear_8_cast_fp16")]; + tensor var_171_cast_fp16 = mul(x = var_142_cast_fp16_2, y = linear_8_cast_fp16)[name = tensor("op_171_cast_fp16")]; + tensor x_11_cast_fp16 = add(x = linear_0_cast_fp16, y = var_171_cast_fp16)[name = tensor("x_11_cast_fp16")]; + tensor flow_net_res_blocks_1_adaLN_modulation_1_weight_to_fp16 = const()[name = tensor("flow_net_res_blocks_1_adaLN_modulation_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(4768064)))]; + tensor flow_net_res_blocks_1_adaLN_modulation_1_bias_to_fp16 = const()[name = tensor("flow_net_res_blocks_1_adaLN_modulation_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(6340992)))]; + tensor linear_9_cast_fp16 = linear(bias = flow_net_res_blocks_1_adaLN_modulation_1_bias_to_fp16, weight = flow_net_res_blocks_1_adaLN_modulation_1_weight_to_fp16, x = input_15_cast_fp16)[name = tensor("linear_9_cast_fp16")]; + tensor var_181_split_sizes_0 = const()[name = tensor("op_181_split_sizes_0"), val = tensor([512, 512, 512])]; + tensor var_181_axis_0 = const()[name = tensor("op_181_axis_0"), val = tensor(-1)]; + tensor var_181_cast_fp16_0, tensor var_181_cast_fp16_1, tensor var_181_cast_fp16_2 = split(axis = var_181_axis_0, split_sizes = var_181_split_sizes_0, x = linear_9_cast_fp16)[name = tensor("op_181_cast_fp16")]; + tensor mean_3_axes_0 = const()[name = tensor("mean_3_axes_0"), val = tensor([-1])]; + tensor mean_3_keep_dims_0 = const()[name = tensor("mean_3_keep_dims_0"), val = tensor(true)]; + tensor mean_3_cast_fp16 = reduce_mean(axes = mean_3_axes_0, keep_dims = mean_3_keep_dims_0, x = x_11_cast_fp16)[name = tensor("mean_3_cast_fp16")]; + tensor sub_5_cast_fp16 = sub(x = x_11_cast_fp16, y = mean_3_cast_fp16)[name = tensor("sub_5_cast_fp16")]; + tensor square_3_cast_fp16 = square(x = sub_5_cast_fp16)[name = tensor("square_3_cast_fp16")]; + tensor reduce_mean_7_axes_0 = const()[name = tensor("reduce_mean_7_axes_0"), val = tensor([-1])]; + tensor reduce_mean_7_keep_dims_0 = const()[name = tensor("reduce_mean_7_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_7_cast_fp16 = reduce_mean(axes = reduce_mean_7_axes_0, keep_dims = reduce_mean_7_keep_dims_0, x = square_3_cast_fp16)[name = tensor("reduce_mean_7_cast_fp16")]; + tensor var_191_to_fp16 = const()[name = tensor("op_191_to_fp16"), val = tensor(0x1.1p-20)]; + tensor var_192_cast_fp16 = add(x = reduce_mean_7_cast_fp16, y = var_191_to_fp16)[name = tensor("op_192_cast_fp16")]; + tensor var_193_cast_fp16 = sqrt(x = var_192_cast_fp16)[name = tensor("op_193_cast_fp16")]; + tensor x_13_cast_fp16 = real_div(x = sub_5_cast_fp16, y = var_193_cast_fp16)[name = tensor("x_13_cast_fp16")]; + tensor flow_net_res_blocks_1_in_ln_weight_to_fp16 = const()[name = tensor("flow_net_res_blocks_1_in_ln_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(6344128)))]; + tensor var_195_cast_fp16 = mul(x = x_13_cast_fp16, y = flow_net_res_blocks_1_in_ln_weight_to_fp16)[name = tensor("op_195_cast_fp16")]; + tensor flow_net_res_blocks_1_in_ln_bias_to_fp16 = const()[name = tensor("flow_net_res_blocks_1_in_ln_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(6345216)))]; + tensor x_15_cast_fp16 = add(x = var_195_cast_fp16, y = flow_net_res_blocks_1_in_ln_bias_to_fp16)[name = tensor("x_15_cast_fp16")]; + tensor var_197_promoted_to_fp16 = const()[name = tensor("op_197_promoted_to_fp16"), val = tensor(0x1p+0)]; + tensor var_198_cast_fp16 = add(x = var_181_cast_fp16_1, y = var_197_promoted_to_fp16)[name = tensor("op_198_cast_fp16")]; + tensor var_199_cast_fp16 = mul(x = x_15_cast_fp16, y = var_198_cast_fp16)[name = tensor("op_199_cast_fp16")]; + tensor input_25_cast_fp16 = add(x = var_199_cast_fp16, y = var_181_cast_fp16_0)[name = tensor("input_25_cast_fp16")]; + tensor flow_net_res_blocks_1_mlp_0_weight_to_fp16 = const()[name = tensor("flow_net_res_blocks_1_mlp_0_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(6346304)))]; + tensor flow_net_res_blocks_1_mlp_0_bias_to_fp16 = const()[name = tensor("flow_net_res_blocks_1_mlp_0_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(6870656)))]; + tensor linear_10_cast_fp16 = linear(bias = flow_net_res_blocks_1_mlp_0_bias_to_fp16, weight = flow_net_res_blocks_1_mlp_0_weight_to_fp16, x = input_25_cast_fp16)[name = tensor("linear_10_cast_fp16")]; + tensor input_29_cast_fp16 = silu(x = linear_10_cast_fp16)[name = tensor("input_29_cast_fp16")]; + tensor flow_net_res_blocks_1_mlp_2_weight_to_fp16 = const()[name = tensor("flow_net_res_blocks_1_mlp_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(6871744)))]; + tensor flow_net_res_blocks_1_mlp_2_bias_to_fp16 = const()[name = tensor("flow_net_res_blocks_1_mlp_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(7396096)))]; + tensor linear_11_cast_fp16 = linear(bias = flow_net_res_blocks_1_mlp_2_bias_to_fp16, weight = flow_net_res_blocks_1_mlp_2_weight_to_fp16, x = input_29_cast_fp16)[name = tensor("linear_11_cast_fp16")]; + tensor var_210_cast_fp16 = mul(x = var_181_cast_fp16_2, y = linear_11_cast_fp16)[name = tensor("op_210_cast_fp16")]; + tensor x_17_cast_fp16 = add(x = x_11_cast_fp16, y = var_210_cast_fp16)[name = tensor("x_17_cast_fp16")]; + tensor flow_net_res_blocks_2_adaLN_modulation_1_weight_to_fp16 = const()[name = tensor("flow_net_res_blocks_2_adaLN_modulation_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(7397184)))]; + tensor flow_net_res_blocks_2_adaLN_modulation_1_bias_to_fp16 = const()[name = tensor("flow_net_res_blocks_2_adaLN_modulation_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8970112)))]; + tensor linear_12_cast_fp16 = linear(bias = flow_net_res_blocks_2_adaLN_modulation_1_bias_to_fp16, weight = flow_net_res_blocks_2_adaLN_modulation_1_weight_to_fp16, x = input_15_cast_fp16)[name = tensor("linear_12_cast_fp16")]; + tensor var_220_split_sizes_0 = const()[name = tensor("op_220_split_sizes_0"), val = tensor([512, 512, 512])]; + tensor var_220_axis_0 = const()[name = tensor("op_220_axis_0"), val = tensor(-1)]; + tensor var_220_cast_fp16_0, tensor var_220_cast_fp16_1, tensor var_220_cast_fp16_2 = split(axis = var_220_axis_0, split_sizes = var_220_split_sizes_0, x = linear_12_cast_fp16)[name = tensor("op_220_cast_fp16")]; + tensor mean_5_axes_0 = const()[name = tensor("mean_5_axes_0"), val = tensor([-1])]; + tensor mean_5_keep_dims_0 = const()[name = tensor("mean_5_keep_dims_0"), val = tensor(true)]; + tensor mean_5_cast_fp16 = reduce_mean(axes = mean_5_axes_0, keep_dims = mean_5_keep_dims_0, x = x_17_cast_fp16)[name = tensor("mean_5_cast_fp16")]; + tensor sub_6_cast_fp16 = sub(x = x_17_cast_fp16, y = mean_5_cast_fp16)[name = tensor("sub_6_cast_fp16")]; + tensor square_4_cast_fp16 = square(x = sub_6_cast_fp16)[name = tensor("square_4_cast_fp16")]; + tensor reduce_mean_9_axes_0 = const()[name = tensor("reduce_mean_9_axes_0"), val = tensor([-1])]; + tensor reduce_mean_9_keep_dims_0 = const()[name = tensor("reduce_mean_9_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_9_cast_fp16 = reduce_mean(axes = reduce_mean_9_axes_0, keep_dims = reduce_mean_9_keep_dims_0, x = square_4_cast_fp16)[name = tensor("reduce_mean_9_cast_fp16")]; + tensor var_230_to_fp16 = const()[name = tensor("op_230_to_fp16"), val = tensor(0x1.1p-20)]; + tensor var_231_cast_fp16 = add(x = reduce_mean_9_cast_fp16, y = var_230_to_fp16)[name = tensor("op_231_cast_fp16")]; + tensor var_232_cast_fp16 = sqrt(x = var_231_cast_fp16)[name = tensor("op_232_cast_fp16")]; + tensor x_19_cast_fp16 = real_div(x = sub_6_cast_fp16, y = var_232_cast_fp16)[name = tensor("x_19_cast_fp16")]; + tensor flow_net_res_blocks_2_in_ln_weight_to_fp16 = const()[name = tensor("flow_net_res_blocks_2_in_ln_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8973248)))]; + tensor var_234_cast_fp16 = mul(x = x_19_cast_fp16, y = flow_net_res_blocks_2_in_ln_weight_to_fp16)[name = tensor("op_234_cast_fp16")]; + tensor flow_net_res_blocks_2_in_ln_bias_to_fp16 = const()[name = tensor("flow_net_res_blocks_2_in_ln_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8974336)))]; + tensor x_21_cast_fp16 = add(x = var_234_cast_fp16, y = flow_net_res_blocks_2_in_ln_bias_to_fp16)[name = tensor("x_21_cast_fp16")]; + tensor var_236_promoted_to_fp16 = const()[name = tensor("op_236_promoted_to_fp16"), val = tensor(0x1p+0)]; + tensor var_237_cast_fp16 = add(x = var_220_cast_fp16_1, y = var_236_promoted_to_fp16)[name = tensor("op_237_cast_fp16")]; + tensor var_238_cast_fp16 = mul(x = x_21_cast_fp16, y = var_237_cast_fp16)[name = tensor("op_238_cast_fp16")]; + tensor input_33_cast_fp16 = add(x = var_238_cast_fp16, y = var_220_cast_fp16_0)[name = tensor("input_33_cast_fp16")]; + tensor flow_net_res_blocks_2_mlp_0_weight_to_fp16 = const()[name = tensor("flow_net_res_blocks_2_mlp_0_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8975424)))]; + tensor flow_net_res_blocks_2_mlp_0_bias_to_fp16 = const()[name = tensor("flow_net_res_blocks_2_mlp_0_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9499776)))]; + tensor linear_13_cast_fp16 = linear(bias = flow_net_res_blocks_2_mlp_0_bias_to_fp16, weight = flow_net_res_blocks_2_mlp_0_weight_to_fp16, x = input_33_cast_fp16)[name = tensor("linear_13_cast_fp16")]; + tensor input_37_cast_fp16 = silu(x = linear_13_cast_fp16)[name = tensor("input_37_cast_fp16")]; + tensor flow_net_res_blocks_2_mlp_2_weight_to_fp16 = const()[name = tensor("flow_net_res_blocks_2_mlp_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9500864)))]; + tensor flow_net_res_blocks_2_mlp_2_bias_to_fp16 = const()[name = tensor("flow_net_res_blocks_2_mlp_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(10025216)))]; + tensor linear_14_cast_fp16 = linear(bias = flow_net_res_blocks_2_mlp_2_bias_to_fp16, weight = flow_net_res_blocks_2_mlp_2_weight_to_fp16, x = input_37_cast_fp16)[name = tensor("linear_14_cast_fp16")]; + tensor var_249_cast_fp16 = mul(x = var_220_cast_fp16_2, y = linear_14_cast_fp16)[name = tensor("op_249_cast_fp16")]; + tensor x_23_cast_fp16 = add(x = x_17_cast_fp16, y = var_249_cast_fp16)[name = tensor("x_23_cast_fp16")]; + tensor flow_net_res_blocks_3_adaLN_modulation_1_weight_to_fp16 = const()[name = tensor("flow_net_res_blocks_3_adaLN_modulation_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(10026304)))]; + tensor flow_net_res_blocks_3_adaLN_modulation_1_bias_to_fp16 = const()[name = tensor("flow_net_res_blocks_3_adaLN_modulation_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(11599232)))]; + tensor linear_15_cast_fp16 = linear(bias = flow_net_res_blocks_3_adaLN_modulation_1_bias_to_fp16, weight = flow_net_res_blocks_3_adaLN_modulation_1_weight_to_fp16, x = input_15_cast_fp16)[name = tensor("linear_15_cast_fp16")]; + tensor var_259_split_sizes_0 = const()[name = tensor("op_259_split_sizes_0"), val = tensor([512, 512, 512])]; + tensor var_259_axis_0 = const()[name = tensor("op_259_axis_0"), val = tensor(-1)]; + tensor var_259_cast_fp16_0, tensor var_259_cast_fp16_1, tensor var_259_cast_fp16_2 = split(axis = var_259_axis_0, split_sizes = var_259_split_sizes_0, x = linear_15_cast_fp16)[name = tensor("op_259_cast_fp16")]; + tensor mean_7_axes_0 = const()[name = tensor("mean_7_axes_0"), val = tensor([-1])]; + tensor mean_7_keep_dims_0 = const()[name = tensor("mean_7_keep_dims_0"), val = tensor(true)]; + tensor mean_7_cast_fp16 = reduce_mean(axes = mean_7_axes_0, keep_dims = mean_7_keep_dims_0, x = x_23_cast_fp16)[name = tensor("mean_7_cast_fp16")]; + tensor sub_7_cast_fp16 = sub(x = x_23_cast_fp16, y = mean_7_cast_fp16)[name = tensor("sub_7_cast_fp16")]; + tensor square_5_cast_fp16 = square(x = sub_7_cast_fp16)[name = tensor("square_5_cast_fp16")]; + tensor reduce_mean_11_axes_0 = const()[name = tensor("reduce_mean_11_axes_0"), val = tensor([-1])]; + tensor reduce_mean_11_keep_dims_0 = const()[name = tensor("reduce_mean_11_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_11_cast_fp16 = reduce_mean(axes = reduce_mean_11_axes_0, keep_dims = reduce_mean_11_keep_dims_0, x = square_5_cast_fp16)[name = tensor("reduce_mean_11_cast_fp16")]; + tensor var_269_to_fp16 = const()[name = tensor("op_269_to_fp16"), val = tensor(0x1.1p-20)]; + tensor var_270_cast_fp16 = add(x = reduce_mean_11_cast_fp16, y = var_269_to_fp16)[name = tensor("op_270_cast_fp16")]; + tensor var_271_cast_fp16 = sqrt(x = var_270_cast_fp16)[name = tensor("op_271_cast_fp16")]; + tensor x_25_cast_fp16 = real_div(x = sub_7_cast_fp16, y = var_271_cast_fp16)[name = tensor("x_25_cast_fp16")]; + tensor flow_net_res_blocks_3_in_ln_weight_to_fp16 = const()[name = tensor("flow_net_res_blocks_3_in_ln_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(11602368)))]; + tensor var_273_cast_fp16 = mul(x = x_25_cast_fp16, y = flow_net_res_blocks_3_in_ln_weight_to_fp16)[name = tensor("op_273_cast_fp16")]; + tensor flow_net_res_blocks_3_in_ln_bias_to_fp16 = const()[name = tensor("flow_net_res_blocks_3_in_ln_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(11603456)))]; + tensor x_27_cast_fp16 = add(x = var_273_cast_fp16, y = flow_net_res_blocks_3_in_ln_bias_to_fp16)[name = tensor("x_27_cast_fp16")]; + tensor var_275_promoted_to_fp16 = const()[name = tensor("op_275_promoted_to_fp16"), val = tensor(0x1p+0)]; + tensor var_276_cast_fp16 = add(x = var_259_cast_fp16_1, y = var_275_promoted_to_fp16)[name = tensor("op_276_cast_fp16")]; + tensor var_277_cast_fp16 = mul(x = x_27_cast_fp16, y = var_276_cast_fp16)[name = tensor("op_277_cast_fp16")]; + tensor input_41_cast_fp16 = add(x = var_277_cast_fp16, y = var_259_cast_fp16_0)[name = tensor("input_41_cast_fp16")]; + tensor flow_net_res_blocks_3_mlp_0_weight_to_fp16 = const()[name = tensor("flow_net_res_blocks_3_mlp_0_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(11604544)))]; + tensor flow_net_res_blocks_3_mlp_0_bias_to_fp16 = const()[name = tensor("flow_net_res_blocks_3_mlp_0_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(12128896)))]; + tensor linear_16_cast_fp16 = linear(bias = flow_net_res_blocks_3_mlp_0_bias_to_fp16, weight = flow_net_res_blocks_3_mlp_0_weight_to_fp16, x = input_41_cast_fp16)[name = tensor("linear_16_cast_fp16")]; + tensor input_45_cast_fp16 = silu(x = linear_16_cast_fp16)[name = tensor("input_45_cast_fp16")]; + tensor flow_net_res_blocks_3_mlp_2_weight_to_fp16 = const()[name = tensor("flow_net_res_blocks_3_mlp_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(12129984)))]; + tensor flow_net_res_blocks_3_mlp_2_bias_to_fp16 = const()[name = tensor("flow_net_res_blocks_3_mlp_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(12654336)))]; + tensor linear_17_cast_fp16 = linear(bias = flow_net_res_blocks_3_mlp_2_bias_to_fp16, weight = flow_net_res_blocks_3_mlp_2_weight_to_fp16, x = input_45_cast_fp16)[name = tensor("linear_17_cast_fp16")]; + tensor var_288_cast_fp16 = mul(x = var_259_cast_fp16_2, y = linear_17_cast_fp16)[name = tensor("op_288_cast_fp16")]; + tensor x_29_cast_fp16 = add(x = x_23_cast_fp16, y = var_288_cast_fp16)[name = tensor("x_29_cast_fp16")]; + tensor flow_net_res_blocks_4_adaLN_modulation_1_weight_to_fp16 = const()[name = tensor("flow_net_res_blocks_4_adaLN_modulation_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(12655424)))]; + tensor flow_net_res_blocks_4_adaLN_modulation_1_bias_to_fp16 = const()[name = tensor("flow_net_res_blocks_4_adaLN_modulation_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(14228352)))]; + tensor linear_18_cast_fp16 = linear(bias = flow_net_res_blocks_4_adaLN_modulation_1_bias_to_fp16, weight = flow_net_res_blocks_4_adaLN_modulation_1_weight_to_fp16, x = input_15_cast_fp16)[name = tensor("linear_18_cast_fp16")]; + tensor var_298_split_sizes_0 = const()[name = tensor("op_298_split_sizes_0"), val = tensor([512, 512, 512])]; + tensor var_298_axis_0 = const()[name = tensor("op_298_axis_0"), val = tensor(-1)]; + tensor var_298_cast_fp16_0, tensor var_298_cast_fp16_1, tensor var_298_cast_fp16_2 = split(axis = var_298_axis_0, split_sizes = var_298_split_sizes_0, x = linear_18_cast_fp16)[name = tensor("op_298_cast_fp16")]; + tensor mean_9_axes_0 = const()[name = tensor("mean_9_axes_0"), val = tensor([-1])]; + tensor mean_9_keep_dims_0 = const()[name = tensor("mean_9_keep_dims_0"), val = tensor(true)]; + tensor mean_9_cast_fp16 = reduce_mean(axes = mean_9_axes_0, keep_dims = mean_9_keep_dims_0, x = x_29_cast_fp16)[name = tensor("mean_9_cast_fp16")]; + tensor sub_8_cast_fp16 = sub(x = x_29_cast_fp16, y = mean_9_cast_fp16)[name = tensor("sub_8_cast_fp16")]; + tensor square_6_cast_fp16 = square(x = sub_8_cast_fp16)[name = tensor("square_6_cast_fp16")]; + tensor reduce_mean_13_axes_0 = const()[name = tensor("reduce_mean_13_axes_0"), val = tensor([-1])]; + tensor reduce_mean_13_keep_dims_0 = const()[name = tensor("reduce_mean_13_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_13_cast_fp16 = reduce_mean(axes = reduce_mean_13_axes_0, keep_dims = reduce_mean_13_keep_dims_0, x = square_6_cast_fp16)[name = tensor("reduce_mean_13_cast_fp16")]; + tensor var_308_to_fp16 = const()[name = tensor("op_308_to_fp16"), val = tensor(0x1.1p-20)]; + tensor var_309_cast_fp16 = add(x = reduce_mean_13_cast_fp16, y = var_308_to_fp16)[name = tensor("op_309_cast_fp16")]; + tensor var_310_cast_fp16 = sqrt(x = var_309_cast_fp16)[name = tensor("op_310_cast_fp16")]; + tensor x_31_cast_fp16 = real_div(x = sub_8_cast_fp16, y = var_310_cast_fp16)[name = tensor("x_31_cast_fp16")]; + tensor flow_net_res_blocks_4_in_ln_weight_to_fp16 = const()[name = tensor("flow_net_res_blocks_4_in_ln_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(14231488)))]; + tensor var_312_cast_fp16 = mul(x = x_31_cast_fp16, y = flow_net_res_blocks_4_in_ln_weight_to_fp16)[name = tensor("op_312_cast_fp16")]; + tensor flow_net_res_blocks_4_in_ln_bias_to_fp16 = const()[name = tensor("flow_net_res_blocks_4_in_ln_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(14232576)))]; + tensor x_33_cast_fp16 = add(x = var_312_cast_fp16, y = flow_net_res_blocks_4_in_ln_bias_to_fp16)[name = tensor("x_33_cast_fp16")]; + tensor var_314_promoted_to_fp16 = const()[name = tensor("op_314_promoted_to_fp16"), val = tensor(0x1p+0)]; + tensor var_315_cast_fp16 = add(x = var_298_cast_fp16_1, y = var_314_promoted_to_fp16)[name = tensor("op_315_cast_fp16")]; + tensor var_316_cast_fp16 = mul(x = x_33_cast_fp16, y = var_315_cast_fp16)[name = tensor("op_316_cast_fp16")]; + tensor input_49_cast_fp16 = add(x = var_316_cast_fp16, y = var_298_cast_fp16_0)[name = tensor("input_49_cast_fp16")]; + tensor flow_net_res_blocks_4_mlp_0_weight_to_fp16 = const()[name = tensor("flow_net_res_blocks_4_mlp_0_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(14233664)))]; + tensor flow_net_res_blocks_4_mlp_0_bias_to_fp16 = const()[name = tensor("flow_net_res_blocks_4_mlp_0_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(14758016)))]; + tensor linear_19_cast_fp16 = linear(bias = flow_net_res_blocks_4_mlp_0_bias_to_fp16, weight = flow_net_res_blocks_4_mlp_0_weight_to_fp16, x = input_49_cast_fp16)[name = tensor("linear_19_cast_fp16")]; + tensor input_53_cast_fp16 = silu(x = linear_19_cast_fp16)[name = tensor("input_53_cast_fp16")]; + tensor flow_net_res_blocks_4_mlp_2_weight_to_fp16 = const()[name = tensor("flow_net_res_blocks_4_mlp_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(14759104)))]; + tensor flow_net_res_blocks_4_mlp_2_bias_to_fp16 = const()[name = tensor("flow_net_res_blocks_4_mlp_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15283456)))]; + tensor linear_20_cast_fp16 = linear(bias = flow_net_res_blocks_4_mlp_2_bias_to_fp16, weight = flow_net_res_blocks_4_mlp_2_weight_to_fp16, x = input_53_cast_fp16)[name = tensor("linear_20_cast_fp16")]; + tensor var_327_cast_fp16 = mul(x = var_298_cast_fp16_2, y = linear_20_cast_fp16)[name = tensor("op_327_cast_fp16")]; + tensor x_35_cast_fp16 = add(x = x_29_cast_fp16, y = var_327_cast_fp16)[name = tensor("x_35_cast_fp16")]; + tensor flow_net_res_blocks_5_adaLN_modulation_1_weight_to_fp16 = const()[name = tensor("flow_net_res_blocks_5_adaLN_modulation_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15284544)))]; + tensor flow_net_res_blocks_5_adaLN_modulation_1_bias_to_fp16 = const()[name = tensor("flow_net_res_blocks_5_adaLN_modulation_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(16857472)))]; + tensor linear_21_cast_fp16 = linear(bias = flow_net_res_blocks_5_adaLN_modulation_1_bias_to_fp16, weight = flow_net_res_blocks_5_adaLN_modulation_1_weight_to_fp16, x = input_15_cast_fp16)[name = tensor("linear_21_cast_fp16")]; + tensor var_337_split_sizes_0 = const()[name = tensor("op_337_split_sizes_0"), val = tensor([512, 512, 512])]; + tensor var_337_axis_0 = const()[name = tensor("op_337_axis_0"), val = tensor(-1)]; + tensor var_337_cast_fp16_0, tensor var_337_cast_fp16_1, tensor var_337_cast_fp16_2 = split(axis = var_337_axis_0, split_sizes = var_337_split_sizes_0, x = linear_21_cast_fp16)[name = tensor("op_337_cast_fp16")]; + tensor mean_11_axes_0 = const()[name = tensor("mean_11_axes_0"), val = tensor([-1])]; + tensor mean_11_keep_dims_0 = const()[name = tensor("mean_11_keep_dims_0"), val = tensor(true)]; + tensor mean_11_cast_fp16 = reduce_mean(axes = mean_11_axes_0, keep_dims = mean_11_keep_dims_0, x = x_35_cast_fp16)[name = tensor("mean_11_cast_fp16")]; + tensor sub_9_cast_fp16 = sub(x = x_35_cast_fp16, y = mean_11_cast_fp16)[name = tensor("sub_9_cast_fp16")]; + tensor square_7_cast_fp16 = square(x = sub_9_cast_fp16)[name = tensor("square_7_cast_fp16")]; + tensor reduce_mean_15_axes_0 = const()[name = tensor("reduce_mean_15_axes_0"), val = tensor([-1])]; + tensor reduce_mean_15_keep_dims_0 = const()[name = tensor("reduce_mean_15_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_15_cast_fp16 = reduce_mean(axes = reduce_mean_15_axes_0, keep_dims = reduce_mean_15_keep_dims_0, x = square_7_cast_fp16)[name = tensor("reduce_mean_15_cast_fp16")]; + tensor var_347_to_fp16 = const()[name = tensor("op_347_to_fp16"), val = tensor(0x1.1p-20)]; + tensor var_348_cast_fp16 = add(x = reduce_mean_15_cast_fp16, y = var_347_to_fp16)[name = tensor("op_348_cast_fp16")]; + tensor var_349_cast_fp16 = sqrt(x = var_348_cast_fp16)[name = tensor("op_349_cast_fp16")]; + tensor x_37_cast_fp16 = real_div(x = sub_9_cast_fp16, y = var_349_cast_fp16)[name = tensor("x_37_cast_fp16")]; + tensor flow_net_res_blocks_5_in_ln_weight_to_fp16 = const()[name = tensor("flow_net_res_blocks_5_in_ln_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(16860608)))]; + tensor var_351_cast_fp16 = mul(x = x_37_cast_fp16, y = flow_net_res_blocks_5_in_ln_weight_to_fp16)[name = tensor("op_351_cast_fp16")]; + tensor flow_net_res_blocks_5_in_ln_bias_to_fp16 = const()[name = tensor("flow_net_res_blocks_5_in_ln_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(16861696)))]; + tensor x_39_cast_fp16 = add(x = var_351_cast_fp16, y = flow_net_res_blocks_5_in_ln_bias_to_fp16)[name = tensor("x_39_cast_fp16")]; + tensor var_353_promoted_to_fp16 = const()[name = tensor("op_353_promoted_to_fp16"), val = tensor(0x1p+0)]; + tensor var_354_cast_fp16 = add(x = var_337_cast_fp16_1, y = var_353_promoted_to_fp16)[name = tensor("op_354_cast_fp16")]; + tensor var_355_cast_fp16 = mul(x = x_39_cast_fp16, y = var_354_cast_fp16)[name = tensor("op_355_cast_fp16")]; + tensor input_57_cast_fp16 = add(x = var_355_cast_fp16, y = var_337_cast_fp16_0)[name = tensor("input_57_cast_fp16")]; + tensor flow_net_res_blocks_5_mlp_0_weight_to_fp16 = const()[name = tensor("flow_net_res_blocks_5_mlp_0_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(16862784)))]; + tensor flow_net_res_blocks_5_mlp_0_bias_to_fp16 = const()[name = tensor("flow_net_res_blocks_5_mlp_0_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(17387136)))]; + tensor linear_22_cast_fp16 = linear(bias = flow_net_res_blocks_5_mlp_0_bias_to_fp16, weight = flow_net_res_blocks_5_mlp_0_weight_to_fp16, x = input_57_cast_fp16)[name = tensor("linear_22_cast_fp16")]; + tensor input_61_cast_fp16 = silu(x = linear_22_cast_fp16)[name = tensor("input_61_cast_fp16")]; + tensor flow_net_res_blocks_5_mlp_2_weight_to_fp16 = const()[name = tensor("flow_net_res_blocks_5_mlp_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(17388224)))]; + tensor flow_net_res_blocks_5_mlp_2_bias_to_fp16 = const()[name = tensor("flow_net_res_blocks_5_mlp_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(17912576)))]; + tensor linear_23_cast_fp16 = linear(bias = flow_net_res_blocks_5_mlp_2_bias_to_fp16, weight = flow_net_res_blocks_5_mlp_2_weight_to_fp16, x = input_61_cast_fp16)[name = tensor("linear_23_cast_fp16")]; + tensor var_366_cast_fp16 = mul(x = var_337_cast_fp16_2, y = linear_23_cast_fp16)[name = tensor("op_366_cast_fp16")]; + tensor x_41_cast_fp16 = add(x = x_35_cast_fp16, y = var_366_cast_fp16)[name = tensor("x_41_cast_fp16")]; + tensor flow_net_final_layer_adaLN_modulation_1_weight_to_fp16 = const()[name = tensor("flow_net_final_layer_adaLN_modulation_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(17913664)))]; + tensor flow_net_final_layer_adaLN_modulation_1_bias_to_fp16 = const()[name = tensor("flow_net_final_layer_adaLN_modulation_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(18962304)))]; + tensor linear_24_cast_fp16 = linear(bias = flow_net_final_layer_adaLN_modulation_1_bias_to_fp16, weight = flow_net_final_layer_adaLN_modulation_1_weight_to_fp16, x = input_15_cast_fp16)[name = tensor("linear_24_cast_fp16")]; + tensor var_375_split_sizes_0 = const()[name = tensor("op_375_split_sizes_0"), val = tensor([512, 512])]; + tensor var_375_axis_0 = const()[name = tensor("op_375_axis_0"), val = tensor(-1)]; + tensor var_375_cast_fp16_0, tensor var_375_cast_fp16_1 = split(axis = var_375_axis_0, split_sizes = var_375_split_sizes_0, x = linear_24_cast_fp16)[name = tensor("op_375_cast_fp16")]; + tensor mean_13_axes_0 = const()[name = tensor("mean_13_axes_0"), val = tensor([-1])]; + tensor mean_13_keep_dims_0 = const()[name = tensor("mean_13_keep_dims_0"), val = tensor(true)]; + tensor mean_13_cast_fp16 = reduce_mean(axes = mean_13_axes_0, keep_dims = mean_13_keep_dims_0, x = x_41_cast_fp16)[name = tensor("mean_13_cast_fp16")]; + tensor sub_10_cast_fp16 = sub(x = x_41_cast_fp16, y = mean_13_cast_fp16)[name = tensor("sub_10_cast_fp16")]; + tensor square_8_cast_fp16 = square(x = sub_10_cast_fp16)[name = tensor("square_8_cast_fp16")]; + tensor reduce_mean_17_axes_0 = const()[name = tensor("reduce_mean_17_axes_0"), val = tensor([-1])]; + tensor reduce_mean_17_keep_dims_0 = const()[name = tensor("reduce_mean_17_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_17_cast_fp16 = reduce_mean(axes = reduce_mean_17_axes_0, keep_dims = reduce_mean_17_keep_dims_0, x = square_8_cast_fp16)[name = tensor("reduce_mean_17_cast_fp16")]; + tensor var_382_to_fp16 = const()[name = tensor("op_382_to_fp16"), val = tensor(0x1.1p-20)]; + tensor var_383_cast_fp16 = add(x = reduce_mean_17_cast_fp16, y = var_382_to_fp16)[name = tensor("op_383_cast_fp16")]; + tensor var_384_cast_fp16 = sqrt(x = var_383_cast_fp16)[name = tensor("op_384_cast_fp16")]; + tensor x_43_cast_fp16 = real_div(x = sub_10_cast_fp16, y = var_384_cast_fp16)[name = tensor("x_43_cast_fp16")]; + tensor var_386_promoted_to_fp16 = const()[name = tensor("op_386_promoted_to_fp16"), val = tensor(0x1p+0)]; + tensor var_387_cast_fp16 = add(x = var_375_cast_fp16_1, y = var_386_promoted_to_fp16)[name = tensor("op_387_cast_fp16")]; + tensor var_388_cast_fp16 = mul(x = x_43_cast_fp16, y = var_387_cast_fp16)[name = tensor("op_388_cast_fp16")]; + tensor input_65_cast_fp16 = add(x = var_388_cast_fp16, y = var_375_cast_fp16_0)[name = tensor("input_65_cast_fp16")]; + tensor flow_net_final_layer_linear_weight_to_fp16 = const()[name = tensor("flow_net_final_layer_linear_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(18964416)))]; + tensor flow_net_final_layer_linear_bias_to_fp16 = const()[name = tensor("flow_net_final_layer_linear_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(18997248)))]; + tensor linear_25_cast_fp16 = linear(bias = flow_net_final_layer_linear_bias_to_fp16, weight = flow_net_final_layer_linear_weight_to_fp16, x = input_65_cast_fp16)[name = tensor("linear_25_cast_fp16")]; + tensor var_399_to_fp16 = const()[name = tensor("op_399_to_fp16"), val = tensor(0x1p-3)]; + tensor var_400_cast_fp16 = mul(x = linear_25_cast_fp16, y = var_399_to_fp16)[name = tensor("op_400_cast_fp16")]; + tensor input_67_cast_fp16 = add(x = latent_init_to_fp16, y = var_400_cast_fp16)[name = tensor("input_67_cast_fp16")]; + tensor linear_26_cast_fp16 = linear(bias = flow_net_input_proj_bias_to_fp16, weight = flow_net_input_proj_weight_to_fp16, x = input_67_cast_fp16)[name = tensor("linear_26_cast_fp16")]; + tensor input_71_to_fp16 = const()[name = tensor("input_71_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(18997376)))]; + tensor input_73_cast_fp16 = silu(x = input_71_to_fp16)[name = tensor("input_73_cast_fp16")]; + tensor linear_28_cast_fp16 = linear(bias = flow_net_time_embed_0_mlp_2_bias_to_fp16, weight = flow_net_time_embed_0_mlp_2_weight_to_fp16, x = input_73_cast_fp16)[name = tensor("linear_28_cast_fp16")]; + tensor reduce_mean_18_axes_0 = const()[name = tensor("reduce_mean_18_axes_0"), val = tensor([-1])]; + tensor reduce_mean_18_keep_dims_0 = const()[name = tensor("reduce_mean_18_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_18_cast_fp16 = reduce_mean(axes = reduce_mean_18_axes_0, keep_dims = reduce_mean_18_keep_dims_0, x = linear_28_cast_fp16)[name = tensor("reduce_mean_18_cast_fp16")]; + tensor sub_11_cast_fp16 = sub(x = linear_28_cast_fp16, y = reduce_mean_18_cast_fp16)[name = tensor("sub_11_cast_fp16")]; + tensor square_9_cast_fp16 = square(x = sub_11_cast_fp16)[name = tensor("square_9_cast_fp16")]; + tensor reduce_mean_19_axes_0 = const()[name = tensor("reduce_mean_19_axes_0"), val = tensor([-1])]; + tensor reduce_mean_19_keep_dims_0 = const()[name = tensor("reduce_mean_19_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_19_cast_fp16 = reduce_mean(axes = reduce_mean_19_axes_0, keep_dims = reduce_mean_19_keep_dims_0, x = square_9_cast_fp16)[name = tensor("reduce_mean_19_cast_fp16")]; + tensor real_div_2_to_fp16 = const()[name = tensor("real_div_2_to_fp16"), val = tensor(0x1.008p+0)]; + tensor mul_2_cast_fp16 = mul(x = reduce_mean_19_cast_fp16, y = real_div_2_to_fp16)[name = tensor("mul_2_cast_fp16")]; + tensor var_466_to_fp16 = const()[name = tensor("op_466_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_19_cast_fp16 = add(x = mul_2_cast_fp16, y = var_466_to_fp16)[name = tensor("var_19_cast_fp16")]; + tensor var_469_epsilon_0 = const()[name = tensor("op_469_epsilon_0"), val = tensor(0x1.197998p-40)]; + tensor var_469_cast_fp16 = rsqrt(epsilon = var_469_epsilon_0, x = var_19_cast_fp16)[name = tensor("op_469_cast_fp16")]; + tensor var_470_cast_fp16 = mul(x = const_3_to_fp16, y = var_469_cast_fp16)[name = tensor("op_470_cast_fp16")]; + tensor var_471_cast_fp16 = mul(x = linear_28_cast_fp16, y = var_470_cast_fp16)[name = tensor("op_471_cast_fp16")]; + tensor input_77_to_fp16 = const()[name = tensor("input_77_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(18998464)))]; + tensor input_79_cast_fp16 = silu(x = input_77_to_fp16)[name = tensor("input_79_cast_fp16")]; + tensor linear_30_cast_fp16 = linear(bias = flow_net_time_embed_1_mlp_2_bias_to_fp16, weight = flow_net_time_embed_1_mlp_2_weight_to_fp16, x = input_79_cast_fp16)[name = tensor("linear_30_cast_fp16")]; + tensor reduce_mean_20_axes_0 = const()[name = tensor("reduce_mean_20_axes_0"), val = tensor([-1])]; + tensor reduce_mean_20_keep_dims_0 = const()[name = tensor("reduce_mean_20_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_20_cast_fp16 = reduce_mean(axes = reduce_mean_20_axes_0, keep_dims = reduce_mean_20_keep_dims_0, x = linear_30_cast_fp16)[name = tensor("reduce_mean_20_cast_fp16")]; + tensor sub_13_cast_fp16 = sub(x = linear_30_cast_fp16, y = reduce_mean_20_cast_fp16)[name = tensor("sub_13_cast_fp16")]; + tensor square_10_cast_fp16 = square(x = sub_13_cast_fp16)[name = tensor("square_10_cast_fp16")]; + tensor reduce_mean_21_axes_0 = const()[name = tensor("reduce_mean_21_axes_0"), val = tensor([-1])]; + tensor reduce_mean_21_keep_dims_0 = const()[name = tensor("reduce_mean_21_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_21_cast_fp16 = reduce_mean(axes = reduce_mean_21_axes_0, keep_dims = reduce_mean_21_keep_dims_0, x = square_10_cast_fp16)[name = tensor("reduce_mean_21_cast_fp16")]; + tensor real_div_3_to_fp16 = const()[name = tensor("real_div_3_to_fp16"), val = tensor(0x1.008p+0)]; + tensor mul_3_cast_fp16 = mul(x = reduce_mean_21_cast_fp16, y = real_div_3_to_fp16)[name = tensor("mul_3_cast_fp16")]; + tensor var_503_to_fp16 = const()[name = tensor("op_503_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_21_cast_fp16 = add(x = mul_3_cast_fp16, y = var_503_to_fp16)[name = tensor("var_21_cast_fp16")]; + tensor var_506_epsilon_0 = const()[name = tensor("op_506_epsilon_0"), val = tensor(0x1.197998p-40)]; + tensor var_506_cast_fp16 = rsqrt(epsilon = var_506_epsilon_0, x = var_21_cast_fp16)[name = tensor("op_506_cast_fp16")]; + tensor var_507_cast_fp16 = mul(x = const_5_to_fp16, y = var_506_cast_fp16)[name = tensor("op_507_cast_fp16")]; + tensor var_508_cast_fp16 = mul(x = linear_30_cast_fp16, y = var_507_cast_fp16)[name = tensor("op_508_cast_fp16")]; + tensor var_520_cast_fp16 = add(x = var_471_cast_fp16, y = var_508_cast_fp16)[name = tensor("op_520_cast_fp16")]; + tensor _inversed_t_combined_3_y_0_to_fp16 = const()[name = tensor("_inversed_t_combined_3_y_0_to_fp16"), val = tensor(0x1p-1)]; + tensor _inversed_t_combined_3_cast_fp16 = mul(x = var_520_cast_fp16, y = _inversed_t_combined_3_y_0_to_fp16)[name = tensor("_inversed_t_combined_3_cast_fp16")]; + tensor input_81_cast_fp16 = add(x = _inversed_t_combined_3_cast_fp16, y = linear_5_cast_fp16)[name = tensor("input_81_cast_fp16")]; + tensor input_83_cast_fp16 = silu(x = input_81_cast_fp16)[name = tensor("input_83_cast_fp16")]; + tensor linear_32_cast_fp16 = linear(bias = flow_net_res_blocks_0_adaLN_modulation_1_bias_to_fp16, weight = flow_net_res_blocks_0_adaLN_modulation_1_weight_to_fp16, x = input_83_cast_fp16)[name = tensor("linear_32_cast_fp16")]; + tensor var_535_split_sizes_0 = const()[name = tensor("op_535_split_sizes_0"), val = tensor([512, 512, 512])]; + tensor var_535_axis_0 = const()[name = tensor("op_535_axis_0"), val = tensor(-1)]; + tensor var_535_cast_fp16_0, tensor var_535_cast_fp16_1, tensor var_535_cast_fp16_2 = split(axis = var_535_axis_0, split_sizes = var_535_split_sizes_0, x = linear_32_cast_fp16)[name = tensor("op_535_cast_fp16")]; + tensor mean_15_axes_0 = const()[name = tensor("mean_15_axes_0"), val = tensor([-1])]; + tensor mean_15_keep_dims_0 = const()[name = tensor("mean_15_keep_dims_0"), val = tensor(true)]; + tensor mean_15_cast_fp16 = reduce_mean(axes = mean_15_axes_0, keep_dims = mean_15_keep_dims_0, x = linear_26_cast_fp16)[name = tensor("mean_15_cast_fp16")]; + tensor sub_15_cast_fp16 = sub(x = linear_26_cast_fp16, y = mean_15_cast_fp16)[name = tensor("sub_15_cast_fp16")]; + tensor square_11_cast_fp16 = square(x = sub_15_cast_fp16)[name = tensor("square_11_cast_fp16")]; + tensor reduce_mean_23_axes_0 = const()[name = tensor("reduce_mean_23_axes_0"), val = tensor([-1])]; + tensor reduce_mean_23_keep_dims_0 = const()[name = tensor("reduce_mean_23_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_23_cast_fp16 = reduce_mean(axes = reduce_mean_23_axes_0, keep_dims = reduce_mean_23_keep_dims_0, x = square_11_cast_fp16)[name = tensor("reduce_mean_23_cast_fp16")]; + tensor var_545_to_fp16 = const()[name = tensor("op_545_to_fp16"), val = tensor(0x1.1p-20)]; + tensor var_546_cast_fp16 = add(x = reduce_mean_23_cast_fp16, y = var_545_to_fp16)[name = tensor("op_546_cast_fp16")]; + tensor var_547_cast_fp16 = sqrt(x = var_546_cast_fp16)[name = tensor("op_547_cast_fp16")]; + tensor x_51_cast_fp16 = real_div(x = sub_15_cast_fp16, y = var_547_cast_fp16)[name = tensor("x_51_cast_fp16")]; + tensor var_549_cast_fp16 = mul(x = x_51_cast_fp16, y = flow_net_res_blocks_0_in_ln_weight_to_fp16)[name = tensor("op_549_cast_fp16")]; + tensor x_53_cast_fp16 = add(x = var_549_cast_fp16, y = flow_net_res_blocks_0_in_ln_bias_to_fp16)[name = tensor("x_53_cast_fp16")]; + tensor var_551_promoted_to_fp16 = const()[name = tensor("op_551_promoted_to_fp16"), val = tensor(0x1p+0)]; + tensor var_552_cast_fp16 = add(x = var_535_cast_fp16_1, y = var_551_promoted_to_fp16)[name = tensor("op_552_cast_fp16")]; + tensor var_553_cast_fp16 = mul(x = x_53_cast_fp16, y = var_552_cast_fp16)[name = tensor("op_553_cast_fp16")]; + tensor input_85_cast_fp16 = add(x = var_553_cast_fp16, y = var_535_cast_fp16_0)[name = tensor("input_85_cast_fp16")]; + tensor linear_33_cast_fp16 = linear(bias = flow_net_res_blocks_0_mlp_0_bias_to_fp16, weight = flow_net_res_blocks_0_mlp_0_weight_to_fp16, x = input_85_cast_fp16)[name = tensor("linear_33_cast_fp16")]; + tensor input_89_cast_fp16 = silu(x = linear_33_cast_fp16)[name = tensor("input_89_cast_fp16")]; + tensor linear_34_cast_fp16 = linear(bias = flow_net_res_blocks_0_mlp_2_bias_to_fp16, weight = flow_net_res_blocks_0_mlp_2_weight_to_fp16, x = input_89_cast_fp16)[name = tensor("linear_34_cast_fp16")]; + tensor var_564_cast_fp16 = mul(x = var_535_cast_fp16_2, y = linear_34_cast_fp16)[name = tensor("op_564_cast_fp16")]; + tensor x_55_cast_fp16 = add(x = linear_26_cast_fp16, y = var_564_cast_fp16)[name = tensor("x_55_cast_fp16")]; + tensor linear_35_cast_fp16 = linear(bias = flow_net_res_blocks_1_adaLN_modulation_1_bias_to_fp16, weight = flow_net_res_blocks_1_adaLN_modulation_1_weight_to_fp16, x = input_83_cast_fp16)[name = tensor("linear_35_cast_fp16")]; + tensor var_574_split_sizes_0 = const()[name = tensor("op_574_split_sizes_0"), val = tensor([512, 512, 512])]; + tensor var_574_axis_0 = const()[name = tensor("op_574_axis_0"), val = tensor(-1)]; + tensor var_574_cast_fp16_0, tensor var_574_cast_fp16_1, tensor var_574_cast_fp16_2 = split(axis = var_574_axis_0, split_sizes = var_574_split_sizes_0, x = linear_35_cast_fp16)[name = tensor("op_574_cast_fp16")]; + tensor mean_17_axes_0 = const()[name = tensor("mean_17_axes_0"), val = tensor([-1])]; + tensor mean_17_keep_dims_0 = const()[name = tensor("mean_17_keep_dims_0"), val = tensor(true)]; + tensor mean_17_cast_fp16 = reduce_mean(axes = mean_17_axes_0, keep_dims = mean_17_keep_dims_0, x = x_55_cast_fp16)[name = tensor("mean_17_cast_fp16")]; + tensor sub_16_cast_fp16 = sub(x = x_55_cast_fp16, y = mean_17_cast_fp16)[name = tensor("sub_16_cast_fp16")]; + tensor square_12_cast_fp16 = square(x = sub_16_cast_fp16)[name = tensor("square_12_cast_fp16")]; + tensor reduce_mean_25_axes_0 = const()[name = tensor("reduce_mean_25_axes_0"), val = tensor([-1])]; + tensor reduce_mean_25_keep_dims_0 = const()[name = tensor("reduce_mean_25_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_25_cast_fp16 = reduce_mean(axes = reduce_mean_25_axes_0, keep_dims = reduce_mean_25_keep_dims_0, x = square_12_cast_fp16)[name = tensor("reduce_mean_25_cast_fp16")]; + tensor var_584_to_fp16 = const()[name = tensor("op_584_to_fp16"), val = tensor(0x1.1p-20)]; + tensor var_585_cast_fp16 = add(x = reduce_mean_25_cast_fp16, y = var_584_to_fp16)[name = tensor("op_585_cast_fp16")]; + tensor var_586_cast_fp16 = sqrt(x = var_585_cast_fp16)[name = tensor("op_586_cast_fp16")]; + tensor x_57_cast_fp16 = real_div(x = sub_16_cast_fp16, y = var_586_cast_fp16)[name = tensor("x_57_cast_fp16")]; + tensor var_588_cast_fp16 = mul(x = x_57_cast_fp16, y = flow_net_res_blocks_1_in_ln_weight_to_fp16)[name = tensor("op_588_cast_fp16")]; + tensor x_59_cast_fp16 = add(x = var_588_cast_fp16, y = flow_net_res_blocks_1_in_ln_bias_to_fp16)[name = tensor("x_59_cast_fp16")]; + tensor var_590_promoted_to_fp16 = const()[name = tensor("op_590_promoted_to_fp16"), val = tensor(0x1p+0)]; + tensor var_591_cast_fp16 = add(x = var_574_cast_fp16_1, y = var_590_promoted_to_fp16)[name = tensor("op_591_cast_fp16")]; + tensor var_592_cast_fp16 = mul(x = x_59_cast_fp16, y = var_591_cast_fp16)[name = tensor("op_592_cast_fp16")]; + tensor input_93_cast_fp16 = add(x = var_592_cast_fp16, y = var_574_cast_fp16_0)[name = tensor("input_93_cast_fp16")]; + tensor linear_36_cast_fp16 = linear(bias = flow_net_res_blocks_1_mlp_0_bias_to_fp16, weight = flow_net_res_blocks_1_mlp_0_weight_to_fp16, x = input_93_cast_fp16)[name = tensor("linear_36_cast_fp16")]; + tensor input_97_cast_fp16 = silu(x = linear_36_cast_fp16)[name = tensor("input_97_cast_fp16")]; + tensor linear_37_cast_fp16 = linear(bias = flow_net_res_blocks_1_mlp_2_bias_to_fp16, weight = flow_net_res_blocks_1_mlp_2_weight_to_fp16, x = input_97_cast_fp16)[name = tensor("linear_37_cast_fp16")]; + tensor var_603_cast_fp16 = mul(x = var_574_cast_fp16_2, y = linear_37_cast_fp16)[name = tensor("op_603_cast_fp16")]; + tensor x_61_cast_fp16 = add(x = x_55_cast_fp16, y = var_603_cast_fp16)[name = tensor("x_61_cast_fp16")]; + tensor linear_38_cast_fp16 = linear(bias = flow_net_res_blocks_2_adaLN_modulation_1_bias_to_fp16, weight = flow_net_res_blocks_2_adaLN_modulation_1_weight_to_fp16, x = input_83_cast_fp16)[name = tensor("linear_38_cast_fp16")]; + tensor var_613_split_sizes_0 = const()[name = tensor("op_613_split_sizes_0"), val = tensor([512, 512, 512])]; + tensor var_613_axis_0 = const()[name = tensor("op_613_axis_0"), val = tensor(-1)]; + tensor var_613_cast_fp16_0, tensor var_613_cast_fp16_1, tensor var_613_cast_fp16_2 = split(axis = var_613_axis_0, split_sizes = var_613_split_sizes_0, x = linear_38_cast_fp16)[name = tensor("op_613_cast_fp16")]; + tensor mean_19_axes_0 = const()[name = tensor("mean_19_axes_0"), val = tensor([-1])]; + tensor mean_19_keep_dims_0 = const()[name = tensor("mean_19_keep_dims_0"), val = tensor(true)]; + tensor mean_19_cast_fp16 = reduce_mean(axes = mean_19_axes_0, keep_dims = mean_19_keep_dims_0, x = x_61_cast_fp16)[name = tensor("mean_19_cast_fp16")]; + tensor sub_17_cast_fp16 = sub(x = x_61_cast_fp16, y = mean_19_cast_fp16)[name = tensor("sub_17_cast_fp16")]; + tensor square_13_cast_fp16 = square(x = sub_17_cast_fp16)[name = tensor("square_13_cast_fp16")]; + tensor reduce_mean_27_axes_0 = const()[name = tensor("reduce_mean_27_axes_0"), val = tensor([-1])]; + tensor reduce_mean_27_keep_dims_0 = const()[name = tensor("reduce_mean_27_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_27_cast_fp16 = reduce_mean(axes = reduce_mean_27_axes_0, keep_dims = reduce_mean_27_keep_dims_0, x = square_13_cast_fp16)[name = tensor("reduce_mean_27_cast_fp16")]; + tensor var_623_to_fp16 = const()[name = tensor("op_623_to_fp16"), val = tensor(0x1.1p-20)]; + tensor var_624_cast_fp16 = add(x = reduce_mean_27_cast_fp16, y = var_623_to_fp16)[name = tensor("op_624_cast_fp16")]; + tensor var_625_cast_fp16 = sqrt(x = var_624_cast_fp16)[name = tensor("op_625_cast_fp16")]; + tensor x_63_cast_fp16 = real_div(x = sub_17_cast_fp16, y = var_625_cast_fp16)[name = tensor("x_63_cast_fp16")]; + tensor var_627_cast_fp16 = mul(x = x_63_cast_fp16, y = flow_net_res_blocks_2_in_ln_weight_to_fp16)[name = tensor("op_627_cast_fp16")]; + tensor x_65_cast_fp16 = add(x = var_627_cast_fp16, y = flow_net_res_blocks_2_in_ln_bias_to_fp16)[name = tensor("x_65_cast_fp16")]; + tensor var_629_promoted_to_fp16 = const()[name = tensor("op_629_promoted_to_fp16"), val = tensor(0x1p+0)]; + tensor var_630_cast_fp16 = add(x = var_613_cast_fp16_1, y = var_629_promoted_to_fp16)[name = tensor("op_630_cast_fp16")]; + tensor var_631_cast_fp16 = mul(x = x_65_cast_fp16, y = var_630_cast_fp16)[name = tensor("op_631_cast_fp16")]; + tensor input_101_cast_fp16 = add(x = var_631_cast_fp16, y = var_613_cast_fp16_0)[name = tensor("input_101_cast_fp16")]; + tensor linear_39_cast_fp16 = linear(bias = flow_net_res_blocks_2_mlp_0_bias_to_fp16, weight = flow_net_res_blocks_2_mlp_0_weight_to_fp16, x = input_101_cast_fp16)[name = tensor("linear_39_cast_fp16")]; + tensor input_105_cast_fp16 = silu(x = linear_39_cast_fp16)[name = tensor("input_105_cast_fp16")]; + tensor linear_40_cast_fp16 = linear(bias = flow_net_res_blocks_2_mlp_2_bias_to_fp16, weight = flow_net_res_blocks_2_mlp_2_weight_to_fp16, x = input_105_cast_fp16)[name = tensor("linear_40_cast_fp16")]; + tensor var_642_cast_fp16 = mul(x = var_613_cast_fp16_2, y = linear_40_cast_fp16)[name = tensor("op_642_cast_fp16")]; + tensor x_67_cast_fp16 = add(x = x_61_cast_fp16, y = var_642_cast_fp16)[name = tensor("x_67_cast_fp16")]; + tensor linear_41_cast_fp16 = linear(bias = flow_net_res_blocks_3_adaLN_modulation_1_bias_to_fp16, weight = flow_net_res_blocks_3_adaLN_modulation_1_weight_to_fp16, x = input_83_cast_fp16)[name = tensor("linear_41_cast_fp16")]; + tensor var_652_split_sizes_0 = const()[name = tensor("op_652_split_sizes_0"), val = tensor([512, 512, 512])]; + tensor var_652_axis_0 = const()[name = tensor("op_652_axis_0"), val = tensor(-1)]; + tensor var_652_cast_fp16_0, tensor var_652_cast_fp16_1, tensor var_652_cast_fp16_2 = split(axis = var_652_axis_0, split_sizes = var_652_split_sizes_0, x = linear_41_cast_fp16)[name = tensor("op_652_cast_fp16")]; + tensor mean_21_axes_0 = const()[name = tensor("mean_21_axes_0"), val = tensor([-1])]; + tensor mean_21_keep_dims_0 = const()[name = tensor("mean_21_keep_dims_0"), val = tensor(true)]; + tensor mean_21_cast_fp16 = reduce_mean(axes = mean_21_axes_0, keep_dims = mean_21_keep_dims_0, x = x_67_cast_fp16)[name = tensor("mean_21_cast_fp16")]; + tensor sub_18_cast_fp16 = sub(x = x_67_cast_fp16, y = mean_21_cast_fp16)[name = tensor("sub_18_cast_fp16")]; + tensor square_14_cast_fp16 = square(x = sub_18_cast_fp16)[name = tensor("square_14_cast_fp16")]; + tensor reduce_mean_29_axes_0 = const()[name = tensor("reduce_mean_29_axes_0"), val = tensor([-1])]; + tensor reduce_mean_29_keep_dims_0 = const()[name = tensor("reduce_mean_29_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_29_cast_fp16 = reduce_mean(axes = reduce_mean_29_axes_0, keep_dims = reduce_mean_29_keep_dims_0, x = square_14_cast_fp16)[name = tensor("reduce_mean_29_cast_fp16")]; + tensor var_662_to_fp16 = const()[name = tensor("op_662_to_fp16"), val = tensor(0x1.1p-20)]; + tensor var_663_cast_fp16 = add(x = reduce_mean_29_cast_fp16, y = var_662_to_fp16)[name = tensor("op_663_cast_fp16")]; + tensor var_664_cast_fp16 = sqrt(x = var_663_cast_fp16)[name = tensor("op_664_cast_fp16")]; + tensor x_69_cast_fp16 = real_div(x = sub_18_cast_fp16, y = var_664_cast_fp16)[name = tensor("x_69_cast_fp16")]; + tensor var_666_cast_fp16 = mul(x = x_69_cast_fp16, y = flow_net_res_blocks_3_in_ln_weight_to_fp16)[name = tensor("op_666_cast_fp16")]; + tensor x_71_cast_fp16 = add(x = var_666_cast_fp16, y = flow_net_res_blocks_3_in_ln_bias_to_fp16)[name = tensor("x_71_cast_fp16")]; + tensor var_668_promoted_to_fp16 = const()[name = tensor("op_668_promoted_to_fp16"), val = tensor(0x1p+0)]; + tensor var_669_cast_fp16 = add(x = var_652_cast_fp16_1, y = var_668_promoted_to_fp16)[name = tensor("op_669_cast_fp16")]; + tensor var_670_cast_fp16 = mul(x = x_71_cast_fp16, y = var_669_cast_fp16)[name = tensor("op_670_cast_fp16")]; + tensor input_109_cast_fp16 = add(x = var_670_cast_fp16, y = var_652_cast_fp16_0)[name = tensor("input_109_cast_fp16")]; + tensor linear_42_cast_fp16 = linear(bias = flow_net_res_blocks_3_mlp_0_bias_to_fp16, weight = flow_net_res_blocks_3_mlp_0_weight_to_fp16, x = input_109_cast_fp16)[name = tensor("linear_42_cast_fp16")]; + tensor input_113_cast_fp16 = silu(x = linear_42_cast_fp16)[name = tensor("input_113_cast_fp16")]; + tensor linear_43_cast_fp16 = linear(bias = flow_net_res_blocks_3_mlp_2_bias_to_fp16, weight = flow_net_res_blocks_3_mlp_2_weight_to_fp16, x = input_113_cast_fp16)[name = tensor("linear_43_cast_fp16")]; + tensor var_681_cast_fp16 = mul(x = var_652_cast_fp16_2, y = linear_43_cast_fp16)[name = tensor("op_681_cast_fp16")]; + tensor x_73_cast_fp16 = add(x = x_67_cast_fp16, y = var_681_cast_fp16)[name = tensor("x_73_cast_fp16")]; + tensor linear_44_cast_fp16 = linear(bias = flow_net_res_blocks_4_adaLN_modulation_1_bias_to_fp16, weight = flow_net_res_blocks_4_adaLN_modulation_1_weight_to_fp16, x = input_83_cast_fp16)[name = tensor("linear_44_cast_fp16")]; + tensor var_691_split_sizes_0 = const()[name = tensor("op_691_split_sizes_0"), val = tensor([512, 512, 512])]; + tensor var_691_axis_0 = const()[name = tensor("op_691_axis_0"), val = tensor(-1)]; + tensor var_691_cast_fp16_0, tensor var_691_cast_fp16_1, tensor var_691_cast_fp16_2 = split(axis = var_691_axis_0, split_sizes = var_691_split_sizes_0, x = linear_44_cast_fp16)[name = tensor("op_691_cast_fp16")]; + tensor mean_23_axes_0 = const()[name = tensor("mean_23_axes_0"), val = tensor([-1])]; + tensor mean_23_keep_dims_0 = const()[name = tensor("mean_23_keep_dims_0"), val = tensor(true)]; + tensor mean_23_cast_fp16 = reduce_mean(axes = mean_23_axes_0, keep_dims = mean_23_keep_dims_0, x = x_73_cast_fp16)[name = tensor("mean_23_cast_fp16")]; + tensor sub_19_cast_fp16 = sub(x = x_73_cast_fp16, y = mean_23_cast_fp16)[name = tensor("sub_19_cast_fp16")]; + tensor square_15_cast_fp16 = square(x = sub_19_cast_fp16)[name = tensor("square_15_cast_fp16")]; + tensor reduce_mean_31_axes_0 = const()[name = tensor("reduce_mean_31_axes_0"), val = tensor([-1])]; + tensor reduce_mean_31_keep_dims_0 = const()[name = tensor("reduce_mean_31_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_31_cast_fp16 = reduce_mean(axes = reduce_mean_31_axes_0, keep_dims = reduce_mean_31_keep_dims_0, x = square_15_cast_fp16)[name = tensor("reduce_mean_31_cast_fp16")]; + tensor var_701_to_fp16 = const()[name = tensor("op_701_to_fp16"), val = tensor(0x1.1p-20)]; + tensor var_702_cast_fp16 = add(x = reduce_mean_31_cast_fp16, y = var_701_to_fp16)[name = tensor("op_702_cast_fp16")]; + tensor var_703_cast_fp16 = sqrt(x = var_702_cast_fp16)[name = tensor("op_703_cast_fp16")]; + tensor x_75_cast_fp16 = real_div(x = sub_19_cast_fp16, y = var_703_cast_fp16)[name = tensor("x_75_cast_fp16")]; + tensor var_705_cast_fp16 = mul(x = x_75_cast_fp16, y = flow_net_res_blocks_4_in_ln_weight_to_fp16)[name = tensor("op_705_cast_fp16")]; + tensor x_77_cast_fp16 = add(x = var_705_cast_fp16, y = flow_net_res_blocks_4_in_ln_bias_to_fp16)[name = tensor("x_77_cast_fp16")]; + tensor var_707_promoted_to_fp16 = const()[name = tensor("op_707_promoted_to_fp16"), val = tensor(0x1p+0)]; + tensor var_708_cast_fp16 = add(x = var_691_cast_fp16_1, y = var_707_promoted_to_fp16)[name = tensor("op_708_cast_fp16")]; + tensor var_709_cast_fp16 = mul(x = x_77_cast_fp16, y = var_708_cast_fp16)[name = tensor("op_709_cast_fp16")]; + tensor input_117_cast_fp16 = add(x = var_709_cast_fp16, y = var_691_cast_fp16_0)[name = tensor("input_117_cast_fp16")]; + tensor linear_45_cast_fp16 = linear(bias = flow_net_res_blocks_4_mlp_0_bias_to_fp16, weight = flow_net_res_blocks_4_mlp_0_weight_to_fp16, x = input_117_cast_fp16)[name = tensor("linear_45_cast_fp16")]; + tensor input_121_cast_fp16 = silu(x = linear_45_cast_fp16)[name = tensor("input_121_cast_fp16")]; + tensor linear_46_cast_fp16 = linear(bias = flow_net_res_blocks_4_mlp_2_bias_to_fp16, weight = flow_net_res_blocks_4_mlp_2_weight_to_fp16, x = input_121_cast_fp16)[name = tensor("linear_46_cast_fp16")]; + tensor var_720_cast_fp16 = mul(x = var_691_cast_fp16_2, y = linear_46_cast_fp16)[name = tensor("op_720_cast_fp16")]; + tensor x_79_cast_fp16 = add(x = x_73_cast_fp16, y = var_720_cast_fp16)[name = tensor("x_79_cast_fp16")]; + tensor linear_47_cast_fp16 = linear(bias = flow_net_res_blocks_5_adaLN_modulation_1_bias_to_fp16, weight = flow_net_res_blocks_5_adaLN_modulation_1_weight_to_fp16, x = input_83_cast_fp16)[name = tensor("linear_47_cast_fp16")]; + tensor var_730_split_sizes_0 = const()[name = tensor("op_730_split_sizes_0"), val = tensor([512, 512, 512])]; + tensor var_730_axis_0 = const()[name = tensor("op_730_axis_0"), val = tensor(-1)]; + tensor var_730_cast_fp16_0, tensor var_730_cast_fp16_1, tensor var_730_cast_fp16_2 = split(axis = var_730_axis_0, split_sizes = var_730_split_sizes_0, x = linear_47_cast_fp16)[name = tensor("op_730_cast_fp16")]; + tensor mean_25_axes_0 = const()[name = tensor("mean_25_axes_0"), val = tensor([-1])]; + tensor mean_25_keep_dims_0 = const()[name = tensor("mean_25_keep_dims_0"), val = tensor(true)]; + tensor mean_25_cast_fp16 = reduce_mean(axes = mean_25_axes_0, keep_dims = mean_25_keep_dims_0, x = x_79_cast_fp16)[name = tensor("mean_25_cast_fp16")]; + tensor sub_20_cast_fp16 = sub(x = x_79_cast_fp16, y = mean_25_cast_fp16)[name = tensor("sub_20_cast_fp16")]; + tensor square_16_cast_fp16 = square(x = sub_20_cast_fp16)[name = tensor("square_16_cast_fp16")]; + tensor reduce_mean_33_axes_0 = const()[name = tensor("reduce_mean_33_axes_0"), val = tensor([-1])]; + tensor reduce_mean_33_keep_dims_0 = const()[name = tensor("reduce_mean_33_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_33_cast_fp16 = reduce_mean(axes = reduce_mean_33_axes_0, keep_dims = reduce_mean_33_keep_dims_0, x = square_16_cast_fp16)[name = tensor("reduce_mean_33_cast_fp16")]; + tensor var_740_to_fp16 = const()[name = tensor("op_740_to_fp16"), val = tensor(0x1.1p-20)]; + tensor var_741_cast_fp16 = add(x = reduce_mean_33_cast_fp16, y = var_740_to_fp16)[name = tensor("op_741_cast_fp16")]; + tensor var_742_cast_fp16 = sqrt(x = var_741_cast_fp16)[name = tensor("op_742_cast_fp16")]; + tensor x_81_cast_fp16 = real_div(x = sub_20_cast_fp16, y = var_742_cast_fp16)[name = tensor("x_81_cast_fp16")]; + tensor var_744_cast_fp16 = mul(x = x_81_cast_fp16, y = flow_net_res_blocks_5_in_ln_weight_to_fp16)[name = tensor("op_744_cast_fp16")]; + tensor x_83_cast_fp16 = add(x = var_744_cast_fp16, y = flow_net_res_blocks_5_in_ln_bias_to_fp16)[name = tensor("x_83_cast_fp16")]; + tensor var_746_promoted_to_fp16 = const()[name = tensor("op_746_promoted_to_fp16"), val = tensor(0x1p+0)]; + tensor var_747_cast_fp16 = add(x = var_730_cast_fp16_1, y = var_746_promoted_to_fp16)[name = tensor("op_747_cast_fp16")]; + tensor var_748_cast_fp16 = mul(x = x_83_cast_fp16, y = var_747_cast_fp16)[name = tensor("op_748_cast_fp16")]; + tensor input_125_cast_fp16 = add(x = var_748_cast_fp16, y = var_730_cast_fp16_0)[name = tensor("input_125_cast_fp16")]; + tensor linear_48_cast_fp16 = linear(bias = flow_net_res_blocks_5_mlp_0_bias_to_fp16, weight = flow_net_res_blocks_5_mlp_0_weight_to_fp16, x = input_125_cast_fp16)[name = tensor("linear_48_cast_fp16")]; + tensor input_129_cast_fp16 = silu(x = linear_48_cast_fp16)[name = tensor("input_129_cast_fp16")]; + tensor linear_49_cast_fp16 = linear(bias = flow_net_res_blocks_5_mlp_2_bias_to_fp16, weight = flow_net_res_blocks_5_mlp_2_weight_to_fp16, x = input_129_cast_fp16)[name = tensor("linear_49_cast_fp16")]; + tensor var_759_cast_fp16 = mul(x = var_730_cast_fp16_2, y = linear_49_cast_fp16)[name = tensor("op_759_cast_fp16")]; + tensor x_85_cast_fp16 = add(x = x_79_cast_fp16, y = var_759_cast_fp16)[name = tensor("x_85_cast_fp16")]; + tensor linear_50_cast_fp16 = linear(bias = flow_net_final_layer_adaLN_modulation_1_bias_to_fp16, weight = flow_net_final_layer_adaLN_modulation_1_weight_to_fp16, x = input_83_cast_fp16)[name = tensor("linear_50_cast_fp16")]; + tensor var_768_split_sizes_0 = const()[name = tensor("op_768_split_sizes_0"), val = tensor([512, 512])]; + tensor var_768_axis_0 = const()[name = tensor("op_768_axis_0"), val = tensor(-1)]; + tensor var_768_cast_fp16_0, tensor var_768_cast_fp16_1 = split(axis = var_768_axis_0, split_sizes = var_768_split_sizes_0, x = linear_50_cast_fp16)[name = tensor("op_768_cast_fp16")]; + tensor mean_27_axes_0 = const()[name = tensor("mean_27_axes_0"), val = tensor([-1])]; + tensor mean_27_keep_dims_0 = const()[name = tensor("mean_27_keep_dims_0"), val = tensor(true)]; + tensor mean_27_cast_fp16 = reduce_mean(axes = mean_27_axes_0, keep_dims = mean_27_keep_dims_0, x = x_85_cast_fp16)[name = tensor("mean_27_cast_fp16")]; + tensor sub_21_cast_fp16 = sub(x = x_85_cast_fp16, y = mean_27_cast_fp16)[name = tensor("sub_21_cast_fp16")]; + tensor square_17_cast_fp16 = square(x = sub_21_cast_fp16)[name = tensor("square_17_cast_fp16")]; + tensor reduce_mean_35_axes_0 = const()[name = tensor("reduce_mean_35_axes_0"), val = tensor([-1])]; + tensor reduce_mean_35_keep_dims_0 = const()[name = tensor("reduce_mean_35_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_35_cast_fp16 = reduce_mean(axes = reduce_mean_35_axes_0, keep_dims = reduce_mean_35_keep_dims_0, x = square_17_cast_fp16)[name = tensor("reduce_mean_35_cast_fp16")]; + tensor var_775_to_fp16 = const()[name = tensor("op_775_to_fp16"), val = tensor(0x1.1p-20)]; + tensor var_776_cast_fp16 = add(x = reduce_mean_35_cast_fp16, y = var_775_to_fp16)[name = tensor("op_776_cast_fp16")]; + tensor var_777_cast_fp16 = sqrt(x = var_776_cast_fp16)[name = tensor("op_777_cast_fp16")]; + tensor x_87_cast_fp16 = real_div(x = sub_21_cast_fp16, y = var_777_cast_fp16)[name = tensor("x_87_cast_fp16")]; + tensor var_779_promoted_to_fp16 = const()[name = tensor("op_779_promoted_to_fp16"), val = tensor(0x1p+0)]; + tensor var_780_cast_fp16 = add(x = var_768_cast_fp16_1, y = var_779_promoted_to_fp16)[name = tensor("op_780_cast_fp16")]; + tensor var_781_cast_fp16 = mul(x = x_87_cast_fp16, y = var_780_cast_fp16)[name = tensor("op_781_cast_fp16")]; + tensor input_133_cast_fp16 = add(x = var_781_cast_fp16, y = var_768_cast_fp16_0)[name = tensor("input_133_cast_fp16")]; + tensor linear_51_cast_fp16 = linear(bias = flow_net_final_layer_linear_bias_to_fp16, weight = flow_net_final_layer_linear_weight_to_fp16, x = input_133_cast_fp16)[name = tensor("linear_51_cast_fp16")]; + tensor var_792_to_fp16 = const()[name = tensor("op_792_to_fp16"), val = tensor(0x1p-3)]; + tensor var_793_cast_fp16 = mul(x = linear_51_cast_fp16, y = var_792_to_fp16)[name = tensor("op_793_cast_fp16")]; + tensor input_135_cast_fp16 = add(x = input_67_cast_fp16, y = var_793_cast_fp16)[name = tensor("input_135_cast_fp16")]; + tensor linear_52_cast_fp16 = linear(bias = flow_net_input_proj_bias_to_fp16, weight = flow_net_input_proj_weight_to_fp16, x = input_135_cast_fp16)[name = tensor("linear_52_cast_fp16")]; + tensor input_139_to_fp16 = const()[name = tensor("input_139_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(18999552)))]; + tensor input_141_cast_fp16 = silu(x = input_139_to_fp16)[name = tensor("input_141_cast_fp16")]; + tensor linear_54_cast_fp16 = linear(bias = flow_net_time_embed_0_mlp_2_bias_to_fp16, weight = flow_net_time_embed_0_mlp_2_weight_to_fp16, x = input_141_cast_fp16)[name = tensor("linear_54_cast_fp16")]; + tensor reduce_mean_36_axes_0 = const()[name = tensor("reduce_mean_36_axes_0"), val = tensor([-1])]; + tensor reduce_mean_36_keep_dims_0 = const()[name = tensor("reduce_mean_36_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_36_cast_fp16 = reduce_mean(axes = reduce_mean_36_axes_0, keep_dims = reduce_mean_36_keep_dims_0, x = linear_54_cast_fp16)[name = tensor("reduce_mean_36_cast_fp16")]; + tensor sub_22_cast_fp16 = sub(x = linear_54_cast_fp16, y = reduce_mean_36_cast_fp16)[name = tensor("sub_22_cast_fp16")]; + tensor square_18_cast_fp16 = square(x = sub_22_cast_fp16)[name = tensor("square_18_cast_fp16")]; + tensor reduce_mean_37_axes_0 = const()[name = tensor("reduce_mean_37_axes_0"), val = tensor([-1])]; + tensor reduce_mean_37_keep_dims_0 = const()[name = tensor("reduce_mean_37_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_37_cast_fp16 = reduce_mean(axes = reduce_mean_37_axes_0, keep_dims = reduce_mean_37_keep_dims_0, x = square_18_cast_fp16)[name = tensor("reduce_mean_37_cast_fp16")]; + tensor real_div_4_to_fp16 = const()[name = tensor("real_div_4_to_fp16"), val = tensor(0x1.008p+0)]; + tensor mul_4_cast_fp16 = mul(x = reduce_mean_37_cast_fp16, y = real_div_4_to_fp16)[name = tensor("mul_4_cast_fp16")]; + tensor var_859_to_fp16 = const()[name = tensor("op_859_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_37_cast_fp16 = add(x = mul_4_cast_fp16, y = var_859_to_fp16)[name = tensor("var_37_cast_fp16")]; + tensor var_862_epsilon_0 = const()[name = tensor("op_862_epsilon_0"), val = tensor(0x1.197998p-40)]; + tensor var_862_cast_fp16 = rsqrt(epsilon = var_862_epsilon_0, x = var_37_cast_fp16)[name = tensor("op_862_cast_fp16")]; + tensor var_863_cast_fp16 = mul(x = const_3_to_fp16, y = var_862_cast_fp16)[name = tensor("op_863_cast_fp16")]; + tensor var_864_cast_fp16 = mul(x = linear_54_cast_fp16, y = var_863_cast_fp16)[name = tensor("op_864_cast_fp16")]; + tensor input_145_to_fp16 = const()[name = tensor("input_145_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(19000640)))]; + tensor input_147_cast_fp16 = silu(x = input_145_to_fp16)[name = tensor("input_147_cast_fp16")]; + tensor linear_56_cast_fp16 = linear(bias = flow_net_time_embed_1_mlp_2_bias_to_fp16, weight = flow_net_time_embed_1_mlp_2_weight_to_fp16, x = input_147_cast_fp16)[name = tensor("linear_56_cast_fp16")]; + tensor reduce_mean_38_axes_0 = const()[name = tensor("reduce_mean_38_axes_0"), val = tensor([-1])]; + tensor reduce_mean_38_keep_dims_0 = const()[name = tensor("reduce_mean_38_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_38_cast_fp16 = reduce_mean(axes = reduce_mean_38_axes_0, keep_dims = reduce_mean_38_keep_dims_0, x = linear_56_cast_fp16)[name = tensor("reduce_mean_38_cast_fp16")]; + tensor sub_24_cast_fp16 = sub(x = linear_56_cast_fp16, y = reduce_mean_38_cast_fp16)[name = tensor("sub_24_cast_fp16")]; + tensor square_19_cast_fp16 = square(x = sub_24_cast_fp16)[name = tensor("square_19_cast_fp16")]; + tensor reduce_mean_39_axes_0 = const()[name = tensor("reduce_mean_39_axes_0"), val = tensor([-1])]; + tensor reduce_mean_39_keep_dims_0 = const()[name = tensor("reduce_mean_39_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_39_cast_fp16 = reduce_mean(axes = reduce_mean_39_axes_0, keep_dims = reduce_mean_39_keep_dims_0, x = square_19_cast_fp16)[name = tensor("reduce_mean_39_cast_fp16")]; + tensor real_div_5_to_fp16 = const()[name = tensor("real_div_5_to_fp16"), val = tensor(0x1.008p+0)]; + tensor mul_5_cast_fp16 = mul(x = reduce_mean_39_cast_fp16, y = real_div_5_to_fp16)[name = tensor("mul_5_cast_fp16")]; + tensor var_896_to_fp16 = const()[name = tensor("op_896_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_39_cast_fp16 = add(x = mul_5_cast_fp16, y = var_896_to_fp16)[name = tensor("var_39_cast_fp16")]; + tensor var_899_epsilon_0 = const()[name = tensor("op_899_epsilon_0"), val = tensor(0x1.197998p-40)]; + tensor var_899_cast_fp16 = rsqrt(epsilon = var_899_epsilon_0, x = var_39_cast_fp16)[name = tensor("op_899_cast_fp16")]; + tensor var_900_cast_fp16 = mul(x = const_5_to_fp16, y = var_899_cast_fp16)[name = tensor("op_900_cast_fp16")]; + tensor var_901_cast_fp16 = mul(x = linear_56_cast_fp16, y = var_900_cast_fp16)[name = tensor("op_901_cast_fp16")]; + tensor var_913_cast_fp16 = add(x = var_864_cast_fp16, y = var_901_cast_fp16)[name = tensor("op_913_cast_fp16")]; + tensor _inversed_t_combined_5_y_0_to_fp16 = const()[name = tensor("_inversed_t_combined_5_y_0_to_fp16"), val = tensor(0x1p-1)]; + tensor _inversed_t_combined_5_cast_fp16 = mul(x = var_913_cast_fp16, y = _inversed_t_combined_5_y_0_to_fp16)[name = tensor("_inversed_t_combined_5_cast_fp16")]; + tensor input_149_cast_fp16 = add(x = _inversed_t_combined_5_cast_fp16, y = linear_5_cast_fp16)[name = tensor("input_149_cast_fp16")]; + tensor input_151_cast_fp16 = silu(x = input_149_cast_fp16)[name = tensor("input_151_cast_fp16")]; + tensor linear_58_cast_fp16 = linear(bias = flow_net_res_blocks_0_adaLN_modulation_1_bias_to_fp16, weight = flow_net_res_blocks_0_adaLN_modulation_1_weight_to_fp16, x = input_151_cast_fp16)[name = tensor("linear_58_cast_fp16")]; + tensor var_928_split_sizes_0 = const()[name = tensor("op_928_split_sizes_0"), val = tensor([512, 512, 512])]; + tensor var_928_axis_0 = const()[name = tensor("op_928_axis_0"), val = tensor(-1)]; + tensor var_928_cast_fp16_0, tensor var_928_cast_fp16_1, tensor var_928_cast_fp16_2 = split(axis = var_928_axis_0, split_sizes = var_928_split_sizes_0, x = linear_58_cast_fp16)[name = tensor("op_928_cast_fp16")]; + tensor mean_29_axes_0 = const()[name = tensor("mean_29_axes_0"), val = tensor([-1])]; + tensor mean_29_keep_dims_0 = const()[name = tensor("mean_29_keep_dims_0"), val = tensor(true)]; + tensor mean_29_cast_fp16 = reduce_mean(axes = mean_29_axes_0, keep_dims = mean_29_keep_dims_0, x = linear_52_cast_fp16)[name = tensor("mean_29_cast_fp16")]; + tensor sub_26_cast_fp16 = sub(x = linear_52_cast_fp16, y = mean_29_cast_fp16)[name = tensor("sub_26_cast_fp16")]; + tensor square_20_cast_fp16 = square(x = sub_26_cast_fp16)[name = tensor("square_20_cast_fp16")]; + tensor reduce_mean_41_axes_0 = const()[name = tensor("reduce_mean_41_axes_0"), val = tensor([-1])]; + tensor reduce_mean_41_keep_dims_0 = const()[name = tensor("reduce_mean_41_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_41_cast_fp16 = reduce_mean(axes = reduce_mean_41_axes_0, keep_dims = reduce_mean_41_keep_dims_0, x = square_20_cast_fp16)[name = tensor("reduce_mean_41_cast_fp16")]; + tensor var_938_to_fp16 = const()[name = tensor("op_938_to_fp16"), val = tensor(0x1.1p-20)]; + tensor var_939_cast_fp16 = add(x = reduce_mean_41_cast_fp16, y = var_938_to_fp16)[name = tensor("op_939_cast_fp16")]; + tensor var_940_cast_fp16 = sqrt(x = var_939_cast_fp16)[name = tensor("op_940_cast_fp16")]; + tensor x_95_cast_fp16 = real_div(x = sub_26_cast_fp16, y = var_940_cast_fp16)[name = tensor("x_95_cast_fp16")]; + tensor var_942_cast_fp16 = mul(x = x_95_cast_fp16, y = flow_net_res_blocks_0_in_ln_weight_to_fp16)[name = tensor("op_942_cast_fp16")]; + tensor x_97_cast_fp16 = add(x = var_942_cast_fp16, y = flow_net_res_blocks_0_in_ln_bias_to_fp16)[name = tensor("x_97_cast_fp16")]; + tensor var_944_promoted_to_fp16 = const()[name = tensor("op_944_promoted_to_fp16"), val = tensor(0x1p+0)]; + tensor var_945_cast_fp16 = add(x = var_928_cast_fp16_1, y = var_944_promoted_to_fp16)[name = tensor("op_945_cast_fp16")]; + tensor var_946_cast_fp16 = mul(x = x_97_cast_fp16, y = var_945_cast_fp16)[name = tensor("op_946_cast_fp16")]; + tensor input_153_cast_fp16 = add(x = var_946_cast_fp16, y = var_928_cast_fp16_0)[name = tensor("input_153_cast_fp16")]; + tensor linear_59_cast_fp16 = linear(bias = flow_net_res_blocks_0_mlp_0_bias_to_fp16, weight = flow_net_res_blocks_0_mlp_0_weight_to_fp16, x = input_153_cast_fp16)[name = tensor("linear_59_cast_fp16")]; + tensor input_157_cast_fp16 = silu(x = linear_59_cast_fp16)[name = tensor("input_157_cast_fp16")]; + tensor linear_60_cast_fp16 = linear(bias = flow_net_res_blocks_0_mlp_2_bias_to_fp16, weight = flow_net_res_blocks_0_mlp_2_weight_to_fp16, x = input_157_cast_fp16)[name = tensor("linear_60_cast_fp16")]; + tensor var_957_cast_fp16 = mul(x = var_928_cast_fp16_2, y = linear_60_cast_fp16)[name = tensor("op_957_cast_fp16")]; + tensor x_99_cast_fp16 = add(x = linear_52_cast_fp16, y = var_957_cast_fp16)[name = tensor("x_99_cast_fp16")]; + tensor linear_61_cast_fp16 = linear(bias = flow_net_res_blocks_1_adaLN_modulation_1_bias_to_fp16, weight = flow_net_res_blocks_1_adaLN_modulation_1_weight_to_fp16, x = input_151_cast_fp16)[name = tensor("linear_61_cast_fp16")]; + tensor var_967_split_sizes_0 = const()[name = tensor("op_967_split_sizes_0"), val = tensor([512, 512, 512])]; + tensor var_967_axis_0 = const()[name = tensor("op_967_axis_0"), val = tensor(-1)]; + tensor var_967_cast_fp16_0, tensor var_967_cast_fp16_1, tensor var_967_cast_fp16_2 = split(axis = var_967_axis_0, split_sizes = var_967_split_sizes_0, x = linear_61_cast_fp16)[name = tensor("op_967_cast_fp16")]; + tensor mean_31_axes_0 = const()[name = tensor("mean_31_axes_0"), val = tensor([-1])]; + tensor mean_31_keep_dims_0 = const()[name = tensor("mean_31_keep_dims_0"), val = tensor(true)]; + tensor mean_31_cast_fp16 = reduce_mean(axes = mean_31_axes_0, keep_dims = mean_31_keep_dims_0, x = x_99_cast_fp16)[name = tensor("mean_31_cast_fp16")]; + tensor sub_27_cast_fp16 = sub(x = x_99_cast_fp16, y = mean_31_cast_fp16)[name = tensor("sub_27_cast_fp16")]; + tensor square_21_cast_fp16 = square(x = sub_27_cast_fp16)[name = tensor("square_21_cast_fp16")]; + tensor reduce_mean_43_axes_0 = const()[name = tensor("reduce_mean_43_axes_0"), val = tensor([-1])]; + tensor reduce_mean_43_keep_dims_0 = const()[name = tensor("reduce_mean_43_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_43_cast_fp16 = reduce_mean(axes = reduce_mean_43_axes_0, keep_dims = reduce_mean_43_keep_dims_0, x = square_21_cast_fp16)[name = tensor("reduce_mean_43_cast_fp16")]; + tensor var_977_to_fp16 = const()[name = tensor("op_977_to_fp16"), val = tensor(0x1.1p-20)]; + tensor var_978_cast_fp16 = add(x = reduce_mean_43_cast_fp16, y = var_977_to_fp16)[name = tensor("op_978_cast_fp16")]; + tensor var_979_cast_fp16 = sqrt(x = var_978_cast_fp16)[name = tensor("op_979_cast_fp16")]; + tensor x_101_cast_fp16 = real_div(x = sub_27_cast_fp16, y = var_979_cast_fp16)[name = tensor("x_101_cast_fp16")]; + tensor var_981_cast_fp16 = mul(x = x_101_cast_fp16, y = flow_net_res_blocks_1_in_ln_weight_to_fp16)[name = tensor("op_981_cast_fp16")]; + tensor x_103_cast_fp16 = add(x = var_981_cast_fp16, y = flow_net_res_blocks_1_in_ln_bias_to_fp16)[name = tensor("x_103_cast_fp16")]; + tensor var_983_promoted_to_fp16 = const()[name = tensor("op_983_promoted_to_fp16"), val = tensor(0x1p+0)]; + tensor var_984_cast_fp16 = add(x = var_967_cast_fp16_1, y = var_983_promoted_to_fp16)[name = tensor("op_984_cast_fp16")]; + tensor var_985_cast_fp16 = mul(x = x_103_cast_fp16, y = var_984_cast_fp16)[name = tensor("op_985_cast_fp16")]; + tensor input_161_cast_fp16 = add(x = var_985_cast_fp16, y = var_967_cast_fp16_0)[name = tensor("input_161_cast_fp16")]; + tensor linear_62_cast_fp16 = linear(bias = flow_net_res_blocks_1_mlp_0_bias_to_fp16, weight = flow_net_res_blocks_1_mlp_0_weight_to_fp16, x = input_161_cast_fp16)[name = tensor("linear_62_cast_fp16")]; + tensor input_165_cast_fp16 = silu(x = linear_62_cast_fp16)[name = tensor("input_165_cast_fp16")]; + tensor linear_63_cast_fp16 = linear(bias = flow_net_res_blocks_1_mlp_2_bias_to_fp16, weight = flow_net_res_blocks_1_mlp_2_weight_to_fp16, x = input_165_cast_fp16)[name = tensor("linear_63_cast_fp16")]; + tensor var_996_cast_fp16 = mul(x = var_967_cast_fp16_2, y = linear_63_cast_fp16)[name = tensor("op_996_cast_fp16")]; + tensor x_105_cast_fp16 = add(x = x_99_cast_fp16, y = var_996_cast_fp16)[name = tensor("x_105_cast_fp16")]; + tensor linear_64_cast_fp16 = linear(bias = flow_net_res_blocks_2_adaLN_modulation_1_bias_to_fp16, weight = flow_net_res_blocks_2_adaLN_modulation_1_weight_to_fp16, x = input_151_cast_fp16)[name = tensor("linear_64_cast_fp16")]; + tensor var_1006_split_sizes_0 = const()[name = tensor("op_1006_split_sizes_0"), val = tensor([512, 512, 512])]; + tensor var_1006_axis_0 = const()[name = tensor("op_1006_axis_0"), val = tensor(-1)]; + tensor var_1006_cast_fp16_0, tensor var_1006_cast_fp16_1, tensor var_1006_cast_fp16_2 = split(axis = var_1006_axis_0, split_sizes = var_1006_split_sizes_0, x = linear_64_cast_fp16)[name = tensor("op_1006_cast_fp16")]; + tensor mean_33_axes_0 = const()[name = tensor("mean_33_axes_0"), val = tensor([-1])]; + tensor mean_33_keep_dims_0 = const()[name = tensor("mean_33_keep_dims_0"), val = tensor(true)]; + tensor mean_33_cast_fp16 = reduce_mean(axes = mean_33_axes_0, keep_dims = mean_33_keep_dims_0, x = x_105_cast_fp16)[name = tensor("mean_33_cast_fp16")]; + tensor sub_28_cast_fp16 = sub(x = x_105_cast_fp16, y = mean_33_cast_fp16)[name = tensor("sub_28_cast_fp16")]; + tensor square_22_cast_fp16 = square(x = sub_28_cast_fp16)[name = tensor("square_22_cast_fp16")]; + tensor reduce_mean_45_axes_0 = const()[name = tensor("reduce_mean_45_axes_0"), val = tensor([-1])]; + tensor reduce_mean_45_keep_dims_0 = const()[name = tensor("reduce_mean_45_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_45_cast_fp16 = reduce_mean(axes = reduce_mean_45_axes_0, keep_dims = reduce_mean_45_keep_dims_0, x = square_22_cast_fp16)[name = tensor("reduce_mean_45_cast_fp16")]; + tensor var_1016_to_fp16 = const()[name = tensor("op_1016_to_fp16"), val = tensor(0x1.1p-20)]; + tensor var_1017_cast_fp16 = add(x = reduce_mean_45_cast_fp16, y = var_1016_to_fp16)[name = tensor("op_1017_cast_fp16")]; + tensor var_1018_cast_fp16 = sqrt(x = var_1017_cast_fp16)[name = tensor("op_1018_cast_fp16")]; + tensor x_107_cast_fp16 = real_div(x = sub_28_cast_fp16, y = var_1018_cast_fp16)[name = tensor("x_107_cast_fp16")]; + tensor var_1020_cast_fp16 = mul(x = x_107_cast_fp16, y = flow_net_res_blocks_2_in_ln_weight_to_fp16)[name = tensor("op_1020_cast_fp16")]; + tensor x_109_cast_fp16 = add(x = var_1020_cast_fp16, y = flow_net_res_blocks_2_in_ln_bias_to_fp16)[name = tensor("x_109_cast_fp16")]; + tensor var_1022_promoted_to_fp16 = const()[name = tensor("op_1022_promoted_to_fp16"), val = tensor(0x1p+0)]; + tensor var_1023_cast_fp16 = add(x = var_1006_cast_fp16_1, y = var_1022_promoted_to_fp16)[name = tensor("op_1023_cast_fp16")]; + tensor var_1024_cast_fp16 = mul(x = x_109_cast_fp16, y = var_1023_cast_fp16)[name = tensor("op_1024_cast_fp16")]; + tensor input_169_cast_fp16 = add(x = var_1024_cast_fp16, y = var_1006_cast_fp16_0)[name = tensor("input_169_cast_fp16")]; + tensor linear_65_cast_fp16 = linear(bias = flow_net_res_blocks_2_mlp_0_bias_to_fp16, weight = flow_net_res_blocks_2_mlp_0_weight_to_fp16, x = input_169_cast_fp16)[name = tensor("linear_65_cast_fp16")]; + tensor input_173_cast_fp16 = silu(x = linear_65_cast_fp16)[name = tensor("input_173_cast_fp16")]; + tensor linear_66_cast_fp16 = linear(bias = flow_net_res_blocks_2_mlp_2_bias_to_fp16, weight = flow_net_res_blocks_2_mlp_2_weight_to_fp16, x = input_173_cast_fp16)[name = tensor("linear_66_cast_fp16")]; + tensor var_1035_cast_fp16 = mul(x = var_1006_cast_fp16_2, y = linear_66_cast_fp16)[name = tensor("op_1035_cast_fp16")]; + tensor x_111_cast_fp16 = add(x = x_105_cast_fp16, y = var_1035_cast_fp16)[name = tensor("x_111_cast_fp16")]; + tensor linear_67_cast_fp16 = linear(bias = flow_net_res_blocks_3_adaLN_modulation_1_bias_to_fp16, weight = flow_net_res_blocks_3_adaLN_modulation_1_weight_to_fp16, x = input_151_cast_fp16)[name = tensor("linear_67_cast_fp16")]; + tensor var_1045_split_sizes_0 = const()[name = tensor("op_1045_split_sizes_0"), val = tensor([512, 512, 512])]; + tensor var_1045_axis_0 = const()[name = tensor("op_1045_axis_0"), val = tensor(-1)]; + tensor var_1045_cast_fp16_0, tensor var_1045_cast_fp16_1, tensor var_1045_cast_fp16_2 = split(axis = var_1045_axis_0, split_sizes = var_1045_split_sizes_0, x = linear_67_cast_fp16)[name = tensor("op_1045_cast_fp16")]; + tensor mean_35_axes_0 = const()[name = tensor("mean_35_axes_0"), val = tensor([-1])]; + tensor mean_35_keep_dims_0 = const()[name = tensor("mean_35_keep_dims_0"), val = tensor(true)]; + tensor mean_35_cast_fp16 = reduce_mean(axes = mean_35_axes_0, keep_dims = mean_35_keep_dims_0, x = x_111_cast_fp16)[name = tensor("mean_35_cast_fp16")]; + tensor sub_29_cast_fp16 = sub(x = x_111_cast_fp16, y = mean_35_cast_fp16)[name = tensor("sub_29_cast_fp16")]; + tensor square_23_cast_fp16 = square(x = sub_29_cast_fp16)[name = tensor("square_23_cast_fp16")]; + tensor reduce_mean_47_axes_0 = const()[name = tensor("reduce_mean_47_axes_0"), val = tensor([-1])]; + tensor reduce_mean_47_keep_dims_0 = const()[name = tensor("reduce_mean_47_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_47_cast_fp16 = reduce_mean(axes = reduce_mean_47_axes_0, keep_dims = reduce_mean_47_keep_dims_0, x = square_23_cast_fp16)[name = tensor("reduce_mean_47_cast_fp16")]; + tensor var_1055_to_fp16 = const()[name = tensor("op_1055_to_fp16"), val = tensor(0x1.1p-20)]; + tensor var_1056_cast_fp16 = add(x = reduce_mean_47_cast_fp16, y = var_1055_to_fp16)[name = tensor("op_1056_cast_fp16")]; + tensor var_1057_cast_fp16 = sqrt(x = var_1056_cast_fp16)[name = tensor("op_1057_cast_fp16")]; + tensor x_113_cast_fp16 = real_div(x = sub_29_cast_fp16, y = var_1057_cast_fp16)[name = tensor("x_113_cast_fp16")]; + tensor var_1059_cast_fp16 = mul(x = x_113_cast_fp16, y = flow_net_res_blocks_3_in_ln_weight_to_fp16)[name = tensor("op_1059_cast_fp16")]; + tensor x_115_cast_fp16 = add(x = var_1059_cast_fp16, y = flow_net_res_blocks_3_in_ln_bias_to_fp16)[name = tensor("x_115_cast_fp16")]; + tensor var_1061_promoted_to_fp16 = const()[name = tensor("op_1061_promoted_to_fp16"), val = tensor(0x1p+0)]; + tensor var_1062_cast_fp16 = add(x = var_1045_cast_fp16_1, y = var_1061_promoted_to_fp16)[name = tensor("op_1062_cast_fp16")]; + tensor var_1063_cast_fp16 = mul(x = x_115_cast_fp16, y = var_1062_cast_fp16)[name = tensor("op_1063_cast_fp16")]; + tensor input_177_cast_fp16 = add(x = var_1063_cast_fp16, y = var_1045_cast_fp16_0)[name = tensor("input_177_cast_fp16")]; + tensor linear_68_cast_fp16 = linear(bias = flow_net_res_blocks_3_mlp_0_bias_to_fp16, weight = flow_net_res_blocks_3_mlp_0_weight_to_fp16, x = input_177_cast_fp16)[name = tensor("linear_68_cast_fp16")]; + tensor input_181_cast_fp16 = silu(x = linear_68_cast_fp16)[name = tensor("input_181_cast_fp16")]; + tensor linear_69_cast_fp16 = linear(bias = flow_net_res_blocks_3_mlp_2_bias_to_fp16, weight = flow_net_res_blocks_3_mlp_2_weight_to_fp16, x = input_181_cast_fp16)[name = tensor("linear_69_cast_fp16")]; + tensor var_1074_cast_fp16 = mul(x = var_1045_cast_fp16_2, y = linear_69_cast_fp16)[name = tensor("op_1074_cast_fp16")]; + tensor x_117_cast_fp16 = add(x = x_111_cast_fp16, y = var_1074_cast_fp16)[name = tensor("x_117_cast_fp16")]; + tensor linear_70_cast_fp16 = linear(bias = flow_net_res_blocks_4_adaLN_modulation_1_bias_to_fp16, weight = flow_net_res_blocks_4_adaLN_modulation_1_weight_to_fp16, x = input_151_cast_fp16)[name = tensor("linear_70_cast_fp16")]; + tensor var_1084_split_sizes_0 = const()[name = tensor("op_1084_split_sizes_0"), val = tensor([512, 512, 512])]; + tensor var_1084_axis_0 = const()[name = tensor("op_1084_axis_0"), val = tensor(-1)]; + tensor var_1084_cast_fp16_0, tensor var_1084_cast_fp16_1, tensor var_1084_cast_fp16_2 = split(axis = var_1084_axis_0, split_sizes = var_1084_split_sizes_0, x = linear_70_cast_fp16)[name = tensor("op_1084_cast_fp16")]; + tensor mean_37_axes_0 = const()[name = tensor("mean_37_axes_0"), val = tensor([-1])]; + tensor mean_37_keep_dims_0 = const()[name = tensor("mean_37_keep_dims_0"), val = tensor(true)]; + tensor mean_37_cast_fp16 = reduce_mean(axes = mean_37_axes_0, keep_dims = mean_37_keep_dims_0, x = x_117_cast_fp16)[name = tensor("mean_37_cast_fp16")]; + tensor sub_30_cast_fp16 = sub(x = x_117_cast_fp16, y = mean_37_cast_fp16)[name = tensor("sub_30_cast_fp16")]; + tensor square_24_cast_fp16 = square(x = sub_30_cast_fp16)[name = tensor("square_24_cast_fp16")]; + tensor reduce_mean_49_axes_0 = const()[name = tensor("reduce_mean_49_axes_0"), val = tensor([-1])]; + tensor reduce_mean_49_keep_dims_0 = const()[name = tensor("reduce_mean_49_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_49_cast_fp16 = reduce_mean(axes = reduce_mean_49_axes_0, keep_dims = reduce_mean_49_keep_dims_0, x = square_24_cast_fp16)[name = tensor("reduce_mean_49_cast_fp16")]; + tensor var_1094_to_fp16 = const()[name = tensor("op_1094_to_fp16"), val = tensor(0x1.1p-20)]; + tensor var_1095_cast_fp16 = add(x = reduce_mean_49_cast_fp16, y = var_1094_to_fp16)[name = tensor("op_1095_cast_fp16")]; + tensor var_1096_cast_fp16 = sqrt(x = var_1095_cast_fp16)[name = tensor("op_1096_cast_fp16")]; + tensor x_119_cast_fp16 = real_div(x = sub_30_cast_fp16, y = var_1096_cast_fp16)[name = tensor("x_119_cast_fp16")]; + tensor var_1098_cast_fp16 = mul(x = x_119_cast_fp16, y = flow_net_res_blocks_4_in_ln_weight_to_fp16)[name = tensor("op_1098_cast_fp16")]; + tensor x_121_cast_fp16 = add(x = var_1098_cast_fp16, y = flow_net_res_blocks_4_in_ln_bias_to_fp16)[name = tensor("x_121_cast_fp16")]; + tensor var_1100_promoted_to_fp16 = const()[name = tensor("op_1100_promoted_to_fp16"), val = tensor(0x1p+0)]; + tensor var_1101_cast_fp16 = add(x = var_1084_cast_fp16_1, y = var_1100_promoted_to_fp16)[name = tensor("op_1101_cast_fp16")]; + tensor var_1102_cast_fp16 = mul(x = x_121_cast_fp16, y = var_1101_cast_fp16)[name = tensor("op_1102_cast_fp16")]; + tensor input_185_cast_fp16 = add(x = var_1102_cast_fp16, y = var_1084_cast_fp16_0)[name = tensor("input_185_cast_fp16")]; + tensor linear_71_cast_fp16 = linear(bias = flow_net_res_blocks_4_mlp_0_bias_to_fp16, weight = flow_net_res_blocks_4_mlp_0_weight_to_fp16, x = input_185_cast_fp16)[name = tensor("linear_71_cast_fp16")]; + tensor input_189_cast_fp16 = silu(x = linear_71_cast_fp16)[name = tensor("input_189_cast_fp16")]; + tensor linear_72_cast_fp16 = linear(bias = flow_net_res_blocks_4_mlp_2_bias_to_fp16, weight = flow_net_res_blocks_4_mlp_2_weight_to_fp16, x = input_189_cast_fp16)[name = tensor("linear_72_cast_fp16")]; + tensor var_1113_cast_fp16 = mul(x = var_1084_cast_fp16_2, y = linear_72_cast_fp16)[name = tensor("op_1113_cast_fp16")]; + tensor x_123_cast_fp16 = add(x = x_117_cast_fp16, y = var_1113_cast_fp16)[name = tensor("x_123_cast_fp16")]; + tensor linear_73_cast_fp16 = linear(bias = flow_net_res_blocks_5_adaLN_modulation_1_bias_to_fp16, weight = flow_net_res_blocks_5_adaLN_modulation_1_weight_to_fp16, x = input_151_cast_fp16)[name = tensor("linear_73_cast_fp16")]; + tensor var_1123_split_sizes_0 = const()[name = tensor("op_1123_split_sizes_0"), val = tensor([512, 512, 512])]; + tensor var_1123_axis_0 = const()[name = tensor("op_1123_axis_0"), val = tensor(-1)]; + tensor var_1123_cast_fp16_0, tensor var_1123_cast_fp16_1, tensor var_1123_cast_fp16_2 = split(axis = var_1123_axis_0, split_sizes = var_1123_split_sizes_0, x = linear_73_cast_fp16)[name = tensor("op_1123_cast_fp16")]; + tensor mean_39_axes_0 = const()[name = tensor("mean_39_axes_0"), val = tensor([-1])]; + tensor mean_39_keep_dims_0 = const()[name = tensor("mean_39_keep_dims_0"), val = tensor(true)]; + tensor mean_39_cast_fp16 = reduce_mean(axes = mean_39_axes_0, keep_dims = mean_39_keep_dims_0, x = x_123_cast_fp16)[name = tensor("mean_39_cast_fp16")]; + tensor sub_31_cast_fp16 = sub(x = x_123_cast_fp16, y = mean_39_cast_fp16)[name = tensor("sub_31_cast_fp16")]; + tensor square_25_cast_fp16 = square(x = sub_31_cast_fp16)[name = tensor("square_25_cast_fp16")]; + tensor reduce_mean_51_axes_0 = const()[name = tensor("reduce_mean_51_axes_0"), val = tensor([-1])]; + tensor reduce_mean_51_keep_dims_0 = const()[name = tensor("reduce_mean_51_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_51_cast_fp16 = reduce_mean(axes = reduce_mean_51_axes_0, keep_dims = reduce_mean_51_keep_dims_0, x = square_25_cast_fp16)[name = tensor("reduce_mean_51_cast_fp16")]; + tensor var_1133_to_fp16 = const()[name = tensor("op_1133_to_fp16"), val = tensor(0x1.1p-20)]; + tensor var_1134_cast_fp16 = add(x = reduce_mean_51_cast_fp16, y = var_1133_to_fp16)[name = tensor("op_1134_cast_fp16")]; + tensor var_1135_cast_fp16 = sqrt(x = var_1134_cast_fp16)[name = tensor("op_1135_cast_fp16")]; + tensor x_125_cast_fp16 = real_div(x = sub_31_cast_fp16, y = var_1135_cast_fp16)[name = tensor("x_125_cast_fp16")]; + tensor var_1137_cast_fp16 = mul(x = x_125_cast_fp16, y = flow_net_res_blocks_5_in_ln_weight_to_fp16)[name = tensor("op_1137_cast_fp16")]; + tensor x_127_cast_fp16 = add(x = var_1137_cast_fp16, y = flow_net_res_blocks_5_in_ln_bias_to_fp16)[name = tensor("x_127_cast_fp16")]; + tensor var_1139_promoted_to_fp16 = const()[name = tensor("op_1139_promoted_to_fp16"), val = tensor(0x1p+0)]; + tensor var_1140_cast_fp16 = add(x = var_1123_cast_fp16_1, y = var_1139_promoted_to_fp16)[name = tensor("op_1140_cast_fp16")]; + tensor var_1141_cast_fp16 = mul(x = x_127_cast_fp16, y = var_1140_cast_fp16)[name = tensor("op_1141_cast_fp16")]; + tensor input_193_cast_fp16 = add(x = var_1141_cast_fp16, y = var_1123_cast_fp16_0)[name = tensor("input_193_cast_fp16")]; + tensor linear_74_cast_fp16 = linear(bias = flow_net_res_blocks_5_mlp_0_bias_to_fp16, weight = flow_net_res_blocks_5_mlp_0_weight_to_fp16, x = input_193_cast_fp16)[name = tensor("linear_74_cast_fp16")]; + tensor input_197_cast_fp16 = silu(x = linear_74_cast_fp16)[name = tensor("input_197_cast_fp16")]; + tensor linear_75_cast_fp16 = linear(bias = flow_net_res_blocks_5_mlp_2_bias_to_fp16, weight = flow_net_res_blocks_5_mlp_2_weight_to_fp16, x = input_197_cast_fp16)[name = tensor("linear_75_cast_fp16")]; + tensor var_1152_cast_fp16 = mul(x = var_1123_cast_fp16_2, y = linear_75_cast_fp16)[name = tensor("op_1152_cast_fp16")]; + tensor x_129_cast_fp16 = add(x = x_123_cast_fp16, y = var_1152_cast_fp16)[name = tensor("x_129_cast_fp16")]; + tensor linear_76_cast_fp16 = linear(bias = flow_net_final_layer_adaLN_modulation_1_bias_to_fp16, weight = flow_net_final_layer_adaLN_modulation_1_weight_to_fp16, x = input_151_cast_fp16)[name = tensor("linear_76_cast_fp16")]; + tensor var_1161_split_sizes_0 = const()[name = tensor("op_1161_split_sizes_0"), val = tensor([512, 512])]; + tensor var_1161_axis_0 = const()[name = tensor("op_1161_axis_0"), val = tensor(-1)]; + tensor var_1161_cast_fp16_0, tensor var_1161_cast_fp16_1 = split(axis = var_1161_axis_0, split_sizes = var_1161_split_sizes_0, x = linear_76_cast_fp16)[name = tensor("op_1161_cast_fp16")]; + tensor mean_41_axes_0 = const()[name = tensor("mean_41_axes_0"), val = tensor([-1])]; + tensor mean_41_keep_dims_0 = const()[name = tensor("mean_41_keep_dims_0"), val = tensor(true)]; + tensor mean_41_cast_fp16 = reduce_mean(axes = mean_41_axes_0, keep_dims = mean_41_keep_dims_0, x = x_129_cast_fp16)[name = tensor("mean_41_cast_fp16")]; + tensor sub_32_cast_fp16 = sub(x = x_129_cast_fp16, y = mean_41_cast_fp16)[name = tensor("sub_32_cast_fp16")]; + tensor square_26_cast_fp16 = square(x = sub_32_cast_fp16)[name = tensor("square_26_cast_fp16")]; + tensor reduce_mean_53_axes_0 = const()[name = tensor("reduce_mean_53_axes_0"), val = tensor([-1])]; + tensor reduce_mean_53_keep_dims_0 = const()[name = tensor("reduce_mean_53_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_53_cast_fp16 = reduce_mean(axes = reduce_mean_53_axes_0, keep_dims = reduce_mean_53_keep_dims_0, x = square_26_cast_fp16)[name = tensor("reduce_mean_53_cast_fp16")]; + tensor var_1168_to_fp16 = const()[name = tensor("op_1168_to_fp16"), val = tensor(0x1.1p-20)]; + tensor var_1169_cast_fp16 = add(x = reduce_mean_53_cast_fp16, y = var_1168_to_fp16)[name = tensor("op_1169_cast_fp16")]; + tensor var_1170_cast_fp16 = sqrt(x = var_1169_cast_fp16)[name = tensor("op_1170_cast_fp16")]; + tensor x_131_cast_fp16 = real_div(x = sub_32_cast_fp16, y = var_1170_cast_fp16)[name = tensor("x_131_cast_fp16")]; + tensor var_1172_promoted_to_fp16 = const()[name = tensor("op_1172_promoted_to_fp16"), val = tensor(0x1p+0)]; + tensor var_1173_cast_fp16 = add(x = var_1161_cast_fp16_1, y = var_1172_promoted_to_fp16)[name = tensor("op_1173_cast_fp16")]; + tensor var_1174_cast_fp16 = mul(x = x_131_cast_fp16, y = var_1173_cast_fp16)[name = tensor("op_1174_cast_fp16")]; + tensor input_201_cast_fp16 = add(x = var_1174_cast_fp16, y = var_1161_cast_fp16_0)[name = tensor("input_201_cast_fp16")]; + tensor linear_77_cast_fp16 = linear(bias = flow_net_final_layer_linear_bias_to_fp16, weight = flow_net_final_layer_linear_weight_to_fp16, x = input_201_cast_fp16)[name = tensor("linear_77_cast_fp16")]; + tensor var_1185_to_fp16 = const()[name = tensor("op_1185_to_fp16"), val = tensor(0x1p-3)]; + tensor var_1186_cast_fp16 = mul(x = linear_77_cast_fp16, y = var_1185_to_fp16)[name = tensor("op_1186_cast_fp16")]; + tensor input_203_cast_fp16 = add(x = input_135_cast_fp16, y = var_1186_cast_fp16)[name = tensor("input_203_cast_fp16")]; + tensor linear_78_cast_fp16 = linear(bias = flow_net_input_proj_bias_to_fp16, weight = flow_net_input_proj_weight_to_fp16, x = input_203_cast_fp16)[name = tensor("linear_78_cast_fp16")]; + tensor input_207_to_fp16 = const()[name = tensor("input_207_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(19001728)))]; + tensor input_209_cast_fp16 = silu(x = input_207_to_fp16)[name = tensor("input_209_cast_fp16")]; + tensor linear_80_cast_fp16 = linear(bias = flow_net_time_embed_0_mlp_2_bias_to_fp16, weight = flow_net_time_embed_0_mlp_2_weight_to_fp16, x = input_209_cast_fp16)[name = tensor("linear_80_cast_fp16")]; + tensor reduce_mean_54_axes_0 = const()[name = tensor("reduce_mean_54_axes_0"), val = tensor([-1])]; + tensor reduce_mean_54_keep_dims_0 = const()[name = tensor("reduce_mean_54_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_54_cast_fp16 = reduce_mean(axes = reduce_mean_54_axes_0, keep_dims = reduce_mean_54_keep_dims_0, x = linear_80_cast_fp16)[name = tensor("reduce_mean_54_cast_fp16")]; + tensor sub_33_cast_fp16 = sub(x = linear_80_cast_fp16, y = reduce_mean_54_cast_fp16)[name = tensor("sub_33_cast_fp16")]; + tensor square_27_cast_fp16 = square(x = sub_33_cast_fp16)[name = tensor("square_27_cast_fp16")]; + tensor reduce_mean_55_axes_0 = const()[name = tensor("reduce_mean_55_axes_0"), val = tensor([-1])]; + tensor reduce_mean_55_keep_dims_0 = const()[name = tensor("reduce_mean_55_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_55_cast_fp16 = reduce_mean(axes = reduce_mean_55_axes_0, keep_dims = reduce_mean_55_keep_dims_0, x = square_27_cast_fp16)[name = tensor("reduce_mean_55_cast_fp16")]; + tensor real_div_6_to_fp16 = const()[name = tensor("real_div_6_to_fp16"), val = tensor(0x1.008p+0)]; + tensor mul_6_cast_fp16 = mul(x = reduce_mean_55_cast_fp16, y = real_div_6_to_fp16)[name = tensor("mul_6_cast_fp16")]; + tensor var_1252_to_fp16 = const()[name = tensor("op_1252_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_55_cast_fp16 = add(x = mul_6_cast_fp16, y = var_1252_to_fp16)[name = tensor("var_55_cast_fp16")]; + tensor var_1255_epsilon_0 = const()[name = tensor("op_1255_epsilon_0"), val = tensor(0x1.197998p-40)]; + tensor var_1255_cast_fp16 = rsqrt(epsilon = var_1255_epsilon_0, x = var_55_cast_fp16)[name = tensor("op_1255_cast_fp16")]; + tensor var_1256_cast_fp16 = mul(x = const_3_to_fp16, y = var_1255_cast_fp16)[name = tensor("op_1256_cast_fp16")]; + tensor var_1257_cast_fp16 = mul(x = linear_80_cast_fp16, y = var_1256_cast_fp16)[name = tensor("op_1257_cast_fp16")]; + tensor input_213_to_fp16 = const()[name = tensor("input_213_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(19002816)))]; + tensor input_215_cast_fp16 = silu(x = input_213_to_fp16)[name = tensor("input_215_cast_fp16")]; + tensor linear_82_cast_fp16 = linear(bias = flow_net_time_embed_1_mlp_2_bias_to_fp16, weight = flow_net_time_embed_1_mlp_2_weight_to_fp16, x = input_215_cast_fp16)[name = tensor("linear_82_cast_fp16")]; + tensor reduce_mean_56_axes_0 = const()[name = tensor("reduce_mean_56_axes_0"), val = tensor([-1])]; + tensor reduce_mean_56_keep_dims_0 = const()[name = tensor("reduce_mean_56_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_56_cast_fp16 = reduce_mean(axes = reduce_mean_56_axes_0, keep_dims = reduce_mean_56_keep_dims_0, x = linear_82_cast_fp16)[name = tensor("reduce_mean_56_cast_fp16")]; + tensor sub_35_cast_fp16 = sub(x = linear_82_cast_fp16, y = reduce_mean_56_cast_fp16)[name = tensor("sub_35_cast_fp16")]; + tensor square_28_cast_fp16 = square(x = sub_35_cast_fp16)[name = tensor("square_28_cast_fp16")]; + tensor reduce_mean_57_axes_0 = const()[name = tensor("reduce_mean_57_axes_0"), val = tensor([-1])]; + tensor reduce_mean_57_keep_dims_0 = const()[name = tensor("reduce_mean_57_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_57_cast_fp16 = reduce_mean(axes = reduce_mean_57_axes_0, keep_dims = reduce_mean_57_keep_dims_0, x = square_28_cast_fp16)[name = tensor("reduce_mean_57_cast_fp16")]; + tensor real_div_7_to_fp16 = const()[name = tensor("real_div_7_to_fp16"), val = tensor(0x1.008p+0)]; + tensor mul_7_cast_fp16 = mul(x = reduce_mean_57_cast_fp16, y = real_div_7_to_fp16)[name = tensor("mul_7_cast_fp16")]; + tensor var_1289_to_fp16 = const()[name = tensor("op_1289_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_57_cast_fp16 = add(x = mul_7_cast_fp16, y = var_1289_to_fp16)[name = tensor("var_57_cast_fp16")]; + tensor var_1292_epsilon_0 = const()[name = tensor("op_1292_epsilon_0"), val = tensor(0x1.197998p-40)]; + tensor var_1292_cast_fp16 = rsqrt(epsilon = var_1292_epsilon_0, x = var_57_cast_fp16)[name = tensor("op_1292_cast_fp16")]; + tensor var_1293_cast_fp16 = mul(x = const_5_to_fp16, y = var_1292_cast_fp16)[name = tensor("op_1293_cast_fp16")]; + tensor var_1294_cast_fp16 = mul(x = linear_82_cast_fp16, y = var_1293_cast_fp16)[name = tensor("op_1294_cast_fp16")]; + tensor var_1306_cast_fp16 = add(x = var_1257_cast_fp16, y = var_1294_cast_fp16)[name = tensor("op_1306_cast_fp16")]; + tensor _inversed_t_combined_7_y_0_to_fp16 = const()[name = tensor("_inversed_t_combined_7_y_0_to_fp16"), val = tensor(0x1p-1)]; + tensor _inversed_t_combined_7_cast_fp16 = mul(x = var_1306_cast_fp16, y = _inversed_t_combined_7_y_0_to_fp16)[name = tensor("_inversed_t_combined_7_cast_fp16")]; + tensor input_217_cast_fp16 = add(x = _inversed_t_combined_7_cast_fp16, y = linear_5_cast_fp16)[name = tensor("input_217_cast_fp16")]; + tensor input_219_cast_fp16 = silu(x = input_217_cast_fp16)[name = tensor("input_219_cast_fp16")]; + tensor linear_84_cast_fp16 = linear(bias = flow_net_res_blocks_0_adaLN_modulation_1_bias_to_fp16, weight = flow_net_res_blocks_0_adaLN_modulation_1_weight_to_fp16, x = input_219_cast_fp16)[name = tensor("linear_84_cast_fp16")]; + tensor var_1321_split_sizes_0 = const()[name = tensor("op_1321_split_sizes_0"), val = tensor([512, 512, 512])]; + tensor var_1321_axis_0 = const()[name = tensor("op_1321_axis_0"), val = tensor(-1)]; + tensor var_1321_cast_fp16_0, tensor var_1321_cast_fp16_1, tensor var_1321_cast_fp16_2 = split(axis = var_1321_axis_0, split_sizes = var_1321_split_sizes_0, x = linear_84_cast_fp16)[name = tensor("op_1321_cast_fp16")]; + tensor mean_43_axes_0 = const()[name = tensor("mean_43_axes_0"), val = tensor([-1])]; + tensor mean_43_keep_dims_0 = const()[name = tensor("mean_43_keep_dims_0"), val = tensor(true)]; + tensor mean_43_cast_fp16 = reduce_mean(axes = mean_43_axes_0, keep_dims = mean_43_keep_dims_0, x = linear_78_cast_fp16)[name = tensor("mean_43_cast_fp16")]; + tensor sub_37_cast_fp16 = sub(x = linear_78_cast_fp16, y = mean_43_cast_fp16)[name = tensor("sub_37_cast_fp16")]; + tensor square_29_cast_fp16 = square(x = sub_37_cast_fp16)[name = tensor("square_29_cast_fp16")]; + tensor reduce_mean_59_axes_0 = const()[name = tensor("reduce_mean_59_axes_0"), val = tensor([-1])]; + tensor reduce_mean_59_keep_dims_0 = const()[name = tensor("reduce_mean_59_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_59_cast_fp16 = reduce_mean(axes = reduce_mean_59_axes_0, keep_dims = reduce_mean_59_keep_dims_0, x = square_29_cast_fp16)[name = tensor("reduce_mean_59_cast_fp16")]; + tensor var_1331_to_fp16 = const()[name = tensor("op_1331_to_fp16"), val = tensor(0x1.1p-20)]; + tensor var_1332_cast_fp16 = add(x = reduce_mean_59_cast_fp16, y = var_1331_to_fp16)[name = tensor("op_1332_cast_fp16")]; + tensor var_1333_cast_fp16 = sqrt(x = var_1332_cast_fp16)[name = tensor("op_1333_cast_fp16")]; + tensor x_139_cast_fp16 = real_div(x = sub_37_cast_fp16, y = var_1333_cast_fp16)[name = tensor("x_139_cast_fp16")]; + tensor var_1335_cast_fp16 = mul(x = x_139_cast_fp16, y = flow_net_res_blocks_0_in_ln_weight_to_fp16)[name = tensor("op_1335_cast_fp16")]; + tensor x_141_cast_fp16 = add(x = var_1335_cast_fp16, y = flow_net_res_blocks_0_in_ln_bias_to_fp16)[name = tensor("x_141_cast_fp16")]; + tensor var_1337_promoted_to_fp16 = const()[name = tensor("op_1337_promoted_to_fp16"), val = tensor(0x1p+0)]; + tensor var_1338_cast_fp16 = add(x = var_1321_cast_fp16_1, y = var_1337_promoted_to_fp16)[name = tensor("op_1338_cast_fp16")]; + tensor var_1339_cast_fp16 = mul(x = x_141_cast_fp16, y = var_1338_cast_fp16)[name = tensor("op_1339_cast_fp16")]; + tensor input_221_cast_fp16 = add(x = var_1339_cast_fp16, y = var_1321_cast_fp16_0)[name = tensor("input_221_cast_fp16")]; + tensor linear_85_cast_fp16 = linear(bias = flow_net_res_blocks_0_mlp_0_bias_to_fp16, weight = flow_net_res_blocks_0_mlp_0_weight_to_fp16, x = input_221_cast_fp16)[name = tensor("linear_85_cast_fp16")]; + tensor input_225_cast_fp16 = silu(x = linear_85_cast_fp16)[name = tensor("input_225_cast_fp16")]; + tensor linear_86_cast_fp16 = linear(bias = flow_net_res_blocks_0_mlp_2_bias_to_fp16, weight = flow_net_res_blocks_0_mlp_2_weight_to_fp16, x = input_225_cast_fp16)[name = tensor("linear_86_cast_fp16")]; + tensor var_1350_cast_fp16 = mul(x = var_1321_cast_fp16_2, y = linear_86_cast_fp16)[name = tensor("op_1350_cast_fp16")]; + tensor x_143_cast_fp16 = add(x = linear_78_cast_fp16, y = var_1350_cast_fp16)[name = tensor("x_143_cast_fp16")]; + tensor linear_87_cast_fp16 = linear(bias = flow_net_res_blocks_1_adaLN_modulation_1_bias_to_fp16, weight = flow_net_res_blocks_1_adaLN_modulation_1_weight_to_fp16, x = input_219_cast_fp16)[name = tensor("linear_87_cast_fp16")]; + tensor var_1360_split_sizes_0 = const()[name = tensor("op_1360_split_sizes_0"), val = tensor([512, 512, 512])]; + tensor var_1360_axis_0 = const()[name = tensor("op_1360_axis_0"), val = tensor(-1)]; + tensor var_1360_cast_fp16_0, tensor var_1360_cast_fp16_1, tensor var_1360_cast_fp16_2 = split(axis = var_1360_axis_0, split_sizes = var_1360_split_sizes_0, x = linear_87_cast_fp16)[name = tensor("op_1360_cast_fp16")]; + tensor mean_45_axes_0 = const()[name = tensor("mean_45_axes_0"), val = tensor([-1])]; + tensor mean_45_keep_dims_0 = const()[name = tensor("mean_45_keep_dims_0"), val = tensor(true)]; + tensor mean_45_cast_fp16 = reduce_mean(axes = mean_45_axes_0, keep_dims = mean_45_keep_dims_0, x = x_143_cast_fp16)[name = tensor("mean_45_cast_fp16")]; + tensor sub_38_cast_fp16 = sub(x = x_143_cast_fp16, y = mean_45_cast_fp16)[name = tensor("sub_38_cast_fp16")]; + tensor square_30_cast_fp16 = square(x = sub_38_cast_fp16)[name = tensor("square_30_cast_fp16")]; + tensor reduce_mean_61_axes_0 = const()[name = tensor("reduce_mean_61_axes_0"), val = tensor([-1])]; + tensor reduce_mean_61_keep_dims_0 = const()[name = tensor("reduce_mean_61_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_61_cast_fp16 = reduce_mean(axes = reduce_mean_61_axes_0, keep_dims = reduce_mean_61_keep_dims_0, x = square_30_cast_fp16)[name = tensor("reduce_mean_61_cast_fp16")]; + tensor var_1370_to_fp16 = const()[name = tensor("op_1370_to_fp16"), val = tensor(0x1.1p-20)]; + tensor var_1371_cast_fp16 = add(x = reduce_mean_61_cast_fp16, y = var_1370_to_fp16)[name = tensor("op_1371_cast_fp16")]; + tensor var_1372_cast_fp16 = sqrt(x = var_1371_cast_fp16)[name = tensor("op_1372_cast_fp16")]; + tensor x_145_cast_fp16 = real_div(x = sub_38_cast_fp16, y = var_1372_cast_fp16)[name = tensor("x_145_cast_fp16")]; + tensor var_1374_cast_fp16 = mul(x = x_145_cast_fp16, y = flow_net_res_blocks_1_in_ln_weight_to_fp16)[name = tensor("op_1374_cast_fp16")]; + tensor x_147_cast_fp16 = add(x = var_1374_cast_fp16, y = flow_net_res_blocks_1_in_ln_bias_to_fp16)[name = tensor("x_147_cast_fp16")]; + tensor var_1376_promoted_to_fp16 = const()[name = tensor("op_1376_promoted_to_fp16"), val = tensor(0x1p+0)]; + tensor var_1377_cast_fp16 = add(x = var_1360_cast_fp16_1, y = var_1376_promoted_to_fp16)[name = tensor("op_1377_cast_fp16")]; + tensor var_1378_cast_fp16 = mul(x = x_147_cast_fp16, y = var_1377_cast_fp16)[name = tensor("op_1378_cast_fp16")]; + tensor input_229_cast_fp16 = add(x = var_1378_cast_fp16, y = var_1360_cast_fp16_0)[name = tensor("input_229_cast_fp16")]; + tensor linear_88_cast_fp16 = linear(bias = flow_net_res_blocks_1_mlp_0_bias_to_fp16, weight = flow_net_res_blocks_1_mlp_0_weight_to_fp16, x = input_229_cast_fp16)[name = tensor("linear_88_cast_fp16")]; + tensor input_233_cast_fp16 = silu(x = linear_88_cast_fp16)[name = tensor("input_233_cast_fp16")]; + tensor linear_89_cast_fp16 = linear(bias = flow_net_res_blocks_1_mlp_2_bias_to_fp16, weight = flow_net_res_blocks_1_mlp_2_weight_to_fp16, x = input_233_cast_fp16)[name = tensor("linear_89_cast_fp16")]; + tensor var_1389_cast_fp16 = mul(x = var_1360_cast_fp16_2, y = linear_89_cast_fp16)[name = tensor("op_1389_cast_fp16")]; + tensor x_149_cast_fp16 = add(x = x_143_cast_fp16, y = var_1389_cast_fp16)[name = tensor("x_149_cast_fp16")]; + tensor linear_90_cast_fp16 = linear(bias = flow_net_res_blocks_2_adaLN_modulation_1_bias_to_fp16, weight = flow_net_res_blocks_2_adaLN_modulation_1_weight_to_fp16, x = input_219_cast_fp16)[name = tensor("linear_90_cast_fp16")]; + tensor var_1399_split_sizes_0 = const()[name = tensor("op_1399_split_sizes_0"), val = tensor([512, 512, 512])]; + tensor var_1399_axis_0 = const()[name = tensor("op_1399_axis_0"), val = tensor(-1)]; + tensor var_1399_cast_fp16_0, tensor var_1399_cast_fp16_1, tensor var_1399_cast_fp16_2 = split(axis = var_1399_axis_0, split_sizes = var_1399_split_sizes_0, x = linear_90_cast_fp16)[name = tensor("op_1399_cast_fp16")]; + tensor mean_47_axes_0 = const()[name = tensor("mean_47_axes_0"), val = tensor([-1])]; + tensor mean_47_keep_dims_0 = const()[name = tensor("mean_47_keep_dims_0"), val = tensor(true)]; + tensor mean_47_cast_fp16 = reduce_mean(axes = mean_47_axes_0, keep_dims = mean_47_keep_dims_0, x = x_149_cast_fp16)[name = tensor("mean_47_cast_fp16")]; + tensor sub_39_cast_fp16 = sub(x = x_149_cast_fp16, y = mean_47_cast_fp16)[name = tensor("sub_39_cast_fp16")]; + tensor square_31_cast_fp16 = square(x = sub_39_cast_fp16)[name = tensor("square_31_cast_fp16")]; + tensor reduce_mean_63_axes_0 = const()[name = tensor("reduce_mean_63_axes_0"), val = tensor([-1])]; + tensor reduce_mean_63_keep_dims_0 = const()[name = tensor("reduce_mean_63_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_63_cast_fp16 = reduce_mean(axes = reduce_mean_63_axes_0, keep_dims = reduce_mean_63_keep_dims_0, x = square_31_cast_fp16)[name = tensor("reduce_mean_63_cast_fp16")]; + tensor var_1409_to_fp16 = const()[name = tensor("op_1409_to_fp16"), val = tensor(0x1.1p-20)]; + tensor var_1410_cast_fp16 = add(x = reduce_mean_63_cast_fp16, y = var_1409_to_fp16)[name = tensor("op_1410_cast_fp16")]; + tensor var_1411_cast_fp16 = sqrt(x = var_1410_cast_fp16)[name = tensor("op_1411_cast_fp16")]; + tensor x_151_cast_fp16 = real_div(x = sub_39_cast_fp16, y = var_1411_cast_fp16)[name = tensor("x_151_cast_fp16")]; + tensor var_1413_cast_fp16 = mul(x = x_151_cast_fp16, y = flow_net_res_blocks_2_in_ln_weight_to_fp16)[name = tensor("op_1413_cast_fp16")]; + tensor x_153_cast_fp16 = add(x = var_1413_cast_fp16, y = flow_net_res_blocks_2_in_ln_bias_to_fp16)[name = tensor("x_153_cast_fp16")]; + tensor var_1415_promoted_to_fp16 = const()[name = tensor("op_1415_promoted_to_fp16"), val = tensor(0x1p+0)]; + tensor var_1416_cast_fp16 = add(x = var_1399_cast_fp16_1, y = var_1415_promoted_to_fp16)[name = tensor("op_1416_cast_fp16")]; + tensor var_1417_cast_fp16 = mul(x = x_153_cast_fp16, y = var_1416_cast_fp16)[name = tensor("op_1417_cast_fp16")]; + tensor input_237_cast_fp16 = add(x = var_1417_cast_fp16, y = var_1399_cast_fp16_0)[name = tensor("input_237_cast_fp16")]; + tensor linear_91_cast_fp16 = linear(bias = flow_net_res_blocks_2_mlp_0_bias_to_fp16, weight = flow_net_res_blocks_2_mlp_0_weight_to_fp16, x = input_237_cast_fp16)[name = tensor("linear_91_cast_fp16")]; + tensor input_241_cast_fp16 = silu(x = linear_91_cast_fp16)[name = tensor("input_241_cast_fp16")]; + tensor linear_92_cast_fp16 = linear(bias = flow_net_res_blocks_2_mlp_2_bias_to_fp16, weight = flow_net_res_blocks_2_mlp_2_weight_to_fp16, x = input_241_cast_fp16)[name = tensor("linear_92_cast_fp16")]; + tensor var_1428_cast_fp16 = mul(x = var_1399_cast_fp16_2, y = linear_92_cast_fp16)[name = tensor("op_1428_cast_fp16")]; + tensor x_155_cast_fp16 = add(x = x_149_cast_fp16, y = var_1428_cast_fp16)[name = tensor("x_155_cast_fp16")]; + tensor linear_93_cast_fp16 = linear(bias = flow_net_res_blocks_3_adaLN_modulation_1_bias_to_fp16, weight = flow_net_res_blocks_3_adaLN_modulation_1_weight_to_fp16, x = input_219_cast_fp16)[name = tensor("linear_93_cast_fp16")]; + tensor var_1438_split_sizes_0 = const()[name = tensor("op_1438_split_sizes_0"), val = tensor([512, 512, 512])]; + tensor var_1438_axis_0 = const()[name = tensor("op_1438_axis_0"), val = tensor(-1)]; + tensor var_1438_cast_fp16_0, tensor var_1438_cast_fp16_1, tensor var_1438_cast_fp16_2 = split(axis = var_1438_axis_0, split_sizes = var_1438_split_sizes_0, x = linear_93_cast_fp16)[name = tensor("op_1438_cast_fp16")]; + tensor mean_49_axes_0 = const()[name = tensor("mean_49_axes_0"), val = tensor([-1])]; + tensor mean_49_keep_dims_0 = const()[name = tensor("mean_49_keep_dims_0"), val = tensor(true)]; + tensor mean_49_cast_fp16 = reduce_mean(axes = mean_49_axes_0, keep_dims = mean_49_keep_dims_0, x = x_155_cast_fp16)[name = tensor("mean_49_cast_fp16")]; + tensor sub_40_cast_fp16 = sub(x = x_155_cast_fp16, y = mean_49_cast_fp16)[name = tensor("sub_40_cast_fp16")]; + tensor square_32_cast_fp16 = square(x = sub_40_cast_fp16)[name = tensor("square_32_cast_fp16")]; + tensor reduce_mean_65_axes_0 = const()[name = tensor("reduce_mean_65_axes_0"), val = tensor([-1])]; + tensor reduce_mean_65_keep_dims_0 = const()[name = tensor("reduce_mean_65_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_65_cast_fp16 = reduce_mean(axes = reduce_mean_65_axes_0, keep_dims = reduce_mean_65_keep_dims_0, x = square_32_cast_fp16)[name = tensor("reduce_mean_65_cast_fp16")]; + tensor var_1448_to_fp16 = const()[name = tensor("op_1448_to_fp16"), val = tensor(0x1.1p-20)]; + tensor var_1449_cast_fp16 = add(x = reduce_mean_65_cast_fp16, y = var_1448_to_fp16)[name = tensor("op_1449_cast_fp16")]; + tensor var_1450_cast_fp16 = sqrt(x = var_1449_cast_fp16)[name = tensor("op_1450_cast_fp16")]; + tensor x_157_cast_fp16 = real_div(x = sub_40_cast_fp16, y = var_1450_cast_fp16)[name = tensor("x_157_cast_fp16")]; + tensor var_1452_cast_fp16 = mul(x = x_157_cast_fp16, y = flow_net_res_blocks_3_in_ln_weight_to_fp16)[name = tensor("op_1452_cast_fp16")]; + tensor x_159_cast_fp16 = add(x = var_1452_cast_fp16, y = flow_net_res_blocks_3_in_ln_bias_to_fp16)[name = tensor("x_159_cast_fp16")]; + tensor var_1454_promoted_to_fp16 = const()[name = tensor("op_1454_promoted_to_fp16"), val = tensor(0x1p+0)]; + tensor var_1455_cast_fp16 = add(x = var_1438_cast_fp16_1, y = var_1454_promoted_to_fp16)[name = tensor("op_1455_cast_fp16")]; + tensor var_1456_cast_fp16 = mul(x = x_159_cast_fp16, y = var_1455_cast_fp16)[name = tensor("op_1456_cast_fp16")]; + tensor input_245_cast_fp16 = add(x = var_1456_cast_fp16, y = var_1438_cast_fp16_0)[name = tensor("input_245_cast_fp16")]; + tensor linear_94_cast_fp16 = linear(bias = flow_net_res_blocks_3_mlp_0_bias_to_fp16, weight = flow_net_res_blocks_3_mlp_0_weight_to_fp16, x = input_245_cast_fp16)[name = tensor("linear_94_cast_fp16")]; + tensor input_249_cast_fp16 = silu(x = linear_94_cast_fp16)[name = tensor("input_249_cast_fp16")]; + tensor linear_95_cast_fp16 = linear(bias = flow_net_res_blocks_3_mlp_2_bias_to_fp16, weight = flow_net_res_blocks_3_mlp_2_weight_to_fp16, x = input_249_cast_fp16)[name = tensor("linear_95_cast_fp16")]; + tensor var_1467_cast_fp16 = mul(x = var_1438_cast_fp16_2, y = linear_95_cast_fp16)[name = tensor("op_1467_cast_fp16")]; + tensor x_161_cast_fp16 = add(x = x_155_cast_fp16, y = var_1467_cast_fp16)[name = tensor("x_161_cast_fp16")]; + tensor linear_96_cast_fp16 = linear(bias = flow_net_res_blocks_4_adaLN_modulation_1_bias_to_fp16, weight = flow_net_res_blocks_4_adaLN_modulation_1_weight_to_fp16, x = input_219_cast_fp16)[name = tensor("linear_96_cast_fp16")]; + tensor var_1477_split_sizes_0 = const()[name = tensor("op_1477_split_sizes_0"), val = tensor([512, 512, 512])]; + tensor var_1477_axis_0 = const()[name = tensor("op_1477_axis_0"), val = tensor(-1)]; + tensor var_1477_cast_fp16_0, tensor var_1477_cast_fp16_1, tensor var_1477_cast_fp16_2 = split(axis = var_1477_axis_0, split_sizes = var_1477_split_sizes_0, x = linear_96_cast_fp16)[name = tensor("op_1477_cast_fp16")]; + tensor mean_51_axes_0 = const()[name = tensor("mean_51_axes_0"), val = tensor([-1])]; + tensor mean_51_keep_dims_0 = const()[name = tensor("mean_51_keep_dims_0"), val = tensor(true)]; + tensor mean_51_cast_fp16 = reduce_mean(axes = mean_51_axes_0, keep_dims = mean_51_keep_dims_0, x = x_161_cast_fp16)[name = tensor("mean_51_cast_fp16")]; + tensor sub_41_cast_fp16 = sub(x = x_161_cast_fp16, y = mean_51_cast_fp16)[name = tensor("sub_41_cast_fp16")]; + tensor square_33_cast_fp16 = square(x = sub_41_cast_fp16)[name = tensor("square_33_cast_fp16")]; + tensor reduce_mean_67_axes_0 = const()[name = tensor("reduce_mean_67_axes_0"), val = tensor([-1])]; + tensor reduce_mean_67_keep_dims_0 = const()[name = tensor("reduce_mean_67_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_67_cast_fp16 = reduce_mean(axes = reduce_mean_67_axes_0, keep_dims = reduce_mean_67_keep_dims_0, x = square_33_cast_fp16)[name = tensor("reduce_mean_67_cast_fp16")]; + tensor var_1487_to_fp16 = const()[name = tensor("op_1487_to_fp16"), val = tensor(0x1.1p-20)]; + tensor var_1488_cast_fp16 = add(x = reduce_mean_67_cast_fp16, y = var_1487_to_fp16)[name = tensor("op_1488_cast_fp16")]; + tensor var_1489_cast_fp16 = sqrt(x = var_1488_cast_fp16)[name = tensor("op_1489_cast_fp16")]; + tensor x_163_cast_fp16 = real_div(x = sub_41_cast_fp16, y = var_1489_cast_fp16)[name = tensor("x_163_cast_fp16")]; + tensor var_1491_cast_fp16 = mul(x = x_163_cast_fp16, y = flow_net_res_blocks_4_in_ln_weight_to_fp16)[name = tensor("op_1491_cast_fp16")]; + tensor x_165_cast_fp16 = add(x = var_1491_cast_fp16, y = flow_net_res_blocks_4_in_ln_bias_to_fp16)[name = tensor("x_165_cast_fp16")]; + tensor var_1493_promoted_to_fp16 = const()[name = tensor("op_1493_promoted_to_fp16"), val = tensor(0x1p+0)]; + tensor var_1494_cast_fp16 = add(x = var_1477_cast_fp16_1, y = var_1493_promoted_to_fp16)[name = tensor("op_1494_cast_fp16")]; + tensor var_1495_cast_fp16 = mul(x = x_165_cast_fp16, y = var_1494_cast_fp16)[name = tensor("op_1495_cast_fp16")]; + tensor input_253_cast_fp16 = add(x = var_1495_cast_fp16, y = var_1477_cast_fp16_0)[name = tensor("input_253_cast_fp16")]; + tensor linear_97_cast_fp16 = linear(bias = flow_net_res_blocks_4_mlp_0_bias_to_fp16, weight = flow_net_res_blocks_4_mlp_0_weight_to_fp16, x = input_253_cast_fp16)[name = tensor("linear_97_cast_fp16")]; + tensor input_257_cast_fp16 = silu(x = linear_97_cast_fp16)[name = tensor("input_257_cast_fp16")]; + tensor linear_98_cast_fp16 = linear(bias = flow_net_res_blocks_4_mlp_2_bias_to_fp16, weight = flow_net_res_blocks_4_mlp_2_weight_to_fp16, x = input_257_cast_fp16)[name = tensor("linear_98_cast_fp16")]; + tensor var_1506_cast_fp16 = mul(x = var_1477_cast_fp16_2, y = linear_98_cast_fp16)[name = tensor("op_1506_cast_fp16")]; + tensor x_167_cast_fp16 = add(x = x_161_cast_fp16, y = var_1506_cast_fp16)[name = tensor("x_167_cast_fp16")]; + tensor linear_99_cast_fp16 = linear(bias = flow_net_res_blocks_5_adaLN_modulation_1_bias_to_fp16, weight = flow_net_res_blocks_5_adaLN_modulation_1_weight_to_fp16, x = input_219_cast_fp16)[name = tensor("linear_99_cast_fp16")]; + tensor var_1516_split_sizes_0 = const()[name = tensor("op_1516_split_sizes_0"), val = tensor([512, 512, 512])]; + tensor var_1516_axis_0 = const()[name = tensor("op_1516_axis_0"), val = tensor(-1)]; + tensor var_1516_cast_fp16_0, tensor var_1516_cast_fp16_1, tensor var_1516_cast_fp16_2 = split(axis = var_1516_axis_0, split_sizes = var_1516_split_sizes_0, x = linear_99_cast_fp16)[name = tensor("op_1516_cast_fp16")]; + tensor mean_53_axes_0 = const()[name = tensor("mean_53_axes_0"), val = tensor([-1])]; + tensor mean_53_keep_dims_0 = const()[name = tensor("mean_53_keep_dims_0"), val = tensor(true)]; + tensor mean_53_cast_fp16 = reduce_mean(axes = mean_53_axes_0, keep_dims = mean_53_keep_dims_0, x = x_167_cast_fp16)[name = tensor("mean_53_cast_fp16")]; + tensor sub_42_cast_fp16 = sub(x = x_167_cast_fp16, y = mean_53_cast_fp16)[name = tensor("sub_42_cast_fp16")]; + tensor square_34_cast_fp16 = square(x = sub_42_cast_fp16)[name = tensor("square_34_cast_fp16")]; + tensor reduce_mean_69_axes_0 = const()[name = tensor("reduce_mean_69_axes_0"), val = tensor([-1])]; + tensor reduce_mean_69_keep_dims_0 = const()[name = tensor("reduce_mean_69_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_69_cast_fp16 = reduce_mean(axes = reduce_mean_69_axes_0, keep_dims = reduce_mean_69_keep_dims_0, x = square_34_cast_fp16)[name = tensor("reduce_mean_69_cast_fp16")]; + tensor var_1526_to_fp16 = const()[name = tensor("op_1526_to_fp16"), val = tensor(0x1.1p-20)]; + tensor var_1527_cast_fp16 = add(x = reduce_mean_69_cast_fp16, y = var_1526_to_fp16)[name = tensor("op_1527_cast_fp16")]; + tensor var_1528_cast_fp16 = sqrt(x = var_1527_cast_fp16)[name = tensor("op_1528_cast_fp16")]; + tensor x_169_cast_fp16 = real_div(x = sub_42_cast_fp16, y = var_1528_cast_fp16)[name = tensor("x_169_cast_fp16")]; + tensor var_1530_cast_fp16 = mul(x = x_169_cast_fp16, y = flow_net_res_blocks_5_in_ln_weight_to_fp16)[name = tensor("op_1530_cast_fp16")]; + tensor x_171_cast_fp16 = add(x = var_1530_cast_fp16, y = flow_net_res_blocks_5_in_ln_bias_to_fp16)[name = tensor("x_171_cast_fp16")]; + tensor var_1532_promoted_to_fp16 = const()[name = tensor("op_1532_promoted_to_fp16"), val = tensor(0x1p+0)]; + tensor var_1533_cast_fp16 = add(x = var_1516_cast_fp16_1, y = var_1532_promoted_to_fp16)[name = tensor("op_1533_cast_fp16")]; + tensor var_1534_cast_fp16 = mul(x = x_171_cast_fp16, y = var_1533_cast_fp16)[name = tensor("op_1534_cast_fp16")]; + tensor input_261_cast_fp16 = add(x = var_1534_cast_fp16, y = var_1516_cast_fp16_0)[name = tensor("input_261_cast_fp16")]; + tensor linear_100_cast_fp16 = linear(bias = flow_net_res_blocks_5_mlp_0_bias_to_fp16, weight = flow_net_res_blocks_5_mlp_0_weight_to_fp16, x = input_261_cast_fp16)[name = tensor("linear_100_cast_fp16")]; + tensor input_265_cast_fp16 = silu(x = linear_100_cast_fp16)[name = tensor("input_265_cast_fp16")]; + tensor linear_101_cast_fp16 = linear(bias = flow_net_res_blocks_5_mlp_2_bias_to_fp16, weight = flow_net_res_blocks_5_mlp_2_weight_to_fp16, x = input_265_cast_fp16)[name = tensor("linear_101_cast_fp16")]; + tensor var_1545_cast_fp16 = mul(x = var_1516_cast_fp16_2, y = linear_101_cast_fp16)[name = tensor("op_1545_cast_fp16")]; + tensor x_173_cast_fp16 = add(x = x_167_cast_fp16, y = var_1545_cast_fp16)[name = tensor("x_173_cast_fp16")]; + tensor linear_102_cast_fp16 = linear(bias = flow_net_final_layer_adaLN_modulation_1_bias_to_fp16, weight = flow_net_final_layer_adaLN_modulation_1_weight_to_fp16, x = input_219_cast_fp16)[name = tensor("linear_102_cast_fp16")]; + tensor var_1554_split_sizes_0 = const()[name = tensor("op_1554_split_sizes_0"), val = tensor([512, 512])]; + tensor var_1554_axis_0 = const()[name = tensor("op_1554_axis_0"), val = tensor(-1)]; + tensor var_1554_cast_fp16_0, tensor var_1554_cast_fp16_1 = split(axis = var_1554_axis_0, split_sizes = var_1554_split_sizes_0, x = linear_102_cast_fp16)[name = tensor("op_1554_cast_fp16")]; + tensor mean_55_axes_0 = const()[name = tensor("mean_55_axes_0"), val = tensor([-1])]; + tensor mean_55_keep_dims_0 = const()[name = tensor("mean_55_keep_dims_0"), val = tensor(true)]; + tensor mean_55_cast_fp16 = reduce_mean(axes = mean_55_axes_0, keep_dims = mean_55_keep_dims_0, x = x_173_cast_fp16)[name = tensor("mean_55_cast_fp16")]; + tensor sub_43_cast_fp16 = sub(x = x_173_cast_fp16, y = mean_55_cast_fp16)[name = tensor("sub_43_cast_fp16")]; + tensor square_35_cast_fp16 = square(x = sub_43_cast_fp16)[name = tensor("square_35_cast_fp16")]; + tensor reduce_mean_71_axes_0 = const()[name = tensor("reduce_mean_71_axes_0"), val = tensor([-1])]; + tensor reduce_mean_71_keep_dims_0 = const()[name = tensor("reduce_mean_71_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_71_cast_fp16 = reduce_mean(axes = reduce_mean_71_axes_0, keep_dims = reduce_mean_71_keep_dims_0, x = square_35_cast_fp16)[name = tensor("reduce_mean_71_cast_fp16")]; + tensor var_1561_to_fp16 = const()[name = tensor("op_1561_to_fp16"), val = tensor(0x1.1p-20)]; + tensor var_1562_cast_fp16 = add(x = reduce_mean_71_cast_fp16, y = var_1561_to_fp16)[name = tensor("op_1562_cast_fp16")]; + tensor var_1563_cast_fp16 = sqrt(x = var_1562_cast_fp16)[name = tensor("op_1563_cast_fp16")]; + tensor x_175_cast_fp16 = real_div(x = sub_43_cast_fp16, y = var_1563_cast_fp16)[name = tensor("x_175_cast_fp16")]; + tensor var_1565_promoted_to_fp16 = const()[name = tensor("op_1565_promoted_to_fp16"), val = tensor(0x1p+0)]; + tensor var_1566_cast_fp16 = add(x = var_1554_cast_fp16_1, y = var_1565_promoted_to_fp16)[name = tensor("op_1566_cast_fp16")]; + tensor var_1567_cast_fp16 = mul(x = x_175_cast_fp16, y = var_1566_cast_fp16)[name = tensor("op_1567_cast_fp16")]; + tensor input_269_cast_fp16 = add(x = var_1567_cast_fp16, y = var_1554_cast_fp16_0)[name = tensor("input_269_cast_fp16")]; + tensor linear_103_cast_fp16 = linear(bias = flow_net_final_layer_linear_bias_to_fp16, weight = flow_net_final_layer_linear_weight_to_fp16, x = input_269_cast_fp16)[name = tensor("linear_103_cast_fp16")]; + tensor var_1578_to_fp16 = const()[name = tensor("op_1578_to_fp16"), val = tensor(0x1p-3)]; + tensor var_1579_cast_fp16 = mul(x = linear_103_cast_fp16, y = var_1578_to_fp16)[name = tensor("op_1579_cast_fp16")]; + tensor input_271_cast_fp16 = add(x = input_203_cast_fp16, y = var_1579_cast_fp16)[name = tensor("input_271_cast_fp16")]; + tensor linear_104_cast_fp16 = linear(bias = flow_net_input_proj_bias_to_fp16, weight = flow_net_input_proj_weight_to_fp16, x = input_271_cast_fp16)[name = tensor("linear_104_cast_fp16")]; + tensor input_275_to_fp16 = const()[name = tensor("input_275_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(19003904)))]; + tensor input_277_cast_fp16 = silu(x = input_275_to_fp16)[name = tensor("input_277_cast_fp16")]; + tensor linear_106_cast_fp16 = linear(bias = flow_net_time_embed_0_mlp_2_bias_to_fp16, weight = flow_net_time_embed_0_mlp_2_weight_to_fp16, x = input_277_cast_fp16)[name = tensor("linear_106_cast_fp16")]; + tensor reduce_mean_72_axes_0 = const()[name = tensor("reduce_mean_72_axes_0"), val = tensor([-1])]; + tensor reduce_mean_72_keep_dims_0 = const()[name = tensor("reduce_mean_72_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_72_cast_fp16 = reduce_mean(axes = reduce_mean_72_axes_0, keep_dims = reduce_mean_72_keep_dims_0, x = linear_106_cast_fp16)[name = tensor("reduce_mean_72_cast_fp16")]; + tensor sub_44_cast_fp16 = sub(x = linear_106_cast_fp16, y = reduce_mean_72_cast_fp16)[name = tensor("sub_44_cast_fp16")]; + tensor square_36_cast_fp16 = square(x = sub_44_cast_fp16)[name = tensor("square_36_cast_fp16")]; + tensor reduce_mean_73_axes_0 = const()[name = tensor("reduce_mean_73_axes_0"), val = tensor([-1])]; + tensor reduce_mean_73_keep_dims_0 = const()[name = tensor("reduce_mean_73_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_73_cast_fp16 = reduce_mean(axes = reduce_mean_73_axes_0, keep_dims = reduce_mean_73_keep_dims_0, x = square_36_cast_fp16)[name = tensor("reduce_mean_73_cast_fp16")]; + tensor real_div_8_to_fp16 = const()[name = tensor("real_div_8_to_fp16"), val = tensor(0x1.008p+0)]; + tensor mul_8_cast_fp16 = mul(x = reduce_mean_73_cast_fp16, y = real_div_8_to_fp16)[name = tensor("mul_8_cast_fp16")]; + tensor var_1645_to_fp16 = const()[name = tensor("op_1645_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_73_cast_fp16_0 = add(x = mul_8_cast_fp16, y = var_1645_to_fp16)[name = tensor("var_73_cast_fp16")]; + tensor var_1648_epsilon_0 = const()[name = tensor("op_1648_epsilon_0"), val = tensor(0x1.197998p-40)]; + tensor var_1648_cast_fp16 = rsqrt(epsilon = var_1648_epsilon_0, x = var_73_cast_fp16_0)[name = tensor("op_1648_cast_fp16")]; + tensor var_1649_cast_fp16 = mul(x = const_3_to_fp16, y = var_1648_cast_fp16)[name = tensor("op_1649_cast_fp16")]; + tensor var_1650_cast_fp16 = mul(x = linear_106_cast_fp16, y = var_1649_cast_fp16)[name = tensor("op_1650_cast_fp16")]; + tensor input_281_to_fp16 = const()[name = tensor("input_281_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(19004992)))]; + tensor input_283_cast_fp16 = silu(x = input_281_to_fp16)[name = tensor("input_283_cast_fp16")]; + tensor linear_108_cast_fp16 = linear(bias = flow_net_time_embed_1_mlp_2_bias_to_fp16, weight = flow_net_time_embed_1_mlp_2_weight_to_fp16, x = input_283_cast_fp16)[name = tensor("linear_108_cast_fp16")]; + tensor reduce_mean_74_axes_0 = const()[name = tensor("reduce_mean_74_axes_0"), val = tensor([-1])]; + tensor reduce_mean_74_keep_dims_0 = const()[name = tensor("reduce_mean_74_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_74_cast_fp16 = reduce_mean(axes = reduce_mean_74_axes_0, keep_dims = reduce_mean_74_keep_dims_0, x = linear_108_cast_fp16)[name = tensor("reduce_mean_74_cast_fp16")]; + tensor sub_46_cast_fp16 = sub(x = linear_108_cast_fp16, y = reduce_mean_74_cast_fp16)[name = tensor("sub_46_cast_fp16")]; + tensor square_37_cast_fp16 = square(x = sub_46_cast_fp16)[name = tensor("square_37_cast_fp16")]; + tensor reduce_mean_75_axes_0 = const()[name = tensor("reduce_mean_75_axes_0"), val = tensor([-1])]; + tensor reduce_mean_75_keep_dims_0 = const()[name = tensor("reduce_mean_75_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_75_cast_fp16 = reduce_mean(axes = reduce_mean_75_axes_0, keep_dims = reduce_mean_75_keep_dims_0, x = square_37_cast_fp16)[name = tensor("reduce_mean_75_cast_fp16")]; + tensor real_div_9_to_fp16 = const()[name = tensor("real_div_9_to_fp16"), val = tensor(0x1.008p+0)]; + tensor mul_9_cast_fp16 = mul(x = reduce_mean_75_cast_fp16, y = real_div_9_to_fp16)[name = tensor("mul_9_cast_fp16")]; + tensor var_1682_to_fp16 = const()[name = tensor("op_1682_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_75_cast_fp16_0 = add(x = mul_9_cast_fp16, y = var_1682_to_fp16)[name = tensor("var_75_cast_fp16")]; + tensor var_1685_epsilon_0 = const()[name = tensor("op_1685_epsilon_0"), val = tensor(0x1.197998p-40)]; + tensor var_1685_cast_fp16 = rsqrt(epsilon = var_1685_epsilon_0, x = var_75_cast_fp16_0)[name = tensor("op_1685_cast_fp16")]; + tensor var_1686_cast_fp16 = mul(x = const_5_to_fp16, y = var_1685_cast_fp16)[name = tensor("op_1686_cast_fp16")]; + tensor var_1687_cast_fp16 = mul(x = linear_108_cast_fp16, y = var_1686_cast_fp16)[name = tensor("op_1687_cast_fp16")]; + tensor var_1699_cast_fp16 = add(x = var_1650_cast_fp16, y = var_1687_cast_fp16)[name = tensor("op_1699_cast_fp16")]; + tensor _inversed_t_combined_9_y_0_to_fp16 = const()[name = tensor("_inversed_t_combined_9_y_0_to_fp16"), val = tensor(0x1p-1)]; + tensor _inversed_t_combined_9_cast_fp16 = mul(x = var_1699_cast_fp16, y = _inversed_t_combined_9_y_0_to_fp16)[name = tensor("_inversed_t_combined_9_cast_fp16")]; + tensor input_285_cast_fp16 = add(x = _inversed_t_combined_9_cast_fp16, y = linear_5_cast_fp16)[name = tensor("input_285_cast_fp16")]; + tensor input_287_cast_fp16 = silu(x = input_285_cast_fp16)[name = tensor("input_287_cast_fp16")]; + tensor linear_110_cast_fp16 = linear(bias = flow_net_res_blocks_0_adaLN_modulation_1_bias_to_fp16, weight = flow_net_res_blocks_0_adaLN_modulation_1_weight_to_fp16, x = input_287_cast_fp16)[name = tensor("linear_110_cast_fp16")]; + tensor var_1714_split_sizes_0 = const()[name = tensor("op_1714_split_sizes_0"), val = tensor([512, 512, 512])]; + tensor var_1714_axis_0 = const()[name = tensor("op_1714_axis_0"), val = tensor(-1)]; + tensor var_1714_cast_fp16_0, tensor var_1714_cast_fp16_1, tensor var_1714_cast_fp16_2 = split(axis = var_1714_axis_0, split_sizes = var_1714_split_sizes_0, x = linear_110_cast_fp16)[name = tensor("op_1714_cast_fp16")]; + tensor mean_57_axes_0 = const()[name = tensor("mean_57_axes_0"), val = tensor([-1])]; + tensor mean_57_keep_dims_0 = const()[name = tensor("mean_57_keep_dims_0"), val = tensor(true)]; + tensor mean_57_cast_fp16 = reduce_mean(axes = mean_57_axes_0, keep_dims = mean_57_keep_dims_0, x = linear_104_cast_fp16)[name = tensor("mean_57_cast_fp16")]; + tensor sub_48_cast_fp16 = sub(x = linear_104_cast_fp16, y = mean_57_cast_fp16)[name = tensor("sub_48_cast_fp16")]; + tensor square_38_cast_fp16 = square(x = sub_48_cast_fp16)[name = tensor("square_38_cast_fp16")]; + tensor reduce_mean_77_axes_0 = const()[name = tensor("reduce_mean_77_axes_0"), val = tensor([-1])]; + tensor reduce_mean_77_keep_dims_0 = const()[name = tensor("reduce_mean_77_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_77_cast_fp16 = reduce_mean(axes = reduce_mean_77_axes_0, keep_dims = reduce_mean_77_keep_dims_0, x = square_38_cast_fp16)[name = tensor("reduce_mean_77_cast_fp16")]; + tensor var_1724_to_fp16 = const()[name = tensor("op_1724_to_fp16"), val = tensor(0x1.1p-20)]; + tensor var_1725_cast_fp16 = add(x = reduce_mean_77_cast_fp16, y = var_1724_to_fp16)[name = tensor("op_1725_cast_fp16")]; + tensor var_1726_cast_fp16 = sqrt(x = var_1725_cast_fp16)[name = tensor("op_1726_cast_fp16")]; + tensor x_183_cast_fp16 = real_div(x = sub_48_cast_fp16, y = var_1726_cast_fp16)[name = tensor("x_183_cast_fp16")]; + tensor var_1728_cast_fp16 = mul(x = x_183_cast_fp16, y = flow_net_res_blocks_0_in_ln_weight_to_fp16)[name = tensor("op_1728_cast_fp16")]; + tensor x_185_cast_fp16 = add(x = var_1728_cast_fp16, y = flow_net_res_blocks_0_in_ln_bias_to_fp16)[name = tensor("x_185_cast_fp16")]; + tensor var_1730_promoted_to_fp16 = const()[name = tensor("op_1730_promoted_to_fp16"), val = tensor(0x1p+0)]; + tensor var_1731_cast_fp16 = add(x = var_1714_cast_fp16_1, y = var_1730_promoted_to_fp16)[name = tensor("op_1731_cast_fp16")]; + tensor var_1732_cast_fp16 = mul(x = x_185_cast_fp16, y = var_1731_cast_fp16)[name = tensor("op_1732_cast_fp16")]; + tensor input_289_cast_fp16 = add(x = var_1732_cast_fp16, y = var_1714_cast_fp16_0)[name = tensor("input_289_cast_fp16")]; + tensor linear_111_cast_fp16 = linear(bias = flow_net_res_blocks_0_mlp_0_bias_to_fp16, weight = flow_net_res_blocks_0_mlp_0_weight_to_fp16, x = input_289_cast_fp16)[name = tensor("linear_111_cast_fp16")]; + tensor input_293_cast_fp16 = silu(x = linear_111_cast_fp16)[name = tensor("input_293_cast_fp16")]; + tensor linear_112_cast_fp16 = linear(bias = flow_net_res_blocks_0_mlp_2_bias_to_fp16, weight = flow_net_res_blocks_0_mlp_2_weight_to_fp16, x = input_293_cast_fp16)[name = tensor("linear_112_cast_fp16")]; + tensor var_1743_cast_fp16 = mul(x = var_1714_cast_fp16_2, y = linear_112_cast_fp16)[name = tensor("op_1743_cast_fp16")]; + tensor x_187_cast_fp16 = add(x = linear_104_cast_fp16, y = var_1743_cast_fp16)[name = tensor("x_187_cast_fp16")]; + tensor linear_113_cast_fp16 = linear(bias = flow_net_res_blocks_1_adaLN_modulation_1_bias_to_fp16, weight = flow_net_res_blocks_1_adaLN_modulation_1_weight_to_fp16, x = input_287_cast_fp16)[name = tensor("linear_113_cast_fp16")]; + tensor var_1753_split_sizes_0 = const()[name = tensor("op_1753_split_sizes_0"), val = tensor([512, 512, 512])]; + tensor var_1753_axis_0 = const()[name = tensor("op_1753_axis_0"), val = tensor(-1)]; + tensor var_1753_cast_fp16_0, tensor var_1753_cast_fp16_1, tensor var_1753_cast_fp16_2 = split(axis = var_1753_axis_0, split_sizes = var_1753_split_sizes_0, x = linear_113_cast_fp16)[name = tensor("op_1753_cast_fp16")]; + tensor mean_59_axes_0 = const()[name = tensor("mean_59_axes_0"), val = tensor([-1])]; + tensor mean_59_keep_dims_0 = const()[name = tensor("mean_59_keep_dims_0"), val = tensor(true)]; + tensor mean_59_cast_fp16 = reduce_mean(axes = mean_59_axes_0, keep_dims = mean_59_keep_dims_0, x = x_187_cast_fp16)[name = tensor("mean_59_cast_fp16")]; + tensor sub_49_cast_fp16 = sub(x = x_187_cast_fp16, y = mean_59_cast_fp16)[name = tensor("sub_49_cast_fp16")]; + tensor square_39_cast_fp16 = square(x = sub_49_cast_fp16)[name = tensor("square_39_cast_fp16")]; + tensor reduce_mean_79_axes_0 = const()[name = tensor("reduce_mean_79_axes_0"), val = tensor([-1])]; + tensor reduce_mean_79_keep_dims_0 = const()[name = tensor("reduce_mean_79_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_79_cast_fp16 = reduce_mean(axes = reduce_mean_79_axes_0, keep_dims = reduce_mean_79_keep_dims_0, x = square_39_cast_fp16)[name = tensor("reduce_mean_79_cast_fp16")]; + tensor var_1763_to_fp16 = const()[name = tensor("op_1763_to_fp16"), val = tensor(0x1.1p-20)]; + tensor var_1764_cast_fp16 = add(x = reduce_mean_79_cast_fp16, y = var_1763_to_fp16)[name = tensor("op_1764_cast_fp16")]; + tensor var_1765_cast_fp16 = sqrt(x = var_1764_cast_fp16)[name = tensor("op_1765_cast_fp16")]; + tensor x_189_cast_fp16 = real_div(x = sub_49_cast_fp16, y = var_1765_cast_fp16)[name = tensor("x_189_cast_fp16")]; + tensor var_1767_cast_fp16 = mul(x = x_189_cast_fp16, y = flow_net_res_blocks_1_in_ln_weight_to_fp16)[name = tensor("op_1767_cast_fp16")]; + tensor x_191_cast_fp16 = add(x = var_1767_cast_fp16, y = flow_net_res_blocks_1_in_ln_bias_to_fp16)[name = tensor("x_191_cast_fp16")]; + tensor var_1769_promoted_to_fp16 = const()[name = tensor("op_1769_promoted_to_fp16"), val = tensor(0x1p+0)]; + tensor var_1770_cast_fp16 = add(x = var_1753_cast_fp16_1, y = var_1769_promoted_to_fp16)[name = tensor("op_1770_cast_fp16")]; + tensor var_1771_cast_fp16 = mul(x = x_191_cast_fp16, y = var_1770_cast_fp16)[name = tensor("op_1771_cast_fp16")]; + tensor input_297_cast_fp16 = add(x = var_1771_cast_fp16, y = var_1753_cast_fp16_0)[name = tensor("input_297_cast_fp16")]; + tensor linear_114_cast_fp16 = linear(bias = flow_net_res_blocks_1_mlp_0_bias_to_fp16, weight = flow_net_res_blocks_1_mlp_0_weight_to_fp16, x = input_297_cast_fp16)[name = tensor("linear_114_cast_fp16")]; + tensor input_301_cast_fp16 = silu(x = linear_114_cast_fp16)[name = tensor("input_301_cast_fp16")]; + tensor linear_115_cast_fp16 = linear(bias = flow_net_res_blocks_1_mlp_2_bias_to_fp16, weight = flow_net_res_blocks_1_mlp_2_weight_to_fp16, x = input_301_cast_fp16)[name = tensor("linear_115_cast_fp16")]; + tensor var_1782_cast_fp16 = mul(x = var_1753_cast_fp16_2, y = linear_115_cast_fp16)[name = tensor("op_1782_cast_fp16")]; + tensor x_193_cast_fp16 = add(x = x_187_cast_fp16, y = var_1782_cast_fp16)[name = tensor("x_193_cast_fp16")]; + tensor linear_116_cast_fp16 = linear(bias = flow_net_res_blocks_2_adaLN_modulation_1_bias_to_fp16, weight = flow_net_res_blocks_2_adaLN_modulation_1_weight_to_fp16, x = input_287_cast_fp16)[name = tensor("linear_116_cast_fp16")]; + tensor var_1792_split_sizes_0 = const()[name = tensor("op_1792_split_sizes_0"), val = tensor([512, 512, 512])]; + tensor var_1792_axis_0 = const()[name = tensor("op_1792_axis_0"), val = tensor(-1)]; + tensor var_1792_cast_fp16_0, tensor var_1792_cast_fp16_1, tensor var_1792_cast_fp16_2 = split(axis = var_1792_axis_0, split_sizes = var_1792_split_sizes_0, x = linear_116_cast_fp16)[name = tensor("op_1792_cast_fp16")]; + tensor mean_61_axes_0 = const()[name = tensor("mean_61_axes_0"), val = tensor([-1])]; + tensor mean_61_keep_dims_0 = const()[name = tensor("mean_61_keep_dims_0"), val = tensor(true)]; + tensor mean_61_cast_fp16 = reduce_mean(axes = mean_61_axes_0, keep_dims = mean_61_keep_dims_0, x = x_193_cast_fp16)[name = tensor("mean_61_cast_fp16")]; + tensor sub_50_cast_fp16 = sub(x = x_193_cast_fp16, y = mean_61_cast_fp16)[name = tensor("sub_50_cast_fp16")]; + tensor square_40_cast_fp16 = square(x = sub_50_cast_fp16)[name = tensor("square_40_cast_fp16")]; + tensor reduce_mean_81_axes_0 = const()[name = tensor("reduce_mean_81_axes_0"), val = tensor([-1])]; + tensor reduce_mean_81_keep_dims_0 = const()[name = tensor("reduce_mean_81_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_81_cast_fp16 = reduce_mean(axes = reduce_mean_81_axes_0, keep_dims = reduce_mean_81_keep_dims_0, x = square_40_cast_fp16)[name = tensor("reduce_mean_81_cast_fp16")]; + tensor var_1802_to_fp16 = const()[name = tensor("op_1802_to_fp16"), val = tensor(0x1.1p-20)]; + tensor var_1803_cast_fp16 = add(x = reduce_mean_81_cast_fp16, y = var_1802_to_fp16)[name = tensor("op_1803_cast_fp16")]; + tensor var_1804_cast_fp16 = sqrt(x = var_1803_cast_fp16)[name = tensor("op_1804_cast_fp16")]; + tensor x_195_cast_fp16 = real_div(x = sub_50_cast_fp16, y = var_1804_cast_fp16)[name = tensor("x_195_cast_fp16")]; + tensor var_1806_cast_fp16 = mul(x = x_195_cast_fp16, y = flow_net_res_blocks_2_in_ln_weight_to_fp16)[name = tensor("op_1806_cast_fp16")]; + tensor x_197_cast_fp16 = add(x = var_1806_cast_fp16, y = flow_net_res_blocks_2_in_ln_bias_to_fp16)[name = tensor("x_197_cast_fp16")]; + tensor var_1808_promoted_to_fp16 = const()[name = tensor("op_1808_promoted_to_fp16"), val = tensor(0x1p+0)]; + tensor var_1809_cast_fp16 = add(x = var_1792_cast_fp16_1, y = var_1808_promoted_to_fp16)[name = tensor("op_1809_cast_fp16")]; + tensor var_1810_cast_fp16 = mul(x = x_197_cast_fp16, y = var_1809_cast_fp16)[name = tensor("op_1810_cast_fp16")]; + tensor input_305_cast_fp16 = add(x = var_1810_cast_fp16, y = var_1792_cast_fp16_0)[name = tensor("input_305_cast_fp16")]; + tensor linear_117_cast_fp16 = linear(bias = flow_net_res_blocks_2_mlp_0_bias_to_fp16, weight = flow_net_res_blocks_2_mlp_0_weight_to_fp16, x = input_305_cast_fp16)[name = tensor("linear_117_cast_fp16")]; + tensor input_309_cast_fp16 = silu(x = linear_117_cast_fp16)[name = tensor("input_309_cast_fp16")]; + tensor linear_118_cast_fp16 = linear(bias = flow_net_res_blocks_2_mlp_2_bias_to_fp16, weight = flow_net_res_blocks_2_mlp_2_weight_to_fp16, x = input_309_cast_fp16)[name = tensor("linear_118_cast_fp16")]; + tensor var_1821_cast_fp16 = mul(x = var_1792_cast_fp16_2, y = linear_118_cast_fp16)[name = tensor("op_1821_cast_fp16")]; + tensor x_199_cast_fp16 = add(x = x_193_cast_fp16, y = var_1821_cast_fp16)[name = tensor("x_199_cast_fp16")]; + tensor linear_119_cast_fp16 = linear(bias = flow_net_res_blocks_3_adaLN_modulation_1_bias_to_fp16, weight = flow_net_res_blocks_3_adaLN_modulation_1_weight_to_fp16, x = input_287_cast_fp16)[name = tensor("linear_119_cast_fp16")]; + tensor var_1831_split_sizes_0 = const()[name = tensor("op_1831_split_sizes_0"), val = tensor([512, 512, 512])]; + tensor var_1831_axis_0 = const()[name = tensor("op_1831_axis_0"), val = tensor(-1)]; + tensor var_1831_cast_fp16_0, tensor var_1831_cast_fp16_1, tensor var_1831_cast_fp16_2 = split(axis = var_1831_axis_0, split_sizes = var_1831_split_sizes_0, x = linear_119_cast_fp16)[name = tensor("op_1831_cast_fp16")]; + tensor mean_63_axes_0 = const()[name = tensor("mean_63_axes_0"), val = tensor([-1])]; + tensor mean_63_keep_dims_0 = const()[name = tensor("mean_63_keep_dims_0"), val = tensor(true)]; + tensor mean_63_cast_fp16 = reduce_mean(axes = mean_63_axes_0, keep_dims = mean_63_keep_dims_0, x = x_199_cast_fp16)[name = tensor("mean_63_cast_fp16")]; + tensor sub_51_cast_fp16 = sub(x = x_199_cast_fp16, y = mean_63_cast_fp16)[name = tensor("sub_51_cast_fp16")]; + tensor square_41_cast_fp16 = square(x = sub_51_cast_fp16)[name = tensor("square_41_cast_fp16")]; + tensor reduce_mean_83_axes_0 = const()[name = tensor("reduce_mean_83_axes_0"), val = tensor([-1])]; + tensor reduce_mean_83_keep_dims_0 = const()[name = tensor("reduce_mean_83_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_83_cast_fp16 = reduce_mean(axes = reduce_mean_83_axes_0, keep_dims = reduce_mean_83_keep_dims_0, x = square_41_cast_fp16)[name = tensor("reduce_mean_83_cast_fp16")]; + tensor var_1841_to_fp16 = const()[name = tensor("op_1841_to_fp16"), val = tensor(0x1.1p-20)]; + tensor var_1842_cast_fp16 = add(x = reduce_mean_83_cast_fp16, y = var_1841_to_fp16)[name = tensor("op_1842_cast_fp16")]; + tensor var_1843_cast_fp16 = sqrt(x = var_1842_cast_fp16)[name = tensor("op_1843_cast_fp16")]; + tensor x_201_cast_fp16 = real_div(x = sub_51_cast_fp16, y = var_1843_cast_fp16)[name = tensor("x_201_cast_fp16")]; + tensor var_1845_cast_fp16 = mul(x = x_201_cast_fp16, y = flow_net_res_blocks_3_in_ln_weight_to_fp16)[name = tensor("op_1845_cast_fp16")]; + tensor x_203_cast_fp16 = add(x = var_1845_cast_fp16, y = flow_net_res_blocks_3_in_ln_bias_to_fp16)[name = tensor("x_203_cast_fp16")]; + tensor var_1847_promoted_to_fp16 = const()[name = tensor("op_1847_promoted_to_fp16"), val = tensor(0x1p+0)]; + tensor var_1848_cast_fp16 = add(x = var_1831_cast_fp16_1, y = var_1847_promoted_to_fp16)[name = tensor("op_1848_cast_fp16")]; + tensor var_1849_cast_fp16 = mul(x = x_203_cast_fp16, y = var_1848_cast_fp16)[name = tensor("op_1849_cast_fp16")]; + tensor input_313_cast_fp16 = add(x = var_1849_cast_fp16, y = var_1831_cast_fp16_0)[name = tensor("input_313_cast_fp16")]; + tensor linear_120_cast_fp16 = linear(bias = flow_net_res_blocks_3_mlp_0_bias_to_fp16, weight = flow_net_res_blocks_3_mlp_0_weight_to_fp16, x = input_313_cast_fp16)[name = tensor("linear_120_cast_fp16")]; + tensor input_317_cast_fp16 = silu(x = linear_120_cast_fp16)[name = tensor("input_317_cast_fp16")]; + tensor linear_121_cast_fp16 = linear(bias = flow_net_res_blocks_3_mlp_2_bias_to_fp16, weight = flow_net_res_blocks_3_mlp_2_weight_to_fp16, x = input_317_cast_fp16)[name = tensor("linear_121_cast_fp16")]; + tensor var_1860_cast_fp16 = mul(x = var_1831_cast_fp16_2, y = linear_121_cast_fp16)[name = tensor("op_1860_cast_fp16")]; + tensor x_205_cast_fp16 = add(x = x_199_cast_fp16, y = var_1860_cast_fp16)[name = tensor("x_205_cast_fp16")]; + tensor linear_122_cast_fp16 = linear(bias = flow_net_res_blocks_4_adaLN_modulation_1_bias_to_fp16, weight = flow_net_res_blocks_4_adaLN_modulation_1_weight_to_fp16, x = input_287_cast_fp16)[name = tensor("linear_122_cast_fp16")]; + tensor var_1870_split_sizes_0 = const()[name = tensor("op_1870_split_sizes_0"), val = tensor([512, 512, 512])]; + tensor var_1870_axis_0 = const()[name = tensor("op_1870_axis_0"), val = tensor(-1)]; + tensor var_1870_cast_fp16_0, tensor var_1870_cast_fp16_1, tensor var_1870_cast_fp16_2 = split(axis = var_1870_axis_0, split_sizes = var_1870_split_sizes_0, x = linear_122_cast_fp16)[name = tensor("op_1870_cast_fp16")]; + tensor mean_65_axes_0 = const()[name = tensor("mean_65_axes_0"), val = tensor([-1])]; + tensor mean_65_keep_dims_0 = const()[name = tensor("mean_65_keep_dims_0"), val = tensor(true)]; + tensor mean_65_cast_fp16 = reduce_mean(axes = mean_65_axes_0, keep_dims = mean_65_keep_dims_0, x = x_205_cast_fp16)[name = tensor("mean_65_cast_fp16")]; + tensor sub_52_cast_fp16 = sub(x = x_205_cast_fp16, y = mean_65_cast_fp16)[name = tensor("sub_52_cast_fp16")]; + tensor square_42_cast_fp16 = square(x = sub_52_cast_fp16)[name = tensor("square_42_cast_fp16")]; + tensor reduce_mean_85_axes_0 = const()[name = tensor("reduce_mean_85_axes_0"), val = tensor([-1])]; + tensor reduce_mean_85_keep_dims_0 = const()[name = tensor("reduce_mean_85_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_85_cast_fp16 = reduce_mean(axes = reduce_mean_85_axes_0, keep_dims = reduce_mean_85_keep_dims_0, x = square_42_cast_fp16)[name = tensor("reduce_mean_85_cast_fp16")]; + tensor var_1880_to_fp16 = const()[name = tensor("op_1880_to_fp16"), val = tensor(0x1.1p-20)]; + tensor var_1881_cast_fp16 = add(x = reduce_mean_85_cast_fp16, y = var_1880_to_fp16)[name = tensor("op_1881_cast_fp16")]; + tensor var_1882_cast_fp16 = sqrt(x = var_1881_cast_fp16)[name = tensor("op_1882_cast_fp16")]; + tensor x_207_cast_fp16 = real_div(x = sub_52_cast_fp16, y = var_1882_cast_fp16)[name = tensor("x_207_cast_fp16")]; + tensor var_1884_cast_fp16 = mul(x = x_207_cast_fp16, y = flow_net_res_blocks_4_in_ln_weight_to_fp16)[name = tensor("op_1884_cast_fp16")]; + tensor x_209_cast_fp16 = add(x = var_1884_cast_fp16, y = flow_net_res_blocks_4_in_ln_bias_to_fp16)[name = tensor("x_209_cast_fp16")]; + tensor var_1886_promoted_to_fp16 = const()[name = tensor("op_1886_promoted_to_fp16"), val = tensor(0x1p+0)]; + tensor var_1887_cast_fp16 = add(x = var_1870_cast_fp16_1, y = var_1886_promoted_to_fp16)[name = tensor("op_1887_cast_fp16")]; + tensor var_1888_cast_fp16 = mul(x = x_209_cast_fp16, y = var_1887_cast_fp16)[name = tensor("op_1888_cast_fp16")]; + tensor input_321_cast_fp16 = add(x = var_1888_cast_fp16, y = var_1870_cast_fp16_0)[name = tensor("input_321_cast_fp16")]; + tensor linear_123_cast_fp16 = linear(bias = flow_net_res_blocks_4_mlp_0_bias_to_fp16, weight = flow_net_res_blocks_4_mlp_0_weight_to_fp16, x = input_321_cast_fp16)[name = tensor("linear_123_cast_fp16")]; + tensor input_325_cast_fp16 = silu(x = linear_123_cast_fp16)[name = tensor("input_325_cast_fp16")]; + tensor linear_124_cast_fp16 = linear(bias = flow_net_res_blocks_4_mlp_2_bias_to_fp16, weight = flow_net_res_blocks_4_mlp_2_weight_to_fp16, x = input_325_cast_fp16)[name = tensor("linear_124_cast_fp16")]; + tensor var_1899_cast_fp16 = mul(x = var_1870_cast_fp16_2, y = linear_124_cast_fp16)[name = tensor("op_1899_cast_fp16")]; + tensor x_211_cast_fp16 = add(x = x_205_cast_fp16, y = var_1899_cast_fp16)[name = tensor("x_211_cast_fp16")]; + tensor linear_125_cast_fp16 = linear(bias = flow_net_res_blocks_5_adaLN_modulation_1_bias_to_fp16, weight = flow_net_res_blocks_5_adaLN_modulation_1_weight_to_fp16, x = input_287_cast_fp16)[name = tensor("linear_125_cast_fp16")]; + tensor var_1909_split_sizes_0 = const()[name = tensor("op_1909_split_sizes_0"), val = tensor([512, 512, 512])]; + tensor var_1909_axis_0 = const()[name = tensor("op_1909_axis_0"), val = tensor(-1)]; + tensor var_1909_cast_fp16_0, tensor var_1909_cast_fp16_1, tensor var_1909_cast_fp16_2 = split(axis = var_1909_axis_0, split_sizes = var_1909_split_sizes_0, x = linear_125_cast_fp16)[name = tensor("op_1909_cast_fp16")]; + tensor mean_67_axes_0 = const()[name = tensor("mean_67_axes_0"), val = tensor([-1])]; + tensor mean_67_keep_dims_0 = const()[name = tensor("mean_67_keep_dims_0"), val = tensor(true)]; + tensor mean_67_cast_fp16 = reduce_mean(axes = mean_67_axes_0, keep_dims = mean_67_keep_dims_0, x = x_211_cast_fp16)[name = tensor("mean_67_cast_fp16")]; + tensor sub_53_cast_fp16 = sub(x = x_211_cast_fp16, y = mean_67_cast_fp16)[name = tensor("sub_53_cast_fp16")]; + tensor square_43_cast_fp16 = square(x = sub_53_cast_fp16)[name = tensor("square_43_cast_fp16")]; + tensor reduce_mean_87_axes_0 = const()[name = tensor("reduce_mean_87_axes_0"), val = tensor([-1])]; + tensor reduce_mean_87_keep_dims_0 = const()[name = tensor("reduce_mean_87_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_87_cast_fp16 = reduce_mean(axes = reduce_mean_87_axes_0, keep_dims = reduce_mean_87_keep_dims_0, x = square_43_cast_fp16)[name = tensor("reduce_mean_87_cast_fp16")]; + tensor var_1919_to_fp16 = const()[name = tensor("op_1919_to_fp16"), val = tensor(0x1.1p-20)]; + tensor var_1920_cast_fp16 = add(x = reduce_mean_87_cast_fp16, y = var_1919_to_fp16)[name = tensor("op_1920_cast_fp16")]; + tensor var_1921_cast_fp16 = sqrt(x = var_1920_cast_fp16)[name = tensor("op_1921_cast_fp16")]; + tensor x_213_cast_fp16 = real_div(x = sub_53_cast_fp16, y = var_1921_cast_fp16)[name = tensor("x_213_cast_fp16")]; + tensor var_1923_cast_fp16 = mul(x = x_213_cast_fp16, y = flow_net_res_blocks_5_in_ln_weight_to_fp16)[name = tensor("op_1923_cast_fp16")]; + tensor x_215_cast_fp16 = add(x = var_1923_cast_fp16, y = flow_net_res_blocks_5_in_ln_bias_to_fp16)[name = tensor("x_215_cast_fp16")]; + tensor var_1925_promoted_to_fp16 = const()[name = tensor("op_1925_promoted_to_fp16"), val = tensor(0x1p+0)]; + tensor var_1926_cast_fp16 = add(x = var_1909_cast_fp16_1, y = var_1925_promoted_to_fp16)[name = tensor("op_1926_cast_fp16")]; + tensor var_1927_cast_fp16 = mul(x = x_215_cast_fp16, y = var_1926_cast_fp16)[name = tensor("op_1927_cast_fp16")]; + tensor input_329_cast_fp16 = add(x = var_1927_cast_fp16, y = var_1909_cast_fp16_0)[name = tensor("input_329_cast_fp16")]; + tensor linear_126_cast_fp16 = linear(bias = flow_net_res_blocks_5_mlp_0_bias_to_fp16, weight = flow_net_res_blocks_5_mlp_0_weight_to_fp16, x = input_329_cast_fp16)[name = tensor("linear_126_cast_fp16")]; + tensor input_333_cast_fp16 = silu(x = linear_126_cast_fp16)[name = tensor("input_333_cast_fp16")]; + tensor linear_127_cast_fp16 = linear(bias = flow_net_res_blocks_5_mlp_2_bias_to_fp16, weight = flow_net_res_blocks_5_mlp_2_weight_to_fp16, x = input_333_cast_fp16)[name = tensor("linear_127_cast_fp16")]; + tensor var_1938_cast_fp16 = mul(x = var_1909_cast_fp16_2, y = linear_127_cast_fp16)[name = tensor("op_1938_cast_fp16")]; + tensor x_217_cast_fp16 = add(x = x_211_cast_fp16, y = var_1938_cast_fp16)[name = tensor("x_217_cast_fp16")]; + tensor linear_128_cast_fp16 = linear(bias = flow_net_final_layer_adaLN_modulation_1_bias_to_fp16, weight = flow_net_final_layer_adaLN_modulation_1_weight_to_fp16, x = input_287_cast_fp16)[name = tensor("linear_128_cast_fp16")]; + tensor var_1947_split_sizes_0 = const()[name = tensor("op_1947_split_sizes_0"), val = tensor([512, 512])]; + tensor var_1947_axis_0 = const()[name = tensor("op_1947_axis_0"), val = tensor(-1)]; + tensor var_1947_cast_fp16_0, tensor var_1947_cast_fp16_1 = split(axis = var_1947_axis_0, split_sizes = var_1947_split_sizes_0, x = linear_128_cast_fp16)[name = tensor("op_1947_cast_fp16")]; + tensor mean_69_axes_0 = const()[name = tensor("mean_69_axes_0"), val = tensor([-1])]; + tensor mean_69_keep_dims_0 = const()[name = tensor("mean_69_keep_dims_0"), val = tensor(true)]; + tensor mean_69_cast_fp16 = reduce_mean(axes = mean_69_axes_0, keep_dims = mean_69_keep_dims_0, x = x_217_cast_fp16)[name = tensor("mean_69_cast_fp16")]; + tensor sub_54_cast_fp16 = sub(x = x_217_cast_fp16, y = mean_69_cast_fp16)[name = tensor("sub_54_cast_fp16")]; + tensor square_44_cast_fp16 = square(x = sub_54_cast_fp16)[name = tensor("square_44_cast_fp16")]; + tensor reduce_mean_89_axes_0 = const()[name = tensor("reduce_mean_89_axes_0"), val = tensor([-1])]; + tensor reduce_mean_89_keep_dims_0 = const()[name = tensor("reduce_mean_89_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_89_cast_fp16 = reduce_mean(axes = reduce_mean_89_axes_0, keep_dims = reduce_mean_89_keep_dims_0, x = square_44_cast_fp16)[name = tensor("reduce_mean_89_cast_fp16")]; + tensor var_1954_to_fp16 = const()[name = tensor("op_1954_to_fp16"), val = tensor(0x1.1p-20)]; + tensor var_1955_cast_fp16 = add(x = reduce_mean_89_cast_fp16, y = var_1954_to_fp16)[name = tensor("op_1955_cast_fp16")]; + tensor var_1956_cast_fp16 = sqrt(x = var_1955_cast_fp16)[name = tensor("op_1956_cast_fp16")]; + tensor x_219_cast_fp16 = real_div(x = sub_54_cast_fp16, y = var_1956_cast_fp16)[name = tensor("x_219_cast_fp16")]; + tensor var_1958_promoted_to_fp16 = const()[name = tensor("op_1958_promoted_to_fp16"), val = tensor(0x1p+0)]; + tensor var_1959_cast_fp16 = add(x = var_1947_cast_fp16_1, y = var_1958_promoted_to_fp16)[name = tensor("op_1959_cast_fp16")]; + tensor var_1960_cast_fp16 = mul(x = x_219_cast_fp16, y = var_1959_cast_fp16)[name = tensor("op_1960_cast_fp16")]; + tensor input_337_cast_fp16 = add(x = var_1960_cast_fp16, y = var_1947_cast_fp16_0)[name = tensor("input_337_cast_fp16")]; + tensor linear_129_cast_fp16 = linear(bias = flow_net_final_layer_linear_bias_to_fp16, weight = flow_net_final_layer_linear_weight_to_fp16, x = input_337_cast_fp16)[name = tensor("linear_129_cast_fp16")]; + tensor var_1971_to_fp16 = const()[name = tensor("op_1971_to_fp16"), val = tensor(0x1p-3)]; + tensor var_1972_cast_fp16 = mul(x = linear_129_cast_fp16, y = var_1971_to_fp16)[name = tensor("op_1972_cast_fp16")]; + tensor input_339_cast_fp16 = add(x = input_271_cast_fp16, y = var_1972_cast_fp16)[name = tensor("input_339_cast_fp16")]; + tensor linear_130_cast_fp16 = linear(bias = flow_net_input_proj_bias_to_fp16, weight = flow_net_input_proj_weight_to_fp16, x = input_339_cast_fp16)[name = tensor("linear_130_cast_fp16")]; + tensor input_343_to_fp16 = const()[name = tensor("input_343_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(19006080)))]; + tensor input_345_cast_fp16 = silu(x = input_343_to_fp16)[name = tensor("input_345_cast_fp16")]; + tensor linear_132_cast_fp16 = linear(bias = flow_net_time_embed_0_mlp_2_bias_to_fp16, weight = flow_net_time_embed_0_mlp_2_weight_to_fp16, x = input_345_cast_fp16)[name = tensor("linear_132_cast_fp16")]; + tensor reduce_mean_90_axes_0 = const()[name = tensor("reduce_mean_90_axes_0"), val = tensor([-1])]; + tensor reduce_mean_90_keep_dims_0 = const()[name = tensor("reduce_mean_90_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_90_cast_fp16 = reduce_mean(axes = reduce_mean_90_axes_0, keep_dims = reduce_mean_90_keep_dims_0, x = linear_132_cast_fp16)[name = tensor("reduce_mean_90_cast_fp16")]; + tensor sub_55_cast_fp16 = sub(x = linear_132_cast_fp16, y = reduce_mean_90_cast_fp16)[name = tensor("sub_55_cast_fp16")]; + tensor square_45_cast_fp16 = square(x = sub_55_cast_fp16)[name = tensor("square_45_cast_fp16")]; + tensor reduce_mean_91_axes_0 = const()[name = tensor("reduce_mean_91_axes_0"), val = tensor([-1])]; + tensor reduce_mean_91_keep_dims_0 = const()[name = tensor("reduce_mean_91_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_91_cast_fp16 = reduce_mean(axes = reduce_mean_91_axes_0, keep_dims = reduce_mean_91_keep_dims_0, x = square_45_cast_fp16)[name = tensor("reduce_mean_91_cast_fp16")]; + tensor real_div_10_to_fp16 = const()[name = tensor("real_div_10_to_fp16"), val = tensor(0x1.008p+0)]; + tensor mul_10_cast_fp16 = mul(x = reduce_mean_91_cast_fp16, y = real_div_10_to_fp16)[name = tensor("mul_10_cast_fp16")]; + tensor var_2038_to_fp16 = const()[name = tensor("op_2038_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_91_cast_fp16 = add(x = mul_10_cast_fp16, y = var_2038_to_fp16)[name = tensor("var_91_cast_fp16")]; + tensor var_2041_epsilon_0 = const()[name = tensor("op_2041_epsilon_0"), val = tensor(0x1.197998p-40)]; + tensor var_2041_cast_fp16 = rsqrt(epsilon = var_2041_epsilon_0, x = var_91_cast_fp16)[name = tensor("op_2041_cast_fp16")]; + tensor var_2042_cast_fp16 = mul(x = const_3_to_fp16, y = var_2041_cast_fp16)[name = tensor("op_2042_cast_fp16")]; + tensor var_2043_cast_fp16 = mul(x = linear_132_cast_fp16, y = var_2042_cast_fp16)[name = tensor("op_2043_cast_fp16")]; + tensor input_349_to_fp16 = const()[name = tensor("input_349_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(19007168)))]; + tensor input_351_cast_fp16 = silu(x = input_349_to_fp16)[name = tensor("input_351_cast_fp16")]; + tensor linear_134_cast_fp16 = linear(bias = flow_net_time_embed_1_mlp_2_bias_to_fp16, weight = flow_net_time_embed_1_mlp_2_weight_to_fp16, x = input_351_cast_fp16)[name = tensor("linear_134_cast_fp16")]; + tensor reduce_mean_92_axes_0 = const()[name = tensor("reduce_mean_92_axes_0"), val = tensor([-1])]; + tensor reduce_mean_92_keep_dims_0 = const()[name = tensor("reduce_mean_92_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_92_cast_fp16 = reduce_mean(axes = reduce_mean_92_axes_0, keep_dims = reduce_mean_92_keep_dims_0, x = linear_134_cast_fp16)[name = tensor("reduce_mean_92_cast_fp16")]; + tensor sub_57_cast_fp16 = sub(x = linear_134_cast_fp16, y = reduce_mean_92_cast_fp16)[name = tensor("sub_57_cast_fp16")]; + tensor square_46_cast_fp16 = square(x = sub_57_cast_fp16)[name = tensor("square_46_cast_fp16")]; + tensor reduce_mean_93_axes_0 = const()[name = tensor("reduce_mean_93_axes_0"), val = tensor([-1])]; + tensor reduce_mean_93_keep_dims_0 = const()[name = tensor("reduce_mean_93_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_93_cast_fp16 = reduce_mean(axes = reduce_mean_93_axes_0, keep_dims = reduce_mean_93_keep_dims_0, x = square_46_cast_fp16)[name = tensor("reduce_mean_93_cast_fp16")]; + tensor real_div_11_to_fp16 = const()[name = tensor("real_div_11_to_fp16"), val = tensor(0x1.008p+0)]; + tensor mul_11_cast_fp16 = mul(x = reduce_mean_93_cast_fp16, y = real_div_11_to_fp16)[name = tensor("mul_11_cast_fp16")]; + tensor var_2075_to_fp16 = const()[name = tensor("op_2075_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_93_cast_fp16 = add(x = mul_11_cast_fp16, y = var_2075_to_fp16)[name = tensor("var_93_cast_fp16")]; + tensor var_2078_epsilon_0 = const()[name = tensor("op_2078_epsilon_0"), val = tensor(0x1.197998p-40)]; + tensor var_2078_cast_fp16 = rsqrt(epsilon = var_2078_epsilon_0, x = var_93_cast_fp16)[name = tensor("op_2078_cast_fp16")]; + tensor var_2079_cast_fp16 = mul(x = const_5_to_fp16, y = var_2078_cast_fp16)[name = tensor("op_2079_cast_fp16")]; + tensor var_2080_cast_fp16 = mul(x = linear_134_cast_fp16, y = var_2079_cast_fp16)[name = tensor("op_2080_cast_fp16")]; + tensor var_2092_cast_fp16 = add(x = var_2043_cast_fp16, y = var_2080_cast_fp16)[name = tensor("op_2092_cast_fp16")]; + tensor _inversed_t_combined_11_y_0_to_fp16 = const()[name = tensor("_inversed_t_combined_11_y_0_to_fp16"), val = tensor(0x1p-1)]; + tensor _inversed_t_combined_11_cast_fp16 = mul(x = var_2092_cast_fp16, y = _inversed_t_combined_11_y_0_to_fp16)[name = tensor("_inversed_t_combined_11_cast_fp16")]; + tensor input_353_cast_fp16 = add(x = _inversed_t_combined_11_cast_fp16, y = linear_5_cast_fp16)[name = tensor("input_353_cast_fp16")]; + tensor input_355_cast_fp16 = silu(x = input_353_cast_fp16)[name = tensor("input_355_cast_fp16")]; + tensor linear_136_cast_fp16 = linear(bias = flow_net_res_blocks_0_adaLN_modulation_1_bias_to_fp16, weight = flow_net_res_blocks_0_adaLN_modulation_1_weight_to_fp16, x = input_355_cast_fp16)[name = tensor("linear_136_cast_fp16")]; + tensor var_2107_split_sizes_0 = const()[name = tensor("op_2107_split_sizes_0"), val = tensor([512, 512, 512])]; + tensor var_2107_axis_0 = const()[name = tensor("op_2107_axis_0"), val = tensor(-1)]; + tensor var_2107_cast_fp16_0, tensor var_2107_cast_fp16_1, tensor var_2107_cast_fp16_2 = split(axis = var_2107_axis_0, split_sizes = var_2107_split_sizes_0, x = linear_136_cast_fp16)[name = tensor("op_2107_cast_fp16")]; + tensor mean_71_axes_0 = const()[name = tensor("mean_71_axes_0"), val = tensor([-1])]; + tensor mean_71_keep_dims_0 = const()[name = tensor("mean_71_keep_dims_0"), val = tensor(true)]; + tensor mean_71_cast_fp16 = reduce_mean(axes = mean_71_axes_0, keep_dims = mean_71_keep_dims_0, x = linear_130_cast_fp16)[name = tensor("mean_71_cast_fp16")]; + tensor sub_59_cast_fp16 = sub(x = linear_130_cast_fp16, y = mean_71_cast_fp16)[name = tensor("sub_59_cast_fp16")]; + tensor square_47_cast_fp16 = square(x = sub_59_cast_fp16)[name = tensor("square_47_cast_fp16")]; + tensor reduce_mean_95_axes_0 = const()[name = tensor("reduce_mean_95_axes_0"), val = tensor([-1])]; + tensor reduce_mean_95_keep_dims_0 = const()[name = tensor("reduce_mean_95_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_95_cast_fp16 = reduce_mean(axes = reduce_mean_95_axes_0, keep_dims = reduce_mean_95_keep_dims_0, x = square_47_cast_fp16)[name = tensor("reduce_mean_95_cast_fp16")]; + tensor var_2117_to_fp16 = const()[name = tensor("op_2117_to_fp16"), val = tensor(0x1.1p-20)]; + tensor var_2118_cast_fp16 = add(x = reduce_mean_95_cast_fp16, y = var_2117_to_fp16)[name = tensor("op_2118_cast_fp16")]; + tensor var_2119_cast_fp16 = sqrt(x = var_2118_cast_fp16)[name = tensor("op_2119_cast_fp16")]; + tensor x_227_cast_fp16 = real_div(x = sub_59_cast_fp16, y = var_2119_cast_fp16)[name = tensor("x_227_cast_fp16")]; + tensor var_2121_cast_fp16 = mul(x = x_227_cast_fp16, y = flow_net_res_blocks_0_in_ln_weight_to_fp16)[name = tensor("op_2121_cast_fp16")]; + tensor x_229_cast_fp16 = add(x = var_2121_cast_fp16, y = flow_net_res_blocks_0_in_ln_bias_to_fp16)[name = tensor("x_229_cast_fp16")]; + tensor var_2123_promoted_to_fp16 = const()[name = tensor("op_2123_promoted_to_fp16"), val = tensor(0x1p+0)]; + tensor var_2124_cast_fp16 = add(x = var_2107_cast_fp16_1, y = var_2123_promoted_to_fp16)[name = tensor("op_2124_cast_fp16")]; + tensor var_2125_cast_fp16 = mul(x = x_229_cast_fp16, y = var_2124_cast_fp16)[name = tensor("op_2125_cast_fp16")]; + tensor input_357_cast_fp16 = add(x = var_2125_cast_fp16, y = var_2107_cast_fp16_0)[name = tensor("input_357_cast_fp16")]; + tensor linear_137_cast_fp16 = linear(bias = flow_net_res_blocks_0_mlp_0_bias_to_fp16, weight = flow_net_res_blocks_0_mlp_0_weight_to_fp16, x = input_357_cast_fp16)[name = tensor("linear_137_cast_fp16")]; + tensor input_361_cast_fp16 = silu(x = linear_137_cast_fp16)[name = tensor("input_361_cast_fp16")]; + tensor linear_138_cast_fp16 = linear(bias = flow_net_res_blocks_0_mlp_2_bias_to_fp16, weight = flow_net_res_blocks_0_mlp_2_weight_to_fp16, x = input_361_cast_fp16)[name = tensor("linear_138_cast_fp16")]; + tensor var_2136_cast_fp16 = mul(x = var_2107_cast_fp16_2, y = linear_138_cast_fp16)[name = tensor("op_2136_cast_fp16")]; + tensor x_231_cast_fp16 = add(x = linear_130_cast_fp16, y = var_2136_cast_fp16)[name = tensor("x_231_cast_fp16")]; + tensor linear_139_cast_fp16 = linear(bias = flow_net_res_blocks_1_adaLN_modulation_1_bias_to_fp16, weight = flow_net_res_blocks_1_adaLN_modulation_1_weight_to_fp16, x = input_355_cast_fp16)[name = tensor("linear_139_cast_fp16")]; + tensor var_2146_split_sizes_0 = const()[name = tensor("op_2146_split_sizes_0"), val = tensor([512, 512, 512])]; + tensor var_2146_axis_0 = const()[name = tensor("op_2146_axis_0"), val = tensor(-1)]; + tensor var_2146_cast_fp16_0, tensor var_2146_cast_fp16_1, tensor var_2146_cast_fp16_2 = split(axis = var_2146_axis_0, split_sizes = var_2146_split_sizes_0, x = linear_139_cast_fp16)[name = tensor("op_2146_cast_fp16")]; + tensor mean_73_axes_0 = const()[name = tensor("mean_73_axes_0"), val = tensor([-1])]; + tensor mean_73_keep_dims_0 = const()[name = tensor("mean_73_keep_dims_0"), val = tensor(true)]; + tensor mean_73_cast_fp16 = reduce_mean(axes = mean_73_axes_0, keep_dims = mean_73_keep_dims_0, x = x_231_cast_fp16)[name = tensor("mean_73_cast_fp16")]; + tensor sub_60_cast_fp16 = sub(x = x_231_cast_fp16, y = mean_73_cast_fp16)[name = tensor("sub_60_cast_fp16")]; + tensor square_48_cast_fp16 = square(x = sub_60_cast_fp16)[name = tensor("square_48_cast_fp16")]; + tensor reduce_mean_97_axes_0 = const()[name = tensor("reduce_mean_97_axes_0"), val = tensor([-1])]; + tensor reduce_mean_97_keep_dims_0 = const()[name = tensor("reduce_mean_97_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_97_cast_fp16 = reduce_mean(axes = reduce_mean_97_axes_0, keep_dims = reduce_mean_97_keep_dims_0, x = square_48_cast_fp16)[name = tensor("reduce_mean_97_cast_fp16")]; + tensor var_2156_to_fp16 = const()[name = tensor("op_2156_to_fp16"), val = tensor(0x1.1p-20)]; + tensor var_2157_cast_fp16 = add(x = reduce_mean_97_cast_fp16, y = var_2156_to_fp16)[name = tensor("op_2157_cast_fp16")]; + tensor var_2158_cast_fp16 = sqrt(x = var_2157_cast_fp16)[name = tensor("op_2158_cast_fp16")]; + tensor x_233_cast_fp16 = real_div(x = sub_60_cast_fp16, y = var_2158_cast_fp16)[name = tensor("x_233_cast_fp16")]; + tensor var_2160_cast_fp16 = mul(x = x_233_cast_fp16, y = flow_net_res_blocks_1_in_ln_weight_to_fp16)[name = tensor("op_2160_cast_fp16")]; + tensor x_235_cast_fp16 = add(x = var_2160_cast_fp16, y = flow_net_res_blocks_1_in_ln_bias_to_fp16)[name = tensor("x_235_cast_fp16")]; + tensor var_2162_promoted_to_fp16 = const()[name = tensor("op_2162_promoted_to_fp16"), val = tensor(0x1p+0)]; + tensor var_2163_cast_fp16 = add(x = var_2146_cast_fp16_1, y = var_2162_promoted_to_fp16)[name = tensor("op_2163_cast_fp16")]; + tensor var_2164_cast_fp16 = mul(x = x_235_cast_fp16, y = var_2163_cast_fp16)[name = tensor("op_2164_cast_fp16")]; + tensor input_365_cast_fp16 = add(x = var_2164_cast_fp16, y = var_2146_cast_fp16_0)[name = tensor("input_365_cast_fp16")]; + tensor linear_140_cast_fp16 = linear(bias = flow_net_res_blocks_1_mlp_0_bias_to_fp16, weight = flow_net_res_blocks_1_mlp_0_weight_to_fp16, x = input_365_cast_fp16)[name = tensor("linear_140_cast_fp16")]; + tensor input_369_cast_fp16 = silu(x = linear_140_cast_fp16)[name = tensor("input_369_cast_fp16")]; + tensor linear_141_cast_fp16 = linear(bias = flow_net_res_blocks_1_mlp_2_bias_to_fp16, weight = flow_net_res_blocks_1_mlp_2_weight_to_fp16, x = input_369_cast_fp16)[name = tensor("linear_141_cast_fp16")]; + tensor var_2175_cast_fp16 = mul(x = var_2146_cast_fp16_2, y = linear_141_cast_fp16)[name = tensor("op_2175_cast_fp16")]; + tensor x_237_cast_fp16 = add(x = x_231_cast_fp16, y = var_2175_cast_fp16)[name = tensor("x_237_cast_fp16")]; + tensor linear_142_cast_fp16 = linear(bias = flow_net_res_blocks_2_adaLN_modulation_1_bias_to_fp16, weight = flow_net_res_blocks_2_adaLN_modulation_1_weight_to_fp16, x = input_355_cast_fp16)[name = tensor("linear_142_cast_fp16")]; + tensor var_2185_split_sizes_0 = const()[name = tensor("op_2185_split_sizes_0"), val = tensor([512, 512, 512])]; + tensor var_2185_axis_0 = const()[name = tensor("op_2185_axis_0"), val = tensor(-1)]; + tensor var_2185_cast_fp16_0, tensor var_2185_cast_fp16_1, tensor var_2185_cast_fp16_2 = split(axis = var_2185_axis_0, split_sizes = var_2185_split_sizes_0, x = linear_142_cast_fp16)[name = tensor("op_2185_cast_fp16")]; + tensor mean_75_axes_0 = const()[name = tensor("mean_75_axes_0"), val = tensor([-1])]; + tensor mean_75_keep_dims_0 = const()[name = tensor("mean_75_keep_dims_0"), val = tensor(true)]; + tensor mean_75_cast_fp16 = reduce_mean(axes = mean_75_axes_0, keep_dims = mean_75_keep_dims_0, x = x_237_cast_fp16)[name = tensor("mean_75_cast_fp16")]; + tensor sub_61_cast_fp16 = sub(x = x_237_cast_fp16, y = mean_75_cast_fp16)[name = tensor("sub_61_cast_fp16")]; + tensor square_49_cast_fp16 = square(x = sub_61_cast_fp16)[name = tensor("square_49_cast_fp16")]; + tensor reduce_mean_99_axes_0 = const()[name = tensor("reduce_mean_99_axes_0"), val = tensor([-1])]; + tensor reduce_mean_99_keep_dims_0 = const()[name = tensor("reduce_mean_99_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_99_cast_fp16 = reduce_mean(axes = reduce_mean_99_axes_0, keep_dims = reduce_mean_99_keep_dims_0, x = square_49_cast_fp16)[name = tensor("reduce_mean_99_cast_fp16")]; + tensor var_2195_to_fp16 = const()[name = tensor("op_2195_to_fp16"), val = tensor(0x1.1p-20)]; + tensor var_2196_cast_fp16 = add(x = reduce_mean_99_cast_fp16, y = var_2195_to_fp16)[name = tensor("op_2196_cast_fp16")]; + tensor var_2197_cast_fp16 = sqrt(x = var_2196_cast_fp16)[name = tensor("op_2197_cast_fp16")]; + tensor x_239_cast_fp16 = real_div(x = sub_61_cast_fp16, y = var_2197_cast_fp16)[name = tensor("x_239_cast_fp16")]; + tensor var_2199_cast_fp16 = mul(x = x_239_cast_fp16, y = flow_net_res_blocks_2_in_ln_weight_to_fp16)[name = tensor("op_2199_cast_fp16")]; + tensor x_241_cast_fp16 = add(x = var_2199_cast_fp16, y = flow_net_res_blocks_2_in_ln_bias_to_fp16)[name = tensor("x_241_cast_fp16")]; + tensor var_2201_promoted_to_fp16 = const()[name = tensor("op_2201_promoted_to_fp16"), val = tensor(0x1p+0)]; + tensor var_2202_cast_fp16 = add(x = var_2185_cast_fp16_1, y = var_2201_promoted_to_fp16)[name = tensor("op_2202_cast_fp16")]; + tensor var_2203_cast_fp16 = mul(x = x_241_cast_fp16, y = var_2202_cast_fp16)[name = tensor("op_2203_cast_fp16")]; + tensor input_373_cast_fp16 = add(x = var_2203_cast_fp16, y = var_2185_cast_fp16_0)[name = tensor("input_373_cast_fp16")]; + tensor linear_143_cast_fp16 = linear(bias = flow_net_res_blocks_2_mlp_0_bias_to_fp16, weight = flow_net_res_blocks_2_mlp_0_weight_to_fp16, x = input_373_cast_fp16)[name = tensor("linear_143_cast_fp16")]; + tensor input_377_cast_fp16 = silu(x = linear_143_cast_fp16)[name = tensor("input_377_cast_fp16")]; + tensor linear_144_cast_fp16 = linear(bias = flow_net_res_blocks_2_mlp_2_bias_to_fp16, weight = flow_net_res_blocks_2_mlp_2_weight_to_fp16, x = input_377_cast_fp16)[name = tensor("linear_144_cast_fp16")]; + tensor var_2214_cast_fp16 = mul(x = var_2185_cast_fp16_2, y = linear_144_cast_fp16)[name = tensor("op_2214_cast_fp16")]; + tensor x_243_cast_fp16 = add(x = x_237_cast_fp16, y = var_2214_cast_fp16)[name = tensor("x_243_cast_fp16")]; + tensor linear_145_cast_fp16 = linear(bias = flow_net_res_blocks_3_adaLN_modulation_1_bias_to_fp16, weight = flow_net_res_blocks_3_adaLN_modulation_1_weight_to_fp16, x = input_355_cast_fp16)[name = tensor("linear_145_cast_fp16")]; + tensor var_2224_split_sizes_0 = const()[name = tensor("op_2224_split_sizes_0"), val = tensor([512, 512, 512])]; + tensor var_2224_axis_0 = const()[name = tensor("op_2224_axis_0"), val = tensor(-1)]; + tensor var_2224_cast_fp16_0, tensor var_2224_cast_fp16_1, tensor var_2224_cast_fp16_2 = split(axis = var_2224_axis_0, split_sizes = var_2224_split_sizes_0, x = linear_145_cast_fp16)[name = tensor("op_2224_cast_fp16")]; + tensor mean_77_axes_0 = const()[name = tensor("mean_77_axes_0"), val = tensor([-1])]; + tensor mean_77_keep_dims_0 = const()[name = tensor("mean_77_keep_dims_0"), val = tensor(true)]; + tensor mean_77_cast_fp16 = reduce_mean(axes = mean_77_axes_0, keep_dims = mean_77_keep_dims_0, x = x_243_cast_fp16)[name = tensor("mean_77_cast_fp16")]; + tensor sub_62_cast_fp16 = sub(x = x_243_cast_fp16, y = mean_77_cast_fp16)[name = tensor("sub_62_cast_fp16")]; + tensor square_50_cast_fp16 = square(x = sub_62_cast_fp16)[name = tensor("square_50_cast_fp16")]; + tensor reduce_mean_101_axes_0 = const()[name = tensor("reduce_mean_101_axes_0"), val = tensor([-1])]; + tensor reduce_mean_101_keep_dims_0 = const()[name = tensor("reduce_mean_101_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_101_cast_fp16 = reduce_mean(axes = reduce_mean_101_axes_0, keep_dims = reduce_mean_101_keep_dims_0, x = square_50_cast_fp16)[name = tensor("reduce_mean_101_cast_fp16")]; + tensor var_2234_to_fp16 = const()[name = tensor("op_2234_to_fp16"), val = tensor(0x1.1p-20)]; + tensor var_2235_cast_fp16 = add(x = reduce_mean_101_cast_fp16, y = var_2234_to_fp16)[name = tensor("op_2235_cast_fp16")]; + tensor var_2236_cast_fp16 = sqrt(x = var_2235_cast_fp16)[name = tensor("op_2236_cast_fp16")]; + tensor x_245_cast_fp16 = real_div(x = sub_62_cast_fp16, y = var_2236_cast_fp16)[name = tensor("x_245_cast_fp16")]; + tensor var_2238_cast_fp16 = mul(x = x_245_cast_fp16, y = flow_net_res_blocks_3_in_ln_weight_to_fp16)[name = tensor("op_2238_cast_fp16")]; + tensor x_247_cast_fp16 = add(x = var_2238_cast_fp16, y = flow_net_res_blocks_3_in_ln_bias_to_fp16)[name = tensor("x_247_cast_fp16")]; + tensor var_2240_promoted_to_fp16 = const()[name = tensor("op_2240_promoted_to_fp16"), val = tensor(0x1p+0)]; + tensor var_2241_cast_fp16 = add(x = var_2224_cast_fp16_1, y = var_2240_promoted_to_fp16)[name = tensor("op_2241_cast_fp16")]; + tensor var_2242_cast_fp16 = mul(x = x_247_cast_fp16, y = var_2241_cast_fp16)[name = tensor("op_2242_cast_fp16")]; + tensor input_381_cast_fp16 = add(x = var_2242_cast_fp16, y = var_2224_cast_fp16_0)[name = tensor("input_381_cast_fp16")]; + tensor linear_146_cast_fp16 = linear(bias = flow_net_res_blocks_3_mlp_0_bias_to_fp16, weight = flow_net_res_blocks_3_mlp_0_weight_to_fp16, x = input_381_cast_fp16)[name = tensor("linear_146_cast_fp16")]; + tensor input_385_cast_fp16 = silu(x = linear_146_cast_fp16)[name = tensor("input_385_cast_fp16")]; + tensor linear_147_cast_fp16 = linear(bias = flow_net_res_blocks_3_mlp_2_bias_to_fp16, weight = flow_net_res_blocks_3_mlp_2_weight_to_fp16, x = input_385_cast_fp16)[name = tensor("linear_147_cast_fp16")]; + tensor var_2253_cast_fp16 = mul(x = var_2224_cast_fp16_2, y = linear_147_cast_fp16)[name = tensor("op_2253_cast_fp16")]; + tensor x_249_cast_fp16 = add(x = x_243_cast_fp16, y = var_2253_cast_fp16)[name = tensor("x_249_cast_fp16")]; + tensor linear_148_cast_fp16 = linear(bias = flow_net_res_blocks_4_adaLN_modulation_1_bias_to_fp16, weight = flow_net_res_blocks_4_adaLN_modulation_1_weight_to_fp16, x = input_355_cast_fp16)[name = tensor("linear_148_cast_fp16")]; + tensor var_2263_split_sizes_0 = const()[name = tensor("op_2263_split_sizes_0"), val = tensor([512, 512, 512])]; + tensor var_2263_axis_0 = const()[name = tensor("op_2263_axis_0"), val = tensor(-1)]; + tensor var_2263_cast_fp16_0, tensor var_2263_cast_fp16_1, tensor var_2263_cast_fp16_2 = split(axis = var_2263_axis_0, split_sizes = var_2263_split_sizes_0, x = linear_148_cast_fp16)[name = tensor("op_2263_cast_fp16")]; + tensor mean_79_axes_0 = const()[name = tensor("mean_79_axes_0"), val = tensor([-1])]; + tensor mean_79_keep_dims_0 = const()[name = tensor("mean_79_keep_dims_0"), val = tensor(true)]; + tensor mean_79_cast_fp16 = reduce_mean(axes = mean_79_axes_0, keep_dims = mean_79_keep_dims_0, x = x_249_cast_fp16)[name = tensor("mean_79_cast_fp16")]; + tensor sub_63_cast_fp16 = sub(x = x_249_cast_fp16, y = mean_79_cast_fp16)[name = tensor("sub_63_cast_fp16")]; + tensor square_51_cast_fp16 = square(x = sub_63_cast_fp16)[name = tensor("square_51_cast_fp16")]; + tensor reduce_mean_103_axes_0 = const()[name = tensor("reduce_mean_103_axes_0"), val = tensor([-1])]; + tensor reduce_mean_103_keep_dims_0 = const()[name = tensor("reduce_mean_103_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_103_cast_fp16 = reduce_mean(axes = reduce_mean_103_axes_0, keep_dims = reduce_mean_103_keep_dims_0, x = square_51_cast_fp16)[name = tensor("reduce_mean_103_cast_fp16")]; + tensor var_2273_to_fp16 = const()[name = tensor("op_2273_to_fp16"), val = tensor(0x1.1p-20)]; + tensor var_2274_cast_fp16 = add(x = reduce_mean_103_cast_fp16, y = var_2273_to_fp16)[name = tensor("op_2274_cast_fp16")]; + tensor var_2275_cast_fp16 = sqrt(x = var_2274_cast_fp16)[name = tensor("op_2275_cast_fp16")]; + tensor x_251_cast_fp16 = real_div(x = sub_63_cast_fp16, y = var_2275_cast_fp16)[name = tensor("x_251_cast_fp16")]; + tensor var_2277_cast_fp16 = mul(x = x_251_cast_fp16, y = flow_net_res_blocks_4_in_ln_weight_to_fp16)[name = tensor("op_2277_cast_fp16")]; + tensor x_253_cast_fp16 = add(x = var_2277_cast_fp16, y = flow_net_res_blocks_4_in_ln_bias_to_fp16)[name = tensor("x_253_cast_fp16")]; + tensor var_2279_promoted_to_fp16 = const()[name = tensor("op_2279_promoted_to_fp16"), val = tensor(0x1p+0)]; + tensor var_2280_cast_fp16 = add(x = var_2263_cast_fp16_1, y = var_2279_promoted_to_fp16)[name = tensor("op_2280_cast_fp16")]; + tensor var_2281_cast_fp16 = mul(x = x_253_cast_fp16, y = var_2280_cast_fp16)[name = tensor("op_2281_cast_fp16")]; + tensor input_389_cast_fp16 = add(x = var_2281_cast_fp16, y = var_2263_cast_fp16_0)[name = tensor("input_389_cast_fp16")]; + tensor linear_149_cast_fp16 = linear(bias = flow_net_res_blocks_4_mlp_0_bias_to_fp16, weight = flow_net_res_blocks_4_mlp_0_weight_to_fp16, x = input_389_cast_fp16)[name = tensor("linear_149_cast_fp16")]; + tensor input_393_cast_fp16 = silu(x = linear_149_cast_fp16)[name = tensor("input_393_cast_fp16")]; + tensor linear_150_cast_fp16 = linear(bias = flow_net_res_blocks_4_mlp_2_bias_to_fp16, weight = flow_net_res_blocks_4_mlp_2_weight_to_fp16, x = input_393_cast_fp16)[name = tensor("linear_150_cast_fp16")]; + tensor var_2292_cast_fp16 = mul(x = var_2263_cast_fp16_2, y = linear_150_cast_fp16)[name = tensor("op_2292_cast_fp16")]; + tensor x_255_cast_fp16 = add(x = x_249_cast_fp16, y = var_2292_cast_fp16)[name = tensor("x_255_cast_fp16")]; + tensor linear_151_cast_fp16 = linear(bias = flow_net_res_blocks_5_adaLN_modulation_1_bias_to_fp16, weight = flow_net_res_blocks_5_adaLN_modulation_1_weight_to_fp16, x = input_355_cast_fp16)[name = tensor("linear_151_cast_fp16")]; + tensor var_2302_split_sizes_0 = const()[name = tensor("op_2302_split_sizes_0"), val = tensor([512, 512, 512])]; + tensor var_2302_axis_0 = const()[name = tensor("op_2302_axis_0"), val = tensor(-1)]; + tensor var_2302_cast_fp16_0, tensor var_2302_cast_fp16_1, tensor var_2302_cast_fp16_2 = split(axis = var_2302_axis_0, split_sizes = var_2302_split_sizes_0, x = linear_151_cast_fp16)[name = tensor("op_2302_cast_fp16")]; + tensor mean_81_axes_0 = const()[name = tensor("mean_81_axes_0"), val = tensor([-1])]; + tensor mean_81_keep_dims_0 = const()[name = tensor("mean_81_keep_dims_0"), val = tensor(true)]; + tensor mean_81_cast_fp16 = reduce_mean(axes = mean_81_axes_0, keep_dims = mean_81_keep_dims_0, x = x_255_cast_fp16)[name = tensor("mean_81_cast_fp16")]; + tensor sub_64_cast_fp16 = sub(x = x_255_cast_fp16, y = mean_81_cast_fp16)[name = tensor("sub_64_cast_fp16")]; + tensor square_52_cast_fp16 = square(x = sub_64_cast_fp16)[name = tensor("square_52_cast_fp16")]; + tensor reduce_mean_105_axes_0 = const()[name = tensor("reduce_mean_105_axes_0"), val = tensor([-1])]; + tensor reduce_mean_105_keep_dims_0 = const()[name = tensor("reduce_mean_105_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_105_cast_fp16 = reduce_mean(axes = reduce_mean_105_axes_0, keep_dims = reduce_mean_105_keep_dims_0, x = square_52_cast_fp16)[name = tensor("reduce_mean_105_cast_fp16")]; + tensor var_2312_to_fp16 = const()[name = tensor("op_2312_to_fp16"), val = tensor(0x1.1p-20)]; + tensor var_2313_cast_fp16 = add(x = reduce_mean_105_cast_fp16, y = var_2312_to_fp16)[name = tensor("op_2313_cast_fp16")]; + tensor var_2314_cast_fp16 = sqrt(x = var_2313_cast_fp16)[name = tensor("op_2314_cast_fp16")]; + tensor x_257_cast_fp16 = real_div(x = sub_64_cast_fp16, y = var_2314_cast_fp16)[name = tensor("x_257_cast_fp16")]; + tensor var_2316_cast_fp16 = mul(x = x_257_cast_fp16, y = flow_net_res_blocks_5_in_ln_weight_to_fp16)[name = tensor("op_2316_cast_fp16")]; + tensor x_259_cast_fp16 = add(x = var_2316_cast_fp16, y = flow_net_res_blocks_5_in_ln_bias_to_fp16)[name = tensor("x_259_cast_fp16")]; + tensor var_2318_promoted_to_fp16 = const()[name = tensor("op_2318_promoted_to_fp16"), val = tensor(0x1p+0)]; + tensor var_2319_cast_fp16 = add(x = var_2302_cast_fp16_1, y = var_2318_promoted_to_fp16)[name = tensor("op_2319_cast_fp16")]; + tensor var_2320_cast_fp16 = mul(x = x_259_cast_fp16, y = var_2319_cast_fp16)[name = tensor("op_2320_cast_fp16")]; + tensor input_397_cast_fp16 = add(x = var_2320_cast_fp16, y = var_2302_cast_fp16_0)[name = tensor("input_397_cast_fp16")]; + tensor linear_152_cast_fp16 = linear(bias = flow_net_res_blocks_5_mlp_0_bias_to_fp16, weight = flow_net_res_blocks_5_mlp_0_weight_to_fp16, x = input_397_cast_fp16)[name = tensor("linear_152_cast_fp16")]; + tensor input_401_cast_fp16 = silu(x = linear_152_cast_fp16)[name = tensor("input_401_cast_fp16")]; + tensor linear_153_cast_fp16 = linear(bias = flow_net_res_blocks_5_mlp_2_bias_to_fp16, weight = flow_net_res_blocks_5_mlp_2_weight_to_fp16, x = input_401_cast_fp16)[name = tensor("linear_153_cast_fp16")]; + tensor var_2331_cast_fp16 = mul(x = var_2302_cast_fp16_2, y = linear_153_cast_fp16)[name = tensor("op_2331_cast_fp16")]; + tensor x_261_cast_fp16 = add(x = x_255_cast_fp16, y = var_2331_cast_fp16)[name = tensor("x_261_cast_fp16")]; + tensor linear_154_cast_fp16 = linear(bias = flow_net_final_layer_adaLN_modulation_1_bias_to_fp16, weight = flow_net_final_layer_adaLN_modulation_1_weight_to_fp16, x = input_355_cast_fp16)[name = tensor("linear_154_cast_fp16")]; + tensor var_2340_split_sizes_0 = const()[name = tensor("op_2340_split_sizes_0"), val = tensor([512, 512])]; + tensor var_2340_axis_0 = const()[name = tensor("op_2340_axis_0"), val = tensor(-1)]; + tensor var_2340_cast_fp16_0, tensor var_2340_cast_fp16_1 = split(axis = var_2340_axis_0, split_sizes = var_2340_split_sizes_0, x = linear_154_cast_fp16)[name = tensor("op_2340_cast_fp16")]; + tensor mean_83_axes_0 = const()[name = tensor("mean_83_axes_0"), val = tensor([-1])]; + tensor mean_83_keep_dims_0 = const()[name = tensor("mean_83_keep_dims_0"), val = tensor(true)]; + tensor mean_83_cast_fp16 = reduce_mean(axes = mean_83_axes_0, keep_dims = mean_83_keep_dims_0, x = x_261_cast_fp16)[name = tensor("mean_83_cast_fp16")]; + tensor sub_65_cast_fp16 = sub(x = x_261_cast_fp16, y = mean_83_cast_fp16)[name = tensor("sub_65_cast_fp16")]; + tensor square_53_cast_fp16 = square(x = sub_65_cast_fp16)[name = tensor("square_53_cast_fp16")]; + tensor reduce_mean_107_axes_0 = const()[name = tensor("reduce_mean_107_axes_0"), val = tensor([-1])]; + tensor reduce_mean_107_keep_dims_0 = const()[name = tensor("reduce_mean_107_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_107_cast_fp16 = reduce_mean(axes = reduce_mean_107_axes_0, keep_dims = reduce_mean_107_keep_dims_0, x = square_53_cast_fp16)[name = tensor("reduce_mean_107_cast_fp16")]; + tensor var_2347_to_fp16 = const()[name = tensor("op_2347_to_fp16"), val = tensor(0x1.1p-20)]; + tensor var_2348_cast_fp16 = add(x = reduce_mean_107_cast_fp16, y = var_2347_to_fp16)[name = tensor("op_2348_cast_fp16")]; + tensor var_2349_cast_fp16 = sqrt(x = var_2348_cast_fp16)[name = tensor("op_2349_cast_fp16")]; + tensor x_263_cast_fp16 = real_div(x = sub_65_cast_fp16, y = var_2349_cast_fp16)[name = tensor("x_263_cast_fp16")]; + tensor var_2351_promoted_to_fp16 = const()[name = tensor("op_2351_promoted_to_fp16"), val = tensor(0x1p+0)]; + tensor var_2352_cast_fp16 = add(x = var_2340_cast_fp16_1, y = var_2351_promoted_to_fp16)[name = tensor("op_2352_cast_fp16")]; + tensor var_2353_cast_fp16 = mul(x = x_263_cast_fp16, y = var_2352_cast_fp16)[name = tensor("op_2353_cast_fp16")]; + tensor input_405_cast_fp16 = add(x = var_2353_cast_fp16, y = var_2340_cast_fp16_0)[name = tensor("input_405_cast_fp16")]; + tensor linear_155_cast_fp16 = linear(bias = flow_net_final_layer_linear_bias_to_fp16, weight = flow_net_final_layer_linear_weight_to_fp16, x = input_405_cast_fp16)[name = tensor("linear_155_cast_fp16")]; + tensor var_2364_to_fp16 = const()[name = tensor("op_2364_to_fp16"), val = tensor(0x1p-3)]; + tensor var_2365_cast_fp16 = mul(x = linear_155_cast_fp16, y = var_2364_to_fp16)[name = tensor("op_2365_cast_fp16")]; + tensor input_407_cast_fp16 = add(x = input_339_cast_fp16, y = var_2365_cast_fp16)[name = tensor("input_407_cast_fp16")]; + tensor linear_156_cast_fp16 = linear(bias = flow_net_input_proj_bias_to_fp16, weight = flow_net_input_proj_weight_to_fp16, x = input_407_cast_fp16)[name = tensor("linear_156_cast_fp16")]; + tensor input_411_to_fp16 = const()[name = tensor("input_411_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(19008256)))]; + tensor input_413_cast_fp16 = silu(x = input_411_to_fp16)[name = tensor("input_413_cast_fp16")]; + tensor linear_158_cast_fp16 = linear(bias = flow_net_time_embed_0_mlp_2_bias_to_fp16, weight = flow_net_time_embed_0_mlp_2_weight_to_fp16, x = input_413_cast_fp16)[name = tensor("linear_158_cast_fp16")]; + tensor reduce_mean_108_axes_0 = const()[name = tensor("reduce_mean_108_axes_0"), val = tensor([-1])]; + tensor reduce_mean_108_keep_dims_0 = const()[name = tensor("reduce_mean_108_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_108_cast_fp16 = reduce_mean(axes = reduce_mean_108_axes_0, keep_dims = reduce_mean_108_keep_dims_0, x = linear_158_cast_fp16)[name = tensor("reduce_mean_108_cast_fp16")]; + tensor sub_66_cast_fp16 = sub(x = linear_158_cast_fp16, y = reduce_mean_108_cast_fp16)[name = tensor("sub_66_cast_fp16")]; + tensor square_54_cast_fp16 = square(x = sub_66_cast_fp16)[name = tensor("square_54_cast_fp16")]; + tensor reduce_mean_109_axes_0 = const()[name = tensor("reduce_mean_109_axes_0"), val = tensor([-1])]; + tensor reduce_mean_109_keep_dims_0 = const()[name = tensor("reduce_mean_109_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_109_cast_fp16 = reduce_mean(axes = reduce_mean_109_axes_0, keep_dims = reduce_mean_109_keep_dims_0, x = square_54_cast_fp16)[name = tensor("reduce_mean_109_cast_fp16")]; + tensor real_div_12_to_fp16 = const()[name = tensor("real_div_12_to_fp16"), val = tensor(0x1.008p+0)]; + tensor mul_12_cast_fp16 = mul(x = reduce_mean_109_cast_fp16, y = real_div_12_to_fp16)[name = tensor("mul_12_cast_fp16")]; + tensor var_2431_to_fp16 = const()[name = tensor("op_2431_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_109_cast_fp16 = add(x = mul_12_cast_fp16, y = var_2431_to_fp16)[name = tensor("var_109_cast_fp16")]; + tensor var_2434_epsilon_0 = const()[name = tensor("op_2434_epsilon_0"), val = tensor(0x1.197998p-40)]; + tensor var_2434_cast_fp16 = rsqrt(epsilon = var_2434_epsilon_0, x = var_109_cast_fp16)[name = tensor("op_2434_cast_fp16")]; + tensor var_2435_cast_fp16 = mul(x = const_3_to_fp16, y = var_2434_cast_fp16)[name = tensor("op_2435_cast_fp16")]; + tensor var_2436_cast_fp16 = mul(x = linear_158_cast_fp16, y = var_2435_cast_fp16)[name = tensor("op_2436_cast_fp16")]; + tensor input_417_to_fp16 = const()[name = tensor("input_417_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(19009344)))]; + tensor input_419_cast_fp16 = silu(x = input_417_to_fp16)[name = tensor("input_419_cast_fp16")]; + tensor linear_160_cast_fp16 = linear(bias = flow_net_time_embed_1_mlp_2_bias_to_fp16, weight = flow_net_time_embed_1_mlp_2_weight_to_fp16, x = input_419_cast_fp16)[name = tensor("linear_160_cast_fp16")]; + tensor reduce_mean_110_axes_0 = const()[name = tensor("reduce_mean_110_axes_0"), val = tensor([-1])]; + tensor reduce_mean_110_keep_dims_0 = const()[name = tensor("reduce_mean_110_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_110_cast_fp16 = reduce_mean(axes = reduce_mean_110_axes_0, keep_dims = reduce_mean_110_keep_dims_0, x = linear_160_cast_fp16)[name = tensor("reduce_mean_110_cast_fp16")]; + tensor sub_68_cast_fp16 = sub(x = linear_160_cast_fp16, y = reduce_mean_110_cast_fp16)[name = tensor("sub_68_cast_fp16")]; + tensor square_55_cast_fp16 = square(x = sub_68_cast_fp16)[name = tensor("square_55_cast_fp16")]; + tensor reduce_mean_111_axes_0 = const()[name = tensor("reduce_mean_111_axes_0"), val = tensor([-1])]; + tensor reduce_mean_111_keep_dims_0 = const()[name = tensor("reduce_mean_111_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_111_cast_fp16 = reduce_mean(axes = reduce_mean_111_axes_0, keep_dims = reduce_mean_111_keep_dims_0, x = square_55_cast_fp16)[name = tensor("reduce_mean_111_cast_fp16")]; + tensor real_div_13_to_fp16 = const()[name = tensor("real_div_13_to_fp16"), val = tensor(0x1.008p+0)]; + tensor mul_13_cast_fp16 = mul(x = reduce_mean_111_cast_fp16, y = real_div_13_to_fp16)[name = tensor("mul_13_cast_fp16")]; + tensor var_2468_to_fp16 = const()[name = tensor("op_2468_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_111_cast_fp16 = add(x = mul_13_cast_fp16, y = var_2468_to_fp16)[name = tensor("var_111_cast_fp16")]; + tensor var_2471_epsilon_0 = const()[name = tensor("op_2471_epsilon_0"), val = tensor(0x1.197998p-40)]; + tensor var_2471_cast_fp16 = rsqrt(epsilon = var_2471_epsilon_0, x = var_111_cast_fp16)[name = tensor("op_2471_cast_fp16")]; + tensor var_2472_cast_fp16 = mul(x = const_5_to_fp16, y = var_2471_cast_fp16)[name = tensor("op_2472_cast_fp16")]; + tensor var_2473_cast_fp16 = mul(x = linear_160_cast_fp16, y = var_2472_cast_fp16)[name = tensor("op_2473_cast_fp16")]; + tensor var_2485_cast_fp16 = add(x = var_2436_cast_fp16, y = var_2473_cast_fp16)[name = tensor("op_2485_cast_fp16")]; + tensor _inversed_t_combined_13_y_0_to_fp16 = const()[name = tensor("_inversed_t_combined_13_y_0_to_fp16"), val = tensor(0x1p-1)]; + tensor _inversed_t_combined_13_cast_fp16 = mul(x = var_2485_cast_fp16, y = _inversed_t_combined_13_y_0_to_fp16)[name = tensor("_inversed_t_combined_13_cast_fp16")]; + tensor input_421_cast_fp16 = add(x = _inversed_t_combined_13_cast_fp16, y = linear_5_cast_fp16)[name = tensor("input_421_cast_fp16")]; + tensor input_423_cast_fp16 = silu(x = input_421_cast_fp16)[name = tensor("input_423_cast_fp16")]; + tensor linear_162_cast_fp16 = linear(bias = flow_net_res_blocks_0_adaLN_modulation_1_bias_to_fp16, weight = flow_net_res_blocks_0_adaLN_modulation_1_weight_to_fp16, x = input_423_cast_fp16)[name = tensor("linear_162_cast_fp16")]; + tensor var_2500_split_sizes_0 = const()[name = tensor("op_2500_split_sizes_0"), val = tensor([512, 512, 512])]; + tensor var_2500_axis_0 = const()[name = tensor("op_2500_axis_0"), val = tensor(-1)]; + tensor var_2500_cast_fp16_0, tensor var_2500_cast_fp16_1, tensor var_2500_cast_fp16_2 = split(axis = var_2500_axis_0, split_sizes = var_2500_split_sizes_0, x = linear_162_cast_fp16)[name = tensor("op_2500_cast_fp16")]; + tensor mean_85_axes_0 = const()[name = tensor("mean_85_axes_0"), val = tensor([-1])]; + tensor mean_85_keep_dims_0 = const()[name = tensor("mean_85_keep_dims_0"), val = tensor(true)]; + tensor mean_85_cast_fp16 = reduce_mean(axes = mean_85_axes_0, keep_dims = mean_85_keep_dims_0, x = linear_156_cast_fp16)[name = tensor("mean_85_cast_fp16")]; + tensor sub_70_cast_fp16 = sub(x = linear_156_cast_fp16, y = mean_85_cast_fp16)[name = tensor("sub_70_cast_fp16")]; + tensor square_56_cast_fp16 = square(x = sub_70_cast_fp16)[name = tensor("square_56_cast_fp16")]; + tensor reduce_mean_113_axes_0 = const()[name = tensor("reduce_mean_113_axes_0"), val = tensor([-1])]; + tensor reduce_mean_113_keep_dims_0 = const()[name = tensor("reduce_mean_113_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_113_cast_fp16 = reduce_mean(axes = reduce_mean_113_axes_0, keep_dims = reduce_mean_113_keep_dims_0, x = square_56_cast_fp16)[name = tensor("reduce_mean_113_cast_fp16")]; + tensor var_2510_to_fp16 = const()[name = tensor("op_2510_to_fp16"), val = tensor(0x1.1p-20)]; + tensor var_2511_cast_fp16 = add(x = reduce_mean_113_cast_fp16, y = var_2510_to_fp16)[name = tensor("op_2511_cast_fp16")]; + tensor var_2512_cast_fp16 = sqrt(x = var_2511_cast_fp16)[name = tensor("op_2512_cast_fp16")]; + tensor x_271_cast_fp16 = real_div(x = sub_70_cast_fp16, y = var_2512_cast_fp16)[name = tensor("x_271_cast_fp16")]; + tensor var_2514_cast_fp16 = mul(x = x_271_cast_fp16, y = flow_net_res_blocks_0_in_ln_weight_to_fp16)[name = tensor("op_2514_cast_fp16")]; + tensor x_273_cast_fp16 = add(x = var_2514_cast_fp16, y = flow_net_res_blocks_0_in_ln_bias_to_fp16)[name = tensor("x_273_cast_fp16")]; + tensor var_2516_promoted_to_fp16 = const()[name = tensor("op_2516_promoted_to_fp16"), val = tensor(0x1p+0)]; + tensor var_2517_cast_fp16 = add(x = var_2500_cast_fp16_1, y = var_2516_promoted_to_fp16)[name = tensor("op_2517_cast_fp16")]; + tensor var_2518_cast_fp16 = mul(x = x_273_cast_fp16, y = var_2517_cast_fp16)[name = tensor("op_2518_cast_fp16")]; + tensor input_425_cast_fp16 = add(x = var_2518_cast_fp16, y = var_2500_cast_fp16_0)[name = tensor("input_425_cast_fp16")]; + tensor linear_163_cast_fp16 = linear(bias = flow_net_res_blocks_0_mlp_0_bias_to_fp16, weight = flow_net_res_blocks_0_mlp_0_weight_to_fp16, x = input_425_cast_fp16)[name = tensor("linear_163_cast_fp16")]; + tensor input_429_cast_fp16 = silu(x = linear_163_cast_fp16)[name = tensor("input_429_cast_fp16")]; + tensor linear_164_cast_fp16 = linear(bias = flow_net_res_blocks_0_mlp_2_bias_to_fp16, weight = flow_net_res_blocks_0_mlp_2_weight_to_fp16, x = input_429_cast_fp16)[name = tensor("linear_164_cast_fp16")]; + tensor var_2529_cast_fp16 = mul(x = var_2500_cast_fp16_2, y = linear_164_cast_fp16)[name = tensor("op_2529_cast_fp16")]; + tensor x_275_cast_fp16 = add(x = linear_156_cast_fp16, y = var_2529_cast_fp16)[name = tensor("x_275_cast_fp16")]; + tensor linear_165_cast_fp16 = linear(bias = flow_net_res_blocks_1_adaLN_modulation_1_bias_to_fp16, weight = flow_net_res_blocks_1_adaLN_modulation_1_weight_to_fp16, x = input_423_cast_fp16)[name = tensor("linear_165_cast_fp16")]; + tensor var_2539_split_sizes_0 = const()[name = tensor("op_2539_split_sizes_0"), val = tensor([512, 512, 512])]; + tensor var_2539_axis_0 = const()[name = tensor("op_2539_axis_0"), val = tensor(-1)]; + tensor var_2539_cast_fp16_0, tensor var_2539_cast_fp16_1, tensor var_2539_cast_fp16_2 = split(axis = var_2539_axis_0, split_sizes = var_2539_split_sizes_0, x = linear_165_cast_fp16)[name = tensor("op_2539_cast_fp16")]; + tensor mean_87_axes_0 = const()[name = tensor("mean_87_axes_0"), val = tensor([-1])]; + tensor mean_87_keep_dims_0 = const()[name = tensor("mean_87_keep_dims_0"), val = tensor(true)]; + tensor mean_87_cast_fp16 = reduce_mean(axes = mean_87_axes_0, keep_dims = mean_87_keep_dims_0, x = x_275_cast_fp16)[name = tensor("mean_87_cast_fp16")]; + tensor sub_71_cast_fp16 = sub(x = x_275_cast_fp16, y = mean_87_cast_fp16)[name = tensor("sub_71_cast_fp16")]; + tensor square_57_cast_fp16 = square(x = sub_71_cast_fp16)[name = tensor("square_57_cast_fp16")]; + tensor reduce_mean_115_axes_0 = const()[name = tensor("reduce_mean_115_axes_0"), val = tensor([-1])]; + tensor reduce_mean_115_keep_dims_0 = const()[name = tensor("reduce_mean_115_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_115_cast_fp16 = reduce_mean(axes = reduce_mean_115_axes_0, keep_dims = reduce_mean_115_keep_dims_0, x = square_57_cast_fp16)[name = tensor("reduce_mean_115_cast_fp16")]; + tensor var_2549_to_fp16 = const()[name = tensor("op_2549_to_fp16"), val = tensor(0x1.1p-20)]; + tensor var_2550_cast_fp16 = add(x = reduce_mean_115_cast_fp16, y = var_2549_to_fp16)[name = tensor("op_2550_cast_fp16")]; + tensor var_2551_cast_fp16 = sqrt(x = var_2550_cast_fp16)[name = tensor("op_2551_cast_fp16")]; + tensor x_277_cast_fp16 = real_div(x = sub_71_cast_fp16, y = var_2551_cast_fp16)[name = tensor("x_277_cast_fp16")]; + tensor var_2553_cast_fp16 = mul(x = x_277_cast_fp16, y = flow_net_res_blocks_1_in_ln_weight_to_fp16)[name = tensor("op_2553_cast_fp16")]; + tensor x_279_cast_fp16 = add(x = var_2553_cast_fp16, y = flow_net_res_blocks_1_in_ln_bias_to_fp16)[name = tensor("x_279_cast_fp16")]; + tensor var_2555_promoted_to_fp16 = const()[name = tensor("op_2555_promoted_to_fp16"), val = tensor(0x1p+0)]; + tensor var_2556_cast_fp16 = add(x = var_2539_cast_fp16_1, y = var_2555_promoted_to_fp16)[name = tensor("op_2556_cast_fp16")]; + tensor var_2557_cast_fp16 = mul(x = x_279_cast_fp16, y = var_2556_cast_fp16)[name = tensor("op_2557_cast_fp16")]; + tensor input_433_cast_fp16 = add(x = var_2557_cast_fp16, y = var_2539_cast_fp16_0)[name = tensor("input_433_cast_fp16")]; + tensor linear_166_cast_fp16 = linear(bias = flow_net_res_blocks_1_mlp_0_bias_to_fp16, weight = flow_net_res_blocks_1_mlp_0_weight_to_fp16, x = input_433_cast_fp16)[name = tensor("linear_166_cast_fp16")]; + tensor input_437_cast_fp16 = silu(x = linear_166_cast_fp16)[name = tensor("input_437_cast_fp16")]; + tensor linear_167_cast_fp16 = linear(bias = flow_net_res_blocks_1_mlp_2_bias_to_fp16, weight = flow_net_res_blocks_1_mlp_2_weight_to_fp16, x = input_437_cast_fp16)[name = tensor("linear_167_cast_fp16")]; + tensor var_2568_cast_fp16 = mul(x = var_2539_cast_fp16_2, y = linear_167_cast_fp16)[name = tensor("op_2568_cast_fp16")]; + tensor x_281_cast_fp16 = add(x = x_275_cast_fp16, y = var_2568_cast_fp16)[name = tensor("x_281_cast_fp16")]; + tensor linear_168_cast_fp16 = linear(bias = flow_net_res_blocks_2_adaLN_modulation_1_bias_to_fp16, weight = flow_net_res_blocks_2_adaLN_modulation_1_weight_to_fp16, x = input_423_cast_fp16)[name = tensor("linear_168_cast_fp16")]; + tensor var_2578_split_sizes_0 = const()[name = tensor("op_2578_split_sizes_0"), val = tensor([512, 512, 512])]; + tensor var_2578_axis_0 = const()[name = tensor("op_2578_axis_0"), val = tensor(-1)]; + tensor var_2578_cast_fp16_0, tensor var_2578_cast_fp16_1, tensor var_2578_cast_fp16_2 = split(axis = var_2578_axis_0, split_sizes = var_2578_split_sizes_0, x = linear_168_cast_fp16)[name = tensor("op_2578_cast_fp16")]; + tensor mean_89_axes_0 = const()[name = tensor("mean_89_axes_0"), val = tensor([-1])]; + tensor mean_89_keep_dims_0 = const()[name = tensor("mean_89_keep_dims_0"), val = tensor(true)]; + tensor mean_89_cast_fp16 = reduce_mean(axes = mean_89_axes_0, keep_dims = mean_89_keep_dims_0, x = x_281_cast_fp16)[name = tensor("mean_89_cast_fp16")]; + tensor sub_72_cast_fp16 = sub(x = x_281_cast_fp16, y = mean_89_cast_fp16)[name = tensor("sub_72_cast_fp16")]; + tensor square_58_cast_fp16 = square(x = sub_72_cast_fp16)[name = tensor("square_58_cast_fp16")]; + tensor reduce_mean_117_axes_0 = const()[name = tensor("reduce_mean_117_axes_0"), val = tensor([-1])]; + tensor reduce_mean_117_keep_dims_0 = const()[name = tensor("reduce_mean_117_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_117_cast_fp16 = reduce_mean(axes = reduce_mean_117_axes_0, keep_dims = reduce_mean_117_keep_dims_0, x = square_58_cast_fp16)[name = tensor("reduce_mean_117_cast_fp16")]; + tensor var_2588_to_fp16 = const()[name = tensor("op_2588_to_fp16"), val = tensor(0x1.1p-20)]; + tensor var_2589_cast_fp16 = add(x = reduce_mean_117_cast_fp16, y = var_2588_to_fp16)[name = tensor("op_2589_cast_fp16")]; + tensor var_2590_cast_fp16 = sqrt(x = var_2589_cast_fp16)[name = tensor("op_2590_cast_fp16")]; + tensor x_283_cast_fp16 = real_div(x = sub_72_cast_fp16, y = var_2590_cast_fp16)[name = tensor("x_283_cast_fp16")]; + tensor var_2592_cast_fp16 = mul(x = x_283_cast_fp16, y = flow_net_res_blocks_2_in_ln_weight_to_fp16)[name = tensor("op_2592_cast_fp16")]; + tensor x_285_cast_fp16 = add(x = var_2592_cast_fp16, y = flow_net_res_blocks_2_in_ln_bias_to_fp16)[name = tensor("x_285_cast_fp16")]; + tensor var_2594_promoted_to_fp16 = const()[name = tensor("op_2594_promoted_to_fp16"), val = tensor(0x1p+0)]; + tensor var_2595_cast_fp16 = add(x = var_2578_cast_fp16_1, y = var_2594_promoted_to_fp16)[name = tensor("op_2595_cast_fp16")]; + tensor var_2596_cast_fp16 = mul(x = x_285_cast_fp16, y = var_2595_cast_fp16)[name = tensor("op_2596_cast_fp16")]; + tensor input_441_cast_fp16 = add(x = var_2596_cast_fp16, y = var_2578_cast_fp16_0)[name = tensor("input_441_cast_fp16")]; + tensor linear_169_cast_fp16 = linear(bias = flow_net_res_blocks_2_mlp_0_bias_to_fp16, weight = flow_net_res_blocks_2_mlp_0_weight_to_fp16, x = input_441_cast_fp16)[name = tensor("linear_169_cast_fp16")]; + tensor input_445_cast_fp16 = silu(x = linear_169_cast_fp16)[name = tensor("input_445_cast_fp16")]; + tensor linear_170_cast_fp16 = linear(bias = flow_net_res_blocks_2_mlp_2_bias_to_fp16, weight = flow_net_res_blocks_2_mlp_2_weight_to_fp16, x = input_445_cast_fp16)[name = tensor("linear_170_cast_fp16")]; + tensor var_2607_cast_fp16 = mul(x = var_2578_cast_fp16_2, y = linear_170_cast_fp16)[name = tensor("op_2607_cast_fp16")]; + tensor x_287_cast_fp16 = add(x = x_281_cast_fp16, y = var_2607_cast_fp16)[name = tensor("x_287_cast_fp16")]; + tensor linear_171_cast_fp16 = linear(bias = flow_net_res_blocks_3_adaLN_modulation_1_bias_to_fp16, weight = flow_net_res_blocks_3_adaLN_modulation_1_weight_to_fp16, x = input_423_cast_fp16)[name = tensor("linear_171_cast_fp16")]; + tensor var_2617_split_sizes_0 = const()[name = tensor("op_2617_split_sizes_0"), val = tensor([512, 512, 512])]; + tensor var_2617_axis_0 = const()[name = tensor("op_2617_axis_0"), val = tensor(-1)]; + tensor var_2617_cast_fp16_0, tensor var_2617_cast_fp16_1, tensor var_2617_cast_fp16_2 = split(axis = var_2617_axis_0, split_sizes = var_2617_split_sizes_0, x = linear_171_cast_fp16)[name = tensor("op_2617_cast_fp16")]; + tensor mean_91_axes_0 = const()[name = tensor("mean_91_axes_0"), val = tensor([-1])]; + tensor mean_91_keep_dims_0 = const()[name = tensor("mean_91_keep_dims_0"), val = tensor(true)]; + tensor mean_91_cast_fp16 = reduce_mean(axes = mean_91_axes_0, keep_dims = mean_91_keep_dims_0, x = x_287_cast_fp16)[name = tensor("mean_91_cast_fp16")]; + tensor sub_73_cast_fp16 = sub(x = x_287_cast_fp16, y = mean_91_cast_fp16)[name = tensor("sub_73_cast_fp16")]; + tensor square_59_cast_fp16 = square(x = sub_73_cast_fp16)[name = tensor("square_59_cast_fp16")]; + tensor reduce_mean_119_axes_0 = const()[name = tensor("reduce_mean_119_axes_0"), val = tensor([-1])]; + tensor reduce_mean_119_keep_dims_0 = const()[name = tensor("reduce_mean_119_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_119_cast_fp16 = reduce_mean(axes = reduce_mean_119_axes_0, keep_dims = reduce_mean_119_keep_dims_0, x = square_59_cast_fp16)[name = tensor("reduce_mean_119_cast_fp16")]; + tensor var_2627_to_fp16 = const()[name = tensor("op_2627_to_fp16"), val = tensor(0x1.1p-20)]; + tensor var_2628_cast_fp16 = add(x = reduce_mean_119_cast_fp16, y = var_2627_to_fp16)[name = tensor("op_2628_cast_fp16")]; + tensor var_2629_cast_fp16 = sqrt(x = var_2628_cast_fp16)[name = tensor("op_2629_cast_fp16")]; + tensor x_289_cast_fp16 = real_div(x = sub_73_cast_fp16, y = var_2629_cast_fp16)[name = tensor("x_289_cast_fp16")]; + tensor var_2631_cast_fp16 = mul(x = x_289_cast_fp16, y = flow_net_res_blocks_3_in_ln_weight_to_fp16)[name = tensor("op_2631_cast_fp16")]; + tensor x_291_cast_fp16 = add(x = var_2631_cast_fp16, y = flow_net_res_blocks_3_in_ln_bias_to_fp16)[name = tensor("x_291_cast_fp16")]; + tensor var_2633_promoted_to_fp16 = const()[name = tensor("op_2633_promoted_to_fp16"), val = tensor(0x1p+0)]; + tensor var_2634_cast_fp16 = add(x = var_2617_cast_fp16_1, y = var_2633_promoted_to_fp16)[name = tensor("op_2634_cast_fp16")]; + tensor var_2635_cast_fp16 = mul(x = x_291_cast_fp16, y = var_2634_cast_fp16)[name = tensor("op_2635_cast_fp16")]; + tensor input_449_cast_fp16 = add(x = var_2635_cast_fp16, y = var_2617_cast_fp16_0)[name = tensor("input_449_cast_fp16")]; + tensor linear_172_cast_fp16 = linear(bias = flow_net_res_blocks_3_mlp_0_bias_to_fp16, weight = flow_net_res_blocks_3_mlp_0_weight_to_fp16, x = input_449_cast_fp16)[name = tensor("linear_172_cast_fp16")]; + tensor input_453_cast_fp16 = silu(x = linear_172_cast_fp16)[name = tensor("input_453_cast_fp16")]; + tensor linear_173_cast_fp16 = linear(bias = flow_net_res_blocks_3_mlp_2_bias_to_fp16, weight = flow_net_res_blocks_3_mlp_2_weight_to_fp16, x = input_453_cast_fp16)[name = tensor("linear_173_cast_fp16")]; + tensor var_2646_cast_fp16 = mul(x = var_2617_cast_fp16_2, y = linear_173_cast_fp16)[name = tensor("op_2646_cast_fp16")]; + tensor x_293_cast_fp16 = add(x = x_287_cast_fp16, y = var_2646_cast_fp16)[name = tensor("x_293_cast_fp16")]; + tensor linear_174_cast_fp16 = linear(bias = flow_net_res_blocks_4_adaLN_modulation_1_bias_to_fp16, weight = flow_net_res_blocks_4_adaLN_modulation_1_weight_to_fp16, x = input_423_cast_fp16)[name = tensor("linear_174_cast_fp16")]; + tensor var_2656_split_sizes_0 = const()[name = tensor("op_2656_split_sizes_0"), val = tensor([512, 512, 512])]; + tensor var_2656_axis_0 = const()[name = tensor("op_2656_axis_0"), val = tensor(-1)]; + tensor var_2656_cast_fp16_0, tensor var_2656_cast_fp16_1, tensor var_2656_cast_fp16_2 = split(axis = var_2656_axis_0, split_sizes = var_2656_split_sizes_0, x = linear_174_cast_fp16)[name = tensor("op_2656_cast_fp16")]; + tensor mean_93_axes_0 = const()[name = tensor("mean_93_axes_0"), val = tensor([-1])]; + tensor mean_93_keep_dims_0 = const()[name = tensor("mean_93_keep_dims_0"), val = tensor(true)]; + tensor mean_93_cast_fp16 = reduce_mean(axes = mean_93_axes_0, keep_dims = mean_93_keep_dims_0, x = x_293_cast_fp16)[name = tensor("mean_93_cast_fp16")]; + tensor sub_74_cast_fp16 = sub(x = x_293_cast_fp16, y = mean_93_cast_fp16)[name = tensor("sub_74_cast_fp16")]; + tensor square_60_cast_fp16 = square(x = sub_74_cast_fp16)[name = tensor("square_60_cast_fp16")]; + tensor reduce_mean_121_axes_0 = const()[name = tensor("reduce_mean_121_axes_0"), val = tensor([-1])]; + tensor reduce_mean_121_keep_dims_0 = const()[name = tensor("reduce_mean_121_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_121_cast_fp16 = reduce_mean(axes = reduce_mean_121_axes_0, keep_dims = reduce_mean_121_keep_dims_0, x = square_60_cast_fp16)[name = tensor("reduce_mean_121_cast_fp16")]; + tensor var_2666_to_fp16 = const()[name = tensor("op_2666_to_fp16"), val = tensor(0x1.1p-20)]; + tensor var_2667_cast_fp16 = add(x = reduce_mean_121_cast_fp16, y = var_2666_to_fp16)[name = tensor("op_2667_cast_fp16")]; + tensor var_2668_cast_fp16 = sqrt(x = var_2667_cast_fp16)[name = tensor("op_2668_cast_fp16")]; + tensor x_295_cast_fp16 = real_div(x = sub_74_cast_fp16, y = var_2668_cast_fp16)[name = tensor("x_295_cast_fp16")]; + tensor var_2670_cast_fp16 = mul(x = x_295_cast_fp16, y = flow_net_res_blocks_4_in_ln_weight_to_fp16)[name = tensor("op_2670_cast_fp16")]; + tensor x_297_cast_fp16 = add(x = var_2670_cast_fp16, y = flow_net_res_blocks_4_in_ln_bias_to_fp16)[name = tensor("x_297_cast_fp16")]; + tensor var_2672_promoted_to_fp16 = const()[name = tensor("op_2672_promoted_to_fp16"), val = tensor(0x1p+0)]; + tensor var_2673_cast_fp16 = add(x = var_2656_cast_fp16_1, y = var_2672_promoted_to_fp16)[name = tensor("op_2673_cast_fp16")]; + tensor var_2674_cast_fp16 = mul(x = x_297_cast_fp16, y = var_2673_cast_fp16)[name = tensor("op_2674_cast_fp16")]; + tensor input_457_cast_fp16 = add(x = var_2674_cast_fp16, y = var_2656_cast_fp16_0)[name = tensor("input_457_cast_fp16")]; + tensor linear_175_cast_fp16 = linear(bias = flow_net_res_blocks_4_mlp_0_bias_to_fp16, weight = flow_net_res_blocks_4_mlp_0_weight_to_fp16, x = input_457_cast_fp16)[name = tensor("linear_175_cast_fp16")]; + tensor input_461_cast_fp16 = silu(x = linear_175_cast_fp16)[name = tensor("input_461_cast_fp16")]; + tensor linear_176_cast_fp16 = linear(bias = flow_net_res_blocks_4_mlp_2_bias_to_fp16, weight = flow_net_res_blocks_4_mlp_2_weight_to_fp16, x = input_461_cast_fp16)[name = tensor("linear_176_cast_fp16")]; + tensor var_2685_cast_fp16 = mul(x = var_2656_cast_fp16_2, y = linear_176_cast_fp16)[name = tensor("op_2685_cast_fp16")]; + tensor x_299_cast_fp16 = add(x = x_293_cast_fp16, y = var_2685_cast_fp16)[name = tensor("x_299_cast_fp16")]; + tensor linear_177_cast_fp16 = linear(bias = flow_net_res_blocks_5_adaLN_modulation_1_bias_to_fp16, weight = flow_net_res_blocks_5_adaLN_modulation_1_weight_to_fp16, x = input_423_cast_fp16)[name = tensor("linear_177_cast_fp16")]; + tensor var_2695_split_sizes_0 = const()[name = tensor("op_2695_split_sizes_0"), val = tensor([512, 512, 512])]; + tensor var_2695_axis_0 = const()[name = tensor("op_2695_axis_0"), val = tensor(-1)]; + tensor var_2695_cast_fp16_0, tensor var_2695_cast_fp16_1, tensor var_2695_cast_fp16_2 = split(axis = var_2695_axis_0, split_sizes = var_2695_split_sizes_0, x = linear_177_cast_fp16)[name = tensor("op_2695_cast_fp16")]; + tensor mean_95_axes_0 = const()[name = tensor("mean_95_axes_0"), val = tensor([-1])]; + tensor mean_95_keep_dims_0 = const()[name = tensor("mean_95_keep_dims_0"), val = tensor(true)]; + tensor mean_95_cast_fp16 = reduce_mean(axes = mean_95_axes_0, keep_dims = mean_95_keep_dims_0, x = x_299_cast_fp16)[name = tensor("mean_95_cast_fp16")]; + tensor sub_75_cast_fp16 = sub(x = x_299_cast_fp16, y = mean_95_cast_fp16)[name = tensor("sub_75_cast_fp16")]; + tensor square_61_cast_fp16 = square(x = sub_75_cast_fp16)[name = tensor("square_61_cast_fp16")]; + tensor reduce_mean_123_axes_0 = const()[name = tensor("reduce_mean_123_axes_0"), val = tensor([-1])]; + tensor reduce_mean_123_keep_dims_0 = const()[name = tensor("reduce_mean_123_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_123_cast_fp16 = reduce_mean(axes = reduce_mean_123_axes_0, keep_dims = reduce_mean_123_keep_dims_0, x = square_61_cast_fp16)[name = tensor("reduce_mean_123_cast_fp16")]; + tensor var_2705_to_fp16 = const()[name = tensor("op_2705_to_fp16"), val = tensor(0x1.1p-20)]; + tensor var_2706_cast_fp16 = add(x = reduce_mean_123_cast_fp16, y = var_2705_to_fp16)[name = tensor("op_2706_cast_fp16")]; + tensor var_2707_cast_fp16 = sqrt(x = var_2706_cast_fp16)[name = tensor("op_2707_cast_fp16")]; + tensor x_301_cast_fp16 = real_div(x = sub_75_cast_fp16, y = var_2707_cast_fp16)[name = tensor("x_301_cast_fp16")]; + tensor var_2709_cast_fp16 = mul(x = x_301_cast_fp16, y = flow_net_res_blocks_5_in_ln_weight_to_fp16)[name = tensor("op_2709_cast_fp16")]; + tensor x_303_cast_fp16 = add(x = var_2709_cast_fp16, y = flow_net_res_blocks_5_in_ln_bias_to_fp16)[name = tensor("x_303_cast_fp16")]; + tensor var_2711_promoted_to_fp16 = const()[name = tensor("op_2711_promoted_to_fp16"), val = tensor(0x1p+0)]; + tensor var_2712_cast_fp16 = add(x = var_2695_cast_fp16_1, y = var_2711_promoted_to_fp16)[name = tensor("op_2712_cast_fp16")]; + tensor var_2713_cast_fp16 = mul(x = x_303_cast_fp16, y = var_2712_cast_fp16)[name = tensor("op_2713_cast_fp16")]; + tensor input_465_cast_fp16 = add(x = var_2713_cast_fp16, y = var_2695_cast_fp16_0)[name = tensor("input_465_cast_fp16")]; + tensor linear_178_cast_fp16 = linear(bias = flow_net_res_blocks_5_mlp_0_bias_to_fp16, weight = flow_net_res_blocks_5_mlp_0_weight_to_fp16, x = input_465_cast_fp16)[name = tensor("linear_178_cast_fp16")]; + tensor input_469_cast_fp16 = silu(x = linear_178_cast_fp16)[name = tensor("input_469_cast_fp16")]; + tensor linear_179_cast_fp16 = linear(bias = flow_net_res_blocks_5_mlp_2_bias_to_fp16, weight = flow_net_res_blocks_5_mlp_2_weight_to_fp16, x = input_469_cast_fp16)[name = tensor("linear_179_cast_fp16")]; + tensor var_2724_cast_fp16 = mul(x = var_2695_cast_fp16_2, y = linear_179_cast_fp16)[name = tensor("op_2724_cast_fp16")]; + tensor x_305_cast_fp16 = add(x = x_299_cast_fp16, y = var_2724_cast_fp16)[name = tensor("x_305_cast_fp16")]; + tensor linear_180_cast_fp16 = linear(bias = flow_net_final_layer_adaLN_modulation_1_bias_to_fp16, weight = flow_net_final_layer_adaLN_modulation_1_weight_to_fp16, x = input_423_cast_fp16)[name = tensor("linear_180_cast_fp16")]; + tensor var_2733_split_sizes_0 = const()[name = tensor("op_2733_split_sizes_0"), val = tensor([512, 512])]; + tensor var_2733_axis_0 = const()[name = tensor("op_2733_axis_0"), val = tensor(-1)]; + tensor var_2733_cast_fp16_0, tensor var_2733_cast_fp16_1 = split(axis = var_2733_axis_0, split_sizes = var_2733_split_sizes_0, x = linear_180_cast_fp16)[name = tensor("op_2733_cast_fp16")]; + tensor mean_97_axes_0 = const()[name = tensor("mean_97_axes_0"), val = tensor([-1])]; + tensor mean_97_keep_dims_0 = const()[name = tensor("mean_97_keep_dims_0"), val = tensor(true)]; + tensor mean_97_cast_fp16 = reduce_mean(axes = mean_97_axes_0, keep_dims = mean_97_keep_dims_0, x = x_305_cast_fp16)[name = tensor("mean_97_cast_fp16")]; + tensor sub_76_cast_fp16 = sub(x = x_305_cast_fp16, y = mean_97_cast_fp16)[name = tensor("sub_76_cast_fp16")]; + tensor square_62_cast_fp16 = square(x = sub_76_cast_fp16)[name = tensor("square_62_cast_fp16")]; + tensor reduce_mean_125_axes_0 = const()[name = tensor("reduce_mean_125_axes_0"), val = tensor([-1])]; + tensor reduce_mean_125_keep_dims_0 = const()[name = tensor("reduce_mean_125_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_125_cast_fp16 = reduce_mean(axes = reduce_mean_125_axes_0, keep_dims = reduce_mean_125_keep_dims_0, x = square_62_cast_fp16)[name = tensor("reduce_mean_125_cast_fp16")]; + tensor var_2740_to_fp16 = const()[name = tensor("op_2740_to_fp16"), val = tensor(0x1.1p-20)]; + tensor var_2741_cast_fp16 = add(x = reduce_mean_125_cast_fp16, y = var_2740_to_fp16)[name = tensor("op_2741_cast_fp16")]; + tensor var_2742_cast_fp16 = sqrt(x = var_2741_cast_fp16)[name = tensor("op_2742_cast_fp16")]; + tensor x_307_cast_fp16 = real_div(x = sub_76_cast_fp16, y = var_2742_cast_fp16)[name = tensor("x_307_cast_fp16")]; + tensor var_2744_promoted_to_fp16 = const()[name = tensor("op_2744_promoted_to_fp16"), val = tensor(0x1p+0)]; + tensor var_2745_cast_fp16 = add(x = var_2733_cast_fp16_1, y = var_2744_promoted_to_fp16)[name = tensor("op_2745_cast_fp16")]; + tensor var_2746_cast_fp16 = mul(x = x_307_cast_fp16, y = var_2745_cast_fp16)[name = tensor("op_2746_cast_fp16")]; + tensor input_473_cast_fp16 = add(x = var_2746_cast_fp16, y = var_2733_cast_fp16_0)[name = tensor("input_473_cast_fp16")]; + tensor linear_181_cast_fp16 = linear(bias = flow_net_final_layer_linear_bias_to_fp16, weight = flow_net_final_layer_linear_weight_to_fp16, x = input_473_cast_fp16)[name = tensor("linear_181_cast_fp16")]; + tensor var_2757_to_fp16 = const()[name = tensor("op_2757_to_fp16"), val = tensor(0x1p-3)]; + tensor var_2758_cast_fp16 = mul(x = linear_181_cast_fp16, y = var_2757_to_fp16)[name = tensor("op_2758_cast_fp16")]; + tensor input_475_cast_fp16 = add(x = input_407_cast_fp16, y = var_2758_cast_fp16)[name = tensor("input_475_cast_fp16")]; + tensor linear_182_cast_fp16 = linear(bias = flow_net_input_proj_bias_to_fp16, weight = flow_net_input_proj_weight_to_fp16, x = input_475_cast_fp16)[name = tensor("linear_182_cast_fp16")]; + tensor input_479_to_fp16 = const()[name = tensor("input_479_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(19010432)))]; + tensor input_481_cast_fp16 = silu(x = input_479_to_fp16)[name = tensor("input_481_cast_fp16")]; + tensor linear_184_cast_fp16 = linear(bias = flow_net_time_embed_0_mlp_2_bias_to_fp16, weight = flow_net_time_embed_0_mlp_2_weight_to_fp16, x = input_481_cast_fp16)[name = tensor("linear_184_cast_fp16")]; + tensor reduce_mean_126_axes_0 = const()[name = tensor("reduce_mean_126_axes_0"), val = tensor([-1])]; + tensor reduce_mean_126_keep_dims_0 = const()[name = tensor("reduce_mean_126_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_126_cast_fp16 = reduce_mean(axes = reduce_mean_126_axes_0, keep_dims = reduce_mean_126_keep_dims_0, x = linear_184_cast_fp16)[name = tensor("reduce_mean_126_cast_fp16")]; + tensor sub_77_cast_fp16 = sub(x = linear_184_cast_fp16, y = reduce_mean_126_cast_fp16)[name = tensor("sub_77_cast_fp16")]; + tensor square_63_cast_fp16 = square(x = sub_77_cast_fp16)[name = tensor("square_63_cast_fp16")]; + tensor reduce_mean_127_axes_0 = const()[name = tensor("reduce_mean_127_axes_0"), val = tensor([-1])]; + tensor reduce_mean_127_keep_dims_0 = const()[name = tensor("reduce_mean_127_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_127_cast_fp16 = reduce_mean(axes = reduce_mean_127_axes_0, keep_dims = reduce_mean_127_keep_dims_0, x = square_63_cast_fp16)[name = tensor("reduce_mean_127_cast_fp16")]; + tensor real_div_14_to_fp16 = const()[name = tensor("real_div_14_to_fp16"), val = tensor(0x1.008p+0)]; + tensor mul_14_cast_fp16 = mul(x = reduce_mean_127_cast_fp16, y = real_div_14_to_fp16)[name = tensor("mul_14_cast_fp16")]; + tensor var_2824_to_fp16 = const()[name = tensor("op_2824_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_127_cast_fp16_0 = add(x = mul_14_cast_fp16, y = var_2824_to_fp16)[name = tensor("var_127_cast_fp16")]; + tensor var_2827_epsilon_0 = const()[name = tensor("op_2827_epsilon_0"), val = tensor(0x1.197998p-40)]; + tensor var_2827_cast_fp16 = rsqrt(epsilon = var_2827_epsilon_0, x = var_127_cast_fp16_0)[name = tensor("op_2827_cast_fp16")]; + tensor var_2828_cast_fp16 = mul(x = const_3_to_fp16, y = var_2827_cast_fp16)[name = tensor("op_2828_cast_fp16")]; + tensor var_2829_cast_fp16 = mul(x = linear_184_cast_fp16, y = var_2828_cast_fp16)[name = tensor("op_2829_cast_fp16")]; + tensor input_485_to_fp16 = const()[name = tensor("input_485_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(19011520)))]; + tensor input_487_cast_fp16 = silu(x = input_485_to_fp16)[name = tensor("input_487_cast_fp16")]; + tensor linear_186_cast_fp16 = linear(bias = flow_net_time_embed_1_mlp_2_bias_to_fp16, weight = flow_net_time_embed_1_mlp_2_weight_to_fp16, x = input_487_cast_fp16)[name = tensor("linear_186_cast_fp16")]; + tensor reduce_mean_128_axes_0 = const()[name = tensor("reduce_mean_128_axes_0"), val = tensor([-1])]; + tensor reduce_mean_128_keep_dims_0 = const()[name = tensor("reduce_mean_128_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_128_cast_fp16 = reduce_mean(axes = reduce_mean_128_axes_0, keep_dims = reduce_mean_128_keep_dims_0, x = linear_186_cast_fp16)[name = tensor("reduce_mean_128_cast_fp16")]; + tensor sub_79_cast_fp16 = sub(x = linear_186_cast_fp16, y = reduce_mean_128_cast_fp16)[name = tensor("sub_79_cast_fp16")]; + tensor square_64_cast_fp16 = square(x = sub_79_cast_fp16)[name = tensor("square_64_cast_fp16")]; + tensor reduce_mean_129_axes_0 = const()[name = tensor("reduce_mean_129_axes_0"), val = tensor([-1])]; + tensor reduce_mean_129_keep_dims_0 = const()[name = tensor("reduce_mean_129_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_129_cast_fp16 = reduce_mean(axes = reduce_mean_129_axes_0, keep_dims = reduce_mean_129_keep_dims_0, x = square_64_cast_fp16)[name = tensor("reduce_mean_129_cast_fp16")]; + tensor real_div_15_to_fp16 = const()[name = tensor("real_div_15_to_fp16"), val = tensor(0x1.008p+0)]; + tensor mul_15_cast_fp16 = mul(x = reduce_mean_129_cast_fp16, y = real_div_15_to_fp16)[name = tensor("mul_15_cast_fp16")]; + tensor var_2851_to_fp16 = const()[name = tensor("op_2851_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_129_cast_fp16 = add(x = mul_15_cast_fp16, y = var_2851_to_fp16)[name = tensor("var_129_cast_fp16")]; + tensor var_2854_epsilon_0 = const()[name = tensor("op_2854_epsilon_0"), val = tensor(0x1.197998p-40)]; + tensor var_2854_cast_fp16 = rsqrt(epsilon = var_2854_epsilon_0, x = var_129_cast_fp16)[name = tensor("op_2854_cast_fp16")]; + tensor var_2855_cast_fp16 = mul(x = const_5_to_fp16, y = var_2854_cast_fp16)[name = tensor("op_2855_cast_fp16")]; + tensor var_2856_cast_fp16 = mul(x = linear_186_cast_fp16, y = var_2855_cast_fp16)[name = tensor("op_2856_cast_fp16")]; + tensor var_2858_cast_fp16 = add(x = var_2829_cast_fp16, y = var_2856_cast_fp16)[name = tensor("op_2858_cast_fp16")]; + tensor _inversed_t_combined_y_0_to_fp16 = const()[name = tensor("_inversed_t_combined_y_0_to_fp16"), val = tensor(0x1p-1)]; + tensor _inversed_t_combined_cast_fp16 = mul(x = var_2858_cast_fp16, y = _inversed_t_combined_y_0_to_fp16)[name = tensor("_inversed_t_combined_cast_fp16")]; + tensor input_489_cast_fp16 = add(x = _inversed_t_combined_cast_fp16, y = linear_5_cast_fp16)[name = tensor("input_489_cast_fp16")]; + tensor input_491_cast_fp16 = silu(x = input_489_cast_fp16)[name = tensor("input_491_cast_fp16")]; + tensor linear_188_cast_fp16 = linear(bias = flow_net_res_blocks_0_adaLN_modulation_1_bias_to_fp16, weight = flow_net_res_blocks_0_adaLN_modulation_1_weight_to_fp16, x = input_491_cast_fp16)[name = tensor("linear_188_cast_fp16")]; + tensor var_2873_split_sizes_0 = const()[name = tensor("op_2873_split_sizes_0"), val = tensor([512, 512, 512])]; + tensor var_2873_axis_0 = const()[name = tensor("op_2873_axis_0"), val = tensor(-1)]; + tensor var_2873_cast_fp16_0, tensor var_2873_cast_fp16_1, tensor var_2873_cast_fp16_2 = split(axis = var_2873_axis_0, split_sizes = var_2873_split_sizes_0, x = linear_188_cast_fp16)[name = tensor("op_2873_cast_fp16")]; + tensor mean_99_axes_0 = const()[name = tensor("mean_99_axes_0"), val = tensor([-1])]; + tensor mean_99_keep_dims_0 = const()[name = tensor("mean_99_keep_dims_0"), val = tensor(true)]; + tensor mean_99_cast_fp16 = reduce_mean(axes = mean_99_axes_0, keep_dims = mean_99_keep_dims_0, x = linear_182_cast_fp16)[name = tensor("mean_99_cast_fp16")]; + tensor sub_81_cast_fp16 = sub(x = linear_182_cast_fp16, y = mean_99_cast_fp16)[name = tensor("sub_81_cast_fp16")]; + tensor square_65_cast_fp16 = square(x = sub_81_cast_fp16)[name = tensor("square_65_cast_fp16")]; + tensor reduce_mean_131_axes_0 = const()[name = tensor("reduce_mean_131_axes_0"), val = tensor([-1])]; + tensor reduce_mean_131_keep_dims_0 = const()[name = tensor("reduce_mean_131_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_131_cast_fp16 = reduce_mean(axes = reduce_mean_131_axes_0, keep_dims = reduce_mean_131_keep_dims_0, x = square_65_cast_fp16)[name = tensor("reduce_mean_131_cast_fp16")]; + tensor var_2883_to_fp16 = const()[name = tensor("op_2883_to_fp16"), val = tensor(0x1.1p-20)]; + tensor var_2884_cast_fp16 = add(x = reduce_mean_131_cast_fp16, y = var_2883_to_fp16)[name = tensor("op_2884_cast_fp16")]; + tensor var_2885_cast_fp16 = sqrt(x = var_2884_cast_fp16)[name = tensor("op_2885_cast_fp16")]; + tensor x_315_cast_fp16 = real_div(x = sub_81_cast_fp16, y = var_2885_cast_fp16)[name = tensor("x_315_cast_fp16")]; + tensor var_2887_cast_fp16 = mul(x = x_315_cast_fp16, y = flow_net_res_blocks_0_in_ln_weight_to_fp16)[name = tensor("op_2887_cast_fp16")]; + tensor x_317_cast_fp16 = add(x = var_2887_cast_fp16, y = flow_net_res_blocks_0_in_ln_bias_to_fp16)[name = tensor("x_317_cast_fp16")]; + tensor var_2889_promoted_to_fp16 = const()[name = tensor("op_2889_promoted_to_fp16"), val = tensor(0x1p+0)]; + tensor var_2890_cast_fp16 = add(x = var_2873_cast_fp16_1, y = var_2889_promoted_to_fp16)[name = tensor("op_2890_cast_fp16")]; + tensor var_2891_cast_fp16 = mul(x = x_317_cast_fp16, y = var_2890_cast_fp16)[name = tensor("op_2891_cast_fp16")]; + tensor input_493_cast_fp16 = add(x = var_2891_cast_fp16, y = var_2873_cast_fp16_0)[name = tensor("input_493_cast_fp16")]; + tensor linear_189_cast_fp16 = linear(bias = flow_net_res_blocks_0_mlp_0_bias_to_fp16, weight = flow_net_res_blocks_0_mlp_0_weight_to_fp16, x = input_493_cast_fp16)[name = tensor("linear_189_cast_fp16")]; + tensor input_497_cast_fp16 = silu(x = linear_189_cast_fp16)[name = tensor("input_497_cast_fp16")]; + tensor linear_190_cast_fp16 = linear(bias = flow_net_res_blocks_0_mlp_2_bias_to_fp16, weight = flow_net_res_blocks_0_mlp_2_weight_to_fp16, x = input_497_cast_fp16)[name = tensor("linear_190_cast_fp16")]; + tensor var_2902_cast_fp16 = mul(x = var_2873_cast_fp16_2, y = linear_190_cast_fp16)[name = tensor("op_2902_cast_fp16")]; + tensor x_319_cast_fp16 = add(x = linear_182_cast_fp16, y = var_2902_cast_fp16)[name = tensor("x_319_cast_fp16")]; + tensor linear_191_cast_fp16 = linear(bias = flow_net_res_blocks_1_adaLN_modulation_1_bias_to_fp16, weight = flow_net_res_blocks_1_adaLN_modulation_1_weight_to_fp16, x = input_491_cast_fp16)[name = tensor("linear_191_cast_fp16")]; + tensor var_2912_split_sizes_0 = const()[name = tensor("op_2912_split_sizes_0"), val = tensor([512, 512, 512])]; + tensor var_2912_axis_0 = const()[name = tensor("op_2912_axis_0"), val = tensor(-1)]; + tensor var_2912_cast_fp16_0, tensor var_2912_cast_fp16_1, tensor var_2912_cast_fp16_2 = split(axis = var_2912_axis_0, split_sizes = var_2912_split_sizes_0, x = linear_191_cast_fp16)[name = tensor("op_2912_cast_fp16")]; + tensor mean_101_axes_0 = const()[name = tensor("mean_101_axes_0"), val = tensor([-1])]; + tensor mean_101_keep_dims_0 = const()[name = tensor("mean_101_keep_dims_0"), val = tensor(true)]; + tensor mean_101_cast_fp16 = reduce_mean(axes = mean_101_axes_0, keep_dims = mean_101_keep_dims_0, x = x_319_cast_fp16)[name = tensor("mean_101_cast_fp16")]; + tensor sub_82_cast_fp16 = sub(x = x_319_cast_fp16, y = mean_101_cast_fp16)[name = tensor("sub_82_cast_fp16")]; + tensor square_66_cast_fp16 = square(x = sub_82_cast_fp16)[name = tensor("square_66_cast_fp16")]; + tensor reduce_mean_133_axes_0 = const()[name = tensor("reduce_mean_133_axes_0"), val = tensor([-1])]; + tensor reduce_mean_133_keep_dims_0 = const()[name = tensor("reduce_mean_133_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_133_cast_fp16 = reduce_mean(axes = reduce_mean_133_axes_0, keep_dims = reduce_mean_133_keep_dims_0, x = square_66_cast_fp16)[name = tensor("reduce_mean_133_cast_fp16")]; + tensor var_2922_to_fp16 = const()[name = tensor("op_2922_to_fp16"), val = tensor(0x1.1p-20)]; + tensor var_2923_cast_fp16 = add(x = reduce_mean_133_cast_fp16, y = var_2922_to_fp16)[name = tensor("op_2923_cast_fp16")]; + tensor var_2924_cast_fp16 = sqrt(x = var_2923_cast_fp16)[name = tensor("op_2924_cast_fp16")]; + tensor x_321_cast_fp16 = real_div(x = sub_82_cast_fp16, y = var_2924_cast_fp16)[name = tensor("x_321_cast_fp16")]; + tensor var_2926_cast_fp16 = mul(x = x_321_cast_fp16, y = flow_net_res_blocks_1_in_ln_weight_to_fp16)[name = tensor("op_2926_cast_fp16")]; + tensor x_323_cast_fp16 = add(x = var_2926_cast_fp16, y = flow_net_res_blocks_1_in_ln_bias_to_fp16)[name = tensor("x_323_cast_fp16")]; + tensor var_2928_promoted_to_fp16 = const()[name = tensor("op_2928_promoted_to_fp16"), val = tensor(0x1p+0)]; + tensor var_2929_cast_fp16 = add(x = var_2912_cast_fp16_1, y = var_2928_promoted_to_fp16)[name = tensor("op_2929_cast_fp16")]; + tensor var_2930_cast_fp16 = mul(x = x_323_cast_fp16, y = var_2929_cast_fp16)[name = tensor("op_2930_cast_fp16")]; + tensor input_501_cast_fp16 = add(x = var_2930_cast_fp16, y = var_2912_cast_fp16_0)[name = tensor("input_501_cast_fp16")]; + tensor linear_192_cast_fp16 = linear(bias = flow_net_res_blocks_1_mlp_0_bias_to_fp16, weight = flow_net_res_blocks_1_mlp_0_weight_to_fp16, x = input_501_cast_fp16)[name = tensor("linear_192_cast_fp16")]; + tensor input_505_cast_fp16 = silu(x = linear_192_cast_fp16)[name = tensor("input_505_cast_fp16")]; + tensor linear_193_cast_fp16 = linear(bias = flow_net_res_blocks_1_mlp_2_bias_to_fp16, weight = flow_net_res_blocks_1_mlp_2_weight_to_fp16, x = input_505_cast_fp16)[name = tensor("linear_193_cast_fp16")]; + tensor var_2941_cast_fp16 = mul(x = var_2912_cast_fp16_2, y = linear_193_cast_fp16)[name = tensor("op_2941_cast_fp16")]; + tensor x_325_cast_fp16 = add(x = x_319_cast_fp16, y = var_2941_cast_fp16)[name = tensor("x_325_cast_fp16")]; + tensor linear_194_cast_fp16 = linear(bias = flow_net_res_blocks_2_adaLN_modulation_1_bias_to_fp16, weight = flow_net_res_blocks_2_adaLN_modulation_1_weight_to_fp16, x = input_491_cast_fp16)[name = tensor("linear_194_cast_fp16")]; + tensor var_2951_split_sizes_0 = const()[name = tensor("op_2951_split_sizes_0"), val = tensor([512, 512, 512])]; + tensor var_2951_axis_0 = const()[name = tensor("op_2951_axis_0"), val = tensor(-1)]; + tensor var_2951_cast_fp16_0, tensor var_2951_cast_fp16_1, tensor var_2951_cast_fp16_2 = split(axis = var_2951_axis_0, split_sizes = var_2951_split_sizes_0, x = linear_194_cast_fp16)[name = tensor("op_2951_cast_fp16")]; + tensor mean_103_axes_0 = const()[name = tensor("mean_103_axes_0"), val = tensor([-1])]; + tensor mean_103_keep_dims_0 = const()[name = tensor("mean_103_keep_dims_0"), val = tensor(true)]; + tensor mean_103_cast_fp16 = reduce_mean(axes = mean_103_axes_0, keep_dims = mean_103_keep_dims_0, x = x_325_cast_fp16)[name = tensor("mean_103_cast_fp16")]; + tensor sub_83_cast_fp16 = sub(x = x_325_cast_fp16, y = mean_103_cast_fp16)[name = tensor("sub_83_cast_fp16")]; + tensor square_67_cast_fp16 = square(x = sub_83_cast_fp16)[name = tensor("square_67_cast_fp16")]; + tensor reduce_mean_135_axes_0 = const()[name = tensor("reduce_mean_135_axes_0"), val = tensor([-1])]; + tensor reduce_mean_135_keep_dims_0 = const()[name = tensor("reduce_mean_135_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_135_cast_fp16 = reduce_mean(axes = reduce_mean_135_axes_0, keep_dims = reduce_mean_135_keep_dims_0, x = square_67_cast_fp16)[name = tensor("reduce_mean_135_cast_fp16")]; + tensor var_2961_to_fp16 = const()[name = tensor("op_2961_to_fp16"), val = tensor(0x1.1p-20)]; + tensor var_2962_cast_fp16 = add(x = reduce_mean_135_cast_fp16, y = var_2961_to_fp16)[name = tensor("op_2962_cast_fp16")]; + tensor var_2963_cast_fp16 = sqrt(x = var_2962_cast_fp16)[name = tensor("op_2963_cast_fp16")]; + tensor x_327_cast_fp16 = real_div(x = sub_83_cast_fp16, y = var_2963_cast_fp16)[name = tensor("x_327_cast_fp16")]; + tensor var_2965_cast_fp16 = mul(x = x_327_cast_fp16, y = flow_net_res_blocks_2_in_ln_weight_to_fp16)[name = tensor("op_2965_cast_fp16")]; + tensor x_329_cast_fp16 = add(x = var_2965_cast_fp16, y = flow_net_res_blocks_2_in_ln_bias_to_fp16)[name = tensor("x_329_cast_fp16")]; + tensor var_2967_promoted_to_fp16 = const()[name = tensor("op_2967_promoted_to_fp16"), val = tensor(0x1p+0)]; + tensor var_2968_cast_fp16 = add(x = var_2951_cast_fp16_1, y = var_2967_promoted_to_fp16)[name = tensor("op_2968_cast_fp16")]; + tensor var_2969_cast_fp16 = mul(x = x_329_cast_fp16, y = var_2968_cast_fp16)[name = tensor("op_2969_cast_fp16")]; + tensor input_509_cast_fp16 = add(x = var_2969_cast_fp16, y = var_2951_cast_fp16_0)[name = tensor("input_509_cast_fp16")]; + tensor linear_195_cast_fp16 = linear(bias = flow_net_res_blocks_2_mlp_0_bias_to_fp16, weight = flow_net_res_blocks_2_mlp_0_weight_to_fp16, x = input_509_cast_fp16)[name = tensor("linear_195_cast_fp16")]; + tensor input_513_cast_fp16 = silu(x = linear_195_cast_fp16)[name = tensor("input_513_cast_fp16")]; + tensor linear_196_cast_fp16 = linear(bias = flow_net_res_blocks_2_mlp_2_bias_to_fp16, weight = flow_net_res_blocks_2_mlp_2_weight_to_fp16, x = input_513_cast_fp16)[name = tensor("linear_196_cast_fp16")]; + tensor var_2980_cast_fp16 = mul(x = var_2951_cast_fp16_2, y = linear_196_cast_fp16)[name = tensor("op_2980_cast_fp16")]; + tensor x_331_cast_fp16 = add(x = x_325_cast_fp16, y = var_2980_cast_fp16)[name = tensor("x_331_cast_fp16")]; + tensor linear_197_cast_fp16 = linear(bias = flow_net_res_blocks_3_adaLN_modulation_1_bias_to_fp16, weight = flow_net_res_blocks_3_adaLN_modulation_1_weight_to_fp16, x = input_491_cast_fp16)[name = tensor("linear_197_cast_fp16")]; + tensor var_2990_split_sizes_0 = const()[name = tensor("op_2990_split_sizes_0"), val = tensor([512, 512, 512])]; + tensor var_2990_axis_0 = const()[name = tensor("op_2990_axis_0"), val = tensor(-1)]; + tensor var_2990_cast_fp16_0, tensor var_2990_cast_fp16_1, tensor var_2990_cast_fp16_2 = split(axis = var_2990_axis_0, split_sizes = var_2990_split_sizes_0, x = linear_197_cast_fp16)[name = tensor("op_2990_cast_fp16")]; + tensor mean_105_axes_0 = const()[name = tensor("mean_105_axes_0"), val = tensor([-1])]; + tensor mean_105_keep_dims_0 = const()[name = tensor("mean_105_keep_dims_0"), val = tensor(true)]; + tensor mean_105_cast_fp16 = reduce_mean(axes = mean_105_axes_0, keep_dims = mean_105_keep_dims_0, x = x_331_cast_fp16)[name = tensor("mean_105_cast_fp16")]; + tensor sub_84_cast_fp16 = sub(x = x_331_cast_fp16, y = mean_105_cast_fp16)[name = tensor("sub_84_cast_fp16")]; + tensor square_68_cast_fp16 = square(x = sub_84_cast_fp16)[name = tensor("square_68_cast_fp16")]; + tensor reduce_mean_137_axes_0 = const()[name = tensor("reduce_mean_137_axes_0"), val = tensor([-1])]; + tensor reduce_mean_137_keep_dims_0 = const()[name = tensor("reduce_mean_137_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_137_cast_fp16 = reduce_mean(axes = reduce_mean_137_axes_0, keep_dims = reduce_mean_137_keep_dims_0, x = square_68_cast_fp16)[name = tensor("reduce_mean_137_cast_fp16")]; + tensor var_3000_to_fp16 = const()[name = tensor("op_3000_to_fp16"), val = tensor(0x1.1p-20)]; + tensor var_3001_cast_fp16 = add(x = reduce_mean_137_cast_fp16, y = var_3000_to_fp16)[name = tensor("op_3001_cast_fp16")]; + tensor var_3002_cast_fp16 = sqrt(x = var_3001_cast_fp16)[name = tensor("op_3002_cast_fp16")]; + tensor x_333_cast_fp16 = real_div(x = sub_84_cast_fp16, y = var_3002_cast_fp16)[name = tensor("x_333_cast_fp16")]; + tensor var_3004_cast_fp16 = mul(x = x_333_cast_fp16, y = flow_net_res_blocks_3_in_ln_weight_to_fp16)[name = tensor("op_3004_cast_fp16")]; + tensor x_335_cast_fp16 = add(x = var_3004_cast_fp16, y = flow_net_res_blocks_3_in_ln_bias_to_fp16)[name = tensor("x_335_cast_fp16")]; + tensor var_3006_promoted_to_fp16 = const()[name = tensor("op_3006_promoted_to_fp16"), val = tensor(0x1p+0)]; + tensor var_3007_cast_fp16 = add(x = var_2990_cast_fp16_1, y = var_3006_promoted_to_fp16)[name = tensor("op_3007_cast_fp16")]; + tensor var_3008_cast_fp16 = mul(x = x_335_cast_fp16, y = var_3007_cast_fp16)[name = tensor("op_3008_cast_fp16")]; + tensor input_517_cast_fp16 = add(x = var_3008_cast_fp16, y = var_2990_cast_fp16_0)[name = tensor("input_517_cast_fp16")]; + tensor linear_198_cast_fp16 = linear(bias = flow_net_res_blocks_3_mlp_0_bias_to_fp16, weight = flow_net_res_blocks_3_mlp_0_weight_to_fp16, x = input_517_cast_fp16)[name = tensor("linear_198_cast_fp16")]; + tensor input_521_cast_fp16 = silu(x = linear_198_cast_fp16)[name = tensor("input_521_cast_fp16")]; + tensor linear_199_cast_fp16 = linear(bias = flow_net_res_blocks_3_mlp_2_bias_to_fp16, weight = flow_net_res_blocks_3_mlp_2_weight_to_fp16, x = input_521_cast_fp16)[name = tensor("linear_199_cast_fp16")]; + tensor var_3019_cast_fp16 = mul(x = var_2990_cast_fp16_2, y = linear_199_cast_fp16)[name = tensor("op_3019_cast_fp16")]; + tensor x_337_cast_fp16 = add(x = x_331_cast_fp16, y = var_3019_cast_fp16)[name = tensor("x_337_cast_fp16")]; + tensor linear_200_cast_fp16 = linear(bias = flow_net_res_blocks_4_adaLN_modulation_1_bias_to_fp16, weight = flow_net_res_blocks_4_adaLN_modulation_1_weight_to_fp16, x = input_491_cast_fp16)[name = tensor("linear_200_cast_fp16")]; + tensor var_3029_split_sizes_0 = const()[name = tensor("op_3029_split_sizes_0"), val = tensor([512, 512, 512])]; + tensor var_3029_axis_0 = const()[name = tensor("op_3029_axis_0"), val = tensor(-1)]; + tensor var_3029_cast_fp16_0, tensor var_3029_cast_fp16_1, tensor var_3029_cast_fp16_2 = split(axis = var_3029_axis_0, split_sizes = var_3029_split_sizes_0, x = linear_200_cast_fp16)[name = tensor("op_3029_cast_fp16")]; + tensor mean_107_axes_0 = const()[name = tensor("mean_107_axes_0"), val = tensor([-1])]; + tensor mean_107_keep_dims_0 = const()[name = tensor("mean_107_keep_dims_0"), val = tensor(true)]; + tensor mean_107_cast_fp16 = reduce_mean(axes = mean_107_axes_0, keep_dims = mean_107_keep_dims_0, x = x_337_cast_fp16)[name = tensor("mean_107_cast_fp16")]; + tensor sub_85_cast_fp16 = sub(x = x_337_cast_fp16, y = mean_107_cast_fp16)[name = tensor("sub_85_cast_fp16")]; + tensor square_69_cast_fp16 = square(x = sub_85_cast_fp16)[name = tensor("square_69_cast_fp16")]; + tensor reduce_mean_139_axes_0 = const()[name = tensor("reduce_mean_139_axes_0"), val = tensor([-1])]; + tensor reduce_mean_139_keep_dims_0 = const()[name = tensor("reduce_mean_139_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_139_cast_fp16 = reduce_mean(axes = reduce_mean_139_axes_0, keep_dims = reduce_mean_139_keep_dims_0, x = square_69_cast_fp16)[name = tensor("reduce_mean_139_cast_fp16")]; + tensor var_3039_to_fp16 = const()[name = tensor("op_3039_to_fp16"), val = tensor(0x1.1p-20)]; + tensor var_3040_cast_fp16 = add(x = reduce_mean_139_cast_fp16, y = var_3039_to_fp16)[name = tensor("op_3040_cast_fp16")]; + tensor var_3041_cast_fp16 = sqrt(x = var_3040_cast_fp16)[name = tensor("op_3041_cast_fp16")]; + tensor x_339_cast_fp16 = real_div(x = sub_85_cast_fp16, y = var_3041_cast_fp16)[name = tensor("x_339_cast_fp16")]; + tensor var_3043_cast_fp16 = mul(x = x_339_cast_fp16, y = flow_net_res_blocks_4_in_ln_weight_to_fp16)[name = tensor("op_3043_cast_fp16")]; + tensor x_341_cast_fp16 = add(x = var_3043_cast_fp16, y = flow_net_res_blocks_4_in_ln_bias_to_fp16)[name = tensor("x_341_cast_fp16")]; + tensor var_3045_promoted_to_fp16 = const()[name = tensor("op_3045_promoted_to_fp16"), val = tensor(0x1p+0)]; + tensor var_3046_cast_fp16 = add(x = var_3029_cast_fp16_1, y = var_3045_promoted_to_fp16)[name = tensor("op_3046_cast_fp16")]; + tensor var_3047_cast_fp16 = mul(x = x_341_cast_fp16, y = var_3046_cast_fp16)[name = tensor("op_3047_cast_fp16")]; + tensor input_525_cast_fp16 = add(x = var_3047_cast_fp16, y = var_3029_cast_fp16_0)[name = tensor("input_525_cast_fp16")]; + tensor linear_201_cast_fp16 = linear(bias = flow_net_res_blocks_4_mlp_0_bias_to_fp16, weight = flow_net_res_blocks_4_mlp_0_weight_to_fp16, x = input_525_cast_fp16)[name = tensor("linear_201_cast_fp16")]; + tensor input_529_cast_fp16 = silu(x = linear_201_cast_fp16)[name = tensor("input_529_cast_fp16")]; + tensor linear_202_cast_fp16 = linear(bias = flow_net_res_blocks_4_mlp_2_bias_to_fp16, weight = flow_net_res_blocks_4_mlp_2_weight_to_fp16, x = input_529_cast_fp16)[name = tensor("linear_202_cast_fp16")]; + tensor var_3058_cast_fp16 = mul(x = var_3029_cast_fp16_2, y = linear_202_cast_fp16)[name = tensor("op_3058_cast_fp16")]; + tensor x_343_cast_fp16 = add(x = x_337_cast_fp16, y = var_3058_cast_fp16)[name = tensor("x_343_cast_fp16")]; + tensor linear_203_cast_fp16 = linear(bias = flow_net_res_blocks_5_adaLN_modulation_1_bias_to_fp16, weight = flow_net_res_blocks_5_adaLN_modulation_1_weight_to_fp16, x = input_491_cast_fp16)[name = tensor("linear_203_cast_fp16")]; + tensor var_3068_split_sizes_0 = const()[name = tensor("op_3068_split_sizes_0"), val = tensor([512, 512, 512])]; + tensor var_3068_axis_0 = const()[name = tensor("op_3068_axis_0"), val = tensor(-1)]; + tensor var_3068_cast_fp16_0, tensor var_3068_cast_fp16_1, tensor var_3068_cast_fp16_2 = split(axis = var_3068_axis_0, split_sizes = var_3068_split_sizes_0, x = linear_203_cast_fp16)[name = tensor("op_3068_cast_fp16")]; + tensor mean_109_axes_0 = const()[name = tensor("mean_109_axes_0"), val = tensor([-1])]; + tensor mean_109_keep_dims_0 = const()[name = tensor("mean_109_keep_dims_0"), val = tensor(true)]; + tensor mean_109_cast_fp16 = reduce_mean(axes = mean_109_axes_0, keep_dims = mean_109_keep_dims_0, x = x_343_cast_fp16)[name = tensor("mean_109_cast_fp16")]; + tensor sub_86_cast_fp16 = sub(x = x_343_cast_fp16, y = mean_109_cast_fp16)[name = tensor("sub_86_cast_fp16")]; + tensor square_70_cast_fp16 = square(x = sub_86_cast_fp16)[name = tensor("square_70_cast_fp16")]; + tensor reduce_mean_141_axes_0 = const()[name = tensor("reduce_mean_141_axes_0"), val = tensor([-1])]; + tensor reduce_mean_141_keep_dims_0 = const()[name = tensor("reduce_mean_141_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_141_cast_fp16 = reduce_mean(axes = reduce_mean_141_axes_0, keep_dims = reduce_mean_141_keep_dims_0, x = square_70_cast_fp16)[name = tensor("reduce_mean_141_cast_fp16")]; + tensor var_3078_to_fp16 = const()[name = tensor("op_3078_to_fp16"), val = tensor(0x1.1p-20)]; + tensor var_3079_cast_fp16 = add(x = reduce_mean_141_cast_fp16, y = var_3078_to_fp16)[name = tensor("op_3079_cast_fp16")]; + tensor var_3080_cast_fp16 = sqrt(x = var_3079_cast_fp16)[name = tensor("op_3080_cast_fp16")]; + tensor x_345_cast_fp16 = real_div(x = sub_86_cast_fp16, y = var_3080_cast_fp16)[name = tensor("x_345_cast_fp16")]; + tensor var_3082_cast_fp16 = mul(x = x_345_cast_fp16, y = flow_net_res_blocks_5_in_ln_weight_to_fp16)[name = tensor("op_3082_cast_fp16")]; + tensor x_347_cast_fp16 = add(x = var_3082_cast_fp16, y = flow_net_res_blocks_5_in_ln_bias_to_fp16)[name = tensor("x_347_cast_fp16")]; + tensor var_3084_promoted_to_fp16 = const()[name = tensor("op_3084_promoted_to_fp16"), val = tensor(0x1p+0)]; + tensor var_3085_cast_fp16 = add(x = var_3068_cast_fp16_1, y = var_3084_promoted_to_fp16)[name = tensor("op_3085_cast_fp16")]; + tensor var_3086_cast_fp16 = mul(x = x_347_cast_fp16, y = var_3085_cast_fp16)[name = tensor("op_3086_cast_fp16")]; + tensor input_533_cast_fp16 = add(x = var_3086_cast_fp16, y = var_3068_cast_fp16_0)[name = tensor("input_533_cast_fp16")]; + tensor linear_204_cast_fp16 = linear(bias = flow_net_res_blocks_5_mlp_0_bias_to_fp16, weight = flow_net_res_blocks_5_mlp_0_weight_to_fp16, x = input_533_cast_fp16)[name = tensor("linear_204_cast_fp16")]; + tensor input_537_cast_fp16 = silu(x = linear_204_cast_fp16)[name = tensor("input_537_cast_fp16")]; + tensor linear_205_cast_fp16 = linear(bias = flow_net_res_blocks_5_mlp_2_bias_to_fp16, weight = flow_net_res_blocks_5_mlp_2_weight_to_fp16, x = input_537_cast_fp16)[name = tensor("linear_205_cast_fp16")]; + tensor var_3097_cast_fp16 = mul(x = var_3068_cast_fp16_2, y = linear_205_cast_fp16)[name = tensor("op_3097_cast_fp16")]; + tensor x_349_cast_fp16 = add(x = x_343_cast_fp16, y = var_3097_cast_fp16)[name = tensor("x_349_cast_fp16")]; + tensor linear_206_cast_fp16 = linear(bias = flow_net_final_layer_adaLN_modulation_1_bias_to_fp16, weight = flow_net_final_layer_adaLN_modulation_1_weight_to_fp16, x = input_491_cast_fp16)[name = tensor("linear_206_cast_fp16")]; + tensor var_3106_split_sizes_0 = const()[name = tensor("op_3106_split_sizes_0"), val = tensor([512, 512])]; + tensor var_3106_axis_0 = const()[name = tensor("op_3106_axis_0"), val = tensor(-1)]; + tensor var_3106_cast_fp16_0, tensor var_3106_cast_fp16_1 = split(axis = var_3106_axis_0, split_sizes = var_3106_split_sizes_0, x = linear_206_cast_fp16)[name = tensor("op_3106_cast_fp16")]; + tensor mean_axes_0 = const()[name = tensor("mean_axes_0"), val = tensor([-1])]; + tensor mean_keep_dims_0 = const()[name = tensor("mean_keep_dims_0"), val = tensor(true)]; + tensor mean_cast_fp16 = reduce_mean(axes = mean_axes_0, keep_dims = mean_keep_dims_0, x = x_349_cast_fp16)[name = tensor("mean_cast_fp16")]; + tensor sub_87_cast_fp16 = sub(x = x_349_cast_fp16, y = mean_cast_fp16)[name = tensor("sub_87_cast_fp16")]; + tensor square_71_cast_fp16 = square(x = sub_87_cast_fp16)[name = tensor("square_71_cast_fp16")]; + tensor reduce_mean_143_axes_0 = const()[name = tensor("reduce_mean_143_axes_0"), val = tensor([-1])]; + tensor reduce_mean_143_keep_dims_0 = const()[name = tensor("reduce_mean_143_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_143_cast_fp16 = reduce_mean(axes = reduce_mean_143_axes_0, keep_dims = reduce_mean_143_keep_dims_0, x = square_71_cast_fp16)[name = tensor("reduce_mean_143_cast_fp16")]; + tensor var_3113_to_fp16 = const()[name = tensor("op_3113_to_fp16"), val = tensor(0x1.1p-20)]; + tensor var_3114_cast_fp16 = add(x = reduce_mean_143_cast_fp16, y = var_3113_to_fp16)[name = tensor("op_3114_cast_fp16")]; + tensor var_3115_cast_fp16 = sqrt(x = var_3114_cast_fp16)[name = tensor("op_3115_cast_fp16")]; + tensor x_cast_fp16 = real_div(x = sub_87_cast_fp16, y = var_3115_cast_fp16)[name = tensor("x_cast_fp16")]; + tensor var_3117_promoted_to_fp16 = const()[name = tensor("op_3117_promoted_to_fp16"), val = tensor(0x1p+0)]; + tensor var_3118_cast_fp16 = add(x = var_3106_cast_fp16_1, y = var_3117_promoted_to_fp16)[name = tensor("op_3118_cast_fp16")]; + tensor var_3119_cast_fp16 = mul(x = x_cast_fp16, y = var_3118_cast_fp16)[name = tensor("op_3119_cast_fp16")]; + tensor input_cast_fp16 = add(x = var_3119_cast_fp16, y = var_3106_cast_fp16_0)[name = tensor("input_cast_fp16")]; + tensor linear_207_cast_fp16 = linear(bias = flow_net_final_layer_linear_bias_to_fp16, weight = flow_net_final_layer_linear_weight_to_fp16, x = input_cast_fp16)[name = tensor("linear_207_cast_fp16")]; + tensor var_3124_to_fp16 = const()[name = tensor("op_3124_to_fp16"), val = tensor(0x1p-3)]; + tensor var_3125_cast_fp16 = mul(x = linear_207_cast_fp16, y = var_3124_to_fp16)[name = tensor("op_3125_cast_fp16")]; + tensor var_3127_cast_fp16 = add(x = input_475_cast_fp16, y = var_3125_cast_fp16)[name = tensor("op_3127_cast_fp16")]; + tensor var_3127_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_3127_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor latent_final = cast(dtype = var_3127_cast_fp16_to_fp32_dtype_0, x = var_3127_cast_fp16)[name = tensor("cast_144")]; + } -> (latent_final); +} \ No newline at end of file diff --git a/v2.1/italian/flow_decoder_fused.mlmodelc/weights/weight.bin b/v2.1/italian/flow_decoder_fused.mlmodelc/weights/weight.bin new file mode 100644 index 0000000000000000000000000000000000000000..8997311285821476f5de08d0ec031ead9ba1afa4 --- /dev/null +++ b/v2.1/italian/flow_decoder_fused.mlmodelc/weights/weight.bin @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:2dba60db47a97f0d1989398ab403f4992fc243f1043965be1124035e821ee0e2 +size 19012608 diff --git a/v2.1/italian/flow_decoder_fused.mlpackage/Data/com.apple.CoreML/model.mlmodel b/v2.1/italian/flow_decoder_fused.mlpackage/Data/com.apple.CoreML/model.mlmodel new file mode 100644 index 0000000000000000000000000000000000000000..e2c5f5c5bdfd83ce359f9e237830daad398bc237 --- /dev/null +++ b/v2.1/italian/flow_decoder_fused.mlpackage/Data/com.apple.CoreML/model.mlmodel @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:63c099ff92b5c965e30ec891f338575e94701a9289be32e2ff02ee46e6307c9d +size 301733 diff --git a/v2.1/italian/flow_decoder_fused.mlpackage/Data/com.apple.CoreML/weights/weight.bin b/v2.1/italian/flow_decoder_fused.mlpackage/Data/com.apple.CoreML/weights/weight.bin new file mode 100644 index 0000000000000000000000000000000000000000..8997311285821476f5de08d0ec031ead9ba1afa4 --- /dev/null +++ b/v2.1/italian/flow_decoder_fused.mlpackage/Data/com.apple.CoreML/weights/weight.bin @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:2dba60db47a97f0d1989398ab403f4992fc243f1043965be1124035e821ee0e2 +size 19012608 diff --git a/v2.1/italian/flow_decoder_fused.mlpackage/Manifest.json b/v2.1/italian/flow_decoder_fused.mlpackage/Manifest.json new file mode 100644 index 0000000000000000000000000000000000000000..65e833b8a37de9986640256e0f1a902855f4619a --- /dev/null +++ b/v2.1/italian/flow_decoder_fused.mlpackage/Manifest.json @@ -0,0 +1,18 @@ +{ + "fileFormatVersion": "1.0.0", + "itemInfoEntries": { + "90DF0D2D-9967-4A66-85B9-F926E25FB3A0": { + "author": "com.apple.CoreML", + "description": "CoreML Model Weights", + "name": "weights", + "path": "com.apple.CoreML/weights" + }, + "ECBE9011-B808-4315-B445-4105F53EAB97": { + "author": "com.apple.CoreML", + "description": "CoreML Model Specification", + "name": "model.mlmodel", + "path": "com.apple.CoreML/model.mlmodel" + } + }, + "rootModelIdentifier": "ECBE9011-B808-4315-B445-4105F53EAB97" +} diff --git a/v2.1/italian/flowlm_step.mlmodelc/analytics/coremldata.bin b/v2.1/italian/flowlm_step.mlmodelc/analytics/coremldata.bin new file mode 100644 index 0000000000000000000000000000000000000000..90776df14f4dd80ec1f812752c6b0d189da4ecd4 --- /dev/null +++ b/v2.1/italian/flowlm_step.mlmodelc/analytics/coremldata.bin @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:b19f58ae2b2bfbb000159702bd0718f444a162a37597ea0dec967630a41aa9c0 +size 243 diff --git a/v2.1/italian/flowlm_step.mlmodelc/coremldata.bin b/v2.1/italian/flowlm_step.mlmodelc/coremldata.bin new file mode 100644 index 0000000000000000000000000000000000000000..477d47128f2993129f90cd19a0dbb463bbfc07ac --- /dev/null +++ b/v2.1/italian/flowlm_step.mlmodelc/coremldata.bin @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:64429c772090ba85ee444be73a602e47b3f4f5fcd38321cd1d38a915b90e2c52 +size 1191 diff --git a/v2.1/italian/flowlm_step.mlmodelc/model.mil b/v2.1/italian/flowlm_step.mlmodelc/model.mil new file mode 100644 index 0000000000000000000000000000000000000000..fca20c00b7767bb216dbcee198438ee24b8bfca5 --- /dev/null +++ b/v2.1/italian/flowlm_step.mlmodelc/model.mil @@ -0,0 +1,1359 @@ +program(1.0) +[buildInfo = dict, tensor>({{"coremlc-component-MIL", "3520.4.1"}, {"coremlc-version", "3520.5.1"}, {"coremltools-component-torch", "2.12.0"}, {"coremltools-source-dialect", "TorchScript"}, {"coremltools-version", "9.0"}})] +{ + func main(tensor bos_emb, tensor cache0, tensor cache1, tensor cache2, tensor cache3, tensor cache4, tensor cache5, tensor position0, tensor position1, tensor position2, tensor position3, tensor position4, tensor position5, tensor sequence) { + tensor sequence_to_fp16_dtype_0 = const()[name = tensor("sequence_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor sequence_to_fp16 = cast(dtype = sequence_to_fp16_dtype_0, x = sequence)[name = tensor("cast_141")]; + tensor var_54_cast_fp16 = not_equal(x = sequence_to_fp16, y = sequence_to_fp16)[name = tensor("op_54_cast_fp16")]; + tensor expand_dims_0_axes_0 = const()[name = tensor("expand_dims_0_axes_0"), val = tensor([0, 1])]; + tensor bos_emb_to_fp16_dtype_0 = const()[name = tensor("bos_emb_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor bos_emb_to_fp16 = cast(dtype = bos_emb_to_fp16_dtype_0, x = bos_emb)[name = tensor("cast_140")]; + tensor expand_dims_0_cast_fp16 = expand_dims(axes = expand_dims_0_axes_0, x = bos_emb_to_fp16)[name = tensor("expand_dims_0_cast_fp16")]; + tensor input_1_cast_fp16 = select(a = expand_dims_0_cast_fp16, b = sequence_to_fp16, cond = var_54_cast_fp16)[name = tensor("input_1_cast_fp16")]; + tensor input_linear_weight_to_fp16 = const()[name = tensor("input_linear_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(64)))]; + tensor linear_0_bias_0_to_fp16 = const()[name = tensor("linear_0_bias_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(65664)))]; + tensor linear_0_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = input_linear_weight_to_fp16, x = input_1_cast_fp16)[name = tensor("linear_0_cast_fp16")]; + tensor input_5_axes_0 = const()[name = tensor("input_5_axes_0"), val = tensor([-1])]; + tensor norm0_1_weight_to_fp16 = const()[name = tensor("norm0_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(67776)))]; + tensor norm0_1_bias_to_fp16 = const()[name = tensor("norm0_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(69888)))]; + tensor var_60_to_fp16 = const()[name = tensor("op_60_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_5_cast_fp16 = layer_norm(axes = input_5_axes_0, beta = norm0_1_bias_to_fp16, epsilon = var_60_to_fp16, gamma = norm0_1_weight_to_fp16, x = linear_0_cast_fp16)[name = tensor("input_5_cast_fp16")]; + tensor attn0_in_proj_weight_to_fp16 = const()[name = tensor("attn0_in_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(72000)))]; + tensor linear_1_bias_0_to_fp16 = const()[name = tensor("linear_1_bias_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(6363520)))]; + tensor linear_1_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = attn0_in_proj_weight_to_fp16, x = input_5_cast_fp16)[name = tensor("linear_1_cast_fp16")]; + tensor var_74 = const()[name = tensor("op_74"), val = tensor([1, 1, 3, 16, 64])]; + tensor qkv_1_cast_fp16 = reshape(shape = var_74, x = linear_1_cast_fp16)[name = tensor("qkv_1_cast_fp16")]; + tensor q_1_begin_0 = const()[name = tensor("q_1_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor q_1_end_0 = const()[name = tensor("q_1_end_0"), val = tensor([1, 1, 1, 16, 64])]; + tensor q_1_end_mask_0 = const()[name = tensor("q_1_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor q_1_squeeze_mask_0 = const()[name = tensor("q_1_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor q_1_cast_fp16 = slice_by_index(begin = q_1_begin_0, end = q_1_end_0, end_mask = q_1_end_mask_0, squeeze_mask = q_1_squeeze_mask_0, x = qkv_1_cast_fp16)[name = tensor("q_1_cast_fp16")]; + tensor k_1_begin_0 = const()[name = tensor("k_1_begin_0"), val = tensor([0, 0, 1, 0, 0])]; + tensor k_1_end_0 = const()[name = tensor("k_1_end_0"), val = tensor([1, 1, 2, 16, 64])]; + tensor k_1_end_mask_0 = const()[name = tensor("k_1_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor k_1_squeeze_mask_0 = const()[name = tensor("k_1_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor k_1_cast_fp16 = slice_by_index(begin = k_1_begin_0, end = k_1_end_0, end_mask = k_1_end_mask_0, squeeze_mask = k_1_squeeze_mask_0, x = qkv_1_cast_fp16)[name = tensor("k_1_cast_fp16")]; + tensor v_1_begin_0 = const()[name = tensor("v_1_begin_0"), val = tensor([0, 0, 2, 0, 0])]; + tensor v_1_end_0 = const()[name = tensor("v_1_end_0"), val = tensor([1, 1, 3, 16, 64])]; + tensor v_1_end_mask_0 = const()[name = tensor("v_1_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor v_1_squeeze_mask_0 = const()[name = tensor("v_1_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor v_1_cast_fp16 = slice_by_index(begin = v_1_begin_0, end = v_1_end_0, end_mask = v_1_end_mask_0, squeeze_mask = v_1_squeeze_mask_0, x = qkv_1_cast_fp16)[name = tensor("v_1_cast_fp16")]; + tensor var_140 = const()[name = tensor("op_140"), val = tensor([1, 1, 1, 1])]; + tensor position0_to_fp16_dtype_0 = const()[name = tensor("position0_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor position0_to_fp16 = cast(dtype = position0_to_fp16_dtype_0, x = position0)[name = tensor("cast_139")]; + tensor ts_5_cast_fp16 = reshape(shape = var_140, x = position0_to_fp16)[name = tensor("ts_5_cast_fp16")]; + tensor var_147 = const()[name = tensor("op_147"), val = tensor([1, 1, 16, 32, 2])]; + tensor q_complex_1_cast_fp16 = reshape(shape = var_147, x = q_1_cast_fp16)[name = tensor("q_complex_1_cast_fp16")]; + tensor var_154 = const()[name = tensor("op_154"), val = tensor([1, 1, 16, 32, 2])]; + tensor k_complex_1_cast_fp16 = reshape(shape = var_154, x = k_1_cast_fp16)[name = tensor("k_complex_1_cast_fp16")]; + tensor var_158_begin_0 = const()[name = tensor("op_158_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_158_end_0 = const()[name = tensor("op_158_end_0"), val = tensor([1, 1, 16, 32, 1])]; + tensor var_158_end_mask_0 = const()[name = tensor("op_158_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_158_squeeze_mask_0 = const()[name = tensor("op_158_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_158_cast_fp16 = slice_by_index(begin = var_158_begin_0, end = var_158_end_0, end_mask = var_158_end_mask_0, squeeze_mask = var_158_squeeze_mask_0, x = q_complex_1_cast_fp16)[name = tensor("op_158_cast_fp16")]; + tensor var_166_begin_0 = const()[name = tensor("op_166_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_166_end_0 = const()[name = tensor("op_166_end_0"), val = tensor([1, 1, 16, 32, 2])]; + tensor var_166_end_mask_0 = const()[name = tensor("op_166_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_166_squeeze_mask_0 = const()[name = tensor("op_166_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_166_cast_fp16 = slice_by_index(begin = var_166_begin_0, end = var_166_end_0, end_mask = var_166_end_mask_0, squeeze_mask = var_166_squeeze_mask_0, x = q_complex_1_cast_fp16)[name = tensor("op_166_cast_fp16")]; + tensor var_174_begin_0 = const()[name = tensor("op_174_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_174_end_0 = const()[name = tensor("op_174_end_0"), val = tensor([1, 1, 16, 32, 1])]; + tensor var_174_end_mask_0 = const()[name = tensor("op_174_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_174_squeeze_mask_0 = const()[name = tensor("op_174_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_174_cast_fp16 = slice_by_index(begin = var_174_begin_0, end = var_174_end_0, end_mask = var_174_end_mask_0, squeeze_mask = var_174_squeeze_mask_0, x = k_complex_1_cast_fp16)[name = tensor("op_174_cast_fp16")]; + tensor var_182_begin_0 = const()[name = tensor("op_182_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_182_end_0 = const()[name = tensor("op_182_end_0"), val = tensor([1, 1, 16, 32, 2])]; + tensor var_182_end_mask_0 = const()[name = tensor("op_182_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_182_squeeze_mask_0 = const()[name = tensor("op_182_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_182_cast_fp16 = slice_by_index(begin = var_182_begin_0, end = var_182_end_0, end_mask = var_182_end_mask_0, squeeze_mask = var_182_squeeze_mask_0, x = k_complex_1_cast_fp16)[name = tensor("op_182_cast_fp16")]; + tensor freqs_1_to_fp16 = const()[name = tensor("freqs_1_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(6369728)))]; + tensor var_188_cast_fp16 = mul(x = freqs_1_to_fp16, y = ts_5_cast_fp16)[name = tensor("op_188_cast_fp16")]; + tensor rotr_1_cast_fp16 = cos(x = var_188_cast_fp16)[name = tensor("rotr_1_cast_fp16")]; + tensor roti_1_cast_fp16 = sin(x = var_188_cast_fp16)[name = tensor("roti_1_cast_fp16")]; + tensor var_192_cast_fp16 = mul(x = var_158_cast_fp16, y = rotr_1_cast_fp16)[name = tensor("op_192_cast_fp16")]; + tensor var_193_cast_fp16 = mul(x = var_166_cast_fp16, y = roti_1_cast_fp16)[name = tensor("op_193_cast_fp16")]; + tensor qor_1_cast_fp16 = sub(x = var_192_cast_fp16, y = var_193_cast_fp16)[name = tensor("qor_1_cast_fp16")]; + tensor var_196_cast_fp16 = mul(x = var_158_cast_fp16, y = roti_1_cast_fp16)[name = tensor("op_196_cast_fp16")]; + tensor var_197_cast_fp16 = mul(x = var_166_cast_fp16, y = rotr_1_cast_fp16)[name = tensor("op_197_cast_fp16")]; + tensor qoi_1_cast_fp16 = add(x = var_196_cast_fp16, y = var_197_cast_fp16)[name = tensor("qoi_1_cast_fp16")]; + tensor var_200_cast_fp16 = mul(x = var_174_cast_fp16, y = rotr_1_cast_fp16)[name = tensor("op_200_cast_fp16")]; + tensor var_201_cast_fp16 = mul(x = var_182_cast_fp16, y = roti_1_cast_fp16)[name = tensor("op_201_cast_fp16")]; + tensor kor_1_cast_fp16 = sub(x = var_200_cast_fp16, y = var_201_cast_fp16)[name = tensor("kor_1_cast_fp16")]; + tensor var_204_cast_fp16 = mul(x = var_174_cast_fp16, y = roti_1_cast_fp16)[name = tensor("op_204_cast_fp16")]; + tensor var_205_cast_fp16 = mul(x = var_182_cast_fp16, y = rotr_1_cast_fp16)[name = tensor("op_205_cast_fp16")]; + tensor koi_1_cast_fp16 = add(x = var_204_cast_fp16, y = var_205_cast_fp16)[name = tensor("koi_1_cast_fp16")]; + tensor qo_1_axis_0 = const()[name = tensor("qo_1_axis_0"), val = tensor(-1)]; + tensor qo_1_cast_fp16 = stack(axis = qo_1_axis_0, values = (qor_1_cast_fp16, qoi_1_cast_fp16))[name = tensor("qo_1_cast_fp16")]; + tensor ko_1_axis_0 = const()[name = tensor("ko_1_axis_0"), val = tensor(-1)]; + tensor ko_1_cast_fp16 = stack(axis = ko_1_axis_0, values = (kor_1_cast_fp16, koi_1_cast_fp16))[name = tensor("ko_1_cast_fp16")]; + tensor var_238 = const()[name = tensor("op_238"), val = tensor([1, 1, 16, 64])]; + tensor q_3_cast_fp16 = reshape(shape = var_238, x = qo_1_cast_fp16)[name = tensor("q_3_cast_fp16")]; + tensor var_244 = const()[name = tensor("op_244"), val = tensor([1, 1, 16, 64])]; + tensor k_3_cast_fp16 = reshape(shape = var_244, x = ko_1_cast_fp16)[name = tensor("k_3_cast_fp16")]; + tensor _inversed_269_y_0_to_fp16 = const()[name = tensor("_inversed_269_y_0_to_fp16"), val = tensor(0x1p-9)]; + tensor _inversed_269_cast_fp16 = mul(x = ts_5_cast_fp16, y = _inversed_269_y_0_to_fp16)[name = tensor("_inversed_269_cast_fp16")]; + tensor var_270_cast_fp16 = floor(x = _inversed_269_cast_fp16)[name = tensor("op_270_cast_fp16")]; + tensor var_271_to_fp16 = const()[name = tensor("op_271_to_fp16"), val = tensor(0x1p+9)]; + tensor var_272_cast_fp16 = mul(x = var_270_cast_fp16, y = var_271_to_fp16)[name = tensor("op_272_cast_fp16")]; + tensor write_indices_float_3_cast_fp16 = sub(x = ts_5_cast_fp16, y = var_272_cast_fp16)[name = tensor("write_indices_float_3_cast_fp16")]; + tensor var_279_dtype_0 = const()[name = tensor("op_279_dtype_0"), val = tensor("int32")]; + tensor write_indices_1_reps_0 = const()[name = tensor("write_indices_1_reps_0"), val = tensor([1, 1, 16, 64])]; + tensor write_indices_float_3_cast_fp16_to_int32 = cast(dtype = var_279_dtype_0, x = write_indices_float_3_cast_fp16)[name = tensor("cast_138")]; + tensor write_indices_1 = tile(reps = write_indices_1_reps_0, x = write_indices_float_3_cast_fp16_to_int32)[name = tensor("write_indices_1")]; + tensor var_289_begin_0 = const()[name = tensor("op_289_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_289_end_0 = const()[name = tensor("op_289_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor var_289_end_mask_0 = const()[name = tensor("op_289_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_289_squeeze_mask_0 = const()[name = tensor("op_289_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor cache0_to_fp16_dtype_0 = const()[name = tensor("cache0_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor cache0_to_fp16 = cast(dtype = cache0_to_fp16_dtype_0, x = cache0)[name = tensor("cast_137")]; + tensor var_289_cast_fp16 = slice_by_index(begin = var_289_begin_0, end = var_289_end_0, end_mask = var_289_end_mask_0, squeeze_mask = var_289_squeeze_mask_0, x = cache0_to_fp16)[name = tensor("op_289_cast_fp16")]; + tensor var_291_axis_0 = const()[name = tensor("op_291_axis_0"), val = tensor(1)]; + tensor var_291_mode_0 = const()[name = tensor("op_291_mode_0"), val = tensor("update")]; + tensor var_291_validate_indices_0 = const()[name = tensor("op_291_validate_indices_0"), val = tensor(false)]; + tensor var_291_cast_fp16 = scatter_along_axis(axis = var_291_axis_0, data = var_289_cast_fp16, indices = write_indices_1, mode = var_291_mode_0, updates = k_3_cast_fp16, validate_indices = var_291_validate_indices_0)[name = tensor("op_291_cast_fp16")]; + tensor concat_2 = const()[name = tensor("concat_2"), val = tensor([0, 0, 0, 0, 0])]; + tensor concat_3 = const()[name = tensor("concat_3"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_1_internal_tensor_assign_1_stride_0 = const()[name = tensor("new_cache_1_internal_tensor_assign_1_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_1_internal_tensor_assign_1_begin_mask_0 = const()[name = tensor("new_cache_1_internal_tensor_assign_1_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_1_internal_tensor_assign_1_end_mask_0 = const()[name = tensor("new_cache_1_internal_tensor_assign_1_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_1_internal_tensor_assign_1_squeeze_mask_0 = const()[name = tensor("new_cache_1_internal_tensor_assign_1_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_12 = const()[name = tensor("shape_12"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_0 = const()[name = tensor("reduce_prod_0"), val = tensor(1048576)]; + tensor range_1d_0_start_0 = const()[name = tensor("range_1d_0_start_0"), val = tensor(0)]; + tensor range_1d_0_step_0 = const()[name = tensor("range_1d_0_step_0"), val = tensor(1)]; + tensor range_1d_0 = range_1d(end = reduce_prod_0, start = range_1d_0_start_0, step = range_1d_0_step_0)[name = tensor("range_1d_0")]; + tensor reshape_0 = reshape(shape = shape_12, x = range_1d_0)[name = tensor("reshape_0")]; + tensor slice_by_index_0 = slice_by_index(begin = concat_2, begin_mask = new_cache_1_internal_tensor_assign_1_begin_mask_0, end = concat_3, end_mask = new_cache_1_internal_tensor_assign_1_end_mask_0, squeeze_mask = new_cache_1_internal_tensor_assign_1_squeeze_mask_0, stride = new_cache_1_internal_tensor_assign_1_stride_0, x = reshape_0)[name = tensor("slice_by_index_0")]; + tensor reshape_1_shape_0 = const()[name = tensor("reshape_1_shape_0"), val = tensor([-1])]; + tensor reshape_1 = reshape(shape = reshape_1_shape_0, x = slice_by_index_0)[name = tensor("reshape_1")]; + tensor reshape_2_shape_0 = const()[name = tensor("reshape_2_shape_0"), val = tensor([-1])]; + tensor reshape_2_cast_fp16 = reshape(shape = reshape_2_shape_0, x = var_291_cast_fp16)[name = tensor("reshape_2_cast_fp16")]; + tensor reshape_3_shape_0 = const()[name = tensor("reshape_3_shape_0"), val = tensor([-1])]; + tensor reshape_3_cast_fp16 = reshape(shape = reshape_3_shape_0, x = cache0_to_fp16)[name = tensor("reshape_3_cast_fp16")]; + tensor scatter_0_mode_0 = const()[name = tensor("scatter_0_mode_0"), val = tensor("update")]; + tensor scatter_0_axis_0 = const()[name = tensor("scatter_0_axis_0"), val = tensor(0)]; + tensor scatter_0_validate_indices_0 = const()[name = tensor("scatter_0_validate_indices_0"), val = tensor(false)]; + tensor scatter_0_cast_fp16 = scatter(axis = scatter_0_axis_0, data = reshape_3_cast_fp16, indices = reshape_1, mode = scatter_0_mode_0, updates = reshape_2_cast_fp16, validate_indices = scatter_0_validate_indices_0)[name = tensor("scatter_0_cast_fp16")]; + tensor reshape_4_cast_fp16 = reshape(shape = shape_12, x = scatter_0_cast_fp16)[name = tensor("reshape_4_cast_fp16")]; + tensor var_299_begin_0 = const()[name = tensor("op_299_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor var_299_end_0 = const()[name = tensor("op_299_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor var_299_end_mask_0 = const()[name = tensor("op_299_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_299_squeeze_mask_0 = const()[name = tensor("op_299_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor var_299_cast_fp16 = slice_by_index(begin = var_299_begin_0, end = var_299_end_0, end_mask = var_299_end_mask_0, squeeze_mask = var_299_squeeze_mask_0, x = reshape_4_cast_fp16)[name = tensor("op_299_cast_fp16")]; + tensor var_301_axis_0 = const()[name = tensor("op_301_axis_0"), val = tensor(1)]; + tensor var_301_mode_0 = const()[name = tensor("op_301_mode_0"), val = tensor("update")]; + tensor var_301_validate_indices_0 = const()[name = tensor("op_301_validate_indices_0"), val = tensor(false)]; + tensor var_301_cast_fp16 = scatter_along_axis(axis = var_301_axis_0, data = var_299_cast_fp16, indices = write_indices_1, mode = var_301_mode_0, updates = v_1_cast_fp16, validate_indices = var_301_validate_indices_0)[name = tensor("op_301_cast_fp16")]; + tensor concat_4 = const()[name = tensor("concat_4"), val = tensor([1, 0, 0, 0, 0])]; + tensor concat_5 = const()[name = tensor("concat_5"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_1_internal_tensor_assign_2_stride_0 = const()[name = tensor("new_cache_1_internal_tensor_assign_2_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_1_internal_tensor_assign_2_begin_mask_0 = const()[name = tensor("new_cache_1_internal_tensor_assign_2_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_1_internal_tensor_assign_2_end_mask_0 = const()[name = tensor("new_cache_1_internal_tensor_assign_2_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_1_internal_tensor_assign_2_squeeze_mask_0 = const()[name = tensor("new_cache_1_internal_tensor_assign_2_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_13 = const()[name = tensor("shape_13"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_1 = const()[name = tensor("reduce_prod_1"), val = tensor(1048576)]; + tensor range_1d_1_start_0 = const()[name = tensor("range_1d_1_start_0"), val = tensor(0)]; + tensor range_1d_1_step_0 = const()[name = tensor("range_1d_1_step_0"), val = tensor(1)]; + tensor range_1d_1 = range_1d(end = reduce_prod_1, start = range_1d_1_start_0, step = range_1d_1_step_0)[name = tensor("range_1d_1")]; + tensor reshape_5 = reshape(shape = shape_13, x = range_1d_1)[name = tensor("reshape_5")]; + tensor slice_by_index_1 = slice_by_index(begin = concat_4, begin_mask = new_cache_1_internal_tensor_assign_2_begin_mask_0, end = concat_5, end_mask = new_cache_1_internal_tensor_assign_2_end_mask_0, squeeze_mask = new_cache_1_internal_tensor_assign_2_squeeze_mask_0, stride = new_cache_1_internal_tensor_assign_2_stride_0, x = reshape_5)[name = tensor("slice_by_index_1")]; + tensor reshape_6_shape_0 = const()[name = tensor("reshape_6_shape_0"), val = tensor([-1])]; + tensor reshape_6 = reshape(shape = reshape_6_shape_0, x = slice_by_index_1)[name = tensor("reshape_6")]; + tensor reshape_7_shape_0 = const()[name = tensor("reshape_7_shape_0"), val = tensor([-1])]; + tensor reshape_7_cast_fp16 = reshape(shape = reshape_7_shape_0, x = var_301_cast_fp16)[name = tensor("reshape_7_cast_fp16")]; + tensor reshape_8_shape_0 = const()[name = tensor("reshape_8_shape_0"), val = tensor([-1])]; + tensor reshape_8_cast_fp16 = reshape(shape = reshape_8_shape_0, x = reshape_4_cast_fp16)[name = tensor("reshape_8_cast_fp16")]; + tensor scatter_1_mode_0 = const()[name = tensor("scatter_1_mode_0"), val = tensor("update")]; + tensor scatter_1_axis_0 = const()[name = tensor("scatter_1_axis_0"), val = tensor(0)]; + tensor scatter_1_validate_indices_0 = const()[name = tensor("scatter_1_validate_indices_0"), val = tensor(false)]; + tensor scatter_1_cast_fp16 = scatter(axis = scatter_1_axis_0, data = reshape_8_cast_fp16, indices = reshape_6, mode = scatter_1_mode_0, updates = reshape_7_cast_fp16, validate_indices = scatter_1_validate_indices_0)[name = tensor("scatter_1_cast_fp16")]; + tensor reshape_9_cast_fp16 = reshape(shape = shape_13, x = scatter_1_cast_fp16)[name = tensor("reshape_9_cast_fp16")]; + tensor reshape_9_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("reshape_9_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor keys_1_begin_0 = const()[name = tensor("keys_1_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor keys_1_end_0 = const()[name = tensor("keys_1_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor keys_1_end_mask_0 = const()[name = tensor("keys_1_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor keys_1_squeeze_mask_0 = const()[name = tensor("keys_1_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor keys_1_cast_fp16 = slice_by_index(begin = keys_1_begin_0, end = keys_1_end_0, end_mask = keys_1_end_mask_0, squeeze_mask = keys_1_squeeze_mask_0, x = reshape_9_cast_fp16)[name = tensor("keys_1_cast_fp16")]; + tensor values_1_begin_0 = const()[name = tensor("values_1_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor values_1_end_0 = const()[name = tensor("values_1_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor values_1_end_mask_0 = const()[name = tensor("values_1_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor values_1_squeeze_mask_0 = const()[name = tensor("values_1_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor values_1_cast_fp16 = slice_by_index(begin = values_1_begin_0, end = values_1_end_0, end_mask = values_1_end_mask_0, squeeze_mask = values_1_squeeze_mask_0, x = reshape_9_cast_fp16)[name = tensor("values_1_cast_fp16")]; + tensor var_313_cast_fp16 = not_equal(x = keys_1_cast_fp16, y = keys_1_cast_fp16)[name = tensor("op_313_cast_fp16")]; + tensor var_319_to_fp16 = const()[name = tensor("op_319_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(6369856)))]; + tensor keys_3_cast_fp16 = select(a = var_319_to_fp16, b = keys_1_cast_fp16, cond = var_313_cast_fp16)[name = tensor("keys_3_cast_fp16")]; + tensor var_321_cast_fp16 = not_equal(x = values_1_cast_fp16, y = values_1_cast_fp16)[name = tensor("op_321_cast_fp16")]; + tensor values_3_cast_fp16 = select(a = var_319_to_fp16, b = values_1_cast_fp16, cond = var_321_cast_fp16)[name = tensor("values_3_cast_fp16")]; + tensor var_345 = const()[name = tensor("op_345"), val = tensor([0, 2, 1, 3])]; + tensor var_361 = const()[name = tensor("op_361"), val = tensor([1, 1, 1])]; + tensor var_362_cast_fp16 = reshape(shape = var_361, x = position0_to_fp16)[name = tensor("op_362_cast_fp16")]; + tensor var_382_to_fp16 = const()[name = tensor("op_382_to_fp16"), val = tensor(0x1p+0)]; + tensor valid_len_1_cast_fp16 = add(x = var_362_cast_fp16, y = var_382_to_fp16)[name = tensor("valid_len_1_cast_fp16")]; + tensor k_positions_1_promoted_to_fp16 = const()[name = tensor("k_positions_1_promoted_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(7418496)))]; + tensor valid_mask_1_cast_fp16 = less(x = k_positions_1_promoted_to_fp16, y = valid_len_1_cast_fp16)[name = tensor("valid_mask_1_cast_fp16")]; + tensor causal_mask_1_cast_fp16 = less_equal(x = k_positions_1_promoted_to_fp16, y = var_362_cast_fp16)[name = tensor("causal_mask_1_cast_fp16")]; + tensor attn_mask_1 = logical_and(x = valid_mask_1_cast_fp16, y = causal_mask_1_cast_fp16)[name = tensor("attn_mask_1")]; + tensor attn_mask_3_axes_0 = const()[name = tensor("attn_mask_3_axes_0"), val = tensor([1])]; + tensor attn_mask_3 = expand_dims(axes = attn_mask_3_axes_0, x = attn_mask_1)[name = tensor("attn_mask_3")]; + tensor var_400_transpose_x_0 = const()[name = tensor("op_400_transpose_x_0"), val = tensor(false)]; + tensor var_400_transpose_y_0 = const()[name = tensor("op_400_transpose_y_0"), val = tensor(false)]; + tensor transpose_18_perm_0 = const()[name = tensor("transpose_18_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_19_perm_0 = const()[name = tensor("transpose_19_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_19 = transpose(perm = transpose_19_perm_0, x = keys_3_cast_fp16)[name = tensor("transpose_51")]; + tensor transpose_18 = transpose(perm = transpose_18_perm_0, x = q_3_cast_fp16)[name = tensor("transpose_52")]; + tensor var_400_cast_fp16 = matmul(transpose_x = var_400_transpose_x_0, transpose_y = var_400_transpose_y_0, x = transpose_18, y = transpose_19)[name = tensor("op_400_cast_fp16")]; + tensor var_394_to_fp16 = const()[name = tensor("op_394_to_fp16"), val = tensor([0x1p-3])]; + tensor attn_weights_1_cast_fp16 = mul(x = var_400_cast_fp16, y = var_394_to_fp16)[name = tensor("attn_weights_1_cast_fp16")]; + tensor var_402 = logical_not(x = attn_mask_3)[name = tensor("op_402")]; + tensor var_403_to_fp16 = const()[name = tensor("op_403_to_fp16"), val = tensor(-inf)]; + tensor attn_weights_3_cast_fp16 = select(a = var_403_to_fp16, b = attn_weights_1_cast_fp16, cond = var_402)[name = tensor("attn_weights_3_cast_fp16")]; + tensor var_405 = const()[name = tensor("op_405"), val = tensor(-1)]; + tensor attn_weights_5_cast_fp16 = softmax(axis = var_405, x = attn_weights_3_cast_fp16)[name = tensor("attn_weights_5_cast_fp16")]; + tensor attn_output_1_transpose_x_0 = const()[name = tensor("attn_output_1_transpose_x_0"), val = tensor(false)]; + tensor attn_output_1_transpose_y_0 = const()[name = tensor("attn_output_1_transpose_y_0"), val = tensor(false)]; + tensor values_5_cast_fp16 = transpose(perm = var_345, x = values_3_cast_fp16)[name = tensor("transpose_53")]; + tensor attn_output_1_cast_fp16 = matmul(transpose_x = attn_output_1_transpose_x_0, transpose_y = attn_output_1_transpose_y_0, x = attn_weights_5_cast_fp16, y = values_5_cast_fp16)[name = tensor("attn_output_1_cast_fp16")]; + tensor var_413 = const()[name = tensor("op_413"), val = tensor([0, 2, 1, 3])]; + tensor var_418 = const()[name = tensor("op_418"), val = tensor([1, 1, 1024])]; + tensor var_414_cast_fp16 = transpose(perm = var_413, x = attn_output_1_cast_fp16)[name = tensor("transpose_50")]; + tensor input_7_cast_fp16 = reshape(shape = var_418, x = var_414_cast_fp16)[name = tensor("input_7_cast_fp16")]; + tensor attn0_out_proj_weight_to_fp16 = const()[name = tensor("attn0_out_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(7419584)))]; + tensor linear_2_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = attn0_out_proj_weight_to_fp16, x = input_7_cast_fp16)[name = tensor("linear_2_cast_fp16")]; + tensor var_424_to_fp16 = const()[name = tensor("op_424_to_fp16"), val = tensor(0x1p+0)]; + tensor var_425_cast_fp16 = add(x = position0_to_fp16, y = var_424_to_fp16)[name = tensor("op_425_cast_fp16")]; + tensor var_425_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_425_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor input_9_cast_fp16 = add(x = linear_0_cast_fp16, y = linear_2_cast_fp16)[name = tensor("input_9_cast_fp16")]; + tensor input_11_axes_0 = const()[name = tensor("input_11_axes_0"), val = tensor([-1])]; + tensor norm0_2_weight_to_fp16 = const()[name = tensor("norm0_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9516800)))]; + tensor norm0_2_bias_to_fp16 = const()[name = tensor("norm0_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9518912)))]; + tensor var_429_to_fp16 = const()[name = tensor("op_429_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_11_cast_fp16 = layer_norm(axes = input_11_axes_0, beta = norm0_2_bias_to_fp16, epsilon = var_429_to_fp16, gamma = norm0_2_weight_to_fp16, x = input_9_cast_fp16)[name = tensor("input_11_cast_fp16")]; + tensor linear0_1_weight_to_fp16 = const()[name = tensor("linear0_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9521024)))]; + tensor linear_3_bias_0_to_fp16 = const()[name = tensor("linear_3_bias_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(17909696)))]; + tensor linear_3_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = linear0_1_weight_to_fp16, x = input_11_cast_fp16)[name = tensor("linear_3_cast_fp16")]; + tensor input_13_mode_0 = const()[name = tensor("input_13_mode_0"), val = tensor("EXACT")]; + tensor input_13_cast_fp16 = gelu(mode = input_13_mode_0, x = linear_3_cast_fp16)[name = tensor("input_13_cast_fp16")]; + tensor linear0_2_weight_to_fp16 = const()[name = tensor("linear0_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(17917952)))]; + tensor linear_4_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = linear0_2_weight_to_fp16, x = input_13_cast_fp16)[name = tensor("linear_4_cast_fp16")]; + tensor input_15_cast_fp16 = add(x = input_9_cast_fp16, y = linear_4_cast_fp16)[name = tensor("input_15_cast_fp16")]; + tensor input_17_axes_0 = const()[name = tensor("input_17_axes_0"), val = tensor([-1])]; + tensor norm1_1_weight_to_fp16 = const()[name = tensor("norm1_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(26306624)))]; + tensor norm1_1_bias_to_fp16 = const()[name = tensor("norm1_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(26308736)))]; + tensor var_446_to_fp16 = const()[name = tensor("op_446_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_17_cast_fp16 = layer_norm(axes = input_17_axes_0, beta = norm1_1_bias_to_fp16, epsilon = var_446_to_fp16, gamma = norm1_1_weight_to_fp16, x = input_15_cast_fp16)[name = tensor("input_17_cast_fp16")]; + tensor attn1_in_proj_weight_to_fp16 = const()[name = tensor("attn1_in_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(26310848)))]; + tensor linear_5_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = attn1_in_proj_weight_to_fp16, x = input_17_cast_fp16)[name = tensor("linear_5_cast_fp16")]; + tensor var_460 = const()[name = tensor("op_460"), val = tensor([1, 1, 3, 16, 64])]; + tensor qkv_3_cast_fp16 = reshape(shape = var_460, x = linear_5_cast_fp16)[name = tensor("qkv_3_cast_fp16")]; + tensor q_7_begin_0 = const()[name = tensor("q_7_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor q_7_end_0 = const()[name = tensor("q_7_end_0"), val = tensor([1, 1, 1, 16, 64])]; + tensor q_7_end_mask_0 = const()[name = tensor("q_7_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor q_7_squeeze_mask_0 = const()[name = tensor("q_7_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor q_7_cast_fp16 = slice_by_index(begin = q_7_begin_0, end = q_7_end_0, end_mask = q_7_end_mask_0, squeeze_mask = q_7_squeeze_mask_0, x = qkv_3_cast_fp16)[name = tensor("q_7_cast_fp16")]; + tensor k_5_begin_0 = const()[name = tensor("k_5_begin_0"), val = tensor([0, 0, 1, 0, 0])]; + tensor k_5_end_0 = const()[name = tensor("k_5_end_0"), val = tensor([1, 1, 2, 16, 64])]; + tensor k_5_end_mask_0 = const()[name = tensor("k_5_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor k_5_squeeze_mask_0 = const()[name = tensor("k_5_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor k_5_cast_fp16 = slice_by_index(begin = k_5_begin_0, end = k_5_end_0, end_mask = k_5_end_mask_0, squeeze_mask = k_5_squeeze_mask_0, x = qkv_3_cast_fp16)[name = tensor("k_5_cast_fp16")]; + tensor v_3_begin_0 = const()[name = tensor("v_3_begin_0"), val = tensor([0, 0, 2, 0, 0])]; + tensor v_3_end_0 = const()[name = tensor("v_3_end_0"), val = tensor([1, 1, 3, 16, 64])]; + tensor v_3_end_mask_0 = const()[name = tensor("v_3_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor v_3_squeeze_mask_0 = const()[name = tensor("v_3_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor v_3_cast_fp16 = slice_by_index(begin = v_3_begin_0, end = v_3_end_0, end_mask = v_3_end_mask_0, squeeze_mask = v_3_squeeze_mask_0, x = qkv_3_cast_fp16)[name = tensor("v_3_cast_fp16")]; + tensor var_526 = const()[name = tensor("op_526"), val = tensor([1, 1, 1, 1])]; + tensor position1_to_fp16_dtype_0 = const()[name = tensor("position1_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor position1_to_fp16 = cast(dtype = position1_to_fp16_dtype_0, x = position1)[name = tensor("cast_134")]; + tensor ts_11_cast_fp16 = reshape(shape = var_526, x = position1_to_fp16)[name = tensor("ts_11_cast_fp16")]; + tensor var_533 = const()[name = tensor("op_533"), val = tensor([1, 1, 16, 32, 2])]; + tensor q_complex_3_cast_fp16 = reshape(shape = var_533, x = q_7_cast_fp16)[name = tensor("q_complex_3_cast_fp16")]; + tensor var_540 = const()[name = tensor("op_540"), val = tensor([1, 1, 16, 32, 2])]; + tensor k_complex_3_cast_fp16 = reshape(shape = var_540, x = k_5_cast_fp16)[name = tensor("k_complex_3_cast_fp16")]; + tensor var_544_begin_0 = const()[name = tensor("op_544_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_544_end_0 = const()[name = tensor("op_544_end_0"), val = tensor([1, 1, 16, 32, 1])]; + tensor var_544_end_mask_0 = const()[name = tensor("op_544_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_544_squeeze_mask_0 = const()[name = tensor("op_544_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_544_cast_fp16 = slice_by_index(begin = var_544_begin_0, end = var_544_end_0, end_mask = var_544_end_mask_0, squeeze_mask = var_544_squeeze_mask_0, x = q_complex_3_cast_fp16)[name = tensor("op_544_cast_fp16")]; + tensor var_552_begin_0 = const()[name = tensor("op_552_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_552_end_0 = const()[name = tensor("op_552_end_0"), val = tensor([1, 1, 16, 32, 2])]; + tensor var_552_end_mask_0 = const()[name = tensor("op_552_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_552_squeeze_mask_0 = const()[name = tensor("op_552_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_552_cast_fp16 = slice_by_index(begin = var_552_begin_0, end = var_552_end_0, end_mask = var_552_end_mask_0, squeeze_mask = var_552_squeeze_mask_0, x = q_complex_3_cast_fp16)[name = tensor("op_552_cast_fp16")]; + tensor var_560_begin_0 = const()[name = tensor("op_560_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_560_end_0 = const()[name = tensor("op_560_end_0"), val = tensor([1, 1, 16, 32, 1])]; + tensor var_560_end_mask_0 = const()[name = tensor("op_560_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_560_squeeze_mask_0 = const()[name = tensor("op_560_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_560_cast_fp16 = slice_by_index(begin = var_560_begin_0, end = var_560_end_0, end_mask = var_560_end_mask_0, squeeze_mask = var_560_squeeze_mask_0, x = k_complex_3_cast_fp16)[name = tensor("op_560_cast_fp16")]; + tensor var_568_begin_0 = const()[name = tensor("op_568_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_568_end_0 = const()[name = tensor("op_568_end_0"), val = tensor([1, 1, 16, 32, 2])]; + tensor var_568_end_mask_0 = const()[name = tensor("op_568_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_568_squeeze_mask_0 = const()[name = tensor("op_568_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_568_cast_fp16 = slice_by_index(begin = var_568_begin_0, end = var_568_end_0, end_mask = var_568_end_mask_0, squeeze_mask = var_568_squeeze_mask_0, x = k_complex_3_cast_fp16)[name = tensor("op_568_cast_fp16")]; + tensor freqs_3_to_fp16 = const()[name = tensor("freqs_3_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(32602368)))]; + tensor var_574_cast_fp16 = mul(x = freqs_3_to_fp16, y = ts_11_cast_fp16)[name = tensor("op_574_cast_fp16")]; + tensor rotr_3_cast_fp16 = cos(x = var_574_cast_fp16)[name = tensor("rotr_3_cast_fp16")]; + tensor roti_3_cast_fp16 = sin(x = var_574_cast_fp16)[name = tensor("roti_3_cast_fp16")]; + tensor var_578_cast_fp16 = mul(x = var_544_cast_fp16, y = rotr_3_cast_fp16)[name = tensor("op_578_cast_fp16")]; + tensor var_579_cast_fp16 = mul(x = var_552_cast_fp16, y = roti_3_cast_fp16)[name = tensor("op_579_cast_fp16")]; + tensor qor_5_cast_fp16 = sub(x = var_578_cast_fp16, y = var_579_cast_fp16)[name = tensor("qor_5_cast_fp16")]; + tensor var_582_cast_fp16 = mul(x = var_544_cast_fp16, y = roti_3_cast_fp16)[name = tensor("op_582_cast_fp16")]; + tensor var_583_cast_fp16 = mul(x = var_552_cast_fp16, y = rotr_3_cast_fp16)[name = tensor("op_583_cast_fp16")]; + tensor qoi_5_cast_fp16 = add(x = var_582_cast_fp16, y = var_583_cast_fp16)[name = tensor("qoi_5_cast_fp16")]; + tensor var_586_cast_fp16 = mul(x = var_560_cast_fp16, y = rotr_3_cast_fp16)[name = tensor("op_586_cast_fp16")]; + tensor var_587_cast_fp16 = mul(x = var_568_cast_fp16, y = roti_3_cast_fp16)[name = tensor("op_587_cast_fp16")]; + tensor kor_5_cast_fp16 = sub(x = var_586_cast_fp16, y = var_587_cast_fp16)[name = tensor("kor_5_cast_fp16")]; + tensor var_590_cast_fp16 = mul(x = var_560_cast_fp16, y = roti_3_cast_fp16)[name = tensor("op_590_cast_fp16")]; + tensor var_591_cast_fp16 = mul(x = var_568_cast_fp16, y = rotr_3_cast_fp16)[name = tensor("op_591_cast_fp16")]; + tensor koi_5_cast_fp16 = add(x = var_590_cast_fp16, y = var_591_cast_fp16)[name = tensor("koi_5_cast_fp16")]; + tensor qo_3_axis_0 = const()[name = tensor("qo_3_axis_0"), val = tensor(-1)]; + tensor qo_3_cast_fp16 = stack(axis = qo_3_axis_0, values = (qor_5_cast_fp16, qoi_5_cast_fp16))[name = tensor("qo_3_cast_fp16")]; + tensor ko_3_axis_0 = const()[name = tensor("ko_3_axis_0"), val = tensor(-1)]; + tensor ko_3_cast_fp16 = stack(axis = ko_3_axis_0, values = (kor_5_cast_fp16, koi_5_cast_fp16))[name = tensor("ko_3_cast_fp16")]; + tensor var_624 = const()[name = tensor("op_624"), val = tensor([1, 1, 16, 64])]; + tensor q_9_cast_fp16 = reshape(shape = var_624, x = qo_3_cast_fp16)[name = tensor("q_9_cast_fp16")]; + tensor var_630 = const()[name = tensor("op_630"), val = tensor([1, 1, 16, 64])]; + tensor k_7_cast_fp16 = reshape(shape = var_630, x = ko_3_cast_fp16)[name = tensor("k_7_cast_fp16")]; + tensor _inversed_655_y_0_to_fp16 = const()[name = tensor("_inversed_655_y_0_to_fp16"), val = tensor(0x1p-9)]; + tensor _inversed_655_cast_fp16 = mul(x = ts_11_cast_fp16, y = _inversed_655_y_0_to_fp16)[name = tensor("_inversed_655_cast_fp16")]; + tensor var_656_cast_fp16 = floor(x = _inversed_655_cast_fp16)[name = tensor("op_656_cast_fp16")]; + tensor var_657_to_fp16 = const()[name = tensor("op_657_to_fp16"), val = tensor(0x1p+9)]; + tensor var_658_cast_fp16 = mul(x = var_656_cast_fp16, y = var_657_to_fp16)[name = tensor("op_658_cast_fp16")]; + tensor write_indices_float_7_cast_fp16 = sub(x = ts_11_cast_fp16, y = var_658_cast_fp16)[name = tensor("write_indices_float_7_cast_fp16")]; + tensor var_665_dtype_0 = const()[name = tensor("op_665_dtype_0"), val = tensor("int32")]; + tensor write_indices_3_reps_0 = const()[name = tensor("write_indices_3_reps_0"), val = tensor([1, 1, 16, 64])]; + tensor write_indices_float_7_cast_fp16_to_int32 = cast(dtype = var_665_dtype_0, x = write_indices_float_7_cast_fp16)[name = tensor("cast_133")]; + tensor write_indices_3 = tile(reps = write_indices_3_reps_0, x = write_indices_float_7_cast_fp16_to_int32)[name = tensor("write_indices_3")]; + tensor var_675_begin_0 = const()[name = tensor("op_675_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_675_end_0 = const()[name = tensor("op_675_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor var_675_end_mask_0 = const()[name = tensor("op_675_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_675_squeeze_mask_0 = const()[name = tensor("op_675_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor cache1_to_fp16_dtype_0 = const()[name = tensor("cache1_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor cache1_to_fp16 = cast(dtype = cache1_to_fp16_dtype_0, x = cache1)[name = tensor("cast_132")]; + tensor var_675_cast_fp16 = slice_by_index(begin = var_675_begin_0, end = var_675_end_0, end_mask = var_675_end_mask_0, squeeze_mask = var_675_squeeze_mask_0, x = cache1_to_fp16)[name = tensor("op_675_cast_fp16")]; + tensor var_677_axis_0 = const()[name = tensor("op_677_axis_0"), val = tensor(1)]; + tensor var_677_mode_0 = const()[name = tensor("op_677_mode_0"), val = tensor("update")]; + tensor var_677_validate_indices_0 = const()[name = tensor("op_677_validate_indices_0"), val = tensor(false)]; + tensor var_677_cast_fp16 = scatter_along_axis(axis = var_677_axis_0, data = var_675_cast_fp16, indices = write_indices_3, mode = var_677_mode_0, updates = k_7_cast_fp16, validate_indices = var_677_validate_indices_0)[name = tensor("op_677_cast_fp16")]; + tensor concat_9 = const()[name = tensor("concat_9"), val = tensor([0, 0, 0, 0, 0])]; + tensor concat_10 = const()[name = tensor("concat_10"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_3_internal_tensor_assign_1_stride_0 = const()[name = tensor("new_cache_3_internal_tensor_assign_1_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_3_internal_tensor_assign_1_begin_mask_0 = const()[name = tensor("new_cache_3_internal_tensor_assign_1_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_3_internal_tensor_assign_1_end_mask_0 = const()[name = tensor("new_cache_3_internal_tensor_assign_1_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_3_internal_tensor_assign_1_squeeze_mask_0 = const()[name = tensor("new_cache_3_internal_tensor_assign_1_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_14 = const()[name = tensor("shape_14"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_2 = const()[name = tensor("reduce_prod_2"), val = tensor(1048576)]; + tensor range_1d_2_start_0 = const()[name = tensor("range_1d_2_start_0"), val = tensor(0)]; + tensor range_1d_2_step_0 = const()[name = tensor("range_1d_2_step_0"), val = tensor(1)]; + tensor range_1d_2 = range_1d(end = reduce_prod_2, start = range_1d_2_start_0, step = range_1d_2_step_0)[name = tensor("range_1d_2")]; + tensor reshape_10 = reshape(shape = shape_14, x = range_1d_2)[name = tensor("reshape_10")]; + tensor slice_by_index_2 = slice_by_index(begin = concat_9, begin_mask = new_cache_3_internal_tensor_assign_1_begin_mask_0, end = concat_10, end_mask = new_cache_3_internal_tensor_assign_1_end_mask_0, squeeze_mask = new_cache_3_internal_tensor_assign_1_squeeze_mask_0, stride = new_cache_3_internal_tensor_assign_1_stride_0, x = reshape_10)[name = tensor("slice_by_index_2")]; + tensor reshape_11_shape_0 = const()[name = tensor("reshape_11_shape_0"), val = tensor([-1])]; + tensor reshape_11 = reshape(shape = reshape_11_shape_0, x = slice_by_index_2)[name = tensor("reshape_11")]; + tensor reshape_12_shape_0 = const()[name = tensor("reshape_12_shape_0"), val = tensor([-1])]; + tensor reshape_12_cast_fp16 = reshape(shape = reshape_12_shape_0, x = var_677_cast_fp16)[name = tensor("reshape_12_cast_fp16")]; + tensor reshape_13_shape_0 = const()[name = tensor("reshape_13_shape_0"), val = tensor([-1])]; + tensor reshape_13_cast_fp16 = reshape(shape = reshape_13_shape_0, x = cache1_to_fp16)[name = tensor("reshape_13_cast_fp16")]; + tensor scatter_2_mode_0 = const()[name = tensor("scatter_2_mode_0"), val = tensor("update")]; + tensor scatter_2_axis_0 = const()[name = tensor("scatter_2_axis_0"), val = tensor(0)]; + tensor scatter_2_validate_indices_0 = const()[name = tensor("scatter_2_validate_indices_0"), val = tensor(false)]; + tensor scatter_2_cast_fp16 = scatter(axis = scatter_2_axis_0, data = reshape_13_cast_fp16, indices = reshape_11, mode = scatter_2_mode_0, updates = reshape_12_cast_fp16, validate_indices = scatter_2_validate_indices_0)[name = tensor("scatter_2_cast_fp16")]; + tensor reshape_14_cast_fp16 = reshape(shape = shape_14, x = scatter_2_cast_fp16)[name = tensor("reshape_14_cast_fp16")]; + tensor var_685_begin_0 = const()[name = tensor("op_685_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor var_685_end_0 = const()[name = tensor("op_685_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor var_685_end_mask_0 = const()[name = tensor("op_685_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_685_squeeze_mask_0 = const()[name = tensor("op_685_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor var_685_cast_fp16 = slice_by_index(begin = var_685_begin_0, end = var_685_end_0, end_mask = var_685_end_mask_0, squeeze_mask = var_685_squeeze_mask_0, x = reshape_14_cast_fp16)[name = tensor("op_685_cast_fp16")]; + tensor var_687_axis_0 = const()[name = tensor("op_687_axis_0"), val = tensor(1)]; + tensor var_687_mode_0 = const()[name = tensor("op_687_mode_0"), val = tensor("update")]; + tensor var_687_validate_indices_0 = const()[name = tensor("op_687_validate_indices_0"), val = tensor(false)]; + tensor var_687_cast_fp16 = scatter_along_axis(axis = var_687_axis_0, data = var_685_cast_fp16, indices = write_indices_3, mode = var_687_mode_0, updates = v_3_cast_fp16, validate_indices = var_687_validate_indices_0)[name = tensor("op_687_cast_fp16")]; + tensor concat_11 = const()[name = tensor("concat_11"), val = tensor([1, 0, 0, 0, 0])]; + tensor concat_12 = const()[name = tensor("concat_12"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_3_internal_tensor_assign_2_stride_0 = const()[name = tensor("new_cache_3_internal_tensor_assign_2_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_3_internal_tensor_assign_2_begin_mask_0 = const()[name = tensor("new_cache_3_internal_tensor_assign_2_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_3_internal_tensor_assign_2_end_mask_0 = const()[name = tensor("new_cache_3_internal_tensor_assign_2_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_3_internal_tensor_assign_2_squeeze_mask_0 = const()[name = tensor("new_cache_3_internal_tensor_assign_2_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_15 = const()[name = tensor("shape_15"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_3 = const()[name = tensor("reduce_prod_3"), val = tensor(1048576)]; + tensor range_1d_3_start_0 = const()[name = tensor("range_1d_3_start_0"), val = tensor(0)]; + tensor range_1d_3_step_0 = const()[name = tensor("range_1d_3_step_0"), val = tensor(1)]; + tensor range_1d_3 = range_1d(end = reduce_prod_3, start = range_1d_3_start_0, step = range_1d_3_step_0)[name = tensor("range_1d_3")]; + tensor reshape_15 = reshape(shape = shape_15, x = range_1d_3)[name = tensor("reshape_15")]; + tensor slice_by_index_3 = slice_by_index(begin = concat_11, begin_mask = new_cache_3_internal_tensor_assign_2_begin_mask_0, end = concat_12, end_mask = new_cache_3_internal_tensor_assign_2_end_mask_0, squeeze_mask = new_cache_3_internal_tensor_assign_2_squeeze_mask_0, stride = new_cache_3_internal_tensor_assign_2_stride_0, x = reshape_15)[name = tensor("slice_by_index_3")]; + tensor reshape_16_shape_0 = const()[name = tensor("reshape_16_shape_0"), val = tensor([-1])]; + tensor reshape_16 = reshape(shape = reshape_16_shape_0, x = slice_by_index_3)[name = tensor("reshape_16")]; + tensor reshape_17_shape_0 = const()[name = tensor("reshape_17_shape_0"), val = tensor([-1])]; + tensor reshape_17_cast_fp16 = reshape(shape = reshape_17_shape_0, x = var_687_cast_fp16)[name = tensor("reshape_17_cast_fp16")]; + tensor reshape_18_shape_0 = const()[name = tensor("reshape_18_shape_0"), val = tensor([-1])]; + tensor reshape_18_cast_fp16 = reshape(shape = reshape_18_shape_0, x = reshape_14_cast_fp16)[name = tensor("reshape_18_cast_fp16")]; + tensor scatter_3_mode_0 = const()[name = tensor("scatter_3_mode_0"), val = tensor("update")]; + tensor scatter_3_axis_0 = const()[name = tensor("scatter_3_axis_0"), val = tensor(0)]; + tensor scatter_3_validate_indices_0 = const()[name = tensor("scatter_3_validate_indices_0"), val = tensor(false)]; + tensor scatter_3_cast_fp16 = scatter(axis = scatter_3_axis_0, data = reshape_18_cast_fp16, indices = reshape_16, mode = scatter_3_mode_0, updates = reshape_17_cast_fp16, validate_indices = scatter_3_validate_indices_0)[name = tensor("scatter_3_cast_fp16")]; + tensor reshape_19_cast_fp16 = reshape(shape = shape_15, x = scatter_3_cast_fp16)[name = tensor("reshape_19_cast_fp16")]; + tensor reshape_19_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("reshape_19_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor keys_7_begin_0 = const()[name = tensor("keys_7_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor keys_7_end_0 = const()[name = tensor("keys_7_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor keys_7_end_mask_0 = const()[name = tensor("keys_7_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor keys_7_squeeze_mask_0 = const()[name = tensor("keys_7_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor keys_7_cast_fp16 = slice_by_index(begin = keys_7_begin_0, end = keys_7_end_0, end_mask = keys_7_end_mask_0, squeeze_mask = keys_7_squeeze_mask_0, x = reshape_19_cast_fp16)[name = tensor("keys_7_cast_fp16")]; + tensor values_7_begin_0 = const()[name = tensor("values_7_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor values_7_end_0 = const()[name = tensor("values_7_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor values_7_end_mask_0 = const()[name = tensor("values_7_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor values_7_squeeze_mask_0 = const()[name = tensor("values_7_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor values_7_cast_fp16 = slice_by_index(begin = values_7_begin_0, end = values_7_end_0, end_mask = values_7_end_mask_0, squeeze_mask = values_7_squeeze_mask_0, x = reshape_19_cast_fp16)[name = tensor("values_7_cast_fp16")]; + tensor var_699_cast_fp16 = not_equal(x = keys_7_cast_fp16, y = keys_7_cast_fp16)[name = tensor("op_699_cast_fp16")]; + tensor keys_9_cast_fp16 = select(a = var_319_to_fp16, b = keys_7_cast_fp16, cond = var_699_cast_fp16)[name = tensor("keys_9_cast_fp16")]; + tensor var_707_cast_fp16 = not_equal(x = values_7_cast_fp16, y = values_7_cast_fp16)[name = tensor("op_707_cast_fp16")]; + tensor values_9_cast_fp16 = select(a = var_319_to_fp16, b = values_7_cast_fp16, cond = var_707_cast_fp16)[name = tensor("values_9_cast_fp16")]; + tensor var_731 = const()[name = tensor("op_731"), val = tensor([0, 2, 1, 3])]; + tensor var_747 = const()[name = tensor("op_747"), val = tensor([1, 1, 1])]; + tensor var_748_cast_fp16 = reshape(shape = var_747, x = position1_to_fp16)[name = tensor("op_748_cast_fp16")]; + tensor var_768_to_fp16 = const()[name = tensor("op_768_to_fp16"), val = tensor(0x1p+0)]; + tensor valid_len_3_cast_fp16 = add(x = var_748_cast_fp16, y = var_768_to_fp16)[name = tensor("valid_len_3_cast_fp16")]; + tensor valid_mask_3_cast_fp16 = less(x = k_positions_1_promoted_to_fp16, y = valid_len_3_cast_fp16)[name = tensor("valid_mask_3_cast_fp16")]; + tensor causal_mask_3_cast_fp16 = less_equal(x = k_positions_1_promoted_to_fp16, y = var_748_cast_fp16)[name = tensor("causal_mask_3_cast_fp16")]; + tensor attn_mask_5 = logical_and(x = valid_mask_3_cast_fp16, y = causal_mask_3_cast_fp16)[name = tensor("attn_mask_5")]; + tensor attn_mask_7_axes_0 = const()[name = tensor("attn_mask_7_axes_0"), val = tensor([1])]; + tensor attn_mask_7 = expand_dims(axes = attn_mask_7_axes_0, x = attn_mask_5)[name = tensor("attn_mask_7")]; + tensor var_786_transpose_x_0 = const()[name = tensor("op_786_transpose_x_0"), val = tensor(false)]; + tensor var_786_transpose_y_0 = const()[name = tensor("op_786_transpose_y_0"), val = tensor(false)]; + tensor transpose_20_perm_0 = const()[name = tensor("transpose_20_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_21_perm_0 = const()[name = tensor("transpose_21_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_21 = transpose(perm = transpose_21_perm_0, x = keys_9_cast_fp16)[name = tensor("transpose_47")]; + tensor transpose_20 = transpose(perm = transpose_20_perm_0, x = q_9_cast_fp16)[name = tensor("transpose_48")]; + tensor var_786_cast_fp16 = matmul(transpose_x = var_786_transpose_x_0, transpose_y = var_786_transpose_y_0, x = transpose_20, y = transpose_21)[name = tensor("op_786_cast_fp16")]; + tensor var_780_to_fp16 = const()[name = tensor("op_780_to_fp16"), val = tensor([0x1p-3])]; + tensor attn_weights_7_cast_fp16 = mul(x = var_786_cast_fp16, y = var_780_to_fp16)[name = tensor("attn_weights_7_cast_fp16")]; + tensor var_788 = logical_not(x = attn_mask_7)[name = tensor("op_788")]; + tensor var_789_to_fp16 = const()[name = tensor("op_789_to_fp16"), val = tensor(-inf)]; + tensor attn_weights_9_cast_fp16 = select(a = var_789_to_fp16, b = attn_weights_7_cast_fp16, cond = var_788)[name = tensor("attn_weights_9_cast_fp16")]; + tensor var_791 = const()[name = tensor("op_791"), val = tensor(-1)]; + tensor attn_weights_11_cast_fp16 = softmax(axis = var_791, x = attn_weights_9_cast_fp16)[name = tensor("attn_weights_11_cast_fp16")]; + tensor attn_output_3_transpose_x_0 = const()[name = tensor("attn_output_3_transpose_x_0"), val = tensor(false)]; + tensor attn_output_3_transpose_y_0 = const()[name = tensor("attn_output_3_transpose_y_0"), val = tensor(false)]; + tensor values_11_cast_fp16 = transpose(perm = var_731, x = values_9_cast_fp16)[name = tensor("transpose_49")]; + tensor attn_output_3_cast_fp16 = matmul(transpose_x = attn_output_3_transpose_x_0, transpose_y = attn_output_3_transpose_y_0, x = attn_weights_11_cast_fp16, y = values_11_cast_fp16)[name = tensor("attn_output_3_cast_fp16")]; + tensor var_799 = const()[name = tensor("op_799"), val = tensor([0, 2, 1, 3])]; + tensor var_804 = const()[name = tensor("op_804"), val = tensor([1, 1, 1024])]; + tensor var_800_cast_fp16 = transpose(perm = var_799, x = attn_output_3_cast_fp16)[name = tensor("transpose_46")]; + tensor input_19_cast_fp16 = reshape(shape = var_804, x = var_800_cast_fp16)[name = tensor("input_19_cast_fp16")]; + tensor attn1_out_proj_weight_to_fp16 = const()[name = tensor("attn1_out_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(32602496)))]; + tensor linear_6_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = attn1_out_proj_weight_to_fp16, x = input_19_cast_fp16)[name = tensor("linear_6_cast_fp16")]; + tensor var_810_to_fp16 = const()[name = tensor("op_810_to_fp16"), val = tensor(0x1p+0)]; + tensor var_811_cast_fp16 = add(x = position1_to_fp16, y = var_810_to_fp16)[name = tensor("op_811_cast_fp16")]; + tensor var_811_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_811_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor input_21_cast_fp16 = add(x = input_15_cast_fp16, y = linear_6_cast_fp16)[name = tensor("input_21_cast_fp16")]; + tensor input_23_axes_0 = const()[name = tensor("input_23_axes_0"), val = tensor([-1])]; + tensor norm1_2_weight_to_fp16 = const()[name = tensor("norm1_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(34699712)))]; + tensor norm1_2_bias_to_fp16 = const()[name = tensor("norm1_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(34701824)))]; + tensor var_815_to_fp16 = const()[name = tensor("op_815_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_23_cast_fp16 = layer_norm(axes = input_23_axes_0, beta = norm1_2_bias_to_fp16, epsilon = var_815_to_fp16, gamma = norm1_2_weight_to_fp16, x = input_21_cast_fp16)[name = tensor("input_23_cast_fp16")]; + tensor linear1_1_weight_to_fp16 = const()[name = tensor("linear1_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(34703936)))]; + tensor linear_7_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = linear1_1_weight_to_fp16, x = input_23_cast_fp16)[name = tensor("linear_7_cast_fp16")]; + tensor input_25_mode_0 = const()[name = tensor("input_25_mode_0"), val = tensor("EXACT")]; + tensor input_25_cast_fp16 = gelu(mode = input_25_mode_0, x = linear_7_cast_fp16)[name = tensor("input_25_cast_fp16")]; + tensor linear1_2_weight_to_fp16 = const()[name = tensor("linear1_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(43092608)))]; + tensor linear_8_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = linear1_2_weight_to_fp16, x = input_25_cast_fp16)[name = tensor("linear_8_cast_fp16")]; + tensor input_27_cast_fp16 = add(x = input_21_cast_fp16, y = linear_8_cast_fp16)[name = tensor("input_27_cast_fp16")]; + tensor input_29_axes_0 = const()[name = tensor("input_29_axes_0"), val = tensor([-1])]; + tensor norm2_1_weight_to_fp16 = const()[name = tensor("norm2_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(51481280)))]; + tensor norm2_1_bias_to_fp16 = const()[name = tensor("norm2_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(51483392)))]; + tensor var_832_to_fp16 = const()[name = tensor("op_832_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_29_cast_fp16 = layer_norm(axes = input_29_axes_0, beta = norm2_1_bias_to_fp16, epsilon = var_832_to_fp16, gamma = norm2_1_weight_to_fp16, x = input_27_cast_fp16)[name = tensor("input_29_cast_fp16")]; + tensor attn2_in_proj_weight_to_fp16 = const()[name = tensor("attn2_in_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(51485504)))]; + tensor linear_9_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = attn2_in_proj_weight_to_fp16, x = input_29_cast_fp16)[name = tensor("linear_9_cast_fp16")]; + tensor var_846 = const()[name = tensor("op_846"), val = tensor([1, 1, 3, 16, 64])]; + tensor qkv_5_cast_fp16 = reshape(shape = var_846, x = linear_9_cast_fp16)[name = tensor("qkv_5_cast_fp16")]; + tensor q_13_begin_0 = const()[name = tensor("q_13_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor q_13_end_0 = const()[name = tensor("q_13_end_0"), val = tensor([1, 1, 1, 16, 64])]; + tensor q_13_end_mask_0 = const()[name = tensor("q_13_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor q_13_squeeze_mask_0 = const()[name = tensor("q_13_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor q_13_cast_fp16 = slice_by_index(begin = q_13_begin_0, end = q_13_end_0, end_mask = q_13_end_mask_0, squeeze_mask = q_13_squeeze_mask_0, x = qkv_5_cast_fp16)[name = tensor("q_13_cast_fp16")]; + tensor k_9_begin_0 = const()[name = tensor("k_9_begin_0"), val = tensor([0, 0, 1, 0, 0])]; + tensor k_9_end_0 = const()[name = tensor("k_9_end_0"), val = tensor([1, 1, 2, 16, 64])]; + tensor k_9_end_mask_0 = const()[name = tensor("k_9_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor k_9_squeeze_mask_0 = const()[name = tensor("k_9_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor k_9_cast_fp16 = slice_by_index(begin = k_9_begin_0, end = k_9_end_0, end_mask = k_9_end_mask_0, squeeze_mask = k_9_squeeze_mask_0, x = qkv_5_cast_fp16)[name = tensor("k_9_cast_fp16")]; + tensor v_5_begin_0 = const()[name = tensor("v_5_begin_0"), val = tensor([0, 0, 2, 0, 0])]; + tensor v_5_end_0 = const()[name = tensor("v_5_end_0"), val = tensor([1, 1, 3, 16, 64])]; + tensor v_5_end_mask_0 = const()[name = tensor("v_5_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor v_5_squeeze_mask_0 = const()[name = tensor("v_5_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor v_5_cast_fp16 = slice_by_index(begin = v_5_begin_0, end = v_5_end_0, end_mask = v_5_end_mask_0, squeeze_mask = v_5_squeeze_mask_0, x = qkv_5_cast_fp16)[name = tensor("v_5_cast_fp16")]; + tensor var_912 = const()[name = tensor("op_912"), val = tensor([1, 1, 1, 1])]; + tensor position2_to_fp16_dtype_0 = const()[name = tensor("position2_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor position2_to_fp16 = cast(dtype = position2_to_fp16_dtype_0, x = position2)[name = tensor("cast_129")]; + tensor ts_17_cast_fp16 = reshape(shape = var_912, x = position2_to_fp16)[name = tensor("ts_17_cast_fp16")]; + tensor var_919 = const()[name = tensor("op_919"), val = tensor([1, 1, 16, 32, 2])]; + tensor q_complex_5_cast_fp16 = reshape(shape = var_919, x = q_13_cast_fp16)[name = tensor("q_complex_5_cast_fp16")]; + tensor var_926 = const()[name = tensor("op_926"), val = tensor([1, 1, 16, 32, 2])]; + tensor k_complex_5_cast_fp16 = reshape(shape = var_926, x = k_9_cast_fp16)[name = tensor("k_complex_5_cast_fp16")]; + tensor var_930_begin_0 = const()[name = tensor("op_930_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_930_end_0 = const()[name = tensor("op_930_end_0"), val = tensor([1, 1, 16, 32, 1])]; + tensor var_930_end_mask_0 = const()[name = tensor("op_930_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_930_squeeze_mask_0 = const()[name = tensor("op_930_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_930_cast_fp16 = slice_by_index(begin = var_930_begin_0, end = var_930_end_0, end_mask = var_930_end_mask_0, squeeze_mask = var_930_squeeze_mask_0, x = q_complex_5_cast_fp16)[name = tensor("op_930_cast_fp16")]; + tensor var_938_begin_0 = const()[name = tensor("op_938_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_938_end_0 = const()[name = tensor("op_938_end_0"), val = tensor([1, 1, 16, 32, 2])]; + tensor var_938_end_mask_0 = const()[name = tensor("op_938_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_938_squeeze_mask_0 = const()[name = tensor("op_938_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_938_cast_fp16 = slice_by_index(begin = var_938_begin_0, end = var_938_end_0, end_mask = var_938_end_mask_0, squeeze_mask = var_938_squeeze_mask_0, x = q_complex_5_cast_fp16)[name = tensor("op_938_cast_fp16")]; + tensor var_946_begin_0 = const()[name = tensor("op_946_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_946_end_0 = const()[name = tensor("op_946_end_0"), val = tensor([1, 1, 16, 32, 1])]; + tensor var_946_end_mask_0 = const()[name = tensor("op_946_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_946_squeeze_mask_0 = const()[name = tensor("op_946_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_946_cast_fp16 = slice_by_index(begin = var_946_begin_0, end = var_946_end_0, end_mask = var_946_end_mask_0, squeeze_mask = var_946_squeeze_mask_0, x = k_complex_5_cast_fp16)[name = tensor("op_946_cast_fp16")]; + tensor var_954_begin_0 = const()[name = tensor("op_954_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_954_end_0 = const()[name = tensor("op_954_end_0"), val = tensor([1, 1, 16, 32, 2])]; + tensor var_954_end_mask_0 = const()[name = tensor("op_954_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_954_squeeze_mask_0 = const()[name = tensor("op_954_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_954_cast_fp16 = slice_by_index(begin = var_954_begin_0, end = var_954_end_0, end_mask = var_954_end_mask_0, squeeze_mask = var_954_squeeze_mask_0, x = k_complex_5_cast_fp16)[name = tensor("op_954_cast_fp16")]; + tensor freqs_5_to_fp16 = const()[name = tensor("freqs_5_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(57777024)))]; + tensor var_960_cast_fp16 = mul(x = freqs_5_to_fp16, y = ts_17_cast_fp16)[name = tensor("op_960_cast_fp16")]; + tensor rotr_5_cast_fp16 = cos(x = var_960_cast_fp16)[name = tensor("rotr_5_cast_fp16")]; + tensor roti_5_cast_fp16 = sin(x = var_960_cast_fp16)[name = tensor("roti_5_cast_fp16")]; + tensor var_964_cast_fp16 = mul(x = var_930_cast_fp16, y = rotr_5_cast_fp16)[name = tensor("op_964_cast_fp16")]; + tensor var_965_cast_fp16 = mul(x = var_938_cast_fp16, y = roti_5_cast_fp16)[name = tensor("op_965_cast_fp16")]; + tensor qor_9_cast_fp16 = sub(x = var_964_cast_fp16, y = var_965_cast_fp16)[name = tensor("qor_9_cast_fp16")]; + tensor var_968_cast_fp16 = mul(x = var_930_cast_fp16, y = roti_5_cast_fp16)[name = tensor("op_968_cast_fp16")]; + tensor var_969_cast_fp16 = mul(x = var_938_cast_fp16, y = rotr_5_cast_fp16)[name = tensor("op_969_cast_fp16")]; + tensor qoi_9_cast_fp16 = add(x = var_968_cast_fp16, y = var_969_cast_fp16)[name = tensor("qoi_9_cast_fp16")]; + tensor var_972_cast_fp16 = mul(x = var_946_cast_fp16, y = rotr_5_cast_fp16)[name = tensor("op_972_cast_fp16")]; + tensor var_973_cast_fp16 = mul(x = var_954_cast_fp16, y = roti_5_cast_fp16)[name = tensor("op_973_cast_fp16")]; + tensor kor_9_cast_fp16 = sub(x = var_972_cast_fp16, y = var_973_cast_fp16)[name = tensor("kor_9_cast_fp16")]; + tensor var_976_cast_fp16 = mul(x = var_946_cast_fp16, y = roti_5_cast_fp16)[name = tensor("op_976_cast_fp16")]; + tensor var_977_cast_fp16 = mul(x = var_954_cast_fp16, y = rotr_5_cast_fp16)[name = tensor("op_977_cast_fp16")]; + tensor koi_9_cast_fp16 = add(x = var_976_cast_fp16, y = var_977_cast_fp16)[name = tensor("koi_9_cast_fp16")]; + tensor qo_5_axis_0 = const()[name = tensor("qo_5_axis_0"), val = tensor(-1)]; + tensor qo_5_cast_fp16 = stack(axis = qo_5_axis_0, values = (qor_9_cast_fp16, qoi_9_cast_fp16))[name = tensor("qo_5_cast_fp16")]; + tensor ko_5_axis_0 = const()[name = tensor("ko_5_axis_0"), val = tensor(-1)]; + tensor ko_5_cast_fp16 = stack(axis = ko_5_axis_0, values = (kor_9_cast_fp16, koi_9_cast_fp16))[name = tensor("ko_5_cast_fp16")]; + tensor var_1010 = const()[name = tensor("op_1010"), val = tensor([1, 1, 16, 64])]; + tensor q_15_cast_fp16 = reshape(shape = var_1010, x = qo_5_cast_fp16)[name = tensor("q_15_cast_fp16")]; + tensor var_1016 = const()[name = tensor("op_1016"), val = tensor([1, 1, 16, 64])]; + tensor k_11_cast_fp16 = reshape(shape = var_1016, x = ko_5_cast_fp16)[name = tensor("k_11_cast_fp16")]; + tensor _inversed_1041_y_0_to_fp16 = const()[name = tensor("_inversed_1041_y_0_to_fp16"), val = tensor(0x1p-9)]; + tensor _inversed_1041_cast_fp16 = mul(x = ts_17_cast_fp16, y = _inversed_1041_y_0_to_fp16)[name = tensor("_inversed_1041_cast_fp16")]; + tensor var_1042_cast_fp16 = floor(x = _inversed_1041_cast_fp16)[name = tensor("op_1042_cast_fp16")]; + tensor var_1043_to_fp16 = const()[name = tensor("op_1043_to_fp16"), val = tensor(0x1p+9)]; + tensor var_1044_cast_fp16 = mul(x = var_1042_cast_fp16, y = var_1043_to_fp16)[name = tensor("op_1044_cast_fp16")]; + tensor write_indices_float_11_cast_fp16 = sub(x = ts_17_cast_fp16, y = var_1044_cast_fp16)[name = tensor("write_indices_float_11_cast_fp16")]; + tensor var_1051_dtype_0 = const()[name = tensor("op_1051_dtype_0"), val = tensor("int32")]; + tensor write_indices_5_reps_0 = const()[name = tensor("write_indices_5_reps_0"), val = tensor([1, 1, 16, 64])]; + tensor write_indices_float_11_cast_fp16_to_int32 = cast(dtype = var_1051_dtype_0, x = write_indices_float_11_cast_fp16)[name = tensor("cast_128")]; + tensor write_indices_5 = tile(reps = write_indices_5_reps_0, x = write_indices_float_11_cast_fp16_to_int32)[name = tensor("write_indices_5")]; + tensor var_1061_begin_0 = const()[name = tensor("op_1061_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_1061_end_0 = const()[name = tensor("op_1061_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor var_1061_end_mask_0 = const()[name = tensor("op_1061_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_1061_squeeze_mask_0 = const()[name = tensor("op_1061_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor cache2_to_fp16_dtype_0 = const()[name = tensor("cache2_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor cache2_to_fp16 = cast(dtype = cache2_to_fp16_dtype_0, x = cache2)[name = tensor("cast_127")]; + tensor var_1061_cast_fp16 = slice_by_index(begin = var_1061_begin_0, end = var_1061_end_0, end_mask = var_1061_end_mask_0, squeeze_mask = var_1061_squeeze_mask_0, x = cache2_to_fp16)[name = tensor("op_1061_cast_fp16")]; + tensor var_1063_axis_0 = const()[name = tensor("op_1063_axis_0"), val = tensor(1)]; + tensor var_1063_mode_0 = const()[name = tensor("op_1063_mode_0"), val = tensor("update")]; + tensor var_1063_validate_indices_0 = const()[name = tensor("op_1063_validate_indices_0"), val = tensor(false)]; + tensor var_1063_cast_fp16 = scatter_along_axis(axis = var_1063_axis_0, data = var_1061_cast_fp16, indices = write_indices_5, mode = var_1063_mode_0, updates = k_11_cast_fp16, validate_indices = var_1063_validate_indices_0)[name = tensor("op_1063_cast_fp16")]; + tensor concat_16 = const()[name = tensor("concat_16"), val = tensor([0, 0, 0, 0, 0])]; + tensor concat_17 = const()[name = tensor("concat_17"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_5_internal_tensor_assign_1_stride_0 = const()[name = tensor("new_cache_5_internal_tensor_assign_1_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_5_internal_tensor_assign_1_begin_mask_0 = const()[name = tensor("new_cache_5_internal_tensor_assign_1_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_5_internal_tensor_assign_1_end_mask_0 = const()[name = tensor("new_cache_5_internal_tensor_assign_1_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_5_internal_tensor_assign_1_squeeze_mask_0 = const()[name = tensor("new_cache_5_internal_tensor_assign_1_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_16 = const()[name = tensor("shape_16"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_4 = const()[name = tensor("reduce_prod_4"), val = tensor(1048576)]; + tensor range_1d_4_start_0 = const()[name = tensor("range_1d_4_start_0"), val = tensor(0)]; + tensor range_1d_4_step_0 = const()[name = tensor("range_1d_4_step_0"), val = tensor(1)]; + tensor range_1d_4 = range_1d(end = reduce_prod_4, start = range_1d_4_start_0, step = range_1d_4_step_0)[name = tensor("range_1d_4")]; + tensor reshape_20 = reshape(shape = shape_16, x = range_1d_4)[name = tensor("reshape_20")]; + tensor slice_by_index_4 = slice_by_index(begin = concat_16, begin_mask = new_cache_5_internal_tensor_assign_1_begin_mask_0, end = concat_17, end_mask = new_cache_5_internal_tensor_assign_1_end_mask_0, squeeze_mask = new_cache_5_internal_tensor_assign_1_squeeze_mask_0, stride = new_cache_5_internal_tensor_assign_1_stride_0, x = reshape_20)[name = tensor("slice_by_index_4")]; + tensor reshape_21_shape_0 = const()[name = tensor("reshape_21_shape_0"), val = tensor([-1])]; + tensor reshape_21 = reshape(shape = reshape_21_shape_0, x = slice_by_index_4)[name = tensor("reshape_21")]; + tensor reshape_22_shape_0 = const()[name = tensor("reshape_22_shape_0"), val = tensor([-1])]; + tensor reshape_22_cast_fp16 = reshape(shape = reshape_22_shape_0, x = var_1063_cast_fp16)[name = tensor("reshape_22_cast_fp16")]; + tensor reshape_23_shape_0 = const()[name = tensor("reshape_23_shape_0"), val = tensor([-1])]; + tensor reshape_23_cast_fp16 = reshape(shape = reshape_23_shape_0, x = cache2_to_fp16)[name = tensor("reshape_23_cast_fp16")]; + tensor scatter_4_mode_0 = const()[name = tensor("scatter_4_mode_0"), val = tensor("update")]; + tensor scatter_4_axis_0 = const()[name = tensor("scatter_4_axis_0"), val = tensor(0)]; + tensor scatter_4_validate_indices_0 = const()[name = tensor("scatter_4_validate_indices_0"), val = tensor(false)]; + tensor scatter_4_cast_fp16 = scatter(axis = scatter_4_axis_0, data = reshape_23_cast_fp16, indices = reshape_21, mode = scatter_4_mode_0, updates = reshape_22_cast_fp16, validate_indices = scatter_4_validate_indices_0)[name = tensor("scatter_4_cast_fp16")]; + tensor reshape_24_cast_fp16 = reshape(shape = shape_16, x = scatter_4_cast_fp16)[name = tensor("reshape_24_cast_fp16")]; + tensor var_1071_begin_0 = const()[name = tensor("op_1071_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor var_1071_end_0 = const()[name = tensor("op_1071_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor var_1071_end_mask_0 = const()[name = tensor("op_1071_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_1071_squeeze_mask_0 = const()[name = tensor("op_1071_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor var_1071_cast_fp16 = slice_by_index(begin = var_1071_begin_0, end = var_1071_end_0, end_mask = var_1071_end_mask_0, squeeze_mask = var_1071_squeeze_mask_0, x = reshape_24_cast_fp16)[name = tensor("op_1071_cast_fp16")]; + tensor var_1073_axis_0 = const()[name = tensor("op_1073_axis_0"), val = tensor(1)]; + tensor var_1073_mode_0 = const()[name = tensor("op_1073_mode_0"), val = tensor("update")]; + tensor var_1073_validate_indices_0 = const()[name = tensor("op_1073_validate_indices_0"), val = tensor(false)]; + tensor var_1073_cast_fp16 = scatter_along_axis(axis = var_1073_axis_0, data = var_1071_cast_fp16, indices = write_indices_5, mode = var_1073_mode_0, updates = v_5_cast_fp16, validate_indices = var_1073_validate_indices_0)[name = tensor("op_1073_cast_fp16")]; + tensor concat_18 = const()[name = tensor("concat_18"), val = tensor([1, 0, 0, 0, 0])]; + tensor concat_19 = const()[name = tensor("concat_19"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_5_internal_tensor_assign_2_stride_0 = const()[name = tensor("new_cache_5_internal_tensor_assign_2_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_5_internal_tensor_assign_2_begin_mask_0 = const()[name = tensor("new_cache_5_internal_tensor_assign_2_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_5_internal_tensor_assign_2_end_mask_0 = const()[name = tensor("new_cache_5_internal_tensor_assign_2_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_5_internal_tensor_assign_2_squeeze_mask_0 = const()[name = tensor("new_cache_5_internal_tensor_assign_2_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_17 = const()[name = tensor("shape_17"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_5 = const()[name = tensor("reduce_prod_5"), val = tensor(1048576)]; + tensor range_1d_5_start_0 = const()[name = tensor("range_1d_5_start_0"), val = tensor(0)]; + tensor range_1d_5_step_0 = const()[name = tensor("range_1d_5_step_0"), val = tensor(1)]; + tensor range_1d_5 = range_1d(end = reduce_prod_5, start = range_1d_5_start_0, step = range_1d_5_step_0)[name = tensor("range_1d_5")]; + tensor reshape_25 = reshape(shape = shape_17, x = range_1d_5)[name = tensor("reshape_25")]; + tensor slice_by_index_5 = slice_by_index(begin = concat_18, begin_mask = new_cache_5_internal_tensor_assign_2_begin_mask_0, end = concat_19, end_mask = new_cache_5_internal_tensor_assign_2_end_mask_0, squeeze_mask = new_cache_5_internal_tensor_assign_2_squeeze_mask_0, stride = new_cache_5_internal_tensor_assign_2_stride_0, x = reshape_25)[name = tensor("slice_by_index_5")]; + tensor reshape_26_shape_0 = const()[name = tensor("reshape_26_shape_0"), val = tensor([-1])]; + tensor reshape_26 = reshape(shape = reshape_26_shape_0, x = slice_by_index_5)[name = tensor("reshape_26")]; + tensor reshape_27_shape_0 = const()[name = tensor("reshape_27_shape_0"), val = tensor([-1])]; + tensor reshape_27_cast_fp16 = reshape(shape = reshape_27_shape_0, x = var_1073_cast_fp16)[name = tensor("reshape_27_cast_fp16")]; + tensor reshape_28_shape_0 = const()[name = tensor("reshape_28_shape_0"), val = tensor([-1])]; + tensor reshape_28_cast_fp16 = reshape(shape = reshape_28_shape_0, x = reshape_24_cast_fp16)[name = tensor("reshape_28_cast_fp16")]; + tensor scatter_5_mode_0 = const()[name = tensor("scatter_5_mode_0"), val = tensor("update")]; + tensor scatter_5_axis_0 = const()[name = tensor("scatter_5_axis_0"), val = tensor(0)]; + tensor scatter_5_validate_indices_0 = const()[name = tensor("scatter_5_validate_indices_0"), val = tensor(false)]; + tensor scatter_5_cast_fp16 = scatter(axis = scatter_5_axis_0, data = reshape_28_cast_fp16, indices = reshape_26, mode = scatter_5_mode_0, updates = reshape_27_cast_fp16, validate_indices = scatter_5_validate_indices_0)[name = tensor("scatter_5_cast_fp16")]; + tensor reshape_29_cast_fp16 = reshape(shape = shape_17, x = scatter_5_cast_fp16)[name = tensor("reshape_29_cast_fp16")]; + tensor reshape_29_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("reshape_29_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor keys_13_begin_0 = const()[name = tensor("keys_13_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor keys_13_end_0 = const()[name = tensor("keys_13_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor keys_13_end_mask_0 = const()[name = tensor("keys_13_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor keys_13_squeeze_mask_0 = const()[name = tensor("keys_13_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor keys_13_cast_fp16 = slice_by_index(begin = keys_13_begin_0, end = keys_13_end_0, end_mask = keys_13_end_mask_0, squeeze_mask = keys_13_squeeze_mask_0, x = reshape_29_cast_fp16)[name = tensor("keys_13_cast_fp16")]; + tensor values_13_begin_0 = const()[name = tensor("values_13_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor values_13_end_0 = const()[name = tensor("values_13_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor values_13_end_mask_0 = const()[name = tensor("values_13_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor values_13_squeeze_mask_0 = const()[name = tensor("values_13_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor values_13_cast_fp16 = slice_by_index(begin = values_13_begin_0, end = values_13_end_0, end_mask = values_13_end_mask_0, squeeze_mask = values_13_squeeze_mask_0, x = reshape_29_cast_fp16)[name = tensor("values_13_cast_fp16")]; + tensor var_1085_cast_fp16 = not_equal(x = keys_13_cast_fp16, y = keys_13_cast_fp16)[name = tensor("op_1085_cast_fp16")]; + tensor keys_15_cast_fp16 = select(a = var_319_to_fp16, b = keys_13_cast_fp16, cond = var_1085_cast_fp16)[name = tensor("keys_15_cast_fp16")]; + tensor var_1093_cast_fp16 = not_equal(x = values_13_cast_fp16, y = values_13_cast_fp16)[name = tensor("op_1093_cast_fp16")]; + tensor values_15_cast_fp16 = select(a = var_319_to_fp16, b = values_13_cast_fp16, cond = var_1093_cast_fp16)[name = tensor("values_15_cast_fp16")]; + tensor var_1117 = const()[name = tensor("op_1117"), val = tensor([0, 2, 1, 3])]; + tensor var_1133 = const()[name = tensor("op_1133"), val = tensor([1, 1, 1])]; + tensor var_1134_cast_fp16 = reshape(shape = var_1133, x = position2_to_fp16)[name = tensor("op_1134_cast_fp16")]; + tensor var_1154_to_fp16 = const()[name = tensor("op_1154_to_fp16"), val = tensor(0x1p+0)]; + tensor valid_len_5_cast_fp16 = add(x = var_1134_cast_fp16, y = var_1154_to_fp16)[name = tensor("valid_len_5_cast_fp16")]; + tensor valid_mask_5_cast_fp16 = less(x = k_positions_1_promoted_to_fp16, y = valid_len_5_cast_fp16)[name = tensor("valid_mask_5_cast_fp16")]; + tensor causal_mask_5_cast_fp16 = less_equal(x = k_positions_1_promoted_to_fp16, y = var_1134_cast_fp16)[name = tensor("causal_mask_5_cast_fp16")]; + tensor attn_mask_9 = logical_and(x = valid_mask_5_cast_fp16, y = causal_mask_5_cast_fp16)[name = tensor("attn_mask_9")]; + tensor attn_mask_11_axes_0 = const()[name = tensor("attn_mask_11_axes_0"), val = tensor([1])]; + tensor attn_mask_11 = expand_dims(axes = attn_mask_11_axes_0, x = attn_mask_9)[name = tensor("attn_mask_11")]; + tensor var_1172_transpose_x_0 = const()[name = tensor("op_1172_transpose_x_0"), val = tensor(false)]; + tensor var_1172_transpose_y_0 = const()[name = tensor("op_1172_transpose_y_0"), val = tensor(false)]; + tensor transpose_22_perm_0 = const()[name = tensor("transpose_22_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_23_perm_0 = const()[name = tensor("transpose_23_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_23 = transpose(perm = transpose_23_perm_0, x = keys_15_cast_fp16)[name = tensor("transpose_43")]; + tensor transpose_22 = transpose(perm = transpose_22_perm_0, x = q_15_cast_fp16)[name = tensor("transpose_44")]; + tensor var_1172_cast_fp16 = matmul(transpose_x = var_1172_transpose_x_0, transpose_y = var_1172_transpose_y_0, x = transpose_22, y = transpose_23)[name = tensor("op_1172_cast_fp16")]; + tensor var_1166_to_fp16 = const()[name = tensor("op_1166_to_fp16"), val = tensor([0x1p-3])]; + tensor attn_weights_13_cast_fp16 = mul(x = var_1172_cast_fp16, y = var_1166_to_fp16)[name = tensor("attn_weights_13_cast_fp16")]; + tensor var_1174 = logical_not(x = attn_mask_11)[name = tensor("op_1174")]; + tensor var_1175_to_fp16 = const()[name = tensor("op_1175_to_fp16"), val = tensor(-inf)]; + tensor attn_weights_15_cast_fp16 = select(a = var_1175_to_fp16, b = attn_weights_13_cast_fp16, cond = var_1174)[name = tensor("attn_weights_15_cast_fp16")]; + tensor var_1177 = const()[name = tensor("op_1177"), val = tensor(-1)]; + tensor attn_weights_17_cast_fp16 = softmax(axis = var_1177, x = attn_weights_15_cast_fp16)[name = tensor("attn_weights_17_cast_fp16")]; + tensor attn_output_5_transpose_x_0 = const()[name = tensor("attn_output_5_transpose_x_0"), val = tensor(false)]; + tensor attn_output_5_transpose_y_0 = const()[name = tensor("attn_output_5_transpose_y_0"), val = tensor(false)]; + tensor values_17_cast_fp16 = transpose(perm = var_1117, x = values_15_cast_fp16)[name = tensor("transpose_45")]; + tensor attn_output_5_cast_fp16 = matmul(transpose_x = attn_output_5_transpose_x_0, transpose_y = attn_output_5_transpose_y_0, x = attn_weights_17_cast_fp16, y = values_17_cast_fp16)[name = tensor("attn_output_5_cast_fp16")]; + tensor var_1185 = const()[name = tensor("op_1185"), val = tensor([0, 2, 1, 3])]; + tensor var_1190 = const()[name = tensor("op_1190"), val = tensor([1, 1, 1024])]; + tensor var_1186_cast_fp16 = transpose(perm = var_1185, x = attn_output_5_cast_fp16)[name = tensor("transpose_42")]; + tensor input_31_cast_fp16 = reshape(shape = var_1190, x = var_1186_cast_fp16)[name = tensor("input_31_cast_fp16")]; + tensor attn2_out_proj_weight_to_fp16 = const()[name = tensor("attn2_out_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(57777152)))]; + tensor linear_10_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = attn2_out_proj_weight_to_fp16, x = input_31_cast_fp16)[name = tensor("linear_10_cast_fp16")]; + tensor var_1196_to_fp16 = const()[name = tensor("op_1196_to_fp16"), val = tensor(0x1p+0)]; + tensor var_1197_cast_fp16 = add(x = position2_to_fp16, y = var_1196_to_fp16)[name = tensor("op_1197_cast_fp16")]; + tensor var_1197_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_1197_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor input_33_cast_fp16 = add(x = input_27_cast_fp16, y = linear_10_cast_fp16)[name = tensor("input_33_cast_fp16")]; + tensor input_35_axes_0 = const()[name = tensor("input_35_axes_0"), val = tensor([-1])]; + tensor norm2_2_weight_to_fp16 = const()[name = tensor("norm2_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(59874368)))]; + tensor norm2_2_bias_to_fp16 = const()[name = tensor("norm2_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(59876480)))]; + tensor var_1201_to_fp16 = const()[name = tensor("op_1201_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_35_cast_fp16 = layer_norm(axes = input_35_axes_0, beta = norm2_2_bias_to_fp16, epsilon = var_1201_to_fp16, gamma = norm2_2_weight_to_fp16, x = input_33_cast_fp16)[name = tensor("input_35_cast_fp16")]; + tensor linear2_1_weight_to_fp16 = const()[name = tensor("linear2_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(59878592)))]; + tensor linear_11_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = linear2_1_weight_to_fp16, x = input_35_cast_fp16)[name = tensor("linear_11_cast_fp16")]; + tensor input_37_mode_0 = const()[name = tensor("input_37_mode_0"), val = tensor("EXACT")]; + tensor input_37_cast_fp16 = gelu(mode = input_37_mode_0, x = linear_11_cast_fp16)[name = tensor("input_37_cast_fp16")]; + tensor linear2_2_weight_to_fp16 = const()[name = tensor("linear2_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(68267264)))]; + tensor linear_12_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = linear2_2_weight_to_fp16, x = input_37_cast_fp16)[name = tensor("linear_12_cast_fp16")]; + tensor input_39_cast_fp16 = add(x = input_33_cast_fp16, y = linear_12_cast_fp16)[name = tensor("input_39_cast_fp16")]; + tensor input_41_axes_0 = const()[name = tensor("input_41_axes_0"), val = tensor([-1])]; + tensor norm3_1_weight_to_fp16 = const()[name = tensor("norm3_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(76655936)))]; + tensor norm3_1_bias_to_fp16 = const()[name = tensor("norm3_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(76658048)))]; + tensor var_1218_to_fp16 = const()[name = tensor("op_1218_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_41_cast_fp16 = layer_norm(axes = input_41_axes_0, beta = norm3_1_bias_to_fp16, epsilon = var_1218_to_fp16, gamma = norm3_1_weight_to_fp16, x = input_39_cast_fp16)[name = tensor("input_41_cast_fp16")]; + tensor attn3_in_proj_weight_to_fp16 = const()[name = tensor("attn3_in_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(76660160)))]; + tensor linear_13_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = attn3_in_proj_weight_to_fp16, x = input_41_cast_fp16)[name = tensor("linear_13_cast_fp16")]; + tensor var_1232 = const()[name = tensor("op_1232"), val = tensor([1, 1, 3, 16, 64])]; + tensor qkv_7_cast_fp16 = reshape(shape = var_1232, x = linear_13_cast_fp16)[name = tensor("qkv_7_cast_fp16")]; + tensor q_19_begin_0 = const()[name = tensor("q_19_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor q_19_end_0 = const()[name = tensor("q_19_end_0"), val = tensor([1, 1, 1, 16, 64])]; + tensor q_19_end_mask_0 = const()[name = tensor("q_19_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor q_19_squeeze_mask_0 = const()[name = tensor("q_19_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor q_19_cast_fp16 = slice_by_index(begin = q_19_begin_0, end = q_19_end_0, end_mask = q_19_end_mask_0, squeeze_mask = q_19_squeeze_mask_0, x = qkv_7_cast_fp16)[name = tensor("q_19_cast_fp16")]; + tensor k_13_begin_0 = const()[name = tensor("k_13_begin_0"), val = tensor([0, 0, 1, 0, 0])]; + tensor k_13_end_0 = const()[name = tensor("k_13_end_0"), val = tensor([1, 1, 2, 16, 64])]; + tensor k_13_end_mask_0 = const()[name = tensor("k_13_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor k_13_squeeze_mask_0 = const()[name = tensor("k_13_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor k_13_cast_fp16 = slice_by_index(begin = k_13_begin_0, end = k_13_end_0, end_mask = k_13_end_mask_0, squeeze_mask = k_13_squeeze_mask_0, x = qkv_7_cast_fp16)[name = tensor("k_13_cast_fp16")]; + tensor v_7_begin_0 = const()[name = tensor("v_7_begin_0"), val = tensor([0, 0, 2, 0, 0])]; + tensor v_7_end_0 = const()[name = tensor("v_7_end_0"), val = tensor([1, 1, 3, 16, 64])]; + tensor v_7_end_mask_0 = const()[name = tensor("v_7_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor v_7_squeeze_mask_0 = const()[name = tensor("v_7_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor v_7_cast_fp16 = slice_by_index(begin = v_7_begin_0, end = v_7_end_0, end_mask = v_7_end_mask_0, squeeze_mask = v_7_squeeze_mask_0, x = qkv_7_cast_fp16)[name = tensor("v_7_cast_fp16")]; + tensor var_1298 = const()[name = tensor("op_1298"), val = tensor([1, 1, 1, 1])]; + tensor position3_to_fp16_dtype_0 = const()[name = tensor("position3_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor position3_to_fp16 = cast(dtype = position3_to_fp16_dtype_0, x = position3)[name = tensor("cast_124")]; + tensor ts_23_cast_fp16 = reshape(shape = var_1298, x = position3_to_fp16)[name = tensor("ts_23_cast_fp16")]; + tensor var_1305 = const()[name = tensor("op_1305"), val = tensor([1, 1, 16, 32, 2])]; + tensor q_complex_7_cast_fp16 = reshape(shape = var_1305, x = q_19_cast_fp16)[name = tensor("q_complex_7_cast_fp16")]; + tensor var_1312 = const()[name = tensor("op_1312"), val = tensor([1, 1, 16, 32, 2])]; + tensor k_complex_7_cast_fp16 = reshape(shape = var_1312, x = k_13_cast_fp16)[name = tensor("k_complex_7_cast_fp16")]; + tensor var_1316_begin_0 = const()[name = tensor("op_1316_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_1316_end_0 = const()[name = tensor("op_1316_end_0"), val = tensor([1, 1, 16, 32, 1])]; + tensor var_1316_end_mask_0 = const()[name = tensor("op_1316_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_1316_squeeze_mask_0 = const()[name = tensor("op_1316_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_1316_cast_fp16 = slice_by_index(begin = var_1316_begin_0, end = var_1316_end_0, end_mask = var_1316_end_mask_0, squeeze_mask = var_1316_squeeze_mask_0, x = q_complex_7_cast_fp16)[name = tensor("op_1316_cast_fp16")]; + tensor var_1324_begin_0 = const()[name = tensor("op_1324_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_1324_end_0 = const()[name = tensor("op_1324_end_0"), val = tensor([1, 1, 16, 32, 2])]; + tensor var_1324_end_mask_0 = const()[name = tensor("op_1324_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_1324_squeeze_mask_0 = const()[name = tensor("op_1324_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_1324_cast_fp16 = slice_by_index(begin = var_1324_begin_0, end = var_1324_end_0, end_mask = var_1324_end_mask_0, squeeze_mask = var_1324_squeeze_mask_0, x = q_complex_7_cast_fp16)[name = tensor("op_1324_cast_fp16")]; + tensor var_1332_begin_0 = const()[name = tensor("op_1332_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_1332_end_0 = const()[name = tensor("op_1332_end_0"), val = tensor([1, 1, 16, 32, 1])]; + tensor var_1332_end_mask_0 = const()[name = tensor("op_1332_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_1332_squeeze_mask_0 = const()[name = tensor("op_1332_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_1332_cast_fp16 = slice_by_index(begin = var_1332_begin_0, end = var_1332_end_0, end_mask = var_1332_end_mask_0, squeeze_mask = var_1332_squeeze_mask_0, x = k_complex_7_cast_fp16)[name = tensor("op_1332_cast_fp16")]; + tensor var_1340_begin_0 = const()[name = tensor("op_1340_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_1340_end_0 = const()[name = tensor("op_1340_end_0"), val = tensor([1, 1, 16, 32, 2])]; + tensor var_1340_end_mask_0 = const()[name = tensor("op_1340_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_1340_squeeze_mask_0 = const()[name = tensor("op_1340_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_1340_cast_fp16 = slice_by_index(begin = var_1340_begin_0, end = var_1340_end_0, end_mask = var_1340_end_mask_0, squeeze_mask = var_1340_squeeze_mask_0, x = k_complex_7_cast_fp16)[name = tensor("op_1340_cast_fp16")]; + tensor freqs_7_to_fp16 = const()[name = tensor("freqs_7_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(82951680)))]; + tensor var_1346_cast_fp16 = mul(x = freqs_7_to_fp16, y = ts_23_cast_fp16)[name = tensor("op_1346_cast_fp16")]; + tensor rotr_7_cast_fp16 = cos(x = var_1346_cast_fp16)[name = tensor("rotr_7_cast_fp16")]; + tensor roti_7_cast_fp16 = sin(x = var_1346_cast_fp16)[name = tensor("roti_7_cast_fp16")]; + tensor var_1350_cast_fp16 = mul(x = var_1316_cast_fp16, y = rotr_7_cast_fp16)[name = tensor("op_1350_cast_fp16")]; + tensor var_1351_cast_fp16 = mul(x = var_1324_cast_fp16, y = roti_7_cast_fp16)[name = tensor("op_1351_cast_fp16")]; + tensor qor_13_cast_fp16 = sub(x = var_1350_cast_fp16, y = var_1351_cast_fp16)[name = tensor("qor_13_cast_fp16")]; + tensor var_1354_cast_fp16 = mul(x = var_1316_cast_fp16, y = roti_7_cast_fp16)[name = tensor("op_1354_cast_fp16")]; + tensor var_1355_cast_fp16 = mul(x = var_1324_cast_fp16, y = rotr_7_cast_fp16)[name = tensor("op_1355_cast_fp16")]; + tensor qoi_13_cast_fp16 = add(x = var_1354_cast_fp16, y = var_1355_cast_fp16)[name = tensor("qoi_13_cast_fp16")]; + tensor var_1358_cast_fp16 = mul(x = var_1332_cast_fp16, y = rotr_7_cast_fp16)[name = tensor("op_1358_cast_fp16")]; + tensor var_1359_cast_fp16 = mul(x = var_1340_cast_fp16, y = roti_7_cast_fp16)[name = tensor("op_1359_cast_fp16")]; + tensor kor_13_cast_fp16 = sub(x = var_1358_cast_fp16, y = var_1359_cast_fp16)[name = tensor("kor_13_cast_fp16")]; + tensor var_1362_cast_fp16 = mul(x = var_1332_cast_fp16, y = roti_7_cast_fp16)[name = tensor("op_1362_cast_fp16")]; + tensor var_1363_cast_fp16 = mul(x = var_1340_cast_fp16, y = rotr_7_cast_fp16)[name = tensor("op_1363_cast_fp16")]; + tensor koi_13_cast_fp16 = add(x = var_1362_cast_fp16, y = var_1363_cast_fp16)[name = tensor("koi_13_cast_fp16")]; + tensor qo_7_axis_0 = const()[name = tensor("qo_7_axis_0"), val = tensor(-1)]; + tensor qo_7_cast_fp16 = stack(axis = qo_7_axis_0, values = (qor_13_cast_fp16, qoi_13_cast_fp16))[name = tensor("qo_7_cast_fp16")]; + tensor ko_7_axis_0 = const()[name = tensor("ko_7_axis_0"), val = tensor(-1)]; + tensor ko_7_cast_fp16 = stack(axis = ko_7_axis_0, values = (kor_13_cast_fp16, koi_13_cast_fp16))[name = tensor("ko_7_cast_fp16")]; + tensor var_1396 = const()[name = tensor("op_1396"), val = tensor([1, 1, 16, 64])]; + tensor q_21_cast_fp16 = reshape(shape = var_1396, x = qo_7_cast_fp16)[name = tensor("q_21_cast_fp16")]; + tensor var_1402 = const()[name = tensor("op_1402"), val = tensor([1, 1, 16, 64])]; + tensor k_15_cast_fp16 = reshape(shape = var_1402, x = ko_7_cast_fp16)[name = tensor("k_15_cast_fp16")]; + tensor _inversed_1427_y_0_to_fp16 = const()[name = tensor("_inversed_1427_y_0_to_fp16"), val = tensor(0x1p-9)]; + tensor _inversed_1427_cast_fp16 = mul(x = ts_23_cast_fp16, y = _inversed_1427_y_0_to_fp16)[name = tensor("_inversed_1427_cast_fp16")]; + tensor var_1428_cast_fp16 = floor(x = _inversed_1427_cast_fp16)[name = tensor("op_1428_cast_fp16")]; + tensor var_1429_to_fp16 = const()[name = tensor("op_1429_to_fp16"), val = tensor(0x1p+9)]; + tensor var_1430_cast_fp16 = mul(x = var_1428_cast_fp16, y = var_1429_to_fp16)[name = tensor("op_1430_cast_fp16")]; + tensor write_indices_float_15_cast_fp16 = sub(x = ts_23_cast_fp16, y = var_1430_cast_fp16)[name = tensor("write_indices_float_15_cast_fp16")]; + tensor var_1437_dtype_0 = const()[name = tensor("op_1437_dtype_0"), val = tensor("int32")]; + tensor write_indices_7_reps_0 = const()[name = tensor("write_indices_7_reps_0"), val = tensor([1, 1, 16, 64])]; + tensor write_indices_float_15_cast_fp16_to_int32 = cast(dtype = var_1437_dtype_0, x = write_indices_float_15_cast_fp16)[name = tensor("cast_123")]; + tensor write_indices_7 = tile(reps = write_indices_7_reps_0, x = write_indices_float_15_cast_fp16_to_int32)[name = tensor("write_indices_7")]; + tensor var_1447_begin_0 = const()[name = tensor("op_1447_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_1447_end_0 = const()[name = tensor("op_1447_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor var_1447_end_mask_0 = const()[name = tensor("op_1447_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_1447_squeeze_mask_0 = const()[name = tensor("op_1447_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor cache3_to_fp16_dtype_0 = const()[name = tensor("cache3_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor cache3_to_fp16 = cast(dtype = cache3_to_fp16_dtype_0, x = cache3)[name = tensor("cast_122")]; + tensor var_1447_cast_fp16 = slice_by_index(begin = var_1447_begin_0, end = var_1447_end_0, end_mask = var_1447_end_mask_0, squeeze_mask = var_1447_squeeze_mask_0, x = cache3_to_fp16)[name = tensor("op_1447_cast_fp16")]; + tensor var_1449_axis_0 = const()[name = tensor("op_1449_axis_0"), val = tensor(1)]; + tensor var_1449_mode_0 = const()[name = tensor("op_1449_mode_0"), val = tensor("update")]; + tensor var_1449_validate_indices_0 = const()[name = tensor("op_1449_validate_indices_0"), val = tensor(false)]; + tensor var_1449_cast_fp16 = scatter_along_axis(axis = var_1449_axis_0, data = var_1447_cast_fp16, indices = write_indices_7, mode = var_1449_mode_0, updates = k_15_cast_fp16, validate_indices = var_1449_validate_indices_0)[name = tensor("op_1449_cast_fp16")]; + tensor concat_23 = const()[name = tensor("concat_23"), val = tensor([0, 0, 0, 0, 0])]; + tensor concat_24 = const()[name = tensor("concat_24"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_7_internal_tensor_assign_1_stride_0 = const()[name = tensor("new_cache_7_internal_tensor_assign_1_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_7_internal_tensor_assign_1_begin_mask_0 = const()[name = tensor("new_cache_7_internal_tensor_assign_1_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_7_internal_tensor_assign_1_end_mask_0 = const()[name = tensor("new_cache_7_internal_tensor_assign_1_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_7_internal_tensor_assign_1_squeeze_mask_0 = const()[name = tensor("new_cache_7_internal_tensor_assign_1_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_18 = const()[name = tensor("shape_18"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_6 = const()[name = tensor("reduce_prod_6"), val = tensor(1048576)]; + tensor range_1d_6_start_0 = const()[name = tensor("range_1d_6_start_0"), val = tensor(0)]; + tensor range_1d_6_step_0 = const()[name = tensor("range_1d_6_step_0"), val = tensor(1)]; + tensor range_1d_6 = range_1d(end = reduce_prod_6, start = range_1d_6_start_0, step = range_1d_6_step_0)[name = tensor("range_1d_6")]; + tensor reshape_30 = reshape(shape = shape_18, x = range_1d_6)[name = tensor("reshape_30")]; + tensor slice_by_index_6 = slice_by_index(begin = concat_23, begin_mask = new_cache_7_internal_tensor_assign_1_begin_mask_0, end = concat_24, end_mask = new_cache_7_internal_tensor_assign_1_end_mask_0, squeeze_mask = new_cache_7_internal_tensor_assign_1_squeeze_mask_0, stride = new_cache_7_internal_tensor_assign_1_stride_0, x = reshape_30)[name = tensor("slice_by_index_6")]; + tensor reshape_31_shape_0 = const()[name = tensor("reshape_31_shape_0"), val = tensor([-1])]; + tensor reshape_31 = reshape(shape = reshape_31_shape_0, x = slice_by_index_6)[name = tensor("reshape_31")]; + tensor reshape_32_shape_0 = const()[name = tensor("reshape_32_shape_0"), val = tensor([-1])]; + tensor reshape_32_cast_fp16 = reshape(shape = reshape_32_shape_0, x = var_1449_cast_fp16)[name = tensor("reshape_32_cast_fp16")]; + tensor reshape_33_shape_0 = const()[name = tensor("reshape_33_shape_0"), val = tensor([-1])]; + tensor reshape_33_cast_fp16 = reshape(shape = reshape_33_shape_0, x = cache3_to_fp16)[name = tensor("reshape_33_cast_fp16")]; + tensor scatter_6_mode_0 = const()[name = tensor("scatter_6_mode_0"), val = tensor("update")]; + tensor scatter_6_axis_0 = const()[name = tensor("scatter_6_axis_0"), val = tensor(0)]; + tensor scatter_6_validate_indices_0 = const()[name = tensor("scatter_6_validate_indices_0"), val = tensor(false)]; + tensor scatter_6_cast_fp16 = scatter(axis = scatter_6_axis_0, data = reshape_33_cast_fp16, indices = reshape_31, mode = scatter_6_mode_0, updates = reshape_32_cast_fp16, validate_indices = scatter_6_validate_indices_0)[name = tensor("scatter_6_cast_fp16")]; + tensor reshape_34_cast_fp16 = reshape(shape = shape_18, x = scatter_6_cast_fp16)[name = tensor("reshape_34_cast_fp16")]; + tensor var_1457_begin_0 = const()[name = tensor("op_1457_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor var_1457_end_0 = const()[name = tensor("op_1457_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor var_1457_end_mask_0 = const()[name = tensor("op_1457_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_1457_squeeze_mask_0 = const()[name = tensor("op_1457_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor var_1457_cast_fp16 = slice_by_index(begin = var_1457_begin_0, end = var_1457_end_0, end_mask = var_1457_end_mask_0, squeeze_mask = var_1457_squeeze_mask_0, x = reshape_34_cast_fp16)[name = tensor("op_1457_cast_fp16")]; + tensor var_1459_axis_0 = const()[name = tensor("op_1459_axis_0"), val = tensor(1)]; + tensor var_1459_mode_0 = const()[name = tensor("op_1459_mode_0"), val = tensor("update")]; + tensor var_1459_validate_indices_0 = const()[name = tensor("op_1459_validate_indices_0"), val = tensor(false)]; + tensor var_1459_cast_fp16 = scatter_along_axis(axis = var_1459_axis_0, data = var_1457_cast_fp16, indices = write_indices_7, mode = var_1459_mode_0, updates = v_7_cast_fp16, validate_indices = var_1459_validate_indices_0)[name = tensor("op_1459_cast_fp16")]; + tensor concat_25 = const()[name = tensor("concat_25"), val = tensor([1, 0, 0, 0, 0])]; + tensor concat_26 = const()[name = tensor("concat_26"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_7_internal_tensor_assign_2_stride_0 = const()[name = tensor("new_cache_7_internal_tensor_assign_2_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_7_internal_tensor_assign_2_begin_mask_0 = const()[name = tensor("new_cache_7_internal_tensor_assign_2_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_7_internal_tensor_assign_2_end_mask_0 = const()[name = tensor("new_cache_7_internal_tensor_assign_2_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_7_internal_tensor_assign_2_squeeze_mask_0 = const()[name = tensor("new_cache_7_internal_tensor_assign_2_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_19 = const()[name = tensor("shape_19"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_7 = const()[name = tensor("reduce_prod_7"), val = tensor(1048576)]; + tensor range_1d_7_start_0 = const()[name = tensor("range_1d_7_start_0"), val = tensor(0)]; + tensor range_1d_7_step_0 = const()[name = tensor("range_1d_7_step_0"), val = tensor(1)]; + tensor range_1d_7 = range_1d(end = reduce_prod_7, start = range_1d_7_start_0, step = range_1d_7_step_0)[name = tensor("range_1d_7")]; + tensor reshape_35 = reshape(shape = shape_19, x = range_1d_7)[name = tensor("reshape_35")]; + tensor slice_by_index_7 = slice_by_index(begin = concat_25, begin_mask = new_cache_7_internal_tensor_assign_2_begin_mask_0, end = concat_26, end_mask = new_cache_7_internal_tensor_assign_2_end_mask_0, squeeze_mask = new_cache_7_internal_tensor_assign_2_squeeze_mask_0, stride = new_cache_7_internal_tensor_assign_2_stride_0, x = reshape_35)[name = tensor("slice_by_index_7")]; + tensor reshape_36_shape_0 = const()[name = tensor("reshape_36_shape_0"), val = tensor([-1])]; + tensor reshape_36 = reshape(shape = reshape_36_shape_0, x = slice_by_index_7)[name = tensor("reshape_36")]; + tensor reshape_37_shape_0 = const()[name = tensor("reshape_37_shape_0"), val = tensor([-1])]; + tensor reshape_37_cast_fp16 = reshape(shape = reshape_37_shape_0, x = var_1459_cast_fp16)[name = tensor("reshape_37_cast_fp16")]; + tensor reshape_38_shape_0 = const()[name = tensor("reshape_38_shape_0"), val = tensor([-1])]; + tensor reshape_38_cast_fp16 = reshape(shape = reshape_38_shape_0, x = reshape_34_cast_fp16)[name = tensor("reshape_38_cast_fp16")]; + tensor scatter_7_mode_0 = const()[name = tensor("scatter_7_mode_0"), val = tensor("update")]; + tensor scatter_7_axis_0 = const()[name = tensor("scatter_7_axis_0"), val = tensor(0)]; + tensor scatter_7_validate_indices_0 = const()[name = tensor("scatter_7_validate_indices_0"), val = tensor(false)]; + tensor scatter_7_cast_fp16 = scatter(axis = scatter_7_axis_0, data = reshape_38_cast_fp16, indices = reshape_36, mode = scatter_7_mode_0, updates = reshape_37_cast_fp16, validate_indices = scatter_7_validate_indices_0)[name = tensor("scatter_7_cast_fp16")]; + tensor reshape_39_cast_fp16 = reshape(shape = shape_19, x = scatter_7_cast_fp16)[name = tensor("reshape_39_cast_fp16")]; + tensor reshape_39_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("reshape_39_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor keys_19_begin_0 = const()[name = tensor("keys_19_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor keys_19_end_0 = const()[name = tensor("keys_19_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor keys_19_end_mask_0 = const()[name = tensor("keys_19_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor keys_19_squeeze_mask_0 = const()[name = tensor("keys_19_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor keys_19_cast_fp16 = slice_by_index(begin = keys_19_begin_0, end = keys_19_end_0, end_mask = keys_19_end_mask_0, squeeze_mask = keys_19_squeeze_mask_0, x = reshape_39_cast_fp16)[name = tensor("keys_19_cast_fp16")]; + tensor values_19_begin_0 = const()[name = tensor("values_19_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor values_19_end_0 = const()[name = tensor("values_19_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor values_19_end_mask_0 = const()[name = tensor("values_19_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor values_19_squeeze_mask_0 = const()[name = tensor("values_19_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor values_19_cast_fp16 = slice_by_index(begin = values_19_begin_0, end = values_19_end_0, end_mask = values_19_end_mask_0, squeeze_mask = values_19_squeeze_mask_0, x = reshape_39_cast_fp16)[name = tensor("values_19_cast_fp16")]; + tensor var_1471_cast_fp16 = not_equal(x = keys_19_cast_fp16, y = keys_19_cast_fp16)[name = tensor("op_1471_cast_fp16")]; + tensor keys_21_cast_fp16 = select(a = var_319_to_fp16, b = keys_19_cast_fp16, cond = var_1471_cast_fp16)[name = tensor("keys_21_cast_fp16")]; + tensor var_1479_cast_fp16 = not_equal(x = values_19_cast_fp16, y = values_19_cast_fp16)[name = tensor("op_1479_cast_fp16")]; + tensor values_21_cast_fp16 = select(a = var_319_to_fp16, b = values_19_cast_fp16, cond = var_1479_cast_fp16)[name = tensor("values_21_cast_fp16")]; + tensor var_1503 = const()[name = tensor("op_1503"), val = tensor([0, 2, 1, 3])]; + tensor var_1519 = const()[name = tensor("op_1519"), val = tensor([1, 1, 1])]; + tensor var_1520_cast_fp16 = reshape(shape = var_1519, x = position3_to_fp16)[name = tensor("op_1520_cast_fp16")]; + tensor var_1540_to_fp16 = const()[name = tensor("op_1540_to_fp16"), val = tensor(0x1p+0)]; + tensor valid_len_7_cast_fp16 = add(x = var_1520_cast_fp16, y = var_1540_to_fp16)[name = tensor("valid_len_7_cast_fp16")]; + tensor valid_mask_7_cast_fp16 = less(x = k_positions_1_promoted_to_fp16, y = valid_len_7_cast_fp16)[name = tensor("valid_mask_7_cast_fp16")]; + tensor causal_mask_7_cast_fp16 = less_equal(x = k_positions_1_promoted_to_fp16, y = var_1520_cast_fp16)[name = tensor("causal_mask_7_cast_fp16")]; + tensor attn_mask_13 = logical_and(x = valid_mask_7_cast_fp16, y = causal_mask_7_cast_fp16)[name = tensor("attn_mask_13")]; + tensor attn_mask_15_axes_0 = const()[name = tensor("attn_mask_15_axes_0"), val = tensor([1])]; + tensor attn_mask_15 = expand_dims(axes = attn_mask_15_axes_0, x = attn_mask_13)[name = tensor("attn_mask_15")]; + tensor var_1558_transpose_x_0 = const()[name = tensor("op_1558_transpose_x_0"), val = tensor(false)]; + tensor var_1558_transpose_y_0 = const()[name = tensor("op_1558_transpose_y_0"), val = tensor(false)]; + tensor transpose_24_perm_0 = const()[name = tensor("transpose_24_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_25_perm_0 = const()[name = tensor("transpose_25_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_25 = transpose(perm = transpose_25_perm_0, x = keys_21_cast_fp16)[name = tensor("transpose_39")]; + tensor transpose_24 = transpose(perm = transpose_24_perm_0, x = q_21_cast_fp16)[name = tensor("transpose_40")]; + tensor var_1558_cast_fp16 = matmul(transpose_x = var_1558_transpose_x_0, transpose_y = var_1558_transpose_y_0, x = transpose_24, y = transpose_25)[name = tensor("op_1558_cast_fp16")]; + tensor var_1552_to_fp16 = const()[name = tensor("op_1552_to_fp16"), val = tensor([0x1p-3])]; + tensor attn_weights_19_cast_fp16 = mul(x = var_1558_cast_fp16, y = var_1552_to_fp16)[name = tensor("attn_weights_19_cast_fp16")]; + tensor var_1560 = logical_not(x = attn_mask_15)[name = tensor("op_1560")]; + tensor var_1561_to_fp16 = const()[name = tensor("op_1561_to_fp16"), val = tensor(-inf)]; + tensor attn_weights_21_cast_fp16 = select(a = var_1561_to_fp16, b = attn_weights_19_cast_fp16, cond = var_1560)[name = tensor("attn_weights_21_cast_fp16")]; + tensor var_1563 = const()[name = tensor("op_1563"), val = tensor(-1)]; + tensor attn_weights_23_cast_fp16 = softmax(axis = var_1563, x = attn_weights_21_cast_fp16)[name = tensor("attn_weights_23_cast_fp16")]; + tensor attn_output_7_transpose_x_0 = const()[name = tensor("attn_output_7_transpose_x_0"), val = tensor(false)]; + tensor attn_output_7_transpose_y_0 = const()[name = tensor("attn_output_7_transpose_y_0"), val = tensor(false)]; + tensor values_23_cast_fp16 = transpose(perm = var_1503, x = values_21_cast_fp16)[name = tensor("transpose_41")]; + tensor attn_output_7_cast_fp16 = matmul(transpose_x = attn_output_7_transpose_x_0, transpose_y = attn_output_7_transpose_y_0, x = attn_weights_23_cast_fp16, y = values_23_cast_fp16)[name = tensor("attn_output_7_cast_fp16")]; + tensor var_1571 = const()[name = tensor("op_1571"), val = tensor([0, 2, 1, 3])]; + tensor var_1576 = const()[name = tensor("op_1576"), val = tensor([1, 1, 1024])]; + tensor var_1572_cast_fp16 = transpose(perm = var_1571, x = attn_output_7_cast_fp16)[name = tensor("transpose_38")]; + tensor input_43_cast_fp16 = reshape(shape = var_1576, x = var_1572_cast_fp16)[name = tensor("input_43_cast_fp16")]; + tensor attn3_out_proj_weight_to_fp16 = const()[name = tensor("attn3_out_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(82951808)))]; + tensor linear_14_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = attn3_out_proj_weight_to_fp16, x = input_43_cast_fp16)[name = tensor("linear_14_cast_fp16")]; + tensor var_1582_to_fp16 = const()[name = tensor("op_1582_to_fp16"), val = tensor(0x1p+0)]; + tensor var_1583_cast_fp16 = add(x = position3_to_fp16, y = var_1582_to_fp16)[name = tensor("op_1583_cast_fp16")]; + tensor var_1583_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_1583_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor input_45_cast_fp16 = add(x = input_39_cast_fp16, y = linear_14_cast_fp16)[name = tensor("input_45_cast_fp16")]; + tensor input_47_axes_0 = const()[name = tensor("input_47_axes_0"), val = tensor([-1])]; + tensor norm3_2_weight_to_fp16 = const()[name = tensor("norm3_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(85049024)))]; + tensor norm3_2_bias_to_fp16 = const()[name = tensor("norm3_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(85051136)))]; + tensor var_1587_to_fp16 = const()[name = tensor("op_1587_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_47_cast_fp16 = layer_norm(axes = input_47_axes_0, beta = norm3_2_bias_to_fp16, epsilon = var_1587_to_fp16, gamma = norm3_2_weight_to_fp16, x = input_45_cast_fp16)[name = tensor("input_47_cast_fp16")]; + tensor linear3_1_weight_to_fp16 = const()[name = tensor("linear3_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(85053248)))]; + tensor linear_15_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = linear3_1_weight_to_fp16, x = input_47_cast_fp16)[name = tensor("linear_15_cast_fp16")]; + tensor input_49_mode_0 = const()[name = tensor("input_49_mode_0"), val = tensor("EXACT")]; + tensor input_49_cast_fp16 = gelu(mode = input_49_mode_0, x = linear_15_cast_fp16)[name = tensor("input_49_cast_fp16")]; + tensor linear3_2_weight_to_fp16 = const()[name = tensor("linear3_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(93441920)))]; + tensor linear_16_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = linear3_2_weight_to_fp16, x = input_49_cast_fp16)[name = tensor("linear_16_cast_fp16")]; + tensor input_51_cast_fp16 = add(x = input_45_cast_fp16, y = linear_16_cast_fp16)[name = tensor("input_51_cast_fp16")]; + tensor input_53_axes_0 = const()[name = tensor("input_53_axes_0"), val = tensor([-1])]; + tensor norm4_1_weight_to_fp16 = const()[name = tensor("norm4_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(101830592)))]; + tensor norm4_1_bias_to_fp16 = const()[name = tensor("norm4_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(101832704)))]; + tensor var_1604_to_fp16 = const()[name = tensor("op_1604_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_53_cast_fp16 = layer_norm(axes = input_53_axes_0, beta = norm4_1_bias_to_fp16, epsilon = var_1604_to_fp16, gamma = norm4_1_weight_to_fp16, x = input_51_cast_fp16)[name = tensor("input_53_cast_fp16")]; + tensor attn4_in_proj_weight_to_fp16 = const()[name = tensor("attn4_in_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(101834816)))]; + tensor linear_17_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = attn4_in_proj_weight_to_fp16, x = input_53_cast_fp16)[name = tensor("linear_17_cast_fp16")]; + tensor var_1618 = const()[name = tensor("op_1618"), val = tensor([1, 1, 3, 16, 64])]; + tensor qkv_9_cast_fp16 = reshape(shape = var_1618, x = linear_17_cast_fp16)[name = tensor("qkv_9_cast_fp16")]; + tensor q_25_begin_0 = const()[name = tensor("q_25_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor q_25_end_0 = const()[name = tensor("q_25_end_0"), val = tensor([1, 1, 1, 16, 64])]; + tensor q_25_end_mask_0 = const()[name = tensor("q_25_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor q_25_squeeze_mask_0 = const()[name = tensor("q_25_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor q_25_cast_fp16 = slice_by_index(begin = q_25_begin_0, end = q_25_end_0, end_mask = q_25_end_mask_0, squeeze_mask = q_25_squeeze_mask_0, x = qkv_9_cast_fp16)[name = tensor("q_25_cast_fp16")]; + tensor k_17_begin_0 = const()[name = tensor("k_17_begin_0"), val = tensor([0, 0, 1, 0, 0])]; + tensor k_17_end_0 = const()[name = tensor("k_17_end_0"), val = tensor([1, 1, 2, 16, 64])]; + tensor k_17_end_mask_0 = const()[name = tensor("k_17_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor k_17_squeeze_mask_0 = const()[name = tensor("k_17_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor k_17_cast_fp16 = slice_by_index(begin = k_17_begin_0, end = k_17_end_0, end_mask = k_17_end_mask_0, squeeze_mask = k_17_squeeze_mask_0, x = qkv_9_cast_fp16)[name = tensor("k_17_cast_fp16")]; + tensor v_9_begin_0 = const()[name = tensor("v_9_begin_0"), val = tensor([0, 0, 2, 0, 0])]; + tensor v_9_end_0 = const()[name = tensor("v_9_end_0"), val = tensor([1, 1, 3, 16, 64])]; + tensor v_9_end_mask_0 = const()[name = tensor("v_9_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor v_9_squeeze_mask_0 = const()[name = tensor("v_9_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor v_9_cast_fp16 = slice_by_index(begin = v_9_begin_0, end = v_9_end_0, end_mask = v_9_end_mask_0, squeeze_mask = v_9_squeeze_mask_0, x = qkv_9_cast_fp16)[name = tensor("v_9_cast_fp16")]; + tensor var_1684 = const()[name = tensor("op_1684"), val = tensor([1, 1, 1, 1])]; + tensor position4_to_fp16_dtype_0 = const()[name = tensor("position4_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor position4_to_fp16 = cast(dtype = position4_to_fp16_dtype_0, x = position4)[name = tensor("cast_119")]; + tensor ts_29_cast_fp16 = reshape(shape = var_1684, x = position4_to_fp16)[name = tensor("ts_29_cast_fp16")]; + tensor var_1691 = const()[name = tensor("op_1691"), val = tensor([1, 1, 16, 32, 2])]; + tensor q_complex_9_cast_fp16 = reshape(shape = var_1691, x = q_25_cast_fp16)[name = tensor("q_complex_9_cast_fp16")]; + tensor var_1698 = const()[name = tensor("op_1698"), val = tensor([1, 1, 16, 32, 2])]; + tensor k_complex_9_cast_fp16 = reshape(shape = var_1698, x = k_17_cast_fp16)[name = tensor("k_complex_9_cast_fp16")]; + tensor var_1702_begin_0 = const()[name = tensor("op_1702_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_1702_end_0 = const()[name = tensor("op_1702_end_0"), val = tensor([1, 1, 16, 32, 1])]; + tensor var_1702_end_mask_0 = const()[name = tensor("op_1702_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_1702_squeeze_mask_0 = const()[name = tensor("op_1702_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_1702_cast_fp16 = slice_by_index(begin = var_1702_begin_0, end = var_1702_end_0, end_mask = var_1702_end_mask_0, squeeze_mask = var_1702_squeeze_mask_0, x = q_complex_9_cast_fp16)[name = tensor("op_1702_cast_fp16")]; + tensor var_1710_begin_0 = const()[name = tensor("op_1710_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_1710_end_0 = const()[name = tensor("op_1710_end_0"), val = tensor([1, 1, 16, 32, 2])]; + tensor var_1710_end_mask_0 = const()[name = tensor("op_1710_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_1710_squeeze_mask_0 = const()[name = tensor("op_1710_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_1710_cast_fp16 = slice_by_index(begin = var_1710_begin_0, end = var_1710_end_0, end_mask = var_1710_end_mask_0, squeeze_mask = var_1710_squeeze_mask_0, x = q_complex_9_cast_fp16)[name = tensor("op_1710_cast_fp16")]; + tensor var_1718_begin_0 = const()[name = tensor("op_1718_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_1718_end_0 = const()[name = tensor("op_1718_end_0"), val = tensor([1, 1, 16, 32, 1])]; + tensor var_1718_end_mask_0 = const()[name = tensor("op_1718_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_1718_squeeze_mask_0 = const()[name = tensor("op_1718_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_1718_cast_fp16 = slice_by_index(begin = var_1718_begin_0, end = var_1718_end_0, end_mask = var_1718_end_mask_0, squeeze_mask = var_1718_squeeze_mask_0, x = k_complex_9_cast_fp16)[name = tensor("op_1718_cast_fp16")]; + tensor var_1726_begin_0 = const()[name = tensor("op_1726_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_1726_end_0 = const()[name = tensor("op_1726_end_0"), val = tensor([1, 1, 16, 32, 2])]; + tensor var_1726_end_mask_0 = const()[name = tensor("op_1726_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_1726_squeeze_mask_0 = const()[name = tensor("op_1726_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_1726_cast_fp16 = slice_by_index(begin = var_1726_begin_0, end = var_1726_end_0, end_mask = var_1726_end_mask_0, squeeze_mask = var_1726_squeeze_mask_0, x = k_complex_9_cast_fp16)[name = tensor("op_1726_cast_fp16")]; + tensor freqs_9_to_fp16 = const()[name = tensor("freqs_9_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(108126336)))]; + tensor var_1732_cast_fp16 = mul(x = freqs_9_to_fp16, y = ts_29_cast_fp16)[name = tensor("op_1732_cast_fp16")]; + tensor rotr_9_cast_fp16 = cos(x = var_1732_cast_fp16)[name = tensor("rotr_9_cast_fp16")]; + tensor roti_9_cast_fp16 = sin(x = var_1732_cast_fp16)[name = tensor("roti_9_cast_fp16")]; + tensor var_1736_cast_fp16 = mul(x = var_1702_cast_fp16, y = rotr_9_cast_fp16)[name = tensor("op_1736_cast_fp16")]; + tensor var_1737_cast_fp16 = mul(x = var_1710_cast_fp16, y = roti_9_cast_fp16)[name = tensor("op_1737_cast_fp16")]; + tensor qor_17_cast_fp16 = sub(x = var_1736_cast_fp16, y = var_1737_cast_fp16)[name = tensor("qor_17_cast_fp16")]; + tensor var_1740_cast_fp16 = mul(x = var_1702_cast_fp16, y = roti_9_cast_fp16)[name = tensor("op_1740_cast_fp16")]; + tensor var_1741_cast_fp16 = mul(x = var_1710_cast_fp16, y = rotr_9_cast_fp16)[name = tensor("op_1741_cast_fp16")]; + tensor qoi_17_cast_fp16 = add(x = var_1740_cast_fp16, y = var_1741_cast_fp16)[name = tensor("qoi_17_cast_fp16")]; + tensor var_1744_cast_fp16 = mul(x = var_1718_cast_fp16, y = rotr_9_cast_fp16)[name = tensor("op_1744_cast_fp16")]; + tensor var_1745_cast_fp16 = mul(x = var_1726_cast_fp16, y = roti_9_cast_fp16)[name = tensor("op_1745_cast_fp16")]; + tensor kor_17_cast_fp16 = sub(x = var_1744_cast_fp16, y = var_1745_cast_fp16)[name = tensor("kor_17_cast_fp16")]; + tensor var_1748_cast_fp16 = mul(x = var_1718_cast_fp16, y = roti_9_cast_fp16)[name = tensor("op_1748_cast_fp16")]; + tensor var_1749_cast_fp16 = mul(x = var_1726_cast_fp16, y = rotr_9_cast_fp16)[name = tensor("op_1749_cast_fp16")]; + tensor koi_17_cast_fp16 = add(x = var_1748_cast_fp16, y = var_1749_cast_fp16)[name = tensor("koi_17_cast_fp16")]; + tensor qo_9_axis_0 = const()[name = tensor("qo_9_axis_0"), val = tensor(-1)]; + tensor qo_9_cast_fp16 = stack(axis = qo_9_axis_0, values = (qor_17_cast_fp16, qoi_17_cast_fp16))[name = tensor("qo_9_cast_fp16")]; + tensor ko_9_axis_0 = const()[name = tensor("ko_9_axis_0"), val = tensor(-1)]; + tensor ko_9_cast_fp16 = stack(axis = ko_9_axis_0, values = (kor_17_cast_fp16, koi_17_cast_fp16))[name = tensor("ko_9_cast_fp16")]; + tensor var_1782 = const()[name = tensor("op_1782"), val = tensor([1, 1, 16, 64])]; + tensor q_27_cast_fp16 = reshape(shape = var_1782, x = qo_9_cast_fp16)[name = tensor("q_27_cast_fp16")]; + tensor var_1788 = const()[name = tensor("op_1788"), val = tensor([1, 1, 16, 64])]; + tensor k_19_cast_fp16 = reshape(shape = var_1788, x = ko_9_cast_fp16)[name = tensor("k_19_cast_fp16")]; + tensor _inversed_1813_y_0_to_fp16 = const()[name = tensor("_inversed_1813_y_0_to_fp16"), val = tensor(0x1p-9)]; + tensor _inversed_1813_cast_fp16 = mul(x = ts_29_cast_fp16, y = _inversed_1813_y_0_to_fp16)[name = tensor("_inversed_1813_cast_fp16")]; + tensor var_1814_cast_fp16 = floor(x = _inversed_1813_cast_fp16)[name = tensor("op_1814_cast_fp16")]; + tensor var_1815_to_fp16 = const()[name = tensor("op_1815_to_fp16"), val = tensor(0x1p+9)]; + tensor var_1816_cast_fp16 = mul(x = var_1814_cast_fp16, y = var_1815_to_fp16)[name = tensor("op_1816_cast_fp16")]; + tensor write_indices_float_19_cast_fp16 = sub(x = ts_29_cast_fp16, y = var_1816_cast_fp16)[name = tensor("write_indices_float_19_cast_fp16")]; + tensor var_1823_dtype_0 = const()[name = tensor("op_1823_dtype_0"), val = tensor("int32")]; + tensor write_indices_9_reps_0 = const()[name = tensor("write_indices_9_reps_0"), val = tensor([1, 1, 16, 64])]; + tensor write_indices_float_19_cast_fp16_to_int32 = cast(dtype = var_1823_dtype_0, x = write_indices_float_19_cast_fp16)[name = tensor("cast_118")]; + tensor write_indices_9 = tile(reps = write_indices_9_reps_0, x = write_indices_float_19_cast_fp16_to_int32)[name = tensor("write_indices_9")]; + tensor var_1833_begin_0 = const()[name = tensor("op_1833_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_1833_end_0 = const()[name = tensor("op_1833_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor var_1833_end_mask_0 = const()[name = tensor("op_1833_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_1833_squeeze_mask_0 = const()[name = tensor("op_1833_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor cache4_to_fp16_dtype_0 = const()[name = tensor("cache4_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor cache4_to_fp16 = cast(dtype = cache4_to_fp16_dtype_0, x = cache4)[name = tensor("cast_117")]; + tensor var_1833_cast_fp16 = slice_by_index(begin = var_1833_begin_0, end = var_1833_end_0, end_mask = var_1833_end_mask_0, squeeze_mask = var_1833_squeeze_mask_0, x = cache4_to_fp16)[name = tensor("op_1833_cast_fp16")]; + tensor var_1835_axis_0 = const()[name = tensor("op_1835_axis_0"), val = tensor(1)]; + tensor var_1835_mode_0 = const()[name = tensor("op_1835_mode_0"), val = tensor("update")]; + tensor var_1835_validate_indices_0 = const()[name = tensor("op_1835_validate_indices_0"), val = tensor(false)]; + tensor var_1835_cast_fp16 = scatter_along_axis(axis = var_1835_axis_0, data = var_1833_cast_fp16, indices = write_indices_9, mode = var_1835_mode_0, updates = k_19_cast_fp16, validate_indices = var_1835_validate_indices_0)[name = tensor("op_1835_cast_fp16")]; + tensor concat_30 = const()[name = tensor("concat_30"), val = tensor([0, 0, 0, 0, 0])]; + tensor concat_31 = const()[name = tensor("concat_31"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_9_internal_tensor_assign_1_stride_0 = const()[name = tensor("new_cache_9_internal_tensor_assign_1_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_9_internal_tensor_assign_1_begin_mask_0 = const()[name = tensor("new_cache_9_internal_tensor_assign_1_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_9_internal_tensor_assign_1_end_mask_0 = const()[name = tensor("new_cache_9_internal_tensor_assign_1_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_9_internal_tensor_assign_1_squeeze_mask_0 = const()[name = tensor("new_cache_9_internal_tensor_assign_1_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_20 = const()[name = tensor("shape_20"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_8 = const()[name = tensor("reduce_prod_8"), val = tensor(1048576)]; + tensor range_1d_8_start_0 = const()[name = tensor("range_1d_8_start_0"), val = tensor(0)]; + tensor range_1d_8_step_0 = const()[name = tensor("range_1d_8_step_0"), val = tensor(1)]; + tensor range_1d_8 = range_1d(end = reduce_prod_8, start = range_1d_8_start_0, step = range_1d_8_step_0)[name = tensor("range_1d_8")]; + tensor reshape_40 = reshape(shape = shape_20, x = range_1d_8)[name = tensor("reshape_40")]; + tensor slice_by_index_8 = slice_by_index(begin = concat_30, begin_mask = new_cache_9_internal_tensor_assign_1_begin_mask_0, end = concat_31, end_mask = new_cache_9_internal_tensor_assign_1_end_mask_0, squeeze_mask = new_cache_9_internal_tensor_assign_1_squeeze_mask_0, stride = new_cache_9_internal_tensor_assign_1_stride_0, x = reshape_40)[name = tensor("slice_by_index_8")]; + tensor reshape_41_shape_0 = const()[name = tensor("reshape_41_shape_0"), val = tensor([-1])]; + tensor reshape_41 = reshape(shape = reshape_41_shape_0, x = slice_by_index_8)[name = tensor("reshape_41")]; + tensor reshape_42_shape_0 = const()[name = tensor("reshape_42_shape_0"), val = tensor([-1])]; + tensor reshape_42_cast_fp16 = reshape(shape = reshape_42_shape_0, x = var_1835_cast_fp16)[name = tensor("reshape_42_cast_fp16")]; + tensor reshape_43_shape_0 = const()[name = tensor("reshape_43_shape_0"), val = tensor([-1])]; + tensor reshape_43_cast_fp16 = reshape(shape = reshape_43_shape_0, x = cache4_to_fp16)[name = tensor("reshape_43_cast_fp16")]; + tensor scatter_8_mode_0 = const()[name = tensor("scatter_8_mode_0"), val = tensor("update")]; + tensor scatter_8_axis_0 = const()[name = tensor("scatter_8_axis_0"), val = tensor(0)]; + tensor scatter_8_validate_indices_0 = const()[name = tensor("scatter_8_validate_indices_0"), val = tensor(false)]; + tensor scatter_8_cast_fp16 = scatter(axis = scatter_8_axis_0, data = reshape_43_cast_fp16, indices = reshape_41, mode = scatter_8_mode_0, updates = reshape_42_cast_fp16, validate_indices = scatter_8_validate_indices_0)[name = tensor("scatter_8_cast_fp16")]; + tensor reshape_44_cast_fp16 = reshape(shape = shape_20, x = scatter_8_cast_fp16)[name = tensor("reshape_44_cast_fp16")]; + tensor var_1843_begin_0 = const()[name = tensor("op_1843_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor var_1843_end_0 = const()[name = tensor("op_1843_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor var_1843_end_mask_0 = const()[name = tensor("op_1843_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_1843_squeeze_mask_0 = const()[name = tensor("op_1843_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor var_1843_cast_fp16 = slice_by_index(begin = var_1843_begin_0, end = var_1843_end_0, end_mask = var_1843_end_mask_0, squeeze_mask = var_1843_squeeze_mask_0, x = reshape_44_cast_fp16)[name = tensor("op_1843_cast_fp16")]; + tensor var_1845_axis_0 = const()[name = tensor("op_1845_axis_0"), val = tensor(1)]; + tensor var_1845_mode_0 = const()[name = tensor("op_1845_mode_0"), val = tensor("update")]; + tensor var_1845_validate_indices_0 = const()[name = tensor("op_1845_validate_indices_0"), val = tensor(false)]; + tensor var_1845_cast_fp16 = scatter_along_axis(axis = var_1845_axis_0, data = var_1843_cast_fp16, indices = write_indices_9, mode = var_1845_mode_0, updates = v_9_cast_fp16, validate_indices = var_1845_validate_indices_0)[name = tensor("op_1845_cast_fp16")]; + tensor concat_32 = const()[name = tensor("concat_32"), val = tensor([1, 0, 0, 0, 0])]; + tensor concat_33 = const()[name = tensor("concat_33"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_9_internal_tensor_assign_2_stride_0 = const()[name = tensor("new_cache_9_internal_tensor_assign_2_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_9_internal_tensor_assign_2_begin_mask_0 = const()[name = tensor("new_cache_9_internal_tensor_assign_2_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_9_internal_tensor_assign_2_end_mask_0 = const()[name = tensor("new_cache_9_internal_tensor_assign_2_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_9_internal_tensor_assign_2_squeeze_mask_0 = const()[name = tensor("new_cache_9_internal_tensor_assign_2_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_21 = const()[name = tensor("shape_21"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_9 = const()[name = tensor("reduce_prod_9"), val = tensor(1048576)]; + tensor range_1d_9_start_0 = const()[name = tensor("range_1d_9_start_0"), val = tensor(0)]; + tensor range_1d_9_step_0 = const()[name = tensor("range_1d_9_step_0"), val = tensor(1)]; + tensor range_1d_9 = range_1d(end = reduce_prod_9, start = range_1d_9_start_0, step = range_1d_9_step_0)[name = tensor("range_1d_9")]; + tensor reshape_45 = reshape(shape = shape_21, x = range_1d_9)[name = tensor("reshape_45")]; + tensor slice_by_index_9 = slice_by_index(begin = concat_32, begin_mask = new_cache_9_internal_tensor_assign_2_begin_mask_0, end = concat_33, end_mask = new_cache_9_internal_tensor_assign_2_end_mask_0, squeeze_mask = new_cache_9_internal_tensor_assign_2_squeeze_mask_0, stride = new_cache_9_internal_tensor_assign_2_stride_0, x = reshape_45)[name = tensor("slice_by_index_9")]; + tensor reshape_46_shape_0 = const()[name = tensor("reshape_46_shape_0"), val = tensor([-1])]; + tensor reshape_46 = reshape(shape = reshape_46_shape_0, x = slice_by_index_9)[name = tensor("reshape_46")]; + tensor reshape_47_shape_0 = const()[name = tensor("reshape_47_shape_0"), val = tensor([-1])]; + tensor reshape_47_cast_fp16 = reshape(shape = reshape_47_shape_0, x = var_1845_cast_fp16)[name = tensor("reshape_47_cast_fp16")]; + tensor reshape_48_shape_0 = const()[name = tensor("reshape_48_shape_0"), val = tensor([-1])]; + tensor reshape_48_cast_fp16 = reshape(shape = reshape_48_shape_0, x = reshape_44_cast_fp16)[name = tensor("reshape_48_cast_fp16")]; + tensor scatter_9_mode_0 = const()[name = tensor("scatter_9_mode_0"), val = tensor("update")]; + tensor scatter_9_axis_0 = const()[name = tensor("scatter_9_axis_0"), val = tensor(0)]; + tensor scatter_9_validate_indices_0 = const()[name = tensor("scatter_9_validate_indices_0"), val = tensor(false)]; + tensor scatter_9_cast_fp16 = scatter(axis = scatter_9_axis_0, data = reshape_48_cast_fp16, indices = reshape_46, mode = scatter_9_mode_0, updates = reshape_47_cast_fp16, validate_indices = scatter_9_validate_indices_0)[name = tensor("scatter_9_cast_fp16")]; + tensor reshape_49_cast_fp16 = reshape(shape = shape_21, x = scatter_9_cast_fp16)[name = tensor("reshape_49_cast_fp16")]; + tensor reshape_49_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("reshape_49_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor keys_25_begin_0 = const()[name = tensor("keys_25_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor keys_25_end_0 = const()[name = tensor("keys_25_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor keys_25_end_mask_0 = const()[name = tensor("keys_25_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor keys_25_squeeze_mask_0 = const()[name = tensor("keys_25_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor keys_25_cast_fp16 = slice_by_index(begin = keys_25_begin_0, end = keys_25_end_0, end_mask = keys_25_end_mask_0, squeeze_mask = keys_25_squeeze_mask_0, x = reshape_49_cast_fp16)[name = tensor("keys_25_cast_fp16")]; + tensor values_25_begin_0 = const()[name = tensor("values_25_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor values_25_end_0 = const()[name = tensor("values_25_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor values_25_end_mask_0 = const()[name = tensor("values_25_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor values_25_squeeze_mask_0 = const()[name = tensor("values_25_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor values_25_cast_fp16 = slice_by_index(begin = values_25_begin_0, end = values_25_end_0, end_mask = values_25_end_mask_0, squeeze_mask = values_25_squeeze_mask_0, x = reshape_49_cast_fp16)[name = tensor("values_25_cast_fp16")]; + tensor var_1857_cast_fp16 = not_equal(x = keys_25_cast_fp16, y = keys_25_cast_fp16)[name = tensor("op_1857_cast_fp16")]; + tensor keys_27_cast_fp16 = select(a = var_319_to_fp16, b = keys_25_cast_fp16, cond = var_1857_cast_fp16)[name = tensor("keys_27_cast_fp16")]; + tensor var_1865_cast_fp16 = not_equal(x = values_25_cast_fp16, y = values_25_cast_fp16)[name = tensor("op_1865_cast_fp16")]; + tensor values_27_cast_fp16 = select(a = var_319_to_fp16, b = values_25_cast_fp16, cond = var_1865_cast_fp16)[name = tensor("values_27_cast_fp16")]; + tensor var_1889 = const()[name = tensor("op_1889"), val = tensor([0, 2, 1, 3])]; + tensor var_1905 = const()[name = tensor("op_1905"), val = tensor([1, 1, 1])]; + tensor var_1906_cast_fp16 = reshape(shape = var_1905, x = position4_to_fp16)[name = tensor("op_1906_cast_fp16")]; + tensor var_1926_to_fp16 = const()[name = tensor("op_1926_to_fp16"), val = tensor(0x1p+0)]; + tensor valid_len_9_cast_fp16 = add(x = var_1906_cast_fp16, y = var_1926_to_fp16)[name = tensor("valid_len_9_cast_fp16")]; + tensor valid_mask_9_cast_fp16 = less(x = k_positions_1_promoted_to_fp16, y = valid_len_9_cast_fp16)[name = tensor("valid_mask_9_cast_fp16")]; + tensor causal_mask_9_cast_fp16 = less_equal(x = k_positions_1_promoted_to_fp16, y = var_1906_cast_fp16)[name = tensor("causal_mask_9_cast_fp16")]; + tensor attn_mask_17 = logical_and(x = valid_mask_9_cast_fp16, y = causal_mask_9_cast_fp16)[name = tensor("attn_mask_17")]; + tensor attn_mask_19_axes_0 = const()[name = tensor("attn_mask_19_axes_0"), val = tensor([1])]; + tensor attn_mask_19 = expand_dims(axes = attn_mask_19_axes_0, x = attn_mask_17)[name = tensor("attn_mask_19")]; + tensor var_1944_transpose_x_0 = const()[name = tensor("op_1944_transpose_x_0"), val = tensor(false)]; + tensor var_1944_transpose_y_0 = const()[name = tensor("op_1944_transpose_y_0"), val = tensor(false)]; + tensor transpose_26_perm_0 = const()[name = tensor("transpose_26_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_27_perm_0 = const()[name = tensor("transpose_27_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_27 = transpose(perm = transpose_27_perm_0, x = keys_27_cast_fp16)[name = tensor("transpose_35")]; + tensor transpose_26 = transpose(perm = transpose_26_perm_0, x = q_27_cast_fp16)[name = tensor("transpose_36")]; + tensor var_1944_cast_fp16 = matmul(transpose_x = var_1944_transpose_x_0, transpose_y = var_1944_transpose_y_0, x = transpose_26, y = transpose_27)[name = tensor("op_1944_cast_fp16")]; + tensor var_1938_to_fp16 = const()[name = tensor("op_1938_to_fp16"), val = tensor([0x1p-3])]; + tensor attn_weights_25_cast_fp16 = mul(x = var_1944_cast_fp16, y = var_1938_to_fp16)[name = tensor("attn_weights_25_cast_fp16")]; + tensor var_1946 = logical_not(x = attn_mask_19)[name = tensor("op_1946")]; + tensor var_1947_to_fp16 = const()[name = tensor("op_1947_to_fp16"), val = tensor(-inf)]; + tensor attn_weights_27_cast_fp16 = select(a = var_1947_to_fp16, b = attn_weights_25_cast_fp16, cond = var_1946)[name = tensor("attn_weights_27_cast_fp16")]; + tensor var_1949 = const()[name = tensor("op_1949"), val = tensor(-1)]; + tensor attn_weights_29_cast_fp16 = softmax(axis = var_1949, x = attn_weights_27_cast_fp16)[name = tensor("attn_weights_29_cast_fp16")]; + tensor attn_output_9_transpose_x_0 = const()[name = tensor("attn_output_9_transpose_x_0"), val = tensor(false)]; + tensor attn_output_9_transpose_y_0 = const()[name = tensor("attn_output_9_transpose_y_0"), val = tensor(false)]; + tensor values_29_cast_fp16 = transpose(perm = var_1889, x = values_27_cast_fp16)[name = tensor("transpose_37")]; + tensor attn_output_9_cast_fp16 = matmul(transpose_x = attn_output_9_transpose_x_0, transpose_y = attn_output_9_transpose_y_0, x = attn_weights_29_cast_fp16, y = values_29_cast_fp16)[name = tensor("attn_output_9_cast_fp16")]; + tensor var_1957 = const()[name = tensor("op_1957"), val = tensor([0, 2, 1, 3])]; + tensor var_1962 = const()[name = tensor("op_1962"), val = tensor([1, 1, 1024])]; + tensor var_1958_cast_fp16 = transpose(perm = var_1957, x = attn_output_9_cast_fp16)[name = tensor("transpose_34")]; + tensor input_55_cast_fp16 = reshape(shape = var_1962, x = var_1958_cast_fp16)[name = tensor("input_55_cast_fp16")]; + tensor attn4_out_proj_weight_to_fp16 = const()[name = tensor("attn4_out_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(108126464)))]; + tensor linear_18_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = attn4_out_proj_weight_to_fp16, x = input_55_cast_fp16)[name = tensor("linear_18_cast_fp16")]; + tensor var_1968_to_fp16 = const()[name = tensor("op_1968_to_fp16"), val = tensor(0x1p+0)]; + tensor var_1969_cast_fp16 = add(x = position4_to_fp16, y = var_1968_to_fp16)[name = tensor("op_1969_cast_fp16")]; + tensor var_1969_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_1969_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor input_57_cast_fp16 = add(x = input_51_cast_fp16, y = linear_18_cast_fp16)[name = tensor("input_57_cast_fp16")]; + tensor input_59_axes_0 = const()[name = tensor("input_59_axes_0"), val = tensor([-1])]; + tensor norm4_2_weight_to_fp16 = const()[name = tensor("norm4_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(110223680)))]; + tensor norm4_2_bias_to_fp16 = const()[name = tensor("norm4_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(110225792)))]; + tensor var_1973_to_fp16 = const()[name = tensor("op_1973_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_59_cast_fp16 = layer_norm(axes = input_59_axes_0, beta = norm4_2_bias_to_fp16, epsilon = var_1973_to_fp16, gamma = norm4_2_weight_to_fp16, x = input_57_cast_fp16)[name = tensor("input_59_cast_fp16")]; + tensor linear4_1_weight_to_fp16 = const()[name = tensor("linear4_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(110227904)))]; + tensor linear_19_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = linear4_1_weight_to_fp16, x = input_59_cast_fp16)[name = tensor("linear_19_cast_fp16")]; + tensor input_61_mode_0 = const()[name = tensor("input_61_mode_0"), val = tensor("EXACT")]; + tensor input_61_cast_fp16 = gelu(mode = input_61_mode_0, x = linear_19_cast_fp16)[name = tensor("input_61_cast_fp16")]; + tensor linear4_2_weight_to_fp16 = const()[name = tensor("linear4_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(118616576)))]; + tensor linear_20_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = linear4_2_weight_to_fp16, x = input_61_cast_fp16)[name = tensor("linear_20_cast_fp16")]; + tensor input_63_cast_fp16 = add(x = input_57_cast_fp16, y = linear_20_cast_fp16)[name = tensor("input_63_cast_fp16")]; + tensor input_65_axes_0 = const()[name = tensor("input_65_axes_0"), val = tensor([-1])]; + tensor norm5_1_weight_to_fp16 = const()[name = tensor("norm5_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(127005248)))]; + tensor norm5_1_bias_to_fp16 = const()[name = tensor("norm5_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(127007360)))]; + tensor var_1990_to_fp16 = const()[name = tensor("op_1990_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_65_cast_fp16 = layer_norm(axes = input_65_axes_0, beta = norm5_1_bias_to_fp16, epsilon = var_1990_to_fp16, gamma = norm5_1_weight_to_fp16, x = input_63_cast_fp16)[name = tensor("input_65_cast_fp16")]; + tensor attn5_in_proj_weight_to_fp16 = const()[name = tensor("attn5_in_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(127009472)))]; + tensor linear_21_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = attn5_in_proj_weight_to_fp16, x = input_65_cast_fp16)[name = tensor("linear_21_cast_fp16")]; + tensor var_2004 = const()[name = tensor("op_2004"), val = tensor([1, 1, 3, 16, 64])]; + tensor qkv_cast_fp16 = reshape(shape = var_2004, x = linear_21_cast_fp16)[name = tensor("qkv_cast_fp16")]; + tensor q_31_begin_0 = const()[name = tensor("q_31_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor q_31_end_0 = const()[name = tensor("q_31_end_0"), val = tensor([1, 1, 1, 16, 64])]; + tensor q_31_end_mask_0 = const()[name = tensor("q_31_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor q_31_squeeze_mask_0 = const()[name = tensor("q_31_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor q_31_cast_fp16 = slice_by_index(begin = q_31_begin_0, end = q_31_end_0, end_mask = q_31_end_mask_0, squeeze_mask = q_31_squeeze_mask_0, x = qkv_cast_fp16)[name = tensor("q_31_cast_fp16")]; + tensor k_21_begin_0 = const()[name = tensor("k_21_begin_0"), val = tensor([0, 0, 1, 0, 0])]; + tensor k_21_end_0 = const()[name = tensor("k_21_end_0"), val = tensor([1, 1, 2, 16, 64])]; + tensor k_21_end_mask_0 = const()[name = tensor("k_21_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor k_21_squeeze_mask_0 = const()[name = tensor("k_21_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor k_21_cast_fp16 = slice_by_index(begin = k_21_begin_0, end = k_21_end_0, end_mask = k_21_end_mask_0, squeeze_mask = k_21_squeeze_mask_0, x = qkv_cast_fp16)[name = tensor("k_21_cast_fp16")]; + tensor v_begin_0 = const()[name = tensor("v_begin_0"), val = tensor([0, 0, 2, 0, 0])]; + tensor v_end_0 = const()[name = tensor("v_end_0"), val = tensor([1, 1, 3, 16, 64])]; + tensor v_end_mask_0 = const()[name = tensor("v_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor v_squeeze_mask_0 = const()[name = tensor("v_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor v_cast_fp16 = slice_by_index(begin = v_begin_0, end = v_end_0, end_mask = v_end_mask_0, squeeze_mask = v_squeeze_mask_0, x = qkv_cast_fp16)[name = tensor("v_cast_fp16")]; + tensor var_2070 = const()[name = tensor("op_2070"), val = tensor([1, 1, 1, 1])]; + tensor position5_to_fp16_dtype_0 = const()[name = tensor("position5_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor position5_to_fp16 = cast(dtype = position5_to_fp16_dtype_0, x = position5)[name = tensor("cast_114")]; + tensor ts_cast_fp16 = reshape(shape = var_2070, x = position5_to_fp16)[name = tensor("ts_cast_fp16")]; + tensor var_2077 = const()[name = tensor("op_2077"), val = tensor([1, 1, 16, 32, 2])]; + tensor q_complex_cast_fp16 = reshape(shape = var_2077, x = q_31_cast_fp16)[name = tensor("q_complex_cast_fp16")]; + tensor var_2084 = const()[name = tensor("op_2084"), val = tensor([1, 1, 16, 32, 2])]; + tensor k_complex_cast_fp16 = reshape(shape = var_2084, x = k_21_cast_fp16)[name = tensor("k_complex_cast_fp16")]; + tensor var_2088_begin_0 = const()[name = tensor("op_2088_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_2088_end_0 = const()[name = tensor("op_2088_end_0"), val = tensor([1, 1, 16, 32, 1])]; + tensor var_2088_end_mask_0 = const()[name = tensor("op_2088_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_2088_squeeze_mask_0 = const()[name = tensor("op_2088_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_2088_cast_fp16 = slice_by_index(begin = var_2088_begin_0, end = var_2088_end_0, end_mask = var_2088_end_mask_0, squeeze_mask = var_2088_squeeze_mask_0, x = q_complex_cast_fp16)[name = tensor("op_2088_cast_fp16")]; + tensor var_2096_begin_0 = const()[name = tensor("op_2096_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_2096_end_0 = const()[name = tensor("op_2096_end_0"), val = tensor([1, 1, 16, 32, 2])]; + tensor var_2096_end_mask_0 = const()[name = tensor("op_2096_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_2096_squeeze_mask_0 = const()[name = tensor("op_2096_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_2096_cast_fp16 = slice_by_index(begin = var_2096_begin_0, end = var_2096_end_0, end_mask = var_2096_end_mask_0, squeeze_mask = var_2096_squeeze_mask_0, x = q_complex_cast_fp16)[name = tensor("op_2096_cast_fp16")]; + tensor var_2104_begin_0 = const()[name = tensor("op_2104_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_2104_end_0 = const()[name = tensor("op_2104_end_0"), val = tensor([1, 1, 16, 32, 1])]; + tensor var_2104_end_mask_0 = const()[name = tensor("op_2104_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_2104_squeeze_mask_0 = const()[name = tensor("op_2104_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_2104_cast_fp16 = slice_by_index(begin = var_2104_begin_0, end = var_2104_end_0, end_mask = var_2104_end_mask_0, squeeze_mask = var_2104_squeeze_mask_0, x = k_complex_cast_fp16)[name = tensor("op_2104_cast_fp16")]; + tensor var_2112_begin_0 = const()[name = tensor("op_2112_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_2112_end_0 = const()[name = tensor("op_2112_end_0"), val = tensor([1, 1, 16, 32, 2])]; + tensor var_2112_end_mask_0 = const()[name = tensor("op_2112_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_2112_squeeze_mask_0 = const()[name = tensor("op_2112_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_2112_cast_fp16 = slice_by_index(begin = var_2112_begin_0, end = var_2112_end_0, end_mask = var_2112_end_mask_0, squeeze_mask = var_2112_squeeze_mask_0, x = k_complex_cast_fp16)[name = tensor("op_2112_cast_fp16")]; + tensor freqs_to_fp16 = const()[name = tensor("freqs_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(133300992)))]; + tensor var_2118_cast_fp16 = mul(x = freqs_to_fp16, y = ts_cast_fp16)[name = tensor("op_2118_cast_fp16")]; + tensor rotr_cast_fp16 = cos(x = var_2118_cast_fp16)[name = tensor("rotr_cast_fp16")]; + tensor roti_cast_fp16 = sin(x = var_2118_cast_fp16)[name = tensor("roti_cast_fp16")]; + tensor var_2122_cast_fp16 = mul(x = var_2088_cast_fp16, y = rotr_cast_fp16)[name = tensor("op_2122_cast_fp16")]; + tensor var_2123_cast_fp16 = mul(x = var_2096_cast_fp16, y = roti_cast_fp16)[name = tensor("op_2123_cast_fp16")]; + tensor qor_21_cast_fp16 = sub(x = var_2122_cast_fp16, y = var_2123_cast_fp16)[name = tensor("qor_21_cast_fp16")]; + tensor var_2126_cast_fp16 = mul(x = var_2088_cast_fp16, y = roti_cast_fp16)[name = tensor("op_2126_cast_fp16")]; + tensor var_2127_cast_fp16 = mul(x = var_2096_cast_fp16, y = rotr_cast_fp16)[name = tensor("op_2127_cast_fp16")]; + tensor qoi_21_cast_fp16 = add(x = var_2126_cast_fp16, y = var_2127_cast_fp16)[name = tensor("qoi_21_cast_fp16")]; + tensor var_2130_cast_fp16 = mul(x = var_2104_cast_fp16, y = rotr_cast_fp16)[name = tensor("op_2130_cast_fp16")]; + tensor var_2131_cast_fp16 = mul(x = var_2112_cast_fp16, y = roti_cast_fp16)[name = tensor("op_2131_cast_fp16")]; + tensor kor_21_cast_fp16 = sub(x = var_2130_cast_fp16, y = var_2131_cast_fp16)[name = tensor("kor_21_cast_fp16")]; + tensor var_2134_cast_fp16 = mul(x = var_2104_cast_fp16, y = roti_cast_fp16)[name = tensor("op_2134_cast_fp16")]; + tensor var_2135_cast_fp16 = mul(x = var_2112_cast_fp16, y = rotr_cast_fp16)[name = tensor("op_2135_cast_fp16")]; + tensor koi_21_cast_fp16 = add(x = var_2134_cast_fp16, y = var_2135_cast_fp16)[name = tensor("koi_21_cast_fp16")]; + tensor qo_axis_0 = const()[name = tensor("qo_axis_0"), val = tensor(-1)]; + tensor qo_cast_fp16 = stack(axis = qo_axis_0, values = (qor_21_cast_fp16, qoi_21_cast_fp16))[name = tensor("qo_cast_fp16")]; + tensor ko_axis_0 = const()[name = tensor("ko_axis_0"), val = tensor(-1)]; + tensor ko_cast_fp16 = stack(axis = ko_axis_0, values = (kor_21_cast_fp16, koi_21_cast_fp16))[name = tensor("ko_cast_fp16")]; + tensor var_2168 = const()[name = tensor("op_2168"), val = tensor([1, 1, 16, 64])]; + tensor q_33_cast_fp16 = reshape(shape = var_2168, x = qo_cast_fp16)[name = tensor("q_33_cast_fp16")]; + tensor var_2174 = const()[name = tensor("op_2174"), val = tensor([1, 1, 16, 64])]; + tensor k_cast_fp16 = reshape(shape = var_2174, x = ko_cast_fp16)[name = tensor("k_cast_fp16")]; + tensor _inversed_2199_y_0_to_fp16 = const()[name = tensor("_inversed_2199_y_0_to_fp16"), val = tensor(0x1p-9)]; + tensor _inversed_2199_cast_fp16 = mul(x = ts_cast_fp16, y = _inversed_2199_y_0_to_fp16)[name = tensor("_inversed_2199_cast_fp16")]; + tensor var_2200_cast_fp16 = floor(x = _inversed_2199_cast_fp16)[name = tensor("op_2200_cast_fp16")]; + tensor var_2201_to_fp16 = const()[name = tensor("op_2201_to_fp16"), val = tensor(0x1p+9)]; + tensor var_2202_cast_fp16 = mul(x = var_2200_cast_fp16, y = var_2201_to_fp16)[name = tensor("op_2202_cast_fp16")]; + tensor write_indices_float_cast_fp16 = sub(x = ts_cast_fp16, y = var_2202_cast_fp16)[name = tensor("write_indices_float_cast_fp16")]; + tensor var_2209_dtype_0 = const()[name = tensor("op_2209_dtype_0"), val = tensor("int32")]; + tensor write_indices_reps_0 = const()[name = tensor("write_indices_reps_0"), val = tensor([1, 1, 16, 64])]; + tensor write_indices_float_cast_fp16_to_int32 = cast(dtype = var_2209_dtype_0, x = write_indices_float_cast_fp16)[name = tensor("cast_113")]; + tensor write_indices = tile(reps = write_indices_reps_0, x = write_indices_float_cast_fp16_to_int32)[name = tensor("write_indices")]; + tensor var_2219_begin_0 = const()[name = tensor("op_2219_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_2219_end_0 = const()[name = tensor("op_2219_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor var_2219_end_mask_0 = const()[name = tensor("op_2219_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_2219_squeeze_mask_0 = const()[name = tensor("op_2219_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor cache5_to_fp16_dtype_0 = const()[name = tensor("cache5_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor cache5_to_fp16 = cast(dtype = cache5_to_fp16_dtype_0, x = cache5)[name = tensor("cast_112")]; + tensor var_2219_cast_fp16 = slice_by_index(begin = var_2219_begin_0, end = var_2219_end_0, end_mask = var_2219_end_mask_0, squeeze_mask = var_2219_squeeze_mask_0, x = cache5_to_fp16)[name = tensor("op_2219_cast_fp16")]; + tensor var_2221_axis_0 = const()[name = tensor("op_2221_axis_0"), val = tensor(1)]; + tensor var_2221_mode_0 = const()[name = tensor("op_2221_mode_0"), val = tensor("update")]; + tensor var_2221_validate_indices_0 = const()[name = tensor("op_2221_validate_indices_0"), val = tensor(false)]; + tensor var_2221_cast_fp16 = scatter_along_axis(axis = var_2221_axis_0, data = var_2219_cast_fp16, indices = write_indices, mode = var_2221_mode_0, updates = k_cast_fp16, validate_indices = var_2221_validate_indices_0)[name = tensor("op_2221_cast_fp16")]; + tensor concat_37 = const()[name = tensor("concat_37"), val = tensor([0, 0, 0, 0, 0])]; + tensor concat_38 = const()[name = tensor("concat_38"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_internal_tensor_assign_1_stride_0 = const()[name = tensor("new_cache_internal_tensor_assign_1_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_internal_tensor_assign_1_begin_mask_0 = const()[name = tensor("new_cache_internal_tensor_assign_1_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_internal_tensor_assign_1_end_mask_0 = const()[name = tensor("new_cache_internal_tensor_assign_1_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_internal_tensor_assign_1_squeeze_mask_0 = const()[name = tensor("new_cache_internal_tensor_assign_1_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_22 = const()[name = tensor("shape_22"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_10 = const()[name = tensor("reduce_prod_10"), val = tensor(1048576)]; + tensor range_1d_10_start_0 = const()[name = tensor("range_1d_10_start_0"), val = tensor(0)]; + tensor range_1d_10_step_0 = const()[name = tensor("range_1d_10_step_0"), val = tensor(1)]; + tensor range_1d_10 = range_1d(end = reduce_prod_10, start = range_1d_10_start_0, step = range_1d_10_step_0)[name = tensor("range_1d_10")]; + tensor reshape_50 = reshape(shape = shape_22, x = range_1d_10)[name = tensor("reshape_50")]; + tensor slice_by_index_10 = slice_by_index(begin = concat_37, begin_mask = new_cache_internal_tensor_assign_1_begin_mask_0, end = concat_38, end_mask = new_cache_internal_tensor_assign_1_end_mask_0, squeeze_mask = new_cache_internal_tensor_assign_1_squeeze_mask_0, stride = new_cache_internal_tensor_assign_1_stride_0, x = reshape_50)[name = tensor("slice_by_index_10")]; + tensor reshape_51_shape_0 = const()[name = tensor("reshape_51_shape_0"), val = tensor([-1])]; + tensor reshape_51 = reshape(shape = reshape_51_shape_0, x = slice_by_index_10)[name = tensor("reshape_51")]; + tensor reshape_52_shape_0 = const()[name = tensor("reshape_52_shape_0"), val = tensor([-1])]; + tensor reshape_52_cast_fp16 = reshape(shape = reshape_52_shape_0, x = var_2221_cast_fp16)[name = tensor("reshape_52_cast_fp16")]; + tensor reshape_53_shape_0 = const()[name = tensor("reshape_53_shape_0"), val = tensor([-1])]; + tensor reshape_53_cast_fp16 = reshape(shape = reshape_53_shape_0, x = cache5_to_fp16)[name = tensor("reshape_53_cast_fp16")]; + tensor scatter_10_mode_0 = const()[name = tensor("scatter_10_mode_0"), val = tensor("update")]; + tensor scatter_10_axis_0 = const()[name = tensor("scatter_10_axis_0"), val = tensor(0)]; + tensor scatter_10_validate_indices_0 = const()[name = tensor("scatter_10_validate_indices_0"), val = tensor(false)]; + tensor scatter_10_cast_fp16 = scatter(axis = scatter_10_axis_0, data = reshape_53_cast_fp16, indices = reshape_51, mode = scatter_10_mode_0, updates = reshape_52_cast_fp16, validate_indices = scatter_10_validate_indices_0)[name = tensor("scatter_10_cast_fp16")]; + tensor reshape_54_cast_fp16 = reshape(shape = shape_22, x = scatter_10_cast_fp16)[name = tensor("reshape_54_cast_fp16")]; + tensor var_2229_begin_0 = const()[name = tensor("op_2229_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor var_2229_end_0 = const()[name = tensor("op_2229_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor var_2229_end_mask_0 = const()[name = tensor("op_2229_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_2229_squeeze_mask_0 = const()[name = tensor("op_2229_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor var_2229_cast_fp16 = slice_by_index(begin = var_2229_begin_0, end = var_2229_end_0, end_mask = var_2229_end_mask_0, squeeze_mask = var_2229_squeeze_mask_0, x = reshape_54_cast_fp16)[name = tensor("op_2229_cast_fp16")]; + tensor var_2231_axis_0 = const()[name = tensor("op_2231_axis_0"), val = tensor(1)]; + tensor var_2231_mode_0 = const()[name = tensor("op_2231_mode_0"), val = tensor("update")]; + tensor var_2231_validate_indices_0 = const()[name = tensor("op_2231_validate_indices_0"), val = tensor(false)]; + tensor var_2231_cast_fp16 = scatter_along_axis(axis = var_2231_axis_0, data = var_2229_cast_fp16, indices = write_indices, mode = var_2231_mode_0, updates = v_cast_fp16, validate_indices = var_2231_validate_indices_0)[name = tensor("op_2231_cast_fp16")]; + tensor concat_39 = const()[name = tensor("concat_39"), val = tensor([1, 0, 0, 0, 0])]; + tensor concat_40 = const()[name = tensor("concat_40"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_internal_tensor_assign_2_stride_0 = const()[name = tensor("new_cache_internal_tensor_assign_2_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_internal_tensor_assign_2_begin_mask_0 = const()[name = tensor("new_cache_internal_tensor_assign_2_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_internal_tensor_assign_2_end_mask_0 = const()[name = tensor("new_cache_internal_tensor_assign_2_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_internal_tensor_assign_2_squeeze_mask_0 = const()[name = tensor("new_cache_internal_tensor_assign_2_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_23 = const()[name = tensor("shape_23"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_11 = const()[name = tensor("reduce_prod_11"), val = tensor(1048576)]; + tensor range_1d_11_start_0 = const()[name = tensor("range_1d_11_start_0"), val = tensor(0)]; + tensor range_1d_11_step_0 = const()[name = tensor("range_1d_11_step_0"), val = tensor(1)]; + tensor range_1d_11 = range_1d(end = reduce_prod_11, start = range_1d_11_start_0, step = range_1d_11_step_0)[name = tensor("range_1d_11")]; + tensor reshape_55 = reshape(shape = shape_23, x = range_1d_11)[name = tensor("reshape_55")]; + tensor slice_by_index_11 = slice_by_index(begin = concat_39, begin_mask = new_cache_internal_tensor_assign_2_begin_mask_0, end = concat_40, end_mask = new_cache_internal_tensor_assign_2_end_mask_0, squeeze_mask = new_cache_internal_tensor_assign_2_squeeze_mask_0, stride = new_cache_internal_tensor_assign_2_stride_0, x = reshape_55)[name = tensor("slice_by_index_11")]; + tensor reshape_56_shape_0 = const()[name = tensor("reshape_56_shape_0"), val = tensor([-1])]; + tensor reshape_56 = reshape(shape = reshape_56_shape_0, x = slice_by_index_11)[name = tensor("reshape_56")]; + tensor reshape_57_shape_0 = const()[name = tensor("reshape_57_shape_0"), val = tensor([-1])]; + tensor reshape_57_cast_fp16 = reshape(shape = reshape_57_shape_0, x = var_2231_cast_fp16)[name = tensor("reshape_57_cast_fp16")]; + tensor reshape_58_shape_0 = const()[name = tensor("reshape_58_shape_0"), val = tensor([-1])]; + tensor reshape_58_cast_fp16 = reshape(shape = reshape_58_shape_0, x = reshape_54_cast_fp16)[name = tensor("reshape_58_cast_fp16")]; + tensor scatter_11_mode_0 = const()[name = tensor("scatter_11_mode_0"), val = tensor("update")]; + tensor scatter_11_axis_0 = const()[name = tensor("scatter_11_axis_0"), val = tensor(0)]; + tensor scatter_11_validate_indices_0 = const()[name = tensor("scatter_11_validate_indices_0"), val = tensor(false)]; + tensor scatter_11_cast_fp16 = scatter(axis = scatter_11_axis_0, data = reshape_58_cast_fp16, indices = reshape_56, mode = scatter_11_mode_0, updates = reshape_57_cast_fp16, validate_indices = scatter_11_validate_indices_0)[name = tensor("scatter_11_cast_fp16")]; + tensor reshape_59_cast_fp16 = reshape(shape = shape_23, x = scatter_11_cast_fp16)[name = tensor("reshape_59_cast_fp16")]; + tensor reshape_59_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("reshape_59_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor keys_31_begin_0 = const()[name = tensor("keys_31_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor keys_31_end_0 = const()[name = tensor("keys_31_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor keys_31_end_mask_0 = const()[name = tensor("keys_31_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor keys_31_squeeze_mask_0 = const()[name = tensor("keys_31_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor keys_31_cast_fp16 = slice_by_index(begin = keys_31_begin_0, end = keys_31_end_0, end_mask = keys_31_end_mask_0, squeeze_mask = keys_31_squeeze_mask_0, x = reshape_59_cast_fp16)[name = tensor("keys_31_cast_fp16")]; + tensor values_31_begin_0 = const()[name = tensor("values_31_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor values_31_end_0 = const()[name = tensor("values_31_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor values_31_end_mask_0 = const()[name = tensor("values_31_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor values_31_squeeze_mask_0 = const()[name = tensor("values_31_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor values_31_cast_fp16 = slice_by_index(begin = values_31_begin_0, end = values_31_end_0, end_mask = values_31_end_mask_0, squeeze_mask = values_31_squeeze_mask_0, x = reshape_59_cast_fp16)[name = tensor("values_31_cast_fp16")]; + tensor var_2243_cast_fp16 = not_equal(x = keys_31_cast_fp16, y = keys_31_cast_fp16)[name = tensor("op_2243_cast_fp16")]; + tensor keys_33_cast_fp16 = select(a = var_319_to_fp16, b = keys_31_cast_fp16, cond = var_2243_cast_fp16)[name = tensor("keys_33_cast_fp16")]; + tensor var_2251_cast_fp16 = not_equal(x = values_31_cast_fp16, y = values_31_cast_fp16)[name = tensor("op_2251_cast_fp16")]; + tensor values_33_cast_fp16 = select(a = var_319_to_fp16, b = values_31_cast_fp16, cond = var_2251_cast_fp16)[name = tensor("values_33_cast_fp16")]; + tensor var_2275 = const()[name = tensor("op_2275"), val = tensor([0, 2, 1, 3])]; + tensor var_2291 = const()[name = tensor("op_2291"), val = tensor([1, 1, 1])]; + tensor var_2292_cast_fp16 = reshape(shape = var_2291, x = position5_to_fp16)[name = tensor("op_2292_cast_fp16")]; + tensor var_2312_to_fp16 = const()[name = tensor("op_2312_to_fp16"), val = tensor(0x1p+0)]; + tensor valid_len_cast_fp16 = add(x = var_2292_cast_fp16, y = var_2312_to_fp16)[name = tensor("valid_len_cast_fp16")]; + tensor valid_mask_cast_fp16 = less(x = k_positions_1_promoted_to_fp16, y = valid_len_cast_fp16)[name = tensor("valid_mask_cast_fp16")]; + tensor causal_mask_cast_fp16 = less_equal(x = k_positions_1_promoted_to_fp16, y = var_2292_cast_fp16)[name = tensor("causal_mask_cast_fp16")]; + tensor attn_mask_21 = logical_and(x = valid_mask_cast_fp16, y = causal_mask_cast_fp16)[name = tensor("attn_mask_21")]; + tensor attn_mask_axes_0 = const()[name = tensor("attn_mask_axes_0"), val = tensor([1])]; + tensor attn_mask = expand_dims(axes = attn_mask_axes_0, x = attn_mask_21)[name = tensor("attn_mask")]; + tensor var_2330_transpose_x_0 = const()[name = tensor("op_2330_transpose_x_0"), val = tensor(false)]; + tensor var_2330_transpose_y_0 = const()[name = tensor("op_2330_transpose_y_0"), val = tensor(false)]; + tensor transpose_28_perm_0 = const()[name = tensor("transpose_28_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_29_perm_0 = const()[name = tensor("transpose_29_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_29 = transpose(perm = transpose_29_perm_0, x = keys_33_cast_fp16)[name = tensor("transpose_31")]; + tensor transpose_28 = transpose(perm = transpose_28_perm_0, x = q_33_cast_fp16)[name = tensor("transpose_32")]; + tensor var_2330_cast_fp16 = matmul(transpose_x = var_2330_transpose_x_0, transpose_y = var_2330_transpose_y_0, x = transpose_28, y = transpose_29)[name = tensor("op_2330_cast_fp16")]; + tensor var_2324_to_fp16 = const()[name = tensor("op_2324_to_fp16"), val = tensor([0x1p-3])]; + tensor attn_weights_31_cast_fp16 = mul(x = var_2330_cast_fp16, y = var_2324_to_fp16)[name = tensor("attn_weights_31_cast_fp16")]; + tensor var_2332 = logical_not(x = attn_mask)[name = tensor("op_2332")]; + tensor var_2333_to_fp16 = const()[name = tensor("op_2333_to_fp16"), val = tensor(-inf)]; + tensor attn_weights_33_cast_fp16 = select(a = var_2333_to_fp16, b = attn_weights_31_cast_fp16, cond = var_2332)[name = tensor("attn_weights_33_cast_fp16")]; + tensor var_2335 = const()[name = tensor("op_2335"), val = tensor(-1)]; + tensor attn_weights_cast_fp16 = softmax(axis = var_2335, x = attn_weights_33_cast_fp16)[name = tensor("attn_weights_cast_fp16")]; + tensor attn_output_transpose_x_0 = const()[name = tensor("attn_output_transpose_x_0"), val = tensor(false)]; + tensor attn_output_transpose_y_0 = const()[name = tensor("attn_output_transpose_y_0"), val = tensor(false)]; + tensor values_cast_fp16 = transpose(perm = var_2275, x = values_33_cast_fp16)[name = tensor("transpose_33")]; + tensor attn_output_cast_fp16 = matmul(transpose_x = attn_output_transpose_x_0, transpose_y = attn_output_transpose_y_0, x = attn_weights_cast_fp16, y = values_cast_fp16)[name = tensor("attn_output_cast_fp16")]; + tensor var_2343 = const()[name = tensor("op_2343"), val = tensor([0, 2, 1, 3])]; + tensor var_2348 = const()[name = tensor("op_2348"), val = tensor([1, 1, 1024])]; + tensor var_2344_cast_fp16 = transpose(perm = var_2343, x = attn_output_cast_fp16)[name = tensor("transpose_30")]; + tensor input_67_cast_fp16 = reshape(shape = var_2348, x = var_2344_cast_fp16)[name = tensor("input_67_cast_fp16")]; + tensor attn5_out_proj_weight_to_fp16 = const()[name = tensor("attn5_out_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(133301120)))]; + tensor linear_22_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = attn5_out_proj_weight_to_fp16, x = input_67_cast_fp16)[name = tensor("linear_22_cast_fp16")]; + tensor var_2354_to_fp16 = const()[name = tensor("op_2354_to_fp16"), val = tensor(0x1p+0)]; + tensor var_2355_cast_fp16 = add(x = position5_to_fp16, y = var_2354_to_fp16)[name = tensor("op_2355_cast_fp16")]; + tensor var_2355_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_2355_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor input_69_cast_fp16 = add(x = input_63_cast_fp16, y = linear_22_cast_fp16)[name = tensor("input_69_cast_fp16")]; + tensor input_71_axes_0 = const()[name = tensor("input_71_axes_0"), val = tensor([-1])]; + tensor norm5_2_weight_to_fp16 = const()[name = tensor("norm5_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(135398336)))]; + tensor norm5_2_bias_to_fp16 = const()[name = tensor("norm5_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(135400448)))]; + tensor var_2359_to_fp16 = const()[name = tensor("op_2359_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_71_cast_fp16 = layer_norm(axes = input_71_axes_0, beta = norm5_2_bias_to_fp16, epsilon = var_2359_to_fp16, gamma = norm5_2_weight_to_fp16, x = input_69_cast_fp16)[name = tensor("input_71_cast_fp16")]; + tensor linear5_1_weight_to_fp16 = const()[name = tensor("linear5_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(135402560)))]; + tensor linear_23_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = linear5_1_weight_to_fp16, x = input_71_cast_fp16)[name = tensor("linear_23_cast_fp16")]; + tensor input_73_mode_0 = const()[name = tensor("input_73_mode_0"), val = tensor("EXACT")]; + tensor input_73_cast_fp16 = gelu(mode = input_73_mode_0, x = linear_23_cast_fp16)[name = tensor("input_73_cast_fp16")]; + tensor linear5_2_weight_to_fp16 = const()[name = tensor("linear5_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(143791232)))]; + tensor linear_24_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = linear5_2_weight_to_fp16, x = input_73_cast_fp16)[name = tensor("linear_24_cast_fp16")]; + tensor input_75_cast_fp16 = add(x = input_69_cast_fp16, y = linear_24_cast_fp16)[name = tensor("input_75_cast_fp16")]; + tensor input_axes_0 = const()[name = tensor("input_axes_0"), val = tensor([-1])]; + tensor out_norm_weight_to_fp16 = const()[name = tensor("out_norm_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(152179904)))]; + tensor out_norm_bias_to_fp16 = const()[name = tensor("out_norm_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(152182016)))]; + tensor var_2376_to_fp16 = const()[name = tensor("op_2376_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_cast_fp16 = layer_norm(axes = input_axes_0, beta = out_norm_bias_to_fp16, epsilon = var_2376_to_fp16, gamma = out_norm_weight_to_fp16, x = input_75_cast_fp16)[name = tensor("input_cast_fp16")]; + tensor input_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("input_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor out_eos_weight_to_fp16 = const()[name = tensor("out_eos_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(152184128)))]; + tensor out_eos_bias_to_fp16 = const()[name = tensor("out_eos_bias_to_fp16"), val = tensor([-0x1.64p-3])]; + tensor linear_25_cast_fp16 = linear(bias = out_eos_bias_to_fp16, weight = out_eos_weight_to_fp16, x = input_cast_fp16)[name = tensor("linear_25_cast_fp16")]; + tensor linear_25_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("linear_25_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor var_2384 = cast(dtype = linear_25_cast_fp16_to_fp32_dtype_0, x = linear_25_cast_fp16)[name = tensor("cast_108")]; + tensor input = cast(dtype = input_cast_fp16_to_fp32_dtype_0, x = input_cast_fp16)[name = tensor("cast_109")]; + tensor var_2355 = cast(dtype = var_2355_cast_fp16_to_fp32_dtype_0, x = var_2355_cast_fp16)[name = tensor("cast_110")]; + tensor new_cache_internal_tensor_assign_2 = cast(dtype = reshape_59_cast_fp16_to_fp32_dtype_0, x = reshape_59_cast_fp16)[name = tensor("cast_111")]; + tensor var_1969 = cast(dtype = var_1969_cast_fp16_to_fp32_dtype_0, x = var_1969_cast_fp16)[name = tensor("cast_115")]; + tensor new_cache_9_internal_tensor_assign_2 = cast(dtype = reshape_49_cast_fp16_to_fp32_dtype_0, x = reshape_49_cast_fp16)[name = tensor("cast_116")]; + tensor var_1583 = cast(dtype = var_1583_cast_fp16_to_fp32_dtype_0, x = var_1583_cast_fp16)[name = tensor("cast_120")]; + tensor new_cache_7_internal_tensor_assign_2 = cast(dtype = reshape_39_cast_fp16_to_fp32_dtype_0, x = reshape_39_cast_fp16)[name = tensor("cast_121")]; + tensor var_1197 = cast(dtype = var_1197_cast_fp16_to_fp32_dtype_0, x = var_1197_cast_fp16)[name = tensor("cast_125")]; + tensor new_cache_5_internal_tensor_assign_2 = cast(dtype = reshape_29_cast_fp16_to_fp32_dtype_0, x = reshape_29_cast_fp16)[name = tensor("cast_126")]; + tensor var_811 = cast(dtype = var_811_cast_fp16_to_fp32_dtype_0, x = var_811_cast_fp16)[name = tensor("cast_130")]; + tensor new_cache_3_internal_tensor_assign_2 = cast(dtype = reshape_19_cast_fp16_to_fp32_dtype_0, x = reshape_19_cast_fp16)[name = tensor("cast_131")]; + tensor var_425 = cast(dtype = var_425_cast_fp16_to_fp32_dtype_0, x = var_425_cast_fp16)[name = tensor("cast_135")]; + tensor new_cache_1_internal_tensor_assign_2 = cast(dtype = reshape_9_cast_fp16_to_fp32_dtype_0, x = reshape_9_cast_fp16)[name = tensor("cast_136")]; + } -> (input, var_2384, new_cache_1_internal_tensor_assign_2, var_425, new_cache_3_internal_tensor_assign_2, var_811, new_cache_5_internal_tensor_assign_2, var_1197, new_cache_7_internal_tensor_assign_2, var_1583, new_cache_9_internal_tensor_assign_2, var_1969, new_cache_internal_tensor_assign_2, var_2355); +} \ No newline at end of file diff --git a/v2.1/italian/flowlm_step.mlmodelc/weights/weight.bin b/v2.1/italian/flowlm_step.mlmodelc/weights/weight.bin new file mode 100644 index 0000000000000000000000000000000000000000..f6de0e04cae06693252bce8e1a4e1d718ed70df4 --- /dev/null +++ b/v2.1/italian/flowlm_step.mlmodelc/weights/weight.bin @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:6e16becc616f5082bd09936ece09e835e66f20f4385002344089fa17d426ee77 +size 152186240 diff --git a/v2.1/italian/flowlm_step.mlpackage/Data/com.apple.CoreML/model.mlmodel b/v2.1/italian/flowlm_step.mlpackage/Data/com.apple.CoreML/model.mlmodel new file mode 100644 index 0000000000000000000000000000000000000000..6657d9ba11f620de39f867aaf6f2ae69400f2f37 --- /dev/null +++ b/v2.1/italian/flowlm_step.mlpackage/Data/com.apple.CoreML/model.mlmodel @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:412a73187c2fd3baf48434ff2708a814450b00a3a4cab96ed6fb52a4001036df +size 207618 diff --git a/v2.1/italian/flowlm_step.mlpackage/Data/com.apple.CoreML/weights/weight.bin b/v2.1/italian/flowlm_step.mlpackage/Data/com.apple.CoreML/weights/weight.bin new file mode 100644 index 0000000000000000000000000000000000000000..f6de0e04cae06693252bce8e1a4e1d718ed70df4 --- /dev/null +++ b/v2.1/italian/flowlm_step.mlpackage/Data/com.apple.CoreML/weights/weight.bin @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:6e16becc616f5082bd09936ece09e835e66f20f4385002344089fa17d426ee77 +size 152186240 diff --git a/v2.1/italian/flowlm_step.mlpackage/Manifest.json b/v2.1/italian/flowlm_step.mlpackage/Manifest.json new file mode 100644 index 0000000000000000000000000000000000000000..81bb78eedd7b8f182a10f56e0b2846fa48f672de --- /dev/null +++ b/v2.1/italian/flowlm_step.mlpackage/Manifest.json @@ -0,0 +1,18 @@ +{ + "fileFormatVersion": "1.0.0", + "itemInfoEntries": { + "07D4206E-8E00-43BC-BBF4-5CCD6442499A": { + "author": "com.apple.CoreML", + "description": "CoreML Model Weights", + "name": "weights", + "path": "com.apple.CoreML/weights" + }, + "C7835080-440D-4ECB-BDC4-C5A1AC7F6747": { + "author": "com.apple.CoreML", + "description": "CoreML Model Specification", + "name": "model.mlmodel", + "path": "com.apple.CoreML/model.mlmodel" + } + }, + "rootModelIdentifier": "C7835080-440D-4ECB-BDC4-C5A1AC7F6747" +} diff --git a/v2.1/italian/flowlm_stepv2.mlmodelc/analytics/coremldata.bin b/v2.1/italian/flowlm_stepv2.mlmodelc/analytics/coremldata.bin new file mode 100644 index 0000000000000000000000000000000000000000..9c78fe17d045427e3a10c7f8031ab6094d274069 --- /dev/null +++ b/v2.1/italian/flowlm_stepv2.mlmodelc/analytics/coremldata.bin @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:2eb1b88584e4a78a8f93fb0776021bd5e28ae6c42fc37f77f149d625c19a7574 +size 243 diff --git a/v2.1/italian/flowlm_stepv2.mlmodelc/coremldata.bin b/v2.1/italian/flowlm_stepv2.mlmodelc/coremldata.bin new file mode 100644 index 0000000000000000000000000000000000000000..857890f9739029dd9ca13cc6f157ba573793d299 --- /dev/null +++ b/v2.1/italian/flowlm_stepv2.mlmodelc/coremldata.bin @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:a6e5a9e52e1c48d17457417ba298e8e2d20ddfa5ad1214904628183aa9ad0ff6 +size 1190 diff --git a/v2.1/italian/flowlm_stepv2.mlmodelc/metadata.json b/v2.1/italian/flowlm_stepv2.mlmodelc/metadata.json new file mode 100644 index 0000000000000000000000000000000000000000..f2969c83e1cf7a6025c56eb3502b7ec456a89630 --- /dev/null +++ b/v2.1/italian/flowlm_stepv2.mlmodelc/metadata.json @@ -0,0 +1,348 @@ +[ + { + "metadataOutputVersion" : "3.0", + "storagePrecision" : "Mixed (Float16, Int8)", + "outputSchema" : [ + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1 × 1 × 1024)", + "shortDescription" : "", + "shape" : "[1, 1, 1024]", + "name" : "input", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1 × 1 × 1)", + "shortDescription" : "", + "shape" : "[1, 1, 1]", + "name" : "var_2582", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 2 × 1 × 512 × 16 × 64)", + "shortDescription" : "", + "shape" : "[2, 1, 512, 16, 64]", + "name" : "new_cache_1_internal_tensor_assign_2", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1)", + "shortDescription" : "", + "shape" : "[1]", + "name" : "var_458", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 2 × 1 × 512 × 16 × 64)", + "shortDescription" : "", + "shape" : "[2, 1, 512, 16, 64]", + "name" : "new_cache_3_internal_tensor_assign_2", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1)", + "shortDescription" : "", + "shape" : "[1]", + "name" : "var_877", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 2 × 1 × 512 × 16 × 64)", + "shortDescription" : "", + "shape" : "[2, 1, 512, 16, 64]", + "name" : "new_cache_5_internal_tensor_assign_2", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1)", + "shortDescription" : "", + "shape" : "[1]", + "name" : "var_1296", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 2 × 1 × 512 × 16 × 64)", + "shortDescription" : "", + "shape" : "[2, 1, 512, 16, 64]", + "name" : "new_cache_7_internal_tensor_assign_2", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1)", + "shortDescription" : "", + "shape" : "[1]", + "name" : "var_1715", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 2 × 1 × 512 × 16 × 64)", + "shortDescription" : "", + "shape" : "[2, 1, 512, 16, 64]", + "name" : "new_cache_9_internal_tensor_assign_2", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1)", + "shortDescription" : "", + "shape" : "[1]", + "name" : "var_2134", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 2 × 1 × 512 × 16 × 64)", + "shortDescription" : "", + "shape" : "[2, 1, 512, 16, 64]", + "name" : "new_cache_internal_tensor_assign_2", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1)", + "shortDescription" : "", + "shape" : "[1]", + "name" : "var_2553", + "type" : "MultiArray" + } + ], + "modelParameters" : [ + + ], + "specificationVersion" : 8, + "mlProgramOperationTypeHistogram" : { + "Ios17.cos" : 6, + "Range1d" : 12, + "Ios17.notEqual" : 13, + "Ios17.floor" : 6, + "Ios17.reshape" : 108, + "Ios17.logicalAnd" : 6, + "Ios16.softmax" : 6, + "Ios17.scatter" : 12, + "Ios17.matmul" : 12, + "Ios17.sin" : 6, + "Ios17.transpose" : 24, + "Select" : 19, + "Ios17.expandDims" : 7, + "Ios17.add" : 36, + "Tile" : 6, + "Ios17.sliceByIndex" : 78, + "Ios17.lessEqual" : 6, + "Ios17.scatterAlongAxis" : 12, + "Ios17.logicalNot" : 6, + "Ios17.layerNorm" : 13, + "Ios16.constexprAffineDequantize" : 24, + "Ios17.cast" : 34, + "Ios17.sub" : 18, + "Ios17.less" : 6, + "Ios16.gelu" : 6, + "Ios17.linear" : 26, + "Stack" : 12, + "Ios17.mul" : 72 + }, + "computePrecision" : "Mixed (Float16, Float32, Int32)", + "isUpdatable" : "0", + "stateSchema" : [ + + ], + "availability" : { + "macOS" : "14.0", + "tvOS" : "17.0", + "visionOS" : "1.0", + "watchOS" : "10.0", + "iOS" : "17.0", + "macCatalyst" : "17.0" + }, + "modelType" : { + "name" : "MLModelType_mlProgram" + }, + "userDefinedMetadata" : { + "com.github.apple.coremltools.conversion_date" : "2026-04-27", + "com.github.apple.coremltools.source" : "torch==2.9.1", + "com.github.apple.coremltools.version" : "9.0", + "com.github.apple.coremltools.source_dialect" : "TorchScript" + }, + "inputSchema" : [ + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1 × 1 × 32)", + "shortDescription" : "", + "shape" : "[1, 1, 32]", + "name" : "sequence", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 32)", + "shortDescription" : "", + "shape" : "[32]", + "name" : "bos_emb", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 2 × 1 × 512 × 16 × 64)", + "shortDescription" : "", + "shape" : "[2, 1, 512, 16, 64]", + "name" : "cache0", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1)", + "shortDescription" : "", + "shape" : "[1]", + "name" : "position0", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 2 × 1 × 512 × 16 × 64)", + "shortDescription" : "", + "shape" : "[2, 1, 512, 16, 64]", + "name" : "cache1", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1)", + "shortDescription" : "", + "shape" : "[1]", + "name" : "position1", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 2 × 1 × 512 × 16 × 64)", + "shortDescription" : "", + "shape" : "[2, 1, 512, 16, 64]", + "name" : "cache2", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1)", + "shortDescription" : "", + "shape" : "[1]", + "name" : "position2", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 2 × 1 × 512 × 16 × 64)", + "shortDescription" : "", + "shape" : "[2, 1, 512, 16, 64]", + "name" : "cache3", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1)", + "shortDescription" : "", + "shape" : "[1]", + "name" : "position3", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 2 × 1 × 512 × 16 × 64)", + "shortDescription" : "", + "shape" : "[2, 1, 512, 16, 64]", + "name" : "cache4", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1)", + "shortDescription" : "", + "shape" : "[1]", + "name" : "position4", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 2 × 1 × 512 × 16 × 64)", + "shortDescription" : "", + "shape" : "[2, 1, 512, 16, 64]", + "name" : "cache5", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1)", + "shortDescription" : "", + "shape" : "[1]", + "name" : "position5", + "type" : "MultiArray" + } + ], + "generatedClassName" : "flowlm_stepv2", + "method" : "predict" + } +] \ No newline at end of file diff --git a/v2.1/italian/flowlm_stepv2.mlmodelc/model.mil b/v2.1/italian/flowlm_stepv2.mlmodelc/model.mil new file mode 100644 index 0000000000000000000000000000000000000000..21d3cff572e013e90d06f014ddf49f4e52f598e3 --- /dev/null +++ b/v2.1/italian/flowlm_stepv2.mlmodelc/model.mil @@ -0,0 +1,1359 @@ +program(1.0) +[buildInfo = dict, tensor>({{"coremlc-component-MIL", "3520.4.1"}, {"coremlc-version", "3520.5.1"}, {"coremltools-component-torch", "2.9.1"}, {"coremltools-source-dialect", "TorchScript"}, {"coremltools-version", "9.0"}})] +{ + func main(tensor bos_emb, tensor cache0, tensor cache1, tensor cache2, tensor cache3, tensor cache4, tensor cache5, tensor position0, tensor position1, tensor position2, tensor position3, tensor position4, tensor position5, tensor sequence) { + tensor attn0_in_proj_weight_cast_fp16 = constexpr_affine_dequantize()[axis = tensor(0), name = tensor("attn0_in_proj_weight_cast_fp16"), quantized_data = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(64))), scale = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(3148992))), zero_point = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(3145856)))]; + tensor attn0_out_proj_weight_cast_fp16 = constexpr_affine_dequantize()[axis = tensor(0), name = tensor("attn0_out_proj_weight_cast_fp16"), quantized_data = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(3155200))), scale = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(4204928))), zero_point = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(4203840)))]; + tensor linear0_1_weight_cast_fp16 = constexpr_affine_dequantize()[axis = tensor(0), name = tensor("linear0_1_weight_cast_fp16"), quantized_data = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(4207040))), scale = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8405568))), zero_point = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8401408)))]; + tensor linear0_2_weight_cast_fp16 = constexpr_affine_dequantize()[axis = tensor(0), name = tensor("linear0_2_weight_cast_fp16"), quantized_data = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8413824))), scale = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(12608192))), zero_point = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(4203840)))]; + tensor attn1_in_proj_weight_cast_fp16 = constexpr_affine_dequantize()[axis = tensor(0), name = tensor("attn1_in_proj_weight_cast_fp16"), quantized_data = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(12610304))), scale = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15756096))), zero_point = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(3145856)))]; + tensor attn1_out_proj_weight_cast_fp16 = constexpr_affine_dequantize()[axis = tensor(0), name = tensor("attn1_out_proj_weight_cast_fp16"), quantized_data = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15762304))), scale = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(16810944))), zero_point = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(4203840)))]; + tensor linear1_1_weight_cast_fp16 = constexpr_affine_dequantize()[axis = tensor(0), name = tensor("linear1_1_weight_cast_fp16"), quantized_data = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(16813056))), scale = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21007424))), zero_point = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8401408)))]; + tensor linear1_2_weight_cast_fp16 = constexpr_affine_dequantize()[axis = tensor(0), name = tensor("linear1_2_weight_cast_fp16"), quantized_data = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21015680))), scale = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(25210048))), zero_point = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(4203840)))]; + tensor attn2_in_proj_weight_cast_fp16 = constexpr_affine_dequantize()[axis = tensor(0), name = tensor("attn2_in_proj_weight_cast_fp16"), quantized_data = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(25212160))), scale = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28357952))), zero_point = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(3145856)))]; + tensor attn2_out_proj_weight_cast_fp16 = constexpr_affine_dequantize()[axis = tensor(0), name = tensor("attn2_out_proj_weight_cast_fp16"), quantized_data = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28364160))), scale = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(29412800))), zero_point = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(4203840)))]; + tensor linear2_1_weight_cast_fp16 = constexpr_affine_dequantize()[axis = tensor(0), name = tensor("linear2_1_weight_cast_fp16"), quantized_data = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(29414912))), scale = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(33609280))), zero_point = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8401408)))]; + tensor linear2_2_weight_cast_fp16 = constexpr_affine_dequantize()[axis = tensor(0), name = tensor("linear2_2_weight_cast_fp16"), quantized_data = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(33617536))), scale = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(37811904))), zero_point = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(4203840)))]; + tensor attn3_in_proj_weight_cast_fp16 = constexpr_affine_dequantize()[axis = tensor(0), name = tensor("attn3_in_proj_weight_cast_fp16"), quantized_data = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(37814016))), scale = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(40959808))), zero_point = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(3145856)))]; + tensor attn3_out_proj_weight_cast_fp16 = constexpr_affine_dequantize()[axis = tensor(0), name = tensor("attn3_out_proj_weight_cast_fp16"), quantized_data = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(40966016))), scale = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(42014656))), zero_point = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(4203840)))]; + tensor linear3_1_weight_cast_fp16 = constexpr_affine_dequantize()[axis = tensor(0), name = tensor("linear3_1_weight_cast_fp16"), quantized_data = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(42016768))), scale = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(46211136))), zero_point = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8401408)))]; + tensor linear3_2_weight_cast_fp16 = constexpr_affine_dequantize()[axis = tensor(0), name = tensor("linear3_2_weight_cast_fp16"), quantized_data = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(46219392))), scale = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(50413760))), zero_point = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(4203840)))]; + tensor attn4_in_proj_weight_cast_fp16 = constexpr_affine_dequantize()[axis = tensor(0), name = tensor("attn4_in_proj_weight_cast_fp16"), quantized_data = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(50415872))), scale = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(53561664))), zero_point = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(3145856)))]; + tensor attn4_out_proj_weight_cast_fp16 = constexpr_affine_dequantize()[axis = tensor(0), name = tensor("attn4_out_proj_weight_cast_fp16"), quantized_data = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(53567872))), scale = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(54616512))), zero_point = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(4203840)))]; + tensor linear4_1_weight_cast_fp16 = constexpr_affine_dequantize()[axis = tensor(0), name = tensor("linear4_1_weight_cast_fp16"), quantized_data = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(54618624))), scale = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(58812992))), zero_point = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8401408)))]; + tensor linear4_2_weight_cast_fp16 = constexpr_affine_dequantize()[axis = tensor(0), name = tensor("linear4_2_weight_cast_fp16"), quantized_data = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(58821248))), scale = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(63015616))), zero_point = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(4203840)))]; + tensor attn5_in_proj_weight_cast_fp16 = constexpr_affine_dequantize()[axis = tensor(0), name = tensor("attn5_in_proj_weight_cast_fp16"), quantized_data = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(63017728))), scale = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(66163520))), zero_point = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(3145856)))]; + tensor attn5_out_proj_weight_cast_fp16 = constexpr_affine_dequantize()[axis = tensor(0), name = tensor("attn5_out_proj_weight_cast_fp16"), quantized_data = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(66169728))), scale = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(67218368))), zero_point = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(4203840)))]; + tensor linear5_1_weight_cast_fp16 = constexpr_affine_dequantize()[axis = tensor(0), name = tensor("linear5_1_weight_cast_fp16"), quantized_data = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(67220480))), scale = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(71414848))), zero_point = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8401408)))]; + tensor linear5_2_weight_cast_fp16 = constexpr_affine_dequantize()[axis = tensor(0), name = tensor("linear5_2_weight_cast_fp16"), quantized_data = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(71423104))), scale = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(75617472))), zero_point = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(4203840)))]; + tensor sequence_to_fp16_dtype_0 = const()[name = tensor("sequence_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor sequence_to_fp16 = cast(dtype = sequence_to_fp16_dtype_0, x = sequence)[name = tensor("cast_141")]; + tensor var_54_cast_fp16 = not_equal(x = sequence_to_fp16, y = sequence_to_fp16)[name = tensor("op_54_cast_fp16")]; + tensor expand_dims_0_axes_0 = const()[name = tensor("expand_dims_0_axes_0"), val = tensor([0, 1])]; + tensor bos_emb_to_fp16_dtype_0 = const()[name = tensor("bos_emb_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor bos_emb_to_fp16 = cast(dtype = bos_emb_to_fp16_dtype_0, x = bos_emb)[name = tensor("cast_140")]; + tensor expand_dims_0_cast_fp16 = expand_dims(axes = expand_dims_0_axes_0, x = bos_emb_to_fp16)[name = tensor("expand_dims_0_cast_fp16")]; + tensor input_1_cast_fp16 = select(a = expand_dims_0_cast_fp16, b = sequence_to_fp16, cond = var_54_cast_fp16)[name = tensor("input_1_cast_fp16")]; + tensor input_linear_weight_to_fp16 = const()[name = tensor("input_linear_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(75619584)))]; + tensor linear_0_bias_0_to_fp16 = const()[name = tensor("linear_0_bias_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(75685184)))]; + tensor linear_0_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = input_linear_weight_to_fp16, x = input_1_cast_fp16)[name = tensor("linear_0_cast_fp16")]; + tensor x_1_axes_0 = const()[name = tensor("x_1_axes_0"), val = tensor([-1])]; + tensor norm0_1_weight_to_fp16 = const()[name = tensor("norm0_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(75687296)))]; + tensor norm0_1_bias_to_fp16 = const()[name = tensor("norm0_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(75689408)))]; + tensor var_60_to_fp16 = const()[name = tensor("op_60_to_fp16"), val = tensor(0x1.5p-17)]; + tensor x_1_cast_fp16 = layer_norm(axes = x_1_axes_0, beta = norm0_1_bias_to_fp16, epsilon = var_60_to_fp16, gamma = norm0_1_weight_to_fp16, x = linear_0_cast_fp16)[name = tensor("x_1_cast_fp16")]; + tensor linear_1_bias_0_to_fp16 = const()[name = tensor("linear_1_bias_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(75691520)))]; + tensor linear_1_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = attn0_in_proj_weight_cast_fp16, x = x_1_cast_fp16)[name = tensor("linear_1_cast_fp16")]; + tensor var_96 = const()[name = tensor("op_96"), val = tensor([1, 1, 3, 16, 64])]; + tensor qkv_1_cast_fp16 = reshape(shape = var_96, x = linear_1_cast_fp16)[name = tensor("qkv_1_cast_fp16")]; + tensor q_1_begin_0 = const()[name = tensor("q_1_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor q_1_end_0 = const()[name = tensor("q_1_end_0"), val = tensor([1, 1, 1, 16, 64])]; + tensor q_1_end_mask_0 = const()[name = tensor("q_1_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor q_1_squeeze_mask_0 = const()[name = tensor("q_1_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor q_1_cast_fp16 = slice_by_index(begin = q_1_begin_0, end = q_1_end_0, end_mask = q_1_end_mask_0, squeeze_mask = q_1_squeeze_mask_0, x = qkv_1_cast_fp16)[name = tensor("q_1_cast_fp16")]; + tensor k_1_begin_0 = const()[name = tensor("k_1_begin_0"), val = tensor([0, 0, 1, 0, 0])]; + tensor k_1_end_0 = const()[name = tensor("k_1_end_0"), val = tensor([1, 1, 2, 16, 64])]; + tensor k_1_end_mask_0 = const()[name = tensor("k_1_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor k_1_squeeze_mask_0 = const()[name = tensor("k_1_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor k_1_cast_fp16 = slice_by_index(begin = k_1_begin_0, end = k_1_end_0, end_mask = k_1_end_mask_0, squeeze_mask = k_1_squeeze_mask_0, x = qkv_1_cast_fp16)[name = tensor("k_1_cast_fp16")]; + tensor v_1_begin_0 = const()[name = tensor("v_1_begin_0"), val = tensor([0, 0, 2, 0, 0])]; + tensor v_1_end_0 = const()[name = tensor("v_1_end_0"), val = tensor([1, 1, 3, 16, 64])]; + tensor v_1_end_mask_0 = const()[name = tensor("v_1_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor v_1_squeeze_mask_0 = const()[name = tensor("v_1_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor v_1_cast_fp16 = slice_by_index(begin = v_1_begin_0, end = v_1_end_0, end_mask = v_1_end_mask_0, squeeze_mask = v_1_squeeze_mask_0, x = qkv_1_cast_fp16)[name = tensor("v_1_cast_fp16")]; + tensor var_200 = const()[name = tensor("op_200"), val = tensor([1, 1, 1, 1])]; + tensor position0_to_fp16_dtype_0 = const()[name = tensor("position0_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor position0_to_fp16 = cast(dtype = position0_to_fp16_dtype_0, x = position0)[name = tensor("cast_139")]; + tensor ts_5_cast_fp16 = reshape(shape = var_200, x = position0_to_fp16)[name = tensor("ts_5_cast_fp16")]; + tensor var_204 = const()[name = tensor("op_204"), val = tensor([1, 1, 16, 32, 2])]; + tensor q_complex_1_cast_fp16 = reshape(shape = var_204, x = q_1_cast_fp16)[name = tensor("q_complex_1_cast_fp16")]; + tensor var_208 = const()[name = tensor("op_208"), val = tensor([1, 1, 16, 32, 2])]; + tensor k_complex_1_cast_fp16 = reshape(shape = var_208, x = k_1_cast_fp16)[name = tensor("k_complex_1_cast_fp16")]; + tensor var_212_begin_0 = const()[name = tensor("op_212_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_212_end_0 = const()[name = tensor("op_212_end_0"), val = tensor([1, 1, 16, 32, 1])]; + tensor var_212_end_mask_0 = const()[name = tensor("op_212_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_212_squeeze_mask_0 = const()[name = tensor("op_212_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_212_cast_fp16 = slice_by_index(begin = var_212_begin_0, end = var_212_end_0, end_mask = var_212_end_mask_0, squeeze_mask = var_212_squeeze_mask_0, x = q_complex_1_cast_fp16)[name = tensor("op_212_cast_fp16")]; + tensor var_220_begin_0 = const()[name = tensor("op_220_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_220_end_0 = const()[name = tensor("op_220_end_0"), val = tensor([1, 1, 16, 32, 2])]; + tensor var_220_end_mask_0 = const()[name = tensor("op_220_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_220_squeeze_mask_0 = const()[name = tensor("op_220_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_220_cast_fp16 = slice_by_index(begin = var_220_begin_0, end = var_220_end_0, end_mask = var_220_end_mask_0, squeeze_mask = var_220_squeeze_mask_0, x = q_complex_1_cast_fp16)[name = tensor("op_220_cast_fp16")]; + tensor var_228_begin_0 = const()[name = tensor("op_228_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_228_end_0 = const()[name = tensor("op_228_end_0"), val = tensor([1, 1, 16, 32, 1])]; + tensor var_228_end_mask_0 = const()[name = tensor("op_228_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_228_squeeze_mask_0 = const()[name = tensor("op_228_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_228_cast_fp16 = slice_by_index(begin = var_228_begin_0, end = var_228_end_0, end_mask = var_228_end_mask_0, squeeze_mask = var_228_squeeze_mask_0, x = k_complex_1_cast_fp16)[name = tensor("op_228_cast_fp16")]; + tensor var_236_begin_0 = const()[name = tensor("op_236_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_236_end_0 = const()[name = tensor("op_236_end_0"), val = tensor([1, 1, 16, 32, 2])]; + tensor var_236_end_mask_0 = const()[name = tensor("op_236_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_236_squeeze_mask_0 = const()[name = tensor("op_236_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_236_cast_fp16 = slice_by_index(begin = var_236_begin_0, end = var_236_end_0, end_mask = var_236_end_mask_0, squeeze_mask = var_236_squeeze_mask_0, x = k_complex_1_cast_fp16)[name = tensor("op_236_cast_fp16")]; + tensor freqs_1_to_fp16 = const()[name = tensor("freqs_1_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(75697728)))]; + tensor var_242_cast_fp16 = mul(x = freqs_1_to_fp16, y = ts_5_cast_fp16)[name = tensor("op_242_cast_fp16")]; + tensor rotr_1_cast_fp16 = cos(x = var_242_cast_fp16)[name = tensor("rotr_1_cast_fp16")]; + tensor roti_1_cast_fp16 = sin(x = var_242_cast_fp16)[name = tensor("roti_1_cast_fp16")]; + tensor var_246_cast_fp16 = mul(x = var_212_cast_fp16, y = rotr_1_cast_fp16)[name = tensor("op_246_cast_fp16")]; + tensor var_247_cast_fp16 = mul(x = var_220_cast_fp16, y = roti_1_cast_fp16)[name = tensor("op_247_cast_fp16")]; + tensor qor_1_cast_fp16 = sub(x = var_246_cast_fp16, y = var_247_cast_fp16)[name = tensor("qor_1_cast_fp16")]; + tensor var_250_cast_fp16 = mul(x = var_212_cast_fp16, y = roti_1_cast_fp16)[name = tensor("op_250_cast_fp16")]; + tensor var_251_cast_fp16 = mul(x = var_220_cast_fp16, y = rotr_1_cast_fp16)[name = tensor("op_251_cast_fp16")]; + tensor qoi_1_cast_fp16 = add(x = var_250_cast_fp16, y = var_251_cast_fp16)[name = tensor("qoi_1_cast_fp16")]; + tensor var_254_cast_fp16 = mul(x = var_228_cast_fp16, y = rotr_1_cast_fp16)[name = tensor("op_254_cast_fp16")]; + tensor var_255_cast_fp16 = mul(x = var_236_cast_fp16, y = roti_1_cast_fp16)[name = tensor("op_255_cast_fp16")]; + tensor kor_1_cast_fp16 = sub(x = var_254_cast_fp16, y = var_255_cast_fp16)[name = tensor("kor_1_cast_fp16")]; + tensor var_258_cast_fp16 = mul(x = var_228_cast_fp16, y = roti_1_cast_fp16)[name = tensor("op_258_cast_fp16")]; + tensor var_259_cast_fp16 = mul(x = var_236_cast_fp16, y = rotr_1_cast_fp16)[name = tensor("op_259_cast_fp16")]; + tensor koi_1_cast_fp16 = add(x = var_258_cast_fp16, y = var_259_cast_fp16)[name = tensor("koi_1_cast_fp16")]; + tensor qo_1_axis_0 = const()[name = tensor("qo_1_axis_0"), val = tensor(-1)]; + tensor qo_1_cast_fp16 = stack(axis = qo_1_axis_0, values = (qor_1_cast_fp16, qoi_1_cast_fp16))[name = tensor("qo_1_cast_fp16")]; + tensor ko_1_axis_0 = const()[name = tensor("ko_1_axis_0"), val = tensor(-1)]; + tensor ko_1_cast_fp16 = stack(axis = ko_1_axis_0, values = (kor_1_cast_fp16, koi_1_cast_fp16))[name = tensor("ko_1_cast_fp16")]; + tensor var_288 = const()[name = tensor("op_288"), val = tensor([1, 1, 16, 64])]; + tensor q_3_cast_fp16 = reshape(shape = var_288, x = qo_1_cast_fp16)[name = tensor("q_3_cast_fp16")]; + tensor var_290 = const()[name = tensor("op_290"), val = tensor([1, 1, 16, 64])]; + tensor k_3_cast_fp16 = reshape(shape = var_290, x = ko_1_cast_fp16)[name = tensor("k_3_cast_fp16")]; + tensor _inversed_312_y_0_to_fp16 = const()[name = tensor("_inversed_312_y_0_to_fp16"), val = tensor(0x1p-9)]; + tensor _inversed_312_cast_fp16 = mul(x = ts_5_cast_fp16, y = _inversed_312_y_0_to_fp16)[name = tensor("_inversed_312_cast_fp16")]; + tensor var_313_cast_fp16 = floor(x = _inversed_312_cast_fp16)[name = tensor("op_313_cast_fp16")]; + tensor var_314_to_fp16 = const()[name = tensor("op_314_to_fp16"), val = tensor(0x1p+9)]; + tensor var_315_cast_fp16 = mul(x = var_313_cast_fp16, y = var_314_to_fp16)[name = tensor("op_315_cast_fp16")]; + tensor write_indices_float_3_cast_fp16 = sub(x = ts_5_cast_fp16, y = var_315_cast_fp16)[name = tensor("write_indices_float_3_cast_fp16")]; + tensor var_322_dtype_0 = const()[name = tensor("op_322_dtype_0"), val = tensor("int32")]; + tensor write_indices_1_reps_0 = const()[name = tensor("write_indices_1_reps_0"), val = tensor([1, 1, 16, 64])]; + tensor write_indices_float_3_cast_fp16_to_int32 = cast(dtype = var_322_dtype_0, x = write_indices_float_3_cast_fp16)[name = tensor("cast_138")]; + tensor write_indices_1 = tile(reps = write_indices_1_reps_0, x = write_indices_float_3_cast_fp16_to_int32)[name = tensor("write_indices_1")]; + tensor var_330_begin_0 = const()[name = tensor("op_330_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_330_end_0 = const()[name = tensor("op_330_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor var_330_end_mask_0 = const()[name = tensor("op_330_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_330_squeeze_mask_0 = const()[name = tensor("op_330_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor cache0_to_fp16_dtype_0 = const()[name = tensor("cache0_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor cache0_to_fp16 = cast(dtype = cache0_to_fp16_dtype_0, x = cache0)[name = tensor("cast_137")]; + tensor var_330_cast_fp16 = slice_by_index(begin = var_330_begin_0, end = var_330_end_0, end_mask = var_330_end_mask_0, squeeze_mask = var_330_squeeze_mask_0, x = cache0_to_fp16)[name = tensor("op_330_cast_fp16")]; + tensor var_332_axis_0 = const()[name = tensor("op_332_axis_0"), val = tensor(1)]; + tensor var_332_mode_0 = const()[name = tensor("op_332_mode_0"), val = tensor("update")]; + tensor var_332_validate_indices_0 = const()[name = tensor("op_332_validate_indices_0"), val = tensor(false)]; + tensor var_332_cast_fp16 = scatter_along_axis(axis = var_332_axis_0, data = var_330_cast_fp16, indices = write_indices_1, mode = var_332_mode_0, updates = k_3_cast_fp16, validate_indices = var_332_validate_indices_0)[name = tensor("op_332_cast_fp16")]; + tensor concat_2 = const()[name = tensor("concat_2"), val = tensor([0, 0, 0, 0, 0])]; + tensor concat_3 = const()[name = tensor("concat_3"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_1_internal_tensor_assign_1_stride_0 = const()[name = tensor("new_cache_1_internal_tensor_assign_1_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_1_internal_tensor_assign_1_begin_mask_0 = const()[name = tensor("new_cache_1_internal_tensor_assign_1_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_1_internal_tensor_assign_1_end_mask_0 = const()[name = tensor("new_cache_1_internal_tensor_assign_1_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_1_internal_tensor_assign_1_squeeze_mask_0 = const()[name = tensor("new_cache_1_internal_tensor_assign_1_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_12 = const()[name = tensor("shape_12"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_0 = const()[name = tensor("reduce_prod_0"), val = tensor(1048576)]; + tensor range_1d_0_start_0 = const()[name = tensor("range_1d_0_start_0"), val = tensor(0)]; + tensor range_1d_0_step_0 = const()[name = tensor("range_1d_0_step_0"), val = tensor(1)]; + tensor range_1d_0 = range_1d(end = reduce_prod_0, start = range_1d_0_start_0, step = range_1d_0_step_0)[name = tensor("range_1d_0")]; + tensor reshape_0 = reshape(shape = shape_12, x = range_1d_0)[name = tensor("reshape_0")]; + tensor slice_by_index_0 = slice_by_index(begin = concat_2, begin_mask = new_cache_1_internal_tensor_assign_1_begin_mask_0, end = concat_3, end_mask = new_cache_1_internal_tensor_assign_1_end_mask_0, squeeze_mask = new_cache_1_internal_tensor_assign_1_squeeze_mask_0, stride = new_cache_1_internal_tensor_assign_1_stride_0, x = reshape_0)[name = tensor("slice_by_index_0")]; + tensor reshape_1_shape_0 = const()[name = tensor("reshape_1_shape_0"), val = tensor([-1])]; + tensor reshape_1 = reshape(shape = reshape_1_shape_0, x = slice_by_index_0)[name = tensor("reshape_1")]; + tensor reshape_2_shape_0 = const()[name = tensor("reshape_2_shape_0"), val = tensor([-1])]; + tensor reshape_2_cast_fp16 = reshape(shape = reshape_2_shape_0, x = var_332_cast_fp16)[name = tensor("reshape_2_cast_fp16")]; + tensor reshape_3_shape_0 = const()[name = tensor("reshape_3_shape_0"), val = tensor([-1])]; + tensor reshape_3_cast_fp16 = reshape(shape = reshape_3_shape_0, x = cache0_to_fp16)[name = tensor("reshape_3_cast_fp16")]; + tensor scatter_0_mode_0 = const()[name = tensor("scatter_0_mode_0"), val = tensor("update")]; + tensor scatter_0_axis_0 = const()[name = tensor("scatter_0_axis_0"), val = tensor(0)]; + tensor scatter_0_validate_indices_0 = const()[name = tensor("scatter_0_validate_indices_0"), val = tensor(false)]; + tensor scatter_0_cast_fp16 = scatter(axis = scatter_0_axis_0, data = reshape_3_cast_fp16, indices = reshape_1, mode = scatter_0_mode_0, updates = reshape_2_cast_fp16, validate_indices = scatter_0_validate_indices_0)[name = tensor("scatter_0_cast_fp16")]; + tensor reshape_4_cast_fp16 = reshape(shape = shape_12, x = scatter_0_cast_fp16)[name = tensor("reshape_4_cast_fp16")]; + tensor var_340_begin_0 = const()[name = tensor("op_340_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor var_340_end_0 = const()[name = tensor("op_340_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor var_340_end_mask_0 = const()[name = tensor("op_340_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_340_squeeze_mask_0 = const()[name = tensor("op_340_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor var_340_cast_fp16 = slice_by_index(begin = var_340_begin_0, end = var_340_end_0, end_mask = var_340_end_mask_0, squeeze_mask = var_340_squeeze_mask_0, x = reshape_4_cast_fp16)[name = tensor("op_340_cast_fp16")]; + tensor var_342_axis_0 = const()[name = tensor("op_342_axis_0"), val = tensor(1)]; + tensor var_342_mode_0 = const()[name = tensor("op_342_mode_0"), val = tensor("update")]; + tensor var_342_validate_indices_0 = const()[name = tensor("op_342_validate_indices_0"), val = tensor(false)]; + tensor var_342_cast_fp16 = scatter_along_axis(axis = var_342_axis_0, data = var_340_cast_fp16, indices = write_indices_1, mode = var_342_mode_0, updates = v_1_cast_fp16, validate_indices = var_342_validate_indices_0)[name = tensor("op_342_cast_fp16")]; + tensor concat_4 = const()[name = tensor("concat_4"), val = tensor([1, 0, 0, 0, 0])]; + tensor concat_5 = const()[name = tensor("concat_5"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_1_internal_tensor_assign_2_stride_0 = const()[name = tensor("new_cache_1_internal_tensor_assign_2_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_1_internal_tensor_assign_2_begin_mask_0 = const()[name = tensor("new_cache_1_internal_tensor_assign_2_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_1_internal_tensor_assign_2_end_mask_0 = const()[name = tensor("new_cache_1_internal_tensor_assign_2_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_1_internal_tensor_assign_2_squeeze_mask_0 = const()[name = tensor("new_cache_1_internal_tensor_assign_2_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_13 = const()[name = tensor("shape_13"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_1 = const()[name = tensor("reduce_prod_1"), val = tensor(1048576)]; + tensor range_1d_1_start_0 = const()[name = tensor("range_1d_1_start_0"), val = tensor(0)]; + tensor range_1d_1_step_0 = const()[name = tensor("range_1d_1_step_0"), val = tensor(1)]; + tensor range_1d_1 = range_1d(end = reduce_prod_1, start = range_1d_1_start_0, step = range_1d_1_step_0)[name = tensor("range_1d_1")]; + tensor reshape_5 = reshape(shape = shape_13, x = range_1d_1)[name = tensor("reshape_5")]; + tensor slice_by_index_1 = slice_by_index(begin = concat_4, begin_mask = new_cache_1_internal_tensor_assign_2_begin_mask_0, end = concat_5, end_mask = new_cache_1_internal_tensor_assign_2_end_mask_0, squeeze_mask = new_cache_1_internal_tensor_assign_2_squeeze_mask_0, stride = new_cache_1_internal_tensor_assign_2_stride_0, x = reshape_5)[name = tensor("slice_by_index_1")]; + tensor reshape_6_shape_0 = const()[name = tensor("reshape_6_shape_0"), val = tensor([-1])]; + tensor reshape_6 = reshape(shape = reshape_6_shape_0, x = slice_by_index_1)[name = tensor("reshape_6")]; + tensor reshape_7_shape_0 = const()[name = tensor("reshape_7_shape_0"), val = tensor([-1])]; + tensor reshape_7_cast_fp16 = reshape(shape = reshape_7_shape_0, x = var_342_cast_fp16)[name = tensor("reshape_7_cast_fp16")]; + tensor reshape_8_shape_0 = const()[name = tensor("reshape_8_shape_0"), val = tensor([-1])]; + tensor reshape_8_cast_fp16 = reshape(shape = reshape_8_shape_0, x = reshape_4_cast_fp16)[name = tensor("reshape_8_cast_fp16")]; + tensor scatter_1_mode_0 = const()[name = tensor("scatter_1_mode_0"), val = tensor("update")]; + tensor scatter_1_axis_0 = const()[name = tensor("scatter_1_axis_0"), val = tensor(0)]; + tensor scatter_1_validate_indices_0 = const()[name = tensor("scatter_1_validate_indices_0"), val = tensor(false)]; + tensor scatter_1_cast_fp16 = scatter(axis = scatter_1_axis_0, data = reshape_8_cast_fp16, indices = reshape_6, mode = scatter_1_mode_0, updates = reshape_7_cast_fp16, validate_indices = scatter_1_validate_indices_0)[name = tensor("scatter_1_cast_fp16")]; + tensor reshape_9_cast_fp16 = reshape(shape = shape_13, x = scatter_1_cast_fp16)[name = tensor("reshape_9_cast_fp16")]; + tensor reshape_9_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("reshape_9_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor keys_1_begin_0 = const()[name = tensor("keys_1_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor keys_1_end_0 = const()[name = tensor("keys_1_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor keys_1_end_mask_0 = const()[name = tensor("keys_1_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor keys_1_squeeze_mask_0 = const()[name = tensor("keys_1_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor keys_1_cast_fp16 = slice_by_index(begin = keys_1_begin_0, end = keys_1_end_0, end_mask = keys_1_end_mask_0, squeeze_mask = keys_1_squeeze_mask_0, x = reshape_9_cast_fp16)[name = tensor("keys_1_cast_fp16")]; + tensor values_1_begin_0 = const()[name = tensor("values_1_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor values_1_end_0 = const()[name = tensor("values_1_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor values_1_end_mask_0 = const()[name = tensor("values_1_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor values_1_squeeze_mask_0 = const()[name = tensor("values_1_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor values_1_cast_fp16 = slice_by_index(begin = values_1_begin_0, end = values_1_end_0, end_mask = values_1_end_mask_0, squeeze_mask = values_1_squeeze_mask_0, x = reshape_9_cast_fp16)[name = tensor("values_1_cast_fp16")]; + tensor var_354_cast_fp16 = not_equal(x = keys_1_cast_fp16, y = keys_1_cast_fp16)[name = tensor("op_354_cast_fp16")]; + tensor var_360_to_fp16 = const()[name = tensor("op_360_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(75697856)))]; + tensor keys_3_cast_fp16 = select(a = var_360_to_fp16, b = keys_1_cast_fp16, cond = var_354_cast_fp16)[name = tensor("keys_3_cast_fp16")]; + tensor var_362_cast_fp16 = not_equal(x = values_1_cast_fp16, y = values_1_cast_fp16)[name = tensor("op_362_cast_fp16")]; + tensor values_3_cast_fp16 = select(a = var_360_to_fp16, b = values_1_cast_fp16, cond = var_362_cast_fp16)[name = tensor("values_3_cast_fp16")]; + tensor var_386 = const()[name = tensor("op_386"), val = tensor([0, 2, 1, 3])]; + tensor var_399 = const()[name = tensor("op_399"), val = tensor([1, 1, 1])]; + tensor var_400_cast_fp16 = reshape(shape = var_399, x = position0_to_fp16)[name = tensor("op_400_cast_fp16")]; + tensor var_417_to_fp16 = const()[name = tensor("op_417_to_fp16"), val = tensor(0x1p+0)]; + tensor valid_len_1_cast_fp16 = add(x = var_400_cast_fp16, y = var_417_to_fp16)[name = tensor("valid_len_1_cast_fp16")]; + tensor k_positions_1_promoted_to_fp16 = const()[name = tensor("k_positions_1_promoted_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(76746496)))]; + tensor valid_mask_1_cast_fp16 = less(x = k_positions_1_promoted_to_fp16, y = valid_len_1_cast_fp16)[name = tensor("valid_mask_1_cast_fp16")]; + tensor causal_mask_1_cast_fp16 = less_equal(x = k_positions_1_promoted_to_fp16, y = var_400_cast_fp16)[name = tensor("causal_mask_1_cast_fp16")]; + tensor attn_mask_1 = logical_and(x = valid_mask_1_cast_fp16, y = causal_mask_1_cast_fp16)[name = tensor("attn_mask_1")]; + tensor attn_mask_3_axes_0 = const()[name = tensor("attn_mask_3_axes_0"), val = tensor([1])]; + tensor attn_mask_3 = expand_dims(axes = attn_mask_3_axes_0, x = attn_mask_1)[name = tensor("attn_mask_3")]; + tensor var_435_transpose_x_0 = const()[name = tensor("op_435_transpose_x_0"), val = tensor(false)]; + tensor var_435_transpose_y_0 = const()[name = tensor("op_435_transpose_y_0"), val = tensor(false)]; + tensor transpose_18_perm_0 = const()[name = tensor("transpose_18_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_19_perm_0 = const()[name = tensor("transpose_19_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_19 = transpose(perm = transpose_19_perm_0, x = keys_3_cast_fp16)[name = tensor("transpose_51")]; + tensor transpose_18 = transpose(perm = transpose_18_perm_0, x = q_3_cast_fp16)[name = tensor("transpose_52")]; + tensor var_435_cast_fp16 = matmul(transpose_x = var_435_transpose_x_0, transpose_y = var_435_transpose_y_0, x = transpose_18, y = transpose_19)[name = tensor("op_435_cast_fp16")]; + tensor var_429_to_fp16 = const()[name = tensor("op_429_to_fp16"), val = tensor([0x1p-3])]; + tensor attn_weights_1_cast_fp16 = mul(x = var_435_cast_fp16, y = var_429_to_fp16)[name = tensor("attn_weights_1_cast_fp16")]; + tensor var_437 = logical_not(x = attn_mask_3)[name = tensor("op_437")]; + tensor var_438_to_fp16 = const()[name = tensor("op_438_to_fp16"), val = tensor(-inf)]; + tensor attn_weights_3_cast_fp16 = select(a = var_438_to_fp16, b = attn_weights_1_cast_fp16, cond = var_437)[name = tensor("attn_weights_3_cast_fp16")]; + tensor var_440 = const()[name = tensor("op_440"), val = tensor(-1)]; + tensor attn_weights_5_cast_fp16 = softmax(axis = var_440, x = attn_weights_3_cast_fp16)[name = tensor("attn_weights_5_cast_fp16")]; + tensor attn_output_1_transpose_x_0 = const()[name = tensor("attn_output_1_transpose_x_0"), val = tensor(false)]; + tensor attn_output_1_transpose_y_0 = const()[name = tensor("attn_output_1_transpose_y_0"), val = tensor(false)]; + tensor values_5_cast_fp16 = transpose(perm = var_386, x = values_3_cast_fp16)[name = tensor("transpose_53")]; + tensor attn_output_1_cast_fp16 = matmul(transpose_x = attn_output_1_transpose_x_0, transpose_y = attn_output_1_transpose_y_0, x = attn_weights_5_cast_fp16, y = values_5_cast_fp16)[name = tensor("attn_output_1_cast_fp16")]; + tensor var_448 = const()[name = tensor("op_448"), val = tensor([0, 2, 1, 3])]; + tensor var_451 = const()[name = tensor("op_451"), val = tensor([1, 1, 1024])]; + tensor var_449_cast_fp16 = transpose(perm = var_448, x = attn_output_1_cast_fp16)[name = tensor("transpose_50")]; + tensor input_5_cast_fp16 = reshape(shape = var_451, x = var_449_cast_fp16)[name = tensor("input_5_cast_fp16")]; + tensor linear_2_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = attn0_out_proj_weight_cast_fp16, x = input_5_cast_fp16)[name = tensor("linear_2_cast_fp16")]; + tensor var_457_to_fp16 = const()[name = tensor("op_457_to_fp16"), val = tensor(0x1p+0)]; + tensor var_458_cast_fp16 = add(x = position0_to_fp16, y = var_457_to_fp16)[name = tensor("op_458_cast_fp16")]; + tensor var_458_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_458_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor input_7_cast_fp16 = add(x = linear_0_cast_fp16, y = linear_2_cast_fp16)[name = tensor("input_7_cast_fp16")]; + tensor input_9_axes_0 = const()[name = tensor("input_9_axes_0"), val = tensor([-1])]; + tensor norm0_2_weight_to_fp16 = const()[name = tensor("norm0_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(76747584)))]; + tensor norm0_2_bias_to_fp16 = const()[name = tensor("norm0_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(76749696)))]; + tensor var_462_to_fp16 = const()[name = tensor("op_462_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_9_cast_fp16 = layer_norm(axes = input_9_axes_0, beta = norm0_2_bias_to_fp16, epsilon = var_462_to_fp16, gamma = norm0_2_weight_to_fp16, x = input_7_cast_fp16)[name = tensor("input_9_cast_fp16")]; + tensor linear_3_bias_0_to_fp16 = const()[name = tensor("linear_3_bias_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(76751808)))]; + tensor linear_3_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = linear0_1_weight_cast_fp16, x = input_9_cast_fp16)[name = tensor("linear_3_cast_fp16")]; + tensor input_11_mode_0 = const()[name = tensor("input_11_mode_0"), val = tensor("EXACT")]; + tensor input_11_cast_fp16 = gelu(mode = input_11_mode_0, x = linear_3_cast_fp16)[name = tensor("input_11_cast_fp16")]; + tensor linear_4_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = linear0_2_weight_cast_fp16, x = input_11_cast_fp16)[name = tensor("linear_4_cast_fp16")]; + tensor input_13_cast_fp16 = add(x = input_7_cast_fp16, y = linear_4_cast_fp16)[name = tensor("input_13_cast_fp16")]; + tensor x_3_axes_0 = const()[name = tensor("x_3_axes_0"), val = tensor([-1])]; + tensor norm1_1_weight_to_fp16 = const()[name = tensor("norm1_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(76760064)))]; + tensor norm1_1_bias_to_fp16 = const()[name = tensor("norm1_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(76762176)))]; + tensor var_479_to_fp16 = const()[name = tensor("op_479_to_fp16"), val = tensor(0x1.5p-17)]; + tensor x_3_cast_fp16 = layer_norm(axes = x_3_axes_0, beta = norm1_1_bias_to_fp16, epsilon = var_479_to_fp16, gamma = norm1_1_weight_to_fp16, x = input_13_cast_fp16)[name = tensor("x_3_cast_fp16")]; + tensor linear_5_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = attn1_in_proj_weight_cast_fp16, x = x_3_cast_fp16)[name = tensor("linear_5_cast_fp16")]; + tensor var_515 = const()[name = tensor("op_515"), val = tensor([1, 1, 3, 16, 64])]; + tensor qkv_3_cast_fp16 = reshape(shape = var_515, x = linear_5_cast_fp16)[name = tensor("qkv_3_cast_fp16")]; + tensor q_7_begin_0 = const()[name = tensor("q_7_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor q_7_end_0 = const()[name = tensor("q_7_end_0"), val = tensor([1, 1, 1, 16, 64])]; + tensor q_7_end_mask_0 = const()[name = tensor("q_7_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor q_7_squeeze_mask_0 = const()[name = tensor("q_7_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor q_7_cast_fp16 = slice_by_index(begin = q_7_begin_0, end = q_7_end_0, end_mask = q_7_end_mask_0, squeeze_mask = q_7_squeeze_mask_0, x = qkv_3_cast_fp16)[name = tensor("q_7_cast_fp16")]; + tensor k_5_begin_0 = const()[name = tensor("k_5_begin_0"), val = tensor([0, 0, 1, 0, 0])]; + tensor k_5_end_0 = const()[name = tensor("k_5_end_0"), val = tensor([1, 1, 2, 16, 64])]; + tensor k_5_end_mask_0 = const()[name = tensor("k_5_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor k_5_squeeze_mask_0 = const()[name = tensor("k_5_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor k_5_cast_fp16 = slice_by_index(begin = k_5_begin_0, end = k_5_end_0, end_mask = k_5_end_mask_0, squeeze_mask = k_5_squeeze_mask_0, x = qkv_3_cast_fp16)[name = tensor("k_5_cast_fp16")]; + tensor v_3_begin_0 = const()[name = tensor("v_3_begin_0"), val = tensor([0, 0, 2, 0, 0])]; + tensor v_3_end_0 = const()[name = tensor("v_3_end_0"), val = tensor([1, 1, 3, 16, 64])]; + tensor v_3_end_mask_0 = const()[name = tensor("v_3_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor v_3_squeeze_mask_0 = const()[name = tensor("v_3_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor v_3_cast_fp16 = slice_by_index(begin = v_3_begin_0, end = v_3_end_0, end_mask = v_3_end_mask_0, squeeze_mask = v_3_squeeze_mask_0, x = qkv_3_cast_fp16)[name = tensor("v_3_cast_fp16")]; + tensor var_619 = const()[name = tensor("op_619"), val = tensor([1, 1, 1, 1])]; + tensor position1_to_fp16_dtype_0 = const()[name = tensor("position1_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor position1_to_fp16 = cast(dtype = position1_to_fp16_dtype_0, x = position1)[name = tensor("cast_134")]; + tensor ts_11_cast_fp16 = reshape(shape = var_619, x = position1_to_fp16)[name = tensor("ts_11_cast_fp16")]; + tensor var_623 = const()[name = tensor("op_623"), val = tensor([1, 1, 16, 32, 2])]; + tensor q_complex_3_cast_fp16 = reshape(shape = var_623, x = q_7_cast_fp16)[name = tensor("q_complex_3_cast_fp16")]; + tensor var_627 = const()[name = tensor("op_627"), val = tensor([1, 1, 16, 32, 2])]; + tensor k_complex_3_cast_fp16 = reshape(shape = var_627, x = k_5_cast_fp16)[name = tensor("k_complex_3_cast_fp16")]; + tensor var_631_begin_0 = const()[name = tensor("op_631_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_631_end_0 = const()[name = tensor("op_631_end_0"), val = tensor([1, 1, 16, 32, 1])]; + tensor var_631_end_mask_0 = const()[name = tensor("op_631_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_631_squeeze_mask_0 = const()[name = tensor("op_631_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_631_cast_fp16 = slice_by_index(begin = var_631_begin_0, end = var_631_end_0, end_mask = var_631_end_mask_0, squeeze_mask = var_631_squeeze_mask_0, x = q_complex_3_cast_fp16)[name = tensor("op_631_cast_fp16")]; + tensor var_639_begin_0 = const()[name = tensor("op_639_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_639_end_0 = const()[name = tensor("op_639_end_0"), val = tensor([1, 1, 16, 32, 2])]; + tensor var_639_end_mask_0 = const()[name = tensor("op_639_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_639_squeeze_mask_0 = const()[name = tensor("op_639_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_639_cast_fp16 = slice_by_index(begin = var_639_begin_0, end = var_639_end_0, end_mask = var_639_end_mask_0, squeeze_mask = var_639_squeeze_mask_0, x = q_complex_3_cast_fp16)[name = tensor("op_639_cast_fp16")]; + tensor var_647_begin_0 = const()[name = tensor("op_647_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_647_end_0 = const()[name = tensor("op_647_end_0"), val = tensor([1, 1, 16, 32, 1])]; + tensor var_647_end_mask_0 = const()[name = tensor("op_647_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_647_squeeze_mask_0 = const()[name = tensor("op_647_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_647_cast_fp16 = slice_by_index(begin = var_647_begin_0, end = var_647_end_0, end_mask = var_647_end_mask_0, squeeze_mask = var_647_squeeze_mask_0, x = k_complex_3_cast_fp16)[name = tensor("op_647_cast_fp16")]; + tensor var_655_begin_0 = const()[name = tensor("op_655_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_655_end_0 = const()[name = tensor("op_655_end_0"), val = tensor([1, 1, 16, 32, 2])]; + tensor var_655_end_mask_0 = const()[name = tensor("op_655_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_655_squeeze_mask_0 = const()[name = tensor("op_655_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_655_cast_fp16 = slice_by_index(begin = var_655_begin_0, end = var_655_end_0, end_mask = var_655_end_mask_0, squeeze_mask = var_655_squeeze_mask_0, x = k_complex_3_cast_fp16)[name = tensor("op_655_cast_fp16")]; + tensor freqs_3_to_fp16 = const()[name = tensor("freqs_3_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(76764288)))]; + tensor var_661_cast_fp16 = mul(x = freqs_3_to_fp16, y = ts_11_cast_fp16)[name = tensor("op_661_cast_fp16")]; + tensor rotr_3_cast_fp16 = cos(x = var_661_cast_fp16)[name = tensor("rotr_3_cast_fp16")]; + tensor roti_3_cast_fp16 = sin(x = var_661_cast_fp16)[name = tensor("roti_3_cast_fp16")]; + tensor var_665_cast_fp16 = mul(x = var_631_cast_fp16, y = rotr_3_cast_fp16)[name = tensor("op_665_cast_fp16")]; + tensor var_666_cast_fp16 = mul(x = var_639_cast_fp16, y = roti_3_cast_fp16)[name = tensor("op_666_cast_fp16")]; + tensor qor_5_cast_fp16 = sub(x = var_665_cast_fp16, y = var_666_cast_fp16)[name = tensor("qor_5_cast_fp16")]; + tensor var_669_cast_fp16 = mul(x = var_631_cast_fp16, y = roti_3_cast_fp16)[name = tensor("op_669_cast_fp16")]; + tensor var_670_cast_fp16 = mul(x = var_639_cast_fp16, y = rotr_3_cast_fp16)[name = tensor("op_670_cast_fp16")]; + tensor qoi_5_cast_fp16 = add(x = var_669_cast_fp16, y = var_670_cast_fp16)[name = tensor("qoi_5_cast_fp16")]; + tensor var_673_cast_fp16 = mul(x = var_647_cast_fp16, y = rotr_3_cast_fp16)[name = tensor("op_673_cast_fp16")]; + tensor var_674_cast_fp16 = mul(x = var_655_cast_fp16, y = roti_3_cast_fp16)[name = tensor("op_674_cast_fp16")]; + tensor kor_5_cast_fp16 = sub(x = var_673_cast_fp16, y = var_674_cast_fp16)[name = tensor("kor_5_cast_fp16")]; + tensor var_677_cast_fp16 = mul(x = var_647_cast_fp16, y = roti_3_cast_fp16)[name = tensor("op_677_cast_fp16")]; + tensor var_678_cast_fp16 = mul(x = var_655_cast_fp16, y = rotr_3_cast_fp16)[name = tensor("op_678_cast_fp16")]; + tensor koi_5_cast_fp16 = add(x = var_677_cast_fp16, y = var_678_cast_fp16)[name = tensor("koi_5_cast_fp16")]; + tensor qo_3_axis_0 = const()[name = tensor("qo_3_axis_0"), val = tensor(-1)]; + tensor qo_3_cast_fp16 = stack(axis = qo_3_axis_0, values = (qor_5_cast_fp16, qoi_5_cast_fp16))[name = tensor("qo_3_cast_fp16")]; + tensor ko_3_axis_0 = const()[name = tensor("ko_3_axis_0"), val = tensor(-1)]; + tensor ko_3_cast_fp16 = stack(axis = ko_3_axis_0, values = (kor_5_cast_fp16, koi_5_cast_fp16))[name = tensor("ko_3_cast_fp16")]; + tensor var_707 = const()[name = tensor("op_707"), val = tensor([1, 1, 16, 64])]; + tensor q_9_cast_fp16 = reshape(shape = var_707, x = qo_3_cast_fp16)[name = tensor("q_9_cast_fp16")]; + tensor var_709 = const()[name = tensor("op_709"), val = tensor([1, 1, 16, 64])]; + tensor k_7_cast_fp16 = reshape(shape = var_709, x = ko_3_cast_fp16)[name = tensor("k_7_cast_fp16")]; + tensor _inversed_731_y_0_to_fp16 = const()[name = tensor("_inversed_731_y_0_to_fp16"), val = tensor(0x1p-9)]; + tensor _inversed_731_cast_fp16 = mul(x = ts_11_cast_fp16, y = _inversed_731_y_0_to_fp16)[name = tensor("_inversed_731_cast_fp16")]; + tensor var_732_cast_fp16 = floor(x = _inversed_731_cast_fp16)[name = tensor("op_732_cast_fp16")]; + tensor var_733_to_fp16 = const()[name = tensor("op_733_to_fp16"), val = tensor(0x1p+9)]; + tensor var_734_cast_fp16 = mul(x = var_732_cast_fp16, y = var_733_to_fp16)[name = tensor("op_734_cast_fp16")]; + tensor write_indices_float_7_cast_fp16 = sub(x = ts_11_cast_fp16, y = var_734_cast_fp16)[name = tensor("write_indices_float_7_cast_fp16")]; + tensor var_741_dtype_0 = const()[name = tensor("op_741_dtype_0"), val = tensor("int32")]; + tensor write_indices_3_reps_0 = const()[name = tensor("write_indices_3_reps_0"), val = tensor([1, 1, 16, 64])]; + tensor write_indices_float_7_cast_fp16_to_int32 = cast(dtype = var_741_dtype_0, x = write_indices_float_7_cast_fp16)[name = tensor("cast_133")]; + tensor write_indices_3 = tile(reps = write_indices_3_reps_0, x = write_indices_float_7_cast_fp16_to_int32)[name = tensor("write_indices_3")]; + tensor var_749_begin_0 = const()[name = tensor("op_749_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_749_end_0 = const()[name = tensor("op_749_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor var_749_end_mask_0 = const()[name = tensor("op_749_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_749_squeeze_mask_0 = const()[name = tensor("op_749_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor cache1_to_fp16_dtype_0 = const()[name = tensor("cache1_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor cache1_to_fp16 = cast(dtype = cache1_to_fp16_dtype_0, x = cache1)[name = tensor("cast_132")]; + tensor var_749_cast_fp16 = slice_by_index(begin = var_749_begin_0, end = var_749_end_0, end_mask = var_749_end_mask_0, squeeze_mask = var_749_squeeze_mask_0, x = cache1_to_fp16)[name = tensor("op_749_cast_fp16")]; + tensor var_751_axis_0 = const()[name = tensor("op_751_axis_0"), val = tensor(1)]; + tensor var_751_mode_0 = const()[name = tensor("op_751_mode_0"), val = tensor("update")]; + tensor var_751_validate_indices_0 = const()[name = tensor("op_751_validate_indices_0"), val = tensor(false)]; + tensor var_751_cast_fp16 = scatter_along_axis(axis = var_751_axis_0, data = var_749_cast_fp16, indices = write_indices_3, mode = var_751_mode_0, updates = k_7_cast_fp16, validate_indices = var_751_validate_indices_0)[name = tensor("op_751_cast_fp16")]; + tensor concat_9 = const()[name = tensor("concat_9"), val = tensor([0, 0, 0, 0, 0])]; + tensor concat_10 = const()[name = tensor("concat_10"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_3_internal_tensor_assign_1_stride_0 = const()[name = tensor("new_cache_3_internal_tensor_assign_1_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_3_internal_tensor_assign_1_begin_mask_0 = const()[name = tensor("new_cache_3_internal_tensor_assign_1_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_3_internal_tensor_assign_1_end_mask_0 = const()[name = tensor("new_cache_3_internal_tensor_assign_1_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_3_internal_tensor_assign_1_squeeze_mask_0 = const()[name = tensor("new_cache_3_internal_tensor_assign_1_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_14 = const()[name = tensor("shape_14"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_2 = const()[name = tensor("reduce_prod_2"), val = tensor(1048576)]; + tensor range_1d_2_start_0 = const()[name = tensor("range_1d_2_start_0"), val = tensor(0)]; + tensor range_1d_2_step_0 = const()[name = tensor("range_1d_2_step_0"), val = tensor(1)]; + tensor range_1d_2 = range_1d(end = reduce_prod_2, start = range_1d_2_start_0, step = range_1d_2_step_0)[name = tensor("range_1d_2")]; + tensor reshape_10 = reshape(shape = shape_14, x = range_1d_2)[name = tensor("reshape_10")]; + tensor slice_by_index_2 = slice_by_index(begin = concat_9, begin_mask = new_cache_3_internal_tensor_assign_1_begin_mask_0, end = concat_10, end_mask = new_cache_3_internal_tensor_assign_1_end_mask_0, squeeze_mask = new_cache_3_internal_tensor_assign_1_squeeze_mask_0, stride = new_cache_3_internal_tensor_assign_1_stride_0, x = reshape_10)[name = tensor("slice_by_index_2")]; + tensor reshape_11_shape_0 = const()[name = tensor("reshape_11_shape_0"), val = tensor([-1])]; + tensor reshape_11 = reshape(shape = reshape_11_shape_0, x = slice_by_index_2)[name = tensor("reshape_11")]; + tensor reshape_12_shape_0 = const()[name = tensor("reshape_12_shape_0"), val = tensor([-1])]; + tensor reshape_12_cast_fp16 = reshape(shape = reshape_12_shape_0, x = var_751_cast_fp16)[name = tensor("reshape_12_cast_fp16")]; + tensor reshape_13_shape_0 = const()[name = tensor("reshape_13_shape_0"), val = tensor([-1])]; + tensor reshape_13_cast_fp16 = reshape(shape = reshape_13_shape_0, x = cache1_to_fp16)[name = tensor("reshape_13_cast_fp16")]; + tensor scatter_2_mode_0 = const()[name = tensor("scatter_2_mode_0"), val = tensor("update")]; + tensor scatter_2_axis_0 = const()[name = tensor("scatter_2_axis_0"), val = tensor(0)]; + tensor scatter_2_validate_indices_0 = const()[name = tensor("scatter_2_validate_indices_0"), val = tensor(false)]; + tensor scatter_2_cast_fp16 = scatter(axis = scatter_2_axis_0, data = reshape_13_cast_fp16, indices = reshape_11, mode = scatter_2_mode_0, updates = reshape_12_cast_fp16, validate_indices = scatter_2_validate_indices_0)[name = tensor("scatter_2_cast_fp16")]; + tensor reshape_14_cast_fp16 = reshape(shape = shape_14, x = scatter_2_cast_fp16)[name = tensor("reshape_14_cast_fp16")]; + tensor var_759_begin_0 = const()[name = tensor("op_759_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor var_759_end_0 = const()[name = tensor("op_759_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor var_759_end_mask_0 = const()[name = tensor("op_759_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_759_squeeze_mask_0 = const()[name = tensor("op_759_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor var_759_cast_fp16 = slice_by_index(begin = var_759_begin_0, end = var_759_end_0, end_mask = var_759_end_mask_0, squeeze_mask = var_759_squeeze_mask_0, x = reshape_14_cast_fp16)[name = tensor("op_759_cast_fp16")]; + tensor var_761_axis_0 = const()[name = tensor("op_761_axis_0"), val = tensor(1)]; + tensor var_761_mode_0 = const()[name = tensor("op_761_mode_0"), val = tensor("update")]; + tensor var_761_validate_indices_0 = const()[name = tensor("op_761_validate_indices_0"), val = tensor(false)]; + tensor var_761_cast_fp16 = scatter_along_axis(axis = var_761_axis_0, data = var_759_cast_fp16, indices = write_indices_3, mode = var_761_mode_0, updates = v_3_cast_fp16, validate_indices = var_761_validate_indices_0)[name = tensor("op_761_cast_fp16")]; + tensor concat_11 = const()[name = tensor("concat_11"), val = tensor([1, 0, 0, 0, 0])]; + tensor concat_12 = const()[name = tensor("concat_12"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_3_internal_tensor_assign_2_stride_0 = const()[name = tensor("new_cache_3_internal_tensor_assign_2_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_3_internal_tensor_assign_2_begin_mask_0 = const()[name = tensor("new_cache_3_internal_tensor_assign_2_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_3_internal_tensor_assign_2_end_mask_0 = const()[name = tensor("new_cache_3_internal_tensor_assign_2_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_3_internal_tensor_assign_2_squeeze_mask_0 = const()[name = tensor("new_cache_3_internal_tensor_assign_2_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_15 = const()[name = tensor("shape_15"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_3 = const()[name = tensor("reduce_prod_3"), val = tensor(1048576)]; + tensor range_1d_3_start_0 = const()[name = tensor("range_1d_3_start_0"), val = tensor(0)]; + tensor range_1d_3_step_0 = const()[name = tensor("range_1d_3_step_0"), val = tensor(1)]; + tensor range_1d_3 = range_1d(end = reduce_prod_3, start = range_1d_3_start_0, step = range_1d_3_step_0)[name = tensor("range_1d_3")]; + tensor reshape_15 = reshape(shape = shape_15, x = range_1d_3)[name = tensor("reshape_15")]; + tensor slice_by_index_3 = slice_by_index(begin = concat_11, begin_mask = new_cache_3_internal_tensor_assign_2_begin_mask_0, end = concat_12, end_mask = new_cache_3_internal_tensor_assign_2_end_mask_0, squeeze_mask = new_cache_3_internal_tensor_assign_2_squeeze_mask_0, stride = new_cache_3_internal_tensor_assign_2_stride_0, x = reshape_15)[name = tensor("slice_by_index_3")]; + tensor reshape_16_shape_0 = const()[name = tensor("reshape_16_shape_0"), val = tensor([-1])]; + tensor reshape_16 = reshape(shape = reshape_16_shape_0, x = slice_by_index_3)[name = tensor("reshape_16")]; + tensor reshape_17_shape_0 = const()[name = tensor("reshape_17_shape_0"), val = tensor([-1])]; + tensor reshape_17_cast_fp16 = reshape(shape = reshape_17_shape_0, x = var_761_cast_fp16)[name = tensor("reshape_17_cast_fp16")]; + tensor reshape_18_shape_0 = const()[name = tensor("reshape_18_shape_0"), val = tensor([-1])]; + tensor reshape_18_cast_fp16 = reshape(shape = reshape_18_shape_0, x = reshape_14_cast_fp16)[name = tensor("reshape_18_cast_fp16")]; + tensor scatter_3_mode_0 = const()[name = tensor("scatter_3_mode_0"), val = tensor("update")]; + tensor scatter_3_axis_0 = const()[name = tensor("scatter_3_axis_0"), val = tensor(0)]; + tensor scatter_3_validate_indices_0 = const()[name = tensor("scatter_3_validate_indices_0"), val = tensor(false)]; + tensor scatter_3_cast_fp16 = scatter(axis = scatter_3_axis_0, data = reshape_18_cast_fp16, indices = reshape_16, mode = scatter_3_mode_0, updates = reshape_17_cast_fp16, validate_indices = scatter_3_validate_indices_0)[name = tensor("scatter_3_cast_fp16")]; + tensor reshape_19_cast_fp16 = reshape(shape = shape_15, x = scatter_3_cast_fp16)[name = tensor("reshape_19_cast_fp16")]; + tensor reshape_19_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("reshape_19_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor keys_7_begin_0 = const()[name = tensor("keys_7_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor keys_7_end_0 = const()[name = tensor("keys_7_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor keys_7_end_mask_0 = const()[name = tensor("keys_7_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor keys_7_squeeze_mask_0 = const()[name = tensor("keys_7_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor keys_7_cast_fp16 = slice_by_index(begin = keys_7_begin_0, end = keys_7_end_0, end_mask = keys_7_end_mask_0, squeeze_mask = keys_7_squeeze_mask_0, x = reshape_19_cast_fp16)[name = tensor("keys_7_cast_fp16")]; + tensor values_7_begin_0 = const()[name = tensor("values_7_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor values_7_end_0 = const()[name = tensor("values_7_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor values_7_end_mask_0 = const()[name = tensor("values_7_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor values_7_squeeze_mask_0 = const()[name = tensor("values_7_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor values_7_cast_fp16 = slice_by_index(begin = values_7_begin_0, end = values_7_end_0, end_mask = values_7_end_mask_0, squeeze_mask = values_7_squeeze_mask_0, x = reshape_19_cast_fp16)[name = tensor("values_7_cast_fp16")]; + tensor var_773_cast_fp16 = not_equal(x = keys_7_cast_fp16, y = keys_7_cast_fp16)[name = tensor("op_773_cast_fp16")]; + tensor keys_9_cast_fp16 = select(a = var_360_to_fp16, b = keys_7_cast_fp16, cond = var_773_cast_fp16)[name = tensor("keys_9_cast_fp16")]; + tensor var_781_cast_fp16 = not_equal(x = values_7_cast_fp16, y = values_7_cast_fp16)[name = tensor("op_781_cast_fp16")]; + tensor values_9_cast_fp16 = select(a = var_360_to_fp16, b = values_7_cast_fp16, cond = var_781_cast_fp16)[name = tensor("values_9_cast_fp16")]; + tensor var_805 = const()[name = tensor("op_805"), val = tensor([0, 2, 1, 3])]; + tensor var_818 = const()[name = tensor("op_818"), val = tensor([1, 1, 1])]; + tensor var_819_cast_fp16 = reshape(shape = var_818, x = position1_to_fp16)[name = tensor("op_819_cast_fp16")]; + tensor var_836_to_fp16 = const()[name = tensor("op_836_to_fp16"), val = tensor(0x1p+0)]; + tensor valid_len_3_cast_fp16 = add(x = var_819_cast_fp16, y = var_836_to_fp16)[name = tensor("valid_len_3_cast_fp16")]; + tensor valid_mask_3_cast_fp16 = less(x = k_positions_1_promoted_to_fp16, y = valid_len_3_cast_fp16)[name = tensor("valid_mask_3_cast_fp16")]; + tensor causal_mask_3_cast_fp16 = less_equal(x = k_positions_1_promoted_to_fp16, y = var_819_cast_fp16)[name = tensor("causal_mask_3_cast_fp16")]; + tensor attn_mask_5 = logical_and(x = valid_mask_3_cast_fp16, y = causal_mask_3_cast_fp16)[name = tensor("attn_mask_5")]; + tensor attn_mask_7_axes_0 = const()[name = tensor("attn_mask_7_axes_0"), val = tensor([1])]; + tensor attn_mask_7 = expand_dims(axes = attn_mask_7_axes_0, x = attn_mask_5)[name = tensor("attn_mask_7")]; + tensor var_854_transpose_x_0 = const()[name = tensor("op_854_transpose_x_0"), val = tensor(false)]; + tensor var_854_transpose_y_0 = const()[name = tensor("op_854_transpose_y_0"), val = tensor(false)]; + tensor transpose_20_perm_0 = const()[name = tensor("transpose_20_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_21_perm_0 = const()[name = tensor("transpose_21_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_21 = transpose(perm = transpose_21_perm_0, x = keys_9_cast_fp16)[name = tensor("transpose_47")]; + tensor transpose_20 = transpose(perm = transpose_20_perm_0, x = q_9_cast_fp16)[name = tensor("transpose_48")]; + tensor var_854_cast_fp16 = matmul(transpose_x = var_854_transpose_x_0, transpose_y = var_854_transpose_y_0, x = transpose_20, y = transpose_21)[name = tensor("op_854_cast_fp16")]; + tensor var_848_to_fp16 = const()[name = tensor("op_848_to_fp16"), val = tensor([0x1p-3])]; + tensor attn_weights_7_cast_fp16 = mul(x = var_854_cast_fp16, y = var_848_to_fp16)[name = tensor("attn_weights_7_cast_fp16")]; + tensor var_856 = logical_not(x = attn_mask_7)[name = tensor("op_856")]; + tensor var_857_to_fp16 = const()[name = tensor("op_857_to_fp16"), val = tensor(-inf)]; + tensor attn_weights_9_cast_fp16 = select(a = var_857_to_fp16, b = attn_weights_7_cast_fp16, cond = var_856)[name = tensor("attn_weights_9_cast_fp16")]; + tensor var_859 = const()[name = tensor("op_859"), val = tensor(-1)]; + tensor attn_weights_11_cast_fp16 = softmax(axis = var_859, x = attn_weights_9_cast_fp16)[name = tensor("attn_weights_11_cast_fp16")]; + tensor attn_output_3_transpose_x_0 = const()[name = tensor("attn_output_3_transpose_x_0"), val = tensor(false)]; + tensor attn_output_3_transpose_y_0 = const()[name = tensor("attn_output_3_transpose_y_0"), val = tensor(false)]; + tensor values_11_cast_fp16 = transpose(perm = var_805, x = values_9_cast_fp16)[name = tensor("transpose_49")]; + tensor attn_output_3_cast_fp16 = matmul(transpose_x = attn_output_3_transpose_x_0, transpose_y = attn_output_3_transpose_y_0, x = attn_weights_11_cast_fp16, y = values_11_cast_fp16)[name = tensor("attn_output_3_cast_fp16")]; + tensor var_867 = const()[name = tensor("op_867"), val = tensor([0, 2, 1, 3])]; + tensor var_870 = const()[name = tensor("op_870"), val = tensor([1, 1, 1024])]; + tensor var_868_cast_fp16 = transpose(perm = var_867, x = attn_output_3_cast_fp16)[name = tensor("transpose_46")]; + tensor input_15_cast_fp16 = reshape(shape = var_870, x = var_868_cast_fp16)[name = tensor("input_15_cast_fp16")]; + tensor linear_6_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = attn1_out_proj_weight_cast_fp16, x = input_15_cast_fp16)[name = tensor("linear_6_cast_fp16")]; + tensor var_876_to_fp16 = const()[name = tensor("op_876_to_fp16"), val = tensor(0x1p+0)]; + tensor var_877_cast_fp16 = add(x = position1_to_fp16, y = var_876_to_fp16)[name = tensor("op_877_cast_fp16")]; + tensor var_877_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_877_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor input_17_cast_fp16 = add(x = input_13_cast_fp16, y = linear_6_cast_fp16)[name = tensor("input_17_cast_fp16")]; + tensor input_19_axes_0 = const()[name = tensor("input_19_axes_0"), val = tensor([-1])]; + tensor norm1_2_weight_to_fp16 = const()[name = tensor("norm1_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(76764416)))]; + tensor norm1_2_bias_to_fp16 = const()[name = tensor("norm1_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(76766528)))]; + tensor var_881_to_fp16 = const()[name = tensor("op_881_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_19_cast_fp16 = layer_norm(axes = input_19_axes_0, beta = norm1_2_bias_to_fp16, epsilon = var_881_to_fp16, gamma = norm1_2_weight_to_fp16, x = input_17_cast_fp16)[name = tensor("input_19_cast_fp16")]; + tensor linear_7_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = linear1_1_weight_cast_fp16, x = input_19_cast_fp16)[name = tensor("linear_7_cast_fp16")]; + tensor input_21_mode_0 = const()[name = tensor("input_21_mode_0"), val = tensor("EXACT")]; + tensor input_21_cast_fp16 = gelu(mode = input_21_mode_0, x = linear_7_cast_fp16)[name = tensor("input_21_cast_fp16")]; + tensor linear_8_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = linear1_2_weight_cast_fp16, x = input_21_cast_fp16)[name = tensor("linear_8_cast_fp16")]; + tensor input_23_cast_fp16 = add(x = input_17_cast_fp16, y = linear_8_cast_fp16)[name = tensor("input_23_cast_fp16")]; + tensor x_5_axes_0 = const()[name = tensor("x_5_axes_0"), val = tensor([-1])]; + tensor norm2_1_weight_to_fp16 = const()[name = tensor("norm2_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(76768640)))]; + tensor norm2_1_bias_to_fp16 = const()[name = tensor("norm2_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(76770752)))]; + tensor var_898_to_fp16 = const()[name = tensor("op_898_to_fp16"), val = tensor(0x1.5p-17)]; + tensor x_5_cast_fp16 = layer_norm(axes = x_5_axes_0, beta = norm2_1_bias_to_fp16, epsilon = var_898_to_fp16, gamma = norm2_1_weight_to_fp16, x = input_23_cast_fp16)[name = tensor("x_5_cast_fp16")]; + tensor linear_9_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = attn2_in_proj_weight_cast_fp16, x = x_5_cast_fp16)[name = tensor("linear_9_cast_fp16")]; + tensor var_934 = const()[name = tensor("op_934"), val = tensor([1, 1, 3, 16, 64])]; + tensor qkv_5_cast_fp16 = reshape(shape = var_934, x = linear_9_cast_fp16)[name = tensor("qkv_5_cast_fp16")]; + tensor q_13_begin_0 = const()[name = tensor("q_13_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor q_13_end_0 = const()[name = tensor("q_13_end_0"), val = tensor([1, 1, 1, 16, 64])]; + tensor q_13_end_mask_0 = const()[name = tensor("q_13_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor q_13_squeeze_mask_0 = const()[name = tensor("q_13_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor q_13_cast_fp16 = slice_by_index(begin = q_13_begin_0, end = q_13_end_0, end_mask = q_13_end_mask_0, squeeze_mask = q_13_squeeze_mask_0, x = qkv_5_cast_fp16)[name = tensor("q_13_cast_fp16")]; + tensor k_9_begin_0 = const()[name = tensor("k_9_begin_0"), val = tensor([0, 0, 1, 0, 0])]; + tensor k_9_end_0 = const()[name = tensor("k_9_end_0"), val = tensor([1, 1, 2, 16, 64])]; + tensor k_9_end_mask_0 = const()[name = tensor("k_9_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor k_9_squeeze_mask_0 = const()[name = tensor("k_9_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor k_9_cast_fp16 = slice_by_index(begin = k_9_begin_0, end = k_9_end_0, end_mask = k_9_end_mask_0, squeeze_mask = k_9_squeeze_mask_0, x = qkv_5_cast_fp16)[name = tensor("k_9_cast_fp16")]; + tensor v_5_begin_0 = const()[name = tensor("v_5_begin_0"), val = tensor([0, 0, 2, 0, 0])]; + tensor v_5_end_0 = const()[name = tensor("v_5_end_0"), val = tensor([1, 1, 3, 16, 64])]; + tensor v_5_end_mask_0 = const()[name = tensor("v_5_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor v_5_squeeze_mask_0 = const()[name = tensor("v_5_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor v_5_cast_fp16 = slice_by_index(begin = v_5_begin_0, end = v_5_end_0, end_mask = v_5_end_mask_0, squeeze_mask = v_5_squeeze_mask_0, x = qkv_5_cast_fp16)[name = tensor("v_5_cast_fp16")]; + tensor var_1038 = const()[name = tensor("op_1038"), val = tensor([1, 1, 1, 1])]; + tensor position2_to_fp16_dtype_0 = const()[name = tensor("position2_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor position2_to_fp16 = cast(dtype = position2_to_fp16_dtype_0, x = position2)[name = tensor("cast_129")]; + tensor ts_17_cast_fp16 = reshape(shape = var_1038, x = position2_to_fp16)[name = tensor("ts_17_cast_fp16")]; + tensor var_1042 = const()[name = tensor("op_1042"), val = tensor([1, 1, 16, 32, 2])]; + tensor q_complex_5_cast_fp16 = reshape(shape = var_1042, x = q_13_cast_fp16)[name = tensor("q_complex_5_cast_fp16")]; + tensor var_1046 = const()[name = tensor("op_1046"), val = tensor([1, 1, 16, 32, 2])]; + tensor k_complex_5_cast_fp16 = reshape(shape = var_1046, x = k_9_cast_fp16)[name = tensor("k_complex_5_cast_fp16")]; + tensor var_1050_begin_0 = const()[name = tensor("op_1050_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_1050_end_0 = const()[name = tensor("op_1050_end_0"), val = tensor([1, 1, 16, 32, 1])]; + tensor var_1050_end_mask_0 = const()[name = tensor("op_1050_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_1050_squeeze_mask_0 = const()[name = tensor("op_1050_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_1050_cast_fp16 = slice_by_index(begin = var_1050_begin_0, end = var_1050_end_0, end_mask = var_1050_end_mask_0, squeeze_mask = var_1050_squeeze_mask_0, x = q_complex_5_cast_fp16)[name = tensor("op_1050_cast_fp16")]; + tensor var_1058_begin_0 = const()[name = tensor("op_1058_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_1058_end_0 = const()[name = tensor("op_1058_end_0"), val = tensor([1, 1, 16, 32, 2])]; + tensor var_1058_end_mask_0 = const()[name = tensor("op_1058_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_1058_squeeze_mask_0 = const()[name = tensor("op_1058_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_1058_cast_fp16 = slice_by_index(begin = var_1058_begin_0, end = var_1058_end_0, end_mask = var_1058_end_mask_0, squeeze_mask = var_1058_squeeze_mask_0, x = q_complex_5_cast_fp16)[name = tensor("op_1058_cast_fp16")]; + tensor var_1066_begin_0 = const()[name = tensor("op_1066_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_1066_end_0 = const()[name = tensor("op_1066_end_0"), val = tensor([1, 1, 16, 32, 1])]; + tensor var_1066_end_mask_0 = const()[name = tensor("op_1066_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_1066_squeeze_mask_0 = const()[name = tensor("op_1066_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_1066_cast_fp16 = slice_by_index(begin = var_1066_begin_0, end = var_1066_end_0, end_mask = var_1066_end_mask_0, squeeze_mask = var_1066_squeeze_mask_0, x = k_complex_5_cast_fp16)[name = tensor("op_1066_cast_fp16")]; + tensor var_1074_begin_0 = const()[name = tensor("op_1074_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_1074_end_0 = const()[name = tensor("op_1074_end_0"), val = tensor([1, 1, 16, 32, 2])]; + tensor var_1074_end_mask_0 = const()[name = tensor("op_1074_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_1074_squeeze_mask_0 = const()[name = tensor("op_1074_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_1074_cast_fp16 = slice_by_index(begin = var_1074_begin_0, end = var_1074_end_0, end_mask = var_1074_end_mask_0, squeeze_mask = var_1074_squeeze_mask_0, x = k_complex_5_cast_fp16)[name = tensor("op_1074_cast_fp16")]; + tensor freqs_5_to_fp16 = const()[name = tensor("freqs_5_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(76772864)))]; + tensor var_1080_cast_fp16 = mul(x = freqs_5_to_fp16, y = ts_17_cast_fp16)[name = tensor("op_1080_cast_fp16")]; + tensor rotr_5_cast_fp16 = cos(x = var_1080_cast_fp16)[name = tensor("rotr_5_cast_fp16")]; + tensor roti_5_cast_fp16 = sin(x = var_1080_cast_fp16)[name = tensor("roti_5_cast_fp16")]; + tensor var_1084_cast_fp16 = mul(x = var_1050_cast_fp16, y = rotr_5_cast_fp16)[name = tensor("op_1084_cast_fp16")]; + tensor var_1085_cast_fp16 = mul(x = var_1058_cast_fp16, y = roti_5_cast_fp16)[name = tensor("op_1085_cast_fp16")]; + tensor qor_9_cast_fp16 = sub(x = var_1084_cast_fp16, y = var_1085_cast_fp16)[name = tensor("qor_9_cast_fp16")]; + tensor var_1088_cast_fp16 = mul(x = var_1050_cast_fp16, y = roti_5_cast_fp16)[name = tensor("op_1088_cast_fp16")]; + tensor var_1089_cast_fp16 = mul(x = var_1058_cast_fp16, y = rotr_5_cast_fp16)[name = tensor("op_1089_cast_fp16")]; + tensor qoi_9_cast_fp16 = add(x = var_1088_cast_fp16, y = var_1089_cast_fp16)[name = tensor("qoi_9_cast_fp16")]; + tensor var_1092_cast_fp16 = mul(x = var_1066_cast_fp16, y = rotr_5_cast_fp16)[name = tensor("op_1092_cast_fp16")]; + tensor var_1093_cast_fp16 = mul(x = var_1074_cast_fp16, y = roti_5_cast_fp16)[name = tensor("op_1093_cast_fp16")]; + tensor kor_9_cast_fp16 = sub(x = var_1092_cast_fp16, y = var_1093_cast_fp16)[name = tensor("kor_9_cast_fp16")]; + tensor var_1096_cast_fp16 = mul(x = var_1066_cast_fp16, y = roti_5_cast_fp16)[name = tensor("op_1096_cast_fp16")]; + tensor var_1097_cast_fp16 = mul(x = var_1074_cast_fp16, y = rotr_5_cast_fp16)[name = tensor("op_1097_cast_fp16")]; + tensor koi_9_cast_fp16 = add(x = var_1096_cast_fp16, y = var_1097_cast_fp16)[name = tensor("koi_9_cast_fp16")]; + tensor qo_5_axis_0 = const()[name = tensor("qo_5_axis_0"), val = tensor(-1)]; + tensor qo_5_cast_fp16 = stack(axis = qo_5_axis_0, values = (qor_9_cast_fp16, qoi_9_cast_fp16))[name = tensor("qo_5_cast_fp16")]; + tensor ko_5_axis_0 = const()[name = tensor("ko_5_axis_0"), val = tensor(-1)]; + tensor ko_5_cast_fp16 = stack(axis = ko_5_axis_0, values = (kor_9_cast_fp16, koi_9_cast_fp16))[name = tensor("ko_5_cast_fp16")]; + tensor var_1126 = const()[name = tensor("op_1126"), val = tensor([1, 1, 16, 64])]; + tensor q_15_cast_fp16 = reshape(shape = var_1126, x = qo_5_cast_fp16)[name = tensor("q_15_cast_fp16")]; + tensor var_1128 = const()[name = tensor("op_1128"), val = tensor([1, 1, 16, 64])]; + tensor k_11_cast_fp16 = reshape(shape = var_1128, x = ko_5_cast_fp16)[name = tensor("k_11_cast_fp16")]; + tensor _inversed_1150_y_0_to_fp16 = const()[name = tensor("_inversed_1150_y_0_to_fp16"), val = tensor(0x1p-9)]; + tensor _inversed_1150_cast_fp16 = mul(x = ts_17_cast_fp16, y = _inversed_1150_y_0_to_fp16)[name = tensor("_inversed_1150_cast_fp16")]; + tensor var_1151_cast_fp16 = floor(x = _inversed_1150_cast_fp16)[name = tensor("op_1151_cast_fp16")]; + tensor var_1152_to_fp16 = const()[name = tensor("op_1152_to_fp16"), val = tensor(0x1p+9)]; + tensor var_1153_cast_fp16 = mul(x = var_1151_cast_fp16, y = var_1152_to_fp16)[name = tensor("op_1153_cast_fp16")]; + tensor write_indices_float_11_cast_fp16 = sub(x = ts_17_cast_fp16, y = var_1153_cast_fp16)[name = tensor("write_indices_float_11_cast_fp16")]; + tensor var_1160_dtype_0 = const()[name = tensor("op_1160_dtype_0"), val = tensor("int32")]; + tensor write_indices_5_reps_0 = const()[name = tensor("write_indices_5_reps_0"), val = tensor([1, 1, 16, 64])]; + tensor write_indices_float_11_cast_fp16_to_int32 = cast(dtype = var_1160_dtype_0, x = write_indices_float_11_cast_fp16)[name = tensor("cast_128")]; + tensor write_indices_5 = tile(reps = write_indices_5_reps_0, x = write_indices_float_11_cast_fp16_to_int32)[name = tensor("write_indices_5")]; + tensor var_1168_begin_0 = const()[name = tensor("op_1168_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_1168_end_0 = const()[name = tensor("op_1168_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor var_1168_end_mask_0 = const()[name = tensor("op_1168_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_1168_squeeze_mask_0 = const()[name = tensor("op_1168_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor cache2_to_fp16_dtype_0 = const()[name = tensor("cache2_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor cache2_to_fp16 = cast(dtype = cache2_to_fp16_dtype_0, x = cache2)[name = tensor("cast_127")]; + tensor var_1168_cast_fp16 = slice_by_index(begin = var_1168_begin_0, end = var_1168_end_0, end_mask = var_1168_end_mask_0, squeeze_mask = var_1168_squeeze_mask_0, x = cache2_to_fp16)[name = tensor("op_1168_cast_fp16")]; + tensor var_1170_axis_0 = const()[name = tensor("op_1170_axis_0"), val = tensor(1)]; + tensor var_1170_mode_0 = const()[name = tensor("op_1170_mode_0"), val = tensor("update")]; + tensor var_1170_validate_indices_0 = const()[name = tensor("op_1170_validate_indices_0"), val = tensor(false)]; + tensor var_1170_cast_fp16 = scatter_along_axis(axis = var_1170_axis_0, data = var_1168_cast_fp16, indices = write_indices_5, mode = var_1170_mode_0, updates = k_11_cast_fp16, validate_indices = var_1170_validate_indices_0)[name = tensor("op_1170_cast_fp16")]; + tensor concat_16 = const()[name = tensor("concat_16"), val = tensor([0, 0, 0, 0, 0])]; + tensor concat_17 = const()[name = tensor("concat_17"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_5_internal_tensor_assign_1_stride_0 = const()[name = tensor("new_cache_5_internal_tensor_assign_1_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_5_internal_tensor_assign_1_begin_mask_0 = const()[name = tensor("new_cache_5_internal_tensor_assign_1_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_5_internal_tensor_assign_1_end_mask_0 = const()[name = tensor("new_cache_5_internal_tensor_assign_1_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_5_internal_tensor_assign_1_squeeze_mask_0 = const()[name = tensor("new_cache_5_internal_tensor_assign_1_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_16 = const()[name = tensor("shape_16"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_4 = const()[name = tensor("reduce_prod_4"), val = tensor(1048576)]; + tensor range_1d_4_start_0 = const()[name = tensor("range_1d_4_start_0"), val = tensor(0)]; + tensor range_1d_4_step_0 = const()[name = tensor("range_1d_4_step_0"), val = tensor(1)]; + tensor range_1d_4 = range_1d(end = reduce_prod_4, start = range_1d_4_start_0, step = range_1d_4_step_0)[name = tensor("range_1d_4")]; + tensor reshape_20 = reshape(shape = shape_16, x = range_1d_4)[name = tensor("reshape_20")]; + tensor slice_by_index_4 = slice_by_index(begin = concat_16, begin_mask = new_cache_5_internal_tensor_assign_1_begin_mask_0, end = concat_17, end_mask = new_cache_5_internal_tensor_assign_1_end_mask_0, squeeze_mask = new_cache_5_internal_tensor_assign_1_squeeze_mask_0, stride = new_cache_5_internal_tensor_assign_1_stride_0, x = reshape_20)[name = tensor("slice_by_index_4")]; + tensor reshape_21_shape_0 = const()[name = tensor("reshape_21_shape_0"), val = tensor([-1])]; + tensor reshape_21 = reshape(shape = reshape_21_shape_0, x = slice_by_index_4)[name = tensor("reshape_21")]; + tensor reshape_22_shape_0 = const()[name = tensor("reshape_22_shape_0"), val = tensor([-1])]; + tensor reshape_22_cast_fp16 = reshape(shape = reshape_22_shape_0, x = var_1170_cast_fp16)[name = tensor("reshape_22_cast_fp16")]; + tensor reshape_23_shape_0 = const()[name = tensor("reshape_23_shape_0"), val = tensor([-1])]; + tensor reshape_23_cast_fp16 = reshape(shape = reshape_23_shape_0, x = cache2_to_fp16)[name = tensor("reshape_23_cast_fp16")]; + tensor scatter_4_mode_0 = const()[name = tensor("scatter_4_mode_0"), val = tensor("update")]; + tensor scatter_4_axis_0 = const()[name = tensor("scatter_4_axis_0"), val = tensor(0)]; + tensor scatter_4_validate_indices_0 = const()[name = tensor("scatter_4_validate_indices_0"), val = tensor(false)]; + tensor scatter_4_cast_fp16 = scatter(axis = scatter_4_axis_0, data = reshape_23_cast_fp16, indices = reshape_21, mode = scatter_4_mode_0, updates = reshape_22_cast_fp16, validate_indices = scatter_4_validate_indices_0)[name = tensor("scatter_4_cast_fp16")]; + tensor reshape_24_cast_fp16 = reshape(shape = shape_16, x = scatter_4_cast_fp16)[name = tensor("reshape_24_cast_fp16")]; + tensor var_1178_begin_0 = const()[name = tensor("op_1178_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor var_1178_end_0 = const()[name = tensor("op_1178_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor var_1178_end_mask_0 = const()[name = tensor("op_1178_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_1178_squeeze_mask_0 = const()[name = tensor("op_1178_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor var_1178_cast_fp16 = slice_by_index(begin = var_1178_begin_0, end = var_1178_end_0, end_mask = var_1178_end_mask_0, squeeze_mask = var_1178_squeeze_mask_0, x = reshape_24_cast_fp16)[name = tensor("op_1178_cast_fp16")]; + tensor var_1180_axis_0 = const()[name = tensor("op_1180_axis_0"), val = tensor(1)]; + tensor var_1180_mode_0 = const()[name = tensor("op_1180_mode_0"), val = tensor("update")]; + tensor var_1180_validate_indices_0 = const()[name = tensor("op_1180_validate_indices_0"), val = tensor(false)]; + tensor var_1180_cast_fp16 = scatter_along_axis(axis = var_1180_axis_0, data = var_1178_cast_fp16, indices = write_indices_5, mode = var_1180_mode_0, updates = v_5_cast_fp16, validate_indices = var_1180_validate_indices_0)[name = tensor("op_1180_cast_fp16")]; + tensor concat_18 = const()[name = tensor("concat_18"), val = tensor([1, 0, 0, 0, 0])]; + tensor concat_19 = const()[name = tensor("concat_19"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_5_internal_tensor_assign_2_stride_0 = const()[name = tensor("new_cache_5_internal_tensor_assign_2_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_5_internal_tensor_assign_2_begin_mask_0 = const()[name = tensor("new_cache_5_internal_tensor_assign_2_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_5_internal_tensor_assign_2_end_mask_0 = const()[name = tensor("new_cache_5_internal_tensor_assign_2_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_5_internal_tensor_assign_2_squeeze_mask_0 = const()[name = tensor("new_cache_5_internal_tensor_assign_2_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_17 = const()[name = tensor("shape_17"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_5 = const()[name = tensor("reduce_prod_5"), val = tensor(1048576)]; + tensor range_1d_5_start_0 = const()[name = tensor("range_1d_5_start_0"), val = tensor(0)]; + tensor range_1d_5_step_0 = const()[name = tensor("range_1d_5_step_0"), val = tensor(1)]; + tensor range_1d_5 = range_1d(end = reduce_prod_5, start = range_1d_5_start_0, step = range_1d_5_step_0)[name = tensor("range_1d_5")]; + tensor reshape_25 = reshape(shape = shape_17, x = range_1d_5)[name = tensor("reshape_25")]; + tensor slice_by_index_5 = slice_by_index(begin = concat_18, begin_mask = new_cache_5_internal_tensor_assign_2_begin_mask_0, end = concat_19, end_mask = new_cache_5_internal_tensor_assign_2_end_mask_0, squeeze_mask = new_cache_5_internal_tensor_assign_2_squeeze_mask_0, stride = new_cache_5_internal_tensor_assign_2_stride_0, x = reshape_25)[name = tensor("slice_by_index_5")]; + tensor reshape_26_shape_0 = const()[name = tensor("reshape_26_shape_0"), val = tensor([-1])]; + tensor reshape_26 = reshape(shape = reshape_26_shape_0, x = slice_by_index_5)[name = tensor("reshape_26")]; + tensor reshape_27_shape_0 = const()[name = tensor("reshape_27_shape_0"), val = tensor([-1])]; + tensor reshape_27_cast_fp16 = reshape(shape = reshape_27_shape_0, x = var_1180_cast_fp16)[name = tensor("reshape_27_cast_fp16")]; + tensor reshape_28_shape_0 = const()[name = tensor("reshape_28_shape_0"), val = tensor([-1])]; + tensor reshape_28_cast_fp16 = reshape(shape = reshape_28_shape_0, x = reshape_24_cast_fp16)[name = tensor("reshape_28_cast_fp16")]; + tensor scatter_5_mode_0 = const()[name = tensor("scatter_5_mode_0"), val = tensor("update")]; + tensor scatter_5_axis_0 = const()[name = tensor("scatter_5_axis_0"), val = tensor(0)]; + tensor scatter_5_validate_indices_0 = const()[name = tensor("scatter_5_validate_indices_0"), val = tensor(false)]; + tensor scatter_5_cast_fp16 = scatter(axis = scatter_5_axis_0, data = reshape_28_cast_fp16, indices = reshape_26, mode = scatter_5_mode_0, updates = reshape_27_cast_fp16, validate_indices = scatter_5_validate_indices_0)[name = tensor("scatter_5_cast_fp16")]; + tensor reshape_29_cast_fp16 = reshape(shape = shape_17, x = scatter_5_cast_fp16)[name = tensor("reshape_29_cast_fp16")]; + tensor reshape_29_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("reshape_29_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor keys_13_begin_0 = const()[name = tensor("keys_13_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor keys_13_end_0 = const()[name = tensor("keys_13_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor keys_13_end_mask_0 = const()[name = tensor("keys_13_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor keys_13_squeeze_mask_0 = const()[name = tensor("keys_13_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor keys_13_cast_fp16 = slice_by_index(begin = keys_13_begin_0, end = keys_13_end_0, end_mask = keys_13_end_mask_0, squeeze_mask = keys_13_squeeze_mask_0, x = reshape_29_cast_fp16)[name = tensor("keys_13_cast_fp16")]; + tensor values_13_begin_0 = const()[name = tensor("values_13_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor values_13_end_0 = const()[name = tensor("values_13_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor values_13_end_mask_0 = const()[name = tensor("values_13_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor values_13_squeeze_mask_0 = const()[name = tensor("values_13_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor values_13_cast_fp16 = slice_by_index(begin = values_13_begin_0, end = values_13_end_0, end_mask = values_13_end_mask_0, squeeze_mask = values_13_squeeze_mask_0, x = reshape_29_cast_fp16)[name = tensor("values_13_cast_fp16")]; + tensor var_1192_cast_fp16 = not_equal(x = keys_13_cast_fp16, y = keys_13_cast_fp16)[name = tensor("op_1192_cast_fp16")]; + tensor keys_15_cast_fp16 = select(a = var_360_to_fp16, b = keys_13_cast_fp16, cond = var_1192_cast_fp16)[name = tensor("keys_15_cast_fp16")]; + tensor var_1200_cast_fp16 = not_equal(x = values_13_cast_fp16, y = values_13_cast_fp16)[name = tensor("op_1200_cast_fp16")]; + tensor values_15_cast_fp16 = select(a = var_360_to_fp16, b = values_13_cast_fp16, cond = var_1200_cast_fp16)[name = tensor("values_15_cast_fp16")]; + tensor var_1224 = const()[name = tensor("op_1224"), val = tensor([0, 2, 1, 3])]; + tensor var_1237 = const()[name = tensor("op_1237"), val = tensor([1, 1, 1])]; + tensor var_1238_cast_fp16 = reshape(shape = var_1237, x = position2_to_fp16)[name = tensor("op_1238_cast_fp16")]; + tensor var_1255_to_fp16 = const()[name = tensor("op_1255_to_fp16"), val = tensor(0x1p+0)]; + tensor valid_len_5_cast_fp16 = add(x = var_1238_cast_fp16, y = var_1255_to_fp16)[name = tensor("valid_len_5_cast_fp16")]; + tensor valid_mask_5_cast_fp16 = less(x = k_positions_1_promoted_to_fp16, y = valid_len_5_cast_fp16)[name = tensor("valid_mask_5_cast_fp16")]; + tensor causal_mask_5_cast_fp16 = less_equal(x = k_positions_1_promoted_to_fp16, y = var_1238_cast_fp16)[name = tensor("causal_mask_5_cast_fp16")]; + tensor attn_mask_9 = logical_and(x = valid_mask_5_cast_fp16, y = causal_mask_5_cast_fp16)[name = tensor("attn_mask_9")]; + tensor attn_mask_11_axes_0 = const()[name = tensor("attn_mask_11_axes_0"), val = tensor([1])]; + tensor attn_mask_11 = expand_dims(axes = attn_mask_11_axes_0, x = attn_mask_9)[name = tensor("attn_mask_11")]; + tensor var_1273_transpose_x_0 = const()[name = tensor("op_1273_transpose_x_0"), val = tensor(false)]; + tensor var_1273_transpose_y_0 = const()[name = tensor("op_1273_transpose_y_0"), val = tensor(false)]; + tensor transpose_22_perm_0 = const()[name = tensor("transpose_22_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_23_perm_0 = const()[name = tensor("transpose_23_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_23 = transpose(perm = transpose_23_perm_0, x = keys_15_cast_fp16)[name = tensor("transpose_43")]; + tensor transpose_22 = transpose(perm = transpose_22_perm_0, x = q_15_cast_fp16)[name = tensor("transpose_44")]; + tensor var_1273_cast_fp16 = matmul(transpose_x = var_1273_transpose_x_0, transpose_y = var_1273_transpose_y_0, x = transpose_22, y = transpose_23)[name = tensor("op_1273_cast_fp16")]; + tensor var_1267_to_fp16 = const()[name = tensor("op_1267_to_fp16"), val = tensor([0x1p-3])]; + tensor attn_weights_13_cast_fp16 = mul(x = var_1273_cast_fp16, y = var_1267_to_fp16)[name = tensor("attn_weights_13_cast_fp16")]; + tensor var_1275 = logical_not(x = attn_mask_11)[name = tensor("op_1275")]; + tensor var_1276_to_fp16 = const()[name = tensor("op_1276_to_fp16"), val = tensor(-inf)]; + tensor attn_weights_15_cast_fp16 = select(a = var_1276_to_fp16, b = attn_weights_13_cast_fp16, cond = var_1275)[name = tensor("attn_weights_15_cast_fp16")]; + tensor var_1278 = const()[name = tensor("op_1278"), val = tensor(-1)]; + tensor attn_weights_17_cast_fp16 = softmax(axis = var_1278, x = attn_weights_15_cast_fp16)[name = tensor("attn_weights_17_cast_fp16")]; + tensor attn_output_5_transpose_x_0 = const()[name = tensor("attn_output_5_transpose_x_0"), val = tensor(false)]; + tensor attn_output_5_transpose_y_0 = const()[name = tensor("attn_output_5_transpose_y_0"), val = tensor(false)]; + tensor values_17_cast_fp16 = transpose(perm = var_1224, x = values_15_cast_fp16)[name = tensor("transpose_45")]; + tensor attn_output_5_cast_fp16 = matmul(transpose_x = attn_output_5_transpose_x_0, transpose_y = attn_output_5_transpose_y_0, x = attn_weights_17_cast_fp16, y = values_17_cast_fp16)[name = tensor("attn_output_5_cast_fp16")]; + tensor var_1286 = const()[name = tensor("op_1286"), val = tensor([0, 2, 1, 3])]; + tensor var_1289 = const()[name = tensor("op_1289"), val = tensor([1, 1, 1024])]; + tensor var_1287_cast_fp16 = transpose(perm = var_1286, x = attn_output_5_cast_fp16)[name = tensor("transpose_42")]; + tensor input_25_cast_fp16 = reshape(shape = var_1289, x = var_1287_cast_fp16)[name = tensor("input_25_cast_fp16")]; + tensor linear_10_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = attn2_out_proj_weight_cast_fp16, x = input_25_cast_fp16)[name = tensor("linear_10_cast_fp16")]; + tensor var_1295_to_fp16 = const()[name = tensor("op_1295_to_fp16"), val = tensor(0x1p+0)]; + tensor var_1296_cast_fp16 = add(x = position2_to_fp16, y = var_1295_to_fp16)[name = tensor("op_1296_cast_fp16")]; + tensor var_1296_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_1296_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor input_27_cast_fp16 = add(x = input_23_cast_fp16, y = linear_10_cast_fp16)[name = tensor("input_27_cast_fp16")]; + tensor input_29_axes_0 = const()[name = tensor("input_29_axes_0"), val = tensor([-1])]; + tensor norm2_2_weight_to_fp16 = const()[name = tensor("norm2_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(76772992)))]; + tensor norm2_2_bias_to_fp16 = const()[name = tensor("norm2_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(76775104)))]; + tensor var_1300_to_fp16 = const()[name = tensor("op_1300_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_29_cast_fp16 = layer_norm(axes = input_29_axes_0, beta = norm2_2_bias_to_fp16, epsilon = var_1300_to_fp16, gamma = norm2_2_weight_to_fp16, x = input_27_cast_fp16)[name = tensor("input_29_cast_fp16")]; + tensor linear_11_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = linear2_1_weight_cast_fp16, x = input_29_cast_fp16)[name = tensor("linear_11_cast_fp16")]; + tensor input_31_mode_0 = const()[name = tensor("input_31_mode_0"), val = tensor("EXACT")]; + tensor input_31_cast_fp16 = gelu(mode = input_31_mode_0, x = linear_11_cast_fp16)[name = tensor("input_31_cast_fp16")]; + tensor linear_12_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = linear2_2_weight_cast_fp16, x = input_31_cast_fp16)[name = tensor("linear_12_cast_fp16")]; + tensor input_33_cast_fp16 = add(x = input_27_cast_fp16, y = linear_12_cast_fp16)[name = tensor("input_33_cast_fp16")]; + tensor x_7_axes_0 = const()[name = tensor("x_7_axes_0"), val = tensor([-1])]; + tensor norm3_1_weight_to_fp16 = const()[name = tensor("norm3_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(76777216)))]; + tensor norm3_1_bias_to_fp16 = const()[name = tensor("norm3_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(76779328)))]; + tensor var_1317_to_fp16 = const()[name = tensor("op_1317_to_fp16"), val = tensor(0x1.5p-17)]; + tensor x_7_cast_fp16 = layer_norm(axes = x_7_axes_0, beta = norm3_1_bias_to_fp16, epsilon = var_1317_to_fp16, gamma = norm3_1_weight_to_fp16, x = input_33_cast_fp16)[name = tensor("x_7_cast_fp16")]; + tensor linear_13_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = attn3_in_proj_weight_cast_fp16, x = x_7_cast_fp16)[name = tensor("linear_13_cast_fp16")]; + tensor var_1353 = const()[name = tensor("op_1353"), val = tensor([1, 1, 3, 16, 64])]; + tensor qkv_7_cast_fp16 = reshape(shape = var_1353, x = linear_13_cast_fp16)[name = tensor("qkv_7_cast_fp16")]; + tensor q_19_begin_0 = const()[name = tensor("q_19_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor q_19_end_0 = const()[name = tensor("q_19_end_0"), val = tensor([1, 1, 1, 16, 64])]; + tensor q_19_end_mask_0 = const()[name = tensor("q_19_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor q_19_squeeze_mask_0 = const()[name = tensor("q_19_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor q_19_cast_fp16 = slice_by_index(begin = q_19_begin_0, end = q_19_end_0, end_mask = q_19_end_mask_0, squeeze_mask = q_19_squeeze_mask_0, x = qkv_7_cast_fp16)[name = tensor("q_19_cast_fp16")]; + tensor k_13_begin_0 = const()[name = tensor("k_13_begin_0"), val = tensor([0, 0, 1, 0, 0])]; + tensor k_13_end_0 = const()[name = tensor("k_13_end_0"), val = tensor([1, 1, 2, 16, 64])]; + tensor k_13_end_mask_0 = const()[name = tensor("k_13_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor k_13_squeeze_mask_0 = const()[name = tensor("k_13_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor k_13_cast_fp16 = slice_by_index(begin = k_13_begin_0, end = k_13_end_0, end_mask = k_13_end_mask_0, squeeze_mask = k_13_squeeze_mask_0, x = qkv_7_cast_fp16)[name = tensor("k_13_cast_fp16")]; + tensor v_7_begin_0 = const()[name = tensor("v_7_begin_0"), val = tensor([0, 0, 2, 0, 0])]; + tensor v_7_end_0 = const()[name = tensor("v_7_end_0"), val = tensor([1, 1, 3, 16, 64])]; + tensor v_7_end_mask_0 = const()[name = tensor("v_7_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor v_7_squeeze_mask_0 = const()[name = tensor("v_7_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor v_7_cast_fp16 = slice_by_index(begin = v_7_begin_0, end = v_7_end_0, end_mask = v_7_end_mask_0, squeeze_mask = v_7_squeeze_mask_0, x = qkv_7_cast_fp16)[name = tensor("v_7_cast_fp16")]; + tensor var_1457 = const()[name = tensor("op_1457"), val = tensor([1, 1, 1, 1])]; + tensor position3_to_fp16_dtype_0 = const()[name = tensor("position3_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor position3_to_fp16 = cast(dtype = position3_to_fp16_dtype_0, x = position3)[name = tensor("cast_124")]; + tensor ts_23_cast_fp16 = reshape(shape = var_1457, x = position3_to_fp16)[name = tensor("ts_23_cast_fp16")]; + tensor var_1461 = const()[name = tensor("op_1461"), val = tensor([1, 1, 16, 32, 2])]; + tensor q_complex_7_cast_fp16 = reshape(shape = var_1461, x = q_19_cast_fp16)[name = tensor("q_complex_7_cast_fp16")]; + tensor var_1465 = const()[name = tensor("op_1465"), val = tensor([1, 1, 16, 32, 2])]; + tensor k_complex_7_cast_fp16 = reshape(shape = var_1465, x = k_13_cast_fp16)[name = tensor("k_complex_7_cast_fp16")]; + tensor var_1469_begin_0 = const()[name = tensor("op_1469_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_1469_end_0 = const()[name = tensor("op_1469_end_0"), val = tensor([1, 1, 16, 32, 1])]; + tensor var_1469_end_mask_0 = const()[name = tensor("op_1469_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_1469_squeeze_mask_0 = const()[name = tensor("op_1469_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_1469_cast_fp16 = slice_by_index(begin = var_1469_begin_0, end = var_1469_end_0, end_mask = var_1469_end_mask_0, squeeze_mask = var_1469_squeeze_mask_0, x = q_complex_7_cast_fp16)[name = tensor("op_1469_cast_fp16")]; + tensor var_1477_begin_0 = const()[name = tensor("op_1477_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_1477_end_0 = const()[name = tensor("op_1477_end_0"), val = tensor([1, 1, 16, 32, 2])]; + tensor var_1477_end_mask_0 = const()[name = tensor("op_1477_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_1477_squeeze_mask_0 = const()[name = tensor("op_1477_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_1477_cast_fp16 = slice_by_index(begin = var_1477_begin_0, end = var_1477_end_0, end_mask = var_1477_end_mask_0, squeeze_mask = var_1477_squeeze_mask_0, x = q_complex_7_cast_fp16)[name = tensor("op_1477_cast_fp16")]; + tensor var_1485_begin_0 = const()[name = tensor("op_1485_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_1485_end_0 = const()[name = tensor("op_1485_end_0"), val = tensor([1, 1, 16, 32, 1])]; + tensor var_1485_end_mask_0 = const()[name = tensor("op_1485_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_1485_squeeze_mask_0 = const()[name = tensor("op_1485_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_1485_cast_fp16 = slice_by_index(begin = var_1485_begin_0, end = var_1485_end_0, end_mask = var_1485_end_mask_0, squeeze_mask = var_1485_squeeze_mask_0, x = k_complex_7_cast_fp16)[name = tensor("op_1485_cast_fp16")]; + tensor var_1493_begin_0 = const()[name = tensor("op_1493_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_1493_end_0 = const()[name = tensor("op_1493_end_0"), val = tensor([1, 1, 16, 32, 2])]; + tensor var_1493_end_mask_0 = const()[name = tensor("op_1493_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_1493_squeeze_mask_0 = const()[name = tensor("op_1493_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_1493_cast_fp16 = slice_by_index(begin = var_1493_begin_0, end = var_1493_end_0, end_mask = var_1493_end_mask_0, squeeze_mask = var_1493_squeeze_mask_0, x = k_complex_7_cast_fp16)[name = tensor("op_1493_cast_fp16")]; + tensor freqs_7_to_fp16 = const()[name = tensor("freqs_7_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(76781440)))]; + tensor var_1499_cast_fp16 = mul(x = freqs_7_to_fp16, y = ts_23_cast_fp16)[name = tensor("op_1499_cast_fp16")]; + tensor rotr_7_cast_fp16 = cos(x = var_1499_cast_fp16)[name = tensor("rotr_7_cast_fp16")]; + tensor roti_7_cast_fp16 = sin(x = var_1499_cast_fp16)[name = tensor("roti_7_cast_fp16")]; + tensor var_1503_cast_fp16 = mul(x = var_1469_cast_fp16, y = rotr_7_cast_fp16)[name = tensor("op_1503_cast_fp16")]; + tensor var_1504_cast_fp16 = mul(x = var_1477_cast_fp16, y = roti_7_cast_fp16)[name = tensor("op_1504_cast_fp16")]; + tensor qor_13_cast_fp16 = sub(x = var_1503_cast_fp16, y = var_1504_cast_fp16)[name = tensor("qor_13_cast_fp16")]; + tensor var_1507_cast_fp16 = mul(x = var_1469_cast_fp16, y = roti_7_cast_fp16)[name = tensor("op_1507_cast_fp16")]; + tensor var_1508_cast_fp16 = mul(x = var_1477_cast_fp16, y = rotr_7_cast_fp16)[name = tensor("op_1508_cast_fp16")]; + tensor qoi_13_cast_fp16 = add(x = var_1507_cast_fp16, y = var_1508_cast_fp16)[name = tensor("qoi_13_cast_fp16")]; + tensor var_1511_cast_fp16 = mul(x = var_1485_cast_fp16, y = rotr_7_cast_fp16)[name = tensor("op_1511_cast_fp16")]; + tensor var_1512_cast_fp16 = mul(x = var_1493_cast_fp16, y = roti_7_cast_fp16)[name = tensor("op_1512_cast_fp16")]; + tensor kor_13_cast_fp16 = sub(x = var_1511_cast_fp16, y = var_1512_cast_fp16)[name = tensor("kor_13_cast_fp16")]; + tensor var_1515_cast_fp16 = mul(x = var_1485_cast_fp16, y = roti_7_cast_fp16)[name = tensor("op_1515_cast_fp16")]; + tensor var_1516_cast_fp16 = mul(x = var_1493_cast_fp16, y = rotr_7_cast_fp16)[name = tensor("op_1516_cast_fp16")]; + tensor koi_13_cast_fp16 = add(x = var_1515_cast_fp16, y = var_1516_cast_fp16)[name = tensor("koi_13_cast_fp16")]; + tensor qo_7_axis_0 = const()[name = tensor("qo_7_axis_0"), val = tensor(-1)]; + tensor qo_7_cast_fp16 = stack(axis = qo_7_axis_0, values = (qor_13_cast_fp16, qoi_13_cast_fp16))[name = tensor("qo_7_cast_fp16")]; + tensor ko_7_axis_0 = const()[name = tensor("ko_7_axis_0"), val = tensor(-1)]; + tensor ko_7_cast_fp16 = stack(axis = ko_7_axis_0, values = (kor_13_cast_fp16, koi_13_cast_fp16))[name = tensor("ko_7_cast_fp16")]; + tensor var_1545 = const()[name = tensor("op_1545"), val = tensor([1, 1, 16, 64])]; + tensor q_21_cast_fp16 = reshape(shape = var_1545, x = qo_7_cast_fp16)[name = tensor("q_21_cast_fp16")]; + tensor var_1547 = const()[name = tensor("op_1547"), val = tensor([1, 1, 16, 64])]; + tensor k_15_cast_fp16 = reshape(shape = var_1547, x = ko_7_cast_fp16)[name = tensor("k_15_cast_fp16")]; + tensor _inversed_1569_y_0_to_fp16 = const()[name = tensor("_inversed_1569_y_0_to_fp16"), val = tensor(0x1p-9)]; + tensor _inversed_1569_cast_fp16 = mul(x = ts_23_cast_fp16, y = _inversed_1569_y_0_to_fp16)[name = tensor("_inversed_1569_cast_fp16")]; + tensor var_1570_cast_fp16 = floor(x = _inversed_1569_cast_fp16)[name = tensor("op_1570_cast_fp16")]; + tensor var_1571_to_fp16 = const()[name = tensor("op_1571_to_fp16"), val = tensor(0x1p+9)]; + tensor var_1572_cast_fp16 = mul(x = var_1570_cast_fp16, y = var_1571_to_fp16)[name = tensor("op_1572_cast_fp16")]; + tensor write_indices_float_15_cast_fp16 = sub(x = ts_23_cast_fp16, y = var_1572_cast_fp16)[name = tensor("write_indices_float_15_cast_fp16")]; + tensor var_1579_dtype_0 = const()[name = tensor("op_1579_dtype_0"), val = tensor("int32")]; + tensor write_indices_7_reps_0 = const()[name = tensor("write_indices_7_reps_0"), val = tensor([1, 1, 16, 64])]; + tensor write_indices_float_15_cast_fp16_to_int32 = cast(dtype = var_1579_dtype_0, x = write_indices_float_15_cast_fp16)[name = tensor("cast_123")]; + tensor write_indices_7 = tile(reps = write_indices_7_reps_0, x = write_indices_float_15_cast_fp16_to_int32)[name = tensor("write_indices_7")]; + tensor var_1587_begin_0 = const()[name = tensor("op_1587_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_1587_end_0 = const()[name = tensor("op_1587_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor var_1587_end_mask_0 = const()[name = tensor("op_1587_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_1587_squeeze_mask_0 = const()[name = tensor("op_1587_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor cache3_to_fp16_dtype_0 = const()[name = tensor("cache3_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor cache3_to_fp16 = cast(dtype = cache3_to_fp16_dtype_0, x = cache3)[name = tensor("cast_122")]; + tensor var_1587_cast_fp16 = slice_by_index(begin = var_1587_begin_0, end = var_1587_end_0, end_mask = var_1587_end_mask_0, squeeze_mask = var_1587_squeeze_mask_0, x = cache3_to_fp16)[name = tensor("op_1587_cast_fp16")]; + tensor var_1589_axis_0 = const()[name = tensor("op_1589_axis_0"), val = tensor(1)]; + tensor var_1589_mode_0 = const()[name = tensor("op_1589_mode_0"), val = tensor("update")]; + tensor var_1589_validate_indices_0 = const()[name = tensor("op_1589_validate_indices_0"), val = tensor(false)]; + tensor var_1589_cast_fp16 = scatter_along_axis(axis = var_1589_axis_0, data = var_1587_cast_fp16, indices = write_indices_7, mode = var_1589_mode_0, updates = k_15_cast_fp16, validate_indices = var_1589_validate_indices_0)[name = tensor("op_1589_cast_fp16")]; + tensor concat_23 = const()[name = tensor("concat_23"), val = tensor([0, 0, 0, 0, 0])]; + tensor concat_24 = const()[name = tensor("concat_24"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_7_internal_tensor_assign_1_stride_0 = const()[name = tensor("new_cache_7_internal_tensor_assign_1_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_7_internal_tensor_assign_1_begin_mask_0 = const()[name = tensor("new_cache_7_internal_tensor_assign_1_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_7_internal_tensor_assign_1_end_mask_0 = const()[name = tensor("new_cache_7_internal_tensor_assign_1_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_7_internal_tensor_assign_1_squeeze_mask_0 = const()[name = tensor("new_cache_7_internal_tensor_assign_1_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_18 = const()[name = tensor("shape_18"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_6 = const()[name = tensor("reduce_prod_6"), val = tensor(1048576)]; + tensor range_1d_6_start_0 = const()[name = tensor("range_1d_6_start_0"), val = tensor(0)]; + tensor range_1d_6_step_0 = const()[name = tensor("range_1d_6_step_0"), val = tensor(1)]; + tensor range_1d_6 = range_1d(end = reduce_prod_6, start = range_1d_6_start_0, step = range_1d_6_step_0)[name = tensor("range_1d_6")]; + tensor reshape_30 = reshape(shape = shape_18, x = range_1d_6)[name = tensor("reshape_30")]; + tensor slice_by_index_6 = slice_by_index(begin = concat_23, begin_mask = new_cache_7_internal_tensor_assign_1_begin_mask_0, end = concat_24, end_mask = new_cache_7_internal_tensor_assign_1_end_mask_0, squeeze_mask = new_cache_7_internal_tensor_assign_1_squeeze_mask_0, stride = new_cache_7_internal_tensor_assign_1_stride_0, x = reshape_30)[name = tensor("slice_by_index_6")]; + tensor reshape_31_shape_0 = const()[name = tensor("reshape_31_shape_0"), val = tensor([-1])]; + tensor reshape_31 = reshape(shape = reshape_31_shape_0, x = slice_by_index_6)[name = tensor("reshape_31")]; + tensor reshape_32_shape_0 = const()[name = tensor("reshape_32_shape_0"), val = tensor([-1])]; + tensor reshape_32_cast_fp16 = reshape(shape = reshape_32_shape_0, x = var_1589_cast_fp16)[name = tensor("reshape_32_cast_fp16")]; + tensor reshape_33_shape_0 = const()[name = tensor("reshape_33_shape_0"), val = tensor([-1])]; + tensor reshape_33_cast_fp16 = reshape(shape = reshape_33_shape_0, x = cache3_to_fp16)[name = tensor("reshape_33_cast_fp16")]; + tensor scatter_6_mode_0 = const()[name = tensor("scatter_6_mode_0"), val = tensor("update")]; + tensor scatter_6_axis_0 = const()[name = tensor("scatter_6_axis_0"), val = tensor(0)]; + tensor scatter_6_validate_indices_0 = const()[name = tensor("scatter_6_validate_indices_0"), val = tensor(false)]; + tensor scatter_6_cast_fp16 = scatter(axis = scatter_6_axis_0, data = reshape_33_cast_fp16, indices = reshape_31, mode = scatter_6_mode_0, updates = reshape_32_cast_fp16, validate_indices = scatter_6_validate_indices_0)[name = tensor("scatter_6_cast_fp16")]; + tensor reshape_34_cast_fp16 = reshape(shape = shape_18, x = scatter_6_cast_fp16)[name = tensor("reshape_34_cast_fp16")]; + tensor var_1597_begin_0 = const()[name = tensor("op_1597_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor var_1597_end_0 = const()[name = tensor("op_1597_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor var_1597_end_mask_0 = const()[name = tensor("op_1597_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_1597_squeeze_mask_0 = const()[name = tensor("op_1597_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor var_1597_cast_fp16 = slice_by_index(begin = var_1597_begin_0, end = var_1597_end_0, end_mask = var_1597_end_mask_0, squeeze_mask = var_1597_squeeze_mask_0, x = reshape_34_cast_fp16)[name = tensor("op_1597_cast_fp16")]; + tensor var_1599_axis_0 = const()[name = tensor("op_1599_axis_0"), val = tensor(1)]; + tensor var_1599_mode_0 = const()[name = tensor("op_1599_mode_0"), val = tensor("update")]; + tensor var_1599_validate_indices_0 = const()[name = tensor("op_1599_validate_indices_0"), val = tensor(false)]; + tensor var_1599_cast_fp16 = scatter_along_axis(axis = var_1599_axis_0, data = var_1597_cast_fp16, indices = write_indices_7, mode = var_1599_mode_0, updates = v_7_cast_fp16, validate_indices = var_1599_validate_indices_0)[name = tensor("op_1599_cast_fp16")]; + tensor concat_25 = const()[name = tensor("concat_25"), val = tensor([1, 0, 0, 0, 0])]; + tensor concat_26 = const()[name = tensor("concat_26"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_7_internal_tensor_assign_2_stride_0 = const()[name = tensor("new_cache_7_internal_tensor_assign_2_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_7_internal_tensor_assign_2_begin_mask_0 = const()[name = tensor("new_cache_7_internal_tensor_assign_2_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_7_internal_tensor_assign_2_end_mask_0 = const()[name = tensor("new_cache_7_internal_tensor_assign_2_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_7_internal_tensor_assign_2_squeeze_mask_0 = const()[name = tensor("new_cache_7_internal_tensor_assign_2_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_19 = const()[name = tensor("shape_19"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_7 = const()[name = tensor("reduce_prod_7"), val = tensor(1048576)]; + tensor range_1d_7_start_0 = const()[name = tensor("range_1d_7_start_0"), val = tensor(0)]; + tensor range_1d_7_step_0 = const()[name = tensor("range_1d_7_step_0"), val = tensor(1)]; + tensor range_1d_7 = range_1d(end = reduce_prod_7, start = range_1d_7_start_0, step = range_1d_7_step_0)[name = tensor("range_1d_7")]; + tensor reshape_35 = reshape(shape = shape_19, x = range_1d_7)[name = tensor("reshape_35")]; + tensor slice_by_index_7 = slice_by_index(begin = concat_25, begin_mask = new_cache_7_internal_tensor_assign_2_begin_mask_0, end = concat_26, end_mask = new_cache_7_internal_tensor_assign_2_end_mask_0, squeeze_mask = new_cache_7_internal_tensor_assign_2_squeeze_mask_0, stride = new_cache_7_internal_tensor_assign_2_stride_0, x = reshape_35)[name = tensor("slice_by_index_7")]; + tensor reshape_36_shape_0 = const()[name = tensor("reshape_36_shape_0"), val = tensor([-1])]; + tensor reshape_36 = reshape(shape = reshape_36_shape_0, x = slice_by_index_7)[name = tensor("reshape_36")]; + tensor reshape_37_shape_0 = const()[name = tensor("reshape_37_shape_0"), val = tensor([-1])]; + tensor reshape_37_cast_fp16 = reshape(shape = reshape_37_shape_0, x = var_1599_cast_fp16)[name = tensor("reshape_37_cast_fp16")]; + tensor reshape_38_shape_0 = const()[name = tensor("reshape_38_shape_0"), val = tensor([-1])]; + tensor reshape_38_cast_fp16 = reshape(shape = reshape_38_shape_0, x = reshape_34_cast_fp16)[name = tensor("reshape_38_cast_fp16")]; + tensor scatter_7_mode_0 = const()[name = tensor("scatter_7_mode_0"), val = tensor("update")]; + tensor scatter_7_axis_0 = const()[name = tensor("scatter_7_axis_0"), val = tensor(0)]; + tensor scatter_7_validate_indices_0 = const()[name = tensor("scatter_7_validate_indices_0"), val = tensor(false)]; + tensor scatter_7_cast_fp16 = scatter(axis = scatter_7_axis_0, data = reshape_38_cast_fp16, indices = reshape_36, mode = scatter_7_mode_0, updates = reshape_37_cast_fp16, validate_indices = scatter_7_validate_indices_0)[name = tensor("scatter_7_cast_fp16")]; + tensor reshape_39_cast_fp16 = reshape(shape = shape_19, x = scatter_7_cast_fp16)[name = tensor("reshape_39_cast_fp16")]; + tensor reshape_39_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("reshape_39_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor keys_19_begin_0 = const()[name = tensor("keys_19_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor keys_19_end_0 = const()[name = tensor("keys_19_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor keys_19_end_mask_0 = const()[name = tensor("keys_19_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor keys_19_squeeze_mask_0 = const()[name = tensor("keys_19_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor keys_19_cast_fp16 = slice_by_index(begin = keys_19_begin_0, end = keys_19_end_0, end_mask = keys_19_end_mask_0, squeeze_mask = keys_19_squeeze_mask_0, x = reshape_39_cast_fp16)[name = tensor("keys_19_cast_fp16")]; + tensor values_19_begin_0 = const()[name = tensor("values_19_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor values_19_end_0 = const()[name = tensor("values_19_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor values_19_end_mask_0 = const()[name = tensor("values_19_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor values_19_squeeze_mask_0 = const()[name = tensor("values_19_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor values_19_cast_fp16 = slice_by_index(begin = values_19_begin_0, end = values_19_end_0, end_mask = values_19_end_mask_0, squeeze_mask = values_19_squeeze_mask_0, x = reshape_39_cast_fp16)[name = tensor("values_19_cast_fp16")]; + tensor var_1611_cast_fp16 = not_equal(x = keys_19_cast_fp16, y = keys_19_cast_fp16)[name = tensor("op_1611_cast_fp16")]; + tensor keys_21_cast_fp16 = select(a = var_360_to_fp16, b = keys_19_cast_fp16, cond = var_1611_cast_fp16)[name = tensor("keys_21_cast_fp16")]; + tensor var_1619_cast_fp16 = not_equal(x = values_19_cast_fp16, y = values_19_cast_fp16)[name = tensor("op_1619_cast_fp16")]; + tensor values_21_cast_fp16 = select(a = var_360_to_fp16, b = values_19_cast_fp16, cond = var_1619_cast_fp16)[name = tensor("values_21_cast_fp16")]; + tensor var_1643 = const()[name = tensor("op_1643"), val = tensor([0, 2, 1, 3])]; + tensor var_1656 = const()[name = tensor("op_1656"), val = tensor([1, 1, 1])]; + tensor var_1657_cast_fp16 = reshape(shape = var_1656, x = position3_to_fp16)[name = tensor("op_1657_cast_fp16")]; + tensor var_1674_to_fp16 = const()[name = tensor("op_1674_to_fp16"), val = tensor(0x1p+0)]; + tensor valid_len_7_cast_fp16 = add(x = var_1657_cast_fp16, y = var_1674_to_fp16)[name = tensor("valid_len_7_cast_fp16")]; + tensor valid_mask_7_cast_fp16 = less(x = k_positions_1_promoted_to_fp16, y = valid_len_7_cast_fp16)[name = tensor("valid_mask_7_cast_fp16")]; + tensor causal_mask_7_cast_fp16 = less_equal(x = k_positions_1_promoted_to_fp16, y = var_1657_cast_fp16)[name = tensor("causal_mask_7_cast_fp16")]; + tensor attn_mask_13 = logical_and(x = valid_mask_7_cast_fp16, y = causal_mask_7_cast_fp16)[name = tensor("attn_mask_13")]; + tensor attn_mask_15_axes_0 = const()[name = tensor("attn_mask_15_axes_0"), val = tensor([1])]; + tensor attn_mask_15 = expand_dims(axes = attn_mask_15_axes_0, x = attn_mask_13)[name = tensor("attn_mask_15")]; + tensor var_1692_transpose_x_0 = const()[name = tensor("op_1692_transpose_x_0"), val = tensor(false)]; + tensor var_1692_transpose_y_0 = const()[name = tensor("op_1692_transpose_y_0"), val = tensor(false)]; + tensor transpose_24_perm_0 = const()[name = tensor("transpose_24_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_25_perm_0 = const()[name = tensor("transpose_25_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_25 = transpose(perm = transpose_25_perm_0, x = keys_21_cast_fp16)[name = tensor("transpose_39")]; + tensor transpose_24 = transpose(perm = transpose_24_perm_0, x = q_21_cast_fp16)[name = tensor("transpose_40")]; + tensor var_1692_cast_fp16 = matmul(transpose_x = var_1692_transpose_x_0, transpose_y = var_1692_transpose_y_0, x = transpose_24, y = transpose_25)[name = tensor("op_1692_cast_fp16")]; + tensor var_1686_to_fp16 = const()[name = tensor("op_1686_to_fp16"), val = tensor([0x1p-3])]; + tensor attn_weights_19_cast_fp16 = mul(x = var_1692_cast_fp16, y = var_1686_to_fp16)[name = tensor("attn_weights_19_cast_fp16")]; + tensor var_1694 = logical_not(x = attn_mask_15)[name = tensor("op_1694")]; + tensor var_1695_to_fp16 = const()[name = tensor("op_1695_to_fp16"), val = tensor(-inf)]; + tensor attn_weights_21_cast_fp16 = select(a = var_1695_to_fp16, b = attn_weights_19_cast_fp16, cond = var_1694)[name = tensor("attn_weights_21_cast_fp16")]; + tensor var_1697 = const()[name = tensor("op_1697"), val = tensor(-1)]; + tensor attn_weights_23_cast_fp16 = softmax(axis = var_1697, x = attn_weights_21_cast_fp16)[name = tensor("attn_weights_23_cast_fp16")]; + tensor attn_output_7_transpose_x_0 = const()[name = tensor("attn_output_7_transpose_x_0"), val = tensor(false)]; + tensor attn_output_7_transpose_y_0 = const()[name = tensor("attn_output_7_transpose_y_0"), val = tensor(false)]; + tensor values_23_cast_fp16 = transpose(perm = var_1643, x = values_21_cast_fp16)[name = tensor("transpose_41")]; + tensor attn_output_7_cast_fp16 = matmul(transpose_x = attn_output_7_transpose_x_0, transpose_y = attn_output_7_transpose_y_0, x = attn_weights_23_cast_fp16, y = values_23_cast_fp16)[name = tensor("attn_output_7_cast_fp16")]; + tensor var_1705 = const()[name = tensor("op_1705"), val = tensor([0, 2, 1, 3])]; + tensor var_1708 = const()[name = tensor("op_1708"), val = tensor([1, 1, 1024])]; + tensor var_1706_cast_fp16 = transpose(perm = var_1705, x = attn_output_7_cast_fp16)[name = tensor("transpose_38")]; + tensor input_35_cast_fp16 = reshape(shape = var_1708, x = var_1706_cast_fp16)[name = tensor("input_35_cast_fp16")]; + tensor linear_14_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = attn3_out_proj_weight_cast_fp16, x = input_35_cast_fp16)[name = tensor("linear_14_cast_fp16")]; + tensor var_1714_to_fp16 = const()[name = tensor("op_1714_to_fp16"), val = tensor(0x1p+0)]; + tensor var_1715_cast_fp16 = add(x = position3_to_fp16, y = var_1714_to_fp16)[name = tensor("op_1715_cast_fp16")]; + tensor var_1715_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_1715_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor input_37_cast_fp16 = add(x = input_33_cast_fp16, y = linear_14_cast_fp16)[name = tensor("input_37_cast_fp16")]; + tensor input_39_axes_0 = const()[name = tensor("input_39_axes_0"), val = tensor([-1])]; + tensor norm3_2_weight_to_fp16 = const()[name = tensor("norm3_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(76781568)))]; + tensor norm3_2_bias_to_fp16 = const()[name = tensor("norm3_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(76783680)))]; + tensor var_1719_to_fp16 = const()[name = tensor("op_1719_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_39_cast_fp16 = layer_norm(axes = input_39_axes_0, beta = norm3_2_bias_to_fp16, epsilon = var_1719_to_fp16, gamma = norm3_2_weight_to_fp16, x = input_37_cast_fp16)[name = tensor("input_39_cast_fp16")]; + tensor linear_15_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = linear3_1_weight_cast_fp16, x = input_39_cast_fp16)[name = tensor("linear_15_cast_fp16")]; + tensor input_41_mode_0 = const()[name = tensor("input_41_mode_0"), val = tensor("EXACT")]; + tensor input_41_cast_fp16 = gelu(mode = input_41_mode_0, x = linear_15_cast_fp16)[name = tensor("input_41_cast_fp16")]; + tensor linear_16_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = linear3_2_weight_cast_fp16, x = input_41_cast_fp16)[name = tensor("linear_16_cast_fp16")]; + tensor input_43_cast_fp16 = add(x = input_37_cast_fp16, y = linear_16_cast_fp16)[name = tensor("input_43_cast_fp16")]; + tensor x_9_axes_0 = const()[name = tensor("x_9_axes_0"), val = tensor([-1])]; + tensor norm4_1_weight_to_fp16 = const()[name = tensor("norm4_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(76785792)))]; + tensor norm4_1_bias_to_fp16 = const()[name = tensor("norm4_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(76787904)))]; + tensor var_1736_to_fp16 = const()[name = tensor("op_1736_to_fp16"), val = tensor(0x1.5p-17)]; + tensor x_9_cast_fp16 = layer_norm(axes = x_9_axes_0, beta = norm4_1_bias_to_fp16, epsilon = var_1736_to_fp16, gamma = norm4_1_weight_to_fp16, x = input_43_cast_fp16)[name = tensor("x_9_cast_fp16")]; + tensor linear_17_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = attn4_in_proj_weight_cast_fp16, x = x_9_cast_fp16)[name = tensor("linear_17_cast_fp16")]; + tensor var_1772 = const()[name = tensor("op_1772"), val = tensor([1, 1, 3, 16, 64])]; + tensor qkv_9_cast_fp16 = reshape(shape = var_1772, x = linear_17_cast_fp16)[name = tensor("qkv_9_cast_fp16")]; + tensor q_25_begin_0 = const()[name = tensor("q_25_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor q_25_end_0 = const()[name = tensor("q_25_end_0"), val = tensor([1, 1, 1, 16, 64])]; + tensor q_25_end_mask_0 = const()[name = tensor("q_25_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor q_25_squeeze_mask_0 = const()[name = tensor("q_25_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor q_25_cast_fp16 = slice_by_index(begin = q_25_begin_0, end = q_25_end_0, end_mask = q_25_end_mask_0, squeeze_mask = q_25_squeeze_mask_0, x = qkv_9_cast_fp16)[name = tensor("q_25_cast_fp16")]; + tensor k_17_begin_0 = const()[name = tensor("k_17_begin_0"), val = tensor([0, 0, 1, 0, 0])]; + tensor k_17_end_0 = const()[name = tensor("k_17_end_0"), val = tensor([1, 1, 2, 16, 64])]; + tensor k_17_end_mask_0 = const()[name = tensor("k_17_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor k_17_squeeze_mask_0 = const()[name = tensor("k_17_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor k_17_cast_fp16 = slice_by_index(begin = k_17_begin_0, end = k_17_end_0, end_mask = k_17_end_mask_0, squeeze_mask = k_17_squeeze_mask_0, x = qkv_9_cast_fp16)[name = tensor("k_17_cast_fp16")]; + tensor v_9_begin_0 = const()[name = tensor("v_9_begin_0"), val = tensor([0, 0, 2, 0, 0])]; + tensor v_9_end_0 = const()[name = tensor("v_9_end_0"), val = tensor([1, 1, 3, 16, 64])]; + tensor v_9_end_mask_0 = const()[name = tensor("v_9_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor v_9_squeeze_mask_0 = const()[name = tensor("v_9_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor v_9_cast_fp16 = slice_by_index(begin = v_9_begin_0, end = v_9_end_0, end_mask = v_9_end_mask_0, squeeze_mask = v_9_squeeze_mask_0, x = qkv_9_cast_fp16)[name = tensor("v_9_cast_fp16")]; + tensor var_1876 = const()[name = tensor("op_1876"), val = tensor([1, 1, 1, 1])]; + tensor position4_to_fp16_dtype_0 = const()[name = tensor("position4_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor position4_to_fp16 = cast(dtype = position4_to_fp16_dtype_0, x = position4)[name = tensor("cast_119")]; + tensor ts_29_cast_fp16 = reshape(shape = var_1876, x = position4_to_fp16)[name = tensor("ts_29_cast_fp16")]; + tensor var_1880 = const()[name = tensor("op_1880"), val = tensor([1, 1, 16, 32, 2])]; + tensor q_complex_9_cast_fp16 = reshape(shape = var_1880, x = q_25_cast_fp16)[name = tensor("q_complex_9_cast_fp16")]; + tensor var_1884 = const()[name = tensor("op_1884"), val = tensor([1, 1, 16, 32, 2])]; + tensor k_complex_9_cast_fp16 = reshape(shape = var_1884, x = k_17_cast_fp16)[name = tensor("k_complex_9_cast_fp16")]; + tensor var_1888_begin_0 = const()[name = tensor("op_1888_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_1888_end_0 = const()[name = tensor("op_1888_end_0"), val = tensor([1, 1, 16, 32, 1])]; + tensor var_1888_end_mask_0 = const()[name = tensor("op_1888_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_1888_squeeze_mask_0 = const()[name = tensor("op_1888_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_1888_cast_fp16 = slice_by_index(begin = var_1888_begin_0, end = var_1888_end_0, end_mask = var_1888_end_mask_0, squeeze_mask = var_1888_squeeze_mask_0, x = q_complex_9_cast_fp16)[name = tensor("op_1888_cast_fp16")]; + tensor var_1896_begin_0 = const()[name = tensor("op_1896_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_1896_end_0 = const()[name = tensor("op_1896_end_0"), val = tensor([1, 1, 16, 32, 2])]; + tensor var_1896_end_mask_0 = const()[name = tensor("op_1896_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_1896_squeeze_mask_0 = const()[name = tensor("op_1896_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_1896_cast_fp16 = slice_by_index(begin = var_1896_begin_0, end = var_1896_end_0, end_mask = var_1896_end_mask_0, squeeze_mask = var_1896_squeeze_mask_0, x = q_complex_9_cast_fp16)[name = tensor("op_1896_cast_fp16")]; + tensor var_1904_begin_0 = const()[name = tensor("op_1904_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_1904_end_0 = const()[name = tensor("op_1904_end_0"), val = tensor([1, 1, 16, 32, 1])]; + tensor var_1904_end_mask_0 = const()[name = tensor("op_1904_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_1904_squeeze_mask_0 = const()[name = tensor("op_1904_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_1904_cast_fp16 = slice_by_index(begin = var_1904_begin_0, end = var_1904_end_0, end_mask = var_1904_end_mask_0, squeeze_mask = var_1904_squeeze_mask_0, x = k_complex_9_cast_fp16)[name = tensor("op_1904_cast_fp16")]; + tensor var_1912_begin_0 = const()[name = tensor("op_1912_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_1912_end_0 = const()[name = tensor("op_1912_end_0"), val = tensor([1, 1, 16, 32, 2])]; + tensor var_1912_end_mask_0 = const()[name = tensor("op_1912_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_1912_squeeze_mask_0 = const()[name = tensor("op_1912_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_1912_cast_fp16 = slice_by_index(begin = var_1912_begin_0, end = var_1912_end_0, end_mask = var_1912_end_mask_0, squeeze_mask = var_1912_squeeze_mask_0, x = k_complex_9_cast_fp16)[name = tensor("op_1912_cast_fp16")]; + tensor freqs_9_to_fp16 = const()[name = tensor("freqs_9_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(76790016)))]; + tensor var_1918_cast_fp16 = mul(x = freqs_9_to_fp16, y = ts_29_cast_fp16)[name = tensor("op_1918_cast_fp16")]; + tensor rotr_9_cast_fp16 = cos(x = var_1918_cast_fp16)[name = tensor("rotr_9_cast_fp16")]; + tensor roti_9_cast_fp16 = sin(x = var_1918_cast_fp16)[name = tensor("roti_9_cast_fp16")]; + tensor var_1922_cast_fp16 = mul(x = var_1888_cast_fp16, y = rotr_9_cast_fp16)[name = tensor("op_1922_cast_fp16")]; + tensor var_1923_cast_fp16 = mul(x = var_1896_cast_fp16, y = roti_9_cast_fp16)[name = tensor("op_1923_cast_fp16")]; + tensor qor_17_cast_fp16 = sub(x = var_1922_cast_fp16, y = var_1923_cast_fp16)[name = tensor("qor_17_cast_fp16")]; + tensor var_1926_cast_fp16 = mul(x = var_1888_cast_fp16, y = roti_9_cast_fp16)[name = tensor("op_1926_cast_fp16")]; + tensor var_1927_cast_fp16 = mul(x = var_1896_cast_fp16, y = rotr_9_cast_fp16)[name = tensor("op_1927_cast_fp16")]; + tensor qoi_17_cast_fp16 = add(x = var_1926_cast_fp16, y = var_1927_cast_fp16)[name = tensor("qoi_17_cast_fp16")]; + tensor var_1930_cast_fp16 = mul(x = var_1904_cast_fp16, y = rotr_9_cast_fp16)[name = tensor("op_1930_cast_fp16")]; + tensor var_1931_cast_fp16 = mul(x = var_1912_cast_fp16, y = roti_9_cast_fp16)[name = tensor("op_1931_cast_fp16")]; + tensor kor_17_cast_fp16 = sub(x = var_1930_cast_fp16, y = var_1931_cast_fp16)[name = tensor("kor_17_cast_fp16")]; + tensor var_1934_cast_fp16 = mul(x = var_1904_cast_fp16, y = roti_9_cast_fp16)[name = tensor("op_1934_cast_fp16")]; + tensor var_1935_cast_fp16 = mul(x = var_1912_cast_fp16, y = rotr_9_cast_fp16)[name = tensor("op_1935_cast_fp16")]; + tensor koi_17_cast_fp16 = add(x = var_1934_cast_fp16, y = var_1935_cast_fp16)[name = tensor("koi_17_cast_fp16")]; + tensor qo_9_axis_0 = const()[name = tensor("qo_9_axis_0"), val = tensor(-1)]; + tensor qo_9_cast_fp16 = stack(axis = qo_9_axis_0, values = (qor_17_cast_fp16, qoi_17_cast_fp16))[name = tensor("qo_9_cast_fp16")]; + tensor ko_9_axis_0 = const()[name = tensor("ko_9_axis_0"), val = tensor(-1)]; + tensor ko_9_cast_fp16 = stack(axis = ko_9_axis_0, values = (kor_17_cast_fp16, koi_17_cast_fp16))[name = tensor("ko_9_cast_fp16")]; + tensor var_1964 = const()[name = tensor("op_1964"), val = tensor([1, 1, 16, 64])]; + tensor q_27_cast_fp16 = reshape(shape = var_1964, x = qo_9_cast_fp16)[name = tensor("q_27_cast_fp16")]; + tensor var_1966 = const()[name = tensor("op_1966"), val = tensor([1, 1, 16, 64])]; + tensor k_19_cast_fp16 = reshape(shape = var_1966, x = ko_9_cast_fp16)[name = tensor("k_19_cast_fp16")]; + tensor _inversed_1988_y_0_to_fp16 = const()[name = tensor("_inversed_1988_y_0_to_fp16"), val = tensor(0x1p-9)]; + tensor _inversed_1988_cast_fp16 = mul(x = ts_29_cast_fp16, y = _inversed_1988_y_0_to_fp16)[name = tensor("_inversed_1988_cast_fp16")]; + tensor var_1989_cast_fp16 = floor(x = _inversed_1988_cast_fp16)[name = tensor("op_1989_cast_fp16")]; + tensor var_1990_to_fp16 = const()[name = tensor("op_1990_to_fp16"), val = tensor(0x1p+9)]; + tensor var_1991_cast_fp16 = mul(x = var_1989_cast_fp16, y = var_1990_to_fp16)[name = tensor("op_1991_cast_fp16")]; + tensor write_indices_float_19_cast_fp16 = sub(x = ts_29_cast_fp16, y = var_1991_cast_fp16)[name = tensor("write_indices_float_19_cast_fp16")]; + tensor var_1998_dtype_0 = const()[name = tensor("op_1998_dtype_0"), val = tensor("int32")]; + tensor write_indices_9_reps_0 = const()[name = tensor("write_indices_9_reps_0"), val = tensor([1, 1, 16, 64])]; + tensor write_indices_float_19_cast_fp16_to_int32 = cast(dtype = var_1998_dtype_0, x = write_indices_float_19_cast_fp16)[name = tensor("cast_118")]; + tensor write_indices_9 = tile(reps = write_indices_9_reps_0, x = write_indices_float_19_cast_fp16_to_int32)[name = tensor("write_indices_9")]; + tensor var_2006_begin_0 = const()[name = tensor("op_2006_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_2006_end_0 = const()[name = tensor("op_2006_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor var_2006_end_mask_0 = const()[name = tensor("op_2006_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_2006_squeeze_mask_0 = const()[name = tensor("op_2006_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor cache4_to_fp16_dtype_0 = const()[name = tensor("cache4_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor cache4_to_fp16 = cast(dtype = cache4_to_fp16_dtype_0, x = cache4)[name = tensor("cast_117")]; + tensor var_2006_cast_fp16 = slice_by_index(begin = var_2006_begin_0, end = var_2006_end_0, end_mask = var_2006_end_mask_0, squeeze_mask = var_2006_squeeze_mask_0, x = cache4_to_fp16)[name = tensor("op_2006_cast_fp16")]; + tensor var_2008_axis_0 = const()[name = tensor("op_2008_axis_0"), val = tensor(1)]; + tensor var_2008_mode_0 = const()[name = tensor("op_2008_mode_0"), val = tensor("update")]; + tensor var_2008_validate_indices_0 = const()[name = tensor("op_2008_validate_indices_0"), val = tensor(false)]; + tensor var_2008_cast_fp16 = scatter_along_axis(axis = var_2008_axis_0, data = var_2006_cast_fp16, indices = write_indices_9, mode = var_2008_mode_0, updates = k_19_cast_fp16, validate_indices = var_2008_validate_indices_0)[name = tensor("op_2008_cast_fp16")]; + tensor concat_30 = const()[name = tensor("concat_30"), val = tensor([0, 0, 0, 0, 0])]; + tensor concat_31 = const()[name = tensor("concat_31"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_9_internal_tensor_assign_1_stride_0 = const()[name = tensor("new_cache_9_internal_tensor_assign_1_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_9_internal_tensor_assign_1_begin_mask_0 = const()[name = tensor("new_cache_9_internal_tensor_assign_1_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_9_internal_tensor_assign_1_end_mask_0 = const()[name = tensor("new_cache_9_internal_tensor_assign_1_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_9_internal_tensor_assign_1_squeeze_mask_0 = const()[name = tensor("new_cache_9_internal_tensor_assign_1_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_20 = const()[name = tensor("shape_20"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_8 = const()[name = tensor("reduce_prod_8"), val = tensor(1048576)]; + tensor range_1d_8_start_0 = const()[name = tensor("range_1d_8_start_0"), val = tensor(0)]; + tensor range_1d_8_step_0 = const()[name = tensor("range_1d_8_step_0"), val = tensor(1)]; + tensor range_1d_8 = range_1d(end = reduce_prod_8, start = range_1d_8_start_0, step = range_1d_8_step_0)[name = tensor("range_1d_8")]; + tensor reshape_40 = reshape(shape = shape_20, x = range_1d_8)[name = tensor("reshape_40")]; + tensor slice_by_index_8 = slice_by_index(begin = concat_30, begin_mask = new_cache_9_internal_tensor_assign_1_begin_mask_0, end = concat_31, end_mask = new_cache_9_internal_tensor_assign_1_end_mask_0, squeeze_mask = new_cache_9_internal_tensor_assign_1_squeeze_mask_0, stride = new_cache_9_internal_tensor_assign_1_stride_0, x = reshape_40)[name = tensor("slice_by_index_8")]; + tensor reshape_41_shape_0 = const()[name = tensor("reshape_41_shape_0"), val = tensor([-1])]; + tensor reshape_41 = reshape(shape = reshape_41_shape_0, x = slice_by_index_8)[name = tensor("reshape_41")]; + tensor reshape_42_shape_0 = const()[name = tensor("reshape_42_shape_0"), val = tensor([-1])]; + tensor reshape_42_cast_fp16 = reshape(shape = reshape_42_shape_0, x = var_2008_cast_fp16)[name = tensor("reshape_42_cast_fp16")]; + tensor reshape_43_shape_0 = const()[name = tensor("reshape_43_shape_0"), val = tensor([-1])]; + tensor reshape_43_cast_fp16 = reshape(shape = reshape_43_shape_0, x = cache4_to_fp16)[name = tensor("reshape_43_cast_fp16")]; + tensor scatter_8_mode_0 = const()[name = tensor("scatter_8_mode_0"), val = tensor("update")]; + tensor scatter_8_axis_0 = const()[name = tensor("scatter_8_axis_0"), val = tensor(0)]; + tensor scatter_8_validate_indices_0 = const()[name = tensor("scatter_8_validate_indices_0"), val = tensor(false)]; + tensor scatter_8_cast_fp16 = scatter(axis = scatter_8_axis_0, data = reshape_43_cast_fp16, indices = reshape_41, mode = scatter_8_mode_0, updates = reshape_42_cast_fp16, validate_indices = scatter_8_validate_indices_0)[name = tensor("scatter_8_cast_fp16")]; + tensor reshape_44_cast_fp16 = reshape(shape = shape_20, x = scatter_8_cast_fp16)[name = tensor("reshape_44_cast_fp16")]; + tensor var_2016_begin_0 = const()[name = tensor("op_2016_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor var_2016_end_0 = const()[name = tensor("op_2016_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor var_2016_end_mask_0 = const()[name = tensor("op_2016_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_2016_squeeze_mask_0 = const()[name = tensor("op_2016_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor var_2016_cast_fp16 = slice_by_index(begin = var_2016_begin_0, end = var_2016_end_0, end_mask = var_2016_end_mask_0, squeeze_mask = var_2016_squeeze_mask_0, x = reshape_44_cast_fp16)[name = tensor("op_2016_cast_fp16")]; + tensor var_2018_axis_0 = const()[name = tensor("op_2018_axis_0"), val = tensor(1)]; + tensor var_2018_mode_0 = const()[name = tensor("op_2018_mode_0"), val = tensor("update")]; + tensor var_2018_validate_indices_0 = const()[name = tensor("op_2018_validate_indices_0"), val = tensor(false)]; + tensor var_2018_cast_fp16 = scatter_along_axis(axis = var_2018_axis_0, data = var_2016_cast_fp16, indices = write_indices_9, mode = var_2018_mode_0, updates = v_9_cast_fp16, validate_indices = var_2018_validate_indices_0)[name = tensor("op_2018_cast_fp16")]; + tensor concat_32 = const()[name = tensor("concat_32"), val = tensor([1, 0, 0, 0, 0])]; + tensor concat_33 = const()[name = tensor("concat_33"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_9_internal_tensor_assign_2_stride_0 = const()[name = tensor("new_cache_9_internal_tensor_assign_2_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_9_internal_tensor_assign_2_begin_mask_0 = const()[name = tensor("new_cache_9_internal_tensor_assign_2_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_9_internal_tensor_assign_2_end_mask_0 = const()[name = tensor("new_cache_9_internal_tensor_assign_2_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_9_internal_tensor_assign_2_squeeze_mask_0 = const()[name = tensor("new_cache_9_internal_tensor_assign_2_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_21 = const()[name = tensor("shape_21"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_9 = const()[name = tensor("reduce_prod_9"), val = tensor(1048576)]; + tensor range_1d_9_start_0 = const()[name = tensor("range_1d_9_start_0"), val = tensor(0)]; + tensor range_1d_9_step_0 = const()[name = tensor("range_1d_9_step_0"), val = tensor(1)]; + tensor range_1d_9 = range_1d(end = reduce_prod_9, start = range_1d_9_start_0, step = range_1d_9_step_0)[name = tensor("range_1d_9")]; + tensor reshape_45 = reshape(shape = shape_21, x = range_1d_9)[name = tensor("reshape_45")]; + tensor slice_by_index_9 = slice_by_index(begin = concat_32, begin_mask = new_cache_9_internal_tensor_assign_2_begin_mask_0, end = concat_33, end_mask = new_cache_9_internal_tensor_assign_2_end_mask_0, squeeze_mask = new_cache_9_internal_tensor_assign_2_squeeze_mask_0, stride = new_cache_9_internal_tensor_assign_2_stride_0, x = reshape_45)[name = tensor("slice_by_index_9")]; + tensor reshape_46_shape_0 = const()[name = tensor("reshape_46_shape_0"), val = tensor([-1])]; + tensor reshape_46 = reshape(shape = reshape_46_shape_0, x = slice_by_index_9)[name = tensor("reshape_46")]; + tensor reshape_47_shape_0 = const()[name = tensor("reshape_47_shape_0"), val = tensor([-1])]; + tensor reshape_47_cast_fp16 = reshape(shape = reshape_47_shape_0, x = var_2018_cast_fp16)[name = tensor("reshape_47_cast_fp16")]; + tensor reshape_48_shape_0 = const()[name = tensor("reshape_48_shape_0"), val = tensor([-1])]; + tensor reshape_48_cast_fp16 = reshape(shape = reshape_48_shape_0, x = reshape_44_cast_fp16)[name = tensor("reshape_48_cast_fp16")]; + tensor scatter_9_mode_0 = const()[name = tensor("scatter_9_mode_0"), val = tensor("update")]; + tensor scatter_9_axis_0 = const()[name = tensor("scatter_9_axis_0"), val = tensor(0)]; + tensor scatter_9_validate_indices_0 = const()[name = tensor("scatter_9_validate_indices_0"), val = tensor(false)]; + tensor scatter_9_cast_fp16 = scatter(axis = scatter_9_axis_0, data = reshape_48_cast_fp16, indices = reshape_46, mode = scatter_9_mode_0, updates = reshape_47_cast_fp16, validate_indices = scatter_9_validate_indices_0)[name = tensor("scatter_9_cast_fp16")]; + tensor reshape_49_cast_fp16 = reshape(shape = shape_21, x = scatter_9_cast_fp16)[name = tensor("reshape_49_cast_fp16")]; + tensor reshape_49_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("reshape_49_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor keys_25_begin_0 = const()[name = tensor("keys_25_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor keys_25_end_0 = const()[name = tensor("keys_25_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor keys_25_end_mask_0 = const()[name = tensor("keys_25_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor keys_25_squeeze_mask_0 = const()[name = tensor("keys_25_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor keys_25_cast_fp16 = slice_by_index(begin = keys_25_begin_0, end = keys_25_end_0, end_mask = keys_25_end_mask_0, squeeze_mask = keys_25_squeeze_mask_0, x = reshape_49_cast_fp16)[name = tensor("keys_25_cast_fp16")]; + tensor values_25_begin_0 = const()[name = tensor("values_25_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor values_25_end_0 = const()[name = tensor("values_25_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor values_25_end_mask_0 = const()[name = tensor("values_25_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor values_25_squeeze_mask_0 = const()[name = tensor("values_25_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor values_25_cast_fp16 = slice_by_index(begin = values_25_begin_0, end = values_25_end_0, end_mask = values_25_end_mask_0, squeeze_mask = values_25_squeeze_mask_0, x = reshape_49_cast_fp16)[name = tensor("values_25_cast_fp16")]; + tensor var_2030_cast_fp16 = not_equal(x = keys_25_cast_fp16, y = keys_25_cast_fp16)[name = tensor("op_2030_cast_fp16")]; + tensor keys_27_cast_fp16 = select(a = var_360_to_fp16, b = keys_25_cast_fp16, cond = var_2030_cast_fp16)[name = tensor("keys_27_cast_fp16")]; + tensor var_2038_cast_fp16 = not_equal(x = values_25_cast_fp16, y = values_25_cast_fp16)[name = tensor("op_2038_cast_fp16")]; + tensor values_27_cast_fp16 = select(a = var_360_to_fp16, b = values_25_cast_fp16, cond = var_2038_cast_fp16)[name = tensor("values_27_cast_fp16")]; + tensor var_2062 = const()[name = tensor("op_2062"), val = tensor([0, 2, 1, 3])]; + tensor var_2075 = const()[name = tensor("op_2075"), val = tensor([1, 1, 1])]; + tensor var_2076_cast_fp16 = reshape(shape = var_2075, x = position4_to_fp16)[name = tensor("op_2076_cast_fp16")]; + tensor var_2093_to_fp16 = const()[name = tensor("op_2093_to_fp16"), val = tensor(0x1p+0)]; + tensor valid_len_9_cast_fp16 = add(x = var_2076_cast_fp16, y = var_2093_to_fp16)[name = tensor("valid_len_9_cast_fp16")]; + tensor valid_mask_9_cast_fp16 = less(x = k_positions_1_promoted_to_fp16, y = valid_len_9_cast_fp16)[name = tensor("valid_mask_9_cast_fp16")]; + tensor causal_mask_9_cast_fp16 = less_equal(x = k_positions_1_promoted_to_fp16, y = var_2076_cast_fp16)[name = tensor("causal_mask_9_cast_fp16")]; + tensor attn_mask_17 = logical_and(x = valid_mask_9_cast_fp16, y = causal_mask_9_cast_fp16)[name = tensor("attn_mask_17")]; + tensor attn_mask_19_axes_0 = const()[name = tensor("attn_mask_19_axes_0"), val = tensor([1])]; + tensor attn_mask_19 = expand_dims(axes = attn_mask_19_axes_0, x = attn_mask_17)[name = tensor("attn_mask_19")]; + tensor var_2111_transpose_x_0 = const()[name = tensor("op_2111_transpose_x_0"), val = tensor(false)]; + tensor var_2111_transpose_y_0 = const()[name = tensor("op_2111_transpose_y_0"), val = tensor(false)]; + tensor transpose_26_perm_0 = const()[name = tensor("transpose_26_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_27_perm_0 = const()[name = tensor("transpose_27_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_27 = transpose(perm = transpose_27_perm_0, x = keys_27_cast_fp16)[name = tensor("transpose_35")]; + tensor transpose_26 = transpose(perm = transpose_26_perm_0, x = q_27_cast_fp16)[name = tensor("transpose_36")]; + tensor var_2111_cast_fp16 = matmul(transpose_x = var_2111_transpose_x_0, transpose_y = var_2111_transpose_y_0, x = transpose_26, y = transpose_27)[name = tensor("op_2111_cast_fp16")]; + tensor var_2105_to_fp16 = const()[name = tensor("op_2105_to_fp16"), val = tensor([0x1p-3])]; + tensor attn_weights_25_cast_fp16 = mul(x = var_2111_cast_fp16, y = var_2105_to_fp16)[name = tensor("attn_weights_25_cast_fp16")]; + tensor var_2113 = logical_not(x = attn_mask_19)[name = tensor("op_2113")]; + tensor var_2114_to_fp16 = const()[name = tensor("op_2114_to_fp16"), val = tensor(-inf)]; + tensor attn_weights_27_cast_fp16 = select(a = var_2114_to_fp16, b = attn_weights_25_cast_fp16, cond = var_2113)[name = tensor("attn_weights_27_cast_fp16")]; + tensor var_2116 = const()[name = tensor("op_2116"), val = tensor(-1)]; + tensor attn_weights_29_cast_fp16 = softmax(axis = var_2116, x = attn_weights_27_cast_fp16)[name = tensor("attn_weights_29_cast_fp16")]; + tensor attn_output_9_transpose_x_0 = const()[name = tensor("attn_output_9_transpose_x_0"), val = tensor(false)]; + tensor attn_output_9_transpose_y_0 = const()[name = tensor("attn_output_9_transpose_y_0"), val = tensor(false)]; + tensor values_29_cast_fp16 = transpose(perm = var_2062, x = values_27_cast_fp16)[name = tensor("transpose_37")]; + tensor attn_output_9_cast_fp16 = matmul(transpose_x = attn_output_9_transpose_x_0, transpose_y = attn_output_9_transpose_y_0, x = attn_weights_29_cast_fp16, y = values_29_cast_fp16)[name = tensor("attn_output_9_cast_fp16")]; + tensor var_2124 = const()[name = tensor("op_2124"), val = tensor([0, 2, 1, 3])]; + tensor var_2127 = const()[name = tensor("op_2127"), val = tensor([1, 1, 1024])]; + tensor var_2125_cast_fp16 = transpose(perm = var_2124, x = attn_output_9_cast_fp16)[name = tensor("transpose_34")]; + tensor input_45_cast_fp16 = reshape(shape = var_2127, x = var_2125_cast_fp16)[name = tensor("input_45_cast_fp16")]; + tensor linear_18_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = attn4_out_proj_weight_cast_fp16, x = input_45_cast_fp16)[name = tensor("linear_18_cast_fp16")]; + tensor var_2133_to_fp16 = const()[name = tensor("op_2133_to_fp16"), val = tensor(0x1p+0)]; + tensor var_2134_cast_fp16 = add(x = position4_to_fp16, y = var_2133_to_fp16)[name = tensor("op_2134_cast_fp16")]; + tensor var_2134_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_2134_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor input_47_cast_fp16 = add(x = input_43_cast_fp16, y = linear_18_cast_fp16)[name = tensor("input_47_cast_fp16")]; + tensor input_49_axes_0 = const()[name = tensor("input_49_axes_0"), val = tensor([-1])]; + tensor norm4_2_weight_to_fp16 = const()[name = tensor("norm4_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(76790144)))]; + tensor norm4_2_bias_to_fp16 = const()[name = tensor("norm4_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(76792256)))]; + tensor var_2138_to_fp16 = const()[name = tensor("op_2138_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_49_cast_fp16 = layer_norm(axes = input_49_axes_0, beta = norm4_2_bias_to_fp16, epsilon = var_2138_to_fp16, gamma = norm4_2_weight_to_fp16, x = input_47_cast_fp16)[name = tensor("input_49_cast_fp16")]; + tensor linear_19_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = linear4_1_weight_cast_fp16, x = input_49_cast_fp16)[name = tensor("linear_19_cast_fp16")]; + tensor input_51_mode_0 = const()[name = tensor("input_51_mode_0"), val = tensor("EXACT")]; + tensor input_51_cast_fp16 = gelu(mode = input_51_mode_0, x = linear_19_cast_fp16)[name = tensor("input_51_cast_fp16")]; + tensor linear_20_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = linear4_2_weight_cast_fp16, x = input_51_cast_fp16)[name = tensor("linear_20_cast_fp16")]; + tensor input_53_cast_fp16 = add(x = input_47_cast_fp16, y = linear_20_cast_fp16)[name = tensor("input_53_cast_fp16")]; + tensor x_axes_0 = const()[name = tensor("x_axes_0"), val = tensor([-1])]; + tensor norm5_1_weight_to_fp16 = const()[name = tensor("norm5_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(76794368)))]; + tensor norm5_1_bias_to_fp16 = const()[name = tensor("norm5_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(76796480)))]; + tensor var_2155_to_fp16 = const()[name = tensor("op_2155_to_fp16"), val = tensor(0x1.5p-17)]; + tensor x_cast_fp16 = layer_norm(axes = x_axes_0, beta = norm5_1_bias_to_fp16, epsilon = var_2155_to_fp16, gamma = norm5_1_weight_to_fp16, x = input_53_cast_fp16)[name = tensor("x_cast_fp16")]; + tensor linear_21_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = attn5_in_proj_weight_cast_fp16, x = x_cast_fp16)[name = tensor("linear_21_cast_fp16")]; + tensor var_2191 = const()[name = tensor("op_2191"), val = tensor([1, 1, 3, 16, 64])]; + tensor qkv_cast_fp16 = reshape(shape = var_2191, x = linear_21_cast_fp16)[name = tensor("qkv_cast_fp16")]; + tensor q_31_begin_0 = const()[name = tensor("q_31_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor q_31_end_0 = const()[name = tensor("q_31_end_0"), val = tensor([1, 1, 1, 16, 64])]; + tensor q_31_end_mask_0 = const()[name = tensor("q_31_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor q_31_squeeze_mask_0 = const()[name = tensor("q_31_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor q_31_cast_fp16 = slice_by_index(begin = q_31_begin_0, end = q_31_end_0, end_mask = q_31_end_mask_0, squeeze_mask = q_31_squeeze_mask_0, x = qkv_cast_fp16)[name = tensor("q_31_cast_fp16")]; + tensor k_21_begin_0 = const()[name = tensor("k_21_begin_0"), val = tensor([0, 0, 1, 0, 0])]; + tensor k_21_end_0 = const()[name = tensor("k_21_end_0"), val = tensor([1, 1, 2, 16, 64])]; + tensor k_21_end_mask_0 = const()[name = tensor("k_21_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor k_21_squeeze_mask_0 = const()[name = tensor("k_21_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor k_21_cast_fp16 = slice_by_index(begin = k_21_begin_0, end = k_21_end_0, end_mask = k_21_end_mask_0, squeeze_mask = k_21_squeeze_mask_0, x = qkv_cast_fp16)[name = tensor("k_21_cast_fp16")]; + tensor v_begin_0 = const()[name = tensor("v_begin_0"), val = tensor([0, 0, 2, 0, 0])]; + tensor v_end_0 = const()[name = tensor("v_end_0"), val = tensor([1, 1, 3, 16, 64])]; + tensor v_end_mask_0 = const()[name = tensor("v_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor v_squeeze_mask_0 = const()[name = tensor("v_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor v_cast_fp16 = slice_by_index(begin = v_begin_0, end = v_end_0, end_mask = v_end_mask_0, squeeze_mask = v_squeeze_mask_0, x = qkv_cast_fp16)[name = tensor("v_cast_fp16")]; + tensor var_2295 = const()[name = tensor("op_2295"), val = tensor([1, 1, 1, 1])]; + tensor position5_to_fp16_dtype_0 = const()[name = tensor("position5_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor position5_to_fp16 = cast(dtype = position5_to_fp16_dtype_0, x = position5)[name = tensor("cast_114")]; + tensor ts_cast_fp16 = reshape(shape = var_2295, x = position5_to_fp16)[name = tensor("ts_cast_fp16")]; + tensor var_2299 = const()[name = tensor("op_2299"), val = tensor([1, 1, 16, 32, 2])]; + tensor q_complex_cast_fp16 = reshape(shape = var_2299, x = q_31_cast_fp16)[name = tensor("q_complex_cast_fp16")]; + tensor var_2303 = const()[name = tensor("op_2303"), val = tensor([1, 1, 16, 32, 2])]; + tensor k_complex_cast_fp16 = reshape(shape = var_2303, x = k_21_cast_fp16)[name = tensor("k_complex_cast_fp16")]; + tensor var_2307_begin_0 = const()[name = tensor("op_2307_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_2307_end_0 = const()[name = tensor("op_2307_end_0"), val = tensor([1, 1, 16, 32, 1])]; + tensor var_2307_end_mask_0 = const()[name = tensor("op_2307_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_2307_squeeze_mask_0 = const()[name = tensor("op_2307_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_2307_cast_fp16 = slice_by_index(begin = var_2307_begin_0, end = var_2307_end_0, end_mask = var_2307_end_mask_0, squeeze_mask = var_2307_squeeze_mask_0, x = q_complex_cast_fp16)[name = tensor("op_2307_cast_fp16")]; + tensor var_2315_begin_0 = const()[name = tensor("op_2315_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_2315_end_0 = const()[name = tensor("op_2315_end_0"), val = tensor([1, 1, 16, 32, 2])]; + tensor var_2315_end_mask_0 = const()[name = tensor("op_2315_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_2315_squeeze_mask_0 = const()[name = tensor("op_2315_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_2315_cast_fp16 = slice_by_index(begin = var_2315_begin_0, end = var_2315_end_0, end_mask = var_2315_end_mask_0, squeeze_mask = var_2315_squeeze_mask_0, x = q_complex_cast_fp16)[name = tensor("op_2315_cast_fp16")]; + tensor var_2323_begin_0 = const()[name = tensor("op_2323_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_2323_end_0 = const()[name = tensor("op_2323_end_0"), val = tensor([1, 1, 16, 32, 1])]; + tensor var_2323_end_mask_0 = const()[name = tensor("op_2323_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_2323_squeeze_mask_0 = const()[name = tensor("op_2323_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_2323_cast_fp16 = slice_by_index(begin = var_2323_begin_0, end = var_2323_end_0, end_mask = var_2323_end_mask_0, squeeze_mask = var_2323_squeeze_mask_0, x = k_complex_cast_fp16)[name = tensor("op_2323_cast_fp16")]; + tensor var_2331_begin_0 = const()[name = tensor("op_2331_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_2331_end_0 = const()[name = tensor("op_2331_end_0"), val = tensor([1, 1, 16, 32, 2])]; + tensor var_2331_end_mask_0 = const()[name = tensor("op_2331_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_2331_squeeze_mask_0 = const()[name = tensor("op_2331_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_2331_cast_fp16 = slice_by_index(begin = var_2331_begin_0, end = var_2331_end_0, end_mask = var_2331_end_mask_0, squeeze_mask = var_2331_squeeze_mask_0, x = k_complex_cast_fp16)[name = tensor("op_2331_cast_fp16")]; + tensor freqs_to_fp16 = const()[name = tensor("freqs_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(76798592)))]; + tensor var_2337_cast_fp16 = mul(x = freqs_to_fp16, y = ts_cast_fp16)[name = tensor("op_2337_cast_fp16")]; + tensor rotr_cast_fp16 = cos(x = var_2337_cast_fp16)[name = tensor("rotr_cast_fp16")]; + tensor roti_cast_fp16 = sin(x = var_2337_cast_fp16)[name = tensor("roti_cast_fp16")]; + tensor var_2341_cast_fp16 = mul(x = var_2307_cast_fp16, y = rotr_cast_fp16)[name = tensor("op_2341_cast_fp16")]; + tensor var_2342_cast_fp16 = mul(x = var_2315_cast_fp16, y = roti_cast_fp16)[name = tensor("op_2342_cast_fp16")]; + tensor qor_21_cast_fp16 = sub(x = var_2341_cast_fp16, y = var_2342_cast_fp16)[name = tensor("qor_21_cast_fp16")]; + tensor var_2345_cast_fp16 = mul(x = var_2307_cast_fp16, y = roti_cast_fp16)[name = tensor("op_2345_cast_fp16")]; + tensor var_2346_cast_fp16 = mul(x = var_2315_cast_fp16, y = rotr_cast_fp16)[name = tensor("op_2346_cast_fp16")]; + tensor qoi_21_cast_fp16 = add(x = var_2345_cast_fp16, y = var_2346_cast_fp16)[name = tensor("qoi_21_cast_fp16")]; + tensor var_2349_cast_fp16 = mul(x = var_2323_cast_fp16, y = rotr_cast_fp16)[name = tensor("op_2349_cast_fp16")]; + tensor var_2350_cast_fp16 = mul(x = var_2331_cast_fp16, y = roti_cast_fp16)[name = tensor("op_2350_cast_fp16")]; + tensor kor_21_cast_fp16 = sub(x = var_2349_cast_fp16, y = var_2350_cast_fp16)[name = tensor("kor_21_cast_fp16")]; + tensor var_2353_cast_fp16 = mul(x = var_2323_cast_fp16, y = roti_cast_fp16)[name = tensor("op_2353_cast_fp16")]; + tensor var_2354_cast_fp16 = mul(x = var_2331_cast_fp16, y = rotr_cast_fp16)[name = tensor("op_2354_cast_fp16")]; + tensor koi_21_cast_fp16 = add(x = var_2353_cast_fp16, y = var_2354_cast_fp16)[name = tensor("koi_21_cast_fp16")]; + tensor qo_axis_0 = const()[name = tensor("qo_axis_0"), val = tensor(-1)]; + tensor qo_cast_fp16 = stack(axis = qo_axis_0, values = (qor_21_cast_fp16, qoi_21_cast_fp16))[name = tensor("qo_cast_fp16")]; + tensor ko_axis_0 = const()[name = tensor("ko_axis_0"), val = tensor(-1)]; + tensor ko_cast_fp16 = stack(axis = ko_axis_0, values = (kor_21_cast_fp16, koi_21_cast_fp16))[name = tensor("ko_cast_fp16")]; + tensor var_2383 = const()[name = tensor("op_2383"), val = tensor([1, 1, 16, 64])]; + tensor q_33_cast_fp16 = reshape(shape = var_2383, x = qo_cast_fp16)[name = tensor("q_33_cast_fp16")]; + tensor var_2385 = const()[name = tensor("op_2385"), val = tensor([1, 1, 16, 64])]; + tensor k_cast_fp16 = reshape(shape = var_2385, x = ko_cast_fp16)[name = tensor("k_cast_fp16")]; + tensor _inversed_2407_y_0_to_fp16 = const()[name = tensor("_inversed_2407_y_0_to_fp16"), val = tensor(0x1p-9)]; + tensor _inversed_2407_cast_fp16 = mul(x = ts_cast_fp16, y = _inversed_2407_y_0_to_fp16)[name = tensor("_inversed_2407_cast_fp16")]; + tensor var_2408_cast_fp16 = floor(x = _inversed_2407_cast_fp16)[name = tensor("op_2408_cast_fp16")]; + tensor var_2409_to_fp16 = const()[name = tensor("op_2409_to_fp16"), val = tensor(0x1p+9)]; + tensor var_2410_cast_fp16 = mul(x = var_2408_cast_fp16, y = var_2409_to_fp16)[name = tensor("op_2410_cast_fp16")]; + tensor write_indices_float_cast_fp16 = sub(x = ts_cast_fp16, y = var_2410_cast_fp16)[name = tensor("write_indices_float_cast_fp16")]; + tensor var_2417_dtype_0 = const()[name = tensor("op_2417_dtype_0"), val = tensor("int32")]; + tensor write_indices_reps_0 = const()[name = tensor("write_indices_reps_0"), val = tensor([1, 1, 16, 64])]; + tensor write_indices_float_cast_fp16_to_int32 = cast(dtype = var_2417_dtype_0, x = write_indices_float_cast_fp16)[name = tensor("cast_113")]; + tensor write_indices = tile(reps = write_indices_reps_0, x = write_indices_float_cast_fp16_to_int32)[name = tensor("write_indices")]; + tensor var_2425_begin_0 = const()[name = tensor("op_2425_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_2425_end_0 = const()[name = tensor("op_2425_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor var_2425_end_mask_0 = const()[name = tensor("op_2425_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_2425_squeeze_mask_0 = const()[name = tensor("op_2425_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor cache5_to_fp16_dtype_0 = const()[name = tensor("cache5_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor cache5_to_fp16 = cast(dtype = cache5_to_fp16_dtype_0, x = cache5)[name = tensor("cast_112")]; + tensor var_2425_cast_fp16 = slice_by_index(begin = var_2425_begin_0, end = var_2425_end_0, end_mask = var_2425_end_mask_0, squeeze_mask = var_2425_squeeze_mask_0, x = cache5_to_fp16)[name = tensor("op_2425_cast_fp16")]; + tensor var_2427_axis_0 = const()[name = tensor("op_2427_axis_0"), val = tensor(1)]; + tensor var_2427_mode_0 = const()[name = tensor("op_2427_mode_0"), val = tensor("update")]; + tensor var_2427_validate_indices_0 = const()[name = tensor("op_2427_validate_indices_0"), val = tensor(false)]; + tensor var_2427_cast_fp16 = scatter_along_axis(axis = var_2427_axis_0, data = var_2425_cast_fp16, indices = write_indices, mode = var_2427_mode_0, updates = k_cast_fp16, validate_indices = var_2427_validate_indices_0)[name = tensor("op_2427_cast_fp16")]; + tensor concat_37 = const()[name = tensor("concat_37"), val = tensor([0, 0, 0, 0, 0])]; + tensor concat_38 = const()[name = tensor("concat_38"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_internal_tensor_assign_1_stride_0 = const()[name = tensor("new_cache_internal_tensor_assign_1_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_internal_tensor_assign_1_begin_mask_0 = const()[name = tensor("new_cache_internal_tensor_assign_1_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_internal_tensor_assign_1_end_mask_0 = const()[name = tensor("new_cache_internal_tensor_assign_1_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_internal_tensor_assign_1_squeeze_mask_0 = const()[name = tensor("new_cache_internal_tensor_assign_1_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_22 = const()[name = tensor("shape_22"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_10 = const()[name = tensor("reduce_prod_10"), val = tensor(1048576)]; + tensor range_1d_10_start_0 = const()[name = tensor("range_1d_10_start_0"), val = tensor(0)]; + tensor range_1d_10_step_0 = const()[name = tensor("range_1d_10_step_0"), val = tensor(1)]; + tensor range_1d_10 = range_1d(end = reduce_prod_10, start = range_1d_10_start_0, step = range_1d_10_step_0)[name = tensor("range_1d_10")]; + tensor reshape_50 = reshape(shape = shape_22, x = range_1d_10)[name = tensor("reshape_50")]; + tensor slice_by_index_10 = slice_by_index(begin = concat_37, begin_mask = new_cache_internal_tensor_assign_1_begin_mask_0, end = concat_38, end_mask = new_cache_internal_tensor_assign_1_end_mask_0, squeeze_mask = new_cache_internal_tensor_assign_1_squeeze_mask_0, stride = new_cache_internal_tensor_assign_1_stride_0, x = reshape_50)[name = tensor("slice_by_index_10")]; + tensor reshape_51_shape_0 = const()[name = tensor("reshape_51_shape_0"), val = tensor([-1])]; + tensor reshape_51 = reshape(shape = reshape_51_shape_0, x = slice_by_index_10)[name = tensor("reshape_51")]; + tensor reshape_52_shape_0 = const()[name = tensor("reshape_52_shape_0"), val = tensor([-1])]; + tensor reshape_52_cast_fp16 = reshape(shape = reshape_52_shape_0, x = var_2427_cast_fp16)[name = tensor("reshape_52_cast_fp16")]; + tensor reshape_53_shape_0 = const()[name = tensor("reshape_53_shape_0"), val = tensor([-1])]; + tensor reshape_53_cast_fp16 = reshape(shape = reshape_53_shape_0, x = cache5_to_fp16)[name = tensor("reshape_53_cast_fp16")]; + tensor scatter_10_mode_0 = const()[name = tensor("scatter_10_mode_0"), val = tensor("update")]; + tensor scatter_10_axis_0 = const()[name = tensor("scatter_10_axis_0"), val = tensor(0)]; + tensor scatter_10_validate_indices_0 = const()[name = tensor("scatter_10_validate_indices_0"), val = tensor(false)]; + tensor scatter_10_cast_fp16 = scatter(axis = scatter_10_axis_0, data = reshape_53_cast_fp16, indices = reshape_51, mode = scatter_10_mode_0, updates = reshape_52_cast_fp16, validate_indices = scatter_10_validate_indices_0)[name = tensor("scatter_10_cast_fp16")]; + tensor reshape_54_cast_fp16 = reshape(shape = shape_22, x = scatter_10_cast_fp16)[name = tensor("reshape_54_cast_fp16")]; + tensor var_2435_begin_0 = const()[name = tensor("op_2435_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor var_2435_end_0 = const()[name = tensor("op_2435_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor var_2435_end_mask_0 = const()[name = tensor("op_2435_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_2435_squeeze_mask_0 = const()[name = tensor("op_2435_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor var_2435_cast_fp16 = slice_by_index(begin = var_2435_begin_0, end = var_2435_end_0, end_mask = var_2435_end_mask_0, squeeze_mask = var_2435_squeeze_mask_0, x = reshape_54_cast_fp16)[name = tensor("op_2435_cast_fp16")]; + tensor var_2437_axis_0 = const()[name = tensor("op_2437_axis_0"), val = tensor(1)]; + tensor var_2437_mode_0 = const()[name = tensor("op_2437_mode_0"), val = tensor("update")]; + tensor var_2437_validate_indices_0 = const()[name = tensor("op_2437_validate_indices_0"), val = tensor(false)]; + tensor var_2437_cast_fp16 = scatter_along_axis(axis = var_2437_axis_0, data = var_2435_cast_fp16, indices = write_indices, mode = var_2437_mode_0, updates = v_cast_fp16, validate_indices = var_2437_validate_indices_0)[name = tensor("op_2437_cast_fp16")]; + tensor concat_39 = const()[name = tensor("concat_39"), val = tensor([1, 0, 0, 0, 0])]; + tensor concat_40 = const()[name = tensor("concat_40"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_internal_tensor_assign_2_stride_0 = const()[name = tensor("new_cache_internal_tensor_assign_2_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_internal_tensor_assign_2_begin_mask_0 = const()[name = tensor("new_cache_internal_tensor_assign_2_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_internal_tensor_assign_2_end_mask_0 = const()[name = tensor("new_cache_internal_tensor_assign_2_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_internal_tensor_assign_2_squeeze_mask_0 = const()[name = tensor("new_cache_internal_tensor_assign_2_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_23 = const()[name = tensor("shape_23"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_11 = const()[name = tensor("reduce_prod_11"), val = tensor(1048576)]; + tensor range_1d_11_start_0 = const()[name = tensor("range_1d_11_start_0"), val = tensor(0)]; + tensor range_1d_11_step_0 = const()[name = tensor("range_1d_11_step_0"), val = tensor(1)]; + tensor range_1d_11 = range_1d(end = reduce_prod_11, start = range_1d_11_start_0, step = range_1d_11_step_0)[name = tensor("range_1d_11")]; + tensor reshape_55 = reshape(shape = shape_23, x = range_1d_11)[name = tensor("reshape_55")]; + tensor slice_by_index_11 = slice_by_index(begin = concat_39, begin_mask = new_cache_internal_tensor_assign_2_begin_mask_0, end = concat_40, end_mask = new_cache_internal_tensor_assign_2_end_mask_0, squeeze_mask = new_cache_internal_tensor_assign_2_squeeze_mask_0, stride = new_cache_internal_tensor_assign_2_stride_0, x = reshape_55)[name = tensor("slice_by_index_11")]; + tensor reshape_56_shape_0 = const()[name = tensor("reshape_56_shape_0"), val = tensor([-1])]; + tensor reshape_56 = reshape(shape = reshape_56_shape_0, x = slice_by_index_11)[name = tensor("reshape_56")]; + tensor reshape_57_shape_0 = const()[name = tensor("reshape_57_shape_0"), val = tensor([-1])]; + tensor reshape_57_cast_fp16 = reshape(shape = reshape_57_shape_0, x = var_2437_cast_fp16)[name = tensor("reshape_57_cast_fp16")]; + tensor reshape_58_shape_0 = const()[name = tensor("reshape_58_shape_0"), val = tensor([-1])]; + tensor reshape_58_cast_fp16 = reshape(shape = reshape_58_shape_0, x = reshape_54_cast_fp16)[name = tensor("reshape_58_cast_fp16")]; + tensor scatter_11_mode_0 = const()[name = tensor("scatter_11_mode_0"), val = tensor("update")]; + tensor scatter_11_axis_0 = const()[name = tensor("scatter_11_axis_0"), val = tensor(0)]; + tensor scatter_11_validate_indices_0 = const()[name = tensor("scatter_11_validate_indices_0"), val = tensor(false)]; + tensor scatter_11_cast_fp16 = scatter(axis = scatter_11_axis_0, data = reshape_58_cast_fp16, indices = reshape_56, mode = scatter_11_mode_0, updates = reshape_57_cast_fp16, validate_indices = scatter_11_validate_indices_0)[name = tensor("scatter_11_cast_fp16")]; + tensor reshape_59_cast_fp16 = reshape(shape = shape_23, x = scatter_11_cast_fp16)[name = tensor("reshape_59_cast_fp16")]; + tensor reshape_59_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("reshape_59_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor keys_31_begin_0 = const()[name = tensor("keys_31_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor keys_31_end_0 = const()[name = tensor("keys_31_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor keys_31_end_mask_0 = const()[name = tensor("keys_31_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor keys_31_squeeze_mask_0 = const()[name = tensor("keys_31_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor keys_31_cast_fp16 = slice_by_index(begin = keys_31_begin_0, end = keys_31_end_0, end_mask = keys_31_end_mask_0, squeeze_mask = keys_31_squeeze_mask_0, x = reshape_59_cast_fp16)[name = tensor("keys_31_cast_fp16")]; + tensor values_31_begin_0 = const()[name = tensor("values_31_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor values_31_end_0 = const()[name = tensor("values_31_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor values_31_end_mask_0 = const()[name = tensor("values_31_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor values_31_squeeze_mask_0 = const()[name = tensor("values_31_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor values_31_cast_fp16 = slice_by_index(begin = values_31_begin_0, end = values_31_end_0, end_mask = values_31_end_mask_0, squeeze_mask = values_31_squeeze_mask_0, x = reshape_59_cast_fp16)[name = tensor("values_31_cast_fp16")]; + tensor var_2449_cast_fp16 = not_equal(x = keys_31_cast_fp16, y = keys_31_cast_fp16)[name = tensor("op_2449_cast_fp16")]; + tensor keys_33_cast_fp16 = select(a = var_360_to_fp16, b = keys_31_cast_fp16, cond = var_2449_cast_fp16)[name = tensor("keys_33_cast_fp16")]; + tensor var_2457_cast_fp16 = not_equal(x = values_31_cast_fp16, y = values_31_cast_fp16)[name = tensor("op_2457_cast_fp16")]; + tensor values_33_cast_fp16 = select(a = var_360_to_fp16, b = values_31_cast_fp16, cond = var_2457_cast_fp16)[name = tensor("values_33_cast_fp16")]; + tensor var_2481 = const()[name = tensor("op_2481"), val = tensor([0, 2, 1, 3])]; + tensor var_2494 = const()[name = tensor("op_2494"), val = tensor([1, 1, 1])]; + tensor var_2495_cast_fp16 = reshape(shape = var_2494, x = position5_to_fp16)[name = tensor("op_2495_cast_fp16")]; + tensor var_2512_to_fp16 = const()[name = tensor("op_2512_to_fp16"), val = tensor(0x1p+0)]; + tensor valid_len_cast_fp16 = add(x = var_2495_cast_fp16, y = var_2512_to_fp16)[name = tensor("valid_len_cast_fp16")]; + tensor valid_mask_cast_fp16 = less(x = k_positions_1_promoted_to_fp16, y = valid_len_cast_fp16)[name = tensor("valid_mask_cast_fp16")]; + tensor causal_mask_cast_fp16 = less_equal(x = k_positions_1_promoted_to_fp16, y = var_2495_cast_fp16)[name = tensor("causal_mask_cast_fp16")]; + tensor attn_mask_21 = logical_and(x = valid_mask_cast_fp16, y = causal_mask_cast_fp16)[name = tensor("attn_mask_21")]; + tensor attn_mask_axes_0 = const()[name = tensor("attn_mask_axes_0"), val = tensor([1])]; + tensor attn_mask = expand_dims(axes = attn_mask_axes_0, x = attn_mask_21)[name = tensor("attn_mask")]; + tensor var_2530_transpose_x_0 = const()[name = tensor("op_2530_transpose_x_0"), val = tensor(false)]; + tensor var_2530_transpose_y_0 = const()[name = tensor("op_2530_transpose_y_0"), val = tensor(false)]; + tensor transpose_28_perm_0 = const()[name = tensor("transpose_28_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_29_perm_0 = const()[name = tensor("transpose_29_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_29 = transpose(perm = transpose_29_perm_0, x = keys_33_cast_fp16)[name = tensor("transpose_31")]; + tensor transpose_28 = transpose(perm = transpose_28_perm_0, x = q_33_cast_fp16)[name = tensor("transpose_32")]; + tensor var_2530_cast_fp16 = matmul(transpose_x = var_2530_transpose_x_0, transpose_y = var_2530_transpose_y_0, x = transpose_28, y = transpose_29)[name = tensor("op_2530_cast_fp16")]; + tensor var_2524_to_fp16 = const()[name = tensor("op_2524_to_fp16"), val = tensor([0x1p-3])]; + tensor attn_weights_31_cast_fp16 = mul(x = var_2530_cast_fp16, y = var_2524_to_fp16)[name = tensor("attn_weights_31_cast_fp16")]; + tensor var_2532 = logical_not(x = attn_mask)[name = tensor("op_2532")]; + tensor var_2533_to_fp16 = const()[name = tensor("op_2533_to_fp16"), val = tensor(-inf)]; + tensor attn_weights_33_cast_fp16 = select(a = var_2533_to_fp16, b = attn_weights_31_cast_fp16, cond = var_2532)[name = tensor("attn_weights_33_cast_fp16")]; + tensor var_2535 = const()[name = tensor("op_2535"), val = tensor(-1)]; + tensor attn_weights_cast_fp16 = softmax(axis = var_2535, x = attn_weights_33_cast_fp16)[name = tensor("attn_weights_cast_fp16")]; + tensor attn_output_transpose_x_0 = const()[name = tensor("attn_output_transpose_x_0"), val = tensor(false)]; + tensor attn_output_transpose_y_0 = const()[name = tensor("attn_output_transpose_y_0"), val = tensor(false)]; + tensor values_cast_fp16 = transpose(perm = var_2481, x = values_33_cast_fp16)[name = tensor("transpose_33")]; + tensor attn_output_cast_fp16 = matmul(transpose_x = attn_output_transpose_x_0, transpose_y = attn_output_transpose_y_0, x = attn_weights_cast_fp16, y = values_cast_fp16)[name = tensor("attn_output_cast_fp16")]; + tensor var_2543 = const()[name = tensor("op_2543"), val = tensor([0, 2, 1, 3])]; + tensor var_2546 = const()[name = tensor("op_2546"), val = tensor([1, 1, 1024])]; + tensor var_2544_cast_fp16 = transpose(perm = var_2543, x = attn_output_cast_fp16)[name = tensor("transpose_30")]; + tensor input_55_cast_fp16 = reshape(shape = var_2546, x = var_2544_cast_fp16)[name = tensor("input_55_cast_fp16")]; + tensor linear_22_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = attn5_out_proj_weight_cast_fp16, x = input_55_cast_fp16)[name = tensor("linear_22_cast_fp16")]; + tensor var_2552_to_fp16 = const()[name = tensor("op_2552_to_fp16"), val = tensor(0x1p+0)]; + tensor var_2553_cast_fp16 = add(x = position5_to_fp16, y = var_2552_to_fp16)[name = tensor("op_2553_cast_fp16")]; + tensor var_2553_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_2553_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor input_57_cast_fp16 = add(x = input_53_cast_fp16, y = linear_22_cast_fp16)[name = tensor("input_57_cast_fp16")]; + tensor input_59_axes_0 = const()[name = tensor("input_59_axes_0"), val = tensor([-1])]; + tensor norm5_2_weight_to_fp16 = const()[name = tensor("norm5_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(76798720)))]; + tensor norm5_2_bias_to_fp16 = const()[name = tensor("norm5_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(76800832)))]; + tensor var_2557_to_fp16 = const()[name = tensor("op_2557_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_59_cast_fp16 = layer_norm(axes = input_59_axes_0, beta = norm5_2_bias_to_fp16, epsilon = var_2557_to_fp16, gamma = norm5_2_weight_to_fp16, x = input_57_cast_fp16)[name = tensor("input_59_cast_fp16")]; + tensor linear_23_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = linear5_1_weight_cast_fp16, x = input_59_cast_fp16)[name = tensor("linear_23_cast_fp16")]; + tensor input_61_mode_0 = const()[name = tensor("input_61_mode_0"), val = tensor("EXACT")]; + tensor input_61_cast_fp16 = gelu(mode = input_61_mode_0, x = linear_23_cast_fp16)[name = tensor("input_61_cast_fp16")]; + tensor linear_24_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = linear5_2_weight_cast_fp16, x = input_61_cast_fp16)[name = tensor("linear_24_cast_fp16")]; + tensor input_63_cast_fp16 = add(x = input_57_cast_fp16, y = linear_24_cast_fp16)[name = tensor("input_63_cast_fp16")]; + tensor input_axes_0 = const()[name = tensor("input_axes_0"), val = tensor([-1])]; + tensor out_norm_weight_to_fp16 = const()[name = tensor("out_norm_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(76802944)))]; + tensor out_norm_bias_to_fp16 = const()[name = tensor("out_norm_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(76805056)))]; + tensor var_2574_to_fp16 = const()[name = tensor("op_2574_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_cast_fp16 = layer_norm(axes = input_axes_0, beta = out_norm_bias_to_fp16, epsilon = var_2574_to_fp16, gamma = out_norm_weight_to_fp16, x = input_63_cast_fp16)[name = tensor("input_cast_fp16")]; + tensor input_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("input_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor out_eos_weight_to_fp16 = const()[name = tensor("out_eos_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(76807168)))]; + tensor out_eos_bias_to_fp16 = const()[name = tensor("out_eos_bias_to_fp16"), val = tensor([-0x1.64p-3])]; + tensor linear_25_cast_fp16 = linear(bias = out_eos_bias_to_fp16, weight = out_eos_weight_to_fp16, x = input_cast_fp16)[name = tensor("linear_25_cast_fp16")]; + tensor linear_25_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("linear_25_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor var_2582 = cast(dtype = linear_25_cast_fp16_to_fp32_dtype_0, x = linear_25_cast_fp16)[name = tensor("cast_108")]; + tensor input = cast(dtype = input_cast_fp16_to_fp32_dtype_0, x = input_cast_fp16)[name = tensor("cast_109")]; + tensor var_2553 = cast(dtype = var_2553_cast_fp16_to_fp32_dtype_0, x = var_2553_cast_fp16)[name = tensor("cast_110")]; + tensor new_cache_internal_tensor_assign_2 = cast(dtype = reshape_59_cast_fp16_to_fp32_dtype_0, x = reshape_59_cast_fp16)[name = tensor("cast_111")]; + tensor var_2134 = cast(dtype = var_2134_cast_fp16_to_fp32_dtype_0, x = var_2134_cast_fp16)[name = tensor("cast_115")]; + tensor new_cache_9_internal_tensor_assign_2 = cast(dtype = reshape_49_cast_fp16_to_fp32_dtype_0, x = reshape_49_cast_fp16)[name = tensor("cast_116")]; + tensor var_1715 = cast(dtype = var_1715_cast_fp16_to_fp32_dtype_0, x = var_1715_cast_fp16)[name = tensor("cast_120")]; + tensor new_cache_7_internal_tensor_assign_2 = cast(dtype = reshape_39_cast_fp16_to_fp32_dtype_0, x = reshape_39_cast_fp16)[name = tensor("cast_121")]; + tensor var_1296 = cast(dtype = var_1296_cast_fp16_to_fp32_dtype_0, x = var_1296_cast_fp16)[name = tensor("cast_125")]; + tensor new_cache_5_internal_tensor_assign_2 = cast(dtype = reshape_29_cast_fp16_to_fp32_dtype_0, x = reshape_29_cast_fp16)[name = tensor("cast_126")]; + tensor var_877 = cast(dtype = var_877_cast_fp16_to_fp32_dtype_0, x = var_877_cast_fp16)[name = tensor("cast_130")]; + tensor new_cache_3_internal_tensor_assign_2 = cast(dtype = reshape_19_cast_fp16_to_fp32_dtype_0, x = reshape_19_cast_fp16)[name = tensor("cast_131")]; + tensor var_458 = cast(dtype = var_458_cast_fp16_to_fp32_dtype_0, x = var_458_cast_fp16)[name = tensor("cast_135")]; + tensor new_cache_1_internal_tensor_assign_2 = cast(dtype = reshape_9_cast_fp16_to_fp32_dtype_0, x = reshape_9_cast_fp16)[name = tensor("cast_136")]; + } -> (input, var_2582, new_cache_1_internal_tensor_assign_2, var_458, new_cache_3_internal_tensor_assign_2, var_877, new_cache_5_internal_tensor_assign_2, var_1296, new_cache_7_internal_tensor_assign_2, var_1715, new_cache_9_internal_tensor_assign_2, var_2134, new_cache_internal_tensor_assign_2, var_2553); +} \ No newline at end of file diff --git a/v2.1/italian/flowlm_stepv2.mlmodelc/weights/weight.bin b/v2.1/italian/flowlm_stepv2.mlmodelc/weights/weight.bin new file mode 100644 index 0000000000000000000000000000000000000000..327b3563c74d90c632a2902043b7140a6bf9b88f --- /dev/null +++ b/v2.1/italian/flowlm_stepv2.mlmodelc/weights/weight.bin @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:3e8bf78206107b195778090f1b801103056bc6a7e10c9fca645efa1205021fb2 +size 76809280 diff --git a/v2.1/italian/flowlm_stepv2.mlpackage/Data/com.apple.CoreML/model.mlmodel b/v2.1/italian/flowlm_stepv2.mlpackage/Data/com.apple.CoreML/model.mlmodel new file mode 100644 index 0000000000000000000000000000000000000000..262eec4d7175599c978cb62a62cfc313512f74b2 --- /dev/null +++ b/v2.1/italian/flowlm_stepv2.mlpackage/Data/com.apple.CoreML/model.mlmodel @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:ca949d1b6e7294d08d84325267f5f19db764fa8c44959a113c7895443c21bd78 +size 212305 diff --git a/v2.1/italian/flowlm_stepv2.mlpackage/Data/com.apple.CoreML/weights/weight.bin b/v2.1/italian/flowlm_stepv2.mlpackage/Data/com.apple.CoreML/weights/weight.bin new file mode 100644 index 0000000000000000000000000000000000000000..327b3563c74d90c632a2902043b7140a6bf9b88f --- /dev/null +++ b/v2.1/italian/flowlm_stepv2.mlpackage/Data/com.apple.CoreML/weights/weight.bin @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:3e8bf78206107b195778090f1b801103056bc6a7e10c9fca645efa1205021fb2 +size 76809280 diff --git a/v2.1/italian/flowlm_stepv2.mlpackage/Manifest.json b/v2.1/italian/flowlm_stepv2.mlpackage/Manifest.json new file mode 100644 index 0000000000000000000000000000000000000000..856211b5368aafb2c755cd203d7ff8effab58bb5 --- /dev/null +++ b/v2.1/italian/flowlm_stepv2.mlpackage/Manifest.json @@ -0,0 +1,18 @@ +{ + "fileFormatVersion": "1.0.0", + "itemInfoEntries": { + "8D0C92FD-866D-4737-B6D7-3CB862C06540": { + "author": "com.apple.CoreML", + "description": "CoreML Model Weights", + "name": "weights", + "path": "com.apple.CoreML/weights" + }, + "C3046C60-C34D-425B-B441-BEFBDAC6E421": { + "author": "com.apple.CoreML", + "description": "CoreML Model Specification", + "name": "model.mlmodel", + "path": "com.apple.CoreML/model.mlmodel" + } + }, + "rootModelIdentifier": "C3046C60-C34D-425B-B441-BEFBDAC6E421" +} diff --git a/v2.1/italian/manifest.json b/v2.1/italian/manifest.json new file mode 100644 index 0000000000000000000000000000000000000000..fc7da98920126d1261a9c47517b1303db1599db7 --- /dev/null +++ b/v2.1/italian/manifest.json @@ -0,0 +1,27 @@ +{ + "version": "2.1", + "base": "v2", + "kind": "optimization-reconvert", + "self_contained": true, + "note": "Same weights as v2. Re-converted for speed: fused flow decoder (100% ANE), one-shot cond prefill, fp16 flowlm. NOT a finetune.", + "new": [ + "flow_decoder_fused", + "cond_prefill", + "flowlm_step" + ], + "copied_from_v2": [ + "flowlm_stepv2.mlpackage", + "flowlm_stepv2.mlmodelc", + "mimi_decoder.mlpackage", + "mimi_decoder.mlmodelc", + "constants", + "constants_bin" + ], + "compute_units": { + "flow_decoder_fused": "all (ANE)", + "cond_prefill": "all (GPU)", + "flowlm_step": "all (GPU, fp16)", + "flowlm_stepv2": "cpuAndGpu (GPU, int8 \u2014 fastest flowlm)", + "mimi_decoder": "cpuOnly" + } +} diff --git a/v2.1/italian/mimi_decoder.mlmodelc/analytics/coremldata.bin b/v2.1/italian/mimi_decoder.mlmodelc/analytics/coremldata.bin new file mode 100644 index 0000000000000000000000000000000000000000..b302e375d31a07ece125db11fb87c818abb59ea2 --- /dev/null +++ b/v2.1/italian/mimi_decoder.mlmodelc/analytics/coremldata.bin @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:5acd53da51d268c30df8f691e564f40c924047d3c9253c003b021e00fc8dd69d +size 243 diff --git a/v2.1/italian/mimi_decoder.mlmodelc/coremldata.bin b/v2.1/italian/mimi_decoder.mlmodelc/coremldata.bin new file mode 100644 index 0000000000000000000000000000000000000000..2a47bb05123f95efeacec13be8028417eb73395f --- /dev/null +++ b/v2.1/italian/mimi_decoder.mlmodelc/coremldata.bin @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:2efdf24a47db9d0b238c46996854bf587b9b18e745c1848cf48af815f392a632 +size 1777 diff --git a/v2.1/italian/mimi_decoder.mlmodelc/metadata.json b/v2.1/italian/mimi_decoder.mlmodelc/metadata.json new file mode 100644 index 0000000000000000000000000000000000000000..f66f9e7ff5be2ffbc44c8300f3501f67a77afaf6 --- /dev/null +++ b/v2.1/italian/mimi_decoder.mlmodelc/metadata.json @@ -0,0 +1,573 @@ +[ + { + "metadataOutputVersion" : "3.0", + "storagePrecision" : "Float32", + "outputSchema" : [ + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1 × 1 × 1920)", + "shortDescription" : "", + "shape" : "[1, 1, 1920]", + "name" : "var_724", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1 × 512 × 16)", + "shortDescription" : "", + "shape" : "[1, 512, 16]", + "name" : "var_77", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 2 × 1 × 256 × 8 × 64)", + "shortDescription" : "", + "shape" : "[2, 1, 256, 8, 64]", + "name" : "var_210", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1)", + "shortDescription" : "", + "shape" : "[1]", + "name" : "var_741", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 2 × 1 × 256 × 8 × 64)", + "shortDescription" : "", + "shape" : "[2, 1, 256, 8, 64]", + "name" : "var_400", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1)", + "shortDescription" : "", + "shape" : "[1]", + "name" : "var_744", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1 × 512 × 6)", + "shortDescription" : "", + "shape" : "[1, 512, 6]", + "name" : "var_542", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1)", + "shortDescription" : "", + "shape" : "[1]", + "name" : "conv0_first", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1 × 256 × 6)", + "shortDescription" : "", + "shape" : "[1, 256, 6]", + "name" : "var_565", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1 × 256 × 2)", + "shortDescription" : "", + "shape" : "[1, 256, 2]", + "name" : "var_585", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1)", + "shortDescription" : "", + "shape" : "[1]", + "name" : "res0_conv0_first", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1 × 128 × 0)", + "shortDescription" : "", + "shape" : "[1, 128, 0]", + "name" : "res0_conv1_prev", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1)", + "shortDescription" : "", + "shape" : "[1]", + "name" : "res0_conv1_first", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1 × 128 × 5)", + "shortDescription" : "", + "shape" : "[1, 128, 5]", + "name" : "var_621", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1 × 128 × 2)", + "shortDescription" : "", + "shape" : "[1, 128, 2]", + "name" : "var_641", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1)", + "shortDescription" : "", + "shape" : "[1]", + "name" : "res1_conv0_first", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1 × 64 × 0)", + "shortDescription" : "", + "shape" : "[1, 64, 0]", + "name" : "res1_conv1_prev", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1)", + "shortDescription" : "", + "shape" : "[1]", + "name" : "res1_conv1_first", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1 × 64 × 4)", + "shortDescription" : "", + "shape" : "[1, 64, 4]", + "name" : "var_677", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1 × 64 × 2)", + "shortDescription" : "", + "shape" : "[1, 64, 2]", + "name" : "var_697", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1)", + "shortDescription" : "", + "shape" : "[1]", + "name" : "res2_conv0_first", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1 × 32 × 0)", + "shortDescription" : "", + "shape" : "[1, 32, 0]", + "name" : "res2_conv1_prev", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1)", + "shortDescription" : "", + "shape" : "[1]", + "name" : "res2_conv1_first", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1 × 64 × 2)", + "shortDescription" : "", + "shape" : "[1, 64, 2]", + "name" : "var_725", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1)", + "shortDescription" : "", + "shape" : "[1]", + "name" : "conv_final_first", + "type" : "MultiArray" + } + ], + "modelParameters" : [ + + ], + "specificationVersion" : 8, + "mlProgramOperationTypeHistogram" : { + "Ios17.notEqual" : 4, + "Tile" : 2, + "Ios17.lessEqual" : 2, + "Ios17.layerNorm" : 4, + "Stack" : 6, + "Ios17.greaterEqual" : 4, + "Ios17.scatterAlongAxis" : 4, + "Ios17.sub" : 17, + "Identity" : 11, + "Ios17.elu" : 10, + "Ios17.logicalNot" : 2, + "Ios17.sliceByIndex" : 35, + "Ios17.concat" : 9, + "Ios17.convTranspose" : 4, + "Ios17.floorDiv" : 4, + "Ios17.expandDims" : 7, + "Ios17.less" : 2, + "Ios17.mul" : 29, + "Ios17.transpose" : 10, + "Ios17.matmul" : 4, + "Ios16.gelu" : 2, + "Ios17.conv" : 9, + "Ios17.reshape" : 22, + "Split" : 2, + "Ios17.squeeze" : 6, + "Ios17.cast" : 6, + "Ios17.sin" : 2, + "Ios17.linear" : 8, + "Select" : 6, + "Ios16.softmax" : 2, + "Ios17.add" : 28, + "Ios17.cos" : 2, + "Ios17.logicalAnd" : 6 + }, + "computePrecision" : "Mixed (Float32, Int32)", + "isUpdatable" : "0", + "stateSchema" : [ + + ], + "availability" : { + "macOS" : "14.0", + "tvOS" : "17.0", + "visionOS" : "1.0", + "watchOS" : "10.0", + "iOS" : "17.0", + "macCatalyst" : "17.0" + }, + "modelType" : { + "name" : "MLModelType_mlProgram" + }, + "userDefinedMetadata" : { + "com.github.apple.coremltools.conversion_date" : "2026-04-24", + "com.github.apple.coremltools.source" : "torch==2.9.1", + "com.github.apple.coremltools.version" : "9.0", + "com.github.apple.coremltools.source_dialect" : "TorchScript" + }, + "inputSchema" : [ + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1 × 32)", + "shortDescription" : "", + "shape" : "[1, 32]", + "name" : "latent", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1 × 512 × 16)", + "shortDescription" : "", + "shape" : "[1, 512, 16]", + "name" : "upsample_partial", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 2 × 1 × 256 × 8 × 64)", + "shortDescription" : "", + "shape" : "[2, 1, 256, 8, 64]", + "name" : "attn0_cache", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1)", + "shortDescription" : "", + "shape" : "[1]", + "name" : "attn0_offset", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 2 × 1 × 256 × 8 × 64)", + "shortDescription" : "", + "shape" : "[2, 1, 256, 8, 64]", + "name" : "attn1_cache", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1)", + "shortDescription" : "", + "shape" : "[1]", + "name" : "attn1_offset", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1 × 512 × 6)", + "shortDescription" : "", + "shape" : "[1, 512, 6]", + "name" : "conv0_prev", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1)", + "shortDescription" : "", + "shape" : "[1]", + "name" : "conv0_first", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1 × 256 × 6)", + "shortDescription" : "", + "shape" : "[1, 256, 6]", + "name" : "convtr0_partial", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1 × 256 × 2)", + "shortDescription" : "", + "shape" : "[1, 256, 2]", + "name" : "res0_conv0_prev", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1)", + "shortDescription" : "", + "shape" : "[1]", + "name" : "res0_conv0_first", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1 × 128 × 0)", + "shortDescription" : "", + "shape" : "[1, 128, 0]", + "name" : "res0_conv1_prev", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1)", + "shortDescription" : "", + "shape" : "[1]", + "name" : "res0_conv1_first", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1 × 128 × 5)", + "shortDescription" : "", + "shape" : "[1, 128, 5]", + "name" : "convtr1_partial", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1 × 128 × 2)", + "shortDescription" : "", + "shape" : "[1, 128, 2]", + "name" : "res1_conv0_prev", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1)", + "shortDescription" : "", + "shape" : "[1]", + "name" : "res1_conv0_first", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1 × 64 × 0)", + "shortDescription" : "", + "shape" : "[1, 64, 0]", + "name" : "res1_conv1_prev", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1)", + "shortDescription" : "", + "shape" : "[1]", + "name" : "res1_conv1_first", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1 × 64 × 4)", + "shortDescription" : "", + "shape" : "[1, 64, 4]", + "name" : "convtr2_partial", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1 × 64 × 2)", + "shortDescription" : "", + "shape" : "[1, 64, 2]", + "name" : "res2_conv0_prev", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1)", + "shortDescription" : "", + "shape" : "[1]", + "name" : "res2_conv0_first", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1 × 32 × 0)", + "shortDescription" : "", + "shape" : "[1, 32, 0]", + "name" : "res2_conv1_prev", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1)", + "shortDescription" : "", + "shape" : "[1]", + "name" : "res2_conv1_first", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1 × 64 × 2)", + "shortDescription" : "", + "shape" : "[1, 64, 2]", + "name" : "conv_final_prev", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1)", + "shortDescription" : "", + "shape" : "[1]", + "name" : "conv_final_first", + "type" : "MultiArray" + } + ], + "generatedClassName" : "mimi_decoder", + "method" : "predict" + } +] \ No newline at end of file diff --git a/v2.1/italian/mimi_decoder.mlmodelc/model.mil b/v2.1/italian/mimi_decoder.mlmodelc/model.mil new file mode 100644 index 0000000000000000000000000000000000000000..1757e7713b3aa3b787d85968c60ed3fbe4cedb08 --- /dev/null +++ b/v2.1/italian/mimi_decoder.mlmodelc/model.mil @@ -0,0 +1,646 @@ +program(1.0) +[buildInfo = dict, tensor>({{"coremlc-component-MIL", "3520.4.1"}, {"coremlc-version", "3520.5.1"}, {"coremltools-component-torch", "2.9.1"}, {"coremltools-source-dialect", "TorchScript"}, {"coremltools-version", "9.0"}})] +{ + func main(tensor attn0_cache, tensor attn0_offset, tensor attn1_cache, tensor attn1_offset, tensor conv0_first, tensor conv0_prev, tensor conv_final_first, tensor conv_final_prev, tensor convtr0_partial, tensor convtr1_partial, tensor convtr2_partial, tensor latent, tensor res0_conv0_first, tensor res0_conv0_prev, tensor res0_conv1_first, tensor res0_conv1_prev, tensor res1_conv0_first, tensor res1_conv0_prev, tensor res1_conv1_first, tensor res1_conv1_prev, tensor res2_conv0_first, tensor res2_conv0_prev, tensor res2_conv1_first, tensor res2_conv1_prev, tensor upsample_partial) { + tensor emb_mean = const()[name = tensor("emb_mean"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(64)))]; + tensor emb_std = const()[name = tensor("emb_std"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(256)))]; + tensor mimi_quantizer_output_proj_weight = const()[name = tensor("mimi_quantizer_output_proj_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(448)))]; + tensor mimi_upsample_convtr_convtr_weight = const()[name = tensor("mimi_upsample_convtr_convtr_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(66048)))]; + tensor mimi_decoder_transformer_transformer_layers_0_norm1_bias = const()[name = tensor("mimi_decoder_transformer_transformer_layers_0_norm1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(131648)))]; + tensor mimi_decoder_transformer_transformer_layers_0_norm1_weight = const()[name = tensor("mimi_decoder_transformer_transformer_layers_0_norm1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(133760)))]; + tensor mimi_decoder_transformer_transformer_layers_0_self_attn_in_proj_weight = const()[name = tensor("mimi_decoder_transformer_transformer_layers_0_self_attn_in_proj_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(135872)))]; + tensor mimi_decoder_transformer_transformer_layers_0_self_attn_out_proj_weight = const()[name = tensor("mimi_decoder_transformer_transformer_layers_0_self_attn_out_proj_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(3281664)))]; + tensor mimi_decoder_transformer_transformer_layers_0_layer_scale_1_scale = const()[name = tensor("mimi_decoder_transformer_transformer_layers_0_layer_scale_1_scale"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(4330304)))]; + tensor mimi_decoder_transformer_transformer_layers_0_norm2_bias = const()[name = tensor("mimi_decoder_transformer_transformer_layers_0_norm2_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(4332416)))]; + tensor mimi_decoder_transformer_transformer_layers_0_norm2_weight = const()[name = tensor("mimi_decoder_transformer_transformer_layers_0_norm2_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(4334528)))]; + tensor mimi_decoder_transformer_transformer_layers_0_linear1_weight = const()[name = tensor("mimi_decoder_transformer_transformer_layers_0_linear1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(4336640)))]; + tensor mimi_decoder_transformer_transformer_layers_0_linear2_weight = const()[name = tensor("mimi_decoder_transformer_transformer_layers_0_linear2_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8531008)))]; + tensor mimi_decoder_transformer_transformer_layers_0_layer_scale_2_scale = const()[name = tensor("mimi_decoder_transformer_transformer_layers_0_layer_scale_2_scale"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(12725376)))]; + tensor mimi_decoder_transformer_transformer_layers_1_norm1_bias = const()[name = tensor("mimi_decoder_transformer_transformer_layers_1_norm1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(12727488)))]; + tensor mimi_decoder_transformer_transformer_layers_1_norm1_weight = const()[name = tensor("mimi_decoder_transformer_transformer_layers_1_norm1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(12729600)))]; + tensor mimi_decoder_transformer_transformer_layers_1_self_attn_in_proj_weight = const()[name = tensor("mimi_decoder_transformer_transformer_layers_1_self_attn_in_proj_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(12731712)))]; + tensor mimi_decoder_transformer_transformer_layers_1_self_attn_out_proj_weight = const()[name = tensor("mimi_decoder_transformer_transformer_layers_1_self_attn_out_proj_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15877504)))]; + tensor mimi_decoder_transformer_transformer_layers_1_layer_scale_1_scale = const()[name = tensor("mimi_decoder_transformer_transformer_layers_1_layer_scale_1_scale"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(16926144)))]; + tensor mimi_decoder_transformer_transformer_layers_1_norm2_bias = const()[name = tensor("mimi_decoder_transformer_transformer_layers_1_norm2_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(16928256)))]; + tensor mimi_decoder_transformer_transformer_layers_1_norm2_weight = const()[name = tensor("mimi_decoder_transformer_transformer_layers_1_norm2_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(16930368)))]; + tensor mimi_decoder_transformer_transformer_layers_1_linear1_weight = const()[name = tensor("mimi_decoder_transformer_transformer_layers_1_linear1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(16932480)))]; + tensor mimi_decoder_transformer_transformer_layers_1_linear2_weight = const()[name = tensor("mimi_decoder_transformer_transformer_layers_1_linear2_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21126848)))]; + tensor mimi_decoder_transformer_transformer_layers_1_layer_scale_2_scale = const()[name = tensor("mimi_decoder_transformer_transformer_layers_1_layer_scale_2_scale"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(25321216)))]; + tensor mimi_decoder_model_0_conv_bias = const()[name = tensor("mimi_decoder_model_0_conv_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(25323328)))]; + tensor mimi_decoder_model_0_conv_weight = const()[name = tensor("mimi_decoder_model_0_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(25325440)))]; + tensor mimi_decoder_model_2_convtr_bias = const()[name = tensor("mimi_decoder_model_2_convtr_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(32665536)))]; + tensor mimi_decoder_model_2_convtr_weight = const()[name = tensor("mimi_decoder_model_2_convtr_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(32666624)))]; + tensor mimi_decoder_model_3_block_1_conv_bias = const()[name = tensor("mimi_decoder_model_3_block_1_conv_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(38958144)))]; + tensor mimi_decoder_model_3_block_1_conv_weight = const()[name = tensor("mimi_decoder_model_3_block_1_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(38958720)))]; + tensor mimi_decoder_model_3_block_3_conv_bias = const()[name = tensor("mimi_decoder_model_3_block_3_conv_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(39352000)))]; + tensor mimi_decoder_model_3_block_3_conv_weight = const()[name = tensor("mimi_decoder_model_3_block_3_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(39353088)))]; + tensor mimi_decoder_model_5_convtr_bias = const()[name = tensor("mimi_decoder_model_5_convtr_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(39484224)))]; + tensor mimi_decoder_model_5_convtr_weight = const()[name = tensor("mimi_decoder_model_5_convtr_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(39484800)))]; + tensor mimi_decoder_model_6_block_1_conv_bias = const()[name = tensor("mimi_decoder_model_6_block_1_conv_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(40795584)))]; + tensor mimi_decoder_model_6_block_1_conv_weight = const()[name = tensor("mimi_decoder_model_6_block_1_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(40795904)))]; + tensor mimi_decoder_model_6_block_3_conv_bias = const()[name = tensor("mimi_decoder_model_6_block_3_conv_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(40894272)))]; + tensor mimi_decoder_model_6_block_3_conv_weight = const()[name = tensor("mimi_decoder_model_6_block_3_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(40894848)))]; + tensor mimi_decoder_model_8_convtr_bias = const()[name = tensor("mimi_decoder_model_8_convtr_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(40927680)))]; + tensor mimi_decoder_model_8_convtr_weight = const()[name = tensor("mimi_decoder_model_8_convtr_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(40928000)))]; + tensor mimi_decoder_model_9_block_1_conv_bias = const()[name = tensor("mimi_decoder_model_9_block_1_conv_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(41190208)))]; + tensor mimi_decoder_model_9_block_1_conv_weight = const()[name = tensor("mimi_decoder_model_9_block_1_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(41190400)))]; + tensor mimi_decoder_model_9_block_3_conv_bias = const()[name = tensor("mimi_decoder_model_9_block_3_conv_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(41215040)))]; + tensor mimi_decoder_model_9_block_3_conv_weight = const()[name = tensor("mimi_decoder_model_9_block_3_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(41215360)))]; + tensor mimi_decoder_model_11_conv_bias = const()[name = tensor("mimi_decoder_model_11_conv_bias"), val = tensor([0x1.46p-11])]; + tensor mimi_decoder_model_11_conv_weight = const()[name = tensor("mimi_decoder_model_11_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(41223616)))]; + tensor var_38 = mul(x = latent, y = emb_std)[name = tensor("op_38")]; + tensor denorm = add(x = var_38, y = emb_mean)[name = tensor("denorm")]; + tensor input_1_axes_0 = const()[name = tensor("input_1_axes_0"), val = tensor([-1])]; + tensor input_1 = expand_dims(axes = input_1_axes_0, x = denorm)[name = tensor("input_1")]; + tensor x_1_pad_type_0 = const()[name = tensor("x_1_pad_type_0"), val = tensor("valid")]; + tensor x_1_strides_0 = const()[name = tensor("x_1_strides_0"), val = tensor([1])]; + tensor x_1_pad_0 = const()[name = tensor("x_1_pad_0"), val = tensor([0, 0])]; + tensor x_1_dilations_0 = const()[name = tensor("x_1_dilations_0"), val = tensor([1])]; + tensor x_1_groups_0 = const()[name = tensor("x_1_groups_0"), val = tensor(1)]; + tensor x_1 = conv(dilations = x_1_dilations_0, groups = x_1_groups_0, pad = x_1_pad_0, pad_type = x_1_pad_type_0, strides = x_1_strides_0, weight = mimi_quantizer_output_proj_weight, x = input_1)[name = tensor("x_1")]; + tensor var_62 = const()[name = tensor("op_62"), val = tensor(-1)]; + tensor y_1_pad_type_0 = const()[name = tensor("y_1_pad_type_0"), val = tensor("valid")]; + tensor y_1_strides_0 = const()[name = tensor("y_1_strides_0"), val = tensor([16])]; + tensor y_1_groups_0 = const()[name = tensor("y_1_groups_0"), val = tensor(512)]; + tensor y_1_pad_0 = const()[name = tensor("y_1_pad_0"), val = tensor([0, 0])]; + tensor y_1_dilations_0 = const()[name = tensor("y_1_dilations_0"), val = tensor([1])]; + tensor y_1_has_output_shape_output_shape_0 = const()[name = tensor("y_1_has_output_shape_output_shape_0"), val = tensor([1, 512, 32])]; + tensor y_1_has_output_shape = conv_transpose(dilations = y_1_dilations_0, groups = y_1_groups_0, output_shape = y_1_has_output_shape_output_shape_0, pad = y_1_pad_0, pad_type = y_1_pad_type_0, strides = y_1_strides_0, weight = mimi_upsample_convtr_convtr_weight, x = x_1)[name = tensor("y_1_has_output_shape")]; + tensor var_72_begin_0 = const()[name = tensor("op_72_begin_0"), val = tensor([0, 0, 0])]; + tensor var_72_end_0 = const()[name = tensor("op_72_end_0"), val = tensor([1, 512, 16])]; + tensor var_72_end_mask_0 = const()[name = tensor("op_72_end_mask_0"), val = tensor([true, true, false])]; + tensor var_72 = slice_by_index(begin = var_72_begin_0, end = var_72_end_0, end_mask = var_72_end_mask_0, x = y_1_has_output_shape)[name = tensor("op_72")]; + tensor var_73 = add(x = var_72, y = upsample_partial)[name = tensor("op_73")]; + tensor var_74_begin_0 = const()[name = tensor("op_74_begin_0"), val = tensor([0, 0, 16])]; + tensor var_74_end_0 = const()[name = tensor("op_74_end_0"), val = tensor([1, 512, 32])]; + tensor var_74_end_mask_0 = const()[name = tensor("op_74_end_mask_0"), val = tensor([true, true, true])]; + tensor var_74 = slice_by_index(begin = var_74_begin_0, end = var_74_end_0, end_mask = var_74_end_mask_0, x = y_1_has_output_shape)[name = tensor("op_74")]; + tensor y_3_interleave_0 = const()[name = tensor("y_3_interleave_0"), val = tensor(false)]; + tensor y_3 = concat(axis = var_62, interleave = y_3_interleave_0, values = (var_73, var_74))[name = tensor("y_3")]; + tensor var_77_begin_0 = const()[name = tensor("op_77_begin_0"), val = tensor([0, 0, 16])]; + tensor var_77_end_0 = const()[name = tensor("op_77_end_0"), val = tensor([1, 512, 32])]; + tensor var_77_end_mask_0 = const()[name = tensor("op_77_end_mask_0"), val = tensor([true, true, true])]; + tensor var_77 = slice_by_index(begin = var_77_begin_0, end = var_77_end_0, end_mask = var_77_end_mask_0, x = y_3)[name = tensor("op_77")]; + tensor x_3_begin_0 = const()[name = tensor("x_3_begin_0"), val = tensor([0, 0, 0])]; + tensor x_3_end_0 = const()[name = tensor("x_3_end_0"), val = tensor([1, 512, 16])]; + tensor x_3_end_mask_0 = const()[name = tensor("x_3_end_mask_0"), val = tensor([true, true, false])]; + tensor x_3 = slice_by_index(begin = x_3_begin_0, end = x_3_end_0, end_mask = x_3_end_mask_0, x = y_3)[name = tensor("x_3")]; + tensor var_86 = const()[name = tensor("op_86"), val = tensor(0)]; + tensor var_91 = const()[name = tensor("op_91"), val = tensor(-1)]; + tensor var_100 = const()[name = tensor("op_100"), val = tensor(-0x1.ff933cp+127)]; + tensor var_102 = const()[name = tensor("op_102"), val = tensor(0x1.4f8b58p-17)]; + tensor input_3_perm_0 = const()[name = tensor("input_3_perm_0"), val = tensor([0, 2, 1])]; + tensor query_1_axes_0 = const()[name = tensor("query_1_axes_0"), val = tensor([-1])]; + tensor input_3 = transpose(perm = input_3_perm_0, x = x_3)[name = tensor("transpose_19")]; + tensor query_1 = layer_norm(axes = query_1_axes_0, beta = mimi_decoder_transformer_transformer_layers_0_norm1_bias, epsilon = var_102, gamma = mimi_decoder_transformer_transformer_layers_0_norm1_weight, x = input_3)[name = tensor("query_1")]; + tensor linear_0_bias_0 = const()[name = tensor("linear_0_bias_0"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(41224448)))]; + tensor projected_1 = linear(bias = linear_0_bias_0, weight = mimi_decoder_transformer_transformer_layers_0_self_attn_in_proj_weight, x = query_1)[name = tensor("linear_0")]; + tensor var_130 = const()[name = tensor("op_130"), val = tensor([1, 16, 3, 8, 64])]; + tensor packed_1 = reshape(shape = var_130, x = projected_1)[name = tensor("packed_1")]; + tensor var_132_split_sizes_0 = const()[name = tensor("op_132_split_sizes_0"), val = tensor([1, 1, 1])]; + tensor var_132_axis_0 = const()[name = tensor("op_132_axis_0"), val = tensor(2)]; + tensor var_132_0, tensor var_132_1, tensor var_132_2 = split(axis = var_132_axis_0, split_sizes = var_132_split_sizes_0, x = packed_1)[name = tensor("op_132")]; + tensor squeeze_0_axes_0 = const()[name = tensor("squeeze_0_axes_0"), val = tensor([2])]; + tensor squeeze_0 = squeeze(axes = squeeze_0_axes_0, x = var_132_0)[name = tensor("squeeze_0")]; + tensor squeeze_1_axes_0 = const()[name = tensor("squeeze_1_axes_0"), val = tensor([2])]; + tensor squeeze_1 = squeeze(axes = squeeze_1_axes_0, x = var_132_1)[name = tensor("squeeze_1")]; + tensor squeeze_2_axes_0 = const()[name = tensor("squeeze_2_axes_0"), val = tensor([2])]; + tensor squeeze_2 = squeeze(axes = squeeze_2_axes_0, x = var_132_2)[name = tensor("squeeze_2")]; + tensor offset_3_begin_0 = const()[name = tensor("offset_3_begin_0"), val = tensor([0])]; + tensor offset_3_end_0 = const()[name = tensor("offset_3_end_0"), val = tensor([1])]; + tensor offset_3_end_mask_0 = const()[name = tensor("offset_3_end_mask_0"), val = tensor([false])]; + tensor offset_3_squeeze_mask_0 = const()[name = tensor("offset_3_squeeze_mask_0"), val = tensor([true])]; + tensor offset_3 = slice_by_index(begin = offset_3_begin_0, end = offset_3_end_0, end_mask = offset_3_end_mask_0, squeeze_mask = offset_3_squeeze_mask_0, x = attn0_offset)[name = tensor("offset_3")]; + tensor freqs_1 = const()[name = tensor("freqs_1"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(41230656)))]; + tensor ts_1_promoted = const()[name = tensor("ts_1_promoted"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(41230848)))]; + tensor ts_3 = add(x = ts_1_promoted, y = offset_3)[name = tensor("ts_3")]; + tensor var_148 = const()[name = tensor("op_148"), val = tensor([-1, 1, 1])]; + tensor ts_5 = reshape(shape = var_148, x = ts_3)[name = tensor("ts_5")]; + tensor var_150 = const()[name = tensor("op_150"), val = tensor([1, 16, 8, 32, 2])]; + tensor q_3 = reshape(shape = var_150, x = squeeze_0)[name = tensor("q_3")]; + tensor var_152 = const()[name = tensor("op_152"), val = tensor([1, 16, 8, 32, 2])]; + tensor k_3 = reshape(shape = var_152, x = squeeze_1)[name = tensor("k_3")]; + tensor var_154_begin_0 = const()[name = tensor("op_154_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_154_end_0 = const()[name = tensor("op_154_end_0"), val = tensor([1, 16, 8, 32, 1])]; + tensor var_154_end_mask_0 = const()[name = tensor("op_154_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_154_squeeze_mask_0 = const()[name = tensor("op_154_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_154 = slice_by_index(begin = var_154_begin_0, end = var_154_end_0, end_mask = var_154_end_mask_0, squeeze_mask = var_154_squeeze_mask_0, x = q_3)[name = tensor("op_154")]; + tensor var_156_begin_0 = const()[name = tensor("op_156_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_156_end_0 = const()[name = tensor("op_156_end_0"), val = tensor([1, 16, 8, 32, 2])]; + tensor var_156_end_mask_0 = const()[name = tensor("op_156_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_156_squeeze_mask_0 = const()[name = tensor("op_156_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_156 = slice_by_index(begin = var_156_begin_0, end = var_156_end_0, end_mask = var_156_end_mask_0, squeeze_mask = var_156_squeeze_mask_0, x = q_3)[name = tensor("op_156")]; + tensor var_158_begin_0 = const()[name = tensor("op_158_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_158_end_0 = const()[name = tensor("op_158_end_0"), val = tensor([1, 16, 8, 32, 1])]; + tensor var_158_end_mask_0 = const()[name = tensor("op_158_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_158_squeeze_mask_0 = const()[name = tensor("op_158_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_158 = slice_by_index(begin = var_158_begin_0, end = var_158_end_0, end_mask = var_158_end_mask_0, squeeze_mask = var_158_squeeze_mask_0, x = k_3)[name = tensor("op_158")]; + tensor var_160_begin_0 = const()[name = tensor("op_160_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_160_end_0 = const()[name = tensor("op_160_end_0"), val = tensor([1, 16, 8, 32, 2])]; + tensor var_160_end_mask_0 = const()[name = tensor("op_160_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_160_squeeze_mask_0 = const()[name = tensor("op_160_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_160 = slice_by_index(begin = var_160_begin_0, end = var_160_end_0, end_mask = var_160_end_mask_0, squeeze_mask = var_160_squeeze_mask_0, x = k_3)[name = tensor("op_160")]; + tensor var_162 = mul(x = freqs_1, y = ts_5)[name = tensor("op_162")]; + tensor rotr_1 = cos(x = var_162)[name = tensor("rotr_1")]; + tensor roti_1 = sin(x = var_162)[name = tensor("roti_1")]; + tensor var_166 = mul(x = var_154, y = rotr_1)[name = tensor("op_166")]; + tensor var_167 = mul(x = var_156, y = roti_1)[name = tensor("op_167")]; + tensor qor_1 = sub(x = var_166, y = var_167)[name = tensor("qor_1")]; + tensor var_169 = mul(x = var_154, y = roti_1)[name = tensor("op_169")]; + tensor var_170 = mul(x = var_156, y = rotr_1)[name = tensor("op_170")]; + tensor qoi_1 = add(x = var_169, y = var_170)[name = tensor("qoi_1")]; + tensor var_172 = mul(x = var_158, y = rotr_1)[name = tensor("op_172")]; + tensor var_173 = mul(x = var_160, y = roti_1)[name = tensor("op_173")]; + tensor kor_1 = sub(x = var_172, y = var_173)[name = tensor("kor_1")]; + tensor var_175 = mul(x = var_158, y = roti_1)[name = tensor("op_175")]; + tensor var_176 = mul(x = var_160, y = rotr_1)[name = tensor("op_176")]; + tensor koi_1 = add(x = var_175, y = var_176)[name = tensor("koi_1")]; + tensor qo_1_axis_0 = const()[name = tensor("qo_1_axis_0"), val = tensor(-1)]; + tensor qo_1 = stack(axis = qo_1_axis_0, values = (qor_1, qoi_1))[name = tensor("qo_1")]; + tensor ko_1_axis_0 = const()[name = tensor("ko_1_axis_0"), val = tensor(-1)]; + tensor ko_1 = stack(axis = ko_1_axis_0, values = (kor_1, koi_1))[name = tensor("ko_1")]; + tensor var_186 = const()[name = tensor("op_186"), val = tensor([1, 16, 8, 64])]; + tensor q_5 = reshape(shape = var_186, x = qo_1)[name = tensor("q_5")]; + tensor var_188 = const()[name = tensor("op_188"), val = tensor([1, 16, 8, 64])]; + tensor k_5 = reshape(shape = var_188, x = ko_1)[name = tensor("k_5")]; + tensor capacity_1 = const()[name = tensor("capacity_1"), val = tensor([256])]; + tensor var_193_dtype_0 = const()[name = tensor("op_193_dtype_0"), val = tensor("int32")]; + tensor var_194 = const()[name = tensor("op_194"), val = tensor([1, 1])]; + tensor var_193 = cast(dtype = var_193_dtype_0, x = attn0_offset)[name = tensor("cast_49")]; + tensor write_base_1 = reshape(shape = var_194, x = var_193)[name = tensor("write_base_1")]; + tensor write_range_1 = const()[name = tensor("write_range_1"), val = tensor([[0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15]])]; + tensor abs_idx_1 = add(x = write_base_1, y = write_range_1)[name = tensor("abs_idx_1")]; + tensor wrapped_1_div = floor_div(x = abs_idx_1, y = capacity_1)[name = tensor("wrapped_1_div")]; + tensor wrapped_1_div_scaled = mul(x = wrapped_1_div, y = capacity_1)[name = tensor("wrapped_1_div_scaled")]; + tensor wrapped_1 = sub(x = abs_idx_1, y = wrapped_1_div_scaled)[name = tensor("wrapped_1")]; + tensor var_201 = const()[name = tensor("op_201"), val = tensor([1, 16, 1, 1])]; + tensor var_202 = reshape(shape = var_201, x = wrapped_1)[name = tensor("op_202")]; + tensor write_indexes_1_reps_0 = const()[name = tensor("write_indexes_1_reps_0"), val = tensor([1, 1, 8, 64])]; + tensor write_indexes_1 = tile(reps = write_indexes_1_reps_0, x = var_202)[name = tensor("write_indexes_1")]; + tensor var_205_begin_0 = const()[name = tensor("op_205_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_205_end_0 = const()[name = tensor("op_205_end_0"), val = tensor([1, 1, 256, 8, 64])]; + tensor var_205_end_mask_0 = const()[name = tensor("op_205_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_205_squeeze_mask_0 = const()[name = tensor("op_205_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor var_205 = slice_by_index(begin = var_205_begin_0, end = var_205_end_0, end_mask = var_205_end_mask_0, squeeze_mask = var_205_squeeze_mask_0, x = attn0_cache)[name = tensor("op_205")]; + tensor new_k_cache_1_axis_0 = const()[name = tensor("new_k_cache_1_axis_0"), val = tensor(1)]; + tensor new_k_cache_1_mode_0 = const()[name = tensor("new_k_cache_1_mode_0"), val = tensor("update")]; + tensor new_k_cache_1_validate_indices_0 = const()[name = tensor("new_k_cache_1_validate_indices_0"), val = tensor(false)]; + tensor new_k_cache_1 = scatter_along_axis(axis = new_k_cache_1_axis_0, data = var_205, indices = write_indexes_1, mode = new_k_cache_1_mode_0, updates = k_5, validate_indices = new_k_cache_1_validate_indices_0)[name = tensor("new_k_cache_1")]; + tensor var_207_begin_0 = const()[name = tensor("op_207_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor var_207_end_0 = const()[name = tensor("op_207_end_0"), val = tensor([2, 1, 256, 8, 64])]; + tensor var_207_end_mask_0 = const()[name = tensor("op_207_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_207_squeeze_mask_0 = const()[name = tensor("op_207_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor var_207 = slice_by_index(begin = var_207_begin_0, end = var_207_end_0, end_mask = var_207_end_mask_0, squeeze_mask = var_207_squeeze_mask_0, x = attn0_cache)[name = tensor("op_207")]; + tensor new_v_cache_1_axis_0 = const()[name = tensor("new_v_cache_1_axis_0"), val = tensor(1)]; + tensor new_v_cache_1_mode_0 = const()[name = tensor("new_v_cache_1_mode_0"), val = tensor("update")]; + tensor new_v_cache_1_validate_indices_0 = const()[name = tensor("new_v_cache_1_validate_indices_0"), val = tensor(false)]; + tensor new_v_cache_1 = scatter_along_axis(axis = new_v_cache_1_axis_0, data = var_207, indices = write_indexes_1, mode = new_v_cache_1_mode_0, updates = squeeze_2, validate_indices = new_v_cache_1_validate_indices_0)[name = tensor("new_v_cache_1")]; + tensor var_210_axis_0 = const()[name = tensor("op_210_axis_0"), val = tensor(0)]; + tensor var_210 = stack(axis = var_210_axis_0, values = (new_k_cache_1, new_v_cache_1))[name = tensor("op_210")]; + tensor var_211 = not_equal(x = new_k_cache_1, y = new_k_cache_1)[name = tensor("op_211")]; + tensor var_212 = const()[name = tensor("op_212"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(41230976)))]; + tensor new_k_cache_3 = select(a = var_212, b = new_k_cache_1, cond = var_211)[name = tensor("new_k_cache_3")]; + tensor var_214 = not_equal(x = new_v_cache_1, y = new_v_cache_1)[name = tensor("op_214")]; + tensor new_v_cache_3 = select(a = var_212, b = new_v_cache_1, cond = var_214)[name = tensor("new_v_cache_3")]; + tensor var_219 = const()[name = tensor("op_219"), val = tensor([0, 2, 1, 3])]; + tensor var_221 = const()[name = tensor("op_221"), val = tensor([1, 1])]; + tensor var_222 = reshape(shape = var_221, x = attn0_offset)[name = tensor("op_222")]; + tensor var_224_promoted = const()[name = tensor("op_224_promoted"), val = tensor([0x1.ep+3])]; + tensor var_225 = add(x = var_222, y = var_224_promoted)[name = tensor("op_225")]; + tensor last_pos_1_dtype_0 = const()[name = tensor("last_pos_1_dtype_0"), val = tensor("int32")]; + tensor slot_idx_1 = const()[name = tensor("slot_idx_1"), val = tensor([[0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19, 20, 21, 22, 23, 24, 25, 26, 27, 28, 29, 30, 31, 32, 33, 34, 35, 36, 37, 38, 39, 40, 41, 42, 43, 44, 45, 46, 47, 48, 49, 50, 51, 52, 53, 54, 55, 56, 57, 58, 59, 60, 61, 62, 63, 64, 65, 66, 67, 68, 69, 70, 71, 72, 73, 74, 75, 76, 77, 78, 79, 80, 81, 82, 83, 84, 85, 86, 87, 88, 89, 90, 91, 92, 93, 94, 95, 96, 97, 98, 99, 100, 101, 102, 103, 104, 105, 106, 107, 108, 109, 110, 111, 112, 113, 114, 115, 116, 117, 118, 119, 120, 121, 122, 123, 124, 125, 126, 127, 128, 129, 130, 131, 132, 133, 134, 135, 136, 137, 138, 139, 140, 141, 142, 143, 144, 145, 146, 147, 148, 149, 150, 151, 152, 153, 154, 155, 156, 157, 158, 159, 160, 161, 162, 163, 164, 165, 166, 167, 168, 169, 170, 171, 172, 173, 174, 175, 176, 177, 178, 179, 180, 181, 182, 183, 184, 185, 186, 187, 188, 189, 190, 191, 192, 193, 194, 195, 196, 197, 198, 199, 200, 201, 202, 203, 204, 205, 206, 207, 208, 209, 210, 211, 212, 213, 214, 215, 216, 217, 218, 219, 220, 221, 222, 223, 224, 225, 226, 227, 228, 229, 230, 231, 232, 233, 234, 235, 236, 237, 238, 239, 240, 241, 242, 243, 244, 245, 246, 247, 248, 249, 250, 251, 252, 253, 254, 255]])]; + tensor last_pos_1 = cast(dtype = last_pos_1_dtype_0, x = var_225)[name = tensor("cast_48")]; + tensor diff_1 = sub(x = last_pos_1, y = slot_idx_1)[name = tensor("diff_1")]; + tensor var_231_div = floor_div(x = diff_1, y = capacity_1)[name = tensor("op_231_div")]; + tensor var_231_div_scaled = mul(x = var_231_div, y = capacity_1)[name = tensor("op_231_div_scaled")]; + tensor var_231 = sub(x = diff_1, y = var_231_div_scaled)[name = tensor("op_231")]; + tensor pos_k_1 = sub(x = last_pos_1, y = var_231)[name = tensor("pos_k_1")]; + tensor var_237_promoted = const()[name = tensor("op_237_promoted"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(41755328)))]; + tensor pos_q_1 = add(x = var_222, y = var_237_promoted)[name = tensor("pos_q_1")]; + tensor var_241_axes_0 = const()[name = tensor("op_241_axes_0"), val = tensor([2])]; + tensor var_241 = expand_dims(axes = var_241_axes_0, x = pos_q_1)[name = tensor("op_241")]; + tensor var_243_axes_0 = const()[name = tensor("op_243_axes_0"), val = tensor([1])]; + tensor var_243 = expand_dims(axes = var_243_axes_0, x = pos_k_1)[name = tensor("op_243")]; + tensor var_244_promoted_dtype_0 = const()[name = tensor("op_244_promoted_dtype_0"), val = tensor("fp32")]; + tensor var_244_promoted = cast(dtype = var_244_promoted_dtype_0, x = var_243)[name = tensor("cast_47")]; + tensor delta_1 = sub(x = var_241, y = var_244_promoted)[name = tensor("delta_1")]; + tensor valid_1 = greater_equal(x = var_243, y = var_86)[name = tensor("valid_1")]; + tensor var_253 = const()[name = tensor("op_253"), val = tensor([1, 1, 1])]; + tensor var_254 = reshape(shape = var_253, x = attn0_offset)[name = tensor("op_254")]; + tensor var_256_promoted = const()[name = tensor("op_256_promoted"), val = tensor([0x1.ep+3])]; + tensor var_257 = add(x = var_254, y = var_256_promoted)[name = tensor("op_257")]; + tensor var_258 = less_equal(x = var_244_promoted, y = var_257)[name = tensor("op_258")]; + tensor valid_3 = logical_and(x = valid_1, y = var_258)[name = tensor("valid_3")]; + tensor var_86_promoted = const()[name = tensor("op_86_promoted"), val = tensor(0x0p+0)]; + tensor var_260 = greater_equal(x = delta_1, y = var_86_promoted)[name = tensor("op_260")]; + tensor attn_mask_1 = logical_and(x = valid_3, y = var_260)[name = tensor("attn_mask_1")]; + tensor var_98_promoted = const()[name = tensor("op_98_promoted"), val = tensor(0x1.f4p+7)]; + tensor var_262 = less(x = delta_1, y = var_98_promoted)[name = tensor("op_262")]; + tensor attn_mask_3 = logical_and(x = attn_mask_1, y = var_262)[name = tensor("attn_mask_3")]; + tensor attn_mask_5_axes_0 = const()[name = tensor("attn_mask_5_axes_0"), val = tensor([1])]; + tensor attn_mask_5 = expand_dims(axes = attn_mask_5_axes_0, x = attn_mask_3)[name = tensor("attn_mask_5")]; + tensor var_267_transpose_x_0 = const()[name = tensor("op_267_transpose_x_0"), val = tensor(false)]; + tensor var_267_transpose_y_0 = const()[name = tensor("op_267_transpose_y_0"), val = tensor(false)]; + tensor transpose_6_perm_0 = const()[name = tensor("transpose_6_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_7_perm_0 = const()[name = tensor("transpose_7_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_7 = transpose(perm = transpose_7_perm_0, x = new_k_cache_3)[name = tensor("transpose_16")]; + tensor transpose_6 = transpose(perm = transpose_6_perm_0, x = q_5)[name = tensor("transpose_17")]; + tensor var_267 = matmul(transpose_x = var_267_transpose_x_0, transpose_y = var_267_transpose_y_0, x = transpose_6, y = transpose_7)[name = tensor("op_267")]; + tensor var_268 = const()[name = tensor("op_268"), val = tensor(0x1p-3)]; + tensor attn_1 = mul(x = var_267, y = var_268)[name = tensor("attn_1")]; + tensor var_270 = logical_not(x = attn_mask_5)[name = tensor("op_270")]; + tensor attn_3 = select(a = var_100, b = attn_1, cond = var_270)[name = tensor("attn_3")]; + tensor attn_5 = softmax(axis = var_91, x = attn_3)[name = tensor("attn_5")]; + tensor x_5_transpose_x_0 = const()[name = tensor("x_5_transpose_x_0"), val = tensor(false)]; + tensor x_5_transpose_y_0 = const()[name = tensor("x_5_transpose_y_0"), val = tensor(false)]; + tensor v_attn_1 = transpose(perm = var_219, x = new_v_cache_3)[name = tensor("transpose_18")]; + tensor x_5 = matmul(transpose_x = x_5_transpose_x_0, transpose_y = x_5_transpose_y_0, x = attn_5, y = v_attn_1)[name = tensor("x_5")]; + tensor var_274_perm_0 = const()[name = tensor("op_274_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_275 = const()[name = tensor("op_275"), val = tensor([1, 16, 512])]; + tensor var_274 = transpose(perm = var_274_perm_0, x = x_5)[name = tensor("transpose_15")]; + tensor input_5 = reshape(shape = var_275, x = var_274)[name = tensor("input_5")]; + tensor linear_1_bias_0 = const()[name = tensor("linear_1_bias_0"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(41755456)))]; + tensor x_7 = linear(bias = linear_1_bias_0, weight = mimi_decoder_transformer_transformer_layers_0_self_attn_out_proj_weight, x = input_5)[name = tensor("linear_1")]; + tensor var_284 = mul(x = mimi_decoder_transformer_transformer_layers_0_layer_scale_1_scale, y = x_7)[name = tensor("op_284")]; + tensor input_7 = add(x = input_3, y = var_284)[name = tensor("input_7")]; + tensor input_9_axes_0 = const()[name = tensor("input_9_axes_0"), val = tensor([-1])]; + tensor input_9 = layer_norm(axes = input_9_axes_0, beta = mimi_decoder_transformer_transformer_layers_0_norm2_bias, epsilon = var_102, gamma = mimi_decoder_transformer_transformer_layers_0_norm2_weight, x = input_7)[name = tensor("input_9")]; + tensor linear_2_bias_0 = const()[name = tensor("linear_2_bias_0"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(41757568)))]; + tensor var_291 = linear(bias = linear_2_bias_0, weight = mimi_decoder_transformer_transformer_layers_0_linear1_weight, x = input_9)[name = tensor("linear_2")]; + tensor input_11_mode_0 = const()[name = tensor("input_11_mode_0"), val = tensor("EXACT")]; + tensor input_11 = gelu(mode = input_11_mode_0, x = var_291)[name = tensor("input_11")]; + tensor x_9 = linear(bias = linear_1_bias_0, weight = mimi_decoder_transformer_transformer_layers_0_linear2_weight, x = input_11)[name = tensor("linear_3")]; + tensor var_297 = mul(x = mimi_decoder_transformer_transformer_layers_0_layer_scale_2_scale, y = x_9)[name = tensor("op_297")]; + tensor input_13 = add(x = input_7, y = var_297)[name = tensor("input_13")]; + tensor query_axes_0 = const()[name = tensor("query_axes_0"), val = tensor([-1])]; + tensor query = layer_norm(axes = query_axes_0, beta = mimi_decoder_transformer_transformer_layers_1_norm1_bias, epsilon = var_102, gamma = mimi_decoder_transformer_transformer_layers_1_norm1_weight, x = input_13)[name = tensor("query")]; + tensor projected = linear(bias = linear_0_bias_0, weight = mimi_decoder_transformer_transformer_layers_1_self_attn_in_proj_weight, x = query)[name = tensor("linear_4")]; + tensor var_320 = const()[name = tensor("op_320"), val = tensor([1, 16, 3, 8, 64])]; + tensor packed = reshape(shape = var_320, x = projected)[name = tensor("packed")]; + tensor var_322_split_sizes_0 = const()[name = tensor("op_322_split_sizes_0"), val = tensor([1, 1, 1])]; + tensor var_322_axis_0 = const()[name = tensor("op_322_axis_0"), val = tensor(2)]; + tensor var_322_0, tensor var_322_1, tensor var_322_2 = split(axis = var_322_axis_0, split_sizes = var_322_split_sizes_0, x = packed)[name = tensor("op_322")]; + tensor squeeze_3_axes_0 = const()[name = tensor("squeeze_3_axes_0"), val = tensor([2])]; + tensor squeeze_3 = squeeze(axes = squeeze_3_axes_0, x = var_322_0)[name = tensor("squeeze_3")]; + tensor squeeze_4_axes_0 = const()[name = tensor("squeeze_4_axes_0"), val = tensor([2])]; + tensor squeeze_4 = squeeze(axes = squeeze_4_axes_0, x = var_322_1)[name = tensor("squeeze_4")]; + tensor squeeze_5_axes_0 = const()[name = tensor("squeeze_5_axes_0"), val = tensor([2])]; + tensor squeeze_5 = squeeze(axes = squeeze_5_axes_0, x = var_322_2)[name = tensor("squeeze_5")]; + tensor offset_begin_0 = const()[name = tensor("offset_begin_0"), val = tensor([0])]; + tensor offset_end_0 = const()[name = tensor("offset_end_0"), val = tensor([1])]; + tensor offset_end_mask_0 = const()[name = tensor("offset_end_mask_0"), val = tensor([false])]; + tensor offset_squeeze_mask_0 = const()[name = tensor("offset_squeeze_mask_0"), val = tensor([true])]; + tensor offset = slice_by_index(begin = offset_begin_0, end = offset_end_0, end_mask = offset_end_mask_0, squeeze_mask = offset_squeeze_mask_0, x = attn1_offset)[name = tensor("offset")]; + tensor freqs = const()[name = tensor("freqs"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(41765824)))]; + tensor ts_7_promoted = const()[name = tensor("ts_7_promoted"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(41766016)))]; + tensor ts_9 = add(x = ts_7_promoted, y = offset)[name = tensor("ts_9")]; + tensor var_338 = const()[name = tensor("op_338"), val = tensor([-1, 1, 1])]; + tensor ts = reshape(shape = var_338, x = ts_9)[name = tensor("ts")]; + tensor var_340 = const()[name = tensor("op_340"), val = tensor([1, 16, 8, 32, 2])]; + tensor q_9 = reshape(shape = var_340, x = squeeze_3)[name = tensor("q_9")]; + tensor var_342 = const()[name = tensor("op_342"), val = tensor([1, 16, 8, 32, 2])]; + tensor k_9 = reshape(shape = var_342, x = squeeze_4)[name = tensor("k_9")]; + tensor var_344_begin_0 = const()[name = tensor("op_344_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_344_end_0 = const()[name = tensor("op_344_end_0"), val = tensor([1, 16, 8, 32, 1])]; + tensor var_344_end_mask_0 = const()[name = tensor("op_344_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_344_squeeze_mask_0 = const()[name = tensor("op_344_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_344 = slice_by_index(begin = var_344_begin_0, end = var_344_end_0, end_mask = var_344_end_mask_0, squeeze_mask = var_344_squeeze_mask_0, x = q_9)[name = tensor("op_344")]; + tensor var_346_begin_0 = const()[name = tensor("op_346_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_346_end_0 = const()[name = tensor("op_346_end_0"), val = tensor([1, 16, 8, 32, 2])]; + tensor var_346_end_mask_0 = const()[name = tensor("op_346_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_346_squeeze_mask_0 = const()[name = tensor("op_346_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_346 = slice_by_index(begin = var_346_begin_0, end = var_346_end_0, end_mask = var_346_end_mask_0, squeeze_mask = var_346_squeeze_mask_0, x = q_9)[name = tensor("op_346")]; + tensor var_348_begin_0 = const()[name = tensor("op_348_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_348_end_0 = const()[name = tensor("op_348_end_0"), val = tensor([1, 16, 8, 32, 1])]; + tensor var_348_end_mask_0 = const()[name = tensor("op_348_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_348_squeeze_mask_0 = const()[name = tensor("op_348_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_348 = slice_by_index(begin = var_348_begin_0, end = var_348_end_0, end_mask = var_348_end_mask_0, squeeze_mask = var_348_squeeze_mask_0, x = k_9)[name = tensor("op_348")]; + tensor var_350_begin_0 = const()[name = tensor("op_350_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_350_end_0 = const()[name = tensor("op_350_end_0"), val = tensor([1, 16, 8, 32, 2])]; + tensor var_350_end_mask_0 = const()[name = tensor("op_350_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_350_squeeze_mask_0 = const()[name = tensor("op_350_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_350 = slice_by_index(begin = var_350_begin_0, end = var_350_end_0, end_mask = var_350_end_mask_0, squeeze_mask = var_350_squeeze_mask_0, x = k_9)[name = tensor("op_350")]; + tensor var_352 = mul(x = freqs, y = ts)[name = tensor("op_352")]; + tensor rotr = cos(x = var_352)[name = tensor("rotr")]; + tensor roti = sin(x = var_352)[name = tensor("roti")]; + tensor var_356 = mul(x = var_344, y = rotr)[name = tensor("op_356")]; + tensor var_357 = mul(x = var_346, y = roti)[name = tensor("op_357")]; + tensor qor_5 = sub(x = var_356, y = var_357)[name = tensor("qor_5")]; + tensor var_359 = mul(x = var_344, y = roti)[name = tensor("op_359")]; + tensor var_360 = mul(x = var_346, y = rotr)[name = tensor("op_360")]; + tensor qoi_5 = add(x = var_359, y = var_360)[name = tensor("qoi_5")]; + tensor var_362 = mul(x = var_348, y = rotr)[name = tensor("op_362")]; + tensor var_363 = mul(x = var_350, y = roti)[name = tensor("op_363")]; + tensor kor_5 = sub(x = var_362, y = var_363)[name = tensor("kor_5")]; + tensor var_365 = mul(x = var_348, y = roti)[name = tensor("op_365")]; + tensor var_366 = mul(x = var_350, y = rotr)[name = tensor("op_366")]; + tensor koi_5 = add(x = var_365, y = var_366)[name = tensor("koi_5")]; + tensor qo_axis_0 = const()[name = tensor("qo_axis_0"), val = tensor(-1)]; + tensor qo = stack(axis = qo_axis_0, values = (qor_5, qoi_5))[name = tensor("qo")]; + tensor ko_axis_0 = const()[name = tensor("ko_axis_0"), val = tensor(-1)]; + tensor ko = stack(axis = ko_axis_0, values = (kor_5, koi_5))[name = tensor("ko")]; + tensor var_376 = const()[name = tensor("op_376"), val = tensor([1, 16, 8, 64])]; + tensor q = reshape(shape = var_376, x = qo)[name = tensor("q")]; + tensor var_378 = const()[name = tensor("op_378"), val = tensor([1, 16, 8, 64])]; + tensor k = reshape(shape = var_378, x = ko)[name = tensor("k")]; + tensor capacity = const()[name = tensor("capacity"), val = tensor([256])]; + tensor var_383_dtype_0 = const()[name = tensor("op_383_dtype_0"), val = tensor("int32")]; + tensor var_384 = const()[name = tensor("op_384"), val = tensor([1, 1])]; + tensor var_383 = cast(dtype = var_383_dtype_0, x = attn1_offset)[name = tensor("cast_46")]; + tensor write_base = reshape(shape = var_384, x = var_383)[name = tensor("write_base")]; + tensor write_range = const()[name = tensor("write_range"), val = tensor([[0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15]])]; + tensor abs_idx = add(x = write_base, y = write_range)[name = tensor("abs_idx")]; + tensor wrapped_div = floor_div(x = abs_idx, y = capacity)[name = tensor("wrapped_div")]; + tensor wrapped_div_scaled = mul(x = wrapped_div, y = capacity)[name = tensor("wrapped_div_scaled")]; + tensor wrapped = sub(x = abs_idx, y = wrapped_div_scaled)[name = tensor("wrapped")]; + tensor var_391 = const()[name = tensor("op_391"), val = tensor([1, 16, 1, 1])]; + tensor var_392 = reshape(shape = var_391, x = wrapped)[name = tensor("op_392")]; + tensor write_indexes_reps_0 = const()[name = tensor("write_indexes_reps_0"), val = tensor([1, 1, 8, 64])]; + tensor write_indexes = tile(reps = write_indexes_reps_0, x = var_392)[name = tensor("write_indexes")]; + tensor var_395_begin_0 = const()[name = tensor("op_395_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_395_end_0 = const()[name = tensor("op_395_end_0"), val = tensor([1, 1, 256, 8, 64])]; + tensor var_395_end_mask_0 = const()[name = tensor("op_395_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_395_squeeze_mask_0 = const()[name = tensor("op_395_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor var_395 = slice_by_index(begin = var_395_begin_0, end = var_395_end_0, end_mask = var_395_end_mask_0, squeeze_mask = var_395_squeeze_mask_0, x = attn1_cache)[name = tensor("op_395")]; + tensor new_k_cache_5_axis_0 = const()[name = tensor("new_k_cache_5_axis_0"), val = tensor(1)]; + tensor new_k_cache_5_mode_0 = const()[name = tensor("new_k_cache_5_mode_0"), val = tensor("update")]; + tensor new_k_cache_5_validate_indices_0 = const()[name = tensor("new_k_cache_5_validate_indices_0"), val = tensor(false)]; + tensor new_k_cache_5 = scatter_along_axis(axis = new_k_cache_5_axis_0, data = var_395, indices = write_indexes, mode = new_k_cache_5_mode_0, updates = k, validate_indices = new_k_cache_5_validate_indices_0)[name = tensor("new_k_cache_5")]; + tensor var_397_begin_0 = const()[name = tensor("op_397_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor var_397_end_0 = const()[name = tensor("op_397_end_0"), val = tensor([2, 1, 256, 8, 64])]; + tensor var_397_end_mask_0 = const()[name = tensor("op_397_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_397_squeeze_mask_0 = const()[name = tensor("op_397_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor var_397 = slice_by_index(begin = var_397_begin_0, end = var_397_end_0, end_mask = var_397_end_mask_0, squeeze_mask = var_397_squeeze_mask_0, x = attn1_cache)[name = tensor("op_397")]; + tensor new_v_cache_5_axis_0 = const()[name = tensor("new_v_cache_5_axis_0"), val = tensor(1)]; + tensor new_v_cache_5_mode_0 = const()[name = tensor("new_v_cache_5_mode_0"), val = tensor("update")]; + tensor new_v_cache_5_validate_indices_0 = const()[name = tensor("new_v_cache_5_validate_indices_0"), val = tensor(false)]; + tensor new_v_cache_5 = scatter_along_axis(axis = new_v_cache_5_axis_0, data = var_397, indices = write_indexes, mode = new_v_cache_5_mode_0, updates = squeeze_5, validate_indices = new_v_cache_5_validate_indices_0)[name = tensor("new_v_cache_5")]; + tensor var_400_axis_0 = const()[name = tensor("op_400_axis_0"), val = tensor(0)]; + tensor var_400 = stack(axis = var_400_axis_0, values = (new_k_cache_5, new_v_cache_5))[name = tensor("op_400")]; + tensor var_401 = not_equal(x = new_k_cache_5, y = new_k_cache_5)[name = tensor("op_401")]; + tensor new_k_cache = select(a = var_212, b = new_k_cache_5, cond = var_401)[name = tensor("new_k_cache")]; + tensor var_404 = not_equal(x = new_v_cache_5, y = new_v_cache_5)[name = tensor("op_404")]; + tensor new_v_cache = select(a = var_212, b = new_v_cache_5, cond = var_404)[name = tensor("new_v_cache")]; + tensor var_409 = const()[name = tensor("op_409"), val = tensor([0, 2, 1, 3])]; + tensor var_411 = const()[name = tensor("op_411"), val = tensor([1, 1])]; + tensor var_412 = reshape(shape = var_411, x = attn1_offset)[name = tensor("op_412")]; + tensor var_414_promoted = const()[name = tensor("op_414_promoted"), val = tensor([0x1.ep+3])]; + tensor var_415 = add(x = var_412, y = var_414_promoted)[name = tensor("op_415")]; + tensor last_pos_dtype_0 = const()[name = tensor("last_pos_dtype_0"), val = tensor("int32")]; + tensor last_pos = cast(dtype = last_pos_dtype_0, x = var_415)[name = tensor("cast_45")]; + tensor diff = sub(x = last_pos, y = slot_idx_1)[name = tensor("diff")]; + tensor var_421_div = floor_div(x = diff, y = capacity)[name = tensor("op_421_div")]; + tensor var_421_div_scaled = mul(x = var_421_div, y = capacity)[name = tensor("op_421_div_scaled")]; + tensor var_421 = sub(x = diff, y = var_421_div_scaled)[name = tensor("op_421")]; + tensor pos_k = sub(x = last_pos, y = var_421)[name = tensor("pos_k")]; + tensor var_427_promoted = const()[name = tensor("op_427_promoted"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(41766144)))]; + tensor pos_q = add(x = var_412, y = var_427_promoted)[name = tensor("pos_q")]; + tensor var_431_axes_0 = const()[name = tensor("op_431_axes_0"), val = tensor([2])]; + tensor var_431 = expand_dims(axes = var_431_axes_0, x = pos_q)[name = tensor("op_431")]; + tensor var_433_axes_0 = const()[name = tensor("op_433_axes_0"), val = tensor([1])]; + tensor var_433 = expand_dims(axes = var_433_axes_0, x = pos_k)[name = tensor("op_433")]; + tensor var_434_promoted_dtype_0 = const()[name = tensor("op_434_promoted_dtype_0"), val = tensor("fp32")]; + tensor var_434_promoted = cast(dtype = var_434_promoted_dtype_0, x = var_433)[name = tensor("cast_44")]; + tensor delta = sub(x = var_431, y = var_434_promoted)[name = tensor("delta")]; + tensor valid_5 = greater_equal(x = var_433, y = var_86)[name = tensor("valid_5")]; + tensor var_443 = const()[name = tensor("op_443"), val = tensor([1, 1, 1])]; + tensor var_444 = reshape(shape = var_443, x = attn1_offset)[name = tensor("op_444")]; + tensor var_446_promoted = const()[name = tensor("op_446_promoted"), val = tensor([0x1.ep+3])]; + tensor var_447 = add(x = var_444, y = var_446_promoted)[name = tensor("op_447")]; + tensor var_448 = less_equal(x = var_434_promoted, y = var_447)[name = tensor("op_448")]; + tensor valid = logical_and(x = valid_5, y = var_448)[name = tensor("valid")]; + tensor var_86_promoted_1 = const()[name = tensor("op_86_promoted_1"), val = tensor(0x0p+0)]; + tensor var_450 = greater_equal(x = delta, y = var_86_promoted_1)[name = tensor("op_450")]; + tensor attn_mask_7 = logical_and(x = valid, y = var_450)[name = tensor("attn_mask_7")]; + tensor var_98_promoted_1 = const()[name = tensor("op_98_promoted_1"), val = tensor(0x1.f4p+7)]; + tensor var_452 = less(x = delta, y = var_98_promoted_1)[name = tensor("op_452")]; + tensor attn_mask_9 = logical_and(x = attn_mask_7, y = var_452)[name = tensor("attn_mask_9")]; + tensor attn_mask_axes_0 = const()[name = tensor("attn_mask_axes_0"), val = tensor([1])]; + tensor attn_mask = expand_dims(axes = attn_mask_axes_0, x = attn_mask_9)[name = tensor("attn_mask")]; + tensor var_457_transpose_x_0 = const()[name = tensor("op_457_transpose_x_0"), val = tensor(false)]; + tensor var_457_transpose_y_0 = const()[name = tensor("op_457_transpose_y_0"), val = tensor(false)]; + tensor transpose_8_perm_0 = const()[name = tensor("transpose_8_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_9_perm_0 = const()[name = tensor("transpose_9_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_9 = transpose(perm = transpose_9_perm_0, x = new_k_cache)[name = tensor("transpose_12")]; + tensor transpose_8 = transpose(perm = transpose_8_perm_0, x = q)[name = tensor("transpose_13")]; + tensor var_457 = matmul(transpose_x = var_457_transpose_x_0, transpose_y = var_457_transpose_y_0, x = transpose_8, y = transpose_9)[name = tensor("op_457")]; + tensor var_458 = const()[name = tensor("op_458"), val = tensor(0x1p-3)]; + tensor attn_7 = mul(x = var_457, y = var_458)[name = tensor("attn_7")]; + tensor var_460 = logical_not(x = attn_mask)[name = tensor("op_460")]; + tensor attn_9 = select(a = var_100, b = attn_7, cond = var_460)[name = tensor("attn_9")]; + tensor attn = softmax(axis = var_91, x = attn_9)[name = tensor("attn")]; + tensor x_11_transpose_x_0 = const()[name = tensor("x_11_transpose_x_0"), val = tensor(false)]; + tensor x_11_transpose_y_0 = const()[name = tensor("x_11_transpose_y_0"), val = tensor(false)]; + tensor v_attn = transpose(perm = var_409, x = new_v_cache)[name = tensor("transpose_14")]; + tensor x_11 = matmul(transpose_x = x_11_transpose_x_0, transpose_y = x_11_transpose_y_0, x = attn, y = v_attn)[name = tensor("x_11")]; + tensor var_464_perm_0 = const()[name = tensor("op_464_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_465 = const()[name = tensor("op_465"), val = tensor([1, 16, 512])]; + tensor var_464 = transpose(perm = var_464_perm_0, x = x_11)[name = tensor("transpose_11")]; + tensor input_15 = reshape(shape = var_465, x = var_464)[name = tensor("input_15")]; + tensor x_13 = linear(bias = linear_1_bias_0, weight = mimi_decoder_transformer_transformer_layers_1_self_attn_out_proj_weight, x = input_15)[name = tensor("linear_5")]; + tensor var_474 = mul(x = mimi_decoder_transformer_transformer_layers_1_layer_scale_1_scale, y = x_13)[name = tensor("op_474")]; + tensor input_17 = add(x = input_13, y = var_474)[name = tensor("input_17")]; + tensor input_19_axes_0 = const()[name = tensor("input_19_axes_0"), val = tensor([-1])]; + tensor input_19 = layer_norm(axes = input_19_axes_0, beta = mimi_decoder_transformer_transformer_layers_1_norm2_bias, epsilon = var_102, gamma = mimi_decoder_transformer_transformer_layers_1_norm2_weight, x = input_17)[name = tensor("input_19")]; + tensor var_481 = linear(bias = linear_2_bias_0, weight = mimi_decoder_transformer_transformer_layers_1_linear1_weight, x = input_19)[name = tensor("linear_6")]; + tensor input_21_mode_0 = const()[name = tensor("input_21_mode_0"), val = tensor("EXACT")]; + tensor input_21 = gelu(mode = input_21_mode_0, x = var_481)[name = tensor("input_21")]; + tensor x_15 = linear(bias = linear_1_bias_0, weight = mimi_decoder_transformer_transformer_layers_1_linear2_weight, x = input_21)[name = tensor("linear_7")]; + tensor var_487 = mul(x = mimi_decoder_transformer_transformer_layers_1_layer_scale_2_scale, y = x_15)[name = tensor("op_487")]; + tensor z = add(x = input_17, y = var_487)[name = tensor("z")]; + tensor x_17_perm_0 = const()[name = tensor("x_17_perm_0"), val = tensor([0, 2, 1])]; + tensor var_507 = const()[name = tensor("op_507"), val = tensor(0x1p+0)]; + tensor var_508 = const()[name = tensor("op_508"), val = tensor(-1)]; + tensor input_23_interleave_0 = const()[name = tensor("input_23_interleave_0"), val = tensor(false)]; + tensor x_17 = transpose(perm = x_17_perm_0, x = z)[name = tensor("transpose_10")]; + tensor input_23 = concat(axis = var_508, interleave = input_23_interleave_0, values = (conv0_prev, x_17))[name = tensor("input_23")]; + tensor input_25_pad_type_0 = const()[name = tensor("input_25_pad_type_0"), val = tensor("valid")]; + tensor input_25_strides_0 = const()[name = tensor("input_25_strides_0"), val = tensor([1])]; + tensor input_25_pad_0 = const()[name = tensor("input_25_pad_0"), val = tensor([0, 0])]; + tensor input_25_dilations_0 = const()[name = tensor("input_25_dilations_0"), val = tensor([1])]; + tensor input_25_groups_0 = const()[name = tensor("input_25_groups_0"), val = tensor(1)]; + tensor input_25 = conv(bias = mimi_decoder_model_0_conv_bias, dilations = input_25_dilations_0, groups = input_25_groups_0, pad = input_25_pad_0, pad_type = input_25_pad_type_0, strides = input_25_strides_0, weight = mimi_decoder_model_0_conv_weight, x = input_23)[name = tensor("input_25")]; + tensor var_542_begin_0 = const()[name = tensor("op_542_begin_0"), val = tensor([0, 0, 16])]; + tensor var_542_end_0 = const()[name = tensor("op_542_end_0"), val = tensor([1, 512, 22])]; + tensor var_542_end_mask_0 = const()[name = tensor("op_542_end_mask_0"), val = tensor([true, true, true])]; + tensor var_542 = slice_by_index(begin = var_542_begin_0, end = var_542_end_0, end_mask = var_542_end_mask_0, x = input_23)[name = tensor("op_542")]; + tensor input_27 = elu(alpha = var_507, x = input_25)[name = tensor("input_27")]; + tensor y_5_pad_type_0 = const()[name = tensor("y_5_pad_type_0"), val = tensor("valid")]; + tensor y_5_strides_0 = const()[name = tensor("y_5_strides_0"), val = tensor([6])]; + tensor y_5_pad_0 = const()[name = tensor("y_5_pad_0"), val = tensor([0, 0])]; + tensor y_5_dilations_0 = const()[name = tensor("y_5_dilations_0"), val = tensor([1])]; + tensor y_5_groups_0 = const()[name = tensor("y_5_groups_0"), val = tensor(1)]; + tensor y_5_has_output_shape_output_shape_0 = const()[name = tensor("y_5_has_output_shape_output_shape_0"), val = tensor([1, 256, 102])]; + tensor y_5_has_output_shape = conv_transpose(bias = mimi_decoder_model_2_convtr_bias, dilations = y_5_dilations_0, groups = y_5_groups_0, output_shape = y_5_has_output_shape_output_shape_0, pad = y_5_pad_0, pad_type = y_5_pad_type_0, strides = y_5_strides_0, weight = mimi_decoder_model_2_convtr_weight, x = input_27)[name = tensor("y_5_has_output_shape")]; + tensor var_557_begin_0 = const()[name = tensor("op_557_begin_0"), val = tensor([0, 0, 0])]; + tensor var_557_end_0 = const()[name = tensor("op_557_end_0"), val = tensor([1, 256, 6])]; + tensor var_557_end_mask_0 = const()[name = tensor("op_557_end_mask_0"), val = tensor([true, true, false])]; + tensor var_557 = slice_by_index(begin = var_557_begin_0, end = var_557_end_0, end_mask = var_557_end_mask_0, x = y_5_has_output_shape)[name = tensor("op_557")]; + tensor var_558 = add(x = var_557, y = convtr0_partial)[name = tensor("op_558")]; + tensor var_559_begin_0 = const()[name = tensor("op_559_begin_0"), val = tensor([0, 0, 6])]; + tensor var_559_end_0 = const()[name = tensor("op_559_end_0"), val = tensor([1, 256, 102])]; + tensor var_559_end_mask_0 = const()[name = tensor("op_559_end_mask_0"), val = tensor([true, true, true])]; + tensor var_559 = slice_by_index(begin = var_559_begin_0, end = var_559_end_0, end_mask = var_559_end_mask_0, x = y_5_has_output_shape)[name = tensor("op_559")]; + tensor y_7_interleave_0 = const()[name = tensor("y_7_interleave_0"), val = tensor(false)]; + tensor y_7 = concat(axis = var_508, interleave = y_7_interleave_0, values = (var_558, var_559))[name = tensor("y_7")]; + tensor new_partial_1_begin_0 = const()[name = tensor("new_partial_1_begin_0"), val = tensor([0, 0, 96])]; + tensor new_partial_1_end_0 = const()[name = tensor("new_partial_1_end_0"), val = tensor([1, 256, 102])]; + tensor new_partial_1_end_mask_0 = const()[name = tensor("new_partial_1_end_mask_0"), val = tensor([true, true, true])]; + tensor new_partial_1 = slice_by_index(begin = new_partial_1_begin_0, end = new_partial_1_end_0, end_mask = new_partial_1_end_mask_0, x = y_7)[name = tensor("new_partial_1")]; + tensor var_564 = const()[name = tensor("op_564"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(41766272)))]; + tensor var_565 = sub(x = new_partial_1, y = var_564)[name = tensor("op_565")]; + tensor input_29_begin_0 = const()[name = tensor("input_29_begin_0"), val = tensor([0, 0, 0])]; + tensor input_29_end_0 = const()[name = tensor("input_29_end_0"), val = tensor([1, 256, 96])]; + tensor input_29_end_mask_0 = const()[name = tensor("input_29_end_mask_0"), val = tensor([true, true, false])]; + tensor input_29 = slice_by_index(begin = input_29_begin_0, end = input_29_end_0, end_mask = input_29_end_mask_0, x = y_7)[name = tensor("input_29")]; + tensor x_19 = elu(alpha = var_507, x = input_29)[name = tensor("x_19")]; + tensor input_31_interleave_0 = const()[name = tensor("input_31_interleave_0"), val = tensor(false)]; + tensor input_31 = concat(axis = var_508, interleave = input_31_interleave_0, values = (res0_conv0_prev, x_19))[name = tensor("input_31")]; + tensor input_33_pad_type_0 = const()[name = tensor("input_33_pad_type_0"), val = tensor("valid")]; + tensor input_33_strides_0 = const()[name = tensor("input_33_strides_0"), val = tensor([1])]; + tensor input_33_pad_0 = const()[name = tensor("input_33_pad_0"), val = tensor([0, 0])]; + tensor input_33_dilations_0 = const()[name = tensor("input_33_dilations_0"), val = tensor([1])]; + tensor input_33_groups_0 = const()[name = tensor("input_33_groups_0"), val = tensor(1)]; + tensor input_33 = conv(bias = mimi_decoder_model_3_block_1_conv_bias, dilations = input_33_dilations_0, groups = input_33_groups_0, pad = input_33_pad_0, pad_type = input_33_pad_type_0, strides = input_33_strides_0, weight = mimi_decoder_model_3_block_1_conv_weight, x = input_31)[name = tensor("input_33")]; + tensor var_585_begin_0 = const()[name = tensor("op_585_begin_0"), val = tensor([0, 0, 96])]; + tensor var_585_end_0 = const()[name = tensor("op_585_end_0"), val = tensor([1, 256, 98])]; + tensor var_585_end_mask_0 = const()[name = tensor("op_585_end_mask_0"), val = tensor([true, true, true])]; + tensor var_585 = slice_by_index(begin = var_585_begin_0, end = var_585_end_0, end_mask = var_585_end_mask_0, x = input_31)[name = tensor("op_585")]; + tensor x_21 = elu(alpha = var_507, x = input_33)[name = tensor("x_21")]; + tensor v_5_pad_type_0 = const()[name = tensor("v_5_pad_type_0"), val = tensor("valid")]; + tensor v_5_strides_0 = const()[name = tensor("v_5_strides_0"), val = tensor([1])]; + tensor v_5_pad_0 = const()[name = tensor("v_5_pad_0"), val = tensor([0, 0])]; + tensor v_5_dilations_0 = const()[name = tensor("v_5_dilations_0"), val = tensor([1])]; + tensor v_5_groups_0 = const()[name = tensor("v_5_groups_0"), val = tensor(1)]; + tensor v_5 = conv(bias = mimi_decoder_model_3_block_3_conv_bias, dilations = v_5_dilations_0, groups = v_5_groups_0, pad = v_5_pad_0, pad_type = v_5_pad_type_0, strides = v_5_strides_0, weight = mimi_decoder_model_3_block_3_conv_weight, x = x_21)[name = tensor("v_5")]; + tensor input_35 = add(x = input_29, y = v_5)[name = tensor("input_35")]; + tensor input_37 = elu(alpha = var_507, x = input_35)[name = tensor("input_37")]; + tensor y_9_pad_type_0 = const()[name = tensor("y_9_pad_type_0"), val = tensor("valid")]; + tensor y_9_strides_0 = const()[name = tensor("y_9_strides_0"), val = tensor([5])]; + tensor y_9_pad_0 = const()[name = tensor("y_9_pad_0"), val = tensor([0, 0])]; + tensor y_9_dilations_0 = const()[name = tensor("y_9_dilations_0"), val = tensor([1])]; + tensor y_9_groups_0 = const()[name = tensor("y_9_groups_0"), val = tensor(1)]; + tensor y_9_has_output_shape_output_shape_0 = const()[name = tensor("y_9_has_output_shape_output_shape_0"), val = tensor([1, 128, 485])]; + tensor y_9_has_output_shape = conv_transpose(bias = mimi_decoder_model_5_convtr_bias, dilations = y_9_dilations_0, groups = y_9_groups_0, output_shape = y_9_has_output_shape_output_shape_0, pad = y_9_pad_0, pad_type = y_9_pad_type_0, strides = y_9_strides_0, weight = mimi_decoder_model_5_convtr_weight, x = input_37)[name = tensor("y_9_has_output_shape")]; + tensor var_613_begin_0 = const()[name = tensor("op_613_begin_0"), val = tensor([0, 0, 0])]; + tensor var_613_end_0 = const()[name = tensor("op_613_end_0"), val = tensor([1, 128, 5])]; + tensor var_613_end_mask_0 = const()[name = tensor("op_613_end_mask_0"), val = tensor([true, true, false])]; + tensor var_613 = slice_by_index(begin = var_613_begin_0, end = var_613_end_0, end_mask = var_613_end_mask_0, x = y_9_has_output_shape)[name = tensor("op_613")]; + tensor var_614 = add(x = var_613, y = convtr1_partial)[name = tensor("op_614")]; + tensor var_615_begin_0 = const()[name = tensor("op_615_begin_0"), val = tensor([0, 0, 5])]; + tensor var_615_end_0 = const()[name = tensor("op_615_end_0"), val = tensor([1, 128, 485])]; + tensor var_615_end_mask_0 = const()[name = tensor("op_615_end_mask_0"), val = tensor([true, true, true])]; + tensor var_615 = slice_by_index(begin = var_615_begin_0, end = var_615_end_0, end_mask = var_615_end_mask_0, x = y_9_has_output_shape)[name = tensor("op_615")]; + tensor y_11_interleave_0 = const()[name = tensor("y_11_interleave_0"), val = tensor(false)]; + tensor y_11 = concat(axis = var_508, interleave = y_11_interleave_0, values = (var_614, var_615))[name = tensor("y_11")]; + tensor new_partial_3_begin_0 = const()[name = tensor("new_partial_3_begin_0"), val = tensor([0, 0, 480])]; + tensor new_partial_3_end_0 = const()[name = tensor("new_partial_3_end_0"), val = tensor([1, 128, 485])]; + tensor new_partial_3_end_mask_0 = const()[name = tensor("new_partial_3_end_mask_0"), val = tensor([true, true, true])]; + tensor new_partial_3 = slice_by_index(begin = new_partial_3_begin_0, end = new_partial_3_end_0, end_mask = new_partial_3_end_mask_0, x = y_11)[name = tensor("new_partial_3")]; + tensor var_620 = const()[name = tensor("op_620"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(41767360)))]; + tensor var_621 = sub(x = new_partial_3, y = var_620)[name = tensor("op_621")]; + tensor input_39_begin_0 = const()[name = tensor("input_39_begin_0"), val = tensor([0, 0, 0])]; + tensor input_39_end_0 = const()[name = tensor("input_39_end_0"), val = tensor([1, 128, 480])]; + tensor input_39_end_mask_0 = const()[name = tensor("input_39_end_mask_0"), val = tensor([true, true, false])]; + tensor input_39 = slice_by_index(begin = input_39_begin_0, end = input_39_end_0, end_mask = input_39_end_mask_0, x = y_11)[name = tensor("input_39")]; + tensor x_23 = elu(alpha = var_507, x = input_39)[name = tensor("x_23")]; + tensor input_41_interleave_0 = const()[name = tensor("input_41_interleave_0"), val = tensor(false)]; + tensor input_41 = concat(axis = var_508, interleave = input_41_interleave_0, values = (res1_conv0_prev, x_23))[name = tensor("input_41")]; + tensor input_43_pad_type_0 = const()[name = tensor("input_43_pad_type_0"), val = tensor("valid")]; + tensor input_43_strides_0 = const()[name = tensor("input_43_strides_0"), val = tensor([1])]; + tensor input_43_pad_0 = const()[name = tensor("input_43_pad_0"), val = tensor([0, 0])]; + tensor input_43_dilations_0 = const()[name = tensor("input_43_dilations_0"), val = tensor([1])]; + tensor input_43_groups_0 = const()[name = tensor("input_43_groups_0"), val = tensor(1)]; + tensor input_43 = conv(bias = mimi_decoder_model_6_block_1_conv_bias, dilations = input_43_dilations_0, groups = input_43_groups_0, pad = input_43_pad_0, pad_type = input_43_pad_type_0, strides = input_43_strides_0, weight = mimi_decoder_model_6_block_1_conv_weight, x = input_41)[name = tensor("input_43")]; + tensor var_641_begin_0 = const()[name = tensor("op_641_begin_0"), val = tensor([0, 0, 480])]; + tensor var_641_end_0 = const()[name = tensor("op_641_end_0"), val = tensor([1, 128, 482])]; + tensor var_641_end_mask_0 = const()[name = tensor("op_641_end_mask_0"), val = tensor([true, true, true])]; + tensor var_641 = slice_by_index(begin = var_641_begin_0, end = var_641_end_0, end_mask = var_641_end_mask_0, x = input_41)[name = tensor("op_641")]; + tensor x_25 = elu(alpha = var_507, x = input_43)[name = tensor("x_25")]; + tensor v_7_pad_type_0 = const()[name = tensor("v_7_pad_type_0"), val = tensor("valid")]; + tensor v_7_strides_0 = const()[name = tensor("v_7_strides_0"), val = tensor([1])]; + tensor v_7_pad_0 = const()[name = tensor("v_7_pad_0"), val = tensor([0, 0])]; + tensor v_7_dilations_0 = const()[name = tensor("v_7_dilations_0"), val = tensor([1])]; + tensor v_7_groups_0 = const()[name = tensor("v_7_groups_0"), val = tensor(1)]; + tensor v_7 = conv(bias = mimi_decoder_model_6_block_3_conv_bias, dilations = v_7_dilations_0, groups = v_7_groups_0, pad = v_7_pad_0, pad_type = v_7_pad_type_0, strides = v_7_strides_0, weight = mimi_decoder_model_6_block_3_conv_weight, x = x_25)[name = tensor("v_7")]; + tensor input_45 = add(x = input_39, y = v_7)[name = tensor("input_45")]; + tensor input_47 = elu(alpha = var_507, x = input_45)[name = tensor("input_47")]; + tensor y_13_pad_type_0 = const()[name = tensor("y_13_pad_type_0"), val = tensor("valid")]; + tensor y_13_strides_0 = const()[name = tensor("y_13_strides_0"), val = tensor([4])]; + tensor y_13_pad_0 = const()[name = tensor("y_13_pad_0"), val = tensor([0, 0])]; + tensor y_13_dilations_0 = const()[name = tensor("y_13_dilations_0"), val = tensor([1])]; + tensor y_13_groups_0 = const()[name = tensor("y_13_groups_0"), val = tensor(1)]; + tensor y_13_has_output_shape_output_shape_0 = const()[name = tensor("y_13_has_output_shape_output_shape_0"), val = tensor([1, 64, 1924])]; + tensor y_13_has_output_shape = conv_transpose(bias = mimi_decoder_model_8_convtr_bias, dilations = y_13_dilations_0, groups = y_13_groups_0, output_shape = y_13_has_output_shape_output_shape_0, pad = y_13_pad_0, pad_type = y_13_pad_type_0, strides = y_13_strides_0, weight = mimi_decoder_model_8_convtr_weight, x = input_47)[name = tensor("y_13_has_output_shape")]; + tensor var_669_begin_0 = const()[name = tensor("op_669_begin_0"), val = tensor([0, 0, 0])]; + tensor var_669_end_0 = const()[name = tensor("op_669_end_0"), val = tensor([1, 64, 4])]; + tensor var_669_end_mask_0 = const()[name = tensor("op_669_end_mask_0"), val = tensor([true, true, false])]; + tensor var_669 = slice_by_index(begin = var_669_begin_0, end = var_669_end_0, end_mask = var_669_end_mask_0, x = y_13_has_output_shape)[name = tensor("op_669")]; + tensor var_670 = add(x = var_669, y = convtr2_partial)[name = tensor("op_670")]; + tensor var_671_begin_0 = const()[name = tensor("op_671_begin_0"), val = tensor([0, 0, 4])]; + tensor var_671_end_0 = const()[name = tensor("op_671_end_0"), val = tensor([1, 64, 1924])]; + tensor var_671_end_mask_0 = const()[name = tensor("op_671_end_mask_0"), val = tensor([true, true, true])]; + tensor var_671 = slice_by_index(begin = var_671_begin_0, end = var_671_end_0, end_mask = var_671_end_mask_0, x = y_13_has_output_shape)[name = tensor("op_671")]; + tensor y_interleave_0 = const()[name = tensor("y_interleave_0"), val = tensor(false)]; + tensor y = concat(axis = var_508, interleave = y_interleave_0, values = (var_670, var_671))[name = tensor("y")]; + tensor new_partial_begin_0 = const()[name = tensor("new_partial_begin_0"), val = tensor([0, 0, 1920])]; + tensor new_partial_end_0 = const()[name = tensor("new_partial_end_0"), val = tensor([1, 64, 1924])]; + tensor new_partial_end_mask_0 = const()[name = tensor("new_partial_end_mask_0"), val = tensor([true, true, true])]; + tensor new_partial = slice_by_index(begin = new_partial_begin_0, end = new_partial_end_0, end_mask = new_partial_end_mask_0, x = y)[name = tensor("new_partial")]; + tensor var_676 = const()[name = tensor("op_676"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(41767936)))]; + tensor var_677 = sub(x = new_partial, y = var_676)[name = tensor("op_677")]; + tensor input_49_begin_0 = const()[name = tensor("input_49_begin_0"), val = tensor([0, 0, 0])]; + tensor input_49_end_0 = const()[name = tensor("input_49_end_0"), val = tensor([1, 64, 1920])]; + tensor input_49_end_mask_0 = const()[name = tensor("input_49_end_mask_0"), val = tensor([true, true, false])]; + tensor input_49 = slice_by_index(begin = input_49_begin_0, end = input_49_end_0, end_mask = input_49_end_mask_0, x = y)[name = tensor("input_49")]; + tensor x_27 = elu(alpha = var_507, x = input_49)[name = tensor("x_27")]; + tensor input_51_interleave_0 = const()[name = tensor("input_51_interleave_0"), val = tensor(false)]; + tensor input_51 = concat(axis = var_508, interleave = input_51_interleave_0, values = (res2_conv0_prev, x_27))[name = tensor("input_51")]; + tensor input_53_pad_type_0 = const()[name = tensor("input_53_pad_type_0"), val = tensor("valid")]; + tensor input_53_strides_0 = const()[name = tensor("input_53_strides_0"), val = tensor([1])]; + tensor input_53_pad_0 = const()[name = tensor("input_53_pad_0"), val = tensor([0, 0])]; + tensor input_53_dilations_0 = const()[name = tensor("input_53_dilations_0"), val = tensor([1])]; + tensor input_53_groups_0 = const()[name = tensor("input_53_groups_0"), val = tensor(1)]; + tensor input_53 = conv(bias = mimi_decoder_model_9_block_1_conv_bias, dilations = input_53_dilations_0, groups = input_53_groups_0, pad = input_53_pad_0, pad_type = input_53_pad_type_0, strides = input_53_strides_0, weight = mimi_decoder_model_9_block_1_conv_weight, x = input_51)[name = tensor("input_53")]; + tensor var_697_begin_0 = const()[name = tensor("op_697_begin_0"), val = tensor([0, 0, 1920])]; + tensor var_697_end_0 = const()[name = tensor("op_697_end_0"), val = tensor([1, 64, 1922])]; + tensor var_697_end_mask_0 = const()[name = tensor("op_697_end_mask_0"), val = tensor([true, true, true])]; + tensor var_697 = slice_by_index(begin = var_697_begin_0, end = var_697_end_0, end_mask = var_697_end_mask_0, x = input_51)[name = tensor("op_697")]; + tensor x_29 = elu(alpha = var_507, x = input_53)[name = tensor("x_29")]; + tensor v_pad_type_0 = const()[name = tensor("v_pad_type_0"), val = tensor("valid")]; + tensor v_strides_0 = const()[name = tensor("v_strides_0"), val = tensor([1])]; + tensor v_pad_0 = const()[name = tensor("v_pad_0"), val = tensor([0, 0])]; + tensor v_dilations_0 = const()[name = tensor("v_dilations_0"), val = tensor([1])]; + tensor v_groups_0 = const()[name = tensor("v_groups_0"), val = tensor(1)]; + tensor v = conv(bias = mimi_decoder_model_9_block_3_conv_bias, dilations = v_dilations_0, groups = v_groups_0, pad = v_pad_0, pad_type = v_pad_type_0, strides = v_strides_0, weight = mimi_decoder_model_9_block_3_conv_weight, x = x_29)[name = tensor("v")]; + tensor input_55 = add(x = input_49, y = v)[name = tensor("input_55")]; + tensor x = elu(alpha = var_507, x = input_55)[name = tensor("x")]; + tensor input_interleave_0 = const()[name = tensor("input_interleave_0"), val = tensor(false)]; + tensor input = concat(axis = var_508, interleave = input_interleave_0, values = (conv_final_prev, x))[name = tensor("input")]; + tensor var_724_pad_type_0 = const()[name = tensor("op_724_pad_type_0"), val = tensor("valid")]; + tensor var_724_strides_0 = const()[name = tensor("op_724_strides_0"), val = tensor([1])]; + tensor var_724_pad_0 = const()[name = tensor("op_724_pad_0"), val = tensor([0, 0])]; + tensor var_724_dilations_0 = const()[name = tensor("op_724_dilations_0"), val = tensor([1])]; + tensor var_724_groups_0 = const()[name = tensor("op_724_groups_0"), val = tensor(1)]; + tensor var_724 = conv(bias = mimi_decoder_model_11_conv_bias, dilations = var_724_dilations_0, groups = var_724_groups_0, pad = var_724_pad_0, pad_type = var_724_pad_type_0, strides = var_724_strides_0, weight = mimi_decoder_model_11_conv_weight, x = input)[name = tensor("op_724")]; + tensor var_725_begin_0 = const()[name = tensor("op_725_begin_0"), val = tensor([0, 0, 1920])]; + tensor var_725_end_0 = const()[name = tensor("op_725_end_0"), val = tensor([1, 64, 1922])]; + tensor var_725_end_mask_0 = const()[name = tensor("op_725_end_mask_0"), val = tensor([true, true, true])]; + tensor var_725 = slice_by_index(begin = var_725_begin_0, end = var_725_end_0, end_mask = var_725_end_mask_0, x = input)[name = tensor("op_725")]; + tensor var_740_promoted = const()[name = tensor("op_740_promoted"), val = tensor(0x1p+4)]; + tensor var_741 = add(x = attn0_offset, y = var_740_promoted)[name = tensor("op_741")]; + tensor var_743_promoted = const()[name = tensor("op_743_promoted"), val = tensor(0x1p+4)]; + tensor var_744 = add(x = attn1_offset, y = var_743_promoted)[name = tensor("op_744")]; + tensor conv0_first_tmp = identity(x = conv0_first)[name = tensor("conv0_first_tmp")]; + tensor res0_conv0_first_tmp = identity(x = res0_conv0_first)[name = tensor("res0_conv0_first_tmp")]; + tensor res0_conv1_prev_tmp = identity(x = res0_conv1_prev)[name = tensor("res0_conv1_prev_tmp")]; + tensor res0_conv1_first_tmp = identity(x = res0_conv1_first)[name = tensor("res0_conv1_first_tmp")]; + tensor res1_conv0_first_tmp = identity(x = res1_conv0_first)[name = tensor("res1_conv0_first_tmp")]; + tensor res1_conv1_prev_tmp = identity(x = res1_conv1_prev)[name = tensor("res1_conv1_prev_tmp")]; + tensor res1_conv1_first_tmp = identity(x = res1_conv1_first)[name = tensor("res1_conv1_first_tmp")]; + tensor res2_conv0_first_tmp = identity(x = res2_conv0_first)[name = tensor("res2_conv0_first_tmp")]; + tensor res2_conv1_prev_tmp = identity(x = res2_conv1_prev)[name = tensor("res2_conv1_prev_tmp")]; + tensor res2_conv1_first_tmp = identity(x = res2_conv1_first)[name = tensor("res2_conv1_first_tmp")]; + tensor conv_final_first_tmp = identity(x = conv_final_first)[name = tensor("conv_final_first_tmp")]; + } -> (var_724, var_77, var_210, var_741, var_400, var_744, var_542, conv0_first, var_565, var_585, res0_conv0_first, res0_conv1_prev, res0_conv1_first, var_621, var_641, res1_conv0_first, res1_conv1_prev, res1_conv1_first, var_677, var_697, res2_conv0_first, res2_conv1_prev, res2_conv1_first, var_725, conv_final_first); +} \ No newline at end of file diff --git a/v2.1/italian/mimi_decoder.mlmodelc/weights/weight.bin b/v2.1/italian/mimi_decoder.mlmodelc/weights/weight.bin new file mode 100644 index 0000000000000000000000000000000000000000..c2896951f3ea7956cb354933a2e293a51ccf6a93 --- /dev/null +++ b/v2.1/italian/mimi_decoder.mlmodelc/weights/weight.bin @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:42b8f0af28e5488e1de239479690b51b4a2411501ef1a1de38b9244317bb0cc2 +size 41768256 diff --git a/v2.1/italian/mimi_decoder.mlpackage/Data/com.apple.CoreML/model.mlmodel b/v2.1/italian/mimi_decoder.mlpackage/Data/com.apple.CoreML/model.mlmodel new file mode 100644 index 0000000000000000000000000000000000000000..2d99fb17b7cc55d6fcb9ed35e807c55439d82ae3 --- /dev/null +++ b/v2.1/italian/mimi_decoder.mlpackage/Data/com.apple.CoreML/model.mlmodel @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:27e84301b7559dd76ddcd76960133073793bdfbd9fe4ed279c902f71edf14f86 +size 93270 diff --git a/v2.1/italian/mimi_decoder.mlpackage/Data/com.apple.CoreML/weights/weight.bin b/v2.1/italian/mimi_decoder.mlpackage/Data/com.apple.CoreML/weights/weight.bin new file mode 100644 index 0000000000000000000000000000000000000000..c2896951f3ea7956cb354933a2e293a51ccf6a93 --- /dev/null +++ b/v2.1/italian/mimi_decoder.mlpackage/Data/com.apple.CoreML/weights/weight.bin @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:42b8f0af28e5488e1de239479690b51b4a2411501ef1a1de38b9244317bb0cc2 +size 41768256 diff --git a/v2.1/italian/mimi_decoder.mlpackage/Manifest.json b/v2.1/italian/mimi_decoder.mlpackage/Manifest.json new file mode 100644 index 0000000000000000000000000000000000000000..7064bf0cf385b419d03700342577350d4d35cc2d --- /dev/null +++ b/v2.1/italian/mimi_decoder.mlpackage/Manifest.json @@ -0,0 +1,18 @@ +{ + "fileFormatVersion": "1.0.0", + "itemInfoEntries": { + "34608B57-5762-4FAB-A1ED-EAFB8CC22FDD": { + "author": "com.apple.CoreML", + "description": "CoreML Model Specification", + "name": "model.mlmodel", + "path": "com.apple.CoreML/model.mlmodel" + }, + "CD69AEEA-77F8-4B86-81C3-6249634594F3": { + "author": "com.apple.CoreML", + "description": "CoreML Model Weights", + "name": "weights", + "path": "com.apple.CoreML/weights" + } + }, + "rootModelIdentifier": "34608B57-5762-4FAB-A1ED-EAFB8CC22FDD" +} diff --git a/v2.1/italian_24l/cond_prefill.mlmodelc/analytics/coremldata.bin b/v2.1/italian_24l/cond_prefill.mlmodelc/analytics/coremldata.bin new file mode 100644 index 0000000000000000000000000000000000000000..52ae97057885d276abc5ac36696d7bfab801a810 --- /dev/null +++ b/v2.1/italian_24l/cond_prefill.mlmodelc/analytics/coremldata.bin @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:65afe0db6cfeaa56afbb899ca97e437632696ec83f68030e4d86d01cea43b4aa +size 243 diff --git a/v2.1/italian_24l/cond_prefill.mlmodelc/coremldata.bin b/v2.1/italian_24l/cond_prefill.mlmodelc/coremldata.bin new file mode 100644 index 0000000000000000000000000000000000000000..6dd5e23ca4c6ce053bcc3c26673db89b9b7aa33d --- /dev/null +++ b/v2.1/italian_24l/cond_prefill.mlmodelc/coremldata.bin @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:d000085ecd1e08db6a0d668682c868054dcfcfb29cd199d92da26a8162121c5f +size 3519 diff --git a/v2.1/italian_24l/cond_prefill.mlmodelc/model.mil b/v2.1/italian_24l/cond_prefill.mlmodelc/model.mil new file mode 100644 index 0000000000000000000000000000000000000000..7028cd94630c33bb5143ac20b6b86ca016eb1d25 --- /dev/null +++ b/v2.1/italian_24l/cond_prefill.mlmodelc/model.mil @@ -0,0 +1,5431 @@ +program(1.0) +[buildInfo = dict, tensor>({{"coremlc-component-MIL", "3520.4.1"}, {"coremlc-version", "3520.5.1"}, {"coremltools-component-torch", "2.12.0"}, {"coremltools-source-dialect", "TorchScript"}, {"coremltools-version", "9.0"}})] +{ + func main(tensor cache0, tensor cache1, tensor cache10, tensor cache11, tensor cache12, tensor cache13, tensor cache14, tensor cache15, tensor cache16, tensor cache17, tensor cache18, tensor cache19, tensor cache2, tensor cache20, tensor cache21, tensor cache22, tensor cache23, tensor cache3, tensor cache4, tensor cache5, tensor cache6, tensor cache7, tensor cache8, tensor cache9, tensor conditioning, tensor position0, tensor position1, tensor position10, tensor position11, tensor position12, tensor position13, tensor position14, tensor position15, tensor position16, tensor position17, tensor position18, tensor position19, tensor position2, tensor position20, tensor position21, tensor position22, tensor position23, tensor position3, tensor position4, tensor position5, tensor position6, tensor position7, tensor position8, tensor position9, tensor valid_len) { + tensor input_3_axes_0 = const()[name = tensor("input_3_axes_0"), val = tensor([-1])]; + tensor conditioning_to_fp16_dtype_0 = const()[name = tensor("conditioning_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor norm0_1_weight_to_fp16 = const()[name = tensor("norm0_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(64)))]; + tensor norm0_1_bias_to_fp16 = const()[name = tensor("norm0_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(2176)))]; + tensor var_192_to_fp16 = const()[name = tensor("op_192_to_fp16"), val = tensor(0x1.5p-17)]; + tensor conditioning_to_fp16 = cast(dtype = conditioning_to_fp16_dtype_0, x = conditioning)[name = tensor("cast_663")]; + tensor input_3_cast_fp16 = layer_norm(axes = input_3_axes_0, beta = norm0_1_bias_to_fp16, epsilon = var_192_to_fp16, gamma = norm0_1_weight_to_fp16, x = conditioning_to_fp16)[name = tensor("input_3_cast_fp16")]; + tensor attn0_in_proj_weight_to_fp16 = const()[name = tensor("attn0_in_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(4288)))]; + tensor linear_0_bias_0_to_fp16 = const()[name = tensor("linear_0_bias_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(6295808)))]; + tensor linear_0_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = attn0_in_proj_weight_to_fp16, x = input_3_cast_fp16)[name = tensor("linear_0_cast_fp16")]; + tensor var_206 = const()[name = tensor("op_206"), val = tensor([1, 256, 3, 16, 64])]; + tensor qkv_1_cast_fp16 = reshape(shape = var_206, x = linear_0_cast_fp16)[name = tensor("qkv_1_cast_fp16")]; + tensor q_1_begin_0 = const()[name = tensor("q_1_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor q_1_end_0 = const()[name = tensor("q_1_end_0"), val = tensor([1, 256, 1, 16, 64])]; + tensor q_1_end_mask_0 = const()[name = tensor("q_1_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor q_1_squeeze_mask_0 = const()[name = tensor("q_1_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor q_1_cast_fp16 = slice_by_index(begin = q_1_begin_0, end = q_1_end_0, end_mask = q_1_end_mask_0, squeeze_mask = q_1_squeeze_mask_0, x = qkv_1_cast_fp16)[name = tensor("q_1_cast_fp16")]; + tensor k_1_begin_0 = const()[name = tensor("k_1_begin_0"), val = tensor([0, 0, 1, 0, 0])]; + tensor k_1_end_0 = const()[name = tensor("k_1_end_0"), val = tensor([1, 256, 2, 16, 64])]; + tensor k_1_end_mask_0 = const()[name = tensor("k_1_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor k_1_squeeze_mask_0 = const()[name = tensor("k_1_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor k_1_cast_fp16 = slice_by_index(begin = k_1_begin_0, end = k_1_end_0, end_mask = k_1_end_mask_0, squeeze_mask = k_1_squeeze_mask_0, x = qkv_1_cast_fp16)[name = tensor("k_1_cast_fp16")]; + tensor v_1_begin_0 = const()[name = tensor("v_1_begin_0"), val = tensor([0, 0, 2, 0, 0])]; + tensor v_1_end_0 = const()[name = tensor("v_1_end_0"), val = tensor([1, 256, 3, 16, 64])]; + tensor v_1_end_mask_0 = const()[name = tensor("v_1_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor v_1_squeeze_mask_0 = const()[name = tensor("v_1_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor v_1_cast_fp16 = slice_by_index(begin = v_1_begin_0, end = v_1_end_0, end_mask = v_1_end_mask_0, squeeze_mask = v_1_squeeze_mask_0, x = qkv_1_cast_fp16)[name = tensor("v_1_cast_fp16")]; + tensor var_264 = const()[name = tensor("op_264"), val = tensor([1, 1])]; + tensor position0_to_fp16_dtype_0 = const()[name = tensor("position0_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor position0_to_fp16 = cast(dtype = position0_to_fp16_dtype_0, x = position0)[name = tensor("cast_662")]; + tensor var_265_cast_fp16 = reshape(shape = var_264, x = position0_to_fp16)[name = tensor("op_265_cast_fp16")]; + tensor ts_1_promoted_to_fp16 = const()[name = tensor("ts_1_promoted_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(6302016)))]; + tensor ts_3_cast_fp16 = add(x = ts_1_promoted_to_fp16, y = var_265_cast_fp16)[name = tensor("ts_3_cast_fp16")]; + tensor var_272 = const()[name = tensor("op_272"), val = tensor([1, 256, 1, 1])]; + tensor ts_5_cast_fp16 = reshape(shape = var_272, x = ts_3_cast_fp16)[name = tensor("ts_5_cast_fp16")]; + tensor var_279 = const()[name = tensor("op_279"), val = tensor([1, 256, 16, 32, 2])]; + tensor q_complex_1_cast_fp16 = reshape(shape = var_279, x = q_1_cast_fp16)[name = tensor("q_complex_1_cast_fp16")]; + tensor var_286 = const()[name = tensor("op_286"), val = tensor([1, 256, 16, 32, 2])]; + tensor k_complex_1_cast_fp16 = reshape(shape = var_286, x = k_1_cast_fp16)[name = tensor("k_complex_1_cast_fp16")]; + tensor var_290_begin_0 = const()[name = tensor("op_290_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_290_end_0 = const()[name = tensor("op_290_end_0"), val = tensor([1, 256, 16, 32, 1])]; + tensor var_290_end_mask_0 = const()[name = tensor("op_290_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_290_squeeze_mask_0 = const()[name = tensor("op_290_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_290_cast_fp16 = slice_by_index(begin = var_290_begin_0, end = var_290_end_0, end_mask = var_290_end_mask_0, squeeze_mask = var_290_squeeze_mask_0, x = q_complex_1_cast_fp16)[name = tensor("op_290_cast_fp16")]; + tensor var_298_begin_0 = const()[name = tensor("op_298_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_298_end_0 = const()[name = tensor("op_298_end_0"), val = tensor([1, 256, 16, 32, 2])]; + tensor var_298_end_mask_0 = const()[name = tensor("op_298_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_298_squeeze_mask_0 = const()[name = tensor("op_298_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_298_cast_fp16 = slice_by_index(begin = var_298_begin_0, end = var_298_end_0, end_mask = var_298_end_mask_0, squeeze_mask = var_298_squeeze_mask_0, x = q_complex_1_cast_fp16)[name = tensor("op_298_cast_fp16")]; + tensor var_306_begin_0 = const()[name = tensor("op_306_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_306_end_0 = const()[name = tensor("op_306_end_0"), val = tensor([1, 256, 16, 32, 1])]; + tensor var_306_end_mask_0 = const()[name = tensor("op_306_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_306_squeeze_mask_0 = const()[name = tensor("op_306_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_306_cast_fp16 = slice_by_index(begin = var_306_begin_0, end = var_306_end_0, end_mask = var_306_end_mask_0, squeeze_mask = var_306_squeeze_mask_0, x = k_complex_1_cast_fp16)[name = tensor("op_306_cast_fp16")]; + tensor var_314_begin_0 = const()[name = tensor("op_314_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_314_end_0 = const()[name = tensor("op_314_end_0"), val = tensor([1, 256, 16, 32, 2])]; + tensor var_314_end_mask_0 = const()[name = tensor("op_314_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_314_squeeze_mask_0 = const()[name = tensor("op_314_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_314_cast_fp16 = slice_by_index(begin = var_314_begin_0, end = var_314_end_0, end_mask = var_314_end_mask_0, squeeze_mask = var_314_squeeze_mask_0, x = k_complex_1_cast_fp16)[name = tensor("op_314_cast_fp16")]; + tensor freqs_1_to_fp16 = const()[name = tensor("freqs_1_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(6302592)))]; + tensor var_320_cast_fp16 = mul(x = freqs_1_to_fp16, y = ts_5_cast_fp16)[name = tensor("op_320_cast_fp16")]; + tensor rotr_1_cast_fp16 = cos(x = var_320_cast_fp16)[name = tensor("rotr_1_cast_fp16")]; + tensor roti_1_cast_fp16 = sin(x = var_320_cast_fp16)[name = tensor("roti_1_cast_fp16")]; + tensor var_324_cast_fp16 = mul(x = var_290_cast_fp16, y = rotr_1_cast_fp16)[name = tensor("op_324_cast_fp16")]; + tensor var_325_cast_fp16 = mul(x = var_298_cast_fp16, y = roti_1_cast_fp16)[name = tensor("op_325_cast_fp16")]; + tensor qor_1_cast_fp16 = sub(x = var_324_cast_fp16, y = var_325_cast_fp16)[name = tensor("qor_1_cast_fp16")]; + tensor var_328_cast_fp16 = mul(x = var_290_cast_fp16, y = roti_1_cast_fp16)[name = tensor("op_328_cast_fp16")]; + tensor var_329_cast_fp16 = mul(x = var_298_cast_fp16, y = rotr_1_cast_fp16)[name = tensor("op_329_cast_fp16")]; + tensor qoi_1_cast_fp16 = add(x = var_328_cast_fp16, y = var_329_cast_fp16)[name = tensor("qoi_1_cast_fp16")]; + tensor var_332_cast_fp16 = mul(x = var_306_cast_fp16, y = rotr_1_cast_fp16)[name = tensor("op_332_cast_fp16")]; + tensor var_333_cast_fp16 = mul(x = var_314_cast_fp16, y = roti_1_cast_fp16)[name = tensor("op_333_cast_fp16")]; + tensor kor_1_cast_fp16 = sub(x = var_332_cast_fp16, y = var_333_cast_fp16)[name = tensor("kor_1_cast_fp16")]; + tensor var_336_cast_fp16 = mul(x = var_306_cast_fp16, y = roti_1_cast_fp16)[name = tensor("op_336_cast_fp16")]; + tensor var_337_cast_fp16 = mul(x = var_314_cast_fp16, y = rotr_1_cast_fp16)[name = tensor("op_337_cast_fp16")]; + tensor koi_1_cast_fp16 = add(x = var_336_cast_fp16, y = var_337_cast_fp16)[name = tensor("koi_1_cast_fp16")]; + tensor qo_1_axis_0 = const()[name = tensor("qo_1_axis_0"), val = tensor(-1)]; + tensor qo_1_cast_fp16 = stack(axis = qo_1_axis_0, values = (qor_1_cast_fp16, qoi_1_cast_fp16))[name = tensor("qo_1_cast_fp16")]; + tensor ko_1_axis_0 = const()[name = tensor("ko_1_axis_0"), val = tensor(-1)]; + tensor ko_1_cast_fp16 = stack(axis = ko_1_axis_0, values = (kor_1_cast_fp16, koi_1_cast_fp16))[name = tensor("ko_1_cast_fp16")]; + tensor var_370 = const()[name = tensor("op_370"), val = tensor([1, 256, 16, 64])]; + tensor q_3_cast_fp16 = reshape(shape = var_370, x = qo_1_cast_fp16)[name = tensor("q_3_cast_fp16")]; + tensor var_376 = const()[name = tensor("op_376"), val = tensor([1, 256, 16, 64])]; + tensor k_3_cast_fp16 = reshape(shape = var_376, x = ko_1_cast_fp16)[name = tensor("k_3_cast_fp16")]; + tensor var_393_promoted_to_fp16 = const()[name = tensor("op_393_promoted_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(6302720)))]; + tensor write_indices_float_1_cast_fp16 = add(x = var_265_cast_fp16, y = var_393_promoted_to_fp16)[name = tensor("write_indices_float_1_cast_fp16")]; + tensor _inversed_397_y_0_to_fp16 = const()[name = tensor("_inversed_397_y_0_to_fp16"), val = tensor(0x1p-9)]; + tensor _inversed_397_cast_fp16 = mul(x = write_indices_float_1_cast_fp16, y = _inversed_397_y_0_to_fp16)[name = tensor("_inversed_397_cast_fp16")]; + tensor var_398_cast_fp16 = floor(x = _inversed_397_cast_fp16)[name = tensor("op_398_cast_fp16")]; + tensor var_399_to_fp16 = const()[name = tensor("op_399_to_fp16"), val = tensor(0x1p+9)]; + tensor var_400_cast_fp16 = mul(x = var_398_cast_fp16, y = var_399_to_fp16)[name = tensor("op_400_cast_fp16")]; + tensor write_indices_float_3_cast_fp16 = sub(x = write_indices_float_1_cast_fp16, y = var_400_cast_fp16)[name = tensor("write_indices_float_3_cast_fp16")]; + tensor var_409 = const()[name = tensor("op_409"), val = tensor([1, 1])]; + tensor valid_len_to_fp16_dtype_0 = const()[name = tensor("valid_len_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor valid_len_to_fp16 = cast(dtype = valid_len_to_fp16_dtype_0, x = valid_len)[name = tensor("cast_661")]; + tensor var_410_cast_fp16 = reshape(shape = var_409, x = valid_len_to_fp16)[name = tensor("op_410_cast_fp16")]; + tensor token_valid_1_cast_fp16 = less(x = var_393_promoted_to_fp16, y = var_410_cast_fp16)[name = tensor("token_valid_1_cast_fp16")]; + tensor var_418_value_0_to_fp16 = const()[name = tensor("op_418_value_0_to_fp16"), val = tensor(0x1.ffp+8)]; + tensor var_418_cast_fp16 = fill_like(ref_tensor = write_indices_float_3_cast_fp16, value = var_418_value_0_to_fp16)[name = tensor("op_418_cast_fp16")]; + tensor write_indices_float_5_cast_fp16 = select(a = write_indices_float_3_cast_fp16, b = var_418_cast_fp16, cond = token_valid_1_cast_fp16)[name = tensor("write_indices_float_5_cast_fp16")]; + tensor var_424_dtype_0 = const()[name = tensor("op_424_dtype_0"), val = tensor("int32")]; + tensor var_429 = const()[name = tensor("op_429"), val = tensor([1, 256, 1, 1])]; + tensor write_indices_float_5_cast_fp16_to_int32 = cast(dtype = var_424_dtype_0, x = write_indices_float_5_cast_fp16)[name = tensor("cast_660")]; + tensor var_430 = reshape(shape = var_429, x = write_indices_float_5_cast_fp16_to_int32)[name = tensor("op_430")]; + tensor write_indices_1_reps_0 = const()[name = tensor("write_indices_1_reps_0"), val = tensor([1, 1, 16, 64])]; + tensor write_indices_1 = tile(reps = write_indices_1_reps_0, x = var_430)[name = tensor("write_indices_1")]; + tensor var_440_begin_0 = const()[name = tensor("op_440_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_440_end_0 = const()[name = tensor("op_440_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor var_440_end_mask_0 = const()[name = tensor("op_440_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_440_squeeze_mask_0 = const()[name = tensor("op_440_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor cache0_to_fp16_dtype_0 = const()[name = tensor("cache0_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor cache0_to_fp16 = cast(dtype = cache0_to_fp16_dtype_0, x = cache0)[name = tensor("cast_659")]; + tensor var_440_cast_fp16 = slice_by_index(begin = var_440_begin_0, end = var_440_end_0, end_mask = var_440_end_mask_0, squeeze_mask = var_440_squeeze_mask_0, x = cache0_to_fp16)[name = tensor("op_440_cast_fp16")]; + tensor var_442_axis_0 = const()[name = tensor("op_442_axis_0"), val = tensor(1)]; + tensor var_442_mode_0 = const()[name = tensor("op_442_mode_0"), val = tensor("update")]; + tensor var_442_validate_indices_0 = const()[name = tensor("op_442_validate_indices_0"), val = tensor(false)]; + tensor var_442_cast_fp16 = scatter_along_axis(axis = var_442_axis_0, data = var_440_cast_fp16, indices = write_indices_1, mode = var_442_mode_0, updates = k_3_cast_fp16, validate_indices = var_442_validate_indices_0)[name = tensor("op_442_cast_fp16")]; + tensor concat_2 = const()[name = tensor("concat_2"), val = tensor([0, 0, 0, 0, 0])]; + tensor concat_3 = const()[name = tensor("concat_3"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_1_internal_tensor_assign_1_stride_0 = const()[name = tensor("new_cache_1_internal_tensor_assign_1_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_1_internal_tensor_assign_1_begin_mask_0 = const()[name = tensor("new_cache_1_internal_tensor_assign_1_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_1_internal_tensor_assign_1_end_mask_0 = const()[name = tensor("new_cache_1_internal_tensor_assign_1_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_1_internal_tensor_assign_1_squeeze_mask_0 = const()[name = tensor("new_cache_1_internal_tensor_assign_1_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_46 = const()[name = tensor("shape_46"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_0 = const()[name = tensor("reduce_prod_0"), val = tensor(1048576)]; + tensor range_1d_0_start_0 = const()[name = tensor("range_1d_0_start_0"), val = tensor(0)]; + tensor range_1d_0_step_0 = const()[name = tensor("range_1d_0_step_0"), val = tensor(1)]; + tensor range_1d_0 = range_1d(end = reduce_prod_0, start = range_1d_0_start_0, step = range_1d_0_step_0)[name = tensor("range_1d_0")]; + tensor reshape_0 = reshape(shape = shape_46, x = range_1d_0)[name = tensor("reshape_0")]; + tensor slice_by_index_0 = slice_by_index(begin = concat_2, begin_mask = new_cache_1_internal_tensor_assign_1_begin_mask_0, end = concat_3, end_mask = new_cache_1_internal_tensor_assign_1_end_mask_0, squeeze_mask = new_cache_1_internal_tensor_assign_1_squeeze_mask_0, stride = new_cache_1_internal_tensor_assign_1_stride_0, x = reshape_0)[name = tensor("slice_by_index_0")]; + tensor reshape_1_shape_0 = const()[name = tensor("reshape_1_shape_0"), val = tensor([-1])]; + tensor reshape_1 = reshape(shape = reshape_1_shape_0, x = slice_by_index_0)[name = tensor("reshape_1")]; + tensor reshape_2_shape_0 = const()[name = tensor("reshape_2_shape_0"), val = tensor([-1])]; + tensor reshape_2_cast_fp16 = reshape(shape = reshape_2_shape_0, x = var_442_cast_fp16)[name = tensor("reshape_2_cast_fp16")]; + tensor reshape_3_shape_0 = const()[name = tensor("reshape_3_shape_0"), val = tensor([-1])]; + tensor reshape_3_cast_fp16 = reshape(shape = reshape_3_shape_0, x = cache0_to_fp16)[name = tensor("reshape_3_cast_fp16")]; + tensor scatter_0_mode_0 = const()[name = tensor("scatter_0_mode_0"), val = tensor("update")]; + tensor scatter_0_axis_0 = const()[name = tensor("scatter_0_axis_0"), val = tensor(0)]; + tensor scatter_0_validate_indices_0 = const()[name = tensor("scatter_0_validate_indices_0"), val = tensor(false)]; + tensor scatter_0_cast_fp16 = scatter(axis = scatter_0_axis_0, data = reshape_3_cast_fp16, indices = reshape_1, mode = scatter_0_mode_0, updates = reshape_2_cast_fp16, validate_indices = scatter_0_validate_indices_0)[name = tensor("scatter_0_cast_fp16")]; + tensor reshape_4_cast_fp16 = reshape(shape = shape_46, x = scatter_0_cast_fp16)[name = tensor("reshape_4_cast_fp16")]; + tensor var_450_begin_0 = const()[name = tensor("op_450_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor var_450_end_0 = const()[name = tensor("op_450_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor var_450_end_mask_0 = const()[name = tensor("op_450_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_450_squeeze_mask_0 = const()[name = tensor("op_450_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor var_450_cast_fp16 = slice_by_index(begin = var_450_begin_0, end = var_450_end_0, end_mask = var_450_end_mask_0, squeeze_mask = var_450_squeeze_mask_0, x = reshape_4_cast_fp16)[name = tensor("op_450_cast_fp16")]; + tensor var_452_axis_0 = const()[name = tensor("op_452_axis_0"), val = tensor(1)]; + tensor var_452_mode_0 = const()[name = tensor("op_452_mode_0"), val = tensor("update")]; + tensor var_452_validate_indices_0 = const()[name = tensor("op_452_validate_indices_0"), val = tensor(false)]; + tensor var_452_cast_fp16 = scatter_along_axis(axis = var_452_axis_0, data = var_450_cast_fp16, indices = write_indices_1, mode = var_452_mode_0, updates = v_1_cast_fp16, validate_indices = var_452_validate_indices_0)[name = tensor("op_452_cast_fp16")]; + tensor concat_4 = const()[name = tensor("concat_4"), val = tensor([1, 0, 0, 0, 0])]; + tensor concat_5 = const()[name = tensor("concat_5"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_1_internal_tensor_assign_2_stride_0 = const()[name = tensor("new_cache_1_internal_tensor_assign_2_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_1_internal_tensor_assign_2_begin_mask_0 = const()[name = tensor("new_cache_1_internal_tensor_assign_2_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_1_internal_tensor_assign_2_end_mask_0 = const()[name = tensor("new_cache_1_internal_tensor_assign_2_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_1_internal_tensor_assign_2_squeeze_mask_0 = const()[name = tensor("new_cache_1_internal_tensor_assign_2_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_47 = const()[name = tensor("shape_47"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_1 = const()[name = tensor("reduce_prod_1"), val = tensor(1048576)]; + tensor range_1d_1_start_0 = const()[name = tensor("range_1d_1_start_0"), val = tensor(0)]; + tensor range_1d_1_step_0 = const()[name = tensor("range_1d_1_step_0"), val = tensor(1)]; + tensor range_1d_1 = range_1d(end = reduce_prod_1, start = range_1d_1_start_0, step = range_1d_1_step_0)[name = tensor("range_1d_1")]; + tensor reshape_5 = reshape(shape = shape_47, x = range_1d_1)[name = tensor("reshape_5")]; + tensor slice_by_index_1 = slice_by_index(begin = concat_4, begin_mask = new_cache_1_internal_tensor_assign_2_begin_mask_0, end = concat_5, end_mask = new_cache_1_internal_tensor_assign_2_end_mask_0, squeeze_mask = new_cache_1_internal_tensor_assign_2_squeeze_mask_0, stride = new_cache_1_internal_tensor_assign_2_stride_0, x = reshape_5)[name = tensor("slice_by_index_1")]; + tensor reshape_6_shape_0 = const()[name = tensor("reshape_6_shape_0"), val = tensor([-1])]; + tensor reshape_6 = reshape(shape = reshape_6_shape_0, x = slice_by_index_1)[name = tensor("reshape_6")]; + tensor reshape_7_shape_0 = const()[name = tensor("reshape_7_shape_0"), val = tensor([-1])]; + tensor reshape_7_cast_fp16 = reshape(shape = reshape_7_shape_0, x = var_452_cast_fp16)[name = tensor("reshape_7_cast_fp16")]; + tensor reshape_8_shape_0 = const()[name = tensor("reshape_8_shape_0"), val = tensor([-1])]; + tensor reshape_8_cast_fp16 = reshape(shape = reshape_8_shape_0, x = reshape_4_cast_fp16)[name = tensor("reshape_8_cast_fp16")]; + tensor scatter_1_mode_0 = const()[name = tensor("scatter_1_mode_0"), val = tensor("update")]; + tensor scatter_1_axis_0 = const()[name = tensor("scatter_1_axis_0"), val = tensor(0)]; + tensor scatter_1_validate_indices_0 = const()[name = tensor("scatter_1_validate_indices_0"), val = tensor(false)]; + tensor scatter_1_cast_fp16 = scatter(axis = scatter_1_axis_0, data = reshape_8_cast_fp16, indices = reshape_6, mode = scatter_1_mode_0, updates = reshape_7_cast_fp16, validate_indices = scatter_1_validate_indices_0)[name = tensor("scatter_1_cast_fp16")]; + tensor reshape_9_cast_fp16 = reshape(shape = shape_47, x = scatter_1_cast_fp16)[name = tensor("reshape_9_cast_fp16")]; + tensor reshape_9_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("reshape_9_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor keys_1_begin_0 = const()[name = tensor("keys_1_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor keys_1_end_0 = const()[name = tensor("keys_1_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor keys_1_end_mask_0 = const()[name = tensor("keys_1_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor keys_1_squeeze_mask_0 = const()[name = tensor("keys_1_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor keys_1_cast_fp16 = slice_by_index(begin = keys_1_begin_0, end = keys_1_end_0, end_mask = keys_1_end_mask_0, squeeze_mask = keys_1_squeeze_mask_0, x = reshape_9_cast_fp16)[name = tensor("keys_1_cast_fp16")]; + tensor values_1_begin_0 = const()[name = tensor("values_1_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor values_1_end_0 = const()[name = tensor("values_1_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor values_1_end_mask_0 = const()[name = tensor("values_1_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor values_1_squeeze_mask_0 = const()[name = tensor("values_1_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor values_1_cast_fp16 = slice_by_index(begin = values_1_begin_0, end = values_1_end_0, end_mask = values_1_end_mask_0, squeeze_mask = values_1_squeeze_mask_0, x = reshape_9_cast_fp16)[name = tensor("values_1_cast_fp16")]; + tensor var_464_cast_fp16 = not_equal(x = keys_1_cast_fp16, y = keys_1_cast_fp16)[name = tensor("op_464_cast_fp16")]; + tensor var_470_to_fp16 = const()[name = tensor("op_470_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(6303296)))]; + tensor keys_3_cast_fp16 = select(a = var_470_to_fp16, b = keys_1_cast_fp16, cond = var_464_cast_fp16)[name = tensor("keys_3_cast_fp16")]; + tensor var_472_cast_fp16 = not_equal(x = values_1_cast_fp16, y = values_1_cast_fp16)[name = tensor("op_472_cast_fp16")]; + tensor values_3_cast_fp16 = select(a = var_470_to_fp16, b = values_1_cast_fp16, cond = var_472_cast_fp16)[name = tensor("values_3_cast_fp16")]; + tensor var_496 = const()[name = tensor("op_496"), val = tensor([0, 2, 1, 3])]; + tensor var_512 = const()[name = tensor("op_512"), val = tensor([1, 1, 1])]; + tensor var_513_cast_fp16 = reshape(shape = var_512, x = position0_to_fp16)[name = tensor("op_513_cast_fp16")]; + tensor q_offsets_1_promoted_to_fp16 = const()[name = tensor("q_offsets_1_promoted_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(7351936)))]; + tensor q_positions_1_cast_fp16 = add(x = var_513_cast_fp16, y = q_offsets_1_promoted_to_fp16)[name = tensor("q_positions_1_cast_fp16")]; + tensor var_535 = const()[name = tensor("op_535"), val = tensor([1, 1, 1])]; + tensor var_536_cast_fp16 = reshape(shape = var_535, x = valid_len_to_fp16)[name = tensor("op_536_cast_fp16")]; + tensor valid_end_1_cast_fp16 = add(x = var_513_cast_fp16, y = var_536_cast_fp16)[name = tensor("valid_end_1_cast_fp16")]; + tensor k_positions_1_promoted_to_fp16 = const()[name = tensor("k_positions_1_promoted_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(7352512)))]; + tensor valid_mask_1_cast_fp16 = less(x = k_positions_1_promoted_to_fp16, y = valid_end_1_cast_fp16)[name = tensor("valid_mask_1_cast_fp16")]; + tensor causal_mask_1_cast_fp16 = less_equal(x = k_positions_1_promoted_to_fp16, y = q_positions_1_cast_fp16)[name = tensor("causal_mask_1_cast_fp16")]; + tensor var_541 = logical_and(x = valid_mask_1_cast_fp16, y = causal_mask_1_cast_fp16)[name = tensor("op_541")]; + tensor attn_mask_1_axes_0 = const()[name = tensor("attn_mask_1_axes_0"), val = tensor([1])]; + tensor attn_mask_1 = expand_dims(axes = attn_mask_1_axes_0, x = var_541)[name = tensor("attn_mask_1")]; + tensor var_555_transpose_x_0 = const()[name = tensor("op_555_transpose_x_0"), val = tensor(false)]; + tensor var_555_transpose_y_0 = const()[name = tensor("op_555_transpose_y_0"), val = tensor(false)]; + tensor transpose_69_perm_0 = const()[name = tensor("transpose_69_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_70_perm_0 = const()[name = tensor("transpose_70_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_70 = transpose(perm = transpose_70_perm_0, x = keys_3_cast_fp16)[name = tensor("transpose_204")]; + tensor transpose_69 = transpose(perm = transpose_69_perm_0, x = q_3_cast_fp16)[name = tensor("transpose_205")]; + tensor var_555_cast_fp16 = matmul(transpose_x = var_555_transpose_x_0, transpose_y = var_555_transpose_y_0, x = transpose_69, y = transpose_70)[name = tensor("op_555_cast_fp16")]; + tensor var_549_to_fp16 = const()[name = tensor("op_549_to_fp16"), val = tensor([0x1p-3])]; + tensor attn_weights_1_cast_fp16 = mul(x = var_555_cast_fp16, y = var_549_to_fp16)[name = tensor("attn_weights_1_cast_fp16")]; + tensor var_557 = logical_not(x = attn_mask_1)[name = tensor("op_557")]; + tensor var_558_to_fp16 = const()[name = tensor("op_558_to_fp16"), val = tensor(-inf)]; + tensor attn_weights_3_cast_fp16 = select(a = var_558_to_fp16, b = attn_weights_1_cast_fp16, cond = var_557)[name = tensor("attn_weights_3_cast_fp16")]; + tensor var_560 = const()[name = tensor("op_560"), val = tensor(-1)]; + tensor attn_weights_5_cast_fp16 = softmax(axis = var_560, x = attn_weights_3_cast_fp16)[name = tensor("attn_weights_5_cast_fp16")]; + tensor attn_output_1_transpose_x_0 = const()[name = tensor("attn_output_1_transpose_x_0"), val = tensor(false)]; + tensor attn_output_1_transpose_y_0 = const()[name = tensor("attn_output_1_transpose_y_0"), val = tensor(false)]; + tensor values_5_cast_fp16 = transpose(perm = var_496, x = values_3_cast_fp16)[name = tensor("transpose_206")]; + tensor attn_output_1_cast_fp16 = matmul(transpose_x = attn_output_1_transpose_x_0, transpose_y = attn_output_1_transpose_y_0, x = attn_weights_5_cast_fp16, y = values_5_cast_fp16)[name = tensor("attn_output_1_cast_fp16")]; + tensor var_568 = const()[name = tensor("op_568"), val = tensor([0, 2, 1, 3])]; + tensor var_573 = const()[name = tensor("op_573"), val = tensor([1, 256, 1024])]; + tensor var_569_cast_fp16 = transpose(perm = var_568, x = attn_output_1_cast_fp16)[name = tensor("transpose_203")]; + tensor input_5_cast_fp16 = reshape(shape = var_573, x = var_569_cast_fp16)[name = tensor("input_5_cast_fp16")]; + tensor attn0_out_proj_weight_to_fp16 = const()[name = tensor("attn0_out_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(7353600)))]; + tensor linear_1_bias_0_to_fp16 = const()[name = tensor("linear_1_bias_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9450816)))]; + tensor linear_1_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = attn0_out_proj_weight_to_fp16, x = input_5_cast_fp16)[name = tensor("linear_1_cast_fp16")]; + tensor var_582_cast_fp16 = add(x = position0_to_fp16, y = valid_len_to_fp16)[name = tensor("op_582_cast_fp16")]; + tensor var_582_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_582_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor input_7_cast_fp16 = add(x = conditioning_to_fp16, y = linear_1_cast_fp16)[name = tensor("input_7_cast_fp16")]; + tensor input_9_axes_0 = const()[name = tensor("input_9_axes_0"), val = tensor([-1])]; + tensor norm0_2_weight_to_fp16 = const()[name = tensor("norm0_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9452928)))]; + tensor norm0_2_bias_to_fp16 = const()[name = tensor("norm0_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9455040)))]; + tensor var_586_to_fp16 = const()[name = tensor("op_586_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_9_cast_fp16 = layer_norm(axes = input_9_axes_0, beta = norm0_2_bias_to_fp16, epsilon = var_586_to_fp16, gamma = norm0_2_weight_to_fp16, x = input_7_cast_fp16)[name = tensor("input_9_cast_fp16")]; + tensor linear0_1_weight_to_fp16 = const()[name = tensor("linear0_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9457152)))]; + tensor linear_2_bias_0_to_fp16 = const()[name = tensor("linear_2_bias_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(17845824)))]; + tensor linear_2_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = linear0_1_weight_to_fp16, x = input_9_cast_fp16)[name = tensor("linear_2_cast_fp16")]; + tensor input_11_mode_0 = const()[name = tensor("input_11_mode_0"), val = tensor("EXACT")]; + tensor input_11_cast_fp16 = gelu(mode = input_11_mode_0, x = linear_2_cast_fp16)[name = tensor("input_11_cast_fp16")]; + tensor linear0_2_weight_to_fp16 = const()[name = tensor("linear0_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(17854080)))]; + tensor linear_3_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = linear0_2_weight_to_fp16, x = input_11_cast_fp16)[name = tensor("linear_3_cast_fp16")]; + tensor input_13_cast_fp16 = add(x = input_7_cast_fp16, y = linear_3_cast_fp16)[name = tensor("input_13_cast_fp16")]; + tensor input_15_axes_0 = const()[name = tensor("input_15_axes_0"), val = tensor([-1])]; + tensor norm1_1_weight_to_fp16 = const()[name = tensor("norm1_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(26242752)))]; + tensor norm1_1_bias_to_fp16 = const()[name = tensor("norm1_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(26244864)))]; + tensor var_603_to_fp16 = const()[name = tensor("op_603_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_15_cast_fp16 = layer_norm(axes = input_15_axes_0, beta = norm1_1_bias_to_fp16, epsilon = var_603_to_fp16, gamma = norm1_1_weight_to_fp16, x = input_13_cast_fp16)[name = tensor("input_15_cast_fp16")]; + tensor attn1_in_proj_weight_to_fp16 = const()[name = tensor("attn1_in_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(26246976)))]; + tensor linear_4_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = attn1_in_proj_weight_to_fp16, x = input_15_cast_fp16)[name = tensor("linear_4_cast_fp16")]; + tensor var_617 = const()[name = tensor("op_617"), val = tensor([1, 256, 3, 16, 64])]; + tensor qkv_3_cast_fp16 = reshape(shape = var_617, x = linear_4_cast_fp16)[name = tensor("qkv_3_cast_fp16")]; + tensor q_7_begin_0 = const()[name = tensor("q_7_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor q_7_end_0 = const()[name = tensor("q_7_end_0"), val = tensor([1, 256, 1, 16, 64])]; + tensor q_7_end_mask_0 = const()[name = tensor("q_7_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor q_7_squeeze_mask_0 = const()[name = tensor("q_7_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor q_7_cast_fp16 = slice_by_index(begin = q_7_begin_0, end = q_7_end_0, end_mask = q_7_end_mask_0, squeeze_mask = q_7_squeeze_mask_0, x = qkv_3_cast_fp16)[name = tensor("q_7_cast_fp16")]; + tensor k_5_begin_0 = const()[name = tensor("k_5_begin_0"), val = tensor([0, 0, 1, 0, 0])]; + tensor k_5_end_0 = const()[name = tensor("k_5_end_0"), val = tensor([1, 256, 2, 16, 64])]; + tensor k_5_end_mask_0 = const()[name = tensor("k_5_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor k_5_squeeze_mask_0 = const()[name = tensor("k_5_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor k_5_cast_fp16 = slice_by_index(begin = k_5_begin_0, end = k_5_end_0, end_mask = k_5_end_mask_0, squeeze_mask = k_5_squeeze_mask_0, x = qkv_3_cast_fp16)[name = tensor("k_5_cast_fp16")]; + tensor v_3_begin_0 = const()[name = tensor("v_3_begin_0"), val = tensor([0, 0, 2, 0, 0])]; + tensor v_3_end_0 = const()[name = tensor("v_3_end_0"), val = tensor([1, 256, 3, 16, 64])]; + tensor v_3_end_mask_0 = const()[name = tensor("v_3_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor v_3_squeeze_mask_0 = const()[name = tensor("v_3_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor v_3_cast_fp16 = slice_by_index(begin = v_3_begin_0, end = v_3_end_0, end_mask = v_3_end_mask_0, squeeze_mask = v_3_squeeze_mask_0, x = qkv_3_cast_fp16)[name = tensor("v_3_cast_fp16")]; + tensor var_675 = const()[name = tensor("op_675"), val = tensor([1, 1])]; + tensor position1_to_fp16_dtype_0 = const()[name = tensor("position1_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor position1_to_fp16 = cast(dtype = position1_to_fp16_dtype_0, x = position1)[name = tensor("cast_656")]; + tensor var_676_cast_fp16 = reshape(shape = var_675, x = position1_to_fp16)[name = tensor("op_676_cast_fp16")]; + tensor ts_9_cast_fp16 = add(x = ts_1_promoted_to_fp16, y = var_676_cast_fp16)[name = tensor("ts_9_cast_fp16")]; + tensor var_683 = const()[name = tensor("op_683"), val = tensor([1, 256, 1, 1])]; + tensor ts_11_cast_fp16 = reshape(shape = var_683, x = ts_9_cast_fp16)[name = tensor("ts_11_cast_fp16")]; + tensor var_690 = const()[name = tensor("op_690"), val = tensor([1, 256, 16, 32, 2])]; + tensor q_complex_3_cast_fp16 = reshape(shape = var_690, x = q_7_cast_fp16)[name = tensor("q_complex_3_cast_fp16")]; + tensor var_697 = const()[name = tensor("op_697"), val = tensor([1, 256, 16, 32, 2])]; + tensor k_complex_3_cast_fp16 = reshape(shape = var_697, x = k_5_cast_fp16)[name = tensor("k_complex_3_cast_fp16")]; + tensor var_701_begin_0 = const()[name = tensor("op_701_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_701_end_0 = const()[name = tensor("op_701_end_0"), val = tensor([1, 256, 16, 32, 1])]; + tensor var_701_end_mask_0 = const()[name = tensor("op_701_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_701_squeeze_mask_0 = const()[name = tensor("op_701_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_701_cast_fp16 = slice_by_index(begin = var_701_begin_0, end = var_701_end_0, end_mask = var_701_end_mask_0, squeeze_mask = var_701_squeeze_mask_0, x = q_complex_3_cast_fp16)[name = tensor("op_701_cast_fp16")]; + tensor var_709_begin_0 = const()[name = tensor("op_709_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_709_end_0 = const()[name = tensor("op_709_end_0"), val = tensor([1, 256, 16, 32, 2])]; + tensor var_709_end_mask_0 = const()[name = tensor("op_709_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_709_squeeze_mask_0 = const()[name = tensor("op_709_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_709_cast_fp16 = slice_by_index(begin = var_709_begin_0, end = var_709_end_0, end_mask = var_709_end_mask_0, squeeze_mask = var_709_squeeze_mask_0, x = q_complex_3_cast_fp16)[name = tensor("op_709_cast_fp16")]; + tensor var_717_begin_0 = const()[name = tensor("op_717_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_717_end_0 = const()[name = tensor("op_717_end_0"), val = tensor([1, 256, 16, 32, 1])]; + tensor var_717_end_mask_0 = const()[name = tensor("op_717_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_717_squeeze_mask_0 = const()[name = tensor("op_717_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_717_cast_fp16 = slice_by_index(begin = var_717_begin_0, end = var_717_end_0, end_mask = var_717_end_mask_0, squeeze_mask = var_717_squeeze_mask_0, x = k_complex_3_cast_fp16)[name = tensor("op_717_cast_fp16")]; + tensor var_725_begin_0 = const()[name = tensor("op_725_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_725_end_0 = const()[name = tensor("op_725_end_0"), val = tensor([1, 256, 16, 32, 2])]; + tensor var_725_end_mask_0 = const()[name = tensor("op_725_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_725_squeeze_mask_0 = const()[name = tensor("op_725_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_725_cast_fp16 = slice_by_index(begin = var_725_begin_0, end = var_725_end_0, end_mask = var_725_end_mask_0, squeeze_mask = var_725_squeeze_mask_0, x = k_complex_3_cast_fp16)[name = tensor("op_725_cast_fp16")]; + tensor freqs_3_to_fp16 = const()[name = tensor("freqs_3_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(32538496)))]; + tensor var_731_cast_fp16 = mul(x = freqs_3_to_fp16, y = ts_11_cast_fp16)[name = tensor("op_731_cast_fp16")]; + tensor rotr_3_cast_fp16 = cos(x = var_731_cast_fp16)[name = tensor("rotr_3_cast_fp16")]; + tensor roti_3_cast_fp16 = sin(x = var_731_cast_fp16)[name = tensor("roti_3_cast_fp16")]; + tensor var_735_cast_fp16 = mul(x = var_701_cast_fp16, y = rotr_3_cast_fp16)[name = tensor("op_735_cast_fp16")]; + tensor var_736_cast_fp16 = mul(x = var_709_cast_fp16, y = roti_3_cast_fp16)[name = tensor("op_736_cast_fp16")]; + tensor qor_5_cast_fp16 = sub(x = var_735_cast_fp16, y = var_736_cast_fp16)[name = tensor("qor_5_cast_fp16")]; + tensor var_739_cast_fp16 = mul(x = var_701_cast_fp16, y = roti_3_cast_fp16)[name = tensor("op_739_cast_fp16")]; + tensor var_740_cast_fp16 = mul(x = var_709_cast_fp16, y = rotr_3_cast_fp16)[name = tensor("op_740_cast_fp16")]; + tensor qoi_5_cast_fp16 = add(x = var_739_cast_fp16, y = var_740_cast_fp16)[name = tensor("qoi_5_cast_fp16")]; + tensor var_743_cast_fp16 = mul(x = var_717_cast_fp16, y = rotr_3_cast_fp16)[name = tensor("op_743_cast_fp16")]; + tensor var_744_cast_fp16 = mul(x = var_725_cast_fp16, y = roti_3_cast_fp16)[name = tensor("op_744_cast_fp16")]; + tensor kor_5_cast_fp16 = sub(x = var_743_cast_fp16, y = var_744_cast_fp16)[name = tensor("kor_5_cast_fp16")]; + tensor var_747_cast_fp16 = mul(x = var_717_cast_fp16, y = roti_3_cast_fp16)[name = tensor("op_747_cast_fp16")]; + tensor var_748_cast_fp16 = mul(x = var_725_cast_fp16, y = rotr_3_cast_fp16)[name = tensor("op_748_cast_fp16")]; + tensor koi_5_cast_fp16 = add(x = var_747_cast_fp16, y = var_748_cast_fp16)[name = tensor("koi_5_cast_fp16")]; + tensor qo_3_axis_0 = const()[name = tensor("qo_3_axis_0"), val = tensor(-1)]; + tensor qo_3_cast_fp16 = stack(axis = qo_3_axis_0, values = (qor_5_cast_fp16, qoi_5_cast_fp16))[name = tensor("qo_3_cast_fp16")]; + tensor ko_3_axis_0 = const()[name = tensor("ko_3_axis_0"), val = tensor(-1)]; + tensor ko_3_cast_fp16 = stack(axis = ko_3_axis_0, values = (kor_5_cast_fp16, koi_5_cast_fp16))[name = tensor("ko_3_cast_fp16")]; + tensor var_781 = const()[name = tensor("op_781"), val = tensor([1, 256, 16, 64])]; + tensor q_9_cast_fp16 = reshape(shape = var_781, x = qo_3_cast_fp16)[name = tensor("q_9_cast_fp16")]; + tensor var_787 = const()[name = tensor("op_787"), val = tensor([1, 256, 16, 64])]; + tensor k_7_cast_fp16 = reshape(shape = var_787, x = ko_3_cast_fp16)[name = tensor("k_7_cast_fp16")]; + tensor write_indices_float_7_cast_fp16 = add(x = var_676_cast_fp16, y = var_393_promoted_to_fp16)[name = tensor("write_indices_float_7_cast_fp16")]; + tensor _inversed_808_y_0_to_fp16 = const()[name = tensor("_inversed_808_y_0_to_fp16"), val = tensor(0x1p-9)]; + tensor _inversed_808_cast_fp16 = mul(x = write_indices_float_7_cast_fp16, y = _inversed_808_y_0_to_fp16)[name = tensor("_inversed_808_cast_fp16")]; + tensor var_809_cast_fp16 = floor(x = _inversed_808_cast_fp16)[name = tensor("op_809_cast_fp16")]; + tensor var_810_to_fp16 = const()[name = tensor("op_810_to_fp16"), val = tensor(0x1p+9)]; + tensor var_811_cast_fp16 = mul(x = var_809_cast_fp16, y = var_810_to_fp16)[name = tensor("op_811_cast_fp16")]; + tensor write_indices_float_9_cast_fp16 = sub(x = write_indices_float_7_cast_fp16, y = var_811_cast_fp16)[name = tensor("write_indices_float_9_cast_fp16")]; + tensor var_829_value_0_to_fp16 = const()[name = tensor("op_829_value_0_to_fp16"), val = tensor(0x1.ffp+8)]; + tensor var_829_cast_fp16 = fill_like(ref_tensor = write_indices_float_9_cast_fp16, value = var_829_value_0_to_fp16)[name = tensor("op_829_cast_fp16")]; + tensor write_indices_float_11_cast_fp16 = select(a = write_indices_float_9_cast_fp16, b = var_829_cast_fp16, cond = token_valid_1_cast_fp16)[name = tensor("write_indices_float_11_cast_fp16")]; + tensor var_835_dtype_0 = const()[name = tensor("op_835_dtype_0"), val = tensor("int32")]; + tensor var_840 = const()[name = tensor("op_840"), val = tensor([1, 256, 1, 1])]; + tensor write_indices_float_11_cast_fp16_to_int32 = cast(dtype = var_835_dtype_0, x = write_indices_float_11_cast_fp16)[name = tensor("cast_655")]; + tensor var_841 = reshape(shape = var_840, x = write_indices_float_11_cast_fp16_to_int32)[name = tensor("op_841")]; + tensor write_indices_3_reps_0 = const()[name = tensor("write_indices_3_reps_0"), val = tensor([1, 1, 16, 64])]; + tensor write_indices_3 = tile(reps = write_indices_3_reps_0, x = var_841)[name = tensor("write_indices_3")]; + tensor var_851_begin_0 = const()[name = tensor("op_851_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_851_end_0 = const()[name = tensor("op_851_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor var_851_end_mask_0 = const()[name = tensor("op_851_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_851_squeeze_mask_0 = const()[name = tensor("op_851_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor cache1_to_fp16_dtype_0 = const()[name = tensor("cache1_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor cache1_to_fp16 = cast(dtype = cache1_to_fp16_dtype_0, x = cache1)[name = tensor("cast_654")]; + tensor var_851_cast_fp16 = slice_by_index(begin = var_851_begin_0, end = var_851_end_0, end_mask = var_851_end_mask_0, squeeze_mask = var_851_squeeze_mask_0, x = cache1_to_fp16)[name = tensor("op_851_cast_fp16")]; + tensor var_853_axis_0 = const()[name = tensor("op_853_axis_0"), val = tensor(1)]; + tensor var_853_mode_0 = const()[name = tensor("op_853_mode_0"), val = tensor("update")]; + tensor var_853_validate_indices_0 = const()[name = tensor("op_853_validate_indices_0"), val = tensor(false)]; + tensor var_853_cast_fp16 = scatter_along_axis(axis = var_853_axis_0, data = var_851_cast_fp16, indices = write_indices_3, mode = var_853_mode_0, updates = k_7_cast_fp16, validate_indices = var_853_validate_indices_0)[name = tensor("op_853_cast_fp16")]; + tensor concat_10 = const()[name = tensor("concat_10"), val = tensor([0, 0, 0, 0, 0])]; + tensor concat_11 = const()[name = tensor("concat_11"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_3_internal_tensor_assign_1_stride_0 = const()[name = tensor("new_cache_3_internal_tensor_assign_1_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_3_internal_tensor_assign_1_begin_mask_0 = const()[name = tensor("new_cache_3_internal_tensor_assign_1_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_3_internal_tensor_assign_1_end_mask_0 = const()[name = tensor("new_cache_3_internal_tensor_assign_1_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_3_internal_tensor_assign_1_squeeze_mask_0 = const()[name = tensor("new_cache_3_internal_tensor_assign_1_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_48 = const()[name = tensor("shape_48"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_2 = const()[name = tensor("reduce_prod_2"), val = tensor(1048576)]; + tensor range_1d_2_start_0 = const()[name = tensor("range_1d_2_start_0"), val = tensor(0)]; + tensor range_1d_2_step_0 = const()[name = tensor("range_1d_2_step_0"), val = tensor(1)]; + tensor range_1d_2 = range_1d(end = reduce_prod_2, start = range_1d_2_start_0, step = range_1d_2_step_0)[name = tensor("range_1d_2")]; + tensor reshape_10 = reshape(shape = shape_48, x = range_1d_2)[name = tensor("reshape_10")]; + tensor slice_by_index_2 = slice_by_index(begin = concat_10, begin_mask = new_cache_3_internal_tensor_assign_1_begin_mask_0, end = concat_11, end_mask = new_cache_3_internal_tensor_assign_1_end_mask_0, squeeze_mask = new_cache_3_internal_tensor_assign_1_squeeze_mask_0, stride = new_cache_3_internal_tensor_assign_1_stride_0, x = reshape_10)[name = tensor("slice_by_index_2")]; + tensor reshape_11_shape_0 = const()[name = tensor("reshape_11_shape_0"), val = tensor([-1])]; + tensor reshape_11 = reshape(shape = reshape_11_shape_0, x = slice_by_index_2)[name = tensor("reshape_11")]; + tensor reshape_12_shape_0 = const()[name = tensor("reshape_12_shape_0"), val = tensor([-1])]; + tensor reshape_12_cast_fp16 = reshape(shape = reshape_12_shape_0, x = var_853_cast_fp16)[name = tensor("reshape_12_cast_fp16")]; + tensor reshape_13_shape_0 = const()[name = tensor("reshape_13_shape_0"), val = tensor([-1])]; + tensor reshape_13_cast_fp16 = reshape(shape = reshape_13_shape_0, x = cache1_to_fp16)[name = tensor("reshape_13_cast_fp16")]; + tensor scatter_2_mode_0 = const()[name = tensor("scatter_2_mode_0"), val = tensor("update")]; + tensor scatter_2_axis_0 = const()[name = tensor("scatter_2_axis_0"), val = tensor(0)]; + tensor scatter_2_validate_indices_0 = const()[name = tensor("scatter_2_validate_indices_0"), val = tensor(false)]; + tensor scatter_2_cast_fp16 = scatter(axis = scatter_2_axis_0, data = reshape_13_cast_fp16, indices = reshape_11, mode = scatter_2_mode_0, updates = reshape_12_cast_fp16, validate_indices = scatter_2_validate_indices_0)[name = tensor("scatter_2_cast_fp16")]; + tensor reshape_14_cast_fp16 = reshape(shape = shape_48, x = scatter_2_cast_fp16)[name = tensor("reshape_14_cast_fp16")]; + tensor var_861_begin_0 = const()[name = tensor("op_861_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor var_861_end_0 = const()[name = tensor("op_861_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor var_861_end_mask_0 = const()[name = tensor("op_861_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_861_squeeze_mask_0 = const()[name = tensor("op_861_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor var_861_cast_fp16 = slice_by_index(begin = var_861_begin_0, end = var_861_end_0, end_mask = var_861_end_mask_0, squeeze_mask = var_861_squeeze_mask_0, x = reshape_14_cast_fp16)[name = tensor("op_861_cast_fp16")]; + tensor var_863_axis_0 = const()[name = tensor("op_863_axis_0"), val = tensor(1)]; + tensor var_863_mode_0 = const()[name = tensor("op_863_mode_0"), val = tensor("update")]; + tensor var_863_validate_indices_0 = const()[name = tensor("op_863_validate_indices_0"), val = tensor(false)]; + tensor var_863_cast_fp16 = scatter_along_axis(axis = var_863_axis_0, data = var_861_cast_fp16, indices = write_indices_3, mode = var_863_mode_0, updates = v_3_cast_fp16, validate_indices = var_863_validate_indices_0)[name = tensor("op_863_cast_fp16")]; + tensor concat_12 = const()[name = tensor("concat_12"), val = tensor([1, 0, 0, 0, 0])]; + tensor concat_13 = const()[name = tensor("concat_13"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_3_internal_tensor_assign_2_stride_0 = const()[name = tensor("new_cache_3_internal_tensor_assign_2_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_3_internal_tensor_assign_2_begin_mask_0 = const()[name = tensor("new_cache_3_internal_tensor_assign_2_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_3_internal_tensor_assign_2_end_mask_0 = const()[name = tensor("new_cache_3_internal_tensor_assign_2_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_3_internal_tensor_assign_2_squeeze_mask_0 = const()[name = tensor("new_cache_3_internal_tensor_assign_2_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_49 = const()[name = tensor("shape_49"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_3 = const()[name = tensor("reduce_prod_3"), val = tensor(1048576)]; + tensor range_1d_3_start_0 = const()[name = tensor("range_1d_3_start_0"), val = tensor(0)]; + tensor range_1d_3_step_0 = const()[name = tensor("range_1d_3_step_0"), val = tensor(1)]; + tensor range_1d_3 = range_1d(end = reduce_prod_3, start = range_1d_3_start_0, step = range_1d_3_step_0)[name = tensor("range_1d_3")]; + tensor reshape_15 = reshape(shape = shape_49, x = range_1d_3)[name = tensor("reshape_15")]; + tensor slice_by_index_3 = slice_by_index(begin = concat_12, begin_mask = new_cache_3_internal_tensor_assign_2_begin_mask_0, end = concat_13, end_mask = new_cache_3_internal_tensor_assign_2_end_mask_0, squeeze_mask = new_cache_3_internal_tensor_assign_2_squeeze_mask_0, stride = new_cache_3_internal_tensor_assign_2_stride_0, x = reshape_15)[name = tensor("slice_by_index_3")]; + tensor reshape_16_shape_0 = const()[name = tensor("reshape_16_shape_0"), val = tensor([-1])]; + tensor reshape_16 = reshape(shape = reshape_16_shape_0, x = slice_by_index_3)[name = tensor("reshape_16")]; + tensor reshape_17_shape_0 = const()[name = tensor("reshape_17_shape_0"), val = tensor([-1])]; + tensor reshape_17_cast_fp16 = reshape(shape = reshape_17_shape_0, x = var_863_cast_fp16)[name = tensor("reshape_17_cast_fp16")]; + tensor reshape_18_shape_0 = const()[name = tensor("reshape_18_shape_0"), val = tensor([-1])]; + tensor reshape_18_cast_fp16 = reshape(shape = reshape_18_shape_0, x = reshape_14_cast_fp16)[name = tensor("reshape_18_cast_fp16")]; + tensor scatter_3_mode_0 = const()[name = tensor("scatter_3_mode_0"), val = tensor("update")]; + tensor scatter_3_axis_0 = const()[name = tensor("scatter_3_axis_0"), val = tensor(0)]; + tensor scatter_3_validate_indices_0 = const()[name = tensor("scatter_3_validate_indices_0"), val = tensor(false)]; + tensor scatter_3_cast_fp16 = scatter(axis = scatter_3_axis_0, data = reshape_18_cast_fp16, indices = reshape_16, mode = scatter_3_mode_0, updates = reshape_17_cast_fp16, validate_indices = scatter_3_validate_indices_0)[name = tensor("scatter_3_cast_fp16")]; + tensor reshape_19_cast_fp16 = reshape(shape = shape_49, x = scatter_3_cast_fp16)[name = tensor("reshape_19_cast_fp16")]; + tensor reshape_19_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("reshape_19_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor keys_7_begin_0 = const()[name = tensor("keys_7_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor keys_7_end_0 = const()[name = tensor("keys_7_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor keys_7_end_mask_0 = const()[name = tensor("keys_7_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor keys_7_squeeze_mask_0 = const()[name = tensor("keys_7_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor keys_7_cast_fp16 = slice_by_index(begin = keys_7_begin_0, end = keys_7_end_0, end_mask = keys_7_end_mask_0, squeeze_mask = keys_7_squeeze_mask_0, x = reshape_19_cast_fp16)[name = tensor("keys_7_cast_fp16")]; + tensor values_7_begin_0 = const()[name = tensor("values_7_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor values_7_end_0 = const()[name = tensor("values_7_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor values_7_end_mask_0 = const()[name = tensor("values_7_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor values_7_squeeze_mask_0 = const()[name = tensor("values_7_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor values_7_cast_fp16 = slice_by_index(begin = values_7_begin_0, end = values_7_end_0, end_mask = values_7_end_mask_0, squeeze_mask = values_7_squeeze_mask_0, x = reshape_19_cast_fp16)[name = tensor("values_7_cast_fp16")]; + tensor var_875_cast_fp16 = not_equal(x = keys_7_cast_fp16, y = keys_7_cast_fp16)[name = tensor("op_875_cast_fp16")]; + tensor keys_9_cast_fp16 = select(a = var_470_to_fp16, b = keys_7_cast_fp16, cond = var_875_cast_fp16)[name = tensor("keys_9_cast_fp16")]; + tensor var_883_cast_fp16 = not_equal(x = values_7_cast_fp16, y = values_7_cast_fp16)[name = tensor("op_883_cast_fp16")]; + tensor values_9_cast_fp16 = select(a = var_470_to_fp16, b = values_7_cast_fp16, cond = var_883_cast_fp16)[name = tensor("values_9_cast_fp16")]; + tensor var_907 = const()[name = tensor("op_907"), val = tensor([0, 2, 1, 3])]; + tensor var_923 = const()[name = tensor("op_923"), val = tensor([1, 1, 1])]; + tensor var_924_cast_fp16 = reshape(shape = var_923, x = position1_to_fp16)[name = tensor("op_924_cast_fp16")]; + tensor q_positions_3_cast_fp16 = add(x = var_924_cast_fp16, y = q_offsets_1_promoted_to_fp16)[name = tensor("q_positions_3_cast_fp16")]; + tensor valid_end_3_cast_fp16 = add(x = var_924_cast_fp16, y = var_536_cast_fp16)[name = tensor("valid_end_3_cast_fp16")]; + tensor valid_mask_3_cast_fp16 = less(x = k_positions_1_promoted_to_fp16, y = valid_end_3_cast_fp16)[name = tensor("valid_mask_3_cast_fp16")]; + tensor causal_mask_3_cast_fp16 = less_equal(x = k_positions_1_promoted_to_fp16, y = q_positions_3_cast_fp16)[name = tensor("causal_mask_3_cast_fp16")]; + tensor var_952 = logical_and(x = valid_mask_3_cast_fp16, y = causal_mask_3_cast_fp16)[name = tensor("op_952")]; + tensor attn_mask_3_axes_0 = const()[name = tensor("attn_mask_3_axes_0"), val = tensor([1])]; + tensor attn_mask_3 = expand_dims(axes = attn_mask_3_axes_0, x = var_952)[name = tensor("attn_mask_3")]; + tensor var_966_transpose_x_0 = const()[name = tensor("op_966_transpose_x_0"), val = tensor(false)]; + tensor var_966_transpose_y_0 = const()[name = tensor("op_966_transpose_y_0"), val = tensor(false)]; + tensor transpose_71_perm_0 = const()[name = tensor("transpose_71_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_72_perm_0 = const()[name = tensor("transpose_72_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_72 = transpose(perm = transpose_72_perm_0, x = keys_9_cast_fp16)[name = tensor("transpose_200")]; + tensor transpose_71 = transpose(perm = transpose_71_perm_0, x = q_9_cast_fp16)[name = tensor("transpose_201")]; + tensor var_966_cast_fp16 = matmul(transpose_x = var_966_transpose_x_0, transpose_y = var_966_transpose_y_0, x = transpose_71, y = transpose_72)[name = tensor("op_966_cast_fp16")]; + tensor var_960_to_fp16 = const()[name = tensor("op_960_to_fp16"), val = tensor([0x1p-3])]; + tensor attn_weights_7_cast_fp16 = mul(x = var_966_cast_fp16, y = var_960_to_fp16)[name = tensor("attn_weights_7_cast_fp16")]; + tensor var_968 = logical_not(x = attn_mask_3)[name = tensor("op_968")]; + tensor var_969_to_fp16 = const()[name = tensor("op_969_to_fp16"), val = tensor(-inf)]; + tensor attn_weights_9_cast_fp16 = select(a = var_969_to_fp16, b = attn_weights_7_cast_fp16, cond = var_968)[name = tensor("attn_weights_9_cast_fp16")]; + tensor var_971 = const()[name = tensor("op_971"), val = tensor(-1)]; + tensor attn_weights_11_cast_fp16 = softmax(axis = var_971, x = attn_weights_9_cast_fp16)[name = tensor("attn_weights_11_cast_fp16")]; + tensor attn_output_3_transpose_x_0 = const()[name = tensor("attn_output_3_transpose_x_0"), val = tensor(false)]; + tensor attn_output_3_transpose_y_0 = const()[name = tensor("attn_output_3_transpose_y_0"), val = tensor(false)]; + tensor values_11_cast_fp16 = transpose(perm = var_907, x = values_9_cast_fp16)[name = tensor("transpose_202")]; + tensor attn_output_3_cast_fp16 = matmul(transpose_x = attn_output_3_transpose_x_0, transpose_y = attn_output_3_transpose_y_0, x = attn_weights_11_cast_fp16, y = values_11_cast_fp16)[name = tensor("attn_output_3_cast_fp16")]; + tensor var_979 = const()[name = tensor("op_979"), val = tensor([0, 2, 1, 3])]; + tensor var_984 = const()[name = tensor("op_984"), val = tensor([1, 256, 1024])]; + tensor var_980_cast_fp16 = transpose(perm = var_979, x = attn_output_3_cast_fp16)[name = tensor("transpose_199")]; + tensor input_17_cast_fp16 = reshape(shape = var_984, x = var_980_cast_fp16)[name = tensor("input_17_cast_fp16")]; + tensor attn1_out_proj_weight_to_fp16 = const()[name = tensor("attn1_out_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(32538624)))]; + tensor linear_5_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = attn1_out_proj_weight_to_fp16, x = input_17_cast_fp16)[name = tensor("linear_5_cast_fp16")]; + tensor var_993_cast_fp16 = add(x = position1_to_fp16, y = valid_len_to_fp16)[name = tensor("op_993_cast_fp16")]; + tensor var_993_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_993_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor input_19_cast_fp16 = add(x = input_13_cast_fp16, y = linear_5_cast_fp16)[name = tensor("input_19_cast_fp16")]; + tensor input_21_axes_0 = const()[name = tensor("input_21_axes_0"), val = tensor([-1])]; + tensor norm1_2_weight_to_fp16 = const()[name = tensor("norm1_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(34635840)))]; + tensor norm1_2_bias_to_fp16 = const()[name = tensor("norm1_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(34637952)))]; + tensor var_997_to_fp16 = const()[name = tensor("op_997_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_21_cast_fp16 = layer_norm(axes = input_21_axes_0, beta = norm1_2_bias_to_fp16, epsilon = var_997_to_fp16, gamma = norm1_2_weight_to_fp16, x = input_19_cast_fp16)[name = tensor("input_21_cast_fp16")]; + tensor linear1_1_weight_to_fp16 = const()[name = tensor("linear1_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(34640064)))]; + tensor linear_6_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = linear1_1_weight_to_fp16, x = input_21_cast_fp16)[name = tensor("linear_6_cast_fp16")]; + tensor input_23_mode_0 = const()[name = tensor("input_23_mode_0"), val = tensor("EXACT")]; + tensor input_23_cast_fp16 = gelu(mode = input_23_mode_0, x = linear_6_cast_fp16)[name = tensor("input_23_cast_fp16")]; + tensor linear1_2_weight_to_fp16 = const()[name = tensor("linear1_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(43028736)))]; + tensor linear_7_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = linear1_2_weight_to_fp16, x = input_23_cast_fp16)[name = tensor("linear_7_cast_fp16")]; + tensor input_25_cast_fp16 = add(x = input_19_cast_fp16, y = linear_7_cast_fp16)[name = tensor("input_25_cast_fp16")]; + tensor input_27_axes_0 = const()[name = tensor("input_27_axes_0"), val = tensor([-1])]; + tensor norm2_1_weight_to_fp16 = const()[name = tensor("norm2_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(51417408)))]; + tensor norm2_1_bias_to_fp16 = const()[name = tensor("norm2_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(51419520)))]; + tensor var_1014_to_fp16 = const()[name = tensor("op_1014_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_27_cast_fp16 = layer_norm(axes = input_27_axes_0, beta = norm2_1_bias_to_fp16, epsilon = var_1014_to_fp16, gamma = norm2_1_weight_to_fp16, x = input_25_cast_fp16)[name = tensor("input_27_cast_fp16")]; + tensor attn2_in_proj_weight_to_fp16 = const()[name = tensor("attn2_in_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(51421632)))]; + tensor linear_8_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = attn2_in_proj_weight_to_fp16, x = input_27_cast_fp16)[name = tensor("linear_8_cast_fp16")]; + tensor var_1028 = const()[name = tensor("op_1028"), val = tensor([1, 256, 3, 16, 64])]; + tensor qkv_5_cast_fp16 = reshape(shape = var_1028, x = linear_8_cast_fp16)[name = tensor("qkv_5_cast_fp16")]; + tensor q_13_begin_0 = const()[name = tensor("q_13_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor q_13_end_0 = const()[name = tensor("q_13_end_0"), val = tensor([1, 256, 1, 16, 64])]; + tensor q_13_end_mask_0 = const()[name = tensor("q_13_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor q_13_squeeze_mask_0 = const()[name = tensor("q_13_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor q_13_cast_fp16 = slice_by_index(begin = q_13_begin_0, end = q_13_end_0, end_mask = q_13_end_mask_0, squeeze_mask = q_13_squeeze_mask_0, x = qkv_5_cast_fp16)[name = tensor("q_13_cast_fp16")]; + tensor k_9_begin_0 = const()[name = tensor("k_9_begin_0"), val = tensor([0, 0, 1, 0, 0])]; + tensor k_9_end_0 = const()[name = tensor("k_9_end_0"), val = tensor([1, 256, 2, 16, 64])]; + tensor k_9_end_mask_0 = const()[name = tensor("k_9_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor k_9_squeeze_mask_0 = const()[name = tensor("k_9_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor k_9_cast_fp16 = slice_by_index(begin = k_9_begin_0, end = k_9_end_0, end_mask = k_9_end_mask_0, squeeze_mask = k_9_squeeze_mask_0, x = qkv_5_cast_fp16)[name = tensor("k_9_cast_fp16")]; + tensor v_5_begin_0 = const()[name = tensor("v_5_begin_0"), val = tensor([0, 0, 2, 0, 0])]; + tensor v_5_end_0 = const()[name = tensor("v_5_end_0"), val = tensor([1, 256, 3, 16, 64])]; + tensor v_5_end_mask_0 = const()[name = tensor("v_5_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor v_5_squeeze_mask_0 = const()[name = tensor("v_5_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor v_5_cast_fp16 = slice_by_index(begin = v_5_begin_0, end = v_5_end_0, end_mask = v_5_end_mask_0, squeeze_mask = v_5_squeeze_mask_0, x = qkv_5_cast_fp16)[name = tensor("v_5_cast_fp16")]; + tensor var_1086 = const()[name = tensor("op_1086"), val = tensor([1, 1])]; + tensor position2_to_fp16_dtype_0 = const()[name = tensor("position2_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor position2_to_fp16 = cast(dtype = position2_to_fp16_dtype_0, x = position2)[name = tensor("cast_651")]; + tensor var_1087_cast_fp16 = reshape(shape = var_1086, x = position2_to_fp16)[name = tensor("op_1087_cast_fp16")]; + tensor ts_15_cast_fp16 = add(x = ts_1_promoted_to_fp16, y = var_1087_cast_fp16)[name = tensor("ts_15_cast_fp16")]; + tensor var_1094 = const()[name = tensor("op_1094"), val = tensor([1, 256, 1, 1])]; + tensor ts_17_cast_fp16 = reshape(shape = var_1094, x = ts_15_cast_fp16)[name = tensor("ts_17_cast_fp16")]; + tensor var_1101 = const()[name = tensor("op_1101"), val = tensor([1, 256, 16, 32, 2])]; + tensor q_complex_5_cast_fp16 = reshape(shape = var_1101, x = q_13_cast_fp16)[name = tensor("q_complex_5_cast_fp16")]; + tensor var_1108 = const()[name = tensor("op_1108"), val = tensor([1, 256, 16, 32, 2])]; + tensor k_complex_5_cast_fp16 = reshape(shape = var_1108, x = k_9_cast_fp16)[name = tensor("k_complex_5_cast_fp16")]; + tensor var_1112_begin_0 = const()[name = tensor("op_1112_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_1112_end_0 = const()[name = tensor("op_1112_end_0"), val = tensor([1, 256, 16, 32, 1])]; + tensor var_1112_end_mask_0 = const()[name = tensor("op_1112_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_1112_squeeze_mask_0 = const()[name = tensor("op_1112_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_1112_cast_fp16 = slice_by_index(begin = var_1112_begin_0, end = var_1112_end_0, end_mask = var_1112_end_mask_0, squeeze_mask = var_1112_squeeze_mask_0, x = q_complex_5_cast_fp16)[name = tensor("op_1112_cast_fp16")]; + tensor var_1120_begin_0 = const()[name = tensor("op_1120_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_1120_end_0 = const()[name = tensor("op_1120_end_0"), val = tensor([1, 256, 16, 32, 2])]; + tensor var_1120_end_mask_0 = const()[name = tensor("op_1120_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_1120_squeeze_mask_0 = const()[name = tensor("op_1120_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_1120_cast_fp16 = slice_by_index(begin = var_1120_begin_0, end = var_1120_end_0, end_mask = var_1120_end_mask_0, squeeze_mask = var_1120_squeeze_mask_0, x = q_complex_5_cast_fp16)[name = tensor("op_1120_cast_fp16")]; + tensor var_1128_begin_0 = const()[name = tensor("op_1128_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_1128_end_0 = const()[name = tensor("op_1128_end_0"), val = tensor([1, 256, 16, 32, 1])]; + tensor var_1128_end_mask_0 = const()[name = tensor("op_1128_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_1128_squeeze_mask_0 = const()[name = tensor("op_1128_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_1128_cast_fp16 = slice_by_index(begin = var_1128_begin_0, end = var_1128_end_0, end_mask = var_1128_end_mask_0, squeeze_mask = var_1128_squeeze_mask_0, x = k_complex_5_cast_fp16)[name = tensor("op_1128_cast_fp16")]; + tensor var_1136_begin_0 = const()[name = tensor("op_1136_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_1136_end_0 = const()[name = tensor("op_1136_end_0"), val = tensor([1, 256, 16, 32, 2])]; + tensor var_1136_end_mask_0 = const()[name = tensor("op_1136_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_1136_squeeze_mask_0 = const()[name = tensor("op_1136_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_1136_cast_fp16 = slice_by_index(begin = var_1136_begin_0, end = var_1136_end_0, end_mask = var_1136_end_mask_0, squeeze_mask = var_1136_squeeze_mask_0, x = k_complex_5_cast_fp16)[name = tensor("op_1136_cast_fp16")]; + tensor freqs_5_to_fp16 = const()[name = tensor("freqs_5_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(57713152)))]; + tensor var_1142_cast_fp16 = mul(x = freqs_5_to_fp16, y = ts_17_cast_fp16)[name = tensor("op_1142_cast_fp16")]; + tensor rotr_5_cast_fp16 = cos(x = var_1142_cast_fp16)[name = tensor("rotr_5_cast_fp16")]; + tensor roti_5_cast_fp16 = sin(x = var_1142_cast_fp16)[name = tensor("roti_5_cast_fp16")]; + tensor var_1146_cast_fp16 = mul(x = var_1112_cast_fp16, y = rotr_5_cast_fp16)[name = tensor("op_1146_cast_fp16")]; + tensor var_1147_cast_fp16 = mul(x = var_1120_cast_fp16, y = roti_5_cast_fp16)[name = tensor("op_1147_cast_fp16")]; + tensor qor_9_cast_fp16 = sub(x = var_1146_cast_fp16, y = var_1147_cast_fp16)[name = tensor("qor_9_cast_fp16")]; + tensor var_1150_cast_fp16 = mul(x = var_1112_cast_fp16, y = roti_5_cast_fp16)[name = tensor("op_1150_cast_fp16")]; + tensor var_1151_cast_fp16 = mul(x = var_1120_cast_fp16, y = rotr_5_cast_fp16)[name = tensor("op_1151_cast_fp16")]; + tensor qoi_9_cast_fp16 = add(x = var_1150_cast_fp16, y = var_1151_cast_fp16)[name = tensor("qoi_9_cast_fp16")]; + tensor var_1154_cast_fp16 = mul(x = var_1128_cast_fp16, y = rotr_5_cast_fp16)[name = tensor("op_1154_cast_fp16")]; + tensor var_1155_cast_fp16 = mul(x = var_1136_cast_fp16, y = roti_5_cast_fp16)[name = tensor("op_1155_cast_fp16")]; + tensor kor_9_cast_fp16 = sub(x = var_1154_cast_fp16, y = var_1155_cast_fp16)[name = tensor("kor_9_cast_fp16")]; + tensor var_1158_cast_fp16 = mul(x = var_1128_cast_fp16, y = roti_5_cast_fp16)[name = tensor("op_1158_cast_fp16")]; + tensor var_1159_cast_fp16 = mul(x = var_1136_cast_fp16, y = rotr_5_cast_fp16)[name = tensor("op_1159_cast_fp16")]; + tensor koi_9_cast_fp16 = add(x = var_1158_cast_fp16, y = var_1159_cast_fp16)[name = tensor("koi_9_cast_fp16")]; + tensor qo_5_axis_0 = const()[name = tensor("qo_5_axis_0"), val = tensor(-1)]; + tensor qo_5_cast_fp16 = stack(axis = qo_5_axis_0, values = (qor_9_cast_fp16, qoi_9_cast_fp16))[name = tensor("qo_5_cast_fp16")]; + tensor ko_5_axis_0 = const()[name = tensor("ko_5_axis_0"), val = tensor(-1)]; + tensor ko_5_cast_fp16 = stack(axis = ko_5_axis_0, values = (kor_9_cast_fp16, koi_9_cast_fp16))[name = tensor("ko_5_cast_fp16")]; + tensor var_1192 = const()[name = tensor("op_1192"), val = tensor([1, 256, 16, 64])]; + tensor q_15_cast_fp16 = reshape(shape = var_1192, x = qo_5_cast_fp16)[name = tensor("q_15_cast_fp16")]; + tensor var_1198 = const()[name = tensor("op_1198"), val = tensor([1, 256, 16, 64])]; + tensor k_11_cast_fp16 = reshape(shape = var_1198, x = ko_5_cast_fp16)[name = tensor("k_11_cast_fp16")]; + tensor write_indices_float_13_cast_fp16 = add(x = var_1087_cast_fp16, y = var_393_promoted_to_fp16)[name = tensor("write_indices_float_13_cast_fp16")]; + tensor _inversed_1219_y_0_to_fp16 = const()[name = tensor("_inversed_1219_y_0_to_fp16"), val = tensor(0x1p-9)]; + tensor _inversed_1219_cast_fp16 = mul(x = write_indices_float_13_cast_fp16, y = _inversed_1219_y_0_to_fp16)[name = tensor("_inversed_1219_cast_fp16")]; + tensor var_1220_cast_fp16 = floor(x = _inversed_1219_cast_fp16)[name = tensor("op_1220_cast_fp16")]; + tensor var_1221_to_fp16 = const()[name = tensor("op_1221_to_fp16"), val = tensor(0x1p+9)]; + tensor var_1222_cast_fp16 = mul(x = var_1220_cast_fp16, y = var_1221_to_fp16)[name = tensor("op_1222_cast_fp16")]; + tensor write_indices_float_15_cast_fp16 = sub(x = write_indices_float_13_cast_fp16, y = var_1222_cast_fp16)[name = tensor("write_indices_float_15_cast_fp16")]; + tensor var_1240_value_0_to_fp16 = const()[name = tensor("op_1240_value_0_to_fp16"), val = tensor(0x1.ffp+8)]; + tensor var_1240_cast_fp16 = fill_like(ref_tensor = write_indices_float_15_cast_fp16, value = var_1240_value_0_to_fp16)[name = tensor("op_1240_cast_fp16")]; + tensor write_indices_float_17_cast_fp16 = select(a = write_indices_float_15_cast_fp16, b = var_1240_cast_fp16, cond = token_valid_1_cast_fp16)[name = tensor("write_indices_float_17_cast_fp16")]; + tensor var_1246_dtype_0 = const()[name = tensor("op_1246_dtype_0"), val = tensor("int32")]; + tensor var_1251 = const()[name = tensor("op_1251"), val = tensor([1, 256, 1, 1])]; + tensor write_indices_float_17_cast_fp16_to_int32 = cast(dtype = var_1246_dtype_0, x = write_indices_float_17_cast_fp16)[name = tensor("cast_650")]; + tensor var_1252 = reshape(shape = var_1251, x = write_indices_float_17_cast_fp16_to_int32)[name = tensor("op_1252")]; + tensor write_indices_5_reps_0 = const()[name = tensor("write_indices_5_reps_0"), val = tensor([1, 1, 16, 64])]; + tensor write_indices_5 = tile(reps = write_indices_5_reps_0, x = var_1252)[name = tensor("write_indices_5")]; + tensor var_1262_begin_0 = const()[name = tensor("op_1262_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_1262_end_0 = const()[name = tensor("op_1262_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor var_1262_end_mask_0 = const()[name = tensor("op_1262_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_1262_squeeze_mask_0 = const()[name = tensor("op_1262_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor cache2_to_fp16_dtype_0 = const()[name = tensor("cache2_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor cache2_to_fp16 = cast(dtype = cache2_to_fp16_dtype_0, x = cache2)[name = tensor("cast_649")]; + tensor var_1262_cast_fp16 = slice_by_index(begin = var_1262_begin_0, end = var_1262_end_0, end_mask = var_1262_end_mask_0, squeeze_mask = var_1262_squeeze_mask_0, x = cache2_to_fp16)[name = tensor("op_1262_cast_fp16")]; + tensor var_1264_axis_0 = const()[name = tensor("op_1264_axis_0"), val = tensor(1)]; + tensor var_1264_mode_0 = const()[name = tensor("op_1264_mode_0"), val = tensor("update")]; + tensor var_1264_validate_indices_0 = const()[name = tensor("op_1264_validate_indices_0"), val = tensor(false)]; + tensor var_1264_cast_fp16 = scatter_along_axis(axis = var_1264_axis_0, data = var_1262_cast_fp16, indices = write_indices_5, mode = var_1264_mode_0, updates = k_11_cast_fp16, validate_indices = var_1264_validate_indices_0)[name = tensor("op_1264_cast_fp16")]; + tensor concat_18 = const()[name = tensor("concat_18"), val = tensor([0, 0, 0, 0, 0])]; + tensor concat_19 = const()[name = tensor("concat_19"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_5_internal_tensor_assign_1_stride_0 = const()[name = tensor("new_cache_5_internal_tensor_assign_1_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_5_internal_tensor_assign_1_begin_mask_0 = const()[name = tensor("new_cache_5_internal_tensor_assign_1_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_5_internal_tensor_assign_1_end_mask_0 = const()[name = tensor("new_cache_5_internal_tensor_assign_1_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_5_internal_tensor_assign_1_squeeze_mask_0 = const()[name = tensor("new_cache_5_internal_tensor_assign_1_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_50 = const()[name = tensor("shape_50"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_4 = const()[name = tensor("reduce_prod_4"), val = tensor(1048576)]; + tensor range_1d_4_start_0 = const()[name = tensor("range_1d_4_start_0"), val = tensor(0)]; + tensor range_1d_4_step_0 = const()[name = tensor("range_1d_4_step_0"), val = tensor(1)]; + tensor range_1d_4 = range_1d(end = reduce_prod_4, start = range_1d_4_start_0, step = range_1d_4_step_0)[name = tensor("range_1d_4")]; + tensor reshape_20 = reshape(shape = shape_50, x = range_1d_4)[name = tensor("reshape_20")]; + tensor slice_by_index_4 = slice_by_index(begin = concat_18, begin_mask = new_cache_5_internal_tensor_assign_1_begin_mask_0, end = concat_19, end_mask = new_cache_5_internal_tensor_assign_1_end_mask_0, squeeze_mask = new_cache_5_internal_tensor_assign_1_squeeze_mask_0, stride = new_cache_5_internal_tensor_assign_1_stride_0, x = reshape_20)[name = tensor("slice_by_index_4")]; + tensor reshape_21_shape_0 = const()[name = tensor("reshape_21_shape_0"), val = tensor([-1])]; + tensor reshape_21 = reshape(shape = reshape_21_shape_0, x = slice_by_index_4)[name = tensor("reshape_21")]; + tensor reshape_22_shape_0 = const()[name = tensor("reshape_22_shape_0"), val = tensor([-1])]; + tensor reshape_22_cast_fp16 = reshape(shape = reshape_22_shape_0, x = var_1264_cast_fp16)[name = tensor("reshape_22_cast_fp16")]; + tensor reshape_23_shape_0 = const()[name = tensor("reshape_23_shape_0"), val = tensor([-1])]; + tensor reshape_23_cast_fp16 = reshape(shape = reshape_23_shape_0, x = cache2_to_fp16)[name = tensor("reshape_23_cast_fp16")]; + tensor scatter_4_mode_0 = const()[name = tensor("scatter_4_mode_0"), val = tensor("update")]; + tensor scatter_4_axis_0 = const()[name = tensor("scatter_4_axis_0"), val = tensor(0)]; + tensor scatter_4_validate_indices_0 = const()[name = tensor("scatter_4_validate_indices_0"), val = tensor(false)]; + tensor scatter_4_cast_fp16 = scatter(axis = scatter_4_axis_0, data = reshape_23_cast_fp16, indices = reshape_21, mode = scatter_4_mode_0, updates = reshape_22_cast_fp16, validate_indices = scatter_4_validate_indices_0)[name = tensor("scatter_4_cast_fp16")]; + tensor reshape_24_cast_fp16 = reshape(shape = shape_50, x = scatter_4_cast_fp16)[name = tensor("reshape_24_cast_fp16")]; + tensor var_1272_begin_0 = const()[name = tensor("op_1272_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor var_1272_end_0 = const()[name = tensor("op_1272_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor var_1272_end_mask_0 = const()[name = tensor("op_1272_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_1272_squeeze_mask_0 = const()[name = tensor("op_1272_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor var_1272_cast_fp16 = slice_by_index(begin = var_1272_begin_0, end = var_1272_end_0, end_mask = var_1272_end_mask_0, squeeze_mask = var_1272_squeeze_mask_0, x = reshape_24_cast_fp16)[name = tensor("op_1272_cast_fp16")]; + tensor var_1274_axis_0 = const()[name = tensor("op_1274_axis_0"), val = tensor(1)]; + tensor var_1274_mode_0 = const()[name = tensor("op_1274_mode_0"), val = tensor("update")]; + tensor var_1274_validate_indices_0 = const()[name = tensor("op_1274_validate_indices_0"), val = tensor(false)]; + tensor var_1274_cast_fp16 = scatter_along_axis(axis = var_1274_axis_0, data = var_1272_cast_fp16, indices = write_indices_5, mode = var_1274_mode_0, updates = v_5_cast_fp16, validate_indices = var_1274_validate_indices_0)[name = tensor("op_1274_cast_fp16")]; + tensor concat_20 = const()[name = tensor("concat_20"), val = tensor([1, 0, 0, 0, 0])]; + tensor concat_21 = const()[name = tensor("concat_21"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_5_internal_tensor_assign_2_stride_0 = const()[name = tensor("new_cache_5_internal_tensor_assign_2_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_5_internal_tensor_assign_2_begin_mask_0 = const()[name = tensor("new_cache_5_internal_tensor_assign_2_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_5_internal_tensor_assign_2_end_mask_0 = const()[name = tensor("new_cache_5_internal_tensor_assign_2_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_5_internal_tensor_assign_2_squeeze_mask_0 = const()[name = tensor("new_cache_5_internal_tensor_assign_2_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_51 = const()[name = tensor("shape_51"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_5 = const()[name = tensor("reduce_prod_5"), val = tensor(1048576)]; + tensor range_1d_5_start_0 = const()[name = tensor("range_1d_5_start_0"), val = tensor(0)]; + tensor range_1d_5_step_0 = const()[name = tensor("range_1d_5_step_0"), val = tensor(1)]; + tensor range_1d_5 = range_1d(end = reduce_prod_5, start = range_1d_5_start_0, step = range_1d_5_step_0)[name = tensor("range_1d_5")]; + tensor reshape_25 = reshape(shape = shape_51, x = range_1d_5)[name = tensor("reshape_25")]; + tensor slice_by_index_5 = slice_by_index(begin = concat_20, begin_mask = new_cache_5_internal_tensor_assign_2_begin_mask_0, end = concat_21, end_mask = new_cache_5_internal_tensor_assign_2_end_mask_0, squeeze_mask = new_cache_5_internal_tensor_assign_2_squeeze_mask_0, stride = new_cache_5_internal_tensor_assign_2_stride_0, x = reshape_25)[name = tensor("slice_by_index_5")]; + tensor reshape_26_shape_0 = const()[name = tensor("reshape_26_shape_0"), val = tensor([-1])]; + tensor reshape_26 = reshape(shape = reshape_26_shape_0, x = slice_by_index_5)[name = tensor("reshape_26")]; + tensor reshape_27_shape_0 = const()[name = tensor("reshape_27_shape_0"), val = tensor([-1])]; + tensor reshape_27_cast_fp16 = reshape(shape = reshape_27_shape_0, x = var_1274_cast_fp16)[name = tensor("reshape_27_cast_fp16")]; + tensor reshape_28_shape_0 = const()[name = tensor("reshape_28_shape_0"), val = tensor([-1])]; + tensor reshape_28_cast_fp16 = reshape(shape = reshape_28_shape_0, x = reshape_24_cast_fp16)[name = tensor("reshape_28_cast_fp16")]; + tensor scatter_5_mode_0 = const()[name = tensor("scatter_5_mode_0"), val = tensor("update")]; + tensor scatter_5_axis_0 = const()[name = tensor("scatter_5_axis_0"), val = tensor(0)]; + tensor scatter_5_validate_indices_0 = const()[name = tensor("scatter_5_validate_indices_0"), val = tensor(false)]; + tensor scatter_5_cast_fp16 = scatter(axis = scatter_5_axis_0, data = reshape_28_cast_fp16, indices = reshape_26, mode = scatter_5_mode_0, updates = reshape_27_cast_fp16, validate_indices = scatter_5_validate_indices_0)[name = tensor("scatter_5_cast_fp16")]; + tensor reshape_29_cast_fp16 = reshape(shape = shape_51, x = scatter_5_cast_fp16)[name = tensor("reshape_29_cast_fp16")]; + tensor reshape_29_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("reshape_29_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor keys_13_begin_0 = const()[name = tensor("keys_13_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor keys_13_end_0 = const()[name = tensor("keys_13_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor keys_13_end_mask_0 = const()[name = tensor("keys_13_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor keys_13_squeeze_mask_0 = const()[name = tensor("keys_13_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor keys_13_cast_fp16 = slice_by_index(begin = keys_13_begin_0, end = keys_13_end_0, end_mask = keys_13_end_mask_0, squeeze_mask = keys_13_squeeze_mask_0, x = reshape_29_cast_fp16)[name = tensor("keys_13_cast_fp16")]; + tensor values_13_begin_0 = const()[name = tensor("values_13_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor values_13_end_0 = const()[name = tensor("values_13_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor values_13_end_mask_0 = const()[name = tensor("values_13_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor values_13_squeeze_mask_0 = const()[name = tensor("values_13_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor values_13_cast_fp16 = slice_by_index(begin = values_13_begin_0, end = values_13_end_0, end_mask = values_13_end_mask_0, squeeze_mask = values_13_squeeze_mask_0, x = reshape_29_cast_fp16)[name = tensor("values_13_cast_fp16")]; + tensor var_1286_cast_fp16 = not_equal(x = keys_13_cast_fp16, y = keys_13_cast_fp16)[name = tensor("op_1286_cast_fp16")]; + tensor keys_15_cast_fp16 = select(a = var_470_to_fp16, b = keys_13_cast_fp16, cond = var_1286_cast_fp16)[name = tensor("keys_15_cast_fp16")]; + tensor var_1294_cast_fp16 = not_equal(x = values_13_cast_fp16, y = values_13_cast_fp16)[name = tensor("op_1294_cast_fp16")]; + tensor values_15_cast_fp16 = select(a = var_470_to_fp16, b = values_13_cast_fp16, cond = var_1294_cast_fp16)[name = tensor("values_15_cast_fp16")]; + tensor var_1318 = const()[name = tensor("op_1318"), val = tensor([0, 2, 1, 3])]; + tensor var_1334 = const()[name = tensor("op_1334"), val = tensor([1, 1, 1])]; + tensor var_1335_cast_fp16 = reshape(shape = var_1334, x = position2_to_fp16)[name = tensor("op_1335_cast_fp16")]; + tensor q_positions_5_cast_fp16 = add(x = var_1335_cast_fp16, y = q_offsets_1_promoted_to_fp16)[name = tensor("q_positions_5_cast_fp16")]; + tensor valid_end_5_cast_fp16 = add(x = var_1335_cast_fp16, y = var_536_cast_fp16)[name = tensor("valid_end_5_cast_fp16")]; + tensor valid_mask_5_cast_fp16 = less(x = k_positions_1_promoted_to_fp16, y = valid_end_5_cast_fp16)[name = tensor("valid_mask_5_cast_fp16")]; + tensor causal_mask_5_cast_fp16 = less_equal(x = k_positions_1_promoted_to_fp16, y = q_positions_5_cast_fp16)[name = tensor("causal_mask_5_cast_fp16")]; + tensor var_1363 = logical_and(x = valid_mask_5_cast_fp16, y = causal_mask_5_cast_fp16)[name = tensor("op_1363")]; + tensor attn_mask_5_axes_0 = const()[name = tensor("attn_mask_5_axes_0"), val = tensor([1])]; + tensor attn_mask_5 = expand_dims(axes = attn_mask_5_axes_0, x = var_1363)[name = tensor("attn_mask_5")]; + tensor var_1377_transpose_x_0 = const()[name = tensor("op_1377_transpose_x_0"), val = tensor(false)]; + tensor var_1377_transpose_y_0 = const()[name = tensor("op_1377_transpose_y_0"), val = tensor(false)]; + tensor transpose_73_perm_0 = const()[name = tensor("transpose_73_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_74_perm_0 = const()[name = tensor("transpose_74_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_74 = transpose(perm = transpose_74_perm_0, x = keys_15_cast_fp16)[name = tensor("transpose_196")]; + tensor transpose_73 = transpose(perm = transpose_73_perm_0, x = q_15_cast_fp16)[name = tensor("transpose_197")]; + tensor var_1377_cast_fp16 = matmul(transpose_x = var_1377_transpose_x_0, transpose_y = var_1377_transpose_y_0, x = transpose_73, y = transpose_74)[name = tensor("op_1377_cast_fp16")]; + tensor var_1371_to_fp16 = const()[name = tensor("op_1371_to_fp16"), val = tensor([0x1p-3])]; + tensor attn_weights_13_cast_fp16 = mul(x = var_1377_cast_fp16, y = var_1371_to_fp16)[name = tensor("attn_weights_13_cast_fp16")]; + tensor var_1379 = logical_not(x = attn_mask_5)[name = tensor("op_1379")]; + tensor var_1380_to_fp16 = const()[name = tensor("op_1380_to_fp16"), val = tensor(-inf)]; + tensor attn_weights_15_cast_fp16 = select(a = var_1380_to_fp16, b = attn_weights_13_cast_fp16, cond = var_1379)[name = tensor("attn_weights_15_cast_fp16")]; + tensor var_1382 = const()[name = tensor("op_1382"), val = tensor(-1)]; + tensor attn_weights_17_cast_fp16 = softmax(axis = var_1382, x = attn_weights_15_cast_fp16)[name = tensor("attn_weights_17_cast_fp16")]; + tensor attn_output_5_transpose_x_0 = const()[name = tensor("attn_output_5_transpose_x_0"), val = tensor(false)]; + tensor attn_output_5_transpose_y_0 = const()[name = tensor("attn_output_5_transpose_y_0"), val = tensor(false)]; + tensor values_17_cast_fp16 = transpose(perm = var_1318, x = values_15_cast_fp16)[name = tensor("transpose_198")]; + tensor attn_output_5_cast_fp16 = matmul(transpose_x = attn_output_5_transpose_x_0, transpose_y = attn_output_5_transpose_y_0, x = attn_weights_17_cast_fp16, y = values_17_cast_fp16)[name = tensor("attn_output_5_cast_fp16")]; + tensor var_1390 = const()[name = tensor("op_1390"), val = tensor([0, 2, 1, 3])]; + tensor var_1395 = const()[name = tensor("op_1395"), val = tensor([1, 256, 1024])]; + tensor var_1391_cast_fp16 = transpose(perm = var_1390, x = attn_output_5_cast_fp16)[name = tensor("transpose_195")]; + tensor input_29_cast_fp16 = reshape(shape = var_1395, x = var_1391_cast_fp16)[name = tensor("input_29_cast_fp16")]; + tensor attn2_out_proj_weight_to_fp16 = const()[name = tensor("attn2_out_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(57713280)))]; + tensor linear_9_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = attn2_out_proj_weight_to_fp16, x = input_29_cast_fp16)[name = tensor("linear_9_cast_fp16")]; + tensor var_1404_cast_fp16 = add(x = position2_to_fp16, y = valid_len_to_fp16)[name = tensor("op_1404_cast_fp16")]; + tensor var_1404_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_1404_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor input_31_cast_fp16 = add(x = input_25_cast_fp16, y = linear_9_cast_fp16)[name = tensor("input_31_cast_fp16")]; + tensor input_33_axes_0 = const()[name = tensor("input_33_axes_0"), val = tensor([-1])]; + tensor norm2_2_weight_to_fp16 = const()[name = tensor("norm2_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(59810496)))]; + tensor norm2_2_bias_to_fp16 = const()[name = tensor("norm2_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(59812608)))]; + tensor var_1408_to_fp16 = const()[name = tensor("op_1408_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_33_cast_fp16 = layer_norm(axes = input_33_axes_0, beta = norm2_2_bias_to_fp16, epsilon = var_1408_to_fp16, gamma = norm2_2_weight_to_fp16, x = input_31_cast_fp16)[name = tensor("input_33_cast_fp16")]; + tensor linear2_1_weight_to_fp16 = const()[name = tensor("linear2_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(59814720)))]; + tensor linear_10_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = linear2_1_weight_to_fp16, x = input_33_cast_fp16)[name = tensor("linear_10_cast_fp16")]; + tensor input_35_mode_0 = const()[name = tensor("input_35_mode_0"), val = tensor("EXACT")]; + tensor input_35_cast_fp16 = gelu(mode = input_35_mode_0, x = linear_10_cast_fp16)[name = tensor("input_35_cast_fp16")]; + tensor linear2_2_weight_to_fp16 = const()[name = tensor("linear2_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(68203392)))]; + tensor linear_11_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = linear2_2_weight_to_fp16, x = input_35_cast_fp16)[name = tensor("linear_11_cast_fp16")]; + tensor input_37_cast_fp16 = add(x = input_31_cast_fp16, y = linear_11_cast_fp16)[name = tensor("input_37_cast_fp16")]; + tensor input_39_axes_0 = const()[name = tensor("input_39_axes_0"), val = tensor([-1])]; + tensor norm3_1_weight_to_fp16 = const()[name = tensor("norm3_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(76592064)))]; + tensor norm3_1_bias_to_fp16 = const()[name = tensor("norm3_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(76594176)))]; + tensor var_1425_to_fp16 = const()[name = tensor("op_1425_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_39_cast_fp16 = layer_norm(axes = input_39_axes_0, beta = norm3_1_bias_to_fp16, epsilon = var_1425_to_fp16, gamma = norm3_1_weight_to_fp16, x = input_37_cast_fp16)[name = tensor("input_39_cast_fp16")]; + tensor attn3_in_proj_weight_to_fp16 = const()[name = tensor("attn3_in_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(76596288)))]; + tensor linear_12_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = attn3_in_proj_weight_to_fp16, x = input_39_cast_fp16)[name = tensor("linear_12_cast_fp16")]; + tensor var_1439 = const()[name = tensor("op_1439"), val = tensor([1, 256, 3, 16, 64])]; + tensor qkv_7_cast_fp16 = reshape(shape = var_1439, x = linear_12_cast_fp16)[name = tensor("qkv_7_cast_fp16")]; + tensor q_19_begin_0 = const()[name = tensor("q_19_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor q_19_end_0 = const()[name = tensor("q_19_end_0"), val = tensor([1, 256, 1, 16, 64])]; + tensor q_19_end_mask_0 = const()[name = tensor("q_19_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor q_19_squeeze_mask_0 = const()[name = tensor("q_19_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor q_19_cast_fp16 = slice_by_index(begin = q_19_begin_0, end = q_19_end_0, end_mask = q_19_end_mask_0, squeeze_mask = q_19_squeeze_mask_0, x = qkv_7_cast_fp16)[name = tensor("q_19_cast_fp16")]; + tensor k_13_begin_0 = const()[name = tensor("k_13_begin_0"), val = tensor([0, 0, 1, 0, 0])]; + tensor k_13_end_0 = const()[name = tensor("k_13_end_0"), val = tensor([1, 256, 2, 16, 64])]; + tensor k_13_end_mask_0 = const()[name = tensor("k_13_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor k_13_squeeze_mask_0 = const()[name = tensor("k_13_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor k_13_cast_fp16 = slice_by_index(begin = k_13_begin_0, end = k_13_end_0, end_mask = k_13_end_mask_0, squeeze_mask = k_13_squeeze_mask_0, x = qkv_7_cast_fp16)[name = tensor("k_13_cast_fp16")]; + tensor v_7_begin_0 = const()[name = tensor("v_7_begin_0"), val = tensor([0, 0, 2, 0, 0])]; + tensor v_7_end_0 = const()[name = tensor("v_7_end_0"), val = tensor([1, 256, 3, 16, 64])]; + tensor v_7_end_mask_0 = const()[name = tensor("v_7_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor v_7_squeeze_mask_0 = const()[name = tensor("v_7_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor v_7_cast_fp16 = slice_by_index(begin = v_7_begin_0, end = v_7_end_0, end_mask = v_7_end_mask_0, squeeze_mask = v_7_squeeze_mask_0, x = qkv_7_cast_fp16)[name = tensor("v_7_cast_fp16")]; + tensor var_1497 = const()[name = tensor("op_1497"), val = tensor([1, 1])]; + tensor position3_to_fp16_dtype_0 = const()[name = tensor("position3_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor position3_to_fp16 = cast(dtype = position3_to_fp16_dtype_0, x = position3)[name = tensor("cast_646")]; + tensor var_1498_cast_fp16 = reshape(shape = var_1497, x = position3_to_fp16)[name = tensor("op_1498_cast_fp16")]; + tensor ts_21_cast_fp16 = add(x = ts_1_promoted_to_fp16, y = var_1498_cast_fp16)[name = tensor("ts_21_cast_fp16")]; + tensor var_1505 = const()[name = tensor("op_1505"), val = tensor([1, 256, 1, 1])]; + tensor ts_23_cast_fp16 = reshape(shape = var_1505, x = ts_21_cast_fp16)[name = tensor("ts_23_cast_fp16")]; + tensor var_1512 = const()[name = tensor("op_1512"), val = tensor([1, 256, 16, 32, 2])]; + tensor q_complex_7_cast_fp16 = reshape(shape = var_1512, x = q_19_cast_fp16)[name = tensor("q_complex_7_cast_fp16")]; + tensor var_1519 = const()[name = tensor("op_1519"), val = tensor([1, 256, 16, 32, 2])]; + tensor k_complex_7_cast_fp16 = reshape(shape = var_1519, x = k_13_cast_fp16)[name = tensor("k_complex_7_cast_fp16")]; + tensor var_1523_begin_0 = const()[name = tensor("op_1523_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_1523_end_0 = const()[name = tensor("op_1523_end_0"), val = tensor([1, 256, 16, 32, 1])]; + tensor var_1523_end_mask_0 = const()[name = tensor("op_1523_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_1523_squeeze_mask_0 = const()[name = tensor("op_1523_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_1523_cast_fp16 = slice_by_index(begin = var_1523_begin_0, end = var_1523_end_0, end_mask = var_1523_end_mask_0, squeeze_mask = var_1523_squeeze_mask_0, x = q_complex_7_cast_fp16)[name = tensor("op_1523_cast_fp16")]; + tensor var_1531_begin_0 = const()[name = tensor("op_1531_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_1531_end_0 = const()[name = tensor("op_1531_end_0"), val = tensor([1, 256, 16, 32, 2])]; + tensor var_1531_end_mask_0 = const()[name = tensor("op_1531_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_1531_squeeze_mask_0 = const()[name = tensor("op_1531_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_1531_cast_fp16 = slice_by_index(begin = var_1531_begin_0, end = var_1531_end_0, end_mask = var_1531_end_mask_0, squeeze_mask = var_1531_squeeze_mask_0, x = q_complex_7_cast_fp16)[name = tensor("op_1531_cast_fp16")]; + tensor var_1539_begin_0 = const()[name = tensor("op_1539_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_1539_end_0 = const()[name = tensor("op_1539_end_0"), val = tensor([1, 256, 16, 32, 1])]; + tensor var_1539_end_mask_0 = const()[name = tensor("op_1539_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_1539_squeeze_mask_0 = const()[name = tensor("op_1539_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_1539_cast_fp16 = slice_by_index(begin = var_1539_begin_0, end = var_1539_end_0, end_mask = var_1539_end_mask_0, squeeze_mask = var_1539_squeeze_mask_0, x = k_complex_7_cast_fp16)[name = tensor("op_1539_cast_fp16")]; + tensor var_1547_begin_0 = const()[name = tensor("op_1547_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_1547_end_0 = const()[name = tensor("op_1547_end_0"), val = tensor([1, 256, 16, 32, 2])]; + tensor var_1547_end_mask_0 = const()[name = tensor("op_1547_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_1547_squeeze_mask_0 = const()[name = tensor("op_1547_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_1547_cast_fp16 = slice_by_index(begin = var_1547_begin_0, end = var_1547_end_0, end_mask = var_1547_end_mask_0, squeeze_mask = var_1547_squeeze_mask_0, x = k_complex_7_cast_fp16)[name = tensor("op_1547_cast_fp16")]; + tensor freqs_7_to_fp16 = const()[name = tensor("freqs_7_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(82887808)))]; + tensor var_1553_cast_fp16 = mul(x = freqs_7_to_fp16, y = ts_23_cast_fp16)[name = tensor("op_1553_cast_fp16")]; + tensor rotr_7_cast_fp16 = cos(x = var_1553_cast_fp16)[name = tensor("rotr_7_cast_fp16")]; + tensor roti_7_cast_fp16 = sin(x = var_1553_cast_fp16)[name = tensor("roti_7_cast_fp16")]; + tensor var_1557_cast_fp16 = mul(x = var_1523_cast_fp16, y = rotr_7_cast_fp16)[name = tensor("op_1557_cast_fp16")]; + tensor var_1558_cast_fp16 = mul(x = var_1531_cast_fp16, y = roti_7_cast_fp16)[name = tensor("op_1558_cast_fp16")]; + tensor qor_13_cast_fp16 = sub(x = var_1557_cast_fp16, y = var_1558_cast_fp16)[name = tensor("qor_13_cast_fp16")]; + tensor var_1561_cast_fp16 = mul(x = var_1523_cast_fp16, y = roti_7_cast_fp16)[name = tensor("op_1561_cast_fp16")]; + tensor var_1562_cast_fp16 = mul(x = var_1531_cast_fp16, y = rotr_7_cast_fp16)[name = tensor("op_1562_cast_fp16")]; + tensor qoi_13_cast_fp16 = add(x = var_1561_cast_fp16, y = var_1562_cast_fp16)[name = tensor("qoi_13_cast_fp16")]; + tensor var_1565_cast_fp16 = mul(x = var_1539_cast_fp16, y = rotr_7_cast_fp16)[name = tensor("op_1565_cast_fp16")]; + tensor var_1566_cast_fp16 = mul(x = var_1547_cast_fp16, y = roti_7_cast_fp16)[name = tensor("op_1566_cast_fp16")]; + tensor kor_13_cast_fp16 = sub(x = var_1565_cast_fp16, y = var_1566_cast_fp16)[name = tensor("kor_13_cast_fp16")]; + tensor var_1569_cast_fp16 = mul(x = var_1539_cast_fp16, y = roti_7_cast_fp16)[name = tensor("op_1569_cast_fp16")]; + tensor var_1570_cast_fp16 = mul(x = var_1547_cast_fp16, y = rotr_7_cast_fp16)[name = tensor("op_1570_cast_fp16")]; + tensor koi_13_cast_fp16 = add(x = var_1569_cast_fp16, y = var_1570_cast_fp16)[name = tensor("koi_13_cast_fp16")]; + tensor qo_7_axis_0 = const()[name = tensor("qo_7_axis_0"), val = tensor(-1)]; + tensor qo_7_cast_fp16 = stack(axis = qo_7_axis_0, values = (qor_13_cast_fp16, qoi_13_cast_fp16))[name = tensor("qo_7_cast_fp16")]; + tensor ko_7_axis_0 = const()[name = tensor("ko_7_axis_0"), val = tensor(-1)]; + tensor ko_7_cast_fp16 = stack(axis = ko_7_axis_0, values = (kor_13_cast_fp16, koi_13_cast_fp16))[name = tensor("ko_7_cast_fp16")]; + tensor var_1603 = const()[name = tensor("op_1603"), val = tensor([1, 256, 16, 64])]; + tensor q_21_cast_fp16 = reshape(shape = var_1603, x = qo_7_cast_fp16)[name = tensor("q_21_cast_fp16")]; + tensor var_1609 = const()[name = tensor("op_1609"), val = tensor([1, 256, 16, 64])]; + tensor k_15_cast_fp16 = reshape(shape = var_1609, x = ko_7_cast_fp16)[name = tensor("k_15_cast_fp16")]; + tensor write_indices_float_19_cast_fp16 = add(x = var_1498_cast_fp16, y = var_393_promoted_to_fp16)[name = tensor("write_indices_float_19_cast_fp16")]; + tensor _inversed_1630_y_0_to_fp16 = const()[name = tensor("_inversed_1630_y_0_to_fp16"), val = tensor(0x1p-9)]; + tensor _inversed_1630_cast_fp16 = mul(x = write_indices_float_19_cast_fp16, y = _inversed_1630_y_0_to_fp16)[name = tensor("_inversed_1630_cast_fp16")]; + tensor var_1631_cast_fp16 = floor(x = _inversed_1630_cast_fp16)[name = tensor("op_1631_cast_fp16")]; + tensor var_1632_to_fp16 = const()[name = tensor("op_1632_to_fp16"), val = tensor(0x1p+9)]; + tensor var_1633_cast_fp16 = mul(x = var_1631_cast_fp16, y = var_1632_to_fp16)[name = tensor("op_1633_cast_fp16")]; + tensor write_indices_float_21_cast_fp16 = sub(x = write_indices_float_19_cast_fp16, y = var_1633_cast_fp16)[name = tensor("write_indices_float_21_cast_fp16")]; + tensor var_1651_value_0_to_fp16 = const()[name = tensor("op_1651_value_0_to_fp16"), val = tensor(0x1.ffp+8)]; + tensor var_1651_cast_fp16 = fill_like(ref_tensor = write_indices_float_21_cast_fp16, value = var_1651_value_0_to_fp16)[name = tensor("op_1651_cast_fp16")]; + tensor write_indices_float_23_cast_fp16 = select(a = write_indices_float_21_cast_fp16, b = var_1651_cast_fp16, cond = token_valid_1_cast_fp16)[name = tensor("write_indices_float_23_cast_fp16")]; + tensor var_1657_dtype_0 = const()[name = tensor("op_1657_dtype_0"), val = tensor("int32")]; + tensor var_1662 = const()[name = tensor("op_1662"), val = tensor([1, 256, 1, 1])]; + tensor write_indices_float_23_cast_fp16_to_int32 = cast(dtype = var_1657_dtype_0, x = write_indices_float_23_cast_fp16)[name = tensor("cast_645")]; + tensor var_1663 = reshape(shape = var_1662, x = write_indices_float_23_cast_fp16_to_int32)[name = tensor("op_1663")]; + tensor write_indices_7_reps_0 = const()[name = tensor("write_indices_7_reps_0"), val = tensor([1, 1, 16, 64])]; + tensor write_indices_7 = tile(reps = write_indices_7_reps_0, x = var_1663)[name = tensor("write_indices_7")]; + tensor var_1673_begin_0 = const()[name = tensor("op_1673_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_1673_end_0 = const()[name = tensor("op_1673_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor var_1673_end_mask_0 = const()[name = tensor("op_1673_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_1673_squeeze_mask_0 = const()[name = tensor("op_1673_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor cache3_to_fp16_dtype_0 = const()[name = tensor("cache3_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor cache3_to_fp16 = cast(dtype = cache3_to_fp16_dtype_0, x = cache3)[name = tensor("cast_644")]; + tensor var_1673_cast_fp16 = slice_by_index(begin = var_1673_begin_0, end = var_1673_end_0, end_mask = var_1673_end_mask_0, squeeze_mask = var_1673_squeeze_mask_0, x = cache3_to_fp16)[name = tensor("op_1673_cast_fp16")]; + tensor var_1675_axis_0 = const()[name = tensor("op_1675_axis_0"), val = tensor(1)]; + tensor var_1675_mode_0 = const()[name = tensor("op_1675_mode_0"), val = tensor("update")]; + tensor var_1675_validate_indices_0 = const()[name = tensor("op_1675_validate_indices_0"), val = tensor(false)]; + tensor var_1675_cast_fp16 = scatter_along_axis(axis = var_1675_axis_0, data = var_1673_cast_fp16, indices = write_indices_7, mode = var_1675_mode_0, updates = k_15_cast_fp16, validate_indices = var_1675_validate_indices_0)[name = tensor("op_1675_cast_fp16")]; + tensor concat_26 = const()[name = tensor("concat_26"), val = tensor([0, 0, 0, 0, 0])]; + tensor concat_27 = const()[name = tensor("concat_27"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_7_internal_tensor_assign_1_stride_0 = const()[name = tensor("new_cache_7_internal_tensor_assign_1_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_7_internal_tensor_assign_1_begin_mask_0 = const()[name = tensor("new_cache_7_internal_tensor_assign_1_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_7_internal_tensor_assign_1_end_mask_0 = const()[name = tensor("new_cache_7_internal_tensor_assign_1_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_7_internal_tensor_assign_1_squeeze_mask_0 = const()[name = tensor("new_cache_7_internal_tensor_assign_1_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_52 = const()[name = tensor("shape_52"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_6 = const()[name = tensor("reduce_prod_6"), val = tensor(1048576)]; + tensor range_1d_6_start_0 = const()[name = tensor("range_1d_6_start_0"), val = tensor(0)]; + tensor range_1d_6_step_0 = const()[name = tensor("range_1d_6_step_0"), val = tensor(1)]; + tensor range_1d_6 = range_1d(end = reduce_prod_6, start = range_1d_6_start_0, step = range_1d_6_step_0)[name = tensor("range_1d_6")]; + tensor reshape_30 = reshape(shape = shape_52, x = range_1d_6)[name = tensor("reshape_30")]; + tensor slice_by_index_6 = slice_by_index(begin = concat_26, begin_mask = new_cache_7_internal_tensor_assign_1_begin_mask_0, end = concat_27, end_mask = new_cache_7_internal_tensor_assign_1_end_mask_0, squeeze_mask = new_cache_7_internal_tensor_assign_1_squeeze_mask_0, stride = new_cache_7_internal_tensor_assign_1_stride_0, x = reshape_30)[name = tensor("slice_by_index_6")]; + tensor reshape_31_shape_0 = const()[name = tensor("reshape_31_shape_0"), val = tensor([-1])]; + tensor reshape_31 = reshape(shape = reshape_31_shape_0, x = slice_by_index_6)[name = tensor("reshape_31")]; + tensor reshape_32_shape_0 = const()[name = tensor("reshape_32_shape_0"), val = tensor([-1])]; + tensor reshape_32_cast_fp16 = reshape(shape = reshape_32_shape_0, x = var_1675_cast_fp16)[name = tensor("reshape_32_cast_fp16")]; + tensor reshape_33_shape_0 = const()[name = tensor("reshape_33_shape_0"), val = tensor([-1])]; + tensor reshape_33_cast_fp16 = reshape(shape = reshape_33_shape_0, x = cache3_to_fp16)[name = tensor("reshape_33_cast_fp16")]; + tensor scatter_6_mode_0 = const()[name = tensor("scatter_6_mode_0"), val = tensor("update")]; + tensor scatter_6_axis_0 = const()[name = tensor("scatter_6_axis_0"), val = tensor(0)]; + tensor scatter_6_validate_indices_0 = const()[name = tensor("scatter_6_validate_indices_0"), val = tensor(false)]; + tensor scatter_6_cast_fp16 = scatter(axis = scatter_6_axis_0, data = reshape_33_cast_fp16, indices = reshape_31, mode = scatter_6_mode_0, updates = reshape_32_cast_fp16, validate_indices = scatter_6_validate_indices_0)[name = tensor("scatter_6_cast_fp16")]; + tensor reshape_34_cast_fp16 = reshape(shape = shape_52, x = scatter_6_cast_fp16)[name = tensor("reshape_34_cast_fp16")]; + tensor var_1683_begin_0 = const()[name = tensor("op_1683_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor var_1683_end_0 = const()[name = tensor("op_1683_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor var_1683_end_mask_0 = const()[name = tensor("op_1683_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_1683_squeeze_mask_0 = const()[name = tensor("op_1683_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor var_1683_cast_fp16 = slice_by_index(begin = var_1683_begin_0, end = var_1683_end_0, end_mask = var_1683_end_mask_0, squeeze_mask = var_1683_squeeze_mask_0, x = reshape_34_cast_fp16)[name = tensor("op_1683_cast_fp16")]; + tensor var_1685_axis_0 = const()[name = tensor("op_1685_axis_0"), val = tensor(1)]; + tensor var_1685_mode_0 = const()[name = tensor("op_1685_mode_0"), val = tensor("update")]; + tensor var_1685_validate_indices_0 = const()[name = tensor("op_1685_validate_indices_0"), val = tensor(false)]; + tensor var_1685_cast_fp16 = scatter_along_axis(axis = var_1685_axis_0, data = var_1683_cast_fp16, indices = write_indices_7, mode = var_1685_mode_0, updates = v_7_cast_fp16, validate_indices = var_1685_validate_indices_0)[name = tensor("op_1685_cast_fp16")]; + tensor concat_28 = const()[name = tensor("concat_28"), val = tensor([1, 0, 0, 0, 0])]; + tensor concat_29 = const()[name = tensor("concat_29"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_7_internal_tensor_assign_2_stride_0 = const()[name = tensor("new_cache_7_internal_tensor_assign_2_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_7_internal_tensor_assign_2_begin_mask_0 = const()[name = tensor("new_cache_7_internal_tensor_assign_2_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_7_internal_tensor_assign_2_end_mask_0 = const()[name = tensor("new_cache_7_internal_tensor_assign_2_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_7_internal_tensor_assign_2_squeeze_mask_0 = const()[name = tensor("new_cache_7_internal_tensor_assign_2_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_53 = const()[name = tensor("shape_53"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_7 = const()[name = tensor("reduce_prod_7"), val = tensor(1048576)]; + tensor range_1d_7_start_0 = const()[name = tensor("range_1d_7_start_0"), val = tensor(0)]; + tensor range_1d_7_step_0 = const()[name = tensor("range_1d_7_step_0"), val = tensor(1)]; + tensor range_1d_7 = range_1d(end = reduce_prod_7, start = range_1d_7_start_0, step = range_1d_7_step_0)[name = tensor("range_1d_7")]; + tensor reshape_35 = reshape(shape = shape_53, x = range_1d_7)[name = tensor("reshape_35")]; + tensor slice_by_index_7 = slice_by_index(begin = concat_28, begin_mask = new_cache_7_internal_tensor_assign_2_begin_mask_0, end = concat_29, end_mask = new_cache_7_internal_tensor_assign_2_end_mask_0, squeeze_mask = new_cache_7_internal_tensor_assign_2_squeeze_mask_0, stride = new_cache_7_internal_tensor_assign_2_stride_0, x = reshape_35)[name = tensor("slice_by_index_7")]; + tensor reshape_36_shape_0 = const()[name = tensor("reshape_36_shape_0"), val = tensor([-1])]; + tensor reshape_36 = reshape(shape = reshape_36_shape_0, x = slice_by_index_7)[name = tensor("reshape_36")]; + tensor reshape_37_shape_0 = const()[name = tensor("reshape_37_shape_0"), val = tensor([-1])]; + tensor reshape_37_cast_fp16 = reshape(shape = reshape_37_shape_0, x = var_1685_cast_fp16)[name = tensor("reshape_37_cast_fp16")]; + tensor reshape_38_shape_0 = const()[name = tensor("reshape_38_shape_0"), val = tensor([-1])]; + tensor reshape_38_cast_fp16 = reshape(shape = reshape_38_shape_0, x = reshape_34_cast_fp16)[name = tensor("reshape_38_cast_fp16")]; + tensor scatter_7_mode_0 = const()[name = tensor("scatter_7_mode_0"), val = tensor("update")]; + tensor scatter_7_axis_0 = const()[name = tensor("scatter_7_axis_0"), val = tensor(0)]; + tensor scatter_7_validate_indices_0 = const()[name = tensor("scatter_7_validate_indices_0"), val = tensor(false)]; + tensor scatter_7_cast_fp16 = scatter(axis = scatter_7_axis_0, data = reshape_38_cast_fp16, indices = reshape_36, mode = scatter_7_mode_0, updates = reshape_37_cast_fp16, validate_indices = scatter_7_validate_indices_0)[name = tensor("scatter_7_cast_fp16")]; + tensor reshape_39_cast_fp16 = reshape(shape = shape_53, x = scatter_7_cast_fp16)[name = tensor("reshape_39_cast_fp16")]; + tensor reshape_39_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("reshape_39_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor keys_19_begin_0 = const()[name = tensor("keys_19_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor keys_19_end_0 = const()[name = tensor("keys_19_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor keys_19_end_mask_0 = const()[name = tensor("keys_19_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor keys_19_squeeze_mask_0 = const()[name = tensor("keys_19_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor keys_19_cast_fp16 = slice_by_index(begin = keys_19_begin_0, end = keys_19_end_0, end_mask = keys_19_end_mask_0, squeeze_mask = keys_19_squeeze_mask_0, x = reshape_39_cast_fp16)[name = tensor("keys_19_cast_fp16")]; + tensor values_19_begin_0 = const()[name = tensor("values_19_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor values_19_end_0 = const()[name = tensor("values_19_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor values_19_end_mask_0 = const()[name = tensor("values_19_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor values_19_squeeze_mask_0 = const()[name = tensor("values_19_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor values_19_cast_fp16 = slice_by_index(begin = values_19_begin_0, end = values_19_end_0, end_mask = values_19_end_mask_0, squeeze_mask = values_19_squeeze_mask_0, x = reshape_39_cast_fp16)[name = tensor("values_19_cast_fp16")]; + tensor var_1697_cast_fp16 = not_equal(x = keys_19_cast_fp16, y = keys_19_cast_fp16)[name = tensor("op_1697_cast_fp16")]; + tensor keys_21_cast_fp16 = select(a = var_470_to_fp16, b = keys_19_cast_fp16, cond = var_1697_cast_fp16)[name = tensor("keys_21_cast_fp16")]; + tensor var_1705_cast_fp16 = not_equal(x = values_19_cast_fp16, y = values_19_cast_fp16)[name = tensor("op_1705_cast_fp16")]; + tensor values_21_cast_fp16 = select(a = var_470_to_fp16, b = values_19_cast_fp16, cond = var_1705_cast_fp16)[name = tensor("values_21_cast_fp16")]; + tensor var_1729 = const()[name = tensor("op_1729"), val = tensor([0, 2, 1, 3])]; + tensor var_1745 = const()[name = tensor("op_1745"), val = tensor([1, 1, 1])]; + tensor var_1746_cast_fp16 = reshape(shape = var_1745, x = position3_to_fp16)[name = tensor("op_1746_cast_fp16")]; + tensor q_positions_7_cast_fp16 = add(x = var_1746_cast_fp16, y = q_offsets_1_promoted_to_fp16)[name = tensor("q_positions_7_cast_fp16")]; + tensor valid_end_7_cast_fp16 = add(x = var_1746_cast_fp16, y = var_536_cast_fp16)[name = tensor("valid_end_7_cast_fp16")]; + tensor valid_mask_7_cast_fp16 = less(x = k_positions_1_promoted_to_fp16, y = valid_end_7_cast_fp16)[name = tensor("valid_mask_7_cast_fp16")]; + tensor causal_mask_7_cast_fp16 = less_equal(x = k_positions_1_promoted_to_fp16, y = q_positions_7_cast_fp16)[name = tensor("causal_mask_7_cast_fp16")]; + tensor var_1774 = logical_and(x = valid_mask_7_cast_fp16, y = causal_mask_7_cast_fp16)[name = tensor("op_1774")]; + tensor attn_mask_7_axes_0 = const()[name = tensor("attn_mask_7_axes_0"), val = tensor([1])]; + tensor attn_mask_7 = expand_dims(axes = attn_mask_7_axes_0, x = var_1774)[name = tensor("attn_mask_7")]; + tensor var_1788_transpose_x_0 = const()[name = tensor("op_1788_transpose_x_0"), val = tensor(false)]; + tensor var_1788_transpose_y_0 = const()[name = tensor("op_1788_transpose_y_0"), val = tensor(false)]; + tensor transpose_75_perm_0 = const()[name = tensor("transpose_75_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_76_perm_0 = const()[name = tensor("transpose_76_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_76 = transpose(perm = transpose_76_perm_0, x = keys_21_cast_fp16)[name = tensor("transpose_192")]; + tensor transpose_75 = transpose(perm = transpose_75_perm_0, x = q_21_cast_fp16)[name = tensor("transpose_193")]; + tensor var_1788_cast_fp16 = matmul(transpose_x = var_1788_transpose_x_0, transpose_y = var_1788_transpose_y_0, x = transpose_75, y = transpose_76)[name = tensor("op_1788_cast_fp16")]; + tensor var_1782_to_fp16 = const()[name = tensor("op_1782_to_fp16"), val = tensor([0x1p-3])]; + tensor attn_weights_19_cast_fp16 = mul(x = var_1788_cast_fp16, y = var_1782_to_fp16)[name = tensor("attn_weights_19_cast_fp16")]; + tensor var_1790 = logical_not(x = attn_mask_7)[name = tensor("op_1790")]; + tensor var_1791_to_fp16 = const()[name = tensor("op_1791_to_fp16"), val = tensor(-inf)]; + tensor attn_weights_21_cast_fp16 = select(a = var_1791_to_fp16, b = attn_weights_19_cast_fp16, cond = var_1790)[name = tensor("attn_weights_21_cast_fp16")]; + tensor var_1793 = const()[name = tensor("op_1793"), val = tensor(-1)]; + tensor attn_weights_23_cast_fp16 = softmax(axis = var_1793, x = attn_weights_21_cast_fp16)[name = tensor("attn_weights_23_cast_fp16")]; + tensor attn_output_7_transpose_x_0 = const()[name = tensor("attn_output_7_transpose_x_0"), val = tensor(false)]; + tensor attn_output_7_transpose_y_0 = const()[name = tensor("attn_output_7_transpose_y_0"), val = tensor(false)]; + tensor values_23_cast_fp16 = transpose(perm = var_1729, x = values_21_cast_fp16)[name = tensor("transpose_194")]; + tensor attn_output_7_cast_fp16 = matmul(transpose_x = attn_output_7_transpose_x_0, transpose_y = attn_output_7_transpose_y_0, x = attn_weights_23_cast_fp16, y = values_23_cast_fp16)[name = tensor("attn_output_7_cast_fp16")]; + tensor var_1801 = const()[name = tensor("op_1801"), val = tensor([0, 2, 1, 3])]; + tensor var_1806 = const()[name = tensor("op_1806"), val = tensor([1, 256, 1024])]; + tensor var_1802_cast_fp16 = transpose(perm = var_1801, x = attn_output_7_cast_fp16)[name = tensor("transpose_191")]; + tensor input_41_cast_fp16 = reshape(shape = var_1806, x = var_1802_cast_fp16)[name = tensor("input_41_cast_fp16")]; + tensor attn3_out_proj_weight_to_fp16 = const()[name = tensor("attn3_out_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(82887936)))]; + tensor linear_13_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = attn3_out_proj_weight_to_fp16, x = input_41_cast_fp16)[name = tensor("linear_13_cast_fp16")]; + tensor var_1815_cast_fp16 = add(x = position3_to_fp16, y = valid_len_to_fp16)[name = tensor("op_1815_cast_fp16")]; + tensor var_1815_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_1815_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor input_43_cast_fp16 = add(x = input_37_cast_fp16, y = linear_13_cast_fp16)[name = tensor("input_43_cast_fp16")]; + tensor input_45_axes_0 = const()[name = tensor("input_45_axes_0"), val = tensor([-1])]; + tensor norm3_2_weight_to_fp16 = const()[name = tensor("norm3_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(84985152)))]; + tensor norm3_2_bias_to_fp16 = const()[name = tensor("norm3_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(84987264)))]; + tensor var_1819_to_fp16 = const()[name = tensor("op_1819_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_45_cast_fp16 = layer_norm(axes = input_45_axes_0, beta = norm3_2_bias_to_fp16, epsilon = var_1819_to_fp16, gamma = norm3_2_weight_to_fp16, x = input_43_cast_fp16)[name = tensor("input_45_cast_fp16")]; + tensor linear3_1_weight_to_fp16 = const()[name = tensor("linear3_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(84989376)))]; + tensor linear_14_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = linear3_1_weight_to_fp16, x = input_45_cast_fp16)[name = tensor("linear_14_cast_fp16")]; + tensor input_47_mode_0 = const()[name = tensor("input_47_mode_0"), val = tensor("EXACT")]; + tensor input_47_cast_fp16 = gelu(mode = input_47_mode_0, x = linear_14_cast_fp16)[name = tensor("input_47_cast_fp16")]; + tensor linear3_2_weight_to_fp16 = const()[name = tensor("linear3_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(93378048)))]; + tensor linear_15_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = linear3_2_weight_to_fp16, x = input_47_cast_fp16)[name = tensor("linear_15_cast_fp16")]; + tensor input_49_cast_fp16 = add(x = input_43_cast_fp16, y = linear_15_cast_fp16)[name = tensor("input_49_cast_fp16")]; + tensor input_51_axes_0 = const()[name = tensor("input_51_axes_0"), val = tensor([-1])]; + tensor norm4_1_weight_to_fp16 = const()[name = tensor("norm4_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(101766720)))]; + tensor norm4_1_bias_to_fp16 = const()[name = tensor("norm4_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(101768832)))]; + tensor var_1836_to_fp16 = const()[name = tensor("op_1836_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_51_cast_fp16 = layer_norm(axes = input_51_axes_0, beta = norm4_1_bias_to_fp16, epsilon = var_1836_to_fp16, gamma = norm4_1_weight_to_fp16, x = input_49_cast_fp16)[name = tensor("input_51_cast_fp16")]; + tensor attn4_in_proj_weight_to_fp16 = const()[name = tensor("attn4_in_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(101770944)))]; + tensor linear_16_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = attn4_in_proj_weight_to_fp16, x = input_51_cast_fp16)[name = tensor("linear_16_cast_fp16")]; + tensor var_1850 = const()[name = tensor("op_1850"), val = tensor([1, 256, 3, 16, 64])]; + tensor qkv_9_cast_fp16 = reshape(shape = var_1850, x = linear_16_cast_fp16)[name = tensor("qkv_9_cast_fp16")]; + tensor q_25_begin_0 = const()[name = tensor("q_25_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor q_25_end_0 = const()[name = tensor("q_25_end_0"), val = tensor([1, 256, 1, 16, 64])]; + tensor q_25_end_mask_0 = const()[name = tensor("q_25_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor q_25_squeeze_mask_0 = const()[name = tensor("q_25_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor q_25_cast_fp16 = slice_by_index(begin = q_25_begin_0, end = q_25_end_0, end_mask = q_25_end_mask_0, squeeze_mask = q_25_squeeze_mask_0, x = qkv_9_cast_fp16)[name = tensor("q_25_cast_fp16")]; + tensor k_17_begin_0 = const()[name = tensor("k_17_begin_0"), val = tensor([0, 0, 1, 0, 0])]; + tensor k_17_end_0 = const()[name = tensor("k_17_end_0"), val = tensor([1, 256, 2, 16, 64])]; + tensor k_17_end_mask_0 = const()[name = tensor("k_17_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor k_17_squeeze_mask_0 = const()[name = tensor("k_17_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor k_17_cast_fp16 = slice_by_index(begin = k_17_begin_0, end = k_17_end_0, end_mask = k_17_end_mask_0, squeeze_mask = k_17_squeeze_mask_0, x = qkv_9_cast_fp16)[name = tensor("k_17_cast_fp16")]; + tensor v_9_begin_0 = const()[name = tensor("v_9_begin_0"), val = tensor([0, 0, 2, 0, 0])]; + tensor v_9_end_0 = const()[name = tensor("v_9_end_0"), val = tensor([1, 256, 3, 16, 64])]; + tensor v_9_end_mask_0 = const()[name = tensor("v_9_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor v_9_squeeze_mask_0 = const()[name = tensor("v_9_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor v_9_cast_fp16 = slice_by_index(begin = v_9_begin_0, end = v_9_end_0, end_mask = v_9_end_mask_0, squeeze_mask = v_9_squeeze_mask_0, x = qkv_9_cast_fp16)[name = tensor("v_9_cast_fp16")]; + tensor var_1908 = const()[name = tensor("op_1908"), val = tensor([1, 1])]; + tensor position4_to_fp16_dtype_0 = const()[name = tensor("position4_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor position4_to_fp16 = cast(dtype = position4_to_fp16_dtype_0, x = position4)[name = tensor("cast_641")]; + tensor var_1909_cast_fp16 = reshape(shape = var_1908, x = position4_to_fp16)[name = tensor("op_1909_cast_fp16")]; + tensor ts_27_cast_fp16 = add(x = ts_1_promoted_to_fp16, y = var_1909_cast_fp16)[name = tensor("ts_27_cast_fp16")]; + tensor var_1916 = const()[name = tensor("op_1916"), val = tensor([1, 256, 1, 1])]; + tensor ts_29_cast_fp16 = reshape(shape = var_1916, x = ts_27_cast_fp16)[name = tensor("ts_29_cast_fp16")]; + tensor var_1923 = const()[name = tensor("op_1923"), val = tensor([1, 256, 16, 32, 2])]; + tensor q_complex_9_cast_fp16 = reshape(shape = var_1923, x = q_25_cast_fp16)[name = tensor("q_complex_9_cast_fp16")]; + tensor var_1930 = const()[name = tensor("op_1930"), val = tensor([1, 256, 16, 32, 2])]; + tensor k_complex_9_cast_fp16 = reshape(shape = var_1930, x = k_17_cast_fp16)[name = tensor("k_complex_9_cast_fp16")]; + tensor var_1934_begin_0 = const()[name = tensor("op_1934_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_1934_end_0 = const()[name = tensor("op_1934_end_0"), val = tensor([1, 256, 16, 32, 1])]; + tensor var_1934_end_mask_0 = const()[name = tensor("op_1934_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_1934_squeeze_mask_0 = const()[name = tensor("op_1934_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_1934_cast_fp16 = slice_by_index(begin = var_1934_begin_0, end = var_1934_end_0, end_mask = var_1934_end_mask_0, squeeze_mask = var_1934_squeeze_mask_0, x = q_complex_9_cast_fp16)[name = tensor("op_1934_cast_fp16")]; + tensor var_1942_begin_0 = const()[name = tensor("op_1942_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_1942_end_0 = const()[name = tensor("op_1942_end_0"), val = tensor([1, 256, 16, 32, 2])]; + tensor var_1942_end_mask_0 = const()[name = tensor("op_1942_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_1942_squeeze_mask_0 = const()[name = tensor("op_1942_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_1942_cast_fp16 = slice_by_index(begin = var_1942_begin_0, end = var_1942_end_0, end_mask = var_1942_end_mask_0, squeeze_mask = var_1942_squeeze_mask_0, x = q_complex_9_cast_fp16)[name = tensor("op_1942_cast_fp16")]; + tensor var_1950_begin_0 = const()[name = tensor("op_1950_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_1950_end_0 = const()[name = tensor("op_1950_end_0"), val = tensor([1, 256, 16, 32, 1])]; + tensor var_1950_end_mask_0 = const()[name = tensor("op_1950_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_1950_squeeze_mask_0 = const()[name = tensor("op_1950_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_1950_cast_fp16 = slice_by_index(begin = var_1950_begin_0, end = var_1950_end_0, end_mask = var_1950_end_mask_0, squeeze_mask = var_1950_squeeze_mask_0, x = k_complex_9_cast_fp16)[name = tensor("op_1950_cast_fp16")]; + tensor var_1958_begin_0 = const()[name = tensor("op_1958_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_1958_end_0 = const()[name = tensor("op_1958_end_0"), val = tensor([1, 256, 16, 32, 2])]; + tensor var_1958_end_mask_0 = const()[name = tensor("op_1958_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_1958_squeeze_mask_0 = const()[name = tensor("op_1958_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_1958_cast_fp16 = slice_by_index(begin = var_1958_begin_0, end = var_1958_end_0, end_mask = var_1958_end_mask_0, squeeze_mask = var_1958_squeeze_mask_0, x = k_complex_9_cast_fp16)[name = tensor("op_1958_cast_fp16")]; + tensor freqs_9_to_fp16 = const()[name = tensor("freqs_9_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(108062464)))]; + tensor var_1964_cast_fp16 = mul(x = freqs_9_to_fp16, y = ts_29_cast_fp16)[name = tensor("op_1964_cast_fp16")]; + tensor rotr_9_cast_fp16 = cos(x = var_1964_cast_fp16)[name = tensor("rotr_9_cast_fp16")]; + tensor roti_9_cast_fp16 = sin(x = var_1964_cast_fp16)[name = tensor("roti_9_cast_fp16")]; + tensor var_1968_cast_fp16 = mul(x = var_1934_cast_fp16, y = rotr_9_cast_fp16)[name = tensor("op_1968_cast_fp16")]; + tensor var_1969_cast_fp16 = mul(x = var_1942_cast_fp16, y = roti_9_cast_fp16)[name = tensor("op_1969_cast_fp16")]; + tensor qor_17_cast_fp16 = sub(x = var_1968_cast_fp16, y = var_1969_cast_fp16)[name = tensor("qor_17_cast_fp16")]; + tensor var_1972_cast_fp16 = mul(x = var_1934_cast_fp16, y = roti_9_cast_fp16)[name = tensor("op_1972_cast_fp16")]; + tensor var_1973_cast_fp16 = mul(x = var_1942_cast_fp16, y = rotr_9_cast_fp16)[name = tensor("op_1973_cast_fp16")]; + tensor qoi_17_cast_fp16 = add(x = var_1972_cast_fp16, y = var_1973_cast_fp16)[name = tensor("qoi_17_cast_fp16")]; + tensor var_1976_cast_fp16 = mul(x = var_1950_cast_fp16, y = rotr_9_cast_fp16)[name = tensor("op_1976_cast_fp16")]; + tensor var_1977_cast_fp16 = mul(x = var_1958_cast_fp16, y = roti_9_cast_fp16)[name = tensor("op_1977_cast_fp16")]; + tensor kor_17_cast_fp16 = sub(x = var_1976_cast_fp16, y = var_1977_cast_fp16)[name = tensor("kor_17_cast_fp16")]; + tensor var_1980_cast_fp16 = mul(x = var_1950_cast_fp16, y = roti_9_cast_fp16)[name = tensor("op_1980_cast_fp16")]; + tensor var_1981_cast_fp16 = mul(x = var_1958_cast_fp16, y = rotr_9_cast_fp16)[name = tensor("op_1981_cast_fp16")]; + tensor koi_17_cast_fp16 = add(x = var_1980_cast_fp16, y = var_1981_cast_fp16)[name = tensor("koi_17_cast_fp16")]; + tensor qo_9_axis_0 = const()[name = tensor("qo_9_axis_0"), val = tensor(-1)]; + tensor qo_9_cast_fp16 = stack(axis = qo_9_axis_0, values = (qor_17_cast_fp16, qoi_17_cast_fp16))[name = tensor("qo_9_cast_fp16")]; + tensor ko_9_axis_0 = const()[name = tensor("ko_9_axis_0"), val = tensor(-1)]; + tensor ko_9_cast_fp16 = stack(axis = ko_9_axis_0, values = (kor_17_cast_fp16, koi_17_cast_fp16))[name = tensor("ko_9_cast_fp16")]; + tensor var_2014 = const()[name = tensor("op_2014"), val = tensor([1, 256, 16, 64])]; + tensor q_27_cast_fp16 = reshape(shape = var_2014, x = qo_9_cast_fp16)[name = tensor("q_27_cast_fp16")]; + tensor var_2020 = const()[name = tensor("op_2020"), val = tensor([1, 256, 16, 64])]; + tensor k_19_cast_fp16 = reshape(shape = var_2020, x = ko_9_cast_fp16)[name = tensor("k_19_cast_fp16")]; + tensor write_indices_float_25_cast_fp16 = add(x = var_1909_cast_fp16, y = var_393_promoted_to_fp16)[name = tensor("write_indices_float_25_cast_fp16")]; + tensor _inversed_2041_y_0_to_fp16 = const()[name = tensor("_inversed_2041_y_0_to_fp16"), val = tensor(0x1p-9)]; + tensor _inversed_2041_cast_fp16 = mul(x = write_indices_float_25_cast_fp16, y = _inversed_2041_y_0_to_fp16)[name = tensor("_inversed_2041_cast_fp16")]; + tensor var_2042_cast_fp16 = floor(x = _inversed_2041_cast_fp16)[name = tensor("op_2042_cast_fp16")]; + tensor var_2043_to_fp16 = const()[name = tensor("op_2043_to_fp16"), val = tensor(0x1p+9)]; + tensor var_2044_cast_fp16 = mul(x = var_2042_cast_fp16, y = var_2043_to_fp16)[name = tensor("op_2044_cast_fp16")]; + tensor write_indices_float_27_cast_fp16 = sub(x = write_indices_float_25_cast_fp16, y = var_2044_cast_fp16)[name = tensor("write_indices_float_27_cast_fp16")]; + tensor var_2062_value_0_to_fp16 = const()[name = tensor("op_2062_value_0_to_fp16"), val = tensor(0x1.ffp+8)]; + tensor var_2062_cast_fp16 = fill_like(ref_tensor = write_indices_float_27_cast_fp16, value = var_2062_value_0_to_fp16)[name = tensor("op_2062_cast_fp16")]; + tensor write_indices_float_29_cast_fp16 = select(a = write_indices_float_27_cast_fp16, b = var_2062_cast_fp16, cond = token_valid_1_cast_fp16)[name = tensor("write_indices_float_29_cast_fp16")]; + tensor var_2068_dtype_0 = const()[name = tensor("op_2068_dtype_0"), val = tensor("int32")]; + tensor var_2073 = const()[name = tensor("op_2073"), val = tensor([1, 256, 1, 1])]; + tensor write_indices_float_29_cast_fp16_to_int32 = cast(dtype = var_2068_dtype_0, x = write_indices_float_29_cast_fp16)[name = tensor("cast_640")]; + tensor var_2074 = reshape(shape = var_2073, x = write_indices_float_29_cast_fp16_to_int32)[name = tensor("op_2074")]; + tensor write_indices_9_reps_0 = const()[name = tensor("write_indices_9_reps_0"), val = tensor([1, 1, 16, 64])]; + tensor write_indices_9 = tile(reps = write_indices_9_reps_0, x = var_2074)[name = tensor("write_indices_9")]; + tensor var_2084_begin_0 = const()[name = tensor("op_2084_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_2084_end_0 = const()[name = tensor("op_2084_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor var_2084_end_mask_0 = const()[name = tensor("op_2084_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_2084_squeeze_mask_0 = const()[name = tensor("op_2084_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor cache4_to_fp16_dtype_0 = const()[name = tensor("cache4_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor cache4_to_fp16 = cast(dtype = cache4_to_fp16_dtype_0, x = cache4)[name = tensor("cast_639")]; + tensor var_2084_cast_fp16 = slice_by_index(begin = var_2084_begin_0, end = var_2084_end_0, end_mask = var_2084_end_mask_0, squeeze_mask = var_2084_squeeze_mask_0, x = cache4_to_fp16)[name = tensor("op_2084_cast_fp16")]; + tensor var_2086_axis_0 = const()[name = tensor("op_2086_axis_0"), val = tensor(1)]; + tensor var_2086_mode_0 = const()[name = tensor("op_2086_mode_0"), val = tensor("update")]; + tensor var_2086_validate_indices_0 = const()[name = tensor("op_2086_validate_indices_0"), val = tensor(false)]; + tensor var_2086_cast_fp16 = scatter_along_axis(axis = var_2086_axis_0, data = var_2084_cast_fp16, indices = write_indices_9, mode = var_2086_mode_0, updates = k_19_cast_fp16, validate_indices = var_2086_validate_indices_0)[name = tensor("op_2086_cast_fp16")]; + tensor concat_34 = const()[name = tensor("concat_34"), val = tensor([0, 0, 0, 0, 0])]; + tensor concat_35 = const()[name = tensor("concat_35"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_9_internal_tensor_assign_1_stride_0 = const()[name = tensor("new_cache_9_internal_tensor_assign_1_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_9_internal_tensor_assign_1_begin_mask_0 = const()[name = tensor("new_cache_9_internal_tensor_assign_1_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_9_internal_tensor_assign_1_end_mask_0 = const()[name = tensor("new_cache_9_internal_tensor_assign_1_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_9_internal_tensor_assign_1_squeeze_mask_0 = const()[name = tensor("new_cache_9_internal_tensor_assign_1_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_54 = const()[name = tensor("shape_54"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_8 = const()[name = tensor("reduce_prod_8"), val = tensor(1048576)]; + tensor range_1d_8_start_0 = const()[name = tensor("range_1d_8_start_0"), val = tensor(0)]; + tensor range_1d_8_step_0 = const()[name = tensor("range_1d_8_step_0"), val = tensor(1)]; + tensor range_1d_8 = range_1d(end = reduce_prod_8, start = range_1d_8_start_0, step = range_1d_8_step_0)[name = tensor("range_1d_8")]; + tensor reshape_40 = reshape(shape = shape_54, x = range_1d_8)[name = tensor("reshape_40")]; + tensor slice_by_index_8 = slice_by_index(begin = concat_34, begin_mask = new_cache_9_internal_tensor_assign_1_begin_mask_0, end = concat_35, end_mask = new_cache_9_internal_tensor_assign_1_end_mask_0, squeeze_mask = new_cache_9_internal_tensor_assign_1_squeeze_mask_0, stride = new_cache_9_internal_tensor_assign_1_stride_0, x = reshape_40)[name = tensor("slice_by_index_8")]; + tensor reshape_41_shape_0 = const()[name = tensor("reshape_41_shape_0"), val = tensor([-1])]; + tensor reshape_41 = reshape(shape = reshape_41_shape_0, x = slice_by_index_8)[name = tensor("reshape_41")]; + tensor reshape_42_shape_0 = const()[name = tensor("reshape_42_shape_0"), val = tensor([-1])]; + tensor reshape_42_cast_fp16 = reshape(shape = reshape_42_shape_0, x = var_2086_cast_fp16)[name = tensor("reshape_42_cast_fp16")]; + tensor reshape_43_shape_0 = const()[name = tensor("reshape_43_shape_0"), val = tensor([-1])]; + tensor reshape_43_cast_fp16 = reshape(shape = reshape_43_shape_0, x = cache4_to_fp16)[name = tensor("reshape_43_cast_fp16")]; + tensor scatter_8_mode_0 = const()[name = tensor("scatter_8_mode_0"), val = tensor("update")]; + tensor scatter_8_axis_0 = const()[name = tensor("scatter_8_axis_0"), val = tensor(0)]; + tensor scatter_8_validate_indices_0 = const()[name = tensor("scatter_8_validate_indices_0"), val = tensor(false)]; + tensor scatter_8_cast_fp16 = scatter(axis = scatter_8_axis_0, data = reshape_43_cast_fp16, indices = reshape_41, mode = scatter_8_mode_0, updates = reshape_42_cast_fp16, validate_indices = scatter_8_validate_indices_0)[name = tensor("scatter_8_cast_fp16")]; + tensor reshape_44_cast_fp16 = reshape(shape = shape_54, x = scatter_8_cast_fp16)[name = tensor("reshape_44_cast_fp16")]; + tensor var_2094_begin_0 = const()[name = tensor("op_2094_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor var_2094_end_0 = const()[name = tensor("op_2094_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor var_2094_end_mask_0 = const()[name = tensor("op_2094_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_2094_squeeze_mask_0 = const()[name = tensor("op_2094_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor var_2094_cast_fp16 = slice_by_index(begin = var_2094_begin_0, end = var_2094_end_0, end_mask = var_2094_end_mask_0, squeeze_mask = var_2094_squeeze_mask_0, x = reshape_44_cast_fp16)[name = tensor("op_2094_cast_fp16")]; + tensor var_2096_axis_0 = const()[name = tensor("op_2096_axis_0"), val = tensor(1)]; + tensor var_2096_mode_0 = const()[name = tensor("op_2096_mode_0"), val = tensor("update")]; + tensor var_2096_validate_indices_0 = const()[name = tensor("op_2096_validate_indices_0"), val = tensor(false)]; + tensor var_2096_cast_fp16 = scatter_along_axis(axis = var_2096_axis_0, data = var_2094_cast_fp16, indices = write_indices_9, mode = var_2096_mode_0, updates = v_9_cast_fp16, validate_indices = var_2096_validate_indices_0)[name = tensor("op_2096_cast_fp16")]; + tensor concat_36 = const()[name = tensor("concat_36"), val = tensor([1, 0, 0, 0, 0])]; + tensor concat_37 = const()[name = tensor("concat_37"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_9_internal_tensor_assign_2_stride_0 = const()[name = tensor("new_cache_9_internal_tensor_assign_2_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_9_internal_tensor_assign_2_begin_mask_0 = const()[name = tensor("new_cache_9_internal_tensor_assign_2_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_9_internal_tensor_assign_2_end_mask_0 = const()[name = tensor("new_cache_9_internal_tensor_assign_2_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_9_internal_tensor_assign_2_squeeze_mask_0 = const()[name = tensor("new_cache_9_internal_tensor_assign_2_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_55 = const()[name = tensor("shape_55"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_9 = const()[name = tensor("reduce_prod_9"), val = tensor(1048576)]; + tensor range_1d_9_start_0 = const()[name = tensor("range_1d_9_start_0"), val = tensor(0)]; + tensor range_1d_9_step_0 = const()[name = tensor("range_1d_9_step_0"), val = tensor(1)]; + tensor range_1d_9 = range_1d(end = reduce_prod_9, start = range_1d_9_start_0, step = range_1d_9_step_0)[name = tensor("range_1d_9")]; + tensor reshape_45 = reshape(shape = shape_55, x = range_1d_9)[name = tensor("reshape_45")]; + tensor slice_by_index_9 = slice_by_index(begin = concat_36, begin_mask = new_cache_9_internal_tensor_assign_2_begin_mask_0, end = concat_37, end_mask = new_cache_9_internal_tensor_assign_2_end_mask_0, squeeze_mask = new_cache_9_internal_tensor_assign_2_squeeze_mask_0, stride = new_cache_9_internal_tensor_assign_2_stride_0, x = reshape_45)[name = tensor("slice_by_index_9")]; + tensor reshape_46_shape_0 = const()[name = tensor("reshape_46_shape_0"), val = tensor([-1])]; + tensor reshape_46 = reshape(shape = reshape_46_shape_0, x = slice_by_index_9)[name = tensor("reshape_46")]; + tensor reshape_47_shape_0 = const()[name = tensor("reshape_47_shape_0"), val = tensor([-1])]; + tensor reshape_47_cast_fp16 = reshape(shape = reshape_47_shape_0, x = var_2096_cast_fp16)[name = tensor("reshape_47_cast_fp16")]; + tensor reshape_48_shape_0 = const()[name = tensor("reshape_48_shape_0"), val = tensor([-1])]; + tensor reshape_48_cast_fp16 = reshape(shape = reshape_48_shape_0, x = reshape_44_cast_fp16)[name = tensor("reshape_48_cast_fp16")]; + tensor scatter_9_mode_0 = const()[name = tensor("scatter_9_mode_0"), val = tensor("update")]; + tensor scatter_9_axis_0 = const()[name = tensor("scatter_9_axis_0"), val = tensor(0)]; + tensor scatter_9_validate_indices_0 = const()[name = tensor("scatter_9_validate_indices_0"), val = tensor(false)]; + tensor scatter_9_cast_fp16 = scatter(axis = scatter_9_axis_0, data = reshape_48_cast_fp16, indices = reshape_46, mode = scatter_9_mode_0, updates = reshape_47_cast_fp16, validate_indices = scatter_9_validate_indices_0)[name = tensor("scatter_9_cast_fp16")]; + tensor reshape_49_cast_fp16 = reshape(shape = shape_55, x = scatter_9_cast_fp16)[name = tensor("reshape_49_cast_fp16")]; + tensor reshape_49_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("reshape_49_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor keys_25_begin_0 = const()[name = tensor("keys_25_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor keys_25_end_0 = const()[name = tensor("keys_25_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor keys_25_end_mask_0 = const()[name = tensor("keys_25_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor keys_25_squeeze_mask_0 = const()[name = tensor("keys_25_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor keys_25_cast_fp16 = slice_by_index(begin = keys_25_begin_0, end = keys_25_end_0, end_mask = keys_25_end_mask_0, squeeze_mask = keys_25_squeeze_mask_0, x = reshape_49_cast_fp16)[name = tensor("keys_25_cast_fp16")]; + tensor values_25_begin_0 = const()[name = tensor("values_25_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor values_25_end_0 = const()[name = tensor("values_25_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor values_25_end_mask_0 = const()[name = tensor("values_25_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor values_25_squeeze_mask_0 = const()[name = tensor("values_25_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor values_25_cast_fp16 = slice_by_index(begin = values_25_begin_0, end = values_25_end_0, end_mask = values_25_end_mask_0, squeeze_mask = values_25_squeeze_mask_0, x = reshape_49_cast_fp16)[name = tensor("values_25_cast_fp16")]; + tensor var_2108_cast_fp16 = not_equal(x = keys_25_cast_fp16, y = keys_25_cast_fp16)[name = tensor("op_2108_cast_fp16")]; + tensor keys_27_cast_fp16 = select(a = var_470_to_fp16, b = keys_25_cast_fp16, cond = var_2108_cast_fp16)[name = tensor("keys_27_cast_fp16")]; + tensor var_2116_cast_fp16 = not_equal(x = values_25_cast_fp16, y = values_25_cast_fp16)[name = tensor("op_2116_cast_fp16")]; + tensor values_27_cast_fp16 = select(a = var_470_to_fp16, b = values_25_cast_fp16, cond = var_2116_cast_fp16)[name = tensor("values_27_cast_fp16")]; + tensor var_2140 = const()[name = tensor("op_2140"), val = tensor([0, 2, 1, 3])]; + tensor var_2156 = const()[name = tensor("op_2156"), val = tensor([1, 1, 1])]; + tensor var_2157_cast_fp16 = reshape(shape = var_2156, x = position4_to_fp16)[name = tensor("op_2157_cast_fp16")]; + tensor q_positions_9_cast_fp16 = add(x = var_2157_cast_fp16, y = q_offsets_1_promoted_to_fp16)[name = tensor("q_positions_9_cast_fp16")]; + tensor valid_end_9_cast_fp16 = add(x = var_2157_cast_fp16, y = var_536_cast_fp16)[name = tensor("valid_end_9_cast_fp16")]; + tensor valid_mask_9_cast_fp16 = less(x = k_positions_1_promoted_to_fp16, y = valid_end_9_cast_fp16)[name = tensor("valid_mask_9_cast_fp16")]; + tensor causal_mask_9_cast_fp16 = less_equal(x = k_positions_1_promoted_to_fp16, y = q_positions_9_cast_fp16)[name = tensor("causal_mask_9_cast_fp16")]; + tensor var_2185 = logical_and(x = valid_mask_9_cast_fp16, y = causal_mask_9_cast_fp16)[name = tensor("op_2185")]; + tensor attn_mask_9_axes_0 = const()[name = tensor("attn_mask_9_axes_0"), val = tensor([1])]; + tensor attn_mask_9 = expand_dims(axes = attn_mask_9_axes_0, x = var_2185)[name = tensor("attn_mask_9")]; + tensor var_2199_transpose_x_0 = const()[name = tensor("op_2199_transpose_x_0"), val = tensor(false)]; + tensor var_2199_transpose_y_0 = const()[name = tensor("op_2199_transpose_y_0"), val = tensor(false)]; + tensor transpose_77_perm_0 = const()[name = tensor("transpose_77_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_78_perm_0 = const()[name = tensor("transpose_78_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_78 = transpose(perm = transpose_78_perm_0, x = keys_27_cast_fp16)[name = tensor("transpose_188")]; + tensor transpose_77 = transpose(perm = transpose_77_perm_0, x = q_27_cast_fp16)[name = tensor("transpose_189")]; + tensor var_2199_cast_fp16 = matmul(transpose_x = var_2199_transpose_x_0, transpose_y = var_2199_transpose_y_0, x = transpose_77, y = transpose_78)[name = tensor("op_2199_cast_fp16")]; + tensor var_2193_to_fp16 = const()[name = tensor("op_2193_to_fp16"), val = tensor([0x1p-3])]; + tensor attn_weights_25_cast_fp16 = mul(x = var_2199_cast_fp16, y = var_2193_to_fp16)[name = tensor("attn_weights_25_cast_fp16")]; + tensor var_2201 = logical_not(x = attn_mask_9)[name = tensor("op_2201")]; + tensor var_2202_to_fp16 = const()[name = tensor("op_2202_to_fp16"), val = tensor(-inf)]; + tensor attn_weights_27_cast_fp16 = select(a = var_2202_to_fp16, b = attn_weights_25_cast_fp16, cond = var_2201)[name = tensor("attn_weights_27_cast_fp16")]; + tensor var_2204 = const()[name = tensor("op_2204"), val = tensor(-1)]; + tensor attn_weights_29_cast_fp16 = softmax(axis = var_2204, x = attn_weights_27_cast_fp16)[name = tensor("attn_weights_29_cast_fp16")]; + tensor attn_output_9_transpose_x_0 = const()[name = tensor("attn_output_9_transpose_x_0"), val = tensor(false)]; + tensor attn_output_9_transpose_y_0 = const()[name = tensor("attn_output_9_transpose_y_0"), val = tensor(false)]; + tensor values_29_cast_fp16 = transpose(perm = var_2140, x = values_27_cast_fp16)[name = tensor("transpose_190")]; + tensor attn_output_9_cast_fp16 = matmul(transpose_x = attn_output_9_transpose_x_0, transpose_y = attn_output_9_transpose_y_0, x = attn_weights_29_cast_fp16, y = values_29_cast_fp16)[name = tensor("attn_output_9_cast_fp16")]; + tensor var_2212 = const()[name = tensor("op_2212"), val = tensor([0, 2, 1, 3])]; + tensor var_2217 = const()[name = tensor("op_2217"), val = tensor([1, 256, 1024])]; + tensor var_2213_cast_fp16 = transpose(perm = var_2212, x = attn_output_9_cast_fp16)[name = tensor("transpose_187")]; + tensor input_53_cast_fp16 = reshape(shape = var_2217, x = var_2213_cast_fp16)[name = tensor("input_53_cast_fp16")]; + tensor attn4_out_proj_weight_to_fp16 = const()[name = tensor("attn4_out_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(108062592)))]; + tensor linear_17_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = attn4_out_proj_weight_to_fp16, x = input_53_cast_fp16)[name = tensor("linear_17_cast_fp16")]; + tensor var_2226_cast_fp16 = add(x = position4_to_fp16, y = valid_len_to_fp16)[name = tensor("op_2226_cast_fp16")]; + tensor var_2226_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_2226_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor input_55_cast_fp16 = add(x = input_49_cast_fp16, y = linear_17_cast_fp16)[name = tensor("input_55_cast_fp16")]; + tensor input_57_axes_0 = const()[name = tensor("input_57_axes_0"), val = tensor([-1])]; + tensor norm4_2_weight_to_fp16 = const()[name = tensor("norm4_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(110159808)))]; + tensor norm4_2_bias_to_fp16 = const()[name = tensor("norm4_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(110161920)))]; + tensor var_2230_to_fp16 = const()[name = tensor("op_2230_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_57_cast_fp16 = layer_norm(axes = input_57_axes_0, beta = norm4_2_bias_to_fp16, epsilon = var_2230_to_fp16, gamma = norm4_2_weight_to_fp16, x = input_55_cast_fp16)[name = tensor("input_57_cast_fp16")]; + tensor linear4_1_weight_to_fp16 = const()[name = tensor("linear4_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(110164032)))]; + tensor linear_18_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = linear4_1_weight_to_fp16, x = input_57_cast_fp16)[name = tensor("linear_18_cast_fp16")]; + tensor input_59_mode_0 = const()[name = tensor("input_59_mode_0"), val = tensor("EXACT")]; + tensor input_59_cast_fp16 = gelu(mode = input_59_mode_0, x = linear_18_cast_fp16)[name = tensor("input_59_cast_fp16")]; + tensor linear4_2_weight_to_fp16 = const()[name = tensor("linear4_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(118552704)))]; + tensor linear_19_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = linear4_2_weight_to_fp16, x = input_59_cast_fp16)[name = tensor("linear_19_cast_fp16")]; + tensor input_61_cast_fp16 = add(x = input_55_cast_fp16, y = linear_19_cast_fp16)[name = tensor("input_61_cast_fp16")]; + tensor input_63_axes_0 = const()[name = tensor("input_63_axes_0"), val = tensor([-1])]; + tensor norm5_1_weight_to_fp16 = const()[name = tensor("norm5_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(126941376)))]; + tensor norm5_1_bias_to_fp16 = const()[name = tensor("norm5_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(126943488)))]; + tensor var_2247_to_fp16 = const()[name = tensor("op_2247_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_63_cast_fp16 = layer_norm(axes = input_63_axes_0, beta = norm5_1_bias_to_fp16, epsilon = var_2247_to_fp16, gamma = norm5_1_weight_to_fp16, x = input_61_cast_fp16)[name = tensor("input_63_cast_fp16")]; + tensor attn5_in_proj_weight_to_fp16 = const()[name = tensor("attn5_in_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(126945600)))]; + tensor linear_20_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = attn5_in_proj_weight_to_fp16, x = input_63_cast_fp16)[name = tensor("linear_20_cast_fp16")]; + tensor var_2261 = const()[name = tensor("op_2261"), val = tensor([1, 256, 3, 16, 64])]; + tensor qkv_11_cast_fp16 = reshape(shape = var_2261, x = linear_20_cast_fp16)[name = tensor("qkv_11_cast_fp16")]; + tensor q_31_begin_0 = const()[name = tensor("q_31_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor q_31_end_0 = const()[name = tensor("q_31_end_0"), val = tensor([1, 256, 1, 16, 64])]; + tensor q_31_end_mask_0 = const()[name = tensor("q_31_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor q_31_squeeze_mask_0 = const()[name = tensor("q_31_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor q_31_cast_fp16 = slice_by_index(begin = q_31_begin_0, end = q_31_end_0, end_mask = q_31_end_mask_0, squeeze_mask = q_31_squeeze_mask_0, x = qkv_11_cast_fp16)[name = tensor("q_31_cast_fp16")]; + tensor k_21_begin_0 = const()[name = tensor("k_21_begin_0"), val = tensor([0, 0, 1, 0, 0])]; + tensor k_21_end_0 = const()[name = tensor("k_21_end_0"), val = tensor([1, 256, 2, 16, 64])]; + tensor k_21_end_mask_0 = const()[name = tensor("k_21_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor k_21_squeeze_mask_0 = const()[name = tensor("k_21_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor k_21_cast_fp16 = slice_by_index(begin = k_21_begin_0, end = k_21_end_0, end_mask = k_21_end_mask_0, squeeze_mask = k_21_squeeze_mask_0, x = qkv_11_cast_fp16)[name = tensor("k_21_cast_fp16")]; + tensor v_11_begin_0 = const()[name = tensor("v_11_begin_0"), val = tensor([0, 0, 2, 0, 0])]; + tensor v_11_end_0 = const()[name = tensor("v_11_end_0"), val = tensor([1, 256, 3, 16, 64])]; + tensor v_11_end_mask_0 = const()[name = tensor("v_11_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor v_11_squeeze_mask_0 = const()[name = tensor("v_11_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor v_11_cast_fp16 = slice_by_index(begin = v_11_begin_0, end = v_11_end_0, end_mask = v_11_end_mask_0, squeeze_mask = v_11_squeeze_mask_0, x = qkv_11_cast_fp16)[name = tensor("v_11_cast_fp16")]; + tensor var_2319 = const()[name = tensor("op_2319"), val = tensor([1, 1])]; + tensor position5_to_fp16_dtype_0 = const()[name = tensor("position5_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor position5_to_fp16 = cast(dtype = position5_to_fp16_dtype_0, x = position5)[name = tensor("cast_636")]; + tensor var_2320_cast_fp16 = reshape(shape = var_2319, x = position5_to_fp16)[name = tensor("op_2320_cast_fp16")]; + tensor ts_33_cast_fp16 = add(x = ts_1_promoted_to_fp16, y = var_2320_cast_fp16)[name = tensor("ts_33_cast_fp16")]; + tensor var_2327 = const()[name = tensor("op_2327"), val = tensor([1, 256, 1, 1])]; + tensor ts_35_cast_fp16 = reshape(shape = var_2327, x = ts_33_cast_fp16)[name = tensor("ts_35_cast_fp16")]; + tensor var_2334 = const()[name = tensor("op_2334"), val = tensor([1, 256, 16, 32, 2])]; + tensor q_complex_11_cast_fp16 = reshape(shape = var_2334, x = q_31_cast_fp16)[name = tensor("q_complex_11_cast_fp16")]; + tensor var_2341 = const()[name = tensor("op_2341"), val = tensor([1, 256, 16, 32, 2])]; + tensor k_complex_11_cast_fp16 = reshape(shape = var_2341, x = k_21_cast_fp16)[name = tensor("k_complex_11_cast_fp16")]; + tensor var_2345_begin_0 = const()[name = tensor("op_2345_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_2345_end_0 = const()[name = tensor("op_2345_end_0"), val = tensor([1, 256, 16, 32, 1])]; + tensor var_2345_end_mask_0 = const()[name = tensor("op_2345_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_2345_squeeze_mask_0 = const()[name = tensor("op_2345_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_2345_cast_fp16 = slice_by_index(begin = var_2345_begin_0, end = var_2345_end_0, end_mask = var_2345_end_mask_0, squeeze_mask = var_2345_squeeze_mask_0, x = q_complex_11_cast_fp16)[name = tensor("op_2345_cast_fp16")]; + tensor var_2353_begin_0 = const()[name = tensor("op_2353_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_2353_end_0 = const()[name = tensor("op_2353_end_0"), val = tensor([1, 256, 16, 32, 2])]; + tensor var_2353_end_mask_0 = const()[name = tensor("op_2353_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_2353_squeeze_mask_0 = const()[name = tensor("op_2353_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_2353_cast_fp16 = slice_by_index(begin = var_2353_begin_0, end = var_2353_end_0, end_mask = var_2353_end_mask_0, squeeze_mask = var_2353_squeeze_mask_0, x = q_complex_11_cast_fp16)[name = tensor("op_2353_cast_fp16")]; + tensor var_2361_begin_0 = const()[name = tensor("op_2361_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_2361_end_0 = const()[name = tensor("op_2361_end_0"), val = tensor([1, 256, 16, 32, 1])]; + tensor var_2361_end_mask_0 = const()[name = tensor("op_2361_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_2361_squeeze_mask_0 = const()[name = tensor("op_2361_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_2361_cast_fp16 = slice_by_index(begin = var_2361_begin_0, end = var_2361_end_0, end_mask = var_2361_end_mask_0, squeeze_mask = var_2361_squeeze_mask_0, x = k_complex_11_cast_fp16)[name = tensor("op_2361_cast_fp16")]; + tensor var_2369_begin_0 = const()[name = tensor("op_2369_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_2369_end_0 = const()[name = tensor("op_2369_end_0"), val = tensor([1, 256, 16, 32, 2])]; + tensor var_2369_end_mask_0 = const()[name = tensor("op_2369_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_2369_squeeze_mask_0 = const()[name = tensor("op_2369_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_2369_cast_fp16 = slice_by_index(begin = var_2369_begin_0, end = var_2369_end_0, end_mask = var_2369_end_mask_0, squeeze_mask = var_2369_squeeze_mask_0, x = k_complex_11_cast_fp16)[name = tensor("op_2369_cast_fp16")]; + tensor freqs_11_to_fp16 = const()[name = tensor("freqs_11_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(133237120)))]; + tensor var_2375_cast_fp16 = mul(x = freqs_11_to_fp16, y = ts_35_cast_fp16)[name = tensor("op_2375_cast_fp16")]; + tensor rotr_11_cast_fp16 = cos(x = var_2375_cast_fp16)[name = tensor("rotr_11_cast_fp16")]; + tensor roti_11_cast_fp16 = sin(x = var_2375_cast_fp16)[name = tensor("roti_11_cast_fp16")]; + tensor var_2379_cast_fp16 = mul(x = var_2345_cast_fp16, y = rotr_11_cast_fp16)[name = tensor("op_2379_cast_fp16")]; + tensor var_2380_cast_fp16 = mul(x = var_2353_cast_fp16, y = roti_11_cast_fp16)[name = tensor("op_2380_cast_fp16")]; + tensor qor_21_cast_fp16 = sub(x = var_2379_cast_fp16, y = var_2380_cast_fp16)[name = tensor("qor_21_cast_fp16")]; + tensor var_2383_cast_fp16 = mul(x = var_2345_cast_fp16, y = roti_11_cast_fp16)[name = tensor("op_2383_cast_fp16")]; + tensor var_2384_cast_fp16 = mul(x = var_2353_cast_fp16, y = rotr_11_cast_fp16)[name = tensor("op_2384_cast_fp16")]; + tensor qoi_21_cast_fp16 = add(x = var_2383_cast_fp16, y = var_2384_cast_fp16)[name = tensor("qoi_21_cast_fp16")]; + tensor var_2387_cast_fp16 = mul(x = var_2361_cast_fp16, y = rotr_11_cast_fp16)[name = tensor("op_2387_cast_fp16")]; + tensor var_2388_cast_fp16 = mul(x = var_2369_cast_fp16, y = roti_11_cast_fp16)[name = tensor("op_2388_cast_fp16")]; + tensor kor_21_cast_fp16 = sub(x = var_2387_cast_fp16, y = var_2388_cast_fp16)[name = tensor("kor_21_cast_fp16")]; + tensor var_2391_cast_fp16 = mul(x = var_2361_cast_fp16, y = roti_11_cast_fp16)[name = tensor("op_2391_cast_fp16")]; + tensor var_2392_cast_fp16 = mul(x = var_2369_cast_fp16, y = rotr_11_cast_fp16)[name = tensor("op_2392_cast_fp16")]; + tensor koi_21_cast_fp16 = add(x = var_2391_cast_fp16, y = var_2392_cast_fp16)[name = tensor("koi_21_cast_fp16")]; + tensor qo_11_axis_0 = const()[name = tensor("qo_11_axis_0"), val = tensor(-1)]; + tensor qo_11_cast_fp16 = stack(axis = qo_11_axis_0, values = (qor_21_cast_fp16, qoi_21_cast_fp16))[name = tensor("qo_11_cast_fp16")]; + tensor ko_11_axis_0 = const()[name = tensor("ko_11_axis_0"), val = tensor(-1)]; + tensor ko_11_cast_fp16 = stack(axis = ko_11_axis_0, values = (kor_21_cast_fp16, koi_21_cast_fp16))[name = tensor("ko_11_cast_fp16")]; + tensor var_2425 = const()[name = tensor("op_2425"), val = tensor([1, 256, 16, 64])]; + tensor q_33_cast_fp16 = reshape(shape = var_2425, x = qo_11_cast_fp16)[name = tensor("q_33_cast_fp16")]; + tensor var_2431 = const()[name = tensor("op_2431"), val = tensor([1, 256, 16, 64])]; + tensor k_23_cast_fp16 = reshape(shape = var_2431, x = ko_11_cast_fp16)[name = tensor("k_23_cast_fp16")]; + tensor write_indices_float_31_cast_fp16 = add(x = var_2320_cast_fp16, y = var_393_promoted_to_fp16)[name = tensor("write_indices_float_31_cast_fp16")]; + tensor _inversed_2452_y_0_to_fp16 = const()[name = tensor("_inversed_2452_y_0_to_fp16"), val = tensor(0x1p-9)]; + tensor _inversed_2452_cast_fp16 = mul(x = write_indices_float_31_cast_fp16, y = _inversed_2452_y_0_to_fp16)[name = tensor("_inversed_2452_cast_fp16")]; + tensor var_2453_cast_fp16 = floor(x = _inversed_2452_cast_fp16)[name = tensor("op_2453_cast_fp16")]; + tensor var_2454_to_fp16 = const()[name = tensor("op_2454_to_fp16"), val = tensor(0x1p+9)]; + tensor var_2455_cast_fp16 = mul(x = var_2453_cast_fp16, y = var_2454_to_fp16)[name = tensor("op_2455_cast_fp16")]; + tensor write_indices_float_33_cast_fp16 = sub(x = write_indices_float_31_cast_fp16, y = var_2455_cast_fp16)[name = tensor("write_indices_float_33_cast_fp16")]; + tensor var_2473_value_0_to_fp16 = const()[name = tensor("op_2473_value_0_to_fp16"), val = tensor(0x1.ffp+8)]; + tensor var_2473_cast_fp16 = fill_like(ref_tensor = write_indices_float_33_cast_fp16, value = var_2473_value_0_to_fp16)[name = tensor("op_2473_cast_fp16")]; + tensor write_indices_float_35_cast_fp16 = select(a = write_indices_float_33_cast_fp16, b = var_2473_cast_fp16, cond = token_valid_1_cast_fp16)[name = tensor("write_indices_float_35_cast_fp16")]; + tensor var_2479_dtype_0 = const()[name = tensor("op_2479_dtype_0"), val = tensor("int32")]; + tensor var_2484 = const()[name = tensor("op_2484"), val = tensor([1, 256, 1, 1])]; + tensor write_indices_float_35_cast_fp16_to_int32 = cast(dtype = var_2479_dtype_0, x = write_indices_float_35_cast_fp16)[name = tensor("cast_635")]; + tensor var_2485 = reshape(shape = var_2484, x = write_indices_float_35_cast_fp16_to_int32)[name = tensor("op_2485")]; + tensor write_indices_11_reps_0 = const()[name = tensor("write_indices_11_reps_0"), val = tensor([1, 1, 16, 64])]; + tensor write_indices_11 = tile(reps = write_indices_11_reps_0, x = var_2485)[name = tensor("write_indices_11")]; + tensor var_2495_begin_0 = const()[name = tensor("op_2495_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_2495_end_0 = const()[name = tensor("op_2495_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor var_2495_end_mask_0 = const()[name = tensor("op_2495_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_2495_squeeze_mask_0 = const()[name = tensor("op_2495_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor cache5_to_fp16_dtype_0 = const()[name = tensor("cache5_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor cache5_to_fp16 = cast(dtype = cache5_to_fp16_dtype_0, x = cache5)[name = tensor("cast_634")]; + tensor var_2495_cast_fp16 = slice_by_index(begin = var_2495_begin_0, end = var_2495_end_0, end_mask = var_2495_end_mask_0, squeeze_mask = var_2495_squeeze_mask_0, x = cache5_to_fp16)[name = tensor("op_2495_cast_fp16")]; + tensor var_2497_axis_0 = const()[name = tensor("op_2497_axis_0"), val = tensor(1)]; + tensor var_2497_mode_0 = const()[name = tensor("op_2497_mode_0"), val = tensor("update")]; + tensor var_2497_validate_indices_0 = const()[name = tensor("op_2497_validate_indices_0"), val = tensor(false)]; + tensor var_2497_cast_fp16 = scatter_along_axis(axis = var_2497_axis_0, data = var_2495_cast_fp16, indices = write_indices_11, mode = var_2497_mode_0, updates = k_23_cast_fp16, validate_indices = var_2497_validate_indices_0)[name = tensor("op_2497_cast_fp16")]; + tensor concat_42 = const()[name = tensor("concat_42"), val = tensor([0, 0, 0, 0, 0])]; + tensor concat_43 = const()[name = tensor("concat_43"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_11_internal_tensor_assign_1_stride_0 = const()[name = tensor("new_cache_11_internal_tensor_assign_1_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_11_internal_tensor_assign_1_begin_mask_0 = const()[name = tensor("new_cache_11_internal_tensor_assign_1_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_11_internal_tensor_assign_1_end_mask_0 = const()[name = tensor("new_cache_11_internal_tensor_assign_1_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_11_internal_tensor_assign_1_squeeze_mask_0 = const()[name = tensor("new_cache_11_internal_tensor_assign_1_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_56 = const()[name = tensor("shape_56"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_10 = const()[name = tensor("reduce_prod_10"), val = tensor(1048576)]; + tensor range_1d_10_start_0 = const()[name = tensor("range_1d_10_start_0"), val = tensor(0)]; + tensor range_1d_10_step_0 = const()[name = tensor("range_1d_10_step_0"), val = tensor(1)]; + tensor range_1d_10 = range_1d(end = reduce_prod_10, start = range_1d_10_start_0, step = range_1d_10_step_0)[name = tensor("range_1d_10")]; + tensor reshape_50 = reshape(shape = shape_56, x = range_1d_10)[name = tensor("reshape_50")]; + tensor slice_by_index_10 = slice_by_index(begin = concat_42, begin_mask = new_cache_11_internal_tensor_assign_1_begin_mask_0, end = concat_43, end_mask = new_cache_11_internal_tensor_assign_1_end_mask_0, squeeze_mask = new_cache_11_internal_tensor_assign_1_squeeze_mask_0, stride = new_cache_11_internal_tensor_assign_1_stride_0, x = reshape_50)[name = tensor("slice_by_index_10")]; + tensor reshape_51_shape_0 = const()[name = tensor("reshape_51_shape_0"), val = tensor([-1])]; + tensor reshape_51 = reshape(shape = reshape_51_shape_0, x = slice_by_index_10)[name = tensor("reshape_51")]; + tensor reshape_52_shape_0 = const()[name = tensor("reshape_52_shape_0"), val = tensor([-1])]; + tensor reshape_52_cast_fp16 = reshape(shape = reshape_52_shape_0, x = var_2497_cast_fp16)[name = tensor("reshape_52_cast_fp16")]; + tensor reshape_53_shape_0 = const()[name = tensor("reshape_53_shape_0"), val = tensor([-1])]; + tensor reshape_53_cast_fp16 = reshape(shape = reshape_53_shape_0, x = cache5_to_fp16)[name = tensor("reshape_53_cast_fp16")]; + tensor scatter_10_mode_0 = const()[name = tensor("scatter_10_mode_0"), val = tensor("update")]; + tensor scatter_10_axis_0 = const()[name = tensor("scatter_10_axis_0"), val = tensor(0)]; + tensor scatter_10_validate_indices_0 = const()[name = tensor("scatter_10_validate_indices_0"), val = tensor(false)]; + tensor scatter_10_cast_fp16 = scatter(axis = scatter_10_axis_0, data = reshape_53_cast_fp16, indices = reshape_51, mode = scatter_10_mode_0, updates = reshape_52_cast_fp16, validate_indices = scatter_10_validate_indices_0)[name = tensor("scatter_10_cast_fp16")]; + tensor reshape_54_cast_fp16 = reshape(shape = shape_56, x = scatter_10_cast_fp16)[name = tensor("reshape_54_cast_fp16")]; + tensor var_2505_begin_0 = const()[name = tensor("op_2505_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor var_2505_end_0 = const()[name = tensor("op_2505_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor var_2505_end_mask_0 = const()[name = tensor("op_2505_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_2505_squeeze_mask_0 = const()[name = tensor("op_2505_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor var_2505_cast_fp16 = slice_by_index(begin = var_2505_begin_0, end = var_2505_end_0, end_mask = var_2505_end_mask_0, squeeze_mask = var_2505_squeeze_mask_0, x = reshape_54_cast_fp16)[name = tensor("op_2505_cast_fp16")]; + tensor var_2507_axis_0 = const()[name = tensor("op_2507_axis_0"), val = tensor(1)]; + tensor var_2507_mode_0 = const()[name = tensor("op_2507_mode_0"), val = tensor("update")]; + tensor var_2507_validate_indices_0 = const()[name = tensor("op_2507_validate_indices_0"), val = tensor(false)]; + tensor var_2507_cast_fp16 = scatter_along_axis(axis = var_2507_axis_0, data = var_2505_cast_fp16, indices = write_indices_11, mode = var_2507_mode_0, updates = v_11_cast_fp16, validate_indices = var_2507_validate_indices_0)[name = tensor("op_2507_cast_fp16")]; + tensor concat_44 = const()[name = tensor("concat_44"), val = tensor([1, 0, 0, 0, 0])]; + tensor concat_45 = const()[name = tensor("concat_45"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_11_internal_tensor_assign_2_stride_0 = const()[name = tensor("new_cache_11_internal_tensor_assign_2_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_11_internal_tensor_assign_2_begin_mask_0 = const()[name = tensor("new_cache_11_internal_tensor_assign_2_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_11_internal_tensor_assign_2_end_mask_0 = const()[name = tensor("new_cache_11_internal_tensor_assign_2_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_11_internal_tensor_assign_2_squeeze_mask_0 = const()[name = tensor("new_cache_11_internal_tensor_assign_2_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_57 = const()[name = tensor("shape_57"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_11 = const()[name = tensor("reduce_prod_11"), val = tensor(1048576)]; + tensor range_1d_11_start_0 = const()[name = tensor("range_1d_11_start_0"), val = tensor(0)]; + tensor range_1d_11_step_0 = const()[name = tensor("range_1d_11_step_0"), val = tensor(1)]; + tensor range_1d_11 = range_1d(end = reduce_prod_11, start = range_1d_11_start_0, step = range_1d_11_step_0)[name = tensor("range_1d_11")]; + tensor reshape_55 = reshape(shape = shape_57, x = range_1d_11)[name = tensor("reshape_55")]; + tensor slice_by_index_11 = slice_by_index(begin = concat_44, begin_mask = new_cache_11_internal_tensor_assign_2_begin_mask_0, end = concat_45, end_mask = new_cache_11_internal_tensor_assign_2_end_mask_0, squeeze_mask = new_cache_11_internal_tensor_assign_2_squeeze_mask_0, stride = new_cache_11_internal_tensor_assign_2_stride_0, x = reshape_55)[name = tensor("slice_by_index_11")]; + tensor reshape_56_shape_0 = const()[name = tensor("reshape_56_shape_0"), val = tensor([-1])]; + tensor reshape_56 = reshape(shape = reshape_56_shape_0, x = slice_by_index_11)[name = tensor("reshape_56")]; + tensor reshape_57_shape_0 = const()[name = tensor("reshape_57_shape_0"), val = tensor([-1])]; + tensor reshape_57_cast_fp16 = reshape(shape = reshape_57_shape_0, x = var_2507_cast_fp16)[name = tensor("reshape_57_cast_fp16")]; + tensor reshape_58_shape_0 = const()[name = tensor("reshape_58_shape_0"), val = tensor([-1])]; + tensor reshape_58_cast_fp16 = reshape(shape = reshape_58_shape_0, x = reshape_54_cast_fp16)[name = tensor("reshape_58_cast_fp16")]; + tensor scatter_11_mode_0 = const()[name = tensor("scatter_11_mode_0"), val = tensor("update")]; + tensor scatter_11_axis_0 = const()[name = tensor("scatter_11_axis_0"), val = tensor(0)]; + tensor scatter_11_validate_indices_0 = const()[name = tensor("scatter_11_validate_indices_0"), val = tensor(false)]; + tensor scatter_11_cast_fp16 = scatter(axis = scatter_11_axis_0, data = reshape_58_cast_fp16, indices = reshape_56, mode = scatter_11_mode_0, updates = reshape_57_cast_fp16, validate_indices = scatter_11_validate_indices_0)[name = tensor("scatter_11_cast_fp16")]; + tensor reshape_59_cast_fp16 = reshape(shape = shape_57, x = scatter_11_cast_fp16)[name = tensor("reshape_59_cast_fp16")]; + tensor reshape_59_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("reshape_59_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor keys_31_begin_0 = const()[name = tensor("keys_31_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor keys_31_end_0 = const()[name = tensor("keys_31_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor keys_31_end_mask_0 = const()[name = tensor("keys_31_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor keys_31_squeeze_mask_0 = const()[name = tensor("keys_31_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor keys_31_cast_fp16 = slice_by_index(begin = keys_31_begin_0, end = keys_31_end_0, end_mask = keys_31_end_mask_0, squeeze_mask = keys_31_squeeze_mask_0, x = reshape_59_cast_fp16)[name = tensor("keys_31_cast_fp16")]; + tensor values_31_begin_0 = const()[name = tensor("values_31_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor values_31_end_0 = const()[name = tensor("values_31_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor values_31_end_mask_0 = const()[name = tensor("values_31_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor values_31_squeeze_mask_0 = const()[name = tensor("values_31_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor values_31_cast_fp16 = slice_by_index(begin = values_31_begin_0, end = values_31_end_0, end_mask = values_31_end_mask_0, squeeze_mask = values_31_squeeze_mask_0, x = reshape_59_cast_fp16)[name = tensor("values_31_cast_fp16")]; + tensor var_2519_cast_fp16 = not_equal(x = keys_31_cast_fp16, y = keys_31_cast_fp16)[name = tensor("op_2519_cast_fp16")]; + tensor keys_33_cast_fp16 = select(a = var_470_to_fp16, b = keys_31_cast_fp16, cond = var_2519_cast_fp16)[name = tensor("keys_33_cast_fp16")]; + tensor var_2527_cast_fp16 = not_equal(x = values_31_cast_fp16, y = values_31_cast_fp16)[name = tensor("op_2527_cast_fp16")]; + tensor values_33_cast_fp16 = select(a = var_470_to_fp16, b = values_31_cast_fp16, cond = var_2527_cast_fp16)[name = tensor("values_33_cast_fp16")]; + tensor var_2551 = const()[name = tensor("op_2551"), val = tensor([0, 2, 1, 3])]; + tensor var_2567 = const()[name = tensor("op_2567"), val = tensor([1, 1, 1])]; + tensor var_2568_cast_fp16 = reshape(shape = var_2567, x = position5_to_fp16)[name = tensor("op_2568_cast_fp16")]; + tensor q_positions_11_cast_fp16 = add(x = var_2568_cast_fp16, y = q_offsets_1_promoted_to_fp16)[name = tensor("q_positions_11_cast_fp16")]; + tensor valid_end_11_cast_fp16 = add(x = var_2568_cast_fp16, y = var_536_cast_fp16)[name = tensor("valid_end_11_cast_fp16")]; + tensor valid_mask_11_cast_fp16 = less(x = k_positions_1_promoted_to_fp16, y = valid_end_11_cast_fp16)[name = tensor("valid_mask_11_cast_fp16")]; + tensor causal_mask_11_cast_fp16 = less_equal(x = k_positions_1_promoted_to_fp16, y = q_positions_11_cast_fp16)[name = tensor("causal_mask_11_cast_fp16")]; + tensor var_2596 = logical_and(x = valid_mask_11_cast_fp16, y = causal_mask_11_cast_fp16)[name = tensor("op_2596")]; + tensor attn_mask_11_axes_0 = const()[name = tensor("attn_mask_11_axes_0"), val = tensor([1])]; + tensor attn_mask_11 = expand_dims(axes = attn_mask_11_axes_0, x = var_2596)[name = tensor("attn_mask_11")]; + tensor var_2610_transpose_x_0 = const()[name = tensor("op_2610_transpose_x_0"), val = tensor(false)]; + tensor var_2610_transpose_y_0 = const()[name = tensor("op_2610_transpose_y_0"), val = tensor(false)]; + tensor transpose_79_perm_0 = const()[name = tensor("transpose_79_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_80_perm_0 = const()[name = tensor("transpose_80_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_80 = transpose(perm = transpose_80_perm_0, x = keys_33_cast_fp16)[name = tensor("transpose_184")]; + tensor transpose_79 = transpose(perm = transpose_79_perm_0, x = q_33_cast_fp16)[name = tensor("transpose_185")]; + tensor var_2610_cast_fp16 = matmul(transpose_x = var_2610_transpose_x_0, transpose_y = var_2610_transpose_y_0, x = transpose_79, y = transpose_80)[name = tensor("op_2610_cast_fp16")]; + tensor var_2604_to_fp16 = const()[name = tensor("op_2604_to_fp16"), val = tensor([0x1p-3])]; + tensor attn_weights_31_cast_fp16 = mul(x = var_2610_cast_fp16, y = var_2604_to_fp16)[name = tensor("attn_weights_31_cast_fp16")]; + tensor var_2612 = logical_not(x = attn_mask_11)[name = tensor("op_2612")]; + tensor var_2613_to_fp16 = const()[name = tensor("op_2613_to_fp16"), val = tensor(-inf)]; + tensor attn_weights_33_cast_fp16 = select(a = var_2613_to_fp16, b = attn_weights_31_cast_fp16, cond = var_2612)[name = tensor("attn_weights_33_cast_fp16")]; + tensor var_2615 = const()[name = tensor("op_2615"), val = tensor(-1)]; + tensor attn_weights_35_cast_fp16 = softmax(axis = var_2615, x = attn_weights_33_cast_fp16)[name = tensor("attn_weights_35_cast_fp16")]; + tensor attn_output_11_transpose_x_0 = const()[name = tensor("attn_output_11_transpose_x_0"), val = tensor(false)]; + tensor attn_output_11_transpose_y_0 = const()[name = tensor("attn_output_11_transpose_y_0"), val = tensor(false)]; + tensor values_35_cast_fp16 = transpose(perm = var_2551, x = values_33_cast_fp16)[name = tensor("transpose_186")]; + tensor attn_output_11_cast_fp16 = matmul(transpose_x = attn_output_11_transpose_x_0, transpose_y = attn_output_11_transpose_y_0, x = attn_weights_35_cast_fp16, y = values_35_cast_fp16)[name = tensor("attn_output_11_cast_fp16")]; + tensor var_2623 = const()[name = tensor("op_2623"), val = tensor([0, 2, 1, 3])]; + tensor var_2628 = const()[name = tensor("op_2628"), val = tensor([1, 256, 1024])]; + tensor var_2624_cast_fp16 = transpose(perm = var_2623, x = attn_output_11_cast_fp16)[name = tensor("transpose_183")]; + tensor input_65_cast_fp16 = reshape(shape = var_2628, x = var_2624_cast_fp16)[name = tensor("input_65_cast_fp16")]; + tensor attn5_out_proj_weight_to_fp16 = const()[name = tensor("attn5_out_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(133237248)))]; + tensor linear_21_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = attn5_out_proj_weight_to_fp16, x = input_65_cast_fp16)[name = tensor("linear_21_cast_fp16")]; + tensor var_2637_cast_fp16 = add(x = position5_to_fp16, y = valid_len_to_fp16)[name = tensor("op_2637_cast_fp16")]; + tensor var_2637_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_2637_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor input_67_cast_fp16 = add(x = input_61_cast_fp16, y = linear_21_cast_fp16)[name = tensor("input_67_cast_fp16")]; + tensor input_69_axes_0 = const()[name = tensor("input_69_axes_0"), val = tensor([-1])]; + tensor norm5_2_weight_to_fp16 = const()[name = tensor("norm5_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(135334464)))]; + tensor norm5_2_bias_to_fp16 = const()[name = tensor("norm5_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(135336576)))]; + tensor var_2641_to_fp16 = const()[name = tensor("op_2641_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_69_cast_fp16 = layer_norm(axes = input_69_axes_0, beta = norm5_2_bias_to_fp16, epsilon = var_2641_to_fp16, gamma = norm5_2_weight_to_fp16, x = input_67_cast_fp16)[name = tensor("input_69_cast_fp16")]; + tensor linear5_1_weight_to_fp16 = const()[name = tensor("linear5_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(135338688)))]; + tensor linear_22_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = linear5_1_weight_to_fp16, x = input_69_cast_fp16)[name = tensor("linear_22_cast_fp16")]; + tensor input_71_mode_0 = const()[name = tensor("input_71_mode_0"), val = tensor("EXACT")]; + tensor input_71_cast_fp16 = gelu(mode = input_71_mode_0, x = linear_22_cast_fp16)[name = tensor("input_71_cast_fp16")]; + tensor linear5_2_weight_to_fp16 = const()[name = tensor("linear5_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(143727360)))]; + tensor linear_23_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = linear5_2_weight_to_fp16, x = input_71_cast_fp16)[name = tensor("linear_23_cast_fp16")]; + tensor input_73_cast_fp16 = add(x = input_67_cast_fp16, y = linear_23_cast_fp16)[name = tensor("input_73_cast_fp16")]; + tensor input_75_axes_0 = const()[name = tensor("input_75_axes_0"), val = tensor([-1])]; + tensor norm6_1_weight_to_fp16 = const()[name = tensor("norm6_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(152116032)))]; + tensor norm6_1_bias_to_fp16 = const()[name = tensor("norm6_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(152118144)))]; + tensor var_2658_to_fp16 = const()[name = tensor("op_2658_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_75_cast_fp16 = layer_norm(axes = input_75_axes_0, beta = norm6_1_bias_to_fp16, epsilon = var_2658_to_fp16, gamma = norm6_1_weight_to_fp16, x = input_73_cast_fp16)[name = tensor("input_75_cast_fp16")]; + tensor attn6_in_proj_weight_to_fp16 = const()[name = tensor("attn6_in_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(152120256)))]; + tensor linear_24_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = attn6_in_proj_weight_to_fp16, x = input_75_cast_fp16)[name = tensor("linear_24_cast_fp16")]; + tensor var_2672 = const()[name = tensor("op_2672"), val = tensor([1, 256, 3, 16, 64])]; + tensor qkv_13_cast_fp16 = reshape(shape = var_2672, x = linear_24_cast_fp16)[name = tensor("qkv_13_cast_fp16")]; + tensor q_37_begin_0 = const()[name = tensor("q_37_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor q_37_end_0 = const()[name = tensor("q_37_end_0"), val = tensor([1, 256, 1, 16, 64])]; + tensor q_37_end_mask_0 = const()[name = tensor("q_37_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor q_37_squeeze_mask_0 = const()[name = tensor("q_37_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor q_37_cast_fp16 = slice_by_index(begin = q_37_begin_0, end = q_37_end_0, end_mask = q_37_end_mask_0, squeeze_mask = q_37_squeeze_mask_0, x = qkv_13_cast_fp16)[name = tensor("q_37_cast_fp16")]; + tensor k_25_begin_0 = const()[name = tensor("k_25_begin_0"), val = tensor([0, 0, 1, 0, 0])]; + tensor k_25_end_0 = const()[name = tensor("k_25_end_0"), val = tensor([1, 256, 2, 16, 64])]; + tensor k_25_end_mask_0 = const()[name = tensor("k_25_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor k_25_squeeze_mask_0 = const()[name = tensor("k_25_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor k_25_cast_fp16 = slice_by_index(begin = k_25_begin_0, end = k_25_end_0, end_mask = k_25_end_mask_0, squeeze_mask = k_25_squeeze_mask_0, x = qkv_13_cast_fp16)[name = tensor("k_25_cast_fp16")]; + tensor v_13_begin_0 = const()[name = tensor("v_13_begin_0"), val = tensor([0, 0, 2, 0, 0])]; + tensor v_13_end_0 = const()[name = tensor("v_13_end_0"), val = tensor([1, 256, 3, 16, 64])]; + tensor v_13_end_mask_0 = const()[name = tensor("v_13_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor v_13_squeeze_mask_0 = const()[name = tensor("v_13_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor v_13_cast_fp16 = slice_by_index(begin = v_13_begin_0, end = v_13_end_0, end_mask = v_13_end_mask_0, squeeze_mask = v_13_squeeze_mask_0, x = qkv_13_cast_fp16)[name = tensor("v_13_cast_fp16")]; + tensor var_2730 = const()[name = tensor("op_2730"), val = tensor([1, 1])]; + tensor position6_to_fp16_dtype_0 = const()[name = tensor("position6_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor position6_to_fp16 = cast(dtype = position6_to_fp16_dtype_0, x = position6)[name = tensor("cast_631")]; + tensor var_2731_cast_fp16 = reshape(shape = var_2730, x = position6_to_fp16)[name = tensor("op_2731_cast_fp16")]; + tensor ts_39_cast_fp16 = add(x = ts_1_promoted_to_fp16, y = var_2731_cast_fp16)[name = tensor("ts_39_cast_fp16")]; + tensor var_2738 = const()[name = tensor("op_2738"), val = tensor([1, 256, 1, 1])]; + tensor ts_41_cast_fp16 = reshape(shape = var_2738, x = ts_39_cast_fp16)[name = tensor("ts_41_cast_fp16")]; + tensor var_2745 = const()[name = tensor("op_2745"), val = tensor([1, 256, 16, 32, 2])]; + tensor q_complex_13_cast_fp16 = reshape(shape = var_2745, x = q_37_cast_fp16)[name = tensor("q_complex_13_cast_fp16")]; + tensor var_2752 = const()[name = tensor("op_2752"), val = tensor([1, 256, 16, 32, 2])]; + tensor k_complex_13_cast_fp16 = reshape(shape = var_2752, x = k_25_cast_fp16)[name = tensor("k_complex_13_cast_fp16")]; + tensor var_2756_begin_0 = const()[name = tensor("op_2756_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_2756_end_0 = const()[name = tensor("op_2756_end_0"), val = tensor([1, 256, 16, 32, 1])]; + tensor var_2756_end_mask_0 = const()[name = tensor("op_2756_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_2756_squeeze_mask_0 = const()[name = tensor("op_2756_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_2756_cast_fp16 = slice_by_index(begin = var_2756_begin_0, end = var_2756_end_0, end_mask = var_2756_end_mask_0, squeeze_mask = var_2756_squeeze_mask_0, x = q_complex_13_cast_fp16)[name = tensor("op_2756_cast_fp16")]; + tensor var_2764_begin_0 = const()[name = tensor("op_2764_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_2764_end_0 = const()[name = tensor("op_2764_end_0"), val = tensor([1, 256, 16, 32, 2])]; + tensor var_2764_end_mask_0 = const()[name = tensor("op_2764_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_2764_squeeze_mask_0 = const()[name = tensor("op_2764_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_2764_cast_fp16 = slice_by_index(begin = var_2764_begin_0, end = var_2764_end_0, end_mask = var_2764_end_mask_0, squeeze_mask = var_2764_squeeze_mask_0, x = q_complex_13_cast_fp16)[name = tensor("op_2764_cast_fp16")]; + tensor var_2772_begin_0 = const()[name = tensor("op_2772_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_2772_end_0 = const()[name = tensor("op_2772_end_0"), val = tensor([1, 256, 16, 32, 1])]; + tensor var_2772_end_mask_0 = const()[name = tensor("op_2772_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_2772_squeeze_mask_0 = const()[name = tensor("op_2772_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_2772_cast_fp16 = slice_by_index(begin = var_2772_begin_0, end = var_2772_end_0, end_mask = var_2772_end_mask_0, squeeze_mask = var_2772_squeeze_mask_0, x = k_complex_13_cast_fp16)[name = tensor("op_2772_cast_fp16")]; + tensor var_2780_begin_0 = const()[name = tensor("op_2780_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_2780_end_0 = const()[name = tensor("op_2780_end_0"), val = tensor([1, 256, 16, 32, 2])]; + tensor var_2780_end_mask_0 = const()[name = tensor("op_2780_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_2780_squeeze_mask_0 = const()[name = tensor("op_2780_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_2780_cast_fp16 = slice_by_index(begin = var_2780_begin_0, end = var_2780_end_0, end_mask = var_2780_end_mask_0, squeeze_mask = var_2780_squeeze_mask_0, x = k_complex_13_cast_fp16)[name = tensor("op_2780_cast_fp16")]; + tensor freqs_13_to_fp16 = const()[name = tensor("freqs_13_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(158411776)))]; + tensor var_2786_cast_fp16 = mul(x = freqs_13_to_fp16, y = ts_41_cast_fp16)[name = tensor("op_2786_cast_fp16")]; + tensor rotr_13_cast_fp16 = cos(x = var_2786_cast_fp16)[name = tensor("rotr_13_cast_fp16")]; + tensor roti_13_cast_fp16 = sin(x = var_2786_cast_fp16)[name = tensor("roti_13_cast_fp16")]; + tensor var_2790_cast_fp16 = mul(x = var_2756_cast_fp16, y = rotr_13_cast_fp16)[name = tensor("op_2790_cast_fp16")]; + tensor var_2791_cast_fp16 = mul(x = var_2764_cast_fp16, y = roti_13_cast_fp16)[name = tensor("op_2791_cast_fp16")]; + tensor qor_25_cast_fp16 = sub(x = var_2790_cast_fp16, y = var_2791_cast_fp16)[name = tensor("qor_25_cast_fp16")]; + tensor var_2794_cast_fp16 = mul(x = var_2756_cast_fp16, y = roti_13_cast_fp16)[name = tensor("op_2794_cast_fp16")]; + tensor var_2795_cast_fp16 = mul(x = var_2764_cast_fp16, y = rotr_13_cast_fp16)[name = tensor("op_2795_cast_fp16")]; + tensor qoi_25_cast_fp16 = add(x = var_2794_cast_fp16, y = var_2795_cast_fp16)[name = tensor("qoi_25_cast_fp16")]; + tensor var_2798_cast_fp16 = mul(x = var_2772_cast_fp16, y = rotr_13_cast_fp16)[name = tensor("op_2798_cast_fp16")]; + tensor var_2799_cast_fp16 = mul(x = var_2780_cast_fp16, y = roti_13_cast_fp16)[name = tensor("op_2799_cast_fp16")]; + tensor kor_25_cast_fp16 = sub(x = var_2798_cast_fp16, y = var_2799_cast_fp16)[name = tensor("kor_25_cast_fp16")]; + tensor var_2802_cast_fp16 = mul(x = var_2772_cast_fp16, y = roti_13_cast_fp16)[name = tensor("op_2802_cast_fp16")]; + tensor var_2803_cast_fp16 = mul(x = var_2780_cast_fp16, y = rotr_13_cast_fp16)[name = tensor("op_2803_cast_fp16")]; + tensor koi_25_cast_fp16 = add(x = var_2802_cast_fp16, y = var_2803_cast_fp16)[name = tensor("koi_25_cast_fp16")]; + tensor qo_13_axis_0 = const()[name = tensor("qo_13_axis_0"), val = tensor(-1)]; + tensor qo_13_cast_fp16 = stack(axis = qo_13_axis_0, values = (qor_25_cast_fp16, qoi_25_cast_fp16))[name = tensor("qo_13_cast_fp16")]; + tensor ko_13_axis_0 = const()[name = tensor("ko_13_axis_0"), val = tensor(-1)]; + tensor ko_13_cast_fp16 = stack(axis = ko_13_axis_0, values = (kor_25_cast_fp16, koi_25_cast_fp16))[name = tensor("ko_13_cast_fp16")]; + tensor var_2836 = const()[name = tensor("op_2836"), val = tensor([1, 256, 16, 64])]; + tensor q_39_cast_fp16 = reshape(shape = var_2836, x = qo_13_cast_fp16)[name = tensor("q_39_cast_fp16")]; + tensor var_2842 = const()[name = tensor("op_2842"), val = tensor([1, 256, 16, 64])]; + tensor k_27_cast_fp16 = reshape(shape = var_2842, x = ko_13_cast_fp16)[name = tensor("k_27_cast_fp16")]; + tensor write_indices_float_37_cast_fp16 = add(x = var_2731_cast_fp16, y = var_393_promoted_to_fp16)[name = tensor("write_indices_float_37_cast_fp16")]; + tensor _inversed_2863_y_0_to_fp16 = const()[name = tensor("_inversed_2863_y_0_to_fp16"), val = tensor(0x1p-9)]; + tensor _inversed_2863_cast_fp16 = mul(x = write_indices_float_37_cast_fp16, y = _inversed_2863_y_0_to_fp16)[name = tensor("_inversed_2863_cast_fp16")]; + tensor var_2864_cast_fp16 = floor(x = _inversed_2863_cast_fp16)[name = tensor("op_2864_cast_fp16")]; + tensor var_2865_to_fp16 = const()[name = tensor("op_2865_to_fp16"), val = tensor(0x1p+9)]; + tensor var_2866_cast_fp16 = mul(x = var_2864_cast_fp16, y = var_2865_to_fp16)[name = tensor("op_2866_cast_fp16")]; + tensor write_indices_float_39_cast_fp16 = sub(x = write_indices_float_37_cast_fp16, y = var_2866_cast_fp16)[name = tensor("write_indices_float_39_cast_fp16")]; + tensor var_2884_value_0_to_fp16 = const()[name = tensor("op_2884_value_0_to_fp16"), val = tensor(0x1.ffp+8)]; + tensor var_2884_cast_fp16 = fill_like(ref_tensor = write_indices_float_39_cast_fp16, value = var_2884_value_0_to_fp16)[name = tensor("op_2884_cast_fp16")]; + tensor write_indices_float_41_cast_fp16 = select(a = write_indices_float_39_cast_fp16, b = var_2884_cast_fp16, cond = token_valid_1_cast_fp16)[name = tensor("write_indices_float_41_cast_fp16")]; + tensor var_2890_dtype_0 = const()[name = tensor("op_2890_dtype_0"), val = tensor("int32")]; + tensor var_2895 = const()[name = tensor("op_2895"), val = tensor([1, 256, 1, 1])]; + tensor write_indices_float_41_cast_fp16_to_int32 = cast(dtype = var_2890_dtype_0, x = write_indices_float_41_cast_fp16)[name = tensor("cast_630")]; + tensor var_2896 = reshape(shape = var_2895, x = write_indices_float_41_cast_fp16_to_int32)[name = tensor("op_2896")]; + tensor write_indices_13_reps_0 = const()[name = tensor("write_indices_13_reps_0"), val = tensor([1, 1, 16, 64])]; + tensor write_indices_13 = tile(reps = write_indices_13_reps_0, x = var_2896)[name = tensor("write_indices_13")]; + tensor var_2906_begin_0 = const()[name = tensor("op_2906_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_2906_end_0 = const()[name = tensor("op_2906_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor var_2906_end_mask_0 = const()[name = tensor("op_2906_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_2906_squeeze_mask_0 = const()[name = tensor("op_2906_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor cache6_to_fp16_dtype_0 = const()[name = tensor("cache6_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor cache6_to_fp16 = cast(dtype = cache6_to_fp16_dtype_0, x = cache6)[name = tensor("cast_629")]; + tensor var_2906_cast_fp16 = slice_by_index(begin = var_2906_begin_0, end = var_2906_end_0, end_mask = var_2906_end_mask_0, squeeze_mask = var_2906_squeeze_mask_0, x = cache6_to_fp16)[name = tensor("op_2906_cast_fp16")]; + tensor var_2908_axis_0 = const()[name = tensor("op_2908_axis_0"), val = tensor(1)]; + tensor var_2908_mode_0 = const()[name = tensor("op_2908_mode_0"), val = tensor("update")]; + tensor var_2908_validate_indices_0 = const()[name = tensor("op_2908_validate_indices_0"), val = tensor(false)]; + tensor var_2908_cast_fp16 = scatter_along_axis(axis = var_2908_axis_0, data = var_2906_cast_fp16, indices = write_indices_13, mode = var_2908_mode_0, updates = k_27_cast_fp16, validate_indices = var_2908_validate_indices_0)[name = tensor("op_2908_cast_fp16")]; + tensor concat_50 = const()[name = tensor("concat_50"), val = tensor([0, 0, 0, 0, 0])]; + tensor concat_51 = const()[name = tensor("concat_51"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_13_internal_tensor_assign_1_stride_0 = const()[name = tensor("new_cache_13_internal_tensor_assign_1_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_13_internal_tensor_assign_1_begin_mask_0 = const()[name = tensor("new_cache_13_internal_tensor_assign_1_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_13_internal_tensor_assign_1_end_mask_0 = const()[name = tensor("new_cache_13_internal_tensor_assign_1_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_13_internal_tensor_assign_1_squeeze_mask_0 = const()[name = tensor("new_cache_13_internal_tensor_assign_1_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_58 = const()[name = tensor("shape_58"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_12 = const()[name = tensor("reduce_prod_12"), val = tensor(1048576)]; + tensor range_1d_12_start_0 = const()[name = tensor("range_1d_12_start_0"), val = tensor(0)]; + tensor range_1d_12_step_0 = const()[name = tensor("range_1d_12_step_0"), val = tensor(1)]; + tensor range_1d_12 = range_1d(end = reduce_prod_12, start = range_1d_12_start_0, step = range_1d_12_step_0)[name = tensor("range_1d_12")]; + tensor reshape_60 = reshape(shape = shape_58, x = range_1d_12)[name = tensor("reshape_60")]; + tensor slice_by_index_12 = slice_by_index(begin = concat_50, begin_mask = new_cache_13_internal_tensor_assign_1_begin_mask_0, end = concat_51, end_mask = new_cache_13_internal_tensor_assign_1_end_mask_0, squeeze_mask = new_cache_13_internal_tensor_assign_1_squeeze_mask_0, stride = new_cache_13_internal_tensor_assign_1_stride_0, x = reshape_60)[name = tensor("slice_by_index_12")]; + tensor reshape_61_shape_0 = const()[name = tensor("reshape_61_shape_0"), val = tensor([-1])]; + tensor reshape_61 = reshape(shape = reshape_61_shape_0, x = slice_by_index_12)[name = tensor("reshape_61")]; + tensor reshape_62_shape_0 = const()[name = tensor("reshape_62_shape_0"), val = tensor([-1])]; + tensor reshape_62_cast_fp16 = reshape(shape = reshape_62_shape_0, x = var_2908_cast_fp16)[name = tensor("reshape_62_cast_fp16")]; + tensor reshape_63_shape_0 = const()[name = tensor("reshape_63_shape_0"), val = tensor([-1])]; + tensor reshape_63_cast_fp16 = reshape(shape = reshape_63_shape_0, x = cache6_to_fp16)[name = tensor("reshape_63_cast_fp16")]; + tensor scatter_12_mode_0 = const()[name = tensor("scatter_12_mode_0"), val = tensor("update")]; + tensor scatter_12_axis_0 = const()[name = tensor("scatter_12_axis_0"), val = tensor(0)]; + tensor scatter_12_validate_indices_0 = const()[name = tensor("scatter_12_validate_indices_0"), val = tensor(false)]; + tensor scatter_12_cast_fp16 = scatter(axis = scatter_12_axis_0, data = reshape_63_cast_fp16, indices = reshape_61, mode = scatter_12_mode_0, updates = reshape_62_cast_fp16, validate_indices = scatter_12_validate_indices_0)[name = tensor("scatter_12_cast_fp16")]; + tensor reshape_64_cast_fp16 = reshape(shape = shape_58, x = scatter_12_cast_fp16)[name = tensor("reshape_64_cast_fp16")]; + tensor var_2916_begin_0 = const()[name = tensor("op_2916_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor var_2916_end_0 = const()[name = tensor("op_2916_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor var_2916_end_mask_0 = const()[name = tensor("op_2916_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_2916_squeeze_mask_0 = const()[name = tensor("op_2916_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor var_2916_cast_fp16 = slice_by_index(begin = var_2916_begin_0, end = var_2916_end_0, end_mask = var_2916_end_mask_0, squeeze_mask = var_2916_squeeze_mask_0, x = reshape_64_cast_fp16)[name = tensor("op_2916_cast_fp16")]; + tensor var_2918_axis_0 = const()[name = tensor("op_2918_axis_0"), val = tensor(1)]; + tensor var_2918_mode_0 = const()[name = tensor("op_2918_mode_0"), val = tensor("update")]; + tensor var_2918_validate_indices_0 = const()[name = tensor("op_2918_validate_indices_0"), val = tensor(false)]; + tensor var_2918_cast_fp16 = scatter_along_axis(axis = var_2918_axis_0, data = var_2916_cast_fp16, indices = write_indices_13, mode = var_2918_mode_0, updates = v_13_cast_fp16, validate_indices = var_2918_validate_indices_0)[name = tensor("op_2918_cast_fp16")]; + tensor concat_52 = const()[name = tensor("concat_52"), val = tensor([1, 0, 0, 0, 0])]; + tensor concat_53 = const()[name = tensor("concat_53"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_13_internal_tensor_assign_2_stride_0 = const()[name = tensor("new_cache_13_internal_tensor_assign_2_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_13_internal_tensor_assign_2_begin_mask_0 = const()[name = tensor("new_cache_13_internal_tensor_assign_2_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_13_internal_tensor_assign_2_end_mask_0 = const()[name = tensor("new_cache_13_internal_tensor_assign_2_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_13_internal_tensor_assign_2_squeeze_mask_0 = const()[name = tensor("new_cache_13_internal_tensor_assign_2_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_59 = const()[name = tensor("shape_59"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_13 = const()[name = tensor("reduce_prod_13"), val = tensor(1048576)]; + tensor range_1d_13_start_0 = const()[name = tensor("range_1d_13_start_0"), val = tensor(0)]; + tensor range_1d_13_step_0 = const()[name = tensor("range_1d_13_step_0"), val = tensor(1)]; + tensor range_1d_13 = range_1d(end = reduce_prod_13, start = range_1d_13_start_0, step = range_1d_13_step_0)[name = tensor("range_1d_13")]; + tensor reshape_65 = reshape(shape = shape_59, x = range_1d_13)[name = tensor("reshape_65")]; + tensor slice_by_index_13 = slice_by_index(begin = concat_52, begin_mask = new_cache_13_internal_tensor_assign_2_begin_mask_0, end = concat_53, end_mask = new_cache_13_internal_tensor_assign_2_end_mask_0, squeeze_mask = new_cache_13_internal_tensor_assign_2_squeeze_mask_0, stride = new_cache_13_internal_tensor_assign_2_stride_0, x = reshape_65)[name = tensor("slice_by_index_13")]; + tensor reshape_66_shape_0 = const()[name = tensor("reshape_66_shape_0"), val = tensor([-1])]; + tensor reshape_66 = reshape(shape = reshape_66_shape_0, x = slice_by_index_13)[name = tensor("reshape_66")]; + tensor reshape_67_shape_0 = const()[name = tensor("reshape_67_shape_0"), val = tensor([-1])]; + tensor reshape_67_cast_fp16 = reshape(shape = reshape_67_shape_0, x = var_2918_cast_fp16)[name = tensor("reshape_67_cast_fp16")]; + tensor reshape_68_shape_0 = const()[name = tensor("reshape_68_shape_0"), val = tensor([-1])]; + tensor reshape_68_cast_fp16 = reshape(shape = reshape_68_shape_0, x = reshape_64_cast_fp16)[name = tensor("reshape_68_cast_fp16")]; + tensor scatter_13_mode_0 = const()[name = tensor("scatter_13_mode_0"), val = tensor("update")]; + tensor scatter_13_axis_0 = const()[name = tensor("scatter_13_axis_0"), val = tensor(0)]; + tensor scatter_13_validate_indices_0 = const()[name = tensor("scatter_13_validate_indices_0"), val = tensor(false)]; + tensor scatter_13_cast_fp16 = scatter(axis = scatter_13_axis_0, data = reshape_68_cast_fp16, indices = reshape_66, mode = scatter_13_mode_0, updates = reshape_67_cast_fp16, validate_indices = scatter_13_validate_indices_0)[name = tensor("scatter_13_cast_fp16")]; + tensor reshape_69_cast_fp16 = reshape(shape = shape_59, x = scatter_13_cast_fp16)[name = tensor("reshape_69_cast_fp16")]; + tensor reshape_69_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("reshape_69_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor keys_37_begin_0 = const()[name = tensor("keys_37_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor keys_37_end_0 = const()[name = tensor("keys_37_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor keys_37_end_mask_0 = const()[name = tensor("keys_37_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor keys_37_squeeze_mask_0 = const()[name = tensor("keys_37_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor keys_37_cast_fp16 = slice_by_index(begin = keys_37_begin_0, end = keys_37_end_0, end_mask = keys_37_end_mask_0, squeeze_mask = keys_37_squeeze_mask_0, x = reshape_69_cast_fp16)[name = tensor("keys_37_cast_fp16")]; + tensor values_37_begin_0 = const()[name = tensor("values_37_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor values_37_end_0 = const()[name = tensor("values_37_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor values_37_end_mask_0 = const()[name = tensor("values_37_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor values_37_squeeze_mask_0 = const()[name = tensor("values_37_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor values_37_cast_fp16 = slice_by_index(begin = values_37_begin_0, end = values_37_end_0, end_mask = values_37_end_mask_0, squeeze_mask = values_37_squeeze_mask_0, x = reshape_69_cast_fp16)[name = tensor("values_37_cast_fp16")]; + tensor var_2930_cast_fp16 = not_equal(x = keys_37_cast_fp16, y = keys_37_cast_fp16)[name = tensor("op_2930_cast_fp16")]; + tensor keys_39_cast_fp16 = select(a = var_470_to_fp16, b = keys_37_cast_fp16, cond = var_2930_cast_fp16)[name = tensor("keys_39_cast_fp16")]; + tensor var_2938_cast_fp16 = not_equal(x = values_37_cast_fp16, y = values_37_cast_fp16)[name = tensor("op_2938_cast_fp16")]; + tensor values_39_cast_fp16 = select(a = var_470_to_fp16, b = values_37_cast_fp16, cond = var_2938_cast_fp16)[name = tensor("values_39_cast_fp16")]; + tensor var_2962 = const()[name = tensor("op_2962"), val = tensor([0, 2, 1, 3])]; + tensor var_2978 = const()[name = tensor("op_2978"), val = tensor([1, 1, 1])]; + tensor var_2979_cast_fp16 = reshape(shape = var_2978, x = position6_to_fp16)[name = tensor("op_2979_cast_fp16")]; + tensor q_positions_13_cast_fp16 = add(x = var_2979_cast_fp16, y = q_offsets_1_promoted_to_fp16)[name = tensor("q_positions_13_cast_fp16")]; + tensor valid_end_13_cast_fp16 = add(x = var_2979_cast_fp16, y = var_536_cast_fp16)[name = tensor("valid_end_13_cast_fp16")]; + tensor valid_mask_13_cast_fp16 = less(x = k_positions_1_promoted_to_fp16, y = valid_end_13_cast_fp16)[name = tensor("valid_mask_13_cast_fp16")]; + tensor causal_mask_13_cast_fp16 = less_equal(x = k_positions_1_promoted_to_fp16, y = q_positions_13_cast_fp16)[name = tensor("causal_mask_13_cast_fp16")]; + tensor var_3007 = logical_and(x = valid_mask_13_cast_fp16, y = causal_mask_13_cast_fp16)[name = tensor("op_3007")]; + tensor attn_mask_13_axes_0 = const()[name = tensor("attn_mask_13_axes_0"), val = tensor([1])]; + tensor attn_mask_13 = expand_dims(axes = attn_mask_13_axes_0, x = var_3007)[name = tensor("attn_mask_13")]; + tensor var_3021_transpose_x_0 = const()[name = tensor("op_3021_transpose_x_0"), val = tensor(false)]; + tensor var_3021_transpose_y_0 = const()[name = tensor("op_3021_transpose_y_0"), val = tensor(false)]; + tensor transpose_81_perm_0 = const()[name = tensor("transpose_81_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_82_perm_0 = const()[name = tensor("transpose_82_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_82 = transpose(perm = transpose_82_perm_0, x = keys_39_cast_fp16)[name = tensor("transpose_180")]; + tensor transpose_81 = transpose(perm = transpose_81_perm_0, x = q_39_cast_fp16)[name = tensor("transpose_181")]; + tensor var_3021_cast_fp16 = matmul(transpose_x = var_3021_transpose_x_0, transpose_y = var_3021_transpose_y_0, x = transpose_81, y = transpose_82)[name = tensor("op_3021_cast_fp16")]; + tensor var_3015_to_fp16 = const()[name = tensor("op_3015_to_fp16"), val = tensor([0x1p-3])]; + tensor attn_weights_37_cast_fp16 = mul(x = var_3021_cast_fp16, y = var_3015_to_fp16)[name = tensor("attn_weights_37_cast_fp16")]; + tensor var_3023 = logical_not(x = attn_mask_13)[name = tensor("op_3023")]; + tensor var_3024_to_fp16 = const()[name = tensor("op_3024_to_fp16"), val = tensor(-inf)]; + tensor attn_weights_39_cast_fp16 = select(a = var_3024_to_fp16, b = attn_weights_37_cast_fp16, cond = var_3023)[name = tensor("attn_weights_39_cast_fp16")]; + tensor var_3026 = const()[name = tensor("op_3026"), val = tensor(-1)]; + tensor attn_weights_41_cast_fp16 = softmax(axis = var_3026, x = attn_weights_39_cast_fp16)[name = tensor("attn_weights_41_cast_fp16")]; + tensor attn_output_13_transpose_x_0 = const()[name = tensor("attn_output_13_transpose_x_0"), val = tensor(false)]; + tensor attn_output_13_transpose_y_0 = const()[name = tensor("attn_output_13_transpose_y_0"), val = tensor(false)]; + tensor values_41_cast_fp16 = transpose(perm = var_2962, x = values_39_cast_fp16)[name = tensor("transpose_182")]; + tensor attn_output_13_cast_fp16 = matmul(transpose_x = attn_output_13_transpose_x_0, transpose_y = attn_output_13_transpose_y_0, x = attn_weights_41_cast_fp16, y = values_41_cast_fp16)[name = tensor("attn_output_13_cast_fp16")]; + tensor var_3034 = const()[name = tensor("op_3034"), val = tensor([0, 2, 1, 3])]; + tensor var_3039 = const()[name = tensor("op_3039"), val = tensor([1, 256, 1024])]; + tensor var_3035_cast_fp16 = transpose(perm = var_3034, x = attn_output_13_cast_fp16)[name = tensor("transpose_179")]; + tensor input_77_cast_fp16 = reshape(shape = var_3039, x = var_3035_cast_fp16)[name = tensor("input_77_cast_fp16")]; + tensor attn6_out_proj_weight_to_fp16 = const()[name = tensor("attn6_out_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(158411904)))]; + tensor linear_25_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = attn6_out_proj_weight_to_fp16, x = input_77_cast_fp16)[name = tensor("linear_25_cast_fp16")]; + tensor var_3048_cast_fp16 = add(x = position6_to_fp16, y = valid_len_to_fp16)[name = tensor("op_3048_cast_fp16")]; + tensor var_3048_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_3048_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor input_79_cast_fp16 = add(x = input_73_cast_fp16, y = linear_25_cast_fp16)[name = tensor("input_79_cast_fp16")]; + tensor input_81_axes_0 = const()[name = tensor("input_81_axes_0"), val = tensor([-1])]; + tensor norm6_2_weight_to_fp16 = const()[name = tensor("norm6_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(160509120)))]; + tensor norm6_2_bias_to_fp16 = const()[name = tensor("norm6_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(160511232)))]; + tensor var_3052_to_fp16 = const()[name = tensor("op_3052_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_81_cast_fp16 = layer_norm(axes = input_81_axes_0, beta = norm6_2_bias_to_fp16, epsilon = var_3052_to_fp16, gamma = norm6_2_weight_to_fp16, x = input_79_cast_fp16)[name = tensor("input_81_cast_fp16")]; + tensor linear6_1_weight_to_fp16 = const()[name = tensor("linear6_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(160513344)))]; + tensor linear_26_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = linear6_1_weight_to_fp16, x = input_81_cast_fp16)[name = tensor("linear_26_cast_fp16")]; + tensor input_83_mode_0 = const()[name = tensor("input_83_mode_0"), val = tensor("EXACT")]; + tensor input_83_cast_fp16 = gelu(mode = input_83_mode_0, x = linear_26_cast_fp16)[name = tensor("input_83_cast_fp16")]; + tensor linear6_2_weight_to_fp16 = const()[name = tensor("linear6_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(168902016)))]; + tensor linear_27_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = linear6_2_weight_to_fp16, x = input_83_cast_fp16)[name = tensor("linear_27_cast_fp16")]; + tensor input_85_cast_fp16 = add(x = input_79_cast_fp16, y = linear_27_cast_fp16)[name = tensor("input_85_cast_fp16")]; + tensor input_87_axes_0 = const()[name = tensor("input_87_axes_0"), val = tensor([-1])]; + tensor norm7_1_weight_to_fp16 = const()[name = tensor("norm7_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(177290688)))]; + tensor norm7_1_bias_to_fp16 = const()[name = tensor("norm7_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(177292800)))]; + tensor var_3069_to_fp16 = const()[name = tensor("op_3069_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_87_cast_fp16 = layer_norm(axes = input_87_axes_0, beta = norm7_1_bias_to_fp16, epsilon = var_3069_to_fp16, gamma = norm7_1_weight_to_fp16, x = input_85_cast_fp16)[name = tensor("input_87_cast_fp16")]; + tensor attn7_in_proj_weight_to_fp16 = const()[name = tensor("attn7_in_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(177294912)))]; + tensor linear_28_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = attn7_in_proj_weight_to_fp16, x = input_87_cast_fp16)[name = tensor("linear_28_cast_fp16")]; + tensor var_3083 = const()[name = tensor("op_3083"), val = tensor([1, 256, 3, 16, 64])]; + tensor qkv_15_cast_fp16 = reshape(shape = var_3083, x = linear_28_cast_fp16)[name = tensor("qkv_15_cast_fp16")]; + tensor q_43_begin_0 = const()[name = tensor("q_43_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor q_43_end_0 = const()[name = tensor("q_43_end_0"), val = tensor([1, 256, 1, 16, 64])]; + tensor q_43_end_mask_0 = const()[name = tensor("q_43_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor q_43_squeeze_mask_0 = const()[name = tensor("q_43_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor q_43_cast_fp16 = slice_by_index(begin = q_43_begin_0, end = q_43_end_0, end_mask = q_43_end_mask_0, squeeze_mask = q_43_squeeze_mask_0, x = qkv_15_cast_fp16)[name = tensor("q_43_cast_fp16")]; + tensor k_29_begin_0 = const()[name = tensor("k_29_begin_0"), val = tensor([0, 0, 1, 0, 0])]; + tensor k_29_end_0 = const()[name = tensor("k_29_end_0"), val = tensor([1, 256, 2, 16, 64])]; + tensor k_29_end_mask_0 = const()[name = tensor("k_29_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor k_29_squeeze_mask_0 = const()[name = tensor("k_29_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor k_29_cast_fp16 = slice_by_index(begin = k_29_begin_0, end = k_29_end_0, end_mask = k_29_end_mask_0, squeeze_mask = k_29_squeeze_mask_0, x = qkv_15_cast_fp16)[name = tensor("k_29_cast_fp16")]; + tensor v_15_begin_0 = const()[name = tensor("v_15_begin_0"), val = tensor([0, 0, 2, 0, 0])]; + tensor v_15_end_0 = const()[name = tensor("v_15_end_0"), val = tensor([1, 256, 3, 16, 64])]; + tensor v_15_end_mask_0 = const()[name = tensor("v_15_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor v_15_squeeze_mask_0 = const()[name = tensor("v_15_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor v_15_cast_fp16 = slice_by_index(begin = v_15_begin_0, end = v_15_end_0, end_mask = v_15_end_mask_0, squeeze_mask = v_15_squeeze_mask_0, x = qkv_15_cast_fp16)[name = tensor("v_15_cast_fp16")]; + tensor var_3141 = const()[name = tensor("op_3141"), val = tensor([1, 1])]; + tensor position7_to_fp16_dtype_0 = const()[name = tensor("position7_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor position7_to_fp16 = cast(dtype = position7_to_fp16_dtype_0, x = position7)[name = tensor("cast_626")]; + tensor var_3142_cast_fp16 = reshape(shape = var_3141, x = position7_to_fp16)[name = tensor("op_3142_cast_fp16")]; + tensor ts_45_cast_fp16 = add(x = ts_1_promoted_to_fp16, y = var_3142_cast_fp16)[name = tensor("ts_45_cast_fp16")]; + tensor var_3149 = const()[name = tensor("op_3149"), val = tensor([1, 256, 1, 1])]; + tensor ts_47_cast_fp16 = reshape(shape = var_3149, x = ts_45_cast_fp16)[name = tensor("ts_47_cast_fp16")]; + tensor var_3156 = const()[name = tensor("op_3156"), val = tensor([1, 256, 16, 32, 2])]; + tensor q_complex_15_cast_fp16 = reshape(shape = var_3156, x = q_43_cast_fp16)[name = tensor("q_complex_15_cast_fp16")]; + tensor var_3163 = const()[name = tensor("op_3163"), val = tensor([1, 256, 16, 32, 2])]; + tensor k_complex_15_cast_fp16 = reshape(shape = var_3163, x = k_29_cast_fp16)[name = tensor("k_complex_15_cast_fp16")]; + tensor var_3167_begin_0 = const()[name = tensor("op_3167_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_3167_end_0 = const()[name = tensor("op_3167_end_0"), val = tensor([1, 256, 16, 32, 1])]; + tensor var_3167_end_mask_0 = const()[name = tensor("op_3167_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_3167_squeeze_mask_0 = const()[name = tensor("op_3167_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_3167_cast_fp16 = slice_by_index(begin = var_3167_begin_0, end = var_3167_end_0, end_mask = var_3167_end_mask_0, squeeze_mask = var_3167_squeeze_mask_0, x = q_complex_15_cast_fp16)[name = tensor("op_3167_cast_fp16")]; + tensor var_3175_begin_0 = const()[name = tensor("op_3175_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_3175_end_0 = const()[name = tensor("op_3175_end_0"), val = tensor([1, 256, 16, 32, 2])]; + tensor var_3175_end_mask_0 = const()[name = tensor("op_3175_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_3175_squeeze_mask_0 = const()[name = tensor("op_3175_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_3175_cast_fp16 = slice_by_index(begin = var_3175_begin_0, end = var_3175_end_0, end_mask = var_3175_end_mask_0, squeeze_mask = var_3175_squeeze_mask_0, x = q_complex_15_cast_fp16)[name = tensor("op_3175_cast_fp16")]; + tensor var_3183_begin_0 = const()[name = tensor("op_3183_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_3183_end_0 = const()[name = tensor("op_3183_end_0"), val = tensor([1, 256, 16, 32, 1])]; + tensor var_3183_end_mask_0 = const()[name = tensor("op_3183_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_3183_squeeze_mask_0 = const()[name = tensor("op_3183_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_3183_cast_fp16 = slice_by_index(begin = var_3183_begin_0, end = var_3183_end_0, end_mask = var_3183_end_mask_0, squeeze_mask = var_3183_squeeze_mask_0, x = k_complex_15_cast_fp16)[name = tensor("op_3183_cast_fp16")]; + tensor var_3191_begin_0 = const()[name = tensor("op_3191_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_3191_end_0 = const()[name = tensor("op_3191_end_0"), val = tensor([1, 256, 16, 32, 2])]; + tensor var_3191_end_mask_0 = const()[name = tensor("op_3191_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_3191_squeeze_mask_0 = const()[name = tensor("op_3191_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_3191_cast_fp16 = slice_by_index(begin = var_3191_begin_0, end = var_3191_end_0, end_mask = var_3191_end_mask_0, squeeze_mask = var_3191_squeeze_mask_0, x = k_complex_15_cast_fp16)[name = tensor("op_3191_cast_fp16")]; + tensor freqs_15_to_fp16 = const()[name = tensor("freqs_15_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(183586432)))]; + tensor var_3197_cast_fp16 = mul(x = freqs_15_to_fp16, y = ts_47_cast_fp16)[name = tensor("op_3197_cast_fp16")]; + tensor rotr_15_cast_fp16 = cos(x = var_3197_cast_fp16)[name = tensor("rotr_15_cast_fp16")]; + tensor roti_15_cast_fp16 = sin(x = var_3197_cast_fp16)[name = tensor("roti_15_cast_fp16")]; + tensor var_3201_cast_fp16 = mul(x = var_3167_cast_fp16, y = rotr_15_cast_fp16)[name = tensor("op_3201_cast_fp16")]; + tensor var_3202_cast_fp16 = mul(x = var_3175_cast_fp16, y = roti_15_cast_fp16)[name = tensor("op_3202_cast_fp16")]; + tensor qor_29_cast_fp16 = sub(x = var_3201_cast_fp16, y = var_3202_cast_fp16)[name = tensor("qor_29_cast_fp16")]; + tensor var_3205_cast_fp16 = mul(x = var_3167_cast_fp16, y = roti_15_cast_fp16)[name = tensor("op_3205_cast_fp16")]; + tensor var_3206_cast_fp16 = mul(x = var_3175_cast_fp16, y = rotr_15_cast_fp16)[name = tensor("op_3206_cast_fp16")]; + tensor qoi_29_cast_fp16 = add(x = var_3205_cast_fp16, y = var_3206_cast_fp16)[name = tensor("qoi_29_cast_fp16")]; + tensor var_3209_cast_fp16 = mul(x = var_3183_cast_fp16, y = rotr_15_cast_fp16)[name = tensor("op_3209_cast_fp16")]; + tensor var_3210_cast_fp16 = mul(x = var_3191_cast_fp16, y = roti_15_cast_fp16)[name = tensor("op_3210_cast_fp16")]; + tensor kor_29_cast_fp16 = sub(x = var_3209_cast_fp16, y = var_3210_cast_fp16)[name = tensor("kor_29_cast_fp16")]; + tensor var_3213_cast_fp16 = mul(x = var_3183_cast_fp16, y = roti_15_cast_fp16)[name = tensor("op_3213_cast_fp16")]; + tensor var_3214_cast_fp16 = mul(x = var_3191_cast_fp16, y = rotr_15_cast_fp16)[name = tensor("op_3214_cast_fp16")]; + tensor koi_29_cast_fp16 = add(x = var_3213_cast_fp16, y = var_3214_cast_fp16)[name = tensor("koi_29_cast_fp16")]; + tensor qo_15_axis_0 = const()[name = tensor("qo_15_axis_0"), val = tensor(-1)]; + tensor qo_15_cast_fp16 = stack(axis = qo_15_axis_0, values = (qor_29_cast_fp16, qoi_29_cast_fp16))[name = tensor("qo_15_cast_fp16")]; + tensor ko_15_axis_0 = const()[name = tensor("ko_15_axis_0"), val = tensor(-1)]; + tensor ko_15_cast_fp16 = stack(axis = ko_15_axis_0, values = (kor_29_cast_fp16, koi_29_cast_fp16))[name = tensor("ko_15_cast_fp16")]; + tensor var_3247 = const()[name = tensor("op_3247"), val = tensor([1, 256, 16, 64])]; + tensor q_45_cast_fp16 = reshape(shape = var_3247, x = qo_15_cast_fp16)[name = tensor("q_45_cast_fp16")]; + tensor var_3253 = const()[name = tensor("op_3253"), val = tensor([1, 256, 16, 64])]; + tensor k_31_cast_fp16 = reshape(shape = var_3253, x = ko_15_cast_fp16)[name = tensor("k_31_cast_fp16")]; + tensor write_indices_float_43_cast_fp16 = add(x = var_3142_cast_fp16, y = var_393_promoted_to_fp16)[name = tensor("write_indices_float_43_cast_fp16")]; + tensor _inversed_3274_y_0_to_fp16 = const()[name = tensor("_inversed_3274_y_0_to_fp16"), val = tensor(0x1p-9)]; + tensor _inversed_3274_cast_fp16 = mul(x = write_indices_float_43_cast_fp16, y = _inversed_3274_y_0_to_fp16)[name = tensor("_inversed_3274_cast_fp16")]; + tensor var_3275_cast_fp16 = floor(x = _inversed_3274_cast_fp16)[name = tensor("op_3275_cast_fp16")]; + tensor var_3276_to_fp16 = const()[name = tensor("op_3276_to_fp16"), val = tensor(0x1p+9)]; + tensor var_3277_cast_fp16 = mul(x = var_3275_cast_fp16, y = var_3276_to_fp16)[name = tensor("op_3277_cast_fp16")]; + tensor write_indices_float_45_cast_fp16 = sub(x = write_indices_float_43_cast_fp16, y = var_3277_cast_fp16)[name = tensor("write_indices_float_45_cast_fp16")]; + tensor var_3295_value_0_to_fp16 = const()[name = tensor("op_3295_value_0_to_fp16"), val = tensor(0x1.ffp+8)]; + tensor var_3295_cast_fp16 = fill_like(ref_tensor = write_indices_float_45_cast_fp16, value = var_3295_value_0_to_fp16)[name = tensor("op_3295_cast_fp16")]; + tensor write_indices_float_47_cast_fp16 = select(a = write_indices_float_45_cast_fp16, b = var_3295_cast_fp16, cond = token_valid_1_cast_fp16)[name = tensor("write_indices_float_47_cast_fp16")]; + tensor var_3301_dtype_0 = const()[name = tensor("op_3301_dtype_0"), val = tensor("int32")]; + tensor var_3306 = const()[name = tensor("op_3306"), val = tensor([1, 256, 1, 1])]; + tensor write_indices_float_47_cast_fp16_to_int32 = cast(dtype = var_3301_dtype_0, x = write_indices_float_47_cast_fp16)[name = tensor("cast_625")]; + tensor var_3307 = reshape(shape = var_3306, x = write_indices_float_47_cast_fp16_to_int32)[name = tensor("op_3307")]; + tensor write_indices_15_reps_0 = const()[name = tensor("write_indices_15_reps_0"), val = tensor([1, 1, 16, 64])]; + tensor write_indices_15 = tile(reps = write_indices_15_reps_0, x = var_3307)[name = tensor("write_indices_15")]; + tensor var_3317_begin_0 = const()[name = tensor("op_3317_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_3317_end_0 = const()[name = tensor("op_3317_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor var_3317_end_mask_0 = const()[name = tensor("op_3317_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_3317_squeeze_mask_0 = const()[name = tensor("op_3317_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor cache7_to_fp16_dtype_0 = const()[name = tensor("cache7_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor cache7_to_fp16 = cast(dtype = cache7_to_fp16_dtype_0, x = cache7)[name = tensor("cast_624")]; + tensor var_3317_cast_fp16 = slice_by_index(begin = var_3317_begin_0, end = var_3317_end_0, end_mask = var_3317_end_mask_0, squeeze_mask = var_3317_squeeze_mask_0, x = cache7_to_fp16)[name = tensor("op_3317_cast_fp16")]; + tensor var_3319_axis_0 = const()[name = tensor("op_3319_axis_0"), val = tensor(1)]; + tensor var_3319_mode_0 = const()[name = tensor("op_3319_mode_0"), val = tensor("update")]; + tensor var_3319_validate_indices_0 = const()[name = tensor("op_3319_validate_indices_0"), val = tensor(false)]; + tensor var_3319_cast_fp16 = scatter_along_axis(axis = var_3319_axis_0, data = var_3317_cast_fp16, indices = write_indices_15, mode = var_3319_mode_0, updates = k_31_cast_fp16, validate_indices = var_3319_validate_indices_0)[name = tensor("op_3319_cast_fp16")]; + tensor concat_58 = const()[name = tensor("concat_58"), val = tensor([0, 0, 0, 0, 0])]; + tensor concat_59 = const()[name = tensor("concat_59"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_15_internal_tensor_assign_1_stride_0 = const()[name = tensor("new_cache_15_internal_tensor_assign_1_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_15_internal_tensor_assign_1_begin_mask_0 = const()[name = tensor("new_cache_15_internal_tensor_assign_1_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_15_internal_tensor_assign_1_end_mask_0 = const()[name = tensor("new_cache_15_internal_tensor_assign_1_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_15_internal_tensor_assign_1_squeeze_mask_0 = const()[name = tensor("new_cache_15_internal_tensor_assign_1_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_60 = const()[name = tensor("shape_60"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_14 = const()[name = tensor("reduce_prod_14"), val = tensor(1048576)]; + tensor range_1d_14_start_0 = const()[name = tensor("range_1d_14_start_0"), val = tensor(0)]; + tensor range_1d_14_step_0 = const()[name = tensor("range_1d_14_step_0"), val = tensor(1)]; + tensor range_1d_14 = range_1d(end = reduce_prod_14, start = range_1d_14_start_0, step = range_1d_14_step_0)[name = tensor("range_1d_14")]; + tensor reshape_70 = reshape(shape = shape_60, x = range_1d_14)[name = tensor("reshape_70")]; + tensor slice_by_index_14 = slice_by_index(begin = concat_58, begin_mask = new_cache_15_internal_tensor_assign_1_begin_mask_0, end = concat_59, end_mask = new_cache_15_internal_tensor_assign_1_end_mask_0, squeeze_mask = new_cache_15_internal_tensor_assign_1_squeeze_mask_0, stride = new_cache_15_internal_tensor_assign_1_stride_0, x = reshape_70)[name = tensor("slice_by_index_14")]; + tensor reshape_71_shape_0 = const()[name = tensor("reshape_71_shape_0"), val = tensor([-1])]; + tensor reshape_71 = reshape(shape = reshape_71_shape_0, x = slice_by_index_14)[name = tensor("reshape_71")]; + tensor reshape_72_shape_0 = const()[name = tensor("reshape_72_shape_0"), val = tensor([-1])]; + tensor reshape_72_cast_fp16 = reshape(shape = reshape_72_shape_0, x = var_3319_cast_fp16)[name = tensor("reshape_72_cast_fp16")]; + tensor reshape_73_shape_0 = const()[name = tensor("reshape_73_shape_0"), val = tensor([-1])]; + tensor reshape_73_cast_fp16 = reshape(shape = reshape_73_shape_0, x = cache7_to_fp16)[name = tensor("reshape_73_cast_fp16")]; + tensor scatter_14_mode_0 = const()[name = tensor("scatter_14_mode_0"), val = tensor("update")]; + tensor scatter_14_axis_0 = const()[name = tensor("scatter_14_axis_0"), val = tensor(0)]; + tensor scatter_14_validate_indices_0 = const()[name = tensor("scatter_14_validate_indices_0"), val = tensor(false)]; + tensor scatter_14_cast_fp16 = scatter(axis = scatter_14_axis_0, data = reshape_73_cast_fp16, indices = reshape_71, mode = scatter_14_mode_0, updates = reshape_72_cast_fp16, validate_indices = scatter_14_validate_indices_0)[name = tensor("scatter_14_cast_fp16")]; + tensor reshape_74_cast_fp16 = reshape(shape = shape_60, x = scatter_14_cast_fp16)[name = tensor("reshape_74_cast_fp16")]; + tensor var_3327_begin_0 = const()[name = tensor("op_3327_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor var_3327_end_0 = const()[name = tensor("op_3327_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor var_3327_end_mask_0 = const()[name = tensor("op_3327_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_3327_squeeze_mask_0 = const()[name = tensor("op_3327_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor var_3327_cast_fp16 = slice_by_index(begin = var_3327_begin_0, end = var_3327_end_0, end_mask = var_3327_end_mask_0, squeeze_mask = var_3327_squeeze_mask_0, x = reshape_74_cast_fp16)[name = tensor("op_3327_cast_fp16")]; + tensor var_3329_axis_0 = const()[name = tensor("op_3329_axis_0"), val = tensor(1)]; + tensor var_3329_mode_0 = const()[name = tensor("op_3329_mode_0"), val = tensor("update")]; + tensor var_3329_validate_indices_0 = const()[name = tensor("op_3329_validate_indices_0"), val = tensor(false)]; + tensor var_3329_cast_fp16 = scatter_along_axis(axis = var_3329_axis_0, data = var_3327_cast_fp16, indices = write_indices_15, mode = var_3329_mode_0, updates = v_15_cast_fp16, validate_indices = var_3329_validate_indices_0)[name = tensor("op_3329_cast_fp16")]; + tensor concat_60 = const()[name = tensor("concat_60"), val = tensor([1, 0, 0, 0, 0])]; + tensor concat_61 = const()[name = tensor("concat_61"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_15_internal_tensor_assign_2_stride_0 = const()[name = tensor("new_cache_15_internal_tensor_assign_2_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_15_internal_tensor_assign_2_begin_mask_0 = const()[name = tensor("new_cache_15_internal_tensor_assign_2_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_15_internal_tensor_assign_2_end_mask_0 = const()[name = tensor("new_cache_15_internal_tensor_assign_2_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_15_internal_tensor_assign_2_squeeze_mask_0 = const()[name = tensor("new_cache_15_internal_tensor_assign_2_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_61 = const()[name = tensor("shape_61"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_15 = const()[name = tensor("reduce_prod_15"), val = tensor(1048576)]; + tensor range_1d_15_start_0 = const()[name = tensor("range_1d_15_start_0"), val = tensor(0)]; + tensor range_1d_15_step_0 = const()[name = tensor("range_1d_15_step_0"), val = tensor(1)]; + tensor range_1d_15 = range_1d(end = reduce_prod_15, start = range_1d_15_start_0, step = range_1d_15_step_0)[name = tensor("range_1d_15")]; + tensor reshape_75 = reshape(shape = shape_61, x = range_1d_15)[name = tensor("reshape_75")]; + tensor slice_by_index_15 = slice_by_index(begin = concat_60, begin_mask = new_cache_15_internal_tensor_assign_2_begin_mask_0, end = concat_61, end_mask = new_cache_15_internal_tensor_assign_2_end_mask_0, squeeze_mask = new_cache_15_internal_tensor_assign_2_squeeze_mask_0, stride = new_cache_15_internal_tensor_assign_2_stride_0, x = reshape_75)[name = tensor("slice_by_index_15")]; + tensor reshape_76_shape_0 = const()[name = tensor("reshape_76_shape_0"), val = tensor([-1])]; + tensor reshape_76 = reshape(shape = reshape_76_shape_0, x = slice_by_index_15)[name = tensor("reshape_76")]; + tensor reshape_77_shape_0 = const()[name = tensor("reshape_77_shape_0"), val = tensor([-1])]; + tensor reshape_77_cast_fp16 = reshape(shape = reshape_77_shape_0, x = var_3329_cast_fp16)[name = tensor("reshape_77_cast_fp16")]; + tensor reshape_78_shape_0 = const()[name = tensor("reshape_78_shape_0"), val = tensor([-1])]; + tensor reshape_78_cast_fp16 = reshape(shape = reshape_78_shape_0, x = reshape_74_cast_fp16)[name = tensor("reshape_78_cast_fp16")]; + tensor scatter_15_mode_0 = const()[name = tensor("scatter_15_mode_0"), val = tensor("update")]; + tensor scatter_15_axis_0 = const()[name = tensor("scatter_15_axis_0"), val = tensor(0)]; + tensor scatter_15_validate_indices_0 = const()[name = tensor("scatter_15_validate_indices_0"), val = tensor(false)]; + tensor scatter_15_cast_fp16 = scatter(axis = scatter_15_axis_0, data = reshape_78_cast_fp16, indices = reshape_76, mode = scatter_15_mode_0, updates = reshape_77_cast_fp16, validate_indices = scatter_15_validate_indices_0)[name = tensor("scatter_15_cast_fp16")]; + tensor reshape_79_cast_fp16 = reshape(shape = shape_61, x = scatter_15_cast_fp16)[name = tensor("reshape_79_cast_fp16")]; + tensor reshape_79_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("reshape_79_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor keys_43_begin_0 = const()[name = tensor("keys_43_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor keys_43_end_0 = const()[name = tensor("keys_43_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor keys_43_end_mask_0 = const()[name = tensor("keys_43_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor keys_43_squeeze_mask_0 = const()[name = tensor("keys_43_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor keys_43_cast_fp16 = slice_by_index(begin = keys_43_begin_0, end = keys_43_end_0, end_mask = keys_43_end_mask_0, squeeze_mask = keys_43_squeeze_mask_0, x = reshape_79_cast_fp16)[name = tensor("keys_43_cast_fp16")]; + tensor values_43_begin_0 = const()[name = tensor("values_43_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor values_43_end_0 = const()[name = tensor("values_43_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor values_43_end_mask_0 = const()[name = tensor("values_43_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor values_43_squeeze_mask_0 = const()[name = tensor("values_43_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor values_43_cast_fp16 = slice_by_index(begin = values_43_begin_0, end = values_43_end_0, end_mask = values_43_end_mask_0, squeeze_mask = values_43_squeeze_mask_0, x = reshape_79_cast_fp16)[name = tensor("values_43_cast_fp16")]; + tensor var_3341_cast_fp16 = not_equal(x = keys_43_cast_fp16, y = keys_43_cast_fp16)[name = tensor("op_3341_cast_fp16")]; + tensor keys_45_cast_fp16 = select(a = var_470_to_fp16, b = keys_43_cast_fp16, cond = var_3341_cast_fp16)[name = tensor("keys_45_cast_fp16")]; + tensor var_3349_cast_fp16 = not_equal(x = values_43_cast_fp16, y = values_43_cast_fp16)[name = tensor("op_3349_cast_fp16")]; + tensor values_45_cast_fp16 = select(a = var_470_to_fp16, b = values_43_cast_fp16, cond = var_3349_cast_fp16)[name = tensor("values_45_cast_fp16")]; + tensor var_3373 = const()[name = tensor("op_3373"), val = tensor([0, 2, 1, 3])]; + tensor var_3389 = const()[name = tensor("op_3389"), val = tensor([1, 1, 1])]; + tensor var_3390_cast_fp16 = reshape(shape = var_3389, x = position7_to_fp16)[name = tensor("op_3390_cast_fp16")]; + tensor q_positions_15_cast_fp16 = add(x = var_3390_cast_fp16, y = q_offsets_1_promoted_to_fp16)[name = tensor("q_positions_15_cast_fp16")]; + tensor valid_end_15_cast_fp16 = add(x = var_3390_cast_fp16, y = var_536_cast_fp16)[name = tensor("valid_end_15_cast_fp16")]; + tensor valid_mask_15_cast_fp16 = less(x = k_positions_1_promoted_to_fp16, y = valid_end_15_cast_fp16)[name = tensor("valid_mask_15_cast_fp16")]; + tensor causal_mask_15_cast_fp16 = less_equal(x = k_positions_1_promoted_to_fp16, y = q_positions_15_cast_fp16)[name = tensor("causal_mask_15_cast_fp16")]; + tensor var_3418 = logical_and(x = valid_mask_15_cast_fp16, y = causal_mask_15_cast_fp16)[name = tensor("op_3418")]; + tensor attn_mask_15_axes_0 = const()[name = tensor("attn_mask_15_axes_0"), val = tensor([1])]; + tensor attn_mask_15 = expand_dims(axes = attn_mask_15_axes_0, x = var_3418)[name = tensor("attn_mask_15")]; + tensor var_3432_transpose_x_0 = const()[name = tensor("op_3432_transpose_x_0"), val = tensor(false)]; + tensor var_3432_transpose_y_0 = const()[name = tensor("op_3432_transpose_y_0"), val = tensor(false)]; + tensor transpose_83_perm_0 = const()[name = tensor("transpose_83_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_84_perm_0 = const()[name = tensor("transpose_84_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_84 = transpose(perm = transpose_84_perm_0, x = keys_45_cast_fp16)[name = tensor("transpose_176")]; + tensor transpose_83 = transpose(perm = transpose_83_perm_0, x = q_45_cast_fp16)[name = tensor("transpose_177")]; + tensor var_3432_cast_fp16 = matmul(transpose_x = var_3432_transpose_x_0, transpose_y = var_3432_transpose_y_0, x = transpose_83, y = transpose_84)[name = tensor("op_3432_cast_fp16")]; + tensor var_3426_to_fp16 = const()[name = tensor("op_3426_to_fp16"), val = tensor([0x1p-3])]; + tensor attn_weights_43_cast_fp16 = mul(x = var_3432_cast_fp16, y = var_3426_to_fp16)[name = tensor("attn_weights_43_cast_fp16")]; + tensor var_3434 = logical_not(x = attn_mask_15)[name = tensor("op_3434")]; + tensor var_3435_to_fp16 = const()[name = tensor("op_3435_to_fp16"), val = tensor(-inf)]; + tensor attn_weights_45_cast_fp16 = select(a = var_3435_to_fp16, b = attn_weights_43_cast_fp16, cond = var_3434)[name = tensor("attn_weights_45_cast_fp16")]; + tensor var_3437 = const()[name = tensor("op_3437"), val = tensor(-1)]; + tensor attn_weights_47_cast_fp16 = softmax(axis = var_3437, x = attn_weights_45_cast_fp16)[name = tensor("attn_weights_47_cast_fp16")]; + tensor attn_output_15_transpose_x_0 = const()[name = tensor("attn_output_15_transpose_x_0"), val = tensor(false)]; + tensor attn_output_15_transpose_y_0 = const()[name = tensor("attn_output_15_transpose_y_0"), val = tensor(false)]; + tensor values_47_cast_fp16 = transpose(perm = var_3373, x = values_45_cast_fp16)[name = tensor("transpose_178")]; + tensor attn_output_15_cast_fp16 = matmul(transpose_x = attn_output_15_transpose_x_0, transpose_y = attn_output_15_transpose_y_0, x = attn_weights_47_cast_fp16, y = values_47_cast_fp16)[name = tensor("attn_output_15_cast_fp16")]; + tensor var_3445 = const()[name = tensor("op_3445"), val = tensor([0, 2, 1, 3])]; + tensor var_3450 = const()[name = tensor("op_3450"), val = tensor([1, 256, 1024])]; + tensor var_3446_cast_fp16 = transpose(perm = var_3445, x = attn_output_15_cast_fp16)[name = tensor("transpose_175")]; + tensor input_89_cast_fp16 = reshape(shape = var_3450, x = var_3446_cast_fp16)[name = tensor("input_89_cast_fp16")]; + tensor attn7_out_proj_weight_to_fp16 = const()[name = tensor("attn7_out_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(183586560)))]; + tensor linear_29_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = attn7_out_proj_weight_to_fp16, x = input_89_cast_fp16)[name = tensor("linear_29_cast_fp16")]; + tensor var_3459_cast_fp16 = add(x = position7_to_fp16, y = valid_len_to_fp16)[name = tensor("op_3459_cast_fp16")]; + tensor var_3459_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_3459_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor input_91_cast_fp16 = add(x = input_85_cast_fp16, y = linear_29_cast_fp16)[name = tensor("input_91_cast_fp16")]; + tensor input_93_axes_0 = const()[name = tensor("input_93_axes_0"), val = tensor([-1])]; + tensor norm7_2_weight_to_fp16 = const()[name = tensor("norm7_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(185683776)))]; + tensor norm7_2_bias_to_fp16 = const()[name = tensor("norm7_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(185685888)))]; + tensor var_3463_to_fp16 = const()[name = tensor("op_3463_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_93_cast_fp16 = layer_norm(axes = input_93_axes_0, beta = norm7_2_bias_to_fp16, epsilon = var_3463_to_fp16, gamma = norm7_2_weight_to_fp16, x = input_91_cast_fp16)[name = tensor("input_93_cast_fp16")]; + tensor linear7_1_weight_to_fp16 = const()[name = tensor("linear7_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(185688000)))]; + tensor linear_30_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = linear7_1_weight_to_fp16, x = input_93_cast_fp16)[name = tensor("linear_30_cast_fp16")]; + tensor input_95_mode_0 = const()[name = tensor("input_95_mode_0"), val = tensor("EXACT")]; + tensor input_95_cast_fp16 = gelu(mode = input_95_mode_0, x = linear_30_cast_fp16)[name = tensor("input_95_cast_fp16")]; + tensor linear7_2_weight_to_fp16 = const()[name = tensor("linear7_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(194076672)))]; + tensor linear_31_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = linear7_2_weight_to_fp16, x = input_95_cast_fp16)[name = tensor("linear_31_cast_fp16")]; + tensor input_97_cast_fp16 = add(x = input_91_cast_fp16, y = linear_31_cast_fp16)[name = tensor("input_97_cast_fp16")]; + tensor input_99_axes_0 = const()[name = tensor("input_99_axes_0"), val = tensor([-1])]; + tensor norm8_1_weight_to_fp16 = const()[name = tensor("norm8_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(202465344)))]; + tensor norm8_1_bias_to_fp16 = const()[name = tensor("norm8_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(202467456)))]; + tensor var_3480_to_fp16 = const()[name = tensor("op_3480_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_99_cast_fp16 = layer_norm(axes = input_99_axes_0, beta = norm8_1_bias_to_fp16, epsilon = var_3480_to_fp16, gamma = norm8_1_weight_to_fp16, x = input_97_cast_fp16)[name = tensor("input_99_cast_fp16")]; + tensor attn8_in_proj_weight_to_fp16 = const()[name = tensor("attn8_in_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(202469568)))]; + tensor linear_32_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = attn8_in_proj_weight_to_fp16, x = input_99_cast_fp16)[name = tensor("linear_32_cast_fp16")]; + tensor var_3494 = const()[name = tensor("op_3494"), val = tensor([1, 256, 3, 16, 64])]; + tensor qkv_17_cast_fp16 = reshape(shape = var_3494, x = linear_32_cast_fp16)[name = tensor("qkv_17_cast_fp16")]; + tensor q_49_begin_0 = const()[name = tensor("q_49_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor q_49_end_0 = const()[name = tensor("q_49_end_0"), val = tensor([1, 256, 1, 16, 64])]; + tensor q_49_end_mask_0 = const()[name = tensor("q_49_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor q_49_squeeze_mask_0 = const()[name = tensor("q_49_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor q_49_cast_fp16 = slice_by_index(begin = q_49_begin_0, end = q_49_end_0, end_mask = q_49_end_mask_0, squeeze_mask = q_49_squeeze_mask_0, x = qkv_17_cast_fp16)[name = tensor("q_49_cast_fp16")]; + tensor k_33_begin_0 = const()[name = tensor("k_33_begin_0"), val = tensor([0, 0, 1, 0, 0])]; + tensor k_33_end_0 = const()[name = tensor("k_33_end_0"), val = tensor([1, 256, 2, 16, 64])]; + tensor k_33_end_mask_0 = const()[name = tensor("k_33_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor k_33_squeeze_mask_0 = const()[name = tensor("k_33_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor k_33_cast_fp16 = slice_by_index(begin = k_33_begin_0, end = k_33_end_0, end_mask = k_33_end_mask_0, squeeze_mask = k_33_squeeze_mask_0, x = qkv_17_cast_fp16)[name = tensor("k_33_cast_fp16")]; + tensor v_17_begin_0 = const()[name = tensor("v_17_begin_0"), val = tensor([0, 0, 2, 0, 0])]; + tensor v_17_end_0 = const()[name = tensor("v_17_end_0"), val = tensor([1, 256, 3, 16, 64])]; + tensor v_17_end_mask_0 = const()[name = tensor("v_17_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor v_17_squeeze_mask_0 = const()[name = tensor("v_17_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor v_17_cast_fp16 = slice_by_index(begin = v_17_begin_0, end = v_17_end_0, end_mask = v_17_end_mask_0, squeeze_mask = v_17_squeeze_mask_0, x = qkv_17_cast_fp16)[name = tensor("v_17_cast_fp16")]; + tensor var_3552 = const()[name = tensor("op_3552"), val = tensor([1, 1])]; + tensor position8_to_fp16_dtype_0 = const()[name = tensor("position8_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor position8_to_fp16 = cast(dtype = position8_to_fp16_dtype_0, x = position8)[name = tensor("cast_621")]; + tensor var_3553_cast_fp16 = reshape(shape = var_3552, x = position8_to_fp16)[name = tensor("op_3553_cast_fp16")]; + tensor ts_51_cast_fp16 = add(x = ts_1_promoted_to_fp16, y = var_3553_cast_fp16)[name = tensor("ts_51_cast_fp16")]; + tensor var_3560 = const()[name = tensor("op_3560"), val = tensor([1, 256, 1, 1])]; + tensor ts_53_cast_fp16 = reshape(shape = var_3560, x = ts_51_cast_fp16)[name = tensor("ts_53_cast_fp16")]; + tensor var_3567 = const()[name = tensor("op_3567"), val = tensor([1, 256, 16, 32, 2])]; + tensor q_complex_17_cast_fp16 = reshape(shape = var_3567, x = q_49_cast_fp16)[name = tensor("q_complex_17_cast_fp16")]; + tensor var_3574 = const()[name = tensor("op_3574"), val = tensor([1, 256, 16, 32, 2])]; + tensor k_complex_17_cast_fp16 = reshape(shape = var_3574, x = k_33_cast_fp16)[name = tensor("k_complex_17_cast_fp16")]; + tensor var_3578_begin_0 = const()[name = tensor("op_3578_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_3578_end_0 = const()[name = tensor("op_3578_end_0"), val = tensor([1, 256, 16, 32, 1])]; + tensor var_3578_end_mask_0 = const()[name = tensor("op_3578_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_3578_squeeze_mask_0 = const()[name = tensor("op_3578_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_3578_cast_fp16 = slice_by_index(begin = var_3578_begin_0, end = var_3578_end_0, end_mask = var_3578_end_mask_0, squeeze_mask = var_3578_squeeze_mask_0, x = q_complex_17_cast_fp16)[name = tensor("op_3578_cast_fp16")]; + tensor var_3586_begin_0 = const()[name = tensor("op_3586_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_3586_end_0 = const()[name = tensor("op_3586_end_0"), val = tensor([1, 256, 16, 32, 2])]; + tensor var_3586_end_mask_0 = const()[name = tensor("op_3586_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_3586_squeeze_mask_0 = const()[name = tensor("op_3586_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_3586_cast_fp16 = slice_by_index(begin = var_3586_begin_0, end = var_3586_end_0, end_mask = var_3586_end_mask_0, squeeze_mask = var_3586_squeeze_mask_0, x = q_complex_17_cast_fp16)[name = tensor("op_3586_cast_fp16")]; + tensor var_3594_begin_0 = const()[name = tensor("op_3594_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_3594_end_0 = const()[name = tensor("op_3594_end_0"), val = tensor([1, 256, 16, 32, 1])]; + tensor var_3594_end_mask_0 = const()[name = tensor("op_3594_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_3594_squeeze_mask_0 = const()[name = tensor("op_3594_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_3594_cast_fp16 = slice_by_index(begin = var_3594_begin_0, end = var_3594_end_0, end_mask = var_3594_end_mask_0, squeeze_mask = var_3594_squeeze_mask_0, x = k_complex_17_cast_fp16)[name = tensor("op_3594_cast_fp16")]; + tensor var_3602_begin_0 = const()[name = tensor("op_3602_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_3602_end_0 = const()[name = tensor("op_3602_end_0"), val = tensor([1, 256, 16, 32, 2])]; + tensor var_3602_end_mask_0 = const()[name = tensor("op_3602_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_3602_squeeze_mask_0 = const()[name = tensor("op_3602_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_3602_cast_fp16 = slice_by_index(begin = var_3602_begin_0, end = var_3602_end_0, end_mask = var_3602_end_mask_0, squeeze_mask = var_3602_squeeze_mask_0, x = k_complex_17_cast_fp16)[name = tensor("op_3602_cast_fp16")]; + tensor freqs_17_to_fp16 = const()[name = tensor("freqs_17_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(208761088)))]; + tensor var_3608_cast_fp16 = mul(x = freqs_17_to_fp16, y = ts_53_cast_fp16)[name = tensor("op_3608_cast_fp16")]; + tensor rotr_17_cast_fp16 = cos(x = var_3608_cast_fp16)[name = tensor("rotr_17_cast_fp16")]; + tensor roti_17_cast_fp16 = sin(x = var_3608_cast_fp16)[name = tensor("roti_17_cast_fp16")]; + tensor var_3612_cast_fp16 = mul(x = var_3578_cast_fp16, y = rotr_17_cast_fp16)[name = tensor("op_3612_cast_fp16")]; + tensor var_3613_cast_fp16 = mul(x = var_3586_cast_fp16, y = roti_17_cast_fp16)[name = tensor("op_3613_cast_fp16")]; + tensor qor_33_cast_fp16 = sub(x = var_3612_cast_fp16, y = var_3613_cast_fp16)[name = tensor("qor_33_cast_fp16")]; + tensor var_3616_cast_fp16 = mul(x = var_3578_cast_fp16, y = roti_17_cast_fp16)[name = tensor("op_3616_cast_fp16")]; + tensor var_3617_cast_fp16 = mul(x = var_3586_cast_fp16, y = rotr_17_cast_fp16)[name = tensor("op_3617_cast_fp16")]; + tensor qoi_33_cast_fp16 = add(x = var_3616_cast_fp16, y = var_3617_cast_fp16)[name = tensor("qoi_33_cast_fp16")]; + tensor var_3620_cast_fp16 = mul(x = var_3594_cast_fp16, y = rotr_17_cast_fp16)[name = tensor("op_3620_cast_fp16")]; + tensor var_3621_cast_fp16 = mul(x = var_3602_cast_fp16, y = roti_17_cast_fp16)[name = tensor("op_3621_cast_fp16")]; + tensor kor_33_cast_fp16 = sub(x = var_3620_cast_fp16, y = var_3621_cast_fp16)[name = tensor("kor_33_cast_fp16")]; + tensor var_3624_cast_fp16 = mul(x = var_3594_cast_fp16, y = roti_17_cast_fp16)[name = tensor("op_3624_cast_fp16")]; + tensor var_3625_cast_fp16 = mul(x = var_3602_cast_fp16, y = rotr_17_cast_fp16)[name = tensor("op_3625_cast_fp16")]; + tensor koi_33_cast_fp16 = add(x = var_3624_cast_fp16, y = var_3625_cast_fp16)[name = tensor("koi_33_cast_fp16")]; + tensor qo_17_axis_0 = const()[name = tensor("qo_17_axis_0"), val = tensor(-1)]; + tensor qo_17_cast_fp16 = stack(axis = qo_17_axis_0, values = (qor_33_cast_fp16, qoi_33_cast_fp16))[name = tensor("qo_17_cast_fp16")]; + tensor ko_17_axis_0 = const()[name = tensor("ko_17_axis_0"), val = tensor(-1)]; + tensor ko_17_cast_fp16 = stack(axis = ko_17_axis_0, values = (kor_33_cast_fp16, koi_33_cast_fp16))[name = tensor("ko_17_cast_fp16")]; + tensor var_3658 = const()[name = tensor("op_3658"), val = tensor([1, 256, 16, 64])]; + tensor q_51_cast_fp16 = reshape(shape = var_3658, x = qo_17_cast_fp16)[name = tensor("q_51_cast_fp16")]; + tensor var_3664 = const()[name = tensor("op_3664"), val = tensor([1, 256, 16, 64])]; + tensor k_35_cast_fp16 = reshape(shape = var_3664, x = ko_17_cast_fp16)[name = tensor("k_35_cast_fp16")]; + tensor write_indices_float_49_cast_fp16 = add(x = var_3553_cast_fp16, y = var_393_promoted_to_fp16)[name = tensor("write_indices_float_49_cast_fp16")]; + tensor _inversed_3685_y_0_to_fp16 = const()[name = tensor("_inversed_3685_y_0_to_fp16"), val = tensor(0x1p-9)]; + tensor _inversed_3685_cast_fp16 = mul(x = write_indices_float_49_cast_fp16, y = _inversed_3685_y_0_to_fp16)[name = tensor("_inversed_3685_cast_fp16")]; + tensor var_3686_cast_fp16 = floor(x = _inversed_3685_cast_fp16)[name = tensor("op_3686_cast_fp16")]; + tensor var_3687_to_fp16 = const()[name = tensor("op_3687_to_fp16"), val = tensor(0x1p+9)]; + tensor var_3688_cast_fp16 = mul(x = var_3686_cast_fp16, y = var_3687_to_fp16)[name = tensor("op_3688_cast_fp16")]; + tensor write_indices_float_51_cast_fp16 = sub(x = write_indices_float_49_cast_fp16, y = var_3688_cast_fp16)[name = tensor("write_indices_float_51_cast_fp16")]; + tensor var_3706_value_0_to_fp16 = const()[name = tensor("op_3706_value_0_to_fp16"), val = tensor(0x1.ffp+8)]; + tensor var_3706_cast_fp16 = fill_like(ref_tensor = write_indices_float_51_cast_fp16, value = var_3706_value_0_to_fp16)[name = tensor("op_3706_cast_fp16")]; + tensor write_indices_float_53_cast_fp16 = select(a = write_indices_float_51_cast_fp16, b = var_3706_cast_fp16, cond = token_valid_1_cast_fp16)[name = tensor("write_indices_float_53_cast_fp16")]; + tensor var_3712_dtype_0 = const()[name = tensor("op_3712_dtype_0"), val = tensor("int32")]; + tensor var_3717 = const()[name = tensor("op_3717"), val = tensor([1, 256, 1, 1])]; + tensor write_indices_float_53_cast_fp16_to_int32 = cast(dtype = var_3712_dtype_0, x = write_indices_float_53_cast_fp16)[name = tensor("cast_620")]; + tensor var_3718 = reshape(shape = var_3717, x = write_indices_float_53_cast_fp16_to_int32)[name = tensor("op_3718")]; + tensor write_indices_17_reps_0 = const()[name = tensor("write_indices_17_reps_0"), val = tensor([1, 1, 16, 64])]; + tensor write_indices_17 = tile(reps = write_indices_17_reps_0, x = var_3718)[name = tensor("write_indices_17")]; + tensor var_3728_begin_0 = const()[name = tensor("op_3728_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_3728_end_0 = const()[name = tensor("op_3728_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor var_3728_end_mask_0 = const()[name = tensor("op_3728_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_3728_squeeze_mask_0 = const()[name = tensor("op_3728_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor cache8_to_fp16_dtype_0 = const()[name = tensor("cache8_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor cache8_to_fp16 = cast(dtype = cache8_to_fp16_dtype_0, x = cache8)[name = tensor("cast_619")]; + tensor var_3728_cast_fp16 = slice_by_index(begin = var_3728_begin_0, end = var_3728_end_0, end_mask = var_3728_end_mask_0, squeeze_mask = var_3728_squeeze_mask_0, x = cache8_to_fp16)[name = tensor("op_3728_cast_fp16")]; + tensor var_3730_axis_0 = const()[name = tensor("op_3730_axis_0"), val = tensor(1)]; + tensor var_3730_mode_0 = const()[name = tensor("op_3730_mode_0"), val = tensor("update")]; + tensor var_3730_validate_indices_0 = const()[name = tensor("op_3730_validate_indices_0"), val = tensor(false)]; + tensor var_3730_cast_fp16 = scatter_along_axis(axis = var_3730_axis_0, data = var_3728_cast_fp16, indices = write_indices_17, mode = var_3730_mode_0, updates = k_35_cast_fp16, validate_indices = var_3730_validate_indices_0)[name = tensor("op_3730_cast_fp16")]; + tensor concat_66 = const()[name = tensor("concat_66"), val = tensor([0, 0, 0, 0, 0])]; + tensor concat_67 = const()[name = tensor("concat_67"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_17_internal_tensor_assign_1_stride_0 = const()[name = tensor("new_cache_17_internal_tensor_assign_1_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_17_internal_tensor_assign_1_begin_mask_0 = const()[name = tensor("new_cache_17_internal_tensor_assign_1_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_17_internal_tensor_assign_1_end_mask_0 = const()[name = tensor("new_cache_17_internal_tensor_assign_1_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_17_internal_tensor_assign_1_squeeze_mask_0 = const()[name = tensor("new_cache_17_internal_tensor_assign_1_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_62 = const()[name = tensor("shape_62"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_16 = const()[name = tensor("reduce_prod_16"), val = tensor(1048576)]; + tensor range_1d_16_start_0 = const()[name = tensor("range_1d_16_start_0"), val = tensor(0)]; + tensor range_1d_16_step_0 = const()[name = tensor("range_1d_16_step_0"), val = tensor(1)]; + tensor range_1d_16 = range_1d(end = reduce_prod_16, start = range_1d_16_start_0, step = range_1d_16_step_0)[name = tensor("range_1d_16")]; + tensor reshape_80 = reshape(shape = shape_62, x = range_1d_16)[name = tensor("reshape_80")]; + tensor slice_by_index_16 = slice_by_index(begin = concat_66, begin_mask = new_cache_17_internal_tensor_assign_1_begin_mask_0, end = concat_67, end_mask = new_cache_17_internal_tensor_assign_1_end_mask_0, squeeze_mask = new_cache_17_internal_tensor_assign_1_squeeze_mask_0, stride = new_cache_17_internal_tensor_assign_1_stride_0, x = reshape_80)[name = tensor("slice_by_index_16")]; + tensor reshape_81_shape_0 = const()[name = tensor("reshape_81_shape_0"), val = tensor([-1])]; + tensor reshape_81 = reshape(shape = reshape_81_shape_0, x = slice_by_index_16)[name = tensor("reshape_81")]; + tensor reshape_82_shape_0 = const()[name = tensor("reshape_82_shape_0"), val = tensor([-1])]; + tensor reshape_82_cast_fp16 = reshape(shape = reshape_82_shape_0, x = var_3730_cast_fp16)[name = tensor("reshape_82_cast_fp16")]; + tensor reshape_83_shape_0 = const()[name = tensor("reshape_83_shape_0"), val = tensor([-1])]; + tensor reshape_83_cast_fp16 = reshape(shape = reshape_83_shape_0, x = cache8_to_fp16)[name = tensor("reshape_83_cast_fp16")]; + tensor scatter_16_mode_0 = const()[name = tensor("scatter_16_mode_0"), val = tensor("update")]; + tensor scatter_16_axis_0 = const()[name = tensor("scatter_16_axis_0"), val = tensor(0)]; + tensor scatter_16_validate_indices_0 = const()[name = tensor("scatter_16_validate_indices_0"), val = tensor(false)]; + tensor scatter_16_cast_fp16 = scatter(axis = scatter_16_axis_0, data = reshape_83_cast_fp16, indices = reshape_81, mode = scatter_16_mode_0, updates = reshape_82_cast_fp16, validate_indices = scatter_16_validate_indices_0)[name = tensor("scatter_16_cast_fp16")]; + tensor reshape_84_cast_fp16 = reshape(shape = shape_62, x = scatter_16_cast_fp16)[name = tensor("reshape_84_cast_fp16")]; + tensor var_3738_begin_0 = const()[name = tensor("op_3738_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor var_3738_end_0 = const()[name = tensor("op_3738_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor var_3738_end_mask_0 = const()[name = tensor("op_3738_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_3738_squeeze_mask_0 = const()[name = tensor("op_3738_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor var_3738_cast_fp16 = slice_by_index(begin = var_3738_begin_0, end = var_3738_end_0, end_mask = var_3738_end_mask_0, squeeze_mask = var_3738_squeeze_mask_0, x = reshape_84_cast_fp16)[name = tensor("op_3738_cast_fp16")]; + tensor var_3740_axis_0 = const()[name = tensor("op_3740_axis_0"), val = tensor(1)]; + tensor var_3740_mode_0 = const()[name = tensor("op_3740_mode_0"), val = tensor("update")]; + tensor var_3740_validate_indices_0 = const()[name = tensor("op_3740_validate_indices_0"), val = tensor(false)]; + tensor var_3740_cast_fp16 = scatter_along_axis(axis = var_3740_axis_0, data = var_3738_cast_fp16, indices = write_indices_17, mode = var_3740_mode_0, updates = v_17_cast_fp16, validate_indices = var_3740_validate_indices_0)[name = tensor("op_3740_cast_fp16")]; + tensor concat_68 = const()[name = tensor("concat_68"), val = tensor([1, 0, 0, 0, 0])]; + tensor concat_69 = const()[name = tensor("concat_69"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_17_internal_tensor_assign_2_stride_0 = const()[name = tensor("new_cache_17_internal_tensor_assign_2_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_17_internal_tensor_assign_2_begin_mask_0 = const()[name = tensor("new_cache_17_internal_tensor_assign_2_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_17_internal_tensor_assign_2_end_mask_0 = const()[name = tensor("new_cache_17_internal_tensor_assign_2_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_17_internal_tensor_assign_2_squeeze_mask_0 = const()[name = tensor("new_cache_17_internal_tensor_assign_2_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_63 = const()[name = tensor("shape_63"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_17 = const()[name = tensor("reduce_prod_17"), val = tensor(1048576)]; + tensor range_1d_17_start_0 = const()[name = tensor("range_1d_17_start_0"), val = tensor(0)]; + tensor range_1d_17_step_0 = const()[name = tensor("range_1d_17_step_0"), val = tensor(1)]; + tensor range_1d_17 = range_1d(end = reduce_prod_17, start = range_1d_17_start_0, step = range_1d_17_step_0)[name = tensor("range_1d_17")]; + tensor reshape_85 = reshape(shape = shape_63, x = range_1d_17)[name = tensor("reshape_85")]; + tensor slice_by_index_17 = slice_by_index(begin = concat_68, begin_mask = new_cache_17_internal_tensor_assign_2_begin_mask_0, end = concat_69, end_mask = new_cache_17_internal_tensor_assign_2_end_mask_0, squeeze_mask = new_cache_17_internal_tensor_assign_2_squeeze_mask_0, stride = new_cache_17_internal_tensor_assign_2_stride_0, x = reshape_85)[name = tensor("slice_by_index_17")]; + tensor reshape_86_shape_0 = const()[name = tensor("reshape_86_shape_0"), val = tensor([-1])]; + tensor reshape_86 = reshape(shape = reshape_86_shape_0, x = slice_by_index_17)[name = tensor("reshape_86")]; + tensor reshape_87_shape_0 = const()[name = tensor("reshape_87_shape_0"), val = tensor([-1])]; + tensor reshape_87_cast_fp16 = reshape(shape = reshape_87_shape_0, x = var_3740_cast_fp16)[name = tensor("reshape_87_cast_fp16")]; + tensor reshape_88_shape_0 = const()[name = tensor("reshape_88_shape_0"), val = tensor([-1])]; + tensor reshape_88_cast_fp16 = reshape(shape = reshape_88_shape_0, x = reshape_84_cast_fp16)[name = tensor("reshape_88_cast_fp16")]; + tensor scatter_17_mode_0 = const()[name = tensor("scatter_17_mode_0"), val = tensor("update")]; + tensor scatter_17_axis_0 = const()[name = tensor("scatter_17_axis_0"), val = tensor(0)]; + tensor scatter_17_validate_indices_0 = const()[name = tensor("scatter_17_validate_indices_0"), val = tensor(false)]; + tensor scatter_17_cast_fp16 = scatter(axis = scatter_17_axis_0, data = reshape_88_cast_fp16, indices = reshape_86, mode = scatter_17_mode_0, updates = reshape_87_cast_fp16, validate_indices = scatter_17_validate_indices_0)[name = tensor("scatter_17_cast_fp16")]; + tensor reshape_89_cast_fp16 = reshape(shape = shape_63, x = scatter_17_cast_fp16)[name = tensor("reshape_89_cast_fp16")]; + tensor reshape_89_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("reshape_89_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor keys_49_begin_0 = const()[name = tensor("keys_49_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor keys_49_end_0 = const()[name = tensor("keys_49_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor keys_49_end_mask_0 = const()[name = tensor("keys_49_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor keys_49_squeeze_mask_0 = const()[name = tensor("keys_49_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor keys_49_cast_fp16 = slice_by_index(begin = keys_49_begin_0, end = keys_49_end_0, end_mask = keys_49_end_mask_0, squeeze_mask = keys_49_squeeze_mask_0, x = reshape_89_cast_fp16)[name = tensor("keys_49_cast_fp16")]; + tensor values_49_begin_0 = const()[name = tensor("values_49_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor values_49_end_0 = const()[name = tensor("values_49_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor values_49_end_mask_0 = const()[name = tensor("values_49_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor values_49_squeeze_mask_0 = const()[name = tensor("values_49_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor values_49_cast_fp16 = slice_by_index(begin = values_49_begin_0, end = values_49_end_0, end_mask = values_49_end_mask_0, squeeze_mask = values_49_squeeze_mask_0, x = reshape_89_cast_fp16)[name = tensor("values_49_cast_fp16")]; + tensor var_3752_cast_fp16 = not_equal(x = keys_49_cast_fp16, y = keys_49_cast_fp16)[name = tensor("op_3752_cast_fp16")]; + tensor keys_51_cast_fp16 = select(a = var_470_to_fp16, b = keys_49_cast_fp16, cond = var_3752_cast_fp16)[name = tensor("keys_51_cast_fp16")]; + tensor var_3760_cast_fp16 = not_equal(x = values_49_cast_fp16, y = values_49_cast_fp16)[name = tensor("op_3760_cast_fp16")]; + tensor values_51_cast_fp16 = select(a = var_470_to_fp16, b = values_49_cast_fp16, cond = var_3760_cast_fp16)[name = tensor("values_51_cast_fp16")]; + tensor var_3784 = const()[name = tensor("op_3784"), val = tensor([0, 2, 1, 3])]; + tensor var_3800 = const()[name = tensor("op_3800"), val = tensor([1, 1, 1])]; + tensor var_3801_cast_fp16 = reshape(shape = var_3800, x = position8_to_fp16)[name = tensor("op_3801_cast_fp16")]; + tensor q_positions_17_cast_fp16 = add(x = var_3801_cast_fp16, y = q_offsets_1_promoted_to_fp16)[name = tensor("q_positions_17_cast_fp16")]; + tensor valid_end_17_cast_fp16 = add(x = var_3801_cast_fp16, y = var_536_cast_fp16)[name = tensor("valid_end_17_cast_fp16")]; + tensor valid_mask_17_cast_fp16 = less(x = k_positions_1_promoted_to_fp16, y = valid_end_17_cast_fp16)[name = tensor("valid_mask_17_cast_fp16")]; + tensor causal_mask_17_cast_fp16 = less_equal(x = k_positions_1_promoted_to_fp16, y = q_positions_17_cast_fp16)[name = tensor("causal_mask_17_cast_fp16")]; + tensor var_3829 = logical_and(x = valid_mask_17_cast_fp16, y = causal_mask_17_cast_fp16)[name = tensor("op_3829")]; + tensor attn_mask_17_axes_0 = const()[name = tensor("attn_mask_17_axes_0"), val = tensor([1])]; + tensor attn_mask_17 = expand_dims(axes = attn_mask_17_axes_0, x = var_3829)[name = tensor("attn_mask_17")]; + tensor var_3843_transpose_x_0 = const()[name = tensor("op_3843_transpose_x_0"), val = tensor(false)]; + tensor var_3843_transpose_y_0 = const()[name = tensor("op_3843_transpose_y_0"), val = tensor(false)]; + tensor transpose_85_perm_0 = const()[name = tensor("transpose_85_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_86_perm_0 = const()[name = tensor("transpose_86_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_86 = transpose(perm = transpose_86_perm_0, x = keys_51_cast_fp16)[name = tensor("transpose_172")]; + tensor transpose_85 = transpose(perm = transpose_85_perm_0, x = q_51_cast_fp16)[name = tensor("transpose_173")]; + tensor var_3843_cast_fp16 = matmul(transpose_x = var_3843_transpose_x_0, transpose_y = var_3843_transpose_y_0, x = transpose_85, y = transpose_86)[name = tensor("op_3843_cast_fp16")]; + tensor var_3837_to_fp16 = const()[name = tensor("op_3837_to_fp16"), val = tensor([0x1p-3])]; + tensor attn_weights_49_cast_fp16 = mul(x = var_3843_cast_fp16, y = var_3837_to_fp16)[name = tensor("attn_weights_49_cast_fp16")]; + tensor var_3845 = logical_not(x = attn_mask_17)[name = tensor("op_3845")]; + tensor var_3846_to_fp16 = const()[name = tensor("op_3846_to_fp16"), val = tensor(-inf)]; + tensor attn_weights_51_cast_fp16 = select(a = var_3846_to_fp16, b = attn_weights_49_cast_fp16, cond = var_3845)[name = tensor("attn_weights_51_cast_fp16")]; + tensor var_3848 = const()[name = tensor("op_3848"), val = tensor(-1)]; + tensor attn_weights_53_cast_fp16 = softmax(axis = var_3848, x = attn_weights_51_cast_fp16)[name = tensor("attn_weights_53_cast_fp16")]; + tensor attn_output_17_transpose_x_0 = const()[name = tensor("attn_output_17_transpose_x_0"), val = tensor(false)]; + tensor attn_output_17_transpose_y_0 = const()[name = tensor("attn_output_17_transpose_y_0"), val = tensor(false)]; + tensor values_53_cast_fp16 = transpose(perm = var_3784, x = values_51_cast_fp16)[name = tensor("transpose_174")]; + tensor attn_output_17_cast_fp16 = matmul(transpose_x = attn_output_17_transpose_x_0, transpose_y = attn_output_17_transpose_y_0, x = attn_weights_53_cast_fp16, y = values_53_cast_fp16)[name = tensor("attn_output_17_cast_fp16")]; + tensor var_3856 = const()[name = tensor("op_3856"), val = tensor([0, 2, 1, 3])]; + tensor var_3861 = const()[name = tensor("op_3861"), val = tensor([1, 256, 1024])]; + tensor var_3857_cast_fp16 = transpose(perm = var_3856, x = attn_output_17_cast_fp16)[name = tensor("transpose_171")]; + tensor input_101_cast_fp16 = reshape(shape = var_3861, x = var_3857_cast_fp16)[name = tensor("input_101_cast_fp16")]; + tensor attn8_out_proj_weight_to_fp16 = const()[name = tensor("attn8_out_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(208761216)))]; + tensor linear_33_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = attn8_out_proj_weight_to_fp16, x = input_101_cast_fp16)[name = tensor("linear_33_cast_fp16")]; + tensor var_3870_cast_fp16 = add(x = position8_to_fp16, y = valid_len_to_fp16)[name = tensor("op_3870_cast_fp16")]; + tensor var_3870_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_3870_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor input_103_cast_fp16 = add(x = input_97_cast_fp16, y = linear_33_cast_fp16)[name = tensor("input_103_cast_fp16")]; + tensor input_105_axes_0 = const()[name = tensor("input_105_axes_0"), val = tensor([-1])]; + tensor norm8_2_weight_to_fp16 = const()[name = tensor("norm8_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(210858432)))]; + tensor norm8_2_bias_to_fp16 = const()[name = tensor("norm8_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(210860544)))]; + tensor var_3874_to_fp16 = const()[name = tensor("op_3874_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_105_cast_fp16 = layer_norm(axes = input_105_axes_0, beta = norm8_2_bias_to_fp16, epsilon = var_3874_to_fp16, gamma = norm8_2_weight_to_fp16, x = input_103_cast_fp16)[name = tensor("input_105_cast_fp16")]; + tensor linear8_1_weight_to_fp16 = const()[name = tensor("linear8_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(210862656)))]; + tensor linear_34_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = linear8_1_weight_to_fp16, x = input_105_cast_fp16)[name = tensor("linear_34_cast_fp16")]; + tensor input_107_mode_0 = const()[name = tensor("input_107_mode_0"), val = tensor("EXACT")]; + tensor input_107_cast_fp16 = gelu(mode = input_107_mode_0, x = linear_34_cast_fp16)[name = tensor("input_107_cast_fp16")]; + tensor linear8_2_weight_to_fp16 = const()[name = tensor("linear8_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(219251328)))]; + tensor linear_35_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = linear8_2_weight_to_fp16, x = input_107_cast_fp16)[name = tensor("linear_35_cast_fp16")]; + tensor input_109_cast_fp16 = add(x = input_103_cast_fp16, y = linear_35_cast_fp16)[name = tensor("input_109_cast_fp16")]; + tensor input_111_axes_0 = const()[name = tensor("input_111_axes_0"), val = tensor([-1])]; + tensor norm9_1_weight_to_fp16 = const()[name = tensor("norm9_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(227640000)))]; + tensor norm9_1_bias_to_fp16 = const()[name = tensor("norm9_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(227642112)))]; + tensor var_3891_to_fp16 = const()[name = tensor("op_3891_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_111_cast_fp16 = layer_norm(axes = input_111_axes_0, beta = norm9_1_bias_to_fp16, epsilon = var_3891_to_fp16, gamma = norm9_1_weight_to_fp16, x = input_109_cast_fp16)[name = tensor("input_111_cast_fp16")]; + tensor attn9_in_proj_weight_to_fp16 = const()[name = tensor("attn9_in_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(227644224)))]; + tensor linear_36_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = attn9_in_proj_weight_to_fp16, x = input_111_cast_fp16)[name = tensor("linear_36_cast_fp16")]; + tensor var_3905 = const()[name = tensor("op_3905"), val = tensor([1, 256, 3, 16, 64])]; + tensor qkv_19_cast_fp16 = reshape(shape = var_3905, x = linear_36_cast_fp16)[name = tensor("qkv_19_cast_fp16")]; + tensor q_55_begin_0 = const()[name = tensor("q_55_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor q_55_end_0 = const()[name = tensor("q_55_end_0"), val = tensor([1, 256, 1, 16, 64])]; + tensor q_55_end_mask_0 = const()[name = tensor("q_55_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor q_55_squeeze_mask_0 = const()[name = tensor("q_55_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor q_55_cast_fp16 = slice_by_index(begin = q_55_begin_0, end = q_55_end_0, end_mask = q_55_end_mask_0, squeeze_mask = q_55_squeeze_mask_0, x = qkv_19_cast_fp16)[name = tensor("q_55_cast_fp16")]; + tensor k_37_begin_0 = const()[name = tensor("k_37_begin_0"), val = tensor([0, 0, 1, 0, 0])]; + tensor k_37_end_0 = const()[name = tensor("k_37_end_0"), val = tensor([1, 256, 2, 16, 64])]; + tensor k_37_end_mask_0 = const()[name = tensor("k_37_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor k_37_squeeze_mask_0 = const()[name = tensor("k_37_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor k_37_cast_fp16 = slice_by_index(begin = k_37_begin_0, end = k_37_end_0, end_mask = k_37_end_mask_0, squeeze_mask = k_37_squeeze_mask_0, x = qkv_19_cast_fp16)[name = tensor("k_37_cast_fp16")]; + tensor v_19_begin_0 = const()[name = tensor("v_19_begin_0"), val = tensor([0, 0, 2, 0, 0])]; + tensor v_19_end_0 = const()[name = tensor("v_19_end_0"), val = tensor([1, 256, 3, 16, 64])]; + tensor v_19_end_mask_0 = const()[name = tensor("v_19_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor v_19_squeeze_mask_0 = const()[name = tensor("v_19_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor v_19_cast_fp16 = slice_by_index(begin = v_19_begin_0, end = v_19_end_0, end_mask = v_19_end_mask_0, squeeze_mask = v_19_squeeze_mask_0, x = qkv_19_cast_fp16)[name = tensor("v_19_cast_fp16")]; + tensor var_3963 = const()[name = tensor("op_3963"), val = tensor([1, 1])]; + tensor position9_to_fp16_dtype_0 = const()[name = tensor("position9_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor position9_to_fp16 = cast(dtype = position9_to_fp16_dtype_0, x = position9)[name = tensor("cast_616")]; + tensor var_3964_cast_fp16 = reshape(shape = var_3963, x = position9_to_fp16)[name = tensor("op_3964_cast_fp16")]; + tensor ts_57_cast_fp16 = add(x = ts_1_promoted_to_fp16, y = var_3964_cast_fp16)[name = tensor("ts_57_cast_fp16")]; + tensor var_3971 = const()[name = tensor("op_3971"), val = tensor([1, 256, 1, 1])]; + tensor ts_59_cast_fp16 = reshape(shape = var_3971, x = ts_57_cast_fp16)[name = tensor("ts_59_cast_fp16")]; + tensor var_3978 = const()[name = tensor("op_3978"), val = tensor([1, 256, 16, 32, 2])]; + tensor q_complex_19_cast_fp16 = reshape(shape = var_3978, x = q_55_cast_fp16)[name = tensor("q_complex_19_cast_fp16")]; + tensor var_3985 = const()[name = tensor("op_3985"), val = tensor([1, 256, 16, 32, 2])]; + tensor k_complex_19_cast_fp16 = reshape(shape = var_3985, x = k_37_cast_fp16)[name = tensor("k_complex_19_cast_fp16")]; + tensor var_3989_begin_0 = const()[name = tensor("op_3989_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_3989_end_0 = const()[name = tensor("op_3989_end_0"), val = tensor([1, 256, 16, 32, 1])]; + tensor var_3989_end_mask_0 = const()[name = tensor("op_3989_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_3989_squeeze_mask_0 = const()[name = tensor("op_3989_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_3989_cast_fp16 = slice_by_index(begin = var_3989_begin_0, end = var_3989_end_0, end_mask = var_3989_end_mask_0, squeeze_mask = var_3989_squeeze_mask_0, x = q_complex_19_cast_fp16)[name = tensor("op_3989_cast_fp16")]; + tensor var_3997_begin_0 = const()[name = tensor("op_3997_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_3997_end_0 = const()[name = tensor("op_3997_end_0"), val = tensor([1, 256, 16, 32, 2])]; + tensor var_3997_end_mask_0 = const()[name = tensor("op_3997_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_3997_squeeze_mask_0 = const()[name = tensor("op_3997_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_3997_cast_fp16 = slice_by_index(begin = var_3997_begin_0, end = var_3997_end_0, end_mask = var_3997_end_mask_0, squeeze_mask = var_3997_squeeze_mask_0, x = q_complex_19_cast_fp16)[name = tensor("op_3997_cast_fp16")]; + tensor var_4005_begin_0 = const()[name = tensor("op_4005_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_4005_end_0 = const()[name = tensor("op_4005_end_0"), val = tensor([1, 256, 16, 32, 1])]; + tensor var_4005_end_mask_0 = const()[name = tensor("op_4005_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_4005_squeeze_mask_0 = const()[name = tensor("op_4005_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_4005_cast_fp16 = slice_by_index(begin = var_4005_begin_0, end = var_4005_end_0, end_mask = var_4005_end_mask_0, squeeze_mask = var_4005_squeeze_mask_0, x = k_complex_19_cast_fp16)[name = tensor("op_4005_cast_fp16")]; + tensor var_4013_begin_0 = const()[name = tensor("op_4013_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_4013_end_0 = const()[name = tensor("op_4013_end_0"), val = tensor([1, 256, 16, 32, 2])]; + tensor var_4013_end_mask_0 = const()[name = tensor("op_4013_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_4013_squeeze_mask_0 = const()[name = tensor("op_4013_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_4013_cast_fp16 = slice_by_index(begin = var_4013_begin_0, end = var_4013_end_0, end_mask = var_4013_end_mask_0, squeeze_mask = var_4013_squeeze_mask_0, x = k_complex_19_cast_fp16)[name = tensor("op_4013_cast_fp16")]; + tensor freqs_19_to_fp16 = const()[name = tensor("freqs_19_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(233935744)))]; + tensor var_4019_cast_fp16 = mul(x = freqs_19_to_fp16, y = ts_59_cast_fp16)[name = tensor("op_4019_cast_fp16")]; + tensor rotr_19_cast_fp16 = cos(x = var_4019_cast_fp16)[name = tensor("rotr_19_cast_fp16")]; + tensor roti_19_cast_fp16 = sin(x = var_4019_cast_fp16)[name = tensor("roti_19_cast_fp16")]; + tensor var_4023_cast_fp16 = mul(x = var_3989_cast_fp16, y = rotr_19_cast_fp16)[name = tensor("op_4023_cast_fp16")]; + tensor var_4024_cast_fp16 = mul(x = var_3997_cast_fp16, y = roti_19_cast_fp16)[name = tensor("op_4024_cast_fp16")]; + tensor qor_37_cast_fp16 = sub(x = var_4023_cast_fp16, y = var_4024_cast_fp16)[name = tensor("qor_37_cast_fp16")]; + tensor var_4027_cast_fp16 = mul(x = var_3989_cast_fp16, y = roti_19_cast_fp16)[name = tensor("op_4027_cast_fp16")]; + tensor var_4028_cast_fp16 = mul(x = var_3997_cast_fp16, y = rotr_19_cast_fp16)[name = tensor("op_4028_cast_fp16")]; + tensor qoi_37_cast_fp16 = add(x = var_4027_cast_fp16, y = var_4028_cast_fp16)[name = tensor("qoi_37_cast_fp16")]; + tensor var_4031_cast_fp16 = mul(x = var_4005_cast_fp16, y = rotr_19_cast_fp16)[name = tensor("op_4031_cast_fp16")]; + tensor var_4032_cast_fp16 = mul(x = var_4013_cast_fp16, y = roti_19_cast_fp16)[name = tensor("op_4032_cast_fp16")]; + tensor kor_37_cast_fp16 = sub(x = var_4031_cast_fp16, y = var_4032_cast_fp16)[name = tensor("kor_37_cast_fp16")]; + tensor var_4035_cast_fp16 = mul(x = var_4005_cast_fp16, y = roti_19_cast_fp16)[name = tensor("op_4035_cast_fp16")]; + tensor var_4036_cast_fp16 = mul(x = var_4013_cast_fp16, y = rotr_19_cast_fp16)[name = tensor("op_4036_cast_fp16")]; + tensor koi_37_cast_fp16 = add(x = var_4035_cast_fp16, y = var_4036_cast_fp16)[name = tensor("koi_37_cast_fp16")]; + tensor qo_19_axis_0 = const()[name = tensor("qo_19_axis_0"), val = tensor(-1)]; + tensor qo_19_cast_fp16 = stack(axis = qo_19_axis_0, values = (qor_37_cast_fp16, qoi_37_cast_fp16))[name = tensor("qo_19_cast_fp16")]; + tensor ko_19_axis_0 = const()[name = tensor("ko_19_axis_0"), val = tensor(-1)]; + tensor ko_19_cast_fp16 = stack(axis = ko_19_axis_0, values = (kor_37_cast_fp16, koi_37_cast_fp16))[name = tensor("ko_19_cast_fp16")]; + tensor var_4069 = const()[name = tensor("op_4069"), val = tensor([1, 256, 16, 64])]; + tensor q_57_cast_fp16 = reshape(shape = var_4069, x = qo_19_cast_fp16)[name = tensor("q_57_cast_fp16")]; + tensor var_4075 = const()[name = tensor("op_4075"), val = tensor([1, 256, 16, 64])]; + tensor k_39_cast_fp16 = reshape(shape = var_4075, x = ko_19_cast_fp16)[name = tensor("k_39_cast_fp16")]; + tensor write_indices_float_55_cast_fp16 = add(x = var_3964_cast_fp16, y = var_393_promoted_to_fp16)[name = tensor("write_indices_float_55_cast_fp16")]; + tensor _inversed_4096_y_0_to_fp16 = const()[name = tensor("_inversed_4096_y_0_to_fp16"), val = tensor(0x1p-9)]; + tensor _inversed_4096_cast_fp16 = mul(x = write_indices_float_55_cast_fp16, y = _inversed_4096_y_0_to_fp16)[name = tensor("_inversed_4096_cast_fp16")]; + tensor var_4097_cast_fp16 = floor(x = _inversed_4096_cast_fp16)[name = tensor("op_4097_cast_fp16")]; + tensor var_4098_to_fp16 = const()[name = tensor("op_4098_to_fp16"), val = tensor(0x1p+9)]; + tensor var_4099_cast_fp16 = mul(x = var_4097_cast_fp16, y = var_4098_to_fp16)[name = tensor("op_4099_cast_fp16")]; + tensor write_indices_float_57_cast_fp16 = sub(x = write_indices_float_55_cast_fp16, y = var_4099_cast_fp16)[name = tensor("write_indices_float_57_cast_fp16")]; + tensor var_4117_value_0_to_fp16 = const()[name = tensor("op_4117_value_0_to_fp16"), val = tensor(0x1.ffp+8)]; + tensor var_4117_cast_fp16 = fill_like(ref_tensor = write_indices_float_57_cast_fp16, value = var_4117_value_0_to_fp16)[name = tensor("op_4117_cast_fp16")]; + tensor write_indices_float_59_cast_fp16 = select(a = write_indices_float_57_cast_fp16, b = var_4117_cast_fp16, cond = token_valid_1_cast_fp16)[name = tensor("write_indices_float_59_cast_fp16")]; + tensor var_4123_dtype_0 = const()[name = tensor("op_4123_dtype_0"), val = tensor("int32")]; + tensor var_4128 = const()[name = tensor("op_4128"), val = tensor([1, 256, 1, 1])]; + tensor write_indices_float_59_cast_fp16_to_int32 = cast(dtype = var_4123_dtype_0, x = write_indices_float_59_cast_fp16)[name = tensor("cast_615")]; + tensor var_4129 = reshape(shape = var_4128, x = write_indices_float_59_cast_fp16_to_int32)[name = tensor("op_4129")]; + tensor write_indices_19_reps_0 = const()[name = tensor("write_indices_19_reps_0"), val = tensor([1, 1, 16, 64])]; + tensor write_indices_19 = tile(reps = write_indices_19_reps_0, x = var_4129)[name = tensor("write_indices_19")]; + tensor var_4139_begin_0 = const()[name = tensor("op_4139_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_4139_end_0 = const()[name = tensor("op_4139_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor var_4139_end_mask_0 = const()[name = tensor("op_4139_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_4139_squeeze_mask_0 = const()[name = tensor("op_4139_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor cache9_to_fp16_dtype_0 = const()[name = tensor("cache9_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor cache9_to_fp16 = cast(dtype = cache9_to_fp16_dtype_0, x = cache9)[name = tensor("cast_614")]; + tensor var_4139_cast_fp16 = slice_by_index(begin = var_4139_begin_0, end = var_4139_end_0, end_mask = var_4139_end_mask_0, squeeze_mask = var_4139_squeeze_mask_0, x = cache9_to_fp16)[name = tensor("op_4139_cast_fp16")]; + tensor var_4141_axis_0 = const()[name = tensor("op_4141_axis_0"), val = tensor(1)]; + tensor var_4141_mode_0 = const()[name = tensor("op_4141_mode_0"), val = tensor("update")]; + tensor var_4141_validate_indices_0 = const()[name = tensor("op_4141_validate_indices_0"), val = tensor(false)]; + tensor var_4141_cast_fp16 = scatter_along_axis(axis = var_4141_axis_0, data = var_4139_cast_fp16, indices = write_indices_19, mode = var_4141_mode_0, updates = k_39_cast_fp16, validate_indices = var_4141_validate_indices_0)[name = tensor("op_4141_cast_fp16")]; + tensor concat_74 = const()[name = tensor("concat_74"), val = tensor([0, 0, 0, 0, 0])]; + tensor concat_75 = const()[name = tensor("concat_75"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_19_internal_tensor_assign_1_stride_0 = const()[name = tensor("new_cache_19_internal_tensor_assign_1_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_19_internal_tensor_assign_1_begin_mask_0 = const()[name = tensor("new_cache_19_internal_tensor_assign_1_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_19_internal_tensor_assign_1_end_mask_0 = const()[name = tensor("new_cache_19_internal_tensor_assign_1_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_19_internal_tensor_assign_1_squeeze_mask_0 = const()[name = tensor("new_cache_19_internal_tensor_assign_1_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_64 = const()[name = tensor("shape_64"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_18 = const()[name = tensor("reduce_prod_18"), val = tensor(1048576)]; + tensor range_1d_18_start_0 = const()[name = tensor("range_1d_18_start_0"), val = tensor(0)]; + tensor range_1d_18_step_0 = const()[name = tensor("range_1d_18_step_0"), val = tensor(1)]; + tensor range_1d_18 = range_1d(end = reduce_prod_18, start = range_1d_18_start_0, step = range_1d_18_step_0)[name = tensor("range_1d_18")]; + tensor reshape_90 = reshape(shape = shape_64, x = range_1d_18)[name = tensor("reshape_90")]; + tensor slice_by_index_18 = slice_by_index(begin = concat_74, begin_mask = new_cache_19_internal_tensor_assign_1_begin_mask_0, end = concat_75, end_mask = new_cache_19_internal_tensor_assign_1_end_mask_0, squeeze_mask = new_cache_19_internal_tensor_assign_1_squeeze_mask_0, stride = new_cache_19_internal_tensor_assign_1_stride_0, x = reshape_90)[name = tensor("slice_by_index_18")]; + tensor reshape_91_shape_0 = const()[name = tensor("reshape_91_shape_0"), val = tensor([-1])]; + tensor reshape_91 = reshape(shape = reshape_91_shape_0, x = slice_by_index_18)[name = tensor("reshape_91")]; + tensor reshape_92_shape_0 = const()[name = tensor("reshape_92_shape_0"), val = tensor([-1])]; + tensor reshape_92_cast_fp16 = reshape(shape = reshape_92_shape_0, x = var_4141_cast_fp16)[name = tensor("reshape_92_cast_fp16")]; + tensor reshape_93_shape_0 = const()[name = tensor("reshape_93_shape_0"), val = tensor([-1])]; + tensor reshape_93_cast_fp16 = reshape(shape = reshape_93_shape_0, x = cache9_to_fp16)[name = tensor("reshape_93_cast_fp16")]; + tensor scatter_18_mode_0 = const()[name = tensor("scatter_18_mode_0"), val = tensor("update")]; + tensor scatter_18_axis_0 = const()[name = tensor("scatter_18_axis_0"), val = tensor(0)]; + tensor scatter_18_validate_indices_0 = const()[name = tensor("scatter_18_validate_indices_0"), val = tensor(false)]; + tensor scatter_18_cast_fp16 = scatter(axis = scatter_18_axis_0, data = reshape_93_cast_fp16, indices = reshape_91, mode = scatter_18_mode_0, updates = reshape_92_cast_fp16, validate_indices = scatter_18_validate_indices_0)[name = tensor("scatter_18_cast_fp16")]; + tensor reshape_94_cast_fp16 = reshape(shape = shape_64, x = scatter_18_cast_fp16)[name = tensor("reshape_94_cast_fp16")]; + tensor var_4149_begin_0 = const()[name = tensor("op_4149_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor var_4149_end_0 = const()[name = tensor("op_4149_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor var_4149_end_mask_0 = const()[name = tensor("op_4149_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_4149_squeeze_mask_0 = const()[name = tensor("op_4149_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor var_4149_cast_fp16 = slice_by_index(begin = var_4149_begin_0, end = var_4149_end_0, end_mask = var_4149_end_mask_0, squeeze_mask = var_4149_squeeze_mask_0, x = reshape_94_cast_fp16)[name = tensor("op_4149_cast_fp16")]; + tensor var_4151_axis_0 = const()[name = tensor("op_4151_axis_0"), val = tensor(1)]; + tensor var_4151_mode_0 = const()[name = tensor("op_4151_mode_0"), val = tensor("update")]; + tensor var_4151_validate_indices_0 = const()[name = tensor("op_4151_validate_indices_0"), val = tensor(false)]; + tensor var_4151_cast_fp16 = scatter_along_axis(axis = var_4151_axis_0, data = var_4149_cast_fp16, indices = write_indices_19, mode = var_4151_mode_0, updates = v_19_cast_fp16, validate_indices = var_4151_validate_indices_0)[name = tensor("op_4151_cast_fp16")]; + tensor concat_76 = const()[name = tensor("concat_76"), val = tensor([1, 0, 0, 0, 0])]; + tensor concat_77 = const()[name = tensor("concat_77"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_19_internal_tensor_assign_2_stride_0 = const()[name = tensor("new_cache_19_internal_tensor_assign_2_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_19_internal_tensor_assign_2_begin_mask_0 = const()[name = tensor("new_cache_19_internal_tensor_assign_2_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_19_internal_tensor_assign_2_end_mask_0 = const()[name = tensor("new_cache_19_internal_tensor_assign_2_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_19_internal_tensor_assign_2_squeeze_mask_0 = const()[name = tensor("new_cache_19_internal_tensor_assign_2_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_65 = const()[name = tensor("shape_65"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_19 = const()[name = tensor("reduce_prod_19"), val = tensor(1048576)]; + tensor range_1d_19_start_0 = const()[name = tensor("range_1d_19_start_0"), val = tensor(0)]; + tensor range_1d_19_step_0 = const()[name = tensor("range_1d_19_step_0"), val = tensor(1)]; + tensor range_1d_19 = range_1d(end = reduce_prod_19, start = range_1d_19_start_0, step = range_1d_19_step_0)[name = tensor("range_1d_19")]; + tensor reshape_95 = reshape(shape = shape_65, x = range_1d_19)[name = tensor("reshape_95")]; + tensor slice_by_index_19 = slice_by_index(begin = concat_76, begin_mask = new_cache_19_internal_tensor_assign_2_begin_mask_0, end = concat_77, end_mask = new_cache_19_internal_tensor_assign_2_end_mask_0, squeeze_mask = new_cache_19_internal_tensor_assign_2_squeeze_mask_0, stride = new_cache_19_internal_tensor_assign_2_stride_0, x = reshape_95)[name = tensor("slice_by_index_19")]; + tensor reshape_96_shape_0 = const()[name = tensor("reshape_96_shape_0"), val = tensor([-1])]; + tensor reshape_96 = reshape(shape = reshape_96_shape_0, x = slice_by_index_19)[name = tensor("reshape_96")]; + tensor reshape_97_shape_0 = const()[name = tensor("reshape_97_shape_0"), val = tensor([-1])]; + tensor reshape_97_cast_fp16 = reshape(shape = reshape_97_shape_0, x = var_4151_cast_fp16)[name = tensor("reshape_97_cast_fp16")]; + tensor reshape_98_shape_0 = const()[name = tensor("reshape_98_shape_0"), val = tensor([-1])]; + tensor reshape_98_cast_fp16 = reshape(shape = reshape_98_shape_0, x = reshape_94_cast_fp16)[name = tensor("reshape_98_cast_fp16")]; + tensor scatter_19_mode_0 = const()[name = tensor("scatter_19_mode_0"), val = tensor("update")]; + tensor scatter_19_axis_0 = const()[name = tensor("scatter_19_axis_0"), val = tensor(0)]; + tensor scatter_19_validate_indices_0 = const()[name = tensor("scatter_19_validate_indices_0"), val = tensor(false)]; + tensor scatter_19_cast_fp16 = scatter(axis = scatter_19_axis_0, data = reshape_98_cast_fp16, indices = reshape_96, mode = scatter_19_mode_0, updates = reshape_97_cast_fp16, validate_indices = scatter_19_validate_indices_0)[name = tensor("scatter_19_cast_fp16")]; + tensor reshape_99_cast_fp16 = reshape(shape = shape_65, x = scatter_19_cast_fp16)[name = tensor("reshape_99_cast_fp16")]; + tensor reshape_99_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("reshape_99_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor keys_55_begin_0 = const()[name = tensor("keys_55_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor keys_55_end_0 = const()[name = tensor("keys_55_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor keys_55_end_mask_0 = const()[name = tensor("keys_55_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor keys_55_squeeze_mask_0 = const()[name = tensor("keys_55_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor keys_55_cast_fp16 = slice_by_index(begin = keys_55_begin_0, end = keys_55_end_0, end_mask = keys_55_end_mask_0, squeeze_mask = keys_55_squeeze_mask_0, x = reshape_99_cast_fp16)[name = tensor("keys_55_cast_fp16")]; + tensor values_55_begin_0 = const()[name = tensor("values_55_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor values_55_end_0 = const()[name = tensor("values_55_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor values_55_end_mask_0 = const()[name = tensor("values_55_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor values_55_squeeze_mask_0 = const()[name = tensor("values_55_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor values_55_cast_fp16 = slice_by_index(begin = values_55_begin_0, end = values_55_end_0, end_mask = values_55_end_mask_0, squeeze_mask = values_55_squeeze_mask_0, x = reshape_99_cast_fp16)[name = tensor("values_55_cast_fp16")]; + tensor var_4163_cast_fp16 = not_equal(x = keys_55_cast_fp16, y = keys_55_cast_fp16)[name = tensor("op_4163_cast_fp16")]; + tensor keys_57_cast_fp16 = select(a = var_470_to_fp16, b = keys_55_cast_fp16, cond = var_4163_cast_fp16)[name = tensor("keys_57_cast_fp16")]; + tensor var_4171_cast_fp16 = not_equal(x = values_55_cast_fp16, y = values_55_cast_fp16)[name = tensor("op_4171_cast_fp16")]; + tensor values_57_cast_fp16 = select(a = var_470_to_fp16, b = values_55_cast_fp16, cond = var_4171_cast_fp16)[name = tensor("values_57_cast_fp16")]; + tensor var_4195 = const()[name = tensor("op_4195"), val = tensor([0, 2, 1, 3])]; + tensor var_4211 = const()[name = tensor("op_4211"), val = tensor([1, 1, 1])]; + tensor var_4212_cast_fp16 = reshape(shape = var_4211, x = position9_to_fp16)[name = tensor("op_4212_cast_fp16")]; + tensor q_positions_19_cast_fp16 = add(x = var_4212_cast_fp16, y = q_offsets_1_promoted_to_fp16)[name = tensor("q_positions_19_cast_fp16")]; + tensor valid_end_19_cast_fp16 = add(x = var_4212_cast_fp16, y = var_536_cast_fp16)[name = tensor("valid_end_19_cast_fp16")]; + tensor valid_mask_19_cast_fp16 = less(x = k_positions_1_promoted_to_fp16, y = valid_end_19_cast_fp16)[name = tensor("valid_mask_19_cast_fp16")]; + tensor causal_mask_19_cast_fp16 = less_equal(x = k_positions_1_promoted_to_fp16, y = q_positions_19_cast_fp16)[name = tensor("causal_mask_19_cast_fp16")]; + tensor var_4240 = logical_and(x = valid_mask_19_cast_fp16, y = causal_mask_19_cast_fp16)[name = tensor("op_4240")]; + tensor attn_mask_19_axes_0 = const()[name = tensor("attn_mask_19_axes_0"), val = tensor([1])]; + tensor attn_mask_19 = expand_dims(axes = attn_mask_19_axes_0, x = var_4240)[name = tensor("attn_mask_19")]; + tensor var_4254_transpose_x_0 = const()[name = tensor("op_4254_transpose_x_0"), val = tensor(false)]; + tensor var_4254_transpose_y_0 = const()[name = tensor("op_4254_transpose_y_0"), val = tensor(false)]; + tensor transpose_87_perm_0 = const()[name = tensor("transpose_87_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_88_perm_0 = const()[name = tensor("transpose_88_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_88 = transpose(perm = transpose_88_perm_0, x = keys_57_cast_fp16)[name = tensor("transpose_168")]; + tensor transpose_87 = transpose(perm = transpose_87_perm_0, x = q_57_cast_fp16)[name = tensor("transpose_169")]; + tensor var_4254_cast_fp16 = matmul(transpose_x = var_4254_transpose_x_0, transpose_y = var_4254_transpose_y_0, x = transpose_87, y = transpose_88)[name = tensor("op_4254_cast_fp16")]; + tensor var_4248_to_fp16 = const()[name = tensor("op_4248_to_fp16"), val = tensor([0x1p-3])]; + tensor attn_weights_55_cast_fp16 = mul(x = var_4254_cast_fp16, y = var_4248_to_fp16)[name = tensor("attn_weights_55_cast_fp16")]; + tensor var_4256 = logical_not(x = attn_mask_19)[name = tensor("op_4256")]; + tensor var_4257_to_fp16 = const()[name = tensor("op_4257_to_fp16"), val = tensor(-inf)]; + tensor attn_weights_57_cast_fp16 = select(a = var_4257_to_fp16, b = attn_weights_55_cast_fp16, cond = var_4256)[name = tensor("attn_weights_57_cast_fp16")]; + tensor var_4259 = const()[name = tensor("op_4259"), val = tensor(-1)]; + tensor attn_weights_59_cast_fp16 = softmax(axis = var_4259, x = attn_weights_57_cast_fp16)[name = tensor("attn_weights_59_cast_fp16")]; + tensor attn_output_19_transpose_x_0 = const()[name = tensor("attn_output_19_transpose_x_0"), val = tensor(false)]; + tensor attn_output_19_transpose_y_0 = const()[name = tensor("attn_output_19_transpose_y_0"), val = tensor(false)]; + tensor values_59_cast_fp16 = transpose(perm = var_4195, x = values_57_cast_fp16)[name = tensor("transpose_170")]; + tensor attn_output_19_cast_fp16 = matmul(transpose_x = attn_output_19_transpose_x_0, transpose_y = attn_output_19_transpose_y_0, x = attn_weights_59_cast_fp16, y = values_59_cast_fp16)[name = tensor("attn_output_19_cast_fp16")]; + tensor var_4267 = const()[name = tensor("op_4267"), val = tensor([0, 2, 1, 3])]; + tensor var_4272 = const()[name = tensor("op_4272"), val = tensor([1, 256, 1024])]; + tensor var_4268_cast_fp16 = transpose(perm = var_4267, x = attn_output_19_cast_fp16)[name = tensor("transpose_167")]; + tensor input_113_cast_fp16 = reshape(shape = var_4272, x = var_4268_cast_fp16)[name = tensor("input_113_cast_fp16")]; + tensor attn9_out_proj_weight_to_fp16 = const()[name = tensor("attn9_out_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(233935872)))]; + tensor linear_37_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = attn9_out_proj_weight_to_fp16, x = input_113_cast_fp16)[name = tensor("linear_37_cast_fp16")]; + tensor var_4281_cast_fp16 = add(x = position9_to_fp16, y = valid_len_to_fp16)[name = tensor("op_4281_cast_fp16")]; + tensor var_4281_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_4281_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor input_115_cast_fp16 = add(x = input_109_cast_fp16, y = linear_37_cast_fp16)[name = tensor("input_115_cast_fp16")]; + tensor input_117_axes_0 = const()[name = tensor("input_117_axes_0"), val = tensor([-1])]; + tensor norm9_2_weight_to_fp16 = const()[name = tensor("norm9_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(236033088)))]; + tensor norm9_2_bias_to_fp16 = const()[name = tensor("norm9_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(236035200)))]; + tensor var_4285_to_fp16 = const()[name = tensor("op_4285_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_117_cast_fp16 = layer_norm(axes = input_117_axes_0, beta = norm9_2_bias_to_fp16, epsilon = var_4285_to_fp16, gamma = norm9_2_weight_to_fp16, x = input_115_cast_fp16)[name = tensor("input_117_cast_fp16")]; + tensor linear9_1_weight_to_fp16 = const()[name = tensor("linear9_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(236037312)))]; + tensor linear_38_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = linear9_1_weight_to_fp16, x = input_117_cast_fp16)[name = tensor("linear_38_cast_fp16")]; + tensor input_119_mode_0 = const()[name = tensor("input_119_mode_0"), val = tensor("EXACT")]; + tensor input_119_cast_fp16 = gelu(mode = input_119_mode_0, x = linear_38_cast_fp16)[name = tensor("input_119_cast_fp16")]; + tensor linear9_2_weight_to_fp16 = const()[name = tensor("linear9_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(244425984)))]; + tensor linear_39_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = linear9_2_weight_to_fp16, x = input_119_cast_fp16)[name = tensor("linear_39_cast_fp16")]; + tensor input_121_cast_fp16 = add(x = input_115_cast_fp16, y = linear_39_cast_fp16)[name = tensor("input_121_cast_fp16")]; + tensor input_123_axes_0 = const()[name = tensor("input_123_axes_0"), val = tensor([-1])]; + tensor norm10_1_weight_to_fp16 = const()[name = tensor("norm10_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(252814656)))]; + tensor norm10_1_bias_to_fp16 = const()[name = tensor("norm10_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(252816768)))]; + tensor var_4302_to_fp16 = const()[name = tensor("op_4302_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_123_cast_fp16 = layer_norm(axes = input_123_axes_0, beta = norm10_1_bias_to_fp16, epsilon = var_4302_to_fp16, gamma = norm10_1_weight_to_fp16, x = input_121_cast_fp16)[name = tensor("input_123_cast_fp16")]; + tensor attn10_in_proj_weight_to_fp16 = const()[name = tensor("attn10_in_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(252818880)))]; + tensor linear_40_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = attn10_in_proj_weight_to_fp16, x = input_123_cast_fp16)[name = tensor("linear_40_cast_fp16")]; + tensor var_4316 = const()[name = tensor("op_4316"), val = tensor([1, 256, 3, 16, 64])]; + tensor qkv_21_cast_fp16 = reshape(shape = var_4316, x = linear_40_cast_fp16)[name = tensor("qkv_21_cast_fp16")]; + tensor q_61_begin_0 = const()[name = tensor("q_61_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor q_61_end_0 = const()[name = tensor("q_61_end_0"), val = tensor([1, 256, 1, 16, 64])]; + tensor q_61_end_mask_0 = const()[name = tensor("q_61_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor q_61_squeeze_mask_0 = const()[name = tensor("q_61_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor q_61_cast_fp16 = slice_by_index(begin = q_61_begin_0, end = q_61_end_0, end_mask = q_61_end_mask_0, squeeze_mask = q_61_squeeze_mask_0, x = qkv_21_cast_fp16)[name = tensor("q_61_cast_fp16")]; + tensor k_41_begin_0 = const()[name = tensor("k_41_begin_0"), val = tensor([0, 0, 1, 0, 0])]; + tensor k_41_end_0 = const()[name = tensor("k_41_end_0"), val = tensor([1, 256, 2, 16, 64])]; + tensor k_41_end_mask_0 = const()[name = tensor("k_41_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor k_41_squeeze_mask_0 = const()[name = tensor("k_41_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor k_41_cast_fp16 = slice_by_index(begin = k_41_begin_0, end = k_41_end_0, end_mask = k_41_end_mask_0, squeeze_mask = k_41_squeeze_mask_0, x = qkv_21_cast_fp16)[name = tensor("k_41_cast_fp16")]; + tensor v_21_begin_0 = const()[name = tensor("v_21_begin_0"), val = tensor([0, 0, 2, 0, 0])]; + tensor v_21_end_0 = const()[name = tensor("v_21_end_0"), val = tensor([1, 256, 3, 16, 64])]; + tensor v_21_end_mask_0 = const()[name = tensor("v_21_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor v_21_squeeze_mask_0 = const()[name = tensor("v_21_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor v_21_cast_fp16 = slice_by_index(begin = v_21_begin_0, end = v_21_end_0, end_mask = v_21_end_mask_0, squeeze_mask = v_21_squeeze_mask_0, x = qkv_21_cast_fp16)[name = tensor("v_21_cast_fp16")]; + tensor var_4374 = const()[name = tensor("op_4374"), val = tensor([1, 1])]; + tensor position10_to_fp16_dtype_0 = const()[name = tensor("position10_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor position10_to_fp16 = cast(dtype = position10_to_fp16_dtype_0, x = position10)[name = tensor("cast_611")]; + tensor var_4375_cast_fp16 = reshape(shape = var_4374, x = position10_to_fp16)[name = tensor("op_4375_cast_fp16")]; + tensor ts_63_cast_fp16 = add(x = ts_1_promoted_to_fp16, y = var_4375_cast_fp16)[name = tensor("ts_63_cast_fp16")]; + tensor var_4382 = const()[name = tensor("op_4382"), val = tensor([1, 256, 1, 1])]; + tensor ts_65_cast_fp16 = reshape(shape = var_4382, x = ts_63_cast_fp16)[name = tensor("ts_65_cast_fp16")]; + tensor var_4389 = const()[name = tensor("op_4389"), val = tensor([1, 256, 16, 32, 2])]; + tensor q_complex_21_cast_fp16 = reshape(shape = var_4389, x = q_61_cast_fp16)[name = tensor("q_complex_21_cast_fp16")]; + tensor var_4396 = const()[name = tensor("op_4396"), val = tensor([1, 256, 16, 32, 2])]; + tensor k_complex_21_cast_fp16 = reshape(shape = var_4396, x = k_41_cast_fp16)[name = tensor("k_complex_21_cast_fp16")]; + tensor var_4400_begin_0 = const()[name = tensor("op_4400_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_4400_end_0 = const()[name = tensor("op_4400_end_0"), val = tensor([1, 256, 16, 32, 1])]; + tensor var_4400_end_mask_0 = const()[name = tensor("op_4400_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_4400_squeeze_mask_0 = const()[name = tensor("op_4400_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_4400_cast_fp16 = slice_by_index(begin = var_4400_begin_0, end = var_4400_end_0, end_mask = var_4400_end_mask_0, squeeze_mask = var_4400_squeeze_mask_0, x = q_complex_21_cast_fp16)[name = tensor("op_4400_cast_fp16")]; + tensor var_4408_begin_0 = const()[name = tensor("op_4408_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_4408_end_0 = const()[name = tensor("op_4408_end_0"), val = tensor([1, 256, 16, 32, 2])]; + tensor var_4408_end_mask_0 = const()[name = tensor("op_4408_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_4408_squeeze_mask_0 = const()[name = tensor("op_4408_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_4408_cast_fp16 = slice_by_index(begin = var_4408_begin_0, end = var_4408_end_0, end_mask = var_4408_end_mask_0, squeeze_mask = var_4408_squeeze_mask_0, x = q_complex_21_cast_fp16)[name = tensor("op_4408_cast_fp16")]; + tensor var_4416_begin_0 = const()[name = tensor("op_4416_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_4416_end_0 = const()[name = tensor("op_4416_end_0"), val = tensor([1, 256, 16, 32, 1])]; + tensor var_4416_end_mask_0 = const()[name = tensor("op_4416_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_4416_squeeze_mask_0 = const()[name = tensor("op_4416_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_4416_cast_fp16 = slice_by_index(begin = var_4416_begin_0, end = var_4416_end_0, end_mask = var_4416_end_mask_0, squeeze_mask = var_4416_squeeze_mask_0, x = k_complex_21_cast_fp16)[name = tensor("op_4416_cast_fp16")]; + tensor var_4424_begin_0 = const()[name = tensor("op_4424_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_4424_end_0 = const()[name = tensor("op_4424_end_0"), val = tensor([1, 256, 16, 32, 2])]; + tensor var_4424_end_mask_0 = const()[name = tensor("op_4424_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_4424_squeeze_mask_0 = const()[name = tensor("op_4424_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_4424_cast_fp16 = slice_by_index(begin = var_4424_begin_0, end = var_4424_end_0, end_mask = var_4424_end_mask_0, squeeze_mask = var_4424_squeeze_mask_0, x = k_complex_21_cast_fp16)[name = tensor("op_4424_cast_fp16")]; + tensor freqs_21_to_fp16 = const()[name = tensor("freqs_21_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(259110400)))]; + tensor var_4430_cast_fp16 = mul(x = freqs_21_to_fp16, y = ts_65_cast_fp16)[name = tensor("op_4430_cast_fp16")]; + tensor rotr_21_cast_fp16 = cos(x = var_4430_cast_fp16)[name = tensor("rotr_21_cast_fp16")]; + tensor roti_21_cast_fp16 = sin(x = var_4430_cast_fp16)[name = tensor("roti_21_cast_fp16")]; + tensor var_4434_cast_fp16 = mul(x = var_4400_cast_fp16, y = rotr_21_cast_fp16)[name = tensor("op_4434_cast_fp16")]; + tensor var_4435_cast_fp16 = mul(x = var_4408_cast_fp16, y = roti_21_cast_fp16)[name = tensor("op_4435_cast_fp16")]; + tensor qor_41_cast_fp16 = sub(x = var_4434_cast_fp16, y = var_4435_cast_fp16)[name = tensor("qor_41_cast_fp16")]; + tensor var_4438_cast_fp16 = mul(x = var_4400_cast_fp16, y = roti_21_cast_fp16)[name = tensor("op_4438_cast_fp16")]; + tensor var_4439_cast_fp16 = mul(x = var_4408_cast_fp16, y = rotr_21_cast_fp16)[name = tensor("op_4439_cast_fp16")]; + tensor qoi_41_cast_fp16 = add(x = var_4438_cast_fp16, y = var_4439_cast_fp16)[name = tensor("qoi_41_cast_fp16")]; + tensor var_4442_cast_fp16 = mul(x = var_4416_cast_fp16, y = rotr_21_cast_fp16)[name = tensor("op_4442_cast_fp16")]; + tensor var_4443_cast_fp16 = mul(x = var_4424_cast_fp16, y = roti_21_cast_fp16)[name = tensor("op_4443_cast_fp16")]; + tensor kor_41_cast_fp16 = sub(x = var_4442_cast_fp16, y = var_4443_cast_fp16)[name = tensor("kor_41_cast_fp16")]; + tensor var_4446_cast_fp16 = mul(x = var_4416_cast_fp16, y = roti_21_cast_fp16)[name = tensor("op_4446_cast_fp16")]; + tensor var_4447_cast_fp16 = mul(x = var_4424_cast_fp16, y = rotr_21_cast_fp16)[name = tensor("op_4447_cast_fp16")]; + tensor koi_41_cast_fp16 = add(x = var_4446_cast_fp16, y = var_4447_cast_fp16)[name = tensor("koi_41_cast_fp16")]; + tensor qo_21_axis_0 = const()[name = tensor("qo_21_axis_0"), val = tensor(-1)]; + tensor qo_21_cast_fp16 = stack(axis = qo_21_axis_0, values = (qor_41_cast_fp16, qoi_41_cast_fp16))[name = tensor("qo_21_cast_fp16")]; + tensor ko_21_axis_0 = const()[name = tensor("ko_21_axis_0"), val = tensor(-1)]; + tensor ko_21_cast_fp16 = stack(axis = ko_21_axis_0, values = (kor_41_cast_fp16, koi_41_cast_fp16))[name = tensor("ko_21_cast_fp16")]; + tensor var_4480 = const()[name = tensor("op_4480"), val = tensor([1, 256, 16, 64])]; + tensor q_63_cast_fp16 = reshape(shape = var_4480, x = qo_21_cast_fp16)[name = tensor("q_63_cast_fp16")]; + tensor var_4486 = const()[name = tensor("op_4486"), val = tensor([1, 256, 16, 64])]; + tensor k_43_cast_fp16 = reshape(shape = var_4486, x = ko_21_cast_fp16)[name = tensor("k_43_cast_fp16")]; + tensor write_indices_float_61_cast_fp16 = add(x = var_4375_cast_fp16, y = var_393_promoted_to_fp16)[name = tensor("write_indices_float_61_cast_fp16")]; + tensor _inversed_4507_y_0_to_fp16 = const()[name = tensor("_inversed_4507_y_0_to_fp16"), val = tensor(0x1p-9)]; + tensor _inversed_4507_cast_fp16 = mul(x = write_indices_float_61_cast_fp16, y = _inversed_4507_y_0_to_fp16)[name = tensor("_inversed_4507_cast_fp16")]; + tensor var_4508_cast_fp16 = floor(x = _inversed_4507_cast_fp16)[name = tensor("op_4508_cast_fp16")]; + tensor var_4509_to_fp16 = const()[name = tensor("op_4509_to_fp16"), val = tensor(0x1p+9)]; + tensor var_4510_cast_fp16 = mul(x = var_4508_cast_fp16, y = var_4509_to_fp16)[name = tensor("op_4510_cast_fp16")]; + tensor write_indices_float_63_cast_fp16 = sub(x = write_indices_float_61_cast_fp16, y = var_4510_cast_fp16)[name = tensor("write_indices_float_63_cast_fp16")]; + tensor var_4528_value_0_to_fp16 = const()[name = tensor("op_4528_value_0_to_fp16"), val = tensor(0x1.ffp+8)]; + tensor var_4528_cast_fp16 = fill_like(ref_tensor = write_indices_float_63_cast_fp16, value = var_4528_value_0_to_fp16)[name = tensor("op_4528_cast_fp16")]; + tensor write_indices_float_65_cast_fp16 = select(a = write_indices_float_63_cast_fp16, b = var_4528_cast_fp16, cond = token_valid_1_cast_fp16)[name = tensor("write_indices_float_65_cast_fp16")]; + tensor var_4534_dtype_0 = const()[name = tensor("op_4534_dtype_0"), val = tensor("int32")]; + tensor var_4539 = const()[name = tensor("op_4539"), val = tensor([1, 256, 1, 1])]; + tensor write_indices_float_65_cast_fp16_to_int32 = cast(dtype = var_4534_dtype_0, x = write_indices_float_65_cast_fp16)[name = tensor("cast_610")]; + tensor var_4540 = reshape(shape = var_4539, x = write_indices_float_65_cast_fp16_to_int32)[name = tensor("op_4540")]; + tensor write_indices_21_reps_0 = const()[name = tensor("write_indices_21_reps_0"), val = tensor([1, 1, 16, 64])]; + tensor write_indices_21 = tile(reps = write_indices_21_reps_0, x = var_4540)[name = tensor("write_indices_21")]; + tensor var_4550_begin_0 = const()[name = tensor("op_4550_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_4550_end_0 = const()[name = tensor("op_4550_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor var_4550_end_mask_0 = const()[name = tensor("op_4550_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_4550_squeeze_mask_0 = const()[name = tensor("op_4550_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor cache10_to_fp16_dtype_0 = const()[name = tensor("cache10_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor cache10_to_fp16 = cast(dtype = cache10_to_fp16_dtype_0, x = cache10)[name = tensor("cast_609")]; + tensor var_4550_cast_fp16 = slice_by_index(begin = var_4550_begin_0, end = var_4550_end_0, end_mask = var_4550_end_mask_0, squeeze_mask = var_4550_squeeze_mask_0, x = cache10_to_fp16)[name = tensor("op_4550_cast_fp16")]; + tensor var_4552_axis_0 = const()[name = tensor("op_4552_axis_0"), val = tensor(1)]; + tensor var_4552_mode_0 = const()[name = tensor("op_4552_mode_0"), val = tensor("update")]; + tensor var_4552_validate_indices_0 = const()[name = tensor("op_4552_validate_indices_0"), val = tensor(false)]; + tensor var_4552_cast_fp16 = scatter_along_axis(axis = var_4552_axis_0, data = var_4550_cast_fp16, indices = write_indices_21, mode = var_4552_mode_0, updates = k_43_cast_fp16, validate_indices = var_4552_validate_indices_0)[name = tensor("op_4552_cast_fp16")]; + tensor concat_82 = const()[name = tensor("concat_82"), val = tensor([0, 0, 0, 0, 0])]; + tensor concat_83 = const()[name = tensor("concat_83"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_21_internal_tensor_assign_1_stride_0 = const()[name = tensor("new_cache_21_internal_tensor_assign_1_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_21_internal_tensor_assign_1_begin_mask_0 = const()[name = tensor("new_cache_21_internal_tensor_assign_1_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_21_internal_tensor_assign_1_end_mask_0 = const()[name = tensor("new_cache_21_internal_tensor_assign_1_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_21_internal_tensor_assign_1_squeeze_mask_0 = const()[name = tensor("new_cache_21_internal_tensor_assign_1_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_66 = const()[name = tensor("shape_66"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_20 = const()[name = tensor("reduce_prod_20"), val = tensor(1048576)]; + tensor range_1d_20_start_0 = const()[name = tensor("range_1d_20_start_0"), val = tensor(0)]; + tensor range_1d_20_step_0 = const()[name = tensor("range_1d_20_step_0"), val = tensor(1)]; + tensor range_1d_20 = range_1d(end = reduce_prod_20, start = range_1d_20_start_0, step = range_1d_20_step_0)[name = tensor("range_1d_20")]; + tensor reshape_100 = reshape(shape = shape_66, x = range_1d_20)[name = tensor("reshape_100")]; + tensor slice_by_index_20 = slice_by_index(begin = concat_82, begin_mask = new_cache_21_internal_tensor_assign_1_begin_mask_0, end = concat_83, end_mask = new_cache_21_internal_tensor_assign_1_end_mask_0, squeeze_mask = new_cache_21_internal_tensor_assign_1_squeeze_mask_0, stride = new_cache_21_internal_tensor_assign_1_stride_0, x = reshape_100)[name = tensor("slice_by_index_20")]; + tensor reshape_101_shape_0 = const()[name = tensor("reshape_101_shape_0"), val = tensor([-1])]; + tensor reshape_101 = reshape(shape = reshape_101_shape_0, x = slice_by_index_20)[name = tensor("reshape_101")]; + tensor reshape_102_shape_0 = const()[name = tensor("reshape_102_shape_0"), val = tensor([-1])]; + tensor reshape_102_cast_fp16 = reshape(shape = reshape_102_shape_0, x = var_4552_cast_fp16)[name = tensor("reshape_102_cast_fp16")]; + tensor reshape_103_shape_0 = const()[name = tensor("reshape_103_shape_0"), val = tensor([-1])]; + tensor reshape_103_cast_fp16 = reshape(shape = reshape_103_shape_0, x = cache10_to_fp16)[name = tensor("reshape_103_cast_fp16")]; + tensor scatter_20_mode_0 = const()[name = tensor("scatter_20_mode_0"), val = tensor("update")]; + tensor scatter_20_axis_0 = const()[name = tensor("scatter_20_axis_0"), val = tensor(0)]; + tensor scatter_20_validate_indices_0 = const()[name = tensor("scatter_20_validate_indices_0"), val = tensor(false)]; + tensor scatter_20_cast_fp16 = scatter(axis = scatter_20_axis_0, data = reshape_103_cast_fp16, indices = reshape_101, mode = scatter_20_mode_0, updates = reshape_102_cast_fp16, validate_indices = scatter_20_validate_indices_0)[name = tensor("scatter_20_cast_fp16")]; + tensor reshape_104_cast_fp16 = reshape(shape = shape_66, x = scatter_20_cast_fp16)[name = tensor("reshape_104_cast_fp16")]; + tensor var_4560_begin_0 = const()[name = tensor("op_4560_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor var_4560_end_0 = const()[name = tensor("op_4560_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor var_4560_end_mask_0 = const()[name = tensor("op_4560_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_4560_squeeze_mask_0 = const()[name = tensor("op_4560_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor var_4560_cast_fp16 = slice_by_index(begin = var_4560_begin_0, end = var_4560_end_0, end_mask = var_4560_end_mask_0, squeeze_mask = var_4560_squeeze_mask_0, x = reshape_104_cast_fp16)[name = tensor("op_4560_cast_fp16")]; + tensor var_4562_axis_0 = const()[name = tensor("op_4562_axis_0"), val = tensor(1)]; + tensor var_4562_mode_0 = const()[name = tensor("op_4562_mode_0"), val = tensor("update")]; + tensor var_4562_validate_indices_0 = const()[name = tensor("op_4562_validate_indices_0"), val = tensor(false)]; + tensor var_4562_cast_fp16 = scatter_along_axis(axis = var_4562_axis_0, data = var_4560_cast_fp16, indices = write_indices_21, mode = var_4562_mode_0, updates = v_21_cast_fp16, validate_indices = var_4562_validate_indices_0)[name = tensor("op_4562_cast_fp16")]; + tensor concat_84 = const()[name = tensor("concat_84"), val = tensor([1, 0, 0, 0, 0])]; + tensor concat_85 = const()[name = tensor("concat_85"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_21_internal_tensor_assign_2_stride_0 = const()[name = tensor("new_cache_21_internal_tensor_assign_2_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_21_internal_tensor_assign_2_begin_mask_0 = const()[name = tensor("new_cache_21_internal_tensor_assign_2_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_21_internal_tensor_assign_2_end_mask_0 = const()[name = tensor("new_cache_21_internal_tensor_assign_2_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_21_internal_tensor_assign_2_squeeze_mask_0 = const()[name = tensor("new_cache_21_internal_tensor_assign_2_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_67 = const()[name = tensor("shape_67"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_21 = const()[name = tensor("reduce_prod_21"), val = tensor(1048576)]; + tensor range_1d_21_start_0 = const()[name = tensor("range_1d_21_start_0"), val = tensor(0)]; + tensor range_1d_21_step_0 = const()[name = tensor("range_1d_21_step_0"), val = tensor(1)]; + tensor range_1d_21 = range_1d(end = reduce_prod_21, start = range_1d_21_start_0, step = range_1d_21_step_0)[name = tensor("range_1d_21")]; + tensor reshape_105 = reshape(shape = shape_67, x = range_1d_21)[name = tensor("reshape_105")]; + tensor slice_by_index_21 = slice_by_index(begin = concat_84, begin_mask = new_cache_21_internal_tensor_assign_2_begin_mask_0, end = concat_85, end_mask = new_cache_21_internal_tensor_assign_2_end_mask_0, squeeze_mask = new_cache_21_internal_tensor_assign_2_squeeze_mask_0, stride = new_cache_21_internal_tensor_assign_2_stride_0, x = reshape_105)[name = tensor("slice_by_index_21")]; + tensor reshape_106_shape_0 = const()[name = tensor("reshape_106_shape_0"), val = tensor([-1])]; + tensor reshape_106 = reshape(shape = reshape_106_shape_0, x = slice_by_index_21)[name = tensor("reshape_106")]; + tensor reshape_107_shape_0 = const()[name = tensor("reshape_107_shape_0"), val = tensor([-1])]; + tensor reshape_107_cast_fp16 = reshape(shape = reshape_107_shape_0, x = var_4562_cast_fp16)[name = tensor("reshape_107_cast_fp16")]; + tensor reshape_108_shape_0 = const()[name = tensor("reshape_108_shape_0"), val = tensor([-1])]; + tensor reshape_108_cast_fp16 = reshape(shape = reshape_108_shape_0, x = reshape_104_cast_fp16)[name = tensor("reshape_108_cast_fp16")]; + tensor scatter_21_mode_0 = const()[name = tensor("scatter_21_mode_0"), val = tensor("update")]; + tensor scatter_21_axis_0 = const()[name = tensor("scatter_21_axis_0"), val = tensor(0)]; + tensor scatter_21_validate_indices_0 = const()[name = tensor("scatter_21_validate_indices_0"), val = tensor(false)]; + tensor scatter_21_cast_fp16 = scatter(axis = scatter_21_axis_0, data = reshape_108_cast_fp16, indices = reshape_106, mode = scatter_21_mode_0, updates = reshape_107_cast_fp16, validate_indices = scatter_21_validate_indices_0)[name = tensor("scatter_21_cast_fp16")]; + tensor reshape_109_cast_fp16 = reshape(shape = shape_67, x = scatter_21_cast_fp16)[name = tensor("reshape_109_cast_fp16")]; + tensor reshape_109_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("reshape_109_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor keys_61_begin_0 = const()[name = tensor("keys_61_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor keys_61_end_0 = const()[name = tensor("keys_61_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor keys_61_end_mask_0 = const()[name = tensor("keys_61_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor keys_61_squeeze_mask_0 = const()[name = tensor("keys_61_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor keys_61_cast_fp16 = slice_by_index(begin = keys_61_begin_0, end = keys_61_end_0, end_mask = keys_61_end_mask_0, squeeze_mask = keys_61_squeeze_mask_0, x = reshape_109_cast_fp16)[name = tensor("keys_61_cast_fp16")]; + tensor values_61_begin_0 = const()[name = tensor("values_61_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor values_61_end_0 = const()[name = tensor("values_61_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor values_61_end_mask_0 = const()[name = tensor("values_61_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor values_61_squeeze_mask_0 = const()[name = tensor("values_61_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor values_61_cast_fp16 = slice_by_index(begin = values_61_begin_0, end = values_61_end_0, end_mask = values_61_end_mask_0, squeeze_mask = values_61_squeeze_mask_0, x = reshape_109_cast_fp16)[name = tensor("values_61_cast_fp16")]; + tensor var_4574_cast_fp16 = not_equal(x = keys_61_cast_fp16, y = keys_61_cast_fp16)[name = tensor("op_4574_cast_fp16")]; + tensor keys_63_cast_fp16 = select(a = var_470_to_fp16, b = keys_61_cast_fp16, cond = var_4574_cast_fp16)[name = tensor("keys_63_cast_fp16")]; + tensor var_4582_cast_fp16 = not_equal(x = values_61_cast_fp16, y = values_61_cast_fp16)[name = tensor("op_4582_cast_fp16")]; + tensor values_63_cast_fp16 = select(a = var_470_to_fp16, b = values_61_cast_fp16, cond = var_4582_cast_fp16)[name = tensor("values_63_cast_fp16")]; + tensor var_4606 = const()[name = tensor("op_4606"), val = tensor([0, 2, 1, 3])]; + tensor var_4622 = const()[name = tensor("op_4622"), val = tensor([1, 1, 1])]; + tensor var_4623_cast_fp16 = reshape(shape = var_4622, x = position10_to_fp16)[name = tensor("op_4623_cast_fp16")]; + tensor q_positions_21_cast_fp16 = add(x = var_4623_cast_fp16, y = q_offsets_1_promoted_to_fp16)[name = tensor("q_positions_21_cast_fp16")]; + tensor valid_end_21_cast_fp16 = add(x = var_4623_cast_fp16, y = var_536_cast_fp16)[name = tensor("valid_end_21_cast_fp16")]; + tensor valid_mask_21_cast_fp16 = less(x = k_positions_1_promoted_to_fp16, y = valid_end_21_cast_fp16)[name = tensor("valid_mask_21_cast_fp16")]; + tensor causal_mask_21_cast_fp16 = less_equal(x = k_positions_1_promoted_to_fp16, y = q_positions_21_cast_fp16)[name = tensor("causal_mask_21_cast_fp16")]; + tensor var_4651 = logical_and(x = valid_mask_21_cast_fp16, y = causal_mask_21_cast_fp16)[name = tensor("op_4651")]; + tensor attn_mask_21_axes_0 = const()[name = tensor("attn_mask_21_axes_0"), val = tensor([1])]; + tensor attn_mask_21 = expand_dims(axes = attn_mask_21_axes_0, x = var_4651)[name = tensor("attn_mask_21")]; + tensor var_4665_transpose_x_0 = const()[name = tensor("op_4665_transpose_x_0"), val = tensor(false)]; + tensor var_4665_transpose_y_0 = const()[name = tensor("op_4665_transpose_y_0"), val = tensor(false)]; + tensor transpose_89_perm_0 = const()[name = tensor("transpose_89_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_90_perm_0 = const()[name = tensor("transpose_90_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_90 = transpose(perm = transpose_90_perm_0, x = keys_63_cast_fp16)[name = tensor("transpose_164")]; + tensor transpose_89 = transpose(perm = transpose_89_perm_0, x = q_63_cast_fp16)[name = tensor("transpose_165")]; + tensor var_4665_cast_fp16 = matmul(transpose_x = var_4665_transpose_x_0, transpose_y = var_4665_transpose_y_0, x = transpose_89, y = transpose_90)[name = tensor("op_4665_cast_fp16")]; + tensor var_4659_to_fp16 = const()[name = tensor("op_4659_to_fp16"), val = tensor([0x1p-3])]; + tensor attn_weights_61_cast_fp16 = mul(x = var_4665_cast_fp16, y = var_4659_to_fp16)[name = tensor("attn_weights_61_cast_fp16")]; + tensor var_4667 = logical_not(x = attn_mask_21)[name = tensor("op_4667")]; + tensor var_4668_to_fp16 = const()[name = tensor("op_4668_to_fp16"), val = tensor(-inf)]; + tensor attn_weights_63_cast_fp16 = select(a = var_4668_to_fp16, b = attn_weights_61_cast_fp16, cond = var_4667)[name = tensor("attn_weights_63_cast_fp16")]; + tensor var_4670 = const()[name = tensor("op_4670"), val = tensor(-1)]; + tensor attn_weights_65_cast_fp16 = softmax(axis = var_4670, x = attn_weights_63_cast_fp16)[name = tensor("attn_weights_65_cast_fp16")]; + tensor attn_output_21_transpose_x_0 = const()[name = tensor("attn_output_21_transpose_x_0"), val = tensor(false)]; + tensor attn_output_21_transpose_y_0 = const()[name = tensor("attn_output_21_transpose_y_0"), val = tensor(false)]; + tensor values_65_cast_fp16 = transpose(perm = var_4606, x = values_63_cast_fp16)[name = tensor("transpose_166")]; + tensor attn_output_21_cast_fp16 = matmul(transpose_x = attn_output_21_transpose_x_0, transpose_y = attn_output_21_transpose_y_0, x = attn_weights_65_cast_fp16, y = values_65_cast_fp16)[name = tensor("attn_output_21_cast_fp16")]; + tensor var_4678 = const()[name = tensor("op_4678"), val = tensor([0, 2, 1, 3])]; + tensor var_4683 = const()[name = tensor("op_4683"), val = tensor([1, 256, 1024])]; + tensor var_4679_cast_fp16 = transpose(perm = var_4678, x = attn_output_21_cast_fp16)[name = tensor("transpose_163")]; + tensor input_125_cast_fp16 = reshape(shape = var_4683, x = var_4679_cast_fp16)[name = tensor("input_125_cast_fp16")]; + tensor attn10_out_proj_weight_to_fp16 = const()[name = tensor("attn10_out_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(259110528)))]; + tensor linear_41_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = attn10_out_proj_weight_to_fp16, x = input_125_cast_fp16)[name = tensor("linear_41_cast_fp16")]; + tensor var_4692_cast_fp16 = add(x = position10_to_fp16, y = valid_len_to_fp16)[name = tensor("op_4692_cast_fp16")]; + tensor var_4692_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_4692_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor input_127_cast_fp16 = add(x = input_121_cast_fp16, y = linear_41_cast_fp16)[name = tensor("input_127_cast_fp16")]; + tensor input_129_axes_0 = const()[name = tensor("input_129_axes_0"), val = tensor([-1])]; + tensor norm10_2_weight_to_fp16 = const()[name = tensor("norm10_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(261207744)))]; + tensor norm10_2_bias_to_fp16 = const()[name = tensor("norm10_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(261209856)))]; + tensor var_4696_to_fp16 = const()[name = tensor("op_4696_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_129_cast_fp16 = layer_norm(axes = input_129_axes_0, beta = norm10_2_bias_to_fp16, epsilon = var_4696_to_fp16, gamma = norm10_2_weight_to_fp16, x = input_127_cast_fp16)[name = tensor("input_129_cast_fp16")]; + tensor linear10_1_weight_to_fp16 = const()[name = tensor("linear10_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(261211968)))]; + tensor linear_42_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = linear10_1_weight_to_fp16, x = input_129_cast_fp16)[name = tensor("linear_42_cast_fp16")]; + tensor input_131_mode_0 = const()[name = tensor("input_131_mode_0"), val = tensor("EXACT")]; + tensor input_131_cast_fp16 = gelu(mode = input_131_mode_0, x = linear_42_cast_fp16)[name = tensor("input_131_cast_fp16")]; + tensor linear10_2_weight_to_fp16 = const()[name = tensor("linear10_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(269600640)))]; + tensor linear_43_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = linear10_2_weight_to_fp16, x = input_131_cast_fp16)[name = tensor("linear_43_cast_fp16")]; + tensor input_133_cast_fp16 = add(x = input_127_cast_fp16, y = linear_43_cast_fp16)[name = tensor("input_133_cast_fp16")]; + tensor input_135_axes_0 = const()[name = tensor("input_135_axes_0"), val = tensor([-1])]; + tensor norm11_1_weight_to_fp16 = const()[name = tensor("norm11_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(277989312)))]; + tensor norm11_1_bias_to_fp16 = const()[name = tensor("norm11_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(277991424)))]; + tensor var_4713_to_fp16 = const()[name = tensor("op_4713_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_135_cast_fp16 = layer_norm(axes = input_135_axes_0, beta = norm11_1_bias_to_fp16, epsilon = var_4713_to_fp16, gamma = norm11_1_weight_to_fp16, x = input_133_cast_fp16)[name = tensor("input_135_cast_fp16")]; + tensor attn11_in_proj_weight_to_fp16 = const()[name = tensor("attn11_in_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(277993536)))]; + tensor linear_44_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = attn11_in_proj_weight_to_fp16, x = input_135_cast_fp16)[name = tensor("linear_44_cast_fp16")]; + tensor var_4727 = const()[name = tensor("op_4727"), val = tensor([1, 256, 3, 16, 64])]; + tensor qkv_23_cast_fp16 = reshape(shape = var_4727, x = linear_44_cast_fp16)[name = tensor("qkv_23_cast_fp16")]; + tensor q_67_begin_0 = const()[name = tensor("q_67_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor q_67_end_0 = const()[name = tensor("q_67_end_0"), val = tensor([1, 256, 1, 16, 64])]; + tensor q_67_end_mask_0 = const()[name = tensor("q_67_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor q_67_squeeze_mask_0 = const()[name = tensor("q_67_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor q_67_cast_fp16 = slice_by_index(begin = q_67_begin_0, end = q_67_end_0, end_mask = q_67_end_mask_0, squeeze_mask = q_67_squeeze_mask_0, x = qkv_23_cast_fp16)[name = tensor("q_67_cast_fp16")]; + tensor k_45_begin_0 = const()[name = tensor("k_45_begin_0"), val = tensor([0, 0, 1, 0, 0])]; + tensor k_45_end_0 = const()[name = tensor("k_45_end_0"), val = tensor([1, 256, 2, 16, 64])]; + tensor k_45_end_mask_0 = const()[name = tensor("k_45_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor k_45_squeeze_mask_0 = const()[name = tensor("k_45_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor k_45_cast_fp16 = slice_by_index(begin = k_45_begin_0, end = k_45_end_0, end_mask = k_45_end_mask_0, squeeze_mask = k_45_squeeze_mask_0, x = qkv_23_cast_fp16)[name = tensor("k_45_cast_fp16")]; + tensor v_23_begin_0 = const()[name = tensor("v_23_begin_0"), val = tensor([0, 0, 2, 0, 0])]; + tensor v_23_end_0 = const()[name = tensor("v_23_end_0"), val = tensor([1, 256, 3, 16, 64])]; + tensor v_23_end_mask_0 = const()[name = tensor("v_23_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor v_23_squeeze_mask_0 = const()[name = tensor("v_23_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor v_23_cast_fp16 = slice_by_index(begin = v_23_begin_0, end = v_23_end_0, end_mask = v_23_end_mask_0, squeeze_mask = v_23_squeeze_mask_0, x = qkv_23_cast_fp16)[name = tensor("v_23_cast_fp16")]; + tensor var_4785 = const()[name = tensor("op_4785"), val = tensor([1, 1])]; + tensor position11_to_fp16_dtype_0 = const()[name = tensor("position11_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor position11_to_fp16 = cast(dtype = position11_to_fp16_dtype_0, x = position11)[name = tensor("cast_606")]; + tensor var_4786_cast_fp16 = reshape(shape = var_4785, x = position11_to_fp16)[name = tensor("op_4786_cast_fp16")]; + tensor ts_69_cast_fp16 = add(x = ts_1_promoted_to_fp16, y = var_4786_cast_fp16)[name = tensor("ts_69_cast_fp16")]; + tensor var_4793 = const()[name = tensor("op_4793"), val = tensor([1, 256, 1, 1])]; + tensor ts_71_cast_fp16 = reshape(shape = var_4793, x = ts_69_cast_fp16)[name = tensor("ts_71_cast_fp16")]; + tensor var_4800 = const()[name = tensor("op_4800"), val = tensor([1, 256, 16, 32, 2])]; + tensor q_complex_23_cast_fp16 = reshape(shape = var_4800, x = q_67_cast_fp16)[name = tensor("q_complex_23_cast_fp16")]; + tensor var_4807 = const()[name = tensor("op_4807"), val = tensor([1, 256, 16, 32, 2])]; + tensor k_complex_23_cast_fp16 = reshape(shape = var_4807, x = k_45_cast_fp16)[name = tensor("k_complex_23_cast_fp16")]; + tensor var_4811_begin_0 = const()[name = tensor("op_4811_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_4811_end_0 = const()[name = tensor("op_4811_end_0"), val = tensor([1, 256, 16, 32, 1])]; + tensor var_4811_end_mask_0 = const()[name = tensor("op_4811_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_4811_squeeze_mask_0 = const()[name = tensor("op_4811_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_4811_cast_fp16 = slice_by_index(begin = var_4811_begin_0, end = var_4811_end_0, end_mask = var_4811_end_mask_0, squeeze_mask = var_4811_squeeze_mask_0, x = q_complex_23_cast_fp16)[name = tensor("op_4811_cast_fp16")]; + tensor var_4819_begin_0 = const()[name = tensor("op_4819_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_4819_end_0 = const()[name = tensor("op_4819_end_0"), val = tensor([1, 256, 16, 32, 2])]; + tensor var_4819_end_mask_0 = const()[name = tensor("op_4819_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_4819_squeeze_mask_0 = const()[name = tensor("op_4819_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_4819_cast_fp16 = slice_by_index(begin = var_4819_begin_0, end = var_4819_end_0, end_mask = var_4819_end_mask_0, squeeze_mask = var_4819_squeeze_mask_0, x = q_complex_23_cast_fp16)[name = tensor("op_4819_cast_fp16")]; + tensor var_4827_begin_0 = const()[name = tensor("op_4827_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_4827_end_0 = const()[name = tensor("op_4827_end_0"), val = tensor([1, 256, 16, 32, 1])]; + tensor var_4827_end_mask_0 = const()[name = tensor("op_4827_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_4827_squeeze_mask_0 = const()[name = tensor("op_4827_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_4827_cast_fp16 = slice_by_index(begin = var_4827_begin_0, end = var_4827_end_0, end_mask = var_4827_end_mask_0, squeeze_mask = var_4827_squeeze_mask_0, x = k_complex_23_cast_fp16)[name = tensor("op_4827_cast_fp16")]; + tensor var_4835_begin_0 = const()[name = tensor("op_4835_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_4835_end_0 = const()[name = tensor("op_4835_end_0"), val = tensor([1, 256, 16, 32, 2])]; + tensor var_4835_end_mask_0 = const()[name = tensor("op_4835_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_4835_squeeze_mask_0 = const()[name = tensor("op_4835_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_4835_cast_fp16 = slice_by_index(begin = var_4835_begin_0, end = var_4835_end_0, end_mask = var_4835_end_mask_0, squeeze_mask = var_4835_squeeze_mask_0, x = k_complex_23_cast_fp16)[name = tensor("op_4835_cast_fp16")]; + tensor freqs_23_to_fp16 = const()[name = tensor("freqs_23_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(284285056)))]; + tensor var_4841_cast_fp16 = mul(x = freqs_23_to_fp16, y = ts_71_cast_fp16)[name = tensor("op_4841_cast_fp16")]; + tensor rotr_23_cast_fp16 = cos(x = var_4841_cast_fp16)[name = tensor("rotr_23_cast_fp16")]; + tensor roti_23_cast_fp16 = sin(x = var_4841_cast_fp16)[name = tensor("roti_23_cast_fp16")]; + tensor var_4845_cast_fp16 = mul(x = var_4811_cast_fp16, y = rotr_23_cast_fp16)[name = tensor("op_4845_cast_fp16")]; + tensor var_4846_cast_fp16 = mul(x = var_4819_cast_fp16, y = roti_23_cast_fp16)[name = tensor("op_4846_cast_fp16")]; + tensor qor_45_cast_fp16 = sub(x = var_4845_cast_fp16, y = var_4846_cast_fp16)[name = tensor("qor_45_cast_fp16")]; + tensor var_4849_cast_fp16 = mul(x = var_4811_cast_fp16, y = roti_23_cast_fp16)[name = tensor("op_4849_cast_fp16")]; + tensor var_4850_cast_fp16 = mul(x = var_4819_cast_fp16, y = rotr_23_cast_fp16)[name = tensor("op_4850_cast_fp16")]; + tensor qoi_45_cast_fp16 = add(x = var_4849_cast_fp16, y = var_4850_cast_fp16)[name = tensor("qoi_45_cast_fp16")]; + tensor var_4853_cast_fp16 = mul(x = var_4827_cast_fp16, y = rotr_23_cast_fp16)[name = tensor("op_4853_cast_fp16")]; + tensor var_4854_cast_fp16 = mul(x = var_4835_cast_fp16, y = roti_23_cast_fp16)[name = tensor("op_4854_cast_fp16")]; + tensor kor_45_cast_fp16 = sub(x = var_4853_cast_fp16, y = var_4854_cast_fp16)[name = tensor("kor_45_cast_fp16")]; + tensor var_4857_cast_fp16 = mul(x = var_4827_cast_fp16, y = roti_23_cast_fp16)[name = tensor("op_4857_cast_fp16")]; + tensor var_4858_cast_fp16 = mul(x = var_4835_cast_fp16, y = rotr_23_cast_fp16)[name = tensor("op_4858_cast_fp16")]; + tensor koi_45_cast_fp16 = add(x = var_4857_cast_fp16, y = var_4858_cast_fp16)[name = tensor("koi_45_cast_fp16")]; + tensor qo_23_axis_0 = const()[name = tensor("qo_23_axis_0"), val = tensor(-1)]; + tensor qo_23_cast_fp16 = stack(axis = qo_23_axis_0, values = (qor_45_cast_fp16, qoi_45_cast_fp16))[name = tensor("qo_23_cast_fp16")]; + tensor ko_23_axis_0 = const()[name = tensor("ko_23_axis_0"), val = tensor(-1)]; + tensor ko_23_cast_fp16 = stack(axis = ko_23_axis_0, values = (kor_45_cast_fp16, koi_45_cast_fp16))[name = tensor("ko_23_cast_fp16")]; + tensor var_4891 = const()[name = tensor("op_4891"), val = tensor([1, 256, 16, 64])]; + tensor q_69_cast_fp16 = reshape(shape = var_4891, x = qo_23_cast_fp16)[name = tensor("q_69_cast_fp16")]; + tensor var_4897 = const()[name = tensor("op_4897"), val = tensor([1, 256, 16, 64])]; + tensor k_47_cast_fp16 = reshape(shape = var_4897, x = ko_23_cast_fp16)[name = tensor("k_47_cast_fp16")]; + tensor write_indices_float_67_cast_fp16 = add(x = var_4786_cast_fp16, y = var_393_promoted_to_fp16)[name = tensor("write_indices_float_67_cast_fp16")]; + tensor _inversed_4918_y_0_to_fp16 = const()[name = tensor("_inversed_4918_y_0_to_fp16"), val = tensor(0x1p-9)]; + tensor _inversed_4918_cast_fp16 = mul(x = write_indices_float_67_cast_fp16, y = _inversed_4918_y_0_to_fp16)[name = tensor("_inversed_4918_cast_fp16")]; + tensor var_4919_cast_fp16 = floor(x = _inversed_4918_cast_fp16)[name = tensor("op_4919_cast_fp16")]; + tensor var_4920_to_fp16 = const()[name = tensor("op_4920_to_fp16"), val = tensor(0x1p+9)]; + tensor var_4921_cast_fp16 = mul(x = var_4919_cast_fp16, y = var_4920_to_fp16)[name = tensor("op_4921_cast_fp16")]; + tensor write_indices_float_69_cast_fp16 = sub(x = write_indices_float_67_cast_fp16, y = var_4921_cast_fp16)[name = tensor("write_indices_float_69_cast_fp16")]; + tensor var_4939_value_0_to_fp16 = const()[name = tensor("op_4939_value_0_to_fp16"), val = tensor(0x1.ffp+8)]; + tensor var_4939_cast_fp16 = fill_like(ref_tensor = write_indices_float_69_cast_fp16, value = var_4939_value_0_to_fp16)[name = tensor("op_4939_cast_fp16")]; + tensor write_indices_float_71_cast_fp16 = select(a = write_indices_float_69_cast_fp16, b = var_4939_cast_fp16, cond = token_valid_1_cast_fp16)[name = tensor("write_indices_float_71_cast_fp16")]; + tensor var_4945_dtype_0 = const()[name = tensor("op_4945_dtype_0"), val = tensor("int32")]; + tensor var_4950 = const()[name = tensor("op_4950"), val = tensor([1, 256, 1, 1])]; + tensor write_indices_float_71_cast_fp16_to_int32 = cast(dtype = var_4945_dtype_0, x = write_indices_float_71_cast_fp16)[name = tensor("cast_605")]; + tensor var_4951 = reshape(shape = var_4950, x = write_indices_float_71_cast_fp16_to_int32)[name = tensor("op_4951")]; + tensor write_indices_23_reps_0 = const()[name = tensor("write_indices_23_reps_0"), val = tensor([1, 1, 16, 64])]; + tensor write_indices_23 = tile(reps = write_indices_23_reps_0, x = var_4951)[name = tensor("write_indices_23")]; + tensor var_4961_begin_0 = const()[name = tensor("op_4961_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_4961_end_0 = const()[name = tensor("op_4961_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor var_4961_end_mask_0 = const()[name = tensor("op_4961_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_4961_squeeze_mask_0 = const()[name = tensor("op_4961_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor cache11_to_fp16_dtype_0 = const()[name = tensor("cache11_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor cache11_to_fp16 = cast(dtype = cache11_to_fp16_dtype_0, x = cache11)[name = tensor("cast_604")]; + tensor var_4961_cast_fp16 = slice_by_index(begin = var_4961_begin_0, end = var_4961_end_0, end_mask = var_4961_end_mask_0, squeeze_mask = var_4961_squeeze_mask_0, x = cache11_to_fp16)[name = tensor("op_4961_cast_fp16")]; + tensor var_4963_axis_0 = const()[name = tensor("op_4963_axis_0"), val = tensor(1)]; + tensor var_4963_mode_0 = const()[name = tensor("op_4963_mode_0"), val = tensor("update")]; + tensor var_4963_validate_indices_0 = const()[name = tensor("op_4963_validate_indices_0"), val = tensor(false)]; + tensor var_4963_cast_fp16 = scatter_along_axis(axis = var_4963_axis_0, data = var_4961_cast_fp16, indices = write_indices_23, mode = var_4963_mode_0, updates = k_47_cast_fp16, validate_indices = var_4963_validate_indices_0)[name = tensor("op_4963_cast_fp16")]; + tensor concat_90 = const()[name = tensor("concat_90"), val = tensor([0, 0, 0, 0, 0])]; + tensor concat_91 = const()[name = tensor("concat_91"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_23_internal_tensor_assign_1_stride_0 = const()[name = tensor("new_cache_23_internal_tensor_assign_1_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_23_internal_tensor_assign_1_begin_mask_0 = const()[name = tensor("new_cache_23_internal_tensor_assign_1_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_23_internal_tensor_assign_1_end_mask_0 = const()[name = tensor("new_cache_23_internal_tensor_assign_1_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_23_internal_tensor_assign_1_squeeze_mask_0 = const()[name = tensor("new_cache_23_internal_tensor_assign_1_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_68 = const()[name = tensor("shape_68"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_22 = const()[name = tensor("reduce_prod_22"), val = tensor(1048576)]; + tensor range_1d_22_start_0 = const()[name = tensor("range_1d_22_start_0"), val = tensor(0)]; + tensor range_1d_22_step_0 = const()[name = tensor("range_1d_22_step_0"), val = tensor(1)]; + tensor range_1d_22 = range_1d(end = reduce_prod_22, start = range_1d_22_start_0, step = range_1d_22_step_0)[name = tensor("range_1d_22")]; + tensor reshape_110 = reshape(shape = shape_68, x = range_1d_22)[name = tensor("reshape_110")]; + tensor slice_by_index_22 = slice_by_index(begin = concat_90, begin_mask = new_cache_23_internal_tensor_assign_1_begin_mask_0, end = concat_91, end_mask = new_cache_23_internal_tensor_assign_1_end_mask_0, squeeze_mask = new_cache_23_internal_tensor_assign_1_squeeze_mask_0, stride = new_cache_23_internal_tensor_assign_1_stride_0, x = reshape_110)[name = tensor("slice_by_index_22")]; + tensor reshape_111_shape_0 = const()[name = tensor("reshape_111_shape_0"), val = tensor([-1])]; + tensor reshape_111 = reshape(shape = reshape_111_shape_0, x = slice_by_index_22)[name = tensor("reshape_111")]; + tensor reshape_112_shape_0 = const()[name = tensor("reshape_112_shape_0"), val = tensor([-1])]; + tensor reshape_112_cast_fp16 = reshape(shape = reshape_112_shape_0, x = var_4963_cast_fp16)[name = tensor("reshape_112_cast_fp16")]; + tensor reshape_113_shape_0 = const()[name = tensor("reshape_113_shape_0"), val = tensor([-1])]; + tensor reshape_113_cast_fp16 = reshape(shape = reshape_113_shape_0, x = cache11_to_fp16)[name = tensor("reshape_113_cast_fp16")]; + tensor scatter_22_mode_0 = const()[name = tensor("scatter_22_mode_0"), val = tensor("update")]; + tensor scatter_22_axis_0 = const()[name = tensor("scatter_22_axis_0"), val = tensor(0)]; + tensor scatter_22_validate_indices_0 = const()[name = tensor("scatter_22_validate_indices_0"), val = tensor(false)]; + tensor scatter_22_cast_fp16 = scatter(axis = scatter_22_axis_0, data = reshape_113_cast_fp16, indices = reshape_111, mode = scatter_22_mode_0, updates = reshape_112_cast_fp16, validate_indices = scatter_22_validate_indices_0)[name = tensor("scatter_22_cast_fp16")]; + tensor reshape_114_cast_fp16 = reshape(shape = shape_68, x = scatter_22_cast_fp16)[name = tensor("reshape_114_cast_fp16")]; + tensor var_4971_begin_0 = const()[name = tensor("op_4971_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor var_4971_end_0 = const()[name = tensor("op_4971_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor var_4971_end_mask_0 = const()[name = tensor("op_4971_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_4971_squeeze_mask_0 = const()[name = tensor("op_4971_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor var_4971_cast_fp16 = slice_by_index(begin = var_4971_begin_0, end = var_4971_end_0, end_mask = var_4971_end_mask_0, squeeze_mask = var_4971_squeeze_mask_0, x = reshape_114_cast_fp16)[name = tensor("op_4971_cast_fp16")]; + tensor var_4973_axis_0 = const()[name = tensor("op_4973_axis_0"), val = tensor(1)]; + tensor var_4973_mode_0 = const()[name = tensor("op_4973_mode_0"), val = tensor("update")]; + tensor var_4973_validate_indices_0 = const()[name = tensor("op_4973_validate_indices_0"), val = tensor(false)]; + tensor var_4973_cast_fp16 = scatter_along_axis(axis = var_4973_axis_0, data = var_4971_cast_fp16, indices = write_indices_23, mode = var_4973_mode_0, updates = v_23_cast_fp16, validate_indices = var_4973_validate_indices_0)[name = tensor("op_4973_cast_fp16")]; + tensor concat_92 = const()[name = tensor("concat_92"), val = tensor([1, 0, 0, 0, 0])]; + tensor concat_93 = const()[name = tensor("concat_93"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_23_internal_tensor_assign_2_stride_0 = const()[name = tensor("new_cache_23_internal_tensor_assign_2_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_23_internal_tensor_assign_2_begin_mask_0 = const()[name = tensor("new_cache_23_internal_tensor_assign_2_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_23_internal_tensor_assign_2_end_mask_0 = const()[name = tensor("new_cache_23_internal_tensor_assign_2_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_23_internal_tensor_assign_2_squeeze_mask_0 = const()[name = tensor("new_cache_23_internal_tensor_assign_2_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_69 = const()[name = tensor("shape_69"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_23 = const()[name = tensor("reduce_prod_23"), val = tensor(1048576)]; + tensor range_1d_23_start_0 = const()[name = tensor("range_1d_23_start_0"), val = tensor(0)]; + tensor range_1d_23_step_0 = const()[name = tensor("range_1d_23_step_0"), val = tensor(1)]; + tensor range_1d_23 = range_1d(end = reduce_prod_23, start = range_1d_23_start_0, step = range_1d_23_step_0)[name = tensor("range_1d_23")]; + tensor reshape_115 = reshape(shape = shape_69, x = range_1d_23)[name = tensor("reshape_115")]; + tensor slice_by_index_23 = slice_by_index(begin = concat_92, begin_mask = new_cache_23_internal_tensor_assign_2_begin_mask_0, end = concat_93, end_mask = new_cache_23_internal_tensor_assign_2_end_mask_0, squeeze_mask = new_cache_23_internal_tensor_assign_2_squeeze_mask_0, stride = new_cache_23_internal_tensor_assign_2_stride_0, x = reshape_115)[name = tensor("slice_by_index_23")]; + tensor reshape_116_shape_0 = const()[name = tensor("reshape_116_shape_0"), val = tensor([-1])]; + tensor reshape_116 = reshape(shape = reshape_116_shape_0, x = slice_by_index_23)[name = tensor("reshape_116")]; + tensor reshape_117_shape_0 = const()[name = tensor("reshape_117_shape_0"), val = tensor([-1])]; + tensor reshape_117_cast_fp16 = reshape(shape = reshape_117_shape_0, x = var_4973_cast_fp16)[name = tensor("reshape_117_cast_fp16")]; + tensor reshape_118_shape_0 = const()[name = tensor("reshape_118_shape_0"), val = tensor([-1])]; + tensor reshape_118_cast_fp16 = reshape(shape = reshape_118_shape_0, x = reshape_114_cast_fp16)[name = tensor("reshape_118_cast_fp16")]; + tensor scatter_23_mode_0 = const()[name = tensor("scatter_23_mode_0"), val = tensor("update")]; + tensor scatter_23_axis_0 = const()[name = tensor("scatter_23_axis_0"), val = tensor(0)]; + tensor scatter_23_validate_indices_0 = const()[name = tensor("scatter_23_validate_indices_0"), val = tensor(false)]; + tensor scatter_23_cast_fp16 = scatter(axis = scatter_23_axis_0, data = reshape_118_cast_fp16, indices = reshape_116, mode = scatter_23_mode_0, updates = reshape_117_cast_fp16, validate_indices = scatter_23_validate_indices_0)[name = tensor("scatter_23_cast_fp16")]; + tensor reshape_119_cast_fp16 = reshape(shape = shape_69, x = scatter_23_cast_fp16)[name = tensor("reshape_119_cast_fp16")]; + tensor reshape_119_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("reshape_119_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor keys_67_begin_0 = const()[name = tensor("keys_67_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor keys_67_end_0 = const()[name = tensor("keys_67_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor keys_67_end_mask_0 = const()[name = tensor("keys_67_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor keys_67_squeeze_mask_0 = const()[name = tensor("keys_67_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor keys_67_cast_fp16 = slice_by_index(begin = keys_67_begin_0, end = keys_67_end_0, end_mask = keys_67_end_mask_0, squeeze_mask = keys_67_squeeze_mask_0, x = reshape_119_cast_fp16)[name = tensor("keys_67_cast_fp16")]; + tensor values_67_begin_0 = const()[name = tensor("values_67_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor values_67_end_0 = const()[name = tensor("values_67_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor values_67_end_mask_0 = const()[name = tensor("values_67_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor values_67_squeeze_mask_0 = const()[name = tensor("values_67_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor values_67_cast_fp16 = slice_by_index(begin = values_67_begin_0, end = values_67_end_0, end_mask = values_67_end_mask_0, squeeze_mask = values_67_squeeze_mask_0, x = reshape_119_cast_fp16)[name = tensor("values_67_cast_fp16")]; + tensor var_4985_cast_fp16 = not_equal(x = keys_67_cast_fp16, y = keys_67_cast_fp16)[name = tensor("op_4985_cast_fp16")]; + tensor keys_69_cast_fp16 = select(a = var_470_to_fp16, b = keys_67_cast_fp16, cond = var_4985_cast_fp16)[name = tensor("keys_69_cast_fp16")]; + tensor var_4993_cast_fp16 = not_equal(x = values_67_cast_fp16, y = values_67_cast_fp16)[name = tensor("op_4993_cast_fp16")]; + tensor values_69_cast_fp16 = select(a = var_470_to_fp16, b = values_67_cast_fp16, cond = var_4993_cast_fp16)[name = tensor("values_69_cast_fp16")]; + tensor var_5017 = const()[name = tensor("op_5017"), val = tensor([0, 2, 1, 3])]; + tensor var_5033 = const()[name = tensor("op_5033"), val = tensor([1, 1, 1])]; + tensor var_5034_cast_fp16 = reshape(shape = var_5033, x = position11_to_fp16)[name = tensor("op_5034_cast_fp16")]; + tensor q_positions_23_cast_fp16 = add(x = var_5034_cast_fp16, y = q_offsets_1_promoted_to_fp16)[name = tensor("q_positions_23_cast_fp16")]; + tensor valid_end_23_cast_fp16 = add(x = var_5034_cast_fp16, y = var_536_cast_fp16)[name = tensor("valid_end_23_cast_fp16")]; + tensor valid_mask_23_cast_fp16 = less(x = k_positions_1_promoted_to_fp16, y = valid_end_23_cast_fp16)[name = tensor("valid_mask_23_cast_fp16")]; + tensor causal_mask_23_cast_fp16 = less_equal(x = k_positions_1_promoted_to_fp16, y = q_positions_23_cast_fp16)[name = tensor("causal_mask_23_cast_fp16")]; + tensor var_5062 = logical_and(x = valid_mask_23_cast_fp16, y = causal_mask_23_cast_fp16)[name = tensor("op_5062")]; + tensor attn_mask_23_axes_0 = const()[name = tensor("attn_mask_23_axes_0"), val = tensor([1])]; + tensor attn_mask_23 = expand_dims(axes = attn_mask_23_axes_0, x = var_5062)[name = tensor("attn_mask_23")]; + tensor var_5076_transpose_x_0 = const()[name = tensor("op_5076_transpose_x_0"), val = tensor(false)]; + tensor var_5076_transpose_y_0 = const()[name = tensor("op_5076_transpose_y_0"), val = tensor(false)]; + tensor transpose_91_perm_0 = const()[name = tensor("transpose_91_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_92_perm_0 = const()[name = tensor("transpose_92_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_92 = transpose(perm = transpose_92_perm_0, x = keys_69_cast_fp16)[name = tensor("transpose_160")]; + tensor transpose_91 = transpose(perm = transpose_91_perm_0, x = q_69_cast_fp16)[name = tensor("transpose_161")]; + tensor var_5076_cast_fp16 = matmul(transpose_x = var_5076_transpose_x_0, transpose_y = var_5076_transpose_y_0, x = transpose_91, y = transpose_92)[name = tensor("op_5076_cast_fp16")]; + tensor var_5070_to_fp16 = const()[name = tensor("op_5070_to_fp16"), val = tensor([0x1p-3])]; + tensor attn_weights_67_cast_fp16 = mul(x = var_5076_cast_fp16, y = var_5070_to_fp16)[name = tensor("attn_weights_67_cast_fp16")]; + tensor var_5078 = logical_not(x = attn_mask_23)[name = tensor("op_5078")]; + tensor var_5079_to_fp16 = const()[name = tensor("op_5079_to_fp16"), val = tensor(-inf)]; + tensor attn_weights_69_cast_fp16 = select(a = var_5079_to_fp16, b = attn_weights_67_cast_fp16, cond = var_5078)[name = tensor("attn_weights_69_cast_fp16")]; + tensor var_5081 = const()[name = tensor("op_5081"), val = tensor(-1)]; + tensor attn_weights_71_cast_fp16 = softmax(axis = var_5081, x = attn_weights_69_cast_fp16)[name = tensor("attn_weights_71_cast_fp16")]; + tensor attn_output_23_transpose_x_0 = const()[name = tensor("attn_output_23_transpose_x_0"), val = tensor(false)]; + tensor attn_output_23_transpose_y_0 = const()[name = tensor("attn_output_23_transpose_y_0"), val = tensor(false)]; + tensor values_71_cast_fp16 = transpose(perm = var_5017, x = values_69_cast_fp16)[name = tensor("transpose_162")]; + tensor attn_output_23_cast_fp16 = matmul(transpose_x = attn_output_23_transpose_x_0, transpose_y = attn_output_23_transpose_y_0, x = attn_weights_71_cast_fp16, y = values_71_cast_fp16)[name = tensor("attn_output_23_cast_fp16")]; + tensor var_5089 = const()[name = tensor("op_5089"), val = tensor([0, 2, 1, 3])]; + tensor var_5094 = const()[name = tensor("op_5094"), val = tensor([1, 256, 1024])]; + tensor var_5090_cast_fp16 = transpose(perm = var_5089, x = attn_output_23_cast_fp16)[name = tensor("transpose_159")]; + tensor input_137_cast_fp16 = reshape(shape = var_5094, x = var_5090_cast_fp16)[name = tensor("input_137_cast_fp16")]; + tensor attn11_out_proj_weight_to_fp16 = const()[name = tensor("attn11_out_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(284285184)))]; + tensor linear_45_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = attn11_out_proj_weight_to_fp16, x = input_137_cast_fp16)[name = tensor("linear_45_cast_fp16")]; + tensor var_5103_cast_fp16 = add(x = position11_to_fp16, y = valid_len_to_fp16)[name = tensor("op_5103_cast_fp16")]; + tensor var_5103_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_5103_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor input_139_cast_fp16 = add(x = input_133_cast_fp16, y = linear_45_cast_fp16)[name = tensor("input_139_cast_fp16")]; + tensor input_141_axes_0 = const()[name = tensor("input_141_axes_0"), val = tensor([-1])]; + tensor norm11_2_weight_to_fp16 = const()[name = tensor("norm11_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(286382400)))]; + tensor norm11_2_bias_to_fp16 = const()[name = tensor("norm11_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(286384512)))]; + tensor var_5107_to_fp16 = const()[name = tensor("op_5107_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_141_cast_fp16 = layer_norm(axes = input_141_axes_0, beta = norm11_2_bias_to_fp16, epsilon = var_5107_to_fp16, gamma = norm11_2_weight_to_fp16, x = input_139_cast_fp16)[name = tensor("input_141_cast_fp16")]; + tensor linear11_1_weight_to_fp16 = const()[name = tensor("linear11_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(286386624)))]; + tensor linear_46_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = linear11_1_weight_to_fp16, x = input_141_cast_fp16)[name = tensor("linear_46_cast_fp16")]; + tensor input_143_mode_0 = const()[name = tensor("input_143_mode_0"), val = tensor("EXACT")]; + tensor input_143_cast_fp16 = gelu(mode = input_143_mode_0, x = linear_46_cast_fp16)[name = tensor("input_143_cast_fp16")]; + tensor linear11_2_weight_to_fp16 = const()[name = tensor("linear11_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(294775296)))]; + tensor linear_47_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = linear11_2_weight_to_fp16, x = input_143_cast_fp16)[name = tensor("linear_47_cast_fp16")]; + tensor input_145_cast_fp16 = add(x = input_139_cast_fp16, y = linear_47_cast_fp16)[name = tensor("input_145_cast_fp16")]; + tensor input_147_axes_0 = const()[name = tensor("input_147_axes_0"), val = tensor([-1])]; + tensor norm12_1_weight_to_fp16 = const()[name = tensor("norm12_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(303163968)))]; + tensor norm12_1_bias_to_fp16 = const()[name = tensor("norm12_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(303166080)))]; + tensor var_5124_to_fp16 = const()[name = tensor("op_5124_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_147_cast_fp16 = layer_norm(axes = input_147_axes_0, beta = norm12_1_bias_to_fp16, epsilon = var_5124_to_fp16, gamma = norm12_1_weight_to_fp16, x = input_145_cast_fp16)[name = tensor("input_147_cast_fp16")]; + tensor attn12_in_proj_weight_to_fp16 = const()[name = tensor("attn12_in_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(303168192)))]; + tensor linear_48_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = attn12_in_proj_weight_to_fp16, x = input_147_cast_fp16)[name = tensor("linear_48_cast_fp16")]; + tensor var_5138 = const()[name = tensor("op_5138"), val = tensor([1, 256, 3, 16, 64])]; + tensor qkv_25_cast_fp16 = reshape(shape = var_5138, x = linear_48_cast_fp16)[name = tensor("qkv_25_cast_fp16")]; + tensor q_73_begin_0 = const()[name = tensor("q_73_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor q_73_end_0 = const()[name = tensor("q_73_end_0"), val = tensor([1, 256, 1, 16, 64])]; + tensor q_73_end_mask_0 = const()[name = tensor("q_73_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor q_73_squeeze_mask_0 = const()[name = tensor("q_73_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor q_73_cast_fp16 = slice_by_index(begin = q_73_begin_0, end = q_73_end_0, end_mask = q_73_end_mask_0, squeeze_mask = q_73_squeeze_mask_0, x = qkv_25_cast_fp16)[name = tensor("q_73_cast_fp16")]; + tensor k_49_begin_0 = const()[name = tensor("k_49_begin_0"), val = tensor([0, 0, 1, 0, 0])]; + tensor k_49_end_0 = const()[name = tensor("k_49_end_0"), val = tensor([1, 256, 2, 16, 64])]; + tensor k_49_end_mask_0 = const()[name = tensor("k_49_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor k_49_squeeze_mask_0 = const()[name = tensor("k_49_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor k_49_cast_fp16 = slice_by_index(begin = k_49_begin_0, end = k_49_end_0, end_mask = k_49_end_mask_0, squeeze_mask = k_49_squeeze_mask_0, x = qkv_25_cast_fp16)[name = tensor("k_49_cast_fp16")]; + tensor v_25_begin_0 = const()[name = tensor("v_25_begin_0"), val = tensor([0, 0, 2, 0, 0])]; + tensor v_25_end_0 = const()[name = tensor("v_25_end_0"), val = tensor([1, 256, 3, 16, 64])]; + tensor v_25_end_mask_0 = const()[name = tensor("v_25_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor v_25_squeeze_mask_0 = const()[name = tensor("v_25_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor v_25_cast_fp16 = slice_by_index(begin = v_25_begin_0, end = v_25_end_0, end_mask = v_25_end_mask_0, squeeze_mask = v_25_squeeze_mask_0, x = qkv_25_cast_fp16)[name = tensor("v_25_cast_fp16")]; + tensor var_5196 = const()[name = tensor("op_5196"), val = tensor([1, 1])]; + tensor position12_to_fp16_dtype_0 = const()[name = tensor("position12_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor position12_to_fp16 = cast(dtype = position12_to_fp16_dtype_0, x = position12)[name = tensor("cast_601")]; + tensor var_5197_cast_fp16 = reshape(shape = var_5196, x = position12_to_fp16)[name = tensor("op_5197_cast_fp16")]; + tensor ts_75_cast_fp16 = add(x = ts_1_promoted_to_fp16, y = var_5197_cast_fp16)[name = tensor("ts_75_cast_fp16")]; + tensor var_5204 = const()[name = tensor("op_5204"), val = tensor([1, 256, 1, 1])]; + tensor ts_77_cast_fp16 = reshape(shape = var_5204, x = ts_75_cast_fp16)[name = tensor("ts_77_cast_fp16")]; + tensor var_5211 = const()[name = tensor("op_5211"), val = tensor([1, 256, 16, 32, 2])]; + tensor q_complex_25_cast_fp16 = reshape(shape = var_5211, x = q_73_cast_fp16)[name = tensor("q_complex_25_cast_fp16")]; + tensor var_5218 = const()[name = tensor("op_5218"), val = tensor([1, 256, 16, 32, 2])]; + tensor k_complex_25_cast_fp16 = reshape(shape = var_5218, x = k_49_cast_fp16)[name = tensor("k_complex_25_cast_fp16")]; + tensor var_5222_begin_0 = const()[name = tensor("op_5222_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_5222_end_0 = const()[name = tensor("op_5222_end_0"), val = tensor([1, 256, 16, 32, 1])]; + tensor var_5222_end_mask_0 = const()[name = tensor("op_5222_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_5222_squeeze_mask_0 = const()[name = tensor("op_5222_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_5222_cast_fp16 = slice_by_index(begin = var_5222_begin_0, end = var_5222_end_0, end_mask = var_5222_end_mask_0, squeeze_mask = var_5222_squeeze_mask_0, x = q_complex_25_cast_fp16)[name = tensor("op_5222_cast_fp16")]; + tensor var_5230_begin_0 = const()[name = tensor("op_5230_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_5230_end_0 = const()[name = tensor("op_5230_end_0"), val = tensor([1, 256, 16, 32, 2])]; + tensor var_5230_end_mask_0 = const()[name = tensor("op_5230_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_5230_squeeze_mask_0 = const()[name = tensor("op_5230_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_5230_cast_fp16 = slice_by_index(begin = var_5230_begin_0, end = var_5230_end_0, end_mask = var_5230_end_mask_0, squeeze_mask = var_5230_squeeze_mask_0, x = q_complex_25_cast_fp16)[name = tensor("op_5230_cast_fp16")]; + tensor var_5238_begin_0 = const()[name = tensor("op_5238_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_5238_end_0 = const()[name = tensor("op_5238_end_0"), val = tensor([1, 256, 16, 32, 1])]; + tensor var_5238_end_mask_0 = const()[name = tensor("op_5238_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_5238_squeeze_mask_0 = const()[name = tensor("op_5238_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_5238_cast_fp16 = slice_by_index(begin = var_5238_begin_0, end = var_5238_end_0, end_mask = var_5238_end_mask_0, squeeze_mask = var_5238_squeeze_mask_0, x = k_complex_25_cast_fp16)[name = tensor("op_5238_cast_fp16")]; + tensor var_5246_begin_0 = const()[name = tensor("op_5246_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_5246_end_0 = const()[name = tensor("op_5246_end_0"), val = tensor([1, 256, 16, 32, 2])]; + tensor var_5246_end_mask_0 = const()[name = tensor("op_5246_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_5246_squeeze_mask_0 = const()[name = tensor("op_5246_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_5246_cast_fp16 = slice_by_index(begin = var_5246_begin_0, end = var_5246_end_0, end_mask = var_5246_end_mask_0, squeeze_mask = var_5246_squeeze_mask_0, x = k_complex_25_cast_fp16)[name = tensor("op_5246_cast_fp16")]; + tensor freqs_25_to_fp16 = const()[name = tensor("freqs_25_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(309459712)))]; + tensor var_5252_cast_fp16 = mul(x = freqs_25_to_fp16, y = ts_77_cast_fp16)[name = tensor("op_5252_cast_fp16")]; + tensor rotr_25_cast_fp16 = cos(x = var_5252_cast_fp16)[name = tensor("rotr_25_cast_fp16")]; + tensor roti_25_cast_fp16 = sin(x = var_5252_cast_fp16)[name = tensor("roti_25_cast_fp16")]; + tensor var_5256_cast_fp16 = mul(x = var_5222_cast_fp16, y = rotr_25_cast_fp16)[name = tensor("op_5256_cast_fp16")]; + tensor var_5257_cast_fp16 = mul(x = var_5230_cast_fp16, y = roti_25_cast_fp16)[name = tensor("op_5257_cast_fp16")]; + tensor qor_49_cast_fp16 = sub(x = var_5256_cast_fp16, y = var_5257_cast_fp16)[name = tensor("qor_49_cast_fp16")]; + tensor var_5260_cast_fp16 = mul(x = var_5222_cast_fp16, y = roti_25_cast_fp16)[name = tensor("op_5260_cast_fp16")]; + tensor var_5261_cast_fp16 = mul(x = var_5230_cast_fp16, y = rotr_25_cast_fp16)[name = tensor("op_5261_cast_fp16")]; + tensor qoi_49_cast_fp16 = add(x = var_5260_cast_fp16, y = var_5261_cast_fp16)[name = tensor("qoi_49_cast_fp16")]; + tensor var_5264_cast_fp16 = mul(x = var_5238_cast_fp16, y = rotr_25_cast_fp16)[name = tensor("op_5264_cast_fp16")]; + tensor var_5265_cast_fp16 = mul(x = var_5246_cast_fp16, y = roti_25_cast_fp16)[name = tensor("op_5265_cast_fp16")]; + tensor kor_49_cast_fp16 = sub(x = var_5264_cast_fp16, y = var_5265_cast_fp16)[name = tensor("kor_49_cast_fp16")]; + tensor var_5268_cast_fp16 = mul(x = var_5238_cast_fp16, y = roti_25_cast_fp16)[name = tensor("op_5268_cast_fp16")]; + tensor var_5269_cast_fp16 = mul(x = var_5246_cast_fp16, y = rotr_25_cast_fp16)[name = tensor("op_5269_cast_fp16")]; + tensor koi_49_cast_fp16 = add(x = var_5268_cast_fp16, y = var_5269_cast_fp16)[name = tensor("koi_49_cast_fp16")]; + tensor qo_25_axis_0 = const()[name = tensor("qo_25_axis_0"), val = tensor(-1)]; + tensor qo_25_cast_fp16 = stack(axis = qo_25_axis_0, values = (qor_49_cast_fp16, qoi_49_cast_fp16))[name = tensor("qo_25_cast_fp16")]; + tensor ko_25_axis_0 = const()[name = tensor("ko_25_axis_0"), val = tensor(-1)]; + tensor ko_25_cast_fp16 = stack(axis = ko_25_axis_0, values = (kor_49_cast_fp16, koi_49_cast_fp16))[name = tensor("ko_25_cast_fp16")]; + tensor var_5302 = const()[name = tensor("op_5302"), val = tensor([1, 256, 16, 64])]; + tensor q_75_cast_fp16 = reshape(shape = var_5302, x = qo_25_cast_fp16)[name = tensor("q_75_cast_fp16")]; + tensor var_5308 = const()[name = tensor("op_5308"), val = tensor([1, 256, 16, 64])]; + tensor k_51_cast_fp16 = reshape(shape = var_5308, x = ko_25_cast_fp16)[name = tensor("k_51_cast_fp16")]; + tensor write_indices_float_73_cast_fp16 = add(x = var_5197_cast_fp16, y = var_393_promoted_to_fp16)[name = tensor("write_indices_float_73_cast_fp16")]; + tensor _inversed_5329_y_0_to_fp16 = const()[name = tensor("_inversed_5329_y_0_to_fp16"), val = tensor(0x1p-9)]; + tensor _inversed_5329_cast_fp16 = mul(x = write_indices_float_73_cast_fp16, y = _inversed_5329_y_0_to_fp16)[name = tensor("_inversed_5329_cast_fp16")]; + tensor var_5330_cast_fp16 = floor(x = _inversed_5329_cast_fp16)[name = tensor("op_5330_cast_fp16")]; + tensor var_5331_to_fp16 = const()[name = tensor("op_5331_to_fp16"), val = tensor(0x1p+9)]; + tensor var_5332_cast_fp16 = mul(x = var_5330_cast_fp16, y = var_5331_to_fp16)[name = tensor("op_5332_cast_fp16")]; + tensor write_indices_float_75_cast_fp16 = sub(x = write_indices_float_73_cast_fp16, y = var_5332_cast_fp16)[name = tensor("write_indices_float_75_cast_fp16")]; + tensor var_5350_value_0_to_fp16 = const()[name = tensor("op_5350_value_0_to_fp16"), val = tensor(0x1.ffp+8)]; + tensor var_5350_cast_fp16 = fill_like(ref_tensor = write_indices_float_75_cast_fp16, value = var_5350_value_0_to_fp16)[name = tensor("op_5350_cast_fp16")]; + tensor write_indices_float_77_cast_fp16 = select(a = write_indices_float_75_cast_fp16, b = var_5350_cast_fp16, cond = token_valid_1_cast_fp16)[name = tensor("write_indices_float_77_cast_fp16")]; + tensor var_5356_dtype_0 = const()[name = tensor("op_5356_dtype_0"), val = tensor("int32")]; + tensor var_5361 = const()[name = tensor("op_5361"), val = tensor([1, 256, 1, 1])]; + tensor write_indices_float_77_cast_fp16_to_int32 = cast(dtype = var_5356_dtype_0, x = write_indices_float_77_cast_fp16)[name = tensor("cast_600")]; + tensor var_5362 = reshape(shape = var_5361, x = write_indices_float_77_cast_fp16_to_int32)[name = tensor("op_5362")]; + tensor write_indices_25_reps_0 = const()[name = tensor("write_indices_25_reps_0"), val = tensor([1, 1, 16, 64])]; + tensor write_indices_25 = tile(reps = write_indices_25_reps_0, x = var_5362)[name = tensor("write_indices_25")]; + tensor var_5372_begin_0 = const()[name = tensor("op_5372_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_5372_end_0 = const()[name = tensor("op_5372_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor var_5372_end_mask_0 = const()[name = tensor("op_5372_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_5372_squeeze_mask_0 = const()[name = tensor("op_5372_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor cache12_to_fp16_dtype_0 = const()[name = tensor("cache12_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor cache12_to_fp16 = cast(dtype = cache12_to_fp16_dtype_0, x = cache12)[name = tensor("cast_599")]; + tensor var_5372_cast_fp16 = slice_by_index(begin = var_5372_begin_0, end = var_5372_end_0, end_mask = var_5372_end_mask_0, squeeze_mask = var_5372_squeeze_mask_0, x = cache12_to_fp16)[name = tensor("op_5372_cast_fp16")]; + tensor var_5374_axis_0 = const()[name = tensor("op_5374_axis_0"), val = tensor(1)]; + tensor var_5374_mode_0 = const()[name = tensor("op_5374_mode_0"), val = tensor("update")]; + tensor var_5374_validate_indices_0 = const()[name = tensor("op_5374_validate_indices_0"), val = tensor(false)]; + tensor var_5374_cast_fp16 = scatter_along_axis(axis = var_5374_axis_0, data = var_5372_cast_fp16, indices = write_indices_25, mode = var_5374_mode_0, updates = k_51_cast_fp16, validate_indices = var_5374_validate_indices_0)[name = tensor("op_5374_cast_fp16")]; + tensor concat_98 = const()[name = tensor("concat_98"), val = tensor([0, 0, 0, 0, 0])]; + tensor concat_99 = const()[name = tensor("concat_99"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_25_internal_tensor_assign_1_stride_0 = const()[name = tensor("new_cache_25_internal_tensor_assign_1_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_25_internal_tensor_assign_1_begin_mask_0 = const()[name = tensor("new_cache_25_internal_tensor_assign_1_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_25_internal_tensor_assign_1_end_mask_0 = const()[name = tensor("new_cache_25_internal_tensor_assign_1_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_25_internal_tensor_assign_1_squeeze_mask_0 = const()[name = tensor("new_cache_25_internal_tensor_assign_1_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_70 = const()[name = tensor("shape_70"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_24 = const()[name = tensor("reduce_prod_24"), val = tensor(1048576)]; + tensor range_1d_24_start_0 = const()[name = tensor("range_1d_24_start_0"), val = tensor(0)]; + tensor range_1d_24_step_0 = const()[name = tensor("range_1d_24_step_0"), val = tensor(1)]; + tensor range_1d_24 = range_1d(end = reduce_prod_24, start = range_1d_24_start_0, step = range_1d_24_step_0)[name = tensor("range_1d_24")]; + tensor reshape_120 = reshape(shape = shape_70, x = range_1d_24)[name = tensor("reshape_120")]; + tensor slice_by_index_24 = slice_by_index(begin = concat_98, begin_mask = new_cache_25_internal_tensor_assign_1_begin_mask_0, end = concat_99, end_mask = new_cache_25_internal_tensor_assign_1_end_mask_0, squeeze_mask = new_cache_25_internal_tensor_assign_1_squeeze_mask_0, stride = new_cache_25_internal_tensor_assign_1_stride_0, x = reshape_120)[name = tensor("slice_by_index_24")]; + tensor reshape_121_shape_0 = const()[name = tensor("reshape_121_shape_0"), val = tensor([-1])]; + tensor reshape_121 = reshape(shape = reshape_121_shape_0, x = slice_by_index_24)[name = tensor("reshape_121")]; + tensor reshape_122_shape_0 = const()[name = tensor("reshape_122_shape_0"), val = tensor([-1])]; + tensor reshape_122_cast_fp16 = reshape(shape = reshape_122_shape_0, x = var_5374_cast_fp16)[name = tensor("reshape_122_cast_fp16")]; + tensor reshape_123_shape_0 = const()[name = tensor("reshape_123_shape_0"), val = tensor([-1])]; + tensor reshape_123_cast_fp16 = reshape(shape = reshape_123_shape_0, x = cache12_to_fp16)[name = tensor("reshape_123_cast_fp16")]; + tensor scatter_24_mode_0 = const()[name = tensor("scatter_24_mode_0"), val = tensor("update")]; + tensor scatter_24_axis_0 = const()[name = tensor("scatter_24_axis_0"), val = tensor(0)]; + tensor scatter_24_validate_indices_0 = const()[name = tensor("scatter_24_validate_indices_0"), val = tensor(false)]; + tensor scatter_24_cast_fp16 = scatter(axis = scatter_24_axis_0, data = reshape_123_cast_fp16, indices = reshape_121, mode = scatter_24_mode_0, updates = reshape_122_cast_fp16, validate_indices = scatter_24_validate_indices_0)[name = tensor("scatter_24_cast_fp16")]; + tensor reshape_124_cast_fp16 = reshape(shape = shape_70, x = scatter_24_cast_fp16)[name = tensor("reshape_124_cast_fp16")]; + tensor var_5382_begin_0 = const()[name = tensor("op_5382_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor var_5382_end_0 = const()[name = tensor("op_5382_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor var_5382_end_mask_0 = const()[name = tensor("op_5382_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_5382_squeeze_mask_0 = const()[name = tensor("op_5382_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor var_5382_cast_fp16 = slice_by_index(begin = var_5382_begin_0, end = var_5382_end_0, end_mask = var_5382_end_mask_0, squeeze_mask = var_5382_squeeze_mask_0, x = reshape_124_cast_fp16)[name = tensor("op_5382_cast_fp16")]; + tensor var_5384_axis_0 = const()[name = tensor("op_5384_axis_0"), val = tensor(1)]; + tensor var_5384_mode_0 = const()[name = tensor("op_5384_mode_0"), val = tensor("update")]; + tensor var_5384_validate_indices_0 = const()[name = tensor("op_5384_validate_indices_0"), val = tensor(false)]; + tensor var_5384_cast_fp16 = scatter_along_axis(axis = var_5384_axis_0, data = var_5382_cast_fp16, indices = write_indices_25, mode = var_5384_mode_0, updates = v_25_cast_fp16, validate_indices = var_5384_validate_indices_0)[name = tensor("op_5384_cast_fp16")]; + tensor concat_100 = const()[name = tensor("concat_100"), val = tensor([1, 0, 0, 0, 0])]; + tensor concat_101 = const()[name = tensor("concat_101"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_25_internal_tensor_assign_2_stride_0 = const()[name = tensor("new_cache_25_internal_tensor_assign_2_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_25_internal_tensor_assign_2_begin_mask_0 = const()[name = tensor("new_cache_25_internal_tensor_assign_2_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_25_internal_tensor_assign_2_end_mask_0 = const()[name = tensor("new_cache_25_internal_tensor_assign_2_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_25_internal_tensor_assign_2_squeeze_mask_0 = const()[name = tensor("new_cache_25_internal_tensor_assign_2_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_71 = const()[name = tensor("shape_71"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_25 = const()[name = tensor("reduce_prod_25"), val = tensor(1048576)]; + tensor range_1d_25_start_0 = const()[name = tensor("range_1d_25_start_0"), val = tensor(0)]; + tensor range_1d_25_step_0 = const()[name = tensor("range_1d_25_step_0"), val = tensor(1)]; + tensor range_1d_25 = range_1d(end = reduce_prod_25, start = range_1d_25_start_0, step = range_1d_25_step_0)[name = tensor("range_1d_25")]; + tensor reshape_125 = reshape(shape = shape_71, x = range_1d_25)[name = tensor("reshape_125")]; + tensor slice_by_index_25 = slice_by_index(begin = concat_100, begin_mask = new_cache_25_internal_tensor_assign_2_begin_mask_0, end = concat_101, end_mask = new_cache_25_internal_tensor_assign_2_end_mask_0, squeeze_mask = new_cache_25_internal_tensor_assign_2_squeeze_mask_0, stride = new_cache_25_internal_tensor_assign_2_stride_0, x = reshape_125)[name = tensor("slice_by_index_25")]; + tensor reshape_126_shape_0 = const()[name = tensor("reshape_126_shape_0"), val = tensor([-1])]; + tensor reshape_126 = reshape(shape = reshape_126_shape_0, x = slice_by_index_25)[name = tensor("reshape_126")]; + tensor reshape_127_shape_0 = const()[name = tensor("reshape_127_shape_0"), val = tensor([-1])]; + tensor reshape_127_cast_fp16 = reshape(shape = reshape_127_shape_0, x = var_5384_cast_fp16)[name = tensor("reshape_127_cast_fp16")]; + tensor reshape_128_shape_0 = const()[name = tensor("reshape_128_shape_0"), val = tensor([-1])]; + tensor reshape_128_cast_fp16 = reshape(shape = reshape_128_shape_0, x = reshape_124_cast_fp16)[name = tensor("reshape_128_cast_fp16")]; + tensor scatter_25_mode_0 = const()[name = tensor("scatter_25_mode_0"), val = tensor("update")]; + tensor scatter_25_axis_0 = const()[name = tensor("scatter_25_axis_0"), val = tensor(0)]; + tensor scatter_25_validate_indices_0 = const()[name = tensor("scatter_25_validate_indices_0"), val = tensor(false)]; + tensor scatter_25_cast_fp16 = scatter(axis = scatter_25_axis_0, data = reshape_128_cast_fp16, indices = reshape_126, mode = scatter_25_mode_0, updates = reshape_127_cast_fp16, validate_indices = scatter_25_validate_indices_0)[name = tensor("scatter_25_cast_fp16")]; + tensor reshape_129_cast_fp16 = reshape(shape = shape_71, x = scatter_25_cast_fp16)[name = tensor("reshape_129_cast_fp16")]; + tensor reshape_129_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("reshape_129_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor keys_73_begin_0 = const()[name = tensor("keys_73_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor keys_73_end_0 = const()[name = tensor("keys_73_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor keys_73_end_mask_0 = const()[name = tensor("keys_73_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor keys_73_squeeze_mask_0 = const()[name = tensor("keys_73_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor keys_73_cast_fp16 = slice_by_index(begin = keys_73_begin_0, end = keys_73_end_0, end_mask = keys_73_end_mask_0, squeeze_mask = keys_73_squeeze_mask_0, x = reshape_129_cast_fp16)[name = tensor("keys_73_cast_fp16")]; + tensor values_73_begin_0 = const()[name = tensor("values_73_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor values_73_end_0 = const()[name = tensor("values_73_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor values_73_end_mask_0 = const()[name = tensor("values_73_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor values_73_squeeze_mask_0 = const()[name = tensor("values_73_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor values_73_cast_fp16 = slice_by_index(begin = values_73_begin_0, end = values_73_end_0, end_mask = values_73_end_mask_0, squeeze_mask = values_73_squeeze_mask_0, x = reshape_129_cast_fp16)[name = tensor("values_73_cast_fp16")]; + tensor var_5396_cast_fp16 = not_equal(x = keys_73_cast_fp16, y = keys_73_cast_fp16)[name = tensor("op_5396_cast_fp16")]; + tensor keys_75_cast_fp16 = select(a = var_470_to_fp16, b = keys_73_cast_fp16, cond = var_5396_cast_fp16)[name = tensor("keys_75_cast_fp16")]; + tensor var_5404_cast_fp16 = not_equal(x = values_73_cast_fp16, y = values_73_cast_fp16)[name = tensor("op_5404_cast_fp16")]; + tensor values_75_cast_fp16 = select(a = var_470_to_fp16, b = values_73_cast_fp16, cond = var_5404_cast_fp16)[name = tensor("values_75_cast_fp16")]; + tensor var_5428 = const()[name = tensor("op_5428"), val = tensor([0, 2, 1, 3])]; + tensor var_5444 = const()[name = tensor("op_5444"), val = tensor([1, 1, 1])]; + tensor var_5445_cast_fp16 = reshape(shape = var_5444, x = position12_to_fp16)[name = tensor("op_5445_cast_fp16")]; + tensor q_positions_25_cast_fp16 = add(x = var_5445_cast_fp16, y = q_offsets_1_promoted_to_fp16)[name = tensor("q_positions_25_cast_fp16")]; + tensor valid_end_25_cast_fp16 = add(x = var_5445_cast_fp16, y = var_536_cast_fp16)[name = tensor("valid_end_25_cast_fp16")]; + tensor valid_mask_25_cast_fp16 = less(x = k_positions_1_promoted_to_fp16, y = valid_end_25_cast_fp16)[name = tensor("valid_mask_25_cast_fp16")]; + tensor causal_mask_25_cast_fp16 = less_equal(x = k_positions_1_promoted_to_fp16, y = q_positions_25_cast_fp16)[name = tensor("causal_mask_25_cast_fp16")]; + tensor var_5473 = logical_and(x = valid_mask_25_cast_fp16, y = causal_mask_25_cast_fp16)[name = tensor("op_5473")]; + tensor attn_mask_25_axes_0 = const()[name = tensor("attn_mask_25_axes_0"), val = tensor([1])]; + tensor attn_mask_25 = expand_dims(axes = attn_mask_25_axes_0, x = var_5473)[name = tensor("attn_mask_25")]; + tensor var_5487_transpose_x_0 = const()[name = tensor("op_5487_transpose_x_0"), val = tensor(false)]; + tensor var_5487_transpose_y_0 = const()[name = tensor("op_5487_transpose_y_0"), val = tensor(false)]; + tensor transpose_93_perm_0 = const()[name = tensor("transpose_93_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_94_perm_0 = const()[name = tensor("transpose_94_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_94 = transpose(perm = transpose_94_perm_0, x = keys_75_cast_fp16)[name = tensor("transpose_156")]; + tensor transpose_93 = transpose(perm = transpose_93_perm_0, x = q_75_cast_fp16)[name = tensor("transpose_157")]; + tensor var_5487_cast_fp16 = matmul(transpose_x = var_5487_transpose_x_0, transpose_y = var_5487_transpose_y_0, x = transpose_93, y = transpose_94)[name = tensor("op_5487_cast_fp16")]; + tensor var_5481_to_fp16 = const()[name = tensor("op_5481_to_fp16"), val = tensor([0x1p-3])]; + tensor attn_weights_73_cast_fp16 = mul(x = var_5487_cast_fp16, y = var_5481_to_fp16)[name = tensor("attn_weights_73_cast_fp16")]; + tensor var_5489 = logical_not(x = attn_mask_25)[name = tensor("op_5489")]; + tensor var_5490_to_fp16 = const()[name = tensor("op_5490_to_fp16"), val = tensor(-inf)]; + tensor attn_weights_75_cast_fp16 = select(a = var_5490_to_fp16, b = attn_weights_73_cast_fp16, cond = var_5489)[name = tensor("attn_weights_75_cast_fp16")]; + tensor var_5492 = const()[name = tensor("op_5492"), val = tensor(-1)]; + tensor attn_weights_77_cast_fp16 = softmax(axis = var_5492, x = attn_weights_75_cast_fp16)[name = tensor("attn_weights_77_cast_fp16")]; + tensor attn_output_25_transpose_x_0 = const()[name = tensor("attn_output_25_transpose_x_0"), val = tensor(false)]; + tensor attn_output_25_transpose_y_0 = const()[name = tensor("attn_output_25_transpose_y_0"), val = tensor(false)]; + tensor values_77_cast_fp16 = transpose(perm = var_5428, x = values_75_cast_fp16)[name = tensor("transpose_158")]; + tensor attn_output_25_cast_fp16 = matmul(transpose_x = attn_output_25_transpose_x_0, transpose_y = attn_output_25_transpose_y_0, x = attn_weights_77_cast_fp16, y = values_77_cast_fp16)[name = tensor("attn_output_25_cast_fp16")]; + tensor var_5500 = const()[name = tensor("op_5500"), val = tensor([0, 2, 1, 3])]; + tensor var_5505 = const()[name = tensor("op_5505"), val = tensor([1, 256, 1024])]; + tensor var_5501_cast_fp16 = transpose(perm = var_5500, x = attn_output_25_cast_fp16)[name = tensor("transpose_155")]; + tensor input_149_cast_fp16 = reshape(shape = var_5505, x = var_5501_cast_fp16)[name = tensor("input_149_cast_fp16")]; + tensor attn12_out_proj_weight_to_fp16 = const()[name = tensor("attn12_out_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(309459840)))]; + tensor linear_49_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = attn12_out_proj_weight_to_fp16, x = input_149_cast_fp16)[name = tensor("linear_49_cast_fp16")]; + tensor var_5514_cast_fp16 = add(x = position12_to_fp16, y = valid_len_to_fp16)[name = tensor("op_5514_cast_fp16")]; + tensor var_5514_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_5514_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor input_151_cast_fp16 = add(x = input_145_cast_fp16, y = linear_49_cast_fp16)[name = tensor("input_151_cast_fp16")]; + tensor input_153_axes_0 = const()[name = tensor("input_153_axes_0"), val = tensor([-1])]; + tensor norm12_2_weight_to_fp16 = const()[name = tensor("norm12_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(311557056)))]; + tensor norm12_2_bias_to_fp16 = const()[name = tensor("norm12_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(311559168)))]; + tensor var_5518_to_fp16 = const()[name = tensor("op_5518_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_153_cast_fp16 = layer_norm(axes = input_153_axes_0, beta = norm12_2_bias_to_fp16, epsilon = var_5518_to_fp16, gamma = norm12_2_weight_to_fp16, x = input_151_cast_fp16)[name = tensor("input_153_cast_fp16")]; + tensor linear12_1_weight_to_fp16 = const()[name = tensor("linear12_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(311561280)))]; + tensor linear_50_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = linear12_1_weight_to_fp16, x = input_153_cast_fp16)[name = tensor("linear_50_cast_fp16")]; + tensor input_155_mode_0 = const()[name = tensor("input_155_mode_0"), val = tensor("EXACT")]; + tensor input_155_cast_fp16 = gelu(mode = input_155_mode_0, x = linear_50_cast_fp16)[name = tensor("input_155_cast_fp16")]; + tensor linear12_2_weight_to_fp16 = const()[name = tensor("linear12_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(319949952)))]; + tensor linear_51_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = linear12_2_weight_to_fp16, x = input_155_cast_fp16)[name = tensor("linear_51_cast_fp16")]; + tensor input_157_cast_fp16 = add(x = input_151_cast_fp16, y = linear_51_cast_fp16)[name = tensor("input_157_cast_fp16")]; + tensor input_159_axes_0 = const()[name = tensor("input_159_axes_0"), val = tensor([-1])]; + tensor norm13_1_weight_to_fp16 = const()[name = tensor("norm13_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(328338624)))]; + tensor norm13_1_bias_to_fp16 = const()[name = tensor("norm13_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(328340736)))]; + tensor var_5535_to_fp16 = const()[name = tensor("op_5535_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_159_cast_fp16 = layer_norm(axes = input_159_axes_0, beta = norm13_1_bias_to_fp16, epsilon = var_5535_to_fp16, gamma = norm13_1_weight_to_fp16, x = input_157_cast_fp16)[name = tensor("input_159_cast_fp16")]; + tensor attn13_in_proj_weight_to_fp16 = const()[name = tensor("attn13_in_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(328342848)))]; + tensor linear_52_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = attn13_in_proj_weight_to_fp16, x = input_159_cast_fp16)[name = tensor("linear_52_cast_fp16")]; + tensor var_5549 = const()[name = tensor("op_5549"), val = tensor([1, 256, 3, 16, 64])]; + tensor qkv_27_cast_fp16 = reshape(shape = var_5549, x = linear_52_cast_fp16)[name = tensor("qkv_27_cast_fp16")]; + tensor q_79_begin_0 = const()[name = tensor("q_79_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor q_79_end_0 = const()[name = tensor("q_79_end_0"), val = tensor([1, 256, 1, 16, 64])]; + tensor q_79_end_mask_0 = const()[name = tensor("q_79_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor q_79_squeeze_mask_0 = const()[name = tensor("q_79_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor q_79_cast_fp16 = slice_by_index(begin = q_79_begin_0, end = q_79_end_0, end_mask = q_79_end_mask_0, squeeze_mask = q_79_squeeze_mask_0, x = qkv_27_cast_fp16)[name = tensor("q_79_cast_fp16")]; + tensor k_53_begin_0 = const()[name = tensor("k_53_begin_0"), val = tensor([0, 0, 1, 0, 0])]; + tensor k_53_end_0 = const()[name = tensor("k_53_end_0"), val = tensor([1, 256, 2, 16, 64])]; + tensor k_53_end_mask_0 = const()[name = tensor("k_53_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor k_53_squeeze_mask_0 = const()[name = tensor("k_53_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor k_53_cast_fp16 = slice_by_index(begin = k_53_begin_0, end = k_53_end_0, end_mask = k_53_end_mask_0, squeeze_mask = k_53_squeeze_mask_0, x = qkv_27_cast_fp16)[name = tensor("k_53_cast_fp16")]; + tensor v_27_begin_0 = const()[name = tensor("v_27_begin_0"), val = tensor([0, 0, 2, 0, 0])]; + tensor v_27_end_0 = const()[name = tensor("v_27_end_0"), val = tensor([1, 256, 3, 16, 64])]; + tensor v_27_end_mask_0 = const()[name = tensor("v_27_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor v_27_squeeze_mask_0 = const()[name = tensor("v_27_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor v_27_cast_fp16 = slice_by_index(begin = v_27_begin_0, end = v_27_end_0, end_mask = v_27_end_mask_0, squeeze_mask = v_27_squeeze_mask_0, x = qkv_27_cast_fp16)[name = tensor("v_27_cast_fp16")]; + tensor var_5607 = const()[name = tensor("op_5607"), val = tensor([1, 1])]; + tensor position13_to_fp16_dtype_0 = const()[name = tensor("position13_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor position13_to_fp16 = cast(dtype = position13_to_fp16_dtype_0, x = position13)[name = tensor("cast_596")]; + tensor var_5608_cast_fp16 = reshape(shape = var_5607, x = position13_to_fp16)[name = tensor("op_5608_cast_fp16")]; + tensor ts_81_cast_fp16 = add(x = ts_1_promoted_to_fp16, y = var_5608_cast_fp16)[name = tensor("ts_81_cast_fp16")]; + tensor var_5615 = const()[name = tensor("op_5615"), val = tensor([1, 256, 1, 1])]; + tensor ts_83_cast_fp16 = reshape(shape = var_5615, x = ts_81_cast_fp16)[name = tensor("ts_83_cast_fp16")]; + tensor var_5622 = const()[name = tensor("op_5622"), val = tensor([1, 256, 16, 32, 2])]; + tensor q_complex_27_cast_fp16 = reshape(shape = var_5622, x = q_79_cast_fp16)[name = tensor("q_complex_27_cast_fp16")]; + tensor var_5629 = const()[name = tensor("op_5629"), val = tensor([1, 256, 16, 32, 2])]; + tensor k_complex_27_cast_fp16 = reshape(shape = var_5629, x = k_53_cast_fp16)[name = tensor("k_complex_27_cast_fp16")]; + tensor var_5633_begin_0 = const()[name = tensor("op_5633_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_5633_end_0 = const()[name = tensor("op_5633_end_0"), val = tensor([1, 256, 16, 32, 1])]; + tensor var_5633_end_mask_0 = const()[name = tensor("op_5633_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_5633_squeeze_mask_0 = const()[name = tensor("op_5633_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_5633_cast_fp16 = slice_by_index(begin = var_5633_begin_0, end = var_5633_end_0, end_mask = var_5633_end_mask_0, squeeze_mask = var_5633_squeeze_mask_0, x = q_complex_27_cast_fp16)[name = tensor("op_5633_cast_fp16")]; + tensor var_5641_begin_0 = const()[name = tensor("op_5641_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_5641_end_0 = const()[name = tensor("op_5641_end_0"), val = tensor([1, 256, 16, 32, 2])]; + tensor var_5641_end_mask_0 = const()[name = tensor("op_5641_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_5641_squeeze_mask_0 = const()[name = tensor("op_5641_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_5641_cast_fp16 = slice_by_index(begin = var_5641_begin_0, end = var_5641_end_0, end_mask = var_5641_end_mask_0, squeeze_mask = var_5641_squeeze_mask_0, x = q_complex_27_cast_fp16)[name = tensor("op_5641_cast_fp16")]; + tensor var_5649_begin_0 = const()[name = tensor("op_5649_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_5649_end_0 = const()[name = tensor("op_5649_end_0"), val = tensor([1, 256, 16, 32, 1])]; + tensor var_5649_end_mask_0 = const()[name = tensor("op_5649_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_5649_squeeze_mask_0 = const()[name = tensor("op_5649_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_5649_cast_fp16 = slice_by_index(begin = var_5649_begin_0, end = var_5649_end_0, end_mask = var_5649_end_mask_0, squeeze_mask = var_5649_squeeze_mask_0, x = k_complex_27_cast_fp16)[name = tensor("op_5649_cast_fp16")]; + tensor var_5657_begin_0 = const()[name = tensor("op_5657_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_5657_end_0 = const()[name = tensor("op_5657_end_0"), val = tensor([1, 256, 16, 32, 2])]; + tensor var_5657_end_mask_0 = const()[name = tensor("op_5657_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_5657_squeeze_mask_0 = const()[name = tensor("op_5657_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_5657_cast_fp16 = slice_by_index(begin = var_5657_begin_0, end = var_5657_end_0, end_mask = var_5657_end_mask_0, squeeze_mask = var_5657_squeeze_mask_0, x = k_complex_27_cast_fp16)[name = tensor("op_5657_cast_fp16")]; + tensor freqs_27_to_fp16 = const()[name = tensor("freqs_27_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(334634368)))]; + tensor var_5663_cast_fp16 = mul(x = freqs_27_to_fp16, y = ts_83_cast_fp16)[name = tensor("op_5663_cast_fp16")]; + tensor rotr_27_cast_fp16 = cos(x = var_5663_cast_fp16)[name = tensor("rotr_27_cast_fp16")]; + tensor roti_27_cast_fp16 = sin(x = var_5663_cast_fp16)[name = tensor("roti_27_cast_fp16")]; + tensor var_5667_cast_fp16 = mul(x = var_5633_cast_fp16, y = rotr_27_cast_fp16)[name = tensor("op_5667_cast_fp16")]; + tensor var_5668_cast_fp16 = mul(x = var_5641_cast_fp16, y = roti_27_cast_fp16)[name = tensor("op_5668_cast_fp16")]; + tensor qor_53_cast_fp16 = sub(x = var_5667_cast_fp16, y = var_5668_cast_fp16)[name = tensor("qor_53_cast_fp16")]; + tensor var_5671_cast_fp16 = mul(x = var_5633_cast_fp16, y = roti_27_cast_fp16)[name = tensor("op_5671_cast_fp16")]; + tensor var_5672_cast_fp16 = mul(x = var_5641_cast_fp16, y = rotr_27_cast_fp16)[name = tensor("op_5672_cast_fp16")]; + tensor qoi_53_cast_fp16 = add(x = var_5671_cast_fp16, y = var_5672_cast_fp16)[name = tensor("qoi_53_cast_fp16")]; + tensor var_5675_cast_fp16 = mul(x = var_5649_cast_fp16, y = rotr_27_cast_fp16)[name = tensor("op_5675_cast_fp16")]; + tensor var_5676_cast_fp16 = mul(x = var_5657_cast_fp16, y = roti_27_cast_fp16)[name = tensor("op_5676_cast_fp16")]; + tensor kor_53_cast_fp16 = sub(x = var_5675_cast_fp16, y = var_5676_cast_fp16)[name = tensor("kor_53_cast_fp16")]; + tensor var_5679_cast_fp16 = mul(x = var_5649_cast_fp16, y = roti_27_cast_fp16)[name = tensor("op_5679_cast_fp16")]; + tensor var_5680_cast_fp16 = mul(x = var_5657_cast_fp16, y = rotr_27_cast_fp16)[name = tensor("op_5680_cast_fp16")]; + tensor koi_53_cast_fp16 = add(x = var_5679_cast_fp16, y = var_5680_cast_fp16)[name = tensor("koi_53_cast_fp16")]; + tensor qo_27_axis_0 = const()[name = tensor("qo_27_axis_0"), val = tensor(-1)]; + tensor qo_27_cast_fp16 = stack(axis = qo_27_axis_0, values = (qor_53_cast_fp16, qoi_53_cast_fp16))[name = tensor("qo_27_cast_fp16")]; + tensor ko_27_axis_0 = const()[name = tensor("ko_27_axis_0"), val = tensor(-1)]; + tensor ko_27_cast_fp16 = stack(axis = ko_27_axis_0, values = (kor_53_cast_fp16, koi_53_cast_fp16))[name = tensor("ko_27_cast_fp16")]; + tensor var_5713 = const()[name = tensor("op_5713"), val = tensor([1, 256, 16, 64])]; + tensor q_81_cast_fp16 = reshape(shape = var_5713, x = qo_27_cast_fp16)[name = tensor("q_81_cast_fp16")]; + tensor var_5719 = const()[name = tensor("op_5719"), val = tensor([1, 256, 16, 64])]; + tensor k_55_cast_fp16 = reshape(shape = var_5719, x = ko_27_cast_fp16)[name = tensor("k_55_cast_fp16")]; + tensor write_indices_float_79_cast_fp16 = add(x = var_5608_cast_fp16, y = var_393_promoted_to_fp16)[name = tensor("write_indices_float_79_cast_fp16")]; + tensor _inversed_5740_y_0_to_fp16 = const()[name = tensor("_inversed_5740_y_0_to_fp16"), val = tensor(0x1p-9)]; + tensor _inversed_5740_cast_fp16 = mul(x = write_indices_float_79_cast_fp16, y = _inversed_5740_y_0_to_fp16)[name = tensor("_inversed_5740_cast_fp16")]; + tensor var_5741_cast_fp16 = floor(x = _inversed_5740_cast_fp16)[name = tensor("op_5741_cast_fp16")]; + tensor var_5742_to_fp16 = const()[name = tensor("op_5742_to_fp16"), val = tensor(0x1p+9)]; + tensor var_5743_cast_fp16 = mul(x = var_5741_cast_fp16, y = var_5742_to_fp16)[name = tensor("op_5743_cast_fp16")]; + tensor write_indices_float_81_cast_fp16 = sub(x = write_indices_float_79_cast_fp16, y = var_5743_cast_fp16)[name = tensor("write_indices_float_81_cast_fp16")]; + tensor var_5761_value_0_to_fp16 = const()[name = tensor("op_5761_value_0_to_fp16"), val = tensor(0x1.ffp+8)]; + tensor var_5761_cast_fp16 = fill_like(ref_tensor = write_indices_float_81_cast_fp16, value = var_5761_value_0_to_fp16)[name = tensor("op_5761_cast_fp16")]; + tensor write_indices_float_83_cast_fp16 = select(a = write_indices_float_81_cast_fp16, b = var_5761_cast_fp16, cond = token_valid_1_cast_fp16)[name = tensor("write_indices_float_83_cast_fp16")]; + tensor var_5767_dtype_0 = const()[name = tensor("op_5767_dtype_0"), val = tensor("int32")]; + tensor var_5772 = const()[name = tensor("op_5772"), val = tensor([1, 256, 1, 1])]; + tensor write_indices_float_83_cast_fp16_to_int32 = cast(dtype = var_5767_dtype_0, x = write_indices_float_83_cast_fp16)[name = tensor("cast_595")]; + tensor var_5773 = reshape(shape = var_5772, x = write_indices_float_83_cast_fp16_to_int32)[name = tensor("op_5773")]; + tensor write_indices_27_reps_0 = const()[name = tensor("write_indices_27_reps_0"), val = tensor([1, 1, 16, 64])]; + tensor write_indices_27 = tile(reps = write_indices_27_reps_0, x = var_5773)[name = tensor("write_indices_27")]; + tensor var_5783_begin_0 = const()[name = tensor("op_5783_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_5783_end_0 = const()[name = tensor("op_5783_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor var_5783_end_mask_0 = const()[name = tensor("op_5783_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_5783_squeeze_mask_0 = const()[name = tensor("op_5783_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor cache13_to_fp16_dtype_0 = const()[name = tensor("cache13_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor cache13_to_fp16 = cast(dtype = cache13_to_fp16_dtype_0, x = cache13)[name = tensor("cast_594")]; + tensor var_5783_cast_fp16 = slice_by_index(begin = var_5783_begin_0, end = var_5783_end_0, end_mask = var_5783_end_mask_0, squeeze_mask = var_5783_squeeze_mask_0, x = cache13_to_fp16)[name = tensor("op_5783_cast_fp16")]; + tensor var_5785_axis_0 = const()[name = tensor("op_5785_axis_0"), val = tensor(1)]; + tensor var_5785_mode_0 = const()[name = tensor("op_5785_mode_0"), val = tensor("update")]; + tensor var_5785_validate_indices_0 = const()[name = tensor("op_5785_validate_indices_0"), val = tensor(false)]; + tensor var_5785_cast_fp16 = scatter_along_axis(axis = var_5785_axis_0, data = var_5783_cast_fp16, indices = write_indices_27, mode = var_5785_mode_0, updates = k_55_cast_fp16, validate_indices = var_5785_validate_indices_0)[name = tensor("op_5785_cast_fp16")]; + tensor concat_106 = const()[name = tensor("concat_106"), val = tensor([0, 0, 0, 0, 0])]; + tensor concat_107 = const()[name = tensor("concat_107"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_27_internal_tensor_assign_1_stride_0 = const()[name = tensor("new_cache_27_internal_tensor_assign_1_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_27_internal_tensor_assign_1_begin_mask_0 = const()[name = tensor("new_cache_27_internal_tensor_assign_1_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_27_internal_tensor_assign_1_end_mask_0 = const()[name = tensor("new_cache_27_internal_tensor_assign_1_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_27_internal_tensor_assign_1_squeeze_mask_0 = const()[name = tensor("new_cache_27_internal_tensor_assign_1_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_72 = const()[name = tensor("shape_72"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_26 = const()[name = tensor("reduce_prod_26"), val = tensor(1048576)]; + tensor range_1d_26_start_0 = const()[name = tensor("range_1d_26_start_0"), val = tensor(0)]; + tensor range_1d_26_step_0 = const()[name = tensor("range_1d_26_step_0"), val = tensor(1)]; + tensor range_1d_26 = range_1d(end = reduce_prod_26, start = range_1d_26_start_0, step = range_1d_26_step_0)[name = tensor("range_1d_26")]; + tensor reshape_130 = reshape(shape = shape_72, x = range_1d_26)[name = tensor("reshape_130")]; + tensor slice_by_index_26 = slice_by_index(begin = concat_106, begin_mask = new_cache_27_internal_tensor_assign_1_begin_mask_0, end = concat_107, end_mask = new_cache_27_internal_tensor_assign_1_end_mask_0, squeeze_mask = new_cache_27_internal_tensor_assign_1_squeeze_mask_0, stride = new_cache_27_internal_tensor_assign_1_stride_0, x = reshape_130)[name = tensor("slice_by_index_26")]; + tensor reshape_131_shape_0 = const()[name = tensor("reshape_131_shape_0"), val = tensor([-1])]; + tensor reshape_131 = reshape(shape = reshape_131_shape_0, x = slice_by_index_26)[name = tensor("reshape_131")]; + tensor reshape_132_shape_0 = const()[name = tensor("reshape_132_shape_0"), val = tensor([-1])]; + tensor reshape_132_cast_fp16 = reshape(shape = reshape_132_shape_0, x = var_5785_cast_fp16)[name = tensor("reshape_132_cast_fp16")]; + tensor reshape_133_shape_0 = const()[name = tensor("reshape_133_shape_0"), val = tensor([-1])]; + tensor reshape_133_cast_fp16 = reshape(shape = reshape_133_shape_0, x = cache13_to_fp16)[name = tensor("reshape_133_cast_fp16")]; + tensor scatter_26_mode_0 = const()[name = tensor("scatter_26_mode_0"), val = tensor("update")]; + tensor scatter_26_axis_0 = const()[name = tensor("scatter_26_axis_0"), val = tensor(0)]; + tensor scatter_26_validate_indices_0 = const()[name = tensor("scatter_26_validate_indices_0"), val = tensor(false)]; + tensor scatter_26_cast_fp16 = scatter(axis = scatter_26_axis_0, data = reshape_133_cast_fp16, indices = reshape_131, mode = scatter_26_mode_0, updates = reshape_132_cast_fp16, validate_indices = scatter_26_validate_indices_0)[name = tensor("scatter_26_cast_fp16")]; + tensor reshape_134_cast_fp16 = reshape(shape = shape_72, x = scatter_26_cast_fp16)[name = tensor("reshape_134_cast_fp16")]; + tensor var_5793_begin_0 = const()[name = tensor("op_5793_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor var_5793_end_0 = const()[name = tensor("op_5793_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor var_5793_end_mask_0 = const()[name = tensor("op_5793_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_5793_squeeze_mask_0 = const()[name = tensor("op_5793_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor var_5793_cast_fp16 = slice_by_index(begin = var_5793_begin_0, end = var_5793_end_0, end_mask = var_5793_end_mask_0, squeeze_mask = var_5793_squeeze_mask_0, x = reshape_134_cast_fp16)[name = tensor("op_5793_cast_fp16")]; + tensor var_5795_axis_0 = const()[name = tensor("op_5795_axis_0"), val = tensor(1)]; + tensor var_5795_mode_0 = const()[name = tensor("op_5795_mode_0"), val = tensor("update")]; + tensor var_5795_validate_indices_0 = const()[name = tensor("op_5795_validate_indices_0"), val = tensor(false)]; + tensor var_5795_cast_fp16 = scatter_along_axis(axis = var_5795_axis_0, data = var_5793_cast_fp16, indices = write_indices_27, mode = var_5795_mode_0, updates = v_27_cast_fp16, validate_indices = var_5795_validate_indices_0)[name = tensor("op_5795_cast_fp16")]; + tensor concat_108 = const()[name = tensor("concat_108"), val = tensor([1, 0, 0, 0, 0])]; + tensor concat_109 = const()[name = tensor("concat_109"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_27_internal_tensor_assign_2_stride_0 = const()[name = tensor("new_cache_27_internal_tensor_assign_2_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_27_internal_tensor_assign_2_begin_mask_0 = const()[name = tensor("new_cache_27_internal_tensor_assign_2_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_27_internal_tensor_assign_2_end_mask_0 = const()[name = tensor("new_cache_27_internal_tensor_assign_2_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_27_internal_tensor_assign_2_squeeze_mask_0 = const()[name = tensor("new_cache_27_internal_tensor_assign_2_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_73 = const()[name = tensor("shape_73"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_27 = const()[name = tensor("reduce_prod_27"), val = tensor(1048576)]; + tensor range_1d_27_start_0 = const()[name = tensor("range_1d_27_start_0"), val = tensor(0)]; + tensor range_1d_27_step_0 = const()[name = tensor("range_1d_27_step_0"), val = tensor(1)]; + tensor range_1d_27 = range_1d(end = reduce_prod_27, start = range_1d_27_start_0, step = range_1d_27_step_0)[name = tensor("range_1d_27")]; + tensor reshape_135 = reshape(shape = shape_73, x = range_1d_27)[name = tensor("reshape_135")]; + tensor slice_by_index_27 = slice_by_index(begin = concat_108, begin_mask = new_cache_27_internal_tensor_assign_2_begin_mask_0, end = concat_109, end_mask = new_cache_27_internal_tensor_assign_2_end_mask_0, squeeze_mask = new_cache_27_internal_tensor_assign_2_squeeze_mask_0, stride = new_cache_27_internal_tensor_assign_2_stride_0, x = reshape_135)[name = tensor("slice_by_index_27")]; + tensor reshape_136_shape_0 = const()[name = tensor("reshape_136_shape_0"), val = tensor([-1])]; + tensor reshape_136 = reshape(shape = reshape_136_shape_0, x = slice_by_index_27)[name = tensor("reshape_136")]; + tensor reshape_137_shape_0 = const()[name = tensor("reshape_137_shape_0"), val = tensor([-1])]; + tensor reshape_137_cast_fp16 = reshape(shape = reshape_137_shape_0, x = var_5795_cast_fp16)[name = tensor("reshape_137_cast_fp16")]; + tensor reshape_138_shape_0 = const()[name = tensor("reshape_138_shape_0"), val = tensor([-1])]; + tensor reshape_138_cast_fp16 = reshape(shape = reshape_138_shape_0, x = reshape_134_cast_fp16)[name = tensor("reshape_138_cast_fp16")]; + tensor scatter_27_mode_0 = const()[name = tensor("scatter_27_mode_0"), val = tensor("update")]; + tensor scatter_27_axis_0 = const()[name = tensor("scatter_27_axis_0"), val = tensor(0)]; + tensor scatter_27_validate_indices_0 = const()[name = tensor("scatter_27_validate_indices_0"), val = tensor(false)]; + tensor scatter_27_cast_fp16 = scatter(axis = scatter_27_axis_0, data = reshape_138_cast_fp16, indices = reshape_136, mode = scatter_27_mode_0, updates = reshape_137_cast_fp16, validate_indices = scatter_27_validate_indices_0)[name = tensor("scatter_27_cast_fp16")]; + tensor reshape_139_cast_fp16 = reshape(shape = shape_73, x = scatter_27_cast_fp16)[name = tensor("reshape_139_cast_fp16")]; + tensor reshape_139_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("reshape_139_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor keys_79_begin_0 = const()[name = tensor("keys_79_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor keys_79_end_0 = const()[name = tensor("keys_79_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor keys_79_end_mask_0 = const()[name = tensor("keys_79_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor keys_79_squeeze_mask_0 = const()[name = tensor("keys_79_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor keys_79_cast_fp16 = slice_by_index(begin = keys_79_begin_0, end = keys_79_end_0, end_mask = keys_79_end_mask_0, squeeze_mask = keys_79_squeeze_mask_0, x = reshape_139_cast_fp16)[name = tensor("keys_79_cast_fp16")]; + tensor values_79_begin_0 = const()[name = tensor("values_79_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor values_79_end_0 = const()[name = tensor("values_79_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor values_79_end_mask_0 = const()[name = tensor("values_79_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor values_79_squeeze_mask_0 = const()[name = tensor("values_79_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor values_79_cast_fp16 = slice_by_index(begin = values_79_begin_0, end = values_79_end_0, end_mask = values_79_end_mask_0, squeeze_mask = values_79_squeeze_mask_0, x = reshape_139_cast_fp16)[name = tensor("values_79_cast_fp16")]; + tensor var_5807_cast_fp16 = not_equal(x = keys_79_cast_fp16, y = keys_79_cast_fp16)[name = tensor("op_5807_cast_fp16")]; + tensor keys_81_cast_fp16 = select(a = var_470_to_fp16, b = keys_79_cast_fp16, cond = var_5807_cast_fp16)[name = tensor("keys_81_cast_fp16")]; + tensor var_5815_cast_fp16 = not_equal(x = values_79_cast_fp16, y = values_79_cast_fp16)[name = tensor("op_5815_cast_fp16")]; + tensor values_81_cast_fp16 = select(a = var_470_to_fp16, b = values_79_cast_fp16, cond = var_5815_cast_fp16)[name = tensor("values_81_cast_fp16")]; + tensor var_5839 = const()[name = tensor("op_5839"), val = tensor([0, 2, 1, 3])]; + tensor var_5855 = const()[name = tensor("op_5855"), val = tensor([1, 1, 1])]; + tensor var_5856_cast_fp16 = reshape(shape = var_5855, x = position13_to_fp16)[name = tensor("op_5856_cast_fp16")]; + tensor q_positions_27_cast_fp16 = add(x = var_5856_cast_fp16, y = q_offsets_1_promoted_to_fp16)[name = tensor("q_positions_27_cast_fp16")]; + tensor valid_end_27_cast_fp16 = add(x = var_5856_cast_fp16, y = var_536_cast_fp16)[name = tensor("valid_end_27_cast_fp16")]; + tensor valid_mask_27_cast_fp16 = less(x = k_positions_1_promoted_to_fp16, y = valid_end_27_cast_fp16)[name = tensor("valid_mask_27_cast_fp16")]; + tensor causal_mask_27_cast_fp16 = less_equal(x = k_positions_1_promoted_to_fp16, y = q_positions_27_cast_fp16)[name = tensor("causal_mask_27_cast_fp16")]; + tensor var_5884 = logical_and(x = valid_mask_27_cast_fp16, y = causal_mask_27_cast_fp16)[name = tensor("op_5884")]; + tensor attn_mask_27_axes_0 = const()[name = tensor("attn_mask_27_axes_0"), val = tensor([1])]; + tensor attn_mask_27 = expand_dims(axes = attn_mask_27_axes_0, x = var_5884)[name = tensor("attn_mask_27")]; + tensor var_5898_transpose_x_0 = const()[name = tensor("op_5898_transpose_x_0"), val = tensor(false)]; + tensor var_5898_transpose_y_0 = const()[name = tensor("op_5898_transpose_y_0"), val = tensor(false)]; + tensor transpose_95_perm_0 = const()[name = tensor("transpose_95_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_96_perm_0 = const()[name = tensor("transpose_96_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_96 = transpose(perm = transpose_96_perm_0, x = keys_81_cast_fp16)[name = tensor("transpose_152")]; + tensor transpose_95 = transpose(perm = transpose_95_perm_0, x = q_81_cast_fp16)[name = tensor("transpose_153")]; + tensor var_5898_cast_fp16 = matmul(transpose_x = var_5898_transpose_x_0, transpose_y = var_5898_transpose_y_0, x = transpose_95, y = transpose_96)[name = tensor("op_5898_cast_fp16")]; + tensor var_5892_to_fp16 = const()[name = tensor("op_5892_to_fp16"), val = tensor([0x1p-3])]; + tensor attn_weights_79_cast_fp16 = mul(x = var_5898_cast_fp16, y = var_5892_to_fp16)[name = tensor("attn_weights_79_cast_fp16")]; + tensor var_5900 = logical_not(x = attn_mask_27)[name = tensor("op_5900")]; + tensor var_5901_to_fp16 = const()[name = tensor("op_5901_to_fp16"), val = tensor(-inf)]; + tensor attn_weights_81_cast_fp16 = select(a = var_5901_to_fp16, b = attn_weights_79_cast_fp16, cond = var_5900)[name = tensor("attn_weights_81_cast_fp16")]; + tensor var_5903 = const()[name = tensor("op_5903"), val = tensor(-1)]; + tensor attn_weights_83_cast_fp16 = softmax(axis = var_5903, x = attn_weights_81_cast_fp16)[name = tensor("attn_weights_83_cast_fp16")]; + tensor attn_output_27_transpose_x_0 = const()[name = tensor("attn_output_27_transpose_x_0"), val = tensor(false)]; + tensor attn_output_27_transpose_y_0 = const()[name = tensor("attn_output_27_transpose_y_0"), val = tensor(false)]; + tensor values_83_cast_fp16 = transpose(perm = var_5839, x = values_81_cast_fp16)[name = tensor("transpose_154")]; + tensor attn_output_27_cast_fp16 = matmul(transpose_x = attn_output_27_transpose_x_0, transpose_y = attn_output_27_transpose_y_0, x = attn_weights_83_cast_fp16, y = values_83_cast_fp16)[name = tensor("attn_output_27_cast_fp16")]; + tensor var_5911 = const()[name = tensor("op_5911"), val = tensor([0, 2, 1, 3])]; + tensor var_5916 = const()[name = tensor("op_5916"), val = tensor([1, 256, 1024])]; + tensor var_5912_cast_fp16 = transpose(perm = var_5911, x = attn_output_27_cast_fp16)[name = tensor("transpose_151")]; + tensor input_161_cast_fp16 = reshape(shape = var_5916, x = var_5912_cast_fp16)[name = tensor("input_161_cast_fp16")]; + tensor attn13_out_proj_weight_to_fp16 = const()[name = tensor("attn13_out_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(334634496)))]; + tensor linear_53_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = attn13_out_proj_weight_to_fp16, x = input_161_cast_fp16)[name = tensor("linear_53_cast_fp16")]; + tensor var_5925_cast_fp16 = add(x = position13_to_fp16, y = valid_len_to_fp16)[name = tensor("op_5925_cast_fp16")]; + tensor var_5925_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_5925_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor input_163_cast_fp16 = add(x = input_157_cast_fp16, y = linear_53_cast_fp16)[name = tensor("input_163_cast_fp16")]; + tensor input_165_axes_0 = const()[name = tensor("input_165_axes_0"), val = tensor([-1])]; + tensor norm13_2_weight_to_fp16 = const()[name = tensor("norm13_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(336731712)))]; + tensor norm13_2_bias_to_fp16 = const()[name = tensor("norm13_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(336733824)))]; + tensor var_5929_to_fp16 = const()[name = tensor("op_5929_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_165_cast_fp16 = layer_norm(axes = input_165_axes_0, beta = norm13_2_bias_to_fp16, epsilon = var_5929_to_fp16, gamma = norm13_2_weight_to_fp16, x = input_163_cast_fp16)[name = tensor("input_165_cast_fp16")]; + tensor linear13_1_weight_to_fp16 = const()[name = tensor("linear13_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(336735936)))]; + tensor linear_54_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = linear13_1_weight_to_fp16, x = input_165_cast_fp16)[name = tensor("linear_54_cast_fp16")]; + tensor input_167_mode_0 = const()[name = tensor("input_167_mode_0"), val = tensor("EXACT")]; + tensor input_167_cast_fp16 = gelu(mode = input_167_mode_0, x = linear_54_cast_fp16)[name = tensor("input_167_cast_fp16")]; + tensor linear13_2_weight_to_fp16 = const()[name = tensor("linear13_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(345124608)))]; + tensor linear_55_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = linear13_2_weight_to_fp16, x = input_167_cast_fp16)[name = tensor("linear_55_cast_fp16")]; + tensor input_169_cast_fp16 = add(x = input_163_cast_fp16, y = linear_55_cast_fp16)[name = tensor("input_169_cast_fp16")]; + tensor input_171_axes_0 = const()[name = tensor("input_171_axes_0"), val = tensor([-1])]; + tensor norm14_1_weight_to_fp16 = const()[name = tensor("norm14_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(353513280)))]; + tensor norm14_1_bias_to_fp16 = const()[name = tensor("norm14_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(353515392)))]; + tensor var_5946_to_fp16 = const()[name = tensor("op_5946_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_171_cast_fp16 = layer_norm(axes = input_171_axes_0, beta = norm14_1_bias_to_fp16, epsilon = var_5946_to_fp16, gamma = norm14_1_weight_to_fp16, x = input_169_cast_fp16)[name = tensor("input_171_cast_fp16")]; + tensor attn14_in_proj_weight_to_fp16 = const()[name = tensor("attn14_in_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(353517504)))]; + tensor linear_56_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = attn14_in_proj_weight_to_fp16, x = input_171_cast_fp16)[name = tensor("linear_56_cast_fp16")]; + tensor var_5960 = const()[name = tensor("op_5960"), val = tensor([1, 256, 3, 16, 64])]; + tensor qkv_29_cast_fp16 = reshape(shape = var_5960, x = linear_56_cast_fp16)[name = tensor("qkv_29_cast_fp16")]; + tensor q_85_begin_0 = const()[name = tensor("q_85_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor q_85_end_0 = const()[name = tensor("q_85_end_0"), val = tensor([1, 256, 1, 16, 64])]; + tensor q_85_end_mask_0 = const()[name = tensor("q_85_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor q_85_squeeze_mask_0 = const()[name = tensor("q_85_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor q_85_cast_fp16 = slice_by_index(begin = q_85_begin_0, end = q_85_end_0, end_mask = q_85_end_mask_0, squeeze_mask = q_85_squeeze_mask_0, x = qkv_29_cast_fp16)[name = tensor("q_85_cast_fp16")]; + tensor k_57_begin_0 = const()[name = tensor("k_57_begin_0"), val = tensor([0, 0, 1, 0, 0])]; + tensor k_57_end_0 = const()[name = tensor("k_57_end_0"), val = tensor([1, 256, 2, 16, 64])]; + tensor k_57_end_mask_0 = const()[name = tensor("k_57_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor k_57_squeeze_mask_0 = const()[name = tensor("k_57_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor k_57_cast_fp16 = slice_by_index(begin = k_57_begin_0, end = k_57_end_0, end_mask = k_57_end_mask_0, squeeze_mask = k_57_squeeze_mask_0, x = qkv_29_cast_fp16)[name = tensor("k_57_cast_fp16")]; + tensor v_29_begin_0 = const()[name = tensor("v_29_begin_0"), val = tensor([0, 0, 2, 0, 0])]; + tensor v_29_end_0 = const()[name = tensor("v_29_end_0"), val = tensor([1, 256, 3, 16, 64])]; + tensor v_29_end_mask_0 = const()[name = tensor("v_29_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor v_29_squeeze_mask_0 = const()[name = tensor("v_29_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor v_29_cast_fp16 = slice_by_index(begin = v_29_begin_0, end = v_29_end_0, end_mask = v_29_end_mask_0, squeeze_mask = v_29_squeeze_mask_0, x = qkv_29_cast_fp16)[name = tensor("v_29_cast_fp16")]; + tensor var_6018 = const()[name = tensor("op_6018"), val = tensor([1, 1])]; + tensor position14_to_fp16_dtype_0 = const()[name = tensor("position14_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor position14_to_fp16 = cast(dtype = position14_to_fp16_dtype_0, x = position14)[name = tensor("cast_591")]; + tensor var_6019_cast_fp16 = reshape(shape = var_6018, x = position14_to_fp16)[name = tensor("op_6019_cast_fp16")]; + tensor ts_87_cast_fp16 = add(x = ts_1_promoted_to_fp16, y = var_6019_cast_fp16)[name = tensor("ts_87_cast_fp16")]; + tensor var_6026 = const()[name = tensor("op_6026"), val = tensor([1, 256, 1, 1])]; + tensor ts_89_cast_fp16 = reshape(shape = var_6026, x = ts_87_cast_fp16)[name = tensor("ts_89_cast_fp16")]; + tensor var_6033 = const()[name = tensor("op_6033"), val = tensor([1, 256, 16, 32, 2])]; + tensor q_complex_29_cast_fp16 = reshape(shape = var_6033, x = q_85_cast_fp16)[name = tensor("q_complex_29_cast_fp16")]; + tensor var_6040 = const()[name = tensor("op_6040"), val = tensor([1, 256, 16, 32, 2])]; + tensor k_complex_29_cast_fp16 = reshape(shape = var_6040, x = k_57_cast_fp16)[name = tensor("k_complex_29_cast_fp16")]; + tensor var_6044_begin_0 = const()[name = tensor("op_6044_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_6044_end_0 = const()[name = tensor("op_6044_end_0"), val = tensor([1, 256, 16, 32, 1])]; + tensor var_6044_end_mask_0 = const()[name = tensor("op_6044_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_6044_squeeze_mask_0 = const()[name = tensor("op_6044_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_6044_cast_fp16 = slice_by_index(begin = var_6044_begin_0, end = var_6044_end_0, end_mask = var_6044_end_mask_0, squeeze_mask = var_6044_squeeze_mask_0, x = q_complex_29_cast_fp16)[name = tensor("op_6044_cast_fp16")]; + tensor var_6052_begin_0 = const()[name = tensor("op_6052_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_6052_end_0 = const()[name = tensor("op_6052_end_0"), val = tensor([1, 256, 16, 32, 2])]; + tensor var_6052_end_mask_0 = const()[name = tensor("op_6052_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_6052_squeeze_mask_0 = const()[name = tensor("op_6052_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_6052_cast_fp16 = slice_by_index(begin = var_6052_begin_0, end = var_6052_end_0, end_mask = var_6052_end_mask_0, squeeze_mask = var_6052_squeeze_mask_0, x = q_complex_29_cast_fp16)[name = tensor("op_6052_cast_fp16")]; + tensor var_6060_begin_0 = const()[name = tensor("op_6060_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_6060_end_0 = const()[name = tensor("op_6060_end_0"), val = tensor([1, 256, 16, 32, 1])]; + tensor var_6060_end_mask_0 = const()[name = tensor("op_6060_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_6060_squeeze_mask_0 = const()[name = tensor("op_6060_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_6060_cast_fp16 = slice_by_index(begin = var_6060_begin_0, end = var_6060_end_0, end_mask = var_6060_end_mask_0, squeeze_mask = var_6060_squeeze_mask_0, x = k_complex_29_cast_fp16)[name = tensor("op_6060_cast_fp16")]; + tensor var_6068_begin_0 = const()[name = tensor("op_6068_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_6068_end_0 = const()[name = tensor("op_6068_end_0"), val = tensor([1, 256, 16, 32, 2])]; + tensor var_6068_end_mask_0 = const()[name = tensor("op_6068_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_6068_squeeze_mask_0 = const()[name = tensor("op_6068_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_6068_cast_fp16 = slice_by_index(begin = var_6068_begin_0, end = var_6068_end_0, end_mask = var_6068_end_mask_0, squeeze_mask = var_6068_squeeze_mask_0, x = k_complex_29_cast_fp16)[name = tensor("op_6068_cast_fp16")]; + tensor freqs_29_to_fp16 = const()[name = tensor("freqs_29_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(359809024)))]; + tensor var_6074_cast_fp16 = mul(x = freqs_29_to_fp16, y = ts_89_cast_fp16)[name = tensor("op_6074_cast_fp16")]; + tensor rotr_29_cast_fp16 = cos(x = var_6074_cast_fp16)[name = tensor("rotr_29_cast_fp16")]; + tensor roti_29_cast_fp16 = sin(x = var_6074_cast_fp16)[name = tensor("roti_29_cast_fp16")]; + tensor var_6078_cast_fp16 = mul(x = var_6044_cast_fp16, y = rotr_29_cast_fp16)[name = tensor("op_6078_cast_fp16")]; + tensor var_6079_cast_fp16 = mul(x = var_6052_cast_fp16, y = roti_29_cast_fp16)[name = tensor("op_6079_cast_fp16")]; + tensor qor_57_cast_fp16 = sub(x = var_6078_cast_fp16, y = var_6079_cast_fp16)[name = tensor("qor_57_cast_fp16")]; + tensor var_6082_cast_fp16 = mul(x = var_6044_cast_fp16, y = roti_29_cast_fp16)[name = tensor("op_6082_cast_fp16")]; + tensor var_6083_cast_fp16 = mul(x = var_6052_cast_fp16, y = rotr_29_cast_fp16)[name = tensor("op_6083_cast_fp16")]; + tensor qoi_57_cast_fp16 = add(x = var_6082_cast_fp16, y = var_6083_cast_fp16)[name = tensor("qoi_57_cast_fp16")]; + tensor var_6086_cast_fp16 = mul(x = var_6060_cast_fp16, y = rotr_29_cast_fp16)[name = tensor("op_6086_cast_fp16")]; + tensor var_6087_cast_fp16 = mul(x = var_6068_cast_fp16, y = roti_29_cast_fp16)[name = tensor("op_6087_cast_fp16")]; + tensor kor_57_cast_fp16 = sub(x = var_6086_cast_fp16, y = var_6087_cast_fp16)[name = tensor("kor_57_cast_fp16")]; + tensor var_6090_cast_fp16 = mul(x = var_6060_cast_fp16, y = roti_29_cast_fp16)[name = tensor("op_6090_cast_fp16")]; + tensor var_6091_cast_fp16 = mul(x = var_6068_cast_fp16, y = rotr_29_cast_fp16)[name = tensor("op_6091_cast_fp16")]; + tensor koi_57_cast_fp16 = add(x = var_6090_cast_fp16, y = var_6091_cast_fp16)[name = tensor("koi_57_cast_fp16")]; + tensor qo_29_axis_0 = const()[name = tensor("qo_29_axis_0"), val = tensor(-1)]; + tensor qo_29_cast_fp16 = stack(axis = qo_29_axis_0, values = (qor_57_cast_fp16, qoi_57_cast_fp16))[name = tensor("qo_29_cast_fp16")]; + tensor ko_29_axis_0 = const()[name = tensor("ko_29_axis_0"), val = tensor(-1)]; + tensor ko_29_cast_fp16 = stack(axis = ko_29_axis_0, values = (kor_57_cast_fp16, koi_57_cast_fp16))[name = tensor("ko_29_cast_fp16")]; + tensor var_6124 = const()[name = tensor("op_6124"), val = tensor([1, 256, 16, 64])]; + tensor q_87_cast_fp16 = reshape(shape = var_6124, x = qo_29_cast_fp16)[name = tensor("q_87_cast_fp16")]; + tensor var_6130 = const()[name = tensor("op_6130"), val = tensor([1, 256, 16, 64])]; + tensor k_59_cast_fp16 = reshape(shape = var_6130, x = ko_29_cast_fp16)[name = tensor("k_59_cast_fp16")]; + tensor write_indices_float_85_cast_fp16 = add(x = var_6019_cast_fp16, y = var_393_promoted_to_fp16)[name = tensor("write_indices_float_85_cast_fp16")]; + tensor _inversed_6151_y_0_to_fp16 = const()[name = tensor("_inversed_6151_y_0_to_fp16"), val = tensor(0x1p-9)]; + tensor _inversed_6151_cast_fp16 = mul(x = write_indices_float_85_cast_fp16, y = _inversed_6151_y_0_to_fp16)[name = tensor("_inversed_6151_cast_fp16")]; + tensor var_6152_cast_fp16 = floor(x = _inversed_6151_cast_fp16)[name = tensor("op_6152_cast_fp16")]; + tensor var_6153_to_fp16 = const()[name = tensor("op_6153_to_fp16"), val = tensor(0x1p+9)]; + tensor var_6154_cast_fp16 = mul(x = var_6152_cast_fp16, y = var_6153_to_fp16)[name = tensor("op_6154_cast_fp16")]; + tensor write_indices_float_87_cast_fp16 = sub(x = write_indices_float_85_cast_fp16, y = var_6154_cast_fp16)[name = tensor("write_indices_float_87_cast_fp16")]; + tensor var_6172_value_0_to_fp16 = const()[name = tensor("op_6172_value_0_to_fp16"), val = tensor(0x1.ffp+8)]; + tensor var_6172_cast_fp16 = fill_like(ref_tensor = write_indices_float_87_cast_fp16, value = var_6172_value_0_to_fp16)[name = tensor("op_6172_cast_fp16")]; + tensor write_indices_float_89_cast_fp16 = select(a = write_indices_float_87_cast_fp16, b = var_6172_cast_fp16, cond = token_valid_1_cast_fp16)[name = tensor("write_indices_float_89_cast_fp16")]; + tensor var_6178_dtype_0 = const()[name = tensor("op_6178_dtype_0"), val = tensor("int32")]; + tensor var_6183 = const()[name = tensor("op_6183"), val = tensor([1, 256, 1, 1])]; + tensor write_indices_float_89_cast_fp16_to_int32 = cast(dtype = var_6178_dtype_0, x = write_indices_float_89_cast_fp16)[name = tensor("cast_590")]; + tensor var_6184 = reshape(shape = var_6183, x = write_indices_float_89_cast_fp16_to_int32)[name = tensor("op_6184")]; + tensor write_indices_29_reps_0 = const()[name = tensor("write_indices_29_reps_0"), val = tensor([1, 1, 16, 64])]; + tensor write_indices_29 = tile(reps = write_indices_29_reps_0, x = var_6184)[name = tensor("write_indices_29")]; + tensor var_6194_begin_0 = const()[name = tensor("op_6194_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_6194_end_0 = const()[name = tensor("op_6194_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor var_6194_end_mask_0 = const()[name = tensor("op_6194_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_6194_squeeze_mask_0 = const()[name = tensor("op_6194_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor cache14_to_fp16_dtype_0 = const()[name = tensor("cache14_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor cache14_to_fp16 = cast(dtype = cache14_to_fp16_dtype_0, x = cache14)[name = tensor("cast_589")]; + tensor var_6194_cast_fp16 = slice_by_index(begin = var_6194_begin_0, end = var_6194_end_0, end_mask = var_6194_end_mask_0, squeeze_mask = var_6194_squeeze_mask_0, x = cache14_to_fp16)[name = tensor("op_6194_cast_fp16")]; + tensor var_6196_axis_0 = const()[name = tensor("op_6196_axis_0"), val = tensor(1)]; + tensor var_6196_mode_0 = const()[name = tensor("op_6196_mode_0"), val = tensor("update")]; + tensor var_6196_validate_indices_0 = const()[name = tensor("op_6196_validate_indices_0"), val = tensor(false)]; + tensor var_6196_cast_fp16 = scatter_along_axis(axis = var_6196_axis_0, data = var_6194_cast_fp16, indices = write_indices_29, mode = var_6196_mode_0, updates = k_59_cast_fp16, validate_indices = var_6196_validate_indices_0)[name = tensor("op_6196_cast_fp16")]; + tensor concat_114 = const()[name = tensor("concat_114"), val = tensor([0, 0, 0, 0, 0])]; + tensor concat_115 = const()[name = tensor("concat_115"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_29_internal_tensor_assign_1_stride_0 = const()[name = tensor("new_cache_29_internal_tensor_assign_1_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_29_internal_tensor_assign_1_begin_mask_0 = const()[name = tensor("new_cache_29_internal_tensor_assign_1_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_29_internal_tensor_assign_1_end_mask_0 = const()[name = tensor("new_cache_29_internal_tensor_assign_1_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_29_internal_tensor_assign_1_squeeze_mask_0 = const()[name = tensor("new_cache_29_internal_tensor_assign_1_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_74 = const()[name = tensor("shape_74"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_28 = const()[name = tensor("reduce_prod_28"), val = tensor(1048576)]; + tensor range_1d_28_start_0 = const()[name = tensor("range_1d_28_start_0"), val = tensor(0)]; + tensor range_1d_28_step_0 = const()[name = tensor("range_1d_28_step_0"), val = tensor(1)]; + tensor range_1d_28 = range_1d(end = reduce_prod_28, start = range_1d_28_start_0, step = range_1d_28_step_0)[name = tensor("range_1d_28")]; + tensor reshape_140 = reshape(shape = shape_74, x = range_1d_28)[name = tensor("reshape_140")]; + tensor slice_by_index_28 = slice_by_index(begin = concat_114, begin_mask = new_cache_29_internal_tensor_assign_1_begin_mask_0, end = concat_115, end_mask = new_cache_29_internal_tensor_assign_1_end_mask_0, squeeze_mask = new_cache_29_internal_tensor_assign_1_squeeze_mask_0, stride = new_cache_29_internal_tensor_assign_1_stride_0, x = reshape_140)[name = tensor("slice_by_index_28")]; + tensor reshape_141_shape_0 = const()[name = tensor("reshape_141_shape_0"), val = tensor([-1])]; + tensor reshape_141 = reshape(shape = reshape_141_shape_0, x = slice_by_index_28)[name = tensor("reshape_141")]; + tensor reshape_142_shape_0 = const()[name = tensor("reshape_142_shape_0"), val = tensor([-1])]; + tensor reshape_142_cast_fp16 = reshape(shape = reshape_142_shape_0, x = var_6196_cast_fp16)[name = tensor("reshape_142_cast_fp16")]; + tensor reshape_143_shape_0 = const()[name = tensor("reshape_143_shape_0"), val = tensor([-1])]; + tensor reshape_143_cast_fp16 = reshape(shape = reshape_143_shape_0, x = cache14_to_fp16)[name = tensor("reshape_143_cast_fp16")]; + tensor scatter_28_mode_0 = const()[name = tensor("scatter_28_mode_0"), val = tensor("update")]; + tensor scatter_28_axis_0 = const()[name = tensor("scatter_28_axis_0"), val = tensor(0)]; + tensor scatter_28_validate_indices_0 = const()[name = tensor("scatter_28_validate_indices_0"), val = tensor(false)]; + tensor scatter_28_cast_fp16 = scatter(axis = scatter_28_axis_0, data = reshape_143_cast_fp16, indices = reshape_141, mode = scatter_28_mode_0, updates = reshape_142_cast_fp16, validate_indices = scatter_28_validate_indices_0)[name = tensor("scatter_28_cast_fp16")]; + tensor reshape_144_cast_fp16 = reshape(shape = shape_74, x = scatter_28_cast_fp16)[name = tensor("reshape_144_cast_fp16")]; + tensor var_6204_begin_0 = const()[name = tensor("op_6204_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor var_6204_end_0 = const()[name = tensor("op_6204_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor var_6204_end_mask_0 = const()[name = tensor("op_6204_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_6204_squeeze_mask_0 = const()[name = tensor("op_6204_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor var_6204_cast_fp16 = slice_by_index(begin = var_6204_begin_0, end = var_6204_end_0, end_mask = var_6204_end_mask_0, squeeze_mask = var_6204_squeeze_mask_0, x = reshape_144_cast_fp16)[name = tensor("op_6204_cast_fp16")]; + tensor var_6206_axis_0 = const()[name = tensor("op_6206_axis_0"), val = tensor(1)]; + tensor var_6206_mode_0 = const()[name = tensor("op_6206_mode_0"), val = tensor("update")]; + tensor var_6206_validate_indices_0 = const()[name = tensor("op_6206_validate_indices_0"), val = tensor(false)]; + tensor var_6206_cast_fp16 = scatter_along_axis(axis = var_6206_axis_0, data = var_6204_cast_fp16, indices = write_indices_29, mode = var_6206_mode_0, updates = v_29_cast_fp16, validate_indices = var_6206_validate_indices_0)[name = tensor("op_6206_cast_fp16")]; + tensor concat_116 = const()[name = tensor("concat_116"), val = tensor([1, 0, 0, 0, 0])]; + tensor concat_117 = const()[name = tensor("concat_117"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_29_internal_tensor_assign_2_stride_0 = const()[name = tensor("new_cache_29_internal_tensor_assign_2_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_29_internal_tensor_assign_2_begin_mask_0 = const()[name = tensor("new_cache_29_internal_tensor_assign_2_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_29_internal_tensor_assign_2_end_mask_0 = const()[name = tensor("new_cache_29_internal_tensor_assign_2_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_29_internal_tensor_assign_2_squeeze_mask_0 = const()[name = tensor("new_cache_29_internal_tensor_assign_2_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_75 = const()[name = tensor("shape_75"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_29 = const()[name = tensor("reduce_prod_29"), val = tensor(1048576)]; + tensor range_1d_29_start_0 = const()[name = tensor("range_1d_29_start_0"), val = tensor(0)]; + tensor range_1d_29_step_0 = const()[name = tensor("range_1d_29_step_0"), val = tensor(1)]; + tensor range_1d_29 = range_1d(end = reduce_prod_29, start = range_1d_29_start_0, step = range_1d_29_step_0)[name = tensor("range_1d_29")]; + tensor reshape_145 = reshape(shape = shape_75, x = range_1d_29)[name = tensor("reshape_145")]; + tensor slice_by_index_29 = slice_by_index(begin = concat_116, begin_mask = new_cache_29_internal_tensor_assign_2_begin_mask_0, end = concat_117, end_mask = new_cache_29_internal_tensor_assign_2_end_mask_0, squeeze_mask = new_cache_29_internal_tensor_assign_2_squeeze_mask_0, stride = new_cache_29_internal_tensor_assign_2_stride_0, x = reshape_145)[name = tensor("slice_by_index_29")]; + tensor reshape_146_shape_0 = const()[name = tensor("reshape_146_shape_0"), val = tensor([-1])]; + tensor reshape_146 = reshape(shape = reshape_146_shape_0, x = slice_by_index_29)[name = tensor("reshape_146")]; + tensor reshape_147_shape_0 = const()[name = tensor("reshape_147_shape_0"), val = tensor([-1])]; + tensor reshape_147_cast_fp16 = reshape(shape = reshape_147_shape_0, x = var_6206_cast_fp16)[name = tensor("reshape_147_cast_fp16")]; + tensor reshape_148_shape_0 = const()[name = tensor("reshape_148_shape_0"), val = tensor([-1])]; + tensor reshape_148_cast_fp16 = reshape(shape = reshape_148_shape_0, x = reshape_144_cast_fp16)[name = tensor("reshape_148_cast_fp16")]; + tensor scatter_29_mode_0 = const()[name = tensor("scatter_29_mode_0"), val = tensor("update")]; + tensor scatter_29_axis_0 = const()[name = tensor("scatter_29_axis_0"), val = tensor(0)]; + tensor scatter_29_validate_indices_0 = const()[name = tensor("scatter_29_validate_indices_0"), val = tensor(false)]; + tensor scatter_29_cast_fp16 = scatter(axis = scatter_29_axis_0, data = reshape_148_cast_fp16, indices = reshape_146, mode = scatter_29_mode_0, updates = reshape_147_cast_fp16, validate_indices = scatter_29_validate_indices_0)[name = tensor("scatter_29_cast_fp16")]; + tensor reshape_149_cast_fp16 = reshape(shape = shape_75, x = scatter_29_cast_fp16)[name = tensor("reshape_149_cast_fp16")]; + tensor reshape_149_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("reshape_149_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor keys_85_begin_0 = const()[name = tensor("keys_85_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor keys_85_end_0 = const()[name = tensor("keys_85_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor keys_85_end_mask_0 = const()[name = tensor("keys_85_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor keys_85_squeeze_mask_0 = const()[name = tensor("keys_85_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor keys_85_cast_fp16 = slice_by_index(begin = keys_85_begin_0, end = keys_85_end_0, end_mask = keys_85_end_mask_0, squeeze_mask = keys_85_squeeze_mask_0, x = reshape_149_cast_fp16)[name = tensor("keys_85_cast_fp16")]; + tensor values_85_begin_0 = const()[name = tensor("values_85_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor values_85_end_0 = const()[name = tensor("values_85_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor values_85_end_mask_0 = const()[name = tensor("values_85_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor values_85_squeeze_mask_0 = const()[name = tensor("values_85_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor values_85_cast_fp16 = slice_by_index(begin = values_85_begin_0, end = values_85_end_0, end_mask = values_85_end_mask_0, squeeze_mask = values_85_squeeze_mask_0, x = reshape_149_cast_fp16)[name = tensor("values_85_cast_fp16")]; + tensor var_6218_cast_fp16 = not_equal(x = keys_85_cast_fp16, y = keys_85_cast_fp16)[name = tensor("op_6218_cast_fp16")]; + tensor keys_87_cast_fp16 = select(a = var_470_to_fp16, b = keys_85_cast_fp16, cond = var_6218_cast_fp16)[name = tensor("keys_87_cast_fp16")]; + tensor var_6226_cast_fp16 = not_equal(x = values_85_cast_fp16, y = values_85_cast_fp16)[name = tensor("op_6226_cast_fp16")]; + tensor values_87_cast_fp16 = select(a = var_470_to_fp16, b = values_85_cast_fp16, cond = var_6226_cast_fp16)[name = tensor("values_87_cast_fp16")]; + tensor var_6250 = const()[name = tensor("op_6250"), val = tensor([0, 2, 1, 3])]; + tensor var_6266 = const()[name = tensor("op_6266"), val = tensor([1, 1, 1])]; + tensor var_6267_cast_fp16 = reshape(shape = var_6266, x = position14_to_fp16)[name = tensor("op_6267_cast_fp16")]; + tensor q_positions_29_cast_fp16 = add(x = var_6267_cast_fp16, y = q_offsets_1_promoted_to_fp16)[name = tensor("q_positions_29_cast_fp16")]; + tensor valid_end_29_cast_fp16 = add(x = var_6267_cast_fp16, y = var_536_cast_fp16)[name = tensor("valid_end_29_cast_fp16")]; + tensor valid_mask_29_cast_fp16 = less(x = k_positions_1_promoted_to_fp16, y = valid_end_29_cast_fp16)[name = tensor("valid_mask_29_cast_fp16")]; + tensor causal_mask_29_cast_fp16 = less_equal(x = k_positions_1_promoted_to_fp16, y = q_positions_29_cast_fp16)[name = tensor("causal_mask_29_cast_fp16")]; + tensor var_6295 = logical_and(x = valid_mask_29_cast_fp16, y = causal_mask_29_cast_fp16)[name = tensor("op_6295")]; + tensor attn_mask_29_axes_0 = const()[name = tensor("attn_mask_29_axes_0"), val = tensor([1])]; + tensor attn_mask_29 = expand_dims(axes = attn_mask_29_axes_0, x = var_6295)[name = tensor("attn_mask_29")]; + tensor var_6309_transpose_x_0 = const()[name = tensor("op_6309_transpose_x_0"), val = tensor(false)]; + tensor var_6309_transpose_y_0 = const()[name = tensor("op_6309_transpose_y_0"), val = tensor(false)]; + tensor transpose_97_perm_0 = const()[name = tensor("transpose_97_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_98_perm_0 = const()[name = tensor("transpose_98_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_98 = transpose(perm = transpose_98_perm_0, x = keys_87_cast_fp16)[name = tensor("transpose_148")]; + tensor transpose_97 = transpose(perm = transpose_97_perm_0, x = q_87_cast_fp16)[name = tensor("transpose_149")]; + tensor var_6309_cast_fp16 = matmul(transpose_x = var_6309_transpose_x_0, transpose_y = var_6309_transpose_y_0, x = transpose_97, y = transpose_98)[name = tensor("op_6309_cast_fp16")]; + tensor var_6303_to_fp16 = const()[name = tensor("op_6303_to_fp16"), val = tensor([0x1p-3])]; + tensor attn_weights_85_cast_fp16 = mul(x = var_6309_cast_fp16, y = var_6303_to_fp16)[name = tensor("attn_weights_85_cast_fp16")]; + tensor var_6311 = logical_not(x = attn_mask_29)[name = tensor("op_6311")]; + tensor var_6312_to_fp16 = const()[name = tensor("op_6312_to_fp16"), val = tensor(-inf)]; + tensor attn_weights_87_cast_fp16 = select(a = var_6312_to_fp16, b = attn_weights_85_cast_fp16, cond = var_6311)[name = tensor("attn_weights_87_cast_fp16")]; + tensor var_6314 = const()[name = tensor("op_6314"), val = tensor(-1)]; + tensor attn_weights_89_cast_fp16 = softmax(axis = var_6314, x = attn_weights_87_cast_fp16)[name = tensor("attn_weights_89_cast_fp16")]; + tensor attn_output_29_transpose_x_0 = const()[name = tensor("attn_output_29_transpose_x_0"), val = tensor(false)]; + tensor attn_output_29_transpose_y_0 = const()[name = tensor("attn_output_29_transpose_y_0"), val = tensor(false)]; + tensor values_89_cast_fp16 = transpose(perm = var_6250, x = values_87_cast_fp16)[name = tensor("transpose_150")]; + tensor attn_output_29_cast_fp16 = matmul(transpose_x = attn_output_29_transpose_x_0, transpose_y = attn_output_29_transpose_y_0, x = attn_weights_89_cast_fp16, y = values_89_cast_fp16)[name = tensor("attn_output_29_cast_fp16")]; + tensor var_6322 = const()[name = tensor("op_6322"), val = tensor([0, 2, 1, 3])]; + tensor var_6327 = const()[name = tensor("op_6327"), val = tensor([1, 256, 1024])]; + tensor var_6323_cast_fp16 = transpose(perm = var_6322, x = attn_output_29_cast_fp16)[name = tensor("transpose_147")]; + tensor input_173_cast_fp16 = reshape(shape = var_6327, x = var_6323_cast_fp16)[name = tensor("input_173_cast_fp16")]; + tensor attn14_out_proj_weight_to_fp16 = const()[name = tensor("attn14_out_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(359809152)))]; + tensor linear_57_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = attn14_out_proj_weight_to_fp16, x = input_173_cast_fp16)[name = tensor("linear_57_cast_fp16")]; + tensor var_6336_cast_fp16 = add(x = position14_to_fp16, y = valid_len_to_fp16)[name = tensor("op_6336_cast_fp16")]; + tensor var_6336_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_6336_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor input_175_cast_fp16 = add(x = input_169_cast_fp16, y = linear_57_cast_fp16)[name = tensor("input_175_cast_fp16")]; + tensor input_177_axes_0 = const()[name = tensor("input_177_axes_0"), val = tensor([-1])]; + tensor norm14_2_weight_to_fp16 = const()[name = tensor("norm14_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(361906368)))]; + tensor norm14_2_bias_to_fp16 = const()[name = tensor("norm14_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(361908480)))]; + tensor var_6340_to_fp16 = const()[name = tensor("op_6340_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_177_cast_fp16 = layer_norm(axes = input_177_axes_0, beta = norm14_2_bias_to_fp16, epsilon = var_6340_to_fp16, gamma = norm14_2_weight_to_fp16, x = input_175_cast_fp16)[name = tensor("input_177_cast_fp16")]; + tensor linear14_1_weight_to_fp16 = const()[name = tensor("linear14_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(361910592)))]; + tensor linear_58_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = linear14_1_weight_to_fp16, x = input_177_cast_fp16)[name = tensor("linear_58_cast_fp16")]; + tensor input_179_mode_0 = const()[name = tensor("input_179_mode_0"), val = tensor("EXACT")]; + tensor input_179_cast_fp16 = gelu(mode = input_179_mode_0, x = linear_58_cast_fp16)[name = tensor("input_179_cast_fp16")]; + tensor linear14_2_weight_to_fp16 = const()[name = tensor("linear14_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(370299264)))]; + tensor linear_59_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = linear14_2_weight_to_fp16, x = input_179_cast_fp16)[name = tensor("linear_59_cast_fp16")]; + tensor input_181_cast_fp16 = add(x = input_175_cast_fp16, y = linear_59_cast_fp16)[name = tensor("input_181_cast_fp16")]; + tensor input_183_axes_0 = const()[name = tensor("input_183_axes_0"), val = tensor([-1])]; + tensor norm15_1_weight_to_fp16 = const()[name = tensor("norm15_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(378687936)))]; + tensor norm15_1_bias_to_fp16 = const()[name = tensor("norm15_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(378690048)))]; + tensor var_6357_to_fp16 = const()[name = tensor("op_6357_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_183_cast_fp16 = layer_norm(axes = input_183_axes_0, beta = norm15_1_bias_to_fp16, epsilon = var_6357_to_fp16, gamma = norm15_1_weight_to_fp16, x = input_181_cast_fp16)[name = tensor("input_183_cast_fp16")]; + tensor attn15_in_proj_weight_to_fp16 = const()[name = tensor("attn15_in_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(378692160)))]; + tensor linear_60_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = attn15_in_proj_weight_to_fp16, x = input_183_cast_fp16)[name = tensor("linear_60_cast_fp16")]; + tensor var_6371 = const()[name = tensor("op_6371"), val = tensor([1, 256, 3, 16, 64])]; + tensor qkv_31_cast_fp16 = reshape(shape = var_6371, x = linear_60_cast_fp16)[name = tensor("qkv_31_cast_fp16")]; + tensor q_91_begin_0 = const()[name = tensor("q_91_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor q_91_end_0 = const()[name = tensor("q_91_end_0"), val = tensor([1, 256, 1, 16, 64])]; + tensor q_91_end_mask_0 = const()[name = tensor("q_91_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor q_91_squeeze_mask_0 = const()[name = tensor("q_91_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor q_91_cast_fp16 = slice_by_index(begin = q_91_begin_0, end = q_91_end_0, end_mask = q_91_end_mask_0, squeeze_mask = q_91_squeeze_mask_0, x = qkv_31_cast_fp16)[name = tensor("q_91_cast_fp16")]; + tensor k_61_begin_0 = const()[name = tensor("k_61_begin_0"), val = tensor([0, 0, 1, 0, 0])]; + tensor k_61_end_0 = const()[name = tensor("k_61_end_0"), val = tensor([1, 256, 2, 16, 64])]; + tensor k_61_end_mask_0 = const()[name = tensor("k_61_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor k_61_squeeze_mask_0 = const()[name = tensor("k_61_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor k_61_cast_fp16 = slice_by_index(begin = k_61_begin_0, end = k_61_end_0, end_mask = k_61_end_mask_0, squeeze_mask = k_61_squeeze_mask_0, x = qkv_31_cast_fp16)[name = tensor("k_61_cast_fp16")]; + tensor v_31_begin_0 = const()[name = tensor("v_31_begin_0"), val = tensor([0, 0, 2, 0, 0])]; + tensor v_31_end_0 = const()[name = tensor("v_31_end_0"), val = tensor([1, 256, 3, 16, 64])]; + tensor v_31_end_mask_0 = const()[name = tensor("v_31_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor v_31_squeeze_mask_0 = const()[name = tensor("v_31_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor v_31_cast_fp16 = slice_by_index(begin = v_31_begin_0, end = v_31_end_0, end_mask = v_31_end_mask_0, squeeze_mask = v_31_squeeze_mask_0, x = qkv_31_cast_fp16)[name = tensor("v_31_cast_fp16")]; + tensor var_6429 = const()[name = tensor("op_6429"), val = tensor([1, 1])]; + tensor position15_to_fp16_dtype_0 = const()[name = tensor("position15_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor position15_to_fp16 = cast(dtype = position15_to_fp16_dtype_0, x = position15)[name = tensor("cast_586")]; + tensor var_6430_cast_fp16 = reshape(shape = var_6429, x = position15_to_fp16)[name = tensor("op_6430_cast_fp16")]; + tensor ts_93_cast_fp16 = add(x = ts_1_promoted_to_fp16, y = var_6430_cast_fp16)[name = tensor("ts_93_cast_fp16")]; + tensor var_6437 = const()[name = tensor("op_6437"), val = tensor([1, 256, 1, 1])]; + tensor ts_95_cast_fp16 = reshape(shape = var_6437, x = ts_93_cast_fp16)[name = tensor("ts_95_cast_fp16")]; + tensor var_6444 = const()[name = tensor("op_6444"), val = tensor([1, 256, 16, 32, 2])]; + tensor q_complex_31_cast_fp16 = reshape(shape = var_6444, x = q_91_cast_fp16)[name = tensor("q_complex_31_cast_fp16")]; + tensor var_6451 = const()[name = tensor("op_6451"), val = tensor([1, 256, 16, 32, 2])]; + tensor k_complex_31_cast_fp16 = reshape(shape = var_6451, x = k_61_cast_fp16)[name = tensor("k_complex_31_cast_fp16")]; + tensor var_6455_begin_0 = const()[name = tensor("op_6455_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_6455_end_0 = const()[name = tensor("op_6455_end_0"), val = tensor([1, 256, 16, 32, 1])]; + tensor var_6455_end_mask_0 = const()[name = tensor("op_6455_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_6455_squeeze_mask_0 = const()[name = tensor("op_6455_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_6455_cast_fp16 = slice_by_index(begin = var_6455_begin_0, end = var_6455_end_0, end_mask = var_6455_end_mask_0, squeeze_mask = var_6455_squeeze_mask_0, x = q_complex_31_cast_fp16)[name = tensor("op_6455_cast_fp16")]; + tensor var_6463_begin_0 = const()[name = tensor("op_6463_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_6463_end_0 = const()[name = tensor("op_6463_end_0"), val = tensor([1, 256, 16, 32, 2])]; + tensor var_6463_end_mask_0 = const()[name = tensor("op_6463_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_6463_squeeze_mask_0 = const()[name = tensor("op_6463_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_6463_cast_fp16 = slice_by_index(begin = var_6463_begin_0, end = var_6463_end_0, end_mask = var_6463_end_mask_0, squeeze_mask = var_6463_squeeze_mask_0, x = q_complex_31_cast_fp16)[name = tensor("op_6463_cast_fp16")]; + tensor var_6471_begin_0 = const()[name = tensor("op_6471_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_6471_end_0 = const()[name = tensor("op_6471_end_0"), val = tensor([1, 256, 16, 32, 1])]; + tensor var_6471_end_mask_0 = const()[name = tensor("op_6471_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_6471_squeeze_mask_0 = const()[name = tensor("op_6471_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_6471_cast_fp16 = slice_by_index(begin = var_6471_begin_0, end = var_6471_end_0, end_mask = var_6471_end_mask_0, squeeze_mask = var_6471_squeeze_mask_0, x = k_complex_31_cast_fp16)[name = tensor("op_6471_cast_fp16")]; + tensor var_6479_begin_0 = const()[name = tensor("op_6479_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_6479_end_0 = const()[name = tensor("op_6479_end_0"), val = tensor([1, 256, 16, 32, 2])]; + tensor var_6479_end_mask_0 = const()[name = tensor("op_6479_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_6479_squeeze_mask_0 = const()[name = tensor("op_6479_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_6479_cast_fp16 = slice_by_index(begin = var_6479_begin_0, end = var_6479_end_0, end_mask = var_6479_end_mask_0, squeeze_mask = var_6479_squeeze_mask_0, x = k_complex_31_cast_fp16)[name = tensor("op_6479_cast_fp16")]; + tensor freqs_31_to_fp16 = const()[name = tensor("freqs_31_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(384983680)))]; + tensor var_6485_cast_fp16 = mul(x = freqs_31_to_fp16, y = ts_95_cast_fp16)[name = tensor("op_6485_cast_fp16")]; + tensor rotr_31_cast_fp16 = cos(x = var_6485_cast_fp16)[name = tensor("rotr_31_cast_fp16")]; + tensor roti_31_cast_fp16 = sin(x = var_6485_cast_fp16)[name = tensor("roti_31_cast_fp16")]; + tensor var_6489_cast_fp16 = mul(x = var_6455_cast_fp16, y = rotr_31_cast_fp16)[name = tensor("op_6489_cast_fp16")]; + tensor var_6490_cast_fp16 = mul(x = var_6463_cast_fp16, y = roti_31_cast_fp16)[name = tensor("op_6490_cast_fp16")]; + tensor qor_61_cast_fp16 = sub(x = var_6489_cast_fp16, y = var_6490_cast_fp16)[name = tensor("qor_61_cast_fp16")]; + tensor var_6493_cast_fp16 = mul(x = var_6455_cast_fp16, y = roti_31_cast_fp16)[name = tensor("op_6493_cast_fp16")]; + tensor var_6494_cast_fp16 = mul(x = var_6463_cast_fp16, y = rotr_31_cast_fp16)[name = tensor("op_6494_cast_fp16")]; + tensor qoi_61_cast_fp16 = add(x = var_6493_cast_fp16, y = var_6494_cast_fp16)[name = tensor("qoi_61_cast_fp16")]; + tensor var_6497_cast_fp16 = mul(x = var_6471_cast_fp16, y = rotr_31_cast_fp16)[name = tensor("op_6497_cast_fp16")]; + tensor var_6498_cast_fp16 = mul(x = var_6479_cast_fp16, y = roti_31_cast_fp16)[name = tensor("op_6498_cast_fp16")]; + tensor kor_61_cast_fp16 = sub(x = var_6497_cast_fp16, y = var_6498_cast_fp16)[name = tensor("kor_61_cast_fp16")]; + tensor var_6501_cast_fp16 = mul(x = var_6471_cast_fp16, y = roti_31_cast_fp16)[name = tensor("op_6501_cast_fp16")]; + tensor var_6502_cast_fp16 = mul(x = var_6479_cast_fp16, y = rotr_31_cast_fp16)[name = tensor("op_6502_cast_fp16")]; + tensor koi_61_cast_fp16 = add(x = var_6501_cast_fp16, y = var_6502_cast_fp16)[name = tensor("koi_61_cast_fp16")]; + tensor qo_31_axis_0 = const()[name = tensor("qo_31_axis_0"), val = tensor(-1)]; + tensor qo_31_cast_fp16 = stack(axis = qo_31_axis_0, values = (qor_61_cast_fp16, qoi_61_cast_fp16))[name = tensor("qo_31_cast_fp16")]; + tensor ko_31_axis_0 = const()[name = tensor("ko_31_axis_0"), val = tensor(-1)]; + tensor ko_31_cast_fp16 = stack(axis = ko_31_axis_0, values = (kor_61_cast_fp16, koi_61_cast_fp16))[name = tensor("ko_31_cast_fp16")]; + tensor var_6535 = const()[name = tensor("op_6535"), val = tensor([1, 256, 16, 64])]; + tensor q_93_cast_fp16 = reshape(shape = var_6535, x = qo_31_cast_fp16)[name = tensor("q_93_cast_fp16")]; + tensor var_6541 = const()[name = tensor("op_6541"), val = tensor([1, 256, 16, 64])]; + tensor k_63_cast_fp16 = reshape(shape = var_6541, x = ko_31_cast_fp16)[name = tensor("k_63_cast_fp16")]; + tensor write_indices_float_91_cast_fp16 = add(x = var_6430_cast_fp16, y = var_393_promoted_to_fp16)[name = tensor("write_indices_float_91_cast_fp16")]; + tensor _inversed_6562_y_0_to_fp16 = const()[name = tensor("_inversed_6562_y_0_to_fp16"), val = tensor(0x1p-9)]; + tensor _inversed_6562_cast_fp16 = mul(x = write_indices_float_91_cast_fp16, y = _inversed_6562_y_0_to_fp16)[name = tensor("_inversed_6562_cast_fp16")]; + tensor var_6563_cast_fp16 = floor(x = _inversed_6562_cast_fp16)[name = tensor("op_6563_cast_fp16")]; + tensor var_6564_to_fp16 = const()[name = tensor("op_6564_to_fp16"), val = tensor(0x1p+9)]; + tensor var_6565_cast_fp16 = mul(x = var_6563_cast_fp16, y = var_6564_to_fp16)[name = tensor("op_6565_cast_fp16")]; + tensor write_indices_float_93_cast_fp16 = sub(x = write_indices_float_91_cast_fp16, y = var_6565_cast_fp16)[name = tensor("write_indices_float_93_cast_fp16")]; + tensor var_6583_value_0_to_fp16 = const()[name = tensor("op_6583_value_0_to_fp16"), val = tensor(0x1.ffp+8)]; + tensor var_6583_cast_fp16 = fill_like(ref_tensor = write_indices_float_93_cast_fp16, value = var_6583_value_0_to_fp16)[name = tensor("op_6583_cast_fp16")]; + tensor write_indices_float_95_cast_fp16 = select(a = write_indices_float_93_cast_fp16, b = var_6583_cast_fp16, cond = token_valid_1_cast_fp16)[name = tensor("write_indices_float_95_cast_fp16")]; + tensor var_6589_dtype_0 = const()[name = tensor("op_6589_dtype_0"), val = tensor("int32")]; + tensor var_6594 = const()[name = tensor("op_6594"), val = tensor([1, 256, 1, 1])]; + tensor write_indices_float_95_cast_fp16_to_int32 = cast(dtype = var_6589_dtype_0, x = write_indices_float_95_cast_fp16)[name = tensor("cast_585")]; + tensor var_6595 = reshape(shape = var_6594, x = write_indices_float_95_cast_fp16_to_int32)[name = tensor("op_6595")]; + tensor write_indices_31_reps_0 = const()[name = tensor("write_indices_31_reps_0"), val = tensor([1, 1, 16, 64])]; + tensor write_indices_31 = tile(reps = write_indices_31_reps_0, x = var_6595)[name = tensor("write_indices_31")]; + tensor var_6605_begin_0 = const()[name = tensor("op_6605_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_6605_end_0 = const()[name = tensor("op_6605_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor var_6605_end_mask_0 = const()[name = tensor("op_6605_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_6605_squeeze_mask_0 = const()[name = tensor("op_6605_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor cache15_to_fp16_dtype_0 = const()[name = tensor("cache15_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor cache15_to_fp16 = cast(dtype = cache15_to_fp16_dtype_0, x = cache15)[name = tensor("cast_584")]; + tensor var_6605_cast_fp16 = slice_by_index(begin = var_6605_begin_0, end = var_6605_end_0, end_mask = var_6605_end_mask_0, squeeze_mask = var_6605_squeeze_mask_0, x = cache15_to_fp16)[name = tensor("op_6605_cast_fp16")]; + tensor var_6607_axis_0 = const()[name = tensor("op_6607_axis_0"), val = tensor(1)]; + tensor var_6607_mode_0 = const()[name = tensor("op_6607_mode_0"), val = tensor("update")]; + tensor var_6607_validate_indices_0 = const()[name = tensor("op_6607_validate_indices_0"), val = tensor(false)]; + tensor var_6607_cast_fp16 = scatter_along_axis(axis = var_6607_axis_0, data = var_6605_cast_fp16, indices = write_indices_31, mode = var_6607_mode_0, updates = k_63_cast_fp16, validate_indices = var_6607_validate_indices_0)[name = tensor("op_6607_cast_fp16")]; + tensor concat_122 = const()[name = tensor("concat_122"), val = tensor([0, 0, 0, 0, 0])]; + tensor concat_123 = const()[name = tensor("concat_123"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_31_internal_tensor_assign_1_stride_0 = const()[name = tensor("new_cache_31_internal_tensor_assign_1_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_31_internal_tensor_assign_1_begin_mask_0 = const()[name = tensor("new_cache_31_internal_tensor_assign_1_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_31_internal_tensor_assign_1_end_mask_0 = const()[name = tensor("new_cache_31_internal_tensor_assign_1_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_31_internal_tensor_assign_1_squeeze_mask_0 = const()[name = tensor("new_cache_31_internal_tensor_assign_1_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_76 = const()[name = tensor("shape_76"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_30 = const()[name = tensor("reduce_prod_30"), val = tensor(1048576)]; + tensor range_1d_30_start_0 = const()[name = tensor("range_1d_30_start_0"), val = tensor(0)]; + tensor range_1d_30_step_0 = const()[name = tensor("range_1d_30_step_0"), val = tensor(1)]; + tensor range_1d_30 = range_1d(end = reduce_prod_30, start = range_1d_30_start_0, step = range_1d_30_step_0)[name = tensor("range_1d_30")]; + tensor reshape_150 = reshape(shape = shape_76, x = range_1d_30)[name = tensor("reshape_150")]; + tensor slice_by_index_30 = slice_by_index(begin = concat_122, begin_mask = new_cache_31_internal_tensor_assign_1_begin_mask_0, end = concat_123, end_mask = new_cache_31_internal_tensor_assign_1_end_mask_0, squeeze_mask = new_cache_31_internal_tensor_assign_1_squeeze_mask_0, stride = new_cache_31_internal_tensor_assign_1_stride_0, x = reshape_150)[name = tensor("slice_by_index_30")]; + tensor reshape_151_shape_0 = const()[name = tensor("reshape_151_shape_0"), val = tensor([-1])]; + tensor reshape_151 = reshape(shape = reshape_151_shape_0, x = slice_by_index_30)[name = tensor("reshape_151")]; + tensor reshape_152_shape_0 = const()[name = tensor("reshape_152_shape_0"), val = tensor([-1])]; + tensor reshape_152_cast_fp16 = reshape(shape = reshape_152_shape_0, x = var_6607_cast_fp16)[name = tensor("reshape_152_cast_fp16")]; + tensor reshape_153_shape_0 = const()[name = tensor("reshape_153_shape_0"), val = tensor([-1])]; + tensor reshape_153_cast_fp16 = reshape(shape = reshape_153_shape_0, x = cache15_to_fp16)[name = tensor("reshape_153_cast_fp16")]; + tensor scatter_30_mode_0 = const()[name = tensor("scatter_30_mode_0"), val = tensor("update")]; + tensor scatter_30_axis_0 = const()[name = tensor("scatter_30_axis_0"), val = tensor(0)]; + tensor scatter_30_validate_indices_0 = const()[name = tensor("scatter_30_validate_indices_0"), val = tensor(false)]; + tensor scatter_30_cast_fp16 = scatter(axis = scatter_30_axis_0, data = reshape_153_cast_fp16, indices = reshape_151, mode = scatter_30_mode_0, updates = reshape_152_cast_fp16, validate_indices = scatter_30_validate_indices_0)[name = tensor("scatter_30_cast_fp16")]; + tensor reshape_154_cast_fp16 = reshape(shape = shape_76, x = scatter_30_cast_fp16)[name = tensor("reshape_154_cast_fp16")]; + tensor var_6615_begin_0 = const()[name = tensor("op_6615_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor var_6615_end_0 = const()[name = tensor("op_6615_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor var_6615_end_mask_0 = const()[name = tensor("op_6615_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_6615_squeeze_mask_0 = const()[name = tensor("op_6615_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor var_6615_cast_fp16 = slice_by_index(begin = var_6615_begin_0, end = var_6615_end_0, end_mask = var_6615_end_mask_0, squeeze_mask = var_6615_squeeze_mask_0, x = reshape_154_cast_fp16)[name = tensor("op_6615_cast_fp16")]; + tensor var_6617_axis_0 = const()[name = tensor("op_6617_axis_0"), val = tensor(1)]; + tensor var_6617_mode_0 = const()[name = tensor("op_6617_mode_0"), val = tensor("update")]; + tensor var_6617_validate_indices_0 = const()[name = tensor("op_6617_validate_indices_0"), val = tensor(false)]; + tensor var_6617_cast_fp16 = scatter_along_axis(axis = var_6617_axis_0, data = var_6615_cast_fp16, indices = write_indices_31, mode = var_6617_mode_0, updates = v_31_cast_fp16, validate_indices = var_6617_validate_indices_0)[name = tensor("op_6617_cast_fp16")]; + tensor concat_124 = const()[name = tensor("concat_124"), val = tensor([1, 0, 0, 0, 0])]; + tensor concat_125 = const()[name = tensor("concat_125"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_31_internal_tensor_assign_2_stride_0 = const()[name = tensor("new_cache_31_internal_tensor_assign_2_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_31_internal_tensor_assign_2_begin_mask_0 = const()[name = tensor("new_cache_31_internal_tensor_assign_2_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_31_internal_tensor_assign_2_end_mask_0 = const()[name = tensor("new_cache_31_internal_tensor_assign_2_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_31_internal_tensor_assign_2_squeeze_mask_0 = const()[name = tensor("new_cache_31_internal_tensor_assign_2_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_77 = const()[name = tensor("shape_77"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_31 = const()[name = tensor("reduce_prod_31"), val = tensor(1048576)]; + tensor range_1d_31_start_0 = const()[name = tensor("range_1d_31_start_0"), val = tensor(0)]; + tensor range_1d_31_step_0 = const()[name = tensor("range_1d_31_step_0"), val = tensor(1)]; + tensor range_1d_31 = range_1d(end = reduce_prod_31, start = range_1d_31_start_0, step = range_1d_31_step_0)[name = tensor("range_1d_31")]; + tensor reshape_155 = reshape(shape = shape_77, x = range_1d_31)[name = tensor("reshape_155")]; + tensor slice_by_index_31 = slice_by_index(begin = concat_124, begin_mask = new_cache_31_internal_tensor_assign_2_begin_mask_0, end = concat_125, end_mask = new_cache_31_internal_tensor_assign_2_end_mask_0, squeeze_mask = new_cache_31_internal_tensor_assign_2_squeeze_mask_0, stride = new_cache_31_internal_tensor_assign_2_stride_0, x = reshape_155)[name = tensor("slice_by_index_31")]; + tensor reshape_156_shape_0 = const()[name = tensor("reshape_156_shape_0"), val = tensor([-1])]; + tensor reshape_156 = reshape(shape = reshape_156_shape_0, x = slice_by_index_31)[name = tensor("reshape_156")]; + tensor reshape_157_shape_0 = const()[name = tensor("reshape_157_shape_0"), val = tensor([-1])]; + tensor reshape_157_cast_fp16 = reshape(shape = reshape_157_shape_0, x = var_6617_cast_fp16)[name = tensor("reshape_157_cast_fp16")]; + tensor reshape_158_shape_0 = const()[name = tensor("reshape_158_shape_0"), val = tensor([-1])]; + tensor reshape_158_cast_fp16 = reshape(shape = reshape_158_shape_0, x = reshape_154_cast_fp16)[name = tensor("reshape_158_cast_fp16")]; + tensor scatter_31_mode_0 = const()[name = tensor("scatter_31_mode_0"), val = tensor("update")]; + tensor scatter_31_axis_0 = const()[name = tensor("scatter_31_axis_0"), val = tensor(0)]; + tensor scatter_31_validate_indices_0 = const()[name = tensor("scatter_31_validate_indices_0"), val = tensor(false)]; + tensor scatter_31_cast_fp16 = scatter(axis = scatter_31_axis_0, data = reshape_158_cast_fp16, indices = reshape_156, mode = scatter_31_mode_0, updates = reshape_157_cast_fp16, validate_indices = scatter_31_validate_indices_0)[name = tensor("scatter_31_cast_fp16")]; + tensor reshape_159_cast_fp16 = reshape(shape = shape_77, x = scatter_31_cast_fp16)[name = tensor("reshape_159_cast_fp16")]; + tensor reshape_159_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("reshape_159_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor keys_91_begin_0 = const()[name = tensor("keys_91_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor keys_91_end_0 = const()[name = tensor("keys_91_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor keys_91_end_mask_0 = const()[name = tensor("keys_91_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor keys_91_squeeze_mask_0 = const()[name = tensor("keys_91_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor keys_91_cast_fp16 = slice_by_index(begin = keys_91_begin_0, end = keys_91_end_0, end_mask = keys_91_end_mask_0, squeeze_mask = keys_91_squeeze_mask_0, x = reshape_159_cast_fp16)[name = tensor("keys_91_cast_fp16")]; + tensor values_91_begin_0 = const()[name = tensor("values_91_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor values_91_end_0 = const()[name = tensor("values_91_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor values_91_end_mask_0 = const()[name = tensor("values_91_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor values_91_squeeze_mask_0 = const()[name = tensor("values_91_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor values_91_cast_fp16 = slice_by_index(begin = values_91_begin_0, end = values_91_end_0, end_mask = values_91_end_mask_0, squeeze_mask = values_91_squeeze_mask_0, x = reshape_159_cast_fp16)[name = tensor("values_91_cast_fp16")]; + tensor var_6629_cast_fp16 = not_equal(x = keys_91_cast_fp16, y = keys_91_cast_fp16)[name = tensor("op_6629_cast_fp16")]; + tensor keys_93_cast_fp16 = select(a = var_470_to_fp16, b = keys_91_cast_fp16, cond = var_6629_cast_fp16)[name = tensor("keys_93_cast_fp16")]; + tensor var_6637_cast_fp16 = not_equal(x = values_91_cast_fp16, y = values_91_cast_fp16)[name = tensor("op_6637_cast_fp16")]; + tensor values_93_cast_fp16 = select(a = var_470_to_fp16, b = values_91_cast_fp16, cond = var_6637_cast_fp16)[name = tensor("values_93_cast_fp16")]; + tensor var_6661 = const()[name = tensor("op_6661"), val = tensor([0, 2, 1, 3])]; + tensor var_6677 = const()[name = tensor("op_6677"), val = tensor([1, 1, 1])]; + tensor var_6678_cast_fp16 = reshape(shape = var_6677, x = position15_to_fp16)[name = tensor("op_6678_cast_fp16")]; + tensor q_positions_31_cast_fp16 = add(x = var_6678_cast_fp16, y = q_offsets_1_promoted_to_fp16)[name = tensor("q_positions_31_cast_fp16")]; + tensor valid_end_31_cast_fp16 = add(x = var_6678_cast_fp16, y = var_536_cast_fp16)[name = tensor("valid_end_31_cast_fp16")]; + tensor valid_mask_31_cast_fp16 = less(x = k_positions_1_promoted_to_fp16, y = valid_end_31_cast_fp16)[name = tensor("valid_mask_31_cast_fp16")]; + tensor causal_mask_31_cast_fp16 = less_equal(x = k_positions_1_promoted_to_fp16, y = q_positions_31_cast_fp16)[name = tensor("causal_mask_31_cast_fp16")]; + tensor var_6706 = logical_and(x = valid_mask_31_cast_fp16, y = causal_mask_31_cast_fp16)[name = tensor("op_6706")]; + tensor attn_mask_31_axes_0 = const()[name = tensor("attn_mask_31_axes_0"), val = tensor([1])]; + tensor attn_mask_31 = expand_dims(axes = attn_mask_31_axes_0, x = var_6706)[name = tensor("attn_mask_31")]; + tensor var_6720_transpose_x_0 = const()[name = tensor("op_6720_transpose_x_0"), val = tensor(false)]; + tensor var_6720_transpose_y_0 = const()[name = tensor("op_6720_transpose_y_0"), val = tensor(false)]; + tensor transpose_99_perm_0 = const()[name = tensor("transpose_99_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_100_perm_0 = const()[name = tensor("transpose_100_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_100 = transpose(perm = transpose_100_perm_0, x = keys_93_cast_fp16)[name = tensor("transpose_144")]; + tensor transpose_99 = transpose(perm = transpose_99_perm_0, x = q_93_cast_fp16)[name = tensor("transpose_145")]; + tensor var_6720_cast_fp16 = matmul(transpose_x = var_6720_transpose_x_0, transpose_y = var_6720_transpose_y_0, x = transpose_99, y = transpose_100)[name = tensor("op_6720_cast_fp16")]; + tensor var_6714_to_fp16 = const()[name = tensor("op_6714_to_fp16"), val = tensor([0x1p-3])]; + tensor attn_weights_91_cast_fp16 = mul(x = var_6720_cast_fp16, y = var_6714_to_fp16)[name = tensor("attn_weights_91_cast_fp16")]; + tensor var_6722 = logical_not(x = attn_mask_31)[name = tensor("op_6722")]; + tensor var_6723_to_fp16 = const()[name = tensor("op_6723_to_fp16"), val = tensor(-inf)]; + tensor attn_weights_93_cast_fp16 = select(a = var_6723_to_fp16, b = attn_weights_91_cast_fp16, cond = var_6722)[name = tensor("attn_weights_93_cast_fp16")]; + tensor var_6725 = const()[name = tensor("op_6725"), val = tensor(-1)]; + tensor attn_weights_95_cast_fp16 = softmax(axis = var_6725, x = attn_weights_93_cast_fp16)[name = tensor("attn_weights_95_cast_fp16")]; + tensor attn_output_31_transpose_x_0 = const()[name = tensor("attn_output_31_transpose_x_0"), val = tensor(false)]; + tensor attn_output_31_transpose_y_0 = const()[name = tensor("attn_output_31_transpose_y_0"), val = tensor(false)]; + tensor values_95_cast_fp16 = transpose(perm = var_6661, x = values_93_cast_fp16)[name = tensor("transpose_146")]; + tensor attn_output_31_cast_fp16 = matmul(transpose_x = attn_output_31_transpose_x_0, transpose_y = attn_output_31_transpose_y_0, x = attn_weights_95_cast_fp16, y = values_95_cast_fp16)[name = tensor("attn_output_31_cast_fp16")]; + tensor var_6733 = const()[name = tensor("op_6733"), val = tensor([0, 2, 1, 3])]; + tensor var_6738 = const()[name = tensor("op_6738"), val = tensor([1, 256, 1024])]; + tensor var_6734_cast_fp16 = transpose(perm = var_6733, x = attn_output_31_cast_fp16)[name = tensor("transpose_143")]; + tensor input_185_cast_fp16 = reshape(shape = var_6738, x = var_6734_cast_fp16)[name = tensor("input_185_cast_fp16")]; + tensor attn15_out_proj_weight_to_fp16 = const()[name = tensor("attn15_out_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(384983808)))]; + tensor linear_61_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = attn15_out_proj_weight_to_fp16, x = input_185_cast_fp16)[name = tensor("linear_61_cast_fp16")]; + tensor var_6747_cast_fp16 = add(x = position15_to_fp16, y = valid_len_to_fp16)[name = tensor("op_6747_cast_fp16")]; + tensor var_6747_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_6747_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor input_187_cast_fp16 = add(x = input_181_cast_fp16, y = linear_61_cast_fp16)[name = tensor("input_187_cast_fp16")]; + tensor input_189_axes_0 = const()[name = tensor("input_189_axes_0"), val = tensor([-1])]; + tensor norm15_2_weight_to_fp16 = const()[name = tensor("norm15_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(387081024)))]; + tensor norm15_2_bias_to_fp16 = const()[name = tensor("norm15_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(387083136)))]; + tensor var_6751_to_fp16 = const()[name = tensor("op_6751_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_189_cast_fp16 = layer_norm(axes = input_189_axes_0, beta = norm15_2_bias_to_fp16, epsilon = var_6751_to_fp16, gamma = norm15_2_weight_to_fp16, x = input_187_cast_fp16)[name = tensor("input_189_cast_fp16")]; + tensor linear15_1_weight_to_fp16 = const()[name = tensor("linear15_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(387085248)))]; + tensor linear_62_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = linear15_1_weight_to_fp16, x = input_189_cast_fp16)[name = tensor("linear_62_cast_fp16")]; + tensor input_191_mode_0 = const()[name = tensor("input_191_mode_0"), val = tensor("EXACT")]; + tensor input_191_cast_fp16 = gelu(mode = input_191_mode_0, x = linear_62_cast_fp16)[name = tensor("input_191_cast_fp16")]; + tensor linear15_2_weight_to_fp16 = const()[name = tensor("linear15_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(395473920)))]; + tensor linear_63_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = linear15_2_weight_to_fp16, x = input_191_cast_fp16)[name = tensor("linear_63_cast_fp16")]; + tensor input_193_cast_fp16 = add(x = input_187_cast_fp16, y = linear_63_cast_fp16)[name = tensor("input_193_cast_fp16")]; + tensor input_195_axes_0 = const()[name = tensor("input_195_axes_0"), val = tensor([-1])]; + tensor norm16_1_weight_to_fp16 = const()[name = tensor("norm16_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(403862592)))]; + tensor norm16_1_bias_to_fp16 = const()[name = tensor("norm16_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(403864704)))]; + tensor var_6768_to_fp16 = const()[name = tensor("op_6768_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_195_cast_fp16 = layer_norm(axes = input_195_axes_0, beta = norm16_1_bias_to_fp16, epsilon = var_6768_to_fp16, gamma = norm16_1_weight_to_fp16, x = input_193_cast_fp16)[name = tensor("input_195_cast_fp16")]; + tensor attn16_in_proj_weight_to_fp16 = const()[name = tensor("attn16_in_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(403866816)))]; + tensor linear_64_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = attn16_in_proj_weight_to_fp16, x = input_195_cast_fp16)[name = tensor("linear_64_cast_fp16")]; + tensor var_6782 = const()[name = tensor("op_6782"), val = tensor([1, 256, 3, 16, 64])]; + tensor qkv_33_cast_fp16 = reshape(shape = var_6782, x = linear_64_cast_fp16)[name = tensor("qkv_33_cast_fp16")]; + tensor q_97_begin_0 = const()[name = tensor("q_97_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor q_97_end_0 = const()[name = tensor("q_97_end_0"), val = tensor([1, 256, 1, 16, 64])]; + tensor q_97_end_mask_0 = const()[name = tensor("q_97_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor q_97_squeeze_mask_0 = const()[name = tensor("q_97_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor q_97_cast_fp16 = slice_by_index(begin = q_97_begin_0, end = q_97_end_0, end_mask = q_97_end_mask_0, squeeze_mask = q_97_squeeze_mask_0, x = qkv_33_cast_fp16)[name = tensor("q_97_cast_fp16")]; + tensor k_65_begin_0 = const()[name = tensor("k_65_begin_0"), val = tensor([0, 0, 1, 0, 0])]; + tensor k_65_end_0 = const()[name = tensor("k_65_end_0"), val = tensor([1, 256, 2, 16, 64])]; + tensor k_65_end_mask_0 = const()[name = tensor("k_65_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor k_65_squeeze_mask_0 = const()[name = tensor("k_65_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor k_65_cast_fp16 = slice_by_index(begin = k_65_begin_0, end = k_65_end_0, end_mask = k_65_end_mask_0, squeeze_mask = k_65_squeeze_mask_0, x = qkv_33_cast_fp16)[name = tensor("k_65_cast_fp16")]; + tensor v_33_begin_0 = const()[name = tensor("v_33_begin_0"), val = tensor([0, 0, 2, 0, 0])]; + tensor v_33_end_0 = const()[name = tensor("v_33_end_0"), val = tensor([1, 256, 3, 16, 64])]; + tensor v_33_end_mask_0 = const()[name = tensor("v_33_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor v_33_squeeze_mask_0 = const()[name = tensor("v_33_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor v_33_cast_fp16 = slice_by_index(begin = v_33_begin_0, end = v_33_end_0, end_mask = v_33_end_mask_0, squeeze_mask = v_33_squeeze_mask_0, x = qkv_33_cast_fp16)[name = tensor("v_33_cast_fp16")]; + tensor var_6840 = const()[name = tensor("op_6840"), val = tensor([1, 1])]; + tensor position16_to_fp16_dtype_0 = const()[name = tensor("position16_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor position16_to_fp16 = cast(dtype = position16_to_fp16_dtype_0, x = position16)[name = tensor("cast_581")]; + tensor var_6841_cast_fp16 = reshape(shape = var_6840, x = position16_to_fp16)[name = tensor("op_6841_cast_fp16")]; + tensor ts_99_cast_fp16 = add(x = ts_1_promoted_to_fp16, y = var_6841_cast_fp16)[name = tensor("ts_99_cast_fp16")]; + tensor var_6848 = const()[name = tensor("op_6848"), val = tensor([1, 256, 1, 1])]; + tensor ts_101_cast_fp16 = reshape(shape = var_6848, x = ts_99_cast_fp16)[name = tensor("ts_101_cast_fp16")]; + tensor var_6855 = const()[name = tensor("op_6855"), val = tensor([1, 256, 16, 32, 2])]; + tensor q_complex_33_cast_fp16 = reshape(shape = var_6855, x = q_97_cast_fp16)[name = tensor("q_complex_33_cast_fp16")]; + tensor var_6862 = const()[name = tensor("op_6862"), val = tensor([1, 256, 16, 32, 2])]; + tensor k_complex_33_cast_fp16 = reshape(shape = var_6862, x = k_65_cast_fp16)[name = tensor("k_complex_33_cast_fp16")]; + tensor var_6866_begin_0 = const()[name = tensor("op_6866_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_6866_end_0 = const()[name = tensor("op_6866_end_0"), val = tensor([1, 256, 16, 32, 1])]; + tensor var_6866_end_mask_0 = const()[name = tensor("op_6866_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_6866_squeeze_mask_0 = const()[name = tensor("op_6866_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_6866_cast_fp16 = slice_by_index(begin = var_6866_begin_0, end = var_6866_end_0, end_mask = var_6866_end_mask_0, squeeze_mask = var_6866_squeeze_mask_0, x = q_complex_33_cast_fp16)[name = tensor("op_6866_cast_fp16")]; + tensor var_6874_begin_0 = const()[name = tensor("op_6874_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_6874_end_0 = const()[name = tensor("op_6874_end_0"), val = tensor([1, 256, 16, 32, 2])]; + tensor var_6874_end_mask_0 = const()[name = tensor("op_6874_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_6874_squeeze_mask_0 = const()[name = tensor("op_6874_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_6874_cast_fp16 = slice_by_index(begin = var_6874_begin_0, end = var_6874_end_0, end_mask = var_6874_end_mask_0, squeeze_mask = var_6874_squeeze_mask_0, x = q_complex_33_cast_fp16)[name = tensor("op_6874_cast_fp16")]; + tensor var_6882_begin_0 = const()[name = tensor("op_6882_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_6882_end_0 = const()[name = tensor("op_6882_end_0"), val = tensor([1, 256, 16, 32, 1])]; + tensor var_6882_end_mask_0 = const()[name = tensor("op_6882_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_6882_squeeze_mask_0 = const()[name = tensor("op_6882_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_6882_cast_fp16 = slice_by_index(begin = var_6882_begin_0, end = var_6882_end_0, end_mask = var_6882_end_mask_0, squeeze_mask = var_6882_squeeze_mask_0, x = k_complex_33_cast_fp16)[name = tensor("op_6882_cast_fp16")]; + tensor var_6890_begin_0 = const()[name = tensor("op_6890_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_6890_end_0 = const()[name = tensor("op_6890_end_0"), val = tensor([1, 256, 16, 32, 2])]; + tensor var_6890_end_mask_0 = const()[name = tensor("op_6890_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_6890_squeeze_mask_0 = const()[name = tensor("op_6890_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_6890_cast_fp16 = slice_by_index(begin = var_6890_begin_0, end = var_6890_end_0, end_mask = var_6890_end_mask_0, squeeze_mask = var_6890_squeeze_mask_0, x = k_complex_33_cast_fp16)[name = tensor("op_6890_cast_fp16")]; + tensor freqs_33_to_fp16 = const()[name = tensor("freqs_33_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(410158336)))]; + tensor var_6896_cast_fp16 = mul(x = freqs_33_to_fp16, y = ts_101_cast_fp16)[name = tensor("op_6896_cast_fp16")]; + tensor rotr_33_cast_fp16 = cos(x = var_6896_cast_fp16)[name = tensor("rotr_33_cast_fp16")]; + tensor roti_33_cast_fp16 = sin(x = var_6896_cast_fp16)[name = tensor("roti_33_cast_fp16")]; + tensor var_6900_cast_fp16 = mul(x = var_6866_cast_fp16, y = rotr_33_cast_fp16)[name = tensor("op_6900_cast_fp16")]; + tensor var_6901_cast_fp16 = mul(x = var_6874_cast_fp16, y = roti_33_cast_fp16)[name = tensor("op_6901_cast_fp16")]; + tensor qor_65_cast_fp16 = sub(x = var_6900_cast_fp16, y = var_6901_cast_fp16)[name = tensor("qor_65_cast_fp16")]; + tensor var_6904_cast_fp16 = mul(x = var_6866_cast_fp16, y = roti_33_cast_fp16)[name = tensor("op_6904_cast_fp16")]; + tensor var_6905_cast_fp16 = mul(x = var_6874_cast_fp16, y = rotr_33_cast_fp16)[name = tensor("op_6905_cast_fp16")]; + tensor qoi_65_cast_fp16 = add(x = var_6904_cast_fp16, y = var_6905_cast_fp16)[name = tensor("qoi_65_cast_fp16")]; + tensor var_6908_cast_fp16 = mul(x = var_6882_cast_fp16, y = rotr_33_cast_fp16)[name = tensor("op_6908_cast_fp16")]; + tensor var_6909_cast_fp16 = mul(x = var_6890_cast_fp16, y = roti_33_cast_fp16)[name = tensor("op_6909_cast_fp16")]; + tensor kor_65_cast_fp16 = sub(x = var_6908_cast_fp16, y = var_6909_cast_fp16)[name = tensor("kor_65_cast_fp16")]; + tensor var_6912_cast_fp16 = mul(x = var_6882_cast_fp16, y = roti_33_cast_fp16)[name = tensor("op_6912_cast_fp16")]; + tensor var_6913_cast_fp16 = mul(x = var_6890_cast_fp16, y = rotr_33_cast_fp16)[name = tensor("op_6913_cast_fp16")]; + tensor koi_65_cast_fp16 = add(x = var_6912_cast_fp16, y = var_6913_cast_fp16)[name = tensor("koi_65_cast_fp16")]; + tensor qo_33_axis_0 = const()[name = tensor("qo_33_axis_0"), val = tensor(-1)]; + tensor qo_33_cast_fp16 = stack(axis = qo_33_axis_0, values = (qor_65_cast_fp16, qoi_65_cast_fp16))[name = tensor("qo_33_cast_fp16")]; + tensor ko_33_axis_0 = const()[name = tensor("ko_33_axis_0"), val = tensor(-1)]; + tensor ko_33_cast_fp16 = stack(axis = ko_33_axis_0, values = (kor_65_cast_fp16, koi_65_cast_fp16))[name = tensor("ko_33_cast_fp16")]; + tensor var_6946 = const()[name = tensor("op_6946"), val = tensor([1, 256, 16, 64])]; + tensor q_99_cast_fp16 = reshape(shape = var_6946, x = qo_33_cast_fp16)[name = tensor("q_99_cast_fp16")]; + tensor var_6952 = const()[name = tensor("op_6952"), val = tensor([1, 256, 16, 64])]; + tensor k_67_cast_fp16 = reshape(shape = var_6952, x = ko_33_cast_fp16)[name = tensor("k_67_cast_fp16")]; + tensor write_indices_float_97_cast_fp16 = add(x = var_6841_cast_fp16, y = var_393_promoted_to_fp16)[name = tensor("write_indices_float_97_cast_fp16")]; + tensor _inversed_6973_y_0_to_fp16 = const()[name = tensor("_inversed_6973_y_0_to_fp16"), val = tensor(0x1p-9)]; + tensor _inversed_6973_cast_fp16 = mul(x = write_indices_float_97_cast_fp16, y = _inversed_6973_y_0_to_fp16)[name = tensor("_inversed_6973_cast_fp16")]; + tensor var_6974_cast_fp16 = floor(x = _inversed_6973_cast_fp16)[name = tensor("op_6974_cast_fp16")]; + tensor var_6975_to_fp16 = const()[name = tensor("op_6975_to_fp16"), val = tensor(0x1p+9)]; + tensor var_6976_cast_fp16 = mul(x = var_6974_cast_fp16, y = var_6975_to_fp16)[name = tensor("op_6976_cast_fp16")]; + tensor write_indices_float_99_cast_fp16 = sub(x = write_indices_float_97_cast_fp16, y = var_6976_cast_fp16)[name = tensor("write_indices_float_99_cast_fp16")]; + tensor var_6994_value_0_to_fp16 = const()[name = tensor("op_6994_value_0_to_fp16"), val = tensor(0x1.ffp+8)]; + tensor var_6994_cast_fp16 = fill_like(ref_tensor = write_indices_float_99_cast_fp16, value = var_6994_value_0_to_fp16)[name = tensor("op_6994_cast_fp16")]; + tensor write_indices_float_101_cast_fp16 = select(a = write_indices_float_99_cast_fp16, b = var_6994_cast_fp16, cond = token_valid_1_cast_fp16)[name = tensor("write_indices_float_101_cast_fp16")]; + tensor var_7000_dtype_0 = const()[name = tensor("op_7000_dtype_0"), val = tensor("int32")]; + tensor var_7005 = const()[name = tensor("op_7005"), val = tensor([1, 256, 1, 1])]; + tensor write_indices_float_101_cast_fp16_to_int32 = cast(dtype = var_7000_dtype_0, x = write_indices_float_101_cast_fp16)[name = tensor("cast_580")]; + tensor var_7006 = reshape(shape = var_7005, x = write_indices_float_101_cast_fp16_to_int32)[name = tensor("op_7006")]; + tensor write_indices_33_reps_0 = const()[name = tensor("write_indices_33_reps_0"), val = tensor([1, 1, 16, 64])]; + tensor write_indices_33 = tile(reps = write_indices_33_reps_0, x = var_7006)[name = tensor("write_indices_33")]; + tensor var_7016_begin_0 = const()[name = tensor("op_7016_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_7016_end_0 = const()[name = tensor("op_7016_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor var_7016_end_mask_0 = const()[name = tensor("op_7016_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_7016_squeeze_mask_0 = const()[name = tensor("op_7016_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor cache16_to_fp16_dtype_0 = const()[name = tensor("cache16_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor cache16_to_fp16 = cast(dtype = cache16_to_fp16_dtype_0, x = cache16)[name = tensor("cast_579")]; + tensor var_7016_cast_fp16 = slice_by_index(begin = var_7016_begin_0, end = var_7016_end_0, end_mask = var_7016_end_mask_0, squeeze_mask = var_7016_squeeze_mask_0, x = cache16_to_fp16)[name = tensor("op_7016_cast_fp16")]; + tensor var_7018_axis_0 = const()[name = tensor("op_7018_axis_0"), val = tensor(1)]; + tensor var_7018_mode_0 = const()[name = tensor("op_7018_mode_0"), val = tensor("update")]; + tensor var_7018_validate_indices_0 = const()[name = tensor("op_7018_validate_indices_0"), val = tensor(false)]; + tensor var_7018_cast_fp16 = scatter_along_axis(axis = var_7018_axis_0, data = var_7016_cast_fp16, indices = write_indices_33, mode = var_7018_mode_0, updates = k_67_cast_fp16, validate_indices = var_7018_validate_indices_0)[name = tensor("op_7018_cast_fp16")]; + tensor concat_130 = const()[name = tensor("concat_130"), val = tensor([0, 0, 0, 0, 0])]; + tensor concat_131 = const()[name = tensor("concat_131"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_33_internal_tensor_assign_1_stride_0 = const()[name = tensor("new_cache_33_internal_tensor_assign_1_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_33_internal_tensor_assign_1_begin_mask_0 = const()[name = tensor("new_cache_33_internal_tensor_assign_1_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_33_internal_tensor_assign_1_end_mask_0 = const()[name = tensor("new_cache_33_internal_tensor_assign_1_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_33_internal_tensor_assign_1_squeeze_mask_0 = const()[name = tensor("new_cache_33_internal_tensor_assign_1_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_78 = const()[name = tensor("shape_78"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_32 = const()[name = tensor("reduce_prod_32"), val = tensor(1048576)]; + tensor range_1d_32_start_0 = const()[name = tensor("range_1d_32_start_0"), val = tensor(0)]; + tensor range_1d_32_step_0 = const()[name = tensor("range_1d_32_step_0"), val = tensor(1)]; + tensor range_1d_32 = range_1d(end = reduce_prod_32, start = range_1d_32_start_0, step = range_1d_32_step_0)[name = tensor("range_1d_32")]; + tensor reshape_160 = reshape(shape = shape_78, x = range_1d_32)[name = tensor("reshape_160")]; + tensor slice_by_index_32 = slice_by_index(begin = concat_130, begin_mask = new_cache_33_internal_tensor_assign_1_begin_mask_0, end = concat_131, end_mask = new_cache_33_internal_tensor_assign_1_end_mask_0, squeeze_mask = new_cache_33_internal_tensor_assign_1_squeeze_mask_0, stride = new_cache_33_internal_tensor_assign_1_stride_0, x = reshape_160)[name = tensor("slice_by_index_32")]; + tensor reshape_161_shape_0 = const()[name = tensor("reshape_161_shape_0"), val = tensor([-1])]; + tensor reshape_161 = reshape(shape = reshape_161_shape_0, x = slice_by_index_32)[name = tensor("reshape_161")]; + tensor reshape_162_shape_0 = const()[name = tensor("reshape_162_shape_0"), val = tensor([-1])]; + tensor reshape_162_cast_fp16 = reshape(shape = reshape_162_shape_0, x = var_7018_cast_fp16)[name = tensor("reshape_162_cast_fp16")]; + tensor reshape_163_shape_0 = const()[name = tensor("reshape_163_shape_0"), val = tensor([-1])]; + tensor reshape_163_cast_fp16 = reshape(shape = reshape_163_shape_0, x = cache16_to_fp16)[name = tensor("reshape_163_cast_fp16")]; + tensor scatter_32_mode_0 = const()[name = tensor("scatter_32_mode_0"), val = tensor("update")]; + tensor scatter_32_axis_0 = const()[name = tensor("scatter_32_axis_0"), val = tensor(0)]; + tensor scatter_32_validate_indices_0 = const()[name = tensor("scatter_32_validate_indices_0"), val = tensor(false)]; + tensor scatter_32_cast_fp16 = scatter(axis = scatter_32_axis_0, data = reshape_163_cast_fp16, indices = reshape_161, mode = scatter_32_mode_0, updates = reshape_162_cast_fp16, validate_indices = scatter_32_validate_indices_0)[name = tensor("scatter_32_cast_fp16")]; + tensor reshape_164_cast_fp16 = reshape(shape = shape_78, x = scatter_32_cast_fp16)[name = tensor("reshape_164_cast_fp16")]; + tensor var_7026_begin_0 = const()[name = tensor("op_7026_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor var_7026_end_0 = const()[name = tensor("op_7026_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor var_7026_end_mask_0 = const()[name = tensor("op_7026_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_7026_squeeze_mask_0 = const()[name = tensor("op_7026_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor var_7026_cast_fp16 = slice_by_index(begin = var_7026_begin_0, end = var_7026_end_0, end_mask = var_7026_end_mask_0, squeeze_mask = var_7026_squeeze_mask_0, x = reshape_164_cast_fp16)[name = tensor("op_7026_cast_fp16")]; + tensor var_7028_axis_0 = const()[name = tensor("op_7028_axis_0"), val = tensor(1)]; + tensor var_7028_mode_0 = const()[name = tensor("op_7028_mode_0"), val = tensor("update")]; + tensor var_7028_validate_indices_0 = const()[name = tensor("op_7028_validate_indices_0"), val = tensor(false)]; + tensor var_7028_cast_fp16 = scatter_along_axis(axis = var_7028_axis_0, data = var_7026_cast_fp16, indices = write_indices_33, mode = var_7028_mode_0, updates = v_33_cast_fp16, validate_indices = var_7028_validate_indices_0)[name = tensor("op_7028_cast_fp16")]; + tensor concat_132 = const()[name = tensor("concat_132"), val = tensor([1, 0, 0, 0, 0])]; + tensor concat_133 = const()[name = tensor("concat_133"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_33_internal_tensor_assign_2_stride_0 = const()[name = tensor("new_cache_33_internal_tensor_assign_2_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_33_internal_tensor_assign_2_begin_mask_0 = const()[name = tensor("new_cache_33_internal_tensor_assign_2_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_33_internal_tensor_assign_2_end_mask_0 = const()[name = tensor("new_cache_33_internal_tensor_assign_2_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_33_internal_tensor_assign_2_squeeze_mask_0 = const()[name = tensor("new_cache_33_internal_tensor_assign_2_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_79 = const()[name = tensor("shape_79"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_33 = const()[name = tensor("reduce_prod_33"), val = tensor(1048576)]; + tensor range_1d_33_start_0 = const()[name = tensor("range_1d_33_start_0"), val = tensor(0)]; + tensor range_1d_33_step_0 = const()[name = tensor("range_1d_33_step_0"), val = tensor(1)]; + tensor range_1d_33 = range_1d(end = reduce_prod_33, start = range_1d_33_start_0, step = range_1d_33_step_0)[name = tensor("range_1d_33")]; + tensor reshape_165 = reshape(shape = shape_79, x = range_1d_33)[name = tensor("reshape_165")]; + tensor slice_by_index_33 = slice_by_index(begin = concat_132, begin_mask = new_cache_33_internal_tensor_assign_2_begin_mask_0, end = concat_133, end_mask = new_cache_33_internal_tensor_assign_2_end_mask_0, squeeze_mask = new_cache_33_internal_tensor_assign_2_squeeze_mask_0, stride = new_cache_33_internal_tensor_assign_2_stride_0, x = reshape_165)[name = tensor("slice_by_index_33")]; + tensor reshape_166_shape_0 = const()[name = tensor("reshape_166_shape_0"), val = tensor([-1])]; + tensor reshape_166 = reshape(shape = reshape_166_shape_0, x = slice_by_index_33)[name = tensor("reshape_166")]; + tensor reshape_167_shape_0 = const()[name = tensor("reshape_167_shape_0"), val = tensor([-1])]; + tensor reshape_167_cast_fp16 = reshape(shape = reshape_167_shape_0, x = var_7028_cast_fp16)[name = tensor("reshape_167_cast_fp16")]; + tensor reshape_168_shape_0 = const()[name = tensor("reshape_168_shape_0"), val = tensor([-1])]; + tensor reshape_168_cast_fp16 = reshape(shape = reshape_168_shape_0, x = reshape_164_cast_fp16)[name = tensor("reshape_168_cast_fp16")]; + tensor scatter_33_mode_0 = const()[name = tensor("scatter_33_mode_0"), val = tensor("update")]; + tensor scatter_33_axis_0 = const()[name = tensor("scatter_33_axis_0"), val = tensor(0)]; + tensor scatter_33_validate_indices_0 = const()[name = tensor("scatter_33_validate_indices_0"), val = tensor(false)]; + tensor scatter_33_cast_fp16 = scatter(axis = scatter_33_axis_0, data = reshape_168_cast_fp16, indices = reshape_166, mode = scatter_33_mode_0, updates = reshape_167_cast_fp16, validate_indices = scatter_33_validate_indices_0)[name = tensor("scatter_33_cast_fp16")]; + tensor reshape_169_cast_fp16 = reshape(shape = shape_79, x = scatter_33_cast_fp16)[name = tensor("reshape_169_cast_fp16")]; + tensor reshape_169_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("reshape_169_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor keys_97_begin_0 = const()[name = tensor("keys_97_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor keys_97_end_0 = const()[name = tensor("keys_97_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor keys_97_end_mask_0 = const()[name = tensor("keys_97_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor keys_97_squeeze_mask_0 = const()[name = tensor("keys_97_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor keys_97_cast_fp16 = slice_by_index(begin = keys_97_begin_0, end = keys_97_end_0, end_mask = keys_97_end_mask_0, squeeze_mask = keys_97_squeeze_mask_0, x = reshape_169_cast_fp16)[name = tensor("keys_97_cast_fp16")]; + tensor values_97_begin_0 = const()[name = tensor("values_97_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor values_97_end_0 = const()[name = tensor("values_97_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor values_97_end_mask_0 = const()[name = tensor("values_97_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor values_97_squeeze_mask_0 = const()[name = tensor("values_97_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor values_97_cast_fp16 = slice_by_index(begin = values_97_begin_0, end = values_97_end_0, end_mask = values_97_end_mask_0, squeeze_mask = values_97_squeeze_mask_0, x = reshape_169_cast_fp16)[name = tensor("values_97_cast_fp16")]; + tensor var_7040_cast_fp16 = not_equal(x = keys_97_cast_fp16, y = keys_97_cast_fp16)[name = tensor("op_7040_cast_fp16")]; + tensor keys_99_cast_fp16 = select(a = var_470_to_fp16, b = keys_97_cast_fp16, cond = var_7040_cast_fp16)[name = tensor("keys_99_cast_fp16")]; + tensor var_7048_cast_fp16 = not_equal(x = values_97_cast_fp16, y = values_97_cast_fp16)[name = tensor("op_7048_cast_fp16")]; + tensor values_99_cast_fp16 = select(a = var_470_to_fp16, b = values_97_cast_fp16, cond = var_7048_cast_fp16)[name = tensor("values_99_cast_fp16")]; + tensor var_7072 = const()[name = tensor("op_7072"), val = tensor([0, 2, 1, 3])]; + tensor var_7088 = const()[name = tensor("op_7088"), val = tensor([1, 1, 1])]; + tensor var_7089_cast_fp16 = reshape(shape = var_7088, x = position16_to_fp16)[name = tensor("op_7089_cast_fp16")]; + tensor q_positions_33_cast_fp16 = add(x = var_7089_cast_fp16, y = q_offsets_1_promoted_to_fp16)[name = tensor("q_positions_33_cast_fp16")]; + tensor valid_end_33_cast_fp16 = add(x = var_7089_cast_fp16, y = var_536_cast_fp16)[name = tensor("valid_end_33_cast_fp16")]; + tensor valid_mask_33_cast_fp16 = less(x = k_positions_1_promoted_to_fp16, y = valid_end_33_cast_fp16)[name = tensor("valid_mask_33_cast_fp16")]; + tensor causal_mask_33_cast_fp16 = less_equal(x = k_positions_1_promoted_to_fp16, y = q_positions_33_cast_fp16)[name = tensor("causal_mask_33_cast_fp16")]; + tensor var_7117 = logical_and(x = valid_mask_33_cast_fp16, y = causal_mask_33_cast_fp16)[name = tensor("op_7117")]; + tensor attn_mask_33_axes_0 = const()[name = tensor("attn_mask_33_axes_0"), val = tensor([1])]; + tensor attn_mask_33 = expand_dims(axes = attn_mask_33_axes_0, x = var_7117)[name = tensor("attn_mask_33")]; + tensor var_7131_transpose_x_0 = const()[name = tensor("op_7131_transpose_x_0"), val = tensor(false)]; + tensor var_7131_transpose_y_0 = const()[name = tensor("op_7131_transpose_y_0"), val = tensor(false)]; + tensor transpose_101_perm_0 = const()[name = tensor("transpose_101_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_102_perm_0 = const()[name = tensor("transpose_102_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_102 = transpose(perm = transpose_102_perm_0, x = keys_99_cast_fp16)[name = tensor("transpose_140")]; + tensor transpose_101 = transpose(perm = transpose_101_perm_0, x = q_99_cast_fp16)[name = tensor("transpose_141")]; + tensor var_7131_cast_fp16 = matmul(transpose_x = var_7131_transpose_x_0, transpose_y = var_7131_transpose_y_0, x = transpose_101, y = transpose_102)[name = tensor("op_7131_cast_fp16")]; + tensor var_7125_to_fp16 = const()[name = tensor("op_7125_to_fp16"), val = tensor([0x1p-3])]; + tensor attn_weights_97_cast_fp16 = mul(x = var_7131_cast_fp16, y = var_7125_to_fp16)[name = tensor("attn_weights_97_cast_fp16")]; + tensor var_7133 = logical_not(x = attn_mask_33)[name = tensor("op_7133")]; + tensor var_7134_to_fp16 = const()[name = tensor("op_7134_to_fp16"), val = tensor(-inf)]; + tensor attn_weights_99_cast_fp16 = select(a = var_7134_to_fp16, b = attn_weights_97_cast_fp16, cond = var_7133)[name = tensor("attn_weights_99_cast_fp16")]; + tensor var_7136 = const()[name = tensor("op_7136"), val = tensor(-1)]; + tensor attn_weights_101_cast_fp16 = softmax(axis = var_7136, x = attn_weights_99_cast_fp16)[name = tensor("attn_weights_101_cast_fp16")]; + tensor attn_output_33_transpose_x_0 = const()[name = tensor("attn_output_33_transpose_x_0"), val = tensor(false)]; + tensor attn_output_33_transpose_y_0 = const()[name = tensor("attn_output_33_transpose_y_0"), val = tensor(false)]; + tensor values_101_cast_fp16 = transpose(perm = var_7072, x = values_99_cast_fp16)[name = tensor("transpose_142")]; + tensor attn_output_33_cast_fp16 = matmul(transpose_x = attn_output_33_transpose_x_0, transpose_y = attn_output_33_transpose_y_0, x = attn_weights_101_cast_fp16, y = values_101_cast_fp16)[name = tensor("attn_output_33_cast_fp16")]; + tensor var_7144 = const()[name = tensor("op_7144"), val = tensor([0, 2, 1, 3])]; + tensor var_7149 = const()[name = tensor("op_7149"), val = tensor([1, 256, 1024])]; + tensor var_7145_cast_fp16 = transpose(perm = var_7144, x = attn_output_33_cast_fp16)[name = tensor("transpose_139")]; + tensor input_197_cast_fp16 = reshape(shape = var_7149, x = var_7145_cast_fp16)[name = tensor("input_197_cast_fp16")]; + tensor attn16_out_proj_weight_to_fp16 = const()[name = tensor("attn16_out_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(410158464)))]; + tensor linear_65_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = attn16_out_proj_weight_to_fp16, x = input_197_cast_fp16)[name = tensor("linear_65_cast_fp16")]; + tensor var_7158_cast_fp16 = add(x = position16_to_fp16, y = valid_len_to_fp16)[name = tensor("op_7158_cast_fp16")]; + tensor var_7158_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_7158_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor input_199_cast_fp16 = add(x = input_193_cast_fp16, y = linear_65_cast_fp16)[name = tensor("input_199_cast_fp16")]; + tensor input_201_axes_0 = const()[name = tensor("input_201_axes_0"), val = tensor([-1])]; + tensor norm16_2_weight_to_fp16 = const()[name = tensor("norm16_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(412255680)))]; + tensor norm16_2_bias_to_fp16 = const()[name = tensor("norm16_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(412257792)))]; + tensor var_7162_to_fp16 = const()[name = tensor("op_7162_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_201_cast_fp16 = layer_norm(axes = input_201_axes_0, beta = norm16_2_bias_to_fp16, epsilon = var_7162_to_fp16, gamma = norm16_2_weight_to_fp16, x = input_199_cast_fp16)[name = tensor("input_201_cast_fp16")]; + tensor linear16_1_weight_to_fp16 = const()[name = tensor("linear16_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(412259904)))]; + tensor linear_66_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = linear16_1_weight_to_fp16, x = input_201_cast_fp16)[name = tensor("linear_66_cast_fp16")]; + tensor input_203_mode_0 = const()[name = tensor("input_203_mode_0"), val = tensor("EXACT")]; + tensor input_203_cast_fp16 = gelu(mode = input_203_mode_0, x = linear_66_cast_fp16)[name = tensor("input_203_cast_fp16")]; + tensor linear16_2_weight_to_fp16 = const()[name = tensor("linear16_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(420648576)))]; + tensor linear_67_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = linear16_2_weight_to_fp16, x = input_203_cast_fp16)[name = tensor("linear_67_cast_fp16")]; + tensor input_205_cast_fp16 = add(x = input_199_cast_fp16, y = linear_67_cast_fp16)[name = tensor("input_205_cast_fp16")]; + tensor input_207_axes_0 = const()[name = tensor("input_207_axes_0"), val = tensor([-1])]; + tensor norm17_1_weight_to_fp16 = const()[name = tensor("norm17_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(429037248)))]; + tensor norm17_1_bias_to_fp16 = const()[name = tensor("norm17_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(429039360)))]; + tensor var_7179_to_fp16 = const()[name = tensor("op_7179_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_207_cast_fp16 = layer_norm(axes = input_207_axes_0, beta = norm17_1_bias_to_fp16, epsilon = var_7179_to_fp16, gamma = norm17_1_weight_to_fp16, x = input_205_cast_fp16)[name = tensor("input_207_cast_fp16")]; + tensor attn17_in_proj_weight_to_fp16 = const()[name = tensor("attn17_in_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(429041472)))]; + tensor linear_68_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = attn17_in_proj_weight_to_fp16, x = input_207_cast_fp16)[name = tensor("linear_68_cast_fp16")]; + tensor var_7193 = const()[name = tensor("op_7193"), val = tensor([1, 256, 3, 16, 64])]; + tensor qkv_35_cast_fp16 = reshape(shape = var_7193, x = linear_68_cast_fp16)[name = tensor("qkv_35_cast_fp16")]; + tensor q_103_begin_0 = const()[name = tensor("q_103_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor q_103_end_0 = const()[name = tensor("q_103_end_0"), val = tensor([1, 256, 1, 16, 64])]; + tensor q_103_end_mask_0 = const()[name = tensor("q_103_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor q_103_squeeze_mask_0 = const()[name = tensor("q_103_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor q_103_cast_fp16 = slice_by_index(begin = q_103_begin_0, end = q_103_end_0, end_mask = q_103_end_mask_0, squeeze_mask = q_103_squeeze_mask_0, x = qkv_35_cast_fp16)[name = tensor("q_103_cast_fp16")]; + tensor k_69_begin_0 = const()[name = tensor("k_69_begin_0"), val = tensor([0, 0, 1, 0, 0])]; + tensor k_69_end_0 = const()[name = tensor("k_69_end_0"), val = tensor([1, 256, 2, 16, 64])]; + tensor k_69_end_mask_0 = const()[name = tensor("k_69_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor k_69_squeeze_mask_0 = const()[name = tensor("k_69_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor k_69_cast_fp16 = slice_by_index(begin = k_69_begin_0, end = k_69_end_0, end_mask = k_69_end_mask_0, squeeze_mask = k_69_squeeze_mask_0, x = qkv_35_cast_fp16)[name = tensor("k_69_cast_fp16")]; + tensor v_35_begin_0 = const()[name = tensor("v_35_begin_0"), val = tensor([0, 0, 2, 0, 0])]; + tensor v_35_end_0 = const()[name = tensor("v_35_end_0"), val = tensor([1, 256, 3, 16, 64])]; + tensor v_35_end_mask_0 = const()[name = tensor("v_35_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor v_35_squeeze_mask_0 = const()[name = tensor("v_35_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor v_35_cast_fp16 = slice_by_index(begin = v_35_begin_0, end = v_35_end_0, end_mask = v_35_end_mask_0, squeeze_mask = v_35_squeeze_mask_0, x = qkv_35_cast_fp16)[name = tensor("v_35_cast_fp16")]; + tensor var_7251 = const()[name = tensor("op_7251"), val = tensor([1, 1])]; + tensor position17_to_fp16_dtype_0 = const()[name = tensor("position17_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor position17_to_fp16 = cast(dtype = position17_to_fp16_dtype_0, x = position17)[name = tensor("cast_576")]; + tensor var_7252_cast_fp16 = reshape(shape = var_7251, x = position17_to_fp16)[name = tensor("op_7252_cast_fp16")]; + tensor ts_105_cast_fp16 = add(x = ts_1_promoted_to_fp16, y = var_7252_cast_fp16)[name = tensor("ts_105_cast_fp16")]; + tensor var_7259 = const()[name = tensor("op_7259"), val = tensor([1, 256, 1, 1])]; + tensor ts_107_cast_fp16 = reshape(shape = var_7259, x = ts_105_cast_fp16)[name = tensor("ts_107_cast_fp16")]; + tensor var_7266 = const()[name = tensor("op_7266"), val = tensor([1, 256, 16, 32, 2])]; + tensor q_complex_35_cast_fp16 = reshape(shape = var_7266, x = q_103_cast_fp16)[name = tensor("q_complex_35_cast_fp16")]; + tensor var_7273 = const()[name = tensor("op_7273"), val = tensor([1, 256, 16, 32, 2])]; + tensor k_complex_35_cast_fp16 = reshape(shape = var_7273, x = k_69_cast_fp16)[name = tensor("k_complex_35_cast_fp16")]; + tensor var_7277_begin_0 = const()[name = tensor("op_7277_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_7277_end_0 = const()[name = tensor("op_7277_end_0"), val = tensor([1, 256, 16, 32, 1])]; + tensor var_7277_end_mask_0 = const()[name = tensor("op_7277_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_7277_squeeze_mask_0 = const()[name = tensor("op_7277_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_7277_cast_fp16 = slice_by_index(begin = var_7277_begin_0, end = var_7277_end_0, end_mask = var_7277_end_mask_0, squeeze_mask = var_7277_squeeze_mask_0, x = q_complex_35_cast_fp16)[name = tensor("op_7277_cast_fp16")]; + tensor var_7285_begin_0 = const()[name = tensor("op_7285_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_7285_end_0 = const()[name = tensor("op_7285_end_0"), val = tensor([1, 256, 16, 32, 2])]; + tensor var_7285_end_mask_0 = const()[name = tensor("op_7285_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_7285_squeeze_mask_0 = const()[name = tensor("op_7285_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_7285_cast_fp16 = slice_by_index(begin = var_7285_begin_0, end = var_7285_end_0, end_mask = var_7285_end_mask_0, squeeze_mask = var_7285_squeeze_mask_0, x = q_complex_35_cast_fp16)[name = tensor("op_7285_cast_fp16")]; + tensor var_7293_begin_0 = const()[name = tensor("op_7293_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_7293_end_0 = const()[name = tensor("op_7293_end_0"), val = tensor([1, 256, 16, 32, 1])]; + tensor var_7293_end_mask_0 = const()[name = tensor("op_7293_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_7293_squeeze_mask_0 = const()[name = tensor("op_7293_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_7293_cast_fp16 = slice_by_index(begin = var_7293_begin_0, end = var_7293_end_0, end_mask = var_7293_end_mask_0, squeeze_mask = var_7293_squeeze_mask_0, x = k_complex_35_cast_fp16)[name = tensor("op_7293_cast_fp16")]; + tensor var_7301_begin_0 = const()[name = tensor("op_7301_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_7301_end_0 = const()[name = tensor("op_7301_end_0"), val = tensor([1, 256, 16, 32, 2])]; + tensor var_7301_end_mask_0 = const()[name = tensor("op_7301_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_7301_squeeze_mask_0 = const()[name = tensor("op_7301_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_7301_cast_fp16 = slice_by_index(begin = var_7301_begin_0, end = var_7301_end_0, end_mask = var_7301_end_mask_0, squeeze_mask = var_7301_squeeze_mask_0, x = k_complex_35_cast_fp16)[name = tensor("op_7301_cast_fp16")]; + tensor freqs_35_to_fp16 = const()[name = tensor("freqs_35_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(435332992)))]; + tensor var_7307_cast_fp16 = mul(x = freqs_35_to_fp16, y = ts_107_cast_fp16)[name = tensor("op_7307_cast_fp16")]; + tensor rotr_35_cast_fp16 = cos(x = var_7307_cast_fp16)[name = tensor("rotr_35_cast_fp16")]; + tensor roti_35_cast_fp16 = sin(x = var_7307_cast_fp16)[name = tensor("roti_35_cast_fp16")]; + tensor var_7311_cast_fp16 = mul(x = var_7277_cast_fp16, y = rotr_35_cast_fp16)[name = tensor("op_7311_cast_fp16")]; + tensor var_7312_cast_fp16 = mul(x = var_7285_cast_fp16, y = roti_35_cast_fp16)[name = tensor("op_7312_cast_fp16")]; + tensor qor_69_cast_fp16 = sub(x = var_7311_cast_fp16, y = var_7312_cast_fp16)[name = tensor("qor_69_cast_fp16")]; + tensor var_7315_cast_fp16 = mul(x = var_7277_cast_fp16, y = roti_35_cast_fp16)[name = tensor("op_7315_cast_fp16")]; + tensor var_7316_cast_fp16 = mul(x = var_7285_cast_fp16, y = rotr_35_cast_fp16)[name = tensor("op_7316_cast_fp16")]; + tensor qoi_69_cast_fp16 = add(x = var_7315_cast_fp16, y = var_7316_cast_fp16)[name = tensor("qoi_69_cast_fp16")]; + tensor var_7319_cast_fp16 = mul(x = var_7293_cast_fp16, y = rotr_35_cast_fp16)[name = tensor("op_7319_cast_fp16")]; + tensor var_7320_cast_fp16 = mul(x = var_7301_cast_fp16, y = roti_35_cast_fp16)[name = tensor("op_7320_cast_fp16")]; + tensor kor_69_cast_fp16 = sub(x = var_7319_cast_fp16, y = var_7320_cast_fp16)[name = tensor("kor_69_cast_fp16")]; + tensor var_7323_cast_fp16 = mul(x = var_7293_cast_fp16, y = roti_35_cast_fp16)[name = tensor("op_7323_cast_fp16")]; + tensor var_7324_cast_fp16 = mul(x = var_7301_cast_fp16, y = rotr_35_cast_fp16)[name = tensor("op_7324_cast_fp16")]; + tensor koi_69_cast_fp16 = add(x = var_7323_cast_fp16, y = var_7324_cast_fp16)[name = tensor("koi_69_cast_fp16")]; + tensor qo_35_axis_0 = const()[name = tensor("qo_35_axis_0"), val = tensor(-1)]; + tensor qo_35_cast_fp16 = stack(axis = qo_35_axis_0, values = (qor_69_cast_fp16, qoi_69_cast_fp16))[name = tensor("qo_35_cast_fp16")]; + tensor ko_35_axis_0 = const()[name = tensor("ko_35_axis_0"), val = tensor(-1)]; + tensor ko_35_cast_fp16 = stack(axis = ko_35_axis_0, values = (kor_69_cast_fp16, koi_69_cast_fp16))[name = tensor("ko_35_cast_fp16")]; + tensor var_7357 = const()[name = tensor("op_7357"), val = tensor([1, 256, 16, 64])]; + tensor q_105_cast_fp16 = reshape(shape = var_7357, x = qo_35_cast_fp16)[name = tensor("q_105_cast_fp16")]; + tensor var_7363 = const()[name = tensor("op_7363"), val = tensor([1, 256, 16, 64])]; + tensor k_71_cast_fp16 = reshape(shape = var_7363, x = ko_35_cast_fp16)[name = tensor("k_71_cast_fp16")]; + tensor write_indices_float_103_cast_fp16 = add(x = var_7252_cast_fp16, y = var_393_promoted_to_fp16)[name = tensor("write_indices_float_103_cast_fp16")]; + tensor _inversed_7384_y_0_to_fp16 = const()[name = tensor("_inversed_7384_y_0_to_fp16"), val = tensor(0x1p-9)]; + tensor _inversed_7384_cast_fp16 = mul(x = write_indices_float_103_cast_fp16, y = _inversed_7384_y_0_to_fp16)[name = tensor("_inversed_7384_cast_fp16")]; + tensor var_7385_cast_fp16 = floor(x = _inversed_7384_cast_fp16)[name = tensor("op_7385_cast_fp16")]; + tensor var_7386_to_fp16 = const()[name = tensor("op_7386_to_fp16"), val = tensor(0x1p+9)]; + tensor var_7387_cast_fp16 = mul(x = var_7385_cast_fp16, y = var_7386_to_fp16)[name = tensor("op_7387_cast_fp16")]; + tensor write_indices_float_105_cast_fp16 = sub(x = write_indices_float_103_cast_fp16, y = var_7387_cast_fp16)[name = tensor("write_indices_float_105_cast_fp16")]; + tensor var_7405_value_0_to_fp16 = const()[name = tensor("op_7405_value_0_to_fp16"), val = tensor(0x1.ffp+8)]; + tensor var_7405_cast_fp16 = fill_like(ref_tensor = write_indices_float_105_cast_fp16, value = var_7405_value_0_to_fp16)[name = tensor("op_7405_cast_fp16")]; + tensor write_indices_float_107_cast_fp16 = select(a = write_indices_float_105_cast_fp16, b = var_7405_cast_fp16, cond = token_valid_1_cast_fp16)[name = tensor("write_indices_float_107_cast_fp16")]; + tensor var_7411_dtype_0 = const()[name = tensor("op_7411_dtype_0"), val = tensor("int32")]; + tensor var_7416 = const()[name = tensor("op_7416"), val = tensor([1, 256, 1, 1])]; + tensor write_indices_float_107_cast_fp16_to_int32 = cast(dtype = var_7411_dtype_0, x = write_indices_float_107_cast_fp16)[name = tensor("cast_575")]; + tensor var_7417 = reshape(shape = var_7416, x = write_indices_float_107_cast_fp16_to_int32)[name = tensor("op_7417")]; + tensor write_indices_35_reps_0 = const()[name = tensor("write_indices_35_reps_0"), val = tensor([1, 1, 16, 64])]; + tensor write_indices_35 = tile(reps = write_indices_35_reps_0, x = var_7417)[name = tensor("write_indices_35")]; + tensor var_7427_begin_0 = const()[name = tensor("op_7427_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_7427_end_0 = const()[name = tensor("op_7427_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor var_7427_end_mask_0 = const()[name = tensor("op_7427_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_7427_squeeze_mask_0 = const()[name = tensor("op_7427_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor cache17_to_fp16_dtype_0 = const()[name = tensor("cache17_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor cache17_to_fp16 = cast(dtype = cache17_to_fp16_dtype_0, x = cache17)[name = tensor("cast_574")]; + tensor var_7427_cast_fp16 = slice_by_index(begin = var_7427_begin_0, end = var_7427_end_0, end_mask = var_7427_end_mask_0, squeeze_mask = var_7427_squeeze_mask_0, x = cache17_to_fp16)[name = tensor("op_7427_cast_fp16")]; + tensor var_7429_axis_0 = const()[name = tensor("op_7429_axis_0"), val = tensor(1)]; + tensor var_7429_mode_0 = const()[name = tensor("op_7429_mode_0"), val = tensor("update")]; + tensor var_7429_validate_indices_0 = const()[name = tensor("op_7429_validate_indices_0"), val = tensor(false)]; + tensor var_7429_cast_fp16 = scatter_along_axis(axis = var_7429_axis_0, data = var_7427_cast_fp16, indices = write_indices_35, mode = var_7429_mode_0, updates = k_71_cast_fp16, validate_indices = var_7429_validate_indices_0)[name = tensor("op_7429_cast_fp16")]; + tensor concat_138 = const()[name = tensor("concat_138"), val = tensor([0, 0, 0, 0, 0])]; + tensor concat_139 = const()[name = tensor("concat_139"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_35_internal_tensor_assign_1_stride_0 = const()[name = tensor("new_cache_35_internal_tensor_assign_1_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_35_internal_tensor_assign_1_begin_mask_0 = const()[name = tensor("new_cache_35_internal_tensor_assign_1_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_35_internal_tensor_assign_1_end_mask_0 = const()[name = tensor("new_cache_35_internal_tensor_assign_1_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_35_internal_tensor_assign_1_squeeze_mask_0 = const()[name = tensor("new_cache_35_internal_tensor_assign_1_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_80 = const()[name = tensor("shape_80"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_34 = const()[name = tensor("reduce_prod_34"), val = tensor(1048576)]; + tensor range_1d_34_start_0 = const()[name = tensor("range_1d_34_start_0"), val = tensor(0)]; + tensor range_1d_34_step_0 = const()[name = tensor("range_1d_34_step_0"), val = tensor(1)]; + tensor range_1d_34 = range_1d(end = reduce_prod_34, start = range_1d_34_start_0, step = range_1d_34_step_0)[name = tensor("range_1d_34")]; + tensor reshape_170 = reshape(shape = shape_80, x = range_1d_34)[name = tensor("reshape_170")]; + tensor slice_by_index_34 = slice_by_index(begin = concat_138, begin_mask = new_cache_35_internal_tensor_assign_1_begin_mask_0, end = concat_139, end_mask = new_cache_35_internal_tensor_assign_1_end_mask_0, squeeze_mask = new_cache_35_internal_tensor_assign_1_squeeze_mask_0, stride = new_cache_35_internal_tensor_assign_1_stride_0, x = reshape_170)[name = tensor("slice_by_index_34")]; + tensor reshape_171_shape_0 = const()[name = tensor("reshape_171_shape_0"), val = tensor([-1])]; + tensor reshape_171 = reshape(shape = reshape_171_shape_0, x = slice_by_index_34)[name = tensor("reshape_171")]; + tensor reshape_172_shape_0 = const()[name = tensor("reshape_172_shape_0"), val = tensor([-1])]; + tensor reshape_172_cast_fp16 = reshape(shape = reshape_172_shape_0, x = var_7429_cast_fp16)[name = tensor("reshape_172_cast_fp16")]; + tensor reshape_173_shape_0 = const()[name = tensor("reshape_173_shape_0"), val = tensor([-1])]; + tensor reshape_173_cast_fp16 = reshape(shape = reshape_173_shape_0, x = cache17_to_fp16)[name = tensor("reshape_173_cast_fp16")]; + tensor scatter_34_mode_0 = const()[name = tensor("scatter_34_mode_0"), val = tensor("update")]; + tensor scatter_34_axis_0 = const()[name = tensor("scatter_34_axis_0"), val = tensor(0)]; + tensor scatter_34_validate_indices_0 = const()[name = tensor("scatter_34_validate_indices_0"), val = tensor(false)]; + tensor scatter_34_cast_fp16 = scatter(axis = scatter_34_axis_0, data = reshape_173_cast_fp16, indices = reshape_171, mode = scatter_34_mode_0, updates = reshape_172_cast_fp16, validate_indices = scatter_34_validate_indices_0)[name = tensor("scatter_34_cast_fp16")]; + tensor reshape_174_cast_fp16 = reshape(shape = shape_80, x = scatter_34_cast_fp16)[name = tensor("reshape_174_cast_fp16")]; + tensor var_7437_begin_0 = const()[name = tensor("op_7437_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor var_7437_end_0 = const()[name = tensor("op_7437_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor var_7437_end_mask_0 = const()[name = tensor("op_7437_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_7437_squeeze_mask_0 = const()[name = tensor("op_7437_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor var_7437_cast_fp16 = slice_by_index(begin = var_7437_begin_0, end = var_7437_end_0, end_mask = var_7437_end_mask_0, squeeze_mask = var_7437_squeeze_mask_0, x = reshape_174_cast_fp16)[name = tensor("op_7437_cast_fp16")]; + tensor var_7439_axis_0 = const()[name = tensor("op_7439_axis_0"), val = tensor(1)]; + tensor var_7439_mode_0 = const()[name = tensor("op_7439_mode_0"), val = tensor("update")]; + tensor var_7439_validate_indices_0 = const()[name = tensor("op_7439_validate_indices_0"), val = tensor(false)]; + tensor var_7439_cast_fp16 = scatter_along_axis(axis = var_7439_axis_0, data = var_7437_cast_fp16, indices = write_indices_35, mode = var_7439_mode_0, updates = v_35_cast_fp16, validate_indices = var_7439_validate_indices_0)[name = tensor("op_7439_cast_fp16")]; + tensor concat_140 = const()[name = tensor("concat_140"), val = tensor([1, 0, 0, 0, 0])]; + tensor concat_141 = const()[name = tensor("concat_141"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_35_internal_tensor_assign_2_stride_0 = const()[name = tensor("new_cache_35_internal_tensor_assign_2_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_35_internal_tensor_assign_2_begin_mask_0 = const()[name = tensor("new_cache_35_internal_tensor_assign_2_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_35_internal_tensor_assign_2_end_mask_0 = const()[name = tensor("new_cache_35_internal_tensor_assign_2_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_35_internal_tensor_assign_2_squeeze_mask_0 = const()[name = tensor("new_cache_35_internal_tensor_assign_2_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_81 = const()[name = tensor("shape_81"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_35 = const()[name = tensor("reduce_prod_35"), val = tensor(1048576)]; + tensor range_1d_35_start_0 = const()[name = tensor("range_1d_35_start_0"), val = tensor(0)]; + tensor range_1d_35_step_0 = const()[name = tensor("range_1d_35_step_0"), val = tensor(1)]; + tensor range_1d_35 = range_1d(end = reduce_prod_35, start = range_1d_35_start_0, step = range_1d_35_step_0)[name = tensor("range_1d_35")]; + tensor reshape_175 = reshape(shape = shape_81, x = range_1d_35)[name = tensor("reshape_175")]; + tensor slice_by_index_35 = slice_by_index(begin = concat_140, begin_mask = new_cache_35_internal_tensor_assign_2_begin_mask_0, end = concat_141, end_mask = new_cache_35_internal_tensor_assign_2_end_mask_0, squeeze_mask = new_cache_35_internal_tensor_assign_2_squeeze_mask_0, stride = new_cache_35_internal_tensor_assign_2_stride_0, x = reshape_175)[name = tensor("slice_by_index_35")]; + tensor reshape_176_shape_0 = const()[name = tensor("reshape_176_shape_0"), val = tensor([-1])]; + tensor reshape_176 = reshape(shape = reshape_176_shape_0, x = slice_by_index_35)[name = tensor("reshape_176")]; + tensor reshape_177_shape_0 = const()[name = tensor("reshape_177_shape_0"), val = tensor([-1])]; + tensor reshape_177_cast_fp16 = reshape(shape = reshape_177_shape_0, x = var_7439_cast_fp16)[name = tensor("reshape_177_cast_fp16")]; + tensor reshape_178_shape_0 = const()[name = tensor("reshape_178_shape_0"), val = tensor([-1])]; + tensor reshape_178_cast_fp16 = reshape(shape = reshape_178_shape_0, x = reshape_174_cast_fp16)[name = tensor("reshape_178_cast_fp16")]; + tensor scatter_35_mode_0 = const()[name = tensor("scatter_35_mode_0"), val = tensor("update")]; + tensor scatter_35_axis_0 = const()[name = tensor("scatter_35_axis_0"), val = tensor(0)]; + tensor scatter_35_validate_indices_0 = const()[name = tensor("scatter_35_validate_indices_0"), val = tensor(false)]; + tensor scatter_35_cast_fp16 = scatter(axis = scatter_35_axis_0, data = reshape_178_cast_fp16, indices = reshape_176, mode = scatter_35_mode_0, updates = reshape_177_cast_fp16, validate_indices = scatter_35_validate_indices_0)[name = tensor("scatter_35_cast_fp16")]; + tensor reshape_179_cast_fp16 = reshape(shape = shape_81, x = scatter_35_cast_fp16)[name = tensor("reshape_179_cast_fp16")]; + tensor reshape_179_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("reshape_179_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor keys_103_begin_0 = const()[name = tensor("keys_103_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor keys_103_end_0 = const()[name = tensor("keys_103_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor keys_103_end_mask_0 = const()[name = tensor("keys_103_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor keys_103_squeeze_mask_0 = const()[name = tensor("keys_103_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor keys_103_cast_fp16 = slice_by_index(begin = keys_103_begin_0, end = keys_103_end_0, end_mask = keys_103_end_mask_0, squeeze_mask = keys_103_squeeze_mask_0, x = reshape_179_cast_fp16)[name = tensor("keys_103_cast_fp16")]; + tensor values_103_begin_0 = const()[name = tensor("values_103_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor values_103_end_0 = const()[name = tensor("values_103_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor values_103_end_mask_0 = const()[name = tensor("values_103_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor values_103_squeeze_mask_0 = const()[name = tensor("values_103_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor values_103_cast_fp16 = slice_by_index(begin = values_103_begin_0, end = values_103_end_0, end_mask = values_103_end_mask_0, squeeze_mask = values_103_squeeze_mask_0, x = reshape_179_cast_fp16)[name = tensor("values_103_cast_fp16")]; + tensor var_7451_cast_fp16 = not_equal(x = keys_103_cast_fp16, y = keys_103_cast_fp16)[name = tensor("op_7451_cast_fp16")]; + tensor keys_105_cast_fp16 = select(a = var_470_to_fp16, b = keys_103_cast_fp16, cond = var_7451_cast_fp16)[name = tensor("keys_105_cast_fp16")]; + tensor var_7459_cast_fp16 = not_equal(x = values_103_cast_fp16, y = values_103_cast_fp16)[name = tensor("op_7459_cast_fp16")]; + tensor values_105_cast_fp16 = select(a = var_470_to_fp16, b = values_103_cast_fp16, cond = var_7459_cast_fp16)[name = tensor("values_105_cast_fp16")]; + tensor var_7483 = const()[name = tensor("op_7483"), val = tensor([0, 2, 1, 3])]; + tensor var_7499 = const()[name = tensor("op_7499"), val = tensor([1, 1, 1])]; + tensor var_7500_cast_fp16 = reshape(shape = var_7499, x = position17_to_fp16)[name = tensor("op_7500_cast_fp16")]; + tensor q_positions_35_cast_fp16 = add(x = var_7500_cast_fp16, y = q_offsets_1_promoted_to_fp16)[name = tensor("q_positions_35_cast_fp16")]; + tensor valid_end_35_cast_fp16 = add(x = var_7500_cast_fp16, y = var_536_cast_fp16)[name = tensor("valid_end_35_cast_fp16")]; + tensor valid_mask_35_cast_fp16 = less(x = k_positions_1_promoted_to_fp16, y = valid_end_35_cast_fp16)[name = tensor("valid_mask_35_cast_fp16")]; + tensor causal_mask_35_cast_fp16 = less_equal(x = k_positions_1_promoted_to_fp16, y = q_positions_35_cast_fp16)[name = tensor("causal_mask_35_cast_fp16")]; + tensor var_7528 = logical_and(x = valid_mask_35_cast_fp16, y = causal_mask_35_cast_fp16)[name = tensor("op_7528")]; + tensor attn_mask_35_axes_0 = const()[name = tensor("attn_mask_35_axes_0"), val = tensor([1])]; + tensor attn_mask_35 = expand_dims(axes = attn_mask_35_axes_0, x = var_7528)[name = tensor("attn_mask_35")]; + tensor var_7542_transpose_x_0 = const()[name = tensor("op_7542_transpose_x_0"), val = tensor(false)]; + tensor var_7542_transpose_y_0 = const()[name = tensor("op_7542_transpose_y_0"), val = tensor(false)]; + tensor transpose_103_perm_0 = const()[name = tensor("transpose_103_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_104_perm_0 = const()[name = tensor("transpose_104_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_104 = transpose(perm = transpose_104_perm_0, x = keys_105_cast_fp16)[name = tensor("transpose_136")]; + tensor transpose_103 = transpose(perm = transpose_103_perm_0, x = q_105_cast_fp16)[name = tensor("transpose_137")]; + tensor var_7542_cast_fp16 = matmul(transpose_x = var_7542_transpose_x_0, transpose_y = var_7542_transpose_y_0, x = transpose_103, y = transpose_104)[name = tensor("op_7542_cast_fp16")]; + tensor var_7536_to_fp16 = const()[name = tensor("op_7536_to_fp16"), val = tensor([0x1p-3])]; + tensor attn_weights_103_cast_fp16 = mul(x = var_7542_cast_fp16, y = var_7536_to_fp16)[name = tensor("attn_weights_103_cast_fp16")]; + tensor var_7544 = logical_not(x = attn_mask_35)[name = tensor("op_7544")]; + tensor var_7545_to_fp16 = const()[name = tensor("op_7545_to_fp16"), val = tensor(-inf)]; + tensor attn_weights_105_cast_fp16 = select(a = var_7545_to_fp16, b = attn_weights_103_cast_fp16, cond = var_7544)[name = tensor("attn_weights_105_cast_fp16")]; + tensor var_7547 = const()[name = tensor("op_7547"), val = tensor(-1)]; + tensor attn_weights_107_cast_fp16 = softmax(axis = var_7547, x = attn_weights_105_cast_fp16)[name = tensor("attn_weights_107_cast_fp16")]; + tensor attn_output_35_transpose_x_0 = const()[name = tensor("attn_output_35_transpose_x_0"), val = tensor(false)]; + tensor attn_output_35_transpose_y_0 = const()[name = tensor("attn_output_35_transpose_y_0"), val = tensor(false)]; + tensor values_107_cast_fp16 = transpose(perm = var_7483, x = values_105_cast_fp16)[name = tensor("transpose_138")]; + tensor attn_output_35_cast_fp16 = matmul(transpose_x = attn_output_35_transpose_x_0, transpose_y = attn_output_35_transpose_y_0, x = attn_weights_107_cast_fp16, y = values_107_cast_fp16)[name = tensor("attn_output_35_cast_fp16")]; + tensor var_7555 = const()[name = tensor("op_7555"), val = tensor([0, 2, 1, 3])]; + tensor var_7560 = const()[name = tensor("op_7560"), val = tensor([1, 256, 1024])]; + tensor var_7556_cast_fp16 = transpose(perm = var_7555, x = attn_output_35_cast_fp16)[name = tensor("transpose_135")]; + tensor input_209_cast_fp16 = reshape(shape = var_7560, x = var_7556_cast_fp16)[name = tensor("input_209_cast_fp16")]; + tensor attn17_out_proj_weight_to_fp16 = const()[name = tensor("attn17_out_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(435333120)))]; + tensor linear_69_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = attn17_out_proj_weight_to_fp16, x = input_209_cast_fp16)[name = tensor("linear_69_cast_fp16")]; + tensor var_7569_cast_fp16 = add(x = position17_to_fp16, y = valid_len_to_fp16)[name = tensor("op_7569_cast_fp16")]; + tensor var_7569_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_7569_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor input_211_cast_fp16 = add(x = input_205_cast_fp16, y = linear_69_cast_fp16)[name = tensor("input_211_cast_fp16")]; + tensor input_213_axes_0 = const()[name = tensor("input_213_axes_0"), val = tensor([-1])]; + tensor norm17_2_weight_to_fp16 = const()[name = tensor("norm17_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(437430336)))]; + tensor norm17_2_bias_to_fp16 = const()[name = tensor("norm17_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(437432448)))]; + tensor var_7573_to_fp16 = const()[name = tensor("op_7573_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_213_cast_fp16 = layer_norm(axes = input_213_axes_0, beta = norm17_2_bias_to_fp16, epsilon = var_7573_to_fp16, gamma = norm17_2_weight_to_fp16, x = input_211_cast_fp16)[name = tensor("input_213_cast_fp16")]; + tensor linear17_1_weight_to_fp16 = const()[name = tensor("linear17_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(437434560)))]; + tensor linear_70_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = linear17_1_weight_to_fp16, x = input_213_cast_fp16)[name = tensor("linear_70_cast_fp16")]; + tensor input_215_mode_0 = const()[name = tensor("input_215_mode_0"), val = tensor("EXACT")]; + tensor input_215_cast_fp16 = gelu(mode = input_215_mode_0, x = linear_70_cast_fp16)[name = tensor("input_215_cast_fp16")]; + tensor linear17_2_weight_to_fp16 = const()[name = tensor("linear17_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(445823232)))]; + tensor linear_71_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = linear17_2_weight_to_fp16, x = input_215_cast_fp16)[name = tensor("linear_71_cast_fp16")]; + tensor input_217_cast_fp16 = add(x = input_211_cast_fp16, y = linear_71_cast_fp16)[name = tensor("input_217_cast_fp16")]; + tensor input_219_axes_0 = const()[name = tensor("input_219_axes_0"), val = tensor([-1])]; + tensor norm18_1_weight_to_fp16 = const()[name = tensor("norm18_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(454211904)))]; + tensor norm18_1_bias_to_fp16 = const()[name = tensor("norm18_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(454214016)))]; + tensor var_7590_to_fp16 = const()[name = tensor("op_7590_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_219_cast_fp16 = layer_norm(axes = input_219_axes_0, beta = norm18_1_bias_to_fp16, epsilon = var_7590_to_fp16, gamma = norm18_1_weight_to_fp16, x = input_217_cast_fp16)[name = tensor("input_219_cast_fp16")]; + tensor attn18_in_proj_weight_to_fp16 = const()[name = tensor("attn18_in_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(454216128)))]; + tensor linear_72_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = attn18_in_proj_weight_to_fp16, x = input_219_cast_fp16)[name = tensor("linear_72_cast_fp16")]; + tensor var_7604 = const()[name = tensor("op_7604"), val = tensor([1, 256, 3, 16, 64])]; + tensor qkv_37_cast_fp16 = reshape(shape = var_7604, x = linear_72_cast_fp16)[name = tensor("qkv_37_cast_fp16")]; + tensor q_109_begin_0 = const()[name = tensor("q_109_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor q_109_end_0 = const()[name = tensor("q_109_end_0"), val = tensor([1, 256, 1, 16, 64])]; + tensor q_109_end_mask_0 = const()[name = tensor("q_109_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor q_109_squeeze_mask_0 = const()[name = tensor("q_109_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor q_109_cast_fp16 = slice_by_index(begin = q_109_begin_0, end = q_109_end_0, end_mask = q_109_end_mask_0, squeeze_mask = q_109_squeeze_mask_0, x = qkv_37_cast_fp16)[name = tensor("q_109_cast_fp16")]; + tensor k_73_begin_0 = const()[name = tensor("k_73_begin_0"), val = tensor([0, 0, 1, 0, 0])]; + tensor k_73_end_0 = const()[name = tensor("k_73_end_0"), val = tensor([1, 256, 2, 16, 64])]; + tensor k_73_end_mask_0 = const()[name = tensor("k_73_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor k_73_squeeze_mask_0 = const()[name = tensor("k_73_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor k_73_cast_fp16 = slice_by_index(begin = k_73_begin_0, end = k_73_end_0, end_mask = k_73_end_mask_0, squeeze_mask = k_73_squeeze_mask_0, x = qkv_37_cast_fp16)[name = tensor("k_73_cast_fp16")]; + tensor v_37_begin_0 = const()[name = tensor("v_37_begin_0"), val = tensor([0, 0, 2, 0, 0])]; + tensor v_37_end_0 = const()[name = tensor("v_37_end_0"), val = tensor([1, 256, 3, 16, 64])]; + tensor v_37_end_mask_0 = const()[name = tensor("v_37_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor v_37_squeeze_mask_0 = const()[name = tensor("v_37_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor v_37_cast_fp16 = slice_by_index(begin = v_37_begin_0, end = v_37_end_0, end_mask = v_37_end_mask_0, squeeze_mask = v_37_squeeze_mask_0, x = qkv_37_cast_fp16)[name = tensor("v_37_cast_fp16")]; + tensor var_7662 = const()[name = tensor("op_7662"), val = tensor([1, 1])]; + tensor position18_to_fp16_dtype_0 = const()[name = tensor("position18_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor position18_to_fp16 = cast(dtype = position18_to_fp16_dtype_0, x = position18)[name = tensor("cast_571")]; + tensor var_7663_cast_fp16 = reshape(shape = var_7662, x = position18_to_fp16)[name = tensor("op_7663_cast_fp16")]; + tensor ts_111_cast_fp16 = add(x = ts_1_promoted_to_fp16, y = var_7663_cast_fp16)[name = tensor("ts_111_cast_fp16")]; + tensor var_7670 = const()[name = tensor("op_7670"), val = tensor([1, 256, 1, 1])]; + tensor ts_113_cast_fp16 = reshape(shape = var_7670, x = ts_111_cast_fp16)[name = tensor("ts_113_cast_fp16")]; + tensor var_7677 = const()[name = tensor("op_7677"), val = tensor([1, 256, 16, 32, 2])]; + tensor q_complex_37_cast_fp16 = reshape(shape = var_7677, x = q_109_cast_fp16)[name = tensor("q_complex_37_cast_fp16")]; + tensor var_7684 = const()[name = tensor("op_7684"), val = tensor([1, 256, 16, 32, 2])]; + tensor k_complex_37_cast_fp16 = reshape(shape = var_7684, x = k_73_cast_fp16)[name = tensor("k_complex_37_cast_fp16")]; + tensor var_7688_begin_0 = const()[name = tensor("op_7688_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_7688_end_0 = const()[name = tensor("op_7688_end_0"), val = tensor([1, 256, 16, 32, 1])]; + tensor var_7688_end_mask_0 = const()[name = tensor("op_7688_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_7688_squeeze_mask_0 = const()[name = tensor("op_7688_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_7688_cast_fp16 = slice_by_index(begin = var_7688_begin_0, end = var_7688_end_0, end_mask = var_7688_end_mask_0, squeeze_mask = var_7688_squeeze_mask_0, x = q_complex_37_cast_fp16)[name = tensor("op_7688_cast_fp16")]; + tensor var_7696_begin_0 = const()[name = tensor("op_7696_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_7696_end_0 = const()[name = tensor("op_7696_end_0"), val = tensor([1, 256, 16, 32, 2])]; + tensor var_7696_end_mask_0 = const()[name = tensor("op_7696_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_7696_squeeze_mask_0 = const()[name = tensor("op_7696_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_7696_cast_fp16 = slice_by_index(begin = var_7696_begin_0, end = var_7696_end_0, end_mask = var_7696_end_mask_0, squeeze_mask = var_7696_squeeze_mask_0, x = q_complex_37_cast_fp16)[name = tensor("op_7696_cast_fp16")]; + tensor var_7704_begin_0 = const()[name = tensor("op_7704_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_7704_end_0 = const()[name = tensor("op_7704_end_0"), val = tensor([1, 256, 16, 32, 1])]; + tensor var_7704_end_mask_0 = const()[name = tensor("op_7704_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_7704_squeeze_mask_0 = const()[name = tensor("op_7704_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_7704_cast_fp16 = slice_by_index(begin = var_7704_begin_0, end = var_7704_end_0, end_mask = var_7704_end_mask_0, squeeze_mask = var_7704_squeeze_mask_0, x = k_complex_37_cast_fp16)[name = tensor("op_7704_cast_fp16")]; + tensor var_7712_begin_0 = const()[name = tensor("op_7712_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_7712_end_0 = const()[name = tensor("op_7712_end_0"), val = tensor([1, 256, 16, 32, 2])]; + tensor var_7712_end_mask_0 = const()[name = tensor("op_7712_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_7712_squeeze_mask_0 = const()[name = tensor("op_7712_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_7712_cast_fp16 = slice_by_index(begin = var_7712_begin_0, end = var_7712_end_0, end_mask = var_7712_end_mask_0, squeeze_mask = var_7712_squeeze_mask_0, x = k_complex_37_cast_fp16)[name = tensor("op_7712_cast_fp16")]; + tensor freqs_37_to_fp16 = const()[name = tensor("freqs_37_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(460507648)))]; + tensor var_7718_cast_fp16 = mul(x = freqs_37_to_fp16, y = ts_113_cast_fp16)[name = tensor("op_7718_cast_fp16")]; + tensor rotr_37_cast_fp16 = cos(x = var_7718_cast_fp16)[name = tensor("rotr_37_cast_fp16")]; + tensor roti_37_cast_fp16 = sin(x = var_7718_cast_fp16)[name = tensor("roti_37_cast_fp16")]; + tensor var_7722_cast_fp16 = mul(x = var_7688_cast_fp16, y = rotr_37_cast_fp16)[name = tensor("op_7722_cast_fp16")]; + tensor var_7723_cast_fp16 = mul(x = var_7696_cast_fp16, y = roti_37_cast_fp16)[name = tensor("op_7723_cast_fp16")]; + tensor qor_73_cast_fp16 = sub(x = var_7722_cast_fp16, y = var_7723_cast_fp16)[name = tensor("qor_73_cast_fp16")]; + tensor var_7726_cast_fp16 = mul(x = var_7688_cast_fp16, y = roti_37_cast_fp16)[name = tensor("op_7726_cast_fp16")]; + tensor var_7727_cast_fp16 = mul(x = var_7696_cast_fp16, y = rotr_37_cast_fp16)[name = tensor("op_7727_cast_fp16")]; + tensor qoi_73_cast_fp16 = add(x = var_7726_cast_fp16, y = var_7727_cast_fp16)[name = tensor("qoi_73_cast_fp16")]; + tensor var_7730_cast_fp16 = mul(x = var_7704_cast_fp16, y = rotr_37_cast_fp16)[name = tensor("op_7730_cast_fp16")]; + tensor var_7731_cast_fp16 = mul(x = var_7712_cast_fp16, y = roti_37_cast_fp16)[name = tensor("op_7731_cast_fp16")]; + tensor kor_73_cast_fp16 = sub(x = var_7730_cast_fp16, y = var_7731_cast_fp16)[name = tensor("kor_73_cast_fp16")]; + tensor var_7734_cast_fp16 = mul(x = var_7704_cast_fp16, y = roti_37_cast_fp16)[name = tensor("op_7734_cast_fp16")]; + tensor var_7735_cast_fp16 = mul(x = var_7712_cast_fp16, y = rotr_37_cast_fp16)[name = tensor("op_7735_cast_fp16")]; + tensor koi_73_cast_fp16 = add(x = var_7734_cast_fp16, y = var_7735_cast_fp16)[name = tensor("koi_73_cast_fp16")]; + tensor qo_37_axis_0 = const()[name = tensor("qo_37_axis_0"), val = tensor(-1)]; + tensor qo_37_cast_fp16 = stack(axis = qo_37_axis_0, values = (qor_73_cast_fp16, qoi_73_cast_fp16))[name = tensor("qo_37_cast_fp16")]; + tensor ko_37_axis_0 = const()[name = tensor("ko_37_axis_0"), val = tensor(-1)]; + tensor ko_37_cast_fp16 = stack(axis = ko_37_axis_0, values = (kor_73_cast_fp16, koi_73_cast_fp16))[name = tensor("ko_37_cast_fp16")]; + tensor var_7768 = const()[name = tensor("op_7768"), val = tensor([1, 256, 16, 64])]; + tensor q_111_cast_fp16 = reshape(shape = var_7768, x = qo_37_cast_fp16)[name = tensor("q_111_cast_fp16")]; + tensor var_7774 = const()[name = tensor("op_7774"), val = tensor([1, 256, 16, 64])]; + tensor k_75_cast_fp16 = reshape(shape = var_7774, x = ko_37_cast_fp16)[name = tensor("k_75_cast_fp16")]; + tensor write_indices_float_109_cast_fp16 = add(x = var_7663_cast_fp16, y = var_393_promoted_to_fp16)[name = tensor("write_indices_float_109_cast_fp16")]; + tensor _inversed_7795_y_0_to_fp16 = const()[name = tensor("_inversed_7795_y_0_to_fp16"), val = tensor(0x1p-9)]; + tensor _inversed_7795_cast_fp16 = mul(x = write_indices_float_109_cast_fp16, y = _inversed_7795_y_0_to_fp16)[name = tensor("_inversed_7795_cast_fp16")]; + tensor var_7796_cast_fp16 = floor(x = _inversed_7795_cast_fp16)[name = tensor("op_7796_cast_fp16")]; + tensor var_7797_to_fp16 = const()[name = tensor("op_7797_to_fp16"), val = tensor(0x1p+9)]; + tensor var_7798_cast_fp16 = mul(x = var_7796_cast_fp16, y = var_7797_to_fp16)[name = tensor("op_7798_cast_fp16")]; + tensor write_indices_float_111_cast_fp16 = sub(x = write_indices_float_109_cast_fp16, y = var_7798_cast_fp16)[name = tensor("write_indices_float_111_cast_fp16")]; + tensor var_7816_value_0_to_fp16 = const()[name = tensor("op_7816_value_0_to_fp16"), val = tensor(0x1.ffp+8)]; + tensor var_7816_cast_fp16 = fill_like(ref_tensor = write_indices_float_111_cast_fp16, value = var_7816_value_0_to_fp16)[name = tensor("op_7816_cast_fp16")]; + tensor write_indices_float_113_cast_fp16 = select(a = write_indices_float_111_cast_fp16, b = var_7816_cast_fp16, cond = token_valid_1_cast_fp16)[name = tensor("write_indices_float_113_cast_fp16")]; + tensor var_7822_dtype_0 = const()[name = tensor("op_7822_dtype_0"), val = tensor("int32")]; + tensor var_7827 = const()[name = tensor("op_7827"), val = tensor([1, 256, 1, 1])]; + tensor write_indices_float_113_cast_fp16_to_int32 = cast(dtype = var_7822_dtype_0, x = write_indices_float_113_cast_fp16)[name = tensor("cast_570")]; + tensor var_7828 = reshape(shape = var_7827, x = write_indices_float_113_cast_fp16_to_int32)[name = tensor("op_7828")]; + tensor write_indices_37_reps_0 = const()[name = tensor("write_indices_37_reps_0"), val = tensor([1, 1, 16, 64])]; + tensor write_indices_37 = tile(reps = write_indices_37_reps_0, x = var_7828)[name = tensor("write_indices_37")]; + tensor var_7838_begin_0 = const()[name = tensor("op_7838_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_7838_end_0 = const()[name = tensor("op_7838_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor var_7838_end_mask_0 = const()[name = tensor("op_7838_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_7838_squeeze_mask_0 = const()[name = tensor("op_7838_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor cache18_to_fp16_dtype_0 = const()[name = tensor("cache18_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor cache18_to_fp16 = cast(dtype = cache18_to_fp16_dtype_0, x = cache18)[name = tensor("cast_569")]; + tensor var_7838_cast_fp16 = slice_by_index(begin = var_7838_begin_0, end = var_7838_end_0, end_mask = var_7838_end_mask_0, squeeze_mask = var_7838_squeeze_mask_0, x = cache18_to_fp16)[name = tensor("op_7838_cast_fp16")]; + tensor var_7840_axis_0 = const()[name = tensor("op_7840_axis_0"), val = tensor(1)]; + tensor var_7840_mode_0 = const()[name = tensor("op_7840_mode_0"), val = tensor("update")]; + tensor var_7840_validate_indices_0 = const()[name = tensor("op_7840_validate_indices_0"), val = tensor(false)]; + tensor var_7840_cast_fp16 = scatter_along_axis(axis = var_7840_axis_0, data = var_7838_cast_fp16, indices = write_indices_37, mode = var_7840_mode_0, updates = k_75_cast_fp16, validate_indices = var_7840_validate_indices_0)[name = tensor("op_7840_cast_fp16")]; + tensor concat_146 = const()[name = tensor("concat_146"), val = tensor([0, 0, 0, 0, 0])]; + tensor concat_147 = const()[name = tensor("concat_147"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_37_internal_tensor_assign_1_stride_0 = const()[name = tensor("new_cache_37_internal_tensor_assign_1_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_37_internal_tensor_assign_1_begin_mask_0 = const()[name = tensor("new_cache_37_internal_tensor_assign_1_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_37_internal_tensor_assign_1_end_mask_0 = const()[name = tensor("new_cache_37_internal_tensor_assign_1_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_37_internal_tensor_assign_1_squeeze_mask_0 = const()[name = tensor("new_cache_37_internal_tensor_assign_1_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_82 = const()[name = tensor("shape_82"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_36 = const()[name = tensor("reduce_prod_36"), val = tensor(1048576)]; + tensor range_1d_36_start_0 = const()[name = tensor("range_1d_36_start_0"), val = tensor(0)]; + tensor range_1d_36_step_0 = const()[name = tensor("range_1d_36_step_0"), val = tensor(1)]; + tensor range_1d_36 = range_1d(end = reduce_prod_36, start = range_1d_36_start_0, step = range_1d_36_step_0)[name = tensor("range_1d_36")]; + tensor reshape_180 = reshape(shape = shape_82, x = range_1d_36)[name = tensor("reshape_180")]; + tensor slice_by_index_36 = slice_by_index(begin = concat_146, begin_mask = new_cache_37_internal_tensor_assign_1_begin_mask_0, end = concat_147, end_mask = new_cache_37_internal_tensor_assign_1_end_mask_0, squeeze_mask = new_cache_37_internal_tensor_assign_1_squeeze_mask_0, stride = new_cache_37_internal_tensor_assign_1_stride_0, x = reshape_180)[name = tensor("slice_by_index_36")]; + tensor reshape_181_shape_0 = const()[name = tensor("reshape_181_shape_0"), val = tensor([-1])]; + tensor reshape_181 = reshape(shape = reshape_181_shape_0, x = slice_by_index_36)[name = tensor("reshape_181")]; + tensor reshape_182_shape_0 = const()[name = tensor("reshape_182_shape_0"), val = tensor([-1])]; + tensor reshape_182_cast_fp16 = reshape(shape = reshape_182_shape_0, x = var_7840_cast_fp16)[name = tensor("reshape_182_cast_fp16")]; + tensor reshape_183_shape_0 = const()[name = tensor("reshape_183_shape_0"), val = tensor([-1])]; + tensor reshape_183_cast_fp16 = reshape(shape = reshape_183_shape_0, x = cache18_to_fp16)[name = tensor("reshape_183_cast_fp16")]; + tensor scatter_36_mode_0 = const()[name = tensor("scatter_36_mode_0"), val = tensor("update")]; + tensor scatter_36_axis_0 = const()[name = tensor("scatter_36_axis_0"), val = tensor(0)]; + tensor scatter_36_validate_indices_0 = const()[name = tensor("scatter_36_validate_indices_0"), val = tensor(false)]; + tensor scatter_36_cast_fp16 = scatter(axis = scatter_36_axis_0, data = reshape_183_cast_fp16, indices = reshape_181, mode = scatter_36_mode_0, updates = reshape_182_cast_fp16, validate_indices = scatter_36_validate_indices_0)[name = tensor("scatter_36_cast_fp16")]; + tensor reshape_184_cast_fp16 = reshape(shape = shape_82, x = scatter_36_cast_fp16)[name = tensor("reshape_184_cast_fp16")]; + tensor var_7848_begin_0 = const()[name = tensor("op_7848_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor var_7848_end_0 = const()[name = tensor("op_7848_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor var_7848_end_mask_0 = const()[name = tensor("op_7848_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_7848_squeeze_mask_0 = const()[name = tensor("op_7848_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor var_7848_cast_fp16 = slice_by_index(begin = var_7848_begin_0, end = var_7848_end_0, end_mask = var_7848_end_mask_0, squeeze_mask = var_7848_squeeze_mask_0, x = reshape_184_cast_fp16)[name = tensor("op_7848_cast_fp16")]; + tensor var_7850_axis_0 = const()[name = tensor("op_7850_axis_0"), val = tensor(1)]; + tensor var_7850_mode_0 = const()[name = tensor("op_7850_mode_0"), val = tensor("update")]; + tensor var_7850_validate_indices_0 = const()[name = tensor("op_7850_validate_indices_0"), val = tensor(false)]; + tensor var_7850_cast_fp16 = scatter_along_axis(axis = var_7850_axis_0, data = var_7848_cast_fp16, indices = write_indices_37, mode = var_7850_mode_0, updates = v_37_cast_fp16, validate_indices = var_7850_validate_indices_0)[name = tensor("op_7850_cast_fp16")]; + tensor concat_148 = const()[name = tensor("concat_148"), val = tensor([1, 0, 0, 0, 0])]; + tensor concat_149 = const()[name = tensor("concat_149"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_37_internal_tensor_assign_2_stride_0 = const()[name = tensor("new_cache_37_internal_tensor_assign_2_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_37_internal_tensor_assign_2_begin_mask_0 = const()[name = tensor("new_cache_37_internal_tensor_assign_2_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_37_internal_tensor_assign_2_end_mask_0 = const()[name = tensor("new_cache_37_internal_tensor_assign_2_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_37_internal_tensor_assign_2_squeeze_mask_0 = const()[name = tensor("new_cache_37_internal_tensor_assign_2_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_83 = const()[name = tensor("shape_83"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_37 = const()[name = tensor("reduce_prod_37"), val = tensor(1048576)]; + tensor range_1d_37_start_0 = const()[name = tensor("range_1d_37_start_0"), val = tensor(0)]; + tensor range_1d_37_step_0 = const()[name = tensor("range_1d_37_step_0"), val = tensor(1)]; + tensor range_1d_37 = range_1d(end = reduce_prod_37, start = range_1d_37_start_0, step = range_1d_37_step_0)[name = tensor("range_1d_37")]; + tensor reshape_185 = reshape(shape = shape_83, x = range_1d_37)[name = tensor("reshape_185")]; + tensor slice_by_index_37 = slice_by_index(begin = concat_148, begin_mask = new_cache_37_internal_tensor_assign_2_begin_mask_0, end = concat_149, end_mask = new_cache_37_internal_tensor_assign_2_end_mask_0, squeeze_mask = new_cache_37_internal_tensor_assign_2_squeeze_mask_0, stride = new_cache_37_internal_tensor_assign_2_stride_0, x = reshape_185)[name = tensor("slice_by_index_37")]; + tensor reshape_186_shape_0 = const()[name = tensor("reshape_186_shape_0"), val = tensor([-1])]; + tensor reshape_186 = reshape(shape = reshape_186_shape_0, x = slice_by_index_37)[name = tensor("reshape_186")]; + tensor reshape_187_shape_0 = const()[name = tensor("reshape_187_shape_0"), val = tensor([-1])]; + tensor reshape_187_cast_fp16 = reshape(shape = reshape_187_shape_0, x = var_7850_cast_fp16)[name = tensor("reshape_187_cast_fp16")]; + tensor reshape_188_shape_0 = const()[name = tensor("reshape_188_shape_0"), val = tensor([-1])]; + tensor reshape_188_cast_fp16 = reshape(shape = reshape_188_shape_0, x = reshape_184_cast_fp16)[name = tensor("reshape_188_cast_fp16")]; + tensor scatter_37_mode_0 = const()[name = tensor("scatter_37_mode_0"), val = tensor("update")]; + tensor scatter_37_axis_0 = const()[name = tensor("scatter_37_axis_0"), val = tensor(0)]; + tensor scatter_37_validate_indices_0 = const()[name = tensor("scatter_37_validate_indices_0"), val = tensor(false)]; + tensor scatter_37_cast_fp16 = scatter(axis = scatter_37_axis_0, data = reshape_188_cast_fp16, indices = reshape_186, mode = scatter_37_mode_0, updates = reshape_187_cast_fp16, validate_indices = scatter_37_validate_indices_0)[name = tensor("scatter_37_cast_fp16")]; + tensor reshape_189_cast_fp16 = reshape(shape = shape_83, x = scatter_37_cast_fp16)[name = tensor("reshape_189_cast_fp16")]; + tensor reshape_189_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("reshape_189_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor keys_109_begin_0 = const()[name = tensor("keys_109_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor keys_109_end_0 = const()[name = tensor("keys_109_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor keys_109_end_mask_0 = const()[name = tensor("keys_109_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor keys_109_squeeze_mask_0 = const()[name = tensor("keys_109_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor keys_109_cast_fp16 = slice_by_index(begin = keys_109_begin_0, end = keys_109_end_0, end_mask = keys_109_end_mask_0, squeeze_mask = keys_109_squeeze_mask_0, x = reshape_189_cast_fp16)[name = tensor("keys_109_cast_fp16")]; + tensor values_109_begin_0 = const()[name = tensor("values_109_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor values_109_end_0 = const()[name = tensor("values_109_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor values_109_end_mask_0 = const()[name = tensor("values_109_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor values_109_squeeze_mask_0 = const()[name = tensor("values_109_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor values_109_cast_fp16 = slice_by_index(begin = values_109_begin_0, end = values_109_end_0, end_mask = values_109_end_mask_0, squeeze_mask = values_109_squeeze_mask_0, x = reshape_189_cast_fp16)[name = tensor("values_109_cast_fp16")]; + tensor var_7862_cast_fp16 = not_equal(x = keys_109_cast_fp16, y = keys_109_cast_fp16)[name = tensor("op_7862_cast_fp16")]; + tensor keys_111_cast_fp16 = select(a = var_470_to_fp16, b = keys_109_cast_fp16, cond = var_7862_cast_fp16)[name = tensor("keys_111_cast_fp16")]; + tensor var_7870_cast_fp16 = not_equal(x = values_109_cast_fp16, y = values_109_cast_fp16)[name = tensor("op_7870_cast_fp16")]; + tensor values_111_cast_fp16 = select(a = var_470_to_fp16, b = values_109_cast_fp16, cond = var_7870_cast_fp16)[name = tensor("values_111_cast_fp16")]; + tensor var_7894 = const()[name = tensor("op_7894"), val = tensor([0, 2, 1, 3])]; + tensor var_7910 = const()[name = tensor("op_7910"), val = tensor([1, 1, 1])]; + tensor var_7911_cast_fp16 = reshape(shape = var_7910, x = position18_to_fp16)[name = tensor("op_7911_cast_fp16")]; + tensor q_positions_37_cast_fp16 = add(x = var_7911_cast_fp16, y = q_offsets_1_promoted_to_fp16)[name = tensor("q_positions_37_cast_fp16")]; + tensor valid_end_37_cast_fp16 = add(x = var_7911_cast_fp16, y = var_536_cast_fp16)[name = tensor("valid_end_37_cast_fp16")]; + tensor valid_mask_37_cast_fp16 = less(x = k_positions_1_promoted_to_fp16, y = valid_end_37_cast_fp16)[name = tensor("valid_mask_37_cast_fp16")]; + tensor causal_mask_37_cast_fp16 = less_equal(x = k_positions_1_promoted_to_fp16, y = q_positions_37_cast_fp16)[name = tensor("causal_mask_37_cast_fp16")]; + tensor var_7939 = logical_and(x = valid_mask_37_cast_fp16, y = causal_mask_37_cast_fp16)[name = tensor("op_7939")]; + tensor attn_mask_37_axes_0 = const()[name = tensor("attn_mask_37_axes_0"), val = tensor([1])]; + tensor attn_mask_37 = expand_dims(axes = attn_mask_37_axes_0, x = var_7939)[name = tensor("attn_mask_37")]; + tensor var_7953_transpose_x_0 = const()[name = tensor("op_7953_transpose_x_0"), val = tensor(false)]; + tensor var_7953_transpose_y_0 = const()[name = tensor("op_7953_transpose_y_0"), val = tensor(false)]; + tensor transpose_105_perm_0 = const()[name = tensor("transpose_105_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_106_perm_0 = const()[name = tensor("transpose_106_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_106 = transpose(perm = transpose_106_perm_0, x = keys_111_cast_fp16)[name = tensor("transpose_132")]; + tensor transpose_105 = transpose(perm = transpose_105_perm_0, x = q_111_cast_fp16)[name = tensor("transpose_133")]; + tensor var_7953_cast_fp16 = matmul(transpose_x = var_7953_transpose_x_0, transpose_y = var_7953_transpose_y_0, x = transpose_105, y = transpose_106)[name = tensor("op_7953_cast_fp16")]; + tensor var_7947_to_fp16 = const()[name = tensor("op_7947_to_fp16"), val = tensor([0x1p-3])]; + tensor attn_weights_109_cast_fp16 = mul(x = var_7953_cast_fp16, y = var_7947_to_fp16)[name = tensor("attn_weights_109_cast_fp16")]; + tensor var_7955 = logical_not(x = attn_mask_37)[name = tensor("op_7955")]; + tensor var_7956_to_fp16 = const()[name = tensor("op_7956_to_fp16"), val = tensor(-inf)]; + tensor attn_weights_111_cast_fp16 = select(a = var_7956_to_fp16, b = attn_weights_109_cast_fp16, cond = var_7955)[name = tensor("attn_weights_111_cast_fp16")]; + tensor var_7958 = const()[name = tensor("op_7958"), val = tensor(-1)]; + tensor attn_weights_113_cast_fp16 = softmax(axis = var_7958, x = attn_weights_111_cast_fp16)[name = tensor("attn_weights_113_cast_fp16")]; + tensor attn_output_37_transpose_x_0 = const()[name = tensor("attn_output_37_transpose_x_0"), val = tensor(false)]; + tensor attn_output_37_transpose_y_0 = const()[name = tensor("attn_output_37_transpose_y_0"), val = tensor(false)]; + tensor values_113_cast_fp16 = transpose(perm = var_7894, x = values_111_cast_fp16)[name = tensor("transpose_134")]; + tensor attn_output_37_cast_fp16 = matmul(transpose_x = attn_output_37_transpose_x_0, transpose_y = attn_output_37_transpose_y_0, x = attn_weights_113_cast_fp16, y = values_113_cast_fp16)[name = tensor("attn_output_37_cast_fp16")]; + tensor var_7966 = const()[name = tensor("op_7966"), val = tensor([0, 2, 1, 3])]; + tensor var_7971 = const()[name = tensor("op_7971"), val = tensor([1, 256, 1024])]; + tensor var_7967_cast_fp16 = transpose(perm = var_7966, x = attn_output_37_cast_fp16)[name = tensor("transpose_131")]; + tensor input_221_cast_fp16 = reshape(shape = var_7971, x = var_7967_cast_fp16)[name = tensor("input_221_cast_fp16")]; + tensor attn18_out_proj_weight_to_fp16 = const()[name = tensor("attn18_out_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(460507776)))]; + tensor linear_73_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = attn18_out_proj_weight_to_fp16, x = input_221_cast_fp16)[name = tensor("linear_73_cast_fp16")]; + tensor var_7980_cast_fp16 = add(x = position18_to_fp16, y = valid_len_to_fp16)[name = tensor("op_7980_cast_fp16")]; + tensor var_7980_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_7980_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor input_223_cast_fp16 = add(x = input_217_cast_fp16, y = linear_73_cast_fp16)[name = tensor("input_223_cast_fp16")]; + tensor input_225_axes_0 = const()[name = tensor("input_225_axes_0"), val = tensor([-1])]; + tensor norm18_2_weight_to_fp16 = const()[name = tensor("norm18_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(462604992)))]; + tensor norm18_2_bias_to_fp16 = const()[name = tensor("norm18_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(462607104)))]; + tensor var_7984_to_fp16 = const()[name = tensor("op_7984_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_225_cast_fp16 = layer_norm(axes = input_225_axes_0, beta = norm18_2_bias_to_fp16, epsilon = var_7984_to_fp16, gamma = norm18_2_weight_to_fp16, x = input_223_cast_fp16)[name = tensor("input_225_cast_fp16")]; + tensor linear18_1_weight_to_fp16 = const()[name = tensor("linear18_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(462609216)))]; + tensor linear_74_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = linear18_1_weight_to_fp16, x = input_225_cast_fp16)[name = tensor("linear_74_cast_fp16")]; + tensor input_227_mode_0 = const()[name = tensor("input_227_mode_0"), val = tensor("EXACT")]; + tensor input_227_cast_fp16 = gelu(mode = input_227_mode_0, x = linear_74_cast_fp16)[name = tensor("input_227_cast_fp16")]; + tensor linear18_2_weight_to_fp16 = const()[name = tensor("linear18_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(470997888)))]; + tensor linear_75_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = linear18_2_weight_to_fp16, x = input_227_cast_fp16)[name = tensor("linear_75_cast_fp16")]; + tensor input_229_cast_fp16 = add(x = input_223_cast_fp16, y = linear_75_cast_fp16)[name = tensor("input_229_cast_fp16")]; + tensor input_231_axes_0 = const()[name = tensor("input_231_axes_0"), val = tensor([-1])]; + tensor norm19_1_weight_to_fp16 = const()[name = tensor("norm19_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(479386560)))]; + tensor norm19_1_bias_to_fp16 = const()[name = tensor("norm19_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(479388672)))]; + tensor var_8001_to_fp16 = const()[name = tensor("op_8001_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_231_cast_fp16 = layer_norm(axes = input_231_axes_0, beta = norm19_1_bias_to_fp16, epsilon = var_8001_to_fp16, gamma = norm19_1_weight_to_fp16, x = input_229_cast_fp16)[name = tensor("input_231_cast_fp16")]; + tensor attn19_in_proj_weight_to_fp16 = const()[name = tensor("attn19_in_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(479390784)))]; + tensor linear_76_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = attn19_in_proj_weight_to_fp16, x = input_231_cast_fp16)[name = tensor("linear_76_cast_fp16")]; + tensor var_8015 = const()[name = tensor("op_8015"), val = tensor([1, 256, 3, 16, 64])]; + tensor qkv_39_cast_fp16 = reshape(shape = var_8015, x = linear_76_cast_fp16)[name = tensor("qkv_39_cast_fp16")]; + tensor q_115_begin_0 = const()[name = tensor("q_115_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor q_115_end_0 = const()[name = tensor("q_115_end_0"), val = tensor([1, 256, 1, 16, 64])]; + tensor q_115_end_mask_0 = const()[name = tensor("q_115_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor q_115_squeeze_mask_0 = const()[name = tensor("q_115_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor q_115_cast_fp16 = slice_by_index(begin = q_115_begin_0, end = q_115_end_0, end_mask = q_115_end_mask_0, squeeze_mask = q_115_squeeze_mask_0, x = qkv_39_cast_fp16)[name = tensor("q_115_cast_fp16")]; + tensor k_77_begin_0 = const()[name = tensor("k_77_begin_0"), val = tensor([0, 0, 1, 0, 0])]; + tensor k_77_end_0 = const()[name = tensor("k_77_end_0"), val = tensor([1, 256, 2, 16, 64])]; + tensor k_77_end_mask_0 = const()[name = tensor("k_77_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor k_77_squeeze_mask_0 = const()[name = tensor("k_77_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor k_77_cast_fp16 = slice_by_index(begin = k_77_begin_0, end = k_77_end_0, end_mask = k_77_end_mask_0, squeeze_mask = k_77_squeeze_mask_0, x = qkv_39_cast_fp16)[name = tensor("k_77_cast_fp16")]; + tensor v_39_begin_0 = const()[name = tensor("v_39_begin_0"), val = tensor([0, 0, 2, 0, 0])]; + tensor v_39_end_0 = const()[name = tensor("v_39_end_0"), val = tensor([1, 256, 3, 16, 64])]; + tensor v_39_end_mask_0 = const()[name = tensor("v_39_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor v_39_squeeze_mask_0 = const()[name = tensor("v_39_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor v_39_cast_fp16 = slice_by_index(begin = v_39_begin_0, end = v_39_end_0, end_mask = v_39_end_mask_0, squeeze_mask = v_39_squeeze_mask_0, x = qkv_39_cast_fp16)[name = tensor("v_39_cast_fp16")]; + tensor var_8073 = const()[name = tensor("op_8073"), val = tensor([1, 1])]; + tensor position19_to_fp16_dtype_0 = const()[name = tensor("position19_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor position19_to_fp16 = cast(dtype = position19_to_fp16_dtype_0, x = position19)[name = tensor("cast_566")]; + tensor var_8074_cast_fp16 = reshape(shape = var_8073, x = position19_to_fp16)[name = tensor("op_8074_cast_fp16")]; + tensor ts_117_cast_fp16 = add(x = ts_1_promoted_to_fp16, y = var_8074_cast_fp16)[name = tensor("ts_117_cast_fp16")]; + tensor var_8081 = const()[name = tensor("op_8081"), val = tensor([1, 256, 1, 1])]; + tensor ts_119_cast_fp16 = reshape(shape = var_8081, x = ts_117_cast_fp16)[name = tensor("ts_119_cast_fp16")]; + tensor var_8088 = const()[name = tensor("op_8088"), val = tensor([1, 256, 16, 32, 2])]; + tensor q_complex_39_cast_fp16 = reshape(shape = var_8088, x = q_115_cast_fp16)[name = tensor("q_complex_39_cast_fp16")]; + tensor var_8095 = const()[name = tensor("op_8095"), val = tensor([1, 256, 16, 32, 2])]; + tensor k_complex_39_cast_fp16 = reshape(shape = var_8095, x = k_77_cast_fp16)[name = tensor("k_complex_39_cast_fp16")]; + tensor var_8099_begin_0 = const()[name = tensor("op_8099_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_8099_end_0 = const()[name = tensor("op_8099_end_0"), val = tensor([1, 256, 16, 32, 1])]; + tensor var_8099_end_mask_0 = const()[name = tensor("op_8099_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_8099_squeeze_mask_0 = const()[name = tensor("op_8099_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_8099_cast_fp16 = slice_by_index(begin = var_8099_begin_0, end = var_8099_end_0, end_mask = var_8099_end_mask_0, squeeze_mask = var_8099_squeeze_mask_0, x = q_complex_39_cast_fp16)[name = tensor("op_8099_cast_fp16")]; + tensor var_8107_begin_0 = const()[name = tensor("op_8107_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_8107_end_0 = const()[name = tensor("op_8107_end_0"), val = tensor([1, 256, 16, 32, 2])]; + tensor var_8107_end_mask_0 = const()[name = tensor("op_8107_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_8107_squeeze_mask_0 = const()[name = tensor("op_8107_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_8107_cast_fp16 = slice_by_index(begin = var_8107_begin_0, end = var_8107_end_0, end_mask = var_8107_end_mask_0, squeeze_mask = var_8107_squeeze_mask_0, x = q_complex_39_cast_fp16)[name = tensor("op_8107_cast_fp16")]; + tensor var_8115_begin_0 = const()[name = tensor("op_8115_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_8115_end_0 = const()[name = tensor("op_8115_end_0"), val = tensor([1, 256, 16, 32, 1])]; + tensor var_8115_end_mask_0 = const()[name = tensor("op_8115_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_8115_squeeze_mask_0 = const()[name = tensor("op_8115_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_8115_cast_fp16 = slice_by_index(begin = var_8115_begin_0, end = var_8115_end_0, end_mask = var_8115_end_mask_0, squeeze_mask = var_8115_squeeze_mask_0, x = k_complex_39_cast_fp16)[name = tensor("op_8115_cast_fp16")]; + tensor var_8123_begin_0 = const()[name = tensor("op_8123_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_8123_end_0 = const()[name = tensor("op_8123_end_0"), val = tensor([1, 256, 16, 32, 2])]; + tensor var_8123_end_mask_0 = const()[name = tensor("op_8123_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_8123_squeeze_mask_0 = const()[name = tensor("op_8123_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_8123_cast_fp16 = slice_by_index(begin = var_8123_begin_0, end = var_8123_end_0, end_mask = var_8123_end_mask_0, squeeze_mask = var_8123_squeeze_mask_0, x = k_complex_39_cast_fp16)[name = tensor("op_8123_cast_fp16")]; + tensor freqs_39_to_fp16 = const()[name = tensor("freqs_39_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(485682304)))]; + tensor var_8129_cast_fp16 = mul(x = freqs_39_to_fp16, y = ts_119_cast_fp16)[name = tensor("op_8129_cast_fp16")]; + tensor rotr_39_cast_fp16 = cos(x = var_8129_cast_fp16)[name = tensor("rotr_39_cast_fp16")]; + tensor roti_39_cast_fp16 = sin(x = var_8129_cast_fp16)[name = tensor("roti_39_cast_fp16")]; + tensor var_8133_cast_fp16 = mul(x = var_8099_cast_fp16, y = rotr_39_cast_fp16)[name = tensor("op_8133_cast_fp16")]; + tensor var_8134_cast_fp16 = mul(x = var_8107_cast_fp16, y = roti_39_cast_fp16)[name = tensor("op_8134_cast_fp16")]; + tensor qor_77_cast_fp16 = sub(x = var_8133_cast_fp16, y = var_8134_cast_fp16)[name = tensor("qor_77_cast_fp16")]; + tensor var_8137_cast_fp16 = mul(x = var_8099_cast_fp16, y = roti_39_cast_fp16)[name = tensor("op_8137_cast_fp16")]; + tensor var_8138_cast_fp16 = mul(x = var_8107_cast_fp16, y = rotr_39_cast_fp16)[name = tensor("op_8138_cast_fp16")]; + tensor qoi_77_cast_fp16 = add(x = var_8137_cast_fp16, y = var_8138_cast_fp16)[name = tensor("qoi_77_cast_fp16")]; + tensor var_8141_cast_fp16 = mul(x = var_8115_cast_fp16, y = rotr_39_cast_fp16)[name = tensor("op_8141_cast_fp16")]; + tensor var_8142_cast_fp16 = mul(x = var_8123_cast_fp16, y = roti_39_cast_fp16)[name = tensor("op_8142_cast_fp16")]; + tensor kor_77_cast_fp16 = sub(x = var_8141_cast_fp16, y = var_8142_cast_fp16)[name = tensor("kor_77_cast_fp16")]; + tensor var_8145_cast_fp16 = mul(x = var_8115_cast_fp16, y = roti_39_cast_fp16)[name = tensor("op_8145_cast_fp16")]; + tensor var_8146_cast_fp16 = mul(x = var_8123_cast_fp16, y = rotr_39_cast_fp16)[name = tensor("op_8146_cast_fp16")]; + tensor koi_77_cast_fp16 = add(x = var_8145_cast_fp16, y = var_8146_cast_fp16)[name = tensor("koi_77_cast_fp16")]; + tensor qo_39_axis_0 = const()[name = tensor("qo_39_axis_0"), val = tensor(-1)]; + tensor qo_39_cast_fp16 = stack(axis = qo_39_axis_0, values = (qor_77_cast_fp16, qoi_77_cast_fp16))[name = tensor("qo_39_cast_fp16")]; + tensor ko_39_axis_0 = const()[name = tensor("ko_39_axis_0"), val = tensor(-1)]; + tensor ko_39_cast_fp16 = stack(axis = ko_39_axis_0, values = (kor_77_cast_fp16, koi_77_cast_fp16))[name = tensor("ko_39_cast_fp16")]; + tensor var_8179 = const()[name = tensor("op_8179"), val = tensor([1, 256, 16, 64])]; + tensor q_117_cast_fp16 = reshape(shape = var_8179, x = qo_39_cast_fp16)[name = tensor("q_117_cast_fp16")]; + tensor var_8185 = const()[name = tensor("op_8185"), val = tensor([1, 256, 16, 64])]; + tensor k_79_cast_fp16 = reshape(shape = var_8185, x = ko_39_cast_fp16)[name = tensor("k_79_cast_fp16")]; + tensor write_indices_float_115_cast_fp16 = add(x = var_8074_cast_fp16, y = var_393_promoted_to_fp16)[name = tensor("write_indices_float_115_cast_fp16")]; + tensor _inversed_8206_y_0_to_fp16 = const()[name = tensor("_inversed_8206_y_0_to_fp16"), val = tensor(0x1p-9)]; + tensor _inversed_8206_cast_fp16 = mul(x = write_indices_float_115_cast_fp16, y = _inversed_8206_y_0_to_fp16)[name = tensor("_inversed_8206_cast_fp16")]; + tensor var_8207_cast_fp16 = floor(x = _inversed_8206_cast_fp16)[name = tensor("op_8207_cast_fp16")]; + tensor var_8208_to_fp16 = const()[name = tensor("op_8208_to_fp16"), val = tensor(0x1p+9)]; + tensor var_8209_cast_fp16 = mul(x = var_8207_cast_fp16, y = var_8208_to_fp16)[name = tensor("op_8209_cast_fp16")]; + tensor write_indices_float_117_cast_fp16 = sub(x = write_indices_float_115_cast_fp16, y = var_8209_cast_fp16)[name = tensor("write_indices_float_117_cast_fp16")]; + tensor var_8227_value_0_to_fp16 = const()[name = tensor("op_8227_value_0_to_fp16"), val = tensor(0x1.ffp+8)]; + tensor var_8227_cast_fp16 = fill_like(ref_tensor = write_indices_float_117_cast_fp16, value = var_8227_value_0_to_fp16)[name = tensor("op_8227_cast_fp16")]; + tensor write_indices_float_119_cast_fp16 = select(a = write_indices_float_117_cast_fp16, b = var_8227_cast_fp16, cond = token_valid_1_cast_fp16)[name = tensor("write_indices_float_119_cast_fp16")]; + tensor var_8233_dtype_0 = const()[name = tensor("op_8233_dtype_0"), val = tensor("int32")]; + tensor var_8238 = const()[name = tensor("op_8238"), val = tensor([1, 256, 1, 1])]; + tensor write_indices_float_119_cast_fp16_to_int32 = cast(dtype = var_8233_dtype_0, x = write_indices_float_119_cast_fp16)[name = tensor("cast_565")]; + tensor var_8239 = reshape(shape = var_8238, x = write_indices_float_119_cast_fp16_to_int32)[name = tensor("op_8239")]; + tensor write_indices_39_reps_0 = const()[name = tensor("write_indices_39_reps_0"), val = tensor([1, 1, 16, 64])]; + tensor write_indices_39 = tile(reps = write_indices_39_reps_0, x = var_8239)[name = tensor("write_indices_39")]; + tensor var_8249_begin_0 = const()[name = tensor("op_8249_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_8249_end_0 = const()[name = tensor("op_8249_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor var_8249_end_mask_0 = const()[name = tensor("op_8249_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_8249_squeeze_mask_0 = const()[name = tensor("op_8249_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor cache19_to_fp16_dtype_0 = const()[name = tensor("cache19_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor cache19_to_fp16 = cast(dtype = cache19_to_fp16_dtype_0, x = cache19)[name = tensor("cast_564")]; + tensor var_8249_cast_fp16 = slice_by_index(begin = var_8249_begin_0, end = var_8249_end_0, end_mask = var_8249_end_mask_0, squeeze_mask = var_8249_squeeze_mask_0, x = cache19_to_fp16)[name = tensor("op_8249_cast_fp16")]; + tensor var_8251_axis_0 = const()[name = tensor("op_8251_axis_0"), val = tensor(1)]; + tensor var_8251_mode_0 = const()[name = tensor("op_8251_mode_0"), val = tensor("update")]; + tensor var_8251_validate_indices_0 = const()[name = tensor("op_8251_validate_indices_0"), val = tensor(false)]; + tensor var_8251_cast_fp16 = scatter_along_axis(axis = var_8251_axis_0, data = var_8249_cast_fp16, indices = write_indices_39, mode = var_8251_mode_0, updates = k_79_cast_fp16, validate_indices = var_8251_validate_indices_0)[name = tensor("op_8251_cast_fp16")]; + tensor concat_154 = const()[name = tensor("concat_154"), val = tensor([0, 0, 0, 0, 0])]; + tensor concat_155 = const()[name = tensor("concat_155"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_39_internal_tensor_assign_1_stride_0 = const()[name = tensor("new_cache_39_internal_tensor_assign_1_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_39_internal_tensor_assign_1_begin_mask_0 = const()[name = tensor("new_cache_39_internal_tensor_assign_1_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_39_internal_tensor_assign_1_end_mask_0 = const()[name = tensor("new_cache_39_internal_tensor_assign_1_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_39_internal_tensor_assign_1_squeeze_mask_0 = const()[name = tensor("new_cache_39_internal_tensor_assign_1_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_84 = const()[name = tensor("shape_84"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_38 = const()[name = tensor("reduce_prod_38"), val = tensor(1048576)]; + tensor range_1d_38_start_0 = const()[name = tensor("range_1d_38_start_0"), val = tensor(0)]; + tensor range_1d_38_step_0 = const()[name = tensor("range_1d_38_step_0"), val = tensor(1)]; + tensor range_1d_38 = range_1d(end = reduce_prod_38, start = range_1d_38_start_0, step = range_1d_38_step_0)[name = tensor("range_1d_38")]; + tensor reshape_190 = reshape(shape = shape_84, x = range_1d_38)[name = tensor("reshape_190")]; + tensor slice_by_index_38 = slice_by_index(begin = concat_154, begin_mask = new_cache_39_internal_tensor_assign_1_begin_mask_0, end = concat_155, end_mask = new_cache_39_internal_tensor_assign_1_end_mask_0, squeeze_mask = new_cache_39_internal_tensor_assign_1_squeeze_mask_0, stride = new_cache_39_internal_tensor_assign_1_stride_0, x = reshape_190)[name = tensor("slice_by_index_38")]; + tensor reshape_191_shape_0 = const()[name = tensor("reshape_191_shape_0"), val = tensor([-1])]; + tensor reshape_191 = reshape(shape = reshape_191_shape_0, x = slice_by_index_38)[name = tensor("reshape_191")]; + tensor reshape_192_shape_0 = const()[name = tensor("reshape_192_shape_0"), val = tensor([-1])]; + tensor reshape_192_cast_fp16 = reshape(shape = reshape_192_shape_0, x = var_8251_cast_fp16)[name = tensor("reshape_192_cast_fp16")]; + tensor reshape_193_shape_0 = const()[name = tensor("reshape_193_shape_0"), val = tensor([-1])]; + tensor reshape_193_cast_fp16 = reshape(shape = reshape_193_shape_0, x = cache19_to_fp16)[name = tensor("reshape_193_cast_fp16")]; + tensor scatter_38_mode_0 = const()[name = tensor("scatter_38_mode_0"), val = tensor("update")]; + tensor scatter_38_axis_0 = const()[name = tensor("scatter_38_axis_0"), val = tensor(0)]; + tensor scatter_38_validate_indices_0 = const()[name = tensor("scatter_38_validate_indices_0"), val = tensor(false)]; + tensor scatter_38_cast_fp16 = scatter(axis = scatter_38_axis_0, data = reshape_193_cast_fp16, indices = reshape_191, mode = scatter_38_mode_0, updates = reshape_192_cast_fp16, validate_indices = scatter_38_validate_indices_0)[name = tensor("scatter_38_cast_fp16")]; + tensor reshape_194_cast_fp16 = reshape(shape = shape_84, x = scatter_38_cast_fp16)[name = tensor("reshape_194_cast_fp16")]; + tensor var_8259_begin_0 = const()[name = tensor("op_8259_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor var_8259_end_0 = const()[name = tensor("op_8259_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor var_8259_end_mask_0 = const()[name = tensor("op_8259_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_8259_squeeze_mask_0 = const()[name = tensor("op_8259_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor var_8259_cast_fp16 = slice_by_index(begin = var_8259_begin_0, end = var_8259_end_0, end_mask = var_8259_end_mask_0, squeeze_mask = var_8259_squeeze_mask_0, x = reshape_194_cast_fp16)[name = tensor("op_8259_cast_fp16")]; + tensor var_8261_axis_0 = const()[name = tensor("op_8261_axis_0"), val = tensor(1)]; + tensor var_8261_mode_0 = const()[name = tensor("op_8261_mode_0"), val = tensor("update")]; + tensor var_8261_validate_indices_0 = const()[name = tensor("op_8261_validate_indices_0"), val = tensor(false)]; + tensor var_8261_cast_fp16 = scatter_along_axis(axis = var_8261_axis_0, data = var_8259_cast_fp16, indices = write_indices_39, mode = var_8261_mode_0, updates = v_39_cast_fp16, validate_indices = var_8261_validate_indices_0)[name = tensor("op_8261_cast_fp16")]; + tensor concat_156 = const()[name = tensor("concat_156"), val = tensor([1, 0, 0, 0, 0])]; + tensor concat_157 = const()[name = tensor("concat_157"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_39_internal_tensor_assign_2_stride_0 = const()[name = tensor("new_cache_39_internal_tensor_assign_2_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_39_internal_tensor_assign_2_begin_mask_0 = const()[name = tensor("new_cache_39_internal_tensor_assign_2_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_39_internal_tensor_assign_2_end_mask_0 = const()[name = tensor("new_cache_39_internal_tensor_assign_2_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_39_internal_tensor_assign_2_squeeze_mask_0 = const()[name = tensor("new_cache_39_internal_tensor_assign_2_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_85 = const()[name = tensor("shape_85"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_39 = const()[name = tensor("reduce_prod_39"), val = tensor(1048576)]; + tensor range_1d_39_start_0 = const()[name = tensor("range_1d_39_start_0"), val = tensor(0)]; + tensor range_1d_39_step_0 = const()[name = tensor("range_1d_39_step_0"), val = tensor(1)]; + tensor range_1d_39 = range_1d(end = reduce_prod_39, start = range_1d_39_start_0, step = range_1d_39_step_0)[name = tensor("range_1d_39")]; + tensor reshape_195 = reshape(shape = shape_85, x = range_1d_39)[name = tensor("reshape_195")]; + tensor slice_by_index_39 = slice_by_index(begin = concat_156, begin_mask = new_cache_39_internal_tensor_assign_2_begin_mask_0, end = concat_157, end_mask = new_cache_39_internal_tensor_assign_2_end_mask_0, squeeze_mask = new_cache_39_internal_tensor_assign_2_squeeze_mask_0, stride = new_cache_39_internal_tensor_assign_2_stride_0, x = reshape_195)[name = tensor("slice_by_index_39")]; + tensor reshape_196_shape_0 = const()[name = tensor("reshape_196_shape_0"), val = tensor([-1])]; + tensor reshape_196 = reshape(shape = reshape_196_shape_0, x = slice_by_index_39)[name = tensor("reshape_196")]; + tensor reshape_197_shape_0 = const()[name = tensor("reshape_197_shape_0"), val = tensor([-1])]; + tensor reshape_197_cast_fp16 = reshape(shape = reshape_197_shape_0, x = var_8261_cast_fp16)[name = tensor("reshape_197_cast_fp16")]; + tensor reshape_198_shape_0 = const()[name = tensor("reshape_198_shape_0"), val = tensor([-1])]; + tensor reshape_198_cast_fp16 = reshape(shape = reshape_198_shape_0, x = reshape_194_cast_fp16)[name = tensor("reshape_198_cast_fp16")]; + tensor scatter_39_mode_0 = const()[name = tensor("scatter_39_mode_0"), val = tensor("update")]; + tensor scatter_39_axis_0 = const()[name = tensor("scatter_39_axis_0"), val = tensor(0)]; + tensor scatter_39_validate_indices_0 = const()[name = tensor("scatter_39_validate_indices_0"), val = tensor(false)]; + tensor scatter_39_cast_fp16 = scatter(axis = scatter_39_axis_0, data = reshape_198_cast_fp16, indices = reshape_196, mode = scatter_39_mode_0, updates = reshape_197_cast_fp16, validate_indices = scatter_39_validate_indices_0)[name = tensor("scatter_39_cast_fp16")]; + tensor reshape_199_cast_fp16 = reshape(shape = shape_85, x = scatter_39_cast_fp16)[name = tensor("reshape_199_cast_fp16")]; + tensor reshape_199_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("reshape_199_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor keys_115_begin_0 = const()[name = tensor("keys_115_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor keys_115_end_0 = const()[name = tensor("keys_115_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor keys_115_end_mask_0 = const()[name = tensor("keys_115_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor keys_115_squeeze_mask_0 = const()[name = tensor("keys_115_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor keys_115_cast_fp16 = slice_by_index(begin = keys_115_begin_0, end = keys_115_end_0, end_mask = keys_115_end_mask_0, squeeze_mask = keys_115_squeeze_mask_0, x = reshape_199_cast_fp16)[name = tensor("keys_115_cast_fp16")]; + tensor values_115_begin_0 = const()[name = tensor("values_115_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor values_115_end_0 = const()[name = tensor("values_115_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor values_115_end_mask_0 = const()[name = tensor("values_115_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor values_115_squeeze_mask_0 = const()[name = tensor("values_115_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor values_115_cast_fp16 = slice_by_index(begin = values_115_begin_0, end = values_115_end_0, end_mask = values_115_end_mask_0, squeeze_mask = values_115_squeeze_mask_0, x = reshape_199_cast_fp16)[name = tensor("values_115_cast_fp16")]; + tensor var_8273_cast_fp16 = not_equal(x = keys_115_cast_fp16, y = keys_115_cast_fp16)[name = tensor("op_8273_cast_fp16")]; + tensor keys_117_cast_fp16 = select(a = var_470_to_fp16, b = keys_115_cast_fp16, cond = var_8273_cast_fp16)[name = tensor("keys_117_cast_fp16")]; + tensor var_8281_cast_fp16 = not_equal(x = values_115_cast_fp16, y = values_115_cast_fp16)[name = tensor("op_8281_cast_fp16")]; + tensor values_117_cast_fp16 = select(a = var_470_to_fp16, b = values_115_cast_fp16, cond = var_8281_cast_fp16)[name = tensor("values_117_cast_fp16")]; + tensor var_8305 = const()[name = tensor("op_8305"), val = tensor([0, 2, 1, 3])]; + tensor var_8321 = const()[name = tensor("op_8321"), val = tensor([1, 1, 1])]; + tensor var_8322_cast_fp16 = reshape(shape = var_8321, x = position19_to_fp16)[name = tensor("op_8322_cast_fp16")]; + tensor q_positions_39_cast_fp16 = add(x = var_8322_cast_fp16, y = q_offsets_1_promoted_to_fp16)[name = tensor("q_positions_39_cast_fp16")]; + tensor valid_end_39_cast_fp16 = add(x = var_8322_cast_fp16, y = var_536_cast_fp16)[name = tensor("valid_end_39_cast_fp16")]; + tensor valid_mask_39_cast_fp16 = less(x = k_positions_1_promoted_to_fp16, y = valid_end_39_cast_fp16)[name = tensor("valid_mask_39_cast_fp16")]; + tensor causal_mask_39_cast_fp16 = less_equal(x = k_positions_1_promoted_to_fp16, y = q_positions_39_cast_fp16)[name = tensor("causal_mask_39_cast_fp16")]; + tensor var_8350 = logical_and(x = valid_mask_39_cast_fp16, y = causal_mask_39_cast_fp16)[name = tensor("op_8350")]; + tensor attn_mask_39_axes_0 = const()[name = tensor("attn_mask_39_axes_0"), val = tensor([1])]; + tensor attn_mask_39 = expand_dims(axes = attn_mask_39_axes_0, x = var_8350)[name = tensor("attn_mask_39")]; + tensor var_8364_transpose_x_0 = const()[name = tensor("op_8364_transpose_x_0"), val = tensor(false)]; + tensor var_8364_transpose_y_0 = const()[name = tensor("op_8364_transpose_y_0"), val = tensor(false)]; + tensor transpose_107_perm_0 = const()[name = tensor("transpose_107_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_108_perm_0 = const()[name = tensor("transpose_108_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_108 = transpose(perm = transpose_108_perm_0, x = keys_117_cast_fp16)[name = tensor("transpose_128")]; + tensor transpose_107 = transpose(perm = transpose_107_perm_0, x = q_117_cast_fp16)[name = tensor("transpose_129")]; + tensor var_8364_cast_fp16 = matmul(transpose_x = var_8364_transpose_x_0, transpose_y = var_8364_transpose_y_0, x = transpose_107, y = transpose_108)[name = tensor("op_8364_cast_fp16")]; + tensor var_8358_to_fp16 = const()[name = tensor("op_8358_to_fp16"), val = tensor([0x1p-3])]; + tensor attn_weights_115_cast_fp16 = mul(x = var_8364_cast_fp16, y = var_8358_to_fp16)[name = tensor("attn_weights_115_cast_fp16")]; + tensor var_8366 = logical_not(x = attn_mask_39)[name = tensor("op_8366")]; + tensor var_8367_to_fp16 = const()[name = tensor("op_8367_to_fp16"), val = tensor(-inf)]; + tensor attn_weights_117_cast_fp16 = select(a = var_8367_to_fp16, b = attn_weights_115_cast_fp16, cond = var_8366)[name = tensor("attn_weights_117_cast_fp16")]; + tensor var_8369 = const()[name = tensor("op_8369"), val = tensor(-1)]; + tensor attn_weights_119_cast_fp16 = softmax(axis = var_8369, x = attn_weights_117_cast_fp16)[name = tensor("attn_weights_119_cast_fp16")]; + tensor attn_output_39_transpose_x_0 = const()[name = tensor("attn_output_39_transpose_x_0"), val = tensor(false)]; + tensor attn_output_39_transpose_y_0 = const()[name = tensor("attn_output_39_transpose_y_0"), val = tensor(false)]; + tensor values_119_cast_fp16 = transpose(perm = var_8305, x = values_117_cast_fp16)[name = tensor("transpose_130")]; + tensor attn_output_39_cast_fp16 = matmul(transpose_x = attn_output_39_transpose_x_0, transpose_y = attn_output_39_transpose_y_0, x = attn_weights_119_cast_fp16, y = values_119_cast_fp16)[name = tensor("attn_output_39_cast_fp16")]; + tensor var_8377 = const()[name = tensor("op_8377"), val = tensor([0, 2, 1, 3])]; + tensor var_8382 = const()[name = tensor("op_8382"), val = tensor([1, 256, 1024])]; + tensor var_8378_cast_fp16 = transpose(perm = var_8377, x = attn_output_39_cast_fp16)[name = tensor("transpose_127")]; + tensor input_233_cast_fp16 = reshape(shape = var_8382, x = var_8378_cast_fp16)[name = tensor("input_233_cast_fp16")]; + tensor attn19_out_proj_weight_to_fp16 = const()[name = tensor("attn19_out_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(485682432)))]; + tensor linear_77_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = attn19_out_proj_weight_to_fp16, x = input_233_cast_fp16)[name = tensor("linear_77_cast_fp16")]; + tensor var_8391_cast_fp16 = add(x = position19_to_fp16, y = valid_len_to_fp16)[name = tensor("op_8391_cast_fp16")]; + tensor var_8391_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_8391_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor input_235_cast_fp16 = add(x = input_229_cast_fp16, y = linear_77_cast_fp16)[name = tensor("input_235_cast_fp16")]; + tensor input_237_axes_0 = const()[name = tensor("input_237_axes_0"), val = tensor([-1])]; + tensor norm19_2_weight_to_fp16 = const()[name = tensor("norm19_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(487779648)))]; + tensor norm19_2_bias_to_fp16 = const()[name = tensor("norm19_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(487781760)))]; + tensor var_8395_to_fp16 = const()[name = tensor("op_8395_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_237_cast_fp16 = layer_norm(axes = input_237_axes_0, beta = norm19_2_bias_to_fp16, epsilon = var_8395_to_fp16, gamma = norm19_2_weight_to_fp16, x = input_235_cast_fp16)[name = tensor("input_237_cast_fp16")]; + tensor linear19_1_weight_to_fp16 = const()[name = tensor("linear19_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(487783872)))]; + tensor linear_78_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = linear19_1_weight_to_fp16, x = input_237_cast_fp16)[name = tensor("linear_78_cast_fp16")]; + tensor input_239_mode_0 = const()[name = tensor("input_239_mode_0"), val = tensor("EXACT")]; + tensor input_239_cast_fp16 = gelu(mode = input_239_mode_0, x = linear_78_cast_fp16)[name = tensor("input_239_cast_fp16")]; + tensor linear19_2_weight_to_fp16 = const()[name = tensor("linear19_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(496172544)))]; + tensor linear_79_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = linear19_2_weight_to_fp16, x = input_239_cast_fp16)[name = tensor("linear_79_cast_fp16")]; + tensor input_241_cast_fp16 = add(x = input_235_cast_fp16, y = linear_79_cast_fp16)[name = tensor("input_241_cast_fp16")]; + tensor input_243_axes_0 = const()[name = tensor("input_243_axes_0"), val = tensor([-1])]; + tensor norm20_1_weight_to_fp16 = const()[name = tensor("norm20_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(504561216)))]; + tensor norm20_1_bias_to_fp16 = const()[name = tensor("norm20_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(504563328)))]; + tensor var_8412_to_fp16 = const()[name = tensor("op_8412_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_243_cast_fp16 = layer_norm(axes = input_243_axes_0, beta = norm20_1_bias_to_fp16, epsilon = var_8412_to_fp16, gamma = norm20_1_weight_to_fp16, x = input_241_cast_fp16)[name = tensor("input_243_cast_fp16")]; + tensor attn20_in_proj_weight_to_fp16 = const()[name = tensor("attn20_in_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(504565440)))]; + tensor linear_80_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = attn20_in_proj_weight_to_fp16, x = input_243_cast_fp16)[name = tensor("linear_80_cast_fp16")]; + tensor var_8426 = const()[name = tensor("op_8426"), val = tensor([1, 256, 3, 16, 64])]; + tensor qkv_41_cast_fp16 = reshape(shape = var_8426, x = linear_80_cast_fp16)[name = tensor("qkv_41_cast_fp16")]; + tensor q_121_begin_0 = const()[name = tensor("q_121_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor q_121_end_0 = const()[name = tensor("q_121_end_0"), val = tensor([1, 256, 1, 16, 64])]; + tensor q_121_end_mask_0 = const()[name = tensor("q_121_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor q_121_squeeze_mask_0 = const()[name = tensor("q_121_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor q_121_cast_fp16 = slice_by_index(begin = q_121_begin_0, end = q_121_end_0, end_mask = q_121_end_mask_0, squeeze_mask = q_121_squeeze_mask_0, x = qkv_41_cast_fp16)[name = tensor("q_121_cast_fp16")]; + tensor k_81_begin_0 = const()[name = tensor("k_81_begin_0"), val = tensor([0, 0, 1, 0, 0])]; + tensor k_81_end_0 = const()[name = tensor("k_81_end_0"), val = tensor([1, 256, 2, 16, 64])]; + tensor k_81_end_mask_0 = const()[name = tensor("k_81_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor k_81_squeeze_mask_0 = const()[name = tensor("k_81_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor k_81_cast_fp16 = slice_by_index(begin = k_81_begin_0, end = k_81_end_0, end_mask = k_81_end_mask_0, squeeze_mask = k_81_squeeze_mask_0, x = qkv_41_cast_fp16)[name = tensor("k_81_cast_fp16")]; + tensor v_41_begin_0 = const()[name = tensor("v_41_begin_0"), val = tensor([0, 0, 2, 0, 0])]; + tensor v_41_end_0 = const()[name = tensor("v_41_end_0"), val = tensor([1, 256, 3, 16, 64])]; + tensor v_41_end_mask_0 = const()[name = tensor("v_41_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor v_41_squeeze_mask_0 = const()[name = tensor("v_41_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor v_41_cast_fp16 = slice_by_index(begin = v_41_begin_0, end = v_41_end_0, end_mask = v_41_end_mask_0, squeeze_mask = v_41_squeeze_mask_0, x = qkv_41_cast_fp16)[name = tensor("v_41_cast_fp16")]; + tensor var_8484 = const()[name = tensor("op_8484"), val = tensor([1, 1])]; + tensor position20_to_fp16_dtype_0 = const()[name = tensor("position20_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor position20_to_fp16 = cast(dtype = position20_to_fp16_dtype_0, x = position20)[name = tensor("cast_561")]; + tensor var_8485_cast_fp16 = reshape(shape = var_8484, x = position20_to_fp16)[name = tensor("op_8485_cast_fp16")]; + tensor ts_123_cast_fp16 = add(x = ts_1_promoted_to_fp16, y = var_8485_cast_fp16)[name = tensor("ts_123_cast_fp16")]; + tensor var_8492 = const()[name = tensor("op_8492"), val = tensor([1, 256, 1, 1])]; + tensor ts_125_cast_fp16 = reshape(shape = var_8492, x = ts_123_cast_fp16)[name = tensor("ts_125_cast_fp16")]; + tensor var_8499 = const()[name = tensor("op_8499"), val = tensor([1, 256, 16, 32, 2])]; + tensor q_complex_41_cast_fp16 = reshape(shape = var_8499, x = q_121_cast_fp16)[name = tensor("q_complex_41_cast_fp16")]; + tensor var_8506 = const()[name = tensor("op_8506"), val = tensor([1, 256, 16, 32, 2])]; + tensor k_complex_41_cast_fp16 = reshape(shape = var_8506, x = k_81_cast_fp16)[name = tensor("k_complex_41_cast_fp16")]; + tensor var_8510_begin_0 = const()[name = tensor("op_8510_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_8510_end_0 = const()[name = tensor("op_8510_end_0"), val = tensor([1, 256, 16, 32, 1])]; + tensor var_8510_end_mask_0 = const()[name = tensor("op_8510_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_8510_squeeze_mask_0 = const()[name = tensor("op_8510_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_8510_cast_fp16 = slice_by_index(begin = var_8510_begin_0, end = var_8510_end_0, end_mask = var_8510_end_mask_0, squeeze_mask = var_8510_squeeze_mask_0, x = q_complex_41_cast_fp16)[name = tensor("op_8510_cast_fp16")]; + tensor var_8518_begin_0 = const()[name = tensor("op_8518_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_8518_end_0 = const()[name = tensor("op_8518_end_0"), val = tensor([1, 256, 16, 32, 2])]; + tensor var_8518_end_mask_0 = const()[name = tensor("op_8518_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_8518_squeeze_mask_0 = const()[name = tensor("op_8518_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_8518_cast_fp16 = slice_by_index(begin = var_8518_begin_0, end = var_8518_end_0, end_mask = var_8518_end_mask_0, squeeze_mask = var_8518_squeeze_mask_0, x = q_complex_41_cast_fp16)[name = tensor("op_8518_cast_fp16")]; + tensor var_8526_begin_0 = const()[name = tensor("op_8526_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_8526_end_0 = const()[name = tensor("op_8526_end_0"), val = tensor([1, 256, 16, 32, 1])]; + tensor var_8526_end_mask_0 = const()[name = tensor("op_8526_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_8526_squeeze_mask_0 = const()[name = tensor("op_8526_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_8526_cast_fp16 = slice_by_index(begin = var_8526_begin_0, end = var_8526_end_0, end_mask = var_8526_end_mask_0, squeeze_mask = var_8526_squeeze_mask_0, x = k_complex_41_cast_fp16)[name = tensor("op_8526_cast_fp16")]; + tensor var_8534_begin_0 = const()[name = tensor("op_8534_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_8534_end_0 = const()[name = tensor("op_8534_end_0"), val = tensor([1, 256, 16, 32, 2])]; + tensor var_8534_end_mask_0 = const()[name = tensor("op_8534_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_8534_squeeze_mask_0 = const()[name = tensor("op_8534_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_8534_cast_fp16 = slice_by_index(begin = var_8534_begin_0, end = var_8534_end_0, end_mask = var_8534_end_mask_0, squeeze_mask = var_8534_squeeze_mask_0, x = k_complex_41_cast_fp16)[name = tensor("op_8534_cast_fp16")]; + tensor freqs_41_to_fp16 = const()[name = tensor("freqs_41_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(510856960)))]; + tensor var_8540_cast_fp16 = mul(x = freqs_41_to_fp16, y = ts_125_cast_fp16)[name = tensor("op_8540_cast_fp16")]; + tensor rotr_41_cast_fp16 = cos(x = var_8540_cast_fp16)[name = tensor("rotr_41_cast_fp16")]; + tensor roti_41_cast_fp16 = sin(x = var_8540_cast_fp16)[name = tensor("roti_41_cast_fp16")]; + tensor var_8544_cast_fp16 = mul(x = var_8510_cast_fp16, y = rotr_41_cast_fp16)[name = tensor("op_8544_cast_fp16")]; + tensor var_8545_cast_fp16 = mul(x = var_8518_cast_fp16, y = roti_41_cast_fp16)[name = tensor("op_8545_cast_fp16")]; + tensor qor_81_cast_fp16 = sub(x = var_8544_cast_fp16, y = var_8545_cast_fp16)[name = tensor("qor_81_cast_fp16")]; + tensor var_8548_cast_fp16 = mul(x = var_8510_cast_fp16, y = roti_41_cast_fp16)[name = tensor("op_8548_cast_fp16")]; + tensor var_8549_cast_fp16 = mul(x = var_8518_cast_fp16, y = rotr_41_cast_fp16)[name = tensor("op_8549_cast_fp16")]; + tensor qoi_81_cast_fp16 = add(x = var_8548_cast_fp16, y = var_8549_cast_fp16)[name = tensor("qoi_81_cast_fp16")]; + tensor var_8552_cast_fp16 = mul(x = var_8526_cast_fp16, y = rotr_41_cast_fp16)[name = tensor("op_8552_cast_fp16")]; + tensor var_8553_cast_fp16 = mul(x = var_8534_cast_fp16, y = roti_41_cast_fp16)[name = tensor("op_8553_cast_fp16")]; + tensor kor_81_cast_fp16 = sub(x = var_8552_cast_fp16, y = var_8553_cast_fp16)[name = tensor("kor_81_cast_fp16")]; + tensor var_8556_cast_fp16 = mul(x = var_8526_cast_fp16, y = roti_41_cast_fp16)[name = tensor("op_8556_cast_fp16")]; + tensor var_8557_cast_fp16 = mul(x = var_8534_cast_fp16, y = rotr_41_cast_fp16)[name = tensor("op_8557_cast_fp16")]; + tensor koi_81_cast_fp16 = add(x = var_8556_cast_fp16, y = var_8557_cast_fp16)[name = tensor("koi_81_cast_fp16")]; + tensor qo_41_axis_0 = const()[name = tensor("qo_41_axis_0"), val = tensor(-1)]; + tensor qo_41_cast_fp16 = stack(axis = qo_41_axis_0, values = (qor_81_cast_fp16, qoi_81_cast_fp16))[name = tensor("qo_41_cast_fp16")]; + tensor ko_41_axis_0 = const()[name = tensor("ko_41_axis_0"), val = tensor(-1)]; + tensor ko_41_cast_fp16 = stack(axis = ko_41_axis_0, values = (kor_81_cast_fp16, koi_81_cast_fp16))[name = tensor("ko_41_cast_fp16")]; + tensor var_8590 = const()[name = tensor("op_8590"), val = tensor([1, 256, 16, 64])]; + tensor q_123_cast_fp16 = reshape(shape = var_8590, x = qo_41_cast_fp16)[name = tensor("q_123_cast_fp16")]; + tensor var_8596 = const()[name = tensor("op_8596"), val = tensor([1, 256, 16, 64])]; + tensor k_83_cast_fp16 = reshape(shape = var_8596, x = ko_41_cast_fp16)[name = tensor("k_83_cast_fp16")]; + tensor write_indices_float_121_cast_fp16 = add(x = var_8485_cast_fp16, y = var_393_promoted_to_fp16)[name = tensor("write_indices_float_121_cast_fp16")]; + tensor _inversed_8617_y_0_to_fp16 = const()[name = tensor("_inversed_8617_y_0_to_fp16"), val = tensor(0x1p-9)]; + tensor _inversed_8617_cast_fp16 = mul(x = write_indices_float_121_cast_fp16, y = _inversed_8617_y_0_to_fp16)[name = tensor("_inversed_8617_cast_fp16")]; + tensor var_8618_cast_fp16 = floor(x = _inversed_8617_cast_fp16)[name = tensor("op_8618_cast_fp16")]; + tensor var_8619_to_fp16 = const()[name = tensor("op_8619_to_fp16"), val = tensor(0x1p+9)]; + tensor var_8620_cast_fp16 = mul(x = var_8618_cast_fp16, y = var_8619_to_fp16)[name = tensor("op_8620_cast_fp16")]; + tensor write_indices_float_123_cast_fp16 = sub(x = write_indices_float_121_cast_fp16, y = var_8620_cast_fp16)[name = tensor("write_indices_float_123_cast_fp16")]; + tensor var_8638_value_0_to_fp16 = const()[name = tensor("op_8638_value_0_to_fp16"), val = tensor(0x1.ffp+8)]; + tensor var_8638_cast_fp16 = fill_like(ref_tensor = write_indices_float_123_cast_fp16, value = var_8638_value_0_to_fp16)[name = tensor("op_8638_cast_fp16")]; + tensor write_indices_float_125_cast_fp16 = select(a = write_indices_float_123_cast_fp16, b = var_8638_cast_fp16, cond = token_valid_1_cast_fp16)[name = tensor("write_indices_float_125_cast_fp16")]; + tensor var_8644_dtype_0 = const()[name = tensor("op_8644_dtype_0"), val = tensor("int32")]; + tensor var_8649 = const()[name = tensor("op_8649"), val = tensor([1, 256, 1, 1])]; + tensor write_indices_float_125_cast_fp16_to_int32 = cast(dtype = var_8644_dtype_0, x = write_indices_float_125_cast_fp16)[name = tensor("cast_560")]; + tensor var_8650 = reshape(shape = var_8649, x = write_indices_float_125_cast_fp16_to_int32)[name = tensor("op_8650")]; + tensor write_indices_41_reps_0 = const()[name = tensor("write_indices_41_reps_0"), val = tensor([1, 1, 16, 64])]; + tensor write_indices_41 = tile(reps = write_indices_41_reps_0, x = var_8650)[name = tensor("write_indices_41")]; + tensor var_8660_begin_0 = const()[name = tensor("op_8660_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_8660_end_0 = const()[name = tensor("op_8660_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor var_8660_end_mask_0 = const()[name = tensor("op_8660_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_8660_squeeze_mask_0 = const()[name = tensor("op_8660_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor cache20_to_fp16_dtype_0 = const()[name = tensor("cache20_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor cache20_to_fp16 = cast(dtype = cache20_to_fp16_dtype_0, x = cache20)[name = tensor("cast_559")]; + tensor var_8660_cast_fp16 = slice_by_index(begin = var_8660_begin_0, end = var_8660_end_0, end_mask = var_8660_end_mask_0, squeeze_mask = var_8660_squeeze_mask_0, x = cache20_to_fp16)[name = tensor("op_8660_cast_fp16")]; + tensor var_8662_axis_0 = const()[name = tensor("op_8662_axis_0"), val = tensor(1)]; + tensor var_8662_mode_0 = const()[name = tensor("op_8662_mode_0"), val = tensor("update")]; + tensor var_8662_validate_indices_0 = const()[name = tensor("op_8662_validate_indices_0"), val = tensor(false)]; + tensor var_8662_cast_fp16 = scatter_along_axis(axis = var_8662_axis_0, data = var_8660_cast_fp16, indices = write_indices_41, mode = var_8662_mode_0, updates = k_83_cast_fp16, validate_indices = var_8662_validate_indices_0)[name = tensor("op_8662_cast_fp16")]; + tensor concat_162 = const()[name = tensor("concat_162"), val = tensor([0, 0, 0, 0, 0])]; + tensor concat_163 = const()[name = tensor("concat_163"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_41_internal_tensor_assign_1_stride_0 = const()[name = tensor("new_cache_41_internal_tensor_assign_1_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_41_internal_tensor_assign_1_begin_mask_0 = const()[name = tensor("new_cache_41_internal_tensor_assign_1_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_41_internal_tensor_assign_1_end_mask_0 = const()[name = tensor("new_cache_41_internal_tensor_assign_1_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_41_internal_tensor_assign_1_squeeze_mask_0 = const()[name = tensor("new_cache_41_internal_tensor_assign_1_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_86 = const()[name = tensor("shape_86"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_40 = const()[name = tensor("reduce_prod_40"), val = tensor(1048576)]; + tensor range_1d_40_start_0 = const()[name = tensor("range_1d_40_start_0"), val = tensor(0)]; + tensor range_1d_40_step_0 = const()[name = tensor("range_1d_40_step_0"), val = tensor(1)]; + tensor range_1d_40 = range_1d(end = reduce_prod_40, start = range_1d_40_start_0, step = range_1d_40_step_0)[name = tensor("range_1d_40")]; + tensor reshape_200 = reshape(shape = shape_86, x = range_1d_40)[name = tensor("reshape_200")]; + tensor slice_by_index_40 = slice_by_index(begin = concat_162, begin_mask = new_cache_41_internal_tensor_assign_1_begin_mask_0, end = concat_163, end_mask = new_cache_41_internal_tensor_assign_1_end_mask_0, squeeze_mask = new_cache_41_internal_tensor_assign_1_squeeze_mask_0, stride = new_cache_41_internal_tensor_assign_1_stride_0, x = reshape_200)[name = tensor("slice_by_index_40")]; + tensor reshape_201_shape_0 = const()[name = tensor("reshape_201_shape_0"), val = tensor([-1])]; + tensor reshape_201 = reshape(shape = reshape_201_shape_0, x = slice_by_index_40)[name = tensor("reshape_201")]; + tensor reshape_202_shape_0 = const()[name = tensor("reshape_202_shape_0"), val = tensor([-1])]; + tensor reshape_202_cast_fp16 = reshape(shape = reshape_202_shape_0, x = var_8662_cast_fp16)[name = tensor("reshape_202_cast_fp16")]; + tensor reshape_203_shape_0 = const()[name = tensor("reshape_203_shape_0"), val = tensor([-1])]; + tensor reshape_203_cast_fp16 = reshape(shape = reshape_203_shape_0, x = cache20_to_fp16)[name = tensor("reshape_203_cast_fp16")]; + tensor scatter_40_mode_0 = const()[name = tensor("scatter_40_mode_0"), val = tensor("update")]; + tensor scatter_40_axis_0 = const()[name = tensor("scatter_40_axis_0"), val = tensor(0)]; + tensor scatter_40_validate_indices_0 = const()[name = tensor("scatter_40_validate_indices_0"), val = tensor(false)]; + tensor scatter_40_cast_fp16 = scatter(axis = scatter_40_axis_0, data = reshape_203_cast_fp16, indices = reshape_201, mode = scatter_40_mode_0, updates = reshape_202_cast_fp16, validate_indices = scatter_40_validate_indices_0)[name = tensor("scatter_40_cast_fp16")]; + tensor reshape_204_cast_fp16 = reshape(shape = shape_86, x = scatter_40_cast_fp16)[name = tensor("reshape_204_cast_fp16")]; + tensor var_8670_begin_0 = const()[name = tensor("op_8670_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor var_8670_end_0 = const()[name = tensor("op_8670_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor var_8670_end_mask_0 = const()[name = tensor("op_8670_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_8670_squeeze_mask_0 = const()[name = tensor("op_8670_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor var_8670_cast_fp16 = slice_by_index(begin = var_8670_begin_0, end = var_8670_end_0, end_mask = var_8670_end_mask_0, squeeze_mask = var_8670_squeeze_mask_0, x = reshape_204_cast_fp16)[name = tensor("op_8670_cast_fp16")]; + tensor var_8672_axis_0 = const()[name = tensor("op_8672_axis_0"), val = tensor(1)]; + tensor var_8672_mode_0 = const()[name = tensor("op_8672_mode_0"), val = tensor("update")]; + tensor var_8672_validate_indices_0 = const()[name = tensor("op_8672_validate_indices_0"), val = tensor(false)]; + tensor var_8672_cast_fp16 = scatter_along_axis(axis = var_8672_axis_0, data = var_8670_cast_fp16, indices = write_indices_41, mode = var_8672_mode_0, updates = v_41_cast_fp16, validate_indices = var_8672_validate_indices_0)[name = tensor("op_8672_cast_fp16")]; + tensor concat_164 = const()[name = tensor("concat_164"), val = tensor([1, 0, 0, 0, 0])]; + tensor concat_165 = const()[name = tensor("concat_165"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_41_internal_tensor_assign_2_stride_0 = const()[name = tensor("new_cache_41_internal_tensor_assign_2_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_41_internal_tensor_assign_2_begin_mask_0 = const()[name = tensor("new_cache_41_internal_tensor_assign_2_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_41_internal_tensor_assign_2_end_mask_0 = const()[name = tensor("new_cache_41_internal_tensor_assign_2_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_41_internal_tensor_assign_2_squeeze_mask_0 = const()[name = tensor("new_cache_41_internal_tensor_assign_2_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_87 = const()[name = tensor("shape_87"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_41 = const()[name = tensor("reduce_prod_41"), val = tensor(1048576)]; + tensor range_1d_41_start_0 = const()[name = tensor("range_1d_41_start_0"), val = tensor(0)]; + tensor range_1d_41_step_0 = const()[name = tensor("range_1d_41_step_0"), val = tensor(1)]; + tensor range_1d_41 = range_1d(end = reduce_prod_41, start = range_1d_41_start_0, step = range_1d_41_step_0)[name = tensor("range_1d_41")]; + tensor reshape_205 = reshape(shape = shape_87, x = range_1d_41)[name = tensor("reshape_205")]; + tensor slice_by_index_41 = slice_by_index(begin = concat_164, begin_mask = new_cache_41_internal_tensor_assign_2_begin_mask_0, end = concat_165, end_mask = new_cache_41_internal_tensor_assign_2_end_mask_0, squeeze_mask = new_cache_41_internal_tensor_assign_2_squeeze_mask_0, stride = new_cache_41_internal_tensor_assign_2_stride_0, x = reshape_205)[name = tensor("slice_by_index_41")]; + tensor reshape_206_shape_0 = const()[name = tensor("reshape_206_shape_0"), val = tensor([-1])]; + tensor reshape_206 = reshape(shape = reshape_206_shape_0, x = slice_by_index_41)[name = tensor("reshape_206")]; + tensor reshape_207_shape_0 = const()[name = tensor("reshape_207_shape_0"), val = tensor([-1])]; + tensor reshape_207_cast_fp16 = reshape(shape = reshape_207_shape_0, x = var_8672_cast_fp16)[name = tensor("reshape_207_cast_fp16")]; + tensor reshape_208_shape_0 = const()[name = tensor("reshape_208_shape_0"), val = tensor([-1])]; + tensor reshape_208_cast_fp16 = reshape(shape = reshape_208_shape_0, x = reshape_204_cast_fp16)[name = tensor("reshape_208_cast_fp16")]; + tensor scatter_41_mode_0 = const()[name = tensor("scatter_41_mode_0"), val = tensor("update")]; + tensor scatter_41_axis_0 = const()[name = tensor("scatter_41_axis_0"), val = tensor(0)]; + tensor scatter_41_validate_indices_0 = const()[name = tensor("scatter_41_validate_indices_0"), val = tensor(false)]; + tensor scatter_41_cast_fp16 = scatter(axis = scatter_41_axis_0, data = reshape_208_cast_fp16, indices = reshape_206, mode = scatter_41_mode_0, updates = reshape_207_cast_fp16, validate_indices = scatter_41_validate_indices_0)[name = tensor("scatter_41_cast_fp16")]; + tensor reshape_209_cast_fp16 = reshape(shape = shape_87, x = scatter_41_cast_fp16)[name = tensor("reshape_209_cast_fp16")]; + tensor reshape_209_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("reshape_209_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor keys_121_begin_0 = const()[name = tensor("keys_121_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor keys_121_end_0 = const()[name = tensor("keys_121_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor keys_121_end_mask_0 = const()[name = tensor("keys_121_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor keys_121_squeeze_mask_0 = const()[name = tensor("keys_121_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor keys_121_cast_fp16 = slice_by_index(begin = keys_121_begin_0, end = keys_121_end_0, end_mask = keys_121_end_mask_0, squeeze_mask = keys_121_squeeze_mask_0, x = reshape_209_cast_fp16)[name = tensor("keys_121_cast_fp16")]; + tensor values_121_begin_0 = const()[name = tensor("values_121_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor values_121_end_0 = const()[name = tensor("values_121_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor values_121_end_mask_0 = const()[name = tensor("values_121_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor values_121_squeeze_mask_0 = const()[name = tensor("values_121_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor values_121_cast_fp16 = slice_by_index(begin = values_121_begin_0, end = values_121_end_0, end_mask = values_121_end_mask_0, squeeze_mask = values_121_squeeze_mask_0, x = reshape_209_cast_fp16)[name = tensor("values_121_cast_fp16")]; + tensor var_8684_cast_fp16 = not_equal(x = keys_121_cast_fp16, y = keys_121_cast_fp16)[name = tensor("op_8684_cast_fp16")]; + tensor keys_123_cast_fp16 = select(a = var_470_to_fp16, b = keys_121_cast_fp16, cond = var_8684_cast_fp16)[name = tensor("keys_123_cast_fp16")]; + tensor var_8692_cast_fp16 = not_equal(x = values_121_cast_fp16, y = values_121_cast_fp16)[name = tensor("op_8692_cast_fp16")]; + tensor values_123_cast_fp16 = select(a = var_470_to_fp16, b = values_121_cast_fp16, cond = var_8692_cast_fp16)[name = tensor("values_123_cast_fp16")]; + tensor var_8716 = const()[name = tensor("op_8716"), val = tensor([0, 2, 1, 3])]; + tensor var_8732 = const()[name = tensor("op_8732"), val = tensor([1, 1, 1])]; + tensor var_8733_cast_fp16 = reshape(shape = var_8732, x = position20_to_fp16)[name = tensor("op_8733_cast_fp16")]; + tensor q_positions_41_cast_fp16 = add(x = var_8733_cast_fp16, y = q_offsets_1_promoted_to_fp16)[name = tensor("q_positions_41_cast_fp16")]; + tensor valid_end_41_cast_fp16 = add(x = var_8733_cast_fp16, y = var_536_cast_fp16)[name = tensor("valid_end_41_cast_fp16")]; + tensor valid_mask_41_cast_fp16 = less(x = k_positions_1_promoted_to_fp16, y = valid_end_41_cast_fp16)[name = tensor("valid_mask_41_cast_fp16")]; + tensor causal_mask_41_cast_fp16 = less_equal(x = k_positions_1_promoted_to_fp16, y = q_positions_41_cast_fp16)[name = tensor("causal_mask_41_cast_fp16")]; + tensor var_8761 = logical_and(x = valid_mask_41_cast_fp16, y = causal_mask_41_cast_fp16)[name = tensor("op_8761")]; + tensor attn_mask_41_axes_0 = const()[name = tensor("attn_mask_41_axes_0"), val = tensor([1])]; + tensor attn_mask_41 = expand_dims(axes = attn_mask_41_axes_0, x = var_8761)[name = tensor("attn_mask_41")]; + tensor var_8775_transpose_x_0 = const()[name = tensor("op_8775_transpose_x_0"), val = tensor(false)]; + tensor var_8775_transpose_y_0 = const()[name = tensor("op_8775_transpose_y_0"), val = tensor(false)]; + tensor transpose_109_perm_0 = const()[name = tensor("transpose_109_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_110_perm_0 = const()[name = tensor("transpose_110_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_110 = transpose(perm = transpose_110_perm_0, x = keys_123_cast_fp16)[name = tensor("transpose_124")]; + tensor transpose_109 = transpose(perm = transpose_109_perm_0, x = q_123_cast_fp16)[name = tensor("transpose_125")]; + tensor var_8775_cast_fp16 = matmul(transpose_x = var_8775_transpose_x_0, transpose_y = var_8775_transpose_y_0, x = transpose_109, y = transpose_110)[name = tensor("op_8775_cast_fp16")]; + tensor var_8769_to_fp16 = const()[name = tensor("op_8769_to_fp16"), val = tensor([0x1p-3])]; + tensor attn_weights_121_cast_fp16 = mul(x = var_8775_cast_fp16, y = var_8769_to_fp16)[name = tensor("attn_weights_121_cast_fp16")]; + tensor var_8777 = logical_not(x = attn_mask_41)[name = tensor("op_8777")]; + tensor var_8778_to_fp16 = const()[name = tensor("op_8778_to_fp16"), val = tensor(-inf)]; + tensor attn_weights_123_cast_fp16 = select(a = var_8778_to_fp16, b = attn_weights_121_cast_fp16, cond = var_8777)[name = tensor("attn_weights_123_cast_fp16")]; + tensor var_8780 = const()[name = tensor("op_8780"), val = tensor(-1)]; + tensor attn_weights_125_cast_fp16 = softmax(axis = var_8780, x = attn_weights_123_cast_fp16)[name = tensor("attn_weights_125_cast_fp16")]; + tensor attn_output_41_transpose_x_0 = const()[name = tensor("attn_output_41_transpose_x_0"), val = tensor(false)]; + tensor attn_output_41_transpose_y_0 = const()[name = tensor("attn_output_41_transpose_y_0"), val = tensor(false)]; + tensor values_125_cast_fp16 = transpose(perm = var_8716, x = values_123_cast_fp16)[name = tensor("transpose_126")]; + tensor attn_output_41_cast_fp16 = matmul(transpose_x = attn_output_41_transpose_x_0, transpose_y = attn_output_41_transpose_y_0, x = attn_weights_125_cast_fp16, y = values_125_cast_fp16)[name = tensor("attn_output_41_cast_fp16")]; + tensor var_8788 = const()[name = tensor("op_8788"), val = tensor([0, 2, 1, 3])]; + tensor var_8793 = const()[name = tensor("op_8793"), val = tensor([1, 256, 1024])]; + tensor var_8789_cast_fp16 = transpose(perm = var_8788, x = attn_output_41_cast_fp16)[name = tensor("transpose_123")]; + tensor input_245_cast_fp16 = reshape(shape = var_8793, x = var_8789_cast_fp16)[name = tensor("input_245_cast_fp16")]; + tensor attn20_out_proj_weight_to_fp16 = const()[name = tensor("attn20_out_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(510857088)))]; + tensor linear_81_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = attn20_out_proj_weight_to_fp16, x = input_245_cast_fp16)[name = tensor("linear_81_cast_fp16")]; + tensor var_8802_cast_fp16 = add(x = position20_to_fp16, y = valid_len_to_fp16)[name = tensor("op_8802_cast_fp16")]; + tensor var_8802_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_8802_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor input_247_cast_fp16 = add(x = input_241_cast_fp16, y = linear_81_cast_fp16)[name = tensor("input_247_cast_fp16")]; + tensor input_249_axes_0 = const()[name = tensor("input_249_axes_0"), val = tensor([-1])]; + tensor norm20_2_weight_to_fp16 = const()[name = tensor("norm20_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(512954304)))]; + tensor norm20_2_bias_to_fp16 = const()[name = tensor("norm20_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(512956416)))]; + tensor var_8806_to_fp16 = const()[name = tensor("op_8806_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_249_cast_fp16 = layer_norm(axes = input_249_axes_0, beta = norm20_2_bias_to_fp16, epsilon = var_8806_to_fp16, gamma = norm20_2_weight_to_fp16, x = input_247_cast_fp16)[name = tensor("input_249_cast_fp16")]; + tensor linear20_1_weight_to_fp16 = const()[name = tensor("linear20_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(512958528)))]; + tensor linear_82_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = linear20_1_weight_to_fp16, x = input_249_cast_fp16)[name = tensor("linear_82_cast_fp16")]; + tensor input_251_mode_0 = const()[name = tensor("input_251_mode_0"), val = tensor("EXACT")]; + tensor input_251_cast_fp16 = gelu(mode = input_251_mode_0, x = linear_82_cast_fp16)[name = tensor("input_251_cast_fp16")]; + tensor linear20_2_weight_to_fp16 = const()[name = tensor("linear20_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(521347200)))]; + tensor linear_83_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = linear20_2_weight_to_fp16, x = input_251_cast_fp16)[name = tensor("linear_83_cast_fp16")]; + tensor input_253_cast_fp16 = add(x = input_247_cast_fp16, y = linear_83_cast_fp16)[name = tensor("input_253_cast_fp16")]; + tensor input_255_axes_0 = const()[name = tensor("input_255_axes_0"), val = tensor([-1])]; + tensor norm21_1_weight_to_fp16 = const()[name = tensor("norm21_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(529735872)))]; + tensor norm21_1_bias_to_fp16 = const()[name = tensor("norm21_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(529737984)))]; + tensor var_8823_to_fp16 = const()[name = tensor("op_8823_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_255_cast_fp16 = layer_norm(axes = input_255_axes_0, beta = norm21_1_bias_to_fp16, epsilon = var_8823_to_fp16, gamma = norm21_1_weight_to_fp16, x = input_253_cast_fp16)[name = tensor("input_255_cast_fp16")]; + tensor attn21_in_proj_weight_to_fp16 = const()[name = tensor("attn21_in_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(529740096)))]; + tensor linear_84_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = attn21_in_proj_weight_to_fp16, x = input_255_cast_fp16)[name = tensor("linear_84_cast_fp16")]; + tensor var_8837 = const()[name = tensor("op_8837"), val = tensor([1, 256, 3, 16, 64])]; + tensor qkv_43_cast_fp16 = reshape(shape = var_8837, x = linear_84_cast_fp16)[name = tensor("qkv_43_cast_fp16")]; + tensor q_127_begin_0 = const()[name = tensor("q_127_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor q_127_end_0 = const()[name = tensor("q_127_end_0"), val = tensor([1, 256, 1, 16, 64])]; + tensor q_127_end_mask_0 = const()[name = tensor("q_127_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor q_127_squeeze_mask_0 = const()[name = tensor("q_127_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor q_127_cast_fp16 = slice_by_index(begin = q_127_begin_0, end = q_127_end_0, end_mask = q_127_end_mask_0, squeeze_mask = q_127_squeeze_mask_0, x = qkv_43_cast_fp16)[name = tensor("q_127_cast_fp16")]; + tensor k_85_begin_0 = const()[name = tensor("k_85_begin_0"), val = tensor([0, 0, 1, 0, 0])]; + tensor k_85_end_0 = const()[name = tensor("k_85_end_0"), val = tensor([1, 256, 2, 16, 64])]; + tensor k_85_end_mask_0 = const()[name = tensor("k_85_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor k_85_squeeze_mask_0 = const()[name = tensor("k_85_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor k_85_cast_fp16 = slice_by_index(begin = k_85_begin_0, end = k_85_end_0, end_mask = k_85_end_mask_0, squeeze_mask = k_85_squeeze_mask_0, x = qkv_43_cast_fp16)[name = tensor("k_85_cast_fp16")]; + tensor v_43_begin_0 = const()[name = tensor("v_43_begin_0"), val = tensor([0, 0, 2, 0, 0])]; + tensor v_43_end_0 = const()[name = tensor("v_43_end_0"), val = tensor([1, 256, 3, 16, 64])]; + tensor v_43_end_mask_0 = const()[name = tensor("v_43_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor v_43_squeeze_mask_0 = const()[name = tensor("v_43_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor v_43_cast_fp16 = slice_by_index(begin = v_43_begin_0, end = v_43_end_0, end_mask = v_43_end_mask_0, squeeze_mask = v_43_squeeze_mask_0, x = qkv_43_cast_fp16)[name = tensor("v_43_cast_fp16")]; + tensor var_8895 = const()[name = tensor("op_8895"), val = tensor([1, 1])]; + tensor position21_to_fp16_dtype_0 = const()[name = tensor("position21_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor position21_to_fp16 = cast(dtype = position21_to_fp16_dtype_0, x = position21)[name = tensor("cast_556")]; + tensor var_8896_cast_fp16 = reshape(shape = var_8895, x = position21_to_fp16)[name = tensor("op_8896_cast_fp16")]; + tensor ts_129_cast_fp16 = add(x = ts_1_promoted_to_fp16, y = var_8896_cast_fp16)[name = tensor("ts_129_cast_fp16")]; + tensor var_8903 = const()[name = tensor("op_8903"), val = tensor([1, 256, 1, 1])]; + tensor ts_131_cast_fp16 = reshape(shape = var_8903, x = ts_129_cast_fp16)[name = tensor("ts_131_cast_fp16")]; + tensor var_8910 = const()[name = tensor("op_8910"), val = tensor([1, 256, 16, 32, 2])]; + tensor q_complex_43_cast_fp16 = reshape(shape = var_8910, x = q_127_cast_fp16)[name = tensor("q_complex_43_cast_fp16")]; + tensor var_8917 = const()[name = tensor("op_8917"), val = tensor([1, 256, 16, 32, 2])]; + tensor k_complex_43_cast_fp16 = reshape(shape = var_8917, x = k_85_cast_fp16)[name = tensor("k_complex_43_cast_fp16")]; + tensor var_8921_begin_0 = const()[name = tensor("op_8921_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_8921_end_0 = const()[name = tensor("op_8921_end_0"), val = tensor([1, 256, 16, 32, 1])]; + tensor var_8921_end_mask_0 = const()[name = tensor("op_8921_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_8921_squeeze_mask_0 = const()[name = tensor("op_8921_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_8921_cast_fp16 = slice_by_index(begin = var_8921_begin_0, end = var_8921_end_0, end_mask = var_8921_end_mask_0, squeeze_mask = var_8921_squeeze_mask_0, x = q_complex_43_cast_fp16)[name = tensor("op_8921_cast_fp16")]; + tensor var_8929_begin_0 = const()[name = tensor("op_8929_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_8929_end_0 = const()[name = tensor("op_8929_end_0"), val = tensor([1, 256, 16, 32, 2])]; + tensor var_8929_end_mask_0 = const()[name = tensor("op_8929_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_8929_squeeze_mask_0 = const()[name = tensor("op_8929_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_8929_cast_fp16 = slice_by_index(begin = var_8929_begin_0, end = var_8929_end_0, end_mask = var_8929_end_mask_0, squeeze_mask = var_8929_squeeze_mask_0, x = q_complex_43_cast_fp16)[name = tensor("op_8929_cast_fp16")]; + tensor var_8937_begin_0 = const()[name = tensor("op_8937_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_8937_end_0 = const()[name = tensor("op_8937_end_0"), val = tensor([1, 256, 16, 32, 1])]; + tensor var_8937_end_mask_0 = const()[name = tensor("op_8937_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_8937_squeeze_mask_0 = const()[name = tensor("op_8937_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_8937_cast_fp16 = slice_by_index(begin = var_8937_begin_0, end = var_8937_end_0, end_mask = var_8937_end_mask_0, squeeze_mask = var_8937_squeeze_mask_0, x = k_complex_43_cast_fp16)[name = tensor("op_8937_cast_fp16")]; + tensor var_8945_begin_0 = const()[name = tensor("op_8945_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_8945_end_0 = const()[name = tensor("op_8945_end_0"), val = tensor([1, 256, 16, 32, 2])]; + tensor var_8945_end_mask_0 = const()[name = tensor("op_8945_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_8945_squeeze_mask_0 = const()[name = tensor("op_8945_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_8945_cast_fp16 = slice_by_index(begin = var_8945_begin_0, end = var_8945_end_0, end_mask = var_8945_end_mask_0, squeeze_mask = var_8945_squeeze_mask_0, x = k_complex_43_cast_fp16)[name = tensor("op_8945_cast_fp16")]; + tensor freqs_43_to_fp16 = const()[name = tensor("freqs_43_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(536031616)))]; + tensor var_8951_cast_fp16 = mul(x = freqs_43_to_fp16, y = ts_131_cast_fp16)[name = tensor("op_8951_cast_fp16")]; + tensor rotr_43_cast_fp16 = cos(x = var_8951_cast_fp16)[name = tensor("rotr_43_cast_fp16")]; + tensor roti_43_cast_fp16 = sin(x = var_8951_cast_fp16)[name = tensor("roti_43_cast_fp16")]; + tensor var_8955_cast_fp16 = mul(x = var_8921_cast_fp16, y = rotr_43_cast_fp16)[name = tensor("op_8955_cast_fp16")]; + tensor var_8956_cast_fp16 = mul(x = var_8929_cast_fp16, y = roti_43_cast_fp16)[name = tensor("op_8956_cast_fp16")]; + tensor qor_85_cast_fp16 = sub(x = var_8955_cast_fp16, y = var_8956_cast_fp16)[name = tensor("qor_85_cast_fp16")]; + tensor var_8959_cast_fp16 = mul(x = var_8921_cast_fp16, y = roti_43_cast_fp16)[name = tensor("op_8959_cast_fp16")]; + tensor var_8960_cast_fp16 = mul(x = var_8929_cast_fp16, y = rotr_43_cast_fp16)[name = tensor("op_8960_cast_fp16")]; + tensor qoi_85_cast_fp16 = add(x = var_8959_cast_fp16, y = var_8960_cast_fp16)[name = tensor("qoi_85_cast_fp16")]; + tensor var_8963_cast_fp16 = mul(x = var_8937_cast_fp16, y = rotr_43_cast_fp16)[name = tensor("op_8963_cast_fp16")]; + tensor var_8964_cast_fp16 = mul(x = var_8945_cast_fp16, y = roti_43_cast_fp16)[name = tensor("op_8964_cast_fp16")]; + tensor kor_85_cast_fp16 = sub(x = var_8963_cast_fp16, y = var_8964_cast_fp16)[name = tensor("kor_85_cast_fp16")]; + tensor var_8967_cast_fp16 = mul(x = var_8937_cast_fp16, y = roti_43_cast_fp16)[name = tensor("op_8967_cast_fp16")]; + tensor var_8968_cast_fp16 = mul(x = var_8945_cast_fp16, y = rotr_43_cast_fp16)[name = tensor("op_8968_cast_fp16")]; + tensor koi_85_cast_fp16 = add(x = var_8967_cast_fp16, y = var_8968_cast_fp16)[name = tensor("koi_85_cast_fp16")]; + tensor qo_43_axis_0 = const()[name = tensor("qo_43_axis_0"), val = tensor(-1)]; + tensor qo_43_cast_fp16 = stack(axis = qo_43_axis_0, values = (qor_85_cast_fp16, qoi_85_cast_fp16))[name = tensor("qo_43_cast_fp16")]; + tensor ko_43_axis_0 = const()[name = tensor("ko_43_axis_0"), val = tensor(-1)]; + tensor ko_43_cast_fp16 = stack(axis = ko_43_axis_0, values = (kor_85_cast_fp16, koi_85_cast_fp16))[name = tensor("ko_43_cast_fp16")]; + tensor var_9001 = const()[name = tensor("op_9001"), val = tensor([1, 256, 16, 64])]; + tensor q_129_cast_fp16 = reshape(shape = var_9001, x = qo_43_cast_fp16)[name = tensor("q_129_cast_fp16")]; + tensor var_9007 = const()[name = tensor("op_9007"), val = tensor([1, 256, 16, 64])]; + tensor k_87_cast_fp16 = reshape(shape = var_9007, x = ko_43_cast_fp16)[name = tensor("k_87_cast_fp16")]; + tensor write_indices_float_127_cast_fp16 = add(x = var_8896_cast_fp16, y = var_393_promoted_to_fp16)[name = tensor("write_indices_float_127_cast_fp16")]; + tensor _inversed_9028_y_0_to_fp16 = const()[name = tensor("_inversed_9028_y_0_to_fp16"), val = tensor(0x1p-9)]; + tensor _inversed_9028_cast_fp16 = mul(x = write_indices_float_127_cast_fp16, y = _inversed_9028_y_0_to_fp16)[name = tensor("_inversed_9028_cast_fp16")]; + tensor var_9029_cast_fp16 = floor(x = _inversed_9028_cast_fp16)[name = tensor("op_9029_cast_fp16")]; + tensor var_9030_to_fp16 = const()[name = tensor("op_9030_to_fp16"), val = tensor(0x1p+9)]; + tensor var_9031_cast_fp16 = mul(x = var_9029_cast_fp16, y = var_9030_to_fp16)[name = tensor("op_9031_cast_fp16")]; + tensor write_indices_float_129_cast_fp16 = sub(x = write_indices_float_127_cast_fp16, y = var_9031_cast_fp16)[name = tensor("write_indices_float_129_cast_fp16")]; + tensor var_9049_value_0_to_fp16 = const()[name = tensor("op_9049_value_0_to_fp16"), val = tensor(0x1.ffp+8)]; + tensor var_9049_cast_fp16 = fill_like(ref_tensor = write_indices_float_129_cast_fp16, value = var_9049_value_0_to_fp16)[name = tensor("op_9049_cast_fp16")]; + tensor write_indices_float_131_cast_fp16 = select(a = write_indices_float_129_cast_fp16, b = var_9049_cast_fp16, cond = token_valid_1_cast_fp16)[name = tensor("write_indices_float_131_cast_fp16")]; + tensor var_9055_dtype_0 = const()[name = tensor("op_9055_dtype_0"), val = tensor("int32")]; + tensor var_9060 = const()[name = tensor("op_9060"), val = tensor([1, 256, 1, 1])]; + tensor write_indices_float_131_cast_fp16_to_int32 = cast(dtype = var_9055_dtype_0, x = write_indices_float_131_cast_fp16)[name = tensor("cast_555")]; + tensor var_9061 = reshape(shape = var_9060, x = write_indices_float_131_cast_fp16_to_int32)[name = tensor("op_9061")]; + tensor write_indices_43_reps_0 = const()[name = tensor("write_indices_43_reps_0"), val = tensor([1, 1, 16, 64])]; + tensor write_indices_43 = tile(reps = write_indices_43_reps_0, x = var_9061)[name = tensor("write_indices_43")]; + tensor var_9071_begin_0 = const()[name = tensor("op_9071_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_9071_end_0 = const()[name = tensor("op_9071_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor var_9071_end_mask_0 = const()[name = tensor("op_9071_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_9071_squeeze_mask_0 = const()[name = tensor("op_9071_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor cache21_to_fp16_dtype_0 = const()[name = tensor("cache21_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor cache21_to_fp16 = cast(dtype = cache21_to_fp16_dtype_0, x = cache21)[name = tensor("cast_554")]; + tensor var_9071_cast_fp16 = slice_by_index(begin = var_9071_begin_0, end = var_9071_end_0, end_mask = var_9071_end_mask_0, squeeze_mask = var_9071_squeeze_mask_0, x = cache21_to_fp16)[name = tensor("op_9071_cast_fp16")]; + tensor var_9073_axis_0 = const()[name = tensor("op_9073_axis_0"), val = tensor(1)]; + tensor var_9073_mode_0 = const()[name = tensor("op_9073_mode_0"), val = tensor("update")]; + tensor var_9073_validate_indices_0 = const()[name = tensor("op_9073_validate_indices_0"), val = tensor(false)]; + tensor var_9073_cast_fp16 = scatter_along_axis(axis = var_9073_axis_0, data = var_9071_cast_fp16, indices = write_indices_43, mode = var_9073_mode_0, updates = k_87_cast_fp16, validate_indices = var_9073_validate_indices_0)[name = tensor("op_9073_cast_fp16")]; + tensor concat_170 = const()[name = tensor("concat_170"), val = tensor([0, 0, 0, 0, 0])]; + tensor concat_171 = const()[name = tensor("concat_171"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_43_internal_tensor_assign_1_stride_0 = const()[name = tensor("new_cache_43_internal_tensor_assign_1_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_43_internal_tensor_assign_1_begin_mask_0 = const()[name = tensor("new_cache_43_internal_tensor_assign_1_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_43_internal_tensor_assign_1_end_mask_0 = const()[name = tensor("new_cache_43_internal_tensor_assign_1_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_43_internal_tensor_assign_1_squeeze_mask_0 = const()[name = tensor("new_cache_43_internal_tensor_assign_1_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_88 = const()[name = tensor("shape_88"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_42 = const()[name = tensor("reduce_prod_42"), val = tensor(1048576)]; + tensor range_1d_42_start_0 = const()[name = tensor("range_1d_42_start_0"), val = tensor(0)]; + tensor range_1d_42_step_0 = const()[name = tensor("range_1d_42_step_0"), val = tensor(1)]; + tensor range_1d_42 = range_1d(end = reduce_prod_42, start = range_1d_42_start_0, step = range_1d_42_step_0)[name = tensor("range_1d_42")]; + tensor reshape_210 = reshape(shape = shape_88, x = range_1d_42)[name = tensor("reshape_210")]; + tensor slice_by_index_42 = slice_by_index(begin = concat_170, begin_mask = new_cache_43_internal_tensor_assign_1_begin_mask_0, end = concat_171, end_mask = new_cache_43_internal_tensor_assign_1_end_mask_0, squeeze_mask = new_cache_43_internal_tensor_assign_1_squeeze_mask_0, stride = new_cache_43_internal_tensor_assign_1_stride_0, x = reshape_210)[name = tensor("slice_by_index_42")]; + tensor reshape_211_shape_0 = const()[name = tensor("reshape_211_shape_0"), val = tensor([-1])]; + tensor reshape_211 = reshape(shape = reshape_211_shape_0, x = slice_by_index_42)[name = tensor("reshape_211")]; + tensor reshape_212_shape_0 = const()[name = tensor("reshape_212_shape_0"), val = tensor([-1])]; + tensor reshape_212_cast_fp16 = reshape(shape = reshape_212_shape_0, x = var_9073_cast_fp16)[name = tensor("reshape_212_cast_fp16")]; + tensor reshape_213_shape_0 = const()[name = tensor("reshape_213_shape_0"), val = tensor([-1])]; + tensor reshape_213_cast_fp16 = reshape(shape = reshape_213_shape_0, x = cache21_to_fp16)[name = tensor("reshape_213_cast_fp16")]; + tensor scatter_42_mode_0 = const()[name = tensor("scatter_42_mode_0"), val = tensor("update")]; + tensor scatter_42_axis_0 = const()[name = tensor("scatter_42_axis_0"), val = tensor(0)]; + tensor scatter_42_validate_indices_0 = const()[name = tensor("scatter_42_validate_indices_0"), val = tensor(false)]; + tensor scatter_42_cast_fp16 = scatter(axis = scatter_42_axis_0, data = reshape_213_cast_fp16, indices = reshape_211, mode = scatter_42_mode_0, updates = reshape_212_cast_fp16, validate_indices = scatter_42_validate_indices_0)[name = tensor("scatter_42_cast_fp16")]; + tensor reshape_214_cast_fp16 = reshape(shape = shape_88, x = scatter_42_cast_fp16)[name = tensor("reshape_214_cast_fp16")]; + tensor var_9081_begin_0 = const()[name = tensor("op_9081_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor var_9081_end_0 = const()[name = tensor("op_9081_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor var_9081_end_mask_0 = const()[name = tensor("op_9081_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_9081_squeeze_mask_0 = const()[name = tensor("op_9081_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor var_9081_cast_fp16 = slice_by_index(begin = var_9081_begin_0, end = var_9081_end_0, end_mask = var_9081_end_mask_0, squeeze_mask = var_9081_squeeze_mask_0, x = reshape_214_cast_fp16)[name = tensor("op_9081_cast_fp16")]; + tensor var_9083_axis_0 = const()[name = tensor("op_9083_axis_0"), val = tensor(1)]; + tensor var_9083_mode_0 = const()[name = tensor("op_9083_mode_0"), val = tensor("update")]; + tensor var_9083_validate_indices_0 = const()[name = tensor("op_9083_validate_indices_0"), val = tensor(false)]; + tensor var_9083_cast_fp16 = scatter_along_axis(axis = var_9083_axis_0, data = var_9081_cast_fp16, indices = write_indices_43, mode = var_9083_mode_0, updates = v_43_cast_fp16, validate_indices = var_9083_validate_indices_0)[name = tensor("op_9083_cast_fp16")]; + tensor concat_172 = const()[name = tensor("concat_172"), val = tensor([1, 0, 0, 0, 0])]; + tensor concat_173 = const()[name = tensor("concat_173"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_43_internal_tensor_assign_2_stride_0 = const()[name = tensor("new_cache_43_internal_tensor_assign_2_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_43_internal_tensor_assign_2_begin_mask_0 = const()[name = tensor("new_cache_43_internal_tensor_assign_2_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_43_internal_tensor_assign_2_end_mask_0 = const()[name = tensor("new_cache_43_internal_tensor_assign_2_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_43_internal_tensor_assign_2_squeeze_mask_0 = const()[name = tensor("new_cache_43_internal_tensor_assign_2_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_89 = const()[name = tensor("shape_89"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_43 = const()[name = tensor("reduce_prod_43"), val = tensor(1048576)]; + tensor range_1d_43_start_0 = const()[name = tensor("range_1d_43_start_0"), val = tensor(0)]; + tensor range_1d_43_step_0 = const()[name = tensor("range_1d_43_step_0"), val = tensor(1)]; + tensor range_1d_43 = range_1d(end = reduce_prod_43, start = range_1d_43_start_0, step = range_1d_43_step_0)[name = tensor("range_1d_43")]; + tensor reshape_215 = reshape(shape = shape_89, x = range_1d_43)[name = tensor("reshape_215")]; + tensor slice_by_index_43 = slice_by_index(begin = concat_172, begin_mask = new_cache_43_internal_tensor_assign_2_begin_mask_0, end = concat_173, end_mask = new_cache_43_internal_tensor_assign_2_end_mask_0, squeeze_mask = new_cache_43_internal_tensor_assign_2_squeeze_mask_0, stride = new_cache_43_internal_tensor_assign_2_stride_0, x = reshape_215)[name = tensor("slice_by_index_43")]; + tensor reshape_216_shape_0 = const()[name = tensor("reshape_216_shape_0"), val = tensor([-1])]; + tensor reshape_216 = reshape(shape = reshape_216_shape_0, x = slice_by_index_43)[name = tensor("reshape_216")]; + tensor reshape_217_shape_0 = const()[name = tensor("reshape_217_shape_0"), val = tensor([-1])]; + tensor reshape_217_cast_fp16 = reshape(shape = reshape_217_shape_0, x = var_9083_cast_fp16)[name = tensor("reshape_217_cast_fp16")]; + tensor reshape_218_shape_0 = const()[name = tensor("reshape_218_shape_0"), val = tensor([-1])]; + tensor reshape_218_cast_fp16 = reshape(shape = reshape_218_shape_0, x = reshape_214_cast_fp16)[name = tensor("reshape_218_cast_fp16")]; + tensor scatter_43_mode_0 = const()[name = tensor("scatter_43_mode_0"), val = tensor("update")]; + tensor scatter_43_axis_0 = const()[name = tensor("scatter_43_axis_0"), val = tensor(0)]; + tensor scatter_43_validate_indices_0 = const()[name = tensor("scatter_43_validate_indices_0"), val = tensor(false)]; + tensor scatter_43_cast_fp16 = scatter(axis = scatter_43_axis_0, data = reshape_218_cast_fp16, indices = reshape_216, mode = scatter_43_mode_0, updates = reshape_217_cast_fp16, validate_indices = scatter_43_validate_indices_0)[name = tensor("scatter_43_cast_fp16")]; + tensor reshape_219_cast_fp16 = reshape(shape = shape_89, x = scatter_43_cast_fp16)[name = tensor("reshape_219_cast_fp16")]; + tensor reshape_219_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("reshape_219_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor keys_127_begin_0 = const()[name = tensor("keys_127_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor keys_127_end_0 = const()[name = tensor("keys_127_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor keys_127_end_mask_0 = const()[name = tensor("keys_127_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor keys_127_squeeze_mask_0 = const()[name = tensor("keys_127_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor keys_127_cast_fp16 = slice_by_index(begin = keys_127_begin_0, end = keys_127_end_0, end_mask = keys_127_end_mask_0, squeeze_mask = keys_127_squeeze_mask_0, x = reshape_219_cast_fp16)[name = tensor("keys_127_cast_fp16")]; + tensor values_127_begin_0 = const()[name = tensor("values_127_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor values_127_end_0 = const()[name = tensor("values_127_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor values_127_end_mask_0 = const()[name = tensor("values_127_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor values_127_squeeze_mask_0 = const()[name = tensor("values_127_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor values_127_cast_fp16 = slice_by_index(begin = values_127_begin_0, end = values_127_end_0, end_mask = values_127_end_mask_0, squeeze_mask = values_127_squeeze_mask_0, x = reshape_219_cast_fp16)[name = tensor("values_127_cast_fp16")]; + tensor var_9095_cast_fp16 = not_equal(x = keys_127_cast_fp16, y = keys_127_cast_fp16)[name = tensor("op_9095_cast_fp16")]; + tensor keys_129_cast_fp16 = select(a = var_470_to_fp16, b = keys_127_cast_fp16, cond = var_9095_cast_fp16)[name = tensor("keys_129_cast_fp16")]; + tensor var_9103_cast_fp16 = not_equal(x = values_127_cast_fp16, y = values_127_cast_fp16)[name = tensor("op_9103_cast_fp16")]; + tensor values_129_cast_fp16 = select(a = var_470_to_fp16, b = values_127_cast_fp16, cond = var_9103_cast_fp16)[name = tensor("values_129_cast_fp16")]; + tensor var_9127 = const()[name = tensor("op_9127"), val = tensor([0, 2, 1, 3])]; + tensor var_9143 = const()[name = tensor("op_9143"), val = tensor([1, 1, 1])]; + tensor var_9144_cast_fp16 = reshape(shape = var_9143, x = position21_to_fp16)[name = tensor("op_9144_cast_fp16")]; + tensor q_positions_43_cast_fp16 = add(x = var_9144_cast_fp16, y = q_offsets_1_promoted_to_fp16)[name = tensor("q_positions_43_cast_fp16")]; + tensor valid_end_43_cast_fp16 = add(x = var_9144_cast_fp16, y = var_536_cast_fp16)[name = tensor("valid_end_43_cast_fp16")]; + tensor valid_mask_43_cast_fp16 = less(x = k_positions_1_promoted_to_fp16, y = valid_end_43_cast_fp16)[name = tensor("valid_mask_43_cast_fp16")]; + tensor causal_mask_43_cast_fp16 = less_equal(x = k_positions_1_promoted_to_fp16, y = q_positions_43_cast_fp16)[name = tensor("causal_mask_43_cast_fp16")]; + tensor var_9172 = logical_and(x = valid_mask_43_cast_fp16, y = causal_mask_43_cast_fp16)[name = tensor("op_9172")]; + tensor attn_mask_43_axes_0 = const()[name = tensor("attn_mask_43_axes_0"), val = tensor([1])]; + tensor attn_mask_43 = expand_dims(axes = attn_mask_43_axes_0, x = var_9172)[name = tensor("attn_mask_43")]; + tensor var_9186_transpose_x_0 = const()[name = tensor("op_9186_transpose_x_0"), val = tensor(false)]; + tensor var_9186_transpose_y_0 = const()[name = tensor("op_9186_transpose_y_0"), val = tensor(false)]; + tensor transpose_111_perm_0 = const()[name = tensor("transpose_111_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_112_perm_0 = const()[name = tensor("transpose_112_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_112 = transpose(perm = transpose_112_perm_0, x = keys_129_cast_fp16)[name = tensor("transpose_120")]; + tensor transpose_111 = transpose(perm = transpose_111_perm_0, x = q_129_cast_fp16)[name = tensor("transpose_121")]; + tensor var_9186_cast_fp16 = matmul(transpose_x = var_9186_transpose_x_0, transpose_y = var_9186_transpose_y_0, x = transpose_111, y = transpose_112)[name = tensor("op_9186_cast_fp16")]; + tensor var_9180_to_fp16 = const()[name = tensor("op_9180_to_fp16"), val = tensor([0x1p-3])]; + tensor attn_weights_127_cast_fp16 = mul(x = var_9186_cast_fp16, y = var_9180_to_fp16)[name = tensor("attn_weights_127_cast_fp16")]; + tensor var_9188 = logical_not(x = attn_mask_43)[name = tensor("op_9188")]; + tensor var_9189_to_fp16 = const()[name = tensor("op_9189_to_fp16"), val = tensor(-inf)]; + tensor attn_weights_129_cast_fp16 = select(a = var_9189_to_fp16, b = attn_weights_127_cast_fp16, cond = var_9188)[name = tensor("attn_weights_129_cast_fp16")]; + tensor var_9191 = const()[name = tensor("op_9191"), val = tensor(-1)]; + tensor attn_weights_131_cast_fp16 = softmax(axis = var_9191, x = attn_weights_129_cast_fp16)[name = tensor("attn_weights_131_cast_fp16")]; + tensor attn_output_43_transpose_x_0 = const()[name = tensor("attn_output_43_transpose_x_0"), val = tensor(false)]; + tensor attn_output_43_transpose_y_0 = const()[name = tensor("attn_output_43_transpose_y_0"), val = tensor(false)]; + tensor values_131_cast_fp16 = transpose(perm = var_9127, x = values_129_cast_fp16)[name = tensor("transpose_122")]; + tensor attn_output_43_cast_fp16 = matmul(transpose_x = attn_output_43_transpose_x_0, transpose_y = attn_output_43_transpose_y_0, x = attn_weights_131_cast_fp16, y = values_131_cast_fp16)[name = tensor("attn_output_43_cast_fp16")]; + tensor var_9199 = const()[name = tensor("op_9199"), val = tensor([0, 2, 1, 3])]; + tensor var_9204 = const()[name = tensor("op_9204"), val = tensor([1, 256, 1024])]; + tensor var_9200_cast_fp16 = transpose(perm = var_9199, x = attn_output_43_cast_fp16)[name = tensor("transpose_119")]; + tensor input_257_cast_fp16 = reshape(shape = var_9204, x = var_9200_cast_fp16)[name = tensor("input_257_cast_fp16")]; + tensor attn21_out_proj_weight_to_fp16 = const()[name = tensor("attn21_out_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(536031744)))]; + tensor linear_85_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = attn21_out_proj_weight_to_fp16, x = input_257_cast_fp16)[name = tensor("linear_85_cast_fp16")]; + tensor var_9213_cast_fp16 = add(x = position21_to_fp16, y = valid_len_to_fp16)[name = tensor("op_9213_cast_fp16")]; + tensor var_9213_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_9213_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor input_259_cast_fp16 = add(x = input_253_cast_fp16, y = linear_85_cast_fp16)[name = tensor("input_259_cast_fp16")]; + tensor input_261_axes_0 = const()[name = tensor("input_261_axes_0"), val = tensor([-1])]; + tensor norm21_2_weight_to_fp16 = const()[name = tensor("norm21_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(538128960)))]; + tensor norm21_2_bias_to_fp16 = const()[name = tensor("norm21_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(538131072)))]; + tensor var_9217_to_fp16 = const()[name = tensor("op_9217_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_261_cast_fp16 = layer_norm(axes = input_261_axes_0, beta = norm21_2_bias_to_fp16, epsilon = var_9217_to_fp16, gamma = norm21_2_weight_to_fp16, x = input_259_cast_fp16)[name = tensor("input_261_cast_fp16")]; + tensor linear21_1_weight_to_fp16 = const()[name = tensor("linear21_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(538133184)))]; + tensor linear_86_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = linear21_1_weight_to_fp16, x = input_261_cast_fp16)[name = tensor("linear_86_cast_fp16")]; + tensor input_263_mode_0 = const()[name = tensor("input_263_mode_0"), val = tensor("EXACT")]; + tensor input_263_cast_fp16 = gelu(mode = input_263_mode_0, x = linear_86_cast_fp16)[name = tensor("input_263_cast_fp16")]; + tensor linear21_2_weight_to_fp16 = const()[name = tensor("linear21_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(546521856)))]; + tensor linear_87_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = linear21_2_weight_to_fp16, x = input_263_cast_fp16)[name = tensor("linear_87_cast_fp16")]; + tensor input_265_cast_fp16 = add(x = input_259_cast_fp16, y = linear_87_cast_fp16)[name = tensor("input_265_cast_fp16")]; + tensor input_267_axes_0 = const()[name = tensor("input_267_axes_0"), val = tensor([-1])]; + tensor norm22_1_weight_to_fp16 = const()[name = tensor("norm22_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(554910528)))]; + tensor norm22_1_bias_to_fp16 = const()[name = tensor("norm22_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(554912640)))]; + tensor var_9234_to_fp16 = const()[name = tensor("op_9234_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_267_cast_fp16 = layer_norm(axes = input_267_axes_0, beta = norm22_1_bias_to_fp16, epsilon = var_9234_to_fp16, gamma = norm22_1_weight_to_fp16, x = input_265_cast_fp16)[name = tensor("input_267_cast_fp16")]; + tensor attn22_in_proj_weight_to_fp16 = const()[name = tensor("attn22_in_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(554914752)))]; + tensor linear_88_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = attn22_in_proj_weight_to_fp16, x = input_267_cast_fp16)[name = tensor("linear_88_cast_fp16")]; + tensor var_9248 = const()[name = tensor("op_9248"), val = tensor([1, 256, 3, 16, 64])]; + tensor qkv_45_cast_fp16 = reshape(shape = var_9248, x = linear_88_cast_fp16)[name = tensor("qkv_45_cast_fp16")]; + tensor q_133_begin_0 = const()[name = tensor("q_133_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor q_133_end_0 = const()[name = tensor("q_133_end_0"), val = tensor([1, 256, 1, 16, 64])]; + tensor q_133_end_mask_0 = const()[name = tensor("q_133_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor q_133_squeeze_mask_0 = const()[name = tensor("q_133_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor q_133_cast_fp16 = slice_by_index(begin = q_133_begin_0, end = q_133_end_0, end_mask = q_133_end_mask_0, squeeze_mask = q_133_squeeze_mask_0, x = qkv_45_cast_fp16)[name = tensor("q_133_cast_fp16")]; + tensor k_89_begin_0 = const()[name = tensor("k_89_begin_0"), val = tensor([0, 0, 1, 0, 0])]; + tensor k_89_end_0 = const()[name = tensor("k_89_end_0"), val = tensor([1, 256, 2, 16, 64])]; + tensor k_89_end_mask_0 = const()[name = tensor("k_89_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor k_89_squeeze_mask_0 = const()[name = tensor("k_89_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor k_89_cast_fp16 = slice_by_index(begin = k_89_begin_0, end = k_89_end_0, end_mask = k_89_end_mask_0, squeeze_mask = k_89_squeeze_mask_0, x = qkv_45_cast_fp16)[name = tensor("k_89_cast_fp16")]; + tensor v_45_begin_0 = const()[name = tensor("v_45_begin_0"), val = tensor([0, 0, 2, 0, 0])]; + tensor v_45_end_0 = const()[name = tensor("v_45_end_0"), val = tensor([1, 256, 3, 16, 64])]; + tensor v_45_end_mask_0 = const()[name = tensor("v_45_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor v_45_squeeze_mask_0 = const()[name = tensor("v_45_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor v_45_cast_fp16 = slice_by_index(begin = v_45_begin_0, end = v_45_end_0, end_mask = v_45_end_mask_0, squeeze_mask = v_45_squeeze_mask_0, x = qkv_45_cast_fp16)[name = tensor("v_45_cast_fp16")]; + tensor var_9306 = const()[name = tensor("op_9306"), val = tensor([1, 1])]; + tensor position22_to_fp16_dtype_0 = const()[name = tensor("position22_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor position22_to_fp16 = cast(dtype = position22_to_fp16_dtype_0, x = position22)[name = tensor("cast_551")]; + tensor var_9307_cast_fp16 = reshape(shape = var_9306, x = position22_to_fp16)[name = tensor("op_9307_cast_fp16")]; + tensor ts_135_cast_fp16 = add(x = ts_1_promoted_to_fp16, y = var_9307_cast_fp16)[name = tensor("ts_135_cast_fp16")]; + tensor var_9314 = const()[name = tensor("op_9314"), val = tensor([1, 256, 1, 1])]; + tensor ts_137_cast_fp16 = reshape(shape = var_9314, x = ts_135_cast_fp16)[name = tensor("ts_137_cast_fp16")]; + tensor var_9321 = const()[name = tensor("op_9321"), val = tensor([1, 256, 16, 32, 2])]; + tensor q_complex_45_cast_fp16 = reshape(shape = var_9321, x = q_133_cast_fp16)[name = tensor("q_complex_45_cast_fp16")]; + tensor var_9328 = const()[name = tensor("op_9328"), val = tensor([1, 256, 16, 32, 2])]; + tensor k_complex_45_cast_fp16 = reshape(shape = var_9328, x = k_89_cast_fp16)[name = tensor("k_complex_45_cast_fp16")]; + tensor var_9332_begin_0 = const()[name = tensor("op_9332_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_9332_end_0 = const()[name = tensor("op_9332_end_0"), val = tensor([1, 256, 16, 32, 1])]; + tensor var_9332_end_mask_0 = const()[name = tensor("op_9332_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_9332_squeeze_mask_0 = const()[name = tensor("op_9332_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_9332_cast_fp16 = slice_by_index(begin = var_9332_begin_0, end = var_9332_end_0, end_mask = var_9332_end_mask_0, squeeze_mask = var_9332_squeeze_mask_0, x = q_complex_45_cast_fp16)[name = tensor("op_9332_cast_fp16")]; + tensor var_9340_begin_0 = const()[name = tensor("op_9340_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_9340_end_0 = const()[name = tensor("op_9340_end_0"), val = tensor([1, 256, 16, 32, 2])]; + tensor var_9340_end_mask_0 = const()[name = tensor("op_9340_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_9340_squeeze_mask_0 = const()[name = tensor("op_9340_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_9340_cast_fp16 = slice_by_index(begin = var_9340_begin_0, end = var_9340_end_0, end_mask = var_9340_end_mask_0, squeeze_mask = var_9340_squeeze_mask_0, x = q_complex_45_cast_fp16)[name = tensor("op_9340_cast_fp16")]; + tensor var_9348_begin_0 = const()[name = tensor("op_9348_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_9348_end_0 = const()[name = tensor("op_9348_end_0"), val = tensor([1, 256, 16, 32, 1])]; + tensor var_9348_end_mask_0 = const()[name = tensor("op_9348_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_9348_squeeze_mask_0 = const()[name = tensor("op_9348_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_9348_cast_fp16 = slice_by_index(begin = var_9348_begin_0, end = var_9348_end_0, end_mask = var_9348_end_mask_0, squeeze_mask = var_9348_squeeze_mask_0, x = k_complex_45_cast_fp16)[name = tensor("op_9348_cast_fp16")]; + tensor var_9356_begin_0 = const()[name = tensor("op_9356_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_9356_end_0 = const()[name = tensor("op_9356_end_0"), val = tensor([1, 256, 16, 32, 2])]; + tensor var_9356_end_mask_0 = const()[name = tensor("op_9356_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_9356_squeeze_mask_0 = const()[name = tensor("op_9356_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_9356_cast_fp16 = slice_by_index(begin = var_9356_begin_0, end = var_9356_end_0, end_mask = var_9356_end_mask_0, squeeze_mask = var_9356_squeeze_mask_0, x = k_complex_45_cast_fp16)[name = tensor("op_9356_cast_fp16")]; + tensor freqs_45_to_fp16 = const()[name = tensor("freqs_45_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(561206272)))]; + tensor var_9362_cast_fp16 = mul(x = freqs_45_to_fp16, y = ts_137_cast_fp16)[name = tensor("op_9362_cast_fp16")]; + tensor rotr_45_cast_fp16 = cos(x = var_9362_cast_fp16)[name = tensor("rotr_45_cast_fp16")]; + tensor roti_45_cast_fp16 = sin(x = var_9362_cast_fp16)[name = tensor("roti_45_cast_fp16")]; + tensor var_9366_cast_fp16 = mul(x = var_9332_cast_fp16, y = rotr_45_cast_fp16)[name = tensor("op_9366_cast_fp16")]; + tensor var_9367_cast_fp16 = mul(x = var_9340_cast_fp16, y = roti_45_cast_fp16)[name = tensor("op_9367_cast_fp16")]; + tensor qor_89_cast_fp16 = sub(x = var_9366_cast_fp16, y = var_9367_cast_fp16)[name = tensor("qor_89_cast_fp16")]; + tensor var_9370_cast_fp16 = mul(x = var_9332_cast_fp16, y = roti_45_cast_fp16)[name = tensor("op_9370_cast_fp16")]; + tensor var_9371_cast_fp16 = mul(x = var_9340_cast_fp16, y = rotr_45_cast_fp16)[name = tensor("op_9371_cast_fp16")]; + tensor qoi_89_cast_fp16 = add(x = var_9370_cast_fp16, y = var_9371_cast_fp16)[name = tensor("qoi_89_cast_fp16")]; + tensor var_9374_cast_fp16 = mul(x = var_9348_cast_fp16, y = rotr_45_cast_fp16)[name = tensor("op_9374_cast_fp16")]; + tensor var_9375_cast_fp16 = mul(x = var_9356_cast_fp16, y = roti_45_cast_fp16)[name = tensor("op_9375_cast_fp16")]; + tensor kor_89_cast_fp16 = sub(x = var_9374_cast_fp16, y = var_9375_cast_fp16)[name = tensor("kor_89_cast_fp16")]; + tensor var_9378_cast_fp16 = mul(x = var_9348_cast_fp16, y = roti_45_cast_fp16)[name = tensor("op_9378_cast_fp16")]; + tensor var_9379_cast_fp16 = mul(x = var_9356_cast_fp16, y = rotr_45_cast_fp16)[name = tensor("op_9379_cast_fp16")]; + tensor koi_89_cast_fp16 = add(x = var_9378_cast_fp16, y = var_9379_cast_fp16)[name = tensor("koi_89_cast_fp16")]; + tensor qo_45_axis_0 = const()[name = tensor("qo_45_axis_0"), val = tensor(-1)]; + tensor qo_45_cast_fp16 = stack(axis = qo_45_axis_0, values = (qor_89_cast_fp16, qoi_89_cast_fp16))[name = tensor("qo_45_cast_fp16")]; + tensor ko_45_axis_0 = const()[name = tensor("ko_45_axis_0"), val = tensor(-1)]; + tensor ko_45_cast_fp16 = stack(axis = ko_45_axis_0, values = (kor_89_cast_fp16, koi_89_cast_fp16))[name = tensor("ko_45_cast_fp16")]; + tensor var_9412 = const()[name = tensor("op_9412"), val = tensor([1, 256, 16, 64])]; + tensor q_135_cast_fp16 = reshape(shape = var_9412, x = qo_45_cast_fp16)[name = tensor("q_135_cast_fp16")]; + tensor var_9418 = const()[name = tensor("op_9418"), val = tensor([1, 256, 16, 64])]; + tensor k_91_cast_fp16 = reshape(shape = var_9418, x = ko_45_cast_fp16)[name = tensor("k_91_cast_fp16")]; + tensor write_indices_float_133_cast_fp16 = add(x = var_9307_cast_fp16, y = var_393_promoted_to_fp16)[name = tensor("write_indices_float_133_cast_fp16")]; + tensor _inversed_9439_y_0_to_fp16 = const()[name = tensor("_inversed_9439_y_0_to_fp16"), val = tensor(0x1p-9)]; + tensor _inversed_9439_cast_fp16 = mul(x = write_indices_float_133_cast_fp16, y = _inversed_9439_y_0_to_fp16)[name = tensor("_inversed_9439_cast_fp16")]; + tensor var_9440_cast_fp16 = floor(x = _inversed_9439_cast_fp16)[name = tensor("op_9440_cast_fp16")]; + tensor var_9441_to_fp16 = const()[name = tensor("op_9441_to_fp16"), val = tensor(0x1p+9)]; + tensor var_9442_cast_fp16 = mul(x = var_9440_cast_fp16, y = var_9441_to_fp16)[name = tensor("op_9442_cast_fp16")]; + tensor write_indices_float_135_cast_fp16 = sub(x = write_indices_float_133_cast_fp16, y = var_9442_cast_fp16)[name = tensor("write_indices_float_135_cast_fp16")]; + tensor var_9460_value_0_to_fp16 = const()[name = tensor("op_9460_value_0_to_fp16"), val = tensor(0x1.ffp+8)]; + tensor var_9460_cast_fp16 = fill_like(ref_tensor = write_indices_float_135_cast_fp16, value = var_9460_value_0_to_fp16)[name = tensor("op_9460_cast_fp16")]; + tensor write_indices_float_137_cast_fp16 = select(a = write_indices_float_135_cast_fp16, b = var_9460_cast_fp16, cond = token_valid_1_cast_fp16)[name = tensor("write_indices_float_137_cast_fp16")]; + tensor var_9466_dtype_0 = const()[name = tensor("op_9466_dtype_0"), val = tensor("int32")]; + tensor var_9471 = const()[name = tensor("op_9471"), val = tensor([1, 256, 1, 1])]; + tensor write_indices_float_137_cast_fp16_to_int32 = cast(dtype = var_9466_dtype_0, x = write_indices_float_137_cast_fp16)[name = tensor("cast_550")]; + tensor var_9472 = reshape(shape = var_9471, x = write_indices_float_137_cast_fp16_to_int32)[name = tensor("op_9472")]; + tensor write_indices_45_reps_0 = const()[name = tensor("write_indices_45_reps_0"), val = tensor([1, 1, 16, 64])]; + tensor write_indices_45 = tile(reps = write_indices_45_reps_0, x = var_9472)[name = tensor("write_indices_45")]; + tensor var_9482_begin_0 = const()[name = tensor("op_9482_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_9482_end_0 = const()[name = tensor("op_9482_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor var_9482_end_mask_0 = const()[name = tensor("op_9482_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_9482_squeeze_mask_0 = const()[name = tensor("op_9482_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor cache22_to_fp16_dtype_0 = const()[name = tensor("cache22_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor cache22_to_fp16 = cast(dtype = cache22_to_fp16_dtype_0, x = cache22)[name = tensor("cast_549")]; + tensor var_9482_cast_fp16 = slice_by_index(begin = var_9482_begin_0, end = var_9482_end_0, end_mask = var_9482_end_mask_0, squeeze_mask = var_9482_squeeze_mask_0, x = cache22_to_fp16)[name = tensor("op_9482_cast_fp16")]; + tensor var_9484_axis_0 = const()[name = tensor("op_9484_axis_0"), val = tensor(1)]; + tensor var_9484_mode_0 = const()[name = tensor("op_9484_mode_0"), val = tensor("update")]; + tensor var_9484_validate_indices_0 = const()[name = tensor("op_9484_validate_indices_0"), val = tensor(false)]; + tensor var_9484_cast_fp16 = scatter_along_axis(axis = var_9484_axis_0, data = var_9482_cast_fp16, indices = write_indices_45, mode = var_9484_mode_0, updates = k_91_cast_fp16, validate_indices = var_9484_validate_indices_0)[name = tensor("op_9484_cast_fp16")]; + tensor concat_178 = const()[name = tensor("concat_178"), val = tensor([0, 0, 0, 0, 0])]; + tensor concat_179 = const()[name = tensor("concat_179"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_45_internal_tensor_assign_1_stride_0 = const()[name = tensor("new_cache_45_internal_tensor_assign_1_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_45_internal_tensor_assign_1_begin_mask_0 = const()[name = tensor("new_cache_45_internal_tensor_assign_1_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_45_internal_tensor_assign_1_end_mask_0 = const()[name = tensor("new_cache_45_internal_tensor_assign_1_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_45_internal_tensor_assign_1_squeeze_mask_0 = const()[name = tensor("new_cache_45_internal_tensor_assign_1_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_90 = const()[name = tensor("shape_90"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_44 = const()[name = tensor("reduce_prod_44"), val = tensor(1048576)]; + tensor range_1d_44_start_0 = const()[name = tensor("range_1d_44_start_0"), val = tensor(0)]; + tensor range_1d_44_step_0 = const()[name = tensor("range_1d_44_step_0"), val = tensor(1)]; + tensor range_1d_44 = range_1d(end = reduce_prod_44, start = range_1d_44_start_0, step = range_1d_44_step_0)[name = tensor("range_1d_44")]; + tensor reshape_220 = reshape(shape = shape_90, x = range_1d_44)[name = tensor("reshape_220")]; + tensor slice_by_index_44 = slice_by_index(begin = concat_178, begin_mask = new_cache_45_internal_tensor_assign_1_begin_mask_0, end = concat_179, end_mask = new_cache_45_internal_tensor_assign_1_end_mask_0, squeeze_mask = new_cache_45_internal_tensor_assign_1_squeeze_mask_0, stride = new_cache_45_internal_tensor_assign_1_stride_0, x = reshape_220)[name = tensor("slice_by_index_44")]; + tensor reshape_221_shape_0 = const()[name = tensor("reshape_221_shape_0"), val = tensor([-1])]; + tensor reshape_221 = reshape(shape = reshape_221_shape_0, x = slice_by_index_44)[name = tensor("reshape_221")]; + tensor reshape_222_shape_0 = const()[name = tensor("reshape_222_shape_0"), val = tensor([-1])]; + tensor reshape_222_cast_fp16 = reshape(shape = reshape_222_shape_0, x = var_9484_cast_fp16)[name = tensor("reshape_222_cast_fp16")]; + tensor reshape_223_shape_0 = const()[name = tensor("reshape_223_shape_0"), val = tensor([-1])]; + tensor reshape_223_cast_fp16 = reshape(shape = reshape_223_shape_0, x = cache22_to_fp16)[name = tensor("reshape_223_cast_fp16")]; + tensor scatter_44_mode_0 = const()[name = tensor("scatter_44_mode_0"), val = tensor("update")]; + tensor scatter_44_axis_0 = const()[name = tensor("scatter_44_axis_0"), val = tensor(0)]; + tensor scatter_44_validate_indices_0 = const()[name = tensor("scatter_44_validate_indices_0"), val = tensor(false)]; + tensor scatter_44_cast_fp16 = scatter(axis = scatter_44_axis_0, data = reshape_223_cast_fp16, indices = reshape_221, mode = scatter_44_mode_0, updates = reshape_222_cast_fp16, validate_indices = scatter_44_validate_indices_0)[name = tensor("scatter_44_cast_fp16")]; + tensor reshape_224_cast_fp16 = reshape(shape = shape_90, x = scatter_44_cast_fp16)[name = tensor("reshape_224_cast_fp16")]; + tensor var_9492_begin_0 = const()[name = tensor("op_9492_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor var_9492_end_0 = const()[name = tensor("op_9492_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor var_9492_end_mask_0 = const()[name = tensor("op_9492_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_9492_squeeze_mask_0 = const()[name = tensor("op_9492_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor var_9492_cast_fp16 = slice_by_index(begin = var_9492_begin_0, end = var_9492_end_0, end_mask = var_9492_end_mask_0, squeeze_mask = var_9492_squeeze_mask_0, x = reshape_224_cast_fp16)[name = tensor("op_9492_cast_fp16")]; + tensor var_9494_axis_0 = const()[name = tensor("op_9494_axis_0"), val = tensor(1)]; + tensor var_9494_mode_0 = const()[name = tensor("op_9494_mode_0"), val = tensor("update")]; + tensor var_9494_validate_indices_0 = const()[name = tensor("op_9494_validate_indices_0"), val = tensor(false)]; + tensor var_9494_cast_fp16 = scatter_along_axis(axis = var_9494_axis_0, data = var_9492_cast_fp16, indices = write_indices_45, mode = var_9494_mode_0, updates = v_45_cast_fp16, validate_indices = var_9494_validate_indices_0)[name = tensor("op_9494_cast_fp16")]; + tensor concat_180 = const()[name = tensor("concat_180"), val = tensor([1, 0, 0, 0, 0])]; + tensor concat_181 = const()[name = tensor("concat_181"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_45_internal_tensor_assign_2_stride_0 = const()[name = tensor("new_cache_45_internal_tensor_assign_2_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_45_internal_tensor_assign_2_begin_mask_0 = const()[name = tensor("new_cache_45_internal_tensor_assign_2_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_45_internal_tensor_assign_2_end_mask_0 = const()[name = tensor("new_cache_45_internal_tensor_assign_2_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_45_internal_tensor_assign_2_squeeze_mask_0 = const()[name = tensor("new_cache_45_internal_tensor_assign_2_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_91 = const()[name = tensor("shape_91"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_45 = const()[name = tensor("reduce_prod_45"), val = tensor(1048576)]; + tensor range_1d_45_start_0 = const()[name = tensor("range_1d_45_start_0"), val = tensor(0)]; + tensor range_1d_45_step_0 = const()[name = tensor("range_1d_45_step_0"), val = tensor(1)]; + tensor range_1d_45 = range_1d(end = reduce_prod_45, start = range_1d_45_start_0, step = range_1d_45_step_0)[name = tensor("range_1d_45")]; + tensor reshape_225 = reshape(shape = shape_91, x = range_1d_45)[name = tensor("reshape_225")]; + tensor slice_by_index_45 = slice_by_index(begin = concat_180, begin_mask = new_cache_45_internal_tensor_assign_2_begin_mask_0, end = concat_181, end_mask = new_cache_45_internal_tensor_assign_2_end_mask_0, squeeze_mask = new_cache_45_internal_tensor_assign_2_squeeze_mask_0, stride = new_cache_45_internal_tensor_assign_2_stride_0, x = reshape_225)[name = tensor("slice_by_index_45")]; + tensor reshape_226_shape_0 = const()[name = tensor("reshape_226_shape_0"), val = tensor([-1])]; + tensor reshape_226 = reshape(shape = reshape_226_shape_0, x = slice_by_index_45)[name = tensor("reshape_226")]; + tensor reshape_227_shape_0 = const()[name = tensor("reshape_227_shape_0"), val = tensor([-1])]; + tensor reshape_227_cast_fp16 = reshape(shape = reshape_227_shape_0, x = var_9494_cast_fp16)[name = tensor("reshape_227_cast_fp16")]; + tensor reshape_228_shape_0 = const()[name = tensor("reshape_228_shape_0"), val = tensor([-1])]; + tensor reshape_228_cast_fp16 = reshape(shape = reshape_228_shape_0, x = reshape_224_cast_fp16)[name = tensor("reshape_228_cast_fp16")]; + tensor scatter_45_mode_0 = const()[name = tensor("scatter_45_mode_0"), val = tensor("update")]; + tensor scatter_45_axis_0 = const()[name = tensor("scatter_45_axis_0"), val = tensor(0)]; + tensor scatter_45_validate_indices_0 = const()[name = tensor("scatter_45_validate_indices_0"), val = tensor(false)]; + tensor scatter_45_cast_fp16 = scatter(axis = scatter_45_axis_0, data = reshape_228_cast_fp16, indices = reshape_226, mode = scatter_45_mode_0, updates = reshape_227_cast_fp16, validate_indices = scatter_45_validate_indices_0)[name = tensor("scatter_45_cast_fp16")]; + tensor reshape_229_cast_fp16 = reshape(shape = shape_91, x = scatter_45_cast_fp16)[name = tensor("reshape_229_cast_fp16")]; + tensor reshape_229_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("reshape_229_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor keys_133_begin_0 = const()[name = tensor("keys_133_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor keys_133_end_0 = const()[name = tensor("keys_133_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor keys_133_end_mask_0 = const()[name = tensor("keys_133_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor keys_133_squeeze_mask_0 = const()[name = tensor("keys_133_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor keys_133_cast_fp16 = slice_by_index(begin = keys_133_begin_0, end = keys_133_end_0, end_mask = keys_133_end_mask_0, squeeze_mask = keys_133_squeeze_mask_0, x = reshape_229_cast_fp16)[name = tensor("keys_133_cast_fp16")]; + tensor values_133_begin_0 = const()[name = tensor("values_133_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor values_133_end_0 = const()[name = tensor("values_133_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor values_133_end_mask_0 = const()[name = tensor("values_133_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor values_133_squeeze_mask_0 = const()[name = tensor("values_133_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor values_133_cast_fp16 = slice_by_index(begin = values_133_begin_0, end = values_133_end_0, end_mask = values_133_end_mask_0, squeeze_mask = values_133_squeeze_mask_0, x = reshape_229_cast_fp16)[name = tensor("values_133_cast_fp16")]; + tensor var_9506_cast_fp16 = not_equal(x = keys_133_cast_fp16, y = keys_133_cast_fp16)[name = tensor("op_9506_cast_fp16")]; + tensor keys_135_cast_fp16 = select(a = var_470_to_fp16, b = keys_133_cast_fp16, cond = var_9506_cast_fp16)[name = tensor("keys_135_cast_fp16")]; + tensor var_9514_cast_fp16 = not_equal(x = values_133_cast_fp16, y = values_133_cast_fp16)[name = tensor("op_9514_cast_fp16")]; + tensor values_135_cast_fp16 = select(a = var_470_to_fp16, b = values_133_cast_fp16, cond = var_9514_cast_fp16)[name = tensor("values_135_cast_fp16")]; + tensor var_9538 = const()[name = tensor("op_9538"), val = tensor([0, 2, 1, 3])]; + tensor var_9554 = const()[name = tensor("op_9554"), val = tensor([1, 1, 1])]; + tensor var_9555_cast_fp16 = reshape(shape = var_9554, x = position22_to_fp16)[name = tensor("op_9555_cast_fp16")]; + tensor q_positions_45_cast_fp16 = add(x = var_9555_cast_fp16, y = q_offsets_1_promoted_to_fp16)[name = tensor("q_positions_45_cast_fp16")]; + tensor valid_end_45_cast_fp16 = add(x = var_9555_cast_fp16, y = var_536_cast_fp16)[name = tensor("valid_end_45_cast_fp16")]; + tensor valid_mask_45_cast_fp16 = less(x = k_positions_1_promoted_to_fp16, y = valid_end_45_cast_fp16)[name = tensor("valid_mask_45_cast_fp16")]; + tensor causal_mask_45_cast_fp16 = less_equal(x = k_positions_1_promoted_to_fp16, y = q_positions_45_cast_fp16)[name = tensor("causal_mask_45_cast_fp16")]; + tensor var_9583 = logical_and(x = valid_mask_45_cast_fp16, y = causal_mask_45_cast_fp16)[name = tensor("op_9583")]; + tensor attn_mask_45_axes_0 = const()[name = tensor("attn_mask_45_axes_0"), val = tensor([1])]; + tensor attn_mask_45 = expand_dims(axes = attn_mask_45_axes_0, x = var_9583)[name = tensor("attn_mask_45")]; + tensor var_9597_transpose_x_0 = const()[name = tensor("op_9597_transpose_x_0"), val = tensor(false)]; + tensor var_9597_transpose_y_0 = const()[name = tensor("op_9597_transpose_y_0"), val = tensor(false)]; + tensor transpose_113_perm_0 = const()[name = tensor("transpose_113_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_114_perm_0 = const()[name = tensor("transpose_114_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_114 = transpose(perm = transpose_114_perm_0, x = keys_135_cast_fp16)[name = tensor("transpose_116")]; + tensor transpose_113 = transpose(perm = transpose_113_perm_0, x = q_135_cast_fp16)[name = tensor("transpose_117")]; + tensor var_9597_cast_fp16 = matmul(transpose_x = var_9597_transpose_x_0, transpose_y = var_9597_transpose_y_0, x = transpose_113, y = transpose_114)[name = tensor("op_9597_cast_fp16")]; + tensor var_9591_to_fp16 = const()[name = tensor("op_9591_to_fp16"), val = tensor([0x1p-3])]; + tensor attn_weights_133_cast_fp16 = mul(x = var_9597_cast_fp16, y = var_9591_to_fp16)[name = tensor("attn_weights_133_cast_fp16")]; + tensor var_9599 = logical_not(x = attn_mask_45)[name = tensor("op_9599")]; + tensor var_9600_to_fp16 = const()[name = tensor("op_9600_to_fp16"), val = tensor(-inf)]; + tensor attn_weights_135_cast_fp16 = select(a = var_9600_to_fp16, b = attn_weights_133_cast_fp16, cond = var_9599)[name = tensor("attn_weights_135_cast_fp16")]; + tensor var_9602 = const()[name = tensor("op_9602"), val = tensor(-1)]; + tensor attn_weights_137_cast_fp16 = softmax(axis = var_9602, x = attn_weights_135_cast_fp16)[name = tensor("attn_weights_137_cast_fp16")]; + tensor attn_output_45_transpose_x_0 = const()[name = tensor("attn_output_45_transpose_x_0"), val = tensor(false)]; + tensor attn_output_45_transpose_y_0 = const()[name = tensor("attn_output_45_transpose_y_0"), val = tensor(false)]; + tensor values_137_cast_fp16 = transpose(perm = var_9538, x = values_135_cast_fp16)[name = tensor("transpose_118")]; + tensor attn_output_45_cast_fp16 = matmul(transpose_x = attn_output_45_transpose_x_0, transpose_y = attn_output_45_transpose_y_0, x = attn_weights_137_cast_fp16, y = values_137_cast_fp16)[name = tensor("attn_output_45_cast_fp16")]; + tensor var_9610 = const()[name = tensor("op_9610"), val = tensor([0, 2, 1, 3])]; + tensor var_9615 = const()[name = tensor("op_9615"), val = tensor([1, 256, 1024])]; + tensor var_9611_cast_fp16 = transpose(perm = var_9610, x = attn_output_45_cast_fp16)[name = tensor("transpose_115")]; + tensor input_269_cast_fp16 = reshape(shape = var_9615, x = var_9611_cast_fp16)[name = tensor("input_269_cast_fp16")]; + tensor attn22_out_proj_weight_to_fp16 = const()[name = tensor("attn22_out_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(561206400)))]; + tensor linear_89_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = attn22_out_proj_weight_to_fp16, x = input_269_cast_fp16)[name = tensor("linear_89_cast_fp16")]; + tensor var_9624_cast_fp16 = add(x = position22_to_fp16, y = valid_len_to_fp16)[name = tensor("op_9624_cast_fp16")]; + tensor var_9624_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_9624_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor input_271_cast_fp16 = add(x = input_265_cast_fp16, y = linear_89_cast_fp16)[name = tensor("input_271_cast_fp16")]; + tensor input_273_axes_0 = const()[name = tensor("input_273_axes_0"), val = tensor([-1])]; + tensor norm22_2_weight_to_fp16 = const()[name = tensor("norm22_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(563303616)))]; + tensor norm22_2_bias_to_fp16 = const()[name = tensor("norm22_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(563305728)))]; + tensor var_9628_to_fp16 = const()[name = tensor("op_9628_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_273_cast_fp16 = layer_norm(axes = input_273_axes_0, beta = norm22_2_bias_to_fp16, epsilon = var_9628_to_fp16, gamma = norm22_2_weight_to_fp16, x = input_271_cast_fp16)[name = tensor("input_273_cast_fp16")]; + tensor linear22_1_weight_to_fp16 = const()[name = tensor("linear22_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(563307840)))]; + tensor linear_90_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = linear22_1_weight_to_fp16, x = input_273_cast_fp16)[name = tensor("linear_90_cast_fp16")]; + tensor input_275_mode_0 = const()[name = tensor("input_275_mode_0"), val = tensor("EXACT")]; + tensor input_275_cast_fp16 = gelu(mode = input_275_mode_0, x = linear_90_cast_fp16)[name = tensor("input_275_cast_fp16")]; + tensor linear22_2_weight_to_fp16 = const()[name = tensor("linear22_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(571696512)))]; + tensor linear_91_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = linear22_2_weight_to_fp16, x = input_275_cast_fp16)[name = tensor("linear_91_cast_fp16")]; + tensor input_277_cast_fp16 = add(x = input_271_cast_fp16, y = linear_91_cast_fp16)[name = tensor("input_277_cast_fp16")]; + tensor input_279_axes_0 = const()[name = tensor("input_279_axes_0"), val = tensor([-1])]; + tensor norm23_1_weight_to_fp16 = const()[name = tensor("norm23_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(580085184)))]; + tensor norm23_1_bias_to_fp16 = const()[name = tensor("norm23_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(580087296)))]; + tensor var_9645_to_fp16 = const()[name = tensor("op_9645_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_279_cast_fp16 = layer_norm(axes = input_279_axes_0, beta = norm23_1_bias_to_fp16, epsilon = var_9645_to_fp16, gamma = norm23_1_weight_to_fp16, x = input_277_cast_fp16)[name = tensor("input_279_cast_fp16")]; + tensor attn23_in_proj_weight_to_fp16 = const()[name = tensor("attn23_in_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(580089408)))]; + tensor linear_92_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = attn23_in_proj_weight_to_fp16, x = input_279_cast_fp16)[name = tensor("linear_92_cast_fp16")]; + tensor var_9659 = const()[name = tensor("op_9659"), val = tensor([1, 256, 3, 16, 64])]; + tensor qkv_cast_fp16 = reshape(shape = var_9659, x = linear_92_cast_fp16)[name = tensor("qkv_cast_fp16")]; + tensor k_93_begin_0 = const()[name = tensor("k_93_begin_0"), val = tensor([0, 0, 1, 0, 0])]; + tensor k_93_end_0 = const()[name = tensor("k_93_end_0"), val = tensor([1, 256, 2, 16, 64])]; + tensor k_93_end_mask_0 = const()[name = tensor("k_93_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor k_93_squeeze_mask_0 = const()[name = tensor("k_93_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor k_93_cast_fp16 = slice_by_index(begin = k_93_begin_0, end = k_93_end_0, end_mask = k_93_end_mask_0, squeeze_mask = k_93_squeeze_mask_0, x = qkv_cast_fp16)[name = tensor("k_93_cast_fp16")]; + tensor v_begin_0 = const()[name = tensor("v_begin_0"), val = tensor([0, 0, 2, 0, 0])]; + tensor v_end_0 = const()[name = tensor("v_end_0"), val = tensor([1, 256, 3, 16, 64])]; + tensor v_end_mask_0 = const()[name = tensor("v_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor v_squeeze_mask_0 = const()[name = tensor("v_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor v_cast_fp16 = slice_by_index(begin = v_begin_0, end = v_end_0, end_mask = v_end_mask_0, squeeze_mask = v_squeeze_mask_0, x = qkv_cast_fp16)[name = tensor("v_cast_fp16")]; + tensor var_9704 = const()[name = tensor("op_9704"), val = tensor([1, 1])]; + tensor position23_to_fp16_dtype_0 = const()[name = tensor("position23_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor position23_to_fp16 = cast(dtype = position23_to_fp16_dtype_0, x = position23)[name = tensor("cast_546")]; + tensor var_9705_cast_fp16 = reshape(shape = var_9704, x = position23_to_fp16)[name = tensor("op_9705_cast_fp16")]; + tensor ts_141_cast_fp16 = add(x = ts_1_promoted_to_fp16, y = var_9705_cast_fp16)[name = tensor("ts_141_cast_fp16")]; + tensor var_9712 = const()[name = tensor("op_9712"), val = tensor([1, 256, 1, 1])]; + tensor ts_cast_fp16 = reshape(shape = var_9712, x = ts_141_cast_fp16)[name = tensor("ts_cast_fp16")]; + tensor var_9719 = const()[name = tensor("op_9719"), val = tensor([1, 256, 16, 32, 2])]; + tensor k_complex_cast_fp16 = reshape(shape = var_9719, x = k_93_cast_fp16)[name = tensor("k_complex_cast_fp16")]; + tensor var_9723_begin_0 = const()[name = tensor("op_9723_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_9723_end_0 = const()[name = tensor("op_9723_end_0"), val = tensor([1, 256, 16, 32, 1])]; + tensor var_9723_end_mask_0 = const()[name = tensor("op_9723_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_9723_squeeze_mask_0 = const()[name = tensor("op_9723_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_9723_cast_fp16 = slice_by_index(begin = var_9723_begin_0, end = var_9723_end_0, end_mask = var_9723_end_mask_0, squeeze_mask = var_9723_squeeze_mask_0, x = k_complex_cast_fp16)[name = tensor("op_9723_cast_fp16")]; + tensor var_9731_begin_0 = const()[name = tensor("op_9731_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_9731_end_0 = const()[name = tensor("op_9731_end_0"), val = tensor([1, 256, 16, 32, 2])]; + tensor var_9731_end_mask_0 = const()[name = tensor("op_9731_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_9731_squeeze_mask_0 = const()[name = tensor("op_9731_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_9731_cast_fp16 = slice_by_index(begin = var_9731_begin_0, end = var_9731_end_0, end_mask = var_9731_end_mask_0, squeeze_mask = var_9731_squeeze_mask_0, x = k_complex_cast_fp16)[name = tensor("op_9731_cast_fp16")]; + tensor freqs_to_fp16 = const()[name = tensor("freqs_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(586380928)))]; + tensor var_9737_cast_fp16 = mul(x = freqs_to_fp16, y = ts_cast_fp16)[name = tensor("op_9737_cast_fp16")]; + tensor rotr_cast_fp16 = cos(x = var_9737_cast_fp16)[name = tensor("rotr_cast_fp16")]; + tensor roti_cast_fp16 = sin(x = var_9737_cast_fp16)[name = tensor("roti_cast_fp16")]; + tensor var_9741_cast_fp16 = mul(x = var_9723_cast_fp16, y = rotr_cast_fp16)[name = tensor("op_9741_cast_fp16")]; + tensor var_9742_cast_fp16 = mul(x = var_9731_cast_fp16, y = roti_cast_fp16)[name = tensor("op_9742_cast_fp16")]; + tensor kor_93_cast_fp16 = sub(x = var_9741_cast_fp16, y = var_9742_cast_fp16)[name = tensor("kor_93_cast_fp16")]; + tensor var_9745_cast_fp16 = mul(x = var_9723_cast_fp16, y = roti_cast_fp16)[name = tensor("op_9745_cast_fp16")]; + tensor var_9746_cast_fp16 = mul(x = var_9731_cast_fp16, y = rotr_cast_fp16)[name = tensor("op_9746_cast_fp16")]; + tensor koi_93_cast_fp16 = add(x = var_9745_cast_fp16, y = var_9746_cast_fp16)[name = tensor("koi_93_cast_fp16")]; + tensor ko_axis_0 = const()[name = tensor("ko_axis_0"), val = tensor(-1)]; + tensor ko_cast_fp16 = stack(axis = ko_axis_0, values = (kor_93_cast_fp16, koi_93_cast_fp16))[name = tensor("ko_cast_fp16")]; + tensor var_9766 = const()[name = tensor("op_9766"), val = tensor([1, 256, 16, 64])]; + tensor k_cast_fp16 = reshape(shape = var_9766, x = ko_cast_fp16)[name = tensor("k_cast_fp16")]; + tensor write_indices_float_139_cast_fp16 = add(x = var_9705_cast_fp16, y = var_393_promoted_to_fp16)[name = tensor("write_indices_float_139_cast_fp16")]; + tensor _inversed_9787_y_0_to_fp16 = const()[name = tensor("_inversed_9787_y_0_to_fp16"), val = tensor(0x1p-9)]; + tensor _inversed_9787_cast_fp16 = mul(x = write_indices_float_139_cast_fp16, y = _inversed_9787_y_0_to_fp16)[name = tensor("_inversed_9787_cast_fp16")]; + tensor var_9788_cast_fp16 = floor(x = _inversed_9787_cast_fp16)[name = tensor("op_9788_cast_fp16")]; + tensor var_9789_to_fp16 = const()[name = tensor("op_9789_to_fp16"), val = tensor(0x1p+9)]; + tensor var_9790_cast_fp16 = mul(x = var_9788_cast_fp16, y = var_9789_to_fp16)[name = tensor("op_9790_cast_fp16")]; + tensor write_indices_float_141_cast_fp16 = sub(x = write_indices_float_139_cast_fp16, y = var_9790_cast_fp16)[name = tensor("write_indices_float_141_cast_fp16")]; + tensor var_9808_value_0_to_fp16 = const()[name = tensor("op_9808_value_0_to_fp16"), val = tensor(0x1.ffp+8)]; + tensor var_9808_cast_fp16 = fill_like(ref_tensor = write_indices_float_141_cast_fp16, value = var_9808_value_0_to_fp16)[name = tensor("op_9808_cast_fp16")]; + tensor write_indices_float_cast_fp16 = select(a = write_indices_float_141_cast_fp16, b = var_9808_cast_fp16, cond = token_valid_1_cast_fp16)[name = tensor("write_indices_float_cast_fp16")]; + tensor var_9814_dtype_0 = const()[name = tensor("op_9814_dtype_0"), val = tensor("int32")]; + tensor var_9819 = const()[name = tensor("op_9819"), val = tensor([1, 256, 1, 1])]; + tensor write_indices_float_cast_fp16_to_int32 = cast(dtype = var_9814_dtype_0, x = write_indices_float_cast_fp16)[name = tensor("cast_545")]; + tensor var_9820 = reshape(shape = var_9819, x = write_indices_float_cast_fp16_to_int32)[name = tensor("op_9820")]; + tensor write_indices_reps_0 = const()[name = tensor("write_indices_reps_0"), val = tensor([1, 1, 16, 64])]; + tensor write_indices = tile(reps = write_indices_reps_0, x = var_9820)[name = tensor("write_indices")]; + tensor var_9830_begin_0 = const()[name = tensor("op_9830_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_9830_end_0 = const()[name = tensor("op_9830_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor var_9830_end_mask_0 = const()[name = tensor("op_9830_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_9830_squeeze_mask_0 = const()[name = tensor("op_9830_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor cache23_to_fp16_dtype_0 = const()[name = tensor("cache23_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor cache23_to_fp16 = cast(dtype = cache23_to_fp16_dtype_0, x = cache23)[name = tensor("cast_544")]; + tensor var_9830_cast_fp16 = slice_by_index(begin = var_9830_begin_0, end = var_9830_end_0, end_mask = var_9830_end_mask_0, squeeze_mask = var_9830_squeeze_mask_0, x = cache23_to_fp16)[name = tensor("op_9830_cast_fp16")]; + tensor var_9832_axis_0 = const()[name = tensor("op_9832_axis_0"), val = tensor(1)]; + tensor var_9832_mode_0 = const()[name = tensor("op_9832_mode_0"), val = tensor("update")]; + tensor var_9832_validate_indices_0 = const()[name = tensor("op_9832_validate_indices_0"), val = tensor(false)]; + tensor var_9832_cast_fp16 = scatter_along_axis(axis = var_9832_axis_0, data = var_9830_cast_fp16, indices = write_indices, mode = var_9832_mode_0, updates = k_cast_fp16, validate_indices = var_9832_validate_indices_0)[name = tensor("op_9832_cast_fp16")]; + tensor concat_186 = const()[name = tensor("concat_186"), val = tensor([0, 0, 0, 0, 0])]; + tensor concat_187 = const()[name = tensor("concat_187"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_internal_tensor_assign_1_stride_0 = const()[name = tensor("new_cache_internal_tensor_assign_1_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_internal_tensor_assign_1_begin_mask_0 = const()[name = tensor("new_cache_internal_tensor_assign_1_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_internal_tensor_assign_1_end_mask_0 = const()[name = tensor("new_cache_internal_tensor_assign_1_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_internal_tensor_assign_1_squeeze_mask_0 = const()[name = tensor("new_cache_internal_tensor_assign_1_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_92 = const()[name = tensor("shape_92"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_46 = const()[name = tensor("reduce_prod_46"), val = tensor(1048576)]; + tensor range_1d_46_start_0 = const()[name = tensor("range_1d_46_start_0"), val = tensor(0)]; + tensor range_1d_46_step_0 = const()[name = tensor("range_1d_46_step_0"), val = tensor(1)]; + tensor range_1d_46 = range_1d(end = reduce_prod_46, start = range_1d_46_start_0, step = range_1d_46_step_0)[name = tensor("range_1d_46")]; + tensor reshape_230 = reshape(shape = shape_92, x = range_1d_46)[name = tensor("reshape_230")]; + tensor slice_by_index_46 = slice_by_index(begin = concat_186, begin_mask = new_cache_internal_tensor_assign_1_begin_mask_0, end = concat_187, end_mask = new_cache_internal_tensor_assign_1_end_mask_0, squeeze_mask = new_cache_internal_tensor_assign_1_squeeze_mask_0, stride = new_cache_internal_tensor_assign_1_stride_0, x = reshape_230)[name = tensor("slice_by_index_46")]; + tensor reshape_231_shape_0 = const()[name = tensor("reshape_231_shape_0"), val = tensor([-1])]; + tensor reshape_231 = reshape(shape = reshape_231_shape_0, x = slice_by_index_46)[name = tensor("reshape_231")]; + tensor reshape_232_shape_0 = const()[name = tensor("reshape_232_shape_0"), val = tensor([-1])]; + tensor reshape_232_cast_fp16 = reshape(shape = reshape_232_shape_0, x = var_9832_cast_fp16)[name = tensor("reshape_232_cast_fp16")]; + tensor reshape_233_shape_0 = const()[name = tensor("reshape_233_shape_0"), val = tensor([-1])]; + tensor reshape_233_cast_fp16 = reshape(shape = reshape_233_shape_0, x = cache23_to_fp16)[name = tensor("reshape_233_cast_fp16")]; + tensor scatter_46_mode_0 = const()[name = tensor("scatter_46_mode_0"), val = tensor("update")]; + tensor scatter_46_axis_0 = const()[name = tensor("scatter_46_axis_0"), val = tensor(0)]; + tensor scatter_46_validate_indices_0 = const()[name = tensor("scatter_46_validate_indices_0"), val = tensor(false)]; + tensor scatter_46_cast_fp16 = scatter(axis = scatter_46_axis_0, data = reshape_233_cast_fp16, indices = reshape_231, mode = scatter_46_mode_0, updates = reshape_232_cast_fp16, validate_indices = scatter_46_validate_indices_0)[name = tensor("scatter_46_cast_fp16")]; + tensor reshape_234_cast_fp16 = reshape(shape = shape_92, x = scatter_46_cast_fp16)[name = tensor("reshape_234_cast_fp16")]; + tensor var_9840_begin_0 = const()[name = tensor("op_9840_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor var_9840_end_0 = const()[name = tensor("op_9840_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor var_9840_end_mask_0 = const()[name = tensor("op_9840_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_9840_squeeze_mask_0 = const()[name = tensor("op_9840_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor var_9840_cast_fp16 = slice_by_index(begin = var_9840_begin_0, end = var_9840_end_0, end_mask = var_9840_end_mask_0, squeeze_mask = var_9840_squeeze_mask_0, x = reshape_234_cast_fp16)[name = tensor("op_9840_cast_fp16")]; + tensor var_9842_axis_0 = const()[name = tensor("op_9842_axis_0"), val = tensor(1)]; + tensor var_9842_mode_0 = const()[name = tensor("op_9842_mode_0"), val = tensor("update")]; + tensor var_9842_validate_indices_0 = const()[name = tensor("op_9842_validate_indices_0"), val = tensor(false)]; + tensor var_9842_cast_fp16 = scatter_along_axis(axis = var_9842_axis_0, data = var_9840_cast_fp16, indices = write_indices, mode = var_9842_mode_0, updates = v_cast_fp16, validate_indices = var_9842_validate_indices_0)[name = tensor("op_9842_cast_fp16")]; + tensor concat_188 = const()[name = tensor("concat_188"), val = tensor([1, 0, 0, 0, 0])]; + tensor concat_189 = const()[name = tensor("concat_189"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_internal_tensor_assign_2_stride_0 = const()[name = tensor("new_cache_internal_tensor_assign_2_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_internal_tensor_assign_2_begin_mask_0 = const()[name = tensor("new_cache_internal_tensor_assign_2_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_internal_tensor_assign_2_end_mask_0 = const()[name = tensor("new_cache_internal_tensor_assign_2_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_internal_tensor_assign_2_squeeze_mask_0 = const()[name = tensor("new_cache_internal_tensor_assign_2_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_93 = const()[name = tensor("shape_93"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_47 = const()[name = tensor("reduce_prod_47"), val = tensor(1048576)]; + tensor range_1d_47_start_0 = const()[name = tensor("range_1d_47_start_0"), val = tensor(0)]; + tensor range_1d_47_step_0 = const()[name = tensor("range_1d_47_step_0"), val = tensor(1)]; + tensor range_1d_47 = range_1d(end = reduce_prod_47, start = range_1d_47_start_0, step = range_1d_47_step_0)[name = tensor("range_1d_47")]; + tensor reshape_235 = reshape(shape = shape_93, x = range_1d_47)[name = tensor("reshape_235")]; + tensor slice_by_index_47 = slice_by_index(begin = concat_188, begin_mask = new_cache_internal_tensor_assign_2_begin_mask_0, end = concat_189, end_mask = new_cache_internal_tensor_assign_2_end_mask_0, squeeze_mask = new_cache_internal_tensor_assign_2_squeeze_mask_0, stride = new_cache_internal_tensor_assign_2_stride_0, x = reshape_235)[name = tensor("slice_by_index_47")]; + tensor reshape_236_shape_0 = const()[name = tensor("reshape_236_shape_0"), val = tensor([-1])]; + tensor reshape_236 = reshape(shape = reshape_236_shape_0, x = slice_by_index_47)[name = tensor("reshape_236")]; + tensor reshape_237_shape_0 = const()[name = tensor("reshape_237_shape_0"), val = tensor([-1])]; + tensor reshape_237_cast_fp16 = reshape(shape = reshape_237_shape_0, x = var_9842_cast_fp16)[name = tensor("reshape_237_cast_fp16")]; + tensor reshape_238_shape_0 = const()[name = tensor("reshape_238_shape_0"), val = tensor([-1])]; + tensor reshape_238_cast_fp16 = reshape(shape = reshape_238_shape_0, x = reshape_234_cast_fp16)[name = tensor("reshape_238_cast_fp16")]; + tensor scatter_47_mode_0 = const()[name = tensor("scatter_47_mode_0"), val = tensor("update")]; + tensor scatter_47_axis_0 = const()[name = tensor("scatter_47_axis_0"), val = tensor(0)]; + tensor scatter_47_validate_indices_0 = const()[name = tensor("scatter_47_validate_indices_0"), val = tensor(false)]; + tensor scatter_47_cast_fp16 = scatter(axis = scatter_47_axis_0, data = reshape_238_cast_fp16, indices = reshape_236, mode = scatter_47_mode_0, updates = reshape_237_cast_fp16, validate_indices = scatter_47_validate_indices_0)[name = tensor("scatter_47_cast_fp16")]; + tensor reshape_239_cast_fp16 = reshape(shape = shape_93, x = scatter_47_cast_fp16)[name = tensor("reshape_239_cast_fp16")]; + tensor reshape_239_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("reshape_239_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor var_9852_cast_fp16 = add(x = position23_to_fp16, y = valid_len_to_fp16)[name = tensor("op_9852_cast_fp16")]; + tensor var_9852_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_9852_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor var_9852 = cast(dtype = var_9852_cast_fp16_to_fp32_dtype_0, x = var_9852_cast_fp16)[name = tensor("cast_542")]; + tensor new_cache_internal_tensor_assign_2 = cast(dtype = reshape_239_cast_fp16_to_fp32_dtype_0, x = reshape_239_cast_fp16)[name = tensor("cast_543")]; + tensor var_9624 = cast(dtype = var_9624_cast_fp16_to_fp32_dtype_0, x = var_9624_cast_fp16)[name = tensor("cast_547")]; + tensor new_cache_45_internal_tensor_assign_2 = cast(dtype = reshape_229_cast_fp16_to_fp32_dtype_0, x = reshape_229_cast_fp16)[name = tensor("cast_548")]; + tensor var_9213 = cast(dtype = var_9213_cast_fp16_to_fp32_dtype_0, x = var_9213_cast_fp16)[name = tensor("cast_552")]; + tensor new_cache_43_internal_tensor_assign_2 = cast(dtype = reshape_219_cast_fp16_to_fp32_dtype_0, x = reshape_219_cast_fp16)[name = tensor("cast_553")]; + tensor var_8802 = cast(dtype = var_8802_cast_fp16_to_fp32_dtype_0, x = var_8802_cast_fp16)[name = tensor("cast_557")]; + tensor new_cache_41_internal_tensor_assign_2 = cast(dtype = reshape_209_cast_fp16_to_fp32_dtype_0, x = reshape_209_cast_fp16)[name = tensor("cast_558")]; + tensor var_8391 = cast(dtype = var_8391_cast_fp16_to_fp32_dtype_0, x = var_8391_cast_fp16)[name = tensor("cast_562")]; + tensor new_cache_39_internal_tensor_assign_2 = cast(dtype = reshape_199_cast_fp16_to_fp32_dtype_0, x = reshape_199_cast_fp16)[name = tensor("cast_563")]; + tensor var_7980 = cast(dtype = var_7980_cast_fp16_to_fp32_dtype_0, x = var_7980_cast_fp16)[name = tensor("cast_567")]; + tensor new_cache_37_internal_tensor_assign_2 = cast(dtype = reshape_189_cast_fp16_to_fp32_dtype_0, x = reshape_189_cast_fp16)[name = tensor("cast_568")]; + tensor var_7569 = cast(dtype = var_7569_cast_fp16_to_fp32_dtype_0, x = var_7569_cast_fp16)[name = tensor("cast_572")]; + tensor new_cache_35_internal_tensor_assign_2 = cast(dtype = reshape_179_cast_fp16_to_fp32_dtype_0, x = reshape_179_cast_fp16)[name = tensor("cast_573")]; + tensor var_7158 = cast(dtype = var_7158_cast_fp16_to_fp32_dtype_0, x = var_7158_cast_fp16)[name = tensor("cast_577")]; + tensor new_cache_33_internal_tensor_assign_2 = cast(dtype = reshape_169_cast_fp16_to_fp32_dtype_0, x = reshape_169_cast_fp16)[name = tensor("cast_578")]; + tensor var_6747 = cast(dtype = var_6747_cast_fp16_to_fp32_dtype_0, x = var_6747_cast_fp16)[name = tensor("cast_582")]; + tensor new_cache_31_internal_tensor_assign_2 = cast(dtype = reshape_159_cast_fp16_to_fp32_dtype_0, x = reshape_159_cast_fp16)[name = tensor("cast_583")]; + tensor var_6336 = cast(dtype = var_6336_cast_fp16_to_fp32_dtype_0, x = var_6336_cast_fp16)[name = tensor("cast_587")]; + tensor new_cache_29_internal_tensor_assign_2 = cast(dtype = reshape_149_cast_fp16_to_fp32_dtype_0, x = reshape_149_cast_fp16)[name = tensor("cast_588")]; + tensor var_5925 = cast(dtype = var_5925_cast_fp16_to_fp32_dtype_0, x = var_5925_cast_fp16)[name = tensor("cast_592")]; + tensor new_cache_27_internal_tensor_assign_2 = cast(dtype = reshape_139_cast_fp16_to_fp32_dtype_0, x = reshape_139_cast_fp16)[name = tensor("cast_593")]; + tensor var_5514 = cast(dtype = var_5514_cast_fp16_to_fp32_dtype_0, x = var_5514_cast_fp16)[name = tensor("cast_597")]; + tensor new_cache_25_internal_tensor_assign_2 = cast(dtype = reshape_129_cast_fp16_to_fp32_dtype_0, x = reshape_129_cast_fp16)[name = tensor("cast_598")]; + tensor var_5103 = cast(dtype = var_5103_cast_fp16_to_fp32_dtype_0, x = var_5103_cast_fp16)[name = tensor("cast_602")]; + tensor new_cache_23_internal_tensor_assign_2 = cast(dtype = reshape_119_cast_fp16_to_fp32_dtype_0, x = reshape_119_cast_fp16)[name = tensor("cast_603")]; + tensor var_4692 = cast(dtype = var_4692_cast_fp16_to_fp32_dtype_0, x = var_4692_cast_fp16)[name = tensor("cast_607")]; + tensor new_cache_21_internal_tensor_assign_2 = cast(dtype = reshape_109_cast_fp16_to_fp32_dtype_0, x = reshape_109_cast_fp16)[name = tensor("cast_608")]; + tensor var_4281 = cast(dtype = var_4281_cast_fp16_to_fp32_dtype_0, x = var_4281_cast_fp16)[name = tensor("cast_612")]; + tensor new_cache_19_internal_tensor_assign_2 = cast(dtype = reshape_99_cast_fp16_to_fp32_dtype_0, x = reshape_99_cast_fp16)[name = tensor("cast_613")]; + tensor var_3870 = cast(dtype = var_3870_cast_fp16_to_fp32_dtype_0, x = var_3870_cast_fp16)[name = tensor("cast_617")]; + tensor new_cache_17_internal_tensor_assign_2 = cast(dtype = reshape_89_cast_fp16_to_fp32_dtype_0, x = reshape_89_cast_fp16)[name = tensor("cast_618")]; + tensor var_3459 = cast(dtype = var_3459_cast_fp16_to_fp32_dtype_0, x = var_3459_cast_fp16)[name = tensor("cast_622")]; + tensor new_cache_15_internal_tensor_assign_2 = cast(dtype = reshape_79_cast_fp16_to_fp32_dtype_0, x = reshape_79_cast_fp16)[name = tensor("cast_623")]; + tensor var_3048 = cast(dtype = var_3048_cast_fp16_to_fp32_dtype_0, x = var_3048_cast_fp16)[name = tensor("cast_627")]; + tensor new_cache_13_internal_tensor_assign_2 = cast(dtype = reshape_69_cast_fp16_to_fp32_dtype_0, x = reshape_69_cast_fp16)[name = tensor("cast_628")]; + tensor var_2637 = cast(dtype = var_2637_cast_fp16_to_fp32_dtype_0, x = var_2637_cast_fp16)[name = tensor("cast_632")]; + tensor new_cache_11_internal_tensor_assign_2 = cast(dtype = reshape_59_cast_fp16_to_fp32_dtype_0, x = reshape_59_cast_fp16)[name = tensor("cast_633")]; + tensor var_2226 = cast(dtype = var_2226_cast_fp16_to_fp32_dtype_0, x = var_2226_cast_fp16)[name = tensor("cast_637")]; + tensor new_cache_9_internal_tensor_assign_2 = cast(dtype = reshape_49_cast_fp16_to_fp32_dtype_0, x = reshape_49_cast_fp16)[name = tensor("cast_638")]; + tensor var_1815 = cast(dtype = var_1815_cast_fp16_to_fp32_dtype_0, x = var_1815_cast_fp16)[name = tensor("cast_642")]; + tensor new_cache_7_internal_tensor_assign_2 = cast(dtype = reshape_39_cast_fp16_to_fp32_dtype_0, x = reshape_39_cast_fp16)[name = tensor("cast_643")]; + tensor var_1404 = cast(dtype = var_1404_cast_fp16_to_fp32_dtype_0, x = var_1404_cast_fp16)[name = tensor("cast_647")]; + tensor new_cache_5_internal_tensor_assign_2 = cast(dtype = reshape_29_cast_fp16_to_fp32_dtype_0, x = reshape_29_cast_fp16)[name = tensor("cast_648")]; + tensor var_993 = cast(dtype = var_993_cast_fp16_to_fp32_dtype_0, x = var_993_cast_fp16)[name = tensor("cast_652")]; + tensor new_cache_3_internal_tensor_assign_2 = cast(dtype = reshape_19_cast_fp16_to_fp32_dtype_0, x = reshape_19_cast_fp16)[name = tensor("cast_653")]; + tensor var_582 = cast(dtype = var_582_cast_fp16_to_fp32_dtype_0, x = var_582_cast_fp16)[name = tensor("cast_657")]; + tensor new_cache_1_internal_tensor_assign_2 = cast(dtype = reshape_9_cast_fp16_to_fp32_dtype_0, x = reshape_9_cast_fp16)[name = tensor("cast_658")]; + } -> (new_cache_1_internal_tensor_assign_2, var_582, new_cache_3_internal_tensor_assign_2, var_993, new_cache_5_internal_tensor_assign_2, var_1404, new_cache_7_internal_tensor_assign_2, var_1815, new_cache_9_internal_tensor_assign_2, var_2226, new_cache_11_internal_tensor_assign_2, var_2637, new_cache_13_internal_tensor_assign_2, var_3048, new_cache_15_internal_tensor_assign_2, var_3459, new_cache_17_internal_tensor_assign_2, var_3870, new_cache_19_internal_tensor_assign_2, var_4281, new_cache_21_internal_tensor_assign_2, var_4692, new_cache_23_internal_tensor_assign_2, var_5103, new_cache_25_internal_tensor_assign_2, var_5514, new_cache_27_internal_tensor_assign_2, var_5925, new_cache_29_internal_tensor_assign_2, var_6336, new_cache_31_internal_tensor_assign_2, var_6747, new_cache_33_internal_tensor_assign_2, var_7158, new_cache_35_internal_tensor_assign_2, var_7569, new_cache_37_internal_tensor_assign_2, var_7980, new_cache_39_internal_tensor_assign_2, var_8391, new_cache_41_internal_tensor_assign_2, var_8802, new_cache_43_internal_tensor_assign_2, var_9213, new_cache_45_internal_tensor_assign_2, var_9624, new_cache_internal_tensor_assign_2, var_9852); +} \ No newline at end of file diff --git a/v2.1/italian_24l/cond_prefill.mlmodelc/weights/weight.bin b/v2.1/italian_24l/cond_prefill.mlmodelc/weights/weight.bin new file mode 100644 index 0000000000000000000000000000000000000000..98fefe3aa7b4d827bb30ae5b7fe2daab009b19eb --- /dev/null +++ b/v2.1/italian_24l/cond_prefill.mlmodelc/weights/weight.bin @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:068de0b7dfbc5ce75cb11ef21cfa83fb46ea8dc24f260d2e2c9b141b5530ad96 +size 586381056 diff --git a/v2.1/italian_24l/cond_prefill.mlpackage/Data/com.apple.CoreML/model.mlmodel b/v2.1/italian_24l/cond_prefill.mlpackage/Data/com.apple.CoreML/model.mlmodel new file mode 100644 index 0000000000000000000000000000000000000000..268da7efeea355a10c5ab4596433eecee0210f29 --- /dev/null +++ b/v2.1/italian_24l/cond_prefill.mlpackage/Data/com.apple.CoreML/model.mlmodel @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:d9c9bc5c1cbf02734055941beb4a6b63b9e4237078662c47bd6f5011ec87c906 +size 842494 diff --git a/v2.1/italian_24l/cond_prefill.mlpackage/Data/com.apple.CoreML/weights/weight.bin b/v2.1/italian_24l/cond_prefill.mlpackage/Data/com.apple.CoreML/weights/weight.bin new file mode 100644 index 0000000000000000000000000000000000000000..98fefe3aa7b4d827bb30ae5b7fe2daab009b19eb --- /dev/null +++ b/v2.1/italian_24l/cond_prefill.mlpackage/Data/com.apple.CoreML/weights/weight.bin @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:068de0b7dfbc5ce75cb11ef21cfa83fb46ea8dc24f260d2e2c9b141b5530ad96 +size 586381056 diff --git a/v2.1/italian_24l/cond_prefill.mlpackage/Manifest.json b/v2.1/italian_24l/cond_prefill.mlpackage/Manifest.json new file mode 100644 index 0000000000000000000000000000000000000000..ba2028a3844aec99d29c7f91b806bc8542f4ab25 --- /dev/null +++ b/v2.1/italian_24l/cond_prefill.mlpackage/Manifest.json @@ -0,0 +1,18 @@ +{ + "fileFormatVersion": "1.0.0", + "itemInfoEntries": { + "39D8B6D1-8B05-4DA4-8B0B-A0AB5FE83970": { + "author": "com.apple.CoreML", + "description": "CoreML Model Specification", + "name": "model.mlmodel", + "path": "com.apple.CoreML/model.mlmodel" + }, + "662460B8-C1CD-4203-A841-7D1F915C7946": { + "author": "com.apple.CoreML", + "description": "CoreML Model Weights", + "name": "weights", + "path": "com.apple.CoreML/weights" + } + }, + "rootModelIdentifier": "39D8B6D1-8B05-4DA4-8B0B-A0AB5FE83970" +} diff --git a/v2.1/italian_24l/constants/bos_emb.npy b/v2.1/italian_24l/constants/bos_emb.npy new file mode 100644 index 0000000000000000000000000000000000000000..dfa073b09190015e4ef4327cfb77b221dfea365b --- /dev/null +++ b/v2.1/italian_24l/constants/bos_emb.npy @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:ba976ce845cc8087beda3659cbf40ddcc1aa21be6cc55c055b097286ca5d7462 +size 256 diff --git a/v2.1/italian_24l/constants/emb_mean.npy b/v2.1/italian_24l/constants/emb_mean.npy new file mode 100644 index 0000000000000000000000000000000000000000..caab83ff55325b240ef55fa9cd40b5edf04a0032 --- /dev/null +++ b/v2.1/italian_24l/constants/emb_mean.npy @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:80bbe48ccafadccde914579a7b6607cb9639f76f87ed3b29ae329b2ef3b7d2fd +size 256 diff --git a/v2.1/italian_24l/constants/emb_std.npy b/v2.1/italian_24l/constants/emb_std.npy new file mode 100644 index 0000000000000000000000000000000000000000..722ab578df338c4aea23acf36ce4beccbc7d5956 --- /dev/null +++ b/v2.1/italian_24l/constants/emb_std.npy @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:c3c68664acd8a79629d900c7a7155a1f0813b5118ae51a063b22d64d92ca8f71 +size 256 diff --git a/v2.1/italian_24l/constants/mimi_init_state.npz b/v2.1/italian_24l/constants/mimi_init_state.npz new file mode 100644 index 0000000000000000000000000000000000000000..3b52f383a86ef0439f9993506e4529c4c1234828 --- /dev/null +++ b/v2.1/italian_24l/constants/mimi_init_state.npz @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:515588e10e9c14a6b009f1cfda9fcd592542b1640ad856d5fc69e50024702a39 +size 1083108 diff --git a/v2.1/italian_24l/constants/quantizer_weight.npy b/v2.1/italian_24l/constants/quantizer_weight.npy new file mode 100644 index 0000000000000000000000000000000000000000..f3c0009e88c203a64687e9e0e278ef50f3e50f1e --- /dev/null +++ b/v2.1/italian_24l/constants/quantizer_weight.npy @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:c40a96dbde34920b6d468078ccbf796da40ebb94d259a20bf96b86e8635e6372 +size 65664 diff --git a/v2.1/italian_24l/constants/text_embed_table.npy b/v2.1/italian_24l/constants/text_embed_table.npy new file mode 100644 index 0000000000000000000000000000000000000000..482a3f299748ea0086c15b7703fcdca9b73d1b44 --- /dev/null +++ b/v2.1/italian_24l/constants/text_embed_table.npy @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:d9d340919a7d896c3750b805e9a1c7140f50d9871a0ec90c78f4d230867a9fe2 +size 16388224 diff --git a/v2.1/italian_24l/constants_bin/alba.safetensors b/v2.1/italian_24l/constants_bin/alba.safetensors new file mode 100644 index 0000000000000000000000000000000000000000..3d076def65a8863b7d4676cc2b8fb4c7ffd8fde1 --- /dev/null +++ b/v2.1/italian_24l/constants_bin/alba.safetensors @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:323a2f5757df5c003c54c34d3aaefb993d80565302683df6f9e883f8e4155d4e +size 24777760 diff --git a/v2.1/italian_24l/constants_bin/anna.safetensors b/v2.1/italian_24l/constants_bin/anna.safetensors new file mode 100644 index 0000000000000000000000000000000000000000..a0c4fab493c74afa77bc08f604159f6983140813 --- /dev/null +++ b/v2.1/italian_24l/constants_bin/anna.safetensors @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:79b769775905fc6308538f024945c4df826009aab899c1265814aa728e85fab2 +size 31265824 diff --git a/v2.1/italian_24l/constants_bin/azelma.safetensors b/v2.1/italian_24l/constants_bin/azelma.safetensors new file mode 100644 index 0000000000000000000000000000000000000000..aae61fc64d7cca49d6dd1f40f99d7ff131032f35 --- /dev/null +++ b/v2.1/italian_24l/constants_bin/azelma.safetensors @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:bb8b885c84cf25f027b65cf59dc00d9e362d4eb83340472647608ef59369f1db +size 31855648 diff --git a/v2.1/italian_24l/constants_bin/bill_boerst.safetensors b/v2.1/italian_24l/constants_bin/bill_boerst.safetensors new file mode 100644 index 0000000000000000000000000000000000000000..1357cce9067309fd0d424f8249c435192ff84384 --- /dev/null +++ b/v2.1/italian_24l/constants_bin/bill_boerst.safetensors @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:fa3f82bd776d5c6979760f81d0257cbeb07db20a3ebdd23efbc0175b62b7fa7d +size 26940448 diff --git a/v2.1/italian_24l/constants_bin/bos_before_voice.bin b/v2.1/italian_24l/constants_bin/bos_before_voice.bin new file mode 100644 index 0000000000000000000000000000000000000000..741720143b7f39781c7acb3c211adc41cc084ece --- /dev/null +++ b/v2.1/italian_24l/constants_bin/bos_before_voice.bin @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:152095d37393d3fab6eacd19430374a3df3b13400d581f1a56be61afba22382f +size 4096 diff --git a/v2.1/italian_24l/constants_bin/bos_emb.bin b/v2.1/italian_24l/constants_bin/bos_emb.bin new file mode 100644 index 0000000000000000000000000000000000000000..7447c36181b54fc00f8e8370fbebaebcdc96e62a --- /dev/null +++ b/v2.1/italian_24l/constants_bin/bos_emb.bin @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:8c2837f5e736445e37a0beb857d1cc710c55ef76727f38d8d4adb542a2d0c0d0 +size 128 diff --git a/v2.1/italian_24l/constants_bin/caro_davy.safetensors b/v2.1/italian_24l/constants_bin/caro_davy.safetensors new file mode 100644 index 0000000000000000000000000000000000000000..2890210db5d6ae273ee433e8de0ec441d9545ee8 --- /dev/null +++ b/v2.1/italian_24l/constants_bin/caro_davy.safetensors @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:b94a629320414ae7492147b24f4dcf613c55e6df22b0eb15b0d3bccf574f7766 +size 21042200 diff --git a/v2.1/italian_24l/constants_bin/charles.safetensors b/v2.1/italian_24l/constants_bin/charles.safetensors new file mode 100644 index 0000000000000000000000000000000000000000..64672a904fea7bf2b3fff7fb3c217f6854c9a27e --- /dev/null +++ b/v2.1/italian_24l/constants_bin/charles.safetensors @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:031a242165fff8e9c8f4750adbe5dabe0e481c98e4fb44a9734f023d8125bcb4 +size 24777760 diff --git a/v2.1/italian_24l/constants_bin/cosette.safetensors b/v2.1/italian_24l/constants_bin/cosette.safetensors new file mode 100644 index 0000000000000000000000000000000000000000..dba3a9be709668ff2d65ca6dc5e71981a2534725 --- /dev/null +++ b/v2.1/italian_24l/constants_bin/cosette.safetensors @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:9638432133eed16be5cd51a841910412c7939d846832bf225f8373848fb34c8a +size 24777760 diff --git a/v2.1/italian_24l/constants_bin/eponine.safetensors b/v2.1/italian_24l/constants_bin/eponine.safetensors new file mode 100644 index 0000000000000000000000000000000000000000..9bf57027817206c11163608dfa65cd5ec1a686a5 --- /dev/null +++ b/v2.1/italian_24l/constants_bin/eponine.safetensors @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:e105aab69178dee5cf4d09f8d532e6d25d97b3fa84ba71366628943d0f1031ca +size 27726880 diff --git a/v2.1/italian_24l/constants_bin/estelle.safetensors b/v2.1/italian_24l/constants_bin/estelle.safetensors new file mode 100644 index 0000000000000000000000000000000000000000..ab797312f4365b7852857256c3dbf64b13029ca1 --- /dev/null +++ b/v2.1/italian_24l/constants_bin/estelle.safetensors @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:bdbf9dd65b3872fd365851794f30074d41428299e9ae2a75f23353478282610e +size 33035296 diff --git a/v2.1/italian_24l/constants_bin/eve.safetensors b/v2.1/italian_24l/constants_bin/eve.safetensors new file mode 100644 index 0000000000000000000000000000000000000000..ee4050faaa94b445b5e005ae57c2406befb84d3c --- /dev/null +++ b/v2.1/italian_24l/constants_bin/eve.safetensors @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:4af73ac08e055f763553c023d15706ed0b70b718f2366e80eb11c5ab716b99e2 +size 26154016 diff --git a/v2.1/italian_24l/constants_bin/fantine.safetensors b/v2.1/italian_24l/constants_bin/fantine.safetensors new file mode 100644 index 0000000000000000000000000000000000000000..17fc69bf9b7316034c7d24cfaeea9b7db2d6f137 --- /dev/null +++ b/v2.1/italian_24l/constants_bin/fantine.safetensors @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:009a6ff7a582e8082a39c7363dc504b812f2ba41e45e8e6f1bb47fde35783cde +size 26154016 diff --git a/v2.1/italian_24l/constants_bin/george.safetensors b/v2.1/italian_24l/constants_bin/george.safetensors new file mode 100644 index 0000000000000000000000000000000000000000..2e4b05000ba22a73fd9f5f91228418c05d7bbc61 --- /dev/null +++ b/v2.1/italian_24l/constants_bin/george.safetensors @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:4a38a515b74b297edcbd4b91413aefaf0cba0f77e625cadbb678cee7c361913a +size 24974368 diff --git a/v2.1/italian_24l/constants_bin/giovanni.safetensors b/v2.1/italian_24l/constants_bin/giovanni.safetensors new file mode 100644 index 0000000000000000000000000000000000000000..c03f500351326406df61994e077099ff45327996 --- /dev/null +++ b/v2.1/italian_24l/constants_bin/giovanni.safetensors @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:3bd832fc10cfacc752956288101725626d133b456cb3c47551c2776fed6fe087 +size 18486272 diff --git a/v2.1/italian_24l/constants_bin/jane.safetensors b/v2.1/italian_24l/constants_bin/jane.safetensors new file mode 100644 index 0000000000000000000000000000000000000000..546373fb400f5b38ed54083154d37b14ad901192 --- /dev/null +++ b/v2.1/italian_24l/constants_bin/jane.safetensors @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:7db2a66e736084ec6385baa11d499360476843d2bee8bb5f23678d16d4c795a3 +size 29496352 diff --git a/v2.1/italian_24l/constants_bin/javert.safetensors b/v2.1/italian_24l/constants_bin/javert.safetensors new file mode 100644 index 0000000000000000000000000000000000000000..47d4b4abd533951ee9e123602aa2963729368515 --- /dev/null +++ b/v2.1/italian_24l/constants_bin/javert.safetensors @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:2bb9683a08c65d10af9beb820bebc65190e4ee601026e7d882ab8f5627e8e762 +size 24777760 diff --git a/v2.1/italian_24l/constants_bin/jean.safetensors b/v2.1/italian_24l/constants_bin/jean.safetensors new file mode 100644 index 0000000000000000000000000000000000000000..a741dd64cf8df95225fda0a7a0e3b904a389ea64 --- /dev/null +++ b/v2.1/italian_24l/constants_bin/jean.safetensors @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:76fdaaabc711f6f692bc09af60a2bcb96f6976c49a4f9608fa2a6140a8360ea0 +size 24777760 diff --git a/v2.1/italian_24l/constants_bin/juergen.safetensors b/v2.1/italian_24l/constants_bin/juergen.safetensors new file mode 100644 index 0000000000000000000000000000000000000000..a7953b82f75655e188ae288f946620d2ec73a33e --- /dev/null +++ b/v2.1/italian_24l/constants_bin/juergen.safetensors @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:c98a01e032690b30fc7ab7c914338684a6feda701739a63125e49fb84f0d6749 +size 24974368 diff --git a/v2.1/italian_24l/constants_bin/lola.safetensors b/v2.1/italian_24l/constants_bin/lola.safetensors new file mode 100644 index 0000000000000000000000000000000000000000..ac737118623e3437df8476aa36729325d8070d51 --- /dev/null +++ b/v2.1/italian_24l/constants_bin/lola.safetensors @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:7b46ad33f31ec4f46619f16dbb6052861c114a9eba9b033504f15bae7b2b69fd +size 23794712 diff --git a/v2.1/italian_24l/constants_bin/marius.safetensors b/v2.1/italian_24l/constants_bin/marius.safetensors new file mode 100644 index 0000000000000000000000000000000000000000..e832b3d2dc8a1b3f5510591bee1d3a6d278cab08 --- /dev/null +++ b/v2.1/italian_24l/constants_bin/marius.safetensors @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:647474a57af63a32243f7172ed26d0f8d8af69051eae0b606ab2ce6c72a5a292 +size 24777760 diff --git a/v2.1/italian_24l/constants_bin/mary.safetensors b/v2.1/italian_24l/constants_bin/mary.safetensors new file mode 100644 index 0000000000000000000000000000000000000000..8112230be1377810ae86ce811aae1d9c8ff428c7 --- /dev/null +++ b/v2.1/italian_24l/constants_bin/mary.safetensors @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:c1db72858c30459abed015fcf768afa85edf9fd84779a00a85f025ba65448bc6 +size 24777760 diff --git a/v2.1/italian_24l/constants_bin/michael.safetensors b/v2.1/italian_24l/constants_bin/michael.safetensors new file mode 100644 index 0000000000000000000000000000000000000000..f481fef2b3b80e4d67741871bb748ea9e6b78174 --- /dev/null +++ b/v2.1/italian_24l/constants_bin/michael.safetensors @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:ddb9751e053ce8087c760fab30a7e7926843dedc57f7c1ba6e894c31cbc1a07d +size 29103136 diff --git a/v2.1/italian_24l/constants_bin/paul.safetensors b/v2.1/italian_24l/constants_bin/paul.safetensors new file mode 100644 index 0000000000000000000000000000000000000000..15b9a6331fc6590f418c315b2b70cbebc21c66f5 --- /dev/null +++ b/v2.1/italian_24l/constants_bin/paul.safetensors @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:23dd4a5c1517e3fd5a7b54d16c2fe66442d80762d77ff8231f9b4cc551d3cf21 +size 27923488 diff --git a/v2.1/italian_24l/constants_bin/peter_yearsley.safetensors b/v2.1/italian_24l/constants_bin/peter_yearsley.safetensors new file mode 100644 index 0000000000000000000000000000000000000000..97a6f26c6bb36e86bde8ab3f5fe59b5f01b42d8f --- /dev/null +++ b/v2.1/italian_24l/constants_bin/peter_yearsley.safetensors @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:24d98b4ec96899d2b386c3868e3a3c27f9ac73c546012c91448e2c399b7ddbf5 +size 14947320 diff --git a/v2.1/italian_24l/constants_bin/rafael.safetensors b/v2.1/italian_24l/constants_bin/rafael.safetensors new file mode 100644 index 0000000000000000000000000000000000000000..e8f243cd02920d259b636875ed903d7368be884b --- /dev/null +++ b/v2.1/italian_24l/constants_bin/rafael.safetensors @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:b61aa92012eebe3e0de858a384a0043c01628bf7205b56f55f4dfa3752841b10 +size 24777760 diff --git a/v2.1/italian_24l/constants_bin/stuart_bell.safetensors b/v2.1/italian_24l/constants_bin/stuart_bell.safetensors new file mode 100644 index 0000000000000000000000000000000000000000..5f86c8769b9d0497d337dc78617dd72a6e0b9114 --- /dev/null +++ b/v2.1/italian_24l/constants_bin/stuart_bell.safetensors @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:dede3d55bddb07f8f5de215a19419ebb3861da20376b184242e01cc423666e77 +size 21042200 diff --git a/v2.1/italian_24l/constants_bin/text_embed_table.bin b/v2.1/italian_24l/constants_bin/text_embed_table.bin new file mode 100644 index 0000000000000000000000000000000000000000..28316a3053d488a63be870f01cdbbf97a6fbd479 --- /dev/null +++ b/v2.1/italian_24l/constants_bin/text_embed_table.bin @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:f8713722d79f795b24d0e40eb5b1cb60de4ae9349ffbfa502a2e9a34d85215a3 +size 16388096 diff --git a/v2.1/italian_24l/constants_bin/tokenizer.model b/v2.1/italian_24l/constants_bin/tokenizer.model new file mode 100644 index 0000000000000000000000000000000000000000..c2dab86ea3c7a064b38d657237cec3544371fe8b --- /dev/null +++ b/v2.1/italian_24l/constants_bin/tokenizer.model @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:6583b974a11b90e14d8a4c8e9c43f06c3861b9ede6e5023a4c27ab5a3a7d4c39 +size 60078 diff --git a/v2.1/italian_24l/constants_bin/vera.safetensors b/v2.1/italian_24l/constants_bin/vera.safetensors new file mode 100644 index 0000000000000000000000000000000000000000..f1442f7f5847ddde87016e523782cda12fdf195b --- /dev/null +++ b/v2.1/italian_24l/constants_bin/vera.safetensors @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:055b734607a1341b792f25c98d78c9af23f3dab0eb9e6261719865e47b9d7731 +size 26940448 diff --git a/v2.1/italian_24l/flow_decoder_fused.mlmodelc/analytics/coremldata.bin b/v2.1/italian_24l/flow_decoder_fused.mlmodelc/analytics/coremldata.bin new file mode 100644 index 0000000000000000000000000000000000000000..1272c49a22ce9a984442cd50a4f884e47cceaf6c --- /dev/null +++ b/v2.1/italian_24l/flow_decoder_fused.mlmodelc/analytics/coremldata.bin @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:b87f8097421e0252cbaa6e5bd773e033fb50a581aa47c1bf65ad394ce356ec65 +size 243 diff --git a/v2.1/italian_24l/flow_decoder_fused.mlmodelc/coremldata.bin b/v2.1/italian_24l/flow_decoder_fused.mlmodelc/coremldata.bin new file mode 100644 index 0000000000000000000000000000000000000000..0466e0f81a32f207e428000a6f0e2a0acba7a6e2 --- /dev/null +++ b/v2.1/italian_24l/flow_decoder_fused.mlmodelc/coremldata.bin @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:e2e537bd0fbdeebd68ea0bce9f4ac0b8acd9962cb671e049f864aef7c49e2294 +size 413 diff --git a/v2.1/italian_24l/flow_decoder_fused.mlmodelc/model.mil b/v2.1/italian_24l/flow_decoder_fused.mlmodelc/model.mil new file mode 100644 index 0000000000000000000000000000000000000000..c64a1f34555e7bbf9c8d4502d29f9ef52dd6b0ec --- /dev/null +++ b/v2.1/italian_24l/flow_decoder_fused.mlmodelc/model.mil @@ -0,0 +1,1915 @@ +program(1.0) +[buildInfo = dict, tensor>({{"coremlc-component-MIL", "3520.4.1"}, {"coremlc-version", "3520.5.1"}, {"coremltools-component-torch", "2.12.0"}, {"coremltools-source-dialect", "TorchScript"}, {"coremltools-version", "9.0"}})] +{ + func main(tensor latent_init, tensor transformer_out) { + tensor latent_init_to_fp16_dtype_0 = const()[name = tensor("latent_init_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor flow_net_input_proj_weight_to_fp16 = const()[name = tensor("flow_net_input_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(64)))]; + tensor flow_net_input_proj_bias_to_fp16 = const()[name = tensor("flow_net_input_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(32896)))]; + tensor latent_init_to_fp16 = cast(dtype = latent_init_to_fp16_dtype_0, x = latent_init)[name = tensor("cast_146")]; + tensor linear_0_cast_fp16 = linear(bias = flow_net_input_proj_bias_to_fp16, weight = flow_net_input_proj_weight_to_fp16, x = latent_init_to_fp16)[name = tensor("linear_0_cast_fp16")]; + tensor input_3_to_fp16 = const()[name = tensor("input_3_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(33984)))]; + tensor input_5_cast_fp16 = silu(x = input_3_to_fp16)[name = tensor("input_5_cast_fp16")]; + tensor flow_net_time_embed_0_mlp_2_weight_to_fp16 = const()[name = tensor("flow_net_time_embed_0_mlp_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(35072)))]; + tensor flow_net_time_embed_0_mlp_2_bias_to_fp16 = const()[name = tensor("flow_net_time_embed_0_mlp_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(559424)))]; + tensor linear_2_cast_fp16 = linear(bias = flow_net_time_embed_0_mlp_2_bias_to_fp16, weight = flow_net_time_embed_0_mlp_2_weight_to_fp16, x = input_5_cast_fp16)[name = tensor("linear_2_cast_fp16")]; + tensor reduce_mean_0_axes_0 = const()[name = tensor("reduce_mean_0_axes_0"), val = tensor([-1])]; + tensor reduce_mean_0_keep_dims_0 = const()[name = tensor("reduce_mean_0_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_0_cast_fp16 = reduce_mean(axes = reduce_mean_0_axes_0, keep_dims = reduce_mean_0_keep_dims_0, x = linear_2_cast_fp16)[name = tensor("reduce_mean_0_cast_fp16")]; + tensor sub_0_cast_fp16 = sub(x = linear_2_cast_fp16, y = reduce_mean_0_cast_fp16)[name = tensor("sub_0_cast_fp16")]; + tensor square_0_cast_fp16 = square(x = sub_0_cast_fp16)[name = tensor("square_0_cast_fp16")]; + tensor reduce_mean_1_axes_0 = const()[name = tensor("reduce_mean_1_axes_0"), val = tensor([-1])]; + tensor reduce_mean_1_keep_dims_0 = const()[name = tensor("reduce_mean_1_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_1_cast_fp16 = reduce_mean(axes = reduce_mean_1_axes_0, keep_dims = reduce_mean_1_keep_dims_0, x = square_0_cast_fp16)[name = tensor("reduce_mean_1_cast_fp16")]; + tensor real_div_0_to_fp16 = const()[name = tensor("real_div_0_to_fp16"), val = tensor(0x1.008p+0)]; + tensor mul_0_cast_fp16 = mul(x = reduce_mean_1_cast_fp16, y = real_div_0_to_fp16)[name = tensor("mul_0_cast_fp16")]; + tensor var_70_to_fp16 = const()[name = tensor("op_70_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_1_cast_fp16 = add(x = mul_0_cast_fp16, y = var_70_to_fp16)[name = tensor("var_1_cast_fp16")]; + tensor var_73_epsilon_0 = const()[name = tensor("op_73_epsilon_0"), val = tensor(0x1.197998p-40)]; + tensor var_73_cast_fp16 = rsqrt(epsilon = var_73_epsilon_0, x = var_1_cast_fp16)[name = tensor("op_73_cast_fp16")]; + tensor const_3_to_fp16 = const()[name = tensor("const_3_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(560512)))]; + tensor var_74_cast_fp16 = mul(x = const_3_to_fp16, y = var_73_cast_fp16)[name = tensor("op_74_cast_fp16")]; + tensor var_75_cast_fp16 = mul(x = linear_2_cast_fp16, y = var_74_cast_fp16)[name = tensor("op_75_cast_fp16")]; + tensor input_9_to_fp16 = const()[name = tensor("input_9_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(561600)))]; + tensor input_11_cast_fp16 = silu(x = input_9_to_fp16)[name = tensor("input_11_cast_fp16")]; + tensor flow_net_time_embed_1_mlp_2_weight_to_fp16 = const()[name = tensor("flow_net_time_embed_1_mlp_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(562688)))]; + tensor flow_net_time_embed_1_mlp_2_bias_to_fp16 = const()[name = tensor("flow_net_time_embed_1_mlp_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1087040)))]; + tensor linear_4_cast_fp16 = linear(bias = flow_net_time_embed_1_mlp_2_bias_to_fp16, weight = flow_net_time_embed_1_mlp_2_weight_to_fp16, x = input_11_cast_fp16)[name = tensor("linear_4_cast_fp16")]; + tensor reduce_mean_2_axes_0 = const()[name = tensor("reduce_mean_2_axes_0"), val = tensor([-1])]; + tensor reduce_mean_2_keep_dims_0 = const()[name = tensor("reduce_mean_2_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_2_cast_fp16 = reduce_mean(axes = reduce_mean_2_axes_0, keep_dims = reduce_mean_2_keep_dims_0, x = linear_4_cast_fp16)[name = tensor("reduce_mean_2_cast_fp16")]; + tensor sub_2_cast_fp16 = sub(x = linear_4_cast_fp16, y = reduce_mean_2_cast_fp16)[name = tensor("sub_2_cast_fp16")]; + tensor square_1_cast_fp16 = square(x = sub_2_cast_fp16)[name = tensor("square_1_cast_fp16")]; + tensor reduce_mean_3_axes_0 = const()[name = tensor("reduce_mean_3_axes_0"), val = tensor([-1])]; + tensor reduce_mean_3_keep_dims_0 = const()[name = tensor("reduce_mean_3_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_3_cast_fp16 = reduce_mean(axes = reduce_mean_3_axes_0, keep_dims = reduce_mean_3_keep_dims_0, x = square_1_cast_fp16)[name = tensor("reduce_mean_3_cast_fp16")]; + tensor real_div_1_to_fp16 = const()[name = tensor("real_div_1_to_fp16"), val = tensor(0x1.008p+0)]; + tensor mul_1_cast_fp16 = mul(x = reduce_mean_3_cast_fp16, y = real_div_1_to_fp16)[name = tensor("mul_1_cast_fp16")]; + tensor var_110_to_fp16 = const()[name = tensor("op_110_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_3_cast_fp16 = add(x = mul_1_cast_fp16, y = var_110_to_fp16)[name = tensor("var_3_cast_fp16")]; + tensor var_113_epsilon_0 = const()[name = tensor("op_113_epsilon_0"), val = tensor(0x1.197998p-40)]; + tensor var_113_cast_fp16 = rsqrt(epsilon = var_113_epsilon_0, x = var_3_cast_fp16)[name = tensor("op_113_cast_fp16")]; + tensor const_5_to_fp16 = const()[name = tensor("const_5_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1088128)))]; + tensor var_114_cast_fp16 = mul(x = const_5_to_fp16, y = var_113_cast_fp16)[name = tensor("op_114_cast_fp16")]; + tensor var_115_cast_fp16 = mul(x = linear_4_cast_fp16, y = var_114_cast_fp16)[name = tensor("op_115_cast_fp16")]; + tensor var_127_cast_fp16 = add(x = var_75_cast_fp16, y = var_115_cast_fp16)[name = tensor("op_127_cast_fp16")]; + tensor _inversed_t_combined_1_y_0_to_fp16 = const()[name = tensor("_inversed_t_combined_1_y_0_to_fp16"), val = tensor(0x1p-1)]; + tensor _inversed_t_combined_1_cast_fp16 = mul(x = var_127_cast_fp16, y = _inversed_t_combined_1_y_0_to_fp16)[name = tensor("_inversed_t_combined_1_cast_fp16")]; + tensor transformer_out_to_fp16_dtype_0 = const()[name = tensor("transformer_out_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor flow_net_cond_embed_weight_to_fp16 = const()[name = tensor("flow_net_cond_embed_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1089216)))]; + tensor flow_net_cond_embed_bias_to_fp16 = const()[name = tensor("flow_net_cond_embed_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(2137856)))]; + tensor transformer_out_to_fp16 = cast(dtype = transformer_out_to_fp16_dtype_0, x = transformer_out)[name = tensor("cast_145")]; + tensor linear_5_cast_fp16 = linear(bias = flow_net_cond_embed_bias_to_fp16, weight = flow_net_cond_embed_weight_to_fp16, x = transformer_out_to_fp16)[name = tensor("linear_5_cast_fp16")]; + tensor input_13_cast_fp16 = add(x = _inversed_t_combined_1_cast_fp16, y = linear_5_cast_fp16)[name = tensor("input_13_cast_fp16")]; + tensor input_15_cast_fp16 = silu(x = input_13_cast_fp16)[name = tensor("input_15_cast_fp16")]; + tensor flow_net_res_blocks_0_adaLN_modulation_1_weight_to_fp16 = const()[name = tensor("flow_net_res_blocks_0_adaLN_modulation_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(2138944)))]; + tensor flow_net_res_blocks_0_adaLN_modulation_1_bias_to_fp16 = const()[name = tensor("flow_net_res_blocks_0_adaLN_modulation_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(3711872)))]; + tensor linear_6_cast_fp16 = linear(bias = flow_net_res_blocks_0_adaLN_modulation_1_bias_to_fp16, weight = flow_net_res_blocks_0_adaLN_modulation_1_weight_to_fp16, x = input_15_cast_fp16)[name = tensor("linear_6_cast_fp16")]; + tensor var_142_split_sizes_0 = const()[name = tensor("op_142_split_sizes_0"), val = tensor([512, 512, 512])]; + tensor var_142_axis_0 = const()[name = tensor("op_142_axis_0"), val = tensor(-1)]; + tensor var_142_cast_fp16_0, tensor var_142_cast_fp16_1, tensor var_142_cast_fp16_2 = split(axis = var_142_axis_0, split_sizes = var_142_split_sizes_0, x = linear_6_cast_fp16)[name = tensor("op_142_cast_fp16")]; + tensor mean_1_axes_0 = const()[name = tensor("mean_1_axes_0"), val = tensor([-1])]; + tensor mean_1_keep_dims_0 = const()[name = tensor("mean_1_keep_dims_0"), val = tensor(true)]; + tensor mean_1_cast_fp16 = reduce_mean(axes = mean_1_axes_0, keep_dims = mean_1_keep_dims_0, x = linear_0_cast_fp16)[name = tensor("mean_1_cast_fp16")]; + tensor sub_4_cast_fp16 = sub(x = linear_0_cast_fp16, y = mean_1_cast_fp16)[name = tensor("sub_4_cast_fp16")]; + tensor square_2_cast_fp16 = square(x = sub_4_cast_fp16)[name = tensor("square_2_cast_fp16")]; + tensor reduce_mean_5_axes_0 = const()[name = tensor("reduce_mean_5_axes_0"), val = tensor([-1])]; + tensor reduce_mean_5_keep_dims_0 = const()[name = tensor("reduce_mean_5_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_5_cast_fp16 = reduce_mean(axes = reduce_mean_5_axes_0, keep_dims = reduce_mean_5_keep_dims_0, x = square_2_cast_fp16)[name = tensor("reduce_mean_5_cast_fp16")]; + tensor var_152_to_fp16 = const()[name = tensor("op_152_to_fp16"), val = tensor(0x1.1p-20)]; + tensor var_153_cast_fp16 = add(x = reduce_mean_5_cast_fp16, y = var_152_to_fp16)[name = tensor("op_153_cast_fp16")]; + tensor var_154_cast_fp16 = sqrt(x = var_153_cast_fp16)[name = tensor("op_154_cast_fp16")]; + tensor x_7_cast_fp16 = real_div(x = sub_4_cast_fp16, y = var_154_cast_fp16)[name = tensor("x_7_cast_fp16")]; + tensor flow_net_res_blocks_0_in_ln_weight_to_fp16 = const()[name = tensor("flow_net_res_blocks_0_in_ln_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(3715008)))]; + tensor var_156_cast_fp16 = mul(x = x_7_cast_fp16, y = flow_net_res_blocks_0_in_ln_weight_to_fp16)[name = tensor("op_156_cast_fp16")]; + tensor flow_net_res_blocks_0_in_ln_bias_to_fp16 = const()[name = tensor("flow_net_res_blocks_0_in_ln_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(3716096)))]; + tensor x_9_cast_fp16 = add(x = var_156_cast_fp16, y = flow_net_res_blocks_0_in_ln_bias_to_fp16)[name = tensor("x_9_cast_fp16")]; + tensor var_158_promoted_to_fp16 = const()[name = tensor("op_158_promoted_to_fp16"), val = tensor(0x1p+0)]; + tensor var_159_cast_fp16 = add(x = var_142_cast_fp16_1, y = var_158_promoted_to_fp16)[name = tensor("op_159_cast_fp16")]; + tensor var_160_cast_fp16 = mul(x = x_9_cast_fp16, y = var_159_cast_fp16)[name = tensor("op_160_cast_fp16")]; + tensor input_17_cast_fp16 = add(x = var_160_cast_fp16, y = var_142_cast_fp16_0)[name = tensor("input_17_cast_fp16")]; + tensor flow_net_res_blocks_0_mlp_0_weight_to_fp16 = const()[name = tensor("flow_net_res_blocks_0_mlp_0_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(3717184)))]; + tensor flow_net_res_blocks_0_mlp_0_bias_to_fp16 = const()[name = tensor("flow_net_res_blocks_0_mlp_0_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(4241536)))]; + tensor linear_7_cast_fp16 = linear(bias = flow_net_res_blocks_0_mlp_0_bias_to_fp16, weight = flow_net_res_blocks_0_mlp_0_weight_to_fp16, x = input_17_cast_fp16)[name = tensor("linear_7_cast_fp16")]; + tensor input_21_cast_fp16 = silu(x = linear_7_cast_fp16)[name = tensor("input_21_cast_fp16")]; + tensor flow_net_res_blocks_0_mlp_2_weight_to_fp16 = const()[name = tensor("flow_net_res_blocks_0_mlp_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(4242624)))]; + tensor flow_net_res_blocks_0_mlp_2_bias_to_fp16 = const()[name = tensor("flow_net_res_blocks_0_mlp_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(4766976)))]; + tensor linear_8_cast_fp16 = linear(bias = flow_net_res_blocks_0_mlp_2_bias_to_fp16, weight = flow_net_res_blocks_0_mlp_2_weight_to_fp16, x = input_21_cast_fp16)[name = tensor("linear_8_cast_fp16")]; + tensor var_171_cast_fp16 = mul(x = var_142_cast_fp16_2, y = linear_8_cast_fp16)[name = tensor("op_171_cast_fp16")]; + tensor x_11_cast_fp16 = add(x = linear_0_cast_fp16, y = var_171_cast_fp16)[name = tensor("x_11_cast_fp16")]; + tensor flow_net_res_blocks_1_adaLN_modulation_1_weight_to_fp16 = const()[name = tensor("flow_net_res_blocks_1_adaLN_modulation_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(4768064)))]; + tensor flow_net_res_blocks_1_adaLN_modulation_1_bias_to_fp16 = const()[name = tensor("flow_net_res_blocks_1_adaLN_modulation_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(6340992)))]; + tensor linear_9_cast_fp16 = linear(bias = flow_net_res_blocks_1_adaLN_modulation_1_bias_to_fp16, weight = flow_net_res_blocks_1_adaLN_modulation_1_weight_to_fp16, x = input_15_cast_fp16)[name = tensor("linear_9_cast_fp16")]; + tensor var_181_split_sizes_0 = const()[name = tensor("op_181_split_sizes_0"), val = tensor([512, 512, 512])]; + tensor var_181_axis_0 = const()[name = tensor("op_181_axis_0"), val = tensor(-1)]; + tensor var_181_cast_fp16_0, tensor var_181_cast_fp16_1, tensor var_181_cast_fp16_2 = split(axis = var_181_axis_0, split_sizes = var_181_split_sizes_0, x = linear_9_cast_fp16)[name = tensor("op_181_cast_fp16")]; + tensor mean_3_axes_0 = const()[name = tensor("mean_3_axes_0"), val = tensor([-1])]; + tensor mean_3_keep_dims_0 = const()[name = tensor("mean_3_keep_dims_0"), val = tensor(true)]; + tensor mean_3_cast_fp16 = reduce_mean(axes = mean_3_axes_0, keep_dims = mean_3_keep_dims_0, x = x_11_cast_fp16)[name = tensor("mean_3_cast_fp16")]; + tensor sub_5_cast_fp16 = sub(x = x_11_cast_fp16, y = mean_3_cast_fp16)[name = tensor("sub_5_cast_fp16")]; + tensor square_3_cast_fp16 = square(x = sub_5_cast_fp16)[name = tensor("square_3_cast_fp16")]; + tensor reduce_mean_7_axes_0 = const()[name = tensor("reduce_mean_7_axes_0"), val = tensor([-1])]; + tensor reduce_mean_7_keep_dims_0 = const()[name = tensor("reduce_mean_7_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_7_cast_fp16 = reduce_mean(axes = reduce_mean_7_axes_0, keep_dims = reduce_mean_7_keep_dims_0, x = square_3_cast_fp16)[name = tensor("reduce_mean_7_cast_fp16")]; + tensor var_191_to_fp16 = const()[name = tensor("op_191_to_fp16"), val = tensor(0x1.1p-20)]; + tensor var_192_cast_fp16 = add(x = reduce_mean_7_cast_fp16, y = var_191_to_fp16)[name = tensor("op_192_cast_fp16")]; + tensor var_193_cast_fp16 = sqrt(x = var_192_cast_fp16)[name = tensor("op_193_cast_fp16")]; + tensor x_13_cast_fp16 = real_div(x = sub_5_cast_fp16, y = var_193_cast_fp16)[name = tensor("x_13_cast_fp16")]; + tensor flow_net_res_blocks_1_in_ln_weight_to_fp16 = const()[name = tensor("flow_net_res_blocks_1_in_ln_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(6344128)))]; + tensor var_195_cast_fp16 = mul(x = x_13_cast_fp16, y = flow_net_res_blocks_1_in_ln_weight_to_fp16)[name = tensor("op_195_cast_fp16")]; + tensor flow_net_res_blocks_1_in_ln_bias_to_fp16 = const()[name = tensor("flow_net_res_blocks_1_in_ln_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(6345216)))]; + tensor x_15_cast_fp16 = add(x = var_195_cast_fp16, y = flow_net_res_blocks_1_in_ln_bias_to_fp16)[name = tensor("x_15_cast_fp16")]; + tensor var_197_promoted_to_fp16 = const()[name = tensor("op_197_promoted_to_fp16"), val = tensor(0x1p+0)]; + tensor var_198_cast_fp16 = add(x = var_181_cast_fp16_1, y = var_197_promoted_to_fp16)[name = tensor("op_198_cast_fp16")]; + tensor var_199_cast_fp16 = mul(x = x_15_cast_fp16, y = var_198_cast_fp16)[name = tensor("op_199_cast_fp16")]; + tensor input_25_cast_fp16 = add(x = var_199_cast_fp16, y = var_181_cast_fp16_0)[name = tensor("input_25_cast_fp16")]; + tensor flow_net_res_blocks_1_mlp_0_weight_to_fp16 = const()[name = tensor("flow_net_res_blocks_1_mlp_0_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(6346304)))]; + tensor flow_net_res_blocks_1_mlp_0_bias_to_fp16 = const()[name = tensor("flow_net_res_blocks_1_mlp_0_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(6870656)))]; + tensor linear_10_cast_fp16 = linear(bias = flow_net_res_blocks_1_mlp_0_bias_to_fp16, weight = flow_net_res_blocks_1_mlp_0_weight_to_fp16, x = input_25_cast_fp16)[name = tensor("linear_10_cast_fp16")]; + tensor input_29_cast_fp16 = silu(x = linear_10_cast_fp16)[name = tensor("input_29_cast_fp16")]; + tensor flow_net_res_blocks_1_mlp_2_weight_to_fp16 = const()[name = tensor("flow_net_res_blocks_1_mlp_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(6871744)))]; + tensor flow_net_res_blocks_1_mlp_2_bias_to_fp16 = const()[name = tensor("flow_net_res_blocks_1_mlp_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(7396096)))]; + tensor linear_11_cast_fp16 = linear(bias = flow_net_res_blocks_1_mlp_2_bias_to_fp16, weight = flow_net_res_blocks_1_mlp_2_weight_to_fp16, x = input_29_cast_fp16)[name = tensor("linear_11_cast_fp16")]; + tensor var_210_cast_fp16 = mul(x = var_181_cast_fp16_2, y = linear_11_cast_fp16)[name = tensor("op_210_cast_fp16")]; + tensor x_17_cast_fp16 = add(x = x_11_cast_fp16, y = var_210_cast_fp16)[name = tensor("x_17_cast_fp16")]; + tensor flow_net_res_blocks_2_adaLN_modulation_1_weight_to_fp16 = const()[name = tensor("flow_net_res_blocks_2_adaLN_modulation_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(7397184)))]; + tensor flow_net_res_blocks_2_adaLN_modulation_1_bias_to_fp16 = const()[name = tensor("flow_net_res_blocks_2_adaLN_modulation_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8970112)))]; + tensor linear_12_cast_fp16 = linear(bias = flow_net_res_blocks_2_adaLN_modulation_1_bias_to_fp16, weight = flow_net_res_blocks_2_adaLN_modulation_1_weight_to_fp16, x = input_15_cast_fp16)[name = tensor("linear_12_cast_fp16")]; + tensor var_220_split_sizes_0 = const()[name = tensor("op_220_split_sizes_0"), val = tensor([512, 512, 512])]; + tensor var_220_axis_0 = const()[name = tensor("op_220_axis_0"), val = tensor(-1)]; + tensor var_220_cast_fp16_0, tensor var_220_cast_fp16_1, tensor var_220_cast_fp16_2 = split(axis = var_220_axis_0, split_sizes = var_220_split_sizes_0, x = linear_12_cast_fp16)[name = tensor("op_220_cast_fp16")]; + tensor mean_5_axes_0 = const()[name = tensor("mean_5_axes_0"), val = tensor([-1])]; + tensor mean_5_keep_dims_0 = const()[name = tensor("mean_5_keep_dims_0"), val = tensor(true)]; + tensor mean_5_cast_fp16 = reduce_mean(axes = mean_5_axes_0, keep_dims = mean_5_keep_dims_0, x = x_17_cast_fp16)[name = tensor("mean_5_cast_fp16")]; + tensor sub_6_cast_fp16 = sub(x = x_17_cast_fp16, y = mean_5_cast_fp16)[name = tensor("sub_6_cast_fp16")]; + tensor square_4_cast_fp16 = square(x = sub_6_cast_fp16)[name = tensor("square_4_cast_fp16")]; + tensor reduce_mean_9_axes_0 = const()[name = tensor("reduce_mean_9_axes_0"), val = tensor([-1])]; + tensor reduce_mean_9_keep_dims_0 = const()[name = tensor("reduce_mean_9_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_9_cast_fp16 = reduce_mean(axes = reduce_mean_9_axes_0, keep_dims = reduce_mean_9_keep_dims_0, x = square_4_cast_fp16)[name = tensor("reduce_mean_9_cast_fp16")]; + tensor var_230_to_fp16 = const()[name = tensor("op_230_to_fp16"), val = tensor(0x1.1p-20)]; + tensor var_231_cast_fp16 = add(x = reduce_mean_9_cast_fp16, y = var_230_to_fp16)[name = tensor("op_231_cast_fp16")]; + tensor var_232_cast_fp16 = sqrt(x = var_231_cast_fp16)[name = tensor("op_232_cast_fp16")]; + tensor x_19_cast_fp16 = real_div(x = sub_6_cast_fp16, y = var_232_cast_fp16)[name = tensor("x_19_cast_fp16")]; + tensor flow_net_res_blocks_2_in_ln_weight_to_fp16 = const()[name = tensor("flow_net_res_blocks_2_in_ln_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8973248)))]; + tensor var_234_cast_fp16 = mul(x = x_19_cast_fp16, y = flow_net_res_blocks_2_in_ln_weight_to_fp16)[name = tensor("op_234_cast_fp16")]; + tensor flow_net_res_blocks_2_in_ln_bias_to_fp16 = const()[name = tensor("flow_net_res_blocks_2_in_ln_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8974336)))]; + tensor x_21_cast_fp16 = add(x = var_234_cast_fp16, y = flow_net_res_blocks_2_in_ln_bias_to_fp16)[name = tensor("x_21_cast_fp16")]; + tensor var_236_promoted_to_fp16 = const()[name = tensor("op_236_promoted_to_fp16"), val = tensor(0x1p+0)]; + tensor var_237_cast_fp16 = add(x = var_220_cast_fp16_1, y = var_236_promoted_to_fp16)[name = tensor("op_237_cast_fp16")]; + tensor var_238_cast_fp16 = mul(x = x_21_cast_fp16, y = var_237_cast_fp16)[name = tensor("op_238_cast_fp16")]; + tensor input_33_cast_fp16 = add(x = var_238_cast_fp16, y = var_220_cast_fp16_0)[name = tensor("input_33_cast_fp16")]; + tensor flow_net_res_blocks_2_mlp_0_weight_to_fp16 = const()[name = tensor("flow_net_res_blocks_2_mlp_0_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8975424)))]; + tensor flow_net_res_blocks_2_mlp_0_bias_to_fp16 = const()[name = tensor("flow_net_res_blocks_2_mlp_0_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9499776)))]; + tensor linear_13_cast_fp16 = linear(bias = flow_net_res_blocks_2_mlp_0_bias_to_fp16, weight = flow_net_res_blocks_2_mlp_0_weight_to_fp16, x = input_33_cast_fp16)[name = tensor("linear_13_cast_fp16")]; + tensor input_37_cast_fp16 = silu(x = linear_13_cast_fp16)[name = tensor("input_37_cast_fp16")]; + tensor flow_net_res_blocks_2_mlp_2_weight_to_fp16 = const()[name = tensor("flow_net_res_blocks_2_mlp_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9500864)))]; + tensor flow_net_res_blocks_2_mlp_2_bias_to_fp16 = const()[name = tensor("flow_net_res_blocks_2_mlp_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(10025216)))]; + tensor linear_14_cast_fp16 = linear(bias = flow_net_res_blocks_2_mlp_2_bias_to_fp16, weight = flow_net_res_blocks_2_mlp_2_weight_to_fp16, x = input_37_cast_fp16)[name = tensor("linear_14_cast_fp16")]; + tensor var_249_cast_fp16 = mul(x = var_220_cast_fp16_2, y = linear_14_cast_fp16)[name = tensor("op_249_cast_fp16")]; + tensor x_23_cast_fp16 = add(x = x_17_cast_fp16, y = var_249_cast_fp16)[name = tensor("x_23_cast_fp16")]; + tensor flow_net_res_blocks_3_adaLN_modulation_1_weight_to_fp16 = const()[name = tensor("flow_net_res_blocks_3_adaLN_modulation_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(10026304)))]; + tensor flow_net_res_blocks_3_adaLN_modulation_1_bias_to_fp16 = const()[name = tensor("flow_net_res_blocks_3_adaLN_modulation_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(11599232)))]; + tensor linear_15_cast_fp16 = linear(bias = flow_net_res_blocks_3_adaLN_modulation_1_bias_to_fp16, weight = flow_net_res_blocks_3_adaLN_modulation_1_weight_to_fp16, x = input_15_cast_fp16)[name = tensor("linear_15_cast_fp16")]; + tensor var_259_split_sizes_0 = const()[name = tensor("op_259_split_sizes_0"), val = tensor([512, 512, 512])]; + tensor var_259_axis_0 = const()[name = tensor("op_259_axis_0"), val = tensor(-1)]; + tensor var_259_cast_fp16_0, tensor var_259_cast_fp16_1, tensor var_259_cast_fp16_2 = split(axis = var_259_axis_0, split_sizes = var_259_split_sizes_0, x = linear_15_cast_fp16)[name = tensor("op_259_cast_fp16")]; + tensor mean_7_axes_0 = const()[name = tensor("mean_7_axes_0"), val = tensor([-1])]; + tensor mean_7_keep_dims_0 = const()[name = tensor("mean_7_keep_dims_0"), val = tensor(true)]; + tensor mean_7_cast_fp16 = reduce_mean(axes = mean_7_axes_0, keep_dims = mean_7_keep_dims_0, x = x_23_cast_fp16)[name = tensor("mean_7_cast_fp16")]; + tensor sub_7_cast_fp16 = sub(x = x_23_cast_fp16, y = mean_7_cast_fp16)[name = tensor("sub_7_cast_fp16")]; + tensor square_5_cast_fp16 = square(x = sub_7_cast_fp16)[name = tensor("square_5_cast_fp16")]; + tensor reduce_mean_11_axes_0 = const()[name = tensor("reduce_mean_11_axes_0"), val = tensor([-1])]; + tensor reduce_mean_11_keep_dims_0 = const()[name = tensor("reduce_mean_11_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_11_cast_fp16 = reduce_mean(axes = reduce_mean_11_axes_0, keep_dims = reduce_mean_11_keep_dims_0, x = square_5_cast_fp16)[name = tensor("reduce_mean_11_cast_fp16")]; + tensor var_269_to_fp16 = const()[name = tensor("op_269_to_fp16"), val = tensor(0x1.1p-20)]; + tensor var_270_cast_fp16 = add(x = reduce_mean_11_cast_fp16, y = var_269_to_fp16)[name = tensor("op_270_cast_fp16")]; + tensor var_271_cast_fp16 = sqrt(x = var_270_cast_fp16)[name = tensor("op_271_cast_fp16")]; + tensor x_25_cast_fp16 = real_div(x = sub_7_cast_fp16, y = var_271_cast_fp16)[name = tensor("x_25_cast_fp16")]; + tensor flow_net_res_blocks_3_in_ln_weight_to_fp16 = const()[name = tensor("flow_net_res_blocks_3_in_ln_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(11602368)))]; + tensor var_273_cast_fp16 = mul(x = x_25_cast_fp16, y = flow_net_res_blocks_3_in_ln_weight_to_fp16)[name = tensor("op_273_cast_fp16")]; + tensor flow_net_res_blocks_3_in_ln_bias_to_fp16 = const()[name = tensor("flow_net_res_blocks_3_in_ln_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(11603456)))]; + tensor x_27_cast_fp16 = add(x = var_273_cast_fp16, y = flow_net_res_blocks_3_in_ln_bias_to_fp16)[name = tensor("x_27_cast_fp16")]; + tensor var_275_promoted_to_fp16 = const()[name = tensor("op_275_promoted_to_fp16"), val = tensor(0x1p+0)]; + tensor var_276_cast_fp16 = add(x = var_259_cast_fp16_1, y = var_275_promoted_to_fp16)[name = tensor("op_276_cast_fp16")]; + tensor var_277_cast_fp16 = mul(x = x_27_cast_fp16, y = var_276_cast_fp16)[name = tensor("op_277_cast_fp16")]; + tensor input_41_cast_fp16 = add(x = var_277_cast_fp16, y = var_259_cast_fp16_0)[name = tensor("input_41_cast_fp16")]; + tensor flow_net_res_blocks_3_mlp_0_weight_to_fp16 = const()[name = tensor("flow_net_res_blocks_3_mlp_0_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(11604544)))]; + tensor flow_net_res_blocks_3_mlp_0_bias_to_fp16 = const()[name = tensor("flow_net_res_blocks_3_mlp_0_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(12128896)))]; + tensor linear_16_cast_fp16 = linear(bias = flow_net_res_blocks_3_mlp_0_bias_to_fp16, weight = flow_net_res_blocks_3_mlp_0_weight_to_fp16, x = input_41_cast_fp16)[name = tensor("linear_16_cast_fp16")]; + tensor input_45_cast_fp16 = silu(x = linear_16_cast_fp16)[name = tensor("input_45_cast_fp16")]; + tensor flow_net_res_blocks_3_mlp_2_weight_to_fp16 = const()[name = tensor("flow_net_res_blocks_3_mlp_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(12129984)))]; + tensor flow_net_res_blocks_3_mlp_2_bias_to_fp16 = const()[name = tensor("flow_net_res_blocks_3_mlp_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(12654336)))]; + tensor linear_17_cast_fp16 = linear(bias = flow_net_res_blocks_3_mlp_2_bias_to_fp16, weight = flow_net_res_blocks_3_mlp_2_weight_to_fp16, x = input_45_cast_fp16)[name = tensor("linear_17_cast_fp16")]; + tensor var_288_cast_fp16 = mul(x = var_259_cast_fp16_2, y = linear_17_cast_fp16)[name = tensor("op_288_cast_fp16")]; + tensor x_29_cast_fp16 = add(x = x_23_cast_fp16, y = var_288_cast_fp16)[name = tensor("x_29_cast_fp16")]; + tensor flow_net_res_blocks_4_adaLN_modulation_1_weight_to_fp16 = const()[name = tensor("flow_net_res_blocks_4_adaLN_modulation_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(12655424)))]; + tensor flow_net_res_blocks_4_adaLN_modulation_1_bias_to_fp16 = const()[name = tensor("flow_net_res_blocks_4_adaLN_modulation_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(14228352)))]; + tensor linear_18_cast_fp16 = linear(bias = flow_net_res_blocks_4_adaLN_modulation_1_bias_to_fp16, weight = flow_net_res_blocks_4_adaLN_modulation_1_weight_to_fp16, x = input_15_cast_fp16)[name = tensor("linear_18_cast_fp16")]; + tensor var_298_split_sizes_0 = const()[name = tensor("op_298_split_sizes_0"), val = tensor([512, 512, 512])]; + tensor var_298_axis_0 = const()[name = tensor("op_298_axis_0"), val = tensor(-1)]; + tensor var_298_cast_fp16_0, tensor var_298_cast_fp16_1, tensor var_298_cast_fp16_2 = split(axis = var_298_axis_0, split_sizes = var_298_split_sizes_0, x = linear_18_cast_fp16)[name = tensor("op_298_cast_fp16")]; + tensor mean_9_axes_0 = const()[name = tensor("mean_9_axes_0"), val = tensor([-1])]; + tensor mean_9_keep_dims_0 = const()[name = tensor("mean_9_keep_dims_0"), val = tensor(true)]; + tensor mean_9_cast_fp16 = reduce_mean(axes = mean_9_axes_0, keep_dims = mean_9_keep_dims_0, x = x_29_cast_fp16)[name = tensor("mean_9_cast_fp16")]; + tensor sub_8_cast_fp16 = sub(x = x_29_cast_fp16, y = mean_9_cast_fp16)[name = tensor("sub_8_cast_fp16")]; + tensor square_6_cast_fp16 = square(x = sub_8_cast_fp16)[name = tensor("square_6_cast_fp16")]; + tensor reduce_mean_13_axes_0 = const()[name = tensor("reduce_mean_13_axes_0"), val = tensor([-1])]; + tensor reduce_mean_13_keep_dims_0 = const()[name = tensor("reduce_mean_13_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_13_cast_fp16 = reduce_mean(axes = reduce_mean_13_axes_0, keep_dims = reduce_mean_13_keep_dims_0, x = square_6_cast_fp16)[name = tensor("reduce_mean_13_cast_fp16")]; + tensor var_308_to_fp16 = const()[name = tensor("op_308_to_fp16"), val = tensor(0x1.1p-20)]; + tensor var_309_cast_fp16 = add(x = reduce_mean_13_cast_fp16, y = var_308_to_fp16)[name = tensor("op_309_cast_fp16")]; + tensor var_310_cast_fp16 = sqrt(x = var_309_cast_fp16)[name = tensor("op_310_cast_fp16")]; + tensor x_31_cast_fp16 = real_div(x = sub_8_cast_fp16, y = var_310_cast_fp16)[name = tensor("x_31_cast_fp16")]; + tensor flow_net_res_blocks_4_in_ln_weight_to_fp16 = const()[name = tensor("flow_net_res_blocks_4_in_ln_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(14231488)))]; + tensor var_312_cast_fp16 = mul(x = x_31_cast_fp16, y = flow_net_res_blocks_4_in_ln_weight_to_fp16)[name = tensor("op_312_cast_fp16")]; + tensor flow_net_res_blocks_4_in_ln_bias_to_fp16 = const()[name = tensor("flow_net_res_blocks_4_in_ln_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(14232576)))]; + tensor x_33_cast_fp16 = add(x = var_312_cast_fp16, y = flow_net_res_blocks_4_in_ln_bias_to_fp16)[name = tensor("x_33_cast_fp16")]; + tensor var_314_promoted_to_fp16 = const()[name = tensor("op_314_promoted_to_fp16"), val = tensor(0x1p+0)]; + tensor var_315_cast_fp16 = add(x = var_298_cast_fp16_1, y = var_314_promoted_to_fp16)[name = tensor("op_315_cast_fp16")]; + tensor var_316_cast_fp16 = mul(x = x_33_cast_fp16, y = var_315_cast_fp16)[name = tensor("op_316_cast_fp16")]; + tensor input_49_cast_fp16 = add(x = var_316_cast_fp16, y = var_298_cast_fp16_0)[name = tensor("input_49_cast_fp16")]; + tensor flow_net_res_blocks_4_mlp_0_weight_to_fp16 = const()[name = tensor("flow_net_res_blocks_4_mlp_0_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(14233664)))]; + tensor flow_net_res_blocks_4_mlp_0_bias_to_fp16 = const()[name = tensor("flow_net_res_blocks_4_mlp_0_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(14758016)))]; + tensor linear_19_cast_fp16 = linear(bias = flow_net_res_blocks_4_mlp_0_bias_to_fp16, weight = flow_net_res_blocks_4_mlp_0_weight_to_fp16, x = input_49_cast_fp16)[name = tensor("linear_19_cast_fp16")]; + tensor input_53_cast_fp16 = silu(x = linear_19_cast_fp16)[name = tensor("input_53_cast_fp16")]; + tensor flow_net_res_blocks_4_mlp_2_weight_to_fp16 = const()[name = tensor("flow_net_res_blocks_4_mlp_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(14759104)))]; + tensor flow_net_res_blocks_4_mlp_2_bias_to_fp16 = const()[name = tensor("flow_net_res_blocks_4_mlp_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15283456)))]; + tensor linear_20_cast_fp16 = linear(bias = flow_net_res_blocks_4_mlp_2_bias_to_fp16, weight = flow_net_res_blocks_4_mlp_2_weight_to_fp16, x = input_53_cast_fp16)[name = tensor("linear_20_cast_fp16")]; + tensor var_327_cast_fp16 = mul(x = var_298_cast_fp16_2, y = linear_20_cast_fp16)[name = tensor("op_327_cast_fp16")]; + tensor x_35_cast_fp16 = add(x = x_29_cast_fp16, y = var_327_cast_fp16)[name = tensor("x_35_cast_fp16")]; + tensor flow_net_res_blocks_5_adaLN_modulation_1_weight_to_fp16 = const()[name = tensor("flow_net_res_blocks_5_adaLN_modulation_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15284544)))]; + tensor flow_net_res_blocks_5_adaLN_modulation_1_bias_to_fp16 = const()[name = tensor("flow_net_res_blocks_5_adaLN_modulation_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(16857472)))]; + tensor linear_21_cast_fp16 = linear(bias = flow_net_res_blocks_5_adaLN_modulation_1_bias_to_fp16, weight = flow_net_res_blocks_5_adaLN_modulation_1_weight_to_fp16, x = input_15_cast_fp16)[name = tensor("linear_21_cast_fp16")]; + tensor var_337_split_sizes_0 = const()[name = tensor("op_337_split_sizes_0"), val = tensor([512, 512, 512])]; + tensor var_337_axis_0 = const()[name = tensor("op_337_axis_0"), val = tensor(-1)]; + tensor var_337_cast_fp16_0, tensor var_337_cast_fp16_1, tensor var_337_cast_fp16_2 = split(axis = var_337_axis_0, split_sizes = var_337_split_sizes_0, x = linear_21_cast_fp16)[name = tensor("op_337_cast_fp16")]; + tensor mean_11_axes_0 = const()[name = tensor("mean_11_axes_0"), val = tensor([-1])]; + tensor mean_11_keep_dims_0 = const()[name = tensor("mean_11_keep_dims_0"), val = tensor(true)]; + tensor mean_11_cast_fp16 = reduce_mean(axes = mean_11_axes_0, keep_dims = mean_11_keep_dims_0, x = x_35_cast_fp16)[name = tensor("mean_11_cast_fp16")]; + tensor sub_9_cast_fp16 = sub(x = x_35_cast_fp16, y = mean_11_cast_fp16)[name = tensor("sub_9_cast_fp16")]; + tensor square_7_cast_fp16 = square(x = sub_9_cast_fp16)[name = tensor("square_7_cast_fp16")]; + tensor reduce_mean_15_axes_0 = const()[name = tensor("reduce_mean_15_axes_0"), val = tensor([-1])]; + tensor reduce_mean_15_keep_dims_0 = const()[name = tensor("reduce_mean_15_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_15_cast_fp16 = reduce_mean(axes = reduce_mean_15_axes_0, keep_dims = reduce_mean_15_keep_dims_0, x = square_7_cast_fp16)[name = tensor("reduce_mean_15_cast_fp16")]; + tensor var_347_to_fp16 = const()[name = tensor("op_347_to_fp16"), val = tensor(0x1.1p-20)]; + tensor var_348_cast_fp16 = add(x = reduce_mean_15_cast_fp16, y = var_347_to_fp16)[name = tensor("op_348_cast_fp16")]; + tensor var_349_cast_fp16 = sqrt(x = var_348_cast_fp16)[name = tensor("op_349_cast_fp16")]; + tensor x_37_cast_fp16 = real_div(x = sub_9_cast_fp16, y = var_349_cast_fp16)[name = tensor("x_37_cast_fp16")]; + tensor flow_net_res_blocks_5_in_ln_weight_to_fp16 = const()[name = tensor("flow_net_res_blocks_5_in_ln_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(16860608)))]; + tensor var_351_cast_fp16 = mul(x = x_37_cast_fp16, y = flow_net_res_blocks_5_in_ln_weight_to_fp16)[name = tensor("op_351_cast_fp16")]; + tensor flow_net_res_blocks_5_in_ln_bias_to_fp16 = const()[name = tensor("flow_net_res_blocks_5_in_ln_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(16861696)))]; + tensor x_39_cast_fp16 = add(x = var_351_cast_fp16, y = flow_net_res_blocks_5_in_ln_bias_to_fp16)[name = tensor("x_39_cast_fp16")]; + tensor var_353_promoted_to_fp16 = const()[name = tensor("op_353_promoted_to_fp16"), val = tensor(0x1p+0)]; + tensor var_354_cast_fp16 = add(x = var_337_cast_fp16_1, y = var_353_promoted_to_fp16)[name = tensor("op_354_cast_fp16")]; + tensor var_355_cast_fp16 = mul(x = x_39_cast_fp16, y = var_354_cast_fp16)[name = tensor("op_355_cast_fp16")]; + tensor input_57_cast_fp16 = add(x = var_355_cast_fp16, y = var_337_cast_fp16_0)[name = tensor("input_57_cast_fp16")]; + tensor flow_net_res_blocks_5_mlp_0_weight_to_fp16 = const()[name = tensor("flow_net_res_blocks_5_mlp_0_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(16862784)))]; + tensor flow_net_res_blocks_5_mlp_0_bias_to_fp16 = const()[name = tensor("flow_net_res_blocks_5_mlp_0_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(17387136)))]; + tensor linear_22_cast_fp16 = linear(bias = flow_net_res_blocks_5_mlp_0_bias_to_fp16, weight = flow_net_res_blocks_5_mlp_0_weight_to_fp16, x = input_57_cast_fp16)[name = tensor("linear_22_cast_fp16")]; + tensor input_61_cast_fp16 = silu(x = linear_22_cast_fp16)[name = tensor("input_61_cast_fp16")]; + tensor flow_net_res_blocks_5_mlp_2_weight_to_fp16 = const()[name = tensor("flow_net_res_blocks_5_mlp_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(17388224)))]; + tensor flow_net_res_blocks_5_mlp_2_bias_to_fp16 = const()[name = tensor("flow_net_res_blocks_5_mlp_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(17912576)))]; + tensor linear_23_cast_fp16 = linear(bias = flow_net_res_blocks_5_mlp_2_bias_to_fp16, weight = flow_net_res_blocks_5_mlp_2_weight_to_fp16, x = input_61_cast_fp16)[name = tensor("linear_23_cast_fp16")]; + tensor var_366_cast_fp16 = mul(x = var_337_cast_fp16_2, y = linear_23_cast_fp16)[name = tensor("op_366_cast_fp16")]; + tensor x_41_cast_fp16 = add(x = x_35_cast_fp16, y = var_366_cast_fp16)[name = tensor("x_41_cast_fp16")]; + tensor flow_net_final_layer_adaLN_modulation_1_weight_to_fp16 = const()[name = tensor("flow_net_final_layer_adaLN_modulation_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(17913664)))]; + tensor flow_net_final_layer_adaLN_modulation_1_bias_to_fp16 = const()[name = tensor("flow_net_final_layer_adaLN_modulation_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(18962304)))]; + tensor linear_24_cast_fp16 = linear(bias = flow_net_final_layer_adaLN_modulation_1_bias_to_fp16, weight = flow_net_final_layer_adaLN_modulation_1_weight_to_fp16, x = input_15_cast_fp16)[name = tensor("linear_24_cast_fp16")]; + tensor var_375_split_sizes_0 = const()[name = tensor("op_375_split_sizes_0"), val = tensor([512, 512])]; + tensor var_375_axis_0 = const()[name = tensor("op_375_axis_0"), val = tensor(-1)]; + tensor var_375_cast_fp16_0, tensor var_375_cast_fp16_1 = split(axis = var_375_axis_0, split_sizes = var_375_split_sizes_0, x = linear_24_cast_fp16)[name = tensor("op_375_cast_fp16")]; + tensor mean_13_axes_0 = const()[name = tensor("mean_13_axes_0"), val = tensor([-1])]; + tensor mean_13_keep_dims_0 = const()[name = tensor("mean_13_keep_dims_0"), val = tensor(true)]; + tensor mean_13_cast_fp16 = reduce_mean(axes = mean_13_axes_0, keep_dims = mean_13_keep_dims_0, x = x_41_cast_fp16)[name = tensor("mean_13_cast_fp16")]; + tensor sub_10_cast_fp16 = sub(x = x_41_cast_fp16, y = mean_13_cast_fp16)[name = tensor("sub_10_cast_fp16")]; + tensor square_8_cast_fp16 = square(x = sub_10_cast_fp16)[name = tensor("square_8_cast_fp16")]; + tensor reduce_mean_17_axes_0 = const()[name = tensor("reduce_mean_17_axes_0"), val = tensor([-1])]; + tensor reduce_mean_17_keep_dims_0 = const()[name = tensor("reduce_mean_17_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_17_cast_fp16 = reduce_mean(axes = reduce_mean_17_axes_0, keep_dims = reduce_mean_17_keep_dims_0, x = square_8_cast_fp16)[name = tensor("reduce_mean_17_cast_fp16")]; + tensor var_382_to_fp16 = const()[name = tensor("op_382_to_fp16"), val = tensor(0x1.1p-20)]; + tensor var_383_cast_fp16 = add(x = reduce_mean_17_cast_fp16, y = var_382_to_fp16)[name = tensor("op_383_cast_fp16")]; + tensor var_384_cast_fp16 = sqrt(x = var_383_cast_fp16)[name = tensor("op_384_cast_fp16")]; + tensor x_43_cast_fp16 = real_div(x = sub_10_cast_fp16, y = var_384_cast_fp16)[name = tensor("x_43_cast_fp16")]; + tensor var_386_promoted_to_fp16 = const()[name = tensor("op_386_promoted_to_fp16"), val = tensor(0x1p+0)]; + tensor var_387_cast_fp16 = add(x = var_375_cast_fp16_1, y = var_386_promoted_to_fp16)[name = tensor("op_387_cast_fp16")]; + tensor var_388_cast_fp16 = mul(x = x_43_cast_fp16, y = var_387_cast_fp16)[name = tensor("op_388_cast_fp16")]; + tensor input_65_cast_fp16 = add(x = var_388_cast_fp16, y = var_375_cast_fp16_0)[name = tensor("input_65_cast_fp16")]; + tensor flow_net_final_layer_linear_weight_to_fp16 = const()[name = tensor("flow_net_final_layer_linear_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(18964416)))]; + tensor flow_net_final_layer_linear_bias_to_fp16 = const()[name = tensor("flow_net_final_layer_linear_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(18997248)))]; + tensor linear_25_cast_fp16 = linear(bias = flow_net_final_layer_linear_bias_to_fp16, weight = flow_net_final_layer_linear_weight_to_fp16, x = input_65_cast_fp16)[name = tensor("linear_25_cast_fp16")]; + tensor var_399_to_fp16 = const()[name = tensor("op_399_to_fp16"), val = tensor(0x1p-3)]; + tensor var_400_cast_fp16 = mul(x = linear_25_cast_fp16, y = var_399_to_fp16)[name = tensor("op_400_cast_fp16")]; + tensor input_67_cast_fp16 = add(x = latent_init_to_fp16, y = var_400_cast_fp16)[name = tensor("input_67_cast_fp16")]; + tensor linear_26_cast_fp16 = linear(bias = flow_net_input_proj_bias_to_fp16, weight = flow_net_input_proj_weight_to_fp16, x = input_67_cast_fp16)[name = tensor("linear_26_cast_fp16")]; + tensor input_71_to_fp16 = const()[name = tensor("input_71_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(18997376)))]; + tensor input_73_cast_fp16 = silu(x = input_71_to_fp16)[name = tensor("input_73_cast_fp16")]; + tensor linear_28_cast_fp16 = linear(bias = flow_net_time_embed_0_mlp_2_bias_to_fp16, weight = flow_net_time_embed_0_mlp_2_weight_to_fp16, x = input_73_cast_fp16)[name = tensor("linear_28_cast_fp16")]; + tensor reduce_mean_18_axes_0 = const()[name = tensor("reduce_mean_18_axes_0"), val = tensor([-1])]; + tensor reduce_mean_18_keep_dims_0 = const()[name = tensor("reduce_mean_18_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_18_cast_fp16 = reduce_mean(axes = reduce_mean_18_axes_0, keep_dims = reduce_mean_18_keep_dims_0, x = linear_28_cast_fp16)[name = tensor("reduce_mean_18_cast_fp16")]; + tensor sub_11_cast_fp16 = sub(x = linear_28_cast_fp16, y = reduce_mean_18_cast_fp16)[name = tensor("sub_11_cast_fp16")]; + tensor square_9_cast_fp16 = square(x = sub_11_cast_fp16)[name = tensor("square_9_cast_fp16")]; + tensor reduce_mean_19_axes_0 = const()[name = tensor("reduce_mean_19_axes_0"), val = tensor([-1])]; + tensor reduce_mean_19_keep_dims_0 = const()[name = tensor("reduce_mean_19_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_19_cast_fp16 = reduce_mean(axes = reduce_mean_19_axes_0, keep_dims = reduce_mean_19_keep_dims_0, x = square_9_cast_fp16)[name = tensor("reduce_mean_19_cast_fp16")]; + tensor real_div_2_to_fp16 = const()[name = tensor("real_div_2_to_fp16"), val = tensor(0x1.008p+0)]; + tensor mul_2_cast_fp16 = mul(x = reduce_mean_19_cast_fp16, y = real_div_2_to_fp16)[name = tensor("mul_2_cast_fp16")]; + tensor var_466_to_fp16 = const()[name = tensor("op_466_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_19_cast_fp16 = add(x = mul_2_cast_fp16, y = var_466_to_fp16)[name = tensor("var_19_cast_fp16")]; + tensor var_469_epsilon_0 = const()[name = tensor("op_469_epsilon_0"), val = tensor(0x1.197998p-40)]; + tensor var_469_cast_fp16 = rsqrt(epsilon = var_469_epsilon_0, x = var_19_cast_fp16)[name = tensor("op_469_cast_fp16")]; + tensor var_470_cast_fp16 = mul(x = const_3_to_fp16, y = var_469_cast_fp16)[name = tensor("op_470_cast_fp16")]; + tensor var_471_cast_fp16 = mul(x = linear_28_cast_fp16, y = var_470_cast_fp16)[name = tensor("op_471_cast_fp16")]; + tensor input_77_to_fp16 = const()[name = tensor("input_77_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(18998464)))]; + tensor input_79_cast_fp16 = silu(x = input_77_to_fp16)[name = tensor("input_79_cast_fp16")]; + tensor linear_30_cast_fp16 = linear(bias = flow_net_time_embed_1_mlp_2_bias_to_fp16, weight = flow_net_time_embed_1_mlp_2_weight_to_fp16, x = input_79_cast_fp16)[name = tensor("linear_30_cast_fp16")]; + tensor reduce_mean_20_axes_0 = const()[name = tensor("reduce_mean_20_axes_0"), val = tensor([-1])]; + tensor reduce_mean_20_keep_dims_0 = const()[name = tensor("reduce_mean_20_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_20_cast_fp16 = reduce_mean(axes = reduce_mean_20_axes_0, keep_dims = reduce_mean_20_keep_dims_0, x = linear_30_cast_fp16)[name = tensor("reduce_mean_20_cast_fp16")]; + tensor sub_13_cast_fp16 = sub(x = linear_30_cast_fp16, y = reduce_mean_20_cast_fp16)[name = tensor("sub_13_cast_fp16")]; + tensor square_10_cast_fp16 = square(x = sub_13_cast_fp16)[name = tensor("square_10_cast_fp16")]; + tensor reduce_mean_21_axes_0 = const()[name = tensor("reduce_mean_21_axes_0"), val = tensor([-1])]; + tensor reduce_mean_21_keep_dims_0 = const()[name = tensor("reduce_mean_21_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_21_cast_fp16 = reduce_mean(axes = reduce_mean_21_axes_0, keep_dims = reduce_mean_21_keep_dims_0, x = square_10_cast_fp16)[name = tensor("reduce_mean_21_cast_fp16")]; + tensor real_div_3_to_fp16 = const()[name = tensor("real_div_3_to_fp16"), val = tensor(0x1.008p+0)]; + tensor mul_3_cast_fp16 = mul(x = reduce_mean_21_cast_fp16, y = real_div_3_to_fp16)[name = tensor("mul_3_cast_fp16")]; + tensor var_503_to_fp16 = const()[name = tensor("op_503_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_21_cast_fp16 = add(x = mul_3_cast_fp16, y = var_503_to_fp16)[name = tensor("var_21_cast_fp16")]; + tensor var_506_epsilon_0 = const()[name = tensor("op_506_epsilon_0"), val = tensor(0x1.197998p-40)]; + tensor var_506_cast_fp16 = rsqrt(epsilon = var_506_epsilon_0, x = var_21_cast_fp16)[name = tensor("op_506_cast_fp16")]; + tensor var_507_cast_fp16 = mul(x = const_5_to_fp16, y = var_506_cast_fp16)[name = tensor("op_507_cast_fp16")]; + tensor var_508_cast_fp16 = mul(x = linear_30_cast_fp16, y = var_507_cast_fp16)[name = tensor("op_508_cast_fp16")]; + tensor var_520_cast_fp16 = add(x = var_471_cast_fp16, y = var_508_cast_fp16)[name = tensor("op_520_cast_fp16")]; + tensor _inversed_t_combined_3_y_0_to_fp16 = const()[name = tensor("_inversed_t_combined_3_y_0_to_fp16"), val = tensor(0x1p-1)]; + tensor _inversed_t_combined_3_cast_fp16 = mul(x = var_520_cast_fp16, y = _inversed_t_combined_3_y_0_to_fp16)[name = tensor("_inversed_t_combined_3_cast_fp16")]; + tensor input_81_cast_fp16 = add(x = _inversed_t_combined_3_cast_fp16, y = linear_5_cast_fp16)[name = tensor("input_81_cast_fp16")]; + tensor input_83_cast_fp16 = silu(x = input_81_cast_fp16)[name = tensor("input_83_cast_fp16")]; + tensor linear_32_cast_fp16 = linear(bias = flow_net_res_blocks_0_adaLN_modulation_1_bias_to_fp16, weight = flow_net_res_blocks_0_adaLN_modulation_1_weight_to_fp16, x = input_83_cast_fp16)[name = tensor("linear_32_cast_fp16")]; + tensor var_535_split_sizes_0 = const()[name = tensor("op_535_split_sizes_0"), val = tensor([512, 512, 512])]; + tensor var_535_axis_0 = const()[name = tensor("op_535_axis_0"), val = tensor(-1)]; + tensor var_535_cast_fp16_0, tensor var_535_cast_fp16_1, tensor var_535_cast_fp16_2 = split(axis = var_535_axis_0, split_sizes = var_535_split_sizes_0, x = linear_32_cast_fp16)[name = tensor("op_535_cast_fp16")]; + tensor mean_15_axes_0 = const()[name = tensor("mean_15_axes_0"), val = tensor([-1])]; + tensor mean_15_keep_dims_0 = const()[name = tensor("mean_15_keep_dims_0"), val = tensor(true)]; + tensor mean_15_cast_fp16 = reduce_mean(axes = mean_15_axes_0, keep_dims = mean_15_keep_dims_0, x = linear_26_cast_fp16)[name = tensor("mean_15_cast_fp16")]; + tensor sub_15_cast_fp16 = sub(x = linear_26_cast_fp16, y = mean_15_cast_fp16)[name = tensor("sub_15_cast_fp16")]; + tensor square_11_cast_fp16 = square(x = sub_15_cast_fp16)[name = tensor("square_11_cast_fp16")]; + tensor reduce_mean_23_axes_0 = const()[name = tensor("reduce_mean_23_axes_0"), val = tensor([-1])]; + tensor reduce_mean_23_keep_dims_0 = const()[name = tensor("reduce_mean_23_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_23_cast_fp16 = reduce_mean(axes = reduce_mean_23_axes_0, keep_dims = reduce_mean_23_keep_dims_0, x = square_11_cast_fp16)[name = tensor("reduce_mean_23_cast_fp16")]; + tensor var_545_to_fp16 = const()[name = tensor("op_545_to_fp16"), val = tensor(0x1.1p-20)]; + tensor var_546_cast_fp16 = add(x = reduce_mean_23_cast_fp16, y = var_545_to_fp16)[name = tensor("op_546_cast_fp16")]; + tensor var_547_cast_fp16 = sqrt(x = var_546_cast_fp16)[name = tensor("op_547_cast_fp16")]; + tensor x_51_cast_fp16 = real_div(x = sub_15_cast_fp16, y = var_547_cast_fp16)[name = tensor("x_51_cast_fp16")]; + tensor var_549_cast_fp16 = mul(x = x_51_cast_fp16, y = flow_net_res_blocks_0_in_ln_weight_to_fp16)[name = tensor("op_549_cast_fp16")]; + tensor x_53_cast_fp16 = add(x = var_549_cast_fp16, y = flow_net_res_blocks_0_in_ln_bias_to_fp16)[name = tensor("x_53_cast_fp16")]; + tensor var_551_promoted_to_fp16 = const()[name = tensor("op_551_promoted_to_fp16"), val = tensor(0x1p+0)]; + tensor var_552_cast_fp16 = add(x = var_535_cast_fp16_1, y = var_551_promoted_to_fp16)[name = tensor("op_552_cast_fp16")]; + tensor var_553_cast_fp16 = mul(x = x_53_cast_fp16, y = var_552_cast_fp16)[name = tensor("op_553_cast_fp16")]; + tensor input_85_cast_fp16 = add(x = var_553_cast_fp16, y = var_535_cast_fp16_0)[name = tensor("input_85_cast_fp16")]; + tensor linear_33_cast_fp16 = linear(bias = flow_net_res_blocks_0_mlp_0_bias_to_fp16, weight = flow_net_res_blocks_0_mlp_0_weight_to_fp16, x = input_85_cast_fp16)[name = tensor("linear_33_cast_fp16")]; + tensor input_89_cast_fp16 = silu(x = linear_33_cast_fp16)[name = tensor("input_89_cast_fp16")]; + tensor linear_34_cast_fp16 = linear(bias = flow_net_res_blocks_0_mlp_2_bias_to_fp16, weight = flow_net_res_blocks_0_mlp_2_weight_to_fp16, x = input_89_cast_fp16)[name = tensor("linear_34_cast_fp16")]; + tensor var_564_cast_fp16 = mul(x = var_535_cast_fp16_2, y = linear_34_cast_fp16)[name = tensor("op_564_cast_fp16")]; + tensor x_55_cast_fp16 = add(x = linear_26_cast_fp16, y = var_564_cast_fp16)[name = tensor("x_55_cast_fp16")]; + tensor linear_35_cast_fp16 = linear(bias = flow_net_res_blocks_1_adaLN_modulation_1_bias_to_fp16, weight = flow_net_res_blocks_1_adaLN_modulation_1_weight_to_fp16, x = input_83_cast_fp16)[name = tensor("linear_35_cast_fp16")]; + tensor var_574_split_sizes_0 = const()[name = tensor("op_574_split_sizes_0"), val = tensor([512, 512, 512])]; + tensor var_574_axis_0 = const()[name = tensor("op_574_axis_0"), val = tensor(-1)]; + tensor var_574_cast_fp16_0, tensor var_574_cast_fp16_1, tensor var_574_cast_fp16_2 = split(axis = var_574_axis_0, split_sizes = var_574_split_sizes_0, x = linear_35_cast_fp16)[name = tensor("op_574_cast_fp16")]; + tensor mean_17_axes_0 = const()[name = tensor("mean_17_axes_0"), val = tensor([-1])]; + tensor mean_17_keep_dims_0 = const()[name = tensor("mean_17_keep_dims_0"), val = tensor(true)]; + tensor mean_17_cast_fp16 = reduce_mean(axes = mean_17_axes_0, keep_dims = mean_17_keep_dims_0, x = x_55_cast_fp16)[name = tensor("mean_17_cast_fp16")]; + tensor sub_16_cast_fp16 = sub(x = x_55_cast_fp16, y = mean_17_cast_fp16)[name = tensor("sub_16_cast_fp16")]; + tensor square_12_cast_fp16 = square(x = sub_16_cast_fp16)[name = tensor("square_12_cast_fp16")]; + tensor reduce_mean_25_axes_0 = const()[name = tensor("reduce_mean_25_axes_0"), val = tensor([-1])]; + tensor reduce_mean_25_keep_dims_0 = const()[name = tensor("reduce_mean_25_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_25_cast_fp16 = reduce_mean(axes = reduce_mean_25_axes_0, keep_dims = reduce_mean_25_keep_dims_0, x = square_12_cast_fp16)[name = tensor("reduce_mean_25_cast_fp16")]; + tensor var_584_to_fp16 = const()[name = tensor("op_584_to_fp16"), val = tensor(0x1.1p-20)]; + tensor var_585_cast_fp16 = add(x = reduce_mean_25_cast_fp16, y = var_584_to_fp16)[name = tensor("op_585_cast_fp16")]; + tensor var_586_cast_fp16 = sqrt(x = var_585_cast_fp16)[name = tensor("op_586_cast_fp16")]; + tensor x_57_cast_fp16 = real_div(x = sub_16_cast_fp16, y = var_586_cast_fp16)[name = tensor("x_57_cast_fp16")]; + tensor var_588_cast_fp16 = mul(x = x_57_cast_fp16, y = flow_net_res_blocks_1_in_ln_weight_to_fp16)[name = tensor("op_588_cast_fp16")]; + tensor x_59_cast_fp16 = add(x = var_588_cast_fp16, y = flow_net_res_blocks_1_in_ln_bias_to_fp16)[name = tensor("x_59_cast_fp16")]; + tensor var_590_promoted_to_fp16 = const()[name = tensor("op_590_promoted_to_fp16"), val = tensor(0x1p+0)]; + tensor var_591_cast_fp16 = add(x = var_574_cast_fp16_1, y = var_590_promoted_to_fp16)[name = tensor("op_591_cast_fp16")]; + tensor var_592_cast_fp16 = mul(x = x_59_cast_fp16, y = var_591_cast_fp16)[name = tensor("op_592_cast_fp16")]; + tensor input_93_cast_fp16 = add(x = var_592_cast_fp16, y = var_574_cast_fp16_0)[name = tensor("input_93_cast_fp16")]; + tensor linear_36_cast_fp16 = linear(bias = flow_net_res_blocks_1_mlp_0_bias_to_fp16, weight = flow_net_res_blocks_1_mlp_0_weight_to_fp16, x = input_93_cast_fp16)[name = tensor("linear_36_cast_fp16")]; + tensor input_97_cast_fp16 = silu(x = linear_36_cast_fp16)[name = tensor("input_97_cast_fp16")]; + tensor linear_37_cast_fp16 = linear(bias = flow_net_res_blocks_1_mlp_2_bias_to_fp16, weight = flow_net_res_blocks_1_mlp_2_weight_to_fp16, x = input_97_cast_fp16)[name = tensor("linear_37_cast_fp16")]; + tensor var_603_cast_fp16 = mul(x = var_574_cast_fp16_2, y = linear_37_cast_fp16)[name = tensor("op_603_cast_fp16")]; + tensor x_61_cast_fp16 = add(x = x_55_cast_fp16, y = var_603_cast_fp16)[name = tensor("x_61_cast_fp16")]; + tensor linear_38_cast_fp16 = linear(bias = flow_net_res_blocks_2_adaLN_modulation_1_bias_to_fp16, weight = flow_net_res_blocks_2_adaLN_modulation_1_weight_to_fp16, x = input_83_cast_fp16)[name = tensor("linear_38_cast_fp16")]; + tensor var_613_split_sizes_0 = const()[name = tensor("op_613_split_sizes_0"), val = tensor([512, 512, 512])]; + tensor var_613_axis_0 = const()[name = tensor("op_613_axis_0"), val = tensor(-1)]; + tensor var_613_cast_fp16_0, tensor var_613_cast_fp16_1, tensor var_613_cast_fp16_2 = split(axis = var_613_axis_0, split_sizes = var_613_split_sizes_0, x = linear_38_cast_fp16)[name = tensor("op_613_cast_fp16")]; + tensor mean_19_axes_0 = const()[name = tensor("mean_19_axes_0"), val = tensor([-1])]; + tensor mean_19_keep_dims_0 = const()[name = tensor("mean_19_keep_dims_0"), val = tensor(true)]; + tensor mean_19_cast_fp16 = reduce_mean(axes = mean_19_axes_0, keep_dims = mean_19_keep_dims_0, x = x_61_cast_fp16)[name = tensor("mean_19_cast_fp16")]; + tensor sub_17_cast_fp16 = sub(x = x_61_cast_fp16, y = mean_19_cast_fp16)[name = tensor("sub_17_cast_fp16")]; + tensor square_13_cast_fp16 = square(x = sub_17_cast_fp16)[name = tensor("square_13_cast_fp16")]; + tensor reduce_mean_27_axes_0 = const()[name = tensor("reduce_mean_27_axes_0"), val = tensor([-1])]; + tensor reduce_mean_27_keep_dims_0 = const()[name = tensor("reduce_mean_27_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_27_cast_fp16 = reduce_mean(axes = reduce_mean_27_axes_0, keep_dims = reduce_mean_27_keep_dims_0, x = square_13_cast_fp16)[name = tensor("reduce_mean_27_cast_fp16")]; + tensor var_623_to_fp16 = const()[name = tensor("op_623_to_fp16"), val = tensor(0x1.1p-20)]; + tensor var_624_cast_fp16 = add(x = reduce_mean_27_cast_fp16, y = var_623_to_fp16)[name = tensor("op_624_cast_fp16")]; + tensor var_625_cast_fp16 = sqrt(x = var_624_cast_fp16)[name = tensor("op_625_cast_fp16")]; + tensor x_63_cast_fp16 = real_div(x = sub_17_cast_fp16, y = var_625_cast_fp16)[name = tensor("x_63_cast_fp16")]; + tensor var_627_cast_fp16 = mul(x = x_63_cast_fp16, y = flow_net_res_blocks_2_in_ln_weight_to_fp16)[name = tensor("op_627_cast_fp16")]; + tensor x_65_cast_fp16 = add(x = var_627_cast_fp16, y = flow_net_res_blocks_2_in_ln_bias_to_fp16)[name = tensor("x_65_cast_fp16")]; + tensor var_629_promoted_to_fp16 = const()[name = tensor("op_629_promoted_to_fp16"), val = tensor(0x1p+0)]; + tensor var_630_cast_fp16 = add(x = var_613_cast_fp16_1, y = var_629_promoted_to_fp16)[name = tensor("op_630_cast_fp16")]; + tensor var_631_cast_fp16 = mul(x = x_65_cast_fp16, y = var_630_cast_fp16)[name = tensor("op_631_cast_fp16")]; + tensor input_101_cast_fp16 = add(x = var_631_cast_fp16, y = var_613_cast_fp16_0)[name = tensor("input_101_cast_fp16")]; + tensor linear_39_cast_fp16 = linear(bias = flow_net_res_blocks_2_mlp_0_bias_to_fp16, weight = flow_net_res_blocks_2_mlp_0_weight_to_fp16, x = input_101_cast_fp16)[name = tensor("linear_39_cast_fp16")]; + tensor input_105_cast_fp16 = silu(x = linear_39_cast_fp16)[name = tensor("input_105_cast_fp16")]; + tensor linear_40_cast_fp16 = linear(bias = flow_net_res_blocks_2_mlp_2_bias_to_fp16, weight = flow_net_res_blocks_2_mlp_2_weight_to_fp16, x = input_105_cast_fp16)[name = tensor("linear_40_cast_fp16")]; + tensor var_642_cast_fp16 = mul(x = var_613_cast_fp16_2, y = linear_40_cast_fp16)[name = tensor("op_642_cast_fp16")]; + tensor x_67_cast_fp16 = add(x = x_61_cast_fp16, y = var_642_cast_fp16)[name = tensor("x_67_cast_fp16")]; + tensor linear_41_cast_fp16 = linear(bias = flow_net_res_blocks_3_adaLN_modulation_1_bias_to_fp16, weight = flow_net_res_blocks_3_adaLN_modulation_1_weight_to_fp16, x = input_83_cast_fp16)[name = tensor("linear_41_cast_fp16")]; + tensor var_652_split_sizes_0 = const()[name = tensor("op_652_split_sizes_0"), val = tensor([512, 512, 512])]; + tensor var_652_axis_0 = const()[name = tensor("op_652_axis_0"), val = tensor(-1)]; + tensor var_652_cast_fp16_0, tensor var_652_cast_fp16_1, tensor var_652_cast_fp16_2 = split(axis = var_652_axis_0, split_sizes = var_652_split_sizes_0, x = linear_41_cast_fp16)[name = tensor("op_652_cast_fp16")]; + tensor mean_21_axes_0 = const()[name = tensor("mean_21_axes_0"), val = tensor([-1])]; + tensor mean_21_keep_dims_0 = const()[name = tensor("mean_21_keep_dims_0"), val = tensor(true)]; + tensor mean_21_cast_fp16 = reduce_mean(axes = mean_21_axes_0, keep_dims = mean_21_keep_dims_0, x = x_67_cast_fp16)[name = tensor("mean_21_cast_fp16")]; + tensor sub_18_cast_fp16 = sub(x = x_67_cast_fp16, y = mean_21_cast_fp16)[name = tensor("sub_18_cast_fp16")]; + tensor square_14_cast_fp16 = square(x = sub_18_cast_fp16)[name = tensor("square_14_cast_fp16")]; + tensor reduce_mean_29_axes_0 = const()[name = tensor("reduce_mean_29_axes_0"), val = tensor([-1])]; + tensor reduce_mean_29_keep_dims_0 = const()[name = tensor("reduce_mean_29_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_29_cast_fp16 = reduce_mean(axes = reduce_mean_29_axes_0, keep_dims = reduce_mean_29_keep_dims_0, x = square_14_cast_fp16)[name = tensor("reduce_mean_29_cast_fp16")]; + tensor var_662_to_fp16 = const()[name = tensor("op_662_to_fp16"), val = tensor(0x1.1p-20)]; + tensor var_663_cast_fp16 = add(x = reduce_mean_29_cast_fp16, y = var_662_to_fp16)[name = tensor("op_663_cast_fp16")]; + tensor var_664_cast_fp16 = sqrt(x = var_663_cast_fp16)[name = tensor("op_664_cast_fp16")]; + tensor x_69_cast_fp16 = real_div(x = sub_18_cast_fp16, y = var_664_cast_fp16)[name = tensor("x_69_cast_fp16")]; + tensor var_666_cast_fp16 = mul(x = x_69_cast_fp16, y = flow_net_res_blocks_3_in_ln_weight_to_fp16)[name = tensor("op_666_cast_fp16")]; + tensor x_71_cast_fp16 = add(x = var_666_cast_fp16, y = flow_net_res_blocks_3_in_ln_bias_to_fp16)[name = tensor("x_71_cast_fp16")]; + tensor var_668_promoted_to_fp16 = const()[name = tensor("op_668_promoted_to_fp16"), val = tensor(0x1p+0)]; + tensor var_669_cast_fp16 = add(x = var_652_cast_fp16_1, y = var_668_promoted_to_fp16)[name = tensor("op_669_cast_fp16")]; + tensor var_670_cast_fp16 = mul(x = x_71_cast_fp16, y = var_669_cast_fp16)[name = tensor("op_670_cast_fp16")]; + tensor input_109_cast_fp16 = add(x = var_670_cast_fp16, y = var_652_cast_fp16_0)[name = tensor("input_109_cast_fp16")]; + tensor linear_42_cast_fp16 = linear(bias = flow_net_res_blocks_3_mlp_0_bias_to_fp16, weight = flow_net_res_blocks_3_mlp_0_weight_to_fp16, x = input_109_cast_fp16)[name = tensor("linear_42_cast_fp16")]; + tensor input_113_cast_fp16 = silu(x = linear_42_cast_fp16)[name = tensor("input_113_cast_fp16")]; + tensor linear_43_cast_fp16 = linear(bias = flow_net_res_blocks_3_mlp_2_bias_to_fp16, weight = flow_net_res_blocks_3_mlp_2_weight_to_fp16, x = input_113_cast_fp16)[name = tensor("linear_43_cast_fp16")]; + tensor var_681_cast_fp16 = mul(x = var_652_cast_fp16_2, y = linear_43_cast_fp16)[name = tensor("op_681_cast_fp16")]; + tensor x_73_cast_fp16 = add(x = x_67_cast_fp16, y = var_681_cast_fp16)[name = tensor("x_73_cast_fp16")]; + tensor linear_44_cast_fp16 = linear(bias = flow_net_res_blocks_4_adaLN_modulation_1_bias_to_fp16, weight = flow_net_res_blocks_4_adaLN_modulation_1_weight_to_fp16, x = input_83_cast_fp16)[name = tensor("linear_44_cast_fp16")]; + tensor var_691_split_sizes_0 = const()[name = tensor("op_691_split_sizes_0"), val = tensor([512, 512, 512])]; + tensor var_691_axis_0 = const()[name = tensor("op_691_axis_0"), val = tensor(-1)]; + tensor var_691_cast_fp16_0, tensor var_691_cast_fp16_1, tensor var_691_cast_fp16_2 = split(axis = var_691_axis_0, split_sizes = var_691_split_sizes_0, x = linear_44_cast_fp16)[name = tensor("op_691_cast_fp16")]; + tensor mean_23_axes_0 = const()[name = tensor("mean_23_axes_0"), val = tensor([-1])]; + tensor mean_23_keep_dims_0 = const()[name = tensor("mean_23_keep_dims_0"), val = tensor(true)]; + tensor mean_23_cast_fp16 = reduce_mean(axes = mean_23_axes_0, keep_dims = mean_23_keep_dims_0, x = x_73_cast_fp16)[name = tensor("mean_23_cast_fp16")]; + tensor sub_19_cast_fp16 = sub(x = x_73_cast_fp16, y = mean_23_cast_fp16)[name = tensor("sub_19_cast_fp16")]; + tensor square_15_cast_fp16 = square(x = sub_19_cast_fp16)[name = tensor("square_15_cast_fp16")]; + tensor reduce_mean_31_axes_0 = const()[name = tensor("reduce_mean_31_axes_0"), val = tensor([-1])]; + tensor reduce_mean_31_keep_dims_0 = const()[name = tensor("reduce_mean_31_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_31_cast_fp16 = reduce_mean(axes = reduce_mean_31_axes_0, keep_dims = reduce_mean_31_keep_dims_0, x = square_15_cast_fp16)[name = tensor("reduce_mean_31_cast_fp16")]; + tensor var_701_to_fp16 = const()[name = tensor("op_701_to_fp16"), val = tensor(0x1.1p-20)]; + tensor var_702_cast_fp16 = add(x = reduce_mean_31_cast_fp16, y = var_701_to_fp16)[name = tensor("op_702_cast_fp16")]; + tensor var_703_cast_fp16 = sqrt(x = var_702_cast_fp16)[name = tensor("op_703_cast_fp16")]; + tensor x_75_cast_fp16 = real_div(x = sub_19_cast_fp16, y = var_703_cast_fp16)[name = tensor("x_75_cast_fp16")]; + tensor var_705_cast_fp16 = mul(x = x_75_cast_fp16, y = flow_net_res_blocks_4_in_ln_weight_to_fp16)[name = tensor("op_705_cast_fp16")]; + tensor x_77_cast_fp16 = add(x = var_705_cast_fp16, y = flow_net_res_blocks_4_in_ln_bias_to_fp16)[name = tensor("x_77_cast_fp16")]; + tensor var_707_promoted_to_fp16 = const()[name = tensor("op_707_promoted_to_fp16"), val = tensor(0x1p+0)]; + tensor var_708_cast_fp16 = add(x = var_691_cast_fp16_1, y = var_707_promoted_to_fp16)[name = tensor("op_708_cast_fp16")]; + tensor var_709_cast_fp16 = mul(x = x_77_cast_fp16, y = var_708_cast_fp16)[name = tensor("op_709_cast_fp16")]; + tensor input_117_cast_fp16 = add(x = var_709_cast_fp16, y = var_691_cast_fp16_0)[name = tensor("input_117_cast_fp16")]; + tensor linear_45_cast_fp16 = linear(bias = flow_net_res_blocks_4_mlp_0_bias_to_fp16, weight = flow_net_res_blocks_4_mlp_0_weight_to_fp16, x = input_117_cast_fp16)[name = tensor("linear_45_cast_fp16")]; + tensor input_121_cast_fp16 = silu(x = linear_45_cast_fp16)[name = tensor("input_121_cast_fp16")]; + tensor linear_46_cast_fp16 = linear(bias = flow_net_res_blocks_4_mlp_2_bias_to_fp16, weight = flow_net_res_blocks_4_mlp_2_weight_to_fp16, x = input_121_cast_fp16)[name = tensor("linear_46_cast_fp16")]; + tensor var_720_cast_fp16 = mul(x = var_691_cast_fp16_2, y = linear_46_cast_fp16)[name = tensor("op_720_cast_fp16")]; + tensor x_79_cast_fp16 = add(x = x_73_cast_fp16, y = var_720_cast_fp16)[name = tensor("x_79_cast_fp16")]; + tensor linear_47_cast_fp16 = linear(bias = flow_net_res_blocks_5_adaLN_modulation_1_bias_to_fp16, weight = flow_net_res_blocks_5_adaLN_modulation_1_weight_to_fp16, x = input_83_cast_fp16)[name = tensor("linear_47_cast_fp16")]; + tensor var_730_split_sizes_0 = const()[name = tensor("op_730_split_sizes_0"), val = tensor([512, 512, 512])]; + tensor var_730_axis_0 = const()[name = tensor("op_730_axis_0"), val = tensor(-1)]; + tensor var_730_cast_fp16_0, tensor var_730_cast_fp16_1, tensor var_730_cast_fp16_2 = split(axis = var_730_axis_0, split_sizes = var_730_split_sizes_0, x = linear_47_cast_fp16)[name = tensor("op_730_cast_fp16")]; + tensor mean_25_axes_0 = const()[name = tensor("mean_25_axes_0"), val = tensor([-1])]; + tensor mean_25_keep_dims_0 = const()[name = tensor("mean_25_keep_dims_0"), val = tensor(true)]; + tensor mean_25_cast_fp16 = reduce_mean(axes = mean_25_axes_0, keep_dims = mean_25_keep_dims_0, x = x_79_cast_fp16)[name = tensor("mean_25_cast_fp16")]; + tensor sub_20_cast_fp16 = sub(x = x_79_cast_fp16, y = mean_25_cast_fp16)[name = tensor("sub_20_cast_fp16")]; + tensor square_16_cast_fp16 = square(x = sub_20_cast_fp16)[name = tensor("square_16_cast_fp16")]; + tensor reduce_mean_33_axes_0 = const()[name = tensor("reduce_mean_33_axes_0"), val = tensor([-1])]; + tensor reduce_mean_33_keep_dims_0 = const()[name = tensor("reduce_mean_33_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_33_cast_fp16 = reduce_mean(axes = reduce_mean_33_axes_0, keep_dims = reduce_mean_33_keep_dims_0, x = square_16_cast_fp16)[name = tensor("reduce_mean_33_cast_fp16")]; + tensor var_740_to_fp16 = const()[name = tensor("op_740_to_fp16"), val = tensor(0x1.1p-20)]; + tensor var_741_cast_fp16 = add(x = reduce_mean_33_cast_fp16, y = var_740_to_fp16)[name = tensor("op_741_cast_fp16")]; + tensor var_742_cast_fp16 = sqrt(x = var_741_cast_fp16)[name = tensor("op_742_cast_fp16")]; + tensor x_81_cast_fp16 = real_div(x = sub_20_cast_fp16, y = var_742_cast_fp16)[name = tensor("x_81_cast_fp16")]; + tensor var_744_cast_fp16 = mul(x = x_81_cast_fp16, y = flow_net_res_blocks_5_in_ln_weight_to_fp16)[name = tensor("op_744_cast_fp16")]; + tensor x_83_cast_fp16 = add(x = var_744_cast_fp16, y = flow_net_res_blocks_5_in_ln_bias_to_fp16)[name = tensor("x_83_cast_fp16")]; + tensor var_746_promoted_to_fp16 = const()[name = tensor("op_746_promoted_to_fp16"), val = tensor(0x1p+0)]; + tensor var_747_cast_fp16 = add(x = var_730_cast_fp16_1, y = var_746_promoted_to_fp16)[name = tensor("op_747_cast_fp16")]; + tensor var_748_cast_fp16 = mul(x = x_83_cast_fp16, y = var_747_cast_fp16)[name = tensor("op_748_cast_fp16")]; + tensor input_125_cast_fp16 = add(x = var_748_cast_fp16, y = var_730_cast_fp16_0)[name = tensor("input_125_cast_fp16")]; + tensor linear_48_cast_fp16 = linear(bias = flow_net_res_blocks_5_mlp_0_bias_to_fp16, weight = flow_net_res_blocks_5_mlp_0_weight_to_fp16, x = input_125_cast_fp16)[name = tensor("linear_48_cast_fp16")]; + tensor input_129_cast_fp16 = silu(x = linear_48_cast_fp16)[name = tensor("input_129_cast_fp16")]; + tensor linear_49_cast_fp16 = linear(bias = flow_net_res_blocks_5_mlp_2_bias_to_fp16, weight = flow_net_res_blocks_5_mlp_2_weight_to_fp16, x = input_129_cast_fp16)[name = tensor("linear_49_cast_fp16")]; + tensor var_759_cast_fp16 = mul(x = var_730_cast_fp16_2, y = linear_49_cast_fp16)[name = tensor("op_759_cast_fp16")]; + tensor x_85_cast_fp16 = add(x = x_79_cast_fp16, y = var_759_cast_fp16)[name = tensor("x_85_cast_fp16")]; + tensor linear_50_cast_fp16 = linear(bias = flow_net_final_layer_adaLN_modulation_1_bias_to_fp16, weight = flow_net_final_layer_adaLN_modulation_1_weight_to_fp16, x = input_83_cast_fp16)[name = tensor("linear_50_cast_fp16")]; + tensor var_768_split_sizes_0 = const()[name = tensor("op_768_split_sizes_0"), val = tensor([512, 512])]; + tensor var_768_axis_0 = const()[name = tensor("op_768_axis_0"), val = tensor(-1)]; + tensor var_768_cast_fp16_0, tensor var_768_cast_fp16_1 = split(axis = var_768_axis_0, split_sizes = var_768_split_sizes_0, x = linear_50_cast_fp16)[name = tensor("op_768_cast_fp16")]; + tensor mean_27_axes_0 = const()[name = tensor("mean_27_axes_0"), val = tensor([-1])]; + tensor mean_27_keep_dims_0 = const()[name = tensor("mean_27_keep_dims_0"), val = tensor(true)]; + tensor mean_27_cast_fp16 = reduce_mean(axes = mean_27_axes_0, keep_dims = mean_27_keep_dims_0, x = x_85_cast_fp16)[name = tensor("mean_27_cast_fp16")]; + tensor sub_21_cast_fp16 = sub(x = x_85_cast_fp16, y = mean_27_cast_fp16)[name = tensor("sub_21_cast_fp16")]; + tensor square_17_cast_fp16 = square(x = sub_21_cast_fp16)[name = tensor("square_17_cast_fp16")]; + tensor reduce_mean_35_axes_0 = const()[name = tensor("reduce_mean_35_axes_0"), val = tensor([-1])]; + tensor reduce_mean_35_keep_dims_0 = const()[name = tensor("reduce_mean_35_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_35_cast_fp16 = reduce_mean(axes = reduce_mean_35_axes_0, keep_dims = reduce_mean_35_keep_dims_0, x = square_17_cast_fp16)[name = tensor("reduce_mean_35_cast_fp16")]; + tensor var_775_to_fp16 = const()[name = tensor("op_775_to_fp16"), val = tensor(0x1.1p-20)]; + tensor var_776_cast_fp16 = add(x = reduce_mean_35_cast_fp16, y = var_775_to_fp16)[name = tensor("op_776_cast_fp16")]; + tensor var_777_cast_fp16 = sqrt(x = var_776_cast_fp16)[name = tensor("op_777_cast_fp16")]; + tensor x_87_cast_fp16 = real_div(x = sub_21_cast_fp16, y = var_777_cast_fp16)[name = tensor("x_87_cast_fp16")]; + tensor var_779_promoted_to_fp16 = const()[name = tensor("op_779_promoted_to_fp16"), val = tensor(0x1p+0)]; + tensor var_780_cast_fp16 = add(x = var_768_cast_fp16_1, y = var_779_promoted_to_fp16)[name = tensor("op_780_cast_fp16")]; + tensor var_781_cast_fp16 = mul(x = x_87_cast_fp16, y = var_780_cast_fp16)[name = tensor("op_781_cast_fp16")]; + tensor input_133_cast_fp16 = add(x = var_781_cast_fp16, y = var_768_cast_fp16_0)[name = tensor("input_133_cast_fp16")]; + tensor linear_51_cast_fp16 = linear(bias = flow_net_final_layer_linear_bias_to_fp16, weight = flow_net_final_layer_linear_weight_to_fp16, x = input_133_cast_fp16)[name = tensor("linear_51_cast_fp16")]; + tensor var_792_to_fp16 = const()[name = tensor("op_792_to_fp16"), val = tensor(0x1p-3)]; + tensor var_793_cast_fp16 = mul(x = linear_51_cast_fp16, y = var_792_to_fp16)[name = tensor("op_793_cast_fp16")]; + tensor input_135_cast_fp16 = add(x = input_67_cast_fp16, y = var_793_cast_fp16)[name = tensor("input_135_cast_fp16")]; + tensor linear_52_cast_fp16 = linear(bias = flow_net_input_proj_bias_to_fp16, weight = flow_net_input_proj_weight_to_fp16, x = input_135_cast_fp16)[name = tensor("linear_52_cast_fp16")]; + tensor input_139_to_fp16 = const()[name = tensor("input_139_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(18999552)))]; + tensor input_141_cast_fp16 = silu(x = input_139_to_fp16)[name = tensor("input_141_cast_fp16")]; + tensor linear_54_cast_fp16 = linear(bias = flow_net_time_embed_0_mlp_2_bias_to_fp16, weight = flow_net_time_embed_0_mlp_2_weight_to_fp16, x = input_141_cast_fp16)[name = tensor("linear_54_cast_fp16")]; + tensor reduce_mean_36_axes_0 = const()[name = tensor("reduce_mean_36_axes_0"), val = tensor([-1])]; + tensor reduce_mean_36_keep_dims_0 = const()[name = tensor("reduce_mean_36_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_36_cast_fp16 = reduce_mean(axes = reduce_mean_36_axes_0, keep_dims = reduce_mean_36_keep_dims_0, x = linear_54_cast_fp16)[name = tensor("reduce_mean_36_cast_fp16")]; + tensor sub_22_cast_fp16 = sub(x = linear_54_cast_fp16, y = reduce_mean_36_cast_fp16)[name = tensor("sub_22_cast_fp16")]; + tensor square_18_cast_fp16 = square(x = sub_22_cast_fp16)[name = tensor("square_18_cast_fp16")]; + tensor reduce_mean_37_axes_0 = const()[name = tensor("reduce_mean_37_axes_0"), val = tensor([-1])]; + tensor reduce_mean_37_keep_dims_0 = const()[name = tensor("reduce_mean_37_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_37_cast_fp16 = reduce_mean(axes = reduce_mean_37_axes_0, keep_dims = reduce_mean_37_keep_dims_0, x = square_18_cast_fp16)[name = tensor("reduce_mean_37_cast_fp16")]; + tensor real_div_4_to_fp16 = const()[name = tensor("real_div_4_to_fp16"), val = tensor(0x1.008p+0)]; + tensor mul_4_cast_fp16 = mul(x = reduce_mean_37_cast_fp16, y = real_div_4_to_fp16)[name = tensor("mul_4_cast_fp16")]; + tensor var_859_to_fp16 = const()[name = tensor("op_859_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_37_cast_fp16 = add(x = mul_4_cast_fp16, y = var_859_to_fp16)[name = tensor("var_37_cast_fp16")]; + tensor var_862_epsilon_0 = const()[name = tensor("op_862_epsilon_0"), val = tensor(0x1.197998p-40)]; + tensor var_862_cast_fp16 = rsqrt(epsilon = var_862_epsilon_0, x = var_37_cast_fp16)[name = tensor("op_862_cast_fp16")]; + tensor var_863_cast_fp16 = mul(x = const_3_to_fp16, y = var_862_cast_fp16)[name = tensor("op_863_cast_fp16")]; + tensor var_864_cast_fp16 = mul(x = linear_54_cast_fp16, y = var_863_cast_fp16)[name = tensor("op_864_cast_fp16")]; + tensor input_145_to_fp16 = const()[name = tensor("input_145_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(19000640)))]; + tensor input_147_cast_fp16 = silu(x = input_145_to_fp16)[name = tensor("input_147_cast_fp16")]; + tensor linear_56_cast_fp16 = linear(bias = flow_net_time_embed_1_mlp_2_bias_to_fp16, weight = flow_net_time_embed_1_mlp_2_weight_to_fp16, x = input_147_cast_fp16)[name = tensor("linear_56_cast_fp16")]; + tensor reduce_mean_38_axes_0 = const()[name = tensor("reduce_mean_38_axes_0"), val = tensor([-1])]; + tensor reduce_mean_38_keep_dims_0 = const()[name = tensor("reduce_mean_38_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_38_cast_fp16 = reduce_mean(axes = reduce_mean_38_axes_0, keep_dims = reduce_mean_38_keep_dims_0, x = linear_56_cast_fp16)[name = tensor("reduce_mean_38_cast_fp16")]; + tensor sub_24_cast_fp16 = sub(x = linear_56_cast_fp16, y = reduce_mean_38_cast_fp16)[name = tensor("sub_24_cast_fp16")]; + tensor square_19_cast_fp16 = square(x = sub_24_cast_fp16)[name = tensor("square_19_cast_fp16")]; + tensor reduce_mean_39_axes_0 = const()[name = tensor("reduce_mean_39_axes_0"), val = tensor([-1])]; + tensor reduce_mean_39_keep_dims_0 = const()[name = tensor("reduce_mean_39_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_39_cast_fp16 = reduce_mean(axes = reduce_mean_39_axes_0, keep_dims = reduce_mean_39_keep_dims_0, x = square_19_cast_fp16)[name = tensor("reduce_mean_39_cast_fp16")]; + tensor real_div_5_to_fp16 = const()[name = tensor("real_div_5_to_fp16"), val = tensor(0x1.008p+0)]; + tensor mul_5_cast_fp16 = mul(x = reduce_mean_39_cast_fp16, y = real_div_5_to_fp16)[name = tensor("mul_5_cast_fp16")]; + tensor var_896_to_fp16 = const()[name = tensor("op_896_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_39_cast_fp16 = add(x = mul_5_cast_fp16, y = var_896_to_fp16)[name = tensor("var_39_cast_fp16")]; + tensor var_899_epsilon_0 = const()[name = tensor("op_899_epsilon_0"), val = tensor(0x1.197998p-40)]; + tensor var_899_cast_fp16 = rsqrt(epsilon = var_899_epsilon_0, x = var_39_cast_fp16)[name = tensor("op_899_cast_fp16")]; + tensor var_900_cast_fp16 = mul(x = const_5_to_fp16, y = var_899_cast_fp16)[name = tensor("op_900_cast_fp16")]; + tensor var_901_cast_fp16 = mul(x = linear_56_cast_fp16, y = var_900_cast_fp16)[name = tensor("op_901_cast_fp16")]; + tensor var_913_cast_fp16 = add(x = var_864_cast_fp16, y = var_901_cast_fp16)[name = tensor("op_913_cast_fp16")]; + tensor _inversed_t_combined_5_y_0_to_fp16 = const()[name = tensor("_inversed_t_combined_5_y_0_to_fp16"), val = tensor(0x1p-1)]; + tensor _inversed_t_combined_5_cast_fp16 = mul(x = var_913_cast_fp16, y = _inversed_t_combined_5_y_0_to_fp16)[name = tensor("_inversed_t_combined_5_cast_fp16")]; + tensor input_149_cast_fp16 = add(x = _inversed_t_combined_5_cast_fp16, y = linear_5_cast_fp16)[name = tensor("input_149_cast_fp16")]; + tensor input_151_cast_fp16 = silu(x = input_149_cast_fp16)[name = tensor("input_151_cast_fp16")]; + tensor linear_58_cast_fp16 = linear(bias = flow_net_res_blocks_0_adaLN_modulation_1_bias_to_fp16, weight = flow_net_res_blocks_0_adaLN_modulation_1_weight_to_fp16, x = input_151_cast_fp16)[name = tensor("linear_58_cast_fp16")]; + tensor var_928_split_sizes_0 = const()[name = tensor("op_928_split_sizes_0"), val = tensor([512, 512, 512])]; + tensor var_928_axis_0 = const()[name = tensor("op_928_axis_0"), val = tensor(-1)]; + tensor var_928_cast_fp16_0, tensor var_928_cast_fp16_1, tensor var_928_cast_fp16_2 = split(axis = var_928_axis_0, split_sizes = var_928_split_sizes_0, x = linear_58_cast_fp16)[name = tensor("op_928_cast_fp16")]; + tensor mean_29_axes_0 = const()[name = tensor("mean_29_axes_0"), val = tensor([-1])]; + tensor mean_29_keep_dims_0 = const()[name = tensor("mean_29_keep_dims_0"), val = tensor(true)]; + tensor mean_29_cast_fp16 = reduce_mean(axes = mean_29_axes_0, keep_dims = mean_29_keep_dims_0, x = linear_52_cast_fp16)[name = tensor("mean_29_cast_fp16")]; + tensor sub_26_cast_fp16 = sub(x = linear_52_cast_fp16, y = mean_29_cast_fp16)[name = tensor("sub_26_cast_fp16")]; + tensor square_20_cast_fp16 = square(x = sub_26_cast_fp16)[name = tensor("square_20_cast_fp16")]; + tensor reduce_mean_41_axes_0 = const()[name = tensor("reduce_mean_41_axes_0"), val = tensor([-1])]; + tensor reduce_mean_41_keep_dims_0 = const()[name = tensor("reduce_mean_41_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_41_cast_fp16 = reduce_mean(axes = reduce_mean_41_axes_0, keep_dims = reduce_mean_41_keep_dims_0, x = square_20_cast_fp16)[name = tensor("reduce_mean_41_cast_fp16")]; + tensor var_938_to_fp16 = const()[name = tensor("op_938_to_fp16"), val = tensor(0x1.1p-20)]; + tensor var_939_cast_fp16 = add(x = reduce_mean_41_cast_fp16, y = var_938_to_fp16)[name = tensor("op_939_cast_fp16")]; + tensor var_940_cast_fp16 = sqrt(x = var_939_cast_fp16)[name = tensor("op_940_cast_fp16")]; + tensor x_95_cast_fp16 = real_div(x = sub_26_cast_fp16, y = var_940_cast_fp16)[name = tensor("x_95_cast_fp16")]; + tensor var_942_cast_fp16 = mul(x = x_95_cast_fp16, y = flow_net_res_blocks_0_in_ln_weight_to_fp16)[name = tensor("op_942_cast_fp16")]; + tensor x_97_cast_fp16 = add(x = var_942_cast_fp16, y = flow_net_res_blocks_0_in_ln_bias_to_fp16)[name = tensor("x_97_cast_fp16")]; + tensor var_944_promoted_to_fp16 = const()[name = tensor("op_944_promoted_to_fp16"), val = tensor(0x1p+0)]; + tensor var_945_cast_fp16 = add(x = var_928_cast_fp16_1, y = var_944_promoted_to_fp16)[name = tensor("op_945_cast_fp16")]; + tensor var_946_cast_fp16 = mul(x = x_97_cast_fp16, y = var_945_cast_fp16)[name = tensor("op_946_cast_fp16")]; + tensor input_153_cast_fp16 = add(x = var_946_cast_fp16, y = var_928_cast_fp16_0)[name = tensor("input_153_cast_fp16")]; + tensor linear_59_cast_fp16 = linear(bias = flow_net_res_blocks_0_mlp_0_bias_to_fp16, weight = flow_net_res_blocks_0_mlp_0_weight_to_fp16, x = input_153_cast_fp16)[name = tensor("linear_59_cast_fp16")]; + tensor input_157_cast_fp16 = silu(x = linear_59_cast_fp16)[name = tensor("input_157_cast_fp16")]; + tensor linear_60_cast_fp16 = linear(bias = flow_net_res_blocks_0_mlp_2_bias_to_fp16, weight = flow_net_res_blocks_0_mlp_2_weight_to_fp16, x = input_157_cast_fp16)[name = tensor("linear_60_cast_fp16")]; + tensor var_957_cast_fp16 = mul(x = var_928_cast_fp16_2, y = linear_60_cast_fp16)[name = tensor("op_957_cast_fp16")]; + tensor x_99_cast_fp16 = add(x = linear_52_cast_fp16, y = var_957_cast_fp16)[name = tensor("x_99_cast_fp16")]; + tensor linear_61_cast_fp16 = linear(bias = flow_net_res_blocks_1_adaLN_modulation_1_bias_to_fp16, weight = flow_net_res_blocks_1_adaLN_modulation_1_weight_to_fp16, x = input_151_cast_fp16)[name = tensor("linear_61_cast_fp16")]; + tensor var_967_split_sizes_0 = const()[name = tensor("op_967_split_sizes_0"), val = tensor([512, 512, 512])]; + tensor var_967_axis_0 = const()[name = tensor("op_967_axis_0"), val = tensor(-1)]; + tensor var_967_cast_fp16_0, tensor var_967_cast_fp16_1, tensor var_967_cast_fp16_2 = split(axis = var_967_axis_0, split_sizes = var_967_split_sizes_0, x = linear_61_cast_fp16)[name = tensor("op_967_cast_fp16")]; + tensor mean_31_axes_0 = const()[name = tensor("mean_31_axes_0"), val = tensor([-1])]; + tensor mean_31_keep_dims_0 = const()[name = tensor("mean_31_keep_dims_0"), val = tensor(true)]; + tensor mean_31_cast_fp16 = reduce_mean(axes = mean_31_axes_0, keep_dims = mean_31_keep_dims_0, x = x_99_cast_fp16)[name = tensor("mean_31_cast_fp16")]; + tensor sub_27_cast_fp16 = sub(x = x_99_cast_fp16, y = mean_31_cast_fp16)[name = tensor("sub_27_cast_fp16")]; + tensor square_21_cast_fp16 = square(x = sub_27_cast_fp16)[name = tensor("square_21_cast_fp16")]; + tensor reduce_mean_43_axes_0 = const()[name = tensor("reduce_mean_43_axes_0"), val = tensor([-1])]; + tensor reduce_mean_43_keep_dims_0 = const()[name = tensor("reduce_mean_43_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_43_cast_fp16 = reduce_mean(axes = reduce_mean_43_axes_0, keep_dims = reduce_mean_43_keep_dims_0, x = square_21_cast_fp16)[name = tensor("reduce_mean_43_cast_fp16")]; + tensor var_977_to_fp16 = const()[name = tensor("op_977_to_fp16"), val = tensor(0x1.1p-20)]; + tensor var_978_cast_fp16 = add(x = reduce_mean_43_cast_fp16, y = var_977_to_fp16)[name = tensor("op_978_cast_fp16")]; + tensor var_979_cast_fp16 = sqrt(x = var_978_cast_fp16)[name = tensor("op_979_cast_fp16")]; + tensor x_101_cast_fp16 = real_div(x = sub_27_cast_fp16, y = var_979_cast_fp16)[name = tensor("x_101_cast_fp16")]; + tensor var_981_cast_fp16 = mul(x = x_101_cast_fp16, y = flow_net_res_blocks_1_in_ln_weight_to_fp16)[name = tensor("op_981_cast_fp16")]; + tensor x_103_cast_fp16 = add(x = var_981_cast_fp16, y = flow_net_res_blocks_1_in_ln_bias_to_fp16)[name = tensor("x_103_cast_fp16")]; + tensor var_983_promoted_to_fp16 = const()[name = tensor("op_983_promoted_to_fp16"), val = tensor(0x1p+0)]; + tensor var_984_cast_fp16 = add(x = var_967_cast_fp16_1, y = var_983_promoted_to_fp16)[name = tensor("op_984_cast_fp16")]; + tensor var_985_cast_fp16 = mul(x = x_103_cast_fp16, y = var_984_cast_fp16)[name = tensor("op_985_cast_fp16")]; + tensor input_161_cast_fp16 = add(x = var_985_cast_fp16, y = var_967_cast_fp16_0)[name = tensor("input_161_cast_fp16")]; + tensor linear_62_cast_fp16 = linear(bias = flow_net_res_blocks_1_mlp_0_bias_to_fp16, weight = flow_net_res_blocks_1_mlp_0_weight_to_fp16, x = input_161_cast_fp16)[name = tensor("linear_62_cast_fp16")]; + tensor input_165_cast_fp16 = silu(x = linear_62_cast_fp16)[name = tensor("input_165_cast_fp16")]; + tensor linear_63_cast_fp16 = linear(bias = flow_net_res_blocks_1_mlp_2_bias_to_fp16, weight = flow_net_res_blocks_1_mlp_2_weight_to_fp16, x = input_165_cast_fp16)[name = tensor("linear_63_cast_fp16")]; + tensor var_996_cast_fp16 = mul(x = var_967_cast_fp16_2, y = linear_63_cast_fp16)[name = tensor("op_996_cast_fp16")]; + tensor x_105_cast_fp16 = add(x = x_99_cast_fp16, y = var_996_cast_fp16)[name = tensor("x_105_cast_fp16")]; + tensor linear_64_cast_fp16 = linear(bias = flow_net_res_blocks_2_adaLN_modulation_1_bias_to_fp16, weight = flow_net_res_blocks_2_adaLN_modulation_1_weight_to_fp16, x = input_151_cast_fp16)[name = tensor("linear_64_cast_fp16")]; + tensor var_1006_split_sizes_0 = const()[name = tensor("op_1006_split_sizes_0"), val = tensor([512, 512, 512])]; + tensor var_1006_axis_0 = const()[name = tensor("op_1006_axis_0"), val = tensor(-1)]; + tensor var_1006_cast_fp16_0, tensor var_1006_cast_fp16_1, tensor var_1006_cast_fp16_2 = split(axis = var_1006_axis_0, split_sizes = var_1006_split_sizes_0, x = linear_64_cast_fp16)[name = tensor("op_1006_cast_fp16")]; + tensor mean_33_axes_0 = const()[name = tensor("mean_33_axes_0"), val = tensor([-1])]; + tensor mean_33_keep_dims_0 = const()[name = tensor("mean_33_keep_dims_0"), val = tensor(true)]; + tensor mean_33_cast_fp16 = reduce_mean(axes = mean_33_axes_0, keep_dims = mean_33_keep_dims_0, x = x_105_cast_fp16)[name = tensor("mean_33_cast_fp16")]; + tensor sub_28_cast_fp16 = sub(x = x_105_cast_fp16, y = mean_33_cast_fp16)[name = tensor("sub_28_cast_fp16")]; + tensor square_22_cast_fp16 = square(x = sub_28_cast_fp16)[name = tensor("square_22_cast_fp16")]; + tensor reduce_mean_45_axes_0 = const()[name = tensor("reduce_mean_45_axes_0"), val = tensor([-1])]; + tensor reduce_mean_45_keep_dims_0 = const()[name = tensor("reduce_mean_45_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_45_cast_fp16 = reduce_mean(axes = reduce_mean_45_axes_0, keep_dims = reduce_mean_45_keep_dims_0, x = square_22_cast_fp16)[name = tensor("reduce_mean_45_cast_fp16")]; + tensor var_1016_to_fp16 = const()[name = tensor("op_1016_to_fp16"), val = tensor(0x1.1p-20)]; + tensor var_1017_cast_fp16 = add(x = reduce_mean_45_cast_fp16, y = var_1016_to_fp16)[name = tensor("op_1017_cast_fp16")]; + tensor var_1018_cast_fp16 = sqrt(x = var_1017_cast_fp16)[name = tensor("op_1018_cast_fp16")]; + tensor x_107_cast_fp16 = real_div(x = sub_28_cast_fp16, y = var_1018_cast_fp16)[name = tensor("x_107_cast_fp16")]; + tensor var_1020_cast_fp16 = mul(x = x_107_cast_fp16, y = flow_net_res_blocks_2_in_ln_weight_to_fp16)[name = tensor("op_1020_cast_fp16")]; + tensor x_109_cast_fp16 = add(x = var_1020_cast_fp16, y = flow_net_res_blocks_2_in_ln_bias_to_fp16)[name = tensor("x_109_cast_fp16")]; + tensor var_1022_promoted_to_fp16 = const()[name = tensor("op_1022_promoted_to_fp16"), val = tensor(0x1p+0)]; + tensor var_1023_cast_fp16 = add(x = var_1006_cast_fp16_1, y = var_1022_promoted_to_fp16)[name = tensor("op_1023_cast_fp16")]; + tensor var_1024_cast_fp16 = mul(x = x_109_cast_fp16, y = var_1023_cast_fp16)[name = tensor("op_1024_cast_fp16")]; + tensor input_169_cast_fp16 = add(x = var_1024_cast_fp16, y = var_1006_cast_fp16_0)[name = tensor("input_169_cast_fp16")]; + tensor linear_65_cast_fp16 = linear(bias = flow_net_res_blocks_2_mlp_0_bias_to_fp16, weight = flow_net_res_blocks_2_mlp_0_weight_to_fp16, x = input_169_cast_fp16)[name = tensor("linear_65_cast_fp16")]; + tensor input_173_cast_fp16 = silu(x = linear_65_cast_fp16)[name = tensor("input_173_cast_fp16")]; + tensor linear_66_cast_fp16 = linear(bias = flow_net_res_blocks_2_mlp_2_bias_to_fp16, weight = flow_net_res_blocks_2_mlp_2_weight_to_fp16, x = input_173_cast_fp16)[name = tensor("linear_66_cast_fp16")]; + tensor var_1035_cast_fp16 = mul(x = var_1006_cast_fp16_2, y = linear_66_cast_fp16)[name = tensor("op_1035_cast_fp16")]; + tensor x_111_cast_fp16 = add(x = x_105_cast_fp16, y = var_1035_cast_fp16)[name = tensor("x_111_cast_fp16")]; + tensor linear_67_cast_fp16 = linear(bias = flow_net_res_blocks_3_adaLN_modulation_1_bias_to_fp16, weight = flow_net_res_blocks_3_adaLN_modulation_1_weight_to_fp16, x = input_151_cast_fp16)[name = tensor("linear_67_cast_fp16")]; + tensor var_1045_split_sizes_0 = const()[name = tensor("op_1045_split_sizes_0"), val = tensor([512, 512, 512])]; + tensor var_1045_axis_0 = const()[name = tensor("op_1045_axis_0"), val = tensor(-1)]; + tensor var_1045_cast_fp16_0, tensor var_1045_cast_fp16_1, tensor var_1045_cast_fp16_2 = split(axis = var_1045_axis_0, split_sizes = var_1045_split_sizes_0, x = linear_67_cast_fp16)[name = tensor("op_1045_cast_fp16")]; + tensor mean_35_axes_0 = const()[name = tensor("mean_35_axes_0"), val = tensor([-1])]; + tensor mean_35_keep_dims_0 = const()[name = tensor("mean_35_keep_dims_0"), val = tensor(true)]; + tensor mean_35_cast_fp16 = reduce_mean(axes = mean_35_axes_0, keep_dims = mean_35_keep_dims_0, x = x_111_cast_fp16)[name = tensor("mean_35_cast_fp16")]; + tensor sub_29_cast_fp16 = sub(x = x_111_cast_fp16, y = mean_35_cast_fp16)[name = tensor("sub_29_cast_fp16")]; + tensor square_23_cast_fp16 = square(x = sub_29_cast_fp16)[name = tensor("square_23_cast_fp16")]; + tensor reduce_mean_47_axes_0 = const()[name = tensor("reduce_mean_47_axes_0"), val = tensor([-1])]; + tensor reduce_mean_47_keep_dims_0 = const()[name = tensor("reduce_mean_47_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_47_cast_fp16 = reduce_mean(axes = reduce_mean_47_axes_0, keep_dims = reduce_mean_47_keep_dims_0, x = square_23_cast_fp16)[name = tensor("reduce_mean_47_cast_fp16")]; + tensor var_1055_to_fp16 = const()[name = tensor("op_1055_to_fp16"), val = tensor(0x1.1p-20)]; + tensor var_1056_cast_fp16 = add(x = reduce_mean_47_cast_fp16, y = var_1055_to_fp16)[name = tensor("op_1056_cast_fp16")]; + tensor var_1057_cast_fp16 = sqrt(x = var_1056_cast_fp16)[name = tensor("op_1057_cast_fp16")]; + tensor x_113_cast_fp16 = real_div(x = sub_29_cast_fp16, y = var_1057_cast_fp16)[name = tensor("x_113_cast_fp16")]; + tensor var_1059_cast_fp16 = mul(x = x_113_cast_fp16, y = flow_net_res_blocks_3_in_ln_weight_to_fp16)[name = tensor("op_1059_cast_fp16")]; + tensor x_115_cast_fp16 = add(x = var_1059_cast_fp16, y = flow_net_res_blocks_3_in_ln_bias_to_fp16)[name = tensor("x_115_cast_fp16")]; + tensor var_1061_promoted_to_fp16 = const()[name = tensor("op_1061_promoted_to_fp16"), val = tensor(0x1p+0)]; + tensor var_1062_cast_fp16 = add(x = var_1045_cast_fp16_1, y = var_1061_promoted_to_fp16)[name = tensor("op_1062_cast_fp16")]; + tensor var_1063_cast_fp16 = mul(x = x_115_cast_fp16, y = var_1062_cast_fp16)[name = tensor("op_1063_cast_fp16")]; + tensor input_177_cast_fp16 = add(x = var_1063_cast_fp16, y = var_1045_cast_fp16_0)[name = tensor("input_177_cast_fp16")]; + tensor linear_68_cast_fp16 = linear(bias = flow_net_res_blocks_3_mlp_0_bias_to_fp16, weight = flow_net_res_blocks_3_mlp_0_weight_to_fp16, x = input_177_cast_fp16)[name = tensor("linear_68_cast_fp16")]; + tensor input_181_cast_fp16 = silu(x = linear_68_cast_fp16)[name = tensor("input_181_cast_fp16")]; + tensor linear_69_cast_fp16 = linear(bias = flow_net_res_blocks_3_mlp_2_bias_to_fp16, weight = flow_net_res_blocks_3_mlp_2_weight_to_fp16, x = input_181_cast_fp16)[name = tensor("linear_69_cast_fp16")]; + tensor var_1074_cast_fp16 = mul(x = var_1045_cast_fp16_2, y = linear_69_cast_fp16)[name = tensor("op_1074_cast_fp16")]; + tensor x_117_cast_fp16 = add(x = x_111_cast_fp16, y = var_1074_cast_fp16)[name = tensor("x_117_cast_fp16")]; + tensor linear_70_cast_fp16 = linear(bias = flow_net_res_blocks_4_adaLN_modulation_1_bias_to_fp16, weight = flow_net_res_blocks_4_adaLN_modulation_1_weight_to_fp16, x = input_151_cast_fp16)[name = tensor("linear_70_cast_fp16")]; + tensor var_1084_split_sizes_0 = const()[name = tensor("op_1084_split_sizes_0"), val = tensor([512, 512, 512])]; + tensor var_1084_axis_0 = const()[name = tensor("op_1084_axis_0"), val = tensor(-1)]; + tensor var_1084_cast_fp16_0, tensor var_1084_cast_fp16_1, tensor var_1084_cast_fp16_2 = split(axis = var_1084_axis_0, split_sizes = var_1084_split_sizes_0, x = linear_70_cast_fp16)[name = tensor("op_1084_cast_fp16")]; + tensor mean_37_axes_0 = const()[name = tensor("mean_37_axes_0"), val = tensor([-1])]; + tensor mean_37_keep_dims_0 = const()[name = tensor("mean_37_keep_dims_0"), val = tensor(true)]; + tensor mean_37_cast_fp16 = reduce_mean(axes = mean_37_axes_0, keep_dims = mean_37_keep_dims_0, x = x_117_cast_fp16)[name = tensor("mean_37_cast_fp16")]; + tensor sub_30_cast_fp16 = sub(x = x_117_cast_fp16, y = mean_37_cast_fp16)[name = tensor("sub_30_cast_fp16")]; + tensor square_24_cast_fp16 = square(x = sub_30_cast_fp16)[name = tensor("square_24_cast_fp16")]; + tensor reduce_mean_49_axes_0 = const()[name = tensor("reduce_mean_49_axes_0"), val = tensor([-1])]; + tensor reduce_mean_49_keep_dims_0 = const()[name = tensor("reduce_mean_49_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_49_cast_fp16 = reduce_mean(axes = reduce_mean_49_axes_0, keep_dims = reduce_mean_49_keep_dims_0, x = square_24_cast_fp16)[name = tensor("reduce_mean_49_cast_fp16")]; + tensor var_1094_to_fp16 = const()[name = tensor("op_1094_to_fp16"), val = tensor(0x1.1p-20)]; + tensor var_1095_cast_fp16 = add(x = reduce_mean_49_cast_fp16, y = var_1094_to_fp16)[name = tensor("op_1095_cast_fp16")]; + tensor var_1096_cast_fp16 = sqrt(x = var_1095_cast_fp16)[name = tensor("op_1096_cast_fp16")]; + tensor x_119_cast_fp16 = real_div(x = sub_30_cast_fp16, y = var_1096_cast_fp16)[name = tensor("x_119_cast_fp16")]; + tensor var_1098_cast_fp16 = mul(x = x_119_cast_fp16, y = flow_net_res_blocks_4_in_ln_weight_to_fp16)[name = tensor("op_1098_cast_fp16")]; + tensor x_121_cast_fp16 = add(x = var_1098_cast_fp16, y = flow_net_res_blocks_4_in_ln_bias_to_fp16)[name = tensor("x_121_cast_fp16")]; + tensor var_1100_promoted_to_fp16 = const()[name = tensor("op_1100_promoted_to_fp16"), val = tensor(0x1p+0)]; + tensor var_1101_cast_fp16 = add(x = var_1084_cast_fp16_1, y = var_1100_promoted_to_fp16)[name = tensor("op_1101_cast_fp16")]; + tensor var_1102_cast_fp16 = mul(x = x_121_cast_fp16, y = var_1101_cast_fp16)[name = tensor("op_1102_cast_fp16")]; + tensor input_185_cast_fp16 = add(x = var_1102_cast_fp16, y = var_1084_cast_fp16_0)[name = tensor("input_185_cast_fp16")]; + tensor linear_71_cast_fp16 = linear(bias = flow_net_res_blocks_4_mlp_0_bias_to_fp16, weight = flow_net_res_blocks_4_mlp_0_weight_to_fp16, x = input_185_cast_fp16)[name = tensor("linear_71_cast_fp16")]; + tensor input_189_cast_fp16 = silu(x = linear_71_cast_fp16)[name = tensor("input_189_cast_fp16")]; + tensor linear_72_cast_fp16 = linear(bias = flow_net_res_blocks_4_mlp_2_bias_to_fp16, weight = flow_net_res_blocks_4_mlp_2_weight_to_fp16, x = input_189_cast_fp16)[name = tensor("linear_72_cast_fp16")]; + tensor var_1113_cast_fp16 = mul(x = var_1084_cast_fp16_2, y = linear_72_cast_fp16)[name = tensor("op_1113_cast_fp16")]; + tensor x_123_cast_fp16 = add(x = x_117_cast_fp16, y = var_1113_cast_fp16)[name = tensor("x_123_cast_fp16")]; + tensor linear_73_cast_fp16 = linear(bias = flow_net_res_blocks_5_adaLN_modulation_1_bias_to_fp16, weight = flow_net_res_blocks_5_adaLN_modulation_1_weight_to_fp16, x = input_151_cast_fp16)[name = tensor("linear_73_cast_fp16")]; + tensor var_1123_split_sizes_0 = const()[name = tensor("op_1123_split_sizes_0"), val = tensor([512, 512, 512])]; + tensor var_1123_axis_0 = const()[name = tensor("op_1123_axis_0"), val = tensor(-1)]; + tensor var_1123_cast_fp16_0, tensor var_1123_cast_fp16_1, tensor var_1123_cast_fp16_2 = split(axis = var_1123_axis_0, split_sizes = var_1123_split_sizes_0, x = linear_73_cast_fp16)[name = tensor("op_1123_cast_fp16")]; + tensor mean_39_axes_0 = const()[name = tensor("mean_39_axes_0"), val = tensor([-1])]; + tensor mean_39_keep_dims_0 = const()[name = tensor("mean_39_keep_dims_0"), val = tensor(true)]; + tensor mean_39_cast_fp16 = reduce_mean(axes = mean_39_axes_0, keep_dims = mean_39_keep_dims_0, x = x_123_cast_fp16)[name = tensor("mean_39_cast_fp16")]; + tensor sub_31_cast_fp16 = sub(x = x_123_cast_fp16, y = mean_39_cast_fp16)[name = tensor("sub_31_cast_fp16")]; + tensor square_25_cast_fp16 = square(x = sub_31_cast_fp16)[name = tensor("square_25_cast_fp16")]; + tensor reduce_mean_51_axes_0 = const()[name = tensor("reduce_mean_51_axes_0"), val = tensor([-1])]; + tensor reduce_mean_51_keep_dims_0 = const()[name = tensor("reduce_mean_51_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_51_cast_fp16 = reduce_mean(axes = reduce_mean_51_axes_0, keep_dims = reduce_mean_51_keep_dims_0, x = square_25_cast_fp16)[name = tensor("reduce_mean_51_cast_fp16")]; + tensor var_1133_to_fp16 = const()[name = tensor("op_1133_to_fp16"), val = tensor(0x1.1p-20)]; + tensor var_1134_cast_fp16 = add(x = reduce_mean_51_cast_fp16, y = var_1133_to_fp16)[name = tensor("op_1134_cast_fp16")]; + tensor var_1135_cast_fp16 = sqrt(x = var_1134_cast_fp16)[name = tensor("op_1135_cast_fp16")]; + tensor x_125_cast_fp16 = real_div(x = sub_31_cast_fp16, y = var_1135_cast_fp16)[name = tensor("x_125_cast_fp16")]; + tensor var_1137_cast_fp16 = mul(x = x_125_cast_fp16, y = flow_net_res_blocks_5_in_ln_weight_to_fp16)[name = tensor("op_1137_cast_fp16")]; + tensor x_127_cast_fp16 = add(x = var_1137_cast_fp16, y = flow_net_res_blocks_5_in_ln_bias_to_fp16)[name = tensor("x_127_cast_fp16")]; + tensor var_1139_promoted_to_fp16 = const()[name = tensor("op_1139_promoted_to_fp16"), val = tensor(0x1p+0)]; + tensor var_1140_cast_fp16 = add(x = var_1123_cast_fp16_1, y = var_1139_promoted_to_fp16)[name = tensor("op_1140_cast_fp16")]; + tensor var_1141_cast_fp16 = mul(x = x_127_cast_fp16, y = var_1140_cast_fp16)[name = tensor("op_1141_cast_fp16")]; + tensor input_193_cast_fp16 = add(x = var_1141_cast_fp16, y = var_1123_cast_fp16_0)[name = tensor("input_193_cast_fp16")]; + tensor linear_74_cast_fp16 = linear(bias = flow_net_res_blocks_5_mlp_0_bias_to_fp16, weight = flow_net_res_blocks_5_mlp_0_weight_to_fp16, x = input_193_cast_fp16)[name = tensor("linear_74_cast_fp16")]; + tensor input_197_cast_fp16 = silu(x = linear_74_cast_fp16)[name = tensor("input_197_cast_fp16")]; + tensor linear_75_cast_fp16 = linear(bias = flow_net_res_blocks_5_mlp_2_bias_to_fp16, weight = flow_net_res_blocks_5_mlp_2_weight_to_fp16, x = input_197_cast_fp16)[name = tensor("linear_75_cast_fp16")]; + tensor var_1152_cast_fp16 = mul(x = var_1123_cast_fp16_2, y = linear_75_cast_fp16)[name = tensor("op_1152_cast_fp16")]; + tensor x_129_cast_fp16 = add(x = x_123_cast_fp16, y = var_1152_cast_fp16)[name = tensor("x_129_cast_fp16")]; + tensor linear_76_cast_fp16 = linear(bias = flow_net_final_layer_adaLN_modulation_1_bias_to_fp16, weight = flow_net_final_layer_adaLN_modulation_1_weight_to_fp16, x = input_151_cast_fp16)[name = tensor("linear_76_cast_fp16")]; + tensor var_1161_split_sizes_0 = const()[name = tensor("op_1161_split_sizes_0"), val = tensor([512, 512])]; + tensor var_1161_axis_0 = const()[name = tensor("op_1161_axis_0"), val = tensor(-1)]; + tensor var_1161_cast_fp16_0, tensor var_1161_cast_fp16_1 = split(axis = var_1161_axis_0, split_sizes = var_1161_split_sizes_0, x = linear_76_cast_fp16)[name = tensor("op_1161_cast_fp16")]; + tensor mean_41_axes_0 = const()[name = tensor("mean_41_axes_0"), val = tensor([-1])]; + tensor mean_41_keep_dims_0 = const()[name = tensor("mean_41_keep_dims_0"), val = tensor(true)]; + tensor mean_41_cast_fp16 = reduce_mean(axes = mean_41_axes_0, keep_dims = mean_41_keep_dims_0, x = x_129_cast_fp16)[name = tensor("mean_41_cast_fp16")]; + tensor sub_32_cast_fp16 = sub(x = x_129_cast_fp16, y = mean_41_cast_fp16)[name = tensor("sub_32_cast_fp16")]; + tensor square_26_cast_fp16 = square(x = sub_32_cast_fp16)[name = tensor("square_26_cast_fp16")]; + tensor reduce_mean_53_axes_0 = const()[name = tensor("reduce_mean_53_axes_0"), val = tensor([-1])]; + tensor reduce_mean_53_keep_dims_0 = const()[name = tensor("reduce_mean_53_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_53_cast_fp16 = reduce_mean(axes = reduce_mean_53_axes_0, keep_dims = reduce_mean_53_keep_dims_0, x = square_26_cast_fp16)[name = tensor("reduce_mean_53_cast_fp16")]; + tensor var_1168_to_fp16 = const()[name = tensor("op_1168_to_fp16"), val = tensor(0x1.1p-20)]; + tensor var_1169_cast_fp16 = add(x = reduce_mean_53_cast_fp16, y = var_1168_to_fp16)[name = tensor("op_1169_cast_fp16")]; + tensor var_1170_cast_fp16 = sqrt(x = var_1169_cast_fp16)[name = tensor("op_1170_cast_fp16")]; + tensor x_131_cast_fp16 = real_div(x = sub_32_cast_fp16, y = var_1170_cast_fp16)[name = tensor("x_131_cast_fp16")]; + tensor var_1172_promoted_to_fp16 = const()[name = tensor("op_1172_promoted_to_fp16"), val = tensor(0x1p+0)]; + tensor var_1173_cast_fp16 = add(x = var_1161_cast_fp16_1, y = var_1172_promoted_to_fp16)[name = tensor("op_1173_cast_fp16")]; + tensor var_1174_cast_fp16 = mul(x = x_131_cast_fp16, y = var_1173_cast_fp16)[name = tensor("op_1174_cast_fp16")]; + tensor input_201_cast_fp16 = add(x = var_1174_cast_fp16, y = var_1161_cast_fp16_0)[name = tensor("input_201_cast_fp16")]; + tensor linear_77_cast_fp16 = linear(bias = flow_net_final_layer_linear_bias_to_fp16, weight = flow_net_final_layer_linear_weight_to_fp16, x = input_201_cast_fp16)[name = tensor("linear_77_cast_fp16")]; + tensor var_1185_to_fp16 = const()[name = tensor("op_1185_to_fp16"), val = tensor(0x1p-3)]; + tensor var_1186_cast_fp16 = mul(x = linear_77_cast_fp16, y = var_1185_to_fp16)[name = tensor("op_1186_cast_fp16")]; + tensor input_203_cast_fp16 = add(x = input_135_cast_fp16, y = var_1186_cast_fp16)[name = tensor("input_203_cast_fp16")]; + tensor linear_78_cast_fp16 = linear(bias = flow_net_input_proj_bias_to_fp16, weight = flow_net_input_proj_weight_to_fp16, x = input_203_cast_fp16)[name = tensor("linear_78_cast_fp16")]; + tensor input_207_to_fp16 = const()[name = tensor("input_207_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(19001728)))]; + tensor input_209_cast_fp16 = silu(x = input_207_to_fp16)[name = tensor("input_209_cast_fp16")]; + tensor linear_80_cast_fp16 = linear(bias = flow_net_time_embed_0_mlp_2_bias_to_fp16, weight = flow_net_time_embed_0_mlp_2_weight_to_fp16, x = input_209_cast_fp16)[name = tensor("linear_80_cast_fp16")]; + tensor reduce_mean_54_axes_0 = const()[name = tensor("reduce_mean_54_axes_0"), val = tensor([-1])]; + tensor reduce_mean_54_keep_dims_0 = const()[name = tensor("reduce_mean_54_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_54_cast_fp16 = reduce_mean(axes = reduce_mean_54_axes_0, keep_dims = reduce_mean_54_keep_dims_0, x = linear_80_cast_fp16)[name = tensor("reduce_mean_54_cast_fp16")]; + tensor sub_33_cast_fp16 = sub(x = linear_80_cast_fp16, y = reduce_mean_54_cast_fp16)[name = tensor("sub_33_cast_fp16")]; + tensor square_27_cast_fp16 = square(x = sub_33_cast_fp16)[name = tensor("square_27_cast_fp16")]; + tensor reduce_mean_55_axes_0 = const()[name = tensor("reduce_mean_55_axes_0"), val = tensor([-1])]; + tensor reduce_mean_55_keep_dims_0 = const()[name = tensor("reduce_mean_55_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_55_cast_fp16 = reduce_mean(axes = reduce_mean_55_axes_0, keep_dims = reduce_mean_55_keep_dims_0, x = square_27_cast_fp16)[name = tensor("reduce_mean_55_cast_fp16")]; + tensor real_div_6_to_fp16 = const()[name = tensor("real_div_6_to_fp16"), val = tensor(0x1.008p+0)]; + tensor mul_6_cast_fp16 = mul(x = reduce_mean_55_cast_fp16, y = real_div_6_to_fp16)[name = tensor("mul_6_cast_fp16")]; + tensor var_1252_to_fp16 = const()[name = tensor("op_1252_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_55_cast_fp16 = add(x = mul_6_cast_fp16, y = var_1252_to_fp16)[name = tensor("var_55_cast_fp16")]; + tensor var_1255_epsilon_0 = const()[name = tensor("op_1255_epsilon_0"), val = tensor(0x1.197998p-40)]; + tensor var_1255_cast_fp16 = rsqrt(epsilon = var_1255_epsilon_0, x = var_55_cast_fp16)[name = tensor("op_1255_cast_fp16")]; + tensor var_1256_cast_fp16 = mul(x = const_3_to_fp16, y = var_1255_cast_fp16)[name = tensor("op_1256_cast_fp16")]; + tensor var_1257_cast_fp16 = mul(x = linear_80_cast_fp16, y = var_1256_cast_fp16)[name = tensor("op_1257_cast_fp16")]; + tensor input_213_to_fp16 = const()[name = tensor("input_213_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(19002816)))]; + tensor input_215_cast_fp16 = silu(x = input_213_to_fp16)[name = tensor("input_215_cast_fp16")]; + tensor linear_82_cast_fp16 = linear(bias = flow_net_time_embed_1_mlp_2_bias_to_fp16, weight = flow_net_time_embed_1_mlp_2_weight_to_fp16, x = input_215_cast_fp16)[name = tensor("linear_82_cast_fp16")]; + tensor reduce_mean_56_axes_0 = const()[name = tensor("reduce_mean_56_axes_0"), val = tensor([-1])]; + tensor reduce_mean_56_keep_dims_0 = const()[name = tensor("reduce_mean_56_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_56_cast_fp16 = reduce_mean(axes = reduce_mean_56_axes_0, keep_dims = reduce_mean_56_keep_dims_0, x = linear_82_cast_fp16)[name = tensor("reduce_mean_56_cast_fp16")]; + tensor sub_35_cast_fp16 = sub(x = linear_82_cast_fp16, y = reduce_mean_56_cast_fp16)[name = tensor("sub_35_cast_fp16")]; + tensor square_28_cast_fp16 = square(x = sub_35_cast_fp16)[name = tensor("square_28_cast_fp16")]; + tensor reduce_mean_57_axes_0 = const()[name = tensor("reduce_mean_57_axes_0"), val = tensor([-1])]; + tensor reduce_mean_57_keep_dims_0 = const()[name = tensor("reduce_mean_57_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_57_cast_fp16 = reduce_mean(axes = reduce_mean_57_axes_0, keep_dims = reduce_mean_57_keep_dims_0, x = square_28_cast_fp16)[name = tensor("reduce_mean_57_cast_fp16")]; + tensor real_div_7_to_fp16 = const()[name = tensor("real_div_7_to_fp16"), val = tensor(0x1.008p+0)]; + tensor mul_7_cast_fp16 = mul(x = reduce_mean_57_cast_fp16, y = real_div_7_to_fp16)[name = tensor("mul_7_cast_fp16")]; + tensor var_1289_to_fp16 = const()[name = tensor("op_1289_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_57_cast_fp16 = add(x = mul_7_cast_fp16, y = var_1289_to_fp16)[name = tensor("var_57_cast_fp16")]; + tensor var_1292_epsilon_0 = const()[name = tensor("op_1292_epsilon_0"), val = tensor(0x1.197998p-40)]; + tensor var_1292_cast_fp16 = rsqrt(epsilon = var_1292_epsilon_0, x = var_57_cast_fp16)[name = tensor("op_1292_cast_fp16")]; + tensor var_1293_cast_fp16 = mul(x = const_5_to_fp16, y = var_1292_cast_fp16)[name = tensor("op_1293_cast_fp16")]; + tensor var_1294_cast_fp16 = mul(x = linear_82_cast_fp16, y = var_1293_cast_fp16)[name = tensor("op_1294_cast_fp16")]; + tensor var_1306_cast_fp16 = add(x = var_1257_cast_fp16, y = var_1294_cast_fp16)[name = tensor("op_1306_cast_fp16")]; + tensor _inversed_t_combined_7_y_0_to_fp16 = const()[name = tensor("_inversed_t_combined_7_y_0_to_fp16"), val = tensor(0x1p-1)]; + tensor _inversed_t_combined_7_cast_fp16 = mul(x = var_1306_cast_fp16, y = _inversed_t_combined_7_y_0_to_fp16)[name = tensor("_inversed_t_combined_7_cast_fp16")]; + tensor input_217_cast_fp16 = add(x = _inversed_t_combined_7_cast_fp16, y = linear_5_cast_fp16)[name = tensor("input_217_cast_fp16")]; + tensor input_219_cast_fp16 = silu(x = input_217_cast_fp16)[name = tensor("input_219_cast_fp16")]; + tensor linear_84_cast_fp16 = linear(bias = flow_net_res_blocks_0_adaLN_modulation_1_bias_to_fp16, weight = flow_net_res_blocks_0_adaLN_modulation_1_weight_to_fp16, x = input_219_cast_fp16)[name = tensor("linear_84_cast_fp16")]; + tensor var_1321_split_sizes_0 = const()[name = tensor("op_1321_split_sizes_0"), val = tensor([512, 512, 512])]; + tensor var_1321_axis_0 = const()[name = tensor("op_1321_axis_0"), val = tensor(-1)]; + tensor var_1321_cast_fp16_0, tensor var_1321_cast_fp16_1, tensor var_1321_cast_fp16_2 = split(axis = var_1321_axis_0, split_sizes = var_1321_split_sizes_0, x = linear_84_cast_fp16)[name = tensor("op_1321_cast_fp16")]; + tensor mean_43_axes_0 = const()[name = tensor("mean_43_axes_0"), val = tensor([-1])]; + tensor mean_43_keep_dims_0 = const()[name = tensor("mean_43_keep_dims_0"), val = tensor(true)]; + tensor mean_43_cast_fp16 = reduce_mean(axes = mean_43_axes_0, keep_dims = mean_43_keep_dims_0, x = linear_78_cast_fp16)[name = tensor("mean_43_cast_fp16")]; + tensor sub_37_cast_fp16 = sub(x = linear_78_cast_fp16, y = mean_43_cast_fp16)[name = tensor("sub_37_cast_fp16")]; + tensor square_29_cast_fp16 = square(x = sub_37_cast_fp16)[name = tensor("square_29_cast_fp16")]; + tensor reduce_mean_59_axes_0 = const()[name = tensor("reduce_mean_59_axes_0"), val = tensor([-1])]; + tensor reduce_mean_59_keep_dims_0 = const()[name = tensor("reduce_mean_59_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_59_cast_fp16 = reduce_mean(axes = reduce_mean_59_axes_0, keep_dims = reduce_mean_59_keep_dims_0, x = square_29_cast_fp16)[name = tensor("reduce_mean_59_cast_fp16")]; + tensor var_1331_to_fp16 = const()[name = tensor("op_1331_to_fp16"), val = tensor(0x1.1p-20)]; + tensor var_1332_cast_fp16 = add(x = reduce_mean_59_cast_fp16, y = var_1331_to_fp16)[name = tensor("op_1332_cast_fp16")]; + tensor var_1333_cast_fp16 = sqrt(x = var_1332_cast_fp16)[name = tensor("op_1333_cast_fp16")]; + tensor x_139_cast_fp16 = real_div(x = sub_37_cast_fp16, y = var_1333_cast_fp16)[name = tensor("x_139_cast_fp16")]; + tensor var_1335_cast_fp16 = mul(x = x_139_cast_fp16, y = flow_net_res_blocks_0_in_ln_weight_to_fp16)[name = tensor("op_1335_cast_fp16")]; + tensor x_141_cast_fp16 = add(x = var_1335_cast_fp16, y = flow_net_res_blocks_0_in_ln_bias_to_fp16)[name = tensor("x_141_cast_fp16")]; + tensor var_1337_promoted_to_fp16 = const()[name = tensor("op_1337_promoted_to_fp16"), val = tensor(0x1p+0)]; + tensor var_1338_cast_fp16 = add(x = var_1321_cast_fp16_1, y = var_1337_promoted_to_fp16)[name = tensor("op_1338_cast_fp16")]; + tensor var_1339_cast_fp16 = mul(x = x_141_cast_fp16, y = var_1338_cast_fp16)[name = tensor("op_1339_cast_fp16")]; + tensor input_221_cast_fp16 = add(x = var_1339_cast_fp16, y = var_1321_cast_fp16_0)[name = tensor("input_221_cast_fp16")]; + tensor linear_85_cast_fp16 = linear(bias = flow_net_res_blocks_0_mlp_0_bias_to_fp16, weight = flow_net_res_blocks_0_mlp_0_weight_to_fp16, x = input_221_cast_fp16)[name = tensor("linear_85_cast_fp16")]; + tensor input_225_cast_fp16 = silu(x = linear_85_cast_fp16)[name = tensor("input_225_cast_fp16")]; + tensor linear_86_cast_fp16 = linear(bias = flow_net_res_blocks_0_mlp_2_bias_to_fp16, weight = flow_net_res_blocks_0_mlp_2_weight_to_fp16, x = input_225_cast_fp16)[name = tensor("linear_86_cast_fp16")]; + tensor var_1350_cast_fp16 = mul(x = var_1321_cast_fp16_2, y = linear_86_cast_fp16)[name = tensor("op_1350_cast_fp16")]; + tensor x_143_cast_fp16 = add(x = linear_78_cast_fp16, y = var_1350_cast_fp16)[name = tensor("x_143_cast_fp16")]; + tensor linear_87_cast_fp16 = linear(bias = flow_net_res_blocks_1_adaLN_modulation_1_bias_to_fp16, weight = flow_net_res_blocks_1_adaLN_modulation_1_weight_to_fp16, x = input_219_cast_fp16)[name = tensor("linear_87_cast_fp16")]; + tensor var_1360_split_sizes_0 = const()[name = tensor("op_1360_split_sizes_0"), val = tensor([512, 512, 512])]; + tensor var_1360_axis_0 = const()[name = tensor("op_1360_axis_0"), val = tensor(-1)]; + tensor var_1360_cast_fp16_0, tensor var_1360_cast_fp16_1, tensor var_1360_cast_fp16_2 = split(axis = var_1360_axis_0, split_sizes = var_1360_split_sizes_0, x = linear_87_cast_fp16)[name = tensor("op_1360_cast_fp16")]; + tensor mean_45_axes_0 = const()[name = tensor("mean_45_axes_0"), val = tensor([-1])]; + tensor mean_45_keep_dims_0 = const()[name = tensor("mean_45_keep_dims_0"), val = tensor(true)]; + tensor mean_45_cast_fp16 = reduce_mean(axes = mean_45_axes_0, keep_dims = mean_45_keep_dims_0, x = x_143_cast_fp16)[name = tensor("mean_45_cast_fp16")]; + tensor sub_38_cast_fp16 = sub(x = x_143_cast_fp16, y = mean_45_cast_fp16)[name = tensor("sub_38_cast_fp16")]; + tensor square_30_cast_fp16 = square(x = sub_38_cast_fp16)[name = tensor("square_30_cast_fp16")]; + tensor reduce_mean_61_axes_0 = const()[name = tensor("reduce_mean_61_axes_0"), val = tensor([-1])]; + tensor reduce_mean_61_keep_dims_0 = const()[name = tensor("reduce_mean_61_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_61_cast_fp16 = reduce_mean(axes = reduce_mean_61_axes_0, keep_dims = reduce_mean_61_keep_dims_0, x = square_30_cast_fp16)[name = tensor("reduce_mean_61_cast_fp16")]; + tensor var_1370_to_fp16 = const()[name = tensor("op_1370_to_fp16"), val = tensor(0x1.1p-20)]; + tensor var_1371_cast_fp16 = add(x = reduce_mean_61_cast_fp16, y = var_1370_to_fp16)[name = tensor("op_1371_cast_fp16")]; + tensor var_1372_cast_fp16 = sqrt(x = var_1371_cast_fp16)[name = tensor("op_1372_cast_fp16")]; + tensor x_145_cast_fp16 = real_div(x = sub_38_cast_fp16, y = var_1372_cast_fp16)[name = tensor("x_145_cast_fp16")]; + tensor var_1374_cast_fp16 = mul(x = x_145_cast_fp16, y = flow_net_res_blocks_1_in_ln_weight_to_fp16)[name = tensor("op_1374_cast_fp16")]; + tensor x_147_cast_fp16 = add(x = var_1374_cast_fp16, y = flow_net_res_blocks_1_in_ln_bias_to_fp16)[name = tensor("x_147_cast_fp16")]; + tensor var_1376_promoted_to_fp16 = const()[name = tensor("op_1376_promoted_to_fp16"), val = tensor(0x1p+0)]; + tensor var_1377_cast_fp16 = add(x = var_1360_cast_fp16_1, y = var_1376_promoted_to_fp16)[name = tensor("op_1377_cast_fp16")]; + tensor var_1378_cast_fp16 = mul(x = x_147_cast_fp16, y = var_1377_cast_fp16)[name = tensor("op_1378_cast_fp16")]; + tensor input_229_cast_fp16 = add(x = var_1378_cast_fp16, y = var_1360_cast_fp16_0)[name = tensor("input_229_cast_fp16")]; + tensor linear_88_cast_fp16 = linear(bias = flow_net_res_blocks_1_mlp_0_bias_to_fp16, weight = flow_net_res_blocks_1_mlp_0_weight_to_fp16, x = input_229_cast_fp16)[name = tensor("linear_88_cast_fp16")]; + tensor input_233_cast_fp16 = silu(x = linear_88_cast_fp16)[name = tensor("input_233_cast_fp16")]; + tensor linear_89_cast_fp16 = linear(bias = flow_net_res_blocks_1_mlp_2_bias_to_fp16, weight = flow_net_res_blocks_1_mlp_2_weight_to_fp16, x = input_233_cast_fp16)[name = tensor("linear_89_cast_fp16")]; + tensor var_1389_cast_fp16 = mul(x = var_1360_cast_fp16_2, y = linear_89_cast_fp16)[name = tensor("op_1389_cast_fp16")]; + tensor x_149_cast_fp16 = add(x = x_143_cast_fp16, y = var_1389_cast_fp16)[name = tensor("x_149_cast_fp16")]; + tensor linear_90_cast_fp16 = linear(bias = flow_net_res_blocks_2_adaLN_modulation_1_bias_to_fp16, weight = flow_net_res_blocks_2_adaLN_modulation_1_weight_to_fp16, x = input_219_cast_fp16)[name = tensor("linear_90_cast_fp16")]; + tensor var_1399_split_sizes_0 = const()[name = tensor("op_1399_split_sizes_0"), val = tensor([512, 512, 512])]; + tensor var_1399_axis_0 = const()[name = tensor("op_1399_axis_0"), val = tensor(-1)]; + tensor var_1399_cast_fp16_0, tensor var_1399_cast_fp16_1, tensor var_1399_cast_fp16_2 = split(axis = var_1399_axis_0, split_sizes = var_1399_split_sizes_0, x = linear_90_cast_fp16)[name = tensor("op_1399_cast_fp16")]; + tensor mean_47_axes_0 = const()[name = tensor("mean_47_axes_0"), val = tensor([-1])]; + tensor mean_47_keep_dims_0 = const()[name = tensor("mean_47_keep_dims_0"), val = tensor(true)]; + tensor mean_47_cast_fp16 = reduce_mean(axes = mean_47_axes_0, keep_dims = mean_47_keep_dims_0, x = x_149_cast_fp16)[name = tensor("mean_47_cast_fp16")]; + tensor sub_39_cast_fp16 = sub(x = x_149_cast_fp16, y = mean_47_cast_fp16)[name = tensor("sub_39_cast_fp16")]; + tensor square_31_cast_fp16 = square(x = sub_39_cast_fp16)[name = tensor("square_31_cast_fp16")]; + tensor reduce_mean_63_axes_0 = const()[name = tensor("reduce_mean_63_axes_0"), val = tensor([-1])]; + tensor reduce_mean_63_keep_dims_0 = const()[name = tensor("reduce_mean_63_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_63_cast_fp16 = reduce_mean(axes = reduce_mean_63_axes_0, keep_dims = reduce_mean_63_keep_dims_0, x = square_31_cast_fp16)[name = tensor("reduce_mean_63_cast_fp16")]; + tensor var_1409_to_fp16 = const()[name = tensor("op_1409_to_fp16"), val = tensor(0x1.1p-20)]; + tensor var_1410_cast_fp16 = add(x = reduce_mean_63_cast_fp16, y = var_1409_to_fp16)[name = tensor("op_1410_cast_fp16")]; + tensor var_1411_cast_fp16 = sqrt(x = var_1410_cast_fp16)[name = tensor("op_1411_cast_fp16")]; + tensor x_151_cast_fp16 = real_div(x = sub_39_cast_fp16, y = var_1411_cast_fp16)[name = tensor("x_151_cast_fp16")]; + tensor var_1413_cast_fp16 = mul(x = x_151_cast_fp16, y = flow_net_res_blocks_2_in_ln_weight_to_fp16)[name = tensor("op_1413_cast_fp16")]; + tensor x_153_cast_fp16 = add(x = var_1413_cast_fp16, y = flow_net_res_blocks_2_in_ln_bias_to_fp16)[name = tensor("x_153_cast_fp16")]; + tensor var_1415_promoted_to_fp16 = const()[name = tensor("op_1415_promoted_to_fp16"), val = tensor(0x1p+0)]; + tensor var_1416_cast_fp16 = add(x = var_1399_cast_fp16_1, y = var_1415_promoted_to_fp16)[name = tensor("op_1416_cast_fp16")]; + tensor var_1417_cast_fp16 = mul(x = x_153_cast_fp16, y = var_1416_cast_fp16)[name = tensor("op_1417_cast_fp16")]; + tensor input_237_cast_fp16 = add(x = var_1417_cast_fp16, y = var_1399_cast_fp16_0)[name = tensor("input_237_cast_fp16")]; + tensor linear_91_cast_fp16 = linear(bias = flow_net_res_blocks_2_mlp_0_bias_to_fp16, weight = flow_net_res_blocks_2_mlp_0_weight_to_fp16, x = input_237_cast_fp16)[name = tensor("linear_91_cast_fp16")]; + tensor input_241_cast_fp16 = silu(x = linear_91_cast_fp16)[name = tensor("input_241_cast_fp16")]; + tensor linear_92_cast_fp16 = linear(bias = flow_net_res_blocks_2_mlp_2_bias_to_fp16, weight = flow_net_res_blocks_2_mlp_2_weight_to_fp16, x = input_241_cast_fp16)[name = tensor("linear_92_cast_fp16")]; + tensor var_1428_cast_fp16 = mul(x = var_1399_cast_fp16_2, y = linear_92_cast_fp16)[name = tensor("op_1428_cast_fp16")]; + tensor x_155_cast_fp16 = add(x = x_149_cast_fp16, y = var_1428_cast_fp16)[name = tensor("x_155_cast_fp16")]; + tensor linear_93_cast_fp16 = linear(bias = flow_net_res_blocks_3_adaLN_modulation_1_bias_to_fp16, weight = flow_net_res_blocks_3_adaLN_modulation_1_weight_to_fp16, x = input_219_cast_fp16)[name = tensor("linear_93_cast_fp16")]; + tensor var_1438_split_sizes_0 = const()[name = tensor("op_1438_split_sizes_0"), val = tensor([512, 512, 512])]; + tensor var_1438_axis_0 = const()[name = tensor("op_1438_axis_0"), val = tensor(-1)]; + tensor var_1438_cast_fp16_0, tensor var_1438_cast_fp16_1, tensor var_1438_cast_fp16_2 = split(axis = var_1438_axis_0, split_sizes = var_1438_split_sizes_0, x = linear_93_cast_fp16)[name = tensor("op_1438_cast_fp16")]; + tensor mean_49_axes_0 = const()[name = tensor("mean_49_axes_0"), val = tensor([-1])]; + tensor mean_49_keep_dims_0 = const()[name = tensor("mean_49_keep_dims_0"), val = tensor(true)]; + tensor mean_49_cast_fp16 = reduce_mean(axes = mean_49_axes_0, keep_dims = mean_49_keep_dims_0, x = x_155_cast_fp16)[name = tensor("mean_49_cast_fp16")]; + tensor sub_40_cast_fp16 = sub(x = x_155_cast_fp16, y = mean_49_cast_fp16)[name = tensor("sub_40_cast_fp16")]; + tensor square_32_cast_fp16 = square(x = sub_40_cast_fp16)[name = tensor("square_32_cast_fp16")]; + tensor reduce_mean_65_axes_0 = const()[name = tensor("reduce_mean_65_axes_0"), val = tensor([-1])]; + tensor reduce_mean_65_keep_dims_0 = const()[name = tensor("reduce_mean_65_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_65_cast_fp16 = reduce_mean(axes = reduce_mean_65_axes_0, keep_dims = reduce_mean_65_keep_dims_0, x = square_32_cast_fp16)[name = tensor("reduce_mean_65_cast_fp16")]; + tensor var_1448_to_fp16 = const()[name = tensor("op_1448_to_fp16"), val = tensor(0x1.1p-20)]; + tensor var_1449_cast_fp16 = add(x = reduce_mean_65_cast_fp16, y = var_1448_to_fp16)[name = tensor("op_1449_cast_fp16")]; + tensor var_1450_cast_fp16 = sqrt(x = var_1449_cast_fp16)[name = tensor("op_1450_cast_fp16")]; + tensor x_157_cast_fp16 = real_div(x = sub_40_cast_fp16, y = var_1450_cast_fp16)[name = tensor("x_157_cast_fp16")]; + tensor var_1452_cast_fp16 = mul(x = x_157_cast_fp16, y = flow_net_res_blocks_3_in_ln_weight_to_fp16)[name = tensor("op_1452_cast_fp16")]; + tensor x_159_cast_fp16 = add(x = var_1452_cast_fp16, y = flow_net_res_blocks_3_in_ln_bias_to_fp16)[name = tensor("x_159_cast_fp16")]; + tensor var_1454_promoted_to_fp16 = const()[name = tensor("op_1454_promoted_to_fp16"), val = tensor(0x1p+0)]; + tensor var_1455_cast_fp16 = add(x = var_1438_cast_fp16_1, y = var_1454_promoted_to_fp16)[name = tensor("op_1455_cast_fp16")]; + tensor var_1456_cast_fp16 = mul(x = x_159_cast_fp16, y = var_1455_cast_fp16)[name = tensor("op_1456_cast_fp16")]; + tensor input_245_cast_fp16 = add(x = var_1456_cast_fp16, y = var_1438_cast_fp16_0)[name = tensor("input_245_cast_fp16")]; + tensor linear_94_cast_fp16 = linear(bias = flow_net_res_blocks_3_mlp_0_bias_to_fp16, weight = flow_net_res_blocks_3_mlp_0_weight_to_fp16, x = input_245_cast_fp16)[name = tensor("linear_94_cast_fp16")]; + tensor input_249_cast_fp16 = silu(x = linear_94_cast_fp16)[name = tensor("input_249_cast_fp16")]; + tensor linear_95_cast_fp16 = linear(bias = flow_net_res_blocks_3_mlp_2_bias_to_fp16, weight = flow_net_res_blocks_3_mlp_2_weight_to_fp16, x = input_249_cast_fp16)[name = tensor("linear_95_cast_fp16")]; + tensor var_1467_cast_fp16 = mul(x = var_1438_cast_fp16_2, y = linear_95_cast_fp16)[name = tensor("op_1467_cast_fp16")]; + tensor x_161_cast_fp16 = add(x = x_155_cast_fp16, y = var_1467_cast_fp16)[name = tensor("x_161_cast_fp16")]; + tensor linear_96_cast_fp16 = linear(bias = flow_net_res_blocks_4_adaLN_modulation_1_bias_to_fp16, weight = flow_net_res_blocks_4_adaLN_modulation_1_weight_to_fp16, x = input_219_cast_fp16)[name = tensor("linear_96_cast_fp16")]; + tensor var_1477_split_sizes_0 = const()[name = tensor("op_1477_split_sizes_0"), val = tensor([512, 512, 512])]; + tensor var_1477_axis_0 = const()[name = tensor("op_1477_axis_0"), val = tensor(-1)]; + tensor var_1477_cast_fp16_0, tensor var_1477_cast_fp16_1, tensor var_1477_cast_fp16_2 = split(axis = var_1477_axis_0, split_sizes = var_1477_split_sizes_0, x = linear_96_cast_fp16)[name = tensor("op_1477_cast_fp16")]; + tensor mean_51_axes_0 = const()[name = tensor("mean_51_axes_0"), val = tensor([-1])]; + tensor mean_51_keep_dims_0 = const()[name = tensor("mean_51_keep_dims_0"), val = tensor(true)]; + tensor mean_51_cast_fp16 = reduce_mean(axes = mean_51_axes_0, keep_dims = mean_51_keep_dims_0, x = x_161_cast_fp16)[name = tensor("mean_51_cast_fp16")]; + tensor sub_41_cast_fp16 = sub(x = x_161_cast_fp16, y = mean_51_cast_fp16)[name = tensor("sub_41_cast_fp16")]; + tensor square_33_cast_fp16 = square(x = sub_41_cast_fp16)[name = tensor("square_33_cast_fp16")]; + tensor reduce_mean_67_axes_0 = const()[name = tensor("reduce_mean_67_axes_0"), val = tensor([-1])]; + tensor reduce_mean_67_keep_dims_0 = const()[name = tensor("reduce_mean_67_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_67_cast_fp16 = reduce_mean(axes = reduce_mean_67_axes_0, keep_dims = reduce_mean_67_keep_dims_0, x = square_33_cast_fp16)[name = tensor("reduce_mean_67_cast_fp16")]; + tensor var_1487_to_fp16 = const()[name = tensor("op_1487_to_fp16"), val = tensor(0x1.1p-20)]; + tensor var_1488_cast_fp16 = add(x = reduce_mean_67_cast_fp16, y = var_1487_to_fp16)[name = tensor("op_1488_cast_fp16")]; + tensor var_1489_cast_fp16 = sqrt(x = var_1488_cast_fp16)[name = tensor("op_1489_cast_fp16")]; + tensor x_163_cast_fp16 = real_div(x = sub_41_cast_fp16, y = var_1489_cast_fp16)[name = tensor("x_163_cast_fp16")]; + tensor var_1491_cast_fp16 = mul(x = x_163_cast_fp16, y = flow_net_res_blocks_4_in_ln_weight_to_fp16)[name = tensor("op_1491_cast_fp16")]; + tensor x_165_cast_fp16 = add(x = var_1491_cast_fp16, y = flow_net_res_blocks_4_in_ln_bias_to_fp16)[name = tensor("x_165_cast_fp16")]; + tensor var_1493_promoted_to_fp16 = const()[name = tensor("op_1493_promoted_to_fp16"), val = tensor(0x1p+0)]; + tensor var_1494_cast_fp16 = add(x = var_1477_cast_fp16_1, y = var_1493_promoted_to_fp16)[name = tensor("op_1494_cast_fp16")]; + tensor var_1495_cast_fp16 = mul(x = x_165_cast_fp16, y = var_1494_cast_fp16)[name = tensor("op_1495_cast_fp16")]; + tensor input_253_cast_fp16 = add(x = var_1495_cast_fp16, y = var_1477_cast_fp16_0)[name = tensor("input_253_cast_fp16")]; + tensor linear_97_cast_fp16 = linear(bias = flow_net_res_blocks_4_mlp_0_bias_to_fp16, weight = flow_net_res_blocks_4_mlp_0_weight_to_fp16, x = input_253_cast_fp16)[name = tensor("linear_97_cast_fp16")]; + tensor input_257_cast_fp16 = silu(x = linear_97_cast_fp16)[name = tensor("input_257_cast_fp16")]; + tensor linear_98_cast_fp16 = linear(bias = flow_net_res_blocks_4_mlp_2_bias_to_fp16, weight = flow_net_res_blocks_4_mlp_2_weight_to_fp16, x = input_257_cast_fp16)[name = tensor("linear_98_cast_fp16")]; + tensor var_1506_cast_fp16 = mul(x = var_1477_cast_fp16_2, y = linear_98_cast_fp16)[name = tensor("op_1506_cast_fp16")]; + tensor x_167_cast_fp16 = add(x = x_161_cast_fp16, y = var_1506_cast_fp16)[name = tensor("x_167_cast_fp16")]; + tensor linear_99_cast_fp16 = linear(bias = flow_net_res_blocks_5_adaLN_modulation_1_bias_to_fp16, weight = flow_net_res_blocks_5_adaLN_modulation_1_weight_to_fp16, x = input_219_cast_fp16)[name = tensor("linear_99_cast_fp16")]; + tensor var_1516_split_sizes_0 = const()[name = tensor("op_1516_split_sizes_0"), val = tensor([512, 512, 512])]; + tensor var_1516_axis_0 = const()[name = tensor("op_1516_axis_0"), val = tensor(-1)]; + tensor var_1516_cast_fp16_0, tensor var_1516_cast_fp16_1, tensor var_1516_cast_fp16_2 = split(axis = var_1516_axis_0, split_sizes = var_1516_split_sizes_0, x = linear_99_cast_fp16)[name = tensor("op_1516_cast_fp16")]; + tensor mean_53_axes_0 = const()[name = tensor("mean_53_axes_0"), val = tensor([-1])]; + tensor mean_53_keep_dims_0 = const()[name = tensor("mean_53_keep_dims_0"), val = tensor(true)]; + tensor mean_53_cast_fp16 = reduce_mean(axes = mean_53_axes_0, keep_dims = mean_53_keep_dims_0, x = x_167_cast_fp16)[name = tensor("mean_53_cast_fp16")]; + tensor sub_42_cast_fp16 = sub(x = x_167_cast_fp16, y = mean_53_cast_fp16)[name = tensor("sub_42_cast_fp16")]; + tensor square_34_cast_fp16 = square(x = sub_42_cast_fp16)[name = tensor("square_34_cast_fp16")]; + tensor reduce_mean_69_axes_0 = const()[name = tensor("reduce_mean_69_axes_0"), val = tensor([-1])]; + tensor reduce_mean_69_keep_dims_0 = const()[name = tensor("reduce_mean_69_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_69_cast_fp16 = reduce_mean(axes = reduce_mean_69_axes_0, keep_dims = reduce_mean_69_keep_dims_0, x = square_34_cast_fp16)[name = tensor("reduce_mean_69_cast_fp16")]; + tensor var_1526_to_fp16 = const()[name = tensor("op_1526_to_fp16"), val = tensor(0x1.1p-20)]; + tensor var_1527_cast_fp16 = add(x = reduce_mean_69_cast_fp16, y = var_1526_to_fp16)[name = tensor("op_1527_cast_fp16")]; + tensor var_1528_cast_fp16 = sqrt(x = var_1527_cast_fp16)[name = tensor("op_1528_cast_fp16")]; + tensor x_169_cast_fp16 = real_div(x = sub_42_cast_fp16, y = var_1528_cast_fp16)[name = tensor("x_169_cast_fp16")]; + tensor var_1530_cast_fp16 = mul(x = x_169_cast_fp16, y = flow_net_res_blocks_5_in_ln_weight_to_fp16)[name = tensor("op_1530_cast_fp16")]; + tensor x_171_cast_fp16 = add(x = var_1530_cast_fp16, y = flow_net_res_blocks_5_in_ln_bias_to_fp16)[name = tensor("x_171_cast_fp16")]; + tensor var_1532_promoted_to_fp16 = const()[name = tensor("op_1532_promoted_to_fp16"), val = tensor(0x1p+0)]; + tensor var_1533_cast_fp16 = add(x = var_1516_cast_fp16_1, y = var_1532_promoted_to_fp16)[name = tensor("op_1533_cast_fp16")]; + tensor var_1534_cast_fp16 = mul(x = x_171_cast_fp16, y = var_1533_cast_fp16)[name = tensor("op_1534_cast_fp16")]; + tensor input_261_cast_fp16 = add(x = var_1534_cast_fp16, y = var_1516_cast_fp16_0)[name = tensor("input_261_cast_fp16")]; + tensor linear_100_cast_fp16 = linear(bias = flow_net_res_blocks_5_mlp_0_bias_to_fp16, weight = flow_net_res_blocks_5_mlp_0_weight_to_fp16, x = input_261_cast_fp16)[name = tensor("linear_100_cast_fp16")]; + tensor input_265_cast_fp16 = silu(x = linear_100_cast_fp16)[name = tensor("input_265_cast_fp16")]; + tensor linear_101_cast_fp16 = linear(bias = flow_net_res_blocks_5_mlp_2_bias_to_fp16, weight = flow_net_res_blocks_5_mlp_2_weight_to_fp16, x = input_265_cast_fp16)[name = tensor("linear_101_cast_fp16")]; + tensor var_1545_cast_fp16 = mul(x = var_1516_cast_fp16_2, y = linear_101_cast_fp16)[name = tensor("op_1545_cast_fp16")]; + tensor x_173_cast_fp16 = add(x = x_167_cast_fp16, y = var_1545_cast_fp16)[name = tensor("x_173_cast_fp16")]; + tensor linear_102_cast_fp16 = linear(bias = flow_net_final_layer_adaLN_modulation_1_bias_to_fp16, weight = flow_net_final_layer_adaLN_modulation_1_weight_to_fp16, x = input_219_cast_fp16)[name = tensor("linear_102_cast_fp16")]; + tensor var_1554_split_sizes_0 = const()[name = tensor("op_1554_split_sizes_0"), val = tensor([512, 512])]; + tensor var_1554_axis_0 = const()[name = tensor("op_1554_axis_0"), val = tensor(-1)]; + tensor var_1554_cast_fp16_0, tensor var_1554_cast_fp16_1 = split(axis = var_1554_axis_0, split_sizes = var_1554_split_sizes_0, x = linear_102_cast_fp16)[name = tensor("op_1554_cast_fp16")]; + tensor mean_55_axes_0 = const()[name = tensor("mean_55_axes_0"), val = tensor([-1])]; + tensor mean_55_keep_dims_0 = const()[name = tensor("mean_55_keep_dims_0"), val = tensor(true)]; + tensor mean_55_cast_fp16 = reduce_mean(axes = mean_55_axes_0, keep_dims = mean_55_keep_dims_0, x = x_173_cast_fp16)[name = tensor("mean_55_cast_fp16")]; + tensor sub_43_cast_fp16 = sub(x = x_173_cast_fp16, y = mean_55_cast_fp16)[name = tensor("sub_43_cast_fp16")]; + tensor square_35_cast_fp16 = square(x = sub_43_cast_fp16)[name = tensor("square_35_cast_fp16")]; + tensor reduce_mean_71_axes_0 = const()[name = tensor("reduce_mean_71_axes_0"), val = tensor([-1])]; + tensor reduce_mean_71_keep_dims_0 = const()[name = tensor("reduce_mean_71_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_71_cast_fp16 = reduce_mean(axes = reduce_mean_71_axes_0, keep_dims = reduce_mean_71_keep_dims_0, x = square_35_cast_fp16)[name = tensor("reduce_mean_71_cast_fp16")]; + tensor var_1561_to_fp16 = const()[name = tensor("op_1561_to_fp16"), val = tensor(0x1.1p-20)]; + tensor var_1562_cast_fp16 = add(x = reduce_mean_71_cast_fp16, y = var_1561_to_fp16)[name = tensor("op_1562_cast_fp16")]; + tensor var_1563_cast_fp16 = sqrt(x = var_1562_cast_fp16)[name = tensor("op_1563_cast_fp16")]; + tensor x_175_cast_fp16 = real_div(x = sub_43_cast_fp16, y = var_1563_cast_fp16)[name = tensor("x_175_cast_fp16")]; + tensor var_1565_promoted_to_fp16 = const()[name = tensor("op_1565_promoted_to_fp16"), val = tensor(0x1p+0)]; + tensor var_1566_cast_fp16 = add(x = var_1554_cast_fp16_1, y = var_1565_promoted_to_fp16)[name = tensor("op_1566_cast_fp16")]; + tensor var_1567_cast_fp16 = mul(x = x_175_cast_fp16, y = var_1566_cast_fp16)[name = tensor("op_1567_cast_fp16")]; + tensor input_269_cast_fp16 = add(x = var_1567_cast_fp16, y = var_1554_cast_fp16_0)[name = tensor("input_269_cast_fp16")]; + tensor linear_103_cast_fp16 = linear(bias = flow_net_final_layer_linear_bias_to_fp16, weight = flow_net_final_layer_linear_weight_to_fp16, x = input_269_cast_fp16)[name = tensor("linear_103_cast_fp16")]; + tensor var_1578_to_fp16 = const()[name = tensor("op_1578_to_fp16"), val = tensor(0x1p-3)]; + tensor var_1579_cast_fp16 = mul(x = linear_103_cast_fp16, y = var_1578_to_fp16)[name = tensor("op_1579_cast_fp16")]; + tensor input_271_cast_fp16 = add(x = input_203_cast_fp16, y = var_1579_cast_fp16)[name = tensor("input_271_cast_fp16")]; + tensor linear_104_cast_fp16 = linear(bias = flow_net_input_proj_bias_to_fp16, weight = flow_net_input_proj_weight_to_fp16, x = input_271_cast_fp16)[name = tensor("linear_104_cast_fp16")]; + tensor input_275_to_fp16 = const()[name = tensor("input_275_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(19003904)))]; + tensor input_277_cast_fp16 = silu(x = input_275_to_fp16)[name = tensor("input_277_cast_fp16")]; + tensor linear_106_cast_fp16 = linear(bias = flow_net_time_embed_0_mlp_2_bias_to_fp16, weight = flow_net_time_embed_0_mlp_2_weight_to_fp16, x = input_277_cast_fp16)[name = tensor("linear_106_cast_fp16")]; + tensor reduce_mean_72_axes_0 = const()[name = tensor("reduce_mean_72_axes_0"), val = tensor([-1])]; + tensor reduce_mean_72_keep_dims_0 = const()[name = tensor("reduce_mean_72_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_72_cast_fp16 = reduce_mean(axes = reduce_mean_72_axes_0, keep_dims = reduce_mean_72_keep_dims_0, x = linear_106_cast_fp16)[name = tensor("reduce_mean_72_cast_fp16")]; + tensor sub_44_cast_fp16 = sub(x = linear_106_cast_fp16, y = reduce_mean_72_cast_fp16)[name = tensor("sub_44_cast_fp16")]; + tensor square_36_cast_fp16 = square(x = sub_44_cast_fp16)[name = tensor("square_36_cast_fp16")]; + tensor reduce_mean_73_axes_0 = const()[name = tensor("reduce_mean_73_axes_0"), val = tensor([-1])]; + tensor reduce_mean_73_keep_dims_0 = const()[name = tensor("reduce_mean_73_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_73_cast_fp16 = reduce_mean(axes = reduce_mean_73_axes_0, keep_dims = reduce_mean_73_keep_dims_0, x = square_36_cast_fp16)[name = tensor("reduce_mean_73_cast_fp16")]; + tensor real_div_8_to_fp16 = const()[name = tensor("real_div_8_to_fp16"), val = tensor(0x1.008p+0)]; + tensor mul_8_cast_fp16 = mul(x = reduce_mean_73_cast_fp16, y = real_div_8_to_fp16)[name = tensor("mul_8_cast_fp16")]; + tensor var_1645_to_fp16 = const()[name = tensor("op_1645_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_73_cast_fp16_0 = add(x = mul_8_cast_fp16, y = var_1645_to_fp16)[name = tensor("var_73_cast_fp16")]; + tensor var_1648_epsilon_0 = const()[name = tensor("op_1648_epsilon_0"), val = tensor(0x1.197998p-40)]; + tensor var_1648_cast_fp16 = rsqrt(epsilon = var_1648_epsilon_0, x = var_73_cast_fp16_0)[name = tensor("op_1648_cast_fp16")]; + tensor var_1649_cast_fp16 = mul(x = const_3_to_fp16, y = var_1648_cast_fp16)[name = tensor("op_1649_cast_fp16")]; + tensor var_1650_cast_fp16 = mul(x = linear_106_cast_fp16, y = var_1649_cast_fp16)[name = tensor("op_1650_cast_fp16")]; + tensor input_281_to_fp16 = const()[name = tensor("input_281_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(19004992)))]; + tensor input_283_cast_fp16 = silu(x = input_281_to_fp16)[name = tensor("input_283_cast_fp16")]; + tensor linear_108_cast_fp16 = linear(bias = flow_net_time_embed_1_mlp_2_bias_to_fp16, weight = flow_net_time_embed_1_mlp_2_weight_to_fp16, x = input_283_cast_fp16)[name = tensor("linear_108_cast_fp16")]; + tensor reduce_mean_74_axes_0 = const()[name = tensor("reduce_mean_74_axes_0"), val = tensor([-1])]; + tensor reduce_mean_74_keep_dims_0 = const()[name = tensor("reduce_mean_74_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_74_cast_fp16 = reduce_mean(axes = reduce_mean_74_axes_0, keep_dims = reduce_mean_74_keep_dims_0, x = linear_108_cast_fp16)[name = tensor("reduce_mean_74_cast_fp16")]; + tensor sub_46_cast_fp16 = sub(x = linear_108_cast_fp16, y = reduce_mean_74_cast_fp16)[name = tensor("sub_46_cast_fp16")]; + tensor square_37_cast_fp16 = square(x = sub_46_cast_fp16)[name = tensor("square_37_cast_fp16")]; + tensor reduce_mean_75_axes_0 = const()[name = tensor("reduce_mean_75_axes_0"), val = tensor([-1])]; + tensor reduce_mean_75_keep_dims_0 = const()[name = tensor("reduce_mean_75_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_75_cast_fp16 = reduce_mean(axes = reduce_mean_75_axes_0, keep_dims = reduce_mean_75_keep_dims_0, x = square_37_cast_fp16)[name = tensor("reduce_mean_75_cast_fp16")]; + tensor real_div_9_to_fp16 = const()[name = tensor("real_div_9_to_fp16"), val = tensor(0x1.008p+0)]; + tensor mul_9_cast_fp16 = mul(x = reduce_mean_75_cast_fp16, y = real_div_9_to_fp16)[name = tensor("mul_9_cast_fp16")]; + tensor var_1682_to_fp16 = const()[name = tensor("op_1682_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_75_cast_fp16_0 = add(x = mul_9_cast_fp16, y = var_1682_to_fp16)[name = tensor("var_75_cast_fp16")]; + tensor var_1685_epsilon_0 = const()[name = tensor("op_1685_epsilon_0"), val = tensor(0x1.197998p-40)]; + tensor var_1685_cast_fp16 = rsqrt(epsilon = var_1685_epsilon_0, x = var_75_cast_fp16_0)[name = tensor("op_1685_cast_fp16")]; + tensor var_1686_cast_fp16 = mul(x = const_5_to_fp16, y = var_1685_cast_fp16)[name = tensor("op_1686_cast_fp16")]; + tensor var_1687_cast_fp16 = mul(x = linear_108_cast_fp16, y = var_1686_cast_fp16)[name = tensor("op_1687_cast_fp16")]; + tensor var_1699_cast_fp16 = add(x = var_1650_cast_fp16, y = var_1687_cast_fp16)[name = tensor("op_1699_cast_fp16")]; + tensor _inversed_t_combined_9_y_0_to_fp16 = const()[name = tensor("_inversed_t_combined_9_y_0_to_fp16"), val = tensor(0x1p-1)]; + tensor _inversed_t_combined_9_cast_fp16 = mul(x = var_1699_cast_fp16, y = _inversed_t_combined_9_y_0_to_fp16)[name = tensor("_inversed_t_combined_9_cast_fp16")]; + tensor input_285_cast_fp16 = add(x = _inversed_t_combined_9_cast_fp16, y = linear_5_cast_fp16)[name = tensor("input_285_cast_fp16")]; + tensor input_287_cast_fp16 = silu(x = input_285_cast_fp16)[name = tensor("input_287_cast_fp16")]; + tensor linear_110_cast_fp16 = linear(bias = flow_net_res_blocks_0_adaLN_modulation_1_bias_to_fp16, weight = flow_net_res_blocks_0_adaLN_modulation_1_weight_to_fp16, x = input_287_cast_fp16)[name = tensor("linear_110_cast_fp16")]; + tensor var_1714_split_sizes_0 = const()[name = tensor("op_1714_split_sizes_0"), val = tensor([512, 512, 512])]; + tensor var_1714_axis_0 = const()[name = tensor("op_1714_axis_0"), val = tensor(-1)]; + tensor var_1714_cast_fp16_0, tensor var_1714_cast_fp16_1, tensor var_1714_cast_fp16_2 = split(axis = var_1714_axis_0, split_sizes = var_1714_split_sizes_0, x = linear_110_cast_fp16)[name = tensor("op_1714_cast_fp16")]; + tensor mean_57_axes_0 = const()[name = tensor("mean_57_axes_0"), val = tensor([-1])]; + tensor mean_57_keep_dims_0 = const()[name = tensor("mean_57_keep_dims_0"), val = tensor(true)]; + tensor mean_57_cast_fp16 = reduce_mean(axes = mean_57_axes_0, keep_dims = mean_57_keep_dims_0, x = linear_104_cast_fp16)[name = tensor("mean_57_cast_fp16")]; + tensor sub_48_cast_fp16 = sub(x = linear_104_cast_fp16, y = mean_57_cast_fp16)[name = tensor("sub_48_cast_fp16")]; + tensor square_38_cast_fp16 = square(x = sub_48_cast_fp16)[name = tensor("square_38_cast_fp16")]; + tensor reduce_mean_77_axes_0 = const()[name = tensor("reduce_mean_77_axes_0"), val = tensor([-1])]; + tensor reduce_mean_77_keep_dims_0 = const()[name = tensor("reduce_mean_77_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_77_cast_fp16 = reduce_mean(axes = reduce_mean_77_axes_0, keep_dims = reduce_mean_77_keep_dims_0, x = square_38_cast_fp16)[name = tensor("reduce_mean_77_cast_fp16")]; + tensor var_1724_to_fp16 = const()[name = tensor("op_1724_to_fp16"), val = tensor(0x1.1p-20)]; + tensor var_1725_cast_fp16 = add(x = reduce_mean_77_cast_fp16, y = var_1724_to_fp16)[name = tensor("op_1725_cast_fp16")]; + tensor var_1726_cast_fp16 = sqrt(x = var_1725_cast_fp16)[name = tensor("op_1726_cast_fp16")]; + tensor x_183_cast_fp16 = real_div(x = sub_48_cast_fp16, y = var_1726_cast_fp16)[name = tensor("x_183_cast_fp16")]; + tensor var_1728_cast_fp16 = mul(x = x_183_cast_fp16, y = flow_net_res_blocks_0_in_ln_weight_to_fp16)[name = tensor("op_1728_cast_fp16")]; + tensor x_185_cast_fp16 = add(x = var_1728_cast_fp16, y = flow_net_res_blocks_0_in_ln_bias_to_fp16)[name = tensor("x_185_cast_fp16")]; + tensor var_1730_promoted_to_fp16 = const()[name = tensor("op_1730_promoted_to_fp16"), val = tensor(0x1p+0)]; + tensor var_1731_cast_fp16 = add(x = var_1714_cast_fp16_1, y = var_1730_promoted_to_fp16)[name = tensor("op_1731_cast_fp16")]; + tensor var_1732_cast_fp16 = mul(x = x_185_cast_fp16, y = var_1731_cast_fp16)[name = tensor("op_1732_cast_fp16")]; + tensor input_289_cast_fp16 = add(x = var_1732_cast_fp16, y = var_1714_cast_fp16_0)[name = tensor("input_289_cast_fp16")]; + tensor linear_111_cast_fp16 = linear(bias = flow_net_res_blocks_0_mlp_0_bias_to_fp16, weight = flow_net_res_blocks_0_mlp_0_weight_to_fp16, x = input_289_cast_fp16)[name = tensor("linear_111_cast_fp16")]; + tensor input_293_cast_fp16 = silu(x = linear_111_cast_fp16)[name = tensor("input_293_cast_fp16")]; + tensor linear_112_cast_fp16 = linear(bias = flow_net_res_blocks_0_mlp_2_bias_to_fp16, weight = flow_net_res_blocks_0_mlp_2_weight_to_fp16, x = input_293_cast_fp16)[name = tensor("linear_112_cast_fp16")]; + tensor var_1743_cast_fp16 = mul(x = var_1714_cast_fp16_2, y = linear_112_cast_fp16)[name = tensor("op_1743_cast_fp16")]; + tensor x_187_cast_fp16 = add(x = linear_104_cast_fp16, y = var_1743_cast_fp16)[name = tensor("x_187_cast_fp16")]; + tensor linear_113_cast_fp16 = linear(bias = flow_net_res_blocks_1_adaLN_modulation_1_bias_to_fp16, weight = flow_net_res_blocks_1_adaLN_modulation_1_weight_to_fp16, x = input_287_cast_fp16)[name = tensor("linear_113_cast_fp16")]; + tensor var_1753_split_sizes_0 = const()[name = tensor("op_1753_split_sizes_0"), val = tensor([512, 512, 512])]; + tensor var_1753_axis_0 = const()[name = tensor("op_1753_axis_0"), val = tensor(-1)]; + tensor var_1753_cast_fp16_0, tensor var_1753_cast_fp16_1, tensor var_1753_cast_fp16_2 = split(axis = var_1753_axis_0, split_sizes = var_1753_split_sizes_0, x = linear_113_cast_fp16)[name = tensor("op_1753_cast_fp16")]; + tensor mean_59_axes_0 = const()[name = tensor("mean_59_axes_0"), val = tensor([-1])]; + tensor mean_59_keep_dims_0 = const()[name = tensor("mean_59_keep_dims_0"), val = tensor(true)]; + tensor mean_59_cast_fp16 = reduce_mean(axes = mean_59_axes_0, keep_dims = mean_59_keep_dims_0, x = x_187_cast_fp16)[name = tensor("mean_59_cast_fp16")]; + tensor sub_49_cast_fp16 = sub(x = x_187_cast_fp16, y = mean_59_cast_fp16)[name = tensor("sub_49_cast_fp16")]; + tensor square_39_cast_fp16 = square(x = sub_49_cast_fp16)[name = tensor("square_39_cast_fp16")]; + tensor reduce_mean_79_axes_0 = const()[name = tensor("reduce_mean_79_axes_0"), val = tensor([-1])]; + tensor reduce_mean_79_keep_dims_0 = const()[name = tensor("reduce_mean_79_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_79_cast_fp16 = reduce_mean(axes = reduce_mean_79_axes_0, keep_dims = reduce_mean_79_keep_dims_0, x = square_39_cast_fp16)[name = tensor("reduce_mean_79_cast_fp16")]; + tensor var_1763_to_fp16 = const()[name = tensor("op_1763_to_fp16"), val = tensor(0x1.1p-20)]; + tensor var_1764_cast_fp16 = add(x = reduce_mean_79_cast_fp16, y = var_1763_to_fp16)[name = tensor("op_1764_cast_fp16")]; + tensor var_1765_cast_fp16 = sqrt(x = var_1764_cast_fp16)[name = tensor("op_1765_cast_fp16")]; + tensor x_189_cast_fp16 = real_div(x = sub_49_cast_fp16, y = var_1765_cast_fp16)[name = tensor("x_189_cast_fp16")]; + tensor var_1767_cast_fp16 = mul(x = x_189_cast_fp16, y = flow_net_res_blocks_1_in_ln_weight_to_fp16)[name = tensor("op_1767_cast_fp16")]; + tensor x_191_cast_fp16 = add(x = var_1767_cast_fp16, y = flow_net_res_blocks_1_in_ln_bias_to_fp16)[name = tensor("x_191_cast_fp16")]; + tensor var_1769_promoted_to_fp16 = const()[name = tensor("op_1769_promoted_to_fp16"), val = tensor(0x1p+0)]; + tensor var_1770_cast_fp16 = add(x = var_1753_cast_fp16_1, y = var_1769_promoted_to_fp16)[name = tensor("op_1770_cast_fp16")]; + tensor var_1771_cast_fp16 = mul(x = x_191_cast_fp16, y = var_1770_cast_fp16)[name = tensor("op_1771_cast_fp16")]; + tensor input_297_cast_fp16 = add(x = var_1771_cast_fp16, y = var_1753_cast_fp16_0)[name = tensor("input_297_cast_fp16")]; + tensor linear_114_cast_fp16 = linear(bias = flow_net_res_blocks_1_mlp_0_bias_to_fp16, weight = flow_net_res_blocks_1_mlp_0_weight_to_fp16, x = input_297_cast_fp16)[name = tensor("linear_114_cast_fp16")]; + tensor input_301_cast_fp16 = silu(x = linear_114_cast_fp16)[name = tensor("input_301_cast_fp16")]; + tensor linear_115_cast_fp16 = linear(bias = flow_net_res_blocks_1_mlp_2_bias_to_fp16, weight = flow_net_res_blocks_1_mlp_2_weight_to_fp16, x = input_301_cast_fp16)[name = tensor("linear_115_cast_fp16")]; + tensor var_1782_cast_fp16 = mul(x = var_1753_cast_fp16_2, y = linear_115_cast_fp16)[name = tensor("op_1782_cast_fp16")]; + tensor x_193_cast_fp16 = add(x = x_187_cast_fp16, y = var_1782_cast_fp16)[name = tensor("x_193_cast_fp16")]; + tensor linear_116_cast_fp16 = linear(bias = flow_net_res_blocks_2_adaLN_modulation_1_bias_to_fp16, weight = flow_net_res_blocks_2_adaLN_modulation_1_weight_to_fp16, x = input_287_cast_fp16)[name = tensor("linear_116_cast_fp16")]; + tensor var_1792_split_sizes_0 = const()[name = tensor("op_1792_split_sizes_0"), val = tensor([512, 512, 512])]; + tensor var_1792_axis_0 = const()[name = tensor("op_1792_axis_0"), val = tensor(-1)]; + tensor var_1792_cast_fp16_0, tensor var_1792_cast_fp16_1, tensor var_1792_cast_fp16_2 = split(axis = var_1792_axis_0, split_sizes = var_1792_split_sizes_0, x = linear_116_cast_fp16)[name = tensor("op_1792_cast_fp16")]; + tensor mean_61_axes_0 = const()[name = tensor("mean_61_axes_0"), val = tensor([-1])]; + tensor mean_61_keep_dims_0 = const()[name = tensor("mean_61_keep_dims_0"), val = tensor(true)]; + tensor mean_61_cast_fp16 = reduce_mean(axes = mean_61_axes_0, keep_dims = mean_61_keep_dims_0, x = x_193_cast_fp16)[name = tensor("mean_61_cast_fp16")]; + tensor sub_50_cast_fp16 = sub(x = x_193_cast_fp16, y = mean_61_cast_fp16)[name = tensor("sub_50_cast_fp16")]; + tensor square_40_cast_fp16 = square(x = sub_50_cast_fp16)[name = tensor("square_40_cast_fp16")]; + tensor reduce_mean_81_axes_0 = const()[name = tensor("reduce_mean_81_axes_0"), val = tensor([-1])]; + tensor reduce_mean_81_keep_dims_0 = const()[name = tensor("reduce_mean_81_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_81_cast_fp16 = reduce_mean(axes = reduce_mean_81_axes_0, keep_dims = reduce_mean_81_keep_dims_0, x = square_40_cast_fp16)[name = tensor("reduce_mean_81_cast_fp16")]; + tensor var_1802_to_fp16 = const()[name = tensor("op_1802_to_fp16"), val = tensor(0x1.1p-20)]; + tensor var_1803_cast_fp16 = add(x = reduce_mean_81_cast_fp16, y = var_1802_to_fp16)[name = tensor("op_1803_cast_fp16")]; + tensor var_1804_cast_fp16 = sqrt(x = var_1803_cast_fp16)[name = tensor("op_1804_cast_fp16")]; + tensor x_195_cast_fp16 = real_div(x = sub_50_cast_fp16, y = var_1804_cast_fp16)[name = tensor("x_195_cast_fp16")]; + tensor var_1806_cast_fp16 = mul(x = x_195_cast_fp16, y = flow_net_res_blocks_2_in_ln_weight_to_fp16)[name = tensor("op_1806_cast_fp16")]; + tensor x_197_cast_fp16 = add(x = var_1806_cast_fp16, y = flow_net_res_blocks_2_in_ln_bias_to_fp16)[name = tensor("x_197_cast_fp16")]; + tensor var_1808_promoted_to_fp16 = const()[name = tensor("op_1808_promoted_to_fp16"), val = tensor(0x1p+0)]; + tensor var_1809_cast_fp16 = add(x = var_1792_cast_fp16_1, y = var_1808_promoted_to_fp16)[name = tensor("op_1809_cast_fp16")]; + tensor var_1810_cast_fp16 = mul(x = x_197_cast_fp16, y = var_1809_cast_fp16)[name = tensor("op_1810_cast_fp16")]; + tensor input_305_cast_fp16 = add(x = var_1810_cast_fp16, y = var_1792_cast_fp16_0)[name = tensor("input_305_cast_fp16")]; + tensor linear_117_cast_fp16 = linear(bias = flow_net_res_blocks_2_mlp_0_bias_to_fp16, weight = flow_net_res_blocks_2_mlp_0_weight_to_fp16, x = input_305_cast_fp16)[name = tensor("linear_117_cast_fp16")]; + tensor input_309_cast_fp16 = silu(x = linear_117_cast_fp16)[name = tensor("input_309_cast_fp16")]; + tensor linear_118_cast_fp16 = linear(bias = flow_net_res_blocks_2_mlp_2_bias_to_fp16, weight = flow_net_res_blocks_2_mlp_2_weight_to_fp16, x = input_309_cast_fp16)[name = tensor("linear_118_cast_fp16")]; + tensor var_1821_cast_fp16 = mul(x = var_1792_cast_fp16_2, y = linear_118_cast_fp16)[name = tensor("op_1821_cast_fp16")]; + tensor x_199_cast_fp16 = add(x = x_193_cast_fp16, y = var_1821_cast_fp16)[name = tensor("x_199_cast_fp16")]; + tensor linear_119_cast_fp16 = linear(bias = flow_net_res_blocks_3_adaLN_modulation_1_bias_to_fp16, weight = flow_net_res_blocks_3_adaLN_modulation_1_weight_to_fp16, x = input_287_cast_fp16)[name = tensor("linear_119_cast_fp16")]; + tensor var_1831_split_sizes_0 = const()[name = tensor("op_1831_split_sizes_0"), val = tensor([512, 512, 512])]; + tensor var_1831_axis_0 = const()[name = tensor("op_1831_axis_0"), val = tensor(-1)]; + tensor var_1831_cast_fp16_0, tensor var_1831_cast_fp16_1, tensor var_1831_cast_fp16_2 = split(axis = var_1831_axis_0, split_sizes = var_1831_split_sizes_0, x = linear_119_cast_fp16)[name = tensor("op_1831_cast_fp16")]; + tensor mean_63_axes_0 = const()[name = tensor("mean_63_axes_0"), val = tensor([-1])]; + tensor mean_63_keep_dims_0 = const()[name = tensor("mean_63_keep_dims_0"), val = tensor(true)]; + tensor mean_63_cast_fp16 = reduce_mean(axes = mean_63_axes_0, keep_dims = mean_63_keep_dims_0, x = x_199_cast_fp16)[name = tensor("mean_63_cast_fp16")]; + tensor sub_51_cast_fp16 = sub(x = x_199_cast_fp16, y = mean_63_cast_fp16)[name = tensor("sub_51_cast_fp16")]; + tensor square_41_cast_fp16 = square(x = sub_51_cast_fp16)[name = tensor("square_41_cast_fp16")]; + tensor reduce_mean_83_axes_0 = const()[name = tensor("reduce_mean_83_axes_0"), val = tensor([-1])]; + tensor reduce_mean_83_keep_dims_0 = const()[name = tensor("reduce_mean_83_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_83_cast_fp16 = reduce_mean(axes = reduce_mean_83_axes_0, keep_dims = reduce_mean_83_keep_dims_0, x = square_41_cast_fp16)[name = tensor("reduce_mean_83_cast_fp16")]; + tensor var_1841_to_fp16 = const()[name = tensor("op_1841_to_fp16"), val = tensor(0x1.1p-20)]; + tensor var_1842_cast_fp16 = add(x = reduce_mean_83_cast_fp16, y = var_1841_to_fp16)[name = tensor("op_1842_cast_fp16")]; + tensor var_1843_cast_fp16 = sqrt(x = var_1842_cast_fp16)[name = tensor("op_1843_cast_fp16")]; + tensor x_201_cast_fp16 = real_div(x = sub_51_cast_fp16, y = var_1843_cast_fp16)[name = tensor("x_201_cast_fp16")]; + tensor var_1845_cast_fp16 = mul(x = x_201_cast_fp16, y = flow_net_res_blocks_3_in_ln_weight_to_fp16)[name = tensor("op_1845_cast_fp16")]; + tensor x_203_cast_fp16 = add(x = var_1845_cast_fp16, y = flow_net_res_blocks_3_in_ln_bias_to_fp16)[name = tensor("x_203_cast_fp16")]; + tensor var_1847_promoted_to_fp16 = const()[name = tensor("op_1847_promoted_to_fp16"), val = tensor(0x1p+0)]; + tensor var_1848_cast_fp16 = add(x = var_1831_cast_fp16_1, y = var_1847_promoted_to_fp16)[name = tensor("op_1848_cast_fp16")]; + tensor var_1849_cast_fp16 = mul(x = x_203_cast_fp16, y = var_1848_cast_fp16)[name = tensor("op_1849_cast_fp16")]; + tensor input_313_cast_fp16 = add(x = var_1849_cast_fp16, y = var_1831_cast_fp16_0)[name = tensor("input_313_cast_fp16")]; + tensor linear_120_cast_fp16 = linear(bias = flow_net_res_blocks_3_mlp_0_bias_to_fp16, weight = flow_net_res_blocks_3_mlp_0_weight_to_fp16, x = input_313_cast_fp16)[name = tensor("linear_120_cast_fp16")]; + tensor input_317_cast_fp16 = silu(x = linear_120_cast_fp16)[name = tensor("input_317_cast_fp16")]; + tensor linear_121_cast_fp16 = linear(bias = flow_net_res_blocks_3_mlp_2_bias_to_fp16, weight = flow_net_res_blocks_3_mlp_2_weight_to_fp16, x = input_317_cast_fp16)[name = tensor("linear_121_cast_fp16")]; + tensor var_1860_cast_fp16 = mul(x = var_1831_cast_fp16_2, y = linear_121_cast_fp16)[name = tensor("op_1860_cast_fp16")]; + tensor x_205_cast_fp16 = add(x = x_199_cast_fp16, y = var_1860_cast_fp16)[name = tensor("x_205_cast_fp16")]; + tensor linear_122_cast_fp16 = linear(bias = flow_net_res_blocks_4_adaLN_modulation_1_bias_to_fp16, weight = flow_net_res_blocks_4_adaLN_modulation_1_weight_to_fp16, x = input_287_cast_fp16)[name = tensor("linear_122_cast_fp16")]; + tensor var_1870_split_sizes_0 = const()[name = tensor("op_1870_split_sizes_0"), val = tensor([512, 512, 512])]; + tensor var_1870_axis_0 = const()[name = tensor("op_1870_axis_0"), val = tensor(-1)]; + tensor var_1870_cast_fp16_0, tensor var_1870_cast_fp16_1, tensor var_1870_cast_fp16_2 = split(axis = var_1870_axis_0, split_sizes = var_1870_split_sizes_0, x = linear_122_cast_fp16)[name = tensor("op_1870_cast_fp16")]; + tensor mean_65_axes_0 = const()[name = tensor("mean_65_axes_0"), val = tensor([-1])]; + tensor mean_65_keep_dims_0 = const()[name = tensor("mean_65_keep_dims_0"), val = tensor(true)]; + tensor mean_65_cast_fp16 = reduce_mean(axes = mean_65_axes_0, keep_dims = mean_65_keep_dims_0, x = x_205_cast_fp16)[name = tensor("mean_65_cast_fp16")]; + tensor sub_52_cast_fp16 = sub(x = x_205_cast_fp16, y = mean_65_cast_fp16)[name = tensor("sub_52_cast_fp16")]; + tensor square_42_cast_fp16 = square(x = sub_52_cast_fp16)[name = tensor("square_42_cast_fp16")]; + tensor reduce_mean_85_axes_0 = const()[name = tensor("reduce_mean_85_axes_0"), val = tensor([-1])]; + tensor reduce_mean_85_keep_dims_0 = const()[name = tensor("reduce_mean_85_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_85_cast_fp16 = reduce_mean(axes = reduce_mean_85_axes_0, keep_dims = reduce_mean_85_keep_dims_0, x = square_42_cast_fp16)[name = tensor("reduce_mean_85_cast_fp16")]; + tensor var_1880_to_fp16 = const()[name = tensor("op_1880_to_fp16"), val = tensor(0x1.1p-20)]; + tensor var_1881_cast_fp16 = add(x = reduce_mean_85_cast_fp16, y = var_1880_to_fp16)[name = tensor("op_1881_cast_fp16")]; + tensor var_1882_cast_fp16 = sqrt(x = var_1881_cast_fp16)[name = tensor("op_1882_cast_fp16")]; + tensor x_207_cast_fp16 = real_div(x = sub_52_cast_fp16, y = var_1882_cast_fp16)[name = tensor("x_207_cast_fp16")]; + tensor var_1884_cast_fp16 = mul(x = x_207_cast_fp16, y = flow_net_res_blocks_4_in_ln_weight_to_fp16)[name = tensor("op_1884_cast_fp16")]; + tensor x_209_cast_fp16 = add(x = var_1884_cast_fp16, y = flow_net_res_blocks_4_in_ln_bias_to_fp16)[name = tensor("x_209_cast_fp16")]; + tensor var_1886_promoted_to_fp16 = const()[name = tensor("op_1886_promoted_to_fp16"), val = tensor(0x1p+0)]; + tensor var_1887_cast_fp16 = add(x = var_1870_cast_fp16_1, y = var_1886_promoted_to_fp16)[name = tensor("op_1887_cast_fp16")]; + tensor var_1888_cast_fp16 = mul(x = x_209_cast_fp16, y = var_1887_cast_fp16)[name = tensor("op_1888_cast_fp16")]; + tensor input_321_cast_fp16 = add(x = var_1888_cast_fp16, y = var_1870_cast_fp16_0)[name = tensor("input_321_cast_fp16")]; + tensor linear_123_cast_fp16 = linear(bias = flow_net_res_blocks_4_mlp_0_bias_to_fp16, weight = flow_net_res_blocks_4_mlp_0_weight_to_fp16, x = input_321_cast_fp16)[name = tensor("linear_123_cast_fp16")]; + tensor input_325_cast_fp16 = silu(x = linear_123_cast_fp16)[name = tensor("input_325_cast_fp16")]; + tensor linear_124_cast_fp16 = linear(bias = flow_net_res_blocks_4_mlp_2_bias_to_fp16, weight = flow_net_res_blocks_4_mlp_2_weight_to_fp16, x = input_325_cast_fp16)[name = tensor("linear_124_cast_fp16")]; + tensor var_1899_cast_fp16 = mul(x = var_1870_cast_fp16_2, y = linear_124_cast_fp16)[name = tensor("op_1899_cast_fp16")]; + tensor x_211_cast_fp16 = add(x = x_205_cast_fp16, y = var_1899_cast_fp16)[name = tensor("x_211_cast_fp16")]; + tensor linear_125_cast_fp16 = linear(bias = flow_net_res_blocks_5_adaLN_modulation_1_bias_to_fp16, weight = flow_net_res_blocks_5_adaLN_modulation_1_weight_to_fp16, x = input_287_cast_fp16)[name = tensor("linear_125_cast_fp16")]; + tensor var_1909_split_sizes_0 = const()[name = tensor("op_1909_split_sizes_0"), val = tensor([512, 512, 512])]; + tensor var_1909_axis_0 = const()[name = tensor("op_1909_axis_0"), val = tensor(-1)]; + tensor var_1909_cast_fp16_0, tensor var_1909_cast_fp16_1, tensor var_1909_cast_fp16_2 = split(axis = var_1909_axis_0, split_sizes = var_1909_split_sizes_0, x = linear_125_cast_fp16)[name = tensor("op_1909_cast_fp16")]; + tensor mean_67_axes_0 = const()[name = tensor("mean_67_axes_0"), val = tensor([-1])]; + tensor mean_67_keep_dims_0 = const()[name = tensor("mean_67_keep_dims_0"), val = tensor(true)]; + tensor mean_67_cast_fp16 = reduce_mean(axes = mean_67_axes_0, keep_dims = mean_67_keep_dims_0, x = x_211_cast_fp16)[name = tensor("mean_67_cast_fp16")]; + tensor sub_53_cast_fp16 = sub(x = x_211_cast_fp16, y = mean_67_cast_fp16)[name = tensor("sub_53_cast_fp16")]; + tensor square_43_cast_fp16 = square(x = sub_53_cast_fp16)[name = tensor("square_43_cast_fp16")]; + tensor reduce_mean_87_axes_0 = const()[name = tensor("reduce_mean_87_axes_0"), val = tensor([-1])]; + tensor reduce_mean_87_keep_dims_0 = const()[name = tensor("reduce_mean_87_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_87_cast_fp16 = reduce_mean(axes = reduce_mean_87_axes_0, keep_dims = reduce_mean_87_keep_dims_0, x = square_43_cast_fp16)[name = tensor("reduce_mean_87_cast_fp16")]; + tensor var_1919_to_fp16 = const()[name = tensor("op_1919_to_fp16"), val = tensor(0x1.1p-20)]; + tensor var_1920_cast_fp16 = add(x = reduce_mean_87_cast_fp16, y = var_1919_to_fp16)[name = tensor("op_1920_cast_fp16")]; + tensor var_1921_cast_fp16 = sqrt(x = var_1920_cast_fp16)[name = tensor("op_1921_cast_fp16")]; + tensor x_213_cast_fp16 = real_div(x = sub_53_cast_fp16, y = var_1921_cast_fp16)[name = tensor("x_213_cast_fp16")]; + tensor var_1923_cast_fp16 = mul(x = x_213_cast_fp16, y = flow_net_res_blocks_5_in_ln_weight_to_fp16)[name = tensor("op_1923_cast_fp16")]; + tensor x_215_cast_fp16 = add(x = var_1923_cast_fp16, y = flow_net_res_blocks_5_in_ln_bias_to_fp16)[name = tensor("x_215_cast_fp16")]; + tensor var_1925_promoted_to_fp16 = const()[name = tensor("op_1925_promoted_to_fp16"), val = tensor(0x1p+0)]; + tensor var_1926_cast_fp16 = add(x = var_1909_cast_fp16_1, y = var_1925_promoted_to_fp16)[name = tensor("op_1926_cast_fp16")]; + tensor var_1927_cast_fp16 = mul(x = x_215_cast_fp16, y = var_1926_cast_fp16)[name = tensor("op_1927_cast_fp16")]; + tensor input_329_cast_fp16 = add(x = var_1927_cast_fp16, y = var_1909_cast_fp16_0)[name = tensor("input_329_cast_fp16")]; + tensor linear_126_cast_fp16 = linear(bias = flow_net_res_blocks_5_mlp_0_bias_to_fp16, weight = flow_net_res_blocks_5_mlp_0_weight_to_fp16, x = input_329_cast_fp16)[name = tensor("linear_126_cast_fp16")]; + tensor input_333_cast_fp16 = silu(x = linear_126_cast_fp16)[name = tensor("input_333_cast_fp16")]; + tensor linear_127_cast_fp16 = linear(bias = flow_net_res_blocks_5_mlp_2_bias_to_fp16, weight = flow_net_res_blocks_5_mlp_2_weight_to_fp16, x = input_333_cast_fp16)[name = tensor("linear_127_cast_fp16")]; + tensor var_1938_cast_fp16 = mul(x = var_1909_cast_fp16_2, y = linear_127_cast_fp16)[name = tensor("op_1938_cast_fp16")]; + tensor x_217_cast_fp16 = add(x = x_211_cast_fp16, y = var_1938_cast_fp16)[name = tensor("x_217_cast_fp16")]; + tensor linear_128_cast_fp16 = linear(bias = flow_net_final_layer_adaLN_modulation_1_bias_to_fp16, weight = flow_net_final_layer_adaLN_modulation_1_weight_to_fp16, x = input_287_cast_fp16)[name = tensor("linear_128_cast_fp16")]; + tensor var_1947_split_sizes_0 = const()[name = tensor("op_1947_split_sizes_0"), val = tensor([512, 512])]; + tensor var_1947_axis_0 = const()[name = tensor("op_1947_axis_0"), val = tensor(-1)]; + tensor var_1947_cast_fp16_0, tensor var_1947_cast_fp16_1 = split(axis = var_1947_axis_0, split_sizes = var_1947_split_sizes_0, x = linear_128_cast_fp16)[name = tensor("op_1947_cast_fp16")]; + tensor mean_69_axes_0 = const()[name = tensor("mean_69_axes_0"), val = tensor([-1])]; + tensor mean_69_keep_dims_0 = const()[name = tensor("mean_69_keep_dims_0"), val = tensor(true)]; + tensor mean_69_cast_fp16 = reduce_mean(axes = mean_69_axes_0, keep_dims = mean_69_keep_dims_0, x = x_217_cast_fp16)[name = tensor("mean_69_cast_fp16")]; + tensor sub_54_cast_fp16 = sub(x = x_217_cast_fp16, y = mean_69_cast_fp16)[name = tensor("sub_54_cast_fp16")]; + tensor square_44_cast_fp16 = square(x = sub_54_cast_fp16)[name = tensor("square_44_cast_fp16")]; + tensor reduce_mean_89_axes_0 = const()[name = tensor("reduce_mean_89_axes_0"), val = tensor([-1])]; + tensor reduce_mean_89_keep_dims_0 = const()[name = tensor("reduce_mean_89_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_89_cast_fp16 = reduce_mean(axes = reduce_mean_89_axes_0, keep_dims = reduce_mean_89_keep_dims_0, x = square_44_cast_fp16)[name = tensor("reduce_mean_89_cast_fp16")]; + tensor var_1954_to_fp16 = const()[name = tensor("op_1954_to_fp16"), val = tensor(0x1.1p-20)]; + tensor var_1955_cast_fp16 = add(x = reduce_mean_89_cast_fp16, y = var_1954_to_fp16)[name = tensor("op_1955_cast_fp16")]; + tensor var_1956_cast_fp16 = sqrt(x = var_1955_cast_fp16)[name = tensor("op_1956_cast_fp16")]; + tensor x_219_cast_fp16 = real_div(x = sub_54_cast_fp16, y = var_1956_cast_fp16)[name = tensor("x_219_cast_fp16")]; + tensor var_1958_promoted_to_fp16 = const()[name = tensor("op_1958_promoted_to_fp16"), val = tensor(0x1p+0)]; + tensor var_1959_cast_fp16 = add(x = var_1947_cast_fp16_1, y = var_1958_promoted_to_fp16)[name = tensor("op_1959_cast_fp16")]; + tensor var_1960_cast_fp16 = mul(x = x_219_cast_fp16, y = var_1959_cast_fp16)[name = tensor("op_1960_cast_fp16")]; + tensor input_337_cast_fp16 = add(x = var_1960_cast_fp16, y = var_1947_cast_fp16_0)[name = tensor("input_337_cast_fp16")]; + tensor linear_129_cast_fp16 = linear(bias = flow_net_final_layer_linear_bias_to_fp16, weight = flow_net_final_layer_linear_weight_to_fp16, x = input_337_cast_fp16)[name = tensor("linear_129_cast_fp16")]; + tensor var_1971_to_fp16 = const()[name = tensor("op_1971_to_fp16"), val = tensor(0x1p-3)]; + tensor var_1972_cast_fp16 = mul(x = linear_129_cast_fp16, y = var_1971_to_fp16)[name = tensor("op_1972_cast_fp16")]; + tensor input_339_cast_fp16 = add(x = input_271_cast_fp16, y = var_1972_cast_fp16)[name = tensor("input_339_cast_fp16")]; + tensor linear_130_cast_fp16 = linear(bias = flow_net_input_proj_bias_to_fp16, weight = flow_net_input_proj_weight_to_fp16, x = input_339_cast_fp16)[name = tensor("linear_130_cast_fp16")]; + tensor input_343_to_fp16 = const()[name = tensor("input_343_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(19006080)))]; + tensor input_345_cast_fp16 = silu(x = input_343_to_fp16)[name = tensor("input_345_cast_fp16")]; + tensor linear_132_cast_fp16 = linear(bias = flow_net_time_embed_0_mlp_2_bias_to_fp16, weight = flow_net_time_embed_0_mlp_2_weight_to_fp16, x = input_345_cast_fp16)[name = tensor("linear_132_cast_fp16")]; + tensor reduce_mean_90_axes_0 = const()[name = tensor("reduce_mean_90_axes_0"), val = tensor([-1])]; + tensor reduce_mean_90_keep_dims_0 = const()[name = tensor("reduce_mean_90_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_90_cast_fp16 = reduce_mean(axes = reduce_mean_90_axes_0, keep_dims = reduce_mean_90_keep_dims_0, x = linear_132_cast_fp16)[name = tensor("reduce_mean_90_cast_fp16")]; + tensor sub_55_cast_fp16 = sub(x = linear_132_cast_fp16, y = reduce_mean_90_cast_fp16)[name = tensor("sub_55_cast_fp16")]; + tensor square_45_cast_fp16 = square(x = sub_55_cast_fp16)[name = tensor("square_45_cast_fp16")]; + tensor reduce_mean_91_axes_0 = const()[name = tensor("reduce_mean_91_axes_0"), val = tensor([-1])]; + tensor reduce_mean_91_keep_dims_0 = const()[name = tensor("reduce_mean_91_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_91_cast_fp16 = reduce_mean(axes = reduce_mean_91_axes_0, keep_dims = reduce_mean_91_keep_dims_0, x = square_45_cast_fp16)[name = tensor("reduce_mean_91_cast_fp16")]; + tensor real_div_10_to_fp16 = const()[name = tensor("real_div_10_to_fp16"), val = tensor(0x1.008p+0)]; + tensor mul_10_cast_fp16 = mul(x = reduce_mean_91_cast_fp16, y = real_div_10_to_fp16)[name = tensor("mul_10_cast_fp16")]; + tensor var_2038_to_fp16 = const()[name = tensor("op_2038_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_91_cast_fp16 = add(x = mul_10_cast_fp16, y = var_2038_to_fp16)[name = tensor("var_91_cast_fp16")]; + tensor var_2041_epsilon_0 = const()[name = tensor("op_2041_epsilon_0"), val = tensor(0x1.197998p-40)]; + tensor var_2041_cast_fp16 = rsqrt(epsilon = var_2041_epsilon_0, x = var_91_cast_fp16)[name = tensor("op_2041_cast_fp16")]; + tensor var_2042_cast_fp16 = mul(x = const_3_to_fp16, y = var_2041_cast_fp16)[name = tensor("op_2042_cast_fp16")]; + tensor var_2043_cast_fp16 = mul(x = linear_132_cast_fp16, y = var_2042_cast_fp16)[name = tensor("op_2043_cast_fp16")]; + tensor input_349_to_fp16 = const()[name = tensor("input_349_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(19007168)))]; + tensor input_351_cast_fp16 = silu(x = input_349_to_fp16)[name = tensor("input_351_cast_fp16")]; + tensor linear_134_cast_fp16 = linear(bias = flow_net_time_embed_1_mlp_2_bias_to_fp16, weight = flow_net_time_embed_1_mlp_2_weight_to_fp16, x = input_351_cast_fp16)[name = tensor("linear_134_cast_fp16")]; + tensor reduce_mean_92_axes_0 = const()[name = tensor("reduce_mean_92_axes_0"), val = tensor([-1])]; + tensor reduce_mean_92_keep_dims_0 = const()[name = tensor("reduce_mean_92_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_92_cast_fp16 = reduce_mean(axes = reduce_mean_92_axes_0, keep_dims = reduce_mean_92_keep_dims_0, x = linear_134_cast_fp16)[name = tensor("reduce_mean_92_cast_fp16")]; + tensor sub_57_cast_fp16 = sub(x = linear_134_cast_fp16, y = reduce_mean_92_cast_fp16)[name = tensor("sub_57_cast_fp16")]; + tensor square_46_cast_fp16 = square(x = sub_57_cast_fp16)[name = tensor("square_46_cast_fp16")]; + tensor reduce_mean_93_axes_0 = const()[name = tensor("reduce_mean_93_axes_0"), val = tensor([-1])]; + tensor reduce_mean_93_keep_dims_0 = const()[name = tensor("reduce_mean_93_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_93_cast_fp16 = reduce_mean(axes = reduce_mean_93_axes_0, keep_dims = reduce_mean_93_keep_dims_0, x = square_46_cast_fp16)[name = tensor("reduce_mean_93_cast_fp16")]; + tensor real_div_11_to_fp16 = const()[name = tensor("real_div_11_to_fp16"), val = tensor(0x1.008p+0)]; + tensor mul_11_cast_fp16 = mul(x = reduce_mean_93_cast_fp16, y = real_div_11_to_fp16)[name = tensor("mul_11_cast_fp16")]; + tensor var_2075_to_fp16 = const()[name = tensor("op_2075_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_93_cast_fp16 = add(x = mul_11_cast_fp16, y = var_2075_to_fp16)[name = tensor("var_93_cast_fp16")]; + tensor var_2078_epsilon_0 = const()[name = tensor("op_2078_epsilon_0"), val = tensor(0x1.197998p-40)]; + tensor var_2078_cast_fp16 = rsqrt(epsilon = var_2078_epsilon_0, x = var_93_cast_fp16)[name = tensor("op_2078_cast_fp16")]; + tensor var_2079_cast_fp16 = mul(x = const_5_to_fp16, y = var_2078_cast_fp16)[name = tensor("op_2079_cast_fp16")]; + tensor var_2080_cast_fp16 = mul(x = linear_134_cast_fp16, y = var_2079_cast_fp16)[name = tensor("op_2080_cast_fp16")]; + tensor var_2092_cast_fp16 = add(x = var_2043_cast_fp16, y = var_2080_cast_fp16)[name = tensor("op_2092_cast_fp16")]; + tensor _inversed_t_combined_11_y_0_to_fp16 = const()[name = tensor("_inversed_t_combined_11_y_0_to_fp16"), val = tensor(0x1p-1)]; + tensor _inversed_t_combined_11_cast_fp16 = mul(x = var_2092_cast_fp16, y = _inversed_t_combined_11_y_0_to_fp16)[name = tensor("_inversed_t_combined_11_cast_fp16")]; + tensor input_353_cast_fp16 = add(x = _inversed_t_combined_11_cast_fp16, y = linear_5_cast_fp16)[name = tensor("input_353_cast_fp16")]; + tensor input_355_cast_fp16 = silu(x = input_353_cast_fp16)[name = tensor("input_355_cast_fp16")]; + tensor linear_136_cast_fp16 = linear(bias = flow_net_res_blocks_0_adaLN_modulation_1_bias_to_fp16, weight = flow_net_res_blocks_0_adaLN_modulation_1_weight_to_fp16, x = input_355_cast_fp16)[name = tensor("linear_136_cast_fp16")]; + tensor var_2107_split_sizes_0 = const()[name = tensor("op_2107_split_sizes_0"), val = tensor([512, 512, 512])]; + tensor var_2107_axis_0 = const()[name = tensor("op_2107_axis_0"), val = tensor(-1)]; + tensor var_2107_cast_fp16_0, tensor var_2107_cast_fp16_1, tensor var_2107_cast_fp16_2 = split(axis = var_2107_axis_0, split_sizes = var_2107_split_sizes_0, x = linear_136_cast_fp16)[name = tensor("op_2107_cast_fp16")]; + tensor mean_71_axes_0 = const()[name = tensor("mean_71_axes_0"), val = tensor([-1])]; + tensor mean_71_keep_dims_0 = const()[name = tensor("mean_71_keep_dims_0"), val = tensor(true)]; + tensor mean_71_cast_fp16 = reduce_mean(axes = mean_71_axes_0, keep_dims = mean_71_keep_dims_0, x = linear_130_cast_fp16)[name = tensor("mean_71_cast_fp16")]; + tensor sub_59_cast_fp16 = sub(x = linear_130_cast_fp16, y = mean_71_cast_fp16)[name = tensor("sub_59_cast_fp16")]; + tensor square_47_cast_fp16 = square(x = sub_59_cast_fp16)[name = tensor("square_47_cast_fp16")]; + tensor reduce_mean_95_axes_0 = const()[name = tensor("reduce_mean_95_axes_0"), val = tensor([-1])]; + tensor reduce_mean_95_keep_dims_0 = const()[name = tensor("reduce_mean_95_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_95_cast_fp16 = reduce_mean(axes = reduce_mean_95_axes_0, keep_dims = reduce_mean_95_keep_dims_0, x = square_47_cast_fp16)[name = tensor("reduce_mean_95_cast_fp16")]; + tensor var_2117_to_fp16 = const()[name = tensor("op_2117_to_fp16"), val = tensor(0x1.1p-20)]; + tensor var_2118_cast_fp16 = add(x = reduce_mean_95_cast_fp16, y = var_2117_to_fp16)[name = tensor("op_2118_cast_fp16")]; + tensor var_2119_cast_fp16 = sqrt(x = var_2118_cast_fp16)[name = tensor("op_2119_cast_fp16")]; + tensor x_227_cast_fp16 = real_div(x = sub_59_cast_fp16, y = var_2119_cast_fp16)[name = tensor("x_227_cast_fp16")]; + tensor var_2121_cast_fp16 = mul(x = x_227_cast_fp16, y = flow_net_res_blocks_0_in_ln_weight_to_fp16)[name = tensor("op_2121_cast_fp16")]; + tensor x_229_cast_fp16 = add(x = var_2121_cast_fp16, y = flow_net_res_blocks_0_in_ln_bias_to_fp16)[name = tensor("x_229_cast_fp16")]; + tensor var_2123_promoted_to_fp16 = const()[name = tensor("op_2123_promoted_to_fp16"), val = tensor(0x1p+0)]; + tensor var_2124_cast_fp16 = add(x = var_2107_cast_fp16_1, y = var_2123_promoted_to_fp16)[name = tensor("op_2124_cast_fp16")]; + tensor var_2125_cast_fp16 = mul(x = x_229_cast_fp16, y = var_2124_cast_fp16)[name = tensor("op_2125_cast_fp16")]; + tensor input_357_cast_fp16 = add(x = var_2125_cast_fp16, y = var_2107_cast_fp16_0)[name = tensor("input_357_cast_fp16")]; + tensor linear_137_cast_fp16 = linear(bias = flow_net_res_blocks_0_mlp_0_bias_to_fp16, weight = flow_net_res_blocks_0_mlp_0_weight_to_fp16, x = input_357_cast_fp16)[name = tensor("linear_137_cast_fp16")]; + tensor input_361_cast_fp16 = silu(x = linear_137_cast_fp16)[name = tensor("input_361_cast_fp16")]; + tensor linear_138_cast_fp16 = linear(bias = flow_net_res_blocks_0_mlp_2_bias_to_fp16, weight = flow_net_res_blocks_0_mlp_2_weight_to_fp16, x = input_361_cast_fp16)[name = tensor("linear_138_cast_fp16")]; + tensor var_2136_cast_fp16 = mul(x = var_2107_cast_fp16_2, y = linear_138_cast_fp16)[name = tensor("op_2136_cast_fp16")]; + tensor x_231_cast_fp16 = add(x = linear_130_cast_fp16, y = var_2136_cast_fp16)[name = tensor("x_231_cast_fp16")]; + tensor linear_139_cast_fp16 = linear(bias = flow_net_res_blocks_1_adaLN_modulation_1_bias_to_fp16, weight = flow_net_res_blocks_1_adaLN_modulation_1_weight_to_fp16, x = input_355_cast_fp16)[name = tensor("linear_139_cast_fp16")]; + tensor var_2146_split_sizes_0 = const()[name = tensor("op_2146_split_sizes_0"), val = tensor([512, 512, 512])]; + tensor var_2146_axis_0 = const()[name = tensor("op_2146_axis_0"), val = tensor(-1)]; + tensor var_2146_cast_fp16_0, tensor var_2146_cast_fp16_1, tensor var_2146_cast_fp16_2 = split(axis = var_2146_axis_0, split_sizes = var_2146_split_sizes_0, x = linear_139_cast_fp16)[name = tensor("op_2146_cast_fp16")]; + tensor mean_73_axes_0 = const()[name = tensor("mean_73_axes_0"), val = tensor([-1])]; + tensor mean_73_keep_dims_0 = const()[name = tensor("mean_73_keep_dims_0"), val = tensor(true)]; + tensor mean_73_cast_fp16 = reduce_mean(axes = mean_73_axes_0, keep_dims = mean_73_keep_dims_0, x = x_231_cast_fp16)[name = tensor("mean_73_cast_fp16")]; + tensor sub_60_cast_fp16 = sub(x = x_231_cast_fp16, y = mean_73_cast_fp16)[name = tensor("sub_60_cast_fp16")]; + tensor square_48_cast_fp16 = square(x = sub_60_cast_fp16)[name = tensor("square_48_cast_fp16")]; + tensor reduce_mean_97_axes_0 = const()[name = tensor("reduce_mean_97_axes_0"), val = tensor([-1])]; + tensor reduce_mean_97_keep_dims_0 = const()[name = tensor("reduce_mean_97_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_97_cast_fp16 = reduce_mean(axes = reduce_mean_97_axes_0, keep_dims = reduce_mean_97_keep_dims_0, x = square_48_cast_fp16)[name = tensor("reduce_mean_97_cast_fp16")]; + tensor var_2156_to_fp16 = const()[name = tensor("op_2156_to_fp16"), val = tensor(0x1.1p-20)]; + tensor var_2157_cast_fp16 = add(x = reduce_mean_97_cast_fp16, y = var_2156_to_fp16)[name = tensor("op_2157_cast_fp16")]; + tensor var_2158_cast_fp16 = sqrt(x = var_2157_cast_fp16)[name = tensor("op_2158_cast_fp16")]; + tensor x_233_cast_fp16 = real_div(x = sub_60_cast_fp16, y = var_2158_cast_fp16)[name = tensor("x_233_cast_fp16")]; + tensor var_2160_cast_fp16 = mul(x = x_233_cast_fp16, y = flow_net_res_blocks_1_in_ln_weight_to_fp16)[name = tensor("op_2160_cast_fp16")]; + tensor x_235_cast_fp16 = add(x = var_2160_cast_fp16, y = flow_net_res_blocks_1_in_ln_bias_to_fp16)[name = tensor("x_235_cast_fp16")]; + tensor var_2162_promoted_to_fp16 = const()[name = tensor("op_2162_promoted_to_fp16"), val = tensor(0x1p+0)]; + tensor var_2163_cast_fp16 = add(x = var_2146_cast_fp16_1, y = var_2162_promoted_to_fp16)[name = tensor("op_2163_cast_fp16")]; + tensor var_2164_cast_fp16 = mul(x = x_235_cast_fp16, y = var_2163_cast_fp16)[name = tensor("op_2164_cast_fp16")]; + tensor input_365_cast_fp16 = add(x = var_2164_cast_fp16, y = var_2146_cast_fp16_0)[name = tensor("input_365_cast_fp16")]; + tensor linear_140_cast_fp16 = linear(bias = flow_net_res_blocks_1_mlp_0_bias_to_fp16, weight = flow_net_res_blocks_1_mlp_0_weight_to_fp16, x = input_365_cast_fp16)[name = tensor("linear_140_cast_fp16")]; + tensor input_369_cast_fp16 = silu(x = linear_140_cast_fp16)[name = tensor("input_369_cast_fp16")]; + tensor linear_141_cast_fp16 = linear(bias = flow_net_res_blocks_1_mlp_2_bias_to_fp16, weight = flow_net_res_blocks_1_mlp_2_weight_to_fp16, x = input_369_cast_fp16)[name = tensor("linear_141_cast_fp16")]; + tensor var_2175_cast_fp16 = mul(x = var_2146_cast_fp16_2, y = linear_141_cast_fp16)[name = tensor("op_2175_cast_fp16")]; + tensor x_237_cast_fp16 = add(x = x_231_cast_fp16, y = var_2175_cast_fp16)[name = tensor("x_237_cast_fp16")]; + tensor linear_142_cast_fp16 = linear(bias = flow_net_res_blocks_2_adaLN_modulation_1_bias_to_fp16, weight = flow_net_res_blocks_2_adaLN_modulation_1_weight_to_fp16, x = input_355_cast_fp16)[name = tensor("linear_142_cast_fp16")]; + tensor var_2185_split_sizes_0 = const()[name = tensor("op_2185_split_sizes_0"), val = tensor([512, 512, 512])]; + tensor var_2185_axis_0 = const()[name = tensor("op_2185_axis_0"), val = tensor(-1)]; + tensor var_2185_cast_fp16_0, tensor var_2185_cast_fp16_1, tensor var_2185_cast_fp16_2 = split(axis = var_2185_axis_0, split_sizes = var_2185_split_sizes_0, x = linear_142_cast_fp16)[name = tensor("op_2185_cast_fp16")]; + tensor mean_75_axes_0 = const()[name = tensor("mean_75_axes_0"), val = tensor([-1])]; + tensor mean_75_keep_dims_0 = const()[name = tensor("mean_75_keep_dims_0"), val = tensor(true)]; + tensor mean_75_cast_fp16 = reduce_mean(axes = mean_75_axes_0, keep_dims = mean_75_keep_dims_0, x = x_237_cast_fp16)[name = tensor("mean_75_cast_fp16")]; + tensor sub_61_cast_fp16 = sub(x = x_237_cast_fp16, y = mean_75_cast_fp16)[name = tensor("sub_61_cast_fp16")]; + tensor square_49_cast_fp16 = square(x = sub_61_cast_fp16)[name = tensor("square_49_cast_fp16")]; + tensor reduce_mean_99_axes_0 = const()[name = tensor("reduce_mean_99_axes_0"), val = tensor([-1])]; + tensor reduce_mean_99_keep_dims_0 = const()[name = tensor("reduce_mean_99_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_99_cast_fp16 = reduce_mean(axes = reduce_mean_99_axes_0, keep_dims = reduce_mean_99_keep_dims_0, x = square_49_cast_fp16)[name = tensor("reduce_mean_99_cast_fp16")]; + tensor var_2195_to_fp16 = const()[name = tensor("op_2195_to_fp16"), val = tensor(0x1.1p-20)]; + tensor var_2196_cast_fp16 = add(x = reduce_mean_99_cast_fp16, y = var_2195_to_fp16)[name = tensor("op_2196_cast_fp16")]; + tensor var_2197_cast_fp16 = sqrt(x = var_2196_cast_fp16)[name = tensor("op_2197_cast_fp16")]; + tensor x_239_cast_fp16 = real_div(x = sub_61_cast_fp16, y = var_2197_cast_fp16)[name = tensor("x_239_cast_fp16")]; + tensor var_2199_cast_fp16 = mul(x = x_239_cast_fp16, y = flow_net_res_blocks_2_in_ln_weight_to_fp16)[name = tensor("op_2199_cast_fp16")]; + tensor x_241_cast_fp16 = add(x = var_2199_cast_fp16, y = flow_net_res_blocks_2_in_ln_bias_to_fp16)[name = tensor("x_241_cast_fp16")]; + tensor var_2201_promoted_to_fp16 = const()[name = tensor("op_2201_promoted_to_fp16"), val = tensor(0x1p+0)]; + tensor var_2202_cast_fp16 = add(x = var_2185_cast_fp16_1, y = var_2201_promoted_to_fp16)[name = tensor("op_2202_cast_fp16")]; + tensor var_2203_cast_fp16 = mul(x = x_241_cast_fp16, y = var_2202_cast_fp16)[name = tensor("op_2203_cast_fp16")]; + tensor input_373_cast_fp16 = add(x = var_2203_cast_fp16, y = var_2185_cast_fp16_0)[name = tensor("input_373_cast_fp16")]; + tensor linear_143_cast_fp16 = linear(bias = flow_net_res_blocks_2_mlp_0_bias_to_fp16, weight = flow_net_res_blocks_2_mlp_0_weight_to_fp16, x = input_373_cast_fp16)[name = tensor("linear_143_cast_fp16")]; + tensor input_377_cast_fp16 = silu(x = linear_143_cast_fp16)[name = tensor("input_377_cast_fp16")]; + tensor linear_144_cast_fp16 = linear(bias = flow_net_res_blocks_2_mlp_2_bias_to_fp16, weight = flow_net_res_blocks_2_mlp_2_weight_to_fp16, x = input_377_cast_fp16)[name = tensor("linear_144_cast_fp16")]; + tensor var_2214_cast_fp16 = mul(x = var_2185_cast_fp16_2, y = linear_144_cast_fp16)[name = tensor("op_2214_cast_fp16")]; + tensor x_243_cast_fp16 = add(x = x_237_cast_fp16, y = var_2214_cast_fp16)[name = tensor("x_243_cast_fp16")]; + tensor linear_145_cast_fp16 = linear(bias = flow_net_res_blocks_3_adaLN_modulation_1_bias_to_fp16, weight = flow_net_res_blocks_3_adaLN_modulation_1_weight_to_fp16, x = input_355_cast_fp16)[name = tensor("linear_145_cast_fp16")]; + tensor var_2224_split_sizes_0 = const()[name = tensor("op_2224_split_sizes_0"), val = tensor([512, 512, 512])]; + tensor var_2224_axis_0 = const()[name = tensor("op_2224_axis_0"), val = tensor(-1)]; + tensor var_2224_cast_fp16_0, tensor var_2224_cast_fp16_1, tensor var_2224_cast_fp16_2 = split(axis = var_2224_axis_0, split_sizes = var_2224_split_sizes_0, x = linear_145_cast_fp16)[name = tensor("op_2224_cast_fp16")]; + tensor mean_77_axes_0 = const()[name = tensor("mean_77_axes_0"), val = tensor([-1])]; + tensor mean_77_keep_dims_0 = const()[name = tensor("mean_77_keep_dims_0"), val = tensor(true)]; + tensor mean_77_cast_fp16 = reduce_mean(axes = mean_77_axes_0, keep_dims = mean_77_keep_dims_0, x = x_243_cast_fp16)[name = tensor("mean_77_cast_fp16")]; + tensor sub_62_cast_fp16 = sub(x = x_243_cast_fp16, y = mean_77_cast_fp16)[name = tensor("sub_62_cast_fp16")]; + tensor square_50_cast_fp16 = square(x = sub_62_cast_fp16)[name = tensor("square_50_cast_fp16")]; + tensor reduce_mean_101_axes_0 = const()[name = tensor("reduce_mean_101_axes_0"), val = tensor([-1])]; + tensor reduce_mean_101_keep_dims_0 = const()[name = tensor("reduce_mean_101_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_101_cast_fp16 = reduce_mean(axes = reduce_mean_101_axes_0, keep_dims = reduce_mean_101_keep_dims_0, x = square_50_cast_fp16)[name = tensor("reduce_mean_101_cast_fp16")]; + tensor var_2234_to_fp16 = const()[name = tensor("op_2234_to_fp16"), val = tensor(0x1.1p-20)]; + tensor var_2235_cast_fp16 = add(x = reduce_mean_101_cast_fp16, y = var_2234_to_fp16)[name = tensor("op_2235_cast_fp16")]; + tensor var_2236_cast_fp16 = sqrt(x = var_2235_cast_fp16)[name = tensor("op_2236_cast_fp16")]; + tensor x_245_cast_fp16 = real_div(x = sub_62_cast_fp16, y = var_2236_cast_fp16)[name = tensor("x_245_cast_fp16")]; + tensor var_2238_cast_fp16 = mul(x = x_245_cast_fp16, y = flow_net_res_blocks_3_in_ln_weight_to_fp16)[name = tensor("op_2238_cast_fp16")]; + tensor x_247_cast_fp16 = add(x = var_2238_cast_fp16, y = flow_net_res_blocks_3_in_ln_bias_to_fp16)[name = tensor("x_247_cast_fp16")]; + tensor var_2240_promoted_to_fp16 = const()[name = tensor("op_2240_promoted_to_fp16"), val = tensor(0x1p+0)]; + tensor var_2241_cast_fp16 = add(x = var_2224_cast_fp16_1, y = var_2240_promoted_to_fp16)[name = tensor("op_2241_cast_fp16")]; + tensor var_2242_cast_fp16 = mul(x = x_247_cast_fp16, y = var_2241_cast_fp16)[name = tensor("op_2242_cast_fp16")]; + tensor input_381_cast_fp16 = add(x = var_2242_cast_fp16, y = var_2224_cast_fp16_0)[name = tensor("input_381_cast_fp16")]; + tensor linear_146_cast_fp16 = linear(bias = flow_net_res_blocks_3_mlp_0_bias_to_fp16, weight = flow_net_res_blocks_3_mlp_0_weight_to_fp16, x = input_381_cast_fp16)[name = tensor("linear_146_cast_fp16")]; + tensor input_385_cast_fp16 = silu(x = linear_146_cast_fp16)[name = tensor("input_385_cast_fp16")]; + tensor linear_147_cast_fp16 = linear(bias = flow_net_res_blocks_3_mlp_2_bias_to_fp16, weight = flow_net_res_blocks_3_mlp_2_weight_to_fp16, x = input_385_cast_fp16)[name = tensor("linear_147_cast_fp16")]; + tensor var_2253_cast_fp16 = mul(x = var_2224_cast_fp16_2, y = linear_147_cast_fp16)[name = tensor("op_2253_cast_fp16")]; + tensor x_249_cast_fp16 = add(x = x_243_cast_fp16, y = var_2253_cast_fp16)[name = tensor("x_249_cast_fp16")]; + tensor linear_148_cast_fp16 = linear(bias = flow_net_res_blocks_4_adaLN_modulation_1_bias_to_fp16, weight = flow_net_res_blocks_4_adaLN_modulation_1_weight_to_fp16, x = input_355_cast_fp16)[name = tensor("linear_148_cast_fp16")]; + tensor var_2263_split_sizes_0 = const()[name = tensor("op_2263_split_sizes_0"), val = tensor([512, 512, 512])]; + tensor var_2263_axis_0 = const()[name = tensor("op_2263_axis_0"), val = tensor(-1)]; + tensor var_2263_cast_fp16_0, tensor var_2263_cast_fp16_1, tensor var_2263_cast_fp16_2 = split(axis = var_2263_axis_0, split_sizes = var_2263_split_sizes_0, x = linear_148_cast_fp16)[name = tensor("op_2263_cast_fp16")]; + tensor mean_79_axes_0 = const()[name = tensor("mean_79_axes_0"), val = tensor([-1])]; + tensor mean_79_keep_dims_0 = const()[name = tensor("mean_79_keep_dims_0"), val = tensor(true)]; + tensor mean_79_cast_fp16 = reduce_mean(axes = mean_79_axes_0, keep_dims = mean_79_keep_dims_0, x = x_249_cast_fp16)[name = tensor("mean_79_cast_fp16")]; + tensor sub_63_cast_fp16 = sub(x = x_249_cast_fp16, y = mean_79_cast_fp16)[name = tensor("sub_63_cast_fp16")]; + tensor square_51_cast_fp16 = square(x = sub_63_cast_fp16)[name = tensor("square_51_cast_fp16")]; + tensor reduce_mean_103_axes_0 = const()[name = tensor("reduce_mean_103_axes_0"), val = tensor([-1])]; + tensor reduce_mean_103_keep_dims_0 = const()[name = tensor("reduce_mean_103_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_103_cast_fp16 = reduce_mean(axes = reduce_mean_103_axes_0, keep_dims = reduce_mean_103_keep_dims_0, x = square_51_cast_fp16)[name = tensor("reduce_mean_103_cast_fp16")]; + tensor var_2273_to_fp16 = const()[name = tensor("op_2273_to_fp16"), val = tensor(0x1.1p-20)]; + tensor var_2274_cast_fp16 = add(x = reduce_mean_103_cast_fp16, y = var_2273_to_fp16)[name = tensor("op_2274_cast_fp16")]; + tensor var_2275_cast_fp16 = sqrt(x = var_2274_cast_fp16)[name = tensor("op_2275_cast_fp16")]; + tensor x_251_cast_fp16 = real_div(x = sub_63_cast_fp16, y = var_2275_cast_fp16)[name = tensor("x_251_cast_fp16")]; + tensor var_2277_cast_fp16 = mul(x = x_251_cast_fp16, y = flow_net_res_blocks_4_in_ln_weight_to_fp16)[name = tensor("op_2277_cast_fp16")]; + tensor x_253_cast_fp16 = add(x = var_2277_cast_fp16, y = flow_net_res_blocks_4_in_ln_bias_to_fp16)[name = tensor("x_253_cast_fp16")]; + tensor var_2279_promoted_to_fp16 = const()[name = tensor("op_2279_promoted_to_fp16"), val = tensor(0x1p+0)]; + tensor var_2280_cast_fp16 = add(x = var_2263_cast_fp16_1, y = var_2279_promoted_to_fp16)[name = tensor("op_2280_cast_fp16")]; + tensor var_2281_cast_fp16 = mul(x = x_253_cast_fp16, y = var_2280_cast_fp16)[name = tensor("op_2281_cast_fp16")]; + tensor input_389_cast_fp16 = add(x = var_2281_cast_fp16, y = var_2263_cast_fp16_0)[name = tensor("input_389_cast_fp16")]; + tensor linear_149_cast_fp16 = linear(bias = flow_net_res_blocks_4_mlp_0_bias_to_fp16, weight = flow_net_res_blocks_4_mlp_0_weight_to_fp16, x = input_389_cast_fp16)[name = tensor("linear_149_cast_fp16")]; + tensor input_393_cast_fp16 = silu(x = linear_149_cast_fp16)[name = tensor("input_393_cast_fp16")]; + tensor linear_150_cast_fp16 = linear(bias = flow_net_res_blocks_4_mlp_2_bias_to_fp16, weight = flow_net_res_blocks_4_mlp_2_weight_to_fp16, x = input_393_cast_fp16)[name = tensor("linear_150_cast_fp16")]; + tensor var_2292_cast_fp16 = mul(x = var_2263_cast_fp16_2, y = linear_150_cast_fp16)[name = tensor("op_2292_cast_fp16")]; + tensor x_255_cast_fp16 = add(x = x_249_cast_fp16, y = var_2292_cast_fp16)[name = tensor("x_255_cast_fp16")]; + tensor linear_151_cast_fp16 = linear(bias = flow_net_res_blocks_5_adaLN_modulation_1_bias_to_fp16, weight = flow_net_res_blocks_5_adaLN_modulation_1_weight_to_fp16, x = input_355_cast_fp16)[name = tensor("linear_151_cast_fp16")]; + tensor var_2302_split_sizes_0 = const()[name = tensor("op_2302_split_sizes_0"), val = tensor([512, 512, 512])]; + tensor var_2302_axis_0 = const()[name = tensor("op_2302_axis_0"), val = tensor(-1)]; + tensor var_2302_cast_fp16_0, tensor var_2302_cast_fp16_1, tensor var_2302_cast_fp16_2 = split(axis = var_2302_axis_0, split_sizes = var_2302_split_sizes_0, x = linear_151_cast_fp16)[name = tensor("op_2302_cast_fp16")]; + tensor mean_81_axes_0 = const()[name = tensor("mean_81_axes_0"), val = tensor([-1])]; + tensor mean_81_keep_dims_0 = const()[name = tensor("mean_81_keep_dims_0"), val = tensor(true)]; + tensor mean_81_cast_fp16 = reduce_mean(axes = mean_81_axes_0, keep_dims = mean_81_keep_dims_0, x = x_255_cast_fp16)[name = tensor("mean_81_cast_fp16")]; + tensor sub_64_cast_fp16 = sub(x = x_255_cast_fp16, y = mean_81_cast_fp16)[name = tensor("sub_64_cast_fp16")]; + tensor square_52_cast_fp16 = square(x = sub_64_cast_fp16)[name = tensor("square_52_cast_fp16")]; + tensor reduce_mean_105_axes_0 = const()[name = tensor("reduce_mean_105_axes_0"), val = tensor([-1])]; + tensor reduce_mean_105_keep_dims_0 = const()[name = tensor("reduce_mean_105_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_105_cast_fp16 = reduce_mean(axes = reduce_mean_105_axes_0, keep_dims = reduce_mean_105_keep_dims_0, x = square_52_cast_fp16)[name = tensor("reduce_mean_105_cast_fp16")]; + tensor var_2312_to_fp16 = const()[name = tensor("op_2312_to_fp16"), val = tensor(0x1.1p-20)]; + tensor var_2313_cast_fp16 = add(x = reduce_mean_105_cast_fp16, y = var_2312_to_fp16)[name = tensor("op_2313_cast_fp16")]; + tensor var_2314_cast_fp16 = sqrt(x = var_2313_cast_fp16)[name = tensor("op_2314_cast_fp16")]; + tensor x_257_cast_fp16 = real_div(x = sub_64_cast_fp16, y = var_2314_cast_fp16)[name = tensor("x_257_cast_fp16")]; + tensor var_2316_cast_fp16 = mul(x = x_257_cast_fp16, y = flow_net_res_blocks_5_in_ln_weight_to_fp16)[name = tensor("op_2316_cast_fp16")]; + tensor x_259_cast_fp16 = add(x = var_2316_cast_fp16, y = flow_net_res_blocks_5_in_ln_bias_to_fp16)[name = tensor("x_259_cast_fp16")]; + tensor var_2318_promoted_to_fp16 = const()[name = tensor("op_2318_promoted_to_fp16"), val = tensor(0x1p+0)]; + tensor var_2319_cast_fp16 = add(x = var_2302_cast_fp16_1, y = var_2318_promoted_to_fp16)[name = tensor("op_2319_cast_fp16")]; + tensor var_2320_cast_fp16 = mul(x = x_259_cast_fp16, y = var_2319_cast_fp16)[name = tensor("op_2320_cast_fp16")]; + tensor input_397_cast_fp16 = add(x = var_2320_cast_fp16, y = var_2302_cast_fp16_0)[name = tensor("input_397_cast_fp16")]; + tensor linear_152_cast_fp16 = linear(bias = flow_net_res_blocks_5_mlp_0_bias_to_fp16, weight = flow_net_res_blocks_5_mlp_0_weight_to_fp16, x = input_397_cast_fp16)[name = tensor("linear_152_cast_fp16")]; + tensor input_401_cast_fp16 = silu(x = linear_152_cast_fp16)[name = tensor("input_401_cast_fp16")]; + tensor linear_153_cast_fp16 = linear(bias = flow_net_res_blocks_5_mlp_2_bias_to_fp16, weight = flow_net_res_blocks_5_mlp_2_weight_to_fp16, x = input_401_cast_fp16)[name = tensor("linear_153_cast_fp16")]; + tensor var_2331_cast_fp16 = mul(x = var_2302_cast_fp16_2, y = linear_153_cast_fp16)[name = tensor("op_2331_cast_fp16")]; + tensor x_261_cast_fp16 = add(x = x_255_cast_fp16, y = var_2331_cast_fp16)[name = tensor("x_261_cast_fp16")]; + tensor linear_154_cast_fp16 = linear(bias = flow_net_final_layer_adaLN_modulation_1_bias_to_fp16, weight = flow_net_final_layer_adaLN_modulation_1_weight_to_fp16, x = input_355_cast_fp16)[name = tensor("linear_154_cast_fp16")]; + tensor var_2340_split_sizes_0 = const()[name = tensor("op_2340_split_sizes_0"), val = tensor([512, 512])]; + tensor var_2340_axis_0 = const()[name = tensor("op_2340_axis_0"), val = tensor(-1)]; + tensor var_2340_cast_fp16_0, tensor var_2340_cast_fp16_1 = split(axis = var_2340_axis_0, split_sizes = var_2340_split_sizes_0, x = linear_154_cast_fp16)[name = tensor("op_2340_cast_fp16")]; + tensor mean_83_axes_0 = const()[name = tensor("mean_83_axes_0"), val = tensor([-1])]; + tensor mean_83_keep_dims_0 = const()[name = tensor("mean_83_keep_dims_0"), val = tensor(true)]; + tensor mean_83_cast_fp16 = reduce_mean(axes = mean_83_axes_0, keep_dims = mean_83_keep_dims_0, x = x_261_cast_fp16)[name = tensor("mean_83_cast_fp16")]; + tensor sub_65_cast_fp16 = sub(x = x_261_cast_fp16, y = mean_83_cast_fp16)[name = tensor("sub_65_cast_fp16")]; + tensor square_53_cast_fp16 = square(x = sub_65_cast_fp16)[name = tensor("square_53_cast_fp16")]; + tensor reduce_mean_107_axes_0 = const()[name = tensor("reduce_mean_107_axes_0"), val = tensor([-1])]; + tensor reduce_mean_107_keep_dims_0 = const()[name = tensor("reduce_mean_107_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_107_cast_fp16 = reduce_mean(axes = reduce_mean_107_axes_0, keep_dims = reduce_mean_107_keep_dims_0, x = square_53_cast_fp16)[name = tensor("reduce_mean_107_cast_fp16")]; + tensor var_2347_to_fp16 = const()[name = tensor("op_2347_to_fp16"), val = tensor(0x1.1p-20)]; + tensor var_2348_cast_fp16 = add(x = reduce_mean_107_cast_fp16, y = var_2347_to_fp16)[name = tensor("op_2348_cast_fp16")]; + tensor var_2349_cast_fp16 = sqrt(x = var_2348_cast_fp16)[name = tensor("op_2349_cast_fp16")]; + tensor x_263_cast_fp16 = real_div(x = sub_65_cast_fp16, y = var_2349_cast_fp16)[name = tensor("x_263_cast_fp16")]; + tensor var_2351_promoted_to_fp16 = const()[name = tensor("op_2351_promoted_to_fp16"), val = tensor(0x1p+0)]; + tensor var_2352_cast_fp16 = add(x = var_2340_cast_fp16_1, y = var_2351_promoted_to_fp16)[name = tensor("op_2352_cast_fp16")]; + tensor var_2353_cast_fp16 = mul(x = x_263_cast_fp16, y = var_2352_cast_fp16)[name = tensor("op_2353_cast_fp16")]; + tensor input_405_cast_fp16 = add(x = var_2353_cast_fp16, y = var_2340_cast_fp16_0)[name = tensor("input_405_cast_fp16")]; + tensor linear_155_cast_fp16 = linear(bias = flow_net_final_layer_linear_bias_to_fp16, weight = flow_net_final_layer_linear_weight_to_fp16, x = input_405_cast_fp16)[name = tensor("linear_155_cast_fp16")]; + tensor var_2364_to_fp16 = const()[name = tensor("op_2364_to_fp16"), val = tensor(0x1p-3)]; + tensor var_2365_cast_fp16 = mul(x = linear_155_cast_fp16, y = var_2364_to_fp16)[name = tensor("op_2365_cast_fp16")]; + tensor input_407_cast_fp16 = add(x = input_339_cast_fp16, y = var_2365_cast_fp16)[name = tensor("input_407_cast_fp16")]; + tensor linear_156_cast_fp16 = linear(bias = flow_net_input_proj_bias_to_fp16, weight = flow_net_input_proj_weight_to_fp16, x = input_407_cast_fp16)[name = tensor("linear_156_cast_fp16")]; + tensor input_411_to_fp16 = const()[name = tensor("input_411_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(19008256)))]; + tensor input_413_cast_fp16 = silu(x = input_411_to_fp16)[name = tensor("input_413_cast_fp16")]; + tensor linear_158_cast_fp16 = linear(bias = flow_net_time_embed_0_mlp_2_bias_to_fp16, weight = flow_net_time_embed_0_mlp_2_weight_to_fp16, x = input_413_cast_fp16)[name = tensor("linear_158_cast_fp16")]; + tensor reduce_mean_108_axes_0 = const()[name = tensor("reduce_mean_108_axes_0"), val = tensor([-1])]; + tensor reduce_mean_108_keep_dims_0 = const()[name = tensor("reduce_mean_108_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_108_cast_fp16 = reduce_mean(axes = reduce_mean_108_axes_0, keep_dims = reduce_mean_108_keep_dims_0, x = linear_158_cast_fp16)[name = tensor("reduce_mean_108_cast_fp16")]; + tensor sub_66_cast_fp16 = sub(x = linear_158_cast_fp16, y = reduce_mean_108_cast_fp16)[name = tensor("sub_66_cast_fp16")]; + tensor square_54_cast_fp16 = square(x = sub_66_cast_fp16)[name = tensor("square_54_cast_fp16")]; + tensor reduce_mean_109_axes_0 = const()[name = tensor("reduce_mean_109_axes_0"), val = tensor([-1])]; + tensor reduce_mean_109_keep_dims_0 = const()[name = tensor("reduce_mean_109_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_109_cast_fp16 = reduce_mean(axes = reduce_mean_109_axes_0, keep_dims = reduce_mean_109_keep_dims_0, x = square_54_cast_fp16)[name = tensor("reduce_mean_109_cast_fp16")]; + tensor real_div_12_to_fp16 = const()[name = tensor("real_div_12_to_fp16"), val = tensor(0x1.008p+0)]; + tensor mul_12_cast_fp16 = mul(x = reduce_mean_109_cast_fp16, y = real_div_12_to_fp16)[name = tensor("mul_12_cast_fp16")]; + tensor var_2431_to_fp16 = const()[name = tensor("op_2431_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_109_cast_fp16 = add(x = mul_12_cast_fp16, y = var_2431_to_fp16)[name = tensor("var_109_cast_fp16")]; + tensor var_2434_epsilon_0 = const()[name = tensor("op_2434_epsilon_0"), val = tensor(0x1.197998p-40)]; + tensor var_2434_cast_fp16 = rsqrt(epsilon = var_2434_epsilon_0, x = var_109_cast_fp16)[name = tensor("op_2434_cast_fp16")]; + tensor var_2435_cast_fp16 = mul(x = const_3_to_fp16, y = var_2434_cast_fp16)[name = tensor("op_2435_cast_fp16")]; + tensor var_2436_cast_fp16 = mul(x = linear_158_cast_fp16, y = var_2435_cast_fp16)[name = tensor("op_2436_cast_fp16")]; + tensor input_417_to_fp16 = const()[name = tensor("input_417_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(19009344)))]; + tensor input_419_cast_fp16 = silu(x = input_417_to_fp16)[name = tensor("input_419_cast_fp16")]; + tensor linear_160_cast_fp16 = linear(bias = flow_net_time_embed_1_mlp_2_bias_to_fp16, weight = flow_net_time_embed_1_mlp_2_weight_to_fp16, x = input_419_cast_fp16)[name = tensor("linear_160_cast_fp16")]; + tensor reduce_mean_110_axes_0 = const()[name = tensor("reduce_mean_110_axes_0"), val = tensor([-1])]; + tensor reduce_mean_110_keep_dims_0 = const()[name = tensor("reduce_mean_110_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_110_cast_fp16 = reduce_mean(axes = reduce_mean_110_axes_0, keep_dims = reduce_mean_110_keep_dims_0, x = linear_160_cast_fp16)[name = tensor("reduce_mean_110_cast_fp16")]; + tensor sub_68_cast_fp16 = sub(x = linear_160_cast_fp16, y = reduce_mean_110_cast_fp16)[name = tensor("sub_68_cast_fp16")]; + tensor square_55_cast_fp16 = square(x = sub_68_cast_fp16)[name = tensor("square_55_cast_fp16")]; + tensor reduce_mean_111_axes_0 = const()[name = tensor("reduce_mean_111_axes_0"), val = tensor([-1])]; + tensor reduce_mean_111_keep_dims_0 = const()[name = tensor("reduce_mean_111_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_111_cast_fp16 = reduce_mean(axes = reduce_mean_111_axes_0, keep_dims = reduce_mean_111_keep_dims_0, x = square_55_cast_fp16)[name = tensor("reduce_mean_111_cast_fp16")]; + tensor real_div_13_to_fp16 = const()[name = tensor("real_div_13_to_fp16"), val = tensor(0x1.008p+0)]; + tensor mul_13_cast_fp16 = mul(x = reduce_mean_111_cast_fp16, y = real_div_13_to_fp16)[name = tensor("mul_13_cast_fp16")]; + tensor var_2468_to_fp16 = const()[name = tensor("op_2468_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_111_cast_fp16 = add(x = mul_13_cast_fp16, y = var_2468_to_fp16)[name = tensor("var_111_cast_fp16")]; + tensor var_2471_epsilon_0 = const()[name = tensor("op_2471_epsilon_0"), val = tensor(0x1.197998p-40)]; + tensor var_2471_cast_fp16 = rsqrt(epsilon = var_2471_epsilon_0, x = var_111_cast_fp16)[name = tensor("op_2471_cast_fp16")]; + tensor var_2472_cast_fp16 = mul(x = const_5_to_fp16, y = var_2471_cast_fp16)[name = tensor("op_2472_cast_fp16")]; + tensor var_2473_cast_fp16 = mul(x = linear_160_cast_fp16, y = var_2472_cast_fp16)[name = tensor("op_2473_cast_fp16")]; + tensor var_2485_cast_fp16 = add(x = var_2436_cast_fp16, y = var_2473_cast_fp16)[name = tensor("op_2485_cast_fp16")]; + tensor _inversed_t_combined_13_y_0_to_fp16 = const()[name = tensor("_inversed_t_combined_13_y_0_to_fp16"), val = tensor(0x1p-1)]; + tensor _inversed_t_combined_13_cast_fp16 = mul(x = var_2485_cast_fp16, y = _inversed_t_combined_13_y_0_to_fp16)[name = tensor("_inversed_t_combined_13_cast_fp16")]; + tensor input_421_cast_fp16 = add(x = _inversed_t_combined_13_cast_fp16, y = linear_5_cast_fp16)[name = tensor("input_421_cast_fp16")]; + tensor input_423_cast_fp16 = silu(x = input_421_cast_fp16)[name = tensor("input_423_cast_fp16")]; + tensor linear_162_cast_fp16 = linear(bias = flow_net_res_blocks_0_adaLN_modulation_1_bias_to_fp16, weight = flow_net_res_blocks_0_adaLN_modulation_1_weight_to_fp16, x = input_423_cast_fp16)[name = tensor("linear_162_cast_fp16")]; + tensor var_2500_split_sizes_0 = const()[name = tensor("op_2500_split_sizes_0"), val = tensor([512, 512, 512])]; + tensor var_2500_axis_0 = const()[name = tensor("op_2500_axis_0"), val = tensor(-1)]; + tensor var_2500_cast_fp16_0, tensor var_2500_cast_fp16_1, tensor var_2500_cast_fp16_2 = split(axis = var_2500_axis_0, split_sizes = var_2500_split_sizes_0, x = linear_162_cast_fp16)[name = tensor("op_2500_cast_fp16")]; + tensor mean_85_axes_0 = const()[name = tensor("mean_85_axes_0"), val = tensor([-1])]; + tensor mean_85_keep_dims_0 = const()[name = tensor("mean_85_keep_dims_0"), val = tensor(true)]; + tensor mean_85_cast_fp16 = reduce_mean(axes = mean_85_axes_0, keep_dims = mean_85_keep_dims_0, x = linear_156_cast_fp16)[name = tensor("mean_85_cast_fp16")]; + tensor sub_70_cast_fp16 = sub(x = linear_156_cast_fp16, y = mean_85_cast_fp16)[name = tensor("sub_70_cast_fp16")]; + tensor square_56_cast_fp16 = square(x = sub_70_cast_fp16)[name = tensor("square_56_cast_fp16")]; + tensor reduce_mean_113_axes_0 = const()[name = tensor("reduce_mean_113_axes_0"), val = tensor([-1])]; + tensor reduce_mean_113_keep_dims_0 = const()[name = tensor("reduce_mean_113_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_113_cast_fp16 = reduce_mean(axes = reduce_mean_113_axes_0, keep_dims = reduce_mean_113_keep_dims_0, x = square_56_cast_fp16)[name = tensor("reduce_mean_113_cast_fp16")]; + tensor var_2510_to_fp16 = const()[name = tensor("op_2510_to_fp16"), val = tensor(0x1.1p-20)]; + tensor var_2511_cast_fp16 = add(x = reduce_mean_113_cast_fp16, y = var_2510_to_fp16)[name = tensor("op_2511_cast_fp16")]; + tensor var_2512_cast_fp16 = sqrt(x = var_2511_cast_fp16)[name = tensor("op_2512_cast_fp16")]; + tensor x_271_cast_fp16 = real_div(x = sub_70_cast_fp16, y = var_2512_cast_fp16)[name = tensor("x_271_cast_fp16")]; + tensor var_2514_cast_fp16 = mul(x = x_271_cast_fp16, y = flow_net_res_blocks_0_in_ln_weight_to_fp16)[name = tensor("op_2514_cast_fp16")]; + tensor x_273_cast_fp16 = add(x = var_2514_cast_fp16, y = flow_net_res_blocks_0_in_ln_bias_to_fp16)[name = tensor("x_273_cast_fp16")]; + tensor var_2516_promoted_to_fp16 = const()[name = tensor("op_2516_promoted_to_fp16"), val = tensor(0x1p+0)]; + tensor var_2517_cast_fp16 = add(x = var_2500_cast_fp16_1, y = var_2516_promoted_to_fp16)[name = tensor("op_2517_cast_fp16")]; + tensor var_2518_cast_fp16 = mul(x = x_273_cast_fp16, y = var_2517_cast_fp16)[name = tensor("op_2518_cast_fp16")]; + tensor input_425_cast_fp16 = add(x = var_2518_cast_fp16, y = var_2500_cast_fp16_0)[name = tensor("input_425_cast_fp16")]; + tensor linear_163_cast_fp16 = linear(bias = flow_net_res_blocks_0_mlp_0_bias_to_fp16, weight = flow_net_res_blocks_0_mlp_0_weight_to_fp16, x = input_425_cast_fp16)[name = tensor("linear_163_cast_fp16")]; + tensor input_429_cast_fp16 = silu(x = linear_163_cast_fp16)[name = tensor("input_429_cast_fp16")]; + tensor linear_164_cast_fp16 = linear(bias = flow_net_res_blocks_0_mlp_2_bias_to_fp16, weight = flow_net_res_blocks_0_mlp_2_weight_to_fp16, x = input_429_cast_fp16)[name = tensor("linear_164_cast_fp16")]; + tensor var_2529_cast_fp16 = mul(x = var_2500_cast_fp16_2, y = linear_164_cast_fp16)[name = tensor("op_2529_cast_fp16")]; + tensor x_275_cast_fp16 = add(x = linear_156_cast_fp16, y = var_2529_cast_fp16)[name = tensor("x_275_cast_fp16")]; + tensor linear_165_cast_fp16 = linear(bias = flow_net_res_blocks_1_adaLN_modulation_1_bias_to_fp16, weight = flow_net_res_blocks_1_adaLN_modulation_1_weight_to_fp16, x = input_423_cast_fp16)[name = tensor("linear_165_cast_fp16")]; + tensor var_2539_split_sizes_0 = const()[name = tensor("op_2539_split_sizes_0"), val = tensor([512, 512, 512])]; + tensor var_2539_axis_0 = const()[name = tensor("op_2539_axis_0"), val = tensor(-1)]; + tensor var_2539_cast_fp16_0, tensor var_2539_cast_fp16_1, tensor var_2539_cast_fp16_2 = split(axis = var_2539_axis_0, split_sizes = var_2539_split_sizes_0, x = linear_165_cast_fp16)[name = tensor("op_2539_cast_fp16")]; + tensor mean_87_axes_0 = const()[name = tensor("mean_87_axes_0"), val = tensor([-1])]; + tensor mean_87_keep_dims_0 = const()[name = tensor("mean_87_keep_dims_0"), val = tensor(true)]; + tensor mean_87_cast_fp16 = reduce_mean(axes = mean_87_axes_0, keep_dims = mean_87_keep_dims_0, x = x_275_cast_fp16)[name = tensor("mean_87_cast_fp16")]; + tensor sub_71_cast_fp16 = sub(x = x_275_cast_fp16, y = mean_87_cast_fp16)[name = tensor("sub_71_cast_fp16")]; + tensor square_57_cast_fp16 = square(x = sub_71_cast_fp16)[name = tensor("square_57_cast_fp16")]; + tensor reduce_mean_115_axes_0 = const()[name = tensor("reduce_mean_115_axes_0"), val = tensor([-1])]; + tensor reduce_mean_115_keep_dims_0 = const()[name = tensor("reduce_mean_115_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_115_cast_fp16 = reduce_mean(axes = reduce_mean_115_axes_0, keep_dims = reduce_mean_115_keep_dims_0, x = square_57_cast_fp16)[name = tensor("reduce_mean_115_cast_fp16")]; + tensor var_2549_to_fp16 = const()[name = tensor("op_2549_to_fp16"), val = tensor(0x1.1p-20)]; + tensor var_2550_cast_fp16 = add(x = reduce_mean_115_cast_fp16, y = var_2549_to_fp16)[name = tensor("op_2550_cast_fp16")]; + tensor var_2551_cast_fp16 = sqrt(x = var_2550_cast_fp16)[name = tensor("op_2551_cast_fp16")]; + tensor x_277_cast_fp16 = real_div(x = sub_71_cast_fp16, y = var_2551_cast_fp16)[name = tensor("x_277_cast_fp16")]; + tensor var_2553_cast_fp16 = mul(x = x_277_cast_fp16, y = flow_net_res_blocks_1_in_ln_weight_to_fp16)[name = tensor("op_2553_cast_fp16")]; + tensor x_279_cast_fp16 = add(x = var_2553_cast_fp16, y = flow_net_res_blocks_1_in_ln_bias_to_fp16)[name = tensor("x_279_cast_fp16")]; + tensor var_2555_promoted_to_fp16 = const()[name = tensor("op_2555_promoted_to_fp16"), val = tensor(0x1p+0)]; + tensor var_2556_cast_fp16 = add(x = var_2539_cast_fp16_1, y = var_2555_promoted_to_fp16)[name = tensor("op_2556_cast_fp16")]; + tensor var_2557_cast_fp16 = mul(x = x_279_cast_fp16, y = var_2556_cast_fp16)[name = tensor("op_2557_cast_fp16")]; + tensor input_433_cast_fp16 = add(x = var_2557_cast_fp16, y = var_2539_cast_fp16_0)[name = tensor("input_433_cast_fp16")]; + tensor linear_166_cast_fp16 = linear(bias = flow_net_res_blocks_1_mlp_0_bias_to_fp16, weight = flow_net_res_blocks_1_mlp_0_weight_to_fp16, x = input_433_cast_fp16)[name = tensor("linear_166_cast_fp16")]; + tensor input_437_cast_fp16 = silu(x = linear_166_cast_fp16)[name = tensor("input_437_cast_fp16")]; + tensor linear_167_cast_fp16 = linear(bias = flow_net_res_blocks_1_mlp_2_bias_to_fp16, weight = flow_net_res_blocks_1_mlp_2_weight_to_fp16, x = input_437_cast_fp16)[name = tensor("linear_167_cast_fp16")]; + tensor var_2568_cast_fp16 = mul(x = var_2539_cast_fp16_2, y = linear_167_cast_fp16)[name = tensor("op_2568_cast_fp16")]; + tensor x_281_cast_fp16 = add(x = x_275_cast_fp16, y = var_2568_cast_fp16)[name = tensor("x_281_cast_fp16")]; + tensor linear_168_cast_fp16 = linear(bias = flow_net_res_blocks_2_adaLN_modulation_1_bias_to_fp16, weight = flow_net_res_blocks_2_adaLN_modulation_1_weight_to_fp16, x = input_423_cast_fp16)[name = tensor("linear_168_cast_fp16")]; + tensor var_2578_split_sizes_0 = const()[name = tensor("op_2578_split_sizes_0"), val = tensor([512, 512, 512])]; + tensor var_2578_axis_0 = const()[name = tensor("op_2578_axis_0"), val = tensor(-1)]; + tensor var_2578_cast_fp16_0, tensor var_2578_cast_fp16_1, tensor var_2578_cast_fp16_2 = split(axis = var_2578_axis_0, split_sizes = var_2578_split_sizes_0, x = linear_168_cast_fp16)[name = tensor("op_2578_cast_fp16")]; + tensor mean_89_axes_0 = const()[name = tensor("mean_89_axes_0"), val = tensor([-1])]; + tensor mean_89_keep_dims_0 = const()[name = tensor("mean_89_keep_dims_0"), val = tensor(true)]; + tensor mean_89_cast_fp16 = reduce_mean(axes = mean_89_axes_0, keep_dims = mean_89_keep_dims_0, x = x_281_cast_fp16)[name = tensor("mean_89_cast_fp16")]; + tensor sub_72_cast_fp16 = sub(x = x_281_cast_fp16, y = mean_89_cast_fp16)[name = tensor("sub_72_cast_fp16")]; + tensor square_58_cast_fp16 = square(x = sub_72_cast_fp16)[name = tensor("square_58_cast_fp16")]; + tensor reduce_mean_117_axes_0 = const()[name = tensor("reduce_mean_117_axes_0"), val = tensor([-1])]; + tensor reduce_mean_117_keep_dims_0 = const()[name = tensor("reduce_mean_117_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_117_cast_fp16 = reduce_mean(axes = reduce_mean_117_axes_0, keep_dims = reduce_mean_117_keep_dims_0, x = square_58_cast_fp16)[name = tensor("reduce_mean_117_cast_fp16")]; + tensor var_2588_to_fp16 = const()[name = tensor("op_2588_to_fp16"), val = tensor(0x1.1p-20)]; + tensor var_2589_cast_fp16 = add(x = reduce_mean_117_cast_fp16, y = var_2588_to_fp16)[name = tensor("op_2589_cast_fp16")]; + tensor var_2590_cast_fp16 = sqrt(x = var_2589_cast_fp16)[name = tensor("op_2590_cast_fp16")]; + tensor x_283_cast_fp16 = real_div(x = sub_72_cast_fp16, y = var_2590_cast_fp16)[name = tensor("x_283_cast_fp16")]; + tensor var_2592_cast_fp16 = mul(x = x_283_cast_fp16, y = flow_net_res_blocks_2_in_ln_weight_to_fp16)[name = tensor("op_2592_cast_fp16")]; + tensor x_285_cast_fp16 = add(x = var_2592_cast_fp16, y = flow_net_res_blocks_2_in_ln_bias_to_fp16)[name = tensor("x_285_cast_fp16")]; + tensor var_2594_promoted_to_fp16 = const()[name = tensor("op_2594_promoted_to_fp16"), val = tensor(0x1p+0)]; + tensor var_2595_cast_fp16 = add(x = var_2578_cast_fp16_1, y = var_2594_promoted_to_fp16)[name = tensor("op_2595_cast_fp16")]; + tensor var_2596_cast_fp16 = mul(x = x_285_cast_fp16, y = var_2595_cast_fp16)[name = tensor("op_2596_cast_fp16")]; + tensor input_441_cast_fp16 = add(x = var_2596_cast_fp16, y = var_2578_cast_fp16_0)[name = tensor("input_441_cast_fp16")]; + tensor linear_169_cast_fp16 = linear(bias = flow_net_res_blocks_2_mlp_0_bias_to_fp16, weight = flow_net_res_blocks_2_mlp_0_weight_to_fp16, x = input_441_cast_fp16)[name = tensor("linear_169_cast_fp16")]; + tensor input_445_cast_fp16 = silu(x = linear_169_cast_fp16)[name = tensor("input_445_cast_fp16")]; + tensor linear_170_cast_fp16 = linear(bias = flow_net_res_blocks_2_mlp_2_bias_to_fp16, weight = flow_net_res_blocks_2_mlp_2_weight_to_fp16, x = input_445_cast_fp16)[name = tensor("linear_170_cast_fp16")]; + tensor var_2607_cast_fp16 = mul(x = var_2578_cast_fp16_2, y = linear_170_cast_fp16)[name = tensor("op_2607_cast_fp16")]; + tensor x_287_cast_fp16 = add(x = x_281_cast_fp16, y = var_2607_cast_fp16)[name = tensor("x_287_cast_fp16")]; + tensor linear_171_cast_fp16 = linear(bias = flow_net_res_blocks_3_adaLN_modulation_1_bias_to_fp16, weight = flow_net_res_blocks_3_adaLN_modulation_1_weight_to_fp16, x = input_423_cast_fp16)[name = tensor("linear_171_cast_fp16")]; + tensor var_2617_split_sizes_0 = const()[name = tensor("op_2617_split_sizes_0"), val = tensor([512, 512, 512])]; + tensor var_2617_axis_0 = const()[name = tensor("op_2617_axis_0"), val = tensor(-1)]; + tensor var_2617_cast_fp16_0, tensor var_2617_cast_fp16_1, tensor var_2617_cast_fp16_2 = split(axis = var_2617_axis_0, split_sizes = var_2617_split_sizes_0, x = linear_171_cast_fp16)[name = tensor("op_2617_cast_fp16")]; + tensor mean_91_axes_0 = const()[name = tensor("mean_91_axes_0"), val = tensor([-1])]; + tensor mean_91_keep_dims_0 = const()[name = tensor("mean_91_keep_dims_0"), val = tensor(true)]; + tensor mean_91_cast_fp16 = reduce_mean(axes = mean_91_axes_0, keep_dims = mean_91_keep_dims_0, x = x_287_cast_fp16)[name = tensor("mean_91_cast_fp16")]; + tensor sub_73_cast_fp16 = sub(x = x_287_cast_fp16, y = mean_91_cast_fp16)[name = tensor("sub_73_cast_fp16")]; + tensor square_59_cast_fp16 = square(x = sub_73_cast_fp16)[name = tensor("square_59_cast_fp16")]; + tensor reduce_mean_119_axes_0 = const()[name = tensor("reduce_mean_119_axes_0"), val = tensor([-1])]; + tensor reduce_mean_119_keep_dims_0 = const()[name = tensor("reduce_mean_119_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_119_cast_fp16 = reduce_mean(axes = reduce_mean_119_axes_0, keep_dims = reduce_mean_119_keep_dims_0, x = square_59_cast_fp16)[name = tensor("reduce_mean_119_cast_fp16")]; + tensor var_2627_to_fp16 = const()[name = tensor("op_2627_to_fp16"), val = tensor(0x1.1p-20)]; + tensor var_2628_cast_fp16 = add(x = reduce_mean_119_cast_fp16, y = var_2627_to_fp16)[name = tensor("op_2628_cast_fp16")]; + tensor var_2629_cast_fp16 = sqrt(x = var_2628_cast_fp16)[name = tensor("op_2629_cast_fp16")]; + tensor x_289_cast_fp16 = real_div(x = sub_73_cast_fp16, y = var_2629_cast_fp16)[name = tensor("x_289_cast_fp16")]; + tensor var_2631_cast_fp16 = mul(x = x_289_cast_fp16, y = flow_net_res_blocks_3_in_ln_weight_to_fp16)[name = tensor("op_2631_cast_fp16")]; + tensor x_291_cast_fp16 = add(x = var_2631_cast_fp16, y = flow_net_res_blocks_3_in_ln_bias_to_fp16)[name = tensor("x_291_cast_fp16")]; + tensor var_2633_promoted_to_fp16 = const()[name = tensor("op_2633_promoted_to_fp16"), val = tensor(0x1p+0)]; + tensor var_2634_cast_fp16 = add(x = var_2617_cast_fp16_1, y = var_2633_promoted_to_fp16)[name = tensor("op_2634_cast_fp16")]; + tensor var_2635_cast_fp16 = mul(x = x_291_cast_fp16, y = var_2634_cast_fp16)[name = tensor("op_2635_cast_fp16")]; + tensor input_449_cast_fp16 = add(x = var_2635_cast_fp16, y = var_2617_cast_fp16_0)[name = tensor("input_449_cast_fp16")]; + tensor linear_172_cast_fp16 = linear(bias = flow_net_res_blocks_3_mlp_0_bias_to_fp16, weight = flow_net_res_blocks_3_mlp_0_weight_to_fp16, x = input_449_cast_fp16)[name = tensor("linear_172_cast_fp16")]; + tensor input_453_cast_fp16 = silu(x = linear_172_cast_fp16)[name = tensor("input_453_cast_fp16")]; + tensor linear_173_cast_fp16 = linear(bias = flow_net_res_blocks_3_mlp_2_bias_to_fp16, weight = flow_net_res_blocks_3_mlp_2_weight_to_fp16, x = input_453_cast_fp16)[name = tensor("linear_173_cast_fp16")]; + tensor var_2646_cast_fp16 = mul(x = var_2617_cast_fp16_2, y = linear_173_cast_fp16)[name = tensor("op_2646_cast_fp16")]; + tensor x_293_cast_fp16 = add(x = x_287_cast_fp16, y = var_2646_cast_fp16)[name = tensor("x_293_cast_fp16")]; + tensor linear_174_cast_fp16 = linear(bias = flow_net_res_blocks_4_adaLN_modulation_1_bias_to_fp16, weight = flow_net_res_blocks_4_adaLN_modulation_1_weight_to_fp16, x = input_423_cast_fp16)[name = tensor("linear_174_cast_fp16")]; + tensor var_2656_split_sizes_0 = const()[name = tensor("op_2656_split_sizes_0"), val = tensor([512, 512, 512])]; + tensor var_2656_axis_0 = const()[name = tensor("op_2656_axis_0"), val = tensor(-1)]; + tensor var_2656_cast_fp16_0, tensor var_2656_cast_fp16_1, tensor var_2656_cast_fp16_2 = split(axis = var_2656_axis_0, split_sizes = var_2656_split_sizes_0, x = linear_174_cast_fp16)[name = tensor("op_2656_cast_fp16")]; + tensor mean_93_axes_0 = const()[name = tensor("mean_93_axes_0"), val = tensor([-1])]; + tensor mean_93_keep_dims_0 = const()[name = tensor("mean_93_keep_dims_0"), val = tensor(true)]; + tensor mean_93_cast_fp16 = reduce_mean(axes = mean_93_axes_0, keep_dims = mean_93_keep_dims_0, x = x_293_cast_fp16)[name = tensor("mean_93_cast_fp16")]; + tensor sub_74_cast_fp16 = sub(x = x_293_cast_fp16, y = mean_93_cast_fp16)[name = tensor("sub_74_cast_fp16")]; + tensor square_60_cast_fp16 = square(x = sub_74_cast_fp16)[name = tensor("square_60_cast_fp16")]; + tensor reduce_mean_121_axes_0 = const()[name = tensor("reduce_mean_121_axes_0"), val = tensor([-1])]; + tensor reduce_mean_121_keep_dims_0 = const()[name = tensor("reduce_mean_121_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_121_cast_fp16 = reduce_mean(axes = reduce_mean_121_axes_0, keep_dims = reduce_mean_121_keep_dims_0, x = square_60_cast_fp16)[name = tensor("reduce_mean_121_cast_fp16")]; + tensor var_2666_to_fp16 = const()[name = tensor("op_2666_to_fp16"), val = tensor(0x1.1p-20)]; + tensor var_2667_cast_fp16 = add(x = reduce_mean_121_cast_fp16, y = var_2666_to_fp16)[name = tensor("op_2667_cast_fp16")]; + tensor var_2668_cast_fp16 = sqrt(x = var_2667_cast_fp16)[name = tensor("op_2668_cast_fp16")]; + tensor x_295_cast_fp16 = real_div(x = sub_74_cast_fp16, y = var_2668_cast_fp16)[name = tensor("x_295_cast_fp16")]; + tensor var_2670_cast_fp16 = mul(x = x_295_cast_fp16, y = flow_net_res_blocks_4_in_ln_weight_to_fp16)[name = tensor("op_2670_cast_fp16")]; + tensor x_297_cast_fp16 = add(x = var_2670_cast_fp16, y = flow_net_res_blocks_4_in_ln_bias_to_fp16)[name = tensor("x_297_cast_fp16")]; + tensor var_2672_promoted_to_fp16 = const()[name = tensor("op_2672_promoted_to_fp16"), val = tensor(0x1p+0)]; + tensor var_2673_cast_fp16 = add(x = var_2656_cast_fp16_1, y = var_2672_promoted_to_fp16)[name = tensor("op_2673_cast_fp16")]; + tensor var_2674_cast_fp16 = mul(x = x_297_cast_fp16, y = var_2673_cast_fp16)[name = tensor("op_2674_cast_fp16")]; + tensor input_457_cast_fp16 = add(x = var_2674_cast_fp16, y = var_2656_cast_fp16_0)[name = tensor("input_457_cast_fp16")]; + tensor linear_175_cast_fp16 = linear(bias = flow_net_res_blocks_4_mlp_0_bias_to_fp16, weight = flow_net_res_blocks_4_mlp_0_weight_to_fp16, x = input_457_cast_fp16)[name = tensor("linear_175_cast_fp16")]; + tensor input_461_cast_fp16 = silu(x = linear_175_cast_fp16)[name = tensor("input_461_cast_fp16")]; + tensor linear_176_cast_fp16 = linear(bias = flow_net_res_blocks_4_mlp_2_bias_to_fp16, weight = flow_net_res_blocks_4_mlp_2_weight_to_fp16, x = input_461_cast_fp16)[name = tensor("linear_176_cast_fp16")]; + tensor var_2685_cast_fp16 = mul(x = var_2656_cast_fp16_2, y = linear_176_cast_fp16)[name = tensor("op_2685_cast_fp16")]; + tensor x_299_cast_fp16 = add(x = x_293_cast_fp16, y = var_2685_cast_fp16)[name = tensor("x_299_cast_fp16")]; + tensor linear_177_cast_fp16 = linear(bias = flow_net_res_blocks_5_adaLN_modulation_1_bias_to_fp16, weight = flow_net_res_blocks_5_adaLN_modulation_1_weight_to_fp16, x = input_423_cast_fp16)[name = tensor("linear_177_cast_fp16")]; + tensor var_2695_split_sizes_0 = const()[name = tensor("op_2695_split_sizes_0"), val = tensor([512, 512, 512])]; + tensor var_2695_axis_0 = const()[name = tensor("op_2695_axis_0"), val = tensor(-1)]; + tensor var_2695_cast_fp16_0, tensor var_2695_cast_fp16_1, tensor var_2695_cast_fp16_2 = split(axis = var_2695_axis_0, split_sizes = var_2695_split_sizes_0, x = linear_177_cast_fp16)[name = tensor("op_2695_cast_fp16")]; + tensor mean_95_axes_0 = const()[name = tensor("mean_95_axes_0"), val = tensor([-1])]; + tensor mean_95_keep_dims_0 = const()[name = tensor("mean_95_keep_dims_0"), val = tensor(true)]; + tensor mean_95_cast_fp16 = reduce_mean(axes = mean_95_axes_0, keep_dims = mean_95_keep_dims_0, x = x_299_cast_fp16)[name = tensor("mean_95_cast_fp16")]; + tensor sub_75_cast_fp16 = sub(x = x_299_cast_fp16, y = mean_95_cast_fp16)[name = tensor("sub_75_cast_fp16")]; + tensor square_61_cast_fp16 = square(x = sub_75_cast_fp16)[name = tensor("square_61_cast_fp16")]; + tensor reduce_mean_123_axes_0 = const()[name = tensor("reduce_mean_123_axes_0"), val = tensor([-1])]; + tensor reduce_mean_123_keep_dims_0 = const()[name = tensor("reduce_mean_123_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_123_cast_fp16 = reduce_mean(axes = reduce_mean_123_axes_0, keep_dims = reduce_mean_123_keep_dims_0, x = square_61_cast_fp16)[name = tensor("reduce_mean_123_cast_fp16")]; + tensor var_2705_to_fp16 = const()[name = tensor("op_2705_to_fp16"), val = tensor(0x1.1p-20)]; + tensor var_2706_cast_fp16 = add(x = reduce_mean_123_cast_fp16, y = var_2705_to_fp16)[name = tensor("op_2706_cast_fp16")]; + tensor var_2707_cast_fp16 = sqrt(x = var_2706_cast_fp16)[name = tensor("op_2707_cast_fp16")]; + tensor x_301_cast_fp16 = real_div(x = sub_75_cast_fp16, y = var_2707_cast_fp16)[name = tensor("x_301_cast_fp16")]; + tensor var_2709_cast_fp16 = mul(x = x_301_cast_fp16, y = flow_net_res_blocks_5_in_ln_weight_to_fp16)[name = tensor("op_2709_cast_fp16")]; + tensor x_303_cast_fp16 = add(x = var_2709_cast_fp16, y = flow_net_res_blocks_5_in_ln_bias_to_fp16)[name = tensor("x_303_cast_fp16")]; + tensor var_2711_promoted_to_fp16 = const()[name = tensor("op_2711_promoted_to_fp16"), val = tensor(0x1p+0)]; + tensor var_2712_cast_fp16 = add(x = var_2695_cast_fp16_1, y = var_2711_promoted_to_fp16)[name = tensor("op_2712_cast_fp16")]; + tensor var_2713_cast_fp16 = mul(x = x_303_cast_fp16, y = var_2712_cast_fp16)[name = tensor("op_2713_cast_fp16")]; + tensor input_465_cast_fp16 = add(x = var_2713_cast_fp16, y = var_2695_cast_fp16_0)[name = tensor("input_465_cast_fp16")]; + tensor linear_178_cast_fp16 = linear(bias = flow_net_res_blocks_5_mlp_0_bias_to_fp16, weight = flow_net_res_blocks_5_mlp_0_weight_to_fp16, x = input_465_cast_fp16)[name = tensor("linear_178_cast_fp16")]; + tensor input_469_cast_fp16 = silu(x = linear_178_cast_fp16)[name = tensor("input_469_cast_fp16")]; + tensor linear_179_cast_fp16 = linear(bias = flow_net_res_blocks_5_mlp_2_bias_to_fp16, weight = flow_net_res_blocks_5_mlp_2_weight_to_fp16, x = input_469_cast_fp16)[name = tensor("linear_179_cast_fp16")]; + tensor var_2724_cast_fp16 = mul(x = var_2695_cast_fp16_2, y = linear_179_cast_fp16)[name = tensor("op_2724_cast_fp16")]; + tensor x_305_cast_fp16 = add(x = x_299_cast_fp16, y = var_2724_cast_fp16)[name = tensor("x_305_cast_fp16")]; + tensor linear_180_cast_fp16 = linear(bias = flow_net_final_layer_adaLN_modulation_1_bias_to_fp16, weight = flow_net_final_layer_adaLN_modulation_1_weight_to_fp16, x = input_423_cast_fp16)[name = tensor("linear_180_cast_fp16")]; + tensor var_2733_split_sizes_0 = const()[name = tensor("op_2733_split_sizes_0"), val = tensor([512, 512])]; + tensor var_2733_axis_0 = const()[name = tensor("op_2733_axis_0"), val = tensor(-1)]; + tensor var_2733_cast_fp16_0, tensor var_2733_cast_fp16_1 = split(axis = var_2733_axis_0, split_sizes = var_2733_split_sizes_0, x = linear_180_cast_fp16)[name = tensor("op_2733_cast_fp16")]; + tensor mean_97_axes_0 = const()[name = tensor("mean_97_axes_0"), val = tensor([-1])]; + tensor mean_97_keep_dims_0 = const()[name = tensor("mean_97_keep_dims_0"), val = tensor(true)]; + tensor mean_97_cast_fp16 = reduce_mean(axes = mean_97_axes_0, keep_dims = mean_97_keep_dims_0, x = x_305_cast_fp16)[name = tensor("mean_97_cast_fp16")]; + tensor sub_76_cast_fp16 = sub(x = x_305_cast_fp16, y = mean_97_cast_fp16)[name = tensor("sub_76_cast_fp16")]; + tensor square_62_cast_fp16 = square(x = sub_76_cast_fp16)[name = tensor("square_62_cast_fp16")]; + tensor reduce_mean_125_axes_0 = const()[name = tensor("reduce_mean_125_axes_0"), val = tensor([-1])]; + tensor reduce_mean_125_keep_dims_0 = const()[name = tensor("reduce_mean_125_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_125_cast_fp16 = reduce_mean(axes = reduce_mean_125_axes_0, keep_dims = reduce_mean_125_keep_dims_0, x = square_62_cast_fp16)[name = tensor("reduce_mean_125_cast_fp16")]; + tensor var_2740_to_fp16 = const()[name = tensor("op_2740_to_fp16"), val = tensor(0x1.1p-20)]; + tensor var_2741_cast_fp16 = add(x = reduce_mean_125_cast_fp16, y = var_2740_to_fp16)[name = tensor("op_2741_cast_fp16")]; + tensor var_2742_cast_fp16 = sqrt(x = var_2741_cast_fp16)[name = tensor("op_2742_cast_fp16")]; + tensor x_307_cast_fp16 = real_div(x = sub_76_cast_fp16, y = var_2742_cast_fp16)[name = tensor("x_307_cast_fp16")]; + tensor var_2744_promoted_to_fp16 = const()[name = tensor("op_2744_promoted_to_fp16"), val = tensor(0x1p+0)]; + tensor var_2745_cast_fp16 = add(x = var_2733_cast_fp16_1, y = var_2744_promoted_to_fp16)[name = tensor("op_2745_cast_fp16")]; + tensor var_2746_cast_fp16 = mul(x = x_307_cast_fp16, y = var_2745_cast_fp16)[name = tensor("op_2746_cast_fp16")]; + tensor input_473_cast_fp16 = add(x = var_2746_cast_fp16, y = var_2733_cast_fp16_0)[name = tensor("input_473_cast_fp16")]; + tensor linear_181_cast_fp16 = linear(bias = flow_net_final_layer_linear_bias_to_fp16, weight = flow_net_final_layer_linear_weight_to_fp16, x = input_473_cast_fp16)[name = tensor("linear_181_cast_fp16")]; + tensor var_2757_to_fp16 = const()[name = tensor("op_2757_to_fp16"), val = tensor(0x1p-3)]; + tensor var_2758_cast_fp16 = mul(x = linear_181_cast_fp16, y = var_2757_to_fp16)[name = tensor("op_2758_cast_fp16")]; + tensor input_475_cast_fp16 = add(x = input_407_cast_fp16, y = var_2758_cast_fp16)[name = tensor("input_475_cast_fp16")]; + tensor linear_182_cast_fp16 = linear(bias = flow_net_input_proj_bias_to_fp16, weight = flow_net_input_proj_weight_to_fp16, x = input_475_cast_fp16)[name = tensor("linear_182_cast_fp16")]; + tensor input_479_to_fp16 = const()[name = tensor("input_479_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(19010432)))]; + tensor input_481_cast_fp16 = silu(x = input_479_to_fp16)[name = tensor("input_481_cast_fp16")]; + tensor linear_184_cast_fp16 = linear(bias = flow_net_time_embed_0_mlp_2_bias_to_fp16, weight = flow_net_time_embed_0_mlp_2_weight_to_fp16, x = input_481_cast_fp16)[name = tensor("linear_184_cast_fp16")]; + tensor reduce_mean_126_axes_0 = const()[name = tensor("reduce_mean_126_axes_0"), val = tensor([-1])]; + tensor reduce_mean_126_keep_dims_0 = const()[name = tensor("reduce_mean_126_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_126_cast_fp16 = reduce_mean(axes = reduce_mean_126_axes_0, keep_dims = reduce_mean_126_keep_dims_0, x = linear_184_cast_fp16)[name = tensor("reduce_mean_126_cast_fp16")]; + tensor sub_77_cast_fp16 = sub(x = linear_184_cast_fp16, y = reduce_mean_126_cast_fp16)[name = tensor("sub_77_cast_fp16")]; + tensor square_63_cast_fp16 = square(x = sub_77_cast_fp16)[name = tensor("square_63_cast_fp16")]; + tensor reduce_mean_127_axes_0 = const()[name = tensor("reduce_mean_127_axes_0"), val = tensor([-1])]; + tensor reduce_mean_127_keep_dims_0 = const()[name = tensor("reduce_mean_127_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_127_cast_fp16 = reduce_mean(axes = reduce_mean_127_axes_0, keep_dims = reduce_mean_127_keep_dims_0, x = square_63_cast_fp16)[name = tensor("reduce_mean_127_cast_fp16")]; + tensor real_div_14_to_fp16 = const()[name = tensor("real_div_14_to_fp16"), val = tensor(0x1.008p+0)]; + tensor mul_14_cast_fp16 = mul(x = reduce_mean_127_cast_fp16, y = real_div_14_to_fp16)[name = tensor("mul_14_cast_fp16")]; + tensor var_2824_to_fp16 = const()[name = tensor("op_2824_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_127_cast_fp16_0 = add(x = mul_14_cast_fp16, y = var_2824_to_fp16)[name = tensor("var_127_cast_fp16")]; + tensor var_2827_epsilon_0 = const()[name = tensor("op_2827_epsilon_0"), val = tensor(0x1.197998p-40)]; + tensor var_2827_cast_fp16 = rsqrt(epsilon = var_2827_epsilon_0, x = var_127_cast_fp16_0)[name = tensor("op_2827_cast_fp16")]; + tensor var_2828_cast_fp16 = mul(x = const_3_to_fp16, y = var_2827_cast_fp16)[name = tensor("op_2828_cast_fp16")]; + tensor var_2829_cast_fp16 = mul(x = linear_184_cast_fp16, y = var_2828_cast_fp16)[name = tensor("op_2829_cast_fp16")]; + tensor input_485_to_fp16 = const()[name = tensor("input_485_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(19011520)))]; + tensor input_487_cast_fp16 = silu(x = input_485_to_fp16)[name = tensor("input_487_cast_fp16")]; + tensor linear_186_cast_fp16 = linear(bias = flow_net_time_embed_1_mlp_2_bias_to_fp16, weight = flow_net_time_embed_1_mlp_2_weight_to_fp16, x = input_487_cast_fp16)[name = tensor("linear_186_cast_fp16")]; + tensor reduce_mean_128_axes_0 = const()[name = tensor("reduce_mean_128_axes_0"), val = tensor([-1])]; + tensor reduce_mean_128_keep_dims_0 = const()[name = tensor("reduce_mean_128_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_128_cast_fp16 = reduce_mean(axes = reduce_mean_128_axes_0, keep_dims = reduce_mean_128_keep_dims_0, x = linear_186_cast_fp16)[name = tensor("reduce_mean_128_cast_fp16")]; + tensor sub_79_cast_fp16 = sub(x = linear_186_cast_fp16, y = reduce_mean_128_cast_fp16)[name = tensor("sub_79_cast_fp16")]; + tensor square_64_cast_fp16 = square(x = sub_79_cast_fp16)[name = tensor("square_64_cast_fp16")]; + tensor reduce_mean_129_axes_0 = const()[name = tensor("reduce_mean_129_axes_0"), val = tensor([-1])]; + tensor reduce_mean_129_keep_dims_0 = const()[name = tensor("reduce_mean_129_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_129_cast_fp16 = reduce_mean(axes = reduce_mean_129_axes_0, keep_dims = reduce_mean_129_keep_dims_0, x = square_64_cast_fp16)[name = tensor("reduce_mean_129_cast_fp16")]; + tensor real_div_15_to_fp16 = const()[name = tensor("real_div_15_to_fp16"), val = tensor(0x1.008p+0)]; + tensor mul_15_cast_fp16 = mul(x = reduce_mean_129_cast_fp16, y = real_div_15_to_fp16)[name = tensor("mul_15_cast_fp16")]; + tensor var_2851_to_fp16 = const()[name = tensor("op_2851_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_129_cast_fp16 = add(x = mul_15_cast_fp16, y = var_2851_to_fp16)[name = tensor("var_129_cast_fp16")]; + tensor var_2854_epsilon_0 = const()[name = tensor("op_2854_epsilon_0"), val = tensor(0x1.197998p-40)]; + tensor var_2854_cast_fp16 = rsqrt(epsilon = var_2854_epsilon_0, x = var_129_cast_fp16)[name = tensor("op_2854_cast_fp16")]; + tensor var_2855_cast_fp16 = mul(x = const_5_to_fp16, y = var_2854_cast_fp16)[name = tensor("op_2855_cast_fp16")]; + tensor var_2856_cast_fp16 = mul(x = linear_186_cast_fp16, y = var_2855_cast_fp16)[name = tensor("op_2856_cast_fp16")]; + tensor var_2858_cast_fp16 = add(x = var_2829_cast_fp16, y = var_2856_cast_fp16)[name = tensor("op_2858_cast_fp16")]; + tensor _inversed_t_combined_y_0_to_fp16 = const()[name = tensor("_inversed_t_combined_y_0_to_fp16"), val = tensor(0x1p-1)]; + tensor _inversed_t_combined_cast_fp16 = mul(x = var_2858_cast_fp16, y = _inversed_t_combined_y_0_to_fp16)[name = tensor("_inversed_t_combined_cast_fp16")]; + tensor input_489_cast_fp16 = add(x = _inversed_t_combined_cast_fp16, y = linear_5_cast_fp16)[name = tensor("input_489_cast_fp16")]; + tensor input_491_cast_fp16 = silu(x = input_489_cast_fp16)[name = tensor("input_491_cast_fp16")]; + tensor linear_188_cast_fp16 = linear(bias = flow_net_res_blocks_0_adaLN_modulation_1_bias_to_fp16, weight = flow_net_res_blocks_0_adaLN_modulation_1_weight_to_fp16, x = input_491_cast_fp16)[name = tensor("linear_188_cast_fp16")]; + tensor var_2873_split_sizes_0 = const()[name = tensor("op_2873_split_sizes_0"), val = tensor([512, 512, 512])]; + tensor var_2873_axis_0 = const()[name = tensor("op_2873_axis_0"), val = tensor(-1)]; + tensor var_2873_cast_fp16_0, tensor var_2873_cast_fp16_1, tensor var_2873_cast_fp16_2 = split(axis = var_2873_axis_0, split_sizes = var_2873_split_sizes_0, x = linear_188_cast_fp16)[name = tensor("op_2873_cast_fp16")]; + tensor mean_99_axes_0 = const()[name = tensor("mean_99_axes_0"), val = tensor([-1])]; + tensor mean_99_keep_dims_0 = const()[name = tensor("mean_99_keep_dims_0"), val = tensor(true)]; + tensor mean_99_cast_fp16 = reduce_mean(axes = mean_99_axes_0, keep_dims = mean_99_keep_dims_0, x = linear_182_cast_fp16)[name = tensor("mean_99_cast_fp16")]; + tensor sub_81_cast_fp16 = sub(x = linear_182_cast_fp16, y = mean_99_cast_fp16)[name = tensor("sub_81_cast_fp16")]; + tensor square_65_cast_fp16 = square(x = sub_81_cast_fp16)[name = tensor("square_65_cast_fp16")]; + tensor reduce_mean_131_axes_0 = const()[name = tensor("reduce_mean_131_axes_0"), val = tensor([-1])]; + tensor reduce_mean_131_keep_dims_0 = const()[name = tensor("reduce_mean_131_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_131_cast_fp16 = reduce_mean(axes = reduce_mean_131_axes_0, keep_dims = reduce_mean_131_keep_dims_0, x = square_65_cast_fp16)[name = tensor("reduce_mean_131_cast_fp16")]; + tensor var_2883_to_fp16 = const()[name = tensor("op_2883_to_fp16"), val = tensor(0x1.1p-20)]; + tensor var_2884_cast_fp16 = add(x = reduce_mean_131_cast_fp16, y = var_2883_to_fp16)[name = tensor("op_2884_cast_fp16")]; + tensor var_2885_cast_fp16 = sqrt(x = var_2884_cast_fp16)[name = tensor("op_2885_cast_fp16")]; + tensor x_315_cast_fp16 = real_div(x = sub_81_cast_fp16, y = var_2885_cast_fp16)[name = tensor("x_315_cast_fp16")]; + tensor var_2887_cast_fp16 = mul(x = x_315_cast_fp16, y = flow_net_res_blocks_0_in_ln_weight_to_fp16)[name = tensor("op_2887_cast_fp16")]; + tensor x_317_cast_fp16 = add(x = var_2887_cast_fp16, y = flow_net_res_blocks_0_in_ln_bias_to_fp16)[name = tensor("x_317_cast_fp16")]; + tensor var_2889_promoted_to_fp16 = const()[name = tensor("op_2889_promoted_to_fp16"), val = tensor(0x1p+0)]; + tensor var_2890_cast_fp16 = add(x = var_2873_cast_fp16_1, y = var_2889_promoted_to_fp16)[name = tensor("op_2890_cast_fp16")]; + tensor var_2891_cast_fp16 = mul(x = x_317_cast_fp16, y = var_2890_cast_fp16)[name = tensor("op_2891_cast_fp16")]; + tensor input_493_cast_fp16 = add(x = var_2891_cast_fp16, y = var_2873_cast_fp16_0)[name = tensor("input_493_cast_fp16")]; + tensor linear_189_cast_fp16 = linear(bias = flow_net_res_blocks_0_mlp_0_bias_to_fp16, weight = flow_net_res_blocks_0_mlp_0_weight_to_fp16, x = input_493_cast_fp16)[name = tensor("linear_189_cast_fp16")]; + tensor input_497_cast_fp16 = silu(x = linear_189_cast_fp16)[name = tensor("input_497_cast_fp16")]; + tensor linear_190_cast_fp16 = linear(bias = flow_net_res_blocks_0_mlp_2_bias_to_fp16, weight = flow_net_res_blocks_0_mlp_2_weight_to_fp16, x = input_497_cast_fp16)[name = tensor("linear_190_cast_fp16")]; + tensor var_2902_cast_fp16 = mul(x = var_2873_cast_fp16_2, y = linear_190_cast_fp16)[name = tensor("op_2902_cast_fp16")]; + tensor x_319_cast_fp16 = add(x = linear_182_cast_fp16, y = var_2902_cast_fp16)[name = tensor("x_319_cast_fp16")]; + tensor linear_191_cast_fp16 = linear(bias = flow_net_res_blocks_1_adaLN_modulation_1_bias_to_fp16, weight = flow_net_res_blocks_1_adaLN_modulation_1_weight_to_fp16, x = input_491_cast_fp16)[name = tensor("linear_191_cast_fp16")]; + tensor var_2912_split_sizes_0 = const()[name = tensor("op_2912_split_sizes_0"), val = tensor([512, 512, 512])]; + tensor var_2912_axis_0 = const()[name = tensor("op_2912_axis_0"), val = tensor(-1)]; + tensor var_2912_cast_fp16_0, tensor var_2912_cast_fp16_1, tensor var_2912_cast_fp16_2 = split(axis = var_2912_axis_0, split_sizes = var_2912_split_sizes_0, x = linear_191_cast_fp16)[name = tensor("op_2912_cast_fp16")]; + tensor mean_101_axes_0 = const()[name = tensor("mean_101_axes_0"), val = tensor([-1])]; + tensor mean_101_keep_dims_0 = const()[name = tensor("mean_101_keep_dims_0"), val = tensor(true)]; + tensor mean_101_cast_fp16 = reduce_mean(axes = mean_101_axes_0, keep_dims = mean_101_keep_dims_0, x = x_319_cast_fp16)[name = tensor("mean_101_cast_fp16")]; + tensor sub_82_cast_fp16 = sub(x = x_319_cast_fp16, y = mean_101_cast_fp16)[name = tensor("sub_82_cast_fp16")]; + tensor square_66_cast_fp16 = square(x = sub_82_cast_fp16)[name = tensor("square_66_cast_fp16")]; + tensor reduce_mean_133_axes_0 = const()[name = tensor("reduce_mean_133_axes_0"), val = tensor([-1])]; + tensor reduce_mean_133_keep_dims_0 = const()[name = tensor("reduce_mean_133_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_133_cast_fp16 = reduce_mean(axes = reduce_mean_133_axes_0, keep_dims = reduce_mean_133_keep_dims_0, x = square_66_cast_fp16)[name = tensor("reduce_mean_133_cast_fp16")]; + tensor var_2922_to_fp16 = const()[name = tensor("op_2922_to_fp16"), val = tensor(0x1.1p-20)]; + tensor var_2923_cast_fp16 = add(x = reduce_mean_133_cast_fp16, y = var_2922_to_fp16)[name = tensor("op_2923_cast_fp16")]; + tensor var_2924_cast_fp16 = sqrt(x = var_2923_cast_fp16)[name = tensor("op_2924_cast_fp16")]; + tensor x_321_cast_fp16 = real_div(x = sub_82_cast_fp16, y = var_2924_cast_fp16)[name = tensor("x_321_cast_fp16")]; + tensor var_2926_cast_fp16 = mul(x = x_321_cast_fp16, y = flow_net_res_blocks_1_in_ln_weight_to_fp16)[name = tensor("op_2926_cast_fp16")]; + tensor x_323_cast_fp16 = add(x = var_2926_cast_fp16, y = flow_net_res_blocks_1_in_ln_bias_to_fp16)[name = tensor("x_323_cast_fp16")]; + tensor var_2928_promoted_to_fp16 = const()[name = tensor("op_2928_promoted_to_fp16"), val = tensor(0x1p+0)]; + tensor var_2929_cast_fp16 = add(x = var_2912_cast_fp16_1, y = var_2928_promoted_to_fp16)[name = tensor("op_2929_cast_fp16")]; + tensor var_2930_cast_fp16 = mul(x = x_323_cast_fp16, y = var_2929_cast_fp16)[name = tensor("op_2930_cast_fp16")]; + tensor input_501_cast_fp16 = add(x = var_2930_cast_fp16, y = var_2912_cast_fp16_0)[name = tensor("input_501_cast_fp16")]; + tensor linear_192_cast_fp16 = linear(bias = flow_net_res_blocks_1_mlp_0_bias_to_fp16, weight = flow_net_res_blocks_1_mlp_0_weight_to_fp16, x = input_501_cast_fp16)[name = tensor("linear_192_cast_fp16")]; + tensor input_505_cast_fp16 = silu(x = linear_192_cast_fp16)[name = tensor("input_505_cast_fp16")]; + tensor linear_193_cast_fp16 = linear(bias = flow_net_res_blocks_1_mlp_2_bias_to_fp16, weight = flow_net_res_blocks_1_mlp_2_weight_to_fp16, x = input_505_cast_fp16)[name = tensor("linear_193_cast_fp16")]; + tensor var_2941_cast_fp16 = mul(x = var_2912_cast_fp16_2, y = linear_193_cast_fp16)[name = tensor("op_2941_cast_fp16")]; + tensor x_325_cast_fp16 = add(x = x_319_cast_fp16, y = var_2941_cast_fp16)[name = tensor("x_325_cast_fp16")]; + tensor linear_194_cast_fp16 = linear(bias = flow_net_res_blocks_2_adaLN_modulation_1_bias_to_fp16, weight = flow_net_res_blocks_2_adaLN_modulation_1_weight_to_fp16, x = input_491_cast_fp16)[name = tensor("linear_194_cast_fp16")]; + tensor var_2951_split_sizes_0 = const()[name = tensor("op_2951_split_sizes_0"), val = tensor([512, 512, 512])]; + tensor var_2951_axis_0 = const()[name = tensor("op_2951_axis_0"), val = tensor(-1)]; + tensor var_2951_cast_fp16_0, tensor var_2951_cast_fp16_1, tensor var_2951_cast_fp16_2 = split(axis = var_2951_axis_0, split_sizes = var_2951_split_sizes_0, x = linear_194_cast_fp16)[name = tensor("op_2951_cast_fp16")]; + tensor mean_103_axes_0 = const()[name = tensor("mean_103_axes_0"), val = tensor([-1])]; + tensor mean_103_keep_dims_0 = const()[name = tensor("mean_103_keep_dims_0"), val = tensor(true)]; + tensor mean_103_cast_fp16 = reduce_mean(axes = mean_103_axes_0, keep_dims = mean_103_keep_dims_0, x = x_325_cast_fp16)[name = tensor("mean_103_cast_fp16")]; + tensor sub_83_cast_fp16 = sub(x = x_325_cast_fp16, y = mean_103_cast_fp16)[name = tensor("sub_83_cast_fp16")]; + tensor square_67_cast_fp16 = square(x = sub_83_cast_fp16)[name = tensor("square_67_cast_fp16")]; + tensor reduce_mean_135_axes_0 = const()[name = tensor("reduce_mean_135_axes_0"), val = tensor([-1])]; + tensor reduce_mean_135_keep_dims_0 = const()[name = tensor("reduce_mean_135_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_135_cast_fp16 = reduce_mean(axes = reduce_mean_135_axes_0, keep_dims = reduce_mean_135_keep_dims_0, x = square_67_cast_fp16)[name = tensor("reduce_mean_135_cast_fp16")]; + tensor var_2961_to_fp16 = const()[name = tensor("op_2961_to_fp16"), val = tensor(0x1.1p-20)]; + tensor var_2962_cast_fp16 = add(x = reduce_mean_135_cast_fp16, y = var_2961_to_fp16)[name = tensor("op_2962_cast_fp16")]; + tensor var_2963_cast_fp16 = sqrt(x = var_2962_cast_fp16)[name = tensor("op_2963_cast_fp16")]; + tensor x_327_cast_fp16 = real_div(x = sub_83_cast_fp16, y = var_2963_cast_fp16)[name = tensor("x_327_cast_fp16")]; + tensor var_2965_cast_fp16 = mul(x = x_327_cast_fp16, y = flow_net_res_blocks_2_in_ln_weight_to_fp16)[name = tensor("op_2965_cast_fp16")]; + tensor x_329_cast_fp16 = add(x = var_2965_cast_fp16, y = flow_net_res_blocks_2_in_ln_bias_to_fp16)[name = tensor("x_329_cast_fp16")]; + tensor var_2967_promoted_to_fp16 = const()[name = tensor("op_2967_promoted_to_fp16"), val = tensor(0x1p+0)]; + tensor var_2968_cast_fp16 = add(x = var_2951_cast_fp16_1, y = var_2967_promoted_to_fp16)[name = tensor("op_2968_cast_fp16")]; + tensor var_2969_cast_fp16 = mul(x = x_329_cast_fp16, y = var_2968_cast_fp16)[name = tensor("op_2969_cast_fp16")]; + tensor input_509_cast_fp16 = add(x = var_2969_cast_fp16, y = var_2951_cast_fp16_0)[name = tensor("input_509_cast_fp16")]; + tensor linear_195_cast_fp16 = linear(bias = flow_net_res_blocks_2_mlp_0_bias_to_fp16, weight = flow_net_res_blocks_2_mlp_0_weight_to_fp16, x = input_509_cast_fp16)[name = tensor("linear_195_cast_fp16")]; + tensor input_513_cast_fp16 = silu(x = linear_195_cast_fp16)[name = tensor("input_513_cast_fp16")]; + tensor linear_196_cast_fp16 = linear(bias = flow_net_res_blocks_2_mlp_2_bias_to_fp16, weight = flow_net_res_blocks_2_mlp_2_weight_to_fp16, x = input_513_cast_fp16)[name = tensor("linear_196_cast_fp16")]; + tensor var_2980_cast_fp16 = mul(x = var_2951_cast_fp16_2, y = linear_196_cast_fp16)[name = tensor("op_2980_cast_fp16")]; + tensor x_331_cast_fp16 = add(x = x_325_cast_fp16, y = var_2980_cast_fp16)[name = tensor("x_331_cast_fp16")]; + tensor linear_197_cast_fp16 = linear(bias = flow_net_res_blocks_3_adaLN_modulation_1_bias_to_fp16, weight = flow_net_res_blocks_3_adaLN_modulation_1_weight_to_fp16, x = input_491_cast_fp16)[name = tensor("linear_197_cast_fp16")]; + tensor var_2990_split_sizes_0 = const()[name = tensor("op_2990_split_sizes_0"), val = tensor([512, 512, 512])]; + tensor var_2990_axis_0 = const()[name = tensor("op_2990_axis_0"), val = tensor(-1)]; + tensor var_2990_cast_fp16_0, tensor var_2990_cast_fp16_1, tensor var_2990_cast_fp16_2 = split(axis = var_2990_axis_0, split_sizes = var_2990_split_sizes_0, x = linear_197_cast_fp16)[name = tensor("op_2990_cast_fp16")]; + tensor mean_105_axes_0 = const()[name = tensor("mean_105_axes_0"), val = tensor([-1])]; + tensor mean_105_keep_dims_0 = const()[name = tensor("mean_105_keep_dims_0"), val = tensor(true)]; + tensor mean_105_cast_fp16 = reduce_mean(axes = mean_105_axes_0, keep_dims = mean_105_keep_dims_0, x = x_331_cast_fp16)[name = tensor("mean_105_cast_fp16")]; + tensor sub_84_cast_fp16 = sub(x = x_331_cast_fp16, y = mean_105_cast_fp16)[name = tensor("sub_84_cast_fp16")]; + tensor square_68_cast_fp16 = square(x = sub_84_cast_fp16)[name = tensor("square_68_cast_fp16")]; + tensor reduce_mean_137_axes_0 = const()[name = tensor("reduce_mean_137_axes_0"), val = tensor([-1])]; + tensor reduce_mean_137_keep_dims_0 = const()[name = tensor("reduce_mean_137_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_137_cast_fp16 = reduce_mean(axes = reduce_mean_137_axes_0, keep_dims = reduce_mean_137_keep_dims_0, x = square_68_cast_fp16)[name = tensor("reduce_mean_137_cast_fp16")]; + tensor var_3000_to_fp16 = const()[name = tensor("op_3000_to_fp16"), val = tensor(0x1.1p-20)]; + tensor var_3001_cast_fp16 = add(x = reduce_mean_137_cast_fp16, y = var_3000_to_fp16)[name = tensor("op_3001_cast_fp16")]; + tensor var_3002_cast_fp16 = sqrt(x = var_3001_cast_fp16)[name = tensor("op_3002_cast_fp16")]; + tensor x_333_cast_fp16 = real_div(x = sub_84_cast_fp16, y = var_3002_cast_fp16)[name = tensor("x_333_cast_fp16")]; + tensor var_3004_cast_fp16 = mul(x = x_333_cast_fp16, y = flow_net_res_blocks_3_in_ln_weight_to_fp16)[name = tensor("op_3004_cast_fp16")]; + tensor x_335_cast_fp16 = add(x = var_3004_cast_fp16, y = flow_net_res_blocks_3_in_ln_bias_to_fp16)[name = tensor("x_335_cast_fp16")]; + tensor var_3006_promoted_to_fp16 = const()[name = tensor("op_3006_promoted_to_fp16"), val = tensor(0x1p+0)]; + tensor var_3007_cast_fp16 = add(x = var_2990_cast_fp16_1, y = var_3006_promoted_to_fp16)[name = tensor("op_3007_cast_fp16")]; + tensor var_3008_cast_fp16 = mul(x = x_335_cast_fp16, y = var_3007_cast_fp16)[name = tensor("op_3008_cast_fp16")]; + tensor input_517_cast_fp16 = add(x = var_3008_cast_fp16, y = var_2990_cast_fp16_0)[name = tensor("input_517_cast_fp16")]; + tensor linear_198_cast_fp16 = linear(bias = flow_net_res_blocks_3_mlp_0_bias_to_fp16, weight = flow_net_res_blocks_3_mlp_0_weight_to_fp16, x = input_517_cast_fp16)[name = tensor("linear_198_cast_fp16")]; + tensor input_521_cast_fp16 = silu(x = linear_198_cast_fp16)[name = tensor("input_521_cast_fp16")]; + tensor linear_199_cast_fp16 = linear(bias = flow_net_res_blocks_3_mlp_2_bias_to_fp16, weight = flow_net_res_blocks_3_mlp_2_weight_to_fp16, x = input_521_cast_fp16)[name = tensor("linear_199_cast_fp16")]; + tensor var_3019_cast_fp16 = mul(x = var_2990_cast_fp16_2, y = linear_199_cast_fp16)[name = tensor("op_3019_cast_fp16")]; + tensor x_337_cast_fp16 = add(x = x_331_cast_fp16, y = var_3019_cast_fp16)[name = tensor("x_337_cast_fp16")]; + tensor linear_200_cast_fp16 = linear(bias = flow_net_res_blocks_4_adaLN_modulation_1_bias_to_fp16, weight = flow_net_res_blocks_4_adaLN_modulation_1_weight_to_fp16, x = input_491_cast_fp16)[name = tensor("linear_200_cast_fp16")]; + tensor var_3029_split_sizes_0 = const()[name = tensor("op_3029_split_sizes_0"), val = tensor([512, 512, 512])]; + tensor var_3029_axis_0 = const()[name = tensor("op_3029_axis_0"), val = tensor(-1)]; + tensor var_3029_cast_fp16_0, tensor var_3029_cast_fp16_1, tensor var_3029_cast_fp16_2 = split(axis = var_3029_axis_0, split_sizes = var_3029_split_sizes_0, x = linear_200_cast_fp16)[name = tensor("op_3029_cast_fp16")]; + tensor mean_107_axes_0 = const()[name = tensor("mean_107_axes_0"), val = tensor([-1])]; + tensor mean_107_keep_dims_0 = const()[name = tensor("mean_107_keep_dims_0"), val = tensor(true)]; + tensor mean_107_cast_fp16 = reduce_mean(axes = mean_107_axes_0, keep_dims = mean_107_keep_dims_0, x = x_337_cast_fp16)[name = tensor("mean_107_cast_fp16")]; + tensor sub_85_cast_fp16 = sub(x = x_337_cast_fp16, y = mean_107_cast_fp16)[name = tensor("sub_85_cast_fp16")]; + tensor square_69_cast_fp16 = square(x = sub_85_cast_fp16)[name = tensor("square_69_cast_fp16")]; + tensor reduce_mean_139_axes_0 = const()[name = tensor("reduce_mean_139_axes_0"), val = tensor([-1])]; + tensor reduce_mean_139_keep_dims_0 = const()[name = tensor("reduce_mean_139_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_139_cast_fp16 = reduce_mean(axes = reduce_mean_139_axes_0, keep_dims = reduce_mean_139_keep_dims_0, x = square_69_cast_fp16)[name = tensor("reduce_mean_139_cast_fp16")]; + tensor var_3039_to_fp16 = const()[name = tensor("op_3039_to_fp16"), val = tensor(0x1.1p-20)]; + tensor var_3040_cast_fp16 = add(x = reduce_mean_139_cast_fp16, y = var_3039_to_fp16)[name = tensor("op_3040_cast_fp16")]; + tensor var_3041_cast_fp16 = sqrt(x = var_3040_cast_fp16)[name = tensor("op_3041_cast_fp16")]; + tensor x_339_cast_fp16 = real_div(x = sub_85_cast_fp16, y = var_3041_cast_fp16)[name = tensor("x_339_cast_fp16")]; + tensor var_3043_cast_fp16 = mul(x = x_339_cast_fp16, y = flow_net_res_blocks_4_in_ln_weight_to_fp16)[name = tensor("op_3043_cast_fp16")]; + tensor x_341_cast_fp16 = add(x = var_3043_cast_fp16, y = flow_net_res_blocks_4_in_ln_bias_to_fp16)[name = tensor("x_341_cast_fp16")]; + tensor var_3045_promoted_to_fp16 = const()[name = tensor("op_3045_promoted_to_fp16"), val = tensor(0x1p+0)]; + tensor var_3046_cast_fp16 = add(x = var_3029_cast_fp16_1, y = var_3045_promoted_to_fp16)[name = tensor("op_3046_cast_fp16")]; + tensor var_3047_cast_fp16 = mul(x = x_341_cast_fp16, y = var_3046_cast_fp16)[name = tensor("op_3047_cast_fp16")]; + tensor input_525_cast_fp16 = add(x = var_3047_cast_fp16, y = var_3029_cast_fp16_0)[name = tensor("input_525_cast_fp16")]; + tensor linear_201_cast_fp16 = linear(bias = flow_net_res_blocks_4_mlp_0_bias_to_fp16, weight = flow_net_res_blocks_4_mlp_0_weight_to_fp16, x = input_525_cast_fp16)[name = tensor("linear_201_cast_fp16")]; + tensor input_529_cast_fp16 = silu(x = linear_201_cast_fp16)[name = tensor("input_529_cast_fp16")]; + tensor linear_202_cast_fp16 = linear(bias = flow_net_res_blocks_4_mlp_2_bias_to_fp16, weight = flow_net_res_blocks_4_mlp_2_weight_to_fp16, x = input_529_cast_fp16)[name = tensor("linear_202_cast_fp16")]; + tensor var_3058_cast_fp16 = mul(x = var_3029_cast_fp16_2, y = linear_202_cast_fp16)[name = tensor("op_3058_cast_fp16")]; + tensor x_343_cast_fp16 = add(x = x_337_cast_fp16, y = var_3058_cast_fp16)[name = tensor("x_343_cast_fp16")]; + tensor linear_203_cast_fp16 = linear(bias = flow_net_res_blocks_5_adaLN_modulation_1_bias_to_fp16, weight = flow_net_res_blocks_5_adaLN_modulation_1_weight_to_fp16, x = input_491_cast_fp16)[name = tensor("linear_203_cast_fp16")]; + tensor var_3068_split_sizes_0 = const()[name = tensor("op_3068_split_sizes_0"), val = tensor([512, 512, 512])]; + tensor var_3068_axis_0 = const()[name = tensor("op_3068_axis_0"), val = tensor(-1)]; + tensor var_3068_cast_fp16_0, tensor var_3068_cast_fp16_1, tensor var_3068_cast_fp16_2 = split(axis = var_3068_axis_0, split_sizes = var_3068_split_sizes_0, x = linear_203_cast_fp16)[name = tensor("op_3068_cast_fp16")]; + tensor mean_109_axes_0 = const()[name = tensor("mean_109_axes_0"), val = tensor([-1])]; + tensor mean_109_keep_dims_0 = const()[name = tensor("mean_109_keep_dims_0"), val = tensor(true)]; + tensor mean_109_cast_fp16 = reduce_mean(axes = mean_109_axes_0, keep_dims = mean_109_keep_dims_0, x = x_343_cast_fp16)[name = tensor("mean_109_cast_fp16")]; + tensor sub_86_cast_fp16 = sub(x = x_343_cast_fp16, y = mean_109_cast_fp16)[name = tensor("sub_86_cast_fp16")]; + tensor square_70_cast_fp16 = square(x = sub_86_cast_fp16)[name = tensor("square_70_cast_fp16")]; + tensor reduce_mean_141_axes_0 = const()[name = tensor("reduce_mean_141_axes_0"), val = tensor([-1])]; + tensor reduce_mean_141_keep_dims_0 = const()[name = tensor("reduce_mean_141_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_141_cast_fp16 = reduce_mean(axes = reduce_mean_141_axes_0, keep_dims = reduce_mean_141_keep_dims_0, x = square_70_cast_fp16)[name = tensor("reduce_mean_141_cast_fp16")]; + tensor var_3078_to_fp16 = const()[name = tensor("op_3078_to_fp16"), val = tensor(0x1.1p-20)]; + tensor var_3079_cast_fp16 = add(x = reduce_mean_141_cast_fp16, y = var_3078_to_fp16)[name = tensor("op_3079_cast_fp16")]; + tensor var_3080_cast_fp16 = sqrt(x = var_3079_cast_fp16)[name = tensor("op_3080_cast_fp16")]; + tensor x_345_cast_fp16 = real_div(x = sub_86_cast_fp16, y = var_3080_cast_fp16)[name = tensor("x_345_cast_fp16")]; + tensor var_3082_cast_fp16 = mul(x = x_345_cast_fp16, y = flow_net_res_blocks_5_in_ln_weight_to_fp16)[name = tensor("op_3082_cast_fp16")]; + tensor x_347_cast_fp16 = add(x = var_3082_cast_fp16, y = flow_net_res_blocks_5_in_ln_bias_to_fp16)[name = tensor("x_347_cast_fp16")]; + tensor var_3084_promoted_to_fp16 = const()[name = tensor("op_3084_promoted_to_fp16"), val = tensor(0x1p+0)]; + tensor var_3085_cast_fp16 = add(x = var_3068_cast_fp16_1, y = var_3084_promoted_to_fp16)[name = tensor("op_3085_cast_fp16")]; + tensor var_3086_cast_fp16 = mul(x = x_347_cast_fp16, y = var_3085_cast_fp16)[name = tensor("op_3086_cast_fp16")]; + tensor input_533_cast_fp16 = add(x = var_3086_cast_fp16, y = var_3068_cast_fp16_0)[name = tensor("input_533_cast_fp16")]; + tensor linear_204_cast_fp16 = linear(bias = flow_net_res_blocks_5_mlp_0_bias_to_fp16, weight = flow_net_res_blocks_5_mlp_0_weight_to_fp16, x = input_533_cast_fp16)[name = tensor("linear_204_cast_fp16")]; + tensor input_537_cast_fp16 = silu(x = linear_204_cast_fp16)[name = tensor("input_537_cast_fp16")]; + tensor linear_205_cast_fp16 = linear(bias = flow_net_res_blocks_5_mlp_2_bias_to_fp16, weight = flow_net_res_blocks_5_mlp_2_weight_to_fp16, x = input_537_cast_fp16)[name = tensor("linear_205_cast_fp16")]; + tensor var_3097_cast_fp16 = mul(x = var_3068_cast_fp16_2, y = linear_205_cast_fp16)[name = tensor("op_3097_cast_fp16")]; + tensor x_349_cast_fp16 = add(x = x_343_cast_fp16, y = var_3097_cast_fp16)[name = tensor("x_349_cast_fp16")]; + tensor linear_206_cast_fp16 = linear(bias = flow_net_final_layer_adaLN_modulation_1_bias_to_fp16, weight = flow_net_final_layer_adaLN_modulation_1_weight_to_fp16, x = input_491_cast_fp16)[name = tensor("linear_206_cast_fp16")]; + tensor var_3106_split_sizes_0 = const()[name = tensor("op_3106_split_sizes_0"), val = tensor([512, 512])]; + tensor var_3106_axis_0 = const()[name = tensor("op_3106_axis_0"), val = tensor(-1)]; + tensor var_3106_cast_fp16_0, tensor var_3106_cast_fp16_1 = split(axis = var_3106_axis_0, split_sizes = var_3106_split_sizes_0, x = linear_206_cast_fp16)[name = tensor("op_3106_cast_fp16")]; + tensor mean_axes_0 = const()[name = tensor("mean_axes_0"), val = tensor([-1])]; + tensor mean_keep_dims_0 = const()[name = tensor("mean_keep_dims_0"), val = tensor(true)]; + tensor mean_cast_fp16 = reduce_mean(axes = mean_axes_0, keep_dims = mean_keep_dims_0, x = x_349_cast_fp16)[name = tensor("mean_cast_fp16")]; + tensor sub_87_cast_fp16 = sub(x = x_349_cast_fp16, y = mean_cast_fp16)[name = tensor("sub_87_cast_fp16")]; + tensor square_71_cast_fp16 = square(x = sub_87_cast_fp16)[name = tensor("square_71_cast_fp16")]; + tensor reduce_mean_143_axes_0 = const()[name = tensor("reduce_mean_143_axes_0"), val = tensor([-1])]; + tensor reduce_mean_143_keep_dims_0 = const()[name = tensor("reduce_mean_143_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_143_cast_fp16 = reduce_mean(axes = reduce_mean_143_axes_0, keep_dims = reduce_mean_143_keep_dims_0, x = square_71_cast_fp16)[name = tensor("reduce_mean_143_cast_fp16")]; + tensor var_3113_to_fp16 = const()[name = tensor("op_3113_to_fp16"), val = tensor(0x1.1p-20)]; + tensor var_3114_cast_fp16 = add(x = reduce_mean_143_cast_fp16, y = var_3113_to_fp16)[name = tensor("op_3114_cast_fp16")]; + tensor var_3115_cast_fp16 = sqrt(x = var_3114_cast_fp16)[name = tensor("op_3115_cast_fp16")]; + tensor x_cast_fp16 = real_div(x = sub_87_cast_fp16, y = var_3115_cast_fp16)[name = tensor("x_cast_fp16")]; + tensor var_3117_promoted_to_fp16 = const()[name = tensor("op_3117_promoted_to_fp16"), val = tensor(0x1p+0)]; + tensor var_3118_cast_fp16 = add(x = var_3106_cast_fp16_1, y = var_3117_promoted_to_fp16)[name = tensor("op_3118_cast_fp16")]; + tensor var_3119_cast_fp16 = mul(x = x_cast_fp16, y = var_3118_cast_fp16)[name = tensor("op_3119_cast_fp16")]; + tensor input_cast_fp16 = add(x = var_3119_cast_fp16, y = var_3106_cast_fp16_0)[name = tensor("input_cast_fp16")]; + tensor linear_207_cast_fp16 = linear(bias = flow_net_final_layer_linear_bias_to_fp16, weight = flow_net_final_layer_linear_weight_to_fp16, x = input_cast_fp16)[name = tensor("linear_207_cast_fp16")]; + tensor var_3124_to_fp16 = const()[name = tensor("op_3124_to_fp16"), val = tensor(0x1p-3)]; + tensor var_3125_cast_fp16 = mul(x = linear_207_cast_fp16, y = var_3124_to_fp16)[name = tensor("op_3125_cast_fp16")]; + tensor var_3127_cast_fp16 = add(x = input_475_cast_fp16, y = var_3125_cast_fp16)[name = tensor("op_3127_cast_fp16")]; + tensor var_3127_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_3127_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor latent_final = cast(dtype = var_3127_cast_fp16_to_fp32_dtype_0, x = var_3127_cast_fp16)[name = tensor("cast_144")]; + } -> (latent_final); +} \ No newline at end of file diff --git a/v2.1/italian_24l/flow_decoder_fused.mlmodelc/weights/weight.bin b/v2.1/italian_24l/flow_decoder_fused.mlmodelc/weights/weight.bin new file mode 100644 index 0000000000000000000000000000000000000000..8997311285821476f5de08d0ec031ead9ba1afa4 --- /dev/null +++ b/v2.1/italian_24l/flow_decoder_fused.mlmodelc/weights/weight.bin @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:2dba60db47a97f0d1989398ab403f4992fc243f1043965be1124035e821ee0e2 +size 19012608 diff --git a/v2.1/italian_24l/flow_decoder_fused.mlpackage/Data/com.apple.CoreML/model.mlmodel b/v2.1/italian_24l/flow_decoder_fused.mlpackage/Data/com.apple.CoreML/model.mlmodel new file mode 100644 index 0000000000000000000000000000000000000000..7ecb15155ec829c5320fcae6d6871f48031f65fd --- /dev/null +++ b/v2.1/italian_24l/flow_decoder_fused.mlpackage/Data/com.apple.CoreML/model.mlmodel @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:683f6c745313a4798084123615e220acd8b0835c1d1dd2fa7dd0e8e2459fb75e +size 301733 diff --git a/v2.1/italian_24l/flow_decoder_fused.mlpackage/Data/com.apple.CoreML/weights/weight.bin b/v2.1/italian_24l/flow_decoder_fused.mlpackage/Data/com.apple.CoreML/weights/weight.bin new file mode 100644 index 0000000000000000000000000000000000000000..8997311285821476f5de08d0ec031ead9ba1afa4 --- /dev/null +++ b/v2.1/italian_24l/flow_decoder_fused.mlpackage/Data/com.apple.CoreML/weights/weight.bin @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:2dba60db47a97f0d1989398ab403f4992fc243f1043965be1124035e821ee0e2 +size 19012608 diff --git a/v2.1/italian_24l/flow_decoder_fused.mlpackage/Manifest.json b/v2.1/italian_24l/flow_decoder_fused.mlpackage/Manifest.json new file mode 100644 index 0000000000000000000000000000000000000000..d265bd0511287a5de9bf5fb6521250c7d9344f8a --- /dev/null +++ b/v2.1/italian_24l/flow_decoder_fused.mlpackage/Manifest.json @@ -0,0 +1,18 @@ +{ + "fileFormatVersion": "1.0.0", + "itemInfoEntries": { + "2A40C91F-BE37-42BC-9D45-3E5D2096C6B5": { + "author": "com.apple.CoreML", + "description": "CoreML Model Specification", + "name": "model.mlmodel", + "path": "com.apple.CoreML/model.mlmodel" + }, + "AA7F82F6-83BA-431F-9907-B9B70BE8BEA9": { + "author": "com.apple.CoreML", + "description": "CoreML Model Weights", + "name": "weights", + "path": "com.apple.CoreML/weights" + } + }, + "rootModelIdentifier": "2A40C91F-BE37-42BC-9D45-3E5D2096C6B5" +} diff --git a/v2.1/italian_24l/flowlm_step.mlmodelc/analytics/coremldata.bin b/v2.1/italian_24l/flowlm_step.mlmodelc/analytics/coremldata.bin new file mode 100644 index 0000000000000000000000000000000000000000..53b5aa45378c6d1ee620b6d16b002e2530c9d510 --- /dev/null +++ b/v2.1/italian_24l/flowlm_step.mlmodelc/analytics/coremldata.bin @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:db131e2142dc728bbd206b388311f9dcee40acaee36ae4868dacac5179981e1d +size 243 diff --git a/v2.1/italian_24l/flowlm_step.mlmodelc/coremldata.bin b/v2.1/italian_24l/flowlm_step.mlmodelc/coremldata.bin new file mode 100644 index 0000000000000000000000000000000000000000..2c78027057a61eb0d7b69438a08c973b4a93ce34 --- /dev/null +++ b/v2.1/italian_24l/flowlm_step.mlmodelc/coremldata.bin @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:1c8be9511a8cf574698c1bf1eb0b94289c1f64635299ac0496a6fb19e6ed4fc3 +size 3559 diff --git a/v2.1/italian_24l/flowlm_step.mlmodelc/model.mil b/v2.1/italian_24l/flowlm_step.mlmodelc/model.mil new file mode 100644 index 0000000000000000000000000000000000000000..644c9fd8d551ba10454b84375e5f87007dc0ff27 --- /dev/null +++ b/v2.1/italian_24l/flowlm_step.mlmodelc/model.mil @@ -0,0 +1,5337 @@ +program(1.0) +[buildInfo = dict, tensor>({{"coremlc-component-MIL", "3520.4.1"}, {"coremlc-version", "3520.5.1"}, {"coremltools-component-torch", "2.12.0"}, {"coremltools-source-dialect", "TorchScript"}, {"coremltools-version", "9.0"}})] +{ + func main(tensor bos_emb, tensor cache0, tensor cache1, tensor cache10, tensor cache11, tensor cache12, tensor cache13, tensor cache14, tensor cache15, tensor cache16, tensor cache17, tensor cache18, tensor cache19, tensor cache2, tensor cache20, tensor cache21, tensor cache22, tensor cache23, tensor cache3, tensor cache4, tensor cache5, tensor cache6, tensor cache7, tensor cache8, tensor cache9, tensor position0, tensor position1, tensor position10, tensor position11, tensor position12, tensor position13, tensor position14, tensor position15, tensor position16, tensor position17, tensor position18, tensor position19, tensor position2, tensor position20, tensor position21, tensor position22, tensor position23, tensor position3, tensor position4, tensor position5, tensor position6, tensor position7, tensor position8, tensor position9, tensor sequence) { + tensor sequence_to_fp16_dtype_0 = const()[name = tensor("sequence_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor sequence_to_fp16 = cast(dtype = sequence_to_fp16_dtype_0, x = sequence)[name = tensor("cast_555")]; + tensor var_198_cast_fp16 = not_equal(x = sequence_to_fp16, y = sequence_to_fp16)[name = tensor("op_198_cast_fp16")]; + tensor expand_dims_0_axes_0 = const()[name = tensor("expand_dims_0_axes_0"), val = tensor([0, 1])]; + tensor bos_emb_to_fp16_dtype_0 = const()[name = tensor("bos_emb_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor bos_emb_to_fp16 = cast(dtype = bos_emb_to_fp16_dtype_0, x = bos_emb)[name = tensor("cast_554")]; + tensor expand_dims_0_cast_fp16 = expand_dims(axes = expand_dims_0_axes_0, x = bos_emb_to_fp16)[name = tensor("expand_dims_0_cast_fp16")]; + tensor input_1_cast_fp16 = select(a = expand_dims_0_cast_fp16, b = sequence_to_fp16, cond = var_198_cast_fp16)[name = tensor("input_1_cast_fp16")]; + tensor input_linear_weight_to_fp16 = const()[name = tensor("input_linear_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(64)))]; + tensor linear_0_bias_0_to_fp16 = const()[name = tensor("linear_0_bias_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(65664)))]; + tensor linear_0_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = input_linear_weight_to_fp16, x = input_1_cast_fp16)[name = tensor("linear_0_cast_fp16")]; + tensor input_5_axes_0 = const()[name = tensor("input_5_axes_0"), val = tensor([-1])]; + tensor norm0_1_weight_to_fp16 = const()[name = tensor("norm0_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(67776)))]; + tensor norm0_1_bias_to_fp16 = const()[name = tensor("norm0_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(69888)))]; + tensor var_204_to_fp16 = const()[name = tensor("op_204_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_5_cast_fp16 = layer_norm(axes = input_5_axes_0, beta = norm0_1_bias_to_fp16, epsilon = var_204_to_fp16, gamma = norm0_1_weight_to_fp16, x = linear_0_cast_fp16)[name = tensor("input_5_cast_fp16")]; + tensor attn0_in_proj_weight_to_fp16 = const()[name = tensor("attn0_in_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(72000)))]; + tensor linear_1_bias_0_to_fp16 = const()[name = tensor("linear_1_bias_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(6363520)))]; + tensor linear_1_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = attn0_in_proj_weight_to_fp16, x = input_5_cast_fp16)[name = tensor("linear_1_cast_fp16")]; + tensor var_218 = const()[name = tensor("op_218"), val = tensor([1, 1, 3, 16, 64])]; + tensor qkv_1_cast_fp16 = reshape(shape = var_218, x = linear_1_cast_fp16)[name = tensor("qkv_1_cast_fp16")]; + tensor q_1_begin_0 = const()[name = tensor("q_1_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor q_1_end_0 = const()[name = tensor("q_1_end_0"), val = tensor([1, 1, 1, 16, 64])]; + tensor q_1_end_mask_0 = const()[name = tensor("q_1_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor q_1_squeeze_mask_0 = const()[name = tensor("q_1_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor q_1_cast_fp16 = slice_by_index(begin = q_1_begin_0, end = q_1_end_0, end_mask = q_1_end_mask_0, squeeze_mask = q_1_squeeze_mask_0, x = qkv_1_cast_fp16)[name = tensor("q_1_cast_fp16")]; + tensor k_1_begin_0 = const()[name = tensor("k_1_begin_0"), val = tensor([0, 0, 1, 0, 0])]; + tensor k_1_end_0 = const()[name = tensor("k_1_end_0"), val = tensor([1, 1, 2, 16, 64])]; + tensor k_1_end_mask_0 = const()[name = tensor("k_1_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor k_1_squeeze_mask_0 = const()[name = tensor("k_1_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor k_1_cast_fp16 = slice_by_index(begin = k_1_begin_0, end = k_1_end_0, end_mask = k_1_end_mask_0, squeeze_mask = k_1_squeeze_mask_0, x = qkv_1_cast_fp16)[name = tensor("k_1_cast_fp16")]; + tensor v_1_begin_0 = const()[name = tensor("v_1_begin_0"), val = tensor([0, 0, 2, 0, 0])]; + tensor v_1_end_0 = const()[name = tensor("v_1_end_0"), val = tensor([1, 1, 3, 16, 64])]; + tensor v_1_end_mask_0 = const()[name = tensor("v_1_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor v_1_squeeze_mask_0 = const()[name = tensor("v_1_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor v_1_cast_fp16 = slice_by_index(begin = v_1_begin_0, end = v_1_end_0, end_mask = v_1_end_mask_0, squeeze_mask = v_1_squeeze_mask_0, x = qkv_1_cast_fp16)[name = tensor("v_1_cast_fp16")]; + tensor var_284 = const()[name = tensor("op_284"), val = tensor([1, 1, 1, 1])]; + tensor position0_to_fp16_dtype_0 = const()[name = tensor("position0_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor position0_to_fp16 = cast(dtype = position0_to_fp16_dtype_0, x = position0)[name = tensor("cast_553")]; + tensor ts_5_cast_fp16 = reshape(shape = var_284, x = position0_to_fp16)[name = tensor("ts_5_cast_fp16")]; + tensor var_291 = const()[name = tensor("op_291"), val = tensor([1, 1, 16, 32, 2])]; + tensor q_complex_1_cast_fp16 = reshape(shape = var_291, x = q_1_cast_fp16)[name = tensor("q_complex_1_cast_fp16")]; + tensor var_298 = const()[name = tensor("op_298"), val = tensor([1, 1, 16, 32, 2])]; + tensor k_complex_1_cast_fp16 = reshape(shape = var_298, x = k_1_cast_fp16)[name = tensor("k_complex_1_cast_fp16")]; + tensor var_302_begin_0 = const()[name = tensor("op_302_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_302_end_0 = const()[name = tensor("op_302_end_0"), val = tensor([1, 1, 16, 32, 1])]; + tensor var_302_end_mask_0 = const()[name = tensor("op_302_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_302_squeeze_mask_0 = const()[name = tensor("op_302_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_302_cast_fp16 = slice_by_index(begin = var_302_begin_0, end = var_302_end_0, end_mask = var_302_end_mask_0, squeeze_mask = var_302_squeeze_mask_0, x = q_complex_1_cast_fp16)[name = tensor("op_302_cast_fp16")]; + tensor var_310_begin_0 = const()[name = tensor("op_310_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_310_end_0 = const()[name = tensor("op_310_end_0"), val = tensor([1, 1, 16, 32, 2])]; + tensor var_310_end_mask_0 = const()[name = tensor("op_310_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_310_squeeze_mask_0 = const()[name = tensor("op_310_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_310_cast_fp16 = slice_by_index(begin = var_310_begin_0, end = var_310_end_0, end_mask = var_310_end_mask_0, squeeze_mask = var_310_squeeze_mask_0, x = q_complex_1_cast_fp16)[name = tensor("op_310_cast_fp16")]; + tensor var_318_begin_0 = const()[name = tensor("op_318_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_318_end_0 = const()[name = tensor("op_318_end_0"), val = tensor([1, 1, 16, 32, 1])]; + tensor var_318_end_mask_0 = const()[name = tensor("op_318_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_318_squeeze_mask_0 = const()[name = tensor("op_318_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_318_cast_fp16 = slice_by_index(begin = var_318_begin_0, end = var_318_end_0, end_mask = var_318_end_mask_0, squeeze_mask = var_318_squeeze_mask_0, x = k_complex_1_cast_fp16)[name = tensor("op_318_cast_fp16")]; + tensor var_326_begin_0 = const()[name = tensor("op_326_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_326_end_0 = const()[name = tensor("op_326_end_0"), val = tensor([1, 1, 16, 32, 2])]; + tensor var_326_end_mask_0 = const()[name = tensor("op_326_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_326_squeeze_mask_0 = const()[name = tensor("op_326_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_326_cast_fp16 = slice_by_index(begin = var_326_begin_0, end = var_326_end_0, end_mask = var_326_end_mask_0, squeeze_mask = var_326_squeeze_mask_0, x = k_complex_1_cast_fp16)[name = tensor("op_326_cast_fp16")]; + tensor freqs_1_to_fp16 = const()[name = tensor("freqs_1_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(6369728)))]; + tensor var_332_cast_fp16 = mul(x = freqs_1_to_fp16, y = ts_5_cast_fp16)[name = tensor("op_332_cast_fp16")]; + tensor rotr_1_cast_fp16 = cos(x = var_332_cast_fp16)[name = tensor("rotr_1_cast_fp16")]; + tensor roti_1_cast_fp16 = sin(x = var_332_cast_fp16)[name = tensor("roti_1_cast_fp16")]; + tensor var_336_cast_fp16 = mul(x = var_302_cast_fp16, y = rotr_1_cast_fp16)[name = tensor("op_336_cast_fp16")]; + tensor var_337_cast_fp16 = mul(x = var_310_cast_fp16, y = roti_1_cast_fp16)[name = tensor("op_337_cast_fp16")]; + tensor qor_1_cast_fp16 = sub(x = var_336_cast_fp16, y = var_337_cast_fp16)[name = tensor("qor_1_cast_fp16")]; + tensor var_340_cast_fp16 = mul(x = var_302_cast_fp16, y = roti_1_cast_fp16)[name = tensor("op_340_cast_fp16")]; + tensor var_341_cast_fp16 = mul(x = var_310_cast_fp16, y = rotr_1_cast_fp16)[name = tensor("op_341_cast_fp16")]; + tensor qoi_1_cast_fp16 = add(x = var_340_cast_fp16, y = var_341_cast_fp16)[name = tensor("qoi_1_cast_fp16")]; + tensor var_344_cast_fp16 = mul(x = var_318_cast_fp16, y = rotr_1_cast_fp16)[name = tensor("op_344_cast_fp16")]; + tensor var_345_cast_fp16 = mul(x = var_326_cast_fp16, y = roti_1_cast_fp16)[name = tensor("op_345_cast_fp16")]; + tensor kor_1_cast_fp16 = sub(x = var_344_cast_fp16, y = var_345_cast_fp16)[name = tensor("kor_1_cast_fp16")]; + tensor var_348_cast_fp16 = mul(x = var_318_cast_fp16, y = roti_1_cast_fp16)[name = tensor("op_348_cast_fp16")]; + tensor var_349_cast_fp16 = mul(x = var_326_cast_fp16, y = rotr_1_cast_fp16)[name = tensor("op_349_cast_fp16")]; + tensor koi_1_cast_fp16 = add(x = var_348_cast_fp16, y = var_349_cast_fp16)[name = tensor("koi_1_cast_fp16")]; + tensor qo_1_axis_0 = const()[name = tensor("qo_1_axis_0"), val = tensor(-1)]; + tensor qo_1_cast_fp16 = stack(axis = qo_1_axis_0, values = (qor_1_cast_fp16, qoi_1_cast_fp16))[name = tensor("qo_1_cast_fp16")]; + tensor ko_1_axis_0 = const()[name = tensor("ko_1_axis_0"), val = tensor(-1)]; + tensor ko_1_cast_fp16 = stack(axis = ko_1_axis_0, values = (kor_1_cast_fp16, koi_1_cast_fp16))[name = tensor("ko_1_cast_fp16")]; + tensor var_382 = const()[name = tensor("op_382"), val = tensor([1, 1, 16, 64])]; + tensor q_3_cast_fp16 = reshape(shape = var_382, x = qo_1_cast_fp16)[name = tensor("q_3_cast_fp16")]; + tensor var_388 = const()[name = tensor("op_388"), val = tensor([1, 1, 16, 64])]; + tensor k_3_cast_fp16 = reshape(shape = var_388, x = ko_1_cast_fp16)[name = tensor("k_3_cast_fp16")]; + tensor _inversed_413_y_0_to_fp16 = const()[name = tensor("_inversed_413_y_0_to_fp16"), val = tensor(0x1p-9)]; + tensor _inversed_413_cast_fp16 = mul(x = ts_5_cast_fp16, y = _inversed_413_y_0_to_fp16)[name = tensor("_inversed_413_cast_fp16")]; + tensor var_414_cast_fp16 = floor(x = _inversed_413_cast_fp16)[name = tensor("op_414_cast_fp16")]; + tensor var_415_to_fp16 = const()[name = tensor("op_415_to_fp16"), val = tensor(0x1p+9)]; + tensor var_416_cast_fp16 = mul(x = var_414_cast_fp16, y = var_415_to_fp16)[name = tensor("op_416_cast_fp16")]; + tensor write_indices_float_3_cast_fp16 = sub(x = ts_5_cast_fp16, y = var_416_cast_fp16)[name = tensor("write_indices_float_3_cast_fp16")]; + tensor var_423_dtype_0 = const()[name = tensor("op_423_dtype_0"), val = tensor("int32")]; + tensor write_indices_1_reps_0 = const()[name = tensor("write_indices_1_reps_0"), val = tensor([1, 1, 16, 64])]; + tensor write_indices_float_3_cast_fp16_to_int32 = cast(dtype = var_423_dtype_0, x = write_indices_float_3_cast_fp16)[name = tensor("cast_552")]; + tensor write_indices_1 = tile(reps = write_indices_1_reps_0, x = write_indices_float_3_cast_fp16_to_int32)[name = tensor("write_indices_1")]; + tensor var_433_begin_0 = const()[name = tensor("op_433_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_433_end_0 = const()[name = tensor("op_433_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor var_433_end_mask_0 = const()[name = tensor("op_433_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_433_squeeze_mask_0 = const()[name = tensor("op_433_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor cache0_to_fp16_dtype_0 = const()[name = tensor("cache0_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor cache0_to_fp16 = cast(dtype = cache0_to_fp16_dtype_0, x = cache0)[name = tensor("cast_551")]; + tensor var_433_cast_fp16 = slice_by_index(begin = var_433_begin_0, end = var_433_end_0, end_mask = var_433_end_mask_0, squeeze_mask = var_433_squeeze_mask_0, x = cache0_to_fp16)[name = tensor("op_433_cast_fp16")]; + tensor var_435_axis_0 = const()[name = tensor("op_435_axis_0"), val = tensor(1)]; + tensor var_435_mode_0 = const()[name = tensor("op_435_mode_0"), val = tensor("update")]; + tensor var_435_validate_indices_0 = const()[name = tensor("op_435_validate_indices_0"), val = tensor(false)]; + tensor var_435_cast_fp16 = scatter_along_axis(axis = var_435_axis_0, data = var_433_cast_fp16, indices = write_indices_1, mode = var_435_mode_0, updates = k_3_cast_fp16, validate_indices = var_435_validate_indices_0)[name = tensor("op_435_cast_fp16")]; + tensor concat_2 = const()[name = tensor("concat_2"), val = tensor([0, 0, 0, 0, 0])]; + tensor concat_3 = const()[name = tensor("concat_3"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_1_internal_tensor_assign_1_stride_0 = const()[name = tensor("new_cache_1_internal_tensor_assign_1_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_1_internal_tensor_assign_1_begin_mask_0 = const()[name = tensor("new_cache_1_internal_tensor_assign_1_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_1_internal_tensor_assign_1_end_mask_0 = const()[name = tensor("new_cache_1_internal_tensor_assign_1_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_1_internal_tensor_assign_1_squeeze_mask_0 = const()[name = tensor("new_cache_1_internal_tensor_assign_1_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_48 = const()[name = tensor("shape_48"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_0 = const()[name = tensor("reduce_prod_0"), val = tensor(1048576)]; + tensor range_1d_0_start_0 = const()[name = tensor("range_1d_0_start_0"), val = tensor(0)]; + tensor range_1d_0_step_0 = const()[name = tensor("range_1d_0_step_0"), val = tensor(1)]; + tensor range_1d_0 = range_1d(end = reduce_prod_0, start = range_1d_0_start_0, step = range_1d_0_step_0)[name = tensor("range_1d_0")]; + tensor reshape_0 = reshape(shape = shape_48, x = range_1d_0)[name = tensor("reshape_0")]; + tensor slice_by_index_0 = slice_by_index(begin = concat_2, begin_mask = new_cache_1_internal_tensor_assign_1_begin_mask_0, end = concat_3, end_mask = new_cache_1_internal_tensor_assign_1_end_mask_0, squeeze_mask = new_cache_1_internal_tensor_assign_1_squeeze_mask_0, stride = new_cache_1_internal_tensor_assign_1_stride_0, x = reshape_0)[name = tensor("slice_by_index_0")]; + tensor reshape_1_shape_0 = const()[name = tensor("reshape_1_shape_0"), val = tensor([-1])]; + tensor reshape_1 = reshape(shape = reshape_1_shape_0, x = slice_by_index_0)[name = tensor("reshape_1")]; + tensor reshape_2_shape_0 = const()[name = tensor("reshape_2_shape_0"), val = tensor([-1])]; + tensor reshape_2_cast_fp16 = reshape(shape = reshape_2_shape_0, x = var_435_cast_fp16)[name = tensor("reshape_2_cast_fp16")]; + tensor reshape_3_shape_0 = const()[name = tensor("reshape_3_shape_0"), val = tensor([-1])]; + tensor reshape_3_cast_fp16 = reshape(shape = reshape_3_shape_0, x = cache0_to_fp16)[name = tensor("reshape_3_cast_fp16")]; + tensor scatter_0_mode_0 = const()[name = tensor("scatter_0_mode_0"), val = tensor("update")]; + tensor scatter_0_axis_0 = const()[name = tensor("scatter_0_axis_0"), val = tensor(0)]; + tensor scatter_0_validate_indices_0 = const()[name = tensor("scatter_0_validate_indices_0"), val = tensor(false)]; + tensor scatter_0_cast_fp16 = scatter(axis = scatter_0_axis_0, data = reshape_3_cast_fp16, indices = reshape_1, mode = scatter_0_mode_0, updates = reshape_2_cast_fp16, validate_indices = scatter_0_validate_indices_0)[name = tensor("scatter_0_cast_fp16")]; + tensor reshape_4_cast_fp16 = reshape(shape = shape_48, x = scatter_0_cast_fp16)[name = tensor("reshape_4_cast_fp16")]; + tensor var_443_begin_0 = const()[name = tensor("op_443_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor var_443_end_0 = const()[name = tensor("op_443_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor var_443_end_mask_0 = const()[name = tensor("op_443_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_443_squeeze_mask_0 = const()[name = tensor("op_443_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor var_443_cast_fp16 = slice_by_index(begin = var_443_begin_0, end = var_443_end_0, end_mask = var_443_end_mask_0, squeeze_mask = var_443_squeeze_mask_0, x = reshape_4_cast_fp16)[name = tensor("op_443_cast_fp16")]; + tensor var_445_axis_0 = const()[name = tensor("op_445_axis_0"), val = tensor(1)]; + tensor var_445_mode_0 = const()[name = tensor("op_445_mode_0"), val = tensor("update")]; + tensor var_445_validate_indices_0 = const()[name = tensor("op_445_validate_indices_0"), val = tensor(false)]; + tensor var_445_cast_fp16 = scatter_along_axis(axis = var_445_axis_0, data = var_443_cast_fp16, indices = write_indices_1, mode = var_445_mode_0, updates = v_1_cast_fp16, validate_indices = var_445_validate_indices_0)[name = tensor("op_445_cast_fp16")]; + tensor concat_4 = const()[name = tensor("concat_4"), val = tensor([1, 0, 0, 0, 0])]; + tensor concat_5 = const()[name = tensor("concat_5"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_1_internal_tensor_assign_2_stride_0 = const()[name = tensor("new_cache_1_internal_tensor_assign_2_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_1_internal_tensor_assign_2_begin_mask_0 = const()[name = tensor("new_cache_1_internal_tensor_assign_2_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_1_internal_tensor_assign_2_end_mask_0 = const()[name = tensor("new_cache_1_internal_tensor_assign_2_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_1_internal_tensor_assign_2_squeeze_mask_0 = const()[name = tensor("new_cache_1_internal_tensor_assign_2_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_49 = const()[name = tensor("shape_49"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_1 = const()[name = tensor("reduce_prod_1"), val = tensor(1048576)]; + tensor range_1d_1_start_0 = const()[name = tensor("range_1d_1_start_0"), val = tensor(0)]; + tensor range_1d_1_step_0 = const()[name = tensor("range_1d_1_step_0"), val = tensor(1)]; + tensor range_1d_1 = range_1d(end = reduce_prod_1, start = range_1d_1_start_0, step = range_1d_1_step_0)[name = tensor("range_1d_1")]; + tensor reshape_5 = reshape(shape = shape_49, x = range_1d_1)[name = tensor("reshape_5")]; + tensor slice_by_index_1 = slice_by_index(begin = concat_4, begin_mask = new_cache_1_internal_tensor_assign_2_begin_mask_0, end = concat_5, end_mask = new_cache_1_internal_tensor_assign_2_end_mask_0, squeeze_mask = new_cache_1_internal_tensor_assign_2_squeeze_mask_0, stride = new_cache_1_internal_tensor_assign_2_stride_0, x = reshape_5)[name = tensor("slice_by_index_1")]; + tensor reshape_6_shape_0 = const()[name = tensor("reshape_6_shape_0"), val = tensor([-1])]; + tensor reshape_6 = reshape(shape = reshape_6_shape_0, x = slice_by_index_1)[name = tensor("reshape_6")]; + tensor reshape_7_shape_0 = const()[name = tensor("reshape_7_shape_0"), val = tensor([-1])]; + tensor reshape_7_cast_fp16 = reshape(shape = reshape_7_shape_0, x = var_445_cast_fp16)[name = tensor("reshape_7_cast_fp16")]; + tensor reshape_8_shape_0 = const()[name = tensor("reshape_8_shape_0"), val = tensor([-1])]; + tensor reshape_8_cast_fp16 = reshape(shape = reshape_8_shape_0, x = reshape_4_cast_fp16)[name = tensor("reshape_8_cast_fp16")]; + tensor scatter_1_mode_0 = const()[name = tensor("scatter_1_mode_0"), val = tensor("update")]; + tensor scatter_1_axis_0 = const()[name = tensor("scatter_1_axis_0"), val = tensor(0)]; + tensor scatter_1_validate_indices_0 = const()[name = tensor("scatter_1_validate_indices_0"), val = tensor(false)]; + tensor scatter_1_cast_fp16 = scatter(axis = scatter_1_axis_0, data = reshape_8_cast_fp16, indices = reshape_6, mode = scatter_1_mode_0, updates = reshape_7_cast_fp16, validate_indices = scatter_1_validate_indices_0)[name = tensor("scatter_1_cast_fp16")]; + tensor reshape_9_cast_fp16 = reshape(shape = shape_49, x = scatter_1_cast_fp16)[name = tensor("reshape_9_cast_fp16")]; + tensor reshape_9_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("reshape_9_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor keys_1_begin_0 = const()[name = tensor("keys_1_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor keys_1_end_0 = const()[name = tensor("keys_1_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor keys_1_end_mask_0 = const()[name = tensor("keys_1_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor keys_1_squeeze_mask_0 = const()[name = tensor("keys_1_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor keys_1_cast_fp16 = slice_by_index(begin = keys_1_begin_0, end = keys_1_end_0, end_mask = keys_1_end_mask_0, squeeze_mask = keys_1_squeeze_mask_0, x = reshape_9_cast_fp16)[name = tensor("keys_1_cast_fp16")]; + tensor values_1_begin_0 = const()[name = tensor("values_1_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor values_1_end_0 = const()[name = tensor("values_1_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor values_1_end_mask_0 = const()[name = tensor("values_1_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor values_1_squeeze_mask_0 = const()[name = tensor("values_1_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor values_1_cast_fp16 = slice_by_index(begin = values_1_begin_0, end = values_1_end_0, end_mask = values_1_end_mask_0, squeeze_mask = values_1_squeeze_mask_0, x = reshape_9_cast_fp16)[name = tensor("values_1_cast_fp16")]; + tensor var_457_cast_fp16 = not_equal(x = keys_1_cast_fp16, y = keys_1_cast_fp16)[name = tensor("op_457_cast_fp16")]; + tensor var_463_to_fp16 = const()[name = tensor("op_463_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(6369856)))]; + tensor keys_3_cast_fp16 = select(a = var_463_to_fp16, b = keys_1_cast_fp16, cond = var_457_cast_fp16)[name = tensor("keys_3_cast_fp16")]; + tensor var_465_cast_fp16 = not_equal(x = values_1_cast_fp16, y = values_1_cast_fp16)[name = tensor("op_465_cast_fp16")]; + tensor values_3_cast_fp16 = select(a = var_463_to_fp16, b = values_1_cast_fp16, cond = var_465_cast_fp16)[name = tensor("values_3_cast_fp16")]; + tensor var_489 = const()[name = tensor("op_489"), val = tensor([0, 2, 1, 3])]; + tensor var_505 = const()[name = tensor("op_505"), val = tensor([1, 1, 1])]; + tensor var_506_cast_fp16 = reshape(shape = var_505, x = position0_to_fp16)[name = tensor("op_506_cast_fp16")]; + tensor var_526_to_fp16 = const()[name = tensor("op_526_to_fp16"), val = tensor(0x1p+0)]; + tensor valid_len_1_cast_fp16 = add(x = var_506_cast_fp16, y = var_526_to_fp16)[name = tensor("valid_len_1_cast_fp16")]; + tensor k_positions_1_promoted_to_fp16 = const()[name = tensor("k_positions_1_promoted_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(7418496)))]; + tensor valid_mask_1_cast_fp16 = less(x = k_positions_1_promoted_to_fp16, y = valid_len_1_cast_fp16)[name = tensor("valid_mask_1_cast_fp16")]; + tensor causal_mask_1_cast_fp16 = less_equal(x = k_positions_1_promoted_to_fp16, y = var_506_cast_fp16)[name = tensor("causal_mask_1_cast_fp16")]; + tensor attn_mask_1 = logical_and(x = valid_mask_1_cast_fp16, y = causal_mask_1_cast_fp16)[name = tensor("attn_mask_1")]; + tensor attn_mask_3_axes_0 = const()[name = tensor("attn_mask_3_axes_0"), val = tensor([1])]; + tensor attn_mask_3 = expand_dims(axes = attn_mask_3_axes_0, x = attn_mask_1)[name = tensor("attn_mask_3")]; + tensor var_544_transpose_x_0 = const()[name = tensor("op_544_transpose_x_0"), val = tensor(false)]; + tensor var_544_transpose_y_0 = const()[name = tensor("op_544_transpose_y_0"), val = tensor(false)]; + tensor transpose_72_perm_0 = const()[name = tensor("transpose_72_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_73_perm_0 = const()[name = tensor("transpose_73_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_73 = transpose(perm = transpose_73_perm_0, x = keys_3_cast_fp16)[name = tensor("transpose_213")]; + tensor transpose_72 = transpose(perm = transpose_72_perm_0, x = q_3_cast_fp16)[name = tensor("transpose_214")]; + tensor var_544_cast_fp16 = matmul(transpose_x = var_544_transpose_x_0, transpose_y = var_544_transpose_y_0, x = transpose_72, y = transpose_73)[name = tensor("op_544_cast_fp16")]; + tensor var_538_to_fp16 = const()[name = tensor("op_538_to_fp16"), val = tensor([0x1p-3])]; + tensor attn_weights_1_cast_fp16 = mul(x = var_544_cast_fp16, y = var_538_to_fp16)[name = tensor("attn_weights_1_cast_fp16")]; + tensor var_546 = logical_not(x = attn_mask_3)[name = tensor("op_546")]; + tensor var_547_to_fp16 = const()[name = tensor("op_547_to_fp16"), val = tensor(-inf)]; + tensor attn_weights_3_cast_fp16 = select(a = var_547_to_fp16, b = attn_weights_1_cast_fp16, cond = var_546)[name = tensor("attn_weights_3_cast_fp16")]; + tensor var_549 = const()[name = tensor("op_549"), val = tensor(-1)]; + tensor attn_weights_5_cast_fp16 = softmax(axis = var_549, x = attn_weights_3_cast_fp16)[name = tensor("attn_weights_5_cast_fp16")]; + tensor attn_output_1_transpose_x_0 = const()[name = tensor("attn_output_1_transpose_x_0"), val = tensor(false)]; + tensor attn_output_1_transpose_y_0 = const()[name = tensor("attn_output_1_transpose_y_0"), val = tensor(false)]; + tensor values_5_cast_fp16 = transpose(perm = var_489, x = values_3_cast_fp16)[name = tensor("transpose_215")]; + tensor attn_output_1_cast_fp16 = matmul(transpose_x = attn_output_1_transpose_x_0, transpose_y = attn_output_1_transpose_y_0, x = attn_weights_5_cast_fp16, y = values_5_cast_fp16)[name = tensor("attn_output_1_cast_fp16")]; + tensor var_557 = const()[name = tensor("op_557"), val = tensor([0, 2, 1, 3])]; + tensor var_562 = const()[name = tensor("op_562"), val = tensor([1, 1, 1024])]; + tensor var_558_cast_fp16 = transpose(perm = var_557, x = attn_output_1_cast_fp16)[name = tensor("transpose_212")]; + tensor input_7_cast_fp16 = reshape(shape = var_562, x = var_558_cast_fp16)[name = tensor("input_7_cast_fp16")]; + tensor attn0_out_proj_weight_to_fp16 = const()[name = tensor("attn0_out_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(7419584)))]; + tensor linear_2_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = attn0_out_proj_weight_to_fp16, x = input_7_cast_fp16)[name = tensor("linear_2_cast_fp16")]; + tensor var_568_to_fp16 = const()[name = tensor("op_568_to_fp16"), val = tensor(0x1p+0)]; + tensor var_569_cast_fp16 = add(x = position0_to_fp16, y = var_568_to_fp16)[name = tensor("op_569_cast_fp16")]; + tensor var_569_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_569_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor input_9_cast_fp16 = add(x = linear_0_cast_fp16, y = linear_2_cast_fp16)[name = tensor("input_9_cast_fp16")]; + tensor input_11_axes_0 = const()[name = tensor("input_11_axes_0"), val = tensor([-1])]; + tensor norm0_2_weight_to_fp16 = const()[name = tensor("norm0_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9516800)))]; + tensor norm0_2_bias_to_fp16 = const()[name = tensor("norm0_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9518912)))]; + tensor var_573_to_fp16 = const()[name = tensor("op_573_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_11_cast_fp16 = layer_norm(axes = input_11_axes_0, beta = norm0_2_bias_to_fp16, epsilon = var_573_to_fp16, gamma = norm0_2_weight_to_fp16, x = input_9_cast_fp16)[name = tensor("input_11_cast_fp16")]; + tensor linear0_1_weight_to_fp16 = const()[name = tensor("linear0_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9521024)))]; + tensor linear_3_bias_0_to_fp16 = const()[name = tensor("linear_3_bias_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(17909696)))]; + tensor linear_3_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = linear0_1_weight_to_fp16, x = input_11_cast_fp16)[name = tensor("linear_3_cast_fp16")]; + tensor input_13_mode_0 = const()[name = tensor("input_13_mode_0"), val = tensor("EXACT")]; + tensor input_13_cast_fp16 = gelu(mode = input_13_mode_0, x = linear_3_cast_fp16)[name = tensor("input_13_cast_fp16")]; + tensor linear0_2_weight_to_fp16 = const()[name = tensor("linear0_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(17917952)))]; + tensor linear_4_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = linear0_2_weight_to_fp16, x = input_13_cast_fp16)[name = tensor("linear_4_cast_fp16")]; + tensor input_15_cast_fp16 = add(x = input_9_cast_fp16, y = linear_4_cast_fp16)[name = tensor("input_15_cast_fp16")]; + tensor input_17_axes_0 = const()[name = tensor("input_17_axes_0"), val = tensor([-1])]; + tensor norm1_1_weight_to_fp16 = const()[name = tensor("norm1_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(26306624)))]; + tensor norm1_1_bias_to_fp16 = const()[name = tensor("norm1_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(26308736)))]; + tensor var_590_to_fp16 = const()[name = tensor("op_590_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_17_cast_fp16 = layer_norm(axes = input_17_axes_0, beta = norm1_1_bias_to_fp16, epsilon = var_590_to_fp16, gamma = norm1_1_weight_to_fp16, x = input_15_cast_fp16)[name = tensor("input_17_cast_fp16")]; + tensor attn1_in_proj_weight_to_fp16 = const()[name = tensor("attn1_in_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(26310848)))]; + tensor linear_5_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = attn1_in_proj_weight_to_fp16, x = input_17_cast_fp16)[name = tensor("linear_5_cast_fp16")]; + tensor var_604 = const()[name = tensor("op_604"), val = tensor([1, 1, 3, 16, 64])]; + tensor qkv_3_cast_fp16 = reshape(shape = var_604, x = linear_5_cast_fp16)[name = tensor("qkv_3_cast_fp16")]; + tensor q_7_begin_0 = const()[name = tensor("q_7_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor q_7_end_0 = const()[name = tensor("q_7_end_0"), val = tensor([1, 1, 1, 16, 64])]; + tensor q_7_end_mask_0 = const()[name = tensor("q_7_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor q_7_squeeze_mask_0 = const()[name = tensor("q_7_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor q_7_cast_fp16 = slice_by_index(begin = q_7_begin_0, end = q_7_end_0, end_mask = q_7_end_mask_0, squeeze_mask = q_7_squeeze_mask_0, x = qkv_3_cast_fp16)[name = tensor("q_7_cast_fp16")]; + tensor k_5_begin_0 = const()[name = tensor("k_5_begin_0"), val = tensor([0, 0, 1, 0, 0])]; + tensor k_5_end_0 = const()[name = tensor("k_5_end_0"), val = tensor([1, 1, 2, 16, 64])]; + tensor k_5_end_mask_0 = const()[name = tensor("k_5_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor k_5_squeeze_mask_0 = const()[name = tensor("k_5_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor k_5_cast_fp16 = slice_by_index(begin = k_5_begin_0, end = k_5_end_0, end_mask = k_5_end_mask_0, squeeze_mask = k_5_squeeze_mask_0, x = qkv_3_cast_fp16)[name = tensor("k_5_cast_fp16")]; + tensor v_3_begin_0 = const()[name = tensor("v_3_begin_0"), val = tensor([0, 0, 2, 0, 0])]; + tensor v_3_end_0 = const()[name = tensor("v_3_end_0"), val = tensor([1, 1, 3, 16, 64])]; + tensor v_3_end_mask_0 = const()[name = tensor("v_3_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor v_3_squeeze_mask_0 = const()[name = tensor("v_3_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor v_3_cast_fp16 = slice_by_index(begin = v_3_begin_0, end = v_3_end_0, end_mask = v_3_end_mask_0, squeeze_mask = v_3_squeeze_mask_0, x = qkv_3_cast_fp16)[name = tensor("v_3_cast_fp16")]; + tensor var_670 = const()[name = tensor("op_670"), val = tensor([1, 1, 1, 1])]; + tensor position1_to_fp16_dtype_0 = const()[name = tensor("position1_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor position1_to_fp16 = cast(dtype = position1_to_fp16_dtype_0, x = position1)[name = tensor("cast_548")]; + tensor ts_11_cast_fp16 = reshape(shape = var_670, x = position1_to_fp16)[name = tensor("ts_11_cast_fp16")]; + tensor var_677 = const()[name = tensor("op_677"), val = tensor([1, 1, 16, 32, 2])]; + tensor q_complex_3_cast_fp16 = reshape(shape = var_677, x = q_7_cast_fp16)[name = tensor("q_complex_3_cast_fp16")]; + tensor var_684 = const()[name = tensor("op_684"), val = tensor([1, 1, 16, 32, 2])]; + tensor k_complex_3_cast_fp16 = reshape(shape = var_684, x = k_5_cast_fp16)[name = tensor("k_complex_3_cast_fp16")]; + tensor var_688_begin_0 = const()[name = tensor("op_688_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_688_end_0 = const()[name = tensor("op_688_end_0"), val = tensor([1, 1, 16, 32, 1])]; + tensor var_688_end_mask_0 = const()[name = tensor("op_688_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_688_squeeze_mask_0 = const()[name = tensor("op_688_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_688_cast_fp16 = slice_by_index(begin = var_688_begin_0, end = var_688_end_0, end_mask = var_688_end_mask_0, squeeze_mask = var_688_squeeze_mask_0, x = q_complex_3_cast_fp16)[name = tensor("op_688_cast_fp16")]; + tensor var_696_begin_0 = const()[name = tensor("op_696_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_696_end_0 = const()[name = tensor("op_696_end_0"), val = tensor([1, 1, 16, 32, 2])]; + tensor var_696_end_mask_0 = const()[name = tensor("op_696_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_696_squeeze_mask_0 = const()[name = tensor("op_696_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_696_cast_fp16 = slice_by_index(begin = var_696_begin_0, end = var_696_end_0, end_mask = var_696_end_mask_0, squeeze_mask = var_696_squeeze_mask_0, x = q_complex_3_cast_fp16)[name = tensor("op_696_cast_fp16")]; + tensor var_704_begin_0 = const()[name = tensor("op_704_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_704_end_0 = const()[name = tensor("op_704_end_0"), val = tensor([1, 1, 16, 32, 1])]; + tensor var_704_end_mask_0 = const()[name = tensor("op_704_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_704_squeeze_mask_0 = const()[name = tensor("op_704_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_704_cast_fp16 = slice_by_index(begin = var_704_begin_0, end = var_704_end_0, end_mask = var_704_end_mask_0, squeeze_mask = var_704_squeeze_mask_0, x = k_complex_3_cast_fp16)[name = tensor("op_704_cast_fp16")]; + tensor var_712_begin_0 = const()[name = tensor("op_712_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_712_end_0 = const()[name = tensor("op_712_end_0"), val = tensor([1, 1, 16, 32, 2])]; + tensor var_712_end_mask_0 = const()[name = tensor("op_712_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_712_squeeze_mask_0 = const()[name = tensor("op_712_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_712_cast_fp16 = slice_by_index(begin = var_712_begin_0, end = var_712_end_0, end_mask = var_712_end_mask_0, squeeze_mask = var_712_squeeze_mask_0, x = k_complex_3_cast_fp16)[name = tensor("op_712_cast_fp16")]; + tensor freqs_3_to_fp16 = const()[name = tensor("freqs_3_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(32602368)))]; + tensor var_718_cast_fp16 = mul(x = freqs_3_to_fp16, y = ts_11_cast_fp16)[name = tensor("op_718_cast_fp16")]; + tensor rotr_3_cast_fp16 = cos(x = var_718_cast_fp16)[name = tensor("rotr_3_cast_fp16")]; + tensor roti_3_cast_fp16 = sin(x = var_718_cast_fp16)[name = tensor("roti_3_cast_fp16")]; + tensor var_722_cast_fp16 = mul(x = var_688_cast_fp16, y = rotr_3_cast_fp16)[name = tensor("op_722_cast_fp16")]; + tensor var_723_cast_fp16 = mul(x = var_696_cast_fp16, y = roti_3_cast_fp16)[name = tensor("op_723_cast_fp16")]; + tensor qor_5_cast_fp16 = sub(x = var_722_cast_fp16, y = var_723_cast_fp16)[name = tensor("qor_5_cast_fp16")]; + tensor var_726_cast_fp16 = mul(x = var_688_cast_fp16, y = roti_3_cast_fp16)[name = tensor("op_726_cast_fp16")]; + tensor var_727_cast_fp16 = mul(x = var_696_cast_fp16, y = rotr_3_cast_fp16)[name = tensor("op_727_cast_fp16")]; + tensor qoi_5_cast_fp16 = add(x = var_726_cast_fp16, y = var_727_cast_fp16)[name = tensor("qoi_5_cast_fp16")]; + tensor var_730_cast_fp16 = mul(x = var_704_cast_fp16, y = rotr_3_cast_fp16)[name = tensor("op_730_cast_fp16")]; + tensor var_731_cast_fp16 = mul(x = var_712_cast_fp16, y = roti_3_cast_fp16)[name = tensor("op_731_cast_fp16")]; + tensor kor_5_cast_fp16 = sub(x = var_730_cast_fp16, y = var_731_cast_fp16)[name = tensor("kor_5_cast_fp16")]; + tensor var_734_cast_fp16 = mul(x = var_704_cast_fp16, y = roti_3_cast_fp16)[name = tensor("op_734_cast_fp16")]; + tensor var_735_cast_fp16 = mul(x = var_712_cast_fp16, y = rotr_3_cast_fp16)[name = tensor("op_735_cast_fp16")]; + tensor koi_5_cast_fp16 = add(x = var_734_cast_fp16, y = var_735_cast_fp16)[name = tensor("koi_5_cast_fp16")]; + tensor qo_3_axis_0 = const()[name = tensor("qo_3_axis_0"), val = tensor(-1)]; + tensor qo_3_cast_fp16 = stack(axis = qo_3_axis_0, values = (qor_5_cast_fp16, qoi_5_cast_fp16))[name = tensor("qo_3_cast_fp16")]; + tensor ko_3_axis_0 = const()[name = tensor("ko_3_axis_0"), val = tensor(-1)]; + tensor ko_3_cast_fp16 = stack(axis = ko_3_axis_0, values = (kor_5_cast_fp16, koi_5_cast_fp16))[name = tensor("ko_3_cast_fp16")]; + tensor var_768 = const()[name = tensor("op_768"), val = tensor([1, 1, 16, 64])]; + tensor q_9_cast_fp16 = reshape(shape = var_768, x = qo_3_cast_fp16)[name = tensor("q_9_cast_fp16")]; + tensor var_774 = const()[name = tensor("op_774"), val = tensor([1, 1, 16, 64])]; + tensor k_7_cast_fp16 = reshape(shape = var_774, x = ko_3_cast_fp16)[name = tensor("k_7_cast_fp16")]; + tensor _inversed_799_y_0_to_fp16 = const()[name = tensor("_inversed_799_y_0_to_fp16"), val = tensor(0x1p-9)]; + tensor _inversed_799_cast_fp16 = mul(x = ts_11_cast_fp16, y = _inversed_799_y_0_to_fp16)[name = tensor("_inversed_799_cast_fp16")]; + tensor var_800_cast_fp16 = floor(x = _inversed_799_cast_fp16)[name = tensor("op_800_cast_fp16")]; + tensor var_801_to_fp16 = const()[name = tensor("op_801_to_fp16"), val = tensor(0x1p+9)]; + tensor var_802_cast_fp16 = mul(x = var_800_cast_fp16, y = var_801_to_fp16)[name = tensor("op_802_cast_fp16")]; + tensor write_indices_float_7_cast_fp16 = sub(x = ts_11_cast_fp16, y = var_802_cast_fp16)[name = tensor("write_indices_float_7_cast_fp16")]; + tensor var_809_dtype_0 = const()[name = tensor("op_809_dtype_0"), val = tensor("int32")]; + tensor write_indices_3_reps_0 = const()[name = tensor("write_indices_3_reps_0"), val = tensor([1, 1, 16, 64])]; + tensor write_indices_float_7_cast_fp16_to_int32 = cast(dtype = var_809_dtype_0, x = write_indices_float_7_cast_fp16)[name = tensor("cast_547")]; + tensor write_indices_3 = tile(reps = write_indices_3_reps_0, x = write_indices_float_7_cast_fp16_to_int32)[name = tensor("write_indices_3")]; + tensor var_819_begin_0 = const()[name = tensor("op_819_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_819_end_0 = const()[name = tensor("op_819_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor var_819_end_mask_0 = const()[name = tensor("op_819_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_819_squeeze_mask_0 = const()[name = tensor("op_819_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor cache1_to_fp16_dtype_0 = const()[name = tensor("cache1_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor cache1_to_fp16 = cast(dtype = cache1_to_fp16_dtype_0, x = cache1)[name = tensor("cast_546")]; + tensor var_819_cast_fp16 = slice_by_index(begin = var_819_begin_0, end = var_819_end_0, end_mask = var_819_end_mask_0, squeeze_mask = var_819_squeeze_mask_0, x = cache1_to_fp16)[name = tensor("op_819_cast_fp16")]; + tensor var_821_axis_0 = const()[name = tensor("op_821_axis_0"), val = tensor(1)]; + tensor var_821_mode_0 = const()[name = tensor("op_821_mode_0"), val = tensor("update")]; + tensor var_821_validate_indices_0 = const()[name = tensor("op_821_validate_indices_0"), val = tensor(false)]; + tensor var_821_cast_fp16 = scatter_along_axis(axis = var_821_axis_0, data = var_819_cast_fp16, indices = write_indices_3, mode = var_821_mode_0, updates = k_7_cast_fp16, validate_indices = var_821_validate_indices_0)[name = tensor("op_821_cast_fp16")]; + tensor concat_9 = const()[name = tensor("concat_9"), val = tensor([0, 0, 0, 0, 0])]; + tensor concat_10 = const()[name = tensor("concat_10"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_3_internal_tensor_assign_1_stride_0 = const()[name = tensor("new_cache_3_internal_tensor_assign_1_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_3_internal_tensor_assign_1_begin_mask_0 = const()[name = tensor("new_cache_3_internal_tensor_assign_1_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_3_internal_tensor_assign_1_end_mask_0 = const()[name = tensor("new_cache_3_internal_tensor_assign_1_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_3_internal_tensor_assign_1_squeeze_mask_0 = const()[name = tensor("new_cache_3_internal_tensor_assign_1_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_50 = const()[name = tensor("shape_50"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_2 = const()[name = tensor("reduce_prod_2"), val = tensor(1048576)]; + tensor range_1d_2_start_0 = const()[name = tensor("range_1d_2_start_0"), val = tensor(0)]; + tensor range_1d_2_step_0 = const()[name = tensor("range_1d_2_step_0"), val = tensor(1)]; + tensor range_1d_2 = range_1d(end = reduce_prod_2, start = range_1d_2_start_0, step = range_1d_2_step_0)[name = tensor("range_1d_2")]; + tensor reshape_10 = reshape(shape = shape_50, x = range_1d_2)[name = tensor("reshape_10")]; + tensor slice_by_index_2 = slice_by_index(begin = concat_9, begin_mask = new_cache_3_internal_tensor_assign_1_begin_mask_0, end = concat_10, end_mask = new_cache_3_internal_tensor_assign_1_end_mask_0, squeeze_mask = new_cache_3_internal_tensor_assign_1_squeeze_mask_0, stride = new_cache_3_internal_tensor_assign_1_stride_0, x = reshape_10)[name = tensor("slice_by_index_2")]; + tensor reshape_11_shape_0 = const()[name = tensor("reshape_11_shape_0"), val = tensor([-1])]; + tensor reshape_11 = reshape(shape = reshape_11_shape_0, x = slice_by_index_2)[name = tensor("reshape_11")]; + tensor reshape_12_shape_0 = const()[name = tensor("reshape_12_shape_0"), val = tensor([-1])]; + tensor reshape_12_cast_fp16 = reshape(shape = reshape_12_shape_0, x = var_821_cast_fp16)[name = tensor("reshape_12_cast_fp16")]; + tensor reshape_13_shape_0 = const()[name = tensor("reshape_13_shape_0"), val = tensor([-1])]; + tensor reshape_13_cast_fp16 = reshape(shape = reshape_13_shape_0, x = cache1_to_fp16)[name = tensor("reshape_13_cast_fp16")]; + tensor scatter_2_mode_0 = const()[name = tensor("scatter_2_mode_0"), val = tensor("update")]; + tensor scatter_2_axis_0 = const()[name = tensor("scatter_2_axis_0"), val = tensor(0)]; + tensor scatter_2_validate_indices_0 = const()[name = tensor("scatter_2_validate_indices_0"), val = tensor(false)]; + tensor scatter_2_cast_fp16 = scatter(axis = scatter_2_axis_0, data = reshape_13_cast_fp16, indices = reshape_11, mode = scatter_2_mode_0, updates = reshape_12_cast_fp16, validate_indices = scatter_2_validate_indices_0)[name = tensor("scatter_2_cast_fp16")]; + tensor reshape_14_cast_fp16 = reshape(shape = shape_50, x = scatter_2_cast_fp16)[name = tensor("reshape_14_cast_fp16")]; + tensor var_829_begin_0 = const()[name = tensor("op_829_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor var_829_end_0 = const()[name = tensor("op_829_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor var_829_end_mask_0 = const()[name = tensor("op_829_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_829_squeeze_mask_0 = const()[name = tensor("op_829_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor var_829_cast_fp16 = slice_by_index(begin = var_829_begin_0, end = var_829_end_0, end_mask = var_829_end_mask_0, squeeze_mask = var_829_squeeze_mask_0, x = reshape_14_cast_fp16)[name = tensor("op_829_cast_fp16")]; + tensor var_831_axis_0 = const()[name = tensor("op_831_axis_0"), val = tensor(1)]; + tensor var_831_mode_0 = const()[name = tensor("op_831_mode_0"), val = tensor("update")]; + tensor var_831_validate_indices_0 = const()[name = tensor("op_831_validate_indices_0"), val = tensor(false)]; + tensor var_831_cast_fp16 = scatter_along_axis(axis = var_831_axis_0, data = var_829_cast_fp16, indices = write_indices_3, mode = var_831_mode_0, updates = v_3_cast_fp16, validate_indices = var_831_validate_indices_0)[name = tensor("op_831_cast_fp16")]; + tensor concat_11 = const()[name = tensor("concat_11"), val = tensor([1, 0, 0, 0, 0])]; + tensor concat_12 = const()[name = tensor("concat_12"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_3_internal_tensor_assign_2_stride_0 = const()[name = tensor("new_cache_3_internal_tensor_assign_2_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_3_internal_tensor_assign_2_begin_mask_0 = const()[name = tensor("new_cache_3_internal_tensor_assign_2_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_3_internal_tensor_assign_2_end_mask_0 = const()[name = tensor("new_cache_3_internal_tensor_assign_2_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_3_internal_tensor_assign_2_squeeze_mask_0 = const()[name = tensor("new_cache_3_internal_tensor_assign_2_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_51 = const()[name = tensor("shape_51"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_3 = const()[name = tensor("reduce_prod_3"), val = tensor(1048576)]; + tensor range_1d_3_start_0 = const()[name = tensor("range_1d_3_start_0"), val = tensor(0)]; + tensor range_1d_3_step_0 = const()[name = tensor("range_1d_3_step_0"), val = tensor(1)]; + tensor range_1d_3 = range_1d(end = reduce_prod_3, start = range_1d_3_start_0, step = range_1d_3_step_0)[name = tensor("range_1d_3")]; + tensor reshape_15 = reshape(shape = shape_51, x = range_1d_3)[name = tensor("reshape_15")]; + tensor slice_by_index_3 = slice_by_index(begin = concat_11, begin_mask = new_cache_3_internal_tensor_assign_2_begin_mask_0, end = concat_12, end_mask = new_cache_3_internal_tensor_assign_2_end_mask_0, squeeze_mask = new_cache_3_internal_tensor_assign_2_squeeze_mask_0, stride = new_cache_3_internal_tensor_assign_2_stride_0, x = reshape_15)[name = tensor("slice_by_index_3")]; + tensor reshape_16_shape_0 = const()[name = tensor("reshape_16_shape_0"), val = tensor([-1])]; + tensor reshape_16 = reshape(shape = reshape_16_shape_0, x = slice_by_index_3)[name = tensor("reshape_16")]; + tensor reshape_17_shape_0 = const()[name = tensor("reshape_17_shape_0"), val = tensor([-1])]; + tensor reshape_17_cast_fp16 = reshape(shape = reshape_17_shape_0, x = var_831_cast_fp16)[name = tensor("reshape_17_cast_fp16")]; + tensor reshape_18_shape_0 = const()[name = tensor("reshape_18_shape_0"), val = tensor([-1])]; + tensor reshape_18_cast_fp16 = reshape(shape = reshape_18_shape_0, x = reshape_14_cast_fp16)[name = tensor("reshape_18_cast_fp16")]; + tensor scatter_3_mode_0 = const()[name = tensor("scatter_3_mode_0"), val = tensor("update")]; + tensor scatter_3_axis_0 = const()[name = tensor("scatter_3_axis_0"), val = tensor(0)]; + tensor scatter_3_validate_indices_0 = const()[name = tensor("scatter_3_validate_indices_0"), val = tensor(false)]; + tensor scatter_3_cast_fp16 = scatter(axis = scatter_3_axis_0, data = reshape_18_cast_fp16, indices = reshape_16, mode = scatter_3_mode_0, updates = reshape_17_cast_fp16, validate_indices = scatter_3_validate_indices_0)[name = tensor("scatter_3_cast_fp16")]; + tensor reshape_19_cast_fp16 = reshape(shape = shape_51, x = scatter_3_cast_fp16)[name = tensor("reshape_19_cast_fp16")]; + tensor reshape_19_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("reshape_19_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor keys_7_begin_0 = const()[name = tensor("keys_7_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor keys_7_end_0 = const()[name = tensor("keys_7_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor keys_7_end_mask_0 = const()[name = tensor("keys_7_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor keys_7_squeeze_mask_0 = const()[name = tensor("keys_7_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor keys_7_cast_fp16 = slice_by_index(begin = keys_7_begin_0, end = keys_7_end_0, end_mask = keys_7_end_mask_0, squeeze_mask = keys_7_squeeze_mask_0, x = reshape_19_cast_fp16)[name = tensor("keys_7_cast_fp16")]; + tensor values_7_begin_0 = const()[name = tensor("values_7_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor values_7_end_0 = const()[name = tensor("values_7_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor values_7_end_mask_0 = const()[name = tensor("values_7_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor values_7_squeeze_mask_0 = const()[name = tensor("values_7_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor values_7_cast_fp16 = slice_by_index(begin = values_7_begin_0, end = values_7_end_0, end_mask = values_7_end_mask_0, squeeze_mask = values_7_squeeze_mask_0, x = reshape_19_cast_fp16)[name = tensor("values_7_cast_fp16")]; + tensor var_843_cast_fp16 = not_equal(x = keys_7_cast_fp16, y = keys_7_cast_fp16)[name = tensor("op_843_cast_fp16")]; + tensor keys_9_cast_fp16 = select(a = var_463_to_fp16, b = keys_7_cast_fp16, cond = var_843_cast_fp16)[name = tensor("keys_9_cast_fp16")]; + tensor var_851_cast_fp16 = not_equal(x = values_7_cast_fp16, y = values_7_cast_fp16)[name = tensor("op_851_cast_fp16")]; + tensor values_9_cast_fp16 = select(a = var_463_to_fp16, b = values_7_cast_fp16, cond = var_851_cast_fp16)[name = tensor("values_9_cast_fp16")]; + tensor var_875 = const()[name = tensor("op_875"), val = tensor([0, 2, 1, 3])]; + tensor var_891 = const()[name = tensor("op_891"), val = tensor([1, 1, 1])]; + tensor var_892_cast_fp16 = reshape(shape = var_891, x = position1_to_fp16)[name = tensor("op_892_cast_fp16")]; + tensor var_912_to_fp16 = const()[name = tensor("op_912_to_fp16"), val = tensor(0x1p+0)]; + tensor valid_len_3_cast_fp16 = add(x = var_892_cast_fp16, y = var_912_to_fp16)[name = tensor("valid_len_3_cast_fp16")]; + tensor valid_mask_3_cast_fp16 = less(x = k_positions_1_promoted_to_fp16, y = valid_len_3_cast_fp16)[name = tensor("valid_mask_3_cast_fp16")]; + tensor causal_mask_3_cast_fp16 = less_equal(x = k_positions_1_promoted_to_fp16, y = var_892_cast_fp16)[name = tensor("causal_mask_3_cast_fp16")]; + tensor attn_mask_5 = logical_and(x = valid_mask_3_cast_fp16, y = causal_mask_3_cast_fp16)[name = tensor("attn_mask_5")]; + tensor attn_mask_7_axes_0 = const()[name = tensor("attn_mask_7_axes_0"), val = tensor([1])]; + tensor attn_mask_7 = expand_dims(axes = attn_mask_7_axes_0, x = attn_mask_5)[name = tensor("attn_mask_7")]; + tensor var_930_transpose_x_0 = const()[name = tensor("op_930_transpose_x_0"), val = tensor(false)]; + tensor var_930_transpose_y_0 = const()[name = tensor("op_930_transpose_y_0"), val = tensor(false)]; + tensor transpose_74_perm_0 = const()[name = tensor("transpose_74_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_75_perm_0 = const()[name = tensor("transpose_75_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_75 = transpose(perm = transpose_75_perm_0, x = keys_9_cast_fp16)[name = tensor("transpose_209")]; + tensor transpose_74 = transpose(perm = transpose_74_perm_0, x = q_9_cast_fp16)[name = tensor("transpose_210")]; + tensor var_930_cast_fp16 = matmul(transpose_x = var_930_transpose_x_0, transpose_y = var_930_transpose_y_0, x = transpose_74, y = transpose_75)[name = tensor("op_930_cast_fp16")]; + tensor var_924_to_fp16 = const()[name = tensor("op_924_to_fp16"), val = tensor([0x1p-3])]; + tensor attn_weights_7_cast_fp16 = mul(x = var_930_cast_fp16, y = var_924_to_fp16)[name = tensor("attn_weights_7_cast_fp16")]; + tensor var_932 = logical_not(x = attn_mask_7)[name = tensor("op_932")]; + tensor var_933_to_fp16 = const()[name = tensor("op_933_to_fp16"), val = tensor(-inf)]; + tensor attn_weights_9_cast_fp16 = select(a = var_933_to_fp16, b = attn_weights_7_cast_fp16, cond = var_932)[name = tensor("attn_weights_9_cast_fp16")]; + tensor var_935 = const()[name = tensor("op_935"), val = tensor(-1)]; + tensor attn_weights_11_cast_fp16 = softmax(axis = var_935, x = attn_weights_9_cast_fp16)[name = tensor("attn_weights_11_cast_fp16")]; + tensor attn_output_3_transpose_x_0 = const()[name = tensor("attn_output_3_transpose_x_0"), val = tensor(false)]; + tensor attn_output_3_transpose_y_0 = const()[name = tensor("attn_output_3_transpose_y_0"), val = tensor(false)]; + tensor values_11_cast_fp16 = transpose(perm = var_875, x = values_9_cast_fp16)[name = tensor("transpose_211")]; + tensor attn_output_3_cast_fp16 = matmul(transpose_x = attn_output_3_transpose_x_0, transpose_y = attn_output_3_transpose_y_0, x = attn_weights_11_cast_fp16, y = values_11_cast_fp16)[name = tensor("attn_output_3_cast_fp16")]; + tensor var_943 = const()[name = tensor("op_943"), val = tensor([0, 2, 1, 3])]; + tensor var_948 = const()[name = tensor("op_948"), val = tensor([1, 1, 1024])]; + tensor var_944_cast_fp16 = transpose(perm = var_943, x = attn_output_3_cast_fp16)[name = tensor("transpose_208")]; + tensor input_19_cast_fp16 = reshape(shape = var_948, x = var_944_cast_fp16)[name = tensor("input_19_cast_fp16")]; + tensor attn1_out_proj_weight_to_fp16 = const()[name = tensor("attn1_out_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(32602496)))]; + tensor linear_6_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = attn1_out_proj_weight_to_fp16, x = input_19_cast_fp16)[name = tensor("linear_6_cast_fp16")]; + tensor var_954_to_fp16 = const()[name = tensor("op_954_to_fp16"), val = tensor(0x1p+0)]; + tensor var_955_cast_fp16 = add(x = position1_to_fp16, y = var_954_to_fp16)[name = tensor("op_955_cast_fp16")]; + tensor var_955_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_955_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor input_21_cast_fp16 = add(x = input_15_cast_fp16, y = linear_6_cast_fp16)[name = tensor("input_21_cast_fp16")]; + tensor input_23_axes_0 = const()[name = tensor("input_23_axes_0"), val = tensor([-1])]; + tensor norm1_2_weight_to_fp16 = const()[name = tensor("norm1_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(34699712)))]; + tensor norm1_2_bias_to_fp16 = const()[name = tensor("norm1_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(34701824)))]; + tensor var_959_to_fp16 = const()[name = tensor("op_959_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_23_cast_fp16 = layer_norm(axes = input_23_axes_0, beta = norm1_2_bias_to_fp16, epsilon = var_959_to_fp16, gamma = norm1_2_weight_to_fp16, x = input_21_cast_fp16)[name = tensor("input_23_cast_fp16")]; + tensor linear1_1_weight_to_fp16 = const()[name = tensor("linear1_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(34703936)))]; + tensor linear_7_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = linear1_1_weight_to_fp16, x = input_23_cast_fp16)[name = tensor("linear_7_cast_fp16")]; + tensor input_25_mode_0 = const()[name = tensor("input_25_mode_0"), val = tensor("EXACT")]; + tensor input_25_cast_fp16 = gelu(mode = input_25_mode_0, x = linear_7_cast_fp16)[name = tensor("input_25_cast_fp16")]; + tensor linear1_2_weight_to_fp16 = const()[name = tensor("linear1_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(43092608)))]; + tensor linear_8_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = linear1_2_weight_to_fp16, x = input_25_cast_fp16)[name = tensor("linear_8_cast_fp16")]; + tensor input_27_cast_fp16 = add(x = input_21_cast_fp16, y = linear_8_cast_fp16)[name = tensor("input_27_cast_fp16")]; + tensor input_29_axes_0 = const()[name = tensor("input_29_axes_0"), val = tensor([-1])]; + tensor norm2_1_weight_to_fp16 = const()[name = tensor("norm2_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(51481280)))]; + tensor norm2_1_bias_to_fp16 = const()[name = tensor("norm2_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(51483392)))]; + tensor var_976_to_fp16 = const()[name = tensor("op_976_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_29_cast_fp16 = layer_norm(axes = input_29_axes_0, beta = norm2_1_bias_to_fp16, epsilon = var_976_to_fp16, gamma = norm2_1_weight_to_fp16, x = input_27_cast_fp16)[name = tensor("input_29_cast_fp16")]; + tensor attn2_in_proj_weight_to_fp16 = const()[name = tensor("attn2_in_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(51485504)))]; + tensor linear_9_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = attn2_in_proj_weight_to_fp16, x = input_29_cast_fp16)[name = tensor("linear_9_cast_fp16")]; + tensor var_990 = const()[name = tensor("op_990"), val = tensor([1, 1, 3, 16, 64])]; + tensor qkv_5_cast_fp16 = reshape(shape = var_990, x = linear_9_cast_fp16)[name = tensor("qkv_5_cast_fp16")]; + tensor q_13_begin_0 = const()[name = tensor("q_13_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor q_13_end_0 = const()[name = tensor("q_13_end_0"), val = tensor([1, 1, 1, 16, 64])]; + tensor q_13_end_mask_0 = const()[name = tensor("q_13_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor q_13_squeeze_mask_0 = const()[name = tensor("q_13_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor q_13_cast_fp16 = slice_by_index(begin = q_13_begin_0, end = q_13_end_0, end_mask = q_13_end_mask_0, squeeze_mask = q_13_squeeze_mask_0, x = qkv_5_cast_fp16)[name = tensor("q_13_cast_fp16")]; + tensor k_9_begin_0 = const()[name = tensor("k_9_begin_0"), val = tensor([0, 0, 1, 0, 0])]; + tensor k_9_end_0 = const()[name = tensor("k_9_end_0"), val = tensor([1, 1, 2, 16, 64])]; + tensor k_9_end_mask_0 = const()[name = tensor("k_9_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor k_9_squeeze_mask_0 = const()[name = tensor("k_9_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor k_9_cast_fp16 = slice_by_index(begin = k_9_begin_0, end = k_9_end_0, end_mask = k_9_end_mask_0, squeeze_mask = k_9_squeeze_mask_0, x = qkv_5_cast_fp16)[name = tensor("k_9_cast_fp16")]; + tensor v_5_begin_0 = const()[name = tensor("v_5_begin_0"), val = tensor([0, 0, 2, 0, 0])]; + tensor v_5_end_0 = const()[name = tensor("v_5_end_0"), val = tensor([1, 1, 3, 16, 64])]; + tensor v_5_end_mask_0 = const()[name = tensor("v_5_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor v_5_squeeze_mask_0 = const()[name = tensor("v_5_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor v_5_cast_fp16 = slice_by_index(begin = v_5_begin_0, end = v_5_end_0, end_mask = v_5_end_mask_0, squeeze_mask = v_5_squeeze_mask_0, x = qkv_5_cast_fp16)[name = tensor("v_5_cast_fp16")]; + tensor var_1056 = const()[name = tensor("op_1056"), val = tensor([1, 1, 1, 1])]; + tensor position2_to_fp16_dtype_0 = const()[name = tensor("position2_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor position2_to_fp16 = cast(dtype = position2_to_fp16_dtype_0, x = position2)[name = tensor("cast_543")]; + tensor ts_17_cast_fp16 = reshape(shape = var_1056, x = position2_to_fp16)[name = tensor("ts_17_cast_fp16")]; + tensor var_1063 = const()[name = tensor("op_1063"), val = tensor([1, 1, 16, 32, 2])]; + tensor q_complex_5_cast_fp16 = reshape(shape = var_1063, x = q_13_cast_fp16)[name = tensor("q_complex_5_cast_fp16")]; + tensor var_1070 = const()[name = tensor("op_1070"), val = tensor([1, 1, 16, 32, 2])]; + tensor k_complex_5_cast_fp16 = reshape(shape = var_1070, x = k_9_cast_fp16)[name = tensor("k_complex_5_cast_fp16")]; + tensor var_1074_begin_0 = const()[name = tensor("op_1074_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_1074_end_0 = const()[name = tensor("op_1074_end_0"), val = tensor([1, 1, 16, 32, 1])]; + tensor var_1074_end_mask_0 = const()[name = tensor("op_1074_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_1074_squeeze_mask_0 = const()[name = tensor("op_1074_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_1074_cast_fp16 = slice_by_index(begin = var_1074_begin_0, end = var_1074_end_0, end_mask = var_1074_end_mask_0, squeeze_mask = var_1074_squeeze_mask_0, x = q_complex_5_cast_fp16)[name = tensor("op_1074_cast_fp16")]; + tensor var_1082_begin_0 = const()[name = tensor("op_1082_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_1082_end_0 = const()[name = tensor("op_1082_end_0"), val = tensor([1, 1, 16, 32, 2])]; + tensor var_1082_end_mask_0 = const()[name = tensor("op_1082_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_1082_squeeze_mask_0 = const()[name = tensor("op_1082_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_1082_cast_fp16 = slice_by_index(begin = var_1082_begin_0, end = var_1082_end_0, end_mask = var_1082_end_mask_0, squeeze_mask = var_1082_squeeze_mask_0, x = q_complex_5_cast_fp16)[name = tensor("op_1082_cast_fp16")]; + tensor var_1090_begin_0 = const()[name = tensor("op_1090_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_1090_end_0 = const()[name = tensor("op_1090_end_0"), val = tensor([1, 1, 16, 32, 1])]; + tensor var_1090_end_mask_0 = const()[name = tensor("op_1090_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_1090_squeeze_mask_0 = const()[name = tensor("op_1090_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_1090_cast_fp16 = slice_by_index(begin = var_1090_begin_0, end = var_1090_end_0, end_mask = var_1090_end_mask_0, squeeze_mask = var_1090_squeeze_mask_0, x = k_complex_5_cast_fp16)[name = tensor("op_1090_cast_fp16")]; + tensor var_1098_begin_0 = const()[name = tensor("op_1098_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_1098_end_0 = const()[name = tensor("op_1098_end_0"), val = tensor([1, 1, 16, 32, 2])]; + tensor var_1098_end_mask_0 = const()[name = tensor("op_1098_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_1098_squeeze_mask_0 = const()[name = tensor("op_1098_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_1098_cast_fp16 = slice_by_index(begin = var_1098_begin_0, end = var_1098_end_0, end_mask = var_1098_end_mask_0, squeeze_mask = var_1098_squeeze_mask_0, x = k_complex_5_cast_fp16)[name = tensor("op_1098_cast_fp16")]; + tensor freqs_5_to_fp16 = const()[name = tensor("freqs_5_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(57777024)))]; + tensor var_1104_cast_fp16 = mul(x = freqs_5_to_fp16, y = ts_17_cast_fp16)[name = tensor("op_1104_cast_fp16")]; + tensor rotr_5_cast_fp16 = cos(x = var_1104_cast_fp16)[name = tensor("rotr_5_cast_fp16")]; + tensor roti_5_cast_fp16 = sin(x = var_1104_cast_fp16)[name = tensor("roti_5_cast_fp16")]; + tensor var_1108_cast_fp16 = mul(x = var_1074_cast_fp16, y = rotr_5_cast_fp16)[name = tensor("op_1108_cast_fp16")]; + tensor var_1109_cast_fp16 = mul(x = var_1082_cast_fp16, y = roti_5_cast_fp16)[name = tensor("op_1109_cast_fp16")]; + tensor qor_9_cast_fp16 = sub(x = var_1108_cast_fp16, y = var_1109_cast_fp16)[name = tensor("qor_9_cast_fp16")]; + tensor var_1112_cast_fp16 = mul(x = var_1074_cast_fp16, y = roti_5_cast_fp16)[name = tensor("op_1112_cast_fp16")]; + tensor var_1113_cast_fp16 = mul(x = var_1082_cast_fp16, y = rotr_5_cast_fp16)[name = tensor("op_1113_cast_fp16")]; + tensor qoi_9_cast_fp16 = add(x = var_1112_cast_fp16, y = var_1113_cast_fp16)[name = tensor("qoi_9_cast_fp16")]; + tensor var_1116_cast_fp16 = mul(x = var_1090_cast_fp16, y = rotr_5_cast_fp16)[name = tensor("op_1116_cast_fp16")]; + tensor var_1117_cast_fp16 = mul(x = var_1098_cast_fp16, y = roti_5_cast_fp16)[name = tensor("op_1117_cast_fp16")]; + tensor kor_9_cast_fp16 = sub(x = var_1116_cast_fp16, y = var_1117_cast_fp16)[name = tensor("kor_9_cast_fp16")]; + tensor var_1120_cast_fp16 = mul(x = var_1090_cast_fp16, y = roti_5_cast_fp16)[name = tensor("op_1120_cast_fp16")]; + tensor var_1121_cast_fp16 = mul(x = var_1098_cast_fp16, y = rotr_5_cast_fp16)[name = tensor("op_1121_cast_fp16")]; + tensor koi_9_cast_fp16 = add(x = var_1120_cast_fp16, y = var_1121_cast_fp16)[name = tensor("koi_9_cast_fp16")]; + tensor qo_5_axis_0 = const()[name = tensor("qo_5_axis_0"), val = tensor(-1)]; + tensor qo_5_cast_fp16 = stack(axis = qo_5_axis_0, values = (qor_9_cast_fp16, qoi_9_cast_fp16))[name = tensor("qo_5_cast_fp16")]; + tensor ko_5_axis_0 = const()[name = tensor("ko_5_axis_0"), val = tensor(-1)]; + tensor ko_5_cast_fp16 = stack(axis = ko_5_axis_0, values = (kor_9_cast_fp16, koi_9_cast_fp16))[name = tensor("ko_5_cast_fp16")]; + tensor var_1154 = const()[name = tensor("op_1154"), val = tensor([1, 1, 16, 64])]; + tensor q_15_cast_fp16 = reshape(shape = var_1154, x = qo_5_cast_fp16)[name = tensor("q_15_cast_fp16")]; + tensor var_1160 = const()[name = tensor("op_1160"), val = tensor([1, 1, 16, 64])]; + tensor k_11_cast_fp16 = reshape(shape = var_1160, x = ko_5_cast_fp16)[name = tensor("k_11_cast_fp16")]; + tensor _inversed_1185_y_0_to_fp16 = const()[name = tensor("_inversed_1185_y_0_to_fp16"), val = tensor(0x1p-9)]; + tensor _inversed_1185_cast_fp16 = mul(x = ts_17_cast_fp16, y = _inversed_1185_y_0_to_fp16)[name = tensor("_inversed_1185_cast_fp16")]; + tensor var_1186_cast_fp16 = floor(x = _inversed_1185_cast_fp16)[name = tensor("op_1186_cast_fp16")]; + tensor var_1187_to_fp16 = const()[name = tensor("op_1187_to_fp16"), val = tensor(0x1p+9)]; + tensor var_1188_cast_fp16 = mul(x = var_1186_cast_fp16, y = var_1187_to_fp16)[name = tensor("op_1188_cast_fp16")]; + tensor write_indices_float_11_cast_fp16 = sub(x = ts_17_cast_fp16, y = var_1188_cast_fp16)[name = tensor("write_indices_float_11_cast_fp16")]; + tensor var_1195_dtype_0 = const()[name = tensor("op_1195_dtype_0"), val = tensor("int32")]; + tensor write_indices_5_reps_0 = const()[name = tensor("write_indices_5_reps_0"), val = tensor([1, 1, 16, 64])]; + tensor write_indices_float_11_cast_fp16_to_int32 = cast(dtype = var_1195_dtype_0, x = write_indices_float_11_cast_fp16)[name = tensor("cast_542")]; + tensor write_indices_5 = tile(reps = write_indices_5_reps_0, x = write_indices_float_11_cast_fp16_to_int32)[name = tensor("write_indices_5")]; + tensor var_1205_begin_0 = const()[name = tensor("op_1205_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_1205_end_0 = const()[name = tensor("op_1205_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor var_1205_end_mask_0 = const()[name = tensor("op_1205_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_1205_squeeze_mask_0 = const()[name = tensor("op_1205_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor cache2_to_fp16_dtype_0 = const()[name = tensor("cache2_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor cache2_to_fp16 = cast(dtype = cache2_to_fp16_dtype_0, x = cache2)[name = tensor("cast_541")]; + tensor var_1205_cast_fp16 = slice_by_index(begin = var_1205_begin_0, end = var_1205_end_0, end_mask = var_1205_end_mask_0, squeeze_mask = var_1205_squeeze_mask_0, x = cache2_to_fp16)[name = tensor("op_1205_cast_fp16")]; + tensor var_1207_axis_0 = const()[name = tensor("op_1207_axis_0"), val = tensor(1)]; + tensor var_1207_mode_0 = const()[name = tensor("op_1207_mode_0"), val = tensor("update")]; + tensor var_1207_validate_indices_0 = const()[name = tensor("op_1207_validate_indices_0"), val = tensor(false)]; + tensor var_1207_cast_fp16 = scatter_along_axis(axis = var_1207_axis_0, data = var_1205_cast_fp16, indices = write_indices_5, mode = var_1207_mode_0, updates = k_11_cast_fp16, validate_indices = var_1207_validate_indices_0)[name = tensor("op_1207_cast_fp16")]; + tensor concat_16 = const()[name = tensor("concat_16"), val = tensor([0, 0, 0, 0, 0])]; + tensor concat_17 = const()[name = tensor("concat_17"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_5_internal_tensor_assign_1_stride_0 = const()[name = tensor("new_cache_5_internal_tensor_assign_1_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_5_internal_tensor_assign_1_begin_mask_0 = const()[name = tensor("new_cache_5_internal_tensor_assign_1_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_5_internal_tensor_assign_1_end_mask_0 = const()[name = tensor("new_cache_5_internal_tensor_assign_1_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_5_internal_tensor_assign_1_squeeze_mask_0 = const()[name = tensor("new_cache_5_internal_tensor_assign_1_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_52 = const()[name = tensor("shape_52"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_4 = const()[name = tensor("reduce_prod_4"), val = tensor(1048576)]; + tensor range_1d_4_start_0 = const()[name = tensor("range_1d_4_start_0"), val = tensor(0)]; + tensor range_1d_4_step_0 = const()[name = tensor("range_1d_4_step_0"), val = tensor(1)]; + tensor range_1d_4 = range_1d(end = reduce_prod_4, start = range_1d_4_start_0, step = range_1d_4_step_0)[name = tensor("range_1d_4")]; + tensor reshape_20 = reshape(shape = shape_52, x = range_1d_4)[name = tensor("reshape_20")]; + tensor slice_by_index_4 = slice_by_index(begin = concat_16, begin_mask = new_cache_5_internal_tensor_assign_1_begin_mask_0, end = concat_17, end_mask = new_cache_5_internal_tensor_assign_1_end_mask_0, squeeze_mask = new_cache_5_internal_tensor_assign_1_squeeze_mask_0, stride = new_cache_5_internal_tensor_assign_1_stride_0, x = reshape_20)[name = tensor("slice_by_index_4")]; + tensor reshape_21_shape_0 = const()[name = tensor("reshape_21_shape_0"), val = tensor([-1])]; + tensor reshape_21 = reshape(shape = reshape_21_shape_0, x = slice_by_index_4)[name = tensor("reshape_21")]; + tensor reshape_22_shape_0 = const()[name = tensor("reshape_22_shape_0"), val = tensor([-1])]; + tensor reshape_22_cast_fp16 = reshape(shape = reshape_22_shape_0, x = var_1207_cast_fp16)[name = tensor("reshape_22_cast_fp16")]; + tensor reshape_23_shape_0 = const()[name = tensor("reshape_23_shape_0"), val = tensor([-1])]; + tensor reshape_23_cast_fp16 = reshape(shape = reshape_23_shape_0, x = cache2_to_fp16)[name = tensor("reshape_23_cast_fp16")]; + tensor scatter_4_mode_0 = const()[name = tensor("scatter_4_mode_0"), val = tensor("update")]; + tensor scatter_4_axis_0 = const()[name = tensor("scatter_4_axis_0"), val = tensor(0)]; + tensor scatter_4_validate_indices_0 = const()[name = tensor("scatter_4_validate_indices_0"), val = tensor(false)]; + tensor scatter_4_cast_fp16 = scatter(axis = scatter_4_axis_0, data = reshape_23_cast_fp16, indices = reshape_21, mode = scatter_4_mode_0, updates = reshape_22_cast_fp16, validate_indices = scatter_4_validate_indices_0)[name = tensor("scatter_4_cast_fp16")]; + tensor reshape_24_cast_fp16 = reshape(shape = shape_52, x = scatter_4_cast_fp16)[name = tensor("reshape_24_cast_fp16")]; + tensor var_1215_begin_0 = const()[name = tensor("op_1215_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor var_1215_end_0 = const()[name = tensor("op_1215_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor var_1215_end_mask_0 = const()[name = tensor("op_1215_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_1215_squeeze_mask_0 = const()[name = tensor("op_1215_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor var_1215_cast_fp16 = slice_by_index(begin = var_1215_begin_0, end = var_1215_end_0, end_mask = var_1215_end_mask_0, squeeze_mask = var_1215_squeeze_mask_0, x = reshape_24_cast_fp16)[name = tensor("op_1215_cast_fp16")]; + tensor var_1217_axis_0 = const()[name = tensor("op_1217_axis_0"), val = tensor(1)]; + tensor var_1217_mode_0 = const()[name = tensor("op_1217_mode_0"), val = tensor("update")]; + tensor var_1217_validate_indices_0 = const()[name = tensor("op_1217_validate_indices_0"), val = tensor(false)]; + tensor var_1217_cast_fp16 = scatter_along_axis(axis = var_1217_axis_0, data = var_1215_cast_fp16, indices = write_indices_5, mode = var_1217_mode_0, updates = v_5_cast_fp16, validate_indices = var_1217_validate_indices_0)[name = tensor("op_1217_cast_fp16")]; + tensor concat_18 = const()[name = tensor("concat_18"), val = tensor([1, 0, 0, 0, 0])]; + tensor concat_19 = const()[name = tensor("concat_19"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_5_internal_tensor_assign_2_stride_0 = const()[name = tensor("new_cache_5_internal_tensor_assign_2_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_5_internal_tensor_assign_2_begin_mask_0 = const()[name = tensor("new_cache_5_internal_tensor_assign_2_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_5_internal_tensor_assign_2_end_mask_0 = const()[name = tensor("new_cache_5_internal_tensor_assign_2_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_5_internal_tensor_assign_2_squeeze_mask_0 = const()[name = tensor("new_cache_5_internal_tensor_assign_2_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_53 = const()[name = tensor("shape_53"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_5 = const()[name = tensor("reduce_prod_5"), val = tensor(1048576)]; + tensor range_1d_5_start_0 = const()[name = tensor("range_1d_5_start_0"), val = tensor(0)]; + tensor range_1d_5_step_0 = const()[name = tensor("range_1d_5_step_0"), val = tensor(1)]; + tensor range_1d_5 = range_1d(end = reduce_prod_5, start = range_1d_5_start_0, step = range_1d_5_step_0)[name = tensor("range_1d_5")]; + tensor reshape_25 = reshape(shape = shape_53, x = range_1d_5)[name = tensor("reshape_25")]; + tensor slice_by_index_5 = slice_by_index(begin = concat_18, begin_mask = new_cache_5_internal_tensor_assign_2_begin_mask_0, end = concat_19, end_mask = new_cache_5_internal_tensor_assign_2_end_mask_0, squeeze_mask = new_cache_5_internal_tensor_assign_2_squeeze_mask_0, stride = new_cache_5_internal_tensor_assign_2_stride_0, x = reshape_25)[name = tensor("slice_by_index_5")]; + tensor reshape_26_shape_0 = const()[name = tensor("reshape_26_shape_0"), val = tensor([-1])]; + tensor reshape_26 = reshape(shape = reshape_26_shape_0, x = slice_by_index_5)[name = tensor("reshape_26")]; + tensor reshape_27_shape_0 = const()[name = tensor("reshape_27_shape_0"), val = tensor([-1])]; + tensor reshape_27_cast_fp16 = reshape(shape = reshape_27_shape_0, x = var_1217_cast_fp16)[name = tensor("reshape_27_cast_fp16")]; + tensor reshape_28_shape_0 = const()[name = tensor("reshape_28_shape_0"), val = tensor([-1])]; + tensor reshape_28_cast_fp16 = reshape(shape = reshape_28_shape_0, x = reshape_24_cast_fp16)[name = tensor("reshape_28_cast_fp16")]; + tensor scatter_5_mode_0 = const()[name = tensor("scatter_5_mode_0"), val = tensor("update")]; + tensor scatter_5_axis_0 = const()[name = tensor("scatter_5_axis_0"), val = tensor(0)]; + tensor scatter_5_validate_indices_0 = const()[name = tensor("scatter_5_validate_indices_0"), val = tensor(false)]; + tensor scatter_5_cast_fp16 = scatter(axis = scatter_5_axis_0, data = reshape_28_cast_fp16, indices = reshape_26, mode = scatter_5_mode_0, updates = reshape_27_cast_fp16, validate_indices = scatter_5_validate_indices_0)[name = tensor("scatter_5_cast_fp16")]; + tensor reshape_29_cast_fp16 = reshape(shape = shape_53, x = scatter_5_cast_fp16)[name = tensor("reshape_29_cast_fp16")]; + tensor reshape_29_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("reshape_29_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor keys_13_begin_0 = const()[name = tensor("keys_13_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor keys_13_end_0 = const()[name = tensor("keys_13_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor keys_13_end_mask_0 = const()[name = tensor("keys_13_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor keys_13_squeeze_mask_0 = const()[name = tensor("keys_13_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor keys_13_cast_fp16 = slice_by_index(begin = keys_13_begin_0, end = keys_13_end_0, end_mask = keys_13_end_mask_0, squeeze_mask = keys_13_squeeze_mask_0, x = reshape_29_cast_fp16)[name = tensor("keys_13_cast_fp16")]; + tensor values_13_begin_0 = const()[name = tensor("values_13_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor values_13_end_0 = const()[name = tensor("values_13_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor values_13_end_mask_0 = const()[name = tensor("values_13_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor values_13_squeeze_mask_0 = const()[name = tensor("values_13_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor values_13_cast_fp16 = slice_by_index(begin = values_13_begin_0, end = values_13_end_0, end_mask = values_13_end_mask_0, squeeze_mask = values_13_squeeze_mask_0, x = reshape_29_cast_fp16)[name = tensor("values_13_cast_fp16")]; + tensor var_1229_cast_fp16 = not_equal(x = keys_13_cast_fp16, y = keys_13_cast_fp16)[name = tensor("op_1229_cast_fp16")]; + tensor keys_15_cast_fp16 = select(a = var_463_to_fp16, b = keys_13_cast_fp16, cond = var_1229_cast_fp16)[name = tensor("keys_15_cast_fp16")]; + tensor var_1237_cast_fp16 = not_equal(x = values_13_cast_fp16, y = values_13_cast_fp16)[name = tensor("op_1237_cast_fp16")]; + tensor values_15_cast_fp16 = select(a = var_463_to_fp16, b = values_13_cast_fp16, cond = var_1237_cast_fp16)[name = tensor("values_15_cast_fp16")]; + tensor var_1261 = const()[name = tensor("op_1261"), val = tensor([0, 2, 1, 3])]; + tensor var_1277 = const()[name = tensor("op_1277"), val = tensor([1, 1, 1])]; + tensor var_1278_cast_fp16 = reshape(shape = var_1277, x = position2_to_fp16)[name = tensor("op_1278_cast_fp16")]; + tensor var_1298_to_fp16 = const()[name = tensor("op_1298_to_fp16"), val = tensor(0x1p+0)]; + tensor valid_len_5_cast_fp16 = add(x = var_1278_cast_fp16, y = var_1298_to_fp16)[name = tensor("valid_len_5_cast_fp16")]; + tensor valid_mask_5_cast_fp16 = less(x = k_positions_1_promoted_to_fp16, y = valid_len_5_cast_fp16)[name = tensor("valid_mask_5_cast_fp16")]; + tensor causal_mask_5_cast_fp16 = less_equal(x = k_positions_1_promoted_to_fp16, y = var_1278_cast_fp16)[name = tensor("causal_mask_5_cast_fp16")]; + tensor attn_mask_9 = logical_and(x = valid_mask_5_cast_fp16, y = causal_mask_5_cast_fp16)[name = tensor("attn_mask_9")]; + tensor attn_mask_11_axes_0 = const()[name = tensor("attn_mask_11_axes_0"), val = tensor([1])]; + tensor attn_mask_11 = expand_dims(axes = attn_mask_11_axes_0, x = attn_mask_9)[name = tensor("attn_mask_11")]; + tensor var_1316_transpose_x_0 = const()[name = tensor("op_1316_transpose_x_0"), val = tensor(false)]; + tensor var_1316_transpose_y_0 = const()[name = tensor("op_1316_transpose_y_0"), val = tensor(false)]; + tensor transpose_76_perm_0 = const()[name = tensor("transpose_76_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_77_perm_0 = const()[name = tensor("transpose_77_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_77 = transpose(perm = transpose_77_perm_0, x = keys_15_cast_fp16)[name = tensor("transpose_205")]; + tensor transpose_76 = transpose(perm = transpose_76_perm_0, x = q_15_cast_fp16)[name = tensor("transpose_206")]; + tensor var_1316_cast_fp16 = matmul(transpose_x = var_1316_transpose_x_0, transpose_y = var_1316_transpose_y_0, x = transpose_76, y = transpose_77)[name = tensor("op_1316_cast_fp16")]; + tensor var_1310_to_fp16 = const()[name = tensor("op_1310_to_fp16"), val = tensor([0x1p-3])]; + tensor attn_weights_13_cast_fp16 = mul(x = var_1316_cast_fp16, y = var_1310_to_fp16)[name = tensor("attn_weights_13_cast_fp16")]; + tensor var_1318 = logical_not(x = attn_mask_11)[name = tensor("op_1318")]; + tensor var_1319_to_fp16 = const()[name = tensor("op_1319_to_fp16"), val = tensor(-inf)]; + tensor attn_weights_15_cast_fp16 = select(a = var_1319_to_fp16, b = attn_weights_13_cast_fp16, cond = var_1318)[name = tensor("attn_weights_15_cast_fp16")]; + tensor var_1321 = const()[name = tensor("op_1321"), val = tensor(-1)]; + tensor attn_weights_17_cast_fp16 = softmax(axis = var_1321, x = attn_weights_15_cast_fp16)[name = tensor("attn_weights_17_cast_fp16")]; + tensor attn_output_5_transpose_x_0 = const()[name = tensor("attn_output_5_transpose_x_0"), val = tensor(false)]; + tensor attn_output_5_transpose_y_0 = const()[name = tensor("attn_output_5_transpose_y_0"), val = tensor(false)]; + tensor values_17_cast_fp16 = transpose(perm = var_1261, x = values_15_cast_fp16)[name = tensor("transpose_207")]; + tensor attn_output_5_cast_fp16 = matmul(transpose_x = attn_output_5_transpose_x_0, transpose_y = attn_output_5_transpose_y_0, x = attn_weights_17_cast_fp16, y = values_17_cast_fp16)[name = tensor("attn_output_5_cast_fp16")]; + tensor var_1329 = const()[name = tensor("op_1329"), val = tensor([0, 2, 1, 3])]; + tensor var_1334 = const()[name = tensor("op_1334"), val = tensor([1, 1, 1024])]; + tensor var_1330_cast_fp16 = transpose(perm = var_1329, x = attn_output_5_cast_fp16)[name = tensor("transpose_204")]; + tensor input_31_cast_fp16 = reshape(shape = var_1334, x = var_1330_cast_fp16)[name = tensor("input_31_cast_fp16")]; + tensor attn2_out_proj_weight_to_fp16 = const()[name = tensor("attn2_out_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(57777152)))]; + tensor linear_10_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = attn2_out_proj_weight_to_fp16, x = input_31_cast_fp16)[name = tensor("linear_10_cast_fp16")]; + tensor var_1340_to_fp16 = const()[name = tensor("op_1340_to_fp16"), val = tensor(0x1p+0)]; + tensor var_1341_cast_fp16 = add(x = position2_to_fp16, y = var_1340_to_fp16)[name = tensor("op_1341_cast_fp16")]; + tensor var_1341_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_1341_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor input_33_cast_fp16 = add(x = input_27_cast_fp16, y = linear_10_cast_fp16)[name = tensor("input_33_cast_fp16")]; + tensor input_35_axes_0 = const()[name = tensor("input_35_axes_0"), val = tensor([-1])]; + tensor norm2_2_weight_to_fp16 = const()[name = tensor("norm2_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(59874368)))]; + tensor norm2_2_bias_to_fp16 = const()[name = tensor("norm2_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(59876480)))]; + tensor var_1345_to_fp16 = const()[name = tensor("op_1345_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_35_cast_fp16 = layer_norm(axes = input_35_axes_0, beta = norm2_2_bias_to_fp16, epsilon = var_1345_to_fp16, gamma = norm2_2_weight_to_fp16, x = input_33_cast_fp16)[name = tensor("input_35_cast_fp16")]; + tensor linear2_1_weight_to_fp16 = const()[name = tensor("linear2_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(59878592)))]; + tensor linear_11_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = linear2_1_weight_to_fp16, x = input_35_cast_fp16)[name = tensor("linear_11_cast_fp16")]; + tensor input_37_mode_0 = const()[name = tensor("input_37_mode_0"), val = tensor("EXACT")]; + tensor input_37_cast_fp16 = gelu(mode = input_37_mode_0, x = linear_11_cast_fp16)[name = tensor("input_37_cast_fp16")]; + tensor linear2_2_weight_to_fp16 = const()[name = tensor("linear2_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(68267264)))]; + tensor linear_12_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = linear2_2_weight_to_fp16, x = input_37_cast_fp16)[name = tensor("linear_12_cast_fp16")]; + tensor input_39_cast_fp16 = add(x = input_33_cast_fp16, y = linear_12_cast_fp16)[name = tensor("input_39_cast_fp16")]; + tensor input_41_axes_0 = const()[name = tensor("input_41_axes_0"), val = tensor([-1])]; + tensor norm3_1_weight_to_fp16 = const()[name = tensor("norm3_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(76655936)))]; + tensor norm3_1_bias_to_fp16 = const()[name = tensor("norm3_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(76658048)))]; + tensor var_1362_to_fp16 = const()[name = tensor("op_1362_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_41_cast_fp16 = layer_norm(axes = input_41_axes_0, beta = norm3_1_bias_to_fp16, epsilon = var_1362_to_fp16, gamma = norm3_1_weight_to_fp16, x = input_39_cast_fp16)[name = tensor("input_41_cast_fp16")]; + tensor attn3_in_proj_weight_to_fp16 = const()[name = tensor("attn3_in_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(76660160)))]; + tensor linear_13_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = attn3_in_proj_weight_to_fp16, x = input_41_cast_fp16)[name = tensor("linear_13_cast_fp16")]; + tensor var_1376 = const()[name = tensor("op_1376"), val = tensor([1, 1, 3, 16, 64])]; + tensor qkv_7_cast_fp16 = reshape(shape = var_1376, x = linear_13_cast_fp16)[name = tensor("qkv_7_cast_fp16")]; + tensor q_19_begin_0 = const()[name = tensor("q_19_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor q_19_end_0 = const()[name = tensor("q_19_end_0"), val = tensor([1, 1, 1, 16, 64])]; + tensor q_19_end_mask_0 = const()[name = tensor("q_19_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor q_19_squeeze_mask_0 = const()[name = tensor("q_19_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor q_19_cast_fp16 = slice_by_index(begin = q_19_begin_0, end = q_19_end_0, end_mask = q_19_end_mask_0, squeeze_mask = q_19_squeeze_mask_0, x = qkv_7_cast_fp16)[name = tensor("q_19_cast_fp16")]; + tensor k_13_begin_0 = const()[name = tensor("k_13_begin_0"), val = tensor([0, 0, 1, 0, 0])]; + tensor k_13_end_0 = const()[name = tensor("k_13_end_0"), val = tensor([1, 1, 2, 16, 64])]; + tensor k_13_end_mask_0 = const()[name = tensor("k_13_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor k_13_squeeze_mask_0 = const()[name = tensor("k_13_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor k_13_cast_fp16 = slice_by_index(begin = k_13_begin_0, end = k_13_end_0, end_mask = k_13_end_mask_0, squeeze_mask = k_13_squeeze_mask_0, x = qkv_7_cast_fp16)[name = tensor("k_13_cast_fp16")]; + tensor v_7_begin_0 = const()[name = tensor("v_7_begin_0"), val = tensor([0, 0, 2, 0, 0])]; + tensor v_7_end_0 = const()[name = tensor("v_7_end_0"), val = tensor([1, 1, 3, 16, 64])]; + tensor v_7_end_mask_0 = const()[name = tensor("v_7_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor v_7_squeeze_mask_0 = const()[name = tensor("v_7_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor v_7_cast_fp16 = slice_by_index(begin = v_7_begin_0, end = v_7_end_0, end_mask = v_7_end_mask_0, squeeze_mask = v_7_squeeze_mask_0, x = qkv_7_cast_fp16)[name = tensor("v_7_cast_fp16")]; + tensor var_1442 = const()[name = tensor("op_1442"), val = tensor([1, 1, 1, 1])]; + tensor position3_to_fp16_dtype_0 = const()[name = tensor("position3_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor position3_to_fp16 = cast(dtype = position3_to_fp16_dtype_0, x = position3)[name = tensor("cast_538")]; + tensor ts_23_cast_fp16 = reshape(shape = var_1442, x = position3_to_fp16)[name = tensor("ts_23_cast_fp16")]; + tensor var_1449 = const()[name = tensor("op_1449"), val = tensor([1, 1, 16, 32, 2])]; + tensor q_complex_7_cast_fp16 = reshape(shape = var_1449, x = q_19_cast_fp16)[name = tensor("q_complex_7_cast_fp16")]; + tensor var_1456 = const()[name = tensor("op_1456"), val = tensor([1, 1, 16, 32, 2])]; + tensor k_complex_7_cast_fp16 = reshape(shape = var_1456, x = k_13_cast_fp16)[name = tensor("k_complex_7_cast_fp16")]; + tensor var_1460_begin_0 = const()[name = tensor("op_1460_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_1460_end_0 = const()[name = tensor("op_1460_end_0"), val = tensor([1, 1, 16, 32, 1])]; + tensor var_1460_end_mask_0 = const()[name = tensor("op_1460_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_1460_squeeze_mask_0 = const()[name = tensor("op_1460_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_1460_cast_fp16 = slice_by_index(begin = var_1460_begin_0, end = var_1460_end_0, end_mask = var_1460_end_mask_0, squeeze_mask = var_1460_squeeze_mask_0, x = q_complex_7_cast_fp16)[name = tensor("op_1460_cast_fp16")]; + tensor var_1468_begin_0 = const()[name = tensor("op_1468_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_1468_end_0 = const()[name = tensor("op_1468_end_0"), val = tensor([1, 1, 16, 32, 2])]; + tensor var_1468_end_mask_0 = const()[name = tensor("op_1468_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_1468_squeeze_mask_0 = const()[name = tensor("op_1468_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_1468_cast_fp16 = slice_by_index(begin = var_1468_begin_0, end = var_1468_end_0, end_mask = var_1468_end_mask_0, squeeze_mask = var_1468_squeeze_mask_0, x = q_complex_7_cast_fp16)[name = tensor("op_1468_cast_fp16")]; + tensor var_1476_begin_0 = const()[name = tensor("op_1476_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_1476_end_0 = const()[name = tensor("op_1476_end_0"), val = tensor([1, 1, 16, 32, 1])]; + tensor var_1476_end_mask_0 = const()[name = tensor("op_1476_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_1476_squeeze_mask_0 = const()[name = tensor("op_1476_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_1476_cast_fp16 = slice_by_index(begin = var_1476_begin_0, end = var_1476_end_0, end_mask = var_1476_end_mask_0, squeeze_mask = var_1476_squeeze_mask_0, x = k_complex_7_cast_fp16)[name = tensor("op_1476_cast_fp16")]; + tensor var_1484_begin_0 = const()[name = tensor("op_1484_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_1484_end_0 = const()[name = tensor("op_1484_end_0"), val = tensor([1, 1, 16, 32, 2])]; + tensor var_1484_end_mask_0 = const()[name = tensor("op_1484_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_1484_squeeze_mask_0 = const()[name = tensor("op_1484_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_1484_cast_fp16 = slice_by_index(begin = var_1484_begin_0, end = var_1484_end_0, end_mask = var_1484_end_mask_0, squeeze_mask = var_1484_squeeze_mask_0, x = k_complex_7_cast_fp16)[name = tensor("op_1484_cast_fp16")]; + tensor freqs_7_to_fp16 = const()[name = tensor("freqs_7_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(82951680)))]; + tensor var_1490_cast_fp16 = mul(x = freqs_7_to_fp16, y = ts_23_cast_fp16)[name = tensor("op_1490_cast_fp16")]; + tensor rotr_7_cast_fp16 = cos(x = var_1490_cast_fp16)[name = tensor("rotr_7_cast_fp16")]; + tensor roti_7_cast_fp16 = sin(x = var_1490_cast_fp16)[name = tensor("roti_7_cast_fp16")]; + tensor var_1494_cast_fp16 = mul(x = var_1460_cast_fp16, y = rotr_7_cast_fp16)[name = tensor("op_1494_cast_fp16")]; + tensor var_1495_cast_fp16 = mul(x = var_1468_cast_fp16, y = roti_7_cast_fp16)[name = tensor("op_1495_cast_fp16")]; + tensor qor_13_cast_fp16 = sub(x = var_1494_cast_fp16, y = var_1495_cast_fp16)[name = tensor("qor_13_cast_fp16")]; + tensor var_1498_cast_fp16 = mul(x = var_1460_cast_fp16, y = roti_7_cast_fp16)[name = tensor("op_1498_cast_fp16")]; + tensor var_1499_cast_fp16 = mul(x = var_1468_cast_fp16, y = rotr_7_cast_fp16)[name = tensor("op_1499_cast_fp16")]; + tensor qoi_13_cast_fp16 = add(x = var_1498_cast_fp16, y = var_1499_cast_fp16)[name = tensor("qoi_13_cast_fp16")]; + tensor var_1502_cast_fp16 = mul(x = var_1476_cast_fp16, y = rotr_7_cast_fp16)[name = tensor("op_1502_cast_fp16")]; + tensor var_1503_cast_fp16 = mul(x = var_1484_cast_fp16, y = roti_7_cast_fp16)[name = tensor("op_1503_cast_fp16")]; + tensor kor_13_cast_fp16 = sub(x = var_1502_cast_fp16, y = var_1503_cast_fp16)[name = tensor("kor_13_cast_fp16")]; + tensor var_1506_cast_fp16 = mul(x = var_1476_cast_fp16, y = roti_7_cast_fp16)[name = tensor("op_1506_cast_fp16")]; + tensor var_1507_cast_fp16 = mul(x = var_1484_cast_fp16, y = rotr_7_cast_fp16)[name = tensor("op_1507_cast_fp16")]; + tensor koi_13_cast_fp16 = add(x = var_1506_cast_fp16, y = var_1507_cast_fp16)[name = tensor("koi_13_cast_fp16")]; + tensor qo_7_axis_0 = const()[name = tensor("qo_7_axis_0"), val = tensor(-1)]; + tensor qo_7_cast_fp16 = stack(axis = qo_7_axis_0, values = (qor_13_cast_fp16, qoi_13_cast_fp16))[name = tensor("qo_7_cast_fp16")]; + tensor ko_7_axis_0 = const()[name = tensor("ko_7_axis_0"), val = tensor(-1)]; + tensor ko_7_cast_fp16 = stack(axis = ko_7_axis_0, values = (kor_13_cast_fp16, koi_13_cast_fp16))[name = tensor("ko_7_cast_fp16")]; + tensor var_1540 = const()[name = tensor("op_1540"), val = tensor([1, 1, 16, 64])]; + tensor q_21_cast_fp16 = reshape(shape = var_1540, x = qo_7_cast_fp16)[name = tensor("q_21_cast_fp16")]; + tensor var_1546 = const()[name = tensor("op_1546"), val = tensor([1, 1, 16, 64])]; + tensor k_15_cast_fp16 = reshape(shape = var_1546, x = ko_7_cast_fp16)[name = tensor("k_15_cast_fp16")]; + tensor _inversed_1571_y_0_to_fp16 = const()[name = tensor("_inversed_1571_y_0_to_fp16"), val = tensor(0x1p-9)]; + tensor _inversed_1571_cast_fp16 = mul(x = ts_23_cast_fp16, y = _inversed_1571_y_0_to_fp16)[name = tensor("_inversed_1571_cast_fp16")]; + tensor var_1572_cast_fp16 = floor(x = _inversed_1571_cast_fp16)[name = tensor("op_1572_cast_fp16")]; + tensor var_1573_to_fp16 = const()[name = tensor("op_1573_to_fp16"), val = tensor(0x1p+9)]; + tensor var_1574_cast_fp16 = mul(x = var_1572_cast_fp16, y = var_1573_to_fp16)[name = tensor("op_1574_cast_fp16")]; + tensor write_indices_float_15_cast_fp16 = sub(x = ts_23_cast_fp16, y = var_1574_cast_fp16)[name = tensor("write_indices_float_15_cast_fp16")]; + tensor var_1581_dtype_0 = const()[name = tensor("op_1581_dtype_0"), val = tensor("int32")]; + tensor write_indices_7_reps_0 = const()[name = tensor("write_indices_7_reps_0"), val = tensor([1, 1, 16, 64])]; + tensor write_indices_float_15_cast_fp16_to_int32 = cast(dtype = var_1581_dtype_0, x = write_indices_float_15_cast_fp16)[name = tensor("cast_537")]; + tensor write_indices_7 = tile(reps = write_indices_7_reps_0, x = write_indices_float_15_cast_fp16_to_int32)[name = tensor("write_indices_7")]; + tensor var_1591_begin_0 = const()[name = tensor("op_1591_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_1591_end_0 = const()[name = tensor("op_1591_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor var_1591_end_mask_0 = const()[name = tensor("op_1591_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_1591_squeeze_mask_0 = const()[name = tensor("op_1591_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor cache3_to_fp16_dtype_0 = const()[name = tensor("cache3_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor cache3_to_fp16 = cast(dtype = cache3_to_fp16_dtype_0, x = cache3)[name = tensor("cast_536")]; + tensor var_1591_cast_fp16 = slice_by_index(begin = var_1591_begin_0, end = var_1591_end_0, end_mask = var_1591_end_mask_0, squeeze_mask = var_1591_squeeze_mask_0, x = cache3_to_fp16)[name = tensor("op_1591_cast_fp16")]; + tensor var_1593_axis_0 = const()[name = tensor("op_1593_axis_0"), val = tensor(1)]; + tensor var_1593_mode_0 = const()[name = tensor("op_1593_mode_0"), val = tensor("update")]; + tensor var_1593_validate_indices_0 = const()[name = tensor("op_1593_validate_indices_0"), val = tensor(false)]; + tensor var_1593_cast_fp16 = scatter_along_axis(axis = var_1593_axis_0, data = var_1591_cast_fp16, indices = write_indices_7, mode = var_1593_mode_0, updates = k_15_cast_fp16, validate_indices = var_1593_validate_indices_0)[name = tensor("op_1593_cast_fp16")]; + tensor concat_23 = const()[name = tensor("concat_23"), val = tensor([0, 0, 0, 0, 0])]; + tensor concat_24 = const()[name = tensor("concat_24"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_7_internal_tensor_assign_1_stride_0 = const()[name = tensor("new_cache_7_internal_tensor_assign_1_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_7_internal_tensor_assign_1_begin_mask_0 = const()[name = tensor("new_cache_7_internal_tensor_assign_1_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_7_internal_tensor_assign_1_end_mask_0 = const()[name = tensor("new_cache_7_internal_tensor_assign_1_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_7_internal_tensor_assign_1_squeeze_mask_0 = const()[name = tensor("new_cache_7_internal_tensor_assign_1_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_54 = const()[name = tensor("shape_54"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_6 = const()[name = tensor("reduce_prod_6"), val = tensor(1048576)]; + tensor range_1d_6_start_0 = const()[name = tensor("range_1d_6_start_0"), val = tensor(0)]; + tensor range_1d_6_step_0 = const()[name = tensor("range_1d_6_step_0"), val = tensor(1)]; + tensor range_1d_6 = range_1d(end = reduce_prod_6, start = range_1d_6_start_0, step = range_1d_6_step_0)[name = tensor("range_1d_6")]; + tensor reshape_30 = reshape(shape = shape_54, x = range_1d_6)[name = tensor("reshape_30")]; + tensor slice_by_index_6 = slice_by_index(begin = concat_23, begin_mask = new_cache_7_internal_tensor_assign_1_begin_mask_0, end = concat_24, end_mask = new_cache_7_internal_tensor_assign_1_end_mask_0, squeeze_mask = new_cache_7_internal_tensor_assign_1_squeeze_mask_0, stride = new_cache_7_internal_tensor_assign_1_stride_0, x = reshape_30)[name = tensor("slice_by_index_6")]; + tensor reshape_31_shape_0 = const()[name = tensor("reshape_31_shape_0"), val = tensor([-1])]; + tensor reshape_31 = reshape(shape = reshape_31_shape_0, x = slice_by_index_6)[name = tensor("reshape_31")]; + tensor reshape_32_shape_0 = const()[name = tensor("reshape_32_shape_0"), val = tensor([-1])]; + tensor reshape_32_cast_fp16 = reshape(shape = reshape_32_shape_0, x = var_1593_cast_fp16)[name = tensor("reshape_32_cast_fp16")]; + tensor reshape_33_shape_0 = const()[name = tensor("reshape_33_shape_0"), val = tensor([-1])]; + tensor reshape_33_cast_fp16 = reshape(shape = reshape_33_shape_0, x = cache3_to_fp16)[name = tensor("reshape_33_cast_fp16")]; + tensor scatter_6_mode_0 = const()[name = tensor("scatter_6_mode_0"), val = tensor("update")]; + tensor scatter_6_axis_0 = const()[name = tensor("scatter_6_axis_0"), val = tensor(0)]; + tensor scatter_6_validate_indices_0 = const()[name = tensor("scatter_6_validate_indices_0"), val = tensor(false)]; + tensor scatter_6_cast_fp16 = scatter(axis = scatter_6_axis_0, data = reshape_33_cast_fp16, indices = reshape_31, mode = scatter_6_mode_0, updates = reshape_32_cast_fp16, validate_indices = scatter_6_validate_indices_0)[name = tensor("scatter_6_cast_fp16")]; + tensor reshape_34_cast_fp16 = reshape(shape = shape_54, x = scatter_6_cast_fp16)[name = tensor("reshape_34_cast_fp16")]; + tensor var_1601_begin_0 = const()[name = tensor("op_1601_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor var_1601_end_0 = const()[name = tensor("op_1601_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor var_1601_end_mask_0 = const()[name = tensor("op_1601_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_1601_squeeze_mask_0 = const()[name = tensor("op_1601_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor var_1601_cast_fp16 = slice_by_index(begin = var_1601_begin_0, end = var_1601_end_0, end_mask = var_1601_end_mask_0, squeeze_mask = var_1601_squeeze_mask_0, x = reshape_34_cast_fp16)[name = tensor("op_1601_cast_fp16")]; + tensor var_1603_axis_0 = const()[name = tensor("op_1603_axis_0"), val = tensor(1)]; + tensor var_1603_mode_0 = const()[name = tensor("op_1603_mode_0"), val = tensor("update")]; + tensor var_1603_validate_indices_0 = const()[name = tensor("op_1603_validate_indices_0"), val = tensor(false)]; + tensor var_1603_cast_fp16 = scatter_along_axis(axis = var_1603_axis_0, data = var_1601_cast_fp16, indices = write_indices_7, mode = var_1603_mode_0, updates = v_7_cast_fp16, validate_indices = var_1603_validate_indices_0)[name = tensor("op_1603_cast_fp16")]; + tensor concat_25 = const()[name = tensor("concat_25"), val = tensor([1, 0, 0, 0, 0])]; + tensor concat_26 = const()[name = tensor("concat_26"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_7_internal_tensor_assign_2_stride_0 = const()[name = tensor("new_cache_7_internal_tensor_assign_2_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_7_internal_tensor_assign_2_begin_mask_0 = const()[name = tensor("new_cache_7_internal_tensor_assign_2_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_7_internal_tensor_assign_2_end_mask_0 = const()[name = tensor("new_cache_7_internal_tensor_assign_2_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_7_internal_tensor_assign_2_squeeze_mask_0 = const()[name = tensor("new_cache_7_internal_tensor_assign_2_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_55 = const()[name = tensor("shape_55"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_7 = const()[name = tensor("reduce_prod_7"), val = tensor(1048576)]; + tensor range_1d_7_start_0 = const()[name = tensor("range_1d_7_start_0"), val = tensor(0)]; + tensor range_1d_7_step_0 = const()[name = tensor("range_1d_7_step_0"), val = tensor(1)]; + tensor range_1d_7 = range_1d(end = reduce_prod_7, start = range_1d_7_start_0, step = range_1d_7_step_0)[name = tensor("range_1d_7")]; + tensor reshape_35 = reshape(shape = shape_55, x = range_1d_7)[name = tensor("reshape_35")]; + tensor slice_by_index_7 = slice_by_index(begin = concat_25, begin_mask = new_cache_7_internal_tensor_assign_2_begin_mask_0, end = concat_26, end_mask = new_cache_7_internal_tensor_assign_2_end_mask_0, squeeze_mask = new_cache_7_internal_tensor_assign_2_squeeze_mask_0, stride = new_cache_7_internal_tensor_assign_2_stride_0, x = reshape_35)[name = tensor("slice_by_index_7")]; + tensor reshape_36_shape_0 = const()[name = tensor("reshape_36_shape_0"), val = tensor([-1])]; + tensor reshape_36 = reshape(shape = reshape_36_shape_0, x = slice_by_index_7)[name = tensor("reshape_36")]; + tensor reshape_37_shape_0 = const()[name = tensor("reshape_37_shape_0"), val = tensor([-1])]; + tensor reshape_37_cast_fp16 = reshape(shape = reshape_37_shape_0, x = var_1603_cast_fp16)[name = tensor("reshape_37_cast_fp16")]; + tensor reshape_38_shape_0 = const()[name = tensor("reshape_38_shape_0"), val = tensor([-1])]; + tensor reshape_38_cast_fp16 = reshape(shape = reshape_38_shape_0, x = reshape_34_cast_fp16)[name = tensor("reshape_38_cast_fp16")]; + tensor scatter_7_mode_0 = const()[name = tensor("scatter_7_mode_0"), val = tensor("update")]; + tensor scatter_7_axis_0 = const()[name = tensor("scatter_7_axis_0"), val = tensor(0)]; + tensor scatter_7_validate_indices_0 = const()[name = tensor("scatter_7_validate_indices_0"), val = tensor(false)]; + tensor scatter_7_cast_fp16 = scatter(axis = scatter_7_axis_0, data = reshape_38_cast_fp16, indices = reshape_36, mode = scatter_7_mode_0, updates = reshape_37_cast_fp16, validate_indices = scatter_7_validate_indices_0)[name = tensor("scatter_7_cast_fp16")]; + tensor reshape_39_cast_fp16 = reshape(shape = shape_55, x = scatter_7_cast_fp16)[name = tensor("reshape_39_cast_fp16")]; + tensor reshape_39_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("reshape_39_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor keys_19_begin_0 = const()[name = tensor("keys_19_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor keys_19_end_0 = const()[name = tensor("keys_19_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor keys_19_end_mask_0 = const()[name = tensor("keys_19_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor keys_19_squeeze_mask_0 = const()[name = tensor("keys_19_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor keys_19_cast_fp16 = slice_by_index(begin = keys_19_begin_0, end = keys_19_end_0, end_mask = keys_19_end_mask_0, squeeze_mask = keys_19_squeeze_mask_0, x = reshape_39_cast_fp16)[name = tensor("keys_19_cast_fp16")]; + tensor values_19_begin_0 = const()[name = tensor("values_19_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor values_19_end_0 = const()[name = tensor("values_19_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor values_19_end_mask_0 = const()[name = tensor("values_19_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor values_19_squeeze_mask_0 = const()[name = tensor("values_19_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor values_19_cast_fp16 = slice_by_index(begin = values_19_begin_0, end = values_19_end_0, end_mask = values_19_end_mask_0, squeeze_mask = values_19_squeeze_mask_0, x = reshape_39_cast_fp16)[name = tensor("values_19_cast_fp16")]; + tensor var_1615_cast_fp16 = not_equal(x = keys_19_cast_fp16, y = keys_19_cast_fp16)[name = tensor("op_1615_cast_fp16")]; + tensor keys_21_cast_fp16 = select(a = var_463_to_fp16, b = keys_19_cast_fp16, cond = var_1615_cast_fp16)[name = tensor("keys_21_cast_fp16")]; + tensor var_1623_cast_fp16 = not_equal(x = values_19_cast_fp16, y = values_19_cast_fp16)[name = tensor("op_1623_cast_fp16")]; + tensor values_21_cast_fp16 = select(a = var_463_to_fp16, b = values_19_cast_fp16, cond = var_1623_cast_fp16)[name = tensor("values_21_cast_fp16")]; + tensor var_1647 = const()[name = tensor("op_1647"), val = tensor([0, 2, 1, 3])]; + tensor var_1663 = const()[name = tensor("op_1663"), val = tensor([1, 1, 1])]; + tensor var_1664_cast_fp16 = reshape(shape = var_1663, x = position3_to_fp16)[name = tensor("op_1664_cast_fp16")]; + tensor var_1684_to_fp16 = const()[name = tensor("op_1684_to_fp16"), val = tensor(0x1p+0)]; + tensor valid_len_7_cast_fp16 = add(x = var_1664_cast_fp16, y = var_1684_to_fp16)[name = tensor("valid_len_7_cast_fp16")]; + tensor valid_mask_7_cast_fp16 = less(x = k_positions_1_promoted_to_fp16, y = valid_len_7_cast_fp16)[name = tensor("valid_mask_7_cast_fp16")]; + tensor causal_mask_7_cast_fp16 = less_equal(x = k_positions_1_promoted_to_fp16, y = var_1664_cast_fp16)[name = tensor("causal_mask_7_cast_fp16")]; + tensor attn_mask_13 = logical_and(x = valid_mask_7_cast_fp16, y = causal_mask_7_cast_fp16)[name = tensor("attn_mask_13")]; + tensor attn_mask_15_axes_0 = const()[name = tensor("attn_mask_15_axes_0"), val = tensor([1])]; + tensor attn_mask_15 = expand_dims(axes = attn_mask_15_axes_0, x = attn_mask_13)[name = tensor("attn_mask_15")]; + tensor var_1702_transpose_x_0 = const()[name = tensor("op_1702_transpose_x_0"), val = tensor(false)]; + tensor var_1702_transpose_y_0 = const()[name = tensor("op_1702_transpose_y_0"), val = tensor(false)]; + tensor transpose_78_perm_0 = const()[name = tensor("transpose_78_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_79_perm_0 = const()[name = tensor("transpose_79_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_79 = transpose(perm = transpose_79_perm_0, x = keys_21_cast_fp16)[name = tensor("transpose_201")]; + tensor transpose_78 = transpose(perm = transpose_78_perm_0, x = q_21_cast_fp16)[name = tensor("transpose_202")]; + tensor var_1702_cast_fp16 = matmul(transpose_x = var_1702_transpose_x_0, transpose_y = var_1702_transpose_y_0, x = transpose_78, y = transpose_79)[name = tensor("op_1702_cast_fp16")]; + tensor var_1696_to_fp16 = const()[name = tensor("op_1696_to_fp16"), val = tensor([0x1p-3])]; + tensor attn_weights_19_cast_fp16 = mul(x = var_1702_cast_fp16, y = var_1696_to_fp16)[name = tensor("attn_weights_19_cast_fp16")]; + tensor var_1704 = logical_not(x = attn_mask_15)[name = tensor("op_1704")]; + tensor var_1705_to_fp16 = const()[name = tensor("op_1705_to_fp16"), val = tensor(-inf)]; + tensor attn_weights_21_cast_fp16 = select(a = var_1705_to_fp16, b = attn_weights_19_cast_fp16, cond = var_1704)[name = tensor("attn_weights_21_cast_fp16")]; + tensor var_1707 = const()[name = tensor("op_1707"), val = tensor(-1)]; + tensor attn_weights_23_cast_fp16 = softmax(axis = var_1707, x = attn_weights_21_cast_fp16)[name = tensor("attn_weights_23_cast_fp16")]; + tensor attn_output_7_transpose_x_0 = const()[name = tensor("attn_output_7_transpose_x_0"), val = tensor(false)]; + tensor attn_output_7_transpose_y_0 = const()[name = tensor("attn_output_7_transpose_y_0"), val = tensor(false)]; + tensor values_23_cast_fp16 = transpose(perm = var_1647, x = values_21_cast_fp16)[name = tensor("transpose_203")]; + tensor attn_output_7_cast_fp16 = matmul(transpose_x = attn_output_7_transpose_x_0, transpose_y = attn_output_7_transpose_y_0, x = attn_weights_23_cast_fp16, y = values_23_cast_fp16)[name = tensor("attn_output_7_cast_fp16")]; + tensor var_1715 = const()[name = tensor("op_1715"), val = tensor([0, 2, 1, 3])]; + tensor var_1720 = const()[name = tensor("op_1720"), val = tensor([1, 1, 1024])]; + tensor var_1716_cast_fp16 = transpose(perm = var_1715, x = attn_output_7_cast_fp16)[name = tensor("transpose_200")]; + tensor input_43_cast_fp16 = reshape(shape = var_1720, x = var_1716_cast_fp16)[name = tensor("input_43_cast_fp16")]; + tensor attn3_out_proj_weight_to_fp16 = const()[name = tensor("attn3_out_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(82951808)))]; + tensor linear_14_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = attn3_out_proj_weight_to_fp16, x = input_43_cast_fp16)[name = tensor("linear_14_cast_fp16")]; + tensor var_1726_to_fp16 = const()[name = tensor("op_1726_to_fp16"), val = tensor(0x1p+0)]; + tensor var_1727_cast_fp16 = add(x = position3_to_fp16, y = var_1726_to_fp16)[name = tensor("op_1727_cast_fp16")]; + tensor var_1727_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_1727_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor input_45_cast_fp16 = add(x = input_39_cast_fp16, y = linear_14_cast_fp16)[name = tensor("input_45_cast_fp16")]; + tensor input_47_axes_0 = const()[name = tensor("input_47_axes_0"), val = tensor([-1])]; + tensor norm3_2_weight_to_fp16 = const()[name = tensor("norm3_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(85049024)))]; + tensor norm3_2_bias_to_fp16 = const()[name = tensor("norm3_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(85051136)))]; + tensor var_1731_to_fp16 = const()[name = tensor("op_1731_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_47_cast_fp16 = layer_norm(axes = input_47_axes_0, beta = norm3_2_bias_to_fp16, epsilon = var_1731_to_fp16, gamma = norm3_2_weight_to_fp16, x = input_45_cast_fp16)[name = tensor("input_47_cast_fp16")]; + tensor linear3_1_weight_to_fp16 = const()[name = tensor("linear3_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(85053248)))]; + tensor linear_15_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = linear3_1_weight_to_fp16, x = input_47_cast_fp16)[name = tensor("linear_15_cast_fp16")]; + tensor input_49_mode_0 = const()[name = tensor("input_49_mode_0"), val = tensor("EXACT")]; + tensor input_49_cast_fp16 = gelu(mode = input_49_mode_0, x = linear_15_cast_fp16)[name = tensor("input_49_cast_fp16")]; + tensor linear3_2_weight_to_fp16 = const()[name = tensor("linear3_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(93441920)))]; + tensor linear_16_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = linear3_2_weight_to_fp16, x = input_49_cast_fp16)[name = tensor("linear_16_cast_fp16")]; + tensor input_51_cast_fp16 = add(x = input_45_cast_fp16, y = linear_16_cast_fp16)[name = tensor("input_51_cast_fp16")]; + tensor input_53_axes_0 = const()[name = tensor("input_53_axes_0"), val = tensor([-1])]; + tensor norm4_1_weight_to_fp16 = const()[name = tensor("norm4_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(101830592)))]; + tensor norm4_1_bias_to_fp16 = const()[name = tensor("norm4_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(101832704)))]; + tensor var_1748_to_fp16 = const()[name = tensor("op_1748_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_53_cast_fp16 = layer_norm(axes = input_53_axes_0, beta = norm4_1_bias_to_fp16, epsilon = var_1748_to_fp16, gamma = norm4_1_weight_to_fp16, x = input_51_cast_fp16)[name = tensor("input_53_cast_fp16")]; + tensor attn4_in_proj_weight_to_fp16 = const()[name = tensor("attn4_in_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(101834816)))]; + tensor linear_17_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = attn4_in_proj_weight_to_fp16, x = input_53_cast_fp16)[name = tensor("linear_17_cast_fp16")]; + tensor var_1762 = const()[name = tensor("op_1762"), val = tensor([1, 1, 3, 16, 64])]; + tensor qkv_9_cast_fp16 = reshape(shape = var_1762, x = linear_17_cast_fp16)[name = tensor("qkv_9_cast_fp16")]; + tensor q_25_begin_0 = const()[name = tensor("q_25_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor q_25_end_0 = const()[name = tensor("q_25_end_0"), val = tensor([1, 1, 1, 16, 64])]; + tensor q_25_end_mask_0 = const()[name = tensor("q_25_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor q_25_squeeze_mask_0 = const()[name = tensor("q_25_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor q_25_cast_fp16 = slice_by_index(begin = q_25_begin_0, end = q_25_end_0, end_mask = q_25_end_mask_0, squeeze_mask = q_25_squeeze_mask_0, x = qkv_9_cast_fp16)[name = tensor("q_25_cast_fp16")]; + tensor k_17_begin_0 = const()[name = tensor("k_17_begin_0"), val = tensor([0, 0, 1, 0, 0])]; + tensor k_17_end_0 = const()[name = tensor("k_17_end_0"), val = tensor([1, 1, 2, 16, 64])]; + tensor k_17_end_mask_0 = const()[name = tensor("k_17_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor k_17_squeeze_mask_0 = const()[name = tensor("k_17_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor k_17_cast_fp16 = slice_by_index(begin = k_17_begin_0, end = k_17_end_0, end_mask = k_17_end_mask_0, squeeze_mask = k_17_squeeze_mask_0, x = qkv_9_cast_fp16)[name = tensor("k_17_cast_fp16")]; + tensor v_9_begin_0 = const()[name = tensor("v_9_begin_0"), val = tensor([0, 0, 2, 0, 0])]; + tensor v_9_end_0 = const()[name = tensor("v_9_end_0"), val = tensor([1, 1, 3, 16, 64])]; + tensor v_9_end_mask_0 = const()[name = tensor("v_9_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor v_9_squeeze_mask_0 = const()[name = tensor("v_9_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor v_9_cast_fp16 = slice_by_index(begin = v_9_begin_0, end = v_9_end_0, end_mask = v_9_end_mask_0, squeeze_mask = v_9_squeeze_mask_0, x = qkv_9_cast_fp16)[name = tensor("v_9_cast_fp16")]; + tensor var_1828 = const()[name = tensor("op_1828"), val = tensor([1, 1, 1, 1])]; + tensor position4_to_fp16_dtype_0 = const()[name = tensor("position4_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor position4_to_fp16 = cast(dtype = position4_to_fp16_dtype_0, x = position4)[name = tensor("cast_533")]; + tensor ts_29_cast_fp16 = reshape(shape = var_1828, x = position4_to_fp16)[name = tensor("ts_29_cast_fp16")]; + tensor var_1835 = const()[name = tensor("op_1835"), val = tensor([1, 1, 16, 32, 2])]; + tensor q_complex_9_cast_fp16 = reshape(shape = var_1835, x = q_25_cast_fp16)[name = tensor("q_complex_9_cast_fp16")]; + tensor var_1842 = const()[name = tensor("op_1842"), val = tensor([1, 1, 16, 32, 2])]; + tensor k_complex_9_cast_fp16 = reshape(shape = var_1842, x = k_17_cast_fp16)[name = tensor("k_complex_9_cast_fp16")]; + tensor var_1846_begin_0 = const()[name = tensor("op_1846_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_1846_end_0 = const()[name = tensor("op_1846_end_0"), val = tensor([1, 1, 16, 32, 1])]; + tensor var_1846_end_mask_0 = const()[name = tensor("op_1846_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_1846_squeeze_mask_0 = const()[name = tensor("op_1846_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_1846_cast_fp16 = slice_by_index(begin = var_1846_begin_0, end = var_1846_end_0, end_mask = var_1846_end_mask_0, squeeze_mask = var_1846_squeeze_mask_0, x = q_complex_9_cast_fp16)[name = tensor("op_1846_cast_fp16")]; + tensor var_1854_begin_0 = const()[name = tensor("op_1854_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_1854_end_0 = const()[name = tensor("op_1854_end_0"), val = tensor([1, 1, 16, 32, 2])]; + tensor var_1854_end_mask_0 = const()[name = tensor("op_1854_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_1854_squeeze_mask_0 = const()[name = tensor("op_1854_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_1854_cast_fp16 = slice_by_index(begin = var_1854_begin_0, end = var_1854_end_0, end_mask = var_1854_end_mask_0, squeeze_mask = var_1854_squeeze_mask_0, x = q_complex_9_cast_fp16)[name = tensor("op_1854_cast_fp16")]; + tensor var_1862_begin_0 = const()[name = tensor("op_1862_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_1862_end_0 = const()[name = tensor("op_1862_end_0"), val = tensor([1, 1, 16, 32, 1])]; + tensor var_1862_end_mask_0 = const()[name = tensor("op_1862_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_1862_squeeze_mask_0 = const()[name = tensor("op_1862_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_1862_cast_fp16 = slice_by_index(begin = var_1862_begin_0, end = var_1862_end_0, end_mask = var_1862_end_mask_0, squeeze_mask = var_1862_squeeze_mask_0, x = k_complex_9_cast_fp16)[name = tensor("op_1862_cast_fp16")]; + tensor var_1870_begin_0 = const()[name = tensor("op_1870_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_1870_end_0 = const()[name = tensor("op_1870_end_0"), val = tensor([1, 1, 16, 32, 2])]; + tensor var_1870_end_mask_0 = const()[name = tensor("op_1870_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_1870_squeeze_mask_0 = const()[name = tensor("op_1870_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_1870_cast_fp16 = slice_by_index(begin = var_1870_begin_0, end = var_1870_end_0, end_mask = var_1870_end_mask_0, squeeze_mask = var_1870_squeeze_mask_0, x = k_complex_9_cast_fp16)[name = tensor("op_1870_cast_fp16")]; + tensor freqs_9_to_fp16 = const()[name = tensor("freqs_9_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(108126336)))]; + tensor var_1876_cast_fp16 = mul(x = freqs_9_to_fp16, y = ts_29_cast_fp16)[name = tensor("op_1876_cast_fp16")]; + tensor rotr_9_cast_fp16 = cos(x = var_1876_cast_fp16)[name = tensor("rotr_9_cast_fp16")]; + tensor roti_9_cast_fp16 = sin(x = var_1876_cast_fp16)[name = tensor("roti_9_cast_fp16")]; + tensor var_1880_cast_fp16 = mul(x = var_1846_cast_fp16, y = rotr_9_cast_fp16)[name = tensor("op_1880_cast_fp16")]; + tensor var_1881_cast_fp16 = mul(x = var_1854_cast_fp16, y = roti_9_cast_fp16)[name = tensor("op_1881_cast_fp16")]; + tensor qor_17_cast_fp16 = sub(x = var_1880_cast_fp16, y = var_1881_cast_fp16)[name = tensor("qor_17_cast_fp16")]; + tensor var_1884_cast_fp16 = mul(x = var_1846_cast_fp16, y = roti_9_cast_fp16)[name = tensor("op_1884_cast_fp16")]; + tensor var_1885_cast_fp16 = mul(x = var_1854_cast_fp16, y = rotr_9_cast_fp16)[name = tensor("op_1885_cast_fp16")]; + tensor qoi_17_cast_fp16 = add(x = var_1884_cast_fp16, y = var_1885_cast_fp16)[name = tensor("qoi_17_cast_fp16")]; + tensor var_1888_cast_fp16 = mul(x = var_1862_cast_fp16, y = rotr_9_cast_fp16)[name = tensor("op_1888_cast_fp16")]; + tensor var_1889_cast_fp16 = mul(x = var_1870_cast_fp16, y = roti_9_cast_fp16)[name = tensor("op_1889_cast_fp16")]; + tensor kor_17_cast_fp16 = sub(x = var_1888_cast_fp16, y = var_1889_cast_fp16)[name = tensor("kor_17_cast_fp16")]; + tensor var_1892_cast_fp16 = mul(x = var_1862_cast_fp16, y = roti_9_cast_fp16)[name = tensor("op_1892_cast_fp16")]; + tensor var_1893_cast_fp16 = mul(x = var_1870_cast_fp16, y = rotr_9_cast_fp16)[name = tensor("op_1893_cast_fp16")]; + tensor koi_17_cast_fp16 = add(x = var_1892_cast_fp16, y = var_1893_cast_fp16)[name = tensor("koi_17_cast_fp16")]; + tensor qo_9_axis_0 = const()[name = tensor("qo_9_axis_0"), val = tensor(-1)]; + tensor qo_9_cast_fp16 = stack(axis = qo_9_axis_0, values = (qor_17_cast_fp16, qoi_17_cast_fp16))[name = tensor("qo_9_cast_fp16")]; + tensor ko_9_axis_0 = const()[name = tensor("ko_9_axis_0"), val = tensor(-1)]; + tensor ko_9_cast_fp16 = stack(axis = ko_9_axis_0, values = (kor_17_cast_fp16, koi_17_cast_fp16))[name = tensor("ko_9_cast_fp16")]; + tensor var_1926 = const()[name = tensor("op_1926"), val = tensor([1, 1, 16, 64])]; + tensor q_27_cast_fp16 = reshape(shape = var_1926, x = qo_9_cast_fp16)[name = tensor("q_27_cast_fp16")]; + tensor var_1932 = const()[name = tensor("op_1932"), val = tensor([1, 1, 16, 64])]; + tensor k_19_cast_fp16 = reshape(shape = var_1932, x = ko_9_cast_fp16)[name = tensor("k_19_cast_fp16")]; + tensor _inversed_1957_y_0_to_fp16 = const()[name = tensor("_inversed_1957_y_0_to_fp16"), val = tensor(0x1p-9)]; + tensor _inversed_1957_cast_fp16 = mul(x = ts_29_cast_fp16, y = _inversed_1957_y_0_to_fp16)[name = tensor("_inversed_1957_cast_fp16")]; + tensor var_1958_cast_fp16 = floor(x = _inversed_1957_cast_fp16)[name = tensor("op_1958_cast_fp16")]; + tensor var_1959_to_fp16 = const()[name = tensor("op_1959_to_fp16"), val = tensor(0x1p+9)]; + tensor var_1960_cast_fp16 = mul(x = var_1958_cast_fp16, y = var_1959_to_fp16)[name = tensor("op_1960_cast_fp16")]; + tensor write_indices_float_19_cast_fp16 = sub(x = ts_29_cast_fp16, y = var_1960_cast_fp16)[name = tensor("write_indices_float_19_cast_fp16")]; + tensor var_1967_dtype_0 = const()[name = tensor("op_1967_dtype_0"), val = tensor("int32")]; + tensor write_indices_9_reps_0 = const()[name = tensor("write_indices_9_reps_0"), val = tensor([1, 1, 16, 64])]; + tensor write_indices_float_19_cast_fp16_to_int32 = cast(dtype = var_1967_dtype_0, x = write_indices_float_19_cast_fp16)[name = tensor("cast_532")]; + tensor write_indices_9 = tile(reps = write_indices_9_reps_0, x = write_indices_float_19_cast_fp16_to_int32)[name = tensor("write_indices_9")]; + tensor var_1977_begin_0 = const()[name = tensor("op_1977_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_1977_end_0 = const()[name = tensor("op_1977_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor var_1977_end_mask_0 = const()[name = tensor("op_1977_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_1977_squeeze_mask_0 = const()[name = tensor("op_1977_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor cache4_to_fp16_dtype_0 = const()[name = tensor("cache4_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor cache4_to_fp16 = cast(dtype = cache4_to_fp16_dtype_0, x = cache4)[name = tensor("cast_531")]; + tensor var_1977_cast_fp16 = slice_by_index(begin = var_1977_begin_0, end = var_1977_end_0, end_mask = var_1977_end_mask_0, squeeze_mask = var_1977_squeeze_mask_0, x = cache4_to_fp16)[name = tensor("op_1977_cast_fp16")]; + tensor var_1979_axis_0 = const()[name = tensor("op_1979_axis_0"), val = tensor(1)]; + tensor var_1979_mode_0 = const()[name = tensor("op_1979_mode_0"), val = tensor("update")]; + tensor var_1979_validate_indices_0 = const()[name = tensor("op_1979_validate_indices_0"), val = tensor(false)]; + tensor var_1979_cast_fp16 = scatter_along_axis(axis = var_1979_axis_0, data = var_1977_cast_fp16, indices = write_indices_9, mode = var_1979_mode_0, updates = k_19_cast_fp16, validate_indices = var_1979_validate_indices_0)[name = tensor("op_1979_cast_fp16")]; + tensor concat_30 = const()[name = tensor("concat_30"), val = tensor([0, 0, 0, 0, 0])]; + tensor concat_31 = const()[name = tensor("concat_31"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_9_internal_tensor_assign_1_stride_0 = const()[name = tensor("new_cache_9_internal_tensor_assign_1_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_9_internal_tensor_assign_1_begin_mask_0 = const()[name = tensor("new_cache_9_internal_tensor_assign_1_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_9_internal_tensor_assign_1_end_mask_0 = const()[name = tensor("new_cache_9_internal_tensor_assign_1_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_9_internal_tensor_assign_1_squeeze_mask_0 = const()[name = tensor("new_cache_9_internal_tensor_assign_1_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_56 = const()[name = tensor("shape_56"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_8 = const()[name = tensor("reduce_prod_8"), val = tensor(1048576)]; + tensor range_1d_8_start_0 = const()[name = tensor("range_1d_8_start_0"), val = tensor(0)]; + tensor range_1d_8_step_0 = const()[name = tensor("range_1d_8_step_0"), val = tensor(1)]; + tensor range_1d_8 = range_1d(end = reduce_prod_8, start = range_1d_8_start_0, step = range_1d_8_step_0)[name = tensor("range_1d_8")]; + tensor reshape_40 = reshape(shape = shape_56, x = range_1d_8)[name = tensor("reshape_40")]; + tensor slice_by_index_8 = slice_by_index(begin = concat_30, begin_mask = new_cache_9_internal_tensor_assign_1_begin_mask_0, end = concat_31, end_mask = new_cache_9_internal_tensor_assign_1_end_mask_0, squeeze_mask = new_cache_9_internal_tensor_assign_1_squeeze_mask_0, stride = new_cache_9_internal_tensor_assign_1_stride_0, x = reshape_40)[name = tensor("slice_by_index_8")]; + tensor reshape_41_shape_0 = const()[name = tensor("reshape_41_shape_0"), val = tensor([-1])]; + tensor reshape_41 = reshape(shape = reshape_41_shape_0, x = slice_by_index_8)[name = tensor("reshape_41")]; + tensor reshape_42_shape_0 = const()[name = tensor("reshape_42_shape_0"), val = tensor([-1])]; + tensor reshape_42_cast_fp16 = reshape(shape = reshape_42_shape_0, x = var_1979_cast_fp16)[name = tensor("reshape_42_cast_fp16")]; + tensor reshape_43_shape_0 = const()[name = tensor("reshape_43_shape_0"), val = tensor([-1])]; + tensor reshape_43_cast_fp16 = reshape(shape = reshape_43_shape_0, x = cache4_to_fp16)[name = tensor("reshape_43_cast_fp16")]; + tensor scatter_8_mode_0 = const()[name = tensor("scatter_8_mode_0"), val = tensor("update")]; + tensor scatter_8_axis_0 = const()[name = tensor("scatter_8_axis_0"), val = tensor(0)]; + tensor scatter_8_validate_indices_0 = const()[name = tensor("scatter_8_validate_indices_0"), val = tensor(false)]; + tensor scatter_8_cast_fp16 = scatter(axis = scatter_8_axis_0, data = reshape_43_cast_fp16, indices = reshape_41, mode = scatter_8_mode_0, updates = reshape_42_cast_fp16, validate_indices = scatter_8_validate_indices_0)[name = tensor("scatter_8_cast_fp16")]; + tensor reshape_44_cast_fp16 = reshape(shape = shape_56, x = scatter_8_cast_fp16)[name = tensor("reshape_44_cast_fp16")]; + tensor var_1987_begin_0 = const()[name = tensor("op_1987_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor var_1987_end_0 = const()[name = tensor("op_1987_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor var_1987_end_mask_0 = const()[name = tensor("op_1987_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_1987_squeeze_mask_0 = const()[name = tensor("op_1987_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor var_1987_cast_fp16 = slice_by_index(begin = var_1987_begin_0, end = var_1987_end_0, end_mask = var_1987_end_mask_0, squeeze_mask = var_1987_squeeze_mask_0, x = reshape_44_cast_fp16)[name = tensor("op_1987_cast_fp16")]; + tensor var_1989_axis_0 = const()[name = tensor("op_1989_axis_0"), val = tensor(1)]; + tensor var_1989_mode_0 = const()[name = tensor("op_1989_mode_0"), val = tensor("update")]; + tensor var_1989_validate_indices_0 = const()[name = tensor("op_1989_validate_indices_0"), val = tensor(false)]; + tensor var_1989_cast_fp16 = scatter_along_axis(axis = var_1989_axis_0, data = var_1987_cast_fp16, indices = write_indices_9, mode = var_1989_mode_0, updates = v_9_cast_fp16, validate_indices = var_1989_validate_indices_0)[name = tensor("op_1989_cast_fp16")]; + tensor concat_32 = const()[name = tensor("concat_32"), val = tensor([1, 0, 0, 0, 0])]; + tensor concat_33 = const()[name = tensor("concat_33"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_9_internal_tensor_assign_2_stride_0 = const()[name = tensor("new_cache_9_internal_tensor_assign_2_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_9_internal_tensor_assign_2_begin_mask_0 = const()[name = tensor("new_cache_9_internal_tensor_assign_2_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_9_internal_tensor_assign_2_end_mask_0 = const()[name = tensor("new_cache_9_internal_tensor_assign_2_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_9_internal_tensor_assign_2_squeeze_mask_0 = const()[name = tensor("new_cache_9_internal_tensor_assign_2_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_57 = const()[name = tensor("shape_57"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_9 = const()[name = tensor("reduce_prod_9"), val = tensor(1048576)]; + tensor range_1d_9_start_0 = const()[name = tensor("range_1d_9_start_0"), val = tensor(0)]; + tensor range_1d_9_step_0 = const()[name = tensor("range_1d_9_step_0"), val = tensor(1)]; + tensor range_1d_9 = range_1d(end = reduce_prod_9, start = range_1d_9_start_0, step = range_1d_9_step_0)[name = tensor("range_1d_9")]; + tensor reshape_45 = reshape(shape = shape_57, x = range_1d_9)[name = tensor("reshape_45")]; + tensor slice_by_index_9 = slice_by_index(begin = concat_32, begin_mask = new_cache_9_internal_tensor_assign_2_begin_mask_0, end = concat_33, end_mask = new_cache_9_internal_tensor_assign_2_end_mask_0, squeeze_mask = new_cache_9_internal_tensor_assign_2_squeeze_mask_0, stride = new_cache_9_internal_tensor_assign_2_stride_0, x = reshape_45)[name = tensor("slice_by_index_9")]; + tensor reshape_46_shape_0 = const()[name = tensor("reshape_46_shape_0"), val = tensor([-1])]; + tensor reshape_46 = reshape(shape = reshape_46_shape_0, x = slice_by_index_9)[name = tensor("reshape_46")]; + tensor reshape_47_shape_0 = const()[name = tensor("reshape_47_shape_0"), val = tensor([-1])]; + tensor reshape_47_cast_fp16 = reshape(shape = reshape_47_shape_0, x = var_1989_cast_fp16)[name = tensor("reshape_47_cast_fp16")]; + tensor reshape_48_shape_0 = const()[name = tensor("reshape_48_shape_0"), val = tensor([-1])]; + tensor reshape_48_cast_fp16 = reshape(shape = reshape_48_shape_0, x = reshape_44_cast_fp16)[name = tensor("reshape_48_cast_fp16")]; + tensor scatter_9_mode_0 = const()[name = tensor("scatter_9_mode_0"), val = tensor("update")]; + tensor scatter_9_axis_0 = const()[name = tensor("scatter_9_axis_0"), val = tensor(0)]; + tensor scatter_9_validate_indices_0 = const()[name = tensor("scatter_9_validate_indices_0"), val = tensor(false)]; + tensor scatter_9_cast_fp16 = scatter(axis = scatter_9_axis_0, data = reshape_48_cast_fp16, indices = reshape_46, mode = scatter_9_mode_0, updates = reshape_47_cast_fp16, validate_indices = scatter_9_validate_indices_0)[name = tensor("scatter_9_cast_fp16")]; + tensor reshape_49_cast_fp16 = reshape(shape = shape_57, x = scatter_9_cast_fp16)[name = tensor("reshape_49_cast_fp16")]; + tensor reshape_49_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("reshape_49_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor keys_25_begin_0 = const()[name = tensor("keys_25_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor keys_25_end_0 = const()[name = tensor("keys_25_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor keys_25_end_mask_0 = const()[name = tensor("keys_25_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor keys_25_squeeze_mask_0 = const()[name = tensor("keys_25_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor keys_25_cast_fp16 = slice_by_index(begin = keys_25_begin_0, end = keys_25_end_0, end_mask = keys_25_end_mask_0, squeeze_mask = keys_25_squeeze_mask_0, x = reshape_49_cast_fp16)[name = tensor("keys_25_cast_fp16")]; + tensor values_25_begin_0 = const()[name = tensor("values_25_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor values_25_end_0 = const()[name = tensor("values_25_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor values_25_end_mask_0 = const()[name = tensor("values_25_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor values_25_squeeze_mask_0 = const()[name = tensor("values_25_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor values_25_cast_fp16 = slice_by_index(begin = values_25_begin_0, end = values_25_end_0, end_mask = values_25_end_mask_0, squeeze_mask = values_25_squeeze_mask_0, x = reshape_49_cast_fp16)[name = tensor("values_25_cast_fp16")]; + tensor var_2001_cast_fp16 = not_equal(x = keys_25_cast_fp16, y = keys_25_cast_fp16)[name = tensor("op_2001_cast_fp16")]; + tensor keys_27_cast_fp16 = select(a = var_463_to_fp16, b = keys_25_cast_fp16, cond = var_2001_cast_fp16)[name = tensor("keys_27_cast_fp16")]; + tensor var_2009_cast_fp16 = not_equal(x = values_25_cast_fp16, y = values_25_cast_fp16)[name = tensor("op_2009_cast_fp16")]; + tensor values_27_cast_fp16 = select(a = var_463_to_fp16, b = values_25_cast_fp16, cond = var_2009_cast_fp16)[name = tensor("values_27_cast_fp16")]; + tensor var_2033 = const()[name = tensor("op_2033"), val = tensor([0, 2, 1, 3])]; + tensor var_2049 = const()[name = tensor("op_2049"), val = tensor([1, 1, 1])]; + tensor var_2050_cast_fp16 = reshape(shape = var_2049, x = position4_to_fp16)[name = tensor("op_2050_cast_fp16")]; + tensor var_2070_to_fp16 = const()[name = tensor("op_2070_to_fp16"), val = tensor(0x1p+0)]; + tensor valid_len_9_cast_fp16 = add(x = var_2050_cast_fp16, y = var_2070_to_fp16)[name = tensor("valid_len_9_cast_fp16")]; + tensor valid_mask_9_cast_fp16 = less(x = k_positions_1_promoted_to_fp16, y = valid_len_9_cast_fp16)[name = tensor("valid_mask_9_cast_fp16")]; + tensor causal_mask_9_cast_fp16 = less_equal(x = k_positions_1_promoted_to_fp16, y = var_2050_cast_fp16)[name = tensor("causal_mask_9_cast_fp16")]; + tensor attn_mask_17 = logical_and(x = valid_mask_9_cast_fp16, y = causal_mask_9_cast_fp16)[name = tensor("attn_mask_17")]; + tensor attn_mask_19_axes_0 = const()[name = tensor("attn_mask_19_axes_0"), val = tensor([1])]; + tensor attn_mask_19 = expand_dims(axes = attn_mask_19_axes_0, x = attn_mask_17)[name = tensor("attn_mask_19")]; + tensor var_2088_transpose_x_0 = const()[name = tensor("op_2088_transpose_x_0"), val = tensor(false)]; + tensor var_2088_transpose_y_0 = const()[name = tensor("op_2088_transpose_y_0"), val = tensor(false)]; + tensor transpose_80_perm_0 = const()[name = tensor("transpose_80_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_81_perm_0 = const()[name = tensor("transpose_81_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_81 = transpose(perm = transpose_81_perm_0, x = keys_27_cast_fp16)[name = tensor("transpose_197")]; + tensor transpose_80 = transpose(perm = transpose_80_perm_0, x = q_27_cast_fp16)[name = tensor("transpose_198")]; + tensor var_2088_cast_fp16 = matmul(transpose_x = var_2088_transpose_x_0, transpose_y = var_2088_transpose_y_0, x = transpose_80, y = transpose_81)[name = tensor("op_2088_cast_fp16")]; + tensor var_2082_to_fp16 = const()[name = tensor("op_2082_to_fp16"), val = tensor([0x1p-3])]; + tensor attn_weights_25_cast_fp16 = mul(x = var_2088_cast_fp16, y = var_2082_to_fp16)[name = tensor("attn_weights_25_cast_fp16")]; + tensor var_2090 = logical_not(x = attn_mask_19)[name = tensor("op_2090")]; + tensor var_2091_to_fp16 = const()[name = tensor("op_2091_to_fp16"), val = tensor(-inf)]; + tensor attn_weights_27_cast_fp16 = select(a = var_2091_to_fp16, b = attn_weights_25_cast_fp16, cond = var_2090)[name = tensor("attn_weights_27_cast_fp16")]; + tensor var_2093 = const()[name = tensor("op_2093"), val = tensor(-1)]; + tensor attn_weights_29_cast_fp16 = softmax(axis = var_2093, x = attn_weights_27_cast_fp16)[name = tensor("attn_weights_29_cast_fp16")]; + tensor attn_output_9_transpose_x_0 = const()[name = tensor("attn_output_9_transpose_x_0"), val = tensor(false)]; + tensor attn_output_9_transpose_y_0 = const()[name = tensor("attn_output_9_transpose_y_0"), val = tensor(false)]; + tensor values_29_cast_fp16 = transpose(perm = var_2033, x = values_27_cast_fp16)[name = tensor("transpose_199")]; + tensor attn_output_9_cast_fp16 = matmul(transpose_x = attn_output_9_transpose_x_0, transpose_y = attn_output_9_transpose_y_0, x = attn_weights_29_cast_fp16, y = values_29_cast_fp16)[name = tensor("attn_output_9_cast_fp16")]; + tensor var_2101 = const()[name = tensor("op_2101"), val = tensor([0, 2, 1, 3])]; + tensor var_2106 = const()[name = tensor("op_2106"), val = tensor([1, 1, 1024])]; + tensor var_2102_cast_fp16 = transpose(perm = var_2101, x = attn_output_9_cast_fp16)[name = tensor("transpose_196")]; + tensor input_55_cast_fp16 = reshape(shape = var_2106, x = var_2102_cast_fp16)[name = tensor("input_55_cast_fp16")]; + tensor attn4_out_proj_weight_to_fp16 = const()[name = tensor("attn4_out_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(108126464)))]; + tensor linear_18_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = attn4_out_proj_weight_to_fp16, x = input_55_cast_fp16)[name = tensor("linear_18_cast_fp16")]; + tensor var_2112_to_fp16 = const()[name = tensor("op_2112_to_fp16"), val = tensor(0x1p+0)]; + tensor var_2113_cast_fp16 = add(x = position4_to_fp16, y = var_2112_to_fp16)[name = tensor("op_2113_cast_fp16")]; + tensor var_2113_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_2113_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor input_57_cast_fp16 = add(x = input_51_cast_fp16, y = linear_18_cast_fp16)[name = tensor("input_57_cast_fp16")]; + tensor input_59_axes_0 = const()[name = tensor("input_59_axes_0"), val = tensor([-1])]; + tensor norm4_2_weight_to_fp16 = const()[name = tensor("norm4_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(110223680)))]; + tensor norm4_2_bias_to_fp16 = const()[name = tensor("norm4_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(110225792)))]; + tensor var_2117_to_fp16 = const()[name = tensor("op_2117_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_59_cast_fp16 = layer_norm(axes = input_59_axes_0, beta = norm4_2_bias_to_fp16, epsilon = var_2117_to_fp16, gamma = norm4_2_weight_to_fp16, x = input_57_cast_fp16)[name = tensor("input_59_cast_fp16")]; + tensor linear4_1_weight_to_fp16 = const()[name = tensor("linear4_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(110227904)))]; + tensor linear_19_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = linear4_1_weight_to_fp16, x = input_59_cast_fp16)[name = tensor("linear_19_cast_fp16")]; + tensor input_61_mode_0 = const()[name = tensor("input_61_mode_0"), val = tensor("EXACT")]; + tensor input_61_cast_fp16 = gelu(mode = input_61_mode_0, x = linear_19_cast_fp16)[name = tensor("input_61_cast_fp16")]; + tensor linear4_2_weight_to_fp16 = const()[name = tensor("linear4_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(118616576)))]; + tensor linear_20_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = linear4_2_weight_to_fp16, x = input_61_cast_fp16)[name = tensor("linear_20_cast_fp16")]; + tensor input_63_cast_fp16 = add(x = input_57_cast_fp16, y = linear_20_cast_fp16)[name = tensor("input_63_cast_fp16")]; + tensor input_65_axes_0 = const()[name = tensor("input_65_axes_0"), val = tensor([-1])]; + tensor norm5_1_weight_to_fp16 = const()[name = tensor("norm5_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(127005248)))]; + tensor norm5_1_bias_to_fp16 = const()[name = tensor("norm5_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(127007360)))]; + tensor var_2134_to_fp16 = const()[name = tensor("op_2134_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_65_cast_fp16 = layer_norm(axes = input_65_axes_0, beta = norm5_1_bias_to_fp16, epsilon = var_2134_to_fp16, gamma = norm5_1_weight_to_fp16, x = input_63_cast_fp16)[name = tensor("input_65_cast_fp16")]; + tensor attn5_in_proj_weight_to_fp16 = const()[name = tensor("attn5_in_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(127009472)))]; + tensor linear_21_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = attn5_in_proj_weight_to_fp16, x = input_65_cast_fp16)[name = tensor("linear_21_cast_fp16")]; + tensor var_2148 = const()[name = tensor("op_2148"), val = tensor([1, 1, 3, 16, 64])]; + tensor qkv_11_cast_fp16 = reshape(shape = var_2148, x = linear_21_cast_fp16)[name = tensor("qkv_11_cast_fp16")]; + tensor q_31_begin_0 = const()[name = tensor("q_31_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor q_31_end_0 = const()[name = tensor("q_31_end_0"), val = tensor([1, 1, 1, 16, 64])]; + tensor q_31_end_mask_0 = const()[name = tensor("q_31_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor q_31_squeeze_mask_0 = const()[name = tensor("q_31_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor q_31_cast_fp16 = slice_by_index(begin = q_31_begin_0, end = q_31_end_0, end_mask = q_31_end_mask_0, squeeze_mask = q_31_squeeze_mask_0, x = qkv_11_cast_fp16)[name = tensor("q_31_cast_fp16")]; + tensor k_21_begin_0 = const()[name = tensor("k_21_begin_0"), val = tensor([0, 0, 1, 0, 0])]; + tensor k_21_end_0 = const()[name = tensor("k_21_end_0"), val = tensor([1, 1, 2, 16, 64])]; + tensor k_21_end_mask_0 = const()[name = tensor("k_21_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor k_21_squeeze_mask_0 = const()[name = tensor("k_21_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor k_21_cast_fp16 = slice_by_index(begin = k_21_begin_0, end = k_21_end_0, end_mask = k_21_end_mask_0, squeeze_mask = k_21_squeeze_mask_0, x = qkv_11_cast_fp16)[name = tensor("k_21_cast_fp16")]; + tensor v_11_begin_0 = const()[name = tensor("v_11_begin_0"), val = tensor([0, 0, 2, 0, 0])]; + tensor v_11_end_0 = const()[name = tensor("v_11_end_0"), val = tensor([1, 1, 3, 16, 64])]; + tensor v_11_end_mask_0 = const()[name = tensor("v_11_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor v_11_squeeze_mask_0 = const()[name = tensor("v_11_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor v_11_cast_fp16 = slice_by_index(begin = v_11_begin_0, end = v_11_end_0, end_mask = v_11_end_mask_0, squeeze_mask = v_11_squeeze_mask_0, x = qkv_11_cast_fp16)[name = tensor("v_11_cast_fp16")]; + tensor var_2214 = const()[name = tensor("op_2214"), val = tensor([1, 1, 1, 1])]; + tensor position5_to_fp16_dtype_0 = const()[name = tensor("position5_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor position5_to_fp16 = cast(dtype = position5_to_fp16_dtype_0, x = position5)[name = tensor("cast_528")]; + tensor ts_35_cast_fp16 = reshape(shape = var_2214, x = position5_to_fp16)[name = tensor("ts_35_cast_fp16")]; + tensor var_2221 = const()[name = tensor("op_2221"), val = tensor([1, 1, 16, 32, 2])]; + tensor q_complex_11_cast_fp16 = reshape(shape = var_2221, x = q_31_cast_fp16)[name = tensor("q_complex_11_cast_fp16")]; + tensor var_2228 = const()[name = tensor("op_2228"), val = tensor([1, 1, 16, 32, 2])]; + tensor k_complex_11_cast_fp16 = reshape(shape = var_2228, x = k_21_cast_fp16)[name = tensor("k_complex_11_cast_fp16")]; + tensor var_2232_begin_0 = const()[name = tensor("op_2232_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_2232_end_0 = const()[name = tensor("op_2232_end_0"), val = tensor([1, 1, 16, 32, 1])]; + tensor var_2232_end_mask_0 = const()[name = tensor("op_2232_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_2232_squeeze_mask_0 = const()[name = tensor("op_2232_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_2232_cast_fp16 = slice_by_index(begin = var_2232_begin_0, end = var_2232_end_0, end_mask = var_2232_end_mask_0, squeeze_mask = var_2232_squeeze_mask_0, x = q_complex_11_cast_fp16)[name = tensor("op_2232_cast_fp16")]; + tensor var_2240_begin_0 = const()[name = tensor("op_2240_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_2240_end_0 = const()[name = tensor("op_2240_end_0"), val = tensor([1, 1, 16, 32, 2])]; + tensor var_2240_end_mask_0 = const()[name = tensor("op_2240_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_2240_squeeze_mask_0 = const()[name = tensor("op_2240_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_2240_cast_fp16 = slice_by_index(begin = var_2240_begin_0, end = var_2240_end_0, end_mask = var_2240_end_mask_0, squeeze_mask = var_2240_squeeze_mask_0, x = q_complex_11_cast_fp16)[name = tensor("op_2240_cast_fp16")]; + tensor var_2248_begin_0 = const()[name = tensor("op_2248_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_2248_end_0 = const()[name = tensor("op_2248_end_0"), val = tensor([1, 1, 16, 32, 1])]; + tensor var_2248_end_mask_0 = const()[name = tensor("op_2248_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_2248_squeeze_mask_0 = const()[name = tensor("op_2248_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_2248_cast_fp16 = slice_by_index(begin = var_2248_begin_0, end = var_2248_end_0, end_mask = var_2248_end_mask_0, squeeze_mask = var_2248_squeeze_mask_0, x = k_complex_11_cast_fp16)[name = tensor("op_2248_cast_fp16")]; + tensor var_2256_begin_0 = const()[name = tensor("op_2256_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_2256_end_0 = const()[name = tensor("op_2256_end_0"), val = tensor([1, 1, 16, 32, 2])]; + tensor var_2256_end_mask_0 = const()[name = tensor("op_2256_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_2256_squeeze_mask_0 = const()[name = tensor("op_2256_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_2256_cast_fp16 = slice_by_index(begin = var_2256_begin_0, end = var_2256_end_0, end_mask = var_2256_end_mask_0, squeeze_mask = var_2256_squeeze_mask_0, x = k_complex_11_cast_fp16)[name = tensor("op_2256_cast_fp16")]; + tensor freqs_11_to_fp16 = const()[name = tensor("freqs_11_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(133300992)))]; + tensor var_2262_cast_fp16 = mul(x = freqs_11_to_fp16, y = ts_35_cast_fp16)[name = tensor("op_2262_cast_fp16")]; + tensor rotr_11_cast_fp16 = cos(x = var_2262_cast_fp16)[name = tensor("rotr_11_cast_fp16")]; + tensor roti_11_cast_fp16 = sin(x = var_2262_cast_fp16)[name = tensor("roti_11_cast_fp16")]; + tensor var_2266_cast_fp16 = mul(x = var_2232_cast_fp16, y = rotr_11_cast_fp16)[name = tensor("op_2266_cast_fp16")]; + tensor var_2267_cast_fp16 = mul(x = var_2240_cast_fp16, y = roti_11_cast_fp16)[name = tensor("op_2267_cast_fp16")]; + tensor qor_21_cast_fp16 = sub(x = var_2266_cast_fp16, y = var_2267_cast_fp16)[name = tensor("qor_21_cast_fp16")]; + tensor var_2270_cast_fp16 = mul(x = var_2232_cast_fp16, y = roti_11_cast_fp16)[name = tensor("op_2270_cast_fp16")]; + tensor var_2271_cast_fp16 = mul(x = var_2240_cast_fp16, y = rotr_11_cast_fp16)[name = tensor("op_2271_cast_fp16")]; + tensor qoi_21_cast_fp16 = add(x = var_2270_cast_fp16, y = var_2271_cast_fp16)[name = tensor("qoi_21_cast_fp16")]; + tensor var_2274_cast_fp16 = mul(x = var_2248_cast_fp16, y = rotr_11_cast_fp16)[name = tensor("op_2274_cast_fp16")]; + tensor var_2275_cast_fp16 = mul(x = var_2256_cast_fp16, y = roti_11_cast_fp16)[name = tensor("op_2275_cast_fp16")]; + tensor kor_21_cast_fp16 = sub(x = var_2274_cast_fp16, y = var_2275_cast_fp16)[name = tensor("kor_21_cast_fp16")]; + tensor var_2278_cast_fp16 = mul(x = var_2248_cast_fp16, y = roti_11_cast_fp16)[name = tensor("op_2278_cast_fp16")]; + tensor var_2279_cast_fp16 = mul(x = var_2256_cast_fp16, y = rotr_11_cast_fp16)[name = tensor("op_2279_cast_fp16")]; + tensor koi_21_cast_fp16 = add(x = var_2278_cast_fp16, y = var_2279_cast_fp16)[name = tensor("koi_21_cast_fp16")]; + tensor qo_11_axis_0 = const()[name = tensor("qo_11_axis_0"), val = tensor(-1)]; + tensor qo_11_cast_fp16 = stack(axis = qo_11_axis_0, values = (qor_21_cast_fp16, qoi_21_cast_fp16))[name = tensor("qo_11_cast_fp16")]; + tensor ko_11_axis_0 = const()[name = tensor("ko_11_axis_0"), val = tensor(-1)]; + tensor ko_11_cast_fp16 = stack(axis = ko_11_axis_0, values = (kor_21_cast_fp16, koi_21_cast_fp16))[name = tensor("ko_11_cast_fp16")]; + tensor var_2312 = const()[name = tensor("op_2312"), val = tensor([1, 1, 16, 64])]; + tensor q_33_cast_fp16 = reshape(shape = var_2312, x = qo_11_cast_fp16)[name = tensor("q_33_cast_fp16")]; + tensor var_2318 = const()[name = tensor("op_2318"), val = tensor([1, 1, 16, 64])]; + tensor k_23_cast_fp16 = reshape(shape = var_2318, x = ko_11_cast_fp16)[name = tensor("k_23_cast_fp16")]; + tensor _inversed_2343_y_0_to_fp16 = const()[name = tensor("_inversed_2343_y_0_to_fp16"), val = tensor(0x1p-9)]; + tensor _inversed_2343_cast_fp16 = mul(x = ts_35_cast_fp16, y = _inversed_2343_y_0_to_fp16)[name = tensor("_inversed_2343_cast_fp16")]; + tensor var_2344_cast_fp16 = floor(x = _inversed_2343_cast_fp16)[name = tensor("op_2344_cast_fp16")]; + tensor var_2345_to_fp16 = const()[name = tensor("op_2345_to_fp16"), val = tensor(0x1p+9)]; + tensor var_2346_cast_fp16 = mul(x = var_2344_cast_fp16, y = var_2345_to_fp16)[name = tensor("op_2346_cast_fp16")]; + tensor write_indices_float_23_cast_fp16 = sub(x = ts_35_cast_fp16, y = var_2346_cast_fp16)[name = tensor("write_indices_float_23_cast_fp16")]; + tensor var_2353_dtype_0 = const()[name = tensor("op_2353_dtype_0"), val = tensor("int32")]; + tensor write_indices_11_reps_0 = const()[name = tensor("write_indices_11_reps_0"), val = tensor([1, 1, 16, 64])]; + tensor write_indices_float_23_cast_fp16_to_int32 = cast(dtype = var_2353_dtype_0, x = write_indices_float_23_cast_fp16)[name = tensor("cast_527")]; + tensor write_indices_11 = tile(reps = write_indices_11_reps_0, x = write_indices_float_23_cast_fp16_to_int32)[name = tensor("write_indices_11")]; + tensor var_2363_begin_0 = const()[name = tensor("op_2363_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_2363_end_0 = const()[name = tensor("op_2363_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor var_2363_end_mask_0 = const()[name = tensor("op_2363_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_2363_squeeze_mask_0 = const()[name = tensor("op_2363_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor cache5_to_fp16_dtype_0 = const()[name = tensor("cache5_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor cache5_to_fp16 = cast(dtype = cache5_to_fp16_dtype_0, x = cache5)[name = tensor("cast_526")]; + tensor var_2363_cast_fp16 = slice_by_index(begin = var_2363_begin_0, end = var_2363_end_0, end_mask = var_2363_end_mask_0, squeeze_mask = var_2363_squeeze_mask_0, x = cache5_to_fp16)[name = tensor("op_2363_cast_fp16")]; + tensor var_2365_axis_0 = const()[name = tensor("op_2365_axis_0"), val = tensor(1)]; + tensor var_2365_mode_0 = const()[name = tensor("op_2365_mode_0"), val = tensor("update")]; + tensor var_2365_validate_indices_0 = const()[name = tensor("op_2365_validate_indices_0"), val = tensor(false)]; + tensor var_2365_cast_fp16 = scatter_along_axis(axis = var_2365_axis_0, data = var_2363_cast_fp16, indices = write_indices_11, mode = var_2365_mode_0, updates = k_23_cast_fp16, validate_indices = var_2365_validate_indices_0)[name = tensor("op_2365_cast_fp16")]; + tensor concat_37 = const()[name = tensor("concat_37"), val = tensor([0, 0, 0, 0, 0])]; + tensor concat_38 = const()[name = tensor("concat_38"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_11_internal_tensor_assign_1_stride_0 = const()[name = tensor("new_cache_11_internal_tensor_assign_1_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_11_internal_tensor_assign_1_begin_mask_0 = const()[name = tensor("new_cache_11_internal_tensor_assign_1_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_11_internal_tensor_assign_1_end_mask_0 = const()[name = tensor("new_cache_11_internal_tensor_assign_1_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_11_internal_tensor_assign_1_squeeze_mask_0 = const()[name = tensor("new_cache_11_internal_tensor_assign_1_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_58 = const()[name = tensor("shape_58"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_10 = const()[name = tensor("reduce_prod_10"), val = tensor(1048576)]; + tensor range_1d_10_start_0 = const()[name = tensor("range_1d_10_start_0"), val = tensor(0)]; + tensor range_1d_10_step_0 = const()[name = tensor("range_1d_10_step_0"), val = tensor(1)]; + tensor range_1d_10 = range_1d(end = reduce_prod_10, start = range_1d_10_start_0, step = range_1d_10_step_0)[name = tensor("range_1d_10")]; + tensor reshape_50 = reshape(shape = shape_58, x = range_1d_10)[name = tensor("reshape_50")]; + tensor slice_by_index_10 = slice_by_index(begin = concat_37, begin_mask = new_cache_11_internal_tensor_assign_1_begin_mask_0, end = concat_38, end_mask = new_cache_11_internal_tensor_assign_1_end_mask_0, squeeze_mask = new_cache_11_internal_tensor_assign_1_squeeze_mask_0, stride = new_cache_11_internal_tensor_assign_1_stride_0, x = reshape_50)[name = tensor("slice_by_index_10")]; + tensor reshape_51_shape_0 = const()[name = tensor("reshape_51_shape_0"), val = tensor([-1])]; + tensor reshape_51 = reshape(shape = reshape_51_shape_0, x = slice_by_index_10)[name = tensor("reshape_51")]; + tensor reshape_52_shape_0 = const()[name = tensor("reshape_52_shape_0"), val = tensor([-1])]; + tensor reshape_52_cast_fp16 = reshape(shape = reshape_52_shape_0, x = var_2365_cast_fp16)[name = tensor("reshape_52_cast_fp16")]; + tensor reshape_53_shape_0 = const()[name = tensor("reshape_53_shape_0"), val = tensor([-1])]; + tensor reshape_53_cast_fp16 = reshape(shape = reshape_53_shape_0, x = cache5_to_fp16)[name = tensor("reshape_53_cast_fp16")]; + tensor scatter_10_mode_0 = const()[name = tensor("scatter_10_mode_0"), val = tensor("update")]; + tensor scatter_10_axis_0 = const()[name = tensor("scatter_10_axis_0"), val = tensor(0)]; + tensor scatter_10_validate_indices_0 = const()[name = tensor("scatter_10_validate_indices_0"), val = tensor(false)]; + tensor scatter_10_cast_fp16 = scatter(axis = scatter_10_axis_0, data = reshape_53_cast_fp16, indices = reshape_51, mode = scatter_10_mode_0, updates = reshape_52_cast_fp16, validate_indices = scatter_10_validate_indices_0)[name = tensor("scatter_10_cast_fp16")]; + tensor reshape_54_cast_fp16 = reshape(shape = shape_58, x = scatter_10_cast_fp16)[name = tensor("reshape_54_cast_fp16")]; + tensor var_2373_begin_0 = const()[name = tensor("op_2373_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor var_2373_end_0 = const()[name = tensor("op_2373_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor var_2373_end_mask_0 = const()[name = tensor("op_2373_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_2373_squeeze_mask_0 = const()[name = tensor("op_2373_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor var_2373_cast_fp16 = slice_by_index(begin = var_2373_begin_0, end = var_2373_end_0, end_mask = var_2373_end_mask_0, squeeze_mask = var_2373_squeeze_mask_0, x = reshape_54_cast_fp16)[name = tensor("op_2373_cast_fp16")]; + tensor var_2375_axis_0 = const()[name = tensor("op_2375_axis_0"), val = tensor(1)]; + tensor var_2375_mode_0 = const()[name = tensor("op_2375_mode_0"), val = tensor("update")]; + tensor var_2375_validate_indices_0 = const()[name = tensor("op_2375_validate_indices_0"), val = tensor(false)]; + tensor var_2375_cast_fp16 = scatter_along_axis(axis = var_2375_axis_0, data = var_2373_cast_fp16, indices = write_indices_11, mode = var_2375_mode_0, updates = v_11_cast_fp16, validate_indices = var_2375_validate_indices_0)[name = tensor("op_2375_cast_fp16")]; + tensor concat_39 = const()[name = tensor("concat_39"), val = tensor([1, 0, 0, 0, 0])]; + tensor concat_40 = const()[name = tensor("concat_40"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_11_internal_tensor_assign_2_stride_0 = const()[name = tensor("new_cache_11_internal_tensor_assign_2_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_11_internal_tensor_assign_2_begin_mask_0 = const()[name = tensor("new_cache_11_internal_tensor_assign_2_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_11_internal_tensor_assign_2_end_mask_0 = const()[name = tensor("new_cache_11_internal_tensor_assign_2_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_11_internal_tensor_assign_2_squeeze_mask_0 = const()[name = tensor("new_cache_11_internal_tensor_assign_2_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_59 = const()[name = tensor("shape_59"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_11 = const()[name = tensor("reduce_prod_11"), val = tensor(1048576)]; + tensor range_1d_11_start_0 = const()[name = tensor("range_1d_11_start_0"), val = tensor(0)]; + tensor range_1d_11_step_0 = const()[name = tensor("range_1d_11_step_0"), val = tensor(1)]; + tensor range_1d_11 = range_1d(end = reduce_prod_11, start = range_1d_11_start_0, step = range_1d_11_step_0)[name = tensor("range_1d_11")]; + tensor reshape_55 = reshape(shape = shape_59, x = range_1d_11)[name = tensor("reshape_55")]; + tensor slice_by_index_11 = slice_by_index(begin = concat_39, begin_mask = new_cache_11_internal_tensor_assign_2_begin_mask_0, end = concat_40, end_mask = new_cache_11_internal_tensor_assign_2_end_mask_0, squeeze_mask = new_cache_11_internal_tensor_assign_2_squeeze_mask_0, stride = new_cache_11_internal_tensor_assign_2_stride_0, x = reshape_55)[name = tensor("slice_by_index_11")]; + tensor reshape_56_shape_0 = const()[name = tensor("reshape_56_shape_0"), val = tensor([-1])]; + tensor reshape_56 = reshape(shape = reshape_56_shape_0, x = slice_by_index_11)[name = tensor("reshape_56")]; + tensor reshape_57_shape_0 = const()[name = tensor("reshape_57_shape_0"), val = tensor([-1])]; + tensor reshape_57_cast_fp16 = reshape(shape = reshape_57_shape_0, x = var_2375_cast_fp16)[name = tensor("reshape_57_cast_fp16")]; + tensor reshape_58_shape_0 = const()[name = tensor("reshape_58_shape_0"), val = tensor([-1])]; + tensor reshape_58_cast_fp16 = reshape(shape = reshape_58_shape_0, x = reshape_54_cast_fp16)[name = tensor("reshape_58_cast_fp16")]; + tensor scatter_11_mode_0 = const()[name = tensor("scatter_11_mode_0"), val = tensor("update")]; + tensor scatter_11_axis_0 = const()[name = tensor("scatter_11_axis_0"), val = tensor(0)]; + tensor scatter_11_validate_indices_0 = const()[name = tensor("scatter_11_validate_indices_0"), val = tensor(false)]; + tensor scatter_11_cast_fp16 = scatter(axis = scatter_11_axis_0, data = reshape_58_cast_fp16, indices = reshape_56, mode = scatter_11_mode_0, updates = reshape_57_cast_fp16, validate_indices = scatter_11_validate_indices_0)[name = tensor("scatter_11_cast_fp16")]; + tensor reshape_59_cast_fp16 = reshape(shape = shape_59, x = scatter_11_cast_fp16)[name = tensor("reshape_59_cast_fp16")]; + tensor reshape_59_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("reshape_59_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor keys_31_begin_0 = const()[name = tensor("keys_31_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor keys_31_end_0 = const()[name = tensor("keys_31_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor keys_31_end_mask_0 = const()[name = tensor("keys_31_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor keys_31_squeeze_mask_0 = const()[name = tensor("keys_31_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor keys_31_cast_fp16 = slice_by_index(begin = keys_31_begin_0, end = keys_31_end_0, end_mask = keys_31_end_mask_0, squeeze_mask = keys_31_squeeze_mask_0, x = reshape_59_cast_fp16)[name = tensor("keys_31_cast_fp16")]; + tensor values_31_begin_0 = const()[name = tensor("values_31_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor values_31_end_0 = const()[name = tensor("values_31_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor values_31_end_mask_0 = const()[name = tensor("values_31_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor values_31_squeeze_mask_0 = const()[name = tensor("values_31_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor values_31_cast_fp16 = slice_by_index(begin = values_31_begin_0, end = values_31_end_0, end_mask = values_31_end_mask_0, squeeze_mask = values_31_squeeze_mask_0, x = reshape_59_cast_fp16)[name = tensor("values_31_cast_fp16")]; + tensor var_2387_cast_fp16 = not_equal(x = keys_31_cast_fp16, y = keys_31_cast_fp16)[name = tensor("op_2387_cast_fp16")]; + tensor keys_33_cast_fp16 = select(a = var_463_to_fp16, b = keys_31_cast_fp16, cond = var_2387_cast_fp16)[name = tensor("keys_33_cast_fp16")]; + tensor var_2395_cast_fp16 = not_equal(x = values_31_cast_fp16, y = values_31_cast_fp16)[name = tensor("op_2395_cast_fp16")]; + tensor values_33_cast_fp16 = select(a = var_463_to_fp16, b = values_31_cast_fp16, cond = var_2395_cast_fp16)[name = tensor("values_33_cast_fp16")]; + tensor var_2419 = const()[name = tensor("op_2419"), val = tensor([0, 2, 1, 3])]; + tensor var_2435 = const()[name = tensor("op_2435"), val = tensor([1, 1, 1])]; + tensor var_2436_cast_fp16 = reshape(shape = var_2435, x = position5_to_fp16)[name = tensor("op_2436_cast_fp16")]; + tensor var_2456_to_fp16 = const()[name = tensor("op_2456_to_fp16"), val = tensor(0x1p+0)]; + tensor valid_len_11_cast_fp16 = add(x = var_2436_cast_fp16, y = var_2456_to_fp16)[name = tensor("valid_len_11_cast_fp16")]; + tensor valid_mask_11_cast_fp16 = less(x = k_positions_1_promoted_to_fp16, y = valid_len_11_cast_fp16)[name = tensor("valid_mask_11_cast_fp16")]; + tensor causal_mask_11_cast_fp16 = less_equal(x = k_positions_1_promoted_to_fp16, y = var_2436_cast_fp16)[name = tensor("causal_mask_11_cast_fp16")]; + tensor attn_mask_21 = logical_and(x = valid_mask_11_cast_fp16, y = causal_mask_11_cast_fp16)[name = tensor("attn_mask_21")]; + tensor attn_mask_23_axes_0 = const()[name = tensor("attn_mask_23_axes_0"), val = tensor([1])]; + tensor attn_mask_23 = expand_dims(axes = attn_mask_23_axes_0, x = attn_mask_21)[name = tensor("attn_mask_23")]; + tensor var_2474_transpose_x_0 = const()[name = tensor("op_2474_transpose_x_0"), val = tensor(false)]; + tensor var_2474_transpose_y_0 = const()[name = tensor("op_2474_transpose_y_0"), val = tensor(false)]; + tensor transpose_82_perm_0 = const()[name = tensor("transpose_82_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_83_perm_0 = const()[name = tensor("transpose_83_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_83 = transpose(perm = transpose_83_perm_0, x = keys_33_cast_fp16)[name = tensor("transpose_193")]; + tensor transpose_82 = transpose(perm = transpose_82_perm_0, x = q_33_cast_fp16)[name = tensor("transpose_194")]; + tensor var_2474_cast_fp16 = matmul(transpose_x = var_2474_transpose_x_0, transpose_y = var_2474_transpose_y_0, x = transpose_82, y = transpose_83)[name = tensor("op_2474_cast_fp16")]; + tensor var_2468_to_fp16 = const()[name = tensor("op_2468_to_fp16"), val = tensor([0x1p-3])]; + tensor attn_weights_31_cast_fp16 = mul(x = var_2474_cast_fp16, y = var_2468_to_fp16)[name = tensor("attn_weights_31_cast_fp16")]; + tensor var_2476 = logical_not(x = attn_mask_23)[name = tensor("op_2476")]; + tensor var_2477_to_fp16 = const()[name = tensor("op_2477_to_fp16"), val = tensor(-inf)]; + tensor attn_weights_33_cast_fp16 = select(a = var_2477_to_fp16, b = attn_weights_31_cast_fp16, cond = var_2476)[name = tensor("attn_weights_33_cast_fp16")]; + tensor var_2479 = const()[name = tensor("op_2479"), val = tensor(-1)]; + tensor attn_weights_35_cast_fp16 = softmax(axis = var_2479, x = attn_weights_33_cast_fp16)[name = tensor("attn_weights_35_cast_fp16")]; + tensor attn_output_11_transpose_x_0 = const()[name = tensor("attn_output_11_transpose_x_0"), val = tensor(false)]; + tensor attn_output_11_transpose_y_0 = const()[name = tensor("attn_output_11_transpose_y_0"), val = tensor(false)]; + tensor values_35_cast_fp16 = transpose(perm = var_2419, x = values_33_cast_fp16)[name = tensor("transpose_195")]; + tensor attn_output_11_cast_fp16 = matmul(transpose_x = attn_output_11_transpose_x_0, transpose_y = attn_output_11_transpose_y_0, x = attn_weights_35_cast_fp16, y = values_35_cast_fp16)[name = tensor("attn_output_11_cast_fp16")]; + tensor var_2487 = const()[name = tensor("op_2487"), val = tensor([0, 2, 1, 3])]; + tensor var_2492 = const()[name = tensor("op_2492"), val = tensor([1, 1, 1024])]; + tensor var_2488_cast_fp16 = transpose(perm = var_2487, x = attn_output_11_cast_fp16)[name = tensor("transpose_192")]; + tensor input_67_cast_fp16 = reshape(shape = var_2492, x = var_2488_cast_fp16)[name = tensor("input_67_cast_fp16")]; + tensor attn5_out_proj_weight_to_fp16 = const()[name = tensor("attn5_out_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(133301120)))]; + tensor linear_22_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = attn5_out_proj_weight_to_fp16, x = input_67_cast_fp16)[name = tensor("linear_22_cast_fp16")]; + tensor var_2498_to_fp16 = const()[name = tensor("op_2498_to_fp16"), val = tensor(0x1p+0)]; + tensor var_2499_cast_fp16 = add(x = position5_to_fp16, y = var_2498_to_fp16)[name = tensor("op_2499_cast_fp16")]; + tensor var_2499_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_2499_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor input_69_cast_fp16 = add(x = input_63_cast_fp16, y = linear_22_cast_fp16)[name = tensor("input_69_cast_fp16")]; + tensor input_71_axes_0 = const()[name = tensor("input_71_axes_0"), val = tensor([-1])]; + tensor norm5_2_weight_to_fp16 = const()[name = tensor("norm5_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(135398336)))]; + tensor norm5_2_bias_to_fp16 = const()[name = tensor("norm5_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(135400448)))]; + tensor var_2503_to_fp16 = const()[name = tensor("op_2503_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_71_cast_fp16 = layer_norm(axes = input_71_axes_0, beta = norm5_2_bias_to_fp16, epsilon = var_2503_to_fp16, gamma = norm5_2_weight_to_fp16, x = input_69_cast_fp16)[name = tensor("input_71_cast_fp16")]; + tensor linear5_1_weight_to_fp16 = const()[name = tensor("linear5_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(135402560)))]; + tensor linear_23_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = linear5_1_weight_to_fp16, x = input_71_cast_fp16)[name = tensor("linear_23_cast_fp16")]; + tensor input_73_mode_0 = const()[name = tensor("input_73_mode_0"), val = tensor("EXACT")]; + tensor input_73_cast_fp16 = gelu(mode = input_73_mode_0, x = linear_23_cast_fp16)[name = tensor("input_73_cast_fp16")]; + tensor linear5_2_weight_to_fp16 = const()[name = tensor("linear5_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(143791232)))]; + tensor linear_24_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = linear5_2_weight_to_fp16, x = input_73_cast_fp16)[name = tensor("linear_24_cast_fp16")]; + tensor input_75_cast_fp16 = add(x = input_69_cast_fp16, y = linear_24_cast_fp16)[name = tensor("input_75_cast_fp16")]; + tensor input_77_axes_0 = const()[name = tensor("input_77_axes_0"), val = tensor([-1])]; + tensor norm6_1_weight_to_fp16 = const()[name = tensor("norm6_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(152179904)))]; + tensor norm6_1_bias_to_fp16 = const()[name = tensor("norm6_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(152182016)))]; + tensor var_2520_to_fp16 = const()[name = tensor("op_2520_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_77_cast_fp16 = layer_norm(axes = input_77_axes_0, beta = norm6_1_bias_to_fp16, epsilon = var_2520_to_fp16, gamma = norm6_1_weight_to_fp16, x = input_75_cast_fp16)[name = tensor("input_77_cast_fp16")]; + tensor attn6_in_proj_weight_to_fp16 = const()[name = tensor("attn6_in_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(152184128)))]; + tensor linear_25_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = attn6_in_proj_weight_to_fp16, x = input_77_cast_fp16)[name = tensor("linear_25_cast_fp16")]; + tensor var_2534 = const()[name = tensor("op_2534"), val = tensor([1, 1, 3, 16, 64])]; + tensor qkv_13_cast_fp16 = reshape(shape = var_2534, x = linear_25_cast_fp16)[name = tensor("qkv_13_cast_fp16")]; + tensor q_37_begin_0 = const()[name = tensor("q_37_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor q_37_end_0 = const()[name = tensor("q_37_end_0"), val = tensor([1, 1, 1, 16, 64])]; + tensor q_37_end_mask_0 = const()[name = tensor("q_37_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor q_37_squeeze_mask_0 = const()[name = tensor("q_37_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor q_37_cast_fp16 = slice_by_index(begin = q_37_begin_0, end = q_37_end_0, end_mask = q_37_end_mask_0, squeeze_mask = q_37_squeeze_mask_0, x = qkv_13_cast_fp16)[name = tensor("q_37_cast_fp16")]; + tensor k_25_begin_0 = const()[name = tensor("k_25_begin_0"), val = tensor([0, 0, 1, 0, 0])]; + tensor k_25_end_0 = const()[name = tensor("k_25_end_0"), val = tensor([1, 1, 2, 16, 64])]; + tensor k_25_end_mask_0 = const()[name = tensor("k_25_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor k_25_squeeze_mask_0 = const()[name = tensor("k_25_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor k_25_cast_fp16 = slice_by_index(begin = k_25_begin_0, end = k_25_end_0, end_mask = k_25_end_mask_0, squeeze_mask = k_25_squeeze_mask_0, x = qkv_13_cast_fp16)[name = tensor("k_25_cast_fp16")]; + tensor v_13_begin_0 = const()[name = tensor("v_13_begin_0"), val = tensor([0, 0, 2, 0, 0])]; + tensor v_13_end_0 = const()[name = tensor("v_13_end_0"), val = tensor([1, 1, 3, 16, 64])]; + tensor v_13_end_mask_0 = const()[name = tensor("v_13_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor v_13_squeeze_mask_0 = const()[name = tensor("v_13_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor v_13_cast_fp16 = slice_by_index(begin = v_13_begin_0, end = v_13_end_0, end_mask = v_13_end_mask_0, squeeze_mask = v_13_squeeze_mask_0, x = qkv_13_cast_fp16)[name = tensor("v_13_cast_fp16")]; + tensor var_2600 = const()[name = tensor("op_2600"), val = tensor([1, 1, 1, 1])]; + tensor position6_to_fp16_dtype_0 = const()[name = tensor("position6_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor position6_to_fp16 = cast(dtype = position6_to_fp16_dtype_0, x = position6)[name = tensor("cast_523")]; + tensor ts_41_cast_fp16 = reshape(shape = var_2600, x = position6_to_fp16)[name = tensor("ts_41_cast_fp16")]; + tensor var_2607 = const()[name = tensor("op_2607"), val = tensor([1, 1, 16, 32, 2])]; + tensor q_complex_13_cast_fp16 = reshape(shape = var_2607, x = q_37_cast_fp16)[name = tensor("q_complex_13_cast_fp16")]; + tensor var_2614 = const()[name = tensor("op_2614"), val = tensor([1, 1, 16, 32, 2])]; + tensor k_complex_13_cast_fp16 = reshape(shape = var_2614, x = k_25_cast_fp16)[name = tensor("k_complex_13_cast_fp16")]; + tensor var_2618_begin_0 = const()[name = tensor("op_2618_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_2618_end_0 = const()[name = tensor("op_2618_end_0"), val = tensor([1, 1, 16, 32, 1])]; + tensor var_2618_end_mask_0 = const()[name = tensor("op_2618_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_2618_squeeze_mask_0 = const()[name = tensor("op_2618_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_2618_cast_fp16 = slice_by_index(begin = var_2618_begin_0, end = var_2618_end_0, end_mask = var_2618_end_mask_0, squeeze_mask = var_2618_squeeze_mask_0, x = q_complex_13_cast_fp16)[name = tensor("op_2618_cast_fp16")]; + tensor var_2626_begin_0 = const()[name = tensor("op_2626_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_2626_end_0 = const()[name = tensor("op_2626_end_0"), val = tensor([1, 1, 16, 32, 2])]; + tensor var_2626_end_mask_0 = const()[name = tensor("op_2626_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_2626_squeeze_mask_0 = const()[name = tensor("op_2626_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_2626_cast_fp16 = slice_by_index(begin = var_2626_begin_0, end = var_2626_end_0, end_mask = var_2626_end_mask_0, squeeze_mask = var_2626_squeeze_mask_0, x = q_complex_13_cast_fp16)[name = tensor("op_2626_cast_fp16")]; + tensor var_2634_begin_0 = const()[name = tensor("op_2634_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_2634_end_0 = const()[name = tensor("op_2634_end_0"), val = tensor([1, 1, 16, 32, 1])]; + tensor var_2634_end_mask_0 = const()[name = tensor("op_2634_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_2634_squeeze_mask_0 = const()[name = tensor("op_2634_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_2634_cast_fp16 = slice_by_index(begin = var_2634_begin_0, end = var_2634_end_0, end_mask = var_2634_end_mask_0, squeeze_mask = var_2634_squeeze_mask_0, x = k_complex_13_cast_fp16)[name = tensor("op_2634_cast_fp16")]; + tensor var_2642_begin_0 = const()[name = tensor("op_2642_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_2642_end_0 = const()[name = tensor("op_2642_end_0"), val = tensor([1, 1, 16, 32, 2])]; + tensor var_2642_end_mask_0 = const()[name = tensor("op_2642_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_2642_squeeze_mask_0 = const()[name = tensor("op_2642_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_2642_cast_fp16 = slice_by_index(begin = var_2642_begin_0, end = var_2642_end_0, end_mask = var_2642_end_mask_0, squeeze_mask = var_2642_squeeze_mask_0, x = k_complex_13_cast_fp16)[name = tensor("op_2642_cast_fp16")]; + tensor freqs_13_to_fp16 = const()[name = tensor("freqs_13_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(158475648)))]; + tensor var_2648_cast_fp16 = mul(x = freqs_13_to_fp16, y = ts_41_cast_fp16)[name = tensor("op_2648_cast_fp16")]; + tensor rotr_13_cast_fp16 = cos(x = var_2648_cast_fp16)[name = tensor("rotr_13_cast_fp16")]; + tensor roti_13_cast_fp16 = sin(x = var_2648_cast_fp16)[name = tensor("roti_13_cast_fp16")]; + tensor var_2652_cast_fp16 = mul(x = var_2618_cast_fp16, y = rotr_13_cast_fp16)[name = tensor("op_2652_cast_fp16")]; + tensor var_2653_cast_fp16 = mul(x = var_2626_cast_fp16, y = roti_13_cast_fp16)[name = tensor("op_2653_cast_fp16")]; + tensor qor_25_cast_fp16 = sub(x = var_2652_cast_fp16, y = var_2653_cast_fp16)[name = tensor("qor_25_cast_fp16")]; + tensor var_2656_cast_fp16 = mul(x = var_2618_cast_fp16, y = roti_13_cast_fp16)[name = tensor("op_2656_cast_fp16")]; + tensor var_2657_cast_fp16 = mul(x = var_2626_cast_fp16, y = rotr_13_cast_fp16)[name = tensor("op_2657_cast_fp16")]; + tensor qoi_25_cast_fp16 = add(x = var_2656_cast_fp16, y = var_2657_cast_fp16)[name = tensor("qoi_25_cast_fp16")]; + tensor var_2660_cast_fp16 = mul(x = var_2634_cast_fp16, y = rotr_13_cast_fp16)[name = tensor("op_2660_cast_fp16")]; + tensor var_2661_cast_fp16 = mul(x = var_2642_cast_fp16, y = roti_13_cast_fp16)[name = tensor("op_2661_cast_fp16")]; + tensor kor_25_cast_fp16 = sub(x = var_2660_cast_fp16, y = var_2661_cast_fp16)[name = tensor("kor_25_cast_fp16")]; + tensor var_2664_cast_fp16 = mul(x = var_2634_cast_fp16, y = roti_13_cast_fp16)[name = tensor("op_2664_cast_fp16")]; + tensor var_2665_cast_fp16 = mul(x = var_2642_cast_fp16, y = rotr_13_cast_fp16)[name = tensor("op_2665_cast_fp16")]; + tensor koi_25_cast_fp16 = add(x = var_2664_cast_fp16, y = var_2665_cast_fp16)[name = tensor("koi_25_cast_fp16")]; + tensor qo_13_axis_0 = const()[name = tensor("qo_13_axis_0"), val = tensor(-1)]; + tensor qo_13_cast_fp16 = stack(axis = qo_13_axis_0, values = (qor_25_cast_fp16, qoi_25_cast_fp16))[name = tensor("qo_13_cast_fp16")]; + tensor ko_13_axis_0 = const()[name = tensor("ko_13_axis_0"), val = tensor(-1)]; + tensor ko_13_cast_fp16 = stack(axis = ko_13_axis_0, values = (kor_25_cast_fp16, koi_25_cast_fp16))[name = tensor("ko_13_cast_fp16")]; + tensor var_2698 = const()[name = tensor("op_2698"), val = tensor([1, 1, 16, 64])]; + tensor q_39_cast_fp16 = reshape(shape = var_2698, x = qo_13_cast_fp16)[name = tensor("q_39_cast_fp16")]; + tensor var_2704 = const()[name = tensor("op_2704"), val = tensor([1, 1, 16, 64])]; + tensor k_27_cast_fp16 = reshape(shape = var_2704, x = ko_13_cast_fp16)[name = tensor("k_27_cast_fp16")]; + tensor _inversed_2729_y_0_to_fp16 = const()[name = tensor("_inversed_2729_y_0_to_fp16"), val = tensor(0x1p-9)]; + tensor _inversed_2729_cast_fp16 = mul(x = ts_41_cast_fp16, y = _inversed_2729_y_0_to_fp16)[name = tensor("_inversed_2729_cast_fp16")]; + tensor var_2730_cast_fp16 = floor(x = _inversed_2729_cast_fp16)[name = tensor("op_2730_cast_fp16")]; + tensor var_2731_to_fp16 = const()[name = tensor("op_2731_to_fp16"), val = tensor(0x1p+9)]; + tensor var_2732_cast_fp16 = mul(x = var_2730_cast_fp16, y = var_2731_to_fp16)[name = tensor("op_2732_cast_fp16")]; + tensor write_indices_float_27_cast_fp16 = sub(x = ts_41_cast_fp16, y = var_2732_cast_fp16)[name = tensor("write_indices_float_27_cast_fp16")]; + tensor var_2739_dtype_0 = const()[name = tensor("op_2739_dtype_0"), val = tensor("int32")]; + tensor write_indices_13_reps_0 = const()[name = tensor("write_indices_13_reps_0"), val = tensor([1, 1, 16, 64])]; + tensor write_indices_float_27_cast_fp16_to_int32 = cast(dtype = var_2739_dtype_0, x = write_indices_float_27_cast_fp16)[name = tensor("cast_522")]; + tensor write_indices_13 = tile(reps = write_indices_13_reps_0, x = write_indices_float_27_cast_fp16_to_int32)[name = tensor("write_indices_13")]; + tensor var_2749_begin_0 = const()[name = tensor("op_2749_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_2749_end_0 = const()[name = tensor("op_2749_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor var_2749_end_mask_0 = const()[name = tensor("op_2749_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_2749_squeeze_mask_0 = const()[name = tensor("op_2749_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor cache6_to_fp16_dtype_0 = const()[name = tensor("cache6_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor cache6_to_fp16 = cast(dtype = cache6_to_fp16_dtype_0, x = cache6)[name = tensor("cast_521")]; + tensor var_2749_cast_fp16 = slice_by_index(begin = var_2749_begin_0, end = var_2749_end_0, end_mask = var_2749_end_mask_0, squeeze_mask = var_2749_squeeze_mask_0, x = cache6_to_fp16)[name = tensor("op_2749_cast_fp16")]; + tensor var_2751_axis_0 = const()[name = tensor("op_2751_axis_0"), val = tensor(1)]; + tensor var_2751_mode_0 = const()[name = tensor("op_2751_mode_0"), val = tensor("update")]; + tensor var_2751_validate_indices_0 = const()[name = tensor("op_2751_validate_indices_0"), val = tensor(false)]; + tensor var_2751_cast_fp16 = scatter_along_axis(axis = var_2751_axis_0, data = var_2749_cast_fp16, indices = write_indices_13, mode = var_2751_mode_0, updates = k_27_cast_fp16, validate_indices = var_2751_validate_indices_0)[name = tensor("op_2751_cast_fp16")]; + tensor concat_44 = const()[name = tensor("concat_44"), val = tensor([0, 0, 0, 0, 0])]; + tensor concat_45 = const()[name = tensor("concat_45"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_13_internal_tensor_assign_1_stride_0 = const()[name = tensor("new_cache_13_internal_tensor_assign_1_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_13_internal_tensor_assign_1_begin_mask_0 = const()[name = tensor("new_cache_13_internal_tensor_assign_1_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_13_internal_tensor_assign_1_end_mask_0 = const()[name = tensor("new_cache_13_internal_tensor_assign_1_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_13_internal_tensor_assign_1_squeeze_mask_0 = const()[name = tensor("new_cache_13_internal_tensor_assign_1_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_60 = const()[name = tensor("shape_60"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_12 = const()[name = tensor("reduce_prod_12"), val = tensor(1048576)]; + tensor range_1d_12_start_0 = const()[name = tensor("range_1d_12_start_0"), val = tensor(0)]; + tensor range_1d_12_step_0 = const()[name = tensor("range_1d_12_step_0"), val = tensor(1)]; + tensor range_1d_12 = range_1d(end = reduce_prod_12, start = range_1d_12_start_0, step = range_1d_12_step_0)[name = tensor("range_1d_12")]; + tensor reshape_60 = reshape(shape = shape_60, x = range_1d_12)[name = tensor("reshape_60")]; + tensor slice_by_index_12 = slice_by_index(begin = concat_44, begin_mask = new_cache_13_internal_tensor_assign_1_begin_mask_0, end = concat_45, end_mask = new_cache_13_internal_tensor_assign_1_end_mask_0, squeeze_mask = new_cache_13_internal_tensor_assign_1_squeeze_mask_0, stride = new_cache_13_internal_tensor_assign_1_stride_0, x = reshape_60)[name = tensor("slice_by_index_12")]; + tensor reshape_61_shape_0 = const()[name = tensor("reshape_61_shape_0"), val = tensor([-1])]; + tensor reshape_61 = reshape(shape = reshape_61_shape_0, x = slice_by_index_12)[name = tensor("reshape_61")]; + tensor reshape_62_shape_0 = const()[name = tensor("reshape_62_shape_0"), val = tensor([-1])]; + tensor reshape_62_cast_fp16 = reshape(shape = reshape_62_shape_0, x = var_2751_cast_fp16)[name = tensor("reshape_62_cast_fp16")]; + tensor reshape_63_shape_0 = const()[name = tensor("reshape_63_shape_0"), val = tensor([-1])]; + tensor reshape_63_cast_fp16 = reshape(shape = reshape_63_shape_0, x = cache6_to_fp16)[name = tensor("reshape_63_cast_fp16")]; + tensor scatter_12_mode_0 = const()[name = tensor("scatter_12_mode_0"), val = tensor("update")]; + tensor scatter_12_axis_0 = const()[name = tensor("scatter_12_axis_0"), val = tensor(0)]; + tensor scatter_12_validate_indices_0 = const()[name = tensor("scatter_12_validate_indices_0"), val = tensor(false)]; + tensor scatter_12_cast_fp16 = scatter(axis = scatter_12_axis_0, data = reshape_63_cast_fp16, indices = reshape_61, mode = scatter_12_mode_0, updates = reshape_62_cast_fp16, validate_indices = scatter_12_validate_indices_0)[name = tensor("scatter_12_cast_fp16")]; + tensor reshape_64_cast_fp16 = reshape(shape = shape_60, x = scatter_12_cast_fp16)[name = tensor("reshape_64_cast_fp16")]; + tensor var_2759_begin_0 = const()[name = tensor("op_2759_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor var_2759_end_0 = const()[name = tensor("op_2759_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor var_2759_end_mask_0 = const()[name = tensor("op_2759_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_2759_squeeze_mask_0 = const()[name = tensor("op_2759_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor var_2759_cast_fp16 = slice_by_index(begin = var_2759_begin_0, end = var_2759_end_0, end_mask = var_2759_end_mask_0, squeeze_mask = var_2759_squeeze_mask_0, x = reshape_64_cast_fp16)[name = tensor("op_2759_cast_fp16")]; + tensor var_2761_axis_0 = const()[name = tensor("op_2761_axis_0"), val = tensor(1)]; + tensor var_2761_mode_0 = const()[name = tensor("op_2761_mode_0"), val = tensor("update")]; + tensor var_2761_validate_indices_0 = const()[name = tensor("op_2761_validate_indices_0"), val = tensor(false)]; + tensor var_2761_cast_fp16 = scatter_along_axis(axis = var_2761_axis_0, data = var_2759_cast_fp16, indices = write_indices_13, mode = var_2761_mode_0, updates = v_13_cast_fp16, validate_indices = var_2761_validate_indices_0)[name = tensor("op_2761_cast_fp16")]; + tensor concat_46 = const()[name = tensor("concat_46"), val = tensor([1, 0, 0, 0, 0])]; + tensor concat_47 = const()[name = tensor("concat_47"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_13_internal_tensor_assign_2_stride_0 = const()[name = tensor("new_cache_13_internal_tensor_assign_2_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_13_internal_tensor_assign_2_begin_mask_0 = const()[name = tensor("new_cache_13_internal_tensor_assign_2_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_13_internal_tensor_assign_2_end_mask_0 = const()[name = tensor("new_cache_13_internal_tensor_assign_2_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_13_internal_tensor_assign_2_squeeze_mask_0 = const()[name = tensor("new_cache_13_internal_tensor_assign_2_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_61 = const()[name = tensor("shape_61"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_13 = const()[name = tensor("reduce_prod_13"), val = tensor(1048576)]; + tensor range_1d_13_start_0 = const()[name = tensor("range_1d_13_start_0"), val = tensor(0)]; + tensor range_1d_13_step_0 = const()[name = tensor("range_1d_13_step_0"), val = tensor(1)]; + tensor range_1d_13 = range_1d(end = reduce_prod_13, start = range_1d_13_start_0, step = range_1d_13_step_0)[name = tensor("range_1d_13")]; + tensor reshape_65 = reshape(shape = shape_61, x = range_1d_13)[name = tensor("reshape_65")]; + tensor slice_by_index_13 = slice_by_index(begin = concat_46, begin_mask = new_cache_13_internal_tensor_assign_2_begin_mask_0, end = concat_47, end_mask = new_cache_13_internal_tensor_assign_2_end_mask_0, squeeze_mask = new_cache_13_internal_tensor_assign_2_squeeze_mask_0, stride = new_cache_13_internal_tensor_assign_2_stride_0, x = reshape_65)[name = tensor("slice_by_index_13")]; + tensor reshape_66_shape_0 = const()[name = tensor("reshape_66_shape_0"), val = tensor([-1])]; + tensor reshape_66 = reshape(shape = reshape_66_shape_0, x = slice_by_index_13)[name = tensor("reshape_66")]; + tensor reshape_67_shape_0 = const()[name = tensor("reshape_67_shape_0"), val = tensor([-1])]; + tensor reshape_67_cast_fp16 = reshape(shape = reshape_67_shape_0, x = var_2761_cast_fp16)[name = tensor("reshape_67_cast_fp16")]; + tensor reshape_68_shape_0 = const()[name = tensor("reshape_68_shape_0"), val = tensor([-1])]; + tensor reshape_68_cast_fp16 = reshape(shape = reshape_68_shape_0, x = reshape_64_cast_fp16)[name = tensor("reshape_68_cast_fp16")]; + tensor scatter_13_mode_0 = const()[name = tensor("scatter_13_mode_0"), val = tensor("update")]; + tensor scatter_13_axis_0 = const()[name = tensor("scatter_13_axis_0"), val = tensor(0)]; + tensor scatter_13_validate_indices_0 = const()[name = tensor("scatter_13_validate_indices_0"), val = tensor(false)]; + tensor scatter_13_cast_fp16 = scatter(axis = scatter_13_axis_0, data = reshape_68_cast_fp16, indices = reshape_66, mode = scatter_13_mode_0, updates = reshape_67_cast_fp16, validate_indices = scatter_13_validate_indices_0)[name = tensor("scatter_13_cast_fp16")]; + tensor reshape_69_cast_fp16 = reshape(shape = shape_61, x = scatter_13_cast_fp16)[name = tensor("reshape_69_cast_fp16")]; + tensor reshape_69_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("reshape_69_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor keys_37_begin_0 = const()[name = tensor("keys_37_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor keys_37_end_0 = const()[name = tensor("keys_37_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor keys_37_end_mask_0 = const()[name = tensor("keys_37_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor keys_37_squeeze_mask_0 = const()[name = tensor("keys_37_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor keys_37_cast_fp16 = slice_by_index(begin = keys_37_begin_0, end = keys_37_end_0, end_mask = keys_37_end_mask_0, squeeze_mask = keys_37_squeeze_mask_0, x = reshape_69_cast_fp16)[name = tensor("keys_37_cast_fp16")]; + tensor values_37_begin_0 = const()[name = tensor("values_37_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor values_37_end_0 = const()[name = tensor("values_37_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor values_37_end_mask_0 = const()[name = tensor("values_37_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor values_37_squeeze_mask_0 = const()[name = tensor("values_37_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor values_37_cast_fp16 = slice_by_index(begin = values_37_begin_0, end = values_37_end_0, end_mask = values_37_end_mask_0, squeeze_mask = values_37_squeeze_mask_0, x = reshape_69_cast_fp16)[name = tensor("values_37_cast_fp16")]; + tensor var_2773_cast_fp16 = not_equal(x = keys_37_cast_fp16, y = keys_37_cast_fp16)[name = tensor("op_2773_cast_fp16")]; + tensor keys_39_cast_fp16 = select(a = var_463_to_fp16, b = keys_37_cast_fp16, cond = var_2773_cast_fp16)[name = tensor("keys_39_cast_fp16")]; + tensor var_2781_cast_fp16 = not_equal(x = values_37_cast_fp16, y = values_37_cast_fp16)[name = tensor("op_2781_cast_fp16")]; + tensor values_39_cast_fp16 = select(a = var_463_to_fp16, b = values_37_cast_fp16, cond = var_2781_cast_fp16)[name = tensor("values_39_cast_fp16")]; + tensor var_2805 = const()[name = tensor("op_2805"), val = tensor([0, 2, 1, 3])]; + tensor var_2821 = const()[name = tensor("op_2821"), val = tensor([1, 1, 1])]; + tensor var_2822_cast_fp16 = reshape(shape = var_2821, x = position6_to_fp16)[name = tensor("op_2822_cast_fp16")]; + tensor var_2842_to_fp16 = const()[name = tensor("op_2842_to_fp16"), val = tensor(0x1p+0)]; + tensor valid_len_13_cast_fp16 = add(x = var_2822_cast_fp16, y = var_2842_to_fp16)[name = tensor("valid_len_13_cast_fp16")]; + tensor valid_mask_13_cast_fp16 = less(x = k_positions_1_promoted_to_fp16, y = valid_len_13_cast_fp16)[name = tensor("valid_mask_13_cast_fp16")]; + tensor causal_mask_13_cast_fp16 = less_equal(x = k_positions_1_promoted_to_fp16, y = var_2822_cast_fp16)[name = tensor("causal_mask_13_cast_fp16")]; + tensor attn_mask_25 = logical_and(x = valid_mask_13_cast_fp16, y = causal_mask_13_cast_fp16)[name = tensor("attn_mask_25")]; + tensor attn_mask_27_axes_0 = const()[name = tensor("attn_mask_27_axes_0"), val = tensor([1])]; + tensor attn_mask_27 = expand_dims(axes = attn_mask_27_axes_0, x = attn_mask_25)[name = tensor("attn_mask_27")]; + tensor var_2860_transpose_x_0 = const()[name = tensor("op_2860_transpose_x_0"), val = tensor(false)]; + tensor var_2860_transpose_y_0 = const()[name = tensor("op_2860_transpose_y_0"), val = tensor(false)]; + tensor transpose_84_perm_0 = const()[name = tensor("transpose_84_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_85_perm_0 = const()[name = tensor("transpose_85_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_85 = transpose(perm = transpose_85_perm_0, x = keys_39_cast_fp16)[name = tensor("transpose_189")]; + tensor transpose_84 = transpose(perm = transpose_84_perm_0, x = q_39_cast_fp16)[name = tensor("transpose_190")]; + tensor var_2860_cast_fp16 = matmul(transpose_x = var_2860_transpose_x_0, transpose_y = var_2860_transpose_y_0, x = transpose_84, y = transpose_85)[name = tensor("op_2860_cast_fp16")]; + tensor var_2854_to_fp16 = const()[name = tensor("op_2854_to_fp16"), val = tensor([0x1p-3])]; + tensor attn_weights_37_cast_fp16 = mul(x = var_2860_cast_fp16, y = var_2854_to_fp16)[name = tensor("attn_weights_37_cast_fp16")]; + tensor var_2862 = logical_not(x = attn_mask_27)[name = tensor("op_2862")]; + tensor var_2863_to_fp16 = const()[name = tensor("op_2863_to_fp16"), val = tensor(-inf)]; + tensor attn_weights_39_cast_fp16 = select(a = var_2863_to_fp16, b = attn_weights_37_cast_fp16, cond = var_2862)[name = tensor("attn_weights_39_cast_fp16")]; + tensor var_2865 = const()[name = tensor("op_2865"), val = tensor(-1)]; + tensor attn_weights_41_cast_fp16 = softmax(axis = var_2865, x = attn_weights_39_cast_fp16)[name = tensor("attn_weights_41_cast_fp16")]; + tensor attn_output_13_transpose_x_0 = const()[name = tensor("attn_output_13_transpose_x_0"), val = tensor(false)]; + tensor attn_output_13_transpose_y_0 = const()[name = tensor("attn_output_13_transpose_y_0"), val = tensor(false)]; + tensor values_41_cast_fp16 = transpose(perm = var_2805, x = values_39_cast_fp16)[name = tensor("transpose_191")]; + tensor attn_output_13_cast_fp16 = matmul(transpose_x = attn_output_13_transpose_x_0, transpose_y = attn_output_13_transpose_y_0, x = attn_weights_41_cast_fp16, y = values_41_cast_fp16)[name = tensor("attn_output_13_cast_fp16")]; + tensor var_2873 = const()[name = tensor("op_2873"), val = tensor([0, 2, 1, 3])]; + tensor var_2878 = const()[name = tensor("op_2878"), val = tensor([1, 1, 1024])]; + tensor var_2874_cast_fp16 = transpose(perm = var_2873, x = attn_output_13_cast_fp16)[name = tensor("transpose_188")]; + tensor input_79_cast_fp16 = reshape(shape = var_2878, x = var_2874_cast_fp16)[name = tensor("input_79_cast_fp16")]; + tensor attn6_out_proj_weight_to_fp16 = const()[name = tensor("attn6_out_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(158475776)))]; + tensor linear_26_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = attn6_out_proj_weight_to_fp16, x = input_79_cast_fp16)[name = tensor("linear_26_cast_fp16")]; + tensor var_2884_to_fp16 = const()[name = tensor("op_2884_to_fp16"), val = tensor(0x1p+0)]; + tensor var_2885_cast_fp16 = add(x = position6_to_fp16, y = var_2884_to_fp16)[name = tensor("op_2885_cast_fp16")]; + tensor var_2885_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_2885_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor input_81_cast_fp16 = add(x = input_75_cast_fp16, y = linear_26_cast_fp16)[name = tensor("input_81_cast_fp16")]; + tensor input_83_axes_0 = const()[name = tensor("input_83_axes_0"), val = tensor([-1])]; + tensor norm6_2_weight_to_fp16 = const()[name = tensor("norm6_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(160572992)))]; + tensor norm6_2_bias_to_fp16 = const()[name = tensor("norm6_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(160575104)))]; + tensor var_2889_to_fp16 = const()[name = tensor("op_2889_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_83_cast_fp16 = layer_norm(axes = input_83_axes_0, beta = norm6_2_bias_to_fp16, epsilon = var_2889_to_fp16, gamma = norm6_2_weight_to_fp16, x = input_81_cast_fp16)[name = tensor("input_83_cast_fp16")]; + tensor linear6_1_weight_to_fp16 = const()[name = tensor("linear6_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(160577216)))]; + tensor linear_27_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = linear6_1_weight_to_fp16, x = input_83_cast_fp16)[name = tensor("linear_27_cast_fp16")]; + tensor input_85_mode_0 = const()[name = tensor("input_85_mode_0"), val = tensor("EXACT")]; + tensor input_85_cast_fp16 = gelu(mode = input_85_mode_0, x = linear_27_cast_fp16)[name = tensor("input_85_cast_fp16")]; + tensor linear6_2_weight_to_fp16 = const()[name = tensor("linear6_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(168965888)))]; + tensor linear_28_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = linear6_2_weight_to_fp16, x = input_85_cast_fp16)[name = tensor("linear_28_cast_fp16")]; + tensor input_87_cast_fp16 = add(x = input_81_cast_fp16, y = linear_28_cast_fp16)[name = tensor("input_87_cast_fp16")]; + tensor input_89_axes_0 = const()[name = tensor("input_89_axes_0"), val = tensor([-1])]; + tensor norm7_1_weight_to_fp16 = const()[name = tensor("norm7_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(177354560)))]; + tensor norm7_1_bias_to_fp16 = const()[name = tensor("norm7_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(177356672)))]; + tensor var_2906_to_fp16 = const()[name = tensor("op_2906_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_89_cast_fp16 = layer_norm(axes = input_89_axes_0, beta = norm7_1_bias_to_fp16, epsilon = var_2906_to_fp16, gamma = norm7_1_weight_to_fp16, x = input_87_cast_fp16)[name = tensor("input_89_cast_fp16")]; + tensor attn7_in_proj_weight_to_fp16 = const()[name = tensor("attn7_in_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(177358784)))]; + tensor linear_29_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = attn7_in_proj_weight_to_fp16, x = input_89_cast_fp16)[name = tensor("linear_29_cast_fp16")]; + tensor var_2920 = const()[name = tensor("op_2920"), val = tensor([1, 1, 3, 16, 64])]; + tensor qkv_15_cast_fp16 = reshape(shape = var_2920, x = linear_29_cast_fp16)[name = tensor("qkv_15_cast_fp16")]; + tensor q_43_begin_0 = const()[name = tensor("q_43_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor q_43_end_0 = const()[name = tensor("q_43_end_0"), val = tensor([1, 1, 1, 16, 64])]; + tensor q_43_end_mask_0 = const()[name = tensor("q_43_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor q_43_squeeze_mask_0 = const()[name = tensor("q_43_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor q_43_cast_fp16 = slice_by_index(begin = q_43_begin_0, end = q_43_end_0, end_mask = q_43_end_mask_0, squeeze_mask = q_43_squeeze_mask_0, x = qkv_15_cast_fp16)[name = tensor("q_43_cast_fp16")]; + tensor k_29_begin_0 = const()[name = tensor("k_29_begin_0"), val = tensor([0, 0, 1, 0, 0])]; + tensor k_29_end_0 = const()[name = tensor("k_29_end_0"), val = tensor([1, 1, 2, 16, 64])]; + tensor k_29_end_mask_0 = const()[name = tensor("k_29_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor k_29_squeeze_mask_0 = const()[name = tensor("k_29_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor k_29_cast_fp16 = slice_by_index(begin = k_29_begin_0, end = k_29_end_0, end_mask = k_29_end_mask_0, squeeze_mask = k_29_squeeze_mask_0, x = qkv_15_cast_fp16)[name = tensor("k_29_cast_fp16")]; + tensor v_15_begin_0 = const()[name = tensor("v_15_begin_0"), val = tensor([0, 0, 2, 0, 0])]; + tensor v_15_end_0 = const()[name = tensor("v_15_end_0"), val = tensor([1, 1, 3, 16, 64])]; + tensor v_15_end_mask_0 = const()[name = tensor("v_15_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor v_15_squeeze_mask_0 = const()[name = tensor("v_15_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor v_15_cast_fp16 = slice_by_index(begin = v_15_begin_0, end = v_15_end_0, end_mask = v_15_end_mask_0, squeeze_mask = v_15_squeeze_mask_0, x = qkv_15_cast_fp16)[name = tensor("v_15_cast_fp16")]; + tensor var_2986 = const()[name = tensor("op_2986"), val = tensor([1, 1, 1, 1])]; + tensor position7_to_fp16_dtype_0 = const()[name = tensor("position7_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor position7_to_fp16 = cast(dtype = position7_to_fp16_dtype_0, x = position7)[name = tensor("cast_518")]; + tensor ts_47_cast_fp16 = reshape(shape = var_2986, x = position7_to_fp16)[name = tensor("ts_47_cast_fp16")]; + tensor var_2993 = const()[name = tensor("op_2993"), val = tensor([1, 1, 16, 32, 2])]; + tensor q_complex_15_cast_fp16 = reshape(shape = var_2993, x = q_43_cast_fp16)[name = tensor("q_complex_15_cast_fp16")]; + tensor var_3000 = const()[name = tensor("op_3000"), val = tensor([1, 1, 16, 32, 2])]; + tensor k_complex_15_cast_fp16 = reshape(shape = var_3000, x = k_29_cast_fp16)[name = tensor("k_complex_15_cast_fp16")]; + tensor var_3004_begin_0 = const()[name = tensor("op_3004_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_3004_end_0 = const()[name = tensor("op_3004_end_0"), val = tensor([1, 1, 16, 32, 1])]; + tensor var_3004_end_mask_0 = const()[name = tensor("op_3004_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_3004_squeeze_mask_0 = const()[name = tensor("op_3004_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_3004_cast_fp16 = slice_by_index(begin = var_3004_begin_0, end = var_3004_end_0, end_mask = var_3004_end_mask_0, squeeze_mask = var_3004_squeeze_mask_0, x = q_complex_15_cast_fp16)[name = tensor("op_3004_cast_fp16")]; + tensor var_3012_begin_0 = const()[name = tensor("op_3012_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_3012_end_0 = const()[name = tensor("op_3012_end_0"), val = tensor([1, 1, 16, 32, 2])]; + tensor var_3012_end_mask_0 = const()[name = tensor("op_3012_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_3012_squeeze_mask_0 = const()[name = tensor("op_3012_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_3012_cast_fp16 = slice_by_index(begin = var_3012_begin_0, end = var_3012_end_0, end_mask = var_3012_end_mask_0, squeeze_mask = var_3012_squeeze_mask_0, x = q_complex_15_cast_fp16)[name = tensor("op_3012_cast_fp16")]; + tensor var_3020_begin_0 = const()[name = tensor("op_3020_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_3020_end_0 = const()[name = tensor("op_3020_end_0"), val = tensor([1, 1, 16, 32, 1])]; + tensor var_3020_end_mask_0 = const()[name = tensor("op_3020_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_3020_squeeze_mask_0 = const()[name = tensor("op_3020_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_3020_cast_fp16 = slice_by_index(begin = var_3020_begin_0, end = var_3020_end_0, end_mask = var_3020_end_mask_0, squeeze_mask = var_3020_squeeze_mask_0, x = k_complex_15_cast_fp16)[name = tensor("op_3020_cast_fp16")]; + tensor var_3028_begin_0 = const()[name = tensor("op_3028_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_3028_end_0 = const()[name = tensor("op_3028_end_0"), val = tensor([1, 1, 16, 32, 2])]; + tensor var_3028_end_mask_0 = const()[name = tensor("op_3028_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_3028_squeeze_mask_0 = const()[name = tensor("op_3028_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_3028_cast_fp16 = slice_by_index(begin = var_3028_begin_0, end = var_3028_end_0, end_mask = var_3028_end_mask_0, squeeze_mask = var_3028_squeeze_mask_0, x = k_complex_15_cast_fp16)[name = tensor("op_3028_cast_fp16")]; + tensor freqs_15_to_fp16 = const()[name = tensor("freqs_15_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(183650304)))]; + tensor var_3034_cast_fp16 = mul(x = freqs_15_to_fp16, y = ts_47_cast_fp16)[name = tensor("op_3034_cast_fp16")]; + tensor rotr_15_cast_fp16 = cos(x = var_3034_cast_fp16)[name = tensor("rotr_15_cast_fp16")]; + tensor roti_15_cast_fp16 = sin(x = var_3034_cast_fp16)[name = tensor("roti_15_cast_fp16")]; + tensor var_3038_cast_fp16 = mul(x = var_3004_cast_fp16, y = rotr_15_cast_fp16)[name = tensor("op_3038_cast_fp16")]; + tensor var_3039_cast_fp16 = mul(x = var_3012_cast_fp16, y = roti_15_cast_fp16)[name = tensor("op_3039_cast_fp16")]; + tensor qor_29_cast_fp16 = sub(x = var_3038_cast_fp16, y = var_3039_cast_fp16)[name = tensor("qor_29_cast_fp16")]; + tensor var_3042_cast_fp16 = mul(x = var_3004_cast_fp16, y = roti_15_cast_fp16)[name = tensor("op_3042_cast_fp16")]; + tensor var_3043_cast_fp16 = mul(x = var_3012_cast_fp16, y = rotr_15_cast_fp16)[name = tensor("op_3043_cast_fp16")]; + tensor qoi_29_cast_fp16 = add(x = var_3042_cast_fp16, y = var_3043_cast_fp16)[name = tensor("qoi_29_cast_fp16")]; + tensor var_3046_cast_fp16 = mul(x = var_3020_cast_fp16, y = rotr_15_cast_fp16)[name = tensor("op_3046_cast_fp16")]; + tensor var_3047_cast_fp16 = mul(x = var_3028_cast_fp16, y = roti_15_cast_fp16)[name = tensor("op_3047_cast_fp16")]; + tensor kor_29_cast_fp16 = sub(x = var_3046_cast_fp16, y = var_3047_cast_fp16)[name = tensor("kor_29_cast_fp16")]; + tensor var_3050_cast_fp16 = mul(x = var_3020_cast_fp16, y = roti_15_cast_fp16)[name = tensor("op_3050_cast_fp16")]; + tensor var_3051_cast_fp16 = mul(x = var_3028_cast_fp16, y = rotr_15_cast_fp16)[name = tensor("op_3051_cast_fp16")]; + tensor koi_29_cast_fp16 = add(x = var_3050_cast_fp16, y = var_3051_cast_fp16)[name = tensor("koi_29_cast_fp16")]; + tensor qo_15_axis_0 = const()[name = tensor("qo_15_axis_0"), val = tensor(-1)]; + tensor qo_15_cast_fp16 = stack(axis = qo_15_axis_0, values = (qor_29_cast_fp16, qoi_29_cast_fp16))[name = tensor("qo_15_cast_fp16")]; + tensor ko_15_axis_0 = const()[name = tensor("ko_15_axis_0"), val = tensor(-1)]; + tensor ko_15_cast_fp16 = stack(axis = ko_15_axis_0, values = (kor_29_cast_fp16, koi_29_cast_fp16))[name = tensor("ko_15_cast_fp16")]; + tensor var_3084 = const()[name = tensor("op_3084"), val = tensor([1, 1, 16, 64])]; + tensor q_45_cast_fp16 = reshape(shape = var_3084, x = qo_15_cast_fp16)[name = tensor("q_45_cast_fp16")]; + tensor var_3090 = const()[name = tensor("op_3090"), val = tensor([1, 1, 16, 64])]; + tensor k_31_cast_fp16 = reshape(shape = var_3090, x = ko_15_cast_fp16)[name = tensor("k_31_cast_fp16")]; + tensor _inversed_3115_y_0_to_fp16 = const()[name = tensor("_inversed_3115_y_0_to_fp16"), val = tensor(0x1p-9)]; + tensor _inversed_3115_cast_fp16 = mul(x = ts_47_cast_fp16, y = _inversed_3115_y_0_to_fp16)[name = tensor("_inversed_3115_cast_fp16")]; + tensor var_3116_cast_fp16 = floor(x = _inversed_3115_cast_fp16)[name = tensor("op_3116_cast_fp16")]; + tensor var_3117_to_fp16 = const()[name = tensor("op_3117_to_fp16"), val = tensor(0x1p+9)]; + tensor var_3118_cast_fp16 = mul(x = var_3116_cast_fp16, y = var_3117_to_fp16)[name = tensor("op_3118_cast_fp16")]; + tensor write_indices_float_31_cast_fp16 = sub(x = ts_47_cast_fp16, y = var_3118_cast_fp16)[name = tensor("write_indices_float_31_cast_fp16")]; + tensor var_3125_dtype_0 = const()[name = tensor("op_3125_dtype_0"), val = tensor("int32")]; + tensor write_indices_15_reps_0 = const()[name = tensor("write_indices_15_reps_0"), val = tensor([1, 1, 16, 64])]; + tensor write_indices_float_31_cast_fp16_to_int32 = cast(dtype = var_3125_dtype_0, x = write_indices_float_31_cast_fp16)[name = tensor("cast_517")]; + tensor write_indices_15 = tile(reps = write_indices_15_reps_0, x = write_indices_float_31_cast_fp16_to_int32)[name = tensor("write_indices_15")]; + tensor var_3135_begin_0 = const()[name = tensor("op_3135_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_3135_end_0 = const()[name = tensor("op_3135_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor var_3135_end_mask_0 = const()[name = tensor("op_3135_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_3135_squeeze_mask_0 = const()[name = tensor("op_3135_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor cache7_to_fp16_dtype_0 = const()[name = tensor("cache7_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor cache7_to_fp16 = cast(dtype = cache7_to_fp16_dtype_0, x = cache7)[name = tensor("cast_516")]; + tensor var_3135_cast_fp16 = slice_by_index(begin = var_3135_begin_0, end = var_3135_end_0, end_mask = var_3135_end_mask_0, squeeze_mask = var_3135_squeeze_mask_0, x = cache7_to_fp16)[name = tensor("op_3135_cast_fp16")]; + tensor var_3137_axis_0 = const()[name = tensor("op_3137_axis_0"), val = tensor(1)]; + tensor var_3137_mode_0 = const()[name = tensor("op_3137_mode_0"), val = tensor("update")]; + tensor var_3137_validate_indices_0 = const()[name = tensor("op_3137_validate_indices_0"), val = tensor(false)]; + tensor var_3137_cast_fp16 = scatter_along_axis(axis = var_3137_axis_0, data = var_3135_cast_fp16, indices = write_indices_15, mode = var_3137_mode_0, updates = k_31_cast_fp16, validate_indices = var_3137_validate_indices_0)[name = tensor("op_3137_cast_fp16")]; + tensor concat_51 = const()[name = tensor("concat_51"), val = tensor([0, 0, 0, 0, 0])]; + tensor concat_52 = const()[name = tensor("concat_52"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_15_internal_tensor_assign_1_stride_0 = const()[name = tensor("new_cache_15_internal_tensor_assign_1_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_15_internal_tensor_assign_1_begin_mask_0 = const()[name = tensor("new_cache_15_internal_tensor_assign_1_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_15_internal_tensor_assign_1_end_mask_0 = const()[name = tensor("new_cache_15_internal_tensor_assign_1_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_15_internal_tensor_assign_1_squeeze_mask_0 = const()[name = tensor("new_cache_15_internal_tensor_assign_1_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_62 = const()[name = tensor("shape_62"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_14 = const()[name = tensor("reduce_prod_14"), val = tensor(1048576)]; + tensor range_1d_14_start_0 = const()[name = tensor("range_1d_14_start_0"), val = tensor(0)]; + tensor range_1d_14_step_0 = const()[name = tensor("range_1d_14_step_0"), val = tensor(1)]; + tensor range_1d_14 = range_1d(end = reduce_prod_14, start = range_1d_14_start_0, step = range_1d_14_step_0)[name = tensor("range_1d_14")]; + tensor reshape_70 = reshape(shape = shape_62, x = range_1d_14)[name = tensor("reshape_70")]; + tensor slice_by_index_14 = slice_by_index(begin = concat_51, begin_mask = new_cache_15_internal_tensor_assign_1_begin_mask_0, end = concat_52, end_mask = new_cache_15_internal_tensor_assign_1_end_mask_0, squeeze_mask = new_cache_15_internal_tensor_assign_1_squeeze_mask_0, stride = new_cache_15_internal_tensor_assign_1_stride_0, x = reshape_70)[name = tensor("slice_by_index_14")]; + tensor reshape_71_shape_0 = const()[name = tensor("reshape_71_shape_0"), val = tensor([-1])]; + tensor reshape_71 = reshape(shape = reshape_71_shape_0, x = slice_by_index_14)[name = tensor("reshape_71")]; + tensor reshape_72_shape_0 = const()[name = tensor("reshape_72_shape_0"), val = tensor([-1])]; + tensor reshape_72_cast_fp16 = reshape(shape = reshape_72_shape_0, x = var_3137_cast_fp16)[name = tensor("reshape_72_cast_fp16")]; + tensor reshape_73_shape_0 = const()[name = tensor("reshape_73_shape_0"), val = tensor([-1])]; + tensor reshape_73_cast_fp16 = reshape(shape = reshape_73_shape_0, x = cache7_to_fp16)[name = tensor("reshape_73_cast_fp16")]; + tensor scatter_14_mode_0 = const()[name = tensor("scatter_14_mode_0"), val = tensor("update")]; + tensor scatter_14_axis_0 = const()[name = tensor("scatter_14_axis_0"), val = tensor(0)]; + tensor scatter_14_validate_indices_0 = const()[name = tensor("scatter_14_validate_indices_0"), val = tensor(false)]; + tensor scatter_14_cast_fp16 = scatter(axis = scatter_14_axis_0, data = reshape_73_cast_fp16, indices = reshape_71, mode = scatter_14_mode_0, updates = reshape_72_cast_fp16, validate_indices = scatter_14_validate_indices_0)[name = tensor("scatter_14_cast_fp16")]; + tensor reshape_74_cast_fp16 = reshape(shape = shape_62, x = scatter_14_cast_fp16)[name = tensor("reshape_74_cast_fp16")]; + tensor var_3145_begin_0 = const()[name = tensor("op_3145_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor var_3145_end_0 = const()[name = tensor("op_3145_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor var_3145_end_mask_0 = const()[name = tensor("op_3145_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_3145_squeeze_mask_0 = const()[name = tensor("op_3145_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor var_3145_cast_fp16 = slice_by_index(begin = var_3145_begin_0, end = var_3145_end_0, end_mask = var_3145_end_mask_0, squeeze_mask = var_3145_squeeze_mask_0, x = reshape_74_cast_fp16)[name = tensor("op_3145_cast_fp16")]; + tensor var_3147_axis_0 = const()[name = tensor("op_3147_axis_0"), val = tensor(1)]; + tensor var_3147_mode_0 = const()[name = tensor("op_3147_mode_0"), val = tensor("update")]; + tensor var_3147_validate_indices_0 = const()[name = tensor("op_3147_validate_indices_0"), val = tensor(false)]; + tensor var_3147_cast_fp16 = scatter_along_axis(axis = var_3147_axis_0, data = var_3145_cast_fp16, indices = write_indices_15, mode = var_3147_mode_0, updates = v_15_cast_fp16, validate_indices = var_3147_validate_indices_0)[name = tensor("op_3147_cast_fp16")]; + tensor concat_53 = const()[name = tensor("concat_53"), val = tensor([1, 0, 0, 0, 0])]; + tensor concat_54 = const()[name = tensor("concat_54"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_15_internal_tensor_assign_2_stride_0 = const()[name = tensor("new_cache_15_internal_tensor_assign_2_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_15_internal_tensor_assign_2_begin_mask_0 = const()[name = tensor("new_cache_15_internal_tensor_assign_2_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_15_internal_tensor_assign_2_end_mask_0 = const()[name = tensor("new_cache_15_internal_tensor_assign_2_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_15_internal_tensor_assign_2_squeeze_mask_0 = const()[name = tensor("new_cache_15_internal_tensor_assign_2_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_63 = const()[name = tensor("shape_63"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_15 = const()[name = tensor("reduce_prod_15"), val = tensor(1048576)]; + tensor range_1d_15_start_0 = const()[name = tensor("range_1d_15_start_0"), val = tensor(0)]; + tensor range_1d_15_step_0 = const()[name = tensor("range_1d_15_step_0"), val = tensor(1)]; + tensor range_1d_15 = range_1d(end = reduce_prod_15, start = range_1d_15_start_0, step = range_1d_15_step_0)[name = tensor("range_1d_15")]; + tensor reshape_75 = reshape(shape = shape_63, x = range_1d_15)[name = tensor("reshape_75")]; + tensor slice_by_index_15 = slice_by_index(begin = concat_53, begin_mask = new_cache_15_internal_tensor_assign_2_begin_mask_0, end = concat_54, end_mask = new_cache_15_internal_tensor_assign_2_end_mask_0, squeeze_mask = new_cache_15_internal_tensor_assign_2_squeeze_mask_0, stride = new_cache_15_internal_tensor_assign_2_stride_0, x = reshape_75)[name = tensor("slice_by_index_15")]; + tensor reshape_76_shape_0 = const()[name = tensor("reshape_76_shape_0"), val = tensor([-1])]; + tensor reshape_76 = reshape(shape = reshape_76_shape_0, x = slice_by_index_15)[name = tensor("reshape_76")]; + tensor reshape_77_shape_0 = const()[name = tensor("reshape_77_shape_0"), val = tensor([-1])]; + tensor reshape_77_cast_fp16 = reshape(shape = reshape_77_shape_0, x = var_3147_cast_fp16)[name = tensor("reshape_77_cast_fp16")]; + tensor reshape_78_shape_0 = const()[name = tensor("reshape_78_shape_0"), val = tensor([-1])]; + tensor reshape_78_cast_fp16 = reshape(shape = reshape_78_shape_0, x = reshape_74_cast_fp16)[name = tensor("reshape_78_cast_fp16")]; + tensor scatter_15_mode_0 = const()[name = tensor("scatter_15_mode_0"), val = tensor("update")]; + tensor scatter_15_axis_0 = const()[name = tensor("scatter_15_axis_0"), val = tensor(0)]; + tensor scatter_15_validate_indices_0 = const()[name = tensor("scatter_15_validate_indices_0"), val = tensor(false)]; + tensor scatter_15_cast_fp16 = scatter(axis = scatter_15_axis_0, data = reshape_78_cast_fp16, indices = reshape_76, mode = scatter_15_mode_0, updates = reshape_77_cast_fp16, validate_indices = scatter_15_validate_indices_0)[name = tensor("scatter_15_cast_fp16")]; + tensor reshape_79_cast_fp16 = reshape(shape = shape_63, x = scatter_15_cast_fp16)[name = tensor("reshape_79_cast_fp16")]; + tensor reshape_79_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("reshape_79_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor keys_43_begin_0 = const()[name = tensor("keys_43_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor keys_43_end_0 = const()[name = tensor("keys_43_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor keys_43_end_mask_0 = const()[name = tensor("keys_43_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor keys_43_squeeze_mask_0 = const()[name = tensor("keys_43_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor keys_43_cast_fp16 = slice_by_index(begin = keys_43_begin_0, end = keys_43_end_0, end_mask = keys_43_end_mask_0, squeeze_mask = keys_43_squeeze_mask_0, x = reshape_79_cast_fp16)[name = tensor("keys_43_cast_fp16")]; + tensor values_43_begin_0 = const()[name = tensor("values_43_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor values_43_end_0 = const()[name = tensor("values_43_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor values_43_end_mask_0 = const()[name = tensor("values_43_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor values_43_squeeze_mask_0 = const()[name = tensor("values_43_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor values_43_cast_fp16 = slice_by_index(begin = values_43_begin_0, end = values_43_end_0, end_mask = values_43_end_mask_0, squeeze_mask = values_43_squeeze_mask_0, x = reshape_79_cast_fp16)[name = tensor("values_43_cast_fp16")]; + tensor var_3159_cast_fp16 = not_equal(x = keys_43_cast_fp16, y = keys_43_cast_fp16)[name = tensor("op_3159_cast_fp16")]; + tensor keys_45_cast_fp16 = select(a = var_463_to_fp16, b = keys_43_cast_fp16, cond = var_3159_cast_fp16)[name = tensor("keys_45_cast_fp16")]; + tensor var_3167_cast_fp16 = not_equal(x = values_43_cast_fp16, y = values_43_cast_fp16)[name = tensor("op_3167_cast_fp16")]; + tensor values_45_cast_fp16 = select(a = var_463_to_fp16, b = values_43_cast_fp16, cond = var_3167_cast_fp16)[name = tensor("values_45_cast_fp16")]; + tensor var_3191 = const()[name = tensor("op_3191"), val = tensor([0, 2, 1, 3])]; + tensor var_3207 = const()[name = tensor("op_3207"), val = tensor([1, 1, 1])]; + tensor var_3208_cast_fp16 = reshape(shape = var_3207, x = position7_to_fp16)[name = tensor("op_3208_cast_fp16")]; + tensor var_3228_to_fp16 = const()[name = tensor("op_3228_to_fp16"), val = tensor(0x1p+0)]; + tensor valid_len_15_cast_fp16 = add(x = var_3208_cast_fp16, y = var_3228_to_fp16)[name = tensor("valid_len_15_cast_fp16")]; + tensor valid_mask_15_cast_fp16 = less(x = k_positions_1_promoted_to_fp16, y = valid_len_15_cast_fp16)[name = tensor("valid_mask_15_cast_fp16")]; + tensor causal_mask_15_cast_fp16 = less_equal(x = k_positions_1_promoted_to_fp16, y = var_3208_cast_fp16)[name = tensor("causal_mask_15_cast_fp16")]; + tensor attn_mask_29 = logical_and(x = valid_mask_15_cast_fp16, y = causal_mask_15_cast_fp16)[name = tensor("attn_mask_29")]; + tensor attn_mask_31_axes_0 = const()[name = tensor("attn_mask_31_axes_0"), val = tensor([1])]; + tensor attn_mask_31 = expand_dims(axes = attn_mask_31_axes_0, x = attn_mask_29)[name = tensor("attn_mask_31")]; + tensor var_3246_transpose_x_0 = const()[name = tensor("op_3246_transpose_x_0"), val = tensor(false)]; + tensor var_3246_transpose_y_0 = const()[name = tensor("op_3246_transpose_y_0"), val = tensor(false)]; + tensor transpose_86_perm_0 = const()[name = tensor("transpose_86_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_87_perm_0 = const()[name = tensor("transpose_87_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_87 = transpose(perm = transpose_87_perm_0, x = keys_45_cast_fp16)[name = tensor("transpose_185")]; + tensor transpose_86 = transpose(perm = transpose_86_perm_0, x = q_45_cast_fp16)[name = tensor("transpose_186")]; + tensor var_3246_cast_fp16 = matmul(transpose_x = var_3246_transpose_x_0, transpose_y = var_3246_transpose_y_0, x = transpose_86, y = transpose_87)[name = tensor("op_3246_cast_fp16")]; + tensor var_3240_to_fp16 = const()[name = tensor("op_3240_to_fp16"), val = tensor([0x1p-3])]; + tensor attn_weights_43_cast_fp16 = mul(x = var_3246_cast_fp16, y = var_3240_to_fp16)[name = tensor("attn_weights_43_cast_fp16")]; + tensor var_3248 = logical_not(x = attn_mask_31)[name = tensor("op_3248")]; + tensor var_3249_to_fp16 = const()[name = tensor("op_3249_to_fp16"), val = tensor(-inf)]; + tensor attn_weights_45_cast_fp16 = select(a = var_3249_to_fp16, b = attn_weights_43_cast_fp16, cond = var_3248)[name = tensor("attn_weights_45_cast_fp16")]; + tensor var_3251 = const()[name = tensor("op_3251"), val = tensor(-1)]; + tensor attn_weights_47_cast_fp16 = softmax(axis = var_3251, x = attn_weights_45_cast_fp16)[name = tensor("attn_weights_47_cast_fp16")]; + tensor attn_output_15_transpose_x_0 = const()[name = tensor("attn_output_15_transpose_x_0"), val = tensor(false)]; + tensor attn_output_15_transpose_y_0 = const()[name = tensor("attn_output_15_transpose_y_0"), val = tensor(false)]; + tensor values_47_cast_fp16 = transpose(perm = var_3191, x = values_45_cast_fp16)[name = tensor("transpose_187")]; + tensor attn_output_15_cast_fp16 = matmul(transpose_x = attn_output_15_transpose_x_0, transpose_y = attn_output_15_transpose_y_0, x = attn_weights_47_cast_fp16, y = values_47_cast_fp16)[name = tensor("attn_output_15_cast_fp16")]; + tensor var_3259 = const()[name = tensor("op_3259"), val = tensor([0, 2, 1, 3])]; + tensor var_3264 = const()[name = tensor("op_3264"), val = tensor([1, 1, 1024])]; + tensor var_3260_cast_fp16 = transpose(perm = var_3259, x = attn_output_15_cast_fp16)[name = tensor("transpose_184")]; + tensor input_91_cast_fp16 = reshape(shape = var_3264, x = var_3260_cast_fp16)[name = tensor("input_91_cast_fp16")]; + tensor attn7_out_proj_weight_to_fp16 = const()[name = tensor("attn7_out_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(183650432)))]; + tensor linear_30_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = attn7_out_proj_weight_to_fp16, x = input_91_cast_fp16)[name = tensor("linear_30_cast_fp16")]; + tensor var_3270_to_fp16 = const()[name = tensor("op_3270_to_fp16"), val = tensor(0x1p+0)]; + tensor var_3271_cast_fp16 = add(x = position7_to_fp16, y = var_3270_to_fp16)[name = tensor("op_3271_cast_fp16")]; + tensor var_3271_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_3271_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor input_93_cast_fp16 = add(x = input_87_cast_fp16, y = linear_30_cast_fp16)[name = tensor("input_93_cast_fp16")]; + tensor input_95_axes_0 = const()[name = tensor("input_95_axes_0"), val = tensor([-1])]; + tensor norm7_2_weight_to_fp16 = const()[name = tensor("norm7_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(185747648)))]; + tensor norm7_2_bias_to_fp16 = const()[name = tensor("norm7_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(185749760)))]; + tensor var_3275_to_fp16 = const()[name = tensor("op_3275_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_95_cast_fp16 = layer_norm(axes = input_95_axes_0, beta = norm7_2_bias_to_fp16, epsilon = var_3275_to_fp16, gamma = norm7_2_weight_to_fp16, x = input_93_cast_fp16)[name = tensor("input_95_cast_fp16")]; + tensor linear7_1_weight_to_fp16 = const()[name = tensor("linear7_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(185751872)))]; + tensor linear_31_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = linear7_1_weight_to_fp16, x = input_95_cast_fp16)[name = tensor("linear_31_cast_fp16")]; + tensor input_97_mode_0 = const()[name = tensor("input_97_mode_0"), val = tensor("EXACT")]; + tensor input_97_cast_fp16 = gelu(mode = input_97_mode_0, x = linear_31_cast_fp16)[name = tensor("input_97_cast_fp16")]; + tensor linear7_2_weight_to_fp16 = const()[name = tensor("linear7_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(194140544)))]; + tensor linear_32_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = linear7_2_weight_to_fp16, x = input_97_cast_fp16)[name = tensor("linear_32_cast_fp16")]; + tensor input_99_cast_fp16 = add(x = input_93_cast_fp16, y = linear_32_cast_fp16)[name = tensor("input_99_cast_fp16")]; + tensor input_101_axes_0 = const()[name = tensor("input_101_axes_0"), val = tensor([-1])]; + tensor norm8_1_weight_to_fp16 = const()[name = tensor("norm8_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(202529216)))]; + tensor norm8_1_bias_to_fp16 = const()[name = tensor("norm8_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(202531328)))]; + tensor var_3292_to_fp16 = const()[name = tensor("op_3292_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_101_cast_fp16 = layer_norm(axes = input_101_axes_0, beta = norm8_1_bias_to_fp16, epsilon = var_3292_to_fp16, gamma = norm8_1_weight_to_fp16, x = input_99_cast_fp16)[name = tensor("input_101_cast_fp16")]; + tensor attn8_in_proj_weight_to_fp16 = const()[name = tensor("attn8_in_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(202533440)))]; + tensor linear_33_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = attn8_in_proj_weight_to_fp16, x = input_101_cast_fp16)[name = tensor("linear_33_cast_fp16")]; + tensor var_3306 = const()[name = tensor("op_3306"), val = tensor([1, 1, 3, 16, 64])]; + tensor qkv_17_cast_fp16 = reshape(shape = var_3306, x = linear_33_cast_fp16)[name = tensor("qkv_17_cast_fp16")]; + tensor q_49_begin_0 = const()[name = tensor("q_49_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor q_49_end_0 = const()[name = tensor("q_49_end_0"), val = tensor([1, 1, 1, 16, 64])]; + tensor q_49_end_mask_0 = const()[name = tensor("q_49_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor q_49_squeeze_mask_0 = const()[name = tensor("q_49_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor q_49_cast_fp16 = slice_by_index(begin = q_49_begin_0, end = q_49_end_0, end_mask = q_49_end_mask_0, squeeze_mask = q_49_squeeze_mask_0, x = qkv_17_cast_fp16)[name = tensor("q_49_cast_fp16")]; + tensor k_33_begin_0 = const()[name = tensor("k_33_begin_0"), val = tensor([0, 0, 1, 0, 0])]; + tensor k_33_end_0 = const()[name = tensor("k_33_end_0"), val = tensor([1, 1, 2, 16, 64])]; + tensor k_33_end_mask_0 = const()[name = tensor("k_33_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor k_33_squeeze_mask_0 = const()[name = tensor("k_33_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor k_33_cast_fp16 = slice_by_index(begin = k_33_begin_0, end = k_33_end_0, end_mask = k_33_end_mask_0, squeeze_mask = k_33_squeeze_mask_0, x = qkv_17_cast_fp16)[name = tensor("k_33_cast_fp16")]; + tensor v_17_begin_0 = const()[name = tensor("v_17_begin_0"), val = tensor([0, 0, 2, 0, 0])]; + tensor v_17_end_0 = const()[name = tensor("v_17_end_0"), val = tensor([1, 1, 3, 16, 64])]; + tensor v_17_end_mask_0 = const()[name = tensor("v_17_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor v_17_squeeze_mask_0 = const()[name = tensor("v_17_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor v_17_cast_fp16 = slice_by_index(begin = v_17_begin_0, end = v_17_end_0, end_mask = v_17_end_mask_0, squeeze_mask = v_17_squeeze_mask_0, x = qkv_17_cast_fp16)[name = tensor("v_17_cast_fp16")]; + tensor var_3372 = const()[name = tensor("op_3372"), val = tensor([1, 1, 1, 1])]; + tensor position8_to_fp16_dtype_0 = const()[name = tensor("position8_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor position8_to_fp16 = cast(dtype = position8_to_fp16_dtype_0, x = position8)[name = tensor("cast_513")]; + tensor ts_53_cast_fp16 = reshape(shape = var_3372, x = position8_to_fp16)[name = tensor("ts_53_cast_fp16")]; + tensor var_3379 = const()[name = tensor("op_3379"), val = tensor([1, 1, 16, 32, 2])]; + tensor q_complex_17_cast_fp16 = reshape(shape = var_3379, x = q_49_cast_fp16)[name = tensor("q_complex_17_cast_fp16")]; + tensor var_3386 = const()[name = tensor("op_3386"), val = tensor([1, 1, 16, 32, 2])]; + tensor k_complex_17_cast_fp16 = reshape(shape = var_3386, x = k_33_cast_fp16)[name = tensor("k_complex_17_cast_fp16")]; + tensor var_3390_begin_0 = const()[name = tensor("op_3390_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_3390_end_0 = const()[name = tensor("op_3390_end_0"), val = tensor([1, 1, 16, 32, 1])]; + tensor var_3390_end_mask_0 = const()[name = tensor("op_3390_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_3390_squeeze_mask_0 = const()[name = tensor("op_3390_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_3390_cast_fp16 = slice_by_index(begin = var_3390_begin_0, end = var_3390_end_0, end_mask = var_3390_end_mask_0, squeeze_mask = var_3390_squeeze_mask_0, x = q_complex_17_cast_fp16)[name = tensor("op_3390_cast_fp16")]; + tensor var_3398_begin_0 = const()[name = tensor("op_3398_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_3398_end_0 = const()[name = tensor("op_3398_end_0"), val = tensor([1, 1, 16, 32, 2])]; + tensor var_3398_end_mask_0 = const()[name = tensor("op_3398_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_3398_squeeze_mask_0 = const()[name = tensor("op_3398_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_3398_cast_fp16 = slice_by_index(begin = var_3398_begin_0, end = var_3398_end_0, end_mask = var_3398_end_mask_0, squeeze_mask = var_3398_squeeze_mask_0, x = q_complex_17_cast_fp16)[name = tensor("op_3398_cast_fp16")]; + tensor var_3406_begin_0 = const()[name = tensor("op_3406_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_3406_end_0 = const()[name = tensor("op_3406_end_0"), val = tensor([1, 1, 16, 32, 1])]; + tensor var_3406_end_mask_0 = const()[name = tensor("op_3406_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_3406_squeeze_mask_0 = const()[name = tensor("op_3406_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_3406_cast_fp16 = slice_by_index(begin = var_3406_begin_0, end = var_3406_end_0, end_mask = var_3406_end_mask_0, squeeze_mask = var_3406_squeeze_mask_0, x = k_complex_17_cast_fp16)[name = tensor("op_3406_cast_fp16")]; + tensor var_3414_begin_0 = const()[name = tensor("op_3414_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_3414_end_0 = const()[name = tensor("op_3414_end_0"), val = tensor([1, 1, 16, 32, 2])]; + tensor var_3414_end_mask_0 = const()[name = tensor("op_3414_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_3414_squeeze_mask_0 = const()[name = tensor("op_3414_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_3414_cast_fp16 = slice_by_index(begin = var_3414_begin_0, end = var_3414_end_0, end_mask = var_3414_end_mask_0, squeeze_mask = var_3414_squeeze_mask_0, x = k_complex_17_cast_fp16)[name = tensor("op_3414_cast_fp16")]; + tensor freqs_17_to_fp16 = const()[name = tensor("freqs_17_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(208824960)))]; + tensor var_3420_cast_fp16 = mul(x = freqs_17_to_fp16, y = ts_53_cast_fp16)[name = tensor("op_3420_cast_fp16")]; + tensor rotr_17_cast_fp16 = cos(x = var_3420_cast_fp16)[name = tensor("rotr_17_cast_fp16")]; + tensor roti_17_cast_fp16 = sin(x = var_3420_cast_fp16)[name = tensor("roti_17_cast_fp16")]; + tensor var_3424_cast_fp16 = mul(x = var_3390_cast_fp16, y = rotr_17_cast_fp16)[name = tensor("op_3424_cast_fp16")]; + tensor var_3425_cast_fp16 = mul(x = var_3398_cast_fp16, y = roti_17_cast_fp16)[name = tensor("op_3425_cast_fp16")]; + tensor qor_33_cast_fp16 = sub(x = var_3424_cast_fp16, y = var_3425_cast_fp16)[name = tensor("qor_33_cast_fp16")]; + tensor var_3428_cast_fp16 = mul(x = var_3390_cast_fp16, y = roti_17_cast_fp16)[name = tensor("op_3428_cast_fp16")]; + tensor var_3429_cast_fp16 = mul(x = var_3398_cast_fp16, y = rotr_17_cast_fp16)[name = tensor("op_3429_cast_fp16")]; + tensor qoi_33_cast_fp16 = add(x = var_3428_cast_fp16, y = var_3429_cast_fp16)[name = tensor("qoi_33_cast_fp16")]; + tensor var_3432_cast_fp16 = mul(x = var_3406_cast_fp16, y = rotr_17_cast_fp16)[name = tensor("op_3432_cast_fp16")]; + tensor var_3433_cast_fp16 = mul(x = var_3414_cast_fp16, y = roti_17_cast_fp16)[name = tensor("op_3433_cast_fp16")]; + tensor kor_33_cast_fp16 = sub(x = var_3432_cast_fp16, y = var_3433_cast_fp16)[name = tensor("kor_33_cast_fp16")]; + tensor var_3436_cast_fp16 = mul(x = var_3406_cast_fp16, y = roti_17_cast_fp16)[name = tensor("op_3436_cast_fp16")]; + tensor var_3437_cast_fp16 = mul(x = var_3414_cast_fp16, y = rotr_17_cast_fp16)[name = tensor("op_3437_cast_fp16")]; + tensor koi_33_cast_fp16 = add(x = var_3436_cast_fp16, y = var_3437_cast_fp16)[name = tensor("koi_33_cast_fp16")]; + tensor qo_17_axis_0 = const()[name = tensor("qo_17_axis_0"), val = tensor(-1)]; + tensor qo_17_cast_fp16 = stack(axis = qo_17_axis_0, values = (qor_33_cast_fp16, qoi_33_cast_fp16))[name = tensor("qo_17_cast_fp16")]; + tensor ko_17_axis_0 = const()[name = tensor("ko_17_axis_0"), val = tensor(-1)]; + tensor ko_17_cast_fp16 = stack(axis = ko_17_axis_0, values = (kor_33_cast_fp16, koi_33_cast_fp16))[name = tensor("ko_17_cast_fp16")]; + tensor var_3470 = const()[name = tensor("op_3470"), val = tensor([1, 1, 16, 64])]; + tensor q_51_cast_fp16 = reshape(shape = var_3470, x = qo_17_cast_fp16)[name = tensor("q_51_cast_fp16")]; + tensor var_3476 = const()[name = tensor("op_3476"), val = tensor([1, 1, 16, 64])]; + tensor k_35_cast_fp16 = reshape(shape = var_3476, x = ko_17_cast_fp16)[name = tensor("k_35_cast_fp16")]; + tensor _inversed_3501_y_0_to_fp16 = const()[name = tensor("_inversed_3501_y_0_to_fp16"), val = tensor(0x1p-9)]; + tensor _inversed_3501_cast_fp16 = mul(x = ts_53_cast_fp16, y = _inversed_3501_y_0_to_fp16)[name = tensor("_inversed_3501_cast_fp16")]; + tensor var_3502_cast_fp16 = floor(x = _inversed_3501_cast_fp16)[name = tensor("op_3502_cast_fp16")]; + tensor var_3503_to_fp16 = const()[name = tensor("op_3503_to_fp16"), val = tensor(0x1p+9)]; + tensor var_3504_cast_fp16 = mul(x = var_3502_cast_fp16, y = var_3503_to_fp16)[name = tensor("op_3504_cast_fp16")]; + tensor write_indices_float_35_cast_fp16 = sub(x = ts_53_cast_fp16, y = var_3504_cast_fp16)[name = tensor("write_indices_float_35_cast_fp16")]; + tensor var_3511_dtype_0 = const()[name = tensor("op_3511_dtype_0"), val = tensor("int32")]; + tensor write_indices_17_reps_0 = const()[name = tensor("write_indices_17_reps_0"), val = tensor([1, 1, 16, 64])]; + tensor write_indices_float_35_cast_fp16_to_int32 = cast(dtype = var_3511_dtype_0, x = write_indices_float_35_cast_fp16)[name = tensor("cast_512")]; + tensor write_indices_17 = tile(reps = write_indices_17_reps_0, x = write_indices_float_35_cast_fp16_to_int32)[name = tensor("write_indices_17")]; + tensor var_3521_begin_0 = const()[name = tensor("op_3521_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_3521_end_0 = const()[name = tensor("op_3521_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor var_3521_end_mask_0 = const()[name = tensor("op_3521_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_3521_squeeze_mask_0 = const()[name = tensor("op_3521_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor cache8_to_fp16_dtype_0 = const()[name = tensor("cache8_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor cache8_to_fp16 = cast(dtype = cache8_to_fp16_dtype_0, x = cache8)[name = tensor("cast_511")]; + tensor var_3521_cast_fp16 = slice_by_index(begin = var_3521_begin_0, end = var_3521_end_0, end_mask = var_3521_end_mask_0, squeeze_mask = var_3521_squeeze_mask_0, x = cache8_to_fp16)[name = tensor("op_3521_cast_fp16")]; + tensor var_3523_axis_0 = const()[name = tensor("op_3523_axis_0"), val = tensor(1)]; + tensor var_3523_mode_0 = const()[name = tensor("op_3523_mode_0"), val = tensor("update")]; + tensor var_3523_validate_indices_0 = const()[name = tensor("op_3523_validate_indices_0"), val = tensor(false)]; + tensor var_3523_cast_fp16 = scatter_along_axis(axis = var_3523_axis_0, data = var_3521_cast_fp16, indices = write_indices_17, mode = var_3523_mode_0, updates = k_35_cast_fp16, validate_indices = var_3523_validate_indices_0)[name = tensor("op_3523_cast_fp16")]; + tensor concat_58 = const()[name = tensor("concat_58"), val = tensor([0, 0, 0, 0, 0])]; + tensor concat_59 = const()[name = tensor("concat_59"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_17_internal_tensor_assign_1_stride_0 = const()[name = tensor("new_cache_17_internal_tensor_assign_1_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_17_internal_tensor_assign_1_begin_mask_0 = const()[name = tensor("new_cache_17_internal_tensor_assign_1_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_17_internal_tensor_assign_1_end_mask_0 = const()[name = tensor("new_cache_17_internal_tensor_assign_1_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_17_internal_tensor_assign_1_squeeze_mask_0 = const()[name = tensor("new_cache_17_internal_tensor_assign_1_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_64 = const()[name = tensor("shape_64"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_16 = const()[name = tensor("reduce_prod_16"), val = tensor(1048576)]; + tensor range_1d_16_start_0 = const()[name = tensor("range_1d_16_start_0"), val = tensor(0)]; + tensor range_1d_16_step_0 = const()[name = tensor("range_1d_16_step_0"), val = tensor(1)]; + tensor range_1d_16 = range_1d(end = reduce_prod_16, start = range_1d_16_start_0, step = range_1d_16_step_0)[name = tensor("range_1d_16")]; + tensor reshape_80 = reshape(shape = shape_64, x = range_1d_16)[name = tensor("reshape_80")]; + tensor slice_by_index_16 = slice_by_index(begin = concat_58, begin_mask = new_cache_17_internal_tensor_assign_1_begin_mask_0, end = concat_59, end_mask = new_cache_17_internal_tensor_assign_1_end_mask_0, squeeze_mask = new_cache_17_internal_tensor_assign_1_squeeze_mask_0, stride = new_cache_17_internal_tensor_assign_1_stride_0, x = reshape_80)[name = tensor("slice_by_index_16")]; + tensor reshape_81_shape_0 = const()[name = tensor("reshape_81_shape_0"), val = tensor([-1])]; + tensor reshape_81 = reshape(shape = reshape_81_shape_0, x = slice_by_index_16)[name = tensor("reshape_81")]; + tensor reshape_82_shape_0 = const()[name = tensor("reshape_82_shape_0"), val = tensor([-1])]; + tensor reshape_82_cast_fp16 = reshape(shape = reshape_82_shape_0, x = var_3523_cast_fp16)[name = tensor("reshape_82_cast_fp16")]; + tensor reshape_83_shape_0 = const()[name = tensor("reshape_83_shape_0"), val = tensor([-1])]; + tensor reshape_83_cast_fp16 = reshape(shape = reshape_83_shape_0, x = cache8_to_fp16)[name = tensor("reshape_83_cast_fp16")]; + tensor scatter_16_mode_0 = const()[name = tensor("scatter_16_mode_0"), val = tensor("update")]; + tensor scatter_16_axis_0 = const()[name = tensor("scatter_16_axis_0"), val = tensor(0)]; + tensor scatter_16_validate_indices_0 = const()[name = tensor("scatter_16_validate_indices_0"), val = tensor(false)]; + tensor scatter_16_cast_fp16 = scatter(axis = scatter_16_axis_0, data = reshape_83_cast_fp16, indices = reshape_81, mode = scatter_16_mode_0, updates = reshape_82_cast_fp16, validate_indices = scatter_16_validate_indices_0)[name = tensor("scatter_16_cast_fp16")]; + tensor reshape_84_cast_fp16 = reshape(shape = shape_64, x = scatter_16_cast_fp16)[name = tensor("reshape_84_cast_fp16")]; + tensor var_3531_begin_0 = const()[name = tensor("op_3531_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor var_3531_end_0 = const()[name = tensor("op_3531_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor var_3531_end_mask_0 = const()[name = tensor("op_3531_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_3531_squeeze_mask_0 = const()[name = tensor("op_3531_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor var_3531_cast_fp16 = slice_by_index(begin = var_3531_begin_0, end = var_3531_end_0, end_mask = var_3531_end_mask_0, squeeze_mask = var_3531_squeeze_mask_0, x = reshape_84_cast_fp16)[name = tensor("op_3531_cast_fp16")]; + tensor var_3533_axis_0 = const()[name = tensor("op_3533_axis_0"), val = tensor(1)]; + tensor var_3533_mode_0 = const()[name = tensor("op_3533_mode_0"), val = tensor("update")]; + tensor var_3533_validate_indices_0 = const()[name = tensor("op_3533_validate_indices_0"), val = tensor(false)]; + tensor var_3533_cast_fp16 = scatter_along_axis(axis = var_3533_axis_0, data = var_3531_cast_fp16, indices = write_indices_17, mode = var_3533_mode_0, updates = v_17_cast_fp16, validate_indices = var_3533_validate_indices_0)[name = tensor("op_3533_cast_fp16")]; + tensor concat_60 = const()[name = tensor("concat_60"), val = tensor([1, 0, 0, 0, 0])]; + tensor concat_61 = const()[name = tensor("concat_61"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_17_internal_tensor_assign_2_stride_0 = const()[name = tensor("new_cache_17_internal_tensor_assign_2_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_17_internal_tensor_assign_2_begin_mask_0 = const()[name = tensor("new_cache_17_internal_tensor_assign_2_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_17_internal_tensor_assign_2_end_mask_0 = const()[name = tensor("new_cache_17_internal_tensor_assign_2_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_17_internal_tensor_assign_2_squeeze_mask_0 = const()[name = tensor("new_cache_17_internal_tensor_assign_2_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_65 = const()[name = tensor("shape_65"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_17 = const()[name = tensor("reduce_prod_17"), val = tensor(1048576)]; + tensor range_1d_17_start_0 = const()[name = tensor("range_1d_17_start_0"), val = tensor(0)]; + tensor range_1d_17_step_0 = const()[name = tensor("range_1d_17_step_0"), val = tensor(1)]; + tensor range_1d_17 = range_1d(end = reduce_prod_17, start = range_1d_17_start_0, step = range_1d_17_step_0)[name = tensor("range_1d_17")]; + tensor reshape_85 = reshape(shape = shape_65, x = range_1d_17)[name = tensor("reshape_85")]; + tensor slice_by_index_17 = slice_by_index(begin = concat_60, begin_mask = new_cache_17_internal_tensor_assign_2_begin_mask_0, end = concat_61, end_mask = new_cache_17_internal_tensor_assign_2_end_mask_0, squeeze_mask = new_cache_17_internal_tensor_assign_2_squeeze_mask_0, stride = new_cache_17_internal_tensor_assign_2_stride_0, x = reshape_85)[name = tensor("slice_by_index_17")]; + tensor reshape_86_shape_0 = const()[name = tensor("reshape_86_shape_0"), val = tensor([-1])]; + tensor reshape_86 = reshape(shape = reshape_86_shape_0, x = slice_by_index_17)[name = tensor("reshape_86")]; + tensor reshape_87_shape_0 = const()[name = tensor("reshape_87_shape_0"), val = tensor([-1])]; + tensor reshape_87_cast_fp16 = reshape(shape = reshape_87_shape_0, x = var_3533_cast_fp16)[name = tensor("reshape_87_cast_fp16")]; + tensor reshape_88_shape_0 = const()[name = tensor("reshape_88_shape_0"), val = tensor([-1])]; + tensor reshape_88_cast_fp16 = reshape(shape = reshape_88_shape_0, x = reshape_84_cast_fp16)[name = tensor("reshape_88_cast_fp16")]; + tensor scatter_17_mode_0 = const()[name = tensor("scatter_17_mode_0"), val = tensor("update")]; + tensor scatter_17_axis_0 = const()[name = tensor("scatter_17_axis_0"), val = tensor(0)]; + tensor scatter_17_validate_indices_0 = const()[name = tensor("scatter_17_validate_indices_0"), val = tensor(false)]; + tensor scatter_17_cast_fp16 = scatter(axis = scatter_17_axis_0, data = reshape_88_cast_fp16, indices = reshape_86, mode = scatter_17_mode_0, updates = reshape_87_cast_fp16, validate_indices = scatter_17_validate_indices_0)[name = tensor("scatter_17_cast_fp16")]; + tensor reshape_89_cast_fp16 = reshape(shape = shape_65, x = scatter_17_cast_fp16)[name = tensor("reshape_89_cast_fp16")]; + tensor reshape_89_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("reshape_89_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor keys_49_begin_0 = const()[name = tensor("keys_49_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor keys_49_end_0 = const()[name = tensor("keys_49_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor keys_49_end_mask_0 = const()[name = tensor("keys_49_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor keys_49_squeeze_mask_0 = const()[name = tensor("keys_49_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor keys_49_cast_fp16 = slice_by_index(begin = keys_49_begin_0, end = keys_49_end_0, end_mask = keys_49_end_mask_0, squeeze_mask = keys_49_squeeze_mask_0, x = reshape_89_cast_fp16)[name = tensor("keys_49_cast_fp16")]; + tensor values_49_begin_0 = const()[name = tensor("values_49_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor values_49_end_0 = const()[name = tensor("values_49_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor values_49_end_mask_0 = const()[name = tensor("values_49_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor values_49_squeeze_mask_0 = const()[name = tensor("values_49_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor values_49_cast_fp16 = slice_by_index(begin = values_49_begin_0, end = values_49_end_0, end_mask = values_49_end_mask_0, squeeze_mask = values_49_squeeze_mask_0, x = reshape_89_cast_fp16)[name = tensor("values_49_cast_fp16")]; + tensor var_3545_cast_fp16 = not_equal(x = keys_49_cast_fp16, y = keys_49_cast_fp16)[name = tensor("op_3545_cast_fp16")]; + tensor keys_51_cast_fp16 = select(a = var_463_to_fp16, b = keys_49_cast_fp16, cond = var_3545_cast_fp16)[name = tensor("keys_51_cast_fp16")]; + tensor var_3553_cast_fp16 = not_equal(x = values_49_cast_fp16, y = values_49_cast_fp16)[name = tensor("op_3553_cast_fp16")]; + tensor values_51_cast_fp16 = select(a = var_463_to_fp16, b = values_49_cast_fp16, cond = var_3553_cast_fp16)[name = tensor("values_51_cast_fp16")]; + tensor var_3577 = const()[name = tensor("op_3577"), val = tensor([0, 2, 1, 3])]; + tensor var_3593 = const()[name = tensor("op_3593"), val = tensor([1, 1, 1])]; + tensor var_3594_cast_fp16 = reshape(shape = var_3593, x = position8_to_fp16)[name = tensor("op_3594_cast_fp16")]; + tensor var_3614_to_fp16 = const()[name = tensor("op_3614_to_fp16"), val = tensor(0x1p+0)]; + tensor valid_len_17_cast_fp16 = add(x = var_3594_cast_fp16, y = var_3614_to_fp16)[name = tensor("valid_len_17_cast_fp16")]; + tensor valid_mask_17_cast_fp16 = less(x = k_positions_1_promoted_to_fp16, y = valid_len_17_cast_fp16)[name = tensor("valid_mask_17_cast_fp16")]; + tensor causal_mask_17_cast_fp16 = less_equal(x = k_positions_1_promoted_to_fp16, y = var_3594_cast_fp16)[name = tensor("causal_mask_17_cast_fp16")]; + tensor attn_mask_33 = logical_and(x = valid_mask_17_cast_fp16, y = causal_mask_17_cast_fp16)[name = tensor("attn_mask_33")]; + tensor attn_mask_35_axes_0 = const()[name = tensor("attn_mask_35_axes_0"), val = tensor([1])]; + tensor attn_mask_35 = expand_dims(axes = attn_mask_35_axes_0, x = attn_mask_33)[name = tensor("attn_mask_35")]; + tensor var_3632_transpose_x_0 = const()[name = tensor("op_3632_transpose_x_0"), val = tensor(false)]; + tensor var_3632_transpose_y_0 = const()[name = tensor("op_3632_transpose_y_0"), val = tensor(false)]; + tensor transpose_88_perm_0 = const()[name = tensor("transpose_88_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_89_perm_0 = const()[name = tensor("transpose_89_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_89 = transpose(perm = transpose_89_perm_0, x = keys_51_cast_fp16)[name = tensor("transpose_181")]; + tensor transpose_88 = transpose(perm = transpose_88_perm_0, x = q_51_cast_fp16)[name = tensor("transpose_182")]; + tensor var_3632_cast_fp16 = matmul(transpose_x = var_3632_transpose_x_0, transpose_y = var_3632_transpose_y_0, x = transpose_88, y = transpose_89)[name = tensor("op_3632_cast_fp16")]; + tensor var_3626_to_fp16 = const()[name = tensor("op_3626_to_fp16"), val = tensor([0x1p-3])]; + tensor attn_weights_49_cast_fp16 = mul(x = var_3632_cast_fp16, y = var_3626_to_fp16)[name = tensor("attn_weights_49_cast_fp16")]; + tensor var_3634 = logical_not(x = attn_mask_35)[name = tensor("op_3634")]; + tensor var_3635_to_fp16 = const()[name = tensor("op_3635_to_fp16"), val = tensor(-inf)]; + tensor attn_weights_51_cast_fp16 = select(a = var_3635_to_fp16, b = attn_weights_49_cast_fp16, cond = var_3634)[name = tensor("attn_weights_51_cast_fp16")]; + tensor var_3637 = const()[name = tensor("op_3637"), val = tensor(-1)]; + tensor attn_weights_53_cast_fp16 = softmax(axis = var_3637, x = attn_weights_51_cast_fp16)[name = tensor("attn_weights_53_cast_fp16")]; + tensor attn_output_17_transpose_x_0 = const()[name = tensor("attn_output_17_transpose_x_0"), val = tensor(false)]; + tensor attn_output_17_transpose_y_0 = const()[name = tensor("attn_output_17_transpose_y_0"), val = tensor(false)]; + tensor values_53_cast_fp16 = transpose(perm = var_3577, x = values_51_cast_fp16)[name = tensor("transpose_183")]; + tensor attn_output_17_cast_fp16 = matmul(transpose_x = attn_output_17_transpose_x_0, transpose_y = attn_output_17_transpose_y_0, x = attn_weights_53_cast_fp16, y = values_53_cast_fp16)[name = tensor("attn_output_17_cast_fp16")]; + tensor var_3645 = const()[name = tensor("op_3645"), val = tensor([0, 2, 1, 3])]; + tensor var_3650 = const()[name = tensor("op_3650"), val = tensor([1, 1, 1024])]; + tensor var_3646_cast_fp16 = transpose(perm = var_3645, x = attn_output_17_cast_fp16)[name = tensor("transpose_180")]; + tensor input_103_cast_fp16 = reshape(shape = var_3650, x = var_3646_cast_fp16)[name = tensor("input_103_cast_fp16")]; + tensor attn8_out_proj_weight_to_fp16 = const()[name = tensor("attn8_out_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(208825088)))]; + tensor linear_34_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = attn8_out_proj_weight_to_fp16, x = input_103_cast_fp16)[name = tensor("linear_34_cast_fp16")]; + tensor var_3656_to_fp16 = const()[name = tensor("op_3656_to_fp16"), val = tensor(0x1p+0)]; + tensor var_3657_cast_fp16 = add(x = position8_to_fp16, y = var_3656_to_fp16)[name = tensor("op_3657_cast_fp16")]; + tensor var_3657_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_3657_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor input_105_cast_fp16 = add(x = input_99_cast_fp16, y = linear_34_cast_fp16)[name = tensor("input_105_cast_fp16")]; + tensor input_107_axes_0 = const()[name = tensor("input_107_axes_0"), val = tensor([-1])]; + tensor norm8_2_weight_to_fp16 = const()[name = tensor("norm8_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(210922304)))]; + tensor norm8_2_bias_to_fp16 = const()[name = tensor("norm8_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(210924416)))]; + tensor var_3661_to_fp16 = const()[name = tensor("op_3661_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_107_cast_fp16 = layer_norm(axes = input_107_axes_0, beta = norm8_2_bias_to_fp16, epsilon = var_3661_to_fp16, gamma = norm8_2_weight_to_fp16, x = input_105_cast_fp16)[name = tensor("input_107_cast_fp16")]; + tensor linear8_1_weight_to_fp16 = const()[name = tensor("linear8_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(210926528)))]; + tensor linear_35_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = linear8_1_weight_to_fp16, x = input_107_cast_fp16)[name = tensor("linear_35_cast_fp16")]; + tensor input_109_mode_0 = const()[name = tensor("input_109_mode_0"), val = tensor("EXACT")]; + tensor input_109_cast_fp16 = gelu(mode = input_109_mode_0, x = linear_35_cast_fp16)[name = tensor("input_109_cast_fp16")]; + tensor linear8_2_weight_to_fp16 = const()[name = tensor("linear8_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(219315200)))]; + tensor linear_36_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = linear8_2_weight_to_fp16, x = input_109_cast_fp16)[name = tensor("linear_36_cast_fp16")]; + tensor input_111_cast_fp16 = add(x = input_105_cast_fp16, y = linear_36_cast_fp16)[name = tensor("input_111_cast_fp16")]; + tensor input_113_axes_0 = const()[name = tensor("input_113_axes_0"), val = tensor([-1])]; + tensor norm9_1_weight_to_fp16 = const()[name = tensor("norm9_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(227703872)))]; + tensor norm9_1_bias_to_fp16 = const()[name = tensor("norm9_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(227705984)))]; + tensor var_3678_to_fp16 = const()[name = tensor("op_3678_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_113_cast_fp16 = layer_norm(axes = input_113_axes_0, beta = norm9_1_bias_to_fp16, epsilon = var_3678_to_fp16, gamma = norm9_1_weight_to_fp16, x = input_111_cast_fp16)[name = tensor("input_113_cast_fp16")]; + tensor attn9_in_proj_weight_to_fp16 = const()[name = tensor("attn9_in_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(227708096)))]; + tensor linear_37_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = attn9_in_proj_weight_to_fp16, x = input_113_cast_fp16)[name = tensor("linear_37_cast_fp16")]; + tensor var_3692 = const()[name = tensor("op_3692"), val = tensor([1, 1, 3, 16, 64])]; + tensor qkv_19_cast_fp16 = reshape(shape = var_3692, x = linear_37_cast_fp16)[name = tensor("qkv_19_cast_fp16")]; + tensor q_55_begin_0 = const()[name = tensor("q_55_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor q_55_end_0 = const()[name = tensor("q_55_end_0"), val = tensor([1, 1, 1, 16, 64])]; + tensor q_55_end_mask_0 = const()[name = tensor("q_55_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor q_55_squeeze_mask_0 = const()[name = tensor("q_55_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor q_55_cast_fp16 = slice_by_index(begin = q_55_begin_0, end = q_55_end_0, end_mask = q_55_end_mask_0, squeeze_mask = q_55_squeeze_mask_0, x = qkv_19_cast_fp16)[name = tensor("q_55_cast_fp16")]; + tensor k_37_begin_0 = const()[name = tensor("k_37_begin_0"), val = tensor([0, 0, 1, 0, 0])]; + tensor k_37_end_0 = const()[name = tensor("k_37_end_0"), val = tensor([1, 1, 2, 16, 64])]; + tensor k_37_end_mask_0 = const()[name = tensor("k_37_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor k_37_squeeze_mask_0 = const()[name = tensor("k_37_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor k_37_cast_fp16 = slice_by_index(begin = k_37_begin_0, end = k_37_end_0, end_mask = k_37_end_mask_0, squeeze_mask = k_37_squeeze_mask_0, x = qkv_19_cast_fp16)[name = tensor("k_37_cast_fp16")]; + tensor v_19_begin_0 = const()[name = tensor("v_19_begin_0"), val = tensor([0, 0, 2, 0, 0])]; + tensor v_19_end_0 = const()[name = tensor("v_19_end_0"), val = tensor([1, 1, 3, 16, 64])]; + tensor v_19_end_mask_0 = const()[name = tensor("v_19_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor v_19_squeeze_mask_0 = const()[name = tensor("v_19_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor v_19_cast_fp16 = slice_by_index(begin = v_19_begin_0, end = v_19_end_0, end_mask = v_19_end_mask_0, squeeze_mask = v_19_squeeze_mask_0, x = qkv_19_cast_fp16)[name = tensor("v_19_cast_fp16")]; + tensor var_3758 = const()[name = tensor("op_3758"), val = tensor([1, 1, 1, 1])]; + tensor position9_to_fp16_dtype_0 = const()[name = tensor("position9_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor position9_to_fp16 = cast(dtype = position9_to_fp16_dtype_0, x = position9)[name = tensor("cast_508")]; + tensor ts_59_cast_fp16 = reshape(shape = var_3758, x = position9_to_fp16)[name = tensor("ts_59_cast_fp16")]; + tensor var_3765 = const()[name = tensor("op_3765"), val = tensor([1, 1, 16, 32, 2])]; + tensor q_complex_19_cast_fp16 = reshape(shape = var_3765, x = q_55_cast_fp16)[name = tensor("q_complex_19_cast_fp16")]; + tensor var_3772 = const()[name = tensor("op_3772"), val = tensor([1, 1, 16, 32, 2])]; + tensor k_complex_19_cast_fp16 = reshape(shape = var_3772, x = k_37_cast_fp16)[name = tensor("k_complex_19_cast_fp16")]; + tensor var_3776_begin_0 = const()[name = tensor("op_3776_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_3776_end_0 = const()[name = tensor("op_3776_end_0"), val = tensor([1, 1, 16, 32, 1])]; + tensor var_3776_end_mask_0 = const()[name = tensor("op_3776_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_3776_squeeze_mask_0 = const()[name = tensor("op_3776_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_3776_cast_fp16 = slice_by_index(begin = var_3776_begin_0, end = var_3776_end_0, end_mask = var_3776_end_mask_0, squeeze_mask = var_3776_squeeze_mask_0, x = q_complex_19_cast_fp16)[name = tensor("op_3776_cast_fp16")]; + tensor var_3784_begin_0 = const()[name = tensor("op_3784_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_3784_end_0 = const()[name = tensor("op_3784_end_0"), val = tensor([1, 1, 16, 32, 2])]; + tensor var_3784_end_mask_0 = const()[name = tensor("op_3784_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_3784_squeeze_mask_0 = const()[name = tensor("op_3784_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_3784_cast_fp16 = slice_by_index(begin = var_3784_begin_0, end = var_3784_end_0, end_mask = var_3784_end_mask_0, squeeze_mask = var_3784_squeeze_mask_0, x = q_complex_19_cast_fp16)[name = tensor("op_3784_cast_fp16")]; + tensor var_3792_begin_0 = const()[name = tensor("op_3792_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_3792_end_0 = const()[name = tensor("op_3792_end_0"), val = tensor([1, 1, 16, 32, 1])]; + tensor var_3792_end_mask_0 = const()[name = tensor("op_3792_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_3792_squeeze_mask_0 = const()[name = tensor("op_3792_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_3792_cast_fp16 = slice_by_index(begin = var_3792_begin_0, end = var_3792_end_0, end_mask = var_3792_end_mask_0, squeeze_mask = var_3792_squeeze_mask_0, x = k_complex_19_cast_fp16)[name = tensor("op_3792_cast_fp16")]; + tensor var_3800_begin_0 = const()[name = tensor("op_3800_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_3800_end_0 = const()[name = tensor("op_3800_end_0"), val = tensor([1, 1, 16, 32, 2])]; + tensor var_3800_end_mask_0 = const()[name = tensor("op_3800_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_3800_squeeze_mask_0 = const()[name = tensor("op_3800_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_3800_cast_fp16 = slice_by_index(begin = var_3800_begin_0, end = var_3800_end_0, end_mask = var_3800_end_mask_0, squeeze_mask = var_3800_squeeze_mask_0, x = k_complex_19_cast_fp16)[name = tensor("op_3800_cast_fp16")]; + tensor freqs_19_to_fp16 = const()[name = tensor("freqs_19_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(233999616)))]; + tensor var_3806_cast_fp16 = mul(x = freqs_19_to_fp16, y = ts_59_cast_fp16)[name = tensor("op_3806_cast_fp16")]; + tensor rotr_19_cast_fp16 = cos(x = var_3806_cast_fp16)[name = tensor("rotr_19_cast_fp16")]; + tensor roti_19_cast_fp16 = sin(x = var_3806_cast_fp16)[name = tensor("roti_19_cast_fp16")]; + tensor var_3810_cast_fp16 = mul(x = var_3776_cast_fp16, y = rotr_19_cast_fp16)[name = tensor("op_3810_cast_fp16")]; + tensor var_3811_cast_fp16 = mul(x = var_3784_cast_fp16, y = roti_19_cast_fp16)[name = tensor("op_3811_cast_fp16")]; + tensor qor_37_cast_fp16 = sub(x = var_3810_cast_fp16, y = var_3811_cast_fp16)[name = tensor("qor_37_cast_fp16")]; + tensor var_3814_cast_fp16 = mul(x = var_3776_cast_fp16, y = roti_19_cast_fp16)[name = tensor("op_3814_cast_fp16")]; + tensor var_3815_cast_fp16 = mul(x = var_3784_cast_fp16, y = rotr_19_cast_fp16)[name = tensor("op_3815_cast_fp16")]; + tensor qoi_37_cast_fp16 = add(x = var_3814_cast_fp16, y = var_3815_cast_fp16)[name = tensor("qoi_37_cast_fp16")]; + tensor var_3818_cast_fp16 = mul(x = var_3792_cast_fp16, y = rotr_19_cast_fp16)[name = tensor("op_3818_cast_fp16")]; + tensor var_3819_cast_fp16 = mul(x = var_3800_cast_fp16, y = roti_19_cast_fp16)[name = tensor("op_3819_cast_fp16")]; + tensor kor_37_cast_fp16 = sub(x = var_3818_cast_fp16, y = var_3819_cast_fp16)[name = tensor("kor_37_cast_fp16")]; + tensor var_3822_cast_fp16 = mul(x = var_3792_cast_fp16, y = roti_19_cast_fp16)[name = tensor("op_3822_cast_fp16")]; + tensor var_3823_cast_fp16 = mul(x = var_3800_cast_fp16, y = rotr_19_cast_fp16)[name = tensor("op_3823_cast_fp16")]; + tensor koi_37_cast_fp16 = add(x = var_3822_cast_fp16, y = var_3823_cast_fp16)[name = tensor("koi_37_cast_fp16")]; + tensor qo_19_axis_0 = const()[name = tensor("qo_19_axis_0"), val = tensor(-1)]; + tensor qo_19_cast_fp16 = stack(axis = qo_19_axis_0, values = (qor_37_cast_fp16, qoi_37_cast_fp16))[name = tensor("qo_19_cast_fp16")]; + tensor ko_19_axis_0 = const()[name = tensor("ko_19_axis_0"), val = tensor(-1)]; + tensor ko_19_cast_fp16 = stack(axis = ko_19_axis_0, values = (kor_37_cast_fp16, koi_37_cast_fp16))[name = tensor("ko_19_cast_fp16")]; + tensor var_3856 = const()[name = tensor("op_3856"), val = tensor([1, 1, 16, 64])]; + tensor q_57_cast_fp16 = reshape(shape = var_3856, x = qo_19_cast_fp16)[name = tensor("q_57_cast_fp16")]; + tensor var_3862 = const()[name = tensor("op_3862"), val = tensor([1, 1, 16, 64])]; + tensor k_39_cast_fp16 = reshape(shape = var_3862, x = ko_19_cast_fp16)[name = tensor("k_39_cast_fp16")]; + tensor _inversed_3887_y_0_to_fp16 = const()[name = tensor("_inversed_3887_y_0_to_fp16"), val = tensor(0x1p-9)]; + tensor _inversed_3887_cast_fp16 = mul(x = ts_59_cast_fp16, y = _inversed_3887_y_0_to_fp16)[name = tensor("_inversed_3887_cast_fp16")]; + tensor var_3888_cast_fp16 = floor(x = _inversed_3887_cast_fp16)[name = tensor("op_3888_cast_fp16")]; + tensor var_3889_to_fp16 = const()[name = tensor("op_3889_to_fp16"), val = tensor(0x1p+9)]; + tensor var_3890_cast_fp16 = mul(x = var_3888_cast_fp16, y = var_3889_to_fp16)[name = tensor("op_3890_cast_fp16")]; + tensor write_indices_float_39_cast_fp16 = sub(x = ts_59_cast_fp16, y = var_3890_cast_fp16)[name = tensor("write_indices_float_39_cast_fp16")]; + tensor var_3897_dtype_0 = const()[name = tensor("op_3897_dtype_0"), val = tensor("int32")]; + tensor write_indices_19_reps_0 = const()[name = tensor("write_indices_19_reps_0"), val = tensor([1, 1, 16, 64])]; + tensor write_indices_float_39_cast_fp16_to_int32 = cast(dtype = var_3897_dtype_0, x = write_indices_float_39_cast_fp16)[name = tensor("cast_507")]; + tensor write_indices_19 = tile(reps = write_indices_19_reps_0, x = write_indices_float_39_cast_fp16_to_int32)[name = tensor("write_indices_19")]; + tensor var_3907_begin_0 = const()[name = tensor("op_3907_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_3907_end_0 = const()[name = tensor("op_3907_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor var_3907_end_mask_0 = const()[name = tensor("op_3907_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_3907_squeeze_mask_0 = const()[name = tensor("op_3907_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor cache9_to_fp16_dtype_0 = const()[name = tensor("cache9_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor cache9_to_fp16 = cast(dtype = cache9_to_fp16_dtype_0, x = cache9)[name = tensor("cast_506")]; + tensor var_3907_cast_fp16 = slice_by_index(begin = var_3907_begin_0, end = var_3907_end_0, end_mask = var_3907_end_mask_0, squeeze_mask = var_3907_squeeze_mask_0, x = cache9_to_fp16)[name = tensor("op_3907_cast_fp16")]; + tensor var_3909_axis_0 = const()[name = tensor("op_3909_axis_0"), val = tensor(1)]; + tensor var_3909_mode_0 = const()[name = tensor("op_3909_mode_0"), val = tensor("update")]; + tensor var_3909_validate_indices_0 = const()[name = tensor("op_3909_validate_indices_0"), val = tensor(false)]; + tensor var_3909_cast_fp16 = scatter_along_axis(axis = var_3909_axis_0, data = var_3907_cast_fp16, indices = write_indices_19, mode = var_3909_mode_0, updates = k_39_cast_fp16, validate_indices = var_3909_validate_indices_0)[name = tensor("op_3909_cast_fp16")]; + tensor concat_65 = const()[name = tensor("concat_65"), val = tensor([0, 0, 0, 0, 0])]; + tensor concat_66 = const()[name = tensor("concat_66"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_19_internal_tensor_assign_1_stride_0 = const()[name = tensor("new_cache_19_internal_tensor_assign_1_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_19_internal_tensor_assign_1_begin_mask_0 = const()[name = tensor("new_cache_19_internal_tensor_assign_1_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_19_internal_tensor_assign_1_end_mask_0 = const()[name = tensor("new_cache_19_internal_tensor_assign_1_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_19_internal_tensor_assign_1_squeeze_mask_0 = const()[name = tensor("new_cache_19_internal_tensor_assign_1_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_66 = const()[name = tensor("shape_66"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_18 = const()[name = tensor("reduce_prod_18"), val = tensor(1048576)]; + tensor range_1d_18_start_0 = const()[name = tensor("range_1d_18_start_0"), val = tensor(0)]; + tensor range_1d_18_step_0 = const()[name = tensor("range_1d_18_step_0"), val = tensor(1)]; + tensor range_1d_18 = range_1d(end = reduce_prod_18, start = range_1d_18_start_0, step = range_1d_18_step_0)[name = tensor("range_1d_18")]; + tensor reshape_90 = reshape(shape = shape_66, x = range_1d_18)[name = tensor("reshape_90")]; + tensor slice_by_index_18 = slice_by_index(begin = concat_65, begin_mask = new_cache_19_internal_tensor_assign_1_begin_mask_0, end = concat_66, end_mask = new_cache_19_internal_tensor_assign_1_end_mask_0, squeeze_mask = new_cache_19_internal_tensor_assign_1_squeeze_mask_0, stride = new_cache_19_internal_tensor_assign_1_stride_0, x = reshape_90)[name = tensor("slice_by_index_18")]; + tensor reshape_91_shape_0 = const()[name = tensor("reshape_91_shape_0"), val = tensor([-1])]; + tensor reshape_91 = reshape(shape = reshape_91_shape_0, x = slice_by_index_18)[name = tensor("reshape_91")]; + tensor reshape_92_shape_0 = const()[name = tensor("reshape_92_shape_0"), val = tensor([-1])]; + tensor reshape_92_cast_fp16 = reshape(shape = reshape_92_shape_0, x = var_3909_cast_fp16)[name = tensor("reshape_92_cast_fp16")]; + tensor reshape_93_shape_0 = const()[name = tensor("reshape_93_shape_0"), val = tensor([-1])]; + tensor reshape_93_cast_fp16 = reshape(shape = reshape_93_shape_0, x = cache9_to_fp16)[name = tensor("reshape_93_cast_fp16")]; + tensor scatter_18_mode_0 = const()[name = tensor("scatter_18_mode_0"), val = tensor("update")]; + tensor scatter_18_axis_0 = const()[name = tensor("scatter_18_axis_0"), val = tensor(0)]; + tensor scatter_18_validate_indices_0 = const()[name = tensor("scatter_18_validate_indices_0"), val = tensor(false)]; + tensor scatter_18_cast_fp16 = scatter(axis = scatter_18_axis_0, data = reshape_93_cast_fp16, indices = reshape_91, mode = scatter_18_mode_0, updates = reshape_92_cast_fp16, validate_indices = scatter_18_validate_indices_0)[name = tensor("scatter_18_cast_fp16")]; + tensor reshape_94_cast_fp16 = reshape(shape = shape_66, x = scatter_18_cast_fp16)[name = tensor("reshape_94_cast_fp16")]; + tensor var_3917_begin_0 = const()[name = tensor("op_3917_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor var_3917_end_0 = const()[name = tensor("op_3917_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor var_3917_end_mask_0 = const()[name = tensor("op_3917_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_3917_squeeze_mask_0 = const()[name = tensor("op_3917_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor var_3917_cast_fp16 = slice_by_index(begin = var_3917_begin_0, end = var_3917_end_0, end_mask = var_3917_end_mask_0, squeeze_mask = var_3917_squeeze_mask_0, x = reshape_94_cast_fp16)[name = tensor("op_3917_cast_fp16")]; + tensor var_3919_axis_0 = const()[name = tensor("op_3919_axis_0"), val = tensor(1)]; + tensor var_3919_mode_0 = const()[name = tensor("op_3919_mode_0"), val = tensor("update")]; + tensor var_3919_validate_indices_0 = const()[name = tensor("op_3919_validate_indices_0"), val = tensor(false)]; + tensor var_3919_cast_fp16 = scatter_along_axis(axis = var_3919_axis_0, data = var_3917_cast_fp16, indices = write_indices_19, mode = var_3919_mode_0, updates = v_19_cast_fp16, validate_indices = var_3919_validate_indices_0)[name = tensor("op_3919_cast_fp16")]; + tensor concat_67 = const()[name = tensor("concat_67"), val = tensor([1, 0, 0, 0, 0])]; + tensor concat_68 = const()[name = tensor("concat_68"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_19_internal_tensor_assign_2_stride_0 = const()[name = tensor("new_cache_19_internal_tensor_assign_2_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_19_internal_tensor_assign_2_begin_mask_0 = const()[name = tensor("new_cache_19_internal_tensor_assign_2_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_19_internal_tensor_assign_2_end_mask_0 = const()[name = tensor("new_cache_19_internal_tensor_assign_2_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_19_internal_tensor_assign_2_squeeze_mask_0 = const()[name = tensor("new_cache_19_internal_tensor_assign_2_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_67 = const()[name = tensor("shape_67"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_19 = const()[name = tensor("reduce_prod_19"), val = tensor(1048576)]; + tensor range_1d_19_start_0 = const()[name = tensor("range_1d_19_start_0"), val = tensor(0)]; + tensor range_1d_19_step_0 = const()[name = tensor("range_1d_19_step_0"), val = tensor(1)]; + tensor range_1d_19 = range_1d(end = reduce_prod_19, start = range_1d_19_start_0, step = range_1d_19_step_0)[name = tensor("range_1d_19")]; + tensor reshape_95 = reshape(shape = shape_67, x = range_1d_19)[name = tensor("reshape_95")]; + tensor slice_by_index_19 = slice_by_index(begin = concat_67, begin_mask = new_cache_19_internal_tensor_assign_2_begin_mask_0, end = concat_68, end_mask = new_cache_19_internal_tensor_assign_2_end_mask_0, squeeze_mask = new_cache_19_internal_tensor_assign_2_squeeze_mask_0, stride = new_cache_19_internal_tensor_assign_2_stride_0, x = reshape_95)[name = tensor("slice_by_index_19")]; + tensor reshape_96_shape_0 = const()[name = tensor("reshape_96_shape_0"), val = tensor([-1])]; + tensor reshape_96 = reshape(shape = reshape_96_shape_0, x = slice_by_index_19)[name = tensor("reshape_96")]; + tensor reshape_97_shape_0 = const()[name = tensor("reshape_97_shape_0"), val = tensor([-1])]; + tensor reshape_97_cast_fp16 = reshape(shape = reshape_97_shape_0, x = var_3919_cast_fp16)[name = tensor("reshape_97_cast_fp16")]; + tensor reshape_98_shape_0 = const()[name = tensor("reshape_98_shape_0"), val = tensor([-1])]; + tensor reshape_98_cast_fp16 = reshape(shape = reshape_98_shape_0, x = reshape_94_cast_fp16)[name = tensor("reshape_98_cast_fp16")]; + tensor scatter_19_mode_0 = const()[name = tensor("scatter_19_mode_0"), val = tensor("update")]; + tensor scatter_19_axis_0 = const()[name = tensor("scatter_19_axis_0"), val = tensor(0)]; + tensor scatter_19_validate_indices_0 = const()[name = tensor("scatter_19_validate_indices_0"), val = tensor(false)]; + tensor scatter_19_cast_fp16 = scatter(axis = scatter_19_axis_0, data = reshape_98_cast_fp16, indices = reshape_96, mode = scatter_19_mode_0, updates = reshape_97_cast_fp16, validate_indices = scatter_19_validate_indices_0)[name = tensor("scatter_19_cast_fp16")]; + tensor reshape_99_cast_fp16 = reshape(shape = shape_67, x = scatter_19_cast_fp16)[name = tensor("reshape_99_cast_fp16")]; + tensor reshape_99_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("reshape_99_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor keys_55_begin_0 = const()[name = tensor("keys_55_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor keys_55_end_0 = const()[name = tensor("keys_55_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor keys_55_end_mask_0 = const()[name = tensor("keys_55_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor keys_55_squeeze_mask_0 = const()[name = tensor("keys_55_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor keys_55_cast_fp16 = slice_by_index(begin = keys_55_begin_0, end = keys_55_end_0, end_mask = keys_55_end_mask_0, squeeze_mask = keys_55_squeeze_mask_0, x = reshape_99_cast_fp16)[name = tensor("keys_55_cast_fp16")]; + tensor values_55_begin_0 = const()[name = tensor("values_55_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor values_55_end_0 = const()[name = tensor("values_55_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor values_55_end_mask_0 = const()[name = tensor("values_55_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor values_55_squeeze_mask_0 = const()[name = tensor("values_55_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor values_55_cast_fp16 = slice_by_index(begin = values_55_begin_0, end = values_55_end_0, end_mask = values_55_end_mask_0, squeeze_mask = values_55_squeeze_mask_0, x = reshape_99_cast_fp16)[name = tensor("values_55_cast_fp16")]; + tensor var_3931_cast_fp16 = not_equal(x = keys_55_cast_fp16, y = keys_55_cast_fp16)[name = tensor("op_3931_cast_fp16")]; + tensor keys_57_cast_fp16 = select(a = var_463_to_fp16, b = keys_55_cast_fp16, cond = var_3931_cast_fp16)[name = tensor("keys_57_cast_fp16")]; + tensor var_3939_cast_fp16 = not_equal(x = values_55_cast_fp16, y = values_55_cast_fp16)[name = tensor("op_3939_cast_fp16")]; + tensor values_57_cast_fp16 = select(a = var_463_to_fp16, b = values_55_cast_fp16, cond = var_3939_cast_fp16)[name = tensor("values_57_cast_fp16")]; + tensor var_3963 = const()[name = tensor("op_3963"), val = tensor([0, 2, 1, 3])]; + tensor var_3979 = const()[name = tensor("op_3979"), val = tensor([1, 1, 1])]; + tensor var_3980_cast_fp16 = reshape(shape = var_3979, x = position9_to_fp16)[name = tensor("op_3980_cast_fp16")]; + tensor var_4000_to_fp16 = const()[name = tensor("op_4000_to_fp16"), val = tensor(0x1p+0)]; + tensor valid_len_19_cast_fp16 = add(x = var_3980_cast_fp16, y = var_4000_to_fp16)[name = tensor("valid_len_19_cast_fp16")]; + tensor valid_mask_19_cast_fp16 = less(x = k_positions_1_promoted_to_fp16, y = valid_len_19_cast_fp16)[name = tensor("valid_mask_19_cast_fp16")]; + tensor causal_mask_19_cast_fp16 = less_equal(x = k_positions_1_promoted_to_fp16, y = var_3980_cast_fp16)[name = tensor("causal_mask_19_cast_fp16")]; + tensor attn_mask_37 = logical_and(x = valid_mask_19_cast_fp16, y = causal_mask_19_cast_fp16)[name = tensor("attn_mask_37")]; + tensor attn_mask_39_axes_0 = const()[name = tensor("attn_mask_39_axes_0"), val = tensor([1])]; + tensor attn_mask_39 = expand_dims(axes = attn_mask_39_axes_0, x = attn_mask_37)[name = tensor("attn_mask_39")]; + tensor var_4018_transpose_x_0 = const()[name = tensor("op_4018_transpose_x_0"), val = tensor(false)]; + tensor var_4018_transpose_y_0 = const()[name = tensor("op_4018_transpose_y_0"), val = tensor(false)]; + tensor transpose_90_perm_0 = const()[name = tensor("transpose_90_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_91_perm_0 = const()[name = tensor("transpose_91_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_91 = transpose(perm = transpose_91_perm_0, x = keys_57_cast_fp16)[name = tensor("transpose_177")]; + tensor transpose_90 = transpose(perm = transpose_90_perm_0, x = q_57_cast_fp16)[name = tensor("transpose_178")]; + tensor var_4018_cast_fp16 = matmul(transpose_x = var_4018_transpose_x_0, transpose_y = var_4018_transpose_y_0, x = transpose_90, y = transpose_91)[name = tensor("op_4018_cast_fp16")]; + tensor var_4012_to_fp16 = const()[name = tensor("op_4012_to_fp16"), val = tensor([0x1p-3])]; + tensor attn_weights_55_cast_fp16 = mul(x = var_4018_cast_fp16, y = var_4012_to_fp16)[name = tensor("attn_weights_55_cast_fp16")]; + tensor var_4020 = logical_not(x = attn_mask_39)[name = tensor("op_4020")]; + tensor var_4021_to_fp16 = const()[name = tensor("op_4021_to_fp16"), val = tensor(-inf)]; + tensor attn_weights_57_cast_fp16 = select(a = var_4021_to_fp16, b = attn_weights_55_cast_fp16, cond = var_4020)[name = tensor("attn_weights_57_cast_fp16")]; + tensor var_4023 = const()[name = tensor("op_4023"), val = tensor(-1)]; + tensor attn_weights_59_cast_fp16 = softmax(axis = var_4023, x = attn_weights_57_cast_fp16)[name = tensor("attn_weights_59_cast_fp16")]; + tensor attn_output_19_transpose_x_0 = const()[name = tensor("attn_output_19_transpose_x_0"), val = tensor(false)]; + tensor attn_output_19_transpose_y_0 = const()[name = tensor("attn_output_19_transpose_y_0"), val = tensor(false)]; + tensor values_59_cast_fp16 = transpose(perm = var_3963, x = values_57_cast_fp16)[name = tensor("transpose_179")]; + tensor attn_output_19_cast_fp16 = matmul(transpose_x = attn_output_19_transpose_x_0, transpose_y = attn_output_19_transpose_y_0, x = attn_weights_59_cast_fp16, y = values_59_cast_fp16)[name = tensor("attn_output_19_cast_fp16")]; + tensor var_4031 = const()[name = tensor("op_4031"), val = tensor([0, 2, 1, 3])]; + tensor var_4036 = const()[name = tensor("op_4036"), val = tensor([1, 1, 1024])]; + tensor var_4032_cast_fp16 = transpose(perm = var_4031, x = attn_output_19_cast_fp16)[name = tensor("transpose_176")]; + tensor input_115_cast_fp16 = reshape(shape = var_4036, x = var_4032_cast_fp16)[name = tensor("input_115_cast_fp16")]; + tensor attn9_out_proj_weight_to_fp16 = const()[name = tensor("attn9_out_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(233999744)))]; + tensor linear_38_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = attn9_out_proj_weight_to_fp16, x = input_115_cast_fp16)[name = tensor("linear_38_cast_fp16")]; + tensor var_4042_to_fp16 = const()[name = tensor("op_4042_to_fp16"), val = tensor(0x1p+0)]; + tensor var_4043_cast_fp16 = add(x = position9_to_fp16, y = var_4042_to_fp16)[name = tensor("op_4043_cast_fp16")]; + tensor var_4043_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_4043_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor input_117_cast_fp16 = add(x = input_111_cast_fp16, y = linear_38_cast_fp16)[name = tensor("input_117_cast_fp16")]; + tensor input_119_axes_0 = const()[name = tensor("input_119_axes_0"), val = tensor([-1])]; + tensor norm9_2_weight_to_fp16 = const()[name = tensor("norm9_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(236096960)))]; + tensor norm9_2_bias_to_fp16 = const()[name = tensor("norm9_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(236099072)))]; + tensor var_4047_to_fp16 = const()[name = tensor("op_4047_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_119_cast_fp16 = layer_norm(axes = input_119_axes_0, beta = norm9_2_bias_to_fp16, epsilon = var_4047_to_fp16, gamma = norm9_2_weight_to_fp16, x = input_117_cast_fp16)[name = tensor("input_119_cast_fp16")]; + tensor linear9_1_weight_to_fp16 = const()[name = tensor("linear9_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(236101184)))]; + tensor linear_39_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = linear9_1_weight_to_fp16, x = input_119_cast_fp16)[name = tensor("linear_39_cast_fp16")]; + tensor input_121_mode_0 = const()[name = tensor("input_121_mode_0"), val = tensor("EXACT")]; + tensor input_121_cast_fp16 = gelu(mode = input_121_mode_0, x = linear_39_cast_fp16)[name = tensor("input_121_cast_fp16")]; + tensor linear9_2_weight_to_fp16 = const()[name = tensor("linear9_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(244489856)))]; + tensor linear_40_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = linear9_2_weight_to_fp16, x = input_121_cast_fp16)[name = tensor("linear_40_cast_fp16")]; + tensor input_123_cast_fp16 = add(x = input_117_cast_fp16, y = linear_40_cast_fp16)[name = tensor("input_123_cast_fp16")]; + tensor input_125_axes_0 = const()[name = tensor("input_125_axes_0"), val = tensor([-1])]; + tensor norm10_1_weight_to_fp16 = const()[name = tensor("norm10_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(252878528)))]; + tensor norm10_1_bias_to_fp16 = const()[name = tensor("norm10_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(252880640)))]; + tensor var_4064_to_fp16 = const()[name = tensor("op_4064_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_125_cast_fp16 = layer_norm(axes = input_125_axes_0, beta = norm10_1_bias_to_fp16, epsilon = var_4064_to_fp16, gamma = norm10_1_weight_to_fp16, x = input_123_cast_fp16)[name = tensor("input_125_cast_fp16")]; + tensor attn10_in_proj_weight_to_fp16 = const()[name = tensor("attn10_in_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(252882752)))]; + tensor linear_41_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = attn10_in_proj_weight_to_fp16, x = input_125_cast_fp16)[name = tensor("linear_41_cast_fp16")]; + tensor var_4078 = const()[name = tensor("op_4078"), val = tensor([1, 1, 3, 16, 64])]; + tensor qkv_21_cast_fp16 = reshape(shape = var_4078, x = linear_41_cast_fp16)[name = tensor("qkv_21_cast_fp16")]; + tensor q_61_begin_0 = const()[name = tensor("q_61_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor q_61_end_0 = const()[name = tensor("q_61_end_0"), val = tensor([1, 1, 1, 16, 64])]; + tensor q_61_end_mask_0 = const()[name = tensor("q_61_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor q_61_squeeze_mask_0 = const()[name = tensor("q_61_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor q_61_cast_fp16 = slice_by_index(begin = q_61_begin_0, end = q_61_end_0, end_mask = q_61_end_mask_0, squeeze_mask = q_61_squeeze_mask_0, x = qkv_21_cast_fp16)[name = tensor("q_61_cast_fp16")]; + tensor k_41_begin_0 = const()[name = tensor("k_41_begin_0"), val = tensor([0, 0, 1, 0, 0])]; + tensor k_41_end_0 = const()[name = tensor("k_41_end_0"), val = tensor([1, 1, 2, 16, 64])]; + tensor k_41_end_mask_0 = const()[name = tensor("k_41_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor k_41_squeeze_mask_0 = const()[name = tensor("k_41_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor k_41_cast_fp16 = slice_by_index(begin = k_41_begin_0, end = k_41_end_0, end_mask = k_41_end_mask_0, squeeze_mask = k_41_squeeze_mask_0, x = qkv_21_cast_fp16)[name = tensor("k_41_cast_fp16")]; + tensor v_21_begin_0 = const()[name = tensor("v_21_begin_0"), val = tensor([0, 0, 2, 0, 0])]; + tensor v_21_end_0 = const()[name = tensor("v_21_end_0"), val = tensor([1, 1, 3, 16, 64])]; + tensor v_21_end_mask_0 = const()[name = tensor("v_21_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor v_21_squeeze_mask_0 = const()[name = tensor("v_21_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor v_21_cast_fp16 = slice_by_index(begin = v_21_begin_0, end = v_21_end_0, end_mask = v_21_end_mask_0, squeeze_mask = v_21_squeeze_mask_0, x = qkv_21_cast_fp16)[name = tensor("v_21_cast_fp16")]; + tensor var_4144 = const()[name = tensor("op_4144"), val = tensor([1, 1, 1, 1])]; + tensor position10_to_fp16_dtype_0 = const()[name = tensor("position10_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor position10_to_fp16 = cast(dtype = position10_to_fp16_dtype_0, x = position10)[name = tensor("cast_503")]; + tensor ts_65_cast_fp16 = reshape(shape = var_4144, x = position10_to_fp16)[name = tensor("ts_65_cast_fp16")]; + tensor var_4151 = const()[name = tensor("op_4151"), val = tensor([1, 1, 16, 32, 2])]; + tensor q_complex_21_cast_fp16 = reshape(shape = var_4151, x = q_61_cast_fp16)[name = tensor("q_complex_21_cast_fp16")]; + tensor var_4158 = const()[name = tensor("op_4158"), val = tensor([1, 1, 16, 32, 2])]; + tensor k_complex_21_cast_fp16 = reshape(shape = var_4158, x = k_41_cast_fp16)[name = tensor("k_complex_21_cast_fp16")]; + tensor var_4162_begin_0 = const()[name = tensor("op_4162_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_4162_end_0 = const()[name = tensor("op_4162_end_0"), val = tensor([1, 1, 16, 32, 1])]; + tensor var_4162_end_mask_0 = const()[name = tensor("op_4162_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_4162_squeeze_mask_0 = const()[name = tensor("op_4162_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_4162_cast_fp16 = slice_by_index(begin = var_4162_begin_0, end = var_4162_end_0, end_mask = var_4162_end_mask_0, squeeze_mask = var_4162_squeeze_mask_0, x = q_complex_21_cast_fp16)[name = tensor("op_4162_cast_fp16")]; + tensor var_4170_begin_0 = const()[name = tensor("op_4170_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_4170_end_0 = const()[name = tensor("op_4170_end_0"), val = tensor([1, 1, 16, 32, 2])]; + tensor var_4170_end_mask_0 = const()[name = tensor("op_4170_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_4170_squeeze_mask_0 = const()[name = tensor("op_4170_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_4170_cast_fp16 = slice_by_index(begin = var_4170_begin_0, end = var_4170_end_0, end_mask = var_4170_end_mask_0, squeeze_mask = var_4170_squeeze_mask_0, x = q_complex_21_cast_fp16)[name = tensor("op_4170_cast_fp16")]; + tensor var_4178_begin_0 = const()[name = tensor("op_4178_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_4178_end_0 = const()[name = tensor("op_4178_end_0"), val = tensor([1, 1, 16, 32, 1])]; + tensor var_4178_end_mask_0 = const()[name = tensor("op_4178_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_4178_squeeze_mask_0 = const()[name = tensor("op_4178_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_4178_cast_fp16 = slice_by_index(begin = var_4178_begin_0, end = var_4178_end_0, end_mask = var_4178_end_mask_0, squeeze_mask = var_4178_squeeze_mask_0, x = k_complex_21_cast_fp16)[name = tensor("op_4178_cast_fp16")]; + tensor var_4186_begin_0 = const()[name = tensor("op_4186_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_4186_end_0 = const()[name = tensor("op_4186_end_0"), val = tensor([1, 1, 16, 32, 2])]; + tensor var_4186_end_mask_0 = const()[name = tensor("op_4186_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_4186_squeeze_mask_0 = const()[name = tensor("op_4186_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_4186_cast_fp16 = slice_by_index(begin = var_4186_begin_0, end = var_4186_end_0, end_mask = var_4186_end_mask_0, squeeze_mask = var_4186_squeeze_mask_0, x = k_complex_21_cast_fp16)[name = tensor("op_4186_cast_fp16")]; + tensor freqs_21_to_fp16 = const()[name = tensor("freqs_21_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(259174272)))]; + tensor var_4192_cast_fp16 = mul(x = freqs_21_to_fp16, y = ts_65_cast_fp16)[name = tensor("op_4192_cast_fp16")]; + tensor rotr_21_cast_fp16 = cos(x = var_4192_cast_fp16)[name = tensor("rotr_21_cast_fp16")]; + tensor roti_21_cast_fp16 = sin(x = var_4192_cast_fp16)[name = tensor("roti_21_cast_fp16")]; + tensor var_4196_cast_fp16 = mul(x = var_4162_cast_fp16, y = rotr_21_cast_fp16)[name = tensor("op_4196_cast_fp16")]; + tensor var_4197_cast_fp16 = mul(x = var_4170_cast_fp16, y = roti_21_cast_fp16)[name = tensor("op_4197_cast_fp16")]; + tensor qor_41_cast_fp16 = sub(x = var_4196_cast_fp16, y = var_4197_cast_fp16)[name = tensor("qor_41_cast_fp16")]; + tensor var_4200_cast_fp16 = mul(x = var_4162_cast_fp16, y = roti_21_cast_fp16)[name = tensor("op_4200_cast_fp16")]; + tensor var_4201_cast_fp16 = mul(x = var_4170_cast_fp16, y = rotr_21_cast_fp16)[name = tensor("op_4201_cast_fp16")]; + tensor qoi_41_cast_fp16 = add(x = var_4200_cast_fp16, y = var_4201_cast_fp16)[name = tensor("qoi_41_cast_fp16")]; + tensor var_4204_cast_fp16 = mul(x = var_4178_cast_fp16, y = rotr_21_cast_fp16)[name = tensor("op_4204_cast_fp16")]; + tensor var_4205_cast_fp16 = mul(x = var_4186_cast_fp16, y = roti_21_cast_fp16)[name = tensor("op_4205_cast_fp16")]; + tensor kor_41_cast_fp16 = sub(x = var_4204_cast_fp16, y = var_4205_cast_fp16)[name = tensor("kor_41_cast_fp16")]; + tensor var_4208_cast_fp16 = mul(x = var_4178_cast_fp16, y = roti_21_cast_fp16)[name = tensor("op_4208_cast_fp16")]; + tensor var_4209_cast_fp16 = mul(x = var_4186_cast_fp16, y = rotr_21_cast_fp16)[name = tensor("op_4209_cast_fp16")]; + tensor koi_41_cast_fp16 = add(x = var_4208_cast_fp16, y = var_4209_cast_fp16)[name = tensor("koi_41_cast_fp16")]; + tensor qo_21_axis_0 = const()[name = tensor("qo_21_axis_0"), val = tensor(-1)]; + tensor qo_21_cast_fp16 = stack(axis = qo_21_axis_0, values = (qor_41_cast_fp16, qoi_41_cast_fp16))[name = tensor("qo_21_cast_fp16")]; + tensor ko_21_axis_0 = const()[name = tensor("ko_21_axis_0"), val = tensor(-1)]; + tensor ko_21_cast_fp16 = stack(axis = ko_21_axis_0, values = (kor_41_cast_fp16, koi_41_cast_fp16))[name = tensor("ko_21_cast_fp16")]; + tensor var_4242 = const()[name = tensor("op_4242"), val = tensor([1, 1, 16, 64])]; + tensor q_63_cast_fp16 = reshape(shape = var_4242, x = qo_21_cast_fp16)[name = tensor("q_63_cast_fp16")]; + tensor var_4248 = const()[name = tensor("op_4248"), val = tensor([1, 1, 16, 64])]; + tensor k_43_cast_fp16 = reshape(shape = var_4248, x = ko_21_cast_fp16)[name = tensor("k_43_cast_fp16")]; + tensor _inversed_4273_y_0_to_fp16 = const()[name = tensor("_inversed_4273_y_0_to_fp16"), val = tensor(0x1p-9)]; + tensor _inversed_4273_cast_fp16 = mul(x = ts_65_cast_fp16, y = _inversed_4273_y_0_to_fp16)[name = tensor("_inversed_4273_cast_fp16")]; + tensor var_4274_cast_fp16 = floor(x = _inversed_4273_cast_fp16)[name = tensor("op_4274_cast_fp16")]; + tensor var_4275_to_fp16 = const()[name = tensor("op_4275_to_fp16"), val = tensor(0x1p+9)]; + tensor var_4276_cast_fp16 = mul(x = var_4274_cast_fp16, y = var_4275_to_fp16)[name = tensor("op_4276_cast_fp16")]; + tensor write_indices_float_43_cast_fp16 = sub(x = ts_65_cast_fp16, y = var_4276_cast_fp16)[name = tensor("write_indices_float_43_cast_fp16")]; + tensor var_4283_dtype_0 = const()[name = tensor("op_4283_dtype_0"), val = tensor("int32")]; + tensor write_indices_21_reps_0 = const()[name = tensor("write_indices_21_reps_0"), val = tensor([1, 1, 16, 64])]; + tensor write_indices_float_43_cast_fp16_to_int32 = cast(dtype = var_4283_dtype_0, x = write_indices_float_43_cast_fp16)[name = tensor("cast_502")]; + tensor write_indices_21 = tile(reps = write_indices_21_reps_0, x = write_indices_float_43_cast_fp16_to_int32)[name = tensor("write_indices_21")]; + tensor var_4293_begin_0 = const()[name = tensor("op_4293_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_4293_end_0 = const()[name = tensor("op_4293_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor var_4293_end_mask_0 = const()[name = tensor("op_4293_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_4293_squeeze_mask_0 = const()[name = tensor("op_4293_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor cache10_to_fp16_dtype_0 = const()[name = tensor("cache10_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor cache10_to_fp16 = cast(dtype = cache10_to_fp16_dtype_0, x = cache10)[name = tensor("cast_501")]; + tensor var_4293_cast_fp16 = slice_by_index(begin = var_4293_begin_0, end = var_4293_end_0, end_mask = var_4293_end_mask_0, squeeze_mask = var_4293_squeeze_mask_0, x = cache10_to_fp16)[name = tensor("op_4293_cast_fp16")]; + tensor var_4295_axis_0 = const()[name = tensor("op_4295_axis_0"), val = tensor(1)]; + tensor var_4295_mode_0 = const()[name = tensor("op_4295_mode_0"), val = tensor("update")]; + tensor var_4295_validate_indices_0 = const()[name = tensor("op_4295_validate_indices_0"), val = tensor(false)]; + tensor var_4295_cast_fp16 = scatter_along_axis(axis = var_4295_axis_0, data = var_4293_cast_fp16, indices = write_indices_21, mode = var_4295_mode_0, updates = k_43_cast_fp16, validate_indices = var_4295_validate_indices_0)[name = tensor("op_4295_cast_fp16")]; + tensor concat_72 = const()[name = tensor("concat_72"), val = tensor([0, 0, 0, 0, 0])]; + tensor concat_73 = const()[name = tensor("concat_73"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_21_internal_tensor_assign_1_stride_0 = const()[name = tensor("new_cache_21_internal_tensor_assign_1_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_21_internal_tensor_assign_1_begin_mask_0 = const()[name = tensor("new_cache_21_internal_tensor_assign_1_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_21_internal_tensor_assign_1_end_mask_0 = const()[name = tensor("new_cache_21_internal_tensor_assign_1_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_21_internal_tensor_assign_1_squeeze_mask_0 = const()[name = tensor("new_cache_21_internal_tensor_assign_1_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_68 = const()[name = tensor("shape_68"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_20 = const()[name = tensor("reduce_prod_20"), val = tensor(1048576)]; + tensor range_1d_20_start_0 = const()[name = tensor("range_1d_20_start_0"), val = tensor(0)]; + tensor range_1d_20_step_0 = const()[name = tensor("range_1d_20_step_0"), val = tensor(1)]; + tensor range_1d_20 = range_1d(end = reduce_prod_20, start = range_1d_20_start_0, step = range_1d_20_step_0)[name = tensor("range_1d_20")]; + tensor reshape_100 = reshape(shape = shape_68, x = range_1d_20)[name = tensor("reshape_100")]; + tensor slice_by_index_20 = slice_by_index(begin = concat_72, begin_mask = new_cache_21_internal_tensor_assign_1_begin_mask_0, end = concat_73, end_mask = new_cache_21_internal_tensor_assign_1_end_mask_0, squeeze_mask = new_cache_21_internal_tensor_assign_1_squeeze_mask_0, stride = new_cache_21_internal_tensor_assign_1_stride_0, x = reshape_100)[name = tensor("slice_by_index_20")]; + tensor reshape_101_shape_0 = const()[name = tensor("reshape_101_shape_0"), val = tensor([-1])]; + tensor reshape_101 = reshape(shape = reshape_101_shape_0, x = slice_by_index_20)[name = tensor("reshape_101")]; + tensor reshape_102_shape_0 = const()[name = tensor("reshape_102_shape_0"), val = tensor([-1])]; + tensor reshape_102_cast_fp16 = reshape(shape = reshape_102_shape_0, x = var_4295_cast_fp16)[name = tensor("reshape_102_cast_fp16")]; + tensor reshape_103_shape_0 = const()[name = tensor("reshape_103_shape_0"), val = tensor([-1])]; + tensor reshape_103_cast_fp16 = reshape(shape = reshape_103_shape_0, x = cache10_to_fp16)[name = tensor("reshape_103_cast_fp16")]; + tensor scatter_20_mode_0 = const()[name = tensor("scatter_20_mode_0"), val = tensor("update")]; + tensor scatter_20_axis_0 = const()[name = tensor("scatter_20_axis_0"), val = tensor(0)]; + tensor scatter_20_validate_indices_0 = const()[name = tensor("scatter_20_validate_indices_0"), val = tensor(false)]; + tensor scatter_20_cast_fp16 = scatter(axis = scatter_20_axis_0, data = reshape_103_cast_fp16, indices = reshape_101, mode = scatter_20_mode_0, updates = reshape_102_cast_fp16, validate_indices = scatter_20_validate_indices_0)[name = tensor("scatter_20_cast_fp16")]; + tensor reshape_104_cast_fp16 = reshape(shape = shape_68, x = scatter_20_cast_fp16)[name = tensor("reshape_104_cast_fp16")]; + tensor var_4303_begin_0 = const()[name = tensor("op_4303_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor var_4303_end_0 = const()[name = tensor("op_4303_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor var_4303_end_mask_0 = const()[name = tensor("op_4303_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_4303_squeeze_mask_0 = const()[name = tensor("op_4303_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor var_4303_cast_fp16 = slice_by_index(begin = var_4303_begin_0, end = var_4303_end_0, end_mask = var_4303_end_mask_0, squeeze_mask = var_4303_squeeze_mask_0, x = reshape_104_cast_fp16)[name = tensor("op_4303_cast_fp16")]; + tensor var_4305_axis_0 = const()[name = tensor("op_4305_axis_0"), val = tensor(1)]; + tensor var_4305_mode_0 = const()[name = tensor("op_4305_mode_0"), val = tensor("update")]; + tensor var_4305_validate_indices_0 = const()[name = tensor("op_4305_validate_indices_0"), val = tensor(false)]; + tensor var_4305_cast_fp16 = scatter_along_axis(axis = var_4305_axis_0, data = var_4303_cast_fp16, indices = write_indices_21, mode = var_4305_mode_0, updates = v_21_cast_fp16, validate_indices = var_4305_validate_indices_0)[name = tensor("op_4305_cast_fp16")]; + tensor concat_74 = const()[name = tensor("concat_74"), val = tensor([1, 0, 0, 0, 0])]; + tensor concat_75 = const()[name = tensor("concat_75"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_21_internal_tensor_assign_2_stride_0 = const()[name = tensor("new_cache_21_internal_tensor_assign_2_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_21_internal_tensor_assign_2_begin_mask_0 = const()[name = tensor("new_cache_21_internal_tensor_assign_2_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_21_internal_tensor_assign_2_end_mask_0 = const()[name = tensor("new_cache_21_internal_tensor_assign_2_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_21_internal_tensor_assign_2_squeeze_mask_0 = const()[name = tensor("new_cache_21_internal_tensor_assign_2_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_69 = const()[name = tensor("shape_69"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_21 = const()[name = tensor("reduce_prod_21"), val = tensor(1048576)]; + tensor range_1d_21_start_0 = const()[name = tensor("range_1d_21_start_0"), val = tensor(0)]; + tensor range_1d_21_step_0 = const()[name = tensor("range_1d_21_step_0"), val = tensor(1)]; + tensor range_1d_21 = range_1d(end = reduce_prod_21, start = range_1d_21_start_0, step = range_1d_21_step_0)[name = tensor("range_1d_21")]; + tensor reshape_105 = reshape(shape = shape_69, x = range_1d_21)[name = tensor("reshape_105")]; + tensor slice_by_index_21 = slice_by_index(begin = concat_74, begin_mask = new_cache_21_internal_tensor_assign_2_begin_mask_0, end = concat_75, end_mask = new_cache_21_internal_tensor_assign_2_end_mask_0, squeeze_mask = new_cache_21_internal_tensor_assign_2_squeeze_mask_0, stride = new_cache_21_internal_tensor_assign_2_stride_0, x = reshape_105)[name = tensor("slice_by_index_21")]; + tensor reshape_106_shape_0 = const()[name = tensor("reshape_106_shape_0"), val = tensor([-1])]; + tensor reshape_106 = reshape(shape = reshape_106_shape_0, x = slice_by_index_21)[name = tensor("reshape_106")]; + tensor reshape_107_shape_0 = const()[name = tensor("reshape_107_shape_0"), val = tensor([-1])]; + tensor reshape_107_cast_fp16 = reshape(shape = reshape_107_shape_0, x = var_4305_cast_fp16)[name = tensor("reshape_107_cast_fp16")]; + tensor reshape_108_shape_0 = const()[name = tensor("reshape_108_shape_0"), val = tensor([-1])]; + tensor reshape_108_cast_fp16 = reshape(shape = reshape_108_shape_0, x = reshape_104_cast_fp16)[name = tensor("reshape_108_cast_fp16")]; + tensor scatter_21_mode_0 = const()[name = tensor("scatter_21_mode_0"), val = tensor("update")]; + tensor scatter_21_axis_0 = const()[name = tensor("scatter_21_axis_0"), val = tensor(0)]; + tensor scatter_21_validate_indices_0 = const()[name = tensor("scatter_21_validate_indices_0"), val = tensor(false)]; + tensor scatter_21_cast_fp16 = scatter(axis = scatter_21_axis_0, data = reshape_108_cast_fp16, indices = reshape_106, mode = scatter_21_mode_0, updates = reshape_107_cast_fp16, validate_indices = scatter_21_validate_indices_0)[name = tensor("scatter_21_cast_fp16")]; + tensor reshape_109_cast_fp16 = reshape(shape = shape_69, x = scatter_21_cast_fp16)[name = tensor("reshape_109_cast_fp16")]; + tensor reshape_109_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("reshape_109_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor keys_61_begin_0 = const()[name = tensor("keys_61_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor keys_61_end_0 = const()[name = tensor("keys_61_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor keys_61_end_mask_0 = const()[name = tensor("keys_61_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor keys_61_squeeze_mask_0 = const()[name = tensor("keys_61_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor keys_61_cast_fp16 = slice_by_index(begin = keys_61_begin_0, end = keys_61_end_0, end_mask = keys_61_end_mask_0, squeeze_mask = keys_61_squeeze_mask_0, x = reshape_109_cast_fp16)[name = tensor("keys_61_cast_fp16")]; + tensor values_61_begin_0 = const()[name = tensor("values_61_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor values_61_end_0 = const()[name = tensor("values_61_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor values_61_end_mask_0 = const()[name = tensor("values_61_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor values_61_squeeze_mask_0 = const()[name = tensor("values_61_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor values_61_cast_fp16 = slice_by_index(begin = values_61_begin_0, end = values_61_end_0, end_mask = values_61_end_mask_0, squeeze_mask = values_61_squeeze_mask_0, x = reshape_109_cast_fp16)[name = tensor("values_61_cast_fp16")]; + tensor var_4317_cast_fp16 = not_equal(x = keys_61_cast_fp16, y = keys_61_cast_fp16)[name = tensor("op_4317_cast_fp16")]; + tensor keys_63_cast_fp16 = select(a = var_463_to_fp16, b = keys_61_cast_fp16, cond = var_4317_cast_fp16)[name = tensor("keys_63_cast_fp16")]; + tensor var_4325_cast_fp16 = not_equal(x = values_61_cast_fp16, y = values_61_cast_fp16)[name = tensor("op_4325_cast_fp16")]; + tensor values_63_cast_fp16 = select(a = var_463_to_fp16, b = values_61_cast_fp16, cond = var_4325_cast_fp16)[name = tensor("values_63_cast_fp16")]; + tensor var_4349 = const()[name = tensor("op_4349"), val = tensor([0, 2, 1, 3])]; + tensor var_4365 = const()[name = tensor("op_4365"), val = tensor([1, 1, 1])]; + tensor var_4366_cast_fp16 = reshape(shape = var_4365, x = position10_to_fp16)[name = tensor("op_4366_cast_fp16")]; + tensor var_4386_to_fp16 = const()[name = tensor("op_4386_to_fp16"), val = tensor(0x1p+0)]; + tensor valid_len_21_cast_fp16 = add(x = var_4366_cast_fp16, y = var_4386_to_fp16)[name = tensor("valid_len_21_cast_fp16")]; + tensor valid_mask_21_cast_fp16 = less(x = k_positions_1_promoted_to_fp16, y = valid_len_21_cast_fp16)[name = tensor("valid_mask_21_cast_fp16")]; + tensor causal_mask_21_cast_fp16 = less_equal(x = k_positions_1_promoted_to_fp16, y = var_4366_cast_fp16)[name = tensor("causal_mask_21_cast_fp16")]; + tensor attn_mask_41 = logical_and(x = valid_mask_21_cast_fp16, y = causal_mask_21_cast_fp16)[name = tensor("attn_mask_41")]; + tensor attn_mask_43_axes_0 = const()[name = tensor("attn_mask_43_axes_0"), val = tensor([1])]; + tensor attn_mask_43 = expand_dims(axes = attn_mask_43_axes_0, x = attn_mask_41)[name = tensor("attn_mask_43")]; + tensor var_4404_transpose_x_0 = const()[name = tensor("op_4404_transpose_x_0"), val = tensor(false)]; + tensor var_4404_transpose_y_0 = const()[name = tensor("op_4404_transpose_y_0"), val = tensor(false)]; + tensor transpose_92_perm_0 = const()[name = tensor("transpose_92_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_93_perm_0 = const()[name = tensor("transpose_93_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_93 = transpose(perm = transpose_93_perm_0, x = keys_63_cast_fp16)[name = tensor("transpose_173")]; + tensor transpose_92 = transpose(perm = transpose_92_perm_0, x = q_63_cast_fp16)[name = tensor("transpose_174")]; + tensor var_4404_cast_fp16 = matmul(transpose_x = var_4404_transpose_x_0, transpose_y = var_4404_transpose_y_0, x = transpose_92, y = transpose_93)[name = tensor("op_4404_cast_fp16")]; + tensor var_4398_to_fp16 = const()[name = tensor("op_4398_to_fp16"), val = tensor([0x1p-3])]; + tensor attn_weights_61_cast_fp16 = mul(x = var_4404_cast_fp16, y = var_4398_to_fp16)[name = tensor("attn_weights_61_cast_fp16")]; + tensor var_4406 = logical_not(x = attn_mask_43)[name = tensor("op_4406")]; + tensor var_4407_to_fp16 = const()[name = tensor("op_4407_to_fp16"), val = tensor(-inf)]; + tensor attn_weights_63_cast_fp16 = select(a = var_4407_to_fp16, b = attn_weights_61_cast_fp16, cond = var_4406)[name = tensor("attn_weights_63_cast_fp16")]; + tensor var_4409 = const()[name = tensor("op_4409"), val = tensor(-1)]; + tensor attn_weights_65_cast_fp16 = softmax(axis = var_4409, x = attn_weights_63_cast_fp16)[name = tensor("attn_weights_65_cast_fp16")]; + tensor attn_output_21_transpose_x_0 = const()[name = tensor("attn_output_21_transpose_x_0"), val = tensor(false)]; + tensor attn_output_21_transpose_y_0 = const()[name = tensor("attn_output_21_transpose_y_0"), val = tensor(false)]; + tensor values_65_cast_fp16 = transpose(perm = var_4349, x = values_63_cast_fp16)[name = tensor("transpose_175")]; + tensor attn_output_21_cast_fp16 = matmul(transpose_x = attn_output_21_transpose_x_0, transpose_y = attn_output_21_transpose_y_0, x = attn_weights_65_cast_fp16, y = values_65_cast_fp16)[name = tensor("attn_output_21_cast_fp16")]; + tensor var_4417 = const()[name = tensor("op_4417"), val = tensor([0, 2, 1, 3])]; + tensor var_4422 = const()[name = tensor("op_4422"), val = tensor([1, 1, 1024])]; + tensor var_4418_cast_fp16 = transpose(perm = var_4417, x = attn_output_21_cast_fp16)[name = tensor("transpose_172")]; + tensor input_127_cast_fp16 = reshape(shape = var_4422, x = var_4418_cast_fp16)[name = tensor("input_127_cast_fp16")]; + tensor attn10_out_proj_weight_to_fp16 = const()[name = tensor("attn10_out_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(259174400)))]; + tensor linear_42_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = attn10_out_proj_weight_to_fp16, x = input_127_cast_fp16)[name = tensor("linear_42_cast_fp16")]; + tensor var_4428_to_fp16 = const()[name = tensor("op_4428_to_fp16"), val = tensor(0x1p+0)]; + tensor var_4429_cast_fp16 = add(x = position10_to_fp16, y = var_4428_to_fp16)[name = tensor("op_4429_cast_fp16")]; + tensor var_4429_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_4429_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor input_129_cast_fp16 = add(x = input_123_cast_fp16, y = linear_42_cast_fp16)[name = tensor("input_129_cast_fp16")]; + tensor input_131_axes_0 = const()[name = tensor("input_131_axes_0"), val = tensor([-1])]; + tensor norm10_2_weight_to_fp16 = const()[name = tensor("norm10_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(261271616)))]; + tensor norm10_2_bias_to_fp16 = const()[name = tensor("norm10_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(261273728)))]; + tensor var_4433_to_fp16 = const()[name = tensor("op_4433_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_131_cast_fp16 = layer_norm(axes = input_131_axes_0, beta = norm10_2_bias_to_fp16, epsilon = var_4433_to_fp16, gamma = norm10_2_weight_to_fp16, x = input_129_cast_fp16)[name = tensor("input_131_cast_fp16")]; + tensor linear10_1_weight_to_fp16 = const()[name = tensor("linear10_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(261275840)))]; + tensor linear_43_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = linear10_1_weight_to_fp16, x = input_131_cast_fp16)[name = tensor("linear_43_cast_fp16")]; + tensor input_133_mode_0 = const()[name = tensor("input_133_mode_0"), val = tensor("EXACT")]; + tensor input_133_cast_fp16 = gelu(mode = input_133_mode_0, x = linear_43_cast_fp16)[name = tensor("input_133_cast_fp16")]; + tensor linear10_2_weight_to_fp16 = const()[name = tensor("linear10_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(269664512)))]; + tensor linear_44_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = linear10_2_weight_to_fp16, x = input_133_cast_fp16)[name = tensor("linear_44_cast_fp16")]; + tensor input_135_cast_fp16 = add(x = input_129_cast_fp16, y = linear_44_cast_fp16)[name = tensor("input_135_cast_fp16")]; + tensor input_137_axes_0 = const()[name = tensor("input_137_axes_0"), val = tensor([-1])]; + tensor norm11_1_weight_to_fp16 = const()[name = tensor("norm11_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(278053184)))]; + tensor norm11_1_bias_to_fp16 = const()[name = tensor("norm11_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(278055296)))]; + tensor var_4450_to_fp16 = const()[name = tensor("op_4450_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_137_cast_fp16 = layer_norm(axes = input_137_axes_0, beta = norm11_1_bias_to_fp16, epsilon = var_4450_to_fp16, gamma = norm11_1_weight_to_fp16, x = input_135_cast_fp16)[name = tensor("input_137_cast_fp16")]; + tensor attn11_in_proj_weight_to_fp16 = const()[name = tensor("attn11_in_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(278057408)))]; + tensor linear_45_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = attn11_in_proj_weight_to_fp16, x = input_137_cast_fp16)[name = tensor("linear_45_cast_fp16")]; + tensor var_4464 = const()[name = tensor("op_4464"), val = tensor([1, 1, 3, 16, 64])]; + tensor qkv_23_cast_fp16 = reshape(shape = var_4464, x = linear_45_cast_fp16)[name = tensor("qkv_23_cast_fp16")]; + tensor q_67_begin_0 = const()[name = tensor("q_67_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor q_67_end_0 = const()[name = tensor("q_67_end_0"), val = tensor([1, 1, 1, 16, 64])]; + tensor q_67_end_mask_0 = const()[name = tensor("q_67_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor q_67_squeeze_mask_0 = const()[name = tensor("q_67_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor q_67_cast_fp16 = slice_by_index(begin = q_67_begin_0, end = q_67_end_0, end_mask = q_67_end_mask_0, squeeze_mask = q_67_squeeze_mask_0, x = qkv_23_cast_fp16)[name = tensor("q_67_cast_fp16")]; + tensor k_45_begin_0 = const()[name = tensor("k_45_begin_0"), val = tensor([0, 0, 1, 0, 0])]; + tensor k_45_end_0 = const()[name = tensor("k_45_end_0"), val = tensor([1, 1, 2, 16, 64])]; + tensor k_45_end_mask_0 = const()[name = tensor("k_45_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor k_45_squeeze_mask_0 = const()[name = tensor("k_45_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor k_45_cast_fp16 = slice_by_index(begin = k_45_begin_0, end = k_45_end_0, end_mask = k_45_end_mask_0, squeeze_mask = k_45_squeeze_mask_0, x = qkv_23_cast_fp16)[name = tensor("k_45_cast_fp16")]; + tensor v_23_begin_0 = const()[name = tensor("v_23_begin_0"), val = tensor([0, 0, 2, 0, 0])]; + tensor v_23_end_0 = const()[name = tensor("v_23_end_0"), val = tensor([1, 1, 3, 16, 64])]; + tensor v_23_end_mask_0 = const()[name = tensor("v_23_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor v_23_squeeze_mask_0 = const()[name = tensor("v_23_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor v_23_cast_fp16 = slice_by_index(begin = v_23_begin_0, end = v_23_end_0, end_mask = v_23_end_mask_0, squeeze_mask = v_23_squeeze_mask_0, x = qkv_23_cast_fp16)[name = tensor("v_23_cast_fp16")]; + tensor var_4530 = const()[name = tensor("op_4530"), val = tensor([1, 1, 1, 1])]; + tensor position11_to_fp16_dtype_0 = const()[name = tensor("position11_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor position11_to_fp16 = cast(dtype = position11_to_fp16_dtype_0, x = position11)[name = tensor("cast_498")]; + tensor ts_71_cast_fp16 = reshape(shape = var_4530, x = position11_to_fp16)[name = tensor("ts_71_cast_fp16")]; + tensor var_4537 = const()[name = tensor("op_4537"), val = tensor([1, 1, 16, 32, 2])]; + tensor q_complex_23_cast_fp16 = reshape(shape = var_4537, x = q_67_cast_fp16)[name = tensor("q_complex_23_cast_fp16")]; + tensor var_4544 = const()[name = tensor("op_4544"), val = tensor([1, 1, 16, 32, 2])]; + tensor k_complex_23_cast_fp16 = reshape(shape = var_4544, x = k_45_cast_fp16)[name = tensor("k_complex_23_cast_fp16")]; + tensor var_4548_begin_0 = const()[name = tensor("op_4548_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_4548_end_0 = const()[name = tensor("op_4548_end_0"), val = tensor([1, 1, 16, 32, 1])]; + tensor var_4548_end_mask_0 = const()[name = tensor("op_4548_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_4548_squeeze_mask_0 = const()[name = tensor("op_4548_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_4548_cast_fp16 = slice_by_index(begin = var_4548_begin_0, end = var_4548_end_0, end_mask = var_4548_end_mask_0, squeeze_mask = var_4548_squeeze_mask_0, x = q_complex_23_cast_fp16)[name = tensor("op_4548_cast_fp16")]; + tensor var_4556_begin_0 = const()[name = tensor("op_4556_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_4556_end_0 = const()[name = tensor("op_4556_end_0"), val = tensor([1, 1, 16, 32, 2])]; + tensor var_4556_end_mask_0 = const()[name = tensor("op_4556_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_4556_squeeze_mask_0 = const()[name = tensor("op_4556_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_4556_cast_fp16 = slice_by_index(begin = var_4556_begin_0, end = var_4556_end_0, end_mask = var_4556_end_mask_0, squeeze_mask = var_4556_squeeze_mask_0, x = q_complex_23_cast_fp16)[name = tensor("op_4556_cast_fp16")]; + tensor var_4564_begin_0 = const()[name = tensor("op_4564_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_4564_end_0 = const()[name = tensor("op_4564_end_0"), val = tensor([1, 1, 16, 32, 1])]; + tensor var_4564_end_mask_0 = const()[name = tensor("op_4564_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_4564_squeeze_mask_0 = const()[name = tensor("op_4564_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_4564_cast_fp16 = slice_by_index(begin = var_4564_begin_0, end = var_4564_end_0, end_mask = var_4564_end_mask_0, squeeze_mask = var_4564_squeeze_mask_0, x = k_complex_23_cast_fp16)[name = tensor("op_4564_cast_fp16")]; + tensor var_4572_begin_0 = const()[name = tensor("op_4572_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_4572_end_0 = const()[name = tensor("op_4572_end_0"), val = tensor([1, 1, 16, 32, 2])]; + tensor var_4572_end_mask_0 = const()[name = tensor("op_4572_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_4572_squeeze_mask_0 = const()[name = tensor("op_4572_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_4572_cast_fp16 = slice_by_index(begin = var_4572_begin_0, end = var_4572_end_0, end_mask = var_4572_end_mask_0, squeeze_mask = var_4572_squeeze_mask_0, x = k_complex_23_cast_fp16)[name = tensor("op_4572_cast_fp16")]; + tensor freqs_23_to_fp16 = const()[name = tensor("freqs_23_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(284348928)))]; + tensor var_4578_cast_fp16 = mul(x = freqs_23_to_fp16, y = ts_71_cast_fp16)[name = tensor("op_4578_cast_fp16")]; + tensor rotr_23_cast_fp16 = cos(x = var_4578_cast_fp16)[name = tensor("rotr_23_cast_fp16")]; + tensor roti_23_cast_fp16 = sin(x = var_4578_cast_fp16)[name = tensor("roti_23_cast_fp16")]; + tensor var_4582_cast_fp16 = mul(x = var_4548_cast_fp16, y = rotr_23_cast_fp16)[name = tensor("op_4582_cast_fp16")]; + tensor var_4583_cast_fp16 = mul(x = var_4556_cast_fp16, y = roti_23_cast_fp16)[name = tensor("op_4583_cast_fp16")]; + tensor qor_45_cast_fp16 = sub(x = var_4582_cast_fp16, y = var_4583_cast_fp16)[name = tensor("qor_45_cast_fp16")]; + tensor var_4586_cast_fp16 = mul(x = var_4548_cast_fp16, y = roti_23_cast_fp16)[name = tensor("op_4586_cast_fp16")]; + tensor var_4587_cast_fp16 = mul(x = var_4556_cast_fp16, y = rotr_23_cast_fp16)[name = tensor("op_4587_cast_fp16")]; + tensor qoi_45_cast_fp16 = add(x = var_4586_cast_fp16, y = var_4587_cast_fp16)[name = tensor("qoi_45_cast_fp16")]; + tensor var_4590_cast_fp16 = mul(x = var_4564_cast_fp16, y = rotr_23_cast_fp16)[name = tensor("op_4590_cast_fp16")]; + tensor var_4591_cast_fp16 = mul(x = var_4572_cast_fp16, y = roti_23_cast_fp16)[name = tensor("op_4591_cast_fp16")]; + tensor kor_45_cast_fp16 = sub(x = var_4590_cast_fp16, y = var_4591_cast_fp16)[name = tensor("kor_45_cast_fp16")]; + tensor var_4594_cast_fp16 = mul(x = var_4564_cast_fp16, y = roti_23_cast_fp16)[name = tensor("op_4594_cast_fp16")]; + tensor var_4595_cast_fp16 = mul(x = var_4572_cast_fp16, y = rotr_23_cast_fp16)[name = tensor("op_4595_cast_fp16")]; + tensor koi_45_cast_fp16 = add(x = var_4594_cast_fp16, y = var_4595_cast_fp16)[name = tensor("koi_45_cast_fp16")]; + tensor qo_23_axis_0 = const()[name = tensor("qo_23_axis_0"), val = tensor(-1)]; + tensor qo_23_cast_fp16 = stack(axis = qo_23_axis_0, values = (qor_45_cast_fp16, qoi_45_cast_fp16))[name = tensor("qo_23_cast_fp16")]; + tensor ko_23_axis_0 = const()[name = tensor("ko_23_axis_0"), val = tensor(-1)]; + tensor ko_23_cast_fp16 = stack(axis = ko_23_axis_0, values = (kor_45_cast_fp16, koi_45_cast_fp16))[name = tensor("ko_23_cast_fp16")]; + tensor var_4628 = const()[name = tensor("op_4628"), val = tensor([1, 1, 16, 64])]; + tensor q_69_cast_fp16 = reshape(shape = var_4628, x = qo_23_cast_fp16)[name = tensor("q_69_cast_fp16")]; + tensor var_4634 = const()[name = tensor("op_4634"), val = tensor([1, 1, 16, 64])]; + tensor k_47_cast_fp16 = reshape(shape = var_4634, x = ko_23_cast_fp16)[name = tensor("k_47_cast_fp16")]; + tensor _inversed_4659_y_0_to_fp16 = const()[name = tensor("_inversed_4659_y_0_to_fp16"), val = tensor(0x1p-9)]; + tensor _inversed_4659_cast_fp16 = mul(x = ts_71_cast_fp16, y = _inversed_4659_y_0_to_fp16)[name = tensor("_inversed_4659_cast_fp16")]; + tensor var_4660_cast_fp16 = floor(x = _inversed_4659_cast_fp16)[name = tensor("op_4660_cast_fp16")]; + tensor var_4661_to_fp16 = const()[name = tensor("op_4661_to_fp16"), val = tensor(0x1p+9)]; + tensor var_4662_cast_fp16 = mul(x = var_4660_cast_fp16, y = var_4661_to_fp16)[name = tensor("op_4662_cast_fp16")]; + tensor write_indices_float_47_cast_fp16 = sub(x = ts_71_cast_fp16, y = var_4662_cast_fp16)[name = tensor("write_indices_float_47_cast_fp16")]; + tensor var_4669_dtype_0 = const()[name = tensor("op_4669_dtype_0"), val = tensor("int32")]; + tensor write_indices_23_reps_0 = const()[name = tensor("write_indices_23_reps_0"), val = tensor([1, 1, 16, 64])]; + tensor write_indices_float_47_cast_fp16_to_int32 = cast(dtype = var_4669_dtype_0, x = write_indices_float_47_cast_fp16)[name = tensor("cast_497")]; + tensor write_indices_23 = tile(reps = write_indices_23_reps_0, x = write_indices_float_47_cast_fp16_to_int32)[name = tensor("write_indices_23")]; + tensor var_4679_begin_0 = const()[name = tensor("op_4679_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_4679_end_0 = const()[name = tensor("op_4679_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor var_4679_end_mask_0 = const()[name = tensor("op_4679_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_4679_squeeze_mask_0 = const()[name = tensor("op_4679_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor cache11_to_fp16_dtype_0 = const()[name = tensor("cache11_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor cache11_to_fp16 = cast(dtype = cache11_to_fp16_dtype_0, x = cache11)[name = tensor("cast_496")]; + tensor var_4679_cast_fp16 = slice_by_index(begin = var_4679_begin_0, end = var_4679_end_0, end_mask = var_4679_end_mask_0, squeeze_mask = var_4679_squeeze_mask_0, x = cache11_to_fp16)[name = tensor("op_4679_cast_fp16")]; + tensor var_4681_axis_0 = const()[name = tensor("op_4681_axis_0"), val = tensor(1)]; + tensor var_4681_mode_0 = const()[name = tensor("op_4681_mode_0"), val = tensor("update")]; + tensor var_4681_validate_indices_0 = const()[name = tensor("op_4681_validate_indices_0"), val = tensor(false)]; + tensor var_4681_cast_fp16 = scatter_along_axis(axis = var_4681_axis_0, data = var_4679_cast_fp16, indices = write_indices_23, mode = var_4681_mode_0, updates = k_47_cast_fp16, validate_indices = var_4681_validate_indices_0)[name = tensor("op_4681_cast_fp16")]; + tensor concat_79 = const()[name = tensor("concat_79"), val = tensor([0, 0, 0, 0, 0])]; + tensor concat_80 = const()[name = tensor("concat_80"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_23_internal_tensor_assign_1_stride_0 = const()[name = tensor("new_cache_23_internal_tensor_assign_1_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_23_internal_tensor_assign_1_begin_mask_0 = const()[name = tensor("new_cache_23_internal_tensor_assign_1_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_23_internal_tensor_assign_1_end_mask_0 = const()[name = tensor("new_cache_23_internal_tensor_assign_1_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_23_internal_tensor_assign_1_squeeze_mask_0 = const()[name = tensor("new_cache_23_internal_tensor_assign_1_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_70 = const()[name = tensor("shape_70"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_22 = const()[name = tensor("reduce_prod_22"), val = tensor(1048576)]; + tensor range_1d_22_start_0 = const()[name = tensor("range_1d_22_start_0"), val = tensor(0)]; + tensor range_1d_22_step_0 = const()[name = tensor("range_1d_22_step_0"), val = tensor(1)]; + tensor range_1d_22 = range_1d(end = reduce_prod_22, start = range_1d_22_start_0, step = range_1d_22_step_0)[name = tensor("range_1d_22")]; + tensor reshape_110 = reshape(shape = shape_70, x = range_1d_22)[name = tensor("reshape_110")]; + tensor slice_by_index_22 = slice_by_index(begin = concat_79, begin_mask = new_cache_23_internal_tensor_assign_1_begin_mask_0, end = concat_80, end_mask = new_cache_23_internal_tensor_assign_1_end_mask_0, squeeze_mask = new_cache_23_internal_tensor_assign_1_squeeze_mask_0, stride = new_cache_23_internal_tensor_assign_1_stride_0, x = reshape_110)[name = tensor("slice_by_index_22")]; + tensor reshape_111_shape_0 = const()[name = tensor("reshape_111_shape_0"), val = tensor([-1])]; + tensor reshape_111 = reshape(shape = reshape_111_shape_0, x = slice_by_index_22)[name = tensor("reshape_111")]; + tensor reshape_112_shape_0 = const()[name = tensor("reshape_112_shape_0"), val = tensor([-1])]; + tensor reshape_112_cast_fp16 = reshape(shape = reshape_112_shape_0, x = var_4681_cast_fp16)[name = tensor("reshape_112_cast_fp16")]; + tensor reshape_113_shape_0 = const()[name = tensor("reshape_113_shape_0"), val = tensor([-1])]; + tensor reshape_113_cast_fp16 = reshape(shape = reshape_113_shape_0, x = cache11_to_fp16)[name = tensor("reshape_113_cast_fp16")]; + tensor scatter_22_mode_0 = const()[name = tensor("scatter_22_mode_0"), val = tensor("update")]; + tensor scatter_22_axis_0 = const()[name = tensor("scatter_22_axis_0"), val = tensor(0)]; + tensor scatter_22_validate_indices_0 = const()[name = tensor("scatter_22_validate_indices_0"), val = tensor(false)]; + tensor scatter_22_cast_fp16 = scatter(axis = scatter_22_axis_0, data = reshape_113_cast_fp16, indices = reshape_111, mode = scatter_22_mode_0, updates = reshape_112_cast_fp16, validate_indices = scatter_22_validate_indices_0)[name = tensor("scatter_22_cast_fp16")]; + tensor reshape_114_cast_fp16 = reshape(shape = shape_70, x = scatter_22_cast_fp16)[name = tensor("reshape_114_cast_fp16")]; + tensor var_4689_begin_0 = const()[name = tensor("op_4689_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor var_4689_end_0 = const()[name = tensor("op_4689_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor var_4689_end_mask_0 = const()[name = tensor("op_4689_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_4689_squeeze_mask_0 = const()[name = tensor("op_4689_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor var_4689_cast_fp16 = slice_by_index(begin = var_4689_begin_0, end = var_4689_end_0, end_mask = var_4689_end_mask_0, squeeze_mask = var_4689_squeeze_mask_0, x = reshape_114_cast_fp16)[name = tensor("op_4689_cast_fp16")]; + tensor var_4691_axis_0 = const()[name = tensor("op_4691_axis_0"), val = tensor(1)]; + tensor var_4691_mode_0 = const()[name = tensor("op_4691_mode_0"), val = tensor("update")]; + tensor var_4691_validate_indices_0 = const()[name = tensor("op_4691_validate_indices_0"), val = tensor(false)]; + tensor var_4691_cast_fp16 = scatter_along_axis(axis = var_4691_axis_0, data = var_4689_cast_fp16, indices = write_indices_23, mode = var_4691_mode_0, updates = v_23_cast_fp16, validate_indices = var_4691_validate_indices_0)[name = tensor("op_4691_cast_fp16")]; + tensor concat_81 = const()[name = tensor("concat_81"), val = tensor([1, 0, 0, 0, 0])]; + tensor concat_82 = const()[name = tensor("concat_82"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_23_internal_tensor_assign_2_stride_0 = const()[name = tensor("new_cache_23_internal_tensor_assign_2_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_23_internal_tensor_assign_2_begin_mask_0 = const()[name = tensor("new_cache_23_internal_tensor_assign_2_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_23_internal_tensor_assign_2_end_mask_0 = const()[name = tensor("new_cache_23_internal_tensor_assign_2_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_23_internal_tensor_assign_2_squeeze_mask_0 = const()[name = tensor("new_cache_23_internal_tensor_assign_2_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_71 = const()[name = tensor("shape_71"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_23 = const()[name = tensor("reduce_prod_23"), val = tensor(1048576)]; + tensor range_1d_23_start_0 = const()[name = tensor("range_1d_23_start_0"), val = tensor(0)]; + tensor range_1d_23_step_0 = const()[name = tensor("range_1d_23_step_0"), val = tensor(1)]; + tensor range_1d_23 = range_1d(end = reduce_prod_23, start = range_1d_23_start_0, step = range_1d_23_step_0)[name = tensor("range_1d_23")]; + tensor reshape_115 = reshape(shape = shape_71, x = range_1d_23)[name = tensor("reshape_115")]; + tensor slice_by_index_23 = slice_by_index(begin = concat_81, begin_mask = new_cache_23_internal_tensor_assign_2_begin_mask_0, end = concat_82, end_mask = new_cache_23_internal_tensor_assign_2_end_mask_0, squeeze_mask = new_cache_23_internal_tensor_assign_2_squeeze_mask_0, stride = new_cache_23_internal_tensor_assign_2_stride_0, x = reshape_115)[name = tensor("slice_by_index_23")]; + tensor reshape_116_shape_0 = const()[name = tensor("reshape_116_shape_0"), val = tensor([-1])]; + tensor reshape_116 = reshape(shape = reshape_116_shape_0, x = slice_by_index_23)[name = tensor("reshape_116")]; + tensor reshape_117_shape_0 = const()[name = tensor("reshape_117_shape_0"), val = tensor([-1])]; + tensor reshape_117_cast_fp16 = reshape(shape = reshape_117_shape_0, x = var_4691_cast_fp16)[name = tensor("reshape_117_cast_fp16")]; + tensor reshape_118_shape_0 = const()[name = tensor("reshape_118_shape_0"), val = tensor([-1])]; + tensor reshape_118_cast_fp16 = reshape(shape = reshape_118_shape_0, x = reshape_114_cast_fp16)[name = tensor("reshape_118_cast_fp16")]; + tensor scatter_23_mode_0 = const()[name = tensor("scatter_23_mode_0"), val = tensor("update")]; + tensor scatter_23_axis_0 = const()[name = tensor("scatter_23_axis_0"), val = tensor(0)]; + tensor scatter_23_validate_indices_0 = const()[name = tensor("scatter_23_validate_indices_0"), val = tensor(false)]; + tensor scatter_23_cast_fp16 = scatter(axis = scatter_23_axis_0, data = reshape_118_cast_fp16, indices = reshape_116, mode = scatter_23_mode_0, updates = reshape_117_cast_fp16, validate_indices = scatter_23_validate_indices_0)[name = tensor("scatter_23_cast_fp16")]; + tensor reshape_119_cast_fp16 = reshape(shape = shape_71, x = scatter_23_cast_fp16)[name = tensor("reshape_119_cast_fp16")]; + tensor reshape_119_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("reshape_119_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor keys_67_begin_0 = const()[name = tensor("keys_67_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor keys_67_end_0 = const()[name = tensor("keys_67_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor keys_67_end_mask_0 = const()[name = tensor("keys_67_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor keys_67_squeeze_mask_0 = const()[name = tensor("keys_67_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor keys_67_cast_fp16 = slice_by_index(begin = keys_67_begin_0, end = keys_67_end_0, end_mask = keys_67_end_mask_0, squeeze_mask = keys_67_squeeze_mask_0, x = reshape_119_cast_fp16)[name = tensor("keys_67_cast_fp16")]; + tensor values_67_begin_0 = const()[name = tensor("values_67_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor values_67_end_0 = const()[name = tensor("values_67_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor values_67_end_mask_0 = const()[name = tensor("values_67_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor values_67_squeeze_mask_0 = const()[name = tensor("values_67_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor values_67_cast_fp16 = slice_by_index(begin = values_67_begin_0, end = values_67_end_0, end_mask = values_67_end_mask_0, squeeze_mask = values_67_squeeze_mask_0, x = reshape_119_cast_fp16)[name = tensor("values_67_cast_fp16")]; + tensor var_4703_cast_fp16 = not_equal(x = keys_67_cast_fp16, y = keys_67_cast_fp16)[name = tensor("op_4703_cast_fp16")]; + tensor keys_69_cast_fp16 = select(a = var_463_to_fp16, b = keys_67_cast_fp16, cond = var_4703_cast_fp16)[name = tensor("keys_69_cast_fp16")]; + tensor var_4711_cast_fp16 = not_equal(x = values_67_cast_fp16, y = values_67_cast_fp16)[name = tensor("op_4711_cast_fp16")]; + tensor values_69_cast_fp16 = select(a = var_463_to_fp16, b = values_67_cast_fp16, cond = var_4711_cast_fp16)[name = tensor("values_69_cast_fp16")]; + tensor var_4735 = const()[name = tensor("op_4735"), val = tensor([0, 2, 1, 3])]; + tensor var_4751 = const()[name = tensor("op_4751"), val = tensor([1, 1, 1])]; + tensor var_4752_cast_fp16 = reshape(shape = var_4751, x = position11_to_fp16)[name = tensor("op_4752_cast_fp16")]; + tensor var_4772_to_fp16 = const()[name = tensor("op_4772_to_fp16"), val = tensor(0x1p+0)]; + tensor valid_len_23_cast_fp16 = add(x = var_4752_cast_fp16, y = var_4772_to_fp16)[name = tensor("valid_len_23_cast_fp16")]; + tensor valid_mask_23_cast_fp16 = less(x = k_positions_1_promoted_to_fp16, y = valid_len_23_cast_fp16)[name = tensor("valid_mask_23_cast_fp16")]; + tensor causal_mask_23_cast_fp16 = less_equal(x = k_positions_1_promoted_to_fp16, y = var_4752_cast_fp16)[name = tensor("causal_mask_23_cast_fp16")]; + tensor attn_mask_45 = logical_and(x = valid_mask_23_cast_fp16, y = causal_mask_23_cast_fp16)[name = tensor("attn_mask_45")]; + tensor attn_mask_47_axes_0 = const()[name = tensor("attn_mask_47_axes_0"), val = tensor([1])]; + tensor attn_mask_47 = expand_dims(axes = attn_mask_47_axes_0, x = attn_mask_45)[name = tensor("attn_mask_47")]; + tensor var_4790_transpose_x_0 = const()[name = tensor("op_4790_transpose_x_0"), val = tensor(false)]; + tensor var_4790_transpose_y_0 = const()[name = tensor("op_4790_transpose_y_0"), val = tensor(false)]; + tensor transpose_94_perm_0 = const()[name = tensor("transpose_94_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_95_perm_0 = const()[name = tensor("transpose_95_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_95 = transpose(perm = transpose_95_perm_0, x = keys_69_cast_fp16)[name = tensor("transpose_169")]; + tensor transpose_94 = transpose(perm = transpose_94_perm_0, x = q_69_cast_fp16)[name = tensor("transpose_170")]; + tensor var_4790_cast_fp16 = matmul(transpose_x = var_4790_transpose_x_0, transpose_y = var_4790_transpose_y_0, x = transpose_94, y = transpose_95)[name = tensor("op_4790_cast_fp16")]; + tensor var_4784_to_fp16 = const()[name = tensor("op_4784_to_fp16"), val = tensor([0x1p-3])]; + tensor attn_weights_67_cast_fp16 = mul(x = var_4790_cast_fp16, y = var_4784_to_fp16)[name = tensor("attn_weights_67_cast_fp16")]; + tensor var_4792 = logical_not(x = attn_mask_47)[name = tensor("op_4792")]; + tensor var_4793_to_fp16 = const()[name = tensor("op_4793_to_fp16"), val = tensor(-inf)]; + tensor attn_weights_69_cast_fp16 = select(a = var_4793_to_fp16, b = attn_weights_67_cast_fp16, cond = var_4792)[name = tensor("attn_weights_69_cast_fp16")]; + tensor var_4795 = const()[name = tensor("op_4795"), val = tensor(-1)]; + tensor attn_weights_71_cast_fp16 = softmax(axis = var_4795, x = attn_weights_69_cast_fp16)[name = tensor("attn_weights_71_cast_fp16")]; + tensor attn_output_23_transpose_x_0 = const()[name = tensor("attn_output_23_transpose_x_0"), val = tensor(false)]; + tensor attn_output_23_transpose_y_0 = const()[name = tensor("attn_output_23_transpose_y_0"), val = tensor(false)]; + tensor values_71_cast_fp16 = transpose(perm = var_4735, x = values_69_cast_fp16)[name = tensor("transpose_171")]; + tensor attn_output_23_cast_fp16 = matmul(transpose_x = attn_output_23_transpose_x_0, transpose_y = attn_output_23_transpose_y_0, x = attn_weights_71_cast_fp16, y = values_71_cast_fp16)[name = tensor("attn_output_23_cast_fp16")]; + tensor var_4803 = const()[name = tensor("op_4803"), val = tensor([0, 2, 1, 3])]; + tensor var_4808 = const()[name = tensor("op_4808"), val = tensor([1, 1, 1024])]; + tensor var_4804_cast_fp16 = transpose(perm = var_4803, x = attn_output_23_cast_fp16)[name = tensor("transpose_168")]; + tensor input_139_cast_fp16 = reshape(shape = var_4808, x = var_4804_cast_fp16)[name = tensor("input_139_cast_fp16")]; + tensor attn11_out_proj_weight_to_fp16 = const()[name = tensor("attn11_out_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(284349056)))]; + tensor linear_46_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = attn11_out_proj_weight_to_fp16, x = input_139_cast_fp16)[name = tensor("linear_46_cast_fp16")]; + tensor var_4814_to_fp16 = const()[name = tensor("op_4814_to_fp16"), val = tensor(0x1p+0)]; + tensor var_4815_cast_fp16 = add(x = position11_to_fp16, y = var_4814_to_fp16)[name = tensor("op_4815_cast_fp16")]; + tensor var_4815_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_4815_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor input_141_cast_fp16 = add(x = input_135_cast_fp16, y = linear_46_cast_fp16)[name = tensor("input_141_cast_fp16")]; + tensor input_143_axes_0 = const()[name = tensor("input_143_axes_0"), val = tensor([-1])]; + tensor norm11_2_weight_to_fp16 = const()[name = tensor("norm11_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(286446272)))]; + tensor norm11_2_bias_to_fp16 = const()[name = tensor("norm11_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(286448384)))]; + tensor var_4819_to_fp16 = const()[name = tensor("op_4819_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_143_cast_fp16 = layer_norm(axes = input_143_axes_0, beta = norm11_2_bias_to_fp16, epsilon = var_4819_to_fp16, gamma = norm11_2_weight_to_fp16, x = input_141_cast_fp16)[name = tensor("input_143_cast_fp16")]; + tensor linear11_1_weight_to_fp16 = const()[name = tensor("linear11_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(286450496)))]; + tensor linear_47_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = linear11_1_weight_to_fp16, x = input_143_cast_fp16)[name = tensor("linear_47_cast_fp16")]; + tensor input_145_mode_0 = const()[name = tensor("input_145_mode_0"), val = tensor("EXACT")]; + tensor input_145_cast_fp16 = gelu(mode = input_145_mode_0, x = linear_47_cast_fp16)[name = tensor("input_145_cast_fp16")]; + tensor linear11_2_weight_to_fp16 = const()[name = tensor("linear11_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(294839168)))]; + tensor linear_48_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = linear11_2_weight_to_fp16, x = input_145_cast_fp16)[name = tensor("linear_48_cast_fp16")]; + tensor input_147_cast_fp16 = add(x = input_141_cast_fp16, y = linear_48_cast_fp16)[name = tensor("input_147_cast_fp16")]; + tensor input_149_axes_0 = const()[name = tensor("input_149_axes_0"), val = tensor([-1])]; + tensor norm12_1_weight_to_fp16 = const()[name = tensor("norm12_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(303227840)))]; + tensor norm12_1_bias_to_fp16 = const()[name = tensor("norm12_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(303229952)))]; + tensor var_4836_to_fp16 = const()[name = tensor("op_4836_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_149_cast_fp16 = layer_norm(axes = input_149_axes_0, beta = norm12_1_bias_to_fp16, epsilon = var_4836_to_fp16, gamma = norm12_1_weight_to_fp16, x = input_147_cast_fp16)[name = tensor("input_149_cast_fp16")]; + tensor attn12_in_proj_weight_to_fp16 = const()[name = tensor("attn12_in_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(303232064)))]; + tensor linear_49_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = attn12_in_proj_weight_to_fp16, x = input_149_cast_fp16)[name = tensor("linear_49_cast_fp16")]; + tensor var_4850 = const()[name = tensor("op_4850"), val = tensor([1, 1, 3, 16, 64])]; + tensor qkv_25_cast_fp16 = reshape(shape = var_4850, x = linear_49_cast_fp16)[name = tensor("qkv_25_cast_fp16")]; + tensor q_73_begin_0 = const()[name = tensor("q_73_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor q_73_end_0 = const()[name = tensor("q_73_end_0"), val = tensor([1, 1, 1, 16, 64])]; + tensor q_73_end_mask_0 = const()[name = tensor("q_73_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor q_73_squeeze_mask_0 = const()[name = tensor("q_73_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor q_73_cast_fp16 = slice_by_index(begin = q_73_begin_0, end = q_73_end_0, end_mask = q_73_end_mask_0, squeeze_mask = q_73_squeeze_mask_0, x = qkv_25_cast_fp16)[name = tensor("q_73_cast_fp16")]; + tensor k_49_begin_0 = const()[name = tensor("k_49_begin_0"), val = tensor([0, 0, 1, 0, 0])]; + tensor k_49_end_0 = const()[name = tensor("k_49_end_0"), val = tensor([1, 1, 2, 16, 64])]; + tensor k_49_end_mask_0 = const()[name = tensor("k_49_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor k_49_squeeze_mask_0 = const()[name = tensor("k_49_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor k_49_cast_fp16 = slice_by_index(begin = k_49_begin_0, end = k_49_end_0, end_mask = k_49_end_mask_0, squeeze_mask = k_49_squeeze_mask_0, x = qkv_25_cast_fp16)[name = tensor("k_49_cast_fp16")]; + tensor v_25_begin_0 = const()[name = tensor("v_25_begin_0"), val = tensor([0, 0, 2, 0, 0])]; + tensor v_25_end_0 = const()[name = tensor("v_25_end_0"), val = tensor([1, 1, 3, 16, 64])]; + tensor v_25_end_mask_0 = const()[name = tensor("v_25_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor v_25_squeeze_mask_0 = const()[name = tensor("v_25_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor v_25_cast_fp16 = slice_by_index(begin = v_25_begin_0, end = v_25_end_0, end_mask = v_25_end_mask_0, squeeze_mask = v_25_squeeze_mask_0, x = qkv_25_cast_fp16)[name = tensor("v_25_cast_fp16")]; + tensor var_4916 = const()[name = tensor("op_4916"), val = tensor([1, 1, 1, 1])]; + tensor position12_to_fp16_dtype_0 = const()[name = tensor("position12_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor position12_to_fp16 = cast(dtype = position12_to_fp16_dtype_0, x = position12)[name = tensor("cast_493")]; + tensor ts_77_cast_fp16 = reshape(shape = var_4916, x = position12_to_fp16)[name = tensor("ts_77_cast_fp16")]; + tensor var_4923 = const()[name = tensor("op_4923"), val = tensor([1, 1, 16, 32, 2])]; + tensor q_complex_25_cast_fp16 = reshape(shape = var_4923, x = q_73_cast_fp16)[name = tensor("q_complex_25_cast_fp16")]; + tensor var_4930 = const()[name = tensor("op_4930"), val = tensor([1, 1, 16, 32, 2])]; + tensor k_complex_25_cast_fp16 = reshape(shape = var_4930, x = k_49_cast_fp16)[name = tensor("k_complex_25_cast_fp16")]; + tensor var_4934_begin_0 = const()[name = tensor("op_4934_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_4934_end_0 = const()[name = tensor("op_4934_end_0"), val = tensor([1, 1, 16, 32, 1])]; + tensor var_4934_end_mask_0 = const()[name = tensor("op_4934_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_4934_squeeze_mask_0 = const()[name = tensor("op_4934_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_4934_cast_fp16 = slice_by_index(begin = var_4934_begin_0, end = var_4934_end_0, end_mask = var_4934_end_mask_0, squeeze_mask = var_4934_squeeze_mask_0, x = q_complex_25_cast_fp16)[name = tensor("op_4934_cast_fp16")]; + tensor var_4942_begin_0 = const()[name = tensor("op_4942_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_4942_end_0 = const()[name = tensor("op_4942_end_0"), val = tensor([1, 1, 16, 32, 2])]; + tensor var_4942_end_mask_0 = const()[name = tensor("op_4942_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_4942_squeeze_mask_0 = const()[name = tensor("op_4942_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_4942_cast_fp16 = slice_by_index(begin = var_4942_begin_0, end = var_4942_end_0, end_mask = var_4942_end_mask_0, squeeze_mask = var_4942_squeeze_mask_0, x = q_complex_25_cast_fp16)[name = tensor("op_4942_cast_fp16")]; + tensor var_4950_begin_0 = const()[name = tensor("op_4950_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_4950_end_0 = const()[name = tensor("op_4950_end_0"), val = tensor([1, 1, 16, 32, 1])]; + tensor var_4950_end_mask_0 = const()[name = tensor("op_4950_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_4950_squeeze_mask_0 = const()[name = tensor("op_4950_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_4950_cast_fp16 = slice_by_index(begin = var_4950_begin_0, end = var_4950_end_0, end_mask = var_4950_end_mask_0, squeeze_mask = var_4950_squeeze_mask_0, x = k_complex_25_cast_fp16)[name = tensor("op_4950_cast_fp16")]; + tensor var_4958_begin_0 = const()[name = tensor("op_4958_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_4958_end_0 = const()[name = tensor("op_4958_end_0"), val = tensor([1, 1, 16, 32, 2])]; + tensor var_4958_end_mask_0 = const()[name = tensor("op_4958_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_4958_squeeze_mask_0 = const()[name = tensor("op_4958_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_4958_cast_fp16 = slice_by_index(begin = var_4958_begin_0, end = var_4958_end_0, end_mask = var_4958_end_mask_0, squeeze_mask = var_4958_squeeze_mask_0, x = k_complex_25_cast_fp16)[name = tensor("op_4958_cast_fp16")]; + tensor freqs_25_to_fp16 = const()[name = tensor("freqs_25_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(309523584)))]; + tensor var_4964_cast_fp16 = mul(x = freqs_25_to_fp16, y = ts_77_cast_fp16)[name = tensor("op_4964_cast_fp16")]; + tensor rotr_25_cast_fp16 = cos(x = var_4964_cast_fp16)[name = tensor("rotr_25_cast_fp16")]; + tensor roti_25_cast_fp16 = sin(x = var_4964_cast_fp16)[name = tensor("roti_25_cast_fp16")]; + tensor var_4968_cast_fp16 = mul(x = var_4934_cast_fp16, y = rotr_25_cast_fp16)[name = tensor("op_4968_cast_fp16")]; + tensor var_4969_cast_fp16 = mul(x = var_4942_cast_fp16, y = roti_25_cast_fp16)[name = tensor("op_4969_cast_fp16")]; + tensor qor_49_cast_fp16 = sub(x = var_4968_cast_fp16, y = var_4969_cast_fp16)[name = tensor("qor_49_cast_fp16")]; + tensor var_4972_cast_fp16 = mul(x = var_4934_cast_fp16, y = roti_25_cast_fp16)[name = tensor("op_4972_cast_fp16")]; + tensor var_4973_cast_fp16 = mul(x = var_4942_cast_fp16, y = rotr_25_cast_fp16)[name = tensor("op_4973_cast_fp16")]; + tensor qoi_49_cast_fp16 = add(x = var_4972_cast_fp16, y = var_4973_cast_fp16)[name = tensor("qoi_49_cast_fp16")]; + tensor var_4976_cast_fp16 = mul(x = var_4950_cast_fp16, y = rotr_25_cast_fp16)[name = tensor("op_4976_cast_fp16")]; + tensor var_4977_cast_fp16 = mul(x = var_4958_cast_fp16, y = roti_25_cast_fp16)[name = tensor("op_4977_cast_fp16")]; + tensor kor_49_cast_fp16 = sub(x = var_4976_cast_fp16, y = var_4977_cast_fp16)[name = tensor("kor_49_cast_fp16")]; + tensor var_4980_cast_fp16 = mul(x = var_4950_cast_fp16, y = roti_25_cast_fp16)[name = tensor("op_4980_cast_fp16")]; + tensor var_4981_cast_fp16 = mul(x = var_4958_cast_fp16, y = rotr_25_cast_fp16)[name = tensor("op_4981_cast_fp16")]; + tensor koi_49_cast_fp16 = add(x = var_4980_cast_fp16, y = var_4981_cast_fp16)[name = tensor("koi_49_cast_fp16")]; + tensor qo_25_axis_0 = const()[name = tensor("qo_25_axis_0"), val = tensor(-1)]; + tensor qo_25_cast_fp16 = stack(axis = qo_25_axis_0, values = (qor_49_cast_fp16, qoi_49_cast_fp16))[name = tensor("qo_25_cast_fp16")]; + tensor ko_25_axis_0 = const()[name = tensor("ko_25_axis_0"), val = tensor(-1)]; + tensor ko_25_cast_fp16 = stack(axis = ko_25_axis_0, values = (kor_49_cast_fp16, koi_49_cast_fp16))[name = tensor("ko_25_cast_fp16")]; + tensor var_5014 = const()[name = tensor("op_5014"), val = tensor([1, 1, 16, 64])]; + tensor q_75_cast_fp16 = reshape(shape = var_5014, x = qo_25_cast_fp16)[name = tensor("q_75_cast_fp16")]; + tensor var_5020 = const()[name = tensor("op_5020"), val = tensor([1, 1, 16, 64])]; + tensor k_51_cast_fp16 = reshape(shape = var_5020, x = ko_25_cast_fp16)[name = tensor("k_51_cast_fp16")]; + tensor _inversed_5045_y_0_to_fp16 = const()[name = tensor("_inversed_5045_y_0_to_fp16"), val = tensor(0x1p-9)]; + tensor _inversed_5045_cast_fp16 = mul(x = ts_77_cast_fp16, y = _inversed_5045_y_0_to_fp16)[name = tensor("_inversed_5045_cast_fp16")]; + tensor var_5046_cast_fp16 = floor(x = _inversed_5045_cast_fp16)[name = tensor("op_5046_cast_fp16")]; + tensor var_5047_to_fp16 = const()[name = tensor("op_5047_to_fp16"), val = tensor(0x1p+9)]; + tensor var_5048_cast_fp16 = mul(x = var_5046_cast_fp16, y = var_5047_to_fp16)[name = tensor("op_5048_cast_fp16")]; + tensor write_indices_float_51_cast_fp16 = sub(x = ts_77_cast_fp16, y = var_5048_cast_fp16)[name = tensor("write_indices_float_51_cast_fp16")]; + tensor var_5055_dtype_0 = const()[name = tensor("op_5055_dtype_0"), val = tensor("int32")]; + tensor write_indices_25_reps_0 = const()[name = tensor("write_indices_25_reps_0"), val = tensor([1, 1, 16, 64])]; + tensor write_indices_float_51_cast_fp16_to_int32 = cast(dtype = var_5055_dtype_0, x = write_indices_float_51_cast_fp16)[name = tensor("cast_492")]; + tensor write_indices_25 = tile(reps = write_indices_25_reps_0, x = write_indices_float_51_cast_fp16_to_int32)[name = tensor("write_indices_25")]; + tensor var_5065_begin_0 = const()[name = tensor("op_5065_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_5065_end_0 = const()[name = tensor("op_5065_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor var_5065_end_mask_0 = const()[name = tensor("op_5065_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_5065_squeeze_mask_0 = const()[name = tensor("op_5065_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor cache12_to_fp16_dtype_0 = const()[name = tensor("cache12_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor cache12_to_fp16 = cast(dtype = cache12_to_fp16_dtype_0, x = cache12)[name = tensor("cast_491")]; + tensor var_5065_cast_fp16 = slice_by_index(begin = var_5065_begin_0, end = var_5065_end_0, end_mask = var_5065_end_mask_0, squeeze_mask = var_5065_squeeze_mask_0, x = cache12_to_fp16)[name = tensor("op_5065_cast_fp16")]; + tensor var_5067_axis_0 = const()[name = tensor("op_5067_axis_0"), val = tensor(1)]; + tensor var_5067_mode_0 = const()[name = tensor("op_5067_mode_0"), val = tensor("update")]; + tensor var_5067_validate_indices_0 = const()[name = tensor("op_5067_validate_indices_0"), val = tensor(false)]; + tensor var_5067_cast_fp16 = scatter_along_axis(axis = var_5067_axis_0, data = var_5065_cast_fp16, indices = write_indices_25, mode = var_5067_mode_0, updates = k_51_cast_fp16, validate_indices = var_5067_validate_indices_0)[name = tensor("op_5067_cast_fp16")]; + tensor concat_86 = const()[name = tensor("concat_86"), val = tensor([0, 0, 0, 0, 0])]; + tensor concat_87 = const()[name = tensor("concat_87"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_25_internal_tensor_assign_1_stride_0 = const()[name = tensor("new_cache_25_internal_tensor_assign_1_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_25_internal_tensor_assign_1_begin_mask_0 = const()[name = tensor("new_cache_25_internal_tensor_assign_1_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_25_internal_tensor_assign_1_end_mask_0 = const()[name = tensor("new_cache_25_internal_tensor_assign_1_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_25_internal_tensor_assign_1_squeeze_mask_0 = const()[name = tensor("new_cache_25_internal_tensor_assign_1_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_72 = const()[name = tensor("shape_72"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_24 = const()[name = tensor("reduce_prod_24"), val = tensor(1048576)]; + tensor range_1d_24_start_0 = const()[name = tensor("range_1d_24_start_0"), val = tensor(0)]; + tensor range_1d_24_step_0 = const()[name = tensor("range_1d_24_step_0"), val = tensor(1)]; + tensor range_1d_24 = range_1d(end = reduce_prod_24, start = range_1d_24_start_0, step = range_1d_24_step_0)[name = tensor("range_1d_24")]; + tensor reshape_120 = reshape(shape = shape_72, x = range_1d_24)[name = tensor("reshape_120")]; + tensor slice_by_index_24 = slice_by_index(begin = concat_86, begin_mask = new_cache_25_internal_tensor_assign_1_begin_mask_0, end = concat_87, end_mask = new_cache_25_internal_tensor_assign_1_end_mask_0, squeeze_mask = new_cache_25_internal_tensor_assign_1_squeeze_mask_0, stride = new_cache_25_internal_tensor_assign_1_stride_0, x = reshape_120)[name = tensor("slice_by_index_24")]; + tensor reshape_121_shape_0 = const()[name = tensor("reshape_121_shape_0"), val = tensor([-1])]; + tensor reshape_121 = reshape(shape = reshape_121_shape_0, x = slice_by_index_24)[name = tensor("reshape_121")]; + tensor reshape_122_shape_0 = const()[name = tensor("reshape_122_shape_0"), val = tensor([-1])]; + tensor reshape_122_cast_fp16 = reshape(shape = reshape_122_shape_0, x = var_5067_cast_fp16)[name = tensor("reshape_122_cast_fp16")]; + tensor reshape_123_shape_0 = const()[name = tensor("reshape_123_shape_0"), val = tensor([-1])]; + tensor reshape_123_cast_fp16 = reshape(shape = reshape_123_shape_0, x = cache12_to_fp16)[name = tensor("reshape_123_cast_fp16")]; + tensor scatter_24_mode_0 = const()[name = tensor("scatter_24_mode_0"), val = tensor("update")]; + tensor scatter_24_axis_0 = const()[name = tensor("scatter_24_axis_0"), val = tensor(0)]; + tensor scatter_24_validate_indices_0 = const()[name = tensor("scatter_24_validate_indices_0"), val = tensor(false)]; + tensor scatter_24_cast_fp16 = scatter(axis = scatter_24_axis_0, data = reshape_123_cast_fp16, indices = reshape_121, mode = scatter_24_mode_0, updates = reshape_122_cast_fp16, validate_indices = scatter_24_validate_indices_0)[name = tensor("scatter_24_cast_fp16")]; + tensor reshape_124_cast_fp16 = reshape(shape = shape_72, x = scatter_24_cast_fp16)[name = tensor("reshape_124_cast_fp16")]; + tensor var_5075_begin_0 = const()[name = tensor("op_5075_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor var_5075_end_0 = const()[name = tensor("op_5075_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor var_5075_end_mask_0 = const()[name = tensor("op_5075_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_5075_squeeze_mask_0 = const()[name = tensor("op_5075_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor var_5075_cast_fp16 = slice_by_index(begin = var_5075_begin_0, end = var_5075_end_0, end_mask = var_5075_end_mask_0, squeeze_mask = var_5075_squeeze_mask_0, x = reshape_124_cast_fp16)[name = tensor("op_5075_cast_fp16")]; + tensor var_5077_axis_0 = const()[name = tensor("op_5077_axis_0"), val = tensor(1)]; + tensor var_5077_mode_0 = const()[name = tensor("op_5077_mode_0"), val = tensor("update")]; + tensor var_5077_validate_indices_0 = const()[name = tensor("op_5077_validate_indices_0"), val = tensor(false)]; + tensor var_5077_cast_fp16 = scatter_along_axis(axis = var_5077_axis_0, data = var_5075_cast_fp16, indices = write_indices_25, mode = var_5077_mode_0, updates = v_25_cast_fp16, validate_indices = var_5077_validate_indices_0)[name = tensor("op_5077_cast_fp16")]; + tensor concat_88 = const()[name = tensor("concat_88"), val = tensor([1, 0, 0, 0, 0])]; + tensor concat_89 = const()[name = tensor("concat_89"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_25_internal_tensor_assign_2_stride_0 = const()[name = tensor("new_cache_25_internal_tensor_assign_2_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_25_internal_tensor_assign_2_begin_mask_0 = const()[name = tensor("new_cache_25_internal_tensor_assign_2_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_25_internal_tensor_assign_2_end_mask_0 = const()[name = tensor("new_cache_25_internal_tensor_assign_2_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_25_internal_tensor_assign_2_squeeze_mask_0 = const()[name = tensor("new_cache_25_internal_tensor_assign_2_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_73 = const()[name = tensor("shape_73"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_25 = const()[name = tensor("reduce_prod_25"), val = tensor(1048576)]; + tensor range_1d_25_start_0 = const()[name = tensor("range_1d_25_start_0"), val = tensor(0)]; + tensor range_1d_25_step_0 = const()[name = tensor("range_1d_25_step_0"), val = tensor(1)]; + tensor range_1d_25 = range_1d(end = reduce_prod_25, start = range_1d_25_start_0, step = range_1d_25_step_0)[name = tensor("range_1d_25")]; + tensor reshape_125 = reshape(shape = shape_73, x = range_1d_25)[name = tensor("reshape_125")]; + tensor slice_by_index_25 = slice_by_index(begin = concat_88, begin_mask = new_cache_25_internal_tensor_assign_2_begin_mask_0, end = concat_89, end_mask = new_cache_25_internal_tensor_assign_2_end_mask_0, squeeze_mask = new_cache_25_internal_tensor_assign_2_squeeze_mask_0, stride = new_cache_25_internal_tensor_assign_2_stride_0, x = reshape_125)[name = tensor("slice_by_index_25")]; + tensor reshape_126_shape_0 = const()[name = tensor("reshape_126_shape_0"), val = tensor([-1])]; + tensor reshape_126 = reshape(shape = reshape_126_shape_0, x = slice_by_index_25)[name = tensor("reshape_126")]; + tensor reshape_127_shape_0 = const()[name = tensor("reshape_127_shape_0"), val = tensor([-1])]; + tensor reshape_127_cast_fp16 = reshape(shape = reshape_127_shape_0, x = var_5077_cast_fp16)[name = tensor("reshape_127_cast_fp16")]; + tensor reshape_128_shape_0 = const()[name = tensor("reshape_128_shape_0"), val = tensor([-1])]; + tensor reshape_128_cast_fp16 = reshape(shape = reshape_128_shape_0, x = reshape_124_cast_fp16)[name = tensor("reshape_128_cast_fp16")]; + tensor scatter_25_mode_0 = const()[name = tensor("scatter_25_mode_0"), val = tensor("update")]; + tensor scatter_25_axis_0 = const()[name = tensor("scatter_25_axis_0"), val = tensor(0)]; + tensor scatter_25_validate_indices_0 = const()[name = tensor("scatter_25_validate_indices_0"), val = tensor(false)]; + tensor scatter_25_cast_fp16 = scatter(axis = scatter_25_axis_0, data = reshape_128_cast_fp16, indices = reshape_126, mode = scatter_25_mode_0, updates = reshape_127_cast_fp16, validate_indices = scatter_25_validate_indices_0)[name = tensor("scatter_25_cast_fp16")]; + tensor reshape_129_cast_fp16 = reshape(shape = shape_73, x = scatter_25_cast_fp16)[name = tensor("reshape_129_cast_fp16")]; + tensor reshape_129_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("reshape_129_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor keys_73_begin_0 = const()[name = tensor("keys_73_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor keys_73_end_0 = const()[name = tensor("keys_73_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor keys_73_end_mask_0 = const()[name = tensor("keys_73_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor keys_73_squeeze_mask_0 = const()[name = tensor("keys_73_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor keys_73_cast_fp16 = slice_by_index(begin = keys_73_begin_0, end = keys_73_end_0, end_mask = keys_73_end_mask_0, squeeze_mask = keys_73_squeeze_mask_0, x = reshape_129_cast_fp16)[name = tensor("keys_73_cast_fp16")]; + tensor values_73_begin_0 = const()[name = tensor("values_73_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor values_73_end_0 = const()[name = tensor("values_73_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor values_73_end_mask_0 = const()[name = tensor("values_73_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor values_73_squeeze_mask_0 = const()[name = tensor("values_73_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor values_73_cast_fp16 = slice_by_index(begin = values_73_begin_0, end = values_73_end_0, end_mask = values_73_end_mask_0, squeeze_mask = values_73_squeeze_mask_0, x = reshape_129_cast_fp16)[name = tensor("values_73_cast_fp16")]; + tensor var_5089_cast_fp16 = not_equal(x = keys_73_cast_fp16, y = keys_73_cast_fp16)[name = tensor("op_5089_cast_fp16")]; + tensor keys_75_cast_fp16 = select(a = var_463_to_fp16, b = keys_73_cast_fp16, cond = var_5089_cast_fp16)[name = tensor("keys_75_cast_fp16")]; + tensor var_5097_cast_fp16 = not_equal(x = values_73_cast_fp16, y = values_73_cast_fp16)[name = tensor("op_5097_cast_fp16")]; + tensor values_75_cast_fp16 = select(a = var_463_to_fp16, b = values_73_cast_fp16, cond = var_5097_cast_fp16)[name = tensor("values_75_cast_fp16")]; + tensor var_5121 = const()[name = tensor("op_5121"), val = tensor([0, 2, 1, 3])]; + tensor var_5137 = const()[name = tensor("op_5137"), val = tensor([1, 1, 1])]; + tensor var_5138_cast_fp16 = reshape(shape = var_5137, x = position12_to_fp16)[name = tensor("op_5138_cast_fp16")]; + tensor var_5158_to_fp16 = const()[name = tensor("op_5158_to_fp16"), val = tensor(0x1p+0)]; + tensor valid_len_25_cast_fp16 = add(x = var_5138_cast_fp16, y = var_5158_to_fp16)[name = tensor("valid_len_25_cast_fp16")]; + tensor valid_mask_25_cast_fp16 = less(x = k_positions_1_promoted_to_fp16, y = valid_len_25_cast_fp16)[name = tensor("valid_mask_25_cast_fp16")]; + tensor causal_mask_25_cast_fp16 = less_equal(x = k_positions_1_promoted_to_fp16, y = var_5138_cast_fp16)[name = tensor("causal_mask_25_cast_fp16")]; + tensor attn_mask_49 = logical_and(x = valid_mask_25_cast_fp16, y = causal_mask_25_cast_fp16)[name = tensor("attn_mask_49")]; + tensor attn_mask_51_axes_0 = const()[name = tensor("attn_mask_51_axes_0"), val = tensor([1])]; + tensor attn_mask_51 = expand_dims(axes = attn_mask_51_axes_0, x = attn_mask_49)[name = tensor("attn_mask_51")]; + tensor var_5176_transpose_x_0 = const()[name = tensor("op_5176_transpose_x_0"), val = tensor(false)]; + tensor var_5176_transpose_y_0 = const()[name = tensor("op_5176_transpose_y_0"), val = tensor(false)]; + tensor transpose_96_perm_0 = const()[name = tensor("transpose_96_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_97_perm_0 = const()[name = tensor("transpose_97_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_97 = transpose(perm = transpose_97_perm_0, x = keys_75_cast_fp16)[name = tensor("transpose_165")]; + tensor transpose_96 = transpose(perm = transpose_96_perm_0, x = q_75_cast_fp16)[name = tensor("transpose_166")]; + tensor var_5176_cast_fp16 = matmul(transpose_x = var_5176_transpose_x_0, transpose_y = var_5176_transpose_y_0, x = transpose_96, y = transpose_97)[name = tensor("op_5176_cast_fp16")]; + tensor var_5170_to_fp16 = const()[name = tensor("op_5170_to_fp16"), val = tensor([0x1p-3])]; + tensor attn_weights_73_cast_fp16 = mul(x = var_5176_cast_fp16, y = var_5170_to_fp16)[name = tensor("attn_weights_73_cast_fp16")]; + tensor var_5178 = logical_not(x = attn_mask_51)[name = tensor("op_5178")]; + tensor var_5179_to_fp16 = const()[name = tensor("op_5179_to_fp16"), val = tensor(-inf)]; + tensor attn_weights_75_cast_fp16 = select(a = var_5179_to_fp16, b = attn_weights_73_cast_fp16, cond = var_5178)[name = tensor("attn_weights_75_cast_fp16")]; + tensor var_5181 = const()[name = tensor("op_5181"), val = tensor(-1)]; + tensor attn_weights_77_cast_fp16 = softmax(axis = var_5181, x = attn_weights_75_cast_fp16)[name = tensor("attn_weights_77_cast_fp16")]; + tensor attn_output_25_transpose_x_0 = const()[name = tensor("attn_output_25_transpose_x_0"), val = tensor(false)]; + tensor attn_output_25_transpose_y_0 = const()[name = tensor("attn_output_25_transpose_y_0"), val = tensor(false)]; + tensor values_77_cast_fp16 = transpose(perm = var_5121, x = values_75_cast_fp16)[name = tensor("transpose_167")]; + tensor attn_output_25_cast_fp16 = matmul(transpose_x = attn_output_25_transpose_x_0, transpose_y = attn_output_25_transpose_y_0, x = attn_weights_77_cast_fp16, y = values_77_cast_fp16)[name = tensor("attn_output_25_cast_fp16")]; + tensor var_5189 = const()[name = tensor("op_5189"), val = tensor([0, 2, 1, 3])]; + tensor var_5194 = const()[name = tensor("op_5194"), val = tensor([1, 1, 1024])]; + tensor var_5190_cast_fp16 = transpose(perm = var_5189, x = attn_output_25_cast_fp16)[name = tensor("transpose_164")]; + tensor input_151_cast_fp16 = reshape(shape = var_5194, x = var_5190_cast_fp16)[name = tensor("input_151_cast_fp16")]; + tensor attn12_out_proj_weight_to_fp16 = const()[name = tensor("attn12_out_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(309523712)))]; + tensor linear_50_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = attn12_out_proj_weight_to_fp16, x = input_151_cast_fp16)[name = tensor("linear_50_cast_fp16")]; + tensor var_5200_to_fp16 = const()[name = tensor("op_5200_to_fp16"), val = tensor(0x1p+0)]; + tensor var_5201_cast_fp16 = add(x = position12_to_fp16, y = var_5200_to_fp16)[name = tensor("op_5201_cast_fp16")]; + tensor var_5201_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_5201_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor input_153_cast_fp16 = add(x = input_147_cast_fp16, y = linear_50_cast_fp16)[name = tensor("input_153_cast_fp16")]; + tensor input_155_axes_0 = const()[name = tensor("input_155_axes_0"), val = tensor([-1])]; + tensor norm12_2_weight_to_fp16 = const()[name = tensor("norm12_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(311620928)))]; + tensor norm12_2_bias_to_fp16 = const()[name = tensor("norm12_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(311623040)))]; + tensor var_5205_to_fp16 = const()[name = tensor("op_5205_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_155_cast_fp16 = layer_norm(axes = input_155_axes_0, beta = norm12_2_bias_to_fp16, epsilon = var_5205_to_fp16, gamma = norm12_2_weight_to_fp16, x = input_153_cast_fp16)[name = tensor("input_155_cast_fp16")]; + tensor linear12_1_weight_to_fp16 = const()[name = tensor("linear12_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(311625152)))]; + tensor linear_51_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = linear12_1_weight_to_fp16, x = input_155_cast_fp16)[name = tensor("linear_51_cast_fp16")]; + tensor input_157_mode_0 = const()[name = tensor("input_157_mode_0"), val = tensor("EXACT")]; + tensor input_157_cast_fp16 = gelu(mode = input_157_mode_0, x = linear_51_cast_fp16)[name = tensor("input_157_cast_fp16")]; + tensor linear12_2_weight_to_fp16 = const()[name = tensor("linear12_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(320013824)))]; + tensor linear_52_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = linear12_2_weight_to_fp16, x = input_157_cast_fp16)[name = tensor("linear_52_cast_fp16")]; + tensor input_159_cast_fp16 = add(x = input_153_cast_fp16, y = linear_52_cast_fp16)[name = tensor("input_159_cast_fp16")]; + tensor input_161_axes_0 = const()[name = tensor("input_161_axes_0"), val = tensor([-1])]; + tensor norm13_1_weight_to_fp16 = const()[name = tensor("norm13_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(328402496)))]; + tensor norm13_1_bias_to_fp16 = const()[name = tensor("norm13_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(328404608)))]; + tensor var_5222_to_fp16 = const()[name = tensor("op_5222_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_161_cast_fp16 = layer_norm(axes = input_161_axes_0, beta = norm13_1_bias_to_fp16, epsilon = var_5222_to_fp16, gamma = norm13_1_weight_to_fp16, x = input_159_cast_fp16)[name = tensor("input_161_cast_fp16")]; + tensor attn13_in_proj_weight_to_fp16 = const()[name = tensor("attn13_in_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(328406720)))]; + tensor linear_53_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = attn13_in_proj_weight_to_fp16, x = input_161_cast_fp16)[name = tensor("linear_53_cast_fp16")]; + tensor var_5236 = const()[name = tensor("op_5236"), val = tensor([1, 1, 3, 16, 64])]; + tensor qkv_27_cast_fp16 = reshape(shape = var_5236, x = linear_53_cast_fp16)[name = tensor("qkv_27_cast_fp16")]; + tensor q_79_begin_0 = const()[name = tensor("q_79_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor q_79_end_0 = const()[name = tensor("q_79_end_0"), val = tensor([1, 1, 1, 16, 64])]; + tensor q_79_end_mask_0 = const()[name = tensor("q_79_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor q_79_squeeze_mask_0 = const()[name = tensor("q_79_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor q_79_cast_fp16 = slice_by_index(begin = q_79_begin_0, end = q_79_end_0, end_mask = q_79_end_mask_0, squeeze_mask = q_79_squeeze_mask_0, x = qkv_27_cast_fp16)[name = tensor("q_79_cast_fp16")]; + tensor k_53_begin_0 = const()[name = tensor("k_53_begin_0"), val = tensor([0, 0, 1, 0, 0])]; + tensor k_53_end_0 = const()[name = tensor("k_53_end_0"), val = tensor([1, 1, 2, 16, 64])]; + tensor k_53_end_mask_0 = const()[name = tensor("k_53_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor k_53_squeeze_mask_0 = const()[name = tensor("k_53_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor k_53_cast_fp16 = slice_by_index(begin = k_53_begin_0, end = k_53_end_0, end_mask = k_53_end_mask_0, squeeze_mask = k_53_squeeze_mask_0, x = qkv_27_cast_fp16)[name = tensor("k_53_cast_fp16")]; + tensor v_27_begin_0 = const()[name = tensor("v_27_begin_0"), val = tensor([0, 0, 2, 0, 0])]; + tensor v_27_end_0 = const()[name = tensor("v_27_end_0"), val = tensor([1, 1, 3, 16, 64])]; + tensor v_27_end_mask_0 = const()[name = tensor("v_27_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor v_27_squeeze_mask_0 = const()[name = tensor("v_27_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor v_27_cast_fp16 = slice_by_index(begin = v_27_begin_0, end = v_27_end_0, end_mask = v_27_end_mask_0, squeeze_mask = v_27_squeeze_mask_0, x = qkv_27_cast_fp16)[name = tensor("v_27_cast_fp16")]; + tensor var_5302 = const()[name = tensor("op_5302"), val = tensor([1, 1, 1, 1])]; + tensor position13_to_fp16_dtype_0 = const()[name = tensor("position13_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor position13_to_fp16 = cast(dtype = position13_to_fp16_dtype_0, x = position13)[name = tensor("cast_488")]; + tensor ts_83_cast_fp16 = reshape(shape = var_5302, x = position13_to_fp16)[name = tensor("ts_83_cast_fp16")]; + tensor var_5309 = const()[name = tensor("op_5309"), val = tensor([1, 1, 16, 32, 2])]; + tensor q_complex_27_cast_fp16 = reshape(shape = var_5309, x = q_79_cast_fp16)[name = tensor("q_complex_27_cast_fp16")]; + tensor var_5316 = const()[name = tensor("op_5316"), val = tensor([1, 1, 16, 32, 2])]; + tensor k_complex_27_cast_fp16 = reshape(shape = var_5316, x = k_53_cast_fp16)[name = tensor("k_complex_27_cast_fp16")]; + tensor var_5320_begin_0 = const()[name = tensor("op_5320_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_5320_end_0 = const()[name = tensor("op_5320_end_0"), val = tensor([1, 1, 16, 32, 1])]; + tensor var_5320_end_mask_0 = const()[name = tensor("op_5320_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_5320_squeeze_mask_0 = const()[name = tensor("op_5320_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_5320_cast_fp16 = slice_by_index(begin = var_5320_begin_0, end = var_5320_end_0, end_mask = var_5320_end_mask_0, squeeze_mask = var_5320_squeeze_mask_0, x = q_complex_27_cast_fp16)[name = tensor("op_5320_cast_fp16")]; + tensor var_5328_begin_0 = const()[name = tensor("op_5328_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_5328_end_0 = const()[name = tensor("op_5328_end_0"), val = tensor([1, 1, 16, 32, 2])]; + tensor var_5328_end_mask_0 = const()[name = tensor("op_5328_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_5328_squeeze_mask_0 = const()[name = tensor("op_5328_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_5328_cast_fp16 = slice_by_index(begin = var_5328_begin_0, end = var_5328_end_0, end_mask = var_5328_end_mask_0, squeeze_mask = var_5328_squeeze_mask_0, x = q_complex_27_cast_fp16)[name = tensor("op_5328_cast_fp16")]; + tensor var_5336_begin_0 = const()[name = tensor("op_5336_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_5336_end_0 = const()[name = tensor("op_5336_end_0"), val = tensor([1, 1, 16, 32, 1])]; + tensor var_5336_end_mask_0 = const()[name = tensor("op_5336_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_5336_squeeze_mask_0 = const()[name = tensor("op_5336_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_5336_cast_fp16 = slice_by_index(begin = var_5336_begin_0, end = var_5336_end_0, end_mask = var_5336_end_mask_0, squeeze_mask = var_5336_squeeze_mask_0, x = k_complex_27_cast_fp16)[name = tensor("op_5336_cast_fp16")]; + tensor var_5344_begin_0 = const()[name = tensor("op_5344_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_5344_end_0 = const()[name = tensor("op_5344_end_0"), val = tensor([1, 1, 16, 32, 2])]; + tensor var_5344_end_mask_0 = const()[name = tensor("op_5344_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_5344_squeeze_mask_0 = const()[name = tensor("op_5344_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_5344_cast_fp16 = slice_by_index(begin = var_5344_begin_0, end = var_5344_end_0, end_mask = var_5344_end_mask_0, squeeze_mask = var_5344_squeeze_mask_0, x = k_complex_27_cast_fp16)[name = tensor("op_5344_cast_fp16")]; + tensor freqs_27_to_fp16 = const()[name = tensor("freqs_27_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(334698240)))]; + tensor var_5350_cast_fp16 = mul(x = freqs_27_to_fp16, y = ts_83_cast_fp16)[name = tensor("op_5350_cast_fp16")]; + tensor rotr_27_cast_fp16 = cos(x = var_5350_cast_fp16)[name = tensor("rotr_27_cast_fp16")]; + tensor roti_27_cast_fp16 = sin(x = var_5350_cast_fp16)[name = tensor("roti_27_cast_fp16")]; + tensor var_5354_cast_fp16 = mul(x = var_5320_cast_fp16, y = rotr_27_cast_fp16)[name = tensor("op_5354_cast_fp16")]; + tensor var_5355_cast_fp16 = mul(x = var_5328_cast_fp16, y = roti_27_cast_fp16)[name = tensor("op_5355_cast_fp16")]; + tensor qor_53_cast_fp16 = sub(x = var_5354_cast_fp16, y = var_5355_cast_fp16)[name = tensor("qor_53_cast_fp16")]; + tensor var_5358_cast_fp16 = mul(x = var_5320_cast_fp16, y = roti_27_cast_fp16)[name = tensor("op_5358_cast_fp16")]; + tensor var_5359_cast_fp16 = mul(x = var_5328_cast_fp16, y = rotr_27_cast_fp16)[name = tensor("op_5359_cast_fp16")]; + tensor qoi_53_cast_fp16 = add(x = var_5358_cast_fp16, y = var_5359_cast_fp16)[name = tensor("qoi_53_cast_fp16")]; + tensor var_5362_cast_fp16 = mul(x = var_5336_cast_fp16, y = rotr_27_cast_fp16)[name = tensor("op_5362_cast_fp16")]; + tensor var_5363_cast_fp16 = mul(x = var_5344_cast_fp16, y = roti_27_cast_fp16)[name = tensor("op_5363_cast_fp16")]; + tensor kor_53_cast_fp16 = sub(x = var_5362_cast_fp16, y = var_5363_cast_fp16)[name = tensor("kor_53_cast_fp16")]; + tensor var_5366_cast_fp16 = mul(x = var_5336_cast_fp16, y = roti_27_cast_fp16)[name = tensor("op_5366_cast_fp16")]; + tensor var_5367_cast_fp16 = mul(x = var_5344_cast_fp16, y = rotr_27_cast_fp16)[name = tensor("op_5367_cast_fp16")]; + tensor koi_53_cast_fp16 = add(x = var_5366_cast_fp16, y = var_5367_cast_fp16)[name = tensor("koi_53_cast_fp16")]; + tensor qo_27_axis_0 = const()[name = tensor("qo_27_axis_0"), val = tensor(-1)]; + tensor qo_27_cast_fp16 = stack(axis = qo_27_axis_0, values = (qor_53_cast_fp16, qoi_53_cast_fp16))[name = tensor("qo_27_cast_fp16")]; + tensor ko_27_axis_0 = const()[name = tensor("ko_27_axis_0"), val = tensor(-1)]; + tensor ko_27_cast_fp16 = stack(axis = ko_27_axis_0, values = (kor_53_cast_fp16, koi_53_cast_fp16))[name = tensor("ko_27_cast_fp16")]; + tensor var_5400 = const()[name = tensor("op_5400"), val = tensor([1, 1, 16, 64])]; + tensor q_81_cast_fp16 = reshape(shape = var_5400, x = qo_27_cast_fp16)[name = tensor("q_81_cast_fp16")]; + tensor var_5406 = const()[name = tensor("op_5406"), val = tensor([1, 1, 16, 64])]; + tensor k_55_cast_fp16 = reshape(shape = var_5406, x = ko_27_cast_fp16)[name = tensor("k_55_cast_fp16")]; + tensor _inversed_5431_y_0_to_fp16 = const()[name = tensor("_inversed_5431_y_0_to_fp16"), val = tensor(0x1p-9)]; + tensor _inversed_5431_cast_fp16 = mul(x = ts_83_cast_fp16, y = _inversed_5431_y_0_to_fp16)[name = tensor("_inversed_5431_cast_fp16")]; + tensor var_5432_cast_fp16 = floor(x = _inversed_5431_cast_fp16)[name = tensor("op_5432_cast_fp16")]; + tensor var_5433_to_fp16 = const()[name = tensor("op_5433_to_fp16"), val = tensor(0x1p+9)]; + tensor var_5434_cast_fp16 = mul(x = var_5432_cast_fp16, y = var_5433_to_fp16)[name = tensor("op_5434_cast_fp16")]; + tensor write_indices_float_55_cast_fp16 = sub(x = ts_83_cast_fp16, y = var_5434_cast_fp16)[name = tensor("write_indices_float_55_cast_fp16")]; + tensor var_5441_dtype_0 = const()[name = tensor("op_5441_dtype_0"), val = tensor("int32")]; + tensor write_indices_27_reps_0 = const()[name = tensor("write_indices_27_reps_0"), val = tensor([1, 1, 16, 64])]; + tensor write_indices_float_55_cast_fp16_to_int32 = cast(dtype = var_5441_dtype_0, x = write_indices_float_55_cast_fp16)[name = tensor("cast_487")]; + tensor write_indices_27 = tile(reps = write_indices_27_reps_0, x = write_indices_float_55_cast_fp16_to_int32)[name = tensor("write_indices_27")]; + tensor var_5451_begin_0 = const()[name = tensor("op_5451_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_5451_end_0 = const()[name = tensor("op_5451_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor var_5451_end_mask_0 = const()[name = tensor("op_5451_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_5451_squeeze_mask_0 = const()[name = tensor("op_5451_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor cache13_to_fp16_dtype_0 = const()[name = tensor("cache13_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor cache13_to_fp16 = cast(dtype = cache13_to_fp16_dtype_0, x = cache13)[name = tensor("cast_486")]; + tensor var_5451_cast_fp16 = slice_by_index(begin = var_5451_begin_0, end = var_5451_end_0, end_mask = var_5451_end_mask_0, squeeze_mask = var_5451_squeeze_mask_0, x = cache13_to_fp16)[name = tensor("op_5451_cast_fp16")]; + tensor var_5453_axis_0 = const()[name = tensor("op_5453_axis_0"), val = tensor(1)]; + tensor var_5453_mode_0 = const()[name = tensor("op_5453_mode_0"), val = tensor("update")]; + tensor var_5453_validate_indices_0 = const()[name = tensor("op_5453_validate_indices_0"), val = tensor(false)]; + tensor var_5453_cast_fp16 = scatter_along_axis(axis = var_5453_axis_0, data = var_5451_cast_fp16, indices = write_indices_27, mode = var_5453_mode_0, updates = k_55_cast_fp16, validate_indices = var_5453_validate_indices_0)[name = tensor("op_5453_cast_fp16")]; + tensor concat_93 = const()[name = tensor("concat_93"), val = tensor([0, 0, 0, 0, 0])]; + tensor concat_94 = const()[name = tensor("concat_94"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_27_internal_tensor_assign_1_stride_0 = const()[name = tensor("new_cache_27_internal_tensor_assign_1_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_27_internal_tensor_assign_1_begin_mask_0 = const()[name = tensor("new_cache_27_internal_tensor_assign_1_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_27_internal_tensor_assign_1_end_mask_0 = const()[name = tensor("new_cache_27_internal_tensor_assign_1_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_27_internal_tensor_assign_1_squeeze_mask_0 = const()[name = tensor("new_cache_27_internal_tensor_assign_1_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_74 = const()[name = tensor("shape_74"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_26 = const()[name = tensor("reduce_prod_26"), val = tensor(1048576)]; + tensor range_1d_26_start_0 = const()[name = tensor("range_1d_26_start_0"), val = tensor(0)]; + tensor range_1d_26_step_0 = const()[name = tensor("range_1d_26_step_0"), val = tensor(1)]; + tensor range_1d_26 = range_1d(end = reduce_prod_26, start = range_1d_26_start_0, step = range_1d_26_step_0)[name = tensor("range_1d_26")]; + tensor reshape_130 = reshape(shape = shape_74, x = range_1d_26)[name = tensor("reshape_130")]; + tensor slice_by_index_26 = slice_by_index(begin = concat_93, begin_mask = new_cache_27_internal_tensor_assign_1_begin_mask_0, end = concat_94, end_mask = new_cache_27_internal_tensor_assign_1_end_mask_0, squeeze_mask = new_cache_27_internal_tensor_assign_1_squeeze_mask_0, stride = new_cache_27_internal_tensor_assign_1_stride_0, x = reshape_130)[name = tensor("slice_by_index_26")]; + tensor reshape_131_shape_0 = const()[name = tensor("reshape_131_shape_0"), val = tensor([-1])]; + tensor reshape_131 = reshape(shape = reshape_131_shape_0, x = slice_by_index_26)[name = tensor("reshape_131")]; + tensor reshape_132_shape_0 = const()[name = tensor("reshape_132_shape_0"), val = tensor([-1])]; + tensor reshape_132_cast_fp16 = reshape(shape = reshape_132_shape_0, x = var_5453_cast_fp16)[name = tensor("reshape_132_cast_fp16")]; + tensor reshape_133_shape_0 = const()[name = tensor("reshape_133_shape_0"), val = tensor([-1])]; + tensor reshape_133_cast_fp16 = reshape(shape = reshape_133_shape_0, x = cache13_to_fp16)[name = tensor("reshape_133_cast_fp16")]; + tensor scatter_26_mode_0 = const()[name = tensor("scatter_26_mode_0"), val = tensor("update")]; + tensor scatter_26_axis_0 = const()[name = tensor("scatter_26_axis_0"), val = tensor(0)]; + tensor scatter_26_validate_indices_0 = const()[name = tensor("scatter_26_validate_indices_0"), val = tensor(false)]; + tensor scatter_26_cast_fp16 = scatter(axis = scatter_26_axis_0, data = reshape_133_cast_fp16, indices = reshape_131, mode = scatter_26_mode_0, updates = reshape_132_cast_fp16, validate_indices = scatter_26_validate_indices_0)[name = tensor("scatter_26_cast_fp16")]; + tensor reshape_134_cast_fp16 = reshape(shape = shape_74, x = scatter_26_cast_fp16)[name = tensor("reshape_134_cast_fp16")]; + tensor var_5461_begin_0 = const()[name = tensor("op_5461_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor var_5461_end_0 = const()[name = tensor("op_5461_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor var_5461_end_mask_0 = const()[name = tensor("op_5461_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_5461_squeeze_mask_0 = const()[name = tensor("op_5461_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor var_5461_cast_fp16 = slice_by_index(begin = var_5461_begin_0, end = var_5461_end_0, end_mask = var_5461_end_mask_0, squeeze_mask = var_5461_squeeze_mask_0, x = reshape_134_cast_fp16)[name = tensor("op_5461_cast_fp16")]; + tensor var_5463_axis_0 = const()[name = tensor("op_5463_axis_0"), val = tensor(1)]; + tensor var_5463_mode_0 = const()[name = tensor("op_5463_mode_0"), val = tensor("update")]; + tensor var_5463_validate_indices_0 = const()[name = tensor("op_5463_validate_indices_0"), val = tensor(false)]; + tensor var_5463_cast_fp16 = scatter_along_axis(axis = var_5463_axis_0, data = var_5461_cast_fp16, indices = write_indices_27, mode = var_5463_mode_0, updates = v_27_cast_fp16, validate_indices = var_5463_validate_indices_0)[name = tensor("op_5463_cast_fp16")]; + tensor concat_95 = const()[name = tensor("concat_95"), val = tensor([1, 0, 0, 0, 0])]; + tensor concat_96 = const()[name = tensor("concat_96"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_27_internal_tensor_assign_2_stride_0 = const()[name = tensor("new_cache_27_internal_tensor_assign_2_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_27_internal_tensor_assign_2_begin_mask_0 = const()[name = tensor("new_cache_27_internal_tensor_assign_2_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_27_internal_tensor_assign_2_end_mask_0 = const()[name = tensor("new_cache_27_internal_tensor_assign_2_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_27_internal_tensor_assign_2_squeeze_mask_0 = const()[name = tensor("new_cache_27_internal_tensor_assign_2_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_75 = const()[name = tensor("shape_75"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_27 = const()[name = tensor("reduce_prod_27"), val = tensor(1048576)]; + tensor range_1d_27_start_0 = const()[name = tensor("range_1d_27_start_0"), val = tensor(0)]; + tensor range_1d_27_step_0 = const()[name = tensor("range_1d_27_step_0"), val = tensor(1)]; + tensor range_1d_27 = range_1d(end = reduce_prod_27, start = range_1d_27_start_0, step = range_1d_27_step_0)[name = tensor("range_1d_27")]; + tensor reshape_135 = reshape(shape = shape_75, x = range_1d_27)[name = tensor("reshape_135")]; + tensor slice_by_index_27 = slice_by_index(begin = concat_95, begin_mask = new_cache_27_internal_tensor_assign_2_begin_mask_0, end = concat_96, end_mask = new_cache_27_internal_tensor_assign_2_end_mask_0, squeeze_mask = new_cache_27_internal_tensor_assign_2_squeeze_mask_0, stride = new_cache_27_internal_tensor_assign_2_stride_0, x = reshape_135)[name = tensor("slice_by_index_27")]; + tensor reshape_136_shape_0 = const()[name = tensor("reshape_136_shape_0"), val = tensor([-1])]; + tensor reshape_136 = reshape(shape = reshape_136_shape_0, x = slice_by_index_27)[name = tensor("reshape_136")]; + tensor reshape_137_shape_0 = const()[name = tensor("reshape_137_shape_0"), val = tensor([-1])]; + tensor reshape_137_cast_fp16 = reshape(shape = reshape_137_shape_0, x = var_5463_cast_fp16)[name = tensor("reshape_137_cast_fp16")]; + tensor reshape_138_shape_0 = const()[name = tensor("reshape_138_shape_0"), val = tensor([-1])]; + tensor reshape_138_cast_fp16 = reshape(shape = reshape_138_shape_0, x = reshape_134_cast_fp16)[name = tensor("reshape_138_cast_fp16")]; + tensor scatter_27_mode_0 = const()[name = tensor("scatter_27_mode_0"), val = tensor("update")]; + tensor scatter_27_axis_0 = const()[name = tensor("scatter_27_axis_0"), val = tensor(0)]; + tensor scatter_27_validate_indices_0 = const()[name = tensor("scatter_27_validate_indices_0"), val = tensor(false)]; + tensor scatter_27_cast_fp16 = scatter(axis = scatter_27_axis_0, data = reshape_138_cast_fp16, indices = reshape_136, mode = scatter_27_mode_0, updates = reshape_137_cast_fp16, validate_indices = scatter_27_validate_indices_0)[name = tensor("scatter_27_cast_fp16")]; + tensor reshape_139_cast_fp16 = reshape(shape = shape_75, x = scatter_27_cast_fp16)[name = tensor("reshape_139_cast_fp16")]; + tensor reshape_139_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("reshape_139_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor keys_79_begin_0 = const()[name = tensor("keys_79_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor keys_79_end_0 = const()[name = tensor("keys_79_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor keys_79_end_mask_0 = const()[name = tensor("keys_79_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor keys_79_squeeze_mask_0 = const()[name = tensor("keys_79_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor keys_79_cast_fp16 = slice_by_index(begin = keys_79_begin_0, end = keys_79_end_0, end_mask = keys_79_end_mask_0, squeeze_mask = keys_79_squeeze_mask_0, x = reshape_139_cast_fp16)[name = tensor("keys_79_cast_fp16")]; + tensor values_79_begin_0 = const()[name = tensor("values_79_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor values_79_end_0 = const()[name = tensor("values_79_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor values_79_end_mask_0 = const()[name = tensor("values_79_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor values_79_squeeze_mask_0 = const()[name = tensor("values_79_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor values_79_cast_fp16 = slice_by_index(begin = values_79_begin_0, end = values_79_end_0, end_mask = values_79_end_mask_0, squeeze_mask = values_79_squeeze_mask_0, x = reshape_139_cast_fp16)[name = tensor("values_79_cast_fp16")]; + tensor var_5475_cast_fp16 = not_equal(x = keys_79_cast_fp16, y = keys_79_cast_fp16)[name = tensor("op_5475_cast_fp16")]; + tensor keys_81_cast_fp16 = select(a = var_463_to_fp16, b = keys_79_cast_fp16, cond = var_5475_cast_fp16)[name = tensor("keys_81_cast_fp16")]; + tensor var_5483_cast_fp16 = not_equal(x = values_79_cast_fp16, y = values_79_cast_fp16)[name = tensor("op_5483_cast_fp16")]; + tensor values_81_cast_fp16 = select(a = var_463_to_fp16, b = values_79_cast_fp16, cond = var_5483_cast_fp16)[name = tensor("values_81_cast_fp16")]; + tensor var_5507 = const()[name = tensor("op_5507"), val = tensor([0, 2, 1, 3])]; + tensor var_5523 = const()[name = tensor("op_5523"), val = tensor([1, 1, 1])]; + tensor var_5524_cast_fp16 = reshape(shape = var_5523, x = position13_to_fp16)[name = tensor("op_5524_cast_fp16")]; + tensor var_5544_to_fp16 = const()[name = tensor("op_5544_to_fp16"), val = tensor(0x1p+0)]; + tensor valid_len_27_cast_fp16 = add(x = var_5524_cast_fp16, y = var_5544_to_fp16)[name = tensor("valid_len_27_cast_fp16")]; + tensor valid_mask_27_cast_fp16 = less(x = k_positions_1_promoted_to_fp16, y = valid_len_27_cast_fp16)[name = tensor("valid_mask_27_cast_fp16")]; + tensor causal_mask_27_cast_fp16 = less_equal(x = k_positions_1_promoted_to_fp16, y = var_5524_cast_fp16)[name = tensor("causal_mask_27_cast_fp16")]; + tensor attn_mask_53 = logical_and(x = valid_mask_27_cast_fp16, y = causal_mask_27_cast_fp16)[name = tensor("attn_mask_53")]; + tensor attn_mask_55_axes_0 = const()[name = tensor("attn_mask_55_axes_0"), val = tensor([1])]; + tensor attn_mask_55 = expand_dims(axes = attn_mask_55_axes_0, x = attn_mask_53)[name = tensor("attn_mask_55")]; + tensor var_5562_transpose_x_0 = const()[name = tensor("op_5562_transpose_x_0"), val = tensor(false)]; + tensor var_5562_transpose_y_0 = const()[name = tensor("op_5562_transpose_y_0"), val = tensor(false)]; + tensor transpose_98_perm_0 = const()[name = tensor("transpose_98_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_99_perm_0 = const()[name = tensor("transpose_99_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_99 = transpose(perm = transpose_99_perm_0, x = keys_81_cast_fp16)[name = tensor("transpose_161")]; + tensor transpose_98 = transpose(perm = transpose_98_perm_0, x = q_81_cast_fp16)[name = tensor("transpose_162")]; + tensor var_5562_cast_fp16 = matmul(transpose_x = var_5562_transpose_x_0, transpose_y = var_5562_transpose_y_0, x = transpose_98, y = transpose_99)[name = tensor("op_5562_cast_fp16")]; + tensor var_5556_to_fp16 = const()[name = tensor("op_5556_to_fp16"), val = tensor([0x1p-3])]; + tensor attn_weights_79_cast_fp16 = mul(x = var_5562_cast_fp16, y = var_5556_to_fp16)[name = tensor("attn_weights_79_cast_fp16")]; + tensor var_5564 = logical_not(x = attn_mask_55)[name = tensor("op_5564")]; + tensor var_5565_to_fp16 = const()[name = tensor("op_5565_to_fp16"), val = tensor(-inf)]; + tensor attn_weights_81_cast_fp16 = select(a = var_5565_to_fp16, b = attn_weights_79_cast_fp16, cond = var_5564)[name = tensor("attn_weights_81_cast_fp16")]; + tensor var_5567 = const()[name = tensor("op_5567"), val = tensor(-1)]; + tensor attn_weights_83_cast_fp16 = softmax(axis = var_5567, x = attn_weights_81_cast_fp16)[name = tensor("attn_weights_83_cast_fp16")]; + tensor attn_output_27_transpose_x_0 = const()[name = tensor("attn_output_27_transpose_x_0"), val = tensor(false)]; + tensor attn_output_27_transpose_y_0 = const()[name = tensor("attn_output_27_transpose_y_0"), val = tensor(false)]; + tensor values_83_cast_fp16 = transpose(perm = var_5507, x = values_81_cast_fp16)[name = tensor("transpose_163")]; + tensor attn_output_27_cast_fp16 = matmul(transpose_x = attn_output_27_transpose_x_0, transpose_y = attn_output_27_transpose_y_0, x = attn_weights_83_cast_fp16, y = values_83_cast_fp16)[name = tensor("attn_output_27_cast_fp16")]; + tensor var_5575 = const()[name = tensor("op_5575"), val = tensor([0, 2, 1, 3])]; + tensor var_5580 = const()[name = tensor("op_5580"), val = tensor([1, 1, 1024])]; + tensor var_5576_cast_fp16 = transpose(perm = var_5575, x = attn_output_27_cast_fp16)[name = tensor("transpose_160")]; + tensor input_163_cast_fp16 = reshape(shape = var_5580, x = var_5576_cast_fp16)[name = tensor("input_163_cast_fp16")]; + tensor attn13_out_proj_weight_to_fp16 = const()[name = tensor("attn13_out_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(334698368)))]; + tensor linear_54_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = attn13_out_proj_weight_to_fp16, x = input_163_cast_fp16)[name = tensor("linear_54_cast_fp16")]; + tensor var_5586_to_fp16 = const()[name = tensor("op_5586_to_fp16"), val = tensor(0x1p+0)]; + tensor var_5587_cast_fp16 = add(x = position13_to_fp16, y = var_5586_to_fp16)[name = tensor("op_5587_cast_fp16")]; + tensor var_5587_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_5587_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor input_165_cast_fp16 = add(x = input_159_cast_fp16, y = linear_54_cast_fp16)[name = tensor("input_165_cast_fp16")]; + tensor input_167_axes_0 = const()[name = tensor("input_167_axes_0"), val = tensor([-1])]; + tensor norm13_2_weight_to_fp16 = const()[name = tensor("norm13_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(336795584)))]; + tensor norm13_2_bias_to_fp16 = const()[name = tensor("norm13_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(336797696)))]; + tensor var_5591_to_fp16 = const()[name = tensor("op_5591_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_167_cast_fp16 = layer_norm(axes = input_167_axes_0, beta = norm13_2_bias_to_fp16, epsilon = var_5591_to_fp16, gamma = norm13_2_weight_to_fp16, x = input_165_cast_fp16)[name = tensor("input_167_cast_fp16")]; + tensor linear13_1_weight_to_fp16 = const()[name = tensor("linear13_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(336799808)))]; + tensor linear_55_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = linear13_1_weight_to_fp16, x = input_167_cast_fp16)[name = tensor("linear_55_cast_fp16")]; + tensor input_169_mode_0 = const()[name = tensor("input_169_mode_0"), val = tensor("EXACT")]; + tensor input_169_cast_fp16 = gelu(mode = input_169_mode_0, x = linear_55_cast_fp16)[name = tensor("input_169_cast_fp16")]; + tensor linear13_2_weight_to_fp16 = const()[name = tensor("linear13_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(345188480)))]; + tensor linear_56_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = linear13_2_weight_to_fp16, x = input_169_cast_fp16)[name = tensor("linear_56_cast_fp16")]; + tensor input_171_cast_fp16 = add(x = input_165_cast_fp16, y = linear_56_cast_fp16)[name = tensor("input_171_cast_fp16")]; + tensor input_173_axes_0 = const()[name = tensor("input_173_axes_0"), val = tensor([-1])]; + tensor norm14_1_weight_to_fp16 = const()[name = tensor("norm14_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(353577152)))]; + tensor norm14_1_bias_to_fp16 = const()[name = tensor("norm14_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(353579264)))]; + tensor var_5608_to_fp16 = const()[name = tensor("op_5608_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_173_cast_fp16 = layer_norm(axes = input_173_axes_0, beta = norm14_1_bias_to_fp16, epsilon = var_5608_to_fp16, gamma = norm14_1_weight_to_fp16, x = input_171_cast_fp16)[name = tensor("input_173_cast_fp16")]; + tensor attn14_in_proj_weight_to_fp16 = const()[name = tensor("attn14_in_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(353581376)))]; + tensor linear_57_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = attn14_in_proj_weight_to_fp16, x = input_173_cast_fp16)[name = tensor("linear_57_cast_fp16")]; + tensor var_5622 = const()[name = tensor("op_5622"), val = tensor([1, 1, 3, 16, 64])]; + tensor qkv_29_cast_fp16 = reshape(shape = var_5622, x = linear_57_cast_fp16)[name = tensor("qkv_29_cast_fp16")]; + tensor q_85_begin_0 = const()[name = tensor("q_85_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor q_85_end_0 = const()[name = tensor("q_85_end_0"), val = tensor([1, 1, 1, 16, 64])]; + tensor q_85_end_mask_0 = const()[name = tensor("q_85_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor q_85_squeeze_mask_0 = const()[name = tensor("q_85_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor q_85_cast_fp16 = slice_by_index(begin = q_85_begin_0, end = q_85_end_0, end_mask = q_85_end_mask_0, squeeze_mask = q_85_squeeze_mask_0, x = qkv_29_cast_fp16)[name = tensor("q_85_cast_fp16")]; + tensor k_57_begin_0 = const()[name = tensor("k_57_begin_0"), val = tensor([0, 0, 1, 0, 0])]; + tensor k_57_end_0 = const()[name = tensor("k_57_end_0"), val = tensor([1, 1, 2, 16, 64])]; + tensor k_57_end_mask_0 = const()[name = tensor("k_57_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor k_57_squeeze_mask_0 = const()[name = tensor("k_57_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor k_57_cast_fp16 = slice_by_index(begin = k_57_begin_0, end = k_57_end_0, end_mask = k_57_end_mask_0, squeeze_mask = k_57_squeeze_mask_0, x = qkv_29_cast_fp16)[name = tensor("k_57_cast_fp16")]; + tensor v_29_begin_0 = const()[name = tensor("v_29_begin_0"), val = tensor([0, 0, 2, 0, 0])]; + tensor v_29_end_0 = const()[name = tensor("v_29_end_0"), val = tensor([1, 1, 3, 16, 64])]; + tensor v_29_end_mask_0 = const()[name = tensor("v_29_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor v_29_squeeze_mask_0 = const()[name = tensor("v_29_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor v_29_cast_fp16 = slice_by_index(begin = v_29_begin_0, end = v_29_end_0, end_mask = v_29_end_mask_0, squeeze_mask = v_29_squeeze_mask_0, x = qkv_29_cast_fp16)[name = tensor("v_29_cast_fp16")]; + tensor var_5688 = const()[name = tensor("op_5688"), val = tensor([1, 1, 1, 1])]; + tensor position14_to_fp16_dtype_0 = const()[name = tensor("position14_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor position14_to_fp16 = cast(dtype = position14_to_fp16_dtype_0, x = position14)[name = tensor("cast_483")]; + tensor ts_89_cast_fp16 = reshape(shape = var_5688, x = position14_to_fp16)[name = tensor("ts_89_cast_fp16")]; + tensor var_5695 = const()[name = tensor("op_5695"), val = tensor([1, 1, 16, 32, 2])]; + tensor q_complex_29_cast_fp16 = reshape(shape = var_5695, x = q_85_cast_fp16)[name = tensor("q_complex_29_cast_fp16")]; + tensor var_5702 = const()[name = tensor("op_5702"), val = tensor([1, 1, 16, 32, 2])]; + tensor k_complex_29_cast_fp16 = reshape(shape = var_5702, x = k_57_cast_fp16)[name = tensor("k_complex_29_cast_fp16")]; + tensor var_5706_begin_0 = const()[name = tensor("op_5706_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_5706_end_0 = const()[name = tensor("op_5706_end_0"), val = tensor([1, 1, 16, 32, 1])]; + tensor var_5706_end_mask_0 = const()[name = tensor("op_5706_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_5706_squeeze_mask_0 = const()[name = tensor("op_5706_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_5706_cast_fp16 = slice_by_index(begin = var_5706_begin_0, end = var_5706_end_0, end_mask = var_5706_end_mask_0, squeeze_mask = var_5706_squeeze_mask_0, x = q_complex_29_cast_fp16)[name = tensor("op_5706_cast_fp16")]; + tensor var_5714_begin_0 = const()[name = tensor("op_5714_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_5714_end_0 = const()[name = tensor("op_5714_end_0"), val = tensor([1, 1, 16, 32, 2])]; + tensor var_5714_end_mask_0 = const()[name = tensor("op_5714_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_5714_squeeze_mask_0 = const()[name = tensor("op_5714_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_5714_cast_fp16 = slice_by_index(begin = var_5714_begin_0, end = var_5714_end_0, end_mask = var_5714_end_mask_0, squeeze_mask = var_5714_squeeze_mask_0, x = q_complex_29_cast_fp16)[name = tensor("op_5714_cast_fp16")]; + tensor var_5722_begin_0 = const()[name = tensor("op_5722_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_5722_end_0 = const()[name = tensor("op_5722_end_0"), val = tensor([1, 1, 16, 32, 1])]; + tensor var_5722_end_mask_0 = const()[name = tensor("op_5722_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_5722_squeeze_mask_0 = const()[name = tensor("op_5722_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_5722_cast_fp16 = slice_by_index(begin = var_5722_begin_0, end = var_5722_end_0, end_mask = var_5722_end_mask_0, squeeze_mask = var_5722_squeeze_mask_0, x = k_complex_29_cast_fp16)[name = tensor("op_5722_cast_fp16")]; + tensor var_5730_begin_0 = const()[name = tensor("op_5730_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_5730_end_0 = const()[name = tensor("op_5730_end_0"), val = tensor([1, 1, 16, 32, 2])]; + tensor var_5730_end_mask_0 = const()[name = tensor("op_5730_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_5730_squeeze_mask_0 = const()[name = tensor("op_5730_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_5730_cast_fp16 = slice_by_index(begin = var_5730_begin_0, end = var_5730_end_0, end_mask = var_5730_end_mask_0, squeeze_mask = var_5730_squeeze_mask_0, x = k_complex_29_cast_fp16)[name = tensor("op_5730_cast_fp16")]; + tensor freqs_29_to_fp16 = const()[name = tensor("freqs_29_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(359872896)))]; + tensor var_5736_cast_fp16 = mul(x = freqs_29_to_fp16, y = ts_89_cast_fp16)[name = tensor("op_5736_cast_fp16")]; + tensor rotr_29_cast_fp16 = cos(x = var_5736_cast_fp16)[name = tensor("rotr_29_cast_fp16")]; + tensor roti_29_cast_fp16 = sin(x = var_5736_cast_fp16)[name = tensor("roti_29_cast_fp16")]; + tensor var_5740_cast_fp16 = mul(x = var_5706_cast_fp16, y = rotr_29_cast_fp16)[name = tensor("op_5740_cast_fp16")]; + tensor var_5741_cast_fp16 = mul(x = var_5714_cast_fp16, y = roti_29_cast_fp16)[name = tensor("op_5741_cast_fp16")]; + tensor qor_57_cast_fp16 = sub(x = var_5740_cast_fp16, y = var_5741_cast_fp16)[name = tensor("qor_57_cast_fp16")]; + tensor var_5744_cast_fp16 = mul(x = var_5706_cast_fp16, y = roti_29_cast_fp16)[name = tensor("op_5744_cast_fp16")]; + tensor var_5745_cast_fp16 = mul(x = var_5714_cast_fp16, y = rotr_29_cast_fp16)[name = tensor("op_5745_cast_fp16")]; + tensor qoi_57_cast_fp16 = add(x = var_5744_cast_fp16, y = var_5745_cast_fp16)[name = tensor("qoi_57_cast_fp16")]; + tensor var_5748_cast_fp16 = mul(x = var_5722_cast_fp16, y = rotr_29_cast_fp16)[name = tensor("op_5748_cast_fp16")]; + tensor var_5749_cast_fp16 = mul(x = var_5730_cast_fp16, y = roti_29_cast_fp16)[name = tensor("op_5749_cast_fp16")]; + tensor kor_57_cast_fp16 = sub(x = var_5748_cast_fp16, y = var_5749_cast_fp16)[name = tensor("kor_57_cast_fp16")]; + tensor var_5752_cast_fp16 = mul(x = var_5722_cast_fp16, y = roti_29_cast_fp16)[name = tensor("op_5752_cast_fp16")]; + tensor var_5753_cast_fp16 = mul(x = var_5730_cast_fp16, y = rotr_29_cast_fp16)[name = tensor("op_5753_cast_fp16")]; + tensor koi_57_cast_fp16 = add(x = var_5752_cast_fp16, y = var_5753_cast_fp16)[name = tensor("koi_57_cast_fp16")]; + tensor qo_29_axis_0 = const()[name = tensor("qo_29_axis_0"), val = tensor(-1)]; + tensor qo_29_cast_fp16 = stack(axis = qo_29_axis_0, values = (qor_57_cast_fp16, qoi_57_cast_fp16))[name = tensor("qo_29_cast_fp16")]; + tensor ko_29_axis_0 = const()[name = tensor("ko_29_axis_0"), val = tensor(-1)]; + tensor ko_29_cast_fp16 = stack(axis = ko_29_axis_0, values = (kor_57_cast_fp16, koi_57_cast_fp16))[name = tensor("ko_29_cast_fp16")]; + tensor var_5786 = const()[name = tensor("op_5786"), val = tensor([1, 1, 16, 64])]; + tensor q_87_cast_fp16 = reshape(shape = var_5786, x = qo_29_cast_fp16)[name = tensor("q_87_cast_fp16")]; + tensor var_5792 = const()[name = tensor("op_5792"), val = tensor([1, 1, 16, 64])]; + tensor k_59_cast_fp16 = reshape(shape = var_5792, x = ko_29_cast_fp16)[name = tensor("k_59_cast_fp16")]; + tensor _inversed_5817_y_0_to_fp16 = const()[name = tensor("_inversed_5817_y_0_to_fp16"), val = tensor(0x1p-9)]; + tensor _inversed_5817_cast_fp16 = mul(x = ts_89_cast_fp16, y = _inversed_5817_y_0_to_fp16)[name = tensor("_inversed_5817_cast_fp16")]; + tensor var_5818_cast_fp16 = floor(x = _inversed_5817_cast_fp16)[name = tensor("op_5818_cast_fp16")]; + tensor var_5819_to_fp16 = const()[name = tensor("op_5819_to_fp16"), val = tensor(0x1p+9)]; + tensor var_5820_cast_fp16 = mul(x = var_5818_cast_fp16, y = var_5819_to_fp16)[name = tensor("op_5820_cast_fp16")]; + tensor write_indices_float_59_cast_fp16 = sub(x = ts_89_cast_fp16, y = var_5820_cast_fp16)[name = tensor("write_indices_float_59_cast_fp16")]; + tensor var_5827_dtype_0 = const()[name = tensor("op_5827_dtype_0"), val = tensor("int32")]; + tensor write_indices_29_reps_0 = const()[name = tensor("write_indices_29_reps_0"), val = tensor([1, 1, 16, 64])]; + tensor write_indices_float_59_cast_fp16_to_int32 = cast(dtype = var_5827_dtype_0, x = write_indices_float_59_cast_fp16)[name = tensor("cast_482")]; + tensor write_indices_29 = tile(reps = write_indices_29_reps_0, x = write_indices_float_59_cast_fp16_to_int32)[name = tensor("write_indices_29")]; + tensor var_5837_begin_0 = const()[name = tensor("op_5837_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_5837_end_0 = const()[name = tensor("op_5837_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor var_5837_end_mask_0 = const()[name = tensor("op_5837_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_5837_squeeze_mask_0 = const()[name = tensor("op_5837_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor cache14_to_fp16_dtype_0 = const()[name = tensor("cache14_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor cache14_to_fp16 = cast(dtype = cache14_to_fp16_dtype_0, x = cache14)[name = tensor("cast_481")]; + tensor var_5837_cast_fp16 = slice_by_index(begin = var_5837_begin_0, end = var_5837_end_0, end_mask = var_5837_end_mask_0, squeeze_mask = var_5837_squeeze_mask_0, x = cache14_to_fp16)[name = tensor("op_5837_cast_fp16")]; + tensor var_5839_axis_0 = const()[name = tensor("op_5839_axis_0"), val = tensor(1)]; + tensor var_5839_mode_0 = const()[name = tensor("op_5839_mode_0"), val = tensor("update")]; + tensor var_5839_validate_indices_0 = const()[name = tensor("op_5839_validate_indices_0"), val = tensor(false)]; + tensor var_5839_cast_fp16 = scatter_along_axis(axis = var_5839_axis_0, data = var_5837_cast_fp16, indices = write_indices_29, mode = var_5839_mode_0, updates = k_59_cast_fp16, validate_indices = var_5839_validate_indices_0)[name = tensor("op_5839_cast_fp16")]; + tensor concat_100 = const()[name = tensor("concat_100"), val = tensor([0, 0, 0, 0, 0])]; + tensor concat_101 = const()[name = tensor("concat_101"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_29_internal_tensor_assign_1_stride_0 = const()[name = tensor("new_cache_29_internal_tensor_assign_1_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_29_internal_tensor_assign_1_begin_mask_0 = const()[name = tensor("new_cache_29_internal_tensor_assign_1_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_29_internal_tensor_assign_1_end_mask_0 = const()[name = tensor("new_cache_29_internal_tensor_assign_1_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_29_internal_tensor_assign_1_squeeze_mask_0 = const()[name = tensor("new_cache_29_internal_tensor_assign_1_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_76 = const()[name = tensor("shape_76"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_28 = const()[name = tensor("reduce_prod_28"), val = tensor(1048576)]; + tensor range_1d_28_start_0 = const()[name = tensor("range_1d_28_start_0"), val = tensor(0)]; + tensor range_1d_28_step_0 = const()[name = tensor("range_1d_28_step_0"), val = tensor(1)]; + tensor range_1d_28 = range_1d(end = reduce_prod_28, start = range_1d_28_start_0, step = range_1d_28_step_0)[name = tensor("range_1d_28")]; + tensor reshape_140 = reshape(shape = shape_76, x = range_1d_28)[name = tensor("reshape_140")]; + tensor slice_by_index_28 = slice_by_index(begin = concat_100, begin_mask = new_cache_29_internal_tensor_assign_1_begin_mask_0, end = concat_101, end_mask = new_cache_29_internal_tensor_assign_1_end_mask_0, squeeze_mask = new_cache_29_internal_tensor_assign_1_squeeze_mask_0, stride = new_cache_29_internal_tensor_assign_1_stride_0, x = reshape_140)[name = tensor("slice_by_index_28")]; + tensor reshape_141_shape_0 = const()[name = tensor("reshape_141_shape_0"), val = tensor([-1])]; + tensor reshape_141 = reshape(shape = reshape_141_shape_0, x = slice_by_index_28)[name = tensor("reshape_141")]; + tensor reshape_142_shape_0 = const()[name = tensor("reshape_142_shape_0"), val = tensor([-1])]; + tensor reshape_142_cast_fp16 = reshape(shape = reshape_142_shape_0, x = var_5839_cast_fp16)[name = tensor("reshape_142_cast_fp16")]; + tensor reshape_143_shape_0 = const()[name = tensor("reshape_143_shape_0"), val = tensor([-1])]; + tensor reshape_143_cast_fp16 = reshape(shape = reshape_143_shape_0, x = cache14_to_fp16)[name = tensor("reshape_143_cast_fp16")]; + tensor scatter_28_mode_0 = const()[name = tensor("scatter_28_mode_0"), val = tensor("update")]; + tensor scatter_28_axis_0 = const()[name = tensor("scatter_28_axis_0"), val = tensor(0)]; + tensor scatter_28_validate_indices_0 = const()[name = tensor("scatter_28_validate_indices_0"), val = tensor(false)]; + tensor scatter_28_cast_fp16 = scatter(axis = scatter_28_axis_0, data = reshape_143_cast_fp16, indices = reshape_141, mode = scatter_28_mode_0, updates = reshape_142_cast_fp16, validate_indices = scatter_28_validate_indices_0)[name = tensor("scatter_28_cast_fp16")]; + tensor reshape_144_cast_fp16 = reshape(shape = shape_76, x = scatter_28_cast_fp16)[name = tensor("reshape_144_cast_fp16")]; + tensor var_5847_begin_0 = const()[name = tensor("op_5847_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor var_5847_end_0 = const()[name = tensor("op_5847_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor var_5847_end_mask_0 = const()[name = tensor("op_5847_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_5847_squeeze_mask_0 = const()[name = tensor("op_5847_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor var_5847_cast_fp16 = slice_by_index(begin = var_5847_begin_0, end = var_5847_end_0, end_mask = var_5847_end_mask_0, squeeze_mask = var_5847_squeeze_mask_0, x = reshape_144_cast_fp16)[name = tensor("op_5847_cast_fp16")]; + tensor var_5849_axis_0 = const()[name = tensor("op_5849_axis_0"), val = tensor(1)]; + tensor var_5849_mode_0 = const()[name = tensor("op_5849_mode_0"), val = tensor("update")]; + tensor var_5849_validate_indices_0 = const()[name = tensor("op_5849_validate_indices_0"), val = tensor(false)]; + tensor var_5849_cast_fp16 = scatter_along_axis(axis = var_5849_axis_0, data = var_5847_cast_fp16, indices = write_indices_29, mode = var_5849_mode_0, updates = v_29_cast_fp16, validate_indices = var_5849_validate_indices_0)[name = tensor("op_5849_cast_fp16")]; + tensor concat_102 = const()[name = tensor("concat_102"), val = tensor([1, 0, 0, 0, 0])]; + tensor concat_103 = const()[name = tensor("concat_103"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_29_internal_tensor_assign_2_stride_0 = const()[name = tensor("new_cache_29_internal_tensor_assign_2_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_29_internal_tensor_assign_2_begin_mask_0 = const()[name = tensor("new_cache_29_internal_tensor_assign_2_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_29_internal_tensor_assign_2_end_mask_0 = const()[name = tensor("new_cache_29_internal_tensor_assign_2_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_29_internal_tensor_assign_2_squeeze_mask_0 = const()[name = tensor("new_cache_29_internal_tensor_assign_2_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_77 = const()[name = tensor("shape_77"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_29 = const()[name = tensor("reduce_prod_29"), val = tensor(1048576)]; + tensor range_1d_29_start_0 = const()[name = tensor("range_1d_29_start_0"), val = tensor(0)]; + tensor range_1d_29_step_0 = const()[name = tensor("range_1d_29_step_0"), val = tensor(1)]; + tensor range_1d_29 = range_1d(end = reduce_prod_29, start = range_1d_29_start_0, step = range_1d_29_step_0)[name = tensor("range_1d_29")]; + tensor reshape_145 = reshape(shape = shape_77, x = range_1d_29)[name = tensor("reshape_145")]; + tensor slice_by_index_29 = slice_by_index(begin = concat_102, begin_mask = new_cache_29_internal_tensor_assign_2_begin_mask_0, end = concat_103, end_mask = new_cache_29_internal_tensor_assign_2_end_mask_0, squeeze_mask = new_cache_29_internal_tensor_assign_2_squeeze_mask_0, stride = new_cache_29_internal_tensor_assign_2_stride_0, x = reshape_145)[name = tensor("slice_by_index_29")]; + tensor reshape_146_shape_0 = const()[name = tensor("reshape_146_shape_0"), val = tensor([-1])]; + tensor reshape_146 = reshape(shape = reshape_146_shape_0, x = slice_by_index_29)[name = tensor("reshape_146")]; + tensor reshape_147_shape_0 = const()[name = tensor("reshape_147_shape_0"), val = tensor([-1])]; + tensor reshape_147_cast_fp16 = reshape(shape = reshape_147_shape_0, x = var_5849_cast_fp16)[name = tensor("reshape_147_cast_fp16")]; + tensor reshape_148_shape_0 = const()[name = tensor("reshape_148_shape_0"), val = tensor([-1])]; + tensor reshape_148_cast_fp16 = reshape(shape = reshape_148_shape_0, x = reshape_144_cast_fp16)[name = tensor("reshape_148_cast_fp16")]; + tensor scatter_29_mode_0 = const()[name = tensor("scatter_29_mode_0"), val = tensor("update")]; + tensor scatter_29_axis_0 = const()[name = tensor("scatter_29_axis_0"), val = tensor(0)]; + tensor scatter_29_validate_indices_0 = const()[name = tensor("scatter_29_validate_indices_0"), val = tensor(false)]; + tensor scatter_29_cast_fp16 = scatter(axis = scatter_29_axis_0, data = reshape_148_cast_fp16, indices = reshape_146, mode = scatter_29_mode_0, updates = reshape_147_cast_fp16, validate_indices = scatter_29_validate_indices_0)[name = tensor("scatter_29_cast_fp16")]; + tensor reshape_149_cast_fp16 = reshape(shape = shape_77, x = scatter_29_cast_fp16)[name = tensor("reshape_149_cast_fp16")]; + tensor reshape_149_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("reshape_149_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor keys_85_begin_0 = const()[name = tensor("keys_85_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor keys_85_end_0 = const()[name = tensor("keys_85_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor keys_85_end_mask_0 = const()[name = tensor("keys_85_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor keys_85_squeeze_mask_0 = const()[name = tensor("keys_85_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor keys_85_cast_fp16 = slice_by_index(begin = keys_85_begin_0, end = keys_85_end_0, end_mask = keys_85_end_mask_0, squeeze_mask = keys_85_squeeze_mask_0, x = reshape_149_cast_fp16)[name = tensor("keys_85_cast_fp16")]; + tensor values_85_begin_0 = const()[name = tensor("values_85_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor values_85_end_0 = const()[name = tensor("values_85_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor values_85_end_mask_0 = const()[name = tensor("values_85_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor values_85_squeeze_mask_0 = const()[name = tensor("values_85_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor values_85_cast_fp16 = slice_by_index(begin = values_85_begin_0, end = values_85_end_0, end_mask = values_85_end_mask_0, squeeze_mask = values_85_squeeze_mask_0, x = reshape_149_cast_fp16)[name = tensor("values_85_cast_fp16")]; + tensor var_5861_cast_fp16 = not_equal(x = keys_85_cast_fp16, y = keys_85_cast_fp16)[name = tensor("op_5861_cast_fp16")]; + tensor keys_87_cast_fp16 = select(a = var_463_to_fp16, b = keys_85_cast_fp16, cond = var_5861_cast_fp16)[name = tensor("keys_87_cast_fp16")]; + tensor var_5869_cast_fp16 = not_equal(x = values_85_cast_fp16, y = values_85_cast_fp16)[name = tensor("op_5869_cast_fp16")]; + tensor values_87_cast_fp16 = select(a = var_463_to_fp16, b = values_85_cast_fp16, cond = var_5869_cast_fp16)[name = tensor("values_87_cast_fp16")]; + tensor var_5893 = const()[name = tensor("op_5893"), val = tensor([0, 2, 1, 3])]; + tensor var_5909 = const()[name = tensor("op_5909"), val = tensor([1, 1, 1])]; + tensor var_5910_cast_fp16 = reshape(shape = var_5909, x = position14_to_fp16)[name = tensor("op_5910_cast_fp16")]; + tensor var_5930_to_fp16 = const()[name = tensor("op_5930_to_fp16"), val = tensor(0x1p+0)]; + tensor valid_len_29_cast_fp16 = add(x = var_5910_cast_fp16, y = var_5930_to_fp16)[name = tensor("valid_len_29_cast_fp16")]; + tensor valid_mask_29_cast_fp16 = less(x = k_positions_1_promoted_to_fp16, y = valid_len_29_cast_fp16)[name = tensor("valid_mask_29_cast_fp16")]; + tensor causal_mask_29_cast_fp16 = less_equal(x = k_positions_1_promoted_to_fp16, y = var_5910_cast_fp16)[name = tensor("causal_mask_29_cast_fp16")]; + tensor attn_mask_57 = logical_and(x = valid_mask_29_cast_fp16, y = causal_mask_29_cast_fp16)[name = tensor("attn_mask_57")]; + tensor attn_mask_59_axes_0 = const()[name = tensor("attn_mask_59_axes_0"), val = tensor([1])]; + tensor attn_mask_59 = expand_dims(axes = attn_mask_59_axes_0, x = attn_mask_57)[name = tensor("attn_mask_59")]; + tensor var_5948_transpose_x_0 = const()[name = tensor("op_5948_transpose_x_0"), val = tensor(false)]; + tensor var_5948_transpose_y_0 = const()[name = tensor("op_5948_transpose_y_0"), val = tensor(false)]; + tensor transpose_100_perm_0 = const()[name = tensor("transpose_100_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_101_perm_0 = const()[name = tensor("transpose_101_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_101 = transpose(perm = transpose_101_perm_0, x = keys_87_cast_fp16)[name = tensor("transpose_157")]; + tensor transpose_100 = transpose(perm = transpose_100_perm_0, x = q_87_cast_fp16)[name = tensor("transpose_158")]; + tensor var_5948_cast_fp16 = matmul(transpose_x = var_5948_transpose_x_0, transpose_y = var_5948_transpose_y_0, x = transpose_100, y = transpose_101)[name = tensor("op_5948_cast_fp16")]; + tensor var_5942_to_fp16 = const()[name = tensor("op_5942_to_fp16"), val = tensor([0x1p-3])]; + tensor attn_weights_85_cast_fp16 = mul(x = var_5948_cast_fp16, y = var_5942_to_fp16)[name = tensor("attn_weights_85_cast_fp16")]; + tensor var_5950 = logical_not(x = attn_mask_59)[name = tensor("op_5950")]; + tensor var_5951_to_fp16 = const()[name = tensor("op_5951_to_fp16"), val = tensor(-inf)]; + tensor attn_weights_87_cast_fp16 = select(a = var_5951_to_fp16, b = attn_weights_85_cast_fp16, cond = var_5950)[name = tensor("attn_weights_87_cast_fp16")]; + tensor var_5953 = const()[name = tensor("op_5953"), val = tensor(-1)]; + tensor attn_weights_89_cast_fp16 = softmax(axis = var_5953, x = attn_weights_87_cast_fp16)[name = tensor("attn_weights_89_cast_fp16")]; + tensor attn_output_29_transpose_x_0 = const()[name = tensor("attn_output_29_transpose_x_0"), val = tensor(false)]; + tensor attn_output_29_transpose_y_0 = const()[name = tensor("attn_output_29_transpose_y_0"), val = tensor(false)]; + tensor values_89_cast_fp16 = transpose(perm = var_5893, x = values_87_cast_fp16)[name = tensor("transpose_159")]; + tensor attn_output_29_cast_fp16 = matmul(transpose_x = attn_output_29_transpose_x_0, transpose_y = attn_output_29_transpose_y_0, x = attn_weights_89_cast_fp16, y = values_89_cast_fp16)[name = tensor("attn_output_29_cast_fp16")]; + tensor var_5961 = const()[name = tensor("op_5961"), val = tensor([0, 2, 1, 3])]; + tensor var_5966 = const()[name = tensor("op_5966"), val = tensor([1, 1, 1024])]; + tensor var_5962_cast_fp16 = transpose(perm = var_5961, x = attn_output_29_cast_fp16)[name = tensor("transpose_156")]; + tensor input_175_cast_fp16 = reshape(shape = var_5966, x = var_5962_cast_fp16)[name = tensor("input_175_cast_fp16")]; + tensor attn14_out_proj_weight_to_fp16 = const()[name = tensor("attn14_out_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(359873024)))]; + tensor linear_58_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = attn14_out_proj_weight_to_fp16, x = input_175_cast_fp16)[name = tensor("linear_58_cast_fp16")]; + tensor var_5972_to_fp16 = const()[name = tensor("op_5972_to_fp16"), val = tensor(0x1p+0)]; + tensor var_5973_cast_fp16 = add(x = position14_to_fp16, y = var_5972_to_fp16)[name = tensor("op_5973_cast_fp16")]; + tensor var_5973_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_5973_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor input_177_cast_fp16 = add(x = input_171_cast_fp16, y = linear_58_cast_fp16)[name = tensor("input_177_cast_fp16")]; + tensor input_179_axes_0 = const()[name = tensor("input_179_axes_0"), val = tensor([-1])]; + tensor norm14_2_weight_to_fp16 = const()[name = tensor("norm14_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(361970240)))]; + tensor norm14_2_bias_to_fp16 = const()[name = tensor("norm14_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(361972352)))]; + tensor var_5977_to_fp16 = const()[name = tensor("op_5977_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_179_cast_fp16 = layer_norm(axes = input_179_axes_0, beta = norm14_2_bias_to_fp16, epsilon = var_5977_to_fp16, gamma = norm14_2_weight_to_fp16, x = input_177_cast_fp16)[name = tensor("input_179_cast_fp16")]; + tensor linear14_1_weight_to_fp16 = const()[name = tensor("linear14_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(361974464)))]; + tensor linear_59_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = linear14_1_weight_to_fp16, x = input_179_cast_fp16)[name = tensor("linear_59_cast_fp16")]; + tensor input_181_mode_0 = const()[name = tensor("input_181_mode_0"), val = tensor("EXACT")]; + tensor input_181_cast_fp16 = gelu(mode = input_181_mode_0, x = linear_59_cast_fp16)[name = tensor("input_181_cast_fp16")]; + tensor linear14_2_weight_to_fp16 = const()[name = tensor("linear14_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(370363136)))]; + tensor linear_60_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = linear14_2_weight_to_fp16, x = input_181_cast_fp16)[name = tensor("linear_60_cast_fp16")]; + tensor input_183_cast_fp16 = add(x = input_177_cast_fp16, y = linear_60_cast_fp16)[name = tensor("input_183_cast_fp16")]; + tensor input_185_axes_0 = const()[name = tensor("input_185_axes_0"), val = tensor([-1])]; + tensor norm15_1_weight_to_fp16 = const()[name = tensor("norm15_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(378751808)))]; + tensor norm15_1_bias_to_fp16 = const()[name = tensor("norm15_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(378753920)))]; + tensor var_5994_to_fp16 = const()[name = tensor("op_5994_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_185_cast_fp16 = layer_norm(axes = input_185_axes_0, beta = norm15_1_bias_to_fp16, epsilon = var_5994_to_fp16, gamma = norm15_1_weight_to_fp16, x = input_183_cast_fp16)[name = tensor("input_185_cast_fp16")]; + tensor attn15_in_proj_weight_to_fp16 = const()[name = tensor("attn15_in_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(378756032)))]; + tensor linear_61_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = attn15_in_proj_weight_to_fp16, x = input_185_cast_fp16)[name = tensor("linear_61_cast_fp16")]; + tensor var_6008 = const()[name = tensor("op_6008"), val = tensor([1, 1, 3, 16, 64])]; + tensor qkv_31_cast_fp16 = reshape(shape = var_6008, x = linear_61_cast_fp16)[name = tensor("qkv_31_cast_fp16")]; + tensor q_91_begin_0 = const()[name = tensor("q_91_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor q_91_end_0 = const()[name = tensor("q_91_end_0"), val = tensor([1, 1, 1, 16, 64])]; + tensor q_91_end_mask_0 = const()[name = tensor("q_91_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor q_91_squeeze_mask_0 = const()[name = tensor("q_91_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor q_91_cast_fp16 = slice_by_index(begin = q_91_begin_0, end = q_91_end_0, end_mask = q_91_end_mask_0, squeeze_mask = q_91_squeeze_mask_0, x = qkv_31_cast_fp16)[name = tensor("q_91_cast_fp16")]; + tensor k_61_begin_0 = const()[name = tensor("k_61_begin_0"), val = tensor([0, 0, 1, 0, 0])]; + tensor k_61_end_0 = const()[name = tensor("k_61_end_0"), val = tensor([1, 1, 2, 16, 64])]; + tensor k_61_end_mask_0 = const()[name = tensor("k_61_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor k_61_squeeze_mask_0 = const()[name = tensor("k_61_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor k_61_cast_fp16 = slice_by_index(begin = k_61_begin_0, end = k_61_end_0, end_mask = k_61_end_mask_0, squeeze_mask = k_61_squeeze_mask_0, x = qkv_31_cast_fp16)[name = tensor("k_61_cast_fp16")]; + tensor v_31_begin_0 = const()[name = tensor("v_31_begin_0"), val = tensor([0, 0, 2, 0, 0])]; + tensor v_31_end_0 = const()[name = tensor("v_31_end_0"), val = tensor([1, 1, 3, 16, 64])]; + tensor v_31_end_mask_0 = const()[name = tensor("v_31_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor v_31_squeeze_mask_0 = const()[name = tensor("v_31_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor v_31_cast_fp16 = slice_by_index(begin = v_31_begin_0, end = v_31_end_0, end_mask = v_31_end_mask_0, squeeze_mask = v_31_squeeze_mask_0, x = qkv_31_cast_fp16)[name = tensor("v_31_cast_fp16")]; + tensor var_6074 = const()[name = tensor("op_6074"), val = tensor([1, 1, 1, 1])]; + tensor position15_to_fp16_dtype_0 = const()[name = tensor("position15_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor position15_to_fp16 = cast(dtype = position15_to_fp16_dtype_0, x = position15)[name = tensor("cast_478")]; + tensor ts_95_cast_fp16 = reshape(shape = var_6074, x = position15_to_fp16)[name = tensor("ts_95_cast_fp16")]; + tensor var_6081 = const()[name = tensor("op_6081"), val = tensor([1, 1, 16, 32, 2])]; + tensor q_complex_31_cast_fp16 = reshape(shape = var_6081, x = q_91_cast_fp16)[name = tensor("q_complex_31_cast_fp16")]; + tensor var_6088 = const()[name = tensor("op_6088"), val = tensor([1, 1, 16, 32, 2])]; + tensor k_complex_31_cast_fp16 = reshape(shape = var_6088, x = k_61_cast_fp16)[name = tensor("k_complex_31_cast_fp16")]; + tensor var_6092_begin_0 = const()[name = tensor("op_6092_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_6092_end_0 = const()[name = tensor("op_6092_end_0"), val = tensor([1, 1, 16, 32, 1])]; + tensor var_6092_end_mask_0 = const()[name = tensor("op_6092_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_6092_squeeze_mask_0 = const()[name = tensor("op_6092_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_6092_cast_fp16 = slice_by_index(begin = var_6092_begin_0, end = var_6092_end_0, end_mask = var_6092_end_mask_0, squeeze_mask = var_6092_squeeze_mask_0, x = q_complex_31_cast_fp16)[name = tensor("op_6092_cast_fp16")]; + tensor var_6100_begin_0 = const()[name = tensor("op_6100_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_6100_end_0 = const()[name = tensor("op_6100_end_0"), val = tensor([1, 1, 16, 32, 2])]; + tensor var_6100_end_mask_0 = const()[name = tensor("op_6100_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_6100_squeeze_mask_0 = const()[name = tensor("op_6100_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_6100_cast_fp16 = slice_by_index(begin = var_6100_begin_0, end = var_6100_end_0, end_mask = var_6100_end_mask_0, squeeze_mask = var_6100_squeeze_mask_0, x = q_complex_31_cast_fp16)[name = tensor("op_6100_cast_fp16")]; + tensor var_6108_begin_0 = const()[name = tensor("op_6108_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_6108_end_0 = const()[name = tensor("op_6108_end_0"), val = tensor([1, 1, 16, 32, 1])]; + tensor var_6108_end_mask_0 = const()[name = tensor("op_6108_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_6108_squeeze_mask_0 = const()[name = tensor("op_6108_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_6108_cast_fp16 = slice_by_index(begin = var_6108_begin_0, end = var_6108_end_0, end_mask = var_6108_end_mask_0, squeeze_mask = var_6108_squeeze_mask_0, x = k_complex_31_cast_fp16)[name = tensor("op_6108_cast_fp16")]; + tensor var_6116_begin_0 = const()[name = tensor("op_6116_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_6116_end_0 = const()[name = tensor("op_6116_end_0"), val = tensor([1, 1, 16, 32, 2])]; + tensor var_6116_end_mask_0 = const()[name = tensor("op_6116_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_6116_squeeze_mask_0 = const()[name = tensor("op_6116_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_6116_cast_fp16 = slice_by_index(begin = var_6116_begin_0, end = var_6116_end_0, end_mask = var_6116_end_mask_0, squeeze_mask = var_6116_squeeze_mask_0, x = k_complex_31_cast_fp16)[name = tensor("op_6116_cast_fp16")]; + tensor freqs_31_to_fp16 = const()[name = tensor("freqs_31_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(385047552)))]; + tensor var_6122_cast_fp16 = mul(x = freqs_31_to_fp16, y = ts_95_cast_fp16)[name = tensor("op_6122_cast_fp16")]; + tensor rotr_31_cast_fp16 = cos(x = var_6122_cast_fp16)[name = tensor("rotr_31_cast_fp16")]; + tensor roti_31_cast_fp16 = sin(x = var_6122_cast_fp16)[name = tensor("roti_31_cast_fp16")]; + tensor var_6126_cast_fp16 = mul(x = var_6092_cast_fp16, y = rotr_31_cast_fp16)[name = tensor("op_6126_cast_fp16")]; + tensor var_6127_cast_fp16 = mul(x = var_6100_cast_fp16, y = roti_31_cast_fp16)[name = tensor("op_6127_cast_fp16")]; + tensor qor_61_cast_fp16 = sub(x = var_6126_cast_fp16, y = var_6127_cast_fp16)[name = tensor("qor_61_cast_fp16")]; + tensor var_6130_cast_fp16 = mul(x = var_6092_cast_fp16, y = roti_31_cast_fp16)[name = tensor("op_6130_cast_fp16")]; + tensor var_6131_cast_fp16 = mul(x = var_6100_cast_fp16, y = rotr_31_cast_fp16)[name = tensor("op_6131_cast_fp16")]; + tensor qoi_61_cast_fp16 = add(x = var_6130_cast_fp16, y = var_6131_cast_fp16)[name = tensor("qoi_61_cast_fp16")]; + tensor var_6134_cast_fp16 = mul(x = var_6108_cast_fp16, y = rotr_31_cast_fp16)[name = tensor("op_6134_cast_fp16")]; + tensor var_6135_cast_fp16 = mul(x = var_6116_cast_fp16, y = roti_31_cast_fp16)[name = tensor("op_6135_cast_fp16")]; + tensor kor_61_cast_fp16 = sub(x = var_6134_cast_fp16, y = var_6135_cast_fp16)[name = tensor("kor_61_cast_fp16")]; + tensor var_6138_cast_fp16 = mul(x = var_6108_cast_fp16, y = roti_31_cast_fp16)[name = tensor("op_6138_cast_fp16")]; + tensor var_6139_cast_fp16 = mul(x = var_6116_cast_fp16, y = rotr_31_cast_fp16)[name = tensor("op_6139_cast_fp16")]; + tensor koi_61_cast_fp16 = add(x = var_6138_cast_fp16, y = var_6139_cast_fp16)[name = tensor("koi_61_cast_fp16")]; + tensor qo_31_axis_0 = const()[name = tensor("qo_31_axis_0"), val = tensor(-1)]; + tensor qo_31_cast_fp16 = stack(axis = qo_31_axis_0, values = (qor_61_cast_fp16, qoi_61_cast_fp16))[name = tensor("qo_31_cast_fp16")]; + tensor ko_31_axis_0 = const()[name = tensor("ko_31_axis_0"), val = tensor(-1)]; + tensor ko_31_cast_fp16 = stack(axis = ko_31_axis_0, values = (kor_61_cast_fp16, koi_61_cast_fp16))[name = tensor("ko_31_cast_fp16")]; + tensor var_6172 = const()[name = tensor("op_6172"), val = tensor([1, 1, 16, 64])]; + tensor q_93_cast_fp16 = reshape(shape = var_6172, x = qo_31_cast_fp16)[name = tensor("q_93_cast_fp16")]; + tensor var_6178 = const()[name = tensor("op_6178"), val = tensor([1, 1, 16, 64])]; + tensor k_63_cast_fp16 = reshape(shape = var_6178, x = ko_31_cast_fp16)[name = tensor("k_63_cast_fp16")]; + tensor _inversed_6203_y_0_to_fp16 = const()[name = tensor("_inversed_6203_y_0_to_fp16"), val = tensor(0x1p-9)]; + tensor _inversed_6203_cast_fp16 = mul(x = ts_95_cast_fp16, y = _inversed_6203_y_0_to_fp16)[name = tensor("_inversed_6203_cast_fp16")]; + tensor var_6204_cast_fp16 = floor(x = _inversed_6203_cast_fp16)[name = tensor("op_6204_cast_fp16")]; + tensor var_6205_to_fp16 = const()[name = tensor("op_6205_to_fp16"), val = tensor(0x1p+9)]; + tensor var_6206_cast_fp16 = mul(x = var_6204_cast_fp16, y = var_6205_to_fp16)[name = tensor("op_6206_cast_fp16")]; + tensor write_indices_float_63_cast_fp16 = sub(x = ts_95_cast_fp16, y = var_6206_cast_fp16)[name = tensor("write_indices_float_63_cast_fp16")]; + tensor var_6213_dtype_0 = const()[name = tensor("op_6213_dtype_0"), val = tensor("int32")]; + tensor write_indices_31_reps_0 = const()[name = tensor("write_indices_31_reps_0"), val = tensor([1, 1, 16, 64])]; + tensor write_indices_float_63_cast_fp16_to_int32 = cast(dtype = var_6213_dtype_0, x = write_indices_float_63_cast_fp16)[name = tensor("cast_477")]; + tensor write_indices_31 = tile(reps = write_indices_31_reps_0, x = write_indices_float_63_cast_fp16_to_int32)[name = tensor("write_indices_31")]; + tensor var_6223_begin_0 = const()[name = tensor("op_6223_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_6223_end_0 = const()[name = tensor("op_6223_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor var_6223_end_mask_0 = const()[name = tensor("op_6223_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_6223_squeeze_mask_0 = const()[name = tensor("op_6223_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor cache15_to_fp16_dtype_0 = const()[name = tensor("cache15_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor cache15_to_fp16 = cast(dtype = cache15_to_fp16_dtype_0, x = cache15)[name = tensor("cast_476")]; + tensor var_6223_cast_fp16 = slice_by_index(begin = var_6223_begin_0, end = var_6223_end_0, end_mask = var_6223_end_mask_0, squeeze_mask = var_6223_squeeze_mask_0, x = cache15_to_fp16)[name = tensor("op_6223_cast_fp16")]; + tensor var_6225_axis_0 = const()[name = tensor("op_6225_axis_0"), val = tensor(1)]; + tensor var_6225_mode_0 = const()[name = tensor("op_6225_mode_0"), val = tensor("update")]; + tensor var_6225_validate_indices_0 = const()[name = tensor("op_6225_validate_indices_0"), val = tensor(false)]; + tensor var_6225_cast_fp16 = scatter_along_axis(axis = var_6225_axis_0, data = var_6223_cast_fp16, indices = write_indices_31, mode = var_6225_mode_0, updates = k_63_cast_fp16, validate_indices = var_6225_validate_indices_0)[name = tensor("op_6225_cast_fp16")]; + tensor concat_107 = const()[name = tensor("concat_107"), val = tensor([0, 0, 0, 0, 0])]; + tensor concat_108 = const()[name = tensor("concat_108"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_31_internal_tensor_assign_1_stride_0 = const()[name = tensor("new_cache_31_internal_tensor_assign_1_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_31_internal_tensor_assign_1_begin_mask_0 = const()[name = tensor("new_cache_31_internal_tensor_assign_1_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_31_internal_tensor_assign_1_end_mask_0 = const()[name = tensor("new_cache_31_internal_tensor_assign_1_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_31_internal_tensor_assign_1_squeeze_mask_0 = const()[name = tensor("new_cache_31_internal_tensor_assign_1_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_78 = const()[name = tensor("shape_78"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_30 = const()[name = tensor("reduce_prod_30"), val = tensor(1048576)]; + tensor range_1d_30_start_0 = const()[name = tensor("range_1d_30_start_0"), val = tensor(0)]; + tensor range_1d_30_step_0 = const()[name = tensor("range_1d_30_step_0"), val = tensor(1)]; + tensor range_1d_30 = range_1d(end = reduce_prod_30, start = range_1d_30_start_0, step = range_1d_30_step_0)[name = tensor("range_1d_30")]; + tensor reshape_150 = reshape(shape = shape_78, x = range_1d_30)[name = tensor("reshape_150")]; + tensor slice_by_index_30 = slice_by_index(begin = concat_107, begin_mask = new_cache_31_internal_tensor_assign_1_begin_mask_0, end = concat_108, end_mask = new_cache_31_internal_tensor_assign_1_end_mask_0, squeeze_mask = new_cache_31_internal_tensor_assign_1_squeeze_mask_0, stride = new_cache_31_internal_tensor_assign_1_stride_0, x = reshape_150)[name = tensor("slice_by_index_30")]; + tensor reshape_151_shape_0 = const()[name = tensor("reshape_151_shape_0"), val = tensor([-1])]; + tensor reshape_151 = reshape(shape = reshape_151_shape_0, x = slice_by_index_30)[name = tensor("reshape_151")]; + tensor reshape_152_shape_0 = const()[name = tensor("reshape_152_shape_0"), val = tensor([-1])]; + tensor reshape_152_cast_fp16 = reshape(shape = reshape_152_shape_0, x = var_6225_cast_fp16)[name = tensor("reshape_152_cast_fp16")]; + tensor reshape_153_shape_0 = const()[name = tensor("reshape_153_shape_0"), val = tensor([-1])]; + tensor reshape_153_cast_fp16 = reshape(shape = reshape_153_shape_0, x = cache15_to_fp16)[name = tensor("reshape_153_cast_fp16")]; + tensor scatter_30_mode_0 = const()[name = tensor("scatter_30_mode_0"), val = tensor("update")]; + tensor scatter_30_axis_0 = const()[name = tensor("scatter_30_axis_0"), val = tensor(0)]; + tensor scatter_30_validate_indices_0 = const()[name = tensor("scatter_30_validate_indices_0"), val = tensor(false)]; + tensor scatter_30_cast_fp16 = scatter(axis = scatter_30_axis_0, data = reshape_153_cast_fp16, indices = reshape_151, mode = scatter_30_mode_0, updates = reshape_152_cast_fp16, validate_indices = scatter_30_validate_indices_0)[name = tensor("scatter_30_cast_fp16")]; + tensor reshape_154_cast_fp16 = reshape(shape = shape_78, x = scatter_30_cast_fp16)[name = tensor("reshape_154_cast_fp16")]; + tensor var_6233_begin_0 = const()[name = tensor("op_6233_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor var_6233_end_0 = const()[name = tensor("op_6233_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor var_6233_end_mask_0 = const()[name = tensor("op_6233_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_6233_squeeze_mask_0 = const()[name = tensor("op_6233_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor var_6233_cast_fp16 = slice_by_index(begin = var_6233_begin_0, end = var_6233_end_0, end_mask = var_6233_end_mask_0, squeeze_mask = var_6233_squeeze_mask_0, x = reshape_154_cast_fp16)[name = tensor("op_6233_cast_fp16")]; + tensor var_6235_axis_0 = const()[name = tensor("op_6235_axis_0"), val = tensor(1)]; + tensor var_6235_mode_0 = const()[name = tensor("op_6235_mode_0"), val = tensor("update")]; + tensor var_6235_validate_indices_0 = const()[name = tensor("op_6235_validate_indices_0"), val = tensor(false)]; + tensor var_6235_cast_fp16 = scatter_along_axis(axis = var_6235_axis_0, data = var_6233_cast_fp16, indices = write_indices_31, mode = var_6235_mode_0, updates = v_31_cast_fp16, validate_indices = var_6235_validate_indices_0)[name = tensor("op_6235_cast_fp16")]; + tensor concat_109 = const()[name = tensor("concat_109"), val = tensor([1, 0, 0, 0, 0])]; + tensor concat_110 = const()[name = tensor("concat_110"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_31_internal_tensor_assign_2_stride_0 = const()[name = tensor("new_cache_31_internal_tensor_assign_2_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_31_internal_tensor_assign_2_begin_mask_0 = const()[name = tensor("new_cache_31_internal_tensor_assign_2_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_31_internal_tensor_assign_2_end_mask_0 = const()[name = tensor("new_cache_31_internal_tensor_assign_2_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_31_internal_tensor_assign_2_squeeze_mask_0 = const()[name = tensor("new_cache_31_internal_tensor_assign_2_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_79 = const()[name = tensor("shape_79"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_31 = const()[name = tensor("reduce_prod_31"), val = tensor(1048576)]; + tensor range_1d_31_start_0 = const()[name = tensor("range_1d_31_start_0"), val = tensor(0)]; + tensor range_1d_31_step_0 = const()[name = tensor("range_1d_31_step_0"), val = tensor(1)]; + tensor range_1d_31 = range_1d(end = reduce_prod_31, start = range_1d_31_start_0, step = range_1d_31_step_0)[name = tensor("range_1d_31")]; + tensor reshape_155 = reshape(shape = shape_79, x = range_1d_31)[name = tensor("reshape_155")]; + tensor slice_by_index_31 = slice_by_index(begin = concat_109, begin_mask = new_cache_31_internal_tensor_assign_2_begin_mask_0, end = concat_110, end_mask = new_cache_31_internal_tensor_assign_2_end_mask_0, squeeze_mask = new_cache_31_internal_tensor_assign_2_squeeze_mask_0, stride = new_cache_31_internal_tensor_assign_2_stride_0, x = reshape_155)[name = tensor("slice_by_index_31")]; + tensor reshape_156_shape_0 = const()[name = tensor("reshape_156_shape_0"), val = tensor([-1])]; + tensor reshape_156 = reshape(shape = reshape_156_shape_0, x = slice_by_index_31)[name = tensor("reshape_156")]; + tensor reshape_157_shape_0 = const()[name = tensor("reshape_157_shape_0"), val = tensor([-1])]; + tensor reshape_157_cast_fp16 = reshape(shape = reshape_157_shape_0, x = var_6235_cast_fp16)[name = tensor("reshape_157_cast_fp16")]; + tensor reshape_158_shape_0 = const()[name = tensor("reshape_158_shape_0"), val = tensor([-1])]; + tensor reshape_158_cast_fp16 = reshape(shape = reshape_158_shape_0, x = reshape_154_cast_fp16)[name = tensor("reshape_158_cast_fp16")]; + tensor scatter_31_mode_0 = const()[name = tensor("scatter_31_mode_0"), val = tensor("update")]; + tensor scatter_31_axis_0 = const()[name = tensor("scatter_31_axis_0"), val = tensor(0)]; + tensor scatter_31_validate_indices_0 = const()[name = tensor("scatter_31_validate_indices_0"), val = tensor(false)]; + tensor scatter_31_cast_fp16 = scatter(axis = scatter_31_axis_0, data = reshape_158_cast_fp16, indices = reshape_156, mode = scatter_31_mode_0, updates = reshape_157_cast_fp16, validate_indices = scatter_31_validate_indices_0)[name = tensor("scatter_31_cast_fp16")]; + tensor reshape_159_cast_fp16 = reshape(shape = shape_79, x = scatter_31_cast_fp16)[name = tensor("reshape_159_cast_fp16")]; + tensor reshape_159_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("reshape_159_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor keys_91_begin_0 = const()[name = tensor("keys_91_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor keys_91_end_0 = const()[name = tensor("keys_91_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor keys_91_end_mask_0 = const()[name = tensor("keys_91_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor keys_91_squeeze_mask_0 = const()[name = tensor("keys_91_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor keys_91_cast_fp16 = slice_by_index(begin = keys_91_begin_0, end = keys_91_end_0, end_mask = keys_91_end_mask_0, squeeze_mask = keys_91_squeeze_mask_0, x = reshape_159_cast_fp16)[name = tensor("keys_91_cast_fp16")]; + tensor values_91_begin_0 = const()[name = tensor("values_91_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor values_91_end_0 = const()[name = tensor("values_91_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor values_91_end_mask_0 = const()[name = tensor("values_91_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor values_91_squeeze_mask_0 = const()[name = tensor("values_91_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor values_91_cast_fp16 = slice_by_index(begin = values_91_begin_0, end = values_91_end_0, end_mask = values_91_end_mask_0, squeeze_mask = values_91_squeeze_mask_0, x = reshape_159_cast_fp16)[name = tensor("values_91_cast_fp16")]; + tensor var_6247_cast_fp16 = not_equal(x = keys_91_cast_fp16, y = keys_91_cast_fp16)[name = tensor("op_6247_cast_fp16")]; + tensor keys_93_cast_fp16 = select(a = var_463_to_fp16, b = keys_91_cast_fp16, cond = var_6247_cast_fp16)[name = tensor("keys_93_cast_fp16")]; + tensor var_6255_cast_fp16 = not_equal(x = values_91_cast_fp16, y = values_91_cast_fp16)[name = tensor("op_6255_cast_fp16")]; + tensor values_93_cast_fp16 = select(a = var_463_to_fp16, b = values_91_cast_fp16, cond = var_6255_cast_fp16)[name = tensor("values_93_cast_fp16")]; + tensor var_6279 = const()[name = tensor("op_6279"), val = tensor([0, 2, 1, 3])]; + tensor var_6295 = const()[name = tensor("op_6295"), val = tensor([1, 1, 1])]; + tensor var_6296_cast_fp16 = reshape(shape = var_6295, x = position15_to_fp16)[name = tensor("op_6296_cast_fp16")]; + tensor var_6316_to_fp16 = const()[name = tensor("op_6316_to_fp16"), val = tensor(0x1p+0)]; + tensor valid_len_31_cast_fp16 = add(x = var_6296_cast_fp16, y = var_6316_to_fp16)[name = tensor("valid_len_31_cast_fp16")]; + tensor valid_mask_31_cast_fp16 = less(x = k_positions_1_promoted_to_fp16, y = valid_len_31_cast_fp16)[name = tensor("valid_mask_31_cast_fp16")]; + tensor causal_mask_31_cast_fp16 = less_equal(x = k_positions_1_promoted_to_fp16, y = var_6296_cast_fp16)[name = tensor("causal_mask_31_cast_fp16")]; + tensor attn_mask_61 = logical_and(x = valid_mask_31_cast_fp16, y = causal_mask_31_cast_fp16)[name = tensor("attn_mask_61")]; + tensor attn_mask_63_axes_0 = const()[name = tensor("attn_mask_63_axes_0"), val = tensor([1])]; + tensor attn_mask_63 = expand_dims(axes = attn_mask_63_axes_0, x = attn_mask_61)[name = tensor("attn_mask_63")]; + tensor var_6334_transpose_x_0 = const()[name = tensor("op_6334_transpose_x_0"), val = tensor(false)]; + tensor var_6334_transpose_y_0 = const()[name = tensor("op_6334_transpose_y_0"), val = tensor(false)]; + tensor transpose_102_perm_0 = const()[name = tensor("transpose_102_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_103_perm_0 = const()[name = tensor("transpose_103_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_103 = transpose(perm = transpose_103_perm_0, x = keys_93_cast_fp16)[name = tensor("transpose_153")]; + tensor transpose_102 = transpose(perm = transpose_102_perm_0, x = q_93_cast_fp16)[name = tensor("transpose_154")]; + tensor var_6334_cast_fp16 = matmul(transpose_x = var_6334_transpose_x_0, transpose_y = var_6334_transpose_y_0, x = transpose_102, y = transpose_103)[name = tensor("op_6334_cast_fp16")]; + tensor var_6328_to_fp16 = const()[name = tensor("op_6328_to_fp16"), val = tensor([0x1p-3])]; + tensor attn_weights_91_cast_fp16 = mul(x = var_6334_cast_fp16, y = var_6328_to_fp16)[name = tensor("attn_weights_91_cast_fp16")]; + tensor var_6336 = logical_not(x = attn_mask_63)[name = tensor("op_6336")]; + tensor var_6337_to_fp16 = const()[name = tensor("op_6337_to_fp16"), val = tensor(-inf)]; + tensor attn_weights_93_cast_fp16 = select(a = var_6337_to_fp16, b = attn_weights_91_cast_fp16, cond = var_6336)[name = tensor("attn_weights_93_cast_fp16")]; + tensor var_6339 = const()[name = tensor("op_6339"), val = tensor(-1)]; + tensor attn_weights_95_cast_fp16 = softmax(axis = var_6339, x = attn_weights_93_cast_fp16)[name = tensor("attn_weights_95_cast_fp16")]; + tensor attn_output_31_transpose_x_0 = const()[name = tensor("attn_output_31_transpose_x_0"), val = tensor(false)]; + tensor attn_output_31_transpose_y_0 = const()[name = tensor("attn_output_31_transpose_y_0"), val = tensor(false)]; + tensor values_95_cast_fp16 = transpose(perm = var_6279, x = values_93_cast_fp16)[name = tensor("transpose_155")]; + tensor attn_output_31_cast_fp16 = matmul(transpose_x = attn_output_31_transpose_x_0, transpose_y = attn_output_31_transpose_y_0, x = attn_weights_95_cast_fp16, y = values_95_cast_fp16)[name = tensor("attn_output_31_cast_fp16")]; + tensor var_6347 = const()[name = tensor("op_6347"), val = tensor([0, 2, 1, 3])]; + tensor var_6352 = const()[name = tensor("op_6352"), val = tensor([1, 1, 1024])]; + tensor var_6348_cast_fp16 = transpose(perm = var_6347, x = attn_output_31_cast_fp16)[name = tensor("transpose_152")]; + tensor input_187_cast_fp16 = reshape(shape = var_6352, x = var_6348_cast_fp16)[name = tensor("input_187_cast_fp16")]; + tensor attn15_out_proj_weight_to_fp16 = const()[name = tensor("attn15_out_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(385047680)))]; + tensor linear_62_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = attn15_out_proj_weight_to_fp16, x = input_187_cast_fp16)[name = tensor("linear_62_cast_fp16")]; + tensor var_6358_to_fp16 = const()[name = tensor("op_6358_to_fp16"), val = tensor(0x1p+0)]; + tensor var_6359_cast_fp16 = add(x = position15_to_fp16, y = var_6358_to_fp16)[name = tensor("op_6359_cast_fp16")]; + tensor var_6359_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_6359_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor input_189_cast_fp16 = add(x = input_183_cast_fp16, y = linear_62_cast_fp16)[name = tensor("input_189_cast_fp16")]; + tensor input_191_axes_0 = const()[name = tensor("input_191_axes_0"), val = tensor([-1])]; + tensor norm15_2_weight_to_fp16 = const()[name = tensor("norm15_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(387144896)))]; + tensor norm15_2_bias_to_fp16 = const()[name = tensor("norm15_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(387147008)))]; + tensor var_6363_to_fp16 = const()[name = tensor("op_6363_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_191_cast_fp16 = layer_norm(axes = input_191_axes_0, beta = norm15_2_bias_to_fp16, epsilon = var_6363_to_fp16, gamma = norm15_2_weight_to_fp16, x = input_189_cast_fp16)[name = tensor("input_191_cast_fp16")]; + tensor linear15_1_weight_to_fp16 = const()[name = tensor("linear15_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(387149120)))]; + tensor linear_63_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = linear15_1_weight_to_fp16, x = input_191_cast_fp16)[name = tensor("linear_63_cast_fp16")]; + tensor input_193_mode_0 = const()[name = tensor("input_193_mode_0"), val = tensor("EXACT")]; + tensor input_193_cast_fp16 = gelu(mode = input_193_mode_0, x = linear_63_cast_fp16)[name = tensor("input_193_cast_fp16")]; + tensor linear15_2_weight_to_fp16 = const()[name = tensor("linear15_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(395537792)))]; + tensor linear_64_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = linear15_2_weight_to_fp16, x = input_193_cast_fp16)[name = tensor("linear_64_cast_fp16")]; + tensor input_195_cast_fp16 = add(x = input_189_cast_fp16, y = linear_64_cast_fp16)[name = tensor("input_195_cast_fp16")]; + tensor input_197_axes_0 = const()[name = tensor("input_197_axes_0"), val = tensor([-1])]; + tensor norm16_1_weight_to_fp16 = const()[name = tensor("norm16_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(403926464)))]; + tensor norm16_1_bias_to_fp16 = const()[name = tensor("norm16_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(403928576)))]; + tensor var_6380_to_fp16 = const()[name = tensor("op_6380_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_197_cast_fp16 = layer_norm(axes = input_197_axes_0, beta = norm16_1_bias_to_fp16, epsilon = var_6380_to_fp16, gamma = norm16_1_weight_to_fp16, x = input_195_cast_fp16)[name = tensor("input_197_cast_fp16")]; + tensor attn16_in_proj_weight_to_fp16 = const()[name = tensor("attn16_in_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(403930688)))]; + tensor linear_65_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = attn16_in_proj_weight_to_fp16, x = input_197_cast_fp16)[name = tensor("linear_65_cast_fp16")]; + tensor var_6394 = const()[name = tensor("op_6394"), val = tensor([1, 1, 3, 16, 64])]; + tensor qkv_33_cast_fp16 = reshape(shape = var_6394, x = linear_65_cast_fp16)[name = tensor("qkv_33_cast_fp16")]; + tensor q_97_begin_0 = const()[name = tensor("q_97_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor q_97_end_0 = const()[name = tensor("q_97_end_0"), val = tensor([1, 1, 1, 16, 64])]; + tensor q_97_end_mask_0 = const()[name = tensor("q_97_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor q_97_squeeze_mask_0 = const()[name = tensor("q_97_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor q_97_cast_fp16 = slice_by_index(begin = q_97_begin_0, end = q_97_end_0, end_mask = q_97_end_mask_0, squeeze_mask = q_97_squeeze_mask_0, x = qkv_33_cast_fp16)[name = tensor("q_97_cast_fp16")]; + tensor k_65_begin_0 = const()[name = tensor("k_65_begin_0"), val = tensor([0, 0, 1, 0, 0])]; + tensor k_65_end_0 = const()[name = tensor("k_65_end_0"), val = tensor([1, 1, 2, 16, 64])]; + tensor k_65_end_mask_0 = const()[name = tensor("k_65_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor k_65_squeeze_mask_0 = const()[name = tensor("k_65_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor k_65_cast_fp16 = slice_by_index(begin = k_65_begin_0, end = k_65_end_0, end_mask = k_65_end_mask_0, squeeze_mask = k_65_squeeze_mask_0, x = qkv_33_cast_fp16)[name = tensor("k_65_cast_fp16")]; + tensor v_33_begin_0 = const()[name = tensor("v_33_begin_0"), val = tensor([0, 0, 2, 0, 0])]; + tensor v_33_end_0 = const()[name = tensor("v_33_end_0"), val = tensor([1, 1, 3, 16, 64])]; + tensor v_33_end_mask_0 = const()[name = tensor("v_33_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor v_33_squeeze_mask_0 = const()[name = tensor("v_33_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor v_33_cast_fp16 = slice_by_index(begin = v_33_begin_0, end = v_33_end_0, end_mask = v_33_end_mask_0, squeeze_mask = v_33_squeeze_mask_0, x = qkv_33_cast_fp16)[name = tensor("v_33_cast_fp16")]; + tensor var_6460 = const()[name = tensor("op_6460"), val = tensor([1, 1, 1, 1])]; + tensor position16_to_fp16_dtype_0 = const()[name = tensor("position16_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor position16_to_fp16 = cast(dtype = position16_to_fp16_dtype_0, x = position16)[name = tensor("cast_473")]; + tensor ts_101_cast_fp16 = reshape(shape = var_6460, x = position16_to_fp16)[name = tensor("ts_101_cast_fp16")]; + tensor var_6467 = const()[name = tensor("op_6467"), val = tensor([1, 1, 16, 32, 2])]; + tensor q_complex_33_cast_fp16 = reshape(shape = var_6467, x = q_97_cast_fp16)[name = tensor("q_complex_33_cast_fp16")]; + tensor var_6474 = const()[name = tensor("op_6474"), val = tensor([1, 1, 16, 32, 2])]; + tensor k_complex_33_cast_fp16 = reshape(shape = var_6474, x = k_65_cast_fp16)[name = tensor("k_complex_33_cast_fp16")]; + tensor var_6478_begin_0 = const()[name = tensor("op_6478_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_6478_end_0 = const()[name = tensor("op_6478_end_0"), val = tensor([1, 1, 16, 32, 1])]; + tensor var_6478_end_mask_0 = const()[name = tensor("op_6478_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_6478_squeeze_mask_0 = const()[name = tensor("op_6478_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_6478_cast_fp16 = slice_by_index(begin = var_6478_begin_0, end = var_6478_end_0, end_mask = var_6478_end_mask_0, squeeze_mask = var_6478_squeeze_mask_0, x = q_complex_33_cast_fp16)[name = tensor("op_6478_cast_fp16")]; + tensor var_6486_begin_0 = const()[name = tensor("op_6486_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_6486_end_0 = const()[name = tensor("op_6486_end_0"), val = tensor([1, 1, 16, 32, 2])]; + tensor var_6486_end_mask_0 = const()[name = tensor("op_6486_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_6486_squeeze_mask_0 = const()[name = tensor("op_6486_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_6486_cast_fp16 = slice_by_index(begin = var_6486_begin_0, end = var_6486_end_0, end_mask = var_6486_end_mask_0, squeeze_mask = var_6486_squeeze_mask_0, x = q_complex_33_cast_fp16)[name = tensor("op_6486_cast_fp16")]; + tensor var_6494_begin_0 = const()[name = tensor("op_6494_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_6494_end_0 = const()[name = tensor("op_6494_end_0"), val = tensor([1, 1, 16, 32, 1])]; + tensor var_6494_end_mask_0 = const()[name = tensor("op_6494_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_6494_squeeze_mask_0 = const()[name = tensor("op_6494_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_6494_cast_fp16 = slice_by_index(begin = var_6494_begin_0, end = var_6494_end_0, end_mask = var_6494_end_mask_0, squeeze_mask = var_6494_squeeze_mask_0, x = k_complex_33_cast_fp16)[name = tensor("op_6494_cast_fp16")]; + tensor var_6502_begin_0 = const()[name = tensor("op_6502_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_6502_end_0 = const()[name = tensor("op_6502_end_0"), val = tensor([1, 1, 16, 32, 2])]; + tensor var_6502_end_mask_0 = const()[name = tensor("op_6502_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_6502_squeeze_mask_0 = const()[name = tensor("op_6502_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_6502_cast_fp16 = slice_by_index(begin = var_6502_begin_0, end = var_6502_end_0, end_mask = var_6502_end_mask_0, squeeze_mask = var_6502_squeeze_mask_0, x = k_complex_33_cast_fp16)[name = tensor("op_6502_cast_fp16")]; + tensor freqs_33_to_fp16 = const()[name = tensor("freqs_33_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(410222208)))]; + tensor var_6508_cast_fp16 = mul(x = freqs_33_to_fp16, y = ts_101_cast_fp16)[name = tensor("op_6508_cast_fp16")]; + tensor rotr_33_cast_fp16 = cos(x = var_6508_cast_fp16)[name = tensor("rotr_33_cast_fp16")]; + tensor roti_33_cast_fp16 = sin(x = var_6508_cast_fp16)[name = tensor("roti_33_cast_fp16")]; + tensor var_6512_cast_fp16 = mul(x = var_6478_cast_fp16, y = rotr_33_cast_fp16)[name = tensor("op_6512_cast_fp16")]; + tensor var_6513_cast_fp16 = mul(x = var_6486_cast_fp16, y = roti_33_cast_fp16)[name = tensor("op_6513_cast_fp16")]; + tensor qor_65_cast_fp16 = sub(x = var_6512_cast_fp16, y = var_6513_cast_fp16)[name = tensor("qor_65_cast_fp16")]; + tensor var_6516_cast_fp16 = mul(x = var_6478_cast_fp16, y = roti_33_cast_fp16)[name = tensor("op_6516_cast_fp16")]; + tensor var_6517_cast_fp16 = mul(x = var_6486_cast_fp16, y = rotr_33_cast_fp16)[name = tensor("op_6517_cast_fp16")]; + tensor qoi_65_cast_fp16 = add(x = var_6516_cast_fp16, y = var_6517_cast_fp16)[name = tensor("qoi_65_cast_fp16")]; + tensor var_6520_cast_fp16 = mul(x = var_6494_cast_fp16, y = rotr_33_cast_fp16)[name = tensor("op_6520_cast_fp16")]; + tensor var_6521_cast_fp16 = mul(x = var_6502_cast_fp16, y = roti_33_cast_fp16)[name = tensor("op_6521_cast_fp16")]; + tensor kor_65_cast_fp16 = sub(x = var_6520_cast_fp16, y = var_6521_cast_fp16)[name = tensor("kor_65_cast_fp16")]; + tensor var_6524_cast_fp16 = mul(x = var_6494_cast_fp16, y = roti_33_cast_fp16)[name = tensor("op_6524_cast_fp16")]; + tensor var_6525_cast_fp16 = mul(x = var_6502_cast_fp16, y = rotr_33_cast_fp16)[name = tensor("op_6525_cast_fp16")]; + tensor koi_65_cast_fp16 = add(x = var_6524_cast_fp16, y = var_6525_cast_fp16)[name = tensor("koi_65_cast_fp16")]; + tensor qo_33_axis_0 = const()[name = tensor("qo_33_axis_0"), val = tensor(-1)]; + tensor qo_33_cast_fp16 = stack(axis = qo_33_axis_0, values = (qor_65_cast_fp16, qoi_65_cast_fp16))[name = tensor("qo_33_cast_fp16")]; + tensor ko_33_axis_0 = const()[name = tensor("ko_33_axis_0"), val = tensor(-1)]; + tensor ko_33_cast_fp16 = stack(axis = ko_33_axis_0, values = (kor_65_cast_fp16, koi_65_cast_fp16))[name = tensor("ko_33_cast_fp16")]; + tensor var_6558 = const()[name = tensor("op_6558"), val = tensor([1, 1, 16, 64])]; + tensor q_99_cast_fp16 = reshape(shape = var_6558, x = qo_33_cast_fp16)[name = tensor("q_99_cast_fp16")]; + tensor var_6564 = const()[name = tensor("op_6564"), val = tensor([1, 1, 16, 64])]; + tensor k_67_cast_fp16 = reshape(shape = var_6564, x = ko_33_cast_fp16)[name = tensor("k_67_cast_fp16")]; + tensor _inversed_6589_y_0_to_fp16 = const()[name = tensor("_inversed_6589_y_0_to_fp16"), val = tensor(0x1p-9)]; + tensor _inversed_6589_cast_fp16 = mul(x = ts_101_cast_fp16, y = _inversed_6589_y_0_to_fp16)[name = tensor("_inversed_6589_cast_fp16")]; + tensor var_6590_cast_fp16 = floor(x = _inversed_6589_cast_fp16)[name = tensor("op_6590_cast_fp16")]; + tensor var_6591_to_fp16 = const()[name = tensor("op_6591_to_fp16"), val = tensor(0x1p+9)]; + tensor var_6592_cast_fp16 = mul(x = var_6590_cast_fp16, y = var_6591_to_fp16)[name = tensor("op_6592_cast_fp16")]; + tensor write_indices_float_67_cast_fp16 = sub(x = ts_101_cast_fp16, y = var_6592_cast_fp16)[name = tensor("write_indices_float_67_cast_fp16")]; + tensor var_6599_dtype_0 = const()[name = tensor("op_6599_dtype_0"), val = tensor("int32")]; + tensor write_indices_33_reps_0 = const()[name = tensor("write_indices_33_reps_0"), val = tensor([1, 1, 16, 64])]; + tensor write_indices_float_67_cast_fp16_to_int32 = cast(dtype = var_6599_dtype_0, x = write_indices_float_67_cast_fp16)[name = tensor("cast_472")]; + tensor write_indices_33 = tile(reps = write_indices_33_reps_0, x = write_indices_float_67_cast_fp16_to_int32)[name = tensor("write_indices_33")]; + tensor var_6609_begin_0 = const()[name = tensor("op_6609_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_6609_end_0 = const()[name = tensor("op_6609_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor var_6609_end_mask_0 = const()[name = tensor("op_6609_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_6609_squeeze_mask_0 = const()[name = tensor("op_6609_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor cache16_to_fp16_dtype_0 = const()[name = tensor("cache16_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor cache16_to_fp16 = cast(dtype = cache16_to_fp16_dtype_0, x = cache16)[name = tensor("cast_471")]; + tensor var_6609_cast_fp16 = slice_by_index(begin = var_6609_begin_0, end = var_6609_end_0, end_mask = var_6609_end_mask_0, squeeze_mask = var_6609_squeeze_mask_0, x = cache16_to_fp16)[name = tensor("op_6609_cast_fp16")]; + tensor var_6611_axis_0 = const()[name = tensor("op_6611_axis_0"), val = tensor(1)]; + tensor var_6611_mode_0 = const()[name = tensor("op_6611_mode_0"), val = tensor("update")]; + tensor var_6611_validate_indices_0 = const()[name = tensor("op_6611_validate_indices_0"), val = tensor(false)]; + tensor var_6611_cast_fp16 = scatter_along_axis(axis = var_6611_axis_0, data = var_6609_cast_fp16, indices = write_indices_33, mode = var_6611_mode_0, updates = k_67_cast_fp16, validate_indices = var_6611_validate_indices_0)[name = tensor("op_6611_cast_fp16")]; + tensor concat_114 = const()[name = tensor("concat_114"), val = tensor([0, 0, 0, 0, 0])]; + tensor concat_115 = const()[name = tensor("concat_115"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_33_internal_tensor_assign_1_stride_0 = const()[name = tensor("new_cache_33_internal_tensor_assign_1_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_33_internal_tensor_assign_1_begin_mask_0 = const()[name = tensor("new_cache_33_internal_tensor_assign_1_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_33_internal_tensor_assign_1_end_mask_0 = const()[name = tensor("new_cache_33_internal_tensor_assign_1_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_33_internal_tensor_assign_1_squeeze_mask_0 = const()[name = tensor("new_cache_33_internal_tensor_assign_1_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_80 = const()[name = tensor("shape_80"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_32 = const()[name = tensor("reduce_prod_32"), val = tensor(1048576)]; + tensor range_1d_32_start_0 = const()[name = tensor("range_1d_32_start_0"), val = tensor(0)]; + tensor range_1d_32_step_0 = const()[name = tensor("range_1d_32_step_0"), val = tensor(1)]; + tensor range_1d_32 = range_1d(end = reduce_prod_32, start = range_1d_32_start_0, step = range_1d_32_step_0)[name = tensor("range_1d_32")]; + tensor reshape_160 = reshape(shape = shape_80, x = range_1d_32)[name = tensor("reshape_160")]; + tensor slice_by_index_32 = slice_by_index(begin = concat_114, begin_mask = new_cache_33_internal_tensor_assign_1_begin_mask_0, end = concat_115, end_mask = new_cache_33_internal_tensor_assign_1_end_mask_0, squeeze_mask = new_cache_33_internal_tensor_assign_1_squeeze_mask_0, stride = new_cache_33_internal_tensor_assign_1_stride_0, x = reshape_160)[name = tensor("slice_by_index_32")]; + tensor reshape_161_shape_0 = const()[name = tensor("reshape_161_shape_0"), val = tensor([-1])]; + tensor reshape_161 = reshape(shape = reshape_161_shape_0, x = slice_by_index_32)[name = tensor("reshape_161")]; + tensor reshape_162_shape_0 = const()[name = tensor("reshape_162_shape_0"), val = tensor([-1])]; + tensor reshape_162_cast_fp16 = reshape(shape = reshape_162_shape_0, x = var_6611_cast_fp16)[name = tensor("reshape_162_cast_fp16")]; + tensor reshape_163_shape_0 = const()[name = tensor("reshape_163_shape_0"), val = tensor([-1])]; + tensor reshape_163_cast_fp16 = reshape(shape = reshape_163_shape_0, x = cache16_to_fp16)[name = tensor("reshape_163_cast_fp16")]; + tensor scatter_32_mode_0 = const()[name = tensor("scatter_32_mode_0"), val = tensor("update")]; + tensor scatter_32_axis_0 = const()[name = tensor("scatter_32_axis_0"), val = tensor(0)]; + tensor scatter_32_validate_indices_0 = const()[name = tensor("scatter_32_validate_indices_0"), val = tensor(false)]; + tensor scatter_32_cast_fp16 = scatter(axis = scatter_32_axis_0, data = reshape_163_cast_fp16, indices = reshape_161, mode = scatter_32_mode_0, updates = reshape_162_cast_fp16, validate_indices = scatter_32_validate_indices_0)[name = tensor("scatter_32_cast_fp16")]; + tensor reshape_164_cast_fp16 = reshape(shape = shape_80, x = scatter_32_cast_fp16)[name = tensor("reshape_164_cast_fp16")]; + tensor var_6619_begin_0 = const()[name = tensor("op_6619_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor var_6619_end_0 = const()[name = tensor("op_6619_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor var_6619_end_mask_0 = const()[name = tensor("op_6619_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_6619_squeeze_mask_0 = const()[name = tensor("op_6619_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor var_6619_cast_fp16 = slice_by_index(begin = var_6619_begin_0, end = var_6619_end_0, end_mask = var_6619_end_mask_0, squeeze_mask = var_6619_squeeze_mask_0, x = reshape_164_cast_fp16)[name = tensor("op_6619_cast_fp16")]; + tensor var_6621_axis_0 = const()[name = tensor("op_6621_axis_0"), val = tensor(1)]; + tensor var_6621_mode_0 = const()[name = tensor("op_6621_mode_0"), val = tensor("update")]; + tensor var_6621_validate_indices_0 = const()[name = tensor("op_6621_validate_indices_0"), val = tensor(false)]; + tensor var_6621_cast_fp16 = scatter_along_axis(axis = var_6621_axis_0, data = var_6619_cast_fp16, indices = write_indices_33, mode = var_6621_mode_0, updates = v_33_cast_fp16, validate_indices = var_6621_validate_indices_0)[name = tensor("op_6621_cast_fp16")]; + tensor concat_116 = const()[name = tensor("concat_116"), val = tensor([1, 0, 0, 0, 0])]; + tensor concat_117 = const()[name = tensor("concat_117"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_33_internal_tensor_assign_2_stride_0 = const()[name = tensor("new_cache_33_internal_tensor_assign_2_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_33_internal_tensor_assign_2_begin_mask_0 = const()[name = tensor("new_cache_33_internal_tensor_assign_2_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_33_internal_tensor_assign_2_end_mask_0 = const()[name = tensor("new_cache_33_internal_tensor_assign_2_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_33_internal_tensor_assign_2_squeeze_mask_0 = const()[name = tensor("new_cache_33_internal_tensor_assign_2_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_81 = const()[name = tensor("shape_81"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_33 = const()[name = tensor("reduce_prod_33"), val = tensor(1048576)]; + tensor range_1d_33_start_0 = const()[name = tensor("range_1d_33_start_0"), val = tensor(0)]; + tensor range_1d_33_step_0 = const()[name = tensor("range_1d_33_step_0"), val = tensor(1)]; + tensor range_1d_33 = range_1d(end = reduce_prod_33, start = range_1d_33_start_0, step = range_1d_33_step_0)[name = tensor("range_1d_33")]; + tensor reshape_165 = reshape(shape = shape_81, x = range_1d_33)[name = tensor("reshape_165")]; + tensor slice_by_index_33 = slice_by_index(begin = concat_116, begin_mask = new_cache_33_internal_tensor_assign_2_begin_mask_0, end = concat_117, end_mask = new_cache_33_internal_tensor_assign_2_end_mask_0, squeeze_mask = new_cache_33_internal_tensor_assign_2_squeeze_mask_0, stride = new_cache_33_internal_tensor_assign_2_stride_0, x = reshape_165)[name = tensor("slice_by_index_33")]; + tensor reshape_166_shape_0 = const()[name = tensor("reshape_166_shape_0"), val = tensor([-1])]; + tensor reshape_166 = reshape(shape = reshape_166_shape_0, x = slice_by_index_33)[name = tensor("reshape_166")]; + tensor reshape_167_shape_0 = const()[name = tensor("reshape_167_shape_0"), val = tensor([-1])]; + tensor reshape_167_cast_fp16 = reshape(shape = reshape_167_shape_0, x = var_6621_cast_fp16)[name = tensor("reshape_167_cast_fp16")]; + tensor reshape_168_shape_0 = const()[name = tensor("reshape_168_shape_0"), val = tensor([-1])]; + tensor reshape_168_cast_fp16 = reshape(shape = reshape_168_shape_0, x = reshape_164_cast_fp16)[name = tensor("reshape_168_cast_fp16")]; + tensor scatter_33_mode_0 = const()[name = tensor("scatter_33_mode_0"), val = tensor("update")]; + tensor scatter_33_axis_0 = const()[name = tensor("scatter_33_axis_0"), val = tensor(0)]; + tensor scatter_33_validate_indices_0 = const()[name = tensor("scatter_33_validate_indices_0"), val = tensor(false)]; + tensor scatter_33_cast_fp16 = scatter(axis = scatter_33_axis_0, data = reshape_168_cast_fp16, indices = reshape_166, mode = scatter_33_mode_0, updates = reshape_167_cast_fp16, validate_indices = scatter_33_validate_indices_0)[name = tensor("scatter_33_cast_fp16")]; + tensor reshape_169_cast_fp16 = reshape(shape = shape_81, x = scatter_33_cast_fp16)[name = tensor("reshape_169_cast_fp16")]; + tensor reshape_169_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("reshape_169_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor keys_97_begin_0 = const()[name = tensor("keys_97_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor keys_97_end_0 = const()[name = tensor("keys_97_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor keys_97_end_mask_0 = const()[name = tensor("keys_97_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor keys_97_squeeze_mask_0 = const()[name = tensor("keys_97_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor keys_97_cast_fp16 = slice_by_index(begin = keys_97_begin_0, end = keys_97_end_0, end_mask = keys_97_end_mask_0, squeeze_mask = keys_97_squeeze_mask_0, x = reshape_169_cast_fp16)[name = tensor("keys_97_cast_fp16")]; + tensor values_97_begin_0 = const()[name = tensor("values_97_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor values_97_end_0 = const()[name = tensor("values_97_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor values_97_end_mask_0 = const()[name = tensor("values_97_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor values_97_squeeze_mask_0 = const()[name = tensor("values_97_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor values_97_cast_fp16 = slice_by_index(begin = values_97_begin_0, end = values_97_end_0, end_mask = values_97_end_mask_0, squeeze_mask = values_97_squeeze_mask_0, x = reshape_169_cast_fp16)[name = tensor("values_97_cast_fp16")]; + tensor var_6633_cast_fp16 = not_equal(x = keys_97_cast_fp16, y = keys_97_cast_fp16)[name = tensor("op_6633_cast_fp16")]; + tensor keys_99_cast_fp16 = select(a = var_463_to_fp16, b = keys_97_cast_fp16, cond = var_6633_cast_fp16)[name = tensor("keys_99_cast_fp16")]; + tensor var_6641_cast_fp16 = not_equal(x = values_97_cast_fp16, y = values_97_cast_fp16)[name = tensor("op_6641_cast_fp16")]; + tensor values_99_cast_fp16 = select(a = var_463_to_fp16, b = values_97_cast_fp16, cond = var_6641_cast_fp16)[name = tensor("values_99_cast_fp16")]; + tensor var_6665 = const()[name = tensor("op_6665"), val = tensor([0, 2, 1, 3])]; + tensor var_6681 = const()[name = tensor("op_6681"), val = tensor([1, 1, 1])]; + tensor var_6682_cast_fp16 = reshape(shape = var_6681, x = position16_to_fp16)[name = tensor("op_6682_cast_fp16")]; + tensor var_6702_to_fp16 = const()[name = tensor("op_6702_to_fp16"), val = tensor(0x1p+0)]; + tensor valid_len_33_cast_fp16 = add(x = var_6682_cast_fp16, y = var_6702_to_fp16)[name = tensor("valid_len_33_cast_fp16")]; + tensor valid_mask_33_cast_fp16 = less(x = k_positions_1_promoted_to_fp16, y = valid_len_33_cast_fp16)[name = tensor("valid_mask_33_cast_fp16")]; + tensor causal_mask_33_cast_fp16 = less_equal(x = k_positions_1_promoted_to_fp16, y = var_6682_cast_fp16)[name = tensor("causal_mask_33_cast_fp16")]; + tensor attn_mask_65 = logical_and(x = valid_mask_33_cast_fp16, y = causal_mask_33_cast_fp16)[name = tensor("attn_mask_65")]; + tensor attn_mask_67_axes_0 = const()[name = tensor("attn_mask_67_axes_0"), val = tensor([1])]; + tensor attn_mask_67 = expand_dims(axes = attn_mask_67_axes_0, x = attn_mask_65)[name = tensor("attn_mask_67")]; + tensor var_6720_transpose_x_0 = const()[name = tensor("op_6720_transpose_x_0"), val = tensor(false)]; + tensor var_6720_transpose_y_0 = const()[name = tensor("op_6720_transpose_y_0"), val = tensor(false)]; + tensor transpose_104_perm_0 = const()[name = tensor("transpose_104_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_105_perm_0 = const()[name = tensor("transpose_105_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_105 = transpose(perm = transpose_105_perm_0, x = keys_99_cast_fp16)[name = tensor("transpose_149")]; + tensor transpose_104 = transpose(perm = transpose_104_perm_0, x = q_99_cast_fp16)[name = tensor("transpose_150")]; + tensor var_6720_cast_fp16 = matmul(transpose_x = var_6720_transpose_x_0, transpose_y = var_6720_transpose_y_0, x = transpose_104, y = transpose_105)[name = tensor("op_6720_cast_fp16")]; + tensor var_6714_to_fp16 = const()[name = tensor("op_6714_to_fp16"), val = tensor([0x1p-3])]; + tensor attn_weights_97_cast_fp16 = mul(x = var_6720_cast_fp16, y = var_6714_to_fp16)[name = tensor("attn_weights_97_cast_fp16")]; + tensor var_6722 = logical_not(x = attn_mask_67)[name = tensor("op_6722")]; + tensor var_6723_to_fp16 = const()[name = tensor("op_6723_to_fp16"), val = tensor(-inf)]; + tensor attn_weights_99_cast_fp16 = select(a = var_6723_to_fp16, b = attn_weights_97_cast_fp16, cond = var_6722)[name = tensor("attn_weights_99_cast_fp16")]; + tensor var_6725 = const()[name = tensor("op_6725"), val = tensor(-1)]; + tensor attn_weights_101_cast_fp16 = softmax(axis = var_6725, x = attn_weights_99_cast_fp16)[name = tensor("attn_weights_101_cast_fp16")]; + tensor attn_output_33_transpose_x_0 = const()[name = tensor("attn_output_33_transpose_x_0"), val = tensor(false)]; + tensor attn_output_33_transpose_y_0 = const()[name = tensor("attn_output_33_transpose_y_0"), val = tensor(false)]; + tensor values_101_cast_fp16 = transpose(perm = var_6665, x = values_99_cast_fp16)[name = tensor("transpose_151")]; + tensor attn_output_33_cast_fp16 = matmul(transpose_x = attn_output_33_transpose_x_0, transpose_y = attn_output_33_transpose_y_0, x = attn_weights_101_cast_fp16, y = values_101_cast_fp16)[name = tensor("attn_output_33_cast_fp16")]; + tensor var_6733 = const()[name = tensor("op_6733"), val = tensor([0, 2, 1, 3])]; + tensor var_6738 = const()[name = tensor("op_6738"), val = tensor([1, 1, 1024])]; + tensor var_6734_cast_fp16 = transpose(perm = var_6733, x = attn_output_33_cast_fp16)[name = tensor("transpose_148")]; + tensor input_199_cast_fp16 = reshape(shape = var_6738, x = var_6734_cast_fp16)[name = tensor("input_199_cast_fp16")]; + tensor attn16_out_proj_weight_to_fp16 = const()[name = tensor("attn16_out_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(410222336)))]; + tensor linear_66_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = attn16_out_proj_weight_to_fp16, x = input_199_cast_fp16)[name = tensor("linear_66_cast_fp16")]; + tensor var_6744_to_fp16 = const()[name = tensor("op_6744_to_fp16"), val = tensor(0x1p+0)]; + tensor var_6745_cast_fp16 = add(x = position16_to_fp16, y = var_6744_to_fp16)[name = tensor("op_6745_cast_fp16")]; + tensor var_6745_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_6745_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor input_201_cast_fp16 = add(x = input_195_cast_fp16, y = linear_66_cast_fp16)[name = tensor("input_201_cast_fp16")]; + tensor input_203_axes_0 = const()[name = tensor("input_203_axes_0"), val = tensor([-1])]; + tensor norm16_2_weight_to_fp16 = const()[name = tensor("norm16_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(412319552)))]; + tensor norm16_2_bias_to_fp16 = const()[name = tensor("norm16_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(412321664)))]; + tensor var_6749_to_fp16 = const()[name = tensor("op_6749_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_203_cast_fp16 = layer_norm(axes = input_203_axes_0, beta = norm16_2_bias_to_fp16, epsilon = var_6749_to_fp16, gamma = norm16_2_weight_to_fp16, x = input_201_cast_fp16)[name = tensor("input_203_cast_fp16")]; + tensor linear16_1_weight_to_fp16 = const()[name = tensor("linear16_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(412323776)))]; + tensor linear_67_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = linear16_1_weight_to_fp16, x = input_203_cast_fp16)[name = tensor("linear_67_cast_fp16")]; + tensor input_205_mode_0 = const()[name = tensor("input_205_mode_0"), val = tensor("EXACT")]; + tensor input_205_cast_fp16 = gelu(mode = input_205_mode_0, x = linear_67_cast_fp16)[name = tensor("input_205_cast_fp16")]; + tensor linear16_2_weight_to_fp16 = const()[name = tensor("linear16_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(420712448)))]; + tensor linear_68_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = linear16_2_weight_to_fp16, x = input_205_cast_fp16)[name = tensor("linear_68_cast_fp16")]; + tensor input_207_cast_fp16 = add(x = input_201_cast_fp16, y = linear_68_cast_fp16)[name = tensor("input_207_cast_fp16")]; + tensor input_209_axes_0 = const()[name = tensor("input_209_axes_0"), val = tensor([-1])]; + tensor norm17_1_weight_to_fp16 = const()[name = tensor("norm17_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(429101120)))]; + tensor norm17_1_bias_to_fp16 = const()[name = tensor("norm17_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(429103232)))]; + tensor var_6766_to_fp16 = const()[name = tensor("op_6766_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_209_cast_fp16 = layer_norm(axes = input_209_axes_0, beta = norm17_1_bias_to_fp16, epsilon = var_6766_to_fp16, gamma = norm17_1_weight_to_fp16, x = input_207_cast_fp16)[name = tensor("input_209_cast_fp16")]; + tensor attn17_in_proj_weight_to_fp16 = const()[name = tensor("attn17_in_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(429105344)))]; + tensor linear_69_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = attn17_in_proj_weight_to_fp16, x = input_209_cast_fp16)[name = tensor("linear_69_cast_fp16")]; + tensor var_6780 = const()[name = tensor("op_6780"), val = tensor([1, 1, 3, 16, 64])]; + tensor qkv_35_cast_fp16 = reshape(shape = var_6780, x = linear_69_cast_fp16)[name = tensor("qkv_35_cast_fp16")]; + tensor q_103_begin_0 = const()[name = tensor("q_103_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor q_103_end_0 = const()[name = tensor("q_103_end_0"), val = tensor([1, 1, 1, 16, 64])]; + tensor q_103_end_mask_0 = const()[name = tensor("q_103_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor q_103_squeeze_mask_0 = const()[name = tensor("q_103_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor q_103_cast_fp16 = slice_by_index(begin = q_103_begin_0, end = q_103_end_0, end_mask = q_103_end_mask_0, squeeze_mask = q_103_squeeze_mask_0, x = qkv_35_cast_fp16)[name = tensor("q_103_cast_fp16")]; + tensor k_69_begin_0 = const()[name = tensor("k_69_begin_0"), val = tensor([0, 0, 1, 0, 0])]; + tensor k_69_end_0 = const()[name = tensor("k_69_end_0"), val = tensor([1, 1, 2, 16, 64])]; + tensor k_69_end_mask_0 = const()[name = tensor("k_69_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor k_69_squeeze_mask_0 = const()[name = tensor("k_69_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor k_69_cast_fp16 = slice_by_index(begin = k_69_begin_0, end = k_69_end_0, end_mask = k_69_end_mask_0, squeeze_mask = k_69_squeeze_mask_0, x = qkv_35_cast_fp16)[name = tensor("k_69_cast_fp16")]; + tensor v_35_begin_0 = const()[name = tensor("v_35_begin_0"), val = tensor([0, 0, 2, 0, 0])]; + tensor v_35_end_0 = const()[name = tensor("v_35_end_0"), val = tensor([1, 1, 3, 16, 64])]; + tensor v_35_end_mask_0 = const()[name = tensor("v_35_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor v_35_squeeze_mask_0 = const()[name = tensor("v_35_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor v_35_cast_fp16 = slice_by_index(begin = v_35_begin_0, end = v_35_end_0, end_mask = v_35_end_mask_0, squeeze_mask = v_35_squeeze_mask_0, x = qkv_35_cast_fp16)[name = tensor("v_35_cast_fp16")]; + tensor var_6846 = const()[name = tensor("op_6846"), val = tensor([1, 1, 1, 1])]; + tensor position17_to_fp16_dtype_0 = const()[name = tensor("position17_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor position17_to_fp16 = cast(dtype = position17_to_fp16_dtype_0, x = position17)[name = tensor("cast_468")]; + tensor ts_107_cast_fp16 = reshape(shape = var_6846, x = position17_to_fp16)[name = tensor("ts_107_cast_fp16")]; + tensor var_6853 = const()[name = tensor("op_6853"), val = tensor([1, 1, 16, 32, 2])]; + tensor q_complex_35_cast_fp16 = reshape(shape = var_6853, x = q_103_cast_fp16)[name = tensor("q_complex_35_cast_fp16")]; + tensor var_6860 = const()[name = tensor("op_6860"), val = tensor([1, 1, 16, 32, 2])]; + tensor k_complex_35_cast_fp16 = reshape(shape = var_6860, x = k_69_cast_fp16)[name = tensor("k_complex_35_cast_fp16")]; + tensor var_6864_begin_0 = const()[name = tensor("op_6864_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_6864_end_0 = const()[name = tensor("op_6864_end_0"), val = tensor([1, 1, 16, 32, 1])]; + tensor var_6864_end_mask_0 = const()[name = tensor("op_6864_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_6864_squeeze_mask_0 = const()[name = tensor("op_6864_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_6864_cast_fp16 = slice_by_index(begin = var_6864_begin_0, end = var_6864_end_0, end_mask = var_6864_end_mask_0, squeeze_mask = var_6864_squeeze_mask_0, x = q_complex_35_cast_fp16)[name = tensor("op_6864_cast_fp16")]; + tensor var_6872_begin_0 = const()[name = tensor("op_6872_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_6872_end_0 = const()[name = tensor("op_6872_end_0"), val = tensor([1, 1, 16, 32, 2])]; + tensor var_6872_end_mask_0 = const()[name = tensor("op_6872_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_6872_squeeze_mask_0 = const()[name = tensor("op_6872_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_6872_cast_fp16 = slice_by_index(begin = var_6872_begin_0, end = var_6872_end_0, end_mask = var_6872_end_mask_0, squeeze_mask = var_6872_squeeze_mask_0, x = q_complex_35_cast_fp16)[name = tensor("op_6872_cast_fp16")]; + tensor var_6880_begin_0 = const()[name = tensor("op_6880_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_6880_end_0 = const()[name = tensor("op_6880_end_0"), val = tensor([1, 1, 16, 32, 1])]; + tensor var_6880_end_mask_0 = const()[name = tensor("op_6880_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_6880_squeeze_mask_0 = const()[name = tensor("op_6880_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_6880_cast_fp16 = slice_by_index(begin = var_6880_begin_0, end = var_6880_end_0, end_mask = var_6880_end_mask_0, squeeze_mask = var_6880_squeeze_mask_0, x = k_complex_35_cast_fp16)[name = tensor("op_6880_cast_fp16")]; + tensor var_6888_begin_0 = const()[name = tensor("op_6888_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_6888_end_0 = const()[name = tensor("op_6888_end_0"), val = tensor([1, 1, 16, 32, 2])]; + tensor var_6888_end_mask_0 = const()[name = tensor("op_6888_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_6888_squeeze_mask_0 = const()[name = tensor("op_6888_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_6888_cast_fp16 = slice_by_index(begin = var_6888_begin_0, end = var_6888_end_0, end_mask = var_6888_end_mask_0, squeeze_mask = var_6888_squeeze_mask_0, x = k_complex_35_cast_fp16)[name = tensor("op_6888_cast_fp16")]; + tensor freqs_35_to_fp16 = const()[name = tensor("freqs_35_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(435396864)))]; + tensor var_6894_cast_fp16 = mul(x = freqs_35_to_fp16, y = ts_107_cast_fp16)[name = tensor("op_6894_cast_fp16")]; + tensor rotr_35_cast_fp16 = cos(x = var_6894_cast_fp16)[name = tensor("rotr_35_cast_fp16")]; + tensor roti_35_cast_fp16 = sin(x = var_6894_cast_fp16)[name = tensor("roti_35_cast_fp16")]; + tensor var_6898_cast_fp16 = mul(x = var_6864_cast_fp16, y = rotr_35_cast_fp16)[name = tensor("op_6898_cast_fp16")]; + tensor var_6899_cast_fp16 = mul(x = var_6872_cast_fp16, y = roti_35_cast_fp16)[name = tensor("op_6899_cast_fp16")]; + tensor qor_69_cast_fp16 = sub(x = var_6898_cast_fp16, y = var_6899_cast_fp16)[name = tensor("qor_69_cast_fp16")]; + tensor var_6902_cast_fp16 = mul(x = var_6864_cast_fp16, y = roti_35_cast_fp16)[name = tensor("op_6902_cast_fp16")]; + tensor var_6903_cast_fp16 = mul(x = var_6872_cast_fp16, y = rotr_35_cast_fp16)[name = tensor("op_6903_cast_fp16")]; + tensor qoi_69_cast_fp16 = add(x = var_6902_cast_fp16, y = var_6903_cast_fp16)[name = tensor("qoi_69_cast_fp16")]; + tensor var_6906_cast_fp16 = mul(x = var_6880_cast_fp16, y = rotr_35_cast_fp16)[name = tensor("op_6906_cast_fp16")]; + tensor var_6907_cast_fp16 = mul(x = var_6888_cast_fp16, y = roti_35_cast_fp16)[name = tensor("op_6907_cast_fp16")]; + tensor kor_69_cast_fp16 = sub(x = var_6906_cast_fp16, y = var_6907_cast_fp16)[name = tensor("kor_69_cast_fp16")]; + tensor var_6910_cast_fp16 = mul(x = var_6880_cast_fp16, y = roti_35_cast_fp16)[name = tensor("op_6910_cast_fp16")]; + tensor var_6911_cast_fp16 = mul(x = var_6888_cast_fp16, y = rotr_35_cast_fp16)[name = tensor("op_6911_cast_fp16")]; + tensor koi_69_cast_fp16 = add(x = var_6910_cast_fp16, y = var_6911_cast_fp16)[name = tensor("koi_69_cast_fp16")]; + tensor qo_35_axis_0 = const()[name = tensor("qo_35_axis_0"), val = tensor(-1)]; + tensor qo_35_cast_fp16 = stack(axis = qo_35_axis_0, values = (qor_69_cast_fp16, qoi_69_cast_fp16))[name = tensor("qo_35_cast_fp16")]; + tensor ko_35_axis_0 = const()[name = tensor("ko_35_axis_0"), val = tensor(-1)]; + tensor ko_35_cast_fp16 = stack(axis = ko_35_axis_0, values = (kor_69_cast_fp16, koi_69_cast_fp16))[name = tensor("ko_35_cast_fp16")]; + tensor var_6944 = const()[name = tensor("op_6944"), val = tensor([1, 1, 16, 64])]; + tensor q_105_cast_fp16 = reshape(shape = var_6944, x = qo_35_cast_fp16)[name = tensor("q_105_cast_fp16")]; + tensor var_6950 = const()[name = tensor("op_6950"), val = tensor([1, 1, 16, 64])]; + tensor k_71_cast_fp16 = reshape(shape = var_6950, x = ko_35_cast_fp16)[name = tensor("k_71_cast_fp16")]; + tensor _inversed_6975_y_0_to_fp16 = const()[name = tensor("_inversed_6975_y_0_to_fp16"), val = tensor(0x1p-9)]; + tensor _inversed_6975_cast_fp16 = mul(x = ts_107_cast_fp16, y = _inversed_6975_y_0_to_fp16)[name = tensor("_inversed_6975_cast_fp16")]; + tensor var_6976_cast_fp16 = floor(x = _inversed_6975_cast_fp16)[name = tensor("op_6976_cast_fp16")]; + tensor var_6977_to_fp16 = const()[name = tensor("op_6977_to_fp16"), val = tensor(0x1p+9)]; + tensor var_6978_cast_fp16 = mul(x = var_6976_cast_fp16, y = var_6977_to_fp16)[name = tensor("op_6978_cast_fp16")]; + tensor write_indices_float_71_cast_fp16 = sub(x = ts_107_cast_fp16, y = var_6978_cast_fp16)[name = tensor("write_indices_float_71_cast_fp16")]; + tensor var_6985_dtype_0 = const()[name = tensor("op_6985_dtype_0"), val = tensor("int32")]; + tensor write_indices_35_reps_0 = const()[name = tensor("write_indices_35_reps_0"), val = tensor([1, 1, 16, 64])]; + tensor write_indices_float_71_cast_fp16_to_int32 = cast(dtype = var_6985_dtype_0, x = write_indices_float_71_cast_fp16)[name = tensor("cast_467")]; + tensor write_indices_35 = tile(reps = write_indices_35_reps_0, x = write_indices_float_71_cast_fp16_to_int32)[name = tensor("write_indices_35")]; + tensor var_6995_begin_0 = const()[name = tensor("op_6995_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_6995_end_0 = const()[name = tensor("op_6995_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor var_6995_end_mask_0 = const()[name = tensor("op_6995_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_6995_squeeze_mask_0 = const()[name = tensor("op_6995_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor cache17_to_fp16_dtype_0 = const()[name = tensor("cache17_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor cache17_to_fp16 = cast(dtype = cache17_to_fp16_dtype_0, x = cache17)[name = tensor("cast_466")]; + tensor var_6995_cast_fp16 = slice_by_index(begin = var_6995_begin_0, end = var_6995_end_0, end_mask = var_6995_end_mask_0, squeeze_mask = var_6995_squeeze_mask_0, x = cache17_to_fp16)[name = tensor("op_6995_cast_fp16")]; + tensor var_6997_axis_0 = const()[name = tensor("op_6997_axis_0"), val = tensor(1)]; + tensor var_6997_mode_0 = const()[name = tensor("op_6997_mode_0"), val = tensor("update")]; + tensor var_6997_validate_indices_0 = const()[name = tensor("op_6997_validate_indices_0"), val = tensor(false)]; + tensor var_6997_cast_fp16 = scatter_along_axis(axis = var_6997_axis_0, data = var_6995_cast_fp16, indices = write_indices_35, mode = var_6997_mode_0, updates = k_71_cast_fp16, validate_indices = var_6997_validate_indices_0)[name = tensor("op_6997_cast_fp16")]; + tensor concat_121 = const()[name = tensor("concat_121"), val = tensor([0, 0, 0, 0, 0])]; + tensor concat_122 = const()[name = tensor("concat_122"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_35_internal_tensor_assign_1_stride_0 = const()[name = tensor("new_cache_35_internal_tensor_assign_1_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_35_internal_tensor_assign_1_begin_mask_0 = const()[name = tensor("new_cache_35_internal_tensor_assign_1_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_35_internal_tensor_assign_1_end_mask_0 = const()[name = tensor("new_cache_35_internal_tensor_assign_1_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_35_internal_tensor_assign_1_squeeze_mask_0 = const()[name = tensor("new_cache_35_internal_tensor_assign_1_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_82 = const()[name = tensor("shape_82"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_34 = const()[name = tensor("reduce_prod_34"), val = tensor(1048576)]; + tensor range_1d_34_start_0 = const()[name = tensor("range_1d_34_start_0"), val = tensor(0)]; + tensor range_1d_34_step_0 = const()[name = tensor("range_1d_34_step_0"), val = tensor(1)]; + tensor range_1d_34 = range_1d(end = reduce_prod_34, start = range_1d_34_start_0, step = range_1d_34_step_0)[name = tensor("range_1d_34")]; + tensor reshape_170 = reshape(shape = shape_82, x = range_1d_34)[name = tensor("reshape_170")]; + tensor slice_by_index_34 = slice_by_index(begin = concat_121, begin_mask = new_cache_35_internal_tensor_assign_1_begin_mask_0, end = concat_122, end_mask = new_cache_35_internal_tensor_assign_1_end_mask_0, squeeze_mask = new_cache_35_internal_tensor_assign_1_squeeze_mask_0, stride = new_cache_35_internal_tensor_assign_1_stride_0, x = reshape_170)[name = tensor("slice_by_index_34")]; + tensor reshape_171_shape_0 = const()[name = tensor("reshape_171_shape_0"), val = tensor([-1])]; + tensor reshape_171 = reshape(shape = reshape_171_shape_0, x = slice_by_index_34)[name = tensor("reshape_171")]; + tensor reshape_172_shape_0 = const()[name = tensor("reshape_172_shape_0"), val = tensor([-1])]; + tensor reshape_172_cast_fp16 = reshape(shape = reshape_172_shape_0, x = var_6997_cast_fp16)[name = tensor("reshape_172_cast_fp16")]; + tensor reshape_173_shape_0 = const()[name = tensor("reshape_173_shape_0"), val = tensor([-1])]; + tensor reshape_173_cast_fp16 = reshape(shape = reshape_173_shape_0, x = cache17_to_fp16)[name = tensor("reshape_173_cast_fp16")]; + tensor scatter_34_mode_0 = const()[name = tensor("scatter_34_mode_0"), val = tensor("update")]; + tensor scatter_34_axis_0 = const()[name = tensor("scatter_34_axis_0"), val = tensor(0)]; + tensor scatter_34_validate_indices_0 = const()[name = tensor("scatter_34_validate_indices_0"), val = tensor(false)]; + tensor scatter_34_cast_fp16 = scatter(axis = scatter_34_axis_0, data = reshape_173_cast_fp16, indices = reshape_171, mode = scatter_34_mode_0, updates = reshape_172_cast_fp16, validate_indices = scatter_34_validate_indices_0)[name = tensor("scatter_34_cast_fp16")]; + tensor reshape_174_cast_fp16 = reshape(shape = shape_82, x = scatter_34_cast_fp16)[name = tensor("reshape_174_cast_fp16")]; + tensor var_7005_begin_0 = const()[name = tensor("op_7005_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor var_7005_end_0 = const()[name = tensor("op_7005_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor var_7005_end_mask_0 = const()[name = tensor("op_7005_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_7005_squeeze_mask_0 = const()[name = tensor("op_7005_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor var_7005_cast_fp16 = slice_by_index(begin = var_7005_begin_0, end = var_7005_end_0, end_mask = var_7005_end_mask_0, squeeze_mask = var_7005_squeeze_mask_0, x = reshape_174_cast_fp16)[name = tensor("op_7005_cast_fp16")]; + tensor var_7007_axis_0 = const()[name = tensor("op_7007_axis_0"), val = tensor(1)]; + tensor var_7007_mode_0 = const()[name = tensor("op_7007_mode_0"), val = tensor("update")]; + tensor var_7007_validate_indices_0 = const()[name = tensor("op_7007_validate_indices_0"), val = tensor(false)]; + tensor var_7007_cast_fp16 = scatter_along_axis(axis = var_7007_axis_0, data = var_7005_cast_fp16, indices = write_indices_35, mode = var_7007_mode_0, updates = v_35_cast_fp16, validate_indices = var_7007_validate_indices_0)[name = tensor("op_7007_cast_fp16")]; + tensor concat_123 = const()[name = tensor("concat_123"), val = tensor([1, 0, 0, 0, 0])]; + tensor concat_124 = const()[name = tensor("concat_124"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_35_internal_tensor_assign_2_stride_0 = const()[name = tensor("new_cache_35_internal_tensor_assign_2_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_35_internal_tensor_assign_2_begin_mask_0 = const()[name = tensor("new_cache_35_internal_tensor_assign_2_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_35_internal_tensor_assign_2_end_mask_0 = const()[name = tensor("new_cache_35_internal_tensor_assign_2_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_35_internal_tensor_assign_2_squeeze_mask_0 = const()[name = tensor("new_cache_35_internal_tensor_assign_2_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_83 = const()[name = tensor("shape_83"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_35 = const()[name = tensor("reduce_prod_35"), val = tensor(1048576)]; + tensor range_1d_35_start_0 = const()[name = tensor("range_1d_35_start_0"), val = tensor(0)]; + tensor range_1d_35_step_0 = const()[name = tensor("range_1d_35_step_0"), val = tensor(1)]; + tensor range_1d_35 = range_1d(end = reduce_prod_35, start = range_1d_35_start_0, step = range_1d_35_step_0)[name = tensor("range_1d_35")]; + tensor reshape_175 = reshape(shape = shape_83, x = range_1d_35)[name = tensor("reshape_175")]; + tensor slice_by_index_35 = slice_by_index(begin = concat_123, begin_mask = new_cache_35_internal_tensor_assign_2_begin_mask_0, end = concat_124, end_mask = new_cache_35_internal_tensor_assign_2_end_mask_0, squeeze_mask = new_cache_35_internal_tensor_assign_2_squeeze_mask_0, stride = new_cache_35_internal_tensor_assign_2_stride_0, x = reshape_175)[name = tensor("slice_by_index_35")]; + tensor reshape_176_shape_0 = const()[name = tensor("reshape_176_shape_0"), val = tensor([-1])]; + tensor reshape_176 = reshape(shape = reshape_176_shape_0, x = slice_by_index_35)[name = tensor("reshape_176")]; + tensor reshape_177_shape_0 = const()[name = tensor("reshape_177_shape_0"), val = tensor([-1])]; + tensor reshape_177_cast_fp16 = reshape(shape = reshape_177_shape_0, x = var_7007_cast_fp16)[name = tensor("reshape_177_cast_fp16")]; + tensor reshape_178_shape_0 = const()[name = tensor("reshape_178_shape_0"), val = tensor([-1])]; + tensor reshape_178_cast_fp16 = reshape(shape = reshape_178_shape_0, x = reshape_174_cast_fp16)[name = tensor("reshape_178_cast_fp16")]; + tensor scatter_35_mode_0 = const()[name = tensor("scatter_35_mode_0"), val = tensor("update")]; + tensor scatter_35_axis_0 = const()[name = tensor("scatter_35_axis_0"), val = tensor(0)]; + tensor scatter_35_validate_indices_0 = const()[name = tensor("scatter_35_validate_indices_0"), val = tensor(false)]; + tensor scatter_35_cast_fp16 = scatter(axis = scatter_35_axis_0, data = reshape_178_cast_fp16, indices = reshape_176, mode = scatter_35_mode_0, updates = reshape_177_cast_fp16, validate_indices = scatter_35_validate_indices_0)[name = tensor("scatter_35_cast_fp16")]; + tensor reshape_179_cast_fp16 = reshape(shape = shape_83, x = scatter_35_cast_fp16)[name = tensor("reshape_179_cast_fp16")]; + tensor reshape_179_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("reshape_179_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor keys_103_begin_0 = const()[name = tensor("keys_103_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor keys_103_end_0 = const()[name = tensor("keys_103_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor keys_103_end_mask_0 = const()[name = tensor("keys_103_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor keys_103_squeeze_mask_0 = const()[name = tensor("keys_103_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor keys_103_cast_fp16 = slice_by_index(begin = keys_103_begin_0, end = keys_103_end_0, end_mask = keys_103_end_mask_0, squeeze_mask = keys_103_squeeze_mask_0, x = reshape_179_cast_fp16)[name = tensor("keys_103_cast_fp16")]; + tensor values_103_begin_0 = const()[name = tensor("values_103_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor values_103_end_0 = const()[name = tensor("values_103_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor values_103_end_mask_0 = const()[name = tensor("values_103_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor values_103_squeeze_mask_0 = const()[name = tensor("values_103_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor values_103_cast_fp16 = slice_by_index(begin = values_103_begin_0, end = values_103_end_0, end_mask = values_103_end_mask_0, squeeze_mask = values_103_squeeze_mask_0, x = reshape_179_cast_fp16)[name = tensor("values_103_cast_fp16")]; + tensor var_7019_cast_fp16 = not_equal(x = keys_103_cast_fp16, y = keys_103_cast_fp16)[name = tensor("op_7019_cast_fp16")]; + tensor keys_105_cast_fp16 = select(a = var_463_to_fp16, b = keys_103_cast_fp16, cond = var_7019_cast_fp16)[name = tensor("keys_105_cast_fp16")]; + tensor var_7027_cast_fp16 = not_equal(x = values_103_cast_fp16, y = values_103_cast_fp16)[name = tensor("op_7027_cast_fp16")]; + tensor values_105_cast_fp16 = select(a = var_463_to_fp16, b = values_103_cast_fp16, cond = var_7027_cast_fp16)[name = tensor("values_105_cast_fp16")]; + tensor var_7051 = const()[name = tensor("op_7051"), val = tensor([0, 2, 1, 3])]; + tensor var_7067 = const()[name = tensor("op_7067"), val = tensor([1, 1, 1])]; + tensor var_7068_cast_fp16 = reshape(shape = var_7067, x = position17_to_fp16)[name = tensor("op_7068_cast_fp16")]; + tensor var_7088_to_fp16 = const()[name = tensor("op_7088_to_fp16"), val = tensor(0x1p+0)]; + tensor valid_len_35_cast_fp16 = add(x = var_7068_cast_fp16, y = var_7088_to_fp16)[name = tensor("valid_len_35_cast_fp16")]; + tensor valid_mask_35_cast_fp16 = less(x = k_positions_1_promoted_to_fp16, y = valid_len_35_cast_fp16)[name = tensor("valid_mask_35_cast_fp16")]; + tensor causal_mask_35_cast_fp16 = less_equal(x = k_positions_1_promoted_to_fp16, y = var_7068_cast_fp16)[name = tensor("causal_mask_35_cast_fp16")]; + tensor attn_mask_69 = logical_and(x = valid_mask_35_cast_fp16, y = causal_mask_35_cast_fp16)[name = tensor("attn_mask_69")]; + tensor attn_mask_71_axes_0 = const()[name = tensor("attn_mask_71_axes_0"), val = tensor([1])]; + tensor attn_mask_71 = expand_dims(axes = attn_mask_71_axes_0, x = attn_mask_69)[name = tensor("attn_mask_71")]; + tensor var_7106_transpose_x_0 = const()[name = tensor("op_7106_transpose_x_0"), val = tensor(false)]; + tensor var_7106_transpose_y_0 = const()[name = tensor("op_7106_transpose_y_0"), val = tensor(false)]; + tensor transpose_106_perm_0 = const()[name = tensor("transpose_106_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_107_perm_0 = const()[name = tensor("transpose_107_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_107 = transpose(perm = transpose_107_perm_0, x = keys_105_cast_fp16)[name = tensor("transpose_145")]; + tensor transpose_106 = transpose(perm = transpose_106_perm_0, x = q_105_cast_fp16)[name = tensor("transpose_146")]; + tensor var_7106_cast_fp16 = matmul(transpose_x = var_7106_transpose_x_0, transpose_y = var_7106_transpose_y_0, x = transpose_106, y = transpose_107)[name = tensor("op_7106_cast_fp16")]; + tensor var_7100_to_fp16 = const()[name = tensor("op_7100_to_fp16"), val = tensor([0x1p-3])]; + tensor attn_weights_103_cast_fp16 = mul(x = var_7106_cast_fp16, y = var_7100_to_fp16)[name = tensor("attn_weights_103_cast_fp16")]; + tensor var_7108 = logical_not(x = attn_mask_71)[name = tensor("op_7108")]; + tensor var_7109_to_fp16 = const()[name = tensor("op_7109_to_fp16"), val = tensor(-inf)]; + tensor attn_weights_105_cast_fp16 = select(a = var_7109_to_fp16, b = attn_weights_103_cast_fp16, cond = var_7108)[name = tensor("attn_weights_105_cast_fp16")]; + tensor var_7111 = const()[name = tensor("op_7111"), val = tensor(-1)]; + tensor attn_weights_107_cast_fp16 = softmax(axis = var_7111, x = attn_weights_105_cast_fp16)[name = tensor("attn_weights_107_cast_fp16")]; + tensor attn_output_35_transpose_x_0 = const()[name = tensor("attn_output_35_transpose_x_0"), val = tensor(false)]; + tensor attn_output_35_transpose_y_0 = const()[name = tensor("attn_output_35_transpose_y_0"), val = tensor(false)]; + tensor values_107_cast_fp16 = transpose(perm = var_7051, x = values_105_cast_fp16)[name = tensor("transpose_147")]; + tensor attn_output_35_cast_fp16 = matmul(transpose_x = attn_output_35_transpose_x_0, transpose_y = attn_output_35_transpose_y_0, x = attn_weights_107_cast_fp16, y = values_107_cast_fp16)[name = tensor("attn_output_35_cast_fp16")]; + tensor var_7119 = const()[name = tensor("op_7119"), val = tensor([0, 2, 1, 3])]; + tensor var_7124 = const()[name = tensor("op_7124"), val = tensor([1, 1, 1024])]; + tensor var_7120_cast_fp16 = transpose(perm = var_7119, x = attn_output_35_cast_fp16)[name = tensor("transpose_144")]; + tensor input_211_cast_fp16 = reshape(shape = var_7124, x = var_7120_cast_fp16)[name = tensor("input_211_cast_fp16")]; + tensor attn17_out_proj_weight_to_fp16 = const()[name = tensor("attn17_out_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(435396992)))]; + tensor linear_70_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = attn17_out_proj_weight_to_fp16, x = input_211_cast_fp16)[name = tensor("linear_70_cast_fp16")]; + tensor var_7130_to_fp16 = const()[name = tensor("op_7130_to_fp16"), val = tensor(0x1p+0)]; + tensor var_7131_cast_fp16 = add(x = position17_to_fp16, y = var_7130_to_fp16)[name = tensor("op_7131_cast_fp16")]; + tensor var_7131_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_7131_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor input_213_cast_fp16 = add(x = input_207_cast_fp16, y = linear_70_cast_fp16)[name = tensor("input_213_cast_fp16")]; + tensor input_215_axes_0 = const()[name = tensor("input_215_axes_0"), val = tensor([-1])]; + tensor norm17_2_weight_to_fp16 = const()[name = tensor("norm17_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(437494208)))]; + tensor norm17_2_bias_to_fp16 = const()[name = tensor("norm17_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(437496320)))]; + tensor var_7135_to_fp16 = const()[name = tensor("op_7135_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_215_cast_fp16 = layer_norm(axes = input_215_axes_0, beta = norm17_2_bias_to_fp16, epsilon = var_7135_to_fp16, gamma = norm17_2_weight_to_fp16, x = input_213_cast_fp16)[name = tensor("input_215_cast_fp16")]; + tensor linear17_1_weight_to_fp16 = const()[name = tensor("linear17_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(437498432)))]; + tensor linear_71_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = linear17_1_weight_to_fp16, x = input_215_cast_fp16)[name = tensor("linear_71_cast_fp16")]; + tensor input_217_mode_0 = const()[name = tensor("input_217_mode_0"), val = tensor("EXACT")]; + tensor input_217_cast_fp16 = gelu(mode = input_217_mode_0, x = linear_71_cast_fp16)[name = tensor("input_217_cast_fp16")]; + tensor linear17_2_weight_to_fp16 = const()[name = tensor("linear17_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(445887104)))]; + tensor linear_72_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = linear17_2_weight_to_fp16, x = input_217_cast_fp16)[name = tensor("linear_72_cast_fp16")]; + tensor input_219_cast_fp16 = add(x = input_213_cast_fp16, y = linear_72_cast_fp16)[name = tensor("input_219_cast_fp16")]; + tensor input_221_axes_0 = const()[name = tensor("input_221_axes_0"), val = tensor([-1])]; + tensor norm18_1_weight_to_fp16 = const()[name = tensor("norm18_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(454275776)))]; + tensor norm18_1_bias_to_fp16 = const()[name = tensor("norm18_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(454277888)))]; + tensor var_7152_to_fp16 = const()[name = tensor("op_7152_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_221_cast_fp16 = layer_norm(axes = input_221_axes_0, beta = norm18_1_bias_to_fp16, epsilon = var_7152_to_fp16, gamma = norm18_1_weight_to_fp16, x = input_219_cast_fp16)[name = tensor("input_221_cast_fp16")]; + tensor attn18_in_proj_weight_to_fp16 = const()[name = tensor("attn18_in_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(454280000)))]; + tensor linear_73_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = attn18_in_proj_weight_to_fp16, x = input_221_cast_fp16)[name = tensor("linear_73_cast_fp16")]; + tensor var_7166 = const()[name = tensor("op_7166"), val = tensor([1, 1, 3, 16, 64])]; + tensor qkv_37_cast_fp16 = reshape(shape = var_7166, x = linear_73_cast_fp16)[name = tensor("qkv_37_cast_fp16")]; + tensor q_109_begin_0 = const()[name = tensor("q_109_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor q_109_end_0 = const()[name = tensor("q_109_end_0"), val = tensor([1, 1, 1, 16, 64])]; + tensor q_109_end_mask_0 = const()[name = tensor("q_109_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor q_109_squeeze_mask_0 = const()[name = tensor("q_109_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor q_109_cast_fp16 = slice_by_index(begin = q_109_begin_0, end = q_109_end_0, end_mask = q_109_end_mask_0, squeeze_mask = q_109_squeeze_mask_0, x = qkv_37_cast_fp16)[name = tensor("q_109_cast_fp16")]; + tensor k_73_begin_0 = const()[name = tensor("k_73_begin_0"), val = tensor([0, 0, 1, 0, 0])]; + tensor k_73_end_0 = const()[name = tensor("k_73_end_0"), val = tensor([1, 1, 2, 16, 64])]; + tensor k_73_end_mask_0 = const()[name = tensor("k_73_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor k_73_squeeze_mask_0 = const()[name = tensor("k_73_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor k_73_cast_fp16 = slice_by_index(begin = k_73_begin_0, end = k_73_end_0, end_mask = k_73_end_mask_0, squeeze_mask = k_73_squeeze_mask_0, x = qkv_37_cast_fp16)[name = tensor("k_73_cast_fp16")]; + tensor v_37_begin_0 = const()[name = tensor("v_37_begin_0"), val = tensor([0, 0, 2, 0, 0])]; + tensor v_37_end_0 = const()[name = tensor("v_37_end_0"), val = tensor([1, 1, 3, 16, 64])]; + tensor v_37_end_mask_0 = const()[name = tensor("v_37_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor v_37_squeeze_mask_0 = const()[name = tensor("v_37_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor v_37_cast_fp16 = slice_by_index(begin = v_37_begin_0, end = v_37_end_0, end_mask = v_37_end_mask_0, squeeze_mask = v_37_squeeze_mask_0, x = qkv_37_cast_fp16)[name = tensor("v_37_cast_fp16")]; + tensor var_7232 = const()[name = tensor("op_7232"), val = tensor([1, 1, 1, 1])]; + tensor position18_to_fp16_dtype_0 = const()[name = tensor("position18_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor position18_to_fp16 = cast(dtype = position18_to_fp16_dtype_0, x = position18)[name = tensor("cast_463")]; + tensor ts_113_cast_fp16 = reshape(shape = var_7232, x = position18_to_fp16)[name = tensor("ts_113_cast_fp16")]; + tensor var_7239 = const()[name = tensor("op_7239"), val = tensor([1, 1, 16, 32, 2])]; + tensor q_complex_37_cast_fp16 = reshape(shape = var_7239, x = q_109_cast_fp16)[name = tensor("q_complex_37_cast_fp16")]; + tensor var_7246 = const()[name = tensor("op_7246"), val = tensor([1, 1, 16, 32, 2])]; + tensor k_complex_37_cast_fp16 = reshape(shape = var_7246, x = k_73_cast_fp16)[name = tensor("k_complex_37_cast_fp16")]; + tensor var_7250_begin_0 = const()[name = tensor("op_7250_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_7250_end_0 = const()[name = tensor("op_7250_end_0"), val = tensor([1, 1, 16, 32, 1])]; + tensor var_7250_end_mask_0 = const()[name = tensor("op_7250_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_7250_squeeze_mask_0 = const()[name = tensor("op_7250_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_7250_cast_fp16 = slice_by_index(begin = var_7250_begin_0, end = var_7250_end_0, end_mask = var_7250_end_mask_0, squeeze_mask = var_7250_squeeze_mask_0, x = q_complex_37_cast_fp16)[name = tensor("op_7250_cast_fp16")]; + tensor var_7258_begin_0 = const()[name = tensor("op_7258_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_7258_end_0 = const()[name = tensor("op_7258_end_0"), val = tensor([1, 1, 16, 32, 2])]; + tensor var_7258_end_mask_0 = const()[name = tensor("op_7258_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_7258_squeeze_mask_0 = const()[name = tensor("op_7258_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_7258_cast_fp16 = slice_by_index(begin = var_7258_begin_0, end = var_7258_end_0, end_mask = var_7258_end_mask_0, squeeze_mask = var_7258_squeeze_mask_0, x = q_complex_37_cast_fp16)[name = tensor("op_7258_cast_fp16")]; + tensor var_7266_begin_0 = const()[name = tensor("op_7266_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_7266_end_0 = const()[name = tensor("op_7266_end_0"), val = tensor([1, 1, 16, 32, 1])]; + tensor var_7266_end_mask_0 = const()[name = tensor("op_7266_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_7266_squeeze_mask_0 = const()[name = tensor("op_7266_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_7266_cast_fp16 = slice_by_index(begin = var_7266_begin_0, end = var_7266_end_0, end_mask = var_7266_end_mask_0, squeeze_mask = var_7266_squeeze_mask_0, x = k_complex_37_cast_fp16)[name = tensor("op_7266_cast_fp16")]; + tensor var_7274_begin_0 = const()[name = tensor("op_7274_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_7274_end_0 = const()[name = tensor("op_7274_end_0"), val = tensor([1, 1, 16, 32, 2])]; + tensor var_7274_end_mask_0 = const()[name = tensor("op_7274_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_7274_squeeze_mask_0 = const()[name = tensor("op_7274_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_7274_cast_fp16 = slice_by_index(begin = var_7274_begin_0, end = var_7274_end_0, end_mask = var_7274_end_mask_0, squeeze_mask = var_7274_squeeze_mask_0, x = k_complex_37_cast_fp16)[name = tensor("op_7274_cast_fp16")]; + tensor freqs_37_to_fp16 = const()[name = tensor("freqs_37_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(460571520)))]; + tensor var_7280_cast_fp16 = mul(x = freqs_37_to_fp16, y = ts_113_cast_fp16)[name = tensor("op_7280_cast_fp16")]; + tensor rotr_37_cast_fp16 = cos(x = var_7280_cast_fp16)[name = tensor("rotr_37_cast_fp16")]; + tensor roti_37_cast_fp16 = sin(x = var_7280_cast_fp16)[name = tensor("roti_37_cast_fp16")]; + tensor var_7284_cast_fp16 = mul(x = var_7250_cast_fp16, y = rotr_37_cast_fp16)[name = tensor("op_7284_cast_fp16")]; + tensor var_7285_cast_fp16 = mul(x = var_7258_cast_fp16, y = roti_37_cast_fp16)[name = tensor("op_7285_cast_fp16")]; + tensor qor_73_cast_fp16 = sub(x = var_7284_cast_fp16, y = var_7285_cast_fp16)[name = tensor("qor_73_cast_fp16")]; + tensor var_7288_cast_fp16 = mul(x = var_7250_cast_fp16, y = roti_37_cast_fp16)[name = tensor("op_7288_cast_fp16")]; + tensor var_7289_cast_fp16 = mul(x = var_7258_cast_fp16, y = rotr_37_cast_fp16)[name = tensor("op_7289_cast_fp16")]; + tensor qoi_73_cast_fp16 = add(x = var_7288_cast_fp16, y = var_7289_cast_fp16)[name = tensor("qoi_73_cast_fp16")]; + tensor var_7292_cast_fp16 = mul(x = var_7266_cast_fp16, y = rotr_37_cast_fp16)[name = tensor("op_7292_cast_fp16")]; + tensor var_7293_cast_fp16 = mul(x = var_7274_cast_fp16, y = roti_37_cast_fp16)[name = tensor("op_7293_cast_fp16")]; + tensor kor_73_cast_fp16 = sub(x = var_7292_cast_fp16, y = var_7293_cast_fp16)[name = tensor("kor_73_cast_fp16")]; + tensor var_7296_cast_fp16 = mul(x = var_7266_cast_fp16, y = roti_37_cast_fp16)[name = tensor("op_7296_cast_fp16")]; + tensor var_7297_cast_fp16 = mul(x = var_7274_cast_fp16, y = rotr_37_cast_fp16)[name = tensor("op_7297_cast_fp16")]; + tensor koi_73_cast_fp16 = add(x = var_7296_cast_fp16, y = var_7297_cast_fp16)[name = tensor("koi_73_cast_fp16")]; + tensor qo_37_axis_0 = const()[name = tensor("qo_37_axis_0"), val = tensor(-1)]; + tensor qo_37_cast_fp16 = stack(axis = qo_37_axis_0, values = (qor_73_cast_fp16, qoi_73_cast_fp16))[name = tensor("qo_37_cast_fp16")]; + tensor ko_37_axis_0 = const()[name = tensor("ko_37_axis_0"), val = tensor(-1)]; + tensor ko_37_cast_fp16 = stack(axis = ko_37_axis_0, values = (kor_73_cast_fp16, koi_73_cast_fp16))[name = tensor("ko_37_cast_fp16")]; + tensor var_7330 = const()[name = tensor("op_7330"), val = tensor([1, 1, 16, 64])]; + tensor q_111_cast_fp16 = reshape(shape = var_7330, x = qo_37_cast_fp16)[name = tensor("q_111_cast_fp16")]; + tensor var_7336 = const()[name = tensor("op_7336"), val = tensor([1, 1, 16, 64])]; + tensor k_75_cast_fp16 = reshape(shape = var_7336, x = ko_37_cast_fp16)[name = tensor("k_75_cast_fp16")]; + tensor _inversed_7361_y_0_to_fp16 = const()[name = tensor("_inversed_7361_y_0_to_fp16"), val = tensor(0x1p-9)]; + tensor _inversed_7361_cast_fp16 = mul(x = ts_113_cast_fp16, y = _inversed_7361_y_0_to_fp16)[name = tensor("_inversed_7361_cast_fp16")]; + tensor var_7362_cast_fp16 = floor(x = _inversed_7361_cast_fp16)[name = tensor("op_7362_cast_fp16")]; + tensor var_7363_to_fp16 = const()[name = tensor("op_7363_to_fp16"), val = tensor(0x1p+9)]; + tensor var_7364_cast_fp16 = mul(x = var_7362_cast_fp16, y = var_7363_to_fp16)[name = tensor("op_7364_cast_fp16")]; + tensor write_indices_float_75_cast_fp16 = sub(x = ts_113_cast_fp16, y = var_7364_cast_fp16)[name = tensor("write_indices_float_75_cast_fp16")]; + tensor var_7371_dtype_0 = const()[name = tensor("op_7371_dtype_0"), val = tensor("int32")]; + tensor write_indices_37_reps_0 = const()[name = tensor("write_indices_37_reps_0"), val = tensor([1, 1, 16, 64])]; + tensor write_indices_float_75_cast_fp16_to_int32 = cast(dtype = var_7371_dtype_0, x = write_indices_float_75_cast_fp16)[name = tensor("cast_462")]; + tensor write_indices_37 = tile(reps = write_indices_37_reps_0, x = write_indices_float_75_cast_fp16_to_int32)[name = tensor("write_indices_37")]; + tensor var_7381_begin_0 = const()[name = tensor("op_7381_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_7381_end_0 = const()[name = tensor("op_7381_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor var_7381_end_mask_0 = const()[name = tensor("op_7381_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_7381_squeeze_mask_0 = const()[name = tensor("op_7381_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor cache18_to_fp16_dtype_0 = const()[name = tensor("cache18_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor cache18_to_fp16 = cast(dtype = cache18_to_fp16_dtype_0, x = cache18)[name = tensor("cast_461")]; + tensor var_7381_cast_fp16 = slice_by_index(begin = var_7381_begin_0, end = var_7381_end_0, end_mask = var_7381_end_mask_0, squeeze_mask = var_7381_squeeze_mask_0, x = cache18_to_fp16)[name = tensor("op_7381_cast_fp16")]; + tensor var_7383_axis_0 = const()[name = tensor("op_7383_axis_0"), val = tensor(1)]; + tensor var_7383_mode_0 = const()[name = tensor("op_7383_mode_0"), val = tensor("update")]; + tensor var_7383_validate_indices_0 = const()[name = tensor("op_7383_validate_indices_0"), val = tensor(false)]; + tensor var_7383_cast_fp16 = scatter_along_axis(axis = var_7383_axis_0, data = var_7381_cast_fp16, indices = write_indices_37, mode = var_7383_mode_0, updates = k_75_cast_fp16, validate_indices = var_7383_validate_indices_0)[name = tensor("op_7383_cast_fp16")]; + tensor concat_128 = const()[name = tensor("concat_128"), val = tensor([0, 0, 0, 0, 0])]; + tensor concat_129 = const()[name = tensor("concat_129"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_37_internal_tensor_assign_1_stride_0 = const()[name = tensor("new_cache_37_internal_tensor_assign_1_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_37_internal_tensor_assign_1_begin_mask_0 = const()[name = tensor("new_cache_37_internal_tensor_assign_1_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_37_internal_tensor_assign_1_end_mask_0 = const()[name = tensor("new_cache_37_internal_tensor_assign_1_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_37_internal_tensor_assign_1_squeeze_mask_0 = const()[name = tensor("new_cache_37_internal_tensor_assign_1_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_84 = const()[name = tensor("shape_84"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_36 = const()[name = tensor("reduce_prod_36"), val = tensor(1048576)]; + tensor range_1d_36_start_0 = const()[name = tensor("range_1d_36_start_0"), val = tensor(0)]; + tensor range_1d_36_step_0 = const()[name = tensor("range_1d_36_step_0"), val = tensor(1)]; + tensor range_1d_36 = range_1d(end = reduce_prod_36, start = range_1d_36_start_0, step = range_1d_36_step_0)[name = tensor("range_1d_36")]; + tensor reshape_180 = reshape(shape = shape_84, x = range_1d_36)[name = tensor("reshape_180")]; + tensor slice_by_index_36 = slice_by_index(begin = concat_128, begin_mask = new_cache_37_internal_tensor_assign_1_begin_mask_0, end = concat_129, end_mask = new_cache_37_internal_tensor_assign_1_end_mask_0, squeeze_mask = new_cache_37_internal_tensor_assign_1_squeeze_mask_0, stride = new_cache_37_internal_tensor_assign_1_stride_0, x = reshape_180)[name = tensor("slice_by_index_36")]; + tensor reshape_181_shape_0 = const()[name = tensor("reshape_181_shape_0"), val = tensor([-1])]; + tensor reshape_181 = reshape(shape = reshape_181_shape_0, x = slice_by_index_36)[name = tensor("reshape_181")]; + tensor reshape_182_shape_0 = const()[name = tensor("reshape_182_shape_0"), val = tensor([-1])]; + tensor reshape_182_cast_fp16 = reshape(shape = reshape_182_shape_0, x = var_7383_cast_fp16)[name = tensor("reshape_182_cast_fp16")]; + tensor reshape_183_shape_0 = const()[name = tensor("reshape_183_shape_0"), val = tensor([-1])]; + tensor reshape_183_cast_fp16 = reshape(shape = reshape_183_shape_0, x = cache18_to_fp16)[name = tensor("reshape_183_cast_fp16")]; + tensor scatter_36_mode_0 = const()[name = tensor("scatter_36_mode_0"), val = tensor("update")]; + tensor scatter_36_axis_0 = const()[name = tensor("scatter_36_axis_0"), val = tensor(0)]; + tensor scatter_36_validate_indices_0 = const()[name = tensor("scatter_36_validate_indices_0"), val = tensor(false)]; + tensor scatter_36_cast_fp16 = scatter(axis = scatter_36_axis_0, data = reshape_183_cast_fp16, indices = reshape_181, mode = scatter_36_mode_0, updates = reshape_182_cast_fp16, validate_indices = scatter_36_validate_indices_0)[name = tensor("scatter_36_cast_fp16")]; + tensor reshape_184_cast_fp16 = reshape(shape = shape_84, x = scatter_36_cast_fp16)[name = tensor("reshape_184_cast_fp16")]; + tensor var_7391_begin_0 = const()[name = tensor("op_7391_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor var_7391_end_0 = const()[name = tensor("op_7391_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor var_7391_end_mask_0 = const()[name = tensor("op_7391_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_7391_squeeze_mask_0 = const()[name = tensor("op_7391_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor var_7391_cast_fp16 = slice_by_index(begin = var_7391_begin_0, end = var_7391_end_0, end_mask = var_7391_end_mask_0, squeeze_mask = var_7391_squeeze_mask_0, x = reshape_184_cast_fp16)[name = tensor("op_7391_cast_fp16")]; + tensor var_7393_axis_0 = const()[name = tensor("op_7393_axis_0"), val = tensor(1)]; + tensor var_7393_mode_0 = const()[name = tensor("op_7393_mode_0"), val = tensor("update")]; + tensor var_7393_validate_indices_0 = const()[name = tensor("op_7393_validate_indices_0"), val = tensor(false)]; + tensor var_7393_cast_fp16 = scatter_along_axis(axis = var_7393_axis_0, data = var_7391_cast_fp16, indices = write_indices_37, mode = var_7393_mode_0, updates = v_37_cast_fp16, validate_indices = var_7393_validate_indices_0)[name = tensor("op_7393_cast_fp16")]; + tensor concat_130 = const()[name = tensor("concat_130"), val = tensor([1, 0, 0, 0, 0])]; + tensor concat_131 = const()[name = tensor("concat_131"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_37_internal_tensor_assign_2_stride_0 = const()[name = tensor("new_cache_37_internal_tensor_assign_2_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_37_internal_tensor_assign_2_begin_mask_0 = const()[name = tensor("new_cache_37_internal_tensor_assign_2_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_37_internal_tensor_assign_2_end_mask_0 = const()[name = tensor("new_cache_37_internal_tensor_assign_2_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_37_internal_tensor_assign_2_squeeze_mask_0 = const()[name = tensor("new_cache_37_internal_tensor_assign_2_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_85 = const()[name = tensor("shape_85"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_37 = const()[name = tensor("reduce_prod_37"), val = tensor(1048576)]; + tensor range_1d_37_start_0 = const()[name = tensor("range_1d_37_start_0"), val = tensor(0)]; + tensor range_1d_37_step_0 = const()[name = tensor("range_1d_37_step_0"), val = tensor(1)]; + tensor range_1d_37 = range_1d(end = reduce_prod_37, start = range_1d_37_start_0, step = range_1d_37_step_0)[name = tensor("range_1d_37")]; + tensor reshape_185 = reshape(shape = shape_85, x = range_1d_37)[name = tensor("reshape_185")]; + tensor slice_by_index_37 = slice_by_index(begin = concat_130, begin_mask = new_cache_37_internal_tensor_assign_2_begin_mask_0, end = concat_131, end_mask = new_cache_37_internal_tensor_assign_2_end_mask_0, squeeze_mask = new_cache_37_internal_tensor_assign_2_squeeze_mask_0, stride = new_cache_37_internal_tensor_assign_2_stride_0, x = reshape_185)[name = tensor("slice_by_index_37")]; + tensor reshape_186_shape_0 = const()[name = tensor("reshape_186_shape_0"), val = tensor([-1])]; + tensor reshape_186 = reshape(shape = reshape_186_shape_0, x = slice_by_index_37)[name = tensor("reshape_186")]; + tensor reshape_187_shape_0 = const()[name = tensor("reshape_187_shape_0"), val = tensor([-1])]; + tensor reshape_187_cast_fp16 = reshape(shape = reshape_187_shape_0, x = var_7393_cast_fp16)[name = tensor("reshape_187_cast_fp16")]; + tensor reshape_188_shape_0 = const()[name = tensor("reshape_188_shape_0"), val = tensor([-1])]; + tensor reshape_188_cast_fp16 = reshape(shape = reshape_188_shape_0, x = reshape_184_cast_fp16)[name = tensor("reshape_188_cast_fp16")]; + tensor scatter_37_mode_0 = const()[name = tensor("scatter_37_mode_0"), val = tensor("update")]; + tensor scatter_37_axis_0 = const()[name = tensor("scatter_37_axis_0"), val = tensor(0)]; + tensor scatter_37_validate_indices_0 = const()[name = tensor("scatter_37_validate_indices_0"), val = tensor(false)]; + tensor scatter_37_cast_fp16 = scatter(axis = scatter_37_axis_0, data = reshape_188_cast_fp16, indices = reshape_186, mode = scatter_37_mode_0, updates = reshape_187_cast_fp16, validate_indices = scatter_37_validate_indices_0)[name = tensor("scatter_37_cast_fp16")]; + tensor reshape_189_cast_fp16 = reshape(shape = shape_85, x = scatter_37_cast_fp16)[name = tensor("reshape_189_cast_fp16")]; + tensor reshape_189_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("reshape_189_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor keys_109_begin_0 = const()[name = tensor("keys_109_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor keys_109_end_0 = const()[name = tensor("keys_109_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor keys_109_end_mask_0 = const()[name = tensor("keys_109_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor keys_109_squeeze_mask_0 = const()[name = tensor("keys_109_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor keys_109_cast_fp16 = slice_by_index(begin = keys_109_begin_0, end = keys_109_end_0, end_mask = keys_109_end_mask_0, squeeze_mask = keys_109_squeeze_mask_0, x = reshape_189_cast_fp16)[name = tensor("keys_109_cast_fp16")]; + tensor values_109_begin_0 = const()[name = tensor("values_109_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor values_109_end_0 = const()[name = tensor("values_109_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor values_109_end_mask_0 = const()[name = tensor("values_109_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor values_109_squeeze_mask_0 = const()[name = tensor("values_109_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor values_109_cast_fp16 = slice_by_index(begin = values_109_begin_0, end = values_109_end_0, end_mask = values_109_end_mask_0, squeeze_mask = values_109_squeeze_mask_0, x = reshape_189_cast_fp16)[name = tensor("values_109_cast_fp16")]; + tensor var_7405_cast_fp16 = not_equal(x = keys_109_cast_fp16, y = keys_109_cast_fp16)[name = tensor("op_7405_cast_fp16")]; + tensor keys_111_cast_fp16 = select(a = var_463_to_fp16, b = keys_109_cast_fp16, cond = var_7405_cast_fp16)[name = tensor("keys_111_cast_fp16")]; + tensor var_7413_cast_fp16 = not_equal(x = values_109_cast_fp16, y = values_109_cast_fp16)[name = tensor("op_7413_cast_fp16")]; + tensor values_111_cast_fp16 = select(a = var_463_to_fp16, b = values_109_cast_fp16, cond = var_7413_cast_fp16)[name = tensor("values_111_cast_fp16")]; + tensor var_7437 = const()[name = tensor("op_7437"), val = tensor([0, 2, 1, 3])]; + tensor var_7453 = const()[name = tensor("op_7453"), val = tensor([1, 1, 1])]; + tensor var_7454_cast_fp16 = reshape(shape = var_7453, x = position18_to_fp16)[name = tensor("op_7454_cast_fp16")]; + tensor var_7474_to_fp16 = const()[name = tensor("op_7474_to_fp16"), val = tensor(0x1p+0)]; + tensor valid_len_37_cast_fp16 = add(x = var_7454_cast_fp16, y = var_7474_to_fp16)[name = tensor("valid_len_37_cast_fp16")]; + tensor valid_mask_37_cast_fp16 = less(x = k_positions_1_promoted_to_fp16, y = valid_len_37_cast_fp16)[name = tensor("valid_mask_37_cast_fp16")]; + tensor causal_mask_37_cast_fp16 = less_equal(x = k_positions_1_promoted_to_fp16, y = var_7454_cast_fp16)[name = tensor("causal_mask_37_cast_fp16")]; + tensor attn_mask_73 = logical_and(x = valid_mask_37_cast_fp16, y = causal_mask_37_cast_fp16)[name = tensor("attn_mask_73")]; + tensor attn_mask_75_axes_0 = const()[name = tensor("attn_mask_75_axes_0"), val = tensor([1])]; + tensor attn_mask_75 = expand_dims(axes = attn_mask_75_axes_0, x = attn_mask_73)[name = tensor("attn_mask_75")]; + tensor var_7492_transpose_x_0 = const()[name = tensor("op_7492_transpose_x_0"), val = tensor(false)]; + tensor var_7492_transpose_y_0 = const()[name = tensor("op_7492_transpose_y_0"), val = tensor(false)]; + tensor transpose_108_perm_0 = const()[name = tensor("transpose_108_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_109_perm_0 = const()[name = tensor("transpose_109_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_109 = transpose(perm = transpose_109_perm_0, x = keys_111_cast_fp16)[name = tensor("transpose_141")]; + tensor transpose_108 = transpose(perm = transpose_108_perm_0, x = q_111_cast_fp16)[name = tensor("transpose_142")]; + tensor var_7492_cast_fp16 = matmul(transpose_x = var_7492_transpose_x_0, transpose_y = var_7492_transpose_y_0, x = transpose_108, y = transpose_109)[name = tensor("op_7492_cast_fp16")]; + tensor var_7486_to_fp16 = const()[name = tensor("op_7486_to_fp16"), val = tensor([0x1p-3])]; + tensor attn_weights_109_cast_fp16 = mul(x = var_7492_cast_fp16, y = var_7486_to_fp16)[name = tensor("attn_weights_109_cast_fp16")]; + tensor var_7494 = logical_not(x = attn_mask_75)[name = tensor("op_7494")]; + tensor var_7495_to_fp16 = const()[name = tensor("op_7495_to_fp16"), val = tensor(-inf)]; + tensor attn_weights_111_cast_fp16 = select(a = var_7495_to_fp16, b = attn_weights_109_cast_fp16, cond = var_7494)[name = tensor("attn_weights_111_cast_fp16")]; + tensor var_7497 = const()[name = tensor("op_7497"), val = tensor(-1)]; + tensor attn_weights_113_cast_fp16 = softmax(axis = var_7497, x = attn_weights_111_cast_fp16)[name = tensor("attn_weights_113_cast_fp16")]; + tensor attn_output_37_transpose_x_0 = const()[name = tensor("attn_output_37_transpose_x_0"), val = tensor(false)]; + tensor attn_output_37_transpose_y_0 = const()[name = tensor("attn_output_37_transpose_y_0"), val = tensor(false)]; + tensor values_113_cast_fp16 = transpose(perm = var_7437, x = values_111_cast_fp16)[name = tensor("transpose_143")]; + tensor attn_output_37_cast_fp16 = matmul(transpose_x = attn_output_37_transpose_x_0, transpose_y = attn_output_37_transpose_y_0, x = attn_weights_113_cast_fp16, y = values_113_cast_fp16)[name = tensor("attn_output_37_cast_fp16")]; + tensor var_7505 = const()[name = tensor("op_7505"), val = tensor([0, 2, 1, 3])]; + tensor var_7510 = const()[name = tensor("op_7510"), val = tensor([1, 1, 1024])]; + tensor var_7506_cast_fp16 = transpose(perm = var_7505, x = attn_output_37_cast_fp16)[name = tensor("transpose_140")]; + tensor input_223_cast_fp16 = reshape(shape = var_7510, x = var_7506_cast_fp16)[name = tensor("input_223_cast_fp16")]; + tensor attn18_out_proj_weight_to_fp16 = const()[name = tensor("attn18_out_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(460571648)))]; + tensor linear_74_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = attn18_out_proj_weight_to_fp16, x = input_223_cast_fp16)[name = tensor("linear_74_cast_fp16")]; + tensor var_7516_to_fp16 = const()[name = tensor("op_7516_to_fp16"), val = tensor(0x1p+0)]; + tensor var_7517_cast_fp16 = add(x = position18_to_fp16, y = var_7516_to_fp16)[name = tensor("op_7517_cast_fp16")]; + tensor var_7517_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_7517_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor input_225_cast_fp16 = add(x = input_219_cast_fp16, y = linear_74_cast_fp16)[name = tensor("input_225_cast_fp16")]; + tensor input_227_axes_0 = const()[name = tensor("input_227_axes_0"), val = tensor([-1])]; + tensor norm18_2_weight_to_fp16 = const()[name = tensor("norm18_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(462668864)))]; + tensor norm18_2_bias_to_fp16 = const()[name = tensor("norm18_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(462670976)))]; + tensor var_7521_to_fp16 = const()[name = tensor("op_7521_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_227_cast_fp16 = layer_norm(axes = input_227_axes_0, beta = norm18_2_bias_to_fp16, epsilon = var_7521_to_fp16, gamma = norm18_2_weight_to_fp16, x = input_225_cast_fp16)[name = tensor("input_227_cast_fp16")]; + tensor linear18_1_weight_to_fp16 = const()[name = tensor("linear18_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(462673088)))]; + tensor linear_75_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = linear18_1_weight_to_fp16, x = input_227_cast_fp16)[name = tensor("linear_75_cast_fp16")]; + tensor input_229_mode_0 = const()[name = tensor("input_229_mode_0"), val = tensor("EXACT")]; + tensor input_229_cast_fp16 = gelu(mode = input_229_mode_0, x = linear_75_cast_fp16)[name = tensor("input_229_cast_fp16")]; + tensor linear18_2_weight_to_fp16 = const()[name = tensor("linear18_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(471061760)))]; + tensor linear_76_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = linear18_2_weight_to_fp16, x = input_229_cast_fp16)[name = tensor("linear_76_cast_fp16")]; + tensor input_231_cast_fp16 = add(x = input_225_cast_fp16, y = linear_76_cast_fp16)[name = tensor("input_231_cast_fp16")]; + tensor input_233_axes_0 = const()[name = tensor("input_233_axes_0"), val = tensor([-1])]; + tensor norm19_1_weight_to_fp16 = const()[name = tensor("norm19_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(479450432)))]; + tensor norm19_1_bias_to_fp16 = const()[name = tensor("norm19_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(479452544)))]; + tensor var_7538_to_fp16 = const()[name = tensor("op_7538_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_233_cast_fp16 = layer_norm(axes = input_233_axes_0, beta = norm19_1_bias_to_fp16, epsilon = var_7538_to_fp16, gamma = norm19_1_weight_to_fp16, x = input_231_cast_fp16)[name = tensor("input_233_cast_fp16")]; + tensor attn19_in_proj_weight_to_fp16 = const()[name = tensor("attn19_in_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(479454656)))]; + tensor linear_77_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = attn19_in_proj_weight_to_fp16, x = input_233_cast_fp16)[name = tensor("linear_77_cast_fp16")]; + tensor var_7552 = const()[name = tensor("op_7552"), val = tensor([1, 1, 3, 16, 64])]; + tensor qkv_39_cast_fp16 = reshape(shape = var_7552, x = linear_77_cast_fp16)[name = tensor("qkv_39_cast_fp16")]; + tensor q_115_begin_0 = const()[name = tensor("q_115_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor q_115_end_0 = const()[name = tensor("q_115_end_0"), val = tensor([1, 1, 1, 16, 64])]; + tensor q_115_end_mask_0 = const()[name = tensor("q_115_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor q_115_squeeze_mask_0 = const()[name = tensor("q_115_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor q_115_cast_fp16 = slice_by_index(begin = q_115_begin_0, end = q_115_end_0, end_mask = q_115_end_mask_0, squeeze_mask = q_115_squeeze_mask_0, x = qkv_39_cast_fp16)[name = tensor("q_115_cast_fp16")]; + tensor k_77_begin_0 = const()[name = tensor("k_77_begin_0"), val = tensor([0, 0, 1, 0, 0])]; + tensor k_77_end_0 = const()[name = tensor("k_77_end_0"), val = tensor([1, 1, 2, 16, 64])]; + tensor k_77_end_mask_0 = const()[name = tensor("k_77_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor k_77_squeeze_mask_0 = const()[name = tensor("k_77_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor k_77_cast_fp16 = slice_by_index(begin = k_77_begin_0, end = k_77_end_0, end_mask = k_77_end_mask_0, squeeze_mask = k_77_squeeze_mask_0, x = qkv_39_cast_fp16)[name = tensor("k_77_cast_fp16")]; + tensor v_39_begin_0 = const()[name = tensor("v_39_begin_0"), val = tensor([0, 0, 2, 0, 0])]; + tensor v_39_end_0 = const()[name = tensor("v_39_end_0"), val = tensor([1, 1, 3, 16, 64])]; + tensor v_39_end_mask_0 = const()[name = tensor("v_39_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor v_39_squeeze_mask_0 = const()[name = tensor("v_39_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor v_39_cast_fp16 = slice_by_index(begin = v_39_begin_0, end = v_39_end_0, end_mask = v_39_end_mask_0, squeeze_mask = v_39_squeeze_mask_0, x = qkv_39_cast_fp16)[name = tensor("v_39_cast_fp16")]; + tensor var_7618 = const()[name = tensor("op_7618"), val = tensor([1, 1, 1, 1])]; + tensor position19_to_fp16_dtype_0 = const()[name = tensor("position19_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor position19_to_fp16 = cast(dtype = position19_to_fp16_dtype_0, x = position19)[name = tensor("cast_458")]; + tensor ts_119_cast_fp16 = reshape(shape = var_7618, x = position19_to_fp16)[name = tensor("ts_119_cast_fp16")]; + tensor var_7625 = const()[name = tensor("op_7625"), val = tensor([1, 1, 16, 32, 2])]; + tensor q_complex_39_cast_fp16 = reshape(shape = var_7625, x = q_115_cast_fp16)[name = tensor("q_complex_39_cast_fp16")]; + tensor var_7632 = const()[name = tensor("op_7632"), val = tensor([1, 1, 16, 32, 2])]; + tensor k_complex_39_cast_fp16 = reshape(shape = var_7632, x = k_77_cast_fp16)[name = tensor("k_complex_39_cast_fp16")]; + tensor var_7636_begin_0 = const()[name = tensor("op_7636_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_7636_end_0 = const()[name = tensor("op_7636_end_0"), val = tensor([1, 1, 16, 32, 1])]; + tensor var_7636_end_mask_0 = const()[name = tensor("op_7636_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_7636_squeeze_mask_0 = const()[name = tensor("op_7636_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_7636_cast_fp16 = slice_by_index(begin = var_7636_begin_0, end = var_7636_end_0, end_mask = var_7636_end_mask_0, squeeze_mask = var_7636_squeeze_mask_0, x = q_complex_39_cast_fp16)[name = tensor("op_7636_cast_fp16")]; + tensor var_7644_begin_0 = const()[name = tensor("op_7644_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_7644_end_0 = const()[name = tensor("op_7644_end_0"), val = tensor([1, 1, 16, 32, 2])]; + tensor var_7644_end_mask_0 = const()[name = tensor("op_7644_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_7644_squeeze_mask_0 = const()[name = tensor("op_7644_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_7644_cast_fp16 = slice_by_index(begin = var_7644_begin_0, end = var_7644_end_0, end_mask = var_7644_end_mask_0, squeeze_mask = var_7644_squeeze_mask_0, x = q_complex_39_cast_fp16)[name = tensor("op_7644_cast_fp16")]; + tensor var_7652_begin_0 = const()[name = tensor("op_7652_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_7652_end_0 = const()[name = tensor("op_7652_end_0"), val = tensor([1, 1, 16, 32, 1])]; + tensor var_7652_end_mask_0 = const()[name = tensor("op_7652_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_7652_squeeze_mask_0 = const()[name = tensor("op_7652_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_7652_cast_fp16 = slice_by_index(begin = var_7652_begin_0, end = var_7652_end_0, end_mask = var_7652_end_mask_0, squeeze_mask = var_7652_squeeze_mask_0, x = k_complex_39_cast_fp16)[name = tensor("op_7652_cast_fp16")]; + tensor var_7660_begin_0 = const()[name = tensor("op_7660_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_7660_end_0 = const()[name = tensor("op_7660_end_0"), val = tensor([1, 1, 16, 32, 2])]; + tensor var_7660_end_mask_0 = const()[name = tensor("op_7660_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_7660_squeeze_mask_0 = const()[name = tensor("op_7660_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_7660_cast_fp16 = slice_by_index(begin = var_7660_begin_0, end = var_7660_end_0, end_mask = var_7660_end_mask_0, squeeze_mask = var_7660_squeeze_mask_0, x = k_complex_39_cast_fp16)[name = tensor("op_7660_cast_fp16")]; + tensor freqs_39_to_fp16 = const()[name = tensor("freqs_39_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(485746176)))]; + tensor var_7666_cast_fp16 = mul(x = freqs_39_to_fp16, y = ts_119_cast_fp16)[name = tensor("op_7666_cast_fp16")]; + tensor rotr_39_cast_fp16 = cos(x = var_7666_cast_fp16)[name = tensor("rotr_39_cast_fp16")]; + tensor roti_39_cast_fp16 = sin(x = var_7666_cast_fp16)[name = tensor("roti_39_cast_fp16")]; + tensor var_7670_cast_fp16 = mul(x = var_7636_cast_fp16, y = rotr_39_cast_fp16)[name = tensor("op_7670_cast_fp16")]; + tensor var_7671_cast_fp16 = mul(x = var_7644_cast_fp16, y = roti_39_cast_fp16)[name = tensor("op_7671_cast_fp16")]; + tensor qor_77_cast_fp16 = sub(x = var_7670_cast_fp16, y = var_7671_cast_fp16)[name = tensor("qor_77_cast_fp16")]; + tensor var_7674_cast_fp16 = mul(x = var_7636_cast_fp16, y = roti_39_cast_fp16)[name = tensor("op_7674_cast_fp16")]; + tensor var_7675_cast_fp16 = mul(x = var_7644_cast_fp16, y = rotr_39_cast_fp16)[name = tensor("op_7675_cast_fp16")]; + tensor qoi_77_cast_fp16 = add(x = var_7674_cast_fp16, y = var_7675_cast_fp16)[name = tensor("qoi_77_cast_fp16")]; + tensor var_7678_cast_fp16 = mul(x = var_7652_cast_fp16, y = rotr_39_cast_fp16)[name = tensor("op_7678_cast_fp16")]; + tensor var_7679_cast_fp16 = mul(x = var_7660_cast_fp16, y = roti_39_cast_fp16)[name = tensor("op_7679_cast_fp16")]; + tensor kor_77_cast_fp16 = sub(x = var_7678_cast_fp16, y = var_7679_cast_fp16)[name = tensor("kor_77_cast_fp16")]; + tensor var_7682_cast_fp16 = mul(x = var_7652_cast_fp16, y = roti_39_cast_fp16)[name = tensor("op_7682_cast_fp16")]; + tensor var_7683_cast_fp16 = mul(x = var_7660_cast_fp16, y = rotr_39_cast_fp16)[name = tensor("op_7683_cast_fp16")]; + tensor koi_77_cast_fp16 = add(x = var_7682_cast_fp16, y = var_7683_cast_fp16)[name = tensor("koi_77_cast_fp16")]; + tensor qo_39_axis_0 = const()[name = tensor("qo_39_axis_0"), val = tensor(-1)]; + tensor qo_39_cast_fp16 = stack(axis = qo_39_axis_0, values = (qor_77_cast_fp16, qoi_77_cast_fp16))[name = tensor("qo_39_cast_fp16")]; + tensor ko_39_axis_0 = const()[name = tensor("ko_39_axis_0"), val = tensor(-1)]; + tensor ko_39_cast_fp16 = stack(axis = ko_39_axis_0, values = (kor_77_cast_fp16, koi_77_cast_fp16))[name = tensor("ko_39_cast_fp16")]; + tensor var_7716 = const()[name = tensor("op_7716"), val = tensor([1, 1, 16, 64])]; + tensor q_117_cast_fp16 = reshape(shape = var_7716, x = qo_39_cast_fp16)[name = tensor("q_117_cast_fp16")]; + tensor var_7722 = const()[name = tensor("op_7722"), val = tensor([1, 1, 16, 64])]; + tensor k_79_cast_fp16 = reshape(shape = var_7722, x = ko_39_cast_fp16)[name = tensor("k_79_cast_fp16")]; + tensor _inversed_7747_y_0_to_fp16 = const()[name = tensor("_inversed_7747_y_0_to_fp16"), val = tensor(0x1p-9)]; + tensor _inversed_7747_cast_fp16 = mul(x = ts_119_cast_fp16, y = _inversed_7747_y_0_to_fp16)[name = tensor("_inversed_7747_cast_fp16")]; + tensor var_7748_cast_fp16 = floor(x = _inversed_7747_cast_fp16)[name = tensor("op_7748_cast_fp16")]; + tensor var_7749_to_fp16 = const()[name = tensor("op_7749_to_fp16"), val = tensor(0x1p+9)]; + tensor var_7750_cast_fp16 = mul(x = var_7748_cast_fp16, y = var_7749_to_fp16)[name = tensor("op_7750_cast_fp16")]; + tensor write_indices_float_79_cast_fp16 = sub(x = ts_119_cast_fp16, y = var_7750_cast_fp16)[name = tensor("write_indices_float_79_cast_fp16")]; + tensor var_7757_dtype_0 = const()[name = tensor("op_7757_dtype_0"), val = tensor("int32")]; + tensor write_indices_39_reps_0 = const()[name = tensor("write_indices_39_reps_0"), val = tensor([1, 1, 16, 64])]; + tensor write_indices_float_79_cast_fp16_to_int32 = cast(dtype = var_7757_dtype_0, x = write_indices_float_79_cast_fp16)[name = tensor("cast_457")]; + tensor write_indices_39 = tile(reps = write_indices_39_reps_0, x = write_indices_float_79_cast_fp16_to_int32)[name = tensor("write_indices_39")]; + tensor var_7767_begin_0 = const()[name = tensor("op_7767_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_7767_end_0 = const()[name = tensor("op_7767_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor var_7767_end_mask_0 = const()[name = tensor("op_7767_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_7767_squeeze_mask_0 = const()[name = tensor("op_7767_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor cache19_to_fp16_dtype_0 = const()[name = tensor("cache19_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor cache19_to_fp16 = cast(dtype = cache19_to_fp16_dtype_0, x = cache19)[name = tensor("cast_456")]; + tensor var_7767_cast_fp16 = slice_by_index(begin = var_7767_begin_0, end = var_7767_end_0, end_mask = var_7767_end_mask_0, squeeze_mask = var_7767_squeeze_mask_0, x = cache19_to_fp16)[name = tensor("op_7767_cast_fp16")]; + tensor var_7769_axis_0 = const()[name = tensor("op_7769_axis_0"), val = tensor(1)]; + tensor var_7769_mode_0 = const()[name = tensor("op_7769_mode_0"), val = tensor("update")]; + tensor var_7769_validate_indices_0 = const()[name = tensor("op_7769_validate_indices_0"), val = tensor(false)]; + tensor var_7769_cast_fp16 = scatter_along_axis(axis = var_7769_axis_0, data = var_7767_cast_fp16, indices = write_indices_39, mode = var_7769_mode_0, updates = k_79_cast_fp16, validate_indices = var_7769_validate_indices_0)[name = tensor("op_7769_cast_fp16")]; + tensor concat_135 = const()[name = tensor("concat_135"), val = tensor([0, 0, 0, 0, 0])]; + tensor concat_136 = const()[name = tensor("concat_136"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_39_internal_tensor_assign_1_stride_0 = const()[name = tensor("new_cache_39_internal_tensor_assign_1_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_39_internal_tensor_assign_1_begin_mask_0 = const()[name = tensor("new_cache_39_internal_tensor_assign_1_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_39_internal_tensor_assign_1_end_mask_0 = const()[name = tensor("new_cache_39_internal_tensor_assign_1_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_39_internal_tensor_assign_1_squeeze_mask_0 = const()[name = tensor("new_cache_39_internal_tensor_assign_1_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_86 = const()[name = tensor("shape_86"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_38 = const()[name = tensor("reduce_prod_38"), val = tensor(1048576)]; + tensor range_1d_38_start_0 = const()[name = tensor("range_1d_38_start_0"), val = tensor(0)]; + tensor range_1d_38_step_0 = const()[name = tensor("range_1d_38_step_0"), val = tensor(1)]; + tensor range_1d_38 = range_1d(end = reduce_prod_38, start = range_1d_38_start_0, step = range_1d_38_step_0)[name = tensor("range_1d_38")]; + tensor reshape_190 = reshape(shape = shape_86, x = range_1d_38)[name = tensor("reshape_190")]; + tensor slice_by_index_38 = slice_by_index(begin = concat_135, begin_mask = new_cache_39_internal_tensor_assign_1_begin_mask_0, end = concat_136, end_mask = new_cache_39_internal_tensor_assign_1_end_mask_0, squeeze_mask = new_cache_39_internal_tensor_assign_1_squeeze_mask_0, stride = new_cache_39_internal_tensor_assign_1_stride_0, x = reshape_190)[name = tensor("slice_by_index_38")]; + tensor reshape_191_shape_0 = const()[name = tensor("reshape_191_shape_0"), val = tensor([-1])]; + tensor reshape_191 = reshape(shape = reshape_191_shape_0, x = slice_by_index_38)[name = tensor("reshape_191")]; + tensor reshape_192_shape_0 = const()[name = tensor("reshape_192_shape_0"), val = tensor([-1])]; + tensor reshape_192_cast_fp16 = reshape(shape = reshape_192_shape_0, x = var_7769_cast_fp16)[name = tensor("reshape_192_cast_fp16")]; + tensor reshape_193_shape_0 = const()[name = tensor("reshape_193_shape_0"), val = tensor([-1])]; + tensor reshape_193_cast_fp16 = reshape(shape = reshape_193_shape_0, x = cache19_to_fp16)[name = tensor("reshape_193_cast_fp16")]; + tensor scatter_38_mode_0 = const()[name = tensor("scatter_38_mode_0"), val = tensor("update")]; + tensor scatter_38_axis_0 = const()[name = tensor("scatter_38_axis_0"), val = tensor(0)]; + tensor scatter_38_validate_indices_0 = const()[name = tensor("scatter_38_validate_indices_0"), val = tensor(false)]; + tensor scatter_38_cast_fp16 = scatter(axis = scatter_38_axis_0, data = reshape_193_cast_fp16, indices = reshape_191, mode = scatter_38_mode_0, updates = reshape_192_cast_fp16, validate_indices = scatter_38_validate_indices_0)[name = tensor("scatter_38_cast_fp16")]; + tensor reshape_194_cast_fp16 = reshape(shape = shape_86, x = scatter_38_cast_fp16)[name = tensor("reshape_194_cast_fp16")]; + tensor var_7777_begin_0 = const()[name = tensor("op_7777_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor var_7777_end_0 = const()[name = tensor("op_7777_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor var_7777_end_mask_0 = const()[name = tensor("op_7777_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_7777_squeeze_mask_0 = const()[name = tensor("op_7777_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor var_7777_cast_fp16 = slice_by_index(begin = var_7777_begin_0, end = var_7777_end_0, end_mask = var_7777_end_mask_0, squeeze_mask = var_7777_squeeze_mask_0, x = reshape_194_cast_fp16)[name = tensor("op_7777_cast_fp16")]; + tensor var_7779_axis_0 = const()[name = tensor("op_7779_axis_0"), val = tensor(1)]; + tensor var_7779_mode_0 = const()[name = tensor("op_7779_mode_0"), val = tensor("update")]; + tensor var_7779_validate_indices_0 = const()[name = tensor("op_7779_validate_indices_0"), val = tensor(false)]; + tensor var_7779_cast_fp16 = scatter_along_axis(axis = var_7779_axis_0, data = var_7777_cast_fp16, indices = write_indices_39, mode = var_7779_mode_0, updates = v_39_cast_fp16, validate_indices = var_7779_validate_indices_0)[name = tensor("op_7779_cast_fp16")]; + tensor concat_137 = const()[name = tensor("concat_137"), val = tensor([1, 0, 0, 0, 0])]; + tensor concat_138 = const()[name = tensor("concat_138"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_39_internal_tensor_assign_2_stride_0 = const()[name = tensor("new_cache_39_internal_tensor_assign_2_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_39_internal_tensor_assign_2_begin_mask_0 = const()[name = tensor("new_cache_39_internal_tensor_assign_2_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_39_internal_tensor_assign_2_end_mask_0 = const()[name = tensor("new_cache_39_internal_tensor_assign_2_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_39_internal_tensor_assign_2_squeeze_mask_0 = const()[name = tensor("new_cache_39_internal_tensor_assign_2_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_87 = const()[name = tensor("shape_87"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_39 = const()[name = tensor("reduce_prod_39"), val = tensor(1048576)]; + tensor range_1d_39_start_0 = const()[name = tensor("range_1d_39_start_0"), val = tensor(0)]; + tensor range_1d_39_step_0 = const()[name = tensor("range_1d_39_step_0"), val = tensor(1)]; + tensor range_1d_39 = range_1d(end = reduce_prod_39, start = range_1d_39_start_0, step = range_1d_39_step_0)[name = tensor("range_1d_39")]; + tensor reshape_195 = reshape(shape = shape_87, x = range_1d_39)[name = tensor("reshape_195")]; + tensor slice_by_index_39 = slice_by_index(begin = concat_137, begin_mask = new_cache_39_internal_tensor_assign_2_begin_mask_0, end = concat_138, end_mask = new_cache_39_internal_tensor_assign_2_end_mask_0, squeeze_mask = new_cache_39_internal_tensor_assign_2_squeeze_mask_0, stride = new_cache_39_internal_tensor_assign_2_stride_0, x = reshape_195)[name = tensor("slice_by_index_39")]; + tensor reshape_196_shape_0 = const()[name = tensor("reshape_196_shape_0"), val = tensor([-1])]; + tensor reshape_196 = reshape(shape = reshape_196_shape_0, x = slice_by_index_39)[name = tensor("reshape_196")]; + tensor reshape_197_shape_0 = const()[name = tensor("reshape_197_shape_0"), val = tensor([-1])]; + tensor reshape_197_cast_fp16 = reshape(shape = reshape_197_shape_0, x = var_7779_cast_fp16)[name = tensor("reshape_197_cast_fp16")]; + tensor reshape_198_shape_0 = const()[name = tensor("reshape_198_shape_0"), val = tensor([-1])]; + tensor reshape_198_cast_fp16 = reshape(shape = reshape_198_shape_0, x = reshape_194_cast_fp16)[name = tensor("reshape_198_cast_fp16")]; + tensor scatter_39_mode_0 = const()[name = tensor("scatter_39_mode_0"), val = tensor("update")]; + tensor scatter_39_axis_0 = const()[name = tensor("scatter_39_axis_0"), val = tensor(0)]; + tensor scatter_39_validate_indices_0 = const()[name = tensor("scatter_39_validate_indices_0"), val = tensor(false)]; + tensor scatter_39_cast_fp16 = scatter(axis = scatter_39_axis_0, data = reshape_198_cast_fp16, indices = reshape_196, mode = scatter_39_mode_0, updates = reshape_197_cast_fp16, validate_indices = scatter_39_validate_indices_0)[name = tensor("scatter_39_cast_fp16")]; + tensor reshape_199_cast_fp16 = reshape(shape = shape_87, x = scatter_39_cast_fp16)[name = tensor("reshape_199_cast_fp16")]; + tensor reshape_199_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("reshape_199_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor keys_115_begin_0 = const()[name = tensor("keys_115_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor keys_115_end_0 = const()[name = tensor("keys_115_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor keys_115_end_mask_0 = const()[name = tensor("keys_115_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor keys_115_squeeze_mask_0 = const()[name = tensor("keys_115_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor keys_115_cast_fp16 = slice_by_index(begin = keys_115_begin_0, end = keys_115_end_0, end_mask = keys_115_end_mask_0, squeeze_mask = keys_115_squeeze_mask_0, x = reshape_199_cast_fp16)[name = tensor("keys_115_cast_fp16")]; + tensor values_115_begin_0 = const()[name = tensor("values_115_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor values_115_end_0 = const()[name = tensor("values_115_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor values_115_end_mask_0 = const()[name = tensor("values_115_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor values_115_squeeze_mask_0 = const()[name = tensor("values_115_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor values_115_cast_fp16 = slice_by_index(begin = values_115_begin_0, end = values_115_end_0, end_mask = values_115_end_mask_0, squeeze_mask = values_115_squeeze_mask_0, x = reshape_199_cast_fp16)[name = tensor("values_115_cast_fp16")]; + tensor var_7791_cast_fp16 = not_equal(x = keys_115_cast_fp16, y = keys_115_cast_fp16)[name = tensor("op_7791_cast_fp16")]; + tensor keys_117_cast_fp16 = select(a = var_463_to_fp16, b = keys_115_cast_fp16, cond = var_7791_cast_fp16)[name = tensor("keys_117_cast_fp16")]; + tensor var_7799_cast_fp16 = not_equal(x = values_115_cast_fp16, y = values_115_cast_fp16)[name = tensor("op_7799_cast_fp16")]; + tensor values_117_cast_fp16 = select(a = var_463_to_fp16, b = values_115_cast_fp16, cond = var_7799_cast_fp16)[name = tensor("values_117_cast_fp16")]; + tensor var_7823 = const()[name = tensor("op_7823"), val = tensor([0, 2, 1, 3])]; + tensor var_7839 = const()[name = tensor("op_7839"), val = tensor([1, 1, 1])]; + tensor var_7840_cast_fp16 = reshape(shape = var_7839, x = position19_to_fp16)[name = tensor("op_7840_cast_fp16")]; + tensor var_7860_to_fp16 = const()[name = tensor("op_7860_to_fp16"), val = tensor(0x1p+0)]; + tensor valid_len_39_cast_fp16 = add(x = var_7840_cast_fp16, y = var_7860_to_fp16)[name = tensor("valid_len_39_cast_fp16")]; + tensor valid_mask_39_cast_fp16 = less(x = k_positions_1_promoted_to_fp16, y = valid_len_39_cast_fp16)[name = tensor("valid_mask_39_cast_fp16")]; + tensor causal_mask_39_cast_fp16 = less_equal(x = k_positions_1_promoted_to_fp16, y = var_7840_cast_fp16)[name = tensor("causal_mask_39_cast_fp16")]; + tensor attn_mask_77 = logical_and(x = valid_mask_39_cast_fp16, y = causal_mask_39_cast_fp16)[name = tensor("attn_mask_77")]; + tensor attn_mask_79_axes_0 = const()[name = tensor("attn_mask_79_axes_0"), val = tensor([1])]; + tensor attn_mask_79 = expand_dims(axes = attn_mask_79_axes_0, x = attn_mask_77)[name = tensor("attn_mask_79")]; + tensor var_7878_transpose_x_0 = const()[name = tensor("op_7878_transpose_x_0"), val = tensor(false)]; + tensor var_7878_transpose_y_0 = const()[name = tensor("op_7878_transpose_y_0"), val = tensor(false)]; + tensor transpose_110_perm_0 = const()[name = tensor("transpose_110_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_111_perm_0 = const()[name = tensor("transpose_111_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_111 = transpose(perm = transpose_111_perm_0, x = keys_117_cast_fp16)[name = tensor("transpose_137")]; + tensor transpose_110 = transpose(perm = transpose_110_perm_0, x = q_117_cast_fp16)[name = tensor("transpose_138")]; + tensor var_7878_cast_fp16 = matmul(transpose_x = var_7878_transpose_x_0, transpose_y = var_7878_transpose_y_0, x = transpose_110, y = transpose_111)[name = tensor("op_7878_cast_fp16")]; + tensor var_7872_to_fp16 = const()[name = tensor("op_7872_to_fp16"), val = tensor([0x1p-3])]; + tensor attn_weights_115_cast_fp16 = mul(x = var_7878_cast_fp16, y = var_7872_to_fp16)[name = tensor("attn_weights_115_cast_fp16")]; + tensor var_7880 = logical_not(x = attn_mask_79)[name = tensor("op_7880")]; + tensor var_7881_to_fp16 = const()[name = tensor("op_7881_to_fp16"), val = tensor(-inf)]; + tensor attn_weights_117_cast_fp16 = select(a = var_7881_to_fp16, b = attn_weights_115_cast_fp16, cond = var_7880)[name = tensor("attn_weights_117_cast_fp16")]; + tensor var_7883 = const()[name = tensor("op_7883"), val = tensor(-1)]; + tensor attn_weights_119_cast_fp16 = softmax(axis = var_7883, x = attn_weights_117_cast_fp16)[name = tensor("attn_weights_119_cast_fp16")]; + tensor attn_output_39_transpose_x_0 = const()[name = tensor("attn_output_39_transpose_x_0"), val = tensor(false)]; + tensor attn_output_39_transpose_y_0 = const()[name = tensor("attn_output_39_transpose_y_0"), val = tensor(false)]; + tensor values_119_cast_fp16 = transpose(perm = var_7823, x = values_117_cast_fp16)[name = tensor("transpose_139")]; + tensor attn_output_39_cast_fp16 = matmul(transpose_x = attn_output_39_transpose_x_0, transpose_y = attn_output_39_transpose_y_0, x = attn_weights_119_cast_fp16, y = values_119_cast_fp16)[name = tensor("attn_output_39_cast_fp16")]; + tensor var_7891 = const()[name = tensor("op_7891"), val = tensor([0, 2, 1, 3])]; + tensor var_7896 = const()[name = tensor("op_7896"), val = tensor([1, 1, 1024])]; + tensor var_7892_cast_fp16 = transpose(perm = var_7891, x = attn_output_39_cast_fp16)[name = tensor("transpose_136")]; + tensor input_235_cast_fp16 = reshape(shape = var_7896, x = var_7892_cast_fp16)[name = tensor("input_235_cast_fp16")]; + tensor attn19_out_proj_weight_to_fp16 = const()[name = tensor("attn19_out_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(485746304)))]; + tensor linear_78_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = attn19_out_proj_weight_to_fp16, x = input_235_cast_fp16)[name = tensor("linear_78_cast_fp16")]; + tensor var_7902_to_fp16 = const()[name = tensor("op_7902_to_fp16"), val = tensor(0x1p+0)]; + tensor var_7903_cast_fp16 = add(x = position19_to_fp16, y = var_7902_to_fp16)[name = tensor("op_7903_cast_fp16")]; + tensor var_7903_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_7903_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor input_237_cast_fp16 = add(x = input_231_cast_fp16, y = linear_78_cast_fp16)[name = tensor("input_237_cast_fp16")]; + tensor input_239_axes_0 = const()[name = tensor("input_239_axes_0"), val = tensor([-1])]; + tensor norm19_2_weight_to_fp16 = const()[name = tensor("norm19_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(487843520)))]; + tensor norm19_2_bias_to_fp16 = const()[name = tensor("norm19_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(487845632)))]; + tensor var_7907_to_fp16 = const()[name = tensor("op_7907_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_239_cast_fp16 = layer_norm(axes = input_239_axes_0, beta = norm19_2_bias_to_fp16, epsilon = var_7907_to_fp16, gamma = norm19_2_weight_to_fp16, x = input_237_cast_fp16)[name = tensor("input_239_cast_fp16")]; + tensor linear19_1_weight_to_fp16 = const()[name = tensor("linear19_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(487847744)))]; + tensor linear_79_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = linear19_1_weight_to_fp16, x = input_239_cast_fp16)[name = tensor("linear_79_cast_fp16")]; + tensor input_241_mode_0 = const()[name = tensor("input_241_mode_0"), val = tensor("EXACT")]; + tensor input_241_cast_fp16 = gelu(mode = input_241_mode_0, x = linear_79_cast_fp16)[name = tensor("input_241_cast_fp16")]; + tensor linear19_2_weight_to_fp16 = const()[name = tensor("linear19_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(496236416)))]; + tensor linear_80_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = linear19_2_weight_to_fp16, x = input_241_cast_fp16)[name = tensor("linear_80_cast_fp16")]; + tensor input_243_cast_fp16 = add(x = input_237_cast_fp16, y = linear_80_cast_fp16)[name = tensor("input_243_cast_fp16")]; + tensor input_245_axes_0 = const()[name = tensor("input_245_axes_0"), val = tensor([-1])]; + tensor norm20_1_weight_to_fp16 = const()[name = tensor("norm20_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(504625088)))]; + tensor norm20_1_bias_to_fp16 = const()[name = tensor("norm20_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(504627200)))]; + tensor var_7924_to_fp16 = const()[name = tensor("op_7924_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_245_cast_fp16 = layer_norm(axes = input_245_axes_0, beta = norm20_1_bias_to_fp16, epsilon = var_7924_to_fp16, gamma = norm20_1_weight_to_fp16, x = input_243_cast_fp16)[name = tensor("input_245_cast_fp16")]; + tensor attn20_in_proj_weight_to_fp16 = const()[name = tensor("attn20_in_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(504629312)))]; + tensor linear_81_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = attn20_in_proj_weight_to_fp16, x = input_245_cast_fp16)[name = tensor("linear_81_cast_fp16")]; + tensor var_7938 = const()[name = tensor("op_7938"), val = tensor([1, 1, 3, 16, 64])]; + tensor qkv_41_cast_fp16 = reshape(shape = var_7938, x = linear_81_cast_fp16)[name = tensor("qkv_41_cast_fp16")]; + tensor q_121_begin_0 = const()[name = tensor("q_121_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor q_121_end_0 = const()[name = tensor("q_121_end_0"), val = tensor([1, 1, 1, 16, 64])]; + tensor q_121_end_mask_0 = const()[name = tensor("q_121_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor q_121_squeeze_mask_0 = const()[name = tensor("q_121_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor q_121_cast_fp16 = slice_by_index(begin = q_121_begin_0, end = q_121_end_0, end_mask = q_121_end_mask_0, squeeze_mask = q_121_squeeze_mask_0, x = qkv_41_cast_fp16)[name = tensor("q_121_cast_fp16")]; + tensor k_81_begin_0 = const()[name = tensor("k_81_begin_0"), val = tensor([0, 0, 1, 0, 0])]; + tensor k_81_end_0 = const()[name = tensor("k_81_end_0"), val = tensor([1, 1, 2, 16, 64])]; + tensor k_81_end_mask_0 = const()[name = tensor("k_81_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor k_81_squeeze_mask_0 = const()[name = tensor("k_81_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor k_81_cast_fp16 = slice_by_index(begin = k_81_begin_0, end = k_81_end_0, end_mask = k_81_end_mask_0, squeeze_mask = k_81_squeeze_mask_0, x = qkv_41_cast_fp16)[name = tensor("k_81_cast_fp16")]; + tensor v_41_begin_0 = const()[name = tensor("v_41_begin_0"), val = tensor([0, 0, 2, 0, 0])]; + tensor v_41_end_0 = const()[name = tensor("v_41_end_0"), val = tensor([1, 1, 3, 16, 64])]; + tensor v_41_end_mask_0 = const()[name = tensor("v_41_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor v_41_squeeze_mask_0 = const()[name = tensor("v_41_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor v_41_cast_fp16 = slice_by_index(begin = v_41_begin_0, end = v_41_end_0, end_mask = v_41_end_mask_0, squeeze_mask = v_41_squeeze_mask_0, x = qkv_41_cast_fp16)[name = tensor("v_41_cast_fp16")]; + tensor var_8004 = const()[name = tensor("op_8004"), val = tensor([1, 1, 1, 1])]; + tensor position20_to_fp16_dtype_0 = const()[name = tensor("position20_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor position20_to_fp16 = cast(dtype = position20_to_fp16_dtype_0, x = position20)[name = tensor("cast_453")]; + tensor ts_125_cast_fp16 = reshape(shape = var_8004, x = position20_to_fp16)[name = tensor("ts_125_cast_fp16")]; + tensor var_8011 = const()[name = tensor("op_8011"), val = tensor([1, 1, 16, 32, 2])]; + tensor q_complex_41_cast_fp16 = reshape(shape = var_8011, x = q_121_cast_fp16)[name = tensor("q_complex_41_cast_fp16")]; + tensor var_8018 = const()[name = tensor("op_8018"), val = tensor([1, 1, 16, 32, 2])]; + tensor k_complex_41_cast_fp16 = reshape(shape = var_8018, x = k_81_cast_fp16)[name = tensor("k_complex_41_cast_fp16")]; + tensor var_8022_begin_0 = const()[name = tensor("op_8022_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_8022_end_0 = const()[name = tensor("op_8022_end_0"), val = tensor([1, 1, 16, 32, 1])]; + tensor var_8022_end_mask_0 = const()[name = tensor("op_8022_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_8022_squeeze_mask_0 = const()[name = tensor("op_8022_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_8022_cast_fp16 = slice_by_index(begin = var_8022_begin_0, end = var_8022_end_0, end_mask = var_8022_end_mask_0, squeeze_mask = var_8022_squeeze_mask_0, x = q_complex_41_cast_fp16)[name = tensor("op_8022_cast_fp16")]; + tensor var_8030_begin_0 = const()[name = tensor("op_8030_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_8030_end_0 = const()[name = tensor("op_8030_end_0"), val = tensor([1, 1, 16, 32, 2])]; + tensor var_8030_end_mask_0 = const()[name = tensor("op_8030_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_8030_squeeze_mask_0 = const()[name = tensor("op_8030_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_8030_cast_fp16 = slice_by_index(begin = var_8030_begin_0, end = var_8030_end_0, end_mask = var_8030_end_mask_0, squeeze_mask = var_8030_squeeze_mask_0, x = q_complex_41_cast_fp16)[name = tensor("op_8030_cast_fp16")]; + tensor var_8038_begin_0 = const()[name = tensor("op_8038_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_8038_end_0 = const()[name = tensor("op_8038_end_0"), val = tensor([1, 1, 16, 32, 1])]; + tensor var_8038_end_mask_0 = const()[name = tensor("op_8038_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_8038_squeeze_mask_0 = const()[name = tensor("op_8038_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_8038_cast_fp16 = slice_by_index(begin = var_8038_begin_0, end = var_8038_end_0, end_mask = var_8038_end_mask_0, squeeze_mask = var_8038_squeeze_mask_0, x = k_complex_41_cast_fp16)[name = tensor("op_8038_cast_fp16")]; + tensor var_8046_begin_0 = const()[name = tensor("op_8046_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_8046_end_0 = const()[name = tensor("op_8046_end_0"), val = tensor([1, 1, 16, 32, 2])]; + tensor var_8046_end_mask_0 = const()[name = tensor("op_8046_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_8046_squeeze_mask_0 = const()[name = tensor("op_8046_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_8046_cast_fp16 = slice_by_index(begin = var_8046_begin_0, end = var_8046_end_0, end_mask = var_8046_end_mask_0, squeeze_mask = var_8046_squeeze_mask_0, x = k_complex_41_cast_fp16)[name = tensor("op_8046_cast_fp16")]; + tensor freqs_41_to_fp16 = const()[name = tensor("freqs_41_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(510920832)))]; + tensor var_8052_cast_fp16 = mul(x = freqs_41_to_fp16, y = ts_125_cast_fp16)[name = tensor("op_8052_cast_fp16")]; + tensor rotr_41_cast_fp16 = cos(x = var_8052_cast_fp16)[name = tensor("rotr_41_cast_fp16")]; + tensor roti_41_cast_fp16 = sin(x = var_8052_cast_fp16)[name = tensor("roti_41_cast_fp16")]; + tensor var_8056_cast_fp16 = mul(x = var_8022_cast_fp16, y = rotr_41_cast_fp16)[name = tensor("op_8056_cast_fp16")]; + tensor var_8057_cast_fp16 = mul(x = var_8030_cast_fp16, y = roti_41_cast_fp16)[name = tensor("op_8057_cast_fp16")]; + tensor qor_81_cast_fp16 = sub(x = var_8056_cast_fp16, y = var_8057_cast_fp16)[name = tensor("qor_81_cast_fp16")]; + tensor var_8060_cast_fp16 = mul(x = var_8022_cast_fp16, y = roti_41_cast_fp16)[name = tensor("op_8060_cast_fp16")]; + tensor var_8061_cast_fp16 = mul(x = var_8030_cast_fp16, y = rotr_41_cast_fp16)[name = tensor("op_8061_cast_fp16")]; + tensor qoi_81_cast_fp16 = add(x = var_8060_cast_fp16, y = var_8061_cast_fp16)[name = tensor("qoi_81_cast_fp16")]; + tensor var_8064_cast_fp16 = mul(x = var_8038_cast_fp16, y = rotr_41_cast_fp16)[name = tensor("op_8064_cast_fp16")]; + tensor var_8065_cast_fp16 = mul(x = var_8046_cast_fp16, y = roti_41_cast_fp16)[name = tensor("op_8065_cast_fp16")]; + tensor kor_81_cast_fp16 = sub(x = var_8064_cast_fp16, y = var_8065_cast_fp16)[name = tensor("kor_81_cast_fp16")]; + tensor var_8068_cast_fp16 = mul(x = var_8038_cast_fp16, y = roti_41_cast_fp16)[name = tensor("op_8068_cast_fp16")]; + tensor var_8069_cast_fp16 = mul(x = var_8046_cast_fp16, y = rotr_41_cast_fp16)[name = tensor("op_8069_cast_fp16")]; + tensor koi_81_cast_fp16 = add(x = var_8068_cast_fp16, y = var_8069_cast_fp16)[name = tensor("koi_81_cast_fp16")]; + tensor qo_41_axis_0 = const()[name = tensor("qo_41_axis_0"), val = tensor(-1)]; + tensor qo_41_cast_fp16 = stack(axis = qo_41_axis_0, values = (qor_81_cast_fp16, qoi_81_cast_fp16))[name = tensor("qo_41_cast_fp16")]; + tensor ko_41_axis_0 = const()[name = tensor("ko_41_axis_0"), val = tensor(-1)]; + tensor ko_41_cast_fp16 = stack(axis = ko_41_axis_0, values = (kor_81_cast_fp16, koi_81_cast_fp16))[name = tensor("ko_41_cast_fp16")]; + tensor var_8102 = const()[name = tensor("op_8102"), val = tensor([1, 1, 16, 64])]; + tensor q_123_cast_fp16 = reshape(shape = var_8102, x = qo_41_cast_fp16)[name = tensor("q_123_cast_fp16")]; + tensor var_8108 = const()[name = tensor("op_8108"), val = tensor([1, 1, 16, 64])]; + tensor k_83_cast_fp16 = reshape(shape = var_8108, x = ko_41_cast_fp16)[name = tensor("k_83_cast_fp16")]; + tensor _inversed_8133_y_0_to_fp16 = const()[name = tensor("_inversed_8133_y_0_to_fp16"), val = tensor(0x1p-9)]; + tensor _inversed_8133_cast_fp16 = mul(x = ts_125_cast_fp16, y = _inversed_8133_y_0_to_fp16)[name = tensor("_inversed_8133_cast_fp16")]; + tensor var_8134_cast_fp16 = floor(x = _inversed_8133_cast_fp16)[name = tensor("op_8134_cast_fp16")]; + tensor var_8135_to_fp16 = const()[name = tensor("op_8135_to_fp16"), val = tensor(0x1p+9)]; + tensor var_8136_cast_fp16 = mul(x = var_8134_cast_fp16, y = var_8135_to_fp16)[name = tensor("op_8136_cast_fp16")]; + tensor write_indices_float_83_cast_fp16 = sub(x = ts_125_cast_fp16, y = var_8136_cast_fp16)[name = tensor("write_indices_float_83_cast_fp16")]; + tensor var_8143_dtype_0 = const()[name = tensor("op_8143_dtype_0"), val = tensor("int32")]; + tensor write_indices_41_reps_0 = const()[name = tensor("write_indices_41_reps_0"), val = tensor([1, 1, 16, 64])]; + tensor write_indices_float_83_cast_fp16_to_int32 = cast(dtype = var_8143_dtype_0, x = write_indices_float_83_cast_fp16)[name = tensor("cast_452")]; + tensor write_indices_41 = tile(reps = write_indices_41_reps_0, x = write_indices_float_83_cast_fp16_to_int32)[name = tensor("write_indices_41")]; + tensor var_8153_begin_0 = const()[name = tensor("op_8153_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_8153_end_0 = const()[name = tensor("op_8153_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor var_8153_end_mask_0 = const()[name = tensor("op_8153_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_8153_squeeze_mask_0 = const()[name = tensor("op_8153_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor cache20_to_fp16_dtype_0 = const()[name = tensor("cache20_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor cache20_to_fp16 = cast(dtype = cache20_to_fp16_dtype_0, x = cache20)[name = tensor("cast_451")]; + tensor var_8153_cast_fp16 = slice_by_index(begin = var_8153_begin_0, end = var_8153_end_0, end_mask = var_8153_end_mask_0, squeeze_mask = var_8153_squeeze_mask_0, x = cache20_to_fp16)[name = tensor("op_8153_cast_fp16")]; + tensor var_8155_axis_0 = const()[name = tensor("op_8155_axis_0"), val = tensor(1)]; + tensor var_8155_mode_0 = const()[name = tensor("op_8155_mode_0"), val = tensor("update")]; + tensor var_8155_validate_indices_0 = const()[name = tensor("op_8155_validate_indices_0"), val = tensor(false)]; + tensor var_8155_cast_fp16 = scatter_along_axis(axis = var_8155_axis_0, data = var_8153_cast_fp16, indices = write_indices_41, mode = var_8155_mode_0, updates = k_83_cast_fp16, validate_indices = var_8155_validate_indices_0)[name = tensor("op_8155_cast_fp16")]; + tensor concat_142 = const()[name = tensor("concat_142"), val = tensor([0, 0, 0, 0, 0])]; + tensor concat_143 = const()[name = tensor("concat_143"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_41_internal_tensor_assign_1_stride_0 = const()[name = tensor("new_cache_41_internal_tensor_assign_1_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_41_internal_tensor_assign_1_begin_mask_0 = const()[name = tensor("new_cache_41_internal_tensor_assign_1_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_41_internal_tensor_assign_1_end_mask_0 = const()[name = tensor("new_cache_41_internal_tensor_assign_1_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_41_internal_tensor_assign_1_squeeze_mask_0 = const()[name = tensor("new_cache_41_internal_tensor_assign_1_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_88 = const()[name = tensor("shape_88"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_40 = const()[name = tensor("reduce_prod_40"), val = tensor(1048576)]; + tensor range_1d_40_start_0 = const()[name = tensor("range_1d_40_start_0"), val = tensor(0)]; + tensor range_1d_40_step_0 = const()[name = tensor("range_1d_40_step_0"), val = tensor(1)]; + tensor range_1d_40 = range_1d(end = reduce_prod_40, start = range_1d_40_start_0, step = range_1d_40_step_0)[name = tensor("range_1d_40")]; + tensor reshape_200 = reshape(shape = shape_88, x = range_1d_40)[name = tensor("reshape_200")]; + tensor slice_by_index_40 = slice_by_index(begin = concat_142, begin_mask = new_cache_41_internal_tensor_assign_1_begin_mask_0, end = concat_143, end_mask = new_cache_41_internal_tensor_assign_1_end_mask_0, squeeze_mask = new_cache_41_internal_tensor_assign_1_squeeze_mask_0, stride = new_cache_41_internal_tensor_assign_1_stride_0, x = reshape_200)[name = tensor("slice_by_index_40")]; + tensor reshape_201_shape_0 = const()[name = tensor("reshape_201_shape_0"), val = tensor([-1])]; + tensor reshape_201 = reshape(shape = reshape_201_shape_0, x = slice_by_index_40)[name = tensor("reshape_201")]; + tensor reshape_202_shape_0 = const()[name = tensor("reshape_202_shape_0"), val = tensor([-1])]; + tensor reshape_202_cast_fp16 = reshape(shape = reshape_202_shape_0, x = var_8155_cast_fp16)[name = tensor("reshape_202_cast_fp16")]; + tensor reshape_203_shape_0 = const()[name = tensor("reshape_203_shape_0"), val = tensor([-1])]; + tensor reshape_203_cast_fp16 = reshape(shape = reshape_203_shape_0, x = cache20_to_fp16)[name = tensor("reshape_203_cast_fp16")]; + tensor scatter_40_mode_0 = const()[name = tensor("scatter_40_mode_0"), val = tensor("update")]; + tensor scatter_40_axis_0 = const()[name = tensor("scatter_40_axis_0"), val = tensor(0)]; + tensor scatter_40_validate_indices_0 = const()[name = tensor("scatter_40_validate_indices_0"), val = tensor(false)]; + tensor scatter_40_cast_fp16 = scatter(axis = scatter_40_axis_0, data = reshape_203_cast_fp16, indices = reshape_201, mode = scatter_40_mode_0, updates = reshape_202_cast_fp16, validate_indices = scatter_40_validate_indices_0)[name = tensor("scatter_40_cast_fp16")]; + tensor reshape_204_cast_fp16 = reshape(shape = shape_88, x = scatter_40_cast_fp16)[name = tensor("reshape_204_cast_fp16")]; + tensor var_8163_begin_0 = const()[name = tensor("op_8163_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor var_8163_end_0 = const()[name = tensor("op_8163_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor var_8163_end_mask_0 = const()[name = tensor("op_8163_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_8163_squeeze_mask_0 = const()[name = tensor("op_8163_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor var_8163_cast_fp16 = slice_by_index(begin = var_8163_begin_0, end = var_8163_end_0, end_mask = var_8163_end_mask_0, squeeze_mask = var_8163_squeeze_mask_0, x = reshape_204_cast_fp16)[name = tensor("op_8163_cast_fp16")]; + tensor var_8165_axis_0 = const()[name = tensor("op_8165_axis_0"), val = tensor(1)]; + tensor var_8165_mode_0 = const()[name = tensor("op_8165_mode_0"), val = tensor("update")]; + tensor var_8165_validate_indices_0 = const()[name = tensor("op_8165_validate_indices_0"), val = tensor(false)]; + tensor var_8165_cast_fp16 = scatter_along_axis(axis = var_8165_axis_0, data = var_8163_cast_fp16, indices = write_indices_41, mode = var_8165_mode_0, updates = v_41_cast_fp16, validate_indices = var_8165_validate_indices_0)[name = tensor("op_8165_cast_fp16")]; + tensor concat_144 = const()[name = tensor("concat_144"), val = tensor([1, 0, 0, 0, 0])]; + tensor concat_145 = const()[name = tensor("concat_145"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_41_internal_tensor_assign_2_stride_0 = const()[name = tensor("new_cache_41_internal_tensor_assign_2_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_41_internal_tensor_assign_2_begin_mask_0 = const()[name = tensor("new_cache_41_internal_tensor_assign_2_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_41_internal_tensor_assign_2_end_mask_0 = const()[name = tensor("new_cache_41_internal_tensor_assign_2_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_41_internal_tensor_assign_2_squeeze_mask_0 = const()[name = tensor("new_cache_41_internal_tensor_assign_2_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_89 = const()[name = tensor("shape_89"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_41 = const()[name = tensor("reduce_prod_41"), val = tensor(1048576)]; + tensor range_1d_41_start_0 = const()[name = tensor("range_1d_41_start_0"), val = tensor(0)]; + tensor range_1d_41_step_0 = const()[name = tensor("range_1d_41_step_0"), val = tensor(1)]; + tensor range_1d_41 = range_1d(end = reduce_prod_41, start = range_1d_41_start_0, step = range_1d_41_step_0)[name = tensor("range_1d_41")]; + tensor reshape_205 = reshape(shape = shape_89, x = range_1d_41)[name = tensor("reshape_205")]; + tensor slice_by_index_41 = slice_by_index(begin = concat_144, begin_mask = new_cache_41_internal_tensor_assign_2_begin_mask_0, end = concat_145, end_mask = new_cache_41_internal_tensor_assign_2_end_mask_0, squeeze_mask = new_cache_41_internal_tensor_assign_2_squeeze_mask_0, stride = new_cache_41_internal_tensor_assign_2_stride_0, x = reshape_205)[name = tensor("slice_by_index_41")]; + tensor reshape_206_shape_0 = const()[name = tensor("reshape_206_shape_0"), val = tensor([-1])]; + tensor reshape_206 = reshape(shape = reshape_206_shape_0, x = slice_by_index_41)[name = tensor("reshape_206")]; + tensor reshape_207_shape_0 = const()[name = tensor("reshape_207_shape_0"), val = tensor([-1])]; + tensor reshape_207_cast_fp16 = reshape(shape = reshape_207_shape_0, x = var_8165_cast_fp16)[name = tensor("reshape_207_cast_fp16")]; + tensor reshape_208_shape_0 = const()[name = tensor("reshape_208_shape_0"), val = tensor([-1])]; + tensor reshape_208_cast_fp16 = reshape(shape = reshape_208_shape_0, x = reshape_204_cast_fp16)[name = tensor("reshape_208_cast_fp16")]; + tensor scatter_41_mode_0 = const()[name = tensor("scatter_41_mode_0"), val = tensor("update")]; + tensor scatter_41_axis_0 = const()[name = tensor("scatter_41_axis_0"), val = tensor(0)]; + tensor scatter_41_validate_indices_0 = const()[name = tensor("scatter_41_validate_indices_0"), val = tensor(false)]; + tensor scatter_41_cast_fp16 = scatter(axis = scatter_41_axis_0, data = reshape_208_cast_fp16, indices = reshape_206, mode = scatter_41_mode_0, updates = reshape_207_cast_fp16, validate_indices = scatter_41_validate_indices_0)[name = tensor("scatter_41_cast_fp16")]; + tensor reshape_209_cast_fp16 = reshape(shape = shape_89, x = scatter_41_cast_fp16)[name = tensor("reshape_209_cast_fp16")]; + tensor reshape_209_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("reshape_209_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor keys_121_begin_0 = const()[name = tensor("keys_121_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor keys_121_end_0 = const()[name = tensor("keys_121_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor keys_121_end_mask_0 = const()[name = tensor("keys_121_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor keys_121_squeeze_mask_0 = const()[name = tensor("keys_121_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor keys_121_cast_fp16 = slice_by_index(begin = keys_121_begin_0, end = keys_121_end_0, end_mask = keys_121_end_mask_0, squeeze_mask = keys_121_squeeze_mask_0, x = reshape_209_cast_fp16)[name = tensor("keys_121_cast_fp16")]; + tensor values_121_begin_0 = const()[name = tensor("values_121_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor values_121_end_0 = const()[name = tensor("values_121_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor values_121_end_mask_0 = const()[name = tensor("values_121_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor values_121_squeeze_mask_0 = const()[name = tensor("values_121_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor values_121_cast_fp16 = slice_by_index(begin = values_121_begin_0, end = values_121_end_0, end_mask = values_121_end_mask_0, squeeze_mask = values_121_squeeze_mask_0, x = reshape_209_cast_fp16)[name = tensor("values_121_cast_fp16")]; + tensor var_8177_cast_fp16 = not_equal(x = keys_121_cast_fp16, y = keys_121_cast_fp16)[name = tensor("op_8177_cast_fp16")]; + tensor keys_123_cast_fp16 = select(a = var_463_to_fp16, b = keys_121_cast_fp16, cond = var_8177_cast_fp16)[name = tensor("keys_123_cast_fp16")]; + tensor var_8185_cast_fp16 = not_equal(x = values_121_cast_fp16, y = values_121_cast_fp16)[name = tensor("op_8185_cast_fp16")]; + tensor values_123_cast_fp16 = select(a = var_463_to_fp16, b = values_121_cast_fp16, cond = var_8185_cast_fp16)[name = tensor("values_123_cast_fp16")]; + tensor var_8209 = const()[name = tensor("op_8209"), val = tensor([0, 2, 1, 3])]; + tensor var_8225 = const()[name = tensor("op_8225"), val = tensor([1, 1, 1])]; + tensor var_8226_cast_fp16 = reshape(shape = var_8225, x = position20_to_fp16)[name = tensor("op_8226_cast_fp16")]; + tensor var_8246_to_fp16 = const()[name = tensor("op_8246_to_fp16"), val = tensor(0x1p+0)]; + tensor valid_len_41_cast_fp16 = add(x = var_8226_cast_fp16, y = var_8246_to_fp16)[name = tensor("valid_len_41_cast_fp16")]; + tensor valid_mask_41_cast_fp16 = less(x = k_positions_1_promoted_to_fp16, y = valid_len_41_cast_fp16)[name = tensor("valid_mask_41_cast_fp16")]; + tensor causal_mask_41_cast_fp16 = less_equal(x = k_positions_1_promoted_to_fp16, y = var_8226_cast_fp16)[name = tensor("causal_mask_41_cast_fp16")]; + tensor attn_mask_81 = logical_and(x = valid_mask_41_cast_fp16, y = causal_mask_41_cast_fp16)[name = tensor("attn_mask_81")]; + tensor attn_mask_83_axes_0 = const()[name = tensor("attn_mask_83_axes_0"), val = tensor([1])]; + tensor attn_mask_83 = expand_dims(axes = attn_mask_83_axes_0, x = attn_mask_81)[name = tensor("attn_mask_83")]; + tensor var_8264_transpose_x_0 = const()[name = tensor("op_8264_transpose_x_0"), val = tensor(false)]; + tensor var_8264_transpose_y_0 = const()[name = tensor("op_8264_transpose_y_0"), val = tensor(false)]; + tensor transpose_112_perm_0 = const()[name = tensor("transpose_112_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_113_perm_0 = const()[name = tensor("transpose_113_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_113 = transpose(perm = transpose_113_perm_0, x = keys_123_cast_fp16)[name = tensor("transpose_133")]; + tensor transpose_112 = transpose(perm = transpose_112_perm_0, x = q_123_cast_fp16)[name = tensor("transpose_134")]; + tensor var_8264_cast_fp16 = matmul(transpose_x = var_8264_transpose_x_0, transpose_y = var_8264_transpose_y_0, x = transpose_112, y = transpose_113)[name = tensor("op_8264_cast_fp16")]; + tensor var_8258_to_fp16 = const()[name = tensor("op_8258_to_fp16"), val = tensor([0x1p-3])]; + tensor attn_weights_121_cast_fp16 = mul(x = var_8264_cast_fp16, y = var_8258_to_fp16)[name = tensor("attn_weights_121_cast_fp16")]; + tensor var_8266 = logical_not(x = attn_mask_83)[name = tensor("op_8266")]; + tensor var_8267_to_fp16 = const()[name = tensor("op_8267_to_fp16"), val = tensor(-inf)]; + tensor attn_weights_123_cast_fp16 = select(a = var_8267_to_fp16, b = attn_weights_121_cast_fp16, cond = var_8266)[name = tensor("attn_weights_123_cast_fp16")]; + tensor var_8269 = const()[name = tensor("op_8269"), val = tensor(-1)]; + tensor attn_weights_125_cast_fp16 = softmax(axis = var_8269, x = attn_weights_123_cast_fp16)[name = tensor("attn_weights_125_cast_fp16")]; + tensor attn_output_41_transpose_x_0 = const()[name = tensor("attn_output_41_transpose_x_0"), val = tensor(false)]; + tensor attn_output_41_transpose_y_0 = const()[name = tensor("attn_output_41_transpose_y_0"), val = tensor(false)]; + tensor values_125_cast_fp16 = transpose(perm = var_8209, x = values_123_cast_fp16)[name = tensor("transpose_135")]; + tensor attn_output_41_cast_fp16 = matmul(transpose_x = attn_output_41_transpose_x_0, transpose_y = attn_output_41_transpose_y_0, x = attn_weights_125_cast_fp16, y = values_125_cast_fp16)[name = tensor("attn_output_41_cast_fp16")]; + tensor var_8277 = const()[name = tensor("op_8277"), val = tensor([0, 2, 1, 3])]; + tensor var_8282 = const()[name = tensor("op_8282"), val = tensor([1, 1, 1024])]; + tensor var_8278_cast_fp16 = transpose(perm = var_8277, x = attn_output_41_cast_fp16)[name = tensor("transpose_132")]; + tensor input_247_cast_fp16 = reshape(shape = var_8282, x = var_8278_cast_fp16)[name = tensor("input_247_cast_fp16")]; + tensor attn20_out_proj_weight_to_fp16 = const()[name = tensor("attn20_out_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(510920960)))]; + tensor linear_82_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = attn20_out_proj_weight_to_fp16, x = input_247_cast_fp16)[name = tensor("linear_82_cast_fp16")]; + tensor var_8288_to_fp16 = const()[name = tensor("op_8288_to_fp16"), val = tensor(0x1p+0)]; + tensor var_8289_cast_fp16 = add(x = position20_to_fp16, y = var_8288_to_fp16)[name = tensor("op_8289_cast_fp16")]; + tensor var_8289_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_8289_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor input_249_cast_fp16 = add(x = input_243_cast_fp16, y = linear_82_cast_fp16)[name = tensor("input_249_cast_fp16")]; + tensor input_251_axes_0 = const()[name = tensor("input_251_axes_0"), val = tensor([-1])]; + tensor norm20_2_weight_to_fp16 = const()[name = tensor("norm20_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(513018176)))]; + tensor norm20_2_bias_to_fp16 = const()[name = tensor("norm20_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(513020288)))]; + tensor var_8293_to_fp16 = const()[name = tensor("op_8293_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_251_cast_fp16 = layer_norm(axes = input_251_axes_0, beta = norm20_2_bias_to_fp16, epsilon = var_8293_to_fp16, gamma = norm20_2_weight_to_fp16, x = input_249_cast_fp16)[name = tensor("input_251_cast_fp16")]; + tensor linear20_1_weight_to_fp16 = const()[name = tensor("linear20_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(513022400)))]; + tensor linear_83_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = linear20_1_weight_to_fp16, x = input_251_cast_fp16)[name = tensor("linear_83_cast_fp16")]; + tensor input_253_mode_0 = const()[name = tensor("input_253_mode_0"), val = tensor("EXACT")]; + tensor input_253_cast_fp16 = gelu(mode = input_253_mode_0, x = linear_83_cast_fp16)[name = tensor("input_253_cast_fp16")]; + tensor linear20_2_weight_to_fp16 = const()[name = tensor("linear20_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(521411072)))]; + tensor linear_84_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = linear20_2_weight_to_fp16, x = input_253_cast_fp16)[name = tensor("linear_84_cast_fp16")]; + tensor input_255_cast_fp16 = add(x = input_249_cast_fp16, y = linear_84_cast_fp16)[name = tensor("input_255_cast_fp16")]; + tensor input_257_axes_0 = const()[name = tensor("input_257_axes_0"), val = tensor([-1])]; + tensor norm21_1_weight_to_fp16 = const()[name = tensor("norm21_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(529799744)))]; + tensor norm21_1_bias_to_fp16 = const()[name = tensor("norm21_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(529801856)))]; + tensor var_8310_to_fp16 = const()[name = tensor("op_8310_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_257_cast_fp16 = layer_norm(axes = input_257_axes_0, beta = norm21_1_bias_to_fp16, epsilon = var_8310_to_fp16, gamma = norm21_1_weight_to_fp16, x = input_255_cast_fp16)[name = tensor("input_257_cast_fp16")]; + tensor attn21_in_proj_weight_to_fp16 = const()[name = tensor("attn21_in_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(529803968)))]; + tensor linear_85_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = attn21_in_proj_weight_to_fp16, x = input_257_cast_fp16)[name = tensor("linear_85_cast_fp16")]; + tensor var_8324 = const()[name = tensor("op_8324"), val = tensor([1, 1, 3, 16, 64])]; + tensor qkv_43_cast_fp16 = reshape(shape = var_8324, x = linear_85_cast_fp16)[name = tensor("qkv_43_cast_fp16")]; + tensor q_127_begin_0 = const()[name = tensor("q_127_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor q_127_end_0 = const()[name = tensor("q_127_end_0"), val = tensor([1, 1, 1, 16, 64])]; + tensor q_127_end_mask_0 = const()[name = tensor("q_127_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor q_127_squeeze_mask_0 = const()[name = tensor("q_127_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor q_127_cast_fp16 = slice_by_index(begin = q_127_begin_0, end = q_127_end_0, end_mask = q_127_end_mask_0, squeeze_mask = q_127_squeeze_mask_0, x = qkv_43_cast_fp16)[name = tensor("q_127_cast_fp16")]; + tensor k_85_begin_0 = const()[name = tensor("k_85_begin_0"), val = tensor([0, 0, 1, 0, 0])]; + tensor k_85_end_0 = const()[name = tensor("k_85_end_0"), val = tensor([1, 1, 2, 16, 64])]; + tensor k_85_end_mask_0 = const()[name = tensor("k_85_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor k_85_squeeze_mask_0 = const()[name = tensor("k_85_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor k_85_cast_fp16 = slice_by_index(begin = k_85_begin_0, end = k_85_end_0, end_mask = k_85_end_mask_0, squeeze_mask = k_85_squeeze_mask_0, x = qkv_43_cast_fp16)[name = tensor("k_85_cast_fp16")]; + tensor v_43_begin_0 = const()[name = tensor("v_43_begin_0"), val = tensor([0, 0, 2, 0, 0])]; + tensor v_43_end_0 = const()[name = tensor("v_43_end_0"), val = tensor([1, 1, 3, 16, 64])]; + tensor v_43_end_mask_0 = const()[name = tensor("v_43_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor v_43_squeeze_mask_0 = const()[name = tensor("v_43_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor v_43_cast_fp16 = slice_by_index(begin = v_43_begin_0, end = v_43_end_0, end_mask = v_43_end_mask_0, squeeze_mask = v_43_squeeze_mask_0, x = qkv_43_cast_fp16)[name = tensor("v_43_cast_fp16")]; + tensor var_8390 = const()[name = tensor("op_8390"), val = tensor([1, 1, 1, 1])]; + tensor position21_to_fp16_dtype_0 = const()[name = tensor("position21_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor position21_to_fp16 = cast(dtype = position21_to_fp16_dtype_0, x = position21)[name = tensor("cast_448")]; + tensor ts_131_cast_fp16 = reshape(shape = var_8390, x = position21_to_fp16)[name = tensor("ts_131_cast_fp16")]; + tensor var_8397 = const()[name = tensor("op_8397"), val = tensor([1, 1, 16, 32, 2])]; + tensor q_complex_43_cast_fp16 = reshape(shape = var_8397, x = q_127_cast_fp16)[name = tensor("q_complex_43_cast_fp16")]; + tensor var_8404 = const()[name = tensor("op_8404"), val = tensor([1, 1, 16, 32, 2])]; + tensor k_complex_43_cast_fp16 = reshape(shape = var_8404, x = k_85_cast_fp16)[name = tensor("k_complex_43_cast_fp16")]; + tensor var_8408_begin_0 = const()[name = tensor("op_8408_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_8408_end_0 = const()[name = tensor("op_8408_end_0"), val = tensor([1, 1, 16, 32, 1])]; + tensor var_8408_end_mask_0 = const()[name = tensor("op_8408_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_8408_squeeze_mask_0 = const()[name = tensor("op_8408_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_8408_cast_fp16 = slice_by_index(begin = var_8408_begin_0, end = var_8408_end_0, end_mask = var_8408_end_mask_0, squeeze_mask = var_8408_squeeze_mask_0, x = q_complex_43_cast_fp16)[name = tensor("op_8408_cast_fp16")]; + tensor var_8416_begin_0 = const()[name = tensor("op_8416_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_8416_end_0 = const()[name = tensor("op_8416_end_0"), val = tensor([1, 1, 16, 32, 2])]; + tensor var_8416_end_mask_0 = const()[name = tensor("op_8416_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_8416_squeeze_mask_0 = const()[name = tensor("op_8416_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_8416_cast_fp16 = slice_by_index(begin = var_8416_begin_0, end = var_8416_end_0, end_mask = var_8416_end_mask_0, squeeze_mask = var_8416_squeeze_mask_0, x = q_complex_43_cast_fp16)[name = tensor("op_8416_cast_fp16")]; + tensor var_8424_begin_0 = const()[name = tensor("op_8424_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_8424_end_0 = const()[name = tensor("op_8424_end_0"), val = tensor([1, 1, 16, 32, 1])]; + tensor var_8424_end_mask_0 = const()[name = tensor("op_8424_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_8424_squeeze_mask_0 = const()[name = tensor("op_8424_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_8424_cast_fp16 = slice_by_index(begin = var_8424_begin_0, end = var_8424_end_0, end_mask = var_8424_end_mask_0, squeeze_mask = var_8424_squeeze_mask_0, x = k_complex_43_cast_fp16)[name = tensor("op_8424_cast_fp16")]; + tensor var_8432_begin_0 = const()[name = tensor("op_8432_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_8432_end_0 = const()[name = tensor("op_8432_end_0"), val = tensor([1, 1, 16, 32, 2])]; + tensor var_8432_end_mask_0 = const()[name = tensor("op_8432_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_8432_squeeze_mask_0 = const()[name = tensor("op_8432_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_8432_cast_fp16 = slice_by_index(begin = var_8432_begin_0, end = var_8432_end_0, end_mask = var_8432_end_mask_0, squeeze_mask = var_8432_squeeze_mask_0, x = k_complex_43_cast_fp16)[name = tensor("op_8432_cast_fp16")]; + tensor freqs_43_to_fp16 = const()[name = tensor("freqs_43_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(536095488)))]; + tensor var_8438_cast_fp16 = mul(x = freqs_43_to_fp16, y = ts_131_cast_fp16)[name = tensor("op_8438_cast_fp16")]; + tensor rotr_43_cast_fp16 = cos(x = var_8438_cast_fp16)[name = tensor("rotr_43_cast_fp16")]; + tensor roti_43_cast_fp16 = sin(x = var_8438_cast_fp16)[name = tensor("roti_43_cast_fp16")]; + tensor var_8442_cast_fp16 = mul(x = var_8408_cast_fp16, y = rotr_43_cast_fp16)[name = tensor("op_8442_cast_fp16")]; + tensor var_8443_cast_fp16 = mul(x = var_8416_cast_fp16, y = roti_43_cast_fp16)[name = tensor("op_8443_cast_fp16")]; + tensor qor_85_cast_fp16 = sub(x = var_8442_cast_fp16, y = var_8443_cast_fp16)[name = tensor("qor_85_cast_fp16")]; + tensor var_8446_cast_fp16 = mul(x = var_8408_cast_fp16, y = roti_43_cast_fp16)[name = tensor("op_8446_cast_fp16")]; + tensor var_8447_cast_fp16 = mul(x = var_8416_cast_fp16, y = rotr_43_cast_fp16)[name = tensor("op_8447_cast_fp16")]; + tensor qoi_85_cast_fp16 = add(x = var_8446_cast_fp16, y = var_8447_cast_fp16)[name = tensor("qoi_85_cast_fp16")]; + tensor var_8450_cast_fp16 = mul(x = var_8424_cast_fp16, y = rotr_43_cast_fp16)[name = tensor("op_8450_cast_fp16")]; + tensor var_8451_cast_fp16 = mul(x = var_8432_cast_fp16, y = roti_43_cast_fp16)[name = tensor("op_8451_cast_fp16")]; + tensor kor_85_cast_fp16 = sub(x = var_8450_cast_fp16, y = var_8451_cast_fp16)[name = tensor("kor_85_cast_fp16")]; + tensor var_8454_cast_fp16 = mul(x = var_8424_cast_fp16, y = roti_43_cast_fp16)[name = tensor("op_8454_cast_fp16")]; + tensor var_8455_cast_fp16 = mul(x = var_8432_cast_fp16, y = rotr_43_cast_fp16)[name = tensor("op_8455_cast_fp16")]; + tensor koi_85_cast_fp16 = add(x = var_8454_cast_fp16, y = var_8455_cast_fp16)[name = tensor("koi_85_cast_fp16")]; + tensor qo_43_axis_0 = const()[name = tensor("qo_43_axis_0"), val = tensor(-1)]; + tensor qo_43_cast_fp16 = stack(axis = qo_43_axis_0, values = (qor_85_cast_fp16, qoi_85_cast_fp16))[name = tensor("qo_43_cast_fp16")]; + tensor ko_43_axis_0 = const()[name = tensor("ko_43_axis_0"), val = tensor(-1)]; + tensor ko_43_cast_fp16 = stack(axis = ko_43_axis_0, values = (kor_85_cast_fp16, koi_85_cast_fp16))[name = tensor("ko_43_cast_fp16")]; + tensor var_8488 = const()[name = tensor("op_8488"), val = tensor([1, 1, 16, 64])]; + tensor q_129_cast_fp16 = reshape(shape = var_8488, x = qo_43_cast_fp16)[name = tensor("q_129_cast_fp16")]; + tensor var_8494 = const()[name = tensor("op_8494"), val = tensor([1, 1, 16, 64])]; + tensor k_87_cast_fp16 = reshape(shape = var_8494, x = ko_43_cast_fp16)[name = tensor("k_87_cast_fp16")]; + tensor _inversed_8519_y_0_to_fp16 = const()[name = tensor("_inversed_8519_y_0_to_fp16"), val = tensor(0x1p-9)]; + tensor _inversed_8519_cast_fp16 = mul(x = ts_131_cast_fp16, y = _inversed_8519_y_0_to_fp16)[name = tensor("_inversed_8519_cast_fp16")]; + tensor var_8520_cast_fp16 = floor(x = _inversed_8519_cast_fp16)[name = tensor("op_8520_cast_fp16")]; + tensor var_8521_to_fp16 = const()[name = tensor("op_8521_to_fp16"), val = tensor(0x1p+9)]; + tensor var_8522_cast_fp16 = mul(x = var_8520_cast_fp16, y = var_8521_to_fp16)[name = tensor("op_8522_cast_fp16")]; + tensor write_indices_float_87_cast_fp16 = sub(x = ts_131_cast_fp16, y = var_8522_cast_fp16)[name = tensor("write_indices_float_87_cast_fp16")]; + tensor var_8529_dtype_0 = const()[name = tensor("op_8529_dtype_0"), val = tensor("int32")]; + tensor write_indices_43_reps_0 = const()[name = tensor("write_indices_43_reps_0"), val = tensor([1, 1, 16, 64])]; + tensor write_indices_float_87_cast_fp16_to_int32 = cast(dtype = var_8529_dtype_0, x = write_indices_float_87_cast_fp16)[name = tensor("cast_447")]; + tensor write_indices_43 = tile(reps = write_indices_43_reps_0, x = write_indices_float_87_cast_fp16_to_int32)[name = tensor("write_indices_43")]; + tensor var_8539_begin_0 = const()[name = tensor("op_8539_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_8539_end_0 = const()[name = tensor("op_8539_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor var_8539_end_mask_0 = const()[name = tensor("op_8539_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_8539_squeeze_mask_0 = const()[name = tensor("op_8539_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor cache21_to_fp16_dtype_0 = const()[name = tensor("cache21_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor cache21_to_fp16 = cast(dtype = cache21_to_fp16_dtype_0, x = cache21)[name = tensor("cast_446")]; + tensor var_8539_cast_fp16 = slice_by_index(begin = var_8539_begin_0, end = var_8539_end_0, end_mask = var_8539_end_mask_0, squeeze_mask = var_8539_squeeze_mask_0, x = cache21_to_fp16)[name = tensor("op_8539_cast_fp16")]; + tensor var_8541_axis_0 = const()[name = tensor("op_8541_axis_0"), val = tensor(1)]; + tensor var_8541_mode_0 = const()[name = tensor("op_8541_mode_0"), val = tensor("update")]; + tensor var_8541_validate_indices_0 = const()[name = tensor("op_8541_validate_indices_0"), val = tensor(false)]; + tensor var_8541_cast_fp16 = scatter_along_axis(axis = var_8541_axis_0, data = var_8539_cast_fp16, indices = write_indices_43, mode = var_8541_mode_0, updates = k_87_cast_fp16, validate_indices = var_8541_validate_indices_0)[name = tensor("op_8541_cast_fp16")]; + tensor concat_149 = const()[name = tensor("concat_149"), val = tensor([0, 0, 0, 0, 0])]; + tensor concat_150 = const()[name = tensor("concat_150"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_43_internal_tensor_assign_1_stride_0 = const()[name = tensor("new_cache_43_internal_tensor_assign_1_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_43_internal_tensor_assign_1_begin_mask_0 = const()[name = tensor("new_cache_43_internal_tensor_assign_1_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_43_internal_tensor_assign_1_end_mask_0 = const()[name = tensor("new_cache_43_internal_tensor_assign_1_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_43_internal_tensor_assign_1_squeeze_mask_0 = const()[name = tensor("new_cache_43_internal_tensor_assign_1_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_90 = const()[name = tensor("shape_90"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_42 = const()[name = tensor("reduce_prod_42"), val = tensor(1048576)]; + tensor range_1d_42_start_0 = const()[name = tensor("range_1d_42_start_0"), val = tensor(0)]; + tensor range_1d_42_step_0 = const()[name = tensor("range_1d_42_step_0"), val = tensor(1)]; + tensor range_1d_42 = range_1d(end = reduce_prod_42, start = range_1d_42_start_0, step = range_1d_42_step_0)[name = tensor("range_1d_42")]; + tensor reshape_210 = reshape(shape = shape_90, x = range_1d_42)[name = tensor("reshape_210")]; + tensor slice_by_index_42 = slice_by_index(begin = concat_149, begin_mask = new_cache_43_internal_tensor_assign_1_begin_mask_0, end = concat_150, end_mask = new_cache_43_internal_tensor_assign_1_end_mask_0, squeeze_mask = new_cache_43_internal_tensor_assign_1_squeeze_mask_0, stride = new_cache_43_internal_tensor_assign_1_stride_0, x = reshape_210)[name = tensor("slice_by_index_42")]; + tensor reshape_211_shape_0 = const()[name = tensor("reshape_211_shape_0"), val = tensor([-1])]; + tensor reshape_211 = reshape(shape = reshape_211_shape_0, x = slice_by_index_42)[name = tensor("reshape_211")]; + tensor reshape_212_shape_0 = const()[name = tensor("reshape_212_shape_0"), val = tensor([-1])]; + tensor reshape_212_cast_fp16 = reshape(shape = reshape_212_shape_0, x = var_8541_cast_fp16)[name = tensor("reshape_212_cast_fp16")]; + tensor reshape_213_shape_0 = const()[name = tensor("reshape_213_shape_0"), val = tensor([-1])]; + tensor reshape_213_cast_fp16 = reshape(shape = reshape_213_shape_0, x = cache21_to_fp16)[name = tensor("reshape_213_cast_fp16")]; + tensor scatter_42_mode_0 = const()[name = tensor("scatter_42_mode_0"), val = tensor("update")]; + tensor scatter_42_axis_0 = const()[name = tensor("scatter_42_axis_0"), val = tensor(0)]; + tensor scatter_42_validate_indices_0 = const()[name = tensor("scatter_42_validate_indices_0"), val = tensor(false)]; + tensor scatter_42_cast_fp16 = scatter(axis = scatter_42_axis_0, data = reshape_213_cast_fp16, indices = reshape_211, mode = scatter_42_mode_0, updates = reshape_212_cast_fp16, validate_indices = scatter_42_validate_indices_0)[name = tensor("scatter_42_cast_fp16")]; + tensor reshape_214_cast_fp16 = reshape(shape = shape_90, x = scatter_42_cast_fp16)[name = tensor("reshape_214_cast_fp16")]; + tensor var_8549_begin_0 = const()[name = tensor("op_8549_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor var_8549_end_0 = const()[name = tensor("op_8549_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor var_8549_end_mask_0 = const()[name = tensor("op_8549_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_8549_squeeze_mask_0 = const()[name = tensor("op_8549_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor var_8549_cast_fp16 = slice_by_index(begin = var_8549_begin_0, end = var_8549_end_0, end_mask = var_8549_end_mask_0, squeeze_mask = var_8549_squeeze_mask_0, x = reshape_214_cast_fp16)[name = tensor("op_8549_cast_fp16")]; + tensor var_8551_axis_0 = const()[name = tensor("op_8551_axis_0"), val = tensor(1)]; + tensor var_8551_mode_0 = const()[name = tensor("op_8551_mode_0"), val = tensor("update")]; + tensor var_8551_validate_indices_0 = const()[name = tensor("op_8551_validate_indices_0"), val = tensor(false)]; + tensor var_8551_cast_fp16 = scatter_along_axis(axis = var_8551_axis_0, data = var_8549_cast_fp16, indices = write_indices_43, mode = var_8551_mode_0, updates = v_43_cast_fp16, validate_indices = var_8551_validate_indices_0)[name = tensor("op_8551_cast_fp16")]; + tensor concat_151 = const()[name = tensor("concat_151"), val = tensor([1, 0, 0, 0, 0])]; + tensor concat_152 = const()[name = tensor("concat_152"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_43_internal_tensor_assign_2_stride_0 = const()[name = tensor("new_cache_43_internal_tensor_assign_2_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_43_internal_tensor_assign_2_begin_mask_0 = const()[name = tensor("new_cache_43_internal_tensor_assign_2_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_43_internal_tensor_assign_2_end_mask_0 = const()[name = tensor("new_cache_43_internal_tensor_assign_2_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_43_internal_tensor_assign_2_squeeze_mask_0 = const()[name = tensor("new_cache_43_internal_tensor_assign_2_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_91 = const()[name = tensor("shape_91"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_43 = const()[name = tensor("reduce_prod_43"), val = tensor(1048576)]; + tensor range_1d_43_start_0 = const()[name = tensor("range_1d_43_start_0"), val = tensor(0)]; + tensor range_1d_43_step_0 = const()[name = tensor("range_1d_43_step_0"), val = tensor(1)]; + tensor range_1d_43 = range_1d(end = reduce_prod_43, start = range_1d_43_start_0, step = range_1d_43_step_0)[name = tensor("range_1d_43")]; + tensor reshape_215 = reshape(shape = shape_91, x = range_1d_43)[name = tensor("reshape_215")]; + tensor slice_by_index_43 = slice_by_index(begin = concat_151, begin_mask = new_cache_43_internal_tensor_assign_2_begin_mask_0, end = concat_152, end_mask = new_cache_43_internal_tensor_assign_2_end_mask_0, squeeze_mask = new_cache_43_internal_tensor_assign_2_squeeze_mask_0, stride = new_cache_43_internal_tensor_assign_2_stride_0, x = reshape_215)[name = tensor("slice_by_index_43")]; + tensor reshape_216_shape_0 = const()[name = tensor("reshape_216_shape_0"), val = tensor([-1])]; + tensor reshape_216 = reshape(shape = reshape_216_shape_0, x = slice_by_index_43)[name = tensor("reshape_216")]; + tensor reshape_217_shape_0 = const()[name = tensor("reshape_217_shape_0"), val = tensor([-1])]; + tensor reshape_217_cast_fp16 = reshape(shape = reshape_217_shape_0, x = var_8551_cast_fp16)[name = tensor("reshape_217_cast_fp16")]; + tensor reshape_218_shape_0 = const()[name = tensor("reshape_218_shape_0"), val = tensor([-1])]; + tensor reshape_218_cast_fp16 = reshape(shape = reshape_218_shape_0, x = reshape_214_cast_fp16)[name = tensor("reshape_218_cast_fp16")]; + tensor scatter_43_mode_0 = const()[name = tensor("scatter_43_mode_0"), val = tensor("update")]; + tensor scatter_43_axis_0 = const()[name = tensor("scatter_43_axis_0"), val = tensor(0)]; + tensor scatter_43_validate_indices_0 = const()[name = tensor("scatter_43_validate_indices_0"), val = tensor(false)]; + tensor scatter_43_cast_fp16 = scatter(axis = scatter_43_axis_0, data = reshape_218_cast_fp16, indices = reshape_216, mode = scatter_43_mode_0, updates = reshape_217_cast_fp16, validate_indices = scatter_43_validate_indices_0)[name = tensor("scatter_43_cast_fp16")]; + tensor reshape_219_cast_fp16 = reshape(shape = shape_91, x = scatter_43_cast_fp16)[name = tensor("reshape_219_cast_fp16")]; + tensor reshape_219_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("reshape_219_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor keys_127_begin_0 = const()[name = tensor("keys_127_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor keys_127_end_0 = const()[name = tensor("keys_127_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor keys_127_end_mask_0 = const()[name = tensor("keys_127_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor keys_127_squeeze_mask_0 = const()[name = tensor("keys_127_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor keys_127_cast_fp16 = slice_by_index(begin = keys_127_begin_0, end = keys_127_end_0, end_mask = keys_127_end_mask_0, squeeze_mask = keys_127_squeeze_mask_0, x = reshape_219_cast_fp16)[name = tensor("keys_127_cast_fp16")]; + tensor values_127_begin_0 = const()[name = tensor("values_127_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor values_127_end_0 = const()[name = tensor("values_127_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor values_127_end_mask_0 = const()[name = tensor("values_127_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor values_127_squeeze_mask_0 = const()[name = tensor("values_127_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor values_127_cast_fp16 = slice_by_index(begin = values_127_begin_0, end = values_127_end_0, end_mask = values_127_end_mask_0, squeeze_mask = values_127_squeeze_mask_0, x = reshape_219_cast_fp16)[name = tensor("values_127_cast_fp16")]; + tensor var_8563_cast_fp16 = not_equal(x = keys_127_cast_fp16, y = keys_127_cast_fp16)[name = tensor("op_8563_cast_fp16")]; + tensor keys_129_cast_fp16 = select(a = var_463_to_fp16, b = keys_127_cast_fp16, cond = var_8563_cast_fp16)[name = tensor("keys_129_cast_fp16")]; + tensor var_8571_cast_fp16 = not_equal(x = values_127_cast_fp16, y = values_127_cast_fp16)[name = tensor("op_8571_cast_fp16")]; + tensor values_129_cast_fp16 = select(a = var_463_to_fp16, b = values_127_cast_fp16, cond = var_8571_cast_fp16)[name = tensor("values_129_cast_fp16")]; + tensor var_8595 = const()[name = tensor("op_8595"), val = tensor([0, 2, 1, 3])]; + tensor var_8611 = const()[name = tensor("op_8611"), val = tensor([1, 1, 1])]; + tensor var_8612_cast_fp16 = reshape(shape = var_8611, x = position21_to_fp16)[name = tensor("op_8612_cast_fp16")]; + tensor var_8632_to_fp16 = const()[name = tensor("op_8632_to_fp16"), val = tensor(0x1p+0)]; + tensor valid_len_43_cast_fp16 = add(x = var_8612_cast_fp16, y = var_8632_to_fp16)[name = tensor("valid_len_43_cast_fp16")]; + tensor valid_mask_43_cast_fp16 = less(x = k_positions_1_promoted_to_fp16, y = valid_len_43_cast_fp16)[name = tensor("valid_mask_43_cast_fp16")]; + tensor causal_mask_43_cast_fp16 = less_equal(x = k_positions_1_promoted_to_fp16, y = var_8612_cast_fp16)[name = tensor("causal_mask_43_cast_fp16")]; + tensor attn_mask_85 = logical_and(x = valid_mask_43_cast_fp16, y = causal_mask_43_cast_fp16)[name = tensor("attn_mask_85")]; + tensor attn_mask_87_axes_0 = const()[name = tensor("attn_mask_87_axes_0"), val = tensor([1])]; + tensor attn_mask_87 = expand_dims(axes = attn_mask_87_axes_0, x = attn_mask_85)[name = tensor("attn_mask_87")]; + tensor var_8650_transpose_x_0 = const()[name = tensor("op_8650_transpose_x_0"), val = tensor(false)]; + tensor var_8650_transpose_y_0 = const()[name = tensor("op_8650_transpose_y_0"), val = tensor(false)]; + tensor transpose_114_perm_0 = const()[name = tensor("transpose_114_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_115_perm_0 = const()[name = tensor("transpose_115_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_115 = transpose(perm = transpose_115_perm_0, x = keys_129_cast_fp16)[name = tensor("transpose_129")]; + tensor transpose_114 = transpose(perm = transpose_114_perm_0, x = q_129_cast_fp16)[name = tensor("transpose_130")]; + tensor var_8650_cast_fp16 = matmul(transpose_x = var_8650_transpose_x_0, transpose_y = var_8650_transpose_y_0, x = transpose_114, y = transpose_115)[name = tensor("op_8650_cast_fp16")]; + tensor var_8644_to_fp16 = const()[name = tensor("op_8644_to_fp16"), val = tensor([0x1p-3])]; + tensor attn_weights_127_cast_fp16 = mul(x = var_8650_cast_fp16, y = var_8644_to_fp16)[name = tensor("attn_weights_127_cast_fp16")]; + tensor var_8652 = logical_not(x = attn_mask_87)[name = tensor("op_8652")]; + tensor var_8653_to_fp16 = const()[name = tensor("op_8653_to_fp16"), val = tensor(-inf)]; + tensor attn_weights_129_cast_fp16 = select(a = var_8653_to_fp16, b = attn_weights_127_cast_fp16, cond = var_8652)[name = tensor("attn_weights_129_cast_fp16")]; + tensor var_8655 = const()[name = tensor("op_8655"), val = tensor(-1)]; + tensor attn_weights_131_cast_fp16 = softmax(axis = var_8655, x = attn_weights_129_cast_fp16)[name = tensor("attn_weights_131_cast_fp16")]; + tensor attn_output_43_transpose_x_0 = const()[name = tensor("attn_output_43_transpose_x_0"), val = tensor(false)]; + tensor attn_output_43_transpose_y_0 = const()[name = tensor("attn_output_43_transpose_y_0"), val = tensor(false)]; + tensor values_131_cast_fp16 = transpose(perm = var_8595, x = values_129_cast_fp16)[name = tensor("transpose_131")]; + tensor attn_output_43_cast_fp16 = matmul(transpose_x = attn_output_43_transpose_x_0, transpose_y = attn_output_43_transpose_y_0, x = attn_weights_131_cast_fp16, y = values_131_cast_fp16)[name = tensor("attn_output_43_cast_fp16")]; + tensor var_8663 = const()[name = tensor("op_8663"), val = tensor([0, 2, 1, 3])]; + tensor var_8668 = const()[name = tensor("op_8668"), val = tensor([1, 1, 1024])]; + tensor var_8664_cast_fp16 = transpose(perm = var_8663, x = attn_output_43_cast_fp16)[name = tensor("transpose_128")]; + tensor input_259_cast_fp16 = reshape(shape = var_8668, x = var_8664_cast_fp16)[name = tensor("input_259_cast_fp16")]; + tensor attn21_out_proj_weight_to_fp16 = const()[name = tensor("attn21_out_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(536095616)))]; + tensor linear_86_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = attn21_out_proj_weight_to_fp16, x = input_259_cast_fp16)[name = tensor("linear_86_cast_fp16")]; + tensor var_8674_to_fp16 = const()[name = tensor("op_8674_to_fp16"), val = tensor(0x1p+0)]; + tensor var_8675_cast_fp16 = add(x = position21_to_fp16, y = var_8674_to_fp16)[name = tensor("op_8675_cast_fp16")]; + tensor var_8675_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_8675_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor input_261_cast_fp16 = add(x = input_255_cast_fp16, y = linear_86_cast_fp16)[name = tensor("input_261_cast_fp16")]; + tensor input_263_axes_0 = const()[name = tensor("input_263_axes_0"), val = tensor([-1])]; + tensor norm21_2_weight_to_fp16 = const()[name = tensor("norm21_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(538192832)))]; + tensor norm21_2_bias_to_fp16 = const()[name = tensor("norm21_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(538194944)))]; + tensor var_8679_to_fp16 = const()[name = tensor("op_8679_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_263_cast_fp16 = layer_norm(axes = input_263_axes_0, beta = norm21_2_bias_to_fp16, epsilon = var_8679_to_fp16, gamma = norm21_2_weight_to_fp16, x = input_261_cast_fp16)[name = tensor("input_263_cast_fp16")]; + tensor linear21_1_weight_to_fp16 = const()[name = tensor("linear21_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(538197056)))]; + tensor linear_87_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = linear21_1_weight_to_fp16, x = input_263_cast_fp16)[name = tensor("linear_87_cast_fp16")]; + tensor input_265_mode_0 = const()[name = tensor("input_265_mode_0"), val = tensor("EXACT")]; + tensor input_265_cast_fp16 = gelu(mode = input_265_mode_0, x = linear_87_cast_fp16)[name = tensor("input_265_cast_fp16")]; + tensor linear21_2_weight_to_fp16 = const()[name = tensor("linear21_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(546585728)))]; + tensor linear_88_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = linear21_2_weight_to_fp16, x = input_265_cast_fp16)[name = tensor("linear_88_cast_fp16")]; + tensor input_267_cast_fp16 = add(x = input_261_cast_fp16, y = linear_88_cast_fp16)[name = tensor("input_267_cast_fp16")]; + tensor input_269_axes_0 = const()[name = tensor("input_269_axes_0"), val = tensor([-1])]; + tensor norm22_1_weight_to_fp16 = const()[name = tensor("norm22_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(554974400)))]; + tensor norm22_1_bias_to_fp16 = const()[name = tensor("norm22_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(554976512)))]; + tensor var_8696_to_fp16 = const()[name = tensor("op_8696_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_269_cast_fp16 = layer_norm(axes = input_269_axes_0, beta = norm22_1_bias_to_fp16, epsilon = var_8696_to_fp16, gamma = norm22_1_weight_to_fp16, x = input_267_cast_fp16)[name = tensor("input_269_cast_fp16")]; + tensor attn22_in_proj_weight_to_fp16 = const()[name = tensor("attn22_in_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(554978624)))]; + tensor linear_89_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = attn22_in_proj_weight_to_fp16, x = input_269_cast_fp16)[name = tensor("linear_89_cast_fp16")]; + tensor var_8710 = const()[name = tensor("op_8710"), val = tensor([1, 1, 3, 16, 64])]; + tensor qkv_45_cast_fp16 = reshape(shape = var_8710, x = linear_89_cast_fp16)[name = tensor("qkv_45_cast_fp16")]; + tensor q_133_begin_0 = const()[name = tensor("q_133_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor q_133_end_0 = const()[name = tensor("q_133_end_0"), val = tensor([1, 1, 1, 16, 64])]; + tensor q_133_end_mask_0 = const()[name = tensor("q_133_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor q_133_squeeze_mask_0 = const()[name = tensor("q_133_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor q_133_cast_fp16 = slice_by_index(begin = q_133_begin_0, end = q_133_end_0, end_mask = q_133_end_mask_0, squeeze_mask = q_133_squeeze_mask_0, x = qkv_45_cast_fp16)[name = tensor("q_133_cast_fp16")]; + tensor k_89_begin_0 = const()[name = tensor("k_89_begin_0"), val = tensor([0, 0, 1, 0, 0])]; + tensor k_89_end_0 = const()[name = tensor("k_89_end_0"), val = tensor([1, 1, 2, 16, 64])]; + tensor k_89_end_mask_0 = const()[name = tensor("k_89_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor k_89_squeeze_mask_0 = const()[name = tensor("k_89_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor k_89_cast_fp16 = slice_by_index(begin = k_89_begin_0, end = k_89_end_0, end_mask = k_89_end_mask_0, squeeze_mask = k_89_squeeze_mask_0, x = qkv_45_cast_fp16)[name = tensor("k_89_cast_fp16")]; + tensor v_45_begin_0 = const()[name = tensor("v_45_begin_0"), val = tensor([0, 0, 2, 0, 0])]; + tensor v_45_end_0 = const()[name = tensor("v_45_end_0"), val = tensor([1, 1, 3, 16, 64])]; + tensor v_45_end_mask_0 = const()[name = tensor("v_45_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor v_45_squeeze_mask_0 = const()[name = tensor("v_45_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor v_45_cast_fp16 = slice_by_index(begin = v_45_begin_0, end = v_45_end_0, end_mask = v_45_end_mask_0, squeeze_mask = v_45_squeeze_mask_0, x = qkv_45_cast_fp16)[name = tensor("v_45_cast_fp16")]; + tensor var_8776 = const()[name = tensor("op_8776"), val = tensor([1, 1, 1, 1])]; + tensor position22_to_fp16_dtype_0 = const()[name = tensor("position22_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor position22_to_fp16 = cast(dtype = position22_to_fp16_dtype_0, x = position22)[name = tensor("cast_443")]; + tensor ts_137_cast_fp16 = reshape(shape = var_8776, x = position22_to_fp16)[name = tensor("ts_137_cast_fp16")]; + tensor var_8783 = const()[name = tensor("op_8783"), val = tensor([1, 1, 16, 32, 2])]; + tensor q_complex_45_cast_fp16 = reshape(shape = var_8783, x = q_133_cast_fp16)[name = tensor("q_complex_45_cast_fp16")]; + tensor var_8790 = const()[name = tensor("op_8790"), val = tensor([1, 1, 16, 32, 2])]; + tensor k_complex_45_cast_fp16 = reshape(shape = var_8790, x = k_89_cast_fp16)[name = tensor("k_complex_45_cast_fp16")]; + tensor var_8794_begin_0 = const()[name = tensor("op_8794_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_8794_end_0 = const()[name = tensor("op_8794_end_0"), val = tensor([1, 1, 16, 32, 1])]; + tensor var_8794_end_mask_0 = const()[name = tensor("op_8794_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_8794_squeeze_mask_0 = const()[name = tensor("op_8794_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_8794_cast_fp16 = slice_by_index(begin = var_8794_begin_0, end = var_8794_end_0, end_mask = var_8794_end_mask_0, squeeze_mask = var_8794_squeeze_mask_0, x = q_complex_45_cast_fp16)[name = tensor("op_8794_cast_fp16")]; + tensor var_8802_begin_0 = const()[name = tensor("op_8802_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_8802_end_0 = const()[name = tensor("op_8802_end_0"), val = tensor([1, 1, 16, 32, 2])]; + tensor var_8802_end_mask_0 = const()[name = tensor("op_8802_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_8802_squeeze_mask_0 = const()[name = tensor("op_8802_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_8802_cast_fp16 = slice_by_index(begin = var_8802_begin_0, end = var_8802_end_0, end_mask = var_8802_end_mask_0, squeeze_mask = var_8802_squeeze_mask_0, x = q_complex_45_cast_fp16)[name = tensor("op_8802_cast_fp16")]; + tensor var_8810_begin_0 = const()[name = tensor("op_8810_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_8810_end_0 = const()[name = tensor("op_8810_end_0"), val = tensor([1, 1, 16, 32, 1])]; + tensor var_8810_end_mask_0 = const()[name = tensor("op_8810_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_8810_squeeze_mask_0 = const()[name = tensor("op_8810_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_8810_cast_fp16 = slice_by_index(begin = var_8810_begin_0, end = var_8810_end_0, end_mask = var_8810_end_mask_0, squeeze_mask = var_8810_squeeze_mask_0, x = k_complex_45_cast_fp16)[name = tensor("op_8810_cast_fp16")]; + tensor var_8818_begin_0 = const()[name = tensor("op_8818_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_8818_end_0 = const()[name = tensor("op_8818_end_0"), val = tensor([1, 1, 16, 32, 2])]; + tensor var_8818_end_mask_0 = const()[name = tensor("op_8818_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_8818_squeeze_mask_0 = const()[name = tensor("op_8818_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_8818_cast_fp16 = slice_by_index(begin = var_8818_begin_0, end = var_8818_end_0, end_mask = var_8818_end_mask_0, squeeze_mask = var_8818_squeeze_mask_0, x = k_complex_45_cast_fp16)[name = tensor("op_8818_cast_fp16")]; + tensor freqs_45_to_fp16 = const()[name = tensor("freqs_45_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(561270144)))]; + tensor var_8824_cast_fp16 = mul(x = freqs_45_to_fp16, y = ts_137_cast_fp16)[name = tensor("op_8824_cast_fp16")]; + tensor rotr_45_cast_fp16 = cos(x = var_8824_cast_fp16)[name = tensor("rotr_45_cast_fp16")]; + tensor roti_45_cast_fp16 = sin(x = var_8824_cast_fp16)[name = tensor("roti_45_cast_fp16")]; + tensor var_8828_cast_fp16 = mul(x = var_8794_cast_fp16, y = rotr_45_cast_fp16)[name = tensor("op_8828_cast_fp16")]; + tensor var_8829_cast_fp16 = mul(x = var_8802_cast_fp16, y = roti_45_cast_fp16)[name = tensor("op_8829_cast_fp16")]; + tensor qor_89_cast_fp16 = sub(x = var_8828_cast_fp16, y = var_8829_cast_fp16)[name = tensor("qor_89_cast_fp16")]; + tensor var_8832_cast_fp16 = mul(x = var_8794_cast_fp16, y = roti_45_cast_fp16)[name = tensor("op_8832_cast_fp16")]; + tensor var_8833_cast_fp16 = mul(x = var_8802_cast_fp16, y = rotr_45_cast_fp16)[name = tensor("op_8833_cast_fp16")]; + tensor qoi_89_cast_fp16 = add(x = var_8832_cast_fp16, y = var_8833_cast_fp16)[name = tensor("qoi_89_cast_fp16")]; + tensor var_8836_cast_fp16 = mul(x = var_8810_cast_fp16, y = rotr_45_cast_fp16)[name = tensor("op_8836_cast_fp16")]; + tensor var_8837_cast_fp16 = mul(x = var_8818_cast_fp16, y = roti_45_cast_fp16)[name = tensor("op_8837_cast_fp16")]; + tensor kor_89_cast_fp16 = sub(x = var_8836_cast_fp16, y = var_8837_cast_fp16)[name = tensor("kor_89_cast_fp16")]; + tensor var_8840_cast_fp16 = mul(x = var_8810_cast_fp16, y = roti_45_cast_fp16)[name = tensor("op_8840_cast_fp16")]; + tensor var_8841_cast_fp16 = mul(x = var_8818_cast_fp16, y = rotr_45_cast_fp16)[name = tensor("op_8841_cast_fp16")]; + tensor koi_89_cast_fp16 = add(x = var_8840_cast_fp16, y = var_8841_cast_fp16)[name = tensor("koi_89_cast_fp16")]; + tensor qo_45_axis_0 = const()[name = tensor("qo_45_axis_0"), val = tensor(-1)]; + tensor qo_45_cast_fp16 = stack(axis = qo_45_axis_0, values = (qor_89_cast_fp16, qoi_89_cast_fp16))[name = tensor("qo_45_cast_fp16")]; + tensor ko_45_axis_0 = const()[name = tensor("ko_45_axis_0"), val = tensor(-1)]; + tensor ko_45_cast_fp16 = stack(axis = ko_45_axis_0, values = (kor_89_cast_fp16, koi_89_cast_fp16))[name = tensor("ko_45_cast_fp16")]; + tensor var_8874 = const()[name = tensor("op_8874"), val = tensor([1, 1, 16, 64])]; + tensor q_135_cast_fp16 = reshape(shape = var_8874, x = qo_45_cast_fp16)[name = tensor("q_135_cast_fp16")]; + tensor var_8880 = const()[name = tensor("op_8880"), val = tensor([1, 1, 16, 64])]; + tensor k_91_cast_fp16 = reshape(shape = var_8880, x = ko_45_cast_fp16)[name = tensor("k_91_cast_fp16")]; + tensor _inversed_8905_y_0_to_fp16 = const()[name = tensor("_inversed_8905_y_0_to_fp16"), val = tensor(0x1p-9)]; + tensor _inversed_8905_cast_fp16 = mul(x = ts_137_cast_fp16, y = _inversed_8905_y_0_to_fp16)[name = tensor("_inversed_8905_cast_fp16")]; + tensor var_8906_cast_fp16 = floor(x = _inversed_8905_cast_fp16)[name = tensor("op_8906_cast_fp16")]; + tensor var_8907_to_fp16 = const()[name = tensor("op_8907_to_fp16"), val = tensor(0x1p+9)]; + tensor var_8908_cast_fp16 = mul(x = var_8906_cast_fp16, y = var_8907_to_fp16)[name = tensor("op_8908_cast_fp16")]; + tensor write_indices_float_91_cast_fp16 = sub(x = ts_137_cast_fp16, y = var_8908_cast_fp16)[name = tensor("write_indices_float_91_cast_fp16")]; + tensor var_8915_dtype_0 = const()[name = tensor("op_8915_dtype_0"), val = tensor("int32")]; + tensor write_indices_45_reps_0 = const()[name = tensor("write_indices_45_reps_0"), val = tensor([1, 1, 16, 64])]; + tensor write_indices_float_91_cast_fp16_to_int32 = cast(dtype = var_8915_dtype_0, x = write_indices_float_91_cast_fp16)[name = tensor("cast_442")]; + tensor write_indices_45 = tile(reps = write_indices_45_reps_0, x = write_indices_float_91_cast_fp16_to_int32)[name = tensor("write_indices_45")]; + tensor var_8925_begin_0 = const()[name = tensor("op_8925_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_8925_end_0 = const()[name = tensor("op_8925_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor var_8925_end_mask_0 = const()[name = tensor("op_8925_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_8925_squeeze_mask_0 = const()[name = tensor("op_8925_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor cache22_to_fp16_dtype_0 = const()[name = tensor("cache22_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor cache22_to_fp16 = cast(dtype = cache22_to_fp16_dtype_0, x = cache22)[name = tensor("cast_441")]; + tensor var_8925_cast_fp16 = slice_by_index(begin = var_8925_begin_0, end = var_8925_end_0, end_mask = var_8925_end_mask_0, squeeze_mask = var_8925_squeeze_mask_0, x = cache22_to_fp16)[name = tensor("op_8925_cast_fp16")]; + tensor var_8927_axis_0 = const()[name = tensor("op_8927_axis_0"), val = tensor(1)]; + tensor var_8927_mode_0 = const()[name = tensor("op_8927_mode_0"), val = tensor("update")]; + tensor var_8927_validate_indices_0 = const()[name = tensor("op_8927_validate_indices_0"), val = tensor(false)]; + tensor var_8927_cast_fp16 = scatter_along_axis(axis = var_8927_axis_0, data = var_8925_cast_fp16, indices = write_indices_45, mode = var_8927_mode_0, updates = k_91_cast_fp16, validate_indices = var_8927_validate_indices_0)[name = tensor("op_8927_cast_fp16")]; + tensor concat_156 = const()[name = tensor("concat_156"), val = tensor([0, 0, 0, 0, 0])]; + tensor concat_157 = const()[name = tensor("concat_157"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_45_internal_tensor_assign_1_stride_0 = const()[name = tensor("new_cache_45_internal_tensor_assign_1_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_45_internal_tensor_assign_1_begin_mask_0 = const()[name = tensor("new_cache_45_internal_tensor_assign_1_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_45_internal_tensor_assign_1_end_mask_0 = const()[name = tensor("new_cache_45_internal_tensor_assign_1_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_45_internal_tensor_assign_1_squeeze_mask_0 = const()[name = tensor("new_cache_45_internal_tensor_assign_1_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_92 = const()[name = tensor("shape_92"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_44 = const()[name = tensor("reduce_prod_44"), val = tensor(1048576)]; + tensor range_1d_44_start_0 = const()[name = tensor("range_1d_44_start_0"), val = tensor(0)]; + tensor range_1d_44_step_0 = const()[name = tensor("range_1d_44_step_0"), val = tensor(1)]; + tensor range_1d_44 = range_1d(end = reduce_prod_44, start = range_1d_44_start_0, step = range_1d_44_step_0)[name = tensor("range_1d_44")]; + tensor reshape_220 = reshape(shape = shape_92, x = range_1d_44)[name = tensor("reshape_220")]; + tensor slice_by_index_44 = slice_by_index(begin = concat_156, begin_mask = new_cache_45_internal_tensor_assign_1_begin_mask_0, end = concat_157, end_mask = new_cache_45_internal_tensor_assign_1_end_mask_0, squeeze_mask = new_cache_45_internal_tensor_assign_1_squeeze_mask_0, stride = new_cache_45_internal_tensor_assign_1_stride_0, x = reshape_220)[name = tensor("slice_by_index_44")]; + tensor reshape_221_shape_0 = const()[name = tensor("reshape_221_shape_0"), val = tensor([-1])]; + tensor reshape_221 = reshape(shape = reshape_221_shape_0, x = slice_by_index_44)[name = tensor("reshape_221")]; + tensor reshape_222_shape_0 = const()[name = tensor("reshape_222_shape_0"), val = tensor([-1])]; + tensor reshape_222_cast_fp16 = reshape(shape = reshape_222_shape_0, x = var_8927_cast_fp16)[name = tensor("reshape_222_cast_fp16")]; + tensor reshape_223_shape_0 = const()[name = tensor("reshape_223_shape_0"), val = tensor([-1])]; + tensor reshape_223_cast_fp16 = reshape(shape = reshape_223_shape_0, x = cache22_to_fp16)[name = tensor("reshape_223_cast_fp16")]; + tensor scatter_44_mode_0 = const()[name = tensor("scatter_44_mode_0"), val = tensor("update")]; + tensor scatter_44_axis_0 = const()[name = tensor("scatter_44_axis_0"), val = tensor(0)]; + tensor scatter_44_validate_indices_0 = const()[name = tensor("scatter_44_validate_indices_0"), val = tensor(false)]; + tensor scatter_44_cast_fp16 = scatter(axis = scatter_44_axis_0, data = reshape_223_cast_fp16, indices = reshape_221, mode = scatter_44_mode_0, updates = reshape_222_cast_fp16, validate_indices = scatter_44_validate_indices_0)[name = tensor("scatter_44_cast_fp16")]; + tensor reshape_224_cast_fp16 = reshape(shape = shape_92, x = scatter_44_cast_fp16)[name = tensor("reshape_224_cast_fp16")]; + tensor var_8935_begin_0 = const()[name = tensor("op_8935_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor var_8935_end_0 = const()[name = tensor("op_8935_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor var_8935_end_mask_0 = const()[name = tensor("op_8935_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_8935_squeeze_mask_0 = const()[name = tensor("op_8935_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor var_8935_cast_fp16 = slice_by_index(begin = var_8935_begin_0, end = var_8935_end_0, end_mask = var_8935_end_mask_0, squeeze_mask = var_8935_squeeze_mask_0, x = reshape_224_cast_fp16)[name = tensor("op_8935_cast_fp16")]; + tensor var_8937_axis_0 = const()[name = tensor("op_8937_axis_0"), val = tensor(1)]; + tensor var_8937_mode_0 = const()[name = tensor("op_8937_mode_0"), val = tensor("update")]; + tensor var_8937_validate_indices_0 = const()[name = tensor("op_8937_validate_indices_0"), val = tensor(false)]; + tensor var_8937_cast_fp16 = scatter_along_axis(axis = var_8937_axis_0, data = var_8935_cast_fp16, indices = write_indices_45, mode = var_8937_mode_0, updates = v_45_cast_fp16, validate_indices = var_8937_validate_indices_0)[name = tensor("op_8937_cast_fp16")]; + tensor concat_158 = const()[name = tensor("concat_158"), val = tensor([1, 0, 0, 0, 0])]; + tensor concat_159 = const()[name = tensor("concat_159"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_45_internal_tensor_assign_2_stride_0 = const()[name = tensor("new_cache_45_internal_tensor_assign_2_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_45_internal_tensor_assign_2_begin_mask_0 = const()[name = tensor("new_cache_45_internal_tensor_assign_2_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_45_internal_tensor_assign_2_end_mask_0 = const()[name = tensor("new_cache_45_internal_tensor_assign_2_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_45_internal_tensor_assign_2_squeeze_mask_0 = const()[name = tensor("new_cache_45_internal_tensor_assign_2_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_93 = const()[name = tensor("shape_93"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_45 = const()[name = tensor("reduce_prod_45"), val = tensor(1048576)]; + tensor range_1d_45_start_0 = const()[name = tensor("range_1d_45_start_0"), val = tensor(0)]; + tensor range_1d_45_step_0 = const()[name = tensor("range_1d_45_step_0"), val = tensor(1)]; + tensor range_1d_45 = range_1d(end = reduce_prod_45, start = range_1d_45_start_0, step = range_1d_45_step_0)[name = tensor("range_1d_45")]; + tensor reshape_225 = reshape(shape = shape_93, x = range_1d_45)[name = tensor("reshape_225")]; + tensor slice_by_index_45 = slice_by_index(begin = concat_158, begin_mask = new_cache_45_internal_tensor_assign_2_begin_mask_0, end = concat_159, end_mask = new_cache_45_internal_tensor_assign_2_end_mask_0, squeeze_mask = new_cache_45_internal_tensor_assign_2_squeeze_mask_0, stride = new_cache_45_internal_tensor_assign_2_stride_0, x = reshape_225)[name = tensor("slice_by_index_45")]; + tensor reshape_226_shape_0 = const()[name = tensor("reshape_226_shape_0"), val = tensor([-1])]; + tensor reshape_226 = reshape(shape = reshape_226_shape_0, x = slice_by_index_45)[name = tensor("reshape_226")]; + tensor reshape_227_shape_0 = const()[name = tensor("reshape_227_shape_0"), val = tensor([-1])]; + tensor reshape_227_cast_fp16 = reshape(shape = reshape_227_shape_0, x = var_8937_cast_fp16)[name = tensor("reshape_227_cast_fp16")]; + tensor reshape_228_shape_0 = const()[name = tensor("reshape_228_shape_0"), val = tensor([-1])]; + tensor reshape_228_cast_fp16 = reshape(shape = reshape_228_shape_0, x = reshape_224_cast_fp16)[name = tensor("reshape_228_cast_fp16")]; + tensor scatter_45_mode_0 = const()[name = tensor("scatter_45_mode_0"), val = tensor("update")]; + tensor scatter_45_axis_0 = const()[name = tensor("scatter_45_axis_0"), val = tensor(0)]; + tensor scatter_45_validate_indices_0 = const()[name = tensor("scatter_45_validate_indices_0"), val = tensor(false)]; + tensor scatter_45_cast_fp16 = scatter(axis = scatter_45_axis_0, data = reshape_228_cast_fp16, indices = reshape_226, mode = scatter_45_mode_0, updates = reshape_227_cast_fp16, validate_indices = scatter_45_validate_indices_0)[name = tensor("scatter_45_cast_fp16")]; + tensor reshape_229_cast_fp16 = reshape(shape = shape_93, x = scatter_45_cast_fp16)[name = tensor("reshape_229_cast_fp16")]; + tensor reshape_229_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("reshape_229_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor keys_133_begin_0 = const()[name = tensor("keys_133_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor keys_133_end_0 = const()[name = tensor("keys_133_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor keys_133_end_mask_0 = const()[name = tensor("keys_133_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor keys_133_squeeze_mask_0 = const()[name = tensor("keys_133_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor keys_133_cast_fp16 = slice_by_index(begin = keys_133_begin_0, end = keys_133_end_0, end_mask = keys_133_end_mask_0, squeeze_mask = keys_133_squeeze_mask_0, x = reshape_229_cast_fp16)[name = tensor("keys_133_cast_fp16")]; + tensor values_133_begin_0 = const()[name = tensor("values_133_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor values_133_end_0 = const()[name = tensor("values_133_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor values_133_end_mask_0 = const()[name = tensor("values_133_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor values_133_squeeze_mask_0 = const()[name = tensor("values_133_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor values_133_cast_fp16 = slice_by_index(begin = values_133_begin_0, end = values_133_end_0, end_mask = values_133_end_mask_0, squeeze_mask = values_133_squeeze_mask_0, x = reshape_229_cast_fp16)[name = tensor("values_133_cast_fp16")]; + tensor var_8949_cast_fp16 = not_equal(x = keys_133_cast_fp16, y = keys_133_cast_fp16)[name = tensor("op_8949_cast_fp16")]; + tensor keys_135_cast_fp16 = select(a = var_463_to_fp16, b = keys_133_cast_fp16, cond = var_8949_cast_fp16)[name = tensor("keys_135_cast_fp16")]; + tensor var_8957_cast_fp16 = not_equal(x = values_133_cast_fp16, y = values_133_cast_fp16)[name = tensor("op_8957_cast_fp16")]; + tensor values_135_cast_fp16 = select(a = var_463_to_fp16, b = values_133_cast_fp16, cond = var_8957_cast_fp16)[name = tensor("values_135_cast_fp16")]; + tensor var_8981 = const()[name = tensor("op_8981"), val = tensor([0, 2, 1, 3])]; + tensor var_8997 = const()[name = tensor("op_8997"), val = tensor([1, 1, 1])]; + tensor var_8998_cast_fp16 = reshape(shape = var_8997, x = position22_to_fp16)[name = tensor("op_8998_cast_fp16")]; + tensor var_9018_to_fp16 = const()[name = tensor("op_9018_to_fp16"), val = tensor(0x1p+0)]; + tensor valid_len_45_cast_fp16 = add(x = var_8998_cast_fp16, y = var_9018_to_fp16)[name = tensor("valid_len_45_cast_fp16")]; + tensor valid_mask_45_cast_fp16 = less(x = k_positions_1_promoted_to_fp16, y = valid_len_45_cast_fp16)[name = tensor("valid_mask_45_cast_fp16")]; + tensor causal_mask_45_cast_fp16 = less_equal(x = k_positions_1_promoted_to_fp16, y = var_8998_cast_fp16)[name = tensor("causal_mask_45_cast_fp16")]; + tensor attn_mask_89 = logical_and(x = valid_mask_45_cast_fp16, y = causal_mask_45_cast_fp16)[name = tensor("attn_mask_89")]; + tensor attn_mask_91_axes_0 = const()[name = tensor("attn_mask_91_axes_0"), val = tensor([1])]; + tensor attn_mask_91 = expand_dims(axes = attn_mask_91_axes_0, x = attn_mask_89)[name = tensor("attn_mask_91")]; + tensor var_9036_transpose_x_0 = const()[name = tensor("op_9036_transpose_x_0"), val = tensor(false)]; + tensor var_9036_transpose_y_0 = const()[name = tensor("op_9036_transpose_y_0"), val = tensor(false)]; + tensor transpose_116_perm_0 = const()[name = tensor("transpose_116_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_117_perm_0 = const()[name = tensor("transpose_117_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_117 = transpose(perm = transpose_117_perm_0, x = keys_135_cast_fp16)[name = tensor("transpose_125")]; + tensor transpose_116 = transpose(perm = transpose_116_perm_0, x = q_135_cast_fp16)[name = tensor("transpose_126")]; + tensor var_9036_cast_fp16 = matmul(transpose_x = var_9036_transpose_x_0, transpose_y = var_9036_transpose_y_0, x = transpose_116, y = transpose_117)[name = tensor("op_9036_cast_fp16")]; + tensor var_9030_to_fp16 = const()[name = tensor("op_9030_to_fp16"), val = tensor([0x1p-3])]; + tensor attn_weights_133_cast_fp16 = mul(x = var_9036_cast_fp16, y = var_9030_to_fp16)[name = tensor("attn_weights_133_cast_fp16")]; + tensor var_9038 = logical_not(x = attn_mask_91)[name = tensor("op_9038")]; + tensor var_9039_to_fp16 = const()[name = tensor("op_9039_to_fp16"), val = tensor(-inf)]; + tensor attn_weights_135_cast_fp16 = select(a = var_9039_to_fp16, b = attn_weights_133_cast_fp16, cond = var_9038)[name = tensor("attn_weights_135_cast_fp16")]; + tensor var_9041 = const()[name = tensor("op_9041"), val = tensor(-1)]; + tensor attn_weights_137_cast_fp16 = softmax(axis = var_9041, x = attn_weights_135_cast_fp16)[name = tensor("attn_weights_137_cast_fp16")]; + tensor attn_output_45_transpose_x_0 = const()[name = tensor("attn_output_45_transpose_x_0"), val = tensor(false)]; + tensor attn_output_45_transpose_y_0 = const()[name = tensor("attn_output_45_transpose_y_0"), val = tensor(false)]; + tensor values_137_cast_fp16 = transpose(perm = var_8981, x = values_135_cast_fp16)[name = tensor("transpose_127")]; + tensor attn_output_45_cast_fp16 = matmul(transpose_x = attn_output_45_transpose_x_0, transpose_y = attn_output_45_transpose_y_0, x = attn_weights_137_cast_fp16, y = values_137_cast_fp16)[name = tensor("attn_output_45_cast_fp16")]; + tensor var_9049 = const()[name = tensor("op_9049"), val = tensor([0, 2, 1, 3])]; + tensor var_9054 = const()[name = tensor("op_9054"), val = tensor([1, 1, 1024])]; + tensor var_9050_cast_fp16 = transpose(perm = var_9049, x = attn_output_45_cast_fp16)[name = tensor("transpose_124")]; + tensor input_271_cast_fp16 = reshape(shape = var_9054, x = var_9050_cast_fp16)[name = tensor("input_271_cast_fp16")]; + tensor attn22_out_proj_weight_to_fp16 = const()[name = tensor("attn22_out_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(561270272)))]; + tensor linear_90_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = attn22_out_proj_weight_to_fp16, x = input_271_cast_fp16)[name = tensor("linear_90_cast_fp16")]; + tensor var_9060_to_fp16 = const()[name = tensor("op_9060_to_fp16"), val = tensor(0x1p+0)]; + tensor var_9061_cast_fp16 = add(x = position22_to_fp16, y = var_9060_to_fp16)[name = tensor("op_9061_cast_fp16")]; + tensor var_9061_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_9061_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor input_273_cast_fp16 = add(x = input_267_cast_fp16, y = linear_90_cast_fp16)[name = tensor("input_273_cast_fp16")]; + tensor input_275_axes_0 = const()[name = tensor("input_275_axes_0"), val = tensor([-1])]; + tensor norm22_2_weight_to_fp16 = const()[name = tensor("norm22_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(563367488)))]; + tensor norm22_2_bias_to_fp16 = const()[name = tensor("norm22_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(563369600)))]; + tensor var_9065_to_fp16 = const()[name = tensor("op_9065_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_275_cast_fp16 = layer_norm(axes = input_275_axes_0, beta = norm22_2_bias_to_fp16, epsilon = var_9065_to_fp16, gamma = norm22_2_weight_to_fp16, x = input_273_cast_fp16)[name = tensor("input_275_cast_fp16")]; + tensor linear22_1_weight_to_fp16 = const()[name = tensor("linear22_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(563371712)))]; + tensor linear_91_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = linear22_1_weight_to_fp16, x = input_275_cast_fp16)[name = tensor("linear_91_cast_fp16")]; + tensor input_277_mode_0 = const()[name = tensor("input_277_mode_0"), val = tensor("EXACT")]; + tensor input_277_cast_fp16 = gelu(mode = input_277_mode_0, x = linear_91_cast_fp16)[name = tensor("input_277_cast_fp16")]; + tensor linear22_2_weight_to_fp16 = const()[name = tensor("linear22_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(571760384)))]; + tensor linear_92_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = linear22_2_weight_to_fp16, x = input_277_cast_fp16)[name = tensor("linear_92_cast_fp16")]; + tensor input_279_cast_fp16 = add(x = input_273_cast_fp16, y = linear_92_cast_fp16)[name = tensor("input_279_cast_fp16")]; + tensor input_281_axes_0 = const()[name = tensor("input_281_axes_0"), val = tensor([-1])]; + tensor norm23_1_weight_to_fp16 = const()[name = tensor("norm23_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(580149056)))]; + tensor norm23_1_bias_to_fp16 = const()[name = tensor("norm23_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(580151168)))]; + tensor var_9082_to_fp16 = const()[name = tensor("op_9082_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_281_cast_fp16 = layer_norm(axes = input_281_axes_0, beta = norm23_1_bias_to_fp16, epsilon = var_9082_to_fp16, gamma = norm23_1_weight_to_fp16, x = input_279_cast_fp16)[name = tensor("input_281_cast_fp16")]; + tensor attn23_in_proj_weight_to_fp16 = const()[name = tensor("attn23_in_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(580153280)))]; + tensor linear_93_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = attn23_in_proj_weight_to_fp16, x = input_281_cast_fp16)[name = tensor("linear_93_cast_fp16")]; + tensor var_9096 = const()[name = tensor("op_9096"), val = tensor([1, 1, 3, 16, 64])]; + tensor qkv_cast_fp16 = reshape(shape = var_9096, x = linear_93_cast_fp16)[name = tensor("qkv_cast_fp16")]; + tensor q_139_begin_0 = const()[name = tensor("q_139_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor q_139_end_0 = const()[name = tensor("q_139_end_0"), val = tensor([1, 1, 1, 16, 64])]; + tensor q_139_end_mask_0 = const()[name = tensor("q_139_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor q_139_squeeze_mask_0 = const()[name = tensor("q_139_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor q_139_cast_fp16 = slice_by_index(begin = q_139_begin_0, end = q_139_end_0, end_mask = q_139_end_mask_0, squeeze_mask = q_139_squeeze_mask_0, x = qkv_cast_fp16)[name = tensor("q_139_cast_fp16")]; + tensor k_93_begin_0 = const()[name = tensor("k_93_begin_0"), val = tensor([0, 0, 1, 0, 0])]; + tensor k_93_end_0 = const()[name = tensor("k_93_end_0"), val = tensor([1, 1, 2, 16, 64])]; + tensor k_93_end_mask_0 = const()[name = tensor("k_93_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor k_93_squeeze_mask_0 = const()[name = tensor("k_93_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor k_93_cast_fp16 = slice_by_index(begin = k_93_begin_0, end = k_93_end_0, end_mask = k_93_end_mask_0, squeeze_mask = k_93_squeeze_mask_0, x = qkv_cast_fp16)[name = tensor("k_93_cast_fp16")]; + tensor v_begin_0 = const()[name = tensor("v_begin_0"), val = tensor([0, 0, 2, 0, 0])]; + tensor v_end_0 = const()[name = tensor("v_end_0"), val = tensor([1, 1, 3, 16, 64])]; + tensor v_end_mask_0 = const()[name = tensor("v_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor v_squeeze_mask_0 = const()[name = tensor("v_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor v_cast_fp16 = slice_by_index(begin = v_begin_0, end = v_end_0, end_mask = v_end_mask_0, squeeze_mask = v_squeeze_mask_0, x = qkv_cast_fp16)[name = tensor("v_cast_fp16")]; + tensor var_9162 = const()[name = tensor("op_9162"), val = tensor([1, 1, 1, 1])]; + tensor position23_to_fp16_dtype_0 = const()[name = tensor("position23_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor position23_to_fp16 = cast(dtype = position23_to_fp16_dtype_0, x = position23)[name = tensor("cast_438")]; + tensor ts_cast_fp16 = reshape(shape = var_9162, x = position23_to_fp16)[name = tensor("ts_cast_fp16")]; + tensor var_9169 = const()[name = tensor("op_9169"), val = tensor([1, 1, 16, 32, 2])]; + tensor q_complex_cast_fp16 = reshape(shape = var_9169, x = q_139_cast_fp16)[name = tensor("q_complex_cast_fp16")]; + tensor var_9176 = const()[name = tensor("op_9176"), val = tensor([1, 1, 16, 32, 2])]; + tensor k_complex_cast_fp16 = reshape(shape = var_9176, x = k_93_cast_fp16)[name = tensor("k_complex_cast_fp16")]; + tensor var_9180_begin_0 = const()[name = tensor("op_9180_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_9180_end_0 = const()[name = tensor("op_9180_end_0"), val = tensor([1, 1, 16, 32, 1])]; + tensor var_9180_end_mask_0 = const()[name = tensor("op_9180_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_9180_squeeze_mask_0 = const()[name = tensor("op_9180_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_9180_cast_fp16 = slice_by_index(begin = var_9180_begin_0, end = var_9180_end_0, end_mask = var_9180_end_mask_0, squeeze_mask = var_9180_squeeze_mask_0, x = q_complex_cast_fp16)[name = tensor("op_9180_cast_fp16")]; + tensor var_9188_begin_0 = const()[name = tensor("op_9188_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_9188_end_0 = const()[name = tensor("op_9188_end_0"), val = tensor([1, 1, 16, 32, 2])]; + tensor var_9188_end_mask_0 = const()[name = tensor("op_9188_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_9188_squeeze_mask_0 = const()[name = tensor("op_9188_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_9188_cast_fp16 = slice_by_index(begin = var_9188_begin_0, end = var_9188_end_0, end_mask = var_9188_end_mask_0, squeeze_mask = var_9188_squeeze_mask_0, x = q_complex_cast_fp16)[name = tensor("op_9188_cast_fp16")]; + tensor var_9196_begin_0 = const()[name = tensor("op_9196_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_9196_end_0 = const()[name = tensor("op_9196_end_0"), val = tensor([1, 1, 16, 32, 1])]; + tensor var_9196_end_mask_0 = const()[name = tensor("op_9196_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_9196_squeeze_mask_0 = const()[name = tensor("op_9196_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_9196_cast_fp16 = slice_by_index(begin = var_9196_begin_0, end = var_9196_end_0, end_mask = var_9196_end_mask_0, squeeze_mask = var_9196_squeeze_mask_0, x = k_complex_cast_fp16)[name = tensor("op_9196_cast_fp16")]; + tensor var_9204_begin_0 = const()[name = tensor("op_9204_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_9204_end_0 = const()[name = tensor("op_9204_end_0"), val = tensor([1, 1, 16, 32, 2])]; + tensor var_9204_end_mask_0 = const()[name = tensor("op_9204_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_9204_squeeze_mask_0 = const()[name = tensor("op_9204_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_9204_cast_fp16 = slice_by_index(begin = var_9204_begin_0, end = var_9204_end_0, end_mask = var_9204_end_mask_0, squeeze_mask = var_9204_squeeze_mask_0, x = k_complex_cast_fp16)[name = tensor("op_9204_cast_fp16")]; + tensor freqs_to_fp16 = const()[name = tensor("freqs_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(586444800)))]; + tensor var_9210_cast_fp16 = mul(x = freqs_to_fp16, y = ts_cast_fp16)[name = tensor("op_9210_cast_fp16")]; + tensor rotr_cast_fp16 = cos(x = var_9210_cast_fp16)[name = tensor("rotr_cast_fp16")]; + tensor roti_cast_fp16 = sin(x = var_9210_cast_fp16)[name = tensor("roti_cast_fp16")]; + tensor var_9214_cast_fp16 = mul(x = var_9180_cast_fp16, y = rotr_cast_fp16)[name = tensor("op_9214_cast_fp16")]; + tensor var_9215_cast_fp16 = mul(x = var_9188_cast_fp16, y = roti_cast_fp16)[name = tensor("op_9215_cast_fp16")]; + tensor qor_93_cast_fp16 = sub(x = var_9214_cast_fp16, y = var_9215_cast_fp16)[name = tensor("qor_93_cast_fp16")]; + tensor var_9218_cast_fp16 = mul(x = var_9180_cast_fp16, y = roti_cast_fp16)[name = tensor("op_9218_cast_fp16")]; + tensor var_9219_cast_fp16 = mul(x = var_9188_cast_fp16, y = rotr_cast_fp16)[name = tensor("op_9219_cast_fp16")]; + tensor qoi_93_cast_fp16 = add(x = var_9218_cast_fp16, y = var_9219_cast_fp16)[name = tensor("qoi_93_cast_fp16")]; + tensor var_9222_cast_fp16 = mul(x = var_9196_cast_fp16, y = rotr_cast_fp16)[name = tensor("op_9222_cast_fp16")]; + tensor var_9223_cast_fp16 = mul(x = var_9204_cast_fp16, y = roti_cast_fp16)[name = tensor("op_9223_cast_fp16")]; + tensor kor_93_cast_fp16 = sub(x = var_9222_cast_fp16, y = var_9223_cast_fp16)[name = tensor("kor_93_cast_fp16")]; + tensor var_9226_cast_fp16 = mul(x = var_9196_cast_fp16, y = roti_cast_fp16)[name = tensor("op_9226_cast_fp16")]; + tensor var_9227_cast_fp16 = mul(x = var_9204_cast_fp16, y = rotr_cast_fp16)[name = tensor("op_9227_cast_fp16")]; + tensor koi_93_cast_fp16 = add(x = var_9226_cast_fp16, y = var_9227_cast_fp16)[name = tensor("koi_93_cast_fp16")]; + tensor qo_axis_0 = const()[name = tensor("qo_axis_0"), val = tensor(-1)]; + tensor qo_cast_fp16 = stack(axis = qo_axis_0, values = (qor_93_cast_fp16, qoi_93_cast_fp16))[name = tensor("qo_cast_fp16")]; + tensor ko_axis_0 = const()[name = tensor("ko_axis_0"), val = tensor(-1)]; + tensor ko_cast_fp16 = stack(axis = ko_axis_0, values = (kor_93_cast_fp16, koi_93_cast_fp16))[name = tensor("ko_cast_fp16")]; + tensor var_9260 = const()[name = tensor("op_9260"), val = tensor([1, 1, 16, 64])]; + tensor q_141_cast_fp16 = reshape(shape = var_9260, x = qo_cast_fp16)[name = tensor("q_141_cast_fp16")]; + tensor var_9266 = const()[name = tensor("op_9266"), val = tensor([1, 1, 16, 64])]; + tensor k_cast_fp16 = reshape(shape = var_9266, x = ko_cast_fp16)[name = tensor("k_cast_fp16")]; + tensor _inversed_9291_y_0_to_fp16 = const()[name = tensor("_inversed_9291_y_0_to_fp16"), val = tensor(0x1p-9)]; + tensor _inversed_9291_cast_fp16 = mul(x = ts_cast_fp16, y = _inversed_9291_y_0_to_fp16)[name = tensor("_inversed_9291_cast_fp16")]; + tensor var_9292_cast_fp16 = floor(x = _inversed_9291_cast_fp16)[name = tensor("op_9292_cast_fp16")]; + tensor var_9293_to_fp16 = const()[name = tensor("op_9293_to_fp16"), val = tensor(0x1p+9)]; + tensor var_9294_cast_fp16 = mul(x = var_9292_cast_fp16, y = var_9293_to_fp16)[name = tensor("op_9294_cast_fp16")]; + tensor write_indices_float_cast_fp16 = sub(x = ts_cast_fp16, y = var_9294_cast_fp16)[name = tensor("write_indices_float_cast_fp16")]; + tensor var_9301_dtype_0 = const()[name = tensor("op_9301_dtype_0"), val = tensor("int32")]; + tensor write_indices_reps_0 = const()[name = tensor("write_indices_reps_0"), val = tensor([1, 1, 16, 64])]; + tensor write_indices_float_cast_fp16_to_int32 = cast(dtype = var_9301_dtype_0, x = write_indices_float_cast_fp16)[name = tensor("cast_437")]; + tensor write_indices = tile(reps = write_indices_reps_0, x = write_indices_float_cast_fp16_to_int32)[name = tensor("write_indices")]; + tensor var_9311_begin_0 = const()[name = tensor("op_9311_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_9311_end_0 = const()[name = tensor("op_9311_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor var_9311_end_mask_0 = const()[name = tensor("op_9311_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_9311_squeeze_mask_0 = const()[name = tensor("op_9311_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor cache23_to_fp16_dtype_0 = const()[name = tensor("cache23_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor cache23_to_fp16 = cast(dtype = cache23_to_fp16_dtype_0, x = cache23)[name = tensor("cast_436")]; + tensor var_9311_cast_fp16 = slice_by_index(begin = var_9311_begin_0, end = var_9311_end_0, end_mask = var_9311_end_mask_0, squeeze_mask = var_9311_squeeze_mask_0, x = cache23_to_fp16)[name = tensor("op_9311_cast_fp16")]; + tensor var_9313_axis_0 = const()[name = tensor("op_9313_axis_0"), val = tensor(1)]; + tensor var_9313_mode_0 = const()[name = tensor("op_9313_mode_0"), val = tensor("update")]; + tensor var_9313_validate_indices_0 = const()[name = tensor("op_9313_validate_indices_0"), val = tensor(false)]; + tensor var_9313_cast_fp16 = scatter_along_axis(axis = var_9313_axis_0, data = var_9311_cast_fp16, indices = write_indices, mode = var_9313_mode_0, updates = k_cast_fp16, validate_indices = var_9313_validate_indices_0)[name = tensor("op_9313_cast_fp16")]; + tensor concat_163 = const()[name = tensor("concat_163"), val = tensor([0, 0, 0, 0, 0])]; + tensor concat_164 = const()[name = tensor("concat_164"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_internal_tensor_assign_1_stride_0 = const()[name = tensor("new_cache_internal_tensor_assign_1_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_internal_tensor_assign_1_begin_mask_0 = const()[name = tensor("new_cache_internal_tensor_assign_1_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_internal_tensor_assign_1_end_mask_0 = const()[name = tensor("new_cache_internal_tensor_assign_1_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_internal_tensor_assign_1_squeeze_mask_0 = const()[name = tensor("new_cache_internal_tensor_assign_1_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_94 = const()[name = tensor("shape_94"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_46 = const()[name = tensor("reduce_prod_46"), val = tensor(1048576)]; + tensor range_1d_46_start_0 = const()[name = tensor("range_1d_46_start_0"), val = tensor(0)]; + tensor range_1d_46_step_0 = const()[name = tensor("range_1d_46_step_0"), val = tensor(1)]; + tensor range_1d_46 = range_1d(end = reduce_prod_46, start = range_1d_46_start_0, step = range_1d_46_step_0)[name = tensor("range_1d_46")]; + tensor reshape_230 = reshape(shape = shape_94, x = range_1d_46)[name = tensor("reshape_230")]; + tensor slice_by_index_46 = slice_by_index(begin = concat_163, begin_mask = new_cache_internal_tensor_assign_1_begin_mask_0, end = concat_164, end_mask = new_cache_internal_tensor_assign_1_end_mask_0, squeeze_mask = new_cache_internal_tensor_assign_1_squeeze_mask_0, stride = new_cache_internal_tensor_assign_1_stride_0, x = reshape_230)[name = tensor("slice_by_index_46")]; + tensor reshape_231_shape_0 = const()[name = tensor("reshape_231_shape_0"), val = tensor([-1])]; + tensor reshape_231 = reshape(shape = reshape_231_shape_0, x = slice_by_index_46)[name = tensor("reshape_231")]; + tensor reshape_232_shape_0 = const()[name = tensor("reshape_232_shape_0"), val = tensor([-1])]; + tensor reshape_232_cast_fp16 = reshape(shape = reshape_232_shape_0, x = var_9313_cast_fp16)[name = tensor("reshape_232_cast_fp16")]; + tensor reshape_233_shape_0 = const()[name = tensor("reshape_233_shape_0"), val = tensor([-1])]; + tensor reshape_233_cast_fp16 = reshape(shape = reshape_233_shape_0, x = cache23_to_fp16)[name = tensor("reshape_233_cast_fp16")]; + tensor scatter_46_mode_0 = const()[name = tensor("scatter_46_mode_0"), val = tensor("update")]; + tensor scatter_46_axis_0 = const()[name = tensor("scatter_46_axis_0"), val = tensor(0)]; + tensor scatter_46_validate_indices_0 = const()[name = tensor("scatter_46_validate_indices_0"), val = tensor(false)]; + tensor scatter_46_cast_fp16 = scatter(axis = scatter_46_axis_0, data = reshape_233_cast_fp16, indices = reshape_231, mode = scatter_46_mode_0, updates = reshape_232_cast_fp16, validate_indices = scatter_46_validate_indices_0)[name = tensor("scatter_46_cast_fp16")]; + tensor reshape_234_cast_fp16 = reshape(shape = shape_94, x = scatter_46_cast_fp16)[name = tensor("reshape_234_cast_fp16")]; + tensor var_9321_begin_0 = const()[name = tensor("op_9321_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor var_9321_end_0 = const()[name = tensor("op_9321_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor var_9321_end_mask_0 = const()[name = tensor("op_9321_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_9321_squeeze_mask_0 = const()[name = tensor("op_9321_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor var_9321_cast_fp16 = slice_by_index(begin = var_9321_begin_0, end = var_9321_end_0, end_mask = var_9321_end_mask_0, squeeze_mask = var_9321_squeeze_mask_0, x = reshape_234_cast_fp16)[name = tensor("op_9321_cast_fp16")]; + tensor var_9323_axis_0 = const()[name = tensor("op_9323_axis_0"), val = tensor(1)]; + tensor var_9323_mode_0 = const()[name = tensor("op_9323_mode_0"), val = tensor("update")]; + tensor var_9323_validate_indices_0 = const()[name = tensor("op_9323_validate_indices_0"), val = tensor(false)]; + tensor var_9323_cast_fp16 = scatter_along_axis(axis = var_9323_axis_0, data = var_9321_cast_fp16, indices = write_indices, mode = var_9323_mode_0, updates = v_cast_fp16, validate_indices = var_9323_validate_indices_0)[name = tensor("op_9323_cast_fp16")]; + tensor concat_165 = const()[name = tensor("concat_165"), val = tensor([1, 0, 0, 0, 0])]; + tensor concat_166 = const()[name = tensor("concat_166"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_internal_tensor_assign_2_stride_0 = const()[name = tensor("new_cache_internal_tensor_assign_2_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_internal_tensor_assign_2_begin_mask_0 = const()[name = tensor("new_cache_internal_tensor_assign_2_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_internal_tensor_assign_2_end_mask_0 = const()[name = tensor("new_cache_internal_tensor_assign_2_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_internal_tensor_assign_2_squeeze_mask_0 = const()[name = tensor("new_cache_internal_tensor_assign_2_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_95 = const()[name = tensor("shape_95"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_47 = const()[name = tensor("reduce_prod_47"), val = tensor(1048576)]; + tensor range_1d_47_start_0 = const()[name = tensor("range_1d_47_start_0"), val = tensor(0)]; + tensor range_1d_47_step_0 = const()[name = tensor("range_1d_47_step_0"), val = tensor(1)]; + tensor range_1d_47 = range_1d(end = reduce_prod_47, start = range_1d_47_start_0, step = range_1d_47_step_0)[name = tensor("range_1d_47")]; + tensor reshape_235 = reshape(shape = shape_95, x = range_1d_47)[name = tensor("reshape_235")]; + tensor slice_by_index_47 = slice_by_index(begin = concat_165, begin_mask = new_cache_internal_tensor_assign_2_begin_mask_0, end = concat_166, end_mask = new_cache_internal_tensor_assign_2_end_mask_0, squeeze_mask = new_cache_internal_tensor_assign_2_squeeze_mask_0, stride = new_cache_internal_tensor_assign_2_stride_0, x = reshape_235)[name = tensor("slice_by_index_47")]; + tensor reshape_236_shape_0 = const()[name = tensor("reshape_236_shape_0"), val = tensor([-1])]; + tensor reshape_236 = reshape(shape = reshape_236_shape_0, x = slice_by_index_47)[name = tensor("reshape_236")]; + tensor reshape_237_shape_0 = const()[name = tensor("reshape_237_shape_0"), val = tensor([-1])]; + tensor reshape_237_cast_fp16 = reshape(shape = reshape_237_shape_0, x = var_9323_cast_fp16)[name = tensor("reshape_237_cast_fp16")]; + tensor reshape_238_shape_0 = const()[name = tensor("reshape_238_shape_0"), val = tensor([-1])]; + tensor reshape_238_cast_fp16 = reshape(shape = reshape_238_shape_0, x = reshape_234_cast_fp16)[name = tensor("reshape_238_cast_fp16")]; + tensor scatter_47_mode_0 = const()[name = tensor("scatter_47_mode_0"), val = tensor("update")]; + tensor scatter_47_axis_0 = const()[name = tensor("scatter_47_axis_0"), val = tensor(0)]; + tensor scatter_47_validate_indices_0 = const()[name = tensor("scatter_47_validate_indices_0"), val = tensor(false)]; + tensor scatter_47_cast_fp16 = scatter(axis = scatter_47_axis_0, data = reshape_238_cast_fp16, indices = reshape_236, mode = scatter_47_mode_0, updates = reshape_237_cast_fp16, validate_indices = scatter_47_validate_indices_0)[name = tensor("scatter_47_cast_fp16")]; + tensor reshape_239_cast_fp16 = reshape(shape = shape_95, x = scatter_47_cast_fp16)[name = tensor("reshape_239_cast_fp16")]; + tensor reshape_239_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("reshape_239_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor keys_139_begin_0 = const()[name = tensor("keys_139_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor keys_139_end_0 = const()[name = tensor("keys_139_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor keys_139_end_mask_0 = const()[name = tensor("keys_139_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor keys_139_squeeze_mask_0 = const()[name = tensor("keys_139_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor keys_139_cast_fp16 = slice_by_index(begin = keys_139_begin_0, end = keys_139_end_0, end_mask = keys_139_end_mask_0, squeeze_mask = keys_139_squeeze_mask_0, x = reshape_239_cast_fp16)[name = tensor("keys_139_cast_fp16")]; + tensor values_139_begin_0 = const()[name = tensor("values_139_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor values_139_end_0 = const()[name = tensor("values_139_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor values_139_end_mask_0 = const()[name = tensor("values_139_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor values_139_squeeze_mask_0 = const()[name = tensor("values_139_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor values_139_cast_fp16 = slice_by_index(begin = values_139_begin_0, end = values_139_end_0, end_mask = values_139_end_mask_0, squeeze_mask = values_139_squeeze_mask_0, x = reshape_239_cast_fp16)[name = tensor("values_139_cast_fp16")]; + tensor var_9335_cast_fp16 = not_equal(x = keys_139_cast_fp16, y = keys_139_cast_fp16)[name = tensor("op_9335_cast_fp16")]; + tensor keys_141_cast_fp16 = select(a = var_463_to_fp16, b = keys_139_cast_fp16, cond = var_9335_cast_fp16)[name = tensor("keys_141_cast_fp16")]; + tensor var_9343_cast_fp16 = not_equal(x = values_139_cast_fp16, y = values_139_cast_fp16)[name = tensor("op_9343_cast_fp16")]; + tensor values_141_cast_fp16 = select(a = var_463_to_fp16, b = values_139_cast_fp16, cond = var_9343_cast_fp16)[name = tensor("values_141_cast_fp16")]; + tensor var_9367 = const()[name = tensor("op_9367"), val = tensor([0, 2, 1, 3])]; + tensor var_9383 = const()[name = tensor("op_9383"), val = tensor([1, 1, 1])]; + tensor var_9384_cast_fp16 = reshape(shape = var_9383, x = position23_to_fp16)[name = tensor("op_9384_cast_fp16")]; + tensor var_9404_to_fp16 = const()[name = tensor("op_9404_to_fp16"), val = tensor(0x1p+0)]; + tensor valid_len_cast_fp16 = add(x = var_9384_cast_fp16, y = var_9404_to_fp16)[name = tensor("valid_len_cast_fp16")]; + tensor valid_mask_cast_fp16 = less(x = k_positions_1_promoted_to_fp16, y = valid_len_cast_fp16)[name = tensor("valid_mask_cast_fp16")]; + tensor causal_mask_cast_fp16 = less_equal(x = k_positions_1_promoted_to_fp16, y = var_9384_cast_fp16)[name = tensor("causal_mask_cast_fp16")]; + tensor attn_mask_93 = logical_and(x = valid_mask_cast_fp16, y = causal_mask_cast_fp16)[name = tensor("attn_mask_93")]; + tensor attn_mask_axes_0 = const()[name = tensor("attn_mask_axes_0"), val = tensor([1])]; + tensor attn_mask = expand_dims(axes = attn_mask_axes_0, x = attn_mask_93)[name = tensor("attn_mask")]; + tensor var_9422_transpose_x_0 = const()[name = tensor("op_9422_transpose_x_0"), val = tensor(false)]; + tensor var_9422_transpose_y_0 = const()[name = tensor("op_9422_transpose_y_0"), val = tensor(false)]; + tensor transpose_118_perm_0 = const()[name = tensor("transpose_118_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_119_perm_0 = const()[name = tensor("transpose_119_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_119 = transpose(perm = transpose_119_perm_0, x = keys_141_cast_fp16)[name = tensor("transpose_121")]; + tensor transpose_118 = transpose(perm = transpose_118_perm_0, x = q_141_cast_fp16)[name = tensor("transpose_122")]; + tensor var_9422_cast_fp16 = matmul(transpose_x = var_9422_transpose_x_0, transpose_y = var_9422_transpose_y_0, x = transpose_118, y = transpose_119)[name = tensor("op_9422_cast_fp16")]; + tensor var_9416_to_fp16 = const()[name = tensor("op_9416_to_fp16"), val = tensor([0x1p-3])]; + tensor attn_weights_139_cast_fp16 = mul(x = var_9422_cast_fp16, y = var_9416_to_fp16)[name = tensor("attn_weights_139_cast_fp16")]; + tensor var_9424 = logical_not(x = attn_mask)[name = tensor("op_9424")]; + tensor var_9425_to_fp16 = const()[name = tensor("op_9425_to_fp16"), val = tensor(-inf)]; + tensor attn_weights_141_cast_fp16 = select(a = var_9425_to_fp16, b = attn_weights_139_cast_fp16, cond = var_9424)[name = tensor("attn_weights_141_cast_fp16")]; + tensor var_9427 = const()[name = tensor("op_9427"), val = tensor(-1)]; + tensor attn_weights_cast_fp16 = softmax(axis = var_9427, x = attn_weights_141_cast_fp16)[name = tensor("attn_weights_cast_fp16")]; + tensor attn_output_transpose_x_0 = const()[name = tensor("attn_output_transpose_x_0"), val = tensor(false)]; + tensor attn_output_transpose_y_0 = const()[name = tensor("attn_output_transpose_y_0"), val = tensor(false)]; + tensor values_cast_fp16 = transpose(perm = var_9367, x = values_141_cast_fp16)[name = tensor("transpose_123")]; + tensor attn_output_cast_fp16 = matmul(transpose_x = attn_output_transpose_x_0, transpose_y = attn_output_transpose_y_0, x = attn_weights_cast_fp16, y = values_cast_fp16)[name = tensor("attn_output_cast_fp16")]; + tensor var_9435 = const()[name = tensor("op_9435"), val = tensor([0, 2, 1, 3])]; + tensor var_9440 = const()[name = tensor("op_9440"), val = tensor([1, 1, 1024])]; + tensor var_9436_cast_fp16 = transpose(perm = var_9435, x = attn_output_cast_fp16)[name = tensor("transpose_120")]; + tensor input_283_cast_fp16 = reshape(shape = var_9440, x = var_9436_cast_fp16)[name = tensor("input_283_cast_fp16")]; + tensor attn23_out_proj_weight_to_fp16 = const()[name = tensor("attn23_out_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(586444928)))]; + tensor linear_94_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = attn23_out_proj_weight_to_fp16, x = input_283_cast_fp16)[name = tensor("linear_94_cast_fp16")]; + tensor var_9446_to_fp16 = const()[name = tensor("op_9446_to_fp16"), val = tensor(0x1p+0)]; + tensor var_9447_cast_fp16 = add(x = position23_to_fp16, y = var_9446_to_fp16)[name = tensor("op_9447_cast_fp16")]; + tensor var_9447_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_9447_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor input_285_cast_fp16 = add(x = input_279_cast_fp16, y = linear_94_cast_fp16)[name = tensor("input_285_cast_fp16")]; + tensor input_287_axes_0 = const()[name = tensor("input_287_axes_0"), val = tensor([-1])]; + tensor norm23_2_weight_to_fp16 = const()[name = tensor("norm23_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(588542144)))]; + tensor norm23_2_bias_to_fp16 = const()[name = tensor("norm23_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(588544256)))]; + tensor var_9451_to_fp16 = const()[name = tensor("op_9451_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_287_cast_fp16 = layer_norm(axes = input_287_axes_0, beta = norm23_2_bias_to_fp16, epsilon = var_9451_to_fp16, gamma = norm23_2_weight_to_fp16, x = input_285_cast_fp16)[name = tensor("input_287_cast_fp16")]; + tensor linear23_1_weight_to_fp16 = const()[name = tensor("linear23_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(588546368)))]; + tensor linear_95_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = linear23_1_weight_to_fp16, x = input_287_cast_fp16)[name = tensor("linear_95_cast_fp16")]; + tensor input_289_mode_0 = const()[name = tensor("input_289_mode_0"), val = tensor("EXACT")]; + tensor input_289_cast_fp16 = gelu(mode = input_289_mode_0, x = linear_95_cast_fp16)[name = tensor("input_289_cast_fp16")]; + tensor linear23_2_weight_to_fp16 = const()[name = tensor("linear23_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(596935040)))]; + tensor linear_96_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = linear23_2_weight_to_fp16, x = input_289_cast_fp16)[name = tensor("linear_96_cast_fp16")]; + tensor input_291_cast_fp16 = add(x = input_285_cast_fp16, y = linear_96_cast_fp16)[name = tensor("input_291_cast_fp16")]; + tensor input_axes_0 = const()[name = tensor("input_axes_0"), val = tensor([-1])]; + tensor out_norm_weight_to_fp16 = const()[name = tensor("out_norm_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(605323712)))]; + tensor out_norm_bias_to_fp16 = const()[name = tensor("out_norm_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(605325824)))]; + tensor var_9468_to_fp16 = const()[name = tensor("op_9468_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_cast_fp16 = layer_norm(axes = input_axes_0, beta = out_norm_bias_to_fp16, epsilon = var_9468_to_fp16, gamma = out_norm_weight_to_fp16, x = input_291_cast_fp16)[name = tensor("input_cast_fp16")]; + tensor input_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("input_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor out_eos_weight_to_fp16 = const()[name = tensor("out_eos_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(605327936)))]; + tensor out_eos_bias_to_fp16 = const()[name = tensor("out_eos_bias_to_fp16"), val = tensor([-0x1.64p-3])]; + tensor linear_97_cast_fp16 = linear(bias = out_eos_bias_to_fp16, weight = out_eos_weight_to_fp16, x = input_cast_fp16)[name = tensor("linear_97_cast_fp16")]; + tensor linear_97_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("linear_97_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor var_9476 = cast(dtype = linear_97_cast_fp16_to_fp32_dtype_0, x = linear_97_cast_fp16)[name = tensor("cast_432")]; + tensor input = cast(dtype = input_cast_fp16_to_fp32_dtype_0, x = input_cast_fp16)[name = tensor("cast_433")]; + tensor var_9447 = cast(dtype = var_9447_cast_fp16_to_fp32_dtype_0, x = var_9447_cast_fp16)[name = tensor("cast_434")]; + tensor new_cache_internal_tensor_assign_2 = cast(dtype = reshape_239_cast_fp16_to_fp32_dtype_0, x = reshape_239_cast_fp16)[name = tensor("cast_435")]; + tensor var_9061 = cast(dtype = var_9061_cast_fp16_to_fp32_dtype_0, x = var_9061_cast_fp16)[name = tensor("cast_439")]; + tensor new_cache_45_internal_tensor_assign_2 = cast(dtype = reshape_229_cast_fp16_to_fp32_dtype_0, x = reshape_229_cast_fp16)[name = tensor("cast_440")]; + tensor var_8675 = cast(dtype = var_8675_cast_fp16_to_fp32_dtype_0, x = var_8675_cast_fp16)[name = tensor("cast_444")]; + tensor new_cache_43_internal_tensor_assign_2 = cast(dtype = reshape_219_cast_fp16_to_fp32_dtype_0, x = reshape_219_cast_fp16)[name = tensor("cast_445")]; + tensor var_8289 = cast(dtype = var_8289_cast_fp16_to_fp32_dtype_0, x = var_8289_cast_fp16)[name = tensor("cast_449")]; + tensor new_cache_41_internal_tensor_assign_2 = cast(dtype = reshape_209_cast_fp16_to_fp32_dtype_0, x = reshape_209_cast_fp16)[name = tensor("cast_450")]; + tensor var_7903 = cast(dtype = var_7903_cast_fp16_to_fp32_dtype_0, x = var_7903_cast_fp16)[name = tensor("cast_454")]; + tensor new_cache_39_internal_tensor_assign_2 = cast(dtype = reshape_199_cast_fp16_to_fp32_dtype_0, x = reshape_199_cast_fp16)[name = tensor("cast_455")]; + tensor var_7517 = cast(dtype = var_7517_cast_fp16_to_fp32_dtype_0, x = var_7517_cast_fp16)[name = tensor("cast_459")]; + tensor new_cache_37_internal_tensor_assign_2 = cast(dtype = reshape_189_cast_fp16_to_fp32_dtype_0, x = reshape_189_cast_fp16)[name = tensor("cast_460")]; + tensor var_7131 = cast(dtype = var_7131_cast_fp16_to_fp32_dtype_0, x = var_7131_cast_fp16)[name = tensor("cast_464")]; + tensor new_cache_35_internal_tensor_assign_2 = cast(dtype = reshape_179_cast_fp16_to_fp32_dtype_0, x = reshape_179_cast_fp16)[name = tensor("cast_465")]; + tensor var_6745 = cast(dtype = var_6745_cast_fp16_to_fp32_dtype_0, x = var_6745_cast_fp16)[name = tensor("cast_469")]; + tensor new_cache_33_internal_tensor_assign_2 = cast(dtype = reshape_169_cast_fp16_to_fp32_dtype_0, x = reshape_169_cast_fp16)[name = tensor("cast_470")]; + tensor var_6359 = cast(dtype = var_6359_cast_fp16_to_fp32_dtype_0, x = var_6359_cast_fp16)[name = tensor("cast_474")]; + tensor new_cache_31_internal_tensor_assign_2 = cast(dtype = reshape_159_cast_fp16_to_fp32_dtype_0, x = reshape_159_cast_fp16)[name = tensor("cast_475")]; + tensor var_5973 = cast(dtype = var_5973_cast_fp16_to_fp32_dtype_0, x = var_5973_cast_fp16)[name = tensor("cast_479")]; + tensor new_cache_29_internal_tensor_assign_2 = cast(dtype = reshape_149_cast_fp16_to_fp32_dtype_0, x = reshape_149_cast_fp16)[name = tensor("cast_480")]; + tensor var_5587 = cast(dtype = var_5587_cast_fp16_to_fp32_dtype_0, x = var_5587_cast_fp16)[name = tensor("cast_484")]; + tensor new_cache_27_internal_tensor_assign_2 = cast(dtype = reshape_139_cast_fp16_to_fp32_dtype_0, x = reshape_139_cast_fp16)[name = tensor("cast_485")]; + tensor var_5201 = cast(dtype = var_5201_cast_fp16_to_fp32_dtype_0, x = var_5201_cast_fp16)[name = tensor("cast_489")]; + tensor new_cache_25_internal_tensor_assign_2 = cast(dtype = reshape_129_cast_fp16_to_fp32_dtype_0, x = reshape_129_cast_fp16)[name = tensor("cast_490")]; + tensor var_4815 = cast(dtype = var_4815_cast_fp16_to_fp32_dtype_0, x = var_4815_cast_fp16)[name = tensor("cast_494")]; + tensor new_cache_23_internal_tensor_assign_2 = cast(dtype = reshape_119_cast_fp16_to_fp32_dtype_0, x = reshape_119_cast_fp16)[name = tensor("cast_495")]; + tensor var_4429 = cast(dtype = var_4429_cast_fp16_to_fp32_dtype_0, x = var_4429_cast_fp16)[name = tensor("cast_499")]; + tensor new_cache_21_internal_tensor_assign_2 = cast(dtype = reshape_109_cast_fp16_to_fp32_dtype_0, x = reshape_109_cast_fp16)[name = tensor("cast_500")]; + tensor var_4043 = cast(dtype = var_4043_cast_fp16_to_fp32_dtype_0, x = var_4043_cast_fp16)[name = tensor("cast_504")]; + tensor new_cache_19_internal_tensor_assign_2 = cast(dtype = reshape_99_cast_fp16_to_fp32_dtype_0, x = reshape_99_cast_fp16)[name = tensor("cast_505")]; + tensor var_3657 = cast(dtype = var_3657_cast_fp16_to_fp32_dtype_0, x = var_3657_cast_fp16)[name = tensor("cast_509")]; + tensor new_cache_17_internal_tensor_assign_2 = cast(dtype = reshape_89_cast_fp16_to_fp32_dtype_0, x = reshape_89_cast_fp16)[name = tensor("cast_510")]; + tensor var_3271 = cast(dtype = var_3271_cast_fp16_to_fp32_dtype_0, x = var_3271_cast_fp16)[name = tensor("cast_514")]; + tensor new_cache_15_internal_tensor_assign_2 = cast(dtype = reshape_79_cast_fp16_to_fp32_dtype_0, x = reshape_79_cast_fp16)[name = tensor("cast_515")]; + tensor var_2885 = cast(dtype = var_2885_cast_fp16_to_fp32_dtype_0, x = var_2885_cast_fp16)[name = tensor("cast_519")]; + tensor new_cache_13_internal_tensor_assign_2 = cast(dtype = reshape_69_cast_fp16_to_fp32_dtype_0, x = reshape_69_cast_fp16)[name = tensor("cast_520")]; + tensor var_2499 = cast(dtype = var_2499_cast_fp16_to_fp32_dtype_0, x = var_2499_cast_fp16)[name = tensor("cast_524")]; + tensor new_cache_11_internal_tensor_assign_2 = cast(dtype = reshape_59_cast_fp16_to_fp32_dtype_0, x = reshape_59_cast_fp16)[name = tensor("cast_525")]; + tensor var_2113 = cast(dtype = var_2113_cast_fp16_to_fp32_dtype_0, x = var_2113_cast_fp16)[name = tensor("cast_529")]; + tensor new_cache_9_internal_tensor_assign_2 = cast(dtype = reshape_49_cast_fp16_to_fp32_dtype_0, x = reshape_49_cast_fp16)[name = tensor("cast_530")]; + tensor var_1727 = cast(dtype = var_1727_cast_fp16_to_fp32_dtype_0, x = var_1727_cast_fp16)[name = tensor("cast_534")]; + tensor new_cache_7_internal_tensor_assign_2 = cast(dtype = reshape_39_cast_fp16_to_fp32_dtype_0, x = reshape_39_cast_fp16)[name = tensor("cast_535")]; + tensor var_1341 = cast(dtype = var_1341_cast_fp16_to_fp32_dtype_0, x = var_1341_cast_fp16)[name = tensor("cast_539")]; + tensor new_cache_5_internal_tensor_assign_2 = cast(dtype = reshape_29_cast_fp16_to_fp32_dtype_0, x = reshape_29_cast_fp16)[name = tensor("cast_540")]; + tensor var_955 = cast(dtype = var_955_cast_fp16_to_fp32_dtype_0, x = var_955_cast_fp16)[name = tensor("cast_544")]; + tensor new_cache_3_internal_tensor_assign_2 = cast(dtype = reshape_19_cast_fp16_to_fp32_dtype_0, x = reshape_19_cast_fp16)[name = tensor("cast_545")]; + tensor var_569 = cast(dtype = var_569_cast_fp16_to_fp32_dtype_0, x = var_569_cast_fp16)[name = tensor("cast_549")]; + tensor new_cache_1_internal_tensor_assign_2 = cast(dtype = reshape_9_cast_fp16_to_fp32_dtype_0, x = reshape_9_cast_fp16)[name = tensor("cast_550")]; + } -> (input, var_9476, new_cache_1_internal_tensor_assign_2, var_569, new_cache_3_internal_tensor_assign_2, var_955, new_cache_5_internal_tensor_assign_2, var_1341, new_cache_7_internal_tensor_assign_2, var_1727, new_cache_9_internal_tensor_assign_2, var_2113, new_cache_11_internal_tensor_assign_2, var_2499, new_cache_13_internal_tensor_assign_2, var_2885, new_cache_15_internal_tensor_assign_2, var_3271, new_cache_17_internal_tensor_assign_2, var_3657, new_cache_19_internal_tensor_assign_2, var_4043, new_cache_21_internal_tensor_assign_2, var_4429, new_cache_23_internal_tensor_assign_2, var_4815, new_cache_25_internal_tensor_assign_2, var_5201, new_cache_27_internal_tensor_assign_2, var_5587, new_cache_29_internal_tensor_assign_2, var_5973, new_cache_31_internal_tensor_assign_2, var_6359, new_cache_33_internal_tensor_assign_2, var_6745, new_cache_35_internal_tensor_assign_2, var_7131, new_cache_37_internal_tensor_assign_2, var_7517, new_cache_39_internal_tensor_assign_2, var_7903, new_cache_41_internal_tensor_assign_2, var_8289, new_cache_43_internal_tensor_assign_2, var_8675, new_cache_45_internal_tensor_assign_2, var_9061, new_cache_internal_tensor_assign_2, var_9447); +} \ No newline at end of file diff --git a/v2.1/italian_24l/flowlm_step.mlmodelc/weights/weight.bin b/v2.1/italian_24l/flowlm_step.mlmodelc/weights/weight.bin new file mode 100644 index 0000000000000000000000000000000000000000..d4fc91250ab3e068d7669fd9ba54bb2392e5b5a0 --- /dev/null +++ b/v2.1/italian_24l/flowlm_step.mlmodelc/weights/weight.bin @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:59289273d10c63be904dbc872b22daab4b214125377550d9dbf1b82f14fc6aa3 +size 605330048 diff --git a/v2.1/italian_24l/flowlm_step.mlpackage/Data/com.apple.CoreML/model.mlmodel b/v2.1/italian_24l/flowlm_step.mlpackage/Data/com.apple.CoreML/model.mlmodel new file mode 100644 index 0000000000000000000000000000000000000000..51e026fc2678dc7b8a9081437812dfaaf9f90e4d --- /dev/null +++ b/v2.1/italian_24l/flowlm_step.mlpackage/Data/com.apple.CoreML/model.mlmodel @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:1b65b41b48223df6161151299512c753ded770be448a37b518affff5aeee4dee +size 825108 diff --git a/v2.1/italian_24l/flowlm_step.mlpackage/Data/com.apple.CoreML/weights/weight.bin b/v2.1/italian_24l/flowlm_step.mlpackage/Data/com.apple.CoreML/weights/weight.bin new file mode 100644 index 0000000000000000000000000000000000000000..d4fc91250ab3e068d7669fd9ba54bb2392e5b5a0 --- /dev/null +++ b/v2.1/italian_24l/flowlm_step.mlpackage/Data/com.apple.CoreML/weights/weight.bin @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:59289273d10c63be904dbc872b22daab4b214125377550d9dbf1b82f14fc6aa3 +size 605330048 diff --git a/v2.1/italian_24l/flowlm_step.mlpackage/Manifest.json b/v2.1/italian_24l/flowlm_step.mlpackage/Manifest.json new file mode 100644 index 0000000000000000000000000000000000000000..7fbce2103c0d2b8fd743723d8fd89ecb1dfae346 --- /dev/null +++ b/v2.1/italian_24l/flowlm_step.mlpackage/Manifest.json @@ -0,0 +1,18 @@ +{ + "fileFormatVersion": "1.0.0", + "itemInfoEntries": { + "73B3B87E-294E-4E4D-A852-FD3B791EF31E": { + "author": "com.apple.CoreML", + "description": "CoreML Model Weights", + "name": "weights", + "path": "com.apple.CoreML/weights" + }, + "74045170-2EB6-4575-A35B-BE123EAB4FC1": { + "author": "com.apple.CoreML", + "description": "CoreML Model Specification", + "name": "model.mlmodel", + "path": "com.apple.CoreML/model.mlmodel" + } + }, + "rootModelIdentifier": "74045170-2EB6-4575-A35B-BE123EAB4FC1" +} diff --git a/v2.1/italian_24l/flowlm_stepv2.mlmodelc/analytics/coremldata.bin b/v2.1/italian_24l/flowlm_stepv2.mlmodelc/analytics/coremldata.bin new file mode 100644 index 0000000000000000000000000000000000000000..7b9198616531e8ed9c5e902249e4093b2513696a --- /dev/null +++ b/v2.1/italian_24l/flowlm_stepv2.mlmodelc/analytics/coremldata.bin @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:8d53eed01fbb7a31bb877a74921aabf1dc8ee8f94def9d8b56ff5f44c1dcda35 +size 243 diff --git a/v2.1/italian_24l/flowlm_stepv2.mlmodelc/coremldata.bin b/v2.1/italian_24l/flowlm_stepv2.mlmodelc/coremldata.bin new file mode 100644 index 0000000000000000000000000000000000000000..98660effe1f7c37734f5b9cc13e60ed7e219a1bf --- /dev/null +++ b/v2.1/italian_24l/flowlm_stepv2.mlmodelc/coremldata.bin @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:4e5930318538862799470bd2cc790a6cc6079e9ba9d35249d74fc0850217b096 +size 3561 diff --git a/v2.1/italian_24l/flowlm_stepv2.mlmodelc/metadata.json b/v2.1/italian_24l/flowlm_stepv2.mlmodelc/metadata.json new file mode 100644 index 0000000000000000000000000000000000000000..629da9a69819375413a6c11a7789eb2ca4f04e28 --- /dev/null +++ b/v2.1/italian_24l/flowlm_stepv2.mlmodelc/metadata.json @@ -0,0 +1,1068 @@ +[ + { + "metadataOutputVersion" : "3.0", + "storagePrecision" : "Mixed (Float16, Int8)", + "outputSchema" : [ + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1 × 1 × 1024)", + "shortDescription" : "", + "shape" : "[1, 1, 1024]", + "name" : "input", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1 × 1 × 1)", + "shortDescription" : "", + "shape" : "[1, 1, 1]", + "name" : "var_10268", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 2 × 1 × 512 × 16 × 64)", + "shortDescription" : "", + "shape" : "[2, 1, 512, 16, 64]", + "name" : "new_cache_1_internal_tensor_assign_2", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1)", + "shortDescription" : "", + "shape" : "[1]", + "name" : "var_602", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 2 × 1 × 512 × 16 × 64)", + "shortDescription" : "", + "shape" : "[2, 1, 512, 16, 64]", + "name" : "new_cache_3_internal_tensor_assign_2", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1)", + "shortDescription" : "", + "shape" : "[1]", + "name" : "var_1021", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 2 × 1 × 512 × 16 × 64)", + "shortDescription" : "", + "shape" : "[2, 1, 512, 16, 64]", + "name" : "new_cache_5_internal_tensor_assign_2", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1)", + "shortDescription" : "", + "shape" : "[1]", + "name" : "var_1440", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 2 × 1 × 512 × 16 × 64)", + "shortDescription" : "", + "shape" : "[2, 1, 512, 16, 64]", + "name" : "new_cache_7_internal_tensor_assign_2", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1)", + "shortDescription" : "", + "shape" : "[1]", + "name" : "var_1859", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 2 × 1 × 512 × 16 × 64)", + "shortDescription" : "", + "shape" : "[2, 1, 512, 16, 64]", + "name" : "new_cache_9_internal_tensor_assign_2", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1)", + "shortDescription" : "", + "shape" : "[1]", + "name" : "var_2278", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 2 × 1 × 512 × 16 × 64)", + "shortDescription" : "", + "shape" : "[2, 1, 512, 16, 64]", + "name" : "new_cache_11_internal_tensor_assign_2", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1)", + "shortDescription" : "", + "shape" : "[1]", + "name" : "var_2697", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 2 × 1 × 512 × 16 × 64)", + "shortDescription" : "", + "shape" : "[2, 1, 512, 16, 64]", + "name" : "new_cache_13_internal_tensor_assign_2", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1)", + "shortDescription" : "", + "shape" : "[1]", + "name" : "var_3116", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 2 × 1 × 512 × 16 × 64)", + "shortDescription" : "", + "shape" : "[2, 1, 512, 16, 64]", + "name" : "new_cache_15_internal_tensor_assign_2", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1)", + "shortDescription" : "", + "shape" : "[1]", + "name" : "var_3535", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 2 × 1 × 512 × 16 × 64)", + "shortDescription" : "", + "shape" : "[2, 1, 512, 16, 64]", + "name" : "new_cache_17_internal_tensor_assign_2", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1)", + "shortDescription" : "", + "shape" : "[1]", + "name" : "var_3954", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 2 × 1 × 512 × 16 × 64)", + "shortDescription" : "", + "shape" : "[2, 1, 512, 16, 64]", + "name" : "new_cache_19_internal_tensor_assign_2", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1)", + "shortDescription" : "", + "shape" : "[1]", + "name" : "var_4373", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 2 × 1 × 512 × 16 × 64)", + "shortDescription" : "", + "shape" : "[2, 1, 512, 16, 64]", + "name" : "new_cache_21_internal_tensor_assign_2", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1)", + "shortDescription" : "", + "shape" : "[1]", + "name" : "var_4792", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 2 × 1 × 512 × 16 × 64)", + "shortDescription" : "", + "shape" : "[2, 1, 512, 16, 64]", + "name" : "new_cache_23_internal_tensor_assign_2", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1)", + "shortDescription" : "", + "shape" : "[1]", + "name" : "var_5211", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 2 × 1 × 512 × 16 × 64)", + "shortDescription" : "", + "shape" : "[2, 1, 512, 16, 64]", + "name" : "new_cache_25_internal_tensor_assign_2", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1)", + "shortDescription" : "", + "shape" : "[1]", + "name" : "var_5630", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 2 × 1 × 512 × 16 × 64)", + "shortDescription" : "", + "shape" : "[2, 1, 512, 16, 64]", + "name" : "new_cache_27_internal_tensor_assign_2", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1)", + "shortDescription" : "", + "shape" : "[1]", + "name" : "var_6049", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 2 × 1 × 512 × 16 × 64)", + "shortDescription" : "", + "shape" : "[2, 1, 512, 16, 64]", + "name" : "new_cache_29_internal_tensor_assign_2", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1)", + "shortDescription" : "", + "shape" : "[1]", + "name" : "var_6468", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 2 × 1 × 512 × 16 × 64)", + "shortDescription" : "", + "shape" : "[2, 1, 512, 16, 64]", + "name" : "new_cache_31_internal_tensor_assign_2", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1)", + "shortDescription" : "", + "shape" : "[1]", + "name" : "var_6887", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 2 × 1 × 512 × 16 × 64)", + "shortDescription" : "", + "shape" : "[2, 1, 512, 16, 64]", + "name" : "new_cache_33_internal_tensor_assign_2", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1)", + "shortDescription" : "", + "shape" : "[1]", + "name" : "var_7306", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 2 × 1 × 512 × 16 × 64)", + "shortDescription" : "", + "shape" : "[2, 1, 512, 16, 64]", + "name" : "new_cache_35_internal_tensor_assign_2", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1)", + "shortDescription" : "", + "shape" : "[1]", + "name" : "var_7725", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 2 × 1 × 512 × 16 × 64)", + "shortDescription" : "", + "shape" : "[2, 1, 512, 16, 64]", + "name" : "new_cache_37_internal_tensor_assign_2", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1)", + "shortDescription" : "", + "shape" : "[1]", + "name" : "var_8144", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 2 × 1 × 512 × 16 × 64)", + "shortDescription" : "", + "shape" : "[2, 1, 512, 16, 64]", + "name" : "new_cache_39_internal_tensor_assign_2", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1)", + "shortDescription" : "", + "shape" : "[1]", + "name" : "var_8563", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 2 × 1 × 512 × 16 × 64)", + "shortDescription" : "", + "shape" : "[2, 1, 512, 16, 64]", + "name" : "new_cache_41_internal_tensor_assign_2", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1)", + "shortDescription" : "", + "shape" : "[1]", + "name" : "var_8982", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 2 × 1 × 512 × 16 × 64)", + "shortDescription" : "", + "shape" : "[2, 1, 512, 16, 64]", + "name" : "new_cache_43_internal_tensor_assign_2", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1)", + "shortDescription" : "", + "shape" : "[1]", + "name" : "var_9401", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 2 × 1 × 512 × 16 × 64)", + "shortDescription" : "", + "shape" : "[2, 1, 512, 16, 64]", + "name" : "new_cache_45_internal_tensor_assign_2", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1)", + "shortDescription" : "", + "shape" : "[1]", + "name" : "var_9820", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 2 × 1 × 512 × 16 × 64)", + "shortDescription" : "", + "shape" : "[2, 1, 512, 16, 64]", + "name" : "new_cache_internal_tensor_assign_2", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1)", + "shortDescription" : "", + "shape" : "[1]", + "name" : "var_10239", + "type" : "MultiArray" + } + ], + "modelParameters" : [ + + ], + "specificationVersion" : 8, + "mlProgramOperationTypeHistogram" : { + "Ios17.cos" : 24, + "Range1d" : 48, + "Ios17.notEqual" : 49, + "Ios17.floor" : 24, + "Ios17.reshape" : 432, + "Ios17.logicalAnd" : 24, + "Ios16.softmax" : 24, + "Ios17.scatter" : 48, + "Ios17.matmul" : 48, + "Ios17.sin" : 24, + "Ios17.transpose" : 96, + "Select" : 73, + "Ios17.expandDims" : 25, + "Ios17.add" : 144, + "Tile" : 24, + "Ios17.sliceByIndex" : 312, + "Ios17.lessEqual" : 24, + "Ios17.scatterAlongAxis" : 48, + "Ios17.logicalNot" : 24, + "Ios17.layerNorm" : 49, + "Ios16.constexprAffineDequantize" : 96, + "Ios17.cast" : 124, + "Ios17.sub" : 72, + "Ios17.less" : 24, + "Ios16.gelu" : 24, + "Ios17.linear" : 98, + "Stack" : 48, + "Ios17.mul" : 288 + }, + "computePrecision" : "Mixed (Float16, Float32, Int32)", + "isUpdatable" : "0", + "stateSchema" : [ + + ], + "availability" : { + "macOS" : "14.0", + "tvOS" : "17.0", + "visionOS" : "1.0", + "watchOS" : "10.0", + "iOS" : "17.0", + "macCatalyst" : "17.0" + }, + "modelType" : { + "name" : "MLModelType_mlProgram" + }, + "userDefinedMetadata" : { + "com.github.apple.coremltools.conversion_date" : "2026-04-27", + "com.github.apple.coremltools.source" : "torch==2.9.1", + "com.github.apple.coremltools.version" : "9.0", + "com.github.apple.coremltools.source_dialect" : "TorchScript" + }, + "inputSchema" : [ + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1 × 1 × 32)", + "shortDescription" : "", + "shape" : "[1, 1, 32]", + "name" : "sequence", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 32)", + "shortDescription" : "", + "shape" : "[32]", + "name" : "bos_emb", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 2 × 1 × 512 × 16 × 64)", + "shortDescription" : "", + "shape" : "[2, 1, 512, 16, 64]", + "name" : "cache0", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1)", + "shortDescription" : "", + "shape" : "[1]", + "name" : "position0", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 2 × 1 × 512 × 16 × 64)", + "shortDescription" : "", + "shape" : "[2, 1, 512, 16, 64]", + "name" : "cache1", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1)", + "shortDescription" : "", + "shape" : "[1]", + "name" : "position1", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 2 × 1 × 512 × 16 × 64)", + "shortDescription" : "", + "shape" : "[2, 1, 512, 16, 64]", + "name" : "cache2", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1)", + "shortDescription" : "", + "shape" : "[1]", + "name" : "position2", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 2 × 1 × 512 × 16 × 64)", + "shortDescription" : "", + "shape" : "[2, 1, 512, 16, 64]", + "name" : "cache3", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1)", + "shortDescription" : "", + "shape" : "[1]", + "name" : "position3", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 2 × 1 × 512 × 16 × 64)", + "shortDescription" : "", + "shape" : "[2, 1, 512, 16, 64]", + "name" : "cache4", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1)", + "shortDescription" : "", + "shape" : "[1]", + "name" : "position4", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 2 × 1 × 512 × 16 × 64)", + "shortDescription" : "", + "shape" : "[2, 1, 512, 16, 64]", + "name" : "cache5", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1)", + "shortDescription" : "", + "shape" : "[1]", + "name" : "position5", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 2 × 1 × 512 × 16 × 64)", + "shortDescription" : "", + "shape" : "[2, 1, 512, 16, 64]", + "name" : "cache6", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1)", + "shortDescription" : "", + "shape" : "[1]", + "name" : "position6", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 2 × 1 × 512 × 16 × 64)", + "shortDescription" : "", + "shape" : "[2, 1, 512, 16, 64]", + "name" : "cache7", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1)", + "shortDescription" : "", + "shape" : "[1]", + "name" : "position7", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 2 × 1 × 512 × 16 × 64)", + "shortDescription" : "", + "shape" : "[2, 1, 512, 16, 64]", + "name" : "cache8", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1)", + "shortDescription" : "", + "shape" : "[1]", + "name" : "position8", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 2 × 1 × 512 × 16 × 64)", + "shortDescription" : "", + "shape" : "[2, 1, 512, 16, 64]", + "name" : "cache9", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1)", + "shortDescription" : "", + "shape" : "[1]", + "name" : "position9", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 2 × 1 × 512 × 16 × 64)", + "shortDescription" : "", + "shape" : "[2, 1, 512, 16, 64]", + "name" : "cache10", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1)", + "shortDescription" : "", + "shape" : "[1]", + "name" : "position10", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 2 × 1 × 512 × 16 × 64)", + "shortDescription" : "", + "shape" : "[2, 1, 512, 16, 64]", + "name" : "cache11", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1)", + "shortDescription" : "", + "shape" : "[1]", + "name" : "position11", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 2 × 1 × 512 × 16 × 64)", + "shortDescription" : "", + "shape" : "[2, 1, 512, 16, 64]", + "name" : "cache12", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1)", + "shortDescription" : "", + "shape" : "[1]", + "name" : "position12", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 2 × 1 × 512 × 16 × 64)", + "shortDescription" : "", + "shape" : "[2, 1, 512, 16, 64]", + "name" : "cache13", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1)", + "shortDescription" : "", + "shape" : "[1]", + "name" : "position13", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 2 × 1 × 512 × 16 × 64)", + "shortDescription" : "", + "shape" : "[2, 1, 512, 16, 64]", + "name" : "cache14", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1)", + "shortDescription" : "", + "shape" : "[1]", + "name" : "position14", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 2 × 1 × 512 × 16 × 64)", + "shortDescription" : "", + "shape" : "[2, 1, 512, 16, 64]", + "name" : "cache15", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1)", + "shortDescription" : "", + "shape" : "[1]", + "name" : "position15", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 2 × 1 × 512 × 16 × 64)", + "shortDescription" : "", + "shape" : "[2, 1, 512, 16, 64]", + "name" : "cache16", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1)", + "shortDescription" : "", + "shape" : "[1]", + "name" : "position16", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 2 × 1 × 512 × 16 × 64)", + "shortDescription" : "", + "shape" : "[2, 1, 512, 16, 64]", + "name" : "cache17", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1)", + "shortDescription" : "", + "shape" : "[1]", + "name" : "position17", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 2 × 1 × 512 × 16 × 64)", + "shortDescription" : "", + "shape" : "[2, 1, 512, 16, 64]", + "name" : "cache18", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1)", + "shortDescription" : "", + "shape" : "[1]", + "name" : "position18", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 2 × 1 × 512 × 16 × 64)", + "shortDescription" : "", + "shape" : "[2, 1, 512, 16, 64]", + "name" : "cache19", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1)", + "shortDescription" : "", + "shape" : "[1]", + "name" : "position19", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 2 × 1 × 512 × 16 × 64)", + "shortDescription" : "", + "shape" : "[2, 1, 512, 16, 64]", + "name" : "cache20", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1)", + "shortDescription" : "", + "shape" : "[1]", + "name" : "position20", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 2 × 1 × 512 × 16 × 64)", + "shortDescription" : "", + "shape" : "[2, 1, 512, 16, 64]", + "name" : "cache21", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1)", + "shortDescription" : "", + "shape" : "[1]", + "name" : "position21", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 2 × 1 × 512 × 16 × 64)", + "shortDescription" : "", + "shape" : "[2, 1, 512, 16, 64]", + "name" : "cache22", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1)", + "shortDescription" : "", + "shape" : "[1]", + "name" : "position22", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 2 × 1 × 512 × 16 × 64)", + "shortDescription" : "", + "shape" : "[2, 1, 512, 16, 64]", + "name" : "cache23", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1)", + "shortDescription" : "", + "shape" : "[1]", + "name" : "position23", + "type" : "MultiArray" + } + ], + "generatedClassName" : "flowlm_stepv2", + "method" : "predict" + } +] \ No newline at end of file diff --git a/v2.1/italian_24l/flowlm_stepv2.mlmodelc/model.mil b/v2.1/italian_24l/flowlm_stepv2.mlmodelc/model.mil new file mode 100644 index 0000000000000000000000000000000000000000..cf6f9403eaa6a8f69177704dd4a78d56a7061740 --- /dev/null +++ b/v2.1/italian_24l/flowlm_stepv2.mlmodelc/model.mil @@ -0,0 +1,5337 @@ +program(1.0) +[buildInfo = dict, tensor>({{"coremlc-component-MIL", "3520.4.1"}, {"coremlc-version", "3520.5.1"}, {"coremltools-component-torch", "2.9.1"}, {"coremltools-source-dialect", "TorchScript"}, {"coremltools-version", "9.0"}})] +{ + func main(tensor bos_emb, tensor cache0, tensor cache1, tensor cache10, tensor cache11, tensor cache12, tensor cache13, tensor cache14, tensor cache15, tensor cache16, tensor cache17, tensor cache18, tensor cache19, tensor cache2, tensor cache20, tensor cache21, tensor cache22, tensor cache23, tensor cache3, tensor cache4, tensor cache5, tensor cache6, tensor cache7, tensor cache8, tensor cache9, tensor position0, tensor position1, tensor position10, tensor position11, tensor position12, tensor position13, tensor position14, tensor position15, tensor position16, tensor position17, tensor position18, tensor position19, tensor position2, tensor position20, tensor position21, tensor position22, tensor position23, tensor position3, tensor position4, tensor position5, tensor position6, tensor position7, tensor position8, tensor position9, tensor sequence) { + tensor attn0_in_proj_weight_cast_fp16 = constexpr_affine_dequantize()[axis = tensor(0), name = tensor("attn0_in_proj_weight_cast_fp16"), quantized_data = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(64))), scale = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(3148992))), zero_point = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(3145856)))]; + tensor attn0_out_proj_weight_cast_fp16 = constexpr_affine_dequantize()[axis = tensor(0), name = tensor("attn0_out_proj_weight_cast_fp16"), quantized_data = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(3155200))), scale = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(4204928))), zero_point = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(4203840)))]; + tensor linear0_1_weight_cast_fp16 = constexpr_affine_dequantize()[axis = tensor(0), name = tensor("linear0_1_weight_cast_fp16"), quantized_data = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(4207040))), scale = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8405568))), zero_point = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8401408)))]; + tensor linear0_2_weight_cast_fp16 = constexpr_affine_dequantize()[axis = tensor(0), name = tensor("linear0_2_weight_cast_fp16"), quantized_data = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8413824))), scale = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(12608192))), zero_point = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(4203840)))]; + tensor attn1_in_proj_weight_cast_fp16 = constexpr_affine_dequantize()[axis = tensor(0), name = tensor("attn1_in_proj_weight_cast_fp16"), quantized_data = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(12610304))), scale = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15756096))), zero_point = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(3145856)))]; + tensor attn1_out_proj_weight_cast_fp16 = constexpr_affine_dequantize()[axis = tensor(0), name = tensor("attn1_out_proj_weight_cast_fp16"), quantized_data = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15762304))), scale = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(16810944))), zero_point = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(4203840)))]; + tensor linear1_1_weight_cast_fp16 = constexpr_affine_dequantize()[axis = tensor(0), name = tensor("linear1_1_weight_cast_fp16"), quantized_data = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(16813056))), scale = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21007424))), zero_point = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8401408)))]; + tensor linear1_2_weight_cast_fp16 = constexpr_affine_dequantize()[axis = tensor(0), name = tensor("linear1_2_weight_cast_fp16"), quantized_data = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21015680))), scale = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(25210048))), zero_point = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(4203840)))]; + tensor attn2_in_proj_weight_cast_fp16 = constexpr_affine_dequantize()[axis = tensor(0), name = tensor("attn2_in_proj_weight_cast_fp16"), quantized_data = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(25212160))), scale = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28357952))), zero_point = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(3145856)))]; + tensor attn2_out_proj_weight_cast_fp16 = constexpr_affine_dequantize()[axis = tensor(0), name = tensor("attn2_out_proj_weight_cast_fp16"), quantized_data = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28364160))), scale = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(29412800))), zero_point = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(4203840)))]; + tensor linear2_1_weight_cast_fp16 = constexpr_affine_dequantize()[axis = tensor(0), name = tensor("linear2_1_weight_cast_fp16"), quantized_data = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(29414912))), scale = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(33609280))), zero_point = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8401408)))]; + tensor linear2_2_weight_cast_fp16 = constexpr_affine_dequantize()[axis = tensor(0), name = tensor("linear2_2_weight_cast_fp16"), quantized_data = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(33617536))), scale = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(37811904))), zero_point = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(4203840)))]; + tensor attn3_in_proj_weight_cast_fp16 = constexpr_affine_dequantize()[axis = tensor(0), name = tensor("attn3_in_proj_weight_cast_fp16"), quantized_data = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(37814016))), scale = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(40959808))), zero_point = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(3145856)))]; + tensor attn3_out_proj_weight_cast_fp16 = constexpr_affine_dequantize()[axis = tensor(0), name = tensor("attn3_out_proj_weight_cast_fp16"), quantized_data = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(40966016))), scale = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(42014656))), zero_point = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(4203840)))]; + tensor linear3_1_weight_cast_fp16 = constexpr_affine_dequantize()[axis = tensor(0), name = tensor("linear3_1_weight_cast_fp16"), quantized_data = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(42016768))), scale = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(46211136))), zero_point = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8401408)))]; + tensor linear3_2_weight_cast_fp16 = constexpr_affine_dequantize()[axis = tensor(0), name = tensor("linear3_2_weight_cast_fp16"), quantized_data = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(46219392))), scale = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(50413760))), zero_point = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(4203840)))]; + tensor attn4_in_proj_weight_cast_fp16 = constexpr_affine_dequantize()[axis = tensor(0), name = tensor("attn4_in_proj_weight_cast_fp16"), quantized_data = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(50415872))), scale = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(53561664))), zero_point = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(3145856)))]; + tensor attn4_out_proj_weight_cast_fp16 = constexpr_affine_dequantize()[axis = tensor(0), name = tensor("attn4_out_proj_weight_cast_fp16"), quantized_data = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(53567872))), scale = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(54616512))), zero_point = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(4203840)))]; + tensor linear4_1_weight_cast_fp16 = constexpr_affine_dequantize()[axis = tensor(0), name = tensor("linear4_1_weight_cast_fp16"), quantized_data = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(54618624))), scale = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(58812992))), zero_point = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8401408)))]; + tensor linear4_2_weight_cast_fp16 = constexpr_affine_dequantize()[axis = tensor(0), name = tensor("linear4_2_weight_cast_fp16"), quantized_data = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(58821248))), scale = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(63015616))), zero_point = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(4203840)))]; + tensor attn5_in_proj_weight_cast_fp16 = constexpr_affine_dequantize()[axis = tensor(0), name = tensor("attn5_in_proj_weight_cast_fp16"), quantized_data = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(63017728))), scale = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(66163520))), zero_point = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(3145856)))]; + tensor attn5_out_proj_weight_cast_fp16 = constexpr_affine_dequantize()[axis = tensor(0), name = tensor("attn5_out_proj_weight_cast_fp16"), quantized_data = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(66169728))), scale = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(67218368))), zero_point = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(4203840)))]; + tensor linear5_1_weight_cast_fp16 = constexpr_affine_dequantize()[axis = tensor(0), name = tensor("linear5_1_weight_cast_fp16"), quantized_data = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(67220480))), scale = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(71414848))), zero_point = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8401408)))]; + tensor linear5_2_weight_cast_fp16 = constexpr_affine_dequantize()[axis = tensor(0), name = tensor("linear5_2_weight_cast_fp16"), quantized_data = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(71423104))), scale = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(75617472))), zero_point = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(4203840)))]; + tensor attn6_in_proj_weight_cast_fp16 = constexpr_affine_dequantize()[axis = tensor(0), name = tensor("attn6_in_proj_weight_cast_fp16"), quantized_data = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(75619584))), scale = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(78765376))), zero_point = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(3145856)))]; + tensor attn6_out_proj_weight_cast_fp16 = constexpr_affine_dequantize()[axis = tensor(0), name = tensor("attn6_out_proj_weight_cast_fp16"), quantized_data = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(78771584))), scale = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(79820224))), zero_point = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(4203840)))]; + tensor linear6_1_weight_cast_fp16 = constexpr_affine_dequantize()[axis = tensor(0), name = tensor("linear6_1_weight_cast_fp16"), quantized_data = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(79822336))), scale = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(84016704))), zero_point = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8401408)))]; + tensor linear6_2_weight_cast_fp16 = constexpr_affine_dequantize()[axis = tensor(0), name = tensor("linear6_2_weight_cast_fp16"), quantized_data = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(84024960))), scale = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(88219328))), zero_point = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(4203840)))]; + tensor attn7_in_proj_weight_cast_fp16 = constexpr_affine_dequantize()[axis = tensor(0), name = tensor("attn7_in_proj_weight_cast_fp16"), quantized_data = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(88221440))), scale = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(91367232))), zero_point = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(3145856)))]; + tensor attn7_out_proj_weight_cast_fp16 = constexpr_affine_dequantize()[axis = tensor(0), name = tensor("attn7_out_proj_weight_cast_fp16"), quantized_data = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(91373440))), scale = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(92422080))), zero_point = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(4203840)))]; + tensor linear7_1_weight_cast_fp16 = constexpr_affine_dequantize()[axis = tensor(0), name = tensor("linear7_1_weight_cast_fp16"), quantized_data = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(92424192))), scale = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(96618560))), zero_point = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8401408)))]; + tensor linear7_2_weight_cast_fp16 = constexpr_affine_dequantize()[axis = tensor(0), name = tensor("linear7_2_weight_cast_fp16"), quantized_data = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(96626816))), scale = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(100821184))), zero_point = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(4203840)))]; + tensor attn8_in_proj_weight_cast_fp16 = constexpr_affine_dequantize()[axis = tensor(0), name = tensor("attn8_in_proj_weight_cast_fp16"), quantized_data = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(100823296))), scale = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(103969088))), zero_point = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(3145856)))]; + tensor attn8_out_proj_weight_cast_fp16 = constexpr_affine_dequantize()[axis = tensor(0), name = tensor("attn8_out_proj_weight_cast_fp16"), quantized_data = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(103975296))), scale = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(105023936))), zero_point = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(4203840)))]; + tensor linear8_1_weight_cast_fp16 = constexpr_affine_dequantize()[axis = tensor(0), name = tensor("linear8_1_weight_cast_fp16"), quantized_data = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(105026048))), scale = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(109220416))), zero_point = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8401408)))]; + tensor linear8_2_weight_cast_fp16 = constexpr_affine_dequantize()[axis = tensor(0), name = tensor("linear8_2_weight_cast_fp16"), quantized_data = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(109228672))), scale = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(113423040))), zero_point = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(4203840)))]; + tensor attn9_in_proj_weight_cast_fp16 = constexpr_affine_dequantize()[axis = tensor(0), name = tensor("attn9_in_proj_weight_cast_fp16"), quantized_data = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(113425152))), scale = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(116570944))), zero_point = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(3145856)))]; + tensor attn9_out_proj_weight_cast_fp16 = constexpr_affine_dequantize()[axis = tensor(0), name = tensor("attn9_out_proj_weight_cast_fp16"), quantized_data = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(116577152))), scale = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(117625792))), zero_point = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(4203840)))]; + tensor linear9_1_weight_cast_fp16 = constexpr_affine_dequantize()[axis = tensor(0), name = tensor("linear9_1_weight_cast_fp16"), quantized_data = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(117627904))), scale = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(121822272))), zero_point = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8401408)))]; + tensor linear9_2_weight_cast_fp16 = constexpr_affine_dequantize()[axis = tensor(0), name = tensor("linear9_2_weight_cast_fp16"), quantized_data = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(121830528))), scale = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(126024896))), zero_point = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(4203840)))]; + tensor attn10_in_proj_weight_cast_fp16 = constexpr_affine_dequantize()[axis = tensor(0), name = tensor("attn10_in_proj_weight_cast_fp16"), quantized_data = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(126027008))), scale = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(129172800))), zero_point = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(3145856)))]; + tensor attn10_out_proj_weight_cast_fp16 = constexpr_affine_dequantize()[axis = tensor(0), name = tensor("attn10_out_proj_weight_cast_fp16"), quantized_data = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(129179008))), scale = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(130227648))), zero_point = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(4203840)))]; + tensor linear10_1_weight_cast_fp16 = constexpr_affine_dequantize()[axis = tensor(0), name = tensor("linear10_1_weight_cast_fp16"), quantized_data = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(130229760))), scale = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(134424128))), zero_point = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8401408)))]; + tensor linear10_2_weight_cast_fp16 = constexpr_affine_dequantize()[axis = tensor(0), name = tensor("linear10_2_weight_cast_fp16"), quantized_data = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(134432384))), scale = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(138626752))), zero_point = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(4203840)))]; + tensor attn11_in_proj_weight_cast_fp16 = constexpr_affine_dequantize()[axis = tensor(0), name = tensor("attn11_in_proj_weight_cast_fp16"), quantized_data = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(138628864))), scale = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(141774656))), zero_point = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(3145856)))]; + tensor attn11_out_proj_weight_cast_fp16 = constexpr_affine_dequantize()[axis = tensor(0), name = tensor("attn11_out_proj_weight_cast_fp16"), quantized_data = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(141780864))), scale = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(142829504))), zero_point = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(4203840)))]; + tensor linear11_1_weight_cast_fp16 = constexpr_affine_dequantize()[axis = tensor(0), name = tensor("linear11_1_weight_cast_fp16"), quantized_data = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(142831616))), scale = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(147025984))), zero_point = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8401408)))]; + tensor linear11_2_weight_cast_fp16 = constexpr_affine_dequantize()[axis = tensor(0), name = tensor("linear11_2_weight_cast_fp16"), quantized_data = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(147034240))), scale = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(151228608))), zero_point = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(4203840)))]; + tensor attn12_in_proj_weight_cast_fp16 = constexpr_affine_dequantize()[axis = tensor(0), name = tensor("attn12_in_proj_weight_cast_fp16"), quantized_data = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(151230720))), scale = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(154376512))), zero_point = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(3145856)))]; + tensor attn12_out_proj_weight_cast_fp16 = constexpr_affine_dequantize()[axis = tensor(0), name = tensor("attn12_out_proj_weight_cast_fp16"), quantized_data = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(154382720))), scale = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(155431360))), zero_point = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(4203840)))]; + tensor linear12_1_weight_cast_fp16 = constexpr_affine_dequantize()[axis = tensor(0), name = tensor("linear12_1_weight_cast_fp16"), quantized_data = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(155433472))), scale = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(159627840))), zero_point = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8401408)))]; + tensor linear12_2_weight_cast_fp16 = constexpr_affine_dequantize()[axis = tensor(0), name = tensor("linear12_2_weight_cast_fp16"), quantized_data = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(159636096))), scale = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(163830464))), zero_point = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(4203840)))]; + tensor attn13_in_proj_weight_cast_fp16 = constexpr_affine_dequantize()[axis = tensor(0), name = tensor("attn13_in_proj_weight_cast_fp16"), quantized_data = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(163832576))), scale = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(166978368))), zero_point = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(3145856)))]; + tensor attn13_out_proj_weight_cast_fp16 = constexpr_affine_dequantize()[axis = tensor(0), name = tensor("attn13_out_proj_weight_cast_fp16"), quantized_data = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(166984576))), scale = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(168033216))), zero_point = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(4203840)))]; + tensor linear13_1_weight_cast_fp16 = constexpr_affine_dequantize()[axis = tensor(0), name = tensor("linear13_1_weight_cast_fp16"), quantized_data = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(168035328))), scale = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(172229696))), zero_point = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8401408)))]; + tensor linear13_2_weight_cast_fp16 = constexpr_affine_dequantize()[axis = tensor(0), name = tensor("linear13_2_weight_cast_fp16"), quantized_data = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(172237952))), scale = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(176432320))), zero_point = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(4203840)))]; + tensor attn14_in_proj_weight_cast_fp16 = constexpr_affine_dequantize()[axis = tensor(0), name = tensor("attn14_in_proj_weight_cast_fp16"), quantized_data = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(176434432))), scale = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(179580224))), zero_point = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(3145856)))]; + tensor attn14_out_proj_weight_cast_fp16 = constexpr_affine_dequantize()[axis = tensor(0), name = tensor("attn14_out_proj_weight_cast_fp16"), quantized_data = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(179586432))), scale = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(180635072))), zero_point = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(4203840)))]; + tensor linear14_1_weight_cast_fp16 = constexpr_affine_dequantize()[axis = tensor(0), name = tensor("linear14_1_weight_cast_fp16"), quantized_data = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(180637184))), scale = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(184831552))), zero_point = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8401408)))]; + tensor linear14_2_weight_cast_fp16 = constexpr_affine_dequantize()[axis = tensor(0), name = tensor("linear14_2_weight_cast_fp16"), quantized_data = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(184839808))), scale = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(189034176))), zero_point = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(4203840)))]; + tensor attn15_in_proj_weight_cast_fp16 = constexpr_affine_dequantize()[axis = tensor(0), name = tensor("attn15_in_proj_weight_cast_fp16"), quantized_data = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(189036288))), scale = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(192182080))), zero_point = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(3145856)))]; + tensor attn15_out_proj_weight_cast_fp16 = constexpr_affine_dequantize()[axis = tensor(0), name = tensor("attn15_out_proj_weight_cast_fp16"), quantized_data = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(192188288))), scale = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(193236928))), zero_point = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(4203840)))]; + tensor linear15_1_weight_cast_fp16 = constexpr_affine_dequantize()[axis = tensor(0), name = tensor("linear15_1_weight_cast_fp16"), quantized_data = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(193239040))), scale = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(197433408))), zero_point = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8401408)))]; + tensor linear15_2_weight_cast_fp16 = constexpr_affine_dequantize()[axis = tensor(0), name = tensor("linear15_2_weight_cast_fp16"), quantized_data = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(197441664))), scale = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(201636032))), zero_point = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(4203840)))]; + tensor attn16_in_proj_weight_cast_fp16 = constexpr_affine_dequantize()[axis = tensor(0), name = tensor("attn16_in_proj_weight_cast_fp16"), quantized_data = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(201638144))), scale = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(204783936))), zero_point = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(3145856)))]; + tensor attn16_out_proj_weight_cast_fp16 = constexpr_affine_dequantize()[axis = tensor(0), name = tensor("attn16_out_proj_weight_cast_fp16"), quantized_data = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(204790144))), scale = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(205838784))), zero_point = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(4203840)))]; + tensor linear16_1_weight_cast_fp16 = constexpr_affine_dequantize()[axis = tensor(0), name = tensor("linear16_1_weight_cast_fp16"), quantized_data = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(205840896))), scale = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(210035264))), zero_point = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8401408)))]; + tensor linear16_2_weight_cast_fp16 = constexpr_affine_dequantize()[axis = tensor(0), name = tensor("linear16_2_weight_cast_fp16"), quantized_data = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(210043520))), scale = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(214237888))), zero_point = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(4203840)))]; + tensor attn17_in_proj_weight_cast_fp16 = constexpr_affine_dequantize()[axis = tensor(0), name = tensor("attn17_in_proj_weight_cast_fp16"), quantized_data = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(214240000))), scale = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(217385792))), zero_point = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(3145856)))]; + tensor attn17_out_proj_weight_cast_fp16 = constexpr_affine_dequantize()[axis = tensor(0), name = tensor("attn17_out_proj_weight_cast_fp16"), quantized_data = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(217392000))), scale = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(218440640))), zero_point = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(4203840)))]; + tensor linear17_1_weight_cast_fp16 = constexpr_affine_dequantize()[axis = tensor(0), name = tensor("linear17_1_weight_cast_fp16"), quantized_data = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(218442752))), scale = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(222637120))), zero_point = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8401408)))]; + tensor linear17_2_weight_cast_fp16 = constexpr_affine_dequantize()[axis = tensor(0), name = tensor("linear17_2_weight_cast_fp16"), quantized_data = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(222645376))), scale = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(226839744))), zero_point = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(4203840)))]; + tensor attn18_in_proj_weight_cast_fp16 = constexpr_affine_dequantize()[axis = tensor(0), name = tensor("attn18_in_proj_weight_cast_fp16"), quantized_data = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(226841856))), scale = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(229987648))), zero_point = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(3145856)))]; + tensor attn18_out_proj_weight_cast_fp16 = constexpr_affine_dequantize()[axis = tensor(0), name = tensor("attn18_out_proj_weight_cast_fp16"), quantized_data = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(229993856))), scale = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(231042496))), zero_point = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(4203840)))]; + tensor linear18_1_weight_cast_fp16 = constexpr_affine_dequantize()[axis = tensor(0), name = tensor("linear18_1_weight_cast_fp16"), quantized_data = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(231044608))), scale = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(235238976))), zero_point = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8401408)))]; + tensor linear18_2_weight_cast_fp16 = constexpr_affine_dequantize()[axis = tensor(0), name = tensor("linear18_2_weight_cast_fp16"), quantized_data = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(235247232))), scale = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(239441600))), zero_point = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(4203840)))]; + tensor attn19_in_proj_weight_cast_fp16 = constexpr_affine_dequantize()[axis = tensor(0), name = tensor("attn19_in_proj_weight_cast_fp16"), quantized_data = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(239443712))), scale = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(242589504))), zero_point = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(3145856)))]; + tensor attn19_out_proj_weight_cast_fp16 = constexpr_affine_dequantize()[axis = tensor(0), name = tensor("attn19_out_proj_weight_cast_fp16"), quantized_data = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(242595712))), scale = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(243644352))), zero_point = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(4203840)))]; + tensor linear19_1_weight_cast_fp16 = constexpr_affine_dequantize()[axis = tensor(0), name = tensor("linear19_1_weight_cast_fp16"), quantized_data = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(243646464))), scale = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(247840832))), zero_point = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8401408)))]; + tensor linear19_2_weight_cast_fp16 = constexpr_affine_dequantize()[axis = tensor(0), name = tensor("linear19_2_weight_cast_fp16"), quantized_data = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(247849088))), scale = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(252043456))), zero_point = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(4203840)))]; + tensor attn20_in_proj_weight_cast_fp16 = constexpr_affine_dequantize()[axis = tensor(0), name = tensor("attn20_in_proj_weight_cast_fp16"), quantized_data = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(252045568))), scale = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(255191360))), zero_point = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(3145856)))]; + tensor attn20_out_proj_weight_cast_fp16 = constexpr_affine_dequantize()[axis = tensor(0), name = tensor("attn20_out_proj_weight_cast_fp16"), quantized_data = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(255197568))), scale = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(256246208))), zero_point = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(4203840)))]; + tensor linear20_1_weight_cast_fp16 = constexpr_affine_dequantize()[axis = tensor(0), name = tensor("linear20_1_weight_cast_fp16"), quantized_data = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(256248320))), scale = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(260442688))), zero_point = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8401408)))]; + tensor linear20_2_weight_cast_fp16 = constexpr_affine_dequantize()[axis = tensor(0), name = tensor("linear20_2_weight_cast_fp16"), quantized_data = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(260450944))), scale = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(264645312))), zero_point = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(4203840)))]; + tensor attn21_in_proj_weight_cast_fp16 = constexpr_affine_dequantize()[axis = tensor(0), name = tensor("attn21_in_proj_weight_cast_fp16"), quantized_data = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(264647424))), scale = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(267793216))), zero_point = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(3145856)))]; + tensor attn21_out_proj_weight_cast_fp16 = constexpr_affine_dequantize()[axis = tensor(0), name = tensor("attn21_out_proj_weight_cast_fp16"), quantized_data = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(267799424))), scale = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(268848064))), zero_point = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(4203840)))]; + tensor linear21_1_weight_cast_fp16 = constexpr_affine_dequantize()[axis = tensor(0), name = tensor("linear21_1_weight_cast_fp16"), quantized_data = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(268850176))), scale = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(273044544))), zero_point = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8401408)))]; + tensor linear21_2_weight_cast_fp16 = constexpr_affine_dequantize()[axis = tensor(0), name = tensor("linear21_2_weight_cast_fp16"), quantized_data = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(273052800))), scale = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(277247168))), zero_point = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(4203840)))]; + tensor attn22_in_proj_weight_cast_fp16 = constexpr_affine_dequantize()[axis = tensor(0), name = tensor("attn22_in_proj_weight_cast_fp16"), quantized_data = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(277249280))), scale = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(280395072))), zero_point = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(3145856)))]; + tensor attn22_out_proj_weight_cast_fp16 = constexpr_affine_dequantize()[axis = tensor(0), name = tensor("attn22_out_proj_weight_cast_fp16"), quantized_data = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(280401280))), scale = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(281449920))), zero_point = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(4203840)))]; + tensor linear22_1_weight_cast_fp16 = constexpr_affine_dequantize()[axis = tensor(0), name = tensor("linear22_1_weight_cast_fp16"), quantized_data = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(281452032))), scale = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(285646400))), zero_point = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8401408)))]; + tensor linear22_2_weight_cast_fp16 = constexpr_affine_dequantize()[axis = tensor(0), name = tensor("linear22_2_weight_cast_fp16"), quantized_data = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(285654656))), scale = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(289849024))), zero_point = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(4203840)))]; + tensor attn23_in_proj_weight_cast_fp16 = constexpr_affine_dequantize()[axis = tensor(0), name = tensor("attn23_in_proj_weight_cast_fp16"), quantized_data = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(289851136))), scale = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(292996928))), zero_point = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(3145856)))]; + tensor attn23_out_proj_weight_cast_fp16 = constexpr_affine_dequantize()[axis = tensor(0), name = tensor("attn23_out_proj_weight_cast_fp16"), quantized_data = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(293003136))), scale = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(294051776))), zero_point = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(4203840)))]; + tensor linear23_1_weight_cast_fp16 = constexpr_affine_dequantize()[axis = tensor(0), name = tensor("linear23_1_weight_cast_fp16"), quantized_data = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(294053888))), scale = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(298248256))), zero_point = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8401408)))]; + tensor linear23_2_weight_cast_fp16 = constexpr_affine_dequantize()[axis = tensor(0), name = tensor("linear23_2_weight_cast_fp16"), quantized_data = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(298256512))), scale = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(302450880))), zero_point = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(4203840)))]; + tensor sequence_to_fp16_dtype_0 = const()[name = tensor("sequence_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor sequence_to_fp16 = cast(dtype = sequence_to_fp16_dtype_0, x = sequence)[name = tensor("cast_555")]; + tensor var_198_cast_fp16 = not_equal(x = sequence_to_fp16, y = sequence_to_fp16)[name = tensor("op_198_cast_fp16")]; + tensor expand_dims_0_axes_0 = const()[name = tensor("expand_dims_0_axes_0"), val = tensor([0, 1])]; + tensor bos_emb_to_fp16_dtype_0 = const()[name = tensor("bos_emb_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor bos_emb_to_fp16 = cast(dtype = bos_emb_to_fp16_dtype_0, x = bos_emb)[name = tensor("cast_554")]; + tensor expand_dims_0_cast_fp16 = expand_dims(axes = expand_dims_0_axes_0, x = bos_emb_to_fp16)[name = tensor("expand_dims_0_cast_fp16")]; + tensor input_1_cast_fp16 = select(a = expand_dims_0_cast_fp16, b = sequence_to_fp16, cond = var_198_cast_fp16)[name = tensor("input_1_cast_fp16")]; + tensor input_linear_weight_to_fp16 = const()[name = tensor("input_linear_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(302452992)))]; + tensor linear_0_bias_0_to_fp16 = const()[name = tensor("linear_0_bias_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(302518592)))]; + tensor linear_0_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = input_linear_weight_to_fp16, x = input_1_cast_fp16)[name = tensor("linear_0_cast_fp16")]; + tensor x_1_axes_0 = const()[name = tensor("x_1_axes_0"), val = tensor([-1])]; + tensor norm0_1_weight_to_fp16 = const()[name = tensor("norm0_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(302520704)))]; + tensor norm0_1_bias_to_fp16 = const()[name = tensor("norm0_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(302522816)))]; + tensor var_204_to_fp16 = const()[name = tensor("op_204_to_fp16"), val = tensor(0x1.5p-17)]; + tensor x_1_cast_fp16 = layer_norm(axes = x_1_axes_0, beta = norm0_1_bias_to_fp16, epsilon = var_204_to_fp16, gamma = norm0_1_weight_to_fp16, x = linear_0_cast_fp16)[name = tensor("x_1_cast_fp16")]; + tensor linear_1_bias_0_to_fp16 = const()[name = tensor("linear_1_bias_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(302524928)))]; + tensor linear_1_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = attn0_in_proj_weight_cast_fp16, x = x_1_cast_fp16)[name = tensor("linear_1_cast_fp16")]; + tensor var_240 = const()[name = tensor("op_240"), val = tensor([1, 1, 3, 16, 64])]; + tensor qkv_1_cast_fp16 = reshape(shape = var_240, x = linear_1_cast_fp16)[name = tensor("qkv_1_cast_fp16")]; + tensor q_1_begin_0 = const()[name = tensor("q_1_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor q_1_end_0 = const()[name = tensor("q_1_end_0"), val = tensor([1, 1, 1, 16, 64])]; + tensor q_1_end_mask_0 = const()[name = tensor("q_1_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor q_1_squeeze_mask_0 = const()[name = tensor("q_1_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor q_1_cast_fp16 = slice_by_index(begin = q_1_begin_0, end = q_1_end_0, end_mask = q_1_end_mask_0, squeeze_mask = q_1_squeeze_mask_0, x = qkv_1_cast_fp16)[name = tensor("q_1_cast_fp16")]; + tensor k_1_begin_0 = const()[name = tensor("k_1_begin_0"), val = tensor([0, 0, 1, 0, 0])]; + tensor k_1_end_0 = const()[name = tensor("k_1_end_0"), val = tensor([1, 1, 2, 16, 64])]; + tensor k_1_end_mask_0 = const()[name = tensor("k_1_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor k_1_squeeze_mask_0 = const()[name = tensor("k_1_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor k_1_cast_fp16 = slice_by_index(begin = k_1_begin_0, end = k_1_end_0, end_mask = k_1_end_mask_0, squeeze_mask = k_1_squeeze_mask_0, x = qkv_1_cast_fp16)[name = tensor("k_1_cast_fp16")]; + tensor v_1_begin_0 = const()[name = tensor("v_1_begin_0"), val = tensor([0, 0, 2, 0, 0])]; + tensor v_1_end_0 = const()[name = tensor("v_1_end_0"), val = tensor([1, 1, 3, 16, 64])]; + tensor v_1_end_mask_0 = const()[name = tensor("v_1_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor v_1_squeeze_mask_0 = const()[name = tensor("v_1_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor v_1_cast_fp16 = slice_by_index(begin = v_1_begin_0, end = v_1_end_0, end_mask = v_1_end_mask_0, squeeze_mask = v_1_squeeze_mask_0, x = qkv_1_cast_fp16)[name = tensor("v_1_cast_fp16")]; + tensor var_344 = const()[name = tensor("op_344"), val = tensor([1, 1, 1, 1])]; + tensor position0_to_fp16_dtype_0 = const()[name = tensor("position0_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor position0_to_fp16 = cast(dtype = position0_to_fp16_dtype_0, x = position0)[name = tensor("cast_553")]; + tensor ts_5_cast_fp16 = reshape(shape = var_344, x = position0_to_fp16)[name = tensor("ts_5_cast_fp16")]; + tensor var_348 = const()[name = tensor("op_348"), val = tensor([1, 1, 16, 32, 2])]; + tensor q_complex_1_cast_fp16 = reshape(shape = var_348, x = q_1_cast_fp16)[name = tensor("q_complex_1_cast_fp16")]; + tensor var_352 = const()[name = tensor("op_352"), val = tensor([1, 1, 16, 32, 2])]; + tensor k_complex_1_cast_fp16 = reshape(shape = var_352, x = k_1_cast_fp16)[name = tensor("k_complex_1_cast_fp16")]; + tensor var_356_begin_0 = const()[name = tensor("op_356_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_356_end_0 = const()[name = tensor("op_356_end_0"), val = tensor([1, 1, 16, 32, 1])]; + tensor var_356_end_mask_0 = const()[name = tensor("op_356_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_356_squeeze_mask_0 = const()[name = tensor("op_356_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_356_cast_fp16 = slice_by_index(begin = var_356_begin_0, end = var_356_end_0, end_mask = var_356_end_mask_0, squeeze_mask = var_356_squeeze_mask_0, x = q_complex_1_cast_fp16)[name = tensor("op_356_cast_fp16")]; + tensor var_364_begin_0 = const()[name = tensor("op_364_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_364_end_0 = const()[name = tensor("op_364_end_0"), val = tensor([1, 1, 16, 32, 2])]; + tensor var_364_end_mask_0 = const()[name = tensor("op_364_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_364_squeeze_mask_0 = const()[name = tensor("op_364_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_364_cast_fp16 = slice_by_index(begin = var_364_begin_0, end = var_364_end_0, end_mask = var_364_end_mask_0, squeeze_mask = var_364_squeeze_mask_0, x = q_complex_1_cast_fp16)[name = tensor("op_364_cast_fp16")]; + tensor var_372_begin_0 = const()[name = tensor("op_372_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_372_end_0 = const()[name = tensor("op_372_end_0"), val = tensor([1, 1, 16, 32, 1])]; + tensor var_372_end_mask_0 = const()[name = tensor("op_372_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_372_squeeze_mask_0 = const()[name = tensor("op_372_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_372_cast_fp16 = slice_by_index(begin = var_372_begin_0, end = var_372_end_0, end_mask = var_372_end_mask_0, squeeze_mask = var_372_squeeze_mask_0, x = k_complex_1_cast_fp16)[name = tensor("op_372_cast_fp16")]; + tensor var_380_begin_0 = const()[name = tensor("op_380_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_380_end_0 = const()[name = tensor("op_380_end_0"), val = tensor([1, 1, 16, 32, 2])]; + tensor var_380_end_mask_0 = const()[name = tensor("op_380_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_380_squeeze_mask_0 = const()[name = tensor("op_380_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_380_cast_fp16 = slice_by_index(begin = var_380_begin_0, end = var_380_end_0, end_mask = var_380_end_mask_0, squeeze_mask = var_380_squeeze_mask_0, x = k_complex_1_cast_fp16)[name = tensor("op_380_cast_fp16")]; + tensor freqs_1_to_fp16 = const()[name = tensor("freqs_1_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(302531136)))]; + tensor var_386_cast_fp16 = mul(x = freqs_1_to_fp16, y = ts_5_cast_fp16)[name = tensor("op_386_cast_fp16")]; + tensor rotr_1_cast_fp16 = cos(x = var_386_cast_fp16)[name = tensor("rotr_1_cast_fp16")]; + tensor roti_1_cast_fp16 = sin(x = var_386_cast_fp16)[name = tensor("roti_1_cast_fp16")]; + tensor var_390_cast_fp16 = mul(x = var_356_cast_fp16, y = rotr_1_cast_fp16)[name = tensor("op_390_cast_fp16")]; + tensor var_391_cast_fp16 = mul(x = var_364_cast_fp16, y = roti_1_cast_fp16)[name = tensor("op_391_cast_fp16")]; + tensor qor_1_cast_fp16 = sub(x = var_390_cast_fp16, y = var_391_cast_fp16)[name = tensor("qor_1_cast_fp16")]; + tensor var_394_cast_fp16 = mul(x = var_356_cast_fp16, y = roti_1_cast_fp16)[name = tensor("op_394_cast_fp16")]; + tensor var_395_cast_fp16 = mul(x = var_364_cast_fp16, y = rotr_1_cast_fp16)[name = tensor("op_395_cast_fp16")]; + tensor qoi_1_cast_fp16 = add(x = var_394_cast_fp16, y = var_395_cast_fp16)[name = tensor("qoi_1_cast_fp16")]; + tensor var_398_cast_fp16 = mul(x = var_372_cast_fp16, y = rotr_1_cast_fp16)[name = tensor("op_398_cast_fp16")]; + tensor var_399_cast_fp16 = mul(x = var_380_cast_fp16, y = roti_1_cast_fp16)[name = tensor("op_399_cast_fp16")]; + tensor kor_1_cast_fp16 = sub(x = var_398_cast_fp16, y = var_399_cast_fp16)[name = tensor("kor_1_cast_fp16")]; + tensor var_402_cast_fp16 = mul(x = var_372_cast_fp16, y = roti_1_cast_fp16)[name = tensor("op_402_cast_fp16")]; + tensor var_403_cast_fp16 = mul(x = var_380_cast_fp16, y = rotr_1_cast_fp16)[name = tensor("op_403_cast_fp16")]; + tensor koi_1_cast_fp16 = add(x = var_402_cast_fp16, y = var_403_cast_fp16)[name = tensor("koi_1_cast_fp16")]; + tensor qo_1_axis_0 = const()[name = tensor("qo_1_axis_0"), val = tensor(-1)]; + tensor qo_1_cast_fp16 = stack(axis = qo_1_axis_0, values = (qor_1_cast_fp16, qoi_1_cast_fp16))[name = tensor("qo_1_cast_fp16")]; + tensor ko_1_axis_0 = const()[name = tensor("ko_1_axis_0"), val = tensor(-1)]; + tensor ko_1_cast_fp16 = stack(axis = ko_1_axis_0, values = (kor_1_cast_fp16, koi_1_cast_fp16))[name = tensor("ko_1_cast_fp16")]; + tensor var_432 = const()[name = tensor("op_432"), val = tensor([1, 1, 16, 64])]; + tensor q_3_cast_fp16 = reshape(shape = var_432, x = qo_1_cast_fp16)[name = tensor("q_3_cast_fp16")]; + tensor var_434 = const()[name = tensor("op_434"), val = tensor([1, 1, 16, 64])]; + tensor k_3_cast_fp16 = reshape(shape = var_434, x = ko_1_cast_fp16)[name = tensor("k_3_cast_fp16")]; + tensor _inversed_456_y_0_to_fp16 = const()[name = tensor("_inversed_456_y_0_to_fp16"), val = tensor(0x1p-9)]; + tensor _inversed_456_cast_fp16 = mul(x = ts_5_cast_fp16, y = _inversed_456_y_0_to_fp16)[name = tensor("_inversed_456_cast_fp16")]; + tensor var_457_cast_fp16 = floor(x = _inversed_456_cast_fp16)[name = tensor("op_457_cast_fp16")]; + tensor var_458_to_fp16 = const()[name = tensor("op_458_to_fp16"), val = tensor(0x1p+9)]; + tensor var_459_cast_fp16 = mul(x = var_457_cast_fp16, y = var_458_to_fp16)[name = tensor("op_459_cast_fp16")]; + tensor write_indices_float_3_cast_fp16 = sub(x = ts_5_cast_fp16, y = var_459_cast_fp16)[name = tensor("write_indices_float_3_cast_fp16")]; + tensor var_466_dtype_0 = const()[name = tensor("op_466_dtype_0"), val = tensor("int32")]; + tensor write_indices_1_reps_0 = const()[name = tensor("write_indices_1_reps_0"), val = tensor([1, 1, 16, 64])]; + tensor write_indices_float_3_cast_fp16_to_int32 = cast(dtype = var_466_dtype_0, x = write_indices_float_3_cast_fp16)[name = tensor("cast_552")]; + tensor write_indices_1 = tile(reps = write_indices_1_reps_0, x = write_indices_float_3_cast_fp16_to_int32)[name = tensor("write_indices_1")]; + tensor var_474_begin_0 = const()[name = tensor("op_474_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_474_end_0 = const()[name = tensor("op_474_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor var_474_end_mask_0 = const()[name = tensor("op_474_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_474_squeeze_mask_0 = const()[name = tensor("op_474_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor cache0_to_fp16_dtype_0 = const()[name = tensor("cache0_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor cache0_to_fp16 = cast(dtype = cache0_to_fp16_dtype_0, x = cache0)[name = tensor("cast_551")]; + tensor var_474_cast_fp16 = slice_by_index(begin = var_474_begin_0, end = var_474_end_0, end_mask = var_474_end_mask_0, squeeze_mask = var_474_squeeze_mask_0, x = cache0_to_fp16)[name = tensor("op_474_cast_fp16")]; + tensor var_476_axis_0 = const()[name = tensor("op_476_axis_0"), val = tensor(1)]; + tensor var_476_mode_0 = const()[name = tensor("op_476_mode_0"), val = tensor("update")]; + tensor var_476_validate_indices_0 = const()[name = tensor("op_476_validate_indices_0"), val = tensor(false)]; + tensor var_476_cast_fp16 = scatter_along_axis(axis = var_476_axis_0, data = var_474_cast_fp16, indices = write_indices_1, mode = var_476_mode_0, updates = k_3_cast_fp16, validate_indices = var_476_validate_indices_0)[name = tensor("op_476_cast_fp16")]; + tensor concat_2 = const()[name = tensor("concat_2"), val = tensor([0, 0, 0, 0, 0])]; + tensor concat_3 = const()[name = tensor("concat_3"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_1_internal_tensor_assign_1_stride_0 = const()[name = tensor("new_cache_1_internal_tensor_assign_1_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_1_internal_tensor_assign_1_begin_mask_0 = const()[name = tensor("new_cache_1_internal_tensor_assign_1_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_1_internal_tensor_assign_1_end_mask_0 = const()[name = tensor("new_cache_1_internal_tensor_assign_1_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_1_internal_tensor_assign_1_squeeze_mask_0 = const()[name = tensor("new_cache_1_internal_tensor_assign_1_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_48 = const()[name = tensor("shape_48"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_0 = const()[name = tensor("reduce_prod_0"), val = tensor(1048576)]; + tensor range_1d_0_start_0 = const()[name = tensor("range_1d_0_start_0"), val = tensor(0)]; + tensor range_1d_0_step_0 = const()[name = tensor("range_1d_0_step_0"), val = tensor(1)]; + tensor range_1d_0 = range_1d(end = reduce_prod_0, start = range_1d_0_start_0, step = range_1d_0_step_0)[name = tensor("range_1d_0")]; + tensor reshape_0 = reshape(shape = shape_48, x = range_1d_0)[name = tensor("reshape_0")]; + tensor slice_by_index_0 = slice_by_index(begin = concat_2, begin_mask = new_cache_1_internal_tensor_assign_1_begin_mask_0, end = concat_3, end_mask = new_cache_1_internal_tensor_assign_1_end_mask_0, squeeze_mask = new_cache_1_internal_tensor_assign_1_squeeze_mask_0, stride = new_cache_1_internal_tensor_assign_1_stride_0, x = reshape_0)[name = tensor("slice_by_index_0")]; + tensor reshape_1_shape_0 = const()[name = tensor("reshape_1_shape_0"), val = tensor([-1])]; + tensor reshape_1 = reshape(shape = reshape_1_shape_0, x = slice_by_index_0)[name = tensor("reshape_1")]; + tensor reshape_2_shape_0 = const()[name = tensor("reshape_2_shape_0"), val = tensor([-1])]; + tensor reshape_2_cast_fp16 = reshape(shape = reshape_2_shape_0, x = var_476_cast_fp16)[name = tensor("reshape_2_cast_fp16")]; + tensor reshape_3_shape_0 = const()[name = tensor("reshape_3_shape_0"), val = tensor([-1])]; + tensor reshape_3_cast_fp16 = reshape(shape = reshape_3_shape_0, x = cache0_to_fp16)[name = tensor("reshape_3_cast_fp16")]; + tensor scatter_0_mode_0 = const()[name = tensor("scatter_0_mode_0"), val = tensor("update")]; + tensor scatter_0_axis_0 = const()[name = tensor("scatter_0_axis_0"), val = tensor(0)]; + tensor scatter_0_validate_indices_0 = const()[name = tensor("scatter_0_validate_indices_0"), val = tensor(false)]; + tensor scatter_0_cast_fp16 = scatter(axis = scatter_0_axis_0, data = reshape_3_cast_fp16, indices = reshape_1, mode = scatter_0_mode_0, updates = reshape_2_cast_fp16, validate_indices = scatter_0_validate_indices_0)[name = tensor("scatter_0_cast_fp16")]; + tensor reshape_4_cast_fp16 = reshape(shape = shape_48, x = scatter_0_cast_fp16)[name = tensor("reshape_4_cast_fp16")]; + tensor var_484_begin_0 = const()[name = tensor("op_484_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor var_484_end_0 = const()[name = tensor("op_484_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor var_484_end_mask_0 = const()[name = tensor("op_484_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_484_squeeze_mask_0 = const()[name = tensor("op_484_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor var_484_cast_fp16 = slice_by_index(begin = var_484_begin_0, end = var_484_end_0, end_mask = var_484_end_mask_0, squeeze_mask = var_484_squeeze_mask_0, x = reshape_4_cast_fp16)[name = tensor("op_484_cast_fp16")]; + tensor var_486_axis_0 = const()[name = tensor("op_486_axis_0"), val = tensor(1)]; + tensor var_486_mode_0 = const()[name = tensor("op_486_mode_0"), val = tensor("update")]; + tensor var_486_validate_indices_0 = const()[name = tensor("op_486_validate_indices_0"), val = tensor(false)]; + tensor var_486_cast_fp16 = scatter_along_axis(axis = var_486_axis_0, data = var_484_cast_fp16, indices = write_indices_1, mode = var_486_mode_0, updates = v_1_cast_fp16, validate_indices = var_486_validate_indices_0)[name = tensor("op_486_cast_fp16")]; + tensor concat_4 = const()[name = tensor("concat_4"), val = tensor([1, 0, 0, 0, 0])]; + tensor concat_5 = const()[name = tensor("concat_5"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_1_internal_tensor_assign_2_stride_0 = const()[name = tensor("new_cache_1_internal_tensor_assign_2_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_1_internal_tensor_assign_2_begin_mask_0 = const()[name = tensor("new_cache_1_internal_tensor_assign_2_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_1_internal_tensor_assign_2_end_mask_0 = const()[name = tensor("new_cache_1_internal_tensor_assign_2_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_1_internal_tensor_assign_2_squeeze_mask_0 = const()[name = tensor("new_cache_1_internal_tensor_assign_2_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_49 = const()[name = tensor("shape_49"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_1 = const()[name = tensor("reduce_prod_1"), val = tensor(1048576)]; + tensor range_1d_1_start_0 = const()[name = tensor("range_1d_1_start_0"), val = tensor(0)]; + tensor range_1d_1_step_0 = const()[name = tensor("range_1d_1_step_0"), val = tensor(1)]; + tensor range_1d_1 = range_1d(end = reduce_prod_1, start = range_1d_1_start_0, step = range_1d_1_step_0)[name = tensor("range_1d_1")]; + tensor reshape_5 = reshape(shape = shape_49, x = range_1d_1)[name = tensor("reshape_5")]; + tensor slice_by_index_1 = slice_by_index(begin = concat_4, begin_mask = new_cache_1_internal_tensor_assign_2_begin_mask_0, end = concat_5, end_mask = new_cache_1_internal_tensor_assign_2_end_mask_0, squeeze_mask = new_cache_1_internal_tensor_assign_2_squeeze_mask_0, stride = new_cache_1_internal_tensor_assign_2_stride_0, x = reshape_5)[name = tensor("slice_by_index_1")]; + tensor reshape_6_shape_0 = const()[name = tensor("reshape_6_shape_0"), val = tensor([-1])]; + tensor reshape_6 = reshape(shape = reshape_6_shape_0, x = slice_by_index_1)[name = tensor("reshape_6")]; + tensor reshape_7_shape_0 = const()[name = tensor("reshape_7_shape_0"), val = tensor([-1])]; + tensor reshape_7_cast_fp16 = reshape(shape = reshape_7_shape_0, x = var_486_cast_fp16)[name = tensor("reshape_7_cast_fp16")]; + tensor reshape_8_shape_0 = const()[name = tensor("reshape_8_shape_0"), val = tensor([-1])]; + tensor reshape_8_cast_fp16 = reshape(shape = reshape_8_shape_0, x = reshape_4_cast_fp16)[name = tensor("reshape_8_cast_fp16")]; + tensor scatter_1_mode_0 = const()[name = tensor("scatter_1_mode_0"), val = tensor("update")]; + tensor scatter_1_axis_0 = const()[name = tensor("scatter_1_axis_0"), val = tensor(0)]; + tensor scatter_1_validate_indices_0 = const()[name = tensor("scatter_1_validate_indices_0"), val = tensor(false)]; + tensor scatter_1_cast_fp16 = scatter(axis = scatter_1_axis_0, data = reshape_8_cast_fp16, indices = reshape_6, mode = scatter_1_mode_0, updates = reshape_7_cast_fp16, validate_indices = scatter_1_validate_indices_0)[name = tensor("scatter_1_cast_fp16")]; + tensor reshape_9_cast_fp16 = reshape(shape = shape_49, x = scatter_1_cast_fp16)[name = tensor("reshape_9_cast_fp16")]; + tensor reshape_9_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("reshape_9_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor keys_1_begin_0 = const()[name = tensor("keys_1_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor keys_1_end_0 = const()[name = tensor("keys_1_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor keys_1_end_mask_0 = const()[name = tensor("keys_1_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor keys_1_squeeze_mask_0 = const()[name = tensor("keys_1_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor keys_1_cast_fp16 = slice_by_index(begin = keys_1_begin_0, end = keys_1_end_0, end_mask = keys_1_end_mask_0, squeeze_mask = keys_1_squeeze_mask_0, x = reshape_9_cast_fp16)[name = tensor("keys_1_cast_fp16")]; + tensor values_1_begin_0 = const()[name = tensor("values_1_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor values_1_end_0 = const()[name = tensor("values_1_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor values_1_end_mask_0 = const()[name = tensor("values_1_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor values_1_squeeze_mask_0 = const()[name = tensor("values_1_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor values_1_cast_fp16 = slice_by_index(begin = values_1_begin_0, end = values_1_end_0, end_mask = values_1_end_mask_0, squeeze_mask = values_1_squeeze_mask_0, x = reshape_9_cast_fp16)[name = tensor("values_1_cast_fp16")]; + tensor var_498_cast_fp16 = not_equal(x = keys_1_cast_fp16, y = keys_1_cast_fp16)[name = tensor("op_498_cast_fp16")]; + tensor var_504_to_fp16 = const()[name = tensor("op_504_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(302531264)))]; + tensor keys_3_cast_fp16 = select(a = var_504_to_fp16, b = keys_1_cast_fp16, cond = var_498_cast_fp16)[name = tensor("keys_3_cast_fp16")]; + tensor var_506_cast_fp16 = not_equal(x = values_1_cast_fp16, y = values_1_cast_fp16)[name = tensor("op_506_cast_fp16")]; + tensor values_3_cast_fp16 = select(a = var_504_to_fp16, b = values_1_cast_fp16, cond = var_506_cast_fp16)[name = tensor("values_3_cast_fp16")]; + tensor var_530 = const()[name = tensor("op_530"), val = tensor([0, 2, 1, 3])]; + tensor var_543 = const()[name = tensor("op_543"), val = tensor([1, 1, 1])]; + tensor var_544_cast_fp16 = reshape(shape = var_543, x = position0_to_fp16)[name = tensor("op_544_cast_fp16")]; + tensor var_561_to_fp16 = const()[name = tensor("op_561_to_fp16"), val = tensor(0x1p+0)]; + tensor valid_len_1_cast_fp16 = add(x = var_544_cast_fp16, y = var_561_to_fp16)[name = tensor("valid_len_1_cast_fp16")]; + tensor k_positions_1_promoted_to_fp16 = const()[name = tensor("k_positions_1_promoted_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(303579904)))]; + tensor valid_mask_1_cast_fp16 = less(x = k_positions_1_promoted_to_fp16, y = valid_len_1_cast_fp16)[name = tensor("valid_mask_1_cast_fp16")]; + tensor causal_mask_1_cast_fp16 = less_equal(x = k_positions_1_promoted_to_fp16, y = var_544_cast_fp16)[name = tensor("causal_mask_1_cast_fp16")]; + tensor attn_mask_1 = logical_and(x = valid_mask_1_cast_fp16, y = causal_mask_1_cast_fp16)[name = tensor("attn_mask_1")]; + tensor attn_mask_3_axes_0 = const()[name = tensor("attn_mask_3_axes_0"), val = tensor([1])]; + tensor attn_mask_3 = expand_dims(axes = attn_mask_3_axes_0, x = attn_mask_1)[name = tensor("attn_mask_3")]; + tensor var_579_transpose_x_0 = const()[name = tensor("op_579_transpose_x_0"), val = tensor(false)]; + tensor var_579_transpose_y_0 = const()[name = tensor("op_579_transpose_y_0"), val = tensor(false)]; + tensor transpose_72_perm_0 = const()[name = tensor("transpose_72_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_73_perm_0 = const()[name = tensor("transpose_73_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_73 = transpose(perm = transpose_73_perm_0, x = keys_3_cast_fp16)[name = tensor("transpose_213")]; + tensor transpose_72 = transpose(perm = transpose_72_perm_0, x = q_3_cast_fp16)[name = tensor("transpose_214")]; + tensor var_579_cast_fp16 = matmul(transpose_x = var_579_transpose_x_0, transpose_y = var_579_transpose_y_0, x = transpose_72, y = transpose_73)[name = tensor("op_579_cast_fp16")]; + tensor var_573_to_fp16 = const()[name = tensor("op_573_to_fp16"), val = tensor([0x1p-3])]; + tensor attn_weights_1_cast_fp16 = mul(x = var_579_cast_fp16, y = var_573_to_fp16)[name = tensor("attn_weights_1_cast_fp16")]; + tensor var_581 = logical_not(x = attn_mask_3)[name = tensor("op_581")]; + tensor var_582_to_fp16 = const()[name = tensor("op_582_to_fp16"), val = tensor(-inf)]; + tensor attn_weights_3_cast_fp16 = select(a = var_582_to_fp16, b = attn_weights_1_cast_fp16, cond = var_581)[name = tensor("attn_weights_3_cast_fp16")]; + tensor var_584 = const()[name = tensor("op_584"), val = tensor(-1)]; + tensor attn_weights_5_cast_fp16 = softmax(axis = var_584, x = attn_weights_3_cast_fp16)[name = tensor("attn_weights_5_cast_fp16")]; + tensor attn_output_1_transpose_x_0 = const()[name = tensor("attn_output_1_transpose_x_0"), val = tensor(false)]; + tensor attn_output_1_transpose_y_0 = const()[name = tensor("attn_output_1_transpose_y_0"), val = tensor(false)]; + tensor values_5_cast_fp16 = transpose(perm = var_530, x = values_3_cast_fp16)[name = tensor("transpose_215")]; + tensor attn_output_1_cast_fp16 = matmul(transpose_x = attn_output_1_transpose_x_0, transpose_y = attn_output_1_transpose_y_0, x = attn_weights_5_cast_fp16, y = values_5_cast_fp16)[name = tensor("attn_output_1_cast_fp16")]; + tensor var_592 = const()[name = tensor("op_592"), val = tensor([0, 2, 1, 3])]; + tensor var_595 = const()[name = tensor("op_595"), val = tensor([1, 1, 1024])]; + tensor var_593_cast_fp16 = transpose(perm = var_592, x = attn_output_1_cast_fp16)[name = tensor("transpose_212")]; + tensor input_5_cast_fp16 = reshape(shape = var_595, x = var_593_cast_fp16)[name = tensor("input_5_cast_fp16")]; + tensor linear_2_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = attn0_out_proj_weight_cast_fp16, x = input_5_cast_fp16)[name = tensor("linear_2_cast_fp16")]; + tensor var_601_to_fp16 = const()[name = tensor("op_601_to_fp16"), val = tensor(0x1p+0)]; + tensor var_602_cast_fp16 = add(x = position0_to_fp16, y = var_601_to_fp16)[name = tensor("op_602_cast_fp16")]; + tensor var_602_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_602_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor input_7_cast_fp16 = add(x = linear_0_cast_fp16, y = linear_2_cast_fp16)[name = tensor("input_7_cast_fp16")]; + tensor input_9_axes_0 = const()[name = tensor("input_9_axes_0"), val = tensor([-1])]; + tensor norm0_2_weight_to_fp16 = const()[name = tensor("norm0_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(303580992)))]; + tensor norm0_2_bias_to_fp16 = const()[name = tensor("norm0_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(303583104)))]; + tensor var_606_to_fp16 = const()[name = tensor("op_606_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_9_cast_fp16 = layer_norm(axes = input_9_axes_0, beta = norm0_2_bias_to_fp16, epsilon = var_606_to_fp16, gamma = norm0_2_weight_to_fp16, x = input_7_cast_fp16)[name = tensor("input_9_cast_fp16")]; + tensor linear_3_bias_0_to_fp16 = const()[name = tensor("linear_3_bias_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(303585216)))]; + tensor linear_3_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = linear0_1_weight_cast_fp16, x = input_9_cast_fp16)[name = tensor("linear_3_cast_fp16")]; + tensor input_11_mode_0 = const()[name = tensor("input_11_mode_0"), val = tensor("EXACT")]; + tensor input_11_cast_fp16 = gelu(mode = input_11_mode_0, x = linear_3_cast_fp16)[name = tensor("input_11_cast_fp16")]; + tensor linear_4_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = linear0_2_weight_cast_fp16, x = input_11_cast_fp16)[name = tensor("linear_4_cast_fp16")]; + tensor input_13_cast_fp16 = add(x = input_7_cast_fp16, y = linear_4_cast_fp16)[name = tensor("input_13_cast_fp16")]; + tensor x_3_axes_0 = const()[name = tensor("x_3_axes_0"), val = tensor([-1])]; + tensor norm1_1_weight_to_fp16 = const()[name = tensor("norm1_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(303593472)))]; + tensor norm1_1_bias_to_fp16 = const()[name = tensor("norm1_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(303595584)))]; + tensor var_623_to_fp16 = const()[name = tensor("op_623_to_fp16"), val = tensor(0x1.5p-17)]; + tensor x_3_cast_fp16 = layer_norm(axes = x_3_axes_0, beta = norm1_1_bias_to_fp16, epsilon = var_623_to_fp16, gamma = norm1_1_weight_to_fp16, x = input_13_cast_fp16)[name = tensor("x_3_cast_fp16")]; + tensor linear_5_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = attn1_in_proj_weight_cast_fp16, x = x_3_cast_fp16)[name = tensor("linear_5_cast_fp16")]; + tensor var_659 = const()[name = tensor("op_659"), val = tensor([1, 1, 3, 16, 64])]; + tensor qkv_3_cast_fp16 = reshape(shape = var_659, x = linear_5_cast_fp16)[name = tensor("qkv_3_cast_fp16")]; + tensor q_7_begin_0 = const()[name = tensor("q_7_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor q_7_end_0 = const()[name = tensor("q_7_end_0"), val = tensor([1, 1, 1, 16, 64])]; + tensor q_7_end_mask_0 = const()[name = tensor("q_7_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor q_7_squeeze_mask_0 = const()[name = tensor("q_7_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor q_7_cast_fp16 = slice_by_index(begin = q_7_begin_0, end = q_7_end_0, end_mask = q_7_end_mask_0, squeeze_mask = q_7_squeeze_mask_0, x = qkv_3_cast_fp16)[name = tensor("q_7_cast_fp16")]; + tensor k_5_begin_0 = const()[name = tensor("k_5_begin_0"), val = tensor([0, 0, 1, 0, 0])]; + tensor k_5_end_0 = const()[name = tensor("k_5_end_0"), val = tensor([1, 1, 2, 16, 64])]; + tensor k_5_end_mask_0 = const()[name = tensor("k_5_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor k_5_squeeze_mask_0 = const()[name = tensor("k_5_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor k_5_cast_fp16 = slice_by_index(begin = k_5_begin_0, end = k_5_end_0, end_mask = k_5_end_mask_0, squeeze_mask = k_5_squeeze_mask_0, x = qkv_3_cast_fp16)[name = tensor("k_5_cast_fp16")]; + tensor v_3_begin_0 = const()[name = tensor("v_3_begin_0"), val = tensor([0, 0, 2, 0, 0])]; + tensor v_3_end_0 = const()[name = tensor("v_3_end_0"), val = tensor([1, 1, 3, 16, 64])]; + tensor v_3_end_mask_0 = const()[name = tensor("v_3_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor v_3_squeeze_mask_0 = const()[name = tensor("v_3_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor v_3_cast_fp16 = slice_by_index(begin = v_3_begin_0, end = v_3_end_0, end_mask = v_3_end_mask_0, squeeze_mask = v_3_squeeze_mask_0, x = qkv_3_cast_fp16)[name = tensor("v_3_cast_fp16")]; + tensor var_763 = const()[name = tensor("op_763"), val = tensor([1, 1, 1, 1])]; + tensor position1_to_fp16_dtype_0 = const()[name = tensor("position1_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor position1_to_fp16 = cast(dtype = position1_to_fp16_dtype_0, x = position1)[name = tensor("cast_548")]; + tensor ts_11_cast_fp16 = reshape(shape = var_763, x = position1_to_fp16)[name = tensor("ts_11_cast_fp16")]; + tensor var_767 = const()[name = tensor("op_767"), val = tensor([1, 1, 16, 32, 2])]; + tensor q_complex_3_cast_fp16 = reshape(shape = var_767, x = q_7_cast_fp16)[name = tensor("q_complex_3_cast_fp16")]; + tensor var_771 = const()[name = tensor("op_771"), val = tensor([1, 1, 16, 32, 2])]; + tensor k_complex_3_cast_fp16 = reshape(shape = var_771, x = k_5_cast_fp16)[name = tensor("k_complex_3_cast_fp16")]; + tensor var_775_begin_0 = const()[name = tensor("op_775_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_775_end_0 = const()[name = tensor("op_775_end_0"), val = tensor([1, 1, 16, 32, 1])]; + tensor var_775_end_mask_0 = const()[name = tensor("op_775_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_775_squeeze_mask_0 = const()[name = tensor("op_775_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_775_cast_fp16 = slice_by_index(begin = var_775_begin_0, end = var_775_end_0, end_mask = var_775_end_mask_0, squeeze_mask = var_775_squeeze_mask_0, x = q_complex_3_cast_fp16)[name = tensor("op_775_cast_fp16")]; + tensor var_783_begin_0 = const()[name = tensor("op_783_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_783_end_0 = const()[name = tensor("op_783_end_0"), val = tensor([1, 1, 16, 32, 2])]; + tensor var_783_end_mask_0 = const()[name = tensor("op_783_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_783_squeeze_mask_0 = const()[name = tensor("op_783_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_783_cast_fp16 = slice_by_index(begin = var_783_begin_0, end = var_783_end_0, end_mask = var_783_end_mask_0, squeeze_mask = var_783_squeeze_mask_0, x = q_complex_3_cast_fp16)[name = tensor("op_783_cast_fp16")]; + tensor var_791_begin_0 = const()[name = tensor("op_791_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_791_end_0 = const()[name = tensor("op_791_end_0"), val = tensor([1, 1, 16, 32, 1])]; + tensor var_791_end_mask_0 = const()[name = tensor("op_791_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_791_squeeze_mask_0 = const()[name = tensor("op_791_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_791_cast_fp16 = slice_by_index(begin = var_791_begin_0, end = var_791_end_0, end_mask = var_791_end_mask_0, squeeze_mask = var_791_squeeze_mask_0, x = k_complex_3_cast_fp16)[name = tensor("op_791_cast_fp16")]; + tensor var_799_begin_0 = const()[name = tensor("op_799_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_799_end_0 = const()[name = tensor("op_799_end_0"), val = tensor([1, 1, 16, 32, 2])]; + tensor var_799_end_mask_0 = const()[name = tensor("op_799_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_799_squeeze_mask_0 = const()[name = tensor("op_799_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_799_cast_fp16 = slice_by_index(begin = var_799_begin_0, end = var_799_end_0, end_mask = var_799_end_mask_0, squeeze_mask = var_799_squeeze_mask_0, x = k_complex_3_cast_fp16)[name = tensor("op_799_cast_fp16")]; + tensor freqs_3_to_fp16 = const()[name = tensor("freqs_3_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(303597696)))]; + tensor var_805_cast_fp16 = mul(x = freqs_3_to_fp16, y = ts_11_cast_fp16)[name = tensor("op_805_cast_fp16")]; + tensor rotr_3_cast_fp16 = cos(x = var_805_cast_fp16)[name = tensor("rotr_3_cast_fp16")]; + tensor roti_3_cast_fp16 = sin(x = var_805_cast_fp16)[name = tensor("roti_3_cast_fp16")]; + tensor var_809_cast_fp16 = mul(x = var_775_cast_fp16, y = rotr_3_cast_fp16)[name = tensor("op_809_cast_fp16")]; + tensor var_810_cast_fp16 = mul(x = var_783_cast_fp16, y = roti_3_cast_fp16)[name = tensor("op_810_cast_fp16")]; + tensor qor_5_cast_fp16 = sub(x = var_809_cast_fp16, y = var_810_cast_fp16)[name = tensor("qor_5_cast_fp16")]; + tensor var_813_cast_fp16 = mul(x = var_775_cast_fp16, y = roti_3_cast_fp16)[name = tensor("op_813_cast_fp16")]; + tensor var_814_cast_fp16 = mul(x = var_783_cast_fp16, y = rotr_3_cast_fp16)[name = tensor("op_814_cast_fp16")]; + tensor qoi_5_cast_fp16 = add(x = var_813_cast_fp16, y = var_814_cast_fp16)[name = tensor("qoi_5_cast_fp16")]; + tensor var_817_cast_fp16 = mul(x = var_791_cast_fp16, y = rotr_3_cast_fp16)[name = tensor("op_817_cast_fp16")]; + tensor var_818_cast_fp16 = mul(x = var_799_cast_fp16, y = roti_3_cast_fp16)[name = tensor("op_818_cast_fp16")]; + tensor kor_5_cast_fp16 = sub(x = var_817_cast_fp16, y = var_818_cast_fp16)[name = tensor("kor_5_cast_fp16")]; + tensor var_821_cast_fp16 = mul(x = var_791_cast_fp16, y = roti_3_cast_fp16)[name = tensor("op_821_cast_fp16")]; + tensor var_822_cast_fp16 = mul(x = var_799_cast_fp16, y = rotr_3_cast_fp16)[name = tensor("op_822_cast_fp16")]; + tensor koi_5_cast_fp16 = add(x = var_821_cast_fp16, y = var_822_cast_fp16)[name = tensor("koi_5_cast_fp16")]; + tensor qo_3_axis_0 = const()[name = tensor("qo_3_axis_0"), val = tensor(-1)]; + tensor qo_3_cast_fp16 = stack(axis = qo_3_axis_0, values = (qor_5_cast_fp16, qoi_5_cast_fp16))[name = tensor("qo_3_cast_fp16")]; + tensor ko_3_axis_0 = const()[name = tensor("ko_3_axis_0"), val = tensor(-1)]; + tensor ko_3_cast_fp16 = stack(axis = ko_3_axis_0, values = (kor_5_cast_fp16, koi_5_cast_fp16))[name = tensor("ko_3_cast_fp16")]; + tensor var_851 = const()[name = tensor("op_851"), val = tensor([1, 1, 16, 64])]; + tensor q_9_cast_fp16 = reshape(shape = var_851, x = qo_3_cast_fp16)[name = tensor("q_9_cast_fp16")]; + tensor var_853 = const()[name = tensor("op_853"), val = tensor([1, 1, 16, 64])]; + tensor k_7_cast_fp16 = reshape(shape = var_853, x = ko_3_cast_fp16)[name = tensor("k_7_cast_fp16")]; + tensor _inversed_875_y_0_to_fp16 = const()[name = tensor("_inversed_875_y_0_to_fp16"), val = tensor(0x1p-9)]; + tensor _inversed_875_cast_fp16 = mul(x = ts_11_cast_fp16, y = _inversed_875_y_0_to_fp16)[name = tensor("_inversed_875_cast_fp16")]; + tensor var_876_cast_fp16 = floor(x = _inversed_875_cast_fp16)[name = tensor("op_876_cast_fp16")]; + tensor var_877_to_fp16 = const()[name = tensor("op_877_to_fp16"), val = tensor(0x1p+9)]; + tensor var_878_cast_fp16 = mul(x = var_876_cast_fp16, y = var_877_to_fp16)[name = tensor("op_878_cast_fp16")]; + tensor write_indices_float_7_cast_fp16 = sub(x = ts_11_cast_fp16, y = var_878_cast_fp16)[name = tensor("write_indices_float_7_cast_fp16")]; + tensor var_885_dtype_0 = const()[name = tensor("op_885_dtype_0"), val = tensor("int32")]; + tensor write_indices_3_reps_0 = const()[name = tensor("write_indices_3_reps_0"), val = tensor([1, 1, 16, 64])]; + tensor write_indices_float_7_cast_fp16_to_int32 = cast(dtype = var_885_dtype_0, x = write_indices_float_7_cast_fp16)[name = tensor("cast_547")]; + tensor write_indices_3 = tile(reps = write_indices_3_reps_0, x = write_indices_float_7_cast_fp16_to_int32)[name = tensor("write_indices_3")]; + tensor var_893_begin_0 = const()[name = tensor("op_893_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_893_end_0 = const()[name = tensor("op_893_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor var_893_end_mask_0 = const()[name = tensor("op_893_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_893_squeeze_mask_0 = const()[name = tensor("op_893_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor cache1_to_fp16_dtype_0 = const()[name = tensor("cache1_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor cache1_to_fp16 = cast(dtype = cache1_to_fp16_dtype_0, x = cache1)[name = tensor("cast_546")]; + tensor var_893_cast_fp16 = slice_by_index(begin = var_893_begin_0, end = var_893_end_0, end_mask = var_893_end_mask_0, squeeze_mask = var_893_squeeze_mask_0, x = cache1_to_fp16)[name = tensor("op_893_cast_fp16")]; + tensor var_895_axis_0 = const()[name = tensor("op_895_axis_0"), val = tensor(1)]; + tensor var_895_mode_0 = const()[name = tensor("op_895_mode_0"), val = tensor("update")]; + tensor var_895_validate_indices_0 = const()[name = tensor("op_895_validate_indices_0"), val = tensor(false)]; + tensor var_895_cast_fp16 = scatter_along_axis(axis = var_895_axis_0, data = var_893_cast_fp16, indices = write_indices_3, mode = var_895_mode_0, updates = k_7_cast_fp16, validate_indices = var_895_validate_indices_0)[name = tensor("op_895_cast_fp16")]; + tensor concat_9 = const()[name = tensor("concat_9"), val = tensor([0, 0, 0, 0, 0])]; + tensor concat_10 = const()[name = tensor("concat_10"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_3_internal_tensor_assign_1_stride_0 = const()[name = tensor("new_cache_3_internal_tensor_assign_1_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_3_internal_tensor_assign_1_begin_mask_0 = const()[name = tensor("new_cache_3_internal_tensor_assign_1_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_3_internal_tensor_assign_1_end_mask_0 = const()[name = tensor("new_cache_3_internal_tensor_assign_1_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_3_internal_tensor_assign_1_squeeze_mask_0 = const()[name = tensor("new_cache_3_internal_tensor_assign_1_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_50 = const()[name = tensor("shape_50"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_2 = const()[name = tensor("reduce_prod_2"), val = tensor(1048576)]; + tensor range_1d_2_start_0 = const()[name = tensor("range_1d_2_start_0"), val = tensor(0)]; + tensor range_1d_2_step_0 = const()[name = tensor("range_1d_2_step_0"), val = tensor(1)]; + tensor range_1d_2 = range_1d(end = reduce_prod_2, start = range_1d_2_start_0, step = range_1d_2_step_0)[name = tensor("range_1d_2")]; + tensor reshape_10 = reshape(shape = shape_50, x = range_1d_2)[name = tensor("reshape_10")]; + tensor slice_by_index_2 = slice_by_index(begin = concat_9, begin_mask = new_cache_3_internal_tensor_assign_1_begin_mask_0, end = concat_10, end_mask = new_cache_3_internal_tensor_assign_1_end_mask_0, squeeze_mask = new_cache_3_internal_tensor_assign_1_squeeze_mask_0, stride = new_cache_3_internal_tensor_assign_1_stride_0, x = reshape_10)[name = tensor("slice_by_index_2")]; + tensor reshape_11_shape_0 = const()[name = tensor("reshape_11_shape_0"), val = tensor([-1])]; + tensor reshape_11 = reshape(shape = reshape_11_shape_0, x = slice_by_index_2)[name = tensor("reshape_11")]; + tensor reshape_12_shape_0 = const()[name = tensor("reshape_12_shape_0"), val = tensor([-1])]; + tensor reshape_12_cast_fp16 = reshape(shape = reshape_12_shape_0, x = var_895_cast_fp16)[name = tensor("reshape_12_cast_fp16")]; + tensor reshape_13_shape_0 = const()[name = tensor("reshape_13_shape_0"), val = tensor([-1])]; + tensor reshape_13_cast_fp16 = reshape(shape = reshape_13_shape_0, x = cache1_to_fp16)[name = tensor("reshape_13_cast_fp16")]; + tensor scatter_2_mode_0 = const()[name = tensor("scatter_2_mode_0"), val = tensor("update")]; + tensor scatter_2_axis_0 = const()[name = tensor("scatter_2_axis_0"), val = tensor(0)]; + tensor scatter_2_validate_indices_0 = const()[name = tensor("scatter_2_validate_indices_0"), val = tensor(false)]; + tensor scatter_2_cast_fp16 = scatter(axis = scatter_2_axis_0, data = reshape_13_cast_fp16, indices = reshape_11, mode = scatter_2_mode_0, updates = reshape_12_cast_fp16, validate_indices = scatter_2_validate_indices_0)[name = tensor("scatter_2_cast_fp16")]; + tensor reshape_14_cast_fp16 = reshape(shape = shape_50, x = scatter_2_cast_fp16)[name = tensor("reshape_14_cast_fp16")]; + tensor var_903_begin_0 = const()[name = tensor("op_903_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor var_903_end_0 = const()[name = tensor("op_903_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor var_903_end_mask_0 = const()[name = tensor("op_903_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_903_squeeze_mask_0 = const()[name = tensor("op_903_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor var_903_cast_fp16 = slice_by_index(begin = var_903_begin_0, end = var_903_end_0, end_mask = var_903_end_mask_0, squeeze_mask = var_903_squeeze_mask_0, x = reshape_14_cast_fp16)[name = tensor("op_903_cast_fp16")]; + tensor var_905_axis_0 = const()[name = tensor("op_905_axis_0"), val = tensor(1)]; + tensor var_905_mode_0 = const()[name = tensor("op_905_mode_0"), val = tensor("update")]; + tensor var_905_validate_indices_0 = const()[name = tensor("op_905_validate_indices_0"), val = tensor(false)]; + tensor var_905_cast_fp16 = scatter_along_axis(axis = var_905_axis_0, data = var_903_cast_fp16, indices = write_indices_3, mode = var_905_mode_0, updates = v_3_cast_fp16, validate_indices = var_905_validate_indices_0)[name = tensor("op_905_cast_fp16")]; + tensor concat_11 = const()[name = tensor("concat_11"), val = tensor([1, 0, 0, 0, 0])]; + tensor concat_12 = const()[name = tensor("concat_12"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_3_internal_tensor_assign_2_stride_0 = const()[name = tensor("new_cache_3_internal_tensor_assign_2_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_3_internal_tensor_assign_2_begin_mask_0 = const()[name = tensor("new_cache_3_internal_tensor_assign_2_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_3_internal_tensor_assign_2_end_mask_0 = const()[name = tensor("new_cache_3_internal_tensor_assign_2_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_3_internal_tensor_assign_2_squeeze_mask_0 = const()[name = tensor("new_cache_3_internal_tensor_assign_2_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_51 = const()[name = tensor("shape_51"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_3 = const()[name = tensor("reduce_prod_3"), val = tensor(1048576)]; + tensor range_1d_3_start_0 = const()[name = tensor("range_1d_3_start_0"), val = tensor(0)]; + tensor range_1d_3_step_0 = const()[name = tensor("range_1d_3_step_0"), val = tensor(1)]; + tensor range_1d_3 = range_1d(end = reduce_prod_3, start = range_1d_3_start_0, step = range_1d_3_step_0)[name = tensor("range_1d_3")]; + tensor reshape_15 = reshape(shape = shape_51, x = range_1d_3)[name = tensor("reshape_15")]; + tensor slice_by_index_3 = slice_by_index(begin = concat_11, begin_mask = new_cache_3_internal_tensor_assign_2_begin_mask_0, end = concat_12, end_mask = new_cache_3_internal_tensor_assign_2_end_mask_0, squeeze_mask = new_cache_3_internal_tensor_assign_2_squeeze_mask_0, stride = new_cache_3_internal_tensor_assign_2_stride_0, x = reshape_15)[name = tensor("slice_by_index_3")]; + tensor reshape_16_shape_0 = const()[name = tensor("reshape_16_shape_0"), val = tensor([-1])]; + tensor reshape_16 = reshape(shape = reshape_16_shape_0, x = slice_by_index_3)[name = tensor("reshape_16")]; + tensor reshape_17_shape_0 = const()[name = tensor("reshape_17_shape_0"), val = tensor([-1])]; + tensor reshape_17_cast_fp16 = reshape(shape = reshape_17_shape_0, x = var_905_cast_fp16)[name = tensor("reshape_17_cast_fp16")]; + tensor reshape_18_shape_0 = const()[name = tensor("reshape_18_shape_0"), val = tensor([-1])]; + tensor reshape_18_cast_fp16 = reshape(shape = reshape_18_shape_0, x = reshape_14_cast_fp16)[name = tensor("reshape_18_cast_fp16")]; + tensor scatter_3_mode_0 = const()[name = tensor("scatter_3_mode_0"), val = tensor("update")]; + tensor scatter_3_axis_0 = const()[name = tensor("scatter_3_axis_0"), val = tensor(0)]; + tensor scatter_3_validate_indices_0 = const()[name = tensor("scatter_3_validate_indices_0"), val = tensor(false)]; + tensor scatter_3_cast_fp16 = scatter(axis = scatter_3_axis_0, data = reshape_18_cast_fp16, indices = reshape_16, mode = scatter_3_mode_0, updates = reshape_17_cast_fp16, validate_indices = scatter_3_validate_indices_0)[name = tensor("scatter_3_cast_fp16")]; + tensor reshape_19_cast_fp16 = reshape(shape = shape_51, x = scatter_3_cast_fp16)[name = tensor("reshape_19_cast_fp16")]; + tensor reshape_19_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("reshape_19_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor keys_7_begin_0 = const()[name = tensor("keys_7_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor keys_7_end_0 = const()[name = tensor("keys_7_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor keys_7_end_mask_0 = const()[name = tensor("keys_7_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor keys_7_squeeze_mask_0 = const()[name = tensor("keys_7_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor keys_7_cast_fp16 = slice_by_index(begin = keys_7_begin_0, end = keys_7_end_0, end_mask = keys_7_end_mask_0, squeeze_mask = keys_7_squeeze_mask_0, x = reshape_19_cast_fp16)[name = tensor("keys_7_cast_fp16")]; + tensor values_7_begin_0 = const()[name = tensor("values_7_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor values_7_end_0 = const()[name = tensor("values_7_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor values_7_end_mask_0 = const()[name = tensor("values_7_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor values_7_squeeze_mask_0 = const()[name = tensor("values_7_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor values_7_cast_fp16 = slice_by_index(begin = values_7_begin_0, end = values_7_end_0, end_mask = values_7_end_mask_0, squeeze_mask = values_7_squeeze_mask_0, x = reshape_19_cast_fp16)[name = tensor("values_7_cast_fp16")]; + tensor var_917_cast_fp16 = not_equal(x = keys_7_cast_fp16, y = keys_7_cast_fp16)[name = tensor("op_917_cast_fp16")]; + tensor keys_9_cast_fp16 = select(a = var_504_to_fp16, b = keys_7_cast_fp16, cond = var_917_cast_fp16)[name = tensor("keys_9_cast_fp16")]; + tensor var_925_cast_fp16 = not_equal(x = values_7_cast_fp16, y = values_7_cast_fp16)[name = tensor("op_925_cast_fp16")]; + tensor values_9_cast_fp16 = select(a = var_504_to_fp16, b = values_7_cast_fp16, cond = var_925_cast_fp16)[name = tensor("values_9_cast_fp16")]; + tensor var_949 = const()[name = tensor("op_949"), val = tensor([0, 2, 1, 3])]; + tensor var_962 = const()[name = tensor("op_962"), val = tensor([1, 1, 1])]; + tensor var_963_cast_fp16 = reshape(shape = var_962, x = position1_to_fp16)[name = tensor("op_963_cast_fp16")]; + tensor var_980_to_fp16 = const()[name = tensor("op_980_to_fp16"), val = tensor(0x1p+0)]; + tensor valid_len_3_cast_fp16 = add(x = var_963_cast_fp16, y = var_980_to_fp16)[name = tensor("valid_len_3_cast_fp16")]; + tensor valid_mask_3_cast_fp16 = less(x = k_positions_1_promoted_to_fp16, y = valid_len_3_cast_fp16)[name = tensor("valid_mask_3_cast_fp16")]; + tensor causal_mask_3_cast_fp16 = less_equal(x = k_positions_1_promoted_to_fp16, y = var_963_cast_fp16)[name = tensor("causal_mask_3_cast_fp16")]; + tensor attn_mask_5 = logical_and(x = valid_mask_3_cast_fp16, y = causal_mask_3_cast_fp16)[name = tensor("attn_mask_5")]; + tensor attn_mask_7_axes_0 = const()[name = tensor("attn_mask_7_axes_0"), val = tensor([1])]; + tensor attn_mask_7 = expand_dims(axes = attn_mask_7_axes_0, x = attn_mask_5)[name = tensor("attn_mask_7")]; + tensor var_998_transpose_x_0 = const()[name = tensor("op_998_transpose_x_0"), val = tensor(false)]; + tensor var_998_transpose_y_0 = const()[name = tensor("op_998_transpose_y_0"), val = tensor(false)]; + tensor transpose_74_perm_0 = const()[name = tensor("transpose_74_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_75_perm_0 = const()[name = tensor("transpose_75_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_75 = transpose(perm = transpose_75_perm_0, x = keys_9_cast_fp16)[name = tensor("transpose_209")]; + tensor transpose_74 = transpose(perm = transpose_74_perm_0, x = q_9_cast_fp16)[name = tensor("transpose_210")]; + tensor var_998_cast_fp16 = matmul(transpose_x = var_998_transpose_x_0, transpose_y = var_998_transpose_y_0, x = transpose_74, y = transpose_75)[name = tensor("op_998_cast_fp16")]; + tensor var_992_to_fp16 = const()[name = tensor("op_992_to_fp16"), val = tensor([0x1p-3])]; + tensor attn_weights_7_cast_fp16 = mul(x = var_998_cast_fp16, y = var_992_to_fp16)[name = tensor("attn_weights_7_cast_fp16")]; + tensor var_1000 = logical_not(x = attn_mask_7)[name = tensor("op_1000")]; + tensor var_1001_to_fp16 = const()[name = tensor("op_1001_to_fp16"), val = tensor(-inf)]; + tensor attn_weights_9_cast_fp16 = select(a = var_1001_to_fp16, b = attn_weights_7_cast_fp16, cond = var_1000)[name = tensor("attn_weights_9_cast_fp16")]; + tensor var_1003 = const()[name = tensor("op_1003"), val = tensor(-1)]; + tensor attn_weights_11_cast_fp16 = softmax(axis = var_1003, x = attn_weights_9_cast_fp16)[name = tensor("attn_weights_11_cast_fp16")]; + tensor attn_output_3_transpose_x_0 = const()[name = tensor("attn_output_3_transpose_x_0"), val = tensor(false)]; + tensor attn_output_3_transpose_y_0 = const()[name = tensor("attn_output_3_transpose_y_0"), val = tensor(false)]; + tensor values_11_cast_fp16 = transpose(perm = var_949, x = values_9_cast_fp16)[name = tensor("transpose_211")]; + tensor attn_output_3_cast_fp16 = matmul(transpose_x = attn_output_3_transpose_x_0, transpose_y = attn_output_3_transpose_y_0, x = attn_weights_11_cast_fp16, y = values_11_cast_fp16)[name = tensor("attn_output_3_cast_fp16")]; + tensor var_1011 = const()[name = tensor("op_1011"), val = tensor([0, 2, 1, 3])]; + tensor var_1014 = const()[name = tensor("op_1014"), val = tensor([1, 1, 1024])]; + tensor var_1012_cast_fp16 = transpose(perm = var_1011, x = attn_output_3_cast_fp16)[name = tensor("transpose_208")]; + tensor input_15_cast_fp16 = reshape(shape = var_1014, x = var_1012_cast_fp16)[name = tensor("input_15_cast_fp16")]; + tensor linear_6_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = attn1_out_proj_weight_cast_fp16, x = input_15_cast_fp16)[name = tensor("linear_6_cast_fp16")]; + tensor var_1020_to_fp16 = const()[name = tensor("op_1020_to_fp16"), val = tensor(0x1p+0)]; + tensor var_1021_cast_fp16 = add(x = position1_to_fp16, y = var_1020_to_fp16)[name = tensor("op_1021_cast_fp16")]; + tensor var_1021_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_1021_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor input_17_cast_fp16 = add(x = input_13_cast_fp16, y = linear_6_cast_fp16)[name = tensor("input_17_cast_fp16")]; + tensor input_19_axes_0 = const()[name = tensor("input_19_axes_0"), val = tensor([-1])]; + tensor norm1_2_weight_to_fp16 = const()[name = tensor("norm1_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(303597824)))]; + tensor norm1_2_bias_to_fp16 = const()[name = tensor("norm1_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(303599936)))]; + tensor var_1025_to_fp16 = const()[name = tensor("op_1025_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_19_cast_fp16 = layer_norm(axes = input_19_axes_0, beta = norm1_2_bias_to_fp16, epsilon = var_1025_to_fp16, gamma = norm1_2_weight_to_fp16, x = input_17_cast_fp16)[name = tensor("input_19_cast_fp16")]; + tensor linear_7_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = linear1_1_weight_cast_fp16, x = input_19_cast_fp16)[name = tensor("linear_7_cast_fp16")]; + tensor input_21_mode_0 = const()[name = tensor("input_21_mode_0"), val = tensor("EXACT")]; + tensor input_21_cast_fp16 = gelu(mode = input_21_mode_0, x = linear_7_cast_fp16)[name = tensor("input_21_cast_fp16")]; + tensor linear_8_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = linear1_2_weight_cast_fp16, x = input_21_cast_fp16)[name = tensor("linear_8_cast_fp16")]; + tensor input_23_cast_fp16 = add(x = input_17_cast_fp16, y = linear_8_cast_fp16)[name = tensor("input_23_cast_fp16")]; + tensor x_5_axes_0 = const()[name = tensor("x_5_axes_0"), val = tensor([-1])]; + tensor norm2_1_weight_to_fp16 = const()[name = tensor("norm2_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(303602048)))]; + tensor norm2_1_bias_to_fp16 = const()[name = tensor("norm2_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(303604160)))]; + tensor var_1042_to_fp16 = const()[name = tensor("op_1042_to_fp16"), val = tensor(0x1.5p-17)]; + tensor x_5_cast_fp16 = layer_norm(axes = x_5_axes_0, beta = norm2_1_bias_to_fp16, epsilon = var_1042_to_fp16, gamma = norm2_1_weight_to_fp16, x = input_23_cast_fp16)[name = tensor("x_5_cast_fp16")]; + tensor linear_9_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = attn2_in_proj_weight_cast_fp16, x = x_5_cast_fp16)[name = tensor("linear_9_cast_fp16")]; + tensor var_1078 = const()[name = tensor("op_1078"), val = tensor([1, 1, 3, 16, 64])]; + tensor qkv_5_cast_fp16 = reshape(shape = var_1078, x = linear_9_cast_fp16)[name = tensor("qkv_5_cast_fp16")]; + tensor q_13_begin_0 = const()[name = tensor("q_13_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor q_13_end_0 = const()[name = tensor("q_13_end_0"), val = tensor([1, 1, 1, 16, 64])]; + tensor q_13_end_mask_0 = const()[name = tensor("q_13_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor q_13_squeeze_mask_0 = const()[name = tensor("q_13_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor q_13_cast_fp16 = slice_by_index(begin = q_13_begin_0, end = q_13_end_0, end_mask = q_13_end_mask_0, squeeze_mask = q_13_squeeze_mask_0, x = qkv_5_cast_fp16)[name = tensor("q_13_cast_fp16")]; + tensor k_9_begin_0 = const()[name = tensor("k_9_begin_0"), val = tensor([0, 0, 1, 0, 0])]; + tensor k_9_end_0 = const()[name = tensor("k_9_end_0"), val = tensor([1, 1, 2, 16, 64])]; + tensor k_9_end_mask_0 = const()[name = tensor("k_9_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor k_9_squeeze_mask_0 = const()[name = tensor("k_9_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor k_9_cast_fp16 = slice_by_index(begin = k_9_begin_0, end = k_9_end_0, end_mask = k_9_end_mask_0, squeeze_mask = k_9_squeeze_mask_0, x = qkv_5_cast_fp16)[name = tensor("k_9_cast_fp16")]; + tensor v_5_begin_0 = const()[name = tensor("v_5_begin_0"), val = tensor([0, 0, 2, 0, 0])]; + tensor v_5_end_0 = const()[name = tensor("v_5_end_0"), val = tensor([1, 1, 3, 16, 64])]; + tensor v_5_end_mask_0 = const()[name = tensor("v_5_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor v_5_squeeze_mask_0 = const()[name = tensor("v_5_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor v_5_cast_fp16 = slice_by_index(begin = v_5_begin_0, end = v_5_end_0, end_mask = v_5_end_mask_0, squeeze_mask = v_5_squeeze_mask_0, x = qkv_5_cast_fp16)[name = tensor("v_5_cast_fp16")]; + tensor var_1182 = const()[name = tensor("op_1182"), val = tensor([1, 1, 1, 1])]; + tensor position2_to_fp16_dtype_0 = const()[name = tensor("position2_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor position2_to_fp16 = cast(dtype = position2_to_fp16_dtype_0, x = position2)[name = tensor("cast_543")]; + tensor ts_17_cast_fp16 = reshape(shape = var_1182, x = position2_to_fp16)[name = tensor("ts_17_cast_fp16")]; + tensor var_1186 = const()[name = tensor("op_1186"), val = tensor([1, 1, 16, 32, 2])]; + tensor q_complex_5_cast_fp16 = reshape(shape = var_1186, x = q_13_cast_fp16)[name = tensor("q_complex_5_cast_fp16")]; + tensor var_1190 = const()[name = tensor("op_1190"), val = tensor([1, 1, 16, 32, 2])]; + tensor k_complex_5_cast_fp16 = reshape(shape = var_1190, x = k_9_cast_fp16)[name = tensor("k_complex_5_cast_fp16")]; + tensor var_1194_begin_0 = const()[name = tensor("op_1194_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_1194_end_0 = const()[name = tensor("op_1194_end_0"), val = tensor([1, 1, 16, 32, 1])]; + tensor var_1194_end_mask_0 = const()[name = tensor("op_1194_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_1194_squeeze_mask_0 = const()[name = tensor("op_1194_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_1194_cast_fp16 = slice_by_index(begin = var_1194_begin_0, end = var_1194_end_0, end_mask = var_1194_end_mask_0, squeeze_mask = var_1194_squeeze_mask_0, x = q_complex_5_cast_fp16)[name = tensor("op_1194_cast_fp16")]; + tensor var_1202_begin_0 = const()[name = tensor("op_1202_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_1202_end_0 = const()[name = tensor("op_1202_end_0"), val = tensor([1, 1, 16, 32, 2])]; + tensor var_1202_end_mask_0 = const()[name = tensor("op_1202_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_1202_squeeze_mask_0 = const()[name = tensor("op_1202_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_1202_cast_fp16 = slice_by_index(begin = var_1202_begin_0, end = var_1202_end_0, end_mask = var_1202_end_mask_0, squeeze_mask = var_1202_squeeze_mask_0, x = q_complex_5_cast_fp16)[name = tensor("op_1202_cast_fp16")]; + tensor var_1210_begin_0 = const()[name = tensor("op_1210_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_1210_end_0 = const()[name = tensor("op_1210_end_0"), val = tensor([1, 1, 16, 32, 1])]; + tensor var_1210_end_mask_0 = const()[name = tensor("op_1210_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_1210_squeeze_mask_0 = const()[name = tensor("op_1210_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_1210_cast_fp16 = slice_by_index(begin = var_1210_begin_0, end = var_1210_end_0, end_mask = var_1210_end_mask_0, squeeze_mask = var_1210_squeeze_mask_0, x = k_complex_5_cast_fp16)[name = tensor("op_1210_cast_fp16")]; + tensor var_1218_begin_0 = const()[name = tensor("op_1218_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_1218_end_0 = const()[name = tensor("op_1218_end_0"), val = tensor([1, 1, 16, 32, 2])]; + tensor var_1218_end_mask_0 = const()[name = tensor("op_1218_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_1218_squeeze_mask_0 = const()[name = tensor("op_1218_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_1218_cast_fp16 = slice_by_index(begin = var_1218_begin_0, end = var_1218_end_0, end_mask = var_1218_end_mask_0, squeeze_mask = var_1218_squeeze_mask_0, x = k_complex_5_cast_fp16)[name = tensor("op_1218_cast_fp16")]; + tensor freqs_5_to_fp16 = const()[name = tensor("freqs_5_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(303606272)))]; + tensor var_1224_cast_fp16 = mul(x = freqs_5_to_fp16, y = ts_17_cast_fp16)[name = tensor("op_1224_cast_fp16")]; + tensor rotr_5_cast_fp16 = cos(x = var_1224_cast_fp16)[name = tensor("rotr_5_cast_fp16")]; + tensor roti_5_cast_fp16 = sin(x = var_1224_cast_fp16)[name = tensor("roti_5_cast_fp16")]; + tensor var_1228_cast_fp16 = mul(x = var_1194_cast_fp16, y = rotr_5_cast_fp16)[name = tensor("op_1228_cast_fp16")]; + tensor var_1229_cast_fp16 = mul(x = var_1202_cast_fp16, y = roti_5_cast_fp16)[name = tensor("op_1229_cast_fp16")]; + tensor qor_9_cast_fp16 = sub(x = var_1228_cast_fp16, y = var_1229_cast_fp16)[name = tensor("qor_9_cast_fp16")]; + tensor var_1232_cast_fp16 = mul(x = var_1194_cast_fp16, y = roti_5_cast_fp16)[name = tensor("op_1232_cast_fp16")]; + tensor var_1233_cast_fp16 = mul(x = var_1202_cast_fp16, y = rotr_5_cast_fp16)[name = tensor("op_1233_cast_fp16")]; + tensor qoi_9_cast_fp16 = add(x = var_1232_cast_fp16, y = var_1233_cast_fp16)[name = tensor("qoi_9_cast_fp16")]; + tensor var_1236_cast_fp16 = mul(x = var_1210_cast_fp16, y = rotr_5_cast_fp16)[name = tensor("op_1236_cast_fp16")]; + tensor var_1237_cast_fp16 = mul(x = var_1218_cast_fp16, y = roti_5_cast_fp16)[name = tensor("op_1237_cast_fp16")]; + tensor kor_9_cast_fp16 = sub(x = var_1236_cast_fp16, y = var_1237_cast_fp16)[name = tensor("kor_9_cast_fp16")]; + tensor var_1240_cast_fp16 = mul(x = var_1210_cast_fp16, y = roti_5_cast_fp16)[name = tensor("op_1240_cast_fp16")]; + tensor var_1241_cast_fp16 = mul(x = var_1218_cast_fp16, y = rotr_5_cast_fp16)[name = tensor("op_1241_cast_fp16")]; + tensor koi_9_cast_fp16 = add(x = var_1240_cast_fp16, y = var_1241_cast_fp16)[name = tensor("koi_9_cast_fp16")]; + tensor qo_5_axis_0 = const()[name = tensor("qo_5_axis_0"), val = tensor(-1)]; + tensor qo_5_cast_fp16 = stack(axis = qo_5_axis_0, values = (qor_9_cast_fp16, qoi_9_cast_fp16))[name = tensor("qo_5_cast_fp16")]; + tensor ko_5_axis_0 = const()[name = tensor("ko_5_axis_0"), val = tensor(-1)]; + tensor ko_5_cast_fp16 = stack(axis = ko_5_axis_0, values = (kor_9_cast_fp16, koi_9_cast_fp16))[name = tensor("ko_5_cast_fp16")]; + tensor var_1270 = const()[name = tensor("op_1270"), val = tensor([1, 1, 16, 64])]; + tensor q_15_cast_fp16 = reshape(shape = var_1270, x = qo_5_cast_fp16)[name = tensor("q_15_cast_fp16")]; + tensor var_1272 = const()[name = tensor("op_1272"), val = tensor([1, 1, 16, 64])]; + tensor k_11_cast_fp16 = reshape(shape = var_1272, x = ko_5_cast_fp16)[name = tensor("k_11_cast_fp16")]; + tensor _inversed_1294_y_0_to_fp16 = const()[name = tensor("_inversed_1294_y_0_to_fp16"), val = tensor(0x1p-9)]; + tensor _inversed_1294_cast_fp16 = mul(x = ts_17_cast_fp16, y = _inversed_1294_y_0_to_fp16)[name = tensor("_inversed_1294_cast_fp16")]; + tensor var_1295_cast_fp16 = floor(x = _inversed_1294_cast_fp16)[name = tensor("op_1295_cast_fp16")]; + tensor var_1296_to_fp16 = const()[name = tensor("op_1296_to_fp16"), val = tensor(0x1p+9)]; + tensor var_1297_cast_fp16 = mul(x = var_1295_cast_fp16, y = var_1296_to_fp16)[name = tensor("op_1297_cast_fp16")]; + tensor write_indices_float_11_cast_fp16 = sub(x = ts_17_cast_fp16, y = var_1297_cast_fp16)[name = tensor("write_indices_float_11_cast_fp16")]; + tensor var_1304_dtype_0 = const()[name = tensor("op_1304_dtype_0"), val = tensor("int32")]; + tensor write_indices_5_reps_0 = const()[name = tensor("write_indices_5_reps_0"), val = tensor([1, 1, 16, 64])]; + tensor write_indices_float_11_cast_fp16_to_int32 = cast(dtype = var_1304_dtype_0, x = write_indices_float_11_cast_fp16)[name = tensor("cast_542")]; + tensor write_indices_5 = tile(reps = write_indices_5_reps_0, x = write_indices_float_11_cast_fp16_to_int32)[name = tensor("write_indices_5")]; + tensor var_1312_begin_0 = const()[name = tensor("op_1312_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_1312_end_0 = const()[name = tensor("op_1312_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor var_1312_end_mask_0 = const()[name = tensor("op_1312_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_1312_squeeze_mask_0 = const()[name = tensor("op_1312_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor cache2_to_fp16_dtype_0 = const()[name = tensor("cache2_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor cache2_to_fp16 = cast(dtype = cache2_to_fp16_dtype_0, x = cache2)[name = tensor("cast_541")]; + tensor var_1312_cast_fp16 = slice_by_index(begin = var_1312_begin_0, end = var_1312_end_0, end_mask = var_1312_end_mask_0, squeeze_mask = var_1312_squeeze_mask_0, x = cache2_to_fp16)[name = tensor("op_1312_cast_fp16")]; + tensor var_1314_axis_0 = const()[name = tensor("op_1314_axis_0"), val = tensor(1)]; + tensor var_1314_mode_0 = const()[name = tensor("op_1314_mode_0"), val = tensor("update")]; + tensor var_1314_validate_indices_0 = const()[name = tensor("op_1314_validate_indices_0"), val = tensor(false)]; + tensor var_1314_cast_fp16 = scatter_along_axis(axis = var_1314_axis_0, data = var_1312_cast_fp16, indices = write_indices_5, mode = var_1314_mode_0, updates = k_11_cast_fp16, validate_indices = var_1314_validate_indices_0)[name = tensor("op_1314_cast_fp16")]; + tensor concat_16 = const()[name = tensor("concat_16"), val = tensor([0, 0, 0, 0, 0])]; + tensor concat_17 = const()[name = tensor("concat_17"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_5_internal_tensor_assign_1_stride_0 = const()[name = tensor("new_cache_5_internal_tensor_assign_1_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_5_internal_tensor_assign_1_begin_mask_0 = const()[name = tensor("new_cache_5_internal_tensor_assign_1_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_5_internal_tensor_assign_1_end_mask_0 = const()[name = tensor("new_cache_5_internal_tensor_assign_1_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_5_internal_tensor_assign_1_squeeze_mask_0 = const()[name = tensor("new_cache_5_internal_tensor_assign_1_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_52 = const()[name = tensor("shape_52"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_4 = const()[name = tensor("reduce_prod_4"), val = tensor(1048576)]; + tensor range_1d_4_start_0 = const()[name = tensor("range_1d_4_start_0"), val = tensor(0)]; + tensor range_1d_4_step_0 = const()[name = tensor("range_1d_4_step_0"), val = tensor(1)]; + tensor range_1d_4 = range_1d(end = reduce_prod_4, start = range_1d_4_start_0, step = range_1d_4_step_0)[name = tensor("range_1d_4")]; + tensor reshape_20 = reshape(shape = shape_52, x = range_1d_4)[name = tensor("reshape_20")]; + tensor slice_by_index_4 = slice_by_index(begin = concat_16, begin_mask = new_cache_5_internal_tensor_assign_1_begin_mask_0, end = concat_17, end_mask = new_cache_5_internal_tensor_assign_1_end_mask_0, squeeze_mask = new_cache_5_internal_tensor_assign_1_squeeze_mask_0, stride = new_cache_5_internal_tensor_assign_1_stride_0, x = reshape_20)[name = tensor("slice_by_index_4")]; + tensor reshape_21_shape_0 = const()[name = tensor("reshape_21_shape_0"), val = tensor([-1])]; + tensor reshape_21 = reshape(shape = reshape_21_shape_0, x = slice_by_index_4)[name = tensor("reshape_21")]; + tensor reshape_22_shape_0 = const()[name = tensor("reshape_22_shape_0"), val = tensor([-1])]; + tensor reshape_22_cast_fp16 = reshape(shape = reshape_22_shape_0, x = var_1314_cast_fp16)[name = tensor("reshape_22_cast_fp16")]; + tensor reshape_23_shape_0 = const()[name = tensor("reshape_23_shape_0"), val = tensor([-1])]; + tensor reshape_23_cast_fp16 = reshape(shape = reshape_23_shape_0, x = cache2_to_fp16)[name = tensor("reshape_23_cast_fp16")]; + tensor scatter_4_mode_0 = const()[name = tensor("scatter_4_mode_0"), val = tensor("update")]; + tensor scatter_4_axis_0 = const()[name = tensor("scatter_4_axis_0"), val = tensor(0)]; + tensor scatter_4_validate_indices_0 = const()[name = tensor("scatter_4_validate_indices_0"), val = tensor(false)]; + tensor scatter_4_cast_fp16 = scatter(axis = scatter_4_axis_0, data = reshape_23_cast_fp16, indices = reshape_21, mode = scatter_4_mode_0, updates = reshape_22_cast_fp16, validate_indices = scatter_4_validate_indices_0)[name = tensor("scatter_4_cast_fp16")]; + tensor reshape_24_cast_fp16 = reshape(shape = shape_52, x = scatter_4_cast_fp16)[name = tensor("reshape_24_cast_fp16")]; + tensor var_1322_begin_0 = const()[name = tensor("op_1322_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor var_1322_end_0 = const()[name = tensor("op_1322_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor var_1322_end_mask_0 = const()[name = tensor("op_1322_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_1322_squeeze_mask_0 = const()[name = tensor("op_1322_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor var_1322_cast_fp16 = slice_by_index(begin = var_1322_begin_0, end = var_1322_end_0, end_mask = var_1322_end_mask_0, squeeze_mask = var_1322_squeeze_mask_0, x = reshape_24_cast_fp16)[name = tensor("op_1322_cast_fp16")]; + tensor var_1324_axis_0 = const()[name = tensor("op_1324_axis_0"), val = tensor(1)]; + tensor var_1324_mode_0 = const()[name = tensor("op_1324_mode_0"), val = tensor("update")]; + tensor var_1324_validate_indices_0 = const()[name = tensor("op_1324_validate_indices_0"), val = tensor(false)]; + tensor var_1324_cast_fp16 = scatter_along_axis(axis = var_1324_axis_0, data = var_1322_cast_fp16, indices = write_indices_5, mode = var_1324_mode_0, updates = v_5_cast_fp16, validate_indices = var_1324_validate_indices_0)[name = tensor("op_1324_cast_fp16")]; + tensor concat_18 = const()[name = tensor("concat_18"), val = tensor([1, 0, 0, 0, 0])]; + tensor concat_19 = const()[name = tensor("concat_19"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_5_internal_tensor_assign_2_stride_0 = const()[name = tensor("new_cache_5_internal_tensor_assign_2_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_5_internal_tensor_assign_2_begin_mask_0 = const()[name = tensor("new_cache_5_internal_tensor_assign_2_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_5_internal_tensor_assign_2_end_mask_0 = const()[name = tensor("new_cache_5_internal_tensor_assign_2_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_5_internal_tensor_assign_2_squeeze_mask_0 = const()[name = tensor("new_cache_5_internal_tensor_assign_2_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_53 = const()[name = tensor("shape_53"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_5 = const()[name = tensor("reduce_prod_5"), val = tensor(1048576)]; + tensor range_1d_5_start_0 = const()[name = tensor("range_1d_5_start_0"), val = tensor(0)]; + tensor range_1d_5_step_0 = const()[name = tensor("range_1d_5_step_0"), val = tensor(1)]; + tensor range_1d_5 = range_1d(end = reduce_prod_5, start = range_1d_5_start_0, step = range_1d_5_step_0)[name = tensor("range_1d_5")]; + tensor reshape_25 = reshape(shape = shape_53, x = range_1d_5)[name = tensor("reshape_25")]; + tensor slice_by_index_5 = slice_by_index(begin = concat_18, begin_mask = new_cache_5_internal_tensor_assign_2_begin_mask_0, end = concat_19, end_mask = new_cache_5_internal_tensor_assign_2_end_mask_0, squeeze_mask = new_cache_5_internal_tensor_assign_2_squeeze_mask_0, stride = new_cache_5_internal_tensor_assign_2_stride_0, x = reshape_25)[name = tensor("slice_by_index_5")]; + tensor reshape_26_shape_0 = const()[name = tensor("reshape_26_shape_0"), val = tensor([-1])]; + tensor reshape_26 = reshape(shape = reshape_26_shape_0, x = slice_by_index_5)[name = tensor("reshape_26")]; + tensor reshape_27_shape_0 = const()[name = tensor("reshape_27_shape_0"), val = tensor([-1])]; + tensor reshape_27_cast_fp16 = reshape(shape = reshape_27_shape_0, x = var_1324_cast_fp16)[name = tensor("reshape_27_cast_fp16")]; + tensor reshape_28_shape_0 = const()[name = tensor("reshape_28_shape_0"), val = tensor([-1])]; + tensor reshape_28_cast_fp16 = reshape(shape = reshape_28_shape_0, x = reshape_24_cast_fp16)[name = tensor("reshape_28_cast_fp16")]; + tensor scatter_5_mode_0 = const()[name = tensor("scatter_5_mode_0"), val = tensor("update")]; + tensor scatter_5_axis_0 = const()[name = tensor("scatter_5_axis_0"), val = tensor(0)]; + tensor scatter_5_validate_indices_0 = const()[name = tensor("scatter_5_validate_indices_0"), val = tensor(false)]; + tensor scatter_5_cast_fp16 = scatter(axis = scatter_5_axis_0, data = reshape_28_cast_fp16, indices = reshape_26, mode = scatter_5_mode_0, updates = reshape_27_cast_fp16, validate_indices = scatter_5_validate_indices_0)[name = tensor("scatter_5_cast_fp16")]; + tensor reshape_29_cast_fp16 = reshape(shape = shape_53, x = scatter_5_cast_fp16)[name = tensor("reshape_29_cast_fp16")]; + tensor reshape_29_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("reshape_29_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor keys_13_begin_0 = const()[name = tensor("keys_13_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor keys_13_end_0 = const()[name = tensor("keys_13_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor keys_13_end_mask_0 = const()[name = tensor("keys_13_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor keys_13_squeeze_mask_0 = const()[name = tensor("keys_13_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor keys_13_cast_fp16 = slice_by_index(begin = keys_13_begin_0, end = keys_13_end_0, end_mask = keys_13_end_mask_0, squeeze_mask = keys_13_squeeze_mask_0, x = reshape_29_cast_fp16)[name = tensor("keys_13_cast_fp16")]; + tensor values_13_begin_0 = const()[name = tensor("values_13_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor values_13_end_0 = const()[name = tensor("values_13_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor values_13_end_mask_0 = const()[name = tensor("values_13_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor values_13_squeeze_mask_0 = const()[name = tensor("values_13_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor values_13_cast_fp16 = slice_by_index(begin = values_13_begin_0, end = values_13_end_0, end_mask = values_13_end_mask_0, squeeze_mask = values_13_squeeze_mask_0, x = reshape_29_cast_fp16)[name = tensor("values_13_cast_fp16")]; + tensor var_1336_cast_fp16 = not_equal(x = keys_13_cast_fp16, y = keys_13_cast_fp16)[name = tensor("op_1336_cast_fp16")]; + tensor keys_15_cast_fp16 = select(a = var_504_to_fp16, b = keys_13_cast_fp16, cond = var_1336_cast_fp16)[name = tensor("keys_15_cast_fp16")]; + tensor var_1344_cast_fp16 = not_equal(x = values_13_cast_fp16, y = values_13_cast_fp16)[name = tensor("op_1344_cast_fp16")]; + tensor values_15_cast_fp16 = select(a = var_504_to_fp16, b = values_13_cast_fp16, cond = var_1344_cast_fp16)[name = tensor("values_15_cast_fp16")]; + tensor var_1368 = const()[name = tensor("op_1368"), val = tensor([0, 2, 1, 3])]; + tensor var_1381 = const()[name = tensor("op_1381"), val = tensor([1, 1, 1])]; + tensor var_1382_cast_fp16 = reshape(shape = var_1381, x = position2_to_fp16)[name = tensor("op_1382_cast_fp16")]; + tensor var_1399_to_fp16 = const()[name = tensor("op_1399_to_fp16"), val = tensor(0x1p+0)]; + tensor valid_len_5_cast_fp16 = add(x = var_1382_cast_fp16, y = var_1399_to_fp16)[name = tensor("valid_len_5_cast_fp16")]; + tensor valid_mask_5_cast_fp16 = less(x = k_positions_1_promoted_to_fp16, y = valid_len_5_cast_fp16)[name = tensor("valid_mask_5_cast_fp16")]; + tensor causal_mask_5_cast_fp16 = less_equal(x = k_positions_1_promoted_to_fp16, y = var_1382_cast_fp16)[name = tensor("causal_mask_5_cast_fp16")]; + tensor attn_mask_9 = logical_and(x = valid_mask_5_cast_fp16, y = causal_mask_5_cast_fp16)[name = tensor("attn_mask_9")]; + tensor attn_mask_11_axes_0 = const()[name = tensor("attn_mask_11_axes_0"), val = tensor([1])]; + tensor attn_mask_11 = expand_dims(axes = attn_mask_11_axes_0, x = attn_mask_9)[name = tensor("attn_mask_11")]; + tensor var_1417_transpose_x_0 = const()[name = tensor("op_1417_transpose_x_0"), val = tensor(false)]; + tensor var_1417_transpose_y_0 = const()[name = tensor("op_1417_transpose_y_0"), val = tensor(false)]; + tensor transpose_76_perm_0 = const()[name = tensor("transpose_76_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_77_perm_0 = const()[name = tensor("transpose_77_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_77 = transpose(perm = transpose_77_perm_0, x = keys_15_cast_fp16)[name = tensor("transpose_205")]; + tensor transpose_76 = transpose(perm = transpose_76_perm_0, x = q_15_cast_fp16)[name = tensor("transpose_206")]; + tensor var_1417_cast_fp16 = matmul(transpose_x = var_1417_transpose_x_0, transpose_y = var_1417_transpose_y_0, x = transpose_76, y = transpose_77)[name = tensor("op_1417_cast_fp16")]; + tensor var_1411_to_fp16 = const()[name = tensor("op_1411_to_fp16"), val = tensor([0x1p-3])]; + tensor attn_weights_13_cast_fp16 = mul(x = var_1417_cast_fp16, y = var_1411_to_fp16)[name = tensor("attn_weights_13_cast_fp16")]; + tensor var_1419 = logical_not(x = attn_mask_11)[name = tensor("op_1419")]; + tensor var_1420_to_fp16 = const()[name = tensor("op_1420_to_fp16"), val = tensor(-inf)]; + tensor attn_weights_15_cast_fp16 = select(a = var_1420_to_fp16, b = attn_weights_13_cast_fp16, cond = var_1419)[name = tensor("attn_weights_15_cast_fp16")]; + tensor var_1422 = const()[name = tensor("op_1422"), val = tensor(-1)]; + tensor attn_weights_17_cast_fp16 = softmax(axis = var_1422, x = attn_weights_15_cast_fp16)[name = tensor("attn_weights_17_cast_fp16")]; + tensor attn_output_5_transpose_x_0 = const()[name = tensor("attn_output_5_transpose_x_0"), val = tensor(false)]; + tensor attn_output_5_transpose_y_0 = const()[name = tensor("attn_output_5_transpose_y_0"), val = tensor(false)]; + tensor values_17_cast_fp16 = transpose(perm = var_1368, x = values_15_cast_fp16)[name = tensor("transpose_207")]; + tensor attn_output_5_cast_fp16 = matmul(transpose_x = attn_output_5_transpose_x_0, transpose_y = attn_output_5_transpose_y_0, x = attn_weights_17_cast_fp16, y = values_17_cast_fp16)[name = tensor("attn_output_5_cast_fp16")]; + tensor var_1430 = const()[name = tensor("op_1430"), val = tensor([0, 2, 1, 3])]; + tensor var_1433 = const()[name = tensor("op_1433"), val = tensor([1, 1, 1024])]; + tensor var_1431_cast_fp16 = transpose(perm = var_1430, x = attn_output_5_cast_fp16)[name = tensor("transpose_204")]; + tensor input_25_cast_fp16 = reshape(shape = var_1433, x = var_1431_cast_fp16)[name = tensor("input_25_cast_fp16")]; + tensor linear_10_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = attn2_out_proj_weight_cast_fp16, x = input_25_cast_fp16)[name = tensor("linear_10_cast_fp16")]; + tensor var_1439_to_fp16 = const()[name = tensor("op_1439_to_fp16"), val = tensor(0x1p+0)]; + tensor var_1440_cast_fp16 = add(x = position2_to_fp16, y = var_1439_to_fp16)[name = tensor("op_1440_cast_fp16")]; + tensor var_1440_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_1440_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor input_27_cast_fp16 = add(x = input_23_cast_fp16, y = linear_10_cast_fp16)[name = tensor("input_27_cast_fp16")]; + tensor input_29_axes_0 = const()[name = tensor("input_29_axes_0"), val = tensor([-1])]; + tensor norm2_2_weight_to_fp16 = const()[name = tensor("norm2_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(303606400)))]; + tensor norm2_2_bias_to_fp16 = const()[name = tensor("norm2_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(303608512)))]; + tensor var_1444_to_fp16 = const()[name = tensor("op_1444_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_29_cast_fp16 = layer_norm(axes = input_29_axes_0, beta = norm2_2_bias_to_fp16, epsilon = var_1444_to_fp16, gamma = norm2_2_weight_to_fp16, x = input_27_cast_fp16)[name = tensor("input_29_cast_fp16")]; + tensor linear_11_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = linear2_1_weight_cast_fp16, x = input_29_cast_fp16)[name = tensor("linear_11_cast_fp16")]; + tensor input_31_mode_0 = const()[name = tensor("input_31_mode_0"), val = tensor("EXACT")]; + tensor input_31_cast_fp16 = gelu(mode = input_31_mode_0, x = linear_11_cast_fp16)[name = tensor("input_31_cast_fp16")]; + tensor linear_12_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = linear2_2_weight_cast_fp16, x = input_31_cast_fp16)[name = tensor("linear_12_cast_fp16")]; + tensor input_33_cast_fp16 = add(x = input_27_cast_fp16, y = linear_12_cast_fp16)[name = tensor("input_33_cast_fp16")]; + tensor x_7_axes_0 = const()[name = tensor("x_7_axes_0"), val = tensor([-1])]; + tensor norm3_1_weight_to_fp16 = const()[name = tensor("norm3_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(303610624)))]; + tensor norm3_1_bias_to_fp16 = const()[name = tensor("norm3_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(303612736)))]; + tensor var_1461_to_fp16 = const()[name = tensor("op_1461_to_fp16"), val = tensor(0x1.5p-17)]; + tensor x_7_cast_fp16 = layer_norm(axes = x_7_axes_0, beta = norm3_1_bias_to_fp16, epsilon = var_1461_to_fp16, gamma = norm3_1_weight_to_fp16, x = input_33_cast_fp16)[name = tensor("x_7_cast_fp16")]; + tensor linear_13_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = attn3_in_proj_weight_cast_fp16, x = x_7_cast_fp16)[name = tensor("linear_13_cast_fp16")]; + tensor var_1497 = const()[name = tensor("op_1497"), val = tensor([1, 1, 3, 16, 64])]; + tensor qkv_7_cast_fp16 = reshape(shape = var_1497, x = linear_13_cast_fp16)[name = tensor("qkv_7_cast_fp16")]; + tensor q_19_begin_0 = const()[name = tensor("q_19_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor q_19_end_0 = const()[name = tensor("q_19_end_0"), val = tensor([1, 1, 1, 16, 64])]; + tensor q_19_end_mask_0 = const()[name = tensor("q_19_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor q_19_squeeze_mask_0 = const()[name = tensor("q_19_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor q_19_cast_fp16 = slice_by_index(begin = q_19_begin_0, end = q_19_end_0, end_mask = q_19_end_mask_0, squeeze_mask = q_19_squeeze_mask_0, x = qkv_7_cast_fp16)[name = tensor("q_19_cast_fp16")]; + tensor k_13_begin_0 = const()[name = tensor("k_13_begin_0"), val = tensor([0, 0, 1, 0, 0])]; + tensor k_13_end_0 = const()[name = tensor("k_13_end_0"), val = tensor([1, 1, 2, 16, 64])]; + tensor k_13_end_mask_0 = const()[name = tensor("k_13_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor k_13_squeeze_mask_0 = const()[name = tensor("k_13_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor k_13_cast_fp16 = slice_by_index(begin = k_13_begin_0, end = k_13_end_0, end_mask = k_13_end_mask_0, squeeze_mask = k_13_squeeze_mask_0, x = qkv_7_cast_fp16)[name = tensor("k_13_cast_fp16")]; + tensor v_7_begin_0 = const()[name = tensor("v_7_begin_0"), val = tensor([0, 0, 2, 0, 0])]; + tensor v_7_end_0 = const()[name = tensor("v_7_end_0"), val = tensor([1, 1, 3, 16, 64])]; + tensor v_7_end_mask_0 = const()[name = tensor("v_7_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor v_7_squeeze_mask_0 = const()[name = tensor("v_7_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor v_7_cast_fp16 = slice_by_index(begin = v_7_begin_0, end = v_7_end_0, end_mask = v_7_end_mask_0, squeeze_mask = v_7_squeeze_mask_0, x = qkv_7_cast_fp16)[name = tensor("v_7_cast_fp16")]; + tensor var_1601 = const()[name = tensor("op_1601"), val = tensor([1, 1, 1, 1])]; + tensor position3_to_fp16_dtype_0 = const()[name = tensor("position3_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor position3_to_fp16 = cast(dtype = position3_to_fp16_dtype_0, x = position3)[name = tensor("cast_538")]; + tensor ts_23_cast_fp16 = reshape(shape = var_1601, x = position3_to_fp16)[name = tensor("ts_23_cast_fp16")]; + tensor var_1605 = const()[name = tensor("op_1605"), val = tensor([1, 1, 16, 32, 2])]; + tensor q_complex_7_cast_fp16 = reshape(shape = var_1605, x = q_19_cast_fp16)[name = tensor("q_complex_7_cast_fp16")]; + tensor var_1609 = const()[name = tensor("op_1609"), val = tensor([1, 1, 16, 32, 2])]; + tensor k_complex_7_cast_fp16 = reshape(shape = var_1609, x = k_13_cast_fp16)[name = tensor("k_complex_7_cast_fp16")]; + tensor var_1613_begin_0 = const()[name = tensor("op_1613_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_1613_end_0 = const()[name = tensor("op_1613_end_0"), val = tensor([1, 1, 16, 32, 1])]; + tensor var_1613_end_mask_0 = const()[name = tensor("op_1613_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_1613_squeeze_mask_0 = const()[name = tensor("op_1613_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_1613_cast_fp16 = slice_by_index(begin = var_1613_begin_0, end = var_1613_end_0, end_mask = var_1613_end_mask_0, squeeze_mask = var_1613_squeeze_mask_0, x = q_complex_7_cast_fp16)[name = tensor("op_1613_cast_fp16")]; + tensor var_1621_begin_0 = const()[name = tensor("op_1621_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_1621_end_0 = const()[name = tensor("op_1621_end_0"), val = tensor([1, 1, 16, 32, 2])]; + tensor var_1621_end_mask_0 = const()[name = tensor("op_1621_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_1621_squeeze_mask_0 = const()[name = tensor("op_1621_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_1621_cast_fp16 = slice_by_index(begin = var_1621_begin_0, end = var_1621_end_0, end_mask = var_1621_end_mask_0, squeeze_mask = var_1621_squeeze_mask_0, x = q_complex_7_cast_fp16)[name = tensor("op_1621_cast_fp16")]; + tensor var_1629_begin_0 = const()[name = tensor("op_1629_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_1629_end_0 = const()[name = tensor("op_1629_end_0"), val = tensor([1, 1, 16, 32, 1])]; + tensor var_1629_end_mask_0 = const()[name = tensor("op_1629_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_1629_squeeze_mask_0 = const()[name = tensor("op_1629_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_1629_cast_fp16 = slice_by_index(begin = var_1629_begin_0, end = var_1629_end_0, end_mask = var_1629_end_mask_0, squeeze_mask = var_1629_squeeze_mask_0, x = k_complex_7_cast_fp16)[name = tensor("op_1629_cast_fp16")]; + tensor var_1637_begin_0 = const()[name = tensor("op_1637_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_1637_end_0 = const()[name = tensor("op_1637_end_0"), val = tensor([1, 1, 16, 32, 2])]; + tensor var_1637_end_mask_0 = const()[name = tensor("op_1637_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_1637_squeeze_mask_0 = const()[name = tensor("op_1637_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_1637_cast_fp16 = slice_by_index(begin = var_1637_begin_0, end = var_1637_end_0, end_mask = var_1637_end_mask_0, squeeze_mask = var_1637_squeeze_mask_0, x = k_complex_7_cast_fp16)[name = tensor("op_1637_cast_fp16")]; + tensor freqs_7_to_fp16 = const()[name = tensor("freqs_7_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(303614848)))]; + tensor var_1643_cast_fp16 = mul(x = freqs_7_to_fp16, y = ts_23_cast_fp16)[name = tensor("op_1643_cast_fp16")]; + tensor rotr_7_cast_fp16 = cos(x = var_1643_cast_fp16)[name = tensor("rotr_7_cast_fp16")]; + tensor roti_7_cast_fp16 = sin(x = var_1643_cast_fp16)[name = tensor("roti_7_cast_fp16")]; + tensor var_1647_cast_fp16 = mul(x = var_1613_cast_fp16, y = rotr_7_cast_fp16)[name = tensor("op_1647_cast_fp16")]; + tensor var_1648_cast_fp16 = mul(x = var_1621_cast_fp16, y = roti_7_cast_fp16)[name = tensor("op_1648_cast_fp16")]; + tensor qor_13_cast_fp16 = sub(x = var_1647_cast_fp16, y = var_1648_cast_fp16)[name = tensor("qor_13_cast_fp16")]; + tensor var_1651_cast_fp16 = mul(x = var_1613_cast_fp16, y = roti_7_cast_fp16)[name = tensor("op_1651_cast_fp16")]; + tensor var_1652_cast_fp16 = mul(x = var_1621_cast_fp16, y = rotr_7_cast_fp16)[name = tensor("op_1652_cast_fp16")]; + tensor qoi_13_cast_fp16 = add(x = var_1651_cast_fp16, y = var_1652_cast_fp16)[name = tensor("qoi_13_cast_fp16")]; + tensor var_1655_cast_fp16 = mul(x = var_1629_cast_fp16, y = rotr_7_cast_fp16)[name = tensor("op_1655_cast_fp16")]; + tensor var_1656_cast_fp16 = mul(x = var_1637_cast_fp16, y = roti_7_cast_fp16)[name = tensor("op_1656_cast_fp16")]; + tensor kor_13_cast_fp16 = sub(x = var_1655_cast_fp16, y = var_1656_cast_fp16)[name = tensor("kor_13_cast_fp16")]; + tensor var_1659_cast_fp16 = mul(x = var_1629_cast_fp16, y = roti_7_cast_fp16)[name = tensor("op_1659_cast_fp16")]; + tensor var_1660_cast_fp16 = mul(x = var_1637_cast_fp16, y = rotr_7_cast_fp16)[name = tensor("op_1660_cast_fp16")]; + tensor koi_13_cast_fp16 = add(x = var_1659_cast_fp16, y = var_1660_cast_fp16)[name = tensor("koi_13_cast_fp16")]; + tensor qo_7_axis_0 = const()[name = tensor("qo_7_axis_0"), val = tensor(-1)]; + tensor qo_7_cast_fp16 = stack(axis = qo_7_axis_0, values = (qor_13_cast_fp16, qoi_13_cast_fp16))[name = tensor("qo_7_cast_fp16")]; + tensor ko_7_axis_0 = const()[name = tensor("ko_7_axis_0"), val = tensor(-1)]; + tensor ko_7_cast_fp16 = stack(axis = ko_7_axis_0, values = (kor_13_cast_fp16, koi_13_cast_fp16))[name = tensor("ko_7_cast_fp16")]; + tensor var_1689 = const()[name = tensor("op_1689"), val = tensor([1, 1, 16, 64])]; + tensor q_21_cast_fp16 = reshape(shape = var_1689, x = qo_7_cast_fp16)[name = tensor("q_21_cast_fp16")]; + tensor var_1691 = const()[name = tensor("op_1691"), val = tensor([1, 1, 16, 64])]; + tensor k_15_cast_fp16 = reshape(shape = var_1691, x = ko_7_cast_fp16)[name = tensor("k_15_cast_fp16")]; + tensor _inversed_1713_y_0_to_fp16 = const()[name = tensor("_inversed_1713_y_0_to_fp16"), val = tensor(0x1p-9)]; + tensor _inversed_1713_cast_fp16 = mul(x = ts_23_cast_fp16, y = _inversed_1713_y_0_to_fp16)[name = tensor("_inversed_1713_cast_fp16")]; + tensor var_1714_cast_fp16 = floor(x = _inversed_1713_cast_fp16)[name = tensor("op_1714_cast_fp16")]; + tensor var_1715_to_fp16 = const()[name = tensor("op_1715_to_fp16"), val = tensor(0x1p+9)]; + tensor var_1716_cast_fp16 = mul(x = var_1714_cast_fp16, y = var_1715_to_fp16)[name = tensor("op_1716_cast_fp16")]; + tensor write_indices_float_15_cast_fp16 = sub(x = ts_23_cast_fp16, y = var_1716_cast_fp16)[name = tensor("write_indices_float_15_cast_fp16")]; + tensor var_1723_dtype_0 = const()[name = tensor("op_1723_dtype_0"), val = tensor("int32")]; + tensor write_indices_7_reps_0 = const()[name = tensor("write_indices_7_reps_0"), val = tensor([1, 1, 16, 64])]; + tensor write_indices_float_15_cast_fp16_to_int32 = cast(dtype = var_1723_dtype_0, x = write_indices_float_15_cast_fp16)[name = tensor("cast_537")]; + tensor write_indices_7 = tile(reps = write_indices_7_reps_0, x = write_indices_float_15_cast_fp16_to_int32)[name = tensor("write_indices_7")]; + tensor var_1731_begin_0 = const()[name = tensor("op_1731_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_1731_end_0 = const()[name = tensor("op_1731_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor var_1731_end_mask_0 = const()[name = tensor("op_1731_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_1731_squeeze_mask_0 = const()[name = tensor("op_1731_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor cache3_to_fp16_dtype_0 = const()[name = tensor("cache3_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor cache3_to_fp16 = cast(dtype = cache3_to_fp16_dtype_0, x = cache3)[name = tensor("cast_536")]; + tensor var_1731_cast_fp16 = slice_by_index(begin = var_1731_begin_0, end = var_1731_end_0, end_mask = var_1731_end_mask_0, squeeze_mask = var_1731_squeeze_mask_0, x = cache3_to_fp16)[name = tensor("op_1731_cast_fp16")]; + tensor var_1733_axis_0 = const()[name = tensor("op_1733_axis_0"), val = tensor(1)]; + tensor var_1733_mode_0 = const()[name = tensor("op_1733_mode_0"), val = tensor("update")]; + tensor var_1733_validate_indices_0 = const()[name = tensor("op_1733_validate_indices_0"), val = tensor(false)]; + tensor var_1733_cast_fp16 = scatter_along_axis(axis = var_1733_axis_0, data = var_1731_cast_fp16, indices = write_indices_7, mode = var_1733_mode_0, updates = k_15_cast_fp16, validate_indices = var_1733_validate_indices_0)[name = tensor("op_1733_cast_fp16")]; + tensor concat_23 = const()[name = tensor("concat_23"), val = tensor([0, 0, 0, 0, 0])]; + tensor concat_24 = const()[name = tensor("concat_24"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_7_internal_tensor_assign_1_stride_0 = const()[name = tensor("new_cache_7_internal_tensor_assign_1_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_7_internal_tensor_assign_1_begin_mask_0 = const()[name = tensor("new_cache_7_internal_tensor_assign_1_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_7_internal_tensor_assign_1_end_mask_0 = const()[name = tensor("new_cache_7_internal_tensor_assign_1_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_7_internal_tensor_assign_1_squeeze_mask_0 = const()[name = tensor("new_cache_7_internal_tensor_assign_1_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_54 = const()[name = tensor("shape_54"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_6 = const()[name = tensor("reduce_prod_6"), val = tensor(1048576)]; + tensor range_1d_6_start_0 = const()[name = tensor("range_1d_6_start_0"), val = tensor(0)]; + tensor range_1d_6_step_0 = const()[name = tensor("range_1d_6_step_0"), val = tensor(1)]; + tensor range_1d_6 = range_1d(end = reduce_prod_6, start = range_1d_6_start_0, step = range_1d_6_step_0)[name = tensor("range_1d_6")]; + tensor reshape_30 = reshape(shape = shape_54, x = range_1d_6)[name = tensor("reshape_30")]; + tensor slice_by_index_6 = slice_by_index(begin = concat_23, begin_mask = new_cache_7_internal_tensor_assign_1_begin_mask_0, end = concat_24, end_mask = new_cache_7_internal_tensor_assign_1_end_mask_0, squeeze_mask = new_cache_7_internal_tensor_assign_1_squeeze_mask_0, stride = new_cache_7_internal_tensor_assign_1_stride_0, x = reshape_30)[name = tensor("slice_by_index_6")]; + tensor reshape_31_shape_0 = const()[name = tensor("reshape_31_shape_0"), val = tensor([-1])]; + tensor reshape_31 = reshape(shape = reshape_31_shape_0, x = slice_by_index_6)[name = tensor("reshape_31")]; + tensor reshape_32_shape_0 = const()[name = tensor("reshape_32_shape_0"), val = tensor([-1])]; + tensor reshape_32_cast_fp16 = reshape(shape = reshape_32_shape_0, x = var_1733_cast_fp16)[name = tensor("reshape_32_cast_fp16")]; + tensor reshape_33_shape_0 = const()[name = tensor("reshape_33_shape_0"), val = tensor([-1])]; + tensor reshape_33_cast_fp16 = reshape(shape = reshape_33_shape_0, x = cache3_to_fp16)[name = tensor("reshape_33_cast_fp16")]; + tensor scatter_6_mode_0 = const()[name = tensor("scatter_6_mode_0"), val = tensor("update")]; + tensor scatter_6_axis_0 = const()[name = tensor("scatter_6_axis_0"), val = tensor(0)]; + tensor scatter_6_validate_indices_0 = const()[name = tensor("scatter_6_validate_indices_0"), val = tensor(false)]; + tensor scatter_6_cast_fp16 = scatter(axis = scatter_6_axis_0, data = reshape_33_cast_fp16, indices = reshape_31, mode = scatter_6_mode_0, updates = reshape_32_cast_fp16, validate_indices = scatter_6_validate_indices_0)[name = tensor("scatter_6_cast_fp16")]; + tensor reshape_34_cast_fp16 = reshape(shape = shape_54, x = scatter_6_cast_fp16)[name = tensor("reshape_34_cast_fp16")]; + tensor var_1741_begin_0 = const()[name = tensor("op_1741_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor var_1741_end_0 = const()[name = tensor("op_1741_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor var_1741_end_mask_0 = const()[name = tensor("op_1741_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_1741_squeeze_mask_0 = const()[name = tensor("op_1741_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor var_1741_cast_fp16 = slice_by_index(begin = var_1741_begin_0, end = var_1741_end_0, end_mask = var_1741_end_mask_0, squeeze_mask = var_1741_squeeze_mask_0, x = reshape_34_cast_fp16)[name = tensor("op_1741_cast_fp16")]; + tensor var_1743_axis_0 = const()[name = tensor("op_1743_axis_0"), val = tensor(1)]; + tensor var_1743_mode_0 = const()[name = tensor("op_1743_mode_0"), val = tensor("update")]; + tensor var_1743_validate_indices_0 = const()[name = tensor("op_1743_validate_indices_0"), val = tensor(false)]; + tensor var_1743_cast_fp16 = scatter_along_axis(axis = var_1743_axis_0, data = var_1741_cast_fp16, indices = write_indices_7, mode = var_1743_mode_0, updates = v_7_cast_fp16, validate_indices = var_1743_validate_indices_0)[name = tensor("op_1743_cast_fp16")]; + tensor concat_25 = const()[name = tensor("concat_25"), val = tensor([1, 0, 0, 0, 0])]; + tensor concat_26 = const()[name = tensor("concat_26"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_7_internal_tensor_assign_2_stride_0 = const()[name = tensor("new_cache_7_internal_tensor_assign_2_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_7_internal_tensor_assign_2_begin_mask_0 = const()[name = tensor("new_cache_7_internal_tensor_assign_2_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_7_internal_tensor_assign_2_end_mask_0 = const()[name = tensor("new_cache_7_internal_tensor_assign_2_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_7_internal_tensor_assign_2_squeeze_mask_0 = const()[name = tensor("new_cache_7_internal_tensor_assign_2_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_55 = const()[name = tensor("shape_55"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_7 = const()[name = tensor("reduce_prod_7"), val = tensor(1048576)]; + tensor range_1d_7_start_0 = const()[name = tensor("range_1d_7_start_0"), val = tensor(0)]; + tensor range_1d_7_step_0 = const()[name = tensor("range_1d_7_step_0"), val = tensor(1)]; + tensor range_1d_7 = range_1d(end = reduce_prod_7, start = range_1d_7_start_0, step = range_1d_7_step_0)[name = tensor("range_1d_7")]; + tensor reshape_35 = reshape(shape = shape_55, x = range_1d_7)[name = tensor("reshape_35")]; + tensor slice_by_index_7 = slice_by_index(begin = concat_25, begin_mask = new_cache_7_internal_tensor_assign_2_begin_mask_0, end = concat_26, end_mask = new_cache_7_internal_tensor_assign_2_end_mask_0, squeeze_mask = new_cache_7_internal_tensor_assign_2_squeeze_mask_0, stride = new_cache_7_internal_tensor_assign_2_stride_0, x = reshape_35)[name = tensor("slice_by_index_7")]; + tensor reshape_36_shape_0 = const()[name = tensor("reshape_36_shape_0"), val = tensor([-1])]; + tensor reshape_36 = reshape(shape = reshape_36_shape_0, x = slice_by_index_7)[name = tensor("reshape_36")]; + tensor reshape_37_shape_0 = const()[name = tensor("reshape_37_shape_0"), val = tensor([-1])]; + tensor reshape_37_cast_fp16 = reshape(shape = reshape_37_shape_0, x = var_1743_cast_fp16)[name = tensor("reshape_37_cast_fp16")]; + tensor reshape_38_shape_0 = const()[name = tensor("reshape_38_shape_0"), val = tensor([-1])]; + tensor reshape_38_cast_fp16 = reshape(shape = reshape_38_shape_0, x = reshape_34_cast_fp16)[name = tensor("reshape_38_cast_fp16")]; + tensor scatter_7_mode_0 = const()[name = tensor("scatter_7_mode_0"), val = tensor("update")]; + tensor scatter_7_axis_0 = const()[name = tensor("scatter_7_axis_0"), val = tensor(0)]; + tensor scatter_7_validate_indices_0 = const()[name = tensor("scatter_7_validate_indices_0"), val = tensor(false)]; + tensor scatter_7_cast_fp16 = scatter(axis = scatter_7_axis_0, data = reshape_38_cast_fp16, indices = reshape_36, mode = scatter_7_mode_0, updates = reshape_37_cast_fp16, validate_indices = scatter_7_validate_indices_0)[name = tensor("scatter_7_cast_fp16")]; + tensor reshape_39_cast_fp16 = reshape(shape = shape_55, x = scatter_7_cast_fp16)[name = tensor("reshape_39_cast_fp16")]; + tensor reshape_39_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("reshape_39_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor keys_19_begin_0 = const()[name = tensor("keys_19_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor keys_19_end_0 = const()[name = tensor("keys_19_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor keys_19_end_mask_0 = const()[name = tensor("keys_19_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor keys_19_squeeze_mask_0 = const()[name = tensor("keys_19_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor keys_19_cast_fp16 = slice_by_index(begin = keys_19_begin_0, end = keys_19_end_0, end_mask = keys_19_end_mask_0, squeeze_mask = keys_19_squeeze_mask_0, x = reshape_39_cast_fp16)[name = tensor("keys_19_cast_fp16")]; + tensor values_19_begin_0 = const()[name = tensor("values_19_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor values_19_end_0 = const()[name = tensor("values_19_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor values_19_end_mask_0 = const()[name = tensor("values_19_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor values_19_squeeze_mask_0 = const()[name = tensor("values_19_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor values_19_cast_fp16 = slice_by_index(begin = values_19_begin_0, end = values_19_end_0, end_mask = values_19_end_mask_0, squeeze_mask = values_19_squeeze_mask_0, x = reshape_39_cast_fp16)[name = tensor("values_19_cast_fp16")]; + tensor var_1755_cast_fp16 = not_equal(x = keys_19_cast_fp16, y = keys_19_cast_fp16)[name = tensor("op_1755_cast_fp16")]; + tensor keys_21_cast_fp16 = select(a = var_504_to_fp16, b = keys_19_cast_fp16, cond = var_1755_cast_fp16)[name = tensor("keys_21_cast_fp16")]; + tensor var_1763_cast_fp16 = not_equal(x = values_19_cast_fp16, y = values_19_cast_fp16)[name = tensor("op_1763_cast_fp16")]; + tensor values_21_cast_fp16 = select(a = var_504_to_fp16, b = values_19_cast_fp16, cond = var_1763_cast_fp16)[name = tensor("values_21_cast_fp16")]; + tensor var_1787 = const()[name = tensor("op_1787"), val = tensor([0, 2, 1, 3])]; + tensor var_1800 = const()[name = tensor("op_1800"), val = tensor([1, 1, 1])]; + tensor var_1801_cast_fp16 = reshape(shape = var_1800, x = position3_to_fp16)[name = tensor("op_1801_cast_fp16")]; + tensor var_1818_to_fp16 = const()[name = tensor("op_1818_to_fp16"), val = tensor(0x1p+0)]; + tensor valid_len_7_cast_fp16 = add(x = var_1801_cast_fp16, y = var_1818_to_fp16)[name = tensor("valid_len_7_cast_fp16")]; + tensor valid_mask_7_cast_fp16 = less(x = k_positions_1_promoted_to_fp16, y = valid_len_7_cast_fp16)[name = tensor("valid_mask_7_cast_fp16")]; + tensor causal_mask_7_cast_fp16 = less_equal(x = k_positions_1_promoted_to_fp16, y = var_1801_cast_fp16)[name = tensor("causal_mask_7_cast_fp16")]; + tensor attn_mask_13 = logical_and(x = valid_mask_7_cast_fp16, y = causal_mask_7_cast_fp16)[name = tensor("attn_mask_13")]; + tensor attn_mask_15_axes_0 = const()[name = tensor("attn_mask_15_axes_0"), val = tensor([1])]; + tensor attn_mask_15 = expand_dims(axes = attn_mask_15_axes_0, x = attn_mask_13)[name = tensor("attn_mask_15")]; + tensor var_1836_transpose_x_0 = const()[name = tensor("op_1836_transpose_x_0"), val = tensor(false)]; + tensor var_1836_transpose_y_0 = const()[name = tensor("op_1836_transpose_y_0"), val = tensor(false)]; + tensor transpose_78_perm_0 = const()[name = tensor("transpose_78_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_79_perm_0 = const()[name = tensor("transpose_79_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_79 = transpose(perm = transpose_79_perm_0, x = keys_21_cast_fp16)[name = tensor("transpose_201")]; + tensor transpose_78 = transpose(perm = transpose_78_perm_0, x = q_21_cast_fp16)[name = tensor("transpose_202")]; + tensor var_1836_cast_fp16 = matmul(transpose_x = var_1836_transpose_x_0, transpose_y = var_1836_transpose_y_0, x = transpose_78, y = transpose_79)[name = tensor("op_1836_cast_fp16")]; + tensor var_1830_to_fp16 = const()[name = tensor("op_1830_to_fp16"), val = tensor([0x1p-3])]; + tensor attn_weights_19_cast_fp16 = mul(x = var_1836_cast_fp16, y = var_1830_to_fp16)[name = tensor("attn_weights_19_cast_fp16")]; + tensor var_1838 = logical_not(x = attn_mask_15)[name = tensor("op_1838")]; + tensor var_1839_to_fp16 = const()[name = tensor("op_1839_to_fp16"), val = tensor(-inf)]; + tensor attn_weights_21_cast_fp16 = select(a = var_1839_to_fp16, b = attn_weights_19_cast_fp16, cond = var_1838)[name = tensor("attn_weights_21_cast_fp16")]; + tensor var_1841 = const()[name = tensor("op_1841"), val = tensor(-1)]; + tensor attn_weights_23_cast_fp16 = softmax(axis = var_1841, x = attn_weights_21_cast_fp16)[name = tensor("attn_weights_23_cast_fp16")]; + tensor attn_output_7_transpose_x_0 = const()[name = tensor("attn_output_7_transpose_x_0"), val = tensor(false)]; + tensor attn_output_7_transpose_y_0 = const()[name = tensor("attn_output_7_transpose_y_0"), val = tensor(false)]; + tensor values_23_cast_fp16 = transpose(perm = var_1787, x = values_21_cast_fp16)[name = tensor("transpose_203")]; + tensor attn_output_7_cast_fp16 = matmul(transpose_x = attn_output_7_transpose_x_0, transpose_y = attn_output_7_transpose_y_0, x = attn_weights_23_cast_fp16, y = values_23_cast_fp16)[name = tensor("attn_output_7_cast_fp16")]; + tensor var_1849 = const()[name = tensor("op_1849"), val = tensor([0, 2, 1, 3])]; + tensor var_1852 = const()[name = tensor("op_1852"), val = tensor([1, 1, 1024])]; + tensor var_1850_cast_fp16 = transpose(perm = var_1849, x = attn_output_7_cast_fp16)[name = tensor("transpose_200")]; + tensor input_35_cast_fp16 = reshape(shape = var_1852, x = var_1850_cast_fp16)[name = tensor("input_35_cast_fp16")]; + tensor linear_14_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = attn3_out_proj_weight_cast_fp16, x = input_35_cast_fp16)[name = tensor("linear_14_cast_fp16")]; + tensor var_1858_to_fp16 = const()[name = tensor("op_1858_to_fp16"), val = tensor(0x1p+0)]; + tensor var_1859_cast_fp16 = add(x = position3_to_fp16, y = var_1858_to_fp16)[name = tensor("op_1859_cast_fp16")]; + tensor var_1859_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_1859_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor input_37_cast_fp16 = add(x = input_33_cast_fp16, y = linear_14_cast_fp16)[name = tensor("input_37_cast_fp16")]; + tensor input_39_axes_0 = const()[name = tensor("input_39_axes_0"), val = tensor([-1])]; + tensor norm3_2_weight_to_fp16 = const()[name = tensor("norm3_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(303614976)))]; + tensor norm3_2_bias_to_fp16 = const()[name = tensor("norm3_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(303617088)))]; + tensor var_1863_to_fp16 = const()[name = tensor("op_1863_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_39_cast_fp16 = layer_norm(axes = input_39_axes_0, beta = norm3_2_bias_to_fp16, epsilon = var_1863_to_fp16, gamma = norm3_2_weight_to_fp16, x = input_37_cast_fp16)[name = tensor("input_39_cast_fp16")]; + tensor linear_15_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = linear3_1_weight_cast_fp16, x = input_39_cast_fp16)[name = tensor("linear_15_cast_fp16")]; + tensor input_41_mode_0 = const()[name = tensor("input_41_mode_0"), val = tensor("EXACT")]; + tensor input_41_cast_fp16 = gelu(mode = input_41_mode_0, x = linear_15_cast_fp16)[name = tensor("input_41_cast_fp16")]; + tensor linear_16_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = linear3_2_weight_cast_fp16, x = input_41_cast_fp16)[name = tensor("linear_16_cast_fp16")]; + tensor input_43_cast_fp16 = add(x = input_37_cast_fp16, y = linear_16_cast_fp16)[name = tensor("input_43_cast_fp16")]; + tensor x_9_axes_0 = const()[name = tensor("x_9_axes_0"), val = tensor([-1])]; + tensor norm4_1_weight_to_fp16 = const()[name = tensor("norm4_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(303619200)))]; + tensor norm4_1_bias_to_fp16 = const()[name = tensor("norm4_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(303621312)))]; + tensor var_1880_to_fp16 = const()[name = tensor("op_1880_to_fp16"), val = tensor(0x1.5p-17)]; + tensor x_9_cast_fp16 = layer_norm(axes = x_9_axes_0, beta = norm4_1_bias_to_fp16, epsilon = var_1880_to_fp16, gamma = norm4_1_weight_to_fp16, x = input_43_cast_fp16)[name = tensor("x_9_cast_fp16")]; + tensor linear_17_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = attn4_in_proj_weight_cast_fp16, x = x_9_cast_fp16)[name = tensor("linear_17_cast_fp16")]; + tensor var_1916 = const()[name = tensor("op_1916"), val = tensor([1, 1, 3, 16, 64])]; + tensor qkv_9_cast_fp16 = reshape(shape = var_1916, x = linear_17_cast_fp16)[name = tensor("qkv_9_cast_fp16")]; + tensor q_25_begin_0 = const()[name = tensor("q_25_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor q_25_end_0 = const()[name = tensor("q_25_end_0"), val = tensor([1, 1, 1, 16, 64])]; + tensor q_25_end_mask_0 = const()[name = tensor("q_25_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor q_25_squeeze_mask_0 = const()[name = tensor("q_25_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor q_25_cast_fp16 = slice_by_index(begin = q_25_begin_0, end = q_25_end_0, end_mask = q_25_end_mask_0, squeeze_mask = q_25_squeeze_mask_0, x = qkv_9_cast_fp16)[name = tensor("q_25_cast_fp16")]; + tensor k_17_begin_0 = const()[name = tensor("k_17_begin_0"), val = tensor([0, 0, 1, 0, 0])]; + tensor k_17_end_0 = const()[name = tensor("k_17_end_0"), val = tensor([1, 1, 2, 16, 64])]; + tensor k_17_end_mask_0 = const()[name = tensor("k_17_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor k_17_squeeze_mask_0 = const()[name = tensor("k_17_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor k_17_cast_fp16 = slice_by_index(begin = k_17_begin_0, end = k_17_end_0, end_mask = k_17_end_mask_0, squeeze_mask = k_17_squeeze_mask_0, x = qkv_9_cast_fp16)[name = tensor("k_17_cast_fp16")]; + tensor v_9_begin_0 = const()[name = tensor("v_9_begin_0"), val = tensor([0, 0, 2, 0, 0])]; + tensor v_9_end_0 = const()[name = tensor("v_9_end_0"), val = tensor([1, 1, 3, 16, 64])]; + tensor v_9_end_mask_0 = const()[name = tensor("v_9_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor v_9_squeeze_mask_0 = const()[name = tensor("v_9_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor v_9_cast_fp16 = slice_by_index(begin = v_9_begin_0, end = v_9_end_0, end_mask = v_9_end_mask_0, squeeze_mask = v_9_squeeze_mask_0, x = qkv_9_cast_fp16)[name = tensor("v_9_cast_fp16")]; + tensor var_2020 = const()[name = tensor("op_2020"), val = tensor([1, 1, 1, 1])]; + tensor position4_to_fp16_dtype_0 = const()[name = tensor("position4_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor position4_to_fp16 = cast(dtype = position4_to_fp16_dtype_0, x = position4)[name = tensor("cast_533")]; + tensor ts_29_cast_fp16 = reshape(shape = var_2020, x = position4_to_fp16)[name = tensor("ts_29_cast_fp16")]; + tensor var_2024 = const()[name = tensor("op_2024"), val = tensor([1, 1, 16, 32, 2])]; + tensor q_complex_9_cast_fp16 = reshape(shape = var_2024, x = q_25_cast_fp16)[name = tensor("q_complex_9_cast_fp16")]; + tensor var_2028 = const()[name = tensor("op_2028"), val = tensor([1, 1, 16, 32, 2])]; + tensor k_complex_9_cast_fp16 = reshape(shape = var_2028, x = k_17_cast_fp16)[name = tensor("k_complex_9_cast_fp16")]; + tensor var_2032_begin_0 = const()[name = tensor("op_2032_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_2032_end_0 = const()[name = tensor("op_2032_end_0"), val = tensor([1, 1, 16, 32, 1])]; + tensor var_2032_end_mask_0 = const()[name = tensor("op_2032_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_2032_squeeze_mask_0 = const()[name = tensor("op_2032_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_2032_cast_fp16 = slice_by_index(begin = var_2032_begin_0, end = var_2032_end_0, end_mask = var_2032_end_mask_0, squeeze_mask = var_2032_squeeze_mask_0, x = q_complex_9_cast_fp16)[name = tensor("op_2032_cast_fp16")]; + tensor var_2040_begin_0 = const()[name = tensor("op_2040_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_2040_end_0 = const()[name = tensor("op_2040_end_0"), val = tensor([1, 1, 16, 32, 2])]; + tensor var_2040_end_mask_0 = const()[name = tensor("op_2040_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_2040_squeeze_mask_0 = const()[name = tensor("op_2040_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_2040_cast_fp16 = slice_by_index(begin = var_2040_begin_0, end = var_2040_end_0, end_mask = var_2040_end_mask_0, squeeze_mask = var_2040_squeeze_mask_0, x = q_complex_9_cast_fp16)[name = tensor("op_2040_cast_fp16")]; + tensor var_2048_begin_0 = const()[name = tensor("op_2048_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_2048_end_0 = const()[name = tensor("op_2048_end_0"), val = tensor([1, 1, 16, 32, 1])]; + tensor var_2048_end_mask_0 = const()[name = tensor("op_2048_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_2048_squeeze_mask_0 = const()[name = tensor("op_2048_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_2048_cast_fp16 = slice_by_index(begin = var_2048_begin_0, end = var_2048_end_0, end_mask = var_2048_end_mask_0, squeeze_mask = var_2048_squeeze_mask_0, x = k_complex_9_cast_fp16)[name = tensor("op_2048_cast_fp16")]; + tensor var_2056_begin_0 = const()[name = tensor("op_2056_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_2056_end_0 = const()[name = tensor("op_2056_end_0"), val = tensor([1, 1, 16, 32, 2])]; + tensor var_2056_end_mask_0 = const()[name = tensor("op_2056_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_2056_squeeze_mask_0 = const()[name = tensor("op_2056_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_2056_cast_fp16 = slice_by_index(begin = var_2056_begin_0, end = var_2056_end_0, end_mask = var_2056_end_mask_0, squeeze_mask = var_2056_squeeze_mask_0, x = k_complex_9_cast_fp16)[name = tensor("op_2056_cast_fp16")]; + tensor freqs_9_to_fp16 = const()[name = tensor("freqs_9_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(303623424)))]; + tensor var_2062_cast_fp16 = mul(x = freqs_9_to_fp16, y = ts_29_cast_fp16)[name = tensor("op_2062_cast_fp16")]; + tensor rotr_9_cast_fp16 = cos(x = var_2062_cast_fp16)[name = tensor("rotr_9_cast_fp16")]; + tensor roti_9_cast_fp16 = sin(x = var_2062_cast_fp16)[name = tensor("roti_9_cast_fp16")]; + tensor var_2066_cast_fp16 = mul(x = var_2032_cast_fp16, y = rotr_9_cast_fp16)[name = tensor("op_2066_cast_fp16")]; + tensor var_2067_cast_fp16 = mul(x = var_2040_cast_fp16, y = roti_9_cast_fp16)[name = tensor("op_2067_cast_fp16")]; + tensor qor_17_cast_fp16 = sub(x = var_2066_cast_fp16, y = var_2067_cast_fp16)[name = tensor("qor_17_cast_fp16")]; + tensor var_2070_cast_fp16 = mul(x = var_2032_cast_fp16, y = roti_9_cast_fp16)[name = tensor("op_2070_cast_fp16")]; + tensor var_2071_cast_fp16 = mul(x = var_2040_cast_fp16, y = rotr_9_cast_fp16)[name = tensor("op_2071_cast_fp16")]; + tensor qoi_17_cast_fp16 = add(x = var_2070_cast_fp16, y = var_2071_cast_fp16)[name = tensor("qoi_17_cast_fp16")]; + tensor var_2074_cast_fp16 = mul(x = var_2048_cast_fp16, y = rotr_9_cast_fp16)[name = tensor("op_2074_cast_fp16")]; + tensor var_2075_cast_fp16 = mul(x = var_2056_cast_fp16, y = roti_9_cast_fp16)[name = tensor("op_2075_cast_fp16")]; + tensor kor_17_cast_fp16 = sub(x = var_2074_cast_fp16, y = var_2075_cast_fp16)[name = tensor("kor_17_cast_fp16")]; + tensor var_2078_cast_fp16 = mul(x = var_2048_cast_fp16, y = roti_9_cast_fp16)[name = tensor("op_2078_cast_fp16")]; + tensor var_2079_cast_fp16 = mul(x = var_2056_cast_fp16, y = rotr_9_cast_fp16)[name = tensor("op_2079_cast_fp16")]; + tensor koi_17_cast_fp16 = add(x = var_2078_cast_fp16, y = var_2079_cast_fp16)[name = tensor("koi_17_cast_fp16")]; + tensor qo_9_axis_0 = const()[name = tensor("qo_9_axis_0"), val = tensor(-1)]; + tensor qo_9_cast_fp16 = stack(axis = qo_9_axis_0, values = (qor_17_cast_fp16, qoi_17_cast_fp16))[name = tensor("qo_9_cast_fp16")]; + tensor ko_9_axis_0 = const()[name = tensor("ko_9_axis_0"), val = tensor(-1)]; + tensor ko_9_cast_fp16 = stack(axis = ko_9_axis_0, values = (kor_17_cast_fp16, koi_17_cast_fp16))[name = tensor("ko_9_cast_fp16")]; + tensor var_2108 = const()[name = tensor("op_2108"), val = tensor([1, 1, 16, 64])]; + tensor q_27_cast_fp16 = reshape(shape = var_2108, x = qo_9_cast_fp16)[name = tensor("q_27_cast_fp16")]; + tensor var_2110 = const()[name = tensor("op_2110"), val = tensor([1, 1, 16, 64])]; + tensor k_19_cast_fp16 = reshape(shape = var_2110, x = ko_9_cast_fp16)[name = tensor("k_19_cast_fp16")]; + tensor _inversed_2132_y_0_to_fp16 = const()[name = tensor("_inversed_2132_y_0_to_fp16"), val = tensor(0x1p-9)]; + tensor _inversed_2132_cast_fp16 = mul(x = ts_29_cast_fp16, y = _inversed_2132_y_0_to_fp16)[name = tensor("_inversed_2132_cast_fp16")]; + tensor var_2133_cast_fp16 = floor(x = _inversed_2132_cast_fp16)[name = tensor("op_2133_cast_fp16")]; + tensor var_2134_to_fp16 = const()[name = tensor("op_2134_to_fp16"), val = tensor(0x1p+9)]; + tensor var_2135_cast_fp16 = mul(x = var_2133_cast_fp16, y = var_2134_to_fp16)[name = tensor("op_2135_cast_fp16")]; + tensor write_indices_float_19_cast_fp16 = sub(x = ts_29_cast_fp16, y = var_2135_cast_fp16)[name = tensor("write_indices_float_19_cast_fp16")]; + tensor var_2142_dtype_0 = const()[name = tensor("op_2142_dtype_0"), val = tensor("int32")]; + tensor write_indices_9_reps_0 = const()[name = tensor("write_indices_9_reps_0"), val = tensor([1, 1, 16, 64])]; + tensor write_indices_float_19_cast_fp16_to_int32 = cast(dtype = var_2142_dtype_0, x = write_indices_float_19_cast_fp16)[name = tensor("cast_532")]; + tensor write_indices_9 = tile(reps = write_indices_9_reps_0, x = write_indices_float_19_cast_fp16_to_int32)[name = tensor("write_indices_9")]; + tensor var_2150_begin_0 = const()[name = tensor("op_2150_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_2150_end_0 = const()[name = tensor("op_2150_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor var_2150_end_mask_0 = const()[name = tensor("op_2150_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_2150_squeeze_mask_0 = const()[name = tensor("op_2150_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor cache4_to_fp16_dtype_0 = const()[name = tensor("cache4_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor cache4_to_fp16 = cast(dtype = cache4_to_fp16_dtype_0, x = cache4)[name = tensor("cast_531")]; + tensor var_2150_cast_fp16 = slice_by_index(begin = var_2150_begin_0, end = var_2150_end_0, end_mask = var_2150_end_mask_0, squeeze_mask = var_2150_squeeze_mask_0, x = cache4_to_fp16)[name = tensor("op_2150_cast_fp16")]; + tensor var_2152_axis_0 = const()[name = tensor("op_2152_axis_0"), val = tensor(1)]; + tensor var_2152_mode_0 = const()[name = tensor("op_2152_mode_0"), val = tensor("update")]; + tensor var_2152_validate_indices_0 = const()[name = tensor("op_2152_validate_indices_0"), val = tensor(false)]; + tensor var_2152_cast_fp16 = scatter_along_axis(axis = var_2152_axis_0, data = var_2150_cast_fp16, indices = write_indices_9, mode = var_2152_mode_0, updates = k_19_cast_fp16, validate_indices = var_2152_validate_indices_0)[name = tensor("op_2152_cast_fp16")]; + tensor concat_30 = const()[name = tensor("concat_30"), val = tensor([0, 0, 0, 0, 0])]; + tensor concat_31 = const()[name = tensor("concat_31"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_9_internal_tensor_assign_1_stride_0 = const()[name = tensor("new_cache_9_internal_tensor_assign_1_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_9_internal_tensor_assign_1_begin_mask_0 = const()[name = tensor("new_cache_9_internal_tensor_assign_1_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_9_internal_tensor_assign_1_end_mask_0 = const()[name = tensor("new_cache_9_internal_tensor_assign_1_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_9_internal_tensor_assign_1_squeeze_mask_0 = const()[name = tensor("new_cache_9_internal_tensor_assign_1_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_56 = const()[name = tensor("shape_56"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_8 = const()[name = tensor("reduce_prod_8"), val = tensor(1048576)]; + tensor range_1d_8_start_0 = const()[name = tensor("range_1d_8_start_0"), val = tensor(0)]; + tensor range_1d_8_step_0 = const()[name = tensor("range_1d_8_step_0"), val = tensor(1)]; + tensor range_1d_8 = range_1d(end = reduce_prod_8, start = range_1d_8_start_0, step = range_1d_8_step_0)[name = tensor("range_1d_8")]; + tensor reshape_40 = reshape(shape = shape_56, x = range_1d_8)[name = tensor("reshape_40")]; + tensor slice_by_index_8 = slice_by_index(begin = concat_30, begin_mask = new_cache_9_internal_tensor_assign_1_begin_mask_0, end = concat_31, end_mask = new_cache_9_internal_tensor_assign_1_end_mask_0, squeeze_mask = new_cache_9_internal_tensor_assign_1_squeeze_mask_0, stride = new_cache_9_internal_tensor_assign_1_stride_0, x = reshape_40)[name = tensor("slice_by_index_8")]; + tensor reshape_41_shape_0 = const()[name = tensor("reshape_41_shape_0"), val = tensor([-1])]; + tensor reshape_41 = reshape(shape = reshape_41_shape_0, x = slice_by_index_8)[name = tensor("reshape_41")]; + tensor reshape_42_shape_0 = const()[name = tensor("reshape_42_shape_0"), val = tensor([-1])]; + tensor reshape_42_cast_fp16 = reshape(shape = reshape_42_shape_0, x = var_2152_cast_fp16)[name = tensor("reshape_42_cast_fp16")]; + tensor reshape_43_shape_0 = const()[name = tensor("reshape_43_shape_0"), val = tensor([-1])]; + tensor reshape_43_cast_fp16 = reshape(shape = reshape_43_shape_0, x = cache4_to_fp16)[name = tensor("reshape_43_cast_fp16")]; + tensor scatter_8_mode_0 = const()[name = tensor("scatter_8_mode_0"), val = tensor("update")]; + tensor scatter_8_axis_0 = const()[name = tensor("scatter_8_axis_0"), val = tensor(0)]; + tensor scatter_8_validate_indices_0 = const()[name = tensor("scatter_8_validate_indices_0"), val = tensor(false)]; + tensor scatter_8_cast_fp16 = scatter(axis = scatter_8_axis_0, data = reshape_43_cast_fp16, indices = reshape_41, mode = scatter_8_mode_0, updates = reshape_42_cast_fp16, validate_indices = scatter_8_validate_indices_0)[name = tensor("scatter_8_cast_fp16")]; + tensor reshape_44_cast_fp16 = reshape(shape = shape_56, x = scatter_8_cast_fp16)[name = tensor("reshape_44_cast_fp16")]; + tensor var_2160_begin_0 = const()[name = tensor("op_2160_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor var_2160_end_0 = const()[name = tensor("op_2160_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor var_2160_end_mask_0 = const()[name = tensor("op_2160_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_2160_squeeze_mask_0 = const()[name = tensor("op_2160_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor var_2160_cast_fp16 = slice_by_index(begin = var_2160_begin_0, end = var_2160_end_0, end_mask = var_2160_end_mask_0, squeeze_mask = var_2160_squeeze_mask_0, x = reshape_44_cast_fp16)[name = tensor("op_2160_cast_fp16")]; + tensor var_2162_axis_0 = const()[name = tensor("op_2162_axis_0"), val = tensor(1)]; + tensor var_2162_mode_0 = const()[name = tensor("op_2162_mode_0"), val = tensor("update")]; + tensor var_2162_validate_indices_0 = const()[name = tensor("op_2162_validate_indices_0"), val = tensor(false)]; + tensor var_2162_cast_fp16 = scatter_along_axis(axis = var_2162_axis_0, data = var_2160_cast_fp16, indices = write_indices_9, mode = var_2162_mode_0, updates = v_9_cast_fp16, validate_indices = var_2162_validate_indices_0)[name = tensor("op_2162_cast_fp16")]; + tensor concat_32 = const()[name = tensor("concat_32"), val = tensor([1, 0, 0, 0, 0])]; + tensor concat_33 = const()[name = tensor("concat_33"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_9_internal_tensor_assign_2_stride_0 = const()[name = tensor("new_cache_9_internal_tensor_assign_2_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_9_internal_tensor_assign_2_begin_mask_0 = const()[name = tensor("new_cache_9_internal_tensor_assign_2_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_9_internal_tensor_assign_2_end_mask_0 = const()[name = tensor("new_cache_9_internal_tensor_assign_2_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_9_internal_tensor_assign_2_squeeze_mask_0 = const()[name = tensor("new_cache_9_internal_tensor_assign_2_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_57 = const()[name = tensor("shape_57"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_9 = const()[name = tensor("reduce_prod_9"), val = tensor(1048576)]; + tensor range_1d_9_start_0 = const()[name = tensor("range_1d_9_start_0"), val = tensor(0)]; + tensor range_1d_9_step_0 = const()[name = tensor("range_1d_9_step_0"), val = tensor(1)]; + tensor range_1d_9 = range_1d(end = reduce_prod_9, start = range_1d_9_start_0, step = range_1d_9_step_0)[name = tensor("range_1d_9")]; + tensor reshape_45 = reshape(shape = shape_57, x = range_1d_9)[name = tensor("reshape_45")]; + tensor slice_by_index_9 = slice_by_index(begin = concat_32, begin_mask = new_cache_9_internal_tensor_assign_2_begin_mask_0, end = concat_33, end_mask = new_cache_9_internal_tensor_assign_2_end_mask_0, squeeze_mask = new_cache_9_internal_tensor_assign_2_squeeze_mask_0, stride = new_cache_9_internal_tensor_assign_2_stride_0, x = reshape_45)[name = tensor("slice_by_index_9")]; + tensor reshape_46_shape_0 = const()[name = tensor("reshape_46_shape_0"), val = tensor([-1])]; + tensor reshape_46 = reshape(shape = reshape_46_shape_0, x = slice_by_index_9)[name = tensor("reshape_46")]; + tensor reshape_47_shape_0 = const()[name = tensor("reshape_47_shape_0"), val = tensor([-1])]; + tensor reshape_47_cast_fp16 = reshape(shape = reshape_47_shape_0, x = var_2162_cast_fp16)[name = tensor("reshape_47_cast_fp16")]; + tensor reshape_48_shape_0 = const()[name = tensor("reshape_48_shape_0"), val = tensor([-1])]; + tensor reshape_48_cast_fp16 = reshape(shape = reshape_48_shape_0, x = reshape_44_cast_fp16)[name = tensor("reshape_48_cast_fp16")]; + tensor scatter_9_mode_0 = const()[name = tensor("scatter_9_mode_0"), val = tensor("update")]; + tensor scatter_9_axis_0 = const()[name = tensor("scatter_9_axis_0"), val = tensor(0)]; + tensor scatter_9_validate_indices_0 = const()[name = tensor("scatter_9_validate_indices_0"), val = tensor(false)]; + tensor scatter_9_cast_fp16 = scatter(axis = scatter_9_axis_0, data = reshape_48_cast_fp16, indices = reshape_46, mode = scatter_9_mode_0, updates = reshape_47_cast_fp16, validate_indices = scatter_9_validate_indices_0)[name = tensor("scatter_9_cast_fp16")]; + tensor reshape_49_cast_fp16 = reshape(shape = shape_57, x = scatter_9_cast_fp16)[name = tensor("reshape_49_cast_fp16")]; + tensor reshape_49_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("reshape_49_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor keys_25_begin_0 = const()[name = tensor("keys_25_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor keys_25_end_0 = const()[name = tensor("keys_25_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor keys_25_end_mask_0 = const()[name = tensor("keys_25_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor keys_25_squeeze_mask_0 = const()[name = tensor("keys_25_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor keys_25_cast_fp16 = slice_by_index(begin = keys_25_begin_0, end = keys_25_end_0, end_mask = keys_25_end_mask_0, squeeze_mask = keys_25_squeeze_mask_0, x = reshape_49_cast_fp16)[name = tensor("keys_25_cast_fp16")]; + tensor values_25_begin_0 = const()[name = tensor("values_25_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor values_25_end_0 = const()[name = tensor("values_25_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor values_25_end_mask_0 = const()[name = tensor("values_25_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor values_25_squeeze_mask_0 = const()[name = tensor("values_25_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor values_25_cast_fp16 = slice_by_index(begin = values_25_begin_0, end = values_25_end_0, end_mask = values_25_end_mask_0, squeeze_mask = values_25_squeeze_mask_0, x = reshape_49_cast_fp16)[name = tensor("values_25_cast_fp16")]; + tensor var_2174_cast_fp16 = not_equal(x = keys_25_cast_fp16, y = keys_25_cast_fp16)[name = tensor("op_2174_cast_fp16")]; + tensor keys_27_cast_fp16 = select(a = var_504_to_fp16, b = keys_25_cast_fp16, cond = var_2174_cast_fp16)[name = tensor("keys_27_cast_fp16")]; + tensor var_2182_cast_fp16 = not_equal(x = values_25_cast_fp16, y = values_25_cast_fp16)[name = tensor("op_2182_cast_fp16")]; + tensor values_27_cast_fp16 = select(a = var_504_to_fp16, b = values_25_cast_fp16, cond = var_2182_cast_fp16)[name = tensor("values_27_cast_fp16")]; + tensor var_2206 = const()[name = tensor("op_2206"), val = tensor([0, 2, 1, 3])]; + tensor var_2219 = const()[name = tensor("op_2219"), val = tensor([1, 1, 1])]; + tensor var_2220_cast_fp16 = reshape(shape = var_2219, x = position4_to_fp16)[name = tensor("op_2220_cast_fp16")]; + tensor var_2237_to_fp16 = const()[name = tensor("op_2237_to_fp16"), val = tensor(0x1p+0)]; + tensor valid_len_9_cast_fp16 = add(x = var_2220_cast_fp16, y = var_2237_to_fp16)[name = tensor("valid_len_9_cast_fp16")]; + tensor valid_mask_9_cast_fp16 = less(x = k_positions_1_promoted_to_fp16, y = valid_len_9_cast_fp16)[name = tensor("valid_mask_9_cast_fp16")]; + tensor causal_mask_9_cast_fp16 = less_equal(x = k_positions_1_promoted_to_fp16, y = var_2220_cast_fp16)[name = tensor("causal_mask_9_cast_fp16")]; + tensor attn_mask_17 = logical_and(x = valid_mask_9_cast_fp16, y = causal_mask_9_cast_fp16)[name = tensor("attn_mask_17")]; + tensor attn_mask_19_axes_0 = const()[name = tensor("attn_mask_19_axes_0"), val = tensor([1])]; + tensor attn_mask_19 = expand_dims(axes = attn_mask_19_axes_0, x = attn_mask_17)[name = tensor("attn_mask_19")]; + tensor var_2255_transpose_x_0 = const()[name = tensor("op_2255_transpose_x_0"), val = tensor(false)]; + tensor var_2255_transpose_y_0 = const()[name = tensor("op_2255_transpose_y_0"), val = tensor(false)]; + tensor transpose_80_perm_0 = const()[name = tensor("transpose_80_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_81_perm_0 = const()[name = tensor("transpose_81_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_81 = transpose(perm = transpose_81_perm_0, x = keys_27_cast_fp16)[name = tensor("transpose_197")]; + tensor transpose_80 = transpose(perm = transpose_80_perm_0, x = q_27_cast_fp16)[name = tensor("transpose_198")]; + tensor var_2255_cast_fp16 = matmul(transpose_x = var_2255_transpose_x_0, transpose_y = var_2255_transpose_y_0, x = transpose_80, y = transpose_81)[name = tensor("op_2255_cast_fp16")]; + tensor var_2249_to_fp16 = const()[name = tensor("op_2249_to_fp16"), val = tensor([0x1p-3])]; + tensor attn_weights_25_cast_fp16 = mul(x = var_2255_cast_fp16, y = var_2249_to_fp16)[name = tensor("attn_weights_25_cast_fp16")]; + tensor var_2257 = logical_not(x = attn_mask_19)[name = tensor("op_2257")]; + tensor var_2258_to_fp16 = const()[name = tensor("op_2258_to_fp16"), val = tensor(-inf)]; + tensor attn_weights_27_cast_fp16 = select(a = var_2258_to_fp16, b = attn_weights_25_cast_fp16, cond = var_2257)[name = tensor("attn_weights_27_cast_fp16")]; + tensor var_2260 = const()[name = tensor("op_2260"), val = tensor(-1)]; + tensor attn_weights_29_cast_fp16 = softmax(axis = var_2260, x = attn_weights_27_cast_fp16)[name = tensor("attn_weights_29_cast_fp16")]; + tensor attn_output_9_transpose_x_0 = const()[name = tensor("attn_output_9_transpose_x_0"), val = tensor(false)]; + tensor attn_output_9_transpose_y_0 = const()[name = tensor("attn_output_9_transpose_y_0"), val = tensor(false)]; + tensor values_29_cast_fp16 = transpose(perm = var_2206, x = values_27_cast_fp16)[name = tensor("transpose_199")]; + tensor attn_output_9_cast_fp16 = matmul(transpose_x = attn_output_9_transpose_x_0, transpose_y = attn_output_9_transpose_y_0, x = attn_weights_29_cast_fp16, y = values_29_cast_fp16)[name = tensor("attn_output_9_cast_fp16")]; + tensor var_2268 = const()[name = tensor("op_2268"), val = tensor([0, 2, 1, 3])]; + tensor var_2271 = const()[name = tensor("op_2271"), val = tensor([1, 1, 1024])]; + tensor var_2269_cast_fp16 = transpose(perm = var_2268, x = attn_output_9_cast_fp16)[name = tensor("transpose_196")]; + tensor input_45_cast_fp16 = reshape(shape = var_2271, x = var_2269_cast_fp16)[name = tensor("input_45_cast_fp16")]; + tensor linear_18_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = attn4_out_proj_weight_cast_fp16, x = input_45_cast_fp16)[name = tensor("linear_18_cast_fp16")]; + tensor var_2277_to_fp16 = const()[name = tensor("op_2277_to_fp16"), val = tensor(0x1p+0)]; + tensor var_2278_cast_fp16 = add(x = position4_to_fp16, y = var_2277_to_fp16)[name = tensor("op_2278_cast_fp16")]; + tensor var_2278_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_2278_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor input_47_cast_fp16 = add(x = input_43_cast_fp16, y = linear_18_cast_fp16)[name = tensor("input_47_cast_fp16")]; + tensor input_49_axes_0 = const()[name = tensor("input_49_axes_0"), val = tensor([-1])]; + tensor norm4_2_weight_to_fp16 = const()[name = tensor("norm4_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(303623552)))]; + tensor norm4_2_bias_to_fp16 = const()[name = tensor("norm4_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(303625664)))]; + tensor var_2282_to_fp16 = const()[name = tensor("op_2282_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_49_cast_fp16 = layer_norm(axes = input_49_axes_0, beta = norm4_2_bias_to_fp16, epsilon = var_2282_to_fp16, gamma = norm4_2_weight_to_fp16, x = input_47_cast_fp16)[name = tensor("input_49_cast_fp16")]; + tensor linear_19_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = linear4_1_weight_cast_fp16, x = input_49_cast_fp16)[name = tensor("linear_19_cast_fp16")]; + tensor input_51_mode_0 = const()[name = tensor("input_51_mode_0"), val = tensor("EXACT")]; + tensor input_51_cast_fp16 = gelu(mode = input_51_mode_0, x = linear_19_cast_fp16)[name = tensor("input_51_cast_fp16")]; + tensor linear_20_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = linear4_2_weight_cast_fp16, x = input_51_cast_fp16)[name = tensor("linear_20_cast_fp16")]; + tensor input_53_cast_fp16 = add(x = input_47_cast_fp16, y = linear_20_cast_fp16)[name = tensor("input_53_cast_fp16")]; + tensor x_11_axes_0 = const()[name = tensor("x_11_axes_0"), val = tensor([-1])]; + tensor norm5_1_weight_to_fp16 = const()[name = tensor("norm5_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(303627776)))]; + tensor norm5_1_bias_to_fp16 = const()[name = tensor("norm5_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(303629888)))]; + tensor var_2299_to_fp16 = const()[name = tensor("op_2299_to_fp16"), val = tensor(0x1.5p-17)]; + tensor x_11_cast_fp16 = layer_norm(axes = x_11_axes_0, beta = norm5_1_bias_to_fp16, epsilon = var_2299_to_fp16, gamma = norm5_1_weight_to_fp16, x = input_53_cast_fp16)[name = tensor("x_11_cast_fp16")]; + tensor linear_21_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = attn5_in_proj_weight_cast_fp16, x = x_11_cast_fp16)[name = tensor("linear_21_cast_fp16")]; + tensor var_2335 = const()[name = tensor("op_2335"), val = tensor([1, 1, 3, 16, 64])]; + tensor qkv_11_cast_fp16 = reshape(shape = var_2335, x = linear_21_cast_fp16)[name = tensor("qkv_11_cast_fp16")]; + tensor q_31_begin_0 = const()[name = tensor("q_31_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor q_31_end_0 = const()[name = tensor("q_31_end_0"), val = tensor([1, 1, 1, 16, 64])]; + tensor q_31_end_mask_0 = const()[name = tensor("q_31_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor q_31_squeeze_mask_0 = const()[name = tensor("q_31_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor q_31_cast_fp16 = slice_by_index(begin = q_31_begin_0, end = q_31_end_0, end_mask = q_31_end_mask_0, squeeze_mask = q_31_squeeze_mask_0, x = qkv_11_cast_fp16)[name = tensor("q_31_cast_fp16")]; + tensor k_21_begin_0 = const()[name = tensor("k_21_begin_0"), val = tensor([0, 0, 1, 0, 0])]; + tensor k_21_end_0 = const()[name = tensor("k_21_end_0"), val = tensor([1, 1, 2, 16, 64])]; + tensor k_21_end_mask_0 = const()[name = tensor("k_21_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor k_21_squeeze_mask_0 = const()[name = tensor("k_21_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor k_21_cast_fp16 = slice_by_index(begin = k_21_begin_0, end = k_21_end_0, end_mask = k_21_end_mask_0, squeeze_mask = k_21_squeeze_mask_0, x = qkv_11_cast_fp16)[name = tensor("k_21_cast_fp16")]; + tensor v_11_begin_0 = const()[name = tensor("v_11_begin_0"), val = tensor([0, 0, 2, 0, 0])]; + tensor v_11_end_0 = const()[name = tensor("v_11_end_0"), val = tensor([1, 1, 3, 16, 64])]; + tensor v_11_end_mask_0 = const()[name = tensor("v_11_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor v_11_squeeze_mask_0 = const()[name = tensor("v_11_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor v_11_cast_fp16 = slice_by_index(begin = v_11_begin_0, end = v_11_end_0, end_mask = v_11_end_mask_0, squeeze_mask = v_11_squeeze_mask_0, x = qkv_11_cast_fp16)[name = tensor("v_11_cast_fp16")]; + tensor var_2439 = const()[name = tensor("op_2439"), val = tensor([1, 1, 1, 1])]; + tensor position5_to_fp16_dtype_0 = const()[name = tensor("position5_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor position5_to_fp16 = cast(dtype = position5_to_fp16_dtype_0, x = position5)[name = tensor("cast_528")]; + tensor ts_35_cast_fp16 = reshape(shape = var_2439, x = position5_to_fp16)[name = tensor("ts_35_cast_fp16")]; + tensor var_2443 = const()[name = tensor("op_2443"), val = tensor([1, 1, 16, 32, 2])]; + tensor q_complex_11_cast_fp16 = reshape(shape = var_2443, x = q_31_cast_fp16)[name = tensor("q_complex_11_cast_fp16")]; + tensor var_2447 = const()[name = tensor("op_2447"), val = tensor([1, 1, 16, 32, 2])]; + tensor k_complex_11_cast_fp16 = reshape(shape = var_2447, x = k_21_cast_fp16)[name = tensor("k_complex_11_cast_fp16")]; + tensor var_2451_begin_0 = const()[name = tensor("op_2451_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_2451_end_0 = const()[name = tensor("op_2451_end_0"), val = tensor([1, 1, 16, 32, 1])]; + tensor var_2451_end_mask_0 = const()[name = tensor("op_2451_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_2451_squeeze_mask_0 = const()[name = tensor("op_2451_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_2451_cast_fp16 = slice_by_index(begin = var_2451_begin_0, end = var_2451_end_0, end_mask = var_2451_end_mask_0, squeeze_mask = var_2451_squeeze_mask_0, x = q_complex_11_cast_fp16)[name = tensor("op_2451_cast_fp16")]; + tensor var_2459_begin_0 = const()[name = tensor("op_2459_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_2459_end_0 = const()[name = tensor("op_2459_end_0"), val = tensor([1, 1, 16, 32, 2])]; + tensor var_2459_end_mask_0 = const()[name = tensor("op_2459_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_2459_squeeze_mask_0 = const()[name = tensor("op_2459_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_2459_cast_fp16 = slice_by_index(begin = var_2459_begin_0, end = var_2459_end_0, end_mask = var_2459_end_mask_0, squeeze_mask = var_2459_squeeze_mask_0, x = q_complex_11_cast_fp16)[name = tensor("op_2459_cast_fp16")]; + tensor var_2467_begin_0 = const()[name = tensor("op_2467_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_2467_end_0 = const()[name = tensor("op_2467_end_0"), val = tensor([1, 1, 16, 32, 1])]; + tensor var_2467_end_mask_0 = const()[name = tensor("op_2467_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_2467_squeeze_mask_0 = const()[name = tensor("op_2467_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_2467_cast_fp16 = slice_by_index(begin = var_2467_begin_0, end = var_2467_end_0, end_mask = var_2467_end_mask_0, squeeze_mask = var_2467_squeeze_mask_0, x = k_complex_11_cast_fp16)[name = tensor("op_2467_cast_fp16")]; + tensor var_2475_begin_0 = const()[name = tensor("op_2475_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_2475_end_0 = const()[name = tensor("op_2475_end_0"), val = tensor([1, 1, 16, 32, 2])]; + tensor var_2475_end_mask_0 = const()[name = tensor("op_2475_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_2475_squeeze_mask_0 = const()[name = tensor("op_2475_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_2475_cast_fp16 = slice_by_index(begin = var_2475_begin_0, end = var_2475_end_0, end_mask = var_2475_end_mask_0, squeeze_mask = var_2475_squeeze_mask_0, x = k_complex_11_cast_fp16)[name = tensor("op_2475_cast_fp16")]; + tensor freqs_11_to_fp16 = const()[name = tensor("freqs_11_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(303632000)))]; + tensor var_2481_cast_fp16 = mul(x = freqs_11_to_fp16, y = ts_35_cast_fp16)[name = tensor("op_2481_cast_fp16")]; + tensor rotr_11_cast_fp16 = cos(x = var_2481_cast_fp16)[name = tensor("rotr_11_cast_fp16")]; + tensor roti_11_cast_fp16 = sin(x = var_2481_cast_fp16)[name = tensor("roti_11_cast_fp16")]; + tensor var_2485_cast_fp16 = mul(x = var_2451_cast_fp16, y = rotr_11_cast_fp16)[name = tensor("op_2485_cast_fp16")]; + tensor var_2486_cast_fp16 = mul(x = var_2459_cast_fp16, y = roti_11_cast_fp16)[name = tensor("op_2486_cast_fp16")]; + tensor qor_21_cast_fp16 = sub(x = var_2485_cast_fp16, y = var_2486_cast_fp16)[name = tensor("qor_21_cast_fp16")]; + tensor var_2489_cast_fp16 = mul(x = var_2451_cast_fp16, y = roti_11_cast_fp16)[name = tensor("op_2489_cast_fp16")]; + tensor var_2490_cast_fp16 = mul(x = var_2459_cast_fp16, y = rotr_11_cast_fp16)[name = tensor("op_2490_cast_fp16")]; + tensor qoi_21_cast_fp16 = add(x = var_2489_cast_fp16, y = var_2490_cast_fp16)[name = tensor("qoi_21_cast_fp16")]; + tensor var_2493_cast_fp16 = mul(x = var_2467_cast_fp16, y = rotr_11_cast_fp16)[name = tensor("op_2493_cast_fp16")]; + tensor var_2494_cast_fp16 = mul(x = var_2475_cast_fp16, y = roti_11_cast_fp16)[name = tensor("op_2494_cast_fp16")]; + tensor kor_21_cast_fp16 = sub(x = var_2493_cast_fp16, y = var_2494_cast_fp16)[name = tensor("kor_21_cast_fp16")]; + tensor var_2497_cast_fp16 = mul(x = var_2467_cast_fp16, y = roti_11_cast_fp16)[name = tensor("op_2497_cast_fp16")]; + tensor var_2498_cast_fp16 = mul(x = var_2475_cast_fp16, y = rotr_11_cast_fp16)[name = tensor("op_2498_cast_fp16")]; + tensor koi_21_cast_fp16 = add(x = var_2497_cast_fp16, y = var_2498_cast_fp16)[name = tensor("koi_21_cast_fp16")]; + tensor qo_11_axis_0 = const()[name = tensor("qo_11_axis_0"), val = tensor(-1)]; + tensor qo_11_cast_fp16 = stack(axis = qo_11_axis_0, values = (qor_21_cast_fp16, qoi_21_cast_fp16))[name = tensor("qo_11_cast_fp16")]; + tensor ko_11_axis_0 = const()[name = tensor("ko_11_axis_0"), val = tensor(-1)]; + tensor ko_11_cast_fp16 = stack(axis = ko_11_axis_0, values = (kor_21_cast_fp16, koi_21_cast_fp16))[name = tensor("ko_11_cast_fp16")]; + tensor var_2527 = const()[name = tensor("op_2527"), val = tensor([1, 1, 16, 64])]; + tensor q_33_cast_fp16 = reshape(shape = var_2527, x = qo_11_cast_fp16)[name = tensor("q_33_cast_fp16")]; + tensor var_2529 = const()[name = tensor("op_2529"), val = tensor([1, 1, 16, 64])]; + tensor k_23_cast_fp16 = reshape(shape = var_2529, x = ko_11_cast_fp16)[name = tensor("k_23_cast_fp16")]; + tensor _inversed_2551_y_0_to_fp16 = const()[name = tensor("_inversed_2551_y_0_to_fp16"), val = tensor(0x1p-9)]; + tensor _inversed_2551_cast_fp16 = mul(x = ts_35_cast_fp16, y = _inversed_2551_y_0_to_fp16)[name = tensor("_inversed_2551_cast_fp16")]; + tensor var_2552_cast_fp16 = floor(x = _inversed_2551_cast_fp16)[name = tensor("op_2552_cast_fp16")]; + tensor var_2553_to_fp16 = const()[name = tensor("op_2553_to_fp16"), val = tensor(0x1p+9)]; + tensor var_2554_cast_fp16 = mul(x = var_2552_cast_fp16, y = var_2553_to_fp16)[name = tensor("op_2554_cast_fp16")]; + tensor write_indices_float_23_cast_fp16 = sub(x = ts_35_cast_fp16, y = var_2554_cast_fp16)[name = tensor("write_indices_float_23_cast_fp16")]; + tensor var_2561_dtype_0 = const()[name = tensor("op_2561_dtype_0"), val = tensor("int32")]; + tensor write_indices_11_reps_0 = const()[name = tensor("write_indices_11_reps_0"), val = tensor([1, 1, 16, 64])]; + tensor write_indices_float_23_cast_fp16_to_int32 = cast(dtype = var_2561_dtype_0, x = write_indices_float_23_cast_fp16)[name = tensor("cast_527")]; + tensor write_indices_11 = tile(reps = write_indices_11_reps_0, x = write_indices_float_23_cast_fp16_to_int32)[name = tensor("write_indices_11")]; + tensor var_2569_begin_0 = const()[name = tensor("op_2569_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_2569_end_0 = const()[name = tensor("op_2569_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor var_2569_end_mask_0 = const()[name = tensor("op_2569_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_2569_squeeze_mask_0 = const()[name = tensor("op_2569_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor cache5_to_fp16_dtype_0 = const()[name = tensor("cache5_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor cache5_to_fp16 = cast(dtype = cache5_to_fp16_dtype_0, x = cache5)[name = tensor("cast_526")]; + tensor var_2569_cast_fp16 = slice_by_index(begin = var_2569_begin_0, end = var_2569_end_0, end_mask = var_2569_end_mask_0, squeeze_mask = var_2569_squeeze_mask_0, x = cache5_to_fp16)[name = tensor("op_2569_cast_fp16")]; + tensor var_2571_axis_0 = const()[name = tensor("op_2571_axis_0"), val = tensor(1)]; + tensor var_2571_mode_0 = const()[name = tensor("op_2571_mode_0"), val = tensor("update")]; + tensor var_2571_validate_indices_0 = const()[name = tensor("op_2571_validate_indices_0"), val = tensor(false)]; + tensor var_2571_cast_fp16 = scatter_along_axis(axis = var_2571_axis_0, data = var_2569_cast_fp16, indices = write_indices_11, mode = var_2571_mode_0, updates = k_23_cast_fp16, validate_indices = var_2571_validate_indices_0)[name = tensor("op_2571_cast_fp16")]; + tensor concat_37 = const()[name = tensor("concat_37"), val = tensor([0, 0, 0, 0, 0])]; + tensor concat_38 = const()[name = tensor("concat_38"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_11_internal_tensor_assign_1_stride_0 = const()[name = tensor("new_cache_11_internal_tensor_assign_1_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_11_internal_tensor_assign_1_begin_mask_0 = const()[name = tensor("new_cache_11_internal_tensor_assign_1_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_11_internal_tensor_assign_1_end_mask_0 = const()[name = tensor("new_cache_11_internal_tensor_assign_1_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_11_internal_tensor_assign_1_squeeze_mask_0 = const()[name = tensor("new_cache_11_internal_tensor_assign_1_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_58 = const()[name = tensor("shape_58"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_10 = const()[name = tensor("reduce_prod_10"), val = tensor(1048576)]; + tensor range_1d_10_start_0 = const()[name = tensor("range_1d_10_start_0"), val = tensor(0)]; + tensor range_1d_10_step_0 = const()[name = tensor("range_1d_10_step_0"), val = tensor(1)]; + tensor range_1d_10 = range_1d(end = reduce_prod_10, start = range_1d_10_start_0, step = range_1d_10_step_0)[name = tensor("range_1d_10")]; + tensor reshape_50 = reshape(shape = shape_58, x = range_1d_10)[name = tensor("reshape_50")]; + tensor slice_by_index_10 = slice_by_index(begin = concat_37, begin_mask = new_cache_11_internal_tensor_assign_1_begin_mask_0, end = concat_38, end_mask = new_cache_11_internal_tensor_assign_1_end_mask_0, squeeze_mask = new_cache_11_internal_tensor_assign_1_squeeze_mask_0, stride = new_cache_11_internal_tensor_assign_1_stride_0, x = reshape_50)[name = tensor("slice_by_index_10")]; + tensor reshape_51_shape_0 = const()[name = tensor("reshape_51_shape_0"), val = tensor([-1])]; + tensor reshape_51 = reshape(shape = reshape_51_shape_0, x = slice_by_index_10)[name = tensor("reshape_51")]; + tensor reshape_52_shape_0 = const()[name = tensor("reshape_52_shape_0"), val = tensor([-1])]; + tensor reshape_52_cast_fp16 = reshape(shape = reshape_52_shape_0, x = var_2571_cast_fp16)[name = tensor("reshape_52_cast_fp16")]; + tensor reshape_53_shape_0 = const()[name = tensor("reshape_53_shape_0"), val = tensor([-1])]; + tensor reshape_53_cast_fp16 = reshape(shape = reshape_53_shape_0, x = cache5_to_fp16)[name = tensor("reshape_53_cast_fp16")]; + tensor scatter_10_mode_0 = const()[name = tensor("scatter_10_mode_0"), val = tensor("update")]; + tensor scatter_10_axis_0 = const()[name = tensor("scatter_10_axis_0"), val = tensor(0)]; + tensor scatter_10_validate_indices_0 = const()[name = tensor("scatter_10_validate_indices_0"), val = tensor(false)]; + tensor scatter_10_cast_fp16 = scatter(axis = scatter_10_axis_0, data = reshape_53_cast_fp16, indices = reshape_51, mode = scatter_10_mode_0, updates = reshape_52_cast_fp16, validate_indices = scatter_10_validate_indices_0)[name = tensor("scatter_10_cast_fp16")]; + tensor reshape_54_cast_fp16 = reshape(shape = shape_58, x = scatter_10_cast_fp16)[name = tensor("reshape_54_cast_fp16")]; + tensor var_2579_begin_0 = const()[name = tensor("op_2579_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor var_2579_end_0 = const()[name = tensor("op_2579_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor var_2579_end_mask_0 = const()[name = tensor("op_2579_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_2579_squeeze_mask_0 = const()[name = tensor("op_2579_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor var_2579_cast_fp16 = slice_by_index(begin = var_2579_begin_0, end = var_2579_end_0, end_mask = var_2579_end_mask_0, squeeze_mask = var_2579_squeeze_mask_0, x = reshape_54_cast_fp16)[name = tensor("op_2579_cast_fp16")]; + tensor var_2581_axis_0 = const()[name = tensor("op_2581_axis_0"), val = tensor(1)]; + tensor var_2581_mode_0 = const()[name = tensor("op_2581_mode_0"), val = tensor("update")]; + tensor var_2581_validate_indices_0 = const()[name = tensor("op_2581_validate_indices_0"), val = tensor(false)]; + tensor var_2581_cast_fp16 = scatter_along_axis(axis = var_2581_axis_0, data = var_2579_cast_fp16, indices = write_indices_11, mode = var_2581_mode_0, updates = v_11_cast_fp16, validate_indices = var_2581_validate_indices_0)[name = tensor("op_2581_cast_fp16")]; + tensor concat_39 = const()[name = tensor("concat_39"), val = tensor([1, 0, 0, 0, 0])]; + tensor concat_40 = const()[name = tensor("concat_40"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_11_internal_tensor_assign_2_stride_0 = const()[name = tensor("new_cache_11_internal_tensor_assign_2_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_11_internal_tensor_assign_2_begin_mask_0 = const()[name = tensor("new_cache_11_internal_tensor_assign_2_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_11_internal_tensor_assign_2_end_mask_0 = const()[name = tensor("new_cache_11_internal_tensor_assign_2_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_11_internal_tensor_assign_2_squeeze_mask_0 = const()[name = tensor("new_cache_11_internal_tensor_assign_2_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_59 = const()[name = tensor("shape_59"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_11 = const()[name = tensor("reduce_prod_11"), val = tensor(1048576)]; + tensor range_1d_11_start_0 = const()[name = tensor("range_1d_11_start_0"), val = tensor(0)]; + tensor range_1d_11_step_0 = const()[name = tensor("range_1d_11_step_0"), val = tensor(1)]; + tensor range_1d_11 = range_1d(end = reduce_prod_11, start = range_1d_11_start_0, step = range_1d_11_step_0)[name = tensor("range_1d_11")]; + tensor reshape_55 = reshape(shape = shape_59, x = range_1d_11)[name = tensor("reshape_55")]; + tensor slice_by_index_11 = slice_by_index(begin = concat_39, begin_mask = new_cache_11_internal_tensor_assign_2_begin_mask_0, end = concat_40, end_mask = new_cache_11_internal_tensor_assign_2_end_mask_0, squeeze_mask = new_cache_11_internal_tensor_assign_2_squeeze_mask_0, stride = new_cache_11_internal_tensor_assign_2_stride_0, x = reshape_55)[name = tensor("slice_by_index_11")]; + tensor reshape_56_shape_0 = const()[name = tensor("reshape_56_shape_0"), val = tensor([-1])]; + tensor reshape_56 = reshape(shape = reshape_56_shape_0, x = slice_by_index_11)[name = tensor("reshape_56")]; + tensor reshape_57_shape_0 = const()[name = tensor("reshape_57_shape_0"), val = tensor([-1])]; + tensor reshape_57_cast_fp16 = reshape(shape = reshape_57_shape_0, x = var_2581_cast_fp16)[name = tensor("reshape_57_cast_fp16")]; + tensor reshape_58_shape_0 = const()[name = tensor("reshape_58_shape_0"), val = tensor([-1])]; + tensor reshape_58_cast_fp16 = reshape(shape = reshape_58_shape_0, x = reshape_54_cast_fp16)[name = tensor("reshape_58_cast_fp16")]; + tensor scatter_11_mode_0 = const()[name = tensor("scatter_11_mode_0"), val = tensor("update")]; + tensor scatter_11_axis_0 = const()[name = tensor("scatter_11_axis_0"), val = tensor(0)]; + tensor scatter_11_validate_indices_0 = const()[name = tensor("scatter_11_validate_indices_0"), val = tensor(false)]; + tensor scatter_11_cast_fp16 = scatter(axis = scatter_11_axis_0, data = reshape_58_cast_fp16, indices = reshape_56, mode = scatter_11_mode_0, updates = reshape_57_cast_fp16, validate_indices = scatter_11_validate_indices_0)[name = tensor("scatter_11_cast_fp16")]; + tensor reshape_59_cast_fp16 = reshape(shape = shape_59, x = scatter_11_cast_fp16)[name = tensor("reshape_59_cast_fp16")]; + tensor reshape_59_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("reshape_59_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor keys_31_begin_0 = const()[name = tensor("keys_31_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor keys_31_end_0 = const()[name = tensor("keys_31_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor keys_31_end_mask_0 = const()[name = tensor("keys_31_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor keys_31_squeeze_mask_0 = const()[name = tensor("keys_31_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor keys_31_cast_fp16 = slice_by_index(begin = keys_31_begin_0, end = keys_31_end_0, end_mask = keys_31_end_mask_0, squeeze_mask = keys_31_squeeze_mask_0, x = reshape_59_cast_fp16)[name = tensor("keys_31_cast_fp16")]; + tensor values_31_begin_0 = const()[name = tensor("values_31_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor values_31_end_0 = const()[name = tensor("values_31_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor values_31_end_mask_0 = const()[name = tensor("values_31_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor values_31_squeeze_mask_0 = const()[name = tensor("values_31_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor values_31_cast_fp16 = slice_by_index(begin = values_31_begin_0, end = values_31_end_0, end_mask = values_31_end_mask_0, squeeze_mask = values_31_squeeze_mask_0, x = reshape_59_cast_fp16)[name = tensor("values_31_cast_fp16")]; + tensor var_2593_cast_fp16 = not_equal(x = keys_31_cast_fp16, y = keys_31_cast_fp16)[name = tensor("op_2593_cast_fp16")]; + tensor keys_33_cast_fp16 = select(a = var_504_to_fp16, b = keys_31_cast_fp16, cond = var_2593_cast_fp16)[name = tensor("keys_33_cast_fp16")]; + tensor var_2601_cast_fp16 = not_equal(x = values_31_cast_fp16, y = values_31_cast_fp16)[name = tensor("op_2601_cast_fp16")]; + tensor values_33_cast_fp16 = select(a = var_504_to_fp16, b = values_31_cast_fp16, cond = var_2601_cast_fp16)[name = tensor("values_33_cast_fp16")]; + tensor var_2625 = const()[name = tensor("op_2625"), val = tensor([0, 2, 1, 3])]; + tensor var_2638 = const()[name = tensor("op_2638"), val = tensor([1, 1, 1])]; + tensor var_2639_cast_fp16 = reshape(shape = var_2638, x = position5_to_fp16)[name = tensor("op_2639_cast_fp16")]; + tensor var_2656_to_fp16 = const()[name = tensor("op_2656_to_fp16"), val = tensor(0x1p+0)]; + tensor valid_len_11_cast_fp16 = add(x = var_2639_cast_fp16, y = var_2656_to_fp16)[name = tensor("valid_len_11_cast_fp16")]; + tensor valid_mask_11_cast_fp16 = less(x = k_positions_1_promoted_to_fp16, y = valid_len_11_cast_fp16)[name = tensor("valid_mask_11_cast_fp16")]; + tensor causal_mask_11_cast_fp16 = less_equal(x = k_positions_1_promoted_to_fp16, y = var_2639_cast_fp16)[name = tensor("causal_mask_11_cast_fp16")]; + tensor attn_mask_21 = logical_and(x = valid_mask_11_cast_fp16, y = causal_mask_11_cast_fp16)[name = tensor("attn_mask_21")]; + tensor attn_mask_23_axes_0 = const()[name = tensor("attn_mask_23_axes_0"), val = tensor([1])]; + tensor attn_mask_23 = expand_dims(axes = attn_mask_23_axes_0, x = attn_mask_21)[name = tensor("attn_mask_23")]; + tensor var_2674_transpose_x_0 = const()[name = tensor("op_2674_transpose_x_0"), val = tensor(false)]; + tensor var_2674_transpose_y_0 = const()[name = tensor("op_2674_transpose_y_0"), val = tensor(false)]; + tensor transpose_82_perm_0 = const()[name = tensor("transpose_82_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_83_perm_0 = const()[name = tensor("transpose_83_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_83 = transpose(perm = transpose_83_perm_0, x = keys_33_cast_fp16)[name = tensor("transpose_193")]; + tensor transpose_82 = transpose(perm = transpose_82_perm_0, x = q_33_cast_fp16)[name = tensor("transpose_194")]; + tensor var_2674_cast_fp16 = matmul(transpose_x = var_2674_transpose_x_0, transpose_y = var_2674_transpose_y_0, x = transpose_82, y = transpose_83)[name = tensor("op_2674_cast_fp16")]; + tensor var_2668_to_fp16 = const()[name = tensor("op_2668_to_fp16"), val = tensor([0x1p-3])]; + tensor attn_weights_31_cast_fp16 = mul(x = var_2674_cast_fp16, y = var_2668_to_fp16)[name = tensor("attn_weights_31_cast_fp16")]; + tensor var_2676 = logical_not(x = attn_mask_23)[name = tensor("op_2676")]; + tensor var_2677_to_fp16 = const()[name = tensor("op_2677_to_fp16"), val = tensor(-inf)]; + tensor attn_weights_33_cast_fp16 = select(a = var_2677_to_fp16, b = attn_weights_31_cast_fp16, cond = var_2676)[name = tensor("attn_weights_33_cast_fp16")]; + tensor var_2679 = const()[name = tensor("op_2679"), val = tensor(-1)]; + tensor attn_weights_35_cast_fp16 = softmax(axis = var_2679, x = attn_weights_33_cast_fp16)[name = tensor("attn_weights_35_cast_fp16")]; + tensor attn_output_11_transpose_x_0 = const()[name = tensor("attn_output_11_transpose_x_0"), val = tensor(false)]; + tensor attn_output_11_transpose_y_0 = const()[name = tensor("attn_output_11_transpose_y_0"), val = tensor(false)]; + tensor values_35_cast_fp16 = transpose(perm = var_2625, x = values_33_cast_fp16)[name = tensor("transpose_195")]; + tensor attn_output_11_cast_fp16 = matmul(transpose_x = attn_output_11_transpose_x_0, transpose_y = attn_output_11_transpose_y_0, x = attn_weights_35_cast_fp16, y = values_35_cast_fp16)[name = tensor("attn_output_11_cast_fp16")]; + tensor var_2687 = const()[name = tensor("op_2687"), val = tensor([0, 2, 1, 3])]; + tensor var_2690 = const()[name = tensor("op_2690"), val = tensor([1, 1, 1024])]; + tensor var_2688_cast_fp16 = transpose(perm = var_2687, x = attn_output_11_cast_fp16)[name = tensor("transpose_192")]; + tensor input_55_cast_fp16 = reshape(shape = var_2690, x = var_2688_cast_fp16)[name = tensor("input_55_cast_fp16")]; + tensor linear_22_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = attn5_out_proj_weight_cast_fp16, x = input_55_cast_fp16)[name = tensor("linear_22_cast_fp16")]; + tensor var_2696_to_fp16 = const()[name = tensor("op_2696_to_fp16"), val = tensor(0x1p+0)]; + tensor var_2697_cast_fp16 = add(x = position5_to_fp16, y = var_2696_to_fp16)[name = tensor("op_2697_cast_fp16")]; + tensor var_2697_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_2697_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor input_57_cast_fp16 = add(x = input_53_cast_fp16, y = linear_22_cast_fp16)[name = tensor("input_57_cast_fp16")]; + tensor input_59_axes_0 = const()[name = tensor("input_59_axes_0"), val = tensor([-1])]; + tensor norm5_2_weight_to_fp16 = const()[name = tensor("norm5_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(303632128)))]; + tensor norm5_2_bias_to_fp16 = const()[name = tensor("norm5_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(303634240)))]; + tensor var_2701_to_fp16 = const()[name = tensor("op_2701_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_59_cast_fp16 = layer_norm(axes = input_59_axes_0, beta = norm5_2_bias_to_fp16, epsilon = var_2701_to_fp16, gamma = norm5_2_weight_to_fp16, x = input_57_cast_fp16)[name = tensor("input_59_cast_fp16")]; + tensor linear_23_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = linear5_1_weight_cast_fp16, x = input_59_cast_fp16)[name = tensor("linear_23_cast_fp16")]; + tensor input_61_mode_0 = const()[name = tensor("input_61_mode_0"), val = tensor("EXACT")]; + tensor input_61_cast_fp16 = gelu(mode = input_61_mode_0, x = linear_23_cast_fp16)[name = tensor("input_61_cast_fp16")]; + tensor linear_24_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = linear5_2_weight_cast_fp16, x = input_61_cast_fp16)[name = tensor("linear_24_cast_fp16")]; + tensor input_63_cast_fp16 = add(x = input_57_cast_fp16, y = linear_24_cast_fp16)[name = tensor("input_63_cast_fp16")]; + tensor x_13_axes_0 = const()[name = tensor("x_13_axes_0"), val = tensor([-1])]; + tensor norm6_1_weight_to_fp16 = const()[name = tensor("norm6_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(303636352)))]; + tensor norm6_1_bias_to_fp16 = const()[name = tensor("norm6_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(303638464)))]; + tensor var_2718_to_fp16 = const()[name = tensor("op_2718_to_fp16"), val = tensor(0x1.5p-17)]; + tensor x_13_cast_fp16 = layer_norm(axes = x_13_axes_0, beta = norm6_1_bias_to_fp16, epsilon = var_2718_to_fp16, gamma = norm6_1_weight_to_fp16, x = input_63_cast_fp16)[name = tensor("x_13_cast_fp16")]; + tensor linear_25_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = attn6_in_proj_weight_cast_fp16, x = x_13_cast_fp16)[name = tensor("linear_25_cast_fp16")]; + tensor var_2754 = const()[name = tensor("op_2754"), val = tensor([1, 1, 3, 16, 64])]; + tensor qkv_13_cast_fp16 = reshape(shape = var_2754, x = linear_25_cast_fp16)[name = tensor("qkv_13_cast_fp16")]; + tensor q_37_begin_0 = const()[name = tensor("q_37_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor q_37_end_0 = const()[name = tensor("q_37_end_0"), val = tensor([1, 1, 1, 16, 64])]; + tensor q_37_end_mask_0 = const()[name = tensor("q_37_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor q_37_squeeze_mask_0 = const()[name = tensor("q_37_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor q_37_cast_fp16 = slice_by_index(begin = q_37_begin_0, end = q_37_end_0, end_mask = q_37_end_mask_0, squeeze_mask = q_37_squeeze_mask_0, x = qkv_13_cast_fp16)[name = tensor("q_37_cast_fp16")]; + tensor k_25_begin_0 = const()[name = tensor("k_25_begin_0"), val = tensor([0, 0, 1, 0, 0])]; + tensor k_25_end_0 = const()[name = tensor("k_25_end_0"), val = tensor([1, 1, 2, 16, 64])]; + tensor k_25_end_mask_0 = const()[name = tensor("k_25_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor k_25_squeeze_mask_0 = const()[name = tensor("k_25_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor k_25_cast_fp16 = slice_by_index(begin = k_25_begin_0, end = k_25_end_0, end_mask = k_25_end_mask_0, squeeze_mask = k_25_squeeze_mask_0, x = qkv_13_cast_fp16)[name = tensor("k_25_cast_fp16")]; + tensor v_13_begin_0 = const()[name = tensor("v_13_begin_0"), val = tensor([0, 0, 2, 0, 0])]; + tensor v_13_end_0 = const()[name = tensor("v_13_end_0"), val = tensor([1, 1, 3, 16, 64])]; + tensor v_13_end_mask_0 = const()[name = tensor("v_13_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor v_13_squeeze_mask_0 = const()[name = tensor("v_13_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor v_13_cast_fp16 = slice_by_index(begin = v_13_begin_0, end = v_13_end_0, end_mask = v_13_end_mask_0, squeeze_mask = v_13_squeeze_mask_0, x = qkv_13_cast_fp16)[name = tensor("v_13_cast_fp16")]; + tensor var_2858 = const()[name = tensor("op_2858"), val = tensor([1, 1, 1, 1])]; + tensor position6_to_fp16_dtype_0 = const()[name = tensor("position6_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor position6_to_fp16 = cast(dtype = position6_to_fp16_dtype_0, x = position6)[name = tensor("cast_523")]; + tensor ts_41_cast_fp16 = reshape(shape = var_2858, x = position6_to_fp16)[name = tensor("ts_41_cast_fp16")]; + tensor var_2862 = const()[name = tensor("op_2862"), val = tensor([1, 1, 16, 32, 2])]; + tensor q_complex_13_cast_fp16 = reshape(shape = var_2862, x = q_37_cast_fp16)[name = tensor("q_complex_13_cast_fp16")]; + tensor var_2866 = const()[name = tensor("op_2866"), val = tensor([1, 1, 16, 32, 2])]; + tensor k_complex_13_cast_fp16 = reshape(shape = var_2866, x = k_25_cast_fp16)[name = tensor("k_complex_13_cast_fp16")]; + tensor var_2870_begin_0 = const()[name = tensor("op_2870_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_2870_end_0 = const()[name = tensor("op_2870_end_0"), val = tensor([1, 1, 16, 32, 1])]; + tensor var_2870_end_mask_0 = const()[name = tensor("op_2870_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_2870_squeeze_mask_0 = const()[name = tensor("op_2870_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_2870_cast_fp16 = slice_by_index(begin = var_2870_begin_0, end = var_2870_end_0, end_mask = var_2870_end_mask_0, squeeze_mask = var_2870_squeeze_mask_0, x = q_complex_13_cast_fp16)[name = tensor("op_2870_cast_fp16")]; + tensor var_2878_begin_0 = const()[name = tensor("op_2878_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_2878_end_0 = const()[name = tensor("op_2878_end_0"), val = tensor([1, 1, 16, 32, 2])]; + tensor var_2878_end_mask_0 = const()[name = tensor("op_2878_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_2878_squeeze_mask_0 = const()[name = tensor("op_2878_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_2878_cast_fp16 = slice_by_index(begin = var_2878_begin_0, end = var_2878_end_0, end_mask = var_2878_end_mask_0, squeeze_mask = var_2878_squeeze_mask_0, x = q_complex_13_cast_fp16)[name = tensor("op_2878_cast_fp16")]; + tensor var_2886_begin_0 = const()[name = tensor("op_2886_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_2886_end_0 = const()[name = tensor("op_2886_end_0"), val = tensor([1, 1, 16, 32, 1])]; + tensor var_2886_end_mask_0 = const()[name = tensor("op_2886_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_2886_squeeze_mask_0 = const()[name = tensor("op_2886_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_2886_cast_fp16 = slice_by_index(begin = var_2886_begin_0, end = var_2886_end_0, end_mask = var_2886_end_mask_0, squeeze_mask = var_2886_squeeze_mask_0, x = k_complex_13_cast_fp16)[name = tensor("op_2886_cast_fp16")]; + tensor var_2894_begin_0 = const()[name = tensor("op_2894_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_2894_end_0 = const()[name = tensor("op_2894_end_0"), val = tensor([1, 1, 16, 32, 2])]; + tensor var_2894_end_mask_0 = const()[name = tensor("op_2894_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_2894_squeeze_mask_0 = const()[name = tensor("op_2894_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_2894_cast_fp16 = slice_by_index(begin = var_2894_begin_0, end = var_2894_end_0, end_mask = var_2894_end_mask_0, squeeze_mask = var_2894_squeeze_mask_0, x = k_complex_13_cast_fp16)[name = tensor("op_2894_cast_fp16")]; + tensor freqs_13_to_fp16 = const()[name = tensor("freqs_13_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(303640576)))]; + tensor var_2900_cast_fp16 = mul(x = freqs_13_to_fp16, y = ts_41_cast_fp16)[name = tensor("op_2900_cast_fp16")]; + tensor rotr_13_cast_fp16 = cos(x = var_2900_cast_fp16)[name = tensor("rotr_13_cast_fp16")]; + tensor roti_13_cast_fp16 = sin(x = var_2900_cast_fp16)[name = tensor("roti_13_cast_fp16")]; + tensor var_2904_cast_fp16 = mul(x = var_2870_cast_fp16, y = rotr_13_cast_fp16)[name = tensor("op_2904_cast_fp16")]; + tensor var_2905_cast_fp16 = mul(x = var_2878_cast_fp16, y = roti_13_cast_fp16)[name = tensor("op_2905_cast_fp16")]; + tensor qor_25_cast_fp16 = sub(x = var_2904_cast_fp16, y = var_2905_cast_fp16)[name = tensor("qor_25_cast_fp16")]; + tensor var_2908_cast_fp16 = mul(x = var_2870_cast_fp16, y = roti_13_cast_fp16)[name = tensor("op_2908_cast_fp16")]; + tensor var_2909_cast_fp16 = mul(x = var_2878_cast_fp16, y = rotr_13_cast_fp16)[name = tensor("op_2909_cast_fp16")]; + tensor qoi_25_cast_fp16 = add(x = var_2908_cast_fp16, y = var_2909_cast_fp16)[name = tensor("qoi_25_cast_fp16")]; + tensor var_2912_cast_fp16 = mul(x = var_2886_cast_fp16, y = rotr_13_cast_fp16)[name = tensor("op_2912_cast_fp16")]; + tensor var_2913_cast_fp16 = mul(x = var_2894_cast_fp16, y = roti_13_cast_fp16)[name = tensor("op_2913_cast_fp16")]; + tensor kor_25_cast_fp16 = sub(x = var_2912_cast_fp16, y = var_2913_cast_fp16)[name = tensor("kor_25_cast_fp16")]; + tensor var_2916_cast_fp16 = mul(x = var_2886_cast_fp16, y = roti_13_cast_fp16)[name = tensor("op_2916_cast_fp16")]; + tensor var_2917_cast_fp16 = mul(x = var_2894_cast_fp16, y = rotr_13_cast_fp16)[name = tensor("op_2917_cast_fp16")]; + tensor koi_25_cast_fp16 = add(x = var_2916_cast_fp16, y = var_2917_cast_fp16)[name = tensor("koi_25_cast_fp16")]; + tensor qo_13_axis_0 = const()[name = tensor("qo_13_axis_0"), val = tensor(-1)]; + tensor qo_13_cast_fp16 = stack(axis = qo_13_axis_0, values = (qor_25_cast_fp16, qoi_25_cast_fp16))[name = tensor("qo_13_cast_fp16")]; + tensor ko_13_axis_0 = const()[name = tensor("ko_13_axis_0"), val = tensor(-1)]; + tensor ko_13_cast_fp16 = stack(axis = ko_13_axis_0, values = (kor_25_cast_fp16, koi_25_cast_fp16))[name = tensor("ko_13_cast_fp16")]; + tensor var_2946 = const()[name = tensor("op_2946"), val = tensor([1, 1, 16, 64])]; + tensor q_39_cast_fp16 = reshape(shape = var_2946, x = qo_13_cast_fp16)[name = tensor("q_39_cast_fp16")]; + tensor var_2948 = const()[name = tensor("op_2948"), val = tensor([1, 1, 16, 64])]; + tensor k_27_cast_fp16 = reshape(shape = var_2948, x = ko_13_cast_fp16)[name = tensor("k_27_cast_fp16")]; + tensor _inversed_2970_y_0_to_fp16 = const()[name = tensor("_inversed_2970_y_0_to_fp16"), val = tensor(0x1p-9)]; + tensor _inversed_2970_cast_fp16 = mul(x = ts_41_cast_fp16, y = _inversed_2970_y_0_to_fp16)[name = tensor("_inversed_2970_cast_fp16")]; + tensor var_2971_cast_fp16 = floor(x = _inversed_2970_cast_fp16)[name = tensor("op_2971_cast_fp16")]; + tensor var_2972_to_fp16 = const()[name = tensor("op_2972_to_fp16"), val = tensor(0x1p+9)]; + tensor var_2973_cast_fp16 = mul(x = var_2971_cast_fp16, y = var_2972_to_fp16)[name = tensor("op_2973_cast_fp16")]; + tensor write_indices_float_27_cast_fp16 = sub(x = ts_41_cast_fp16, y = var_2973_cast_fp16)[name = tensor("write_indices_float_27_cast_fp16")]; + tensor var_2980_dtype_0 = const()[name = tensor("op_2980_dtype_0"), val = tensor("int32")]; + tensor write_indices_13_reps_0 = const()[name = tensor("write_indices_13_reps_0"), val = tensor([1, 1, 16, 64])]; + tensor write_indices_float_27_cast_fp16_to_int32 = cast(dtype = var_2980_dtype_0, x = write_indices_float_27_cast_fp16)[name = tensor("cast_522")]; + tensor write_indices_13 = tile(reps = write_indices_13_reps_0, x = write_indices_float_27_cast_fp16_to_int32)[name = tensor("write_indices_13")]; + tensor var_2988_begin_0 = const()[name = tensor("op_2988_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_2988_end_0 = const()[name = tensor("op_2988_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor var_2988_end_mask_0 = const()[name = tensor("op_2988_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_2988_squeeze_mask_0 = const()[name = tensor("op_2988_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor cache6_to_fp16_dtype_0 = const()[name = tensor("cache6_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor cache6_to_fp16 = cast(dtype = cache6_to_fp16_dtype_0, x = cache6)[name = tensor("cast_521")]; + tensor var_2988_cast_fp16 = slice_by_index(begin = var_2988_begin_0, end = var_2988_end_0, end_mask = var_2988_end_mask_0, squeeze_mask = var_2988_squeeze_mask_0, x = cache6_to_fp16)[name = tensor("op_2988_cast_fp16")]; + tensor var_2990_axis_0 = const()[name = tensor("op_2990_axis_0"), val = tensor(1)]; + tensor var_2990_mode_0 = const()[name = tensor("op_2990_mode_0"), val = tensor("update")]; + tensor var_2990_validate_indices_0 = const()[name = tensor("op_2990_validate_indices_0"), val = tensor(false)]; + tensor var_2990_cast_fp16 = scatter_along_axis(axis = var_2990_axis_0, data = var_2988_cast_fp16, indices = write_indices_13, mode = var_2990_mode_0, updates = k_27_cast_fp16, validate_indices = var_2990_validate_indices_0)[name = tensor("op_2990_cast_fp16")]; + tensor concat_44 = const()[name = tensor("concat_44"), val = tensor([0, 0, 0, 0, 0])]; + tensor concat_45 = const()[name = tensor("concat_45"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_13_internal_tensor_assign_1_stride_0 = const()[name = tensor("new_cache_13_internal_tensor_assign_1_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_13_internal_tensor_assign_1_begin_mask_0 = const()[name = tensor("new_cache_13_internal_tensor_assign_1_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_13_internal_tensor_assign_1_end_mask_0 = const()[name = tensor("new_cache_13_internal_tensor_assign_1_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_13_internal_tensor_assign_1_squeeze_mask_0 = const()[name = tensor("new_cache_13_internal_tensor_assign_1_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_60 = const()[name = tensor("shape_60"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_12 = const()[name = tensor("reduce_prod_12"), val = tensor(1048576)]; + tensor range_1d_12_start_0 = const()[name = tensor("range_1d_12_start_0"), val = tensor(0)]; + tensor range_1d_12_step_0 = const()[name = tensor("range_1d_12_step_0"), val = tensor(1)]; + tensor range_1d_12 = range_1d(end = reduce_prod_12, start = range_1d_12_start_0, step = range_1d_12_step_0)[name = tensor("range_1d_12")]; + tensor reshape_60 = reshape(shape = shape_60, x = range_1d_12)[name = tensor("reshape_60")]; + tensor slice_by_index_12 = slice_by_index(begin = concat_44, begin_mask = new_cache_13_internal_tensor_assign_1_begin_mask_0, end = concat_45, end_mask = new_cache_13_internal_tensor_assign_1_end_mask_0, squeeze_mask = new_cache_13_internal_tensor_assign_1_squeeze_mask_0, stride = new_cache_13_internal_tensor_assign_1_stride_0, x = reshape_60)[name = tensor("slice_by_index_12")]; + tensor reshape_61_shape_0 = const()[name = tensor("reshape_61_shape_0"), val = tensor([-1])]; + tensor reshape_61 = reshape(shape = reshape_61_shape_0, x = slice_by_index_12)[name = tensor("reshape_61")]; + tensor reshape_62_shape_0 = const()[name = tensor("reshape_62_shape_0"), val = tensor([-1])]; + tensor reshape_62_cast_fp16 = reshape(shape = reshape_62_shape_0, x = var_2990_cast_fp16)[name = tensor("reshape_62_cast_fp16")]; + tensor reshape_63_shape_0 = const()[name = tensor("reshape_63_shape_0"), val = tensor([-1])]; + tensor reshape_63_cast_fp16 = reshape(shape = reshape_63_shape_0, x = cache6_to_fp16)[name = tensor("reshape_63_cast_fp16")]; + tensor scatter_12_mode_0 = const()[name = tensor("scatter_12_mode_0"), val = tensor("update")]; + tensor scatter_12_axis_0 = const()[name = tensor("scatter_12_axis_0"), val = tensor(0)]; + tensor scatter_12_validate_indices_0 = const()[name = tensor("scatter_12_validate_indices_0"), val = tensor(false)]; + tensor scatter_12_cast_fp16 = scatter(axis = scatter_12_axis_0, data = reshape_63_cast_fp16, indices = reshape_61, mode = scatter_12_mode_0, updates = reshape_62_cast_fp16, validate_indices = scatter_12_validate_indices_0)[name = tensor("scatter_12_cast_fp16")]; + tensor reshape_64_cast_fp16 = reshape(shape = shape_60, x = scatter_12_cast_fp16)[name = tensor("reshape_64_cast_fp16")]; + tensor var_2998_begin_0 = const()[name = tensor("op_2998_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor var_2998_end_0 = const()[name = tensor("op_2998_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor var_2998_end_mask_0 = const()[name = tensor("op_2998_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_2998_squeeze_mask_0 = const()[name = tensor("op_2998_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor var_2998_cast_fp16 = slice_by_index(begin = var_2998_begin_0, end = var_2998_end_0, end_mask = var_2998_end_mask_0, squeeze_mask = var_2998_squeeze_mask_0, x = reshape_64_cast_fp16)[name = tensor("op_2998_cast_fp16")]; + tensor var_3000_axis_0 = const()[name = tensor("op_3000_axis_0"), val = tensor(1)]; + tensor var_3000_mode_0 = const()[name = tensor("op_3000_mode_0"), val = tensor("update")]; + tensor var_3000_validate_indices_0 = const()[name = tensor("op_3000_validate_indices_0"), val = tensor(false)]; + tensor var_3000_cast_fp16 = scatter_along_axis(axis = var_3000_axis_0, data = var_2998_cast_fp16, indices = write_indices_13, mode = var_3000_mode_0, updates = v_13_cast_fp16, validate_indices = var_3000_validate_indices_0)[name = tensor("op_3000_cast_fp16")]; + tensor concat_46 = const()[name = tensor("concat_46"), val = tensor([1, 0, 0, 0, 0])]; + tensor concat_47 = const()[name = tensor("concat_47"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_13_internal_tensor_assign_2_stride_0 = const()[name = tensor("new_cache_13_internal_tensor_assign_2_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_13_internal_tensor_assign_2_begin_mask_0 = const()[name = tensor("new_cache_13_internal_tensor_assign_2_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_13_internal_tensor_assign_2_end_mask_0 = const()[name = tensor("new_cache_13_internal_tensor_assign_2_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_13_internal_tensor_assign_2_squeeze_mask_0 = const()[name = tensor("new_cache_13_internal_tensor_assign_2_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_61 = const()[name = tensor("shape_61"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_13 = const()[name = tensor("reduce_prod_13"), val = tensor(1048576)]; + tensor range_1d_13_start_0 = const()[name = tensor("range_1d_13_start_0"), val = tensor(0)]; + tensor range_1d_13_step_0 = const()[name = tensor("range_1d_13_step_0"), val = tensor(1)]; + tensor range_1d_13 = range_1d(end = reduce_prod_13, start = range_1d_13_start_0, step = range_1d_13_step_0)[name = tensor("range_1d_13")]; + tensor reshape_65 = reshape(shape = shape_61, x = range_1d_13)[name = tensor("reshape_65")]; + tensor slice_by_index_13 = slice_by_index(begin = concat_46, begin_mask = new_cache_13_internal_tensor_assign_2_begin_mask_0, end = concat_47, end_mask = new_cache_13_internal_tensor_assign_2_end_mask_0, squeeze_mask = new_cache_13_internal_tensor_assign_2_squeeze_mask_0, stride = new_cache_13_internal_tensor_assign_2_stride_0, x = reshape_65)[name = tensor("slice_by_index_13")]; + tensor reshape_66_shape_0 = const()[name = tensor("reshape_66_shape_0"), val = tensor([-1])]; + tensor reshape_66 = reshape(shape = reshape_66_shape_0, x = slice_by_index_13)[name = tensor("reshape_66")]; + tensor reshape_67_shape_0 = const()[name = tensor("reshape_67_shape_0"), val = tensor([-1])]; + tensor reshape_67_cast_fp16 = reshape(shape = reshape_67_shape_0, x = var_3000_cast_fp16)[name = tensor("reshape_67_cast_fp16")]; + tensor reshape_68_shape_0 = const()[name = tensor("reshape_68_shape_0"), val = tensor([-1])]; + tensor reshape_68_cast_fp16 = reshape(shape = reshape_68_shape_0, x = reshape_64_cast_fp16)[name = tensor("reshape_68_cast_fp16")]; + tensor scatter_13_mode_0 = const()[name = tensor("scatter_13_mode_0"), val = tensor("update")]; + tensor scatter_13_axis_0 = const()[name = tensor("scatter_13_axis_0"), val = tensor(0)]; + tensor scatter_13_validate_indices_0 = const()[name = tensor("scatter_13_validate_indices_0"), val = tensor(false)]; + tensor scatter_13_cast_fp16 = scatter(axis = scatter_13_axis_0, data = reshape_68_cast_fp16, indices = reshape_66, mode = scatter_13_mode_0, updates = reshape_67_cast_fp16, validate_indices = scatter_13_validate_indices_0)[name = tensor("scatter_13_cast_fp16")]; + tensor reshape_69_cast_fp16 = reshape(shape = shape_61, x = scatter_13_cast_fp16)[name = tensor("reshape_69_cast_fp16")]; + tensor reshape_69_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("reshape_69_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor keys_37_begin_0 = const()[name = tensor("keys_37_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor keys_37_end_0 = const()[name = tensor("keys_37_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor keys_37_end_mask_0 = const()[name = tensor("keys_37_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor keys_37_squeeze_mask_0 = const()[name = tensor("keys_37_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor keys_37_cast_fp16 = slice_by_index(begin = keys_37_begin_0, end = keys_37_end_0, end_mask = keys_37_end_mask_0, squeeze_mask = keys_37_squeeze_mask_0, x = reshape_69_cast_fp16)[name = tensor("keys_37_cast_fp16")]; + tensor values_37_begin_0 = const()[name = tensor("values_37_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor values_37_end_0 = const()[name = tensor("values_37_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor values_37_end_mask_0 = const()[name = tensor("values_37_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor values_37_squeeze_mask_0 = const()[name = tensor("values_37_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor values_37_cast_fp16 = slice_by_index(begin = values_37_begin_0, end = values_37_end_0, end_mask = values_37_end_mask_0, squeeze_mask = values_37_squeeze_mask_0, x = reshape_69_cast_fp16)[name = tensor("values_37_cast_fp16")]; + tensor var_3012_cast_fp16 = not_equal(x = keys_37_cast_fp16, y = keys_37_cast_fp16)[name = tensor("op_3012_cast_fp16")]; + tensor keys_39_cast_fp16 = select(a = var_504_to_fp16, b = keys_37_cast_fp16, cond = var_3012_cast_fp16)[name = tensor("keys_39_cast_fp16")]; + tensor var_3020_cast_fp16 = not_equal(x = values_37_cast_fp16, y = values_37_cast_fp16)[name = tensor("op_3020_cast_fp16")]; + tensor values_39_cast_fp16 = select(a = var_504_to_fp16, b = values_37_cast_fp16, cond = var_3020_cast_fp16)[name = tensor("values_39_cast_fp16")]; + tensor var_3044 = const()[name = tensor("op_3044"), val = tensor([0, 2, 1, 3])]; + tensor var_3057 = const()[name = tensor("op_3057"), val = tensor([1, 1, 1])]; + tensor var_3058_cast_fp16 = reshape(shape = var_3057, x = position6_to_fp16)[name = tensor("op_3058_cast_fp16")]; + tensor var_3075_to_fp16 = const()[name = tensor("op_3075_to_fp16"), val = tensor(0x1p+0)]; + tensor valid_len_13_cast_fp16 = add(x = var_3058_cast_fp16, y = var_3075_to_fp16)[name = tensor("valid_len_13_cast_fp16")]; + tensor valid_mask_13_cast_fp16 = less(x = k_positions_1_promoted_to_fp16, y = valid_len_13_cast_fp16)[name = tensor("valid_mask_13_cast_fp16")]; + tensor causal_mask_13_cast_fp16 = less_equal(x = k_positions_1_promoted_to_fp16, y = var_3058_cast_fp16)[name = tensor("causal_mask_13_cast_fp16")]; + tensor attn_mask_25 = logical_and(x = valid_mask_13_cast_fp16, y = causal_mask_13_cast_fp16)[name = tensor("attn_mask_25")]; + tensor attn_mask_27_axes_0 = const()[name = tensor("attn_mask_27_axes_0"), val = tensor([1])]; + tensor attn_mask_27 = expand_dims(axes = attn_mask_27_axes_0, x = attn_mask_25)[name = tensor("attn_mask_27")]; + tensor var_3093_transpose_x_0 = const()[name = tensor("op_3093_transpose_x_0"), val = tensor(false)]; + tensor var_3093_transpose_y_0 = const()[name = tensor("op_3093_transpose_y_0"), val = tensor(false)]; + tensor transpose_84_perm_0 = const()[name = tensor("transpose_84_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_85_perm_0 = const()[name = tensor("transpose_85_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_85 = transpose(perm = transpose_85_perm_0, x = keys_39_cast_fp16)[name = tensor("transpose_189")]; + tensor transpose_84 = transpose(perm = transpose_84_perm_0, x = q_39_cast_fp16)[name = tensor("transpose_190")]; + tensor var_3093_cast_fp16 = matmul(transpose_x = var_3093_transpose_x_0, transpose_y = var_3093_transpose_y_0, x = transpose_84, y = transpose_85)[name = tensor("op_3093_cast_fp16")]; + tensor var_3087_to_fp16 = const()[name = tensor("op_3087_to_fp16"), val = tensor([0x1p-3])]; + tensor attn_weights_37_cast_fp16 = mul(x = var_3093_cast_fp16, y = var_3087_to_fp16)[name = tensor("attn_weights_37_cast_fp16")]; + tensor var_3095 = logical_not(x = attn_mask_27)[name = tensor("op_3095")]; + tensor var_3096_to_fp16 = const()[name = tensor("op_3096_to_fp16"), val = tensor(-inf)]; + tensor attn_weights_39_cast_fp16 = select(a = var_3096_to_fp16, b = attn_weights_37_cast_fp16, cond = var_3095)[name = tensor("attn_weights_39_cast_fp16")]; + tensor var_3098 = const()[name = tensor("op_3098"), val = tensor(-1)]; + tensor attn_weights_41_cast_fp16 = softmax(axis = var_3098, x = attn_weights_39_cast_fp16)[name = tensor("attn_weights_41_cast_fp16")]; + tensor attn_output_13_transpose_x_0 = const()[name = tensor("attn_output_13_transpose_x_0"), val = tensor(false)]; + tensor attn_output_13_transpose_y_0 = const()[name = tensor("attn_output_13_transpose_y_0"), val = tensor(false)]; + tensor values_41_cast_fp16 = transpose(perm = var_3044, x = values_39_cast_fp16)[name = tensor("transpose_191")]; + tensor attn_output_13_cast_fp16 = matmul(transpose_x = attn_output_13_transpose_x_0, transpose_y = attn_output_13_transpose_y_0, x = attn_weights_41_cast_fp16, y = values_41_cast_fp16)[name = tensor("attn_output_13_cast_fp16")]; + tensor var_3106 = const()[name = tensor("op_3106"), val = tensor([0, 2, 1, 3])]; + tensor var_3109 = const()[name = tensor("op_3109"), val = tensor([1, 1, 1024])]; + tensor var_3107_cast_fp16 = transpose(perm = var_3106, x = attn_output_13_cast_fp16)[name = tensor("transpose_188")]; + tensor input_65_cast_fp16 = reshape(shape = var_3109, x = var_3107_cast_fp16)[name = tensor("input_65_cast_fp16")]; + tensor linear_26_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = attn6_out_proj_weight_cast_fp16, x = input_65_cast_fp16)[name = tensor("linear_26_cast_fp16")]; + tensor var_3115_to_fp16 = const()[name = tensor("op_3115_to_fp16"), val = tensor(0x1p+0)]; + tensor var_3116_cast_fp16 = add(x = position6_to_fp16, y = var_3115_to_fp16)[name = tensor("op_3116_cast_fp16")]; + tensor var_3116_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_3116_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor input_67_cast_fp16 = add(x = input_63_cast_fp16, y = linear_26_cast_fp16)[name = tensor("input_67_cast_fp16")]; + tensor input_69_axes_0 = const()[name = tensor("input_69_axes_0"), val = tensor([-1])]; + tensor norm6_2_weight_to_fp16 = const()[name = tensor("norm6_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(303640704)))]; + tensor norm6_2_bias_to_fp16 = const()[name = tensor("norm6_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(303642816)))]; + tensor var_3120_to_fp16 = const()[name = tensor("op_3120_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_69_cast_fp16 = layer_norm(axes = input_69_axes_0, beta = norm6_2_bias_to_fp16, epsilon = var_3120_to_fp16, gamma = norm6_2_weight_to_fp16, x = input_67_cast_fp16)[name = tensor("input_69_cast_fp16")]; + tensor linear_27_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = linear6_1_weight_cast_fp16, x = input_69_cast_fp16)[name = tensor("linear_27_cast_fp16")]; + tensor input_71_mode_0 = const()[name = tensor("input_71_mode_0"), val = tensor("EXACT")]; + tensor input_71_cast_fp16 = gelu(mode = input_71_mode_0, x = linear_27_cast_fp16)[name = tensor("input_71_cast_fp16")]; + tensor linear_28_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = linear6_2_weight_cast_fp16, x = input_71_cast_fp16)[name = tensor("linear_28_cast_fp16")]; + tensor input_73_cast_fp16 = add(x = input_67_cast_fp16, y = linear_28_cast_fp16)[name = tensor("input_73_cast_fp16")]; + tensor x_15_axes_0 = const()[name = tensor("x_15_axes_0"), val = tensor([-1])]; + tensor norm7_1_weight_to_fp16 = const()[name = tensor("norm7_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(303644928)))]; + tensor norm7_1_bias_to_fp16 = const()[name = tensor("norm7_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(303647040)))]; + tensor var_3137_to_fp16 = const()[name = tensor("op_3137_to_fp16"), val = tensor(0x1.5p-17)]; + tensor x_15_cast_fp16 = layer_norm(axes = x_15_axes_0, beta = norm7_1_bias_to_fp16, epsilon = var_3137_to_fp16, gamma = norm7_1_weight_to_fp16, x = input_73_cast_fp16)[name = tensor("x_15_cast_fp16")]; + tensor linear_29_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = attn7_in_proj_weight_cast_fp16, x = x_15_cast_fp16)[name = tensor("linear_29_cast_fp16")]; + tensor var_3173 = const()[name = tensor("op_3173"), val = tensor([1, 1, 3, 16, 64])]; + tensor qkv_15_cast_fp16 = reshape(shape = var_3173, x = linear_29_cast_fp16)[name = tensor("qkv_15_cast_fp16")]; + tensor q_43_begin_0 = const()[name = tensor("q_43_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor q_43_end_0 = const()[name = tensor("q_43_end_0"), val = tensor([1, 1, 1, 16, 64])]; + tensor q_43_end_mask_0 = const()[name = tensor("q_43_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor q_43_squeeze_mask_0 = const()[name = tensor("q_43_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor q_43_cast_fp16 = slice_by_index(begin = q_43_begin_0, end = q_43_end_0, end_mask = q_43_end_mask_0, squeeze_mask = q_43_squeeze_mask_0, x = qkv_15_cast_fp16)[name = tensor("q_43_cast_fp16")]; + tensor k_29_begin_0 = const()[name = tensor("k_29_begin_0"), val = tensor([0, 0, 1, 0, 0])]; + tensor k_29_end_0 = const()[name = tensor("k_29_end_0"), val = tensor([1, 1, 2, 16, 64])]; + tensor k_29_end_mask_0 = const()[name = tensor("k_29_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor k_29_squeeze_mask_0 = const()[name = tensor("k_29_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor k_29_cast_fp16 = slice_by_index(begin = k_29_begin_0, end = k_29_end_0, end_mask = k_29_end_mask_0, squeeze_mask = k_29_squeeze_mask_0, x = qkv_15_cast_fp16)[name = tensor("k_29_cast_fp16")]; + tensor v_15_begin_0 = const()[name = tensor("v_15_begin_0"), val = tensor([0, 0, 2, 0, 0])]; + tensor v_15_end_0 = const()[name = tensor("v_15_end_0"), val = tensor([1, 1, 3, 16, 64])]; + tensor v_15_end_mask_0 = const()[name = tensor("v_15_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor v_15_squeeze_mask_0 = const()[name = tensor("v_15_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor v_15_cast_fp16 = slice_by_index(begin = v_15_begin_0, end = v_15_end_0, end_mask = v_15_end_mask_0, squeeze_mask = v_15_squeeze_mask_0, x = qkv_15_cast_fp16)[name = tensor("v_15_cast_fp16")]; + tensor var_3277 = const()[name = tensor("op_3277"), val = tensor([1, 1, 1, 1])]; + tensor position7_to_fp16_dtype_0 = const()[name = tensor("position7_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor position7_to_fp16 = cast(dtype = position7_to_fp16_dtype_0, x = position7)[name = tensor("cast_518")]; + tensor ts_47_cast_fp16 = reshape(shape = var_3277, x = position7_to_fp16)[name = tensor("ts_47_cast_fp16")]; + tensor var_3281 = const()[name = tensor("op_3281"), val = tensor([1, 1, 16, 32, 2])]; + tensor q_complex_15_cast_fp16 = reshape(shape = var_3281, x = q_43_cast_fp16)[name = tensor("q_complex_15_cast_fp16")]; + tensor var_3285 = const()[name = tensor("op_3285"), val = tensor([1, 1, 16, 32, 2])]; + tensor k_complex_15_cast_fp16 = reshape(shape = var_3285, x = k_29_cast_fp16)[name = tensor("k_complex_15_cast_fp16")]; + tensor var_3289_begin_0 = const()[name = tensor("op_3289_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_3289_end_0 = const()[name = tensor("op_3289_end_0"), val = tensor([1, 1, 16, 32, 1])]; + tensor var_3289_end_mask_0 = const()[name = tensor("op_3289_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_3289_squeeze_mask_0 = const()[name = tensor("op_3289_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_3289_cast_fp16 = slice_by_index(begin = var_3289_begin_0, end = var_3289_end_0, end_mask = var_3289_end_mask_0, squeeze_mask = var_3289_squeeze_mask_0, x = q_complex_15_cast_fp16)[name = tensor("op_3289_cast_fp16")]; + tensor var_3297_begin_0 = const()[name = tensor("op_3297_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_3297_end_0 = const()[name = tensor("op_3297_end_0"), val = tensor([1, 1, 16, 32, 2])]; + tensor var_3297_end_mask_0 = const()[name = tensor("op_3297_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_3297_squeeze_mask_0 = const()[name = tensor("op_3297_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_3297_cast_fp16 = slice_by_index(begin = var_3297_begin_0, end = var_3297_end_0, end_mask = var_3297_end_mask_0, squeeze_mask = var_3297_squeeze_mask_0, x = q_complex_15_cast_fp16)[name = tensor("op_3297_cast_fp16")]; + tensor var_3305_begin_0 = const()[name = tensor("op_3305_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_3305_end_0 = const()[name = tensor("op_3305_end_0"), val = tensor([1, 1, 16, 32, 1])]; + tensor var_3305_end_mask_0 = const()[name = tensor("op_3305_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_3305_squeeze_mask_0 = const()[name = tensor("op_3305_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_3305_cast_fp16 = slice_by_index(begin = var_3305_begin_0, end = var_3305_end_0, end_mask = var_3305_end_mask_0, squeeze_mask = var_3305_squeeze_mask_0, x = k_complex_15_cast_fp16)[name = tensor("op_3305_cast_fp16")]; + tensor var_3313_begin_0 = const()[name = tensor("op_3313_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_3313_end_0 = const()[name = tensor("op_3313_end_0"), val = tensor([1, 1, 16, 32, 2])]; + tensor var_3313_end_mask_0 = const()[name = tensor("op_3313_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_3313_squeeze_mask_0 = const()[name = tensor("op_3313_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_3313_cast_fp16 = slice_by_index(begin = var_3313_begin_0, end = var_3313_end_0, end_mask = var_3313_end_mask_0, squeeze_mask = var_3313_squeeze_mask_0, x = k_complex_15_cast_fp16)[name = tensor("op_3313_cast_fp16")]; + tensor freqs_15_to_fp16 = const()[name = tensor("freqs_15_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(303649152)))]; + tensor var_3319_cast_fp16 = mul(x = freqs_15_to_fp16, y = ts_47_cast_fp16)[name = tensor("op_3319_cast_fp16")]; + tensor rotr_15_cast_fp16 = cos(x = var_3319_cast_fp16)[name = tensor("rotr_15_cast_fp16")]; + tensor roti_15_cast_fp16 = sin(x = var_3319_cast_fp16)[name = tensor("roti_15_cast_fp16")]; + tensor var_3323_cast_fp16 = mul(x = var_3289_cast_fp16, y = rotr_15_cast_fp16)[name = tensor("op_3323_cast_fp16")]; + tensor var_3324_cast_fp16 = mul(x = var_3297_cast_fp16, y = roti_15_cast_fp16)[name = tensor("op_3324_cast_fp16")]; + tensor qor_29_cast_fp16 = sub(x = var_3323_cast_fp16, y = var_3324_cast_fp16)[name = tensor("qor_29_cast_fp16")]; + tensor var_3327_cast_fp16 = mul(x = var_3289_cast_fp16, y = roti_15_cast_fp16)[name = tensor("op_3327_cast_fp16")]; + tensor var_3328_cast_fp16 = mul(x = var_3297_cast_fp16, y = rotr_15_cast_fp16)[name = tensor("op_3328_cast_fp16")]; + tensor qoi_29_cast_fp16 = add(x = var_3327_cast_fp16, y = var_3328_cast_fp16)[name = tensor("qoi_29_cast_fp16")]; + tensor var_3331_cast_fp16 = mul(x = var_3305_cast_fp16, y = rotr_15_cast_fp16)[name = tensor("op_3331_cast_fp16")]; + tensor var_3332_cast_fp16 = mul(x = var_3313_cast_fp16, y = roti_15_cast_fp16)[name = tensor("op_3332_cast_fp16")]; + tensor kor_29_cast_fp16 = sub(x = var_3331_cast_fp16, y = var_3332_cast_fp16)[name = tensor("kor_29_cast_fp16")]; + tensor var_3335_cast_fp16 = mul(x = var_3305_cast_fp16, y = roti_15_cast_fp16)[name = tensor("op_3335_cast_fp16")]; + tensor var_3336_cast_fp16 = mul(x = var_3313_cast_fp16, y = rotr_15_cast_fp16)[name = tensor("op_3336_cast_fp16")]; + tensor koi_29_cast_fp16 = add(x = var_3335_cast_fp16, y = var_3336_cast_fp16)[name = tensor("koi_29_cast_fp16")]; + tensor qo_15_axis_0 = const()[name = tensor("qo_15_axis_0"), val = tensor(-1)]; + tensor qo_15_cast_fp16 = stack(axis = qo_15_axis_0, values = (qor_29_cast_fp16, qoi_29_cast_fp16))[name = tensor("qo_15_cast_fp16")]; + tensor ko_15_axis_0 = const()[name = tensor("ko_15_axis_0"), val = tensor(-1)]; + tensor ko_15_cast_fp16 = stack(axis = ko_15_axis_0, values = (kor_29_cast_fp16, koi_29_cast_fp16))[name = tensor("ko_15_cast_fp16")]; + tensor var_3365 = const()[name = tensor("op_3365"), val = tensor([1, 1, 16, 64])]; + tensor q_45_cast_fp16 = reshape(shape = var_3365, x = qo_15_cast_fp16)[name = tensor("q_45_cast_fp16")]; + tensor var_3367 = const()[name = tensor("op_3367"), val = tensor([1, 1, 16, 64])]; + tensor k_31_cast_fp16 = reshape(shape = var_3367, x = ko_15_cast_fp16)[name = tensor("k_31_cast_fp16")]; + tensor _inversed_3389_y_0_to_fp16 = const()[name = tensor("_inversed_3389_y_0_to_fp16"), val = tensor(0x1p-9)]; + tensor _inversed_3389_cast_fp16 = mul(x = ts_47_cast_fp16, y = _inversed_3389_y_0_to_fp16)[name = tensor("_inversed_3389_cast_fp16")]; + tensor var_3390_cast_fp16 = floor(x = _inversed_3389_cast_fp16)[name = tensor("op_3390_cast_fp16")]; + tensor var_3391_to_fp16 = const()[name = tensor("op_3391_to_fp16"), val = tensor(0x1p+9)]; + tensor var_3392_cast_fp16 = mul(x = var_3390_cast_fp16, y = var_3391_to_fp16)[name = tensor("op_3392_cast_fp16")]; + tensor write_indices_float_31_cast_fp16 = sub(x = ts_47_cast_fp16, y = var_3392_cast_fp16)[name = tensor("write_indices_float_31_cast_fp16")]; + tensor var_3399_dtype_0 = const()[name = tensor("op_3399_dtype_0"), val = tensor("int32")]; + tensor write_indices_15_reps_0 = const()[name = tensor("write_indices_15_reps_0"), val = tensor([1, 1, 16, 64])]; + tensor write_indices_float_31_cast_fp16_to_int32 = cast(dtype = var_3399_dtype_0, x = write_indices_float_31_cast_fp16)[name = tensor("cast_517")]; + tensor write_indices_15 = tile(reps = write_indices_15_reps_0, x = write_indices_float_31_cast_fp16_to_int32)[name = tensor("write_indices_15")]; + tensor var_3407_begin_0 = const()[name = tensor("op_3407_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_3407_end_0 = const()[name = tensor("op_3407_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor var_3407_end_mask_0 = const()[name = tensor("op_3407_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_3407_squeeze_mask_0 = const()[name = tensor("op_3407_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor cache7_to_fp16_dtype_0 = const()[name = tensor("cache7_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor cache7_to_fp16 = cast(dtype = cache7_to_fp16_dtype_0, x = cache7)[name = tensor("cast_516")]; + tensor var_3407_cast_fp16 = slice_by_index(begin = var_3407_begin_0, end = var_3407_end_0, end_mask = var_3407_end_mask_0, squeeze_mask = var_3407_squeeze_mask_0, x = cache7_to_fp16)[name = tensor("op_3407_cast_fp16")]; + tensor var_3409_axis_0 = const()[name = tensor("op_3409_axis_0"), val = tensor(1)]; + tensor var_3409_mode_0 = const()[name = tensor("op_3409_mode_0"), val = tensor("update")]; + tensor var_3409_validate_indices_0 = const()[name = tensor("op_3409_validate_indices_0"), val = tensor(false)]; + tensor var_3409_cast_fp16 = scatter_along_axis(axis = var_3409_axis_0, data = var_3407_cast_fp16, indices = write_indices_15, mode = var_3409_mode_0, updates = k_31_cast_fp16, validate_indices = var_3409_validate_indices_0)[name = tensor("op_3409_cast_fp16")]; + tensor concat_51 = const()[name = tensor("concat_51"), val = tensor([0, 0, 0, 0, 0])]; + tensor concat_52 = const()[name = tensor("concat_52"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_15_internal_tensor_assign_1_stride_0 = const()[name = tensor("new_cache_15_internal_tensor_assign_1_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_15_internal_tensor_assign_1_begin_mask_0 = const()[name = tensor("new_cache_15_internal_tensor_assign_1_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_15_internal_tensor_assign_1_end_mask_0 = const()[name = tensor("new_cache_15_internal_tensor_assign_1_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_15_internal_tensor_assign_1_squeeze_mask_0 = const()[name = tensor("new_cache_15_internal_tensor_assign_1_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_62 = const()[name = tensor("shape_62"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_14 = const()[name = tensor("reduce_prod_14"), val = tensor(1048576)]; + tensor range_1d_14_start_0 = const()[name = tensor("range_1d_14_start_0"), val = tensor(0)]; + tensor range_1d_14_step_0 = const()[name = tensor("range_1d_14_step_0"), val = tensor(1)]; + tensor range_1d_14 = range_1d(end = reduce_prod_14, start = range_1d_14_start_0, step = range_1d_14_step_0)[name = tensor("range_1d_14")]; + tensor reshape_70 = reshape(shape = shape_62, x = range_1d_14)[name = tensor("reshape_70")]; + tensor slice_by_index_14 = slice_by_index(begin = concat_51, begin_mask = new_cache_15_internal_tensor_assign_1_begin_mask_0, end = concat_52, end_mask = new_cache_15_internal_tensor_assign_1_end_mask_0, squeeze_mask = new_cache_15_internal_tensor_assign_1_squeeze_mask_0, stride = new_cache_15_internal_tensor_assign_1_stride_0, x = reshape_70)[name = tensor("slice_by_index_14")]; + tensor reshape_71_shape_0 = const()[name = tensor("reshape_71_shape_0"), val = tensor([-1])]; + tensor reshape_71 = reshape(shape = reshape_71_shape_0, x = slice_by_index_14)[name = tensor("reshape_71")]; + tensor reshape_72_shape_0 = const()[name = tensor("reshape_72_shape_0"), val = tensor([-1])]; + tensor reshape_72_cast_fp16 = reshape(shape = reshape_72_shape_0, x = var_3409_cast_fp16)[name = tensor("reshape_72_cast_fp16")]; + tensor reshape_73_shape_0 = const()[name = tensor("reshape_73_shape_0"), val = tensor([-1])]; + tensor reshape_73_cast_fp16 = reshape(shape = reshape_73_shape_0, x = cache7_to_fp16)[name = tensor("reshape_73_cast_fp16")]; + tensor scatter_14_mode_0 = const()[name = tensor("scatter_14_mode_0"), val = tensor("update")]; + tensor scatter_14_axis_0 = const()[name = tensor("scatter_14_axis_0"), val = tensor(0)]; + tensor scatter_14_validate_indices_0 = const()[name = tensor("scatter_14_validate_indices_0"), val = tensor(false)]; + tensor scatter_14_cast_fp16 = scatter(axis = scatter_14_axis_0, data = reshape_73_cast_fp16, indices = reshape_71, mode = scatter_14_mode_0, updates = reshape_72_cast_fp16, validate_indices = scatter_14_validate_indices_0)[name = tensor("scatter_14_cast_fp16")]; + tensor reshape_74_cast_fp16 = reshape(shape = shape_62, x = scatter_14_cast_fp16)[name = tensor("reshape_74_cast_fp16")]; + tensor var_3417_begin_0 = const()[name = tensor("op_3417_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor var_3417_end_0 = const()[name = tensor("op_3417_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor var_3417_end_mask_0 = const()[name = tensor("op_3417_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_3417_squeeze_mask_0 = const()[name = tensor("op_3417_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor var_3417_cast_fp16 = slice_by_index(begin = var_3417_begin_0, end = var_3417_end_0, end_mask = var_3417_end_mask_0, squeeze_mask = var_3417_squeeze_mask_0, x = reshape_74_cast_fp16)[name = tensor("op_3417_cast_fp16")]; + tensor var_3419_axis_0 = const()[name = tensor("op_3419_axis_0"), val = tensor(1)]; + tensor var_3419_mode_0 = const()[name = tensor("op_3419_mode_0"), val = tensor("update")]; + tensor var_3419_validate_indices_0 = const()[name = tensor("op_3419_validate_indices_0"), val = tensor(false)]; + tensor var_3419_cast_fp16 = scatter_along_axis(axis = var_3419_axis_0, data = var_3417_cast_fp16, indices = write_indices_15, mode = var_3419_mode_0, updates = v_15_cast_fp16, validate_indices = var_3419_validate_indices_0)[name = tensor("op_3419_cast_fp16")]; + tensor concat_53 = const()[name = tensor("concat_53"), val = tensor([1, 0, 0, 0, 0])]; + tensor concat_54 = const()[name = tensor("concat_54"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_15_internal_tensor_assign_2_stride_0 = const()[name = tensor("new_cache_15_internal_tensor_assign_2_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_15_internal_tensor_assign_2_begin_mask_0 = const()[name = tensor("new_cache_15_internal_tensor_assign_2_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_15_internal_tensor_assign_2_end_mask_0 = const()[name = tensor("new_cache_15_internal_tensor_assign_2_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_15_internal_tensor_assign_2_squeeze_mask_0 = const()[name = tensor("new_cache_15_internal_tensor_assign_2_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_63 = const()[name = tensor("shape_63"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_15 = const()[name = tensor("reduce_prod_15"), val = tensor(1048576)]; + tensor range_1d_15_start_0 = const()[name = tensor("range_1d_15_start_0"), val = tensor(0)]; + tensor range_1d_15_step_0 = const()[name = tensor("range_1d_15_step_0"), val = tensor(1)]; + tensor range_1d_15 = range_1d(end = reduce_prod_15, start = range_1d_15_start_0, step = range_1d_15_step_0)[name = tensor("range_1d_15")]; + tensor reshape_75 = reshape(shape = shape_63, x = range_1d_15)[name = tensor("reshape_75")]; + tensor slice_by_index_15 = slice_by_index(begin = concat_53, begin_mask = new_cache_15_internal_tensor_assign_2_begin_mask_0, end = concat_54, end_mask = new_cache_15_internal_tensor_assign_2_end_mask_0, squeeze_mask = new_cache_15_internal_tensor_assign_2_squeeze_mask_0, stride = new_cache_15_internal_tensor_assign_2_stride_0, x = reshape_75)[name = tensor("slice_by_index_15")]; + tensor reshape_76_shape_0 = const()[name = tensor("reshape_76_shape_0"), val = tensor([-1])]; + tensor reshape_76 = reshape(shape = reshape_76_shape_0, x = slice_by_index_15)[name = tensor("reshape_76")]; + tensor reshape_77_shape_0 = const()[name = tensor("reshape_77_shape_0"), val = tensor([-1])]; + tensor reshape_77_cast_fp16 = reshape(shape = reshape_77_shape_0, x = var_3419_cast_fp16)[name = tensor("reshape_77_cast_fp16")]; + tensor reshape_78_shape_0 = const()[name = tensor("reshape_78_shape_0"), val = tensor([-1])]; + tensor reshape_78_cast_fp16 = reshape(shape = reshape_78_shape_0, x = reshape_74_cast_fp16)[name = tensor("reshape_78_cast_fp16")]; + tensor scatter_15_mode_0 = const()[name = tensor("scatter_15_mode_0"), val = tensor("update")]; + tensor scatter_15_axis_0 = const()[name = tensor("scatter_15_axis_0"), val = tensor(0)]; + tensor scatter_15_validate_indices_0 = const()[name = tensor("scatter_15_validate_indices_0"), val = tensor(false)]; + tensor scatter_15_cast_fp16 = scatter(axis = scatter_15_axis_0, data = reshape_78_cast_fp16, indices = reshape_76, mode = scatter_15_mode_0, updates = reshape_77_cast_fp16, validate_indices = scatter_15_validate_indices_0)[name = tensor("scatter_15_cast_fp16")]; + tensor reshape_79_cast_fp16 = reshape(shape = shape_63, x = scatter_15_cast_fp16)[name = tensor("reshape_79_cast_fp16")]; + tensor reshape_79_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("reshape_79_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor keys_43_begin_0 = const()[name = tensor("keys_43_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor keys_43_end_0 = const()[name = tensor("keys_43_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor keys_43_end_mask_0 = const()[name = tensor("keys_43_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor keys_43_squeeze_mask_0 = const()[name = tensor("keys_43_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor keys_43_cast_fp16 = slice_by_index(begin = keys_43_begin_0, end = keys_43_end_0, end_mask = keys_43_end_mask_0, squeeze_mask = keys_43_squeeze_mask_0, x = reshape_79_cast_fp16)[name = tensor("keys_43_cast_fp16")]; + tensor values_43_begin_0 = const()[name = tensor("values_43_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor values_43_end_0 = const()[name = tensor("values_43_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor values_43_end_mask_0 = const()[name = tensor("values_43_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor values_43_squeeze_mask_0 = const()[name = tensor("values_43_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor values_43_cast_fp16 = slice_by_index(begin = values_43_begin_0, end = values_43_end_0, end_mask = values_43_end_mask_0, squeeze_mask = values_43_squeeze_mask_0, x = reshape_79_cast_fp16)[name = tensor("values_43_cast_fp16")]; + tensor var_3431_cast_fp16 = not_equal(x = keys_43_cast_fp16, y = keys_43_cast_fp16)[name = tensor("op_3431_cast_fp16")]; + tensor keys_45_cast_fp16 = select(a = var_504_to_fp16, b = keys_43_cast_fp16, cond = var_3431_cast_fp16)[name = tensor("keys_45_cast_fp16")]; + tensor var_3439_cast_fp16 = not_equal(x = values_43_cast_fp16, y = values_43_cast_fp16)[name = tensor("op_3439_cast_fp16")]; + tensor values_45_cast_fp16 = select(a = var_504_to_fp16, b = values_43_cast_fp16, cond = var_3439_cast_fp16)[name = tensor("values_45_cast_fp16")]; + tensor var_3463 = const()[name = tensor("op_3463"), val = tensor([0, 2, 1, 3])]; + tensor var_3476 = const()[name = tensor("op_3476"), val = tensor([1, 1, 1])]; + tensor var_3477_cast_fp16 = reshape(shape = var_3476, x = position7_to_fp16)[name = tensor("op_3477_cast_fp16")]; + tensor var_3494_to_fp16 = const()[name = tensor("op_3494_to_fp16"), val = tensor(0x1p+0)]; + tensor valid_len_15_cast_fp16 = add(x = var_3477_cast_fp16, y = var_3494_to_fp16)[name = tensor("valid_len_15_cast_fp16")]; + tensor valid_mask_15_cast_fp16 = less(x = k_positions_1_promoted_to_fp16, y = valid_len_15_cast_fp16)[name = tensor("valid_mask_15_cast_fp16")]; + tensor causal_mask_15_cast_fp16 = less_equal(x = k_positions_1_promoted_to_fp16, y = var_3477_cast_fp16)[name = tensor("causal_mask_15_cast_fp16")]; + tensor attn_mask_29 = logical_and(x = valid_mask_15_cast_fp16, y = causal_mask_15_cast_fp16)[name = tensor("attn_mask_29")]; + tensor attn_mask_31_axes_0 = const()[name = tensor("attn_mask_31_axes_0"), val = tensor([1])]; + tensor attn_mask_31 = expand_dims(axes = attn_mask_31_axes_0, x = attn_mask_29)[name = tensor("attn_mask_31")]; + tensor var_3512_transpose_x_0 = const()[name = tensor("op_3512_transpose_x_0"), val = tensor(false)]; + tensor var_3512_transpose_y_0 = const()[name = tensor("op_3512_transpose_y_0"), val = tensor(false)]; + tensor transpose_86_perm_0 = const()[name = tensor("transpose_86_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_87_perm_0 = const()[name = tensor("transpose_87_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_87 = transpose(perm = transpose_87_perm_0, x = keys_45_cast_fp16)[name = tensor("transpose_185")]; + tensor transpose_86 = transpose(perm = transpose_86_perm_0, x = q_45_cast_fp16)[name = tensor("transpose_186")]; + tensor var_3512_cast_fp16 = matmul(transpose_x = var_3512_transpose_x_0, transpose_y = var_3512_transpose_y_0, x = transpose_86, y = transpose_87)[name = tensor("op_3512_cast_fp16")]; + tensor var_3506_to_fp16 = const()[name = tensor("op_3506_to_fp16"), val = tensor([0x1p-3])]; + tensor attn_weights_43_cast_fp16 = mul(x = var_3512_cast_fp16, y = var_3506_to_fp16)[name = tensor("attn_weights_43_cast_fp16")]; + tensor var_3514 = logical_not(x = attn_mask_31)[name = tensor("op_3514")]; + tensor var_3515_to_fp16 = const()[name = tensor("op_3515_to_fp16"), val = tensor(-inf)]; + tensor attn_weights_45_cast_fp16 = select(a = var_3515_to_fp16, b = attn_weights_43_cast_fp16, cond = var_3514)[name = tensor("attn_weights_45_cast_fp16")]; + tensor var_3517 = const()[name = tensor("op_3517"), val = tensor(-1)]; + tensor attn_weights_47_cast_fp16 = softmax(axis = var_3517, x = attn_weights_45_cast_fp16)[name = tensor("attn_weights_47_cast_fp16")]; + tensor attn_output_15_transpose_x_0 = const()[name = tensor("attn_output_15_transpose_x_0"), val = tensor(false)]; + tensor attn_output_15_transpose_y_0 = const()[name = tensor("attn_output_15_transpose_y_0"), val = tensor(false)]; + tensor values_47_cast_fp16 = transpose(perm = var_3463, x = values_45_cast_fp16)[name = tensor("transpose_187")]; + tensor attn_output_15_cast_fp16 = matmul(transpose_x = attn_output_15_transpose_x_0, transpose_y = attn_output_15_transpose_y_0, x = attn_weights_47_cast_fp16, y = values_47_cast_fp16)[name = tensor("attn_output_15_cast_fp16")]; + tensor var_3525 = const()[name = tensor("op_3525"), val = tensor([0, 2, 1, 3])]; + tensor var_3528 = const()[name = tensor("op_3528"), val = tensor([1, 1, 1024])]; + tensor var_3526_cast_fp16 = transpose(perm = var_3525, x = attn_output_15_cast_fp16)[name = tensor("transpose_184")]; + tensor input_75_cast_fp16 = reshape(shape = var_3528, x = var_3526_cast_fp16)[name = tensor("input_75_cast_fp16")]; + tensor linear_30_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = attn7_out_proj_weight_cast_fp16, x = input_75_cast_fp16)[name = tensor("linear_30_cast_fp16")]; + tensor var_3534_to_fp16 = const()[name = tensor("op_3534_to_fp16"), val = tensor(0x1p+0)]; + tensor var_3535_cast_fp16 = add(x = position7_to_fp16, y = var_3534_to_fp16)[name = tensor("op_3535_cast_fp16")]; + tensor var_3535_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_3535_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor input_77_cast_fp16 = add(x = input_73_cast_fp16, y = linear_30_cast_fp16)[name = tensor("input_77_cast_fp16")]; + tensor input_79_axes_0 = const()[name = tensor("input_79_axes_0"), val = tensor([-1])]; + tensor norm7_2_weight_to_fp16 = const()[name = tensor("norm7_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(303649280)))]; + tensor norm7_2_bias_to_fp16 = const()[name = tensor("norm7_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(303651392)))]; + tensor var_3539_to_fp16 = const()[name = tensor("op_3539_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_79_cast_fp16 = layer_norm(axes = input_79_axes_0, beta = norm7_2_bias_to_fp16, epsilon = var_3539_to_fp16, gamma = norm7_2_weight_to_fp16, x = input_77_cast_fp16)[name = tensor("input_79_cast_fp16")]; + tensor linear_31_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = linear7_1_weight_cast_fp16, x = input_79_cast_fp16)[name = tensor("linear_31_cast_fp16")]; + tensor input_81_mode_0 = const()[name = tensor("input_81_mode_0"), val = tensor("EXACT")]; + tensor input_81_cast_fp16 = gelu(mode = input_81_mode_0, x = linear_31_cast_fp16)[name = tensor("input_81_cast_fp16")]; + tensor linear_32_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = linear7_2_weight_cast_fp16, x = input_81_cast_fp16)[name = tensor("linear_32_cast_fp16")]; + tensor input_83_cast_fp16 = add(x = input_77_cast_fp16, y = linear_32_cast_fp16)[name = tensor("input_83_cast_fp16")]; + tensor x_17_axes_0 = const()[name = tensor("x_17_axes_0"), val = tensor([-1])]; + tensor norm8_1_weight_to_fp16 = const()[name = tensor("norm8_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(303653504)))]; + tensor norm8_1_bias_to_fp16 = const()[name = tensor("norm8_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(303655616)))]; + tensor var_3556_to_fp16 = const()[name = tensor("op_3556_to_fp16"), val = tensor(0x1.5p-17)]; + tensor x_17_cast_fp16 = layer_norm(axes = x_17_axes_0, beta = norm8_1_bias_to_fp16, epsilon = var_3556_to_fp16, gamma = norm8_1_weight_to_fp16, x = input_83_cast_fp16)[name = tensor("x_17_cast_fp16")]; + tensor linear_33_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = attn8_in_proj_weight_cast_fp16, x = x_17_cast_fp16)[name = tensor("linear_33_cast_fp16")]; + tensor var_3592 = const()[name = tensor("op_3592"), val = tensor([1, 1, 3, 16, 64])]; + tensor qkv_17_cast_fp16 = reshape(shape = var_3592, x = linear_33_cast_fp16)[name = tensor("qkv_17_cast_fp16")]; + tensor q_49_begin_0 = const()[name = tensor("q_49_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor q_49_end_0 = const()[name = tensor("q_49_end_0"), val = tensor([1, 1, 1, 16, 64])]; + tensor q_49_end_mask_0 = const()[name = tensor("q_49_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor q_49_squeeze_mask_0 = const()[name = tensor("q_49_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor q_49_cast_fp16 = slice_by_index(begin = q_49_begin_0, end = q_49_end_0, end_mask = q_49_end_mask_0, squeeze_mask = q_49_squeeze_mask_0, x = qkv_17_cast_fp16)[name = tensor("q_49_cast_fp16")]; + tensor k_33_begin_0 = const()[name = tensor("k_33_begin_0"), val = tensor([0, 0, 1, 0, 0])]; + tensor k_33_end_0 = const()[name = tensor("k_33_end_0"), val = tensor([1, 1, 2, 16, 64])]; + tensor k_33_end_mask_0 = const()[name = tensor("k_33_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor k_33_squeeze_mask_0 = const()[name = tensor("k_33_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor k_33_cast_fp16 = slice_by_index(begin = k_33_begin_0, end = k_33_end_0, end_mask = k_33_end_mask_0, squeeze_mask = k_33_squeeze_mask_0, x = qkv_17_cast_fp16)[name = tensor("k_33_cast_fp16")]; + tensor v_17_begin_0 = const()[name = tensor("v_17_begin_0"), val = tensor([0, 0, 2, 0, 0])]; + tensor v_17_end_0 = const()[name = tensor("v_17_end_0"), val = tensor([1, 1, 3, 16, 64])]; + tensor v_17_end_mask_0 = const()[name = tensor("v_17_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor v_17_squeeze_mask_0 = const()[name = tensor("v_17_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor v_17_cast_fp16 = slice_by_index(begin = v_17_begin_0, end = v_17_end_0, end_mask = v_17_end_mask_0, squeeze_mask = v_17_squeeze_mask_0, x = qkv_17_cast_fp16)[name = tensor("v_17_cast_fp16")]; + tensor var_3696 = const()[name = tensor("op_3696"), val = tensor([1, 1, 1, 1])]; + tensor position8_to_fp16_dtype_0 = const()[name = tensor("position8_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor position8_to_fp16 = cast(dtype = position8_to_fp16_dtype_0, x = position8)[name = tensor("cast_513")]; + tensor ts_53_cast_fp16 = reshape(shape = var_3696, x = position8_to_fp16)[name = tensor("ts_53_cast_fp16")]; + tensor var_3700 = const()[name = tensor("op_3700"), val = tensor([1, 1, 16, 32, 2])]; + tensor q_complex_17_cast_fp16 = reshape(shape = var_3700, x = q_49_cast_fp16)[name = tensor("q_complex_17_cast_fp16")]; + tensor var_3704 = const()[name = tensor("op_3704"), val = tensor([1, 1, 16, 32, 2])]; + tensor k_complex_17_cast_fp16 = reshape(shape = var_3704, x = k_33_cast_fp16)[name = tensor("k_complex_17_cast_fp16")]; + tensor var_3708_begin_0 = const()[name = tensor("op_3708_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_3708_end_0 = const()[name = tensor("op_3708_end_0"), val = tensor([1, 1, 16, 32, 1])]; + tensor var_3708_end_mask_0 = const()[name = tensor("op_3708_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_3708_squeeze_mask_0 = const()[name = tensor("op_3708_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_3708_cast_fp16 = slice_by_index(begin = var_3708_begin_0, end = var_3708_end_0, end_mask = var_3708_end_mask_0, squeeze_mask = var_3708_squeeze_mask_0, x = q_complex_17_cast_fp16)[name = tensor("op_3708_cast_fp16")]; + tensor var_3716_begin_0 = const()[name = tensor("op_3716_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_3716_end_0 = const()[name = tensor("op_3716_end_0"), val = tensor([1, 1, 16, 32, 2])]; + tensor var_3716_end_mask_0 = const()[name = tensor("op_3716_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_3716_squeeze_mask_0 = const()[name = tensor("op_3716_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_3716_cast_fp16 = slice_by_index(begin = var_3716_begin_0, end = var_3716_end_0, end_mask = var_3716_end_mask_0, squeeze_mask = var_3716_squeeze_mask_0, x = q_complex_17_cast_fp16)[name = tensor("op_3716_cast_fp16")]; + tensor var_3724_begin_0 = const()[name = tensor("op_3724_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_3724_end_0 = const()[name = tensor("op_3724_end_0"), val = tensor([1, 1, 16, 32, 1])]; + tensor var_3724_end_mask_0 = const()[name = tensor("op_3724_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_3724_squeeze_mask_0 = const()[name = tensor("op_3724_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_3724_cast_fp16 = slice_by_index(begin = var_3724_begin_0, end = var_3724_end_0, end_mask = var_3724_end_mask_0, squeeze_mask = var_3724_squeeze_mask_0, x = k_complex_17_cast_fp16)[name = tensor("op_3724_cast_fp16")]; + tensor var_3732_begin_0 = const()[name = tensor("op_3732_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_3732_end_0 = const()[name = tensor("op_3732_end_0"), val = tensor([1, 1, 16, 32, 2])]; + tensor var_3732_end_mask_0 = const()[name = tensor("op_3732_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_3732_squeeze_mask_0 = const()[name = tensor("op_3732_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_3732_cast_fp16 = slice_by_index(begin = var_3732_begin_0, end = var_3732_end_0, end_mask = var_3732_end_mask_0, squeeze_mask = var_3732_squeeze_mask_0, x = k_complex_17_cast_fp16)[name = tensor("op_3732_cast_fp16")]; + tensor freqs_17_to_fp16 = const()[name = tensor("freqs_17_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(303657728)))]; + tensor var_3738_cast_fp16 = mul(x = freqs_17_to_fp16, y = ts_53_cast_fp16)[name = tensor("op_3738_cast_fp16")]; + tensor rotr_17_cast_fp16 = cos(x = var_3738_cast_fp16)[name = tensor("rotr_17_cast_fp16")]; + tensor roti_17_cast_fp16 = sin(x = var_3738_cast_fp16)[name = tensor("roti_17_cast_fp16")]; + tensor var_3742_cast_fp16 = mul(x = var_3708_cast_fp16, y = rotr_17_cast_fp16)[name = tensor("op_3742_cast_fp16")]; + tensor var_3743_cast_fp16 = mul(x = var_3716_cast_fp16, y = roti_17_cast_fp16)[name = tensor("op_3743_cast_fp16")]; + tensor qor_33_cast_fp16 = sub(x = var_3742_cast_fp16, y = var_3743_cast_fp16)[name = tensor("qor_33_cast_fp16")]; + tensor var_3746_cast_fp16 = mul(x = var_3708_cast_fp16, y = roti_17_cast_fp16)[name = tensor("op_3746_cast_fp16")]; + tensor var_3747_cast_fp16 = mul(x = var_3716_cast_fp16, y = rotr_17_cast_fp16)[name = tensor("op_3747_cast_fp16")]; + tensor qoi_33_cast_fp16 = add(x = var_3746_cast_fp16, y = var_3747_cast_fp16)[name = tensor("qoi_33_cast_fp16")]; + tensor var_3750_cast_fp16 = mul(x = var_3724_cast_fp16, y = rotr_17_cast_fp16)[name = tensor("op_3750_cast_fp16")]; + tensor var_3751_cast_fp16 = mul(x = var_3732_cast_fp16, y = roti_17_cast_fp16)[name = tensor("op_3751_cast_fp16")]; + tensor kor_33_cast_fp16 = sub(x = var_3750_cast_fp16, y = var_3751_cast_fp16)[name = tensor("kor_33_cast_fp16")]; + tensor var_3754_cast_fp16 = mul(x = var_3724_cast_fp16, y = roti_17_cast_fp16)[name = tensor("op_3754_cast_fp16")]; + tensor var_3755_cast_fp16 = mul(x = var_3732_cast_fp16, y = rotr_17_cast_fp16)[name = tensor("op_3755_cast_fp16")]; + tensor koi_33_cast_fp16 = add(x = var_3754_cast_fp16, y = var_3755_cast_fp16)[name = tensor("koi_33_cast_fp16")]; + tensor qo_17_axis_0 = const()[name = tensor("qo_17_axis_0"), val = tensor(-1)]; + tensor qo_17_cast_fp16 = stack(axis = qo_17_axis_0, values = (qor_33_cast_fp16, qoi_33_cast_fp16))[name = tensor("qo_17_cast_fp16")]; + tensor ko_17_axis_0 = const()[name = tensor("ko_17_axis_0"), val = tensor(-1)]; + tensor ko_17_cast_fp16 = stack(axis = ko_17_axis_0, values = (kor_33_cast_fp16, koi_33_cast_fp16))[name = tensor("ko_17_cast_fp16")]; + tensor var_3784 = const()[name = tensor("op_3784"), val = tensor([1, 1, 16, 64])]; + tensor q_51_cast_fp16 = reshape(shape = var_3784, x = qo_17_cast_fp16)[name = tensor("q_51_cast_fp16")]; + tensor var_3786 = const()[name = tensor("op_3786"), val = tensor([1, 1, 16, 64])]; + tensor k_35_cast_fp16 = reshape(shape = var_3786, x = ko_17_cast_fp16)[name = tensor("k_35_cast_fp16")]; + tensor _inversed_3808_y_0_to_fp16 = const()[name = tensor("_inversed_3808_y_0_to_fp16"), val = tensor(0x1p-9)]; + tensor _inversed_3808_cast_fp16 = mul(x = ts_53_cast_fp16, y = _inversed_3808_y_0_to_fp16)[name = tensor("_inversed_3808_cast_fp16")]; + tensor var_3809_cast_fp16 = floor(x = _inversed_3808_cast_fp16)[name = tensor("op_3809_cast_fp16")]; + tensor var_3810_to_fp16 = const()[name = tensor("op_3810_to_fp16"), val = tensor(0x1p+9)]; + tensor var_3811_cast_fp16 = mul(x = var_3809_cast_fp16, y = var_3810_to_fp16)[name = tensor("op_3811_cast_fp16")]; + tensor write_indices_float_35_cast_fp16 = sub(x = ts_53_cast_fp16, y = var_3811_cast_fp16)[name = tensor("write_indices_float_35_cast_fp16")]; + tensor var_3818_dtype_0 = const()[name = tensor("op_3818_dtype_0"), val = tensor("int32")]; + tensor write_indices_17_reps_0 = const()[name = tensor("write_indices_17_reps_0"), val = tensor([1, 1, 16, 64])]; + tensor write_indices_float_35_cast_fp16_to_int32 = cast(dtype = var_3818_dtype_0, x = write_indices_float_35_cast_fp16)[name = tensor("cast_512")]; + tensor write_indices_17 = tile(reps = write_indices_17_reps_0, x = write_indices_float_35_cast_fp16_to_int32)[name = tensor("write_indices_17")]; + tensor var_3826_begin_0 = const()[name = tensor("op_3826_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_3826_end_0 = const()[name = tensor("op_3826_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor var_3826_end_mask_0 = const()[name = tensor("op_3826_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_3826_squeeze_mask_0 = const()[name = tensor("op_3826_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor cache8_to_fp16_dtype_0 = const()[name = tensor("cache8_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor cache8_to_fp16 = cast(dtype = cache8_to_fp16_dtype_0, x = cache8)[name = tensor("cast_511")]; + tensor var_3826_cast_fp16 = slice_by_index(begin = var_3826_begin_0, end = var_3826_end_0, end_mask = var_3826_end_mask_0, squeeze_mask = var_3826_squeeze_mask_0, x = cache8_to_fp16)[name = tensor("op_3826_cast_fp16")]; + tensor var_3828_axis_0 = const()[name = tensor("op_3828_axis_0"), val = tensor(1)]; + tensor var_3828_mode_0 = const()[name = tensor("op_3828_mode_0"), val = tensor("update")]; + tensor var_3828_validate_indices_0 = const()[name = tensor("op_3828_validate_indices_0"), val = tensor(false)]; + tensor var_3828_cast_fp16 = scatter_along_axis(axis = var_3828_axis_0, data = var_3826_cast_fp16, indices = write_indices_17, mode = var_3828_mode_0, updates = k_35_cast_fp16, validate_indices = var_3828_validate_indices_0)[name = tensor("op_3828_cast_fp16")]; + tensor concat_58 = const()[name = tensor("concat_58"), val = tensor([0, 0, 0, 0, 0])]; + tensor concat_59 = const()[name = tensor("concat_59"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_17_internal_tensor_assign_1_stride_0 = const()[name = tensor("new_cache_17_internal_tensor_assign_1_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_17_internal_tensor_assign_1_begin_mask_0 = const()[name = tensor("new_cache_17_internal_tensor_assign_1_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_17_internal_tensor_assign_1_end_mask_0 = const()[name = tensor("new_cache_17_internal_tensor_assign_1_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_17_internal_tensor_assign_1_squeeze_mask_0 = const()[name = tensor("new_cache_17_internal_tensor_assign_1_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_64 = const()[name = tensor("shape_64"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_16 = const()[name = tensor("reduce_prod_16"), val = tensor(1048576)]; + tensor range_1d_16_start_0 = const()[name = tensor("range_1d_16_start_0"), val = tensor(0)]; + tensor range_1d_16_step_0 = const()[name = tensor("range_1d_16_step_0"), val = tensor(1)]; + tensor range_1d_16 = range_1d(end = reduce_prod_16, start = range_1d_16_start_0, step = range_1d_16_step_0)[name = tensor("range_1d_16")]; + tensor reshape_80 = reshape(shape = shape_64, x = range_1d_16)[name = tensor("reshape_80")]; + tensor slice_by_index_16 = slice_by_index(begin = concat_58, begin_mask = new_cache_17_internal_tensor_assign_1_begin_mask_0, end = concat_59, end_mask = new_cache_17_internal_tensor_assign_1_end_mask_0, squeeze_mask = new_cache_17_internal_tensor_assign_1_squeeze_mask_0, stride = new_cache_17_internal_tensor_assign_1_stride_0, x = reshape_80)[name = tensor("slice_by_index_16")]; + tensor reshape_81_shape_0 = const()[name = tensor("reshape_81_shape_0"), val = tensor([-1])]; + tensor reshape_81 = reshape(shape = reshape_81_shape_0, x = slice_by_index_16)[name = tensor("reshape_81")]; + tensor reshape_82_shape_0 = const()[name = tensor("reshape_82_shape_0"), val = tensor([-1])]; + tensor reshape_82_cast_fp16 = reshape(shape = reshape_82_shape_0, x = var_3828_cast_fp16)[name = tensor("reshape_82_cast_fp16")]; + tensor reshape_83_shape_0 = const()[name = tensor("reshape_83_shape_0"), val = tensor([-1])]; + tensor reshape_83_cast_fp16 = reshape(shape = reshape_83_shape_0, x = cache8_to_fp16)[name = tensor("reshape_83_cast_fp16")]; + tensor scatter_16_mode_0 = const()[name = tensor("scatter_16_mode_0"), val = tensor("update")]; + tensor scatter_16_axis_0 = const()[name = tensor("scatter_16_axis_0"), val = tensor(0)]; + tensor scatter_16_validate_indices_0 = const()[name = tensor("scatter_16_validate_indices_0"), val = tensor(false)]; + tensor scatter_16_cast_fp16 = scatter(axis = scatter_16_axis_0, data = reshape_83_cast_fp16, indices = reshape_81, mode = scatter_16_mode_0, updates = reshape_82_cast_fp16, validate_indices = scatter_16_validate_indices_0)[name = tensor("scatter_16_cast_fp16")]; + tensor reshape_84_cast_fp16 = reshape(shape = shape_64, x = scatter_16_cast_fp16)[name = tensor("reshape_84_cast_fp16")]; + tensor var_3836_begin_0 = const()[name = tensor("op_3836_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor var_3836_end_0 = const()[name = tensor("op_3836_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor var_3836_end_mask_0 = const()[name = tensor("op_3836_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_3836_squeeze_mask_0 = const()[name = tensor("op_3836_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor var_3836_cast_fp16 = slice_by_index(begin = var_3836_begin_0, end = var_3836_end_0, end_mask = var_3836_end_mask_0, squeeze_mask = var_3836_squeeze_mask_0, x = reshape_84_cast_fp16)[name = tensor("op_3836_cast_fp16")]; + tensor var_3838_axis_0 = const()[name = tensor("op_3838_axis_0"), val = tensor(1)]; + tensor var_3838_mode_0 = const()[name = tensor("op_3838_mode_0"), val = tensor("update")]; + tensor var_3838_validate_indices_0 = const()[name = tensor("op_3838_validate_indices_0"), val = tensor(false)]; + tensor var_3838_cast_fp16 = scatter_along_axis(axis = var_3838_axis_0, data = var_3836_cast_fp16, indices = write_indices_17, mode = var_3838_mode_0, updates = v_17_cast_fp16, validate_indices = var_3838_validate_indices_0)[name = tensor("op_3838_cast_fp16")]; + tensor concat_60 = const()[name = tensor("concat_60"), val = tensor([1, 0, 0, 0, 0])]; + tensor concat_61 = const()[name = tensor("concat_61"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_17_internal_tensor_assign_2_stride_0 = const()[name = tensor("new_cache_17_internal_tensor_assign_2_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_17_internal_tensor_assign_2_begin_mask_0 = const()[name = tensor("new_cache_17_internal_tensor_assign_2_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_17_internal_tensor_assign_2_end_mask_0 = const()[name = tensor("new_cache_17_internal_tensor_assign_2_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_17_internal_tensor_assign_2_squeeze_mask_0 = const()[name = tensor("new_cache_17_internal_tensor_assign_2_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_65 = const()[name = tensor("shape_65"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_17 = const()[name = tensor("reduce_prod_17"), val = tensor(1048576)]; + tensor range_1d_17_start_0 = const()[name = tensor("range_1d_17_start_0"), val = tensor(0)]; + tensor range_1d_17_step_0 = const()[name = tensor("range_1d_17_step_0"), val = tensor(1)]; + tensor range_1d_17 = range_1d(end = reduce_prod_17, start = range_1d_17_start_0, step = range_1d_17_step_0)[name = tensor("range_1d_17")]; + tensor reshape_85 = reshape(shape = shape_65, x = range_1d_17)[name = tensor("reshape_85")]; + tensor slice_by_index_17 = slice_by_index(begin = concat_60, begin_mask = new_cache_17_internal_tensor_assign_2_begin_mask_0, end = concat_61, end_mask = new_cache_17_internal_tensor_assign_2_end_mask_0, squeeze_mask = new_cache_17_internal_tensor_assign_2_squeeze_mask_0, stride = new_cache_17_internal_tensor_assign_2_stride_0, x = reshape_85)[name = tensor("slice_by_index_17")]; + tensor reshape_86_shape_0 = const()[name = tensor("reshape_86_shape_0"), val = tensor([-1])]; + tensor reshape_86 = reshape(shape = reshape_86_shape_0, x = slice_by_index_17)[name = tensor("reshape_86")]; + tensor reshape_87_shape_0 = const()[name = tensor("reshape_87_shape_0"), val = tensor([-1])]; + tensor reshape_87_cast_fp16 = reshape(shape = reshape_87_shape_0, x = var_3838_cast_fp16)[name = tensor("reshape_87_cast_fp16")]; + tensor reshape_88_shape_0 = const()[name = tensor("reshape_88_shape_0"), val = tensor([-1])]; + tensor reshape_88_cast_fp16 = reshape(shape = reshape_88_shape_0, x = reshape_84_cast_fp16)[name = tensor("reshape_88_cast_fp16")]; + tensor scatter_17_mode_0 = const()[name = tensor("scatter_17_mode_0"), val = tensor("update")]; + tensor scatter_17_axis_0 = const()[name = tensor("scatter_17_axis_0"), val = tensor(0)]; + tensor scatter_17_validate_indices_0 = const()[name = tensor("scatter_17_validate_indices_0"), val = tensor(false)]; + tensor scatter_17_cast_fp16 = scatter(axis = scatter_17_axis_0, data = reshape_88_cast_fp16, indices = reshape_86, mode = scatter_17_mode_0, updates = reshape_87_cast_fp16, validate_indices = scatter_17_validate_indices_0)[name = tensor("scatter_17_cast_fp16")]; + tensor reshape_89_cast_fp16 = reshape(shape = shape_65, x = scatter_17_cast_fp16)[name = tensor("reshape_89_cast_fp16")]; + tensor reshape_89_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("reshape_89_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor keys_49_begin_0 = const()[name = tensor("keys_49_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor keys_49_end_0 = const()[name = tensor("keys_49_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor keys_49_end_mask_0 = const()[name = tensor("keys_49_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor keys_49_squeeze_mask_0 = const()[name = tensor("keys_49_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor keys_49_cast_fp16 = slice_by_index(begin = keys_49_begin_0, end = keys_49_end_0, end_mask = keys_49_end_mask_0, squeeze_mask = keys_49_squeeze_mask_0, x = reshape_89_cast_fp16)[name = tensor("keys_49_cast_fp16")]; + tensor values_49_begin_0 = const()[name = tensor("values_49_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor values_49_end_0 = const()[name = tensor("values_49_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor values_49_end_mask_0 = const()[name = tensor("values_49_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor values_49_squeeze_mask_0 = const()[name = tensor("values_49_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor values_49_cast_fp16 = slice_by_index(begin = values_49_begin_0, end = values_49_end_0, end_mask = values_49_end_mask_0, squeeze_mask = values_49_squeeze_mask_0, x = reshape_89_cast_fp16)[name = tensor("values_49_cast_fp16")]; + tensor var_3850_cast_fp16 = not_equal(x = keys_49_cast_fp16, y = keys_49_cast_fp16)[name = tensor("op_3850_cast_fp16")]; + tensor keys_51_cast_fp16 = select(a = var_504_to_fp16, b = keys_49_cast_fp16, cond = var_3850_cast_fp16)[name = tensor("keys_51_cast_fp16")]; + tensor var_3858_cast_fp16 = not_equal(x = values_49_cast_fp16, y = values_49_cast_fp16)[name = tensor("op_3858_cast_fp16")]; + tensor values_51_cast_fp16 = select(a = var_504_to_fp16, b = values_49_cast_fp16, cond = var_3858_cast_fp16)[name = tensor("values_51_cast_fp16")]; + tensor var_3882 = const()[name = tensor("op_3882"), val = tensor([0, 2, 1, 3])]; + tensor var_3895 = const()[name = tensor("op_3895"), val = tensor([1, 1, 1])]; + tensor var_3896_cast_fp16 = reshape(shape = var_3895, x = position8_to_fp16)[name = tensor("op_3896_cast_fp16")]; + tensor var_3913_to_fp16 = const()[name = tensor("op_3913_to_fp16"), val = tensor(0x1p+0)]; + tensor valid_len_17_cast_fp16 = add(x = var_3896_cast_fp16, y = var_3913_to_fp16)[name = tensor("valid_len_17_cast_fp16")]; + tensor valid_mask_17_cast_fp16 = less(x = k_positions_1_promoted_to_fp16, y = valid_len_17_cast_fp16)[name = tensor("valid_mask_17_cast_fp16")]; + tensor causal_mask_17_cast_fp16 = less_equal(x = k_positions_1_promoted_to_fp16, y = var_3896_cast_fp16)[name = tensor("causal_mask_17_cast_fp16")]; + tensor attn_mask_33 = logical_and(x = valid_mask_17_cast_fp16, y = causal_mask_17_cast_fp16)[name = tensor("attn_mask_33")]; + tensor attn_mask_35_axes_0 = const()[name = tensor("attn_mask_35_axes_0"), val = tensor([1])]; + tensor attn_mask_35 = expand_dims(axes = attn_mask_35_axes_0, x = attn_mask_33)[name = tensor("attn_mask_35")]; + tensor var_3931_transpose_x_0 = const()[name = tensor("op_3931_transpose_x_0"), val = tensor(false)]; + tensor var_3931_transpose_y_0 = const()[name = tensor("op_3931_transpose_y_0"), val = tensor(false)]; + tensor transpose_88_perm_0 = const()[name = tensor("transpose_88_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_89_perm_0 = const()[name = tensor("transpose_89_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_89 = transpose(perm = transpose_89_perm_0, x = keys_51_cast_fp16)[name = tensor("transpose_181")]; + tensor transpose_88 = transpose(perm = transpose_88_perm_0, x = q_51_cast_fp16)[name = tensor("transpose_182")]; + tensor var_3931_cast_fp16 = matmul(transpose_x = var_3931_transpose_x_0, transpose_y = var_3931_transpose_y_0, x = transpose_88, y = transpose_89)[name = tensor("op_3931_cast_fp16")]; + tensor var_3925_to_fp16 = const()[name = tensor("op_3925_to_fp16"), val = tensor([0x1p-3])]; + tensor attn_weights_49_cast_fp16 = mul(x = var_3931_cast_fp16, y = var_3925_to_fp16)[name = tensor("attn_weights_49_cast_fp16")]; + tensor var_3933 = logical_not(x = attn_mask_35)[name = tensor("op_3933")]; + tensor var_3934_to_fp16 = const()[name = tensor("op_3934_to_fp16"), val = tensor(-inf)]; + tensor attn_weights_51_cast_fp16 = select(a = var_3934_to_fp16, b = attn_weights_49_cast_fp16, cond = var_3933)[name = tensor("attn_weights_51_cast_fp16")]; + tensor var_3936 = const()[name = tensor("op_3936"), val = tensor(-1)]; + tensor attn_weights_53_cast_fp16 = softmax(axis = var_3936, x = attn_weights_51_cast_fp16)[name = tensor("attn_weights_53_cast_fp16")]; + tensor attn_output_17_transpose_x_0 = const()[name = tensor("attn_output_17_transpose_x_0"), val = tensor(false)]; + tensor attn_output_17_transpose_y_0 = const()[name = tensor("attn_output_17_transpose_y_0"), val = tensor(false)]; + tensor values_53_cast_fp16 = transpose(perm = var_3882, x = values_51_cast_fp16)[name = tensor("transpose_183")]; + tensor attn_output_17_cast_fp16 = matmul(transpose_x = attn_output_17_transpose_x_0, transpose_y = attn_output_17_transpose_y_0, x = attn_weights_53_cast_fp16, y = values_53_cast_fp16)[name = tensor("attn_output_17_cast_fp16")]; + tensor var_3944 = const()[name = tensor("op_3944"), val = tensor([0, 2, 1, 3])]; + tensor var_3947 = const()[name = tensor("op_3947"), val = tensor([1, 1, 1024])]; + tensor var_3945_cast_fp16 = transpose(perm = var_3944, x = attn_output_17_cast_fp16)[name = tensor("transpose_180")]; + tensor input_85_cast_fp16 = reshape(shape = var_3947, x = var_3945_cast_fp16)[name = tensor("input_85_cast_fp16")]; + tensor linear_34_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = attn8_out_proj_weight_cast_fp16, x = input_85_cast_fp16)[name = tensor("linear_34_cast_fp16")]; + tensor var_3953_to_fp16 = const()[name = tensor("op_3953_to_fp16"), val = tensor(0x1p+0)]; + tensor var_3954_cast_fp16 = add(x = position8_to_fp16, y = var_3953_to_fp16)[name = tensor("op_3954_cast_fp16")]; + tensor var_3954_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_3954_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor input_87_cast_fp16 = add(x = input_83_cast_fp16, y = linear_34_cast_fp16)[name = tensor("input_87_cast_fp16")]; + tensor input_89_axes_0 = const()[name = tensor("input_89_axes_0"), val = tensor([-1])]; + tensor norm8_2_weight_to_fp16 = const()[name = tensor("norm8_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(303657856)))]; + tensor norm8_2_bias_to_fp16 = const()[name = tensor("norm8_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(303659968)))]; + tensor var_3958_to_fp16 = const()[name = tensor("op_3958_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_89_cast_fp16 = layer_norm(axes = input_89_axes_0, beta = norm8_2_bias_to_fp16, epsilon = var_3958_to_fp16, gamma = norm8_2_weight_to_fp16, x = input_87_cast_fp16)[name = tensor("input_89_cast_fp16")]; + tensor linear_35_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = linear8_1_weight_cast_fp16, x = input_89_cast_fp16)[name = tensor("linear_35_cast_fp16")]; + tensor input_91_mode_0 = const()[name = tensor("input_91_mode_0"), val = tensor("EXACT")]; + tensor input_91_cast_fp16 = gelu(mode = input_91_mode_0, x = linear_35_cast_fp16)[name = tensor("input_91_cast_fp16")]; + tensor linear_36_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = linear8_2_weight_cast_fp16, x = input_91_cast_fp16)[name = tensor("linear_36_cast_fp16")]; + tensor input_93_cast_fp16 = add(x = input_87_cast_fp16, y = linear_36_cast_fp16)[name = tensor("input_93_cast_fp16")]; + tensor x_19_axes_0 = const()[name = tensor("x_19_axes_0"), val = tensor([-1])]; + tensor norm9_1_weight_to_fp16 = const()[name = tensor("norm9_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(303662080)))]; + tensor norm9_1_bias_to_fp16 = const()[name = tensor("norm9_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(303664192)))]; + tensor var_3975_to_fp16 = const()[name = tensor("op_3975_to_fp16"), val = tensor(0x1.5p-17)]; + tensor x_19_cast_fp16 = layer_norm(axes = x_19_axes_0, beta = norm9_1_bias_to_fp16, epsilon = var_3975_to_fp16, gamma = norm9_1_weight_to_fp16, x = input_93_cast_fp16)[name = tensor("x_19_cast_fp16")]; + tensor linear_37_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = attn9_in_proj_weight_cast_fp16, x = x_19_cast_fp16)[name = tensor("linear_37_cast_fp16")]; + tensor var_4011 = const()[name = tensor("op_4011"), val = tensor([1, 1, 3, 16, 64])]; + tensor qkv_19_cast_fp16 = reshape(shape = var_4011, x = linear_37_cast_fp16)[name = tensor("qkv_19_cast_fp16")]; + tensor q_55_begin_0 = const()[name = tensor("q_55_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor q_55_end_0 = const()[name = tensor("q_55_end_0"), val = tensor([1, 1, 1, 16, 64])]; + tensor q_55_end_mask_0 = const()[name = tensor("q_55_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor q_55_squeeze_mask_0 = const()[name = tensor("q_55_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor q_55_cast_fp16 = slice_by_index(begin = q_55_begin_0, end = q_55_end_0, end_mask = q_55_end_mask_0, squeeze_mask = q_55_squeeze_mask_0, x = qkv_19_cast_fp16)[name = tensor("q_55_cast_fp16")]; + tensor k_37_begin_0 = const()[name = tensor("k_37_begin_0"), val = tensor([0, 0, 1, 0, 0])]; + tensor k_37_end_0 = const()[name = tensor("k_37_end_0"), val = tensor([1, 1, 2, 16, 64])]; + tensor k_37_end_mask_0 = const()[name = tensor("k_37_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor k_37_squeeze_mask_0 = const()[name = tensor("k_37_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor k_37_cast_fp16 = slice_by_index(begin = k_37_begin_0, end = k_37_end_0, end_mask = k_37_end_mask_0, squeeze_mask = k_37_squeeze_mask_0, x = qkv_19_cast_fp16)[name = tensor("k_37_cast_fp16")]; + tensor v_19_begin_0 = const()[name = tensor("v_19_begin_0"), val = tensor([0, 0, 2, 0, 0])]; + tensor v_19_end_0 = const()[name = tensor("v_19_end_0"), val = tensor([1, 1, 3, 16, 64])]; + tensor v_19_end_mask_0 = const()[name = tensor("v_19_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor v_19_squeeze_mask_0 = const()[name = tensor("v_19_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor v_19_cast_fp16 = slice_by_index(begin = v_19_begin_0, end = v_19_end_0, end_mask = v_19_end_mask_0, squeeze_mask = v_19_squeeze_mask_0, x = qkv_19_cast_fp16)[name = tensor("v_19_cast_fp16")]; + tensor var_4115 = const()[name = tensor("op_4115"), val = tensor([1, 1, 1, 1])]; + tensor position9_to_fp16_dtype_0 = const()[name = tensor("position9_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor position9_to_fp16 = cast(dtype = position9_to_fp16_dtype_0, x = position9)[name = tensor("cast_508")]; + tensor ts_59_cast_fp16 = reshape(shape = var_4115, x = position9_to_fp16)[name = tensor("ts_59_cast_fp16")]; + tensor var_4119 = const()[name = tensor("op_4119"), val = tensor([1, 1, 16, 32, 2])]; + tensor q_complex_19_cast_fp16 = reshape(shape = var_4119, x = q_55_cast_fp16)[name = tensor("q_complex_19_cast_fp16")]; + tensor var_4123 = const()[name = tensor("op_4123"), val = tensor([1, 1, 16, 32, 2])]; + tensor k_complex_19_cast_fp16 = reshape(shape = var_4123, x = k_37_cast_fp16)[name = tensor("k_complex_19_cast_fp16")]; + tensor var_4127_begin_0 = const()[name = tensor("op_4127_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_4127_end_0 = const()[name = tensor("op_4127_end_0"), val = tensor([1, 1, 16, 32, 1])]; + tensor var_4127_end_mask_0 = const()[name = tensor("op_4127_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_4127_squeeze_mask_0 = const()[name = tensor("op_4127_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_4127_cast_fp16 = slice_by_index(begin = var_4127_begin_0, end = var_4127_end_0, end_mask = var_4127_end_mask_0, squeeze_mask = var_4127_squeeze_mask_0, x = q_complex_19_cast_fp16)[name = tensor("op_4127_cast_fp16")]; + tensor var_4135_begin_0 = const()[name = tensor("op_4135_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_4135_end_0 = const()[name = tensor("op_4135_end_0"), val = tensor([1, 1, 16, 32, 2])]; + tensor var_4135_end_mask_0 = const()[name = tensor("op_4135_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_4135_squeeze_mask_0 = const()[name = tensor("op_4135_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_4135_cast_fp16 = slice_by_index(begin = var_4135_begin_0, end = var_4135_end_0, end_mask = var_4135_end_mask_0, squeeze_mask = var_4135_squeeze_mask_0, x = q_complex_19_cast_fp16)[name = tensor("op_4135_cast_fp16")]; + tensor var_4143_begin_0 = const()[name = tensor("op_4143_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_4143_end_0 = const()[name = tensor("op_4143_end_0"), val = tensor([1, 1, 16, 32, 1])]; + tensor var_4143_end_mask_0 = const()[name = tensor("op_4143_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_4143_squeeze_mask_0 = const()[name = tensor("op_4143_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_4143_cast_fp16 = slice_by_index(begin = var_4143_begin_0, end = var_4143_end_0, end_mask = var_4143_end_mask_0, squeeze_mask = var_4143_squeeze_mask_0, x = k_complex_19_cast_fp16)[name = tensor("op_4143_cast_fp16")]; + tensor var_4151_begin_0 = const()[name = tensor("op_4151_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_4151_end_0 = const()[name = tensor("op_4151_end_0"), val = tensor([1, 1, 16, 32, 2])]; + tensor var_4151_end_mask_0 = const()[name = tensor("op_4151_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_4151_squeeze_mask_0 = const()[name = tensor("op_4151_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_4151_cast_fp16 = slice_by_index(begin = var_4151_begin_0, end = var_4151_end_0, end_mask = var_4151_end_mask_0, squeeze_mask = var_4151_squeeze_mask_0, x = k_complex_19_cast_fp16)[name = tensor("op_4151_cast_fp16")]; + tensor freqs_19_to_fp16 = const()[name = tensor("freqs_19_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(303666304)))]; + tensor var_4157_cast_fp16 = mul(x = freqs_19_to_fp16, y = ts_59_cast_fp16)[name = tensor("op_4157_cast_fp16")]; + tensor rotr_19_cast_fp16 = cos(x = var_4157_cast_fp16)[name = tensor("rotr_19_cast_fp16")]; + tensor roti_19_cast_fp16 = sin(x = var_4157_cast_fp16)[name = tensor("roti_19_cast_fp16")]; + tensor var_4161_cast_fp16 = mul(x = var_4127_cast_fp16, y = rotr_19_cast_fp16)[name = tensor("op_4161_cast_fp16")]; + tensor var_4162_cast_fp16 = mul(x = var_4135_cast_fp16, y = roti_19_cast_fp16)[name = tensor("op_4162_cast_fp16")]; + tensor qor_37_cast_fp16 = sub(x = var_4161_cast_fp16, y = var_4162_cast_fp16)[name = tensor("qor_37_cast_fp16")]; + tensor var_4165_cast_fp16 = mul(x = var_4127_cast_fp16, y = roti_19_cast_fp16)[name = tensor("op_4165_cast_fp16")]; + tensor var_4166_cast_fp16 = mul(x = var_4135_cast_fp16, y = rotr_19_cast_fp16)[name = tensor("op_4166_cast_fp16")]; + tensor qoi_37_cast_fp16 = add(x = var_4165_cast_fp16, y = var_4166_cast_fp16)[name = tensor("qoi_37_cast_fp16")]; + tensor var_4169_cast_fp16 = mul(x = var_4143_cast_fp16, y = rotr_19_cast_fp16)[name = tensor("op_4169_cast_fp16")]; + tensor var_4170_cast_fp16 = mul(x = var_4151_cast_fp16, y = roti_19_cast_fp16)[name = tensor("op_4170_cast_fp16")]; + tensor kor_37_cast_fp16 = sub(x = var_4169_cast_fp16, y = var_4170_cast_fp16)[name = tensor("kor_37_cast_fp16")]; + tensor var_4173_cast_fp16 = mul(x = var_4143_cast_fp16, y = roti_19_cast_fp16)[name = tensor("op_4173_cast_fp16")]; + tensor var_4174_cast_fp16 = mul(x = var_4151_cast_fp16, y = rotr_19_cast_fp16)[name = tensor("op_4174_cast_fp16")]; + tensor koi_37_cast_fp16 = add(x = var_4173_cast_fp16, y = var_4174_cast_fp16)[name = tensor("koi_37_cast_fp16")]; + tensor qo_19_axis_0 = const()[name = tensor("qo_19_axis_0"), val = tensor(-1)]; + tensor qo_19_cast_fp16 = stack(axis = qo_19_axis_0, values = (qor_37_cast_fp16, qoi_37_cast_fp16))[name = tensor("qo_19_cast_fp16")]; + tensor ko_19_axis_0 = const()[name = tensor("ko_19_axis_0"), val = tensor(-1)]; + tensor ko_19_cast_fp16 = stack(axis = ko_19_axis_0, values = (kor_37_cast_fp16, koi_37_cast_fp16))[name = tensor("ko_19_cast_fp16")]; + tensor var_4203 = const()[name = tensor("op_4203"), val = tensor([1, 1, 16, 64])]; + tensor q_57_cast_fp16 = reshape(shape = var_4203, x = qo_19_cast_fp16)[name = tensor("q_57_cast_fp16")]; + tensor var_4205 = const()[name = tensor("op_4205"), val = tensor([1, 1, 16, 64])]; + tensor k_39_cast_fp16 = reshape(shape = var_4205, x = ko_19_cast_fp16)[name = tensor("k_39_cast_fp16")]; + tensor _inversed_4227_y_0_to_fp16 = const()[name = tensor("_inversed_4227_y_0_to_fp16"), val = tensor(0x1p-9)]; + tensor _inversed_4227_cast_fp16 = mul(x = ts_59_cast_fp16, y = _inversed_4227_y_0_to_fp16)[name = tensor("_inversed_4227_cast_fp16")]; + tensor var_4228_cast_fp16 = floor(x = _inversed_4227_cast_fp16)[name = tensor("op_4228_cast_fp16")]; + tensor var_4229_to_fp16 = const()[name = tensor("op_4229_to_fp16"), val = tensor(0x1p+9)]; + tensor var_4230_cast_fp16 = mul(x = var_4228_cast_fp16, y = var_4229_to_fp16)[name = tensor("op_4230_cast_fp16")]; + tensor write_indices_float_39_cast_fp16 = sub(x = ts_59_cast_fp16, y = var_4230_cast_fp16)[name = tensor("write_indices_float_39_cast_fp16")]; + tensor var_4237_dtype_0 = const()[name = tensor("op_4237_dtype_0"), val = tensor("int32")]; + tensor write_indices_19_reps_0 = const()[name = tensor("write_indices_19_reps_0"), val = tensor([1, 1, 16, 64])]; + tensor write_indices_float_39_cast_fp16_to_int32 = cast(dtype = var_4237_dtype_0, x = write_indices_float_39_cast_fp16)[name = tensor("cast_507")]; + tensor write_indices_19 = tile(reps = write_indices_19_reps_0, x = write_indices_float_39_cast_fp16_to_int32)[name = tensor("write_indices_19")]; + tensor var_4245_begin_0 = const()[name = tensor("op_4245_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_4245_end_0 = const()[name = tensor("op_4245_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor var_4245_end_mask_0 = const()[name = tensor("op_4245_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_4245_squeeze_mask_0 = const()[name = tensor("op_4245_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor cache9_to_fp16_dtype_0 = const()[name = tensor("cache9_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor cache9_to_fp16 = cast(dtype = cache9_to_fp16_dtype_0, x = cache9)[name = tensor("cast_506")]; + tensor var_4245_cast_fp16 = slice_by_index(begin = var_4245_begin_0, end = var_4245_end_0, end_mask = var_4245_end_mask_0, squeeze_mask = var_4245_squeeze_mask_0, x = cache9_to_fp16)[name = tensor("op_4245_cast_fp16")]; + tensor var_4247_axis_0 = const()[name = tensor("op_4247_axis_0"), val = tensor(1)]; + tensor var_4247_mode_0 = const()[name = tensor("op_4247_mode_0"), val = tensor("update")]; + tensor var_4247_validate_indices_0 = const()[name = tensor("op_4247_validate_indices_0"), val = tensor(false)]; + tensor var_4247_cast_fp16 = scatter_along_axis(axis = var_4247_axis_0, data = var_4245_cast_fp16, indices = write_indices_19, mode = var_4247_mode_0, updates = k_39_cast_fp16, validate_indices = var_4247_validate_indices_0)[name = tensor("op_4247_cast_fp16")]; + tensor concat_65 = const()[name = tensor("concat_65"), val = tensor([0, 0, 0, 0, 0])]; + tensor concat_66 = const()[name = tensor("concat_66"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_19_internal_tensor_assign_1_stride_0 = const()[name = tensor("new_cache_19_internal_tensor_assign_1_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_19_internal_tensor_assign_1_begin_mask_0 = const()[name = tensor("new_cache_19_internal_tensor_assign_1_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_19_internal_tensor_assign_1_end_mask_0 = const()[name = tensor("new_cache_19_internal_tensor_assign_1_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_19_internal_tensor_assign_1_squeeze_mask_0 = const()[name = tensor("new_cache_19_internal_tensor_assign_1_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_66 = const()[name = tensor("shape_66"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_18 = const()[name = tensor("reduce_prod_18"), val = tensor(1048576)]; + tensor range_1d_18_start_0 = const()[name = tensor("range_1d_18_start_0"), val = tensor(0)]; + tensor range_1d_18_step_0 = const()[name = tensor("range_1d_18_step_0"), val = tensor(1)]; + tensor range_1d_18 = range_1d(end = reduce_prod_18, start = range_1d_18_start_0, step = range_1d_18_step_0)[name = tensor("range_1d_18")]; + tensor reshape_90 = reshape(shape = shape_66, x = range_1d_18)[name = tensor("reshape_90")]; + tensor slice_by_index_18 = slice_by_index(begin = concat_65, begin_mask = new_cache_19_internal_tensor_assign_1_begin_mask_0, end = concat_66, end_mask = new_cache_19_internal_tensor_assign_1_end_mask_0, squeeze_mask = new_cache_19_internal_tensor_assign_1_squeeze_mask_0, stride = new_cache_19_internal_tensor_assign_1_stride_0, x = reshape_90)[name = tensor("slice_by_index_18")]; + tensor reshape_91_shape_0 = const()[name = tensor("reshape_91_shape_0"), val = tensor([-1])]; + tensor reshape_91 = reshape(shape = reshape_91_shape_0, x = slice_by_index_18)[name = tensor("reshape_91")]; + tensor reshape_92_shape_0 = const()[name = tensor("reshape_92_shape_0"), val = tensor([-1])]; + tensor reshape_92_cast_fp16 = reshape(shape = reshape_92_shape_0, x = var_4247_cast_fp16)[name = tensor("reshape_92_cast_fp16")]; + tensor reshape_93_shape_0 = const()[name = tensor("reshape_93_shape_0"), val = tensor([-1])]; + tensor reshape_93_cast_fp16 = reshape(shape = reshape_93_shape_0, x = cache9_to_fp16)[name = tensor("reshape_93_cast_fp16")]; + tensor scatter_18_mode_0 = const()[name = tensor("scatter_18_mode_0"), val = tensor("update")]; + tensor scatter_18_axis_0 = const()[name = tensor("scatter_18_axis_0"), val = tensor(0)]; + tensor scatter_18_validate_indices_0 = const()[name = tensor("scatter_18_validate_indices_0"), val = tensor(false)]; + tensor scatter_18_cast_fp16 = scatter(axis = scatter_18_axis_0, data = reshape_93_cast_fp16, indices = reshape_91, mode = scatter_18_mode_0, updates = reshape_92_cast_fp16, validate_indices = scatter_18_validate_indices_0)[name = tensor("scatter_18_cast_fp16")]; + tensor reshape_94_cast_fp16 = reshape(shape = shape_66, x = scatter_18_cast_fp16)[name = tensor("reshape_94_cast_fp16")]; + tensor var_4255_begin_0 = const()[name = tensor("op_4255_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor var_4255_end_0 = const()[name = tensor("op_4255_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor var_4255_end_mask_0 = const()[name = tensor("op_4255_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_4255_squeeze_mask_0 = const()[name = tensor("op_4255_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor var_4255_cast_fp16 = slice_by_index(begin = var_4255_begin_0, end = var_4255_end_0, end_mask = var_4255_end_mask_0, squeeze_mask = var_4255_squeeze_mask_0, x = reshape_94_cast_fp16)[name = tensor("op_4255_cast_fp16")]; + tensor var_4257_axis_0 = const()[name = tensor("op_4257_axis_0"), val = tensor(1)]; + tensor var_4257_mode_0 = const()[name = tensor("op_4257_mode_0"), val = tensor("update")]; + tensor var_4257_validate_indices_0 = const()[name = tensor("op_4257_validate_indices_0"), val = tensor(false)]; + tensor var_4257_cast_fp16 = scatter_along_axis(axis = var_4257_axis_0, data = var_4255_cast_fp16, indices = write_indices_19, mode = var_4257_mode_0, updates = v_19_cast_fp16, validate_indices = var_4257_validate_indices_0)[name = tensor("op_4257_cast_fp16")]; + tensor concat_67 = const()[name = tensor("concat_67"), val = tensor([1, 0, 0, 0, 0])]; + tensor concat_68 = const()[name = tensor("concat_68"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_19_internal_tensor_assign_2_stride_0 = const()[name = tensor("new_cache_19_internal_tensor_assign_2_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_19_internal_tensor_assign_2_begin_mask_0 = const()[name = tensor("new_cache_19_internal_tensor_assign_2_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_19_internal_tensor_assign_2_end_mask_0 = const()[name = tensor("new_cache_19_internal_tensor_assign_2_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_19_internal_tensor_assign_2_squeeze_mask_0 = const()[name = tensor("new_cache_19_internal_tensor_assign_2_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_67 = const()[name = tensor("shape_67"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_19 = const()[name = tensor("reduce_prod_19"), val = tensor(1048576)]; + tensor range_1d_19_start_0 = const()[name = tensor("range_1d_19_start_0"), val = tensor(0)]; + tensor range_1d_19_step_0 = const()[name = tensor("range_1d_19_step_0"), val = tensor(1)]; + tensor range_1d_19 = range_1d(end = reduce_prod_19, start = range_1d_19_start_0, step = range_1d_19_step_0)[name = tensor("range_1d_19")]; + tensor reshape_95 = reshape(shape = shape_67, x = range_1d_19)[name = tensor("reshape_95")]; + tensor slice_by_index_19 = slice_by_index(begin = concat_67, begin_mask = new_cache_19_internal_tensor_assign_2_begin_mask_0, end = concat_68, end_mask = new_cache_19_internal_tensor_assign_2_end_mask_0, squeeze_mask = new_cache_19_internal_tensor_assign_2_squeeze_mask_0, stride = new_cache_19_internal_tensor_assign_2_stride_0, x = reshape_95)[name = tensor("slice_by_index_19")]; + tensor reshape_96_shape_0 = const()[name = tensor("reshape_96_shape_0"), val = tensor([-1])]; + tensor reshape_96 = reshape(shape = reshape_96_shape_0, x = slice_by_index_19)[name = tensor("reshape_96")]; + tensor reshape_97_shape_0 = const()[name = tensor("reshape_97_shape_0"), val = tensor([-1])]; + tensor reshape_97_cast_fp16 = reshape(shape = reshape_97_shape_0, x = var_4257_cast_fp16)[name = tensor("reshape_97_cast_fp16")]; + tensor reshape_98_shape_0 = const()[name = tensor("reshape_98_shape_0"), val = tensor([-1])]; + tensor reshape_98_cast_fp16 = reshape(shape = reshape_98_shape_0, x = reshape_94_cast_fp16)[name = tensor("reshape_98_cast_fp16")]; + tensor scatter_19_mode_0 = const()[name = tensor("scatter_19_mode_0"), val = tensor("update")]; + tensor scatter_19_axis_0 = const()[name = tensor("scatter_19_axis_0"), val = tensor(0)]; + tensor scatter_19_validate_indices_0 = const()[name = tensor("scatter_19_validate_indices_0"), val = tensor(false)]; + tensor scatter_19_cast_fp16 = scatter(axis = scatter_19_axis_0, data = reshape_98_cast_fp16, indices = reshape_96, mode = scatter_19_mode_0, updates = reshape_97_cast_fp16, validate_indices = scatter_19_validate_indices_0)[name = tensor("scatter_19_cast_fp16")]; + tensor reshape_99_cast_fp16 = reshape(shape = shape_67, x = scatter_19_cast_fp16)[name = tensor("reshape_99_cast_fp16")]; + tensor reshape_99_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("reshape_99_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor keys_55_begin_0 = const()[name = tensor("keys_55_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor keys_55_end_0 = const()[name = tensor("keys_55_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor keys_55_end_mask_0 = const()[name = tensor("keys_55_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor keys_55_squeeze_mask_0 = const()[name = tensor("keys_55_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor keys_55_cast_fp16 = slice_by_index(begin = keys_55_begin_0, end = keys_55_end_0, end_mask = keys_55_end_mask_0, squeeze_mask = keys_55_squeeze_mask_0, x = reshape_99_cast_fp16)[name = tensor("keys_55_cast_fp16")]; + tensor values_55_begin_0 = const()[name = tensor("values_55_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor values_55_end_0 = const()[name = tensor("values_55_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor values_55_end_mask_0 = const()[name = tensor("values_55_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor values_55_squeeze_mask_0 = const()[name = tensor("values_55_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor values_55_cast_fp16 = slice_by_index(begin = values_55_begin_0, end = values_55_end_0, end_mask = values_55_end_mask_0, squeeze_mask = values_55_squeeze_mask_0, x = reshape_99_cast_fp16)[name = tensor("values_55_cast_fp16")]; + tensor var_4269_cast_fp16 = not_equal(x = keys_55_cast_fp16, y = keys_55_cast_fp16)[name = tensor("op_4269_cast_fp16")]; + tensor keys_57_cast_fp16 = select(a = var_504_to_fp16, b = keys_55_cast_fp16, cond = var_4269_cast_fp16)[name = tensor("keys_57_cast_fp16")]; + tensor var_4277_cast_fp16 = not_equal(x = values_55_cast_fp16, y = values_55_cast_fp16)[name = tensor("op_4277_cast_fp16")]; + tensor values_57_cast_fp16 = select(a = var_504_to_fp16, b = values_55_cast_fp16, cond = var_4277_cast_fp16)[name = tensor("values_57_cast_fp16")]; + tensor var_4301 = const()[name = tensor("op_4301"), val = tensor([0, 2, 1, 3])]; + tensor var_4314 = const()[name = tensor("op_4314"), val = tensor([1, 1, 1])]; + tensor var_4315_cast_fp16 = reshape(shape = var_4314, x = position9_to_fp16)[name = tensor("op_4315_cast_fp16")]; + tensor var_4332_to_fp16 = const()[name = tensor("op_4332_to_fp16"), val = tensor(0x1p+0)]; + tensor valid_len_19_cast_fp16 = add(x = var_4315_cast_fp16, y = var_4332_to_fp16)[name = tensor("valid_len_19_cast_fp16")]; + tensor valid_mask_19_cast_fp16 = less(x = k_positions_1_promoted_to_fp16, y = valid_len_19_cast_fp16)[name = tensor("valid_mask_19_cast_fp16")]; + tensor causal_mask_19_cast_fp16 = less_equal(x = k_positions_1_promoted_to_fp16, y = var_4315_cast_fp16)[name = tensor("causal_mask_19_cast_fp16")]; + tensor attn_mask_37 = logical_and(x = valid_mask_19_cast_fp16, y = causal_mask_19_cast_fp16)[name = tensor("attn_mask_37")]; + tensor attn_mask_39_axes_0 = const()[name = tensor("attn_mask_39_axes_0"), val = tensor([1])]; + tensor attn_mask_39 = expand_dims(axes = attn_mask_39_axes_0, x = attn_mask_37)[name = tensor("attn_mask_39")]; + tensor var_4350_transpose_x_0 = const()[name = tensor("op_4350_transpose_x_0"), val = tensor(false)]; + tensor var_4350_transpose_y_0 = const()[name = tensor("op_4350_transpose_y_0"), val = tensor(false)]; + tensor transpose_90_perm_0 = const()[name = tensor("transpose_90_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_91_perm_0 = const()[name = tensor("transpose_91_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_91 = transpose(perm = transpose_91_perm_0, x = keys_57_cast_fp16)[name = tensor("transpose_177")]; + tensor transpose_90 = transpose(perm = transpose_90_perm_0, x = q_57_cast_fp16)[name = tensor("transpose_178")]; + tensor var_4350_cast_fp16 = matmul(transpose_x = var_4350_transpose_x_0, transpose_y = var_4350_transpose_y_0, x = transpose_90, y = transpose_91)[name = tensor("op_4350_cast_fp16")]; + tensor var_4344_to_fp16 = const()[name = tensor("op_4344_to_fp16"), val = tensor([0x1p-3])]; + tensor attn_weights_55_cast_fp16 = mul(x = var_4350_cast_fp16, y = var_4344_to_fp16)[name = tensor("attn_weights_55_cast_fp16")]; + tensor var_4352 = logical_not(x = attn_mask_39)[name = tensor("op_4352")]; + tensor var_4353_to_fp16 = const()[name = tensor("op_4353_to_fp16"), val = tensor(-inf)]; + tensor attn_weights_57_cast_fp16 = select(a = var_4353_to_fp16, b = attn_weights_55_cast_fp16, cond = var_4352)[name = tensor("attn_weights_57_cast_fp16")]; + tensor var_4355 = const()[name = tensor("op_4355"), val = tensor(-1)]; + tensor attn_weights_59_cast_fp16 = softmax(axis = var_4355, x = attn_weights_57_cast_fp16)[name = tensor("attn_weights_59_cast_fp16")]; + tensor attn_output_19_transpose_x_0 = const()[name = tensor("attn_output_19_transpose_x_0"), val = tensor(false)]; + tensor attn_output_19_transpose_y_0 = const()[name = tensor("attn_output_19_transpose_y_0"), val = tensor(false)]; + tensor values_59_cast_fp16 = transpose(perm = var_4301, x = values_57_cast_fp16)[name = tensor("transpose_179")]; + tensor attn_output_19_cast_fp16 = matmul(transpose_x = attn_output_19_transpose_x_0, transpose_y = attn_output_19_transpose_y_0, x = attn_weights_59_cast_fp16, y = values_59_cast_fp16)[name = tensor("attn_output_19_cast_fp16")]; + tensor var_4363 = const()[name = tensor("op_4363"), val = tensor([0, 2, 1, 3])]; + tensor var_4366 = const()[name = tensor("op_4366"), val = tensor([1, 1, 1024])]; + tensor var_4364_cast_fp16 = transpose(perm = var_4363, x = attn_output_19_cast_fp16)[name = tensor("transpose_176")]; + tensor input_95_cast_fp16 = reshape(shape = var_4366, x = var_4364_cast_fp16)[name = tensor("input_95_cast_fp16")]; + tensor linear_38_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = attn9_out_proj_weight_cast_fp16, x = input_95_cast_fp16)[name = tensor("linear_38_cast_fp16")]; + tensor var_4372_to_fp16 = const()[name = tensor("op_4372_to_fp16"), val = tensor(0x1p+0)]; + tensor var_4373_cast_fp16 = add(x = position9_to_fp16, y = var_4372_to_fp16)[name = tensor("op_4373_cast_fp16")]; + tensor var_4373_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_4373_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor input_97_cast_fp16 = add(x = input_93_cast_fp16, y = linear_38_cast_fp16)[name = tensor("input_97_cast_fp16")]; + tensor input_99_axes_0 = const()[name = tensor("input_99_axes_0"), val = tensor([-1])]; + tensor norm9_2_weight_to_fp16 = const()[name = tensor("norm9_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(303666432)))]; + tensor norm9_2_bias_to_fp16 = const()[name = tensor("norm9_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(303668544)))]; + tensor var_4377_to_fp16 = const()[name = tensor("op_4377_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_99_cast_fp16 = layer_norm(axes = input_99_axes_0, beta = norm9_2_bias_to_fp16, epsilon = var_4377_to_fp16, gamma = norm9_2_weight_to_fp16, x = input_97_cast_fp16)[name = tensor("input_99_cast_fp16")]; + tensor linear_39_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = linear9_1_weight_cast_fp16, x = input_99_cast_fp16)[name = tensor("linear_39_cast_fp16")]; + tensor input_101_mode_0 = const()[name = tensor("input_101_mode_0"), val = tensor("EXACT")]; + tensor input_101_cast_fp16 = gelu(mode = input_101_mode_0, x = linear_39_cast_fp16)[name = tensor("input_101_cast_fp16")]; + tensor linear_40_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = linear9_2_weight_cast_fp16, x = input_101_cast_fp16)[name = tensor("linear_40_cast_fp16")]; + tensor input_103_cast_fp16 = add(x = input_97_cast_fp16, y = linear_40_cast_fp16)[name = tensor("input_103_cast_fp16")]; + tensor x_21_axes_0 = const()[name = tensor("x_21_axes_0"), val = tensor([-1])]; + tensor norm10_1_weight_to_fp16 = const()[name = tensor("norm10_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(303670656)))]; + tensor norm10_1_bias_to_fp16 = const()[name = tensor("norm10_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(303672768)))]; + tensor var_4394_to_fp16 = const()[name = tensor("op_4394_to_fp16"), val = tensor(0x1.5p-17)]; + tensor x_21_cast_fp16 = layer_norm(axes = x_21_axes_0, beta = norm10_1_bias_to_fp16, epsilon = var_4394_to_fp16, gamma = norm10_1_weight_to_fp16, x = input_103_cast_fp16)[name = tensor("x_21_cast_fp16")]; + tensor linear_41_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = attn10_in_proj_weight_cast_fp16, x = x_21_cast_fp16)[name = tensor("linear_41_cast_fp16")]; + tensor var_4430 = const()[name = tensor("op_4430"), val = tensor([1, 1, 3, 16, 64])]; + tensor qkv_21_cast_fp16 = reshape(shape = var_4430, x = linear_41_cast_fp16)[name = tensor("qkv_21_cast_fp16")]; + tensor q_61_begin_0 = const()[name = tensor("q_61_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor q_61_end_0 = const()[name = tensor("q_61_end_0"), val = tensor([1, 1, 1, 16, 64])]; + tensor q_61_end_mask_0 = const()[name = tensor("q_61_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor q_61_squeeze_mask_0 = const()[name = tensor("q_61_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor q_61_cast_fp16 = slice_by_index(begin = q_61_begin_0, end = q_61_end_0, end_mask = q_61_end_mask_0, squeeze_mask = q_61_squeeze_mask_0, x = qkv_21_cast_fp16)[name = tensor("q_61_cast_fp16")]; + tensor k_41_begin_0 = const()[name = tensor("k_41_begin_0"), val = tensor([0, 0, 1, 0, 0])]; + tensor k_41_end_0 = const()[name = tensor("k_41_end_0"), val = tensor([1, 1, 2, 16, 64])]; + tensor k_41_end_mask_0 = const()[name = tensor("k_41_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor k_41_squeeze_mask_0 = const()[name = tensor("k_41_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor k_41_cast_fp16 = slice_by_index(begin = k_41_begin_0, end = k_41_end_0, end_mask = k_41_end_mask_0, squeeze_mask = k_41_squeeze_mask_0, x = qkv_21_cast_fp16)[name = tensor("k_41_cast_fp16")]; + tensor v_21_begin_0 = const()[name = tensor("v_21_begin_0"), val = tensor([0, 0, 2, 0, 0])]; + tensor v_21_end_0 = const()[name = tensor("v_21_end_0"), val = tensor([1, 1, 3, 16, 64])]; + tensor v_21_end_mask_0 = const()[name = tensor("v_21_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor v_21_squeeze_mask_0 = const()[name = tensor("v_21_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor v_21_cast_fp16 = slice_by_index(begin = v_21_begin_0, end = v_21_end_0, end_mask = v_21_end_mask_0, squeeze_mask = v_21_squeeze_mask_0, x = qkv_21_cast_fp16)[name = tensor("v_21_cast_fp16")]; + tensor var_4534 = const()[name = tensor("op_4534"), val = tensor([1, 1, 1, 1])]; + tensor position10_to_fp16_dtype_0 = const()[name = tensor("position10_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor position10_to_fp16 = cast(dtype = position10_to_fp16_dtype_0, x = position10)[name = tensor("cast_503")]; + tensor ts_65_cast_fp16 = reshape(shape = var_4534, x = position10_to_fp16)[name = tensor("ts_65_cast_fp16")]; + tensor var_4538 = const()[name = tensor("op_4538"), val = tensor([1, 1, 16, 32, 2])]; + tensor q_complex_21_cast_fp16 = reshape(shape = var_4538, x = q_61_cast_fp16)[name = tensor("q_complex_21_cast_fp16")]; + tensor var_4542 = const()[name = tensor("op_4542"), val = tensor([1, 1, 16, 32, 2])]; + tensor k_complex_21_cast_fp16 = reshape(shape = var_4542, x = k_41_cast_fp16)[name = tensor("k_complex_21_cast_fp16")]; + tensor var_4546_begin_0 = const()[name = tensor("op_4546_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_4546_end_0 = const()[name = tensor("op_4546_end_0"), val = tensor([1, 1, 16, 32, 1])]; + tensor var_4546_end_mask_0 = const()[name = tensor("op_4546_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_4546_squeeze_mask_0 = const()[name = tensor("op_4546_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_4546_cast_fp16 = slice_by_index(begin = var_4546_begin_0, end = var_4546_end_0, end_mask = var_4546_end_mask_0, squeeze_mask = var_4546_squeeze_mask_0, x = q_complex_21_cast_fp16)[name = tensor("op_4546_cast_fp16")]; + tensor var_4554_begin_0 = const()[name = tensor("op_4554_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_4554_end_0 = const()[name = tensor("op_4554_end_0"), val = tensor([1, 1, 16, 32, 2])]; + tensor var_4554_end_mask_0 = const()[name = tensor("op_4554_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_4554_squeeze_mask_0 = const()[name = tensor("op_4554_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_4554_cast_fp16 = slice_by_index(begin = var_4554_begin_0, end = var_4554_end_0, end_mask = var_4554_end_mask_0, squeeze_mask = var_4554_squeeze_mask_0, x = q_complex_21_cast_fp16)[name = tensor("op_4554_cast_fp16")]; + tensor var_4562_begin_0 = const()[name = tensor("op_4562_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_4562_end_0 = const()[name = tensor("op_4562_end_0"), val = tensor([1, 1, 16, 32, 1])]; + tensor var_4562_end_mask_0 = const()[name = tensor("op_4562_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_4562_squeeze_mask_0 = const()[name = tensor("op_4562_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_4562_cast_fp16 = slice_by_index(begin = var_4562_begin_0, end = var_4562_end_0, end_mask = var_4562_end_mask_0, squeeze_mask = var_4562_squeeze_mask_0, x = k_complex_21_cast_fp16)[name = tensor("op_4562_cast_fp16")]; + tensor var_4570_begin_0 = const()[name = tensor("op_4570_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_4570_end_0 = const()[name = tensor("op_4570_end_0"), val = tensor([1, 1, 16, 32, 2])]; + tensor var_4570_end_mask_0 = const()[name = tensor("op_4570_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_4570_squeeze_mask_0 = const()[name = tensor("op_4570_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_4570_cast_fp16 = slice_by_index(begin = var_4570_begin_0, end = var_4570_end_0, end_mask = var_4570_end_mask_0, squeeze_mask = var_4570_squeeze_mask_0, x = k_complex_21_cast_fp16)[name = tensor("op_4570_cast_fp16")]; + tensor freqs_21_to_fp16 = const()[name = tensor("freqs_21_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(303674880)))]; + tensor var_4576_cast_fp16 = mul(x = freqs_21_to_fp16, y = ts_65_cast_fp16)[name = tensor("op_4576_cast_fp16")]; + tensor rotr_21_cast_fp16 = cos(x = var_4576_cast_fp16)[name = tensor("rotr_21_cast_fp16")]; + tensor roti_21_cast_fp16 = sin(x = var_4576_cast_fp16)[name = tensor("roti_21_cast_fp16")]; + tensor var_4580_cast_fp16 = mul(x = var_4546_cast_fp16, y = rotr_21_cast_fp16)[name = tensor("op_4580_cast_fp16")]; + tensor var_4581_cast_fp16 = mul(x = var_4554_cast_fp16, y = roti_21_cast_fp16)[name = tensor("op_4581_cast_fp16")]; + tensor qor_41_cast_fp16 = sub(x = var_4580_cast_fp16, y = var_4581_cast_fp16)[name = tensor("qor_41_cast_fp16")]; + tensor var_4584_cast_fp16 = mul(x = var_4546_cast_fp16, y = roti_21_cast_fp16)[name = tensor("op_4584_cast_fp16")]; + tensor var_4585_cast_fp16 = mul(x = var_4554_cast_fp16, y = rotr_21_cast_fp16)[name = tensor("op_4585_cast_fp16")]; + tensor qoi_41_cast_fp16 = add(x = var_4584_cast_fp16, y = var_4585_cast_fp16)[name = tensor("qoi_41_cast_fp16")]; + tensor var_4588_cast_fp16 = mul(x = var_4562_cast_fp16, y = rotr_21_cast_fp16)[name = tensor("op_4588_cast_fp16")]; + tensor var_4589_cast_fp16 = mul(x = var_4570_cast_fp16, y = roti_21_cast_fp16)[name = tensor("op_4589_cast_fp16")]; + tensor kor_41_cast_fp16 = sub(x = var_4588_cast_fp16, y = var_4589_cast_fp16)[name = tensor("kor_41_cast_fp16")]; + tensor var_4592_cast_fp16 = mul(x = var_4562_cast_fp16, y = roti_21_cast_fp16)[name = tensor("op_4592_cast_fp16")]; + tensor var_4593_cast_fp16 = mul(x = var_4570_cast_fp16, y = rotr_21_cast_fp16)[name = tensor("op_4593_cast_fp16")]; + tensor koi_41_cast_fp16 = add(x = var_4592_cast_fp16, y = var_4593_cast_fp16)[name = tensor("koi_41_cast_fp16")]; + tensor qo_21_axis_0 = const()[name = tensor("qo_21_axis_0"), val = tensor(-1)]; + tensor qo_21_cast_fp16 = stack(axis = qo_21_axis_0, values = (qor_41_cast_fp16, qoi_41_cast_fp16))[name = tensor("qo_21_cast_fp16")]; + tensor ko_21_axis_0 = const()[name = tensor("ko_21_axis_0"), val = tensor(-1)]; + tensor ko_21_cast_fp16 = stack(axis = ko_21_axis_0, values = (kor_41_cast_fp16, koi_41_cast_fp16))[name = tensor("ko_21_cast_fp16")]; + tensor var_4622 = const()[name = tensor("op_4622"), val = tensor([1, 1, 16, 64])]; + tensor q_63_cast_fp16 = reshape(shape = var_4622, x = qo_21_cast_fp16)[name = tensor("q_63_cast_fp16")]; + tensor var_4624 = const()[name = tensor("op_4624"), val = tensor([1, 1, 16, 64])]; + tensor k_43_cast_fp16 = reshape(shape = var_4624, x = ko_21_cast_fp16)[name = tensor("k_43_cast_fp16")]; + tensor _inversed_4646_y_0_to_fp16 = const()[name = tensor("_inversed_4646_y_0_to_fp16"), val = tensor(0x1p-9)]; + tensor _inversed_4646_cast_fp16 = mul(x = ts_65_cast_fp16, y = _inversed_4646_y_0_to_fp16)[name = tensor("_inversed_4646_cast_fp16")]; + tensor var_4647_cast_fp16 = floor(x = _inversed_4646_cast_fp16)[name = tensor("op_4647_cast_fp16")]; + tensor var_4648_to_fp16 = const()[name = tensor("op_4648_to_fp16"), val = tensor(0x1p+9)]; + tensor var_4649_cast_fp16 = mul(x = var_4647_cast_fp16, y = var_4648_to_fp16)[name = tensor("op_4649_cast_fp16")]; + tensor write_indices_float_43_cast_fp16 = sub(x = ts_65_cast_fp16, y = var_4649_cast_fp16)[name = tensor("write_indices_float_43_cast_fp16")]; + tensor var_4656_dtype_0 = const()[name = tensor("op_4656_dtype_0"), val = tensor("int32")]; + tensor write_indices_21_reps_0 = const()[name = tensor("write_indices_21_reps_0"), val = tensor([1, 1, 16, 64])]; + tensor write_indices_float_43_cast_fp16_to_int32 = cast(dtype = var_4656_dtype_0, x = write_indices_float_43_cast_fp16)[name = tensor("cast_502")]; + tensor write_indices_21 = tile(reps = write_indices_21_reps_0, x = write_indices_float_43_cast_fp16_to_int32)[name = tensor("write_indices_21")]; + tensor var_4664_begin_0 = const()[name = tensor("op_4664_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_4664_end_0 = const()[name = tensor("op_4664_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor var_4664_end_mask_0 = const()[name = tensor("op_4664_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_4664_squeeze_mask_0 = const()[name = tensor("op_4664_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor cache10_to_fp16_dtype_0 = const()[name = tensor("cache10_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor cache10_to_fp16 = cast(dtype = cache10_to_fp16_dtype_0, x = cache10)[name = tensor("cast_501")]; + tensor var_4664_cast_fp16 = slice_by_index(begin = var_4664_begin_0, end = var_4664_end_0, end_mask = var_4664_end_mask_0, squeeze_mask = var_4664_squeeze_mask_0, x = cache10_to_fp16)[name = tensor("op_4664_cast_fp16")]; + tensor var_4666_axis_0 = const()[name = tensor("op_4666_axis_0"), val = tensor(1)]; + tensor var_4666_mode_0 = const()[name = tensor("op_4666_mode_0"), val = tensor("update")]; + tensor var_4666_validate_indices_0 = const()[name = tensor("op_4666_validate_indices_0"), val = tensor(false)]; + tensor var_4666_cast_fp16 = scatter_along_axis(axis = var_4666_axis_0, data = var_4664_cast_fp16, indices = write_indices_21, mode = var_4666_mode_0, updates = k_43_cast_fp16, validate_indices = var_4666_validate_indices_0)[name = tensor("op_4666_cast_fp16")]; + tensor concat_72 = const()[name = tensor("concat_72"), val = tensor([0, 0, 0, 0, 0])]; + tensor concat_73 = const()[name = tensor("concat_73"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_21_internal_tensor_assign_1_stride_0 = const()[name = tensor("new_cache_21_internal_tensor_assign_1_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_21_internal_tensor_assign_1_begin_mask_0 = const()[name = tensor("new_cache_21_internal_tensor_assign_1_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_21_internal_tensor_assign_1_end_mask_0 = const()[name = tensor("new_cache_21_internal_tensor_assign_1_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_21_internal_tensor_assign_1_squeeze_mask_0 = const()[name = tensor("new_cache_21_internal_tensor_assign_1_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_68 = const()[name = tensor("shape_68"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_20 = const()[name = tensor("reduce_prod_20"), val = tensor(1048576)]; + tensor range_1d_20_start_0 = const()[name = tensor("range_1d_20_start_0"), val = tensor(0)]; + tensor range_1d_20_step_0 = const()[name = tensor("range_1d_20_step_0"), val = tensor(1)]; + tensor range_1d_20 = range_1d(end = reduce_prod_20, start = range_1d_20_start_0, step = range_1d_20_step_0)[name = tensor("range_1d_20")]; + tensor reshape_100 = reshape(shape = shape_68, x = range_1d_20)[name = tensor("reshape_100")]; + tensor slice_by_index_20 = slice_by_index(begin = concat_72, begin_mask = new_cache_21_internal_tensor_assign_1_begin_mask_0, end = concat_73, end_mask = new_cache_21_internal_tensor_assign_1_end_mask_0, squeeze_mask = new_cache_21_internal_tensor_assign_1_squeeze_mask_0, stride = new_cache_21_internal_tensor_assign_1_stride_0, x = reshape_100)[name = tensor("slice_by_index_20")]; + tensor reshape_101_shape_0 = const()[name = tensor("reshape_101_shape_0"), val = tensor([-1])]; + tensor reshape_101 = reshape(shape = reshape_101_shape_0, x = slice_by_index_20)[name = tensor("reshape_101")]; + tensor reshape_102_shape_0 = const()[name = tensor("reshape_102_shape_0"), val = tensor([-1])]; + tensor reshape_102_cast_fp16 = reshape(shape = reshape_102_shape_0, x = var_4666_cast_fp16)[name = tensor("reshape_102_cast_fp16")]; + tensor reshape_103_shape_0 = const()[name = tensor("reshape_103_shape_0"), val = tensor([-1])]; + tensor reshape_103_cast_fp16 = reshape(shape = reshape_103_shape_0, x = cache10_to_fp16)[name = tensor("reshape_103_cast_fp16")]; + tensor scatter_20_mode_0 = const()[name = tensor("scatter_20_mode_0"), val = tensor("update")]; + tensor scatter_20_axis_0 = const()[name = tensor("scatter_20_axis_0"), val = tensor(0)]; + tensor scatter_20_validate_indices_0 = const()[name = tensor("scatter_20_validate_indices_0"), val = tensor(false)]; + tensor scatter_20_cast_fp16 = scatter(axis = scatter_20_axis_0, data = reshape_103_cast_fp16, indices = reshape_101, mode = scatter_20_mode_0, updates = reshape_102_cast_fp16, validate_indices = scatter_20_validate_indices_0)[name = tensor("scatter_20_cast_fp16")]; + tensor reshape_104_cast_fp16 = reshape(shape = shape_68, x = scatter_20_cast_fp16)[name = tensor("reshape_104_cast_fp16")]; + tensor var_4674_begin_0 = const()[name = tensor("op_4674_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor var_4674_end_0 = const()[name = tensor("op_4674_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor var_4674_end_mask_0 = const()[name = tensor("op_4674_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_4674_squeeze_mask_0 = const()[name = tensor("op_4674_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor var_4674_cast_fp16 = slice_by_index(begin = var_4674_begin_0, end = var_4674_end_0, end_mask = var_4674_end_mask_0, squeeze_mask = var_4674_squeeze_mask_0, x = reshape_104_cast_fp16)[name = tensor("op_4674_cast_fp16")]; + tensor var_4676_axis_0 = const()[name = tensor("op_4676_axis_0"), val = tensor(1)]; + tensor var_4676_mode_0 = const()[name = tensor("op_4676_mode_0"), val = tensor("update")]; + tensor var_4676_validate_indices_0 = const()[name = tensor("op_4676_validate_indices_0"), val = tensor(false)]; + tensor var_4676_cast_fp16 = scatter_along_axis(axis = var_4676_axis_0, data = var_4674_cast_fp16, indices = write_indices_21, mode = var_4676_mode_0, updates = v_21_cast_fp16, validate_indices = var_4676_validate_indices_0)[name = tensor("op_4676_cast_fp16")]; + tensor concat_74 = const()[name = tensor("concat_74"), val = tensor([1, 0, 0, 0, 0])]; + tensor concat_75 = const()[name = tensor("concat_75"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_21_internal_tensor_assign_2_stride_0 = const()[name = tensor("new_cache_21_internal_tensor_assign_2_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_21_internal_tensor_assign_2_begin_mask_0 = const()[name = tensor("new_cache_21_internal_tensor_assign_2_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_21_internal_tensor_assign_2_end_mask_0 = const()[name = tensor("new_cache_21_internal_tensor_assign_2_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_21_internal_tensor_assign_2_squeeze_mask_0 = const()[name = tensor("new_cache_21_internal_tensor_assign_2_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_69 = const()[name = tensor("shape_69"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_21 = const()[name = tensor("reduce_prod_21"), val = tensor(1048576)]; + tensor range_1d_21_start_0 = const()[name = tensor("range_1d_21_start_0"), val = tensor(0)]; + tensor range_1d_21_step_0 = const()[name = tensor("range_1d_21_step_0"), val = tensor(1)]; + tensor range_1d_21 = range_1d(end = reduce_prod_21, start = range_1d_21_start_0, step = range_1d_21_step_0)[name = tensor("range_1d_21")]; + tensor reshape_105 = reshape(shape = shape_69, x = range_1d_21)[name = tensor("reshape_105")]; + tensor slice_by_index_21 = slice_by_index(begin = concat_74, begin_mask = new_cache_21_internal_tensor_assign_2_begin_mask_0, end = concat_75, end_mask = new_cache_21_internal_tensor_assign_2_end_mask_0, squeeze_mask = new_cache_21_internal_tensor_assign_2_squeeze_mask_0, stride = new_cache_21_internal_tensor_assign_2_stride_0, x = reshape_105)[name = tensor("slice_by_index_21")]; + tensor reshape_106_shape_0 = const()[name = tensor("reshape_106_shape_0"), val = tensor([-1])]; + tensor reshape_106 = reshape(shape = reshape_106_shape_0, x = slice_by_index_21)[name = tensor("reshape_106")]; + tensor reshape_107_shape_0 = const()[name = tensor("reshape_107_shape_0"), val = tensor([-1])]; + tensor reshape_107_cast_fp16 = reshape(shape = reshape_107_shape_0, x = var_4676_cast_fp16)[name = tensor("reshape_107_cast_fp16")]; + tensor reshape_108_shape_0 = const()[name = tensor("reshape_108_shape_0"), val = tensor([-1])]; + tensor reshape_108_cast_fp16 = reshape(shape = reshape_108_shape_0, x = reshape_104_cast_fp16)[name = tensor("reshape_108_cast_fp16")]; + tensor scatter_21_mode_0 = const()[name = tensor("scatter_21_mode_0"), val = tensor("update")]; + tensor scatter_21_axis_0 = const()[name = tensor("scatter_21_axis_0"), val = tensor(0)]; + tensor scatter_21_validate_indices_0 = const()[name = tensor("scatter_21_validate_indices_0"), val = tensor(false)]; + tensor scatter_21_cast_fp16 = scatter(axis = scatter_21_axis_0, data = reshape_108_cast_fp16, indices = reshape_106, mode = scatter_21_mode_0, updates = reshape_107_cast_fp16, validate_indices = scatter_21_validate_indices_0)[name = tensor("scatter_21_cast_fp16")]; + tensor reshape_109_cast_fp16 = reshape(shape = shape_69, x = scatter_21_cast_fp16)[name = tensor("reshape_109_cast_fp16")]; + tensor reshape_109_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("reshape_109_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor keys_61_begin_0 = const()[name = tensor("keys_61_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor keys_61_end_0 = const()[name = tensor("keys_61_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor keys_61_end_mask_0 = const()[name = tensor("keys_61_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor keys_61_squeeze_mask_0 = const()[name = tensor("keys_61_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor keys_61_cast_fp16 = slice_by_index(begin = keys_61_begin_0, end = keys_61_end_0, end_mask = keys_61_end_mask_0, squeeze_mask = keys_61_squeeze_mask_0, x = reshape_109_cast_fp16)[name = tensor("keys_61_cast_fp16")]; + tensor values_61_begin_0 = const()[name = tensor("values_61_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor values_61_end_0 = const()[name = tensor("values_61_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor values_61_end_mask_0 = const()[name = tensor("values_61_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor values_61_squeeze_mask_0 = const()[name = tensor("values_61_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor values_61_cast_fp16 = slice_by_index(begin = values_61_begin_0, end = values_61_end_0, end_mask = values_61_end_mask_0, squeeze_mask = values_61_squeeze_mask_0, x = reshape_109_cast_fp16)[name = tensor("values_61_cast_fp16")]; + tensor var_4688_cast_fp16 = not_equal(x = keys_61_cast_fp16, y = keys_61_cast_fp16)[name = tensor("op_4688_cast_fp16")]; + tensor keys_63_cast_fp16 = select(a = var_504_to_fp16, b = keys_61_cast_fp16, cond = var_4688_cast_fp16)[name = tensor("keys_63_cast_fp16")]; + tensor var_4696_cast_fp16 = not_equal(x = values_61_cast_fp16, y = values_61_cast_fp16)[name = tensor("op_4696_cast_fp16")]; + tensor values_63_cast_fp16 = select(a = var_504_to_fp16, b = values_61_cast_fp16, cond = var_4696_cast_fp16)[name = tensor("values_63_cast_fp16")]; + tensor var_4720 = const()[name = tensor("op_4720"), val = tensor([0, 2, 1, 3])]; + tensor var_4733 = const()[name = tensor("op_4733"), val = tensor([1, 1, 1])]; + tensor var_4734_cast_fp16 = reshape(shape = var_4733, x = position10_to_fp16)[name = tensor("op_4734_cast_fp16")]; + tensor var_4751_to_fp16 = const()[name = tensor("op_4751_to_fp16"), val = tensor(0x1p+0)]; + tensor valid_len_21_cast_fp16 = add(x = var_4734_cast_fp16, y = var_4751_to_fp16)[name = tensor("valid_len_21_cast_fp16")]; + tensor valid_mask_21_cast_fp16 = less(x = k_positions_1_promoted_to_fp16, y = valid_len_21_cast_fp16)[name = tensor("valid_mask_21_cast_fp16")]; + tensor causal_mask_21_cast_fp16 = less_equal(x = k_positions_1_promoted_to_fp16, y = var_4734_cast_fp16)[name = tensor("causal_mask_21_cast_fp16")]; + tensor attn_mask_41 = logical_and(x = valid_mask_21_cast_fp16, y = causal_mask_21_cast_fp16)[name = tensor("attn_mask_41")]; + tensor attn_mask_43_axes_0 = const()[name = tensor("attn_mask_43_axes_0"), val = tensor([1])]; + tensor attn_mask_43 = expand_dims(axes = attn_mask_43_axes_0, x = attn_mask_41)[name = tensor("attn_mask_43")]; + tensor var_4769_transpose_x_0 = const()[name = tensor("op_4769_transpose_x_0"), val = tensor(false)]; + tensor var_4769_transpose_y_0 = const()[name = tensor("op_4769_transpose_y_0"), val = tensor(false)]; + tensor transpose_92_perm_0 = const()[name = tensor("transpose_92_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_93_perm_0 = const()[name = tensor("transpose_93_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_93 = transpose(perm = transpose_93_perm_0, x = keys_63_cast_fp16)[name = tensor("transpose_173")]; + tensor transpose_92 = transpose(perm = transpose_92_perm_0, x = q_63_cast_fp16)[name = tensor("transpose_174")]; + tensor var_4769_cast_fp16 = matmul(transpose_x = var_4769_transpose_x_0, transpose_y = var_4769_transpose_y_0, x = transpose_92, y = transpose_93)[name = tensor("op_4769_cast_fp16")]; + tensor var_4763_to_fp16 = const()[name = tensor("op_4763_to_fp16"), val = tensor([0x1p-3])]; + tensor attn_weights_61_cast_fp16 = mul(x = var_4769_cast_fp16, y = var_4763_to_fp16)[name = tensor("attn_weights_61_cast_fp16")]; + tensor var_4771 = logical_not(x = attn_mask_43)[name = tensor("op_4771")]; + tensor var_4772_to_fp16 = const()[name = tensor("op_4772_to_fp16"), val = tensor(-inf)]; + tensor attn_weights_63_cast_fp16 = select(a = var_4772_to_fp16, b = attn_weights_61_cast_fp16, cond = var_4771)[name = tensor("attn_weights_63_cast_fp16")]; + tensor var_4774 = const()[name = tensor("op_4774"), val = tensor(-1)]; + tensor attn_weights_65_cast_fp16 = softmax(axis = var_4774, x = attn_weights_63_cast_fp16)[name = tensor("attn_weights_65_cast_fp16")]; + tensor attn_output_21_transpose_x_0 = const()[name = tensor("attn_output_21_transpose_x_0"), val = tensor(false)]; + tensor attn_output_21_transpose_y_0 = const()[name = tensor("attn_output_21_transpose_y_0"), val = tensor(false)]; + tensor values_65_cast_fp16 = transpose(perm = var_4720, x = values_63_cast_fp16)[name = tensor("transpose_175")]; + tensor attn_output_21_cast_fp16 = matmul(transpose_x = attn_output_21_transpose_x_0, transpose_y = attn_output_21_transpose_y_0, x = attn_weights_65_cast_fp16, y = values_65_cast_fp16)[name = tensor("attn_output_21_cast_fp16")]; + tensor var_4782 = const()[name = tensor("op_4782"), val = tensor([0, 2, 1, 3])]; + tensor var_4785 = const()[name = tensor("op_4785"), val = tensor([1, 1, 1024])]; + tensor var_4783_cast_fp16 = transpose(perm = var_4782, x = attn_output_21_cast_fp16)[name = tensor("transpose_172")]; + tensor input_105_cast_fp16 = reshape(shape = var_4785, x = var_4783_cast_fp16)[name = tensor("input_105_cast_fp16")]; + tensor linear_42_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = attn10_out_proj_weight_cast_fp16, x = input_105_cast_fp16)[name = tensor("linear_42_cast_fp16")]; + tensor var_4791_to_fp16 = const()[name = tensor("op_4791_to_fp16"), val = tensor(0x1p+0)]; + tensor var_4792_cast_fp16 = add(x = position10_to_fp16, y = var_4791_to_fp16)[name = tensor("op_4792_cast_fp16")]; + tensor var_4792_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_4792_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor input_107_cast_fp16 = add(x = input_103_cast_fp16, y = linear_42_cast_fp16)[name = tensor("input_107_cast_fp16")]; + tensor input_109_axes_0 = const()[name = tensor("input_109_axes_0"), val = tensor([-1])]; + tensor norm10_2_weight_to_fp16 = const()[name = tensor("norm10_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(303675008)))]; + tensor norm10_2_bias_to_fp16 = const()[name = tensor("norm10_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(303677120)))]; + tensor var_4796_to_fp16 = const()[name = tensor("op_4796_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_109_cast_fp16 = layer_norm(axes = input_109_axes_0, beta = norm10_2_bias_to_fp16, epsilon = var_4796_to_fp16, gamma = norm10_2_weight_to_fp16, x = input_107_cast_fp16)[name = tensor("input_109_cast_fp16")]; + tensor linear_43_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = linear10_1_weight_cast_fp16, x = input_109_cast_fp16)[name = tensor("linear_43_cast_fp16")]; + tensor input_111_mode_0 = const()[name = tensor("input_111_mode_0"), val = tensor("EXACT")]; + tensor input_111_cast_fp16 = gelu(mode = input_111_mode_0, x = linear_43_cast_fp16)[name = tensor("input_111_cast_fp16")]; + tensor linear_44_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = linear10_2_weight_cast_fp16, x = input_111_cast_fp16)[name = tensor("linear_44_cast_fp16")]; + tensor input_113_cast_fp16 = add(x = input_107_cast_fp16, y = linear_44_cast_fp16)[name = tensor("input_113_cast_fp16")]; + tensor x_23_axes_0 = const()[name = tensor("x_23_axes_0"), val = tensor([-1])]; + tensor norm11_1_weight_to_fp16 = const()[name = tensor("norm11_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(303679232)))]; + tensor norm11_1_bias_to_fp16 = const()[name = tensor("norm11_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(303681344)))]; + tensor var_4813_to_fp16 = const()[name = tensor("op_4813_to_fp16"), val = tensor(0x1.5p-17)]; + tensor x_23_cast_fp16 = layer_norm(axes = x_23_axes_0, beta = norm11_1_bias_to_fp16, epsilon = var_4813_to_fp16, gamma = norm11_1_weight_to_fp16, x = input_113_cast_fp16)[name = tensor("x_23_cast_fp16")]; + tensor linear_45_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = attn11_in_proj_weight_cast_fp16, x = x_23_cast_fp16)[name = tensor("linear_45_cast_fp16")]; + tensor var_4849 = const()[name = tensor("op_4849"), val = tensor([1, 1, 3, 16, 64])]; + tensor qkv_23_cast_fp16 = reshape(shape = var_4849, x = linear_45_cast_fp16)[name = tensor("qkv_23_cast_fp16")]; + tensor q_67_begin_0 = const()[name = tensor("q_67_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor q_67_end_0 = const()[name = tensor("q_67_end_0"), val = tensor([1, 1, 1, 16, 64])]; + tensor q_67_end_mask_0 = const()[name = tensor("q_67_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor q_67_squeeze_mask_0 = const()[name = tensor("q_67_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor q_67_cast_fp16 = slice_by_index(begin = q_67_begin_0, end = q_67_end_0, end_mask = q_67_end_mask_0, squeeze_mask = q_67_squeeze_mask_0, x = qkv_23_cast_fp16)[name = tensor("q_67_cast_fp16")]; + tensor k_45_begin_0 = const()[name = tensor("k_45_begin_0"), val = tensor([0, 0, 1, 0, 0])]; + tensor k_45_end_0 = const()[name = tensor("k_45_end_0"), val = tensor([1, 1, 2, 16, 64])]; + tensor k_45_end_mask_0 = const()[name = tensor("k_45_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor k_45_squeeze_mask_0 = const()[name = tensor("k_45_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor k_45_cast_fp16 = slice_by_index(begin = k_45_begin_0, end = k_45_end_0, end_mask = k_45_end_mask_0, squeeze_mask = k_45_squeeze_mask_0, x = qkv_23_cast_fp16)[name = tensor("k_45_cast_fp16")]; + tensor v_23_begin_0 = const()[name = tensor("v_23_begin_0"), val = tensor([0, 0, 2, 0, 0])]; + tensor v_23_end_0 = const()[name = tensor("v_23_end_0"), val = tensor([1, 1, 3, 16, 64])]; + tensor v_23_end_mask_0 = const()[name = tensor("v_23_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor v_23_squeeze_mask_0 = const()[name = tensor("v_23_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor v_23_cast_fp16 = slice_by_index(begin = v_23_begin_0, end = v_23_end_0, end_mask = v_23_end_mask_0, squeeze_mask = v_23_squeeze_mask_0, x = qkv_23_cast_fp16)[name = tensor("v_23_cast_fp16")]; + tensor var_4953 = const()[name = tensor("op_4953"), val = tensor([1, 1, 1, 1])]; + tensor position11_to_fp16_dtype_0 = const()[name = tensor("position11_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor position11_to_fp16 = cast(dtype = position11_to_fp16_dtype_0, x = position11)[name = tensor("cast_498")]; + tensor ts_71_cast_fp16 = reshape(shape = var_4953, x = position11_to_fp16)[name = tensor("ts_71_cast_fp16")]; + tensor var_4957 = const()[name = tensor("op_4957"), val = tensor([1, 1, 16, 32, 2])]; + tensor q_complex_23_cast_fp16 = reshape(shape = var_4957, x = q_67_cast_fp16)[name = tensor("q_complex_23_cast_fp16")]; + tensor var_4961 = const()[name = tensor("op_4961"), val = tensor([1, 1, 16, 32, 2])]; + tensor k_complex_23_cast_fp16 = reshape(shape = var_4961, x = k_45_cast_fp16)[name = tensor("k_complex_23_cast_fp16")]; + tensor var_4965_begin_0 = const()[name = tensor("op_4965_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_4965_end_0 = const()[name = tensor("op_4965_end_0"), val = tensor([1, 1, 16, 32, 1])]; + tensor var_4965_end_mask_0 = const()[name = tensor("op_4965_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_4965_squeeze_mask_0 = const()[name = tensor("op_4965_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_4965_cast_fp16 = slice_by_index(begin = var_4965_begin_0, end = var_4965_end_0, end_mask = var_4965_end_mask_0, squeeze_mask = var_4965_squeeze_mask_0, x = q_complex_23_cast_fp16)[name = tensor("op_4965_cast_fp16")]; + tensor var_4973_begin_0 = const()[name = tensor("op_4973_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_4973_end_0 = const()[name = tensor("op_4973_end_0"), val = tensor([1, 1, 16, 32, 2])]; + tensor var_4973_end_mask_0 = const()[name = tensor("op_4973_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_4973_squeeze_mask_0 = const()[name = tensor("op_4973_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_4973_cast_fp16 = slice_by_index(begin = var_4973_begin_0, end = var_4973_end_0, end_mask = var_4973_end_mask_0, squeeze_mask = var_4973_squeeze_mask_0, x = q_complex_23_cast_fp16)[name = tensor("op_4973_cast_fp16")]; + tensor var_4981_begin_0 = const()[name = tensor("op_4981_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_4981_end_0 = const()[name = tensor("op_4981_end_0"), val = tensor([1, 1, 16, 32, 1])]; + tensor var_4981_end_mask_0 = const()[name = tensor("op_4981_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_4981_squeeze_mask_0 = const()[name = tensor("op_4981_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_4981_cast_fp16 = slice_by_index(begin = var_4981_begin_0, end = var_4981_end_0, end_mask = var_4981_end_mask_0, squeeze_mask = var_4981_squeeze_mask_0, x = k_complex_23_cast_fp16)[name = tensor("op_4981_cast_fp16")]; + tensor var_4989_begin_0 = const()[name = tensor("op_4989_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_4989_end_0 = const()[name = tensor("op_4989_end_0"), val = tensor([1, 1, 16, 32, 2])]; + tensor var_4989_end_mask_0 = const()[name = tensor("op_4989_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_4989_squeeze_mask_0 = const()[name = tensor("op_4989_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_4989_cast_fp16 = slice_by_index(begin = var_4989_begin_0, end = var_4989_end_0, end_mask = var_4989_end_mask_0, squeeze_mask = var_4989_squeeze_mask_0, x = k_complex_23_cast_fp16)[name = tensor("op_4989_cast_fp16")]; + tensor freqs_23_to_fp16 = const()[name = tensor("freqs_23_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(303683456)))]; + tensor var_4995_cast_fp16 = mul(x = freqs_23_to_fp16, y = ts_71_cast_fp16)[name = tensor("op_4995_cast_fp16")]; + tensor rotr_23_cast_fp16 = cos(x = var_4995_cast_fp16)[name = tensor("rotr_23_cast_fp16")]; + tensor roti_23_cast_fp16 = sin(x = var_4995_cast_fp16)[name = tensor("roti_23_cast_fp16")]; + tensor var_4999_cast_fp16 = mul(x = var_4965_cast_fp16, y = rotr_23_cast_fp16)[name = tensor("op_4999_cast_fp16")]; + tensor var_5000_cast_fp16 = mul(x = var_4973_cast_fp16, y = roti_23_cast_fp16)[name = tensor("op_5000_cast_fp16")]; + tensor qor_45_cast_fp16 = sub(x = var_4999_cast_fp16, y = var_5000_cast_fp16)[name = tensor("qor_45_cast_fp16")]; + tensor var_5003_cast_fp16 = mul(x = var_4965_cast_fp16, y = roti_23_cast_fp16)[name = tensor("op_5003_cast_fp16")]; + tensor var_5004_cast_fp16 = mul(x = var_4973_cast_fp16, y = rotr_23_cast_fp16)[name = tensor("op_5004_cast_fp16")]; + tensor qoi_45_cast_fp16 = add(x = var_5003_cast_fp16, y = var_5004_cast_fp16)[name = tensor("qoi_45_cast_fp16")]; + tensor var_5007_cast_fp16 = mul(x = var_4981_cast_fp16, y = rotr_23_cast_fp16)[name = tensor("op_5007_cast_fp16")]; + tensor var_5008_cast_fp16 = mul(x = var_4989_cast_fp16, y = roti_23_cast_fp16)[name = tensor("op_5008_cast_fp16")]; + tensor kor_45_cast_fp16 = sub(x = var_5007_cast_fp16, y = var_5008_cast_fp16)[name = tensor("kor_45_cast_fp16")]; + tensor var_5011_cast_fp16 = mul(x = var_4981_cast_fp16, y = roti_23_cast_fp16)[name = tensor("op_5011_cast_fp16")]; + tensor var_5012_cast_fp16 = mul(x = var_4989_cast_fp16, y = rotr_23_cast_fp16)[name = tensor("op_5012_cast_fp16")]; + tensor koi_45_cast_fp16 = add(x = var_5011_cast_fp16, y = var_5012_cast_fp16)[name = tensor("koi_45_cast_fp16")]; + tensor qo_23_axis_0 = const()[name = tensor("qo_23_axis_0"), val = tensor(-1)]; + tensor qo_23_cast_fp16 = stack(axis = qo_23_axis_0, values = (qor_45_cast_fp16, qoi_45_cast_fp16))[name = tensor("qo_23_cast_fp16")]; + tensor ko_23_axis_0 = const()[name = tensor("ko_23_axis_0"), val = tensor(-1)]; + tensor ko_23_cast_fp16 = stack(axis = ko_23_axis_0, values = (kor_45_cast_fp16, koi_45_cast_fp16))[name = tensor("ko_23_cast_fp16")]; + tensor var_5041 = const()[name = tensor("op_5041"), val = tensor([1, 1, 16, 64])]; + tensor q_69_cast_fp16 = reshape(shape = var_5041, x = qo_23_cast_fp16)[name = tensor("q_69_cast_fp16")]; + tensor var_5043 = const()[name = tensor("op_5043"), val = tensor([1, 1, 16, 64])]; + tensor k_47_cast_fp16 = reshape(shape = var_5043, x = ko_23_cast_fp16)[name = tensor("k_47_cast_fp16")]; + tensor _inversed_5065_y_0_to_fp16 = const()[name = tensor("_inversed_5065_y_0_to_fp16"), val = tensor(0x1p-9)]; + tensor _inversed_5065_cast_fp16 = mul(x = ts_71_cast_fp16, y = _inversed_5065_y_0_to_fp16)[name = tensor("_inversed_5065_cast_fp16")]; + tensor var_5066_cast_fp16 = floor(x = _inversed_5065_cast_fp16)[name = tensor("op_5066_cast_fp16")]; + tensor var_5067_to_fp16 = const()[name = tensor("op_5067_to_fp16"), val = tensor(0x1p+9)]; + tensor var_5068_cast_fp16 = mul(x = var_5066_cast_fp16, y = var_5067_to_fp16)[name = tensor("op_5068_cast_fp16")]; + tensor write_indices_float_47_cast_fp16 = sub(x = ts_71_cast_fp16, y = var_5068_cast_fp16)[name = tensor("write_indices_float_47_cast_fp16")]; + tensor var_5075_dtype_0 = const()[name = tensor("op_5075_dtype_0"), val = tensor("int32")]; + tensor write_indices_23_reps_0 = const()[name = tensor("write_indices_23_reps_0"), val = tensor([1, 1, 16, 64])]; + tensor write_indices_float_47_cast_fp16_to_int32 = cast(dtype = var_5075_dtype_0, x = write_indices_float_47_cast_fp16)[name = tensor("cast_497")]; + tensor write_indices_23 = tile(reps = write_indices_23_reps_0, x = write_indices_float_47_cast_fp16_to_int32)[name = tensor("write_indices_23")]; + tensor var_5083_begin_0 = const()[name = tensor("op_5083_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_5083_end_0 = const()[name = tensor("op_5083_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor var_5083_end_mask_0 = const()[name = tensor("op_5083_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_5083_squeeze_mask_0 = const()[name = tensor("op_5083_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor cache11_to_fp16_dtype_0 = const()[name = tensor("cache11_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor cache11_to_fp16 = cast(dtype = cache11_to_fp16_dtype_0, x = cache11)[name = tensor("cast_496")]; + tensor var_5083_cast_fp16 = slice_by_index(begin = var_5083_begin_0, end = var_5083_end_0, end_mask = var_5083_end_mask_0, squeeze_mask = var_5083_squeeze_mask_0, x = cache11_to_fp16)[name = tensor("op_5083_cast_fp16")]; + tensor var_5085_axis_0 = const()[name = tensor("op_5085_axis_0"), val = tensor(1)]; + tensor var_5085_mode_0 = const()[name = tensor("op_5085_mode_0"), val = tensor("update")]; + tensor var_5085_validate_indices_0 = const()[name = tensor("op_5085_validate_indices_0"), val = tensor(false)]; + tensor var_5085_cast_fp16 = scatter_along_axis(axis = var_5085_axis_0, data = var_5083_cast_fp16, indices = write_indices_23, mode = var_5085_mode_0, updates = k_47_cast_fp16, validate_indices = var_5085_validate_indices_0)[name = tensor("op_5085_cast_fp16")]; + tensor concat_79 = const()[name = tensor("concat_79"), val = tensor([0, 0, 0, 0, 0])]; + tensor concat_80 = const()[name = tensor("concat_80"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_23_internal_tensor_assign_1_stride_0 = const()[name = tensor("new_cache_23_internal_tensor_assign_1_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_23_internal_tensor_assign_1_begin_mask_0 = const()[name = tensor("new_cache_23_internal_tensor_assign_1_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_23_internal_tensor_assign_1_end_mask_0 = const()[name = tensor("new_cache_23_internal_tensor_assign_1_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_23_internal_tensor_assign_1_squeeze_mask_0 = const()[name = tensor("new_cache_23_internal_tensor_assign_1_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_70 = const()[name = tensor("shape_70"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_22 = const()[name = tensor("reduce_prod_22"), val = tensor(1048576)]; + tensor range_1d_22_start_0 = const()[name = tensor("range_1d_22_start_0"), val = tensor(0)]; + tensor range_1d_22_step_0 = const()[name = tensor("range_1d_22_step_0"), val = tensor(1)]; + tensor range_1d_22 = range_1d(end = reduce_prod_22, start = range_1d_22_start_0, step = range_1d_22_step_0)[name = tensor("range_1d_22")]; + tensor reshape_110 = reshape(shape = shape_70, x = range_1d_22)[name = tensor("reshape_110")]; + tensor slice_by_index_22 = slice_by_index(begin = concat_79, begin_mask = new_cache_23_internal_tensor_assign_1_begin_mask_0, end = concat_80, end_mask = new_cache_23_internal_tensor_assign_1_end_mask_0, squeeze_mask = new_cache_23_internal_tensor_assign_1_squeeze_mask_0, stride = new_cache_23_internal_tensor_assign_1_stride_0, x = reshape_110)[name = tensor("slice_by_index_22")]; + tensor reshape_111_shape_0 = const()[name = tensor("reshape_111_shape_0"), val = tensor([-1])]; + tensor reshape_111 = reshape(shape = reshape_111_shape_0, x = slice_by_index_22)[name = tensor("reshape_111")]; + tensor reshape_112_shape_0 = const()[name = tensor("reshape_112_shape_0"), val = tensor([-1])]; + tensor reshape_112_cast_fp16 = reshape(shape = reshape_112_shape_0, x = var_5085_cast_fp16)[name = tensor("reshape_112_cast_fp16")]; + tensor reshape_113_shape_0 = const()[name = tensor("reshape_113_shape_0"), val = tensor([-1])]; + tensor reshape_113_cast_fp16 = reshape(shape = reshape_113_shape_0, x = cache11_to_fp16)[name = tensor("reshape_113_cast_fp16")]; + tensor scatter_22_mode_0 = const()[name = tensor("scatter_22_mode_0"), val = tensor("update")]; + tensor scatter_22_axis_0 = const()[name = tensor("scatter_22_axis_0"), val = tensor(0)]; + tensor scatter_22_validate_indices_0 = const()[name = tensor("scatter_22_validate_indices_0"), val = tensor(false)]; + tensor scatter_22_cast_fp16 = scatter(axis = scatter_22_axis_0, data = reshape_113_cast_fp16, indices = reshape_111, mode = scatter_22_mode_0, updates = reshape_112_cast_fp16, validate_indices = scatter_22_validate_indices_0)[name = tensor("scatter_22_cast_fp16")]; + tensor reshape_114_cast_fp16 = reshape(shape = shape_70, x = scatter_22_cast_fp16)[name = tensor("reshape_114_cast_fp16")]; + tensor var_5093_begin_0 = const()[name = tensor("op_5093_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor var_5093_end_0 = const()[name = tensor("op_5093_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor var_5093_end_mask_0 = const()[name = tensor("op_5093_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_5093_squeeze_mask_0 = const()[name = tensor("op_5093_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor var_5093_cast_fp16 = slice_by_index(begin = var_5093_begin_0, end = var_5093_end_0, end_mask = var_5093_end_mask_0, squeeze_mask = var_5093_squeeze_mask_0, x = reshape_114_cast_fp16)[name = tensor("op_5093_cast_fp16")]; + tensor var_5095_axis_0 = const()[name = tensor("op_5095_axis_0"), val = tensor(1)]; + tensor var_5095_mode_0 = const()[name = tensor("op_5095_mode_0"), val = tensor("update")]; + tensor var_5095_validate_indices_0 = const()[name = tensor("op_5095_validate_indices_0"), val = tensor(false)]; + tensor var_5095_cast_fp16 = scatter_along_axis(axis = var_5095_axis_0, data = var_5093_cast_fp16, indices = write_indices_23, mode = var_5095_mode_0, updates = v_23_cast_fp16, validate_indices = var_5095_validate_indices_0)[name = tensor("op_5095_cast_fp16")]; + tensor concat_81 = const()[name = tensor("concat_81"), val = tensor([1, 0, 0, 0, 0])]; + tensor concat_82 = const()[name = tensor("concat_82"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_23_internal_tensor_assign_2_stride_0 = const()[name = tensor("new_cache_23_internal_tensor_assign_2_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_23_internal_tensor_assign_2_begin_mask_0 = const()[name = tensor("new_cache_23_internal_tensor_assign_2_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_23_internal_tensor_assign_2_end_mask_0 = const()[name = tensor("new_cache_23_internal_tensor_assign_2_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_23_internal_tensor_assign_2_squeeze_mask_0 = const()[name = tensor("new_cache_23_internal_tensor_assign_2_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_71 = const()[name = tensor("shape_71"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_23 = const()[name = tensor("reduce_prod_23"), val = tensor(1048576)]; + tensor range_1d_23_start_0 = const()[name = tensor("range_1d_23_start_0"), val = tensor(0)]; + tensor range_1d_23_step_0 = const()[name = tensor("range_1d_23_step_0"), val = tensor(1)]; + tensor range_1d_23 = range_1d(end = reduce_prod_23, start = range_1d_23_start_0, step = range_1d_23_step_0)[name = tensor("range_1d_23")]; + tensor reshape_115 = reshape(shape = shape_71, x = range_1d_23)[name = tensor("reshape_115")]; + tensor slice_by_index_23 = slice_by_index(begin = concat_81, begin_mask = new_cache_23_internal_tensor_assign_2_begin_mask_0, end = concat_82, end_mask = new_cache_23_internal_tensor_assign_2_end_mask_0, squeeze_mask = new_cache_23_internal_tensor_assign_2_squeeze_mask_0, stride = new_cache_23_internal_tensor_assign_2_stride_0, x = reshape_115)[name = tensor("slice_by_index_23")]; + tensor reshape_116_shape_0 = const()[name = tensor("reshape_116_shape_0"), val = tensor([-1])]; + tensor reshape_116 = reshape(shape = reshape_116_shape_0, x = slice_by_index_23)[name = tensor("reshape_116")]; + tensor reshape_117_shape_0 = const()[name = tensor("reshape_117_shape_0"), val = tensor([-1])]; + tensor reshape_117_cast_fp16 = reshape(shape = reshape_117_shape_0, x = var_5095_cast_fp16)[name = tensor("reshape_117_cast_fp16")]; + tensor reshape_118_shape_0 = const()[name = tensor("reshape_118_shape_0"), val = tensor([-1])]; + tensor reshape_118_cast_fp16 = reshape(shape = reshape_118_shape_0, x = reshape_114_cast_fp16)[name = tensor("reshape_118_cast_fp16")]; + tensor scatter_23_mode_0 = const()[name = tensor("scatter_23_mode_0"), val = tensor("update")]; + tensor scatter_23_axis_0 = const()[name = tensor("scatter_23_axis_0"), val = tensor(0)]; + tensor scatter_23_validate_indices_0 = const()[name = tensor("scatter_23_validate_indices_0"), val = tensor(false)]; + tensor scatter_23_cast_fp16 = scatter(axis = scatter_23_axis_0, data = reshape_118_cast_fp16, indices = reshape_116, mode = scatter_23_mode_0, updates = reshape_117_cast_fp16, validate_indices = scatter_23_validate_indices_0)[name = tensor("scatter_23_cast_fp16")]; + tensor reshape_119_cast_fp16 = reshape(shape = shape_71, x = scatter_23_cast_fp16)[name = tensor("reshape_119_cast_fp16")]; + tensor reshape_119_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("reshape_119_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor keys_67_begin_0 = const()[name = tensor("keys_67_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor keys_67_end_0 = const()[name = tensor("keys_67_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor keys_67_end_mask_0 = const()[name = tensor("keys_67_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor keys_67_squeeze_mask_0 = const()[name = tensor("keys_67_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor keys_67_cast_fp16 = slice_by_index(begin = keys_67_begin_0, end = keys_67_end_0, end_mask = keys_67_end_mask_0, squeeze_mask = keys_67_squeeze_mask_0, x = reshape_119_cast_fp16)[name = tensor("keys_67_cast_fp16")]; + tensor values_67_begin_0 = const()[name = tensor("values_67_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor values_67_end_0 = const()[name = tensor("values_67_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor values_67_end_mask_0 = const()[name = tensor("values_67_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor values_67_squeeze_mask_0 = const()[name = tensor("values_67_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor values_67_cast_fp16 = slice_by_index(begin = values_67_begin_0, end = values_67_end_0, end_mask = values_67_end_mask_0, squeeze_mask = values_67_squeeze_mask_0, x = reshape_119_cast_fp16)[name = tensor("values_67_cast_fp16")]; + tensor var_5107_cast_fp16 = not_equal(x = keys_67_cast_fp16, y = keys_67_cast_fp16)[name = tensor("op_5107_cast_fp16")]; + tensor keys_69_cast_fp16 = select(a = var_504_to_fp16, b = keys_67_cast_fp16, cond = var_5107_cast_fp16)[name = tensor("keys_69_cast_fp16")]; + tensor var_5115_cast_fp16 = not_equal(x = values_67_cast_fp16, y = values_67_cast_fp16)[name = tensor("op_5115_cast_fp16")]; + tensor values_69_cast_fp16 = select(a = var_504_to_fp16, b = values_67_cast_fp16, cond = var_5115_cast_fp16)[name = tensor("values_69_cast_fp16")]; + tensor var_5139 = const()[name = tensor("op_5139"), val = tensor([0, 2, 1, 3])]; + tensor var_5152 = const()[name = tensor("op_5152"), val = tensor([1, 1, 1])]; + tensor var_5153_cast_fp16 = reshape(shape = var_5152, x = position11_to_fp16)[name = tensor("op_5153_cast_fp16")]; + tensor var_5170_to_fp16 = const()[name = tensor("op_5170_to_fp16"), val = tensor(0x1p+0)]; + tensor valid_len_23_cast_fp16 = add(x = var_5153_cast_fp16, y = var_5170_to_fp16)[name = tensor("valid_len_23_cast_fp16")]; + tensor valid_mask_23_cast_fp16 = less(x = k_positions_1_promoted_to_fp16, y = valid_len_23_cast_fp16)[name = tensor("valid_mask_23_cast_fp16")]; + tensor causal_mask_23_cast_fp16 = less_equal(x = k_positions_1_promoted_to_fp16, y = var_5153_cast_fp16)[name = tensor("causal_mask_23_cast_fp16")]; + tensor attn_mask_45 = logical_and(x = valid_mask_23_cast_fp16, y = causal_mask_23_cast_fp16)[name = tensor("attn_mask_45")]; + tensor attn_mask_47_axes_0 = const()[name = tensor("attn_mask_47_axes_0"), val = tensor([1])]; + tensor attn_mask_47 = expand_dims(axes = attn_mask_47_axes_0, x = attn_mask_45)[name = tensor("attn_mask_47")]; + tensor var_5188_transpose_x_0 = const()[name = tensor("op_5188_transpose_x_0"), val = tensor(false)]; + tensor var_5188_transpose_y_0 = const()[name = tensor("op_5188_transpose_y_0"), val = tensor(false)]; + tensor transpose_94_perm_0 = const()[name = tensor("transpose_94_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_95_perm_0 = const()[name = tensor("transpose_95_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_95 = transpose(perm = transpose_95_perm_0, x = keys_69_cast_fp16)[name = tensor("transpose_169")]; + tensor transpose_94 = transpose(perm = transpose_94_perm_0, x = q_69_cast_fp16)[name = tensor("transpose_170")]; + tensor var_5188_cast_fp16 = matmul(transpose_x = var_5188_transpose_x_0, transpose_y = var_5188_transpose_y_0, x = transpose_94, y = transpose_95)[name = tensor("op_5188_cast_fp16")]; + tensor var_5182_to_fp16 = const()[name = tensor("op_5182_to_fp16"), val = tensor([0x1p-3])]; + tensor attn_weights_67_cast_fp16 = mul(x = var_5188_cast_fp16, y = var_5182_to_fp16)[name = tensor("attn_weights_67_cast_fp16")]; + tensor var_5190 = logical_not(x = attn_mask_47)[name = tensor("op_5190")]; + tensor var_5191_to_fp16 = const()[name = tensor("op_5191_to_fp16"), val = tensor(-inf)]; + tensor attn_weights_69_cast_fp16 = select(a = var_5191_to_fp16, b = attn_weights_67_cast_fp16, cond = var_5190)[name = tensor("attn_weights_69_cast_fp16")]; + tensor var_5193 = const()[name = tensor("op_5193"), val = tensor(-1)]; + tensor attn_weights_71_cast_fp16 = softmax(axis = var_5193, x = attn_weights_69_cast_fp16)[name = tensor("attn_weights_71_cast_fp16")]; + tensor attn_output_23_transpose_x_0 = const()[name = tensor("attn_output_23_transpose_x_0"), val = tensor(false)]; + tensor attn_output_23_transpose_y_0 = const()[name = tensor("attn_output_23_transpose_y_0"), val = tensor(false)]; + tensor values_71_cast_fp16 = transpose(perm = var_5139, x = values_69_cast_fp16)[name = tensor("transpose_171")]; + tensor attn_output_23_cast_fp16 = matmul(transpose_x = attn_output_23_transpose_x_0, transpose_y = attn_output_23_transpose_y_0, x = attn_weights_71_cast_fp16, y = values_71_cast_fp16)[name = tensor("attn_output_23_cast_fp16")]; + tensor var_5201 = const()[name = tensor("op_5201"), val = tensor([0, 2, 1, 3])]; + tensor var_5204 = const()[name = tensor("op_5204"), val = tensor([1, 1, 1024])]; + tensor var_5202_cast_fp16 = transpose(perm = var_5201, x = attn_output_23_cast_fp16)[name = tensor("transpose_168")]; + tensor input_115_cast_fp16 = reshape(shape = var_5204, x = var_5202_cast_fp16)[name = tensor("input_115_cast_fp16")]; + tensor linear_46_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = attn11_out_proj_weight_cast_fp16, x = input_115_cast_fp16)[name = tensor("linear_46_cast_fp16")]; + tensor var_5210_to_fp16 = const()[name = tensor("op_5210_to_fp16"), val = tensor(0x1p+0)]; + tensor var_5211_cast_fp16 = add(x = position11_to_fp16, y = var_5210_to_fp16)[name = tensor("op_5211_cast_fp16")]; + tensor var_5211_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_5211_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor input_117_cast_fp16 = add(x = input_113_cast_fp16, y = linear_46_cast_fp16)[name = tensor("input_117_cast_fp16")]; + tensor input_119_axes_0 = const()[name = tensor("input_119_axes_0"), val = tensor([-1])]; + tensor norm11_2_weight_to_fp16 = const()[name = tensor("norm11_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(303683584)))]; + tensor norm11_2_bias_to_fp16 = const()[name = tensor("norm11_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(303685696)))]; + tensor var_5215_to_fp16 = const()[name = tensor("op_5215_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_119_cast_fp16 = layer_norm(axes = input_119_axes_0, beta = norm11_2_bias_to_fp16, epsilon = var_5215_to_fp16, gamma = norm11_2_weight_to_fp16, x = input_117_cast_fp16)[name = tensor("input_119_cast_fp16")]; + tensor linear_47_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = linear11_1_weight_cast_fp16, x = input_119_cast_fp16)[name = tensor("linear_47_cast_fp16")]; + tensor input_121_mode_0 = const()[name = tensor("input_121_mode_0"), val = tensor("EXACT")]; + tensor input_121_cast_fp16 = gelu(mode = input_121_mode_0, x = linear_47_cast_fp16)[name = tensor("input_121_cast_fp16")]; + tensor linear_48_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = linear11_2_weight_cast_fp16, x = input_121_cast_fp16)[name = tensor("linear_48_cast_fp16")]; + tensor input_123_cast_fp16 = add(x = input_117_cast_fp16, y = linear_48_cast_fp16)[name = tensor("input_123_cast_fp16")]; + tensor x_25_axes_0 = const()[name = tensor("x_25_axes_0"), val = tensor([-1])]; + tensor norm12_1_weight_to_fp16 = const()[name = tensor("norm12_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(303687808)))]; + tensor norm12_1_bias_to_fp16 = const()[name = tensor("norm12_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(303689920)))]; + tensor var_5232_to_fp16 = const()[name = tensor("op_5232_to_fp16"), val = tensor(0x1.5p-17)]; + tensor x_25_cast_fp16 = layer_norm(axes = x_25_axes_0, beta = norm12_1_bias_to_fp16, epsilon = var_5232_to_fp16, gamma = norm12_1_weight_to_fp16, x = input_123_cast_fp16)[name = tensor("x_25_cast_fp16")]; + tensor linear_49_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = attn12_in_proj_weight_cast_fp16, x = x_25_cast_fp16)[name = tensor("linear_49_cast_fp16")]; + tensor var_5268 = const()[name = tensor("op_5268"), val = tensor([1, 1, 3, 16, 64])]; + tensor qkv_25_cast_fp16 = reshape(shape = var_5268, x = linear_49_cast_fp16)[name = tensor("qkv_25_cast_fp16")]; + tensor q_73_begin_0 = const()[name = tensor("q_73_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor q_73_end_0 = const()[name = tensor("q_73_end_0"), val = tensor([1, 1, 1, 16, 64])]; + tensor q_73_end_mask_0 = const()[name = tensor("q_73_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor q_73_squeeze_mask_0 = const()[name = tensor("q_73_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor q_73_cast_fp16 = slice_by_index(begin = q_73_begin_0, end = q_73_end_0, end_mask = q_73_end_mask_0, squeeze_mask = q_73_squeeze_mask_0, x = qkv_25_cast_fp16)[name = tensor("q_73_cast_fp16")]; + tensor k_49_begin_0 = const()[name = tensor("k_49_begin_0"), val = tensor([0, 0, 1, 0, 0])]; + tensor k_49_end_0 = const()[name = tensor("k_49_end_0"), val = tensor([1, 1, 2, 16, 64])]; + tensor k_49_end_mask_0 = const()[name = tensor("k_49_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor k_49_squeeze_mask_0 = const()[name = tensor("k_49_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor k_49_cast_fp16 = slice_by_index(begin = k_49_begin_0, end = k_49_end_0, end_mask = k_49_end_mask_0, squeeze_mask = k_49_squeeze_mask_0, x = qkv_25_cast_fp16)[name = tensor("k_49_cast_fp16")]; + tensor v_25_begin_0 = const()[name = tensor("v_25_begin_0"), val = tensor([0, 0, 2, 0, 0])]; + tensor v_25_end_0 = const()[name = tensor("v_25_end_0"), val = tensor([1, 1, 3, 16, 64])]; + tensor v_25_end_mask_0 = const()[name = tensor("v_25_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor v_25_squeeze_mask_0 = const()[name = tensor("v_25_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor v_25_cast_fp16 = slice_by_index(begin = v_25_begin_0, end = v_25_end_0, end_mask = v_25_end_mask_0, squeeze_mask = v_25_squeeze_mask_0, x = qkv_25_cast_fp16)[name = tensor("v_25_cast_fp16")]; + tensor var_5372 = const()[name = tensor("op_5372"), val = tensor([1, 1, 1, 1])]; + tensor position12_to_fp16_dtype_0 = const()[name = tensor("position12_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor position12_to_fp16 = cast(dtype = position12_to_fp16_dtype_0, x = position12)[name = tensor("cast_493")]; + tensor ts_77_cast_fp16 = reshape(shape = var_5372, x = position12_to_fp16)[name = tensor("ts_77_cast_fp16")]; + tensor var_5376 = const()[name = tensor("op_5376"), val = tensor([1, 1, 16, 32, 2])]; + tensor q_complex_25_cast_fp16 = reshape(shape = var_5376, x = q_73_cast_fp16)[name = tensor("q_complex_25_cast_fp16")]; + tensor var_5380 = const()[name = tensor("op_5380"), val = tensor([1, 1, 16, 32, 2])]; + tensor k_complex_25_cast_fp16 = reshape(shape = var_5380, x = k_49_cast_fp16)[name = tensor("k_complex_25_cast_fp16")]; + tensor var_5384_begin_0 = const()[name = tensor("op_5384_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_5384_end_0 = const()[name = tensor("op_5384_end_0"), val = tensor([1, 1, 16, 32, 1])]; + tensor var_5384_end_mask_0 = const()[name = tensor("op_5384_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_5384_squeeze_mask_0 = const()[name = tensor("op_5384_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_5384_cast_fp16 = slice_by_index(begin = var_5384_begin_0, end = var_5384_end_0, end_mask = var_5384_end_mask_0, squeeze_mask = var_5384_squeeze_mask_0, x = q_complex_25_cast_fp16)[name = tensor("op_5384_cast_fp16")]; + tensor var_5392_begin_0 = const()[name = tensor("op_5392_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_5392_end_0 = const()[name = tensor("op_5392_end_0"), val = tensor([1, 1, 16, 32, 2])]; + tensor var_5392_end_mask_0 = const()[name = tensor("op_5392_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_5392_squeeze_mask_0 = const()[name = tensor("op_5392_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_5392_cast_fp16 = slice_by_index(begin = var_5392_begin_0, end = var_5392_end_0, end_mask = var_5392_end_mask_0, squeeze_mask = var_5392_squeeze_mask_0, x = q_complex_25_cast_fp16)[name = tensor("op_5392_cast_fp16")]; + tensor var_5400_begin_0 = const()[name = tensor("op_5400_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_5400_end_0 = const()[name = tensor("op_5400_end_0"), val = tensor([1, 1, 16, 32, 1])]; + tensor var_5400_end_mask_0 = const()[name = tensor("op_5400_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_5400_squeeze_mask_0 = const()[name = tensor("op_5400_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_5400_cast_fp16 = slice_by_index(begin = var_5400_begin_0, end = var_5400_end_0, end_mask = var_5400_end_mask_0, squeeze_mask = var_5400_squeeze_mask_0, x = k_complex_25_cast_fp16)[name = tensor("op_5400_cast_fp16")]; + tensor var_5408_begin_0 = const()[name = tensor("op_5408_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_5408_end_0 = const()[name = tensor("op_5408_end_0"), val = tensor([1, 1, 16, 32, 2])]; + tensor var_5408_end_mask_0 = const()[name = tensor("op_5408_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_5408_squeeze_mask_0 = const()[name = tensor("op_5408_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_5408_cast_fp16 = slice_by_index(begin = var_5408_begin_0, end = var_5408_end_0, end_mask = var_5408_end_mask_0, squeeze_mask = var_5408_squeeze_mask_0, x = k_complex_25_cast_fp16)[name = tensor("op_5408_cast_fp16")]; + tensor freqs_25_to_fp16 = const()[name = tensor("freqs_25_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(303692032)))]; + tensor var_5414_cast_fp16 = mul(x = freqs_25_to_fp16, y = ts_77_cast_fp16)[name = tensor("op_5414_cast_fp16")]; + tensor rotr_25_cast_fp16 = cos(x = var_5414_cast_fp16)[name = tensor("rotr_25_cast_fp16")]; + tensor roti_25_cast_fp16 = sin(x = var_5414_cast_fp16)[name = tensor("roti_25_cast_fp16")]; + tensor var_5418_cast_fp16 = mul(x = var_5384_cast_fp16, y = rotr_25_cast_fp16)[name = tensor("op_5418_cast_fp16")]; + tensor var_5419_cast_fp16 = mul(x = var_5392_cast_fp16, y = roti_25_cast_fp16)[name = tensor("op_5419_cast_fp16")]; + tensor qor_49_cast_fp16 = sub(x = var_5418_cast_fp16, y = var_5419_cast_fp16)[name = tensor("qor_49_cast_fp16")]; + tensor var_5422_cast_fp16 = mul(x = var_5384_cast_fp16, y = roti_25_cast_fp16)[name = tensor("op_5422_cast_fp16")]; + tensor var_5423_cast_fp16 = mul(x = var_5392_cast_fp16, y = rotr_25_cast_fp16)[name = tensor("op_5423_cast_fp16")]; + tensor qoi_49_cast_fp16 = add(x = var_5422_cast_fp16, y = var_5423_cast_fp16)[name = tensor("qoi_49_cast_fp16")]; + tensor var_5426_cast_fp16 = mul(x = var_5400_cast_fp16, y = rotr_25_cast_fp16)[name = tensor("op_5426_cast_fp16")]; + tensor var_5427_cast_fp16 = mul(x = var_5408_cast_fp16, y = roti_25_cast_fp16)[name = tensor("op_5427_cast_fp16")]; + tensor kor_49_cast_fp16 = sub(x = var_5426_cast_fp16, y = var_5427_cast_fp16)[name = tensor("kor_49_cast_fp16")]; + tensor var_5430_cast_fp16 = mul(x = var_5400_cast_fp16, y = roti_25_cast_fp16)[name = tensor("op_5430_cast_fp16")]; + tensor var_5431_cast_fp16 = mul(x = var_5408_cast_fp16, y = rotr_25_cast_fp16)[name = tensor("op_5431_cast_fp16")]; + tensor koi_49_cast_fp16 = add(x = var_5430_cast_fp16, y = var_5431_cast_fp16)[name = tensor("koi_49_cast_fp16")]; + tensor qo_25_axis_0 = const()[name = tensor("qo_25_axis_0"), val = tensor(-1)]; + tensor qo_25_cast_fp16 = stack(axis = qo_25_axis_0, values = (qor_49_cast_fp16, qoi_49_cast_fp16))[name = tensor("qo_25_cast_fp16")]; + tensor ko_25_axis_0 = const()[name = tensor("ko_25_axis_0"), val = tensor(-1)]; + tensor ko_25_cast_fp16 = stack(axis = ko_25_axis_0, values = (kor_49_cast_fp16, koi_49_cast_fp16))[name = tensor("ko_25_cast_fp16")]; + tensor var_5460 = const()[name = tensor("op_5460"), val = tensor([1, 1, 16, 64])]; + tensor q_75_cast_fp16 = reshape(shape = var_5460, x = qo_25_cast_fp16)[name = tensor("q_75_cast_fp16")]; + tensor var_5462 = const()[name = tensor("op_5462"), val = tensor([1, 1, 16, 64])]; + tensor k_51_cast_fp16 = reshape(shape = var_5462, x = ko_25_cast_fp16)[name = tensor("k_51_cast_fp16")]; + tensor _inversed_5484_y_0_to_fp16 = const()[name = tensor("_inversed_5484_y_0_to_fp16"), val = tensor(0x1p-9)]; + tensor _inversed_5484_cast_fp16 = mul(x = ts_77_cast_fp16, y = _inversed_5484_y_0_to_fp16)[name = tensor("_inversed_5484_cast_fp16")]; + tensor var_5485_cast_fp16 = floor(x = _inversed_5484_cast_fp16)[name = tensor("op_5485_cast_fp16")]; + tensor var_5486_to_fp16 = const()[name = tensor("op_5486_to_fp16"), val = tensor(0x1p+9)]; + tensor var_5487_cast_fp16 = mul(x = var_5485_cast_fp16, y = var_5486_to_fp16)[name = tensor("op_5487_cast_fp16")]; + tensor write_indices_float_51_cast_fp16 = sub(x = ts_77_cast_fp16, y = var_5487_cast_fp16)[name = tensor("write_indices_float_51_cast_fp16")]; + tensor var_5494_dtype_0 = const()[name = tensor("op_5494_dtype_0"), val = tensor("int32")]; + tensor write_indices_25_reps_0 = const()[name = tensor("write_indices_25_reps_0"), val = tensor([1, 1, 16, 64])]; + tensor write_indices_float_51_cast_fp16_to_int32 = cast(dtype = var_5494_dtype_0, x = write_indices_float_51_cast_fp16)[name = tensor("cast_492")]; + tensor write_indices_25 = tile(reps = write_indices_25_reps_0, x = write_indices_float_51_cast_fp16_to_int32)[name = tensor("write_indices_25")]; + tensor var_5502_begin_0 = const()[name = tensor("op_5502_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_5502_end_0 = const()[name = tensor("op_5502_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor var_5502_end_mask_0 = const()[name = tensor("op_5502_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_5502_squeeze_mask_0 = const()[name = tensor("op_5502_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor cache12_to_fp16_dtype_0 = const()[name = tensor("cache12_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor cache12_to_fp16 = cast(dtype = cache12_to_fp16_dtype_0, x = cache12)[name = tensor("cast_491")]; + tensor var_5502_cast_fp16 = slice_by_index(begin = var_5502_begin_0, end = var_5502_end_0, end_mask = var_5502_end_mask_0, squeeze_mask = var_5502_squeeze_mask_0, x = cache12_to_fp16)[name = tensor("op_5502_cast_fp16")]; + tensor var_5504_axis_0 = const()[name = tensor("op_5504_axis_0"), val = tensor(1)]; + tensor var_5504_mode_0 = const()[name = tensor("op_5504_mode_0"), val = tensor("update")]; + tensor var_5504_validate_indices_0 = const()[name = tensor("op_5504_validate_indices_0"), val = tensor(false)]; + tensor var_5504_cast_fp16 = scatter_along_axis(axis = var_5504_axis_0, data = var_5502_cast_fp16, indices = write_indices_25, mode = var_5504_mode_0, updates = k_51_cast_fp16, validate_indices = var_5504_validate_indices_0)[name = tensor("op_5504_cast_fp16")]; + tensor concat_86 = const()[name = tensor("concat_86"), val = tensor([0, 0, 0, 0, 0])]; + tensor concat_87 = const()[name = tensor("concat_87"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_25_internal_tensor_assign_1_stride_0 = const()[name = tensor("new_cache_25_internal_tensor_assign_1_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_25_internal_tensor_assign_1_begin_mask_0 = const()[name = tensor("new_cache_25_internal_tensor_assign_1_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_25_internal_tensor_assign_1_end_mask_0 = const()[name = tensor("new_cache_25_internal_tensor_assign_1_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_25_internal_tensor_assign_1_squeeze_mask_0 = const()[name = tensor("new_cache_25_internal_tensor_assign_1_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_72 = const()[name = tensor("shape_72"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_24 = const()[name = tensor("reduce_prod_24"), val = tensor(1048576)]; + tensor range_1d_24_start_0 = const()[name = tensor("range_1d_24_start_0"), val = tensor(0)]; + tensor range_1d_24_step_0 = const()[name = tensor("range_1d_24_step_0"), val = tensor(1)]; + tensor range_1d_24 = range_1d(end = reduce_prod_24, start = range_1d_24_start_0, step = range_1d_24_step_0)[name = tensor("range_1d_24")]; + tensor reshape_120 = reshape(shape = shape_72, x = range_1d_24)[name = tensor("reshape_120")]; + tensor slice_by_index_24 = slice_by_index(begin = concat_86, begin_mask = new_cache_25_internal_tensor_assign_1_begin_mask_0, end = concat_87, end_mask = new_cache_25_internal_tensor_assign_1_end_mask_0, squeeze_mask = new_cache_25_internal_tensor_assign_1_squeeze_mask_0, stride = new_cache_25_internal_tensor_assign_1_stride_0, x = reshape_120)[name = tensor("slice_by_index_24")]; + tensor reshape_121_shape_0 = const()[name = tensor("reshape_121_shape_0"), val = tensor([-1])]; + tensor reshape_121 = reshape(shape = reshape_121_shape_0, x = slice_by_index_24)[name = tensor("reshape_121")]; + tensor reshape_122_shape_0 = const()[name = tensor("reshape_122_shape_0"), val = tensor([-1])]; + tensor reshape_122_cast_fp16 = reshape(shape = reshape_122_shape_0, x = var_5504_cast_fp16)[name = tensor("reshape_122_cast_fp16")]; + tensor reshape_123_shape_0 = const()[name = tensor("reshape_123_shape_0"), val = tensor([-1])]; + tensor reshape_123_cast_fp16 = reshape(shape = reshape_123_shape_0, x = cache12_to_fp16)[name = tensor("reshape_123_cast_fp16")]; + tensor scatter_24_mode_0 = const()[name = tensor("scatter_24_mode_0"), val = tensor("update")]; + tensor scatter_24_axis_0 = const()[name = tensor("scatter_24_axis_0"), val = tensor(0)]; + tensor scatter_24_validate_indices_0 = const()[name = tensor("scatter_24_validate_indices_0"), val = tensor(false)]; + tensor scatter_24_cast_fp16 = scatter(axis = scatter_24_axis_0, data = reshape_123_cast_fp16, indices = reshape_121, mode = scatter_24_mode_0, updates = reshape_122_cast_fp16, validate_indices = scatter_24_validate_indices_0)[name = tensor("scatter_24_cast_fp16")]; + tensor reshape_124_cast_fp16 = reshape(shape = shape_72, x = scatter_24_cast_fp16)[name = tensor("reshape_124_cast_fp16")]; + tensor var_5512_begin_0 = const()[name = tensor("op_5512_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor var_5512_end_0 = const()[name = tensor("op_5512_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor var_5512_end_mask_0 = const()[name = tensor("op_5512_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_5512_squeeze_mask_0 = const()[name = tensor("op_5512_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor var_5512_cast_fp16 = slice_by_index(begin = var_5512_begin_0, end = var_5512_end_0, end_mask = var_5512_end_mask_0, squeeze_mask = var_5512_squeeze_mask_0, x = reshape_124_cast_fp16)[name = tensor("op_5512_cast_fp16")]; + tensor var_5514_axis_0 = const()[name = tensor("op_5514_axis_0"), val = tensor(1)]; + tensor var_5514_mode_0 = const()[name = tensor("op_5514_mode_0"), val = tensor("update")]; + tensor var_5514_validate_indices_0 = const()[name = tensor("op_5514_validate_indices_0"), val = tensor(false)]; + tensor var_5514_cast_fp16 = scatter_along_axis(axis = var_5514_axis_0, data = var_5512_cast_fp16, indices = write_indices_25, mode = var_5514_mode_0, updates = v_25_cast_fp16, validate_indices = var_5514_validate_indices_0)[name = tensor("op_5514_cast_fp16")]; + tensor concat_88 = const()[name = tensor("concat_88"), val = tensor([1, 0, 0, 0, 0])]; + tensor concat_89 = const()[name = tensor("concat_89"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_25_internal_tensor_assign_2_stride_0 = const()[name = tensor("new_cache_25_internal_tensor_assign_2_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_25_internal_tensor_assign_2_begin_mask_0 = const()[name = tensor("new_cache_25_internal_tensor_assign_2_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_25_internal_tensor_assign_2_end_mask_0 = const()[name = tensor("new_cache_25_internal_tensor_assign_2_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_25_internal_tensor_assign_2_squeeze_mask_0 = const()[name = tensor("new_cache_25_internal_tensor_assign_2_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_73 = const()[name = tensor("shape_73"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_25 = const()[name = tensor("reduce_prod_25"), val = tensor(1048576)]; + tensor range_1d_25_start_0 = const()[name = tensor("range_1d_25_start_0"), val = tensor(0)]; + tensor range_1d_25_step_0 = const()[name = tensor("range_1d_25_step_0"), val = tensor(1)]; + tensor range_1d_25 = range_1d(end = reduce_prod_25, start = range_1d_25_start_0, step = range_1d_25_step_0)[name = tensor("range_1d_25")]; + tensor reshape_125 = reshape(shape = shape_73, x = range_1d_25)[name = tensor("reshape_125")]; + tensor slice_by_index_25 = slice_by_index(begin = concat_88, begin_mask = new_cache_25_internal_tensor_assign_2_begin_mask_0, end = concat_89, end_mask = new_cache_25_internal_tensor_assign_2_end_mask_0, squeeze_mask = new_cache_25_internal_tensor_assign_2_squeeze_mask_0, stride = new_cache_25_internal_tensor_assign_2_stride_0, x = reshape_125)[name = tensor("slice_by_index_25")]; + tensor reshape_126_shape_0 = const()[name = tensor("reshape_126_shape_0"), val = tensor([-1])]; + tensor reshape_126 = reshape(shape = reshape_126_shape_0, x = slice_by_index_25)[name = tensor("reshape_126")]; + tensor reshape_127_shape_0 = const()[name = tensor("reshape_127_shape_0"), val = tensor([-1])]; + tensor reshape_127_cast_fp16 = reshape(shape = reshape_127_shape_0, x = var_5514_cast_fp16)[name = tensor("reshape_127_cast_fp16")]; + tensor reshape_128_shape_0 = const()[name = tensor("reshape_128_shape_0"), val = tensor([-1])]; + tensor reshape_128_cast_fp16 = reshape(shape = reshape_128_shape_0, x = reshape_124_cast_fp16)[name = tensor("reshape_128_cast_fp16")]; + tensor scatter_25_mode_0 = const()[name = tensor("scatter_25_mode_0"), val = tensor("update")]; + tensor scatter_25_axis_0 = const()[name = tensor("scatter_25_axis_0"), val = tensor(0)]; + tensor scatter_25_validate_indices_0 = const()[name = tensor("scatter_25_validate_indices_0"), val = tensor(false)]; + tensor scatter_25_cast_fp16 = scatter(axis = scatter_25_axis_0, data = reshape_128_cast_fp16, indices = reshape_126, mode = scatter_25_mode_0, updates = reshape_127_cast_fp16, validate_indices = scatter_25_validate_indices_0)[name = tensor("scatter_25_cast_fp16")]; + tensor reshape_129_cast_fp16 = reshape(shape = shape_73, x = scatter_25_cast_fp16)[name = tensor("reshape_129_cast_fp16")]; + tensor reshape_129_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("reshape_129_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor keys_73_begin_0 = const()[name = tensor("keys_73_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor keys_73_end_0 = const()[name = tensor("keys_73_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor keys_73_end_mask_0 = const()[name = tensor("keys_73_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor keys_73_squeeze_mask_0 = const()[name = tensor("keys_73_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor keys_73_cast_fp16 = slice_by_index(begin = keys_73_begin_0, end = keys_73_end_0, end_mask = keys_73_end_mask_0, squeeze_mask = keys_73_squeeze_mask_0, x = reshape_129_cast_fp16)[name = tensor("keys_73_cast_fp16")]; + tensor values_73_begin_0 = const()[name = tensor("values_73_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor values_73_end_0 = const()[name = tensor("values_73_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor values_73_end_mask_0 = const()[name = tensor("values_73_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor values_73_squeeze_mask_0 = const()[name = tensor("values_73_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor values_73_cast_fp16 = slice_by_index(begin = values_73_begin_0, end = values_73_end_0, end_mask = values_73_end_mask_0, squeeze_mask = values_73_squeeze_mask_0, x = reshape_129_cast_fp16)[name = tensor("values_73_cast_fp16")]; + tensor var_5526_cast_fp16 = not_equal(x = keys_73_cast_fp16, y = keys_73_cast_fp16)[name = tensor("op_5526_cast_fp16")]; + tensor keys_75_cast_fp16 = select(a = var_504_to_fp16, b = keys_73_cast_fp16, cond = var_5526_cast_fp16)[name = tensor("keys_75_cast_fp16")]; + tensor var_5534_cast_fp16 = not_equal(x = values_73_cast_fp16, y = values_73_cast_fp16)[name = tensor("op_5534_cast_fp16")]; + tensor values_75_cast_fp16 = select(a = var_504_to_fp16, b = values_73_cast_fp16, cond = var_5534_cast_fp16)[name = tensor("values_75_cast_fp16")]; + tensor var_5558 = const()[name = tensor("op_5558"), val = tensor([0, 2, 1, 3])]; + tensor var_5571 = const()[name = tensor("op_5571"), val = tensor([1, 1, 1])]; + tensor var_5572_cast_fp16 = reshape(shape = var_5571, x = position12_to_fp16)[name = tensor("op_5572_cast_fp16")]; + tensor var_5589_to_fp16 = const()[name = tensor("op_5589_to_fp16"), val = tensor(0x1p+0)]; + tensor valid_len_25_cast_fp16 = add(x = var_5572_cast_fp16, y = var_5589_to_fp16)[name = tensor("valid_len_25_cast_fp16")]; + tensor valid_mask_25_cast_fp16 = less(x = k_positions_1_promoted_to_fp16, y = valid_len_25_cast_fp16)[name = tensor("valid_mask_25_cast_fp16")]; + tensor causal_mask_25_cast_fp16 = less_equal(x = k_positions_1_promoted_to_fp16, y = var_5572_cast_fp16)[name = tensor("causal_mask_25_cast_fp16")]; + tensor attn_mask_49 = logical_and(x = valid_mask_25_cast_fp16, y = causal_mask_25_cast_fp16)[name = tensor("attn_mask_49")]; + tensor attn_mask_51_axes_0 = const()[name = tensor("attn_mask_51_axes_0"), val = tensor([1])]; + tensor attn_mask_51 = expand_dims(axes = attn_mask_51_axes_0, x = attn_mask_49)[name = tensor("attn_mask_51")]; + tensor var_5607_transpose_x_0 = const()[name = tensor("op_5607_transpose_x_0"), val = tensor(false)]; + tensor var_5607_transpose_y_0 = const()[name = tensor("op_5607_transpose_y_0"), val = tensor(false)]; + tensor transpose_96_perm_0 = const()[name = tensor("transpose_96_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_97_perm_0 = const()[name = tensor("transpose_97_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_97 = transpose(perm = transpose_97_perm_0, x = keys_75_cast_fp16)[name = tensor("transpose_165")]; + tensor transpose_96 = transpose(perm = transpose_96_perm_0, x = q_75_cast_fp16)[name = tensor("transpose_166")]; + tensor var_5607_cast_fp16 = matmul(transpose_x = var_5607_transpose_x_0, transpose_y = var_5607_transpose_y_0, x = transpose_96, y = transpose_97)[name = tensor("op_5607_cast_fp16")]; + tensor var_5601_to_fp16 = const()[name = tensor("op_5601_to_fp16"), val = tensor([0x1p-3])]; + tensor attn_weights_73_cast_fp16 = mul(x = var_5607_cast_fp16, y = var_5601_to_fp16)[name = tensor("attn_weights_73_cast_fp16")]; + tensor var_5609 = logical_not(x = attn_mask_51)[name = tensor("op_5609")]; + tensor var_5610_to_fp16 = const()[name = tensor("op_5610_to_fp16"), val = tensor(-inf)]; + tensor attn_weights_75_cast_fp16 = select(a = var_5610_to_fp16, b = attn_weights_73_cast_fp16, cond = var_5609)[name = tensor("attn_weights_75_cast_fp16")]; + tensor var_5612 = const()[name = tensor("op_5612"), val = tensor(-1)]; + tensor attn_weights_77_cast_fp16 = softmax(axis = var_5612, x = attn_weights_75_cast_fp16)[name = tensor("attn_weights_77_cast_fp16")]; + tensor attn_output_25_transpose_x_0 = const()[name = tensor("attn_output_25_transpose_x_0"), val = tensor(false)]; + tensor attn_output_25_transpose_y_0 = const()[name = tensor("attn_output_25_transpose_y_0"), val = tensor(false)]; + tensor values_77_cast_fp16 = transpose(perm = var_5558, x = values_75_cast_fp16)[name = tensor("transpose_167")]; + tensor attn_output_25_cast_fp16 = matmul(transpose_x = attn_output_25_transpose_x_0, transpose_y = attn_output_25_transpose_y_0, x = attn_weights_77_cast_fp16, y = values_77_cast_fp16)[name = tensor("attn_output_25_cast_fp16")]; + tensor var_5620 = const()[name = tensor("op_5620"), val = tensor([0, 2, 1, 3])]; + tensor var_5623 = const()[name = tensor("op_5623"), val = tensor([1, 1, 1024])]; + tensor var_5621_cast_fp16 = transpose(perm = var_5620, x = attn_output_25_cast_fp16)[name = tensor("transpose_164")]; + tensor input_125_cast_fp16 = reshape(shape = var_5623, x = var_5621_cast_fp16)[name = tensor("input_125_cast_fp16")]; + tensor linear_50_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = attn12_out_proj_weight_cast_fp16, x = input_125_cast_fp16)[name = tensor("linear_50_cast_fp16")]; + tensor var_5629_to_fp16 = const()[name = tensor("op_5629_to_fp16"), val = tensor(0x1p+0)]; + tensor var_5630_cast_fp16 = add(x = position12_to_fp16, y = var_5629_to_fp16)[name = tensor("op_5630_cast_fp16")]; + tensor var_5630_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_5630_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor input_127_cast_fp16 = add(x = input_123_cast_fp16, y = linear_50_cast_fp16)[name = tensor("input_127_cast_fp16")]; + tensor input_129_axes_0 = const()[name = tensor("input_129_axes_0"), val = tensor([-1])]; + tensor norm12_2_weight_to_fp16 = const()[name = tensor("norm12_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(303692160)))]; + tensor norm12_2_bias_to_fp16 = const()[name = tensor("norm12_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(303694272)))]; + tensor var_5634_to_fp16 = const()[name = tensor("op_5634_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_129_cast_fp16 = layer_norm(axes = input_129_axes_0, beta = norm12_2_bias_to_fp16, epsilon = var_5634_to_fp16, gamma = norm12_2_weight_to_fp16, x = input_127_cast_fp16)[name = tensor("input_129_cast_fp16")]; + tensor linear_51_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = linear12_1_weight_cast_fp16, x = input_129_cast_fp16)[name = tensor("linear_51_cast_fp16")]; + tensor input_131_mode_0 = const()[name = tensor("input_131_mode_0"), val = tensor("EXACT")]; + tensor input_131_cast_fp16 = gelu(mode = input_131_mode_0, x = linear_51_cast_fp16)[name = tensor("input_131_cast_fp16")]; + tensor linear_52_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = linear12_2_weight_cast_fp16, x = input_131_cast_fp16)[name = tensor("linear_52_cast_fp16")]; + tensor input_133_cast_fp16 = add(x = input_127_cast_fp16, y = linear_52_cast_fp16)[name = tensor("input_133_cast_fp16")]; + tensor x_27_axes_0 = const()[name = tensor("x_27_axes_0"), val = tensor([-1])]; + tensor norm13_1_weight_to_fp16 = const()[name = tensor("norm13_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(303696384)))]; + tensor norm13_1_bias_to_fp16 = const()[name = tensor("norm13_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(303698496)))]; + tensor var_5651_to_fp16 = const()[name = tensor("op_5651_to_fp16"), val = tensor(0x1.5p-17)]; + tensor x_27_cast_fp16 = layer_norm(axes = x_27_axes_0, beta = norm13_1_bias_to_fp16, epsilon = var_5651_to_fp16, gamma = norm13_1_weight_to_fp16, x = input_133_cast_fp16)[name = tensor("x_27_cast_fp16")]; + tensor linear_53_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = attn13_in_proj_weight_cast_fp16, x = x_27_cast_fp16)[name = tensor("linear_53_cast_fp16")]; + tensor var_5687 = const()[name = tensor("op_5687"), val = tensor([1, 1, 3, 16, 64])]; + tensor qkv_27_cast_fp16 = reshape(shape = var_5687, x = linear_53_cast_fp16)[name = tensor("qkv_27_cast_fp16")]; + tensor q_79_begin_0 = const()[name = tensor("q_79_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor q_79_end_0 = const()[name = tensor("q_79_end_0"), val = tensor([1, 1, 1, 16, 64])]; + tensor q_79_end_mask_0 = const()[name = tensor("q_79_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor q_79_squeeze_mask_0 = const()[name = tensor("q_79_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor q_79_cast_fp16 = slice_by_index(begin = q_79_begin_0, end = q_79_end_0, end_mask = q_79_end_mask_0, squeeze_mask = q_79_squeeze_mask_0, x = qkv_27_cast_fp16)[name = tensor("q_79_cast_fp16")]; + tensor k_53_begin_0 = const()[name = tensor("k_53_begin_0"), val = tensor([0, 0, 1, 0, 0])]; + tensor k_53_end_0 = const()[name = tensor("k_53_end_0"), val = tensor([1, 1, 2, 16, 64])]; + tensor k_53_end_mask_0 = const()[name = tensor("k_53_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor k_53_squeeze_mask_0 = const()[name = tensor("k_53_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor k_53_cast_fp16 = slice_by_index(begin = k_53_begin_0, end = k_53_end_0, end_mask = k_53_end_mask_0, squeeze_mask = k_53_squeeze_mask_0, x = qkv_27_cast_fp16)[name = tensor("k_53_cast_fp16")]; + tensor v_27_begin_0 = const()[name = tensor("v_27_begin_0"), val = tensor([0, 0, 2, 0, 0])]; + tensor v_27_end_0 = const()[name = tensor("v_27_end_0"), val = tensor([1, 1, 3, 16, 64])]; + tensor v_27_end_mask_0 = const()[name = tensor("v_27_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor v_27_squeeze_mask_0 = const()[name = tensor("v_27_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor v_27_cast_fp16 = slice_by_index(begin = v_27_begin_0, end = v_27_end_0, end_mask = v_27_end_mask_0, squeeze_mask = v_27_squeeze_mask_0, x = qkv_27_cast_fp16)[name = tensor("v_27_cast_fp16")]; + tensor var_5791 = const()[name = tensor("op_5791"), val = tensor([1, 1, 1, 1])]; + tensor position13_to_fp16_dtype_0 = const()[name = tensor("position13_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor position13_to_fp16 = cast(dtype = position13_to_fp16_dtype_0, x = position13)[name = tensor("cast_488")]; + tensor ts_83_cast_fp16 = reshape(shape = var_5791, x = position13_to_fp16)[name = tensor("ts_83_cast_fp16")]; + tensor var_5795 = const()[name = tensor("op_5795"), val = tensor([1, 1, 16, 32, 2])]; + tensor q_complex_27_cast_fp16 = reshape(shape = var_5795, x = q_79_cast_fp16)[name = tensor("q_complex_27_cast_fp16")]; + tensor var_5799 = const()[name = tensor("op_5799"), val = tensor([1, 1, 16, 32, 2])]; + tensor k_complex_27_cast_fp16 = reshape(shape = var_5799, x = k_53_cast_fp16)[name = tensor("k_complex_27_cast_fp16")]; + tensor var_5803_begin_0 = const()[name = tensor("op_5803_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_5803_end_0 = const()[name = tensor("op_5803_end_0"), val = tensor([1, 1, 16, 32, 1])]; + tensor var_5803_end_mask_0 = const()[name = tensor("op_5803_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_5803_squeeze_mask_0 = const()[name = tensor("op_5803_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_5803_cast_fp16 = slice_by_index(begin = var_5803_begin_0, end = var_5803_end_0, end_mask = var_5803_end_mask_0, squeeze_mask = var_5803_squeeze_mask_0, x = q_complex_27_cast_fp16)[name = tensor("op_5803_cast_fp16")]; + tensor var_5811_begin_0 = const()[name = tensor("op_5811_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_5811_end_0 = const()[name = tensor("op_5811_end_0"), val = tensor([1, 1, 16, 32, 2])]; + tensor var_5811_end_mask_0 = const()[name = tensor("op_5811_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_5811_squeeze_mask_0 = const()[name = tensor("op_5811_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_5811_cast_fp16 = slice_by_index(begin = var_5811_begin_0, end = var_5811_end_0, end_mask = var_5811_end_mask_0, squeeze_mask = var_5811_squeeze_mask_0, x = q_complex_27_cast_fp16)[name = tensor("op_5811_cast_fp16")]; + tensor var_5819_begin_0 = const()[name = tensor("op_5819_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_5819_end_0 = const()[name = tensor("op_5819_end_0"), val = tensor([1, 1, 16, 32, 1])]; + tensor var_5819_end_mask_0 = const()[name = tensor("op_5819_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_5819_squeeze_mask_0 = const()[name = tensor("op_5819_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_5819_cast_fp16 = slice_by_index(begin = var_5819_begin_0, end = var_5819_end_0, end_mask = var_5819_end_mask_0, squeeze_mask = var_5819_squeeze_mask_0, x = k_complex_27_cast_fp16)[name = tensor("op_5819_cast_fp16")]; + tensor var_5827_begin_0 = const()[name = tensor("op_5827_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_5827_end_0 = const()[name = tensor("op_5827_end_0"), val = tensor([1, 1, 16, 32, 2])]; + tensor var_5827_end_mask_0 = const()[name = tensor("op_5827_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_5827_squeeze_mask_0 = const()[name = tensor("op_5827_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_5827_cast_fp16 = slice_by_index(begin = var_5827_begin_0, end = var_5827_end_0, end_mask = var_5827_end_mask_0, squeeze_mask = var_5827_squeeze_mask_0, x = k_complex_27_cast_fp16)[name = tensor("op_5827_cast_fp16")]; + tensor freqs_27_to_fp16 = const()[name = tensor("freqs_27_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(303700608)))]; + tensor var_5833_cast_fp16 = mul(x = freqs_27_to_fp16, y = ts_83_cast_fp16)[name = tensor("op_5833_cast_fp16")]; + tensor rotr_27_cast_fp16 = cos(x = var_5833_cast_fp16)[name = tensor("rotr_27_cast_fp16")]; + tensor roti_27_cast_fp16 = sin(x = var_5833_cast_fp16)[name = tensor("roti_27_cast_fp16")]; + tensor var_5837_cast_fp16 = mul(x = var_5803_cast_fp16, y = rotr_27_cast_fp16)[name = tensor("op_5837_cast_fp16")]; + tensor var_5838_cast_fp16 = mul(x = var_5811_cast_fp16, y = roti_27_cast_fp16)[name = tensor("op_5838_cast_fp16")]; + tensor qor_53_cast_fp16 = sub(x = var_5837_cast_fp16, y = var_5838_cast_fp16)[name = tensor("qor_53_cast_fp16")]; + tensor var_5841_cast_fp16 = mul(x = var_5803_cast_fp16, y = roti_27_cast_fp16)[name = tensor("op_5841_cast_fp16")]; + tensor var_5842_cast_fp16 = mul(x = var_5811_cast_fp16, y = rotr_27_cast_fp16)[name = tensor("op_5842_cast_fp16")]; + tensor qoi_53_cast_fp16 = add(x = var_5841_cast_fp16, y = var_5842_cast_fp16)[name = tensor("qoi_53_cast_fp16")]; + tensor var_5845_cast_fp16 = mul(x = var_5819_cast_fp16, y = rotr_27_cast_fp16)[name = tensor("op_5845_cast_fp16")]; + tensor var_5846_cast_fp16 = mul(x = var_5827_cast_fp16, y = roti_27_cast_fp16)[name = tensor("op_5846_cast_fp16")]; + tensor kor_53_cast_fp16 = sub(x = var_5845_cast_fp16, y = var_5846_cast_fp16)[name = tensor("kor_53_cast_fp16")]; + tensor var_5849_cast_fp16 = mul(x = var_5819_cast_fp16, y = roti_27_cast_fp16)[name = tensor("op_5849_cast_fp16")]; + tensor var_5850_cast_fp16 = mul(x = var_5827_cast_fp16, y = rotr_27_cast_fp16)[name = tensor("op_5850_cast_fp16")]; + tensor koi_53_cast_fp16 = add(x = var_5849_cast_fp16, y = var_5850_cast_fp16)[name = tensor("koi_53_cast_fp16")]; + tensor qo_27_axis_0 = const()[name = tensor("qo_27_axis_0"), val = tensor(-1)]; + tensor qo_27_cast_fp16 = stack(axis = qo_27_axis_0, values = (qor_53_cast_fp16, qoi_53_cast_fp16))[name = tensor("qo_27_cast_fp16")]; + tensor ko_27_axis_0 = const()[name = tensor("ko_27_axis_0"), val = tensor(-1)]; + tensor ko_27_cast_fp16 = stack(axis = ko_27_axis_0, values = (kor_53_cast_fp16, koi_53_cast_fp16))[name = tensor("ko_27_cast_fp16")]; + tensor var_5879 = const()[name = tensor("op_5879"), val = tensor([1, 1, 16, 64])]; + tensor q_81_cast_fp16 = reshape(shape = var_5879, x = qo_27_cast_fp16)[name = tensor("q_81_cast_fp16")]; + tensor var_5881 = const()[name = tensor("op_5881"), val = tensor([1, 1, 16, 64])]; + tensor k_55_cast_fp16 = reshape(shape = var_5881, x = ko_27_cast_fp16)[name = tensor("k_55_cast_fp16")]; + tensor _inversed_5903_y_0_to_fp16 = const()[name = tensor("_inversed_5903_y_0_to_fp16"), val = tensor(0x1p-9)]; + tensor _inversed_5903_cast_fp16 = mul(x = ts_83_cast_fp16, y = _inversed_5903_y_0_to_fp16)[name = tensor("_inversed_5903_cast_fp16")]; + tensor var_5904_cast_fp16 = floor(x = _inversed_5903_cast_fp16)[name = tensor("op_5904_cast_fp16")]; + tensor var_5905_to_fp16 = const()[name = tensor("op_5905_to_fp16"), val = tensor(0x1p+9)]; + tensor var_5906_cast_fp16 = mul(x = var_5904_cast_fp16, y = var_5905_to_fp16)[name = tensor("op_5906_cast_fp16")]; + tensor write_indices_float_55_cast_fp16 = sub(x = ts_83_cast_fp16, y = var_5906_cast_fp16)[name = tensor("write_indices_float_55_cast_fp16")]; + tensor var_5913_dtype_0 = const()[name = tensor("op_5913_dtype_0"), val = tensor("int32")]; + tensor write_indices_27_reps_0 = const()[name = tensor("write_indices_27_reps_0"), val = tensor([1, 1, 16, 64])]; + tensor write_indices_float_55_cast_fp16_to_int32 = cast(dtype = var_5913_dtype_0, x = write_indices_float_55_cast_fp16)[name = tensor("cast_487")]; + tensor write_indices_27 = tile(reps = write_indices_27_reps_0, x = write_indices_float_55_cast_fp16_to_int32)[name = tensor("write_indices_27")]; + tensor var_5921_begin_0 = const()[name = tensor("op_5921_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_5921_end_0 = const()[name = tensor("op_5921_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor var_5921_end_mask_0 = const()[name = tensor("op_5921_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_5921_squeeze_mask_0 = const()[name = tensor("op_5921_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor cache13_to_fp16_dtype_0 = const()[name = tensor("cache13_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor cache13_to_fp16 = cast(dtype = cache13_to_fp16_dtype_0, x = cache13)[name = tensor("cast_486")]; + tensor var_5921_cast_fp16 = slice_by_index(begin = var_5921_begin_0, end = var_5921_end_0, end_mask = var_5921_end_mask_0, squeeze_mask = var_5921_squeeze_mask_0, x = cache13_to_fp16)[name = tensor("op_5921_cast_fp16")]; + tensor var_5923_axis_0 = const()[name = tensor("op_5923_axis_0"), val = tensor(1)]; + tensor var_5923_mode_0 = const()[name = tensor("op_5923_mode_0"), val = tensor("update")]; + tensor var_5923_validate_indices_0 = const()[name = tensor("op_5923_validate_indices_0"), val = tensor(false)]; + tensor var_5923_cast_fp16 = scatter_along_axis(axis = var_5923_axis_0, data = var_5921_cast_fp16, indices = write_indices_27, mode = var_5923_mode_0, updates = k_55_cast_fp16, validate_indices = var_5923_validate_indices_0)[name = tensor("op_5923_cast_fp16")]; + tensor concat_93 = const()[name = tensor("concat_93"), val = tensor([0, 0, 0, 0, 0])]; + tensor concat_94 = const()[name = tensor("concat_94"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_27_internal_tensor_assign_1_stride_0 = const()[name = tensor("new_cache_27_internal_tensor_assign_1_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_27_internal_tensor_assign_1_begin_mask_0 = const()[name = tensor("new_cache_27_internal_tensor_assign_1_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_27_internal_tensor_assign_1_end_mask_0 = const()[name = tensor("new_cache_27_internal_tensor_assign_1_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_27_internal_tensor_assign_1_squeeze_mask_0 = const()[name = tensor("new_cache_27_internal_tensor_assign_1_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_74 = const()[name = tensor("shape_74"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_26 = const()[name = tensor("reduce_prod_26"), val = tensor(1048576)]; + tensor range_1d_26_start_0 = const()[name = tensor("range_1d_26_start_0"), val = tensor(0)]; + tensor range_1d_26_step_0 = const()[name = tensor("range_1d_26_step_0"), val = tensor(1)]; + tensor range_1d_26 = range_1d(end = reduce_prod_26, start = range_1d_26_start_0, step = range_1d_26_step_0)[name = tensor("range_1d_26")]; + tensor reshape_130 = reshape(shape = shape_74, x = range_1d_26)[name = tensor("reshape_130")]; + tensor slice_by_index_26 = slice_by_index(begin = concat_93, begin_mask = new_cache_27_internal_tensor_assign_1_begin_mask_0, end = concat_94, end_mask = new_cache_27_internal_tensor_assign_1_end_mask_0, squeeze_mask = new_cache_27_internal_tensor_assign_1_squeeze_mask_0, stride = new_cache_27_internal_tensor_assign_1_stride_0, x = reshape_130)[name = tensor("slice_by_index_26")]; + tensor reshape_131_shape_0 = const()[name = tensor("reshape_131_shape_0"), val = tensor([-1])]; + tensor reshape_131 = reshape(shape = reshape_131_shape_0, x = slice_by_index_26)[name = tensor("reshape_131")]; + tensor reshape_132_shape_0 = const()[name = tensor("reshape_132_shape_0"), val = tensor([-1])]; + tensor reshape_132_cast_fp16 = reshape(shape = reshape_132_shape_0, x = var_5923_cast_fp16)[name = tensor("reshape_132_cast_fp16")]; + tensor reshape_133_shape_0 = const()[name = tensor("reshape_133_shape_0"), val = tensor([-1])]; + tensor reshape_133_cast_fp16 = reshape(shape = reshape_133_shape_0, x = cache13_to_fp16)[name = tensor("reshape_133_cast_fp16")]; + tensor scatter_26_mode_0 = const()[name = tensor("scatter_26_mode_0"), val = tensor("update")]; + tensor scatter_26_axis_0 = const()[name = tensor("scatter_26_axis_0"), val = tensor(0)]; + tensor scatter_26_validate_indices_0 = const()[name = tensor("scatter_26_validate_indices_0"), val = tensor(false)]; + tensor scatter_26_cast_fp16 = scatter(axis = scatter_26_axis_0, data = reshape_133_cast_fp16, indices = reshape_131, mode = scatter_26_mode_0, updates = reshape_132_cast_fp16, validate_indices = scatter_26_validate_indices_0)[name = tensor("scatter_26_cast_fp16")]; + tensor reshape_134_cast_fp16 = reshape(shape = shape_74, x = scatter_26_cast_fp16)[name = tensor("reshape_134_cast_fp16")]; + tensor var_5931_begin_0 = const()[name = tensor("op_5931_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor var_5931_end_0 = const()[name = tensor("op_5931_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor var_5931_end_mask_0 = const()[name = tensor("op_5931_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_5931_squeeze_mask_0 = const()[name = tensor("op_5931_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor var_5931_cast_fp16 = slice_by_index(begin = var_5931_begin_0, end = var_5931_end_0, end_mask = var_5931_end_mask_0, squeeze_mask = var_5931_squeeze_mask_0, x = reshape_134_cast_fp16)[name = tensor("op_5931_cast_fp16")]; + tensor var_5933_axis_0 = const()[name = tensor("op_5933_axis_0"), val = tensor(1)]; + tensor var_5933_mode_0 = const()[name = tensor("op_5933_mode_0"), val = tensor("update")]; + tensor var_5933_validate_indices_0 = const()[name = tensor("op_5933_validate_indices_0"), val = tensor(false)]; + tensor var_5933_cast_fp16 = scatter_along_axis(axis = var_5933_axis_0, data = var_5931_cast_fp16, indices = write_indices_27, mode = var_5933_mode_0, updates = v_27_cast_fp16, validate_indices = var_5933_validate_indices_0)[name = tensor("op_5933_cast_fp16")]; + tensor concat_95 = const()[name = tensor("concat_95"), val = tensor([1, 0, 0, 0, 0])]; + tensor concat_96 = const()[name = tensor("concat_96"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_27_internal_tensor_assign_2_stride_0 = const()[name = tensor("new_cache_27_internal_tensor_assign_2_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_27_internal_tensor_assign_2_begin_mask_0 = const()[name = tensor("new_cache_27_internal_tensor_assign_2_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_27_internal_tensor_assign_2_end_mask_0 = const()[name = tensor("new_cache_27_internal_tensor_assign_2_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_27_internal_tensor_assign_2_squeeze_mask_0 = const()[name = tensor("new_cache_27_internal_tensor_assign_2_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_75 = const()[name = tensor("shape_75"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_27 = const()[name = tensor("reduce_prod_27"), val = tensor(1048576)]; + tensor range_1d_27_start_0 = const()[name = tensor("range_1d_27_start_0"), val = tensor(0)]; + tensor range_1d_27_step_0 = const()[name = tensor("range_1d_27_step_0"), val = tensor(1)]; + tensor range_1d_27 = range_1d(end = reduce_prod_27, start = range_1d_27_start_0, step = range_1d_27_step_0)[name = tensor("range_1d_27")]; + tensor reshape_135 = reshape(shape = shape_75, x = range_1d_27)[name = tensor("reshape_135")]; + tensor slice_by_index_27 = slice_by_index(begin = concat_95, begin_mask = new_cache_27_internal_tensor_assign_2_begin_mask_0, end = concat_96, end_mask = new_cache_27_internal_tensor_assign_2_end_mask_0, squeeze_mask = new_cache_27_internal_tensor_assign_2_squeeze_mask_0, stride = new_cache_27_internal_tensor_assign_2_stride_0, x = reshape_135)[name = tensor("slice_by_index_27")]; + tensor reshape_136_shape_0 = const()[name = tensor("reshape_136_shape_0"), val = tensor([-1])]; + tensor reshape_136 = reshape(shape = reshape_136_shape_0, x = slice_by_index_27)[name = tensor("reshape_136")]; + tensor reshape_137_shape_0 = const()[name = tensor("reshape_137_shape_0"), val = tensor([-1])]; + tensor reshape_137_cast_fp16 = reshape(shape = reshape_137_shape_0, x = var_5933_cast_fp16)[name = tensor("reshape_137_cast_fp16")]; + tensor reshape_138_shape_0 = const()[name = tensor("reshape_138_shape_0"), val = tensor([-1])]; + tensor reshape_138_cast_fp16 = reshape(shape = reshape_138_shape_0, x = reshape_134_cast_fp16)[name = tensor("reshape_138_cast_fp16")]; + tensor scatter_27_mode_0 = const()[name = tensor("scatter_27_mode_0"), val = tensor("update")]; + tensor scatter_27_axis_0 = const()[name = tensor("scatter_27_axis_0"), val = tensor(0)]; + tensor scatter_27_validate_indices_0 = const()[name = tensor("scatter_27_validate_indices_0"), val = tensor(false)]; + tensor scatter_27_cast_fp16 = scatter(axis = scatter_27_axis_0, data = reshape_138_cast_fp16, indices = reshape_136, mode = scatter_27_mode_0, updates = reshape_137_cast_fp16, validate_indices = scatter_27_validate_indices_0)[name = tensor("scatter_27_cast_fp16")]; + tensor reshape_139_cast_fp16 = reshape(shape = shape_75, x = scatter_27_cast_fp16)[name = tensor("reshape_139_cast_fp16")]; + tensor reshape_139_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("reshape_139_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor keys_79_begin_0 = const()[name = tensor("keys_79_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor keys_79_end_0 = const()[name = tensor("keys_79_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor keys_79_end_mask_0 = const()[name = tensor("keys_79_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor keys_79_squeeze_mask_0 = const()[name = tensor("keys_79_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor keys_79_cast_fp16 = slice_by_index(begin = keys_79_begin_0, end = keys_79_end_0, end_mask = keys_79_end_mask_0, squeeze_mask = keys_79_squeeze_mask_0, x = reshape_139_cast_fp16)[name = tensor("keys_79_cast_fp16")]; + tensor values_79_begin_0 = const()[name = tensor("values_79_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor values_79_end_0 = const()[name = tensor("values_79_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor values_79_end_mask_0 = const()[name = tensor("values_79_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor values_79_squeeze_mask_0 = const()[name = tensor("values_79_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor values_79_cast_fp16 = slice_by_index(begin = values_79_begin_0, end = values_79_end_0, end_mask = values_79_end_mask_0, squeeze_mask = values_79_squeeze_mask_0, x = reshape_139_cast_fp16)[name = tensor("values_79_cast_fp16")]; + tensor var_5945_cast_fp16 = not_equal(x = keys_79_cast_fp16, y = keys_79_cast_fp16)[name = tensor("op_5945_cast_fp16")]; + tensor keys_81_cast_fp16 = select(a = var_504_to_fp16, b = keys_79_cast_fp16, cond = var_5945_cast_fp16)[name = tensor("keys_81_cast_fp16")]; + tensor var_5953_cast_fp16 = not_equal(x = values_79_cast_fp16, y = values_79_cast_fp16)[name = tensor("op_5953_cast_fp16")]; + tensor values_81_cast_fp16 = select(a = var_504_to_fp16, b = values_79_cast_fp16, cond = var_5953_cast_fp16)[name = tensor("values_81_cast_fp16")]; + tensor var_5977 = const()[name = tensor("op_5977"), val = tensor([0, 2, 1, 3])]; + tensor var_5990 = const()[name = tensor("op_5990"), val = tensor([1, 1, 1])]; + tensor var_5991_cast_fp16 = reshape(shape = var_5990, x = position13_to_fp16)[name = tensor("op_5991_cast_fp16")]; + tensor var_6008_to_fp16 = const()[name = tensor("op_6008_to_fp16"), val = tensor(0x1p+0)]; + tensor valid_len_27_cast_fp16 = add(x = var_5991_cast_fp16, y = var_6008_to_fp16)[name = tensor("valid_len_27_cast_fp16")]; + tensor valid_mask_27_cast_fp16 = less(x = k_positions_1_promoted_to_fp16, y = valid_len_27_cast_fp16)[name = tensor("valid_mask_27_cast_fp16")]; + tensor causal_mask_27_cast_fp16 = less_equal(x = k_positions_1_promoted_to_fp16, y = var_5991_cast_fp16)[name = tensor("causal_mask_27_cast_fp16")]; + tensor attn_mask_53 = logical_and(x = valid_mask_27_cast_fp16, y = causal_mask_27_cast_fp16)[name = tensor("attn_mask_53")]; + tensor attn_mask_55_axes_0 = const()[name = tensor("attn_mask_55_axes_0"), val = tensor([1])]; + tensor attn_mask_55 = expand_dims(axes = attn_mask_55_axes_0, x = attn_mask_53)[name = tensor("attn_mask_55")]; + tensor var_6026_transpose_x_0 = const()[name = tensor("op_6026_transpose_x_0"), val = tensor(false)]; + tensor var_6026_transpose_y_0 = const()[name = tensor("op_6026_transpose_y_0"), val = tensor(false)]; + tensor transpose_98_perm_0 = const()[name = tensor("transpose_98_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_99_perm_0 = const()[name = tensor("transpose_99_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_99 = transpose(perm = transpose_99_perm_0, x = keys_81_cast_fp16)[name = tensor("transpose_161")]; + tensor transpose_98 = transpose(perm = transpose_98_perm_0, x = q_81_cast_fp16)[name = tensor("transpose_162")]; + tensor var_6026_cast_fp16 = matmul(transpose_x = var_6026_transpose_x_0, transpose_y = var_6026_transpose_y_0, x = transpose_98, y = transpose_99)[name = tensor("op_6026_cast_fp16")]; + tensor var_6020_to_fp16 = const()[name = tensor("op_6020_to_fp16"), val = tensor([0x1p-3])]; + tensor attn_weights_79_cast_fp16 = mul(x = var_6026_cast_fp16, y = var_6020_to_fp16)[name = tensor("attn_weights_79_cast_fp16")]; + tensor var_6028 = logical_not(x = attn_mask_55)[name = tensor("op_6028")]; + tensor var_6029_to_fp16 = const()[name = tensor("op_6029_to_fp16"), val = tensor(-inf)]; + tensor attn_weights_81_cast_fp16 = select(a = var_6029_to_fp16, b = attn_weights_79_cast_fp16, cond = var_6028)[name = tensor("attn_weights_81_cast_fp16")]; + tensor var_6031 = const()[name = tensor("op_6031"), val = tensor(-1)]; + tensor attn_weights_83_cast_fp16 = softmax(axis = var_6031, x = attn_weights_81_cast_fp16)[name = tensor("attn_weights_83_cast_fp16")]; + tensor attn_output_27_transpose_x_0 = const()[name = tensor("attn_output_27_transpose_x_0"), val = tensor(false)]; + tensor attn_output_27_transpose_y_0 = const()[name = tensor("attn_output_27_transpose_y_0"), val = tensor(false)]; + tensor values_83_cast_fp16 = transpose(perm = var_5977, x = values_81_cast_fp16)[name = tensor("transpose_163")]; + tensor attn_output_27_cast_fp16 = matmul(transpose_x = attn_output_27_transpose_x_0, transpose_y = attn_output_27_transpose_y_0, x = attn_weights_83_cast_fp16, y = values_83_cast_fp16)[name = tensor("attn_output_27_cast_fp16")]; + tensor var_6039 = const()[name = tensor("op_6039"), val = tensor([0, 2, 1, 3])]; + tensor var_6042 = const()[name = tensor("op_6042"), val = tensor([1, 1, 1024])]; + tensor var_6040_cast_fp16 = transpose(perm = var_6039, x = attn_output_27_cast_fp16)[name = tensor("transpose_160")]; + tensor input_135_cast_fp16 = reshape(shape = var_6042, x = var_6040_cast_fp16)[name = tensor("input_135_cast_fp16")]; + tensor linear_54_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = attn13_out_proj_weight_cast_fp16, x = input_135_cast_fp16)[name = tensor("linear_54_cast_fp16")]; + tensor var_6048_to_fp16 = const()[name = tensor("op_6048_to_fp16"), val = tensor(0x1p+0)]; + tensor var_6049_cast_fp16 = add(x = position13_to_fp16, y = var_6048_to_fp16)[name = tensor("op_6049_cast_fp16")]; + tensor var_6049_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_6049_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor input_137_cast_fp16 = add(x = input_133_cast_fp16, y = linear_54_cast_fp16)[name = tensor("input_137_cast_fp16")]; + tensor input_139_axes_0 = const()[name = tensor("input_139_axes_0"), val = tensor([-1])]; + tensor norm13_2_weight_to_fp16 = const()[name = tensor("norm13_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(303700736)))]; + tensor norm13_2_bias_to_fp16 = const()[name = tensor("norm13_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(303702848)))]; + tensor var_6053_to_fp16 = const()[name = tensor("op_6053_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_139_cast_fp16 = layer_norm(axes = input_139_axes_0, beta = norm13_2_bias_to_fp16, epsilon = var_6053_to_fp16, gamma = norm13_2_weight_to_fp16, x = input_137_cast_fp16)[name = tensor("input_139_cast_fp16")]; + tensor linear_55_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = linear13_1_weight_cast_fp16, x = input_139_cast_fp16)[name = tensor("linear_55_cast_fp16")]; + tensor input_141_mode_0 = const()[name = tensor("input_141_mode_0"), val = tensor("EXACT")]; + tensor input_141_cast_fp16 = gelu(mode = input_141_mode_0, x = linear_55_cast_fp16)[name = tensor("input_141_cast_fp16")]; + tensor linear_56_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = linear13_2_weight_cast_fp16, x = input_141_cast_fp16)[name = tensor("linear_56_cast_fp16")]; + tensor input_143_cast_fp16 = add(x = input_137_cast_fp16, y = linear_56_cast_fp16)[name = tensor("input_143_cast_fp16")]; + tensor x_29_axes_0 = const()[name = tensor("x_29_axes_0"), val = tensor([-1])]; + tensor norm14_1_weight_to_fp16 = const()[name = tensor("norm14_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(303704960)))]; + tensor norm14_1_bias_to_fp16 = const()[name = tensor("norm14_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(303707072)))]; + tensor var_6070_to_fp16 = const()[name = tensor("op_6070_to_fp16"), val = tensor(0x1.5p-17)]; + tensor x_29_cast_fp16 = layer_norm(axes = x_29_axes_0, beta = norm14_1_bias_to_fp16, epsilon = var_6070_to_fp16, gamma = norm14_1_weight_to_fp16, x = input_143_cast_fp16)[name = tensor("x_29_cast_fp16")]; + tensor linear_57_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = attn14_in_proj_weight_cast_fp16, x = x_29_cast_fp16)[name = tensor("linear_57_cast_fp16")]; + tensor var_6106 = const()[name = tensor("op_6106"), val = tensor([1, 1, 3, 16, 64])]; + tensor qkv_29_cast_fp16 = reshape(shape = var_6106, x = linear_57_cast_fp16)[name = tensor("qkv_29_cast_fp16")]; + tensor q_85_begin_0 = const()[name = tensor("q_85_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor q_85_end_0 = const()[name = tensor("q_85_end_0"), val = tensor([1, 1, 1, 16, 64])]; + tensor q_85_end_mask_0 = const()[name = tensor("q_85_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor q_85_squeeze_mask_0 = const()[name = tensor("q_85_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor q_85_cast_fp16 = slice_by_index(begin = q_85_begin_0, end = q_85_end_0, end_mask = q_85_end_mask_0, squeeze_mask = q_85_squeeze_mask_0, x = qkv_29_cast_fp16)[name = tensor("q_85_cast_fp16")]; + tensor k_57_begin_0 = const()[name = tensor("k_57_begin_0"), val = tensor([0, 0, 1, 0, 0])]; + tensor k_57_end_0 = const()[name = tensor("k_57_end_0"), val = tensor([1, 1, 2, 16, 64])]; + tensor k_57_end_mask_0 = const()[name = tensor("k_57_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor k_57_squeeze_mask_0 = const()[name = tensor("k_57_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor k_57_cast_fp16 = slice_by_index(begin = k_57_begin_0, end = k_57_end_0, end_mask = k_57_end_mask_0, squeeze_mask = k_57_squeeze_mask_0, x = qkv_29_cast_fp16)[name = tensor("k_57_cast_fp16")]; + tensor v_29_begin_0 = const()[name = tensor("v_29_begin_0"), val = tensor([0, 0, 2, 0, 0])]; + tensor v_29_end_0 = const()[name = tensor("v_29_end_0"), val = tensor([1, 1, 3, 16, 64])]; + tensor v_29_end_mask_0 = const()[name = tensor("v_29_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor v_29_squeeze_mask_0 = const()[name = tensor("v_29_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor v_29_cast_fp16 = slice_by_index(begin = v_29_begin_0, end = v_29_end_0, end_mask = v_29_end_mask_0, squeeze_mask = v_29_squeeze_mask_0, x = qkv_29_cast_fp16)[name = tensor("v_29_cast_fp16")]; + tensor var_6210 = const()[name = tensor("op_6210"), val = tensor([1, 1, 1, 1])]; + tensor position14_to_fp16_dtype_0 = const()[name = tensor("position14_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor position14_to_fp16 = cast(dtype = position14_to_fp16_dtype_0, x = position14)[name = tensor("cast_483")]; + tensor ts_89_cast_fp16 = reshape(shape = var_6210, x = position14_to_fp16)[name = tensor("ts_89_cast_fp16")]; + tensor var_6214 = const()[name = tensor("op_6214"), val = tensor([1, 1, 16, 32, 2])]; + tensor q_complex_29_cast_fp16 = reshape(shape = var_6214, x = q_85_cast_fp16)[name = tensor("q_complex_29_cast_fp16")]; + tensor var_6218 = const()[name = tensor("op_6218"), val = tensor([1, 1, 16, 32, 2])]; + tensor k_complex_29_cast_fp16 = reshape(shape = var_6218, x = k_57_cast_fp16)[name = tensor("k_complex_29_cast_fp16")]; + tensor var_6222_begin_0 = const()[name = tensor("op_6222_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_6222_end_0 = const()[name = tensor("op_6222_end_0"), val = tensor([1, 1, 16, 32, 1])]; + tensor var_6222_end_mask_0 = const()[name = tensor("op_6222_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_6222_squeeze_mask_0 = const()[name = tensor("op_6222_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_6222_cast_fp16 = slice_by_index(begin = var_6222_begin_0, end = var_6222_end_0, end_mask = var_6222_end_mask_0, squeeze_mask = var_6222_squeeze_mask_0, x = q_complex_29_cast_fp16)[name = tensor("op_6222_cast_fp16")]; + tensor var_6230_begin_0 = const()[name = tensor("op_6230_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_6230_end_0 = const()[name = tensor("op_6230_end_0"), val = tensor([1, 1, 16, 32, 2])]; + tensor var_6230_end_mask_0 = const()[name = tensor("op_6230_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_6230_squeeze_mask_0 = const()[name = tensor("op_6230_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_6230_cast_fp16 = slice_by_index(begin = var_6230_begin_0, end = var_6230_end_0, end_mask = var_6230_end_mask_0, squeeze_mask = var_6230_squeeze_mask_0, x = q_complex_29_cast_fp16)[name = tensor("op_6230_cast_fp16")]; + tensor var_6238_begin_0 = const()[name = tensor("op_6238_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_6238_end_0 = const()[name = tensor("op_6238_end_0"), val = tensor([1, 1, 16, 32, 1])]; + tensor var_6238_end_mask_0 = const()[name = tensor("op_6238_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_6238_squeeze_mask_0 = const()[name = tensor("op_6238_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_6238_cast_fp16 = slice_by_index(begin = var_6238_begin_0, end = var_6238_end_0, end_mask = var_6238_end_mask_0, squeeze_mask = var_6238_squeeze_mask_0, x = k_complex_29_cast_fp16)[name = tensor("op_6238_cast_fp16")]; + tensor var_6246_begin_0 = const()[name = tensor("op_6246_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_6246_end_0 = const()[name = tensor("op_6246_end_0"), val = tensor([1, 1, 16, 32, 2])]; + tensor var_6246_end_mask_0 = const()[name = tensor("op_6246_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_6246_squeeze_mask_0 = const()[name = tensor("op_6246_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_6246_cast_fp16 = slice_by_index(begin = var_6246_begin_0, end = var_6246_end_0, end_mask = var_6246_end_mask_0, squeeze_mask = var_6246_squeeze_mask_0, x = k_complex_29_cast_fp16)[name = tensor("op_6246_cast_fp16")]; + tensor freqs_29_to_fp16 = const()[name = tensor("freqs_29_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(303709184)))]; + tensor var_6252_cast_fp16 = mul(x = freqs_29_to_fp16, y = ts_89_cast_fp16)[name = tensor("op_6252_cast_fp16")]; + tensor rotr_29_cast_fp16 = cos(x = var_6252_cast_fp16)[name = tensor("rotr_29_cast_fp16")]; + tensor roti_29_cast_fp16 = sin(x = var_6252_cast_fp16)[name = tensor("roti_29_cast_fp16")]; + tensor var_6256_cast_fp16 = mul(x = var_6222_cast_fp16, y = rotr_29_cast_fp16)[name = tensor("op_6256_cast_fp16")]; + tensor var_6257_cast_fp16 = mul(x = var_6230_cast_fp16, y = roti_29_cast_fp16)[name = tensor("op_6257_cast_fp16")]; + tensor qor_57_cast_fp16 = sub(x = var_6256_cast_fp16, y = var_6257_cast_fp16)[name = tensor("qor_57_cast_fp16")]; + tensor var_6260_cast_fp16 = mul(x = var_6222_cast_fp16, y = roti_29_cast_fp16)[name = tensor("op_6260_cast_fp16")]; + tensor var_6261_cast_fp16 = mul(x = var_6230_cast_fp16, y = rotr_29_cast_fp16)[name = tensor("op_6261_cast_fp16")]; + tensor qoi_57_cast_fp16 = add(x = var_6260_cast_fp16, y = var_6261_cast_fp16)[name = tensor("qoi_57_cast_fp16")]; + tensor var_6264_cast_fp16 = mul(x = var_6238_cast_fp16, y = rotr_29_cast_fp16)[name = tensor("op_6264_cast_fp16")]; + tensor var_6265_cast_fp16 = mul(x = var_6246_cast_fp16, y = roti_29_cast_fp16)[name = tensor("op_6265_cast_fp16")]; + tensor kor_57_cast_fp16 = sub(x = var_6264_cast_fp16, y = var_6265_cast_fp16)[name = tensor("kor_57_cast_fp16")]; + tensor var_6268_cast_fp16 = mul(x = var_6238_cast_fp16, y = roti_29_cast_fp16)[name = tensor("op_6268_cast_fp16")]; + tensor var_6269_cast_fp16 = mul(x = var_6246_cast_fp16, y = rotr_29_cast_fp16)[name = tensor("op_6269_cast_fp16")]; + tensor koi_57_cast_fp16 = add(x = var_6268_cast_fp16, y = var_6269_cast_fp16)[name = tensor("koi_57_cast_fp16")]; + tensor qo_29_axis_0 = const()[name = tensor("qo_29_axis_0"), val = tensor(-1)]; + tensor qo_29_cast_fp16 = stack(axis = qo_29_axis_0, values = (qor_57_cast_fp16, qoi_57_cast_fp16))[name = tensor("qo_29_cast_fp16")]; + tensor ko_29_axis_0 = const()[name = tensor("ko_29_axis_0"), val = tensor(-1)]; + tensor ko_29_cast_fp16 = stack(axis = ko_29_axis_0, values = (kor_57_cast_fp16, koi_57_cast_fp16))[name = tensor("ko_29_cast_fp16")]; + tensor var_6298 = const()[name = tensor("op_6298"), val = tensor([1, 1, 16, 64])]; + tensor q_87_cast_fp16 = reshape(shape = var_6298, x = qo_29_cast_fp16)[name = tensor("q_87_cast_fp16")]; + tensor var_6300 = const()[name = tensor("op_6300"), val = tensor([1, 1, 16, 64])]; + tensor k_59_cast_fp16 = reshape(shape = var_6300, x = ko_29_cast_fp16)[name = tensor("k_59_cast_fp16")]; + tensor _inversed_6322_y_0_to_fp16 = const()[name = tensor("_inversed_6322_y_0_to_fp16"), val = tensor(0x1p-9)]; + tensor _inversed_6322_cast_fp16 = mul(x = ts_89_cast_fp16, y = _inversed_6322_y_0_to_fp16)[name = tensor("_inversed_6322_cast_fp16")]; + tensor var_6323_cast_fp16 = floor(x = _inversed_6322_cast_fp16)[name = tensor("op_6323_cast_fp16")]; + tensor var_6324_to_fp16 = const()[name = tensor("op_6324_to_fp16"), val = tensor(0x1p+9)]; + tensor var_6325_cast_fp16 = mul(x = var_6323_cast_fp16, y = var_6324_to_fp16)[name = tensor("op_6325_cast_fp16")]; + tensor write_indices_float_59_cast_fp16 = sub(x = ts_89_cast_fp16, y = var_6325_cast_fp16)[name = tensor("write_indices_float_59_cast_fp16")]; + tensor var_6332_dtype_0 = const()[name = tensor("op_6332_dtype_0"), val = tensor("int32")]; + tensor write_indices_29_reps_0 = const()[name = tensor("write_indices_29_reps_0"), val = tensor([1, 1, 16, 64])]; + tensor write_indices_float_59_cast_fp16_to_int32 = cast(dtype = var_6332_dtype_0, x = write_indices_float_59_cast_fp16)[name = tensor("cast_482")]; + tensor write_indices_29 = tile(reps = write_indices_29_reps_0, x = write_indices_float_59_cast_fp16_to_int32)[name = tensor("write_indices_29")]; + tensor var_6340_begin_0 = const()[name = tensor("op_6340_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_6340_end_0 = const()[name = tensor("op_6340_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor var_6340_end_mask_0 = const()[name = tensor("op_6340_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_6340_squeeze_mask_0 = const()[name = tensor("op_6340_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor cache14_to_fp16_dtype_0 = const()[name = tensor("cache14_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor cache14_to_fp16 = cast(dtype = cache14_to_fp16_dtype_0, x = cache14)[name = tensor("cast_481")]; + tensor var_6340_cast_fp16 = slice_by_index(begin = var_6340_begin_0, end = var_6340_end_0, end_mask = var_6340_end_mask_0, squeeze_mask = var_6340_squeeze_mask_0, x = cache14_to_fp16)[name = tensor("op_6340_cast_fp16")]; + tensor var_6342_axis_0 = const()[name = tensor("op_6342_axis_0"), val = tensor(1)]; + tensor var_6342_mode_0 = const()[name = tensor("op_6342_mode_0"), val = tensor("update")]; + tensor var_6342_validate_indices_0 = const()[name = tensor("op_6342_validate_indices_0"), val = tensor(false)]; + tensor var_6342_cast_fp16 = scatter_along_axis(axis = var_6342_axis_0, data = var_6340_cast_fp16, indices = write_indices_29, mode = var_6342_mode_0, updates = k_59_cast_fp16, validate_indices = var_6342_validate_indices_0)[name = tensor("op_6342_cast_fp16")]; + tensor concat_100 = const()[name = tensor("concat_100"), val = tensor([0, 0, 0, 0, 0])]; + tensor concat_101 = const()[name = tensor("concat_101"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_29_internal_tensor_assign_1_stride_0 = const()[name = tensor("new_cache_29_internal_tensor_assign_1_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_29_internal_tensor_assign_1_begin_mask_0 = const()[name = tensor("new_cache_29_internal_tensor_assign_1_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_29_internal_tensor_assign_1_end_mask_0 = const()[name = tensor("new_cache_29_internal_tensor_assign_1_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_29_internal_tensor_assign_1_squeeze_mask_0 = const()[name = tensor("new_cache_29_internal_tensor_assign_1_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_76 = const()[name = tensor("shape_76"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_28 = const()[name = tensor("reduce_prod_28"), val = tensor(1048576)]; + tensor range_1d_28_start_0 = const()[name = tensor("range_1d_28_start_0"), val = tensor(0)]; + tensor range_1d_28_step_0 = const()[name = tensor("range_1d_28_step_0"), val = tensor(1)]; + tensor range_1d_28 = range_1d(end = reduce_prod_28, start = range_1d_28_start_0, step = range_1d_28_step_0)[name = tensor("range_1d_28")]; + tensor reshape_140 = reshape(shape = shape_76, x = range_1d_28)[name = tensor("reshape_140")]; + tensor slice_by_index_28 = slice_by_index(begin = concat_100, begin_mask = new_cache_29_internal_tensor_assign_1_begin_mask_0, end = concat_101, end_mask = new_cache_29_internal_tensor_assign_1_end_mask_0, squeeze_mask = new_cache_29_internal_tensor_assign_1_squeeze_mask_0, stride = new_cache_29_internal_tensor_assign_1_stride_0, x = reshape_140)[name = tensor("slice_by_index_28")]; + tensor reshape_141_shape_0 = const()[name = tensor("reshape_141_shape_0"), val = tensor([-1])]; + tensor reshape_141 = reshape(shape = reshape_141_shape_0, x = slice_by_index_28)[name = tensor("reshape_141")]; + tensor reshape_142_shape_0 = const()[name = tensor("reshape_142_shape_0"), val = tensor([-1])]; + tensor reshape_142_cast_fp16 = reshape(shape = reshape_142_shape_0, x = var_6342_cast_fp16)[name = tensor("reshape_142_cast_fp16")]; + tensor reshape_143_shape_0 = const()[name = tensor("reshape_143_shape_0"), val = tensor([-1])]; + tensor reshape_143_cast_fp16 = reshape(shape = reshape_143_shape_0, x = cache14_to_fp16)[name = tensor("reshape_143_cast_fp16")]; + tensor scatter_28_mode_0 = const()[name = tensor("scatter_28_mode_0"), val = tensor("update")]; + tensor scatter_28_axis_0 = const()[name = tensor("scatter_28_axis_0"), val = tensor(0)]; + tensor scatter_28_validate_indices_0 = const()[name = tensor("scatter_28_validate_indices_0"), val = tensor(false)]; + tensor scatter_28_cast_fp16 = scatter(axis = scatter_28_axis_0, data = reshape_143_cast_fp16, indices = reshape_141, mode = scatter_28_mode_0, updates = reshape_142_cast_fp16, validate_indices = scatter_28_validate_indices_0)[name = tensor("scatter_28_cast_fp16")]; + tensor reshape_144_cast_fp16 = reshape(shape = shape_76, x = scatter_28_cast_fp16)[name = tensor("reshape_144_cast_fp16")]; + tensor var_6350_begin_0 = const()[name = tensor("op_6350_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor var_6350_end_0 = const()[name = tensor("op_6350_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor var_6350_end_mask_0 = const()[name = tensor("op_6350_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_6350_squeeze_mask_0 = const()[name = tensor("op_6350_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor var_6350_cast_fp16 = slice_by_index(begin = var_6350_begin_0, end = var_6350_end_0, end_mask = var_6350_end_mask_0, squeeze_mask = var_6350_squeeze_mask_0, x = reshape_144_cast_fp16)[name = tensor("op_6350_cast_fp16")]; + tensor var_6352_axis_0 = const()[name = tensor("op_6352_axis_0"), val = tensor(1)]; + tensor var_6352_mode_0 = const()[name = tensor("op_6352_mode_0"), val = tensor("update")]; + tensor var_6352_validate_indices_0 = const()[name = tensor("op_6352_validate_indices_0"), val = tensor(false)]; + tensor var_6352_cast_fp16 = scatter_along_axis(axis = var_6352_axis_0, data = var_6350_cast_fp16, indices = write_indices_29, mode = var_6352_mode_0, updates = v_29_cast_fp16, validate_indices = var_6352_validate_indices_0)[name = tensor("op_6352_cast_fp16")]; + tensor concat_102 = const()[name = tensor("concat_102"), val = tensor([1, 0, 0, 0, 0])]; + tensor concat_103 = const()[name = tensor("concat_103"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_29_internal_tensor_assign_2_stride_0 = const()[name = tensor("new_cache_29_internal_tensor_assign_2_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_29_internal_tensor_assign_2_begin_mask_0 = const()[name = tensor("new_cache_29_internal_tensor_assign_2_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_29_internal_tensor_assign_2_end_mask_0 = const()[name = tensor("new_cache_29_internal_tensor_assign_2_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_29_internal_tensor_assign_2_squeeze_mask_0 = const()[name = tensor("new_cache_29_internal_tensor_assign_2_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_77 = const()[name = tensor("shape_77"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_29 = const()[name = tensor("reduce_prod_29"), val = tensor(1048576)]; + tensor range_1d_29_start_0 = const()[name = tensor("range_1d_29_start_0"), val = tensor(0)]; + tensor range_1d_29_step_0 = const()[name = tensor("range_1d_29_step_0"), val = tensor(1)]; + tensor range_1d_29 = range_1d(end = reduce_prod_29, start = range_1d_29_start_0, step = range_1d_29_step_0)[name = tensor("range_1d_29")]; + tensor reshape_145 = reshape(shape = shape_77, x = range_1d_29)[name = tensor("reshape_145")]; + tensor slice_by_index_29 = slice_by_index(begin = concat_102, begin_mask = new_cache_29_internal_tensor_assign_2_begin_mask_0, end = concat_103, end_mask = new_cache_29_internal_tensor_assign_2_end_mask_0, squeeze_mask = new_cache_29_internal_tensor_assign_2_squeeze_mask_0, stride = new_cache_29_internal_tensor_assign_2_stride_0, x = reshape_145)[name = tensor("slice_by_index_29")]; + tensor reshape_146_shape_0 = const()[name = tensor("reshape_146_shape_0"), val = tensor([-1])]; + tensor reshape_146 = reshape(shape = reshape_146_shape_0, x = slice_by_index_29)[name = tensor("reshape_146")]; + tensor reshape_147_shape_0 = const()[name = tensor("reshape_147_shape_0"), val = tensor([-1])]; + tensor reshape_147_cast_fp16 = reshape(shape = reshape_147_shape_0, x = var_6352_cast_fp16)[name = tensor("reshape_147_cast_fp16")]; + tensor reshape_148_shape_0 = const()[name = tensor("reshape_148_shape_0"), val = tensor([-1])]; + tensor reshape_148_cast_fp16 = reshape(shape = reshape_148_shape_0, x = reshape_144_cast_fp16)[name = tensor("reshape_148_cast_fp16")]; + tensor scatter_29_mode_0 = const()[name = tensor("scatter_29_mode_0"), val = tensor("update")]; + tensor scatter_29_axis_0 = const()[name = tensor("scatter_29_axis_0"), val = tensor(0)]; + tensor scatter_29_validate_indices_0 = const()[name = tensor("scatter_29_validate_indices_0"), val = tensor(false)]; + tensor scatter_29_cast_fp16 = scatter(axis = scatter_29_axis_0, data = reshape_148_cast_fp16, indices = reshape_146, mode = scatter_29_mode_0, updates = reshape_147_cast_fp16, validate_indices = scatter_29_validate_indices_0)[name = tensor("scatter_29_cast_fp16")]; + tensor reshape_149_cast_fp16 = reshape(shape = shape_77, x = scatter_29_cast_fp16)[name = tensor("reshape_149_cast_fp16")]; + tensor reshape_149_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("reshape_149_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor keys_85_begin_0 = const()[name = tensor("keys_85_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor keys_85_end_0 = const()[name = tensor("keys_85_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor keys_85_end_mask_0 = const()[name = tensor("keys_85_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor keys_85_squeeze_mask_0 = const()[name = tensor("keys_85_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor keys_85_cast_fp16 = slice_by_index(begin = keys_85_begin_0, end = keys_85_end_0, end_mask = keys_85_end_mask_0, squeeze_mask = keys_85_squeeze_mask_0, x = reshape_149_cast_fp16)[name = tensor("keys_85_cast_fp16")]; + tensor values_85_begin_0 = const()[name = tensor("values_85_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor values_85_end_0 = const()[name = tensor("values_85_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor values_85_end_mask_0 = const()[name = tensor("values_85_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor values_85_squeeze_mask_0 = const()[name = tensor("values_85_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor values_85_cast_fp16 = slice_by_index(begin = values_85_begin_0, end = values_85_end_0, end_mask = values_85_end_mask_0, squeeze_mask = values_85_squeeze_mask_0, x = reshape_149_cast_fp16)[name = tensor("values_85_cast_fp16")]; + tensor var_6364_cast_fp16 = not_equal(x = keys_85_cast_fp16, y = keys_85_cast_fp16)[name = tensor("op_6364_cast_fp16")]; + tensor keys_87_cast_fp16 = select(a = var_504_to_fp16, b = keys_85_cast_fp16, cond = var_6364_cast_fp16)[name = tensor("keys_87_cast_fp16")]; + tensor var_6372_cast_fp16 = not_equal(x = values_85_cast_fp16, y = values_85_cast_fp16)[name = tensor("op_6372_cast_fp16")]; + tensor values_87_cast_fp16 = select(a = var_504_to_fp16, b = values_85_cast_fp16, cond = var_6372_cast_fp16)[name = tensor("values_87_cast_fp16")]; + tensor var_6396 = const()[name = tensor("op_6396"), val = tensor([0, 2, 1, 3])]; + tensor var_6409 = const()[name = tensor("op_6409"), val = tensor([1, 1, 1])]; + tensor var_6410_cast_fp16 = reshape(shape = var_6409, x = position14_to_fp16)[name = tensor("op_6410_cast_fp16")]; + tensor var_6427_to_fp16 = const()[name = tensor("op_6427_to_fp16"), val = tensor(0x1p+0)]; + tensor valid_len_29_cast_fp16 = add(x = var_6410_cast_fp16, y = var_6427_to_fp16)[name = tensor("valid_len_29_cast_fp16")]; + tensor valid_mask_29_cast_fp16 = less(x = k_positions_1_promoted_to_fp16, y = valid_len_29_cast_fp16)[name = tensor("valid_mask_29_cast_fp16")]; + tensor causal_mask_29_cast_fp16 = less_equal(x = k_positions_1_promoted_to_fp16, y = var_6410_cast_fp16)[name = tensor("causal_mask_29_cast_fp16")]; + tensor attn_mask_57 = logical_and(x = valid_mask_29_cast_fp16, y = causal_mask_29_cast_fp16)[name = tensor("attn_mask_57")]; + tensor attn_mask_59_axes_0 = const()[name = tensor("attn_mask_59_axes_0"), val = tensor([1])]; + tensor attn_mask_59 = expand_dims(axes = attn_mask_59_axes_0, x = attn_mask_57)[name = tensor("attn_mask_59")]; + tensor var_6445_transpose_x_0 = const()[name = tensor("op_6445_transpose_x_0"), val = tensor(false)]; + tensor var_6445_transpose_y_0 = const()[name = tensor("op_6445_transpose_y_0"), val = tensor(false)]; + tensor transpose_100_perm_0 = const()[name = tensor("transpose_100_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_101_perm_0 = const()[name = tensor("transpose_101_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_101 = transpose(perm = transpose_101_perm_0, x = keys_87_cast_fp16)[name = tensor("transpose_157")]; + tensor transpose_100 = transpose(perm = transpose_100_perm_0, x = q_87_cast_fp16)[name = tensor("transpose_158")]; + tensor var_6445_cast_fp16 = matmul(transpose_x = var_6445_transpose_x_0, transpose_y = var_6445_transpose_y_0, x = transpose_100, y = transpose_101)[name = tensor("op_6445_cast_fp16")]; + tensor var_6439_to_fp16 = const()[name = tensor("op_6439_to_fp16"), val = tensor([0x1p-3])]; + tensor attn_weights_85_cast_fp16 = mul(x = var_6445_cast_fp16, y = var_6439_to_fp16)[name = tensor("attn_weights_85_cast_fp16")]; + tensor var_6447 = logical_not(x = attn_mask_59)[name = tensor("op_6447")]; + tensor var_6448_to_fp16 = const()[name = tensor("op_6448_to_fp16"), val = tensor(-inf)]; + tensor attn_weights_87_cast_fp16 = select(a = var_6448_to_fp16, b = attn_weights_85_cast_fp16, cond = var_6447)[name = tensor("attn_weights_87_cast_fp16")]; + tensor var_6450 = const()[name = tensor("op_6450"), val = tensor(-1)]; + tensor attn_weights_89_cast_fp16 = softmax(axis = var_6450, x = attn_weights_87_cast_fp16)[name = tensor("attn_weights_89_cast_fp16")]; + tensor attn_output_29_transpose_x_0 = const()[name = tensor("attn_output_29_transpose_x_0"), val = tensor(false)]; + tensor attn_output_29_transpose_y_0 = const()[name = tensor("attn_output_29_transpose_y_0"), val = tensor(false)]; + tensor values_89_cast_fp16 = transpose(perm = var_6396, x = values_87_cast_fp16)[name = tensor("transpose_159")]; + tensor attn_output_29_cast_fp16 = matmul(transpose_x = attn_output_29_transpose_x_0, transpose_y = attn_output_29_transpose_y_0, x = attn_weights_89_cast_fp16, y = values_89_cast_fp16)[name = tensor("attn_output_29_cast_fp16")]; + tensor var_6458 = const()[name = tensor("op_6458"), val = tensor([0, 2, 1, 3])]; + tensor var_6461 = const()[name = tensor("op_6461"), val = tensor([1, 1, 1024])]; + tensor var_6459_cast_fp16 = transpose(perm = var_6458, x = attn_output_29_cast_fp16)[name = tensor("transpose_156")]; + tensor input_145_cast_fp16 = reshape(shape = var_6461, x = var_6459_cast_fp16)[name = tensor("input_145_cast_fp16")]; + tensor linear_58_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = attn14_out_proj_weight_cast_fp16, x = input_145_cast_fp16)[name = tensor("linear_58_cast_fp16")]; + tensor var_6467_to_fp16 = const()[name = tensor("op_6467_to_fp16"), val = tensor(0x1p+0)]; + tensor var_6468_cast_fp16 = add(x = position14_to_fp16, y = var_6467_to_fp16)[name = tensor("op_6468_cast_fp16")]; + tensor var_6468_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_6468_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor input_147_cast_fp16 = add(x = input_143_cast_fp16, y = linear_58_cast_fp16)[name = tensor("input_147_cast_fp16")]; + tensor input_149_axes_0 = const()[name = tensor("input_149_axes_0"), val = tensor([-1])]; + tensor norm14_2_weight_to_fp16 = const()[name = tensor("norm14_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(303709312)))]; + tensor norm14_2_bias_to_fp16 = const()[name = tensor("norm14_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(303711424)))]; + tensor var_6472_to_fp16 = const()[name = tensor("op_6472_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_149_cast_fp16 = layer_norm(axes = input_149_axes_0, beta = norm14_2_bias_to_fp16, epsilon = var_6472_to_fp16, gamma = norm14_2_weight_to_fp16, x = input_147_cast_fp16)[name = tensor("input_149_cast_fp16")]; + tensor linear_59_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = linear14_1_weight_cast_fp16, x = input_149_cast_fp16)[name = tensor("linear_59_cast_fp16")]; + tensor input_151_mode_0 = const()[name = tensor("input_151_mode_0"), val = tensor("EXACT")]; + tensor input_151_cast_fp16 = gelu(mode = input_151_mode_0, x = linear_59_cast_fp16)[name = tensor("input_151_cast_fp16")]; + tensor linear_60_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = linear14_2_weight_cast_fp16, x = input_151_cast_fp16)[name = tensor("linear_60_cast_fp16")]; + tensor input_153_cast_fp16 = add(x = input_147_cast_fp16, y = linear_60_cast_fp16)[name = tensor("input_153_cast_fp16")]; + tensor x_31_axes_0 = const()[name = tensor("x_31_axes_0"), val = tensor([-1])]; + tensor norm15_1_weight_to_fp16 = const()[name = tensor("norm15_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(303713536)))]; + tensor norm15_1_bias_to_fp16 = const()[name = tensor("norm15_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(303715648)))]; + tensor var_6489_to_fp16 = const()[name = tensor("op_6489_to_fp16"), val = tensor(0x1.5p-17)]; + tensor x_31_cast_fp16 = layer_norm(axes = x_31_axes_0, beta = norm15_1_bias_to_fp16, epsilon = var_6489_to_fp16, gamma = norm15_1_weight_to_fp16, x = input_153_cast_fp16)[name = tensor("x_31_cast_fp16")]; + tensor linear_61_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = attn15_in_proj_weight_cast_fp16, x = x_31_cast_fp16)[name = tensor("linear_61_cast_fp16")]; + tensor var_6525 = const()[name = tensor("op_6525"), val = tensor([1, 1, 3, 16, 64])]; + tensor qkv_31_cast_fp16 = reshape(shape = var_6525, x = linear_61_cast_fp16)[name = tensor("qkv_31_cast_fp16")]; + tensor q_91_begin_0 = const()[name = tensor("q_91_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor q_91_end_0 = const()[name = tensor("q_91_end_0"), val = tensor([1, 1, 1, 16, 64])]; + tensor q_91_end_mask_0 = const()[name = tensor("q_91_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor q_91_squeeze_mask_0 = const()[name = tensor("q_91_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor q_91_cast_fp16 = slice_by_index(begin = q_91_begin_0, end = q_91_end_0, end_mask = q_91_end_mask_0, squeeze_mask = q_91_squeeze_mask_0, x = qkv_31_cast_fp16)[name = tensor("q_91_cast_fp16")]; + tensor k_61_begin_0 = const()[name = tensor("k_61_begin_0"), val = tensor([0, 0, 1, 0, 0])]; + tensor k_61_end_0 = const()[name = tensor("k_61_end_0"), val = tensor([1, 1, 2, 16, 64])]; + tensor k_61_end_mask_0 = const()[name = tensor("k_61_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor k_61_squeeze_mask_0 = const()[name = tensor("k_61_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor k_61_cast_fp16 = slice_by_index(begin = k_61_begin_0, end = k_61_end_0, end_mask = k_61_end_mask_0, squeeze_mask = k_61_squeeze_mask_0, x = qkv_31_cast_fp16)[name = tensor("k_61_cast_fp16")]; + tensor v_31_begin_0 = const()[name = tensor("v_31_begin_0"), val = tensor([0, 0, 2, 0, 0])]; + tensor v_31_end_0 = const()[name = tensor("v_31_end_0"), val = tensor([1, 1, 3, 16, 64])]; + tensor v_31_end_mask_0 = const()[name = tensor("v_31_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor v_31_squeeze_mask_0 = const()[name = tensor("v_31_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor v_31_cast_fp16 = slice_by_index(begin = v_31_begin_0, end = v_31_end_0, end_mask = v_31_end_mask_0, squeeze_mask = v_31_squeeze_mask_0, x = qkv_31_cast_fp16)[name = tensor("v_31_cast_fp16")]; + tensor var_6629 = const()[name = tensor("op_6629"), val = tensor([1, 1, 1, 1])]; + tensor position15_to_fp16_dtype_0 = const()[name = tensor("position15_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor position15_to_fp16 = cast(dtype = position15_to_fp16_dtype_0, x = position15)[name = tensor("cast_478")]; + tensor ts_95_cast_fp16 = reshape(shape = var_6629, x = position15_to_fp16)[name = tensor("ts_95_cast_fp16")]; + tensor var_6633 = const()[name = tensor("op_6633"), val = tensor([1, 1, 16, 32, 2])]; + tensor q_complex_31_cast_fp16 = reshape(shape = var_6633, x = q_91_cast_fp16)[name = tensor("q_complex_31_cast_fp16")]; + tensor var_6637 = const()[name = tensor("op_6637"), val = tensor([1, 1, 16, 32, 2])]; + tensor k_complex_31_cast_fp16 = reshape(shape = var_6637, x = k_61_cast_fp16)[name = tensor("k_complex_31_cast_fp16")]; + tensor var_6641_begin_0 = const()[name = tensor("op_6641_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_6641_end_0 = const()[name = tensor("op_6641_end_0"), val = tensor([1, 1, 16, 32, 1])]; + tensor var_6641_end_mask_0 = const()[name = tensor("op_6641_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_6641_squeeze_mask_0 = const()[name = tensor("op_6641_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_6641_cast_fp16 = slice_by_index(begin = var_6641_begin_0, end = var_6641_end_0, end_mask = var_6641_end_mask_0, squeeze_mask = var_6641_squeeze_mask_0, x = q_complex_31_cast_fp16)[name = tensor("op_6641_cast_fp16")]; + tensor var_6649_begin_0 = const()[name = tensor("op_6649_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_6649_end_0 = const()[name = tensor("op_6649_end_0"), val = tensor([1, 1, 16, 32, 2])]; + tensor var_6649_end_mask_0 = const()[name = tensor("op_6649_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_6649_squeeze_mask_0 = const()[name = tensor("op_6649_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_6649_cast_fp16 = slice_by_index(begin = var_6649_begin_0, end = var_6649_end_0, end_mask = var_6649_end_mask_0, squeeze_mask = var_6649_squeeze_mask_0, x = q_complex_31_cast_fp16)[name = tensor("op_6649_cast_fp16")]; + tensor var_6657_begin_0 = const()[name = tensor("op_6657_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_6657_end_0 = const()[name = tensor("op_6657_end_0"), val = tensor([1, 1, 16, 32, 1])]; + tensor var_6657_end_mask_0 = const()[name = tensor("op_6657_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_6657_squeeze_mask_0 = const()[name = tensor("op_6657_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_6657_cast_fp16 = slice_by_index(begin = var_6657_begin_0, end = var_6657_end_0, end_mask = var_6657_end_mask_0, squeeze_mask = var_6657_squeeze_mask_0, x = k_complex_31_cast_fp16)[name = tensor("op_6657_cast_fp16")]; + tensor var_6665_begin_0 = const()[name = tensor("op_6665_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_6665_end_0 = const()[name = tensor("op_6665_end_0"), val = tensor([1, 1, 16, 32, 2])]; + tensor var_6665_end_mask_0 = const()[name = tensor("op_6665_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_6665_squeeze_mask_0 = const()[name = tensor("op_6665_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_6665_cast_fp16 = slice_by_index(begin = var_6665_begin_0, end = var_6665_end_0, end_mask = var_6665_end_mask_0, squeeze_mask = var_6665_squeeze_mask_0, x = k_complex_31_cast_fp16)[name = tensor("op_6665_cast_fp16")]; + tensor freqs_31_to_fp16 = const()[name = tensor("freqs_31_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(303717760)))]; + tensor var_6671_cast_fp16 = mul(x = freqs_31_to_fp16, y = ts_95_cast_fp16)[name = tensor("op_6671_cast_fp16")]; + tensor rotr_31_cast_fp16 = cos(x = var_6671_cast_fp16)[name = tensor("rotr_31_cast_fp16")]; + tensor roti_31_cast_fp16 = sin(x = var_6671_cast_fp16)[name = tensor("roti_31_cast_fp16")]; + tensor var_6675_cast_fp16 = mul(x = var_6641_cast_fp16, y = rotr_31_cast_fp16)[name = tensor("op_6675_cast_fp16")]; + tensor var_6676_cast_fp16 = mul(x = var_6649_cast_fp16, y = roti_31_cast_fp16)[name = tensor("op_6676_cast_fp16")]; + tensor qor_61_cast_fp16 = sub(x = var_6675_cast_fp16, y = var_6676_cast_fp16)[name = tensor("qor_61_cast_fp16")]; + tensor var_6679_cast_fp16 = mul(x = var_6641_cast_fp16, y = roti_31_cast_fp16)[name = tensor("op_6679_cast_fp16")]; + tensor var_6680_cast_fp16 = mul(x = var_6649_cast_fp16, y = rotr_31_cast_fp16)[name = tensor("op_6680_cast_fp16")]; + tensor qoi_61_cast_fp16 = add(x = var_6679_cast_fp16, y = var_6680_cast_fp16)[name = tensor("qoi_61_cast_fp16")]; + tensor var_6683_cast_fp16 = mul(x = var_6657_cast_fp16, y = rotr_31_cast_fp16)[name = tensor("op_6683_cast_fp16")]; + tensor var_6684_cast_fp16 = mul(x = var_6665_cast_fp16, y = roti_31_cast_fp16)[name = tensor("op_6684_cast_fp16")]; + tensor kor_61_cast_fp16 = sub(x = var_6683_cast_fp16, y = var_6684_cast_fp16)[name = tensor("kor_61_cast_fp16")]; + tensor var_6687_cast_fp16 = mul(x = var_6657_cast_fp16, y = roti_31_cast_fp16)[name = tensor("op_6687_cast_fp16")]; + tensor var_6688_cast_fp16 = mul(x = var_6665_cast_fp16, y = rotr_31_cast_fp16)[name = tensor("op_6688_cast_fp16")]; + tensor koi_61_cast_fp16 = add(x = var_6687_cast_fp16, y = var_6688_cast_fp16)[name = tensor("koi_61_cast_fp16")]; + tensor qo_31_axis_0 = const()[name = tensor("qo_31_axis_0"), val = tensor(-1)]; + tensor qo_31_cast_fp16 = stack(axis = qo_31_axis_0, values = (qor_61_cast_fp16, qoi_61_cast_fp16))[name = tensor("qo_31_cast_fp16")]; + tensor ko_31_axis_0 = const()[name = tensor("ko_31_axis_0"), val = tensor(-1)]; + tensor ko_31_cast_fp16 = stack(axis = ko_31_axis_0, values = (kor_61_cast_fp16, koi_61_cast_fp16))[name = tensor("ko_31_cast_fp16")]; + tensor var_6717 = const()[name = tensor("op_6717"), val = tensor([1, 1, 16, 64])]; + tensor q_93_cast_fp16 = reshape(shape = var_6717, x = qo_31_cast_fp16)[name = tensor("q_93_cast_fp16")]; + tensor var_6719 = const()[name = tensor("op_6719"), val = tensor([1, 1, 16, 64])]; + tensor k_63_cast_fp16 = reshape(shape = var_6719, x = ko_31_cast_fp16)[name = tensor("k_63_cast_fp16")]; + tensor _inversed_6741_y_0_to_fp16 = const()[name = tensor("_inversed_6741_y_0_to_fp16"), val = tensor(0x1p-9)]; + tensor _inversed_6741_cast_fp16 = mul(x = ts_95_cast_fp16, y = _inversed_6741_y_0_to_fp16)[name = tensor("_inversed_6741_cast_fp16")]; + tensor var_6742_cast_fp16 = floor(x = _inversed_6741_cast_fp16)[name = tensor("op_6742_cast_fp16")]; + tensor var_6743_to_fp16 = const()[name = tensor("op_6743_to_fp16"), val = tensor(0x1p+9)]; + tensor var_6744_cast_fp16 = mul(x = var_6742_cast_fp16, y = var_6743_to_fp16)[name = tensor("op_6744_cast_fp16")]; + tensor write_indices_float_63_cast_fp16 = sub(x = ts_95_cast_fp16, y = var_6744_cast_fp16)[name = tensor("write_indices_float_63_cast_fp16")]; + tensor var_6751_dtype_0 = const()[name = tensor("op_6751_dtype_0"), val = tensor("int32")]; + tensor write_indices_31_reps_0 = const()[name = tensor("write_indices_31_reps_0"), val = tensor([1, 1, 16, 64])]; + tensor write_indices_float_63_cast_fp16_to_int32 = cast(dtype = var_6751_dtype_0, x = write_indices_float_63_cast_fp16)[name = tensor("cast_477")]; + tensor write_indices_31 = tile(reps = write_indices_31_reps_0, x = write_indices_float_63_cast_fp16_to_int32)[name = tensor("write_indices_31")]; + tensor var_6759_begin_0 = const()[name = tensor("op_6759_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_6759_end_0 = const()[name = tensor("op_6759_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor var_6759_end_mask_0 = const()[name = tensor("op_6759_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_6759_squeeze_mask_0 = const()[name = tensor("op_6759_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor cache15_to_fp16_dtype_0 = const()[name = tensor("cache15_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor cache15_to_fp16 = cast(dtype = cache15_to_fp16_dtype_0, x = cache15)[name = tensor("cast_476")]; + tensor var_6759_cast_fp16 = slice_by_index(begin = var_6759_begin_0, end = var_6759_end_0, end_mask = var_6759_end_mask_0, squeeze_mask = var_6759_squeeze_mask_0, x = cache15_to_fp16)[name = tensor("op_6759_cast_fp16")]; + tensor var_6761_axis_0 = const()[name = tensor("op_6761_axis_0"), val = tensor(1)]; + tensor var_6761_mode_0 = const()[name = tensor("op_6761_mode_0"), val = tensor("update")]; + tensor var_6761_validate_indices_0 = const()[name = tensor("op_6761_validate_indices_0"), val = tensor(false)]; + tensor var_6761_cast_fp16 = scatter_along_axis(axis = var_6761_axis_0, data = var_6759_cast_fp16, indices = write_indices_31, mode = var_6761_mode_0, updates = k_63_cast_fp16, validate_indices = var_6761_validate_indices_0)[name = tensor("op_6761_cast_fp16")]; + tensor concat_107 = const()[name = tensor("concat_107"), val = tensor([0, 0, 0, 0, 0])]; + tensor concat_108 = const()[name = tensor("concat_108"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_31_internal_tensor_assign_1_stride_0 = const()[name = tensor("new_cache_31_internal_tensor_assign_1_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_31_internal_tensor_assign_1_begin_mask_0 = const()[name = tensor("new_cache_31_internal_tensor_assign_1_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_31_internal_tensor_assign_1_end_mask_0 = const()[name = tensor("new_cache_31_internal_tensor_assign_1_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_31_internal_tensor_assign_1_squeeze_mask_0 = const()[name = tensor("new_cache_31_internal_tensor_assign_1_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_78 = const()[name = tensor("shape_78"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_30 = const()[name = tensor("reduce_prod_30"), val = tensor(1048576)]; + tensor range_1d_30_start_0 = const()[name = tensor("range_1d_30_start_0"), val = tensor(0)]; + tensor range_1d_30_step_0 = const()[name = tensor("range_1d_30_step_0"), val = tensor(1)]; + tensor range_1d_30 = range_1d(end = reduce_prod_30, start = range_1d_30_start_0, step = range_1d_30_step_0)[name = tensor("range_1d_30")]; + tensor reshape_150 = reshape(shape = shape_78, x = range_1d_30)[name = tensor("reshape_150")]; + tensor slice_by_index_30 = slice_by_index(begin = concat_107, begin_mask = new_cache_31_internal_tensor_assign_1_begin_mask_0, end = concat_108, end_mask = new_cache_31_internal_tensor_assign_1_end_mask_0, squeeze_mask = new_cache_31_internal_tensor_assign_1_squeeze_mask_0, stride = new_cache_31_internal_tensor_assign_1_stride_0, x = reshape_150)[name = tensor("slice_by_index_30")]; + tensor reshape_151_shape_0 = const()[name = tensor("reshape_151_shape_0"), val = tensor([-1])]; + tensor reshape_151 = reshape(shape = reshape_151_shape_0, x = slice_by_index_30)[name = tensor("reshape_151")]; + tensor reshape_152_shape_0 = const()[name = tensor("reshape_152_shape_0"), val = tensor([-1])]; + tensor reshape_152_cast_fp16 = reshape(shape = reshape_152_shape_0, x = var_6761_cast_fp16)[name = tensor("reshape_152_cast_fp16")]; + tensor reshape_153_shape_0 = const()[name = tensor("reshape_153_shape_0"), val = tensor([-1])]; + tensor reshape_153_cast_fp16 = reshape(shape = reshape_153_shape_0, x = cache15_to_fp16)[name = tensor("reshape_153_cast_fp16")]; + tensor scatter_30_mode_0 = const()[name = tensor("scatter_30_mode_0"), val = tensor("update")]; + tensor scatter_30_axis_0 = const()[name = tensor("scatter_30_axis_0"), val = tensor(0)]; + tensor scatter_30_validate_indices_0 = const()[name = tensor("scatter_30_validate_indices_0"), val = tensor(false)]; + tensor scatter_30_cast_fp16 = scatter(axis = scatter_30_axis_0, data = reshape_153_cast_fp16, indices = reshape_151, mode = scatter_30_mode_0, updates = reshape_152_cast_fp16, validate_indices = scatter_30_validate_indices_0)[name = tensor("scatter_30_cast_fp16")]; + tensor reshape_154_cast_fp16 = reshape(shape = shape_78, x = scatter_30_cast_fp16)[name = tensor("reshape_154_cast_fp16")]; + tensor var_6769_begin_0 = const()[name = tensor("op_6769_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor var_6769_end_0 = const()[name = tensor("op_6769_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor var_6769_end_mask_0 = const()[name = tensor("op_6769_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_6769_squeeze_mask_0 = const()[name = tensor("op_6769_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor var_6769_cast_fp16 = slice_by_index(begin = var_6769_begin_0, end = var_6769_end_0, end_mask = var_6769_end_mask_0, squeeze_mask = var_6769_squeeze_mask_0, x = reshape_154_cast_fp16)[name = tensor("op_6769_cast_fp16")]; + tensor var_6771_axis_0 = const()[name = tensor("op_6771_axis_0"), val = tensor(1)]; + tensor var_6771_mode_0 = const()[name = tensor("op_6771_mode_0"), val = tensor("update")]; + tensor var_6771_validate_indices_0 = const()[name = tensor("op_6771_validate_indices_0"), val = tensor(false)]; + tensor var_6771_cast_fp16 = scatter_along_axis(axis = var_6771_axis_0, data = var_6769_cast_fp16, indices = write_indices_31, mode = var_6771_mode_0, updates = v_31_cast_fp16, validate_indices = var_6771_validate_indices_0)[name = tensor("op_6771_cast_fp16")]; + tensor concat_109 = const()[name = tensor("concat_109"), val = tensor([1, 0, 0, 0, 0])]; + tensor concat_110 = const()[name = tensor("concat_110"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_31_internal_tensor_assign_2_stride_0 = const()[name = tensor("new_cache_31_internal_tensor_assign_2_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_31_internal_tensor_assign_2_begin_mask_0 = const()[name = tensor("new_cache_31_internal_tensor_assign_2_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_31_internal_tensor_assign_2_end_mask_0 = const()[name = tensor("new_cache_31_internal_tensor_assign_2_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_31_internal_tensor_assign_2_squeeze_mask_0 = const()[name = tensor("new_cache_31_internal_tensor_assign_2_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_79 = const()[name = tensor("shape_79"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_31 = const()[name = tensor("reduce_prod_31"), val = tensor(1048576)]; + tensor range_1d_31_start_0 = const()[name = tensor("range_1d_31_start_0"), val = tensor(0)]; + tensor range_1d_31_step_0 = const()[name = tensor("range_1d_31_step_0"), val = tensor(1)]; + tensor range_1d_31 = range_1d(end = reduce_prod_31, start = range_1d_31_start_0, step = range_1d_31_step_0)[name = tensor("range_1d_31")]; + tensor reshape_155 = reshape(shape = shape_79, x = range_1d_31)[name = tensor("reshape_155")]; + tensor slice_by_index_31 = slice_by_index(begin = concat_109, begin_mask = new_cache_31_internal_tensor_assign_2_begin_mask_0, end = concat_110, end_mask = new_cache_31_internal_tensor_assign_2_end_mask_0, squeeze_mask = new_cache_31_internal_tensor_assign_2_squeeze_mask_0, stride = new_cache_31_internal_tensor_assign_2_stride_0, x = reshape_155)[name = tensor("slice_by_index_31")]; + tensor reshape_156_shape_0 = const()[name = tensor("reshape_156_shape_0"), val = tensor([-1])]; + tensor reshape_156 = reshape(shape = reshape_156_shape_0, x = slice_by_index_31)[name = tensor("reshape_156")]; + tensor reshape_157_shape_0 = const()[name = tensor("reshape_157_shape_0"), val = tensor([-1])]; + tensor reshape_157_cast_fp16 = reshape(shape = reshape_157_shape_0, x = var_6771_cast_fp16)[name = tensor("reshape_157_cast_fp16")]; + tensor reshape_158_shape_0 = const()[name = tensor("reshape_158_shape_0"), val = tensor([-1])]; + tensor reshape_158_cast_fp16 = reshape(shape = reshape_158_shape_0, x = reshape_154_cast_fp16)[name = tensor("reshape_158_cast_fp16")]; + tensor scatter_31_mode_0 = const()[name = tensor("scatter_31_mode_0"), val = tensor("update")]; + tensor scatter_31_axis_0 = const()[name = tensor("scatter_31_axis_0"), val = tensor(0)]; + tensor scatter_31_validate_indices_0 = const()[name = tensor("scatter_31_validate_indices_0"), val = tensor(false)]; + tensor scatter_31_cast_fp16 = scatter(axis = scatter_31_axis_0, data = reshape_158_cast_fp16, indices = reshape_156, mode = scatter_31_mode_0, updates = reshape_157_cast_fp16, validate_indices = scatter_31_validate_indices_0)[name = tensor("scatter_31_cast_fp16")]; + tensor reshape_159_cast_fp16 = reshape(shape = shape_79, x = scatter_31_cast_fp16)[name = tensor("reshape_159_cast_fp16")]; + tensor reshape_159_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("reshape_159_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor keys_91_begin_0 = const()[name = tensor("keys_91_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor keys_91_end_0 = const()[name = tensor("keys_91_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor keys_91_end_mask_0 = const()[name = tensor("keys_91_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor keys_91_squeeze_mask_0 = const()[name = tensor("keys_91_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor keys_91_cast_fp16 = slice_by_index(begin = keys_91_begin_0, end = keys_91_end_0, end_mask = keys_91_end_mask_0, squeeze_mask = keys_91_squeeze_mask_0, x = reshape_159_cast_fp16)[name = tensor("keys_91_cast_fp16")]; + tensor values_91_begin_0 = const()[name = tensor("values_91_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor values_91_end_0 = const()[name = tensor("values_91_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor values_91_end_mask_0 = const()[name = tensor("values_91_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor values_91_squeeze_mask_0 = const()[name = tensor("values_91_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor values_91_cast_fp16 = slice_by_index(begin = values_91_begin_0, end = values_91_end_0, end_mask = values_91_end_mask_0, squeeze_mask = values_91_squeeze_mask_0, x = reshape_159_cast_fp16)[name = tensor("values_91_cast_fp16")]; + tensor var_6783_cast_fp16 = not_equal(x = keys_91_cast_fp16, y = keys_91_cast_fp16)[name = tensor("op_6783_cast_fp16")]; + tensor keys_93_cast_fp16 = select(a = var_504_to_fp16, b = keys_91_cast_fp16, cond = var_6783_cast_fp16)[name = tensor("keys_93_cast_fp16")]; + tensor var_6791_cast_fp16 = not_equal(x = values_91_cast_fp16, y = values_91_cast_fp16)[name = tensor("op_6791_cast_fp16")]; + tensor values_93_cast_fp16 = select(a = var_504_to_fp16, b = values_91_cast_fp16, cond = var_6791_cast_fp16)[name = tensor("values_93_cast_fp16")]; + tensor var_6815 = const()[name = tensor("op_6815"), val = tensor([0, 2, 1, 3])]; + tensor var_6828 = const()[name = tensor("op_6828"), val = tensor([1, 1, 1])]; + tensor var_6829_cast_fp16 = reshape(shape = var_6828, x = position15_to_fp16)[name = tensor("op_6829_cast_fp16")]; + tensor var_6846_to_fp16 = const()[name = tensor("op_6846_to_fp16"), val = tensor(0x1p+0)]; + tensor valid_len_31_cast_fp16 = add(x = var_6829_cast_fp16, y = var_6846_to_fp16)[name = tensor("valid_len_31_cast_fp16")]; + tensor valid_mask_31_cast_fp16 = less(x = k_positions_1_promoted_to_fp16, y = valid_len_31_cast_fp16)[name = tensor("valid_mask_31_cast_fp16")]; + tensor causal_mask_31_cast_fp16 = less_equal(x = k_positions_1_promoted_to_fp16, y = var_6829_cast_fp16)[name = tensor("causal_mask_31_cast_fp16")]; + tensor attn_mask_61 = logical_and(x = valid_mask_31_cast_fp16, y = causal_mask_31_cast_fp16)[name = tensor("attn_mask_61")]; + tensor attn_mask_63_axes_0 = const()[name = tensor("attn_mask_63_axes_0"), val = tensor([1])]; + tensor attn_mask_63 = expand_dims(axes = attn_mask_63_axes_0, x = attn_mask_61)[name = tensor("attn_mask_63")]; + tensor var_6864_transpose_x_0 = const()[name = tensor("op_6864_transpose_x_0"), val = tensor(false)]; + tensor var_6864_transpose_y_0 = const()[name = tensor("op_6864_transpose_y_0"), val = tensor(false)]; + tensor transpose_102_perm_0 = const()[name = tensor("transpose_102_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_103_perm_0 = const()[name = tensor("transpose_103_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_103 = transpose(perm = transpose_103_perm_0, x = keys_93_cast_fp16)[name = tensor("transpose_153")]; + tensor transpose_102 = transpose(perm = transpose_102_perm_0, x = q_93_cast_fp16)[name = tensor("transpose_154")]; + tensor var_6864_cast_fp16 = matmul(transpose_x = var_6864_transpose_x_0, transpose_y = var_6864_transpose_y_0, x = transpose_102, y = transpose_103)[name = tensor("op_6864_cast_fp16")]; + tensor var_6858_to_fp16 = const()[name = tensor("op_6858_to_fp16"), val = tensor([0x1p-3])]; + tensor attn_weights_91_cast_fp16 = mul(x = var_6864_cast_fp16, y = var_6858_to_fp16)[name = tensor("attn_weights_91_cast_fp16")]; + tensor var_6866 = logical_not(x = attn_mask_63)[name = tensor("op_6866")]; + tensor var_6867_to_fp16 = const()[name = tensor("op_6867_to_fp16"), val = tensor(-inf)]; + tensor attn_weights_93_cast_fp16 = select(a = var_6867_to_fp16, b = attn_weights_91_cast_fp16, cond = var_6866)[name = tensor("attn_weights_93_cast_fp16")]; + tensor var_6869 = const()[name = tensor("op_6869"), val = tensor(-1)]; + tensor attn_weights_95_cast_fp16 = softmax(axis = var_6869, x = attn_weights_93_cast_fp16)[name = tensor("attn_weights_95_cast_fp16")]; + tensor attn_output_31_transpose_x_0 = const()[name = tensor("attn_output_31_transpose_x_0"), val = tensor(false)]; + tensor attn_output_31_transpose_y_0 = const()[name = tensor("attn_output_31_transpose_y_0"), val = tensor(false)]; + tensor values_95_cast_fp16 = transpose(perm = var_6815, x = values_93_cast_fp16)[name = tensor("transpose_155")]; + tensor attn_output_31_cast_fp16 = matmul(transpose_x = attn_output_31_transpose_x_0, transpose_y = attn_output_31_transpose_y_0, x = attn_weights_95_cast_fp16, y = values_95_cast_fp16)[name = tensor("attn_output_31_cast_fp16")]; + tensor var_6877 = const()[name = tensor("op_6877"), val = tensor([0, 2, 1, 3])]; + tensor var_6880 = const()[name = tensor("op_6880"), val = tensor([1, 1, 1024])]; + tensor var_6878_cast_fp16 = transpose(perm = var_6877, x = attn_output_31_cast_fp16)[name = tensor("transpose_152")]; + tensor input_155_cast_fp16 = reshape(shape = var_6880, x = var_6878_cast_fp16)[name = tensor("input_155_cast_fp16")]; + tensor linear_62_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = attn15_out_proj_weight_cast_fp16, x = input_155_cast_fp16)[name = tensor("linear_62_cast_fp16")]; + tensor var_6886_to_fp16 = const()[name = tensor("op_6886_to_fp16"), val = tensor(0x1p+0)]; + tensor var_6887_cast_fp16 = add(x = position15_to_fp16, y = var_6886_to_fp16)[name = tensor("op_6887_cast_fp16")]; + tensor var_6887_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_6887_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor input_157_cast_fp16 = add(x = input_153_cast_fp16, y = linear_62_cast_fp16)[name = tensor("input_157_cast_fp16")]; + tensor input_159_axes_0 = const()[name = tensor("input_159_axes_0"), val = tensor([-1])]; + tensor norm15_2_weight_to_fp16 = const()[name = tensor("norm15_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(303717888)))]; + tensor norm15_2_bias_to_fp16 = const()[name = tensor("norm15_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(303720000)))]; + tensor var_6891_to_fp16 = const()[name = tensor("op_6891_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_159_cast_fp16 = layer_norm(axes = input_159_axes_0, beta = norm15_2_bias_to_fp16, epsilon = var_6891_to_fp16, gamma = norm15_2_weight_to_fp16, x = input_157_cast_fp16)[name = tensor("input_159_cast_fp16")]; + tensor linear_63_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = linear15_1_weight_cast_fp16, x = input_159_cast_fp16)[name = tensor("linear_63_cast_fp16")]; + tensor input_161_mode_0 = const()[name = tensor("input_161_mode_0"), val = tensor("EXACT")]; + tensor input_161_cast_fp16 = gelu(mode = input_161_mode_0, x = linear_63_cast_fp16)[name = tensor("input_161_cast_fp16")]; + tensor linear_64_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = linear15_2_weight_cast_fp16, x = input_161_cast_fp16)[name = tensor("linear_64_cast_fp16")]; + tensor input_163_cast_fp16 = add(x = input_157_cast_fp16, y = linear_64_cast_fp16)[name = tensor("input_163_cast_fp16")]; + tensor x_33_axes_0 = const()[name = tensor("x_33_axes_0"), val = tensor([-1])]; + tensor norm16_1_weight_to_fp16 = const()[name = tensor("norm16_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(303722112)))]; + tensor norm16_1_bias_to_fp16 = const()[name = tensor("norm16_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(303724224)))]; + tensor var_6908_to_fp16 = const()[name = tensor("op_6908_to_fp16"), val = tensor(0x1.5p-17)]; + tensor x_33_cast_fp16 = layer_norm(axes = x_33_axes_0, beta = norm16_1_bias_to_fp16, epsilon = var_6908_to_fp16, gamma = norm16_1_weight_to_fp16, x = input_163_cast_fp16)[name = tensor("x_33_cast_fp16")]; + tensor linear_65_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = attn16_in_proj_weight_cast_fp16, x = x_33_cast_fp16)[name = tensor("linear_65_cast_fp16")]; + tensor var_6944 = const()[name = tensor("op_6944"), val = tensor([1, 1, 3, 16, 64])]; + tensor qkv_33_cast_fp16 = reshape(shape = var_6944, x = linear_65_cast_fp16)[name = tensor("qkv_33_cast_fp16")]; + tensor q_97_begin_0 = const()[name = tensor("q_97_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor q_97_end_0 = const()[name = tensor("q_97_end_0"), val = tensor([1, 1, 1, 16, 64])]; + tensor q_97_end_mask_0 = const()[name = tensor("q_97_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor q_97_squeeze_mask_0 = const()[name = tensor("q_97_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor q_97_cast_fp16 = slice_by_index(begin = q_97_begin_0, end = q_97_end_0, end_mask = q_97_end_mask_0, squeeze_mask = q_97_squeeze_mask_0, x = qkv_33_cast_fp16)[name = tensor("q_97_cast_fp16")]; + tensor k_65_begin_0 = const()[name = tensor("k_65_begin_0"), val = tensor([0, 0, 1, 0, 0])]; + tensor k_65_end_0 = const()[name = tensor("k_65_end_0"), val = tensor([1, 1, 2, 16, 64])]; + tensor k_65_end_mask_0 = const()[name = tensor("k_65_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor k_65_squeeze_mask_0 = const()[name = tensor("k_65_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor k_65_cast_fp16 = slice_by_index(begin = k_65_begin_0, end = k_65_end_0, end_mask = k_65_end_mask_0, squeeze_mask = k_65_squeeze_mask_0, x = qkv_33_cast_fp16)[name = tensor("k_65_cast_fp16")]; + tensor v_33_begin_0 = const()[name = tensor("v_33_begin_0"), val = tensor([0, 0, 2, 0, 0])]; + tensor v_33_end_0 = const()[name = tensor("v_33_end_0"), val = tensor([1, 1, 3, 16, 64])]; + tensor v_33_end_mask_0 = const()[name = tensor("v_33_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor v_33_squeeze_mask_0 = const()[name = tensor("v_33_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor v_33_cast_fp16 = slice_by_index(begin = v_33_begin_0, end = v_33_end_0, end_mask = v_33_end_mask_0, squeeze_mask = v_33_squeeze_mask_0, x = qkv_33_cast_fp16)[name = tensor("v_33_cast_fp16")]; + tensor var_7048 = const()[name = tensor("op_7048"), val = tensor([1, 1, 1, 1])]; + tensor position16_to_fp16_dtype_0 = const()[name = tensor("position16_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor position16_to_fp16 = cast(dtype = position16_to_fp16_dtype_0, x = position16)[name = tensor("cast_473")]; + tensor ts_101_cast_fp16 = reshape(shape = var_7048, x = position16_to_fp16)[name = tensor("ts_101_cast_fp16")]; + tensor var_7052 = const()[name = tensor("op_7052"), val = tensor([1, 1, 16, 32, 2])]; + tensor q_complex_33_cast_fp16 = reshape(shape = var_7052, x = q_97_cast_fp16)[name = tensor("q_complex_33_cast_fp16")]; + tensor var_7056 = const()[name = tensor("op_7056"), val = tensor([1, 1, 16, 32, 2])]; + tensor k_complex_33_cast_fp16 = reshape(shape = var_7056, x = k_65_cast_fp16)[name = tensor("k_complex_33_cast_fp16")]; + tensor var_7060_begin_0 = const()[name = tensor("op_7060_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_7060_end_0 = const()[name = tensor("op_7060_end_0"), val = tensor([1, 1, 16, 32, 1])]; + tensor var_7060_end_mask_0 = const()[name = tensor("op_7060_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_7060_squeeze_mask_0 = const()[name = tensor("op_7060_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_7060_cast_fp16 = slice_by_index(begin = var_7060_begin_0, end = var_7060_end_0, end_mask = var_7060_end_mask_0, squeeze_mask = var_7060_squeeze_mask_0, x = q_complex_33_cast_fp16)[name = tensor("op_7060_cast_fp16")]; + tensor var_7068_begin_0 = const()[name = tensor("op_7068_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_7068_end_0 = const()[name = tensor("op_7068_end_0"), val = tensor([1, 1, 16, 32, 2])]; + tensor var_7068_end_mask_0 = const()[name = tensor("op_7068_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_7068_squeeze_mask_0 = const()[name = tensor("op_7068_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_7068_cast_fp16 = slice_by_index(begin = var_7068_begin_0, end = var_7068_end_0, end_mask = var_7068_end_mask_0, squeeze_mask = var_7068_squeeze_mask_0, x = q_complex_33_cast_fp16)[name = tensor("op_7068_cast_fp16")]; + tensor var_7076_begin_0 = const()[name = tensor("op_7076_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_7076_end_0 = const()[name = tensor("op_7076_end_0"), val = tensor([1, 1, 16, 32, 1])]; + tensor var_7076_end_mask_0 = const()[name = tensor("op_7076_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_7076_squeeze_mask_0 = const()[name = tensor("op_7076_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_7076_cast_fp16 = slice_by_index(begin = var_7076_begin_0, end = var_7076_end_0, end_mask = var_7076_end_mask_0, squeeze_mask = var_7076_squeeze_mask_0, x = k_complex_33_cast_fp16)[name = tensor("op_7076_cast_fp16")]; + tensor var_7084_begin_0 = const()[name = tensor("op_7084_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_7084_end_0 = const()[name = tensor("op_7084_end_0"), val = tensor([1, 1, 16, 32, 2])]; + tensor var_7084_end_mask_0 = const()[name = tensor("op_7084_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_7084_squeeze_mask_0 = const()[name = tensor("op_7084_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_7084_cast_fp16 = slice_by_index(begin = var_7084_begin_0, end = var_7084_end_0, end_mask = var_7084_end_mask_0, squeeze_mask = var_7084_squeeze_mask_0, x = k_complex_33_cast_fp16)[name = tensor("op_7084_cast_fp16")]; + tensor freqs_33_to_fp16 = const()[name = tensor("freqs_33_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(303726336)))]; + tensor var_7090_cast_fp16 = mul(x = freqs_33_to_fp16, y = ts_101_cast_fp16)[name = tensor("op_7090_cast_fp16")]; + tensor rotr_33_cast_fp16 = cos(x = var_7090_cast_fp16)[name = tensor("rotr_33_cast_fp16")]; + tensor roti_33_cast_fp16 = sin(x = var_7090_cast_fp16)[name = tensor("roti_33_cast_fp16")]; + tensor var_7094_cast_fp16 = mul(x = var_7060_cast_fp16, y = rotr_33_cast_fp16)[name = tensor("op_7094_cast_fp16")]; + tensor var_7095_cast_fp16 = mul(x = var_7068_cast_fp16, y = roti_33_cast_fp16)[name = tensor("op_7095_cast_fp16")]; + tensor qor_65_cast_fp16 = sub(x = var_7094_cast_fp16, y = var_7095_cast_fp16)[name = tensor("qor_65_cast_fp16")]; + tensor var_7098_cast_fp16 = mul(x = var_7060_cast_fp16, y = roti_33_cast_fp16)[name = tensor("op_7098_cast_fp16")]; + tensor var_7099_cast_fp16 = mul(x = var_7068_cast_fp16, y = rotr_33_cast_fp16)[name = tensor("op_7099_cast_fp16")]; + tensor qoi_65_cast_fp16 = add(x = var_7098_cast_fp16, y = var_7099_cast_fp16)[name = tensor("qoi_65_cast_fp16")]; + tensor var_7102_cast_fp16 = mul(x = var_7076_cast_fp16, y = rotr_33_cast_fp16)[name = tensor("op_7102_cast_fp16")]; + tensor var_7103_cast_fp16 = mul(x = var_7084_cast_fp16, y = roti_33_cast_fp16)[name = tensor("op_7103_cast_fp16")]; + tensor kor_65_cast_fp16 = sub(x = var_7102_cast_fp16, y = var_7103_cast_fp16)[name = tensor("kor_65_cast_fp16")]; + tensor var_7106_cast_fp16 = mul(x = var_7076_cast_fp16, y = roti_33_cast_fp16)[name = tensor("op_7106_cast_fp16")]; + tensor var_7107_cast_fp16 = mul(x = var_7084_cast_fp16, y = rotr_33_cast_fp16)[name = tensor("op_7107_cast_fp16")]; + tensor koi_65_cast_fp16 = add(x = var_7106_cast_fp16, y = var_7107_cast_fp16)[name = tensor("koi_65_cast_fp16")]; + tensor qo_33_axis_0 = const()[name = tensor("qo_33_axis_0"), val = tensor(-1)]; + tensor qo_33_cast_fp16 = stack(axis = qo_33_axis_0, values = (qor_65_cast_fp16, qoi_65_cast_fp16))[name = tensor("qo_33_cast_fp16")]; + tensor ko_33_axis_0 = const()[name = tensor("ko_33_axis_0"), val = tensor(-1)]; + tensor ko_33_cast_fp16 = stack(axis = ko_33_axis_0, values = (kor_65_cast_fp16, koi_65_cast_fp16))[name = tensor("ko_33_cast_fp16")]; + tensor var_7136 = const()[name = tensor("op_7136"), val = tensor([1, 1, 16, 64])]; + tensor q_99_cast_fp16 = reshape(shape = var_7136, x = qo_33_cast_fp16)[name = tensor("q_99_cast_fp16")]; + tensor var_7138 = const()[name = tensor("op_7138"), val = tensor([1, 1, 16, 64])]; + tensor k_67_cast_fp16 = reshape(shape = var_7138, x = ko_33_cast_fp16)[name = tensor("k_67_cast_fp16")]; + tensor _inversed_7160_y_0_to_fp16 = const()[name = tensor("_inversed_7160_y_0_to_fp16"), val = tensor(0x1p-9)]; + tensor _inversed_7160_cast_fp16 = mul(x = ts_101_cast_fp16, y = _inversed_7160_y_0_to_fp16)[name = tensor("_inversed_7160_cast_fp16")]; + tensor var_7161_cast_fp16 = floor(x = _inversed_7160_cast_fp16)[name = tensor("op_7161_cast_fp16")]; + tensor var_7162_to_fp16 = const()[name = tensor("op_7162_to_fp16"), val = tensor(0x1p+9)]; + tensor var_7163_cast_fp16 = mul(x = var_7161_cast_fp16, y = var_7162_to_fp16)[name = tensor("op_7163_cast_fp16")]; + tensor write_indices_float_67_cast_fp16 = sub(x = ts_101_cast_fp16, y = var_7163_cast_fp16)[name = tensor("write_indices_float_67_cast_fp16")]; + tensor var_7170_dtype_0 = const()[name = tensor("op_7170_dtype_0"), val = tensor("int32")]; + tensor write_indices_33_reps_0 = const()[name = tensor("write_indices_33_reps_0"), val = tensor([1, 1, 16, 64])]; + tensor write_indices_float_67_cast_fp16_to_int32 = cast(dtype = var_7170_dtype_0, x = write_indices_float_67_cast_fp16)[name = tensor("cast_472")]; + tensor write_indices_33 = tile(reps = write_indices_33_reps_0, x = write_indices_float_67_cast_fp16_to_int32)[name = tensor("write_indices_33")]; + tensor var_7178_begin_0 = const()[name = tensor("op_7178_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_7178_end_0 = const()[name = tensor("op_7178_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor var_7178_end_mask_0 = const()[name = tensor("op_7178_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_7178_squeeze_mask_0 = const()[name = tensor("op_7178_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor cache16_to_fp16_dtype_0 = const()[name = tensor("cache16_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor cache16_to_fp16 = cast(dtype = cache16_to_fp16_dtype_0, x = cache16)[name = tensor("cast_471")]; + tensor var_7178_cast_fp16 = slice_by_index(begin = var_7178_begin_0, end = var_7178_end_0, end_mask = var_7178_end_mask_0, squeeze_mask = var_7178_squeeze_mask_0, x = cache16_to_fp16)[name = tensor("op_7178_cast_fp16")]; + tensor var_7180_axis_0 = const()[name = tensor("op_7180_axis_0"), val = tensor(1)]; + tensor var_7180_mode_0 = const()[name = tensor("op_7180_mode_0"), val = tensor("update")]; + tensor var_7180_validate_indices_0 = const()[name = tensor("op_7180_validate_indices_0"), val = tensor(false)]; + tensor var_7180_cast_fp16 = scatter_along_axis(axis = var_7180_axis_0, data = var_7178_cast_fp16, indices = write_indices_33, mode = var_7180_mode_0, updates = k_67_cast_fp16, validate_indices = var_7180_validate_indices_0)[name = tensor("op_7180_cast_fp16")]; + tensor concat_114 = const()[name = tensor("concat_114"), val = tensor([0, 0, 0, 0, 0])]; + tensor concat_115 = const()[name = tensor("concat_115"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_33_internal_tensor_assign_1_stride_0 = const()[name = tensor("new_cache_33_internal_tensor_assign_1_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_33_internal_tensor_assign_1_begin_mask_0 = const()[name = tensor("new_cache_33_internal_tensor_assign_1_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_33_internal_tensor_assign_1_end_mask_0 = const()[name = tensor("new_cache_33_internal_tensor_assign_1_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_33_internal_tensor_assign_1_squeeze_mask_0 = const()[name = tensor("new_cache_33_internal_tensor_assign_1_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_80 = const()[name = tensor("shape_80"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_32 = const()[name = tensor("reduce_prod_32"), val = tensor(1048576)]; + tensor range_1d_32_start_0 = const()[name = tensor("range_1d_32_start_0"), val = tensor(0)]; + tensor range_1d_32_step_0 = const()[name = tensor("range_1d_32_step_0"), val = tensor(1)]; + tensor range_1d_32 = range_1d(end = reduce_prod_32, start = range_1d_32_start_0, step = range_1d_32_step_0)[name = tensor("range_1d_32")]; + tensor reshape_160 = reshape(shape = shape_80, x = range_1d_32)[name = tensor("reshape_160")]; + tensor slice_by_index_32 = slice_by_index(begin = concat_114, begin_mask = new_cache_33_internal_tensor_assign_1_begin_mask_0, end = concat_115, end_mask = new_cache_33_internal_tensor_assign_1_end_mask_0, squeeze_mask = new_cache_33_internal_tensor_assign_1_squeeze_mask_0, stride = new_cache_33_internal_tensor_assign_1_stride_0, x = reshape_160)[name = tensor("slice_by_index_32")]; + tensor reshape_161_shape_0 = const()[name = tensor("reshape_161_shape_0"), val = tensor([-1])]; + tensor reshape_161 = reshape(shape = reshape_161_shape_0, x = slice_by_index_32)[name = tensor("reshape_161")]; + tensor reshape_162_shape_0 = const()[name = tensor("reshape_162_shape_0"), val = tensor([-1])]; + tensor reshape_162_cast_fp16 = reshape(shape = reshape_162_shape_0, x = var_7180_cast_fp16)[name = tensor("reshape_162_cast_fp16")]; + tensor reshape_163_shape_0 = const()[name = tensor("reshape_163_shape_0"), val = tensor([-1])]; + tensor reshape_163_cast_fp16 = reshape(shape = reshape_163_shape_0, x = cache16_to_fp16)[name = tensor("reshape_163_cast_fp16")]; + tensor scatter_32_mode_0 = const()[name = tensor("scatter_32_mode_0"), val = tensor("update")]; + tensor scatter_32_axis_0 = const()[name = tensor("scatter_32_axis_0"), val = tensor(0)]; + tensor scatter_32_validate_indices_0 = const()[name = tensor("scatter_32_validate_indices_0"), val = tensor(false)]; + tensor scatter_32_cast_fp16 = scatter(axis = scatter_32_axis_0, data = reshape_163_cast_fp16, indices = reshape_161, mode = scatter_32_mode_0, updates = reshape_162_cast_fp16, validate_indices = scatter_32_validate_indices_0)[name = tensor("scatter_32_cast_fp16")]; + tensor reshape_164_cast_fp16 = reshape(shape = shape_80, x = scatter_32_cast_fp16)[name = tensor("reshape_164_cast_fp16")]; + tensor var_7188_begin_0 = const()[name = tensor("op_7188_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor var_7188_end_0 = const()[name = tensor("op_7188_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor var_7188_end_mask_0 = const()[name = tensor("op_7188_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_7188_squeeze_mask_0 = const()[name = tensor("op_7188_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor var_7188_cast_fp16 = slice_by_index(begin = var_7188_begin_0, end = var_7188_end_0, end_mask = var_7188_end_mask_0, squeeze_mask = var_7188_squeeze_mask_0, x = reshape_164_cast_fp16)[name = tensor("op_7188_cast_fp16")]; + tensor var_7190_axis_0 = const()[name = tensor("op_7190_axis_0"), val = tensor(1)]; + tensor var_7190_mode_0 = const()[name = tensor("op_7190_mode_0"), val = tensor("update")]; + tensor var_7190_validate_indices_0 = const()[name = tensor("op_7190_validate_indices_0"), val = tensor(false)]; + tensor var_7190_cast_fp16 = scatter_along_axis(axis = var_7190_axis_0, data = var_7188_cast_fp16, indices = write_indices_33, mode = var_7190_mode_0, updates = v_33_cast_fp16, validate_indices = var_7190_validate_indices_0)[name = tensor("op_7190_cast_fp16")]; + tensor concat_116 = const()[name = tensor("concat_116"), val = tensor([1, 0, 0, 0, 0])]; + tensor concat_117 = const()[name = tensor("concat_117"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_33_internal_tensor_assign_2_stride_0 = const()[name = tensor("new_cache_33_internal_tensor_assign_2_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_33_internal_tensor_assign_2_begin_mask_0 = const()[name = tensor("new_cache_33_internal_tensor_assign_2_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_33_internal_tensor_assign_2_end_mask_0 = const()[name = tensor("new_cache_33_internal_tensor_assign_2_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_33_internal_tensor_assign_2_squeeze_mask_0 = const()[name = tensor("new_cache_33_internal_tensor_assign_2_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_81 = const()[name = tensor("shape_81"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_33 = const()[name = tensor("reduce_prod_33"), val = tensor(1048576)]; + tensor range_1d_33_start_0 = const()[name = tensor("range_1d_33_start_0"), val = tensor(0)]; + tensor range_1d_33_step_0 = const()[name = tensor("range_1d_33_step_0"), val = tensor(1)]; + tensor range_1d_33 = range_1d(end = reduce_prod_33, start = range_1d_33_start_0, step = range_1d_33_step_0)[name = tensor("range_1d_33")]; + tensor reshape_165 = reshape(shape = shape_81, x = range_1d_33)[name = tensor("reshape_165")]; + tensor slice_by_index_33 = slice_by_index(begin = concat_116, begin_mask = new_cache_33_internal_tensor_assign_2_begin_mask_0, end = concat_117, end_mask = new_cache_33_internal_tensor_assign_2_end_mask_0, squeeze_mask = new_cache_33_internal_tensor_assign_2_squeeze_mask_0, stride = new_cache_33_internal_tensor_assign_2_stride_0, x = reshape_165)[name = tensor("slice_by_index_33")]; + tensor reshape_166_shape_0 = const()[name = tensor("reshape_166_shape_0"), val = tensor([-1])]; + tensor reshape_166 = reshape(shape = reshape_166_shape_0, x = slice_by_index_33)[name = tensor("reshape_166")]; + tensor reshape_167_shape_0 = const()[name = tensor("reshape_167_shape_0"), val = tensor([-1])]; + tensor reshape_167_cast_fp16 = reshape(shape = reshape_167_shape_0, x = var_7190_cast_fp16)[name = tensor("reshape_167_cast_fp16")]; + tensor reshape_168_shape_0 = const()[name = tensor("reshape_168_shape_0"), val = tensor([-1])]; + tensor reshape_168_cast_fp16 = reshape(shape = reshape_168_shape_0, x = reshape_164_cast_fp16)[name = tensor("reshape_168_cast_fp16")]; + tensor scatter_33_mode_0 = const()[name = tensor("scatter_33_mode_0"), val = tensor("update")]; + tensor scatter_33_axis_0 = const()[name = tensor("scatter_33_axis_0"), val = tensor(0)]; + tensor scatter_33_validate_indices_0 = const()[name = tensor("scatter_33_validate_indices_0"), val = tensor(false)]; + tensor scatter_33_cast_fp16 = scatter(axis = scatter_33_axis_0, data = reshape_168_cast_fp16, indices = reshape_166, mode = scatter_33_mode_0, updates = reshape_167_cast_fp16, validate_indices = scatter_33_validate_indices_0)[name = tensor("scatter_33_cast_fp16")]; + tensor reshape_169_cast_fp16 = reshape(shape = shape_81, x = scatter_33_cast_fp16)[name = tensor("reshape_169_cast_fp16")]; + tensor reshape_169_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("reshape_169_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor keys_97_begin_0 = const()[name = tensor("keys_97_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor keys_97_end_0 = const()[name = tensor("keys_97_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor keys_97_end_mask_0 = const()[name = tensor("keys_97_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor keys_97_squeeze_mask_0 = const()[name = tensor("keys_97_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor keys_97_cast_fp16 = slice_by_index(begin = keys_97_begin_0, end = keys_97_end_0, end_mask = keys_97_end_mask_0, squeeze_mask = keys_97_squeeze_mask_0, x = reshape_169_cast_fp16)[name = tensor("keys_97_cast_fp16")]; + tensor values_97_begin_0 = const()[name = tensor("values_97_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor values_97_end_0 = const()[name = tensor("values_97_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor values_97_end_mask_0 = const()[name = tensor("values_97_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor values_97_squeeze_mask_0 = const()[name = tensor("values_97_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor values_97_cast_fp16 = slice_by_index(begin = values_97_begin_0, end = values_97_end_0, end_mask = values_97_end_mask_0, squeeze_mask = values_97_squeeze_mask_0, x = reshape_169_cast_fp16)[name = tensor("values_97_cast_fp16")]; + tensor var_7202_cast_fp16 = not_equal(x = keys_97_cast_fp16, y = keys_97_cast_fp16)[name = tensor("op_7202_cast_fp16")]; + tensor keys_99_cast_fp16 = select(a = var_504_to_fp16, b = keys_97_cast_fp16, cond = var_7202_cast_fp16)[name = tensor("keys_99_cast_fp16")]; + tensor var_7210_cast_fp16 = not_equal(x = values_97_cast_fp16, y = values_97_cast_fp16)[name = tensor("op_7210_cast_fp16")]; + tensor values_99_cast_fp16 = select(a = var_504_to_fp16, b = values_97_cast_fp16, cond = var_7210_cast_fp16)[name = tensor("values_99_cast_fp16")]; + tensor var_7234 = const()[name = tensor("op_7234"), val = tensor([0, 2, 1, 3])]; + tensor var_7247 = const()[name = tensor("op_7247"), val = tensor([1, 1, 1])]; + tensor var_7248_cast_fp16 = reshape(shape = var_7247, x = position16_to_fp16)[name = tensor("op_7248_cast_fp16")]; + tensor var_7265_to_fp16 = const()[name = tensor("op_7265_to_fp16"), val = tensor(0x1p+0)]; + tensor valid_len_33_cast_fp16 = add(x = var_7248_cast_fp16, y = var_7265_to_fp16)[name = tensor("valid_len_33_cast_fp16")]; + tensor valid_mask_33_cast_fp16 = less(x = k_positions_1_promoted_to_fp16, y = valid_len_33_cast_fp16)[name = tensor("valid_mask_33_cast_fp16")]; + tensor causal_mask_33_cast_fp16 = less_equal(x = k_positions_1_promoted_to_fp16, y = var_7248_cast_fp16)[name = tensor("causal_mask_33_cast_fp16")]; + tensor attn_mask_65 = logical_and(x = valid_mask_33_cast_fp16, y = causal_mask_33_cast_fp16)[name = tensor("attn_mask_65")]; + tensor attn_mask_67_axes_0 = const()[name = tensor("attn_mask_67_axes_0"), val = tensor([1])]; + tensor attn_mask_67 = expand_dims(axes = attn_mask_67_axes_0, x = attn_mask_65)[name = tensor("attn_mask_67")]; + tensor var_7283_transpose_x_0 = const()[name = tensor("op_7283_transpose_x_0"), val = tensor(false)]; + tensor var_7283_transpose_y_0 = const()[name = tensor("op_7283_transpose_y_0"), val = tensor(false)]; + tensor transpose_104_perm_0 = const()[name = tensor("transpose_104_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_105_perm_0 = const()[name = tensor("transpose_105_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_105 = transpose(perm = transpose_105_perm_0, x = keys_99_cast_fp16)[name = tensor("transpose_149")]; + tensor transpose_104 = transpose(perm = transpose_104_perm_0, x = q_99_cast_fp16)[name = tensor("transpose_150")]; + tensor var_7283_cast_fp16 = matmul(transpose_x = var_7283_transpose_x_0, transpose_y = var_7283_transpose_y_0, x = transpose_104, y = transpose_105)[name = tensor("op_7283_cast_fp16")]; + tensor var_7277_to_fp16 = const()[name = tensor("op_7277_to_fp16"), val = tensor([0x1p-3])]; + tensor attn_weights_97_cast_fp16 = mul(x = var_7283_cast_fp16, y = var_7277_to_fp16)[name = tensor("attn_weights_97_cast_fp16")]; + tensor var_7285 = logical_not(x = attn_mask_67)[name = tensor("op_7285")]; + tensor var_7286_to_fp16 = const()[name = tensor("op_7286_to_fp16"), val = tensor(-inf)]; + tensor attn_weights_99_cast_fp16 = select(a = var_7286_to_fp16, b = attn_weights_97_cast_fp16, cond = var_7285)[name = tensor("attn_weights_99_cast_fp16")]; + tensor var_7288 = const()[name = tensor("op_7288"), val = tensor(-1)]; + tensor attn_weights_101_cast_fp16 = softmax(axis = var_7288, x = attn_weights_99_cast_fp16)[name = tensor("attn_weights_101_cast_fp16")]; + tensor attn_output_33_transpose_x_0 = const()[name = tensor("attn_output_33_transpose_x_0"), val = tensor(false)]; + tensor attn_output_33_transpose_y_0 = const()[name = tensor("attn_output_33_transpose_y_0"), val = tensor(false)]; + tensor values_101_cast_fp16 = transpose(perm = var_7234, x = values_99_cast_fp16)[name = tensor("transpose_151")]; + tensor attn_output_33_cast_fp16 = matmul(transpose_x = attn_output_33_transpose_x_0, transpose_y = attn_output_33_transpose_y_0, x = attn_weights_101_cast_fp16, y = values_101_cast_fp16)[name = tensor("attn_output_33_cast_fp16")]; + tensor var_7296 = const()[name = tensor("op_7296"), val = tensor([0, 2, 1, 3])]; + tensor var_7299 = const()[name = tensor("op_7299"), val = tensor([1, 1, 1024])]; + tensor var_7297_cast_fp16 = transpose(perm = var_7296, x = attn_output_33_cast_fp16)[name = tensor("transpose_148")]; + tensor input_165_cast_fp16 = reshape(shape = var_7299, x = var_7297_cast_fp16)[name = tensor("input_165_cast_fp16")]; + tensor linear_66_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = attn16_out_proj_weight_cast_fp16, x = input_165_cast_fp16)[name = tensor("linear_66_cast_fp16")]; + tensor var_7305_to_fp16 = const()[name = tensor("op_7305_to_fp16"), val = tensor(0x1p+0)]; + tensor var_7306_cast_fp16 = add(x = position16_to_fp16, y = var_7305_to_fp16)[name = tensor("op_7306_cast_fp16")]; + tensor var_7306_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_7306_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor input_167_cast_fp16 = add(x = input_163_cast_fp16, y = linear_66_cast_fp16)[name = tensor("input_167_cast_fp16")]; + tensor input_169_axes_0 = const()[name = tensor("input_169_axes_0"), val = tensor([-1])]; + tensor norm16_2_weight_to_fp16 = const()[name = tensor("norm16_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(303726464)))]; + tensor norm16_2_bias_to_fp16 = const()[name = tensor("norm16_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(303728576)))]; + tensor var_7310_to_fp16 = const()[name = tensor("op_7310_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_169_cast_fp16 = layer_norm(axes = input_169_axes_0, beta = norm16_2_bias_to_fp16, epsilon = var_7310_to_fp16, gamma = norm16_2_weight_to_fp16, x = input_167_cast_fp16)[name = tensor("input_169_cast_fp16")]; + tensor linear_67_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = linear16_1_weight_cast_fp16, x = input_169_cast_fp16)[name = tensor("linear_67_cast_fp16")]; + tensor input_171_mode_0 = const()[name = tensor("input_171_mode_0"), val = tensor("EXACT")]; + tensor input_171_cast_fp16 = gelu(mode = input_171_mode_0, x = linear_67_cast_fp16)[name = tensor("input_171_cast_fp16")]; + tensor linear_68_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = linear16_2_weight_cast_fp16, x = input_171_cast_fp16)[name = tensor("linear_68_cast_fp16")]; + tensor input_173_cast_fp16 = add(x = input_167_cast_fp16, y = linear_68_cast_fp16)[name = tensor("input_173_cast_fp16")]; + tensor x_35_axes_0 = const()[name = tensor("x_35_axes_0"), val = tensor([-1])]; + tensor norm17_1_weight_to_fp16 = const()[name = tensor("norm17_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(303730688)))]; + tensor norm17_1_bias_to_fp16 = const()[name = tensor("norm17_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(303732800)))]; + tensor var_7327_to_fp16 = const()[name = tensor("op_7327_to_fp16"), val = tensor(0x1.5p-17)]; + tensor x_35_cast_fp16 = layer_norm(axes = x_35_axes_0, beta = norm17_1_bias_to_fp16, epsilon = var_7327_to_fp16, gamma = norm17_1_weight_to_fp16, x = input_173_cast_fp16)[name = tensor("x_35_cast_fp16")]; + tensor linear_69_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = attn17_in_proj_weight_cast_fp16, x = x_35_cast_fp16)[name = tensor("linear_69_cast_fp16")]; + tensor var_7363 = const()[name = tensor("op_7363"), val = tensor([1, 1, 3, 16, 64])]; + tensor qkv_35_cast_fp16 = reshape(shape = var_7363, x = linear_69_cast_fp16)[name = tensor("qkv_35_cast_fp16")]; + tensor q_103_begin_0 = const()[name = tensor("q_103_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor q_103_end_0 = const()[name = tensor("q_103_end_0"), val = tensor([1, 1, 1, 16, 64])]; + tensor q_103_end_mask_0 = const()[name = tensor("q_103_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor q_103_squeeze_mask_0 = const()[name = tensor("q_103_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor q_103_cast_fp16 = slice_by_index(begin = q_103_begin_0, end = q_103_end_0, end_mask = q_103_end_mask_0, squeeze_mask = q_103_squeeze_mask_0, x = qkv_35_cast_fp16)[name = tensor("q_103_cast_fp16")]; + tensor k_69_begin_0 = const()[name = tensor("k_69_begin_0"), val = tensor([0, 0, 1, 0, 0])]; + tensor k_69_end_0 = const()[name = tensor("k_69_end_0"), val = tensor([1, 1, 2, 16, 64])]; + tensor k_69_end_mask_0 = const()[name = tensor("k_69_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor k_69_squeeze_mask_0 = const()[name = tensor("k_69_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor k_69_cast_fp16 = slice_by_index(begin = k_69_begin_0, end = k_69_end_0, end_mask = k_69_end_mask_0, squeeze_mask = k_69_squeeze_mask_0, x = qkv_35_cast_fp16)[name = tensor("k_69_cast_fp16")]; + tensor v_35_begin_0 = const()[name = tensor("v_35_begin_0"), val = tensor([0, 0, 2, 0, 0])]; + tensor v_35_end_0 = const()[name = tensor("v_35_end_0"), val = tensor([1, 1, 3, 16, 64])]; + tensor v_35_end_mask_0 = const()[name = tensor("v_35_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor v_35_squeeze_mask_0 = const()[name = tensor("v_35_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor v_35_cast_fp16 = slice_by_index(begin = v_35_begin_0, end = v_35_end_0, end_mask = v_35_end_mask_0, squeeze_mask = v_35_squeeze_mask_0, x = qkv_35_cast_fp16)[name = tensor("v_35_cast_fp16")]; + tensor var_7467 = const()[name = tensor("op_7467"), val = tensor([1, 1, 1, 1])]; + tensor position17_to_fp16_dtype_0 = const()[name = tensor("position17_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor position17_to_fp16 = cast(dtype = position17_to_fp16_dtype_0, x = position17)[name = tensor("cast_468")]; + tensor ts_107_cast_fp16 = reshape(shape = var_7467, x = position17_to_fp16)[name = tensor("ts_107_cast_fp16")]; + tensor var_7471 = const()[name = tensor("op_7471"), val = tensor([1, 1, 16, 32, 2])]; + tensor q_complex_35_cast_fp16 = reshape(shape = var_7471, x = q_103_cast_fp16)[name = tensor("q_complex_35_cast_fp16")]; + tensor var_7475 = const()[name = tensor("op_7475"), val = tensor([1, 1, 16, 32, 2])]; + tensor k_complex_35_cast_fp16 = reshape(shape = var_7475, x = k_69_cast_fp16)[name = tensor("k_complex_35_cast_fp16")]; + tensor var_7479_begin_0 = const()[name = tensor("op_7479_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_7479_end_0 = const()[name = tensor("op_7479_end_0"), val = tensor([1, 1, 16, 32, 1])]; + tensor var_7479_end_mask_0 = const()[name = tensor("op_7479_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_7479_squeeze_mask_0 = const()[name = tensor("op_7479_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_7479_cast_fp16 = slice_by_index(begin = var_7479_begin_0, end = var_7479_end_0, end_mask = var_7479_end_mask_0, squeeze_mask = var_7479_squeeze_mask_0, x = q_complex_35_cast_fp16)[name = tensor("op_7479_cast_fp16")]; + tensor var_7487_begin_0 = const()[name = tensor("op_7487_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_7487_end_0 = const()[name = tensor("op_7487_end_0"), val = tensor([1, 1, 16, 32, 2])]; + tensor var_7487_end_mask_0 = const()[name = tensor("op_7487_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_7487_squeeze_mask_0 = const()[name = tensor("op_7487_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_7487_cast_fp16 = slice_by_index(begin = var_7487_begin_0, end = var_7487_end_0, end_mask = var_7487_end_mask_0, squeeze_mask = var_7487_squeeze_mask_0, x = q_complex_35_cast_fp16)[name = tensor("op_7487_cast_fp16")]; + tensor var_7495_begin_0 = const()[name = tensor("op_7495_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_7495_end_0 = const()[name = tensor("op_7495_end_0"), val = tensor([1, 1, 16, 32, 1])]; + tensor var_7495_end_mask_0 = const()[name = tensor("op_7495_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_7495_squeeze_mask_0 = const()[name = tensor("op_7495_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_7495_cast_fp16 = slice_by_index(begin = var_7495_begin_0, end = var_7495_end_0, end_mask = var_7495_end_mask_0, squeeze_mask = var_7495_squeeze_mask_0, x = k_complex_35_cast_fp16)[name = tensor("op_7495_cast_fp16")]; + tensor var_7503_begin_0 = const()[name = tensor("op_7503_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_7503_end_0 = const()[name = tensor("op_7503_end_0"), val = tensor([1, 1, 16, 32, 2])]; + tensor var_7503_end_mask_0 = const()[name = tensor("op_7503_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_7503_squeeze_mask_0 = const()[name = tensor("op_7503_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_7503_cast_fp16 = slice_by_index(begin = var_7503_begin_0, end = var_7503_end_0, end_mask = var_7503_end_mask_0, squeeze_mask = var_7503_squeeze_mask_0, x = k_complex_35_cast_fp16)[name = tensor("op_7503_cast_fp16")]; + tensor freqs_35_to_fp16 = const()[name = tensor("freqs_35_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(303734912)))]; + tensor var_7509_cast_fp16 = mul(x = freqs_35_to_fp16, y = ts_107_cast_fp16)[name = tensor("op_7509_cast_fp16")]; + tensor rotr_35_cast_fp16 = cos(x = var_7509_cast_fp16)[name = tensor("rotr_35_cast_fp16")]; + tensor roti_35_cast_fp16 = sin(x = var_7509_cast_fp16)[name = tensor("roti_35_cast_fp16")]; + tensor var_7513_cast_fp16 = mul(x = var_7479_cast_fp16, y = rotr_35_cast_fp16)[name = tensor("op_7513_cast_fp16")]; + tensor var_7514_cast_fp16 = mul(x = var_7487_cast_fp16, y = roti_35_cast_fp16)[name = tensor("op_7514_cast_fp16")]; + tensor qor_69_cast_fp16 = sub(x = var_7513_cast_fp16, y = var_7514_cast_fp16)[name = tensor("qor_69_cast_fp16")]; + tensor var_7517_cast_fp16 = mul(x = var_7479_cast_fp16, y = roti_35_cast_fp16)[name = tensor("op_7517_cast_fp16")]; + tensor var_7518_cast_fp16 = mul(x = var_7487_cast_fp16, y = rotr_35_cast_fp16)[name = tensor("op_7518_cast_fp16")]; + tensor qoi_69_cast_fp16 = add(x = var_7517_cast_fp16, y = var_7518_cast_fp16)[name = tensor("qoi_69_cast_fp16")]; + tensor var_7521_cast_fp16 = mul(x = var_7495_cast_fp16, y = rotr_35_cast_fp16)[name = tensor("op_7521_cast_fp16")]; + tensor var_7522_cast_fp16 = mul(x = var_7503_cast_fp16, y = roti_35_cast_fp16)[name = tensor("op_7522_cast_fp16")]; + tensor kor_69_cast_fp16 = sub(x = var_7521_cast_fp16, y = var_7522_cast_fp16)[name = tensor("kor_69_cast_fp16")]; + tensor var_7525_cast_fp16 = mul(x = var_7495_cast_fp16, y = roti_35_cast_fp16)[name = tensor("op_7525_cast_fp16")]; + tensor var_7526_cast_fp16 = mul(x = var_7503_cast_fp16, y = rotr_35_cast_fp16)[name = tensor("op_7526_cast_fp16")]; + tensor koi_69_cast_fp16 = add(x = var_7525_cast_fp16, y = var_7526_cast_fp16)[name = tensor("koi_69_cast_fp16")]; + tensor qo_35_axis_0 = const()[name = tensor("qo_35_axis_0"), val = tensor(-1)]; + tensor qo_35_cast_fp16 = stack(axis = qo_35_axis_0, values = (qor_69_cast_fp16, qoi_69_cast_fp16))[name = tensor("qo_35_cast_fp16")]; + tensor ko_35_axis_0 = const()[name = tensor("ko_35_axis_0"), val = tensor(-1)]; + tensor ko_35_cast_fp16 = stack(axis = ko_35_axis_0, values = (kor_69_cast_fp16, koi_69_cast_fp16))[name = tensor("ko_35_cast_fp16")]; + tensor var_7555 = const()[name = tensor("op_7555"), val = tensor([1, 1, 16, 64])]; + tensor q_105_cast_fp16 = reshape(shape = var_7555, x = qo_35_cast_fp16)[name = tensor("q_105_cast_fp16")]; + tensor var_7557 = const()[name = tensor("op_7557"), val = tensor([1, 1, 16, 64])]; + tensor k_71_cast_fp16 = reshape(shape = var_7557, x = ko_35_cast_fp16)[name = tensor("k_71_cast_fp16")]; + tensor _inversed_7579_y_0_to_fp16 = const()[name = tensor("_inversed_7579_y_0_to_fp16"), val = tensor(0x1p-9)]; + tensor _inversed_7579_cast_fp16 = mul(x = ts_107_cast_fp16, y = _inversed_7579_y_0_to_fp16)[name = tensor("_inversed_7579_cast_fp16")]; + tensor var_7580_cast_fp16 = floor(x = _inversed_7579_cast_fp16)[name = tensor("op_7580_cast_fp16")]; + tensor var_7581_to_fp16 = const()[name = tensor("op_7581_to_fp16"), val = tensor(0x1p+9)]; + tensor var_7582_cast_fp16 = mul(x = var_7580_cast_fp16, y = var_7581_to_fp16)[name = tensor("op_7582_cast_fp16")]; + tensor write_indices_float_71_cast_fp16 = sub(x = ts_107_cast_fp16, y = var_7582_cast_fp16)[name = tensor("write_indices_float_71_cast_fp16")]; + tensor var_7589_dtype_0 = const()[name = tensor("op_7589_dtype_0"), val = tensor("int32")]; + tensor write_indices_35_reps_0 = const()[name = tensor("write_indices_35_reps_0"), val = tensor([1, 1, 16, 64])]; + tensor write_indices_float_71_cast_fp16_to_int32 = cast(dtype = var_7589_dtype_0, x = write_indices_float_71_cast_fp16)[name = tensor("cast_467")]; + tensor write_indices_35 = tile(reps = write_indices_35_reps_0, x = write_indices_float_71_cast_fp16_to_int32)[name = tensor("write_indices_35")]; + tensor var_7597_begin_0 = const()[name = tensor("op_7597_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_7597_end_0 = const()[name = tensor("op_7597_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor var_7597_end_mask_0 = const()[name = tensor("op_7597_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_7597_squeeze_mask_0 = const()[name = tensor("op_7597_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor cache17_to_fp16_dtype_0 = const()[name = tensor("cache17_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor cache17_to_fp16 = cast(dtype = cache17_to_fp16_dtype_0, x = cache17)[name = tensor("cast_466")]; + tensor var_7597_cast_fp16 = slice_by_index(begin = var_7597_begin_0, end = var_7597_end_0, end_mask = var_7597_end_mask_0, squeeze_mask = var_7597_squeeze_mask_0, x = cache17_to_fp16)[name = tensor("op_7597_cast_fp16")]; + tensor var_7599_axis_0 = const()[name = tensor("op_7599_axis_0"), val = tensor(1)]; + tensor var_7599_mode_0 = const()[name = tensor("op_7599_mode_0"), val = tensor("update")]; + tensor var_7599_validate_indices_0 = const()[name = tensor("op_7599_validate_indices_0"), val = tensor(false)]; + tensor var_7599_cast_fp16 = scatter_along_axis(axis = var_7599_axis_0, data = var_7597_cast_fp16, indices = write_indices_35, mode = var_7599_mode_0, updates = k_71_cast_fp16, validate_indices = var_7599_validate_indices_0)[name = tensor("op_7599_cast_fp16")]; + tensor concat_121 = const()[name = tensor("concat_121"), val = tensor([0, 0, 0, 0, 0])]; + tensor concat_122 = const()[name = tensor("concat_122"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_35_internal_tensor_assign_1_stride_0 = const()[name = tensor("new_cache_35_internal_tensor_assign_1_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_35_internal_tensor_assign_1_begin_mask_0 = const()[name = tensor("new_cache_35_internal_tensor_assign_1_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_35_internal_tensor_assign_1_end_mask_0 = const()[name = tensor("new_cache_35_internal_tensor_assign_1_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_35_internal_tensor_assign_1_squeeze_mask_0 = const()[name = tensor("new_cache_35_internal_tensor_assign_1_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_82 = const()[name = tensor("shape_82"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_34 = const()[name = tensor("reduce_prod_34"), val = tensor(1048576)]; + tensor range_1d_34_start_0 = const()[name = tensor("range_1d_34_start_0"), val = tensor(0)]; + tensor range_1d_34_step_0 = const()[name = tensor("range_1d_34_step_0"), val = tensor(1)]; + tensor range_1d_34 = range_1d(end = reduce_prod_34, start = range_1d_34_start_0, step = range_1d_34_step_0)[name = tensor("range_1d_34")]; + tensor reshape_170 = reshape(shape = shape_82, x = range_1d_34)[name = tensor("reshape_170")]; + tensor slice_by_index_34 = slice_by_index(begin = concat_121, begin_mask = new_cache_35_internal_tensor_assign_1_begin_mask_0, end = concat_122, end_mask = new_cache_35_internal_tensor_assign_1_end_mask_0, squeeze_mask = new_cache_35_internal_tensor_assign_1_squeeze_mask_0, stride = new_cache_35_internal_tensor_assign_1_stride_0, x = reshape_170)[name = tensor("slice_by_index_34")]; + tensor reshape_171_shape_0 = const()[name = tensor("reshape_171_shape_0"), val = tensor([-1])]; + tensor reshape_171 = reshape(shape = reshape_171_shape_0, x = slice_by_index_34)[name = tensor("reshape_171")]; + tensor reshape_172_shape_0 = const()[name = tensor("reshape_172_shape_0"), val = tensor([-1])]; + tensor reshape_172_cast_fp16 = reshape(shape = reshape_172_shape_0, x = var_7599_cast_fp16)[name = tensor("reshape_172_cast_fp16")]; + tensor reshape_173_shape_0 = const()[name = tensor("reshape_173_shape_0"), val = tensor([-1])]; + tensor reshape_173_cast_fp16 = reshape(shape = reshape_173_shape_0, x = cache17_to_fp16)[name = tensor("reshape_173_cast_fp16")]; + tensor scatter_34_mode_0 = const()[name = tensor("scatter_34_mode_0"), val = tensor("update")]; + tensor scatter_34_axis_0 = const()[name = tensor("scatter_34_axis_0"), val = tensor(0)]; + tensor scatter_34_validate_indices_0 = const()[name = tensor("scatter_34_validate_indices_0"), val = tensor(false)]; + tensor scatter_34_cast_fp16 = scatter(axis = scatter_34_axis_0, data = reshape_173_cast_fp16, indices = reshape_171, mode = scatter_34_mode_0, updates = reshape_172_cast_fp16, validate_indices = scatter_34_validate_indices_0)[name = tensor("scatter_34_cast_fp16")]; + tensor reshape_174_cast_fp16 = reshape(shape = shape_82, x = scatter_34_cast_fp16)[name = tensor("reshape_174_cast_fp16")]; + tensor var_7607_begin_0 = const()[name = tensor("op_7607_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor var_7607_end_0 = const()[name = tensor("op_7607_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor var_7607_end_mask_0 = const()[name = tensor("op_7607_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_7607_squeeze_mask_0 = const()[name = tensor("op_7607_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor var_7607_cast_fp16 = slice_by_index(begin = var_7607_begin_0, end = var_7607_end_0, end_mask = var_7607_end_mask_0, squeeze_mask = var_7607_squeeze_mask_0, x = reshape_174_cast_fp16)[name = tensor("op_7607_cast_fp16")]; + tensor var_7609_axis_0 = const()[name = tensor("op_7609_axis_0"), val = tensor(1)]; + tensor var_7609_mode_0 = const()[name = tensor("op_7609_mode_0"), val = tensor("update")]; + tensor var_7609_validate_indices_0 = const()[name = tensor("op_7609_validate_indices_0"), val = tensor(false)]; + tensor var_7609_cast_fp16 = scatter_along_axis(axis = var_7609_axis_0, data = var_7607_cast_fp16, indices = write_indices_35, mode = var_7609_mode_0, updates = v_35_cast_fp16, validate_indices = var_7609_validate_indices_0)[name = tensor("op_7609_cast_fp16")]; + tensor concat_123 = const()[name = tensor("concat_123"), val = tensor([1, 0, 0, 0, 0])]; + tensor concat_124 = const()[name = tensor("concat_124"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_35_internal_tensor_assign_2_stride_0 = const()[name = tensor("new_cache_35_internal_tensor_assign_2_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_35_internal_tensor_assign_2_begin_mask_0 = const()[name = tensor("new_cache_35_internal_tensor_assign_2_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_35_internal_tensor_assign_2_end_mask_0 = const()[name = tensor("new_cache_35_internal_tensor_assign_2_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_35_internal_tensor_assign_2_squeeze_mask_0 = const()[name = tensor("new_cache_35_internal_tensor_assign_2_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_83 = const()[name = tensor("shape_83"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_35 = const()[name = tensor("reduce_prod_35"), val = tensor(1048576)]; + tensor range_1d_35_start_0 = const()[name = tensor("range_1d_35_start_0"), val = tensor(0)]; + tensor range_1d_35_step_0 = const()[name = tensor("range_1d_35_step_0"), val = tensor(1)]; + tensor range_1d_35 = range_1d(end = reduce_prod_35, start = range_1d_35_start_0, step = range_1d_35_step_0)[name = tensor("range_1d_35")]; + tensor reshape_175 = reshape(shape = shape_83, x = range_1d_35)[name = tensor("reshape_175")]; + tensor slice_by_index_35 = slice_by_index(begin = concat_123, begin_mask = new_cache_35_internal_tensor_assign_2_begin_mask_0, end = concat_124, end_mask = new_cache_35_internal_tensor_assign_2_end_mask_0, squeeze_mask = new_cache_35_internal_tensor_assign_2_squeeze_mask_0, stride = new_cache_35_internal_tensor_assign_2_stride_0, x = reshape_175)[name = tensor("slice_by_index_35")]; + tensor reshape_176_shape_0 = const()[name = tensor("reshape_176_shape_0"), val = tensor([-1])]; + tensor reshape_176 = reshape(shape = reshape_176_shape_0, x = slice_by_index_35)[name = tensor("reshape_176")]; + tensor reshape_177_shape_0 = const()[name = tensor("reshape_177_shape_0"), val = tensor([-1])]; + tensor reshape_177_cast_fp16 = reshape(shape = reshape_177_shape_0, x = var_7609_cast_fp16)[name = tensor("reshape_177_cast_fp16")]; + tensor reshape_178_shape_0 = const()[name = tensor("reshape_178_shape_0"), val = tensor([-1])]; + tensor reshape_178_cast_fp16 = reshape(shape = reshape_178_shape_0, x = reshape_174_cast_fp16)[name = tensor("reshape_178_cast_fp16")]; + tensor scatter_35_mode_0 = const()[name = tensor("scatter_35_mode_0"), val = tensor("update")]; + tensor scatter_35_axis_0 = const()[name = tensor("scatter_35_axis_0"), val = tensor(0)]; + tensor scatter_35_validate_indices_0 = const()[name = tensor("scatter_35_validate_indices_0"), val = tensor(false)]; + tensor scatter_35_cast_fp16 = scatter(axis = scatter_35_axis_0, data = reshape_178_cast_fp16, indices = reshape_176, mode = scatter_35_mode_0, updates = reshape_177_cast_fp16, validate_indices = scatter_35_validate_indices_0)[name = tensor("scatter_35_cast_fp16")]; + tensor reshape_179_cast_fp16 = reshape(shape = shape_83, x = scatter_35_cast_fp16)[name = tensor("reshape_179_cast_fp16")]; + tensor reshape_179_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("reshape_179_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor keys_103_begin_0 = const()[name = tensor("keys_103_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor keys_103_end_0 = const()[name = tensor("keys_103_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor keys_103_end_mask_0 = const()[name = tensor("keys_103_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor keys_103_squeeze_mask_0 = const()[name = tensor("keys_103_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor keys_103_cast_fp16 = slice_by_index(begin = keys_103_begin_0, end = keys_103_end_0, end_mask = keys_103_end_mask_0, squeeze_mask = keys_103_squeeze_mask_0, x = reshape_179_cast_fp16)[name = tensor("keys_103_cast_fp16")]; + tensor values_103_begin_0 = const()[name = tensor("values_103_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor values_103_end_0 = const()[name = tensor("values_103_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor values_103_end_mask_0 = const()[name = tensor("values_103_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor values_103_squeeze_mask_0 = const()[name = tensor("values_103_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor values_103_cast_fp16 = slice_by_index(begin = values_103_begin_0, end = values_103_end_0, end_mask = values_103_end_mask_0, squeeze_mask = values_103_squeeze_mask_0, x = reshape_179_cast_fp16)[name = tensor("values_103_cast_fp16")]; + tensor var_7621_cast_fp16 = not_equal(x = keys_103_cast_fp16, y = keys_103_cast_fp16)[name = tensor("op_7621_cast_fp16")]; + tensor keys_105_cast_fp16 = select(a = var_504_to_fp16, b = keys_103_cast_fp16, cond = var_7621_cast_fp16)[name = tensor("keys_105_cast_fp16")]; + tensor var_7629_cast_fp16 = not_equal(x = values_103_cast_fp16, y = values_103_cast_fp16)[name = tensor("op_7629_cast_fp16")]; + tensor values_105_cast_fp16 = select(a = var_504_to_fp16, b = values_103_cast_fp16, cond = var_7629_cast_fp16)[name = tensor("values_105_cast_fp16")]; + tensor var_7653 = const()[name = tensor("op_7653"), val = tensor([0, 2, 1, 3])]; + tensor var_7666 = const()[name = tensor("op_7666"), val = tensor([1, 1, 1])]; + tensor var_7667_cast_fp16 = reshape(shape = var_7666, x = position17_to_fp16)[name = tensor("op_7667_cast_fp16")]; + tensor var_7684_to_fp16 = const()[name = tensor("op_7684_to_fp16"), val = tensor(0x1p+0)]; + tensor valid_len_35_cast_fp16 = add(x = var_7667_cast_fp16, y = var_7684_to_fp16)[name = tensor("valid_len_35_cast_fp16")]; + tensor valid_mask_35_cast_fp16 = less(x = k_positions_1_promoted_to_fp16, y = valid_len_35_cast_fp16)[name = tensor("valid_mask_35_cast_fp16")]; + tensor causal_mask_35_cast_fp16 = less_equal(x = k_positions_1_promoted_to_fp16, y = var_7667_cast_fp16)[name = tensor("causal_mask_35_cast_fp16")]; + tensor attn_mask_69 = logical_and(x = valid_mask_35_cast_fp16, y = causal_mask_35_cast_fp16)[name = tensor("attn_mask_69")]; + tensor attn_mask_71_axes_0 = const()[name = tensor("attn_mask_71_axes_0"), val = tensor([1])]; + tensor attn_mask_71 = expand_dims(axes = attn_mask_71_axes_0, x = attn_mask_69)[name = tensor("attn_mask_71")]; + tensor var_7702_transpose_x_0 = const()[name = tensor("op_7702_transpose_x_0"), val = tensor(false)]; + tensor var_7702_transpose_y_0 = const()[name = tensor("op_7702_transpose_y_0"), val = tensor(false)]; + tensor transpose_106_perm_0 = const()[name = tensor("transpose_106_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_107_perm_0 = const()[name = tensor("transpose_107_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_107 = transpose(perm = transpose_107_perm_0, x = keys_105_cast_fp16)[name = tensor("transpose_145")]; + tensor transpose_106 = transpose(perm = transpose_106_perm_0, x = q_105_cast_fp16)[name = tensor("transpose_146")]; + tensor var_7702_cast_fp16 = matmul(transpose_x = var_7702_transpose_x_0, transpose_y = var_7702_transpose_y_0, x = transpose_106, y = transpose_107)[name = tensor("op_7702_cast_fp16")]; + tensor var_7696_to_fp16 = const()[name = tensor("op_7696_to_fp16"), val = tensor([0x1p-3])]; + tensor attn_weights_103_cast_fp16 = mul(x = var_7702_cast_fp16, y = var_7696_to_fp16)[name = tensor("attn_weights_103_cast_fp16")]; + tensor var_7704 = logical_not(x = attn_mask_71)[name = tensor("op_7704")]; + tensor var_7705_to_fp16 = const()[name = tensor("op_7705_to_fp16"), val = tensor(-inf)]; + tensor attn_weights_105_cast_fp16 = select(a = var_7705_to_fp16, b = attn_weights_103_cast_fp16, cond = var_7704)[name = tensor("attn_weights_105_cast_fp16")]; + tensor var_7707 = const()[name = tensor("op_7707"), val = tensor(-1)]; + tensor attn_weights_107_cast_fp16 = softmax(axis = var_7707, x = attn_weights_105_cast_fp16)[name = tensor("attn_weights_107_cast_fp16")]; + tensor attn_output_35_transpose_x_0 = const()[name = tensor("attn_output_35_transpose_x_0"), val = tensor(false)]; + tensor attn_output_35_transpose_y_0 = const()[name = tensor("attn_output_35_transpose_y_0"), val = tensor(false)]; + tensor values_107_cast_fp16 = transpose(perm = var_7653, x = values_105_cast_fp16)[name = tensor("transpose_147")]; + tensor attn_output_35_cast_fp16 = matmul(transpose_x = attn_output_35_transpose_x_0, transpose_y = attn_output_35_transpose_y_0, x = attn_weights_107_cast_fp16, y = values_107_cast_fp16)[name = tensor("attn_output_35_cast_fp16")]; + tensor var_7715 = const()[name = tensor("op_7715"), val = tensor([0, 2, 1, 3])]; + tensor var_7718 = const()[name = tensor("op_7718"), val = tensor([1, 1, 1024])]; + tensor var_7716_cast_fp16 = transpose(perm = var_7715, x = attn_output_35_cast_fp16)[name = tensor("transpose_144")]; + tensor input_175_cast_fp16 = reshape(shape = var_7718, x = var_7716_cast_fp16)[name = tensor("input_175_cast_fp16")]; + tensor linear_70_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = attn17_out_proj_weight_cast_fp16, x = input_175_cast_fp16)[name = tensor("linear_70_cast_fp16")]; + tensor var_7724_to_fp16 = const()[name = tensor("op_7724_to_fp16"), val = tensor(0x1p+0)]; + tensor var_7725_cast_fp16 = add(x = position17_to_fp16, y = var_7724_to_fp16)[name = tensor("op_7725_cast_fp16")]; + tensor var_7725_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_7725_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor input_177_cast_fp16 = add(x = input_173_cast_fp16, y = linear_70_cast_fp16)[name = tensor("input_177_cast_fp16")]; + tensor input_179_axes_0 = const()[name = tensor("input_179_axes_0"), val = tensor([-1])]; + tensor norm17_2_weight_to_fp16 = const()[name = tensor("norm17_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(303735040)))]; + tensor norm17_2_bias_to_fp16 = const()[name = tensor("norm17_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(303737152)))]; + tensor var_7729_to_fp16 = const()[name = tensor("op_7729_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_179_cast_fp16 = layer_norm(axes = input_179_axes_0, beta = norm17_2_bias_to_fp16, epsilon = var_7729_to_fp16, gamma = norm17_2_weight_to_fp16, x = input_177_cast_fp16)[name = tensor("input_179_cast_fp16")]; + tensor linear_71_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = linear17_1_weight_cast_fp16, x = input_179_cast_fp16)[name = tensor("linear_71_cast_fp16")]; + tensor input_181_mode_0 = const()[name = tensor("input_181_mode_0"), val = tensor("EXACT")]; + tensor input_181_cast_fp16 = gelu(mode = input_181_mode_0, x = linear_71_cast_fp16)[name = tensor("input_181_cast_fp16")]; + tensor linear_72_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = linear17_2_weight_cast_fp16, x = input_181_cast_fp16)[name = tensor("linear_72_cast_fp16")]; + tensor input_183_cast_fp16 = add(x = input_177_cast_fp16, y = linear_72_cast_fp16)[name = tensor("input_183_cast_fp16")]; + tensor x_37_axes_0 = const()[name = tensor("x_37_axes_0"), val = tensor([-1])]; + tensor norm18_1_weight_to_fp16 = const()[name = tensor("norm18_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(303739264)))]; + tensor norm18_1_bias_to_fp16 = const()[name = tensor("norm18_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(303741376)))]; + tensor var_7746_to_fp16 = const()[name = tensor("op_7746_to_fp16"), val = tensor(0x1.5p-17)]; + tensor x_37_cast_fp16 = layer_norm(axes = x_37_axes_0, beta = norm18_1_bias_to_fp16, epsilon = var_7746_to_fp16, gamma = norm18_1_weight_to_fp16, x = input_183_cast_fp16)[name = tensor("x_37_cast_fp16")]; + tensor linear_73_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = attn18_in_proj_weight_cast_fp16, x = x_37_cast_fp16)[name = tensor("linear_73_cast_fp16")]; + tensor var_7782 = const()[name = tensor("op_7782"), val = tensor([1, 1, 3, 16, 64])]; + tensor qkv_37_cast_fp16 = reshape(shape = var_7782, x = linear_73_cast_fp16)[name = tensor("qkv_37_cast_fp16")]; + tensor q_109_begin_0 = const()[name = tensor("q_109_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor q_109_end_0 = const()[name = tensor("q_109_end_0"), val = tensor([1, 1, 1, 16, 64])]; + tensor q_109_end_mask_0 = const()[name = tensor("q_109_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor q_109_squeeze_mask_0 = const()[name = tensor("q_109_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor q_109_cast_fp16 = slice_by_index(begin = q_109_begin_0, end = q_109_end_0, end_mask = q_109_end_mask_0, squeeze_mask = q_109_squeeze_mask_0, x = qkv_37_cast_fp16)[name = tensor("q_109_cast_fp16")]; + tensor k_73_begin_0 = const()[name = tensor("k_73_begin_0"), val = tensor([0, 0, 1, 0, 0])]; + tensor k_73_end_0 = const()[name = tensor("k_73_end_0"), val = tensor([1, 1, 2, 16, 64])]; + tensor k_73_end_mask_0 = const()[name = tensor("k_73_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor k_73_squeeze_mask_0 = const()[name = tensor("k_73_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor k_73_cast_fp16 = slice_by_index(begin = k_73_begin_0, end = k_73_end_0, end_mask = k_73_end_mask_0, squeeze_mask = k_73_squeeze_mask_0, x = qkv_37_cast_fp16)[name = tensor("k_73_cast_fp16")]; + tensor v_37_begin_0 = const()[name = tensor("v_37_begin_0"), val = tensor([0, 0, 2, 0, 0])]; + tensor v_37_end_0 = const()[name = tensor("v_37_end_0"), val = tensor([1, 1, 3, 16, 64])]; + tensor v_37_end_mask_0 = const()[name = tensor("v_37_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor v_37_squeeze_mask_0 = const()[name = tensor("v_37_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor v_37_cast_fp16 = slice_by_index(begin = v_37_begin_0, end = v_37_end_0, end_mask = v_37_end_mask_0, squeeze_mask = v_37_squeeze_mask_0, x = qkv_37_cast_fp16)[name = tensor("v_37_cast_fp16")]; + tensor var_7886 = const()[name = tensor("op_7886"), val = tensor([1, 1, 1, 1])]; + tensor position18_to_fp16_dtype_0 = const()[name = tensor("position18_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor position18_to_fp16 = cast(dtype = position18_to_fp16_dtype_0, x = position18)[name = tensor("cast_463")]; + tensor ts_113_cast_fp16 = reshape(shape = var_7886, x = position18_to_fp16)[name = tensor("ts_113_cast_fp16")]; + tensor var_7890 = const()[name = tensor("op_7890"), val = tensor([1, 1, 16, 32, 2])]; + tensor q_complex_37_cast_fp16 = reshape(shape = var_7890, x = q_109_cast_fp16)[name = tensor("q_complex_37_cast_fp16")]; + tensor var_7894 = const()[name = tensor("op_7894"), val = tensor([1, 1, 16, 32, 2])]; + tensor k_complex_37_cast_fp16 = reshape(shape = var_7894, x = k_73_cast_fp16)[name = tensor("k_complex_37_cast_fp16")]; + tensor var_7898_begin_0 = const()[name = tensor("op_7898_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_7898_end_0 = const()[name = tensor("op_7898_end_0"), val = tensor([1, 1, 16, 32, 1])]; + tensor var_7898_end_mask_0 = const()[name = tensor("op_7898_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_7898_squeeze_mask_0 = const()[name = tensor("op_7898_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_7898_cast_fp16 = slice_by_index(begin = var_7898_begin_0, end = var_7898_end_0, end_mask = var_7898_end_mask_0, squeeze_mask = var_7898_squeeze_mask_0, x = q_complex_37_cast_fp16)[name = tensor("op_7898_cast_fp16")]; + tensor var_7906_begin_0 = const()[name = tensor("op_7906_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_7906_end_0 = const()[name = tensor("op_7906_end_0"), val = tensor([1, 1, 16, 32, 2])]; + tensor var_7906_end_mask_0 = const()[name = tensor("op_7906_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_7906_squeeze_mask_0 = const()[name = tensor("op_7906_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_7906_cast_fp16 = slice_by_index(begin = var_7906_begin_0, end = var_7906_end_0, end_mask = var_7906_end_mask_0, squeeze_mask = var_7906_squeeze_mask_0, x = q_complex_37_cast_fp16)[name = tensor("op_7906_cast_fp16")]; + tensor var_7914_begin_0 = const()[name = tensor("op_7914_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_7914_end_0 = const()[name = tensor("op_7914_end_0"), val = tensor([1, 1, 16, 32, 1])]; + tensor var_7914_end_mask_0 = const()[name = tensor("op_7914_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_7914_squeeze_mask_0 = const()[name = tensor("op_7914_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_7914_cast_fp16 = slice_by_index(begin = var_7914_begin_0, end = var_7914_end_0, end_mask = var_7914_end_mask_0, squeeze_mask = var_7914_squeeze_mask_0, x = k_complex_37_cast_fp16)[name = tensor("op_7914_cast_fp16")]; + tensor var_7922_begin_0 = const()[name = tensor("op_7922_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_7922_end_0 = const()[name = tensor("op_7922_end_0"), val = tensor([1, 1, 16, 32, 2])]; + tensor var_7922_end_mask_0 = const()[name = tensor("op_7922_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_7922_squeeze_mask_0 = const()[name = tensor("op_7922_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_7922_cast_fp16 = slice_by_index(begin = var_7922_begin_0, end = var_7922_end_0, end_mask = var_7922_end_mask_0, squeeze_mask = var_7922_squeeze_mask_0, x = k_complex_37_cast_fp16)[name = tensor("op_7922_cast_fp16")]; + tensor freqs_37_to_fp16 = const()[name = tensor("freqs_37_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(303743488)))]; + tensor var_7928_cast_fp16 = mul(x = freqs_37_to_fp16, y = ts_113_cast_fp16)[name = tensor("op_7928_cast_fp16")]; + tensor rotr_37_cast_fp16 = cos(x = var_7928_cast_fp16)[name = tensor("rotr_37_cast_fp16")]; + tensor roti_37_cast_fp16 = sin(x = var_7928_cast_fp16)[name = tensor("roti_37_cast_fp16")]; + tensor var_7932_cast_fp16 = mul(x = var_7898_cast_fp16, y = rotr_37_cast_fp16)[name = tensor("op_7932_cast_fp16")]; + tensor var_7933_cast_fp16 = mul(x = var_7906_cast_fp16, y = roti_37_cast_fp16)[name = tensor("op_7933_cast_fp16")]; + tensor qor_73_cast_fp16 = sub(x = var_7932_cast_fp16, y = var_7933_cast_fp16)[name = tensor("qor_73_cast_fp16")]; + tensor var_7936_cast_fp16 = mul(x = var_7898_cast_fp16, y = roti_37_cast_fp16)[name = tensor("op_7936_cast_fp16")]; + tensor var_7937_cast_fp16 = mul(x = var_7906_cast_fp16, y = rotr_37_cast_fp16)[name = tensor("op_7937_cast_fp16")]; + tensor qoi_73_cast_fp16 = add(x = var_7936_cast_fp16, y = var_7937_cast_fp16)[name = tensor("qoi_73_cast_fp16")]; + tensor var_7940_cast_fp16 = mul(x = var_7914_cast_fp16, y = rotr_37_cast_fp16)[name = tensor("op_7940_cast_fp16")]; + tensor var_7941_cast_fp16 = mul(x = var_7922_cast_fp16, y = roti_37_cast_fp16)[name = tensor("op_7941_cast_fp16")]; + tensor kor_73_cast_fp16 = sub(x = var_7940_cast_fp16, y = var_7941_cast_fp16)[name = tensor("kor_73_cast_fp16")]; + tensor var_7944_cast_fp16 = mul(x = var_7914_cast_fp16, y = roti_37_cast_fp16)[name = tensor("op_7944_cast_fp16")]; + tensor var_7945_cast_fp16 = mul(x = var_7922_cast_fp16, y = rotr_37_cast_fp16)[name = tensor("op_7945_cast_fp16")]; + tensor koi_73_cast_fp16 = add(x = var_7944_cast_fp16, y = var_7945_cast_fp16)[name = tensor("koi_73_cast_fp16")]; + tensor qo_37_axis_0 = const()[name = tensor("qo_37_axis_0"), val = tensor(-1)]; + tensor qo_37_cast_fp16 = stack(axis = qo_37_axis_0, values = (qor_73_cast_fp16, qoi_73_cast_fp16))[name = tensor("qo_37_cast_fp16")]; + tensor ko_37_axis_0 = const()[name = tensor("ko_37_axis_0"), val = tensor(-1)]; + tensor ko_37_cast_fp16 = stack(axis = ko_37_axis_0, values = (kor_73_cast_fp16, koi_73_cast_fp16))[name = tensor("ko_37_cast_fp16")]; + tensor var_7974 = const()[name = tensor("op_7974"), val = tensor([1, 1, 16, 64])]; + tensor q_111_cast_fp16 = reshape(shape = var_7974, x = qo_37_cast_fp16)[name = tensor("q_111_cast_fp16")]; + tensor var_7976 = const()[name = tensor("op_7976"), val = tensor([1, 1, 16, 64])]; + tensor k_75_cast_fp16 = reshape(shape = var_7976, x = ko_37_cast_fp16)[name = tensor("k_75_cast_fp16")]; + tensor _inversed_7998_y_0_to_fp16 = const()[name = tensor("_inversed_7998_y_0_to_fp16"), val = tensor(0x1p-9)]; + tensor _inversed_7998_cast_fp16 = mul(x = ts_113_cast_fp16, y = _inversed_7998_y_0_to_fp16)[name = tensor("_inversed_7998_cast_fp16")]; + tensor var_7999_cast_fp16 = floor(x = _inversed_7998_cast_fp16)[name = tensor("op_7999_cast_fp16")]; + tensor var_8000_to_fp16 = const()[name = tensor("op_8000_to_fp16"), val = tensor(0x1p+9)]; + tensor var_8001_cast_fp16 = mul(x = var_7999_cast_fp16, y = var_8000_to_fp16)[name = tensor("op_8001_cast_fp16")]; + tensor write_indices_float_75_cast_fp16 = sub(x = ts_113_cast_fp16, y = var_8001_cast_fp16)[name = tensor("write_indices_float_75_cast_fp16")]; + tensor var_8008_dtype_0 = const()[name = tensor("op_8008_dtype_0"), val = tensor("int32")]; + tensor write_indices_37_reps_0 = const()[name = tensor("write_indices_37_reps_0"), val = tensor([1, 1, 16, 64])]; + tensor write_indices_float_75_cast_fp16_to_int32 = cast(dtype = var_8008_dtype_0, x = write_indices_float_75_cast_fp16)[name = tensor("cast_462")]; + tensor write_indices_37 = tile(reps = write_indices_37_reps_0, x = write_indices_float_75_cast_fp16_to_int32)[name = tensor("write_indices_37")]; + tensor var_8016_begin_0 = const()[name = tensor("op_8016_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_8016_end_0 = const()[name = tensor("op_8016_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor var_8016_end_mask_0 = const()[name = tensor("op_8016_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_8016_squeeze_mask_0 = const()[name = tensor("op_8016_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor cache18_to_fp16_dtype_0 = const()[name = tensor("cache18_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor cache18_to_fp16 = cast(dtype = cache18_to_fp16_dtype_0, x = cache18)[name = tensor("cast_461")]; + tensor var_8016_cast_fp16 = slice_by_index(begin = var_8016_begin_0, end = var_8016_end_0, end_mask = var_8016_end_mask_0, squeeze_mask = var_8016_squeeze_mask_0, x = cache18_to_fp16)[name = tensor("op_8016_cast_fp16")]; + tensor var_8018_axis_0 = const()[name = tensor("op_8018_axis_0"), val = tensor(1)]; + tensor var_8018_mode_0 = const()[name = tensor("op_8018_mode_0"), val = tensor("update")]; + tensor var_8018_validate_indices_0 = const()[name = tensor("op_8018_validate_indices_0"), val = tensor(false)]; + tensor var_8018_cast_fp16 = scatter_along_axis(axis = var_8018_axis_0, data = var_8016_cast_fp16, indices = write_indices_37, mode = var_8018_mode_0, updates = k_75_cast_fp16, validate_indices = var_8018_validate_indices_0)[name = tensor("op_8018_cast_fp16")]; + tensor concat_128 = const()[name = tensor("concat_128"), val = tensor([0, 0, 0, 0, 0])]; + tensor concat_129 = const()[name = tensor("concat_129"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_37_internal_tensor_assign_1_stride_0 = const()[name = tensor("new_cache_37_internal_tensor_assign_1_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_37_internal_tensor_assign_1_begin_mask_0 = const()[name = tensor("new_cache_37_internal_tensor_assign_1_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_37_internal_tensor_assign_1_end_mask_0 = const()[name = tensor("new_cache_37_internal_tensor_assign_1_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_37_internal_tensor_assign_1_squeeze_mask_0 = const()[name = tensor("new_cache_37_internal_tensor_assign_1_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_84 = const()[name = tensor("shape_84"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_36 = const()[name = tensor("reduce_prod_36"), val = tensor(1048576)]; + tensor range_1d_36_start_0 = const()[name = tensor("range_1d_36_start_0"), val = tensor(0)]; + tensor range_1d_36_step_0 = const()[name = tensor("range_1d_36_step_0"), val = tensor(1)]; + tensor range_1d_36 = range_1d(end = reduce_prod_36, start = range_1d_36_start_0, step = range_1d_36_step_0)[name = tensor("range_1d_36")]; + tensor reshape_180 = reshape(shape = shape_84, x = range_1d_36)[name = tensor("reshape_180")]; + tensor slice_by_index_36 = slice_by_index(begin = concat_128, begin_mask = new_cache_37_internal_tensor_assign_1_begin_mask_0, end = concat_129, end_mask = new_cache_37_internal_tensor_assign_1_end_mask_0, squeeze_mask = new_cache_37_internal_tensor_assign_1_squeeze_mask_0, stride = new_cache_37_internal_tensor_assign_1_stride_0, x = reshape_180)[name = tensor("slice_by_index_36")]; + tensor reshape_181_shape_0 = const()[name = tensor("reshape_181_shape_0"), val = tensor([-1])]; + tensor reshape_181 = reshape(shape = reshape_181_shape_0, x = slice_by_index_36)[name = tensor("reshape_181")]; + tensor reshape_182_shape_0 = const()[name = tensor("reshape_182_shape_0"), val = tensor([-1])]; + tensor reshape_182_cast_fp16 = reshape(shape = reshape_182_shape_0, x = var_8018_cast_fp16)[name = tensor("reshape_182_cast_fp16")]; + tensor reshape_183_shape_0 = const()[name = tensor("reshape_183_shape_0"), val = tensor([-1])]; + tensor reshape_183_cast_fp16 = reshape(shape = reshape_183_shape_0, x = cache18_to_fp16)[name = tensor("reshape_183_cast_fp16")]; + tensor scatter_36_mode_0 = const()[name = tensor("scatter_36_mode_0"), val = tensor("update")]; + tensor scatter_36_axis_0 = const()[name = tensor("scatter_36_axis_0"), val = tensor(0)]; + tensor scatter_36_validate_indices_0 = const()[name = tensor("scatter_36_validate_indices_0"), val = tensor(false)]; + tensor scatter_36_cast_fp16 = scatter(axis = scatter_36_axis_0, data = reshape_183_cast_fp16, indices = reshape_181, mode = scatter_36_mode_0, updates = reshape_182_cast_fp16, validate_indices = scatter_36_validate_indices_0)[name = tensor("scatter_36_cast_fp16")]; + tensor reshape_184_cast_fp16 = reshape(shape = shape_84, x = scatter_36_cast_fp16)[name = tensor("reshape_184_cast_fp16")]; + tensor var_8026_begin_0 = const()[name = tensor("op_8026_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor var_8026_end_0 = const()[name = tensor("op_8026_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor var_8026_end_mask_0 = const()[name = tensor("op_8026_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_8026_squeeze_mask_0 = const()[name = tensor("op_8026_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor var_8026_cast_fp16 = slice_by_index(begin = var_8026_begin_0, end = var_8026_end_0, end_mask = var_8026_end_mask_0, squeeze_mask = var_8026_squeeze_mask_0, x = reshape_184_cast_fp16)[name = tensor("op_8026_cast_fp16")]; + tensor var_8028_axis_0 = const()[name = tensor("op_8028_axis_0"), val = tensor(1)]; + tensor var_8028_mode_0 = const()[name = tensor("op_8028_mode_0"), val = tensor("update")]; + tensor var_8028_validate_indices_0 = const()[name = tensor("op_8028_validate_indices_0"), val = tensor(false)]; + tensor var_8028_cast_fp16 = scatter_along_axis(axis = var_8028_axis_0, data = var_8026_cast_fp16, indices = write_indices_37, mode = var_8028_mode_0, updates = v_37_cast_fp16, validate_indices = var_8028_validate_indices_0)[name = tensor("op_8028_cast_fp16")]; + tensor concat_130 = const()[name = tensor("concat_130"), val = tensor([1, 0, 0, 0, 0])]; + tensor concat_131 = const()[name = tensor("concat_131"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_37_internal_tensor_assign_2_stride_0 = const()[name = tensor("new_cache_37_internal_tensor_assign_2_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_37_internal_tensor_assign_2_begin_mask_0 = const()[name = tensor("new_cache_37_internal_tensor_assign_2_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_37_internal_tensor_assign_2_end_mask_0 = const()[name = tensor("new_cache_37_internal_tensor_assign_2_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_37_internal_tensor_assign_2_squeeze_mask_0 = const()[name = tensor("new_cache_37_internal_tensor_assign_2_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_85 = const()[name = tensor("shape_85"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_37 = const()[name = tensor("reduce_prod_37"), val = tensor(1048576)]; + tensor range_1d_37_start_0 = const()[name = tensor("range_1d_37_start_0"), val = tensor(0)]; + tensor range_1d_37_step_0 = const()[name = tensor("range_1d_37_step_0"), val = tensor(1)]; + tensor range_1d_37 = range_1d(end = reduce_prod_37, start = range_1d_37_start_0, step = range_1d_37_step_0)[name = tensor("range_1d_37")]; + tensor reshape_185 = reshape(shape = shape_85, x = range_1d_37)[name = tensor("reshape_185")]; + tensor slice_by_index_37 = slice_by_index(begin = concat_130, begin_mask = new_cache_37_internal_tensor_assign_2_begin_mask_0, end = concat_131, end_mask = new_cache_37_internal_tensor_assign_2_end_mask_0, squeeze_mask = new_cache_37_internal_tensor_assign_2_squeeze_mask_0, stride = new_cache_37_internal_tensor_assign_2_stride_0, x = reshape_185)[name = tensor("slice_by_index_37")]; + tensor reshape_186_shape_0 = const()[name = tensor("reshape_186_shape_0"), val = tensor([-1])]; + tensor reshape_186 = reshape(shape = reshape_186_shape_0, x = slice_by_index_37)[name = tensor("reshape_186")]; + tensor reshape_187_shape_0 = const()[name = tensor("reshape_187_shape_0"), val = tensor([-1])]; + tensor reshape_187_cast_fp16 = reshape(shape = reshape_187_shape_0, x = var_8028_cast_fp16)[name = tensor("reshape_187_cast_fp16")]; + tensor reshape_188_shape_0 = const()[name = tensor("reshape_188_shape_0"), val = tensor([-1])]; + tensor reshape_188_cast_fp16 = reshape(shape = reshape_188_shape_0, x = reshape_184_cast_fp16)[name = tensor("reshape_188_cast_fp16")]; + tensor scatter_37_mode_0 = const()[name = tensor("scatter_37_mode_0"), val = tensor("update")]; + tensor scatter_37_axis_0 = const()[name = tensor("scatter_37_axis_0"), val = tensor(0)]; + tensor scatter_37_validate_indices_0 = const()[name = tensor("scatter_37_validate_indices_0"), val = tensor(false)]; + tensor scatter_37_cast_fp16 = scatter(axis = scatter_37_axis_0, data = reshape_188_cast_fp16, indices = reshape_186, mode = scatter_37_mode_0, updates = reshape_187_cast_fp16, validate_indices = scatter_37_validate_indices_0)[name = tensor("scatter_37_cast_fp16")]; + tensor reshape_189_cast_fp16 = reshape(shape = shape_85, x = scatter_37_cast_fp16)[name = tensor("reshape_189_cast_fp16")]; + tensor reshape_189_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("reshape_189_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor keys_109_begin_0 = const()[name = tensor("keys_109_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor keys_109_end_0 = const()[name = tensor("keys_109_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor keys_109_end_mask_0 = const()[name = tensor("keys_109_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor keys_109_squeeze_mask_0 = const()[name = tensor("keys_109_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor keys_109_cast_fp16 = slice_by_index(begin = keys_109_begin_0, end = keys_109_end_0, end_mask = keys_109_end_mask_0, squeeze_mask = keys_109_squeeze_mask_0, x = reshape_189_cast_fp16)[name = tensor("keys_109_cast_fp16")]; + tensor values_109_begin_0 = const()[name = tensor("values_109_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor values_109_end_0 = const()[name = tensor("values_109_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor values_109_end_mask_0 = const()[name = tensor("values_109_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor values_109_squeeze_mask_0 = const()[name = tensor("values_109_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor values_109_cast_fp16 = slice_by_index(begin = values_109_begin_0, end = values_109_end_0, end_mask = values_109_end_mask_0, squeeze_mask = values_109_squeeze_mask_0, x = reshape_189_cast_fp16)[name = tensor("values_109_cast_fp16")]; + tensor var_8040_cast_fp16 = not_equal(x = keys_109_cast_fp16, y = keys_109_cast_fp16)[name = tensor("op_8040_cast_fp16")]; + tensor keys_111_cast_fp16 = select(a = var_504_to_fp16, b = keys_109_cast_fp16, cond = var_8040_cast_fp16)[name = tensor("keys_111_cast_fp16")]; + tensor var_8048_cast_fp16 = not_equal(x = values_109_cast_fp16, y = values_109_cast_fp16)[name = tensor("op_8048_cast_fp16")]; + tensor values_111_cast_fp16 = select(a = var_504_to_fp16, b = values_109_cast_fp16, cond = var_8048_cast_fp16)[name = tensor("values_111_cast_fp16")]; + tensor var_8072 = const()[name = tensor("op_8072"), val = tensor([0, 2, 1, 3])]; + tensor var_8085 = const()[name = tensor("op_8085"), val = tensor([1, 1, 1])]; + tensor var_8086_cast_fp16 = reshape(shape = var_8085, x = position18_to_fp16)[name = tensor("op_8086_cast_fp16")]; + tensor var_8103_to_fp16 = const()[name = tensor("op_8103_to_fp16"), val = tensor(0x1p+0)]; + tensor valid_len_37_cast_fp16 = add(x = var_8086_cast_fp16, y = var_8103_to_fp16)[name = tensor("valid_len_37_cast_fp16")]; + tensor valid_mask_37_cast_fp16 = less(x = k_positions_1_promoted_to_fp16, y = valid_len_37_cast_fp16)[name = tensor("valid_mask_37_cast_fp16")]; + tensor causal_mask_37_cast_fp16 = less_equal(x = k_positions_1_promoted_to_fp16, y = var_8086_cast_fp16)[name = tensor("causal_mask_37_cast_fp16")]; + tensor attn_mask_73 = logical_and(x = valid_mask_37_cast_fp16, y = causal_mask_37_cast_fp16)[name = tensor("attn_mask_73")]; + tensor attn_mask_75_axes_0 = const()[name = tensor("attn_mask_75_axes_0"), val = tensor([1])]; + tensor attn_mask_75 = expand_dims(axes = attn_mask_75_axes_0, x = attn_mask_73)[name = tensor("attn_mask_75")]; + tensor var_8121_transpose_x_0 = const()[name = tensor("op_8121_transpose_x_0"), val = tensor(false)]; + tensor var_8121_transpose_y_0 = const()[name = tensor("op_8121_transpose_y_0"), val = tensor(false)]; + tensor transpose_108_perm_0 = const()[name = tensor("transpose_108_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_109_perm_0 = const()[name = tensor("transpose_109_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_109 = transpose(perm = transpose_109_perm_0, x = keys_111_cast_fp16)[name = tensor("transpose_141")]; + tensor transpose_108 = transpose(perm = transpose_108_perm_0, x = q_111_cast_fp16)[name = tensor("transpose_142")]; + tensor var_8121_cast_fp16 = matmul(transpose_x = var_8121_transpose_x_0, transpose_y = var_8121_transpose_y_0, x = transpose_108, y = transpose_109)[name = tensor("op_8121_cast_fp16")]; + tensor var_8115_to_fp16 = const()[name = tensor("op_8115_to_fp16"), val = tensor([0x1p-3])]; + tensor attn_weights_109_cast_fp16 = mul(x = var_8121_cast_fp16, y = var_8115_to_fp16)[name = tensor("attn_weights_109_cast_fp16")]; + tensor var_8123 = logical_not(x = attn_mask_75)[name = tensor("op_8123")]; + tensor var_8124_to_fp16 = const()[name = tensor("op_8124_to_fp16"), val = tensor(-inf)]; + tensor attn_weights_111_cast_fp16 = select(a = var_8124_to_fp16, b = attn_weights_109_cast_fp16, cond = var_8123)[name = tensor("attn_weights_111_cast_fp16")]; + tensor var_8126 = const()[name = tensor("op_8126"), val = tensor(-1)]; + tensor attn_weights_113_cast_fp16 = softmax(axis = var_8126, x = attn_weights_111_cast_fp16)[name = tensor("attn_weights_113_cast_fp16")]; + tensor attn_output_37_transpose_x_0 = const()[name = tensor("attn_output_37_transpose_x_0"), val = tensor(false)]; + tensor attn_output_37_transpose_y_0 = const()[name = tensor("attn_output_37_transpose_y_0"), val = tensor(false)]; + tensor values_113_cast_fp16 = transpose(perm = var_8072, x = values_111_cast_fp16)[name = tensor("transpose_143")]; + tensor attn_output_37_cast_fp16 = matmul(transpose_x = attn_output_37_transpose_x_0, transpose_y = attn_output_37_transpose_y_0, x = attn_weights_113_cast_fp16, y = values_113_cast_fp16)[name = tensor("attn_output_37_cast_fp16")]; + tensor var_8134 = const()[name = tensor("op_8134"), val = tensor([0, 2, 1, 3])]; + tensor var_8137 = const()[name = tensor("op_8137"), val = tensor([1, 1, 1024])]; + tensor var_8135_cast_fp16 = transpose(perm = var_8134, x = attn_output_37_cast_fp16)[name = tensor("transpose_140")]; + tensor input_185_cast_fp16 = reshape(shape = var_8137, x = var_8135_cast_fp16)[name = tensor("input_185_cast_fp16")]; + tensor linear_74_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = attn18_out_proj_weight_cast_fp16, x = input_185_cast_fp16)[name = tensor("linear_74_cast_fp16")]; + tensor var_8143_to_fp16 = const()[name = tensor("op_8143_to_fp16"), val = tensor(0x1p+0)]; + tensor var_8144_cast_fp16 = add(x = position18_to_fp16, y = var_8143_to_fp16)[name = tensor("op_8144_cast_fp16")]; + tensor var_8144_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_8144_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor input_187_cast_fp16 = add(x = input_183_cast_fp16, y = linear_74_cast_fp16)[name = tensor("input_187_cast_fp16")]; + tensor input_189_axes_0 = const()[name = tensor("input_189_axes_0"), val = tensor([-1])]; + tensor norm18_2_weight_to_fp16 = const()[name = tensor("norm18_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(303743616)))]; + tensor norm18_2_bias_to_fp16 = const()[name = tensor("norm18_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(303745728)))]; + tensor var_8148_to_fp16 = const()[name = tensor("op_8148_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_189_cast_fp16 = layer_norm(axes = input_189_axes_0, beta = norm18_2_bias_to_fp16, epsilon = var_8148_to_fp16, gamma = norm18_2_weight_to_fp16, x = input_187_cast_fp16)[name = tensor("input_189_cast_fp16")]; + tensor linear_75_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = linear18_1_weight_cast_fp16, x = input_189_cast_fp16)[name = tensor("linear_75_cast_fp16")]; + tensor input_191_mode_0 = const()[name = tensor("input_191_mode_0"), val = tensor("EXACT")]; + tensor input_191_cast_fp16 = gelu(mode = input_191_mode_0, x = linear_75_cast_fp16)[name = tensor("input_191_cast_fp16")]; + tensor linear_76_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = linear18_2_weight_cast_fp16, x = input_191_cast_fp16)[name = tensor("linear_76_cast_fp16")]; + tensor input_193_cast_fp16 = add(x = input_187_cast_fp16, y = linear_76_cast_fp16)[name = tensor("input_193_cast_fp16")]; + tensor x_39_axes_0 = const()[name = tensor("x_39_axes_0"), val = tensor([-1])]; + tensor norm19_1_weight_to_fp16 = const()[name = tensor("norm19_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(303747840)))]; + tensor norm19_1_bias_to_fp16 = const()[name = tensor("norm19_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(303749952)))]; + tensor var_8165_to_fp16 = const()[name = tensor("op_8165_to_fp16"), val = tensor(0x1.5p-17)]; + tensor x_39_cast_fp16 = layer_norm(axes = x_39_axes_0, beta = norm19_1_bias_to_fp16, epsilon = var_8165_to_fp16, gamma = norm19_1_weight_to_fp16, x = input_193_cast_fp16)[name = tensor("x_39_cast_fp16")]; + tensor linear_77_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = attn19_in_proj_weight_cast_fp16, x = x_39_cast_fp16)[name = tensor("linear_77_cast_fp16")]; + tensor var_8201 = const()[name = tensor("op_8201"), val = tensor([1, 1, 3, 16, 64])]; + tensor qkv_39_cast_fp16 = reshape(shape = var_8201, x = linear_77_cast_fp16)[name = tensor("qkv_39_cast_fp16")]; + tensor q_115_begin_0 = const()[name = tensor("q_115_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor q_115_end_0 = const()[name = tensor("q_115_end_0"), val = tensor([1, 1, 1, 16, 64])]; + tensor q_115_end_mask_0 = const()[name = tensor("q_115_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor q_115_squeeze_mask_0 = const()[name = tensor("q_115_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor q_115_cast_fp16 = slice_by_index(begin = q_115_begin_0, end = q_115_end_0, end_mask = q_115_end_mask_0, squeeze_mask = q_115_squeeze_mask_0, x = qkv_39_cast_fp16)[name = tensor("q_115_cast_fp16")]; + tensor k_77_begin_0 = const()[name = tensor("k_77_begin_0"), val = tensor([0, 0, 1, 0, 0])]; + tensor k_77_end_0 = const()[name = tensor("k_77_end_0"), val = tensor([1, 1, 2, 16, 64])]; + tensor k_77_end_mask_0 = const()[name = tensor("k_77_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor k_77_squeeze_mask_0 = const()[name = tensor("k_77_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor k_77_cast_fp16 = slice_by_index(begin = k_77_begin_0, end = k_77_end_0, end_mask = k_77_end_mask_0, squeeze_mask = k_77_squeeze_mask_0, x = qkv_39_cast_fp16)[name = tensor("k_77_cast_fp16")]; + tensor v_39_begin_0 = const()[name = tensor("v_39_begin_0"), val = tensor([0, 0, 2, 0, 0])]; + tensor v_39_end_0 = const()[name = tensor("v_39_end_0"), val = tensor([1, 1, 3, 16, 64])]; + tensor v_39_end_mask_0 = const()[name = tensor("v_39_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor v_39_squeeze_mask_0 = const()[name = tensor("v_39_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor v_39_cast_fp16 = slice_by_index(begin = v_39_begin_0, end = v_39_end_0, end_mask = v_39_end_mask_0, squeeze_mask = v_39_squeeze_mask_0, x = qkv_39_cast_fp16)[name = tensor("v_39_cast_fp16")]; + tensor var_8305 = const()[name = tensor("op_8305"), val = tensor([1, 1, 1, 1])]; + tensor position19_to_fp16_dtype_0 = const()[name = tensor("position19_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor position19_to_fp16 = cast(dtype = position19_to_fp16_dtype_0, x = position19)[name = tensor("cast_458")]; + tensor ts_119_cast_fp16 = reshape(shape = var_8305, x = position19_to_fp16)[name = tensor("ts_119_cast_fp16")]; + tensor var_8309 = const()[name = tensor("op_8309"), val = tensor([1, 1, 16, 32, 2])]; + tensor q_complex_39_cast_fp16 = reshape(shape = var_8309, x = q_115_cast_fp16)[name = tensor("q_complex_39_cast_fp16")]; + tensor var_8313 = const()[name = tensor("op_8313"), val = tensor([1, 1, 16, 32, 2])]; + tensor k_complex_39_cast_fp16 = reshape(shape = var_8313, x = k_77_cast_fp16)[name = tensor("k_complex_39_cast_fp16")]; + tensor var_8317_begin_0 = const()[name = tensor("op_8317_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_8317_end_0 = const()[name = tensor("op_8317_end_0"), val = tensor([1, 1, 16, 32, 1])]; + tensor var_8317_end_mask_0 = const()[name = tensor("op_8317_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_8317_squeeze_mask_0 = const()[name = tensor("op_8317_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_8317_cast_fp16 = slice_by_index(begin = var_8317_begin_0, end = var_8317_end_0, end_mask = var_8317_end_mask_0, squeeze_mask = var_8317_squeeze_mask_0, x = q_complex_39_cast_fp16)[name = tensor("op_8317_cast_fp16")]; + tensor var_8325_begin_0 = const()[name = tensor("op_8325_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_8325_end_0 = const()[name = tensor("op_8325_end_0"), val = tensor([1, 1, 16, 32, 2])]; + tensor var_8325_end_mask_0 = const()[name = tensor("op_8325_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_8325_squeeze_mask_0 = const()[name = tensor("op_8325_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_8325_cast_fp16 = slice_by_index(begin = var_8325_begin_0, end = var_8325_end_0, end_mask = var_8325_end_mask_0, squeeze_mask = var_8325_squeeze_mask_0, x = q_complex_39_cast_fp16)[name = tensor("op_8325_cast_fp16")]; + tensor var_8333_begin_0 = const()[name = tensor("op_8333_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_8333_end_0 = const()[name = tensor("op_8333_end_0"), val = tensor([1, 1, 16, 32, 1])]; + tensor var_8333_end_mask_0 = const()[name = tensor("op_8333_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_8333_squeeze_mask_0 = const()[name = tensor("op_8333_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_8333_cast_fp16 = slice_by_index(begin = var_8333_begin_0, end = var_8333_end_0, end_mask = var_8333_end_mask_0, squeeze_mask = var_8333_squeeze_mask_0, x = k_complex_39_cast_fp16)[name = tensor("op_8333_cast_fp16")]; + tensor var_8341_begin_0 = const()[name = tensor("op_8341_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_8341_end_0 = const()[name = tensor("op_8341_end_0"), val = tensor([1, 1, 16, 32, 2])]; + tensor var_8341_end_mask_0 = const()[name = tensor("op_8341_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_8341_squeeze_mask_0 = const()[name = tensor("op_8341_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_8341_cast_fp16 = slice_by_index(begin = var_8341_begin_0, end = var_8341_end_0, end_mask = var_8341_end_mask_0, squeeze_mask = var_8341_squeeze_mask_0, x = k_complex_39_cast_fp16)[name = tensor("op_8341_cast_fp16")]; + tensor freqs_39_to_fp16 = const()[name = tensor("freqs_39_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(303752064)))]; + tensor var_8347_cast_fp16 = mul(x = freqs_39_to_fp16, y = ts_119_cast_fp16)[name = tensor("op_8347_cast_fp16")]; + tensor rotr_39_cast_fp16 = cos(x = var_8347_cast_fp16)[name = tensor("rotr_39_cast_fp16")]; + tensor roti_39_cast_fp16 = sin(x = var_8347_cast_fp16)[name = tensor("roti_39_cast_fp16")]; + tensor var_8351_cast_fp16 = mul(x = var_8317_cast_fp16, y = rotr_39_cast_fp16)[name = tensor("op_8351_cast_fp16")]; + tensor var_8352_cast_fp16 = mul(x = var_8325_cast_fp16, y = roti_39_cast_fp16)[name = tensor("op_8352_cast_fp16")]; + tensor qor_77_cast_fp16 = sub(x = var_8351_cast_fp16, y = var_8352_cast_fp16)[name = tensor("qor_77_cast_fp16")]; + tensor var_8355_cast_fp16 = mul(x = var_8317_cast_fp16, y = roti_39_cast_fp16)[name = tensor("op_8355_cast_fp16")]; + tensor var_8356_cast_fp16 = mul(x = var_8325_cast_fp16, y = rotr_39_cast_fp16)[name = tensor("op_8356_cast_fp16")]; + tensor qoi_77_cast_fp16 = add(x = var_8355_cast_fp16, y = var_8356_cast_fp16)[name = tensor("qoi_77_cast_fp16")]; + tensor var_8359_cast_fp16 = mul(x = var_8333_cast_fp16, y = rotr_39_cast_fp16)[name = tensor("op_8359_cast_fp16")]; + tensor var_8360_cast_fp16 = mul(x = var_8341_cast_fp16, y = roti_39_cast_fp16)[name = tensor("op_8360_cast_fp16")]; + tensor kor_77_cast_fp16 = sub(x = var_8359_cast_fp16, y = var_8360_cast_fp16)[name = tensor("kor_77_cast_fp16")]; + tensor var_8363_cast_fp16 = mul(x = var_8333_cast_fp16, y = roti_39_cast_fp16)[name = tensor("op_8363_cast_fp16")]; + tensor var_8364_cast_fp16 = mul(x = var_8341_cast_fp16, y = rotr_39_cast_fp16)[name = tensor("op_8364_cast_fp16")]; + tensor koi_77_cast_fp16 = add(x = var_8363_cast_fp16, y = var_8364_cast_fp16)[name = tensor("koi_77_cast_fp16")]; + tensor qo_39_axis_0 = const()[name = tensor("qo_39_axis_0"), val = tensor(-1)]; + tensor qo_39_cast_fp16 = stack(axis = qo_39_axis_0, values = (qor_77_cast_fp16, qoi_77_cast_fp16))[name = tensor("qo_39_cast_fp16")]; + tensor ko_39_axis_0 = const()[name = tensor("ko_39_axis_0"), val = tensor(-1)]; + tensor ko_39_cast_fp16 = stack(axis = ko_39_axis_0, values = (kor_77_cast_fp16, koi_77_cast_fp16))[name = tensor("ko_39_cast_fp16")]; + tensor var_8393 = const()[name = tensor("op_8393"), val = tensor([1, 1, 16, 64])]; + tensor q_117_cast_fp16 = reshape(shape = var_8393, x = qo_39_cast_fp16)[name = tensor("q_117_cast_fp16")]; + tensor var_8395 = const()[name = tensor("op_8395"), val = tensor([1, 1, 16, 64])]; + tensor k_79_cast_fp16 = reshape(shape = var_8395, x = ko_39_cast_fp16)[name = tensor("k_79_cast_fp16")]; + tensor _inversed_8417_y_0_to_fp16 = const()[name = tensor("_inversed_8417_y_0_to_fp16"), val = tensor(0x1p-9)]; + tensor _inversed_8417_cast_fp16 = mul(x = ts_119_cast_fp16, y = _inversed_8417_y_0_to_fp16)[name = tensor("_inversed_8417_cast_fp16")]; + tensor var_8418_cast_fp16 = floor(x = _inversed_8417_cast_fp16)[name = tensor("op_8418_cast_fp16")]; + tensor var_8419_to_fp16 = const()[name = tensor("op_8419_to_fp16"), val = tensor(0x1p+9)]; + tensor var_8420_cast_fp16 = mul(x = var_8418_cast_fp16, y = var_8419_to_fp16)[name = tensor("op_8420_cast_fp16")]; + tensor write_indices_float_79_cast_fp16 = sub(x = ts_119_cast_fp16, y = var_8420_cast_fp16)[name = tensor("write_indices_float_79_cast_fp16")]; + tensor var_8427_dtype_0 = const()[name = tensor("op_8427_dtype_0"), val = tensor("int32")]; + tensor write_indices_39_reps_0 = const()[name = tensor("write_indices_39_reps_0"), val = tensor([1, 1, 16, 64])]; + tensor write_indices_float_79_cast_fp16_to_int32 = cast(dtype = var_8427_dtype_0, x = write_indices_float_79_cast_fp16)[name = tensor("cast_457")]; + tensor write_indices_39 = tile(reps = write_indices_39_reps_0, x = write_indices_float_79_cast_fp16_to_int32)[name = tensor("write_indices_39")]; + tensor var_8435_begin_0 = const()[name = tensor("op_8435_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_8435_end_0 = const()[name = tensor("op_8435_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor var_8435_end_mask_0 = const()[name = tensor("op_8435_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_8435_squeeze_mask_0 = const()[name = tensor("op_8435_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor cache19_to_fp16_dtype_0 = const()[name = tensor("cache19_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor cache19_to_fp16 = cast(dtype = cache19_to_fp16_dtype_0, x = cache19)[name = tensor("cast_456")]; + tensor var_8435_cast_fp16 = slice_by_index(begin = var_8435_begin_0, end = var_8435_end_0, end_mask = var_8435_end_mask_0, squeeze_mask = var_8435_squeeze_mask_0, x = cache19_to_fp16)[name = tensor("op_8435_cast_fp16")]; + tensor var_8437_axis_0 = const()[name = tensor("op_8437_axis_0"), val = tensor(1)]; + tensor var_8437_mode_0 = const()[name = tensor("op_8437_mode_0"), val = tensor("update")]; + tensor var_8437_validate_indices_0 = const()[name = tensor("op_8437_validate_indices_0"), val = tensor(false)]; + tensor var_8437_cast_fp16 = scatter_along_axis(axis = var_8437_axis_0, data = var_8435_cast_fp16, indices = write_indices_39, mode = var_8437_mode_0, updates = k_79_cast_fp16, validate_indices = var_8437_validate_indices_0)[name = tensor("op_8437_cast_fp16")]; + tensor concat_135 = const()[name = tensor("concat_135"), val = tensor([0, 0, 0, 0, 0])]; + tensor concat_136 = const()[name = tensor("concat_136"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_39_internal_tensor_assign_1_stride_0 = const()[name = tensor("new_cache_39_internal_tensor_assign_1_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_39_internal_tensor_assign_1_begin_mask_0 = const()[name = tensor("new_cache_39_internal_tensor_assign_1_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_39_internal_tensor_assign_1_end_mask_0 = const()[name = tensor("new_cache_39_internal_tensor_assign_1_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_39_internal_tensor_assign_1_squeeze_mask_0 = const()[name = tensor("new_cache_39_internal_tensor_assign_1_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_86 = const()[name = tensor("shape_86"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_38 = const()[name = tensor("reduce_prod_38"), val = tensor(1048576)]; + tensor range_1d_38_start_0 = const()[name = tensor("range_1d_38_start_0"), val = tensor(0)]; + tensor range_1d_38_step_0 = const()[name = tensor("range_1d_38_step_0"), val = tensor(1)]; + tensor range_1d_38 = range_1d(end = reduce_prod_38, start = range_1d_38_start_0, step = range_1d_38_step_0)[name = tensor("range_1d_38")]; + tensor reshape_190 = reshape(shape = shape_86, x = range_1d_38)[name = tensor("reshape_190")]; + tensor slice_by_index_38 = slice_by_index(begin = concat_135, begin_mask = new_cache_39_internal_tensor_assign_1_begin_mask_0, end = concat_136, end_mask = new_cache_39_internal_tensor_assign_1_end_mask_0, squeeze_mask = new_cache_39_internal_tensor_assign_1_squeeze_mask_0, stride = new_cache_39_internal_tensor_assign_1_stride_0, x = reshape_190)[name = tensor("slice_by_index_38")]; + tensor reshape_191_shape_0 = const()[name = tensor("reshape_191_shape_0"), val = tensor([-1])]; + tensor reshape_191 = reshape(shape = reshape_191_shape_0, x = slice_by_index_38)[name = tensor("reshape_191")]; + tensor reshape_192_shape_0 = const()[name = tensor("reshape_192_shape_0"), val = tensor([-1])]; + tensor reshape_192_cast_fp16 = reshape(shape = reshape_192_shape_0, x = var_8437_cast_fp16)[name = tensor("reshape_192_cast_fp16")]; + tensor reshape_193_shape_0 = const()[name = tensor("reshape_193_shape_0"), val = tensor([-1])]; + tensor reshape_193_cast_fp16 = reshape(shape = reshape_193_shape_0, x = cache19_to_fp16)[name = tensor("reshape_193_cast_fp16")]; + tensor scatter_38_mode_0 = const()[name = tensor("scatter_38_mode_0"), val = tensor("update")]; + tensor scatter_38_axis_0 = const()[name = tensor("scatter_38_axis_0"), val = tensor(0)]; + tensor scatter_38_validate_indices_0 = const()[name = tensor("scatter_38_validate_indices_0"), val = tensor(false)]; + tensor scatter_38_cast_fp16 = scatter(axis = scatter_38_axis_0, data = reshape_193_cast_fp16, indices = reshape_191, mode = scatter_38_mode_0, updates = reshape_192_cast_fp16, validate_indices = scatter_38_validate_indices_0)[name = tensor("scatter_38_cast_fp16")]; + tensor reshape_194_cast_fp16 = reshape(shape = shape_86, x = scatter_38_cast_fp16)[name = tensor("reshape_194_cast_fp16")]; + tensor var_8445_begin_0 = const()[name = tensor("op_8445_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor var_8445_end_0 = const()[name = tensor("op_8445_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor var_8445_end_mask_0 = const()[name = tensor("op_8445_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_8445_squeeze_mask_0 = const()[name = tensor("op_8445_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor var_8445_cast_fp16 = slice_by_index(begin = var_8445_begin_0, end = var_8445_end_0, end_mask = var_8445_end_mask_0, squeeze_mask = var_8445_squeeze_mask_0, x = reshape_194_cast_fp16)[name = tensor("op_8445_cast_fp16")]; + tensor var_8447_axis_0 = const()[name = tensor("op_8447_axis_0"), val = tensor(1)]; + tensor var_8447_mode_0 = const()[name = tensor("op_8447_mode_0"), val = tensor("update")]; + tensor var_8447_validate_indices_0 = const()[name = tensor("op_8447_validate_indices_0"), val = tensor(false)]; + tensor var_8447_cast_fp16 = scatter_along_axis(axis = var_8447_axis_0, data = var_8445_cast_fp16, indices = write_indices_39, mode = var_8447_mode_0, updates = v_39_cast_fp16, validate_indices = var_8447_validate_indices_0)[name = tensor("op_8447_cast_fp16")]; + tensor concat_137 = const()[name = tensor("concat_137"), val = tensor([1, 0, 0, 0, 0])]; + tensor concat_138 = const()[name = tensor("concat_138"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_39_internal_tensor_assign_2_stride_0 = const()[name = tensor("new_cache_39_internal_tensor_assign_2_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_39_internal_tensor_assign_2_begin_mask_0 = const()[name = tensor("new_cache_39_internal_tensor_assign_2_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_39_internal_tensor_assign_2_end_mask_0 = const()[name = tensor("new_cache_39_internal_tensor_assign_2_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_39_internal_tensor_assign_2_squeeze_mask_0 = const()[name = tensor("new_cache_39_internal_tensor_assign_2_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_87 = const()[name = tensor("shape_87"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_39 = const()[name = tensor("reduce_prod_39"), val = tensor(1048576)]; + tensor range_1d_39_start_0 = const()[name = tensor("range_1d_39_start_0"), val = tensor(0)]; + tensor range_1d_39_step_0 = const()[name = tensor("range_1d_39_step_0"), val = tensor(1)]; + tensor range_1d_39 = range_1d(end = reduce_prod_39, start = range_1d_39_start_0, step = range_1d_39_step_0)[name = tensor("range_1d_39")]; + tensor reshape_195 = reshape(shape = shape_87, x = range_1d_39)[name = tensor("reshape_195")]; + tensor slice_by_index_39 = slice_by_index(begin = concat_137, begin_mask = new_cache_39_internal_tensor_assign_2_begin_mask_0, end = concat_138, end_mask = new_cache_39_internal_tensor_assign_2_end_mask_0, squeeze_mask = new_cache_39_internal_tensor_assign_2_squeeze_mask_0, stride = new_cache_39_internal_tensor_assign_2_stride_0, x = reshape_195)[name = tensor("slice_by_index_39")]; + tensor reshape_196_shape_0 = const()[name = tensor("reshape_196_shape_0"), val = tensor([-1])]; + tensor reshape_196 = reshape(shape = reshape_196_shape_0, x = slice_by_index_39)[name = tensor("reshape_196")]; + tensor reshape_197_shape_0 = const()[name = tensor("reshape_197_shape_0"), val = tensor([-1])]; + tensor reshape_197_cast_fp16 = reshape(shape = reshape_197_shape_0, x = var_8447_cast_fp16)[name = tensor("reshape_197_cast_fp16")]; + tensor reshape_198_shape_0 = const()[name = tensor("reshape_198_shape_0"), val = tensor([-1])]; + tensor reshape_198_cast_fp16 = reshape(shape = reshape_198_shape_0, x = reshape_194_cast_fp16)[name = tensor("reshape_198_cast_fp16")]; + tensor scatter_39_mode_0 = const()[name = tensor("scatter_39_mode_0"), val = tensor("update")]; + tensor scatter_39_axis_0 = const()[name = tensor("scatter_39_axis_0"), val = tensor(0)]; + tensor scatter_39_validate_indices_0 = const()[name = tensor("scatter_39_validate_indices_0"), val = tensor(false)]; + tensor scatter_39_cast_fp16 = scatter(axis = scatter_39_axis_0, data = reshape_198_cast_fp16, indices = reshape_196, mode = scatter_39_mode_0, updates = reshape_197_cast_fp16, validate_indices = scatter_39_validate_indices_0)[name = tensor("scatter_39_cast_fp16")]; + tensor reshape_199_cast_fp16 = reshape(shape = shape_87, x = scatter_39_cast_fp16)[name = tensor("reshape_199_cast_fp16")]; + tensor reshape_199_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("reshape_199_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor keys_115_begin_0 = const()[name = tensor("keys_115_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor keys_115_end_0 = const()[name = tensor("keys_115_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor keys_115_end_mask_0 = const()[name = tensor("keys_115_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor keys_115_squeeze_mask_0 = const()[name = tensor("keys_115_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor keys_115_cast_fp16 = slice_by_index(begin = keys_115_begin_0, end = keys_115_end_0, end_mask = keys_115_end_mask_0, squeeze_mask = keys_115_squeeze_mask_0, x = reshape_199_cast_fp16)[name = tensor("keys_115_cast_fp16")]; + tensor values_115_begin_0 = const()[name = tensor("values_115_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor values_115_end_0 = const()[name = tensor("values_115_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor values_115_end_mask_0 = const()[name = tensor("values_115_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor values_115_squeeze_mask_0 = const()[name = tensor("values_115_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor values_115_cast_fp16 = slice_by_index(begin = values_115_begin_0, end = values_115_end_0, end_mask = values_115_end_mask_0, squeeze_mask = values_115_squeeze_mask_0, x = reshape_199_cast_fp16)[name = tensor("values_115_cast_fp16")]; + tensor var_8459_cast_fp16 = not_equal(x = keys_115_cast_fp16, y = keys_115_cast_fp16)[name = tensor("op_8459_cast_fp16")]; + tensor keys_117_cast_fp16 = select(a = var_504_to_fp16, b = keys_115_cast_fp16, cond = var_8459_cast_fp16)[name = tensor("keys_117_cast_fp16")]; + tensor var_8467_cast_fp16 = not_equal(x = values_115_cast_fp16, y = values_115_cast_fp16)[name = tensor("op_8467_cast_fp16")]; + tensor values_117_cast_fp16 = select(a = var_504_to_fp16, b = values_115_cast_fp16, cond = var_8467_cast_fp16)[name = tensor("values_117_cast_fp16")]; + tensor var_8491 = const()[name = tensor("op_8491"), val = tensor([0, 2, 1, 3])]; + tensor var_8504 = const()[name = tensor("op_8504"), val = tensor([1, 1, 1])]; + tensor var_8505_cast_fp16 = reshape(shape = var_8504, x = position19_to_fp16)[name = tensor("op_8505_cast_fp16")]; + tensor var_8522_to_fp16 = const()[name = tensor("op_8522_to_fp16"), val = tensor(0x1p+0)]; + tensor valid_len_39_cast_fp16 = add(x = var_8505_cast_fp16, y = var_8522_to_fp16)[name = tensor("valid_len_39_cast_fp16")]; + tensor valid_mask_39_cast_fp16 = less(x = k_positions_1_promoted_to_fp16, y = valid_len_39_cast_fp16)[name = tensor("valid_mask_39_cast_fp16")]; + tensor causal_mask_39_cast_fp16 = less_equal(x = k_positions_1_promoted_to_fp16, y = var_8505_cast_fp16)[name = tensor("causal_mask_39_cast_fp16")]; + tensor attn_mask_77 = logical_and(x = valid_mask_39_cast_fp16, y = causal_mask_39_cast_fp16)[name = tensor("attn_mask_77")]; + tensor attn_mask_79_axes_0 = const()[name = tensor("attn_mask_79_axes_0"), val = tensor([1])]; + tensor attn_mask_79 = expand_dims(axes = attn_mask_79_axes_0, x = attn_mask_77)[name = tensor("attn_mask_79")]; + tensor var_8540_transpose_x_0 = const()[name = tensor("op_8540_transpose_x_0"), val = tensor(false)]; + tensor var_8540_transpose_y_0 = const()[name = tensor("op_8540_transpose_y_0"), val = tensor(false)]; + tensor transpose_110_perm_0 = const()[name = tensor("transpose_110_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_111_perm_0 = const()[name = tensor("transpose_111_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_111 = transpose(perm = transpose_111_perm_0, x = keys_117_cast_fp16)[name = tensor("transpose_137")]; + tensor transpose_110 = transpose(perm = transpose_110_perm_0, x = q_117_cast_fp16)[name = tensor("transpose_138")]; + tensor var_8540_cast_fp16 = matmul(transpose_x = var_8540_transpose_x_0, transpose_y = var_8540_transpose_y_0, x = transpose_110, y = transpose_111)[name = tensor("op_8540_cast_fp16")]; + tensor var_8534_to_fp16 = const()[name = tensor("op_8534_to_fp16"), val = tensor([0x1p-3])]; + tensor attn_weights_115_cast_fp16 = mul(x = var_8540_cast_fp16, y = var_8534_to_fp16)[name = tensor("attn_weights_115_cast_fp16")]; + tensor var_8542 = logical_not(x = attn_mask_79)[name = tensor("op_8542")]; + tensor var_8543_to_fp16 = const()[name = tensor("op_8543_to_fp16"), val = tensor(-inf)]; + tensor attn_weights_117_cast_fp16 = select(a = var_8543_to_fp16, b = attn_weights_115_cast_fp16, cond = var_8542)[name = tensor("attn_weights_117_cast_fp16")]; + tensor var_8545 = const()[name = tensor("op_8545"), val = tensor(-1)]; + tensor attn_weights_119_cast_fp16 = softmax(axis = var_8545, x = attn_weights_117_cast_fp16)[name = tensor("attn_weights_119_cast_fp16")]; + tensor attn_output_39_transpose_x_0 = const()[name = tensor("attn_output_39_transpose_x_0"), val = tensor(false)]; + tensor attn_output_39_transpose_y_0 = const()[name = tensor("attn_output_39_transpose_y_0"), val = tensor(false)]; + tensor values_119_cast_fp16 = transpose(perm = var_8491, x = values_117_cast_fp16)[name = tensor("transpose_139")]; + tensor attn_output_39_cast_fp16 = matmul(transpose_x = attn_output_39_transpose_x_0, transpose_y = attn_output_39_transpose_y_0, x = attn_weights_119_cast_fp16, y = values_119_cast_fp16)[name = tensor("attn_output_39_cast_fp16")]; + tensor var_8553 = const()[name = tensor("op_8553"), val = tensor([0, 2, 1, 3])]; + tensor var_8556 = const()[name = tensor("op_8556"), val = tensor([1, 1, 1024])]; + tensor var_8554_cast_fp16 = transpose(perm = var_8553, x = attn_output_39_cast_fp16)[name = tensor("transpose_136")]; + tensor input_195_cast_fp16 = reshape(shape = var_8556, x = var_8554_cast_fp16)[name = tensor("input_195_cast_fp16")]; + tensor linear_78_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = attn19_out_proj_weight_cast_fp16, x = input_195_cast_fp16)[name = tensor("linear_78_cast_fp16")]; + tensor var_8562_to_fp16 = const()[name = tensor("op_8562_to_fp16"), val = tensor(0x1p+0)]; + tensor var_8563_cast_fp16 = add(x = position19_to_fp16, y = var_8562_to_fp16)[name = tensor("op_8563_cast_fp16")]; + tensor var_8563_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_8563_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor input_197_cast_fp16 = add(x = input_193_cast_fp16, y = linear_78_cast_fp16)[name = tensor("input_197_cast_fp16")]; + tensor input_199_axes_0 = const()[name = tensor("input_199_axes_0"), val = tensor([-1])]; + tensor norm19_2_weight_to_fp16 = const()[name = tensor("norm19_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(303752192)))]; + tensor norm19_2_bias_to_fp16 = const()[name = tensor("norm19_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(303754304)))]; + tensor var_8567_to_fp16 = const()[name = tensor("op_8567_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_199_cast_fp16 = layer_norm(axes = input_199_axes_0, beta = norm19_2_bias_to_fp16, epsilon = var_8567_to_fp16, gamma = norm19_2_weight_to_fp16, x = input_197_cast_fp16)[name = tensor("input_199_cast_fp16")]; + tensor linear_79_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = linear19_1_weight_cast_fp16, x = input_199_cast_fp16)[name = tensor("linear_79_cast_fp16")]; + tensor input_201_mode_0 = const()[name = tensor("input_201_mode_0"), val = tensor("EXACT")]; + tensor input_201_cast_fp16 = gelu(mode = input_201_mode_0, x = linear_79_cast_fp16)[name = tensor("input_201_cast_fp16")]; + tensor linear_80_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = linear19_2_weight_cast_fp16, x = input_201_cast_fp16)[name = tensor("linear_80_cast_fp16")]; + tensor input_203_cast_fp16 = add(x = input_197_cast_fp16, y = linear_80_cast_fp16)[name = tensor("input_203_cast_fp16")]; + tensor x_41_axes_0 = const()[name = tensor("x_41_axes_0"), val = tensor([-1])]; + tensor norm20_1_weight_to_fp16 = const()[name = tensor("norm20_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(303756416)))]; + tensor norm20_1_bias_to_fp16 = const()[name = tensor("norm20_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(303758528)))]; + tensor var_8584_to_fp16 = const()[name = tensor("op_8584_to_fp16"), val = tensor(0x1.5p-17)]; + tensor x_41_cast_fp16 = layer_norm(axes = x_41_axes_0, beta = norm20_1_bias_to_fp16, epsilon = var_8584_to_fp16, gamma = norm20_1_weight_to_fp16, x = input_203_cast_fp16)[name = tensor("x_41_cast_fp16")]; + tensor linear_81_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = attn20_in_proj_weight_cast_fp16, x = x_41_cast_fp16)[name = tensor("linear_81_cast_fp16")]; + tensor var_8620 = const()[name = tensor("op_8620"), val = tensor([1, 1, 3, 16, 64])]; + tensor qkv_41_cast_fp16 = reshape(shape = var_8620, x = linear_81_cast_fp16)[name = tensor("qkv_41_cast_fp16")]; + tensor q_121_begin_0 = const()[name = tensor("q_121_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor q_121_end_0 = const()[name = tensor("q_121_end_0"), val = tensor([1, 1, 1, 16, 64])]; + tensor q_121_end_mask_0 = const()[name = tensor("q_121_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor q_121_squeeze_mask_0 = const()[name = tensor("q_121_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor q_121_cast_fp16 = slice_by_index(begin = q_121_begin_0, end = q_121_end_0, end_mask = q_121_end_mask_0, squeeze_mask = q_121_squeeze_mask_0, x = qkv_41_cast_fp16)[name = tensor("q_121_cast_fp16")]; + tensor k_81_begin_0 = const()[name = tensor("k_81_begin_0"), val = tensor([0, 0, 1, 0, 0])]; + tensor k_81_end_0 = const()[name = tensor("k_81_end_0"), val = tensor([1, 1, 2, 16, 64])]; + tensor k_81_end_mask_0 = const()[name = tensor("k_81_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor k_81_squeeze_mask_0 = const()[name = tensor("k_81_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor k_81_cast_fp16 = slice_by_index(begin = k_81_begin_0, end = k_81_end_0, end_mask = k_81_end_mask_0, squeeze_mask = k_81_squeeze_mask_0, x = qkv_41_cast_fp16)[name = tensor("k_81_cast_fp16")]; + tensor v_41_begin_0 = const()[name = tensor("v_41_begin_0"), val = tensor([0, 0, 2, 0, 0])]; + tensor v_41_end_0 = const()[name = tensor("v_41_end_0"), val = tensor([1, 1, 3, 16, 64])]; + tensor v_41_end_mask_0 = const()[name = tensor("v_41_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor v_41_squeeze_mask_0 = const()[name = tensor("v_41_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor v_41_cast_fp16 = slice_by_index(begin = v_41_begin_0, end = v_41_end_0, end_mask = v_41_end_mask_0, squeeze_mask = v_41_squeeze_mask_0, x = qkv_41_cast_fp16)[name = tensor("v_41_cast_fp16")]; + tensor var_8724 = const()[name = tensor("op_8724"), val = tensor([1, 1, 1, 1])]; + tensor position20_to_fp16_dtype_0 = const()[name = tensor("position20_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor position20_to_fp16 = cast(dtype = position20_to_fp16_dtype_0, x = position20)[name = tensor("cast_453")]; + tensor ts_125_cast_fp16 = reshape(shape = var_8724, x = position20_to_fp16)[name = tensor("ts_125_cast_fp16")]; + tensor var_8728 = const()[name = tensor("op_8728"), val = tensor([1, 1, 16, 32, 2])]; + tensor q_complex_41_cast_fp16 = reshape(shape = var_8728, x = q_121_cast_fp16)[name = tensor("q_complex_41_cast_fp16")]; + tensor var_8732 = const()[name = tensor("op_8732"), val = tensor([1, 1, 16, 32, 2])]; + tensor k_complex_41_cast_fp16 = reshape(shape = var_8732, x = k_81_cast_fp16)[name = tensor("k_complex_41_cast_fp16")]; + tensor var_8736_begin_0 = const()[name = tensor("op_8736_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_8736_end_0 = const()[name = tensor("op_8736_end_0"), val = tensor([1, 1, 16, 32, 1])]; + tensor var_8736_end_mask_0 = const()[name = tensor("op_8736_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_8736_squeeze_mask_0 = const()[name = tensor("op_8736_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_8736_cast_fp16 = slice_by_index(begin = var_8736_begin_0, end = var_8736_end_0, end_mask = var_8736_end_mask_0, squeeze_mask = var_8736_squeeze_mask_0, x = q_complex_41_cast_fp16)[name = tensor("op_8736_cast_fp16")]; + tensor var_8744_begin_0 = const()[name = tensor("op_8744_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_8744_end_0 = const()[name = tensor("op_8744_end_0"), val = tensor([1, 1, 16, 32, 2])]; + tensor var_8744_end_mask_0 = const()[name = tensor("op_8744_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_8744_squeeze_mask_0 = const()[name = tensor("op_8744_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_8744_cast_fp16 = slice_by_index(begin = var_8744_begin_0, end = var_8744_end_0, end_mask = var_8744_end_mask_0, squeeze_mask = var_8744_squeeze_mask_0, x = q_complex_41_cast_fp16)[name = tensor("op_8744_cast_fp16")]; + tensor var_8752_begin_0 = const()[name = tensor("op_8752_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_8752_end_0 = const()[name = tensor("op_8752_end_0"), val = tensor([1, 1, 16, 32, 1])]; + tensor var_8752_end_mask_0 = const()[name = tensor("op_8752_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_8752_squeeze_mask_0 = const()[name = tensor("op_8752_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_8752_cast_fp16 = slice_by_index(begin = var_8752_begin_0, end = var_8752_end_0, end_mask = var_8752_end_mask_0, squeeze_mask = var_8752_squeeze_mask_0, x = k_complex_41_cast_fp16)[name = tensor("op_8752_cast_fp16")]; + tensor var_8760_begin_0 = const()[name = tensor("op_8760_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_8760_end_0 = const()[name = tensor("op_8760_end_0"), val = tensor([1, 1, 16, 32, 2])]; + tensor var_8760_end_mask_0 = const()[name = tensor("op_8760_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_8760_squeeze_mask_0 = const()[name = tensor("op_8760_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_8760_cast_fp16 = slice_by_index(begin = var_8760_begin_0, end = var_8760_end_0, end_mask = var_8760_end_mask_0, squeeze_mask = var_8760_squeeze_mask_0, x = k_complex_41_cast_fp16)[name = tensor("op_8760_cast_fp16")]; + tensor freqs_41_to_fp16 = const()[name = tensor("freqs_41_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(303760640)))]; + tensor var_8766_cast_fp16 = mul(x = freqs_41_to_fp16, y = ts_125_cast_fp16)[name = tensor("op_8766_cast_fp16")]; + tensor rotr_41_cast_fp16 = cos(x = var_8766_cast_fp16)[name = tensor("rotr_41_cast_fp16")]; + tensor roti_41_cast_fp16 = sin(x = var_8766_cast_fp16)[name = tensor("roti_41_cast_fp16")]; + tensor var_8770_cast_fp16 = mul(x = var_8736_cast_fp16, y = rotr_41_cast_fp16)[name = tensor("op_8770_cast_fp16")]; + tensor var_8771_cast_fp16 = mul(x = var_8744_cast_fp16, y = roti_41_cast_fp16)[name = tensor("op_8771_cast_fp16")]; + tensor qor_81_cast_fp16 = sub(x = var_8770_cast_fp16, y = var_8771_cast_fp16)[name = tensor("qor_81_cast_fp16")]; + tensor var_8774_cast_fp16 = mul(x = var_8736_cast_fp16, y = roti_41_cast_fp16)[name = tensor("op_8774_cast_fp16")]; + tensor var_8775_cast_fp16 = mul(x = var_8744_cast_fp16, y = rotr_41_cast_fp16)[name = tensor("op_8775_cast_fp16")]; + tensor qoi_81_cast_fp16 = add(x = var_8774_cast_fp16, y = var_8775_cast_fp16)[name = tensor("qoi_81_cast_fp16")]; + tensor var_8778_cast_fp16 = mul(x = var_8752_cast_fp16, y = rotr_41_cast_fp16)[name = tensor("op_8778_cast_fp16")]; + tensor var_8779_cast_fp16 = mul(x = var_8760_cast_fp16, y = roti_41_cast_fp16)[name = tensor("op_8779_cast_fp16")]; + tensor kor_81_cast_fp16 = sub(x = var_8778_cast_fp16, y = var_8779_cast_fp16)[name = tensor("kor_81_cast_fp16")]; + tensor var_8782_cast_fp16 = mul(x = var_8752_cast_fp16, y = roti_41_cast_fp16)[name = tensor("op_8782_cast_fp16")]; + tensor var_8783_cast_fp16 = mul(x = var_8760_cast_fp16, y = rotr_41_cast_fp16)[name = tensor("op_8783_cast_fp16")]; + tensor koi_81_cast_fp16 = add(x = var_8782_cast_fp16, y = var_8783_cast_fp16)[name = tensor("koi_81_cast_fp16")]; + tensor qo_41_axis_0 = const()[name = tensor("qo_41_axis_0"), val = tensor(-1)]; + tensor qo_41_cast_fp16 = stack(axis = qo_41_axis_0, values = (qor_81_cast_fp16, qoi_81_cast_fp16))[name = tensor("qo_41_cast_fp16")]; + tensor ko_41_axis_0 = const()[name = tensor("ko_41_axis_0"), val = tensor(-1)]; + tensor ko_41_cast_fp16 = stack(axis = ko_41_axis_0, values = (kor_81_cast_fp16, koi_81_cast_fp16))[name = tensor("ko_41_cast_fp16")]; + tensor var_8812 = const()[name = tensor("op_8812"), val = tensor([1, 1, 16, 64])]; + tensor q_123_cast_fp16 = reshape(shape = var_8812, x = qo_41_cast_fp16)[name = tensor("q_123_cast_fp16")]; + tensor var_8814 = const()[name = tensor("op_8814"), val = tensor([1, 1, 16, 64])]; + tensor k_83_cast_fp16 = reshape(shape = var_8814, x = ko_41_cast_fp16)[name = tensor("k_83_cast_fp16")]; + tensor _inversed_8836_y_0_to_fp16 = const()[name = tensor("_inversed_8836_y_0_to_fp16"), val = tensor(0x1p-9)]; + tensor _inversed_8836_cast_fp16 = mul(x = ts_125_cast_fp16, y = _inversed_8836_y_0_to_fp16)[name = tensor("_inversed_8836_cast_fp16")]; + tensor var_8837_cast_fp16 = floor(x = _inversed_8836_cast_fp16)[name = tensor("op_8837_cast_fp16")]; + tensor var_8838_to_fp16 = const()[name = tensor("op_8838_to_fp16"), val = tensor(0x1p+9)]; + tensor var_8839_cast_fp16 = mul(x = var_8837_cast_fp16, y = var_8838_to_fp16)[name = tensor("op_8839_cast_fp16")]; + tensor write_indices_float_83_cast_fp16 = sub(x = ts_125_cast_fp16, y = var_8839_cast_fp16)[name = tensor("write_indices_float_83_cast_fp16")]; + tensor var_8846_dtype_0 = const()[name = tensor("op_8846_dtype_0"), val = tensor("int32")]; + tensor write_indices_41_reps_0 = const()[name = tensor("write_indices_41_reps_0"), val = tensor([1, 1, 16, 64])]; + tensor write_indices_float_83_cast_fp16_to_int32 = cast(dtype = var_8846_dtype_0, x = write_indices_float_83_cast_fp16)[name = tensor("cast_452")]; + tensor write_indices_41 = tile(reps = write_indices_41_reps_0, x = write_indices_float_83_cast_fp16_to_int32)[name = tensor("write_indices_41")]; + tensor var_8854_begin_0 = const()[name = tensor("op_8854_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_8854_end_0 = const()[name = tensor("op_8854_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor var_8854_end_mask_0 = const()[name = tensor("op_8854_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_8854_squeeze_mask_0 = const()[name = tensor("op_8854_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor cache20_to_fp16_dtype_0 = const()[name = tensor("cache20_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor cache20_to_fp16 = cast(dtype = cache20_to_fp16_dtype_0, x = cache20)[name = tensor("cast_451")]; + tensor var_8854_cast_fp16 = slice_by_index(begin = var_8854_begin_0, end = var_8854_end_0, end_mask = var_8854_end_mask_0, squeeze_mask = var_8854_squeeze_mask_0, x = cache20_to_fp16)[name = tensor("op_8854_cast_fp16")]; + tensor var_8856_axis_0 = const()[name = tensor("op_8856_axis_0"), val = tensor(1)]; + tensor var_8856_mode_0 = const()[name = tensor("op_8856_mode_0"), val = tensor("update")]; + tensor var_8856_validate_indices_0 = const()[name = tensor("op_8856_validate_indices_0"), val = tensor(false)]; + tensor var_8856_cast_fp16 = scatter_along_axis(axis = var_8856_axis_0, data = var_8854_cast_fp16, indices = write_indices_41, mode = var_8856_mode_0, updates = k_83_cast_fp16, validate_indices = var_8856_validate_indices_0)[name = tensor("op_8856_cast_fp16")]; + tensor concat_142 = const()[name = tensor("concat_142"), val = tensor([0, 0, 0, 0, 0])]; + tensor concat_143 = const()[name = tensor("concat_143"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_41_internal_tensor_assign_1_stride_0 = const()[name = tensor("new_cache_41_internal_tensor_assign_1_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_41_internal_tensor_assign_1_begin_mask_0 = const()[name = tensor("new_cache_41_internal_tensor_assign_1_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_41_internal_tensor_assign_1_end_mask_0 = const()[name = tensor("new_cache_41_internal_tensor_assign_1_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_41_internal_tensor_assign_1_squeeze_mask_0 = const()[name = tensor("new_cache_41_internal_tensor_assign_1_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_88 = const()[name = tensor("shape_88"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_40 = const()[name = tensor("reduce_prod_40"), val = tensor(1048576)]; + tensor range_1d_40_start_0 = const()[name = tensor("range_1d_40_start_0"), val = tensor(0)]; + tensor range_1d_40_step_0 = const()[name = tensor("range_1d_40_step_0"), val = tensor(1)]; + tensor range_1d_40 = range_1d(end = reduce_prod_40, start = range_1d_40_start_0, step = range_1d_40_step_0)[name = tensor("range_1d_40")]; + tensor reshape_200 = reshape(shape = shape_88, x = range_1d_40)[name = tensor("reshape_200")]; + tensor slice_by_index_40 = slice_by_index(begin = concat_142, begin_mask = new_cache_41_internal_tensor_assign_1_begin_mask_0, end = concat_143, end_mask = new_cache_41_internal_tensor_assign_1_end_mask_0, squeeze_mask = new_cache_41_internal_tensor_assign_1_squeeze_mask_0, stride = new_cache_41_internal_tensor_assign_1_stride_0, x = reshape_200)[name = tensor("slice_by_index_40")]; + tensor reshape_201_shape_0 = const()[name = tensor("reshape_201_shape_0"), val = tensor([-1])]; + tensor reshape_201 = reshape(shape = reshape_201_shape_0, x = slice_by_index_40)[name = tensor("reshape_201")]; + tensor reshape_202_shape_0 = const()[name = tensor("reshape_202_shape_0"), val = tensor([-1])]; + tensor reshape_202_cast_fp16 = reshape(shape = reshape_202_shape_0, x = var_8856_cast_fp16)[name = tensor("reshape_202_cast_fp16")]; + tensor reshape_203_shape_0 = const()[name = tensor("reshape_203_shape_0"), val = tensor([-1])]; + tensor reshape_203_cast_fp16 = reshape(shape = reshape_203_shape_0, x = cache20_to_fp16)[name = tensor("reshape_203_cast_fp16")]; + tensor scatter_40_mode_0 = const()[name = tensor("scatter_40_mode_0"), val = tensor("update")]; + tensor scatter_40_axis_0 = const()[name = tensor("scatter_40_axis_0"), val = tensor(0)]; + tensor scatter_40_validate_indices_0 = const()[name = tensor("scatter_40_validate_indices_0"), val = tensor(false)]; + tensor scatter_40_cast_fp16 = scatter(axis = scatter_40_axis_0, data = reshape_203_cast_fp16, indices = reshape_201, mode = scatter_40_mode_0, updates = reshape_202_cast_fp16, validate_indices = scatter_40_validate_indices_0)[name = tensor("scatter_40_cast_fp16")]; + tensor reshape_204_cast_fp16 = reshape(shape = shape_88, x = scatter_40_cast_fp16)[name = tensor("reshape_204_cast_fp16")]; + tensor var_8864_begin_0 = const()[name = tensor("op_8864_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor var_8864_end_0 = const()[name = tensor("op_8864_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor var_8864_end_mask_0 = const()[name = tensor("op_8864_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_8864_squeeze_mask_0 = const()[name = tensor("op_8864_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor var_8864_cast_fp16 = slice_by_index(begin = var_8864_begin_0, end = var_8864_end_0, end_mask = var_8864_end_mask_0, squeeze_mask = var_8864_squeeze_mask_0, x = reshape_204_cast_fp16)[name = tensor("op_8864_cast_fp16")]; + tensor var_8866_axis_0 = const()[name = tensor("op_8866_axis_0"), val = tensor(1)]; + tensor var_8866_mode_0 = const()[name = tensor("op_8866_mode_0"), val = tensor("update")]; + tensor var_8866_validate_indices_0 = const()[name = tensor("op_8866_validate_indices_0"), val = tensor(false)]; + tensor var_8866_cast_fp16 = scatter_along_axis(axis = var_8866_axis_0, data = var_8864_cast_fp16, indices = write_indices_41, mode = var_8866_mode_0, updates = v_41_cast_fp16, validate_indices = var_8866_validate_indices_0)[name = tensor("op_8866_cast_fp16")]; + tensor concat_144 = const()[name = tensor("concat_144"), val = tensor([1, 0, 0, 0, 0])]; + tensor concat_145 = const()[name = tensor("concat_145"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_41_internal_tensor_assign_2_stride_0 = const()[name = tensor("new_cache_41_internal_tensor_assign_2_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_41_internal_tensor_assign_2_begin_mask_0 = const()[name = tensor("new_cache_41_internal_tensor_assign_2_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_41_internal_tensor_assign_2_end_mask_0 = const()[name = tensor("new_cache_41_internal_tensor_assign_2_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_41_internal_tensor_assign_2_squeeze_mask_0 = const()[name = tensor("new_cache_41_internal_tensor_assign_2_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_89 = const()[name = tensor("shape_89"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_41 = const()[name = tensor("reduce_prod_41"), val = tensor(1048576)]; + tensor range_1d_41_start_0 = const()[name = tensor("range_1d_41_start_0"), val = tensor(0)]; + tensor range_1d_41_step_0 = const()[name = tensor("range_1d_41_step_0"), val = tensor(1)]; + tensor range_1d_41 = range_1d(end = reduce_prod_41, start = range_1d_41_start_0, step = range_1d_41_step_0)[name = tensor("range_1d_41")]; + tensor reshape_205 = reshape(shape = shape_89, x = range_1d_41)[name = tensor("reshape_205")]; + tensor slice_by_index_41 = slice_by_index(begin = concat_144, begin_mask = new_cache_41_internal_tensor_assign_2_begin_mask_0, end = concat_145, end_mask = new_cache_41_internal_tensor_assign_2_end_mask_0, squeeze_mask = new_cache_41_internal_tensor_assign_2_squeeze_mask_0, stride = new_cache_41_internal_tensor_assign_2_stride_0, x = reshape_205)[name = tensor("slice_by_index_41")]; + tensor reshape_206_shape_0 = const()[name = tensor("reshape_206_shape_0"), val = tensor([-1])]; + tensor reshape_206 = reshape(shape = reshape_206_shape_0, x = slice_by_index_41)[name = tensor("reshape_206")]; + tensor reshape_207_shape_0 = const()[name = tensor("reshape_207_shape_0"), val = tensor([-1])]; + tensor reshape_207_cast_fp16 = reshape(shape = reshape_207_shape_0, x = var_8866_cast_fp16)[name = tensor("reshape_207_cast_fp16")]; + tensor reshape_208_shape_0 = const()[name = tensor("reshape_208_shape_0"), val = tensor([-1])]; + tensor reshape_208_cast_fp16 = reshape(shape = reshape_208_shape_0, x = reshape_204_cast_fp16)[name = tensor("reshape_208_cast_fp16")]; + tensor scatter_41_mode_0 = const()[name = tensor("scatter_41_mode_0"), val = tensor("update")]; + tensor scatter_41_axis_0 = const()[name = tensor("scatter_41_axis_0"), val = tensor(0)]; + tensor scatter_41_validate_indices_0 = const()[name = tensor("scatter_41_validate_indices_0"), val = tensor(false)]; + tensor scatter_41_cast_fp16 = scatter(axis = scatter_41_axis_0, data = reshape_208_cast_fp16, indices = reshape_206, mode = scatter_41_mode_0, updates = reshape_207_cast_fp16, validate_indices = scatter_41_validate_indices_0)[name = tensor("scatter_41_cast_fp16")]; + tensor reshape_209_cast_fp16 = reshape(shape = shape_89, x = scatter_41_cast_fp16)[name = tensor("reshape_209_cast_fp16")]; + tensor reshape_209_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("reshape_209_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor keys_121_begin_0 = const()[name = tensor("keys_121_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor keys_121_end_0 = const()[name = tensor("keys_121_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor keys_121_end_mask_0 = const()[name = tensor("keys_121_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor keys_121_squeeze_mask_0 = const()[name = tensor("keys_121_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor keys_121_cast_fp16 = slice_by_index(begin = keys_121_begin_0, end = keys_121_end_0, end_mask = keys_121_end_mask_0, squeeze_mask = keys_121_squeeze_mask_0, x = reshape_209_cast_fp16)[name = tensor("keys_121_cast_fp16")]; + tensor values_121_begin_0 = const()[name = tensor("values_121_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor values_121_end_0 = const()[name = tensor("values_121_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor values_121_end_mask_0 = const()[name = tensor("values_121_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor values_121_squeeze_mask_0 = const()[name = tensor("values_121_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor values_121_cast_fp16 = slice_by_index(begin = values_121_begin_0, end = values_121_end_0, end_mask = values_121_end_mask_0, squeeze_mask = values_121_squeeze_mask_0, x = reshape_209_cast_fp16)[name = tensor("values_121_cast_fp16")]; + tensor var_8878_cast_fp16 = not_equal(x = keys_121_cast_fp16, y = keys_121_cast_fp16)[name = tensor("op_8878_cast_fp16")]; + tensor keys_123_cast_fp16 = select(a = var_504_to_fp16, b = keys_121_cast_fp16, cond = var_8878_cast_fp16)[name = tensor("keys_123_cast_fp16")]; + tensor var_8886_cast_fp16 = not_equal(x = values_121_cast_fp16, y = values_121_cast_fp16)[name = tensor("op_8886_cast_fp16")]; + tensor values_123_cast_fp16 = select(a = var_504_to_fp16, b = values_121_cast_fp16, cond = var_8886_cast_fp16)[name = tensor("values_123_cast_fp16")]; + tensor var_8910 = const()[name = tensor("op_8910"), val = tensor([0, 2, 1, 3])]; + tensor var_8923 = const()[name = tensor("op_8923"), val = tensor([1, 1, 1])]; + tensor var_8924_cast_fp16 = reshape(shape = var_8923, x = position20_to_fp16)[name = tensor("op_8924_cast_fp16")]; + tensor var_8941_to_fp16 = const()[name = tensor("op_8941_to_fp16"), val = tensor(0x1p+0)]; + tensor valid_len_41_cast_fp16 = add(x = var_8924_cast_fp16, y = var_8941_to_fp16)[name = tensor("valid_len_41_cast_fp16")]; + tensor valid_mask_41_cast_fp16 = less(x = k_positions_1_promoted_to_fp16, y = valid_len_41_cast_fp16)[name = tensor("valid_mask_41_cast_fp16")]; + tensor causal_mask_41_cast_fp16 = less_equal(x = k_positions_1_promoted_to_fp16, y = var_8924_cast_fp16)[name = tensor("causal_mask_41_cast_fp16")]; + tensor attn_mask_81 = logical_and(x = valid_mask_41_cast_fp16, y = causal_mask_41_cast_fp16)[name = tensor("attn_mask_81")]; + tensor attn_mask_83_axes_0 = const()[name = tensor("attn_mask_83_axes_0"), val = tensor([1])]; + tensor attn_mask_83 = expand_dims(axes = attn_mask_83_axes_0, x = attn_mask_81)[name = tensor("attn_mask_83")]; + tensor var_8959_transpose_x_0 = const()[name = tensor("op_8959_transpose_x_0"), val = tensor(false)]; + tensor var_8959_transpose_y_0 = const()[name = tensor("op_8959_transpose_y_0"), val = tensor(false)]; + tensor transpose_112_perm_0 = const()[name = tensor("transpose_112_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_113_perm_0 = const()[name = tensor("transpose_113_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_113 = transpose(perm = transpose_113_perm_0, x = keys_123_cast_fp16)[name = tensor("transpose_133")]; + tensor transpose_112 = transpose(perm = transpose_112_perm_0, x = q_123_cast_fp16)[name = tensor("transpose_134")]; + tensor var_8959_cast_fp16 = matmul(transpose_x = var_8959_transpose_x_0, transpose_y = var_8959_transpose_y_0, x = transpose_112, y = transpose_113)[name = tensor("op_8959_cast_fp16")]; + tensor var_8953_to_fp16 = const()[name = tensor("op_8953_to_fp16"), val = tensor([0x1p-3])]; + tensor attn_weights_121_cast_fp16 = mul(x = var_8959_cast_fp16, y = var_8953_to_fp16)[name = tensor("attn_weights_121_cast_fp16")]; + tensor var_8961 = logical_not(x = attn_mask_83)[name = tensor("op_8961")]; + tensor var_8962_to_fp16 = const()[name = tensor("op_8962_to_fp16"), val = tensor(-inf)]; + tensor attn_weights_123_cast_fp16 = select(a = var_8962_to_fp16, b = attn_weights_121_cast_fp16, cond = var_8961)[name = tensor("attn_weights_123_cast_fp16")]; + tensor var_8964 = const()[name = tensor("op_8964"), val = tensor(-1)]; + tensor attn_weights_125_cast_fp16 = softmax(axis = var_8964, x = attn_weights_123_cast_fp16)[name = tensor("attn_weights_125_cast_fp16")]; + tensor attn_output_41_transpose_x_0 = const()[name = tensor("attn_output_41_transpose_x_0"), val = tensor(false)]; + tensor attn_output_41_transpose_y_0 = const()[name = tensor("attn_output_41_transpose_y_0"), val = tensor(false)]; + tensor values_125_cast_fp16 = transpose(perm = var_8910, x = values_123_cast_fp16)[name = tensor("transpose_135")]; + tensor attn_output_41_cast_fp16 = matmul(transpose_x = attn_output_41_transpose_x_0, transpose_y = attn_output_41_transpose_y_0, x = attn_weights_125_cast_fp16, y = values_125_cast_fp16)[name = tensor("attn_output_41_cast_fp16")]; + tensor var_8972 = const()[name = tensor("op_8972"), val = tensor([0, 2, 1, 3])]; + tensor var_8975 = const()[name = tensor("op_8975"), val = tensor([1, 1, 1024])]; + tensor var_8973_cast_fp16 = transpose(perm = var_8972, x = attn_output_41_cast_fp16)[name = tensor("transpose_132")]; + tensor input_205_cast_fp16 = reshape(shape = var_8975, x = var_8973_cast_fp16)[name = tensor("input_205_cast_fp16")]; + tensor linear_82_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = attn20_out_proj_weight_cast_fp16, x = input_205_cast_fp16)[name = tensor("linear_82_cast_fp16")]; + tensor var_8981_to_fp16 = const()[name = tensor("op_8981_to_fp16"), val = tensor(0x1p+0)]; + tensor var_8982_cast_fp16 = add(x = position20_to_fp16, y = var_8981_to_fp16)[name = tensor("op_8982_cast_fp16")]; + tensor var_8982_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_8982_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor input_207_cast_fp16 = add(x = input_203_cast_fp16, y = linear_82_cast_fp16)[name = tensor("input_207_cast_fp16")]; + tensor input_209_axes_0 = const()[name = tensor("input_209_axes_0"), val = tensor([-1])]; + tensor norm20_2_weight_to_fp16 = const()[name = tensor("norm20_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(303760768)))]; + tensor norm20_2_bias_to_fp16 = const()[name = tensor("norm20_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(303762880)))]; + tensor var_8986_to_fp16 = const()[name = tensor("op_8986_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_209_cast_fp16 = layer_norm(axes = input_209_axes_0, beta = norm20_2_bias_to_fp16, epsilon = var_8986_to_fp16, gamma = norm20_2_weight_to_fp16, x = input_207_cast_fp16)[name = tensor("input_209_cast_fp16")]; + tensor linear_83_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = linear20_1_weight_cast_fp16, x = input_209_cast_fp16)[name = tensor("linear_83_cast_fp16")]; + tensor input_211_mode_0 = const()[name = tensor("input_211_mode_0"), val = tensor("EXACT")]; + tensor input_211_cast_fp16 = gelu(mode = input_211_mode_0, x = linear_83_cast_fp16)[name = tensor("input_211_cast_fp16")]; + tensor linear_84_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = linear20_2_weight_cast_fp16, x = input_211_cast_fp16)[name = tensor("linear_84_cast_fp16")]; + tensor input_213_cast_fp16 = add(x = input_207_cast_fp16, y = linear_84_cast_fp16)[name = tensor("input_213_cast_fp16")]; + tensor x_43_axes_0 = const()[name = tensor("x_43_axes_0"), val = tensor([-1])]; + tensor norm21_1_weight_to_fp16 = const()[name = tensor("norm21_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(303764992)))]; + tensor norm21_1_bias_to_fp16 = const()[name = tensor("norm21_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(303767104)))]; + tensor var_9003_to_fp16 = const()[name = tensor("op_9003_to_fp16"), val = tensor(0x1.5p-17)]; + tensor x_43_cast_fp16 = layer_norm(axes = x_43_axes_0, beta = norm21_1_bias_to_fp16, epsilon = var_9003_to_fp16, gamma = norm21_1_weight_to_fp16, x = input_213_cast_fp16)[name = tensor("x_43_cast_fp16")]; + tensor linear_85_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = attn21_in_proj_weight_cast_fp16, x = x_43_cast_fp16)[name = tensor("linear_85_cast_fp16")]; + tensor var_9039 = const()[name = tensor("op_9039"), val = tensor([1, 1, 3, 16, 64])]; + tensor qkv_43_cast_fp16 = reshape(shape = var_9039, x = linear_85_cast_fp16)[name = tensor("qkv_43_cast_fp16")]; + tensor q_127_begin_0 = const()[name = tensor("q_127_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor q_127_end_0 = const()[name = tensor("q_127_end_0"), val = tensor([1, 1, 1, 16, 64])]; + tensor q_127_end_mask_0 = const()[name = tensor("q_127_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor q_127_squeeze_mask_0 = const()[name = tensor("q_127_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor q_127_cast_fp16 = slice_by_index(begin = q_127_begin_0, end = q_127_end_0, end_mask = q_127_end_mask_0, squeeze_mask = q_127_squeeze_mask_0, x = qkv_43_cast_fp16)[name = tensor("q_127_cast_fp16")]; + tensor k_85_begin_0 = const()[name = tensor("k_85_begin_0"), val = tensor([0, 0, 1, 0, 0])]; + tensor k_85_end_0 = const()[name = tensor("k_85_end_0"), val = tensor([1, 1, 2, 16, 64])]; + tensor k_85_end_mask_0 = const()[name = tensor("k_85_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor k_85_squeeze_mask_0 = const()[name = tensor("k_85_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor k_85_cast_fp16 = slice_by_index(begin = k_85_begin_0, end = k_85_end_0, end_mask = k_85_end_mask_0, squeeze_mask = k_85_squeeze_mask_0, x = qkv_43_cast_fp16)[name = tensor("k_85_cast_fp16")]; + tensor v_43_begin_0 = const()[name = tensor("v_43_begin_0"), val = tensor([0, 0, 2, 0, 0])]; + tensor v_43_end_0 = const()[name = tensor("v_43_end_0"), val = tensor([1, 1, 3, 16, 64])]; + tensor v_43_end_mask_0 = const()[name = tensor("v_43_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor v_43_squeeze_mask_0 = const()[name = tensor("v_43_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor v_43_cast_fp16 = slice_by_index(begin = v_43_begin_0, end = v_43_end_0, end_mask = v_43_end_mask_0, squeeze_mask = v_43_squeeze_mask_0, x = qkv_43_cast_fp16)[name = tensor("v_43_cast_fp16")]; + tensor var_9143 = const()[name = tensor("op_9143"), val = tensor([1, 1, 1, 1])]; + tensor position21_to_fp16_dtype_0 = const()[name = tensor("position21_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor position21_to_fp16 = cast(dtype = position21_to_fp16_dtype_0, x = position21)[name = tensor("cast_448")]; + tensor ts_131_cast_fp16 = reshape(shape = var_9143, x = position21_to_fp16)[name = tensor("ts_131_cast_fp16")]; + tensor var_9147 = const()[name = tensor("op_9147"), val = tensor([1, 1, 16, 32, 2])]; + tensor q_complex_43_cast_fp16 = reshape(shape = var_9147, x = q_127_cast_fp16)[name = tensor("q_complex_43_cast_fp16")]; + tensor var_9151 = const()[name = tensor("op_9151"), val = tensor([1, 1, 16, 32, 2])]; + tensor k_complex_43_cast_fp16 = reshape(shape = var_9151, x = k_85_cast_fp16)[name = tensor("k_complex_43_cast_fp16")]; + tensor var_9155_begin_0 = const()[name = tensor("op_9155_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_9155_end_0 = const()[name = tensor("op_9155_end_0"), val = tensor([1, 1, 16, 32, 1])]; + tensor var_9155_end_mask_0 = const()[name = tensor("op_9155_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_9155_squeeze_mask_0 = const()[name = tensor("op_9155_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_9155_cast_fp16 = slice_by_index(begin = var_9155_begin_0, end = var_9155_end_0, end_mask = var_9155_end_mask_0, squeeze_mask = var_9155_squeeze_mask_0, x = q_complex_43_cast_fp16)[name = tensor("op_9155_cast_fp16")]; + tensor var_9163_begin_0 = const()[name = tensor("op_9163_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_9163_end_0 = const()[name = tensor("op_9163_end_0"), val = tensor([1, 1, 16, 32, 2])]; + tensor var_9163_end_mask_0 = const()[name = tensor("op_9163_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_9163_squeeze_mask_0 = const()[name = tensor("op_9163_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_9163_cast_fp16 = slice_by_index(begin = var_9163_begin_0, end = var_9163_end_0, end_mask = var_9163_end_mask_0, squeeze_mask = var_9163_squeeze_mask_0, x = q_complex_43_cast_fp16)[name = tensor("op_9163_cast_fp16")]; + tensor var_9171_begin_0 = const()[name = tensor("op_9171_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_9171_end_0 = const()[name = tensor("op_9171_end_0"), val = tensor([1, 1, 16, 32, 1])]; + tensor var_9171_end_mask_0 = const()[name = tensor("op_9171_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_9171_squeeze_mask_0 = const()[name = tensor("op_9171_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_9171_cast_fp16 = slice_by_index(begin = var_9171_begin_0, end = var_9171_end_0, end_mask = var_9171_end_mask_0, squeeze_mask = var_9171_squeeze_mask_0, x = k_complex_43_cast_fp16)[name = tensor("op_9171_cast_fp16")]; + tensor var_9179_begin_0 = const()[name = tensor("op_9179_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_9179_end_0 = const()[name = tensor("op_9179_end_0"), val = tensor([1, 1, 16, 32, 2])]; + tensor var_9179_end_mask_0 = const()[name = tensor("op_9179_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_9179_squeeze_mask_0 = const()[name = tensor("op_9179_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_9179_cast_fp16 = slice_by_index(begin = var_9179_begin_0, end = var_9179_end_0, end_mask = var_9179_end_mask_0, squeeze_mask = var_9179_squeeze_mask_0, x = k_complex_43_cast_fp16)[name = tensor("op_9179_cast_fp16")]; + tensor freqs_43_to_fp16 = const()[name = tensor("freqs_43_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(303769216)))]; + tensor var_9185_cast_fp16 = mul(x = freqs_43_to_fp16, y = ts_131_cast_fp16)[name = tensor("op_9185_cast_fp16")]; + tensor rotr_43_cast_fp16 = cos(x = var_9185_cast_fp16)[name = tensor("rotr_43_cast_fp16")]; + tensor roti_43_cast_fp16 = sin(x = var_9185_cast_fp16)[name = tensor("roti_43_cast_fp16")]; + tensor var_9189_cast_fp16 = mul(x = var_9155_cast_fp16, y = rotr_43_cast_fp16)[name = tensor("op_9189_cast_fp16")]; + tensor var_9190_cast_fp16 = mul(x = var_9163_cast_fp16, y = roti_43_cast_fp16)[name = tensor("op_9190_cast_fp16")]; + tensor qor_85_cast_fp16 = sub(x = var_9189_cast_fp16, y = var_9190_cast_fp16)[name = tensor("qor_85_cast_fp16")]; + tensor var_9193_cast_fp16 = mul(x = var_9155_cast_fp16, y = roti_43_cast_fp16)[name = tensor("op_9193_cast_fp16")]; + tensor var_9194_cast_fp16 = mul(x = var_9163_cast_fp16, y = rotr_43_cast_fp16)[name = tensor("op_9194_cast_fp16")]; + tensor qoi_85_cast_fp16 = add(x = var_9193_cast_fp16, y = var_9194_cast_fp16)[name = tensor("qoi_85_cast_fp16")]; + tensor var_9197_cast_fp16 = mul(x = var_9171_cast_fp16, y = rotr_43_cast_fp16)[name = tensor("op_9197_cast_fp16")]; + tensor var_9198_cast_fp16 = mul(x = var_9179_cast_fp16, y = roti_43_cast_fp16)[name = tensor("op_9198_cast_fp16")]; + tensor kor_85_cast_fp16 = sub(x = var_9197_cast_fp16, y = var_9198_cast_fp16)[name = tensor("kor_85_cast_fp16")]; + tensor var_9201_cast_fp16 = mul(x = var_9171_cast_fp16, y = roti_43_cast_fp16)[name = tensor("op_9201_cast_fp16")]; + tensor var_9202_cast_fp16 = mul(x = var_9179_cast_fp16, y = rotr_43_cast_fp16)[name = tensor("op_9202_cast_fp16")]; + tensor koi_85_cast_fp16 = add(x = var_9201_cast_fp16, y = var_9202_cast_fp16)[name = tensor("koi_85_cast_fp16")]; + tensor qo_43_axis_0 = const()[name = tensor("qo_43_axis_0"), val = tensor(-1)]; + tensor qo_43_cast_fp16 = stack(axis = qo_43_axis_0, values = (qor_85_cast_fp16, qoi_85_cast_fp16))[name = tensor("qo_43_cast_fp16")]; + tensor ko_43_axis_0 = const()[name = tensor("ko_43_axis_0"), val = tensor(-1)]; + tensor ko_43_cast_fp16 = stack(axis = ko_43_axis_0, values = (kor_85_cast_fp16, koi_85_cast_fp16))[name = tensor("ko_43_cast_fp16")]; + tensor var_9231 = const()[name = tensor("op_9231"), val = tensor([1, 1, 16, 64])]; + tensor q_129_cast_fp16 = reshape(shape = var_9231, x = qo_43_cast_fp16)[name = tensor("q_129_cast_fp16")]; + tensor var_9233 = const()[name = tensor("op_9233"), val = tensor([1, 1, 16, 64])]; + tensor k_87_cast_fp16 = reshape(shape = var_9233, x = ko_43_cast_fp16)[name = tensor("k_87_cast_fp16")]; + tensor _inversed_9255_y_0_to_fp16 = const()[name = tensor("_inversed_9255_y_0_to_fp16"), val = tensor(0x1p-9)]; + tensor _inversed_9255_cast_fp16 = mul(x = ts_131_cast_fp16, y = _inversed_9255_y_0_to_fp16)[name = tensor("_inversed_9255_cast_fp16")]; + tensor var_9256_cast_fp16 = floor(x = _inversed_9255_cast_fp16)[name = tensor("op_9256_cast_fp16")]; + tensor var_9257_to_fp16 = const()[name = tensor("op_9257_to_fp16"), val = tensor(0x1p+9)]; + tensor var_9258_cast_fp16 = mul(x = var_9256_cast_fp16, y = var_9257_to_fp16)[name = tensor("op_9258_cast_fp16")]; + tensor write_indices_float_87_cast_fp16 = sub(x = ts_131_cast_fp16, y = var_9258_cast_fp16)[name = tensor("write_indices_float_87_cast_fp16")]; + tensor var_9265_dtype_0 = const()[name = tensor("op_9265_dtype_0"), val = tensor("int32")]; + tensor write_indices_43_reps_0 = const()[name = tensor("write_indices_43_reps_0"), val = tensor([1, 1, 16, 64])]; + tensor write_indices_float_87_cast_fp16_to_int32 = cast(dtype = var_9265_dtype_0, x = write_indices_float_87_cast_fp16)[name = tensor("cast_447")]; + tensor write_indices_43 = tile(reps = write_indices_43_reps_0, x = write_indices_float_87_cast_fp16_to_int32)[name = tensor("write_indices_43")]; + tensor var_9273_begin_0 = const()[name = tensor("op_9273_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_9273_end_0 = const()[name = tensor("op_9273_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor var_9273_end_mask_0 = const()[name = tensor("op_9273_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_9273_squeeze_mask_0 = const()[name = tensor("op_9273_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor cache21_to_fp16_dtype_0 = const()[name = tensor("cache21_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor cache21_to_fp16 = cast(dtype = cache21_to_fp16_dtype_0, x = cache21)[name = tensor("cast_446")]; + tensor var_9273_cast_fp16 = slice_by_index(begin = var_9273_begin_0, end = var_9273_end_0, end_mask = var_9273_end_mask_0, squeeze_mask = var_9273_squeeze_mask_0, x = cache21_to_fp16)[name = tensor("op_9273_cast_fp16")]; + tensor var_9275_axis_0 = const()[name = tensor("op_9275_axis_0"), val = tensor(1)]; + tensor var_9275_mode_0 = const()[name = tensor("op_9275_mode_0"), val = tensor("update")]; + tensor var_9275_validate_indices_0 = const()[name = tensor("op_9275_validate_indices_0"), val = tensor(false)]; + tensor var_9275_cast_fp16 = scatter_along_axis(axis = var_9275_axis_0, data = var_9273_cast_fp16, indices = write_indices_43, mode = var_9275_mode_0, updates = k_87_cast_fp16, validate_indices = var_9275_validate_indices_0)[name = tensor("op_9275_cast_fp16")]; + tensor concat_149 = const()[name = tensor("concat_149"), val = tensor([0, 0, 0, 0, 0])]; + tensor concat_150 = const()[name = tensor("concat_150"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_43_internal_tensor_assign_1_stride_0 = const()[name = tensor("new_cache_43_internal_tensor_assign_1_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_43_internal_tensor_assign_1_begin_mask_0 = const()[name = tensor("new_cache_43_internal_tensor_assign_1_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_43_internal_tensor_assign_1_end_mask_0 = const()[name = tensor("new_cache_43_internal_tensor_assign_1_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_43_internal_tensor_assign_1_squeeze_mask_0 = const()[name = tensor("new_cache_43_internal_tensor_assign_1_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_90 = const()[name = tensor("shape_90"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_42 = const()[name = tensor("reduce_prod_42"), val = tensor(1048576)]; + tensor range_1d_42_start_0 = const()[name = tensor("range_1d_42_start_0"), val = tensor(0)]; + tensor range_1d_42_step_0 = const()[name = tensor("range_1d_42_step_0"), val = tensor(1)]; + tensor range_1d_42 = range_1d(end = reduce_prod_42, start = range_1d_42_start_0, step = range_1d_42_step_0)[name = tensor("range_1d_42")]; + tensor reshape_210 = reshape(shape = shape_90, x = range_1d_42)[name = tensor("reshape_210")]; + tensor slice_by_index_42 = slice_by_index(begin = concat_149, begin_mask = new_cache_43_internal_tensor_assign_1_begin_mask_0, end = concat_150, end_mask = new_cache_43_internal_tensor_assign_1_end_mask_0, squeeze_mask = new_cache_43_internal_tensor_assign_1_squeeze_mask_0, stride = new_cache_43_internal_tensor_assign_1_stride_0, x = reshape_210)[name = tensor("slice_by_index_42")]; + tensor reshape_211_shape_0 = const()[name = tensor("reshape_211_shape_0"), val = tensor([-1])]; + tensor reshape_211 = reshape(shape = reshape_211_shape_0, x = slice_by_index_42)[name = tensor("reshape_211")]; + tensor reshape_212_shape_0 = const()[name = tensor("reshape_212_shape_0"), val = tensor([-1])]; + tensor reshape_212_cast_fp16 = reshape(shape = reshape_212_shape_0, x = var_9275_cast_fp16)[name = tensor("reshape_212_cast_fp16")]; + tensor reshape_213_shape_0 = const()[name = tensor("reshape_213_shape_0"), val = tensor([-1])]; + tensor reshape_213_cast_fp16 = reshape(shape = reshape_213_shape_0, x = cache21_to_fp16)[name = tensor("reshape_213_cast_fp16")]; + tensor scatter_42_mode_0 = const()[name = tensor("scatter_42_mode_0"), val = tensor("update")]; + tensor scatter_42_axis_0 = const()[name = tensor("scatter_42_axis_0"), val = tensor(0)]; + tensor scatter_42_validate_indices_0 = const()[name = tensor("scatter_42_validate_indices_0"), val = tensor(false)]; + tensor scatter_42_cast_fp16 = scatter(axis = scatter_42_axis_0, data = reshape_213_cast_fp16, indices = reshape_211, mode = scatter_42_mode_0, updates = reshape_212_cast_fp16, validate_indices = scatter_42_validate_indices_0)[name = tensor("scatter_42_cast_fp16")]; + tensor reshape_214_cast_fp16 = reshape(shape = shape_90, x = scatter_42_cast_fp16)[name = tensor("reshape_214_cast_fp16")]; + tensor var_9283_begin_0 = const()[name = tensor("op_9283_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor var_9283_end_0 = const()[name = tensor("op_9283_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor var_9283_end_mask_0 = const()[name = tensor("op_9283_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_9283_squeeze_mask_0 = const()[name = tensor("op_9283_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor var_9283_cast_fp16 = slice_by_index(begin = var_9283_begin_0, end = var_9283_end_0, end_mask = var_9283_end_mask_0, squeeze_mask = var_9283_squeeze_mask_0, x = reshape_214_cast_fp16)[name = tensor("op_9283_cast_fp16")]; + tensor var_9285_axis_0 = const()[name = tensor("op_9285_axis_0"), val = tensor(1)]; + tensor var_9285_mode_0 = const()[name = tensor("op_9285_mode_0"), val = tensor("update")]; + tensor var_9285_validate_indices_0 = const()[name = tensor("op_9285_validate_indices_0"), val = tensor(false)]; + tensor var_9285_cast_fp16 = scatter_along_axis(axis = var_9285_axis_0, data = var_9283_cast_fp16, indices = write_indices_43, mode = var_9285_mode_0, updates = v_43_cast_fp16, validate_indices = var_9285_validate_indices_0)[name = tensor("op_9285_cast_fp16")]; + tensor concat_151 = const()[name = tensor("concat_151"), val = tensor([1, 0, 0, 0, 0])]; + tensor concat_152 = const()[name = tensor("concat_152"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_43_internal_tensor_assign_2_stride_0 = const()[name = tensor("new_cache_43_internal_tensor_assign_2_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_43_internal_tensor_assign_2_begin_mask_0 = const()[name = tensor("new_cache_43_internal_tensor_assign_2_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_43_internal_tensor_assign_2_end_mask_0 = const()[name = tensor("new_cache_43_internal_tensor_assign_2_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_43_internal_tensor_assign_2_squeeze_mask_0 = const()[name = tensor("new_cache_43_internal_tensor_assign_2_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_91 = const()[name = tensor("shape_91"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_43 = const()[name = tensor("reduce_prod_43"), val = tensor(1048576)]; + tensor range_1d_43_start_0 = const()[name = tensor("range_1d_43_start_0"), val = tensor(0)]; + tensor range_1d_43_step_0 = const()[name = tensor("range_1d_43_step_0"), val = tensor(1)]; + tensor range_1d_43 = range_1d(end = reduce_prod_43, start = range_1d_43_start_0, step = range_1d_43_step_0)[name = tensor("range_1d_43")]; + tensor reshape_215 = reshape(shape = shape_91, x = range_1d_43)[name = tensor("reshape_215")]; + tensor slice_by_index_43 = slice_by_index(begin = concat_151, begin_mask = new_cache_43_internal_tensor_assign_2_begin_mask_0, end = concat_152, end_mask = new_cache_43_internal_tensor_assign_2_end_mask_0, squeeze_mask = new_cache_43_internal_tensor_assign_2_squeeze_mask_0, stride = new_cache_43_internal_tensor_assign_2_stride_0, x = reshape_215)[name = tensor("slice_by_index_43")]; + tensor reshape_216_shape_0 = const()[name = tensor("reshape_216_shape_0"), val = tensor([-1])]; + tensor reshape_216 = reshape(shape = reshape_216_shape_0, x = slice_by_index_43)[name = tensor("reshape_216")]; + tensor reshape_217_shape_0 = const()[name = tensor("reshape_217_shape_0"), val = tensor([-1])]; + tensor reshape_217_cast_fp16 = reshape(shape = reshape_217_shape_0, x = var_9285_cast_fp16)[name = tensor("reshape_217_cast_fp16")]; + tensor reshape_218_shape_0 = const()[name = tensor("reshape_218_shape_0"), val = tensor([-1])]; + tensor reshape_218_cast_fp16 = reshape(shape = reshape_218_shape_0, x = reshape_214_cast_fp16)[name = tensor("reshape_218_cast_fp16")]; + tensor scatter_43_mode_0 = const()[name = tensor("scatter_43_mode_0"), val = tensor("update")]; + tensor scatter_43_axis_0 = const()[name = tensor("scatter_43_axis_0"), val = tensor(0)]; + tensor scatter_43_validate_indices_0 = const()[name = tensor("scatter_43_validate_indices_0"), val = tensor(false)]; + tensor scatter_43_cast_fp16 = scatter(axis = scatter_43_axis_0, data = reshape_218_cast_fp16, indices = reshape_216, mode = scatter_43_mode_0, updates = reshape_217_cast_fp16, validate_indices = scatter_43_validate_indices_0)[name = tensor("scatter_43_cast_fp16")]; + tensor reshape_219_cast_fp16 = reshape(shape = shape_91, x = scatter_43_cast_fp16)[name = tensor("reshape_219_cast_fp16")]; + tensor reshape_219_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("reshape_219_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor keys_127_begin_0 = const()[name = tensor("keys_127_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor keys_127_end_0 = const()[name = tensor("keys_127_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor keys_127_end_mask_0 = const()[name = tensor("keys_127_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor keys_127_squeeze_mask_0 = const()[name = tensor("keys_127_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor keys_127_cast_fp16 = slice_by_index(begin = keys_127_begin_0, end = keys_127_end_0, end_mask = keys_127_end_mask_0, squeeze_mask = keys_127_squeeze_mask_0, x = reshape_219_cast_fp16)[name = tensor("keys_127_cast_fp16")]; + tensor values_127_begin_0 = const()[name = tensor("values_127_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor values_127_end_0 = const()[name = tensor("values_127_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor values_127_end_mask_0 = const()[name = tensor("values_127_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor values_127_squeeze_mask_0 = const()[name = tensor("values_127_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor values_127_cast_fp16 = slice_by_index(begin = values_127_begin_0, end = values_127_end_0, end_mask = values_127_end_mask_0, squeeze_mask = values_127_squeeze_mask_0, x = reshape_219_cast_fp16)[name = tensor("values_127_cast_fp16")]; + tensor var_9297_cast_fp16 = not_equal(x = keys_127_cast_fp16, y = keys_127_cast_fp16)[name = tensor("op_9297_cast_fp16")]; + tensor keys_129_cast_fp16 = select(a = var_504_to_fp16, b = keys_127_cast_fp16, cond = var_9297_cast_fp16)[name = tensor("keys_129_cast_fp16")]; + tensor var_9305_cast_fp16 = not_equal(x = values_127_cast_fp16, y = values_127_cast_fp16)[name = tensor("op_9305_cast_fp16")]; + tensor values_129_cast_fp16 = select(a = var_504_to_fp16, b = values_127_cast_fp16, cond = var_9305_cast_fp16)[name = tensor("values_129_cast_fp16")]; + tensor var_9329 = const()[name = tensor("op_9329"), val = tensor([0, 2, 1, 3])]; + tensor var_9342 = const()[name = tensor("op_9342"), val = tensor([1, 1, 1])]; + tensor var_9343_cast_fp16 = reshape(shape = var_9342, x = position21_to_fp16)[name = tensor("op_9343_cast_fp16")]; + tensor var_9360_to_fp16 = const()[name = tensor("op_9360_to_fp16"), val = tensor(0x1p+0)]; + tensor valid_len_43_cast_fp16 = add(x = var_9343_cast_fp16, y = var_9360_to_fp16)[name = tensor("valid_len_43_cast_fp16")]; + tensor valid_mask_43_cast_fp16 = less(x = k_positions_1_promoted_to_fp16, y = valid_len_43_cast_fp16)[name = tensor("valid_mask_43_cast_fp16")]; + tensor causal_mask_43_cast_fp16 = less_equal(x = k_positions_1_promoted_to_fp16, y = var_9343_cast_fp16)[name = tensor("causal_mask_43_cast_fp16")]; + tensor attn_mask_85 = logical_and(x = valid_mask_43_cast_fp16, y = causal_mask_43_cast_fp16)[name = tensor("attn_mask_85")]; + tensor attn_mask_87_axes_0 = const()[name = tensor("attn_mask_87_axes_0"), val = tensor([1])]; + tensor attn_mask_87 = expand_dims(axes = attn_mask_87_axes_0, x = attn_mask_85)[name = tensor("attn_mask_87")]; + tensor var_9378_transpose_x_0 = const()[name = tensor("op_9378_transpose_x_0"), val = tensor(false)]; + tensor var_9378_transpose_y_0 = const()[name = tensor("op_9378_transpose_y_0"), val = tensor(false)]; + tensor transpose_114_perm_0 = const()[name = tensor("transpose_114_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_115_perm_0 = const()[name = tensor("transpose_115_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_115 = transpose(perm = transpose_115_perm_0, x = keys_129_cast_fp16)[name = tensor("transpose_129")]; + tensor transpose_114 = transpose(perm = transpose_114_perm_0, x = q_129_cast_fp16)[name = tensor("transpose_130")]; + tensor var_9378_cast_fp16 = matmul(transpose_x = var_9378_transpose_x_0, transpose_y = var_9378_transpose_y_0, x = transpose_114, y = transpose_115)[name = tensor("op_9378_cast_fp16")]; + tensor var_9372_to_fp16 = const()[name = tensor("op_9372_to_fp16"), val = tensor([0x1p-3])]; + tensor attn_weights_127_cast_fp16 = mul(x = var_9378_cast_fp16, y = var_9372_to_fp16)[name = tensor("attn_weights_127_cast_fp16")]; + tensor var_9380 = logical_not(x = attn_mask_87)[name = tensor("op_9380")]; + tensor var_9381_to_fp16 = const()[name = tensor("op_9381_to_fp16"), val = tensor(-inf)]; + tensor attn_weights_129_cast_fp16 = select(a = var_9381_to_fp16, b = attn_weights_127_cast_fp16, cond = var_9380)[name = tensor("attn_weights_129_cast_fp16")]; + tensor var_9383 = const()[name = tensor("op_9383"), val = tensor(-1)]; + tensor attn_weights_131_cast_fp16 = softmax(axis = var_9383, x = attn_weights_129_cast_fp16)[name = tensor("attn_weights_131_cast_fp16")]; + tensor attn_output_43_transpose_x_0 = const()[name = tensor("attn_output_43_transpose_x_0"), val = tensor(false)]; + tensor attn_output_43_transpose_y_0 = const()[name = tensor("attn_output_43_transpose_y_0"), val = tensor(false)]; + tensor values_131_cast_fp16 = transpose(perm = var_9329, x = values_129_cast_fp16)[name = tensor("transpose_131")]; + tensor attn_output_43_cast_fp16 = matmul(transpose_x = attn_output_43_transpose_x_0, transpose_y = attn_output_43_transpose_y_0, x = attn_weights_131_cast_fp16, y = values_131_cast_fp16)[name = tensor("attn_output_43_cast_fp16")]; + tensor var_9391 = const()[name = tensor("op_9391"), val = tensor([0, 2, 1, 3])]; + tensor var_9394 = const()[name = tensor("op_9394"), val = tensor([1, 1, 1024])]; + tensor var_9392_cast_fp16 = transpose(perm = var_9391, x = attn_output_43_cast_fp16)[name = tensor("transpose_128")]; + tensor input_215_cast_fp16 = reshape(shape = var_9394, x = var_9392_cast_fp16)[name = tensor("input_215_cast_fp16")]; + tensor linear_86_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = attn21_out_proj_weight_cast_fp16, x = input_215_cast_fp16)[name = tensor("linear_86_cast_fp16")]; + tensor var_9400_to_fp16 = const()[name = tensor("op_9400_to_fp16"), val = tensor(0x1p+0)]; + tensor var_9401_cast_fp16 = add(x = position21_to_fp16, y = var_9400_to_fp16)[name = tensor("op_9401_cast_fp16")]; + tensor var_9401_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_9401_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor input_217_cast_fp16 = add(x = input_213_cast_fp16, y = linear_86_cast_fp16)[name = tensor("input_217_cast_fp16")]; + tensor input_219_axes_0 = const()[name = tensor("input_219_axes_0"), val = tensor([-1])]; + tensor norm21_2_weight_to_fp16 = const()[name = tensor("norm21_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(303769344)))]; + tensor norm21_2_bias_to_fp16 = const()[name = tensor("norm21_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(303771456)))]; + tensor var_9405_to_fp16 = const()[name = tensor("op_9405_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_219_cast_fp16 = layer_norm(axes = input_219_axes_0, beta = norm21_2_bias_to_fp16, epsilon = var_9405_to_fp16, gamma = norm21_2_weight_to_fp16, x = input_217_cast_fp16)[name = tensor("input_219_cast_fp16")]; + tensor linear_87_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = linear21_1_weight_cast_fp16, x = input_219_cast_fp16)[name = tensor("linear_87_cast_fp16")]; + tensor input_221_mode_0 = const()[name = tensor("input_221_mode_0"), val = tensor("EXACT")]; + tensor input_221_cast_fp16 = gelu(mode = input_221_mode_0, x = linear_87_cast_fp16)[name = tensor("input_221_cast_fp16")]; + tensor linear_88_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = linear21_2_weight_cast_fp16, x = input_221_cast_fp16)[name = tensor("linear_88_cast_fp16")]; + tensor input_223_cast_fp16 = add(x = input_217_cast_fp16, y = linear_88_cast_fp16)[name = tensor("input_223_cast_fp16")]; + tensor x_45_axes_0 = const()[name = tensor("x_45_axes_0"), val = tensor([-1])]; + tensor norm22_1_weight_to_fp16 = const()[name = tensor("norm22_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(303773568)))]; + tensor norm22_1_bias_to_fp16 = const()[name = tensor("norm22_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(303775680)))]; + tensor var_9422_to_fp16 = const()[name = tensor("op_9422_to_fp16"), val = tensor(0x1.5p-17)]; + tensor x_45_cast_fp16 = layer_norm(axes = x_45_axes_0, beta = norm22_1_bias_to_fp16, epsilon = var_9422_to_fp16, gamma = norm22_1_weight_to_fp16, x = input_223_cast_fp16)[name = tensor("x_45_cast_fp16")]; + tensor linear_89_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = attn22_in_proj_weight_cast_fp16, x = x_45_cast_fp16)[name = tensor("linear_89_cast_fp16")]; + tensor var_9458 = const()[name = tensor("op_9458"), val = tensor([1, 1, 3, 16, 64])]; + tensor qkv_45_cast_fp16 = reshape(shape = var_9458, x = linear_89_cast_fp16)[name = tensor("qkv_45_cast_fp16")]; + tensor q_133_begin_0 = const()[name = tensor("q_133_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor q_133_end_0 = const()[name = tensor("q_133_end_0"), val = tensor([1, 1, 1, 16, 64])]; + tensor q_133_end_mask_0 = const()[name = tensor("q_133_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor q_133_squeeze_mask_0 = const()[name = tensor("q_133_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor q_133_cast_fp16 = slice_by_index(begin = q_133_begin_0, end = q_133_end_0, end_mask = q_133_end_mask_0, squeeze_mask = q_133_squeeze_mask_0, x = qkv_45_cast_fp16)[name = tensor("q_133_cast_fp16")]; + tensor k_89_begin_0 = const()[name = tensor("k_89_begin_0"), val = tensor([0, 0, 1, 0, 0])]; + tensor k_89_end_0 = const()[name = tensor("k_89_end_0"), val = tensor([1, 1, 2, 16, 64])]; + tensor k_89_end_mask_0 = const()[name = tensor("k_89_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor k_89_squeeze_mask_0 = const()[name = tensor("k_89_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor k_89_cast_fp16 = slice_by_index(begin = k_89_begin_0, end = k_89_end_0, end_mask = k_89_end_mask_0, squeeze_mask = k_89_squeeze_mask_0, x = qkv_45_cast_fp16)[name = tensor("k_89_cast_fp16")]; + tensor v_45_begin_0 = const()[name = tensor("v_45_begin_0"), val = tensor([0, 0, 2, 0, 0])]; + tensor v_45_end_0 = const()[name = tensor("v_45_end_0"), val = tensor([1, 1, 3, 16, 64])]; + tensor v_45_end_mask_0 = const()[name = tensor("v_45_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor v_45_squeeze_mask_0 = const()[name = tensor("v_45_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor v_45_cast_fp16 = slice_by_index(begin = v_45_begin_0, end = v_45_end_0, end_mask = v_45_end_mask_0, squeeze_mask = v_45_squeeze_mask_0, x = qkv_45_cast_fp16)[name = tensor("v_45_cast_fp16")]; + tensor var_9562 = const()[name = tensor("op_9562"), val = tensor([1, 1, 1, 1])]; + tensor position22_to_fp16_dtype_0 = const()[name = tensor("position22_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor position22_to_fp16 = cast(dtype = position22_to_fp16_dtype_0, x = position22)[name = tensor("cast_443")]; + tensor ts_137_cast_fp16 = reshape(shape = var_9562, x = position22_to_fp16)[name = tensor("ts_137_cast_fp16")]; + tensor var_9566 = const()[name = tensor("op_9566"), val = tensor([1, 1, 16, 32, 2])]; + tensor q_complex_45_cast_fp16 = reshape(shape = var_9566, x = q_133_cast_fp16)[name = tensor("q_complex_45_cast_fp16")]; + tensor var_9570 = const()[name = tensor("op_9570"), val = tensor([1, 1, 16, 32, 2])]; + tensor k_complex_45_cast_fp16 = reshape(shape = var_9570, x = k_89_cast_fp16)[name = tensor("k_complex_45_cast_fp16")]; + tensor var_9574_begin_0 = const()[name = tensor("op_9574_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_9574_end_0 = const()[name = tensor("op_9574_end_0"), val = tensor([1, 1, 16, 32, 1])]; + tensor var_9574_end_mask_0 = const()[name = tensor("op_9574_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_9574_squeeze_mask_0 = const()[name = tensor("op_9574_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_9574_cast_fp16 = slice_by_index(begin = var_9574_begin_0, end = var_9574_end_0, end_mask = var_9574_end_mask_0, squeeze_mask = var_9574_squeeze_mask_0, x = q_complex_45_cast_fp16)[name = tensor("op_9574_cast_fp16")]; + tensor var_9582_begin_0 = const()[name = tensor("op_9582_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_9582_end_0 = const()[name = tensor("op_9582_end_0"), val = tensor([1, 1, 16, 32, 2])]; + tensor var_9582_end_mask_0 = const()[name = tensor("op_9582_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_9582_squeeze_mask_0 = const()[name = tensor("op_9582_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_9582_cast_fp16 = slice_by_index(begin = var_9582_begin_0, end = var_9582_end_0, end_mask = var_9582_end_mask_0, squeeze_mask = var_9582_squeeze_mask_0, x = q_complex_45_cast_fp16)[name = tensor("op_9582_cast_fp16")]; + tensor var_9590_begin_0 = const()[name = tensor("op_9590_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_9590_end_0 = const()[name = tensor("op_9590_end_0"), val = tensor([1, 1, 16, 32, 1])]; + tensor var_9590_end_mask_0 = const()[name = tensor("op_9590_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_9590_squeeze_mask_0 = const()[name = tensor("op_9590_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_9590_cast_fp16 = slice_by_index(begin = var_9590_begin_0, end = var_9590_end_0, end_mask = var_9590_end_mask_0, squeeze_mask = var_9590_squeeze_mask_0, x = k_complex_45_cast_fp16)[name = tensor("op_9590_cast_fp16")]; + tensor var_9598_begin_0 = const()[name = tensor("op_9598_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_9598_end_0 = const()[name = tensor("op_9598_end_0"), val = tensor([1, 1, 16, 32, 2])]; + tensor var_9598_end_mask_0 = const()[name = tensor("op_9598_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_9598_squeeze_mask_0 = const()[name = tensor("op_9598_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_9598_cast_fp16 = slice_by_index(begin = var_9598_begin_0, end = var_9598_end_0, end_mask = var_9598_end_mask_0, squeeze_mask = var_9598_squeeze_mask_0, x = k_complex_45_cast_fp16)[name = tensor("op_9598_cast_fp16")]; + tensor freqs_45_to_fp16 = const()[name = tensor("freqs_45_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(303777792)))]; + tensor var_9604_cast_fp16 = mul(x = freqs_45_to_fp16, y = ts_137_cast_fp16)[name = tensor("op_9604_cast_fp16")]; + tensor rotr_45_cast_fp16 = cos(x = var_9604_cast_fp16)[name = tensor("rotr_45_cast_fp16")]; + tensor roti_45_cast_fp16 = sin(x = var_9604_cast_fp16)[name = tensor("roti_45_cast_fp16")]; + tensor var_9608_cast_fp16 = mul(x = var_9574_cast_fp16, y = rotr_45_cast_fp16)[name = tensor("op_9608_cast_fp16")]; + tensor var_9609_cast_fp16 = mul(x = var_9582_cast_fp16, y = roti_45_cast_fp16)[name = tensor("op_9609_cast_fp16")]; + tensor qor_89_cast_fp16 = sub(x = var_9608_cast_fp16, y = var_9609_cast_fp16)[name = tensor("qor_89_cast_fp16")]; + tensor var_9612_cast_fp16 = mul(x = var_9574_cast_fp16, y = roti_45_cast_fp16)[name = tensor("op_9612_cast_fp16")]; + tensor var_9613_cast_fp16 = mul(x = var_9582_cast_fp16, y = rotr_45_cast_fp16)[name = tensor("op_9613_cast_fp16")]; + tensor qoi_89_cast_fp16 = add(x = var_9612_cast_fp16, y = var_9613_cast_fp16)[name = tensor("qoi_89_cast_fp16")]; + tensor var_9616_cast_fp16 = mul(x = var_9590_cast_fp16, y = rotr_45_cast_fp16)[name = tensor("op_9616_cast_fp16")]; + tensor var_9617_cast_fp16 = mul(x = var_9598_cast_fp16, y = roti_45_cast_fp16)[name = tensor("op_9617_cast_fp16")]; + tensor kor_89_cast_fp16 = sub(x = var_9616_cast_fp16, y = var_9617_cast_fp16)[name = tensor("kor_89_cast_fp16")]; + tensor var_9620_cast_fp16 = mul(x = var_9590_cast_fp16, y = roti_45_cast_fp16)[name = tensor("op_9620_cast_fp16")]; + tensor var_9621_cast_fp16 = mul(x = var_9598_cast_fp16, y = rotr_45_cast_fp16)[name = tensor("op_9621_cast_fp16")]; + tensor koi_89_cast_fp16 = add(x = var_9620_cast_fp16, y = var_9621_cast_fp16)[name = tensor("koi_89_cast_fp16")]; + tensor qo_45_axis_0 = const()[name = tensor("qo_45_axis_0"), val = tensor(-1)]; + tensor qo_45_cast_fp16 = stack(axis = qo_45_axis_0, values = (qor_89_cast_fp16, qoi_89_cast_fp16))[name = tensor("qo_45_cast_fp16")]; + tensor ko_45_axis_0 = const()[name = tensor("ko_45_axis_0"), val = tensor(-1)]; + tensor ko_45_cast_fp16 = stack(axis = ko_45_axis_0, values = (kor_89_cast_fp16, koi_89_cast_fp16))[name = tensor("ko_45_cast_fp16")]; + tensor var_9650 = const()[name = tensor("op_9650"), val = tensor([1, 1, 16, 64])]; + tensor q_135_cast_fp16 = reshape(shape = var_9650, x = qo_45_cast_fp16)[name = tensor("q_135_cast_fp16")]; + tensor var_9652 = const()[name = tensor("op_9652"), val = tensor([1, 1, 16, 64])]; + tensor k_91_cast_fp16 = reshape(shape = var_9652, x = ko_45_cast_fp16)[name = tensor("k_91_cast_fp16")]; + tensor _inversed_9674_y_0_to_fp16 = const()[name = tensor("_inversed_9674_y_0_to_fp16"), val = tensor(0x1p-9)]; + tensor _inversed_9674_cast_fp16 = mul(x = ts_137_cast_fp16, y = _inversed_9674_y_0_to_fp16)[name = tensor("_inversed_9674_cast_fp16")]; + tensor var_9675_cast_fp16 = floor(x = _inversed_9674_cast_fp16)[name = tensor("op_9675_cast_fp16")]; + tensor var_9676_to_fp16 = const()[name = tensor("op_9676_to_fp16"), val = tensor(0x1p+9)]; + tensor var_9677_cast_fp16 = mul(x = var_9675_cast_fp16, y = var_9676_to_fp16)[name = tensor("op_9677_cast_fp16")]; + tensor write_indices_float_91_cast_fp16 = sub(x = ts_137_cast_fp16, y = var_9677_cast_fp16)[name = tensor("write_indices_float_91_cast_fp16")]; + tensor var_9684_dtype_0 = const()[name = tensor("op_9684_dtype_0"), val = tensor("int32")]; + tensor write_indices_45_reps_0 = const()[name = tensor("write_indices_45_reps_0"), val = tensor([1, 1, 16, 64])]; + tensor write_indices_float_91_cast_fp16_to_int32 = cast(dtype = var_9684_dtype_0, x = write_indices_float_91_cast_fp16)[name = tensor("cast_442")]; + tensor write_indices_45 = tile(reps = write_indices_45_reps_0, x = write_indices_float_91_cast_fp16_to_int32)[name = tensor("write_indices_45")]; + tensor var_9692_begin_0 = const()[name = tensor("op_9692_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_9692_end_0 = const()[name = tensor("op_9692_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor var_9692_end_mask_0 = const()[name = tensor("op_9692_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_9692_squeeze_mask_0 = const()[name = tensor("op_9692_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor cache22_to_fp16_dtype_0 = const()[name = tensor("cache22_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor cache22_to_fp16 = cast(dtype = cache22_to_fp16_dtype_0, x = cache22)[name = tensor("cast_441")]; + tensor var_9692_cast_fp16 = slice_by_index(begin = var_9692_begin_0, end = var_9692_end_0, end_mask = var_9692_end_mask_0, squeeze_mask = var_9692_squeeze_mask_0, x = cache22_to_fp16)[name = tensor("op_9692_cast_fp16")]; + tensor var_9694_axis_0 = const()[name = tensor("op_9694_axis_0"), val = tensor(1)]; + tensor var_9694_mode_0 = const()[name = tensor("op_9694_mode_0"), val = tensor("update")]; + tensor var_9694_validate_indices_0 = const()[name = tensor("op_9694_validate_indices_0"), val = tensor(false)]; + tensor var_9694_cast_fp16 = scatter_along_axis(axis = var_9694_axis_0, data = var_9692_cast_fp16, indices = write_indices_45, mode = var_9694_mode_0, updates = k_91_cast_fp16, validate_indices = var_9694_validate_indices_0)[name = tensor("op_9694_cast_fp16")]; + tensor concat_156 = const()[name = tensor("concat_156"), val = tensor([0, 0, 0, 0, 0])]; + tensor concat_157 = const()[name = tensor("concat_157"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_45_internal_tensor_assign_1_stride_0 = const()[name = tensor("new_cache_45_internal_tensor_assign_1_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_45_internal_tensor_assign_1_begin_mask_0 = const()[name = tensor("new_cache_45_internal_tensor_assign_1_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_45_internal_tensor_assign_1_end_mask_0 = const()[name = tensor("new_cache_45_internal_tensor_assign_1_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_45_internal_tensor_assign_1_squeeze_mask_0 = const()[name = tensor("new_cache_45_internal_tensor_assign_1_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_92 = const()[name = tensor("shape_92"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_44 = const()[name = tensor("reduce_prod_44"), val = tensor(1048576)]; + tensor range_1d_44_start_0 = const()[name = tensor("range_1d_44_start_0"), val = tensor(0)]; + tensor range_1d_44_step_0 = const()[name = tensor("range_1d_44_step_0"), val = tensor(1)]; + tensor range_1d_44 = range_1d(end = reduce_prod_44, start = range_1d_44_start_0, step = range_1d_44_step_0)[name = tensor("range_1d_44")]; + tensor reshape_220 = reshape(shape = shape_92, x = range_1d_44)[name = tensor("reshape_220")]; + tensor slice_by_index_44 = slice_by_index(begin = concat_156, begin_mask = new_cache_45_internal_tensor_assign_1_begin_mask_0, end = concat_157, end_mask = new_cache_45_internal_tensor_assign_1_end_mask_0, squeeze_mask = new_cache_45_internal_tensor_assign_1_squeeze_mask_0, stride = new_cache_45_internal_tensor_assign_1_stride_0, x = reshape_220)[name = tensor("slice_by_index_44")]; + tensor reshape_221_shape_0 = const()[name = tensor("reshape_221_shape_0"), val = tensor([-1])]; + tensor reshape_221 = reshape(shape = reshape_221_shape_0, x = slice_by_index_44)[name = tensor("reshape_221")]; + tensor reshape_222_shape_0 = const()[name = tensor("reshape_222_shape_0"), val = tensor([-1])]; + tensor reshape_222_cast_fp16 = reshape(shape = reshape_222_shape_0, x = var_9694_cast_fp16)[name = tensor("reshape_222_cast_fp16")]; + tensor reshape_223_shape_0 = const()[name = tensor("reshape_223_shape_0"), val = tensor([-1])]; + tensor reshape_223_cast_fp16 = reshape(shape = reshape_223_shape_0, x = cache22_to_fp16)[name = tensor("reshape_223_cast_fp16")]; + tensor scatter_44_mode_0 = const()[name = tensor("scatter_44_mode_0"), val = tensor("update")]; + tensor scatter_44_axis_0 = const()[name = tensor("scatter_44_axis_0"), val = tensor(0)]; + tensor scatter_44_validate_indices_0 = const()[name = tensor("scatter_44_validate_indices_0"), val = tensor(false)]; + tensor scatter_44_cast_fp16 = scatter(axis = scatter_44_axis_0, data = reshape_223_cast_fp16, indices = reshape_221, mode = scatter_44_mode_0, updates = reshape_222_cast_fp16, validate_indices = scatter_44_validate_indices_0)[name = tensor("scatter_44_cast_fp16")]; + tensor reshape_224_cast_fp16 = reshape(shape = shape_92, x = scatter_44_cast_fp16)[name = tensor("reshape_224_cast_fp16")]; + tensor var_9702_begin_0 = const()[name = tensor("op_9702_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor var_9702_end_0 = const()[name = tensor("op_9702_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor var_9702_end_mask_0 = const()[name = tensor("op_9702_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_9702_squeeze_mask_0 = const()[name = tensor("op_9702_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor var_9702_cast_fp16 = slice_by_index(begin = var_9702_begin_0, end = var_9702_end_0, end_mask = var_9702_end_mask_0, squeeze_mask = var_9702_squeeze_mask_0, x = reshape_224_cast_fp16)[name = tensor("op_9702_cast_fp16")]; + tensor var_9704_axis_0 = const()[name = tensor("op_9704_axis_0"), val = tensor(1)]; + tensor var_9704_mode_0 = const()[name = tensor("op_9704_mode_0"), val = tensor("update")]; + tensor var_9704_validate_indices_0 = const()[name = tensor("op_9704_validate_indices_0"), val = tensor(false)]; + tensor var_9704_cast_fp16 = scatter_along_axis(axis = var_9704_axis_0, data = var_9702_cast_fp16, indices = write_indices_45, mode = var_9704_mode_0, updates = v_45_cast_fp16, validate_indices = var_9704_validate_indices_0)[name = tensor("op_9704_cast_fp16")]; + tensor concat_158 = const()[name = tensor("concat_158"), val = tensor([1, 0, 0, 0, 0])]; + tensor concat_159 = const()[name = tensor("concat_159"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_45_internal_tensor_assign_2_stride_0 = const()[name = tensor("new_cache_45_internal_tensor_assign_2_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_45_internal_tensor_assign_2_begin_mask_0 = const()[name = tensor("new_cache_45_internal_tensor_assign_2_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_45_internal_tensor_assign_2_end_mask_0 = const()[name = tensor("new_cache_45_internal_tensor_assign_2_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_45_internal_tensor_assign_2_squeeze_mask_0 = const()[name = tensor("new_cache_45_internal_tensor_assign_2_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_93 = const()[name = tensor("shape_93"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_45 = const()[name = tensor("reduce_prod_45"), val = tensor(1048576)]; + tensor range_1d_45_start_0 = const()[name = tensor("range_1d_45_start_0"), val = tensor(0)]; + tensor range_1d_45_step_0 = const()[name = tensor("range_1d_45_step_0"), val = tensor(1)]; + tensor range_1d_45 = range_1d(end = reduce_prod_45, start = range_1d_45_start_0, step = range_1d_45_step_0)[name = tensor("range_1d_45")]; + tensor reshape_225 = reshape(shape = shape_93, x = range_1d_45)[name = tensor("reshape_225")]; + tensor slice_by_index_45 = slice_by_index(begin = concat_158, begin_mask = new_cache_45_internal_tensor_assign_2_begin_mask_0, end = concat_159, end_mask = new_cache_45_internal_tensor_assign_2_end_mask_0, squeeze_mask = new_cache_45_internal_tensor_assign_2_squeeze_mask_0, stride = new_cache_45_internal_tensor_assign_2_stride_0, x = reshape_225)[name = tensor("slice_by_index_45")]; + tensor reshape_226_shape_0 = const()[name = tensor("reshape_226_shape_0"), val = tensor([-1])]; + tensor reshape_226 = reshape(shape = reshape_226_shape_0, x = slice_by_index_45)[name = tensor("reshape_226")]; + tensor reshape_227_shape_0 = const()[name = tensor("reshape_227_shape_0"), val = tensor([-1])]; + tensor reshape_227_cast_fp16 = reshape(shape = reshape_227_shape_0, x = var_9704_cast_fp16)[name = tensor("reshape_227_cast_fp16")]; + tensor reshape_228_shape_0 = const()[name = tensor("reshape_228_shape_0"), val = tensor([-1])]; + tensor reshape_228_cast_fp16 = reshape(shape = reshape_228_shape_0, x = reshape_224_cast_fp16)[name = tensor("reshape_228_cast_fp16")]; + tensor scatter_45_mode_0 = const()[name = tensor("scatter_45_mode_0"), val = tensor("update")]; + tensor scatter_45_axis_0 = const()[name = tensor("scatter_45_axis_0"), val = tensor(0)]; + tensor scatter_45_validate_indices_0 = const()[name = tensor("scatter_45_validate_indices_0"), val = tensor(false)]; + tensor scatter_45_cast_fp16 = scatter(axis = scatter_45_axis_0, data = reshape_228_cast_fp16, indices = reshape_226, mode = scatter_45_mode_0, updates = reshape_227_cast_fp16, validate_indices = scatter_45_validate_indices_0)[name = tensor("scatter_45_cast_fp16")]; + tensor reshape_229_cast_fp16 = reshape(shape = shape_93, x = scatter_45_cast_fp16)[name = tensor("reshape_229_cast_fp16")]; + tensor reshape_229_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("reshape_229_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor keys_133_begin_0 = const()[name = tensor("keys_133_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor keys_133_end_0 = const()[name = tensor("keys_133_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor keys_133_end_mask_0 = const()[name = tensor("keys_133_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor keys_133_squeeze_mask_0 = const()[name = tensor("keys_133_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor keys_133_cast_fp16 = slice_by_index(begin = keys_133_begin_0, end = keys_133_end_0, end_mask = keys_133_end_mask_0, squeeze_mask = keys_133_squeeze_mask_0, x = reshape_229_cast_fp16)[name = tensor("keys_133_cast_fp16")]; + tensor values_133_begin_0 = const()[name = tensor("values_133_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor values_133_end_0 = const()[name = tensor("values_133_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor values_133_end_mask_0 = const()[name = tensor("values_133_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor values_133_squeeze_mask_0 = const()[name = tensor("values_133_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor values_133_cast_fp16 = slice_by_index(begin = values_133_begin_0, end = values_133_end_0, end_mask = values_133_end_mask_0, squeeze_mask = values_133_squeeze_mask_0, x = reshape_229_cast_fp16)[name = tensor("values_133_cast_fp16")]; + tensor var_9716_cast_fp16 = not_equal(x = keys_133_cast_fp16, y = keys_133_cast_fp16)[name = tensor("op_9716_cast_fp16")]; + tensor keys_135_cast_fp16 = select(a = var_504_to_fp16, b = keys_133_cast_fp16, cond = var_9716_cast_fp16)[name = tensor("keys_135_cast_fp16")]; + tensor var_9724_cast_fp16 = not_equal(x = values_133_cast_fp16, y = values_133_cast_fp16)[name = tensor("op_9724_cast_fp16")]; + tensor values_135_cast_fp16 = select(a = var_504_to_fp16, b = values_133_cast_fp16, cond = var_9724_cast_fp16)[name = tensor("values_135_cast_fp16")]; + tensor var_9748 = const()[name = tensor("op_9748"), val = tensor([0, 2, 1, 3])]; + tensor var_9761 = const()[name = tensor("op_9761"), val = tensor([1, 1, 1])]; + tensor var_9762_cast_fp16 = reshape(shape = var_9761, x = position22_to_fp16)[name = tensor("op_9762_cast_fp16")]; + tensor var_9779_to_fp16 = const()[name = tensor("op_9779_to_fp16"), val = tensor(0x1p+0)]; + tensor valid_len_45_cast_fp16 = add(x = var_9762_cast_fp16, y = var_9779_to_fp16)[name = tensor("valid_len_45_cast_fp16")]; + tensor valid_mask_45_cast_fp16 = less(x = k_positions_1_promoted_to_fp16, y = valid_len_45_cast_fp16)[name = tensor("valid_mask_45_cast_fp16")]; + tensor causal_mask_45_cast_fp16 = less_equal(x = k_positions_1_promoted_to_fp16, y = var_9762_cast_fp16)[name = tensor("causal_mask_45_cast_fp16")]; + tensor attn_mask_89 = logical_and(x = valid_mask_45_cast_fp16, y = causal_mask_45_cast_fp16)[name = tensor("attn_mask_89")]; + tensor attn_mask_91_axes_0 = const()[name = tensor("attn_mask_91_axes_0"), val = tensor([1])]; + tensor attn_mask_91 = expand_dims(axes = attn_mask_91_axes_0, x = attn_mask_89)[name = tensor("attn_mask_91")]; + tensor var_9797_transpose_x_0 = const()[name = tensor("op_9797_transpose_x_0"), val = tensor(false)]; + tensor var_9797_transpose_y_0 = const()[name = tensor("op_9797_transpose_y_0"), val = tensor(false)]; + tensor transpose_116_perm_0 = const()[name = tensor("transpose_116_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_117_perm_0 = const()[name = tensor("transpose_117_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_117 = transpose(perm = transpose_117_perm_0, x = keys_135_cast_fp16)[name = tensor("transpose_125")]; + tensor transpose_116 = transpose(perm = transpose_116_perm_0, x = q_135_cast_fp16)[name = tensor("transpose_126")]; + tensor var_9797_cast_fp16 = matmul(transpose_x = var_9797_transpose_x_0, transpose_y = var_9797_transpose_y_0, x = transpose_116, y = transpose_117)[name = tensor("op_9797_cast_fp16")]; + tensor var_9791_to_fp16 = const()[name = tensor("op_9791_to_fp16"), val = tensor([0x1p-3])]; + tensor attn_weights_133_cast_fp16 = mul(x = var_9797_cast_fp16, y = var_9791_to_fp16)[name = tensor("attn_weights_133_cast_fp16")]; + tensor var_9799 = logical_not(x = attn_mask_91)[name = tensor("op_9799")]; + tensor var_9800_to_fp16 = const()[name = tensor("op_9800_to_fp16"), val = tensor(-inf)]; + tensor attn_weights_135_cast_fp16 = select(a = var_9800_to_fp16, b = attn_weights_133_cast_fp16, cond = var_9799)[name = tensor("attn_weights_135_cast_fp16")]; + tensor var_9802 = const()[name = tensor("op_9802"), val = tensor(-1)]; + tensor attn_weights_137_cast_fp16 = softmax(axis = var_9802, x = attn_weights_135_cast_fp16)[name = tensor("attn_weights_137_cast_fp16")]; + tensor attn_output_45_transpose_x_0 = const()[name = tensor("attn_output_45_transpose_x_0"), val = tensor(false)]; + tensor attn_output_45_transpose_y_0 = const()[name = tensor("attn_output_45_transpose_y_0"), val = tensor(false)]; + tensor values_137_cast_fp16 = transpose(perm = var_9748, x = values_135_cast_fp16)[name = tensor("transpose_127")]; + tensor attn_output_45_cast_fp16 = matmul(transpose_x = attn_output_45_transpose_x_0, transpose_y = attn_output_45_transpose_y_0, x = attn_weights_137_cast_fp16, y = values_137_cast_fp16)[name = tensor("attn_output_45_cast_fp16")]; + tensor var_9810 = const()[name = tensor("op_9810"), val = tensor([0, 2, 1, 3])]; + tensor var_9813 = const()[name = tensor("op_9813"), val = tensor([1, 1, 1024])]; + tensor var_9811_cast_fp16 = transpose(perm = var_9810, x = attn_output_45_cast_fp16)[name = tensor("transpose_124")]; + tensor input_225_cast_fp16 = reshape(shape = var_9813, x = var_9811_cast_fp16)[name = tensor("input_225_cast_fp16")]; + tensor linear_90_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = attn22_out_proj_weight_cast_fp16, x = input_225_cast_fp16)[name = tensor("linear_90_cast_fp16")]; + tensor var_9819_to_fp16 = const()[name = tensor("op_9819_to_fp16"), val = tensor(0x1p+0)]; + tensor var_9820_cast_fp16 = add(x = position22_to_fp16, y = var_9819_to_fp16)[name = tensor("op_9820_cast_fp16")]; + tensor var_9820_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_9820_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor input_227_cast_fp16 = add(x = input_223_cast_fp16, y = linear_90_cast_fp16)[name = tensor("input_227_cast_fp16")]; + tensor input_229_axes_0 = const()[name = tensor("input_229_axes_0"), val = tensor([-1])]; + tensor norm22_2_weight_to_fp16 = const()[name = tensor("norm22_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(303777920)))]; + tensor norm22_2_bias_to_fp16 = const()[name = tensor("norm22_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(303780032)))]; + tensor var_9824_to_fp16 = const()[name = tensor("op_9824_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_229_cast_fp16 = layer_norm(axes = input_229_axes_0, beta = norm22_2_bias_to_fp16, epsilon = var_9824_to_fp16, gamma = norm22_2_weight_to_fp16, x = input_227_cast_fp16)[name = tensor("input_229_cast_fp16")]; + tensor linear_91_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = linear22_1_weight_cast_fp16, x = input_229_cast_fp16)[name = tensor("linear_91_cast_fp16")]; + tensor input_231_mode_0 = const()[name = tensor("input_231_mode_0"), val = tensor("EXACT")]; + tensor input_231_cast_fp16 = gelu(mode = input_231_mode_0, x = linear_91_cast_fp16)[name = tensor("input_231_cast_fp16")]; + tensor linear_92_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = linear22_2_weight_cast_fp16, x = input_231_cast_fp16)[name = tensor("linear_92_cast_fp16")]; + tensor input_233_cast_fp16 = add(x = input_227_cast_fp16, y = linear_92_cast_fp16)[name = tensor("input_233_cast_fp16")]; + tensor x_axes_0 = const()[name = tensor("x_axes_0"), val = tensor([-1])]; + tensor norm23_1_weight_to_fp16 = const()[name = tensor("norm23_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(303782144)))]; + tensor norm23_1_bias_to_fp16 = const()[name = tensor("norm23_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(303784256)))]; + tensor var_9841_to_fp16 = const()[name = tensor("op_9841_to_fp16"), val = tensor(0x1.5p-17)]; + tensor x_cast_fp16 = layer_norm(axes = x_axes_0, beta = norm23_1_bias_to_fp16, epsilon = var_9841_to_fp16, gamma = norm23_1_weight_to_fp16, x = input_233_cast_fp16)[name = tensor("x_cast_fp16")]; + tensor linear_93_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = attn23_in_proj_weight_cast_fp16, x = x_cast_fp16)[name = tensor("linear_93_cast_fp16")]; + tensor var_9877 = const()[name = tensor("op_9877"), val = tensor([1, 1, 3, 16, 64])]; + tensor qkv_cast_fp16 = reshape(shape = var_9877, x = linear_93_cast_fp16)[name = tensor("qkv_cast_fp16")]; + tensor q_139_begin_0 = const()[name = tensor("q_139_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor q_139_end_0 = const()[name = tensor("q_139_end_0"), val = tensor([1, 1, 1, 16, 64])]; + tensor q_139_end_mask_0 = const()[name = tensor("q_139_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor q_139_squeeze_mask_0 = const()[name = tensor("q_139_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor q_139_cast_fp16 = slice_by_index(begin = q_139_begin_0, end = q_139_end_0, end_mask = q_139_end_mask_0, squeeze_mask = q_139_squeeze_mask_0, x = qkv_cast_fp16)[name = tensor("q_139_cast_fp16")]; + tensor k_93_begin_0 = const()[name = tensor("k_93_begin_0"), val = tensor([0, 0, 1, 0, 0])]; + tensor k_93_end_0 = const()[name = tensor("k_93_end_0"), val = tensor([1, 1, 2, 16, 64])]; + tensor k_93_end_mask_0 = const()[name = tensor("k_93_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor k_93_squeeze_mask_0 = const()[name = tensor("k_93_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor k_93_cast_fp16 = slice_by_index(begin = k_93_begin_0, end = k_93_end_0, end_mask = k_93_end_mask_0, squeeze_mask = k_93_squeeze_mask_0, x = qkv_cast_fp16)[name = tensor("k_93_cast_fp16")]; + tensor v_begin_0 = const()[name = tensor("v_begin_0"), val = tensor([0, 0, 2, 0, 0])]; + tensor v_end_0 = const()[name = tensor("v_end_0"), val = tensor([1, 1, 3, 16, 64])]; + tensor v_end_mask_0 = const()[name = tensor("v_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor v_squeeze_mask_0 = const()[name = tensor("v_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor v_cast_fp16 = slice_by_index(begin = v_begin_0, end = v_end_0, end_mask = v_end_mask_0, squeeze_mask = v_squeeze_mask_0, x = qkv_cast_fp16)[name = tensor("v_cast_fp16")]; + tensor var_9981 = const()[name = tensor("op_9981"), val = tensor([1, 1, 1, 1])]; + tensor position23_to_fp16_dtype_0 = const()[name = tensor("position23_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor position23_to_fp16 = cast(dtype = position23_to_fp16_dtype_0, x = position23)[name = tensor("cast_438")]; + tensor ts_cast_fp16 = reshape(shape = var_9981, x = position23_to_fp16)[name = tensor("ts_cast_fp16")]; + tensor var_9985 = const()[name = tensor("op_9985"), val = tensor([1, 1, 16, 32, 2])]; + tensor q_complex_cast_fp16 = reshape(shape = var_9985, x = q_139_cast_fp16)[name = tensor("q_complex_cast_fp16")]; + tensor var_9989 = const()[name = tensor("op_9989"), val = tensor([1, 1, 16, 32, 2])]; + tensor k_complex_cast_fp16 = reshape(shape = var_9989, x = k_93_cast_fp16)[name = tensor("k_complex_cast_fp16")]; + tensor var_9993_begin_0 = const()[name = tensor("op_9993_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_9993_end_0 = const()[name = tensor("op_9993_end_0"), val = tensor([1, 1, 16, 32, 1])]; + tensor var_9993_end_mask_0 = const()[name = tensor("op_9993_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_9993_squeeze_mask_0 = const()[name = tensor("op_9993_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_9993_cast_fp16 = slice_by_index(begin = var_9993_begin_0, end = var_9993_end_0, end_mask = var_9993_end_mask_0, squeeze_mask = var_9993_squeeze_mask_0, x = q_complex_cast_fp16)[name = tensor("op_9993_cast_fp16")]; + tensor var_10001_begin_0 = const()[name = tensor("op_10001_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_10001_end_0 = const()[name = tensor("op_10001_end_0"), val = tensor([1, 1, 16, 32, 2])]; + tensor var_10001_end_mask_0 = const()[name = tensor("op_10001_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_10001_squeeze_mask_0 = const()[name = tensor("op_10001_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_10001_cast_fp16 = slice_by_index(begin = var_10001_begin_0, end = var_10001_end_0, end_mask = var_10001_end_mask_0, squeeze_mask = var_10001_squeeze_mask_0, x = q_complex_cast_fp16)[name = tensor("op_10001_cast_fp16")]; + tensor var_10009_begin_0 = const()[name = tensor("op_10009_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_10009_end_0 = const()[name = tensor("op_10009_end_0"), val = tensor([1, 1, 16, 32, 1])]; + tensor var_10009_end_mask_0 = const()[name = tensor("op_10009_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_10009_squeeze_mask_0 = const()[name = tensor("op_10009_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_10009_cast_fp16 = slice_by_index(begin = var_10009_begin_0, end = var_10009_end_0, end_mask = var_10009_end_mask_0, squeeze_mask = var_10009_squeeze_mask_0, x = k_complex_cast_fp16)[name = tensor("op_10009_cast_fp16")]; + tensor var_10017_begin_0 = const()[name = tensor("op_10017_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_10017_end_0 = const()[name = tensor("op_10017_end_0"), val = tensor([1, 1, 16, 32, 2])]; + tensor var_10017_end_mask_0 = const()[name = tensor("op_10017_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_10017_squeeze_mask_0 = const()[name = tensor("op_10017_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_10017_cast_fp16 = slice_by_index(begin = var_10017_begin_0, end = var_10017_end_0, end_mask = var_10017_end_mask_0, squeeze_mask = var_10017_squeeze_mask_0, x = k_complex_cast_fp16)[name = tensor("op_10017_cast_fp16")]; + tensor freqs_to_fp16 = const()[name = tensor("freqs_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(303786368)))]; + tensor var_10023_cast_fp16 = mul(x = freqs_to_fp16, y = ts_cast_fp16)[name = tensor("op_10023_cast_fp16")]; + tensor rotr_cast_fp16 = cos(x = var_10023_cast_fp16)[name = tensor("rotr_cast_fp16")]; + tensor roti_cast_fp16 = sin(x = var_10023_cast_fp16)[name = tensor("roti_cast_fp16")]; + tensor var_10027_cast_fp16 = mul(x = var_9993_cast_fp16, y = rotr_cast_fp16)[name = tensor("op_10027_cast_fp16")]; + tensor var_10028_cast_fp16 = mul(x = var_10001_cast_fp16, y = roti_cast_fp16)[name = tensor("op_10028_cast_fp16")]; + tensor qor_93_cast_fp16 = sub(x = var_10027_cast_fp16, y = var_10028_cast_fp16)[name = tensor("qor_93_cast_fp16")]; + tensor var_10031_cast_fp16 = mul(x = var_9993_cast_fp16, y = roti_cast_fp16)[name = tensor("op_10031_cast_fp16")]; + tensor var_10032_cast_fp16 = mul(x = var_10001_cast_fp16, y = rotr_cast_fp16)[name = tensor("op_10032_cast_fp16")]; + tensor qoi_93_cast_fp16 = add(x = var_10031_cast_fp16, y = var_10032_cast_fp16)[name = tensor("qoi_93_cast_fp16")]; + tensor var_10035_cast_fp16 = mul(x = var_10009_cast_fp16, y = rotr_cast_fp16)[name = tensor("op_10035_cast_fp16")]; + tensor var_10036_cast_fp16 = mul(x = var_10017_cast_fp16, y = roti_cast_fp16)[name = tensor("op_10036_cast_fp16")]; + tensor kor_93_cast_fp16 = sub(x = var_10035_cast_fp16, y = var_10036_cast_fp16)[name = tensor("kor_93_cast_fp16")]; + tensor var_10039_cast_fp16 = mul(x = var_10009_cast_fp16, y = roti_cast_fp16)[name = tensor("op_10039_cast_fp16")]; + tensor var_10040_cast_fp16 = mul(x = var_10017_cast_fp16, y = rotr_cast_fp16)[name = tensor("op_10040_cast_fp16")]; + tensor koi_93_cast_fp16 = add(x = var_10039_cast_fp16, y = var_10040_cast_fp16)[name = tensor("koi_93_cast_fp16")]; + tensor qo_axis_0 = const()[name = tensor("qo_axis_0"), val = tensor(-1)]; + tensor qo_cast_fp16 = stack(axis = qo_axis_0, values = (qor_93_cast_fp16, qoi_93_cast_fp16))[name = tensor("qo_cast_fp16")]; + tensor ko_axis_0 = const()[name = tensor("ko_axis_0"), val = tensor(-1)]; + tensor ko_cast_fp16 = stack(axis = ko_axis_0, values = (kor_93_cast_fp16, koi_93_cast_fp16))[name = tensor("ko_cast_fp16")]; + tensor var_10069 = const()[name = tensor("op_10069"), val = tensor([1, 1, 16, 64])]; + tensor q_141_cast_fp16 = reshape(shape = var_10069, x = qo_cast_fp16)[name = tensor("q_141_cast_fp16")]; + tensor var_10071 = const()[name = tensor("op_10071"), val = tensor([1, 1, 16, 64])]; + tensor k_cast_fp16 = reshape(shape = var_10071, x = ko_cast_fp16)[name = tensor("k_cast_fp16")]; + tensor _inversed_10093_y_0_to_fp16 = const()[name = tensor("_inversed_10093_y_0_to_fp16"), val = tensor(0x1p-9)]; + tensor _inversed_10093_cast_fp16 = mul(x = ts_cast_fp16, y = _inversed_10093_y_0_to_fp16)[name = tensor("_inversed_10093_cast_fp16")]; + tensor var_10094_cast_fp16 = floor(x = _inversed_10093_cast_fp16)[name = tensor("op_10094_cast_fp16")]; + tensor var_10095_to_fp16 = const()[name = tensor("op_10095_to_fp16"), val = tensor(0x1p+9)]; + tensor var_10096_cast_fp16 = mul(x = var_10094_cast_fp16, y = var_10095_to_fp16)[name = tensor("op_10096_cast_fp16")]; + tensor write_indices_float_cast_fp16 = sub(x = ts_cast_fp16, y = var_10096_cast_fp16)[name = tensor("write_indices_float_cast_fp16")]; + tensor var_10103_dtype_0 = const()[name = tensor("op_10103_dtype_0"), val = tensor("int32")]; + tensor write_indices_reps_0 = const()[name = tensor("write_indices_reps_0"), val = tensor([1, 1, 16, 64])]; + tensor write_indices_float_cast_fp16_to_int32 = cast(dtype = var_10103_dtype_0, x = write_indices_float_cast_fp16)[name = tensor("cast_437")]; + tensor write_indices = tile(reps = write_indices_reps_0, x = write_indices_float_cast_fp16_to_int32)[name = tensor("write_indices")]; + tensor var_10111_begin_0 = const()[name = tensor("op_10111_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_10111_end_0 = const()[name = tensor("op_10111_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor var_10111_end_mask_0 = const()[name = tensor("op_10111_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_10111_squeeze_mask_0 = const()[name = tensor("op_10111_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor cache23_to_fp16_dtype_0 = const()[name = tensor("cache23_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor cache23_to_fp16 = cast(dtype = cache23_to_fp16_dtype_0, x = cache23)[name = tensor("cast_436")]; + tensor var_10111_cast_fp16 = slice_by_index(begin = var_10111_begin_0, end = var_10111_end_0, end_mask = var_10111_end_mask_0, squeeze_mask = var_10111_squeeze_mask_0, x = cache23_to_fp16)[name = tensor("op_10111_cast_fp16")]; + tensor var_10113_axis_0 = const()[name = tensor("op_10113_axis_0"), val = tensor(1)]; + tensor var_10113_mode_0 = const()[name = tensor("op_10113_mode_0"), val = tensor("update")]; + tensor var_10113_validate_indices_0 = const()[name = tensor("op_10113_validate_indices_0"), val = tensor(false)]; + tensor var_10113_cast_fp16 = scatter_along_axis(axis = var_10113_axis_0, data = var_10111_cast_fp16, indices = write_indices, mode = var_10113_mode_0, updates = k_cast_fp16, validate_indices = var_10113_validate_indices_0)[name = tensor("op_10113_cast_fp16")]; + tensor concat_163 = const()[name = tensor("concat_163"), val = tensor([0, 0, 0, 0, 0])]; + tensor concat_164 = const()[name = tensor("concat_164"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_internal_tensor_assign_1_stride_0 = const()[name = tensor("new_cache_internal_tensor_assign_1_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_internal_tensor_assign_1_begin_mask_0 = const()[name = tensor("new_cache_internal_tensor_assign_1_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_internal_tensor_assign_1_end_mask_0 = const()[name = tensor("new_cache_internal_tensor_assign_1_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_internal_tensor_assign_1_squeeze_mask_0 = const()[name = tensor("new_cache_internal_tensor_assign_1_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_94 = const()[name = tensor("shape_94"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_46 = const()[name = tensor("reduce_prod_46"), val = tensor(1048576)]; + tensor range_1d_46_start_0 = const()[name = tensor("range_1d_46_start_0"), val = tensor(0)]; + tensor range_1d_46_step_0 = const()[name = tensor("range_1d_46_step_0"), val = tensor(1)]; + tensor range_1d_46 = range_1d(end = reduce_prod_46, start = range_1d_46_start_0, step = range_1d_46_step_0)[name = tensor("range_1d_46")]; + tensor reshape_230 = reshape(shape = shape_94, x = range_1d_46)[name = tensor("reshape_230")]; + tensor slice_by_index_46 = slice_by_index(begin = concat_163, begin_mask = new_cache_internal_tensor_assign_1_begin_mask_0, end = concat_164, end_mask = new_cache_internal_tensor_assign_1_end_mask_0, squeeze_mask = new_cache_internal_tensor_assign_1_squeeze_mask_0, stride = new_cache_internal_tensor_assign_1_stride_0, x = reshape_230)[name = tensor("slice_by_index_46")]; + tensor reshape_231_shape_0 = const()[name = tensor("reshape_231_shape_0"), val = tensor([-1])]; + tensor reshape_231 = reshape(shape = reshape_231_shape_0, x = slice_by_index_46)[name = tensor("reshape_231")]; + tensor reshape_232_shape_0 = const()[name = tensor("reshape_232_shape_0"), val = tensor([-1])]; + tensor reshape_232_cast_fp16 = reshape(shape = reshape_232_shape_0, x = var_10113_cast_fp16)[name = tensor("reshape_232_cast_fp16")]; + tensor reshape_233_shape_0 = const()[name = tensor("reshape_233_shape_0"), val = tensor([-1])]; + tensor reshape_233_cast_fp16 = reshape(shape = reshape_233_shape_0, x = cache23_to_fp16)[name = tensor("reshape_233_cast_fp16")]; + tensor scatter_46_mode_0 = const()[name = tensor("scatter_46_mode_0"), val = tensor("update")]; + tensor scatter_46_axis_0 = const()[name = tensor("scatter_46_axis_0"), val = tensor(0)]; + tensor scatter_46_validate_indices_0 = const()[name = tensor("scatter_46_validate_indices_0"), val = tensor(false)]; + tensor scatter_46_cast_fp16 = scatter(axis = scatter_46_axis_0, data = reshape_233_cast_fp16, indices = reshape_231, mode = scatter_46_mode_0, updates = reshape_232_cast_fp16, validate_indices = scatter_46_validate_indices_0)[name = tensor("scatter_46_cast_fp16")]; + tensor reshape_234_cast_fp16 = reshape(shape = shape_94, x = scatter_46_cast_fp16)[name = tensor("reshape_234_cast_fp16")]; + tensor var_10121_begin_0 = const()[name = tensor("op_10121_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor var_10121_end_0 = const()[name = tensor("op_10121_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor var_10121_end_mask_0 = const()[name = tensor("op_10121_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_10121_squeeze_mask_0 = const()[name = tensor("op_10121_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor var_10121_cast_fp16 = slice_by_index(begin = var_10121_begin_0, end = var_10121_end_0, end_mask = var_10121_end_mask_0, squeeze_mask = var_10121_squeeze_mask_0, x = reshape_234_cast_fp16)[name = tensor("op_10121_cast_fp16")]; + tensor var_10123_axis_0 = const()[name = tensor("op_10123_axis_0"), val = tensor(1)]; + tensor var_10123_mode_0 = const()[name = tensor("op_10123_mode_0"), val = tensor("update")]; + tensor var_10123_validate_indices_0 = const()[name = tensor("op_10123_validate_indices_0"), val = tensor(false)]; + tensor var_10123_cast_fp16 = scatter_along_axis(axis = var_10123_axis_0, data = var_10121_cast_fp16, indices = write_indices, mode = var_10123_mode_0, updates = v_cast_fp16, validate_indices = var_10123_validate_indices_0)[name = tensor("op_10123_cast_fp16")]; + tensor concat_165 = const()[name = tensor("concat_165"), val = tensor([1, 0, 0, 0, 0])]; + tensor concat_166 = const()[name = tensor("concat_166"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_internal_tensor_assign_2_stride_0 = const()[name = tensor("new_cache_internal_tensor_assign_2_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_internal_tensor_assign_2_begin_mask_0 = const()[name = tensor("new_cache_internal_tensor_assign_2_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_internal_tensor_assign_2_end_mask_0 = const()[name = tensor("new_cache_internal_tensor_assign_2_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_internal_tensor_assign_2_squeeze_mask_0 = const()[name = tensor("new_cache_internal_tensor_assign_2_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_95 = const()[name = tensor("shape_95"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_47 = const()[name = tensor("reduce_prod_47"), val = tensor(1048576)]; + tensor range_1d_47_start_0 = const()[name = tensor("range_1d_47_start_0"), val = tensor(0)]; + tensor range_1d_47_step_0 = const()[name = tensor("range_1d_47_step_0"), val = tensor(1)]; + tensor range_1d_47 = range_1d(end = reduce_prod_47, start = range_1d_47_start_0, step = range_1d_47_step_0)[name = tensor("range_1d_47")]; + tensor reshape_235 = reshape(shape = shape_95, x = range_1d_47)[name = tensor("reshape_235")]; + tensor slice_by_index_47 = slice_by_index(begin = concat_165, begin_mask = new_cache_internal_tensor_assign_2_begin_mask_0, end = concat_166, end_mask = new_cache_internal_tensor_assign_2_end_mask_0, squeeze_mask = new_cache_internal_tensor_assign_2_squeeze_mask_0, stride = new_cache_internal_tensor_assign_2_stride_0, x = reshape_235)[name = tensor("slice_by_index_47")]; + tensor reshape_236_shape_0 = const()[name = tensor("reshape_236_shape_0"), val = tensor([-1])]; + tensor reshape_236 = reshape(shape = reshape_236_shape_0, x = slice_by_index_47)[name = tensor("reshape_236")]; + tensor reshape_237_shape_0 = const()[name = tensor("reshape_237_shape_0"), val = tensor([-1])]; + tensor reshape_237_cast_fp16 = reshape(shape = reshape_237_shape_0, x = var_10123_cast_fp16)[name = tensor("reshape_237_cast_fp16")]; + tensor reshape_238_shape_0 = const()[name = tensor("reshape_238_shape_0"), val = tensor([-1])]; + tensor reshape_238_cast_fp16 = reshape(shape = reshape_238_shape_0, x = reshape_234_cast_fp16)[name = tensor("reshape_238_cast_fp16")]; + tensor scatter_47_mode_0 = const()[name = tensor("scatter_47_mode_0"), val = tensor("update")]; + tensor scatter_47_axis_0 = const()[name = tensor("scatter_47_axis_0"), val = tensor(0)]; + tensor scatter_47_validate_indices_0 = const()[name = tensor("scatter_47_validate_indices_0"), val = tensor(false)]; + tensor scatter_47_cast_fp16 = scatter(axis = scatter_47_axis_0, data = reshape_238_cast_fp16, indices = reshape_236, mode = scatter_47_mode_0, updates = reshape_237_cast_fp16, validate_indices = scatter_47_validate_indices_0)[name = tensor("scatter_47_cast_fp16")]; + tensor reshape_239_cast_fp16 = reshape(shape = shape_95, x = scatter_47_cast_fp16)[name = tensor("reshape_239_cast_fp16")]; + tensor reshape_239_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("reshape_239_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor keys_139_begin_0 = const()[name = tensor("keys_139_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor keys_139_end_0 = const()[name = tensor("keys_139_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor keys_139_end_mask_0 = const()[name = tensor("keys_139_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor keys_139_squeeze_mask_0 = const()[name = tensor("keys_139_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor keys_139_cast_fp16 = slice_by_index(begin = keys_139_begin_0, end = keys_139_end_0, end_mask = keys_139_end_mask_0, squeeze_mask = keys_139_squeeze_mask_0, x = reshape_239_cast_fp16)[name = tensor("keys_139_cast_fp16")]; + tensor values_139_begin_0 = const()[name = tensor("values_139_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor values_139_end_0 = const()[name = tensor("values_139_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor values_139_end_mask_0 = const()[name = tensor("values_139_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor values_139_squeeze_mask_0 = const()[name = tensor("values_139_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor values_139_cast_fp16 = slice_by_index(begin = values_139_begin_0, end = values_139_end_0, end_mask = values_139_end_mask_0, squeeze_mask = values_139_squeeze_mask_0, x = reshape_239_cast_fp16)[name = tensor("values_139_cast_fp16")]; + tensor var_10135_cast_fp16 = not_equal(x = keys_139_cast_fp16, y = keys_139_cast_fp16)[name = tensor("op_10135_cast_fp16")]; + tensor keys_141_cast_fp16 = select(a = var_504_to_fp16, b = keys_139_cast_fp16, cond = var_10135_cast_fp16)[name = tensor("keys_141_cast_fp16")]; + tensor var_10143_cast_fp16 = not_equal(x = values_139_cast_fp16, y = values_139_cast_fp16)[name = tensor("op_10143_cast_fp16")]; + tensor values_141_cast_fp16 = select(a = var_504_to_fp16, b = values_139_cast_fp16, cond = var_10143_cast_fp16)[name = tensor("values_141_cast_fp16")]; + tensor var_10167 = const()[name = tensor("op_10167"), val = tensor([0, 2, 1, 3])]; + tensor var_10180 = const()[name = tensor("op_10180"), val = tensor([1, 1, 1])]; + tensor var_10181_cast_fp16 = reshape(shape = var_10180, x = position23_to_fp16)[name = tensor("op_10181_cast_fp16")]; + tensor var_10198_to_fp16 = const()[name = tensor("op_10198_to_fp16"), val = tensor(0x1p+0)]; + tensor valid_len_cast_fp16 = add(x = var_10181_cast_fp16, y = var_10198_to_fp16)[name = tensor("valid_len_cast_fp16")]; + tensor valid_mask_cast_fp16 = less(x = k_positions_1_promoted_to_fp16, y = valid_len_cast_fp16)[name = tensor("valid_mask_cast_fp16")]; + tensor causal_mask_cast_fp16 = less_equal(x = k_positions_1_promoted_to_fp16, y = var_10181_cast_fp16)[name = tensor("causal_mask_cast_fp16")]; + tensor attn_mask_93 = logical_and(x = valid_mask_cast_fp16, y = causal_mask_cast_fp16)[name = tensor("attn_mask_93")]; + tensor attn_mask_axes_0 = const()[name = tensor("attn_mask_axes_0"), val = tensor([1])]; + tensor attn_mask = expand_dims(axes = attn_mask_axes_0, x = attn_mask_93)[name = tensor("attn_mask")]; + tensor var_10216_transpose_x_0 = const()[name = tensor("op_10216_transpose_x_0"), val = tensor(false)]; + tensor var_10216_transpose_y_0 = const()[name = tensor("op_10216_transpose_y_0"), val = tensor(false)]; + tensor transpose_118_perm_0 = const()[name = tensor("transpose_118_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_119_perm_0 = const()[name = tensor("transpose_119_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_119 = transpose(perm = transpose_119_perm_0, x = keys_141_cast_fp16)[name = tensor("transpose_121")]; + tensor transpose_118 = transpose(perm = transpose_118_perm_0, x = q_141_cast_fp16)[name = tensor("transpose_122")]; + tensor var_10216_cast_fp16 = matmul(transpose_x = var_10216_transpose_x_0, transpose_y = var_10216_transpose_y_0, x = transpose_118, y = transpose_119)[name = tensor("op_10216_cast_fp16")]; + tensor var_10210_to_fp16 = const()[name = tensor("op_10210_to_fp16"), val = tensor([0x1p-3])]; + tensor attn_weights_139_cast_fp16 = mul(x = var_10216_cast_fp16, y = var_10210_to_fp16)[name = tensor("attn_weights_139_cast_fp16")]; + tensor var_10218 = logical_not(x = attn_mask)[name = tensor("op_10218")]; + tensor var_10219_to_fp16 = const()[name = tensor("op_10219_to_fp16"), val = tensor(-inf)]; + tensor attn_weights_141_cast_fp16 = select(a = var_10219_to_fp16, b = attn_weights_139_cast_fp16, cond = var_10218)[name = tensor("attn_weights_141_cast_fp16")]; + tensor var_10221 = const()[name = tensor("op_10221"), val = tensor(-1)]; + tensor attn_weights_cast_fp16 = softmax(axis = var_10221, x = attn_weights_141_cast_fp16)[name = tensor("attn_weights_cast_fp16")]; + tensor attn_output_transpose_x_0 = const()[name = tensor("attn_output_transpose_x_0"), val = tensor(false)]; + tensor attn_output_transpose_y_0 = const()[name = tensor("attn_output_transpose_y_0"), val = tensor(false)]; + tensor values_cast_fp16 = transpose(perm = var_10167, x = values_141_cast_fp16)[name = tensor("transpose_123")]; + tensor attn_output_cast_fp16 = matmul(transpose_x = attn_output_transpose_x_0, transpose_y = attn_output_transpose_y_0, x = attn_weights_cast_fp16, y = values_cast_fp16)[name = tensor("attn_output_cast_fp16")]; + tensor var_10229 = const()[name = tensor("op_10229"), val = tensor([0, 2, 1, 3])]; + tensor var_10232 = const()[name = tensor("op_10232"), val = tensor([1, 1, 1024])]; + tensor var_10230_cast_fp16 = transpose(perm = var_10229, x = attn_output_cast_fp16)[name = tensor("transpose_120")]; + tensor input_235_cast_fp16 = reshape(shape = var_10232, x = var_10230_cast_fp16)[name = tensor("input_235_cast_fp16")]; + tensor linear_94_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = attn23_out_proj_weight_cast_fp16, x = input_235_cast_fp16)[name = tensor("linear_94_cast_fp16")]; + tensor var_10238_to_fp16 = const()[name = tensor("op_10238_to_fp16"), val = tensor(0x1p+0)]; + tensor var_10239_cast_fp16 = add(x = position23_to_fp16, y = var_10238_to_fp16)[name = tensor("op_10239_cast_fp16")]; + tensor var_10239_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_10239_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor input_237_cast_fp16 = add(x = input_233_cast_fp16, y = linear_94_cast_fp16)[name = tensor("input_237_cast_fp16")]; + tensor input_239_axes_0 = const()[name = tensor("input_239_axes_0"), val = tensor([-1])]; + tensor norm23_2_weight_to_fp16 = const()[name = tensor("norm23_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(303786496)))]; + tensor norm23_2_bias_to_fp16 = const()[name = tensor("norm23_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(303788608)))]; + tensor var_10243_to_fp16 = const()[name = tensor("op_10243_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_239_cast_fp16 = layer_norm(axes = input_239_axes_0, beta = norm23_2_bias_to_fp16, epsilon = var_10243_to_fp16, gamma = norm23_2_weight_to_fp16, x = input_237_cast_fp16)[name = tensor("input_239_cast_fp16")]; + tensor linear_95_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = linear23_1_weight_cast_fp16, x = input_239_cast_fp16)[name = tensor("linear_95_cast_fp16")]; + tensor input_241_mode_0 = const()[name = tensor("input_241_mode_0"), val = tensor("EXACT")]; + tensor input_241_cast_fp16 = gelu(mode = input_241_mode_0, x = linear_95_cast_fp16)[name = tensor("input_241_cast_fp16")]; + tensor linear_96_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = linear23_2_weight_cast_fp16, x = input_241_cast_fp16)[name = tensor("linear_96_cast_fp16")]; + tensor input_243_cast_fp16 = add(x = input_237_cast_fp16, y = linear_96_cast_fp16)[name = tensor("input_243_cast_fp16")]; + tensor input_axes_0 = const()[name = tensor("input_axes_0"), val = tensor([-1])]; + tensor out_norm_weight_to_fp16 = const()[name = tensor("out_norm_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(303790720)))]; + tensor out_norm_bias_to_fp16 = const()[name = tensor("out_norm_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(303792832)))]; + tensor var_10260_to_fp16 = const()[name = tensor("op_10260_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_cast_fp16 = layer_norm(axes = input_axes_0, beta = out_norm_bias_to_fp16, epsilon = var_10260_to_fp16, gamma = out_norm_weight_to_fp16, x = input_243_cast_fp16)[name = tensor("input_cast_fp16")]; + tensor input_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("input_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor out_eos_weight_to_fp16 = const()[name = tensor("out_eos_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(303794944)))]; + tensor out_eos_bias_to_fp16 = const()[name = tensor("out_eos_bias_to_fp16"), val = tensor([-0x1.64p-3])]; + tensor linear_97_cast_fp16 = linear(bias = out_eos_bias_to_fp16, weight = out_eos_weight_to_fp16, x = input_cast_fp16)[name = tensor("linear_97_cast_fp16")]; + tensor linear_97_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("linear_97_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor var_10268 = cast(dtype = linear_97_cast_fp16_to_fp32_dtype_0, x = linear_97_cast_fp16)[name = tensor("cast_432")]; + tensor input = cast(dtype = input_cast_fp16_to_fp32_dtype_0, x = input_cast_fp16)[name = tensor("cast_433")]; + tensor var_10239 = cast(dtype = var_10239_cast_fp16_to_fp32_dtype_0, x = var_10239_cast_fp16)[name = tensor("cast_434")]; + tensor new_cache_internal_tensor_assign_2 = cast(dtype = reshape_239_cast_fp16_to_fp32_dtype_0, x = reshape_239_cast_fp16)[name = tensor("cast_435")]; + tensor var_9820 = cast(dtype = var_9820_cast_fp16_to_fp32_dtype_0, x = var_9820_cast_fp16)[name = tensor("cast_439")]; + tensor new_cache_45_internal_tensor_assign_2 = cast(dtype = reshape_229_cast_fp16_to_fp32_dtype_0, x = reshape_229_cast_fp16)[name = tensor("cast_440")]; + tensor var_9401 = cast(dtype = var_9401_cast_fp16_to_fp32_dtype_0, x = var_9401_cast_fp16)[name = tensor("cast_444")]; + tensor new_cache_43_internal_tensor_assign_2 = cast(dtype = reshape_219_cast_fp16_to_fp32_dtype_0, x = reshape_219_cast_fp16)[name = tensor("cast_445")]; + tensor var_8982 = cast(dtype = var_8982_cast_fp16_to_fp32_dtype_0, x = var_8982_cast_fp16)[name = tensor("cast_449")]; + tensor new_cache_41_internal_tensor_assign_2 = cast(dtype = reshape_209_cast_fp16_to_fp32_dtype_0, x = reshape_209_cast_fp16)[name = tensor("cast_450")]; + tensor var_8563 = cast(dtype = var_8563_cast_fp16_to_fp32_dtype_0, x = var_8563_cast_fp16)[name = tensor("cast_454")]; + tensor new_cache_39_internal_tensor_assign_2 = cast(dtype = reshape_199_cast_fp16_to_fp32_dtype_0, x = reshape_199_cast_fp16)[name = tensor("cast_455")]; + tensor var_8144 = cast(dtype = var_8144_cast_fp16_to_fp32_dtype_0, x = var_8144_cast_fp16)[name = tensor("cast_459")]; + tensor new_cache_37_internal_tensor_assign_2 = cast(dtype = reshape_189_cast_fp16_to_fp32_dtype_0, x = reshape_189_cast_fp16)[name = tensor("cast_460")]; + tensor var_7725 = cast(dtype = var_7725_cast_fp16_to_fp32_dtype_0, x = var_7725_cast_fp16)[name = tensor("cast_464")]; + tensor new_cache_35_internal_tensor_assign_2 = cast(dtype = reshape_179_cast_fp16_to_fp32_dtype_0, x = reshape_179_cast_fp16)[name = tensor("cast_465")]; + tensor var_7306 = cast(dtype = var_7306_cast_fp16_to_fp32_dtype_0, x = var_7306_cast_fp16)[name = tensor("cast_469")]; + tensor new_cache_33_internal_tensor_assign_2 = cast(dtype = reshape_169_cast_fp16_to_fp32_dtype_0, x = reshape_169_cast_fp16)[name = tensor("cast_470")]; + tensor var_6887 = cast(dtype = var_6887_cast_fp16_to_fp32_dtype_0, x = var_6887_cast_fp16)[name = tensor("cast_474")]; + tensor new_cache_31_internal_tensor_assign_2 = cast(dtype = reshape_159_cast_fp16_to_fp32_dtype_0, x = reshape_159_cast_fp16)[name = tensor("cast_475")]; + tensor var_6468 = cast(dtype = var_6468_cast_fp16_to_fp32_dtype_0, x = var_6468_cast_fp16)[name = tensor("cast_479")]; + tensor new_cache_29_internal_tensor_assign_2 = cast(dtype = reshape_149_cast_fp16_to_fp32_dtype_0, x = reshape_149_cast_fp16)[name = tensor("cast_480")]; + tensor var_6049 = cast(dtype = var_6049_cast_fp16_to_fp32_dtype_0, x = var_6049_cast_fp16)[name = tensor("cast_484")]; + tensor new_cache_27_internal_tensor_assign_2 = cast(dtype = reshape_139_cast_fp16_to_fp32_dtype_0, x = reshape_139_cast_fp16)[name = tensor("cast_485")]; + tensor var_5630 = cast(dtype = var_5630_cast_fp16_to_fp32_dtype_0, x = var_5630_cast_fp16)[name = tensor("cast_489")]; + tensor new_cache_25_internal_tensor_assign_2 = cast(dtype = reshape_129_cast_fp16_to_fp32_dtype_0, x = reshape_129_cast_fp16)[name = tensor("cast_490")]; + tensor var_5211 = cast(dtype = var_5211_cast_fp16_to_fp32_dtype_0, x = var_5211_cast_fp16)[name = tensor("cast_494")]; + tensor new_cache_23_internal_tensor_assign_2 = cast(dtype = reshape_119_cast_fp16_to_fp32_dtype_0, x = reshape_119_cast_fp16)[name = tensor("cast_495")]; + tensor var_4792 = cast(dtype = var_4792_cast_fp16_to_fp32_dtype_0, x = var_4792_cast_fp16)[name = tensor("cast_499")]; + tensor new_cache_21_internal_tensor_assign_2 = cast(dtype = reshape_109_cast_fp16_to_fp32_dtype_0, x = reshape_109_cast_fp16)[name = tensor("cast_500")]; + tensor var_4373 = cast(dtype = var_4373_cast_fp16_to_fp32_dtype_0, x = var_4373_cast_fp16)[name = tensor("cast_504")]; + tensor new_cache_19_internal_tensor_assign_2 = cast(dtype = reshape_99_cast_fp16_to_fp32_dtype_0, x = reshape_99_cast_fp16)[name = tensor("cast_505")]; + tensor var_3954 = cast(dtype = var_3954_cast_fp16_to_fp32_dtype_0, x = var_3954_cast_fp16)[name = tensor("cast_509")]; + tensor new_cache_17_internal_tensor_assign_2 = cast(dtype = reshape_89_cast_fp16_to_fp32_dtype_0, x = reshape_89_cast_fp16)[name = tensor("cast_510")]; + tensor var_3535 = cast(dtype = var_3535_cast_fp16_to_fp32_dtype_0, x = var_3535_cast_fp16)[name = tensor("cast_514")]; + tensor new_cache_15_internal_tensor_assign_2 = cast(dtype = reshape_79_cast_fp16_to_fp32_dtype_0, x = reshape_79_cast_fp16)[name = tensor("cast_515")]; + tensor var_3116 = cast(dtype = var_3116_cast_fp16_to_fp32_dtype_0, x = var_3116_cast_fp16)[name = tensor("cast_519")]; + tensor new_cache_13_internal_tensor_assign_2 = cast(dtype = reshape_69_cast_fp16_to_fp32_dtype_0, x = reshape_69_cast_fp16)[name = tensor("cast_520")]; + tensor var_2697 = cast(dtype = var_2697_cast_fp16_to_fp32_dtype_0, x = var_2697_cast_fp16)[name = tensor("cast_524")]; + tensor new_cache_11_internal_tensor_assign_2 = cast(dtype = reshape_59_cast_fp16_to_fp32_dtype_0, x = reshape_59_cast_fp16)[name = tensor("cast_525")]; + tensor var_2278 = cast(dtype = var_2278_cast_fp16_to_fp32_dtype_0, x = var_2278_cast_fp16)[name = tensor("cast_529")]; + tensor new_cache_9_internal_tensor_assign_2 = cast(dtype = reshape_49_cast_fp16_to_fp32_dtype_0, x = reshape_49_cast_fp16)[name = tensor("cast_530")]; + tensor var_1859 = cast(dtype = var_1859_cast_fp16_to_fp32_dtype_0, x = var_1859_cast_fp16)[name = tensor("cast_534")]; + tensor new_cache_7_internal_tensor_assign_2 = cast(dtype = reshape_39_cast_fp16_to_fp32_dtype_0, x = reshape_39_cast_fp16)[name = tensor("cast_535")]; + tensor var_1440 = cast(dtype = var_1440_cast_fp16_to_fp32_dtype_0, x = var_1440_cast_fp16)[name = tensor("cast_539")]; + tensor new_cache_5_internal_tensor_assign_2 = cast(dtype = reshape_29_cast_fp16_to_fp32_dtype_0, x = reshape_29_cast_fp16)[name = tensor("cast_540")]; + tensor var_1021 = cast(dtype = var_1021_cast_fp16_to_fp32_dtype_0, x = var_1021_cast_fp16)[name = tensor("cast_544")]; + tensor new_cache_3_internal_tensor_assign_2 = cast(dtype = reshape_19_cast_fp16_to_fp32_dtype_0, x = reshape_19_cast_fp16)[name = tensor("cast_545")]; + tensor var_602 = cast(dtype = var_602_cast_fp16_to_fp32_dtype_0, x = var_602_cast_fp16)[name = tensor("cast_549")]; + tensor new_cache_1_internal_tensor_assign_2 = cast(dtype = reshape_9_cast_fp16_to_fp32_dtype_0, x = reshape_9_cast_fp16)[name = tensor("cast_550")]; + } -> (input, var_10268, new_cache_1_internal_tensor_assign_2, var_602, new_cache_3_internal_tensor_assign_2, var_1021, new_cache_5_internal_tensor_assign_2, var_1440, new_cache_7_internal_tensor_assign_2, var_1859, new_cache_9_internal_tensor_assign_2, var_2278, new_cache_11_internal_tensor_assign_2, var_2697, new_cache_13_internal_tensor_assign_2, var_3116, new_cache_15_internal_tensor_assign_2, var_3535, new_cache_17_internal_tensor_assign_2, var_3954, new_cache_19_internal_tensor_assign_2, var_4373, new_cache_21_internal_tensor_assign_2, var_4792, new_cache_23_internal_tensor_assign_2, var_5211, new_cache_25_internal_tensor_assign_2, var_5630, new_cache_27_internal_tensor_assign_2, var_6049, new_cache_29_internal_tensor_assign_2, var_6468, new_cache_31_internal_tensor_assign_2, var_6887, new_cache_33_internal_tensor_assign_2, var_7306, new_cache_35_internal_tensor_assign_2, var_7725, new_cache_37_internal_tensor_assign_2, var_8144, new_cache_39_internal_tensor_assign_2, var_8563, new_cache_41_internal_tensor_assign_2, var_8982, new_cache_43_internal_tensor_assign_2, var_9401, new_cache_45_internal_tensor_assign_2, var_9820, new_cache_internal_tensor_assign_2, var_10239); +} \ No newline at end of file diff --git a/v2.1/italian_24l/flowlm_stepv2.mlmodelc/weights/weight.bin b/v2.1/italian_24l/flowlm_stepv2.mlmodelc/weights/weight.bin new file mode 100644 index 0000000000000000000000000000000000000000..a9d657e8e9f6975bc10ae46021abf6cc853c8973 --- /dev/null +++ b/v2.1/italian_24l/flowlm_stepv2.mlmodelc/weights/weight.bin @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:a47fcb112b1a5296022980ba8dff9f050948d02f35d6265a8aea49117d871d41 +size 303797056 diff --git a/v2.1/italian_24l/flowlm_stepv2.mlpackage/Data/com.apple.CoreML/model.mlmodel b/v2.1/italian_24l/flowlm_stepv2.mlpackage/Data/com.apple.CoreML/model.mlmodel new file mode 100644 index 0000000000000000000000000000000000000000..399d0e7cdd63bc604b37184ab61bc957680a0a62 --- /dev/null +++ b/v2.1/italian_24l/flowlm_stepv2.mlpackage/Data/com.apple.CoreML/model.mlmodel @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:769a57ce1507881bfe131be53ec849beeed1e6378543ce8286cdd2edefa68f2b +size 843748 diff --git a/v2.1/italian_24l/flowlm_stepv2.mlpackage/Data/com.apple.CoreML/weights/weight.bin b/v2.1/italian_24l/flowlm_stepv2.mlpackage/Data/com.apple.CoreML/weights/weight.bin new file mode 100644 index 0000000000000000000000000000000000000000..a9d657e8e9f6975bc10ae46021abf6cc853c8973 --- /dev/null +++ b/v2.1/italian_24l/flowlm_stepv2.mlpackage/Data/com.apple.CoreML/weights/weight.bin @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:a47fcb112b1a5296022980ba8dff9f050948d02f35d6265a8aea49117d871d41 +size 303797056 diff --git a/v2.1/italian_24l/flowlm_stepv2.mlpackage/Manifest.json b/v2.1/italian_24l/flowlm_stepv2.mlpackage/Manifest.json new file mode 100644 index 0000000000000000000000000000000000000000..13020ca009c4be2e9f5755fd80b800dbc675fe45 --- /dev/null +++ b/v2.1/italian_24l/flowlm_stepv2.mlpackage/Manifest.json @@ -0,0 +1,18 @@ +{ + "fileFormatVersion": "1.0.0", + "itemInfoEntries": { + "31593508-45C7-4F30-B468-60472992616C": { + "author": "com.apple.CoreML", + "description": "CoreML Model Weights", + "name": "weights", + "path": "com.apple.CoreML/weights" + }, + "970ACD6F-280A-4A8B-BCC5-5CA6687CBCDB": { + "author": "com.apple.CoreML", + "description": "CoreML Model Specification", + "name": "model.mlmodel", + "path": "com.apple.CoreML/model.mlmodel" + } + }, + "rootModelIdentifier": "970ACD6F-280A-4A8B-BCC5-5CA6687CBCDB" +} diff --git a/v2.1/italian_24l/manifest.json b/v2.1/italian_24l/manifest.json new file mode 100644 index 0000000000000000000000000000000000000000..fc7da98920126d1261a9c47517b1303db1599db7 --- /dev/null +++ b/v2.1/italian_24l/manifest.json @@ -0,0 +1,27 @@ +{ + "version": "2.1", + "base": "v2", + "kind": "optimization-reconvert", + "self_contained": true, + "note": "Same weights as v2. Re-converted for speed: fused flow decoder (100% ANE), one-shot cond prefill, fp16 flowlm. NOT a finetune.", + "new": [ + "flow_decoder_fused", + "cond_prefill", + "flowlm_step" + ], + "copied_from_v2": [ + "flowlm_stepv2.mlpackage", + "flowlm_stepv2.mlmodelc", + "mimi_decoder.mlpackage", + "mimi_decoder.mlmodelc", + "constants", + "constants_bin" + ], + "compute_units": { + "flow_decoder_fused": "all (ANE)", + "cond_prefill": "all (GPU)", + "flowlm_step": "all (GPU, fp16)", + "flowlm_stepv2": "cpuAndGpu (GPU, int8 \u2014 fastest flowlm)", + "mimi_decoder": "cpuOnly" + } +} diff --git a/v2.1/italian_24l/mimi_decoder.mlmodelc/analytics/coremldata.bin b/v2.1/italian_24l/mimi_decoder.mlmodelc/analytics/coremldata.bin new file mode 100644 index 0000000000000000000000000000000000000000..e2f0d997fc613b29db8322580ec29d0c7d2695e2 --- /dev/null +++ b/v2.1/italian_24l/mimi_decoder.mlmodelc/analytics/coremldata.bin @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:6347d2a7390a5d32c18bf5ad60a4747aae36954cf0fc5ea26af164e202e20b06 +size 243 diff --git a/v2.1/italian_24l/mimi_decoder.mlmodelc/coremldata.bin b/v2.1/italian_24l/mimi_decoder.mlmodelc/coremldata.bin new file mode 100644 index 0000000000000000000000000000000000000000..d6812369b250598af855ebff0f794f7bfb46b995 --- /dev/null +++ b/v2.1/italian_24l/mimi_decoder.mlmodelc/coremldata.bin @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:ef490576856d64b232fbeca8c1e761e262d3d010f55be9ea8959cc9acd8ff31d +size 1777 diff --git a/v2.1/italian_24l/mimi_decoder.mlmodelc/metadata.json b/v2.1/italian_24l/mimi_decoder.mlmodelc/metadata.json new file mode 100644 index 0000000000000000000000000000000000000000..f66f9e7ff5be2ffbc44c8300f3501f67a77afaf6 --- /dev/null +++ b/v2.1/italian_24l/mimi_decoder.mlmodelc/metadata.json @@ -0,0 +1,573 @@ +[ + { + "metadataOutputVersion" : "3.0", + "storagePrecision" : "Float32", + "outputSchema" : [ + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1 × 1 × 1920)", + "shortDescription" : "", + "shape" : "[1, 1, 1920]", + "name" : "var_724", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1 × 512 × 16)", + "shortDescription" : "", + "shape" : "[1, 512, 16]", + "name" : "var_77", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 2 × 1 × 256 × 8 × 64)", + "shortDescription" : "", + "shape" : "[2, 1, 256, 8, 64]", + "name" : "var_210", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1)", + "shortDescription" : "", + "shape" : "[1]", + "name" : "var_741", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 2 × 1 × 256 × 8 × 64)", + "shortDescription" : "", + "shape" : "[2, 1, 256, 8, 64]", + "name" : "var_400", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1)", + "shortDescription" : "", + "shape" : "[1]", + "name" : "var_744", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1 × 512 × 6)", + "shortDescription" : "", + "shape" : "[1, 512, 6]", + "name" : "var_542", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1)", + "shortDescription" : "", + "shape" : "[1]", + "name" : "conv0_first", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1 × 256 × 6)", + "shortDescription" : "", + "shape" : "[1, 256, 6]", + "name" : "var_565", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1 × 256 × 2)", + "shortDescription" : "", + "shape" : "[1, 256, 2]", + "name" : "var_585", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1)", + "shortDescription" : "", + "shape" : "[1]", + "name" : "res0_conv0_first", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1 × 128 × 0)", + "shortDescription" : "", + "shape" : "[1, 128, 0]", + "name" : "res0_conv1_prev", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1)", + "shortDescription" : "", + "shape" : "[1]", + "name" : "res0_conv1_first", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1 × 128 × 5)", + "shortDescription" : "", + "shape" : "[1, 128, 5]", + "name" : "var_621", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1 × 128 × 2)", + "shortDescription" : "", + "shape" : "[1, 128, 2]", + "name" : "var_641", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1)", + "shortDescription" : "", + "shape" : "[1]", + "name" : "res1_conv0_first", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1 × 64 × 0)", + "shortDescription" : "", + "shape" : "[1, 64, 0]", + "name" : "res1_conv1_prev", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1)", + "shortDescription" : "", + "shape" : "[1]", + "name" : "res1_conv1_first", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1 × 64 × 4)", + "shortDescription" : "", + "shape" : "[1, 64, 4]", + "name" : "var_677", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1 × 64 × 2)", + "shortDescription" : "", + "shape" : "[1, 64, 2]", + "name" : "var_697", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1)", + "shortDescription" : "", + "shape" : "[1]", + "name" : "res2_conv0_first", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1 × 32 × 0)", + "shortDescription" : "", + "shape" : "[1, 32, 0]", + "name" : "res2_conv1_prev", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1)", + "shortDescription" : "", + "shape" : "[1]", + "name" : "res2_conv1_first", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1 × 64 × 2)", + "shortDescription" : "", + "shape" : "[1, 64, 2]", + "name" : "var_725", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1)", + "shortDescription" : "", + "shape" : "[1]", + "name" : "conv_final_first", + "type" : "MultiArray" + } + ], + "modelParameters" : [ + + ], + "specificationVersion" : 8, + "mlProgramOperationTypeHistogram" : { + "Ios17.notEqual" : 4, + "Tile" : 2, + "Ios17.lessEqual" : 2, + "Ios17.layerNorm" : 4, + "Stack" : 6, + "Ios17.greaterEqual" : 4, + "Ios17.scatterAlongAxis" : 4, + "Ios17.sub" : 17, + "Identity" : 11, + "Ios17.elu" : 10, + "Ios17.logicalNot" : 2, + "Ios17.sliceByIndex" : 35, + "Ios17.concat" : 9, + "Ios17.convTranspose" : 4, + "Ios17.floorDiv" : 4, + "Ios17.expandDims" : 7, + "Ios17.less" : 2, + "Ios17.mul" : 29, + "Ios17.transpose" : 10, + "Ios17.matmul" : 4, + "Ios16.gelu" : 2, + "Ios17.conv" : 9, + "Ios17.reshape" : 22, + "Split" : 2, + "Ios17.squeeze" : 6, + "Ios17.cast" : 6, + "Ios17.sin" : 2, + "Ios17.linear" : 8, + "Select" : 6, + "Ios16.softmax" : 2, + "Ios17.add" : 28, + "Ios17.cos" : 2, + "Ios17.logicalAnd" : 6 + }, + "computePrecision" : "Mixed (Float32, Int32)", + "isUpdatable" : "0", + "stateSchema" : [ + + ], + "availability" : { + "macOS" : "14.0", + "tvOS" : "17.0", + "visionOS" : "1.0", + "watchOS" : "10.0", + "iOS" : "17.0", + "macCatalyst" : "17.0" + }, + "modelType" : { + "name" : "MLModelType_mlProgram" + }, + "userDefinedMetadata" : { + "com.github.apple.coremltools.conversion_date" : "2026-04-24", + "com.github.apple.coremltools.source" : "torch==2.9.1", + "com.github.apple.coremltools.version" : "9.0", + "com.github.apple.coremltools.source_dialect" : "TorchScript" + }, + "inputSchema" : [ + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1 × 32)", + "shortDescription" : "", + "shape" : "[1, 32]", + "name" : "latent", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1 × 512 × 16)", + "shortDescription" : "", + "shape" : "[1, 512, 16]", + "name" : "upsample_partial", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 2 × 1 × 256 × 8 × 64)", + "shortDescription" : "", + "shape" : "[2, 1, 256, 8, 64]", + "name" : "attn0_cache", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1)", + "shortDescription" : "", + "shape" : "[1]", + "name" : "attn0_offset", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 2 × 1 × 256 × 8 × 64)", + "shortDescription" : "", + "shape" : "[2, 1, 256, 8, 64]", + "name" : "attn1_cache", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1)", + "shortDescription" : "", + "shape" : "[1]", + "name" : "attn1_offset", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1 × 512 × 6)", + "shortDescription" : "", + "shape" : "[1, 512, 6]", + "name" : "conv0_prev", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1)", + "shortDescription" : "", + "shape" : "[1]", + "name" : "conv0_first", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1 × 256 × 6)", + "shortDescription" : "", + "shape" : "[1, 256, 6]", + "name" : "convtr0_partial", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1 × 256 × 2)", + "shortDescription" : "", + "shape" : "[1, 256, 2]", + "name" : "res0_conv0_prev", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1)", + "shortDescription" : "", + "shape" : "[1]", + "name" : "res0_conv0_first", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1 × 128 × 0)", + "shortDescription" : "", + "shape" : "[1, 128, 0]", + "name" : "res0_conv1_prev", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1)", + "shortDescription" : "", + "shape" : "[1]", + "name" : "res0_conv1_first", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1 × 128 × 5)", + "shortDescription" : "", + "shape" : "[1, 128, 5]", + "name" : "convtr1_partial", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1 × 128 × 2)", + "shortDescription" : "", + "shape" : "[1, 128, 2]", + "name" : "res1_conv0_prev", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1)", + "shortDescription" : "", + "shape" : "[1]", + "name" : "res1_conv0_first", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1 × 64 × 0)", + "shortDescription" : "", + "shape" : "[1, 64, 0]", + "name" : "res1_conv1_prev", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1)", + "shortDescription" : "", + "shape" : "[1]", + "name" : "res1_conv1_first", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1 × 64 × 4)", + "shortDescription" : "", + "shape" : "[1, 64, 4]", + "name" : "convtr2_partial", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1 × 64 × 2)", + "shortDescription" : "", + "shape" : "[1, 64, 2]", + "name" : "res2_conv0_prev", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1)", + "shortDescription" : "", + "shape" : "[1]", + "name" : "res2_conv0_first", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1 × 32 × 0)", + "shortDescription" : "", + "shape" : "[1, 32, 0]", + "name" : "res2_conv1_prev", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1)", + "shortDescription" : "", + "shape" : "[1]", + "name" : "res2_conv1_first", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1 × 64 × 2)", + "shortDescription" : "", + "shape" : "[1, 64, 2]", + "name" : "conv_final_prev", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1)", + "shortDescription" : "", + "shape" : "[1]", + "name" : "conv_final_first", + "type" : "MultiArray" + } + ], + "generatedClassName" : "mimi_decoder", + "method" : "predict" + } +] \ No newline at end of file diff --git a/v2.1/italian_24l/mimi_decoder.mlmodelc/model.mil b/v2.1/italian_24l/mimi_decoder.mlmodelc/model.mil new file mode 100644 index 0000000000000000000000000000000000000000..1757e7713b3aa3b787d85968c60ed3fbe4cedb08 --- /dev/null +++ b/v2.1/italian_24l/mimi_decoder.mlmodelc/model.mil @@ -0,0 +1,646 @@ +program(1.0) +[buildInfo = dict, tensor>({{"coremlc-component-MIL", "3520.4.1"}, {"coremlc-version", "3520.5.1"}, {"coremltools-component-torch", "2.9.1"}, {"coremltools-source-dialect", "TorchScript"}, {"coremltools-version", "9.0"}})] +{ + func main(tensor attn0_cache, tensor attn0_offset, tensor attn1_cache, tensor attn1_offset, tensor conv0_first, tensor conv0_prev, tensor conv_final_first, tensor conv_final_prev, tensor convtr0_partial, tensor convtr1_partial, tensor convtr2_partial, tensor latent, tensor res0_conv0_first, tensor res0_conv0_prev, tensor res0_conv1_first, tensor res0_conv1_prev, tensor res1_conv0_first, tensor res1_conv0_prev, tensor res1_conv1_first, tensor res1_conv1_prev, tensor res2_conv0_first, tensor res2_conv0_prev, tensor res2_conv1_first, tensor res2_conv1_prev, tensor upsample_partial) { + tensor emb_mean = const()[name = tensor("emb_mean"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(64)))]; + tensor emb_std = const()[name = tensor("emb_std"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(256)))]; + tensor mimi_quantizer_output_proj_weight = const()[name = tensor("mimi_quantizer_output_proj_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(448)))]; + tensor mimi_upsample_convtr_convtr_weight = const()[name = tensor("mimi_upsample_convtr_convtr_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(66048)))]; + tensor mimi_decoder_transformer_transformer_layers_0_norm1_bias = const()[name = tensor("mimi_decoder_transformer_transformer_layers_0_norm1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(131648)))]; + tensor mimi_decoder_transformer_transformer_layers_0_norm1_weight = const()[name = tensor("mimi_decoder_transformer_transformer_layers_0_norm1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(133760)))]; + tensor mimi_decoder_transformer_transformer_layers_0_self_attn_in_proj_weight = const()[name = tensor("mimi_decoder_transformer_transformer_layers_0_self_attn_in_proj_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(135872)))]; + tensor mimi_decoder_transformer_transformer_layers_0_self_attn_out_proj_weight = const()[name = tensor("mimi_decoder_transformer_transformer_layers_0_self_attn_out_proj_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(3281664)))]; + tensor mimi_decoder_transformer_transformer_layers_0_layer_scale_1_scale = const()[name = tensor("mimi_decoder_transformer_transformer_layers_0_layer_scale_1_scale"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(4330304)))]; + tensor mimi_decoder_transformer_transformer_layers_0_norm2_bias = const()[name = tensor("mimi_decoder_transformer_transformer_layers_0_norm2_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(4332416)))]; + tensor mimi_decoder_transformer_transformer_layers_0_norm2_weight = const()[name = tensor("mimi_decoder_transformer_transformer_layers_0_norm2_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(4334528)))]; + tensor mimi_decoder_transformer_transformer_layers_0_linear1_weight = const()[name = tensor("mimi_decoder_transformer_transformer_layers_0_linear1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(4336640)))]; + tensor mimi_decoder_transformer_transformer_layers_0_linear2_weight = const()[name = tensor("mimi_decoder_transformer_transformer_layers_0_linear2_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8531008)))]; + tensor mimi_decoder_transformer_transformer_layers_0_layer_scale_2_scale = const()[name = tensor("mimi_decoder_transformer_transformer_layers_0_layer_scale_2_scale"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(12725376)))]; + tensor mimi_decoder_transformer_transformer_layers_1_norm1_bias = const()[name = tensor("mimi_decoder_transformer_transformer_layers_1_norm1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(12727488)))]; + tensor mimi_decoder_transformer_transformer_layers_1_norm1_weight = const()[name = tensor("mimi_decoder_transformer_transformer_layers_1_norm1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(12729600)))]; + tensor mimi_decoder_transformer_transformer_layers_1_self_attn_in_proj_weight = const()[name = tensor("mimi_decoder_transformer_transformer_layers_1_self_attn_in_proj_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(12731712)))]; + tensor mimi_decoder_transformer_transformer_layers_1_self_attn_out_proj_weight = const()[name = tensor("mimi_decoder_transformer_transformer_layers_1_self_attn_out_proj_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15877504)))]; + tensor mimi_decoder_transformer_transformer_layers_1_layer_scale_1_scale = const()[name = tensor("mimi_decoder_transformer_transformer_layers_1_layer_scale_1_scale"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(16926144)))]; + tensor mimi_decoder_transformer_transformer_layers_1_norm2_bias = const()[name = tensor("mimi_decoder_transformer_transformer_layers_1_norm2_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(16928256)))]; + tensor mimi_decoder_transformer_transformer_layers_1_norm2_weight = const()[name = tensor("mimi_decoder_transformer_transformer_layers_1_norm2_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(16930368)))]; + tensor mimi_decoder_transformer_transformer_layers_1_linear1_weight = const()[name = tensor("mimi_decoder_transformer_transformer_layers_1_linear1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(16932480)))]; + tensor mimi_decoder_transformer_transformer_layers_1_linear2_weight = const()[name = tensor("mimi_decoder_transformer_transformer_layers_1_linear2_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21126848)))]; + tensor mimi_decoder_transformer_transformer_layers_1_layer_scale_2_scale = const()[name = tensor("mimi_decoder_transformer_transformer_layers_1_layer_scale_2_scale"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(25321216)))]; + tensor mimi_decoder_model_0_conv_bias = const()[name = tensor("mimi_decoder_model_0_conv_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(25323328)))]; + tensor mimi_decoder_model_0_conv_weight = const()[name = tensor("mimi_decoder_model_0_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(25325440)))]; + tensor mimi_decoder_model_2_convtr_bias = const()[name = tensor("mimi_decoder_model_2_convtr_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(32665536)))]; + tensor mimi_decoder_model_2_convtr_weight = const()[name = tensor("mimi_decoder_model_2_convtr_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(32666624)))]; + tensor mimi_decoder_model_3_block_1_conv_bias = const()[name = tensor("mimi_decoder_model_3_block_1_conv_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(38958144)))]; + tensor mimi_decoder_model_3_block_1_conv_weight = const()[name = tensor("mimi_decoder_model_3_block_1_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(38958720)))]; + tensor mimi_decoder_model_3_block_3_conv_bias = const()[name = tensor("mimi_decoder_model_3_block_3_conv_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(39352000)))]; + tensor mimi_decoder_model_3_block_3_conv_weight = const()[name = tensor("mimi_decoder_model_3_block_3_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(39353088)))]; + tensor mimi_decoder_model_5_convtr_bias = const()[name = tensor("mimi_decoder_model_5_convtr_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(39484224)))]; + tensor mimi_decoder_model_5_convtr_weight = const()[name = tensor("mimi_decoder_model_5_convtr_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(39484800)))]; + tensor mimi_decoder_model_6_block_1_conv_bias = const()[name = tensor("mimi_decoder_model_6_block_1_conv_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(40795584)))]; + tensor mimi_decoder_model_6_block_1_conv_weight = const()[name = tensor("mimi_decoder_model_6_block_1_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(40795904)))]; + tensor mimi_decoder_model_6_block_3_conv_bias = const()[name = tensor("mimi_decoder_model_6_block_3_conv_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(40894272)))]; + tensor mimi_decoder_model_6_block_3_conv_weight = const()[name = tensor("mimi_decoder_model_6_block_3_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(40894848)))]; + tensor mimi_decoder_model_8_convtr_bias = const()[name = tensor("mimi_decoder_model_8_convtr_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(40927680)))]; + tensor mimi_decoder_model_8_convtr_weight = const()[name = tensor("mimi_decoder_model_8_convtr_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(40928000)))]; + tensor mimi_decoder_model_9_block_1_conv_bias = const()[name = tensor("mimi_decoder_model_9_block_1_conv_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(41190208)))]; + tensor mimi_decoder_model_9_block_1_conv_weight = const()[name = tensor("mimi_decoder_model_9_block_1_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(41190400)))]; + tensor mimi_decoder_model_9_block_3_conv_bias = const()[name = tensor("mimi_decoder_model_9_block_3_conv_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(41215040)))]; + tensor mimi_decoder_model_9_block_3_conv_weight = const()[name = tensor("mimi_decoder_model_9_block_3_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(41215360)))]; + tensor mimi_decoder_model_11_conv_bias = const()[name = tensor("mimi_decoder_model_11_conv_bias"), val = tensor([0x1.46p-11])]; + tensor mimi_decoder_model_11_conv_weight = const()[name = tensor("mimi_decoder_model_11_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(41223616)))]; + tensor var_38 = mul(x = latent, y = emb_std)[name = tensor("op_38")]; + tensor denorm = add(x = var_38, y = emb_mean)[name = tensor("denorm")]; + tensor input_1_axes_0 = const()[name = tensor("input_1_axes_0"), val = tensor([-1])]; + tensor input_1 = expand_dims(axes = input_1_axes_0, x = denorm)[name = tensor("input_1")]; + tensor x_1_pad_type_0 = const()[name = tensor("x_1_pad_type_0"), val = tensor("valid")]; + tensor x_1_strides_0 = const()[name = tensor("x_1_strides_0"), val = tensor([1])]; + tensor x_1_pad_0 = const()[name = tensor("x_1_pad_0"), val = tensor([0, 0])]; + tensor x_1_dilations_0 = const()[name = tensor("x_1_dilations_0"), val = tensor([1])]; + tensor x_1_groups_0 = const()[name = tensor("x_1_groups_0"), val = tensor(1)]; + tensor x_1 = conv(dilations = x_1_dilations_0, groups = x_1_groups_0, pad = x_1_pad_0, pad_type = x_1_pad_type_0, strides = x_1_strides_0, weight = mimi_quantizer_output_proj_weight, x = input_1)[name = tensor("x_1")]; + tensor var_62 = const()[name = tensor("op_62"), val = tensor(-1)]; + tensor y_1_pad_type_0 = const()[name = tensor("y_1_pad_type_0"), val = tensor("valid")]; + tensor y_1_strides_0 = const()[name = tensor("y_1_strides_0"), val = tensor([16])]; + tensor y_1_groups_0 = const()[name = tensor("y_1_groups_0"), val = tensor(512)]; + tensor y_1_pad_0 = const()[name = tensor("y_1_pad_0"), val = tensor([0, 0])]; + tensor y_1_dilations_0 = const()[name = tensor("y_1_dilations_0"), val = tensor([1])]; + tensor y_1_has_output_shape_output_shape_0 = const()[name = tensor("y_1_has_output_shape_output_shape_0"), val = tensor([1, 512, 32])]; + tensor y_1_has_output_shape = conv_transpose(dilations = y_1_dilations_0, groups = y_1_groups_0, output_shape = y_1_has_output_shape_output_shape_0, pad = y_1_pad_0, pad_type = y_1_pad_type_0, strides = y_1_strides_0, weight = mimi_upsample_convtr_convtr_weight, x = x_1)[name = tensor("y_1_has_output_shape")]; + tensor var_72_begin_0 = const()[name = tensor("op_72_begin_0"), val = tensor([0, 0, 0])]; + tensor var_72_end_0 = const()[name = tensor("op_72_end_0"), val = tensor([1, 512, 16])]; + tensor var_72_end_mask_0 = const()[name = tensor("op_72_end_mask_0"), val = tensor([true, true, false])]; + tensor var_72 = slice_by_index(begin = var_72_begin_0, end = var_72_end_0, end_mask = var_72_end_mask_0, x = y_1_has_output_shape)[name = tensor("op_72")]; + tensor var_73 = add(x = var_72, y = upsample_partial)[name = tensor("op_73")]; + tensor var_74_begin_0 = const()[name = tensor("op_74_begin_0"), val = tensor([0, 0, 16])]; + tensor var_74_end_0 = const()[name = tensor("op_74_end_0"), val = tensor([1, 512, 32])]; + tensor var_74_end_mask_0 = const()[name = tensor("op_74_end_mask_0"), val = tensor([true, true, true])]; + tensor var_74 = slice_by_index(begin = var_74_begin_0, end = var_74_end_0, end_mask = var_74_end_mask_0, x = y_1_has_output_shape)[name = tensor("op_74")]; + tensor y_3_interleave_0 = const()[name = tensor("y_3_interleave_0"), val = tensor(false)]; + tensor y_3 = concat(axis = var_62, interleave = y_3_interleave_0, values = (var_73, var_74))[name = tensor("y_3")]; + tensor var_77_begin_0 = const()[name = tensor("op_77_begin_0"), val = tensor([0, 0, 16])]; + tensor var_77_end_0 = const()[name = tensor("op_77_end_0"), val = tensor([1, 512, 32])]; + tensor var_77_end_mask_0 = const()[name = tensor("op_77_end_mask_0"), val = tensor([true, true, true])]; + tensor var_77 = slice_by_index(begin = var_77_begin_0, end = var_77_end_0, end_mask = var_77_end_mask_0, x = y_3)[name = tensor("op_77")]; + tensor x_3_begin_0 = const()[name = tensor("x_3_begin_0"), val = tensor([0, 0, 0])]; + tensor x_3_end_0 = const()[name = tensor("x_3_end_0"), val = tensor([1, 512, 16])]; + tensor x_3_end_mask_0 = const()[name = tensor("x_3_end_mask_0"), val = tensor([true, true, false])]; + tensor x_3 = slice_by_index(begin = x_3_begin_0, end = x_3_end_0, end_mask = x_3_end_mask_0, x = y_3)[name = tensor("x_3")]; + tensor var_86 = const()[name = tensor("op_86"), val = tensor(0)]; + tensor var_91 = const()[name = tensor("op_91"), val = tensor(-1)]; + tensor var_100 = const()[name = tensor("op_100"), val = tensor(-0x1.ff933cp+127)]; + tensor var_102 = const()[name = tensor("op_102"), val = tensor(0x1.4f8b58p-17)]; + tensor input_3_perm_0 = const()[name = tensor("input_3_perm_0"), val = tensor([0, 2, 1])]; + tensor query_1_axes_0 = const()[name = tensor("query_1_axes_0"), val = tensor([-1])]; + tensor input_3 = transpose(perm = input_3_perm_0, x = x_3)[name = tensor("transpose_19")]; + tensor query_1 = layer_norm(axes = query_1_axes_0, beta = mimi_decoder_transformer_transformer_layers_0_norm1_bias, epsilon = var_102, gamma = mimi_decoder_transformer_transformer_layers_0_norm1_weight, x = input_3)[name = tensor("query_1")]; + tensor linear_0_bias_0 = const()[name = tensor("linear_0_bias_0"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(41224448)))]; + tensor projected_1 = linear(bias = linear_0_bias_0, weight = mimi_decoder_transformer_transformer_layers_0_self_attn_in_proj_weight, x = query_1)[name = tensor("linear_0")]; + tensor var_130 = const()[name = tensor("op_130"), val = tensor([1, 16, 3, 8, 64])]; + tensor packed_1 = reshape(shape = var_130, x = projected_1)[name = tensor("packed_1")]; + tensor var_132_split_sizes_0 = const()[name = tensor("op_132_split_sizes_0"), val = tensor([1, 1, 1])]; + tensor var_132_axis_0 = const()[name = tensor("op_132_axis_0"), val = tensor(2)]; + tensor var_132_0, tensor var_132_1, tensor var_132_2 = split(axis = var_132_axis_0, split_sizes = var_132_split_sizes_0, x = packed_1)[name = tensor("op_132")]; + tensor squeeze_0_axes_0 = const()[name = tensor("squeeze_0_axes_0"), val = tensor([2])]; + tensor squeeze_0 = squeeze(axes = squeeze_0_axes_0, x = var_132_0)[name = tensor("squeeze_0")]; + tensor squeeze_1_axes_0 = const()[name = tensor("squeeze_1_axes_0"), val = tensor([2])]; + tensor squeeze_1 = squeeze(axes = squeeze_1_axes_0, x = var_132_1)[name = tensor("squeeze_1")]; + tensor squeeze_2_axes_0 = const()[name = tensor("squeeze_2_axes_0"), val = tensor([2])]; + tensor squeeze_2 = squeeze(axes = squeeze_2_axes_0, x = var_132_2)[name = tensor("squeeze_2")]; + tensor offset_3_begin_0 = const()[name = tensor("offset_3_begin_0"), val = tensor([0])]; + tensor offset_3_end_0 = const()[name = tensor("offset_3_end_0"), val = tensor([1])]; + tensor offset_3_end_mask_0 = const()[name = tensor("offset_3_end_mask_0"), val = tensor([false])]; + tensor offset_3_squeeze_mask_0 = const()[name = tensor("offset_3_squeeze_mask_0"), val = tensor([true])]; + tensor offset_3 = slice_by_index(begin = offset_3_begin_0, end = offset_3_end_0, end_mask = offset_3_end_mask_0, squeeze_mask = offset_3_squeeze_mask_0, x = attn0_offset)[name = tensor("offset_3")]; + tensor freqs_1 = const()[name = tensor("freqs_1"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(41230656)))]; + tensor ts_1_promoted = const()[name = tensor("ts_1_promoted"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(41230848)))]; + tensor ts_3 = add(x = ts_1_promoted, y = offset_3)[name = tensor("ts_3")]; + tensor var_148 = const()[name = tensor("op_148"), val = tensor([-1, 1, 1])]; + tensor ts_5 = reshape(shape = var_148, x = ts_3)[name = tensor("ts_5")]; + tensor var_150 = const()[name = tensor("op_150"), val = tensor([1, 16, 8, 32, 2])]; + tensor q_3 = reshape(shape = var_150, x = squeeze_0)[name = tensor("q_3")]; + tensor var_152 = const()[name = tensor("op_152"), val = tensor([1, 16, 8, 32, 2])]; + tensor k_3 = reshape(shape = var_152, x = squeeze_1)[name = tensor("k_3")]; + tensor var_154_begin_0 = const()[name = tensor("op_154_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_154_end_0 = const()[name = tensor("op_154_end_0"), val = tensor([1, 16, 8, 32, 1])]; + tensor var_154_end_mask_0 = const()[name = tensor("op_154_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_154_squeeze_mask_0 = const()[name = tensor("op_154_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_154 = slice_by_index(begin = var_154_begin_0, end = var_154_end_0, end_mask = var_154_end_mask_0, squeeze_mask = var_154_squeeze_mask_0, x = q_3)[name = tensor("op_154")]; + tensor var_156_begin_0 = const()[name = tensor("op_156_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_156_end_0 = const()[name = tensor("op_156_end_0"), val = tensor([1, 16, 8, 32, 2])]; + tensor var_156_end_mask_0 = const()[name = tensor("op_156_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_156_squeeze_mask_0 = const()[name = tensor("op_156_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_156 = slice_by_index(begin = var_156_begin_0, end = var_156_end_0, end_mask = var_156_end_mask_0, squeeze_mask = var_156_squeeze_mask_0, x = q_3)[name = tensor("op_156")]; + tensor var_158_begin_0 = const()[name = tensor("op_158_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_158_end_0 = const()[name = tensor("op_158_end_0"), val = tensor([1, 16, 8, 32, 1])]; + tensor var_158_end_mask_0 = const()[name = tensor("op_158_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_158_squeeze_mask_0 = const()[name = tensor("op_158_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_158 = slice_by_index(begin = var_158_begin_0, end = var_158_end_0, end_mask = var_158_end_mask_0, squeeze_mask = var_158_squeeze_mask_0, x = k_3)[name = tensor("op_158")]; + tensor var_160_begin_0 = const()[name = tensor("op_160_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_160_end_0 = const()[name = tensor("op_160_end_0"), val = tensor([1, 16, 8, 32, 2])]; + tensor var_160_end_mask_0 = const()[name = tensor("op_160_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_160_squeeze_mask_0 = const()[name = tensor("op_160_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_160 = slice_by_index(begin = var_160_begin_0, end = var_160_end_0, end_mask = var_160_end_mask_0, squeeze_mask = var_160_squeeze_mask_0, x = k_3)[name = tensor("op_160")]; + tensor var_162 = mul(x = freqs_1, y = ts_5)[name = tensor("op_162")]; + tensor rotr_1 = cos(x = var_162)[name = tensor("rotr_1")]; + tensor roti_1 = sin(x = var_162)[name = tensor("roti_1")]; + tensor var_166 = mul(x = var_154, y = rotr_1)[name = tensor("op_166")]; + tensor var_167 = mul(x = var_156, y = roti_1)[name = tensor("op_167")]; + tensor qor_1 = sub(x = var_166, y = var_167)[name = tensor("qor_1")]; + tensor var_169 = mul(x = var_154, y = roti_1)[name = tensor("op_169")]; + tensor var_170 = mul(x = var_156, y = rotr_1)[name = tensor("op_170")]; + tensor qoi_1 = add(x = var_169, y = var_170)[name = tensor("qoi_1")]; + tensor var_172 = mul(x = var_158, y = rotr_1)[name = tensor("op_172")]; + tensor var_173 = mul(x = var_160, y = roti_1)[name = tensor("op_173")]; + tensor kor_1 = sub(x = var_172, y = var_173)[name = tensor("kor_1")]; + tensor var_175 = mul(x = var_158, y = roti_1)[name = tensor("op_175")]; + tensor var_176 = mul(x = var_160, y = rotr_1)[name = tensor("op_176")]; + tensor koi_1 = add(x = var_175, y = var_176)[name = tensor("koi_1")]; + tensor qo_1_axis_0 = const()[name = tensor("qo_1_axis_0"), val = tensor(-1)]; + tensor qo_1 = stack(axis = qo_1_axis_0, values = (qor_1, qoi_1))[name = tensor("qo_1")]; + tensor ko_1_axis_0 = const()[name = tensor("ko_1_axis_0"), val = tensor(-1)]; + tensor ko_1 = stack(axis = ko_1_axis_0, values = (kor_1, koi_1))[name = tensor("ko_1")]; + tensor var_186 = const()[name = tensor("op_186"), val = tensor([1, 16, 8, 64])]; + tensor q_5 = reshape(shape = var_186, x = qo_1)[name = tensor("q_5")]; + tensor var_188 = const()[name = tensor("op_188"), val = tensor([1, 16, 8, 64])]; + tensor k_5 = reshape(shape = var_188, x = ko_1)[name = tensor("k_5")]; + tensor capacity_1 = const()[name = tensor("capacity_1"), val = tensor([256])]; + tensor var_193_dtype_0 = const()[name = tensor("op_193_dtype_0"), val = tensor("int32")]; + tensor var_194 = const()[name = tensor("op_194"), val = tensor([1, 1])]; + tensor var_193 = cast(dtype = var_193_dtype_0, x = attn0_offset)[name = tensor("cast_49")]; + tensor write_base_1 = reshape(shape = var_194, x = var_193)[name = tensor("write_base_1")]; + tensor write_range_1 = const()[name = tensor("write_range_1"), val = tensor([[0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15]])]; + tensor abs_idx_1 = add(x = write_base_1, y = write_range_1)[name = tensor("abs_idx_1")]; + tensor wrapped_1_div = floor_div(x = abs_idx_1, y = capacity_1)[name = tensor("wrapped_1_div")]; + tensor wrapped_1_div_scaled = mul(x = wrapped_1_div, y = capacity_1)[name = tensor("wrapped_1_div_scaled")]; + tensor wrapped_1 = sub(x = abs_idx_1, y = wrapped_1_div_scaled)[name = tensor("wrapped_1")]; + tensor var_201 = const()[name = tensor("op_201"), val = tensor([1, 16, 1, 1])]; + tensor var_202 = reshape(shape = var_201, x = wrapped_1)[name = tensor("op_202")]; + tensor write_indexes_1_reps_0 = const()[name = tensor("write_indexes_1_reps_0"), val = tensor([1, 1, 8, 64])]; + tensor write_indexes_1 = tile(reps = write_indexes_1_reps_0, x = var_202)[name = tensor("write_indexes_1")]; + tensor var_205_begin_0 = const()[name = tensor("op_205_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_205_end_0 = const()[name = tensor("op_205_end_0"), val = tensor([1, 1, 256, 8, 64])]; + tensor var_205_end_mask_0 = const()[name = tensor("op_205_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_205_squeeze_mask_0 = const()[name = tensor("op_205_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor var_205 = slice_by_index(begin = var_205_begin_0, end = var_205_end_0, end_mask = var_205_end_mask_0, squeeze_mask = var_205_squeeze_mask_0, x = attn0_cache)[name = tensor("op_205")]; + tensor new_k_cache_1_axis_0 = const()[name = tensor("new_k_cache_1_axis_0"), val = tensor(1)]; + tensor new_k_cache_1_mode_0 = const()[name = tensor("new_k_cache_1_mode_0"), val = tensor("update")]; + tensor new_k_cache_1_validate_indices_0 = const()[name = tensor("new_k_cache_1_validate_indices_0"), val = tensor(false)]; + tensor new_k_cache_1 = scatter_along_axis(axis = new_k_cache_1_axis_0, data = var_205, indices = write_indexes_1, mode = new_k_cache_1_mode_0, updates = k_5, validate_indices = new_k_cache_1_validate_indices_0)[name = tensor("new_k_cache_1")]; + tensor var_207_begin_0 = const()[name = tensor("op_207_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor var_207_end_0 = const()[name = tensor("op_207_end_0"), val = tensor([2, 1, 256, 8, 64])]; + tensor var_207_end_mask_0 = const()[name = tensor("op_207_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_207_squeeze_mask_0 = const()[name = tensor("op_207_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor var_207 = slice_by_index(begin = var_207_begin_0, end = var_207_end_0, end_mask = var_207_end_mask_0, squeeze_mask = var_207_squeeze_mask_0, x = attn0_cache)[name = tensor("op_207")]; + tensor new_v_cache_1_axis_0 = const()[name = tensor("new_v_cache_1_axis_0"), val = tensor(1)]; + tensor new_v_cache_1_mode_0 = const()[name = tensor("new_v_cache_1_mode_0"), val = tensor("update")]; + tensor new_v_cache_1_validate_indices_0 = const()[name = tensor("new_v_cache_1_validate_indices_0"), val = tensor(false)]; + tensor new_v_cache_1 = scatter_along_axis(axis = new_v_cache_1_axis_0, data = var_207, indices = write_indexes_1, mode = new_v_cache_1_mode_0, updates = squeeze_2, validate_indices = new_v_cache_1_validate_indices_0)[name = tensor("new_v_cache_1")]; + tensor var_210_axis_0 = const()[name = tensor("op_210_axis_0"), val = tensor(0)]; + tensor var_210 = stack(axis = var_210_axis_0, values = (new_k_cache_1, new_v_cache_1))[name = tensor("op_210")]; + tensor var_211 = not_equal(x = new_k_cache_1, y = new_k_cache_1)[name = tensor("op_211")]; + tensor var_212 = const()[name = tensor("op_212"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(41230976)))]; + tensor new_k_cache_3 = select(a = var_212, b = new_k_cache_1, cond = var_211)[name = tensor("new_k_cache_3")]; + tensor var_214 = not_equal(x = new_v_cache_1, y = new_v_cache_1)[name = tensor("op_214")]; + tensor new_v_cache_3 = select(a = var_212, b = new_v_cache_1, cond = var_214)[name = tensor("new_v_cache_3")]; + tensor var_219 = const()[name = tensor("op_219"), val = tensor([0, 2, 1, 3])]; + tensor var_221 = const()[name = tensor("op_221"), val = tensor([1, 1])]; + tensor var_222 = reshape(shape = var_221, x = attn0_offset)[name = tensor("op_222")]; + tensor var_224_promoted = const()[name = tensor("op_224_promoted"), val = tensor([0x1.ep+3])]; + tensor var_225 = add(x = var_222, y = var_224_promoted)[name = tensor("op_225")]; + tensor last_pos_1_dtype_0 = const()[name = tensor("last_pos_1_dtype_0"), val = tensor("int32")]; + tensor slot_idx_1 = const()[name = tensor("slot_idx_1"), val = tensor([[0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19, 20, 21, 22, 23, 24, 25, 26, 27, 28, 29, 30, 31, 32, 33, 34, 35, 36, 37, 38, 39, 40, 41, 42, 43, 44, 45, 46, 47, 48, 49, 50, 51, 52, 53, 54, 55, 56, 57, 58, 59, 60, 61, 62, 63, 64, 65, 66, 67, 68, 69, 70, 71, 72, 73, 74, 75, 76, 77, 78, 79, 80, 81, 82, 83, 84, 85, 86, 87, 88, 89, 90, 91, 92, 93, 94, 95, 96, 97, 98, 99, 100, 101, 102, 103, 104, 105, 106, 107, 108, 109, 110, 111, 112, 113, 114, 115, 116, 117, 118, 119, 120, 121, 122, 123, 124, 125, 126, 127, 128, 129, 130, 131, 132, 133, 134, 135, 136, 137, 138, 139, 140, 141, 142, 143, 144, 145, 146, 147, 148, 149, 150, 151, 152, 153, 154, 155, 156, 157, 158, 159, 160, 161, 162, 163, 164, 165, 166, 167, 168, 169, 170, 171, 172, 173, 174, 175, 176, 177, 178, 179, 180, 181, 182, 183, 184, 185, 186, 187, 188, 189, 190, 191, 192, 193, 194, 195, 196, 197, 198, 199, 200, 201, 202, 203, 204, 205, 206, 207, 208, 209, 210, 211, 212, 213, 214, 215, 216, 217, 218, 219, 220, 221, 222, 223, 224, 225, 226, 227, 228, 229, 230, 231, 232, 233, 234, 235, 236, 237, 238, 239, 240, 241, 242, 243, 244, 245, 246, 247, 248, 249, 250, 251, 252, 253, 254, 255]])]; + tensor last_pos_1 = cast(dtype = last_pos_1_dtype_0, x = var_225)[name = tensor("cast_48")]; + tensor diff_1 = sub(x = last_pos_1, y = slot_idx_1)[name = tensor("diff_1")]; + tensor var_231_div = floor_div(x = diff_1, y = capacity_1)[name = tensor("op_231_div")]; + tensor var_231_div_scaled = mul(x = var_231_div, y = capacity_1)[name = tensor("op_231_div_scaled")]; + tensor var_231 = sub(x = diff_1, y = var_231_div_scaled)[name = tensor("op_231")]; + tensor pos_k_1 = sub(x = last_pos_1, y = var_231)[name = tensor("pos_k_1")]; + tensor var_237_promoted = const()[name = tensor("op_237_promoted"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(41755328)))]; + tensor pos_q_1 = add(x = var_222, y = var_237_promoted)[name = tensor("pos_q_1")]; + tensor var_241_axes_0 = const()[name = tensor("op_241_axes_0"), val = tensor([2])]; + tensor var_241 = expand_dims(axes = var_241_axes_0, x = pos_q_1)[name = tensor("op_241")]; + tensor var_243_axes_0 = const()[name = tensor("op_243_axes_0"), val = tensor([1])]; + tensor var_243 = expand_dims(axes = var_243_axes_0, x = pos_k_1)[name = tensor("op_243")]; + tensor var_244_promoted_dtype_0 = const()[name = tensor("op_244_promoted_dtype_0"), val = tensor("fp32")]; + tensor var_244_promoted = cast(dtype = var_244_promoted_dtype_0, x = var_243)[name = tensor("cast_47")]; + tensor delta_1 = sub(x = var_241, y = var_244_promoted)[name = tensor("delta_1")]; + tensor valid_1 = greater_equal(x = var_243, y = var_86)[name = tensor("valid_1")]; + tensor var_253 = const()[name = tensor("op_253"), val = tensor([1, 1, 1])]; + tensor var_254 = reshape(shape = var_253, x = attn0_offset)[name = tensor("op_254")]; + tensor var_256_promoted = const()[name = tensor("op_256_promoted"), val = tensor([0x1.ep+3])]; + tensor var_257 = add(x = var_254, y = var_256_promoted)[name = tensor("op_257")]; + tensor var_258 = less_equal(x = var_244_promoted, y = var_257)[name = tensor("op_258")]; + tensor valid_3 = logical_and(x = valid_1, y = var_258)[name = tensor("valid_3")]; + tensor var_86_promoted = const()[name = tensor("op_86_promoted"), val = tensor(0x0p+0)]; + tensor var_260 = greater_equal(x = delta_1, y = var_86_promoted)[name = tensor("op_260")]; + tensor attn_mask_1 = logical_and(x = valid_3, y = var_260)[name = tensor("attn_mask_1")]; + tensor var_98_promoted = const()[name = tensor("op_98_promoted"), val = tensor(0x1.f4p+7)]; + tensor var_262 = less(x = delta_1, y = var_98_promoted)[name = tensor("op_262")]; + tensor attn_mask_3 = logical_and(x = attn_mask_1, y = var_262)[name = tensor("attn_mask_3")]; + tensor attn_mask_5_axes_0 = const()[name = tensor("attn_mask_5_axes_0"), val = tensor([1])]; + tensor attn_mask_5 = expand_dims(axes = attn_mask_5_axes_0, x = attn_mask_3)[name = tensor("attn_mask_5")]; + tensor var_267_transpose_x_0 = const()[name = tensor("op_267_transpose_x_0"), val = tensor(false)]; + tensor var_267_transpose_y_0 = const()[name = tensor("op_267_transpose_y_0"), val = tensor(false)]; + tensor transpose_6_perm_0 = const()[name = tensor("transpose_6_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_7_perm_0 = const()[name = tensor("transpose_7_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_7 = transpose(perm = transpose_7_perm_0, x = new_k_cache_3)[name = tensor("transpose_16")]; + tensor transpose_6 = transpose(perm = transpose_6_perm_0, x = q_5)[name = tensor("transpose_17")]; + tensor var_267 = matmul(transpose_x = var_267_transpose_x_0, transpose_y = var_267_transpose_y_0, x = transpose_6, y = transpose_7)[name = tensor("op_267")]; + tensor var_268 = const()[name = tensor("op_268"), val = tensor(0x1p-3)]; + tensor attn_1 = mul(x = var_267, y = var_268)[name = tensor("attn_1")]; + tensor var_270 = logical_not(x = attn_mask_5)[name = tensor("op_270")]; + tensor attn_3 = select(a = var_100, b = attn_1, cond = var_270)[name = tensor("attn_3")]; + tensor attn_5 = softmax(axis = var_91, x = attn_3)[name = tensor("attn_5")]; + tensor x_5_transpose_x_0 = const()[name = tensor("x_5_transpose_x_0"), val = tensor(false)]; + tensor x_5_transpose_y_0 = const()[name = tensor("x_5_transpose_y_0"), val = tensor(false)]; + tensor v_attn_1 = transpose(perm = var_219, x = new_v_cache_3)[name = tensor("transpose_18")]; + tensor x_5 = matmul(transpose_x = x_5_transpose_x_0, transpose_y = x_5_transpose_y_0, x = attn_5, y = v_attn_1)[name = tensor("x_5")]; + tensor var_274_perm_0 = const()[name = tensor("op_274_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_275 = const()[name = tensor("op_275"), val = tensor([1, 16, 512])]; + tensor var_274 = transpose(perm = var_274_perm_0, x = x_5)[name = tensor("transpose_15")]; + tensor input_5 = reshape(shape = var_275, x = var_274)[name = tensor("input_5")]; + tensor linear_1_bias_0 = const()[name = tensor("linear_1_bias_0"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(41755456)))]; + tensor x_7 = linear(bias = linear_1_bias_0, weight = mimi_decoder_transformer_transformer_layers_0_self_attn_out_proj_weight, x = input_5)[name = tensor("linear_1")]; + tensor var_284 = mul(x = mimi_decoder_transformer_transformer_layers_0_layer_scale_1_scale, y = x_7)[name = tensor("op_284")]; + tensor input_7 = add(x = input_3, y = var_284)[name = tensor("input_7")]; + tensor input_9_axes_0 = const()[name = tensor("input_9_axes_0"), val = tensor([-1])]; + tensor input_9 = layer_norm(axes = input_9_axes_0, beta = mimi_decoder_transformer_transformer_layers_0_norm2_bias, epsilon = var_102, gamma = mimi_decoder_transformer_transformer_layers_0_norm2_weight, x = input_7)[name = tensor("input_9")]; + tensor linear_2_bias_0 = const()[name = tensor("linear_2_bias_0"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(41757568)))]; + tensor var_291 = linear(bias = linear_2_bias_0, weight = mimi_decoder_transformer_transformer_layers_0_linear1_weight, x = input_9)[name = tensor("linear_2")]; + tensor input_11_mode_0 = const()[name = tensor("input_11_mode_0"), val = tensor("EXACT")]; + tensor input_11 = gelu(mode = input_11_mode_0, x = var_291)[name = tensor("input_11")]; + tensor x_9 = linear(bias = linear_1_bias_0, weight = mimi_decoder_transformer_transformer_layers_0_linear2_weight, x = input_11)[name = tensor("linear_3")]; + tensor var_297 = mul(x = mimi_decoder_transformer_transformer_layers_0_layer_scale_2_scale, y = x_9)[name = tensor("op_297")]; + tensor input_13 = add(x = input_7, y = var_297)[name = tensor("input_13")]; + tensor query_axes_0 = const()[name = tensor("query_axes_0"), val = tensor([-1])]; + tensor query = layer_norm(axes = query_axes_0, beta = mimi_decoder_transformer_transformer_layers_1_norm1_bias, epsilon = var_102, gamma = mimi_decoder_transformer_transformer_layers_1_norm1_weight, x = input_13)[name = tensor("query")]; + tensor projected = linear(bias = linear_0_bias_0, weight = mimi_decoder_transformer_transformer_layers_1_self_attn_in_proj_weight, x = query)[name = tensor("linear_4")]; + tensor var_320 = const()[name = tensor("op_320"), val = tensor([1, 16, 3, 8, 64])]; + tensor packed = reshape(shape = var_320, x = projected)[name = tensor("packed")]; + tensor var_322_split_sizes_0 = const()[name = tensor("op_322_split_sizes_0"), val = tensor([1, 1, 1])]; + tensor var_322_axis_0 = const()[name = tensor("op_322_axis_0"), val = tensor(2)]; + tensor var_322_0, tensor var_322_1, tensor var_322_2 = split(axis = var_322_axis_0, split_sizes = var_322_split_sizes_0, x = packed)[name = tensor("op_322")]; + tensor squeeze_3_axes_0 = const()[name = tensor("squeeze_3_axes_0"), val = tensor([2])]; + tensor squeeze_3 = squeeze(axes = squeeze_3_axes_0, x = var_322_0)[name = tensor("squeeze_3")]; + tensor squeeze_4_axes_0 = const()[name = tensor("squeeze_4_axes_0"), val = tensor([2])]; + tensor squeeze_4 = squeeze(axes = squeeze_4_axes_0, x = var_322_1)[name = tensor("squeeze_4")]; + tensor squeeze_5_axes_0 = const()[name = tensor("squeeze_5_axes_0"), val = tensor([2])]; + tensor squeeze_5 = squeeze(axes = squeeze_5_axes_0, x = var_322_2)[name = tensor("squeeze_5")]; + tensor offset_begin_0 = const()[name = tensor("offset_begin_0"), val = tensor([0])]; + tensor offset_end_0 = const()[name = tensor("offset_end_0"), val = tensor([1])]; + tensor offset_end_mask_0 = const()[name = tensor("offset_end_mask_0"), val = tensor([false])]; + tensor offset_squeeze_mask_0 = const()[name = tensor("offset_squeeze_mask_0"), val = tensor([true])]; + tensor offset = slice_by_index(begin = offset_begin_0, end = offset_end_0, end_mask = offset_end_mask_0, squeeze_mask = offset_squeeze_mask_0, x = attn1_offset)[name = tensor("offset")]; + tensor freqs = const()[name = tensor("freqs"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(41765824)))]; + tensor ts_7_promoted = const()[name = tensor("ts_7_promoted"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(41766016)))]; + tensor ts_9 = add(x = ts_7_promoted, y = offset)[name = tensor("ts_9")]; + tensor var_338 = const()[name = tensor("op_338"), val = tensor([-1, 1, 1])]; + tensor ts = reshape(shape = var_338, x = ts_9)[name = tensor("ts")]; + tensor var_340 = const()[name = tensor("op_340"), val = tensor([1, 16, 8, 32, 2])]; + tensor q_9 = reshape(shape = var_340, x = squeeze_3)[name = tensor("q_9")]; + tensor var_342 = const()[name = tensor("op_342"), val = tensor([1, 16, 8, 32, 2])]; + tensor k_9 = reshape(shape = var_342, x = squeeze_4)[name = tensor("k_9")]; + tensor var_344_begin_0 = const()[name = tensor("op_344_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_344_end_0 = const()[name = tensor("op_344_end_0"), val = tensor([1, 16, 8, 32, 1])]; + tensor var_344_end_mask_0 = const()[name = tensor("op_344_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_344_squeeze_mask_0 = const()[name = tensor("op_344_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_344 = slice_by_index(begin = var_344_begin_0, end = var_344_end_0, end_mask = var_344_end_mask_0, squeeze_mask = var_344_squeeze_mask_0, x = q_9)[name = tensor("op_344")]; + tensor var_346_begin_0 = const()[name = tensor("op_346_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_346_end_0 = const()[name = tensor("op_346_end_0"), val = tensor([1, 16, 8, 32, 2])]; + tensor var_346_end_mask_0 = const()[name = tensor("op_346_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_346_squeeze_mask_0 = const()[name = tensor("op_346_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_346 = slice_by_index(begin = var_346_begin_0, end = var_346_end_0, end_mask = var_346_end_mask_0, squeeze_mask = var_346_squeeze_mask_0, x = q_9)[name = tensor("op_346")]; + tensor var_348_begin_0 = const()[name = tensor("op_348_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_348_end_0 = const()[name = tensor("op_348_end_0"), val = tensor([1, 16, 8, 32, 1])]; + tensor var_348_end_mask_0 = const()[name = tensor("op_348_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_348_squeeze_mask_0 = const()[name = tensor("op_348_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_348 = slice_by_index(begin = var_348_begin_0, end = var_348_end_0, end_mask = var_348_end_mask_0, squeeze_mask = var_348_squeeze_mask_0, x = k_9)[name = tensor("op_348")]; + tensor var_350_begin_0 = const()[name = tensor("op_350_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_350_end_0 = const()[name = tensor("op_350_end_0"), val = tensor([1, 16, 8, 32, 2])]; + tensor var_350_end_mask_0 = const()[name = tensor("op_350_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_350_squeeze_mask_0 = const()[name = tensor("op_350_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_350 = slice_by_index(begin = var_350_begin_0, end = var_350_end_0, end_mask = var_350_end_mask_0, squeeze_mask = var_350_squeeze_mask_0, x = k_9)[name = tensor("op_350")]; + tensor var_352 = mul(x = freqs, y = ts)[name = tensor("op_352")]; + tensor rotr = cos(x = var_352)[name = tensor("rotr")]; + tensor roti = sin(x = var_352)[name = tensor("roti")]; + tensor var_356 = mul(x = var_344, y = rotr)[name = tensor("op_356")]; + tensor var_357 = mul(x = var_346, y = roti)[name = tensor("op_357")]; + tensor qor_5 = sub(x = var_356, y = var_357)[name = tensor("qor_5")]; + tensor var_359 = mul(x = var_344, y = roti)[name = tensor("op_359")]; + tensor var_360 = mul(x = var_346, y = rotr)[name = tensor("op_360")]; + tensor qoi_5 = add(x = var_359, y = var_360)[name = tensor("qoi_5")]; + tensor var_362 = mul(x = var_348, y = rotr)[name = tensor("op_362")]; + tensor var_363 = mul(x = var_350, y = roti)[name = tensor("op_363")]; + tensor kor_5 = sub(x = var_362, y = var_363)[name = tensor("kor_5")]; + tensor var_365 = mul(x = var_348, y = roti)[name = tensor("op_365")]; + tensor var_366 = mul(x = var_350, y = rotr)[name = tensor("op_366")]; + tensor koi_5 = add(x = var_365, y = var_366)[name = tensor("koi_5")]; + tensor qo_axis_0 = const()[name = tensor("qo_axis_0"), val = tensor(-1)]; + tensor qo = stack(axis = qo_axis_0, values = (qor_5, qoi_5))[name = tensor("qo")]; + tensor ko_axis_0 = const()[name = tensor("ko_axis_0"), val = tensor(-1)]; + tensor ko = stack(axis = ko_axis_0, values = (kor_5, koi_5))[name = tensor("ko")]; + tensor var_376 = const()[name = tensor("op_376"), val = tensor([1, 16, 8, 64])]; + tensor q = reshape(shape = var_376, x = qo)[name = tensor("q")]; + tensor var_378 = const()[name = tensor("op_378"), val = tensor([1, 16, 8, 64])]; + tensor k = reshape(shape = var_378, x = ko)[name = tensor("k")]; + tensor capacity = const()[name = tensor("capacity"), val = tensor([256])]; + tensor var_383_dtype_0 = const()[name = tensor("op_383_dtype_0"), val = tensor("int32")]; + tensor var_384 = const()[name = tensor("op_384"), val = tensor([1, 1])]; + tensor var_383 = cast(dtype = var_383_dtype_0, x = attn1_offset)[name = tensor("cast_46")]; + tensor write_base = reshape(shape = var_384, x = var_383)[name = tensor("write_base")]; + tensor write_range = const()[name = tensor("write_range"), val = tensor([[0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15]])]; + tensor abs_idx = add(x = write_base, y = write_range)[name = tensor("abs_idx")]; + tensor wrapped_div = floor_div(x = abs_idx, y = capacity)[name = tensor("wrapped_div")]; + tensor wrapped_div_scaled = mul(x = wrapped_div, y = capacity)[name = tensor("wrapped_div_scaled")]; + tensor wrapped = sub(x = abs_idx, y = wrapped_div_scaled)[name = tensor("wrapped")]; + tensor var_391 = const()[name = tensor("op_391"), val = tensor([1, 16, 1, 1])]; + tensor var_392 = reshape(shape = var_391, x = wrapped)[name = tensor("op_392")]; + tensor write_indexes_reps_0 = const()[name = tensor("write_indexes_reps_0"), val = tensor([1, 1, 8, 64])]; + tensor write_indexes = tile(reps = write_indexes_reps_0, x = var_392)[name = tensor("write_indexes")]; + tensor var_395_begin_0 = const()[name = tensor("op_395_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_395_end_0 = const()[name = tensor("op_395_end_0"), val = tensor([1, 1, 256, 8, 64])]; + tensor var_395_end_mask_0 = const()[name = tensor("op_395_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_395_squeeze_mask_0 = const()[name = tensor("op_395_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor var_395 = slice_by_index(begin = var_395_begin_0, end = var_395_end_0, end_mask = var_395_end_mask_0, squeeze_mask = var_395_squeeze_mask_0, x = attn1_cache)[name = tensor("op_395")]; + tensor new_k_cache_5_axis_0 = const()[name = tensor("new_k_cache_5_axis_0"), val = tensor(1)]; + tensor new_k_cache_5_mode_0 = const()[name = tensor("new_k_cache_5_mode_0"), val = tensor("update")]; + tensor new_k_cache_5_validate_indices_0 = const()[name = tensor("new_k_cache_5_validate_indices_0"), val = tensor(false)]; + tensor new_k_cache_5 = scatter_along_axis(axis = new_k_cache_5_axis_0, data = var_395, indices = write_indexes, mode = new_k_cache_5_mode_0, updates = k, validate_indices = new_k_cache_5_validate_indices_0)[name = tensor("new_k_cache_5")]; + tensor var_397_begin_0 = const()[name = tensor("op_397_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor var_397_end_0 = const()[name = tensor("op_397_end_0"), val = tensor([2, 1, 256, 8, 64])]; + tensor var_397_end_mask_0 = const()[name = tensor("op_397_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_397_squeeze_mask_0 = const()[name = tensor("op_397_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor var_397 = slice_by_index(begin = var_397_begin_0, end = var_397_end_0, end_mask = var_397_end_mask_0, squeeze_mask = var_397_squeeze_mask_0, x = attn1_cache)[name = tensor("op_397")]; + tensor new_v_cache_5_axis_0 = const()[name = tensor("new_v_cache_5_axis_0"), val = tensor(1)]; + tensor new_v_cache_5_mode_0 = const()[name = tensor("new_v_cache_5_mode_0"), val = tensor("update")]; + tensor new_v_cache_5_validate_indices_0 = const()[name = tensor("new_v_cache_5_validate_indices_0"), val = tensor(false)]; + tensor new_v_cache_5 = scatter_along_axis(axis = new_v_cache_5_axis_0, data = var_397, indices = write_indexes, mode = new_v_cache_5_mode_0, updates = squeeze_5, validate_indices = new_v_cache_5_validate_indices_0)[name = tensor("new_v_cache_5")]; + tensor var_400_axis_0 = const()[name = tensor("op_400_axis_0"), val = tensor(0)]; + tensor var_400 = stack(axis = var_400_axis_0, values = (new_k_cache_5, new_v_cache_5))[name = tensor("op_400")]; + tensor var_401 = not_equal(x = new_k_cache_5, y = new_k_cache_5)[name = tensor("op_401")]; + tensor new_k_cache = select(a = var_212, b = new_k_cache_5, cond = var_401)[name = tensor("new_k_cache")]; + tensor var_404 = not_equal(x = new_v_cache_5, y = new_v_cache_5)[name = tensor("op_404")]; + tensor new_v_cache = select(a = var_212, b = new_v_cache_5, cond = var_404)[name = tensor("new_v_cache")]; + tensor var_409 = const()[name = tensor("op_409"), val = tensor([0, 2, 1, 3])]; + tensor var_411 = const()[name = tensor("op_411"), val = tensor([1, 1])]; + tensor var_412 = reshape(shape = var_411, x = attn1_offset)[name = tensor("op_412")]; + tensor var_414_promoted = const()[name = tensor("op_414_promoted"), val = tensor([0x1.ep+3])]; + tensor var_415 = add(x = var_412, y = var_414_promoted)[name = tensor("op_415")]; + tensor last_pos_dtype_0 = const()[name = tensor("last_pos_dtype_0"), val = tensor("int32")]; + tensor last_pos = cast(dtype = last_pos_dtype_0, x = var_415)[name = tensor("cast_45")]; + tensor diff = sub(x = last_pos, y = slot_idx_1)[name = tensor("diff")]; + tensor var_421_div = floor_div(x = diff, y = capacity)[name = tensor("op_421_div")]; + tensor var_421_div_scaled = mul(x = var_421_div, y = capacity)[name = tensor("op_421_div_scaled")]; + tensor var_421 = sub(x = diff, y = var_421_div_scaled)[name = tensor("op_421")]; + tensor pos_k = sub(x = last_pos, y = var_421)[name = tensor("pos_k")]; + tensor var_427_promoted = const()[name = tensor("op_427_promoted"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(41766144)))]; + tensor pos_q = add(x = var_412, y = var_427_promoted)[name = tensor("pos_q")]; + tensor var_431_axes_0 = const()[name = tensor("op_431_axes_0"), val = tensor([2])]; + tensor var_431 = expand_dims(axes = var_431_axes_0, x = pos_q)[name = tensor("op_431")]; + tensor var_433_axes_0 = const()[name = tensor("op_433_axes_0"), val = tensor([1])]; + tensor var_433 = expand_dims(axes = var_433_axes_0, x = pos_k)[name = tensor("op_433")]; + tensor var_434_promoted_dtype_0 = const()[name = tensor("op_434_promoted_dtype_0"), val = tensor("fp32")]; + tensor var_434_promoted = cast(dtype = var_434_promoted_dtype_0, x = var_433)[name = tensor("cast_44")]; + tensor delta = sub(x = var_431, y = var_434_promoted)[name = tensor("delta")]; + tensor valid_5 = greater_equal(x = var_433, y = var_86)[name = tensor("valid_5")]; + tensor var_443 = const()[name = tensor("op_443"), val = tensor([1, 1, 1])]; + tensor var_444 = reshape(shape = var_443, x = attn1_offset)[name = tensor("op_444")]; + tensor var_446_promoted = const()[name = tensor("op_446_promoted"), val = tensor([0x1.ep+3])]; + tensor var_447 = add(x = var_444, y = var_446_promoted)[name = tensor("op_447")]; + tensor var_448 = less_equal(x = var_434_promoted, y = var_447)[name = tensor("op_448")]; + tensor valid = logical_and(x = valid_5, y = var_448)[name = tensor("valid")]; + tensor var_86_promoted_1 = const()[name = tensor("op_86_promoted_1"), val = tensor(0x0p+0)]; + tensor var_450 = greater_equal(x = delta, y = var_86_promoted_1)[name = tensor("op_450")]; + tensor attn_mask_7 = logical_and(x = valid, y = var_450)[name = tensor("attn_mask_7")]; + tensor var_98_promoted_1 = const()[name = tensor("op_98_promoted_1"), val = tensor(0x1.f4p+7)]; + tensor var_452 = less(x = delta, y = var_98_promoted_1)[name = tensor("op_452")]; + tensor attn_mask_9 = logical_and(x = attn_mask_7, y = var_452)[name = tensor("attn_mask_9")]; + tensor attn_mask_axes_0 = const()[name = tensor("attn_mask_axes_0"), val = tensor([1])]; + tensor attn_mask = expand_dims(axes = attn_mask_axes_0, x = attn_mask_9)[name = tensor("attn_mask")]; + tensor var_457_transpose_x_0 = const()[name = tensor("op_457_transpose_x_0"), val = tensor(false)]; + tensor var_457_transpose_y_0 = const()[name = tensor("op_457_transpose_y_0"), val = tensor(false)]; + tensor transpose_8_perm_0 = const()[name = tensor("transpose_8_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_9_perm_0 = const()[name = tensor("transpose_9_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_9 = transpose(perm = transpose_9_perm_0, x = new_k_cache)[name = tensor("transpose_12")]; + tensor transpose_8 = transpose(perm = transpose_8_perm_0, x = q)[name = tensor("transpose_13")]; + tensor var_457 = matmul(transpose_x = var_457_transpose_x_0, transpose_y = var_457_transpose_y_0, x = transpose_8, y = transpose_9)[name = tensor("op_457")]; + tensor var_458 = const()[name = tensor("op_458"), val = tensor(0x1p-3)]; + tensor attn_7 = mul(x = var_457, y = var_458)[name = tensor("attn_7")]; + tensor var_460 = logical_not(x = attn_mask)[name = tensor("op_460")]; + tensor attn_9 = select(a = var_100, b = attn_7, cond = var_460)[name = tensor("attn_9")]; + tensor attn = softmax(axis = var_91, x = attn_9)[name = tensor("attn")]; + tensor x_11_transpose_x_0 = const()[name = tensor("x_11_transpose_x_0"), val = tensor(false)]; + tensor x_11_transpose_y_0 = const()[name = tensor("x_11_transpose_y_0"), val = tensor(false)]; + tensor v_attn = transpose(perm = var_409, x = new_v_cache)[name = tensor("transpose_14")]; + tensor x_11 = matmul(transpose_x = x_11_transpose_x_0, transpose_y = x_11_transpose_y_0, x = attn, y = v_attn)[name = tensor("x_11")]; + tensor var_464_perm_0 = const()[name = tensor("op_464_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_465 = const()[name = tensor("op_465"), val = tensor([1, 16, 512])]; + tensor var_464 = transpose(perm = var_464_perm_0, x = x_11)[name = tensor("transpose_11")]; + tensor input_15 = reshape(shape = var_465, x = var_464)[name = tensor("input_15")]; + tensor x_13 = linear(bias = linear_1_bias_0, weight = mimi_decoder_transformer_transformer_layers_1_self_attn_out_proj_weight, x = input_15)[name = tensor("linear_5")]; + tensor var_474 = mul(x = mimi_decoder_transformer_transformer_layers_1_layer_scale_1_scale, y = x_13)[name = tensor("op_474")]; + tensor input_17 = add(x = input_13, y = var_474)[name = tensor("input_17")]; + tensor input_19_axes_0 = const()[name = tensor("input_19_axes_0"), val = tensor([-1])]; + tensor input_19 = layer_norm(axes = input_19_axes_0, beta = mimi_decoder_transformer_transformer_layers_1_norm2_bias, epsilon = var_102, gamma = mimi_decoder_transformer_transformer_layers_1_norm2_weight, x = input_17)[name = tensor("input_19")]; + tensor var_481 = linear(bias = linear_2_bias_0, weight = mimi_decoder_transformer_transformer_layers_1_linear1_weight, x = input_19)[name = tensor("linear_6")]; + tensor input_21_mode_0 = const()[name = tensor("input_21_mode_0"), val = tensor("EXACT")]; + tensor input_21 = gelu(mode = input_21_mode_0, x = var_481)[name = tensor("input_21")]; + tensor x_15 = linear(bias = linear_1_bias_0, weight = mimi_decoder_transformer_transformer_layers_1_linear2_weight, x = input_21)[name = tensor("linear_7")]; + tensor var_487 = mul(x = mimi_decoder_transformer_transformer_layers_1_layer_scale_2_scale, y = x_15)[name = tensor("op_487")]; + tensor z = add(x = input_17, y = var_487)[name = tensor("z")]; + tensor x_17_perm_0 = const()[name = tensor("x_17_perm_0"), val = tensor([0, 2, 1])]; + tensor var_507 = const()[name = tensor("op_507"), val = tensor(0x1p+0)]; + tensor var_508 = const()[name = tensor("op_508"), val = tensor(-1)]; + tensor input_23_interleave_0 = const()[name = tensor("input_23_interleave_0"), val = tensor(false)]; + tensor x_17 = transpose(perm = x_17_perm_0, x = z)[name = tensor("transpose_10")]; + tensor input_23 = concat(axis = var_508, interleave = input_23_interleave_0, values = (conv0_prev, x_17))[name = tensor("input_23")]; + tensor input_25_pad_type_0 = const()[name = tensor("input_25_pad_type_0"), val = tensor("valid")]; + tensor input_25_strides_0 = const()[name = tensor("input_25_strides_0"), val = tensor([1])]; + tensor input_25_pad_0 = const()[name = tensor("input_25_pad_0"), val = tensor([0, 0])]; + tensor input_25_dilations_0 = const()[name = tensor("input_25_dilations_0"), val = tensor([1])]; + tensor input_25_groups_0 = const()[name = tensor("input_25_groups_0"), val = tensor(1)]; + tensor input_25 = conv(bias = mimi_decoder_model_0_conv_bias, dilations = input_25_dilations_0, groups = input_25_groups_0, pad = input_25_pad_0, pad_type = input_25_pad_type_0, strides = input_25_strides_0, weight = mimi_decoder_model_0_conv_weight, x = input_23)[name = tensor("input_25")]; + tensor var_542_begin_0 = const()[name = tensor("op_542_begin_0"), val = tensor([0, 0, 16])]; + tensor var_542_end_0 = const()[name = tensor("op_542_end_0"), val = tensor([1, 512, 22])]; + tensor var_542_end_mask_0 = const()[name = tensor("op_542_end_mask_0"), val = tensor([true, true, true])]; + tensor var_542 = slice_by_index(begin = var_542_begin_0, end = var_542_end_0, end_mask = var_542_end_mask_0, x = input_23)[name = tensor("op_542")]; + tensor input_27 = elu(alpha = var_507, x = input_25)[name = tensor("input_27")]; + tensor y_5_pad_type_0 = const()[name = tensor("y_5_pad_type_0"), val = tensor("valid")]; + tensor y_5_strides_0 = const()[name = tensor("y_5_strides_0"), val = tensor([6])]; + tensor y_5_pad_0 = const()[name = tensor("y_5_pad_0"), val = tensor([0, 0])]; + tensor y_5_dilations_0 = const()[name = tensor("y_5_dilations_0"), val = tensor([1])]; + tensor y_5_groups_0 = const()[name = tensor("y_5_groups_0"), val = tensor(1)]; + tensor y_5_has_output_shape_output_shape_0 = const()[name = tensor("y_5_has_output_shape_output_shape_0"), val = tensor([1, 256, 102])]; + tensor y_5_has_output_shape = conv_transpose(bias = mimi_decoder_model_2_convtr_bias, dilations = y_5_dilations_0, groups = y_5_groups_0, output_shape = y_5_has_output_shape_output_shape_0, pad = y_5_pad_0, pad_type = y_5_pad_type_0, strides = y_5_strides_0, weight = mimi_decoder_model_2_convtr_weight, x = input_27)[name = tensor("y_5_has_output_shape")]; + tensor var_557_begin_0 = const()[name = tensor("op_557_begin_0"), val = tensor([0, 0, 0])]; + tensor var_557_end_0 = const()[name = tensor("op_557_end_0"), val = tensor([1, 256, 6])]; + tensor var_557_end_mask_0 = const()[name = tensor("op_557_end_mask_0"), val = tensor([true, true, false])]; + tensor var_557 = slice_by_index(begin = var_557_begin_0, end = var_557_end_0, end_mask = var_557_end_mask_0, x = y_5_has_output_shape)[name = tensor("op_557")]; + tensor var_558 = add(x = var_557, y = convtr0_partial)[name = tensor("op_558")]; + tensor var_559_begin_0 = const()[name = tensor("op_559_begin_0"), val = tensor([0, 0, 6])]; + tensor var_559_end_0 = const()[name = tensor("op_559_end_0"), val = tensor([1, 256, 102])]; + tensor var_559_end_mask_0 = const()[name = tensor("op_559_end_mask_0"), val = tensor([true, true, true])]; + tensor var_559 = slice_by_index(begin = var_559_begin_0, end = var_559_end_0, end_mask = var_559_end_mask_0, x = y_5_has_output_shape)[name = tensor("op_559")]; + tensor y_7_interleave_0 = const()[name = tensor("y_7_interleave_0"), val = tensor(false)]; + tensor y_7 = concat(axis = var_508, interleave = y_7_interleave_0, values = (var_558, var_559))[name = tensor("y_7")]; + tensor new_partial_1_begin_0 = const()[name = tensor("new_partial_1_begin_0"), val = tensor([0, 0, 96])]; + tensor new_partial_1_end_0 = const()[name = tensor("new_partial_1_end_0"), val = tensor([1, 256, 102])]; + tensor new_partial_1_end_mask_0 = const()[name = tensor("new_partial_1_end_mask_0"), val = tensor([true, true, true])]; + tensor new_partial_1 = slice_by_index(begin = new_partial_1_begin_0, end = new_partial_1_end_0, end_mask = new_partial_1_end_mask_0, x = y_7)[name = tensor("new_partial_1")]; + tensor var_564 = const()[name = tensor("op_564"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(41766272)))]; + tensor var_565 = sub(x = new_partial_1, y = var_564)[name = tensor("op_565")]; + tensor input_29_begin_0 = const()[name = tensor("input_29_begin_0"), val = tensor([0, 0, 0])]; + tensor input_29_end_0 = const()[name = tensor("input_29_end_0"), val = tensor([1, 256, 96])]; + tensor input_29_end_mask_0 = const()[name = tensor("input_29_end_mask_0"), val = tensor([true, true, false])]; + tensor input_29 = slice_by_index(begin = input_29_begin_0, end = input_29_end_0, end_mask = input_29_end_mask_0, x = y_7)[name = tensor("input_29")]; + tensor x_19 = elu(alpha = var_507, x = input_29)[name = tensor("x_19")]; + tensor input_31_interleave_0 = const()[name = tensor("input_31_interleave_0"), val = tensor(false)]; + tensor input_31 = concat(axis = var_508, interleave = input_31_interleave_0, values = (res0_conv0_prev, x_19))[name = tensor("input_31")]; + tensor input_33_pad_type_0 = const()[name = tensor("input_33_pad_type_0"), val = tensor("valid")]; + tensor input_33_strides_0 = const()[name = tensor("input_33_strides_0"), val = tensor([1])]; + tensor input_33_pad_0 = const()[name = tensor("input_33_pad_0"), val = tensor([0, 0])]; + tensor input_33_dilations_0 = const()[name = tensor("input_33_dilations_0"), val = tensor([1])]; + tensor input_33_groups_0 = const()[name = tensor("input_33_groups_0"), val = tensor(1)]; + tensor input_33 = conv(bias = mimi_decoder_model_3_block_1_conv_bias, dilations = input_33_dilations_0, groups = input_33_groups_0, pad = input_33_pad_0, pad_type = input_33_pad_type_0, strides = input_33_strides_0, weight = mimi_decoder_model_3_block_1_conv_weight, x = input_31)[name = tensor("input_33")]; + tensor var_585_begin_0 = const()[name = tensor("op_585_begin_0"), val = tensor([0, 0, 96])]; + tensor var_585_end_0 = const()[name = tensor("op_585_end_0"), val = tensor([1, 256, 98])]; + tensor var_585_end_mask_0 = const()[name = tensor("op_585_end_mask_0"), val = tensor([true, true, true])]; + tensor var_585 = slice_by_index(begin = var_585_begin_0, end = var_585_end_0, end_mask = var_585_end_mask_0, x = input_31)[name = tensor("op_585")]; + tensor x_21 = elu(alpha = var_507, x = input_33)[name = tensor("x_21")]; + tensor v_5_pad_type_0 = const()[name = tensor("v_5_pad_type_0"), val = tensor("valid")]; + tensor v_5_strides_0 = const()[name = tensor("v_5_strides_0"), val = tensor([1])]; + tensor v_5_pad_0 = const()[name = tensor("v_5_pad_0"), val = tensor([0, 0])]; + tensor v_5_dilations_0 = const()[name = tensor("v_5_dilations_0"), val = tensor([1])]; + tensor v_5_groups_0 = const()[name = tensor("v_5_groups_0"), val = tensor(1)]; + tensor v_5 = conv(bias = mimi_decoder_model_3_block_3_conv_bias, dilations = v_5_dilations_0, groups = v_5_groups_0, pad = v_5_pad_0, pad_type = v_5_pad_type_0, strides = v_5_strides_0, weight = mimi_decoder_model_3_block_3_conv_weight, x = x_21)[name = tensor("v_5")]; + tensor input_35 = add(x = input_29, y = v_5)[name = tensor("input_35")]; + tensor input_37 = elu(alpha = var_507, x = input_35)[name = tensor("input_37")]; + tensor y_9_pad_type_0 = const()[name = tensor("y_9_pad_type_0"), val = tensor("valid")]; + tensor y_9_strides_0 = const()[name = tensor("y_9_strides_0"), val = tensor([5])]; + tensor y_9_pad_0 = const()[name = tensor("y_9_pad_0"), val = tensor([0, 0])]; + tensor y_9_dilations_0 = const()[name = tensor("y_9_dilations_0"), val = tensor([1])]; + tensor y_9_groups_0 = const()[name = tensor("y_9_groups_0"), val = tensor(1)]; + tensor y_9_has_output_shape_output_shape_0 = const()[name = tensor("y_9_has_output_shape_output_shape_0"), val = tensor([1, 128, 485])]; + tensor y_9_has_output_shape = conv_transpose(bias = mimi_decoder_model_5_convtr_bias, dilations = y_9_dilations_0, groups = y_9_groups_0, output_shape = y_9_has_output_shape_output_shape_0, pad = y_9_pad_0, pad_type = y_9_pad_type_0, strides = y_9_strides_0, weight = mimi_decoder_model_5_convtr_weight, x = input_37)[name = tensor("y_9_has_output_shape")]; + tensor var_613_begin_0 = const()[name = tensor("op_613_begin_0"), val = tensor([0, 0, 0])]; + tensor var_613_end_0 = const()[name = tensor("op_613_end_0"), val = tensor([1, 128, 5])]; + tensor var_613_end_mask_0 = const()[name = tensor("op_613_end_mask_0"), val = tensor([true, true, false])]; + tensor var_613 = slice_by_index(begin = var_613_begin_0, end = var_613_end_0, end_mask = var_613_end_mask_0, x = y_9_has_output_shape)[name = tensor("op_613")]; + tensor var_614 = add(x = var_613, y = convtr1_partial)[name = tensor("op_614")]; + tensor var_615_begin_0 = const()[name = tensor("op_615_begin_0"), val = tensor([0, 0, 5])]; + tensor var_615_end_0 = const()[name = tensor("op_615_end_0"), val = tensor([1, 128, 485])]; + tensor var_615_end_mask_0 = const()[name = tensor("op_615_end_mask_0"), val = tensor([true, true, true])]; + tensor var_615 = slice_by_index(begin = var_615_begin_0, end = var_615_end_0, end_mask = var_615_end_mask_0, x = y_9_has_output_shape)[name = tensor("op_615")]; + tensor y_11_interleave_0 = const()[name = tensor("y_11_interleave_0"), val = tensor(false)]; + tensor y_11 = concat(axis = var_508, interleave = y_11_interleave_0, values = (var_614, var_615))[name = tensor("y_11")]; + tensor new_partial_3_begin_0 = const()[name = tensor("new_partial_3_begin_0"), val = tensor([0, 0, 480])]; + tensor new_partial_3_end_0 = const()[name = tensor("new_partial_3_end_0"), val = tensor([1, 128, 485])]; + tensor new_partial_3_end_mask_0 = const()[name = tensor("new_partial_3_end_mask_0"), val = tensor([true, true, true])]; + tensor new_partial_3 = slice_by_index(begin = new_partial_3_begin_0, end = new_partial_3_end_0, end_mask = new_partial_3_end_mask_0, x = y_11)[name = tensor("new_partial_3")]; + tensor var_620 = const()[name = tensor("op_620"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(41767360)))]; + tensor var_621 = sub(x = new_partial_3, y = var_620)[name = tensor("op_621")]; + tensor input_39_begin_0 = const()[name = tensor("input_39_begin_0"), val = tensor([0, 0, 0])]; + tensor input_39_end_0 = const()[name = tensor("input_39_end_0"), val = tensor([1, 128, 480])]; + tensor input_39_end_mask_0 = const()[name = tensor("input_39_end_mask_0"), val = tensor([true, true, false])]; + tensor input_39 = slice_by_index(begin = input_39_begin_0, end = input_39_end_0, end_mask = input_39_end_mask_0, x = y_11)[name = tensor("input_39")]; + tensor x_23 = elu(alpha = var_507, x = input_39)[name = tensor("x_23")]; + tensor input_41_interleave_0 = const()[name = tensor("input_41_interleave_0"), val = tensor(false)]; + tensor input_41 = concat(axis = var_508, interleave = input_41_interleave_0, values = (res1_conv0_prev, x_23))[name = tensor("input_41")]; + tensor input_43_pad_type_0 = const()[name = tensor("input_43_pad_type_0"), val = tensor("valid")]; + tensor input_43_strides_0 = const()[name = tensor("input_43_strides_0"), val = tensor([1])]; + tensor input_43_pad_0 = const()[name = tensor("input_43_pad_0"), val = tensor([0, 0])]; + tensor input_43_dilations_0 = const()[name = tensor("input_43_dilations_0"), val = tensor([1])]; + tensor input_43_groups_0 = const()[name = tensor("input_43_groups_0"), val = tensor(1)]; + tensor input_43 = conv(bias = mimi_decoder_model_6_block_1_conv_bias, dilations = input_43_dilations_0, groups = input_43_groups_0, pad = input_43_pad_0, pad_type = input_43_pad_type_0, strides = input_43_strides_0, weight = mimi_decoder_model_6_block_1_conv_weight, x = input_41)[name = tensor("input_43")]; + tensor var_641_begin_0 = const()[name = tensor("op_641_begin_0"), val = tensor([0, 0, 480])]; + tensor var_641_end_0 = const()[name = tensor("op_641_end_0"), val = tensor([1, 128, 482])]; + tensor var_641_end_mask_0 = const()[name = tensor("op_641_end_mask_0"), val = tensor([true, true, true])]; + tensor var_641 = slice_by_index(begin = var_641_begin_0, end = var_641_end_0, end_mask = var_641_end_mask_0, x = input_41)[name = tensor("op_641")]; + tensor x_25 = elu(alpha = var_507, x = input_43)[name = tensor("x_25")]; + tensor v_7_pad_type_0 = const()[name = tensor("v_7_pad_type_0"), val = tensor("valid")]; + tensor v_7_strides_0 = const()[name = tensor("v_7_strides_0"), val = tensor([1])]; + tensor v_7_pad_0 = const()[name = tensor("v_7_pad_0"), val = tensor([0, 0])]; + tensor v_7_dilations_0 = const()[name = tensor("v_7_dilations_0"), val = tensor([1])]; + tensor v_7_groups_0 = const()[name = tensor("v_7_groups_0"), val = tensor(1)]; + tensor v_7 = conv(bias = mimi_decoder_model_6_block_3_conv_bias, dilations = v_7_dilations_0, groups = v_7_groups_0, pad = v_7_pad_0, pad_type = v_7_pad_type_0, strides = v_7_strides_0, weight = mimi_decoder_model_6_block_3_conv_weight, x = x_25)[name = tensor("v_7")]; + tensor input_45 = add(x = input_39, y = v_7)[name = tensor("input_45")]; + tensor input_47 = elu(alpha = var_507, x = input_45)[name = tensor("input_47")]; + tensor y_13_pad_type_0 = const()[name = tensor("y_13_pad_type_0"), val = tensor("valid")]; + tensor y_13_strides_0 = const()[name = tensor("y_13_strides_0"), val = tensor([4])]; + tensor y_13_pad_0 = const()[name = tensor("y_13_pad_0"), val = tensor([0, 0])]; + tensor y_13_dilations_0 = const()[name = tensor("y_13_dilations_0"), val = tensor([1])]; + tensor y_13_groups_0 = const()[name = tensor("y_13_groups_0"), val = tensor(1)]; + tensor y_13_has_output_shape_output_shape_0 = const()[name = tensor("y_13_has_output_shape_output_shape_0"), val = tensor([1, 64, 1924])]; + tensor y_13_has_output_shape = conv_transpose(bias = mimi_decoder_model_8_convtr_bias, dilations = y_13_dilations_0, groups = y_13_groups_0, output_shape = y_13_has_output_shape_output_shape_0, pad = y_13_pad_0, pad_type = y_13_pad_type_0, strides = y_13_strides_0, weight = mimi_decoder_model_8_convtr_weight, x = input_47)[name = tensor("y_13_has_output_shape")]; + tensor var_669_begin_0 = const()[name = tensor("op_669_begin_0"), val = tensor([0, 0, 0])]; + tensor var_669_end_0 = const()[name = tensor("op_669_end_0"), val = tensor([1, 64, 4])]; + tensor var_669_end_mask_0 = const()[name = tensor("op_669_end_mask_0"), val = tensor([true, true, false])]; + tensor var_669 = slice_by_index(begin = var_669_begin_0, end = var_669_end_0, end_mask = var_669_end_mask_0, x = y_13_has_output_shape)[name = tensor("op_669")]; + tensor var_670 = add(x = var_669, y = convtr2_partial)[name = tensor("op_670")]; + tensor var_671_begin_0 = const()[name = tensor("op_671_begin_0"), val = tensor([0, 0, 4])]; + tensor var_671_end_0 = const()[name = tensor("op_671_end_0"), val = tensor([1, 64, 1924])]; + tensor var_671_end_mask_0 = const()[name = tensor("op_671_end_mask_0"), val = tensor([true, true, true])]; + tensor var_671 = slice_by_index(begin = var_671_begin_0, end = var_671_end_0, end_mask = var_671_end_mask_0, x = y_13_has_output_shape)[name = tensor("op_671")]; + tensor y_interleave_0 = const()[name = tensor("y_interleave_0"), val = tensor(false)]; + tensor y = concat(axis = var_508, interleave = y_interleave_0, values = (var_670, var_671))[name = tensor("y")]; + tensor new_partial_begin_0 = const()[name = tensor("new_partial_begin_0"), val = tensor([0, 0, 1920])]; + tensor new_partial_end_0 = const()[name = tensor("new_partial_end_0"), val = tensor([1, 64, 1924])]; + tensor new_partial_end_mask_0 = const()[name = tensor("new_partial_end_mask_0"), val = tensor([true, true, true])]; + tensor new_partial = slice_by_index(begin = new_partial_begin_0, end = new_partial_end_0, end_mask = new_partial_end_mask_0, x = y)[name = tensor("new_partial")]; + tensor var_676 = const()[name = tensor("op_676"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(41767936)))]; + tensor var_677 = sub(x = new_partial, y = var_676)[name = tensor("op_677")]; + tensor input_49_begin_0 = const()[name = tensor("input_49_begin_0"), val = tensor([0, 0, 0])]; + tensor input_49_end_0 = const()[name = tensor("input_49_end_0"), val = tensor([1, 64, 1920])]; + tensor input_49_end_mask_0 = const()[name = tensor("input_49_end_mask_0"), val = tensor([true, true, false])]; + tensor input_49 = slice_by_index(begin = input_49_begin_0, end = input_49_end_0, end_mask = input_49_end_mask_0, x = y)[name = tensor("input_49")]; + tensor x_27 = elu(alpha = var_507, x = input_49)[name = tensor("x_27")]; + tensor input_51_interleave_0 = const()[name = tensor("input_51_interleave_0"), val = tensor(false)]; + tensor input_51 = concat(axis = var_508, interleave = input_51_interleave_0, values = (res2_conv0_prev, x_27))[name = tensor("input_51")]; + tensor input_53_pad_type_0 = const()[name = tensor("input_53_pad_type_0"), val = tensor("valid")]; + tensor input_53_strides_0 = const()[name = tensor("input_53_strides_0"), val = tensor([1])]; + tensor input_53_pad_0 = const()[name = tensor("input_53_pad_0"), val = tensor([0, 0])]; + tensor input_53_dilations_0 = const()[name = tensor("input_53_dilations_0"), val = tensor([1])]; + tensor input_53_groups_0 = const()[name = tensor("input_53_groups_0"), val = tensor(1)]; + tensor input_53 = conv(bias = mimi_decoder_model_9_block_1_conv_bias, dilations = input_53_dilations_0, groups = input_53_groups_0, pad = input_53_pad_0, pad_type = input_53_pad_type_0, strides = input_53_strides_0, weight = mimi_decoder_model_9_block_1_conv_weight, x = input_51)[name = tensor("input_53")]; + tensor var_697_begin_0 = const()[name = tensor("op_697_begin_0"), val = tensor([0, 0, 1920])]; + tensor var_697_end_0 = const()[name = tensor("op_697_end_0"), val = tensor([1, 64, 1922])]; + tensor var_697_end_mask_0 = const()[name = tensor("op_697_end_mask_0"), val = tensor([true, true, true])]; + tensor var_697 = slice_by_index(begin = var_697_begin_0, end = var_697_end_0, end_mask = var_697_end_mask_0, x = input_51)[name = tensor("op_697")]; + tensor x_29 = elu(alpha = var_507, x = input_53)[name = tensor("x_29")]; + tensor v_pad_type_0 = const()[name = tensor("v_pad_type_0"), val = tensor("valid")]; + tensor v_strides_0 = const()[name = tensor("v_strides_0"), val = tensor([1])]; + tensor v_pad_0 = const()[name = tensor("v_pad_0"), val = tensor([0, 0])]; + tensor v_dilations_0 = const()[name = tensor("v_dilations_0"), val = tensor([1])]; + tensor v_groups_0 = const()[name = tensor("v_groups_0"), val = tensor(1)]; + tensor v = conv(bias = mimi_decoder_model_9_block_3_conv_bias, dilations = v_dilations_0, groups = v_groups_0, pad = v_pad_0, pad_type = v_pad_type_0, strides = v_strides_0, weight = mimi_decoder_model_9_block_3_conv_weight, x = x_29)[name = tensor("v")]; + tensor input_55 = add(x = input_49, y = v)[name = tensor("input_55")]; + tensor x = elu(alpha = var_507, x = input_55)[name = tensor("x")]; + tensor input_interleave_0 = const()[name = tensor("input_interleave_0"), val = tensor(false)]; + tensor input = concat(axis = var_508, interleave = input_interleave_0, values = (conv_final_prev, x))[name = tensor("input")]; + tensor var_724_pad_type_0 = const()[name = tensor("op_724_pad_type_0"), val = tensor("valid")]; + tensor var_724_strides_0 = const()[name = tensor("op_724_strides_0"), val = tensor([1])]; + tensor var_724_pad_0 = const()[name = tensor("op_724_pad_0"), val = tensor([0, 0])]; + tensor var_724_dilations_0 = const()[name = tensor("op_724_dilations_0"), val = tensor([1])]; + tensor var_724_groups_0 = const()[name = tensor("op_724_groups_0"), val = tensor(1)]; + tensor var_724 = conv(bias = mimi_decoder_model_11_conv_bias, dilations = var_724_dilations_0, groups = var_724_groups_0, pad = var_724_pad_0, pad_type = var_724_pad_type_0, strides = var_724_strides_0, weight = mimi_decoder_model_11_conv_weight, x = input)[name = tensor("op_724")]; + tensor var_725_begin_0 = const()[name = tensor("op_725_begin_0"), val = tensor([0, 0, 1920])]; + tensor var_725_end_0 = const()[name = tensor("op_725_end_0"), val = tensor([1, 64, 1922])]; + tensor var_725_end_mask_0 = const()[name = tensor("op_725_end_mask_0"), val = tensor([true, true, true])]; + tensor var_725 = slice_by_index(begin = var_725_begin_0, end = var_725_end_0, end_mask = var_725_end_mask_0, x = input)[name = tensor("op_725")]; + tensor var_740_promoted = const()[name = tensor("op_740_promoted"), val = tensor(0x1p+4)]; + tensor var_741 = add(x = attn0_offset, y = var_740_promoted)[name = tensor("op_741")]; + tensor var_743_promoted = const()[name = tensor("op_743_promoted"), val = tensor(0x1p+4)]; + tensor var_744 = add(x = attn1_offset, y = var_743_promoted)[name = tensor("op_744")]; + tensor conv0_first_tmp = identity(x = conv0_first)[name = tensor("conv0_first_tmp")]; + tensor res0_conv0_first_tmp = identity(x = res0_conv0_first)[name = tensor("res0_conv0_first_tmp")]; + tensor res0_conv1_prev_tmp = identity(x = res0_conv1_prev)[name = tensor("res0_conv1_prev_tmp")]; + tensor res0_conv1_first_tmp = identity(x = res0_conv1_first)[name = tensor("res0_conv1_first_tmp")]; + tensor res1_conv0_first_tmp = identity(x = res1_conv0_first)[name = tensor("res1_conv0_first_tmp")]; + tensor res1_conv1_prev_tmp = identity(x = res1_conv1_prev)[name = tensor("res1_conv1_prev_tmp")]; + tensor res1_conv1_first_tmp = identity(x = res1_conv1_first)[name = tensor("res1_conv1_first_tmp")]; + tensor res2_conv0_first_tmp = identity(x = res2_conv0_first)[name = tensor("res2_conv0_first_tmp")]; + tensor res2_conv1_prev_tmp = identity(x = res2_conv1_prev)[name = tensor("res2_conv1_prev_tmp")]; + tensor res2_conv1_first_tmp = identity(x = res2_conv1_first)[name = tensor("res2_conv1_first_tmp")]; + tensor conv_final_first_tmp = identity(x = conv_final_first)[name = tensor("conv_final_first_tmp")]; + } -> (var_724, var_77, var_210, var_741, var_400, var_744, var_542, conv0_first, var_565, var_585, res0_conv0_first, res0_conv1_prev, res0_conv1_first, var_621, var_641, res1_conv0_first, res1_conv1_prev, res1_conv1_first, var_677, var_697, res2_conv0_first, res2_conv1_prev, res2_conv1_first, var_725, conv_final_first); +} \ No newline at end of file diff --git a/v2.1/italian_24l/mimi_decoder.mlmodelc/weights/weight.bin b/v2.1/italian_24l/mimi_decoder.mlmodelc/weights/weight.bin new file mode 100644 index 0000000000000000000000000000000000000000..c2896951f3ea7956cb354933a2e293a51ccf6a93 --- /dev/null +++ b/v2.1/italian_24l/mimi_decoder.mlmodelc/weights/weight.bin @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:42b8f0af28e5488e1de239479690b51b4a2411501ef1a1de38b9244317bb0cc2 +size 41768256 diff --git a/v2.1/italian_24l/mimi_decoder.mlpackage/Data/com.apple.CoreML/model.mlmodel b/v2.1/italian_24l/mimi_decoder.mlpackage/Data/com.apple.CoreML/model.mlmodel new file mode 100644 index 0000000000000000000000000000000000000000..d42a7bdad8c4b2b6edc7ffd6e6747e6b3f2b1dcb --- /dev/null +++ b/v2.1/italian_24l/mimi_decoder.mlpackage/Data/com.apple.CoreML/model.mlmodel @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:65a21db69ffb7efce1f076a9e1eafb5d9f135f2e84420c5908c4c83300a08157 +size 93270 diff --git a/v2.1/italian_24l/mimi_decoder.mlpackage/Data/com.apple.CoreML/weights/weight.bin b/v2.1/italian_24l/mimi_decoder.mlpackage/Data/com.apple.CoreML/weights/weight.bin new file mode 100644 index 0000000000000000000000000000000000000000..c2896951f3ea7956cb354933a2e293a51ccf6a93 --- /dev/null +++ b/v2.1/italian_24l/mimi_decoder.mlpackage/Data/com.apple.CoreML/weights/weight.bin @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:42b8f0af28e5488e1de239479690b51b4a2411501ef1a1de38b9244317bb0cc2 +size 41768256 diff --git a/v2.1/italian_24l/mimi_decoder.mlpackage/Manifest.json b/v2.1/italian_24l/mimi_decoder.mlpackage/Manifest.json new file mode 100644 index 0000000000000000000000000000000000000000..705e43f697ee69a6ddebe5d9b41c929abdd58a8a --- /dev/null +++ b/v2.1/italian_24l/mimi_decoder.mlpackage/Manifest.json @@ -0,0 +1,18 @@ +{ + "fileFormatVersion": "1.0.0", + "itemInfoEntries": { + "041E070C-BA82-4303-952E-3636B3F4CFFD": { + "author": "com.apple.CoreML", + "description": "CoreML Model Specification", + "name": "model.mlmodel", + "path": "com.apple.CoreML/model.mlmodel" + }, + "17D22254-EFDB-4740-BE07-10C20CACA6C8": { + "author": "com.apple.CoreML", + "description": "CoreML Model Weights", + "name": "weights", + "path": "com.apple.CoreML/weights" + } + }, + "rootModelIdentifier": "041E070C-BA82-4303-952E-3636B3F4CFFD" +} diff --git a/v2.1/portuguese/cond_prefill.mlmodelc/analytics/coremldata.bin b/v2.1/portuguese/cond_prefill.mlmodelc/analytics/coremldata.bin new file mode 100644 index 0000000000000000000000000000000000000000..73b74ccac138784214a3471e0bf4b9ffabba9156 --- /dev/null +++ b/v2.1/portuguese/cond_prefill.mlmodelc/analytics/coremldata.bin @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:e4b74e555e2239cd6fd2819f2e99aa24f908f963231ad90e83647752d65f72ee +size 243 diff --git a/v2.1/portuguese/cond_prefill.mlmodelc/coremldata.bin b/v2.1/portuguese/cond_prefill.mlmodelc/coremldata.bin new file mode 100644 index 0000000000000000000000000000000000000000..50cb8792086011205d69b3f1726edc5d5bdb817f --- /dev/null +++ b/v2.1/portuguese/cond_prefill.mlmodelc/coremldata.bin @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:604c2ce9a2df18b11cb60efb5d78ca9313a7f6e8804968e929dbd2c458f53aa9 +size 1151 diff --git a/v2.1/portuguese/cond_prefill.mlmodelc/model.mil b/v2.1/portuguese/cond_prefill.mlmodelc/model.mil new file mode 100644 index 0000000000000000000000000000000000000000..203b091a2c5043d4b8c07f4402d5a5f0319b3c3f --- /dev/null +++ b/v2.1/portuguese/cond_prefill.mlmodelc/model.mil @@ -0,0 +1,1309 @@ +program(1.0) +[buildInfo = dict, tensor>({{"coremlc-component-MIL", "3520.4.1"}, {"coremlc-version", "3520.5.1"}, {"coremltools-component-torch", "2.12.0"}, {"coremltools-source-dialect", "TorchScript"}, {"coremltools-version", "9.0"}})] +{ + func main(tensor cache0, tensor cache1, tensor cache2, tensor cache3, tensor cache4, tensor cache5, tensor conditioning, tensor position0, tensor position1, tensor position2, tensor position3, tensor position4, tensor position5, tensor valid_len) { + tensor input_3_axes_0 = const()[name = tensor("input_3_axes_0"), val = tensor([-1])]; + tensor conditioning_to_fp16_dtype_0 = const()[name = tensor("conditioning_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor norm0_1_weight_to_fp16 = const()[name = tensor("norm0_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(64)))]; + tensor norm0_1_bias_to_fp16 = const()[name = tensor("norm0_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(2176)))]; + tensor var_48_to_fp16 = const()[name = tensor("op_48_to_fp16"), val = tensor(0x1.5p-17)]; + tensor conditioning_to_fp16 = cast(dtype = conditioning_to_fp16_dtype_0, x = conditioning)[name = tensor("cast_159")]; + tensor input_3_cast_fp16 = layer_norm(axes = input_3_axes_0, beta = norm0_1_bias_to_fp16, epsilon = var_48_to_fp16, gamma = norm0_1_weight_to_fp16, x = conditioning_to_fp16)[name = tensor("input_3_cast_fp16")]; + tensor attn0_in_proj_weight_to_fp16 = const()[name = tensor("attn0_in_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(4288)))]; + tensor linear_0_bias_0_to_fp16 = const()[name = tensor("linear_0_bias_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(6295808)))]; + tensor linear_0_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = attn0_in_proj_weight_to_fp16, x = input_3_cast_fp16)[name = tensor("linear_0_cast_fp16")]; + tensor var_62 = const()[name = tensor("op_62"), val = tensor([1, 256, 3, 16, 64])]; + tensor qkv_1_cast_fp16 = reshape(shape = var_62, x = linear_0_cast_fp16)[name = tensor("qkv_1_cast_fp16")]; + tensor q_1_begin_0 = const()[name = tensor("q_1_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor q_1_end_0 = const()[name = tensor("q_1_end_0"), val = tensor([1, 256, 1, 16, 64])]; + tensor q_1_end_mask_0 = const()[name = tensor("q_1_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor q_1_squeeze_mask_0 = const()[name = tensor("q_1_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor q_1_cast_fp16 = slice_by_index(begin = q_1_begin_0, end = q_1_end_0, end_mask = q_1_end_mask_0, squeeze_mask = q_1_squeeze_mask_0, x = qkv_1_cast_fp16)[name = tensor("q_1_cast_fp16")]; + tensor k_1_begin_0 = const()[name = tensor("k_1_begin_0"), val = tensor([0, 0, 1, 0, 0])]; + tensor k_1_end_0 = const()[name = tensor("k_1_end_0"), val = tensor([1, 256, 2, 16, 64])]; + tensor k_1_end_mask_0 = const()[name = tensor("k_1_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor k_1_squeeze_mask_0 = const()[name = tensor("k_1_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor k_1_cast_fp16 = slice_by_index(begin = k_1_begin_0, end = k_1_end_0, end_mask = k_1_end_mask_0, squeeze_mask = k_1_squeeze_mask_0, x = qkv_1_cast_fp16)[name = tensor("k_1_cast_fp16")]; + tensor v_1_begin_0 = const()[name = tensor("v_1_begin_0"), val = tensor([0, 0, 2, 0, 0])]; + tensor v_1_end_0 = const()[name = tensor("v_1_end_0"), val = tensor([1, 256, 3, 16, 64])]; + tensor v_1_end_mask_0 = const()[name = tensor("v_1_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor v_1_squeeze_mask_0 = const()[name = tensor("v_1_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor v_1_cast_fp16 = slice_by_index(begin = v_1_begin_0, end = v_1_end_0, end_mask = v_1_end_mask_0, squeeze_mask = v_1_squeeze_mask_0, x = qkv_1_cast_fp16)[name = tensor("v_1_cast_fp16")]; + tensor var_120 = const()[name = tensor("op_120"), val = tensor([1, 1])]; + tensor position0_to_fp16_dtype_0 = const()[name = tensor("position0_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor position0_to_fp16 = cast(dtype = position0_to_fp16_dtype_0, x = position0)[name = tensor("cast_158")]; + tensor var_121_cast_fp16 = reshape(shape = var_120, x = position0_to_fp16)[name = tensor("op_121_cast_fp16")]; + tensor ts_1_promoted_to_fp16 = const()[name = tensor("ts_1_promoted_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(6302016)))]; + tensor ts_3_cast_fp16 = add(x = ts_1_promoted_to_fp16, y = var_121_cast_fp16)[name = tensor("ts_3_cast_fp16")]; + tensor var_128 = const()[name = tensor("op_128"), val = tensor([1, 256, 1, 1])]; + tensor ts_5_cast_fp16 = reshape(shape = var_128, x = ts_3_cast_fp16)[name = tensor("ts_5_cast_fp16")]; + tensor var_135 = const()[name = tensor("op_135"), val = tensor([1, 256, 16, 32, 2])]; + tensor q_complex_1_cast_fp16 = reshape(shape = var_135, x = q_1_cast_fp16)[name = tensor("q_complex_1_cast_fp16")]; + tensor var_142 = const()[name = tensor("op_142"), val = tensor([1, 256, 16, 32, 2])]; + tensor k_complex_1_cast_fp16 = reshape(shape = var_142, x = k_1_cast_fp16)[name = tensor("k_complex_1_cast_fp16")]; + tensor var_146_begin_0 = const()[name = tensor("op_146_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_146_end_0 = const()[name = tensor("op_146_end_0"), val = tensor([1, 256, 16, 32, 1])]; + tensor var_146_end_mask_0 = const()[name = tensor("op_146_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_146_squeeze_mask_0 = const()[name = tensor("op_146_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_146_cast_fp16 = slice_by_index(begin = var_146_begin_0, end = var_146_end_0, end_mask = var_146_end_mask_0, squeeze_mask = var_146_squeeze_mask_0, x = q_complex_1_cast_fp16)[name = tensor("op_146_cast_fp16")]; + tensor var_154_begin_0 = const()[name = tensor("op_154_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_154_end_0 = const()[name = tensor("op_154_end_0"), val = tensor([1, 256, 16, 32, 2])]; + tensor var_154_end_mask_0 = const()[name = tensor("op_154_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_154_squeeze_mask_0 = const()[name = tensor("op_154_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_154_cast_fp16 = slice_by_index(begin = var_154_begin_0, end = var_154_end_0, end_mask = var_154_end_mask_0, squeeze_mask = var_154_squeeze_mask_0, x = q_complex_1_cast_fp16)[name = tensor("op_154_cast_fp16")]; + tensor var_162_begin_0 = const()[name = tensor("op_162_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_162_end_0 = const()[name = tensor("op_162_end_0"), val = tensor([1, 256, 16, 32, 1])]; + tensor var_162_end_mask_0 = const()[name = tensor("op_162_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_162_squeeze_mask_0 = const()[name = tensor("op_162_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_162_cast_fp16 = slice_by_index(begin = var_162_begin_0, end = var_162_end_0, end_mask = var_162_end_mask_0, squeeze_mask = var_162_squeeze_mask_0, x = k_complex_1_cast_fp16)[name = tensor("op_162_cast_fp16")]; + tensor var_170_begin_0 = const()[name = tensor("op_170_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_170_end_0 = const()[name = tensor("op_170_end_0"), val = tensor([1, 256, 16, 32, 2])]; + tensor var_170_end_mask_0 = const()[name = tensor("op_170_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_170_squeeze_mask_0 = const()[name = tensor("op_170_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_170_cast_fp16 = slice_by_index(begin = var_170_begin_0, end = var_170_end_0, end_mask = var_170_end_mask_0, squeeze_mask = var_170_squeeze_mask_0, x = k_complex_1_cast_fp16)[name = tensor("op_170_cast_fp16")]; + tensor freqs_1_to_fp16 = const()[name = tensor("freqs_1_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(6302592)))]; + tensor var_176_cast_fp16 = mul(x = freqs_1_to_fp16, y = ts_5_cast_fp16)[name = tensor("op_176_cast_fp16")]; + tensor rotr_1_cast_fp16 = cos(x = var_176_cast_fp16)[name = tensor("rotr_1_cast_fp16")]; + tensor roti_1_cast_fp16 = sin(x = var_176_cast_fp16)[name = tensor("roti_1_cast_fp16")]; + tensor var_180_cast_fp16 = mul(x = var_146_cast_fp16, y = rotr_1_cast_fp16)[name = tensor("op_180_cast_fp16")]; + tensor var_181_cast_fp16 = mul(x = var_154_cast_fp16, y = roti_1_cast_fp16)[name = tensor("op_181_cast_fp16")]; + tensor qor_1_cast_fp16 = sub(x = var_180_cast_fp16, y = var_181_cast_fp16)[name = tensor("qor_1_cast_fp16")]; + tensor var_184_cast_fp16 = mul(x = var_146_cast_fp16, y = roti_1_cast_fp16)[name = tensor("op_184_cast_fp16")]; + tensor var_185_cast_fp16 = mul(x = var_154_cast_fp16, y = rotr_1_cast_fp16)[name = tensor("op_185_cast_fp16")]; + tensor qoi_1_cast_fp16 = add(x = var_184_cast_fp16, y = var_185_cast_fp16)[name = tensor("qoi_1_cast_fp16")]; + tensor var_188_cast_fp16 = mul(x = var_162_cast_fp16, y = rotr_1_cast_fp16)[name = tensor("op_188_cast_fp16")]; + tensor var_189_cast_fp16 = mul(x = var_170_cast_fp16, y = roti_1_cast_fp16)[name = tensor("op_189_cast_fp16")]; + tensor kor_1_cast_fp16 = sub(x = var_188_cast_fp16, y = var_189_cast_fp16)[name = tensor("kor_1_cast_fp16")]; + tensor var_192_cast_fp16 = mul(x = var_162_cast_fp16, y = roti_1_cast_fp16)[name = tensor("op_192_cast_fp16")]; + tensor var_193_cast_fp16 = mul(x = var_170_cast_fp16, y = rotr_1_cast_fp16)[name = tensor("op_193_cast_fp16")]; + tensor koi_1_cast_fp16 = add(x = var_192_cast_fp16, y = var_193_cast_fp16)[name = tensor("koi_1_cast_fp16")]; + tensor qo_1_axis_0 = const()[name = tensor("qo_1_axis_0"), val = tensor(-1)]; + tensor qo_1_cast_fp16 = stack(axis = qo_1_axis_0, values = (qor_1_cast_fp16, qoi_1_cast_fp16))[name = tensor("qo_1_cast_fp16")]; + tensor ko_1_axis_0 = const()[name = tensor("ko_1_axis_0"), val = tensor(-1)]; + tensor ko_1_cast_fp16 = stack(axis = ko_1_axis_0, values = (kor_1_cast_fp16, koi_1_cast_fp16))[name = tensor("ko_1_cast_fp16")]; + tensor var_226 = const()[name = tensor("op_226"), val = tensor([1, 256, 16, 64])]; + tensor q_3_cast_fp16 = reshape(shape = var_226, x = qo_1_cast_fp16)[name = tensor("q_3_cast_fp16")]; + tensor var_232 = const()[name = tensor("op_232"), val = tensor([1, 256, 16, 64])]; + tensor k_3_cast_fp16 = reshape(shape = var_232, x = ko_1_cast_fp16)[name = tensor("k_3_cast_fp16")]; + tensor var_249_promoted_to_fp16 = const()[name = tensor("op_249_promoted_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(6302720)))]; + tensor write_indices_float_1_cast_fp16 = add(x = var_121_cast_fp16, y = var_249_promoted_to_fp16)[name = tensor("write_indices_float_1_cast_fp16")]; + tensor _inversed_253_y_0_to_fp16 = const()[name = tensor("_inversed_253_y_0_to_fp16"), val = tensor(0x1p-9)]; + tensor _inversed_253_cast_fp16 = mul(x = write_indices_float_1_cast_fp16, y = _inversed_253_y_0_to_fp16)[name = tensor("_inversed_253_cast_fp16")]; + tensor var_254_cast_fp16 = floor(x = _inversed_253_cast_fp16)[name = tensor("op_254_cast_fp16")]; + tensor var_255_to_fp16 = const()[name = tensor("op_255_to_fp16"), val = tensor(0x1p+9)]; + tensor var_256_cast_fp16 = mul(x = var_254_cast_fp16, y = var_255_to_fp16)[name = tensor("op_256_cast_fp16")]; + tensor write_indices_float_3_cast_fp16 = sub(x = write_indices_float_1_cast_fp16, y = var_256_cast_fp16)[name = tensor("write_indices_float_3_cast_fp16")]; + tensor var_265 = const()[name = tensor("op_265"), val = tensor([1, 1])]; + tensor valid_len_to_fp16_dtype_0 = const()[name = tensor("valid_len_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor valid_len_to_fp16 = cast(dtype = valid_len_to_fp16_dtype_0, x = valid_len)[name = tensor("cast_157")]; + tensor var_266_cast_fp16 = reshape(shape = var_265, x = valid_len_to_fp16)[name = tensor("op_266_cast_fp16")]; + tensor token_valid_1_cast_fp16 = less(x = var_249_promoted_to_fp16, y = var_266_cast_fp16)[name = tensor("token_valid_1_cast_fp16")]; + tensor var_274_value_0_to_fp16 = const()[name = tensor("op_274_value_0_to_fp16"), val = tensor(0x1.ffp+8)]; + tensor var_274_cast_fp16 = fill_like(ref_tensor = write_indices_float_3_cast_fp16, value = var_274_value_0_to_fp16)[name = tensor("op_274_cast_fp16")]; + tensor write_indices_float_5_cast_fp16 = select(a = write_indices_float_3_cast_fp16, b = var_274_cast_fp16, cond = token_valid_1_cast_fp16)[name = tensor("write_indices_float_5_cast_fp16")]; + tensor var_280_dtype_0 = const()[name = tensor("op_280_dtype_0"), val = tensor("int32")]; + tensor var_285 = const()[name = tensor("op_285"), val = tensor([1, 256, 1, 1])]; + tensor write_indices_float_5_cast_fp16_to_int32 = cast(dtype = var_280_dtype_0, x = write_indices_float_5_cast_fp16)[name = tensor("cast_156")]; + tensor var_286 = reshape(shape = var_285, x = write_indices_float_5_cast_fp16_to_int32)[name = tensor("op_286")]; + tensor write_indices_1_reps_0 = const()[name = tensor("write_indices_1_reps_0"), val = tensor([1, 1, 16, 64])]; + tensor write_indices_1 = tile(reps = write_indices_1_reps_0, x = var_286)[name = tensor("write_indices_1")]; + tensor var_296_begin_0 = const()[name = tensor("op_296_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_296_end_0 = const()[name = tensor("op_296_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor var_296_end_mask_0 = const()[name = tensor("op_296_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_296_squeeze_mask_0 = const()[name = tensor("op_296_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor cache0_to_fp16_dtype_0 = const()[name = tensor("cache0_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor cache0_to_fp16 = cast(dtype = cache0_to_fp16_dtype_0, x = cache0)[name = tensor("cast_155")]; + tensor var_296_cast_fp16 = slice_by_index(begin = var_296_begin_0, end = var_296_end_0, end_mask = var_296_end_mask_0, squeeze_mask = var_296_squeeze_mask_0, x = cache0_to_fp16)[name = tensor("op_296_cast_fp16")]; + tensor var_298_axis_0 = const()[name = tensor("op_298_axis_0"), val = tensor(1)]; + tensor var_298_mode_0 = const()[name = tensor("op_298_mode_0"), val = tensor("update")]; + tensor var_298_validate_indices_0 = const()[name = tensor("op_298_validate_indices_0"), val = tensor(false)]; + tensor var_298_cast_fp16 = scatter_along_axis(axis = var_298_axis_0, data = var_296_cast_fp16, indices = write_indices_1, mode = var_298_mode_0, updates = k_3_cast_fp16, validate_indices = var_298_validate_indices_0)[name = tensor("op_298_cast_fp16")]; + tensor concat_2 = const()[name = tensor("concat_2"), val = tensor([0, 0, 0, 0, 0])]; + tensor concat_3 = const()[name = tensor("concat_3"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_1_internal_tensor_assign_1_stride_0 = const()[name = tensor("new_cache_1_internal_tensor_assign_1_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_1_internal_tensor_assign_1_begin_mask_0 = const()[name = tensor("new_cache_1_internal_tensor_assign_1_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_1_internal_tensor_assign_1_end_mask_0 = const()[name = tensor("new_cache_1_internal_tensor_assign_1_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_1_internal_tensor_assign_1_squeeze_mask_0 = const()[name = tensor("new_cache_1_internal_tensor_assign_1_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_10 = const()[name = tensor("shape_10"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_0 = const()[name = tensor("reduce_prod_0"), val = tensor(1048576)]; + tensor range_1d_0_start_0 = const()[name = tensor("range_1d_0_start_0"), val = tensor(0)]; + tensor range_1d_0_step_0 = const()[name = tensor("range_1d_0_step_0"), val = tensor(1)]; + tensor range_1d_0 = range_1d(end = reduce_prod_0, start = range_1d_0_start_0, step = range_1d_0_step_0)[name = tensor("range_1d_0")]; + tensor reshape_0 = reshape(shape = shape_10, x = range_1d_0)[name = tensor("reshape_0")]; + tensor slice_by_index_0 = slice_by_index(begin = concat_2, begin_mask = new_cache_1_internal_tensor_assign_1_begin_mask_0, end = concat_3, end_mask = new_cache_1_internal_tensor_assign_1_end_mask_0, squeeze_mask = new_cache_1_internal_tensor_assign_1_squeeze_mask_0, stride = new_cache_1_internal_tensor_assign_1_stride_0, x = reshape_0)[name = tensor("slice_by_index_0")]; + tensor reshape_1_shape_0 = const()[name = tensor("reshape_1_shape_0"), val = tensor([-1])]; + tensor reshape_1 = reshape(shape = reshape_1_shape_0, x = slice_by_index_0)[name = tensor("reshape_1")]; + tensor reshape_2_shape_0 = const()[name = tensor("reshape_2_shape_0"), val = tensor([-1])]; + tensor reshape_2_cast_fp16 = reshape(shape = reshape_2_shape_0, x = var_298_cast_fp16)[name = tensor("reshape_2_cast_fp16")]; + tensor reshape_3_shape_0 = const()[name = tensor("reshape_3_shape_0"), val = tensor([-1])]; + tensor reshape_3_cast_fp16 = reshape(shape = reshape_3_shape_0, x = cache0_to_fp16)[name = tensor("reshape_3_cast_fp16")]; + tensor scatter_0_mode_0 = const()[name = tensor("scatter_0_mode_0"), val = tensor("update")]; + tensor scatter_0_axis_0 = const()[name = tensor("scatter_0_axis_0"), val = tensor(0)]; + tensor scatter_0_validate_indices_0 = const()[name = tensor("scatter_0_validate_indices_0"), val = tensor(false)]; + tensor scatter_0_cast_fp16 = scatter(axis = scatter_0_axis_0, data = reshape_3_cast_fp16, indices = reshape_1, mode = scatter_0_mode_0, updates = reshape_2_cast_fp16, validate_indices = scatter_0_validate_indices_0)[name = tensor("scatter_0_cast_fp16")]; + tensor reshape_4_cast_fp16 = reshape(shape = shape_10, x = scatter_0_cast_fp16)[name = tensor("reshape_4_cast_fp16")]; + tensor var_306_begin_0 = const()[name = tensor("op_306_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor var_306_end_0 = const()[name = tensor("op_306_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor var_306_end_mask_0 = const()[name = tensor("op_306_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_306_squeeze_mask_0 = const()[name = tensor("op_306_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor var_306_cast_fp16 = slice_by_index(begin = var_306_begin_0, end = var_306_end_0, end_mask = var_306_end_mask_0, squeeze_mask = var_306_squeeze_mask_0, x = reshape_4_cast_fp16)[name = tensor("op_306_cast_fp16")]; + tensor var_308_axis_0 = const()[name = tensor("op_308_axis_0"), val = tensor(1)]; + tensor var_308_mode_0 = const()[name = tensor("op_308_mode_0"), val = tensor("update")]; + tensor var_308_validate_indices_0 = const()[name = tensor("op_308_validate_indices_0"), val = tensor(false)]; + tensor var_308_cast_fp16 = scatter_along_axis(axis = var_308_axis_0, data = var_306_cast_fp16, indices = write_indices_1, mode = var_308_mode_0, updates = v_1_cast_fp16, validate_indices = var_308_validate_indices_0)[name = tensor("op_308_cast_fp16")]; + tensor concat_4 = const()[name = tensor("concat_4"), val = tensor([1, 0, 0, 0, 0])]; + tensor concat_5 = const()[name = tensor("concat_5"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_1_internal_tensor_assign_2_stride_0 = const()[name = tensor("new_cache_1_internal_tensor_assign_2_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_1_internal_tensor_assign_2_begin_mask_0 = const()[name = tensor("new_cache_1_internal_tensor_assign_2_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_1_internal_tensor_assign_2_end_mask_0 = const()[name = tensor("new_cache_1_internal_tensor_assign_2_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_1_internal_tensor_assign_2_squeeze_mask_0 = const()[name = tensor("new_cache_1_internal_tensor_assign_2_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_11 = const()[name = tensor("shape_11"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_1 = const()[name = tensor("reduce_prod_1"), val = tensor(1048576)]; + tensor range_1d_1_start_0 = const()[name = tensor("range_1d_1_start_0"), val = tensor(0)]; + tensor range_1d_1_step_0 = const()[name = tensor("range_1d_1_step_0"), val = tensor(1)]; + tensor range_1d_1 = range_1d(end = reduce_prod_1, start = range_1d_1_start_0, step = range_1d_1_step_0)[name = tensor("range_1d_1")]; + tensor reshape_5 = reshape(shape = shape_11, x = range_1d_1)[name = tensor("reshape_5")]; + tensor slice_by_index_1 = slice_by_index(begin = concat_4, begin_mask = new_cache_1_internal_tensor_assign_2_begin_mask_0, end = concat_5, end_mask = new_cache_1_internal_tensor_assign_2_end_mask_0, squeeze_mask = new_cache_1_internal_tensor_assign_2_squeeze_mask_0, stride = new_cache_1_internal_tensor_assign_2_stride_0, x = reshape_5)[name = tensor("slice_by_index_1")]; + tensor reshape_6_shape_0 = const()[name = tensor("reshape_6_shape_0"), val = tensor([-1])]; + tensor reshape_6 = reshape(shape = reshape_6_shape_0, x = slice_by_index_1)[name = tensor("reshape_6")]; + tensor reshape_7_shape_0 = const()[name = tensor("reshape_7_shape_0"), val = tensor([-1])]; + tensor reshape_7_cast_fp16 = reshape(shape = reshape_7_shape_0, x = var_308_cast_fp16)[name = tensor("reshape_7_cast_fp16")]; + tensor reshape_8_shape_0 = const()[name = tensor("reshape_8_shape_0"), val = tensor([-1])]; + tensor reshape_8_cast_fp16 = reshape(shape = reshape_8_shape_0, x = reshape_4_cast_fp16)[name = tensor("reshape_8_cast_fp16")]; + tensor scatter_1_mode_0 = const()[name = tensor("scatter_1_mode_0"), val = tensor("update")]; + tensor scatter_1_axis_0 = const()[name = tensor("scatter_1_axis_0"), val = tensor(0)]; + tensor scatter_1_validate_indices_0 = const()[name = tensor("scatter_1_validate_indices_0"), val = tensor(false)]; + tensor scatter_1_cast_fp16 = scatter(axis = scatter_1_axis_0, data = reshape_8_cast_fp16, indices = reshape_6, mode = scatter_1_mode_0, updates = reshape_7_cast_fp16, validate_indices = scatter_1_validate_indices_0)[name = tensor("scatter_1_cast_fp16")]; + tensor reshape_9_cast_fp16 = reshape(shape = shape_11, x = scatter_1_cast_fp16)[name = tensor("reshape_9_cast_fp16")]; + tensor reshape_9_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("reshape_9_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor keys_1_begin_0 = const()[name = tensor("keys_1_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor keys_1_end_0 = const()[name = tensor("keys_1_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor keys_1_end_mask_0 = const()[name = tensor("keys_1_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor keys_1_squeeze_mask_0 = const()[name = tensor("keys_1_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor keys_1_cast_fp16 = slice_by_index(begin = keys_1_begin_0, end = keys_1_end_0, end_mask = keys_1_end_mask_0, squeeze_mask = keys_1_squeeze_mask_0, x = reshape_9_cast_fp16)[name = tensor("keys_1_cast_fp16")]; + tensor values_1_begin_0 = const()[name = tensor("values_1_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor values_1_end_0 = const()[name = tensor("values_1_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor values_1_end_mask_0 = const()[name = tensor("values_1_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor values_1_squeeze_mask_0 = const()[name = tensor("values_1_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor values_1_cast_fp16 = slice_by_index(begin = values_1_begin_0, end = values_1_end_0, end_mask = values_1_end_mask_0, squeeze_mask = values_1_squeeze_mask_0, x = reshape_9_cast_fp16)[name = tensor("values_1_cast_fp16")]; + tensor var_320_cast_fp16 = not_equal(x = keys_1_cast_fp16, y = keys_1_cast_fp16)[name = tensor("op_320_cast_fp16")]; + tensor var_326_to_fp16 = const()[name = tensor("op_326_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(6303296)))]; + tensor keys_3_cast_fp16 = select(a = var_326_to_fp16, b = keys_1_cast_fp16, cond = var_320_cast_fp16)[name = tensor("keys_3_cast_fp16")]; + tensor var_328_cast_fp16 = not_equal(x = values_1_cast_fp16, y = values_1_cast_fp16)[name = tensor("op_328_cast_fp16")]; + tensor values_3_cast_fp16 = select(a = var_326_to_fp16, b = values_1_cast_fp16, cond = var_328_cast_fp16)[name = tensor("values_3_cast_fp16")]; + tensor var_352 = const()[name = tensor("op_352"), val = tensor([0, 2, 1, 3])]; + tensor var_368 = const()[name = tensor("op_368"), val = tensor([1, 1, 1])]; + tensor var_369_cast_fp16 = reshape(shape = var_368, x = position0_to_fp16)[name = tensor("op_369_cast_fp16")]; + tensor q_offsets_1_promoted_to_fp16 = const()[name = tensor("q_offsets_1_promoted_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(7351936)))]; + tensor q_positions_1_cast_fp16 = add(x = var_369_cast_fp16, y = q_offsets_1_promoted_to_fp16)[name = tensor("q_positions_1_cast_fp16")]; + tensor var_391 = const()[name = tensor("op_391"), val = tensor([1, 1, 1])]; + tensor var_392_cast_fp16 = reshape(shape = var_391, x = valid_len_to_fp16)[name = tensor("op_392_cast_fp16")]; + tensor valid_end_1_cast_fp16 = add(x = var_369_cast_fp16, y = var_392_cast_fp16)[name = tensor("valid_end_1_cast_fp16")]; + tensor k_positions_1_promoted_to_fp16 = const()[name = tensor("k_positions_1_promoted_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(7352512)))]; + tensor valid_mask_1_cast_fp16 = less(x = k_positions_1_promoted_to_fp16, y = valid_end_1_cast_fp16)[name = tensor("valid_mask_1_cast_fp16")]; + tensor causal_mask_1_cast_fp16 = less_equal(x = k_positions_1_promoted_to_fp16, y = q_positions_1_cast_fp16)[name = tensor("causal_mask_1_cast_fp16")]; + tensor var_397 = logical_and(x = valid_mask_1_cast_fp16, y = causal_mask_1_cast_fp16)[name = tensor("op_397")]; + tensor attn_mask_1_axes_0 = const()[name = tensor("attn_mask_1_axes_0"), val = tensor([1])]; + tensor attn_mask_1 = expand_dims(axes = attn_mask_1_axes_0, x = var_397)[name = tensor("attn_mask_1")]; + tensor var_411_transpose_x_0 = const()[name = tensor("op_411_transpose_x_0"), val = tensor(false)]; + tensor var_411_transpose_y_0 = const()[name = tensor("op_411_transpose_y_0"), val = tensor(false)]; + tensor transpose_15_perm_0 = const()[name = tensor("transpose_15_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_16_perm_0 = const()[name = tensor("transpose_16_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_16 = transpose(perm = transpose_16_perm_0, x = keys_3_cast_fp16)[name = tensor("transpose_42")]; + tensor transpose_15 = transpose(perm = transpose_15_perm_0, x = q_3_cast_fp16)[name = tensor("transpose_43")]; + tensor var_411_cast_fp16 = matmul(transpose_x = var_411_transpose_x_0, transpose_y = var_411_transpose_y_0, x = transpose_15, y = transpose_16)[name = tensor("op_411_cast_fp16")]; + tensor var_405_to_fp16 = const()[name = tensor("op_405_to_fp16"), val = tensor([0x1p-3])]; + tensor attn_weights_1_cast_fp16 = mul(x = var_411_cast_fp16, y = var_405_to_fp16)[name = tensor("attn_weights_1_cast_fp16")]; + tensor var_413 = logical_not(x = attn_mask_1)[name = tensor("op_413")]; + tensor var_414_to_fp16 = const()[name = tensor("op_414_to_fp16"), val = tensor(-inf)]; + tensor attn_weights_3_cast_fp16 = select(a = var_414_to_fp16, b = attn_weights_1_cast_fp16, cond = var_413)[name = tensor("attn_weights_3_cast_fp16")]; + tensor var_416 = const()[name = tensor("op_416"), val = tensor(-1)]; + tensor attn_weights_5_cast_fp16 = softmax(axis = var_416, x = attn_weights_3_cast_fp16)[name = tensor("attn_weights_5_cast_fp16")]; + tensor attn_output_1_transpose_x_0 = const()[name = tensor("attn_output_1_transpose_x_0"), val = tensor(false)]; + tensor attn_output_1_transpose_y_0 = const()[name = tensor("attn_output_1_transpose_y_0"), val = tensor(false)]; + tensor values_5_cast_fp16 = transpose(perm = var_352, x = values_3_cast_fp16)[name = tensor("transpose_44")]; + tensor attn_output_1_cast_fp16 = matmul(transpose_x = attn_output_1_transpose_x_0, transpose_y = attn_output_1_transpose_y_0, x = attn_weights_5_cast_fp16, y = values_5_cast_fp16)[name = tensor("attn_output_1_cast_fp16")]; + tensor var_424 = const()[name = tensor("op_424"), val = tensor([0, 2, 1, 3])]; + tensor var_429 = const()[name = tensor("op_429"), val = tensor([1, 256, 1024])]; + tensor var_425_cast_fp16 = transpose(perm = var_424, x = attn_output_1_cast_fp16)[name = tensor("transpose_41")]; + tensor input_5_cast_fp16 = reshape(shape = var_429, x = var_425_cast_fp16)[name = tensor("input_5_cast_fp16")]; + tensor attn0_out_proj_weight_to_fp16 = const()[name = tensor("attn0_out_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(7353600)))]; + tensor linear_1_bias_0_to_fp16 = const()[name = tensor("linear_1_bias_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9450816)))]; + tensor linear_1_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = attn0_out_proj_weight_to_fp16, x = input_5_cast_fp16)[name = tensor("linear_1_cast_fp16")]; + tensor var_438_cast_fp16 = add(x = position0_to_fp16, y = valid_len_to_fp16)[name = tensor("op_438_cast_fp16")]; + tensor var_438_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_438_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor input_7_cast_fp16 = add(x = conditioning_to_fp16, y = linear_1_cast_fp16)[name = tensor("input_7_cast_fp16")]; + tensor input_9_axes_0 = const()[name = tensor("input_9_axes_0"), val = tensor([-1])]; + tensor norm0_2_weight_to_fp16 = const()[name = tensor("norm0_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9452928)))]; + tensor norm0_2_bias_to_fp16 = const()[name = tensor("norm0_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9455040)))]; + tensor var_442_to_fp16 = const()[name = tensor("op_442_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_9_cast_fp16 = layer_norm(axes = input_9_axes_0, beta = norm0_2_bias_to_fp16, epsilon = var_442_to_fp16, gamma = norm0_2_weight_to_fp16, x = input_7_cast_fp16)[name = tensor("input_9_cast_fp16")]; + tensor linear0_1_weight_to_fp16 = const()[name = tensor("linear0_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9457152)))]; + tensor linear_2_bias_0_to_fp16 = const()[name = tensor("linear_2_bias_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(17845824)))]; + tensor linear_2_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = linear0_1_weight_to_fp16, x = input_9_cast_fp16)[name = tensor("linear_2_cast_fp16")]; + tensor input_11_mode_0 = const()[name = tensor("input_11_mode_0"), val = tensor("EXACT")]; + tensor input_11_cast_fp16 = gelu(mode = input_11_mode_0, x = linear_2_cast_fp16)[name = tensor("input_11_cast_fp16")]; + tensor linear0_2_weight_to_fp16 = const()[name = tensor("linear0_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(17854080)))]; + tensor linear_3_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = linear0_2_weight_to_fp16, x = input_11_cast_fp16)[name = tensor("linear_3_cast_fp16")]; + tensor input_13_cast_fp16 = add(x = input_7_cast_fp16, y = linear_3_cast_fp16)[name = tensor("input_13_cast_fp16")]; + tensor input_15_axes_0 = const()[name = tensor("input_15_axes_0"), val = tensor([-1])]; + tensor norm1_1_weight_to_fp16 = const()[name = tensor("norm1_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(26242752)))]; + tensor norm1_1_bias_to_fp16 = const()[name = tensor("norm1_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(26244864)))]; + tensor var_459_to_fp16 = const()[name = tensor("op_459_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_15_cast_fp16 = layer_norm(axes = input_15_axes_0, beta = norm1_1_bias_to_fp16, epsilon = var_459_to_fp16, gamma = norm1_1_weight_to_fp16, x = input_13_cast_fp16)[name = tensor("input_15_cast_fp16")]; + tensor attn1_in_proj_weight_to_fp16 = const()[name = tensor("attn1_in_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(26246976)))]; + tensor linear_4_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = attn1_in_proj_weight_to_fp16, x = input_15_cast_fp16)[name = tensor("linear_4_cast_fp16")]; + tensor var_473 = const()[name = tensor("op_473"), val = tensor([1, 256, 3, 16, 64])]; + tensor qkv_3_cast_fp16 = reshape(shape = var_473, x = linear_4_cast_fp16)[name = tensor("qkv_3_cast_fp16")]; + tensor q_7_begin_0 = const()[name = tensor("q_7_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor q_7_end_0 = const()[name = tensor("q_7_end_0"), val = tensor([1, 256, 1, 16, 64])]; + tensor q_7_end_mask_0 = const()[name = tensor("q_7_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor q_7_squeeze_mask_0 = const()[name = tensor("q_7_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor q_7_cast_fp16 = slice_by_index(begin = q_7_begin_0, end = q_7_end_0, end_mask = q_7_end_mask_0, squeeze_mask = q_7_squeeze_mask_0, x = qkv_3_cast_fp16)[name = tensor("q_7_cast_fp16")]; + tensor k_5_begin_0 = const()[name = tensor("k_5_begin_0"), val = tensor([0, 0, 1, 0, 0])]; + tensor k_5_end_0 = const()[name = tensor("k_5_end_0"), val = tensor([1, 256, 2, 16, 64])]; + tensor k_5_end_mask_0 = const()[name = tensor("k_5_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor k_5_squeeze_mask_0 = const()[name = tensor("k_5_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor k_5_cast_fp16 = slice_by_index(begin = k_5_begin_0, end = k_5_end_0, end_mask = k_5_end_mask_0, squeeze_mask = k_5_squeeze_mask_0, x = qkv_3_cast_fp16)[name = tensor("k_5_cast_fp16")]; + tensor v_3_begin_0 = const()[name = tensor("v_3_begin_0"), val = tensor([0, 0, 2, 0, 0])]; + tensor v_3_end_0 = const()[name = tensor("v_3_end_0"), val = tensor([1, 256, 3, 16, 64])]; + tensor v_3_end_mask_0 = const()[name = tensor("v_3_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor v_3_squeeze_mask_0 = const()[name = tensor("v_3_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor v_3_cast_fp16 = slice_by_index(begin = v_3_begin_0, end = v_3_end_0, end_mask = v_3_end_mask_0, squeeze_mask = v_3_squeeze_mask_0, x = qkv_3_cast_fp16)[name = tensor("v_3_cast_fp16")]; + tensor var_531 = const()[name = tensor("op_531"), val = tensor([1, 1])]; + tensor position1_to_fp16_dtype_0 = const()[name = tensor("position1_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor position1_to_fp16 = cast(dtype = position1_to_fp16_dtype_0, x = position1)[name = tensor("cast_152")]; + tensor var_532_cast_fp16 = reshape(shape = var_531, x = position1_to_fp16)[name = tensor("op_532_cast_fp16")]; + tensor ts_9_cast_fp16 = add(x = ts_1_promoted_to_fp16, y = var_532_cast_fp16)[name = tensor("ts_9_cast_fp16")]; + tensor var_539 = const()[name = tensor("op_539"), val = tensor([1, 256, 1, 1])]; + tensor ts_11_cast_fp16 = reshape(shape = var_539, x = ts_9_cast_fp16)[name = tensor("ts_11_cast_fp16")]; + tensor var_546 = const()[name = tensor("op_546"), val = tensor([1, 256, 16, 32, 2])]; + tensor q_complex_3_cast_fp16 = reshape(shape = var_546, x = q_7_cast_fp16)[name = tensor("q_complex_3_cast_fp16")]; + tensor var_553 = const()[name = tensor("op_553"), val = tensor([1, 256, 16, 32, 2])]; + tensor k_complex_3_cast_fp16 = reshape(shape = var_553, x = k_5_cast_fp16)[name = tensor("k_complex_3_cast_fp16")]; + tensor var_557_begin_0 = const()[name = tensor("op_557_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_557_end_0 = const()[name = tensor("op_557_end_0"), val = tensor([1, 256, 16, 32, 1])]; + tensor var_557_end_mask_0 = const()[name = tensor("op_557_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_557_squeeze_mask_0 = const()[name = tensor("op_557_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_557_cast_fp16 = slice_by_index(begin = var_557_begin_0, end = var_557_end_0, end_mask = var_557_end_mask_0, squeeze_mask = var_557_squeeze_mask_0, x = q_complex_3_cast_fp16)[name = tensor("op_557_cast_fp16")]; + tensor var_565_begin_0 = const()[name = tensor("op_565_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_565_end_0 = const()[name = tensor("op_565_end_0"), val = tensor([1, 256, 16, 32, 2])]; + tensor var_565_end_mask_0 = const()[name = tensor("op_565_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_565_squeeze_mask_0 = const()[name = tensor("op_565_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_565_cast_fp16 = slice_by_index(begin = var_565_begin_0, end = var_565_end_0, end_mask = var_565_end_mask_0, squeeze_mask = var_565_squeeze_mask_0, x = q_complex_3_cast_fp16)[name = tensor("op_565_cast_fp16")]; + tensor var_573_begin_0 = const()[name = tensor("op_573_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_573_end_0 = const()[name = tensor("op_573_end_0"), val = tensor([1, 256, 16, 32, 1])]; + tensor var_573_end_mask_0 = const()[name = tensor("op_573_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_573_squeeze_mask_0 = const()[name = tensor("op_573_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_573_cast_fp16 = slice_by_index(begin = var_573_begin_0, end = var_573_end_0, end_mask = var_573_end_mask_0, squeeze_mask = var_573_squeeze_mask_0, x = k_complex_3_cast_fp16)[name = tensor("op_573_cast_fp16")]; + tensor var_581_begin_0 = const()[name = tensor("op_581_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_581_end_0 = const()[name = tensor("op_581_end_0"), val = tensor([1, 256, 16, 32, 2])]; + tensor var_581_end_mask_0 = const()[name = tensor("op_581_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_581_squeeze_mask_0 = const()[name = tensor("op_581_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_581_cast_fp16 = slice_by_index(begin = var_581_begin_0, end = var_581_end_0, end_mask = var_581_end_mask_0, squeeze_mask = var_581_squeeze_mask_0, x = k_complex_3_cast_fp16)[name = tensor("op_581_cast_fp16")]; + tensor freqs_3_to_fp16 = const()[name = tensor("freqs_3_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(32538496)))]; + tensor var_587_cast_fp16 = mul(x = freqs_3_to_fp16, y = ts_11_cast_fp16)[name = tensor("op_587_cast_fp16")]; + tensor rotr_3_cast_fp16 = cos(x = var_587_cast_fp16)[name = tensor("rotr_3_cast_fp16")]; + tensor roti_3_cast_fp16 = sin(x = var_587_cast_fp16)[name = tensor("roti_3_cast_fp16")]; + tensor var_591_cast_fp16 = mul(x = var_557_cast_fp16, y = rotr_3_cast_fp16)[name = tensor("op_591_cast_fp16")]; + tensor var_592_cast_fp16 = mul(x = var_565_cast_fp16, y = roti_3_cast_fp16)[name = tensor("op_592_cast_fp16")]; + tensor qor_5_cast_fp16 = sub(x = var_591_cast_fp16, y = var_592_cast_fp16)[name = tensor("qor_5_cast_fp16")]; + tensor var_595_cast_fp16 = mul(x = var_557_cast_fp16, y = roti_3_cast_fp16)[name = tensor("op_595_cast_fp16")]; + tensor var_596_cast_fp16 = mul(x = var_565_cast_fp16, y = rotr_3_cast_fp16)[name = tensor("op_596_cast_fp16")]; + tensor qoi_5_cast_fp16 = add(x = var_595_cast_fp16, y = var_596_cast_fp16)[name = tensor("qoi_5_cast_fp16")]; + tensor var_599_cast_fp16 = mul(x = var_573_cast_fp16, y = rotr_3_cast_fp16)[name = tensor("op_599_cast_fp16")]; + tensor var_600_cast_fp16 = mul(x = var_581_cast_fp16, y = roti_3_cast_fp16)[name = tensor("op_600_cast_fp16")]; + tensor kor_5_cast_fp16 = sub(x = var_599_cast_fp16, y = var_600_cast_fp16)[name = tensor("kor_5_cast_fp16")]; + tensor var_603_cast_fp16 = mul(x = var_573_cast_fp16, y = roti_3_cast_fp16)[name = tensor("op_603_cast_fp16")]; + tensor var_604_cast_fp16 = mul(x = var_581_cast_fp16, y = rotr_3_cast_fp16)[name = tensor("op_604_cast_fp16")]; + tensor koi_5_cast_fp16 = add(x = var_603_cast_fp16, y = var_604_cast_fp16)[name = tensor("koi_5_cast_fp16")]; + tensor qo_3_axis_0 = const()[name = tensor("qo_3_axis_0"), val = tensor(-1)]; + tensor qo_3_cast_fp16 = stack(axis = qo_3_axis_0, values = (qor_5_cast_fp16, qoi_5_cast_fp16))[name = tensor("qo_3_cast_fp16")]; + tensor ko_3_axis_0 = const()[name = tensor("ko_3_axis_0"), val = tensor(-1)]; + tensor ko_3_cast_fp16 = stack(axis = ko_3_axis_0, values = (kor_5_cast_fp16, koi_5_cast_fp16))[name = tensor("ko_3_cast_fp16")]; + tensor var_637 = const()[name = tensor("op_637"), val = tensor([1, 256, 16, 64])]; + tensor q_9_cast_fp16 = reshape(shape = var_637, x = qo_3_cast_fp16)[name = tensor("q_9_cast_fp16")]; + tensor var_643 = const()[name = tensor("op_643"), val = tensor([1, 256, 16, 64])]; + tensor k_7_cast_fp16 = reshape(shape = var_643, x = ko_3_cast_fp16)[name = tensor("k_7_cast_fp16")]; + tensor write_indices_float_7_cast_fp16 = add(x = var_532_cast_fp16, y = var_249_promoted_to_fp16)[name = tensor("write_indices_float_7_cast_fp16")]; + tensor _inversed_664_y_0_to_fp16 = const()[name = tensor("_inversed_664_y_0_to_fp16"), val = tensor(0x1p-9)]; + tensor _inversed_664_cast_fp16 = mul(x = write_indices_float_7_cast_fp16, y = _inversed_664_y_0_to_fp16)[name = tensor("_inversed_664_cast_fp16")]; + tensor var_665_cast_fp16 = floor(x = _inversed_664_cast_fp16)[name = tensor("op_665_cast_fp16")]; + tensor var_666_to_fp16 = const()[name = tensor("op_666_to_fp16"), val = tensor(0x1p+9)]; + tensor var_667_cast_fp16 = mul(x = var_665_cast_fp16, y = var_666_to_fp16)[name = tensor("op_667_cast_fp16")]; + tensor write_indices_float_9_cast_fp16 = sub(x = write_indices_float_7_cast_fp16, y = var_667_cast_fp16)[name = tensor("write_indices_float_9_cast_fp16")]; + tensor var_685_value_0_to_fp16 = const()[name = tensor("op_685_value_0_to_fp16"), val = tensor(0x1.ffp+8)]; + tensor var_685_cast_fp16 = fill_like(ref_tensor = write_indices_float_9_cast_fp16, value = var_685_value_0_to_fp16)[name = tensor("op_685_cast_fp16")]; + tensor write_indices_float_11_cast_fp16 = select(a = write_indices_float_9_cast_fp16, b = var_685_cast_fp16, cond = token_valid_1_cast_fp16)[name = tensor("write_indices_float_11_cast_fp16")]; + tensor var_691_dtype_0 = const()[name = tensor("op_691_dtype_0"), val = tensor("int32")]; + tensor var_696 = const()[name = tensor("op_696"), val = tensor([1, 256, 1, 1])]; + tensor write_indices_float_11_cast_fp16_to_int32 = cast(dtype = var_691_dtype_0, x = write_indices_float_11_cast_fp16)[name = tensor("cast_151")]; + tensor var_697 = reshape(shape = var_696, x = write_indices_float_11_cast_fp16_to_int32)[name = tensor("op_697")]; + tensor write_indices_3_reps_0 = const()[name = tensor("write_indices_3_reps_0"), val = tensor([1, 1, 16, 64])]; + tensor write_indices_3 = tile(reps = write_indices_3_reps_0, x = var_697)[name = tensor("write_indices_3")]; + tensor var_707_begin_0 = const()[name = tensor("op_707_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_707_end_0 = const()[name = tensor("op_707_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor var_707_end_mask_0 = const()[name = tensor("op_707_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_707_squeeze_mask_0 = const()[name = tensor("op_707_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor cache1_to_fp16_dtype_0 = const()[name = tensor("cache1_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor cache1_to_fp16 = cast(dtype = cache1_to_fp16_dtype_0, x = cache1)[name = tensor("cast_150")]; + tensor var_707_cast_fp16 = slice_by_index(begin = var_707_begin_0, end = var_707_end_0, end_mask = var_707_end_mask_0, squeeze_mask = var_707_squeeze_mask_0, x = cache1_to_fp16)[name = tensor("op_707_cast_fp16")]; + tensor var_709_axis_0 = const()[name = tensor("op_709_axis_0"), val = tensor(1)]; + tensor var_709_mode_0 = const()[name = tensor("op_709_mode_0"), val = tensor("update")]; + tensor var_709_validate_indices_0 = const()[name = tensor("op_709_validate_indices_0"), val = tensor(false)]; + tensor var_709_cast_fp16 = scatter_along_axis(axis = var_709_axis_0, data = var_707_cast_fp16, indices = write_indices_3, mode = var_709_mode_0, updates = k_7_cast_fp16, validate_indices = var_709_validate_indices_0)[name = tensor("op_709_cast_fp16")]; + tensor concat_10 = const()[name = tensor("concat_10"), val = tensor([0, 0, 0, 0, 0])]; + tensor concat_11 = const()[name = tensor("concat_11"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_3_internal_tensor_assign_1_stride_0 = const()[name = tensor("new_cache_3_internal_tensor_assign_1_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_3_internal_tensor_assign_1_begin_mask_0 = const()[name = tensor("new_cache_3_internal_tensor_assign_1_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_3_internal_tensor_assign_1_end_mask_0 = const()[name = tensor("new_cache_3_internal_tensor_assign_1_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_3_internal_tensor_assign_1_squeeze_mask_0 = const()[name = tensor("new_cache_3_internal_tensor_assign_1_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_12 = const()[name = tensor("shape_12"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_2 = const()[name = tensor("reduce_prod_2"), val = tensor(1048576)]; + tensor range_1d_2_start_0 = const()[name = tensor("range_1d_2_start_0"), val = tensor(0)]; + tensor range_1d_2_step_0 = const()[name = tensor("range_1d_2_step_0"), val = tensor(1)]; + tensor range_1d_2 = range_1d(end = reduce_prod_2, start = range_1d_2_start_0, step = range_1d_2_step_0)[name = tensor("range_1d_2")]; + tensor reshape_10 = reshape(shape = shape_12, x = range_1d_2)[name = tensor("reshape_10")]; + tensor slice_by_index_2 = slice_by_index(begin = concat_10, begin_mask = new_cache_3_internal_tensor_assign_1_begin_mask_0, end = concat_11, end_mask = new_cache_3_internal_tensor_assign_1_end_mask_0, squeeze_mask = new_cache_3_internal_tensor_assign_1_squeeze_mask_0, stride = new_cache_3_internal_tensor_assign_1_stride_0, x = reshape_10)[name = tensor("slice_by_index_2")]; + tensor reshape_11_shape_0 = const()[name = tensor("reshape_11_shape_0"), val = tensor([-1])]; + tensor reshape_11 = reshape(shape = reshape_11_shape_0, x = slice_by_index_2)[name = tensor("reshape_11")]; + tensor reshape_12_shape_0 = const()[name = tensor("reshape_12_shape_0"), val = tensor([-1])]; + tensor reshape_12_cast_fp16 = reshape(shape = reshape_12_shape_0, x = var_709_cast_fp16)[name = tensor("reshape_12_cast_fp16")]; + tensor reshape_13_shape_0 = const()[name = tensor("reshape_13_shape_0"), val = tensor([-1])]; + tensor reshape_13_cast_fp16 = reshape(shape = reshape_13_shape_0, x = cache1_to_fp16)[name = tensor("reshape_13_cast_fp16")]; + tensor scatter_2_mode_0 = const()[name = tensor("scatter_2_mode_0"), val = tensor("update")]; + tensor scatter_2_axis_0 = const()[name = tensor("scatter_2_axis_0"), val = tensor(0)]; + tensor scatter_2_validate_indices_0 = const()[name = tensor("scatter_2_validate_indices_0"), val = tensor(false)]; + tensor scatter_2_cast_fp16 = scatter(axis = scatter_2_axis_0, data = reshape_13_cast_fp16, indices = reshape_11, mode = scatter_2_mode_0, updates = reshape_12_cast_fp16, validate_indices = scatter_2_validate_indices_0)[name = tensor("scatter_2_cast_fp16")]; + tensor reshape_14_cast_fp16 = reshape(shape = shape_12, x = scatter_2_cast_fp16)[name = tensor("reshape_14_cast_fp16")]; + tensor var_717_begin_0 = const()[name = tensor("op_717_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor var_717_end_0 = const()[name = tensor("op_717_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor var_717_end_mask_0 = const()[name = tensor("op_717_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_717_squeeze_mask_0 = const()[name = tensor("op_717_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor var_717_cast_fp16 = slice_by_index(begin = var_717_begin_0, end = var_717_end_0, end_mask = var_717_end_mask_0, squeeze_mask = var_717_squeeze_mask_0, x = reshape_14_cast_fp16)[name = tensor("op_717_cast_fp16")]; + tensor var_719_axis_0 = const()[name = tensor("op_719_axis_0"), val = tensor(1)]; + tensor var_719_mode_0 = const()[name = tensor("op_719_mode_0"), val = tensor("update")]; + tensor var_719_validate_indices_0 = const()[name = tensor("op_719_validate_indices_0"), val = tensor(false)]; + tensor var_719_cast_fp16 = scatter_along_axis(axis = var_719_axis_0, data = var_717_cast_fp16, indices = write_indices_3, mode = var_719_mode_0, updates = v_3_cast_fp16, validate_indices = var_719_validate_indices_0)[name = tensor("op_719_cast_fp16")]; + tensor concat_12 = const()[name = tensor("concat_12"), val = tensor([1, 0, 0, 0, 0])]; + tensor concat_13 = const()[name = tensor("concat_13"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_3_internal_tensor_assign_2_stride_0 = const()[name = tensor("new_cache_3_internal_tensor_assign_2_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_3_internal_tensor_assign_2_begin_mask_0 = const()[name = tensor("new_cache_3_internal_tensor_assign_2_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_3_internal_tensor_assign_2_end_mask_0 = const()[name = tensor("new_cache_3_internal_tensor_assign_2_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_3_internal_tensor_assign_2_squeeze_mask_0 = const()[name = tensor("new_cache_3_internal_tensor_assign_2_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_13 = const()[name = tensor("shape_13"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_3 = const()[name = tensor("reduce_prod_3"), val = tensor(1048576)]; + tensor range_1d_3_start_0 = const()[name = tensor("range_1d_3_start_0"), val = tensor(0)]; + tensor range_1d_3_step_0 = const()[name = tensor("range_1d_3_step_0"), val = tensor(1)]; + tensor range_1d_3 = range_1d(end = reduce_prod_3, start = range_1d_3_start_0, step = range_1d_3_step_0)[name = tensor("range_1d_3")]; + tensor reshape_15 = reshape(shape = shape_13, x = range_1d_3)[name = tensor("reshape_15")]; + tensor slice_by_index_3 = slice_by_index(begin = concat_12, begin_mask = new_cache_3_internal_tensor_assign_2_begin_mask_0, end = concat_13, end_mask = new_cache_3_internal_tensor_assign_2_end_mask_0, squeeze_mask = new_cache_3_internal_tensor_assign_2_squeeze_mask_0, stride = new_cache_3_internal_tensor_assign_2_stride_0, x = reshape_15)[name = tensor("slice_by_index_3")]; + tensor reshape_16_shape_0 = const()[name = tensor("reshape_16_shape_0"), val = tensor([-1])]; + tensor reshape_16 = reshape(shape = reshape_16_shape_0, x = slice_by_index_3)[name = tensor("reshape_16")]; + tensor reshape_17_shape_0 = const()[name = tensor("reshape_17_shape_0"), val = tensor([-1])]; + tensor reshape_17_cast_fp16 = reshape(shape = reshape_17_shape_0, x = var_719_cast_fp16)[name = tensor("reshape_17_cast_fp16")]; + tensor reshape_18_shape_0 = const()[name = tensor("reshape_18_shape_0"), val = tensor([-1])]; + tensor reshape_18_cast_fp16 = reshape(shape = reshape_18_shape_0, x = reshape_14_cast_fp16)[name = tensor("reshape_18_cast_fp16")]; + tensor scatter_3_mode_0 = const()[name = tensor("scatter_3_mode_0"), val = tensor("update")]; + tensor scatter_3_axis_0 = const()[name = tensor("scatter_3_axis_0"), val = tensor(0)]; + tensor scatter_3_validate_indices_0 = const()[name = tensor("scatter_3_validate_indices_0"), val = tensor(false)]; + tensor scatter_3_cast_fp16 = scatter(axis = scatter_3_axis_0, data = reshape_18_cast_fp16, indices = reshape_16, mode = scatter_3_mode_0, updates = reshape_17_cast_fp16, validate_indices = scatter_3_validate_indices_0)[name = tensor("scatter_3_cast_fp16")]; + tensor reshape_19_cast_fp16 = reshape(shape = shape_13, x = scatter_3_cast_fp16)[name = tensor("reshape_19_cast_fp16")]; + tensor reshape_19_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("reshape_19_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor keys_7_begin_0 = const()[name = tensor("keys_7_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor keys_7_end_0 = const()[name = tensor("keys_7_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor keys_7_end_mask_0 = const()[name = tensor("keys_7_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor keys_7_squeeze_mask_0 = const()[name = tensor("keys_7_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor keys_7_cast_fp16 = slice_by_index(begin = keys_7_begin_0, end = keys_7_end_0, end_mask = keys_7_end_mask_0, squeeze_mask = keys_7_squeeze_mask_0, x = reshape_19_cast_fp16)[name = tensor("keys_7_cast_fp16")]; + tensor values_7_begin_0 = const()[name = tensor("values_7_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor values_7_end_0 = const()[name = tensor("values_7_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor values_7_end_mask_0 = const()[name = tensor("values_7_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor values_7_squeeze_mask_0 = const()[name = tensor("values_7_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor values_7_cast_fp16 = slice_by_index(begin = values_7_begin_0, end = values_7_end_0, end_mask = values_7_end_mask_0, squeeze_mask = values_7_squeeze_mask_0, x = reshape_19_cast_fp16)[name = tensor("values_7_cast_fp16")]; + tensor var_731_cast_fp16 = not_equal(x = keys_7_cast_fp16, y = keys_7_cast_fp16)[name = tensor("op_731_cast_fp16")]; + tensor keys_9_cast_fp16 = select(a = var_326_to_fp16, b = keys_7_cast_fp16, cond = var_731_cast_fp16)[name = tensor("keys_9_cast_fp16")]; + tensor var_739_cast_fp16 = not_equal(x = values_7_cast_fp16, y = values_7_cast_fp16)[name = tensor("op_739_cast_fp16")]; + tensor values_9_cast_fp16 = select(a = var_326_to_fp16, b = values_7_cast_fp16, cond = var_739_cast_fp16)[name = tensor("values_9_cast_fp16")]; + tensor var_763 = const()[name = tensor("op_763"), val = tensor([0, 2, 1, 3])]; + tensor var_779 = const()[name = tensor("op_779"), val = tensor([1, 1, 1])]; + tensor var_780_cast_fp16 = reshape(shape = var_779, x = position1_to_fp16)[name = tensor("op_780_cast_fp16")]; + tensor q_positions_3_cast_fp16 = add(x = var_780_cast_fp16, y = q_offsets_1_promoted_to_fp16)[name = tensor("q_positions_3_cast_fp16")]; + tensor valid_end_3_cast_fp16 = add(x = var_780_cast_fp16, y = var_392_cast_fp16)[name = tensor("valid_end_3_cast_fp16")]; + tensor valid_mask_3_cast_fp16 = less(x = k_positions_1_promoted_to_fp16, y = valid_end_3_cast_fp16)[name = tensor("valid_mask_3_cast_fp16")]; + tensor causal_mask_3_cast_fp16 = less_equal(x = k_positions_1_promoted_to_fp16, y = q_positions_3_cast_fp16)[name = tensor("causal_mask_3_cast_fp16")]; + tensor var_808 = logical_and(x = valid_mask_3_cast_fp16, y = causal_mask_3_cast_fp16)[name = tensor("op_808")]; + tensor attn_mask_3_axes_0 = const()[name = tensor("attn_mask_3_axes_0"), val = tensor([1])]; + tensor attn_mask_3 = expand_dims(axes = attn_mask_3_axes_0, x = var_808)[name = tensor("attn_mask_3")]; + tensor var_822_transpose_x_0 = const()[name = tensor("op_822_transpose_x_0"), val = tensor(false)]; + tensor var_822_transpose_y_0 = const()[name = tensor("op_822_transpose_y_0"), val = tensor(false)]; + tensor transpose_17_perm_0 = const()[name = tensor("transpose_17_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_18_perm_0 = const()[name = tensor("transpose_18_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_18 = transpose(perm = transpose_18_perm_0, x = keys_9_cast_fp16)[name = tensor("transpose_38")]; + tensor transpose_17 = transpose(perm = transpose_17_perm_0, x = q_9_cast_fp16)[name = tensor("transpose_39")]; + tensor var_822_cast_fp16 = matmul(transpose_x = var_822_transpose_x_0, transpose_y = var_822_transpose_y_0, x = transpose_17, y = transpose_18)[name = tensor("op_822_cast_fp16")]; + tensor var_816_to_fp16 = const()[name = tensor("op_816_to_fp16"), val = tensor([0x1p-3])]; + tensor attn_weights_7_cast_fp16 = mul(x = var_822_cast_fp16, y = var_816_to_fp16)[name = tensor("attn_weights_7_cast_fp16")]; + tensor var_824 = logical_not(x = attn_mask_3)[name = tensor("op_824")]; + tensor var_825_to_fp16 = const()[name = tensor("op_825_to_fp16"), val = tensor(-inf)]; + tensor attn_weights_9_cast_fp16 = select(a = var_825_to_fp16, b = attn_weights_7_cast_fp16, cond = var_824)[name = tensor("attn_weights_9_cast_fp16")]; + tensor var_827 = const()[name = tensor("op_827"), val = tensor(-1)]; + tensor attn_weights_11_cast_fp16 = softmax(axis = var_827, x = attn_weights_9_cast_fp16)[name = tensor("attn_weights_11_cast_fp16")]; + tensor attn_output_3_transpose_x_0 = const()[name = tensor("attn_output_3_transpose_x_0"), val = tensor(false)]; + tensor attn_output_3_transpose_y_0 = const()[name = tensor("attn_output_3_transpose_y_0"), val = tensor(false)]; + tensor values_11_cast_fp16 = transpose(perm = var_763, x = values_9_cast_fp16)[name = tensor("transpose_40")]; + tensor attn_output_3_cast_fp16 = matmul(transpose_x = attn_output_3_transpose_x_0, transpose_y = attn_output_3_transpose_y_0, x = attn_weights_11_cast_fp16, y = values_11_cast_fp16)[name = tensor("attn_output_3_cast_fp16")]; + tensor var_835 = const()[name = tensor("op_835"), val = tensor([0, 2, 1, 3])]; + tensor var_840 = const()[name = tensor("op_840"), val = tensor([1, 256, 1024])]; + tensor var_836_cast_fp16 = transpose(perm = var_835, x = attn_output_3_cast_fp16)[name = tensor("transpose_37")]; + tensor input_17_cast_fp16 = reshape(shape = var_840, x = var_836_cast_fp16)[name = tensor("input_17_cast_fp16")]; + tensor attn1_out_proj_weight_to_fp16 = const()[name = tensor("attn1_out_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(32538624)))]; + tensor linear_5_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = attn1_out_proj_weight_to_fp16, x = input_17_cast_fp16)[name = tensor("linear_5_cast_fp16")]; + tensor var_849_cast_fp16 = add(x = position1_to_fp16, y = valid_len_to_fp16)[name = tensor("op_849_cast_fp16")]; + tensor var_849_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_849_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor input_19_cast_fp16 = add(x = input_13_cast_fp16, y = linear_5_cast_fp16)[name = tensor("input_19_cast_fp16")]; + tensor input_21_axes_0 = const()[name = tensor("input_21_axes_0"), val = tensor([-1])]; + tensor norm1_2_weight_to_fp16 = const()[name = tensor("norm1_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(34635840)))]; + tensor norm1_2_bias_to_fp16 = const()[name = tensor("norm1_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(34637952)))]; + tensor var_853_to_fp16 = const()[name = tensor("op_853_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_21_cast_fp16 = layer_norm(axes = input_21_axes_0, beta = norm1_2_bias_to_fp16, epsilon = var_853_to_fp16, gamma = norm1_2_weight_to_fp16, x = input_19_cast_fp16)[name = tensor("input_21_cast_fp16")]; + tensor linear1_1_weight_to_fp16 = const()[name = tensor("linear1_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(34640064)))]; + tensor linear_6_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = linear1_1_weight_to_fp16, x = input_21_cast_fp16)[name = tensor("linear_6_cast_fp16")]; + tensor input_23_mode_0 = const()[name = tensor("input_23_mode_0"), val = tensor("EXACT")]; + tensor input_23_cast_fp16 = gelu(mode = input_23_mode_0, x = linear_6_cast_fp16)[name = tensor("input_23_cast_fp16")]; + tensor linear1_2_weight_to_fp16 = const()[name = tensor("linear1_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(43028736)))]; + tensor linear_7_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = linear1_2_weight_to_fp16, x = input_23_cast_fp16)[name = tensor("linear_7_cast_fp16")]; + tensor input_25_cast_fp16 = add(x = input_19_cast_fp16, y = linear_7_cast_fp16)[name = tensor("input_25_cast_fp16")]; + tensor input_27_axes_0 = const()[name = tensor("input_27_axes_0"), val = tensor([-1])]; + tensor norm2_1_weight_to_fp16 = const()[name = tensor("norm2_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(51417408)))]; + tensor norm2_1_bias_to_fp16 = const()[name = tensor("norm2_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(51419520)))]; + tensor var_870_to_fp16 = const()[name = tensor("op_870_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_27_cast_fp16 = layer_norm(axes = input_27_axes_0, beta = norm2_1_bias_to_fp16, epsilon = var_870_to_fp16, gamma = norm2_1_weight_to_fp16, x = input_25_cast_fp16)[name = tensor("input_27_cast_fp16")]; + tensor attn2_in_proj_weight_to_fp16 = const()[name = tensor("attn2_in_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(51421632)))]; + tensor linear_8_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = attn2_in_proj_weight_to_fp16, x = input_27_cast_fp16)[name = tensor("linear_8_cast_fp16")]; + tensor var_884 = const()[name = tensor("op_884"), val = tensor([1, 256, 3, 16, 64])]; + tensor qkv_5_cast_fp16 = reshape(shape = var_884, x = linear_8_cast_fp16)[name = tensor("qkv_5_cast_fp16")]; + tensor q_13_begin_0 = const()[name = tensor("q_13_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor q_13_end_0 = const()[name = tensor("q_13_end_0"), val = tensor([1, 256, 1, 16, 64])]; + tensor q_13_end_mask_0 = const()[name = tensor("q_13_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor q_13_squeeze_mask_0 = const()[name = tensor("q_13_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor q_13_cast_fp16 = slice_by_index(begin = q_13_begin_0, end = q_13_end_0, end_mask = q_13_end_mask_0, squeeze_mask = q_13_squeeze_mask_0, x = qkv_5_cast_fp16)[name = tensor("q_13_cast_fp16")]; + tensor k_9_begin_0 = const()[name = tensor("k_9_begin_0"), val = tensor([0, 0, 1, 0, 0])]; + tensor k_9_end_0 = const()[name = tensor("k_9_end_0"), val = tensor([1, 256, 2, 16, 64])]; + tensor k_9_end_mask_0 = const()[name = tensor("k_9_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor k_9_squeeze_mask_0 = const()[name = tensor("k_9_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor k_9_cast_fp16 = slice_by_index(begin = k_9_begin_0, end = k_9_end_0, end_mask = k_9_end_mask_0, squeeze_mask = k_9_squeeze_mask_0, x = qkv_5_cast_fp16)[name = tensor("k_9_cast_fp16")]; + tensor v_5_begin_0 = const()[name = tensor("v_5_begin_0"), val = tensor([0, 0, 2, 0, 0])]; + tensor v_5_end_0 = const()[name = tensor("v_5_end_0"), val = tensor([1, 256, 3, 16, 64])]; + tensor v_5_end_mask_0 = const()[name = tensor("v_5_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor v_5_squeeze_mask_0 = const()[name = tensor("v_5_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor v_5_cast_fp16 = slice_by_index(begin = v_5_begin_0, end = v_5_end_0, end_mask = v_5_end_mask_0, squeeze_mask = v_5_squeeze_mask_0, x = qkv_5_cast_fp16)[name = tensor("v_5_cast_fp16")]; + tensor var_942 = const()[name = tensor("op_942"), val = tensor([1, 1])]; + tensor position2_to_fp16_dtype_0 = const()[name = tensor("position2_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor position2_to_fp16 = cast(dtype = position2_to_fp16_dtype_0, x = position2)[name = tensor("cast_147")]; + tensor var_943_cast_fp16 = reshape(shape = var_942, x = position2_to_fp16)[name = tensor("op_943_cast_fp16")]; + tensor ts_15_cast_fp16 = add(x = ts_1_promoted_to_fp16, y = var_943_cast_fp16)[name = tensor("ts_15_cast_fp16")]; + tensor var_950 = const()[name = tensor("op_950"), val = tensor([1, 256, 1, 1])]; + tensor ts_17_cast_fp16 = reshape(shape = var_950, x = ts_15_cast_fp16)[name = tensor("ts_17_cast_fp16")]; + tensor var_957 = const()[name = tensor("op_957"), val = tensor([1, 256, 16, 32, 2])]; + tensor q_complex_5_cast_fp16 = reshape(shape = var_957, x = q_13_cast_fp16)[name = tensor("q_complex_5_cast_fp16")]; + tensor var_964 = const()[name = tensor("op_964"), val = tensor([1, 256, 16, 32, 2])]; + tensor k_complex_5_cast_fp16 = reshape(shape = var_964, x = k_9_cast_fp16)[name = tensor("k_complex_5_cast_fp16")]; + tensor var_968_begin_0 = const()[name = tensor("op_968_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_968_end_0 = const()[name = tensor("op_968_end_0"), val = tensor([1, 256, 16, 32, 1])]; + tensor var_968_end_mask_0 = const()[name = tensor("op_968_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_968_squeeze_mask_0 = const()[name = tensor("op_968_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_968_cast_fp16 = slice_by_index(begin = var_968_begin_0, end = var_968_end_0, end_mask = var_968_end_mask_0, squeeze_mask = var_968_squeeze_mask_0, x = q_complex_5_cast_fp16)[name = tensor("op_968_cast_fp16")]; + tensor var_976_begin_0 = const()[name = tensor("op_976_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_976_end_0 = const()[name = tensor("op_976_end_0"), val = tensor([1, 256, 16, 32, 2])]; + tensor var_976_end_mask_0 = const()[name = tensor("op_976_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_976_squeeze_mask_0 = const()[name = tensor("op_976_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_976_cast_fp16 = slice_by_index(begin = var_976_begin_0, end = var_976_end_0, end_mask = var_976_end_mask_0, squeeze_mask = var_976_squeeze_mask_0, x = q_complex_5_cast_fp16)[name = tensor("op_976_cast_fp16")]; + tensor var_984_begin_0 = const()[name = tensor("op_984_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_984_end_0 = const()[name = tensor("op_984_end_0"), val = tensor([1, 256, 16, 32, 1])]; + tensor var_984_end_mask_0 = const()[name = tensor("op_984_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_984_squeeze_mask_0 = const()[name = tensor("op_984_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_984_cast_fp16 = slice_by_index(begin = var_984_begin_0, end = var_984_end_0, end_mask = var_984_end_mask_0, squeeze_mask = var_984_squeeze_mask_0, x = k_complex_5_cast_fp16)[name = tensor("op_984_cast_fp16")]; + tensor var_992_begin_0 = const()[name = tensor("op_992_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_992_end_0 = const()[name = tensor("op_992_end_0"), val = tensor([1, 256, 16, 32, 2])]; + tensor var_992_end_mask_0 = const()[name = tensor("op_992_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_992_squeeze_mask_0 = const()[name = tensor("op_992_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_992_cast_fp16 = slice_by_index(begin = var_992_begin_0, end = var_992_end_0, end_mask = var_992_end_mask_0, squeeze_mask = var_992_squeeze_mask_0, x = k_complex_5_cast_fp16)[name = tensor("op_992_cast_fp16")]; + tensor freqs_5_to_fp16 = const()[name = tensor("freqs_5_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(57713152)))]; + tensor var_998_cast_fp16 = mul(x = freqs_5_to_fp16, y = ts_17_cast_fp16)[name = tensor("op_998_cast_fp16")]; + tensor rotr_5_cast_fp16 = cos(x = var_998_cast_fp16)[name = tensor("rotr_5_cast_fp16")]; + tensor roti_5_cast_fp16 = sin(x = var_998_cast_fp16)[name = tensor("roti_5_cast_fp16")]; + tensor var_1002_cast_fp16 = mul(x = var_968_cast_fp16, y = rotr_5_cast_fp16)[name = tensor("op_1002_cast_fp16")]; + tensor var_1003_cast_fp16 = mul(x = var_976_cast_fp16, y = roti_5_cast_fp16)[name = tensor("op_1003_cast_fp16")]; + tensor qor_9_cast_fp16 = sub(x = var_1002_cast_fp16, y = var_1003_cast_fp16)[name = tensor("qor_9_cast_fp16")]; + tensor var_1006_cast_fp16 = mul(x = var_968_cast_fp16, y = roti_5_cast_fp16)[name = tensor("op_1006_cast_fp16")]; + tensor var_1007_cast_fp16 = mul(x = var_976_cast_fp16, y = rotr_5_cast_fp16)[name = tensor("op_1007_cast_fp16")]; + tensor qoi_9_cast_fp16 = add(x = var_1006_cast_fp16, y = var_1007_cast_fp16)[name = tensor("qoi_9_cast_fp16")]; + tensor var_1010_cast_fp16 = mul(x = var_984_cast_fp16, y = rotr_5_cast_fp16)[name = tensor("op_1010_cast_fp16")]; + tensor var_1011_cast_fp16 = mul(x = var_992_cast_fp16, y = roti_5_cast_fp16)[name = tensor("op_1011_cast_fp16")]; + tensor kor_9_cast_fp16 = sub(x = var_1010_cast_fp16, y = var_1011_cast_fp16)[name = tensor("kor_9_cast_fp16")]; + tensor var_1014_cast_fp16 = mul(x = var_984_cast_fp16, y = roti_5_cast_fp16)[name = tensor("op_1014_cast_fp16")]; + tensor var_1015_cast_fp16 = mul(x = var_992_cast_fp16, y = rotr_5_cast_fp16)[name = tensor("op_1015_cast_fp16")]; + tensor koi_9_cast_fp16 = add(x = var_1014_cast_fp16, y = var_1015_cast_fp16)[name = tensor("koi_9_cast_fp16")]; + tensor qo_5_axis_0 = const()[name = tensor("qo_5_axis_0"), val = tensor(-1)]; + tensor qo_5_cast_fp16 = stack(axis = qo_5_axis_0, values = (qor_9_cast_fp16, qoi_9_cast_fp16))[name = tensor("qo_5_cast_fp16")]; + tensor ko_5_axis_0 = const()[name = tensor("ko_5_axis_0"), val = tensor(-1)]; + tensor ko_5_cast_fp16 = stack(axis = ko_5_axis_0, values = (kor_9_cast_fp16, koi_9_cast_fp16))[name = tensor("ko_5_cast_fp16")]; + tensor var_1048 = const()[name = tensor("op_1048"), val = tensor([1, 256, 16, 64])]; + tensor q_15_cast_fp16 = reshape(shape = var_1048, x = qo_5_cast_fp16)[name = tensor("q_15_cast_fp16")]; + tensor var_1054 = const()[name = tensor("op_1054"), val = tensor([1, 256, 16, 64])]; + tensor k_11_cast_fp16 = reshape(shape = var_1054, x = ko_5_cast_fp16)[name = tensor("k_11_cast_fp16")]; + tensor write_indices_float_13_cast_fp16 = add(x = var_943_cast_fp16, y = var_249_promoted_to_fp16)[name = tensor("write_indices_float_13_cast_fp16")]; + tensor _inversed_1075_y_0_to_fp16 = const()[name = tensor("_inversed_1075_y_0_to_fp16"), val = tensor(0x1p-9)]; + tensor _inversed_1075_cast_fp16 = mul(x = write_indices_float_13_cast_fp16, y = _inversed_1075_y_0_to_fp16)[name = tensor("_inversed_1075_cast_fp16")]; + tensor var_1076_cast_fp16 = floor(x = _inversed_1075_cast_fp16)[name = tensor("op_1076_cast_fp16")]; + tensor var_1077_to_fp16 = const()[name = tensor("op_1077_to_fp16"), val = tensor(0x1p+9)]; + tensor var_1078_cast_fp16 = mul(x = var_1076_cast_fp16, y = var_1077_to_fp16)[name = tensor("op_1078_cast_fp16")]; + tensor write_indices_float_15_cast_fp16 = sub(x = write_indices_float_13_cast_fp16, y = var_1078_cast_fp16)[name = tensor("write_indices_float_15_cast_fp16")]; + tensor var_1096_value_0_to_fp16 = const()[name = tensor("op_1096_value_0_to_fp16"), val = tensor(0x1.ffp+8)]; + tensor var_1096_cast_fp16 = fill_like(ref_tensor = write_indices_float_15_cast_fp16, value = var_1096_value_0_to_fp16)[name = tensor("op_1096_cast_fp16")]; + tensor write_indices_float_17_cast_fp16 = select(a = write_indices_float_15_cast_fp16, b = var_1096_cast_fp16, cond = token_valid_1_cast_fp16)[name = tensor("write_indices_float_17_cast_fp16")]; + tensor var_1102_dtype_0 = const()[name = tensor("op_1102_dtype_0"), val = tensor("int32")]; + tensor var_1107 = const()[name = tensor("op_1107"), val = tensor([1, 256, 1, 1])]; + tensor write_indices_float_17_cast_fp16_to_int32 = cast(dtype = var_1102_dtype_0, x = write_indices_float_17_cast_fp16)[name = tensor("cast_146")]; + tensor var_1108 = reshape(shape = var_1107, x = write_indices_float_17_cast_fp16_to_int32)[name = tensor("op_1108")]; + tensor write_indices_5_reps_0 = const()[name = tensor("write_indices_5_reps_0"), val = tensor([1, 1, 16, 64])]; + tensor write_indices_5 = tile(reps = write_indices_5_reps_0, x = var_1108)[name = tensor("write_indices_5")]; + tensor var_1118_begin_0 = const()[name = tensor("op_1118_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_1118_end_0 = const()[name = tensor("op_1118_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor var_1118_end_mask_0 = const()[name = tensor("op_1118_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_1118_squeeze_mask_0 = const()[name = tensor("op_1118_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor cache2_to_fp16_dtype_0 = const()[name = tensor("cache2_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor cache2_to_fp16 = cast(dtype = cache2_to_fp16_dtype_0, x = cache2)[name = tensor("cast_145")]; + tensor var_1118_cast_fp16 = slice_by_index(begin = var_1118_begin_0, end = var_1118_end_0, end_mask = var_1118_end_mask_0, squeeze_mask = var_1118_squeeze_mask_0, x = cache2_to_fp16)[name = tensor("op_1118_cast_fp16")]; + tensor var_1120_axis_0 = const()[name = tensor("op_1120_axis_0"), val = tensor(1)]; + tensor var_1120_mode_0 = const()[name = tensor("op_1120_mode_0"), val = tensor("update")]; + tensor var_1120_validate_indices_0 = const()[name = tensor("op_1120_validate_indices_0"), val = tensor(false)]; + tensor var_1120_cast_fp16 = scatter_along_axis(axis = var_1120_axis_0, data = var_1118_cast_fp16, indices = write_indices_5, mode = var_1120_mode_0, updates = k_11_cast_fp16, validate_indices = var_1120_validate_indices_0)[name = tensor("op_1120_cast_fp16")]; + tensor concat_18 = const()[name = tensor("concat_18"), val = tensor([0, 0, 0, 0, 0])]; + tensor concat_19 = const()[name = tensor("concat_19"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_5_internal_tensor_assign_1_stride_0 = const()[name = tensor("new_cache_5_internal_tensor_assign_1_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_5_internal_tensor_assign_1_begin_mask_0 = const()[name = tensor("new_cache_5_internal_tensor_assign_1_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_5_internal_tensor_assign_1_end_mask_0 = const()[name = tensor("new_cache_5_internal_tensor_assign_1_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_5_internal_tensor_assign_1_squeeze_mask_0 = const()[name = tensor("new_cache_5_internal_tensor_assign_1_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_14 = const()[name = tensor("shape_14"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_4 = const()[name = tensor("reduce_prod_4"), val = tensor(1048576)]; + tensor range_1d_4_start_0 = const()[name = tensor("range_1d_4_start_0"), val = tensor(0)]; + tensor range_1d_4_step_0 = const()[name = tensor("range_1d_4_step_0"), val = tensor(1)]; + tensor range_1d_4 = range_1d(end = reduce_prod_4, start = range_1d_4_start_0, step = range_1d_4_step_0)[name = tensor("range_1d_4")]; + tensor reshape_20 = reshape(shape = shape_14, x = range_1d_4)[name = tensor("reshape_20")]; + tensor slice_by_index_4 = slice_by_index(begin = concat_18, begin_mask = new_cache_5_internal_tensor_assign_1_begin_mask_0, end = concat_19, end_mask = new_cache_5_internal_tensor_assign_1_end_mask_0, squeeze_mask = new_cache_5_internal_tensor_assign_1_squeeze_mask_0, stride = new_cache_5_internal_tensor_assign_1_stride_0, x = reshape_20)[name = tensor("slice_by_index_4")]; + tensor reshape_21_shape_0 = const()[name = tensor("reshape_21_shape_0"), val = tensor([-1])]; + tensor reshape_21 = reshape(shape = reshape_21_shape_0, x = slice_by_index_4)[name = tensor("reshape_21")]; + tensor reshape_22_shape_0 = const()[name = tensor("reshape_22_shape_0"), val = tensor([-1])]; + tensor reshape_22_cast_fp16 = reshape(shape = reshape_22_shape_0, x = var_1120_cast_fp16)[name = tensor("reshape_22_cast_fp16")]; + tensor reshape_23_shape_0 = const()[name = tensor("reshape_23_shape_0"), val = tensor([-1])]; + tensor reshape_23_cast_fp16 = reshape(shape = reshape_23_shape_0, x = cache2_to_fp16)[name = tensor("reshape_23_cast_fp16")]; + tensor scatter_4_mode_0 = const()[name = tensor("scatter_4_mode_0"), val = tensor("update")]; + tensor scatter_4_axis_0 = const()[name = tensor("scatter_4_axis_0"), val = tensor(0)]; + tensor scatter_4_validate_indices_0 = const()[name = tensor("scatter_4_validate_indices_0"), val = tensor(false)]; + tensor scatter_4_cast_fp16 = scatter(axis = scatter_4_axis_0, data = reshape_23_cast_fp16, indices = reshape_21, mode = scatter_4_mode_0, updates = reshape_22_cast_fp16, validate_indices = scatter_4_validate_indices_0)[name = tensor("scatter_4_cast_fp16")]; + tensor reshape_24_cast_fp16 = reshape(shape = shape_14, x = scatter_4_cast_fp16)[name = tensor("reshape_24_cast_fp16")]; + tensor var_1128_begin_0 = const()[name = tensor("op_1128_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor var_1128_end_0 = const()[name = tensor("op_1128_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor var_1128_end_mask_0 = const()[name = tensor("op_1128_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_1128_squeeze_mask_0 = const()[name = tensor("op_1128_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor var_1128_cast_fp16 = slice_by_index(begin = var_1128_begin_0, end = var_1128_end_0, end_mask = var_1128_end_mask_0, squeeze_mask = var_1128_squeeze_mask_0, x = reshape_24_cast_fp16)[name = tensor("op_1128_cast_fp16")]; + tensor var_1130_axis_0 = const()[name = tensor("op_1130_axis_0"), val = tensor(1)]; + tensor var_1130_mode_0 = const()[name = tensor("op_1130_mode_0"), val = tensor("update")]; + tensor var_1130_validate_indices_0 = const()[name = tensor("op_1130_validate_indices_0"), val = tensor(false)]; + tensor var_1130_cast_fp16 = scatter_along_axis(axis = var_1130_axis_0, data = var_1128_cast_fp16, indices = write_indices_5, mode = var_1130_mode_0, updates = v_5_cast_fp16, validate_indices = var_1130_validate_indices_0)[name = tensor("op_1130_cast_fp16")]; + tensor concat_20 = const()[name = tensor("concat_20"), val = tensor([1, 0, 0, 0, 0])]; + tensor concat_21 = const()[name = tensor("concat_21"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_5_internal_tensor_assign_2_stride_0 = const()[name = tensor("new_cache_5_internal_tensor_assign_2_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_5_internal_tensor_assign_2_begin_mask_0 = const()[name = tensor("new_cache_5_internal_tensor_assign_2_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_5_internal_tensor_assign_2_end_mask_0 = const()[name = tensor("new_cache_5_internal_tensor_assign_2_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_5_internal_tensor_assign_2_squeeze_mask_0 = const()[name = tensor("new_cache_5_internal_tensor_assign_2_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_15 = const()[name = tensor("shape_15"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_5 = const()[name = tensor("reduce_prod_5"), val = tensor(1048576)]; + tensor range_1d_5_start_0 = const()[name = tensor("range_1d_5_start_0"), val = tensor(0)]; + tensor range_1d_5_step_0 = const()[name = tensor("range_1d_5_step_0"), val = tensor(1)]; + tensor range_1d_5 = range_1d(end = reduce_prod_5, start = range_1d_5_start_0, step = range_1d_5_step_0)[name = tensor("range_1d_5")]; + tensor reshape_25 = reshape(shape = shape_15, x = range_1d_5)[name = tensor("reshape_25")]; + tensor slice_by_index_5 = slice_by_index(begin = concat_20, begin_mask = new_cache_5_internal_tensor_assign_2_begin_mask_0, end = concat_21, end_mask = new_cache_5_internal_tensor_assign_2_end_mask_0, squeeze_mask = new_cache_5_internal_tensor_assign_2_squeeze_mask_0, stride = new_cache_5_internal_tensor_assign_2_stride_0, x = reshape_25)[name = tensor("slice_by_index_5")]; + tensor reshape_26_shape_0 = const()[name = tensor("reshape_26_shape_0"), val = tensor([-1])]; + tensor reshape_26 = reshape(shape = reshape_26_shape_0, x = slice_by_index_5)[name = tensor("reshape_26")]; + tensor reshape_27_shape_0 = const()[name = tensor("reshape_27_shape_0"), val = tensor([-1])]; + tensor reshape_27_cast_fp16 = reshape(shape = reshape_27_shape_0, x = var_1130_cast_fp16)[name = tensor("reshape_27_cast_fp16")]; + tensor reshape_28_shape_0 = const()[name = tensor("reshape_28_shape_0"), val = tensor([-1])]; + tensor reshape_28_cast_fp16 = reshape(shape = reshape_28_shape_0, x = reshape_24_cast_fp16)[name = tensor("reshape_28_cast_fp16")]; + tensor scatter_5_mode_0 = const()[name = tensor("scatter_5_mode_0"), val = tensor("update")]; + tensor scatter_5_axis_0 = const()[name = tensor("scatter_5_axis_0"), val = tensor(0)]; + tensor scatter_5_validate_indices_0 = const()[name = tensor("scatter_5_validate_indices_0"), val = tensor(false)]; + tensor scatter_5_cast_fp16 = scatter(axis = scatter_5_axis_0, data = reshape_28_cast_fp16, indices = reshape_26, mode = scatter_5_mode_0, updates = reshape_27_cast_fp16, validate_indices = scatter_5_validate_indices_0)[name = tensor("scatter_5_cast_fp16")]; + tensor reshape_29_cast_fp16 = reshape(shape = shape_15, x = scatter_5_cast_fp16)[name = tensor("reshape_29_cast_fp16")]; + tensor reshape_29_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("reshape_29_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor keys_13_begin_0 = const()[name = tensor("keys_13_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor keys_13_end_0 = const()[name = tensor("keys_13_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor keys_13_end_mask_0 = const()[name = tensor("keys_13_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor keys_13_squeeze_mask_0 = const()[name = tensor("keys_13_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor keys_13_cast_fp16 = slice_by_index(begin = keys_13_begin_0, end = keys_13_end_0, end_mask = keys_13_end_mask_0, squeeze_mask = keys_13_squeeze_mask_0, x = reshape_29_cast_fp16)[name = tensor("keys_13_cast_fp16")]; + tensor values_13_begin_0 = const()[name = tensor("values_13_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor values_13_end_0 = const()[name = tensor("values_13_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor values_13_end_mask_0 = const()[name = tensor("values_13_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor values_13_squeeze_mask_0 = const()[name = tensor("values_13_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor values_13_cast_fp16 = slice_by_index(begin = values_13_begin_0, end = values_13_end_0, end_mask = values_13_end_mask_0, squeeze_mask = values_13_squeeze_mask_0, x = reshape_29_cast_fp16)[name = tensor("values_13_cast_fp16")]; + tensor var_1142_cast_fp16 = not_equal(x = keys_13_cast_fp16, y = keys_13_cast_fp16)[name = tensor("op_1142_cast_fp16")]; + tensor keys_15_cast_fp16 = select(a = var_326_to_fp16, b = keys_13_cast_fp16, cond = var_1142_cast_fp16)[name = tensor("keys_15_cast_fp16")]; + tensor var_1150_cast_fp16 = not_equal(x = values_13_cast_fp16, y = values_13_cast_fp16)[name = tensor("op_1150_cast_fp16")]; + tensor values_15_cast_fp16 = select(a = var_326_to_fp16, b = values_13_cast_fp16, cond = var_1150_cast_fp16)[name = tensor("values_15_cast_fp16")]; + tensor var_1174 = const()[name = tensor("op_1174"), val = tensor([0, 2, 1, 3])]; + tensor var_1190 = const()[name = tensor("op_1190"), val = tensor([1, 1, 1])]; + tensor var_1191_cast_fp16 = reshape(shape = var_1190, x = position2_to_fp16)[name = tensor("op_1191_cast_fp16")]; + tensor q_positions_5_cast_fp16 = add(x = var_1191_cast_fp16, y = q_offsets_1_promoted_to_fp16)[name = tensor("q_positions_5_cast_fp16")]; + tensor valid_end_5_cast_fp16 = add(x = var_1191_cast_fp16, y = var_392_cast_fp16)[name = tensor("valid_end_5_cast_fp16")]; + tensor valid_mask_5_cast_fp16 = less(x = k_positions_1_promoted_to_fp16, y = valid_end_5_cast_fp16)[name = tensor("valid_mask_5_cast_fp16")]; + tensor causal_mask_5_cast_fp16 = less_equal(x = k_positions_1_promoted_to_fp16, y = q_positions_5_cast_fp16)[name = tensor("causal_mask_5_cast_fp16")]; + tensor var_1219 = logical_and(x = valid_mask_5_cast_fp16, y = causal_mask_5_cast_fp16)[name = tensor("op_1219")]; + tensor attn_mask_5_axes_0 = const()[name = tensor("attn_mask_5_axes_0"), val = tensor([1])]; + tensor attn_mask_5 = expand_dims(axes = attn_mask_5_axes_0, x = var_1219)[name = tensor("attn_mask_5")]; + tensor var_1233_transpose_x_0 = const()[name = tensor("op_1233_transpose_x_0"), val = tensor(false)]; + tensor var_1233_transpose_y_0 = const()[name = tensor("op_1233_transpose_y_0"), val = tensor(false)]; + tensor transpose_19_perm_0 = const()[name = tensor("transpose_19_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_20_perm_0 = const()[name = tensor("transpose_20_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_20 = transpose(perm = transpose_20_perm_0, x = keys_15_cast_fp16)[name = tensor("transpose_34")]; + tensor transpose_19 = transpose(perm = transpose_19_perm_0, x = q_15_cast_fp16)[name = tensor("transpose_35")]; + tensor var_1233_cast_fp16 = matmul(transpose_x = var_1233_transpose_x_0, transpose_y = var_1233_transpose_y_0, x = transpose_19, y = transpose_20)[name = tensor("op_1233_cast_fp16")]; + tensor var_1227_to_fp16 = const()[name = tensor("op_1227_to_fp16"), val = tensor([0x1p-3])]; + tensor attn_weights_13_cast_fp16 = mul(x = var_1233_cast_fp16, y = var_1227_to_fp16)[name = tensor("attn_weights_13_cast_fp16")]; + tensor var_1235 = logical_not(x = attn_mask_5)[name = tensor("op_1235")]; + tensor var_1236_to_fp16 = const()[name = tensor("op_1236_to_fp16"), val = tensor(-inf)]; + tensor attn_weights_15_cast_fp16 = select(a = var_1236_to_fp16, b = attn_weights_13_cast_fp16, cond = var_1235)[name = tensor("attn_weights_15_cast_fp16")]; + tensor var_1238 = const()[name = tensor("op_1238"), val = tensor(-1)]; + tensor attn_weights_17_cast_fp16 = softmax(axis = var_1238, x = attn_weights_15_cast_fp16)[name = tensor("attn_weights_17_cast_fp16")]; + tensor attn_output_5_transpose_x_0 = const()[name = tensor("attn_output_5_transpose_x_0"), val = tensor(false)]; + tensor attn_output_5_transpose_y_0 = const()[name = tensor("attn_output_5_transpose_y_0"), val = tensor(false)]; + tensor values_17_cast_fp16 = transpose(perm = var_1174, x = values_15_cast_fp16)[name = tensor("transpose_36")]; + tensor attn_output_5_cast_fp16 = matmul(transpose_x = attn_output_5_transpose_x_0, transpose_y = attn_output_5_transpose_y_0, x = attn_weights_17_cast_fp16, y = values_17_cast_fp16)[name = tensor("attn_output_5_cast_fp16")]; + tensor var_1246 = const()[name = tensor("op_1246"), val = tensor([0, 2, 1, 3])]; + tensor var_1251 = const()[name = tensor("op_1251"), val = tensor([1, 256, 1024])]; + tensor var_1247_cast_fp16 = transpose(perm = var_1246, x = attn_output_5_cast_fp16)[name = tensor("transpose_33")]; + tensor input_29_cast_fp16 = reshape(shape = var_1251, x = var_1247_cast_fp16)[name = tensor("input_29_cast_fp16")]; + tensor attn2_out_proj_weight_to_fp16 = const()[name = tensor("attn2_out_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(57713280)))]; + tensor linear_9_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = attn2_out_proj_weight_to_fp16, x = input_29_cast_fp16)[name = tensor("linear_9_cast_fp16")]; + tensor var_1260_cast_fp16 = add(x = position2_to_fp16, y = valid_len_to_fp16)[name = tensor("op_1260_cast_fp16")]; + tensor var_1260_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_1260_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor input_31_cast_fp16 = add(x = input_25_cast_fp16, y = linear_9_cast_fp16)[name = tensor("input_31_cast_fp16")]; + tensor input_33_axes_0 = const()[name = tensor("input_33_axes_0"), val = tensor([-1])]; + tensor norm2_2_weight_to_fp16 = const()[name = tensor("norm2_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(59810496)))]; + tensor norm2_2_bias_to_fp16 = const()[name = tensor("norm2_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(59812608)))]; + tensor var_1264_to_fp16 = const()[name = tensor("op_1264_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_33_cast_fp16 = layer_norm(axes = input_33_axes_0, beta = norm2_2_bias_to_fp16, epsilon = var_1264_to_fp16, gamma = norm2_2_weight_to_fp16, x = input_31_cast_fp16)[name = tensor("input_33_cast_fp16")]; + tensor linear2_1_weight_to_fp16 = const()[name = tensor("linear2_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(59814720)))]; + tensor linear_10_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = linear2_1_weight_to_fp16, x = input_33_cast_fp16)[name = tensor("linear_10_cast_fp16")]; + tensor input_35_mode_0 = const()[name = tensor("input_35_mode_0"), val = tensor("EXACT")]; + tensor input_35_cast_fp16 = gelu(mode = input_35_mode_0, x = linear_10_cast_fp16)[name = tensor("input_35_cast_fp16")]; + tensor linear2_2_weight_to_fp16 = const()[name = tensor("linear2_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(68203392)))]; + tensor linear_11_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = linear2_2_weight_to_fp16, x = input_35_cast_fp16)[name = tensor("linear_11_cast_fp16")]; + tensor input_37_cast_fp16 = add(x = input_31_cast_fp16, y = linear_11_cast_fp16)[name = tensor("input_37_cast_fp16")]; + tensor input_39_axes_0 = const()[name = tensor("input_39_axes_0"), val = tensor([-1])]; + tensor norm3_1_weight_to_fp16 = const()[name = tensor("norm3_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(76592064)))]; + tensor norm3_1_bias_to_fp16 = const()[name = tensor("norm3_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(76594176)))]; + tensor var_1281_to_fp16 = const()[name = tensor("op_1281_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_39_cast_fp16 = layer_norm(axes = input_39_axes_0, beta = norm3_1_bias_to_fp16, epsilon = var_1281_to_fp16, gamma = norm3_1_weight_to_fp16, x = input_37_cast_fp16)[name = tensor("input_39_cast_fp16")]; + tensor attn3_in_proj_weight_to_fp16 = const()[name = tensor("attn3_in_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(76596288)))]; + tensor linear_12_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = attn3_in_proj_weight_to_fp16, x = input_39_cast_fp16)[name = tensor("linear_12_cast_fp16")]; + tensor var_1295 = const()[name = tensor("op_1295"), val = tensor([1, 256, 3, 16, 64])]; + tensor qkv_7_cast_fp16 = reshape(shape = var_1295, x = linear_12_cast_fp16)[name = tensor("qkv_7_cast_fp16")]; + tensor q_19_begin_0 = const()[name = tensor("q_19_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor q_19_end_0 = const()[name = tensor("q_19_end_0"), val = tensor([1, 256, 1, 16, 64])]; + tensor q_19_end_mask_0 = const()[name = tensor("q_19_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor q_19_squeeze_mask_0 = const()[name = tensor("q_19_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor q_19_cast_fp16 = slice_by_index(begin = q_19_begin_0, end = q_19_end_0, end_mask = q_19_end_mask_0, squeeze_mask = q_19_squeeze_mask_0, x = qkv_7_cast_fp16)[name = tensor("q_19_cast_fp16")]; + tensor k_13_begin_0 = const()[name = tensor("k_13_begin_0"), val = tensor([0, 0, 1, 0, 0])]; + tensor k_13_end_0 = const()[name = tensor("k_13_end_0"), val = tensor([1, 256, 2, 16, 64])]; + tensor k_13_end_mask_0 = const()[name = tensor("k_13_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor k_13_squeeze_mask_0 = const()[name = tensor("k_13_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor k_13_cast_fp16 = slice_by_index(begin = k_13_begin_0, end = k_13_end_0, end_mask = k_13_end_mask_0, squeeze_mask = k_13_squeeze_mask_0, x = qkv_7_cast_fp16)[name = tensor("k_13_cast_fp16")]; + tensor v_7_begin_0 = const()[name = tensor("v_7_begin_0"), val = tensor([0, 0, 2, 0, 0])]; + tensor v_7_end_0 = const()[name = tensor("v_7_end_0"), val = tensor([1, 256, 3, 16, 64])]; + tensor v_7_end_mask_0 = const()[name = tensor("v_7_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor v_7_squeeze_mask_0 = const()[name = tensor("v_7_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor v_7_cast_fp16 = slice_by_index(begin = v_7_begin_0, end = v_7_end_0, end_mask = v_7_end_mask_0, squeeze_mask = v_7_squeeze_mask_0, x = qkv_7_cast_fp16)[name = tensor("v_7_cast_fp16")]; + tensor var_1353 = const()[name = tensor("op_1353"), val = tensor([1, 1])]; + tensor position3_to_fp16_dtype_0 = const()[name = tensor("position3_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor position3_to_fp16 = cast(dtype = position3_to_fp16_dtype_0, x = position3)[name = tensor("cast_142")]; + tensor var_1354_cast_fp16 = reshape(shape = var_1353, x = position3_to_fp16)[name = tensor("op_1354_cast_fp16")]; + tensor ts_21_cast_fp16 = add(x = ts_1_promoted_to_fp16, y = var_1354_cast_fp16)[name = tensor("ts_21_cast_fp16")]; + tensor var_1361 = const()[name = tensor("op_1361"), val = tensor([1, 256, 1, 1])]; + tensor ts_23_cast_fp16 = reshape(shape = var_1361, x = ts_21_cast_fp16)[name = tensor("ts_23_cast_fp16")]; + tensor var_1368 = const()[name = tensor("op_1368"), val = tensor([1, 256, 16, 32, 2])]; + tensor q_complex_7_cast_fp16 = reshape(shape = var_1368, x = q_19_cast_fp16)[name = tensor("q_complex_7_cast_fp16")]; + tensor var_1375 = const()[name = tensor("op_1375"), val = tensor([1, 256, 16, 32, 2])]; + tensor k_complex_7_cast_fp16 = reshape(shape = var_1375, x = k_13_cast_fp16)[name = tensor("k_complex_7_cast_fp16")]; + tensor var_1379_begin_0 = const()[name = tensor("op_1379_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_1379_end_0 = const()[name = tensor("op_1379_end_0"), val = tensor([1, 256, 16, 32, 1])]; + tensor var_1379_end_mask_0 = const()[name = tensor("op_1379_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_1379_squeeze_mask_0 = const()[name = tensor("op_1379_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_1379_cast_fp16 = slice_by_index(begin = var_1379_begin_0, end = var_1379_end_0, end_mask = var_1379_end_mask_0, squeeze_mask = var_1379_squeeze_mask_0, x = q_complex_7_cast_fp16)[name = tensor("op_1379_cast_fp16")]; + tensor var_1387_begin_0 = const()[name = tensor("op_1387_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_1387_end_0 = const()[name = tensor("op_1387_end_0"), val = tensor([1, 256, 16, 32, 2])]; + tensor var_1387_end_mask_0 = const()[name = tensor("op_1387_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_1387_squeeze_mask_0 = const()[name = tensor("op_1387_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_1387_cast_fp16 = slice_by_index(begin = var_1387_begin_0, end = var_1387_end_0, end_mask = var_1387_end_mask_0, squeeze_mask = var_1387_squeeze_mask_0, x = q_complex_7_cast_fp16)[name = tensor("op_1387_cast_fp16")]; + tensor var_1395_begin_0 = const()[name = tensor("op_1395_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_1395_end_0 = const()[name = tensor("op_1395_end_0"), val = tensor([1, 256, 16, 32, 1])]; + tensor var_1395_end_mask_0 = const()[name = tensor("op_1395_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_1395_squeeze_mask_0 = const()[name = tensor("op_1395_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_1395_cast_fp16 = slice_by_index(begin = var_1395_begin_0, end = var_1395_end_0, end_mask = var_1395_end_mask_0, squeeze_mask = var_1395_squeeze_mask_0, x = k_complex_7_cast_fp16)[name = tensor("op_1395_cast_fp16")]; + tensor var_1403_begin_0 = const()[name = tensor("op_1403_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_1403_end_0 = const()[name = tensor("op_1403_end_0"), val = tensor([1, 256, 16, 32, 2])]; + tensor var_1403_end_mask_0 = const()[name = tensor("op_1403_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_1403_squeeze_mask_0 = const()[name = tensor("op_1403_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_1403_cast_fp16 = slice_by_index(begin = var_1403_begin_0, end = var_1403_end_0, end_mask = var_1403_end_mask_0, squeeze_mask = var_1403_squeeze_mask_0, x = k_complex_7_cast_fp16)[name = tensor("op_1403_cast_fp16")]; + tensor freqs_7_to_fp16 = const()[name = tensor("freqs_7_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(82887808)))]; + tensor var_1409_cast_fp16 = mul(x = freqs_7_to_fp16, y = ts_23_cast_fp16)[name = tensor("op_1409_cast_fp16")]; + tensor rotr_7_cast_fp16 = cos(x = var_1409_cast_fp16)[name = tensor("rotr_7_cast_fp16")]; + tensor roti_7_cast_fp16 = sin(x = var_1409_cast_fp16)[name = tensor("roti_7_cast_fp16")]; + tensor var_1413_cast_fp16 = mul(x = var_1379_cast_fp16, y = rotr_7_cast_fp16)[name = tensor("op_1413_cast_fp16")]; + tensor var_1414_cast_fp16 = mul(x = var_1387_cast_fp16, y = roti_7_cast_fp16)[name = tensor("op_1414_cast_fp16")]; + tensor qor_13_cast_fp16 = sub(x = var_1413_cast_fp16, y = var_1414_cast_fp16)[name = tensor("qor_13_cast_fp16")]; + tensor var_1417_cast_fp16 = mul(x = var_1379_cast_fp16, y = roti_7_cast_fp16)[name = tensor("op_1417_cast_fp16")]; + tensor var_1418_cast_fp16 = mul(x = var_1387_cast_fp16, y = rotr_7_cast_fp16)[name = tensor("op_1418_cast_fp16")]; + tensor qoi_13_cast_fp16 = add(x = var_1417_cast_fp16, y = var_1418_cast_fp16)[name = tensor("qoi_13_cast_fp16")]; + tensor var_1421_cast_fp16 = mul(x = var_1395_cast_fp16, y = rotr_7_cast_fp16)[name = tensor("op_1421_cast_fp16")]; + tensor var_1422_cast_fp16 = mul(x = var_1403_cast_fp16, y = roti_7_cast_fp16)[name = tensor("op_1422_cast_fp16")]; + tensor kor_13_cast_fp16 = sub(x = var_1421_cast_fp16, y = var_1422_cast_fp16)[name = tensor("kor_13_cast_fp16")]; + tensor var_1425_cast_fp16 = mul(x = var_1395_cast_fp16, y = roti_7_cast_fp16)[name = tensor("op_1425_cast_fp16")]; + tensor var_1426_cast_fp16 = mul(x = var_1403_cast_fp16, y = rotr_7_cast_fp16)[name = tensor("op_1426_cast_fp16")]; + tensor koi_13_cast_fp16 = add(x = var_1425_cast_fp16, y = var_1426_cast_fp16)[name = tensor("koi_13_cast_fp16")]; + tensor qo_7_axis_0 = const()[name = tensor("qo_7_axis_0"), val = tensor(-1)]; + tensor qo_7_cast_fp16 = stack(axis = qo_7_axis_0, values = (qor_13_cast_fp16, qoi_13_cast_fp16))[name = tensor("qo_7_cast_fp16")]; + tensor ko_7_axis_0 = const()[name = tensor("ko_7_axis_0"), val = tensor(-1)]; + tensor ko_7_cast_fp16 = stack(axis = ko_7_axis_0, values = (kor_13_cast_fp16, koi_13_cast_fp16))[name = tensor("ko_7_cast_fp16")]; + tensor var_1459 = const()[name = tensor("op_1459"), val = tensor([1, 256, 16, 64])]; + tensor q_21_cast_fp16 = reshape(shape = var_1459, x = qo_7_cast_fp16)[name = tensor("q_21_cast_fp16")]; + tensor var_1465 = const()[name = tensor("op_1465"), val = tensor([1, 256, 16, 64])]; + tensor k_15_cast_fp16 = reshape(shape = var_1465, x = ko_7_cast_fp16)[name = tensor("k_15_cast_fp16")]; + tensor write_indices_float_19_cast_fp16 = add(x = var_1354_cast_fp16, y = var_249_promoted_to_fp16)[name = tensor("write_indices_float_19_cast_fp16")]; + tensor _inversed_1486_y_0_to_fp16 = const()[name = tensor("_inversed_1486_y_0_to_fp16"), val = tensor(0x1p-9)]; + tensor _inversed_1486_cast_fp16 = mul(x = write_indices_float_19_cast_fp16, y = _inversed_1486_y_0_to_fp16)[name = tensor("_inversed_1486_cast_fp16")]; + tensor var_1487_cast_fp16 = floor(x = _inversed_1486_cast_fp16)[name = tensor("op_1487_cast_fp16")]; + tensor var_1488_to_fp16 = const()[name = tensor("op_1488_to_fp16"), val = tensor(0x1p+9)]; + tensor var_1489_cast_fp16 = mul(x = var_1487_cast_fp16, y = var_1488_to_fp16)[name = tensor("op_1489_cast_fp16")]; + tensor write_indices_float_21_cast_fp16 = sub(x = write_indices_float_19_cast_fp16, y = var_1489_cast_fp16)[name = tensor("write_indices_float_21_cast_fp16")]; + tensor var_1507_value_0_to_fp16 = const()[name = tensor("op_1507_value_0_to_fp16"), val = tensor(0x1.ffp+8)]; + tensor var_1507_cast_fp16 = fill_like(ref_tensor = write_indices_float_21_cast_fp16, value = var_1507_value_0_to_fp16)[name = tensor("op_1507_cast_fp16")]; + tensor write_indices_float_23_cast_fp16 = select(a = write_indices_float_21_cast_fp16, b = var_1507_cast_fp16, cond = token_valid_1_cast_fp16)[name = tensor("write_indices_float_23_cast_fp16")]; + tensor var_1513_dtype_0 = const()[name = tensor("op_1513_dtype_0"), val = tensor("int32")]; + tensor var_1518 = const()[name = tensor("op_1518"), val = tensor([1, 256, 1, 1])]; + tensor write_indices_float_23_cast_fp16_to_int32 = cast(dtype = var_1513_dtype_0, x = write_indices_float_23_cast_fp16)[name = tensor("cast_141")]; + tensor var_1519 = reshape(shape = var_1518, x = write_indices_float_23_cast_fp16_to_int32)[name = tensor("op_1519")]; + tensor write_indices_7_reps_0 = const()[name = tensor("write_indices_7_reps_0"), val = tensor([1, 1, 16, 64])]; + tensor write_indices_7 = tile(reps = write_indices_7_reps_0, x = var_1519)[name = tensor("write_indices_7")]; + tensor var_1529_begin_0 = const()[name = tensor("op_1529_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_1529_end_0 = const()[name = tensor("op_1529_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor var_1529_end_mask_0 = const()[name = tensor("op_1529_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_1529_squeeze_mask_0 = const()[name = tensor("op_1529_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor cache3_to_fp16_dtype_0 = const()[name = tensor("cache3_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor cache3_to_fp16 = cast(dtype = cache3_to_fp16_dtype_0, x = cache3)[name = tensor("cast_140")]; + tensor var_1529_cast_fp16 = slice_by_index(begin = var_1529_begin_0, end = var_1529_end_0, end_mask = var_1529_end_mask_0, squeeze_mask = var_1529_squeeze_mask_0, x = cache3_to_fp16)[name = tensor("op_1529_cast_fp16")]; + tensor var_1531_axis_0 = const()[name = tensor("op_1531_axis_0"), val = tensor(1)]; + tensor var_1531_mode_0 = const()[name = tensor("op_1531_mode_0"), val = tensor("update")]; + tensor var_1531_validate_indices_0 = const()[name = tensor("op_1531_validate_indices_0"), val = tensor(false)]; + tensor var_1531_cast_fp16 = scatter_along_axis(axis = var_1531_axis_0, data = var_1529_cast_fp16, indices = write_indices_7, mode = var_1531_mode_0, updates = k_15_cast_fp16, validate_indices = var_1531_validate_indices_0)[name = tensor("op_1531_cast_fp16")]; + tensor concat_26 = const()[name = tensor("concat_26"), val = tensor([0, 0, 0, 0, 0])]; + tensor concat_27 = const()[name = tensor("concat_27"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_7_internal_tensor_assign_1_stride_0 = const()[name = tensor("new_cache_7_internal_tensor_assign_1_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_7_internal_tensor_assign_1_begin_mask_0 = const()[name = tensor("new_cache_7_internal_tensor_assign_1_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_7_internal_tensor_assign_1_end_mask_0 = const()[name = tensor("new_cache_7_internal_tensor_assign_1_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_7_internal_tensor_assign_1_squeeze_mask_0 = const()[name = tensor("new_cache_7_internal_tensor_assign_1_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_16 = const()[name = tensor("shape_16"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_6 = const()[name = tensor("reduce_prod_6"), val = tensor(1048576)]; + tensor range_1d_6_start_0 = const()[name = tensor("range_1d_6_start_0"), val = tensor(0)]; + tensor range_1d_6_step_0 = const()[name = tensor("range_1d_6_step_0"), val = tensor(1)]; + tensor range_1d_6 = range_1d(end = reduce_prod_6, start = range_1d_6_start_0, step = range_1d_6_step_0)[name = tensor("range_1d_6")]; + tensor reshape_30 = reshape(shape = shape_16, x = range_1d_6)[name = tensor("reshape_30")]; + tensor slice_by_index_6 = slice_by_index(begin = concat_26, begin_mask = new_cache_7_internal_tensor_assign_1_begin_mask_0, end = concat_27, end_mask = new_cache_7_internal_tensor_assign_1_end_mask_0, squeeze_mask = new_cache_7_internal_tensor_assign_1_squeeze_mask_0, stride = new_cache_7_internal_tensor_assign_1_stride_0, x = reshape_30)[name = tensor("slice_by_index_6")]; + tensor reshape_31_shape_0 = const()[name = tensor("reshape_31_shape_0"), val = tensor([-1])]; + tensor reshape_31 = reshape(shape = reshape_31_shape_0, x = slice_by_index_6)[name = tensor("reshape_31")]; + tensor reshape_32_shape_0 = const()[name = tensor("reshape_32_shape_0"), val = tensor([-1])]; + tensor reshape_32_cast_fp16 = reshape(shape = reshape_32_shape_0, x = var_1531_cast_fp16)[name = tensor("reshape_32_cast_fp16")]; + tensor reshape_33_shape_0 = const()[name = tensor("reshape_33_shape_0"), val = tensor([-1])]; + tensor reshape_33_cast_fp16 = reshape(shape = reshape_33_shape_0, x = cache3_to_fp16)[name = tensor("reshape_33_cast_fp16")]; + tensor scatter_6_mode_0 = const()[name = tensor("scatter_6_mode_0"), val = tensor("update")]; + tensor scatter_6_axis_0 = const()[name = tensor("scatter_6_axis_0"), val = tensor(0)]; + tensor scatter_6_validate_indices_0 = const()[name = tensor("scatter_6_validate_indices_0"), val = tensor(false)]; + tensor scatter_6_cast_fp16 = scatter(axis = scatter_6_axis_0, data = reshape_33_cast_fp16, indices = reshape_31, mode = scatter_6_mode_0, updates = reshape_32_cast_fp16, validate_indices = scatter_6_validate_indices_0)[name = tensor("scatter_6_cast_fp16")]; + tensor reshape_34_cast_fp16 = reshape(shape = shape_16, x = scatter_6_cast_fp16)[name = tensor("reshape_34_cast_fp16")]; + tensor var_1539_begin_0 = const()[name = tensor("op_1539_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor var_1539_end_0 = const()[name = tensor("op_1539_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor var_1539_end_mask_0 = const()[name = tensor("op_1539_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_1539_squeeze_mask_0 = const()[name = tensor("op_1539_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor var_1539_cast_fp16 = slice_by_index(begin = var_1539_begin_0, end = var_1539_end_0, end_mask = var_1539_end_mask_0, squeeze_mask = var_1539_squeeze_mask_0, x = reshape_34_cast_fp16)[name = tensor("op_1539_cast_fp16")]; + tensor var_1541_axis_0 = const()[name = tensor("op_1541_axis_0"), val = tensor(1)]; + tensor var_1541_mode_0 = const()[name = tensor("op_1541_mode_0"), val = tensor("update")]; + tensor var_1541_validate_indices_0 = const()[name = tensor("op_1541_validate_indices_0"), val = tensor(false)]; + tensor var_1541_cast_fp16 = scatter_along_axis(axis = var_1541_axis_0, data = var_1539_cast_fp16, indices = write_indices_7, mode = var_1541_mode_0, updates = v_7_cast_fp16, validate_indices = var_1541_validate_indices_0)[name = tensor("op_1541_cast_fp16")]; + tensor concat_28 = const()[name = tensor("concat_28"), val = tensor([1, 0, 0, 0, 0])]; + tensor concat_29 = const()[name = tensor("concat_29"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_7_internal_tensor_assign_2_stride_0 = const()[name = tensor("new_cache_7_internal_tensor_assign_2_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_7_internal_tensor_assign_2_begin_mask_0 = const()[name = tensor("new_cache_7_internal_tensor_assign_2_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_7_internal_tensor_assign_2_end_mask_0 = const()[name = tensor("new_cache_7_internal_tensor_assign_2_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_7_internal_tensor_assign_2_squeeze_mask_0 = const()[name = tensor("new_cache_7_internal_tensor_assign_2_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_17 = const()[name = tensor("shape_17"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_7 = const()[name = tensor("reduce_prod_7"), val = tensor(1048576)]; + tensor range_1d_7_start_0 = const()[name = tensor("range_1d_7_start_0"), val = tensor(0)]; + tensor range_1d_7_step_0 = const()[name = tensor("range_1d_7_step_0"), val = tensor(1)]; + tensor range_1d_7 = range_1d(end = reduce_prod_7, start = range_1d_7_start_0, step = range_1d_7_step_0)[name = tensor("range_1d_7")]; + tensor reshape_35 = reshape(shape = shape_17, x = range_1d_7)[name = tensor("reshape_35")]; + tensor slice_by_index_7 = slice_by_index(begin = concat_28, begin_mask = new_cache_7_internal_tensor_assign_2_begin_mask_0, end = concat_29, end_mask = new_cache_7_internal_tensor_assign_2_end_mask_0, squeeze_mask = new_cache_7_internal_tensor_assign_2_squeeze_mask_0, stride = new_cache_7_internal_tensor_assign_2_stride_0, x = reshape_35)[name = tensor("slice_by_index_7")]; + tensor reshape_36_shape_0 = const()[name = tensor("reshape_36_shape_0"), val = tensor([-1])]; + tensor reshape_36 = reshape(shape = reshape_36_shape_0, x = slice_by_index_7)[name = tensor("reshape_36")]; + tensor reshape_37_shape_0 = const()[name = tensor("reshape_37_shape_0"), val = tensor([-1])]; + tensor reshape_37_cast_fp16 = reshape(shape = reshape_37_shape_0, x = var_1541_cast_fp16)[name = tensor("reshape_37_cast_fp16")]; + tensor reshape_38_shape_0 = const()[name = tensor("reshape_38_shape_0"), val = tensor([-1])]; + tensor reshape_38_cast_fp16 = reshape(shape = reshape_38_shape_0, x = reshape_34_cast_fp16)[name = tensor("reshape_38_cast_fp16")]; + tensor scatter_7_mode_0 = const()[name = tensor("scatter_7_mode_0"), val = tensor("update")]; + tensor scatter_7_axis_0 = const()[name = tensor("scatter_7_axis_0"), val = tensor(0)]; + tensor scatter_7_validate_indices_0 = const()[name = tensor("scatter_7_validate_indices_0"), val = tensor(false)]; + tensor scatter_7_cast_fp16 = scatter(axis = scatter_7_axis_0, data = reshape_38_cast_fp16, indices = reshape_36, mode = scatter_7_mode_0, updates = reshape_37_cast_fp16, validate_indices = scatter_7_validate_indices_0)[name = tensor("scatter_7_cast_fp16")]; + tensor reshape_39_cast_fp16 = reshape(shape = shape_17, x = scatter_7_cast_fp16)[name = tensor("reshape_39_cast_fp16")]; + tensor reshape_39_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("reshape_39_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor keys_19_begin_0 = const()[name = tensor("keys_19_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor keys_19_end_0 = const()[name = tensor("keys_19_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor keys_19_end_mask_0 = const()[name = tensor("keys_19_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor keys_19_squeeze_mask_0 = const()[name = tensor("keys_19_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor keys_19_cast_fp16 = slice_by_index(begin = keys_19_begin_0, end = keys_19_end_0, end_mask = keys_19_end_mask_0, squeeze_mask = keys_19_squeeze_mask_0, x = reshape_39_cast_fp16)[name = tensor("keys_19_cast_fp16")]; + tensor values_19_begin_0 = const()[name = tensor("values_19_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor values_19_end_0 = const()[name = tensor("values_19_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor values_19_end_mask_0 = const()[name = tensor("values_19_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor values_19_squeeze_mask_0 = const()[name = tensor("values_19_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor values_19_cast_fp16 = slice_by_index(begin = values_19_begin_0, end = values_19_end_0, end_mask = values_19_end_mask_0, squeeze_mask = values_19_squeeze_mask_0, x = reshape_39_cast_fp16)[name = tensor("values_19_cast_fp16")]; + tensor var_1553_cast_fp16 = not_equal(x = keys_19_cast_fp16, y = keys_19_cast_fp16)[name = tensor("op_1553_cast_fp16")]; + tensor keys_21_cast_fp16 = select(a = var_326_to_fp16, b = keys_19_cast_fp16, cond = var_1553_cast_fp16)[name = tensor("keys_21_cast_fp16")]; + tensor var_1561_cast_fp16 = not_equal(x = values_19_cast_fp16, y = values_19_cast_fp16)[name = tensor("op_1561_cast_fp16")]; + tensor values_21_cast_fp16 = select(a = var_326_to_fp16, b = values_19_cast_fp16, cond = var_1561_cast_fp16)[name = tensor("values_21_cast_fp16")]; + tensor var_1585 = const()[name = tensor("op_1585"), val = tensor([0, 2, 1, 3])]; + tensor var_1601 = const()[name = tensor("op_1601"), val = tensor([1, 1, 1])]; + tensor var_1602_cast_fp16 = reshape(shape = var_1601, x = position3_to_fp16)[name = tensor("op_1602_cast_fp16")]; + tensor q_positions_7_cast_fp16 = add(x = var_1602_cast_fp16, y = q_offsets_1_promoted_to_fp16)[name = tensor("q_positions_7_cast_fp16")]; + tensor valid_end_7_cast_fp16 = add(x = var_1602_cast_fp16, y = var_392_cast_fp16)[name = tensor("valid_end_7_cast_fp16")]; + tensor valid_mask_7_cast_fp16 = less(x = k_positions_1_promoted_to_fp16, y = valid_end_7_cast_fp16)[name = tensor("valid_mask_7_cast_fp16")]; + tensor causal_mask_7_cast_fp16 = less_equal(x = k_positions_1_promoted_to_fp16, y = q_positions_7_cast_fp16)[name = tensor("causal_mask_7_cast_fp16")]; + tensor var_1630 = logical_and(x = valid_mask_7_cast_fp16, y = causal_mask_7_cast_fp16)[name = tensor("op_1630")]; + tensor attn_mask_7_axes_0 = const()[name = tensor("attn_mask_7_axes_0"), val = tensor([1])]; + tensor attn_mask_7 = expand_dims(axes = attn_mask_7_axes_0, x = var_1630)[name = tensor("attn_mask_7")]; + tensor var_1644_transpose_x_0 = const()[name = tensor("op_1644_transpose_x_0"), val = tensor(false)]; + tensor var_1644_transpose_y_0 = const()[name = tensor("op_1644_transpose_y_0"), val = tensor(false)]; + tensor transpose_21_perm_0 = const()[name = tensor("transpose_21_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_22_perm_0 = const()[name = tensor("transpose_22_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_22 = transpose(perm = transpose_22_perm_0, x = keys_21_cast_fp16)[name = tensor("transpose_30")]; + tensor transpose_21 = transpose(perm = transpose_21_perm_0, x = q_21_cast_fp16)[name = tensor("transpose_31")]; + tensor var_1644_cast_fp16 = matmul(transpose_x = var_1644_transpose_x_0, transpose_y = var_1644_transpose_y_0, x = transpose_21, y = transpose_22)[name = tensor("op_1644_cast_fp16")]; + tensor var_1638_to_fp16 = const()[name = tensor("op_1638_to_fp16"), val = tensor([0x1p-3])]; + tensor attn_weights_19_cast_fp16 = mul(x = var_1644_cast_fp16, y = var_1638_to_fp16)[name = tensor("attn_weights_19_cast_fp16")]; + tensor var_1646 = logical_not(x = attn_mask_7)[name = tensor("op_1646")]; + tensor var_1647_to_fp16 = const()[name = tensor("op_1647_to_fp16"), val = tensor(-inf)]; + tensor attn_weights_21_cast_fp16 = select(a = var_1647_to_fp16, b = attn_weights_19_cast_fp16, cond = var_1646)[name = tensor("attn_weights_21_cast_fp16")]; + tensor var_1649 = const()[name = tensor("op_1649"), val = tensor(-1)]; + tensor attn_weights_23_cast_fp16 = softmax(axis = var_1649, x = attn_weights_21_cast_fp16)[name = tensor("attn_weights_23_cast_fp16")]; + tensor attn_output_7_transpose_x_0 = const()[name = tensor("attn_output_7_transpose_x_0"), val = tensor(false)]; + tensor attn_output_7_transpose_y_0 = const()[name = tensor("attn_output_7_transpose_y_0"), val = tensor(false)]; + tensor values_23_cast_fp16 = transpose(perm = var_1585, x = values_21_cast_fp16)[name = tensor("transpose_32")]; + tensor attn_output_7_cast_fp16 = matmul(transpose_x = attn_output_7_transpose_x_0, transpose_y = attn_output_7_transpose_y_0, x = attn_weights_23_cast_fp16, y = values_23_cast_fp16)[name = tensor("attn_output_7_cast_fp16")]; + tensor var_1657 = const()[name = tensor("op_1657"), val = tensor([0, 2, 1, 3])]; + tensor var_1662 = const()[name = tensor("op_1662"), val = tensor([1, 256, 1024])]; + tensor var_1658_cast_fp16 = transpose(perm = var_1657, x = attn_output_7_cast_fp16)[name = tensor("transpose_29")]; + tensor input_41_cast_fp16 = reshape(shape = var_1662, x = var_1658_cast_fp16)[name = tensor("input_41_cast_fp16")]; + tensor attn3_out_proj_weight_to_fp16 = const()[name = tensor("attn3_out_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(82887936)))]; + tensor linear_13_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = attn3_out_proj_weight_to_fp16, x = input_41_cast_fp16)[name = tensor("linear_13_cast_fp16")]; + tensor var_1671_cast_fp16 = add(x = position3_to_fp16, y = valid_len_to_fp16)[name = tensor("op_1671_cast_fp16")]; + tensor var_1671_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_1671_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor input_43_cast_fp16 = add(x = input_37_cast_fp16, y = linear_13_cast_fp16)[name = tensor("input_43_cast_fp16")]; + tensor input_45_axes_0 = const()[name = tensor("input_45_axes_0"), val = tensor([-1])]; + tensor norm3_2_weight_to_fp16 = const()[name = tensor("norm3_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(84985152)))]; + tensor norm3_2_bias_to_fp16 = const()[name = tensor("norm3_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(84987264)))]; + tensor var_1675_to_fp16 = const()[name = tensor("op_1675_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_45_cast_fp16 = layer_norm(axes = input_45_axes_0, beta = norm3_2_bias_to_fp16, epsilon = var_1675_to_fp16, gamma = norm3_2_weight_to_fp16, x = input_43_cast_fp16)[name = tensor("input_45_cast_fp16")]; + tensor linear3_1_weight_to_fp16 = const()[name = tensor("linear3_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(84989376)))]; + tensor linear_14_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = linear3_1_weight_to_fp16, x = input_45_cast_fp16)[name = tensor("linear_14_cast_fp16")]; + tensor input_47_mode_0 = const()[name = tensor("input_47_mode_0"), val = tensor("EXACT")]; + tensor input_47_cast_fp16 = gelu(mode = input_47_mode_0, x = linear_14_cast_fp16)[name = tensor("input_47_cast_fp16")]; + tensor linear3_2_weight_to_fp16 = const()[name = tensor("linear3_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(93378048)))]; + tensor linear_15_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = linear3_2_weight_to_fp16, x = input_47_cast_fp16)[name = tensor("linear_15_cast_fp16")]; + tensor input_49_cast_fp16 = add(x = input_43_cast_fp16, y = linear_15_cast_fp16)[name = tensor("input_49_cast_fp16")]; + tensor input_51_axes_0 = const()[name = tensor("input_51_axes_0"), val = tensor([-1])]; + tensor norm4_1_weight_to_fp16 = const()[name = tensor("norm4_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(101766720)))]; + tensor norm4_1_bias_to_fp16 = const()[name = tensor("norm4_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(101768832)))]; + tensor var_1692_to_fp16 = const()[name = tensor("op_1692_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_51_cast_fp16 = layer_norm(axes = input_51_axes_0, beta = norm4_1_bias_to_fp16, epsilon = var_1692_to_fp16, gamma = norm4_1_weight_to_fp16, x = input_49_cast_fp16)[name = tensor("input_51_cast_fp16")]; + tensor attn4_in_proj_weight_to_fp16 = const()[name = tensor("attn4_in_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(101770944)))]; + tensor linear_16_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = attn4_in_proj_weight_to_fp16, x = input_51_cast_fp16)[name = tensor("linear_16_cast_fp16")]; + tensor var_1706 = const()[name = tensor("op_1706"), val = tensor([1, 256, 3, 16, 64])]; + tensor qkv_9_cast_fp16 = reshape(shape = var_1706, x = linear_16_cast_fp16)[name = tensor("qkv_9_cast_fp16")]; + tensor q_25_begin_0 = const()[name = tensor("q_25_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor q_25_end_0 = const()[name = tensor("q_25_end_0"), val = tensor([1, 256, 1, 16, 64])]; + tensor q_25_end_mask_0 = const()[name = tensor("q_25_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor q_25_squeeze_mask_0 = const()[name = tensor("q_25_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor q_25_cast_fp16 = slice_by_index(begin = q_25_begin_0, end = q_25_end_0, end_mask = q_25_end_mask_0, squeeze_mask = q_25_squeeze_mask_0, x = qkv_9_cast_fp16)[name = tensor("q_25_cast_fp16")]; + tensor k_17_begin_0 = const()[name = tensor("k_17_begin_0"), val = tensor([0, 0, 1, 0, 0])]; + tensor k_17_end_0 = const()[name = tensor("k_17_end_0"), val = tensor([1, 256, 2, 16, 64])]; + tensor k_17_end_mask_0 = const()[name = tensor("k_17_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor k_17_squeeze_mask_0 = const()[name = tensor("k_17_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor k_17_cast_fp16 = slice_by_index(begin = k_17_begin_0, end = k_17_end_0, end_mask = k_17_end_mask_0, squeeze_mask = k_17_squeeze_mask_0, x = qkv_9_cast_fp16)[name = tensor("k_17_cast_fp16")]; + tensor v_9_begin_0 = const()[name = tensor("v_9_begin_0"), val = tensor([0, 0, 2, 0, 0])]; + tensor v_9_end_0 = const()[name = tensor("v_9_end_0"), val = tensor([1, 256, 3, 16, 64])]; + tensor v_9_end_mask_0 = const()[name = tensor("v_9_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor v_9_squeeze_mask_0 = const()[name = tensor("v_9_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor v_9_cast_fp16 = slice_by_index(begin = v_9_begin_0, end = v_9_end_0, end_mask = v_9_end_mask_0, squeeze_mask = v_9_squeeze_mask_0, x = qkv_9_cast_fp16)[name = tensor("v_9_cast_fp16")]; + tensor var_1764 = const()[name = tensor("op_1764"), val = tensor([1, 1])]; + tensor position4_to_fp16_dtype_0 = const()[name = tensor("position4_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor position4_to_fp16 = cast(dtype = position4_to_fp16_dtype_0, x = position4)[name = tensor("cast_137")]; + tensor var_1765_cast_fp16 = reshape(shape = var_1764, x = position4_to_fp16)[name = tensor("op_1765_cast_fp16")]; + tensor ts_27_cast_fp16 = add(x = ts_1_promoted_to_fp16, y = var_1765_cast_fp16)[name = tensor("ts_27_cast_fp16")]; + tensor var_1772 = const()[name = tensor("op_1772"), val = tensor([1, 256, 1, 1])]; + tensor ts_29_cast_fp16 = reshape(shape = var_1772, x = ts_27_cast_fp16)[name = tensor("ts_29_cast_fp16")]; + tensor var_1779 = const()[name = tensor("op_1779"), val = tensor([1, 256, 16, 32, 2])]; + tensor q_complex_9_cast_fp16 = reshape(shape = var_1779, x = q_25_cast_fp16)[name = tensor("q_complex_9_cast_fp16")]; + tensor var_1786 = const()[name = tensor("op_1786"), val = tensor([1, 256, 16, 32, 2])]; + tensor k_complex_9_cast_fp16 = reshape(shape = var_1786, x = k_17_cast_fp16)[name = tensor("k_complex_9_cast_fp16")]; + tensor var_1790_begin_0 = const()[name = tensor("op_1790_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_1790_end_0 = const()[name = tensor("op_1790_end_0"), val = tensor([1, 256, 16, 32, 1])]; + tensor var_1790_end_mask_0 = const()[name = tensor("op_1790_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_1790_squeeze_mask_0 = const()[name = tensor("op_1790_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_1790_cast_fp16 = slice_by_index(begin = var_1790_begin_0, end = var_1790_end_0, end_mask = var_1790_end_mask_0, squeeze_mask = var_1790_squeeze_mask_0, x = q_complex_9_cast_fp16)[name = tensor("op_1790_cast_fp16")]; + tensor var_1798_begin_0 = const()[name = tensor("op_1798_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_1798_end_0 = const()[name = tensor("op_1798_end_0"), val = tensor([1, 256, 16, 32, 2])]; + tensor var_1798_end_mask_0 = const()[name = tensor("op_1798_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_1798_squeeze_mask_0 = const()[name = tensor("op_1798_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_1798_cast_fp16 = slice_by_index(begin = var_1798_begin_0, end = var_1798_end_0, end_mask = var_1798_end_mask_0, squeeze_mask = var_1798_squeeze_mask_0, x = q_complex_9_cast_fp16)[name = tensor("op_1798_cast_fp16")]; + tensor var_1806_begin_0 = const()[name = tensor("op_1806_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_1806_end_0 = const()[name = tensor("op_1806_end_0"), val = tensor([1, 256, 16, 32, 1])]; + tensor var_1806_end_mask_0 = const()[name = tensor("op_1806_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_1806_squeeze_mask_0 = const()[name = tensor("op_1806_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_1806_cast_fp16 = slice_by_index(begin = var_1806_begin_0, end = var_1806_end_0, end_mask = var_1806_end_mask_0, squeeze_mask = var_1806_squeeze_mask_0, x = k_complex_9_cast_fp16)[name = tensor("op_1806_cast_fp16")]; + tensor var_1814_begin_0 = const()[name = tensor("op_1814_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_1814_end_0 = const()[name = tensor("op_1814_end_0"), val = tensor([1, 256, 16, 32, 2])]; + tensor var_1814_end_mask_0 = const()[name = tensor("op_1814_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_1814_squeeze_mask_0 = const()[name = tensor("op_1814_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_1814_cast_fp16 = slice_by_index(begin = var_1814_begin_0, end = var_1814_end_0, end_mask = var_1814_end_mask_0, squeeze_mask = var_1814_squeeze_mask_0, x = k_complex_9_cast_fp16)[name = tensor("op_1814_cast_fp16")]; + tensor freqs_9_to_fp16 = const()[name = tensor("freqs_9_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(108062464)))]; + tensor var_1820_cast_fp16 = mul(x = freqs_9_to_fp16, y = ts_29_cast_fp16)[name = tensor("op_1820_cast_fp16")]; + tensor rotr_9_cast_fp16 = cos(x = var_1820_cast_fp16)[name = tensor("rotr_9_cast_fp16")]; + tensor roti_9_cast_fp16 = sin(x = var_1820_cast_fp16)[name = tensor("roti_9_cast_fp16")]; + tensor var_1824_cast_fp16 = mul(x = var_1790_cast_fp16, y = rotr_9_cast_fp16)[name = tensor("op_1824_cast_fp16")]; + tensor var_1825_cast_fp16 = mul(x = var_1798_cast_fp16, y = roti_9_cast_fp16)[name = tensor("op_1825_cast_fp16")]; + tensor qor_17_cast_fp16 = sub(x = var_1824_cast_fp16, y = var_1825_cast_fp16)[name = tensor("qor_17_cast_fp16")]; + tensor var_1828_cast_fp16 = mul(x = var_1790_cast_fp16, y = roti_9_cast_fp16)[name = tensor("op_1828_cast_fp16")]; + tensor var_1829_cast_fp16 = mul(x = var_1798_cast_fp16, y = rotr_9_cast_fp16)[name = tensor("op_1829_cast_fp16")]; + tensor qoi_17_cast_fp16 = add(x = var_1828_cast_fp16, y = var_1829_cast_fp16)[name = tensor("qoi_17_cast_fp16")]; + tensor var_1832_cast_fp16 = mul(x = var_1806_cast_fp16, y = rotr_9_cast_fp16)[name = tensor("op_1832_cast_fp16")]; + tensor var_1833_cast_fp16 = mul(x = var_1814_cast_fp16, y = roti_9_cast_fp16)[name = tensor("op_1833_cast_fp16")]; + tensor kor_17_cast_fp16 = sub(x = var_1832_cast_fp16, y = var_1833_cast_fp16)[name = tensor("kor_17_cast_fp16")]; + tensor var_1836_cast_fp16 = mul(x = var_1806_cast_fp16, y = roti_9_cast_fp16)[name = tensor("op_1836_cast_fp16")]; + tensor var_1837_cast_fp16 = mul(x = var_1814_cast_fp16, y = rotr_9_cast_fp16)[name = tensor("op_1837_cast_fp16")]; + tensor koi_17_cast_fp16 = add(x = var_1836_cast_fp16, y = var_1837_cast_fp16)[name = tensor("koi_17_cast_fp16")]; + tensor qo_9_axis_0 = const()[name = tensor("qo_9_axis_0"), val = tensor(-1)]; + tensor qo_9_cast_fp16 = stack(axis = qo_9_axis_0, values = (qor_17_cast_fp16, qoi_17_cast_fp16))[name = tensor("qo_9_cast_fp16")]; + tensor ko_9_axis_0 = const()[name = tensor("ko_9_axis_0"), val = tensor(-1)]; + tensor ko_9_cast_fp16 = stack(axis = ko_9_axis_0, values = (kor_17_cast_fp16, koi_17_cast_fp16))[name = tensor("ko_9_cast_fp16")]; + tensor var_1870 = const()[name = tensor("op_1870"), val = tensor([1, 256, 16, 64])]; + tensor q_27_cast_fp16 = reshape(shape = var_1870, x = qo_9_cast_fp16)[name = tensor("q_27_cast_fp16")]; + tensor var_1876 = const()[name = tensor("op_1876"), val = tensor([1, 256, 16, 64])]; + tensor k_19_cast_fp16 = reshape(shape = var_1876, x = ko_9_cast_fp16)[name = tensor("k_19_cast_fp16")]; + tensor write_indices_float_25_cast_fp16 = add(x = var_1765_cast_fp16, y = var_249_promoted_to_fp16)[name = tensor("write_indices_float_25_cast_fp16")]; + tensor _inversed_1897_y_0_to_fp16 = const()[name = tensor("_inversed_1897_y_0_to_fp16"), val = tensor(0x1p-9)]; + tensor _inversed_1897_cast_fp16 = mul(x = write_indices_float_25_cast_fp16, y = _inversed_1897_y_0_to_fp16)[name = tensor("_inversed_1897_cast_fp16")]; + tensor var_1898_cast_fp16 = floor(x = _inversed_1897_cast_fp16)[name = tensor("op_1898_cast_fp16")]; + tensor var_1899_to_fp16 = const()[name = tensor("op_1899_to_fp16"), val = tensor(0x1p+9)]; + tensor var_1900_cast_fp16 = mul(x = var_1898_cast_fp16, y = var_1899_to_fp16)[name = tensor("op_1900_cast_fp16")]; + tensor write_indices_float_27_cast_fp16 = sub(x = write_indices_float_25_cast_fp16, y = var_1900_cast_fp16)[name = tensor("write_indices_float_27_cast_fp16")]; + tensor var_1918_value_0_to_fp16 = const()[name = tensor("op_1918_value_0_to_fp16"), val = tensor(0x1.ffp+8)]; + tensor var_1918_cast_fp16 = fill_like(ref_tensor = write_indices_float_27_cast_fp16, value = var_1918_value_0_to_fp16)[name = tensor("op_1918_cast_fp16")]; + tensor write_indices_float_29_cast_fp16 = select(a = write_indices_float_27_cast_fp16, b = var_1918_cast_fp16, cond = token_valid_1_cast_fp16)[name = tensor("write_indices_float_29_cast_fp16")]; + tensor var_1924_dtype_0 = const()[name = tensor("op_1924_dtype_0"), val = tensor("int32")]; + tensor var_1929 = const()[name = tensor("op_1929"), val = tensor([1, 256, 1, 1])]; + tensor write_indices_float_29_cast_fp16_to_int32 = cast(dtype = var_1924_dtype_0, x = write_indices_float_29_cast_fp16)[name = tensor("cast_136")]; + tensor var_1930 = reshape(shape = var_1929, x = write_indices_float_29_cast_fp16_to_int32)[name = tensor("op_1930")]; + tensor write_indices_9_reps_0 = const()[name = tensor("write_indices_9_reps_0"), val = tensor([1, 1, 16, 64])]; + tensor write_indices_9 = tile(reps = write_indices_9_reps_0, x = var_1930)[name = tensor("write_indices_9")]; + tensor var_1940_begin_0 = const()[name = tensor("op_1940_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_1940_end_0 = const()[name = tensor("op_1940_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor var_1940_end_mask_0 = const()[name = tensor("op_1940_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_1940_squeeze_mask_0 = const()[name = tensor("op_1940_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor cache4_to_fp16_dtype_0 = const()[name = tensor("cache4_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor cache4_to_fp16 = cast(dtype = cache4_to_fp16_dtype_0, x = cache4)[name = tensor("cast_135")]; + tensor var_1940_cast_fp16 = slice_by_index(begin = var_1940_begin_0, end = var_1940_end_0, end_mask = var_1940_end_mask_0, squeeze_mask = var_1940_squeeze_mask_0, x = cache4_to_fp16)[name = tensor("op_1940_cast_fp16")]; + tensor var_1942_axis_0 = const()[name = tensor("op_1942_axis_0"), val = tensor(1)]; + tensor var_1942_mode_0 = const()[name = tensor("op_1942_mode_0"), val = tensor("update")]; + tensor var_1942_validate_indices_0 = const()[name = tensor("op_1942_validate_indices_0"), val = tensor(false)]; + tensor var_1942_cast_fp16 = scatter_along_axis(axis = var_1942_axis_0, data = var_1940_cast_fp16, indices = write_indices_9, mode = var_1942_mode_0, updates = k_19_cast_fp16, validate_indices = var_1942_validate_indices_0)[name = tensor("op_1942_cast_fp16")]; + tensor concat_34 = const()[name = tensor("concat_34"), val = tensor([0, 0, 0, 0, 0])]; + tensor concat_35 = const()[name = tensor("concat_35"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_9_internal_tensor_assign_1_stride_0 = const()[name = tensor("new_cache_9_internal_tensor_assign_1_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_9_internal_tensor_assign_1_begin_mask_0 = const()[name = tensor("new_cache_9_internal_tensor_assign_1_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_9_internal_tensor_assign_1_end_mask_0 = const()[name = tensor("new_cache_9_internal_tensor_assign_1_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_9_internal_tensor_assign_1_squeeze_mask_0 = const()[name = tensor("new_cache_9_internal_tensor_assign_1_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_18 = const()[name = tensor("shape_18"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_8 = const()[name = tensor("reduce_prod_8"), val = tensor(1048576)]; + tensor range_1d_8_start_0 = const()[name = tensor("range_1d_8_start_0"), val = tensor(0)]; + tensor range_1d_8_step_0 = const()[name = tensor("range_1d_8_step_0"), val = tensor(1)]; + tensor range_1d_8 = range_1d(end = reduce_prod_8, start = range_1d_8_start_0, step = range_1d_8_step_0)[name = tensor("range_1d_8")]; + tensor reshape_40 = reshape(shape = shape_18, x = range_1d_8)[name = tensor("reshape_40")]; + tensor slice_by_index_8 = slice_by_index(begin = concat_34, begin_mask = new_cache_9_internal_tensor_assign_1_begin_mask_0, end = concat_35, end_mask = new_cache_9_internal_tensor_assign_1_end_mask_0, squeeze_mask = new_cache_9_internal_tensor_assign_1_squeeze_mask_0, stride = new_cache_9_internal_tensor_assign_1_stride_0, x = reshape_40)[name = tensor("slice_by_index_8")]; + tensor reshape_41_shape_0 = const()[name = tensor("reshape_41_shape_0"), val = tensor([-1])]; + tensor reshape_41 = reshape(shape = reshape_41_shape_0, x = slice_by_index_8)[name = tensor("reshape_41")]; + tensor reshape_42_shape_0 = const()[name = tensor("reshape_42_shape_0"), val = tensor([-1])]; + tensor reshape_42_cast_fp16 = reshape(shape = reshape_42_shape_0, x = var_1942_cast_fp16)[name = tensor("reshape_42_cast_fp16")]; + tensor reshape_43_shape_0 = const()[name = tensor("reshape_43_shape_0"), val = tensor([-1])]; + tensor reshape_43_cast_fp16 = reshape(shape = reshape_43_shape_0, x = cache4_to_fp16)[name = tensor("reshape_43_cast_fp16")]; + tensor scatter_8_mode_0 = const()[name = tensor("scatter_8_mode_0"), val = tensor("update")]; + tensor scatter_8_axis_0 = const()[name = tensor("scatter_8_axis_0"), val = tensor(0)]; + tensor scatter_8_validate_indices_0 = const()[name = tensor("scatter_8_validate_indices_0"), val = tensor(false)]; + tensor scatter_8_cast_fp16 = scatter(axis = scatter_8_axis_0, data = reshape_43_cast_fp16, indices = reshape_41, mode = scatter_8_mode_0, updates = reshape_42_cast_fp16, validate_indices = scatter_8_validate_indices_0)[name = tensor("scatter_8_cast_fp16")]; + tensor reshape_44_cast_fp16 = reshape(shape = shape_18, x = scatter_8_cast_fp16)[name = tensor("reshape_44_cast_fp16")]; + tensor var_1950_begin_0 = const()[name = tensor("op_1950_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor var_1950_end_0 = const()[name = tensor("op_1950_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor var_1950_end_mask_0 = const()[name = tensor("op_1950_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_1950_squeeze_mask_0 = const()[name = tensor("op_1950_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor var_1950_cast_fp16 = slice_by_index(begin = var_1950_begin_0, end = var_1950_end_0, end_mask = var_1950_end_mask_0, squeeze_mask = var_1950_squeeze_mask_0, x = reshape_44_cast_fp16)[name = tensor("op_1950_cast_fp16")]; + tensor var_1952_axis_0 = const()[name = tensor("op_1952_axis_0"), val = tensor(1)]; + tensor var_1952_mode_0 = const()[name = tensor("op_1952_mode_0"), val = tensor("update")]; + tensor var_1952_validate_indices_0 = const()[name = tensor("op_1952_validate_indices_0"), val = tensor(false)]; + tensor var_1952_cast_fp16 = scatter_along_axis(axis = var_1952_axis_0, data = var_1950_cast_fp16, indices = write_indices_9, mode = var_1952_mode_0, updates = v_9_cast_fp16, validate_indices = var_1952_validate_indices_0)[name = tensor("op_1952_cast_fp16")]; + tensor concat_36 = const()[name = tensor("concat_36"), val = tensor([1, 0, 0, 0, 0])]; + tensor concat_37 = const()[name = tensor("concat_37"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_9_internal_tensor_assign_2_stride_0 = const()[name = tensor("new_cache_9_internal_tensor_assign_2_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_9_internal_tensor_assign_2_begin_mask_0 = const()[name = tensor("new_cache_9_internal_tensor_assign_2_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_9_internal_tensor_assign_2_end_mask_0 = const()[name = tensor("new_cache_9_internal_tensor_assign_2_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_9_internal_tensor_assign_2_squeeze_mask_0 = const()[name = tensor("new_cache_9_internal_tensor_assign_2_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_19 = const()[name = tensor("shape_19"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_9 = const()[name = tensor("reduce_prod_9"), val = tensor(1048576)]; + tensor range_1d_9_start_0 = const()[name = tensor("range_1d_9_start_0"), val = tensor(0)]; + tensor range_1d_9_step_0 = const()[name = tensor("range_1d_9_step_0"), val = tensor(1)]; + tensor range_1d_9 = range_1d(end = reduce_prod_9, start = range_1d_9_start_0, step = range_1d_9_step_0)[name = tensor("range_1d_9")]; + tensor reshape_45 = reshape(shape = shape_19, x = range_1d_9)[name = tensor("reshape_45")]; + tensor slice_by_index_9 = slice_by_index(begin = concat_36, begin_mask = new_cache_9_internal_tensor_assign_2_begin_mask_0, end = concat_37, end_mask = new_cache_9_internal_tensor_assign_2_end_mask_0, squeeze_mask = new_cache_9_internal_tensor_assign_2_squeeze_mask_0, stride = new_cache_9_internal_tensor_assign_2_stride_0, x = reshape_45)[name = tensor("slice_by_index_9")]; + tensor reshape_46_shape_0 = const()[name = tensor("reshape_46_shape_0"), val = tensor([-1])]; + tensor reshape_46 = reshape(shape = reshape_46_shape_0, x = slice_by_index_9)[name = tensor("reshape_46")]; + tensor reshape_47_shape_0 = const()[name = tensor("reshape_47_shape_0"), val = tensor([-1])]; + tensor reshape_47_cast_fp16 = reshape(shape = reshape_47_shape_0, x = var_1952_cast_fp16)[name = tensor("reshape_47_cast_fp16")]; + tensor reshape_48_shape_0 = const()[name = tensor("reshape_48_shape_0"), val = tensor([-1])]; + tensor reshape_48_cast_fp16 = reshape(shape = reshape_48_shape_0, x = reshape_44_cast_fp16)[name = tensor("reshape_48_cast_fp16")]; + tensor scatter_9_mode_0 = const()[name = tensor("scatter_9_mode_0"), val = tensor("update")]; + tensor scatter_9_axis_0 = const()[name = tensor("scatter_9_axis_0"), val = tensor(0)]; + tensor scatter_9_validate_indices_0 = const()[name = tensor("scatter_9_validate_indices_0"), val = tensor(false)]; + tensor scatter_9_cast_fp16 = scatter(axis = scatter_9_axis_0, data = reshape_48_cast_fp16, indices = reshape_46, mode = scatter_9_mode_0, updates = reshape_47_cast_fp16, validate_indices = scatter_9_validate_indices_0)[name = tensor("scatter_9_cast_fp16")]; + tensor reshape_49_cast_fp16 = reshape(shape = shape_19, x = scatter_9_cast_fp16)[name = tensor("reshape_49_cast_fp16")]; + tensor reshape_49_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("reshape_49_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor keys_25_begin_0 = const()[name = tensor("keys_25_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor keys_25_end_0 = const()[name = tensor("keys_25_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor keys_25_end_mask_0 = const()[name = tensor("keys_25_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor keys_25_squeeze_mask_0 = const()[name = tensor("keys_25_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor keys_25_cast_fp16 = slice_by_index(begin = keys_25_begin_0, end = keys_25_end_0, end_mask = keys_25_end_mask_0, squeeze_mask = keys_25_squeeze_mask_0, x = reshape_49_cast_fp16)[name = tensor("keys_25_cast_fp16")]; + tensor values_25_begin_0 = const()[name = tensor("values_25_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor values_25_end_0 = const()[name = tensor("values_25_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor values_25_end_mask_0 = const()[name = tensor("values_25_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor values_25_squeeze_mask_0 = const()[name = tensor("values_25_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor values_25_cast_fp16 = slice_by_index(begin = values_25_begin_0, end = values_25_end_0, end_mask = values_25_end_mask_0, squeeze_mask = values_25_squeeze_mask_0, x = reshape_49_cast_fp16)[name = tensor("values_25_cast_fp16")]; + tensor var_1964_cast_fp16 = not_equal(x = keys_25_cast_fp16, y = keys_25_cast_fp16)[name = tensor("op_1964_cast_fp16")]; + tensor keys_27_cast_fp16 = select(a = var_326_to_fp16, b = keys_25_cast_fp16, cond = var_1964_cast_fp16)[name = tensor("keys_27_cast_fp16")]; + tensor var_1972_cast_fp16 = not_equal(x = values_25_cast_fp16, y = values_25_cast_fp16)[name = tensor("op_1972_cast_fp16")]; + tensor values_27_cast_fp16 = select(a = var_326_to_fp16, b = values_25_cast_fp16, cond = var_1972_cast_fp16)[name = tensor("values_27_cast_fp16")]; + tensor var_1996 = const()[name = tensor("op_1996"), val = tensor([0, 2, 1, 3])]; + tensor var_2012 = const()[name = tensor("op_2012"), val = tensor([1, 1, 1])]; + tensor var_2013_cast_fp16 = reshape(shape = var_2012, x = position4_to_fp16)[name = tensor("op_2013_cast_fp16")]; + tensor q_positions_9_cast_fp16 = add(x = var_2013_cast_fp16, y = q_offsets_1_promoted_to_fp16)[name = tensor("q_positions_9_cast_fp16")]; + tensor valid_end_9_cast_fp16 = add(x = var_2013_cast_fp16, y = var_392_cast_fp16)[name = tensor("valid_end_9_cast_fp16")]; + tensor valid_mask_9_cast_fp16 = less(x = k_positions_1_promoted_to_fp16, y = valid_end_9_cast_fp16)[name = tensor("valid_mask_9_cast_fp16")]; + tensor causal_mask_9_cast_fp16 = less_equal(x = k_positions_1_promoted_to_fp16, y = q_positions_9_cast_fp16)[name = tensor("causal_mask_9_cast_fp16")]; + tensor var_2041 = logical_and(x = valid_mask_9_cast_fp16, y = causal_mask_9_cast_fp16)[name = tensor("op_2041")]; + tensor attn_mask_9_axes_0 = const()[name = tensor("attn_mask_9_axes_0"), val = tensor([1])]; + tensor attn_mask_9 = expand_dims(axes = attn_mask_9_axes_0, x = var_2041)[name = tensor("attn_mask_9")]; + tensor var_2055_transpose_x_0 = const()[name = tensor("op_2055_transpose_x_0"), val = tensor(false)]; + tensor var_2055_transpose_y_0 = const()[name = tensor("op_2055_transpose_y_0"), val = tensor(false)]; + tensor transpose_23_perm_0 = const()[name = tensor("transpose_23_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_24_perm_0 = const()[name = tensor("transpose_24_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_24 = transpose(perm = transpose_24_perm_0, x = keys_27_cast_fp16)[name = tensor("transpose_26")]; + tensor transpose_23 = transpose(perm = transpose_23_perm_0, x = q_27_cast_fp16)[name = tensor("transpose_27")]; + tensor var_2055_cast_fp16 = matmul(transpose_x = var_2055_transpose_x_0, transpose_y = var_2055_transpose_y_0, x = transpose_23, y = transpose_24)[name = tensor("op_2055_cast_fp16")]; + tensor var_2049_to_fp16 = const()[name = tensor("op_2049_to_fp16"), val = tensor([0x1p-3])]; + tensor attn_weights_25_cast_fp16 = mul(x = var_2055_cast_fp16, y = var_2049_to_fp16)[name = tensor("attn_weights_25_cast_fp16")]; + tensor var_2057 = logical_not(x = attn_mask_9)[name = tensor("op_2057")]; + tensor var_2058_to_fp16 = const()[name = tensor("op_2058_to_fp16"), val = tensor(-inf)]; + tensor attn_weights_27_cast_fp16 = select(a = var_2058_to_fp16, b = attn_weights_25_cast_fp16, cond = var_2057)[name = tensor("attn_weights_27_cast_fp16")]; + tensor var_2060 = const()[name = tensor("op_2060"), val = tensor(-1)]; + tensor attn_weights_29_cast_fp16 = softmax(axis = var_2060, x = attn_weights_27_cast_fp16)[name = tensor("attn_weights_29_cast_fp16")]; + tensor attn_output_9_transpose_x_0 = const()[name = tensor("attn_output_9_transpose_x_0"), val = tensor(false)]; + tensor attn_output_9_transpose_y_0 = const()[name = tensor("attn_output_9_transpose_y_0"), val = tensor(false)]; + tensor values_29_cast_fp16 = transpose(perm = var_1996, x = values_27_cast_fp16)[name = tensor("transpose_28")]; + tensor attn_output_9_cast_fp16 = matmul(transpose_x = attn_output_9_transpose_x_0, transpose_y = attn_output_9_transpose_y_0, x = attn_weights_29_cast_fp16, y = values_29_cast_fp16)[name = tensor("attn_output_9_cast_fp16")]; + tensor var_2068 = const()[name = tensor("op_2068"), val = tensor([0, 2, 1, 3])]; + tensor var_2073 = const()[name = tensor("op_2073"), val = tensor([1, 256, 1024])]; + tensor var_2069_cast_fp16 = transpose(perm = var_2068, x = attn_output_9_cast_fp16)[name = tensor("transpose_25")]; + tensor input_53_cast_fp16 = reshape(shape = var_2073, x = var_2069_cast_fp16)[name = tensor("input_53_cast_fp16")]; + tensor attn4_out_proj_weight_to_fp16 = const()[name = tensor("attn4_out_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(108062592)))]; + tensor linear_17_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = attn4_out_proj_weight_to_fp16, x = input_53_cast_fp16)[name = tensor("linear_17_cast_fp16")]; + tensor var_2082_cast_fp16 = add(x = position4_to_fp16, y = valid_len_to_fp16)[name = tensor("op_2082_cast_fp16")]; + tensor var_2082_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_2082_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor input_55_cast_fp16 = add(x = input_49_cast_fp16, y = linear_17_cast_fp16)[name = tensor("input_55_cast_fp16")]; + tensor input_57_axes_0 = const()[name = tensor("input_57_axes_0"), val = tensor([-1])]; + tensor norm4_2_weight_to_fp16 = const()[name = tensor("norm4_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(110159808)))]; + tensor norm4_2_bias_to_fp16 = const()[name = tensor("norm4_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(110161920)))]; + tensor var_2086_to_fp16 = const()[name = tensor("op_2086_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_57_cast_fp16 = layer_norm(axes = input_57_axes_0, beta = norm4_2_bias_to_fp16, epsilon = var_2086_to_fp16, gamma = norm4_2_weight_to_fp16, x = input_55_cast_fp16)[name = tensor("input_57_cast_fp16")]; + tensor linear4_1_weight_to_fp16 = const()[name = tensor("linear4_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(110164032)))]; + tensor linear_18_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = linear4_1_weight_to_fp16, x = input_57_cast_fp16)[name = tensor("linear_18_cast_fp16")]; + tensor input_59_mode_0 = const()[name = tensor("input_59_mode_0"), val = tensor("EXACT")]; + tensor input_59_cast_fp16 = gelu(mode = input_59_mode_0, x = linear_18_cast_fp16)[name = tensor("input_59_cast_fp16")]; + tensor linear4_2_weight_to_fp16 = const()[name = tensor("linear4_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(118552704)))]; + tensor linear_19_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = linear4_2_weight_to_fp16, x = input_59_cast_fp16)[name = tensor("linear_19_cast_fp16")]; + tensor input_61_cast_fp16 = add(x = input_55_cast_fp16, y = linear_19_cast_fp16)[name = tensor("input_61_cast_fp16")]; + tensor input_63_axes_0 = const()[name = tensor("input_63_axes_0"), val = tensor([-1])]; + tensor norm5_1_weight_to_fp16 = const()[name = tensor("norm5_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(126941376)))]; + tensor norm5_1_bias_to_fp16 = const()[name = tensor("norm5_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(126943488)))]; + tensor var_2103_to_fp16 = const()[name = tensor("op_2103_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_63_cast_fp16 = layer_norm(axes = input_63_axes_0, beta = norm5_1_bias_to_fp16, epsilon = var_2103_to_fp16, gamma = norm5_1_weight_to_fp16, x = input_61_cast_fp16)[name = tensor("input_63_cast_fp16")]; + tensor attn5_in_proj_weight_to_fp16 = const()[name = tensor("attn5_in_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(126945600)))]; + tensor linear_20_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = attn5_in_proj_weight_to_fp16, x = input_63_cast_fp16)[name = tensor("linear_20_cast_fp16")]; + tensor var_2117 = const()[name = tensor("op_2117"), val = tensor([1, 256, 3, 16, 64])]; + tensor qkv_cast_fp16 = reshape(shape = var_2117, x = linear_20_cast_fp16)[name = tensor("qkv_cast_fp16")]; + tensor k_21_begin_0 = const()[name = tensor("k_21_begin_0"), val = tensor([0, 0, 1, 0, 0])]; + tensor k_21_end_0 = const()[name = tensor("k_21_end_0"), val = tensor([1, 256, 2, 16, 64])]; + tensor k_21_end_mask_0 = const()[name = tensor("k_21_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor k_21_squeeze_mask_0 = const()[name = tensor("k_21_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor k_21_cast_fp16 = slice_by_index(begin = k_21_begin_0, end = k_21_end_0, end_mask = k_21_end_mask_0, squeeze_mask = k_21_squeeze_mask_0, x = qkv_cast_fp16)[name = tensor("k_21_cast_fp16")]; + tensor v_begin_0 = const()[name = tensor("v_begin_0"), val = tensor([0, 0, 2, 0, 0])]; + tensor v_end_0 = const()[name = tensor("v_end_0"), val = tensor([1, 256, 3, 16, 64])]; + tensor v_end_mask_0 = const()[name = tensor("v_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor v_squeeze_mask_0 = const()[name = tensor("v_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor v_cast_fp16 = slice_by_index(begin = v_begin_0, end = v_end_0, end_mask = v_end_mask_0, squeeze_mask = v_squeeze_mask_0, x = qkv_cast_fp16)[name = tensor("v_cast_fp16")]; + tensor var_2162 = const()[name = tensor("op_2162"), val = tensor([1, 1])]; + tensor position5_to_fp16_dtype_0 = const()[name = tensor("position5_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor position5_to_fp16 = cast(dtype = position5_to_fp16_dtype_0, x = position5)[name = tensor("cast_132")]; + tensor var_2163_cast_fp16 = reshape(shape = var_2162, x = position5_to_fp16)[name = tensor("op_2163_cast_fp16")]; + tensor ts_33_cast_fp16 = add(x = ts_1_promoted_to_fp16, y = var_2163_cast_fp16)[name = tensor("ts_33_cast_fp16")]; + tensor var_2170 = const()[name = tensor("op_2170"), val = tensor([1, 256, 1, 1])]; + tensor ts_cast_fp16 = reshape(shape = var_2170, x = ts_33_cast_fp16)[name = tensor("ts_cast_fp16")]; + tensor var_2177 = const()[name = tensor("op_2177"), val = tensor([1, 256, 16, 32, 2])]; + tensor k_complex_cast_fp16 = reshape(shape = var_2177, x = k_21_cast_fp16)[name = tensor("k_complex_cast_fp16")]; + tensor var_2181_begin_0 = const()[name = tensor("op_2181_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_2181_end_0 = const()[name = tensor("op_2181_end_0"), val = tensor([1, 256, 16, 32, 1])]; + tensor var_2181_end_mask_0 = const()[name = tensor("op_2181_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_2181_squeeze_mask_0 = const()[name = tensor("op_2181_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_2181_cast_fp16 = slice_by_index(begin = var_2181_begin_0, end = var_2181_end_0, end_mask = var_2181_end_mask_0, squeeze_mask = var_2181_squeeze_mask_0, x = k_complex_cast_fp16)[name = tensor("op_2181_cast_fp16")]; + tensor var_2189_begin_0 = const()[name = tensor("op_2189_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_2189_end_0 = const()[name = tensor("op_2189_end_0"), val = tensor([1, 256, 16, 32, 2])]; + tensor var_2189_end_mask_0 = const()[name = tensor("op_2189_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_2189_squeeze_mask_0 = const()[name = tensor("op_2189_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_2189_cast_fp16 = slice_by_index(begin = var_2189_begin_0, end = var_2189_end_0, end_mask = var_2189_end_mask_0, squeeze_mask = var_2189_squeeze_mask_0, x = k_complex_cast_fp16)[name = tensor("op_2189_cast_fp16")]; + tensor freqs_to_fp16 = const()[name = tensor("freqs_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(133237120)))]; + tensor var_2195_cast_fp16 = mul(x = freqs_to_fp16, y = ts_cast_fp16)[name = tensor("op_2195_cast_fp16")]; + tensor rotr_cast_fp16 = cos(x = var_2195_cast_fp16)[name = tensor("rotr_cast_fp16")]; + tensor roti_cast_fp16 = sin(x = var_2195_cast_fp16)[name = tensor("roti_cast_fp16")]; + tensor var_2199_cast_fp16 = mul(x = var_2181_cast_fp16, y = rotr_cast_fp16)[name = tensor("op_2199_cast_fp16")]; + tensor var_2200_cast_fp16 = mul(x = var_2189_cast_fp16, y = roti_cast_fp16)[name = tensor("op_2200_cast_fp16")]; + tensor kor_21_cast_fp16 = sub(x = var_2199_cast_fp16, y = var_2200_cast_fp16)[name = tensor("kor_21_cast_fp16")]; + tensor var_2203_cast_fp16 = mul(x = var_2181_cast_fp16, y = roti_cast_fp16)[name = tensor("op_2203_cast_fp16")]; + tensor var_2204_cast_fp16 = mul(x = var_2189_cast_fp16, y = rotr_cast_fp16)[name = tensor("op_2204_cast_fp16")]; + tensor koi_21_cast_fp16 = add(x = var_2203_cast_fp16, y = var_2204_cast_fp16)[name = tensor("koi_21_cast_fp16")]; + tensor ko_axis_0 = const()[name = tensor("ko_axis_0"), val = tensor(-1)]; + tensor ko_cast_fp16 = stack(axis = ko_axis_0, values = (kor_21_cast_fp16, koi_21_cast_fp16))[name = tensor("ko_cast_fp16")]; + tensor var_2224 = const()[name = tensor("op_2224"), val = tensor([1, 256, 16, 64])]; + tensor k_cast_fp16 = reshape(shape = var_2224, x = ko_cast_fp16)[name = tensor("k_cast_fp16")]; + tensor write_indices_float_31_cast_fp16 = add(x = var_2163_cast_fp16, y = var_249_promoted_to_fp16)[name = tensor("write_indices_float_31_cast_fp16")]; + tensor _inversed_2245_y_0_to_fp16 = const()[name = tensor("_inversed_2245_y_0_to_fp16"), val = tensor(0x1p-9)]; + tensor _inversed_2245_cast_fp16 = mul(x = write_indices_float_31_cast_fp16, y = _inversed_2245_y_0_to_fp16)[name = tensor("_inversed_2245_cast_fp16")]; + tensor var_2246_cast_fp16 = floor(x = _inversed_2245_cast_fp16)[name = tensor("op_2246_cast_fp16")]; + tensor var_2247_to_fp16 = const()[name = tensor("op_2247_to_fp16"), val = tensor(0x1p+9)]; + tensor var_2248_cast_fp16 = mul(x = var_2246_cast_fp16, y = var_2247_to_fp16)[name = tensor("op_2248_cast_fp16")]; + tensor write_indices_float_33_cast_fp16 = sub(x = write_indices_float_31_cast_fp16, y = var_2248_cast_fp16)[name = tensor("write_indices_float_33_cast_fp16")]; + tensor var_2266_value_0_to_fp16 = const()[name = tensor("op_2266_value_0_to_fp16"), val = tensor(0x1.ffp+8)]; + tensor var_2266_cast_fp16 = fill_like(ref_tensor = write_indices_float_33_cast_fp16, value = var_2266_value_0_to_fp16)[name = tensor("op_2266_cast_fp16")]; + tensor write_indices_float_cast_fp16 = select(a = write_indices_float_33_cast_fp16, b = var_2266_cast_fp16, cond = token_valid_1_cast_fp16)[name = tensor("write_indices_float_cast_fp16")]; + tensor var_2272_dtype_0 = const()[name = tensor("op_2272_dtype_0"), val = tensor("int32")]; + tensor var_2277 = const()[name = tensor("op_2277"), val = tensor([1, 256, 1, 1])]; + tensor write_indices_float_cast_fp16_to_int32 = cast(dtype = var_2272_dtype_0, x = write_indices_float_cast_fp16)[name = tensor("cast_131")]; + tensor var_2278 = reshape(shape = var_2277, x = write_indices_float_cast_fp16_to_int32)[name = tensor("op_2278")]; + tensor write_indices_reps_0 = const()[name = tensor("write_indices_reps_0"), val = tensor([1, 1, 16, 64])]; + tensor write_indices = tile(reps = write_indices_reps_0, x = var_2278)[name = tensor("write_indices")]; + tensor var_2288_begin_0 = const()[name = tensor("op_2288_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_2288_end_0 = const()[name = tensor("op_2288_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor var_2288_end_mask_0 = const()[name = tensor("op_2288_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_2288_squeeze_mask_0 = const()[name = tensor("op_2288_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor cache5_to_fp16_dtype_0 = const()[name = tensor("cache5_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor cache5_to_fp16 = cast(dtype = cache5_to_fp16_dtype_0, x = cache5)[name = tensor("cast_130")]; + tensor var_2288_cast_fp16 = slice_by_index(begin = var_2288_begin_0, end = var_2288_end_0, end_mask = var_2288_end_mask_0, squeeze_mask = var_2288_squeeze_mask_0, x = cache5_to_fp16)[name = tensor("op_2288_cast_fp16")]; + tensor var_2290_axis_0 = const()[name = tensor("op_2290_axis_0"), val = tensor(1)]; + tensor var_2290_mode_0 = const()[name = tensor("op_2290_mode_0"), val = tensor("update")]; + tensor var_2290_validate_indices_0 = const()[name = tensor("op_2290_validate_indices_0"), val = tensor(false)]; + tensor var_2290_cast_fp16 = scatter_along_axis(axis = var_2290_axis_0, data = var_2288_cast_fp16, indices = write_indices, mode = var_2290_mode_0, updates = k_cast_fp16, validate_indices = var_2290_validate_indices_0)[name = tensor("op_2290_cast_fp16")]; + tensor concat_42 = const()[name = tensor("concat_42"), val = tensor([0, 0, 0, 0, 0])]; + tensor concat_43 = const()[name = tensor("concat_43"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_internal_tensor_assign_1_stride_0 = const()[name = tensor("new_cache_internal_tensor_assign_1_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_internal_tensor_assign_1_begin_mask_0 = const()[name = tensor("new_cache_internal_tensor_assign_1_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_internal_tensor_assign_1_end_mask_0 = const()[name = tensor("new_cache_internal_tensor_assign_1_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_internal_tensor_assign_1_squeeze_mask_0 = const()[name = tensor("new_cache_internal_tensor_assign_1_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_20 = const()[name = tensor("shape_20"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_10 = const()[name = tensor("reduce_prod_10"), val = tensor(1048576)]; + tensor range_1d_10_start_0 = const()[name = tensor("range_1d_10_start_0"), val = tensor(0)]; + tensor range_1d_10_step_0 = const()[name = tensor("range_1d_10_step_0"), val = tensor(1)]; + tensor range_1d_10 = range_1d(end = reduce_prod_10, start = range_1d_10_start_0, step = range_1d_10_step_0)[name = tensor("range_1d_10")]; + tensor reshape_50 = reshape(shape = shape_20, x = range_1d_10)[name = tensor("reshape_50")]; + tensor slice_by_index_10 = slice_by_index(begin = concat_42, begin_mask = new_cache_internal_tensor_assign_1_begin_mask_0, end = concat_43, end_mask = new_cache_internal_tensor_assign_1_end_mask_0, squeeze_mask = new_cache_internal_tensor_assign_1_squeeze_mask_0, stride = new_cache_internal_tensor_assign_1_stride_0, x = reshape_50)[name = tensor("slice_by_index_10")]; + tensor reshape_51_shape_0 = const()[name = tensor("reshape_51_shape_0"), val = tensor([-1])]; + tensor reshape_51 = reshape(shape = reshape_51_shape_0, x = slice_by_index_10)[name = tensor("reshape_51")]; + tensor reshape_52_shape_0 = const()[name = tensor("reshape_52_shape_0"), val = tensor([-1])]; + tensor reshape_52_cast_fp16 = reshape(shape = reshape_52_shape_0, x = var_2290_cast_fp16)[name = tensor("reshape_52_cast_fp16")]; + tensor reshape_53_shape_0 = const()[name = tensor("reshape_53_shape_0"), val = tensor([-1])]; + tensor reshape_53_cast_fp16 = reshape(shape = reshape_53_shape_0, x = cache5_to_fp16)[name = tensor("reshape_53_cast_fp16")]; + tensor scatter_10_mode_0 = const()[name = tensor("scatter_10_mode_0"), val = tensor("update")]; + tensor scatter_10_axis_0 = const()[name = tensor("scatter_10_axis_0"), val = tensor(0)]; + tensor scatter_10_validate_indices_0 = const()[name = tensor("scatter_10_validate_indices_0"), val = tensor(false)]; + tensor scatter_10_cast_fp16 = scatter(axis = scatter_10_axis_0, data = reshape_53_cast_fp16, indices = reshape_51, mode = scatter_10_mode_0, updates = reshape_52_cast_fp16, validate_indices = scatter_10_validate_indices_0)[name = tensor("scatter_10_cast_fp16")]; + tensor reshape_54_cast_fp16 = reshape(shape = shape_20, x = scatter_10_cast_fp16)[name = tensor("reshape_54_cast_fp16")]; + tensor var_2298_begin_0 = const()[name = tensor("op_2298_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor var_2298_end_0 = const()[name = tensor("op_2298_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor var_2298_end_mask_0 = const()[name = tensor("op_2298_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_2298_squeeze_mask_0 = const()[name = tensor("op_2298_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor var_2298_cast_fp16 = slice_by_index(begin = var_2298_begin_0, end = var_2298_end_0, end_mask = var_2298_end_mask_0, squeeze_mask = var_2298_squeeze_mask_0, x = reshape_54_cast_fp16)[name = tensor("op_2298_cast_fp16")]; + tensor var_2300_axis_0 = const()[name = tensor("op_2300_axis_0"), val = tensor(1)]; + tensor var_2300_mode_0 = const()[name = tensor("op_2300_mode_0"), val = tensor("update")]; + tensor var_2300_validate_indices_0 = const()[name = tensor("op_2300_validate_indices_0"), val = tensor(false)]; + tensor var_2300_cast_fp16 = scatter_along_axis(axis = var_2300_axis_0, data = var_2298_cast_fp16, indices = write_indices, mode = var_2300_mode_0, updates = v_cast_fp16, validate_indices = var_2300_validate_indices_0)[name = tensor("op_2300_cast_fp16")]; + tensor concat_44 = const()[name = tensor("concat_44"), val = tensor([1, 0, 0, 0, 0])]; + tensor concat_45 = const()[name = tensor("concat_45"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_internal_tensor_assign_2_stride_0 = const()[name = tensor("new_cache_internal_tensor_assign_2_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_internal_tensor_assign_2_begin_mask_0 = const()[name = tensor("new_cache_internal_tensor_assign_2_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_internal_tensor_assign_2_end_mask_0 = const()[name = tensor("new_cache_internal_tensor_assign_2_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_internal_tensor_assign_2_squeeze_mask_0 = const()[name = tensor("new_cache_internal_tensor_assign_2_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_21 = const()[name = tensor("shape_21"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_11 = const()[name = tensor("reduce_prod_11"), val = tensor(1048576)]; + tensor range_1d_11_start_0 = const()[name = tensor("range_1d_11_start_0"), val = tensor(0)]; + tensor range_1d_11_step_0 = const()[name = tensor("range_1d_11_step_0"), val = tensor(1)]; + tensor range_1d_11 = range_1d(end = reduce_prod_11, start = range_1d_11_start_0, step = range_1d_11_step_0)[name = tensor("range_1d_11")]; + tensor reshape_55 = reshape(shape = shape_21, x = range_1d_11)[name = tensor("reshape_55")]; + tensor slice_by_index_11 = slice_by_index(begin = concat_44, begin_mask = new_cache_internal_tensor_assign_2_begin_mask_0, end = concat_45, end_mask = new_cache_internal_tensor_assign_2_end_mask_0, squeeze_mask = new_cache_internal_tensor_assign_2_squeeze_mask_0, stride = new_cache_internal_tensor_assign_2_stride_0, x = reshape_55)[name = tensor("slice_by_index_11")]; + tensor reshape_56_shape_0 = const()[name = tensor("reshape_56_shape_0"), val = tensor([-1])]; + tensor reshape_56 = reshape(shape = reshape_56_shape_0, x = slice_by_index_11)[name = tensor("reshape_56")]; + tensor reshape_57_shape_0 = const()[name = tensor("reshape_57_shape_0"), val = tensor([-1])]; + tensor reshape_57_cast_fp16 = reshape(shape = reshape_57_shape_0, x = var_2300_cast_fp16)[name = tensor("reshape_57_cast_fp16")]; + tensor reshape_58_shape_0 = const()[name = tensor("reshape_58_shape_0"), val = tensor([-1])]; + tensor reshape_58_cast_fp16 = reshape(shape = reshape_58_shape_0, x = reshape_54_cast_fp16)[name = tensor("reshape_58_cast_fp16")]; + tensor scatter_11_mode_0 = const()[name = tensor("scatter_11_mode_0"), val = tensor("update")]; + tensor scatter_11_axis_0 = const()[name = tensor("scatter_11_axis_0"), val = tensor(0)]; + tensor scatter_11_validate_indices_0 = const()[name = tensor("scatter_11_validate_indices_0"), val = tensor(false)]; + tensor scatter_11_cast_fp16 = scatter(axis = scatter_11_axis_0, data = reshape_58_cast_fp16, indices = reshape_56, mode = scatter_11_mode_0, updates = reshape_57_cast_fp16, validate_indices = scatter_11_validate_indices_0)[name = tensor("scatter_11_cast_fp16")]; + tensor reshape_59_cast_fp16 = reshape(shape = shape_21, x = scatter_11_cast_fp16)[name = tensor("reshape_59_cast_fp16")]; + tensor reshape_59_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("reshape_59_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor var_2310_cast_fp16 = add(x = position5_to_fp16, y = valid_len_to_fp16)[name = tensor("op_2310_cast_fp16")]; + tensor var_2310_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_2310_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor var_2310 = cast(dtype = var_2310_cast_fp16_to_fp32_dtype_0, x = var_2310_cast_fp16)[name = tensor("cast_128")]; + tensor new_cache_internal_tensor_assign_2 = cast(dtype = reshape_59_cast_fp16_to_fp32_dtype_0, x = reshape_59_cast_fp16)[name = tensor("cast_129")]; + tensor var_2082 = cast(dtype = var_2082_cast_fp16_to_fp32_dtype_0, x = var_2082_cast_fp16)[name = tensor("cast_133")]; + tensor new_cache_9_internal_tensor_assign_2 = cast(dtype = reshape_49_cast_fp16_to_fp32_dtype_0, x = reshape_49_cast_fp16)[name = tensor("cast_134")]; + tensor var_1671 = cast(dtype = var_1671_cast_fp16_to_fp32_dtype_0, x = var_1671_cast_fp16)[name = tensor("cast_138")]; + tensor new_cache_7_internal_tensor_assign_2 = cast(dtype = reshape_39_cast_fp16_to_fp32_dtype_0, x = reshape_39_cast_fp16)[name = tensor("cast_139")]; + tensor var_1260 = cast(dtype = var_1260_cast_fp16_to_fp32_dtype_0, x = var_1260_cast_fp16)[name = tensor("cast_143")]; + tensor new_cache_5_internal_tensor_assign_2 = cast(dtype = reshape_29_cast_fp16_to_fp32_dtype_0, x = reshape_29_cast_fp16)[name = tensor("cast_144")]; + tensor var_849 = cast(dtype = var_849_cast_fp16_to_fp32_dtype_0, x = var_849_cast_fp16)[name = tensor("cast_148")]; + tensor new_cache_3_internal_tensor_assign_2 = cast(dtype = reshape_19_cast_fp16_to_fp32_dtype_0, x = reshape_19_cast_fp16)[name = tensor("cast_149")]; + tensor var_438 = cast(dtype = var_438_cast_fp16_to_fp32_dtype_0, x = var_438_cast_fp16)[name = tensor("cast_153")]; + tensor new_cache_1_internal_tensor_assign_2 = cast(dtype = reshape_9_cast_fp16_to_fp32_dtype_0, x = reshape_9_cast_fp16)[name = tensor("cast_154")]; + } -> (new_cache_1_internal_tensor_assign_2, var_438, new_cache_3_internal_tensor_assign_2, var_849, new_cache_5_internal_tensor_assign_2, var_1260, new_cache_7_internal_tensor_assign_2, var_1671, new_cache_9_internal_tensor_assign_2, var_2082, new_cache_internal_tensor_assign_2, var_2310); +} \ No newline at end of file diff --git a/v2.1/portuguese/cond_prefill.mlmodelc/weights/weight.bin b/v2.1/portuguese/cond_prefill.mlmodelc/weights/weight.bin new file mode 100644 index 0000000000000000000000000000000000000000..54aa1a68a3d9f2950f8d58a6dd1ccb430b00dd1b --- /dev/null +++ b/v2.1/portuguese/cond_prefill.mlmodelc/weights/weight.bin @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:0d40b2e26baeac0a03cb98c944644fd9b12a1e1ec307aa177d8f667398198c8d +size 133237248 diff --git a/v2.1/portuguese/cond_prefill.mlpackage/Data/com.apple.CoreML/model.mlmodel b/v2.1/portuguese/cond_prefill.mlpackage/Data/com.apple.CoreML/model.mlmodel new file mode 100644 index 0000000000000000000000000000000000000000..0494c77552be0b0d367a971818af464232c05544 --- /dev/null +++ b/v2.1/portuguese/cond_prefill.mlpackage/Data/com.apple.CoreML/model.mlmodel @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:41c6e60a25ae4ff591146d1e8595d3293c60955e0e3e3208532005efaf07d389 +size 200480 diff --git a/v2.1/portuguese/cond_prefill.mlpackage/Data/com.apple.CoreML/weights/weight.bin b/v2.1/portuguese/cond_prefill.mlpackage/Data/com.apple.CoreML/weights/weight.bin new file mode 100644 index 0000000000000000000000000000000000000000..54aa1a68a3d9f2950f8d58a6dd1ccb430b00dd1b --- /dev/null +++ b/v2.1/portuguese/cond_prefill.mlpackage/Data/com.apple.CoreML/weights/weight.bin @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:0d40b2e26baeac0a03cb98c944644fd9b12a1e1ec307aa177d8f667398198c8d +size 133237248 diff --git a/v2.1/portuguese/cond_prefill.mlpackage/Manifest.json b/v2.1/portuguese/cond_prefill.mlpackage/Manifest.json new file mode 100644 index 0000000000000000000000000000000000000000..9fc211ccb7004e4409067918340adbc323f5861f --- /dev/null +++ b/v2.1/portuguese/cond_prefill.mlpackage/Manifest.json @@ -0,0 +1,18 @@ +{ + "fileFormatVersion": "1.0.0", + "itemInfoEntries": { + "5593C151-F0EE-44B3-BA54-448BE94B5E33": { + "author": "com.apple.CoreML", + "description": "CoreML Model Weights", + "name": "weights", + "path": "com.apple.CoreML/weights" + }, + "994F4BC2-CB1B-44F7-8FD1-673327A147BE": { + "author": "com.apple.CoreML", + "description": "CoreML Model Specification", + "name": "model.mlmodel", + "path": "com.apple.CoreML/model.mlmodel" + } + }, + "rootModelIdentifier": "994F4BC2-CB1B-44F7-8FD1-673327A147BE" +} diff --git a/v2.1/portuguese/constants/bos_emb.npy b/v2.1/portuguese/constants/bos_emb.npy new file mode 100644 index 0000000000000000000000000000000000000000..adbfbfd1ea34895513e45f20be17ddceb2595479 --- /dev/null +++ b/v2.1/portuguese/constants/bos_emb.npy @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:d539a9ffc7c6fb2b75a5bc05a2a2b6e1c9d6c2053e60c137ca1e40c86ec581fc +size 256 diff --git a/v2.1/portuguese/constants/emb_mean.npy b/v2.1/portuguese/constants/emb_mean.npy new file mode 100644 index 0000000000000000000000000000000000000000..f3e81b1fd3c115acf06fc2f8f16c98ab076786de --- /dev/null +++ b/v2.1/portuguese/constants/emb_mean.npy @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:6e8f8d7687fbfb92550df4980660eb445bff75ad6a72017524a15c532f38c0c7 +size 256 diff --git a/v2.1/portuguese/constants/emb_std.npy b/v2.1/portuguese/constants/emb_std.npy new file mode 100644 index 0000000000000000000000000000000000000000..73d8b42712f888ece58633a0648eb11b89f00518 --- /dev/null +++ b/v2.1/portuguese/constants/emb_std.npy @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:f15a323c90a6e6914ac259aa3e5fb32be19a6b40818293c84330aefd60d59bb8 +size 256 diff --git a/v2.1/portuguese/constants/mimi_init_state.npz b/v2.1/portuguese/constants/mimi_init_state.npz new file mode 100644 index 0000000000000000000000000000000000000000..3b52f383a86ef0439f9993506e4529c4c1234828 --- /dev/null +++ b/v2.1/portuguese/constants/mimi_init_state.npz @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:515588e10e9c14a6b009f1cfda9fcd592542b1640ad856d5fc69e50024702a39 +size 1083108 diff --git a/v2.1/portuguese/constants/quantizer_weight.npy b/v2.1/portuguese/constants/quantizer_weight.npy new file mode 100644 index 0000000000000000000000000000000000000000..a6d7797c101669ae052c8f4fa42bf751b61d3694 --- /dev/null +++ b/v2.1/portuguese/constants/quantizer_weight.npy @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:f44dcaa3aae563c566c19291ba101b751765d96e1ed4cb3ded8f8fc655d23cee +size 65664 diff --git a/v2.1/portuguese/constants/text_embed_table.npy b/v2.1/portuguese/constants/text_embed_table.npy new file mode 100644 index 0000000000000000000000000000000000000000..e722d10d764208b09489c126bf703568f237abcc --- /dev/null +++ b/v2.1/portuguese/constants/text_embed_table.npy @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:11dd8f55fbdaeef4fe0db590267f77f8daa5f2469a306d46afe8e5634ace5ca9 +size 16388224 diff --git a/v2.1/portuguese/constants_bin/alba.safetensors b/v2.1/portuguese/constants_bin/alba.safetensors new file mode 100644 index 0000000000000000000000000000000000000000..8f74b3cb7638b07a2117229dc87299b04b6f4b37 --- /dev/null +++ b/v2.1/portuguese/constants_bin/alba.safetensors @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:855b277ffa0e1f1600acacdb72f4e8609af7db7bbb5236dd904d1472c27541d2 +size 6194424 diff --git a/v2.1/portuguese/constants_bin/anna.safetensors b/v2.1/portuguese/constants_bin/anna.safetensors new file mode 100644 index 0000000000000000000000000000000000000000..eeb0964356ca08aa3d5c76157906443107f675b0 --- /dev/null +++ b/v2.1/portuguese/constants_bin/anna.safetensors @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:70ef78480c1d90f502826c10b0512effa34262453e22a76d5992d5fa1e8182bb +size 7816440 diff --git a/v2.1/portuguese/constants_bin/azelma.safetensors b/v2.1/portuguese/constants_bin/azelma.safetensors new file mode 100644 index 0000000000000000000000000000000000000000..84080a6321b138ba8233701a7b7544de3c975669 --- /dev/null +++ b/v2.1/portuguese/constants_bin/azelma.safetensors @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:9abb3ad69fbea053e44faa7c93d62d660b3552157ee02bcefc9157101ef269c2 +size 7963896 diff --git a/v2.1/portuguese/constants_bin/bill_boerst.safetensors b/v2.1/portuguese/constants_bin/bill_boerst.safetensors new file mode 100644 index 0000000000000000000000000000000000000000..7bea01e701554cdbca23a2fdab1c11f8f6e9ad8d --- /dev/null +++ b/v2.1/portuguese/constants_bin/bill_boerst.safetensors @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:e5ed6c383f0ae3c92d86202d5f20b2cdd424ad3221873f4322728bdf61890cc6 +size 6735096 diff --git a/v2.1/portuguese/constants_bin/bos_before_voice.bin b/v2.1/portuguese/constants_bin/bos_before_voice.bin new file mode 100644 index 0000000000000000000000000000000000000000..a1c078695688a88fbd1fe161567863b16911f0f2 --- /dev/null +++ b/v2.1/portuguese/constants_bin/bos_before_voice.bin @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:73964fb959c0b8f011b4a50a519f2c0dc42b16190f2d333502c6e5f160f1e3a5 +size 4096 diff --git a/v2.1/portuguese/constants_bin/bos_emb.bin b/v2.1/portuguese/constants_bin/bos_emb.bin new file mode 100644 index 0000000000000000000000000000000000000000..da9dc9a9d757419b9888c345042595f8bd6b6deb --- /dev/null +++ b/v2.1/portuguese/constants_bin/bos_emb.bin @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:15f7f301111cb594c8ba42516508aa73d3416b49494d5256539e503686ecae4d +size 128 diff --git a/v2.1/portuguese/constants_bin/caro_davy.safetensors b/v2.1/portuguese/constants_bin/caro_davy.safetensors new file mode 100644 index 0000000000000000000000000000000000000000..f699ef99f795bb876885551c7187bf1c946c6483 --- /dev/null +++ b/v2.1/portuguese/constants_bin/caro_davy.safetensors @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:85ff67e73e0f1c58e573eb10786da410e1fb79fdc71a10989e7bc11690cb8c04 +size 5260536 diff --git a/v2.1/portuguese/constants_bin/charles.safetensors b/v2.1/portuguese/constants_bin/charles.safetensors new file mode 100644 index 0000000000000000000000000000000000000000..3de84262b928b9fa6c8a5e53520cbbd29d0a2dac --- /dev/null +++ b/v2.1/portuguese/constants_bin/charles.safetensors @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:e643bddfba4c1c302f0589c09a18ac553af9fa2c7e82109e12a1a89225dd06bb +size 6194424 diff --git a/v2.1/portuguese/constants_bin/cosette.safetensors b/v2.1/portuguese/constants_bin/cosette.safetensors new file mode 100644 index 0000000000000000000000000000000000000000..5b0b463e64f53a3d8a65a0b353c1bf109b7d46eb --- /dev/null +++ b/v2.1/portuguese/constants_bin/cosette.safetensors @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:8aafd3e6b527dbb80123376b988ef2f21a8bee46b66e8b841ef529a298ba2acb +size 6194424 diff --git a/v2.1/portuguese/constants_bin/eponine.safetensors b/v2.1/portuguese/constants_bin/eponine.safetensors new file mode 100644 index 0000000000000000000000000000000000000000..7fee5d0b74a1079238911917a00a1f5428699b59 --- /dev/null +++ b/v2.1/portuguese/constants_bin/eponine.safetensors @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:268bf106463dc70867a965a724af55c7d25c926d5408a6d962cb601c7bdac957 +size 6931704 diff --git a/v2.1/portuguese/constants_bin/estelle.safetensors b/v2.1/portuguese/constants_bin/estelle.safetensors new file mode 100644 index 0000000000000000000000000000000000000000..826e63c0e3be9727cd5d540d21214f5521966792 --- /dev/null +++ b/v2.1/portuguese/constants_bin/estelle.safetensors @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:8d6489742a5d4026f23f115f3430cafc4fe931fe9faa2df69e6c3cca5ba8e3c4 +size 8258808 diff --git a/v2.1/portuguese/constants_bin/eve.safetensors b/v2.1/portuguese/constants_bin/eve.safetensors new file mode 100644 index 0000000000000000000000000000000000000000..da24ac74625b52847397ee8b92d1a38de6f322d8 --- /dev/null +++ b/v2.1/portuguese/constants_bin/eve.safetensors @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:c3d3526bdf0d4f82a3e6cc8bca62148a20c69bbe6596fb72920478688f378488 +size 6538488 diff --git a/v2.1/portuguese/constants_bin/fantine.safetensors b/v2.1/portuguese/constants_bin/fantine.safetensors new file mode 100644 index 0000000000000000000000000000000000000000..ee403d05a841e8370ac823a3cab40ba57e141573 --- /dev/null +++ b/v2.1/portuguese/constants_bin/fantine.safetensors @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:826b7131c94bb43ef9123780189e1f1ffac737c2bfec1c897e1244437b481d06 +size 6538488 diff --git a/v2.1/portuguese/constants_bin/george.safetensors b/v2.1/portuguese/constants_bin/george.safetensors new file mode 100644 index 0000000000000000000000000000000000000000..da4021681f8b0f40fb16b9639c24cb915738b9f5 --- /dev/null +++ b/v2.1/portuguese/constants_bin/george.safetensors @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:f727f2820a343dc8cb0f867478e8feae9ecc8da41d6a84e6e051f1c192d20bb4 +size 6243576 diff --git a/v2.1/portuguese/constants_bin/giovanni.safetensors b/v2.1/portuguese/constants_bin/giovanni.safetensors new file mode 100644 index 0000000000000000000000000000000000000000..62a67f473ffb166f59b4db71eb36d3f8ad72d6d1 --- /dev/null +++ b/v2.1/portuguese/constants_bin/giovanni.safetensors @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:099e94d27a8eb01df59e9284de8ca9fd5dc9c1cd48929b9818bed28f7958bb56 +size 4621552 diff --git a/v2.1/portuguese/constants_bin/jane.safetensors b/v2.1/portuguese/constants_bin/jane.safetensors new file mode 100644 index 0000000000000000000000000000000000000000..74d90e5adb6c138a06aee38260ec1306cfbfd330 --- /dev/null +++ b/v2.1/portuguese/constants_bin/jane.safetensors @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:de8f3516290a8f554785366cd2ce5dac1ad6abf694b6d37a41553b092eeb508e +size 7374072 diff --git a/v2.1/portuguese/constants_bin/javert.safetensors b/v2.1/portuguese/constants_bin/javert.safetensors new file mode 100644 index 0000000000000000000000000000000000000000..7dfddd51d583e9652175e152f4e2e2f3a5637261 --- /dev/null +++ b/v2.1/portuguese/constants_bin/javert.safetensors @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:d4f340f52817471a7460530eddb77e8a6b4c06fea5bf285fb80dae46f1e73254 +size 6194424 diff --git a/v2.1/portuguese/constants_bin/jean.safetensors b/v2.1/portuguese/constants_bin/jean.safetensors new file mode 100644 index 0000000000000000000000000000000000000000..74d804ce751525de9bfb8983630ead568cd97a1e --- /dev/null +++ b/v2.1/portuguese/constants_bin/jean.safetensors @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:e3a3e6d7e19db8dc5336e7dc9770d28e2112482c4a50f822940a19b0833f57e2 +size 6194424 diff --git a/v2.1/portuguese/constants_bin/juergen.safetensors b/v2.1/portuguese/constants_bin/juergen.safetensors new file mode 100644 index 0000000000000000000000000000000000000000..958a383fd6c2e86fd9e414c4eb9ff21bc1ace036 --- /dev/null +++ b/v2.1/portuguese/constants_bin/juergen.safetensors @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:fca811ed558b25da1f651da9b1eaab3e45b158258449060ce7694ccc8faaeca7 +size 6243576 diff --git a/v2.1/portuguese/constants_bin/lola.safetensors b/v2.1/portuguese/constants_bin/lola.safetensors new file mode 100644 index 0000000000000000000000000000000000000000..d64d2f25988d5aba85246b867ae29266960300a0 --- /dev/null +++ b/v2.1/portuguese/constants_bin/lola.safetensors @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:50a3a48ec0b67918e4775894a719ef9d4ce642bd355c275ad84decf14471fc55 +size 5948664 diff --git a/v2.1/portuguese/constants_bin/marius.safetensors b/v2.1/portuguese/constants_bin/marius.safetensors new file mode 100644 index 0000000000000000000000000000000000000000..aa4e4bb147f0a567ee8e3d9e163b0dc67b13836e --- /dev/null +++ b/v2.1/portuguese/constants_bin/marius.safetensors @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:dda7f1265e424907c0260b8b640b4aaae6a2e926e7fa40ceda3ac270f316308f +size 6194424 diff --git a/v2.1/portuguese/constants_bin/mary.safetensors b/v2.1/portuguese/constants_bin/mary.safetensors new file mode 100644 index 0000000000000000000000000000000000000000..590d91c5fb29ec64aeb02d9568ba278f7c594373 --- /dev/null +++ b/v2.1/portuguese/constants_bin/mary.safetensors @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:7c7a02fa20d4581e36b54fef748d189873e2ce398ccc57376bf76a8466a6d9cb +size 6194424 diff --git a/v2.1/portuguese/constants_bin/michael.safetensors b/v2.1/portuguese/constants_bin/michael.safetensors new file mode 100644 index 0000000000000000000000000000000000000000..4ce4c3751ebb57b926e60050621c116423b6e4b4 --- /dev/null +++ b/v2.1/portuguese/constants_bin/michael.safetensors @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:d9348313f6e9d7badd420cffac11a59dc5767eeeccc7a679f8cfaaffecce7f2c +size 7275768 diff --git a/v2.1/portuguese/constants_bin/paul.safetensors b/v2.1/portuguese/constants_bin/paul.safetensors new file mode 100644 index 0000000000000000000000000000000000000000..edf141e509b49f7cb7d031c69c509753f8a841d0 --- /dev/null +++ b/v2.1/portuguese/constants_bin/paul.safetensors @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:ce48b67d732704f38985b38f5d5dea54175fa5bca795379cad0cffd55e534668 +size 6980856 diff --git a/v2.1/portuguese/constants_bin/peter_yearsley.safetensors b/v2.1/portuguese/constants_bin/peter_yearsley.safetensors new file mode 100644 index 0000000000000000000000000000000000000000..881b8b97c3f3b9d215cd51ae15c280689fa3cc34 --- /dev/null +++ b/v2.1/portuguese/constants_bin/peter_yearsley.safetensors @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:fad9560971b95008f09a58b9faf689c6e35dc6e1594d8a049c9ce9d65392a6cf +size 3736816 diff --git a/v2.1/portuguese/constants_bin/rafael.safetensors b/v2.1/portuguese/constants_bin/rafael.safetensors new file mode 100644 index 0000000000000000000000000000000000000000..8ac09ca8c25085aababbcc0d73e781ce218f1ccb --- /dev/null +++ b/v2.1/portuguese/constants_bin/rafael.safetensors @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:7e7194d9ab6f1fb78db6f3a0574a605cd5ac666e6c78825f68028aa5e14592bd +size 6194424 diff --git a/v2.1/portuguese/constants_bin/stuart_bell.safetensors b/v2.1/portuguese/constants_bin/stuart_bell.safetensors new file mode 100644 index 0000000000000000000000000000000000000000..df01d9ada5268b4263ec377f00a4175e92e0b436 --- /dev/null +++ b/v2.1/portuguese/constants_bin/stuart_bell.safetensors @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:4e069c47b4a88d163847763efd654666de6c745f692d56d04a52a6b2e662bfa5 +size 5260536 diff --git a/v2.1/portuguese/constants_bin/text_embed_table.bin b/v2.1/portuguese/constants_bin/text_embed_table.bin new file mode 100644 index 0000000000000000000000000000000000000000..4c1f2cd5d1a1c4ba34f15365d81019b05d0152e9 --- /dev/null +++ b/v2.1/portuguese/constants_bin/text_embed_table.bin @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:55c0f677996139ce5004df6cf3815e3fe327eff79a890a8a1180f8d664abca0f +size 16388096 diff --git a/v2.1/portuguese/constants_bin/tokenizer.model b/v2.1/portuguese/constants_bin/tokenizer.model new file mode 100644 index 0000000000000000000000000000000000000000..1b7208270cda6aa8e2fa350d24d798eb6dee35ad --- /dev/null +++ b/v2.1/portuguese/constants_bin/tokenizer.model @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:3aa51309c55f114771c156aaeb86f6fc325991364aa3c38af74aecf1cbd0fade +size 60995 diff --git a/v2.1/portuguese/constants_bin/vera.safetensors b/v2.1/portuguese/constants_bin/vera.safetensors new file mode 100644 index 0000000000000000000000000000000000000000..a97abdae8c34422d6bdd479d777cddfe2c167a65 --- /dev/null +++ b/v2.1/portuguese/constants_bin/vera.safetensors @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:af141d1203871caef36d734d28b3aaee1acd2db00cf690fedcbc48d8268887de +size 6735096 diff --git a/v2.1/portuguese/flow_decoder_fused.mlmodelc/analytics/coremldata.bin b/v2.1/portuguese/flow_decoder_fused.mlmodelc/analytics/coremldata.bin new file mode 100644 index 0000000000000000000000000000000000000000..6e8c17c373ef36d99f9a47a3995bdcb56f8e9f70 --- /dev/null +++ b/v2.1/portuguese/flow_decoder_fused.mlmodelc/analytics/coremldata.bin @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:aca0117d1b1c046f988379f52856d831731f035df9b7e8531a00bad2855dbee9 +size 243 diff --git a/v2.1/portuguese/flow_decoder_fused.mlmodelc/coremldata.bin b/v2.1/portuguese/flow_decoder_fused.mlmodelc/coremldata.bin new file mode 100644 index 0000000000000000000000000000000000000000..908ae1b2172f997d862befdeb3a7cf778bb02e63 --- /dev/null +++ b/v2.1/portuguese/flow_decoder_fused.mlmodelc/coremldata.bin @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:6d03a9080bf362f31109412a19f6e87ff5d984683386ad85321ce664e75e4542 +size 413 diff --git a/v2.1/portuguese/flow_decoder_fused.mlmodelc/model.mil b/v2.1/portuguese/flow_decoder_fused.mlmodelc/model.mil new file mode 100644 index 0000000000000000000000000000000000000000..c64a1f34555e7bbf9c8d4502d29f9ef52dd6b0ec --- /dev/null +++ b/v2.1/portuguese/flow_decoder_fused.mlmodelc/model.mil @@ -0,0 +1,1915 @@ +program(1.0) +[buildInfo = dict, tensor>({{"coremlc-component-MIL", "3520.4.1"}, {"coremlc-version", "3520.5.1"}, {"coremltools-component-torch", "2.12.0"}, {"coremltools-source-dialect", "TorchScript"}, {"coremltools-version", "9.0"}})] +{ + func main(tensor latent_init, tensor transformer_out) { + tensor latent_init_to_fp16_dtype_0 = const()[name = tensor("latent_init_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor flow_net_input_proj_weight_to_fp16 = const()[name = tensor("flow_net_input_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(64)))]; + tensor flow_net_input_proj_bias_to_fp16 = const()[name = tensor("flow_net_input_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(32896)))]; + tensor latent_init_to_fp16 = cast(dtype = latent_init_to_fp16_dtype_0, x = latent_init)[name = tensor("cast_146")]; + tensor linear_0_cast_fp16 = linear(bias = flow_net_input_proj_bias_to_fp16, weight = flow_net_input_proj_weight_to_fp16, x = latent_init_to_fp16)[name = tensor("linear_0_cast_fp16")]; + tensor input_3_to_fp16 = const()[name = tensor("input_3_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(33984)))]; + tensor input_5_cast_fp16 = silu(x = input_3_to_fp16)[name = tensor("input_5_cast_fp16")]; + tensor flow_net_time_embed_0_mlp_2_weight_to_fp16 = const()[name = tensor("flow_net_time_embed_0_mlp_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(35072)))]; + tensor flow_net_time_embed_0_mlp_2_bias_to_fp16 = const()[name = tensor("flow_net_time_embed_0_mlp_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(559424)))]; + tensor linear_2_cast_fp16 = linear(bias = flow_net_time_embed_0_mlp_2_bias_to_fp16, weight = flow_net_time_embed_0_mlp_2_weight_to_fp16, x = input_5_cast_fp16)[name = tensor("linear_2_cast_fp16")]; + tensor reduce_mean_0_axes_0 = const()[name = tensor("reduce_mean_0_axes_0"), val = tensor([-1])]; + tensor reduce_mean_0_keep_dims_0 = const()[name = tensor("reduce_mean_0_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_0_cast_fp16 = reduce_mean(axes = reduce_mean_0_axes_0, keep_dims = reduce_mean_0_keep_dims_0, x = linear_2_cast_fp16)[name = tensor("reduce_mean_0_cast_fp16")]; + tensor sub_0_cast_fp16 = sub(x = linear_2_cast_fp16, y = reduce_mean_0_cast_fp16)[name = tensor("sub_0_cast_fp16")]; + tensor square_0_cast_fp16 = square(x = sub_0_cast_fp16)[name = tensor("square_0_cast_fp16")]; + tensor reduce_mean_1_axes_0 = const()[name = tensor("reduce_mean_1_axes_0"), val = tensor([-1])]; + tensor reduce_mean_1_keep_dims_0 = const()[name = tensor("reduce_mean_1_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_1_cast_fp16 = reduce_mean(axes = reduce_mean_1_axes_0, keep_dims = reduce_mean_1_keep_dims_0, x = square_0_cast_fp16)[name = tensor("reduce_mean_1_cast_fp16")]; + tensor real_div_0_to_fp16 = const()[name = tensor("real_div_0_to_fp16"), val = tensor(0x1.008p+0)]; + tensor mul_0_cast_fp16 = mul(x = reduce_mean_1_cast_fp16, y = real_div_0_to_fp16)[name = tensor("mul_0_cast_fp16")]; + tensor var_70_to_fp16 = const()[name = tensor("op_70_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_1_cast_fp16 = add(x = mul_0_cast_fp16, y = var_70_to_fp16)[name = tensor("var_1_cast_fp16")]; + tensor var_73_epsilon_0 = const()[name = tensor("op_73_epsilon_0"), val = tensor(0x1.197998p-40)]; + tensor var_73_cast_fp16 = rsqrt(epsilon = var_73_epsilon_0, x = var_1_cast_fp16)[name = tensor("op_73_cast_fp16")]; + tensor const_3_to_fp16 = const()[name = tensor("const_3_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(560512)))]; + tensor var_74_cast_fp16 = mul(x = const_3_to_fp16, y = var_73_cast_fp16)[name = tensor("op_74_cast_fp16")]; + tensor var_75_cast_fp16 = mul(x = linear_2_cast_fp16, y = var_74_cast_fp16)[name = tensor("op_75_cast_fp16")]; + tensor input_9_to_fp16 = const()[name = tensor("input_9_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(561600)))]; + tensor input_11_cast_fp16 = silu(x = input_9_to_fp16)[name = tensor("input_11_cast_fp16")]; + tensor flow_net_time_embed_1_mlp_2_weight_to_fp16 = const()[name = tensor("flow_net_time_embed_1_mlp_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(562688)))]; + tensor flow_net_time_embed_1_mlp_2_bias_to_fp16 = const()[name = tensor("flow_net_time_embed_1_mlp_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1087040)))]; + tensor linear_4_cast_fp16 = linear(bias = flow_net_time_embed_1_mlp_2_bias_to_fp16, weight = flow_net_time_embed_1_mlp_2_weight_to_fp16, x = input_11_cast_fp16)[name = tensor("linear_4_cast_fp16")]; + tensor reduce_mean_2_axes_0 = const()[name = tensor("reduce_mean_2_axes_0"), val = tensor([-1])]; + tensor reduce_mean_2_keep_dims_0 = const()[name = tensor("reduce_mean_2_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_2_cast_fp16 = reduce_mean(axes = reduce_mean_2_axes_0, keep_dims = reduce_mean_2_keep_dims_0, x = linear_4_cast_fp16)[name = tensor("reduce_mean_2_cast_fp16")]; + tensor sub_2_cast_fp16 = sub(x = linear_4_cast_fp16, y = reduce_mean_2_cast_fp16)[name = tensor("sub_2_cast_fp16")]; + tensor square_1_cast_fp16 = square(x = sub_2_cast_fp16)[name = tensor("square_1_cast_fp16")]; + tensor reduce_mean_3_axes_0 = const()[name = tensor("reduce_mean_3_axes_0"), val = tensor([-1])]; + tensor reduce_mean_3_keep_dims_0 = const()[name = tensor("reduce_mean_3_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_3_cast_fp16 = reduce_mean(axes = reduce_mean_3_axes_0, keep_dims = reduce_mean_3_keep_dims_0, x = square_1_cast_fp16)[name = tensor("reduce_mean_3_cast_fp16")]; + tensor real_div_1_to_fp16 = const()[name = tensor("real_div_1_to_fp16"), val = tensor(0x1.008p+0)]; + tensor mul_1_cast_fp16 = mul(x = reduce_mean_3_cast_fp16, y = real_div_1_to_fp16)[name = tensor("mul_1_cast_fp16")]; + tensor var_110_to_fp16 = const()[name = tensor("op_110_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_3_cast_fp16 = add(x = mul_1_cast_fp16, y = var_110_to_fp16)[name = tensor("var_3_cast_fp16")]; + tensor var_113_epsilon_0 = const()[name = tensor("op_113_epsilon_0"), val = tensor(0x1.197998p-40)]; + tensor var_113_cast_fp16 = rsqrt(epsilon = var_113_epsilon_0, x = var_3_cast_fp16)[name = tensor("op_113_cast_fp16")]; + tensor const_5_to_fp16 = const()[name = tensor("const_5_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1088128)))]; + tensor var_114_cast_fp16 = mul(x = const_5_to_fp16, y = var_113_cast_fp16)[name = tensor("op_114_cast_fp16")]; + tensor var_115_cast_fp16 = mul(x = linear_4_cast_fp16, y = var_114_cast_fp16)[name = tensor("op_115_cast_fp16")]; + tensor var_127_cast_fp16 = add(x = var_75_cast_fp16, y = var_115_cast_fp16)[name = tensor("op_127_cast_fp16")]; + tensor _inversed_t_combined_1_y_0_to_fp16 = const()[name = tensor("_inversed_t_combined_1_y_0_to_fp16"), val = tensor(0x1p-1)]; + tensor _inversed_t_combined_1_cast_fp16 = mul(x = var_127_cast_fp16, y = _inversed_t_combined_1_y_0_to_fp16)[name = tensor("_inversed_t_combined_1_cast_fp16")]; + tensor transformer_out_to_fp16_dtype_0 = const()[name = tensor("transformer_out_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor flow_net_cond_embed_weight_to_fp16 = const()[name = tensor("flow_net_cond_embed_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1089216)))]; + tensor flow_net_cond_embed_bias_to_fp16 = const()[name = tensor("flow_net_cond_embed_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(2137856)))]; + tensor transformer_out_to_fp16 = cast(dtype = transformer_out_to_fp16_dtype_0, x = transformer_out)[name = tensor("cast_145")]; + tensor linear_5_cast_fp16 = linear(bias = flow_net_cond_embed_bias_to_fp16, weight = flow_net_cond_embed_weight_to_fp16, x = transformer_out_to_fp16)[name = tensor("linear_5_cast_fp16")]; + tensor input_13_cast_fp16 = add(x = _inversed_t_combined_1_cast_fp16, y = linear_5_cast_fp16)[name = tensor("input_13_cast_fp16")]; + tensor input_15_cast_fp16 = silu(x = input_13_cast_fp16)[name = tensor("input_15_cast_fp16")]; + tensor flow_net_res_blocks_0_adaLN_modulation_1_weight_to_fp16 = const()[name = tensor("flow_net_res_blocks_0_adaLN_modulation_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(2138944)))]; + tensor flow_net_res_blocks_0_adaLN_modulation_1_bias_to_fp16 = const()[name = tensor("flow_net_res_blocks_0_adaLN_modulation_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(3711872)))]; + tensor linear_6_cast_fp16 = linear(bias = flow_net_res_blocks_0_adaLN_modulation_1_bias_to_fp16, weight = flow_net_res_blocks_0_adaLN_modulation_1_weight_to_fp16, x = input_15_cast_fp16)[name = tensor("linear_6_cast_fp16")]; + tensor var_142_split_sizes_0 = const()[name = tensor("op_142_split_sizes_0"), val = tensor([512, 512, 512])]; + tensor var_142_axis_0 = const()[name = tensor("op_142_axis_0"), val = tensor(-1)]; + tensor var_142_cast_fp16_0, tensor var_142_cast_fp16_1, tensor var_142_cast_fp16_2 = split(axis = var_142_axis_0, split_sizes = var_142_split_sizes_0, x = linear_6_cast_fp16)[name = tensor("op_142_cast_fp16")]; + tensor mean_1_axes_0 = const()[name = tensor("mean_1_axes_0"), val = tensor([-1])]; + tensor mean_1_keep_dims_0 = const()[name = tensor("mean_1_keep_dims_0"), val = tensor(true)]; + tensor mean_1_cast_fp16 = reduce_mean(axes = mean_1_axes_0, keep_dims = mean_1_keep_dims_0, x = linear_0_cast_fp16)[name = tensor("mean_1_cast_fp16")]; + tensor sub_4_cast_fp16 = sub(x = linear_0_cast_fp16, y = mean_1_cast_fp16)[name = tensor("sub_4_cast_fp16")]; + tensor square_2_cast_fp16 = square(x = sub_4_cast_fp16)[name = tensor("square_2_cast_fp16")]; + tensor reduce_mean_5_axes_0 = const()[name = tensor("reduce_mean_5_axes_0"), val = tensor([-1])]; + tensor reduce_mean_5_keep_dims_0 = const()[name = tensor("reduce_mean_5_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_5_cast_fp16 = reduce_mean(axes = reduce_mean_5_axes_0, keep_dims = reduce_mean_5_keep_dims_0, x = square_2_cast_fp16)[name = tensor("reduce_mean_5_cast_fp16")]; + tensor var_152_to_fp16 = const()[name = tensor("op_152_to_fp16"), val = tensor(0x1.1p-20)]; + tensor var_153_cast_fp16 = add(x = reduce_mean_5_cast_fp16, y = var_152_to_fp16)[name = tensor("op_153_cast_fp16")]; + tensor var_154_cast_fp16 = sqrt(x = var_153_cast_fp16)[name = tensor("op_154_cast_fp16")]; + tensor x_7_cast_fp16 = real_div(x = sub_4_cast_fp16, y = var_154_cast_fp16)[name = tensor("x_7_cast_fp16")]; + tensor flow_net_res_blocks_0_in_ln_weight_to_fp16 = const()[name = tensor("flow_net_res_blocks_0_in_ln_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(3715008)))]; + tensor var_156_cast_fp16 = mul(x = x_7_cast_fp16, y = flow_net_res_blocks_0_in_ln_weight_to_fp16)[name = tensor("op_156_cast_fp16")]; + tensor flow_net_res_blocks_0_in_ln_bias_to_fp16 = const()[name = tensor("flow_net_res_blocks_0_in_ln_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(3716096)))]; + tensor x_9_cast_fp16 = add(x = var_156_cast_fp16, y = flow_net_res_blocks_0_in_ln_bias_to_fp16)[name = tensor("x_9_cast_fp16")]; + tensor var_158_promoted_to_fp16 = const()[name = tensor("op_158_promoted_to_fp16"), val = tensor(0x1p+0)]; + tensor var_159_cast_fp16 = add(x = var_142_cast_fp16_1, y = var_158_promoted_to_fp16)[name = tensor("op_159_cast_fp16")]; + tensor var_160_cast_fp16 = mul(x = x_9_cast_fp16, y = var_159_cast_fp16)[name = tensor("op_160_cast_fp16")]; + tensor input_17_cast_fp16 = add(x = var_160_cast_fp16, y = var_142_cast_fp16_0)[name = tensor("input_17_cast_fp16")]; + tensor flow_net_res_blocks_0_mlp_0_weight_to_fp16 = const()[name = tensor("flow_net_res_blocks_0_mlp_0_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(3717184)))]; + tensor flow_net_res_blocks_0_mlp_0_bias_to_fp16 = const()[name = tensor("flow_net_res_blocks_0_mlp_0_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(4241536)))]; + tensor linear_7_cast_fp16 = linear(bias = flow_net_res_blocks_0_mlp_0_bias_to_fp16, weight = flow_net_res_blocks_0_mlp_0_weight_to_fp16, x = input_17_cast_fp16)[name = tensor("linear_7_cast_fp16")]; + tensor input_21_cast_fp16 = silu(x = linear_7_cast_fp16)[name = tensor("input_21_cast_fp16")]; + tensor flow_net_res_blocks_0_mlp_2_weight_to_fp16 = const()[name = tensor("flow_net_res_blocks_0_mlp_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(4242624)))]; + tensor flow_net_res_blocks_0_mlp_2_bias_to_fp16 = const()[name = tensor("flow_net_res_blocks_0_mlp_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(4766976)))]; + tensor linear_8_cast_fp16 = linear(bias = flow_net_res_blocks_0_mlp_2_bias_to_fp16, weight = flow_net_res_blocks_0_mlp_2_weight_to_fp16, x = input_21_cast_fp16)[name = tensor("linear_8_cast_fp16")]; + tensor var_171_cast_fp16 = mul(x = var_142_cast_fp16_2, y = linear_8_cast_fp16)[name = tensor("op_171_cast_fp16")]; + tensor x_11_cast_fp16 = add(x = linear_0_cast_fp16, y = var_171_cast_fp16)[name = tensor("x_11_cast_fp16")]; + tensor flow_net_res_blocks_1_adaLN_modulation_1_weight_to_fp16 = const()[name = tensor("flow_net_res_blocks_1_adaLN_modulation_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(4768064)))]; + tensor flow_net_res_blocks_1_adaLN_modulation_1_bias_to_fp16 = const()[name = tensor("flow_net_res_blocks_1_adaLN_modulation_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(6340992)))]; + tensor linear_9_cast_fp16 = linear(bias = flow_net_res_blocks_1_adaLN_modulation_1_bias_to_fp16, weight = flow_net_res_blocks_1_adaLN_modulation_1_weight_to_fp16, x = input_15_cast_fp16)[name = tensor("linear_9_cast_fp16")]; + tensor var_181_split_sizes_0 = const()[name = tensor("op_181_split_sizes_0"), val = tensor([512, 512, 512])]; + tensor var_181_axis_0 = const()[name = tensor("op_181_axis_0"), val = tensor(-1)]; + tensor var_181_cast_fp16_0, tensor var_181_cast_fp16_1, tensor var_181_cast_fp16_2 = split(axis = var_181_axis_0, split_sizes = var_181_split_sizes_0, x = linear_9_cast_fp16)[name = tensor("op_181_cast_fp16")]; + tensor mean_3_axes_0 = const()[name = tensor("mean_3_axes_0"), val = tensor([-1])]; + tensor mean_3_keep_dims_0 = const()[name = tensor("mean_3_keep_dims_0"), val = tensor(true)]; + tensor mean_3_cast_fp16 = reduce_mean(axes = mean_3_axes_0, keep_dims = mean_3_keep_dims_0, x = x_11_cast_fp16)[name = tensor("mean_3_cast_fp16")]; + tensor sub_5_cast_fp16 = sub(x = x_11_cast_fp16, y = mean_3_cast_fp16)[name = tensor("sub_5_cast_fp16")]; + tensor square_3_cast_fp16 = square(x = sub_5_cast_fp16)[name = tensor("square_3_cast_fp16")]; + tensor reduce_mean_7_axes_0 = const()[name = tensor("reduce_mean_7_axes_0"), val = tensor([-1])]; + tensor reduce_mean_7_keep_dims_0 = const()[name = tensor("reduce_mean_7_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_7_cast_fp16 = reduce_mean(axes = reduce_mean_7_axes_0, keep_dims = reduce_mean_7_keep_dims_0, x = square_3_cast_fp16)[name = tensor("reduce_mean_7_cast_fp16")]; + tensor var_191_to_fp16 = const()[name = tensor("op_191_to_fp16"), val = tensor(0x1.1p-20)]; + tensor var_192_cast_fp16 = add(x = reduce_mean_7_cast_fp16, y = var_191_to_fp16)[name = tensor("op_192_cast_fp16")]; + tensor var_193_cast_fp16 = sqrt(x = var_192_cast_fp16)[name = tensor("op_193_cast_fp16")]; + tensor x_13_cast_fp16 = real_div(x = sub_5_cast_fp16, y = var_193_cast_fp16)[name = tensor("x_13_cast_fp16")]; + tensor flow_net_res_blocks_1_in_ln_weight_to_fp16 = const()[name = tensor("flow_net_res_blocks_1_in_ln_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(6344128)))]; + tensor var_195_cast_fp16 = mul(x = x_13_cast_fp16, y = flow_net_res_blocks_1_in_ln_weight_to_fp16)[name = tensor("op_195_cast_fp16")]; + tensor flow_net_res_blocks_1_in_ln_bias_to_fp16 = const()[name = tensor("flow_net_res_blocks_1_in_ln_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(6345216)))]; + tensor x_15_cast_fp16 = add(x = var_195_cast_fp16, y = flow_net_res_blocks_1_in_ln_bias_to_fp16)[name = tensor("x_15_cast_fp16")]; + tensor var_197_promoted_to_fp16 = const()[name = tensor("op_197_promoted_to_fp16"), val = tensor(0x1p+0)]; + tensor var_198_cast_fp16 = add(x = var_181_cast_fp16_1, y = var_197_promoted_to_fp16)[name = tensor("op_198_cast_fp16")]; + tensor var_199_cast_fp16 = mul(x = x_15_cast_fp16, y = var_198_cast_fp16)[name = tensor("op_199_cast_fp16")]; + tensor input_25_cast_fp16 = add(x = var_199_cast_fp16, y = var_181_cast_fp16_0)[name = tensor("input_25_cast_fp16")]; + tensor flow_net_res_blocks_1_mlp_0_weight_to_fp16 = const()[name = tensor("flow_net_res_blocks_1_mlp_0_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(6346304)))]; + tensor flow_net_res_blocks_1_mlp_0_bias_to_fp16 = const()[name = tensor("flow_net_res_blocks_1_mlp_0_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(6870656)))]; + tensor linear_10_cast_fp16 = linear(bias = flow_net_res_blocks_1_mlp_0_bias_to_fp16, weight = flow_net_res_blocks_1_mlp_0_weight_to_fp16, x = input_25_cast_fp16)[name = tensor("linear_10_cast_fp16")]; + tensor input_29_cast_fp16 = silu(x = linear_10_cast_fp16)[name = tensor("input_29_cast_fp16")]; + tensor flow_net_res_blocks_1_mlp_2_weight_to_fp16 = const()[name = tensor("flow_net_res_blocks_1_mlp_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(6871744)))]; + tensor flow_net_res_blocks_1_mlp_2_bias_to_fp16 = const()[name = tensor("flow_net_res_blocks_1_mlp_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(7396096)))]; + tensor linear_11_cast_fp16 = linear(bias = flow_net_res_blocks_1_mlp_2_bias_to_fp16, weight = flow_net_res_blocks_1_mlp_2_weight_to_fp16, x = input_29_cast_fp16)[name = tensor("linear_11_cast_fp16")]; + tensor var_210_cast_fp16 = mul(x = var_181_cast_fp16_2, y = linear_11_cast_fp16)[name = tensor("op_210_cast_fp16")]; + tensor x_17_cast_fp16 = add(x = x_11_cast_fp16, y = var_210_cast_fp16)[name = tensor("x_17_cast_fp16")]; + tensor flow_net_res_blocks_2_adaLN_modulation_1_weight_to_fp16 = const()[name = tensor("flow_net_res_blocks_2_adaLN_modulation_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(7397184)))]; + tensor flow_net_res_blocks_2_adaLN_modulation_1_bias_to_fp16 = const()[name = tensor("flow_net_res_blocks_2_adaLN_modulation_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8970112)))]; + tensor linear_12_cast_fp16 = linear(bias = flow_net_res_blocks_2_adaLN_modulation_1_bias_to_fp16, weight = flow_net_res_blocks_2_adaLN_modulation_1_weight_to_fp16, x = input_15_cast_fp16)[name = tensor("linear_12_cast_fp16")]; + tensor var_220_split_sizes_0 = const()[name = tensor("op_220_split_sizes_0"), val = tensor([512, 512, 512])]; + tensor var_220_axis_0 = const()[name = tensor("op_220_axis_0"), val = tensor(-1)]; + tensor var_220_cast_fp16_0, tensor var_220_cast_fp16_1, tensor var_220_cast_fp16_2 = split(axis = var_220_axis_0, split_sizes = var_220_split_sizes_0, x = linear_12_cast_fp16)[name = tensor("op_220_cast_fp16")]; + tensor mean_5_axes_0 = const()[name = tensor("mean_5_axes_0"), val = tensor([-1])]; + tensor mean_5_keep_dims_0 = const()[name = tensor("mean_5_keep_dims_0"), val = tensor(true)]; + tensor mean_5_cast_fp16 = reduce_mean(axes = mean_5_axes_0, keep_dims = mean_5_keep_dims_0, x = x_17_cast_fp16)[name = tensor("mean_5_cast_fp16")]; + tensor sub_6_cast_fp16 = sub(x = x_17_cast_fp16, y = mean_5_cast_fp16)[name = tensor("sub_6_cast_fp16")]; + tensor square_4_cast_fp16 = square(x = sub_6_cast_fp16)[name = tensor("square_4_cast_fp16")]; + tensor reduce_mean_9_axes_0 = const()[name = tensor("reduce_mean_9_axes_0"), val = tensor([-1])]; + tensor reduce_mean_9_keep_dims_0 = const()[name = tensor("reduce_mean_9_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_9_cast_fp16 = reduce_mean(axes = reduce_mean_9_axes_0, keep_dims = reduce_mean_9_keep_dims_0, x = square_4_cast_fp16)[name = tensor("reduce_mean_9_cast_fp16")]; + tensor var_230_to_fp16 = const()[name = tensor("op_230_to_fp16"), val = tensor(0x1.1p-20)]; + tensor var_231_cast_fp16 = add(x = reduce_mean_9_cast_fp16, y = var_230_to_fp16)[name = tensor("op_231_cast_fp16")]; + tensor var_232_cast_fp16 = sqrt(x = var_231_cast_fp16)[name = tensor("op_232_cast_fp16")]; + tensor x_19_cast_fp16 = real_div(x = sub_6_cast_fp16, y = var_232_cast_fp16)[name = tensor("x_19_cast_fp16")]; + tensor flow_net_res_blocks_2_in_ln_weight_to_fp16 = const()[name = tensor("flow_net_res_blocks_2_in_ln_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8973248)))]; + tensor var_234_cast_fp16 = mul(x = x_19_cast_fp16, y = flow_net_res_blocks_2_in_ln_weight_to_fp16)[name = tensor("op_234_cast_fp16")]; + tensor flow_net_res_blocks_2_in_ln_bias_to_fp16 = const()[name = tensor("flow_net_res_blocks_2_in_ln_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8974336)))]; + tensor x_21_cast_fp16 = add(x = var_234_cast_fp16, y = flow_net_res_blocks_2_in_ln_bias_to_fp16)[name = tensor("x_21_cast_fp16")]; + tensor var_236_promoted_to_fp16 = const()[name = tensor("op_236_promoted_to_fp16"), val = tensor(0x1p+0)]; + tensor var_237_cast_fp16 = add(x = var_220_cast_fp16_1, y = var_236_promoted_to_fp16)[name = tensor("op_237_cast_fp16")]; + tensor var_238_cast_fp16 = mul(x = x_21_cast_fp16, y = var_237_cast_fp16)[name = tensor("op_238_cast_fp16")]; + tensor input_33_cast_fp16 = add(x = var_238_cast_fp16, y = var_220_cast_fp16_0)[name = tensor("input_33_cast_fp16")]; + tensor flow_net_res_blocks_2_mlp_0_weight_to_fp16 = const()[name = tensor("flow_net_res_blocks_2_mlp_0_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8975424)))]; + tensor flow_net_res_blocks_2_mlp_0_bias_to_fp16 = const()[name = tensor("flow_net_res_blocks_2_mlp_0_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9499776)))]; + tensor linear_13_cast_fp16 = linear(bias = flow_net_res_blocks_2_mlp_0_bias_to_fp16, weight = flow_net_res_blocks_2_mlp_0_weight_to_fp16, x = input_33_cast_fp16)[name = tensor("linear_13_cast_fp16")]; + tensor input_37_cast_fp16 = silu(x = linear_13_cast_fp16)[name = tensor("input_37_cast_fp16")]; + tensor flow_net_res_blocks_2_mlp_2_weight_to_fp16 = const()[name = tensor("flow_net_res_blocks_2_mlp_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9500864)))]; + tensor flow_net_res_blocks_2_mlp_2_bias_to_fp16 = const()[name = tensor("flow_net_res_blocks_2_mlp_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(10025216)))]; + tensor linear_14_cast_fp16 = linear(bias = flow_net_res_blocks_2_mlp_2_bias_to_fp16, weight = flow_net_res_blocks_2_mlp_2_weight_to_fp16, x = input_37_cast_fp16)[name = tensor("linear_14_cast_fp16")]; + tensor var_249_cast_fp16 = mul(x = var_220_cast_fp16_2, y = linear_14_cast_fp16)[name = tensor("op_249_cast_fp16")]; + tensor x_23_cast_fp16 = add(x = x_17_cast_fp16, y = var_249_cast_fp16)[name = tensor("x_23_cast_fp16")]; + tensor flow_net_res_blocks_3_adaLN_modulation_1_weight_to_fp16 = const()[name = tensor("flow_net_res_blocks_3_adaLN_modulation_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(10026304)))]; + tensor flow_net_res_blocks_3_adaLN_modulation_1_bias_to_fp16 = const()[name = tensor("flow_net_res_blocks_3_adaLN_modulation_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(11599232)))]; + tensor linear_15_cast_fp16 = linear(bias = flow_net_res_blocks_3_adaLN_modulation_1_bias_to_fp16, weight = flow_net_res_blocks_3_adaLN_modulation_1_weight_to_fp16, x = input_15_cast_fp16)[name = tensor("linear_15_cast_fp16")]; + tensor var_259_split_sizes_0 = const()[name = tensor("op_259_split_sizes_0"), val = tensor([512, 512, 512])]; + tensor var_259_axis_0 = const()[name = tensor("op_259_axis_0"), val = tensor(-1)]; + tensor var_259_cast_fp16_0, tensor var_259_cast_fp16_1, tensor var_259_cast_fp16_2 = split(axis = var_259_axis_0, split_sizes = var_259_split_sizes_0, x = linear_15_cast_fp16)[name = tensor("op_259_cast_fp16")]; + tensor mean_7_axes_0 = const()[name = tensor("mean_7_axes_0"), val = tensor([-1])]; + tensor mean_7_keep_dims_0 = const()[name = tensor("mean_7_keep_dims_0"), val = tensor(true)]; + tensor mean_7_cast_fp16 = reduce_mean(axes = mean_7_axes_0, keep_dims = mean_7_keep_dims_0, x = x_23_cast_fp16)[name = tensor("mean_7_cast_fp16")]; + tensor sub_7_cast_fp16 = sub(x = x_23_cast_fp16, y = mean_7_cast_fp16)[name = tensor("sub_7_cast_fp16")]; + tensor square_5_cast_fp16 = square(x = sub_7_cast_fp16)[name = tensor("square_5_cast_fp16")]; + tensor reduce_mean_11_axes_0 = const()[name = tensor("reduce_mean_11_axes_0"), val = tensor([-1])]; + tensor reduce_mean_11_keep_dims_0 = const()[name = tensor("reduce_mean_11_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_11_cast_fp16 = reduce_mean(axes = reduce_mean_11_axes_0, keep_dims = reduce_mean_11_keep_dims_0, x = square_5_cast_fp16)[name = tensor("reduce_mean_11_cast_fp16")]; + tensor var_269_to_fp16 = const()[name = tensor("op_269_to_fp16"), val = tensor(0x1.1p-20)]; + tensor var_270_cast_fp16 = add(x = reduce_mean_11_cast_fp16, y = var_269_to_fp16)[name = tensor("op_270_cast_fp16")]; + tensor var_271_cast_fp16 = sqrt(x = var_270_cast_fp16)[name = tensor("op_271_cast_fp16")]; + tensor x_25_cast_fp16 = real_div(x = sub_7_cast_fp16, y = var_271_cast_fp16)[name = tensor("x_25_cast_fp16")]; + tensor flow_net_res_blocks_3_in_ln_weight_to_fp16 = const()[name = tensor("flow_net_res_blocks_3_in_ln_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(11602368)))]; + tensor var_273_cast_fp16 = mul(x = x_25_cast_fp16, y = flow_net_res_blocks_3_in_ln_weight_to_fp16)[name = tensor("op_273_cast_fp16")]; + tensor flow_net_res_blocks_3_in_ln_bias_to_fp16 = const()[name = tensor("flow_net_res_blocks_3_in_ln_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(11603456)))]; + tensor x_27_cast_fp16 = add(x = var_273_cast_fp16, y = flow_net_res_blocks_3_in_ln_bias_to_fp16)[name = tensor("x_27_cast_fp16")]; + tensor var_275_promoted_to_fp16 = const()[name = tensor("op_275_promoted_to_fp16"), val = tensor(0x1p+0)]; + tensor var_276_cast_fp16 = add(x = var_259_cast_fp16_1, y = var_275_promoted_to_fp16)[name = tensor("op_276_cast_fp16")]; + tensor var_277_cast_fp16 = mul(x = x_27_cast_fp16, y = var_276_cast_fp16)[name = tensor("op_277_cast_fp16")]; + tensor input_41_cast_fp16 = add(x = var_277_cast_fp16, y = var_259_cast_fp16_0)[name = tensor("input_41_cast_fp16")]; + tensor flow_net_res_blocks_3_mlp_0_weight_to_fp16 = const()[name = tensor("flow_net_res_blocks_3_mlp_0_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(11604544)))]; + tensor flow_net_res_blocks_3_mlp_0_bias_to_fp16 = const()[name = tensor("flow_net_res_blocks_3_mlp_0_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(12128896)))]; + tensor linear_16_cast_fp16 = linear(bias = flow_net_res_blocks_3_mlp_0_bias_to_fp16, weight = flow_net_res_blocks_3_mlp_0_weight_to_fp16, x = input_41_cast_fp16)[name = tensor("linear_16_cast_fp16")]; + tensor input_45_cast_fp16 = silu(x = linear_16_cast_fp16)[name = tensor("input_45_cast_fp16")]; + tensor flow_net_res_blocks_3_mlp_2_weight_to_fp16 = const()[name = tensor("flow_net_res_blocks_3_mlp_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(12129984)))]; + tensor flow_net_res_blocks_3_mlp_2_bias_to_fp16 = const()[name = tensor("flow_net_res_blocks_3_mlp_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(12654336)))]; + tensor linear_17_cast_fp16 = linear(bias = flow_net_res_blocks_3_mlp_2_bias_to_fp16, weight = flow_net_res_blocks_3_mlp_2_weight_to_fp16, x = input_45_cast_fp16)[name = tensor("linear_17_cast_fp16")]; + tensor var_288_cast_fp16 = mul(x = var_259_cast_fp16_2, y = linear_17_cast_fp16)[name = tensor("op_288_cast_fp16")]; + tensor x_29_cast_fp16 = add(x = x_23_cast_fp16, y = var_288_cast_fp16)[name = tensor("x_29_cast_fp16")]; + tensor flow_net_res_blocks_4_adaLN_modulation_1_weight_to_fp16 = const()[name = tensor("flow_net_res_blocks_4_adaLN_modulation_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(12655424)))]; + tensor flow_net_res_blocks_4_adaLN_modulation_1_bias_to_fp16 = const()[name = tensor("flow_net_res_blocks_4_adaLN_modulation_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(14228352)))]; + tensor linear_18_cast_fp16 = linear(bias = flow_net_res_blocks_4_adaLN_modulation_1_bias_to_fp16, weight = flow_net_res_blocks_4_adaLN_modulation_1_weight_to_fp16, x = input_15_cast_fp16)[name = tensor("linear_18_cast_fp16")]; + tensor var_298_split_sizes_0 = const()[name = tensor("op_298_split_sizes_0"), val = tensor([512, 512, 512])]; + tensor var_298_axis_0 = const()[name = tensor("op_298_axis_0"), val = tensor(-1)]; + tensor var_298_cast_fp16_0, tensor var_298_cast_fp16_1, tensor var_298_cast_fp16_2 = split(axis = var_298_axis_0, split_sizes = var_298_split_sizes_0, x = linear_18_cast_fp16)[name = tensor("op_298_cast_fp16")]; + tensor mean_9_axes_0 = const()[name = tensor("mean_9_axes_0"), val = tensor([-1])]; + tensor mean_9_keep_dims_0 = const()[name = tensor("mean_9_keep_dims_0"), val = tensor(true)]; + tensor mean_9_cast_fp16 = reduce_mean(axes = mean_9_axes_0, keep_dims = mean_9_keep_dims_0, x = x_29_cast_fp16)[name = tensor("mean_9_cast_fp16")]; + tensor sub_8_cast_fp16 = sub(x = x_29_cast_fp16, y = mean_9_cast_fp16)[name = tensor("sub_8_cast_fp16")]; + tensor square_6_cast_fp16 = square(x = sub_8_cast_fp16)[name = tensor("square_6_cast_fp16")]; + tensor reduce_mean_13_axes_0 = const()[name = tensor("reduce_mean_13_axes_0"), val = tensor([-1])]; + tensor reduce_mean_13_keep_dims_0 = const()[name = tensor("reduce_mean_13_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_13_cast_fp16 = reduce_mean(axes = reduce_mean_13_axes_0, keep_dims = reduce_mean_13_keep_dims_0, x = square_6_cast_fp16)[name = tensor("reduce_mean_13_cast_fp16")]; + tensor var_308_to_fp16 = const()[name = tensor("op_308_to_fp16"), val = tensor(0x1.1p-20)]; + tensor var_309_cast_fp16 = add(x = reduce_mean_13_cast_fp16, y = var_308_to_fp16)[name = tensor("op_309_cast_fp16")]; + tensor var_310_cast_fp16 = sqrt(x = var_309_cast_fp16)[name = tensor("op_310_cast_fp16")]; + tensor x_31_cast_fp16 = real_div(x = sub_8_cast_fp16, y = var_310_cast_fp16)[name = tensor("x_31_cast_fp16")]; + tensor flow_net_res_blocks_4_in_ln_weight_to_fp16 = const()[name = tensor("flow_net_res_blocks_4_in_ln_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(14231488)))]; + tensor var_312_cast_fp16 = mul(x = x_31_cast_fp16, y = flow_net_res_blocks_4_in_ln_weight_to_fp16)[name = tensor("op_312_cast_fp16")]; + tensor flow_net_res_blocks_4_in_ln_bias_to_fp16 = const()[name = tensor("flow_net_res_blocks_4_in_ln_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(14232576)))]; + tensor x_33_cast_fp16 = add(x = var_312_cast_fp16, y = flow_net_res_blocks_4_in_ln_bias_to_fp16)[name = tensor("x_33_cast_fp16")]; + tensor var_314_promoted_to_fp16 = const()[name = tensor("op_314_promoted_to_fp16"), val = tensor(0x1p+0)]; + tensor var_315_cast_fp16 = add(x = var_298_cast_fp16_1, y = var_314_promoted_to_fp16)[name = tensor("op_315_cast_fp16")]; + tensor var_316_cast_fp16 = mul(x = x_33_cast_fp16, y = var_315_cast_fp16)[name = tensor("op_316_cast_fp16")]; + tensor input_49_cast_fp16 = add(x = var_316_cast_fp16, y = var_298_cast_fp16_0)[name = tensor("input_49_cast_fp16")]; + tensor flow_net_res_blocks_4_mlp_0_weight_to_fp16 = const()[name = tensor("flow_net_res_blocks_4_mlp_0_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(14233664)))]; + tensor flow_net_res_blocks_4_mlp_0_bias_to_fp16 = const()[name = tensor("flow_net_res_blocks_4_mlp_0_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(14758016)))]; + tensor linear_19_cast_fp16 = linear(bias = flow_net_res_blocks_4_mlp_0_bias_to_fp16, weight = flow_net_res_blocks_4_mlp_0_weight_to_fp16, x = input_49_cast_fp16)[name = tensor("linear_19_cast_fp16")]; + tensor input_53_cast_fp16 = silu(x = linear_19_cast_fp16)[name = tensor("input_53_cast_fp16")]; + tensor flow_net_res_blocks_4_mlp_2_weight_to_fp16 = const()[name = tensor("flow_net_res_blocks_4_mlp_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(14759104)))]; + tensor flow_net_res_blocks_4_mlp_2_bias_to_fp16 = const()[name = tensor("flow_net_res_blocks_4_mlp_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15283456)))]; + tensor linear_20_cast_fp16 = linear(bias = flow_net_res_blocks_4_mlp_2_bias_to_fp16, weight = flow_net_res_blocks_4_mlp_2_weight_to_fp16, x = input_53_cast_fp16)[name = tensor("linear_20_cast_fp16")]; + tensor var_327_cast_fp16 = mul(x = var_298_cast_fp16_2, y = linear_20_cast_fp16)[name = tensor("op_327_cast_fp16")]; + tensor x_35_cast_fp16 = add(x = x_29_cast_fp16, y = var_327_cast_fp16)[name = tensor("x_35_cast_fp16")]; + tensor flow_net_res_blocks_5_adaLN_modulation_1_weight_to_fp16 = const()[name = tensor("flow_net_res_blocks_5_adaLN_modulation_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15284544)))]; + tensor flow_net_res_blocks_5_adaLN_modulation_1_bias_to_fp16 = const()[name = tensor("flow_net_res_blocks_5_adaLN_modulation_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(16857472)))]; + tensor linear_21_cast_fp16 = linear(bias = flow_net_res_blocks_5_adaLN_modulation_1_bias_to_fp16, weight = flow_net_res_blocks_5_adaLN_modulation_1_weight_to_fp16, x = input_15_cast_fp16)[name = tensor("linear_21_cast_fp16")]; + tensor var_337_split_sizes_0 = const()[name = tensor("op_337_split_sizes_0"), val = tensor([512, 512, 512])]; + tensor var_337_axis_0 = const()[name = tensor("op_337_axis_0"), val = tensor(-1)]; + tensor var_337_cast_fp16_0, tensor var_337_cast_fp16_1, tensor var_337_cast_fp16_2 = split(axis = var_337_axis_0, split_sizes = var_337_split_sizes_0, x = linear_21_cast_fp16)[name = tensor("op_337_cast_fp16")]; + tensor mean_11_axes_0 = const()[name = tensor("mean_11_axes_0"), val = tensor([-1])]; + tensor mean_11_keep_dims_0 = const()[name = tensor("mean_11_keep_dims_0"), val = tensor(true)]; + tensor mean_11_cast_fp16 = reduce_mean(axes = mean_11_axes_0, keep_dims = mean_11_keep_dims_0, x = x_35_cast_fp16)[name = tensor("mean_11_cast_fp16")]; + tensor sub_9_cast_fp16 = sub(x = x_35_cast_fp16, y = mean_11_cast_fp16)[name = tensor("sub_9_cast_fp16")]; + tensor square_7_cast_fp16 = square(x = sub_9_cast_fp16)[name = tensor("square_7_cast_fp16")]; + tensor reduce_mean_15_axes_0 = const()[name = tensor("reduce_mean_15_axes_0"), val = tensor([-1])]; + tensor reduce_mean_15_keep_dims_0 = const()[name = tensor("reduce_mean_15_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_15_cast_fp16 = reduce_mean(axes = reduce_mean_15_axes_0, keep_dims = reduce_mean_15_keep_dims_0, x = square_7_cast_fp16)[name = tensor("reduce_mean_15_cast_fp16")]; + tensor var_347_to_fp16 = const()[name = tensor("op_347_to_fp16"), val = tensor(0x1.1p-20)]; + tensor var_348_cast_fp16 = add(x = reduce_mean_15_cast_fp16, y = var_347_to_fp16)[name = tensor("op_348_cast_fp16")]; + tensor var_349_cast_fp16 = sqrt(x = var_348_cast_fp16)[name = tensor("op_349_cast_fp16")]; + tensor x_37_cast_fp16 = real_div(x = sub_9_cast_fp16, y = var_349_cast_fp16)[name = tensor("x_37_cast_fp16")]; + tensor flow_net_res_blocks_5_in_ln_weight_to_fp16 = const()[name = tensor("flow_net_res_blocks_5_in_ln_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(16860608)))]; + tensor var_351_cast_fp16 = mul(x = x_37_cast_fp16, y = flow_net_res_blocks_5_in_ln_weight_to_fp16)[name = tensor("op_351_cast_fp16")]; + tensor flow_net_res_blocks_5_in_ln_bias_to_fp16 = const()[name = tensor("flow_net_res_blocks_5_in_ln_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(16861696)))]; + tensor x_39_cast_fp16 = add(x = var_351_cast_fp16, y = flow_net_res_blocks_5_in_ln_bias_to_fp16)[name = tensor("x_39_cast_fp16")]; + tensor var_353_promoted_to_fp16 = const()[name = tensor("op_353_promoted_to_fp16"), val = tensor(0x1p+0)]; + tensor var_354_cast_fp16 = add(x = var_337_cast_fp16_1, y = var_353_promoted_to_fp16)[name = tensor("op_354_cast_fp16")]; + tensor var_355_cast_fp16 = mul(x = x_39_cast_fp16, y = var_354_cast_fp16)[name = tensor("op_355_cast_fp16")]; + tensor input_57_cast_fp16 = add(x = var_355_cast_fp16, y = var_337_cast_fp16_0)[name = tensor("input_57_cast_fp16")]; + tensor flow_net_res_blocks_5_mlp_0_weight_to_fp16 = const()[name = tensor("flow_net_res_blocks_5_mlp_0_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(16862784)))]; + tensor flow_net_res_blocks_5_mlp_0_bias_to_fp16 = const()[name = tensor("flow_net_res_blocks_5_mlp_0_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(17387136)))]; + tensor linear_22_cast_fp16 = linear(bias = flow_net_res_blocks_5_mlp_0_bias_to_fp16, weight = flow_net_res_blocks_5_mlp_0_weight_to_fp16, x = input_57_cast_fp16)[name = tensor("linear_22_cast_fp16")]; + tensor input_61_cast_fp16 = silu(x = linear_22_cast_fp16)[name = tensor("input_61_cast_fp16")]; + tensor flow_net_res_blocks_5_mlp_2_weight_to_fp16 = const()[name = tensor("flow_net_res_blocks_5_mlp_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(17388224)))]; + tensor flow_net_res_blocks_5_mlp_2_bias_to_fp16 = const()[name = tensor("flow_net_res_blocks_5_mlp_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(17912576)))]; + tensor linear_23_cast_fp16 = linear(bias = flow_net_res_blocks_5_mlp_2_bias_to_fp16, weight = flow_net_res_blocks_5_mlp_2_weight_to_fp16, x = input_61_cast_fp16)[name = tensor("linear_23_cast_fp16")]; + tensor var_366_cast_fp16 = mul(x = var_337_cast_fp16_2, y = linear_23_cast_fp16)[name = tensor("op_366_cast_fp16")]; + tensor x_41_cast_fp16 = add(x = x_35_cast_fp16, y = var_366_cast_fp16)[name = tensor("x_41_cast_fp16")]; + tensor flow_net_final_layer_adaLN_modulation_1_weight_to_fp16 = const()[name = tensor("flow_net_final_layer_adaLN_modulation_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(17913664)))]; + tensor flow_net_final_layer_adaLN_modulation_1_bias_to_fp16 = const()[name = tensor("flow_net_final_layer_adaLN_modulation_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(18962304)))]; + tensor linear_24_cast_fp16 = linear(bias = flow_net_final_layer_adaLN_modulation_1_bias_to_fp16, weight = flow_net_final_layer_adaLN_modulation_1_weight_to_fp16, x = input_15_cast_fp16)[name = tensor("linear_24_cast_fp16")]; + tensor var_375_split_sizes_0 = const()[name = tensor("op_375_split_sizes_0"), val = tensor([512, 512])]; + tensor var_375_axis_0 = const()[name = tensor("op_375_axis_0"), val = tensor(-1)]; + tensor var_375_cast_fp16_0, tensor var_375_cast_fp16_1 = split(axis = var_375_axis_0, split_sizes = var_375_split_sizes_0, x = linear_24_cast_fp16)[name = tensor("op_375_cast_fp16")]; + tensor mean_13_axes_0 = const()[name = tensor("mean_13_axes_0"), val = tensor([-1])]; + tensor mean_13_keep_dims_0 = const()[name = tensor("mean_13_keep_dims_0"), val = tensor(true)]; + tensor mean_13_cast_fp16 = reduce_mean(axes = mean_13_axes_0, keep_dims = mean_13_keep_dims_0, x = x_41_cast_fp16)[name = tensor("mean_13_cast_fp16")]; + tensor sub_10_cast_fp16 = sub(x = x_41_cast_fp16, y = mean_13_cast_fp16)[name = tensor("sub_10_cast_fp16")]; + tensor square_8_cast_fp16 = square(x = sub_10_cast_fp16)[name = tensor("square_8_cast_fp16")]; + tensor reduce_mean_17_axes_0 = const()[name = tensor("reduce_mean_17_axes_0"), val = tensor([-1])]; + tensor reduce_mean_17_keep_dims_0 = const()[name = tensor("reduce_mean_17_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_17_cast_fp16 = reduce_mean(axes = reduce_mean_17_axes_0, keep_dims = reduce_mean_17_keep_dims_0, x = square_8_cast_fp16)[name = tensor("reduce_mean_17_cast_fp16")]; + tensor var_382_to_fp16 = const()[name = tensor("op_382_to_fp16"), val = tensor(0x1.1p-20)]; + tensor var_383_cast_fp16 = add(x = reduce_mean_17_cast_fp16, y = var_382_to_fp16)[name = tensor("op_383_cast_fp16")]; + tensor var_384_cast_fp16 = sqrt(x = var_383_cast_fp16)[name = tensor("op_384_cast_fp16")]; + tensor x_43_cast_fp16 = real_div(x = sub_10_cast_fp16, y = var_384_cast_fp16)[name = tensor("x_43_cast_fp16")]; + tensor var_386_promoted_to_fp16 = const()[name = tensor("op_386_promoted_to_fp16"), val = tensor(0x1p+0)]; + tensor var_387_cast_fp16 = add(x = var_375_cast_fp16_1, y = var_386_promoted_to_fp16)[name = tensor("op_387_cast_fp16")]; + tensor var_388_cast_fp16 = mul(x = x_43_cast_fp16, y = var_387_cast_fp16)[name = tensor("op_388_cast_fp16")]; + tensor input_65_cast_fp16 = add(x = var_388_cast_fp16, y = var_375_cast_fp16_0)[name = tensor("input_65_cast_fp16")]; + tensor flow_net_final_layer_linear_weight_to_fp16 = const()[name = tensor("flow_net_final_layer_linear_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(18964416)))]; + tensor flow_net_final_layer_linear_bias_to_fp16 = const()[name = tensor("flow_net_final_layer_linear_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(18997248)))]; + tensor linear_25_cast_fp16 = linear(bias = flow_net_final_layer_linear_bias_to_fp16, weight = flow_net_final_layer_linear_weight_to_fp16, x = input_65_cast_fp16)[name = tensor("linear_25_cast_fp16")]; + tensor var_399_to_fp16 = const()[name = tensor("op_399_to_fp16"), val = tensor(0x1p-3)]; + tensor var_400_cast_fp16 = mul(x = linear_25_cast_fp16, y = var_399_to_fp16)[name = tensor("op_400_cast_fp16")]; + tensor input_67_cast_fp16 = add(x = latent_init_to_fp16, y = var_400_cast_fp16)[name = tensor("input_67_cast_fp16")]; + tensor linear_26_cast_fp16 = linear(bias = flow_net_input_proj_bias_to_fp16, weight = flow_net_input_proj_weight_to_fp16, x = input_67_cast_fp16)[name = tensor("linear_26_cast_fp16")]; + tensor input_71_to_fp16 = const()[name = tensor("input_71_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(18997376)))]; + tensor input_73_cast_fp16 = silu(x = input_71_to_fp16)[name = tensor("input_73_cast_fp16")]; + tensor linear_28_cast_fp16 = linear(bias = flow_net_time_embed_0_mlp_2_bias_to_fp16, weight = flow_net_time_embed_0_mlp_2_weight_to_fp16, x = input_73_cast_fp16)[name = tensor("linear_28_cast_fp16")]; + tensor reduce_mean_18_axes_0 = const()[name = tensor("reduce_mean_18_axes_0"), val = tensor([-1])]; + tensor reduce_mean_18_keep_dims_0 = const()[name = tensor("reduce_mean_18_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_18_cast_fp16 = reduce_mean(axes = reduce_mean_18_axes_0, keep_dims = reduce_mean_18_keep_dims_0, x = linear_28_cast_fp16)[name = tensor("reduce_mean_18_cast_fp16")]; + tensor sub_11_cast_fp16 = sub(x = linear_28_cast_fp16, y = reduce_mean_18_cast_fp16)[name = tensor("sub_11_cast_fp16")]; + tensor square_9_cast_fp16 = square(x = sub_11_cast_fp16)[name = tensor("square_9_cast_fp16")]; + tensor reduce_mean_19_axes_0 = const()[name = tensor("reduce_mean_19_axes_0"), val = tensor([-1])]; + tensor reduce_mean_19_keep_dims_0 = const()[name = tensor("reduce_mean_19_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_19_cast_fp16 = reduce_mean(axes = reduce_mean_19_axes_0, keep_dims = reduce_mean_19_keep_dims_0, x = square_9_cast_fp16)[name = tensor("reduce_mean_19_cast_fp16")]; + tensor real_div_2_to_fp16 = const()[name = tensor("real_div_2_to_fp16"), val = tensor(0x1.008p+0)]; + tensor mul_2_cast_fp16 = mul(x = reduce_mean_19_cast_fp16, y = real_div_2_to_fp16)[name = tensor("mul_2_cast_fp16")]; + tensor var_466_to_fp16 = const()[name = tensor("op_466_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_19_cast_fp16 = add(x = mul_2_cast_fp16, y = var_466_to_fp16)[name = tensor("var_19_cast_fp16")]; + tensor var_469_epsilon_0 = const()[name = tensor("op_469_epsilon_0"), val = tensor(0x1.197998p-40)]; + tensor var_469_cast_fp16 = rsqrt(epsilon = var_469_epsilon_0, x = var_19_cast_fp16)[name = tensor("op_469_cast_fp16")]; + tensor var_470_cast_fp16 = mul(x = const_3_to_fp16, y = var_469_cast_fp16)[name = tensor("op_470_cast_fp16")]; + tensor var_471_cast_fp16 = mul(x = linear_28_cast_fp16, y = var_470_cast_fp16)[name = tensor("op_471_cast_fp16")]; + tensor input_77_to_fp16 = const()[name = tensor("input_77_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(18998464)))]; + tensor input_79_cast_fp16 = silu(x = input_77_to_fp16)[name = tensor("input_79_cast_fp16")]; + tensor linear_30_cast_fp16 = linear(bias = flow_net_time_embed_1_mlp_2_bias_to_fp16, weight = flow_net_time_embed_1_mlp_2_weight_to_fp16, x = input_79_cast_fp16)[name = tensor("linear_30_cast_fp16")]; + tensor reduce_mean_20_axes_0 = const()[name = tensor("reduce_mean_20_axes_0"), val = tensor([-1])]; + tensor reduce_mean_20_keep_dims_0 = const()[name = tensor("reduce_mean_20_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_20_cast_fp16 = reduce_mean(axes = reduce_mean_20_axes_0, keep_dims = reduce_mean_20_keep_dims_0, x = linear_30_cast_fp16)[name = tensor("reduce_mean_20_cast_fp16")]; + tensor sub_13_cast_fp16 = sub(x = linear_30_cast_fp16, y = reduce_mean_20_cast_fp16)[name = tensor("sub_13_cast_fp16")]; + tensor square_10_cast_fp16 = square(x = sub_13_cast_fp16)[name = tensor("square_10_cast_fp16")]; + tensor reduce_mean_21_axes_0 = const()[name = tensor("reduce_mean_21_axes_0"), val = tensor([-1])]; + tensor reduce_mean_21_keep_dims_0 = const()[name = tensor("reduce_mean_21_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_21_cast_fp16 = reduce_mean(axes = reduce_mean_21_axes_0, keep_dims = reduce_mean_21_keep_dims_0, x = square_10_cast_fp16)[name = tensor("reduce_mean_21_cast_fp16")]; + tensor real_div_3_to_fp16 = const()[name = tensor("real_div_3_to_fp16"), val = tensor(0x1.008p+0)]; + tensor mul_3_cast_fp16 = mul(x = reduce_mean_21_cast_fp16, y = real_div_3_to_fp16)[name = tensor("mul_3_cast_fp16")]; + tensor var_503_to_fp16 = const()[name = tensor("op_503_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_21_cast_fp16 = add(x = mul_3_cast_fp16, y = var_503_to_fp16)[name = tensor("var_21_cast_fp16")]; + tensor var_506_epsilon_0 = const()[name = tensor("op_506_epsilon_0"), val = tensor(0x1.197998p-40)]; + tensor var_506_cast_fp16 = rsqrt(epsilon = var_506_epsilon_0, x = var_21_cast_fp16)[name = tensor("op_506_cast_fp16")]; + tensor var_507_cast_fp16 = mul(x = const_5_to_fp16, y = var_506_cast_fp16)[name = tensor("op_507_cast_fp16")]; + tensor var_508_cast_fp16 = mul(x = linear_30_cast_fp16, y = var_507_cast_fp16)[name = tensor("op_508_cast_fp16")]; + tensor var_520_cast_fp16 = add(x = var_471_cast_fp16, y = var_508_cast_fp16)[name = tensor("op_520_cast_fp16")]; + tensor _inversed_t_combined_3_y_0_to_fp16 = const()[name = tensor("_inversed_t_combined_3_y_0_to_fp16"), val = tensor(0x1p-1)]; + tensor _inversed_t_combined_3_cast_fp16 = mul(x = var_520_cast_fp16, y = _inversed_t_combined_3_y_0_to_fp16)[name = tensor("_inversed_t_combined_3_cast_fp16")]; + tensor input_81_cast_fp16 = add(x = _inversed_t_combined_3_cast_fp16, y = linear_5_cast_fp16)[name = tensor("input_81_cast_fp16")]; + tensor input_83_cast_fp16 = silu(x = input_81_cast_fp16)[name = tensor("input_83_cast_fp16")]; + tensor linear_32_cast_fp16 = linear(bias = flow_net_res_blocks_0_adaLN_modulation_1_bias_to_fp16, weight = flow_net_res_blocks_0_adaLN_modulation_1_weight_to_fp16, x = input_83_cast_fp16)[name = tensor("linear_32_cast_fp16")]; + tensor var_535_split_sizes_0 = const()[name = tensor("op_535_split_sizes_0"), val = tensor([512, 512, 512])]; + tensor var_535_axis_0 = const()[name = tensor("op_535_axis_0"), val = tensor(-1)]; + tensor var_535_cast_fp16_0, tensor var_535_cast_fp16_1, tensor var_535_cast_fp16_2 = split(axis = var_535_axis_0, split_sizes = var_535_split_sizes_0, x = linear_32_cast_fp16)[name = tensor("op_535_cast_fp16")]; + tensor mean_15_axes_0 = const()[name = tensor("mean_15_axes_0"), val = tensor([-1])]; + tensor mean_15_keep_dims_0 = const()[name = tensor("mean_15_keep_dims_0"), val = tensor(true)]; + tensor mean_15_cast_fp16 = reduce_mean(axes = mean_15_axes_0, keep_dims = mean_15_keep_dims_0, x = linear_26_cast_fp16)[name = tensor("mean_15_cast_fp16")]; + tensor sub_15_cast_fp16 = sub(x = linear_26_cast_fp16, y = mean_15_cast_fp16)[name = tensor("sub_15_cast_fp16")]; + tensor square_11_cast_fp16 = square(x = sub_15_cast_fp16)[name = tensor("square_11_cast_fp16")]; + tensor reduce_mean_23_axes_0 = const()[name = tensor("reduce_mean_23_axes_0"), val = tensor([-1])]; + tensor reduce_mean_23_keep_dims_0 = const()[name = tensor("reduce_mean_23_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_23_cast_fp16 = reduce_mean(axes = reduce_mean_23_axes_0, keep_dims = reduce_mean_23_keep_dims_0, x = square_11_cast_fp16)[name = tensor("reduce_mean_23_cast_fp16")]; + tensor var_545_to_fp16 = const()[name = tensor("op_545_to_fp16"), val = tensor(0x1.1p-20)]; + tensor var_546_cast_fp16 = add(x = reduce_mean_23_cast_fp16, y = var_545_to_fp16)[name = tensor("op_546_cast_fp16")]; + tensor var_547_cast_fp16 = sqrt(x = var_546_cast_fp16)[name = tensor("op_547_cast_fp16")]; + tensor x_51_cast_fp16 = real_div(x = sub_15_cast_fp16, y = var_547_cast_fp16)[name = tensor("x_51_cast_fp16")]; + tensor var_549_cast_fp16 = mul(x = x_51_cast_fp16, y = flow_net_res_blocks_0_in_ln_weight_to_fp16)[name = tensor("op_549_cast_fp16")]; + tensor x_53_cast_fp16 = add(x = var_549_cast_fp16, y = flow_net_res_blocks_0_in_ln_bias_to_fp16)[name = tensor("x_53_cast_fp16")]; + tensor var_551_promoted_to_fp16 = const()[name = tensor("op_551_promoted_to_fp16"), val = tensor(0x1p+0)]; + tensor var_552_cast_fp16 = add(x = var_535_cast_fp16_1, y = var_551_promoted_to_fp16)[name = tensor("op_552_cast_fp16")]; + tensor var_553_cast_fp16 = mul(x = x_53_cast_fp16, y = var_552_cast_fp16)[name = tensor("op_553_cast_fp16")]; + tensor input_85_cast_fp16 = add(x = var_553_cast_fp16, y = var_535_cast_fp16_0)[name = tensor("input_85_cast_fp16")]; + tensor linear_33_cast_fp16 = linear(bias = flow_net_res_blocks_0_mlp_0_bias_to_fp16, weight = flow_net_res_blocks_0_mlp_0_weight_to_fp16, x = input_85_cast_fp16)[name = tensor("linear_33_cast_fp16")]; + tensor input_89_cast_fp16 = silu(x = linear_33_cast_fp16)[name = tensor("input_89_cast_fp16")]; + tensor linear_34_cast_fp16 = linear(bias = flow_net_res_blocks_0_mlp_2_bias_to_fp16, weight = flow_net_res_blocks_0_mlp_2_weight_to_fp16, x = input_89_cast_fp16)[name = tensor("linear_34_cast_fp16")]; + tensor var_564_cast_fp16 = mul(x = var_535_cast_fp16_2, y = linear_34_cast_fp16)[name = tensor("op_564_cast_fp16")]; + tensor x_55_cast_fp16 = add(x = linear_26_cast_fp16, y = var_564_cast_fp16)[name = tensor("x_55_cast_fp16")]; + tensor linear_35_cast_fp16 = linear(bias = flow_net_res_blocks_1_adaLN_modulation_1_bias_to_fp16, weight = flow_net_res_blocks_1_adaLN_modulation_1_weight_to_fp16, x = input_83_cast_fp16)[name = tensor("linear_35_cast_fp16")]; + tensor var_574_split_sizes_0 = const()[name = tensor("op_574_split_sizes_0"), val = tensor([512, 512, 512])]; + tensor var_574_axis_0 = const()[name = tensor("op_574_axis_0"), val = tensor(-1)]; + tensor var_574_cast_fp16_0, tensor var_574_cast_fp16_1, tensor var_574_cast_fp16_2 = split(axis = var_574_axis_0, split_sizes = var_574_split_sizes_0, x = linear_35_cast_fp16)[name = tensor("op_574_cast_fp16")]; + tensor mean_17_axes_0 = const()[name = tensor("mean_17_axes_0"), val = tensor([-1])]; + tensor mean_17_keep_dims_0 = const()[name = tensor("mean_17_keep_dims_0"), val = tensor(true)]; + tensor mean_17_cast_fp16 = reduce_mean(axes = mean_17_axes_0, keep_dims = mean_17_keep_dims_0, x = x_55_cast_fp16)[name = tensor("mean_17_cast_fp16")]; + tensor sub_16_cast_fp16 = sub(x = x_55_cast_fp16, y = mean_17_cast_fp16)[name = tensor("sub_16_cast_fp16")]; + tensor square_12_cast_fp16 = square(x = sub_16_cast_fp16)[name = tensor("square_12_cast_fp16")]; + tensor reduce_mean_25_axes_0 = const()[name = tensor("reduce_mean_25_axes_0"), val = tensor([-1])]; + tensor reduce_mean_25_keep_dims_0 = const()[name = tensor("reduce_mean_25_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_25_cast_fp16 = reduce_mean(axes = reduce_mean_25_axes_0, keep_dims = reduce_mean_25_keep_dims_0, x = square_12_cast_fp16)[name = tensor("reduce_mean_25_cast_fp16")]; + tensor var_584_to_fp16 = const()[name = tensor("op_584_to_fp16"), val = tensor(0x1.1p-20)]; + tensor var_585_cast_fp16 = add(x = reduce_mean_25_cast_fp16, y = var_584_to_fp16)[name = tensor("op_585_cast_fp16")]; + tensor var_586_cast_fp16 = sqrt(x = var_585_cast_fp16)[name = tensor("op_586_cast_fp16")]; + tensor x_57_cast_fp16 = real_div(x = sub_16_cast_fp16, y = var_586_cast_fp16)[name = tensor("x_57_cast_fp16")]; + tensor var_588_cast_fp16 = mul(x = x_57_cast_fp16, y = flow_net_res_blocks_1_in_ln_weight_to_fp16)[name = tensor("op_588_cast_fp16")]; + tensor x_59_cast_fp16 = add(x = var_588_cast_fp16, y = flow_net_res_blocks_1_in_ln_bias_to_fp16)[name = tensor("x_59_cast_fp16")]; + tensor var_590_promoted_to_fp16 = const()[name = tensor("op_590_promoted_to_fp16"), val = tensor(0x1p+0)]; + tensor var_591_cast_fp16 = add(x = var_574_cast_fp16_1, y = var_590_promoted_to_fp16)[name = tensor("op_591_cast_fp16")]; + tensor var_592_cast_fp16 = mul(x = x_59_cast_fp16, y = var_591_cast_fp16)[name = tensor("op_592_cast_fp16")]; + tensor input_93_cast_fp16 = add(x = var_592_cast_fp16, y = var_574_cast_fp16_0)[name = tensor("input_93_cast_fp16")]; + tensor linear_36_cast_fp16 = linear(bias = flow_net_res_blocks_1_mlp_0_bias_to_fp16, weight = flow_net_res_blocks_1_mlp_0_weight_to_fp16, x = input_93_cast_fp16)[name = tensor("linear_36_cast_fp16")]; + tensor input_97_cast_fp16 = silu(x = linear_36_cast_fp16)[name = tensor("input_97_cast_fp16")]; + tensor linear_37_cast_fp16 = linear(bias = flow_net_res_blocks_1_mlp_2_bias_to_fp16, weight = flow_net_res_blocks_1_mlp_2_weight_to_fp16, x = input_97_cast_fp16)[name = tensor("linear_37_cast_fp16")]; + tensor var_603_cast_fp16 = mul(x = var_574_cast_fp16_2, y = linear_37_cast_fp16)[name = tensor("op_603_cast_fp16")]; + tensor x_61_cast_fp16 = add(x = x_55_cast_fp16, y = var_603_cast_fp16)[name = tensor("x_61_cast_fp16")]; + tensor linear_38_cast_fp16 = linear(bias = flow_net_res_blocks_2_adaLN_modulation_1_bias_to_fp16, weight = flow_net_res_blocks_2_adaLN_modulation_1_weight_to_fp16, x = input_83_cast_fp16)[name = tensor("linear_38_cast_fp16")]; + tensor var_613_split_sizes_0 = const()[name = tensor("op_613_split_sizes_0"), val = tensor([512, 512, 512])]; + tensor var_613_axis_0 = const()[name = tensor("op_613_axis_0"), val = tensor(-1)]; + tensor var_613_cast_fp16_0, tensor var_613_cast_fp16_1, tensor var_613_cast_fp16_2 = split(axis = var_613_axis_0, split_sizes = var_613_split_sizes_0, x = linear_38_cast_fp16)[name = tensor("op_613_cast_fp16")]; + tensor mean_19_axes_0 = const()[name = tensor("mean_19_axes_0"), val = tensor([-1])]; + tensor mean_19_keep_dims_0 = const()[name = tensor("mean_19_keep_dims_0"), val = tensor(true)]; + tensor mean_19_cast_fp16 = reduce_mean(axes = mean_19_axes_0, keep_dims = mean_19_keep_dims_0, x = x_61_cast_fp16)[name = tensor("mean_19_cast_fp16")]; + tensor sub_17_cast_fp16 = sub(x = x_61_cast_fp16, y = mean_19_cast_fp16)[name = tensor("sub_17_cast_fp16")]; + tensor square_13_cast_fp16 = square(x = sub_17_cast_fp16)[name = tensor("square_13_cast_fp16")]; + tensor reduce_mean_27_axes_0 = const()[name = tensor("reduce_mean_27_axes_0"), val = tensor([-1])]; + tensor reduce_mean_27_keep_dims_0 = const()[name = tensor("reduce_mean_27_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_27_cast_fp16 = reduce_mean(axes = reduce_mean_27_axes_0, keep_dims = reduce_mean_27_keep_dims_0, x = square_13_cast_fp16)[name = tensor("reduce_mean_27_cast_fp16")]; + tensor var_623_to_fp16 = const()[name = tensor("op_623_to_fp16"), val = tensor(0x1.1p-20)]; + tensor var_624_cast_fp16 = add(x = reduce_mean_27_cast_fp16, y = var_623_to_fp16)[name = tensor("op_624_cast_fp16")]; + tensor var_625_cast_fp16 = sqrt(x = var_624_cast_fp16)[name = tensor("op_625_cast_fp16")]; + tensor x_63_cast_fp16 = real_div(x = sub_17_cast_fp16, y = var_625_cast_fp16)[name = tensor("x_63_cast_fp16")]; + tensor var_627_cast_fp16 = mul(x = x_63_cast_fp16, y = flow_net_res_blocks_2_in_ln_weight_to_fp16)[name = tensor("op_627_cast_fp16")]; + tensor x_65_cast_fp16 = add(x = var_627_cast_fp16, y = flow_net_res_blocks_2_in_ln_bias_to_fp16)[name = tensor("x_65_cast_fp16")]; + tensor var_629_promoted_to_fp16 = const()[name = tensor("op_629_promoted_to_fp16"), val = tensor(0x1p+0)]; + tensor var_630_cast_fp16 = add(x = var_613_cast_fp16_1, y = var_629_promoted_to_fp16)[name = tensor("op_630_cast_fp16")]; + tensor var_631_cast_fp16 = mul(x = x_65_cast_fp16, y = var_630_cast_fp16)[name = tensor("op_631_cast_fp16")]; + tensor input_101_cast_fp16 = add(x = var_631_cast_fp16, y = var_613_cast_fp16_0)[name = tensor("input_101_cast_fp16")]; + tensor linear_39_cast_fp16 = linear(bias = flow_net_res_blocks_2_mlp_0_bias_to_fp16, weight = flow_net_res_blocks_2_mlp_0_weight_to_fp16, x = input_101_cast_fp16)[name = tensor("linear_39_cast_fp16")]; + tensor input_105_cast_fp16 = silu(x = linear_39_cast_fp16)[name = tensor("input_105_cast_fp16")]; + tensor linear_40_cast_fp16 = linear(bias = flow_net_res_blocks_2_mlp_2_bias_to_fp16, weight = flow_net_res_blocks_2_mlp_2_weight_to_fp16, x = input_105_cast_fp16)[name = tensor("linear_40_cast_fp16")]; + tensor var_642_cast_fp16 = mul(x = var_613_cast_fp16_2, y = linear_40_cast_fp16)[name = tensor("op_642_cast_fp16")]; + tensor x_67_cast_fp16 = add(x = x_61_cast_fp16, y = var_642_cast_fp16)[name = tensor("x_67_cast_fp16")]; + tensor linear_41_cast_fp16 = linear(bias = flow_net_res_blocks_3_adaLN_modulation_1_bias_to_fp16, weight = flow_net_res_blocks_3_adaLN_modulation_1_weight_to_fp16, x = input_83_cast_fp16)[name = tensor("linear_41_cast_fp16")]; + tensor var_652_split_sizes_0 = const()[name = tensor("op_652_split_sizes_0"), val = tensor([512, 512, 512])]; + tensor var_652_axis_0 = const()[name = tensor("op_652_axis_0"), val = tensor(-1)]; + tensor var_652_cast_fp16_0, tensor var_652_cast_fp16_1, tensor var_652_cast_fp16_2 = split(axis = var_652_axis_0, split_sizes = var_652_split_sizes_0, x = linear_41_cast_fp16)[name = tensor("op_652_cast_fp16")]; + tensor mean_21_axes_0 = const()[name = tensor("mean_21_axes_0"), val = tensor([-1])]; + tensor mean_21_keep_dims_0 = const()[name = tensor("mean_21_keep_dims_0"), val = tensor(true)]; + tensor mean_21_cast_fp16 = reduce_mean(axes = mean_21_axes_0, keep_dims = mean_21_keep_dims_0, x = x_67_cast_fp16)[name = tensor("mean_21_cast_fp16")]; + tensor sub_18_cast_fp16 = sub(x = x_67_cast_fp16, y = mean_21_cast_fp16)[name = tensor("sub_18_cast_fp16")]; + tensor square_14_cast_fp16 = square(x = sub_18_cast_fp16)[name = tensor("square_14_cast_fp16")]; + tensor reduce_mean_29_axes_0 = const()[name = tensor("reduce_mean_29_axes_0"), val = tensor([-1])]; + tensor reduce_mean_29_keep_dims_0 = const()[name = tensor("reduce_mean_29_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_29_cast_fp16 = reduce_mean(axes = reduce_mean_29_axes_0, keep_dims = reduce_mean_29_keep_dims_0, x = square_14_cast_fp16)[name = tensor("reduce_mean_29_cast_fp16")]; + tensor var_662_to_fp16 = const()[name = tensor("op_662_to_fp16"), val = tensor(0x1.1p-20)]; + tensor var_663_cast_fp16 = add(x = reduce_mean_29_cast_fp16, y = var_662_to_fp16)[name = tensor("op_663_cast_fp16")]; + tensor var_664_cast_fp16 = sqrt(x = var_663_cast_fp16)[name = tensor("op_664_cast_fp16")]; + tensor x_69_cast_fp16 = real_div(x = sub_18_cast_fp16, y = var_664_cast_fp16)[name = tensor("x_69_cast_fp16")]; + tensor var_666_cast_fp16 = mul(x = x_69_cast_fp16, y = flow_net_res_blocks_3_in_ln_weight_to_fp16)[name = tensor("op_666_cast_fp16")]; + tensor x_71_cast_fp16 = add(x = var_666_cast_fp16, y = flow_net_res_blocks_3_in_ln_bias_to_fp16)[name = tensor("x_71_cast_fp16")]; + tensor var_668_promoted_to_fp16 = const()[name = tensor("op_668_promoted_to_fp16"), val = tensor(0x1p+0)]; + tensor var_669_cast_fp16 = add(x = var_652_cast_fp16_1, y = var_668_promoted_to_fp16)[name = tensor("op_669_cast_fp16")]; + tensor var_670_cast_fp16 = mul(x = x_71_cast_fp16, y = var_669_cast_fp16)[name = tensor("op_670_cast_fp16")]; + tensor input_109_cast_fp16 = add(x = var_670_cast_fp16, y = var_652_cast_fp16_0)[name = tensor("input_109_cast_fp16")]; + tensor linear_42_cast_fp16 = linear(bias = flow_net_res_blocks_3_mlp_0_bias_to_fp16, weight = flow_net_res_blocks_3_mlp_0_weight_to_fp16, x = input_109_cast_fp16)[name = tensor("linear_42_cast_fp16")]; + tensor input_113_cast_fp16 = silu(x = linear_42_cast_fp16)[name = tensor("input_113_cast_fp16")]; + tensor linear_43_cast_fp16 = linear(bias = flow_net_res_blocks_3_mlp_2_bias_to_fp16, weight = flow_net_res_blocks_3_mlp_2_weight_to_fp16, x = input_113_cast_fp16)[name = tensor("linear_43_cast_fp16")]; + tensor var_681_cast_fp16 = mul(x = var_652_cast_fp16_2, y = linear_43_cast_fp16)[name = tensor("op_681_cast_fp16")]; + tensor x_73_cast_fp16 = add(x = x_67_cast_fp16, y = var_681_cast_fp16)[name = tensor("x_73_cast_fp16")]; + tensor linear_44_cast_fp16 = linear(bias = flow_net_res_blocks_4_adaLN_modulation_1_bias_to_fp16, weight = flow_net_res_blocks_4_adaLN_modulation_1_weight_to_fp16, x = input_83_cast_fp16)[name = tensor("linear_44_cast_fp16")]; + tensor var_691_split_sizes_0 = const()[name = tensor("op_691_split_sizes_0"), val = tensor([512, 512, 512])]; + tensor var_691_axis_0 = const()[name = tensor("op_691_axis_0"), val = tensor(-1)]; + tensor var_691_cast_fp16_0, tensor var_691_cast_fp16_1, tensor var_691_cast_fp16_2 = split(axis = var_691_axis_0, split_sizes = var_691_split_sizes_0, x = linear_44_cast_fp16)[name = tensor("op_691_cast_fp16")]; + tensor mean_23_axes_0 = const()[name = tensor("mean_23_axes_0"), val = tensor([-1])]; + tensor mean_23_keep_dims_0 = const()[name = tensor("mean_23_keep_dims_0"), val = tensor(true)]; + tensor mean_23_cast_fp16 = reduce_mean(axes = mean_23_axes_0, keep_dims = mean_23_keep_dims_0, x = x_73_cast_fp16)[name = tensor("mean_23_cast_fp16")]; + tensor sub_19_cast_fp16 = sub(x = x_73_cast_fp16, y = mean_23_cast_fp16)[name = tensor("sub_19_cast_fp16")]; + tensor square_15_cast_fp16 = square(x = sub_19_cast_fp16)[name = tensor("square_15_cast_fp16")]; + tensor reduce_mean_31_axes_0 = const()[name = tensor("reduce_mean_31_axes_0"), val = tensor([-1])]; + tensor reduce_mean_31_keep_dims_0 = const()[name = tensor("reduce_mean_31_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_31_cast_fp16 = reduce_mean(axes = reduce_mean_31_axes_0, keep_dims = reduce_mean_31_keep_dims_0, x = square_15_cast_fp16)[name = tensor("reduce_mean_31_cast_fp16")]; + tensor var_701_to_fp16 = const()[name = tensor("op_701_to_fp16"), val = tensor(0x1.1p-20)]; + tensor var_702_cast_fp16 = add(x = reduce_mean_31_cast_fp16, y = var_701_to_fp16)[name = tensor("op_702_cast_fp16")]; + tensor var_703_cast_fp16 = sqrt(x = var_702_cast_fp16)[name = tensor("op_703_cast_fp16")]; + tensor x_75_cast_fp16 = real_div(x = sub_19_cast_fp16, y = var_703_cast_fp16)[name = tensor("x_75_cast_fp16")]; + tensor var_705_cast_fp16 = mul(x = x_75_cast_fp16, y = flow_net_res_blocks_4_in_ln_weight_to_fp16)[name = tensor("op_705_cast_fp16")]; + tensor x_77_cast_fp16 = add(x = var_705_cast_fp16, y = flow_net_res_blocks_4_in_ln_bias_to_fp16)[name = tensor("x_77_cast_fp16")]; + tensor var_707_promoted_to_fp16 = const()[name = tensor("op_707_promoted_to_fp16"), val = tensor(0x1p+0)]; + tensor var_708_cast_fp16 = add(x = var_691_cast_fp16_1, y = var_707_promoted_to_fp16)[name = tensor("op_708_cast_fp16")]; + tensor var_709_cast_fp16 = mul(x = x_77_cast_fp16, y = var_708_cast_fp16)[name = tensor("op_709_cast_fp16")]; + tensor input_117_cast_fp16 = add(x = var_709_cast_fp16, y = var_691_cast_fp16_0)[name = tensor("input_117_cast_fp16")]; + tensor linear_45_cast_fp16 = linear(bias = flow_net_res_blocks_4_mlp_0_bias_to_fp16, weight = flow_net_res_blocks_4_mlp_0_weight_to_fp16, x = input_117_cast_fp16)[name = tensor("linear_45_cast_fp16")]; + tensor input_121_cast_fp16 = silu(x = linear_45_cast_fp16)[name = tensor("input_121_cast_fp16")]; + tensor linear_46_cast_fp16 = linear(bias = flow_net_res_blocks_4_mlp_2_bias_to_fp16, weight = flow_net_res_blocks_4_mlp_2_weight_to_fp16, x = input_121_cast_fp16)[name = tensor("linear_46_cast_fp16")]; + tensor var_720_cast_fp16 = mul(x = var_691_cast_fp16_2, y = linear_46_cast_fp16)[name = tensor("op_720_cast_fp16")]; + tensor x_79_cast_fp16 = add(x = x_73_cast_fp16, y = var_720_cast_fp16)[name = tensor("x_79_cast_fp16")]; + tensor linear_47_cast_fp16 = linear(bias = flow_net_res_blocks_5_adaLN_modulation_1_bias_to_fp16, weight = flow_net_res_blocks_5_adaLN_modulation_1_weight_to_fp16, x = input_83_cast_fp16)[name = tensor("linear_47_cast_fp16")]; + tensor var_730_split_sizes_0 = const()[name = tensor("op_730_split_sizes_0"), val = tensor([512, 512, 512])]; + tensor var_730_axis_0 = const()[name = tensor("op_730_axis_0"), val = tensor(-1)]; + tensor var_730_cast_fp16_0, tensor var_730_cast_fp16_1, tensor var_730_cast_fp16_2 = split(axis = var_730_axis_0, split_sizes = var_730_split_sizes_0, x = linear_47_cast_fp16)[name = tensor("op_730_cast_fp16")]; + tensor mean_25_axes_0 = const()[name = tensor("mean_25_axes_0"), val = tensor([-1])]; + tensor mean_25_keep_dims_0 = const()[name = tensor("mean_25_keep_dims_0"), val = tensor(true)]; + tensor mean_25_cast_fp16 = reduce_mean(axes = mean_25_axes_0, keep_dims = mean_25_keep_dims_0, x = x_79_cast_fp16)[name = tensor("mean_25_cast_fp16")]; + tensor sub_20_cast_fp16 = sub(x = x_79_cast_fp16, y = mean_25_cast_fp16)[name = tensor("sub_20_cast_fp16")]; + tensor square_16_cast_fp16 = square(x = sub_20_cast_fp16)[name = tensor("square_16_cast_fp16")]; + tensor reduce_mean_33_axes_0 = const()[name = tensor("reduce_mean_33_axes_0"), val = tensor([-1])]; + tensor reduce_mean_33_keep_dims_0 = const()[name = tensor("reduce_mean_33_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_33_cast_fp16 = reduce_mean(axes = reduce_mean_33_axes_0, keep_dims = reduce_mean_33_keep_dims_0, x = square_16_cast_fp16)[name = tensor("reduce_mean_33_cast_fp16")]; + tensor var_740_to_fp16 = const()[name = tensor("op_740_to_fp16"), val = tensor(0x1.1p-20)]; + tensor var_741_cast_fp16 = add(x = reduce_mean_33_cast_fp16, y = var_740_to_fp16)[name = tensor("op_741_cast_fp16")]; + tensor var_742_cast_fp16 = sqrt(x = var_741_cast_fp16)[name = tensor("op_742_cast_fp16")]; + tensor x_81_cast_fp16 = real_div(x = sub_20_cast_fp16, y = var_742_cast_fp16)[name = tensor("x_81_cast_fp16")]; + tensor var_744_cast_fp16 = mul(x = x_81_cast_fp16, y = flow_net_res_blocks_5_in_ln_weight_to_fp16)[name = tensor("op_744_cast_fp16")]; + tensor x_83_cast_fp16 = add(x = var_744_cast_fp16, y = flow_net_res_blocks_5_in_ln_bias_to_fp16)[name = tensor("x_83_cast_fp16")]; + tensor var_746_promoted_to_fp16 = const()[name = tensor("op_746_promoted_to_fp16"), val = tensor(0x1p+0)]; + tensor var_747_cast_fp16 = add(x = var_730_cast_fp16_1, y = var_746_promoted_to_fp16)[name = tensor("op_747_cast_fp16")]; + tensor var_748_cast_fp16 = mul(x = x_83_cast_fp16, y = var_747_cast_fp16)[name = tensor("op_748_cast_fp16")]; + tensor input_125_cast_fp16 = add(x = var_748_cast_fp16, y = var_730_cast_fp16_0)[name = tensor("input_125_cast_fp16")]; + tensor linear_48_cast_fp16 = linear(bias = flow_net_res_blocks_5_mlp_0_bias_to_fp16, weight = flow_net_res_blocks_5_mlp_0_weight_to_fp16, x = input_125_cast_fp16)[name = tensor("linear_48_cast_fp16")]; + tensor input_129_cast_fp16 = silu(x = linear_48_cast_fp16)[name = tensor("input_129_cast_fp16")]; + tensor linear_49_cast_fp16 = linear(bias = flow_net_res_blocks_5_mlp_2_bias_to_fp16, weight = flow_net_res_blocks_5_mlp_2_weight_to_fp16, x = input_129_cast_fp16)[name = tensor("linear_49_cast_fp16")]; + tensor var_759_cast_fp16 = mul(x = var_730_cast_fp16_2, y = linear_49_cast_fp16)[name = tensor("op_759_cast_fp16")]; + tensor x_85_cast_fp16 = add(x = x_79_cast_fp16, y = var_759_cast_fp16)[name = tensor("x_85_cast_fp16")]; + tensor linear_50_cast_fp16 = linear(bias = flow_net_final_layer_adaLN_modulation_1_bias_to_fp16, weight = flow_net_final_layer_adaLN_modulation_1_weight_to_fp16, x = input_83_cast_fp16)[name = tensor("linear_50_cast_fp16")]; + tensor var_768_split_sizes_0 = const()[name = tensor("op_768_split_sizes_0"), val = tensor([512, 512])]; + tensor var_768_axis_0 = const()[name = tensor("op_768_axis_0"), val = tensor(-1)]; + tensor var_768_cast_fp16_0, tensor var_768_cast_fp16_1 = split(axis = var_768_axis_0, split_sizes = var_768_split_sizes_0, x = linear_50_cast_fp16)[name = tensor("op_768_cast_fp16")]; + tensor mean_27_axes_0 = const()[name = tensor("mean_27_axes_0"), val = tensor([-1])]; + tensor mean_27_keep_dims_0 = const()[name = tensor("mean_27_keep_dims_0"), val = tensor(true)]; + tensor mean_27_cast_fp16 = reduce_mean(axes = mean_27_axes_0, keep_dims = mean_27_keep_dims_0, x = x_85_cast_fp16)[name = tensor("mean_27_cast_fp16")]; + tensor sub_21_cast_fp16 = sub(x = x_85_cast_fp16, y = mean_27_cast_fp16)[name = tensor("sub_21_cast_fp16")]; + tensor square_17_cast_fp16 = square(x = sub_21_cast_fp16)[name = tensor("square_17_cast_fp16")]; + tensor reduce_mean_35_axes_0 = const()[name = tensor("reduce_mean_35_axes_0"), val = tensor([-1])]; + tensor reduce_mean_35_keep_dims_0 = const()[name = tensor("reduce_mean_35_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_35_cast_fp16 = reduce_mean(axes = reduce_mean_35_axes_0, keep_dims = reduce_mean_35_keep_dims_0, x = square_17_cast_fp16)[name = tensor("reduce_mean_35_cast_fp16")]; + tensor var_775_to_fp16 = const()[name = tensor("op_775_to_fp16"), val = tensor(0x1.1p-20)]; + tensor var_776_cast_fp16 = add(x = reduce_mean_35_cast_fp16, y = var_775_to_fp16)[name = tensor("op_776_cast_fp16")]; + tensor var_777_cast_fp16 = sqrt(x = var_776_cast_fp16)[name = tensor("op_777_cast_fp16")]; + tensor x_87_cast_fp16 = real_div(x = sub_21_cast_fp16, y = var_777_cast_fp16)[name = tensor("x_87_cast_fp16")]; + tensor var_779_promoted_to_fp16 = const()[name = tensor("op_779_promoted_to_fp16"), val = tensor(0x1p+0)]; + tensor var_780_cast_fp16 = add(x = var_768_cast_fp16_1, y = var_779_promoted_to_fp16)[name = tensor("op_780_cast_fp16")]; + tensor var_781_cast_fp16 = mul(x = x_87_cast_fp16, y = var_780_cast_fp16)[name = tensor("op_781_cast_fp16")]; + tensor input_133_cast_fp16 = add(x = var_781_cast_fp16, y = var_768_cast_fp16_0)[name = tensor("input_133_cast_fp16")]; + tensor linear_51_cast_fp16 = linear(bias = flow_net_final_layer_linear_bias_to_fp16, weight = flow_net_final_layer_linear_weight_to_fp16, x = input_133_cast_fp16)[name = tensor("linear_51_cast_fp16")]; + tensor var_792_to_fp16 = const()[name = tensor("op_792_to_fp16"), val = tensor(0x1p-3)]; + tensor var_793_cast_fp16 = mul(x = linear_51_cast_fp16, y = var_792_to_fp16)[name = tensor("op_793_cast_fp16")]; + tensor input_135_cast_fp16 = add(x = input_67_cast_fp16, y = var_793_cast_fp16)[name = tensor("input_135_cast_fp16")]; + tensor linear_52_cast_fp16 = linear(bias = flow_net_input_proj_bias_to_fp16, weight = flow_net_input_proj_weight_to_fp16, x = input_135_cast_fp16)[name = tensor("linear_52_cast_fp16")]; + tensor input_139_to_fp16 = const()[name = tensor("input_139_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(18999552)))]; + tensor input_141_cast_fp16 = silu(x = input_139_to_fp16)[name = tensor("input_141_cast_fp16")]; + tensor linear_54_cast_fp16 = linear(bias = flow_net_time_embed_0_mlp_2_bias_to_fp16, weight = flow_net_time_embed_0_mlp_2_weight_to_fp16, x = input_141_cast_fp16)[name = tensor("linear_54_cast_fp16")]; + tensor reduce_mean_36_axes_0 = const()[name = tensor("reduce_mean_36_axes_0"), val = tensor([-1])]; + tensor reduce_mean_36_keep_dims_0 = const()[name = tensor("reduce_mean_36_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_36_cast_fp16 = reduce_mean(axes = reduce_mean_36_axes_0, keep_dims = reduce_mean_36_keep_dims_0, x = linear_54_cast_fp16)[name = tensor("reduce_mean_36_cast_fp16")]; + tensor sub_22_cast_fp16 = sub(x = linear_54_cast_fp16, y = reduce_mean_36_cast_fp16)[name = tensor("sub_22_cast_fp16")]; + tensor square_18_cast_fp16 = square(x = sub_22_cast_fp16)[name = tensor("square_18_cast_fp16")]; + tensor reduce_mean_37_axes_0 = const()[name = tensor("reduce_mean_37_axes_0"), val = tensor([-1])]; + tensor reduce_mean_37_keep_dims_0 = const()[name = tensor("reduce_mean_37_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_37_cast_fp16 = reduce_mean(axes = reduce_mean_37_axes_0, keep_dims = reduce_mean_37_keep_dims_0, x = square_18_cast_fp16)[name = tensor("reduce_mean_37_cast_fp16")]; + tensor real_div_4_to_fp16 = const()[name = tensor("real_div_4_to_fp16"), val = tensor(0x1.008p+0)]; + tensor mul_4_cast_fp16 = mul(x = reduce_mean_37_cast_fp16, y = real_div_4_to_fp16)[name = tensor("mul_4_cast_fp16")]; + tensor var_859_to_fp16 = const()[name = tensor("op_859_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_37_cast_fp16 = add(x = mul_4_cast_fp16, y = var_859_to_fp16)[name = tensor("var_37_cast_fp16")]; + tensor var_862_epsilon_0 = const()[name = tensor("op_862_epsilon_0"), val = tensor(0x1.197998p-40)]; + tensor var_862_cast_fp16 = rsqrt(epsilon = var_862_epsilon_0, x = var_37_cast_fp16)[name = tensor("op_862_cast_fp16")]; + tensor var_863_cast_fp16 = mul(x = const_3_to_fp16, y = var_862_cast_fp16)[name = tensor("op_863_cast_fp16")]; + tensor var_864_cast_fp16 = mul(x = linear_54_cast_fp16, y = var_863_cast_fp16)[name = tensor("op_864_cast_fp16")]; + tensor input_145_to_fp16 = const()[name = tensor("input_145_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(19000640)))]; + tensor input_147_cast_fp16 = silu(x = input_145_to_fp16)[name = tensor("input_147_cast_fp16")]; + tensor linear_56_cast_fp16 = linear(bias = flow_net_time_embed_1_mlp_2_bias_to_fp16, weight = flow_net_time_embed_1_mlp_2_weight_to_fp16, x = input_147_cast_fp16)[name = tensor("linear_56_cast_fp16")]; + tensor reduce_mean_38_axes_0 = const()[name = tensor("reduce_mean_38_axes_0"), val = tensor([-1])]; + tensor reduce_mean_38_keep_dims_0 = const()[name = tensor("reduce_mean_38_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_38_cast_fp16 = reduce_mean(axes = reduce_mean_38_axes_0, keep_dims = reduce_mean_38_keep_dims_0, x = linear_56_cast_fp16)[name = tensor("reduce_mean_38_cast_fp16")]; + tensor sub_24_cast_fp16 = sub(x = linear_56_cast_fp16, y = reduce_mean_38_cast_fp16)[name = tensor("sub_24_cast_fp16")]; + tensor square_19_cast_fp16 = square(x = sub_24_cast_fp16)[name = tensor("square_19_cast_fp16")]; + tensor reduce_mean_39_axes_0 = const()[name = tensor("reduce_mean_39_axes_0"), val = tensor([-1])]; + tensor reduce_mean_39_keep_dims_0 = const()[name = tensor("reduce_mean_39_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_39_cast_fp16 = reduce_mean(axes = reduce_mean_39_axes_0, keep_dims = reduce_mean_39_keep_dims_0, x = square_19_cast_fp16)[name = tensor("reduce_mean_39_cast_fp16")]; + tensor real_div_5_to_fp16 = const()[name = tensor("real_div_5_to_fp16"), val = tensor(0x1.008p+0)]; + tensor mul_5_cast_fp16 = mul(x = reduce_mean_39_cast_fp16, y = real_div_5_to_fp16)[name = tensor("mul_5_cast_fp16")]; + tensor var_896_to_fp16 = const()[name = tensor("op_896_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_39_cast_fp16 = add(x = mul_5_cast_fp16, y = var_896_to_fp16)[name = tensor("var_39_cast_fp16")]; + tensor var_899_epsilon_0 = const()[name = tensor("op_899_epsilon_0"), val = tensor(0x1.197998p-40)]; + tensor var_899_cast_fp16 = rsqrt(epsilon = var_899_epsilon_0, x = var_39_cast_fp16)[name = tensor("op_899_cast_fp16")]; + tensor var_900_cast_fp16 = mul(x = const_5_to_fp16, y = var_899_cast_fp16)[name = tensor("op_900_cast_fp16")]; + tensor var_901_cast_fp16 = mul(x = linear_56_cast_fp16, y = var_900_cast_fp16)[name = tensor("op_901_cast_fp16")]; + tensor var_913_cast_fp16 = add(x = var_864_cast_fp16, y = var_901_cast_fp16)[name = tensor("op_913_cast_fp16")]; + tensor _inversed_t_combined_5_y_0_to_fp16 = const()[name = tensor("_inversed_t_combined_5_y_0_to_fp16"), val = tensor(0x1p-1)]; + tensor _inversed_t_combined_5_cast_fp16 = mul(x = var_913_cast_fp16, y = _inversed_t_combined_5_y_0_to_fp16)[name = tensor("_inversed_t_combined_5_cast_fp16")]; + tensor input_149_cast_fp16 = add(x = _inversed_t_combined_5_cast_fp16, y = linear_5_cast_fp16)[name = tensor("input_149_cast_fp16")]; + tensor input_151_cast_fp16 = silu(x = input_149_cast_fp16)[name = tensor("input_151_cast_fp16")]; + tensor linear_58_cast_fp16 = linear(bias = flow_net_res_blocks_0_adaLN_modulation_1_bias_to_fp16, weight = flow_net_res_blocks_0_adaLN_modulation_1_weight_to_fp16, x = input_151_cast_fp16)[name = tensor("linear_58_cast_fp16")]; + tensor var_928_split_sizes_0 = const()[name = tensor("op_928_split_sizes_0"), val = tensor([512, 512, 512])]; + tensor var_928_axis_0 = const()[name = tensor("op_928_axis_0"), val = tensor(-1)]; + tensor var_928_cast_fp16_0, tensor var_928_cast_fp16_1, tensor var_928_cast_fp16_2 = split(axis = var_928_axis_0, split_sizes = var_928_split_sizes_0, x = linear_58_cast_fp16)[name = tensor("op_928_cast_fp16")]; + tensor mean_29_axes_0 = const()[name = tensor("mean_29_axes_0"), val = tensor([-1])]; + tensor mean_29_keep_dims_0 = const()[name = tensor("mean_29_keep_dims_0"), val = tensor(true)]; + tensor mean_29_cast_fp16 = reduce_mean(axes = mean_29_axes_0, keep_dims = mean_29_keep_dims_0, x = linear_52_cast_fp16)[name = tensor("mean_29_cast_fp16")]; + tensor sub_26_cast_fp16 = sub(x = linear_52_cast_fp16, y = mean_29_cast_fp16)[name = tensor("sub_26_cast_fp16")]; + tensor square_20_cast_fp16 = square(x = sub_26_cast_fp16)[name = tensor("square_20_cast_fp16")]; + tensor reduce_mean_41_axes_0 = const()[name = tensor("reduce_mean_41_axes_0"), val = tensor([-1])]; + tensor reduce_mean_41_keep_dims_0 = const()[name = tensor("reduce_mean_41_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_41_cast_fp16 = reduce_mean(axes = reduce_mean_41_axes_0, keep_dims = reduce_mean_41_keep_dims_0, x = square_20_cast_fp16)[name = tensor("reduce_mean_41_cast_fp16")]; + tensor var_938_to_fp16 = const()[name = tensor("op_938_to_fp16"), val = tensor(0x1.1p-20)]; + tensor var_939_cast_fp16 = add(x = reduce_mean_41_cast_fp16, y = var_938_to_fp16)[name = tensor("op_939_cast_fp16")]; + tensor var_940_cast_fp16 = sqrt(x = var_939_cast_fp16)[name = tensor("op_940_cast_fp16")]; + tensor x_95_cast_fp16 = real_div(x = sub_26_cast_fp16, y = var_940_cast_fp16)[name = tensor("x_95_cast_fp16")]; + tensor var_942_cast_fp16 = mul(x = x_95_cast_fp16, y = flow_net_res_blocks_0_in_ln_weight_to_fp16)[name = tensor("op_942_cast_fp16")]; + tensor x_97_cast_fp16 = add(x = var_942_cast_fp16, y = flow_net_res_blocks_0_in_ln_bias_to_fp16)[name = tensor("x_97_cast_fp16")]; + tensor var_944_promoted_to_fp16 = const()[name = tensor("op_944_promoted_to_fp16"), val = tensor(0x1p+0)]; + tensor var_945_cast_fp16 = add(x = var_928_cast_fp16_1, y = var_944_promoted_to_fp16)[name = tensor("op_945_cast_fp16")]; + tensor var_946_cast_fp16 = mul(x = x_97_cast_fp16, y = var_945_cast_fp16)[name = tensor("op_946_cast_fp16")]; + tensor input_153_cast_fp16 = add(x = var_946_cast_fp16, y = var_928_cast_fp16_0)[name = tensor("input_153_cast_fp16")]; + tensor linear_59_cast_fp16 = linear(bias = flow_net_res_blocks_0_mlp_0_bias_to_fp16, weight = flow_net_res_blocks_0_mlp_0_weight_to_fp16, x = input_153_cast_fp16)[name = tensor("linear_59_cast_fp16")]; + tensor input_157_cast_fp16 = silu(x = linear_59_cast_fp16)[name = tensor("input_157_cast_fp16")]; + tensor linear_60_cast_fp16 = linear(bias = flow_net_res_blocks_0_mlp_2_bias_to_fp16, weight = flow_net_res_blocks_0_mlp_2_weight_to_fp16, x = input_157_cast_fp16)[name = tensor("linear_60_cast_fp16")]; + tensor var_957_cast_fp16 = mul(x = var_928_cast_fp16_2, y = linear_60_cast_fp16)[name = tensor("op_957_cast_fp16")]; + tensor x_99_cast_fp16 = add(x = linear_52_cast_fp16, y = var_957_cast_fp16)[name = tensor("x_99_cast_fp16")]; + tensor linear_61_cast_fp16 = linear(bias = flow_net_res_blocks_1_adaLN_modulation_1_bias_to_fp16, weight = flow_net_res_blocks_1_adaLN_modulation_1_weight_to_fp16, x = input_151_cast_fp16)[name = tensor("linear_61_cast_fp16")]; + tensor var_967_split_sizes_0 = const()[name = tensor("op_967_split_sizes_0"), val = tensor([512, 512, 512])]; + tensor var_967_axis_0 = const()[name = tensor("op_967_axis_0"), val = tensor(-1)]; + tensor var_967_cast_fp16_0, tensor var_967_cast_fp16_1, tensor var_967_cast_fp16_2 = split(axis = var_967_axis_0, split_sizes = var_967_split_sizes_0, x = linear_61_cast_fp16)[name = tensor("op_967_cast_fp16")]; + tensor mean_31_axes_0 = const()[name = tensor("mean_31_axes_0"), val = tensor([-1])]; + tensor mean_31_keep_dims_0 = const()[name = tensor("mean_31_keep_dims_0"), val = tensor(true)]; + tensor mean_31_cast_fp16 = reduce_mean(axes = mean_31_axes_0, keep_dims = mean_31_keep_dims_0, x = x_99_cast_fp16)[name = tensor("mean_31_cast_fp16")]; + tensor sub_27_cast_fp16 = sub(x = x_99_cast_fp16, y = mean_31_cast_fp16)[name = tensor("sub_27_cast_fp16")]; + tensor square_21_cast_fp16 = square(x = sub_27_cast_fp16)[name = tensor("square_21_cast_fp16")]; + tensor reduce_mean_43_axes_0 = const()[name = tensor("reduce_mean_43_axes_0"), val = tensor([-1])]; + tensor reduce_mean_43_keep_dims_0 = const()[name = tensor("reduce_mean_43_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_43_cast_fp16 = reduce_mean(axes = reduce_mean_43_axes_0, keep_dims = reduce_mean_43_keep_dims_0, x = square_21_cast_fp16)[name = tensor("reduce_mean_43_cast_fp16")]; + tensor var_977_to_fp16 = const()[name = tensor("op_977_to_fp16"), val = tensor(0x1.1p-20)]; + tensor var_978_cast_fp16 = add(x = reduce_mean_43_cast_fp16, y = var_977_to_fp16)[name = tensor("op_978_cast_fp16")]; + tensor var_979_cast_fp16 = sqrt(x = var_978_cast_fp16)[name = tensor("op_979_cast_fp16")]; + tensor x_101_cast_fp16 = real_div(x = sub_27_cast_fp16, y = var_979_cast_fp16)[name = tensor("x_101_cast_fp16")]; + tensor var_981_cast_fp16 = mul(x = x_101_cast_fp16, y = flow_net_res_blocks_1_in_ln_weight_to_fp16)[name = tensor("op_981_cast_fp16")]; + tensor x_103_cast_fp16 = add(x = var_981_cast_fp16, y = flow_net_res_blocks_1_in_ln_bias_to_fp16)[name = tensor("x_103_cast_fp16")]; + tensor var_983_promoted_to_fp16 = const()[name = tensor("op_983_promoted_to_fp16"), val = tensor(0x1p+0)]; + tensor var_984_cast_fp16 = add(x = var_967_cast_fp16_1, y = var_983_promoted_to_fp16)[name = tensor("op_984_cast_fp16")]; + tensor var_985_cast_fp16 = mul(x = x_103_cast_fp16, y = var_984_cast_fp16)[name = tensor("op_985_cast_fp16")]; + tensor input_161_cast_fp16 = add(x = var_985_cast_fp16, y = var_967_cast_fp16_0)[name = tensor("input_161_cast_fp16")]; + tensor linear_62_cast_fp16 = linear(bias = flow_net_res_blocks_1_mlp_0_bias_to_fp16, weight = flow_net_res_blocks_1_mlp_0_weight_to_fp16, x = input_161_cast_fp16)[name = tensor("linear_62_cast_fp16")]; + tensor input_165_cast_fp16 = silu(x = linear_62_cast_fp16)[name = tensor("input_165_cast_fp16")]; + tensor linear_63_cast_fp16 = linear(bias = flow_net_res_blocks_1_mlp_2_bias_to_fp16, weight = flow_net_res_blocks_1_mlp_2_weight_to_fp16, x = input_165_cast_fp16)[name = tensor("linear_63_cast_fp16")]; + tensor var_996_cast_fp16 = mul(x = var_967_cast_fp16_2, y = linear_63_cast_fp16)[name = tensor("op_996_cast_fp16")]; + tensor x_105_cast_fp16 = add(x = x_99_cast_fp16, y = var_996_cast_fp16)[name = tensor("x_105_cast_fp16")]; + tensor linear_64_cast_fp16 = linear(bias = flow_net_res_blocks_2_adaLN_modulation_1_bias_to_fp16, weight = flow_net_res_blocks_2_adaLN_modulation_1_weight_to_fp16, x = input_151_cast_fp16)[name = tensor("linear_64_cast_fp16")]; + tensor var_1006_split_sizes_0 = const()[name = tensor("op_1006_split_sizes_0"), val = tensor([512, 512, 512])]; + tensor var_1006_axis_0 = const()[name = tensor("op_1006_axis_0"), val = tensor(-1)]; + tensor var_1006_cast_fp16_0, tensor var_1006_cast_fp16_1, tensor var_1006_cast_fp16_2 = split(axis = var_1006_axis_0, split_sizes = var_1006_split_sizes_0, x = linear_64_cast_fp16)[name = tensor("op_1006_cast_fp16")]; + tensor mean_33_axes_0 = const()[name = tensor("mean_33_axes_0"), val = tensor([-1])]; + tensor mean_33_keep_dims_0 = const()[name = tensor("mean_33_keep_dims_0"), val = tensor(true)]; + tensor mean_33_cast_fp16 = reduce_mean(axes = mean_33_axes_0, keep_dims = mean_33_keep_dims_0, x = x_105_cast_fp16)[name = tensor("mean_33_cast_fp16")]; + tensor sub_28_cast_fp16 = sub(x = x_105_cast_fp16, y = mean_33_cast_fp16)[name = tensor("sub_28_cast_fp16")]; + tensor square_22_cast_fp16 = square(x = sub_28_cast_fp16)[name = tensor("square_22_cast_fp16")]; + tensor reduce_mean_45_axes_0 = const()[name = tensor("reduce_mean_45_axes_0"), val = tensor([-1])]; + tensor reduce_mean_45_keep_dims_0 = const()[name = tensor("reduce_mean_45_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_45_cast_fp16 = reduce_mean(axes = reduce_mean_45_axes_0, keep_dims = reduce_mean_45_keep_dims_0, x = square_22_cast_fp16)[name = tensor("reduce_mean_45_cast_fp16")]; + tensor var_1016_to_fp16 = const()[name = tensor("op_1016_to_fp16"), val = tensor(0x1.1p-20)]; + tensor var_1017_cast_fp16 = add(x = reduce_mean_45_cast_fp16, y = var_1016_to_fp16)[name = tensor("op_1017_cast_fp16")]; + tensor var_1018_cast_fp16 = sqrt(x = var_1017_cast_fp16)[name = tensor("op_1018_cast_fp16")]; + tensor x_107_cast_fp16 = real_div(x = sub_28_cast_fp16, y = var_1018_cast_fp16)[name = tensor("x_107_cast_fp16")]; + tensor var_1020_cast_fp16 = mul(x = x_107_cast_fp16, y = flow_net_res_blocks_2_in_ln_weight_to_fp16)[name = tensor("op_1020_cast_fp16")]; + tensor x_109_cast_fp16 = add(x = var_1020_cast_fp16, y = flow_net_res_blocks_2_in_ln_bias_to_fp16)[name = tensor("x_109_cast_fp16")]; + tensor var_1022_promoted_to_fp16 = const()[name = tensor("op_1022_promoted_to_fp16"), val = tensor(0x1p+0)]; + tensor var_1023_cast_fp16 = add(x = var_1006_cast_fp16_1, y = var_1022_promoted_to_fp16)[name = tensor("op_1023_cast_fp16")]; + tensor var_1024_cast_fp16 = mul(x = x_109_cast_fp16, y = var_1023_cast_fp16)[name = tensor("op_1024_cast_fp16")]; + tensor input_169_cast_fp16 = add(x = var_1024_cast_fp16, y = var_1006_cast_fp16_0)[name = tensor("input_169_cast_fp16")]; + tensor linear_65_cast_fp16 = linear(bias = flow_net_res_blocks_2_mlp_0_bias_to_fp16, weight = flow_net_res_blocks_2_mlp_0_weight_to_fp16, x = input_169_cast_fp16)[name = tensor("linear_65_cast_fp16")]; + tensor input_173_cast_fp16 = silu(x = linear_65_cast_fp16)[name = tensor("input_173_cast_fp16")]; + tensor linear_66_cast_fp16 = linear(bias = flow_net_res_blocks_2_mlp_2_bias_to_fp16, weight = flow_net_res_blocks_2_mlp_2_weight_to_fp16, x = input_173_cast_fp16)[name = tensor("linear_66_cast_fp16")]; + tensor var_1035_cast_fp16 = mul(x = var_1006_cast_fp16_2, y = linear_66_cast_fp16)[name = tensor("op_1035_cast_fp16")]; + tensor x_111_cast_fp16 = add(x = x_105_cast_fp16, y = var_1035_cast_fp16)[name = tensor("x_111_cast_fp16")]; + tensor linear_67_cast_fp16 = linear(bias = flow_net_res_blocks_3_adaLN_modulation_1_bias_to_fp16, weight = flow_net_res_blocks_3_adaLN_modulation_1_weight_to_fp16, x = input_151_cast_fp16)[name = tensor("linear_67_cast_fp16")]; + tensor var_1045_split_sizes_0 = const()[name = tensor("op_1045_split_sizes_0"), val = tensor([512, 512, 512])]; + tensor var_1045_axis_0 = const()[name = tensor("op_1045_axis_0"), val = tensor(-1)]; + tensor var_1045_cast_fp16_0, tensor var_1045_cast_fp16_1, tensor var_1045_cast_fp16_2 = split(axis = var_1045_axis_0, split_sizes = var_1045_split_sizes_0, x = linear_67_cast_fp16)[name = tensor("op_1045_cast_fp16")]; + tensor mean_35_axes_0 = const()[name = tensor("mean_35_axes_0"), val = tensor([-1])]; + tensor mean_35_keep_dims_0 = const()[name = tensor("mean_35_keep_dims_0"), val = tensor(true)]; + tensor mean_35_cast_fp16 = reduce_mean(axes = mean_35_axes_0, keep_dims = mean_35_keep_dims_0, x = x_111_cast_fp16)[name = tensor("mean_35_cast_fp16")]; + tensor sub_29_cast_fp16 = sub(x = x_111_cast_fp16, y = mean_35_cast_fp16)[name = tensor("sub_29_cast_fp16")]; + tensor square_23_cast_fp16 = square(x = sub_29_cast_fp16)[name = tensor("square_23_cast_fp16")]; + tensor reduce_mean_47_axes_0 = const()[name = tensor("reduce_mean_47_axes_0"), val = tensor([-1])]; + tensor reduce_mean_47_keep_dims_0 = const()[name = tensor("reduce_mean_47_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_47_cast_fp16 = reduce_mean(axes = reduce_mean_47_axes_0, keep_dims = reduce_mean_47_keep_dims_0, x = square_23_cast_fp16)[name = tensor("reduce_mean_47_cast_fp16")]; + tensor var_1055_to_fp16 = const()[name = tensor("op_1055_to_fp16"), val = tensor(0x1.1p-20)]; + tensor var_1056_cast_fp16 = add(x = reduce_mean_47_cast_fp16, y = var_1055_to_fp16)[name = tensor("op_1056_cast_fp16")]; + tensor var_1057_cast_fp16 = sqrt(x = var_1056_cast_fp16)[name = tensor("op_1057_cast_fp16")]; + tensor x_113_cast_fp16 = real_div(x = sub_29_cast_fp16, y = var_1057_cast_fp16)[name = tensor("x_113_cast_fp16")]; + tensor var_1059_cast_fp16 = mul(x = x_113_cast_fp16, y = flow_net_res_blocks_3_in_ln_weight_to_fp16)[name = tensor("op_1059_cast_fp16")]; + tensor x_115_cast_fp16 = add(x = var_1059_cast_fp16, y = flow_net_res_blocks_3_in_ln_bias_to_fp16)[name = tensor("x_115_cast_fp16")]; + tensor var_1061_promoted_to_fp16 = const()[name = tensor("op_1061_promoted_to_fp16"), val = tensor(0x1p+0)]; + tensor var_1062_cast_fp16 = add(x = var_1045_cast_fp16_1, y = var_1061_promoted_to_fp16)[name = tensor("op_1062_cast_fp16")]; + tensor var_1063_cast_fp16 = mul(x = x_115_cast_fp16, y = var_1062_cast_fp16)[name = tensor("op_1063_cast_fp16")]; + tensor input_177_cast_fp16 = add(x = var_1063_cast_fp16, y = var_1045_cast_fp16_0)[name = tensor("input_177_cast_fp16")]; + tensor linear_68_cast_fp16 = linear(bias = flow_net_res_blocks_3_mlp_0_bias_to_fp16, weight = flow_net_res_blocks_3_mlp_0_weight_to_fp16, x = input_177_cast_fp16)[name = tensor("linear_68_cast_fp16")]; + tensor input_181_cast_fp16 = silu(x = linear_68_cast_fp16)[name = tensor("input_181_cast_fp16")]; + tensor linear_69_cast_fp16 = linear(bias = flow_net_res_blocks_3_mlp_2_bias_to_fp16, weight = flow_net_res_blocks_3_mlp_2_weight_to_fp16, x = input_181_cast_fp16)[name = tensor("linear_69_cast_fp16")]; + tensor var_1074_cast_fp16 = mul(x = var_1045_cast_fp16_2, y = linear_69_cast_fp16)[name = tensor("op_1074_cast_fp16")]; + tensor x_117_cast_fp16 = add(x = x_111_cast_fp16, y = var_1074_cast_fp16)[name = tensor("x_117_cast_fp16")]; + tensor linear_70_cast_fp16 = linear(bias = flow_net_res_blocks_4_adaLN_modulation_1_bias_to_fp16, weight = flow_net_res_blocks_4_adaLN_modulation_1_weight_to_fp16, x = input_151_cast_fp16)[name = tensor("linear_70_cast_fp16")]; + tensor var_1084_split_sizes_0 = const()[name = tensor("op_1084_split_sizes_0"), val = tensor([512, 512, 512])]; + tensor var_1084_axis_0 = const()[name = tensor("op_1084_axis_0"), val = tensor(-1)]; + tensor var_1084_cast_fp16_0, tensor var_1084_cast_fp16_1, tensor var_1084_cast_fp16_2 = split(axis = var_1084_axis_0, split_sizes = var_1084_split_sizes_0, x = linear_70_cast_fp16)[name = tensor("op_1084_cast_fp16")]; + tensor mean_37_axes_0 = const()[name = tensor("mean_37_axes_0"), val = tensor([-1])]; + tensor mean_37_keep_dims_0 = const()[name = tensor("mean_37_keep_dims_0"), val = tensor(true)]; + tensor mean_37_cast_fp16 = reduce_mean(axes = mean_37_axes_0, keep_dims = mean_37_keep_dims_0, x = x_117_cast_fp16)[name = tensor("mean_37_cast_fp16")]; + tensor sub_30_cast_fp16 = sub(x = x_117_cast_fp16, y = mean_37_cast_fp16)[name = tensor("sub_30_cast_fp16")]; + tensor square_24_cast_fp16 = square(x = sub_30_cast_fp16)[name = tensor("square_24_cast_fp16")]; + tensor reduce_mean_49_axes_0 = const()[name = tensor("reduce_mean_49_axes_0"), val = tensor([-1])]; + tensor reduce_mean_49_keep_dims_0 = const()[name = tensor("reduce_mean_49_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_49_cast_fp16 = reduce_mean(axes = reduce_mean_49_axes_0, keep_dims = reduce_mean_49_keep_dims_0, x = square_24_cast_fp16)[name = tensor("reduce_mean_49_cast_fp16")]; + tensor var_1094_to_fp16 = const()[name = tensor("op_1094_to_fp16"), val = tensor(0x1.1p-20)]; + tensor var_1095_cast_fp16 = add(x = reduce_mean_49_cast_fp16, y = var_1094_to_fp16)[name = tensor("op_1095_cast_fp16")]; + tensor var_1096_cast_fp16 = sqrt(x = var_1095_cast_fp16)[name = tensor("op_1096_cast_fp16")]; + tensor x_119_cast_fp16 = real_div(x = sub_30_cast_fp16, y = var_1096_cast_fp16)[name = tensor("x_119_cast_fp16")]; + tensor var_1098_cast_fp16 = mul(x = x_119_cast_fp16, y = flow_net_res_blocks_4_in_ln_weight_to_fp16)[name = tensor("op_1098_cast_fp16")]; + tensor x_121_cast_fp16 = add(x = var_1098_cast_fp16, y = flow_net_res_blocks_4_in_ln_bias_to_fp16)[name = tensor("x_121_cast_fp16")]; + tensor var_1100_promoted_to_fp16 = const()[name = tensor("op_1100_promoted_to_fp16"), val = tensor(0x1p+0)]; + tensor var_1101_cast_fp16 = add(x = var_1084_cast_fp16_1, y = var_1100_promoted_to_fp16)[name = tensor("op_1101_cast_fp16")]; + tensor var_1102_cast_fp16 = mul(x = x_121_cast_fp16, y = var_1101_cast_fp16)[name = tensor("op_1102_cast_fp16")]; + tensor input_185_cast_fp16 = add(x = var_1102_cast_fp16, y = var_1084_cast_fp16_0)[name = tensor("input_185_cast_fp16")]; + tensor linear_71_cast_fp16 = linear(bias = flow_net_res_blocks_4_mlp_0_bias_to_fp16, weight = flow_net_res_blocks_4_mlp_0_weight_to_fp16, x = input_185_cast_fp16)[name = tensor("linear_71_cast_fp16")]; + tensor input_189_cast_fp16 = silu(x = linear_71_cast_fp16)[name = tensor("input_189_cast_fp16")]; + tensor linear_72_cast_fp16 = linear(bias = flow_net_res_blocks_4_mlp_2_bias_to_fp16, weight = flow_net_res_blocks_4_mlp_2_weight_to_fp16, x = input_189_cast_fp16)[name = tensor("linear_72_cast_fp16")]; + tensor var_1113_cast_fp16 = mul(x = var_1084_cast_fp16_2, y = linear_72_cast_fp16)[name = tensor("op_1113_cast_fp16")]; + tensor x_123_cast_fp16 = add(x = x_117_cast_fp16, y = var_1113_cast_fp16)[name = tensor("x_123_cast_fp16")]; + tensor linear_73_cast_fp16 = linear(bias = flow_net_res_blocks_5_adaLN_modulation_1_bias_to_fp16, weight = flow_net_res_blocks_5_adaLN_modulation_1_weight_to_fp16, x = input_151_cast_fp16)[name = tensor("linear_73_cast_fp16")]; + tensor var_1123_split_sizes_0 = const()[name = tensor("op_1123_split_sizes_0"), val = tensor([512, 512, 512])]; + tensor var_1123_axis_0 = const()[name = tensor("op_1123_axis_0"), val = tensor(-1)]; + tensor var_1123_cast_fp16_0, tensor var_1123_cast_fp16_1, tensor var_1123_cast_fp16_2 = split(axis = var_1123_axis_0, split_sizes = var_1123_split_sizes_0, x = linear_73_cast_fp16)[name = tensor("op_1123_cast_fp16")]; + tensor mean_39_axes_0 = const()[name = tensor("mean_39_axes_0"), val = tensor([-1])]; + tensor mean_39_keep_dims_0 = const()[name = tensor("mean_39_keep_dims_0"), val = tensor(true)]; + tensor mean_39_cast_fp16 = reduce_mean(axes = mean_39_axes_0, keep_dims = mean_39_keep_dims_0, x = x_123_cast_fp16)[name = tensor("mean_39_cast_fp16")]; + tensor sub_31_cast_fp16 = sub(x = x_123_cast_fp16, y = mean_39_cast_fp16)[name = tensor("sub_31_cast_fp16")]; + tensor square_25_cast_fp16 = square(x = sub_31_cast_fp16)[name = tensor("square_25_cast_fp16")]; + tensor reduce_mean_51_axes_0 = const()[name = tensor("reduce_mean_51_axes_0"), val = tensor([-1])]; + tensor reduce_mean_51_keep_dims_0 = const()[name = tensor("reduce_mean_51_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_51_cast_fp16 = reduce_mean(axes = reduce_mean_51_axes_0, keep_dims = reduce_mean_51_keep_dims_0, x = square_25_cast_fp16)[name = tensor("reduce_mean_51_cast_fp16")]; + tensor var_1133_to_fp16 = const()[name = tensor("op_1133_to_fp16"), val = tensor(0x1.1p-20)]; + tensor var_1134_cast_fp16 = add(x = reduce_mean_51_cast_fp16, y = var_1133_to_fp16)[name = tensor("op_1134_cast_fp16")]; + tensor var_1135_cast_fp16 = sqrt(x = var_1134_cast_fp16)[name = tensor("op_1135_cast_fp16")]; + tensor x_125_cast_fp16 = real_div(x = sub_31_cast_fp16, y = var_1135_cast_fp16)[name = tensor("x_125_cast_fp16")]; + tensor var_1137_cast_fp16 = mul(x = x_125_cast_fp16, y = flow_net_res_blocks_5_in_ln_weight_to_fp16)[name = tensor("op_1137_cast_fp16")]; + tensor x_127_cast_fp16 = add(x = var_1137_cast_fp16, y = flow_net_res_blocks_5_in_ln_bias_to_fp16)[name = tensor("x_127_cast_fp16")]; + tensor var_1139_promoted_to_fp16 = const()[name = tensor("op_1139_promoted_to_fp16"), val = tensor(0x1p+0)]; + tensor var_1140_cast_fp16 = add(x = var_1123_cast_fp16_1, y = var_1139_promoted_to_fp16)[name = tensor("op_1140_cast_fp16")]; + tensor var_1141_cast_fp16 = mul(x = x_127_cast_fp16, y = var_1140_cast_fp16)[name = tensor("op_1141_cast_fp16")]; + tensor input_193_cast_fp16 = add(x = var_1141_cast_fp16, y = var_1123_cast_fp16_0)[name = tensor("input_193_cast_fp16")]; + tensor linear_74_cast_fp16 = linear(bias = flow_net_res_blocks_5_mlp_0_bias_to_fp16, weight = flow_net_res_blocks_5_mlp_0_weight_to_fp16, x = input_193_cast_fp16)[name = tensor("linear_74_cast_fp16")]; + tensor input_197_cast_fp16 = silu(x = linear_74_cast_fp16)[name = tensor("input_197_cast_fp16")]; + tensor linear_75_cast_fp16 = linear(bias = flow_net_res_blocks_5_mlp_2_bias_to_fp16, weight = flow_net_res_blocks_5_mlp_2_weight_to_fp16, x = input_197_cast_fp16)[name = tensor("linear_75_cast_fp16")]; + tensor var_1152_cast_fp16 = mul(x = var_1123_cast_fp16_2, y = linear_75_cast_fp16)[name = tensor("op_1152_cast_fp16")]; + tensor x_129_cast_fp16 = add(x = x_123_cast_fp16, y = var_1152_cast_fp16)[name = tensor("x_129_cast_fp16")]; + tensor linear_76_cast_fp16 = linear(bias = flow_net_final_layer_adaLN_modulation_1_bias_to_fp16, weight = flow_net_final_layer_adaLN_modulation_1_weight_to_fp16, x = input_151_cast_fp16)[name = tensor("linear_76_cast_fp16")]; + tensor var_1161_split_sizes_0 = const()[name = tensor("op_1161_split_sizes_0"), val = tensor([512, 512])]; + tensor var_1161_axis_0 = const()[name = tensor("op_1161_axis_0"), val = tensor(-1)]; + tensor var_1161_cast_fp16_0, tensor var_1161_cast_fp16_1 = split(axis = var_1161_axis_0, split_sizes = var_1161_split_sizes_0, x = linear_76_cast_fp16)[name = tensor("op_1161_cast_fp16")]; + tensor mean_41_axes_0 = const()[name = tensor("mean_41_axes_0"), val = tensor([-1])]; + tensor mean_41_keep_dims_0 = const()[name = tensor("mean_41_keep_dims_0"), val = tensor(true)]; + tensor mean_41_cast_fp16 = reduce_mean(axes = mean_41_axes_0, keep_dims = mean_41_keep_dims_0, x = x_129_cast_fp16)[name = tensor("mean_41_cast_fp16")]; + tensor sub_32_cast_fp16 = sub(x = x_129_cast_fp16, y = mean_41_cast_fp16)[name = tensor("sub_32_cast_fp16")]; + tensor square_26_cast_fp16 = square(x = sub_32_cast_fp16)[name = tensor("square_26_cast_fp16")]; + tensor reduce_mean_53_axes_0 = const()[name = tensor("reduce_mean_53_axes_0"), val = tensor([-1])]; + tensor reduce_mean_53_keep_dims_0 = const()[name = tensor("reduce_mean_53_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_53_cast_fp16 = reduce_mean(axes = reduce_mean_53_axes_0, keep_dims = reduce_mean_53_keep_dims_0, x = square_26_cast_fp16)[name = tensor("reduce_mean_53_cast_fp16")]; + tensor var_1168_to_fp16 = const()[name = tensor("op_1168_to_fp16"), val = tensor(0x1.1p-20)]; + tensor var_1169_cast_fp16 = add(x = reduce_mean_53_cast_fp16, y = var_1168_to_fp16)[name = tensor("op_1169_cast_fp16")]; + tensor var_1170_cast_fp16 = sqrt(x = var_1169_cast_fp16)[name = tensor("op_1170_cast_fp16")]; + tensor x_131_cast_fp16 = real_div(x = sub_32_cast_fp16, y = var_1170_cast_fp16)[name = tensor("x_131_cast_fp16")]; + tensor var_1172_promoted_to_fp16 = const()[name = tensor("op_1172_promoted_to_fp16"), val = tensor(0x1p+0)]; + tensor var_1173_cast_fp16 = add(x = var_1161_cast_fp16_1, y = var_1172_promoted_to_fp16)[name = tensor("op_1173_cast_fp16")]; + tensor var_1174_cast_fp16 = mul(x = x_131_cast_fp16, y = var_1173_cast_fp16)[name = tensor("op_1174_cast_fp16")]; + tensor input_201_cast_fp16 = add(x = var_1174_cast_fp16, y = var_1161_cast_fp16_0)[name = tensor("input_201_cast_fp16")]; + tensor linear_77_cast_fp16 = linear(bias = flow_net_final_layer_linear_bias_to_fp16, weight = flow_net_final_layer_linear_weight_to_fp16, x = input_201_cast_fp16)[name = tensor("linear_77_cast_fp16")]; + tensor var_1185_to_fp16 = const()[name = tensor("op_1185_to_fp16"), val = tensor(0x1p-3)]; + tensor var_1186_cast_fp16 = mul(x = linear_77_cast_fp16, y = var_1185_to_fp16)[name = tensor("op_1186_cast_fp16")]; + tensor input_203_cast_fp16 = add(x = input_135_cast_fp16, y = var_1186_cast_fp16)[name = tensor("input_203_cast_fp16")]; + tensor linear_78_cast_fp16 = linear(bias = flow_net_input_proj_bias_to_fp16, weight = flow_net_input_proj_weight_to_fp16, x = input_203_cast_fp16)[name = tensor("linear_78_cast_fp16")]; + tensor input_207_to_fp16 = const()[name = tensor("input_207_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(19001728)))]; + tensor input_209_cast_fp16 = silu(x = input_207_to_fp16)[name = tensor("input_209_cast_fp16")]; + tensor linear_80_cast_fp16 = linear(bias = flow_net_time_embed_0_mlp_2_bias_to_fp16, weight = flow_net_time_embed_0_mlp_2_weight_to_fp16, x = input_209_cast_fp16)[name = tensor("linear_80_cast_fp16")]; + tensor reduce_mean_54_axes_0 = const()[name = tensor("reduce_mean_54_axes_0"), val = tensor([-1])]; + tensor reduce_mean_54_keep_dims_0 = const()[name = tensor("reduce_mean_54_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_54_cast_fp16 = reduce_mean(axes = reduce_mean_54_axes_0, keep_dims = reduce_mean_54_keep_dims_0, x = linear_80_cast_fp16)[name = tensor("reduce_mean_54_cast_fp16")]; + tensor sub_33_cast_fp16 = sub(x = linear_80_cast_fp16, y = reduce_mean_54_cast_fp16)[name = tensor("sub_33_cast_fp16")]; + tensor square_27_cast_fp16 = square(x = sub_33_cast_fp16)[name = tensor("square_27_cast_fp16")]; + tensor reduce_mean_55_axes_0 = const()[name = tensor("reduce_mean_55_axes_0"), val = tensor([-1])]; + tensor reduce_mean_55_keep_dims_0 = const()[name = tensor("reduce_mean_55_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_55_cast_fp16 = reduce_mean(axes = reduce_mean_55_axes_0, keep_dims = reduce_mean_55_keep_dims_0, x = square_27_cast_fp16)[name = tensor("reduce_mean_55_cast_fp16")]; + tensor real_div_6_to_fp16 = const()[name = tensor("real_div_6_to_fp16"), val = tensor(0x1.008p+0)]; + tensor mul_6_cast_fp16 = mul(x = reduce_mean_55_cast_fp16, y = real_div_6_to_fp16)[name = tensor("mul_6_cast_fp16")]; + tensor var_1252_to_fp16 = const()[name = tensor("op_1252_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_55_cast_fp16 = add(x = mul_6_cast_fp16, y = var_1252_to_fp16)[name = tensor("var_55_cast_fp16")]; + tensor var_1255_epsilon_0 = const()[name = tensor("op_1255_epsilon_0"), val = tensor(0x1.197998p-40)]; + tensor var_1255_cast_fp16 = rsqrt(epsilon = var_1255_epsilon_0, x = var_55_cast_fp16)[name = tensor("op_1255_cast_fp16")]; + tensor var_1256_cast_fp16 = mul(x = const_3_to_fp16, y = var_1255_cast_fp16)[name = tensor("op_1256_cast_fp16")]; + tensor var_1257_cast_fp16 = mul(x = linear_80_cast_fp16, y = var_1256_cast_fp16)[name = tensor("op_1257_cast_fp16")]; + tensor input_213_to_fp16 = const()[name = tensor("input_213_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(19002816)))]; + tensor input_215_cast_fp16 = silu(x = input_213_to_fp16)[name = tensor("input_215_cast_fp16")]; + tensor linear_82_cast_fp16 = linear(bias = flow_net_time_embed_1_mlp_2_bias_to_fp16, weight = flow_net_time_embed_1_mlp_2_weight_to_fp16, x = input_215_cast_fp16)[name = tensor("linear_82_cast_fp16")]; + tensor reduce_mean_56_axes_0 = const()[name = tensor("reduce_mean_56_axes_0"), val = tensor([-1])]; + tensor reduce_mean_56_keep_dims_0 = const()[name = tensor("reduce_mean_56_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_56_cast_fp16 = reduce_mean(axes = reduce_mean_56_axes_0, keep_dims = reduce_mean_56_keep_dims_0, x = linear_82_cast_fp16)[name = tensor("reduce_mean_56_cast_fp16")]; + tensor sub_35_cast_fp16 = sub(x = linear_82_cast_fp16, y = reduce_mean_56_cast_fp16)[name = tensor("sub_35_cast_fp16")]; + tensor square_28_cast_fp16 = square(x = sub_35_cast_fp16)[name = tensor("square_28_cast_fp16")]; + tensor reduce_mean_57_axes_0 = const()[name = tensor("reduce_mean_57_axes_0"), val = tensor([-1])]; + tensor reduce_mean_57_keep_dims_0 = const()[name = tensor("reduce_mean_57_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_57_cast_fp16 = reduce_mean(axes = reduce_mean_57_axes_0, keep_dims = reduce_mean_57_keep_dims_0, x = square_28_cast_fp16)[name = tensor("reduce_mean_57_cast_fp16")]; + tensor real_div_7_to_fp16 = const()[name = tensor("real_div_7_to_fp16"), val = tensor(0x1.008p+0)]; + tensor mul_7_cast_fp16 = mul(x = reduce_mean_57_cast_fp16, y = real_div_7_to_fp16)[name = tensor("mul_7_cast_fp16")]; + tensor var_1289_to_fp16 = const()[name = tensor("op_1289_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_57_cast_fp16 = add(x = mul_7_cast_fp16, y = var_1289_to_fp16)[name = tensor("var_57_cast_fp16")]; + tensor var_1292_epsilon_0 = const()[name = tensor("op_1292_epsilon_0"), val = tensor(0x1.197998p-40)]; + tensor var_1292_cast_fp16 = rsqrt(epsilon = var_1292_epsilon_0, x = var_57_cast_fp16)[name = tensor("op_1292_cast_fp16")]; + tensor var_1293_cast_fp16 = mul(x = const_5_to_fp16, y = var_1292_cast_fp16)[name = tensor("op_1293_cast_fp16")]; + tensor var_1294_cast_fp16 = mul(x = linear_82_cast_fp16, y = var_1293_cast_fp16)[name = tensor("op_1294_cast_fp16")]; + tensor var_1306_cast_fp16 = add(x = var_1257_cast_fp16, y = var_1294_cast_fp16)[name = tensor("op_1306_cast_fp16")]; + tensor _inversed_t_combined_7_y_0_to_fp16 = const()[name = tensor("_inversed_t_combined_7_y_0_to_fp16"), val = tensor(0x1p-1)]; + tensor _inversed_t_combined_7_cast_fp16 = mul(x = var_1306_cast_fp16, y = _inversed_t_combined_7_y_0_to_fp16)[name = tensor("_inversed_t_combined_7_cast_fp16")]; + tensor input_217_cast_fp16 = add(x = _inversed_t_combined_7_cast_fp16, y = linear_5_cast_fp16)[name = tensor("input_217_cast_fp16")]; + tensor input_219_cast_fp16 = silu(x = input_217_cast_fp16)[name = tensor("input_219_cast_fp16")]; + tensor linear_84_cast_fp16 = linear(bias = flow_net_res_blocks_0_adaLN_modulation_1_bias_to_fp16, weight = flow_net_res_blocks_0_adaLN_modulation_1_weight_to_fp16, x = input_219_cast_fp16)[name = tensor("linear_84_cast_fp16")]; + tensor var_1321_split_sizes_0 = const()[name = tensor("op_1321_split_sizes_0"), val = tensor([512, 512, 512])]; + tensor var_1321_axis_0 = const()[name = tensor("op_1321_axis_0"), val = tensor(-1)]; + tensor var_1321_cast_fp16_0, tensor var_1321_cast_fp16_1, tensor var_1321_cast_fp16_2 = split(axis = var_1321_axis_0, split_sizes = var_1321_split_sizes_0, x = linear_84_cast_fp16)[name = tensor("op_1321_cast_fp16")]; + tensor mean_43_axes_0 = const()[name = tensor("mean_43_axes_0"), val = tensor([-1])]; + tensor mean_43_keep_dims_0 = const()[name = tensor("mean_43_keep_dims_0"), val = tensor(true)]; + tensor mean_43_cast_fp16 = reduce_mean(axes = mean_43_axes_0, keep_dims = mean_43_keep_dims_0, x = linear_78_cast_fp16)[name = tensor("mean_43_cast_fp16")]; + tensor sub_37_cast_fp16 = sub(x = linear_78_cast_fp16, y = mean_43_cast_fp16)[name = tensor("sub_37_cast_fp16")]; + tensor square_29_cast_fp16 = square(x = sub_37_cast_fp16)[name = tensor("square_29_cast_fp16")]; + tensor reduce_mean_59_axes_0 = const()[name = tensor("reduce_mean_59_axes_0"), val = tensor([-1])]; + tensor reduce_mean_59_keep_dims_0 = const()[name = tensor("reduce_mean_59_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_59_cast_fp16 = reduce_mean(axes = reduce_mean_59_axes_0, keep_dims = reduce_mean_59_keep_dims_0, x = square_29_cast_fp16)[name = tensor("reduce_mean_59_cast_fp16")]; + tensor var_1331_to_fp16 = const()[name = tensor("op_1331_to_fp16"), val = tensor(0x1.1p-20)]; + tensor var_1332_cast_fp16 = add(x = reduce_mean_59_cast_fp16, y = var_1331_to_fp16)[name = tensor("op_1332_cast_fp16")]; + tensor var_1333_cast_fp16 = sqrt(x = var_1332_cast_fp16)[name = tensor("op_1333_cast_fp16")]; + tensor x_139_cast_fp16 = real_div(x = sub_37_cast_fp16, y = var_1333_cast_fp16)[name = tensor("x_139_cast_fp16")]; + tensor var_1335_cast_fp16 = mul(x = x_139_cast_fp16, y = flow_net_res_blocks_0_in_ln_weight_to_fp16)[name = tensor("op_1335_cast_fp16")]; + tensor x_141_cast_fp16 = add(x = var_1335_cast_fp16, y = flow_net_res_blocks_0_in_ln_bias_to_fp16)[name = tensor("x_141_cast_fp16")]; + tensor var_1337_promoted_to_fp16 = const()[name = tensor("op_1337_promoted_to_fp16"), val = tensor(0x1p+0)]; + tensor var_1338_cast_fp16 = add(x = var_1321_cast_fp16_1, y = var_1337_promoted_to_fp16)[name = tensor("op_1338_cast_fp16")]; + tensor var_1339_cast_fp16 = mul(x = x_141_cast_fp16, y = var_1338_cast_fp16)[name = tensor("op_1339_cast_fp16")]; + tensor input_221_cast_fp16 = add(x = var_1339_cast_fp16, y = var_1321_cast_fp16_0)[name = tensor("input_221_cast_fp16")]; + tensor linear_85_cast_fp16 = linear(bias = flow_net_res_blocks_0_mlp_0_bias_to_fp16, weight = flow_net_res_blocks_0_mlp_0_weight_to_fp16, x = input_221_cast_fp16)[name = tensor("linear_85_cast_fp16")]; + tensor input_225_cast_fp16 = silu(x = linear_85_cast_fp16)[name = tensor("input_225_cast_fp16")]; + tensor linear_86_cast_fp16 = linear(bias = flow_net_res_blocks_0_mlp_2_bias_to_fp16, weight = flow_net_res_blocks_0_mlp_2_weight_to_fp16, x = input_225_cast_fp16)[name = tensor("linear_86_cast_fp16")]; + tensor var_1350_cast_fp16 = mul(x = var_1321_cast_fp16_2, y = linear_86_cast_fp16)[name = tensor("op_1350_cast_fp16")]; + tensor x_143_cast_fp16 = add(x = linear_78_cast_fp16, y = var_1350_cast_fp16)[name = tensor("x_143_cast_fp16")]; + tensor linear_87_cast_fp16 = linear(bias = flow_net_res_blocks_1_adaLN_modulation_1_bias_to_fp16, weight = flow_net_res_blocks_1_adaLN_modulation_1_weight_to_fp16, x = input_219_cast_fp16)[name = tensor("linear_87_cast_fp16")]; + tensor var_1360_split_sizes_0 = const()[name = tensor("op_1360_split_sizes_0"), val = tensor([512, 512, 512])]; + tensor var_1360_axis_0 = const()[name = tensor("op_1360_axis_0"), val = tensor(-1)]; + tensor var_1360_cast_fp16_0, tensor var_1360_cast_fp16_1, tensor var_1360_cast_fp16_2 = split(axis = var_1360_axis_0, split_sizes = var_1360_split_sizes_0, x = linear_87_cast_fp16)[name = tensor("op_1360_cast_fp16")]; + tensor mean_45_axes_0 = const()[name = tensor("mean_45_axes_0"), val = tensor([-1])]; + tensor mean_45_keep_dims_0 = const()[name = tensor("mean_45_keep_dims_0"), val = tensor(true)]; + tensor mean_45_cast_fp16 = reduce_mean(axes = mean_45_axes_0, keep_dims = mean_45_keep_dims_0, x = x_143_cast_fp16)[name = tensor("mean_45_cast_fp16")]; + tensor sub_38_cast_fp16 = sub(x = x_143_cast_fp16, y = mean_45_cast_fp16)[name = tensor("sub_38_cast_fp16")]; + tensor square_30_cast_fp16 = square(x = sub_38_cast_fp16)[name = tensor("square_30_cast_fp16")]; + tensor reduce_mean_61_axes_0 = const()[name = tensor("reduce_mean_61_axes_0"), val = tensor([-1])]; + tensor reduce_mean_61_keep_dims_0 = const()[name = tensor("reduce_mean_61_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_61_cast_fp16 = reduce_mean(axes = reduce_mean_61_axes_0, keep_dims = reduce_mean_61_keep_dims_0, x = square_30_cast_fp16)[name = tensor("reduce_mean_61_cast_fp16")]; + tensor var_1370_to_fp16 = const()[name = tensor("op_1370_to_fp16"), val = tensor(0x1.1p-20)]; + tensor var_1371_cast_fp16 = add(x = reduce_mean_61_cast_fp16, y = var_1370_to_fp16)[name = tensor("op_1371_cast_fp16")]; + tensor var_1372_cast_fp16 = sqrt(x = var_1371_cast_fp16)[name = tensor("op_1372_cast_fp16")]; + tensor x_145_cast_fp16 = real_div(x = sub_38_cast_fp16, y = var_1372_cast_fp16)[name = tensor("x_145_cast_fp16")]; + tensor var_1374_cast_fp16 = mul(x = x_145_cast_fp16, y = flow_net_res_blocks_1_in_ln_weight_to_fp16)[name = tensor("op_1374_cast_fp16")]; + tensor x_147_cast_fp16 = add(x = var_1374_cast_fp16, y = flow_net_res_blocks_1_in_ln_bias_to_fp16)[name = tensor("x_147_cast_fp16")]; + tensor var_1376_promoted_to_fp16 = const()[name = tensor("op_1376_promoted_to_fp16"), val = tensor(0x1p+0)]; + tensor var_1377_cast_fp16 = add(x = var_1360_cast_fp16_1, y = var_1376_promoted_to_fp16)[name = tensor("op_1377_cast_fp16")]; + tensor var_1378_cast_fp16 = mul(x = x_147_cast_fp16, y = var_1377_cast_fp16)[name = tensor("op_1378_cast_fp16")]; + tensor input_229_cast_fp16 = add(x = var_1378_cast_fp16, y = var_1360_cast_fp16_0)[name = tensor("input_229_cast_fp16")]; + tensor linear_88_cast_fp16 = linear(bias = flow_net_res_blocks_1_mlp_0_bias_to_fp16, weight = flow_net_res_blocks_1_mlp_0_weight_to_fp16, x = input_229_cast_fp16)[name = tensor("linear_88_cast_fp16")]; + tensor input_233_cast_fp16 = silu(x = linear_88_cast_fp16)[name = tensor("input_233_cast_fp16")]; + tensor linear_89_cast_fp16 = linear(bias = flow_net_res_blocks_1_mlp_2_bias_to_fp16, weight = flow_net_res_blocks_1_mlp_2_weight_to_fp16, x = input_233_cast_fp16)[name = tensor("linear_89_cast_fp16")]; + tensor var_1389_cast_fp16 = mul(x = var_1360_cast_fp16_2, y = linear_89_cast_fp16)[name = tensor("op_1389_cast_fp16")]; + tensor x_149_cast_fp16 = add(x = x_143_cast_fp16, y = var_1389_cast_fp16)[name = tensor("x_149_cast_fp16")]; + tensor linear_90_cast_fp16 = linear(bias = flow_net_res_blocks_2_adaLN_modulation_1_bias_to_fp16, weight = flow_net_res_blocks_2_adaLN_modulation_1_weight_to_fp16, x = input_219_cast_fp16)[name = tensor("linear_90_cast_fp16")]; + tensor var_1399_split_sizes_0 = const()[name = tensor("op_1399_split_sizes_0"), val = tensor([512, 512, 512])]; + tensor var_1399_axis_0 = const()[name = tensor("op_1399_axis_0"), val = tensor(-1)]; + tensor var_1399_cast_fp16_0, tensor var_1399_cast_fp16_1, tensor var_1399_cast_fp16_2 = split(axis = var_1399_axis_0, split_sizes = var_1399_split_sizes_0, x = linear_90_cast_fp16)[name = tensor("op_1399_cast_fp16")]; + tensor mean_47_axes_0 = const()[name = tensor("mean_47_axes_0"), val = tensor([-1])]; + tensor mean_47_keep_dims_0 = const()[name = tensor("mean_47_keep_dims_0"), val = tensor(true)]; + tensor mean_47_cast_fp16 = reduce_mean(axes = mean_47_axes_0, keep_dims = mean_47_keep_dims_0, x = x_149_cast_fp16)[name = tensor("mean_47_cast_fp16")]; + tensor sub_39_cast_fp16 = sub(x = x_149_cast_fp16, y = mean_47_cast_fp16)[name = tensor("sub_39_cast_fp16")]; + tensor square_31_cast_fp16 = square(x = sub_39_cast_fp16)[name = tensor("square_31_cast_fp16")]; + tensor reduce_mean_63_axes_0 = const()[name = tensor("reduce_mean_63_axes_0"), val = tensor([-1])]; + tensor reduce_mean_63_keep_dims_0 = const()[name = tensor("reduce_mean_63_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_63_cast_fp16 = reduce_mean(axes = reduce_mean_63_axes_0, keep_dims = reduce_mean_63_keep_dims_0, x = square_31_cast_fp16)[name = tensor("reduce_mean_63_cast_fp16")]; + tensor var_1409_to_fp16 = const()[name = tensor("op_1409_to_fp16"), val = tensor(0x1.1p-20)]; + tensor var_1410_cast_fp16 = add(x = reduce_mean_63_cast_fp16, y = var_1409_to_fp16)[name = tensor("op_1410_cast_fp16")]; + tensor var_1411_cast_fp16 = sqrt(x = var_1410_cast_fp16)[name = tensor("op_1411_cast_fp16")]; + tensor x_151_cast_fp16 = real_div(x = sub_39_cast_fp16, y = var_1411_cast_fp16)[name = tensor("x_151_cast_fp16")]; + tensor var_1413_cast_fp16 = mul(x = x_151_cast_fp16, y = flow_net_res_blocks_2_in_ln_weight_to_fp16)[name = tensor("op_1413_cast_fp16")]; + tensor x_153_cast_fp16 = add(x = var_1413_cast_fp16, y = flow_net_res_blocks_2_in_ln_bias_to_fp16)[name = tensor("x_153_cast_fp16")]; + tensor var_1415_promoted_to_fp16 = const()[name = tensor("op_1415_promoted_to_fp16"), val = tensor(0x1p+0)]; + tensor var_1416_cast_fp16 = add(x = var_1399_cast_fp16_1, y = var_1415_promoted_to_fp16)[name = tensor("op_1416_cast_fp16")]; + tensor var_1417_cast_fp16 = mul(x = x_153_cast_fp16, y = var_1416_cast_fp16)[name = tensor("op_1417_cast_fp16")]; + tensor input_237_cast_fp16 = add(x = var_1417_cast_fp16, y = var_1399_cast_fp16_0)[name = tensor("input_237_cast_fp16")]; + tensor linear_91_cast_fp16 = linear(bias = flow_net_res_blocks_2_mlp_0_bias_to_fp16, weight = flow_net_res_blocks_2_mlp_0_weight_to_fp16, x = input_237_cast_fp16)[name = tensor("linear_91_cast_fp16")]; + tensor input_241_cast_fp16 = silu(x = linear_91_cast_fp16)[name = tensor("input_241_cast_fp16")]; + tensor linear_92_cast_fp16 = linear(bias = flow_net_res_blocks_2_mlp_2_bias_to_fp16, weight = flow_net_res_blocks_2_mlp_2_weight_to_fp16, x = input_241_cast_fp16)[name = tensor("linear_92_cast_fp16")]; + tensor var_1428_cast_fp16 = mul(x = var_1399_cast_fp16_2, y = linear_92_cast_fp16)[name = tensor("op_1428_cast_fp16")]; + tensor x_155_cast_fp16 = add(x = x_149_cast_fp16, y = var_1428_cast_fp16)[name = tensor("x_155_cast_fp16")]; + tensor linear_93_cast_fp16 = linear(bias = flow_net_res_blocks_3_adaLN_modulation_1_bias_to_fp16, weight = flow_net_res_blocks_3_adaLN_modulation_1_weight_to_fp16, x = input_219_cast_fp16)[name = tensor("linear_93_cast_fp16")]; + tensor var_1438_split_sizes_0 = const()[name = tensor("op_1438_split_sizes_0"), val = tensor([512, 512, 512])]; + tensor var_1438_axis_0 = const()[name = tensor("op_1438_axis_0"), val = tensor(-1)]; + tensor var_1438_cast_fp16_0, tensor var_1438_cast_fp16_1, tensor var_1438_cast_fp16_2 = split(axis = var_1438_axis_0, split_sizes = var_1438_split_sizes_0, x = linear_93_cast_fp16)[name = tensor("op_1438_cast_fp16")]; + tensor mean_49_axes_0 = const()[name = tensor("mean_49_axes_0"), val = tensor([-1])]; + tensor mean_49_keep_dims_0 = const()[name = tensor("mean_49_keep_dims_0"), val = tensor(true)]; + tensor mean_49_cast_fp16 = reduce_mean(axes = mean_49_axes_0, keep_dims = mean_49_keep_dims_0, x = x_155_cast_fp16)[name = tensor("mean_49_cast_fp16")]; + tensor sub_40_cast_fp16 = sub(x = x_155_cast_fp16, y = mean_49_cast_fp16)[name = tensor("sub_40_cast_fp16")]; + tensor square_32_cast_fp16 = square(x = sub_40_cast_fp16)[name = tensor("square_32_cast_fp16")]; + tensor reduce_mean_65_axes_0 = const()[name = tensor("reduce_mean_65_axes_0"), val = tensor([-1])]; + tensor reduce_mean_65_keep_dims_0 = const()[name = tensor("reduce_mean_65_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_65_cast_fp16 = reduce_mean(axes = reduce_mean_65_axes_0, keep_dims = reduce_mean_65_keep_dims_0, x = square_32_cast_fp16)[name = tensor("reduce_mean_65_cast_fp16")]; + tensor var_1448_to_fp16 = const()[name = tensor("op_1448_to_fp16"), val = tensor(0x1.1p-20)]; + tensor var_1449_cast_fp16 = add(x = reduce_mean_65_cast_fp16, y = var_1448_to_fp16)[name = tensor("op_1449_cast_fp16")]; + tensor var_1450_cast_fp16 = sqrt(x = var_1449_cast_fp16)[name = tensor("op_1450_cast_fp16")]; + tensor x_157_cast_fp16 = real_div(x = sub_40_cast_fp16, y = var_1450_cast_fp16)[name = tensor("x_157_cast_fp16")]; + tensor var_1452_cast_fp16 = mul(x = x_157_cast_fp16, y = flow_net_res_blocks_3_in_ln_weight_to_fp16)[name = tensor("op_1452_cast_fp16")]; + tensor x_159_cast_fp16 = add(x = var_1452_cast_fp16, y = flow_net_res_blocks_3_in_ln_bias_to_fp16)[name = tensor("x_159_cast_fp16")]; + tensor var_1454_promoted_to_fp16 = const()[name = tensor("op_1454_promoted_to_fp16"), val = tensor(0x1p+0)]; + tensor var_1455_cast_fp16 = add(x = var_1438_cast_fp16_1, y = var_1454_promoted_to_fp16)[name = tensor("op_1455_cast_fp16")]; + tensor var_1456_cast_fp16 = mul(x = x_159_cast_fp16, y = var_1455_cast_fp16)[name = tensor("op_1456_cast_fp16")]; + tensor input_245_cast_fp16 = add(x = var_1456_cast_fp16, y = var_1438_cast_fp16_0)[name = tensor("input_245_cast_fp16")]; + tensor linear_94_cast_fp16 = linear(bias = flow_net_res_blocks_3_mlp_0_bias_to_fp16, weight = flow_net_res_blocks_3_mlp_0_weight_to_fp16, x = input_245_cast_fp16)[name = tensor("linear_94_cast_fp16")]; + tensor input_249_cast_fp16 = silu(x = linear_94_cast_fp16)[name = tensor("input_249_cast_fp16")]; + tensor linear_95_cast_fp16 = linear(bias = flow_net_res_blocks_3_mlp_2_bias_to_fp16, weight = flow_net_res_blocks_3_mlp_2_weight_to_fp16, x = input_249_cast_fp16)[name = tensor("linear_95_cast_fp16")]; + tensor var_1467_cast_fp16 = mul(x = var_1438_cast_fp16_2, y = linear_95_cast_fp16)[name = tensor("op_1467_cast_fp16")]; + tensor x_161_cast_fp16 = add(x = x_155_cast_fp16, y = var_1467_cast_fp16)[name = tensor("x_161_cast_fp16")]; + tensor linear_96_cast_fp16 = linear(bias = flow_net_res_blocks_4_adaLN_modulation_1_bias_to_fp16, weight = flow_net_res_blocks_4_adaLN_modulation_1_weight_to_fp16, x = input_219_cast_fp16)[name = tensor("linear_96_cast_fp16")]; + tensor var_1477_split_sizes_0 = const()[name = tensor("op_1477_split_sizes_0"), val = tensor([512, 512, 512])]; + tensor var_1477_axis_0 = const()[name = tensor("op_1477_axis_0"), val = tensor(-1)]; + tensor var_1477_cast_fp16_0, tensor var_1477_cast_fp16_1, tensor var_1477_cast_fp16_2 = split(axis = var_1477_axis_0, split_sizes = var_1477_split_sizes_0, x = linear_96_cast_fp16)[name = tensor("op_1477_cast_fp16")]; + tensor mean_51_axes_0 = const()[name = tensor("mean_51_axes_0"), val = tensor([-1])]; + tensor mean_51_keep_dims_0 = const()[name = tensor("mean_51_keep_dims_0"), val = tensor(true)]; + tensor mean_51_cast_fp16 = reduce_mean(axes = mean_51_axes_0, keep_dims = mean_51_keep_dims_0, x = x_161_cast_fp16)[name = tensor("mean_51_cast_fp16")]; + tensor sub_41_cast_fp16 = sub(x = x_161_cast_fp16, y = mean_51_cast_fp16)[name = tensor("sub_41_cast_fp16")]; + tensor square_33_cast_fp16 = square(x = sub_41_cast_fp16)[name = tensor("square_33_cast_fp16")]; + tensor reduce_mean_67_axes_0 = const()[name = tensor("reduce_mean_67_axes_0"), val = tensor([-1])]; + tensor reduce_mean_67_keep_dims_0 = const()[name = tensor("reduce_mean_67_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_67_cast_fp16 = reduce_mean(axes = reduce_mean_67_axes_0, keep_dims = reduce_mean_67_keep_dims_0, x = square_33_cast_fp16)[name = tensor("reduce_mean_67_cast_fp16")]; + tensor var_1487_to_fp16 = const()[name = tensor("op_1487_to_fp16"), val = tensor(0x1.1p-20)]; + tensor var_1488_cast_fp16 = add(x = reduce_mean_67_cast_fp16, y = var_1487_to_fp16)[name = tensor("op_1488_cast_fp16")]; + tensor var_1489_cast_fp16 = sqrt(x = var_1488_cast_fp16)[name = tensor("op_1489_cast_fp16")]; + tensor x_163_cast_fp16 = real_div(x = sub_41_cast_fp16, y = var_1489_cast_fp16)[name = tensor("x_163_cast_fp16")]; + tensor var_1491_cast_fp16 = mul(x = x_163_cast_fp16, y = flow_net_res_blocks_4_in_ln_weight_to_fp16)[name = tensor("op_1491_cast_fp16")]; + tensor x_165_cast_fp16 = add(x = var_1491_cast_fp16, y = flow_net_res_blocks_4_in_ln_bias_to_fp16)[name = tensor("x_165_cast_fp16")]; + tensor var_1493_promoted_to_fp16 = const()[name = tensor("op_1493_promoted_to_fp16"), val = tensor(0x1p+0)]; + tensor var_1494_cast_fp16 = add(x = var_1477_cast_fp16_1, y = var_1493_promoted_to_fp16)[name = tensor("op_1494_cast_fp16")]; + tensor var_1495_cast_fp16 = mul(x = x_165_cast_fp16, y = var_1494_cast_fp16)[name = tensor("op_1495_cast_fp16")]; + tensor input_253_cast_fp16 = add(x = var_1495_cast_fp16, y = var_1477_cast_fp16_0)[name = tensor("input_253_cast_fp16")]; + tensor linear_97_cast_fp16 = linear(bias = flow_net_res_blocks_4_mlp_0_bias_to_fp16, weight = flow_net_res_blocks_4_mlp_0_weight_to_fp16, x = input_253_cast_fp16)[name = tensor("linear_97_cast_fp16")]; + tensor input_257_cast_fp16 = silu(x = linear_97_cast_fp16)[name = tensor("input_257_cast_fp16")]; + tensor linear_98_cast_fp16 = linear(bias = flow_net_res_blocks_4_mlp_2_bias_to_fp16, weight = flow_net_res_blocks_4_mlp_2_weight_to_fp16, x = input_257_cast_fp16)[name = tensor("linear_98_cast_fp16")]; + tensor var_1506_cast_fp16 = mul(x = var_1477_cast_fp16_2, y = linear_98_cast_fp16)[name = tensor("op_1506_cast_fp16")]; + tensor x_167_cast_fp16 = add(x = x_161_cast_fp16, y = var_1506_cast_fp16)[name = tensor("x_167_cast_fp16")]; + tensor linear_99_cast_fp16 = linear(bias = flow_net_res_blocks_5_adaLN_modulation_1_bias_to_fp16, weight = flow_net_res_blocks_5_adaLN_modulation_1_weight_to_fp16, x = input_219_cast_fp16)[name = tensor("linear_99_cast_fp16")]; + tensor var_1516_split_sizes_0 = const()[name = tensor("op_1516_split_sizes_0"), val = tensor([512, 512, 512])]; + tensor var_1516_axis_0 = const()[name = tensor("op_1516_axis_0"), val = tensor(-1)]; + tensor var_1516_cast_fp16_0, tensor var_1516_cast_fp16_1, tensor var_1516_cast_fp16_2 = split(axis = var_1516_axis_0, split_sizes = var_1516_split_sizes_0, x = linear_99_cast_fp16)[name = tensor("op_1516_cast_fp16")]; + tensor mean_53_axes_0 = const()[name = tensor("mean_53_axes_0"), val = tensor([-1])]; + tensor mean_53_keep_dims_0 = const()[name = tensor("mean_53_keep_dims_0"), val = tensor(true)]; + tensor mean_53_cast_fp16 = reduce_mean(axes = mean_53_axes_0, keep_dims = mean_53_keep_dims_0, x = x_167_cast_fp16)[name = tensor("mean_53_cast_fp16")]; + tensor sub_42_cast_fp16 = sub(x = x_167_cast_fp16, y = mean_53_cast_fp16)[name = tensor("sub_42_cast_fp16")]; + tensor square_34_cast_fp16 = square(x = sub_42_cast_fp16)[name = tensor("square_34_cast_fp16")]; + tensor reduce_mean_69_axes_0 = const()[name = tensor("reduce_mean_69_axes_0"), val = tensor([-1])]; + tensor reduce_mean_69_keep_dims_0 = const()[name = tensor("reduce_mean_69_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_69_cast_fp16 = reduce_mean(axes = reduce_mean_69_axes_0, keep_dims = reduce_mean_69_keep_dims_0, x = square_34_cast_fp16)[name = tensor("reduce_mean_69_cast_fp16")]; + tensor var_1526_to_fp16 = const()[name = tensor("op_1526_to_fp16"), val = tensor(0x1.1p-20)]; + tensor var_1527_cast_fp16 = add(x = reduce_mean_69_cast_fp16, y = var_1526_to_fp16)[name = tensor("op_1527_cast_fp16")]; + tensor var_1528_cast_fp16 = sqrt(x = var_1527_cast_fp16)[name = tensor("op_1528_cast_fp16")]; + tensor x_169_cast_fp16 = real_div(x = sub_42_cast_fp16, y = var_1528_cast_fp16)[name = tensor("x_169_cast_fp16")]; + tensor var_1530_cast_fp16 = mul(x = x_169_cast_fp16, y = flow_net_res_blocks_5_in_ln_weight_to_fp16)[name = tensor("op_1530_cast_fp16")]; + tensor x_171_cast_fp16 = add(x = var_1530_cast_fp16, y = flow_net_res_blocks_5_in_ln_bias_to_fp16)[name = tensor("x_171_cast_fp16")]; + tensor var_1532_promoted_to_fp16 = const()[name = tensor("op_1532_promoted_to_fp16"), val = tensor(0x1p+0)]; + tensor var_1533_cast_fp16 = add(x = var_1516_cast_fp16_1, y = var_1532_promoted_to_fp16)[name = tensor("op_1533_cast_fp16")]; + tensor var_1534_cast_fp16 = mul(x = x_171_cast_fp16, y = var_1533_cast_fp16)[name = tensor("op_1534_cast_fp16")]; + tensor input_261_cast_fp16 = add(x = var_1534_cast_fp16, y = var_1516_cast_fp16_0)[name = tensor("input_261_cast_fp16")]; + tensor linear_100_cast_fp16 = linear(bias = flow_net_res_blocks_5_mlp_0_bias_to_fp16, weight = flow_net_res_blocks_5_mlp_0_weight_to_fp16, x = input_261_cast_fp16)[name = tensor("linear_100_cast_fp16")]; + tensor input_265_cast_fp16 = silu(x = linear_100_cast_fp16)[name = tensor("input_265_cast_fp16")]; + tensor linear_101_cast_fp16 = linear(bias = flow_net_res_blocks_5_mlp_2_bias_to_fp16, weight = flow_net_res_blocks_5_mlp_2_weight_to_fp16, x = input_265_cast_fp16)[name = tensor("linear_101_cast_fp16")]; + tensor var_1545_cast_fp16 = mul(x = var_1516_cast_fp16_2, y = linear_101_cast_fp16)[name = tensor("op_1545_cast_fp16")]; + tensor x_173_cast_fp16 = add(x = x_167_cast_fp16, y = var_1545_cast_fp16)[name = tensor("x_173_cast_fp16")]; + tensor linear_102_cast_fp16 = linear(bias = flow_net_final_layer_adaLN_modulation_1_bias_to_fp16, weight = flow_net_final_layer_adaLN_modulation_1_weight_to_fp16, x = input_219_cast_fp16)[name = tensor("linear_102_cast_fp16")]; + tensor var_1554_split_sizes_0 = const()[name = tensor("op_1554_split_sizes_0"), val = tensor([512, 512])]; + tensor var_1554_axis_0 = const()[name = tensor("op_1554_axis_0"), val = tensor(-1)]; + tensor var_1554_cast_fp16_0, tensor var_1554_cast_fp16_1 = split(axis = var_1554_axis_0, split_sizes = var_1554_split_sizes_0, x = linear_102_cast_fp16)[name = tensor("op_1554_cast_fp16")]; + tensor mean_55_axes_0 = const()[name = tensor("mean_55_axes_0"), val = tensor([-1])]; + tensor mean_55_keep_dims_0 = const()[name = tensor("mean_55_keep_dims_0"), val = tensor(true)]; + tensor mean_55_cast_fp16 = reduce_mean(axes = mean_55_axes_0, keep_dims = mean_55_keep_dims_0, x = x_173_cast_fp16)[name = tensor("mean_55_cast_fp16")]; + tensor sub_43_cast_fp16 = sub(x = x_173_cast_fp16, y = mean_55_cast_fp16)[name = tensor("sub_43_cast_fp16")]; + tensor square_35_cast_fp16 = square(x = sub_43_cast_fp16)[name = tensor("square_35_cast_fp16")]; + tensor reduce_mean_71_axes_0 = const()[name = tensor("reduce_mean_71_axes_0"), val = tensor([-1])]; + tensor reduce_mean_71_keep_dims_0 = const()[name = tensor("reduce_mean_71_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_71_cast_fp16 = reduce_mean(axes = reduce_mean_71_axes_0, keep_dims = reduce_mean_71_keep_dims_0, x = square_35_cast_fp16)[name = tensor("reduce_mean_71_cast_fp16")]; + tensor var_1561_to_fp16 = const()[name = tensor("op_1561_to_fp16"), val = tensor(0x1.1p-20)]; + tensor var_1562_cast_fp16 = add(x = reduce_mean_71_cast_fp16, y = var_1561_to_fp16)[name = tensor("op_1562_cast_fp16")]; + tensor var_1563_cast_fp16 = sqrt(x = var_1562_cast_fp16)[name = tensor("op_1563_cast_fp16")]; + tensor x_175_cast_fp16 = real_div(x = sub_43_cast_fp16, y = var_1563_cast_fp16)[name = tensor("x_175_cast_fp16")]; + tensor var_1565_promoted_to_fp16 = const()[name = tensor("op_1565_promoted_to_fp16"), val = tensor(0x1p+0)]; + tensor var_1566_cast_fp16 = add(x = var_1554_cast_fp16_1, y = var_1565_promoted_to_fp16)[name = tensor("op_1566_cast_fp16")]; + tensor var_1567_cast_fp16 = mul(x = x_175_cast_fp16, y = var_1566_cast_fp16)[name = tensor("op_1567_cast_fp16")]; + tensor input_269_cast_fp16 = add(x = var_1567_cast_fp16, y = var_1554_cast_fp16_0)[name = tensor("input_269_cast_fp16")]; + tensor linear_103_cast_fp16 = linear(bias = flow_net_final_layer_linear_bias_to_fp16, weight = flow_net_final_layer_linear_weight_to_fp16, x = input_269_cast_fp16)[name = tensor("linear_103_cast_fp16")]; + tensor var_1578_to_fp16 = const()[name = tensor("op_1578_to_fp16"), val = tensor(0x1p-3)]; + tensor var_1579_cast_fp16 = mul(x = linear_103_cast_fp16, y = var_1578_to_fp16)[name = tensor("op_1579_cast_fp16")]; + tensor input_271_cast_fp16 = add(x = input_203_cast_fp16, y = var_1579_cast_fp16)[name = tensor("input_271_cast_fp16")]; + tensor linear_104_cast_fp16 = linear(bias = flow_net_input_proj_bias_to_fp16, weight = flow_net_input_proj_weight_to_fp16, x = input_271_cast_fp16)[name = tensor("linear_104_cast_fp16")]; + tensor input_275_to_fp16 = const()[name = tensor("input_275_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(19003904)))]; + tensor input_277_cast_fp16 = silu(x = input_275_to_fp16)[name = tensor("input_277_cast_fp16")]; + tensor linear_106_cast_fp16 = linear(bias = flow_net_time_embed_0_mlp_2_bias_to_fp16, weight = flow_net_time_embed_0_mlp_2_weight_to_fp16, x = input_277_cast_fp16)[name = tensor("linear_106_cast_fp16")]; + tensor reduce_mean_72_axes_0 = const()[name = tensor("reduce_mean_72_axes_0"), val = tensor([-1])]; + tensor reduce_mean_72_keep_dims_0 = const()[name = tensor("reduce_mean_72_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_72_cast_fp16 = reduce_mean(axes = reduce_mean_72_axes_0, keep_dims = reduce_mean_72_keep_dims_0, x = linear_106_cast_fp16)[name = tensor("reduce_mean_72_cast_fp16")]; + tensor sub_44_cast_fp16 = sub(x = linear_106_cast_fp16, y = reduce_mean_72_cast_fp16)[name = tensor("sub_44_cast_fp16")]; + tensor square_36_cast_fp16 = square(x = sub_44_cast_fp16)[name = tensor("square_36_cast_fp16")]; + tensor reduce_mean_73_axes_0 = const()[name = tensor("reduce_mean_73_axes_0"), val = tensor([-1])]; + tensor reduce_mean_73_keep_dims_0 = const()[name = tensor("reduce_mean_73_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_73_cast_fp16 = reduce_mean(axes = reduce_mean_73_axes_0, keep_dims = reduce_mean_73_keep_dims_0, x = square_36_cast_fp16)[name = tensor("reduce_mean_73_cast_fp16")]; + tensor real_div_8_to_fp16 = const()[name = tensor("real_div_8_to_fp16"), val = tensor(0x1.008p+0)]; + tensor mul_8_cast_fp16 = mul(x = reduce_mean_73_cast_fp16, y = real_div_8_to_fp16)[name = tensor("mul_8_cast_fp16")]; + tensor var_1645_to_fp16 = const()[name = tensor("op_1645_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_73_cast_fp16_0 = add(x = mul_8_cast_fp16, y = var_1645_to_fp16)[name = tensor("var_73_cast_fp16")]; + tensor var_1648_epsilon_0 = const()[name = tensor("op_1648_epsilon_0"), val = tensor(0x1.197998p-40)]; + tensor var_1648_cast_fp16 = rsqrt(epsilon = var_1648_epsilon_0, x = var_73_cast_fp16_0)[name = tensor("op_1648_cast_fp16")]; + tensor var_1649_cast_fp16 = mul(x = const_3_to_fp16, y = var_1648_cast_fp16)[name = tensor("op_1649_cast_fp16")]; + tensor var_1650_cast_fp16 = mul(x = linear_106_cast_fp16, y = var_1649_cast_fp16)[name = tensor("op_1650_cast_fp16")]; + tensor input_281_to_fp16 = const()[name = tensor("input_281_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(19004992)))]; + tensor input_283_cast_fp16 = silu(x = input_281_to_fp16)[name = tensor("input_283_cast_fp16")]; + tensor linear_108_cast_fp16 = linear(bias = flow_net_time_embed_1_mlp_2_bias_to_fp16, weight = flow_net_time_embed_1_mlp_2_weight_to_fp16, x = input_283_cast_fp16)[name = tensor("linear_108_cast_fp16")]; + tensor reduce_mean_74_axes_0 = const()[name = tensor("reduce_mean_74_axes_0"), val = tensor([-1])]; + tensor reduce_mean_74_keep_dims_0 = const()[name = tensor("reduce_mean_74_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_74_cast_fp16 = reduce_mean(axes = reduce_mean_74_axes_0, keep_dims = reduce_mean_74_keep_dims_0, x = linear_108_cast_fp16)[name = tensor("reduce_mean_74_cast_fp16")]; + tensor sub_46_cast_fp16 = sub(x = linear_108_cast_fp16, y = reduce_mean_74_cast_fp16)[name = tensor("sub_46_cast_fp16")]; + tensor square_37_cast_fp16 = square(x = sub_46_cast_fp16)[name = tensor("square_37_cast_fp16")]; + tensor reduce_mean_75_axes_0 = const()[name = tensor("reduce_mean_75_axes_0"), val = tensor([-1])]; + tensor reduce_mean_75_keep_dims_0 = const()[name = tensor("reduce_mean_75_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_75_cast_fp16 = reduce_mean(axes = reduce_mean_75_axes_0, keep_dims = reduce_mean_75_keep_dims_0, x = square_37_cast_fp16)[name = tensor("reduce_mean_75_cast_fp16")]; + tensor real_div_9_to_fp16 = const()[name = tensor("real_div_9_to_fp16"), val = tensor(0x1.008p+0)]; + tensor mul_9_cast_fp16 = mul(x = reduce_mean_75_cast_fp16, y = real_div_9_to_fp16)[name = tensor("mul_9_cast_fp16")]; + tensor var_1682_to_fp16 = const()[name = tensor("op_1682_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_75_cast_fp16_0 = add(x = mul_9_cast_fp16, y = var_1682_to_fp16)[name = tensor("var_75_cast_fp16")]; + tensor var_1685_epsilon_0 = const()[name = tensor("op_1685_epsilon_0"), val = tensor(0x1.197998p-40)]; + tensor var_1685_cast_fp16 = rsqrt(epsilon = var_1685_epsilon_0, x = var_75_cast_fp16_0)[name = tensor("op_1685_cast_fp16")]; + tensor var_1686_cast_fp16 = mul(x = const_5_to_fp16, y = var_1685_cast_fp16)[name = tensor("op_1686_cast_fp16")]; + tensor var_1687_cast_fp16 = mul(x = linear_108_cast_fp16, y = var_1686_cast_fp16)[name = tensor("op_1687_cast_fp16")]; + tensor var_1699_cast_fp16 = add(x = var_1650_cast_fp16, y = var_1687_cast_fp16)[name = tensor("op_1699_cast_fp16")]; + tensor _inversed_t_combined_9_y_0_to_fp16 = const()[name = tensor("_inversed_t_combined_9_y_0_to_fp16"), val = tensor(0x1p-1)]; + tensor _inversed_t_combined_9_cast_fp16 = mul(x = var_1699_cast_fp16, y = _inversed_t_combined_9_y_0_to_fp16)[name = tensor("_inversed_t_combined_9_cast_fp16")]; + tensor input_285_cast_fp16 = add(x = _inversed_t_combined_9_cast_fp16, y = linear_5_cast_fp16)[name = tensor("input_285_cast_fp16")]; + tensor input_287_cast_fp16 = silu(x = input_285_cast_fp16)[name = tensor("input_287_cast_fp16")]; + tensor linear_110_cast_fp16 = linear(bias = flow_net_res_blocks_0_adaLN_modulation_1_bias_to_fp16, weight = flow_net_res_blocks_0_adaLN_modulation_1_weight_to_fp16, x = input_287_cast_fp16)[name = tensor("linear_110_cast_fp16")]; + tensor var_1714_split_sizes_0 = const()[name = tensor("op_1714_split_sizes_0"), val = tensor([512, 512, 512])]; + tensor var_1714_axis_0 = const()[name = tensor("op_1714_axis_0"), val = tensor(-1)]; + tensor var_1714_cast_fp16_0, tensor var_1714_cast_fp16_1, tensor var_1714_cast_fp16_2 = split(axis = var_1714_axis_0, split_sizes = var_1714_split_sizes_0, x = linear_110_cast_fp16)[name = tensor("op_1714_cast_fp16")]; + tensor mean_57_axes_0 = const()[name = tensor("mean_57_axes_0"), val = tensor([-1])]; + tensor mean_57_keep_dims_0 = const()[name = tensor("mean_57_keep_dims_0"), val = tensor(true)]; + tensor mean_57_cast_fp16 = reduce_mean(axes = mean_57_axes_0, keep_dims = mean_57_keep_dims_0, x = linear_104_cast_fp16)[name = tensor("mean_57_cast_fp16")]; + tensor sub_48_cast_fp16 = sub(x = linear_104_cast_fp16, y = mean_57_cast_fp16)[name = tensor("sub_48_cast_fp16")]; + tensor square_38_cast_fp16 = square(x = sub_48_cast_fp16)[name = tensor("square_38_cast_fp16")]; + tensor reduce_mean_77_axes_0 = const()[name = tensor("reduce_mean_77_axes_0"), val = tensor([-1])]; + tensor reduce_mean_77_keep_dims_0 = const()[name = tensor("reduce_mean_77_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_77_cast_fp16 = reduce_mean(axes = reduce_mean_77_axes_0, keep_dims = reduce_mean_77_keep_dims_0, x = square_38_cast_fp16)[name = tensor("reduce_mean_77_cast_fp16")]; + tensor var_1724_to_fp16 = const()[name = tensor("op_1724_to_fp16"), val = tensor(0x1.1p-20)]; + tensor var_1725_cast_fp16 = add(x = reduce_mean_77_cast_fp16, y = var_1724_to_fp16)[name = tensor("op_1725_cast_fp16")]; + tensor var_1726_cast_fp16 = sqrt(x = var_1725_cast_fp16)[name = tensor("op_1726_cast_fp16")]; + tensor x_183_cast_fp16 = real_div(x = sub_48_cast_fp16, y = var_1726_cast_fp16)[name = tensor("x_183_cast_fp16")]; + tensor var_1728_cast_fp16 = mul(x = x_183_cast_fp16, y = flow_net_res_blocks_0_in_ln_weight_to_fp16)[name = tensor("op_1728_cast_fp16")]; + tensor x_185_cast_fp16 = add(x = var_1728_cast_fp16, y = flow_net_res_blocks_0_in_ln_bias_to_fp16)[name = tensor("x_185_cast_fp16")]; + tensor var_1730_promoted_to_fp16 = const()[name = tensor("op_1730_promoted_to_fp16"), val = tensor(0x1p+0)]; + tensor var_1731_cast_fp16 = add(x = var_1714_cast_fp16_1, y = var_1730_promoted_to_fp16)[name = tensor("op_1731_cast_fp16")]; + tensor var_1732_cast_fp16 = mul(x = x_185_cast_fp16, y = var_1731_cast_fp16)[name = tensor("op_1732_cast_fp16")]; + tensor input_289_cast_fp16 = add(x = var_1732_cast_fp16, y = var_1714_cast_fp16_0)[name = tensor("input_289_cast_fp16")]; + tensor linear_111_cast_fp16 = linear(bias = flow_net_res_blocks_0_mlp_0_bias_to_fp16, weight = flow_net_res_blocks_0_mlp_0_weight_to_fp16, x = input_289_cast_fp16)[name = tensor("linear_111_cast_fp16")]; + tensor input_293_cast_fp16 = silu(x = linear_111_cast_fp16)[name = tensor("input_293_cast_fp16")]; + tensor linear_112_cast_fp16 = linear(bias = flow_net_res_blocks_0_mlp_2_bias_to_fp16, weight = flow_net_res_blocks_0_mlp_2_weight_to_fp16, x = input_293_cast_fp16)[name = tensor("linear_112_cast_fp16")]; + tensor var_1743_cast_fp16 = mul(x = var_1714_cast_fp16_2, y = linear_112_cast_fp16)[name = tensor("op_1743_cast_fp16")]; + tensor x_187_cast_fp16 = add(x = linear_104_cast_fp16, y = var_1743_cast_fp16)[name = tensor("x_187_cast_fp16")]; + tensor linear_113_cast_fp16 = linear(bias = flow_net_res_blocks_1_adaLN_modulation_1_bias_to_fp16, weight = flow_net_res_blocks_1_adaLN_modulation_1_weight_to_fp16, x = input_287_cast_fp16)[name = tensor("linear_113_cast_fp16")]; + tensor var_1753_split_sizes_0 = const()[name = tensor("op_1753_split_sizes_0"), val = tensor([512, 512, 512])]; + tensor var_1753_axis_0 = const()[name = tensor("op_1753_axis_0"), val = tensor(-1)]; + tensor var_1753_cast_fp16_0, tensor var_1753_cast_fp16_1, tensor var_1753_cast_fp16_2 = split(axis = var_1753_axis_0, split_sizes = var_1753_split_sizes_0, x = linear_113_cast_fp16)[name = tensor("op_1753_cast_fp16")]; + tensor mean_59_axes_0 = const()[name = tensor("mean_59_axes_0"), val = tensor([-1])]; + tensor mean_59_keep_dims_0 = const()[name = tensor("mean_59_keep_dims_0"), val = tensor(true)]; + tensor mean_59_cast_fp16 = reduce_mean(axes = mean_59_axes_0, keep_dims = mean_59_keep_dims_0, x = x_187_cast_fp16)[name = tensor("mean_59_cast_fp16")]; + tensor sub_49_cast_fp16 = sub(x = x_187_cast_fp16, y = mean_59_cast_fp16)[name = tensor("sub_49_cast_fp16")]; + tensor square_39_cast_fp16 = square(x = sub_49_cast_fp16)[name = tensor("square_39_cast_fp16")]; + tensor reduce_mean_79_axes_0 = const()[name = tensor("reduce_mean_79_axes_0"), val = tensor([-1])]; + tensor reduce_mean_79_keep_dims_0 = const()[name = tensor("reduce_mean_79_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_79_cast_fp16 = reduce_mean(axes = reduce_mean_79_axes_0, keep_dims = reduce_mean_79_keep_dims_0, x = square_39_cast_fp16)[name = tensor("reduce_mean_79_cast_fp16")]; + tensor var_1763_to_fp16 = const()[name = tensor("op_1763_to_fp16"), val = tensor(0x1.1p-20)]; + tensor var_1764_cast_fp16 = add(x = reduce_mean_79_cast_fp16, y = var_1763_to_fp16)[name = tensor("op_1764_cast_fp16")]; + tensor var_1765_cast_fp16 = sqrt(x = var_1764_cast_fp16)[name = tensor("op_1765_cast_fp16")]; + tensor x_189_cast_fp16 = real_div(x = sub_49_cast_fp16, y = var_1765_cast_fp16)[name = tensor("x_189_cast_fp16")]; + tensor var_1767_cast_fp16 = mul(x = x_189_cast_fp16, y = flow_net_res_blocks_1_in_ln_weight_to_fp16)[name = tensor("op_1767_cast_fp16")]; + tensor x_191_cast_fp16 = add(x = var_1767_cast_fp16, y = flow_net_res_blocks_1_in_ln_bias_to_fp16)[name = tensor("x_191_cast_fp16")]; + tensor var_1769_promoted_to_fp16 = const()[name = tensor("op_1769_promoted_to_fp16"), val = tensor(0x1p+0)]; + tensor var_1770_cast_fp16 = add(x = var_1753_cast_fp16_1, y = var_1769_promoted_to_fp16)[name = tensor("op_1770_cast_fp16")]; + tensor var_1771_cast_fp16 = mul(x = x_191_cast_fp16, y = var_1770_cast_fp16)[name = tensor("op_1771_cast_fp16")]; + tensor input_297_cast_fp16 = add(x = var_1771_cast_fp16, y = var_1753_cast_fp16_0)[name = tensor("input_297_cast_fp16")]; + tensor linear_114_cast_fp16 = linear(bias = flow_net_res_blocks_1_mlp_0_bias_to_fp16, weight = flow_net_res_blocks_1_mlp_0_weight_to_fp16, x = input_297_cast_fp16)[name = tensor("linear_114_cast_fp16")]; + tensor input_301_cast_fp16 = silu(x = linear_114_cast_fp16)[name = tensor("input_301_cast_fp16")]; + tensor linear_115_cast_fp16 = linear(bias = flow_net_res_blocks_1_mlp_2_bias_to_fp16, weight = flow_net_res_blocks_1_mlp_2_weight_to_fp16, x = input_301_cast_fp16)[name = tensor("linear_115_cast_fp16")]; + tensor var_1782_cast_fp16 = mul(x = var_1753_cast_fp16_2, y = linear_115_cast_fp16)[name = tensor("op_1782_cast_fp16")]; + tensor x_193_cast_fp16 = add(x = x_187_cast_fp16, y = var_1782_cast_fp16)[name = tensor("x_193_cast_fp16")]; + tensor linear_116_cast_fp16 = linear(bias = flow_net_res_blocks_2_adaLN_modulation_1_bias_to_fp16, weight = flow_net_res_blocks_2_adaLN_modulation_1_weight_to_fp16, x = input_287_cast_fp16)[name = tensor("linear_116_cast_fp16")]; + tensor var_1792_split_sizes_0 = const()[name = tensor("op_1792_split_sizes_0"), val = tensor([512, 512, 512])]; + tensor var_1792_axis_0 = const()[name = tensor("op_1792_axis_0"), val = tensor(-1)]; + tensor var_1792_cast_fp16_0, tensor var_1792_cast_fp16_1, tensor var_1792_cast_fp16_2 = split(axis = var_1792_axis_0, split_sizes = var_1792_split_sizes_0, x = linear_116_cast_fp16)[name = tensor("op_1792_cast_fp16")]; + tensor mean_61_axes_0 = const()[name = tensor("mean_61_axes_0"), val = tensor([-1])]; + tensor mean_61_keep_dims_0 = const()[name = tensor("mean_61_keep_dims_0"), val = tensor(true)]; + tensor mean_61_cast_fp16 = reduce_mean(axes = mean_61_axes_0, keep_dims = mean_61_keep_dims_0, x = x_193_cast_fp16)[name = tensor("mean_61_cast_fp16")]; + tensor sub_50_cast_fp16 = sub(x = x_193_cast_fp16, y = mean_61_cast_fp16)[name = tensor("sub_50_cast_fp16")]; + tensor square_40_cast_fp16 = square(x = sub_50_cast_fp16)[name = tensor("square_40_cast_fp16")]; + tensor reduce_mean_81_axes_0 = const()[name = tensor("reduce_mean_81_axes_0"), val = tensor([-1])]; + tensor reduce_mean_81_keep_dims_0 = const()[name = tensor("reduce_mean_81_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_81_cast_fp16 = reduce_mean(axes = reduce_mean_81_axes_0, keep_dims = reduce_mean_81_keep_dims_0, x = square_40_cast_fp16)[name = tensor("reduce_mean_81_cast_fp16")]; + tensor var_1802_to_fp16 = const()[name = tensor("op_1802_to_fp16"), val = tensor(0x1.1p-20)]; + tensor var_1803_cast_fp16 = add(x = reduce_mean_81_cast_fp16, y = var_1802_to_fp16)[name = tensor("op_1803_cast_fp16")]; + tensor var_1804_cast_fp16 = sqrt(x = var_1803_cast_fp16)[name = tensor("op_1804_cast_fp16")]; + tensor x_195_cast_fp16 = real_div(x = sub_50_cast_fp16, y = var_1804_cast_fp16)[name = tensor("x_195_cast_fp16")]; + tensor var_1806_cast_fp16 = mul(x = x_195_cast_fp16, y = flow_net_res_blocks_2_in_ln_weight_to_fp16)[name = tensor("op_1806_cast_fp16")]; + tensor x_197_cast_fp16 = add(x = var_1806_cast_fp16, y = flow_net_res_blocks_2_in_ln_bias_to_fp16)[name = tensor("x_197_cast_fp16")]; + tensor var_1808_promoted_to_fp16 = const()[name = tensor("op_1808_promoted_to_fp16"), val = tensor(0x1p+0)]; + tensor var_1809_cast_fp16 = add(x = var_1792_cast_fp16_1, y = var_1808_promoted_to_fp16)[name = tensor("op_1809_cast_fp16")]; + tensor var_1810_cast_fp16 = mul(x = x_197_cast_fp16, y = var_1809_cast_fp16)[name = tensor("op_1810_cast_fp16")]; + tensor input_305_cast_fp16 = add(x = var_1810_cast_fp16, y = var_1792_cast_fp16_0)[name = tensor("input_305_cast_fp16")]; + tensor linear_117_cast_fp16 = linear(bias = flow_net_res_blocks_2_mlp_0_bias_to_fp16, weight = flow_net_res_blocks_2_mlp_0_weight_to_fp16, x = input_305_cast_fp16)[name = tensor("linear_117_cast_fp16")]; + tensor input_309_cast_fp16 = silu(x = linear_117_cast_fp16)[name = tensor("input_309_cast_fp16")]; + tensor linear_118_cast_fp16 = linear(bias = flow_net_res_blocks_2_mlp_2_bias_to_fp16, weight = flow_net_res_blocks_2_mlp_2_weight_to_fp16, x = input_309_cast_fp16)[name = tensor("linear_118_cast_fp16")]; + tensor var_1821_cast_fp16 = mul(x = var_1792_cast_fp16_2, y = linear_118_cast_fp16)[name = tensor("op_1821_cast_fp16")]; + tensor x_199_cast_fp16 = add(x = x_193_cast_fp16, y = var_1821_cast_fp16)[name = tensor("x_199_cast_fp16")]; + tensor linear_119_cast_fp16 = linear(bias = flow_net_res_blocks_3_adaLN_modulation_1_bias_to_fp16, weight = flow_net_res_blocks_3_adaLN_modulation_1_weight_to_fp16, x = input_287_cast_fp16)[name = tensor("linear_119_cast_fp16")]; + tensor var_1831_split_sizes_0 = const()[name = tensor("op_1831_split_sizes_0"), val = tensor([512, 512, 512])]; + tensor var_1831_axis_0 = const()[name = tensor("op_1831_axis_0"), val = tensor(-1)]; + tensor var_1831_cast_fp16_0, tensor var_1831_cast_fp16_1, tensor var_1831_cast_fp16_2 = split(axis = var_1831_axis_0, split_sizes = var_1831_split_sizes_0, x = linear_119_cast_fp16)[name = tensor("op_1831_cast_fp16")]; + tensor mean_63_axes_0 = const()[name = tensor("mean_63_axes_0"), val = tensor([-1])]; + tensor mean_63_keep_dims_0 = const()[name = tensor("mean_63_keep_dims_0"), val = tensor(true)]; + tensor mean_63_cast_fp16 = reduce_mean(axes = mean_63_axes_0, keep_dims = mean_63_keep_dims_0, x = x_199_cast_fp16)[name = tensor("mean_63_cast_fp16")]; + tensor sub_51_cast_fp16 = sub(x = x_199_cast_fp16, y = mean_63_cast_fp16)[name = tensor("sub_51_cast_fp16")]; + tensor square_41_cast_fp16 = square(x = sub_51_cast_fp16)[name = tensor("square_41_cast_fp16")]; + tensor reduce_mean_83_axes_0 = const()[name = tensor("reduce_mean_83_axes_0"), val = tensor([-1])]; + tensor reduce_mean_83_keep_dims_0 = const()[name = tensor("reduce_mean_83_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_83_cast_fp16 = reduce_mean(axes = reduce_mean_83_axes_0, keep_dims = reduce_mean_83_keep_dims_0, x = square_41_cast_fp16)[name = tensor("reduce_mean_83_cast_fp16")]; + tensor var_1841_to_fp16 = const()[name = tensor("op_1841_to_fp16"), val = tensor(0x1.1p-20)]; + tensor var_1842_cast_fp16 = add(x = reduce_mean_83_cast_fp16, y = var_1841_to_fp16)[name = tensor("op_1842_cast_fp16")]; + tensor var_1843_cast_fp16 = sqrt(x = var_1842_cast_fp16)[name = tensor("op_1843_cast_fp16")]; + tensor x_201_cast_fp16 = real_div(x = sub_51_cast_fp16, y = var_1843_cast_fp16)[name = tensor("x_201_cast_fp16")]; + tensor var_1845_cast_fp16 = mul(x = x_201_cast_fp16, y = flow_net_res_blocks_3_in_ln_weight_to_fp16)[name = tensor("op_1845_cast_fp16")]; + tensor x_203_cast_fp16 = add(x = var_1845_cast_fp16, y = flow_net_res_blocks_3_in_ln_bias_to_fp16)[name = tensor("x_203_cast_fp16")]; + tensor var_1847_promoted_to_fp16 = const()[name = tensor("op_1847_promoted_to_fp16"), val = tensor(0x1p+0)]; + tensor var_1848_cast_fp16 = add(x = var_1831_cast_fp16_1, y = var_1847_promoted_to_fp16)[name = tensor("op_1848_cast_fp16")]; + tensor var_1849_cast_fp16 = mul(x = x_203_cast_fp16, y = var_1848_cast_fp16)[name = tensor("op_1849_cast_fp16")]; + tensor input_313_cast_fp16 = add(x = var_1849_cast_fp16, y = var_1831_cast_fp16_0)[name = tensor("input_313_cast_fp16")]; + tensor linear_120_cast_fp16 = linear(bias = flow_net_res_blocks_3_mlp_0_bias_to_fp16, weight = flow_net_res_blocks_3_mlp_0_weight_to_fp16, x = input_313_cast_fp16)[name = tensor("linear_120_cast_fp16")]; + tensor input_317_cast_fp16 = silu(x = linear_120_cast_fp16)[name = tensor("input_317_cast_fp16")]; + tensor linear_121_cast_fp16 = linear(bias = flow_net_res_blocks_3_mlp_2_bias_to_fp16, weight = flow_net_res_blocks_3_mlp_2_weight_to_fp16, x = input_317_cast_fp16)[name = tensor("linear_121_cast_fp16")]; + tensor var_1860_cast_fp16 = mul(x = var_1831_cast_fp16_2, y = linear_121_cast_fp16)[name = tensor("op_1860_cast_fp16")]; + tensor x_205_cast_fp16 = add(x = x_199_cast_fp16, y = var_1860_cast_fp16)[name = tensor("x_205_cast_fp16")]; + tensor linear_122_cast_fp16 = linear(bias = flow_net_res_blocks_4_adaLN_modulation_1_bias_to_fp16, weight = flow_net_res_blocks_4_adaLN_modulation_1_weight_to_fp16, x = input_287_cast_fp16)[name = tensor("linear_122_cast_fp16")]; + tensor var_1870_split_sizes_0 = const()[name = tensor("op_1870_split_sizes_0"), val = tensor([512, 512, 512])]; + tensor var_1870_axis_0 = const()[name = tensor("op_1870_axis_0"), val = tensor(-1)]; + tensor var_1870_cast_fp16_0, tensor var_1870_cast_fp16_1, tensor var_1870_cast_fp16_2 = split(axis = var_1870_axis_0, split_sizes = var_1870_split_sizes_0, x = linear_122_cast_fp16)[name = tensor("op_1870_cast_fp16")]; + tensor mean_65_axes_0 = const()[name = tensor("mean_65_axes_0"), val = tensor([-1])]; + tensor mean_65_keep_dims_0 = const()[name = tensor("mean_65_keep_dims_0"), val = tensor(true)]; + tensor mean_65_cast_fp16 = reduce_mean(axes = mean_65_axes_0, keep_dims = mean_65_keep_dims_0, x = x_205_cast_fp16)[name = tensor("mean_65_cast_fp16")]; + tensor sub_52_cast_fp16 = sub(x = x_205_cast_fp16, y = mean_65_cast_fp16)[name = tensor("sub_52_cast_fp16")]; + tensor square_42_cast_fp16 = square(x = sub_52_cast_fp16)[name = tensor("square_42_cast_fp16")]; + tensor reduce_mean_85_axes_0 = const()[name = tensor("reduce_mean_85_axes_0"), val = tensor([-1])]; + tensor reduce_mean_85_keep_dims_0 = const()[name = tensor("reduce_mean_85_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_85_cast_fp16 = reduce_mean(axes = reduce_mean_85_axes_0, keep_dims = reduce_mean_85_keep_dims_0, x = square_42_cast_fp16)[name = tensor("reduce_mean_85_cast_fp16")]; + tensor var_1880_to_fp16 = const()[name = tensor("op_1880_to_fp16"), val = tensor(0x1.1p-20)]; + tensor var_1881_cast_fp16 = add(x = reduce_mean_85_cast_fp16, y = var_1880_to_fp16)[name = tensor("op_1881_cast_fp16")]; + tensor var_1882_cast_fp16 = sqrt(x = var_1881_cast_fp16)[name = tensor("op_1882_cast_fp16")]; + tensor x_207_cast_fp16 = real_div(x = sub_52_cast_fp16, y = var_1882_cast_fp16)[name = tensor("x_207_cast_fp16")]; + tensor var_1884_cast_fp16 = mul(x = x_207_cast_fp16, y = flow_net_res_blocks_4_in_ln_weight_to_fp16)[name = tensor("op_1884_cast_fp16")]; + tensor x_209_cast_fp16 = add(x = var_1884_cast_fp16, y = flow_net_res_blocks_4_in_ln_bias_to_fp16)[name = tensor("x_209_cast_fp16")]; + tensor var_1886_promoted_to_fp16 = const()[name = tensor("op_1886_promoted_to_fp16"), val = tensor(0x1p+0)]; + tensor var_1887_cast_fp16 = add(x = var_1870_cast_fp16_1, y = var_1886_promoted_to_fp16)[name = tensor("op_1887_cast_fp16")]; + tensor var_1888_cast_fp16 = mul(x = x_209_cast_fp16, y = var_1887_cast_fp16)[name = tensor("op_1888_cast_fp16")]; + tensor input_321_cast_fp16 = add(x = var_1888_cast_fp16, y = var_1870_cast_fp16_0)[name = tensor("input_321_cast_fp16")]; + tensor linear_123_cast_fp16 = linear(bias = flow_net_res_blocks_4_mlp_0_bias_to_fp16, weight = flow_net_res_blocks_4_mlp_0_weight_to_fp16, x = input_321_cast_fp16)[name = tensor("linear_123_cast_fp16")]; + tensor input_325_cast_fp16 = silu(x = linear_123_cast_fp16)[name = tensor("input_325_cast_fp16")]; + tensor linear_124_cast_fp16 = linear(bias = flow_net_res_blocks_4_mlp_2_bias_to_fp16, weight = flow_net_res_blocks_4_mlp_2_weight_to_fp16, x = input_325_cast_fp16)[name = tensor("linear_124_cast_fp16")]; + tensor var_1899_cast_fp16 = mul(x = var_1870_cast_fp16_2, y = linear_124_cast_fp16)[name = tensor("op_1899_cast_fp16")]; + tensor x_211_cast_fp16 = add(x = x_205_cast_fp16, y = var_1899_cast_fp16)[name = tensor("x_211_cast_fp16")]; + tensor linear_125_cast_fp16 = linear(bias = flow_net_res_blocks_5_adaLN_modulation_1_bias_to_fp16, weight = flow_net_res_blocks_5_adaLN_modulation_1_weight_to_fp16, x = input_287_cast_fp16)[name = tensor("linear_125_cast_fp16")]; + tensor var_1909_split_sizes_0 = const()[name = tensor("op_1909_split_sizes_0"), val = tensor([512, 512, 512])]; + tensor var_1909_axis_0 = const()[name = tensor("op_1909_axis_0"), val = tensor(-1)]; + tensor var_1909_cast_fp16_0, tensor var_1909_cast_fp16_1, tensor var_1909_cast_fp16_2 = split(axis = var_1909_axis_0, split_sizes = var_1909_split_sizes_0, x = linear_125_cast_fp16)[name = tensor("op_1909_cast_fp16")]; + tensor mean_67_axes_0 = const()[name = tensor("mean_67_axes_0"), val = tensor([-1])]; + tensor mean_67_keep_dims_0 = const()[name = tensor("mean_67_keep_dims_0"), val = tensor(true)]; + tensor mean_67_cast_fp16 = reduce_mean(axes = mean_67_axes_0, keep_dims = mean_67_keep_dims_0, x = x_211_cast_fp16)[name = tensor("mean_67_cast_fp16")]; + tensor sub_53_cast_fp16 = sub(x = x_211_cast_fp16, y = mean_67_cast_fp16)[name = tensor("sub_53_cast_fp16")]; + tensor square_43_cast_fp16 = square(x = sub_53_cast_fp16)[name = tensor("square_43_cast_fp16")]; + tensor reduce_mean_87_axes_0 = const()[name = tensor("reduce_mean_87_axes_0"), val = tensor([-1])]; + tensor reduce_mean_87_keep_dims_0 = const()[name = tensor("reduce_mean_87_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_87_cast_fp16 = reduce_mean(axes = reduce_mean_87_axes_0, keep_dims = reduce_mean_87_keep_dims_0, x = square_43_cast_fp16)[name = tensor("reduce_mean_87_cast_fp16")]; + tensor var_1919_to_fp16 = const()[name = tensor("op_1919_to_fp16"), val = tensor(0x1.1p-20)]; + tensor var_1920_cast_fp16 = add(x = reduce_mean_87_cast_fp16, y = var_1919_to_fp16)[name = tensor("op_1920_cast_fp16")]; + tensor var_1921_cast_fp16 = sqrt(x = var_1920_cast_fp16)[name = tensor("op_1921_cast_fp16")]; + tensor x_213_cast_fp16 = real_div(x = sub_53_cast_fp16, y = var_1921_cast_fp16)[name = tensor("x_213_cast_fp16")]; + tensor var_1923_cast_fp16 = mul(x = x_213_cast_fp16, y = flow_net_res_blocks_5_in_ln_weight_to_fp16)[name = tensor("op_1923_cast_fp16")]; + tensor x_215_cast_fp16 = add(x = var_1923_cast_fp16, y = flow_net_res_blocks_5_in_ln_bias_to_fp16)[name = tensor("x_215_cast_fp16")]; + tensor var_1925_promoted_to_fp16 = const()[name = tensor("op_1925_promoted_to_fp16"), val = tensor(0x1p+0)]; + tensor var_1926_cast_fp16 = add(x = var_1909_cast_fp16_1, y = var_1925_promoted_to_fp16)[name = tensor("op_1926_cast_fp16")]; + tensor var_1927_cast_fp16 = mul(x = x_215_cast_fp16, y = var_1926_cast_fp16)[name = tensor("op_1927_cast_fp16")]; + tensor input_329_cast_fp16 = add(x = var_1927_cast_fp16, y = var_1909_cast_fp16_0)[name = tensor("input_329_cast_fp16")]; + tensor linear_126_cast_fp16 = linear(bias = flow_net_res_blocks_5_mlp_0_bias_to_fp16, weight = flow_net_res_blocks_5_mlp_0_weight_to_fp16, x = input_329_cast_fp16)[name = tensor("linear_126_cast_fp16")]; + tensor input_333_cast_fp16 = silu(x = linear_126_cast_fp16)[name = tensor("input_333_cast_fp16")]; + tensor linear_127_cast_fp16 = linear(bias = flow_net_res_blocks_5_mlp_2_bias_to_fp16, weight = flow_net_res_blocks_5_mlp_2_weight_to_fp16, x = input_333_cast_fp16)[name = tensor("linear_127_cast_fp16")]; + tensor var_1938_cast_fp16 = mul(x = var_1909_cast_fp16_2, y = linear_127_cast_fp16)[name = tensor("op_1938_cast_fp16")]; + tensor x_217_cast_fp16 = add(x = x_211_cast_fp16, y = var_1938_cast_fp16)[name = tensor("x_217_cast_fp16")]; + tensor linear_128_cast_fp16 = linear(bias = flow_net_final_layer_adaLN_modulation_1_bias_to_fp16, weight = flow_net_final_layer_adaLN_modulation_1_weight_to_fp16, x = input_287_cast_fp16)[name = tensor("linear_128_cast_fp16")]; + tensor var_1947_split_sizes_0 = const()[name = tensor("op_1947_split_sizes_0"), val = tensor([512, 512])]; + tensor var_1947_axis_0 = const()[name = tensor("op_1947_axis_0"), val = tensor(-1)]; + tensor var_1947_cast_fp16_0, tensor var_1947_cast_fp16_1 = split(axis = var_1947_axis_0, split_sizes = var_1947_split_sizes_0, x = linear_128_cast_fp16)[name = tensor("op_1947_cast_fp16")]; + tensor mean_69_axes_0 = const()[name = tensor("mean_69_axes_0"), val = tensor([-1])]; + tensor mean_69_keep_dims_0 = const()[name = tensor("mean_69_keep_dims_0"), val = tensor(true)]; + tensor mean_69_cast_fp16 = reduce_mean(axes = mean_69_axes_0, keep_dims = mean_69_keep_dims_0, x = x_217_cast_fp16)[name = tensor("mean_69_cast_fp16")]; + tensor sub_54_cast_fp16 = sub(x = x_217_cast_fp16, y = mean_69_cast_fp16)[name = tensor("sub_54_cast_fp16")]; + tensor square_44_cast_fp16 = square(x = sub_54_cast_fp16)[name = tensor("square_44_cast_fp16")]; + tensor reduce_mean_89_axes_0 = const()[name = tensor("reduce_mean_89_axes_0"), val = tensor([-1])]; + tensor reduce_mean_89_keep_dims_0 = const()[name = tensor("reduce_mean_89_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_89_cast_fp16 = reduce_mean(axes = reduce_mean_89_axes_0, keep_dims = reduce_mean_89_keep_dims_0, x = square_44_cast_fp16)[name = tensor("reduce_mean_89_cast_fp16")]; + tensor var_1954_to_fp16 = const()[name = tensor("op_1954_to_fp16"), val = tensor(0x1.1p-20)]; + tensor var_1955_cast_fp16 = add(x = reduce_mean_89_cast_fp16, y = var_1954_to_fp16)[name = tensor("op_1955_cast_fp16")]; + tensor var_1956_cast_fp16 = sqrt(x = var_1955_cast_fp16)[name = tensor("op_1956_cast_fp16")]; + tensor x_219_cast_fp16 = real_div(x = sub_54_cast_fp16, y = var_1956_cast_fp16)[name = tensor("x_219_cast_fp16")]; + tensor var_1958_promoted_to_fp16 = const()[name = tensor("op_1958_promoted_to_fp16"), val = tensor(0x1p+0)]; + tensor var_1959_cast_fp16 = add(x = var_1947_cast_fp16_1, y = var_1958_promoted_to_fp16)[name = tensor("op_1959_cast_fp16")]; + tensor var_1960_cast_fp16 = mul(x = x_219_cast_fp16, y = var_1959_cast_fp16)[name = tensor("op_1960_cast_fp16")]; + tensor input_337_cast_fp16 = add(x = var_1960_cast_fp16, y = var_1947_cast_fp16_0)[name = tensor("input_337_cast_fp16")]; + tensor linear_129_cast_fp16 = linear(bias = flow_net_final_layer_linear_bias_to_fp16, weight = flow_net_final_layer_linear_weight_to_fp16, x = input_337_cast_fp16)[name = tensor("linear_129_cast_fp16")]; + tensor var_1971_to_fp16 = const()[name = tensor("op_1971_to_fp16"), val = tensor(0x1p-3)]; + tensor var_1972_cast_fp16 = mul(x = linear_129_cast_fp16, y = var_1971_to_fp16)[name = tensor("op_1972_cast_fp16")]; + tensor input_339_cast_fp16 = add(x = input_271_cast_fp16, y = var_1972_cast_fp16)[name = tensor("input_339_cast_fp16")]; + tensor linear_130_cast_fp16 = linear(bias = flow_net_input_proj_bias_to_fp16, weight = flow_net_input_proj_weight_to_fp16, x = input_339_cast_fp16)[name = tensor("linear_130_cast_fp16")]; + tensor input_343_to_fp16 = const()[name = tensor("input_343_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(19006080)))]; + tensor input_345_cast_fp16 = silu(x = input_343_to_fp16)[name = tensor("input_345_cast_fp16")]; + tensor linear_132_cast_fp16 = linear(bias = flow_net_time_embed_0_mlp_2_bias_to_fp16, weight = flow_net_time_embed_0_mlp_2_weight_to_fp16, x = input_345_cast_fp16)[name = tensor("linear_132_cast_fp16")]; + tensor reduce_mean_90_axes_0 = const()[name = tensor("reduce_mean_90_axes_0"), val = tensor([-1])]; + tensor reduce_mean_90_keep_dims_0 = const()[name = tensor("reduce_mean_90_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_90_cast_fp16 = reduce_mean(axes = reduce_mean_90_axes_0, keep_dims = reduce_mean_90_keep_dims_0, x = linear_132_cast_fp16)[name = tensor("reduce_mean_90_cast_fp16")]; + tensor sub_55_cast_fp16 = sub(x = linear_132_cast_fp16, y = reduce_mean_90_cast_fp16)[name = tensor("sub_55_cast_fp16")]; + tensor square_45_cast_fp16 = square(x = sub_55_cast_fp16)[name = tensor("square_45_cast_fp16")]; + tensor reduce_mean_91_axes_0 = const()[name = tensor("reduce_mean_91_axes_0"), val = tensor([-1])]; + tensor reduce_mean_91_keep_dims_0 = const()[name = tensor("reduce_mean_91_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_91_cast_fp16 = reduce_mean(axes = reduce_mean_91_axes_0, keep_dims = reduce_mean_91_keep_dims_0, x = square_45_cast_fp16)[name = tensor("reduce_mean_91_cast_fp16")]; + tensor real_div_10_to_fp16 = const()[name = tensor("real_div_10_to_fp16"), val = tensor(0x1.008p+0)]; + tensor mul_10_cast_fp16 = mul(x = reduce_mean_91_cast_fp16, y = real_div_10_to_fp16)[name = tensor("mul_10_cast_fp16")]; + tensor var_2038_to_fp16 = const()[name = tensor("op_2038_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_91_cast_fp16 = add(x = mul_10_cast_fp16, y = var_2038_to_fp16)[name = tensor("var_91_cast_fp16")]; + tensor var_2041_epsilon_0 = const()[name = tensor("op_2041_epsilon_0"), val = tensor(0x1.197998p-40)]; + tensor var_2041_cast_fp16 = rsqrt(epsilon = var_2041_epsilon_0, x = var_91_cast_fp16)[name = tensor("op_2041_cast_fp16")]; + tensor var_2042_cast_fp16 = mul(x = const_3_to_fp16, y = var_2041_cast_fp16)[name = tensor("op_2042_cast_fp16")]; + tensor var_2043_cast_fp16 = mul(x = linear_132_cast_fp16, y = var_2042_cast_fp16)[name = tensor("op_2043_cast_fp16")]; + tensor input_349_to_fp16 = const()[name = tensor("input_349_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(19007168)))]; + tensor input_351_cast_fp16 = silu(x = input_349_to_fp16)[name = tensor("input_351_cast_fp16")]; + tensor linear_134_cast_fp16 = linear(bias = flow_net_time_embed_1_mlp_2_bias_to_fp16, weight = flow_net_time_embed_1_mlp_2_weight_to_fp16, x = input_351_cast_fp16)[name = tensor("linear_134_cast_fp16")]; + tensor reduce_mean_92_axes_0 = const()[name = tensor("reduce_mean_92_axes_0"), val = tensor([-1])]; + tensor reduce_mean_92_keep_dims_0 = const()[name = tensor("reduce_mean_92_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_92_cast_fp16 = reduce_mean(axes = reduce_mean_92_axes_0, keep_dims = reduce_mean_92_keep_dims_0, x = linear_134_cast_fp16)[name = tensor("reduce_mean_92_cast_fp16")]; + tensor sub_57_cast_fp16 = sub(x = linear_134_cast_fp16, y = reduce_mean_92_cast_fp16)[name = tensor("sub_57_cast_fp16")]; + tensor square_46_cast_fp16 = square(x = sub_57_cast_fp16)[name = tensor("square_46_cast_fp16")]; + tensor reduce_mean_93_axes_0 = const()[name = tensor("reduce_mean_93_axes_0"), val = tensor([-1])]; + tensor reduce_mean_93_keep_dims_0 = const()[name = tensor("reduce_mean_93_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_93_cast_fp16 = reduce_mean(axes = reduce_mean_93_axes_0, keep_dims = reduce_mean_93_keep_dims_0, x = square_46_cast_fp16)[name = tensor("reduce_mean_93_cast_fp16")]; + tensor real_div_11_to_fp16 = const()[name = tensor("real_div_11_to_fp16"), val = tensor(0x1.008p+0)]; + tensor mul_11_cast_fp16 = mul(x = reduce_mean_93_cast_fp16, y = real_div_11_to_fp16)[name = tensor("mul_11_cast_fp16")]; + tensor var_2075_to_fp16 = const()[name = tensor("op_2075_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_93_cast_fp16 = add(x = mul_11_cast_fp16, y = var_2075_to_fp16)[name = tensor("var_93_cast_fp16")]; + tensor var_2078_epsilon_0 = const()[name = tensor("op_2078_epsilon_0"), val = tensor(0x1.197998p-40)]; + tensor var_2078_cast_fp16 = rsqrt(epsilon = var_2078_epsilon_0, x = var_93_cast_fp16)[name = tensor("op_2078_cast_fp16")]; + tensor var_2079_cast_fp16 = mul(x = const_5_to_fp16, y = var_2078_cast_fp16)[name = tensor("op_2079_cast_fp16")]; + tensor var_2080_cast_fp16 = mul(x = linear_134_cast_fp16, y = var_2079_cast_fp16)[name = tensor("op_2080_cast_fp16")]; + tensor var_2092_cast_fp16 = add(x = var_2043_cast_fp16, y = var_2080_cast_fp16)[name = tensor("op_2092_cast_fp16")]; + tensor _inversed_t_combined_11_y_0_to_fp16 = const()[name = tensor("_inversed_t_combined_11_y_0_to_fp16"), val = tensor(0x1p-1)]; + tensor _inversed_t_combined_11_cast_fp16 = mul(x = var_2092_cast_fp16, y = _inversed_t_combined_11_y_0_to_fp16)[name = tensor("_inversed_t_combined_11_cast_fp16")]; + tensor input_353_cast_fp16 = add(x = _inversed_t_combined_11_cast_fp16, y = linear_5_cast_fp16)[name = tensor("input_353_cast_fp16")]; + tensor input_355_cast_fp16 = silu(x = input_353_cast_fp16)[name = tensor("input_355_cast_fp16")]; + tensor linear_136_cast_fp16 = linear(bias = flow_net_res_blocks_0_adaLN_modulation_1_bias_to_fp16, weight = flow_net_res_blocks_0_adaLN_modulation_1_weight_to_fp16, x = input_355_cast_fp16)[name = tensor("linear_136_cast_fp16")]; + tensor var_2107_split_sizes_0 = const()[name = tensor("op_2107_split_sizes_0"), val = tensor([512, 512, 512])]; + tensor var_2107_axis_0 = const()[name = tensor("op_2107_axis_0"), val = tensor(-1)]; + tensor var_2107_cast_fp16_0, tensor var_2107_cast_fp16_1, tensor var_2107_cast_fp16_2 = split(axis = var_2107_axis_0, split_sizes = var_2107_split_sizes_0, x = linear_136_cast_fp16)[name = tensor("op_2107_cast_fp16")]; + tensor mean_71_axes_0 = const()[name = tensor("mean_71_axes_0"), val = tensor([-1])]; + tensor mean_71_keep_dims_0 = const()[name = tensor("mean_71_keep_dims_0"), val = tensor(true)]; + tensor mean_71_cast_fp16 = reduce_mean(axes = mean_71_axes_0, keep_dims = mean_71_keep_dims_0, x = linear_130_cast_fp16)[name = tensor("mean_71_cast_fp16")]; + tensor sub_59_cast_fp16 = sub(x = linear_130_cast_fp16, y = mean_71_cast_fp16)[name = tensor("sub_59_cast_fp16")]; + tensor square_47_cast_fp16 = square(x = sub_59_cast_fp16)[name = tensor("square_47_cast_fp16")]; + tensor reduce_mean_95_axes_0 = const()[name = tensor("reduce_mean_95_axes_0"), val = tensor([-1])]; + tensor reduce_mean_95_keep_dims_0 = const()[name = tensor("reduce_mean_95_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_95_cast_fp16 = reduce_mean(axes = reduce_mean_95_axes_0, keep_dims = reduce_mean_95_keep_dims_0, x = square_47_cast_fp16)[name = tensor("reduce_mean_95_cast_fp16")]; + tensor var_2117_to_fp16 = const()[name = tensor("op_2117_to_fp16"), val = tensor(0x1.1p-20)]; + tensor var_2118_cast_fp16 = add(x = reduce_mean_95_cast_fp16, y = var_2117_to_fp16)[name = tensor("op_2118_cast_fp16")]; + tensor var_2119_cast_fp16 = sqrt(x = var_2118_cast_fp16)[name = tensor("op_2119_cast_fp16")]; + tensor x_227_cast_fp16 = real_div(x = sub_59_cast_fp16, y = var_2119_cast_fp16)[name = tensor("x_227_cast_fp16")]; + tensor var_2121_cast_fp16 = mul(x = x_227_cast_fp16, y = flow_net_res_blocks_0_in_ln_weight_to_fp16)[name = tensor("op_2121_cast_fp16")]; + tensor x_229_cast_fp16 = add(x = var_2121_cast_fp16, y = flow_net_res_blocks_0_in_ln_bias_to_fp16)[name = tensor("x_229_cast_fp16")]; + tensor var_2123_promoted_to_fp16 = const()[name = tensor("op_2123_promoted_to_fp16"), val = tensor(0x1p+0)]; + tensor var_2124_cast_fp16 = add(x = var_2107_cast_fp16_1, y = var_2123_promoted_to_fp16)[name = tensor("op_2124_cast_fp16")]; + tensor var_2125_cast_fp16 = mul(x = x_229_cast_fp16, y = var_2124_cast_fp16)[name = tensor("op_2125_cast_fp16")]; + tensor input_357_cast_fp16 = add(x = var_2125_cast_fp16, y = var_2107_cast_fp16_0)[name = tensor("input_357_cast_fp16")]; + tensor linear_137_cast_fp16 = linear(bias = flow_net_res_blocks_0_mlp_0_bias_to_fp16, weight = flow_net_res_blocks_0_mlp_0_weight_to_fp16, x = input_357_cast_fp16)[name = tensor("linear_137_cast_fp16")]; + tensor input_361_cast_fp16 = silu(x = linear_137_cast_fp16)[name = tensor("input_361_cast_fp16")]; + tensor linear_138_cast_fp16 = linear(bias = flow_net_res_blocks_0_mlp_2_bias_to_fp16, weight = flow_net_res_blocks_0_mlp_2_weight_to_fp16, x = input_361_cast_fp16)[name = tensor("linear_138_cast_fp16")]; + tensor var_2136_cast_fp16 = mul(x = var_2107_cast_fp16_2, y = linear_138_cast_fp16)[name = tensor("op_2136_cast_fp16")]; + tensor x_231_cast_fp16 = add(x = linear_130_cast_fp16, y = var_2136_cast_fp16)[name = tensor("x_231_cast_fp16")]; + tensor linear_139_cast_fp16 = linear(bias = flow_net_res_blocks_1_adaLN_modulation_1_bias_to_fp16, weight = flow_net_res_blocks_1_adaLN_modulation_1_weight_to_fp16, x = input_355_cast_fp16)[name = tensor("linear_139_cast_fp16")]; + tensor var_2146_split_sizes_0 = const()[name = tensor("op_2146_split_sizes_0"), val = tensor([512, 512, 512])]; + tensor var_2146_axis_0 = const()[name = tensor("op_2146_axis_0"), val = tensor(-1)]; + tensor var_2146_cast_fp16_0, tensor var_2146_cast_fp16_1, tensor var_2146_cast_fp16_2 = split(axis = var_2146_axis_0, split_sizes = var_2146_split_sizes_0, x = linear_139_cast_fp16)[name = tensor("op_2146_cast_fp16")]; + tensor mean_73_axes_0 = const()[name = tensor("mean_73_axes_0"), val = tensor([-1])]; + tensor mean_73_keep_dims_0 = const()[name = tensor("mean_73_keep_dims_0"), val = tensor(true)]; + tensor mean_73_cast_fp16 = reduce_mean(axes = mean_73_axes_0, keep_dims = mean_73_keep_dims_0, x = x_231_cast_fp16)[name = tensor("mean_73_cast_fp16")]; + tensor sub_60_cast_fp16 = sub(x = x_231_cast_fp16, y = mean_73_cast_fp16)[name = tensor("sub_60_cast_fp16")]; + tensor square_48_cast_fp16 = square(x = sub_60_cast_fp16)[name = tensor("square_48_cast_fp16")]; + tensor reduce_mean_97_axes_0 = const()[name = tensor("reduce_mean_97_axes_0"), val = tensor([-1])]; + tensor reduce_mean_97_keep_dims_0 = const()[name = tensor("reduce_mean_97_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_97_cast_fp16 = reduce_mean(axes = reduce_mean_97_axes_0, keep_dims = reduce_mean_97_keep_dims_0, x = square_48_cast_fp16)[name = tensor("reduce_mean_97_cast_fp16")]; + tensor var_2156_to_fp16 = const()[name = tensor("op_2156_to_fp16"), val = tensor(0x1.1p-20)]; + tensor var_2157_cast_fp16 = add(x = reduce_mean_97_cast_fp16, y = var_2156_to_fp16)[name = tensor("op_2157_cast_fp16")]; + tensor var_2158_cast_fp16 = sqrt(x = var_2157_cast_fp16)[name = tensor("op_2158_cast_fp16")]; + tensor x_233_cast_fp16 = real_div(x = sub_60_cast_fp16, y = var_2158_cast_fp16)[name = tensor("x_233_cast_fp16")]; + tensor var_2160_cast_fp16 = mul(x = x_233_cast_fp16, y = flow_net_res_blocks_1_in_ln_weight_to_fp16)[name = tensor("op_2160_cast_fp16")]; + tensor x_235_cast_fp16 = add(x = var_2160_cast_fp16, y = flow_net_res_blocks_1_in_ln_bias_to_fp16)[name = tensor("x_235_cast_fp16")]; + tensor var_2162_promoted_to_fp16 = const()[name = tensor("op_2162_promoted_to_fp16"), val = tensor(0x1p+0)]; + tensor var_2163_cast_fp16 = add(x = var_2146_cast_fp16_1, y = var_2162_promoted_to_fp16)[name = tensor("op_2163_cast_fp16")]; + tensor var_2164_cast_fp16 = mul(x = x_235_cast_fp16, y = var_2163_cast_fp16)[name = tensor("op_2164_cast_fp16")]; + tensor input_365_cast_fp16 = add(x = var_2164_cast_fp16, y = var_2146_cast_fp16_0)[name = tensor("input_365_cast_fp16")]; + tensor linear_140_cast_fp16 = linear(bias = flow_net_res_blocks_1_mlp_0_bias_to_fp16, weight = flow_net_res_blocks_1_mlp_0_weight_to_fp16, x = input_365_cast_fp16)[name = tensor("linear_140_cast_fp16")]; + tensor input_369_cast_fp16 = silu(x = linear_140_cast_fp16)[name = tensor("input_369_cast_fp16")]; + tensor linear_141_cast_fp16 = linear(bias = flow_net_res_blocks_1_mlp_2_bias_to_fp16, weight = flow_net_res_blocks_1_mlp_2_weight_to_fp16, x = input_369_cast_fp16)[name = tensor("linear_141_cast_fp16")]; + tensor var_2175_cast_fp16 = mul(x = var_2146_cast_fp16_2, y = linear_141_cast_fp16)[name = tensor("op_2175_cast_fp16")]; + tensor x_237_cast_fp16 = add(x = x_231_cast_fp16, y = var_2175_cast_fp16)[name = tensor("x_237_cast_fp16")]; + tensor linear_142_cast_fp16 = linear(bias = flow_net_res_blocks_2_adaLN_modulation_1_bias_to_fp16, weight = flow_net_res_blocks_2_adaLN_modulation_1_weight_to_fp16, x = input_355_cast_fp16)[name = tensor("linear_142_cast_fp16")]; + tensor var_2185_split_sizes_0 = const()[name = tensor("op_2185_split_sizes_0"), val = tensor([512, 512, 512])]; + tensor var_2185_axis_0 = const()[name = tensor("op_2185_axis_0"), val = tensor(-1)]; + tensor var_2185_cast_fp16_0, tensor var_2185_cast_fp16_1, tensor var_2185_cast_fp16_2 = split(axis = var_2185_axis_0, split_sizes = var_2185_split_sizes_0, x = linear_142_cast_fp16)[name = tensor("op_2185_cast_fp16")]; + tensor mean_75_axes_0 = const()[name = tensor("mean_75_axes_0"), val = tensor([-1])]; + tensor mean_75_keep_dims_0 = const()[name = tensor("mean_75_keep_dims_0"), val = tensor(true)]; + tensor mean_75_cast_fp16 = reduce_mean(axes = mean_75_axes_0, keep_dims = mean_75_keep_dims_0, x = x_237_cast_fp16)[name = tensor("mean_75_cast_fp16")]; + tensor sub_61_cast_fp16 = sub(x = x_237_cast_fp16, y = mean_75_cast_fp16)[name = tensor("sub_61_cast_fp16")]; + tensor square_49_cast_fp16 = square(x = sub_61_cast_fp16)[name = tensor("square_49_cast_fp16")]; + tensor reduce_mean_99_axes_0 = const()[name = tensor("reduce_mean_99_axes_0"), val = tensor([-1])]; + tensor reduce_mean_99_keep_dims_0 = const()[name = tensor("reduce_mean_99_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_99_cast_fp16 = reduce_mean(axes = reduce_mean_99_axes_0, keep_dims = reduce_mean_99_keep_dims_0, x = square_49_cast_fp16)[name = tensor("reduce_mean_99_cast_fp16")]; + tensor var_2195_to_fp16 = const()[name = tensor("op_2195_to_fp16"), val = tensor(0x1.1p-20)]; + tensor var_2196_cast_fp16 = add(x = reduce_mean_99_cast_fp16, y = var_2195_to_fp16)[name = tensor("op_2196_cast_fp16")]; + tensor var_2197_cast_fp16 = sqrt(x = var_2196_cast_fp16)[name = tensor("op_2197_cast_fp16")]; + tensor x_239_cast_fp16 = real_div(x = sub_61_cast_fp16, y = var_2197_cast_fp16)[name = tensor("x_239_cast_fp16")]; + tensor var_2199_cast_fp16 = mul(x = x_239_cast_fp16, y = flow_net_res_blocks_2_in_ln_weight_to_fp16)[name = tensor("op_2199_cast_fp16")]; + tensor x_241_cast_fp16 = add(x = var_2199_cast_fp16, y = flow_net_res_blocks_2_in_ln_bias_to_fp16)[name = tensor("x_241_cast_fp16")]; + tensor var_2201_promoted_to_fp16 = const()[name = tensor("op_2201_promoted_to_fp16"), val = tensor(0x1p+0)]; + tensor var_2202_cast_fp16 = add(x = var_2185_cast_fp16_1, y = var_2201_promoted_to_fp16)[name = tensor("op_2202_cast_fp16")]; + tensor var_2203_cast_fp16 = mul(x = x_241_cast_fp16, y = var_2202_cast_fp16)[name = tensor("op_2203_cast_fp16")]; + tensor input_373_cast_fp16 = add(x = var_2203_cast_fp16, y = var_2185_cast_fp16_0)[name = tensor("input_373_cast_fp16")]; + tensor linear_143_cast_fp16 = linear(bias = flow_net_res_blocks_2_mlp_0_bias_to_fp16, weight = flow_net_res_blocks_2_mlp_0_weight_to_fp16, x = input_373_cast_fp16)[name = tensor("linear_143_cast_fp16")]; + tensor input_377_cast_fp16 = silu(x = linear_143_cast_fp16)[name = tensor("input_377_cast_fp16")]; + tensor linear_144_cast_fp16 = linear(bias = flow_net_res_blocks_2_mlp_2_bias_to_fp16, weight = flow_net_res_blocks_2_mlp_2_weight_to_fp16, x = input_377_cast_fp16)[name = tensor("linear_144_cast_fp16")]; + tensor var_2214_cast_fp16 = mul(x = var_2185_cast_fp16_2, y = linear_144_cast_fp16)[name = tensor("op_2214_cast_fp16")]; + tensor x_243_cast_fp16 = add(x = x_237_cast_fp16, y = var_2214_cast_fp16)[name = tensor("x_243_cast_fp16")]; + tensor linear_145_cast_fp16 = linear(bias = flow_net_res_blocks_3_adaLN_modulation_1_bias_to_fp16, weight = flow_net_res_blocks_3_adaLN_modulation_1_weight_to_fp16, x = input_355_cast_fp16)[name = tensor("linear_145_cast_fp16")]; + tensor var_2224_split_sizes_0 = const()[name = tensor("op_2224_split_sizes_0"), val = tensor([512, 512, 512])]; + tensor var_2224_axis_0 = const()[name = tensor("op_2224_axis_0"), val = tensor(-1)]; + tensor var_2224_cast_fp16_0, tensor var_2224_cast_fp16_1, tensor var_2224_cast_fp16_2 = split(axis = var_2224_axis_0, split_sizes = var_2224_split_sizes_0, x = linear_145_cast_fp16)[name = tensor("op_2224_cast_fp16")]; + tensor mean_77_axes_0 = const()[name = tensor("mean_77_axes_0"), val = tensor([-1])]; + tensor mean_77_keep_dims_0 = const()[name = tensor("mean_77_keep_dims_0"), val = tensor(true)]; + tensor mean_77_cast_fp16 = reduce_mean(axes = mean_77_axes_0, keep_dims = mean_77_keep_dims_0, x = x_243_cast_fp16)[name = tensor("mean_77_cast_fp16")]; + tensor sub_62_cast_fp16 = sub(x = x_243_cast_fp16, y = mean_77_cast_fp16)[name = tensor("sub_62_cast_fp16")]; + tensor square_50_cast_fp16 = square(x = sub_62_cast_fp16)[name = tensor("square_50_cast_fp16")]; + tensor reduce_mean_101_axes_0 = const()[name = tensor("reduce_mean_101_axes_0"), val = tensor([-1])]; + tensor reduce_mean_101_keep_dims_0 = const()[name = tensor("reduce_mean_101_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_101_cast_fp16 = reduce_mean(axes = reduce_mean_101_axes_0, keep_dims = reduce_mean_101_keep_dims_0, x = square_50_cast_fp16)[name = tensor("reduce_mean_101_cast_fp16")]; + tensor var_2234_to_fp16 = const()[name = tensor("op_2234_to_fp16"), val = tensor(0x1.1p-20)]; + tensor var_2235_cast_fp16 = add(x = reduce_mean_101_cast_fp16, y = var_2234_to_fp16)[name = tensor("op_2235_cast_fp16")]; + tensor var_2236_cast_fp16 = sqrt(x = var_2235_cast_fp16)[name = tensor("op_2236_cast_fp16")]; + tensor x_245_cast_fp16 = real_div(x = sub_62_cast_fp16, y = var_2236_cast_fp16)[name = tensor("x_245_cast_fp16")]; + tensor var_2238_cast_fp16 = mul(x = x_245_cast_fp16, y = flow_net_res_blocks_3_in_ln_weight_to_fp16)[name = tensor("op_2238_cast_fp16")]; + tensor x_247_cast_fp16 = add(x = var_2238_cast_fp16, y = flow_net_res_blocks_3_in_ln_bias_to_fp16)[name = tensor("x_247_cast_fp16")]; + tensor var_2240_promoted_to_fp16 = const()[name = tensor("op_2240_promoted_to_fp16"), val = tensor(0x1p+0)]; + tensor var_2241_cast_fp16 = add(x = var_2224_cast_fp16_1, y = var_2240_promoted_to_fp16)[name = tensor("op_2241_cast_fp16")]; + tensor var_2242_cast_fp16 = mul(x = x_247_cast_fp16, y = var_2241_cast_fp16)[name = tensor("op_2242_cast_fp16")]; + tensor input_381_cast_fp16 = add(x = var_2242_cast_fp16, y = var_2224_cast_fp16_0)[name = tensor("input_381_cast_fp16")]; + tensor linear_146_cast_fp16 = linear(bias = flow_net_res_blocks_3_mlp_0_bias_to_fp16, weight = flow_net_res_blocks_3_mlp_0_weight_to_fp16, x = input_381_cast_fp16)[name = tensor("linear_146_cast_fp16")]; + tensor input_385_cast_fp16 = silu(x = linear_146_cast_fp16)[name = tensor("input_385_cast_fp16")]; + tensor linear_147_cast_fp16 = linear(bias = flow_net_res_blocks_3_mlp_2_bias_to_fp16, weight = flow_net_res_blocks_3_mlp_2_weight_to_fp16, x = input_385_cast_fp16)[name = tensor("linear_147_cast_fp16")]; + tensor var_2253_cast_fp16 = mul(x = var_2224_cast_fp16_2, y = linear_147_cast_fp16)[name = tensor("op_2253_cast_fp16")]; + tensor x_249_cast_fp16 = add(x = x_243_cast_fp16, y = var_2253_cast_fp16)[name = tensor("x_249_cast_fp16")]; + tensor linear_148_cast_fp16 = linear(bias = flow_net_res_blocks_4_adaLN_modulation_1_bias_to_fp16, weight = flow_net_res_blocks_4_adaLN_modulation_1_weight_to_fp16, x = input_355_cast_fp16)[name = tensor("linear_148_cast_fp16")]; + tensor var_2263_split_sizes_0 = const()[name = tensor("op_2263_split_sizes_0"), val = tensor([512, 512, 512])]; + tensor var_2263_axis_0 = const()[name = tensor("op_2263_axis_0"), val = tensor(-1)]; + tensor var_2263_cast_fp16_0, tensor var_2263_cast_fp16_1, tensor var_2263_cast_fp16_2 = split(axis = var_2263_axis_0, split_sizes = var_2263_split_sizes_0, x = linear_148_cast_fp16)[name = tensor("op_2263_cast_fp16")]; + tensor mean_79_axes_0 = const()[name = tensor("mean_79_axes_0"), val = tensor([-1])]; + tensor mean_79_keep_dims_0 = const()[name = tensor("mean_79_keep_dims_0"), val = tensor(true)]; + tensor mean_79_cast_fp16 = reduce_mean(axes = mean_79_axes_0, keep_dims = mean_79_keep_dims_0, x = x_249_cast_fp16)[name = tensor("mean_79_cast_fp16")]; + tensor sub_63_cast_fp16 = sub(x = x_249_cast_fp16, y = mean_79_cast_fp16)[name = tensor("sub_63_cast_fp16")]; + tensor square_51_cast_fp16 = square(x = sub_63_cast_fp16)[name = tensor("square_51_cast_fp16")]; + tensor reduce_mean_103_axes_0 = const()[name = tensor("reduce_mean_103_axes_0"), val = tensor([-1])]; + tensor reduce_mean_103_keep_dims_0 = const()[name = tensor("reduce_mean_103_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_103_cast_fp16 = reduce_mean(axes = reduce_mean_103_axes_0, keep_dims = reduce_mean_103_keep_dims_0, x = square_51_cast_fp16)[name = tensor("reduce_mean_103_cast_fp16")]; + tensor var_2273_to_fp16 = const()[name = tensor("op_2273_to_fp16"), val = tensor(0x1.1p-20)]; + tensor var_2274_cast_fp16 = add(x = reduce_mean_103_cast_fp16, y = var_2273_to_fp16)[name = tensor("op_2274_cast_fp16")]; + tensor var_2275_cast_fp16 = sqrt(x = var_2274_cast_fp16)[name = tensor("op_2275_cast_fp16")]; + tensor x_251_cast_fp16 = real_div(x = sub_63_cast_fp16, y = var_2275_cast_fp16)[name = tensor("x_251_cast_fp16")]; + tensor var_2277_cast_fp16 = mul(x = x_251_cast_fp16, y = flow_net_res_blocks_4_in_ln_weight_to_fp16)[name = tensor("op_2277_cast_fp16")]; + tensor x_253_cast_fp16 = add(x = var_2277_cast_fp16, y = flow_net_res_blocks_4_in_ln_bias_to_fp16)[name = tensor("x_253_cast_fp16")]; + tensor var_2279_promoted_to_fp16 = const()[name = tensor("op_2279_promoted_to_fp16"), val = tensor(0x1p+0)]; + tensor var_2280_cast_fp16 = add(x = var_2263_cast_fp16_1, y = var_2279_promoted_to_fp16)[name = tensor("op_2280_cast_fp16")]; + tensor var_2281_cast_fp16 = mul(x = x_253_cast_fp16, y = var_2280_cast_fp16)[name = tensor("op_2281_cast_fp16")]; + tensor input_389_cast_fp16 = add(x = var_2281_cast_fp16, y = var_2263_cast_fp16_0)[name = tensor("input_389_cast_fp16")]; + tensor linear_149_cast_fp16 = linear(bias = flow_net_res_blocks_4_mlp_0_bias_to_fp16, weight = flow_net_res_blocks_4_mlp_0_weight_to_fp16, x = input_389_cast_fp16)[name = tensor("linear_149_cast_fp16")]; + tensor input_393_cast_fp16 = silu(x = linear_149_cast_fp16)[name = tensor("input_393_cast_fp16")]; + tensor linear_150_cast_fp16 = linear(bias = flow_net_res_blocks_4_mlp_2_bias_to_fp16, weight = flow_net_res_blocks_4_mlp_2_weight_to_fp16, x = input_393_cast_fp16)[name = tensor("linear_150_cast_fp16")]; + tensor var_2292_cast_fp16 = mul(x = var_2263_cast_fp16_2, y = linear_150_cast_fp16)[name = tensor("op_2292_cast_fp16")]; + tensor x_255_cast_fp16 = add(x = x_249_cast_fp16, y = var_2292_cast_fp16)[name = tensor("x_255_cast_fp16")]; + tensor linear_151_cast_fp16 = linear(bias = flow_net_res_blocks_5_adaLN_modulation_1_bias_to_fp16, weight = flow_net_res_blocks_5_adaLN_modulation_1_weight_to_fp16, x = input_355_cast_fp16)[name = tensor("linear_151_cast_fp16")]; + tensor var_2302_split_sizes_0 = const()[name = tensor("op_2302_split_sizes_0"), val = tensor([512, 512, 512])]; + tensor var_2302_axis_0 = const()[name = tensor("op_2302_axis_0"), val = tensor(-1)]; + tensor var_2302_cast_fp16_0, tensor var_2302_cast_fp16_1, tensor var_2302_cast_fp16_2 = split(axis = var_2302_axis_0, split_sizes = var_2302_split_sizes_0, x = linear_151_cast_fp16)[name = tensor("op_2302_cast_fp16")]; + tensor mean_81_axes_0 = const()[name = tensor("mean_81_axes_0"), val = tensor([-1])]; + tensor mean_81_keep_dims_0 = const()[name = tensor("mean_81_keep_dims_0"), val = tensor(true)]; + tensor mean_81_cast_fp16 = reduce_mean(axes = mean_81_axes_0, keep_dims = mean_81_keep_dims_0, x = x_255_cast_fp16)[name = tensor("mean_81_cast_fp16")]; + tensor sub_64_cast_fp16 = sub(x = x_255_cast_fp16, y = mean_81_cast_fp16)[name = tensor("sub_64_cast_fp16")]; + tensor square_52_cast_fp16 = square(x = sub_64_cast_fp16)[name = tensor("square_52_cast_fp16")]; + tensor reduce_mean_105_axes_0 = const()[name = tensor("reduce_mean_105_axes_0"), val = tensor([-1])]; + tensor reduce_mean_105_keep_dims_0 = const()[name = tensor("reduce_mean_105_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_105_cast_fp16 = reduce_mean(axes = reduce_mean_105_axes_0, keep_dims = reduce_mean_105_keep_dims_0, x = square_52_cast_fp16)[name = tensor("reduce_mean_105_cast_fp16")]; + tensor var_2312_to_fp16 = const()[name = tensor("op_2312_to_fp16"), val = tensor(0x1.1p-20)]; + tensor var_2313_cast_fp16 = add(x = reduce_mean_105_cast_fp16, y = var_2312_to_fp16)[name = tensor("op_2313_cast_fp16")]; + tensor var_2314_cast_fp16 = sqrt(x = var_2313_cast_fp16)[name = tensor("op_2314_cast_fp16")]; + tensor x_257_cast_fp16 = real_div(x = sub_64_cast_fp16, y = var_2314_cast_fp16)[name = tensor("x_257_cast_fp16")]; + tensor var_2316_cast_fp16 = mul(x = x_257_cast_fp16, y = flow_net_res_blocks_5_in_ln_weight_to_fp16)[name = tensor("op_2316_cast_fp16")]; + tensor x_259_cast_fp16 = add(x = var_2316_cast_fp16, y = flow_net_res_blocks_5_in_ln_bias_to_fp16)[name = tensor("x_259_cast_fp16")]; + tensor var_2318_promoted_to_fp16 = const()[name = tensor("op_2318_promoted_to_fp16"), val = tensor(0x1p+0)]; + tensor var_2319_cast_fp16 = add(x = var_2302_cast_fp16_1, y = var_2318_promoted_to_fp16)[name = tensor("op_2319_cast_fp16")]; + tensor var_2320_cast_fp16 = mul(x = x_259_cast_fp16, y = var_2319_cast_fp16)[name = tensor("op_2320_cast_fp16")]; + tensor input_397_cast_fp16 = add(x = var_2320_cast_fp16, y = var_2302_cast_fp16_0)[name = tensor("input_397_cast_fp16")]; + tensor linear_152_cast_fp16 = linear(bias = flow_net_res_blocks_5_mlp_0_bias_to_fp16, weight = flow_net_res_blocks_5_mlp_0_weight_to_fp16, x = input_397_cast_fp16)[name = tensor("linear_152_cast_fp16")]; + tensor input_401_cast_fp16 = silu(x = linear_152_cast_fp16)[name = tensor("input_401_cast_fp16")]; + tensor linear_153_cast_fp16 = linear(bias = flow_net_res_blocks_5_mlp_2_bias_to_fp16, weight = flow_net_res_blocks_5_mlp_2_weight_to_fp16, x = input_401_cast_fp16)[name = tensor("linear_153_cast_fp16")]; + tensor var_2331_cast_fp16 = mul(x = var_2302_cast_fp16_2, y = linear_153_cast_fp16)[name = tensor("op_2331_cast_fp16")]; + tensor x_261_cast_fp16 = add(x = x_255_cast_fp16, y = var_2331_cast_fp16)[name = tensor("x_261_cast_fp16")]; + tensor linear_154_cast_fp16 = linear(bias = flow_net_final_layer_adaLN_modulation_1_bias_to_fp16, weight = flow_net_final_layer_adaLN_modulation_1_weight_to_fp16, x = input_355_cast_fp16)[name = tensor("linear_154_cast_fp16")]; + tensor var_2340_split_sizes_0 = const()[name = tensor("op_2340_split_sizes_0"), val = tensor([512, 512])]; + tensor var_2340_axis_0 = const()[name = tensor("op_2340_axis_0"), val = tensor(-1)]; + tensor var_2340_cast_fp16_0, tensor var_2340_cast_fp16_1 = split(axis = var_2340_axis_0, split_sizes = var_2340_split_sizes_0, x = linear_154_cast_fp16)[name = tensor("op_2340_cast_fp16")]; + tensor mean_83_axes_0 = const()[name = tensor("mean_83_axes_0"), val = tensor([-1])]; + tensor mean_83_keep_dims_0 = const()[name = tensor("mean_83_keep_dims_0"), val = tensor(true)]; + tensor mean_83_cast_fp16 = reduce_mean(axes = mean_83_axes_0, keep_dims = mean_83_keep_dims_0, x = x_261_cast_fp16)[name = tensor("mean_83_cast_fp16")]; + tensor sub_65_cast_fp16 = sub(x = x_261_cast_fp16, y = mean_83_cast_fp16)[name = tensor("sub_65_cast_fp16")]; + tensor square_53_cast_fp16 = square(x = sub_65_cast_fp16)[name = tensor("square_53_cast_fp16")]; + tensor reduce_mean_107_axes_0 = const()[name = tensor("reduce_mean_107_axes_0"), val = tensor([-1])]; + tensor reduce_mean_107_keep_dims_0 = const()[name = tensor("reduce_mean_107_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_107_cast_fp16 = reduce_mean(axes = reduce_mean_107_axes_0, keep_dims = reduce_mean_107_keep_dims_0, x = square_53_cast_fp16)[name = tensor("reduce_mean_107_cast_fp16")]; + tensor var_2347_to_fp16 = const()[name = tensor("op_2347_to_fp16"), val = tensor(0x1.1p-20)]; + tensor var_2348_cast_fp16 = add(x = reduce_mean_107_cast_fp16, y = var_2347_to_fp16)[name = tensor("op_2348_cast_fp16")]; + tensor var_2349_cast_fp16 = sqrt(x = var_2348_cast_fp16)[name = tensor("op_2349_cast_fp16")]; + tensor x_263_cast_fp16 = real_div(x = sub_65_cast_fp16, y = var_2349_cast_fp16)[name = tensor("x_263_cast_fp16")]; + tensor var_2351_promoted_to_fp16 = const()[name = tensor("op_2351_promoted_to_fp16"), val = tensor(0x1p+0)]; + tensor var_2352_cast_fp16 = add(x = var_2340_cast_fp16_1, y = var_2351_promoted_to_fp16)[name = tensor("op_2352_cast_fp16")]; + tensor var_2353_cast_fp16 = mul(x = x_263_cast_fp16, y = var_2352_cast_fp16)[name = tensor("op_2353_cast_fp16")]; + tensor input_405_cast_fp16 = add(x = var_2353_cast_fp16, y = var_2340_cast_fp16_0)[name = tensor("input_405_cast_fp16")]; + tensor linear_155_cast_fp16 = linear(bias = flow_net_final_layer_linear_bias_to_fp16, weight = flow_net_final_layer_linear_weight_to_fp16, x = input_405_cast_fp16)[name = tensor("linear_155_cast_fp16")]; + tensor var_2364_to_fp16 = const()[name = tensor("op_2364_to_fp16"), val = tensor(0x1p-3)]; + tensor var_2365_cast_fp16 = mul(x = linear_155_cast_fp16, y = var_2364_to_fp16)[name = tensor("op_2365_cast_fp16")]; + tensor input_407_cast_fp16 = add(x = input_339_cast_fp16, y = var_2365_cast_fp16)[name = tensor("input_407_cast_fp16")]; + tensor linear_156_cast_fp16 = linear(bias = flow_net_input_proj_bias_to_fp16, weight = flow_net_input_proj_weight_to_fp16, x = input_407_cast_fp16)[name = tensor("linear_156_cast_fp16")]; + tensor input_411_to_fp16 = const()[name = tensor("input_411_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(19008256)))]; + tensor input_413_cast_fp16 = silu(x = input_411_to_fp16)[name = tensor("input_413_cast_fp16")]; + tensor linear_158_cast_fp16 = linear(bias = flow_net_time_embed_0_mlp_2_bias_to_fp16, weight = flow_net_time_embed_0_mlp_2_weight_to_fp16, x = input_413_cast_fp16)[name = tensor("linear_158_cast_fp16")]; + tensor reduce_mean_108_axes_0 = const()[name = tensor("reduce_mean_108_axes_0"), val = tensor([-1])]; + tensor reduce_mean_108_keep_dims_0 = const()[name = tensor("reduce_mean_108_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_108_cast_fp16 = reduce_mean(axes = reduce_mean_108_axes_0, keep_dims = reduce_mean_108_keep_dims_0, x = linear_158_cast_fp16)[name = tensor("reduce_mean_108_cast_fp16")]; + tensor sub_66_cast_fp16 = sub(x = linear_158_cast_fp16, y = reduce_mean_108_cast_fp16)[name = tensor("sub_66_cast_fp16")]; + tensor square_54_cast_fp16 = square(x = sub_66_cast_fp16)[name = tensor("square_54_cast_fp16")]; + tensor reduce_mean_109_axes_0 = const()[name = tensor("reduce_mean_109_axes_0"), val = tensor([-1])]; + tensor reduce_mean_109_keep_dims_0 = const()[name = tensor("reduce_mean_109_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_109_cast_fp16 = reduce_mean(axes = reduce_mean_109_axes_0, keep_dims = reduce_mean_109_keep_dims_0, x = square_54_cast_fp16)[name = tensor("reduce_mean_109_cast_fp16")]; + tensor real_div_12_to_fp16 = const()[name = tensor("real_div_12_to_fp16"), val = tensor(0x1.008p+0)]; + tensor mul_12_cast_fp16 = mul(x = reduce_mean_109_cast_fp16, y = real_div_12_to_fp16)[name = tensor("mul_12_cast_fp16")]; + tensor var_2431_to_fp16 = const()[name = tensor("op_2431_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_109_cast_fp16 = add(x = mul_12_cast_fp16, y = var_2431_to_fp16)[name = tensor("var_109_cast_fp16")]; + tensor var_2434_epsilon_0 = const()[name = tensor("op_2434_epsilon_0"), val = tensor(0x1.197998p-40)]; + tensor var_2434_cast_fp16 = rsqrt(epsilon = var_2434_epsilon_0, x = var_109_cast_fp16)[name = tensor("op_2434_cast_fp16")]; + tensor var_2435_cast_fp16 = mul(x = const_3_to_fp16, y = var_2434_cast_fp16)[name = tensor("op_2435_cast_fp16")]; + tensor var_2436_cast_fp16 = mul(x = linear_158_cast_fp16, y = var_2435_cast_fp16)[name = tensor("op_2436_cast_fp16")]; + tensor input_417_to_fp16 = const()[name = tensor("input_417_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(19009344)))]; + tensor input_419_cast_fp16 = silu(x = input_417_to_fp16)[name = tensor("input_419_cast_fp16")]; + tensor linear_160_cast_fp16 = linear(bias = flow_net_time_embed_1_mlp_2_bias_to_fp16, weight = flow_net_time_embed_1_mlp_2_weight_to_fp16, x = input_419_cast_fp16)[name = tensor("linear_160_cast_fp16")]; + tensor reduce_mean_110_axes_0 = const()[name = tensor("reduce_mean_110_axes_0"), val = tensor([-1])]; + tensor reduce_mean_110_keep_dims_0 = const()[name = tensor("reduce_mean_110_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_110_cast_fp16 = reduce_mean(axes = reduce_mean_110_axes_0, keep_dims = reduce_mean_110_keep_dims_0, x = linear_160_cast_fp16)[name = tensor("reduce_mean_110_cast_fp16")]; + tensor sub_68_cast_fp16 = sub(x = linear_160_cast_fp16, y = reduce_mean_110_cast_fp16)[name = tensor("sub_68_cast_fp16")]; + tensor square_55_cast_fp16 = square(x = sub_68_cast_fp16)[name = tensor("square_55_cast_fp16")]; + tensor reduce_mean_111_axes_0 = const()[name = tensor("reduce_mean_111_axes_0"), val = tensor([-1])]; + tensor reduce_mean_111_keep_dims_0 = const()[name = tensor("reduce_mean_111_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_111_cast_fp16 = reduce_mean(axes = reduce_mean_111_axes_0, keep_dims = reduce_mean_111_keep_dims_0, x = square_55_cast_fp16)[name = tensor("reduce_mean_111_cast_fp16")]; + tensor real_div_13_to_fp16 = const()[name = tensor("real_div_13_to_fp16"), val = tensor(0x1.008p+0)]; + tensor mul_13_cast_fp16 = mul(x = reduce_mean_111_cast_fp16, y = real_div_13_to_fp16)[name = tensor("mul_13_cast_fp16")]; + tensor var_2468_to_fp16 = const()[name = tensor("op_2468_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_111_cast_fp16 = add(x = mul_13_cast_fp16, y = var_2468_to_fp16)[name = tensor("var_111_cast_fp16")]; + tensor var_2471_epsilon_0 = const()[name = tensor("op_2471_epsilon_0"), val = tensor(0x1.197998p-40)]; + tensor var_2471_cast_fp16 = rsqrt(epsilon = var_2471_epsilon_0, x = var_111_cast_fp16)[name = tensor("op_2471_cast_fp16")]; + tensor var_2472_cast_fp16 = mul(x = const_5_to_fp16, y = var_2471_cast_fp16)[name = tensor("op_2472_cast_fp16")]; + tensor var_2473_cast_fp16 = mul(x = linear_160_cast_fp16, y = var_2472_cast_fp16)[name = tensor("op_2473_cast_fp16")]; + tensor var_2485_cast_fp16 = add(x = var_2436_cast_fp16, y = var_2473_cast_fp16)[name = tensor("op_2485_cast_fp16")]; + tensor _inversed_t_combined_13_y_0_to_fp16 = const()[name = tensor("_inversed_t_combined_13_y_0_to_fp16"), val = tensor(0x1p-1)]; + tensor _inversed_t_combined_13_cast_fp16 = mul(x = var_2485_cast_fp16, y = _inversed_t_combined_13_y_0_to_fp16)[name = tensor("_inversed_t_combined_13_cast_fp16")]; + tensor input_421_cast_fp16 = add(x = _inversed_t_combined_13_cast_fp16, y = linear_5_cast_fp16)[name = tensor("input_421_cast_fp16")]; + tensor input_423_cast_fp16 = silu(x = input_421_cast_fp16)[name = tensor("input_423_cast_fp16")]; + tensor linear_162_cast_fp16 = linear(bias = flow_net_res_blocks_0_adaLN_modulation_1_bias_to_fp16, weight = flow_net_res_blocks_0_adaLN_modulation_1_weight_to_fp16, x = input_423_cast_fp16)[name = tensor("linear_162_cast_fp16")]; + tensor var_2500_split_sizes_0 = const()[name = tensor("op_2500_split_sizes_0"), val = tensor([512, 512, 512])]; + tensor var_2500_axis_0 = const()[name = tensor("op_2500_axis_0"), val = tensor(-1)]; + tensor var_2500_cast_fp16_0, tensor var_2500_cast_fp16_1, tensor var_2500_cast_fp16_2 = split(axis = var_2500_axis_0, split_sizes = var_2500_split_sizes_0, x = linear_162_cast_fp16)[name = tensor("op_2500_cast_fp16")]; + tensor mean_85_axes_0 = const()[name = tensor("mean_85_axes_0"), val = tensor([-1])]; + tensor mean_85_keep_dims_0 = const()[name = tensor("mean_85_keep_dims_0"), val = tensor(true)]; + tensor mean_85_cast_fp16 = reduce_mean(axes = mean_85_axes_0, keep_dims = mean_85_keep_dims_0, x = linear_156_cast_fp16)[name = tensor("mean_85_cast_fp16")]; + tensor sub_70_cast_fp16 = sub(x = linear_156_cast_fp16, y = mean_85_cast_fp16)[name = tensor("sub_70_cast_fp16")]; + tensor square_56_cast_fp16 = square(x = sub_70_cast_fp16)[name = tensor("square_56_cast_fp16")]; + tensor reduce_mean_113_axes_0 = const()[name = tensor("reduce_mean_113_axes_0"), val = tensor([-1])]; + tensor reduce_mean_113_keep_dims_0 = const()[name = tensor("reduce_mean_113_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_113_cast_fp16 = reduce_mean(axes = reduce_mean_113_axes_0, keep_dims = reduce_mean_113_keep_dims_0, x = square_56_cast_fp16)[name = tensor("reduce_mean_113_cast_fp16")]; + tensor var_2510_to_fp16 = const()[name = tensor("op_2510_to_fp16"), val = tensor(0x1.1p-20)]; + tensor var_2511_cast_fp16 = add(x = reduce_mean_113_cast_fp16, y = var_2510_to_fp16)[name = tensor("op_2511_cast_fp16")]; + tensor var_2512_cast_fp16 = sqrt(x = var_2511_cast_fp16)[name = tensor("op_2512_cast_fp16")]; + tensor x_271_cast_fp16 = real_div(x = sub_70_cast_fp16, y = var_2512_cast_fp16)[name = tensor("x_271_cast_fp16")]; + tensor var_2514_cast_fp16 = mul(x = x_271_cast_fp16, y = flow_net_res_blocks_0_in_ln_weight_to_fp16)[name = tensor("op_2514_cast_fp16")]; + tensor x_273_cast_fp16 = add(x = var_2514_cast_fp16, y = flow_net_res_blocks_0_in_ln_bias_to_fp16)[name = tensor("x_273_cast_fp16")]; + tensor var_2516_promoted_to_fp16 = const()[name = tensor("op_2516_promoted_to_fp16"), val = tensor(0x1p+0)]; + tensor var_2517_cast_fp16 = add(x = var_2500_cast_fp16_1, y = var_2516_promoted_to_fp16)[name = tensor("op_2517_cast_fp16")]; + tensor var_2518_cast_fp16 = mul(x = x_273_cast_fp16, y = var_2517_cast_fp16)[name = tensor("op_2518_cast_fp16")]; + tensor input_425_cast_fp16 = add(x = var_2518_cast_fp16, y = var_2500_cast_fp16_0)[name = tensor("input_425_cast_fp16")]; + tensor linear_163_cast_fp16 = linear(bias = flow_net_res_blocks_0_mlp_0_bias_to_fp16, weight = flow_net_res_blocks_0_mlp_0_weight_to_fp16, x = input_425_cast_fp16)[name = tensor("linear_163_cast_fp16")]; + tensor input_429_cast_fp16 = silu(x = linear_163_cast_fp16)[name = tensor("input_429_cast_fp16")]; + tensor linear_164_cast_fp16 = linear(bias = flow_net_res_blocks_0_mlp_2_bias_to_fp16, weight = flow_net_res_blocks_0_mlp_2_weight_to_fp16, x = input_429_cast_fp16)[name = tensor("linear_164_cast_fp16")]; + tensor var_2529_cast_fp16 = mul(x = var_2500_cast_fp16_2, y = linear_164_cast_fp16)[name = tensor("op_2529_cast_fp16")]; + tensor x_275_cast_fp16 = add(x = linear_156_cast_fp16, y = var_2529_cast_fp16)[name = tensor("x_275_cast_fp16")]; + tensor linear_165_cast_fp16 = linear(bias = flow_net_res_blocks_1_adaLN_modulation_1_bias_to_fp16, weight = flow_net_res_blocks_1_adaLN_modulation_1_weight_to_fp16, x = input_423_cast_fp16)[name = tensor("linear_165_cast_fp16")]; + tensor var_2539_split_sizes_0 = const()[name = tensor("op_2539_split_sizes_0"), val = tensor([512, 512, 512])]; + tensor var_2539_axis_0 = const()[name = tensor("op_2539_axis_0"), val = tensor(-1)]; + tensor var_2539_cast_fp16_0, tensor var_2539_cast_fp16_1, tensor var_2539_cast_fp16_2 = split(axis = var_2539_axis_0, split_sizes = var_2539_split_sizes_0, x = linear_165_cast_fp16)[name = tensor("op_2539_cast_fp16")]; + tensor mean_87_axes_0 = const()[name = tensor("mean_87_axes_0"), val = tensor([-1])]; + tensor mean_87_keep_dims_0 = const()[name = tensor("mean_87_keep_dims_0"), val = tensor(true)]; + tensor mean_87_cast_fp16 = reduce_mean(axes = mean_87_axes_0, keep_dims = mean_87_keep_dims_0, x = x_275_cast_fp16)[name = tensor("mean_87_cast_fp16")]; + tensor sub_71_cast_fp16 = sub(x = x_275_cast_fp16, y = mean_87_cast_fp16)[name = tensor("sub_71_cast_fp16")]; + tensor square_57_cast_fp16 = square(x = sub_71_cast_fp16)[name = tensor("square_57_cast_fp16")]; + tensor reduce_mean_115_axes_0 = const()[name = tensor("reduce_mean_115_axes_0"), val = tensor([-1])]; + tensor reduce_mean_115_keep_dims_0 = const()[name = tensor("reduce_mean_115_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_115_cast_fp16 = reduce_mean(axes = reduce_mean_115_axes_0, keep_dims = reduce_mean_115_keep_dims_0, x = square_57_cast_fp16)[name = tensor("reduce_mean_115_cast_fp16")]; + tensor var_2549_to_fp16 = const()[name = tensor("op_2549_to_fp16"), val = tensor(0x1.1p-20)]; + tensor var_2550_cast_fp16 = add(x = reduce_mean_115_cast_fp16, y = var_2549_to_fp16)[name = tensor("op_2550_cast_fp16")]; + tensor var_2551_cast_fp16 = sqrt(x = var_2550_cast_fp16)[name = tensor("op_2551_cast_fp16")]; + tensor x_277_cast_fp16 = real_div(x = sub_71_cast_fp16, y = var_2551_cast_fp16)[name = tensor("x_277_cast_fp16")]; + tensor var_2553_cast_fp16 = mul(x = x_277_cast_fp16, y = flow_net_res_blocks_1_in_ln_weight_to_fp16)[name = tensor("op_2553_cast_fp16")]; + tensor x_279_cast_fp16 = add(x = var_2553_cast_fp16, y = flow_net_res_blocks_1_in_ln_bias_to_fp16)[name = tensor("x_279_cast_fp16")]; + tensor var_2555_promoted_to_fp16 = const()[name = tensor("op_2555_promoted_to_fp16"), val = tensor(0x1p+0)]; + tensor var_2556_cast_fp16 = add(x = var_2539_cast_fp16_1, y = var_2555_promoted_to_fp16)[name = tensor("op_2556_cast_fp16")]; + tensor var_2557_cast_fp16 = mul(x = x_279_cast_fp16, y = var_2556_cast_fp16)[name = tensor("op_2557_cast_fp16")]; + tensor input_433_cast_fp16 = add(x = var_2557_cast_fp16, y = var_2539_cast_fp16_0)[name = tensor("input_433_cast_fp16")]; + tensor linear_166_cast_fp16 = linear(bias = flow_net_res_blocks_1_mlp_0_bias_to_fp16, weight = flow_net_res_blocks_1_mlp_0_weight_to_fp16, x = input_433_cast_fp16)[name = tensor("linear_166_cast_fp16")]; + tensor input_437_cast_fp16 = silu(x = linear_166_cast_fp16)[name = tensor("input_437_cast_fp16")]; + tensor linear_167_cast_fp16 = linear(bias = flow_net_res_blocks_1_mlp_2_bias_to_fp16, weight = flow_net_res_blocks_1_mlp_2_weight_to_fp16, x = input_437_cast_fp16)[name = tensor("linear_167_cast_fp16")]; + tensor var_2568_cast_fp16 = mul(x = var_2539_cast_fp16_2, y = linear_167_cast_fp16)[name = tensor("op_2568_cast_fp16")]; + tensor x_281_cast_fp16 = add(x = x_275_cast_fp16, y = var_2568_cast_fp16)[name = tensor("x_281_cast_fp16")]; + tensor linear_168_cast_fp16 = linear(bias = flow_net_res_blocks_2_adaLN_modulation_1_bias_to_fp16, weight = flow_net_res_blocks_2_adaLN_modulation_1_weight_to_fp16, x = input_423_cast_fp16)[name = tensor("linear_168_cast_fp16")]; + tensor var_2578_split_sizes_0 = const()[name = tensor("op_2578_split_sizes_0"), val = tensor([512, 512, 512])]; + tensor var_2578_axis_0 = const()[name = tensor("op_2578_axis_0"), val = tensor(-1)]; + tensor var_2578_cast_fp16_0, tensor var_2578_cast_fp16_1, tensor var_2578_cast_fp16_2 = split(axis = var_2578_axis_0, split_sizes = var_2578_split_sizes_0, x = linear_168_cast_fp16)[name = tensor("op_2578_cast_fp16")]; + tensor mean_89_axes_0 = const()[name = tensor("mean_89_axes_0"), val = tensor([-1])]; + tensor mean_89_keep_dims_0 = const()[name = tensor("mean_89_keep_dims_0"), val = tensor(true)]; + tensor mean_89_cast_fp16 = reduce_mean(axes = mean_89_axes_0, keep_dims = mean_89_keep_dims_0, x = x_281_cast_fp16)[name = tensor("mean_89_cast_fp16")]; + tensor sub_72_cast_fp16 = sub(x = x_281_cast_fp16, y = mean_89_cast_fp16)[name = tensor("sub_72_cast_fp16")]; + tensor square_58_cast_fp16 = square(x = sub_72_cast_fp16)[name = tensor("square_58_cast_fp16")]; + tensor reduce_mean_117_axes_0 = const()[name = tensor("reduce_mean_117_axes_0"), val = tensor([-1])]; + tensor reduce_mean_117_keep_dims_0 = const()[name = tensor("reduce_mean_117_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_117_cast_fp16 = reduce_mean(axes = reduce_mean_117_axes_0, keep_dims = reduce_mean_117_keep_dims_0, x = square_58_cast_fp16)[name = tensor("reduce_mean_117_cast_fp16")]; + tensor var_2588_to_fp16 = const()[name = tensor("op_2588_to_fp16"), val = tensor(0x1.1p-20)]; + tensor var_2589_cast_fp16 = add(x = reduce_mean_117_cast_fp16, y = var_2588_to_fp16)[name = tensor("op_2589_cast_fp16")]; + tensor var_2590_cast_fp16 = sqrt(x = var_2589_cast_fp16)[name = tensor("op_2590_cast_fp16")]; + tensor x_283_cast_fp16 = real_div(x = sub_72_cast_fp16, y = var_2590_cast_fp16)[name = tensor("x_283_cast_fp16")]; + tensor var_2592_cast_fp16 = mul(x = x_283_cast_fp16, y = flow_net_res_blocks_2_in_ln_weight_to_fp16)[name = tensor("op_2592_cast_fp16")]; + tensor x_285_cast_fp16 = add(x = var_2592_cast_fp16, y = flow_net_res_blocks_2_in_ln_bias_to_fp16)[name = tensor("x_285_cast_fp16")]; + tensor var_2594_promoted_to_fp16 = const()[name = tensor("op_2594_promoted_to_fp16"), val = tensor(0x1p+0)]; + tensor var_2595_cast_fp16 = add(x = var_2578_cast_fp16_1, y = var_2594_promoted_to_fp16)[name = tensor("op_2595_cast_fp16")]; + tensor var_2596_cast_fp16 = mul(x = x_285_cast_fp16, y = var_2595_cast_fp16)[name = tensor("op_2596_cast_fp16")]; + tensor input_441_cast_fp16 = add(x = var_2596_cast_fp16, y = var_2578_cast_fp16_0)[name = tensor("input_441_cast_fp16")]; + tensor linear_169_cast_fp16 = linear(bias = flow_net_res_blocks_2_mlp_0_bias_to_fp16, weight = flow_net_res_blocks_2_mlp_0_weight_to_fp16, x = input_441_cast_fp16)[name = tensor("linear_169_cast_fp16")]; + tensor input_445_cast_fp16 = silu(x = linear_169_cast_fp16)[name = tensor("input_445_cast_fp16")]; + tensor linear_170_cast_fp16 = linear(bias = flow_net_res_blocks_2_mlp_2_bias_to_fp16, weight = flow_net_res_blocks_2_mlp_2_weight_to_fp16, x = input_445_cast_fp16)[name = tensor("linear_170_cast_fp16")]; + tensor var_2607_cast_fp16 = mul(x = var_2578_cast_fp16_2, y = linear_170_cast_fp16)[name = tensor("op_2607_cast_fp16")]; + tensor x_287_cast_fp16 = add(x = x_281_cast_fp16, y = var_2607_cast_fp16)[name = tensor("x_287_cast_fp16")]; + tensor linear_171_cast_fp16 = linear(bias = flow_net_res_blocks_3_adaLN_modulation_1_bias_to_fp16, weight = flow_net_res_blocks_3_adaLN_modulation_1_weight_to_fp16, x = input_423_cast_fp16)[name = tensor("linear_171_cast_fp16")]; + tensor var_2617_split_sizes_0 = const()[name = tensor("op_2617_split_sizes_0"), val = tensor([512, 512, 512])]; + tensor var_2617_axis_0 = const()[name = tensor("op_2617_axis_0"), val = tensor(-1)]; + tensor var_2617_cast_fp16_0, tensor var_2617_cast_fp16_1, tensor var_2617_cast_fp16_2 = split(axis = var_2617_axis_0, split_sizes = var_2617_split_sizes_0, x = linear_171_cast_fp16)[name = tensor("op_2617_cast_fp16")]; + tensor mean_91_axes_0 = const()[name = tensor("mean_91_axes_0"), val = tensor([-1])]; + tensor mean_91_keep_dims_0 = const()[name = tensor("mean_91_keep_dims_0"), val = tensor(true)]; + tensor mean_91_cast_fp16 = reduce_mean(axes = mean_91_axes_0, keep_dims = mean_91_keep_dims_0, x = x_287_cast_fp16)[name = tensor("mean_91_cast_fp16")]; + tensor sub_73_cast_fp16 = sub(x = x_287_cast_fp16, y = mean_91_cast_fp16)[name = tensor("sub_73_cast_fp16")]; + tensor square_59_cast_fp16 = square(x = sub_73_cast_fp16)[name = tensor("square_59_cast_fp16")]; + tensor reduce_mean_119_axes_0 = const()[name = tensor("reduce_mean_119_axes_0"), val = tensor([-1])]; + tensor reduce_mean_119_keep_dims_0 = const()[name = tensor("reduce_mean_119_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_119_cast_fp16 = reduce_mean(axes = reduce_mean_119_axes_0, keep_dims = reduce_mean_119_keep_dims_0, x = square_59_cast_fp16)[name = tensor("reduce_mean_119_cast_fp16")]; + tensor var_2627_to_fp16 = const()[name = tensor("op_2627_to_fp16"), val = tensor(0x1.1p-20)]; + tensor var_2628_cast_fp16 = add(x = reduce_mean_119_cast_fp16, y = var_2627_to_fp16)[name = tensor("op_2628_cast_fp16")]; + tensor var_2629_cast_fp16 = sqrt(x = var_2628_cast_fp16)[name = tensor("op_2629_cast_fp16")]; + tensor x_289_cast_fp16 = real_div(x = sub_73_cast_fp16, y = var_2629_cast_fp16)[name = tensor("x_289_cast_fp16")]; + tensor var_2631_cast_fp16 = mul(x = x_289_cast_fp16, y = flow_net_res_blocks_3_in_ln_weight_to_fp16)[name = tensor("op_2631_cast_fp16")]; + tensor x_291_cast_fp16 = add(x = var_2631_cast_fp16, y = flow_net_res_blocks_3_in_ln_bias_to_fp16)[name = tensor("x_291_cast_fp16")]; + tensor var_2633_promoted_to_fp16 = const()[name = tensor("op_2633_promoted_to_fp16"), val = tensor(0x1p+0)]; + tensor var_2634_cast_fp16 = add(x = var_2617_cast_fp16_1, y = var_2633_promoted_to_fp16)[name = tensor("op_2634_cast_fp16")]; + tensor var_2635_cast_fp16 = mul(x = x_291_cast_fp16, y = var_2634_cast_fp16)[name = tensor("op_2635_cast_fp16")]; + tensor input_449_cast_fp16 = add(x = var_2635_cast_fp16, y = var_2617_cast_fp16_0)[name = tensor("input_449_cast_fp16")]; + tensor linear_172_cast_fp16 = linear(bias = flow_net_res_blocks_3_mlp_0_bias_to_fp16, weight = flow_net_res_blocks_3_mlp_0_weight_to_fp16, x = input_449_cast_fp16)[name = tensor("linear_172_cast_fp16")]; + tensor input_453_cast_fp16 = silu(x = linear_172_cast_fp16)[name = tensor("input_453_cast_fp16")]; + tensor linear_173_cast_fp16 = linear(bias = flow_net_res_blocks_3_mlp_2_bias_to_fp16, weight = flow_net_res_blocks_3_mlp_2_weight_to_fp16, x = input_453_cast_fp16)[name = tensor("linear_173_cast_fp16")]; + tensor var_2646_cast_fp16 = mul(x = var_2617_cast_fp16_2, y = linear_173_cast_fp16)[name = tensor("op_2646_cast_fp16")]; + tensor x_293_cast_fp16 = add(x = x_287_cast_fp16, y = var_2646_cast_fp16)[name = tensor("x_293_cast_fp16")]; + tensor linear_174_cast_fp16 = linear(bias = flow_net_res_blocks_4_adaLN_modulation_1_bias_to_fp16, weight = flow_net_res_blocks_4_adaLN_modulation_1_weight_to_fp16, x = input_423_cast_fp16)[name = tensor("linear_174_cast_fp16")]; + tensor var_2656_split_sizes_0 = const()[name = tensor("op_2656_split_sizes_0"), val = tensor([512, 512, 512])]; + tensor var_2656_axis_0 = const()[name = tensor("op_2656_axis_0"), val = tensor(-1)]; + tensor var_2656_cast_fp16_0, tensor var_2656_cast_fp16_1, tensor var_2656_cast_fp16_2 = split(axis = var_2656_axis_0, split_sizes = var_2656_split_sizes_0, x = linear_174_cast_fp16)[name = tensor("op_2656_cast_fp16")]; + tensor mean_93_axes_0 = const()[name = tensor("mean_93_axes_0"), val = tensor([-1])]; + tensor mean_93_keep_dims_0 = const()[name = tensor("mean_93_keep_dims_0"), val = tensor(true)]; + tensor mean_93_cast_fp16 = reduce_mean(axes = mean_93_axes_0, keep_dims = mean_93_keep_dims_0, x = x_293_cast_fp16)[name = tensor("mean_93_cast_fp16")]; + tensor sub_74_cast_fp16 = sub(x = x_293_cast_fp16, y = mean_93_cast_fp16)[name = tensor("sub_74_cast_fp16")]; + tensor square_60_cast_fp16 = square(x = sub_74_cast_fp16)[name = tensor("square_60_cast_fp16")]; + tensor reduce_mean_121_axes_0 = const()[name = tensor("reduce_mean_121_axes_0"), val = tensor([-1])]; + tensor reduce_mean_121_keep_dims_0 = const()[name = tensor("reduce_mean_121_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_121_cast_fp16 = reduce_mean(axes = reduce_mean_121_axes_0, keep_dims = reduce_mean_121_keep_dims_0, x = square_60_cast_fp16)[name = tensor("reduce_mean_121_cast_fp16")]; + tensor var_2666_to_fp16 = const()[name = tensor("op_2666_to_fp16"), val = tensor(0x1.1p-20)]; + tensor var_2667_cast_fp16 = add(x = reduce_mean_121_cast_fp16, y = var_2666_to_fp16)[name = tensor("op_2667_cast_fp16")]; + tensor var_2668_cast_fp16 = sqrt(x = var_2667_cast_fp16)[name = tensor("op_2668_cast_fp16")]; + tensor x_295_cast_fp16 = real_div(x = sub_74_cast_fp16, y = var_2668_cast_fp16)[name = tensor("x_295_cast_fp16")]; + tensor var_2670_cast_fp16 = mul(x = x_295_cast_fp16, y = flow_net_res_blocks_4_in_ln_weight_to_fp16)[name = tensor("op_2670_cast_fp16")]; + tensor x_297_cast_fp16 = add(x = var_2670_cast_fp16, y = flow_net_res_blocks_4_in_ln_bias_to_fp16)[name = tensor("x_297_cast_fp16")]; + tensor var_2672_promoted_to_fp16 = const()[name = tensor("op_2672_promoted_to_fp16"), val = tensor(0x1p+0)]; + tensor var_2673_cast_fp16 = add(x = var_2656_cast_fp16_1, y = var_2672_promoted_to_fp16)[name = tensor("op_2673_cast_fp16")]; + tensor var_2674_cast_fp16 = mul(x = x_297_cast_fp16, y = var_2673_cast_fp16)[name = tensor("op_2674_cast_fp16")]; + tensor input_457_cast_fp16 = add(x = var_2674_cast_fp16, y = var_2656_cast_fp16_0)[name = tensor("input_457_cast_fp16")]; + tensor linear_175_cast_fp16 = linear(bias = flow_net_res_blocks_4_mlp_0_bias_to_fp16, weight = flow_net_res_blocks_4_mlp_0_weight_to_fp16, x = input_457_cast_fp16)[name = tensor("linear_175_cast_fp16")]; + tensor input_461_cast_fp16 = silu(x = linear_175_cast_fp16)[name = tensor("input_461_cast_fp16")]; + tensor linear_176_cast_fp16 = linear(bias = flow_net_res_blocks_4_mlp_2_bias_to_fp16, weight = flow_net_res_blocks_4_mlp_2_weight_to_fp16, x = input_461_cast_fp16)[name = tensor("linear_176_cast_fp16")]; + tensor var_2685_cast_fp16 = mul(x = var_2656_cast_fp16_2, y = linear_176_cast_fp16)[name = tensor("op_2685_cast_fp16")]; + tensor x_299_cast_fp16 = add(x = x_293_cast_fp16, y = var_2685_cast_fp16)[name = tensor("x_299_cast_fp16")]; + tensor linear_177_cast_fp16 = linear(bias = flow_net_res_blocks_5_adaLN_modulation_1_bias_to_fp16, weight = flow_net_res_blocks_5_adaLN_modulation_1_weight_to_fp16, x = input_423_cast_fp16)[name = tensor("linear_177_cast_fp16")]; + tensor var_2695_split_sizes_0 = const()[name = tensor("op_2695_split_sizes_0"), val = tensor([512, 512, 512])]; + tensor var_2695_axis_0 = const()[name = tensor("op_2695_axis_0"), val = tensor(-1)]; + tensor var_2695_cast_fp16_0, tensor var_2695_cast_fp16_1, tensor var_2695_cast_fp16_2 = split(axis = var_2695_axis_0, split_sizes = var_2695_split_sizes_0, x = linear_177_cast_fp16)[name = tensor("op_2695_cast_fp16")]; + tensor mean_95_axes_0 = const()[name = tensor("mean_95_axes_0"), val = tensor([-1])]; + tensor mean_95_keep_dims_0 = const()[name = tensor("mean_95_keep_dims_0"), val = tensor(true)]; + tensor mean_95_cast_fp16 = reduce_mean(axes = mean_95_axes_0, keep_dims = mean_95_keep_dims_0, x = x_299_cast_fp16)[name = tensor("mean_95_cast_fp16")]; + tensor sub_75_cast_fp16 = sub(x = x_299_cast_fp16, y = mean_95_cast_fp16)[name = tensor("sub_75_cast_fp16")]; + tensor square_61_cast_fp16 = square(x = sub_75_cast_fp16)[name = tensor("square_61_cast_fp16")]; + tensor reduce_mean_123_axes_0 = const()[name = tensor("reduce_mean_123_axes_0"), val = tensor([-1])]; + tensor reduce_mean_123_keep_dims_0 = const()[name = tensor("reduce_mean_123_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_123_cast_fp16 = reduce_mean(axes = reduce_mean_123_axes_0, keep_dims = reduce_mean_123_keep_dims_0, x = square_61_cast_fp16)[name = tensor("reduce_mean_123_cast_fp16")]; + tensor var_2705_to_fp16 = const()[name = tensor("op_2705_to_fp16"), val = tensor(0x1.1p-20)]; + tensor var_2706_cast_fp16 = add(x = reduce_mean_123_cast_fp16, y = var_2705_to_fp16)[name = tensor("op_2706_cast_fp16")]; + tensor var_2707_cast_fp16 = sqrt(x = var_2706_cast_fp16)[name = tensor("op_2707_cast_fp16")]; + tensor x_301_cast_fp16 = real_div(x = sub_75_cast_fp16, y = var_2707_cast_fp16)[name = tensor("x_301_cast_fp16")]; + tensor var_2709_cast_fp16 = mul(x = x_301_cast_fp16, y = flow_net_res_blocks_5_in_ln_weight_to_fp16)[name = tensor("op_2709_cast_fp16")]; + tensor x_303_cast_fp16 = add(x = var_2709_cast_fp16, y = flow_net_res_blocks_5_in_ln_bias_to_fp16)[name = tensor("x_303_cast_fp16")]; + tensor var_2711_promoted_to_fp16 = const()[name = tensor("op_2711_promoted_to_fp16"), val = tensor(0x1p+0)]; + tensor var_2712_cast_fp16 = add(x = var_2695_cast_fp16_1, y = var_2711_promoted_to_fp16)[name = tensor("op_2712_cast_fp16")]; + tensor var_2713_cast_fp16 = mul(x = x_303_cast_fp16, y = var_2712_cast_fp16)[name = tensor("op_2713_cast_fp16")]; + tensor input_465_cast_fp16 = add(x = var_2713_cast_fp16, y = var_2695_cast_fp16_0)[name = tensor("input_465_cast_fp16")]; + tensor linear_178_cast_fp16 = linear(bias = flow_net_res_blocks_5_mlp_0_bias_to_fp16, weight = flow_net_res_blocks_5_mlp_0_weight_to_fp16, x = input_465_cast_fp16)[name = tensor("linear_178_cast_fp16")]; + tensor input_469_cast_fp16 = silu(x = linear_178_cast_fp16)[name = tensor("input_469_cast_fp16")]; + tensor linear_179_cast_fp16 = linear(bias = flow_net_res_blocks_5_mlp_2_bias_to_fp16, weight = flow_net_res_blocks_5_mlp_2_weight_to_fp16, x = input_469_cast_fp16)[name = tensor("linear_179_cast_fp16")]; + tensor var_2724_cast_fp16 = mul(x = var_2695_cast_fp16_2, y = linear_179_cast_fp16)[name = tensor("op_2724_cast_fp16")]; + tensor x_305_cast_fp16 = add(x = x_299_cast_fp16, y = var_2724_cast_fp16)[name = tensor("x_305_cast_fp16")]; + tensor linear_180_cast_fp16 = linear(bias = flow_net_final_layer_adaLN_modulation_1_bias_to_fp16, weight = flow_net_final_layer_adaLN_modulation_1_weight_to_fp16, x = input_423_cast_fp16)[name = tensor("linear_180_cast_fp16")]; + tensor var_2733_split_sizes_0 = const()[name = tensor("op_2733_split_sizes_0"), val = tensor([512, 512])]; + tensor var_2733_axis_0 = const()[name = tensor("op_2733_axis_0"), val = tensor(-1)]; + tensor var_2733_cast_fp16_0, tensor var_2733_cast_fp16_1 = split(axis = var_2733_axis_0, split_sizes = var_2733_split_sizes_0, x = linear_180_cast_fp16)[name = tensor("op_2733_cast_fp16")]; + tensor mean_97_axes_0 = const()[name = tensor("mean_97_axes_0"), val = tensor([-1])]; + tensor mean_97_keep_dims_0 = const()[name = tensor("mean_97_keep_dims_0"), val = tensor(true)]; + tensor mean_97_cast_fp16 = reduce_mean(axes = mean_97_axes_0, keep_dims = mean_97_keep_dims_0, x = x_305_cast_fp16)[name = tensor("mean_97_cast_fp16")]; + tensor sub_76_cast_fp16 = sub(x = x_305_cast_fp16, y = mean_97_cast_fp16)[name = tensor("sub_76_cast_fp16")]; + tensor square_62_cast_fp16 = square(x = sub_76_cast_fp16)[name = tensor("square_62_cast_fp16")]; + tensor reduce_mean_125_axes_0 = const()[name = tensor("reduce_mean_125_axes_0"), val = tensor([-1])]; + tensor reduce_mean_125_keep_dims_0 = const()[name = tensor("reduce_mean_125_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_125_cast_fp16 = reduce_mean(axes = reduce_mean_125_axes_0, keep_dims = reduce_mean_125_keep_dims_0, x = square_62_cast_fp16)[name = tensor("reduce_mean_125_cast_fp16")]; + tensor var_2740_to_fp16 = const()[name = tensor("op_2740_to_fp16"), val = tensor(0x1.1p-20)]; + tensor var_2741_cast_fp16 = add(x = reduce_mean_125_cast_fp16, y = var_2740_to_fp16)[name = tensor("op_2741_cast_fp16")]; + tensor var_2742_cast_fp16 = sqrt(x = var_2741_cast_fp16)[name = tensor("op_2742_cast_fp16")]; + tensor x_307_cast_fp16 = real_div(x = sub_76_cast_fp16, y = var_2742_cast_fp16)[name = tensor("x_307_cast_fp16")]; + tensor var_2744_promoted_to_fp16 = const()[name = tensor("op_2744_promoted_to_fp16"), val = tensor(0x1p+0)]; + tensor var_2745_cast_fp16 = add(x = var_2733_cast_fp16_1, y = var_2744_promoted_to_fp16)[name = tensor("op_2745_cast_fp16")]; + tensor var_2746_cast_fp16 = mul(x = x_307_cast_fp16, y = var_2745_cast_fp16)[name = tensor("op_2746_cast_fp16")]; + tensor input_473_cast_fp16 = add(x = var_2746_cast_fp16, y = var_2733_cast_fp16_0)[name = tensor("input_473_cast_fp16")]; + tensor linear_181_cast_fp16 = linear(bias = flow_net_final_layer_linear_bias_to_fp16, weight = flow_net_final_layer_linear_weight_to_fp16, x = input_473_cast_fp16)[name = tensor("linear_181_cast_fp16")]; + tensor var_2757_to_fp16 = const()[name = tensor("op_2757_to_fp16"), val = tensor(0x1p-3)]; + tensor var_2758_cast_fp16 = mul(x = linear_181_cast_fp16, y = var_2757_to_fp16)[name = tensor("op_2758_cast_fp16")]; + tensor input_475_cast_fp16 = add(x = input_407_cast_fp16, y = var_2758_cast_fp16)[name = tensor("input_475_cast_fp16")]; + tensor linear_182_cast_fp16 = linear(bias = flow_net_input_proj_bias_to_fp16, weight = flow_net_input_proj_weight_to_fp16, x = input_475_cast_fp16)[name = tensor("linear_182_cast_fp16")]; + tensor input_479_to_fp16 = const()[name = tensor("input_479_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(19010432)))]; + tensor input_481_cast_fp16 = silu(x = input_479_to_fp16)[name = tensor("input_481_cast_fp16")]; + tensor linear_184_cast_fp16 = linear(bias = flow_net_time_embed_0_mlp_2_bias_to_fp16, weight = flow_net_time_embed_0_mlp_2_weight_to_fp16, x = input_481_cast_fp16)[name = tensor("linear_184_cast_fp16")]; + tensor reduce_mean_126_axes_0 = const()[name = tensor("reduce_mean_126_axes_0"), val = tensor([-1])]; + tensor reduce_mean_126_keep_dims_0 = const()[name = tensor("reduce_mean_126_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_126_cast_fp16 = reduce_mean(axes = reduce_mean_126_axes_0, keep_dims = reduce_mean_126_keep_dims_0, x = linear_184_cast_fp16)[name = tensor("reduce_mean_126_cast_fp16")]; + tensor sub_77_cast_fp16 = sub(x = linear_184_cast_fp16, y = reduce_mean_126_cast_fp16)[name = tensor("sub_77_cast_fp16")]; + tensor square_63_cast_fp16 = square(x = sub_77_cast_fp16)[name = tensor("square_63_cast_fp16")]; + tensor reduce_mean_127_axes_0 = const()[name = tensor("reduce_mean_127_axes_0"), val = tensor([-1])]; + tensor reduce_mean_127_keep_dims_0 = const()[name = tensor("reduce_mean_127_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_127_cast_fp16 = reduce_mean(axes = reduce_mean_127_axes_0, keep_dims = reduce_mean_127_keep_dims_0, x = square_63_cast_fp16)[name = tensor("reduce_mean_127_cast_fp16")]; + tensor real_div_14_to_fp16 = const()[name = tensor("real_div_14_to_fp16"), val = tensor(0x1.008p+0)]; + tensor mul_14_cast_fp16 = mul(x = reduce_mean_127_cast_fp16, y = real_div_14_to_fp16)[name = tensor("mul_14_cast_fp16")]; + tensor var_2824_to_fp16 = const()[name = tensor("op_2824_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_127_cast_fp16_0 = add(x = mul_14_cast_fp16, y = var_2824_to_fp16)[name = tensor("var_127_cast_fp16")]; + tensor var_2827_epsilon_0 = const()[name = tensor("op_2827_epsilon_0"), val = tensor(0x1.197998p-40)]; + tensor var_2827_cast_fp16 = rsqrt(epsilon = var_2827_epsilon_0, x = var_127_cast_fp16_0)[name = tensor("op_2827_cast_fp16")]; + tensor var_2828_cast_fp16 = mul(x = const_3_to_fp16, y = var_2827_cast_fp16)[name = tensor("op_2828_cast_fp16")]; + tensor var_2829_cast_fp16 = mul(x = linear_184_cast_fp16, y = var_2828_cast_fp16)[name = tensor("op_2829_cast_fp16")]; + tensor input_485_to_fp16 = const()[name = tensor("input_485_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(19011520)))]; + tensor input_487_cast_fp16 = silu(x = input_485_to_fp16)[name = tensor("input_487_cast_fp16")]; + tensor linear_186_cast_fp16 = linear(bias = flow_net_time_embed_1_mlp_2_bias_to_fp16, weight = flow_net_time_embed_1_mlp_2_weight_to_fp16, x = input_487_cast_fp16)[name = tensor("linear_186_cast_fp16")]; + tensor reduce_mean_128_axes_0 = const()[name = tensor("reduce_mean_128_axes_0"), val = tensor([-1])]; + tensor reduce_mean_128_keep_dims_0 = const()[name = tensor("reduce_mean_128_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_128_cast_fp16 = reduce_mean(axes = reduce_mean_128_axes_0, keep_dims = reduce_mean_128_keep_dims_0, x = linear_186_cast_fp16)[name = tensor("reduce_mean_128_cast_fp16")]; + tensor sub_79_cast_fp16 = sub(x = linear_186_cast_fp16, y = reduce_mean_128_cast_fp16)[name = tensor("sub_79_cast_fp16")]; + tensor square_64_cast_fp16 = square(x = sub_79_cast_fp16)[name = tensor("square_64_cast_fp16")]; + tensor reduce_mean_129_axes_0 = const()[name = tensor("reduce_mean_129_axes_0"), val = tensor([-1])]; + tensor reduce_mean_129_keep_dims_0 = const()[name = tensor("reduce_mean_129_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_129_cast_fp16 = reduce_mean(axes = reduce_mean_129_axes_0, keep_dims = reduce_mean_129_keep_dims_0, x = square_64_cast_fp16)[name = tensor("reduce_mean_129_cast_fp16")]; + tensor real_div_15_to_fp16 = const()[name = tensor("real_div_15_to_fp16"), val = tensor(0x1.008p+0)]; + tensor mul_15_cast_fp16 = mul(x = reduce_mean_129_cast_fp16, y = real_div_15_to_fp16)[name = tensor("mul_15_cast_fp16")]; + tensor var_2851_to_fp16 = const()[name = tensor("op_2851_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_129_cast_fp16 = add(x = mul_15_cast_fp16, y = var_2851_to_fp16)[name = tensor("var_129_cast_fp16")]; + tensor var_2854_epsilon_0 = const()[name = tensor("op_2854_epsilon_0"), val = tensor(0x1.197998p-40)]; + tensor var_2854_cast_fp16 = rsqrt(epsilon = var_2854_epsilon_0, x = var_129_cast_fp16)[name = tensor("op_2854_cast_fp16")]; + tensor var_2855_cast_fp16 = mul(x = const_5_to_fp16, y = var_2854_cast_fp16)[name = tensor("op_2855_cast_fp16")]; + tensor var_2856_cast_fp16 = mul(x = linear_186_cast_fp16, y = var_2855_cast_fp16)[name = tensor("op_2856_cast_fp16")]; + tensor var_2858_cast_fp16 = add(x = var_2829_cast_fp16, y = var_2856_cast_fp16)[name = tensor("op_2858_cast_fp16")]; + tensor _inversed_t_combined_y_0_to_fp16 = const()[name = tensor("_inversed_t_combined_y_0_to_fp16"), val = tensor(0x1p-1)]; + tensor _inversed_t_combined_cast_fp16 = mul(x = var_2858_cast_fp16, y = _inversed_t_combined_y_0_to_fp16)[name = tensor("_inversed_t_combined_cast_fp16")]; + tensor input_489_cast_fp16 = add(x = _inversed_t_combined_cast_fp16, y = linear_5_cast_fp16)[name = tensor("input_489_cast_fp16")]; + tensor input_491_cast_fp16 = silu(x = input_489_cast_fp16)[name = tensor("input_491_cast_fp16")]; + tensor linear_188_cast_fp16 = linear(bias = flow_net_res_blocks_0_adaLN_modulation_1_bias_to_fp16, weight = flow_net_res_blocks_0_adaLN_modulation_1_weight_to_fp16, x = input_491_cast_fp16)[name = tensor("linear_188_cast_fp16")]; + tensor var_2873_split_sizes_0 = const()[name = tensor("op_2873_split_sizes_0"), val = tensor([512, 512, 512])]; + tensor var_2873_axis_0 = const()[name = tensor("op_2873_axis_0"), val = tensor(-1)]; + tensor var_2873_cast_fp16_0, tensor var_2873_cast_fp16_1, tensor var_2873_cast_fp16_2 = split(axis = var_2873_axis_0, split_sizes = var_2873_split_sizes_0, x = linear_188_cast_fp16)[name = tensor("op_2873_cast_fp16")]; + tensor mean_99_axes_0 = const()[name = tensor("mean_99_axes_0"), val = tensor([-1])]; + tensor mean_99_keep_dims_0 = const()[name = tensor("mean_99_keep_dims_0"), val = tensor(true)]; + tensor mean_99_cast_fp16 = reduce_mean(axes = mean_99_axes_0, keep_dims = mean_99_keep_dims_0, x = linear_182_cast_fp16)[name = tensor("mean_99_cast_fp16")]; + tensor sub_81_cast_fp16 = sub(x = linear_182_cast_fp16, y = mean_99_cast_fp16)[name = tensor("sub_81_cast_fp16")]; + tensor square_65_cast_fp16 = square(x = sub_81_cast_fp16)[name = tensor("square_65_cast_fp16")]; + tensor reduce_mean_131_axes_0 = const()[name = tensor("reduce_mean_131_axes_0"), val = tensor([-1])]; + tensor reduce_mean_131_keep_dims_0 = const()[name = tensor("reduce_mean_131_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_131_cast_fp16 = reduce_mean(axes = reduce_mean_131_axes_0, keep_dims = reduce_mean_131_keep_dims_0, x = square_65_cast_fp16)[name = tensor("reduce_mean_131_cast_fp16")]; + tensor var_2883_to_fp16 = const()[name = tensor("op_2883_to_fp16"), val = tensor(0x1.1p-20)]; + tensor var_2884_cast_fp16 = add(x = reduce_mean_131_cast_fp16, y = var_2883_to_fp16)[name = tensor("op_2884_cast_fp16")]; + tensor var_2885_cast_fp16 = sqrt(x = var_2884_cast_fp16)[name = tensor("op_2885_cast_fp16")]; + tensor x_315_cast_fp16 = real_div(x = sub_81_cast_fp16, y = var_2885_cast_fp16)[name = tensor("x_315_cast_fp16")]; + tensor var_2887_cast_fp16 = mul(x = x_315_cast_fp16, y = flow_net_res_blocks_0_in_ln_weight_to_fp16)[name = tensor("op_2887_cast_fp16")]; + tensor x_317_cast_fp16 = add(x = var_2887_cast_fp16, y = flow_net_res_blocks_0_in_ln_bias_to_fp16)[name = tensor("x_317_cast_fp16")]; + tensor var_2889_promoted_to_fp16 = const()[name = tensor("op_2889_promoted_to_fp16"), val = tensor(0x1p+0)]; + tensor var_2890_cast_fp16 = add(x = var_2873_cast_fp16_1, y = var_2889_promoted_to_fp16)[name = tensor("op_2890_cast_fp16")]; + tensor var_2891_cast_fp16 = mul(x = x_317_cast_fp16, y = var_2890_cast_fp16)[name = tensor("op_2891_cast_fp16")]; + tensor input_493_cast_fp16 = add(x = var_2891_cast_fp16, y = var_2873_cast_fp16_0)[name = tensor("input_493_cast_fp16")]; + tensor linear_189_cast_fp16 = linear(bias = flow_net_res_blocks_0_mlp_0_bias_to_fp16, weight = flow_net_res_blocks_0_mlp_0_weight_to_fp16, x = input_493_cast_fp16)[name = tensor("linear_189_cast_fp16")]; + tensor input_497_cast_fp16 = silu(x = linear_189_cast_fp16)[name = tensor("input_497_cast_fp16")]; + tensor linear_190_cast_fp16 = linear(bias = flow_net_res_blocks_0_mlp_2_bias_to_fp16, weight = flow_net_res_blocks_0_mlp_2_weight_to_fp16, x = input_497_cast_fp16)[name = tensor("linear_190_cast_fp16")]; + tensor var_2902_cast_fp16 = mul(x = var_2873_cast_fp16_2, y = linear_190_cast_fp16)[name = tensor("op_2902_cast_fp16")]; + tensor x_319_cast_fp16 = add(x = linear_182_cast_fp16, y = var_2902_cast_fp16)[name = tensor("x_319_cast_fp16")]; + tensor linear_191_cast_fp16 = linear(bias = flow_net_res_blocks_1_adaLN_modulation_1_bias_to_fp16, weight = flow_net_res_blocks_1_adaLN_modulation_1_weight_to_fp16, x = input_491_cast_fp16)[name = tensor("linear_191_cast_fp16")]; + tensor var_2912_split_sizes_0 = const()[name = tensor("op_2912_split_sizes_0"), val = tensor([512, 512, 512])]; + tensor var_2912_axis_0 = const()[name = tensor("op_2912_axis_0"), val = tensor(-1)]; + tensor var_2912_cast_fp16_0, tensor var_2912_cast_fp16_1, tensor var_2912_cast_fp16_2 = split(axis = var_2912_axis_0, split_sizes = var_2912_split_sizes_0, x = linear_191_cast_fp16)[name = tensor("op_2912_cast_fp16")]; + tensor mean_101_axes_0 = const()[name = tensor("mean_101_axes_0"), val = tensor([-1])]; + tensor mean_101_keep_dims_0 = const()[name = tensor("mean_101_keep_dims_0"), val = tensor(true)]; + tensor mean_101_cast_fp16 = reduce_mean(axes = mean_101_axes_0, keep_dims = mean_101_keep_dims_0, x = x_319_cast_fp16)[name = tensor("mean_101_cast_fp16")]; + tensor sub_82_cast_fp16 = sub(x = x_319_cast_fp16, y = mean_101_cast_fp16)[name = tensor("sub_82_cast_fp16")]; + tensor square_66_cast_fp16 = square(x = sub_82_cast_fp16)[name = tensor("square_66_cast_fp16")]; + tensor reduce_mean_133_axes_0 = const()[name = tensor("reduce_mean_133_axes_0"), val = tensor([-1])]; + tensor reduce_mean_133_keep_dims_0 = const()[name = tensor("reduce_mean_133_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_133_cast_fp16 = reduce_mean(axes = reduce_mean_133_axes_0, keep_dims = reduce_mean_133_keep_dims_0, x = square_66_cast_fp16)[name = tensor("reduce_mean_133_cast_fp16")]; + tensor var_2922_to_fp16 = const()[name = tensor("op_2922_to_fp16"), val = tensor(0x1.1p-20)]; + tensor var_2923_cast_fp16 = add(x = reduce_mean_133_cast_fp16, y = var_2922_to_fp16)[name = tensor("op_2923_cast_fp16")]; + tensor var_2924_cast_fp16 = sqrt(x = var_2923_cast_fp16)[name = tensor("op_2924_cast_fp16")]; + tensor x_321_cast_fp16 = real_div(x = sub_82_cast_fp16, y = var_2924_cast_fp16)[name = tensor("x_321_cast_fp16")]; + tensor var_2926_cast_fp16 = mul(x = x_321_cast_fp16, y = flow_net_res_blocks_1_in_ln_weight_to_fp16)[name = tensor("op_2926_cast_fp16")]; + tensor x_323_cast_fp16 = add(x = var_2926_cast_fp16, y = flow_net_res_blocks_1_in_ln_bias_to_fp16)[name = tensor("x_323_cast_fp16")]; + tensor var_2928_promoted_to_fp16 = const()[name = tensor("op_2928_promoted_to_fp16"), val = tensor(0x1p+0)]; + tensor var_2929_cast_fp16 = add(x = var_2912_cast_fp16_1, y = var_2928_promoted_to_fp16)[name = tensor("op_2929_cast_fp16")]; + tensor var_2930_cast_fp16 = mul(x = x_323_cast_fp16, y = var_2929_cast_fp16)[name = tensor("op_2930_cast_fp16")]; + tensor input_501_cast_fp16 = add(x = var_2930_cast_fp16, y = var_2912_cast_fp16_0)[name = tensor("input_501_cast_fp16")]; + tensor linear_192_cast_fp16 = linear(bias = flow_net_res_blocks_1_mlp_0_bias_to_fp16, weight = flow_net_res_blocks_1_mlp_0_weight_to_fp16, x = input_501_cast_fp16)[name = tensor("linear_192_cast_fp16")]; + tensor input_505_cast_fp16 = silu(x = linear_192_cast_fp16)[name = tensor("input_505_cast_fp16")]; + tensor linear_193_cast_fp16 = linear(bias = flow_net_res_blocks_1_mlp_2_bias_to_fp16, weight = flow_net_res_blocks_1_mlp_2_weight_to_fp16, x = input_505_cast_fp16)[name = tensor("linear_193_cast_fp16")]; + tensor var_2941_cast_fp16 = mul(x = var_2912_cast_fp16_2, y = linear_193_cast_fp16)[name = tensor("op_2941_cast_fp16")]; + tensor x_325_cast_fp16 = add(x = x_319_cast_fp16, y = var_2941_cast_fp16)[name = tensor("x_325_cast_fp16")]; + tensor linear_194_cast_fp16 = linear(bias = flow_net_res_blocks_2_adaLN_modulation_1_bias_to_fp16, weight = flow_net_res_blocks_2_adaLN_modulation_1_weight_to_fp16, x = input_491_cast_fp16)[name = tensor("linear_194_cast_fp16")]; + tensor var_2951_split_sizes_0 = const()[name = tensor("op_2951_split_sizes_0"), val = tensor([512, 512, 512])]; + tensor var_2951_axis_0 = const()[name = tensor("op_2951_axis_0"), val = tensor(-1)]; + tensor var_2951_cast_fp16_0, tensor var_2951_cast_fp16_1, tensor var_2951_cast_fp16_2 = split(axis = var_2951_axis_0, split_sizes = var_2951_split_sizes_0, x = linear_194_cast_fp16)[name = tensor("op_2951_cast_fp16")]; + tensor mean_103_axes_0 = const()[name = tensor("mean_103_axes_0"), val = tensor([-1])]; + tensor mean_103_keep_dims_0 = const()[name = tensor("mean_103_keep_dims_0"), val = tensor(true)]; + tensor mean_103_cast_fp16 = reduce_mean(axes = mean_103_axes_0, keep_dims = mean_103_keep_dims_0, x = x_325_cast_fp16)[name = tensor("mean_103_cast_fp16")]; + tensor sub_83_cast_fp16 = sub(x = x_325_cast_fp16, y = mean_103_cast_fp16)[name = tensor("sub_83_cast_fp16")]; + tensor square_67_cast_fp16 = square(x = sub_83_cast_fp16)[name = tensor("square_67_cast_fp16")]; + tensor reduce_mean_135_axes_0 = const()[name = tensor("reduce_mean_135_axes_0"), val = tensor([-1])]; + tensor reduce_mean_135_keep_dims_0 = const()[name = tensor("reduce_mean_135_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_135_cast_fp16 = reduce_mean(axes = reduce_mean_135_axes_0, keep_dims = reduce_mean_135_keep_dims_0, x = square_67_cast_fp16)[name = tensor("reduce_mean_135_cast_fp16")]; + tensor var_2961_to_fp16 = const()[name = tensor("op_2961_to_fp16"), val = tensor(0x1.1p-20)]; + tensor var_2962_cast_fp16 = add(x = reduce_mean_135_cast_fp16, y = var_2961_to_fp16)[name = tensor("op_2962_cast_fp16")]; + tensor var_2963_cast_fp16 = sqrt(x = var_2962_cast_fp16)[name = tensor("op_2963_cast_fp16")]; + tensor x_327_cast_fp16 = real_div(x = sub_83_cast_fp16, y = var_2963_cast_fp16)[name = tensor("x_327_cast_fp16")]; + tensor var_2965_cast_fp16 = mul(x = x_327_cast_fp16, y = flow_net_res_blocks_2_in_ln_weight_to_fp16)[name = tensor("op_2965_cast_fp16")]; + tensor x_329_cast_fp16 = add(x = var_2965_cast_fp16, y = flow_net_res_blocks_2_in_ln_bias_to_fp16)[name = tensor("x_329_cast_fp16")]; + tensor var_2967_promoted_to_fp16 = const()[name = tensor("op_2967_promoted_to_fp16"), val = tensor(0x1p+0)]; + tensor var_2968_cast_fp16 = add(x = var_2951_cast_fp16_1, y = var_2967_promoted_to_fp16)[name = tensor("op_2968_cast_fp16")]; + tensor var_2969_cast_fp16 = mul(x = x_329_cast_fp16, y = var_2968_cast_fp16)[name = tensor("op_2969_cast_fp16")]; + tensor input_509_cast_fp16 = add(x = var_2969_cast_fp16, y = var_2951_cast_fp16_0)[name = tensor("input_509_cast_fp16")]; + tensor linear_195_cast_fp16 = linear(bias = flow_net_res_blocks_2_mlp_0_bias_to_fp16, weight = flow_net_res_blocks_2_mlp_0_weight_to_fp16, x = input_509_cast_fp16)[name = tensor("linear_195_cast_fp16")]; + tensor input_513_cast_fp16 = silu(x = linear_195_cast_fp16)[name = tensor("input_513_cast_fp16")]; + tensor linear_196_cast_fp16 = linear(bias = flow_net_res_blocks_2_mlp_2_bias_to_fp16, weight = flow_net_res_blocks_2_mlp_2_weight_to_fp16, x = input_513_cast_fp16)[name = tensor("linear_196_cast_fp16")]; + tensor var_2980_cast_fp16 = mul(x = var_2951_cast_fp16_2, y = linear_196_cast_fp16)[name = tensor("op_2980_cast_fp16")]; + tensor x_331_cast_fp16 = add(x = x_325_cast_fp16, y = var_2980_cast_fp16)[name = tensor("x_331_cast_fp16")]; + tensor linear_197_cast_fp16 = linear(bias = flow_net_res_blocks_3_adaLN_modulation_1_bias_to_fp16, weight = flow_net_res_blocks_3_adaLN_modulation_1_weight_to_fp16, x = input_491_cast_fp16)[name = tensor("linear_197_cast_fp16")]; + tensor var_2990_split_sizes_0 = const()[name = tensor("op_2990_split_sizes_0"), val = tensor([512, 512, 512])]; + tensor var_2990_axis_0 = const()[name = tensor("op_2990_axis_0"), val = tensor(-1)]; + tensor var_2990_cast_fp16_0, tensor var_2990_cast_fp16_1, tensor var_2990_cast_fp16_2 = split(axis = var_2990_axis_0, split_sizes = var_2990_split_sizes_0, x = linear_197_cast_fp16)[name = tensor("op_2990_cast_fp16")]; + tensor mean_105_axes_0 = const()[name = tensor("mean_105_axes_0"), val = tensor([-1])]; + tensor mean_105_keep_dims_0 = const()[name = tensor("mean_105_keep_dims_0"), val = tensor(true)]; + tensor mean_105_cast_fp16 = reduce_mean(axes = mean_105_axes_0, keep_dims = mean_105_keep_dims_0, x = x_331_cast_fp16)[name = tensor("mean_105_cast_fp16")]; + tensor sub_84_cast_fp16 = sub(x = x_331_cast_fp16, y = mean_105_cast_fp16)[name = tensor("sub_84_cast_fp16")]; + tensor square_68_cast_fp16 = square(x = sub_84_cast_fp16)[name = tensor("square_68_cast_fp16")]; + tensor reduce_mean_137_axes_0 = const()[name = tensor("reduce_mean_137_axes_0"), val = tensor([-1])]; + tensor reduce_mean_137_keep_dims_0 = const()[name = tensor("reduce_mean_137_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_137_cast_fp16 = reduce_mean(axes = reduce_mean_137_axes_0, keep_dims = reduce_mean_137_keep_dims_0, x = square_68_cast_fp16)[name = tensor("reduce_mean_137_cast_fp16")]; + tensor var_3000_to_fp16 = const()[name = tensor("op_3000_to_fp16"), val = tensor(0x1.1p-20)]; + tensor var_3001_cast_fp16 = add(x = reduce_mean_137_cast_fp16, y = var_3000_to_fp16)[name = tensor("op_3001_cast_fp16")]; + tensor var_3002_cast_fp16 = sqrt(x = var_3001_cast_fp16)[name = tensor("op_3002_cast_fp16")]; + tensor x_333_cast_fp16 = real_div(x = sub_84_cast_fp16, y = var_3002_cast_fp16)[name = tensor("x_333_cast_fp16")]; + tensor var_3004_cast_fp16 = mul(x = x_333_cast_fp16, y = flow_net_res_blocks_3_in_ln_weight_to_fp16)[name = tensor("op_3004_cast_fp16")]; + tensor x_335_cast_fp16 = add(x = var_3004_cast_fp16, y = flow_net_res_blocks_3_in_ln_bias_to_fp16)[name = tensor("x_335_cast_fp16")]; + tensor var_3006_promoted_to_fp16 = const()[name = tensor("op_3006_promoted_to_fp16"), val = tensor(0x1p+0)]; + tensor var_3007_cast_fp16 = add(x = var_2990_cast_fp16_1, y = var_3006_promoted_to_fp16)[name = tensor("op_3007_cast_fp16")]; + tensor var_3008_cast_fp16 = mul(x = x_335_cast_fp16, y = var_3007_cast_fp16)[name = tensor("op_3008_cast_fp16")]; + tensor input_517_cast_fp16 = add(x = var_3008_cast_fp16, y = var_2990_cast_fp16_0)[name = tensor("input_517_cast_fp16")]; + tensor linear_198_cast_fp16 = linear(bias = flow_net_res_blocks_3_mlp_0_bias_to_fp16, weight = flow_net_res_blocks_3_mlp_0_weight_to_fp16, x = input_517_cast_fp16)[name = tensor("linear_198_cast_fp16")]; + tensor input_521_cast_fp16 = silu(x = linear_198_cast_fp16)[name = tensor("input_521_cast_fp16")]; + tensor linear_199_cast_fp16 = linear(bias = flow_net_res_blocks_3_mlp_2_bias_to_fp16, weight = flow_net_res_blocks_3_mlp_2_weight_to_fp16, x = input_521_cast_fp16)[name = tensor("linear_199_cast_fp16")]; + tensor var_3019_cast_fp16 = mul(x = var_2990_cast_fp16_2, y = linear_199_cast_fp16)[name = tensor("op_3019_cast_fp16")]; + tensor x_337_cast_fp16 = add(x = x_331_cast_fp16, y = var_3019_cast_fp16)[name = tensor("x_337_cast_fp16")]; + tensor linear_200_cast_fp16 = linear(bias = flow_net_res_blocks_4_adaLN_modulation_1_bias_to_fp16, weight = flow_net_res_blocks_4_adaLN_modulation_1_weight_to_fp16, x = input_491_cast_fp16)[name = tensor("linear_200_cast_fp16")]; + tensor var_3029_split_sizes_0 = const()[name = tensor("op_3029_split_sizes_0"), val = tensor([512, 512, 512])]; + tensor var_3029_axis_0 = const()[name = tensor("op_3029_axis_0"), val = tensor(-1)]; + tensor var_3029_cast_fp16_0, tensor var_3029_cast_fp16_1, tensor var_3029_cast_fp16_2 = split(axis = var_3029_axis_0, split_sizes = var_3029_split_sizes_0, x = linear_200_cast_fp16)[name = tensor("op_3029_cast_fp16")]; + tensor mean_107_axes_0 = const()[name = tensor("mean_107_axes_0"), val = tensor([-1])]; + tensor mean_107_keep_dims_0 = const()[name = tensor("mean_107_keep_dims_0"), val = tensor(true)]; + tensor mean_107_cast_fp16 = reduce_mean(axes = mean_107_axes_0, keep_dims = mean_107_keep_dims_0, x = x_337_cast_fp16)[name = tensor("mean_107_cast_fp16")]; + tensor sub_85_cast_fp16 = sub(x = x_337_cast_fp16, y = mean_107_cast_fp16)[name = tensor("sub_85_cast_fp16")]; + tensor square_69_cast_fp16 = square(x = sub_85_cast_fp16)[name = tensor("square_69_cast_fp16")]; + tensor reduce_mean_139_axes_0 = const()[name = tensor("reduce_mean_139_axes_0"), val = tensor([-1])]; + tensor reduce_mean_139_keep_dims_0 = const()[name = tensor("reduce_mean_139_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_139_cast_fp16 = reduce_mean(axes = reduce_mean_139_axes_0, keep_dims = reduce_mean_139_keep_dims_0, x = square_69_cast_fp16)[name = tensor("reduce_mean_139_cast_fp16")]; + tensor var_3039_to_fp16 = const()[name = tensor("op_3039_to_fp16"), val = tensor(0x1.1p-20)]; + tensor var_3040_cast_fp16 = add(x = reduce_mean_139_cast_fp16, y = var_3039_to_fp16)[name = tensor("op_3040_cast_fp16")]; + tensor var_3041_cast_fp16 = sqrt(x = var_3040_cast_fp16)[name = tensor("op_3041_cast_fp16")]; + tensor x_339_cast_fp16 = real_div(x = sub_85_cast_fp16, y = var_3041_cast_fp16)[name = tensor("x_339_cast_fp16")]; + tensor var_3043_cast_fp16 = mul(x = x_339_cast_fp16, y = flow_net_res_blocks_4_in_ln_weight_to_fp16)[name = tensor("op_3043_cast_fp16")]; + tensor x_341_cast_fp16 = add(x = var_3043_cast_fp16, y = flow_net_res_blocks_4_in_ln_bias_to_fp16)[name = tensor("x_341_cast_fp16")]; + tensor var_3045_promoted_to_fp16 = const()[name = tensor("op_3045_promoted_to_fp16"), val = tensor(0x1p+0)]; + tensor var_3046_cast_fp16 = add(x = var_3029_cast_fp16_1, y = var_3045_promoted_to_fp16)[name = tensor("op_3046_cast_fp16")]; + tensor var_3047_cast_fp16 = mul(x = x_341_cast_fp16, y = var_3046_cast_fp16)[name = tensor("op_3047_cast_fp16")]; + tensor input_525_cast_fp16 = add(x = var_3047_cast_fp16, y = var_3029_cast_fp16_0)[name = tensor("input_525_cast_fp16")]; + tensor linear_201_cast_fp16 = linear(bias = flow_net_res_blocks_4_mlp_0_bias_to_fp16, weight = flow_net_res_blocks_4_mlp_0_weight_to_fp16, x = input_525_cast_fp16)[name = tensor("linear_201_cast_fp16")]; + tensor input_529_cast_fp16 = silu(x = linear_201_cast_fp16)[name = tensor("input_529_cast_fp16")]; + tensor linear_202_cast_fp16 = linear(bias = flow_net_res_blocks_4_mlp_2_bias_to_fp16, weight = flow_net_res_blocks_4_mlp_2_weight_to_fp16, x = input_529_cast_fp16)[name = tensor("linear_202_cast_fp16")]; + tensor var_3058_cast_fp16 = mul(x = var_3029_cast_fp16_2, y = linear_202_cast_fp16)[name = tensor("op_3058_cast_fp16")]; + tensor x_343_cast_fp16 = add(x = x_337_cast_fp16, y = var_3058_cast_fp16)[name = tensor("x_343_cast_fp16")]; + tensor linear_203_cast_fp16 = linear(bias = flow_net_res_blocks_5_adaLN_modulation_1_bias_to_fp16, weight = flow_net_res_blocks_5_adaLN_modulation_1_weight_to_fp16, x = input_491_cast_fp16)[name = tensor("linear_203_cast_fp16")]; + tensor var_3068_split_sizes_0 = const()[name = tensor("op_3068_split_sizes_0"), val = tensor([512, 512, 512])]; + tensor var_3068_axis_0 = const()[name = tensor("op_3068_axis_0"), val = tensor(-1)]; + tensor var_3068_cast_fp16_0, tensor var_3068_cast_fp16_1, tensor var_3068_cast_fp16_2 = split(axis = var_3068_axis_0, split_sizes = var_3068_split_sizes_0, x = linear_203_cast_fp16)[name = tensor("op_3068_cast_fp16")]; + tensor mean_109_axes_0 = const()[name = tensor("mean_109_axes_0"), val = tensor([-1])]; + tensor mean_109_keep_dims_0 = const()[name = tensor("mean_109_keep_dims_0"), val = tensor(true)]; + tensor mean_109_cast_fp16 = reduce_mean(axes = mean_109_axes_0, keep_dims = mean_109_keep_dims_0, x = x_343_cast_fp16)[name = tensor("mean_109_cast_fp16")]; + tensor sub_86_cast_fp16 = sub(x = x_343_cast_fp16, y = mean_109_cast_fp16)[name = tensor("sub_86_cast_fp16")]; + tensor square_70_cast_fp16 = square(x = sub_86_cast_fp16)[name = tensor("square_70_cast_fp16")]; + tensor reduce_mean_141_axes_0 = const()[name = tensor("reduce_mean_141_axes_0"), val = tensor([-1])]; + tensor reduce_mean_141_keep_dims_0 = const()[name = tensor("reduce_mean_141_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_141_cast_fp16 = reduce_mean(axes = reduce_mean_141_axes_0, keep_dims = reduce_mean_141_keep_dims_0, x = square_70_cast_fp16)[name = tensor("reduce_mean_141_cast_fp16")]; + tensor var_3078_to_fp16 = const()[name = tensor("op_3078_to_fp16"), val = tensor(0x1.1p-20)]; + tensor var_3079_cast_fp16 = add(x = reduce_mean_141_cast_fp16, y = var_3078_to_fp16)[name = tensor("op_3079_cast_fp16")]; + tensor var_3080_cast_fp16 = sqrt(x = var_3079_cast_fp16)[name = tensor("op_3080_cast_fp16")]; + tensor x_345_cast_fp16 = real_div(x = sub_86_cast_fp16, y = var_3080_cast_fp16)[name = tensor("x_345_cast_fp16")]; + tensor var_3082_cast_fp16 = mul(x = x_345_cast_fp16, y = flow_net_res_blocks_5_in_ln_weight_to_fp16)[name = tensor("op_3082_cast_fp16")]; + tensor x_347_cast_fp16 = add(x = var_3082_cast_fp16, y = flow_net_res_blocks_5_in_ln_bias_to_fp16)[name = tensor("x_347_cast_fp16")]; + tensor var_3084_promoted_to_fp16 = const()[name = tensor("op_3084_promoted_to_fp16"), val = tensor(0x1p+0)]; + tensor var_3085_cast_fp16 = add(x = var_3068_cast_fp16_1, y = var_3084_promoted_to_fp16)[name = tensor("op_3085_cast_fp16")]; + tensor var_3086_cast_fp16 = mul(x = x_347_cast_fp16, y = var_3085_cast_fp16)[name = tensor("op_3086_cast_fp16")]; + tensor input_533_cast_fp16 = add(x = var_3086_cast_fp16, y = var_3068_cast_fp16_0)[name = tensor("input_533_cast_fp16")]; + tensor linear_204_cast_fp16 = linear(bias = flow_net_res_blocks_5_mlp_0_bias_to_fp16, weight = flow_net_res_blocks_5_mlp_0_weight_to_fp16, x = input_533_cast_fp16)[name = tensor("linear_204_cast_fp16")]; + tensor input_537_cast_fp16 = silu(x = linear_204_cast_fp16)[name = tensor("input_537_cast_fp16")]; + tensor linear_205_cast_fp16 = linear(bias = flow_net_res_blocks_5_mlp_2_bias_to_fp16, weight = flow_net_res_blocks_5_mlp_2_weight_to_fp16, x = input_537_cast_fp16)[name = tensor("linear_205_cast_fp16")]; + tensor var_3097_cast_fp16 = mul(x = var_3068_cast_fp16_2, y = linear_205_cast_fp16)[name = tensor("op_3097_cast_fp16")]; + tensor x_349_cast_fp16 = add(x = x_343_cast_fp16, y = var_3097_cast_fp16)[name = tensor("x_349_cast_fp16")]; + tensor linear_206_cast_fp16 = linear(bias = flow_net_final_layer_adaLN_modulation_1_bias_to_fp16, weight = flow_net_final_layer_adaLN_modulation_1_weight_to_fp16, x = input_491_cast_fp16)[name = tensor("linear_206_cast_fp16")]; + tensor var_3106_split_sizes_0 = const()[name = tensor("op_3106_split_sizes_0"), val = tensor([512, 512])]; + tensor var_3106_axis_0 = const()[name = tensor("op_3106_axis_0"), val = tensor(-1)]; + tensor var_3106_cast_fp16_0, tensor var_3106_cast_fp16_1 = split(axis = var_3106_axis_0, split_sizes = var_3106_split_sizes_0, x = linear_206_cast_fp16)[name = tensor("op_3106_cast_fp16")]; + tensor mean_axes_0 = const()[name = tensor("mean_axes_0"), val = tensor([-1])]; + tensor mean_keep_dims_0 = const()[name = tensor("mean_keep_dims_0"), val = tensor(true)]; + tensor mean_cast_fp16 = reduce_mean(axes = mean_axes_0, keep_dims = mean_keep_dims_0, x = x_349_cast_fp16)[name = tensor("mean_cast_fp16")]; + tensor sub_87_cast_fp16 = sub(x = x_349_cast_fp16, y = mean_cast_fp16)[name = tensor("sub_87_cast_fp16")]; + tensor square_71_cast_fp16 = square(x = sub_87_cast_fp16)[name = tensor("square_71_cast_fp16")]; + tensor reduce_mean_143_axes_0 = const()[name = tensor("reduce_mean_143_axes_0"), val = tensor([-1])]; + tensor reduce_mean_143_keep_dims_0 = const()[name = tensor("reduce_mean_143_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_143_cast_fp16 = reduce_mean(axes = reduce_mean_143_axes_0, keep_dims = reduce_mean_143_keep_dims_0, x = square_71_cast_fp16)[name = tensor("reduce_mean_143_cast_fp16")]; + tensor var_3113_to_fp16 = const()[name = tensor("op_3113_to_fp16"), val = tensor(0x1.1p-20)]; + tensor var_3114_cast_fp16 = add(x = reduce_mean_143_cast_fp16, y = var_3113_to_fp16)[name = tensor("op_3114_cast_fp16")]; + tensor var_3115_cast_fp16 = sqrt(x = var_3114_cast_fp16)[name = tensor("op_3115_cast_fp16")]; + tensor x_cast_fp16 = real_div(x = sub_87_cast_fp16, y = var_3115_cast_fp16)[name = tensor("x_cast_fp16")]; + tensor var_3117_promoted_to_fp16 = const()[name = tensor("op_3117_promoted_to_fp16"), val = tensor(0x1p+0)]; + tensor var_3118_cast_fp16 = add(x = var_3106_cast_fp16_1, y = var_3117_promoted_to_fp16)[name = tensor("op_3118_cast_fp16")]; + tensor var_3119_cast_fp16 = mul(x = x_cast_fp16, y = var_3118_cast_fp16)[name = tensor("op_3119_cast_fp16")]; + tensor input_cast_fp16 = add(x = var_3119_cast_fp16, y = var_3106_cast_fp16_0)[name = tensor("input_cast_fp16")]; + tensor linear_207_cast_fp16 = linear(bias = flow_net_final_layer_linear_bias_to_fp16, weight = flow_net_final_layer_linear_weight_to_fp16, x = input_cast_fp16)[name = tensor("linear_207_cast_fp16")]; + tensor var_3124_to_fp16 = const()[name = tensor("op_3124_to_fp16"), val = tensor(0x1p-3)]; + tensor var_3125_cast_fp16 = mul(x = linear_207_cast_fp16, y = var_3124_to_fp16)[name = tensor("op_3125_cast_fp16")]; + tensor var_3127_cast_fp16 = add(x = input_475_cast_fp16, y = var_3125_cast_fp16)[name = tensor("op_3127_cast_fp16")]; + tensor var_3127_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_3127_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor latent_final = cast(dtype = var_3127_cast_fp16_to_fp32_dtype_0, x = var_3127_cast_fp16)[name = tensor("cast_144")]; + } -> (latent_final); +} \ No newline at end of file diff --git a/v2.1/portuguese/flow_decoder_fused.mlmodelc/weights/weight.bin b/v2.1/portuguese/flow_decoder_fused.mlmodelc/weights/weight.bin new file mode 100644 index 0000000000000000000000000000000000000000..cea8cca7f8dee90a55f1d0ca46903c8b578db21d --- /dev/null +++ b/v2.1/portuguese/flow_decoder_fused.mlmodelc/weights/weight.bin @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:a55edc1a818e94cf3f3b00b89be7ddae87aa064b6131986b92f6625e4d5753f6 +size 19012608 diff --git a/v2.1/portuguese/flow_decoder_fused.mlpackage/Data/com.apple.CoreML/model.mlmodel b/v2.1/portuguese/flow_decoder_fused.mlpackage/Data/com.apple.CoreML/model.mlmodel new file mode 100644 index 0000000000000000000000000000000000000000..c7beed6032f200847631fd45b22a576b0ba0dc67 --- /dev/null +++ b/v2.1/portuguese/flow_decoder_fused.mlpackage/Data/com.apple.CoreML/model.mlmodel @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:d249f63a373b54eff422186ba6e469b1cfee37617d7e37744c3766ea84746d97 +size 301733 diff --git a/v2.1/portuguese/flow_decoder_fused.mlpackage/Data/com.apple.CoreML/weights/weight.bin b/v2.1/portuguese/flow_decoder_fused.mlpackage/Data/com.apple.CoreML/weights/weight.bin new file mode 100644 index 0000000000000000000000000000000000000000..cea8cca7f8dee90a55f1d0ca46903c8b578db21d --- /dev/null +++ b/v2.1/portuguese/flow_decoder_fused.mlpackage/Data/com.apple.CoreML/weights/weight.bin @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:a55edc1a818e94cf3f3b00b89be7ddae87aa064b6131986b92f6625e4d5753f6 +size 19012608 diff --git a/v2.1/portuguese/flow_decoder_fused.mlpackage/Manifest.json b/v2.1/portuguese/flow_decoder_fused.mlpackage/Manifest.json new file mode 100644 index 0000000000000000000000000000000000000000..4519f887ddcf50ac2297c02f745144379eefd78e --- /dev/null +++ b/v2.1/portuguese/flow_decoder_fused.mlpackage/Manifest.json @@ -0,0 +1,18 @@ +{ + "fileFormatVersion": "1.0.0", + "itemInfoEntries": { + "5587BD60-0C70-4D0E-B733-4168242B4993": { + "author": "com.apple.CoreML", + "description": "CoreML Model Weights", + "name": "weights", + "path": "com.apple.CoreML/weights" + }, + "C74D9A15-75BB-4175-B93A-AD2BD3804BA8": { + "author": "com.apple.CoreML", + "description": "CoreML Model Specification", + "name": "model.mlmodel", + "path": "com.apple.CoreML/model.mlmodel" + } + }, + "rootModelIdentifier": "C74D9A15-75BB-4175-B93A-AD2BD3804BA8" +} diff --git a/v2.1/portuguese/flowlm_step.mlmodelc/analytics/coremldata.bin b/v2.1/portuguese/flowlm_step.mlmodelc/analytics/coremldata.bin new file mode 100644 index 0000000000000000000000000000000000000000..e934c6a296859f00be7cc361b8ee91ba4a11c6af --- /dev/null +++ b/v2.1/portuguese/flowlm_step.mlmodelc/analytics/coremldata.bin @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:7846ef914dc2092385460ad5aa93b99d5f86133eccde84187a07f9ad17e694f5 +size 243 diff --git a/v2.1/portuguese/flowlm_step.mlmodelc/coremldata.bin b/v2.1/portuguese/flowlm_step.mlmodelc/coremldata.bin new file mode 100644 index 0000000000000000000000000000000000000000..9aa363bc4fe7413eddd9d89bfd02228b54394404 --- /dev/null +++ b/v2.1/portuguese/flowlm_step.mlmodelc/coremldata.bin @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:c86e746ee6542c7333b9ad2c56409167368a796e124dbeb3c824b3464518f34f +size 1191 diff --git a/v2.1/portuguese/flowlm_step.mlmodelc/model.mil b/v2.1/portuguese/flowlm_step.mlmodelc/model.mil new file mode 100644 index 0000000000000000000000000000000000000000..a4abf02de5b305963323f8433669999c6a6430c8 --- /dev/null +++ b/v2.1/portuguese/flowlm_step.mlmodelc/model.mil @@ -0,0 +1,1359 @@ +program(1.0) +[buildInfo = dict, tensor>({{"coremlc-component-MIL", "3520.4.1"}, {"coremlc-version", "3520.5.1"}, {"coremltools-component-torch", "2.12.0"}, {"coremltools-source-dialect", "TorchScript"}, {"coremltools-version", "9.0"}})] +{ + func main(tensor bos_emb, tensor cache0, tensor cache1, tensor cache2, tensor cache3, tensor cache4, tensor cache5, tensor position0, tensor position1, tensor position2, tensor position3, tensor position4, tensor position5, tensor sequence) { + tensor sequence_to_fp16_dtype_0 = const()[name = tensor("sequence_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor sequence_to_fp16 = cast(dtype = sequence_to_fp16_dtype_0, x = sequence)[name = tensor("cast_141")]; + tensor var_54_cast_fp16 = not_equal(x = sequence_to_fp16, y = sequence_to_fp16)[name = tensor("op_54_cast_fp16")]; + tensor expand_dims_0_axes_0 = const()[name = tensor("expand_dims_0_axes_0"), val = tensor([0, 1])]; + tensor bos_emb_to_fp16_dtype_0 = const()[name = tensor("bos_emb_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor bos_emb_to_fp16 = cast(dtype = bos_emb_to_fp16_dtype_0, x = bos_emb)[name = tensor("cast_140")]; + tensor expand_dims_0_cast_fp16 = expand_dims(axes = expand_dims_0_axes_0, x = bos_emb_to_fp16)[name = tensor("expand_dims_0_cast_fp16")]; + tensor input_1_cast_fp16 = select(a = expand_dims_0_cast_fp16, b = sequence_to_fp16, cond = var_54_cast_fp16)[name = tensor("input_1_cast_fp16")]; + tensor input_linear_weight_to_fp16 = const()[name = tensor("input_linear_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(64)))]; + tensor linear_0_bias_0_to_fp16 = const()[name = tensor("linear_0_bias_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(65664)))]; + tensor linear_0_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = input_linear_weight_to_fp16, x = input_1_cast_fp16)[name = tensor("linear_0_cast_fp16")]; + tensor input_5_axes_0 = const()[name = tensor("input_5_axes_0"), val = tensor([-1])]; + tensor norm0_1_weight_to_fp16 = const()[name = tensor("norm0_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(67776)))]; + tensor norm0_1_bias_to_fp16 = const()[name = tensor("norm0_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(69888)))]; + tensor var_60_to_fp16 = const()[name = tensor("op_60_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_5_cast_fp16 = layer_norm(axes = input_5_axes_0, beta = norm0_1_bias_to_fp16, epsilon = var_60_to_fp16, gamma = norm0_1_weight_to_fp16, x = linear_0_cast_fp16)[name = tensor("input_5_cast_fp16")]; + tensor attn0_in_proj_weight_to_fp16 = const()[name = tensor("attn0_in_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(72000)))]; + tensor linear_1_bias_0_to_fp16 = const()[name = tensor("linear_1_bias_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(6363520)))]; + tensor linear_1_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = attn0_in_proj_weight_to_fp16, x = input_5_cast_fp16)[name = tensor("linear_1_cast_fp16")]; + tensor var_74 = const()[name = tensor("op_74"), val = tensor([1, 1, 3, 16, 64])]; + tensor qkv_1_cast_fp16 = reshape(shape = var_74, x = linear_1_cast_fp16)[name = tensor("qkv_1_cast_fp16")]; + tensor q_1_begin_0 = const()[name = tensor("q_1_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor q_1_end_0 = const()[name = tensor("q_1_end_0"), val = tensor([1, 1, 1, 16, 64])]; + tensor q_1_end_mask_0 = const()[name = tensor("q_1_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor q_1_squeeze_mask_0 = const()[name = tensor("q_1_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor q_1_cast_fp16 = slice_by_index(begin = q_1_begin_0, end = q_1_end_0, end_mask = q_1_end_mask_0, squeeze_mask = q_1_squeeze_mask_0, x = qkv_1_cast_fp16)[name = tensor("q_1_cast_fp16")]; + tensor k_1_begin_0 = const()[name = tensor("k_1_begin_0"), val = tensor([0, 0, 1, 0, 0])]; + tensor k_1_end_0 = const()[name = tensor("k_1_end_0"), val = tensor([1, 1, 2, 16, 64])]; + tensor k_1_end_mask_0 = const()[name = tensor("k_1_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor k_1_squeeze_mask_0 = const()[name = tensor("k_1_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor k_1_cast_fp16 = slice_by_index(begin = k_1_begin_0, end = k_1_end_0, end_mask = k_1_end_mask_0, squeeze_mask = k_1_squeeze_mask_0, x = qkv_1_cast_fp16)[name = tensor("k_1_cast_fp16")]; + tensor v_1_begin_0 = const()[name = tensor("v_1_begin_0"), val = tensor([0, 0, 2, 0, 0])]; + tensor v_1_end_0 = const()[name = tensor("v_1_end_0"), val = tensor([1, 1, 3, 16, 64])]; + tensor v_1_end_mask_0 = const()[name = tensor("v_1_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor v_1_squeeze_mask_0 = const()[name = tensor("v_1_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor v_1_cast_fp16 = slice_by_index(begin = v_1_begin_0, end = v_1_end_0, end_mask = v_1_end_mask_0, squeeze_mask = v_1_squeeze_mask_0, x = qkv_1_cast_fp16)[name = tensor("v_1_cast_fp16")]; + tensor var_140 = const()[name = tensor("op_140"), val = tensor([1, 1, 1, 1])]; + tensor position0_to_fp16_dtype_0 = const()[name = tensor("position0_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor position0_to_fp16 = cast(dtype = position0_to_fp16_dtype_0, x = position0)[name = tensor("cast_139")]; + tensor ts_5_cast_fp16 = reshape(shape = var_140, x = position0_to_fp16)[name = tensor("ts_5_cast_fp16")]; + tensor var_147 = const()[name = tensor("op_147"), val = tensor([1, 1, 16, 32, 2])]; + tensor q_complex_1_cast_fp16 = reshape(shape = var_147, x = q_1_cast_fp16)[name = tensor("q_complex_1_cast_fp16")]; + tensor var_154 = const()[name = tensor("op_154"), val = tensor([1, 1, 16, 32, 2])]; + tensor k_complex_1_cast_fp16 = reshape(shape = var_154, x = k_1_cast_fp16)[name = tensor("k_complex_1_cast_fp16")]; + tensor var_158_begin_0 = const()[name = tensor("op_158_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_158_end_0 = const()[name = tensor("op_158_end_0"), val = tensor([1, 1, 16, 32, 1])]; + tensor var_158_end_mask_0 = const()[name = tensor("op_158_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_158_squeeze_mask_0 = const()[name = tensor("op_158_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_158_cast_fp16 = slice_by_index(begin = var_158_begin_0, end = var_158_end_0, end_mask = var_158_end_mask_0, squeeze_mask = var_158_squeeze_mask_0, x = q_complex_1_cast_fp16)[name = tensor("op_158_cast_fp16")]; + tensor var_166_begin_0 = const()[name = tensor("op_166_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_166_end_0 = const()[name = tensor("op_166_end_0"), val = tensor([1, 1, 16, 32, 2])]; + tensor var_166_end_mask_0 = const()[name = tensor("op_166_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_166_squeeze_mask_0 = const()[name = tensor("op_166_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_166_cast_fp16 = slice_by_index(begin = var_166_begin_0, end = var_166_end_0, end_mask = var_166_end_mask_0, squeeze_mask = var_166_squeeze_mask_0, x = q_complex_1_cast_fp16)[name = tensor("op_166_cast_fp16")]; + tensor var_174_begin_0 = const()[name = tensor("op_174_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_174_end_0 = const()[name = tensor("op_174_end_0"), val = tensor([1, 1, 16, 32, 1])]; + tensor var_174_end_mask_0 = const()[name = tensor("op_174_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_174_squeeze_mask_0 = const()[name = tensor("op_174_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_174_cast_fp16 = slice_by_index(begin = var_174_begin_0, end = var_174_end_0, end_mask = var_174_end_mask_0, squeeze_mask = var_174_squeeze_mask_0, x = k_complex_1_cast_fp16)[name = tensor("op_174_cast_fp16")]; + tensor var_182_begin_0 = const()[name = tensor("op_182_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_182_end_0 = const()[name = tensor("op_182_end_0"), val = tensor([1, 1, 16, 32, 2])]; + tensor var_182_end_mask_0 = const()[name = tensor("op_182_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_182_squeeze_mask_0 = const()[name = tensor("op_182_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_182_cast_fp16 = slice_by_index(begin = var_182_begin_0, end = var_182_end_0, end_mask = var_182_end_mask_0, squeeze_mask = var_182_squeeze_mask_0, x = k_complex_1_cast_fp16)[name = tensor("op_182_cast_fp16")]; + tensor freqs_1_to_fp16 = const()[name = tensor("freqs_1_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(6369728)))]; + tensor var_188_cast_fp16 = mul(x = freqs_1_to_fp16, y = ts_5_cast_fp16)[name = tensor("op_188_cast_fp16")]; + tensor rotr_1_cast_fp16 = cos(x = var_188_cast_fp16)[name = tensor("rotr_1_cast_fp16")]; + tensor roti_1_cast_fp16 = sin(x = var_188_cast_fp16)[name = tensor("roti_1_cast_fp16")]; + tensor var_192_cast_fp16 = mul(x = var_158_cast_fp16, y = rotr_1_cast_fp16)[name = tensor("op_192_cast_fp16")]; + tensor var_193_cast_fp16 = mul(x = var_166_cast_fp16, y = roti_1_cast_fp16)[name = tensor("op_193_cast_fp16")]; + tensor qor_1_cast_fp16 = sub(x = var_192_cast_fp16, y = var_193_cast_fp16)[name = tensor("qor_1_cast_fp16")]; + tensor var_196_cast_fp16 = mul(x = var_158_cast_fp16, y = roti_1_cast_fp16)[name = tensor("op_196_cast_fp16")]; + tensor var_197_cast_fp16 = mul(x = var_166_cast_fp16, y = rotr_1_cast_fp16)[name = tensor("op_197_cast_fp16")]; + tensor qoi_1_cast_fp16 = add(x = var_196_cast_fp16, y = var_197_cast_fp16)[name = tensor("qoi_1_cast_fp16")]; + tensor var_200_cast_fp16 = mul(x = var_174_cast_fp16, y = rotr_1_cast_fp16)[name = tensor("op_200_cast_fp16")]; + tensor var_201_cast_fp16 = mul(x = var_182_cast_fp16, y = roti_1_cast_fp16)[name = tensor("op_201_cast_fp16")]; + tensor kor_1_cast_fp16 = sub(x = var_200_cast_fp16, y = var_201_cast_fp16)[name = tensor("kor_1_cast_fp16")]; + tensor var_204_cast_fp16 = mul(x = var_174_cast_fp16, y = roti_1_cast_fp16)[name = tensor("op_204_cast_fp16")]; + tensor var_205_cast_fp16 = mul(x = var_182_cast_fp16, y = rotr_1_cast_fp16)[name = tensor("op_205_cast_fp16")]; + tensor koi_1_cast_fp16 = add(x = var_204_cast_fp16, y = var_205_cast_fp16)[name = tensor("koi_1_cast_fp16")]; + tensor qo_1_axis_0 = const()[name = tensor("qo_1_axis_0"), val = tensor(-1)]; + tensor qo_1_cast_fp16 = stack(axis = qo_1_axis_0, values = (qor_1_cast_fp16, qoi_1_cast_fp16))[name = tensor("qo_1_cast_fp16")]; + tensor ko_1_axis_0 = const()[name = tensor("ko_1_axis_0"), val = tensor(-1)]; + tensor ko_1_cast_fp16 = stack(axis = ko_1_axis_0, values = (kor_1_cast_fp16, koi_1_cast_fp16))[name = tensor("ko_1_cast_fp16")]; + tensor var_238 = const()[name = tensor("op_238"), val = tensor([1, 1, 16, 64])]; + tensor q_3_cast_fp16 = reshape(shape = var_238, x = qo_1_cast_fp16)[name = tensor("q_3_cast_fp16")]; + tensor var_244 = const()[name = tensor("op_244"), val = tensor([1, 1, 16, 64])]; + tensor k_3_cast_fp16 = reshape(shape = var_244, x = ko_1_cast_fp16)[name = tensor("k_3_cast_fp16")]; + tensor _inversed_269_y_0_to_fp16 = const()[name = tensor("_inversed_269_y_0_to_fp16"), val = tensor(0x1p-9)]; + tensor _inversed_269_cast_fp16 = mul(x = ts_5_cast_fp16, y = _inversed_269_y_0_to_fp16)[name = tensor("_inversed_269_cast_fp16")]; + tensor var_270_cast_fp16 = floor(x = _inversed_269_cast_fp16)[name = tensor("op_270_cast_fp16")]; + tensor var_271_to_fp16 = const()[name = tensor("op_271_to_fp16"), val = tensor(0x1p+9)]; + tensor var_272_cast_fp16 = mul(x = var_270_cast_fp16, y = var_271_to_fp16)[name = tensor("op_272_cast_fp16")]; + tensor write_indices_float_3_cast_fp16 = sub(x = ts_5_cast_fp16, y = var_272_cast_fp16)[name = tensor("write_indices_float_3_cast_fp16")]; + tensor var_279_dtype_0 = const()[name = tensor("op_279_dtype_0"), val = tensor("int32")]; + tensor write_indices_1_reps_0 = const()[name = tensor("write_indices_1_reps_0"), val = tensor([1, 1, 16, 64])]; + tensor write_indices_float_3_cast_fp16_to_int32 = cast(dtype = var_279_dtype_0, x = write_indices_float_3_cast_fp16)[name = tensor("cast_138")]; + tensor write_indices_1 = tile(reps = write_indices_1_reps_0, x = write_indices_float_3_cast_fp16_to_int32)[name = tensor("write_indices_1")]; + tensor var_289_begin_0 = const()[name = tensor("op_289_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_289_end_0 = const()[name = tensor("op_289_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor var_289_end_mask_0 = const()[name = tensor("op_289_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_289_squeeze_mask_0 = const()[name = tensor("op_289_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor cache0_to_fp16_dtype_0 = const()[name = tensor("cache0_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor cache0_to_fp16 = cast(dtype = cache0_to_fp16_dtype_0, x = cache0)[name = tensor("cast_137")]; + tensor var_289_cast_fp16 = slice_by_index(begin = var_289_begin_0, end = var_289_end_0, end_mask = var_289_end_mask_0, squeeze_mask = var_289_squeeze_mask_0, x = cache0_to_fp16)[name = tensor("op_289_cast_fp16")]; + tensor var_291_axis_0 = const()[name = tensor("op_291_axis_0"), val = tensor(1)]; + tensor var_291_mode_0 = const()[name = tensor("op_291_mode_0"), val = tensor("update")]; + tensor var_291_validate_indices_0 = const()[name = tensor("op_291_validate_indices_0"), val = tensor(false)]; + tensor var_291_cast_fp16 = scatter_along_axis(axis = var_291_axis_0, data = var_289_cast_fp16, indices = write_indices_1, mode = var_291_mode_0, updates = k_3_cast_fp16, validate_indices = var_291_validate_indices_0)[name = tensor("op_291_cast_fp16")]; + tensor concat_2 = const()[name = tensor("concat_2"), val = tensor([0, 0, 0, 0, 0])]; + tensor concat_3 = const()[name = tensor("concat_3"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_1_internal_tensor_assign_1_stride_0 = const()[name = tensor("new_cache_1_internal_tensor_assign_1_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_1_internal_tensor_assign_1_begin_mask_0 = const()[name = tensor("new_cache_1_internal_tensor_assign_1_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_1_internal_tensor_assign_1_end_mask_0 = const()[name = tensor("new_cache_1_internal_tensor_assign_1_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_1_internal_tensor_assign_1_squeeze_mask_0 = const()[name = tensor("new_cache_1_internal_tensor_assign_1_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_12 = const()[name = tensor("shape_12"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_0 = const()[name = tensor("reduce_prod_0"), val = tensor(1048576)]; + tensor range_1d_0_start_0 = const()[name = tensor("range_1d_0_start_0"), val = tensor(0)]; + tensor range_1d_0_step_0 = const()[name = tensor("range_1d_0_step_0"), val = tensor(1)]; + tensor range_1d_0 = range_1d(end = reduce_prod_0, start = range_1d_0_start_0, step = range_1d_0_step_0)[name = tensor("range_1d_0")]; + tensor reshape_0 = reshape(shape = shape_12, x = range_1d_0)[name = tensor("reshape_0")]; + tensor slice_by_index_0 = slice_by_index(begin = concat_2, begin_mask = new_cache_1_internal_tensor_assign_1_begin_mask_0, end = concat_3, end_mask = new_cache_1_internal_tensor_assign_1_end_mask_0, squeeze_mask = new_cache_1_internal_tensor_assign_1_squeeze_mask_0, stride = new_cache_1_internal_tensor_assign_1_stride_0, x = reshape_0)[name = tensor("slice_by_index_0")]; + tensor reshape_1_shape_0 = const()[name = tensor("reshape_1_shape_0"), val = tensor([-1])]; + tensor reshape_1 = reshape(shape = reshape_1_shape_0, x = slice_by_index_0)[name = tensor("reshape_1")]; + tensor reshape_2_shape_0 = const()[name = tensor("reshape_2_shape_0"), val = tensor([-1])]; + tensor reshape_2_cast_fp16 = reshape(shape = reshape_2_shape_0, x = var_291_cast_fp16)[name = tensor("reshape_2_cast_fp16")]; + tensor reshape_3_shape_0 = const()[name = tensor("reshape_3_shape_0"), val = tensor([-1])]; + tensor reshape_3_cast_fp16 = reshape(shape = reshape_3_shape_0, x = cache0_to_fp16)[name = tensor("reshape_3_cast_fp16")]; + tensor scatter_0_mode_0 = const()[name = tensor("scatter_0_mode_0"), val = tensor("update")]; + tensor scatter_0_axis_0 = const()[name = tensor("scatter_0_axis_0"), val = tensor(0)]; + tensor scatter_0_validate_indices_0 = const()[name = tensor("scatter_0_validate_indices_0"), val = tensor(false)]; + tensor scatter_0_cast_fp16 = scatter(axis = scatter_0_axis_0, data = reshape_3_cast_fp16, indices = reshape_1, mode = scatter_0_mode_0, updates = reshape_2_cast_fp16, validate_indices = scatter_0_validate_indices_0)[name = tensor("scatter_0_cast_fp16")]; + tensor reshape_4_cast_fp16 = reshape(shape = shape_12, x = scatter_0_cast_fp16)[name = tensor("reshape_4_cast_fp16")]; + tensor var_299_begin_0 = const()[name = tensor("op_299_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor var_299_end_0 = const()[name = tensor("op_299_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor var_299_end_mask_0 = const()[name = tensor("op_299_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_299_squeeze_mask_0 = const()[name = tensor("op_299_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor var_299_cast_fp16 = slice_by_index(begin = var_299_begin_0, end = var_299_end_0, end_mask = var_299_end_mask_0, squeeze_mask = var_299_squeeze_mask_0, x = reshape_4_cast_fp16)[name = tensor("op_299_cast_fp16")]; + tensor var_301_axis_0 = const()[name = tensor("op_301_axis_0"), val = tensor(1)]; + tensor var_301_mode_0 = const()[name = tensor("op_301_mode_0"), val = tensor("update")]; + tensor var_301_validate_indices_0 = const()[name = tensor("op_301_validate_indices_0"), val = tensor(false)]; + tensor var_301_cast_fp16 = scatter_along_axis(axis = var_301_axis_0, data = var_299_cast_fp16, indices = write_indices_1, mode = var_301_mode_0, updates = v_1_cast_fp16, validate_indices = var_301_validate_indices_0)[name = tensor("op_301_cast_fp16")]; + tensor concat_4 = const()[name = tensor("concat_4"), val = tensor([1, 0, 0, 0, 0])]; + tensor concat_5 = const()[name = tensor("concat_5"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_1_internal_tensor_assign_2_stride_0 = const()[name = tensor("new_cache_1_internal_tensor_assign_2_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_1_internal_tensor_assign_2_begin_mask_0 = const()[name = tensor("new_cache_1_internal_tensor_assign_2_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_1_internal_tensor_assign_2_end_mask_0 = const()[name = tensor("new_cache_1_internal_tensor_assign_2_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_1_internal_tensor_assign_2_squeeze_mask_0 = const()[name = tensor("new_cache_1_internal_tensor_assign_2_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_13 = const()[name = tensor("shape_13"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_1 = const()[name = tensor("reduce_prod_1"), val = tensor(1048576)]; + tensor range_1d_1_start_0 = const()[name = tensor("range_1d_1_start_0"), val = tensor(0)]; + tensor range_1d_1_step_0 = const()[name = tensor("range_1d_1_step_0"), val = tensor(1)]; + tensor range_1d_1 = range_1d(end = reduce_prod_1, start = range_1d_1_start_0, step = range_1d_1_step_0)[name = tensor("range_1d_1")]; + tensor reshape_5 = reshape(shape = shape_13, x = range_1d_1)[name = tensor("reshape_5")]; + tensor slice_by_index_1 = slice_by_index(begin = concat_4, begin_mask = new_cache_1_internal_tensor_assign_2_begin_mask_0, end = concat_5, end_mask = new_cache_1_internal_tensor_assign_2_end_mask_0, squeeze_mask = new_cache_1_internal_tensor_assign_2_squeeze_mask_0, stride = new_cache_1_internal_tensor_assign_2_stride_0, x = reshape_5)[name = tensor("slice_by_index_1")]; + tensor reshape_6_shape_0 = const()[name = tensor("reshape_6_shape_0"), val = tensor([-1])]; + tensor reshape_6 = reshape(shape = reshape_6_shape_0, x = slice_by_index_1)[name = tensor("reshape_6")]; + tensor reshape_7_shape_0 = const()[name = tensor("reshape_7_shape_0"), val = tensor([-1])]; + tensor reshape_7_cast_fp16 = reshape(shape = reshape_7_shape_0, x = var_301_cast_fp16)[name = tensor("reshape_7_cast_fp16")]; + tensor reshape_8_shape_0 = const()[name = tensor("reshape_8_shape_0"), val = tensor([-1])]; + tensor reshape_8_cast_fp16 = reshape(shape = reshape_8_shape_0, x = reshape_4_cast_fp16)[name = tensor("reshape_8_cast_fp16")]; + tensor scatter_1_mode_0 = const()[name = tensor("scatter_1_mode_0"), val = tensor("update")]; + tensor scatter_1_axis_0 = const()[name = tensor("scatter_1_axis_0"), val = tensor(0)]; + tensor scatter_1_validate_indices_0 = const()[name = tensor("scatter_1_validate_indices_0"), val = tensor(false)]; + tensor scatter_1_cast_fp16 = scatter(axis = scatter_1_axis_0, data = reshape_8_cast_fp16, indices = reshape_6, mode = scatter_1_mode_0, updates = reshape_7_cast_fp16, validate_indices = scatter_1_validate_indices_0)[name = tensor("scatter_1_cast_fp16")]; + tensor reshape_9_cast_fp16 = reshape(shape = shape_13, x = scatter_1_cast_fp16)[name = tensor("reshape_9_cast_fp16")]; + tensor reshape_9_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("reshape_9_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor keys_1_begin_0 = const()[name = tensor("keys_1_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor keys_1_end_0 = const()[name = tensor("keys_1_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor keys_1_end_mask_0 = const()[name = tensor("keys_1_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor keys_1_squeeze_mask_0 = const()[name = tensor("keys_1_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor keys_1_cast_fp16 = slice_by_index(begin = keys_1_begin_0, end = keys_1_end_0, end_mask = keys_1_end_mask_0, squeeze_mask = keys_1_squeeze_mask_0, x = reshape_9_cast_fp16)[name = tensor("keys_1_cast_fp16")]; + tensor values_1_begin_0 = const()[name = tensor("values_1_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor values_1_end_0 = const()[name = tensor("values_1_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor values_1_end_mask_0 = const()[name = tensor("values_1_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor values_1_squeeze_mask_0 = const()[name = tensor("values_1_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor values_1_cast_fp16 = slice_by_index(begin = values_1_begin_0, end = values_1_end_0, end_mask = values_1_end_mask_0, squeeze_mask = values_1_squeeze_mask_0, x = reshape_9_cast_fp16)[name = tensor("values_1_cast_fp16")]; + tensor var_313_cast_fp16 = not_equal(x = keys_1_cast_fp16, y = keys_1_cast_fp16)[name = tensor("op_313_cast_fp16")]; + tensor var_319_to_fp16 = const()[name = tensor("op_319_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(6369856)))]; + tensor keys_3_cast_fp16 = select(a = var_319_to_fp16, b = keys_1_cast_fp16, cond = var_313_cast_fp16)[name = tensor("keys_3_cast_fp16")]; + tensor var_321_cast_fp16 = not_equal(x = values_1_cast_fp16, y = values_1_cast_fp16)[name = tensor("op_321_cast_fp16")]; + tensor values_3_cast_fp16 = select(a = var_319_to_fp16, b = values_1_cast_fp16, cond = var_321_cast_fp16)[name = tensor("values_3_cast_fp16")]; + tensor var_345 = const()[name = tensor("op_345"), val = tensor([0, 2, 1, 3])]; + tensor var_361 = const()[name = tensor("op_361"), val = tensor([1, 1, 1])]; + tensor var_362_cast_fp16 = reshape(shape = var_361, x = position0_to_fp16)[name = tensor("op_362_cast_fp16")]; + tensor var_382_to_fp16 = const()[name = tensor("op_382_to_fp16"), val = tensor(0x1p+0)]; + tensor valid_len_1_cast_fp16 = add(x = var_362_cast_fp16, y = var_382_to_fp16)[name = tensor("valid_len_1_cast_fp16")]; + tensor k_positions_1_promoted_to_fp16 = const()[name = tensor("k_positions_1_promoted_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(7418496)))]; + tensor valid_mask_1_cast_fp16 = less(x = k_positions_1_promoted_to_fp16, y = valid_len_1_cast_fp16)[name = tensor("valid_mask_1_cast_fp16")]; + tensor causal_mask_1_cast_fp16 = less_equal(x = k_positions_1_promoted_to_fp16, y = var_362_cast_fp16)[name = tensor("causal_mask_1_cast_fp16")]; + tensor attn_mask_1 = logical_and(x = valid_mask_1_cast_fp16, y = causal_mask_1_cast_fp16)[name = tensor("attn_mask_1")]; + tensor attn_mask_3_axes_0 = const()[name = tensor("attn_mask_3_axes_0"), val = tensor([1])]; + tensor attn_mask_3 = expand_dims(axes = attn_mask_3_axes_0, x = attn_mask_1)[name = tensor("attn_mask_3")]; + tensor var_400_transpose_x_0 = const()[name = tensor("op_400_transpose_x_0"), val = tensor(false)]; + tensor var_400_transpose_y_0 = const()[name = tensor("op_400_transpose_y_0"), val = tensor(false)]; + tensor transpose_18_perm_0 = const()[name = tensor("transpose_18_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_19_perm_0 = const()[name = tensor("transpose_19_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_19 = transpose(perm = transpose_19_perm_0, x = keys_3_cast_fp16)[name = tensor("transpose_51")]; + tensor transpose_18 = transpose(perm = transpose_18_perm_0, x = q_3_cast_fp16)[name = tensor("transpose_52")]; + tensor var_400_cast_fp16 = matmul(transpose_x = var_400_transpose_x_0, transpose_y = var_400_transpose_y_0, x = transpose_18, y = transpose_19)[name = tensor("op_400_cast_fp16")]; + tensor var_394_to_fp16 = const()[name = tensor("op_394_to_fp16"), val = tensor([0x1p-3])]; + tensor attn_weights_1_cast_fp16 = mul(x = var_400_cast_fp16, y = var_394_to_fp16)[name = tensor("attn_weights_1_cast_fp16")]; + tensor var_402 = logical_not(x = attn_mask_3)[name = tensor("op_402")]; + tensor var_403_to_fp16 = const()[name = tensor("op_403_to_fp16"), val = tensor(-inf)]; + tensor attn_weights_3_cast_fp16 = select(a = var_403_to_fp16, b = attn_weights_1_cast_fp16, cond = var_402)[name = tensor("attn_weights_3_cast_fp16")]; + tensor var_405 = const()[name = tensor("op_405"), val = tensor(-1)]; + tensor attn_weights_5_cast_fp16 = softmax(axis = var_405, x = attn_weights_3_cast_fp16)[name = tensor("attn_weights_5_cast_fp16")]; + tensor attn_output_1_transpose_x_0 = const()[name = tensor("attn_output_1_transpose_x_0"), val = tensor(false)]; + tensor attn_output_1_transpose_y_0 = const()[name = tensor("attn_output_1_transpose_y_0"), val = tensor(false)]; + tensor values_5_cast_fp16 = transpose(perm = var_345, x = values_3_cast_fp16)[name = tensor("transpose_53")]; + tensor attn_output_1_cast_fp16 = matmul(transpose_x = attn_output_1_transpose_x_0, transpose_y = attn_output_1_transpose_y_0, x = attn_weights_5_cast_fp16, y = values_5_cast_fp16)[name = tensor("attn_output_1_cast_fp16")]; + tensor var_413 = const()[name = tensor("op_413"), val = tensor([0, 2, 1, 3])]; + tensor var_418 = const()[name = tensor("op_418"), val = tensor([1, 1, 1024])]; + tensor var_414_cast_fp16 = transpose(perm = var_413, x = attn_output_1_cast_fp16)[name = tensor("transpose_50")]; + tensor input_7_cast_fp16 = reshape(shape = var_418, x = var_414_cast_fp16)[name = tensor("input_7_cast_fp16")]; + tensor attn0_out_proj_weight_to_fp16 = const()[name = tensor("attn0_out_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(7419584)))]; + tensor linear_2_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = attn0_out_proj_weight_to_fp16, x = input_7_cast_fp16)[name = tensor("linear_2_cast_fp16")]; + tensor var_424_to_fp16 = const()[name = tensor("op_424_to_fp16"), val = tensor(0x1p+0)]; + tensor var_425_cast_fp16 = add(x = position0_to_fp16, y = var_424_to_fp16)[name = tensor("op_425_cast_fp16")]; + tensor var_425_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_425_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor input_9_cast_fp16 = add(x = linear_0_cast_fp16, y = linear_2_cast_fp16)[name = tensor("input_9_cast_fp16")]; + tensor input_11_axes_0 = const()[name = tensor("input_11_axes_0"), val = tensor([-1])]; + tensor norm0_2_weight_to_fp16 = const()[name = tensor("norm0_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9516800)))]; + tensor norm0_2_bias_to_fp16 = const()[name = tensor("norm0_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9518912)))]; + tensor var_429_to_fp16 = const()[name = tensor("op_429_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_11_cast_fp16 = layer_norm(axes = input_11_axes_0, beta = norm0_2_bias_to_fp16, epsilon = var_429_to_fp16, gamma = norm0_2_weight_to_fp16, x = input_9_cast_fp16)[name = tensor("input_11_cast_fp16")]; + tensor linear0_1_weight_to_fp16 = const()[name = tensor("linear0_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9521024)))]; + tensor linear_3_bias_0_to_fp16 = const()[name = tensor("linear_3_bias_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(17909696)))]; + tensor linear_3_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = linear0_1_weight_to_fp16, x = input_11_cast_fp16)[name = tensor("linear_3_cast_fp16")]; + tensor input_13_mode_0 = const()[name = tensor("input_13_mode_0"), val = tensor("EXACT")]; + tensor input_13_cast_fp16 = gelu(mode = input_13_mode_0, x = linear_3_cast_fp16)[name = tensor("input_13_cast_fp16")]; + tensor linear0_2_weight_to_fp16 = const()[name = tensor("linear0_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(17917952)))]; + tensor linear_4_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = linear0_2_weight_to_fp16, x = input_13_cast_fp16)[name = tensor("linear_4_cast_fp16")]; + tensor input_15_cast_fp16 = add(x = input_9_cast_fp16, y = linear_4_cast_fp16)[name = tensor("input_15_cast_fp16")]; + tensor input_17_axes_0 = const()[name = tensor("input_17_axes_0"), val = tensor([-1])]; + tensor norm1_1_weight_to_fp16 = const()[name = tensor("norm1_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(26306624)))]; + tensor norm1_1_bias_to_fp16 = const()[name = tensor("norm1_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(26308736)))]; + tensor var_446_to_fp16 = const()[name = tensor("op_446_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_17_cast_fp16 = layer_norm(axes = input_17_axes_0, beta = norm1_1_bias_to_fp16, epsilon = var_446_to_fp16, gamma = norm1_1_weight_to_fp16, x = input_15_cast_fp16)[name = tensor("input_17_cast_fp16")]; + tensor attn1_in_proj_weight_to_fp16 = const()[name = tensor("attn1_in_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(26310848)))]; + tensor linear_5_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = attn1_in_proj_weight_to_fp16, x = input_17_cast_fp16)[name = tensor("linear_5_cast_fp16")]; + tensor var_460 = const()[name = tensor("op_460"), val = tensor([1, 1, 3, 16, 64])]; + tensor qkv_3_cast_fp16 = reshape(shape = var_460, x = linear_5_cast_fp16)[name = tensor("qkv_3_cast_fp16")]; + tensor q_7_begin_0 = const()[name = tensor("q_7_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor q_7_end_0 = const()[name = tensor("q_7_end_0"), val = tensor([1, 1, 1, 16, 64])]; + tensor q_7_end_mask_0 = const()[name = tensor("q_7_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor q_7_squeeze_mask_0 = const()[name = tensor("q_7_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor q_7_cast_fp16 = slice_by_index(begin = q_7_begin_0, end = q_7_end_0, end_mask = q_7_end_mask_0, squeeze_mask = q_7_squeeze_mask_0, x = qkv_3_cast_fp16)[name = tensor("q_7_cast_fp16")]; + tensor k_5_begin_0 = const()[name = tensor("k_5_begin_0"), val = tensor([0, 0, 1, 0, 0])]; + tensor k_5_end_0 = const()[name = tensor("k_5_end_0"), val = tensor([1, 1, 2, 16, 64])]; + tensor k_5_end_mask_0 = const()[name = tensor("k_5_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor k_5_squeeze_mask_0 = const()[name = tensor("k_5_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor k_5_cast_fp16 = slice_by_index(begin = k_5_begin_0, end = k_5_end_0, end_mask = k_5_end_mask_0, squeeze_mask = k_5_squeeze_mask_0, x = qkv_3_cast_fp16)[name = tensor("k_5_cast_fp16")]; + tensor v_3_begin_0 = const()[name = tensor("v_3_begin_0"), val = tensor([0, 0, 2, 0, 0])]; + tensor v_3_end_0 = const()[name = tensor("v_3_end_0"), val = tensor([1, 1, 3, 16, 64])]; + tensor v_3_end_mask_0 = const()[name = tensor("v_3_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor v_3_squeeze_mask_0 = const()[name = tensor("v_3_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor v_3_cast_fp16 = slice_by_index(begin = v_3_begin_0, end = v_3_end_0, end_mask = v_3_end_mask_0, squeeze_mask = v_3_squeeze_mask_0, x = qkv_3_cast_fp16)[name = tensor("v_3_cast_fp16")]; + tensor var_526 = const()[name = tensor("op_526"), val = tensor([1, 1, 1, 1])]; + tensor position1_to_fp16_dtype_0 = const()[name = tensor("position1_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor position1_to_fp16 = cast(dtype = position1_to_fp16_dtype_0, x = position1)[name = tensor("cast_134")]; + tensor ts_11_cast_fp16 = reshape(shape = var_526, x = position1_to_fp16)[name = tensor("ts_11_cast_fp16")]; + tensor var_533 = const()[name = tensor("op_533"), val = tensor([1, 1, 16, 32, 2])]; + tensor q_complex_3_cast_fp16 = reshape(shape = var_533, x = q_7_cast_fp16)[name = tensor("q_complex_3_cast_fp16")]; + tensor var_540 = const()[name = tensor("op_540"), val = tensor([1, 1, 16, 32, 2])]; + tensor k_complex_3_cast_fp16 = reshape(shape = var_540, x = k_5_cast_fp16)[name = tensor("k_complex_3_cast_fp16")]; + tensor var_544_begin_0 = const()[name = tensor("op_544_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_544_end_0 = const()[name = tensor("op_544_end_0"), val = tensor([1, 1, 16, 32, 1])]; + tensor var_544_end_mask_0 = const()[name = tensor("op_544_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_544_squeeze_mask_0 = const()[name = tensor("op_544_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_544_cast_fp16 = slice_by_index(begin = var_544_begin_0, end = var_544_end_0, end_mask = var_544_end_mask_0, squeeze_mask = var_544_squeeze_mask_0, x = q_complex_3_cast_fp16)[name = tensor("op_544_cast_fp16")]; + tensor var_552_begin_0 = const()[name = tensor("op_552_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_552_end_0 = const()[name = tensor("op_552_end_0"), val = tensor([1, 1, 16, 32, 2])]; + tensor var_552_end_mask_0 = const()[name = tensor("op_552_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_552_squeeze_mask_0 = const()[name = tensor("op_552_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_552_cast_fp16 = slice_by_index(begin = var_552_begin_0, end = var_552_end_0, end_mask = var_552_end_mask_0, squeeze_mask = var_552_squeeze_mask_0, x = q_complex_3_cast_fp16)[name = tensor("op_552_cast_fp16")]; + tensor var_560_begin_0 = const()[name = tensor("op_560_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_560_end_0 = const()[name = tensor("op_560_end_0"), val = tensor([1, 1, 16, 32, 1])]; + tensor var_560_end_mask_0 = const()[name = tensor("op_560_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_560_squeeze_mask_0 = const()[name = tensor("op_560_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_560_cast_fp16 = slice_by_index(begin = var_560_begin_0, end = var_560_end_0, end_mask = var_560_end_mask_0, squeeze_mask = var_560_squeeze_mask_0, x = k_complex_3_cast_fp16)[name = tensor("op_560_cast_fp16")]; + tensor var_568_begin_0 = const()[name = tensor("op_568_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_568_end_0 = const()[name = tensor("op_568_end_0"), val = tensor([1, 1, 16, 32, 2])]; + tensor var_568_end_mask_0 = const()[name = tensor("op_568_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_568_squeeze_mask_0 = const()[name = tensor("op_568_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_568_cast_fp16 = slice_by_index(begin = var_568_begin_0, end = var_568_end_0, end_mask = var_568_end_mask_0, squeeze_mask = var_568_squeeze_mask_0, x = k_complex_3_cast_fp16)[name = tensor("op_568_cast_fp16")]; + tensor freqs_3_to_fp16 = const()[name = tensor("freqs_3_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(32602368)))]; + tensor var_574_cast_fp16 = mul(x = freqs_3_to_fp16, y = ts_11_cast_fp16)[name = tensor("op_574_cast_fp16")]; + tensor rotr_3_cast_fp16 = cos(x = var_574_cast_fp16)[name = tensor("rotr_3_cast_fp16")]; + tensor roti_3_cast_fp16 = sin(x = var_574_cast_fp16)[name = tensor("roti_3_cast_fp16")]; + tensor var_578_cast_fp16 = mul(x = var_544_cast_fp16, y = rotr_3_cast_fp16)[name = tensor("op_578_cast_fp16")]; + tensor var_579_cast_fp16 = mul(x = var_552_cast_fp16, y = roti_3_cast_fp16)[name = tensor("op_579_cast_fp16")]; + tensor qor_5_cast_fp16 = sub(x = var_578_cast_fp16, y = var_579_cast_fp16)[name = tensor("qor_5_cast_fp16")]; + tensor var_582_cast_fp16 = mul(x = var_544_cast_fp16, y = roti_3_cast_fp16)[name = tensor("op_582_cast_fp16")]; + tensor var_583_cast_fp16 = mul(x = var_552_cast_fp16, y = rotr_3_cast_fp16)[name = tensor("op_583_cast_fp16")]; + tensor qoi_5_cast_fp16 = add(x = var_582_cast_fp16, y = var_583_cast_fp16)[name = tensor("qoi_5_cast_fp16")]; + tensor var_586_cast_fp16 = mul(x = var_560_cast_fp16, y = rotr_3_cast_fp16)[name = tensor("op_586_cast_fp16")]; + tensor var_587_cast_fp16 = mul(x = var_568_cast_fp16, y = roti_3_cast_fp16)[name = tensor("op_587_cast_fp16")]; + tensor kor_5_cast_fp16 = sub(x = var_586_cast_fp16, y = var_587_cast_fp16)[name = tensor("kor_5_cast_fp16")]; + tensor var_590_cast_fp16 = mul(x = var_560_cast_fp16, y = roti_3_cast_fp16)[name = tensor("op_590_cast_fp16")]; + tensor var_591_cast_fp16 = mul(x = var_568_cast_fp16, y = rotr_3_cast_fp16)[name = tensor("op_591_cast_fp16")]; + tensor koi_5_cast_fp16 = add(x = var_590_cast_fp16, y = var_591_cast_fp16)[name = tensor("koi_5_cast_fp16")]; + tensor qo_3_axis_0 = const()[name = tensor("qo_3_axis_0"), val = tensor(-1)]; + tensor qo_3_cast_fp16 = stack(axis = qo_3_axis_0, values = (qor_5_cast_fp16, qoi_5_cast_fp16))[name = tensor("qo_3_cast_fp16")]; + tensor ko_3_axis_0 = const()[name = tensor("ko_3_axis_0"), val = tensor(-1)]; + tensor ko_3_cast_fp16 = stack(axis = ko_3_axis_0, values = (kor_5_cast_fp16, koi_5_cast_fp16))[name = tensor("ko_3_cast_fp16")]; + tensor var_624 = const()[name = tensor("op_624"), val = tensor([1, 1, 16, 64])]; + tensor q_9_cast_fp16 = reshape(shape = var_624, x = qo_3_cast_fp16)[name = tensor("q_9_cast_fp16")]; + tensor var_630 = const()[name = tensor("op_630"), val = tensor([1, 1, 16, 64])]; + tensor k_7_cast_fp16 = reshape(shape = var_630, x = ko_3_cast_fp16)[name = tensor("k_7_cast_fp16")]; + tensor _inversed_655_y_0_to_fp16 = const()[name = tensor("_inversed_655_y_0_to_fp16"), val = tensor(0x1p-9)]; + tensor _inversed_655_cast_fp16 = mul(x = ts_11_cast_fp16, y = _inversed_655_y_0_to_fp16)[name = tensor("_inversed_655_cast_fp16")]; + tensor var_656_cast_fp16 = floor(x = _inversed_655_cast_fp16)[name = tensor("op_656_cast_fp16")]; + tensor var_657_to_fp16 = const()[name = tensor("op_657_to_fp16"), val = tensor(0x1p+9)]; + tensor var_658_cast_fp16 = mul(x = var_656_cast_fp16, y = var_657_to_fp16)[name = tensor("op_658_cast_fp16")]; + tensor write_indices_float_7_cast_fp16 = sub(x = ts_11_cast_fp16, y = var_658_cast_fp16)[name = tensor("write_indices_float_7_cast_fp16")]; + tensor var_665_dtype_0 = const()[name = tensor("op_665_dtype_0"), val = tensor("int32")]; + tensor write_indices_3_reps_0 = const()[name = tensor("write_indices_3_reps_0"), val = tensor([1, 1, 16, 64])]; + tensor write_indices_float_7_cast_fp16_to_int32 = cast(dtype = var_665_dtype_0, x = write_indices_float_7_cast_fp16)[name = tensor("cast_133")]; + tensor write_indices_3 = tile(reps = write_indices_3_reps_0, x = write_indices_float_7_cast_fp16_to_int32)[name = tensor("write_indices_3")]; + tensor var_675_begin_0 = const()[name = tensor("op_675_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_675_end_0 = const()[name = tensor("op_675_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor var_675_end_mask_0 = const()[name = tensor("op_675_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_675_squeeze_mask_0 = const()[name = tensor("op_675_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor cache1_to_fp16_dtype_0 = const()[name = tensor("cache1_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor cache1_to_fp16 = cast(dtype = cache1_to_fp16_dtype_0, x = cache1)[name = tensor("cast_132")]; + tensor var_675_cast_fp16 = slice_by_index(begin = var_675_begin_0, end = var_675_end_0, end_mask = var_675_end_mask_0, squeeze_mask = var_675_squeeze_mask_0, x = cache1_to_fp16)[name = tensor("op_675_cast_fp16")]; + tensor var_677_axis_0 = const()[name = tensor("op_677_axis_0"), val = tensor(1)]; + tensor var_677_mode_0 = const()[name = tensor("op_677_mode_0"), val = tensor("update")]; + tensor var_677_validate_indices_0 = const()[name = tensor("op_677_validate_indices_0"), val = tensor(false)]; + tensor var_677_cast_fp16 = scatter_along_axis(axis = var_677_axis_0, data = var_675_cast_fp16, indices = write_indices_3, mode = var_677_mode_0, updates = k_7_cast_fp16, validate_indices = var_677_validate_indices_0)[name = tensor("op_677_cast_fp16")]; + tensor concat_9 = const()[name = tensor("concat_9"), val = tensor([0, 0, 0, 0, 0])]; + tensor concat_10 = const()[name = tensor("concat_10"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_3_internal_tensor_assign_1_stride_0 = const()[name = tensor("new_cache_3_internal_tensor_assign_1_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_3_internal_tensor_assign_1_begin_mask_0 = const()[name = tensor("new_cache_3_internal_tensor_assign_1_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_3_internal_tensor_assign_1_end_mask_0 = const()[name = tensor("new_cache_3_internal_tensor_assign_1_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_3_internal_tensor_assign_1_squeeze_mask_0 = const()[name = tensor("new_cache_3_internal_tensor_assign_1_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_14 = const()[name = tensor("shape_14"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_2 = const()[name = tensor("reduce_prod_2"), val = tensor(1048576)]; + tensor range_1d_2_start_0 = const()[name = tensor("range_1d_2_start_0"), val = tensor(0)]; + tensor range_1d_2_step_0 = const()[name = tensor("range_1d_2_step_0"), val = tensor(1)]; + tensor range_1d_2 = range_1d(end = reduce_prod_2, start = range_1d_2_start_0, step = range_1d_2_step_0)[name = tensor("range_1d_2")]; + tensor reshape_10 = reshape(shape = shape_14, x = range_1d_2)[name = tensor("reshape_10")]; + tensor slice_by_index_2 = slice_by_index(begin = concat_9, begin_mask = new_cache_3_internal_tensor_assign_1_begin_mask_0, end = concat_10, end_mask = new_cache_3_internal_tensor_assign_1_end_mask_0, squeeze_mask = new_cache_3_internal_tensor_assign_1_squeeze_mask_0, stride = new_cache_3_internal_tensor_assign_1_stride_0, x = reshape_10)[name = tensor("slice_by_index_2")]; + tensor reshape_11_shape_0 = const()[name = tensor("reshape_11_shape_0"), val = tensor([-1])]; + tensor reshape_11 = reshape(shape = reshape_11_shape_0, x = slice_by_index_2)[name = tensor("reshape_11")]; + tensor reshape_12_shape_0 = const()[name = tensor("reshape_12_shape_0"), val = tensor([-1])]; + tensor reshape_12_cast_fp16 = reshape(shape = reshape_12_shape_0, x = var_677_cast_fp16)[name = tensor("reshape_12_cast_fp16")]; + tensor reshape_13_shape_0 = const()[name = tensor("reshape_13_shape_0"), val = tensor([-1])]; + tensor reshape_13_cast_fp16 = reshape(shape = reshape_13_shape_0, x = cache1_to_fp16)[name = tensor("reshape_13_cast_fp16")]; + tensor scatter_2_mode_0 = const()[name = tensor("scatter_2_mode_0"), val = tensor("update")]; + tensor scatter_2_axis_0 = const()[name = tensor("scatter_2_axis_0"), val = tensor(0)]; + tensor scatter_2_validate_indices_0 = const()[name = tensor("scatter_2_validate_indices_0"), val = tensor(false)]; + tensor scatter_2_cast_fp16 = scatter(axis = scatter_2_axis_0, data = reshape_13_cast_fp16, indices = reshape_11, mode = scatter_2_mode_0, updates = reshape_12_cast_fp16, validate_indices = scatter_2_validate_indices_0)[name = tensor("scatter_2_cast_fp16")]; + tensor reshape_14_cast_fp16 = reshape(shape = shape_14, x = scatter_2_cast_fp16)[name = tensor("reshape_14_cast_fp16")]; + tensor var_685_begin_0 = const()[name = tensor("op_685_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor var_685_end_0 = const()[name = tensor("op_685_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor var_685_end_mask_0 = const()[name = tensor("op_685_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_685_squeeze_mask_0 = const()[name = tensor("op_685_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor var_685_cast_fp16 = slice_by_index(begin = var_685_begin_0, end = var_685_end_0, end_mask = var_685_end_mask_0, squeeze_mask = var_685_squeeze_mask_0, x = reshape_14_cast_fp16)[name = tensor("op_685_cast_fp16")]; + tensor var_687_axis_0 = const()[name = tensor("op_687_axis_0"), val = tensor(1)]; + tensor var_687_mode_0 = const()[name = tensor("op_687_mode_0"), val = tensor("update")]; + tensor var_687_validate_indices_0 = const()[name = tensor("op_687_validate_indices_0"), val = tensor(false)]; + tensor var_687_cast_fp16 = scatter_along_axis(axis = var_687_axis_0, data = var_685_cast_fp16, indices = write_indices_3, mode = var_687_mode_0, updates = v_3_cast_fp16, validate_indices = var_687_validate_indices_0)[name = tensor("op_687_cast_fp16")]; + tensor concat_11 = const()[name = tensor("concat_11"), val = tensor([1, 0, 0, 0, 0])]; + tensor concat_12 = const()[name = tensor("concat_12"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_3_internal_tensor_assign_2_stride_0 = const()[name = tensor("new_cache_3_internal_tensor_assign_2_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_3_internal_tensor_assign_2_begin_mask_0 = const()[name = tensor("new_cache_3_internal_tensor_assign_2_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_3_internal_tensor_assign_2_end_mask_0 = const()[name = tensor("new_cache_3_internal_tensor_assign_2_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_3_internal_tensor_assign_2_squeeze_mask_0 = const()[name = tensor("new_cache_3_internal_tensor_assign_2_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_15 = const()[name = tensor("shape_15"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_3 = const()[name = tensor("reduce_prod_3"), val = tensor(1048576)]; + tensor range_1d_3_start_0 = const()[name = tensor("range_1d_3_start_0"), val = tensor(0)]; + tensor range_1d_3_step_0 = const()[name = tensor("range_1d_3_step_0"), val = tensor(1)]; + tensor range_1d_3 = range_1d(end = reduce_prod_3, start = range_1d_3_start_0, step = range_1d_3_step_0)[name = tensor("range_1d_3")]; + tensor reshape_15 = reshape(shape = shape_15, x = range_1d_3)[name = tensor("reshape_15")]; + tensor slice_by_index_3 = slice_by_index(begin = concat_11, begin_mask = new_cache_3_internal_tensor_assign_2_begin_mask_0, end = concat_12, end_mask = new_cache_3_internal_tensor_assign_2_end_mask_0, squeeze_mask = new_cache_3_internal_tensor_assign_2_squeeze_mask_0, stride = new_cache_3_internal_tensor_assign_2_stride_0, x = reshape_15)[name = tensor("slice_by_index_3")]; + tensor reshape_16_shape_0 = const()[name = tensor("reshape_16_shape_0"), val = tensor([-1])]; + tensor reshape_16 = reshape(shape = reshape_16_shape_0, x = slice_by_index_3)[name = tensor("reshape_16")]; + tensor reshape_17_shape_0 = const()[name = tensor("reshape_17_shape_0"), val = tensor([-1])]; + tensor reshape_17_cast_fp16 = reshape(shape = reshape_17_shape_0, x = var_687_cast_fp16)[name = tensor("reshape_17_cast_fp16")]; + tensor reshape_18_shape_0 = const()[name = tensor("reshape_18_shape_0"), val = tensor([-1])]; + tensor reshape_18_cast_fp16 = reshape(shape = reshape_18_shape_0, x = reshape_14_cast_fp16)[name = tensor("reshape_18_cast_fp16")]; + tensor scatter_3_mode_0 = const()[name = tensor("scatter_3_mode_0"), val = tensor("update")]; + tensor scatter_3_axis_0 = const()[name = tensor("scatter_3_axis_0"), val = tensor(0)]; + tensor scatter_3_validate_indices_0 = const()[name = tensor("scatter_3_validate_indices_0"), val = tensor(false)]; + tensor scatter_3_cast_fp16 = scatter(axis = scatter_3_axis_0, data = reshape_18_cast_fp16, indices = reshape_16, mode = scatter_3_mode_0, updates = reshape_17_cast_fp16, validate_indices = scatter_3_validate_indices_0)[name = tensor("scatter_3_cast_fp16")]; + tensor reshape_19_cast_fp16 = reshape(shape = shape_15, x = scatter_3_cast_fp16)[name = tensor("reshape_19_cast_fp16")]; + tensor reshape_19_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("reshape_19_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor keys_7_begin_0 = const()[name = tensor("keys_7_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor keys_7_end_0 = const()[name = tensor("keys_7_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor keys_7_end_mask_0 = const()[name = tensor("keys_7_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor keys_7_squeeze_mask_0 = const()[name = tensor("keys_7_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor keys_7_cast_fp16 = slice_by_index(begin = keys_7_begin_0, end = keys_7_end_0, end_mask = keys_7_end_mask_0, squeeze_mask = keys_7_squeeze_mask_0, x = reshape_19_cast_fp16)[name = tensor("keys_7_cast_fp16")]; + tensor values_7_begin_0 = const()[name = tensor("values_7_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor values_7_end_0 = const()[name = tensor("values_7_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor values_7_end_mask_0 = const()[name = tensor("values_7_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor values_7_squeeze_mask_0 = const()[name = tensor("values_7_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor values_7_cast_fp16 = slice_by_index(begin = values_7_begin_0, end = values_7_end_0, end_mask = values_7_end_mask_0, squeeze_mask = values_7_squeeze_mask_0, x = reshape_19_cast_fp16)[name = tensor("values_7_cast_fp16")]; + tensor var_699_cast_fp16 = not_equal(x = keys_7_cast_fp16, y = keys_7_cast_fp16)[name = tensor("op_699_cast_fp16")]; + tensor keys_9_cast_fp16 = select(a = var_319_to_fp16, b = keys_7_cast_fp16, cond = var_699_cast_fp16)[name = tensor("keys_9_cast_fp16")]; + tensor var_707_cast_fp16 = not_equal(x = values_7_cast_fp16, y = values_7_cast_fp16)[name = tensor("op_707_cast_fp16")]; + tensor values_9_cast_fp16 = select(a = var_319_to_fp16, b = values_7_cast_fp16, cond = var_707_cast_fp16)[name = tensor("values_9_cast_fp16")]; + tensor var_731 = const()[name = tensor("op_731"), val = tensor([0, 2, 1, 3])]; + tensor var_747 = const()[name = tensor("op_747"), val = tensor([1, 1, 1])]; + tensor var_748_cast_fp16 = reshape(shape = var_747, x = position1_to_fp16)[name = tensor("op_748_cast_fp16")]; + tensor var_768_to_fp16 = const()[name = tensor("op_768_to_fp16"), val = tensor(0x1p+0)]; + tensor valid_len_3_cast_fp16 = add(x = var_748_cast_fp16, y = var_768_to_fp16)[name = tensor("valid_len_3_cast_fp16")]; + tensor valid_mask_3_cast_fp16 = less(x = k_positions_1_promoted_to_fp16, y = valid_len_3_cast_fp16)[name = tensor("valid_mask_3_cast_fp16")]; + tensor causal_mask_3_cast_fp16 = less_equal(x = k_positions_1_promoted_to_fp16, y = var_748_cast_fp16)[name = tensor("causal_mask_3_cast_fp16")]; + tensor attn_mask_5 = logical_and(x = valid_mask_3_cast_fp16, y = causal_mask_3_cast_fp16)[name = tensor("attn_mask_5")]; + tensor attn_mask_7_axes_0 = const()[name = tensor("attn_mask_7_axes_0"), val = tensor([1])]; + tensor attn_mask_7 = expand_dims(axes = attn_mask_7_axes_0, x = attn_mask_5)[name = tensor("attn_mask_7")]; + tensor var_786_transpose_x_0 = const()[name = tensor("op_786_transpose_x_0"), val = tensor(false)]; + tensor var_786_transpose_y_0 = const()[name = tensor("op_786_transpose_y_0"), val = tensor(false)]; + tensor transpose_20_perm_0 = const()[name = tensor("transpose_20_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_21_perm_0 = const()[name = tensor("transpose_21_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_21 = transpose(perm = transpose_21_perm_0, x = keys_9_cast_fp16)[name = tensor("transpose_47")]; + tensor transpose_20 = transpose(perm = transpose_20_perm_0, x = q_9_cast_fp16)[name = tensor("transpose_48")]; + tensor var_786_cast_fp16 = matmul(transpose_x = var_786_transpose_x_0, transpose_y = var_786_transpose_y_0, x = transpose_20, y = transpose_21)[name = tensor("op_786_cast_fp16")]; + tensor var_780_to_fp16 = const()[name = tensor("op_780_to_fp16"), val = tensor([0x1p-3])]; + tensor attn_weights_7_cast_fp16 = mul(x = var_786_cast_fp16, y = var_780_to_fp16)[name = tensor("attn_weights_7_cast_fp16")]; + tensor var_788 = logical_not(x = attn_mask_7)[name = tensor("op_788")]; + tensor var_789_to_fp16 = const()[name = tensor("op_789_to_fp16"), val = tensor(-inf)]; + tensor attn_weights_9_cast_fp16 = select(a = var_789_to_fp16, b = attn_weights_7_cast_fp16, cond = var_788)[name = tensor("attn_weights_9_cast_fp16")]; + tensor var_791 = const()[name = tensor("op_791"), val = tensor(-1)]; + tensor attn_weights_11_cast_fp16 = softmax(axis = var_791, x = attn_weights_9_cast_fp16)[name = tensor("attn_weights_11_cast_fp16")]; + tensor attn_output_3_transpose_x_0 = const()[name = tensor("attn_output_3_transpose_x_0"), val = tensor(false)]; + tensor attn_output_3_transpose_y_0 = const()[name = tensor("attn_output_3_transpose_y_0"), val = tensor(false)]; + tensor values_11_cast_fp16 = transpose(perm = var_731, x = values_9_cast_fp16)[name = tensor("transpose_49")]; + tensor attn_output_3_cast_fp16 = matmul(transpose_x = attn_output_3_transpose_x_0, transpose_y = attn_output_3_transpose_y_0, x = attn_weights_11_cast_fp16, y = values_11_cast_fp16)[name = tensor("attn_output_3_cast_fp16")]; + tensor var_799 = const()[name = tensor("op_799"), val = tensor([0, 2, 1, 3])]; + tensor var_804 = const()[name = tensor("op_804"), val = tensor([1, 1, 1024])]; + tensor var_800_cast_fp16 = transpose(perm = var_799, x = attn_output_3_cast_fp16)[name = tensor("transpose_46")]; + tensor input_19_cast_fp16 = reshape(shape = var_804, x = var_800_cast_fp16)[name = tensor("input_19_cast_fp16")]; + tensor attn1_out_proj_weight_to_fp16 = const()[name = tensor("attn1_out_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(32602496)))]; + tensor linear_6_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = attn1_out_proj_weight_to_fp16, x = input_19_cast_fp16)[name = tensor("linear_6_cast_fp16")]; + tensor var_810_to_fp16 = const()[name = tensor("op_810_to_fp16"), val = tensor(0x1p+0)]; + tensor var_811_cast_fp16 = add(x = position1_to_fp16, y = var_810_to_fp16)[name = tensor("op_811_cast_fp16")]; + tensor var_811_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_811_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor input_21_cast_fp16 = add(x = input_15_cast_fp16, y = linear_6_cast_fp16)[name = tensor("input_21_cast_fp16")]; + tensor input_23_axes_0 = const()[name = tensor("input_23_axes_0"), val = tensor([-1])]; + tensor norm1_2_weight_to_fp16 = const()[name = tensor("norm1_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(34699712)))]; + tensor norm1_2_bias_to_fp16 = const()[name = tensor("norm1_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(34701824)))]; + tensor var_815_to_fp16 = const()[name = tensor("op_815_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_23_cast_fp16 = layer_norm(axes = input_23_axes_0, beta = norm1_2_bias_to_fp16, epsilon = var_815_to_fp16, gamma = norm1_2_weight_to_fp16, x = input_21_cast_fp16)[name = tensor("input_23_cast_fp16")]; + tensor linear1_1_weight_to_fp16 = const()[name = tensor("linear1_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(34703936)))]; + tensor linear_7_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = linear1_1_weight_to_fp16, x = input_23_cast_fp16)[name = tensor("linear_7_cast_fp16")]; + tensor input_25_mode_0 = const()[name = tensor("input_25_mode_0"), val = tensor("EXACT")]; + tensor input_25_cast_fp16 = gelu(mode = input_25_mode_0, x = linear_7_cast_fp16)[name = tensor("input_25_cast_fp16")]; + tensor linear1_2_weight_to_fp16 = const()[name = tensor("linear1_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(43092608)))]; + tensor linear_8_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = linear1_2_weight_to_fp16, x = input_25_cast_fp16)[name = tensor("linear_8_cast_fp16")]; + tensor input_27_cast_fp16 = add(x = input_21_cast_fp16, y = linear_8_cast_fp16)[name = tensor("input_27_cast_fp16")]; + tensor input_29_axes_0 = const()[name = tensor("input_29_axes_0"), val = tensor([-1])]; + tensor norm2_1_weight_to_fp16 = const()[name = tensor("norm2_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(51481280)))]; + tensor norm2_1_bias_to_fp16 = const()[name = tensor("norm2_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(51483392)))]; + tensor var_832_to_fp16 = const()[name = tensor("op_832_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_29_cast_fp16 = layer_norm(axes = input_29_axes_0, beta = norm2_1_bias_to_fp16, epsilon = var_832_to_fp16, gamma = norm2_1_weight_to_fp16, x = input_27_cast_fp16)[name = tensor("input_29_cast_fp16")]; + tensor attn2_in_proj_weight_to_fp16 = const()[name = tensor("attn2_in_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(51485504)))]; + tensor linear_9_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = attn2_in_proj_weight_to_fp16, x = input_29_cast_fp16)[name = tensor("linear_9_cast_fp16")]; + tensor var_846 = const()[name = tensor("op_846"), val = tensor([1, 1, 3, 16, 64])]; + tensor qkv_5_cast_fp16 = reshape(shape = var_846, x = linear_9_cast_fp16)[name = tensor("qkv_5_cast_fp16")]; + tensor q_13_begin_0 = const()[name = tensor("q_13_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor q_13_end_0 = const()[name = tensor("q_13_end_0"), val = tensor([1, 1, 1, 16, 64])]; + tensor q_13_end_mask_0 = const()[name = tensor("q_13_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor q_13_squeeze_mask_0 = const()[name = tensor("q_13_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor q_13_cast_fp16 = slice_by_index(begin = q_13_begin_0, end = q_13_end_0, end_mask = q_13_end_mask_0, squeeze_mask = q_13_squeeze_mask_0, x = qkv_5_cast_fp16)[name = tensor("q_13_cast_fp16")]; + tensor k_9_begin_0 = const()[name = tensor("k_9_begin_0"), val = tensor([0, 0, 1, 0, 0])]; + tensor k_9_end_0 = const()[name = tensor("k_9_end_0"), val = tensor([1, 1, 2, 16, 64])]; + tensor k_9_end_mask_0 = const()[name = tensor("k_9_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor k_9_squeeze_mask_0 = const()[name = tensor("k_9_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor k_9_cast_fp16 = slice_by_index(begin = k_9_begin_0, end = k_9_end_0, end_mask = k_9_end_mask_0, squeeze_mask = k_9_squeeze_mask_0, x = qkv_5_cast_fp16)[name = tensor("k_9_cast_fp16")]; + tensor v_5_begin_0 = const()[name = tensor("v_5_begin_0"), val = tensor([0, 0, 2, 0, 0])]; + tensor v_5_end_0 = const()[name = tensor("v_5_end_0"), val = tensor([1, 1, 3, 16, 64])]; + tensor v_5_end_mask_0 = const()[name = tensor("v_5_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor v_5_squeeze_mask_0 = const()[name = tensor("v_5_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor v_5_cast_fp16 = slice_by_index(begin = v_5_begin_0, end = v_5_end_0, end_mask = v_5_end_mask_0, squeeze_mask = v_5_squeeze_mask_0, x = qkv_5_cast_fp16)[name = tensor("v_5_cast_fp16")]; + tensor var_912 = const()[name = tensor("op_912"), val = tensor([1, 1, 1, 1])]; + tensor position2_to_fp16_dtype_0 = const()[name = tensor("position2_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor position2_to_fp16 = cast(dtype = position2_to_fp16_dtype_0, x = position2)[name = tensor("cast_129")]; + tensor ts_17_cast_fp16 = reshape(shape = var_912, x = position2_to_fp16)[name = tensor("ts_17_cast_fp16")]; + tensor var_919 = const()[name = tensor("op_919"), val = tensor([1, 1, 16, 32, 2])]; + tensor q_complex_5_cast_fp16 = reshape(shape = var_919, x = q_13_cast_fp16)[name = tensor("q_complex_5_cast_fp16")]; + tensor var_926 = const()[name = tensor("op_926"), val = tensor([1, 1, 16, 32, 2])]; + tensor k_complex_5_cast_fp16 = reshape(shape = var_926, x = k_9_cast_fp16)[name = tensor("k_complex_5_cast_fp16")]; + tensor var_930_begin_0 = const()[name = tensor("op_930_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_930_end_0 = const()[name = tensor("op_930_end_0"), val = tensor([1, 1, 16, 32, 1])]; + tensor var_930_end_mask_0 = const()[name = tensor("op_930_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_930_squeeze_mask_0 = const()[name = tensor("op_930_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_930_cast_fp16 = slice_by_index(begin = var_930_begin_0, end = var_930_end_0, end_mask = var_930_end_mask_0, squeeze_mask = var_930_squeeze_mask_0, x = q_complex_5_cast_fp16)[name = tensor("op_930_cast_fp16")]; + tensor var_938_begin_0 = const()[name = tensor("op_938_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_938_end_0 = const()[name = tensor("op_938_end_0"), val = tensor([1, 1, 16, 32, 2])]; + tensor var_938_end_mask_0 = const()[name = tensor("op_938_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_938_squeeze_mask_0 = const()[name = tensor("op_938_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_938_cast_fp16 = slice_by_index(begin = var_938_begin_0, end = var_938_end_0, end_mask = var_938_end_mask_0, squeeze_mask = var_938_squeeze_mask_0, x = q_complex_5_cast_fp16)[name = tensor("op_938_cast_fp16")]; + tensor var_946_begin_0 = const()[name = tensor("op_946_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_946_end_0 = const()[name = tensor("op_946_end_0"), val = tensor([1, 1, 16, 32, 1])]; + tensor var_946_end_mask_0 = const()[name = tensor("op_946_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_946_squeeze_mask_0 = const()[name = tensor("op_946_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_946_cast_fp16 = slice_by_index(begin = var_946_begin_0, end = var_946_end_0, end_mask = var_946_end_mask_0, squeeze_mask = var_946_squeeze_mask_0, x = k_complex_5_cast_fp16)[name = tensor("op_946_cast_fp16")]; + tensor var_954_begin_0 = const()[name = tensor("op_954_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_954_end_0 = const()[name = tensor("op_954_end_0"), val = tensor([1, 1, 16, 32, 2])]; + tensor var_954_end_mask_0 = const()[name = tensor("op_954_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_954_squeeze_mask_0 = const()[name = tensor("op_954_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_954_cast_fp16 = slice_by_index(begin = var_954_begin_0, end = var_954_end_0, end_mask = var_954_end_mask_0, squeeze_mask = var_954_squeeze_mask_0, x = k_complex_5_cast_fp16)[name = tensor("op_954_cast_fp16")]; + tensor freqs_5_to_fp16 = const()[name = tensor("freqs_5_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(57777024)))]; + tensor var_960_cast_fp16 = mul(x = freqs_5_to_fp16, y = ts_17_cast_fp16)[name = tensor("op_960_cast_fp16")]; + tensor rotr_5_cast_fp16 = cos(x = var_960_cast_fp16)[name = tensor("rotr_5_cast_fp16")]; + tensor roti_5_cast_fp16 = sin(x = var_960_cast_fp16)[name = tensor("roti_5_cast_fp16")]; + tensor var_964_cast_fp16 = mul(x = var_930_cast_fp16, y = rotr_5_cast_fp16)[name = tensor("op_964_cast_fp16")]; + tensor var_965_cast_fp16 = mul(x = var_938_cast_fp16, y = roti_5_cast_fp16)[name = tensor("op_965_cast_fp16")]; + tensor qor_9_cast_fp16 = sub(x = var_964_cast_fp16, y = var_965_cast_fp16)[name = tensor("qor_9_cast_fp16")]; + tensor var_968_cast_fp16 = mul(x = var_930_cast_fp16, y = roti_5_cast_fp16)[name = tensor("op_968_cast_fp16")]; + tensor var_969_cast_fp16 = mul(x = var_938_cast_fp16, y = rotr_5_cast_fp16)[name = tensor("op_969_cast_fp16")]; + tensor qoi_9_cast_fp16 = add(x = var_968_cast_fp16, y = var_969_cast_fp16)[name = tensor("qoi_9_cast_fp16")]; + tensor var_972_cast_fp16 = mul(x = var_946_cast_fp16, y = rotr_5_cast_fp16)[name = tensor("op_972_cast_fp16")]; + tensor var_973_cast_fp16 = mul(x = var_954_cast_fp16, y = roti_5_cast_fp16)[name = tensor("op_973_cast_fp16")]; + tensor kor_9_cast_fp16 = sub(x = var_972_cast_fp16, y = var_973_cast_fp16)[name = tensor("kor_9_cast_fp16")]; + tensor var_976_cast_fp16 = mul(x = var_946_cast_fp16, y = roti_5_cast_fp16)[name = tensor("op_976_cast_fp16")]; + tensor var_977_cast_fp16 = mul(x = var_954_cast_fp16, y = rotr_5_cast_fp16)[name = tensor("op_977_cast_fp16")]; + tensor koi_9_cast_fp16 = add(x = var_976_cast_fp16, y = var_977_cast_fp16)[name = tensor("koi_9_cast_fp16")]; + tensor qo_5_axis_0 = const()[name = tensor("qo_5_axis_0"), val = tensor(-1)]; + tensor qo_5_cast_fp16 = stack(axis = qo_5_axis_0, values = (qor_9_cast_fp16, qoi_9_cast_fp16))[name = tensor("qo_5_cast_fp16")]; + tensor ko_5_axis_0 = const()[name = tensor("ko_5_axis_0"), val = tensor(-1)]; + tensor ko_5_cast_fp16 = stack(axis = ko_5_axis_0, values = (kor_9_cast_fp16, koi_9_cast_fp16))[name = tensor("ko_5_cast_fp16")]; + tensor var_1010 = const()[name = tensor("op_1010"), val = tensor([1, 1, 16, 64])]; + tensor q_15_cast_fp16 = reshape(shape = var_1010, x = qo_5_cast_fp16)[name = tensor("q_15_cast_fp16")]; + tensor var_1016 = const()[name = tensor("op_1016"), val = tensor([1, 1, 16, 64])]; + tensor k_11_cast_fp16 = reshape(shape = var_1016, x = ko_5_cast_fp16)[name = tensor("k_11_cast_fp16")]; + tensor _inversed_1041_y_0_to_fp16 = const()[name = tensor("_inversed_1041_y_0_to_fp16"), val = tensor(0x1p-9)]; + tensor _inversed_1041_cast_fp16 = mul(x = ts_17_cast_fp16, y = _inversed_1041_y_0_to_fp16)[name = tensor("_inversed_1041_cast_fp16")]; + tensor var_1042_cast_fp16 = floor(x = _inversed_1041_cast_fp16)[name = tensor("op_1042_cast_fp16")]; + tensor var_1043_to_fp16 = const()[name = tensor("op_1043_to_fp16"), val = tensor(0x1p+9)]; + tensor var_1044_cast_fp16 = mul(x = var_1042_cast_fp16, y = var_1043_to_fp16)[name = tensor("op_1044_cast_fp16")]; + tensor write_indices_float_11_cast_fp16 = sub(x = ts_17_cast_fp16, y = var_1044_cast_fp16)[name = tensor("write_indices_float_11_cast_fp16")]; + tensor var_1051_dtype_0 = const()[name = tensor("op_1051_dtype_0"), val = tensor("int32")]; + tensor write_indices_5_reps_0 = const()[name = tensor("write_indices_5_reps_0"), val = tensor([1, 1, 16, 64])]; + tensor write_indices_float_11_cast_fp16_to_int32 = cast(dtype = var_1051_dtype_0, x = write_indices_float_11_cast_fp16)[name = tensor("cast_128")]; + tensor write_indices_5 = tile(reps = write_indices_5_reps_0, x = write_indices_float_11_cast_fp16_to_int32)[name = tensor("write_indices_5")]; + tensor var_1061_begin_0 = const()[name = tensor("op_1061_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_1061_end_0 = const()[name = tensor("op_1061_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor var_1061_end_mask_0 = const()[name = tensor("op_1061_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_1061_squeeze_mask_0 = const()[name = tensor("op_1061_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor cache2_to_fp16_dtype_0 = const()[name = tensor("cache2_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor cache2_to_fp16 = cast(dtype = cache2_to_fp16_dtype_0, x = cache2)[name = tensor("cast_127")]; + tensor var_1061_cast_fp16 = slice_by_index(begin = var_1061_begin_0, end = var_1061_end_0, end_mask = var_1061_end_mask_0, squeeze_mask = var_1061_squeeze_mask_0, x = cache2_to_fp16)[name = tensor("op_1061_cast_fp16")]; + tensor var_1063_axis_0 = const()[name = tensor("op_1063_axis_0"), val = tensor(1)]; + tensor var_1063_mode_0 = const()[name = tensor("op_1063_mode_0"), val = tensor("update")]; + tensor var_1063_validate_indices_0 = const()[name = tensor("op_1063_validate_indices_0"), val = tensor(false)]; + tensor var_1063_cast_fp16 = scatter_along_axis(axis = var_1063_axis_0, data = var_1061_cast_fp16, indices = write_indices_5, mode = var_1063_mode_0, updates = k_11_cast_fp16, validate_indices = var_1063_validate_indices_0)[name = tensor("op_1063_cast_fp16")]; + tensor concat_16 = const()[name = tensor("concat_16"), val = tensor([0, 0, 0, 0, 0])]; + tensor concat_17 = const()[name = tensor("concat_17"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_5_internal_tensor_assign_1_stride_0 = const()[name = tensor("new_cache_5_internal_tensor_assign_1_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_5_internal_tensor_assign_1_begin_mask_0 = const()[name = tensor("new_cache_5_internal_tensor_assign_1_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_5_internal_tensor_assign_1_end_mask_0 = const()[name = tensor("new_cache_5_internal_tensor_assign_1_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_5_internal_tensor_assign_1_squeeze_mask_0 = const()[name = tensor("new_cache_5_internal_tensor_assign_1_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_16 = const()[name = tensor("shape_16"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_4 = const()[name = tensor("reduce_prod_4"), val = tensor(1048576)]; + tensor range_1d_4_start_0 = const()[name = tensor("range_1d_4_start_0"), val = tensor(0)]; + tensor range_1d_4_step_0 = const()[name = tensor("range_1d_4_step_0"), val = tensor(1)]; + tensor range_1d_4 = range_1d(end = reduce_prod_4, start = range_1d_4_start_0, step = range_1d_4_step_0)[name = tensor("range_1d_4")]; + tensor reshape_20 = reshape(shape = shape_16, x = range_1d_4)[name = tensor("reshape_20")]; + tensor slice_by_index_4 = slice_by_index(begin = concat_16, begin_mask = new_cache_5_internal_tensor_assign_1_begin_mask_0, end = concat_17, end_mask = new_cache_5_internal_tensor_assign_1_end_mask_0, squeeze_mask = new_cache_5_internal_tensor_assign_1_squeeze_mask_0, stride = new_cache_5_internal_tensor_assign_1_stride_0, x = reshape_20)[name = tensor("slice_by_index_4")]; + tensor reshape_21_shape_0 = const()[name = tensor("reshape_21_shape_0"), val = tensor([-1])]; + tensor reshape_21 = reshape(shape = reshape_21_shape_0, x = slice_by_index_4)[name = tensor("reshape_21")]; + tensor reshape_22_shape_0 = const()[name = tensor("reshape_22_shape_0"), val = tensor([-1])]; + tensor reshape_22_cast_fp16 = reshape(shape = reshape_22_shape_0, x = var_1063_cast_fp16)[name = tensor("reshape_22_cast_fp16")]; + tensor reshape_23_shape_0 = const()[name = tensor("reshape_23_shape_0"), val = tensor([-1])]; + tensor reshape_23_cast_fp16 = reshape(shape = reshape_23_shape_0, x = cache2_to_fp16)[name = tensor("reshape_23_cast_fp16")]; + tensor scatter_4_mode_0 = const()[name = tensor("scatter_4_mode_0"), val = tensor("update")]; + tensor scatter_4_axis_0 = const()[name = tensor("scatter_4_axis_0"), val = tensor(0)]; + tensor scatter_4_validate_indices_0 = const()[name = tensor("scatter_4_validate_indices_0"), val = tensor(false)]; + tensor scatter_4_cast_fp16 = scatter(axis = scatter_4_axis_0, data = reshape_23_cast_fp16, indices = reshape_21, mode = scatter_4_mode_0, updates = reshape_22_cast_fp16, validate_indices = scatter_4_validate_indices_0)[name = tensor("scatter_4_cast_fp16")]; + tensor reshape_24_cast_fp16 = reshape(shape = shape_16, x = scatter_4_cast_fp16)[name = tensor("reshape_24_cast_fp16")]; + tensor var_1071_begin_0 = const()[name = tensor("op_1071_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor var_1071_end_0 = const()[name = tensor("op_1071_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor var_1071_end_mask_0 = const()[name = tensor("op_1071_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_1071_squeeze_mask_0 = const()[name = tensor("op_1071_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor var_1071_cast_fp16 = slice_by_index(begin = var_1071_begin_0, end = var_1071_end_0, end_mask = var_1071_end_mask_0, squeeze_mask = var_1071_squeeze_mask_0, x = reshape_24_cast_fp16)[name = tensor("op_1071_cast_fp16")]; + tensor var_1073_axis_0 = const()[name = tensor("op_1073_axis_0"), val = tensor(1)]; + tensor var_1073_mode_0 = const()[name = tensor("op_1073_mode_0"), val = tensor("update")]; + tensor var_1073_validate_indices_0 = const()[name = tensor("op_1073_validate_indices_0"), val = tensor(false)]; + tensor var_1073_cast_fp16 = scatter_along_axis(axis = var_1073_axis_0, data = var_1071_cast_fp16, indices = write_indices_5, mode = var_1073_mode_0, updates = v_5_cast_fp16, validate_indices = var_1073_validate_indices_0)[name = tensor("op_1073_cast_fp16")]; + tensor concat_18 = const()[name = tensor("concat_18"), val = tensor([1, 0, 0, 0, 0])]; + tensor concat_19 = const()[name = tensor("concat_19"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_5_internal_tensor_assign_2_stride_0 = const()[name = tensor("new_cache_5_internal_tensor_assign_2_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_5_internal_tensor_assign_2_begin_mask_0 = const()[name = tensor("new_cache_5_internal_tensor_assign_2_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_5_internal_tensor_assign_2_end_mask_0 = const()[name = tensor("new_cache_5_internal_tensor_assign_2_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_5_internal_tensor_assign_2_squeeze_mask_0 = const()[name = tensor("new_cache_5_internal_tensor_assign_2_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_17 = const()[name = tensor("shape_17"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_5 = const()[name = tensor("reduce_prod_5"), val = tensor(1048576)]; + tensor range_1d_5_start_0 = const()[name = tensor("range_1d_5_start_0"), val = tensor(0)]; + tensor range_1d_5_step_0 = const()[name = tensor("range_1d_5_step_0"), val = tensor(1)]; + tensor range_1d_5 = range_1d(end = reduce_prod_5, start = range_1d_5_start_0, step = range_1d_5_step_0)[name = tensor("range_1d_5")]; + tensor reshape_25 = reshape(shape = shape_17, x = range_1d_5)[name = tensor("reshape_25")]; + tensor slice_by_index_5 = slice_by_index(begin = concat_18, begin_mask = new_cache_5_internal_tensor_assign_2_begin_mask_0, end = concat_19, end_mask = new_cache_5_internal_tensor_assign_2_end_mask_0, squeeze_mask = new_cache_5_internal_tensor_assign_2_squeeze_mask_0, stride = new_cache_5_internal_tensor_assign_2_stride_0, x = reshape_25)[name = tensor("slice_by_index_5")]; + tensor reshape_26_shape_0 = const()[name = tensor("reshape_26_shape_0"), val = tensor([-1])]; + tensor reshape_26 = reshape(shape = reshape_26_shape_0, x = slice_by_index_5)[name = tensor("reshape_26")]; + tensor reshape_27_shape_0 = const()[name = tensor("reshape_27_shape_0"), val = tensor([-1])]; + tensor reshape_27_cast_fp16 = reshape(shape = reshape_27_shape_0, x = var_1073_cast_fp16)[name = tensor("reshape_27_cast_fp16")]; + tensor reshape_28_shape_0 = const()[name = tensor("reshape_28_shape_0"), val = tensor([-1])]; + tensor reshape_28_cast_fp16 = reshape(shape = reshape_28_shape_0, x = reshape_24_cast_fp16)[name = tensor("reshape_28_cast_fp16")]; + tensor scatter_5_mode_0 = const()[name = tensor("scatter_5_mode_0"), val = tensor("update")]; + tensor scatter_5_axis_0 = const()[name = tensor("scatter_5_axis_0"), val = tensor(0)]; + tensor scatter_5_validate_indices_0 = const()[name = tensor("scatter_5_validate_indices_0"), val = tensor(false)]; + tensor scatter_5_cast_fp16 = scatter(axis = scatter_5_axis_0, data = reshape_28_cast_fp16, indices = reshape_26, mode = scatter_5_mode_0, updates = reshape_27_cast_fp16, validate_indices = scatter_5_validate_indices_0)[name = tensor("scatter_5_cast_fp16")]; + tensor reshape_29_cast_fp16 = reshape(shape = shape_17, x = scatter_5_cast_fp16)[name = tensor("reshape_29_cast_fp16")]; + tensor reshape_29_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("reshape_29_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor keys_13_begin_0 = const()[name = tensor("keys_13_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor keys_13_end_0 = const()[name = tensor("keys_13_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor keys_13_end_mask_0 = const()[name = tensor("keys_13_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor keys_13_squeeze_mask_0 = const()[name = tensor("keys_13_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor keys_13_cast_fp16 = slice_by_index(begin = keys_13_begin_0, end = keys_13_end_0, end_mask = keys_13_end_mask_0, squeeze_mask = keys_13_squeeze_mask_0, x = reshape_29_cast_fp16)[name = tensor("keys_13_cast_fp16")]; + tensor values_13_begin_0 = const()[name = tensor("values_13_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor values_13_end_0 = const()[name = tensor("values_13_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor values_13_end_mask_0 = const()[name = tensor("values_13_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor values_13_squeeze_mask_0 = const()[name = tensor("values_13_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor values_13_cast_fp16 = slice_by_index(begin = values_13_begin_0, end = values_13_end_0, end_mask = values_13_end_mask_0, squeeze_mask = values_13_squeeze_mask_0, x = reshape_29_cast_fp16)[name = tensor("values_13_cast_fp16")]; + tensor var_1085_cast_fp16 = not_equal(x = keys_13_cast_fp16, y = keys_13_cast_fp16)[name = tensor("op_1085_cast_fp16")]; + tensor keys_15_cast_fp16 = select(a = var_319_to_fp16, b = keys_13_cast_fp16, cond = var_1085_cast_fp16)[name = tensor("keys_15_cast_fp16")]; + tensor var_1093_cast_fp16 = not_equal(x = values_13_cast_fp16, y = values_13_cast_fp16)[name = tensor("op_1093_cast_fp16")]; + tensor values_15_cast_fp16 = select(a = var_319_to_fp16, b = values_13_cast_fp16, cond = var_1093_cast_fp16)[name = tensor("values_15_cast_fp16")]; + tensor var_1117 = const()[name = tensor("op_1117"), val = tensor([0, 2, 1, 3])]; + tensor var_1133 = const()[name = tensor("op_1133"), val = tensor([1, 1, 1])]; + tensor var_1134_cast_fp16 = reshape(shape = var_1133, x = position2_to_fp16)[name = tensor("op_1134_cast_fp16")]; + tensor var_1154_to_fp16 = const()[name = tensor("op_1154_to_fp16"), val = tensor(0x1p+0)]; + tensor valid_len_5_cast_fp16 = add(x = var_1134_cast_fp16, y = var_1154_to_fp16)[name = tensor("valid_len_5_cast_fp16")]; + tensor valid_mask_5_cast_fp16 = less(x = k_positions_1_promoted_to_fp16, y = valid_len_5_cast_fp16)[name = tensor("valid_mask_5_cast_fp16")]; + tensor causal_mask_5_cast_fp16 = less_equal(x = k_positions_1_promoted_to_fp16, y = var_1134_cast_fp16)[name = tensor("causal_mask_5_cast_fp16")]; + tensor attn_mask_9 = logical_and(x = valid_mask_5_cast_fp16, y = causal_mask_5_cast_fp16)[name = tensor("attn_mask_9")]; + tensor attn_mask_11_axes_0 = const()[name = tensor("attn_mask_11_axes_0"), val = tensor([1])]; + tensor attn_mask_11 = expand_dims(axes = attn_mask_11_axes_0, x = attn_mask_9)[name = tensor("attn_mask_11")]; + tensor var_1172_transpose_x_0 = const()[name = tensor("op_1172_transpose_x_0"), val = tensor(false)]; + tensor var_1172_transpose_y_0 = const()[name = tensor("op_1172_transpose_y_0"), val = tensor(false)]; + tensor transpose_22_perm_0 = const()[name = tensor("transpose_22_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_23_perm_0 = const()[name = tensor("transpose_23_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_23 = transpose(perm = transpose_23_perm_0, x = keys_15_cast_fp16)[name = tensor("transpose_43")]; + tensor transpose_22 = transpose(perm = transpose_22_perm_0, x = q_15_cast_fp16)[name = tensor("transpose_44")]; + tensor var_1172_cast_fp16 = matmul(transpose_x = var_1172_transpose_x_0, transpose_y = var_1172_transpose_y_0, x = transpose_22, y = transpose_23)[name = tensor("op_1172_cast_fp16")]; + tensor var_1166_to_fp16 = const()[name = tensor("op_1166_to_fp16"), val = tensor([0x1p-3])]; + tensor attn_weights_13_cast_fp16 = mul(x = var_1172_cast_fp16, y = var_1166_to_fp16)[name = tensor("attn_weights_13_cast_fp16")]; + tensor var_1174 = logical_not(x = attn_mask_11)[name = tensor("op_1174")]; + tensor var_1175_to_fp16 = const()[name = tensor("op_1175_to_fp16"), val = tensor(-inf)]; + tensor attn_weights_15_cast_fp16 = select(a = var_1175_to_fp16, b = attn_weights_13_cast_fp16, cond = var_1174)[name = tensor("attn_weights_15_cast_fp16")]; + tensor var_1177 = const()[name = tensor("op_1177"), val = tensor(-1)]; + tensor attn_weights_17_cast_fp16 = softmax(axis = var_1177, x = attn_weights_15_cast_fp16)[name = tensor("attn_weights_17_cast_fp16")]; + tensor attn_output_5_transpose_x_0 = const()[name = tensor("attn_output_5_transpose_x_0"), val = tensor(false)]; + tensor attn_output_5_transpose_y_0 = const()[name = tensor("attn_output_5_transpose_y_0"), val = tensor(false)]; + tensor values_17_cast_fp16 = transpose(perm = var_1117, x = values_15_cast_fp16)[name = tensor("transpose_45")]; + tensor attn_output_5_cast_fp16 = matmul(transpose_x = attn_output_5_transpose_x_0, transpose_y = attn_output_5_transpose_y_0, x = attn_weights_17_cast_fp16, y = values_17_cast_fp16)[name = tensor("attn_output_5_cast_fp16")]; + tensor var_1185 = const()[name = tensor("op_1185"), val = tensor([0, 2, 1, 3])]; + tensor var_1190 = const()[name = tensor("op_1190"), val = tensor([1, 1, 1024])]; + tensor var_1186_cast_fp16 = transpose(perm = var_1185, x = attn_output_5_cast_fp16)[name = tensor("transpose_42")]; + tensor input_31_cast_fp16 = reshape(shape = var_1190, x = var_1186_cast_fp16)[name = tensor("input_31_cast_fp16")]; + tensor attn2_out_proj_weight_to_fp16 = const()[name = tensor("attn2_out_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(57777152)))]; + tensor linear_10_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = attn2_out_proj_weight_to_fp16, x = input_31_cast_fp16)[name = tensor("linear_10_cast_fp16")]; + tensor var_1196_to_fp16 = const()[name = tensor("op_1196_to_fp16"), val = tensor(0x1p+0)]; + tensor var_1197_cast_fp16 = add(x = position2_to_fp16, y = var_1196_to_fp16)[name = tensor("op_1197_cast_fp16")]; + tensor var_1197_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_1197_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor input_33_cast_fp16 = add(x = input_27_cast_fp16, y = linear_10_cast_fp16)[name = tensor("input_33_cast_fp16")]; + tensor input_35_axes_0 = const()[name = tensor("input_35_axes_0"), val = tensor([-1])]; + tensor norm2_2_weight_to_fp16 = const()[name = tensor("norm2_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(59874368)))]; + tensor norm2_2_bias_to_fp16 = const()[name = tensor("norm2_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(59876480)))]; + tensor var_1201_to_fp16 = const()[name = tensor("op_1201_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_35_cast_fp16 = layer_norm(axes = input_35_axes_0, beta = norm2_2_bias_to_fp16, epsilon = var_1201_to_fp16, gamma = norm2_2_weight_to_fp16, x = input_33_cast_fp16)[name = tensor("input_35_cast_fp16")]; + tensor linear2_1_weight_to_fp16 = const()[name = tensor("linear2_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(59878592)))]; + tensor linear_11_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = linear2_1_weight_to_fp16, x = input_35_cast_fp16)[name = tensor("linear_11_cast_fp16")]; + tensor input_37_mode_0 = const()[name = tensor("input_37_mode_0"), val = tensor("EXACT")]; + tensor input_37_cast_fp16 = gelu(mode = input_37_mode_0, x = linear_11_cast_fp16)[name = tensor("input_37_cast_fp16")]; + tensor linear2_2_weight_to_fp16 = const()[name = tensor("linear2_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(68267264)))]; + tensor linear_12_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = linear2_2_weight_to_fp16, x = input_37_cast_fp16)[name = tensor("linear_12_cast_fp16")]; + tensor input_39_cast_fp16 = add(x = input_33_cast_fp16, y = linear_12_cast_fp16)[name = tensor("input_39_cast_fp16")]; + tensor input_41_axes_0 = const()[name = tensor("input_41_axes_0"), val = tensor([-1])]; + tensor norm3_1_weight_to_fp16 = const()[name = tensor("norm3_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(76655936)))]; + tensor norm3_1_bias_to_fp16 = const()[name = tensor("norm3_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(76658048)))]; + tensor var_1218_to_fp16 = const()[name = tensor("op_1218_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_41_cast_fp16 = layer_norm(axes = input_41_axes_0, beta = norm3_1_bias_to_fp16, epsilon = var_1218_to_fp16, gamma = norm3_1_weight_to_fp16, x = input_39_cast_fp16)[name = tensor("input_41_cast_fp16")]; + tensor attn3_in_proj_weight_to_fp16 = const()[name = tensor("attn3_in_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(76660160)))]; + tensor linear_13_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = attn3_in_proj_weight_to_fp16, x = input_41_cast_fp16)[name = tensor("linear_13_cast_fp16")]; + tensor var_1232 = const()[name = tensor("op_1232"), val = tensor([1, 1, 3, 16, 64])]; + tensor qkv_7_cast_fp16 = reshape(shape = var_1232, x = linear_13_cast_fp16)[name = tensor("qkv_7_cast_fp16")]; + tensor q_19_begin_0 = const()[name = tensor("q_19_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor q_19_end_0 = const()[name = tensor("q_19_end_0"), val = tensor([1, 1, 1, 16, 64])]; + tensor q_19_end_mask_0 = const()[name = tensor("q_19_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor q_19_squeeze_mask_0 = const()[name = tensor("q_19_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor q_19_cast_fp16 = slice_by_index(begin = q_19_begin_0, end = q_19_end_0, end_mask = q_19_end_mask_0, squeeze_mask = q_19_squeeze_mask_0, x = qkv_7_cast_fp16)[name = tensor("q_19_cast_fp16")]; + tensor k_13_begin_0 = const()[name = tensor("k_13_begin_0"), val = tensor([0, 0, 1, 0, 0])]; + tensor k_13_end_0 = const()[name = tensor("k_13_end_0"), val = tensor([1, 1, 2, 16, 64])]; + tensor k_13_end_mask_0 = const()[name = tensor("k_13_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor k_13_squeeze_mask_0 = const()[name = tensor("k_13_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor k_13_cast_fp16 = slice_by_index(begin = k_13_begin_0, end = k_13_end_0, end_mask = k_13_end_mask_0, squeeze_mask = k_13_squeeze_mask_0, x = qkv_7_cast_fp16)[name = tensor("k_13_cast_fp16")]; + tensor v_7_begin_0 = const()[name = tensor("v_7_begin_0"), val = tensor([0, 0, 2, 0, 0])]; + tensor v_7_end_0 = const()[name = tensor("v_7_end_0"), val = tensor([1, 1, 3, 16, 64])]; + tensor v_7_end_mask_0 = const()[name = tensor("v_7_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor v_7_squeeze_mask_0 = const()[name = tensor("v_7_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor v_7_cast_fp16 = slice_by_index(begin = v_7_begin_0, end = v_7_end_0, end_mask = v_7_end_mask_0, squeeze_mask = v_7_squeeze_mask_0, x = qkv_7_cast_fp16)[name = tensor("v_7_cast_fp16")]; + tensor var_1298 = const()[name = tensor("op_1298"), val = tensor([1, 1, 1, 1])]; + tensor position3_to_fp16_dtype_0 = const()[name = tensor("position3_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor position3_to_fp16 = cast(dtype = position3_to_fp16_dtype_0, x = position3)[name = tensor("cast_124")]; + tensor ts_23_cast_fp16 = reshape(shape = var_1298, x = position3_to_fp16)[name = tensor("ts_23_cast_fp16")]; + tensor var_1305 = const()[name = tensor("op_1305"), val = tensor([1, 1, 16, 32, 2])]; + tensor q_complex_7_cast_fp16 = reshape(shape = var_1305, x = q_19_cast_fp16)[name = tensor("q_complex_7_cast_fp16")]; + tensor var_1312 = const()[name = tensor("op_1312"), val = tensor([1, 1, 16, 32, 2])]; + tensor k_complex_7_cast_fp16 = reshape(shape = var_1312, x = k_13_cast_fp16)[name = tensor("k_complex_7_cast_fp16")]; + tensor var_1316_begin_0 = const()[name = tensor("op_1316_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_1316_end_0 = const()[name = tensor("op_1316_end_0"), val = tensor([1, 1, 16, 32, 1])]; + tensor var_1316_end_mask_0 = const()[name = tensor("op_1316_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_1316_squeeze_mask_0 = const()[name = tensor("op_1316_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_1316_cast_fp16 = slice_by_index(begin = var_1316_begin_0, end = var_1316_end_0, end_mask = var_1316_end_mask_0, squeeze_mask = var_1316_squeeze_mask_0, x = q_complex_7_cast_fp16)[name = tensor("op_1316_cast_fp16")]; + tensor var_1324_begin_0 = const()[name = tensor("op_1324_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_1324_end_0 = const()[name = tensor("op_1324_end_0"), val = tensor([1, 1, 16, 32, 2])]; + tensor var_1324_end_mask_0 = const()[name = tensor("op_1324_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_1324_squeeze_mask_0 = const()[name = tensor("op_1324_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_1324_cast_fp16 = slice_by_index(begin = var_1324_begin_0, end = var_1324_end_0, end_mask = var_1324_end_mask_0, squeeze_mask = var_1324_squeeze_mask_0, x = q_complex_7_cast_fp16)[name = tensor("op_1324_cast_fp16")]; + tensor var_1332_begin_0 = const()[name = tensor("op_1332_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_1332_end_0 = const()[name = tensor("op_1332_end_0"), val = tensor([1, 1, 16, 32, 1])]; + tensor var_1332_end_mask_0 = const()[name = tensor("op_1332_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_1332_squeeze_mask_0 = const()[name = tensor("op_1332_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_1332_cast_fp16 = slice_by_index(begin = var_1332_begin_0, end = var_1332_end_0, end_mask = var_1332_end_mask_0, squeeze_mask = var_1332_squeeze_mask_0, x = k_complex_7_cast_fp16)[name = tensor("op_1332_cast_fp16")]; + tensor var_1340_begin_0 = const()[name = tensor("op_1340_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_1340_end_0 = const()[name = tensor("op_1340_end_0"), val = tensor([1, 1, 16, 32, 2])]; + tensor var_1340_end_mask_0 = const()[name = tensor("op_1340_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_1340_squeeze_mask_0 = const()[name = tensor("op_1340_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_1340_cast_fp16 = slice_by_index(begin = var_1340_begin_0, end = var_1340_end_0, end_mask = var_1340_end_mask_0, squeeze_mask = var_1340_squeeze_mask_0, x = k_complex_7_cast_fp16)[name = tensor("op_1340_cast_fp16")]; + tensor freqs_7_to_fp16 = const()[name = tensor("freqs_7_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(82951680)))]; + tensor var_1346_cast_fp16 = mul(x = freqs_7_to_fp16, y = ts_23_cast_fp16)[name = tensor("op_1346_cast_fp16")]; + tensor rotr_7_cast_fp16 = cos(x = var_1346_cast_fp16)[name = tensor("rotr_7_cast_fp16")]; + tensor roti_7_cast_fp16 = sin(x = var_1346_cast_fp16)[name = tensor("roti_7_cast_fp16")]; + tensor var_1350_cast_fp16 = mul(x = var_1316_cast_fp16, y = rotr_7_cast_fp16)[name = tensor("op_1350_cast_fp16")]; + tensor var_1351_cast_fp16 = mul(x = var_1324_cast_fp16, y = roti_7_cast_fp16)[name = tensor("op_1351_cast_fp16")]; + tensor qor_13_cast_fp16 = sub(x = var_1350_cast_fp16, y = var_1351_cast_fp16)[name = tensor("qor_13_cast_fp16")]; + tensor var_1354_cast_fp16 = mul(x = var_1316_cast_fp16, y = roti_7_cast_fp16)[name = tensor("op_1354_cast_fp16")]; + tensor var_1355_cast_fp16 = mul(x = var_1324_cast_fp16, y = rotr_7_cast_fp16)[name = tensor("op_1355_cast_fp16")]; + tensor qoi_13_cast_fp16 = add(x = var_1354_cast_fp16, y = var_1355_cast_fp16)[name = tensor("qoi_13_cast_fp16")]; + tensor var_1358_cast_fp16 = mul(x = var_1332_cast_fp16, y = rotr_7_cast_fp16)[name = tensor("op_1358_cast_fp16")]; + tensor var_1359_cast_fp16 = mul(x = var_1340_cast_fp16, y = roti_7_cast_fp16)[name = tensor("op_1359_cast_fp16")]; + tensor kor_13_cast_fp16 = sub(x = var_1358_cast_fp16, y = var_1359_cast_fp16)[name = tensor("kor_13_cast_fp16")]; + tensor var_1362_cast_fp16 = mul(x = var_1332_cast_fp16, y = roti_7_cast_fp16)[name = tensor("op_1362_cast_fp16")]; + tensor var_1363_cast_fp16 = mul(x = var_1340_cast_fp16, y = rotr_7_cast_fp16)[name = tensor("op_1363_cast_fp16")]; + tensor koi_13_cast_fp16 = add(x = var_1362_cast_fp16, y = var_1363_cast_fp16)[name = tensor("koi_13_cast_fp16")]; + tensor qo_7_axis_0 = const()[name = tensor("qo_7_axis_0"), val = tensor(-1)]; + tensor qo_7_cast_fp16 = stack(axis = qo_7_axis_0, values = (qor_13_cast_fp16, qoi_13_cast_fp16))[name = tensor("qo_7_cast_fp16")]; + tensor ko_7_axis_0 = const()[name = tensor("ko_7_axis_0"), val = tensor(-1)]; + tensor ko_7_cast_fp16 = stack(axis = ko_7_axis_0, values = (kor_13_cast_fp16, koi_13_cast_fp16))[name = tensor("ko_7_cast_fp16")]; + tensor var_1396 = const()[name = tensor("op_1396"), val = tensor([1, 1, 16, 64])]; + tensor q_21_cast_fp16 = reshape(shape = var_1396, x = qo_7_cast_fp16)[name = tensor("q_21_cast_fp16")]; + tensor var_1402 = const()[name = tensor("op_1402"), val = tensor([1, 1, 16, 64])]; + tensor k_15_cast_fp16 = reshape(shape = var_1402, x = ko_7_cast_fp16)[name = tensor("k_15_cast_fp16")]; + tensor _inversed_1427_y_0_to_fp16 = const()[name = tensor("_inversed_1427_y_0_to_fp16"), val = tensor(0x1p-9)]; + tensor _inversed_1427_cast_fp16 = mul(x = ts_23_cast_fp16, y = _inversed_1427_y_0_to_fp16)[name = tensor("_inversed_1427_cast_fp16")]; + tensor var_1428_cast_fp16 = floor(x = _inversed_1427_cast_fp16)[name = tensor("op_1428_cast_fp16")]; + tensor var_1429_to_fp16 = const()[name = tensor("op_1429_to_fp16"), val = tensor(0x1p+9)]; + tensor var_1430_cast_fp16 = mul(x = var_1428_cast_fp16, y = var_1429_to_fp16)[name = tensor("op_1430_cast_fp16")]; + tensor write_indices_float_15_cast_fp16 = sub(x = ts_23_cast_fp16, y = var_1430_cast_fp16)[name = tensor("write_indices_float_15_cast_fp16")]; + tensor var_1437_dtype_0 = const()[name = tensor("op_1437_dtype_0"), val = tensor("int32")]; + tensor write_indices_7_reps_0 = const()[name = tensor("write_indices_7_reps_0"), val = tensor([1, 1, 16, 64])]; + tensor write_indices_float_15_cast_fp16_to_int32 = cast(dtype = var_1437_dtype_0, x = write_indices_float_15_cast_fp16)[name = tensor("cast_123")]; + tensor write_indices_7 = tile(reps = write_indices_7_reps_0, x = write_indices_float_15_cast_fp16_to_int32)[name = tensor("write_indices_7")]; + tensor var_1447_begin_0 = const()[name = tensor("op_1447_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_1447_end_0 = const()[name = tensor("op_1447_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor var_1447_end_mask_0 = const()[name = tensor("op_1447_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_1447_squeeze_mask_0 = const()[name = tensor("op_1447_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor cache3_to_fp16_dtype_0 = const()[name = tensor("cache3_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor cache3_to_fp16 = cast(dtype = cache3_to_fp16_dtype_0, x = cache3)[name = tensor("cast_122")]; + tensor var_1447_cast_fp16 = slice_by_index(begin = var_1447_begin_0, end = var_1447_end_0, end_mask = var_1447_end_mask_0, squeeze_mask = var_1447_squeeze_mask_0, x = cache3_to_fp16)[name = tensor("op_1447_cast_fp16")]; + tensor var_1449_axis_0 = const()[name = tensor("op_1449_axis_0"), val = tensor(1)]; + tensor var_1449_mode_0 = const()[name = tensor("op_1449_mode_0"), val = tensor("update")]; + tensor var_1449_validate_indices_0 = const()[name = tensor("op_1449_validate_indices_0"), val = tensor(false)]; + tensor var_1449_cast_fp16 = scatter_along_axis(axis = var_1449_axis_0, data = var_1447_cast_fp16, indices = write_indices_7, mode = var_1449_mode_0, updates = k_15_cast_fp16, validate_indices = var_1449_validate_indices_0)[name = tensor("op_1449_cast_fp16")]; + tensor concat_23 = const()[name = tensor("concat_23"), val = tensor([0, 0, 0, 0, 0])]; + tensor concat_24 = const()[name = tensor("concat_24"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_7_internal_tensor_assign_1_stride_0 = const()[name = tensor("new_cache_7_internal_tensor_assign_1_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_7_internal_tensor_assign_1_begin_mask_0 = const()[name = tensor("new_cache_7_internal_tensor_assign_1_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_7_internal_tensor_assign_1_end_mask_0 = const()[name = tensor("new_cache_7_internal_tensor_assign_1_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_7_internal_tensor_assign_1_squeeze_mask_0 = const()[name = tensor("new_cache_7_internal_tensor_assign_1_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_18 = const()[name = tensor("shape_18"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_6 = const()[name = tensor("reduce_prod_6"), val = tensor(1048576)]; + tensor range_1d_6_start_0 = const()[name = tensor("range_1d_6_start_0"), val = tensor(0)]; + tensor range_1d_6_step_0 = const()[name = tensor("range_1d_6_step_0"), val = tensor(1)]; + tensor range_1d_6 = range_1d(end = reduce_prod_6, start = range_1d_6_start_0, step = range_1d_6_step_0)[name = tensor("range_1d_6")]; + tensor reshape_30 = reshape(shape = shape_18, x = range_1d_6)[name = tensor("reshape_30")]; + tensor slice_by_index_6 = slice_by_index(begin = concat_23, begin_mask = new_cache_7_internal_tensor_assign_1_begin_mask_0, end = concat_24, end_mask = new_cache_7_internal_tensor_assign_1_end_mask_0, squeeze_mask = new_cache_7_internal_tensor_assign_1_squeeze_mask_0, stride = new_cache_7_internal_tensor_assign_1_stride_0, x = reshape_30)[name = tensor("slice_by_index_6")]; + tensor reshape_31_shape_0 = const()[name = tensor("reshape_31_shape_0"), val = tensor([-1])]; + tensor reshape_31 = reshape(shape = reshape_31_shape_0, x = slice_by_index_6)[name = tensor("reshape_31")]; + tensor reshape_32_shape_0 = const()[name = tensor("reshape_32_shape_0"), val = tensor([-1])]; + tensor reshape_32_cast_fp16 = reshape(shape = reshape_32_shape_0, x = var_1449_cast_fp16)[name = tensor("reshape_32_cast_fp16")]; + tensor reshape_33_shape_0 = const()[name = tensor("reshape_33_shape_0"), val = tensor([-1])]; + tensor reshape_33_cast_fp16 = reshape(shape = reshape_33_shape_0, x = cache3_to_fp16)[name = tensor("reshape_33_cast_fp16")]; + tensor scatter_6_mode_0 = const()[name = tensor("scatter_6_mode_0"), val = tensor("update")]; + tensor scatter_6_axis_0 = const()[name = tensor("scatter_6_axis_0"), val = tensor(0)]; + tensor scatter_6_validate_indices_0 = const()[name = tensor("scatter_6_validate_indices_0"), val = tensor(false)]; + tensor scatter_6_cast_fp16 = scatter(axis = scatter_6_axis_0, data = reshape_33_cast_fp16, indices = reshape_31, mode = scatter_6_mode_0, updates = reshape_32_cast_fp16, validate_indices = scatter_6_validate_indices_0)[name = tensor("scatter_6_cast_fp16")]; + tensor reshape_34_cast_fp16 = reshape(shape = shape_18, x = scatter_6_cast_fp16)[name = tensor("reshape_34_cast_fp16")]; + tensor var_1457_begin_0 = const()[name = tensor("op_1457_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor var_1457_end_0 = const()[name = tensor("op_1457_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor var_1457_end_mask_0 = const()[name = tensor("op_1457_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_1457_squeeze_mask_0 = const()[name = tensor("op_1457_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor var_1457_cast_fp16 = slice_by_index(begin = var_1457_begin_0, end = var_1457_end_0, end_mask = var_1457_end_mask_0, squeeze_mask = var_1457_squeeze_mask_0, x = reshape_34_cast_fp16)[name = tensor("op_1457_cast_fp16")]; + tensor var_1459_axis_0 = const()[name = tensor("op_1459_axis_0"), val = tensor(1)]; + tensor var_1459_mode_0 = const()[name = tensor("op_1459_mode_0"), val = tensor("update")]; + tensor var_1459_validate_indices_0 = const()[name = tensor("op_1459_validate_indices_0"), val = tensor(false)]; + tensor var_1459_cast_fp16 = scatter_along_axis(axis = var_1459_axis_0, data = var_1457_cast_fp16, indices = write_indices_7, mode = var_1459_mode_0, updates = v_7_cast_fp16, validate_indices = var_1459_validate_indices_0)[name = tensor("op_1459_cast_fp16")]; + tensor concat_25 = const()[name = tensor("concat_25"), val = tensor([1, 0, 0, 0, 0])]; + tensor concat_26 = const()[name = tensor("concat_26"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_7_internal_tensor_assign_2_stride_0 = const()[name = tensor("new_cache_7_internal_tensor_assign_2_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_7_internal_tensor_assign_2_begin_mask_0 = const()[name = tensor("new_cache_7_internal_tensor_assign_2_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_7_internal_tensor_assign_2_end_mask_0 = const()[name = tensor("new_cache_7_internal_tensor_assign_2_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_7_internal_tensor_assign_2_squeeze_mask_0 = const()[name = tensor("new_cache_7_internal_tensor_assign_2_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_19 = const()[name = tensor("shape_19"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_7 = const()[name = tensor("reduce_prod_7"), val = tensor(1048576)]; + tensor range_1d_7_start_0 = const()[name = tensor("range_1d_7_start_0"), val = tensor(0)]; + tensor range_1d_7_step_0 = const()[name = tensor("range_1d_7_step_0"), val = tensor(1)]; + tensor range_1d_7 = range_1d(end = reduce_prod_7, start = range_1d_7_start_0, step = range_1d_7_step_0)[name = tensor("range_1d_7")]; + tensor reshape_35 = reshape(shape = shape_19, x = range_1d_7)[name = tensor("reshape_35")]; + tensor slice_by_index_7 = slice_by_index(begin = concat_25, begin_mask = new_cache_7_internal_tensor_assign_2_begin_mask_0, end = concat_26, end_mask = new_cache_7_internal_tensor_assign_2_end_mask_0, squeeze_mask = new_cache_7_internal_tensor_assign_2_squeeze_mask_0, stride = new_cache_7_internal_tensor_assign_2_stride_0, x = reshape_35)[name = tensor("slice_by_index_7")]; + tensor reshape_36_shape_0 = const()[name = tensor("reshape_36_shape_0"), val = tensor([-1])]; + tensor reshape_36 = reshape(shape = reshape_36_shape_0, x = slice_by_index_7)[name = tensor("reshape_36")]; + tensor reshape_37_shape_0 = const()[name = tensor("reshape_37_shape_0"), val = tensor([-1])]; + tensor reshape_37_cast_fp16 = reshape(shape = reshape_37_shape_0, x = var_1459_cast_fp16)[name = tensor("reshape_37_cast_fp16")]; + tensor reshape_38_shape_0 = const()[name = tensor("reshape_38_shape_0"), val = tensor([-1])]; + tensor reshape_38_cast_fp16 = reshape(shape = reshape_38_shape_0, x = reshape_34_cast_fp16)[name = tensor("reshape_38_cast_fp16")]; + tensor scatter_7_mode_0 = const()[name = tensor("scatter_7_mode_0"), val = tensor("update")]; + tensor scatter_7_axis_0 = const()[name = tensor("scatter_7_axis_0"), val = tensor(0)]; + tensor scatter_7_validate_indices_0 = const()[name = tensor("scatter_7_validate_indices_0"), val = tensor(false)]; + tensor scatter_7_cast_fp16 = scatter(axis = scatter_7_axis_0, data = reshape_38_cast_fp16, indices = reshape_36, mode = scatter_7_mode_0, updates = reshape_37_cast_fp16, validate_indices = scatter_7_validate_indices_0)[name = tensor("scatter_7_cast_fp16")]; + tensor reshape_39_cast_fp16 = reshape(shape = shape_19, x = scatter_7_cast_fp16)[name = tensor("reshape_39_cast_fp16")]; + tensor reshape_39_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("reshape_39_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor keys_19_begin_0 = const()[name = tensor("keys_19_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor keys_19_end_0 = const()[name = tensor("keys_19_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor keys_19_end_mask_0 = const()[name = tensor("keys_19_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor keys_19_squeeze_mask_0 = const()[name = tensor("keys_19_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor keys_19_cast_fp16 = slice_by_index(begin = keys_19_begin_0, end = keys_19_end_0, end_mask = keys_19_end_mask_0, squeeze_mask = keys_19_squeeze_mask_0, x = reshape_39_cast_fp16)[name = tensor("keys_19_cast_fp16")]; + tensor values_19_begin_0 = const()[name = tensor("values_19_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor values_19_end_0 = const()[name = tensor("values_19_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor values_19_end_mask_0 = const()[name = tensor("values_19_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor values_19_squeeze_mask_0 = const()[name = tensor("values_19_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor values_19_cast_fp16 = slice_by_index(begin = values_19_begin_0, end = values_19_end_0, end_mask = values_19_end_mask_0, squeeze_mask = values_19_squeeze_mask_0, x = reshape_39_cast_fp16)[name = tensor("values_19_cast_fp16")]; + tensor var_1471_cast_fp16 = not_equal(x = keys_19_cast_fp16, y = keys_19_cast_fp16)[name = tensor("op_1471_cast_fp16")]; + tensor keys_21_cast_fp16 = select(a = var_319_to_fp16, b = keys_19_cast_fp16, cond = var_1471_cast_fp16)[name = tensor("keys_21_cast_fp16")]; + tensor var_1479_cast_fp16 = not_equal(x = values_19_cast_fp16, y = values_19_cast_fp16)[name = tensor("op_1479_cast_fp16")]; + tensor values_21_cast_fp16 = select(a = var_319_to_fp16, b = values_19_cast_fp16, cond = var_1479_cast_fp16)[name = tensor("values_21_cast_fp16")]; + tensor var_1503 = const()[name = tensor("op_1503"), val = tensor([0, 2, 1, 3])]; + tensor var_1519 = const()[name = tensor("op_1519"), val = tensor([1, 1, 1])]; + tensor var_1520_cast_fp16 = reshape(shape = var_1519, x = position3_to_fp16)[name = tensor("op_1520_cast_fp16")]; + tensor var_1540_to_fp16 = const()[name = tensor("op_1540_to_fp16"), val = tensor(0x1p+0)]; + tensor valid_len_7_cast_fp16 = add(x = var_1520_cast_fp16, y = var_1540_to_fp16)[name = tensor("valid_len_7_cast_fp16")]; + tensor valid_mask_7_cast_fp16 = less(x = k_positions_1_promoted_to_fp16, y = valid_len_7_cast_fp16)[name = tensor("valid_mask_7_cast_fp16")]; + tensor causal_mask_7_cast_fp16 = less_equal(x = k_positions_1_promoted_to_fp16, y = var_1520_cast_fp16)[name = tensor("causal_mask_7_cast_fp16")]; + tensor attn_mask_13 = logical_and(x = valid_mask_7_cast_fp16, y = causal_mask_7_cast_fp16)[name = tensor("attn_mask_13")]; + tensor attn_mask_15_axes_0 = const()[name = tensor("attn_mask_15_axes_0"), val = tensor([1])]; + tensor attn_mask_15 = expand_dims(axes = attn_mask_15_axes_0, x = attn_mask_13)[name = tensor("attn_mask_15")]; + tensor var_1558_transpose_x_0 = const()[name = tensor("op_1558_transpose_x_0"), val = tensor(false)]; + tensor var_1558_transpose_y_0 = const()[name = tensor("op_1558_transpose_y_0"), val = tensor(false)]; + tensor transpose_24_perm_0 = const()[name = tensor("transpose_24_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_25_perm_0 = const()[name = tensor("transpose_25_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_25 = transpose(perm = transpose_25_perm_0, x = keys_21_cast_fp16)[name = tensor("transpose_39")]; + tensor transpose_24 = transpose(perm = transpose_24_perm_0, x = q_21_cast_fp16)[name = tensor("transpose_40")]; + tensor var_1558_cast_fp16 = matmul(transpose_x = var_1558_transpose_x_0, transpose_y = var_1558_transpose_y_0, x = transpose_24, y = transpose_25)[name = tensor("op_1558_cast_fp16")]; + tensor var_1552_to_fp16 = const()[name = tensor("op_1552_to_fp16"), val = tensor([0x1p-3])]; + tensor attn_weights_19_cast_fp16 = mul(x = var_1558_cast_fp16, y = var_1552_to_fp16)[name = tensor("attn_weights_19_cast_fp16")]; + tensor var_1560 = logical_not(x = attn_mask_15)[name = tensor("op_1560")]; + tensor var_1561_to_fp16 = const()[name = tensor("op_1561_to_fp16"), val = tensor(-inf)]; + tensor attn_weights_21_cast_fp16 = select(a = var_1561_to_fp16, b = attn_weights_19_cast_fp16, cond = var_1560)[name = tensor("attn_weights_21_cast_fp16")]; + tensor var_1563 = const()[name = tensor("op_1563"), val = tensor(-1)]; + tensor attn_weights_23_cast_fp16 = softmax(axis = var_1563, x = attn_weights_21_cast_fp16)[name = tensor("attn_weights_23_cast_fp16")]; + tensor attn_output_7_transpose_x_0 = const()[name = tensor("attn_output_7_transpose_x_0"), val = tensor(false)]; + tensor attn_output_7_transpose_y_0 = const()[name = tensor("attn_output_7_transpose_y_0"), val = tensor(false)]; + tensor values_23_cast_fp16 = transpose(perm = var_1503, x = values_21_cast_fp16)[name = tensor("transpose_41")]; + tensor attn_output_7_cast_fp16 = matmul(transpose_x = attn_output_7_transpose_x_0, transpose_y = attn_output_7_transpose_y_0, x = attn_weights_23_cast_fp16, y = values_23_cast_fp16)[name = tensor("attn_output_7_cast_fp16")]; + tensor var_1571 = const()[name = tensor("op_1571"), val = tensor([0, 2, 1, 3])]; + tensor var_1576 = const()[name = tensor("op_1576"), val = tensor([1, 1, 1024])]; + tensor var_1572_cast_fp16 = transpose(perm = var_1571, x = attn_output_7_cast_fp16)[name = tensor("transpose_38")]; + tensor input_43_cast_fp16 = reshape(shape = var_1576, x = var_1572_cast_fp16)[name = tensor("input_43_cast_fp16")]; + tensor attn3_out_proj_weight_to_fp16 = const()[name = tensor("attn3_out_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(82951808)))]; + tensor linear_14_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = attn3_out_proj_weight_to_fp16, x = input_43_cast_fp16)[name = tensor("linear_14_cast_fp16")]; + tensor var_1582_to_fp16 = const()[name = tensor("op_1582_to_fp16"), val = tensor(0x1p+0)]; + tensor var_1583_cast_fp16 = add(x = position3_to_fp16, y = var_1582_to_fp16)[name = tensor("op_1583_cast_fp16")]; + tensor var_1583_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_1583_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor input_45_cast_fp16 = add(x = input_39_cast_fp16, y = linear_14_cast_fp16)[name = tensor("input_45_cast_fp16")]; + tensor input_47_axes_0 = const()[name = tensor("input_47_axes_0"), val = tensor([-1])]; + tensor norm3_2_weight_to_fp16 = const()[name = tensor("norm3_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(85049024)))]; + tensor norm3_2_bias_to_fp16 = const()[name = tensor("norm3_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(85051136)))]; + tensor var_1587_to_fp16 = const()[name = tensor("op_1587_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_47_cast_fp16 = layer_norm(axes = input_47_axes_0, beta = norm3_2_bias_to_fp16, epsilon = var_1587_to_fp16, gamma = norm3_2_weight_to_fp16, x = input_45_cast_fp16)[name = tensor("input_47_cast_fp16")]; + tensor linear3_1_weight_to_fp16 = const()[name = tensor("linear3_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(85053248)))]; + tensor linear_15_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = linear3_1_weight_to_fp16, x = input_47_cast_fp16)[name = tensor("linear_15_cast_fp16")]; + tensor input_49_mode_0 = const()[name = tensor("input_49_mode_0"), val = tensor("EXACT")]; + tensor input_49_cast_fp16 = gelu(mode = input_49_mode_0, x = linear_15_cast_fp16)[name = tensor("input_49_cast_fp16")]; + tensor linear3_2_weight_to_fp16 = const()[name = tensor("linear3_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(93441920)))]; + tensor linear_16_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = linear3_2_weight_to_fp16, x = input_49_cast_fp16)[name = tensor("linear_16_cast_fp16")]; + tensor input_51_cast_fp16 = add(x = input_45_cast_fp16, y = linear_16_cast_fp16)[name = tensor("input_51_cast_fp16")]; + tensor input_53_axes_0 = const()[name = tensor("input_53_axes_0"), val = tensor([-1])]; + tensor norm4_1_weight_to_fp16 = const()[name = tensor("norm4_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(101830592)))]; + tensor norm4_1_bias_to_fp16 = const()[name = tensor("norm4_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(101832704)))]; + tensor var_1604_to_fp16 = const()[name = tensor("op_1604_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_53_cast_fp16 = layer_norm(axes = input_53_axes_0, beta = norm4_1_bias_to_fp16, epsilon = var_1604_to_fp16, gamma = norm4_1_weight_to_fp16, x = input_51_cast_fp16)[name = tensor("input_53_cast_fp16")]; + tensor attn4_in_proj_weight_to_fp16 = const()[name = tensor("attn4_in_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(101834816)))]; + tensor linear_17_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = attn4_in_proj_weight_to_fp16, x = input_53_cast_fp16)[name = tensor("linear_17_cast_fp16")]; + tensor var_1618 = const()[name = tensor("op_1618"), val = tensor([1, 1, 3, 16, 64])]; + tensor qkv_9_cast_fp16 = reshape(shape = var_1618, x = linear_17_cast_fp16)[name = tensor("qkv_9_cast_fp16")]; + tensor q_25_begin_0 = const()[name = tensor("q_25_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor q_25_end_0 = const()[name = tensor("q_25_end_0"), val = tensor([1, 1, 1, 16, 64])]; + tensor q_25_end_mask_0 = const()[name = tensor("q_25_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor q_25_squeeze_mask_0 = const()[name = tensor("q_25_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor q_25_cast_fp16 = slice_by_index(begin = q_25_begin_0, end = q_25_end_0, end_mask = q_25_end_mask_0, squeeze_mask = q_25_squeeze_mask_0, x = qkv_9_cast_fp16)[name = tensor("q_25_cast_fp16")]; + tensor k_17_begin_0 = const()[name = tensor("k_17_begin_0"), val = tensor([0, 0, 1, 0, 0])]; + tensor k_17_end_0 = const()[name = tensor("k_17_end_0"), val = tensor([1, 1, 2, 16, 64])]; + tensor k_17_end_mask_0 = const()[name = tensor("k_17_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor k_17_squeeze_mask_0 = const()[name = tensor("k_17_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor k_17_cast_fp16 = slice_by_index(begin = k_17_begin_0, end = k_17_end_0, end_mask = k_17_end_mask_0, squeeze_mask = k_17_squeeze_mask_0, x = qkv_9_cast_fp16)[name = tensor("k_17_cast_fp16")]; + tensor v_9_begin_0 = const()[name = tensor("v_9_begin_0"), val = tensor([0, 0, 2, 0, 0])]; + tensor v_9_end_0 = const()[name = tensor("v_9_end_0"), val = tensor([1, 1, 3, 16, 64])]; + tensor v_9_end_mask_0 = const()[name = tensor("v_9_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor v_9_squeeze_mask_0 = const()[name = tensor("v_9_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor v_9_cast_fp16 = slice_by_index(begin = v_9_begin_0, end = v_9_end_0, end_mask = v_9_end_mask_0, squeeze_mask = v_9_squeeze_mask_0, x = qkv_9_cast_fp16)[name = tensor("v_9_cast_fp16")]; + tensor var_1684 = const()[name = tensor("op_1684"), val = tensor([1, 1, 1, 1])]; + tensor position4_to_fp16_dtype_0 = const()[name = tensor("position4_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor position4_to_fp16 = cast(dtype = position4_to_fp16_dtype_0, x = position4)[name = tensor("cast_119")]; + tensor ts_29_cast_fp16 = reshape(shape = var_1684, x = position4_to_fp16)[name = tensor("ts_29_cast_fp16")]; + tensor var_1691 = const()[name = tensor("op_1691"), val = tensor([1, 1, 16, 32, 2])]; + tensor q_complex_9_cast_fp16 = reshape(shape = var_1691, x = q_25_cast_fp16)[name = tensor("q_complex_9_cast_fp16")]; + tensor var_1698 = const()[name = tensor("op_1698"), val = tensor([1, 1, 16, 32, 2])]; + tensor k_complex_9_cast_fp16 = reshape(shape = var_1698, x = k_17_cast_fp16)[name = tensor("k_complex_9_cast_fp16")]; + tensor var_1702_begin_0 = const()[name = tensor("op_1702_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_1702_end_0 = const()[name = tensor("op_1702_end_0"), val = tensor([1, 1, 16, 32, 1])]; + tensor var_1702_end_mask_0 = const()[name = tensor("op_1702_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_1702_squeeze_mask_0 = const()[name = tensor("op_1702_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_1702_cast_fp16 = slice_by_index(begin = var_1702_begin_0, end = var_1702_end_0, end_mask = var_1702_end_mask_0, squeeze_mask = var_1702_squeeze_mask_0, x = q_complex_9_cast_fp16)[name = tensor("op_1702_cast_fp16")]; + tensor var_1710_begin_0 = const()[name = tensor("op_1710_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_1710_end_0 = const()[name = tensor("op_1710_end_0"), val = tensor([1, 1, 16, 32, 2])]; + tensor var_1710_end_mask_0 = const()[name = tensor("op_1710_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_1710_squeeze_mask_0 = const()[name = tensor("op_1710_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_1710_cast_fp16 = slice_by_index(begin = var_1710_begin_0, end = var_1710_end_0, end_mask = var_1710_end_mask_0, squeeze_mask = var_1710_squeeze_mask_0, x = q_complex_9_cast_fp16)[name = tensor("op_1710_cast_fp16")]; + tensor var_1718_begin_0 = const()[name = tensor("op_1718_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_1718_end_0 = const()[name = tensor("op_1718_end_0"), val = tensor([1, 1, 16, 32, 1])]; + tensor var_1718_end_mask_0 = const()[name = tensor("op_1718_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_1718_squeeze_mask_0 = const()[name = tensor("op_1718_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_1718_cast_fp16 = slice_by_index(begin = var_1718_begin_0, end = var_1718_end_0, end_mask = var_1718_end_mask_0, squeeze_mask = var_1718_squeeze_mask_0, x = k_complex_9_cast_fp16)[name = tensor("op_1718_cast_fp16")]; + tensor var_1726_begin_0 = const()[name = tensor("op_1726_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_1726_end_0 = const()[name = tensor("op_1726_end_0"), val = tensor([1, 1, 16, 32, 2])]; + tensor var_1726_end_mask_0 = const()[name = tensor("op_1726_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_1726_squeeze_mask_0 = const()[name = tensor("op_1726_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_1726_cast_fp16 = slice_by_index(begin = var_1726_begin_0, end = var_1726_end_0, end_mask = var_1726_end_mask_0, squeeze_mask = var_1726_squeeze_mask_0, x = k_complex_9_cast_fp16)[name = tensor("op_1726_cast_fp16")]; + tensor freqs_9_to_fp16 = const()[name = tensor("freqs_9_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(108126336)))]; + tensor var_1732_cast_fp16 = mul(x = freqs_9_to_fp16, y = ts_29_cast_fp16)[name = tensor("op_1732_cast_fp16")]; + tensor rotr_9_cast_fp16 = cos(x = var_1732_cast_fp16)[name = tensor("rotr_9_cast_fp16")]; + tensor roti_9_cast_fp16 = sin(x = var_1732_cast_fp16)[name = tensor("roti_9_cast_fp16")]; + tensor var_1736_cast_fp16 = mul(x = var_1702_cast_fp16, y = rotr_9_cast_fp16)[name = tensor("op_1736_cast_fp16")]; + tensor var_1737_cast_fp16 = mul(x = var_1710_cast_fp16, y = roti_9_cast_fp16)[name = tensor("op_1737_cast_fp16")]; + tensor qor_17_cast_fp16 = sub(x = var_1736_cast_fp16, y = var_1737_cast_fp16)[name = tensor("qor_17_cast_fp16")]; + tensor var_1740_cast_fp16 = mul(x = var_1702_cast_fp16, y = roti_9_cast_fp16)[name = tensor("op_1740_cast_fp16")]; + tensor var_1741_cast_fp16 = mul(x = var_1710_cast_fp16, y = rotr_9_cast_fp16)[name = tensor("op_1741_cast_fp16")]; + tensor qoi_17_cast_fp16 = add(x = var_1740_cast_fp16, y = var_1741_cast_fp16)[name = tensor("qoi_17_cast_fp16")]; + tensor var_1744_cast_fp16 = mul(x = var_1718_cast_fp16, y = rotr_9_cast_fp16)[name = tensor("op_1744_cast_fp16")]; + tensor var_1745_cast_fp16 = mul(x = var_1726_cast_fp16, y = roti_9_cast_fp16)[name = tensor("op_1745_cast_fp16")]; + tensor kor_17_cast_fp16 = sub(x = var_1744_cast_fp16, y = var_1745_cast_fp16)[name = tensor("kor_17_cast_fp16")]; + tensor var_1748_cast_fp16 = mul(x = var_1718_cast_fp16, y = roti_9_cast_fp16)[name = tensor("op_1748_cast_fp16")]; + tensor var_1749_cast_fp16 = mul(x = var_1726_cast_fp16, y = rotr_9_cast_fp16)[name = tensor("op_1749_cast_fp16")]; + tensor koi_17_cast_fp16 = add(x = var_1748_cast_fp16, y = var_1749_cast_fp16)[name = tensor("koi_17_cast_fp16")]; + tensor qo_9_axis_0 = const()[name = tensor("qo_9_axis_0"), val = tensor(-1)]; + tensor qo_9_cast_fp16 = stack(axis = qo_9_axis_0, values = (qor_17_cast_fp16, qoi_17_cast_fp16))[name = tensor("qo_9_cast_fp16")]; + tensor ko_9_axis_0 = const()[name = tensor("ko_9_axis_0"), val = tensor(-1)]; + tensor ko_9_cast_fp16 = stack(axis = ko_9_axis_0, values = (kor_17_cast_fp16, koi_17_cast_fp16))[name = tensor("ko_9_cast_fp16")]; + tensor var_1782 = const()[name = tensor("op_1782"), val = tensor([1, 1, 16, 64])]; + tensor q_27_cast_fp16 = reshape(shape = var_1782, x = qo_9_cast_fp16)[name = tensor("q_27_cast_fp16")]; + tensor var_1788 = const()[name = tensor("op_1788"), val = tensor([1, 1, 16, 64])]; + tensor k_19_cast_fp16 = reshape(shape = var_1788, x = ko_9_cast_fp16)[name = tensor("k_19_cast_fp16")]; + tensor _inversed_1813_y_0_to_fp16 = const()[name = tensor("_inversed_1813_y_0_to_fp16"), val = tensor(0x1p-9)]; + tensor _inversed_1813_cast_fp16 = mul(x = ts_29_cast_fp16, y = _inversed_1813_y_0_to_fp16)[name = tensor("_inversed_1813_cast_fp16")]; + tensor var_1814_cast_fp16 = floor(x = _inversed_1813_cast_fp16)[name = tensor("op_1814_cast_fp16")]; + tensor var_1815_to_fp16 = const()[name = tensor("op_1815_to_fp16"), val = tensor(0x1p+9)]; + tensor var_1816_cast_fp16 = mul(x = var_1814_cast_fp16, y = var_1815_to_fp16)[name = tensor("op_1816_cast_fp16")]; + tensor write_indices_float_19_cast_fp16 = sub(x = ts_29_cast_fp16, y = var_1816_cast_fp16)[name = tensor("write_indices_float_19_cast_fp16")]; + tensor var_1823_dtype_0 = const()[name = tensor("op_1823_dtype_0"), val = tensor("int32")]; + tensor write_indices_9_reps_0 = const()[name = tensor("write_indices_9_reps_0"), val = tensor([1, 1, 16, 64])]; + tensor write_indices_float_19_cast_fp16_to_int32 = cast(dtype = var_1823_dtype_0, x = write_indices_float_19_cast_fp16)[name = tensor("cast_118")]; + tensor write_indices_9 = tile(reps = write_indices_9_reps_0, x = write_indices_float_19_cast_fp16_to_int32)[name = tensor("write_indices_9")]; + tensor var_1833_begin_0 = const()[name = tensor("op_1833_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_1833_end_0 = const()[name = tensor("op_1833_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor var_1833_end_mask_0 = const()[name = tensor("op_1833_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_1833_squeeze_mask_0 = const()[name = tensor("op_1833_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor cache4_to_fp16_dtype_0 = const()[name = tensor("cache4_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor cache4_to_fp16 = cast(dtype = cache4_to_fp16_dtype_0, x = cache4)[name = tensor("cast_117")]; + tensor var_1833_cast_fp16 = slice_by_index(begin = var_1833_begin_0, end = var_1833_end_0, end_mask = var_1833_end_mask_0, squeeze_mask = var_1833_squeeze_mask_0, x = cache4_to_fp16)[name = tensor("op_1833_cast_fp16")]; + tensor var_1835_axis_0 = const()[name = tensor("op_1835_axis_0"), val = tensor(1)]; + tensor var_1835_mode_0 = const()[name = tensor("op_1835_mode_0"), val = tensor("update")]; + tensor var_1835_validate_indices_0 = const()[name = tensor("op_1835_validate_indices_0"), val = tensor(false)]; + tensor var_1835_cast_fp16 = scatter_along_axis(axis = var_1835_axis_0, data = var_1833_cast_fp16, indices = write_indices_9, mode = var_1835_mode_0, updates = k_19_cast_fp16, validate_indices = var_1835_validate_indices_0)[name = tensor("op_1835_cast_fp16")]; + tensor concat_30 = const()[name = tensor("concat_30"), val = tensor([0, 0, 0, 0, 0])]; + tensor concat_31 = const()[name = tensor("concat_31"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_9_internal_tensor_assign_1_stride_0 = const()[name = tensor("new_cache_9_internal_tensor_assign_1_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_9_internal_tensor_assign_1_begin_mask_0 = const()[name = tensor("new_cache_9_internal_tensor_assign_1_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_9_internal_tensor_assign_1_end_mask_0 = const()[name = tensor("new_cache_9_internal_tensor_assign_1_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_9_internal_tensor_assign_1_squeeze_mask_0 = const()[name = tensor("new_cache_9_internal_tensor_assign_1_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_20 = const()[name = tensor("shape_20"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_8 = const()[name = tensor("reduce_prod_8"), val = tensor(1048576)]; + tensor range_1d_8_start_0 = const()[name = tensor("range_1d_8_start_0"), val = tensor(0)]; + tensor range_1d_8_step_0 = const()[name = tensor("range_1d_8_step_0"), val = tensor(1)]; + tensor range_1d_8 = range_1d(end = reduce_prod_8, start = range_1d_8_start_0, step = range_1d_8_step_0)[name = tensor("range_1d_8")]; + tensor reshape_40 = reshape(shape = shape_20, x = range_1d_8)[name = tensor("reshape_40")]; + tensor slice_by_index_8 = slice_by_index(begin = concat_30, begin_mask = new_cache_9_internal_tensor_assign_1_begin_mask_0, end = concat_31, end_mask = new_cache_9_internal_tensor_assign_1_end_mask_0, squeeze_mask = new_cache_9_internal_tensor_assign_1_squeeze_mask_0, stride = new_cache_9_internal_tensor_assign_1_stride_0, x = reshape_40)[name = tensor("slice_by_index_8")]; + tensor reshape_41_shape_0 = const()[name = tensor("reshape_41_shape_0"), val = tensor([-1])]; + tensor reshape_41 = reshape(shape = reshape_41_shape_0, x = slice_by_index_8)[name = tensor("reshape_41")]; + tensor reshape_42_shape_0 = const()[name = tensor("reshape_42_shape_0"), val = tensor([-1])]; + tensor reshape_42_cast_fp16 = reshape(shape = reshape_42_shape_0, x = var_1835_cast_fp16)[name = tensor("reshape_42_cast_fp16")]; + tensor reshape_43_shape_0 = const()[name = tensor("reshape_43_shape_0"), val = tensor([-1])]; + tensor reshape_43_cast_fp16 = reshape(shape = reshape_43_shape_0, x = cache4_to_fp16)[name = tensor("reshape_43_cast_fp16")]; + tensor scatter_8_mode_0 = const()[name = tensor("scatter_8_mode_0"), val = tensor("update")]; + tensor scatter_8_axis_0 = const()[name = tensor("scatter_8_axis_0"), val = tensor(0)]; + tensor scatter_8_validate_indices_0 = const()[name = tensor("scatter_8_validate_indices_0"), val = tensor(false)]; + tensor scatter_8_cast_fp16 = scatter(axis = scatter_8_axis_0, data = reshape_43_cast_fp16, indices = reshape_41, mode = scatter_8_mode_0, updates = reshape_42_cast_fp16, validate_indices = scatter_8_validate_indices_0)[name = tensor("scatter_8_cast_fp16")]; + tensor reshape_44_cast_fp16 = reshape(shape = shape_20, x = scatter_8_cast_fp16)[name = tensor("reshape_44_cast_fp16")]; + tensor var_1843_begin_0 = const()[name = tensor("op_1843_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor var_1843_end_0 = const()[name = tensor("op_1843_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor var_1843_end_mask_0 = const()[name = tensor("op_1843_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_1843_squeeze_mask_0 = const()[name = tensor("op_1843_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor var_1843_cast_fp16 = slice_by_index(begin = var_1843_begin_0, end = var_1843_end_0, end_mask = var_1843_end_mask_0, squeeze_mask = var_1843_squeeze_mask_0, x = reshape_44_cast_fp16)[name = tensor("op_1843_cast_fp16")]; + tensor var_1845_axis_0 = const()[name = tensor("op_1845_axis_0"), val = tensor(1)]; + tensor var_1845_mode_0 = const()[name = tensor("op_1845_mode_0"), val = tensor("update")]; + tensor var_1845_validate_indices_0 = const()[name = tensor("op_1845_validate_indices_0"), val = tensor(false)]; + tensor var_1845_cast_fp16 = scatter_along_axis(axis = var_1845_axis_0, data = var_1843_cast_fp16, indices = write_indices_9, mode = var_1845_mode_0, updates = v_9_cast_fp16, validate_indices = var_1845_validate_indices_0)[name = tensor("op_1845_cast_fp16")]; + tensor concat_32 = const()[name = tensor("concat_32"), val = tensor([1, 0, 0, 0, 0])]; + tensor concat_33 = const()[name = tensor("concat_33"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_9_internal_tensor_assign_2_stride_0 = const()[name = tensor("new_cache_9_internal_tensor_assign_2_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_9_internal_tensor_assign_2_begin_mask_0 = const()[name = tensor("new_cache_9_internal_tensor_assign_2_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_9_internal_tensor_assign_2_end_mask_0 = const()[name = tensor("new_cache_9_internal_tensor_assign_2_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_9_internal_tensor_assign_2_squeeze_mask_0 = const()[name = tensor("new_cache_9_internal_tensor_assign_2_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_21 = const()[name = tensor("shape_21"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_9 = const()[name = tensor("reduce_prod_9"), val = tensor(1048576)]; + tensor range_1d_9_start_0 = const()[name = tensor("range_1d_9_start_0"), val = tensor(0)]; + tensor range_1d_9_step_0 = const()[name = tensor("range_1d_9_step_0"), val = tensor(1)]; + tensor range_1d_9 = range_1d(end = reduce_prod_9, start = range_1d_9_start_0, step = range_1d_9_step_0)[name = tensor("range_1d_9")]; + tensor reshape_45 = reshape(shape = shape_21, x = range_1d_9)[name = tensor("reshape_45")]; + tensor slice_by_index_9 = slice_by_index(begin = concat_32, begin_mask = new_cache_9_internal_tensor_assign_2_begin_mask_0, end = concat_33, end_mask = new_cache_9_internal_tensor_assign_2_end_mask_0, squeeze_mask = new_cache_9_internal_tensor_assign_2_squeeze_mask_0, stride = new_cache_9_internal_tensor_assign_2_stride_0, x = reshape_45)[name = tensor("slice_by_index_9")]; + tensor reshape_46_shape_0 = const()[name = tensor("reshape_46_shape_0"), val = tensor([-1])]; + tensor reshape_46 = reshape(shape = reshape_46_shape_0, x = slice_by_index_9)[name = tensor("reshape_46")]; + tensor reshape_47_shape_0 = const()[name = tensor("reshape_47_shape_0"), val = tensor([-1])]; + tensor reshape_47_cast_fp16 = reshape(shape = reshape_47_shape_0, x = var_1845_cast_fp16)[name = tensor("reshape_47_cast_fp16")]; + tensor reshape_48_shape_0 = const()[name = tensor("reshape_48_shape_0"), val = tensor([-1])]; + tensor reshape_48_cast_fp16 = reshape(shape = reshape_48_shape_0, x = reshape_44_cast_fp16)[name = tensor("reshape_48_cast_fp16")]; + tensor scatter_9_mode_0 = const()[name = tensor("scatter_9_mode_0"), val = tensor("update")]; + tensor scatter_9_axis_0 = const()[name = tensor("scatter_9_axis_0"), val = tensor(0)]; + tensor scatter_9_validate_indices_0 = const()[name = tensor("scatter_9_validate_indices_0"), val = tensor(false)]; + tensor scatter_9_cast_fp16 = scatter(axis = scatter_9_axis_0, data = reshape_48_cast_fp16, indices = reshape_46, mode = scatter_9_mode_0, updates = reshape_47_cast_fp16, validate_indices = scatter_9_validate_indices_0)[name = tensor("scatter_9_cast_fp16")]; + tensor reshape_49_cast_fp16 = reshape(shape = shape_21, x = scatter_9_cast_fp16)[name = tensor("reshape_49_cast_fp16")]; + tensor reshape_49_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("reshape_49_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor keys_25_begin_0 = const()[name = tensor("keys_25_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor keys_25_end_0 = const()[name = tensor("keys_25_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor keys_25_end_mask_0 = const()[name = tensor("keys_25_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor keys_25_squeeze_mask_0 = const()[name = tensor("keys_25_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor keys_25_cast_fp16 = slice_by_index(begin = keys_25_begin_0, end = keys_25_end_0, end_mask = keys_25_end_mask_0, squeeze_mask = keys_25_squeeze_mask_0, x = reshape_49_cast_fp16)[name = tensor("keys_25_cast_fp16")]; + tensor values_25_begin_0 = const()[name = tensor("values_25_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor values_25_end_0 = const()[name = tensor("values_25_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor values_25_end_mask_0 = const()[name = tensor("values_25_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor values_25_squeeze_mask_0 = const()[name = tensor("values_25_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor values_25_cast_fp16 = slice_by_index(begin = values_25_begin_0, end = values_25_end_0, end_mask = values_25_end_mask_0, squeeze_mask = values_25_squeeze_mask_0, x = reshape_49_cast_fp16)[name = tensor("values_25_cast_fp16")]; + tensor var_1857_cast_fp16 = not_equal(x = keys_25_cast_fp16, y = keys_25_cast_fp16)[name = tensor("op_1857_cast_fp16")]; + tensor keys_27_cast_fp16 = select(a = var_319_to_fp16, b = keys_25_cast_fp16, cond = var_1857_cast_fp16)[name = tensor("keys_27_cast_fp16")]; + tensor var_1865_cast_fp16 = not_equal(x = values_25_cast_fp16, y = values_25_cast_fp16)[name = tensor("op_1865_cast_fp16")]; + tensor values_27_cast_fp16 = select(a = var_319_to_fp16, b = values_25_cast_fp16, cond = var_1865_cast_fp16)[name = tensor("values_27_cast_fp16")]; + tensor var_1889 = const()[name = tensor("op_1889"), val = tensor([0, 2, 1, 3])]; + tensor var_1905 = const()[name = tensor("op_1905"), val = tensor([1, 1, 1])]; + tensor var_1906_cast_fp16 = reshape(shape = var_1905, x = position4_to_fp16)[name = tensor("op_1906_cast_fp16")]; + tensor var_1926_to_fp16 = const()[name = tensor("op_1926_to_fp16"), val = tensor(0x1p+0)]; + tensor valid_len_9_cast_fp16 = add(x = var_1906_cast_fp16, y = var_1926_to_fp16)[name = tensor("valid_len_9_cast_fp16")]; + tensor valid_mask_9_cast_fp16 = less(x = k_positions_1_promoted_to_fp16, y = valid_len_9_cast_fp16)[name = tensor("valid_mask_9_cast_fp16")]; + tensor causal_mask_9_cast_fp16 = less_equal(x = k_positions_1_promoted_to_fp16, y = var_1906_cast_fp16)[name = tensor("causal_mask_9_cast_fp16")]; + tensor attn_mask_17 = logical_and(x = valid_mask_9_cast_fp16, y = causal_mask_9_cast_fp16)[name = tensor("attn_mask_17")]; + tensor attn_mask_19_axes_0 = const()[name = tensor("attn_mask_19_axes_0"), val = tensor([1])]; + tensor attn_mask_19 = expand_dims(axes = attn_mask_19_axes_0, x = attn_mask_17)[name = tensor("attn_mask_19")]; + tensor var_1944_transpose_x_0 = const()[name = tensor("op_1944_transpose_x_0"), val = tensor(false)]; + tensor var_1944_transpose_y_0 = const()[name = tensor("op_1944_transpose_y_0"), val = tensor(false)]; + tensor transpose_26_perm_0 = const()[name = tensor("transpose_26_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_27_perm_0 = const()[name = tensor("transpose_27_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_27 = transpose(perm = transpose_27_perm_0, x = keys_27_cast_fp16)[name = tensor("transpose_35")]; + tensor transpose_26 = transpose(perm = transpose_26_perm_0, x = q_27_cast_fp16)[name = tensor("transpose_36")]; + tensor var_1944_cast_fp16 = matmul(transpose_x = var_1944_transpose_x_0, transpose_y = var_1944_transpose_y_0, x = transpose_26, y = transpose_27)[name = tensor("op_1944_cast_fp16")]; + tensor var_1938_to_fp16 = const()[name = tensor("op_1938_to_fp16"), val = tensor([0x1p-3])]; + tensor attn_weights_25_cast_fp16 = mul(x = var_1944_cast_fp16, y = var_1938_to_fp16)[name = tensor("attn_weights_25_cast_fp16")]; + tensor var_1946 = logical_not(x = attn_mask_19)[name = tensor("op_1946")]; + tensor var_1947_to_fp16 = const()[name = tensor("op_1947_to_fp16"), val = tensor(-inf)]; + tensor attn_weights_27_cast_fp16 = select(a = var_1947_to_fp16, b = attn_weights_25_cast_fp16, cond = var_1946)[name = tensor("attn_weights_27_cast_fp16")]; + tensor var_1949 = const()[name = tensor("op_1949"), val = tensor(-1)]; + tensor attn_weights_29_cast_fp16 = softmax(axis = var_1949, x = attn_weights_27_cast_fp16)[name = tensor("attn_weights_29_cast_fp16")]; + tensor attn_output_9_transpose_x_0 = const()[name = tensor("attn_output_9_transpose_x_0"), val = tensor(false)]; + tensor attn_output_9_transpose_y_0 = const()[name = tensor("attn_output_9_transpose_y_0"), val = tensor(false)]; + tensor values_29_cast_fp16 = transpose(perm = var_1889, x = values_27_cast_fp16)[name = tensor("transpose_37")]; + tensor attn_output_9_cast_fp16 = matmul(transpose_x = attn_output_9_transpose_x_0, transpose_y = attn_output_9_transpose_y_0, x = attn_weights_29_cast_fp16, y = values_29_cast_fp16)[name = tensor("attn_output_9_cast_fp16")]; + tensor var_1957 = const()[name = tensor("op_1957"), val = tensor([0, 2, 1, 3])]; + tensor var_1962 = const()[name = tensor("op_1962"), val = tensor([1, 1, 1024])]; + tensor var_1958_cast_fp16 = transpose(perm = var_1957, x = attn_output_9_cast_fp16)[name = tensor("transpose_34")]; + tensor input_55_cast_fp16 = reshape(shape = var_1962, x = var_1958_cast_fp16)[name = tensor("input_55_cast_fp16")]; + tensor attn4_out_proj_weight_to_fp16 = const()[name = tensor("attn4_out_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(108126464)))]; + tensor linear_18_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = attn4_out_proj_weight_to_fp16, x = input_55_cast_fp16)[name = tensor("linear_18_cast_fp16")]; + tensor var_1968_to_fp16 = const()[name = tensor("op_1968_to_fp16"), val = tensor(0x1p+0)]; + tensor var_1969_cast_fp16 = add(x = position4_to_fp16, y = var_1968_to_fp16)[name = tensor("op_1969_cast_fp16")]; + tensor var_1969_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_1969_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor input_57_cast_fp16 = add(x = input_51_cast_fp16, y = linear_18_cast_fp16)[name = tensor("input_57_cast_fp16")]; + tensor input_59_axes_0 = const()[name = tensor("input_59_axes_0"), val = tensor([-1])]; + tensor norm4_2_weight_to_fp16 = const()[name = tensor("norm4_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(110223680)))]; + tensor norm4_2_bias_to_fp16 = const()[name = tensor("norm4_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(110225792)))]; + tensor var_1973_to_fp16 = const()[name = tensor("op_1973_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_59_cast_fp16 = layer_norm(axes = input_59_axes_0, beta = norm4_2_bias_to_fp16, epsilon = var_1973_to_fp16, gamma = norm4_2_weight_to_fp16, x = input_57_cast_fp16)[name = tensor("input_59_cast_fp16")]; + tensor linear4_1_weight_to_fp16 = const()[name = tensor("linear4_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(110227904)))]; + tensor linear_19_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = linear4_1_weight_to_fp16, x = input_59_cast_fp16)[name = tensor("linear_19_cast_fp16")]; + tensor input_61_mode_0 = const()[name = tensor("input_61_mode_0"), val = tensor("EXACT")]; + tensor input_61_cast_fp16 = gelu(mode = input_61_mode_0, x = linear_19_cast_fp16)[name = tensor("input_61_cast_fp16")]; + tensor linear4_2_weight_to_fp16 = const()[name = tensor("linear4_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(118616576)))]; + tensor linear_20_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = linear4_2_weight_to_fp16, x = input_61_cast_fp16)[name = tensor("linear_20_cast_fp16")]; + tensor input_63_cast_fp16 = add(x = input_57_cast_fp16, y = linear_20_cast_fp16)[name = tensor("input_63_cast_fp16")]; + tensor input_65_axes_0 = const()[name = tensor("input_65_axes_0"), val = tensor([-1])]; + tensor norm5_1_weight_to_fp16 = const()[name = tensor("norm5_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(127005248)))]; + tensor norm5_1_bias_to_fp16 = const()[name = tensor("norm5_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(127007360)))]; + tensor var_1990_to_fp16 = const()[name = tensor("op_1990_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_65_cast_fp16 = layer_norm(axes = input_65_axes_0, beta = norm5_1_bias_to_fp16, epsilon = var_1990_to_fp16, gamma = norm5_1_weight_to_fp16, x = input_63_cast_fp16)[name = tensor("input_65_cast_fp16")]; + tensor attn5_in_proj_weight_to_fp16 = const()[name = tensor("attn5_in_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(127009472)))]; + tensor linear_21_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = attn5_in_proj_weight_to_fp16, x = input_65_cast_fp16)[name = tensor("linear_21_cast_fp16")]; + tensor var_2004 = const()[name = tensor("op_2004"), val = tensor([1, 1, 3, 16, 64])]; + tensor qkv_cast_fp16 = reshape(shape = var_2004, x = linear_21_cast_fp16)[name = tensor("qkv_cast_fp16")]; + tensor q_31_begin_0 = const()[name = tensor("q_31_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor q_31_end_0 = const()[name = tensor("q_31_end_0"), val = tensor([1, 1, 1, 16, 64])]; + tensor q_31_end_mask_0 = const()[name = tensor("q_31_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor q_31_squeeze_mask_0 = const()[name = tensor("q_31_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor q_31_cast_fp16 = slice_by_index(begin = q_31_begin_0, end = q_31_end_0, end_mask = q_31_end_mask_0, squeeze_mask = q_31_squeeze_mask_0, x = qkv_cast_fp16)[name = tensor("q_31_cast_fp16")]; + tensor k_21_begin_0 = const()[name = tensor("k_21_begin_0"), val = tensor([0, 0, 1, 0, 0])]; + tensor k_21_end_0 = const()[name = tensor("k_21_end_0"), val = tensor([1, 1, 2, 16, 64])]; + tensor k_21_end_mask_0 = const()[name = tensor("k_21_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor k_21_squeeze_mask_0 = const()[name = tensor("k_21_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor k_21_cast_fp16 = slice_by_index(begin = k_21_begin_0, end = k_21_end_0, end_mask = k_21_end_mask_0, squeeze_mask = k_21_squeeze_mask_0, x = qkv_cast_fp16)[name = tensor("k_21_cast_fp16")]; + tensor v_begin_0 = const()[name = tensor("v_begin_0"), val = tensor([0, 0, 2, 0, 0])]; + tensor v_end_0 = const()[name = tensor("v_end_0"), val = tensor([1, 1, 3, 16, 64])]; + tensor v_end_mask_0 = const()[name = tensor("v_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor v_squeeze_mask_0 = const()[name = tensor("v_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor v_cast_fp16 = slice_by_index(begin = v_begin_0, end = v_end_0, end_mask = v_end_mask_0, squeeze_mask = v_squeeze_mask_0, x = qkv_cast_fp16)[name = tensor("v_cast_fp16")]; + tensor var_2070 = const()[name = tensor("op_2070"), val = tensor([1, 1, 1, 1])]; + tensor position5_to_fp16_dtype_0 = const()[name = tensor("position5_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor position5_to_fp16 = cast(dtype = position5_to_fp16_dtype_0, x = position5)[name = tensor("cast_114")]; + tensor ts_cast_fp16 = reshape(shape = var_2070, x = position5_to_fp16)[name = tensor("ts_cast_fp16")]; + tensor var_2077 = const()[name = tensor("op_2077"), val = tensor([1, 1, 16, 32, 2])]; + tensor q_complex_cast_fp16 = reshape(shape = var_2077, x = q_31_cast_fp16)[name = tensor("q_complex_cast_fp16")]; + tensor var_2084 = const()[name = tensor("op_2084"), val = tensor([1, 1, 16, 32, 2])]; + tensor k_complex_cast_fp16 = reshape(shape = var_2084, x = k_21_cast_fp16)[name = tensor("k_complex_cast_fp16")]; + tensor var_2088_begin_0 = const()[name = tensor("op_2088_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_2088_end_0 = const()[name = tensor("op_2088_end_0"), val = tensor([1, 1, 16, 32, 1])]; + tensor var_2088_end_mask_0 = const()[name = tensor("op_2088_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_2088_squeeze_mask_0 = const()[name = tensor("op_2088_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_2088_cast_fp16 = slice_by_index(begin = var_2088_begin_0, end = var_2088_end_0, end_mask = var_2088_end_mask_0, squeeze_mask = var_2088_squeeze_mask_0, x = q_complex_cast_fp16)[name = tensor("op_2088_cast_fp16")]; + tensor var_2096_begin_0 = const()[name = tensor("op_2096_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_2096_end_0 = const()[name = tensor("op_2096_end_0"), val = tensor([1, 1, 16, 32, 2])]; + tensor var_2096_end_mask_0 = const()[name = tensor("op_2096_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_2096_squeeze_mask_0 = const()[name = tensor("op_2096_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_2096_cast_fp16 = slice_by_index(begin = var_2096_begin_0, end = var_2096_end_0, end_mask = var_2096_end_mask_0, squeeze_mask = var_2096_squeeze_mask_0, x = q_complex_cast_fp16)[name = tensor("op_2096_cast_fp16")]; + tensor var_2104_begin_0 = const()[name = tensor("op_2104_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_2104_end_0 = const()[name = tensor("op_2104_end_0"), val = tensor([1, 1, 16, 32, 1])]; + tensor var_2104_end_mask_0 = const()[name = tensor("op_2104_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_2104_squeeze_mask_0 = const()[name = tensor("op_2104_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_2104_cast_fp16 = slice_by_index(begin = var_2104_begin_0, end = var_2104_end_0, end_mask = var_2104_end_mask_0, squeeze_mask = var_2104_squeeze_mask_0, x = k_complex_cast_fp16)[name = tensor("op_2104_cast_fp16")]; + tensor var_2112_begin_0 = const()[name = tensor("op_2112_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_2112_end_0 = const()[name = tensor("op_2112_end_0"), val = tensor([1, 1, 16, 32, 2])]; + tensor var_2112_end_mask_0 = const()[name = tensor("op_2112_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_2112_squeeze_mask_0 = const()[name = tensor("op_2112_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_2112_cast_fp16 = slice_by_index(begin = var_2112_begin_0, end = var_2112_end_0, end_mask = var_2112_end_mask_0, squeeze_mask = var_2112_squeeze_mask_0, x = k_complex_cast_fp16)[name = tensor("op_2112_cast_fp16")]; + tensor freqs_to_fp16 = const()[name = tensor("freqs_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(133300992)))]; + tensor var_2118_cast_fp16 = mul(x = freqs_to_fp16, y = ts_cast_fp16)[name = tensor("op_2118_cast_fp16")]; + tensor rotr_cast_fp16 = cos(x = var_2118_cast_fp16)[name = tensor("rotr_cast_fp16")]; + tensor roti_cast_fp16 = sin(x = var_2118_cast_fp16)[name = tensor("roti_cast_fp16")]; + tensor var_2122_cast_fp16 = mul(x = var_2088_cast_fp16, y = rotr_cast_fp16)[name = tensor("op_2122_cast_fp16")]; + tensor var_2123_cast_fp16 = mul(x = var_2096_cast_fp16, y = roti_cast_fp16)[name = tensor("op_2123_cast_fp16")]; + tensor qor_21_cast_fp16 = sub(x = var_2122_cast_fp16, y = var_2123_cast_fp16)[name = tensor("qor_21_cast_fp16")]; + tensor var_2126_cast_fp16 = mul(x = var_2088_cast_fp16, y = roti_cast_fp16)[name = tensor("op_2126_cast_fp16")]; + tensor var_2127_cast_fp16 = mul(x = var_2096_cast_fp16, y = rotr_cast_fp16)[name = tensor("op_2127_cast_fp16")]; + tensor qoi_21_cast_fp16 = add(x = var_2126_cast_fp16, y = var_2127_cast_fp16)[name = tensor("qoi_21_cast_fp16")]; + tensor var_2130_cast_fp16 = mul(x = var_2104_cast_fp16, y = rotr_cast_fp16)[name = tensor("op_2130_cast_fp16")]; + tensor var_2131_cast_fp16 = mul(x = var_2112_cast_fp16, y = roti_cast_fp16)[name = tensor("op_2131_cast_fp16")]; + tensor kor_21_cast_fp16 = sub(x = var_2130_cast_fp16, y = var_2131_cast_fp16)[name = tensor("kor_21_cast_fp16")]; + tensor var_2134_cast_fp16 = mul(x = var_2104_cast_fp16, y = roti_cast_fp16)[name = tensor("op_2134_cast_fp16")]; + tensor var_2135_cast_fp16 = mul(x = var_2112_cast_fp16, y = rotr_cast_fp16)[name = tensor("op_2135_cast_fp16")]; + tensor koi_21_cast_fp16 = add(x = var_2134_cast_fp16, y = var_2135_cast_fp16)[name = tensor("koi_21_cast_fp16")]; + tensor qo_axis_0 = const()[name = tensor("qo_axis_0"), val = tensor(-1)]; + tensor qo_cast_fp16 = stack(axis = qo_axis_0, values = (qor_21_cast_fp16, qoi_21_cast_fp16))[name = tensor("qo_cast_fp16")]; + tensor ko_axis_0 = const()[name = tensor("ko_axis_0"), val = tensor(-1)]; + tensor ko_cast_fp16 = stack(axis = ko_axis_0, values = (kor_21_cast_fp16, koi_21_cast_fp16))[name = tensor("ko_cast_fp16")]; + tensor var_2168 = const()[name = tensor("op_2168"), val = tensor([1, 1, 16, 64])]; + tensor q_33_cast_fp16 = reshape(shape = var_2168, x = qo_cast_fp16)[name = tensor("q_33_cast_fp16")]; + tensor var_2174 = const()[name = tensor("op_2174"), val = tensor([1, 1, 16, 64])]; + tensor k_cast_fp16 = reshape(shape = var_2174, x = ko_cast_fp16)[name = tensor("k_cast_fp16")]; + tensor _inversed_2199_y_0_to_fp16 = const()[name = tensor("_inversed_2199_y_0_to_fp16"), val = tensor(0x1p-9)]; + tensor _inversed_2199_cast_fp16 = mul(x = ts_cast_fp16, y = _inversed_2199_y_0_to_fp16)[name = tensor("_inversed_2199_cast_fp16")]; + tensor var_2200_cast_fp16 = floor(x = _inversed_2199_cast_fp16)[name = tensor("op_2200_cast_fp16")]; + tensor var_2201_to_fp16 = const()[name = tensor("op_2201_to_fp16"), val = tensor(0x1p+9)]; + tensor var_2202_cast_fp16 = mul(x = var_2200_cast_fp16, y = var_2201_to_fp16)[name = tensor("op_2202_cast_fp16")]; + tensor write_indices_float_cast_fp16 = sub(x = ts_cast_fp16, y = var_2202_cast_fp16)[name = tensor("write_indices_float_cast_fp16")]; + tensor var_2209_dtype_0 = const()[name = tensor("op_2209_dtype_0"), val = tensor("int32")]; + tensor write_indices_reps_0 = const()[name = tensor("write_indices_reps_0"), val = tensor([1, 1, 16, 64])]; + tensor write_indices_float_cast_fp16_to_int32 = cast(dtype = var_2209_dtype_0, x = write_indices_float_cast_fp16)[name = tensor("cast_113")]; + tensor write_indices = tile(reps = write_indices_reps_0, x = write_indices_float_cast_fp16_to_int32)[name = tensor("write_indices")]; + tensor var_2219_begin_0 = const()[name = tensor("op_2219_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_2219_end_0 = const()[name = tensor("op_2219_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor var_2219_end_mask_0 = const()[name = tensor("op_2219_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_2219_squeeze_mask_0 = const()[name = tensor("op_2219_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor cache5_to_fp16_dtype_0 = const()[name = tensor("cache5_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor cache5_to_fp16 = cast(dtype = cache5_to_fp16_dtype_0, x = cache5)[name = tensor("cast_112")]; + tensor var_2219_cast_fp16 = slice_by_index(begin = var_2219_begin_0, end = var_2219_end_0, end_mask = var_2219_end_mask_0, squeeze_mask = var_2219_squeeze_mask_0, x = cache5_to_fp16)[name = tensor("op_2219_cast_fp16")]; + tensor var_2221_axis_0 = const()[name = tensor("op_2221_axis_0"), val = tensor(1)]; + tensor var_2221_mode_0 = const()[name = tensor("op_2221_mode_0"), val = tensor("update")]; + tensor var_2221_validate_indices_0 = const()[name = tensor("op_2221_validate_indices_0"), val = tensor(false)]; + tensor var_2221_cast_fp16 = scatter_along_axis(axis = var_2221_axis_0, data = var_2219_cast_fp16, indices = write_indices, mode = var_2221_mode_0, updates = k_cast_fp16, validate_indices = var_2221_validate_indices_0)[name = tensor("op_2221_cast_fp16")]; + tensor concat_37 = const()[name = tensor("concat_37"), val = tensor([0, 0, 0, 0, 0])]; + tensor concat_38 = const()[name = tensor("concat_38"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_internal_tensor_assign_1_stride_0 = const()[name = tensor("new_cache_internal_tensor_assign_1_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_internal_tensor_assign_1_begin_mask_0 = const()[name = tensor("new_cache_internal_tensor_assign_1_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_internal_tensor_assign_1_end_mask_0 = const()[name = tensor("new_cache_internal_tensor_assign_1_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_internal_tensor_assign_1_squeeze_mask_0 = const()[name = tensor("new_cache_internal_tensor_assign_1_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_22 = const()[name = tensor("shape_22"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_10 = const()[name = tensor("reduce_prod_10"), val = tensor(1048576)]; + tensor range_1d_10_start_0 = const()[name = tensor("range_1d_10_start_0"), val = tensor(0)]; + tensor range_1d_10_step_0 = const()[name = tensor("range_1d_10_step_0"), val = tensor(1)]; + tensor range_1d_10 = range_1d(end = reduce_prod_10, start = range_1d_10_start_0, step = range_1d_10_step_0)[name = tensor("range_1d_10")]; + tensor reshape_50 = reshape(shape = shape_22, x = range_1d_10)[name = tensor("reshape_50")]; + tensor slice_by_index_10 = slice_by_index(begin = concat_37, begin_mask = new_cache_internal_tensor_assign_1_begin_mask_0, end = concat_38, end_mask = new_cache_internal_tensor_assign_1_end_mask_0, squeeze_mask = new_cache_internal_tensor_assign_1_squeeze_mask_0, stride = new_cache_internal_tensor_assign_1_stride_0, x = reshape_50)[name = tensor("slice_by_index_10")]; + tensor reshape_51_shape_0 = const()[name = tensor("reshape_51_shape_0"), val = tensor([-1])]; + tensor reshape_51 = reshape(shape = reshape_51_shape_0, x = slice_by_index_10)[name = tensor("reshape_51")]; + tensor reshape_52_shape_0 = const()[name = tensor("reshape_52_shape_0"), val = tensor([-1])]; + tensor reshape_52_cast_fp16 = reshape(shape = reshape_52_shape_0, x = var_2221_cast_fp16)[name = tensor("reshape_52_cast_fp16")]; + tensor reshape_53_shape_0 = const()[name = tensor("reshape_53_shape_0"), val = tensor([-1])]; + tensor reshape_53_cast_fp16 = reshape(shape = reshape_53_shape_0, x = cache5_to_fp16)[name = tensor("reshape_53_cast_fp16")]; + tensor scatter_10_mode_0 = const()[name = tensor("scatter_10_mode_0"), val = tensor("update")]; + tensor scatter_10_axis_0 = const()[name = tensor("scatter_10_axis_0"), val = tensor(0)]; + tensor scatter_10_validate_indices_0 = const()[name = tensor("scatter_10_validate_indices_0"), val = tensor(false)]; + tensor scatter_10_cast_fp16 = scatter(axis = scatter_10_axis_0, data = reshape_53_cast_fp16, indices = reshape_51, mode = scatter_10_mode_0, updates = reshape_52_cast_fp16, validate_indices = scatter_10_validate_indices_0)[name = tensor("scatter_10_cast_fp16")]; + tensor reshape_54_cast_fp16 = reshape(shape = shape_22, x = scatter_10_cast_fp16)[name = tensor("reshape_54_cast_fp16")]; + tensor var_2229_begin_0 = const()[name = tensor("op_2229_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor var_2229_end_0 = const()[name = tensor("op_2229_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor var_2229_end_mask_0 = const()[name = tensor("op_2229_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_2229_squeeze_mask_0 = const()[name = tensor("op_2229_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor var_2229_cast_fp16 = slice_by_index(begin = var_2229_begin_0, end = var_2229_end_0, end_mask = var_2229_end_mask_0, squeeze_mask = var_2229_squeeze_mask_0, x = reshape_54_cast_fp16)[name = tensor("op_2229_cast_fp16")]; + tensor var_2231_axis_0 = const()[name = tensor("op_2231_axis_0"), val = tensor(1)]; + tensor var_2231_mode_0 = const()[name = tensor("op_2231_mode_0"), val = tensor("update")]; + tensor var_2231_validate_indices_0 = const()[name = tensor("op_2231_validate_indices_0"), val = tensor(false)]; + tensor var_2231_cast_fp16 = scatter_along_axis(axis = var_2231_axis_0, data = var_2229_cast_fp16, indices = write_indices, mode = var_2231_mode_0, updates = v_cast_fp16, validate_indices = var_2231_validate_indices_0)[name = tensor("op_2231_cast_fp16")]; + tensor concat_39 = const()[name = tensor("concat_39"), val = tensor([1, 0, 0, 0, 0])]; + tensor concat_40 = const()[name = tensor("concat_40"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_internal_tensor_assign_2_stride_0 = const()[name = tensor("new_cache_internal_tensor_assign_2_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_internal_tensor_assign_2_begin_mask_0 = const()[name = tensor("new_cache_internal_tensor_assign_2_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_internal_tensor_assign_2_end_mask_0 = const()[name = tensor("new_cache_internal_tensor_assign_2_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_internal_tensor_assign_2_squeeze_mask_0 = const()[name = tensor("new_cache_internal_tensor_assign_2_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_23 = const()[name = tensor("shape_23"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_11 = const()[name = tensor("reduce_prod_11"), val = tensor(1048576)]; + tensor range_1d_11_start_0 = const()[name = tensor("range_1d_11_start_0"), val = tensor(0)]; + tensor range_1d_11_step_0 = const()[name = tensor("range_1d_11_step_0"), val = tensor(1)]; + tensor range_1d_11 = range_1d(end = reduce_prod_11, start = range_1d_11_start_0, step = range_1d_11_step_0)[name = tensor("range_1d_11")]; + tensor reshape_55 = reshape(shape = shape_23, x = range_1d_11)[name = tensor("reshape_55")]; + tensor slice_by_index_11 = slice_by_index(begin = concat_39, begin_mask = new_cache_internal_tensor_assign_2_begin_mask_0, end = concat_40, end_mask = new_cache_internal_tensor_assign_2_end_mask_0, squeeze_mask = new_cache_internal_tensor_assign_2_squeeze_mask_0, stride = new_cache_internal_tensor_assign_2_stride_0, x = reshape_55)[name = tensor("slice_by_index_11")]; + tensor reshape_56_shape_0 = const()[name = tensor("reshape_56_shape_0"), val = tensor([-1])]; + tensor reshape_56 = reshape(shape = reshape_56_shape_0, x = slice_by_index_11)[name = tensor("reshape_56")]; + tensor reshape_57_shape_0 = const()[name = tensor("reshape_57_shape_0"), val = tensor([-1])]; + tensor reshape_57_cast_fp16 = reshape(shape = reshape_57_shape_0, x = var_2231_cast_fp16)[name = tensor("reshape_57_cast_fp16")]; + tensor reshape_58_shape_0 = const()[name = tensor("reshape_58_shape_0"), val = tensor([-1])]; + tensor reshape_58_cast_fp16 = reshape(shape = reshape_58_shape_0, x = reshape_54_cast_fp16)[name = tensor("reshape_58_cast_fp16")]; + tensor scatter_11_mode_0 = const()[name = tensor("scatter_11_mode_0"), val = tensor("update")]; + tensor scatter_11_axis_0 = const()[name = tensor("scatter_11_axis_0"), val = tensor(0)]; + tensor scatter_11_validate_indices_0 = const()[name = tensor("scatter_11_validate_indices_0"), val = tensor(false)]; + tensor scatter_11_cast_fp16 = scatter(axis = scatter_11_axis_0, data = reshape_58_cast_fp16, indices = reshape_56, mode = scatter_11_mode_0, updates = reshape_57_cast_fp16, validate_indices = scatter_11_validate_indices_0)[name = tensor("scatter_11_cast_fp16")]; + tensor reshape_59_cast_fp16 = reshape(shape = shape_23, x = scatter_11_cast_fp16)[name = tensor("reshape_59_cast_fp16")]; + tensor reshape_59_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("reshape_59_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor keys_31_begin_0 = const()[name = tensor("keys_31_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor keys_31_end_0 = const()[name = tensor("keys_31_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor keys_31_end_mask_0 = const()[name = tensor("keys_31_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor keys_31_squeeze_mask_0 = const()[name = tensor("keys_31_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor keys_31_cast_fp16 = slice_by_index(begin = keys_31_begin_0, end = keys_31_end_0, end_mask = keys_31_end_mask_0, squeeze_mask = keys_31_squeeze_mask_0, x = reshape_59_cast_fp16)[name = tensor("keys_31_cast_fp16")]; + tensor values_31_begin_0 = const()[name = tensor("values_31_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor values_31_end_0 = const()[name = tensor("values_31_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor values_31_end_mask_0 = const()[name = tensor("values_31_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor values_31_squeeze_mask_0 = const()[name = tensor("values_31_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor values_31_cast_fp16 = slice_by_index(begin = values_31_begin_0, end = values_31_end_0, end_mask = values_31_end_mask_0, squeeze_mask = values_31_squeeze_mask_0, x = reshape_59_cast_fp16)[name = tensor("values_31_cast_fp16")]; + tensor var_2243_cast_fp16 = not_equal(x = keys_31_cast_fp16, y = keys_31_cast_fp16)[name = tensor("op_2243_cast_fp16")]; + tensor keys_33_cast_fp16 = select(a = var_319_to_fp16, b = keys_31_cast_fp16, cond = var_2243_cast_fp16)[name = tensor("keys_33_cast_fp16")]; + tensor var_2251_cast_fp16 = not_equal(x = values_31_cast_fp16, y = values_31_cast_fp16)[name = tensor("op_2251_cast_fp16")]; + tensor values_33_cast_fp16 = select(a = var_319_to_fp16, b = values_31_cast_fp16, cond = var_2251_cast_fp16)[name = tensor("values_33_cast_fp16")]; + tensor var_2275 = const()[name = tensor("op_2275"), val = tensor([0, 2, 1, 3])]; + tensor var_2291 = const()[name = tensor("op_2291"), val = tensor([1, 1, 1])]; + tensor var_2292_cast_fp16 = reshape(shape = var_2291, x = position5_to_fp16)[name = tensor("op_2292_cast_fp16")]; + tensor var_2312_to_fp16 = const()[name = tensor("op_2312_to_fp16"), val = tensor(0x1p+0)]; + tensor valid_len_cast_fp16 = add(x = var_2292_cast_fp16, y = var_2312_to_fp16)[name = tensor("valid_len_cast_fp16")]; + tensor valid_mask_cast_fp16 = less(x = k_positions_1_promoted_to_fp16, y = valid_len_cast_fp16)[name = tensor("valid_mask_cast_fp16")]; + tensor causal_mask_cast_fp16 = less_equal(x = k_positions_1_promoted_to_fp16, y = var_2292_cast_fp16)[name = tensor("causal_mask_cast_fp16")]; + tensor attn_mask_21 = logical_and(x = valid_mask_cast_fp16, y = causal_mask_cast_fp16)[name = tensor("attn_mask_21")]; + tensor attn_mask_axes_0 = const()[name = tensor("attn_mask_axes_0"), val = tensor([1])]; + tensor attn_mask = expand_dims(axes = attn_mask_axes_0, x = attn_mask_21)[name = tensor("attn_mask")]; + tensor var_2330_transpose_x_0 = const()[name = tensor("op_2330_transpose_x_0"), val = tensor(false)]; + tensor var_2330_transpose_y_0 = const()[name = tensor("op_2330_transpose_y_0"), val = tensor(false)]; + tensor transpose_28_perm_0 = const()[name = tensor("transpose_28_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_29_perm_0 = const()[name = tensor("transpose_29_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_29 = transpose(perm = transpose_29_perm_0, x = keys_33_cast_fp16)[name = tensor("transpose_31")]; + tensor transpose_28 = transpose(perm = transpose_28_perm_0, x = q_33_cast_fp16)[name = tensor("transpose_32")]; + tensor var_2330_cast_fp16 = matmul(transpose_x = var_2330_transpose_x_0, transpose_y = var_2330_transpose_y_0, x = transpose_28, y = transpose_29)[name = tensor("op_2330_cast_fp16")]; + tensor var_2324_to_fp16 = const()[name = tensor("op_2324_to_fp16"), val = tensor([0x1p-3])]; + tensor attn_weights_31_cast_fp16 = mul(x = var_2330_cast_fp16, y = var_2324_to_fp16)[name = tensor("attn_weights_31_cast_fp16")]; + tensor var_2332 = logical_not(x = attn_mask)[name = tensor("op_2332")]; + tensor var_2333_to_fp16 = const()[name = tensor("op_2333_to_fp16"), val = tensor(-inf)]; + tensor attn_weights_33_cast_fp16 = select(a = var_2333_to_fp16, b = attn_weights_31_cast_fp16, cond = var_2332)[name = tensor("attn_weights_33_cast_fp16")]; + tensor var_2335 = const()[name = tensor("op_2335"), val = tensor(-1)]; + tensor attn_weights_cast_fp16 = softmax(axis = var_2335, x = attn_weights_33_cast_fp16)[name = tensor("attn_weights_cast_fp16")]; + tensor attn_output_transpose_x_0 = const()[name = tensor("attn_output_transpose_x_0"), val = tensor(false)]; + tensor attn_output_transpose_y_0 = const()[name = tensor("attn_output_transpose_y_0"), val = tensor(false)]; + tensor values_cast_fp16 = transpose(perm = var_2275, x = values_33_cast_fp16)[name = tensor("transpose_33")]; + tensor attn_output_cast_fp16 = matmul(transpose_x = attn_output_transpose_x_0, transpose_y = attn_output_transpose_y_0, x = attn_weights_cast_fp16, y = values_cast_fp16)[name = tensor("attn_output_cast_fp16")]; + tensor var_2343 = const()[name = tensor("op_2343"), val = tensor([0, 2, 1, 3])]; + tensor var_2348 = const()[name = tensor("op_2348"), val = tensor([1, 1, 1024])]; + tensor var_2344_cast_fp16 = transpose(perm = var_2343, x = attn_output_cast_fp16)[name = tensor("transpose_30")]; + tensor input_67_cast_fp16 = reshape(shape = var_2348, x = var_2344_cast_fp16)[name = tensor("input_67_cast_fp16")]; + tensor attn5_out_proj_weight_to_fp16 = const()[name = tensor("attn5_out_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(133301120)))]; + tensor linear_22_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = attn5_out_proj_weight_to_fp16, x = input_67_cast_fp16)[name = tensor("linear_22_cast_fp16")]; + tensor var_2354_to_fp16 = const()[name = tensor("op_2354_to_fp16"), val = tensor(0x1p+0)]; + tensor var_2355_cast_fp16 = add(x = position5_to_fp16, y = var_2354_to_fp16)[name = tensor("op_2355_cast_fp16")]; + tensor var_2355_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_2355_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor input_69_cast_fp16 = add(x = input_63_cast_fp16, y = linear_22_cast_fp16)[name = tensor("input_69_cast_fp16")]; + tensor input_71_axes_0 = const()[name = tensor("input_71_axes_0"), val = tensor([-1])]; + tensor norm5_2_weight_to_fp16 = const()[name = tensor("norm5_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(135398336)))]; + tensor norm5_2_bias_to_fp16 = const()[name = tensor("norm5_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(135400448)))]; + tensor var_2359_to_fp16 = const()[name = tensor("op_2359_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_71_cast_fp16 = layer_norm(axes = input_71_axes_0, beta = norm5_2_bias_to_fp16, epsilon = var_2359_to_fp16, gamma = norm5_2_weight_to_fp16, x = input_69_cast_fp16)[name = tensor("input_71_cast_fp16")]; + tensor linear5_1_weight_to_fp16 = const()[name = tensor("linear5_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(135402560)))]; + tensor linear_23_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = linear5_1_weight_to_fp16, x = input_71_cast_fp16)[name = tensor("linear_23_cast_fp16")]; + tensor input_73_mode_0 = const()[name = tensor("input_73_mode_0"), val = tensor("EXACT")]; + tensor input_73_cast_fp16 = gelu(mode = input_73_mode_0, x = linear_23_cast_fp16)[name = tensor("input_73_cast_fp16")]; + tensor linear5_2_weight_to_fp16 = const()[name = tensor("linear5_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(143791232)))]; + tensor linear_24_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = linear5_2_weight_to_fp16, x = input_73_cast_fp16)[name = tensor("linear_24_cast_fp16")]; + tensor input_75_cast_fp16 = add(x = input_69_cast_fp16, y = linear_24_cast_fp16)[name = tensor("input_75_cast_fp16")]; + tensor input_axes_0 = const()[name = tensor("input_axes_0"), val = tensor([-1])]; + tensor out_norm_weight_to_fp16 = const()[name = tensor("out_norm_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(152179904)))]; + tensor out_norm_bias_to_fp16 = const()[name = tensor("out_norm_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(152182016)))]; + tensor var_2376_to_fp16 = const()[name = tensor("op_2376_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_cast_fp16 = layer_norm(axes = input_axes_0, beta = out_norm_bias_to_fp16, epsilon = var_2376_to_fp16, gamma = out_norm_weight_to_fp16, x = input_75_cast_fp16)[name = tensor("input_cast_fp16")]; + tensor input_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("input_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor out_eos_weight_to_fp16 = const()[name = tensor("out_eos_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(152184128)))]; + tensor out_eos_bias_to_fp16 = const()[name = tensor("out_eos_bias_to_fp16"), val = tensor([-0x1.5ap-3])]; + tensor linear_25_cast_fp16 = linear(bias = out_eos_bias_to_fp16, weight = out_eos_weight_to_fp16, x = input_cast_fp16)[name = tensor("linear_25_cast_fp16")]; + tensor linear_25_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("linear_25_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor var_2384 = cast(dtype = linear_25_cast_fp16_to_fp32_dtype_0, x = linear_25_cast_fp16)[name = tensor("cast_108")]; + tensor input = cast(dtype = input_cast_fp16_to_fp32_dtype_0, x = input_cast_fp16)[name = tensor("cast_109")]; + tensor var_2355 = cast(dtype = var_2355_cast_fp16_to_fp32_dtype_0, x = var_2355_cast_fp16)[name = tensor("cast_110")]; + tensor new_cache_internal_tensor_assign_2 = cast(dtype = reshape_59_cast_fp16_to_fp32_dtype_0, x = reshape_59_cast_fp16)[name = tensor("cast_111")]; + tensor var_1969 = cast(dtype = var_1969_cast_fp16_to_fp32_dtype_0, x = var_1969_cast_fp16)[name = tensor("cast_115")]; + tensor new_cache_9_internal_tensor_assign_2 = cast(dtype = reshape_49_cast_fp16_to_fp32_dtype_0, x = reshape_49_cast_fp16)[name = tensor("cast_116")]; + tensor var_1583 = cast(dtype = var_1583_cast_fp16_to_fp32_dtype_0, x = var_1583_cast_fp16)[name = tensor("cast_120")]; + tensor new_cache_7_internal_tensor_assign_2 = cast(dtype = reshape_39_cast_fp16_to_fp32_dtype_0, x = reshape_39_cast_fp16)[name = tensor("cast_121")]; + tensor var_1197 = cast(dtype = var_1197_cast_fp16_to_fp32_dtype_0, x = var_1197_cast_fp16)[name = tensor("cast_125")]; + tensor new_cache_5_internal_tensor_assign_2 = cast(dtype = reshape_29_cast_fp16_to_fp32_dtype_0, x = reshape_29_cast_fp16)[name = tensor("cast_126")]; + tensor var_811 = cast(dtype = var_811_cast_fp16_to_fp32_dtype_0, x = var_811_cast_fp16)[name = tensor("cast_130")]; + tensor new_cache_3_internal_tensor_assign_2 = cast(dtype = reshape_19_cast_fp16_to_fp32_dtype_0, x = reshape_19_cast_fp16)[name = tensor("cast_131")]; + tensor var_425 = cast(dtype = var_425_cast_fp16_to_fp32_dtype_0, x = var_425_cast_fp16)[name = tensor("cast_135")]; + tensor new_cache_1_internal_tensor_assign_2 = cast(dtype = reshape_9_cast_fp16_to_fp32_dtype_0, x = reshape_9_cast_fp16)[name = tensor("cast_136")]; + } -> (input, var_2384, new_cache_1_internal_tensor_assign_2, var_425, new_cache_3_internal_tensor_assign_2, var_811, new_cache_5_internal_tensor_assign_2, var_1197, new_cache_7_internal_tensor_assign_2, var_1583, new_cache_9_internal_tensor_assign_2, var_1969, new_cache_internal_tensor_assign_2, var_2355); +} \ No newline at end of file diff --git a/v2.1/portuguese/flowlm_step.mlmodelc/weights/weight.bin b/v2.1/portuguese/flowlm_step.mlmodelc/weights/weight.bin new file mode 100644 index 0000000000000000000000000000000000000000..66c2d4adb0850ce0adad51b393157719925cef29 --- /dev/null +++ b/v2.1/portuguese/flowlm_step.mlmodelc/weights/weight.bin @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:69b00e96bbf580f044e08904e5396196f4ea74bc2ac8192dc732b73b5e4cea12 +size 152186240 diff --git a/v2.1/portuguese/flowlm_step.mlpackage/Data/com.apple.CoreML/model.mlmodel b/v2.1/portuguese/flowlm_step.mlpackage/Data/com.apple.CoreML/model.mlmodel new file mode 100644 index 0000000000000000000000000000000000000000..57b5c6443937cf460a1ab6141c7a5f08b3a22816 --- /dev/null +++ b/v2.1/portuguese/flowlm_step.mlpackage/Data/com.apple.CoreML/model.mlmodel @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:0c480f46b1a3570dea8af2856a1890b759b861a10247dd94e3f7b6070f56382b +size 207618 diff --git a/v2.1/portuguese/flowlm_step.mlpackage/Data/com.apple.CoreML/weights/weight.bin b/v2.1/portuguese/flowlm_step.mlpackage/Data/com.apple.CoreML/weights/weight.bin new file mode 100644 index 0000000000000000000000000000000000000000..66c2d4adb0850ce0adad51b393157719925cef29 --- /dev/null +++ b/v2.1/portuguese/flowlm_step.mlpackage/Data/com.apple.CoreML/weights/weight.bin @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:69b00e96bbf580f044e08904e5396196f4ea74bc2ac8192dc732b73b5e4cea12 +size 152186240 diff --git a/v2.1/portuguese/flowlm_step.mlpackage/Manifest.json b/v2.1/portuguese/flowlm_step.mlpackage/Manifest.json new file mode 100644 index 0000000000000000000000000000000000000000..97512fcd861a2ad71db5b2c0618d285777a6818f --- /dev/null +++ b/v2.1/portuguese/flowlm_step.mlpackage/Manifest.json @@ -0,0 +1,18 @@ +{ + "fileFormatVersion": "1.0.0", + "itemInfoEntries": { + "C5B8FF0B-F635-4CD7-8F52-3C8F6B686100": { + "author": "com.apple.CoreML", + "description": "CoreML Model Weights", + "name": "weights", + "path": "com.apple.CoreML/weights" + }, + "DFEB7ED6-45E4-45AF-A65F-E212B93B72FE": { + "author": "com.apple.CoreML", + "description": "CoreML Model Specification", + "name": "model.mlmodel", + "path": "com.apple.CoreML/model.mlmodel" + } + }, + "rootModelIdentifier": "DFEB7ED6-45E4-45AF-A65F-E212B93B72FE" +} diff --git a/v2.1/portuguese/flowlm_stepv2.mlmodelc/analytics/coremldata.bin b/v2.1/portuguese/flowlm_stepv2.mlmodelc/analytics/coremldata.bin new file mode 100644 index 0000000000000000000000000000000000000000..92962054e2bd071bb566c373a961282da6011dfa --- /dev/null +++ b/v2.1/portuguese/flowlm_stepv2.mlmodelc/analytics/coremldata.bin @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:125ced412a243bab041df2348f8cec91ea96e27492f4144f1a71abaeca66f387 +size 243 diff --git a/v2.1/portuguese/flowlm_stepv2.mlmodelc/coremldata.bin b/v2.1/portuguese/flowlm_stepv2.mlmodelc/coremldata.bin new file mode 100644 index 0000000000000000000000000000000000000000..f85f00238ad4d275b2d42d263256131ca81a58e1 --- /dev/null +++ b/v2.1/portuguese/flowlm_stepv2.mlmodelc/coremldata.bin @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:a28ab65e7aab97a93da35e319d1a05633a939ae92c84b77a5a6727a2f84bab43 +size 1190 diff --git a/v2.1/portuguese/flowlm_stepv2.mlmodelc/metadata.json b/v2.1/portuguese/flowlm_stepv2.mlmodelc/metadata.json new file mode 100644 index 0000000000000000000000000000000000000000..f2969c83e1cf7a6025c56eb3502b7ec456a89630 --- /dev/null +++ b/v2.1/portuguese/flowlm_stepv2.mlmodelc/metadata.json @@ -0,0 +1,348 @@ +[ + { + "metadataOutputVersion" : "3.0", + "storagePrecision" : "Mixed (Float16, Int8)", + "outputSchema" : [ + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1 × 1 × 1024)", + "shortDescription" : "", + "shape" : "[1, 1, 1024]", + "name" : "input", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1 × 1 × 1)", + "shortDescription" : "", + "shape" : "[1, 1, 1]", + "name" : "var_2582", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 2 × 1 × 512 × 16 × 64)", + "shortDescription" : "", + "shape" : "[2, 1, 512, 16, 64]", + "name" : "new_cache_1_internal_tensor_assign_2", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1)", + "shortDescription" : "", + "shape" : "[1]", + "name" : "var_458", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 2 × 1 × 512 × 16 × 64)", + "shortDescription" : "", + "shape" : "[2, 1, 512, 16, 64]", + "name" : "new_cache_3_internal_tensor_assign_2", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1)", + "shortDescription" : "", + "shape" : "[1]", + "name" : "var_877", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 2 × 1 × 512 × 16 × 64)", + "shortDescription" : "", + "shape" : "[2, 1, 512, 16, 64]", + "name" : "new_cache_5_internal_tensor_assign_2", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1)", + "shortDescription" : "", + "shape" : "[1]", + "name" : "var_1296", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 2 × 1 × 512 × 16 × 64)", + "shortDescription" : "", + "shape" : "[2, 1, 512, 16, 64]", + "name" : "new_cache_7_internal_tensor_assign_2", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1)", + "shortDescription" : "", + "shape" : "[1]", + "name" : "var_1715", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 2 × 1 × 512 × 16 × 64)", + "shortDescription" : "", + "shape" : "[2, 1, 512, 16, 64]", + "name" : "new_cache_9_internal_tensor_assign_2", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1)", + "shortDescription" : "", + "shape" : "[1]", + "name" : "var_2134", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 2 × 1 × 512 × 16 × 64)", + "shortDescription" : "", + "shape" : "[2, 1, 512, 16, 64]", + "name" : "new_cache_internal_tensor_assign_2", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1)", + "shortDescription" : "", + "shape" : "[1]", + "name" : "var_2553", + "type" : "MultiArray" + } + ], + "modelParameters" : [ + + ], + "specificationVersion" : 8, + "mlProgramOperationTypeHistogram" : { + "Ios17.cos" : 6, + "Range1d" : 12, + "Ios17.notEqual" : 13, + "Ios17.floor" : 6, + "Ios17.reshape" : 108, + "Ios17.logicalAnd" : 6, + "Ios16.softmax" : 6, + "Ios17.scatter" : 12, + "Ios17.matmul" : 12, + "Ios17.sin" : 6, + "Ios17.transpose" : 24, + "Select" : 19, + "Ios17.expandDims" : 7, + "Ios17.add" : 36, + "Tile" : 6, + "Ios17.sliceByIndex" : 78, + "Ios17.lessEqual" : 6, + "Ios17.scatterAlongAxis" : 12, + "Ios17.logicalNot" : 6, + "Ios17.layerNorm" : 13, + "Ios16.constexprAffineDequantize" : 24, + "Ios17.cast" : 34, + "Ios17.sub" : 18, + "Ios17.less" : 6, + "Ios16.gelu" : 6, + "Ios17.linear" : 26, + "Stack" : 12, + "Ios17.mul" : 72 + }, + "computePrecision" : "Mixed (Float16, Float32, Int32)", + "isUpdatable" : "0", + "stateSchema" : [ + + ], + "availability" : { + "macOS" : "14.0", + "tvOS" : "17.0", + "visionOS" : "1.0", + "watchOS" : "10.0", + "iOS" : "17.0", + "macCatalyst" : "17.0" + }, + "modelType" : { + "name" : "MLModelType_mlProgram" + }, + "userDefinedMetadata" : { + "com.github.apple.coremltools.conversion_date" : "2026-04-27", + "com.github.apple.coremltools.source" : "torch==2.9.1", + "com.github.apple.coremltools.version" : "9.0", + "com.github.apple.coremltools.source_dialect" : "TorchScript" + }, + "inputSchema" : [ + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1 × 1 × 32)", + "shortDescription" : "", + "shape" : "[1, 1, 32]", + "name" : "sequence", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 32)", + "shortDescription" : "", + "shape" : "[32]", + "name" : "bos_emb", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 2 × 1 × 512 × 16 × 64)", + "shortDescription" : "", + "shape" : "[2, 1, 512, 16, 64]", + "name" : "cache0", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1)", + "shortDescription" : "", + "shape" : "[1]", + "name" : "position0", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 2 × 1 × 512 × 16 × 64)", + "shortDescription" : "", + "shape" : "[2, 1, 512, 16, 64]", + "name" : "cache1", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1)", + "shortDescription" : "", + "shape" : "[1]", + "name" : "position1", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 2 × 1 × 512 × 16 × 64)", + "shortDescription" : "", + "shape" : "[2, 1, 512, 16, 64]", + "name" : "cache2", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1)", + "shortDescription" : "", + "shape" : "[1]", + "name" : "position2", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 2 × 1 × 512 × 16 × 64)", + "shortDescription" : "", + "shape" : "[2, 1, 512, 16, 64]", + "name" : "cache3", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1)", + "shortDescription" : "", + "shape" : "[1]", + "name" : "position3", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 2 × 1 × 512 × 16 × 64)", + "shortDescription" : "", + "shape" : "[2, 1, 512, 16, 64]", + "name" : "cache4", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1)", + "shortDescription" : "", + "shape" : "[1]", + "name" : "position4", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 2 × 1 × 512 × 16 × 64)", + "shortDescription" : "", + "shape" : "[2, 1, 512, 16, 64]", + "name" : "cache5", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1)", + "shortDescription" : "", + "shape" : "[1]", + "name" : "position5", + "type" : "MultiArray" + } + ], + "generatedClassName" : "flowlm_stepv2", + "method" : "predict" + } +] \ No newline at end of file diff --git a/v2.1/portuguese/flowlm_stepv2.mlmodelc/model.mil b/v2.1/portuguese/flowlm_stepv2.mlmodelc/model.mil new file mode 100644 index 0000000000000000000000000000000000000000..d525b23691342ce77ede3cb55738ea4c310ccac7 --- /dev/null +++ b/v2.1/portuguese/flowlm_stepv2.mlmodelc/model.mil @@ -0,0 +1,1359 @@ +program(1.0) +[buildInfo = dict, tensor>({{"coremlc-component-MIL", "3520.4.1"}, {"coremlc-version", "3520.5.1"}, {"coremltools-component-torch", "2.9.1"}, {"coremltools-source-dialect", "TorchScript"}, {"coremltools-version", "9.0"}})] +{ + func main(tensor bos_emb, tensor cache0, tensor cache1, tensor cache2, tensor cache3, tensor cache4, tensor cache5, tensor position0, tensor position1, tensor position2, tensor position3, tensor position4, tensor position5, tensor sequence) { + tensor attn0_in_proj_weight_cast_fp16 = constexpr_affine_dequantize()[axis = tensor(0), name = tensor("attn0_in_proj_weight_cast_fp16"), quantized_data = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(64))), scale = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(3148992))), zero_point = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(3145856)))]; + tensor attn0_out_proj_weight_cast_fp16 = constexpr_affine_dequantize()[axis = tensor(0), name = tensor("attn0_out_proj_weight_cast_fp16"), quantized_data = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(3155200))), scale = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(4204928))), zero_point = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(4203840)))]; + tensor linear0_1_weight_cast_fp16 = constexpr_affine_dequantize()[axis = tensor(0), name = tensor("linear0_1_weight_cast_fp16"), quantized_data = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(4207040))), scale = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8405568))), zero_point = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8401408)))]; + tensor linear0_2_weight_cast_fp16 = constexpr_affine_dequantize()[axis = tensor(0), name = tensor("linear0_2_weight_cast_fp16"), quantized_data = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8413824))), scale = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(12608192))), zero_point = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(4203840)))]; + tensor attn1_in_proj_weight_cast_fp16 = constexpr_affine_dequantize()[axis = tensor(0), name = tensor("attn1_in_proj_weight_cast_fp16"), quantized_data = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(12610304))), scale = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15756096))), zero_point = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(3145856)))]; + tensor attn1_out_proj_weight_cast_fp16 = constexpr_affine_dequantize()[axis = tensor(0), name = tensor("attn1_out_proj_weight_cast_fp16"), quantized_data = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15762304))), scale = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(16810944))), zero_point = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(4203840)))]; + tensor linear1_1_weight_cast_fp16 = constexpr_affine_dequantize()[axis = tensor(0), name = tensor("linear1_1_weight_cast_fp16"), quantized_data = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(16813056))), scale = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21007424))), zero_point = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8401408)))]; + tensor linear1_2_weight_cast_fp16 = constexpr_affine_dequantize()[axis = tensor(0), name = tensor("linear1_2_weight_cast_fp16"), quantized_data = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21015680))), scale = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(25210048))), zero_point = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(4203840)))]; + tensor attn2_in_proj_weight_cast_fp16 = constexpr_affine_dequantize()[axis = tensor(0), name = tensor("attn2_in_proj_weight_cast_fp16"), quantized_data = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(25212160))), scale = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28357952))), zero_point = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(3145856)))]; + tensor attn2_out_proj_weight_cast_fp16 = constexpr_affine_dequantize()[axis = tensor(0), name = tensor("attn2_out_proj_weight_cast_fp16"), quantized_data = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28364160))), scale = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(29412800))), zero_point = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(4203840)))]; + tensor linear2_1_weight_cast_fp16 = constexpr_affine_dequantize()[axis = tensor(0), name = tensor("linear2_1_weight_cast_fp16"), quantized_data = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(29414912))), scale = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(33609280))), zero_point = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8401408)))]; + tensor linear2_2_weight_cast_fp16 = constexpr_affine_dequantize()[axis = tensor(0), name = tensor("linear2_2_weight_cast_fp16"), quantized_data = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(33617536))), scale = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(37811904))), zero_point = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(4203840)))]; + tensor attn3_in_proj_weight_cast_fp16 = constexpr_affine_dequantize()[axis = tensor(0), name = tensor("attn3_in_proj_weight_cast_fp16"), quantized_data = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(37814016))), scale = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(40959808))), zero_point = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(3145856)))]; + tensor attn3_out_proj_weight_cast_fp16 = constexpr_affine_dequantize()[axis = tensor(0), name = tensor("attn3_out_proj_weight_cast_fp16"), quantized_data = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(40966016))), scale = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(42014656))), zero_point = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(4203840)))]; + tensor linear3_1_weight_cast_fp16 = constexpr_affine_dequantize()[axis = tensor(0), name = tensor("linear3_1_weight_cast_fp16"), quantized_data = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(42016768))), scale = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(46211136))), zero_point = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8401408)))]; + tensor linear3_2_weight_cast_fp16 = constexpr_affine_dequantize()[axis = tensor(0), name = tensor("linear3_2_weight_cast_fp16"), quantized_data = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(46219392))), scale = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(50413760))), zero_point = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(4203840)))]; + tensor attn4_in_proj_weight_cast_fp16 = constexpr_affine_dequantize()[axis = tensor(0), name = tensor("attn4_in_proj_weight_cast_fp16"), quantized_data = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(50415872))), scale = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(53561664))), zero_point = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(3145856)))]; + tensor attn4_out_proj_weight_cast_fp16 = constexpr_affine_dequantize()[axis = tensor(0), name = tensor("attn4_out_proj_weight_cast_fp16"), quantized_data = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(53567872))), scale = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(54616512))), zero_point = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(4203840)))]; + tensor linear4_1_weight_cast_fp16 = constexpr_affine_dequantize()[axis = tensor(0), name = tensor("linear4_1_weight_cast_fp16"), quantized_data = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(54618624))), scale = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(58812992))), zero_point = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8401408)))]; + tensor linear4_2_weight_cast_fp16 = constexpr_affine_dequantize()[axis = tensor(0), name = tensor("linear4_2_weight_cast_fp16"), quantized_data = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(58821248))), scale = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(63015616))), zero_point = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(4203840)))]; + tensor attn5_in_proj_weight_cast_fp16 = constexpr_affine_dequantize()[axis = tensor(0), name = tensor("attn5_in_proj_weight_cast_fp16"), quantized_data = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(63017728))), scale = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(66163520))), zero_point = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(3145856)))]; + tensor attn5_out_proj_weight_cast_fp16 = constexpr_affine_dequantize()[axis = tensor(0), name = tensor("attn5_out_proj_weight_cast_fp16"), quantized_data = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(66169728))), scale = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(67218368))), zero_point = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(4203840)))]; + tensor linear5_1_weight_cast_fp16 = constexpr_affine_dequantize()[axis = tensor(0), name = tensor("linear5_1_weight_cast_fp16"), quantized_data = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(67220480))), scale = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(71414848))), zero_point = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8401408)))]; + tensor linear5_2_weight_cast_fp16 = constexpr_affine_dequantize()[axis = tensor(0), name = tensor("linear5_2_weight_cast_fp16"), quantized_data = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(71423104))), scale = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(75617472))), zero_point = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(4203840)))]; + tensor sequence_to_fp16_dtype_0 = const()[name = tensor("sequence_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor sequence_to_fp16 = cast(dtype = sequence_to_fp16_dtype_0, x = sequence)[name = tensor("cast_141")]; + tensor var_54_cast_fp16 = not_equal(x = sequence_to_fp16, y = sequence_to_fp16)[name = tensor("op_54_cast_fp16")]; + tensor expand_dims_0_axes_0 = const()[name = tensor("expand_dims_0_axes_0"), val = tensor([0, 1])]; + tensor bos_emb_to_fp16_dtype_0 = const()[name = tensor("bos_emb_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor bos_emb_to_fp16 = cast(dtype = bos_emb_to_fp16_dtype_0, x = bos_emb)[name = tensor("cast_140")]; + tensor expand_dims_0_cast_fp16 = expand_dims(axes = expand_dims_0_axes_0, x = bos_emb_to_fp16)[name = tensor("expand_dims_0_cast_fp16")]; + tensor input_1_cast_fp16 = select(a = expand_dims_0_cast_fp16, b = sequence_to_fp16, cond = var_54_cast_fp16)[name = tensor("input_1_cast_fp16")]; + tensor input_linear_weight_to_fp16 = const()[name = tensor("input_linear_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(75619584)))]; + tensor linear_0_bias_0_to_fp16 = const()[name = tensor("linear_0_bias_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(75685184)))]; + tensor linear_0_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = input_linear_weight_to_fp16, x = input_1_cast_fp16)[name = tensor("linear_0_cast_fp16")]; + tensor x_1_axes_0 = const()[name = tensor("x_1_axes_0"), val = tensor([-1])]; + tensor norm0_1_weight_to_fp16 = const()[name = tensor("norm0_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(75687296)))]; + tensor norm0_1_bias_to_fp16 = const()[name = tensor("norm0_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(75689408)))]; + tensor var_60_to_fp16 = const()[name = tensor("op_60_to_fp16"), val = tensor(0x1.5p-17)]; + tensor x_1_cast_fp16 = layer_norm(axes = x_1_axes_0, beta = norm0_1_bias_to_fp16, epsilon = var_60_to_fp16, gamma = norm0_1_weight_to_fp16, x = linear_0_cast_fp16)[name = tensor("x_1_cast_fp16")]; + tensor linear_1_bias_0_to_fp16 = const()[name = tensor("linear_1_bias_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(75691520)))]; + tensor linear_1_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = attn0_in_proj_weight_cast_fp16, x = x_1_cast_fp16)[name = tensor("linear_1_cast_fp16")]; + tensor var_96 = const()[name = tensor("op_96"), val = tensor([1, 1, 3, 16, 64])]; + tensor qkv_1_cast_fp16 = reshape(shape = var_96, x = linear_1_cast_fp16)[name = tensor("qkv_1_cast_fp16")]; + tensor q_1_begin_0 = const()[name = tensor("q_1_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor q_1_end_0 = const()[name = tensor("q_1_end_0"), val = tensor([1, 1, 1, 16, 64])]; + tensor q_1_end_mask_0 = const()[name = tensor("q_1_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor q_1_squeeze_mask_0 = const()[name = tensor("q_1_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor q_1_cast_fp16 = slice_by_index(begin = q_1_begin_0, end = q_1_end_0, end_mask = q_1_end_mask_0, squeeze_mask = q_1_squeeze_mask_0, x = qkv_1_cast_fp16)[name = tensor("q_1_cast_fp16")]; + tensor k_1_begin_0 = const()[name = tensor("k_1_begin_0"), val = tensor([0, 0, 1, 0, 0])]; + tensor k_1_end_0 = const()[name = tensor("k_1_end_0"), val = tensor([1, 1, 2, 16, 64])]; + tensor k_1_end_mask_0 = const()[name = tensor("k_1_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor k_1_squeeze_mask_0 = const()[name = tensor("k_1_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor k_1_cast_fp16 = slice_by_index(begin = k_1_begin_0, end = k_1_end_0, end_mask = k_1_end_mask_0, squeeze_mask = k_1_squeeze_mask_0, x = qkv_1_cast_fp16)[name = tensor("k_1_cast_fp16")]; + tensor v_1_begin_0 = const()[name = tensor("v_1_begin_0"), val = tensor([0, 0, 2, 0, 0])]; + tensor v_1_end_0 = const()[name = tensor("v_1_end_0"), val = tensor([1, 1, 3, 16, 64])]; + tensor v_1_end_mask_0 = const()[name = tensor("v_1_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor v_1_squeeze_mask_0 = const()[name = tensor("v_1_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor v_1_cast_fp16 = slice_by_index(begin = v_1_begin_0, end = v_1_end_0, end_mask = v_1_end_mask_0, squeeze_mask = v_1_squeeze_mask_0, x = qkv_1_cast_fp16)[name = tensor("v_1_cast_fp16")]; + tensor var_200 = const()[name = tensor("op_200"), val = tensor([1, 1, 1, 1])]; + tensor position0_to_fp16_dtype_0 = const()[name = tensor("position0_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor position0_to_fp16 = cast(dtype = position0_to_fp16_dtype_0, x = position0)[name = tensor("cast_139")]; + tensor ts_5_cast_fp16 = reshape(shape = var_200, x = position0_to_fp16)[name = tensor("ts_5_cast_fp16")]; + tensor var_204 = const()[name = tensor("op_204"), val = tensor([1, 1, 16, 32, 2])]; + tensor q_complex_1_cast_fp16 = reshape(shape = var_204, x = q_1_cast_fp16)[name = tensor("q_complex_1_cast_fp16")]; + tensor var_208 = const()[name = tensor("op_208"), val = tensor([1, 1, 16, 32, 2])]; + tensor k_complex_1_cast_fp16 = reshape(shape = var_208, x = k_1_cast_fp16)[name = tensor("k_complex_1_cast_fp16")]; + tensor var_212_begin_0 = const()[name = tensor("op_212_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_212_end_0 = const()[name = tensor("op_212_end_0"), val = tensor([1, 1, 16, 32, 1])]; + tensor var_212_end_mask_0 = const()[name = tensor("op_212_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_212_squeeze_mask_0 = const()[name = tensor("op_212_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_212_cast_fp16 = slice_by_index(begin = var_212_begin_0, end = var_212_end_0, end_mask = var_212_end_mask_0, squeeze_mask = var_212_squeeze_mask_0, x = q_complex_1_cast_fp16)[name = tensor("op_212_cast_fp16")]; + tensor var_220_begin_0 = const()[name = tensor("op_220_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_220_end_0 = const()[name = tensor("op_220_end_0"), val = tensor([1, 1, 16, 32, 2])]; + tensor var_220_end_mask_0 = const()[name = tensor("op_220_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_220_squeeze_mask_0 = const()[name = tensor("op_220_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_220_cast_fp16 = slice_by_index(begin = var_220_begin_0, end = var_220_end_0, end_mask = var_220_end_mask_0, squeeze_mask = var_220_squeeze_mask_0, x = q_complex_1_cast_fp16)[name = tensor("op_220_cast_fp16")]; + tensor var_228_begin_0 = const()[name = tensor("op_228_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_228_end_0 = const()[name = tensor("op_228_end_0"), val = tensor([1, 1, 16, 32, 1])]; + tensor var_228_end_mask_0 = const()[name = tensor("op_228_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_228_squeeze_mask_0 = const()[name = tensor("op_228_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_228_cast_fp16 = slice_by_index(begin = var_228_begin_0, end = var_228_end_0, end_mask = var_228_end_mask_0, squeeze_mask = var_228_squeeze_mask_0, x = k_complex_1_cast_fp16)[name = tensor("op_228_cast_fp16")]; + tensor var_236_begin_0 = const()[name = tensor("op_236_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_236_end_0 = const()[name = tensor("op_236_end_0"), val = tensor([1, 1, 16, 32, 2])]; + tensor var_236_end_mask_0 = const()[name = tensor("op_236_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_236_squeeze_mask_0 = const()[name = tensor("op_236_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_236_cast_fp16 = slice_by_index(begin = var_236_begin_0, end = var_236_end_0, end_mask = var_236_end_mask_0, squeeze_mask = var_236_squeeze_mask_0, x = k_complex_1_cast_fp16)[name = tensor("op_236_cast_fp16")]; + tensor freqs_1_to_fp16 = const()[name = tensor("freqs_1_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(75697728)))]; + tensor var_242_cast_fp16 = mul(x = freqs_1_to_fp16, y = ts_5_cast_fp16)[name = tensor("op_242_cast_fp16")]; + tensor rotr_1_cast_fp16 = cos(x = var_242_cast_fp16)[name = tensor("rotr_1_cast_fp16")]; + tensor roti_1_cast_fp16 = sin(x = var_242_cast_fp16)[name = tensor("roti_1_cast_fp16")]; + tensor var_246_cast_fp16 = mul(x = var_212_cast_fp16, y = rotr_1_cast_fp16)[name = tensor("op_246_cast_fp16")]; + tensor var_247_cast_fp16 = mul(x = var_220_cast_fp16, y = roti_1_cast_fp16)[name = tensor("op_247_cast_fp16")]; + tensor qor_1_cast_fp16 = sub(x = var_246_cast_fp16, y = var_247_cast_fp16)[name = tensor("qor_1_cast_fp16")]; + tensor var_250_cast_fp16 = mul(x = var_212_cast_fp16, y = roti_1_cast_fp16)[name = tensor("op_250_cast_fp16")]; + tensor var_251_cast_fp16 = mul(x = var_220_cast_fp16, y = rotr_1_cast_fp16)[name = tensor("op_251_cast_fp16")]; + tensor qoi_1_cast_fp16 = add(x = var_250_cast_fp16, y = var_251_cast_fp16)[name = tensor("qoi_1_cast_fp16")]; + tensor var_254_cast_fp16 = mul(x = var_228_cast_fp16, y = rotr_1_cast_fp16)[name = tensor("op_254_cast_fp16")]; + tensor var_255_cast_fp16 = mul(x = var_236_cast_fp16, y = roti_1_cast_fp16)[name = tensor("op_255_cast_fp16")]; + tensor kor_1_cast_fp16 = sub(x = var_254_cast_fp16, y = var_255_cast_fp16)[name = tensor("kor_1_cast_fp16")]; + tensor var_258_cast_fp16 = mul(x = var_228_cast_fp16, y = roti_1_cast_fp16)[name = tensor("op_258_cast_fp16")]; + tensor var_259_cast_fp16 = mul(x = var_236_cast_fp16, y = rotr_1_cast_fp16)[name = tensor("op_259_cast_fp16")]; + tensor koi_1_cast_fp16 = add(x = var_258_cast_fp16, y = var_259_cast_fp16)[name = tensor("koi_1_cast_fp16")]; + tensor qo_1_axis_0 = const()[name = tensor("qo_1_axis_0"), val = tensor(-1)]; + tensor qo_1_cast_fp16 = stack(axis = qo_1_axis_0, values = (qor_1_cast_fp16, qoi_1_cast_fp16))[name = tensor("qo_1_cast_fp16")]; + tensor ko_1_axis_0 = const()[name = tensor("ko_1_axis_0"), val = tensor(-1)]; + tensor ko_1_cast_fp16 = stack(axis = ko_1_axis_0, values = (kor_1_cast_fp16, koi_1_cast_fp16))[name = tensor("ko_1_cast_fp16")]; + tensor var_288 = const()[name = tensor("op_288"), val = tensor([1, 1, 16, 64])]; + tensor q_3_cast_fp16 = reshape(shape = var_288, x = qo_1_cast_fp16)[name = tensor("q_3_cast_fp16")]; + tensor var_290 = const()[name = tensor("op_290"), val = tensor([1, 1, 16, 64])]; + tensor k_3_cast_fp16 = reshape(shape = var_290, x = ko_1_cast_fp16)[name = tensor("k_3_cast_fp16")]; + tensor _inversed_312_y_0_to_fp16 = const()[name = tensor("_inversed_312_y_0_to_fp16"), val = tensor(0x1p-9)]; + tensor _inversed_312_cast_fp16 = mul(x = ts_5_cast_fp16, y = _inversed_312_y_0_to_fp16)[name = tensor("_inversed_312_cast_fp16")]; + tensor var_313_cast_fp16 = floor(x = _inversed_312_cast_fp16)[name = tensor("op_313_cast_fp16")]; + tensor var_314_to_fp16 = const()[name = tensor("op_314_to_fp16"), val = tensor(0x1p+9)]; + tensor var_315_cast_fp16 = mul(x = var_313_cast_fp16, y = var_314_to_fp16)[name = tensor("op_315_cast_fp16")]; + tensor write_indices_float_3_cast_fp16 = sub(x = ts_5_cast_fp16, y = var_315_cast_fp16)[name = tensor("write_indices_float_3_cast_fp16")]; + tensor var_322_dtype_0 = const()[name = tensor("op_322_dtype_0"), val = tensor("int32")]; + tensor write_indices_1_reps_0 = const()[name = tensor("write_indices_1_reps_0"), val = tensor([1, 1, 16, 64])]; + tensor write_indices_float_3_cast_fp16_to_int32 = cast(dtype = var_322_dtype_0, x = write_indices_float_3_cast_fp16)[name = tensor("cast_138")]; + tensor write_indices_1 = tile(reps = write_indices_1_reps_0, x = write_indices_float_3_cast_fp16_to_int32)[name = tensor("write_indices_1")]; + tensor var_330_begin_0 = const()[name = tensor("op_330_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_330_end_0 = const()[name = tensor("op_330_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor var_330_end_mask_0 = const()[name = tensor("op_330_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_330_squeeze_mask_0 = const()[name = tensor("op_330_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor cache0_to_fp16_dtype_0 = const()[name = tensor("cache0_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor cache0_to_fp16 = cast(dtype = cache0_to_fp16_dtype_0, x = cache0)[name = tensor("cast_137")]; + tensor var_330_cast_fp16 = slice_by_index(begin = var_330_begin_0, end = var_330_end_0, end_mask = var_330_end_mask_0, squeeze_mask = var_330_squeeze_mask_0, x = cache0_to_fp16)[name = tensor("op_330_cast_fp16")]; + tensor var_332_axis_0 = const()[name = tensor("op_332_axis_0"), val = tensor(1)]; + tensor var_332_mode_0 = const()[name = tensor("op_332_mode_0"), val = tensor("update")]; + tensor var_332_validate_indices_0 = const()[name = tensor("op_332_validate_indices_0"), val = tensor(false)]; + tensor var_332_cast_fp16 = scatter_along_axis(axis = var_332_axis_0, data = var_330_cast_fp16, indices = write_indices_1, mode = var_332_mode_0, updates = k_3_cast_fp16, validate_indices = var_332_validate_indices_0)[name = tensor("op_332_cast_fp16")]; + tensor concat_2 = const()[name = tensor("concat_2"), val = tensor([0, 0, 0, 0, 0])]; + tensor concat_3 = const()[name = tensor("concat_3"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_1_internal_tensor_assign_1_stride_0 = const()[name = tensor("new_cache_1_internal_tensor_assign_1_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_1_internal_tensor_assign_1_begin_mask_0 = const()[name = tensor("new_cache_1_internal_tensor_assign_1_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_1_internal_tensor_assign_1_end_mask_0 = const()[name = tensor("new_cache_1_internal_tensor_assign_1_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_1_internal_tensor_assign_1_squeeze_mask_0 = const()[name = tensor("new_cache_1_internal_tensor_assign_1_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_12 = const()[name = tensor("shape_12"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_0 = const()[name = tensor("reduce_prod_0"), val = tensor(1048576)]; + tensor range_1d_0_start_0 = const()[name = tensor("range_1d_0_start_0"), val = tensor(0)]; + tensor range_1d_0_step_0 = const()[name = tensor("range_1d_0_step_0"), val = tensor(1)]; + tensor range_1d_0 = range_1d(end = reduce_prod_0, start = range_1d_0_start_0, step = range_1d_0_step_0)[name = tensor("range_1d_0")]; + tensor reshape_0 = reshape(shape = shape_12, x = range_1d_0)[name = tensor("reshape_0")]; + tensor slice_by_index_0 = slice_by_index(begin = concat_2, begin_mask = new_cache_1_internal_tensor_assign_1_begin_mask_0, end = concat_3, end_mask = new_cache_1_internal_tensor_assign_1_end_mask_0, squeeze_mask = new_cache_1_internal_tensor_assign_1_squeeze_mask_0, stride = new_cache_1_internal_tensor_assign_1_stride_0, x = reshape_0)[name = tensor("slice_by_index_0")]; + tensor reshape_1_shape_0 = const()[name = tensor("reshape_1_shape_0"), val = tensor([-1])]; + tensor reshape_1 = reshape(shape = reshape_1_shape_0, x = slice_by_index_0)[name = tensor("reshape_1")]; + tensor reshape_2_shape_0 = const()[name = tensor("reshape_2_shape_0"), val = tensor([-1])]; + tensor reshape_2_cast_fp16 = reshape(shape = reshape_2_shape_0, x = var_332_cast_fp16)[name = tensor("reshape_2_cast_fp16")]; + tensor reshape_3_shape_0 = const()[name = tensor("reshape_3_shape_0"), val = tensor([-1])]; + tensor reshape_3_cast_fp16 = reshape(shape = reshape_3_shape_0, x = cache0_to_fp16)[name = tensor("reshape_3_cast_fp16")]; + tensor scatter_0_mode_0 = const()[name = tensor("scatter_0_mode_0"), val = tensor("update")]; + tensor scatter_0_axis_0 = const()[name = tensor("scatter_0_axis_0"), val = tensor(0)]; + tensor scatter_0_validate_indices_0 = const()[name = tensor("scatter_0_validate_indices_0"), val = tensor(false)]; + tensor scatter_0_cast_fp16 = scatter(axis = scatter_0_axis_0, data = reshape_3_cast_fp16, indices = reshape_1, mode = scatter_0_mode_0, updates = reshape_2_cast_fp16, validate_indices = scatter_0_validate_indices_0)[name = tensor("scatter_0_cast_fp16")]; + tensor reshape_4_cast_fp16 = reshape(shape = shape_12, x = scatter_0_cast_fp16)[name = tensor("reshape_4_cast_fp16")]; + tensor var_340_begin_0 = const()[name = tensor("op_340_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor var_340_end_0 = const()[name = tensor("op_340_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor var_340_end_mask_0 = const()[name = tensor("op_340_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_340_squeeze_mask_0 = const()[name = tensor("op_340_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor var_340_cast_fp16 = slice_by_index(begin = var_340_begin_0, end = var_340_end_0, end_mask = var_340_end_mask_0, squeeze_mask = var_340_squeeze_mask_0, x = reshape_4_cast_fp16)[name = tensor("op_340_cast_fp16")]; + tensor var_342_axis_0 = const()[name = tensor("op_342_axis_0"), val = tensor(1)]; + tensor var_342_mode_0 = const()[name = tensor("op_342_mode_0"), val = tensor("update")]; + tensor var_342_validate_indices_0 = const()[name = tensor("op_342_validate_indices_0"), val = tensor(false)]; + tensor var_342_cast_fp16 = scatter_along_axis(axis = var_342_axis_0, data = var_340_cast_fp16, indices = write_indices_1, mode = var_342_mode_0, updates = v_1_cast_fp16, validate_indices = var_342_validate_indices_0)[name = tensor("op_342_cast_fp16")]; + tensor concat_4 = const()[name = tensor("concat_4"), val = tensor([1, 0, 0, 0, 0])]; + tensor concat_5 = const()[name = tensor("concat_5"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_1_internal_tensor_assign_2_stride_0 = const()[name = tensor("new_cache_1_internal_tensor_assign_2_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_1_internal_tensor_assign_2_begin_mask_0 = const()[name = tensor("new_cache_1_internal_tensor_assign_2_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_1_internal_tensor_assign_2_end_mask_0 = const()[name = tensor("new_cache_1_internal_tensor_assign_2_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_1_internal_tensor_assign_2_squeeze_mask_0 = const()[name = tensor("new_cache_1_internal_tensor_assign_2_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_13 = const()[name = tensor("shape_13"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_1 = const()[name = tensor("reduce_prod_1"), val = tensor(1048576)]; + tensor range_1d_1_start_0 = const()[name = tensor("range_1d_1_start_0"), val = tensor(0)]; + tensor range_1d_1_step_0 = const()[name = tensor("range_1d_1_step_0"), val = tensor(1)]; + tensor range_1d_1 = range_1d(end = reduce_prod_1, start = range_1d_1_start_0, step = range_1d_1_step_0)[name = tensor("range_1d_1")]; + tensor reshape_5 = reshape(shape = shape_13, x = range_1d_1)[name = tensor("reshape_5")]; + tensor slice_by_index_1 = slice_by_index(begin = concat_4, begin_mask = new_cache_1_internal_tensor_assign_2_begin_mask_0, end = concat_5, end_mask = new_cache_1_internal_tensor_assign_2_end_mask_0, squeeze_mask = new_cache_1_internal_tensor_assign_2_squeeze_mask_0, stride = new_cache_1_internal_tensor_assign_2_stride_0, x = reshape_5)[name = tensor("slice_by_index_1")]; + tensor reshape_6_shape_0 = const()[name = tensor("reshape_6_shape_0"), val = tensor([-1])]; + tensor reshape_6 = reshape(shape = reshape_6_shape_0, x = slice_by_index_1)[name = tensor("reshape_6")]; + tensor reshape_7_shape_0 = const()[name = tensor("reshape_7_shape_0"), val = tensor([-1])]; + tensor reshape_7_cast_fp16 = reshape(shape = reshape_7_shape_0, x = var_342_cast_fp16)[name = tensor("reshape_7_cast_fp16")]; + tensor reshape_8_shape_0 = const()[name = tensor("reshape_8_shape_0"), val = tensor([-1])]; + tensor reshape_8_cast_fp16 = reshape(shape = reshape_8_shape_0, x = reshape_4_cast_fp16)[name = tensor("reshape_8_cast_fp16")]; + tensor scatter_1_mode_0 = const()[name = tensor("scatter_1_mode_0"), val = tensor("update")]; + tensor scatter_1_axis_0 = const()[name = tensor("scatter_1_axis_0"), val = tensor(0)]; + tensor scatter_1_validate_indices_0 = const()[name = tensor("scatter_1_validate_indices_0"), val = tensor(false)]; + tensor scatter_1_cast_fp16 = scatter(axis = scatter_1_axis_0, data = reshape_8_cast_fp16, indices = reshape_6, mode = scatter_1_mode_0, updates = reshape_7_cast_fp16, validate_indices = scatter_1_validate_indices_0)[name = tensor("scatter_1_cast_fp16")]; + tensor reshape_9_cast_fp16 = reshape(shape = shape_13, x = scatter_1_cast_fp16)[name = tensor("reshape_9_cast_fp16")]; + tensor reshape_9_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("reshape_9_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor keys_1_begin_0 = const()[name = tensor("keys_1_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor keys_1_end_0 = const()[name = tensor("keys_1_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor keys_1_end_mask_0 = const()[name = tensor("keys_1_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor keys_1_squeeze_mask_0 = const()[name = tensor("keys_1_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor keys_1_cast_fp16 = slice_by_index(begin = keys_1_begin_0, end = keys_1_end_0, end_mask = keys_1_end_mask_0, squeeze_mask = keys_1_squeeze_mask_0, x = reshape_9_cast_fp16)[name = tensor("keys_1_cast_fp16")]; + tensor values_1_begin_0 = const()[name = tensor("values_1_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor values_1_end_0 = const()[name = tensor("values_1_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor values_1_end_mask_0 = const()[name = tensor("values_1_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor values_1_squeeze_mask_0 = const()[name = tensor("values_1_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor values_1_cast_fp16 = slice_by_index(begin = values_1_begin_0, end = values_1_end_0, end_mask = values_1_end_mask_0, squeeze_mask = values_1_squeeze_mask_0, x = reshape_9_cast_fp16)[name = tensor("values_1_cast_fp16")]; + tensor var_354_cast_fp16 = not_equal(x = keys_1_cast_fp16, y = keys_1_cast_fp16)[name = tensor("op_354_cast_fp16")]; + tensor var_360_to_fp16 = const()[name = tensor("op_360_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(75697856)))]; + tensor keys_3_cast_fp16 = select(a = var_360_to_fp16, b = keys_1_cast_fp16, cond = var_354_cast_fp16)[name = tensor("keys_3_cast_fp16")]; + tensor var_362_cast_fp16 = not_equal(x = values_1_cast_fp16, y = values_1_cast_fp16)[name = tensor("op_362_cast_fp16")]; + tensor values_3_cast_fp16 = select(a = var_360_to_fp16, b = values_1_cast_fp16, cond = var_362_cast_fp16)[name = tensor("values_3_cast_fp16")]; + tensor var_386 = const()[name = tensor("op_386"), val = tensor([0, 2, 1, 3])]; + tensor var_399 = const()[name = tensor("op_399"), val = tensor([1, 1, 1])]; + tensor var_400_cast_fp16 = reshape(shape = var_399, x = position0_to_fp16)[name = tensor("op_400_cast_fp16")]; + tensor var_417_to_fp16 = const()[name = tensor("op_417_to_fp16"), val = tensor(0x1p+0)]; + tensor valid_len_1_cast_fp16 = add(x = var_400_cast_fp16, y = var_417_to_fp16)[name = tensor("valid_len_1_cast_fp16")]; + tensor k_positions_1_promoted_to_fp16 = const()[name = tensor("k_positions_1_promoted_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(76746496)))]; + tensor valid_mask_1_cast_fp16 = less(x = k_positions_1_promoted_to_fp16, y = valid_len_1_cast_fp16)[name = tensor("valid_mask_1_cast_fp16")]; + tensor causal_mask_1_cast_fp16 = less_equal(x = k_positions_1_promoted_to_fp16, y = var_400_cast_fp16)[name = tensor("causal_mask_1_cast_fp16")]; + tensor attn_mask_1 = logical_and(x = valid_mask_1_cast_fp16, y = causal_mask_1_cast_fp16)[name = tensor("attn_mask_1")]; + tensor attn_mask_3_axes_0 = const()[name = tensor("attn_mask_3_axes_0"), val = tensor([1])]; + tensor attn_mask_3 = expand_dims(axes = attn_mask_3_axes_0, x = attn_mask_1)[name = tensor("attn_mask_3")]; + tensor var_435_transpose_x_0 = const()[name = tensor("op_435_transpose_x_0"), val = tensor(false)]; + tensor var_435_transpose_y_0 = const()[name = tensor("op_435_transpose_y_0"), val = tensor(false)]; + tensor transpose_18_perm_0 = const()[name = tensor("transpose_18_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_19_perm_0 = const()[name = tensor("transpose_19_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_19 = transpose(perm = transpose_19_perm_0, x = keys_3_cast_fp16)[name = tensor("transpose_51")]; + tensor transpose_18 = transpose(perm = transpose_18_perm_0, x = q_3_cast_fp16)[name = tensor("transpose_52")]; + tensor var_435_cast_fp16 = matmul(transpose_x = var_435_transpose_x_0, transpose_y = var_435_transpose_y_0, x = transpose_18, y = transpose_19)[name = tensor("op_435_cast_fp16")]; + tensor var_429_to_fp16 = const()[name = tensor("op_429_to_fp16"), val = tensor([0x1p-3])]; + tensor attn_weights_1_cast_fp16 = mul(x = var_435_cast_fp16, y = var_429_to_fp16)[name = tensor("attn_weights_1_cast_fp16")]; + tensor var_437 = logical_not(x = attn_mask_3)[name = tensor("op_437")]; + tensor var_438_to_fp16 = const()[name = tensor("op_438_to_fp16"), val = tensor(-inf)]; + tensor attn_weights_3_cast_fp16 = select(a = var_438_to_fp16, b = attn_weights_1_cast_fp16, cond = var_437)[name = tensor("attn_weights_3_cast_fp16")]; + tensor var_440 = const()[name = tensor("op_440"), val = tensor(-1)]; + tensor attn_weights_5_cast_fp16 = softmax(axis = var_440, x = attn_weights_3_cast_fp16)[name = tensor("attn_weights_5_cast_fp16")]; + tensor attn_output_1_transpose_x_0 = const()[name = tensor("attn_output_1_transpose_x_0"), val = tensor(false)]; + tensor attn_output_1_transpose_y_0 = const()[name = tensor("attn_output_1_transpose_y_0"), val = tensor(false)]; + tensor values_5_cast_fp16 = transpose(perm = var_386, x = values_3_cast_fp16)[name = tensor("transpose_53")]; + tensor attn_output_1_cast_fp16 = matmul(transpose_x = attn_output_1_transpose_x_0, transpose_y = attn_output_1_transpose_y_0, x = attn_weights_5_cast_fp16, y = values_5_cast_fp16)[name = tensor("attn_output_1_cast_fp16")]; + tensor var_448 = const()[name = tensor("op_448"), val = tensor([0, 2, 1, 3])]; + tensor var_451 = const()[name = tensor("op_451"), val = tensor([1, 1, 1024])]; + tensor var_449_cast_fp16 = transpose(perm = var_448, x = attn_output_1_cast_fp16)[name = tensor("transpose_50")]; + tensor input_5_cast_fp16 = reshape(shape = var_451, x = var_449_cast_fp16)[name = tensor("input_5_cast_fp16")]; + tensor linear_2_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = attn0_out_proj_weight_cast_fp16, x = input_5_cast_fp16)[name = tensor("linear_2_cast_fp16")]; + tensor var_457_to_fp16 = const()[name = tensor("op_457_to_fp16"), val = tensor(0x1p+0)]; + tensor var_458_cast_fp16 = add(x = position0_to_fp16, y = var_457_to_fp16)[name = tensor("op_458_cast_fp16")]; + tensor var_458_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_458_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor input_7_cast_fp16 = add(x = linear_0_cast_fp16, y = linear_2_cast_fp16)[name = tensor("input_7_cast_fp16")]; + tensor input_9_axes_0 = const()[name = tensor("input_9_axes_0"), val = tensor([-1])]; + tensor norm0_2_weight_to_fp16 = const()[name = tensor("norm0_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(76747584)))]; + tensor norm0_2_bias_to_fp16 = const()[name = tensor("norm0_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(76749696)))]; + tensor var_462_to_fp16 = const()[name = tensor("op_462_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_9_cast_fp16 = layer_norm(axes = input_9_axes_0, beta = norm0_2_bias_to_fp16, epsilon = var_462_to_fp16, gamma = norm0_2_weight_to_fp16, x = input_7_cast_fp16)[name = tensor("input_9_cast_fp16")]; + tensor linear_3_bias_0_to_fp16 = const()[name = tensor("linear_3_bias_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(76751808)))]; + tensor linear_3_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = linear0_1_weight_cast_fp16, x = input_9_cast_fp16)[name = tensor("linear_3_cast_fp16")]; + tensor input_11_mode_0 = const()[name = tensor("input_11_mode_0"), val = tensor("EXACT")]; + tensor input_11_cast_fp16 = gelu(mode = input_11_mode_0, x = linear_3_cast_fp16)[name = tensor("input_11_cast_fp16")]; + tensor linear_4_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = linear0_2_weight_cast_fp16, x = input_11_cast_fp16)[name = tensor("linear_4_cast_fp16")]; + tensor input_13_cast_fp16 = add(x = input_7_cast_fp16, y = linear_4_cast_fp16)[name = tensor("input_13_cast_fp16")]; + tensor x_3_axes_0 = const()[name = tensor("x_3_axes_0"), val = tensor([-1])]; + tensor norm1_1_weight_to_fp16 = const()[name = tensor("norm1_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(76760064)))]; + tensor norm1_1_bias_to_fp16 = const()[name = tensor("norm1_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(76762176)))]; + tensor var_479_to_fp16 = const()[name = tensor("op_479_to_fp16"), val = tensor(0x1.5p-17)]; + tensor x_3_cast_fp16 = layer_norm(axes = x_3_axes_0, beta = norm1_1_bias_to_fp16, epsilon = var_479_to_fp16, gamma = norm1_1_weight_to_fp16, x = input_13_cast_fp16)[name = tensor("x_3_cast_fp16")]; + tensor linear_5_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = attn1_in_proj_weight_cast_fp16, x = x_3_cast_fp16)[name = tensor("linear_5_cast_fp16")]; + tensor var_515 = const()[name = tensor("op_515"), val = tensor([1, 1, 3, 16, 64])]; + tensor qkv_3_cast_fp16 = reshape(shape = var_515, x = linear_5_cast_fp16)[name = tensor("qkv_3_cast_fp16")]; + tensor q_7_begin_0 = const()[name = tensor("q_7_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor q_7_end_0 = const()[name = tensor("q_7_end_0"), val = tensor([1, 1, 1, 16, 64])]; + tensor q_7_end_mask_0 = const()[name = tensor("q_7_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor q_7_squeeze_mask_0 = const()[name = tensor("q_7_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor q_7_cast_fp16 = slice_by_index(begin = q_7_begin_0, end = q_7_end_0, end_mask = q_7_end_mask_0, squeeze_mask = q_7_squeeze_mask_0, x = qkv_3_cast_fp16)[name = tensor("q_7_cast_fp16")]; + tensor k_5_begin_0 = const()[name = tensor("k_5_begin_0"), val = tensor([0, 0, 1, 0, 0])]; + tensor k_5_end_0 = const()[name = tensor("k_5_end_0"), val = tensor([1, 1, 2, 16, 64])]; + tensor k_5_end_mask_0 = const()[name = tensor("k_5_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor k_5_squeeze_mask_0 = const()[name = tensor("k_5_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor k_5_cast_fp16 = slice_by_index(begin = k_5_begin_0, end = k_5_end_0, end_mask = k_5_end_mask_0, squeeze_mask = k_5_squeeze_mask_0, x = qkv_3_cast_fp16)[name = tensor("k_5_cast_fp16")]; + tensor v_3_begin_0 = const()[name = tensor("v_3_begin_0"), val = tensor([0, 0, 2, 0, 0])]; + tensor v_3_end_0 = const()[name = tensor("v_3_end_0"), val = tensor([1, 1, 3, 16, 64])]; + tensor v_3_end_mask_0 = const()[name = tensor("v_3_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor v_3_squeeze_mask_0 = const()[name = tensor("v_3_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor v_3_cast_fp16 = slice_by_index(begin = v_3_begin_0, end = v_3_end_0, end_mask = v_3_end_mask_0, squeeze_mask = v_3_squeeze_mask_0, x = qkv_3_cast_fp16)[name = tensor("v_3_cast_fp16")]; + tensor var_619 = const()[name = tensor("op_619"), val = tensor([1, 1, 1, 1])]; + tensor position1_to_fp16_dtype_0 = const()[name = tensor("position1_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor position1_to_fp16 = cast(dtype = position1_to_fp16_dtype_0, x = position1)[name = tensor("cast_134")]; + tensor ts_11_cast_fp16 = reshape(shape = var_619, x = position1_to_fp16)[name = tensor("ts_11_cast_fp16")]; + tensor var_623 = const()[name = tensor("op_623"), val = tensor([1, 1, 16, 32, 2])]; + tensor q_complex_3_cast_fp16 = reshape(shape = var_623, x = q_7_cast_fp16)[name = tensor("q_complex_3_cast_fp16")]; + tensor var_627 = const()[name = tensor("op_627"), val = tensor([1, 1, 16, 32, 2])]; + tensor k_complex_3_cast_fp16 = reshape(shape = var_627, x = k_5_cast_fp16)[name = tensor("k_complex_3_cast_fp16")]; + tensor var_631_begin_0 = const()[name = tensor("op_631_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_631_end_0 = const()[name = tensor("op_631_end_0"), val = tensor([1, 1, 16, 32, 1])]; + tensor var_631_end_mask_0 = const()[name = tensor("op_631_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_631_squeeze_mask_0 = const()[name = tensor("op_631_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_631_cast_fp16 = slice_by_index(begin = var_631_begin_0, end = var_631_end_0, end_mask = var_631_end_mask_0, squeeze_mask = var_631_squeeze_mask_0, x = q_complex_3_cast_fp16)[name = tensor("op_631_cast_fp16")]; + tensor var_639_begin_0 = const()[name = tensor("op_639_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_639_end_0 = const()[name = tensor("op_639_end_0"), val = tensor([1, 1, 16, 32, 2])]; + tensor var_639_end_mask_0 = const()[name = tensor("op_639_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_639_squeeze_mask_0 = const()[name = tensor("op_639_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_639_cast_fp16 = slice_by_index(begin = var_639_begin_0, end = var_639_end_0, end_mask = var_639_end_mask_0, squeeze_mask = var_639_squeeze_mask_0, x = q_complex_3_cast_fp16)[name = tensor("op_639_cast_fp16")]; + tensor var_647_begin_0 = const()[name = tensor("op_647_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_647_end_0 = const()[name = tensor("op_647_end_0"), val = tensor([1, 1, 16, 32, 1])]; + tensor var_647_end_mask_0 = const()[name = tensor("op_647_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_647_squeeze_mask_0 = const()[name = tensor("op_647_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_647_cast_fp16 = slice_by_index(begin = var_647_begin_0, end = var_647_end_0, end_mask = var_647_end_mask_0, squeeze_mask = var_647_squeeze_mask_0, x = k_complex_3_cast_fp16)[name = tensor("op_647_cast_fp16")]; + tensor var_655_begin_0 = const()[name = tensor("op_655_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_655_end_0 = const()[name = tensor("op_655_end_0"), val = tensor([1, 1, 16, 32, 2])]; + tensor var_655_end_mask_0 = const()[name = tensor("op_655_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_655_squeeze_mask_0 = const()[name = tensor("op_655_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_655_cast_fp16 = slice_by_index(begin = var_655_begin_0, end = var_655_end_0, end_mask = var_655_end_mask_0, squeeze_mask = var_655_squeeze_mask_0, x = k_complex_3_cast_fp16)[name = tensor("op_655_cast_fp16")]; + tensor freqs_3_to_fp16 = const()[name = tensor("freqs_3_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(76764288)))]; + tensor var_661_cast_fp16 = mul(x = freqs_3_to_fp16, y = ts_11_cast_fp16)[name = tensor("op_661_cast_fp16")]; + tensor rotr_3_cast_fp16 = cos(x = var_661_cast_fp16)[name = tensor("rotr_3_cast_fp16")]; + tensor roti_3_cast_fp16 = sin(x = var_661_cast_fp16)[name = tensor("roti_3_cast_fp16")]; + tensor var_665_cast_fp16 = mul(x = var_631_cast_fp16, y = rotr_3_cast_fp16)[name = tensor("op_665_cast_fp16")]; + tensor var_666_cast_fp16 = mul(x = var_639_cast_fp16, y = roti_3_cast_fp16)[name = tensor("op_666_cast_fp16")]; + tensor qor_5_cast_fp16 = sub(x = var_665_cast_fp16, y = var_666_cast_fp16)[name = tensor("qor_5_cast_fp16")]; + tensor var_669_cast_fp16 = mul(x = var_631_cast_fp16, y = roti_3_cast_fp16)[name = tensor("op_669_cast_fp16")]; + tensor var_670_cast_fp16 = mul(x = var_639_cast_fp16, y = rotr_3_cast_fp16)[name = tensor("op_670_cast_fp16")]; + tensor qoi_5_cast_fp16 = add(x = var_669_cast_fp16, y = var_670_cast_fp16)[name = tensor("qoi_5_cast_fp16")]; + tensor var_673_cast_fp16 = mul(x = var_647_cast_fp16, y = rotr_3_cast_fp16)[name = tensor("op_673_cast_fp16")]; + tensor var_674_cast_fp16 = mul(x = var_655_cast_fp16, y = roti_3_cast_fp16)[name = tensor("op_674_cast_fp16")]; + tensor kor_5_cast_fp16 = sub(x = var_673_cast_fp16, y = var_674_cast_fp16)[name = tensor("kor_5_cast_fp16")]; + tensor var_677_cast_fp16 = mul(x = var_647_cast_fp16, y = roti_3_cast_fp16)[name = tensor("op_677_cast_fp16")]; + tensor var_678_cast_fp16 = mul(x = var_655_cast_fp16, y = rotr_3_cast_fp16)[name = tensor("op_678_cast_fp16")]; + tensor koi_5_cast_fp16 = add(x = var_677_cast_fp16, y = var_678_cast_fp16)[name = tensor("koi_5_cast_fp16")]; + tensor qo_3_axis_0 = const()[name = tensor("qo_3_axis_0"), val = tensor(-1)]; + tensor qo_3_cast_fp16 = stack(axis = qo_3_axis_0, values = (qor_5_cast_fp16, qoi_5_cast_fp16))[name = tensor("qo_3_cast_fp16")]; + tensor ko_3_axis_0 = const()[name = tensor("ko_3_axis_0"), val = tensor(-1)]; + tensor ko_3_cast_fp16 = stack(axis = ko_3_axis_0, values = (kor_5_cast_fp16, koi_5_cast_fp16))[name = tensor("ko_3_cast_fp16")]; + tensor var_707 = const()[name = tensor("op_707"), val = tensor([1, 1, 16, 64])]; + tensor q_9_cast_fp16 = reshape(shape = var_707, x = qo_3_cast_fp16)[name = tensor("q_9_cast_fp16")]; + tensor var_709 = const()[name = tensor("op_709"), val = tensor([1, 1, 16, 64])]; + tensor k_7_cast_fp16 = reshape(shape = var_709, x = ko_3_cast_fp16)[name = tensor("k_7_cast_fp16")]; + tensor _inversed_731_y_0_to_fp16 = const()[name = tensor("_inversed_731_y_0_to_fp16"), val = tensor(0x1p-9)]; + tensor _inversed_731_cast_fp16 = mul(x = ts_11_cast_fp16, y = _inversed_731_y_0_to_fp16)[name = tensor("_inversed_731_cast_fp16")]; + tensor var_732_cast_fp16 = floor(x = _inversed_731_cast_fp16)[name = tensor("op_732_cast_fp16")]; + tensor var_733_to_fp16 = const()[name = tensor("op_733_to_fp16"), val = tensor(0x1p+9)]; + tensor var_734_cast_fp16 = mul(x = var_732_cast_fp16, y = var_733_to_fp16)[name = tensor("op_734_cast_fp16")]; + tensor write_indices_float_7_cast_fp16 = sub(x = ts_11_cast_fp16, y = var_734_cast_fp16)[name = tensor("write_indices_float_7_cast_fp16")]; + tensor var_741_dtype_0 = const()[name = tensor("op_741_dtype_0"), val = tensor("int32")]; + tensor write_indices_3_reps_0 = const()[name = tensor("write_indices_3_reps_0"), val = tensor([1, 1, 16, 64])]; + tensor write_indices_float_7_cast_fp16_to_int32 = cast(dtype = var_741_dtype_0, x = write_indices_float_7_cast_fp16)[name = tensor("cast_133")]; + tensor write_indices_3 = tile(reps = write_indices_3_reps_0, x = write_indices_float_7_cast_fp16_to_int32)[name = tensor("write_indices_3")]; + tensor var_749_begin_0 = const()[name = tensor("op_749_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_749_end_0 = const()[name = tensor("op_749_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor var_749_end_mask_0 = const()[name = tensor("op_749_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_749_squeeze_mask_0 = const()[name = tensor("op_749_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor cache1_to_fp16_dtype_0 = const()[name = tensor("cache1_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor cache1_to_fp16 = cast(dtype = cache1_to_fp16_dtype_0, x = cache1)[name = tensor("cast_132")]; + tensor var_749_cast_fp16 = slice_by_index(begin = var_749_begin_0, end = var_749_end_0, end_mask = var_749_end_mask_0, squeeze_mask = var_749_squeeze_mask_0, x = cache1_to_fp16)[name = tensor("op_749_cast_fp16")]; + tensor var_751_axis_0 = const()[name = tensor("op_751_axis_0"), val = tensor(1)]; + tensor var_751_mode_0 = const()[name = tensor("op_751_mode_0"), val = tensor("update")]; + tensor var_751_validate_indices_0 = const()[name = tensor("op_751_validate_indices_0"), val = tensor(false)]; + tensor var_751_cast_fp16 = scatter_along_axis(axis = var_751_axis_0, data = var_749_cast_fp16, indices = write_indices_3, mode = var_751_mode_0, updates = k_7_cast_fp16, validate_indices = var_751_validate_indices_0)[name = tensor("op_751_cast_fp16")]; + tensor concat_9 = const()[name = tensor("concat_9"), val = tensor([0, 0, 0, 0, 0])]; + tensor concat_10 = const()[name = tensor("concat_10"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_3_internal_tensor_assign_1_stride_0 = const()[name = tensor("new_cache_3_internal_tensor_assign_1_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_3_internal_tensor_assign_1_begin_mask_0 = const()[name = tensor("new_cache_3_internal_tensor_assign_1_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_3_internal_tensor_assign_1_end_mask_0 = const()[name = tensor("new_cache_3_internal_tensor_assign_1_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_3_internal_tensor_assign_1_squeeze_mask_0 = const()[name = tensor("new_cache_3_internal_tensor_assign_1_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_14 = const()[name = tensor("shape_14"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_2 = const()[name = tensor("reduce_prod_2"), val = tensor(1048576)]; + tensor range_1d_2_start_0 = const()[name = tensor("range_1d_2_start_0"), val = tensor(0)]; + tensor range_1d_2_step_0 = const()[name = tensor("range_1d_2_step_0"), val = tensor(1)]; + tensor range_1d_2 = range_1d(end = reduce_prod_2, start = range_1d_2_start_0, step = range_1d_2_step_0)[name = tensor("range_1d_2")]; + tensor reshape_10 = reshape(shape = shape_14, x = range_1d_2)[name = tensor("reshape_10")]; + tensor slice_by_index_2 = slice_by_index(begin = concat_9, begin_mask = new_cache_3_internal_tensor_assign_1_begin_mask_0, end = concat_10, end_mask = new_cache_3_internal_tensor_assign_1_end_mask_0, squeeze_mask = new_cache_3_internal_tensor_assign_1_squeeze_mask_0, stride = new_cache_3_internal_tensor_assign_1_stride_0, x = reshape_10)[name = tensor("slice_by_index_2")]; + tensor reshape_11_shape_0 = const()[name = tensor("reshape_11_shape_0"), val = tensor([-1])]; + tensor reshape_11 = reshape(shape = reshape_11_shape_0, x = slice_by_index_2)[name = tensor("reshape_11")]; + tensor reshape_12_shape_0 = const()[name = tensor("reshape_12_shape_0"), val = tensor([-1])]; + tensor reshape_12_cast_fp16 = reshape(shape = reshape_12_shape_0, x = var_751_cast_fp16)[name = tensor("reshape_12_cast_fp16")]; + tensor reshape_13_shape_0 = const()[name = tensor("reshape_13_shape_0"), val = tensor([-1])]; + tensor reshape_13_cast_fp16 = reshape(shape = reshape_13_shape_0, x = cache1_to_fp16)[name = tensor("reshape_13_cast_fp16")]; + tensor scatter_2_mode_0 = const()[name = tensor("scatter_2_mode_0"), val = tensor("update")]; + tensor scatter_2_axis_0 = const()[name = tensor("scatter_2_axis_0"), val = tensor(0)]; + tensor scatter_2_validate_indices_0 = const()[name = tensor("scatter_2_validate_indices_0"), val = tensor(false)]; + tensor scatter_2_cast_fp16 = scatter(axis = scatter_2_axis_0, data = reshape_13_cast_fp16, indices = reshape_11, mode = scatter_2_mode_0, updates = reshape_12_cast_fp16, validate_indices = scatter_2_validate_indices_0)[name = tensor("scatter_2_cast_fp16")]; + tensor reshape_14_cast_fp16 = reshape(shape = shape_14, x = scatter_2_cast_fp16)[name = tensor("reshape_14_cast_fp16")]; + tensor var_759_begin_0 = const()[name = tensor("op_759_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor var_759_end_0 = const()[name = tensor("op_759_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor var_759_end_mask_0 = const()[name = tensor("op_759_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_759_squeeze_mask_0 = const()[name = tensor("op_759_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor var_759_cast_fp16 = slice_by_index(begin = var_759_begin_0, end = var_759_end_0, end_mask = var_759_end_mask_0, squeeze_mask = var_759_squeeze_mask_0, x = reshape_14_cast_fp16)[name = tensor("op_759_cast_fp16")]; + tensor var_761_axis_0 = const()[name = tensor("op_761_axis_0"), val = tensor(1)]; + tensor var_761_mode_0 = const()[name = tensor("op_761_mode_0"), val = tensor("update")]; + tensor var_761_validate_indices_0 = const()[name = tensor("op_761_validate_indices_0"), val = tensor(false)]; + tensor var_761_cast_fp16 = scatter_along_axis(axis = var_761_axis_0, data = var_759_cast_fp16, indices = write_indices_3, mode = var_761_mode_0, updates = v_3_cast_fp16, validate_indices = var_761_validate_indices_0)[name = tensor("op_761_cast_fp16")]; + tensor concat_11 = const()[name = tensor("concat_11"), val = tensor([1, 0, 0, 0, 0])]; + tensor concat_12 = const()[name = tensor("concat_12"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_3_internal_tensor_assign_2_stride_0 = const()[name = tensor("new_cache_3_internal_tensor_assign_2_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_3_internal_tensor_assign_2_begin_mask_0 = const()[name = tensor("new_cache_3_internal_tensor_assign_2_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_3_internal_tensor_assign_2_end_mask_0 = const()[name = tensor("new_cache_3_internal_tensor_assign_2_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_3_internal_tensor_assign_2_squeeze_mask_0 = const()[name = tensor("new_cache_3_internal_tensor_assign_2_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_15 = const()[name = tensor("shape_15"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_3 = const()[name = tensor("reduce_prod_3"), val = tensor(1048576)]; + tensor range_1d_3_start_0 = const()[name = tensor("range_1d_3_start_0"), val = tensor(0)]; + tensor range_1d_3_step_0 = const()[name = tensor("range_1d_3_step_0"), val = tensor(1)]; + tensor range_1d_3 = range_1d(end = reduce_prod_3, start = range_1d_3_start_0, step = range_1d_3_step_0)[name = tensor("range_1d_3")]; + tensor reshape_15 = reshape(shape = shape_15, x = range_1d_3)[name = tensor("reshape_15")]; + tensor slice_by_index_3 = slice_by_index(begin = concat_11, begin_mask = new_cache_3_internal_tensor_assign_2_begin_mask_0, end = concat_12, end_mask = new_cache_3_internal_tensor_assign_2_end_mask_0, squeeze_mask = new_cache_3_internal_tensor_assign_2_squeeze_mask_0, stride = new_cache_3_internal_tensor_assign_2_stride_0, x = reshape_15)[name = tensor("slice_by_index_3")]; + tensor reshape_16_shape_0 = const()[name = tensor("reshape_16_shape_0"), val = tensor([-1])]; + tensor reshape_16 = reshape(shape = reshape_16_shape_0, x = slice_by_index_3)[name = tensor("reshape_16")]; + tensor reshape_17_shape_0 = const()[name = tensor("reshape_17_shape_0"), val = tensor([-1])]; + tensor reshape_17_cast_fp16 = reshape(shape = reshape_17_shape_0, x = var_761_cast_fp16)[name = tensor("reshape_17_cast_fp16")]; + tensor reshape_18_shape_0 = const()[name = tensor("reshape_18_shape_0"), val = tensor([-1])]; + tensor reshape_18_cast_fp16 = reshape(shape = reshape_18_shape_0, x = reshape_14_cast_fp16)[name = tensor("reshape_18_cast_fp16")]; + tensor scatter_3_mode_0 = const()[name = tensor("scatter_3_mode_0"), val = tensor("update")]; + tensor scatter_3_axis_0 = const()[name = tensor("scatter_3_axis_0"), val = tensor(0)]; + tensor scatter_3_validate_indices_0 = const()[name = tensor("scatter_3_validate_indices_0"), val = tensor(false)]; + tensor scatter_3_cast_fp16 = scatter(axis = scatter_3_axis_0, data = reshape_18_cast_fp16, indices = reshape_16, mode = scatter_3_mode_0, updates = reshape_17_cast_fp16, validate_indices = scatter_3_validate_indices_0)[name = tensor("scatter_3_cast_fp16")]; + tensor reshape_19_cast_fp16 = reshape(shape = shape_15, x = scatter_3_cast_fp16)[name = tensor("reshape_19_cast_fp16")]; + tensor reshape_19_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("reshape_19_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor keys_7_begin_0 = const()[name = tensor("keys_7_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor keys_7_end_0 = const()[name = tensor("keys_7_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor keys_7_end_mask_0 = const()[name = tensor("keys_7_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor keys_7_squeeze_mask_0 = const()[name = tensor("keys_7_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor keys_7_cast_fp16 = slice_by_index(begin = keys_7_begin_0, end = keys_7_end_0, end_mask = keys_7_end_mask_0, squeeze_mask = keys_7_squeeze_mask_0, x = reshape_19_cast_fp16)[name = tensor("keys_7_cast_fp16")]; + tensor values_7_begin_0 = const()[name = tensor("values_7_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor values_7_end_0 = const()[name = tensor("values_7_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor values_7_end_mask_0 = const()[name = tensor("values_7_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor values_7_squeeze_mask_0 = const()[name = tensor("values_7_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor values_7_cast_fp16 = slice_by_index(begin = values_7_begin_0, end = values_7_end_0, end_mask = values_7_end_mask_0, squeeze_mask = values_7_squeeze_mask_0, x = reshape_19_cast_fp16)[name = tensor("values_7_cast_fp16")]; + tensor var_773_cast_fp16 = not_equal(x = keys_7_cast_fp16, y = keys_7_cast_fp16)[name = tensor("op_773_cast_fp16")]; + tensor keys_9_cast_fp16 = select(a = var_360_to_fp16, b = keys_7_cast_fp16, cond = var_773_cast_fp16)[name = tensor("keys_9_cast_fp16")]; + tensor var_781_cast_fp16 = not_equal(x = values_7_cast_fp16, y = values_7_cast_fp16)[name = tensor("op_781_cast_fp16")]; + tensor values_9_cast_fp16 = select(a = var_360_to_fp16, b = values_7_cast_fp16, cond = var_781_cast_fp16)[name = tensor("values_9_cast_fp16")]; + tensor var_805 = const()[name = tensor("op_805"), val = tensor([0, 2, 1, 3])]; + tensor var_818 = const()[name = tensor("op_818"), val = tensor([1, 1, 1])]; + tensor var_819_cast_fp16 = reshape(shape = var_818, x = position1_to_fp16)[name = tensor("op_819_cast_fp16")]; + tensor var_836_to_fp16 = const()[name = tensor("op_836_to_fp16"), val = tensor(0x1p+0)]; + tensor valid_len_3_cast_fp16 = add(x = var_819_cast_fp16, y = var_836_to_fp16)[name = tensor("valid_len_3_cast_fp16")]; + tensor valid_mask_3_cast_fp16 = less(x = k_positions_1_promoted_to_fp16, y = valid_len_3_cast_fp16)[name = tensor("valid_mask_3_cast_fp16")]; + tensor causal_mask_3_cast_fp16 = less_equal(x = k_positions_1_promoted_to_fp16, y = var_819_cast_fp16)[name = tensor("causal_mask_3_cast_fp16")]; + tensor attn_mask_5 = logical_and(x = valid_mask_3_cast_fp16, y = causal_mask_3_cast_fp16)[name = tensor("attn_mask_5")]; + tensor attn_mask_7_axes_0 = const()[name = tensor("attn_mask_7_axes_0"), val = tensor([1])]; + tensor attn_mask_7 = expand_dims(axes = attn_mask_7_axes_0, x = attn_mask_5)[name = tensor("attn_mask_7")]; + tensor var_854_transpose_x_0 = const()[name = tensor("op_854_transpose_x_0"), val = tensor(false)]; + tensor var_854_transpose_y_0 = const()[name = tensor("op_854_transpose_y_0"), val = tensor(false)]; + tensor transpose_20_perm_0 = const()[name = tensor("transpose_20_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_21_perm_0 = const()[name = tensor("transpose_21_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_21 = transpose(perm = transpose_21_perm_0, x = keys_9_cast_fp16)[name = tensor("transpose_47")]; + tensor transpose_20 = transpose(perm = transpose_20_perm_0, x = q_9_cast_fp16)[name = tensor("transpose_48")]; + tensor var_854_cast_fp16 = matmul(transpose_x = var_854_transpose_x_0, transpose_y = var_854_transpose_y_0, x = transpose_20, y = transpose_21)[name = tensor("op_854_cast_fp16")]; + tensor var_848_to_fp16 = const()[name = tensor("op_848_to_fp16"), val = tensor([0x1p-3])]; + tensor attn_weights_7_cast_fp16 = mul(x = var_854_cast_fp16, y = var_848_to_fp16)[name = tensor("attn_weights_7_cast_fp16")]; + tensor var_856 = logical_not(x = attn_mask_7)[name = tensor("op_856")]; + tensor var_857_to_fp16 = const()[name = tensor("op_857_to_fp16"), val = tensor(-inf)]; + tensor attn_weights_9_cast_fp16 = select(a = var_857_to_fp16, b = attn_weights_7_cast_fp16, cond = var_856)[name = tensor("attn_weights_9_cast_fp16")]; + tensor var_859 = const()[name = tensor("op_859"), val = tensor(-1)]; + tensor attn_weights_11_cast_fp16 = softmax(axis = var_859, x = attn_weights_9_cast_fp16)[name = tensor("attn_weights_11_cast_fp16")]; + tensor attn_output_3_transpose_x_0 = const()[name = tensor("attn_output_3_transpose_x_0"), val = tensor(false)]; + tensor attn_output_3_transpose_y_0 = const()[name = tensor("attn_output_3_transpose_y_0"), val = tensor(false)]; + tensor values_11_cast_fp16 = transpose(perm = var_805, x = values_9_cast_fp16)[name = tensor("transpose_49")]; + tensor attn_output_3_cast_fp16 = matmul(transpose_x = attn_output_3_transpose_x_0, transpose_y = attn_output_3_transpose_y_0, x = attn_weights_11_cast_fp16, y = values_11_cast_fp16)[name = tensor("attn_output_3_cast_fp16")]; + tensor var_867 = const()[name = tensor("op_867"), val = tensor([0, 2, 1, 3])]; + tensor var_870 = const()[name = tensor("op_870"), val = tensor([1, 1, 1024])]; + tensor var_868_cast_fp16 = transpose(perm = var_867, x = attn_output_3_cast_fp16)[name = tensor("transpose_46")]; + tensor input_15_cast_fp16 = reshape(shape = var_870, x = var_868_cast_fp16)[name = tensor("input_15_cast_fp16")]; + tensor linear_6_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = attn1_out_proj_weight_cast_fp16, x = input_15_cast_fp16)[name = tensor("linear_6_cast_fp16")]; + tensor var_876_to_fp16 = const()[name = tensor("op_876_to_fp16"), val = tensor(0x1p+0)]; + tensor var_877_cast_fp16 = add(x = position1_to_fp16, y = var_876_to_fp16)[name = tensor("op_877_cast_fp16")]; + tensor var_877_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_877_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor input_17_cast_fp16 = add(x = input_13_cast_fp16, y = linear_6_cast_fp16)[name = tensor("input_17_cast_fp16")]; + tensor input_19_axes_0 = const()[name = tensor("input_19_axes_0"), val = tensor([-1])]; + tensor norm1_2_weight_to_fp16 = const()[name = tensor("norm1_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(76764416)))]; + tensor norm1_2_bias_to_fp16 = const()[name = tensor("norm1_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(76766528)))]; + tensor var_881_to_fp16 = const()[name = tensor("op_881_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_19_cast_fp16 = layer_norm(axes = input_19_axes_0, beta = norm1_2_bias_to_fp16, epsilon = var_881_to_fp16, gamma = norm1_2_weight_to_fp16, x = input_17_cast_fp16)[name = tensor("input_19_cast_fp16")]; + tensor linear_7_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = linear1_1_weight_cast_fp16, x = input_19_cast_fp16)[name = tensor("linear_7_cast_fp16")]; + tensor input_21_mode_0 = const()[name = tensor("input_21_mode_0"), val = tensor("EXACT")]; + tensor input_21_cast_fp16 = gelu(mode = input_21_mode_0, x = linear_7_cast_fp16)[name = tensor("input_21_cast_fp16")]; + tensor linear_8_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = linear1_2_weight_cast_fp16, x = input_21_cast_fp16)[name = tensor("linear_8_cast_fp16")]; + tensor input_23_cast_fp16 = add(x = input_17_cast_fp16, y = linear_8_cast_fp16)[name = tensor("input_23_cast_fp16")]; + tensor x_5_axes_0 = const()[name = tensor("x_5_axes_0"), val = tensor([-1])]; + tensor norm2_1_weight_to_fp16 = const()[name = tensor("norm2_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(76768640)))]; + tensor norm2_1_bias_to_fp16 = const()[name = tensor("norm2_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(76770752)))]; + tensor var_898_to_fp16 = const()[name = tensor("op_898_to_fp16"), val = tensor(0x1.5p-17)]; + tensor x_5_cast_fp16 = layer_norm(axes = x_5_axes_0, beta = norm2_1_bias_to_fp16, epsilon = var_898_to_fp16, gamma = norm2_1_weight_to_fp16, x = input_23_cast_fp16)[name = tensor("x_5_cast_fp16")]; + tensor linear_9_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = attn2_in_proj_weight_cast_fp16, x = x_5_cast_fp16)[name = tensor("linear_9_cast_fp16")]; + tensor var_934 = const()[name = tensor("op_934"), val = tensor([1, 1, 3, 16, 64])]; + tensor qkv_5_cast_fp16 = reshape(shape = var_934, x = linear_9_cast_fp16)[name = tensor("qkv_5_cast_fp16")]; + tensor q_13_begin_0 = const()[name = tensor("q_13_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor q_13_end_0 = const()[name = tensor("q_13_end_0"), val = tensor([1, 1, 1, 16, 64])]; + tensor q_13_end_mask_0 = const()[name = tensor("q_13_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor q_13_squeeze_mask_0 = const()[name = tensor("q_13_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor q_13_cast_fp16 = slice_by_index(begin = q_13_begin_0, end = q_13_end_0, end_mask = q_13_end_mask_0, squeeze_mask = q_13_squeeze_mask_0, x = qkv_5_cast_fp16)[name = tensor("q_13_cast_fp16")]; + tensor k_9_begin_0 = const()[name = tensor("k_9_begin_0"), val = tensor([0, 0, 1, 0, 0])]; + tensor k_9_end_0 = const()[name = tensor("k_9_end_0"), val = tensor([1, 1, 2, 16, 64])]; + tensor k_9_end_mask_0 = const()[name = tensor("k_9_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor k_9_squeeze_mask_0 = const()[name = tensor("k_9_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor k_9_cast_fp16 = slice_by_index(begin = k_9_begin_0, end = k_9_end_0, end_mask = k_9_end_mask_0, squeeze_mask = k_9_squeeze_mask_0, x = qkv_5_cast_fp16)[name = tensor("k_9_cast_fp16")]; + tensor v_5_begin_0 = const()[name = tensor("v_5_begin_0"), val = tensor([0, 0, 2, 0, 0])]; + tensor v_5_end_0 = const()[name = tensor("v_5_end_0"), val = tensor([1, 1, 3, 16, 64])]; + tensor v_5_end_mask_0 = const()[name = tensor("v_5_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor v_5_squeeze_mask_0 = const()[name = tensor("v_5_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor v_5_cast_fp16 = slice_by_index(begin = v_5_begin_0, end = v_5_end_0, end_mask = v_5_end_mask_0, squeeze_mask = v_5_squeeze_mask_0, x = qkv_5_cast_fp16)[name = tensor("v_5_cast_fp16")]; + tensor var_1038 = const()[name = tensor("op_1038"), val = tensor([1, 1, 1, 1])]; + tensor position2_to_fp16_dtype_0 = const()[name = tensor("position2_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor position2_to_fp16 = cast(dtype = position2_to_fp16_dtype_0, x = position2)[name = tensor("cast_129")]; + tensor ts_17_cast_fp16 = reshape(shape = var_1038, x = position2_to_fp16)[name = tensor("ts_17_cast_fp16")]; + tensor var_1042 = const()[name = tensor("op_1042"), val = tensor([1, 1, 16, 32, 2])]; + tensor q_complex_5_cast_fp16 = reshape(shape = var_1042, x = q_13_cast_fp16)[name = tensor("q_complex_5_cast_fp16")]; + tensor var_1046 = const()[name = tensor("op_1046"), val = tensor([1, 1, 16, 32, 2])]; + tensor k_complex_5_cast_fp16 = reshape(shape = var_1046, x = k_9_cast_fp16)[name = tensor("k_complex_5_cast_fp16")]; + tensor var_1050_begin_0 = const()[name = tensor("op_1050_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_1050_end_0 = const()[name = tensor("op_1050_end_0"), val = tensor([1, 1, 16, 32, 1])]; + tensor var_1050_end_mask_0 = const()[name = tensor("op_1050_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_1050_squeeze_mask_0 = const()[name = tensor("op_1050_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_1050_cast_fp16 = slice_by_index(begin = var_1050_begin_0, end = var_1050_end_0, end_mask = var_1050_end_mask_0, squeeze_mask = var_1050_squeeze_mask_0, x = q_complex_5_cast_fp16)[name = tensor("op_1050_cast_fp16")]; + tensor var_1058_begin_0 = const()[name = tensor("op_1058_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_1058_end_0 = const()[name = tensor("op_1058_end_0"), val = tensor([1, 1, 16, 32, 2])]; + tensor var_1058_end_mask_0 = const()[name = tensor("op_1058_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_1058_squeeze_mask_0 = const()[name = tensor("op_1058_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_1058_cast_fp16 = slice_by_index(begin = var_1058_begin_0, end = var_1058_end_0, end_mask = var_1058_end_mask_0, squeeze_mask = var_1058_squeeze_mask_0, x = q_complex_5_cast_fp16)[name = tensor("op_1058_cast_fp16")]; + tensor var_1066_begin_0 = const()[name = tensor("op_1066_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_1066_end_0 = const()[name = tensor("op_1066_end_0"), val = tensor([1, 1, 16, 32, 1])]; + tensor var_1066_end_mask_0 = const()[name = tensor("op_1066_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_1066_squeeze_mask_0 = const()[name = tensor("op_1066_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_1066_cast_fp16 = slice_by_index(begin = var_1066_begin_0, end = var_1066_end_0, end_mask = var_1066_end_mask_0, squeeze_mask = var_1066_squeeze_mask_0, x = k_complex_5_cast_fp16)[name = tensor("op_1066_cast_fp16")]; + tensor var_1074_begin_0 = const()[name = tensor("op_1074_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_1074_end_0 = const()[name = tensor("op_1074_end_0"), val = tensor([1, 1, 16, 32, 2])]; + tensor var_1074_end_mask_0 = const()[name = tensor("op_1074_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_1074_squeeze_mask_0 = const()[name = tensor("op_1074_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_1074_cast_fp16 = slice_by_index(begin = var_1074_begin_0, end = var_1074_end_0, end_mask = var_1074_end_mask_0, squeeze_mask = var_1074_squeeze_mask_0, x = k_complex_5_cast_fp16)[name = tensor("op_1074_cast_fp16")]; + tensor freqs_5_to_fp16 = const()[name = tensor("freqs_5_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(76772864)))]; + tensor var_1080_cast_fp16 = mul(x = freqs_5_to_fp16, y = ts_17_cast_fp16)[name = tensor("op_1080_cast_fp16")]; + tensor rotr_5_cast_fp16 = cos(x = var_1080_cast_fp16)[name = tensor("rotr_5_cast_fp16")]; + tensor roti_5_cast_fp16 = sin(x = var_1080_cast_fp16)[name = tensor("roti_5_cast_fp16")]; + tensor var_1084_cast_fp16 = mul(x = var_1050_cast_fp16, y = rotr_5_cast_fp16)[name = tensor("op_1084_cast_fp16")]; + tensor var_1085_cast_fp16 = mul(x = var_1058_cast_fp16, y = roti_5_cast_fp16)[name = tensor("op_1085_cast_fp16")]; + tensor qor_9_cast_fp16 = sub(x = var_1084_cast_fp16, y = var_1085_cast_fp16)[name = tensor("qor_9_cast_fp16")]; + tensor var_1088_cast_fp16 = mul(x = var_1050_cast_fp16, y = roti_5_cast_fp16)[name = tensor("op_1088_cast_fp16")]; + tensor var_1089_cast_fp16 = mul(x = var_1058_cast_fp16, y = rotr_5_cast_fp16)[name = tensor("op_1089_cast_fp16")]; + tensor qoi_9_cast_fp16 = add(x = var_1088_cast_fp16, y = var_1089_cast_fp16)[name = tensor("qoi_9_cast_fp16")]; + tensor var_1092_cast_fp16 = mul(x = var_1066_cast_fp16, y = rotr_5_cast_fp16)[name = tensor("op_1092_cast_fp16")]; + tensor var_1093_cast_fp16 = mul(x = var_1074_cast_fp16, y = roti_5_cast_fp16)[name = tensor("op_1093_cast_fp16")]; + tensor kor_9_cast_fp16 = sub(x = var_1092_cast_fp16, y = var_1093_cast_fp16)[name = tensor("kor_9_cast_fp16")]; + tensor var_1096_cast_fp16 = mul(x = var_1066_cast_fp16, y = roti_5_cast_fp16)[name = tensor("op_1096_cast_fp16")]; + tensor var_1097_cast_fp16 = mul(x = var_1074_cast_fp16, y = rotr_5_cast_fp16)[name = tensor("op_1097_cast_fp16")]; + tensor koi_9_cast_fp16 = add(x = var_1096_cast_fp16, y = var_1097_cast_fp16)[name = tensor("koi_9_cast_fp16")]; + tensor qo_5_axis_0 = const()[name = tensor("qo_5_axis_0"), val = tensor(-1)]; + tensor qo_5_cast_fp16 = stack(axis = qo_5_axis_0, values = (qor_9_cast_fp16, qoi_9_cast_fp16))[name = tensor("qo_5_cast_fp16")]; + tensor ko_5_axis_0 = const()[name = tensor("ko_5_axis_0"), val = tensor(-1)]; + tensor ko_5_cast_fp16 = stack(axis = ko_5_axis_0, values = (kor_9_cast_fp16, koi_9_cast_fp16))[name = tensor("ko_5_cast_fp16")]; + tensor var_1126 = const()[name = tensor("op_1126"), val = tensor([1, 1, 16, 64])]; + tensor q_15_cast_fp16 = reshape(shape = var_1126, x = qo_5_cast_fp16)[name = tensor("q_15_cast_fp16")]; + tensor var_1128 = const()[name = tensor("op_1128"), val = tensor([1, 1, 16, 64])]; + tensor k_11_cast_fp16 = reshape(shape = var_1128, x = ko_5_cast_fp16)[name = tensor("k_11_cast_fp16")]; + tensor _inversed_1150_y_0_to_fp16 = const()[name = tensor("_inversed_1150_y_0_to_fp16"), val = tensor(0x1p-9)]; + tensor _inversed_1150_cast_fp16 = mul(x = ts_17_cast_fp16, y = _inversed_1150_y_0_to_fp16)[name = tensor("_inversed_1150_cast_fp16")]; + tensor var_1151_cast_fp16 = floor(x = _inversed_1150_cast_fp16)[name = tensor("op_1151_cast_fp16")]; + tensor var_1152_to_fp16 = const()[name = tensor("op_1152_to_fp16"), val = tensor(0x1p+9)]; + tensor var_1153_cast_fp16 = mul(x = var_1151_cast_fp16, y = var_1152_to_fp16)[name = tensor("op_1153_cast_fp16")]; + tensor write_indices_float_11_cast_fp16 = sub(x = ts_17_cast_fp16, y = var_1153_cast_fp16)[name = tensor("write_indices_float_11_cast_fp16")]; + tensor var_1160_dtype_0 = const()[name = tensor("op_1160_dtype_0"), val = tensor("int32")]; + tensor write_indices_5_reps_0 = const()[name = tensor("write_indices_5_reps_0"), val = tensor([1, 1, 16, 64])]; + tensor write_indices_float_11_cast_fp16_to_int32 = cast(dtype = var_1160_dtype_0, x = write_indices_float_11_cast_fp16)[name = tensor("cast_128")]; + tensor write_indices_5 = tile(reps = write_indices_5_reps_0, x = write_indices_float_11_cast_fp16_to_int32)[name = tensor("write_indices_5")]; + tensor var_1168_begin_0 = const()[name = tensor("op_1168_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_1168_end_0 = const()[name = tensor("op_1168_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor var_1168_end_mask_0 = const()[name = tensor("op_1168_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_1168_squeeze_mask_0 = const()[name = tensor("op_1168_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor cache2_to_fp16_dtype_0 = const()[name = tensor("cache2_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor cache2_to_fp16 = cast(dtype = cache2_to_fp16_dtype_0, x = cache2)[name = tensor("cast_127")]; + tensor var_1168_cast_fp16 = slice_by_index(begin = var_1168_begin_0, end = var_1168_end_0, end_mask = var_1168_end_mask_0, squeeze_mask = var_1168_squeeze_mask_0, x = cache2_to_fp16)[name = tensor("op_1168_cast_fp16")]; + tensor var_1170_axis_0 = const()[name = tensor("op_1170_axis_0"), val = tensor(1)]; + tensor var_1170_mode_0 = const()[name = tensor("op_1170_mode_0"), val = tensor("update")]; + tensor var_1170_validate_indices_0 = const()[name = tensor("op_1170_validate_indices_0"), val = tensor(false)]; + tensor var_1170_cast_fp16 = scatter_along_axis(axis = var_1170_axis_0, data = var_1168_cast_fp16, indices = write_indices_5, mode = var_1170_mode_0, updates = k_11_cast_fp16, validate_indices = var_1170_validate_indices_0)[name = tensor("op_1170_cast_fp16")]; + tensor concat_16 = const()[name = tensor("concat_16"), val = tensor([0, 0, 0, 0, 0])]; + tensor concat_17 = const()[name = tensor("concat_17"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_5_internal_tensor_assign_1_stride_0 = const()[name = tensor("new_cache_5_internal_tensor_assign_1_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_5_internal_tensor_assign_1_begin_mask_0 = const()[name = tensor("new_cache_5_internal_tensor_assign_1_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_5_internal_tensor_assign_1_end_mask_0 = const()[name = tensor("new_cache_5_internal_tensor_assign_1_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_5_internal_tensor_assign_1_squeeze_mask_0 = const()[name = tensor("new_cache_5_internal_tensor_assign_1_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_16 = const()[name = tensor("shape_16"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_4 = const()[name = tensor("reduce_prod_4"), val = tensor(1048576)]; + tensor range_1d_4_start_0 = const()[name = tensor("range_1d_4_start_0"), val = tensor(0)]; + tensor range_1d_4_step_0 = const()[name = tensor("range_1d_4_step_0"), val = tensor(1)]; + tensor range_1d_4 = range_1d(end = reduce_prod_4, start = range_1d_4_start_0, step = range_1d_4_step_0)[name = tensor("range_1d_4")]; + tensor reshape_20 = reshape(shape = shape_16, x = range_1d_4)[name = tensor("reshape_20")]; + tensor slice_by_index_4 = slice_by_index(begin = concat_16, begin_mask = new_cache_5_internal_tensor_assign_1_begin_mask_0, end = concat_17, end_mask = new_cache_5_internal_tensor_assign_1_end_mask_0, squeeze_mask = new_cache_5_internal_tensor_assign_1_squeeze_mask_0, stride = new_cache_5_internal_tensor_assign_1_stride_0, x = reshape_20)[name = tensor("slice_by_index_4")]; + tensor reshape_21_shape_0 = const()[name = tensor("reshape_21_shape_0"), val = tensor([-1])]; + tensor reshape_21 = reshape(shape = reshape_21_shape_0, x = slice_by_index_4)[name = tensor("reshape_21")]; + tensor reshape_22_shape_0 = const()[name = tensor("reshape_22_shape_0"), val = tensor([-1])]; + tensor reshape_22_cast_fp16 = reshape(shape = reshape_22_shape_0, x = var_1170_cast_fp16)[name = tensor("reshape_22_cast_fp16")]; + tensor reshape_23_shape_0 = const()[name = tensor("reshape_23_shape_0"), val = tensor([-1])]; + tensor reshape_23_cast_fp16 = reshape(shape = reshape_23_shape_0, x = cache2_to_fp16)[name = tensor("reshape_23_cast_fp16")]; + tensor scatter_4_mode_0 = const()[name = tensor("scatter_4_mode_0"), val = tensor("update")]; + tensor scatter_4_axis_0 = const()[name = tensor("scatter_4_axis_0"), val = tensor(0)]; + tensor scatter_4_validate_indices_0 = const()[name = tensor("scatter_4_validate_indices_0"), val = tensor(false)]; + tensor scatter_4_cast_fp16 = scatter(axis = scatter_4_axis_0, data = reshape_23_cast_fp16, indices = reshape_21, mode = scatter_4_mode_0, updates = reshape_22_cast_fp16, validate_indices = scatter_4_validate_indices_0)[name = tensor("scatter_4_cast_fp16")]; + tensor reshape_24_cast_fp16 = reshape(shape = shape_16, x = scatter_4_cast_fp16)[name = tensor("reshape_24_cast_fp16")]; + tensor var_1178_begin_0 = const()[name = tensor("op_1178_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor var_1178_end_0 = const()[name = tensor("op_1178_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor var_1178_end_mask_0 = const()[name = tensor("op_1178_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_1178_squeeze_mask_0 = const()[name = tensor("op_1178_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor var_1178_cast_fp16 = slice_by_index(begin = var_1178_begin_0, end = var_1178_end_0, end_mask = var_1178_end_mask_0, squeeze_mask = var_1178_squeeze_mask_0, x = reshape_24_cast_fp16)[name = tensor("op_1178_cast_fp16")]; + tensor var_1180_axis_0 = const()[name = tensor("op_1180_axis_0"), val = tensor(1)]; + tensor var_1180_mode_0 = const()[name = tensor("op_1180_mode_0"), val = tensor("update")]; + tensor var_1180_validate_indices_0 = const()[name = tensor("op_1180_validate_indices_0"), val = tensor(false)]; + tensor var_1180_cast_fp16 = scatter_along_axis(axis = var_1180_axis_0, data = var_1178_cast_fp16, indices = write_indices_5, mode = var_1180_mode_0, updates = v_5_cast_fp16, validate_indices = var_1180_validate_indices_0)[name = tensor("op_1180_cast_fp16")]; + tensor concat_18 = const()[name = tensor("concat_18"), val = tensor([1, 0, 0, 0, 0])]; + tensor concat_19 = const()[name = tensor("concat_19"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_5_internal_tensor_assign_2_stride_0 = const()[name = tensor("new_cache_5_internal_tensor_assign_2_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_5_internal_tensor_assign_2_begin_mask_0 = const()[name = tensor("new_cache_5_internal_tensor_assign_2_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_5_internal_tensor_assign_2_end_mask_0 = const()[name = tensor("new_cache_5_internal_tensor_assign_2_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_5_internal_tensor_assign_2_squeeze_mask_0 = const()[name = tensor("new_cache_5_internal_tensor_assign_2_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_17 = const()[name = tensor("shape_17"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_5 = const()[name = tensor("reduce_prod_5"), val = tensor(1048576)]; + tensor range_1d_5_start_0 = const()[name = tensor("range_1d_5_start_0"), val = tensor(0)]; + tensor range_1d_5_step_0 = const()[name = tensor("range_1d_5_step_0"), val = tensor(1)]; + tensor range_1d_5 = range_1d(end = reduce_prod_5, start = range_1d_5_start_0, step = range_1d_5_step_0)[name = tensor("range_1d_5")]; + tensor reshape_25 = reshape(shape = shape_17, x = range_1d_5)[name = tensor("reshape_25")]; + tensor slice_by_index_5 = slice_by_index(begin = concat_18, begin_mask = new_cache_5_internal_tensor_assign_2_begin_mask_0, end = concat_19, end_mask = new_cache_5_internal_tensor_assign_2_end_mask_0, squeeze_mask = new_cache_5_internal_tensor_assign_2_squeeze_mask_0, stride = new_cache_5_internal_tensor_assign_2_stride_0, x = reshape_25)[name = tensor("slice_by_index_5")]; + tensor reshape_26_shape_0 = const()[name = tensor("reshape_26_shape_0"), val = tensor([-1])]; + tensor reshape_26 = reshape(shape = reshape_26_shape_0, x = slice_by_index_5)[name = tensor("reshape_26")]; + tensor reshape_27_shape_0 = const()[name = tensor("reshape_27_shape_0"), val = tensor([-1])]; + tensor reshape_27_cast_fp16 = reshape(shape = reshape_27_shape_0, x = var_1180_cast_fp16)[name = tensor("reshape_27_cast_fp16")]; + tensor reshape_28_shape_0 = const()[name = tensor("reshape_28_shape_0"), val = tensor([-1])]; + tensor reshape_28_cast_fp16 = reshape(shape = reshape_28_shape_0, x = reshape_24_cast_fp16)[name = tensor("reshape_28_cast_fp16")]; + tensor scatter_5_mode_0 = const()[name = tensor("scatter_5_mode_0"), val = tensor("update")]; + tensor scatter_5_axis_0 = const()[name = tensor("scatter_5_axis_0"), val = tensor(0)]; + tensor scatter_5_validate_indices_0 = const()[name = tensor("scatter_5_validate_indices_0"), val = tensor(false)]; + tensor scatter_5_cast_fp16 = scatter(axis = scatter_5_axis_0, data = reshape_28_cast_fp16, indices = reshape_26, mode = scatter_5_mode_0, updates = reshape_27_cast_fp16, validate_indices = scatter_5_validate_indices_0)[name = tensor("scatter_5_cast_fp16")]; + tensor reshape_29_cast_fp16 = reshape(shape = shape_17, x = scatter_5_cast_fp16)[name = tensor("reshape_29_cast_fp16")]; + tensor reshape_29_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("reshape_29_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor keys_13_begin_0 = const()[name = tensor("keys_13_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor keys_13_end_0 = const()[name = tensor("keys_13_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor keys_13_end_mask_0 = const()[name = tensor("keys_13_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor keys_13_squeeze_mask_0 = const()[name = tensor("keys_13_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor keys_13_cast_fp16 = slice_by_index(begin = keys_13_begin_0, end = keys_13_end_0, end_mask = keys_13_end_mask_0, squeeze_mask = keys_13_squeeze_mask_0, x = reshape_29_cast_fp16)[name = tensor("keys_13_cast_fp16")]; + tensor values_13_begin_0 = const()[name = tensor("values_13_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor values_13_end_0 = const()[name = tensor("values_13_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor values_13_end_mask_0 = const()[name = tensor("values_13_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor values_13_squeeze_mask_0 = const()[name = tensor("values_13_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor values_13_cast_fp16 = slice_by_index(begin = values_13_begin_0, end = values_13_end_0, end_mask = values_13_end_mask_0, squeeze_mask = values_13_squeeze_mask_0, x = reshape_29_cast_fp16)[name = tensor("values_13_cast_fp16")]; + tensor var_1192_cast_fp16 = not_equal(x = keys_13_cast_fp16, y = keys_13_cast_fp16)[name = tensor("op_1192_cast_fp16")]; + tensor keys_15_cast_fp16 = select(a = var_360_to_fp16, b = keys_13_cast_fp16, cond = var_1192_cast_fp16)[name = tensor("keys_15_cast_fp16")]; + tensor var_1200_cast_fp16 = not_equal(x = values_13_cast_fp16, y = values_13_cast_fp16)[name = tensor("op_1200_cast_fp16")]; + tensor values_15_cast_fp16 = select(a = var_360_to_fp16, b = values_13_cast_fp16, cond = var_1200_cast_fp16)[name = tensor("values_15_cast_fp16")]; + tensor var_1224 = const()[name = tensor("op_1224"), val = tensor([0, 2, 1, 3])]; + tensor var_1237 = const()[name = tensor("op_1237"), val = tensor([1, 1, 1])]; + tensor var_1238_cast_fp16 = reshape(shape = var_1237, x = position2_to_fp16)[name = tensor("op_1238_cast_fp16")]; + tensor var_1255_to_fp16 = const()[name = tensor("op_1255_to_fp16"), val = tensor(0x1p+0)]; + tensor valid_len_5_cast_fp16 = add(x = var_1238_cast_fp16, y = var_1255_to_fp16)[name = tensor("valid_len_5_cast_fp16")]; + tensor valid_mask_5_cast_fp16 = less(x = k_positions_1_promoted_to_fp16, y = valid_len_5_cast_fp16)[name = tensor("valid_mask_5_cast_fp16")]; + tensor causal_mask_5_cast_fp16 = less_equal(x = k_positions_1_promoted_to_fp16, y = var_1238_cast_fp16)[name = tensor("causal_mask_5_cast_fp16")]; + tensor attn_mask_9 = logical_and(x = valid_mask_5_cast_fp16, y = causal_mask_5_cast_fp16)[name = tensor("attn_mask_9")]; + tensor attn_mask_11_axes_0 = const()[name = tensor("attn_mask_11_axes_0"), val = tensor([1])]; + tensor attn_mask_11 = expand_dims(axes = attn_mask_11_axes_0, x = attn_mask_9)[name = tensor("attn_mask_11")]; + tensor var_1273_transpose_x_0 = const()[name = tensor("op_1273_transpose_x_0"), val = tensor(false)]; + tensor var_1273_transpose_y_0 = const()[name = tensor("op_1273_transpose_y_0"), val = tensor(false)]; + tensor transpose_22_perm_0 = const()[name = tensor("transpose_22_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_23_perm_0 = const()[name = tensor("transpose_23_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_23 = transpose(perm = transpose_23_perm_0, x = keys_15_cast_fp16)[name = tensor("transpose_43")]; + tensor transpose_22 = transpose(perm = transpose_22_perm_0, x = q_15_cast_fp16)[name = tensor("transpose_44")]; + tensor var_1273_cast_fp16 = matmul(transpose_x = var_1273_transpose_x_0, transpose_y = var_1273_transpose_y_0, x = transpose_22, y = transpose_23)[name = tensor("op_1273_cast_fp16")]; + tensor var_1267_to_fp16 = const()[name = tensor("op_1267_to_fp16"), val = tensor([0x1p-3])]; + tensor attn_weights_13_cast_fp16 = mul(x = var_1273_cast_fp16, y = var_1267_to_fp16)[name = tensor("attn_weights_13_cast_fp16")]; + tensor var_1275 = logical_not(x = attn_mask_11)[name = tensor("op_1275")]; + tensor var_1276_to_fp16 = const()[name = tensor("op_1276_to_fp16"), val = tensor(-inf)]; + tensor attn_weights_15_cast_fp16 = select(a = var_1276_to_fp16, b = attn_weights_13_cast_fp16, cond = var_1275)[name = tensor("attn_weights_15_cast_fp16")]; + tensor var_1278 = const()[name = tensor("op_1278"), val = tensor(-1)]; + tensor attn_weights_17_cast_fp16 = softmax(axis = var_1278, x = attn_weights_15_cast_fp16)[name = tensor("attn_weights_17_cast_fp16")]; + tensor attn_output_5_transpose_x_0 = const()[name = tensor("attn_output_5_transpose_x_0"), val = tensor(false)]; + tensor attn_output_5_transpose_y_0 = const()[name = tensor("attn_output_5_transpose_y_0"), val = tensor(false)]; + tensor values_17_cast_fp16 = transpose(perm = var_1224, x = values_15_cast_fp16)[name = tensor("transpose_45")]; + tensor attn_output_5_cast_fp16 = matmul(transpose_x = attn_output_5_transpose_x_0, transpose_y = attn_output_5_transpose_y_0, x = attn_weights_17_cast_fp16, y = values_17_cast_fp16)[name = tensor("attn_output_5_cast_fp16")]; + tensor var_1286 = const()[name = tensor("op_1286"), val = tensor([0, 2, 1, 3])]; + tensor var_1289 = const()[name = tensor("op_1289"), val = tensor([1, 1, 1024])]; + tensor var_1287_cast_fp16 = transpose(perm = var_1286, x = attn_output_5_cast_fp16)[name = tensor("transpose_42")]; + tensor input_25_cast_fp16 = reshape(shape = var_1289, x = var_1287_cast_fp16)[name = tensor("input_25_cast_fp16")]; + tensor linear_10_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = attn2_out_proj_weight_cast_fp16, x = input_25_cast_fp16)[name = tensor("linear_10_cast_fp16")]; + tensor var_1295_to_fp16 = const()[name = tensor("op_1295_to_fp16"), val = tensor(0x1p+0)]; + tensor var_1296_cast_fp16 = add(x = position2_to_fp16, y = var_1295_to_fp16)[name = tensor("op_1296_cast_fp16")]; + tensor var_1296_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_1296_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor input_27_cast_fp16 = add(x = input_23_cast_fp16, y = linear_10_cast_fp16)[name = tensor("input_27_cast_fp16")]; + tensor input_29_axes_0 = const()[name = tensor("input_29_axes_0"), val = tensor([-1])]; + tensor norm2_2_weight_to_fp16 = const()[name = tensor("norm2_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(76772992)))]; + tensor norm2_2_bias_to_fp16 = const()[name = tensor("norm2_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(76775104)))]; + tensor var_1300_to_fp16 = const()[name = tensor("op_1300_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_29_cast_fp16 = layer_norm(axes = input_29_axes_0, beta = norm2_2_bias_to_fp16, epsilon = var_1300_to_fp16, gamma = norm2_2_weight_to_fp16, x = input_27_cast_fp16)[name = tensor("input_29_cast_fp16")]; + tensor linear_11_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = linear2_1_weight_cast_fp16, x = input_29_cast_fp16)[name = tensor("linear_11_cast_fp16")]; + tensor input_31_mode_0 = const()[name = tensor("input_31_mode_0"), val = tensor("EXACT")]; + tensor input_31_cast_fp16 = gelu(mode = input_31_mode_0, x = linear_11_cast_fp16)[name = tensor("input_31_cast_fp16")]; + tensor linear_12_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = linear2_2_weight_cast_fp16, x = input_31_cast_fp16)[name = tensor("linear_12_cast_fp16")]; + tensor input_33_cast_fp16 = add(x = input_27_cast_fp16, y = linear_12_cast_fp16)[name = tensor("input_33_cast_fp16")]; + tensor x_7_axes_0 = const()[name = tensor("x_7_axes_0"), val = tensor([-1])]; + tensor norm3_1_weight_to_fp16 = const()[name = tensor("norm3_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(76777216)))]; + tensor norm3_1_bias_to_fp16 = const()[name = tensor("norm3_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(76779328)))]; + tensor var_1317_to_fp16 = const()[name = tensor("op_1317_to_fp16"), val = tensor(0x1.5p-17)]; + tensor x_7_cast_fp16 = layer_norm(axes = x_7_axes_0, beta = norm3_1_bias_to_fp16, epsilon = var_1317_to_fp16, gamma = norm3_1_weight_to_fp16, x = input_33_cast_fp16)[name = tensor("x_7_cast_fp16")]; + tensor linear_13_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = attn3_in_proj_weight_cast_fp16, x = x_7_cast_fp16)[name = tensor("linear_13_cast_fp16")]; + tensor var_1353 = const()[name = tensor("op_1353"), val = tensor([1, 1, 3, 16, 64])]; + tensor qkv_7_cast_fp16 = reshape(shape = var_1353, x = linear_13_cast_fp16)[name = tensor("qkv_7_cast_fp16")]; + tensor q_19_begin_0 = const()[name = tensor("q_19_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor q_19_end_0 = const()[name = tensor("q_19_end_0"), val = tensor([1, 1, 1, 16, 64])]; + tensor q_19_end_mask_0 = const()[name = tensor("q_19_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor q_19_squeeze_mask_0 = const()[name = tensor("q_19_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor q_19_cast_fp16 = slice_by_index(begin = q_19_begin_0, end = q_19_end_0, end_mask = q_19_end_mask_0, squeeze_mask = q_19_squeeze_mask_0, x = qkv_7_cast_fp16)[name = tensor("q_19_cast_fp16")]; + tensor k_13_begin_0 = const()[name = tensor("k_13_begin_0"), val = tensor([0, 0, 1, 0, 0])]; + tensor k_13_end_0 = const()[name = tensor("k_13_end_0"), val = tensor([1, 1, 2, 16, 64])]; + tensor k_13_end_mask_0 = const()[name = tensor("k_13_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor k_13_squeeze_mask_0 = const()[name = tensor("k_13_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor k_13_cast_fp16 = slice_by_index(begin = k_13_begin_0, end = k_13_end_0, end_mask = k_13_end_mask_0, squeeze_mask = k_13_squeeze_mask_0, x = qkv_7_cast_fp16)[name = tensor("k_13_cast_fp16")]; + tensor v_7_begin_0 = const()[name = tensor("v_7_begin_0"), val = tensor([0, 0, 2, 0, 0])]; + tensor v_7_end_0 = const()[name = tensor("v_7_end_0"), val = tensor([1, 1, 3, 16, 64])]; + tensor v_7_end_mask_0 = const()[name = tensor("v_7_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor v_7_squeeze_mask_0 = const()[name = tensor("v_7_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor v_7_cast_fp16 = slice_by_index(begin = v_7_begin_0, end = v_7_end_0, end_mask = v_7_end_mask_0, squeeze_mask = v_7_squeeze_mask_0, x = qkv_7_cast_fp16)[name = tensor("v_7_cast_fp16")]; + tensor var_1457 = const()[name = tensor("op_1457"), val = tensor([1, 1, 1, 1])]; + tensor position3_to_fp16_dtype_0 = const()[name = tensor("position3_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor position3_to_fp16 = cast(dtype = position3_to_fp16_dtype_0, x = position3)[name = tensor("cast_124")]; + tensor ts_23_cast_fp16 = reshape(shape = var_1457, x = position3_to_fp16)[name = tensor("ts_23_cast_fp16")]; + tensor var_1461 = const()[name = tensor("op_1461"), val = tensor([1, 1, 16, 32, 2])]; + tensor q_complex_7_cast_fp16 = reshape(shape = var_1461, x = q_19_cast_fp16)[name = tensor("q_complex_7_cast_fp16")]; + tensor var_1465 = const()[name = tensor("op_1465"), val = tensor([1, 1, 16, 32, 2])]; + tensor k_complex_7_cast_fp16 = reshape(shape = var_1465, x = k_13_cast_fp16)[name = tensor("k_complex_7_cast_fp16")]; + tensor var_1469_begin_0 = const()[name = tensor("op_1469_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_1469_end_0 = const()[name = tensor("op_1469_end_0"), val = tensor([1, 1, 16, 32, 1])]; + tensor var_1469_end_mask_0 = const()[name = tensor("op_1469_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_1469_squeeze_mask_0 = const()[name = tensor("op_1469_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_1469_cast_fp16 = slice_by_index(begin = var_1469_begin_0, end = var_1469_end_0, end_mask = var_1469_end_mask_0, squeeze_mask = var_1469_squeeze_mask_0, x = q_complex_7_cast_fp16)[name = tensor("op_1469_cast_fp16")]; + tensor var_1477_begin_0 = const()[name = tensor("op_1477_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_1477_end_0 = const()[name = tensor("op_1477_end_0"), val = tensor([1, 1, 16, 32, 2])]; + tensor var_1477_end_mask_0 = const()[name = tensor("op_1477_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_1477_squeeze_mask_0 = const()[name = tensor("op_1477_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_1477_cast_fp16 = slice_by_index(begin = var_1477_begin_0, end = var_1477_end_0, end_mask = var_1477_end_mask_0, squeeze_mask = var_1477_squeeze_mask_0, x = q_complex_7_cast_fp16)[name = tensor("op_1477_cast_fp16")]; + tensor var_1485_begin_0 = const()[name = tensor("op_1485_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_1485_end_0 = const()[name = tensor("op_1485_end_0"), val = tensor([1, 1, 16, 32, 1])]; + tensor var_1485_end_mask_0 = const()[name = tensor("op_1485_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_1485_squeeze_mask_0 = const()[name = tensor("op_1485_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_1485_cast_fp16 = slice_by_index(begin = var_1485_begin_0, end = var_1485_end_0, end_mask = var_1485_end_mask_0, squeeze_mask = var_1485_squeeze_mask_0, x = k_complex_7_cast_fp16)[name = tensor("op_1485_cast_fp16")]; + tensor var_1493_begin_0 = const()[name = tensor("op_1493_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_1493_end_0 = const()[name = tensor("op_1493_end_0"), val = tensor([1, 1, 16, 32, 2])]; + tensor var_1493_end_mask_0 = const()[name = tensor("op_1493_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_1493_squeeze_mask_0 = const()[name = tensor("op_1493_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_1493_cast_fp16 = slice_by_index(begin = var_1493_begin_0, end = var_1493_end_0, end_mask = var_1493_end_mask_0, squeeze_mask = var_1493_squeeze_mask_0, x = k_complex_7_cast_fp16)[name = tensor("op_1493_cast_fp16")]; + tensor freqs_7_to_fp16 = const()[name = tensor("freqs_7_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(76781440)))]; + tensor var_1499_cast_fp16 = mul(x = freqs_7_to_fp16, y = ts_23_cast_fp16)[name = tensor("op_1499_cast_fp16")]; + tensor rotr_7_cast_fp16 = cos(x = var_1499_cast_fp16)[name = tensor("rotr_7_cast_fp16")]; + tensor roti_7_cast_fp16 = sin(x = var_1499_cast_fp16)[name = tensor("roti_7_cast_fp16")]; + tensor var_1503_cast_fp16 = mul(x = var_1469_cast_fp16, y = rotr_7_cast_fp16)[name = tensor("op_1503_cast_fp16")]; + tensor var_1504_cast_fp16 = mul(x = var_1477_cast_fp16, y = roti_7_cast_fp16)[name = tensor("op_1504_cast_fp16")]; + tensor qor_13_cast_fp16 = sub(x = var_1503_cast_fp16, y = var_1504_cast_fp16)[name = tensor("qor_13_cast_fp16")]; + tensor var_1507_cast_fp16 = mul(x = var_1469_cast_fp16, y = roti_7_cast_fp16)[name = tensor("op_1507_cast_fp16")]; + tensor var_1508_cast_fp16 = mul(x = var_1477_cast_fp16, y = rotr_7_cast_fp16)[name = tensor("op_1508_cast_fp16")]; + tensor qoi_13_cast_fp16 = add(x = var_1507_cast_fp16, y = var_1508_cast_fp16)[name = tensor("qoi_13_cast_fp16")]; + tensor var_1511_cast_fp16 = mul(x = var_1485_cast_fp16, y = rotr_7_cast_fp16)[name = tensor("op_1511_cast_fp16")]; + tensor var_1512_cast_fp16 = mul(x = var_1493_cast_fp16, y = roti_7_cast_fp16)[name = tensor("op_1512_cast_fp16")]; + tensor kor_13_cast_fp16 = sub(x = var_1511_cast_fp16, y = var_1512_cast_fp16)[name = tensor("kor_13_cast_fp16")]; + tensor var_1515_cast_fp16 = mul(x = var_1485_cast_fp16, y = roti_7_cast_fp16)[name = tensor("op_1515_cast_fp16")]; + tensor var_1516_cast_fp16 = mul(x = var_1493_cast_fp16, y = rotr_7_cast_fp16)[name = tensor("op_1516_cast_fp16")]; + tensor koi_13_cast_fp16 = add(x = var_1515_cast_fp16, y = var_1516_cast_fp16)[name = tensor("koi_13_cast_fp16")]; + tensor qo_7_axis_0 = const()[name = tensor("qo_7_axis_0"), val = tensor(-1)]; + tensor qo_7_cast_fp16 = stack(axis = qo_7_axis_0, values = (qor_13_cast_fp16, qoi_13_cast_fp16))[name = tensor("qo_7_cast_fp16")]; + tensor ko_7_axis_0 = const()[name = tensor("ko_7_axis_0"), val = tensor(-1)]; + tensor ko_7_cast_fp16 = stack(axis = ko_7_axis_0, values = (kor_13_cast_fp16, koi_13_cast_fp16))[name = tensor("ko_7_cast_fp16")]; + tensor var_1545 = const()[name = tensor("op_1545"), val = tensor([1, 1, 16, 64])]; + tensor q_21_cast_fp16 = reshape(shape = var_1545, x = qo_7_cast_fp16)[name = tensor("q_21_cast_fp16")]; + tensor var_1547 = const()[name = tensor("op_1547"), val = tensor([1, 1, 16, 64])]; + tensor k_15_cast_fp16 = reshape(shape = var_1547, x = ko_7_cast_fp16)[name = tensor("k_15_cast_fp16")]; + tensor _inversed_1569_y_0_to_fp16 = const()[name = tensor("_inversed_1569_y_0_to_fp16"), val = tensor(0x1p-9)]; + tensor _inversed_1569_cast_fp16 = mul(x = ts_23_cast_fp16, y = _inversed_1569_y_0_to_fp16)[name = tensor("_inversed_1569_cast_fp16")]; + tensor var_1570_cast_fp16 = floor(x = _inversed_1569_cast_fp16)[name = tensor("op_1570_cast_fp16")]; + tensor var_1571_to_fp16 = const()[name = tensor("op_1571_to_fp16"), val = tensor(0x1p+9)]; + tensor var_1572_cast_fp16 = mul(x = var_1570_cast_fp16, y = var_1571_to_fp16)[name = tensor("op_1572_cast_fp16")]; + tensor write_indices_float_15_cast_fp16 = sub(x = ts_23_cast_fp16, y = var_1572_cast_fp16)[name = tensor("write_indices_float_15_cast_fp16")]; + tensor var_1579_dtype_0 = const()[name = tensor("op_1579_dtype_0"), val = tensor("int32")]; + tensor write_indices_7_reps_0 = const()[name = tensor("write_indices_7_reps_0"), val = tensor([1, 1, 16, 64])]; + tensor write_indices_float_15_cast_fp16_to_int32 = cast(dtype = var_1579_dtype_0, x = write_indices_float_15_cast_fp16)[name = tensor("cast_123")]; + tensor write_indices_7 = tile(reps = write_indices_7_reps_0, x = write_indices_float_15_cast_fp16_to_int32)[name = tensor("write_indices_7")]; + tensor var_1587_begin_0 = const()[name = tensor("op_1587_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_1587_end_0 = const()[name = tensor("op_1587_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor var_1587_end_mask_0 = const()[name = tensor("op_1587_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_1587_squeeze_mask_0 = const()[name = tensor("op_1587_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor cache3_to_fp16_dtype_0 = const()[name = tensor("cache3_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor cache3_to_fp16 = cast(dtype = cache3_to_fp16_dtype_0, x = cache3)[name = tensor("cast_122")]; + tensor var_1587_cast_fp16 = slice_by_index(begin = var_1587_begin_0, end = var_1587_end_0, end_mask = var_1587_end_mask_0, squeeze_mask = var_1587_squeeze_mask_0, x = cache3_to_fp16)[name = tensor("op_1587_cast_fp16")]; + tensor var_1589_axis_0 = const()[name = tensor("op_1589_axis_0"), val = tensor(1)]; + tensor var_1589_mode_0 = const()[name = tensor("op_1589_mode_0"), val = tensor("update")]; + tensor var_1589_validate_indices_0 = const()[name = tensor("op_1589_validate_indices_0"), val = tensor(false)]; + tensor var_1589_cast_fp16 = scatter_along_axis(axis = var_1589_axis_0, data = var_1587_cast_fp16, indices = write_indices_7, mode = var_1589_mode_0, updates = k_15_cast_fp16, validate_indices = var_1589_validate_indices_0)[name = tensor("op_1589_cast_fp16")]; + tensor concat_23 = const()[name = tensor("concat_23"), val = tensor([0, 0, 0, 0, 0])]; + tensor concat_24 = const()[name = tensor("concat_24"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_7_internal_tensor_assign_1_stride_0 = const()[name = tensor("new_cache_7_internal_tensor_assign_1_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_7_internal_tensor_assign_1_begin_mask_0 = const()[name = tensor("new_cache_7_internal_tensor_assign_1_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_7_internal_tensor_assign_1_end_mask_0 = const()[name = tensor("new_cache_7_internal_tensor_assign_1_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_7_internal_tensor_assign_1_squeeze_mask_0 = const()[name = tensor("new_cache_7_internal_tensor_assign_1_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_18 = const()[name = tensor("shape_18"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_6 = const()[name = tensor("reduce_prod_6"), val = tensor(1048576)]; + tensor range_1d_6_start_0 = const()[name = tensor("range_1d_6_start_0"), val = tensor(0)]; + tensor range_1d_6_step_0 = const()[name = tensor("range_1d_6_step_0"), val = tensor(1)]; + tensor range_1d_6 = range_1d(end = reduce_prod_6, start = range_1d_6_start_0, step = range_1d_6_step_0)[name = tensor("range_1d_6")]; + tensor reshape_30 = reshape(shape = shape_18, x = range_1d_6)[name = tensor("reshape_30")]; + tensor slice_by_index_6 = slice_by_index(begin = concat_23, begin_mask = new_cache_7_internal_tensor_assign_1_begin_mask_0, end = concat_24, end_mask = new_cache_7_internal_tensor_assign_1_end_mask_0, squeeze_mask = new_cache_7_internal_tensor_assign_1_squeeze_mask_0, stride = new_cache_7_internal_tensor_assign_1_stride_0, x = reshape_30)[name = tensor("slice_by_index_6")]; + tensor reshape_31_shape_0 = const()[name = tensor("reshape_31_shape_0"), val = tensor([-1])]; + tensor reshape_31 = reshape(shape = reshape_31_shape_0, x = slice_by_index_6)[name = tensor("reshape_31")]; + tensor reshape_32_shape_0 = const()[name = tensor("reshape_32_shape_0"), val = tensor([-1])]; + tensor reshape_32_cast_fp16 = reshape(shape = reshape_32_shape_0, x = var_1589_cast_fp16)[name = tensor("reshape_32_cast_fp16")]; + tensor reshape_33_shape_0 = const()[name = tensor("reshape_33_shape_0"), val = tensor([-1])]; + tensor reshape_33_cast_fp16 = reshape(shape = reshape_33_shape_0, x = cache3_to_fp16)[name = tensor("reshape_33_cast_fp16")]; + tensor scatter_6_mode_0 = const()[name = tensor("scatter_6_mode_0"), val = tensor("update")]; + tensor scatter_6_axis_0 = const()[name = tensor("scatter_6_axis_0"), val = tensor(0)]; + tensor scatter_6_validate_indices_0 = const()[name = tensor("scatter_6_validate_indices_0"), val = tensor(false)]; + tensor scatter_6_cast_fp16 = scatter(axis = scatter_6_axis_0, data = reshape_33_cast_fp16, indices = reshape_31, mode = scatter_6_mode_0, updates = reshape_32_cast_fp16, validate_indices = scatter_6_validate_indices_0)[name = tensor("scatter_6_cast_fp16")]; + tensor reshape_34_cast_fp16 = reshape(shape = shape_18, x = scatter_6_cast_fp16)[name = tensor("reshape_34_cast_fp16")]; + tensor var_1597_begin_0 = const()[name = tensor("op_1597_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor var_1597_end_0 = const()[name = tensor("op_1597_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor var_1597_end_mask_0 = const()[name = tensor("op_1597_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_1597_squeeze_mask_0 = const()[name = tensor("op_1597_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor var_1597_cast_fp16 = slice_by_index(begin = var_1597_begin_0, end = var_1597_end_0, end_mask = var_1597_end_mask_0, squeeze_mask = var_1597_squeeze_mask_0, x = reshape_34_cast_fp16)[name = tensor("op_1597_cast_fp16")]; + tensor var_1599_axis_0 = const()[name = tensor("op_1599_axis_0"), val = tensor(1)]; + tensor var_1599_mode_0 = const()[name = tensor("op_1599_mode_0"), val = tensor("update")]; + tensor var_1599_validate_indices_0 = const()[name = tensor("op_1599_validate_indices_0"), val = tensor(false)]; + tensor var_1599_cast_fp16 = scatter_along_axis(axis = var_1599_axis_0, data = var_1597_cast_fp16, indices = write_indices_7, mode = var_1599_mode_0, updates = v_7_cast_fp16, validate_indices = var_1599_validate_indices_0)[name = tensor("op_1599_cast_fp16")]; + tensor concat_25 = const()[name = tensor("concat_25"), val = tensor([1, 0, 0, 0, 0])]; + tensor concat_26 = const()[name = tensor("concat_26"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_7_internal_tensor_assign_2_stride_0 = const()[name = tensor("new_cache_7_internal_tensor_assign_2_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_7_internal_tensor_assign_2_begin_mask_0 = const()[name = tensor("new_cache_7_internal_tensor_assign_2_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_7_internal_tensor_assign_2_end_mask_0 = const()[name = tensor("new_cache_7_internal_tensor_assign_2_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_7_internal_tensor_assign_2_squeeze_mask_0 = const()[name = tensor("new_cache_7_internal_tensor_assign_2_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_19 = const()[name = tensor("shape_19"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_7 = const()[name = tensor("reduce_prod_7"), val = tensor(1048576)]; + tensor range_1d_7_start_0 = const()[name = tensor("range_1d_7_start_0"), val = tensor(0)]; + tensor range_1d_7_step_0 = const()[name = tensor("range_1d_7_step_0"), val = tensor(1)]; + tensor range_1d_7 = range_1d(end = reduce_prod_7, start = range_1d_7_start_0, step = range_1d_7_step_0)[name = tensor("range_1d_7")]; + tensor reshape_35 = reshape(shape = shape_19, x = range_1d_7)[name = tensor("reshape_35")]; + tensor slice_by_index_7 = slice_by_index(begin = concat_25, begin_mask = new_cache_7_internal_tensor_assign_2_begin_mask_0, end = concat_26, end_mask = new_cache_7_internal_tensor_assign_2_end_mask_0, squeeze_mask = new_cache_7_internal_tensor_assign_2_squeeze_mask_0, stride = new_cache_7_internal_tensor_assign_2_stride_0, x = reshape_35)[name = tensor("slice_by_index_7")]; + tensor reshape_36_shape_0 = const()[name = tensor("reshape_36_shape_0"), val = tensor([-1])]; + tensor reshape_36 = reshape(shape = reshape_36_shape_0, x = slice_by_index_7)[name = tensor("reshape_36")]; + tensor reshape_37_shape_0 = const()[name = tensor("reshape_37_shape_0"), val = tensor([-1])]; + tensor reshape_37_cast_fp16 = reshape(shape = reshape_37_shape_0, x = var_1599_cast_fp16)[name = tensor("reshape_37_cast_fp16")]; + tensor reshape_38_shape_0 = const()[name = tensor("reshape_38_shape_0"), val = tensor([-1])]; + tensor reshape_38_cast_fp16 = reshape(shape = reshape_38_shape_0, x = reshape_34_cast_fp16)[name = tensor("reshape_38_cast_fp16")]; + tensor scatter_7_mode_0 = const()[name = tensor("scatter_7_mode_0"), val = tensor("update")]; + tensor scatter_7_axis_0 = const()[name = tensor("scatter_7_axis_0"), val = tensor(0)]; + tensor scatter_7_validate_indices_0 = const()[name = tensor("scatter_7_validate_indices_0"), val = tensor(false)]; + tensor scatter_7_cast_fp16 = scatter(axis = scatter_7_axis_0, data = reshape_38_cast_fp16, indices = reshape_36, mode = scatter_7_mode_0, updates = reshape_37_cast_fp16, validate_indices = scatter_7_validate_indices_0)[name = tensor("scatter_7_cast_fp16")]; + tensor reshape_39_cast_fp16 = reshape(shape = shape_19, x = scatter_7_cast_fp16)[name = tensor("reshape_39_cast_fp16")]; + tensor reshape_39_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("reshape_39_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor keys_19_begin_0 = const()[name = tensor("keys_19_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor keys_19_end_0 = const()[name = tensor("keys_19_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor keys_19_end_mask_0 = const()[name = tensor("keys_19_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor keys_19_squeeze_mask_0 = const()[name = tensor("keys_19_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor keys_19_cast_fp16 = slice_by_index(begin = keys_19_begin_0, end = keys_19_end_0, end_mask = keys_19_end_mask_0, squeeze_mask = keys_19_squeeze_mask_0, x = reshape_39_cast_fp16)[name = tensor("keys_19_cast_fp16")]; + tensor values_19_begin_0 = const()[name = tensor("values_19_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor values_19_end_0 = const()[name = tensor("values_19_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor values_19_end_mask_0 = const()[name = tensor("values_19_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor values_19_squeeze_mask_0 = const()[name = tensor("values_19_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor values_19_cast_fp16 = slice_by_index(begin = values_19_begin_0, end = values_19_end_0, end_mask = values_19_end_mask_0, squeeze_mask = values_19_squeeze_mask_0, x = reshape_39_cast_fp16)[name = tensor("values_19_cast_fp16")]; + tensor var_1611_cast_fp16 = not_equal(x = keys_19_cast_fp16, y = keys_19_cast_fp16)[name = tensor("op_1611_cast_fp16")]; + tensor keys_21_cast_fp16 = select(a = var_360_to_fp16, b = keys_19_cast_fp16, cond = var_1611_cast_fp16)[name = tensor("keys_21_cast_fp16")]; + tensor var_1619_cast_fp16 = not_equal(x = values_19_cast_fp16, y = values_19_cast_fp16)[name = tensor("op_1619_cast_fp16")]; + tensor values_21_cast_fp16 = select(a = var_360_to_fp16, b = values_19_cast_fp16, cond = var_1619_cast_fp16)[name = tensor("values_21_cast_fp16")]; + tensor var_1643 = const()[name = tensor("op_1643"), val = tensor([0, 2, 1, 3])]; + tensor var_1656 = const()[name = tensor("op_1656"), val = tensor([1, 1, 1])]; + tensor var_1657_cast_fp16 = reshape(shape = var_1656, x = position3_to_fp16)[name = tensor("op_1657_cast_fp16")]; + tensor var_1674_to_fp16 = const()[name = tensor("op_1674_to_fp16"), val = tensor(0x1p+0)]; + tensor valid_len_7_cast_fp16 = add(x = var_1657_cast_fp16, y = var_1674_to_fp16)[name = tensor("valid_len_7_cast_fp16")]; + tensor valid_mask_7_cast_fp16 = less(x = k_positions_1_promoted_to_fp16, y = valid_len_7_cast_fp16)[name = tensor("valid_mask_7_cast_fp16")]; + tensor causal_mask_7_cast_fp16 = less_equal(x = k_positions_1_promoted_to_fp16, y = var_1657_cast_fp16)[name = tensor("causal_mask_7_cast_fp16")]; + tensor attn_mask_13 = logical_and(x = valid_mask_7_cast_fp16, y = causal_mask_7_cast_fp16)[name = tensor("attn_mask_13")]; + tensor attn_mask_15_axes_0 = const()[name = tensor("attn_mask_15_axes_0"), val = tensor([1])]; + tensor attn_mask_15 = expand_dims(axes = attn_mask_15_axes_0, x = attn_mask_13)[name = tensor("attn_mask_15")]; + tensor var_1692_transpose_x_0 = const()[name = tensor("op_1692_transpose_x_0"), val = tensor(false)]; + tensor var_1692_transpose_y_0 = const()[name = tensor("op_1692_transpose_y_0"), val = tensor(false)]; + tensor transpose_24_perm_0 = const()[name = tensor("transpose_24_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_25_perm_0 = const()[name = tensor("transpose_25_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_25 = transpose(perm = transpose_25_perm_0, x = keys_21_cast_fp16)[name = tensor("transpose_39")]; + tensor transpose_24 = transpose(perm = transpose_24_perm_0, x = q_21_cast_fp16)[name = tensor("transpose_40")]; + tensor var_1692_cast_fp16 = matmul(transpose_x = var_1692_transpose_x_0, transpose_y = var_1692_transpose_y_0, x = transpose_24, y = transpose_25)[name = tensor("op_1692_cast_fp16")]; + tensor var_1686_to_fp16 = const()[name = tensor("op_1686_to_fp16"), val = tensor([0x1p-3])]; + tensor attn_weights_19_cast_fp16 = mul(x = var_1692_cast_fp16, y = var_1686_to_fp16)[name = tensor("attn_weights_19_cast_fp16")]; + tensor var_1694 = logical_not(x = attn_mask_15)[name = tensor("op_1694")]; + tensor var_1695_to_fp16 = const()[name = tensor("op_1695_to_fp16"), val = tensor(-inf)]; + tensor attn_weights_21_cast_fp16 = select(a = var_1695_to_fp16, b = attn_weights_19_cast_fp16, cond = var_1694)[name = tensor("attn_weights_21_cast_fp16")]; + tensor var_1697 = const()[name = tensor("op_1697"), val = tensor(-1)]; + tensor attn_weights_23_cast_fp16 = softmax(axis = var_1697, x = attn_weights_21_cast_fp16)[name = tensor("attn_weights_23_cast_fp16")]; + tensor attn_output_7_transpose_x_0 = const()[name = tensor("attn_output_7_transpose_x_0"), val = tensor(false)]; + tensor attn_output_7_transpose_y_0 = const()[name = tensor("attn_output_7_transpose_y_0"), val = tensor(false)]; + tensor values_23_cast_fp16 = transpose(perm = var_1643, x = values_21_cast_fp16)[name = tensor("transpose_41")]; + tensor attn_output_7_cast_fp16 = matmul(transpose_x = attn_output_7_transpose_x_0, transpose_y = attn_output_7_transpose_y_0, x = attn_weights_23_cast_fp16, y = values_23_cast_fp16)[name = tensor("attn_output_7_cast_fp16")]; + tensor var_1705 = const()[name = tensor("op_1705"), val = tensor([0, 2, 1, 3])]; + tensor var_1708 = const()[name = tensor("op_1708"), val = tensor([1, 1, 1024])]; + tensor var_1706_cast_fp16 = transpose(perm = var_1705, x = attn_output_7_cast_fp16)[name = tensor("transpose_38")]; + tensor input_35_cast_fp16 = reshape(shape = var_1708, x = var_1706_cast_fp16)[name = tensor("input_35_cast_fp16")]; + tensor linear_14_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = attn3_out_proj_weight_cast_fp16, x = input_35_cast_fp16)[name = tensor("linear_14_cast_fp16")]; + tensor var_1714_to_fp16 = const()[name = tensor("op_1714_to_fp16"), val = tensor(0x1p+0)]; + tensor var_1715_cast_fp16 = add(x = position3_to_fp16, y = var_1714_to_fp16)[name = tensor("op_1715_cast_fp16")]; + tensor var_1715_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_1715_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor input_37_cast_fp16 = add(x = input_33_cast_fp16, y = linear_14_cast_fp16)[name = tensor("input_37_cast_fp16")]; + tensor input_39_axes_0 = const()[name = tensor("input_39_axes_0"), val = tensor([-1])]; + tensor norm3_2_weight_to_fp16 = const()[name = tensor("norm3_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(76781568)))]; + tensor norm3_2_bias_to_fp16 = const()[name = tensor("norm3_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(76783680)))]; + tensor var_1719_to_fp16 = const()[name = tensor("op_1719_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_39_cast_fp16 = layer_norm(axes = input_39_axes_0, beta = norm3_2_bias_to_fp16, epsilon = var_1719_to_fp16, gamma = norm3_2_weight_to_fp16, x = input_37_cast_fp16)[name = tensor("input_39_cast_fp16")]; + tensor linear_15_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = linear3_1_weight_cast_fp16, x = input_39_cast_fp16)[name = tensor("linear_15_cast_fp16")]; + tensor input_41_mode_0 = const()[name = tensor("input_41_mode_0"), val = tensor("EXACT")]; + tensor input_41_cast_fp16 = gelu(mode = input_41_mode_0, x = linear_15_cast_fp16)[name = tensor("input_41_cast_fp16")]; + tensor linear_16_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = linear3_2_weight_cast_fp16, x = input_41_cast_fp16)[name = tensor("linear_16_cast_fp16")]; + tensor input_43_cast_fp16 = add(x = input_37_cast_fp16, y = linear_16_cast_fp16)[name = tensor("input_43_cast_fp16")]; + tensor x_9_axes_0 = const()[name = tensor("x_9_axes_0"), val = tensor([-1])]; + tensor norm4_1_weight_to_fp16 = const()[name = tensor("norm4_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(76785792)))]; + tensor norm4_1_bias_to_fp16 = const()[name = tensor("norm4_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(76787904)))]; + tensor var_1736_to_fp16 = const()[name = tensor("op_1736_to_fp16"), val = tensor(0x1.5p-17)]; + tensor x_9_cast_fp16 = layer_norm(axes = x_9_axes_0, beta = norm4_1_bias_to_fp16, epsilon = var_1736_to_fp16, gamma = norm4_1_weight_to_fp16, x = input_43_cast_fp16)[name = tensor("x_9_cast_fp16")]; + tensor linear_17_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = attn4_in_proj_weight_cast_fp16, x = x_9_cast_fp16)[name = tensor("linear_17_cast_fp16")]; + tensor var_1772 = const()[name = tensor("op_1772"), val = tensor([1, 1, 3, 16, 64])]; + tensor qkv_9_cast_fp16 = reshape(shape = var_1772, x = linear_17_cast_fp16)[name = tensor("qkv_9_cast_fp16")]; + tensor q_25_begin_0 = const()[name = tensor("q_25_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor q_25_end_0 = const()[name = tensor("q_25_end_0"), val = tensor([1, 1, 1, 16, 64])]; + tensor q_25_end_mask_0 = const()[name = tensor("q_25_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor q_25_squeeze_mask_0 = const()[name = tensor("q_25_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor q_25_cast_fp16 = slice_by_index(begin = q_25_begin_0, end = q_25_end_0, end_mask = q_25_end_mask_0, squeeze_mask = q_25_squeeze_mask_0, x = qkv_9_cast_fp16)[name = tensor("q_25_cast_fp16")]; + tensor k_17_begin_0 = const()[name = tensor("k_17_begin_0"), val = tensor([0, 0, 1, 0, 0])]; + tensor k_17_end_0 = const()[name = tensor("k_17_end_0"), val = tensor([1, 1, 2, 16, 64])]; + tensor k_17_end_mask_0 = const()[name = tensor("k_17_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor k_17_squeeze_mask_0 = const()[name = tensor("k_17_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor k_17_cast_fp16 = slice_by_index(begin = k_17_begin_0, end = k_17_end_0, end_mask = k_17_end_mask_0, squeeze_mask = k_17_squeeze_mask_0, x = qkv_9_cast_fp16)[name = tensor("k_17_cast_fp16")]; + tensor v_9_begin_0 = const()[name = tensor("v_9_begin_0"), val = tensor([0, 0, 2, 0, 0])]; + tensor v_9_end_0 = const()[name = tensor("v_9_end_0"), val = tensor([1, 1, 3, 16, 64])]; + tensor v_9_end_mask_0 = const()[name = tensor("v_9_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor v_9_squeeze_mask_0 = const()[name = tensor("v_9_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor v_9_cast_fp16 = slice_by_index(begin = v_9_begin_0, end = v_9_end_0, end_mask = v_9_end_mask_0, squeeze_mask = v_9_squeeze_mask_0, x = qkv_9_cast_fp16)[name = tensor("v_9_cast_fp16")]; + tensor var_1876 = const()[name = tensor("op_1876"), val = tensor([1, 1, 1, 1])]; + tensor position4_to_fp16_dtype_0 = const()[name = tensor("position4_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor position4_to_fp16 = cast(dtype = position4_to_fp16_dtype_0, x = position4)[name = tensor("cast_119")]; + tensor ts_29_cast_fp16 = reshape(shape = var_1876, x = position4_to_fp16)[name = tensor("ts_29_cast_fp16")]; + tensor var_1880 = const()[name = tensor("op_1880"), val = tensor([1, 1, 16, 32, 2])]; + tensor q_complex_9_cast_fp16 = reshape(shape = var_1880, x = q_25_cast_fp16)[name = tensor("q_complex_9_cast_fp16")]; + tensor var_1884 = const()[name = tensor("op_1884"), val = tensor([1, 1, 16, 32, 2])]; + tensor k_complex_9_cast_fp16 = reshape(shape = var_1884, x = k_17_cast_fp16)[name = tensor("k_complex_9_cast_fp16")]; + tensor var_1888_begin_0 = const()[name = tensor("op_1888_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_1888_end_0 = const()[name = tensor("op_1888_end_0"), val = tensor([1, 1, 16, 32, 1])]; + tensor var_1888_end_mask_0 = const()[name = tensor("op_1888_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_1888_squeeze_mask_0 = const()[name = tensor("op_1888_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_1888_cast_fp16 = slice_by_index(begin = var_1888_begin_0, end = var_1888_end_0, end_mask = var_1888_end_mask_0, squeeze_mask = var_1888_squeeze_mask_0, x = q_complex_9_cast_fp16)[name = tensor("op_1888_cast_fp16")]; + tensor var_1896_begin_0 = const()[name = tensor("op_1896_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_1896_end_0 = const()[name = tensor("op_1896_end_0"), val = tensor([1, 1, 16, 32, 2])]; + tensor var_1896_end_mask_0 = const()[name = tensor("op_1896_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_1896_squeeze_mask_0 = const()[name = tensor("op_1896_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_1896_cast_fp16 = slice_by_index(begin = var_1896_begin_0, end = var_1896_end_0, end_mask = var_1896_end_mask_0, squeeze_mask = var_1896_squeeze_mask_0, x = q_complex_9_cast_fp16)[name = tensor("op_1896_cast_fp16")]; + tensor var_1904_begin_0 = const()[name = tensor("op_1904_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_1904_end_0 = const()[name = tensor("op_1904_end_0"), val = tensor([1, 1, 16, 32, 1])]; + tensor var_1904_end_mask_0 = const()[name = tensor("op_1904_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_1904_squeeze_mask_0 = const()[name = tensor("op_1904_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_1904_cast_fp16 = slice_by_index(begin = var_1904_begin_0, end = var_1904_end_0, end_mask = var_1904_end_mask_0, squeeze_mask = var_1904_squeeze_mask_0, x = k_complex_9_cast_fp16)[name = tensor("op_1904_cast_fp16")]; + tensor var_1912_begin_0 = const()[name = tensor("op_1912_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_1912_end_0 = const()[name = tensor("op_1912_end_0"), val = tensor([1, 1, 16, 32, 2])]; + tensor var_1912_end_mask_0 = const()[name = tensor("op_1912_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_1912_squeeze_mask_0 = const()[name = tensor("op_1912_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_1912_cast_fp16 = slice_by_index(begin = var_1912_begin_0, end = var_1912_end_0, end_mask = var_1912_end_mask_0, squeeze_mask = var_1912_squeeze_mask_0, x = k_complex_9_cast_fp16)[name = tensor("op_1912_cast_fp16")]; + tensor freqs_9_to_fp16 = const()[name = tensor("freqs_9_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(76790016)))]; + tensor var_1918_cast_fp16 = mul(x = freqs_9_to_fp16, y = ts_29_cast_fp16)[name = tensor("op_1918_cast_fp16")]; + tensor rotr_9_cast_fp16 = cos(x = var_1918_cast_fp16)[name = tensor("rotr_9_cast_fp16")]; + tensor roti_9_cast_fp16 = sin(x = var_1918_cast_fp16)[name = tensor("roti_9_cast_fp16")]; + tensor var_1922_cast_fp16 = mul(x = var_1888_cast_fp16, y = rotr_9_cast_fp16)[name = tensor("op_1922_cast_fp16")]; + tensor var_1923_cast_fp16 = mul(x = var_1896_cast_fp16, y = roti_9_cast_fp16)[name = tensor("op_1923_cast_fp16")]; + tensor qor_17_cast_fp16 = sub(x = var_1922_cast_fp16, y = var_1923_cast_fp16)[name = tensor("qor_17_cast_fp16")]; + tensor var_1926_cast_fp16 = mul(x = var_1888_cast_fp16, y = roti_9_cast_fp16)[name = tensor("op_1926_cast_fp16")]; + tensor var_1927_cast_fp16 = mul(x = var_1896_cast_fp16, y = rotr_9_cast_fp16)[name = tensor("op_1927_cast_fp16")]; + tensor qoi_17_cast_fp16 = add(x = var_1926_cast_fp16, y = var_1927_cast_fp16)[name = tensor("qoi_17_cast_fp16")]; + tensor var_1930_cast_fp16 = mul(x = var_1904_cast_fp16, y = rotr_9_cast_fp16)[name = tensor("op_1930_cast_fp16")]; + tensor var_1931_cast_fp16 = mul(x = var_1912_cast_fp16, y = roti_9_cast_fp16)[name = tensor("op_1931_cast_fp16")]; + tensor kor_17_cast_fp16 = sub(x = var_1930_cast_fp16, y = var_1931_cast_fp16)[name = tensor("kor_17_cast_fp16")]; + tensor var_1934_cast_fp16 = mul(x = var_1904_cast_fp16, y = roti_9_cast_fp16)[name = tensor("op_1934_cast_fp16")]; + tensor var_1935_cast_fp16 = mul(x = var_1912_cast_fp16, y = rotr_9_cast_fp16)[name = tensor("op_1935_cast_fp16")]; + tensor koi_17_cast_fp16 = add(x = var_1934_cast_fp16, y = var_1935_cast_fp16)[name = tensor("koi_17_cast_fp16")]; + tensor qo_9_axis_0 = const()[name = tensor("qo_9_axis_0"), val = tensor(-1)]; + tensor qo_9_cast_fp16 = stack(axis = qo_9_axis_0, values = (qor_17_cast_fp16, qoi_17_cast_fp16))[name = tensor("qo_9_cast_fp16")]; + tensor ko_9_axis_0 = const()[name = tensor("ko_9_axis_0"), val = tensor(-1)]; + tensor ko_9_cast_fp16 = stack(axis = ko_9_axis_0, values = (kor_17_cast_fp16, koi_17_cast_fp16))[name = tensor("ko_9_cast_fp16")]; + tensor var_1964 = const()[name = tensor("op_1964"), val = tensor([1, 1, 16, 64])]; + tensor q_27_cast_fp16 = reshape(shape = var_1964, x = qo_9_cast_fp16)[name = tensor("q_27_cast_fp16")]; + tensor var_1966 = const()[name = tensor("op_1966"), val = tensor([1, 1, 16, 64])]; + tensor k_19_cast_fp16 = reshape(shape = var_1966, x = ko_9_cast_fp16)[name = tensor("k_19_cast_fp16")]; + tensor _inversed_1988_y_0_to_fp16 = const()[name = tensor("_inversed_1988_y_0_to_fp16"), val = tensor(0x1p-9)]; + tensor _inversed_1988_cast_fp16 = mul(x = ts_29_cast_fp16, y = _inversed_1988_y_0_to_fp16)[name = tensor("_inversed_1988_cast_fp16")]; + tensor var_1989_cast_fp16 = floor(x = _inversed_1988_cast_fp16)[name = tensor("op_1989_cast_fp16")]; + tensor var_1990_to_fp16 = const()[name = tensor("op_1990_to_fp16"), val = tensor(0x1p+9)]; + tensor var_1991_cast_fp16 = mul(x = var_1989_cast_fp16, y = var_1990_to_fp16)[name = tensor("op_1991_cast_fp16")]; + tensor write_indices_float_19_cast_fp16 = sub(x = ts_29_cast_fp16, y = var_1991_cast_fp16)[name = tensor("write_indices_float_19_cast_fp16")]; + tensor var_1998_dtype_0 = const()[name = tensor("op_1998_dtype_0"), val = tensor("int32")]; + tensor write_indices_9_reps_0 = const()[name = tensor("write_indices_9_reps_0"), val = tensor([1, 1, 16, 64])]; + tensor write_indices_float_19_cast_fp16_to_int32 = cast(dtype = var_1998_dtype_0, x = write_indices_float_19_cast_fp16)[name = tensor("cast_118")]; + tensor write_indices_9 = tile(reps = write_indices_9_reps_0, x = write_indices_float_19_cast_fp16_to_int32)[name = tensor("write_indices_9")]; + tensor var_2006_begin_0 = const()[name = tensor("op_2006_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_2006_end_0 = const()[name = tensor("op_2006_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor var_2006_end_mask_0 = const()[name = tensor("op_2006_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_2006_squeeze_mask_0 = const()[name = tensor("op_2006_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor cache4_to_fp16_dtype_0 = const()[name = tensor("cache4_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor cache4_to_fp16 = cast(dtype = cache4_to_fp16_dtype_0, x = cache4)[name = tensor("cast_117")]; + tensor var_2006_cast_fp16 = slice_by_index(begin = var_2006_begin_0, end = var_2006_end_0, end_mask = var_2006_end_mask_0, squeeze_mask = var_2006_squeeze_mask_0, x = cache4_to_fp16)[name = tensor("op_2006_cast_fp16")]; + tensor var_2008_axis_0 = const()[name = tensor("op_2008_axis_0"), val = tensor(1)]; + tensor var_2008_mode_0 = const()[name = tensor("op_2008_mode_0"), val = tensor("update")]; + tensor var_2008_validate_indices_0 = const()[name = tensor("op_2008_validate_indices_0"), val = tensor(false)]; + tensor var_2008_cast_fp16 = scatter_along_axis(axis = var_2008_axis_0, data = var_2006_cast_fp16, indices = write_indices_9, mode = var_2008_mode_0, updates = k_19_cast_fp16, validate_indices = var_2008_validate_indices_0)[name = tensor("op_2008_cast_fp16")]; + tensor concat_30 = const()[name = tensor("concat_30"), val = tensor([0, 0, 0, 0, 0])]; + tensor concat_31 = const()[name = tensor("concat_31"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_9_internal_tensor_assign_1_stride_0 = const()[name = tensor("new_cache_9_internal_tensor_assign_1_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_9_internal_tensor_assign_1_begin_mask_0 = const()[name = tensor("new_cache_9_internal_tensor_assign_1_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_9_internal_tensor_assign_1_end_mask_0 = const()[name = tensor("new_cache_9_internal_tensor_assign_1_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_9_internal_tensor_assign_1_squeeze_mask_0 = const()[name = tensor("new_cache_9_internal_tensor_assign_1_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_20 = const()[name = tensor("shape_20"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_8 = const()[name = tensor("reduce_prod_8"), val = tensor(1048576)]; + tensor range_1d_8_start_0 = const()[name = tensor("range_1d_8_start_0"), val = tensor(0)]; + tensor range_1d_8_step_0 = const()[name = tensor("range_1d_8_step_0"), val = tensor(1)]; + tensor range_1d_8 = range_1d(end = reduce_prod_8, start = range_1d_8_start_0, step = range_1d_8_step_0)[name = tensor("range_1d_8")]; + tensor reshape_40 = reshape(shape = shape_20, x = range_1d_8)[name = tensor("reshape_40")]; + tensor slice_by_index_8 = slice_by_index(begin = concat_30, begin_mask = new_cache_9_internal_tensor_assign_1_begin_mask_0, end = concat_31, end_mask = new_cache_9_internal_tensor_assign_1_end_mask_0, squeeze_mask = new_cache_9_internal_tensor_assign_1_squeeze_mask_0, stride = new_cache_9_internal_tensor_assign_1_stride_0, x = reshape_40)[name = tensor("slice_by_index_8")]; + tensor reshape_41_shape_0 = const()[name = tensor("reshape_41_shape_0"), val = tensor([-1])]; + tensor reshape_41 = reshape(shape = reshape_41_shape_0, x = slice_by_index_8)[name = tensor("reshape_41")]; + tensor reshape_42_shape_0 = const()[name = tensor("reshape_42_shape_0"), val = tensor([-1])]; + tensor reshape_42_cast_fp16 = reshape(shape = reshape_42_shape_0, x = var_2008_cast_fp16)[name = tensor("reshape_42_cast_fp16")]; + tensor reshape_43_shape_0 = const()[name = tensor("reshape_43_shape_0"), val = tensor([-1])]; + tensor reshape_43_cast_fp16 = reshape(shape = reshape_43_shape_0, x = cache4_to_fp16)[name = tensor("reshape_43_cast_fp16")]; + tensor scatter_8_mode_0 = const()[name = tensor("scatter_8_mode_0"), val = tensor("update")]; + tensor scatter_8_axis_0 = const()[name = tensor("scatter_8_axis_0"), val = tensor(0)]; + tensor scatter_8_validate_indices_0 = const()[name = tensor("scatter_8_validate_indices_0"), val = tensor(false)]; + tensor scatter_8_cast_fp16 = scatter(axis = scatter_8_axis_0, data = reshape_43_cast_fp16, indices = reshape_41, mode = scatter_8_mode_0, updates = reshape_42_cast_fp16, validate_indices = scatter_8_validate_indices_0)[name = tensor("scatter_8_cast_fp16")]; + tensor reshape_44_cast_fp16 = reshape(shape = shape_20, x = scatter_8_cast_fp16)[name = tensor("reshape_44_cast_fp16")]; + tensor var_2016_begin_0 = const()[name = tensor("op_2016_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor var_2016_end_0 = const()[name = tensor("op_2016_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor var_2016_end_mask_0 = const()[name = tensor("op_2016_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_2016_squeeze_mask_0 = const()[name = tensor("op_2016_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor var_2016_cast_fp16 = slice_by_index(begin = var_2016_begin_0, end = var_2016_end_0, end_mask = var_2016_end_mask_0, squeeze_mask = var_2016_squeeze_mask_0, x = reshape_44_cast_fp16)[name = tensor("op_2016_cast_fp16")]; + tensor var_2018_axis_0 = const()[name = tensor("op_2018_axis_0"), val = tensor(1)]; + tensor var_2018_mode_0 = const()[name = tensor("op_2018_mode_0"), val = tensor("update")]; + tensor var_2018_validate_indices_0 = const()[name = tensor("op_2018_validate_indices_0"), val = tensor(false)]; + tensor var_2018_cast_fp16 = scatter_along_axis(axis = var_2018_axis_0, data = var_2016_cast_fp16, indices = write_indices_9, mode = var_2018_mode_0, updates = v_9_cast_fp16, validate_indices = var_2018_validate_indices_0)[name = tensor("op_2018_cast_fp16")]; + tensor concat_32 = const()[name = tensor("concat_32"), val = tensor([1, 0, 0, 0, 0])]; + tensor concat_33 = const()[name = tensor("concat_33"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_9_internal_tensor_assign_2_stride_0 = const()[name = tensor("new_cache_9_internal_tensor_assign_2_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_9_internal_tensor_assign_2_begin_mask_0 = const()[name = tensor("new_cache_9_internal_tensor_assign_2_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_9_internal_tensor_assign_2_end_mask_0 = const()[name = tensor("new_cache_9_internal_tensor_assign_2_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_9_internal_tensor_assign_2_squeeze_mask_0 = const()[name = tensor("new_cache_9_internal_tensor_assign_2_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_21 = const()[name = tensor("shape_21"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_9 = const()[name = tensor("reduce_prod_9"), val = tensor(1048576)]; + tensor range_1d_9_start_0 = const()[name = tensor("range_1d_9_start_0"), val = tensor(0)]; + tensor range_1d_9_step_0 = const()[name = tensor("range_1d_9_step_0"), val = tensor(1)]; + tensor range_1d_9 = range_1d(end = reduce_prod_9, start = range_1d_9_start_0, step = range_1d_9_step_0)[name = tensor("range_1d_9")]; + tensor reshape_45 = reshape(shape = shape_21, x = range_1d_9)[name = tensor("reshape_45")]; + tensor slice_by_index_9 = slice_by_index(begin = concat_32, begin_mask = new_cache_9_internal_tensor_assign_2_begin_mask_0, end = concat_33, end_mask = new_cache_9_internal_tensor_assign_2_end_mask_0, squeeze_mask = new_cache_9_internal_tensor_assign_2_squeeze_mask_0, stride = new_cache_9_internal_tensor_assign_2_stride_0, x = reshape_45)[name = tensor("slice_by_index_9")]; + tensor reshape_46_shape_0 = const()[name = tensor("reshape_46_shape_0"), val = tensor([-1])]; + tensor reshape_46 = reshape(shape = reshape_46_shape_0, x = slice_by_index_9)[name = tensor("reshape_46")]; + tensor reshape_47_shape_0 = const()[name = tensor("reshape_47_shape_0"), val = tensor([-1])]; + tensor reshape_47_cast_fp16 = reshape(shape = reshape_47_shape_0, x = var_2018_cast_fp16)[name = tensor("reshape_47_cast_fp16")]; + tensor reshape_48_shape_0 = const()[name = tensor("reshape_48_shape_0"), val = tensor([-1])]; + tensor reshape_48_cast_fp16 = reshape(shape = reshape_48_shape_0, x = reshape_44_cast_fp16)[name = tensor("reshape_48_cast_fp16")]; + tensor scatter_9_mode_0 = const()[name = tensor("scatter_9_mode_0"), val = tensor("update")]; + tensor scatter_9_axis_0 = const()[name = tensor("scatter_9_axis_0"), val = tensor(0)]; + tensor scatter_9_validate_indices_0 = const()[name = tensor("scatter_9_validate_indices_0"), val = tensor(false)]; + tensor scatter_9_cast_fp16 = scatter(axis = scatter_9_axis_0, data = reshape_48_cast_fp16, indices = reshape_46, mode = scatter_9_mode_0, updates = reshape_47_cast_fp16, validate_indices = scatter_9_validate_indices_0)[name = tensor("scatter_9_cast_fp16")]; + tensor reshape_49_cast_fp16 = reshape(shape = shape_21, x = scatter_9_cast_fp16)[name = tensor("reshape_49_cast_fp16")]; + tensor reshape_49_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("reshape_49_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor keys_25_begin_0 = const()[name = tensor("keys_25_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor keys_25_end_0 = const()[name = tensor("keys_25_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor keys_25_end_mask_0 = const()[name = tensor("keys_25_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor keys_25_squeeze_mask_0 = const()[name = tensor("keys_25_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor keys_25_cast_fp16 = slice_by_index(begin = keys_25_begin_0, end = keys_25_end_0, end_mask = keys_25_end_mask_0, squeeze_mask = keys_25_squeeze_mask_0, x = reshape_49_cast_fp16)[name = tensor("keys_25_cast_fp16")]; + tensor values_25_begin_0 = const()[name = tensor("values_25_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor values_25_end_0 = const()[name = tensor("values_25_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor values_25_end_mask_0 = const()[name = tensor("values_25_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor values_25_squeeze_mask_0 = const()[name = tensor("values_25_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor values_25_cast_fp16 = slice_by_index(begin = values_25_begin_0, end = values_25_end_0, end_mask = values_25_end_mask_0, squeeze_mask = values_25_squeeze_mask_0, x = reshape_49_cast_fp16)[name = tensor("values_25_cast_fp16")]; + tensor var_2030_cast_fp16 = not_equal(x = keys_25_cast_fp16, y = keys_25_cast_fp16)[name = tensor("op_2030_cast_fp16")]; + tensor keys_27_cast_fp16 = select(a = var_360_to_fp16, b = keys_25_cast_fp16, cond = var_2030_cast_fp16)[name = tensor("keys_27_cast_fp16")]; + tensor var_2038_cast_fp16 = not_equal(x = values_25_cast_fp16, y = values_25_cast_fp16)[name = tensor("op_2038_cast_fp16")]; + tensor values_27_cast_fp16 = select(a = var_360_to_fp16, b = values_25_cast_fp16, cond = var_2038_cast_fp16)[name = tensor("values_27_cast_fp16")]; + tensor var_2062 = const()[name = tensor("op_2062"), val = tensor([0, 2, 1, 3])]; + tensor var_2075 = const()[name = tensor("op_2075"), val = tensor([1, 1, 1])]; + tensor var_2076_cast_fp16 = reshape(shape = var_2075, x = position4_to_fp16)[name = tensor("op_2076_cast_fp16")]; + tensor var_2093_to_fp16 = const()[name = tensor("op_2093_to_fp16"), val = tensor(0x1p+0)]; + tensor valid_len_9_cast_fp16 = add(x = var_2076_cast_fp16, y = var_2093_to_fp16)[name = tensor("valid_len_9_cast_fp16")]; + tensor valid_mask_9_cast_fp16 = less(x = k_positions_1_promoted_to_fp16, y = valid_len_9_cast_fp16)[name = tensor("valid_mask_9_cast_fp16")]; + tensor causal_mask_9_cast_fp16 = less_equal(x = k_positions_1_promoted_to_fp16, y = var_2076_cast_fp16)[name = tensor("causal_mask_9_cast_fp16")]; + tensor attn_mask_17 = logical_and(x = valid_mask_9_cast_fp16, y = causal_mask_9_cast_fp16)[name = tensor("attn_mask_17")]; + tensor attn_mask_19_axes_0 = const()[name = tensor("attn_mask_19_axes_0"), val = tensor([1])]; + tensor attn_mask_19 = expand_dims(axes = attn_mask_19_axes_0, x = attn_mask_17)[name = tensor("attn_mask_19")]; + tensor var_2111_transpose_x_0 = const()[name = tensor("op_2111_transpose_x_0"), val = tensor(false)]; + tensor var_2111_transpose_y_0 = const()[name = tensor("op_2111_transpose_y_0"), val = tensor(false)]; + tensor transpose_26_perm_0 = const()[name = tensor("transpose_26_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_27_perm_0 = const()[name = tensor("transpose_27_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_27 = transpose(perm = transpose_27_perm_0, x = keys_27_cast_fp16)[name = tensor("transpose_35")]; + tensor transpose_26 = transpose(perm = transpose_26_perm_0, x = q_27_cast_fp16)[name = tensor("transpose_36")]; + tensor var_2111_cast_fp16 = matmul(transpose_x = var_2111_transpose_x_0, transpose_y = var_2111_transpose_y_0, x = transpose_26, y = transpose_27)[name = tensor("op_2111_cast_fp16")]; + tensor var_2105_to_fp16 = const()[name = tensor("op_2105_to_fp16"), val = tensor([0x1p-3])]; + tensor attn_weights_25_cast_fp16 = mul(x = var_2111_cast_fp16, y = var_2105_to_fp16)[name = tensor("attn_weights_25_cast_fp16")]; + tensor var_2113 = logical_not(x = attn_mask_19)[name = tensor("op_2113")]; + tensor var_2114_to_fp16 = const()[name = tensor("op_2114_to_fp16"), val = tensor(-inf)]; + tensor attn_weights_27_cast_fp16 = select(a = var_2114_to_fp16, b = attn_weights_25_cast_fp16, cond = var_2113)[name = tensor("attn_weights_27_cast_fp16")]; + tensor var_2116 = const()[name = tensor("op_2116"), val = tensor(-1)]; + tensor attn_weights_29_cast_fp16 = softmax(axis = var_2116, x = attn_weights_27_cast_fp16)[name = tensor("attn_weights_29_cast_fp16")]; + tensor attn_output_9_transpose_x_0 = const()[name = tensor("attn_output_9_transpose_x_0"), val = tensor(false)]; + tensor attn_output_9_transpose_y_0 = const()[name = tensor("attn_output_9_transpose_y_0"), val = tensor(false)]; + tensor values_29_cast_fp16 = transpose(perm = var_2062, x = values_27_cast_fp16)[name = tensor("transpose_37")]; + tensor attn_output_9_cast_fp16 = matmul(transpose_x = attn_output_9_transpose_x_0, transpose_y = attn_output_9_transpose_y_0, x = attn_weights_29_cast_fp16, y = values_29_cast_fp16)[name = tensor("attn_output_9_cast_fp16")]; + tensor var_2124 = const()[name = tensor("op_2124"), val = tensor([0, 2, 1, 3])]; + tensor var_2127 = const()[name = tensor("op_2127"), val = tensor([1, 1, 1024])]; + tensor var_2125_cast_fp16 = transpose(perm = var_2124, x = attn_output_9_cast_fp16)[name = tensor("transpose_34")]; + tensor input_45_cast_fp16 = reshape(shape = var_2127, x = var_2125_cast_fp16)[name = tensor("input_45_cast_fp16")]; + tensor linear_18_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = attn4_out_proj_weight_cast_fp16, x = input_45_cast_fp16)[name = tensor("linear_18_cast_fp16")]; + tensor var_2133_to_fp16 = const()[name = tensor("op_2133_to_fp16"), val = tensor(0x1p+0)]; + tensor var_2134_cast_fp16 = add(x = position4_to_fp16, y = var_2133_to_fp16)[name = tensor("op_2134_cast_fp16")]; + tensor var_2134_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_2134_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor input_47_cast_fp16 = add(x = input_43_cast_fp16, y = linear_18_cast_fp16)[name = tensor("input_47_cast_fp16")]; + tensor input_49_axes_0 = const()[name = tensor("input_49_axes_0"), val = tensor([-1])]; + tensor norm4_2_weight_to_fp16 = const()[name = tensor("norm4_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(76790144)))]; + tensor norm4_2_bias_to_fp16 = const()[name = tensor("norm4_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(76792256)))]; + tensor var_2138_to_fp16 = const()[name = tensor("op_2138_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_49_cast_fp16 = layer_norm(axes = input_49_axes_0, beta = norm4_2_bias_to_fp16, epsilon = var_2138_to_fp16, gamma = norm4_2_weight_to_fp16, x = input_47_cast_fp16)[name = tensor("input_49_cast_fp16")]; + tensor linear_19_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = linear4_1_weight_cast_fp16, x = input_49_cast_fp16)[name = tensor("linear_19_cast_fp16")]; + tensor input_51_mode_0 = const()[name = tensor("input_51_mode_0"), val = tensor("EXACT")]; + tensor input_51_cast_fp16 = gelu(mode = input_51_mode_0, x = linear_19_cast_fp16)[name = tensor("input_51_cast_fp16")]; + tensor linear_20_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = linear4_2_weight_cast_fp16, x = input_51_cast_fp16)[name = tensor("linear_20_cast_fp16")]; + tensor input_53_cast_fp16 = add(x = input_47_cast_fp16, y = linear_20_cast_fp16)[name = tensor("input_53_cast_fp16")]; + tensor x_axes_0 = const()[name = tensor("x_axes_0"), val = tensor([-1])]; + tensor norm5_1_weight_to_fp16 = const()[name = tensor("norm5_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(76794368)))]; + tensor norm5_1_bias_to_fp16 = const()[name = tensor("norm5_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(76796480)))]; + tensor var_2155_to_fp16 = const()[name = tensor("op_2155_to_fp16"), val = tensor(0x1.5p-17)]; + tensor x_cast_fp16 = layer_norm(axes = x_axes_0, beta = norm5_1_bias_to_fp16, epsilon = var_2155_to_fp16, gamma = norm5_1_weight_to_fp16, x = input_53_cast_fp16)[name = tensor("x_cast_fp16")]; + tensor linear_21_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = attn5_in_proj_weight_cast_fp16, x = x_cast_fp16)[name = tensor("linear_21_cast_fp16")]; + tensor var_2191 = const()[name = tensor("op_2191"), val = tensor([1, 1, 3, 16, 64])]; + tensor qkv_cast_fp16 = reshape(shape = var_2191, x = linear_21_cast_fp16)[name = tensor("qkv_cast_fp16")]; + tensor q_31_begin_0 = const()[name = tensor("q_31_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor q_31_end_0 = const()[name = tensor("q_31_end_0"), val = tensor([1, 1, 1, 16, 64])]; + tensor q_31_end_mask_0 = const()[name = tensor("q_31_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor q_31_squeeze_mask_0 = const()[name = tensor("q_31_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor q_31_cast_fp16 = slice_by_index(begin = q_31_begin_0, end = q_31_end_0, end_mask = q_31_end_mask_0, squeeze_mask = q_31_squeeze_mask_0, x = qkv_cast_fp16)[name = tensor("q_31_cast_fp16")]; + tensor k_21_begin_0 = const()[name = tensor("k_21_begin_0"), val = tensor([0, 0, 1, 0, 0])]; + tensor k_21_end_0 = const()[name = tensor("k_21_end_0"), val = tensor([1, 1, 2, 16, 64])]; + tensor k_21_end_mask_0 = const()[name = tensor("k_21_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor k_21_squeeze_mask_0 = const()[name = tensor("k_21_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor k_21_cast_fp16 = slice_by_index(begin = k_21_begin_0, end = k_21_end_0, end_mask = k_21_end_mask_0, squeeze_mask = k_21_squeeze_mask_0, x = qkv_cast_fp16)[name = tensor("k_21_cast_fp16")]; + tensor v_begin_0 = const()[name = tensor("v_begin_0"), val = tensor([0, 0, 2, 0, 0])]; + tensor v_end_0 = const()[name = tensor("v_end_0"), val = tensor([1, 1, 3, 16, 64])]; + tensor v_end_mask_0 = const()[name = tensor("v_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor v_squeeze_mask_0 = const()[name = tensor("v_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor v_cast_fp16 = slice_by_index(begin = v_begin_0, end = v_end_0, end_mask = v_end_mask_0, squeeze_mask = v_squeeze_mask_0, x = qkv_cast_fp16)[name = tensor("v_cast_fp16")]; + tensor var_2295 = const()[name = tensor("op_2295"), val = tensor([1, 1, 1, 1])]; + tensor position5_to_fp16_dtype_0 = const()[name = tensor("position5_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor position5_to_fp16 = cast(dtype = position5_to_fp16_dtype_0, x = position5)[name = tensor("cast_114")]; + tensor ts_cast_fp16 = reshape(shape = var_2295, x = position5_to_fp16)[name = tensor("ts_cast_fp16")]; + tensor var_2299 = const()[name = tensor("op_2299"), val = tensor([1, 1, 16, 32, 2])]; + tensor q_complex_cast_fp16 = reshape(shape = var_2299, x = q_31_cast_fp16)[name = tensor("q_complex_cast_fp16")]; + tensor var_2303 = const()[name = tensor("op_2303"), val = tensor([1, 1, 16, 32, 2])]; + tensor k_complex_cast_fp16 = reshape(shape = var_2303, x = k_21_cast_fp16)[name = tensor("k_complex_cast_fp16")]; + tensor var_2307_begin_0 = const()[name = tensor("op_2307_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_2307_end_0 = const()[name = tensor("op_2307_end_0"), val = tensor([1, 1, 16, 32, 1])]; + tensor var_2307_end_mask_0 = const()[name = tensor("op_2307_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_2307_squeeze_mask_0 = const()[name = tensor("op_2307_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_2307_cast_fp16 = slice_by_index(begin = var_2307_begin_0, end = var_2307_end_0, end_mask = var_2307_end_mask_0, squeeze_mask = var_2307_squeeze_mask_0, x = q_complex_cast_fp16)[name = tensor("op_2307_cast_fp16")]; + tensor var_2315_begin_0 = const()[name = tensor("op_2315_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_2315_end_0 = const()[name = tensor("op_2315_end_0"), val = tensor([1, 1, 16, 32, 2])]; + tensor var_2315_end_mask_0 = const()[name = tensor("op_2315_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_2315_squeeze_mask_0 = const()[name = tensor("op_2315_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_2315_cast_fp16 = slice_by_index(begin = var_2315_begin_0, end = var_2315_end_0, end_mask = var_2315_end_mask_0, squeeze_mask = var_2315_squeeze_mask_0, x = q_complex_cast_fp16)[name = tensor("op_2315_cast_fp16")]; + tensor var_2323_begin_0 = const()[name = tensor("op_2323_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_2323_end_0 = const()[name = tensor("op_2323_end_0"), val = tensor([1, 1, 16, 32, 1])]; + tensor var_2323_end_mask_0 = const()[name = tensor("op_2323_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_2323_squeeze_mask_0 = const()[name = tensor("op_2323_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_2323_cast_fp16 = slice_by_index(begin = var_2323_begin_0, end = var_2323_end_0, end_mask = var_2323_end_mask_0, squeeze_mask = var_2323_squeeze_mask_0, x = k_complex_cast_fp16)[name = tensor("op_2323_cast_fp16")]; + tensor var_2331_begin_0 = const()[name = tensor("op_2331_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_2331_end_0 = const()[name = tensor("op_2331_end_0"), val = tensor([1, 1, 16, 32, 2])]; + tensor var_2331_end_mask_0 = const()[name = tensor("op_2331_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_2331_squeeze_mask_0 = const()[name = tensor("op_2331_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_2331_cast_fp16 = slice_by_index(begin = var_2331_begin_0, end = var_2331_end_0, end_mask = var_2331_end_mask_0, squeeze_mask = var_2331_squeeze_mask_0, x = k_complex_cast_fp16)[name = tensor("op_2331_cast_fp16")]; + tensor freqs_to_fp16 = const()[name = tensor("freqs_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(76798592)))]; + tensor var_2337_cast_fp16 = mul(x = freqs_to_fp16, y = ts_cast_fp16)[name = tensor("op_2337_cast_fp16")]; + tensor rotr_cast_fp16 = cos(x = var_2337_cast_fp16)[name = tensor("rotr_cast_fp16")]; + tensor roti_cast_fp16 = sin(x = var_2337_cast_fp16)[name = tensor("roti_cast_fp16")]; + tensor var_2341_cast_fp16 = mul(x = var_2307_cast_fp16, y = rotr_cast_fp16)[name = tensor("op_2341_cast_fp16")]; + tensor var_2342_cast_fp16 = mul(x = var_2315_cast_fp16, y = roti_cast_fp16)[name = tensor("op_2342_cast_fp16")]; + tensor qor_21_cast_fp16 = sub(x = var_2341_cast_fp16, y = var_2342_cast_fp16)[name = tensor("qor_21_cast_fp16")]; + tensor var_2345_cast_fp16 = mul(x = var_2307_cast_fp16, y = roti_cast_fp16)[name = tensor("op_2345_cast_fp16")]; + tensor var_2346_cast_fp16 = mul(x = var_2315_cast_fp16, y = rotr_cast_fp16)[name = tensor("op_2346_cast_fp16")]; + tensor qoi_21_cast_fp16 = add(x = var_2345_cast_fp16, y = var_2346_cast_fp16)[name = tensor("qoi_21_cast_fp16")]; + tensor var_2349_cast_fp16 = mul(x = var_2323_cast_fp16, y = rotr_cast_fp16)[name = tensor("op_2349_cast_fp16")]; + tensor var_2350_cast_fp16 = mul(x = var_2331_cast_fp16, y = roti_cast_fp16)[name = tensor("op_2350_cast_fp16")]; + tensor kor_21_cast_fp16 = sub(x = var_2349_cast_fp16, y = var_2350_cast_fp16)[name = tensor("kor_21_cast_fp16")]; + tensor var_2353_cast_fp16 = mul(x = var_2323_cast_fp16, y = roti_cast_fp16)[name = tensor("op_2353_cast_fp16")]; + tensor var_2354_cast_fp16 = mul(x = var_2331_cast_fp16, y = rotr_cast_fp16)[name = tensor("op_2354_cast_fp16")]; + tensor koi_21_cast_fp16 = add(x = var_2353_cast_fp16, y = var_2354_cast_fp16)[name = tensor("koi_21_cast_fp16")]; + tensor qo_axis_0 = const()[name = tensor("qo_axis_0"), val = tensor(-1)]; + tensor qo_cast_fp16 = stack(axis = qo_axis_0, values = (qor_21_cast_fp16, qoi_21_cast_fp16))[name = tensor("qo_cast_fp16")]; + tensor ko_axis_0 = const()[name = tensor("ko_axis_0"), val = tensor(-1)]; + tensor ko_cast_fp16 = stack(axis = ko_axis_0, values = (kor_21_cast_fp16, koi_21_cast_fp16))[name = tensor("ko_cast_fp16")]; + tensor var_2383 = const()[name = tensor("op_2383"), val = tensor([1, 1, 16, 64])]; + tensor q_33_cast_fp16 = reshape(shape = var_2383, x = qo_cast_fp16)[name = tensor("q_33_cast_fp16")]; + tensor var_2385 = const()[name = tensor("op_2385"), val = tensor([1, 1, 16, 64])]; + tensor k_cast_fp16 = reshape(shape = var_2385, x = ko_cast_fp16)[name = tensor("k_cast_fp16")]; + tensor _inversed_2407_y_0_to_fp16 = const()[name = tensor("_inversed_2407_y_0_to_fp16"), val = tensor(0x1p-9)]; + tensor _inversed_2407_cast_fp16 = mul(x = ts_cast_fp16, y = _inversed_2407_y_0_to_fp16)[name = tensor("_inversed_2407_cast_fp16")]; + tensor var_2408_cast_fp16 = floor(x = _inversed_2407_cast_fp16)[name = tensor("op_2408_cast_fp16")]; + tensor var_2409_to_fp16 = const()[name = tensor("op_2409_to_fp16"), val = tensor(0x1p+9)]; + tensor var_2410_cast_fp16 = mul(x = var_2408_cast_fp16, y = var_2409_to_fp16)[name = tensor("op_2410_cast_fp16")]; + tensor write_indices_float_cast_fp16 = sub(x = ts_cast_fp16, y = var_2410_cast_fp16)[name = tensor("write_indices_float_cast_fp16")]; + tensor var_2417_dtype_0 = const()[name = tensor("op_2417_dtype_0"), val = tensor("int32")]; + tensor write_indices_reps_0 = const()[name = tensor("write_indices_reps_0"), val = tensor([1, 1, 16, 64])]; + tensor write_indices_float_cast_fp16_to_int32 = cast(dtype = var_2417_dtype_0, x = write_indices_float_cast_fp16)[name = tensor("cast_113")]; + tensor write_indices = tile(reps = write_indices_reps_0, x = write_indices_float_cast_fp16_to_int32)[name = tensor("write_indices")]; + tensor var_2425_begin_0 = const()[name = tensor("op_2425_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_2425_end_0 = const()[name = tensor("op_2425_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor var_2425_end_mask_0 = const()[name = tensor("op_2425_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_2425_squeeze_mask_0 = const()[name = tensor("op_2425_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor cache5_to_fp16_dtype_0 = const()[name = tensor("cache5_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor cache5_to_fp16 = cast(dtype = cache5_to_fp16_dtype_0, x = cache5)[name = tensor("cast_112")]; + tensor var_2425_cast_fp16 = slice_by_index(begin = var_2425_begin_0, end = var_2425_end_0, end_mask = var_2425_end_mask_0, squeeze_mask = var_2425_squeeze_mask_0, x = cache5_to_fp16)[name = tensor("op_2425_cast_fp16")]; + tensor var_2427_axis_0 = const()[name = tensor("op_2427_axis_0"), val = tensor(1)]; + tensor var_2427_mode_0 = const()[name = tensor("op_2427_mode_0"), val = tensor("update")]; + tensor var_2427_validate_indices_0 = const()[name = tensor("op_2427_validate_indices_0"), val = tensor(false)]; + tensor var_2427_cast_fp16 = scatter_along_axis(axis = var_2427_axis_0, data = var_2425_cast_fp16, indices = write_indices, mode = var_2427_mode_0, updates = k_cast_fp16, validate_indices = var_2427_validate_indices_0)[name = tensor("op_2427_cast_fp16")]; + tensor concat_37 = const()[name = tensor("concat_37"), val = tensor([0, 0, 0, 0, 0])]; + tensor concat_38 = const()[name = tensor("concat_38"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_internal_tensor_assign_1_stride_0 = const()[name = tensor("new_cache_internal_tensor_assign_1_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_internal_tensor_assign_1_begin_mask_0 = const()[name = tensor("new_cache_internal_tensor_assign_1_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_internal_tensor_assign_1_end_mask_0 = const()[name = tensor("new_cache_internal_tensor_assign_1_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_internal_tensor_assign_1_squeeze_mask_0 = const()[name = tensor("new_cache_internal_tensor_assign_1_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_22 = const()[name = tensor("shape_22"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_10 = const()[name = tensor("reduce_prod_10"), val = tensor(1048576)]; + tensor range_1d_10_start_0 = const()[name = tensor("range_1d_10_start_0"), val = tensor(0)]; + tensor range_1d_10_step_0 = const()[name = tensor("range_1d_10_step_0"), val = tensor(1)]; + tensor range_1d_10 = range_1d(end = reduce_prod_10, start = range_1d_10_start_0, step = range_1d_10_step_0)[name = tensor("range_1d_10")]; + tensor reshape_50 = reshape(shape = shape_22, x = range_1d_10)[name = tensor("reshape_50")]; + tensor slice_by_index_10 = slice_by_index(begin = concat_37, begin_mask = new_cache_internal_tensor_assign_1_begin_mask_0, end = concat_38, end_mask = new_cache_internal_tensor_assign_1_end_mask_0, squeeze_mask = new_cache_internal_tensor_assign_1_squeeze_mask_0, stride = new_cache_internal_tensor_assign_1_stride_0, x = reshape_50)[name = tensor("slice_by_index_10")]; + tensor reshape_51_shape_0 = const()[name = tensor("reshape_51_shape_0"), val = tensor([-1])]; + tensor reshape_51 = reshape(shape = reshape_51_shape_0, x = slice_by_index_10)[name = tensor("reshape_51")]; + tensor reshape_52_shape_0 = const()[name = tensor("reshape_52_shape_0"), val = tensor([-1])]; + tensor reshape_52_cast_fp16 = reshape(shape = reshape_52_shape_0, x = var_2427_cast_fp16)[name = tensor("reshape_52_cast_fp16")]; + tensor reshape_53_shape_0 = const()[name = tensor("reshape_53_shape_0"), val = tensor([-1])]; + tensor reshape_53_cast_fp16 = reshape(shape = reshape_53_shape_0, x = cache5_to_fp16)[name = tensor("reshape_53_cast_fp16")]; + tensor scatter_10_mode_0 = const()[name = tensor("scatter_10_mode_0"), val = tensor("update")]; + tensor scatter_10_axis_0 = const()[name = tensor("scatter_10_axis_0"), val = tensor(0)]; + tensor scatter_10_validate_indices_0 = const()[name = tensor("scatter_10_validate_indices_0"), val = tensor(false)]; + tensor scatter_10_cast_fp16 = scatter(axis = scatter_10_axis_0, data = reshape_53_cast_fp16, indices = reshape_51, mode = scatter_10_mode_0, updates = reshape_52_cast_fp16, validate_indices = scatter_10_validate_indices_0)[name = tensor("scatter_10_cast_fp16")]; + tensor reshape_54_cast_fp16 = reshape(shape = shape_22, x = scatter_10_cast_fp16)[name = tensor("reshape_54_cast_fp16")]; + tensor var_2435_begin_0 = const()[name = tensor("op_2435_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor var_2435_end_0 = const()[name = tensor("op_2435_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor var_2435_end_mask_0 = const()[name = tensor("op_2435_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_2435_squeeze_mask_0 = const()[name = tensor("op_2435_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor var_2435_cast_fp16 = slice_by_index(begin = var_2435_begin_0, end = var_2435_end_0, end_mask = var_2435_end_mask_0, squeeze_mask = var_2435_squeeze_mask_0, x = reshape_54_cast_fp16)[name = tensor("op_2435_cast_fp16")]; + tensor var_2437_axis_0 = const()[name = tensor("op_2437_axis_0"), val = tensor(1)]; + tensor var_2437_mode_0 = const()[name = tensor("op_2437_mode_0"), val = tensor("update")]; + tensor var_2437_validate_indices_0 = const()[name = tensor("op_2437_validate_indices_0"), val = tensor(false)]; + tensor var_2437_cast_fp16 = scatter_along_axis(axis = var_2437_axis_0, data = var_2435_cast_fp16, indices = write_indices, mode = var_2437_mode_0, updates = v_cast_fp16, validate_indices = var_2437_validate_indices_0)[name = tensor("op_2437_cast_fp16")]; + tensor concat_39 = const()[name = tensor("concat_39"), val = tensor([1, 0, 0, 0, 0])]; + tensor concat_40 = const()[name = tensor("concat_40"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_internal_tensor_assign_2_stride_0 = const()[name = tensor("new_cache_internal_tensor_assign_2_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_internal_tensor_assign_2_begin_mask_0 = const()[name = tensor("new_cache_internal_tensor_assign_2_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_internal_tensor_assign_2_end_mask_0 = const()[name = tensor("new_cache_internal_tensor_assign_2_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_internal_tensor_assign_2_squeeze_mask_0 = const()[name = tensor("new_cache_internal_tensor_assign_2_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_23 = const()[name = tensor("shape_23"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_11 = const()[name = tensor("reduce_prod_11"), val = tensor(1048576)]; + tensor range_1d_11_start_0 = const()[name = tensor("range_1d_11_start_0"), val = tensor(0)]; + tensor range_1d_11_step_0 = const()[name = tensor("range_1d_11_step_0"), val = tensor(1)]; + tensor range_1d_11 = range_1d(end = reduce_prod_11, start = range_1d_11_start_0, step = range_1d_11_step_0)[name = tensor("range_1d_11")]; + tensor reshape_55 = reshape(shape = shape_23, x = range_1d_11)[name = tensor("reshape_55")]; + tensor slice_by_index_11 = slice_by_index(begin = concat_39, begin_mask = new_cache_internal_tensor_assign_2_begin_mask_0, end = concat_40, end_mask = new_cache_internal_tensor_assign_2_end_mask_0, squeeze_mask = new_cache_internal_tensor_assign_2_squeeze_mask_0, stride = new_cache_internal_tensor_assign_2_stride_0, x = reshape_55)[name = tensor("slice_by_index_11")]; + tensor reshape_56_shape_0 = const()[name = tensor("reshape_56_shape_0"), val = tensor([-1])]; + tensor reshape_56 = reshape(shape = reshape_56_shape_0, x = slice_by_index_11)[name = tensor("reshape_56")]; + tensor reshape_57_shape_0 = const()[name = tensor("reshape_57_shape_0"), val = tensor([-1])]; + tensor reshape_57_cast_fp16 = reshape(shape = reshape_57_shape_0, x = var_2437_cast_fp16)[name = tensor("reshape_57_cast_fp16")]; + tensor reshape_58_shape_0 = const()[name = tensor("reshape_58_shape_0"), val = tensor([-1])]; + tensor reshape_58_cast_fp16 = reshape(shape = reshape_58_shape_0, x = reshape_54_cast_fp16)[name = tensor("reshape_58_cast_fp16")]; + tensor scatter_11_mode_0 = const()[name = tensor("scatter_11_mode_0"), val = tensor("update")]; + tensor scatter_11_axis_0 = const()[name = tensor("scatter_11_axis_0"), val = tensor(0)]; + tensor scatter_11_validate_indices_0 = const()[name = tensor("scatter_11_validate_indices_0"), val = tensor(false)]; + tensor scatter_11_cast_fp16 = scatter(axis = scatter_11_axis_0, data = reshape_58_cast_fp16, indices = reshape_56, mode = scatter_11_mode_0, updates = reshape_57_cast_fp16, validate_indices = scatter_11_validate_indices_0)[name = tensor("scatter_11_cast_fp16")]; + tensor reshape_59_cast_fp16 = reshape(shape = shape_23, x = scatter_11_cast_fp16)[name = tensor("reshape_59_cast_fp16")]; + tensor reshape_59_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("reshape_59_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor keys_31_begin_0 = const()[name = tensor("keys_31_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor keys_31_end_0 = const()[name = tensor("keys_31_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor keys_31_end_mask_0 = const()[name = tensor("keys_31_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor keys_31_squeeze_mask_0 = const()[name = tensor("keys_31_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor keys_31_cast_fp16 = slice_by_index(begin = keys_31_begin_0, end = keys_31_end_0, end_mask = keys_31_end_mask_0, squeeze_mask = keys_31_squeeze_mask_0, x = reshape_59_cast_fp16)[name = tensor("keys_31_cast_fp16")]; + tensor values_31_begin_0 = const()[name = tensor("values_31_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor values_31_end_0 = const()[name = tensor("values_31_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor values_31_end_mask_0 = const()[name = tensor("values_31_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor values_31_squeeze_mask_0 = const()[name = tensor("values_31_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor values_31_cast_fp16 = slice_by_index(begin = values_31_begin_0, end = values_31_end_0, end_mask = values_31_end_mask_0, squeeze_mask = values_31_squeeze_mask_0, x = reshape_59_cast_fp16)[name = tensor("values_31_cast_fp16")]; + tensor var_2449_cast_fp16 = not_equal(x = keys_31_cast_fp16, y = keys_31_cast_fp16)[name = tensor("op_2449_cast_fp16")]; + tensor keys_33_cast_fp16 = select(a = var_360_to_fp16, b = keys_31_cast_fp16, cond = var_2449_cast_fp16)[name = tensor("keys_33_cast_fp16")]; + tensor var_2457_cast_fp16 = not_equal(x = values_31_cast_fp16, y = values_31_cast_fp16)[name = tensor("op_2457_cast_fp16")]; + tensor values_33_cast_fp16 = select(a = var_360_to_fp16, b = values_31_cast_fp16, cond = var_2457_cast_fp16)[name = tensor("values_33_cast_fp16")]; + tensor var_2481 = const()[name = tensor("op_2481"), val = tensor([0, 2, 1, 3])]; + tensor var_2494 = const()[name = tensor("op_2494"), val = tensor([1, 1, 1])]; + tensor var_2495_cast_fp16 = reshape(shape = var_2494, x = position5_to_fp16)[name = tensor("op_2495_cast_fp16")]; + tensor var_2512_to_fp16 = const()[name = tensor("op_2512_to_fp16"), val = tensor(0x1p+0)]; + tensor valid_len_cast_fp16 = add(x = var_2495_cast_fp16, y = var_2512_to_fp16)[name = tensor("valid_len_cast_fp16")]; + tensor valid_mask_cast_fp16 = less(x = k_positions_1_promoted_to_fp16, y = valid_len_cast_fp16)[name = tensor("valid_mask_cast_fp16")]; + tensor causal_mask_cast_fp16 = less_equal(x = k_positions_1_promoted_to_fp16, y = var_2495_cast_fp16)[name = tensor("causal_mask_cast_fp16")]; + tensor attn_mask_21 = logical_and(x = valid_mask_cast_fp16, y = causal_mask_cast_fp16)[name = tensor("attn_mask_21")]; + tensor attn_mask_axes_0 = const()[name = tensor("attn_mask_axes_0"), val = tensor([1])]; + tensor attn_mask = expand_dims(axes = attn_mask_axes_0, x = attn_mask_21)[name = tensor("attn_mask")]; + tensor var_2530_transpose_x_0 = const()[name = tensor("op_2530_transpose_x_0"), val = tensor(false)]; + tensor var_2530_transpose_y_0 = const()[name = tensor("op_2530_transpose_y_0"), val = tensor(false)]; + tensor transpose_28_perm_0 = const()[name = tensor("transpose_28_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_29_perm_0 = const()[name = tensor("transpose_29_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_29 = transpose(perm = transpose_29_perm_0, x = keys_33_cast_fp16)[name = tensor("transpose_31")]; + tensor transpose_28 = transpose(perm = transpose_28_perm_0, x = q_33_cast_fp16)[name = tensor("transpose_32")]; + tensor var_2530_cast_fp16 = matmul(transpose_x = var_2530_transpose_x_0, transpose_y = var_2530_transpose_y_0, x = transpose_28, y = transpose_29)[name = tensor("op_2530_cast_fp16")]; + tensor var_2524_to_fp16 = const()[name = tensor("op_2524_to_fp16"), val = tensor([0x1p-3])]; + tensor attn_weights_31_cast_fp16 = mul(x = var_2530_cast_fp16, y = var_2524_to_fp16)[name = tensor("attn_weights_31_cast_fp16")]; + tensor var_2532 = logical_not(x = attn_mask)[name = tensor("op_2532")]; + tensor var_2533_to_fp16 = const()[name = tensor("op_2533_to_fp16"), val = tensor(-inf)]; + tensor attn_weights_33_cast_fp16 = select(a = var_2533_to_fp16, b = attn_weights_31_cast_fp16, cond = var_2532)[name = tensor("attn_weights_33_cast_fp16")]; + tensor var_2535 = const()[name = tensor("op_2535"), val = tensor(-1)]; + tensor attn_weights_cast_fp16 = softmax(axis = var_2535, x = attn_weights_33_cast_fp16)[name = tensor("attn_weights_cast_fp16")]; + tensor attn_output_transpose_x_0 = const()[name = tensor("attn_output_transpose_x_0"), val = tensor(false)]; + tensor attn_output_transpose_y_0 = const()[name = tensor("attn_output_transpose_y_0"), val = tensor(false)]; + tensor values_cast_fp16 = transpose(perm = var_2481, x = values_33_cast_fp16)[name = tensor("transpose_33")]; + tensor attn_output_cast_fp16 = matmul(transpose_x = attn_output_transpose_x_0, transpose_y = attn_output_transpose_y_0, x = attn_weights_cast_fp16, y = values_cast_fp16)[name = tensor("attn_output_cast_fp16")]; + tensor var_2543 = const()[name = tensor("op_2543"), val = tensor([0, 2, 1, 3])]; + tensor var_2546 = const()[name = tensor("op_2546"), val = tensor([1, 1, 1024])]; + tensor var_2544_cast_fp16 = transpose(perm = var_2543, x = attn_output_cast_fp16)[name = tensor("transpose_30")]; + tensor input_55_cast_fp16 = reshape(shape = var_2546, x = var_2544_cast_fp16)[name = tensor("input_55_cast_fp16")]; + tensor linear_22_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = attn5_out_proj_weight_cast_fp16, x = input_55_cast_fp16)[name = tensor("linear_22_cast_fp16")]; + tensor var_2552_to_fp16 = const()[name = tensor("op_2552_to_fp16"), val = tensor(0x1p+0)]; + tensor var_2553_cast_fp16 = add(x = position5_to_fp16, y = var_2552_to_fp16)[name = tensor("op_2553_cast_fp16")]; + tensor var_2553_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_2553_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor input_57_cast_fp16 = add(x = input_53_cast_fp16, y = linear_22_cast_fp16)[name = tensor("input_57_cast_fp16")]; + tensor input_59_axes_0 = const()[name = tensor("input_59_axes_0"), val = tensor([-1])]; + tensor norm5_2_weight_to_fp16 = const()[name = tensor("norm5_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(76798720)))]; + tensor norm5_2_bias_to_fp16 = const()[name = tensor("norm5_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(76800832)))]; + tensor var_2557_to_fp16 = const()[name = tensor("op_2557_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_59_cast_fp16 = layer_norm(axes = input_59_axes_0, beta = norm5_2_bias_to_fp16, epsilon = var_2557_to_fp16, gamma = norm5_2_weight_to_fp16, x = input_57_cast_fp16)[name = tensor("input_59_cast_fp16")]; + tensor linear_23_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = linear5_1_weight_cast_fp16, x = input_59_cast_fp16)[name = tensor("linear_23_cast_fp16")]; + tensor input_61_mode_0 = const()[name = tensor("input_61_mode_0"), val = tensor("EXACT")]; + tensor input_61_cast_fp16 = gelu(mode = input_61_mode_0, x = linear_23_cast_fp16)[name = tensor("input_61_cast_fp16")]; + tensor linear_24_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = linear5_2_weight_cast_fp16, x = input_61_cast_fp16)[name = tensor("linear_24_cast_fp16")]; + tensor input_63_cast_fp16 = add(x = input_57_cast_fp16, y = linear_24_cast_fp16)[name = tensor("input_63_cast_fp16")]; + tensor input_axes_0 = const()[name = tensor("input_axes_0"), val = tensor([-1])]; + tensor out_norm_weight_to_fp16 = const()[name = tensor("out_norm_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(76802944)))]; + tensor out_norm_bias_to_fp16 = const()[name = tensor("out_norm_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(76805056)))]; + tensor var_2574_to_fp16 = const()[name = tensor("op_2574_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_cast_fp16 = layer_norm(axes = input_axes_0, beta = out_norm_bias_to_fp16, epsilon = var_2574_to_fp16, gamma = out_norm_weight_to_fp16, x = input_63_cast_fp16)[name = tensor("input_cast_fp16")]; + tensor input_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("input_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor out_eos_weight_to_fp16 = const()[name = tensor("out_eos_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(76807168)))]; + tensor out_eos_bias_to_fp16 = const()[name = tensor("out_eos_bias_to_fp16"), val = tensor([-0x1.5ap-3])]; + tensor linear_25_cast_fp16 = linear(bias = out_eos_bias_to_fp16, weight = out_eos_weight_to_fp16, x = input_cast_fp16)[name = tensor("linear_25_cast_fp16")]; + tensor linear_25_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("linear_25_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor var_2582 = cast(dtype = linear_25_cast_fp16_to_fp32_dtype_0, x = linear_25_cast_fp16)[name = tensor("cast_108")]; + tensor input = cast(dtype = input_cast_fp16_to_fp32_dtype_0, x = input_cast_fp16)[name = tensor("cast_109")]; + tensor var_2553 = cast(dtype = var_2553_cast_fp16_to_fp32_dtype_0, x = var_2553_cast_fp16)[name = tensor("cast_110")]; + tensor new_cache_internal_tensor_assign_2 = cast(dtype = reshape_59_cast_fp16_to_fp32_dtype_0, x = reshape_59_cast_fp16)[name = tensor("cast_111")]; + tensor var_2134 = cast(dtype = var_2134_cast_fp16_to_fp32_dtype_0, x = var_2134_cast_fp16)[name = tensor("cast_115")]; + tensor new_cache_9_internal_tensor_assign_2 = cast(dtype = reshape_49_cast_fp16_to_fp32_dtype_0, x = reshape_49_cast_fp16)[name = tensor("cast_116")]; + tensor var_1715 = cast(dtype = var_1715_cast_fp16_to_fp32_dtype_0, x = var_1715_cast_fp16)[name = tensor("cast_120")]; + tensor new_cache_7_internal_tensor_assign_2 = cast(dtype = reshape_39_cast_fp16_to_fp32_dtype_0, x = reshape_39_cast_fp16)[name = tensor("cast_121")]; + tensor var_1296 = cast(dtype = var_1296_cast_fp16_to_fp32_dtype_0, x = var_1296_cast_fp16)[name = tensor("cast_125")]; + tensor new_cache_5_internal_tensor_assign_2 = cast(dtype = reshape_29_cast_fp16_to_fp32_dtype_0, x = reshape_29_cast_fp16)[name = tensor("cast_126")]; + tensor var_877 = cast(dtype = var_877_cast_fp16_to_fp32_dtype_0, x = var_877_cast_fp16)[name = tensor("cast_130")]; + tensor new_cache_3_internal_tensor_assign_2 = cast(dtype = reshape_19_cast_fp16_to_fp32_dtype_0, x = reshape_19_cast_fp16)[name = tensor("cast_131")]; + tensor var_458 = cast(dtype = var_458_cast_fp16_to_fp32_dtype_0, x = var_458_cast_fp16)[name = tensor("cast_135")]; + tensor new_cache_1_internal_tensor_assign_2 = cast(dtype = reshape_9_cast_fp16_to_fp32_dtype_0, x = reshape_9_cast_fp16)[name = tensor("cast_136")]; + } -> (input, var_2582, new_cache_1_internal_tensor_assign_2, var_458, new_cache_3_internal_tensor_assign_2, var_877, new_cache_5_internal_tensor_assign_2, var_1296, new_cache_7_internal_tensor_assign_2, var_1715, new_cache_9_internal_tensor_assign_2, var_2134, new_cache_internal_tensor_assign_2, var_2553); +} \ No newline at end of file diff --git a/v2.1/portuguese/flowlm_stepv2.mlmodelc/weights/weight.bin b/v2.1/portuguese/flowlm_stepv2.mlmodelc/weights/weight.bin new file mode 100644 index 0000000000000000000000000000000000000000..a28ad7d2c9506b38c7a3b4174789a790787d3b92 --- /dev/null +++ b/v2.1/portuguese/flowlm_stepv2.mlmodelc/weights/weight.bin @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:83248f033cf713da0b80824ce160b07b7f3697553086e5b02ee667313d8e1057 +size 76809280 diff --git a/v2.1/portuguese/flowlm_stepv2.mlpackage/Data/com.apple.CoreML/model.mlmodel b/v2.1/portuguese/flowlm_stepv2.mlpackage/Data/com.apple.CoreML/model.mlmodel new file mode 100644 index 0000000000000000000000000000000000000000..43b4e3346eaf917a52f7d1c2a551951b177a0233 --- /dev/null +++ b/v2.1/portuguese/flowlm_stepv2.mlpackage/Data/com.apple.CoreML/model.mlmodel @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:2e4fd8b072b31432cebbb87c861e5af35c377eabde7f36cc541a04bbf2ba20c4 +size 212305 diff --git a/v2.1/portuguese/flowlm_stepv2.mlpackage/Data/com.apple.CoreML/weights/weight.bin b/v2.1/portuguese/flowlm_stepv2.mlpackage/Data/com.apple.CoreML/weights/weight.bin new file mode 100644 index 0000000000000000000000000000000000000000..a28ad7d2c9506b38c7a3b4174789a790787d3b92 --- /dev/null +++ b/v2.1/portuguese/flowlm_stepv2.mlpackage/Data/com.apple.CoreML/weights/weight.bin @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:83248f033cf713da0b80824ce160b07b7f3697553086e5b02ee667313d8e1057 +size 76809280 diff --git a/v2.1/portuguese/flowlm_stepv2.mlpackage/Manifest.json b/v2.1/portuguese/flowlm_stepv2.mlpackage/Manifest.json new file mode 100644 index 0000000000000000000000000000000000000000..8d3751f69569d1614ea2ec7436a456f1a314a3b8 --- /dev/null +++ b/v2.1/portuguese/flowlm_stepv2.mlpackage/Manifest.json @@ -0,0 +1,18 @@ +{ + "fileFormatVersion": "1.0.0", + "itemInfoEntries": { + "BBFFA944-5652-4AB8-85A3-FDD426441F17": { + "author": "com.apple.CoreML", + "description": "CoreML Model Specification", + "name": "model.mlmodel", + "path": "com.apple.CoreML/model.mlmodel" + }, + "E4A8B911-863F-4E2F-AAA2-44A2CFABD6D4": { + "author": "com.apple.CoreML", + "description": "CoreML Model Weights", + "name": "weights", + "path": "com.apple.CoreML/weights" + } + }, + "rootModelIdentifier": "BBFFA944-5652-4AB8-85A3-FDD426441F17" +} diff --git a/v2.1/portuguese/manifest.json b/v2.1/portuguese/manifest.json new file mode 100644 index 0000000000000000000000000000000000000000..fc7da98920126d1261a9c47517b1303db1599db7 --- /dev/null +++ b/v2.1/portuguese/manifest.json @@ -0,0 +1,27 @@ +{ + "version": "2.1", + "base": "v2", + "kind": "optimization-reconvert", + "self_contained": true, + "note": "Same weights as v2. Re-converted for speed: fused flow decoder (100% ANE), one-shot cond prefill, fp16 flowlm. NOT a finetune.", + "new": [ + "flow_decoder_fused", + "cond_prefill", + "flowlm_step" + ], + "copied_from_v2": [ + "flowlm_stepv2.mlpackage", + "flowlm_stepv2.mlmodelc", + "mimi_decoder.mlpackage", + "mimi_decoder.mlmodelc", + "constants", + "constants_bin" + ], + "compute_units": { + "flow_decoder_fused": "all (ANE)", + "cond_prefill": "all (GPU)", + "flowlm_step": "all (GPU, fp16)", + "flowlm_stepv2": "cpuAndGpu (GPU, int8 \u2014 fastest flowlm)", + "mimi_decoder": "cpuOnly" + } +} diff --git a/v2.1/portuguese/mimi_decoder.mlmodelc/analytics/coremldata.bin b/v2.1/portuguese/mimi_decoder.mlmodelc/analytics/coremldata.bin new file mode 100644 index 0000000000000000000000000000000000000000..42e8b8e486f3203ef7ebb2cc5037f37a6847c8a8 --- /dev/null +++ b/v2.1/portuguese/mimi_decoder.mlmodelc/analytics/coremldata.bin @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:c6f4eecd44ba2849ab8893a6b9fbf1c289bbaaa7485a784212139f4d440477fa +size 243 diff --git a/v2.1/portuguese/mimi_decoder.mlmodelc/coremldata.bin b/v2.1/portuguese/mimi_decoder.mlmodelc/coremldata.bin new file mode 100644 index 0000000000000000000000000000000000000000..0e526f7f201e465634f29831a885765281f7a8ba --- /dev/null +++ b/v2.1/portuguese/mimi_decoder.mlmodelc/coremldata.bin @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:4fb4d46e1e9f23979228f63f7cd90ce3092f945661d74f2cb2b81bc616716094 +size 1777 diff --git a/v2.1/portuguese/mimi_decoder.mlmodelc/metadata.json b/v2.1/portuguese/mimi_decoder.mlmodelc/metadata.json new file mode 100644 index 0000000000000000000000000000000000000000..f66f9e7ff5be2ffbc44c8300f3501f67a77afaf6 --- /dev/null +++ b/v2.1/portuguese/mimi_decoder.mlmodelc/metadata.json @@ -0,0 +1,573 @@ +[ + { + "metadataOutputVersion" : "3.0", + "storagePrecision" : "Float32", + "outputSchema" : [ + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1 × 1 × 1920)", + "shortDescription" : "", + "shape" : "[1, 1, 1920]", + "name" : "var_724", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1 × 512 × 16)", + "shortDescription" : "", + "shape" : "[1, 512, 16]", + "name" : "var_77", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 2 × 1 × 256 × 8 × 64)", + "shortDescription" : "", + "shape" : "[2, 1, 256, 8, 64]", + "name" : "var_210", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1)", + "shortDescription" : "", + "shape" : "[1]", + "name" : "var_741", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 2 × 1 × 256 × 8 × 64)", + "shortDescription" : "", + "shape" : "[2, 1, 256, 8, 64]", + "name" : "var_400", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1)", + "shortDescription" : "", + "shape" : "[1]", + "name" : "var_744", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1 × 512 × 6)", + "shortDescription" : "", + "shape" : "[1, 512, 6]", + "name" : "var_542", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1)", + "shortDescription" : "", + "shape" : "[1]", + "name" : "conv0_first", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1 × 256 × 6)", + "shortDescription" : "", + "shape" : "[1, 256, 6]", + "name" : "var_565", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1 × 256 × 2)", + "shortDescription" : "", + "shape" : "[1, 256, 2]", + "name" : "var_585", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1)", + "shortDescription" : "", + "shape" : "[1]", + "name" : "res0_conv0_first", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1 × 128 × 0)", + "shortDescription" : "", + "shape" : "[1, 128, 0]", + "name" : "res0_conv1_prev", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1)", + "shortDescription" : "", + "shape" : "[1]", + "name" : "res0_conv1_first", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1 × 128 × 5)", + "shortDescription" : "", + "shape" : "[1, 128, 5]", + "name" : "var_621", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1 × 128 × 2)", + "shortDescription" : "", + "shape" : "[1, 128, 2]", + "name" : "var_641", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1)", + "shortDescription" : "", + "shape" : "[1]", + "name" : "res1_conv0_first", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1 × 64 × 0)", + "shortDescription" : "", + "shape" : "[1, 64, 0]", + "name" : "res1_conv1_prev", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1)", + "shortDescription" : "", + "shape" : "[1]", + "name" : "res1_conv1_first", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1 × 64 × 4)", + "shortDescription" : "", + "shape" : "[1, 64, 4]", + "name" : "var_677", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1 × 64 × 2)", + "shortDescription" : "", + "shape" : "[1, 64, 2]", + "name" : "var_697", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1)", + "shortDescription" : "", + "shape" : "[1]", + "name" : "res2_conv0_first", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1 × 32 × 0)", + "shortDescription" : "", + "shape" : "[1, 32, 0]", + "name" : "res2_conv1_prev", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1)", + "shortDescription" : "", + "shape" : "[1]", + "name" : "res2_conv1_first", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1 × 64 × 2)", + "shortDescription" : "", + "shape" : "[1, 64, 2]", + "name" : "var_725", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1)", + "shortDescription" : "", + "shape" : "[1]", + "name" : "conv_final_first", + "type" : "MultiArray" + } + ], + "modelParameters" : [ + + ], + "specificationVersion" : 8, + "mlProgramOperationTypeHistogram" : { + "Ios17.notEqual" : 4, + "Tile" : 2, + "Ios17.lessEqual" : 2, + "Ios17.layerNorm" : 4, + "Stack" : 6, + "Ios17.greaterEqual" : 4, + "Ios17.scatterAlongAxis" : 4, + "Ios17.sub" : 17, + "Identity" : 11, + "Ios17.elu" : 10, + "Ios17.logicalNot" : 2, + "Ios17.sliceByIndex" : 35, + "Ios17.concat" : 9, + "Ios17.convTranspose" : 4, + "Ios17.floorDiv" : 4, + "Ios17.expandDims" : 7, + "Ios17.less" : 2, + "Ios17.mul" : 29, + "Ios17.transpose" : 10, + "Ios17.matmul" : 4, + "Ios16.gelu" : 2, + "Ios17.conv" : 9, + "Ios17.reshape" : 22, + "Split" : 2, + "Ios17.squeeze" : 6, + "Ios17.cast" : 6, + "Ios17.sin" : 2, + "Ios17.linear" : 8, + "Select" : 6, + "Ios16.softmax" : 2, + "Ios17.add" : 28, + "Ios17.cos" : 2, + "Ios17.logicalAnd" : 6 + }, + "computePrecision" : "Mixed (Float32, Int32)", + "isUpdatable" : "0", + "stateSchema" : [ + + ], + "availability" : { + "macOS" : "14.0", + "tvOS" : "17.0", + "visionOS" : "1.0", + "watchOS" : "10.0", + "iOS" : "17.0", + "macCatalyst" : "17.0" + }, + "modelType" : { + "name" : "MLModelType_mlProgram" + }, + "userDefinedMetadata" : { + "com.github.apple.coremltools.conversion_date" : "2026-04-24", + "com.github.apple.coremltools.source" : "torch==2.9.1", + "com.github.apple.coremltools.version" : "9.0", + "com.github.apple.coremltools.source_dialect" : "TorchScript" + }, + "inputSchema" : [ + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1 × 32)", + "shortDescription" : "", + "shape" : "[1, 32]", + "name" : "latent", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1 × 512 × 16)", + "shortDescription" : "", + "shape" : "[1, 512, 16]", + "name" : "upsample_partial", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 2 × 1 × 256 × 8 × 64)", + "shortDescription" : "", + "shape" : "[2, 1, 256, 8, 64]", + "name" : "attn0_cache", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1)", + "shortDescription" : "", + "shape" : "[1]", + "name" : "attn0_offset", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 2 × 1 × 256 × 8 × 64)", + "shortDescription" : "", + "shape" : "[2, 1, 256, 8, 64]", + "name" : "attn1_cache", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1)", + "shortDescription" : "", + "shape" : "[1]", + "name" : "attn1_offset", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1 × 512 × 6)", + "shortDescription" : "", + "shape" : "[1, 512, 6]", + "name" : "conv0_prev", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1)", + "shortDescription" : "", + "shape" : "[1]", + "name" : "conv0_first", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1 × 256 × 6)", + "shortDescription" : "", + "shape" : "[1, 256, 6]", + "name" : "convtr0_partial", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1 × 256 × 2)", + "shortDescription" : "", + "shape" : "[1, 256, 2]", + "name" : "res0_conv0_prev", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1)", + "shortDescription" : "", + "shape" : "[1]", + "name" : "res0_conv0_first", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1 × 128 × 0)", + "shortDescription" : "", + "shape" : "[1, 128, 0]", + "name" : "res0_conv1_prev", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1)", + "shortDescription" : "", + "shape" : "[1]", + "name" : "res0_conv1_first", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1 × 128 × 5)", + "shortDescription" : "", + "shape" : "[1, 128, 5]", + "name" : "convtr1_partial", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1 × 128 × 2)", + "shortDescription" : "", + "shape" : "[1, 128, 2]", + "name" : "res1_conv0_prev", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1)", + "shortDescription" : "", + "shape" : "[1]", + "name" : "res1_conv0_first", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1 × 64 × 0)", + "shortDescription" : "", + "shape" : "[1, 64, 0]", + "name" : "res1_conv1_prev", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1)", + "shortDescription" : "", + "shape" : "[1]", + "name" : "res1_conv1_first", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1 × 64 × 4)", + "shortDescription" : "", + "shape" : "[1, 64, 4]", + "name" : "convtr2_partial", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1 × 64 × 2)", + "shortDescription" : "", + "shape" : "[1, 64, 2]", + "name" : "res2_conv0_prev", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1)", + "shortDescription" : "", + "shape" : "[1]", + "name" : "res2_conv0_first", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1 × 32 × 0)", + "shortDescription" : "", + "shape" : "[1, 32, 0]", + "name" : "res2_conv1_prev", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1)", + "shortDescription" : "", + "shape" : "[1]", + "name" : "res2_conv1_first", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1 × 64 × 2)", + "shortDescription" : "", + "shape" : "[1, 64, 2]", + "name" : "conv_final_prev", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1)", + "shortDescription" : "", + "shape" : "[1]", + "name" : "conv_final_first", + "type" : "MultiArray" + } + ], + "generatedClassName" : "mimi_decoder", + "method" : "predict" + } +] \ No newline at end of file diff --git a/v2.1/portuguese/mimi_decoder.mlmodelc/model.mil b/v2.1/portuguese/mimi_decoder.mlmodelc/model.mil new file mode 100644 index 0000000000000000000000000000000000000000..54d22260d0813157c84e216bfdbd2ff3fd7fcaf6 --- /dev/null +++ b/v2.1/portuguese/mimi_decoder.mlmodelc/model.mil @@ -0,0 +1,646 @@ +program(1.0) +[buildInfo = dict, tensor>({{"coremlc-component-MIL", "3520.4.1"}, {"coremlc-version", "3520.5.1"}, {"coremltools-component-torch", "2.9.1"}, {"coremltools-source-dialect", "TorchScript"}, {"coremltools-version", "9.0"}})] +{ + func main(tensor attn0_cache, tensor attn0_offset, tensor attn1_cache, tensor attn1_offset, tensor conv0_first, tensor conv0_prev, tensor conv_final_first, tensor conv_final_prev, tensor convtr0_partial, tensor convtr1_partial, tensor convtr2_partial, tensor latent, tensor res0_conv0_first, tensor res0_conv0_prev, tensor res0_conv1_first, tensor res0_conv1_prev, tensor res1_conv0_first, tensor res1_conv0_prev, tensor res1_conv1_first, tensor res1_conv1_prev, tensor res2_conv0_first, tensor res2_conv0_prev, tensor res2_conv1_first, tensor res2_conv1_prev, tensor upsample_partial) { + tensor emb_mean = const()[name = tensor("emb_mean"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(64)))]; + tensor emb_std = const()[name = tensor("emb_std"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(256)))]; + tensor mimi_quantizer_output_proj_weight = const()[name = tensor("mimi_quantizer_output_proj_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(448)))]; + tensor mimi_upsample_convtr_convtr_weight = const()[name = tensor("mimi_upsample_convtr_convtr_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(66048)))]; + tensor mimi_decoder_transformer_transformer_layers_0_norm1_bias = const()[name = tensor("mimi_decoder_transformer_transformer_layers_0_norm1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(131648)))]; + tensor mimi_decoder_transformer_transformer_layers_0_norm1_weight = const()[name = tensor("mimi_decoder_transformer_transformer_layers_0_norm1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(133760)))]; + tensor mimi_decoder_transformer_transformer_layers_0_self_attn_in_proj_weight = const()[name = tensor("mimi_decoder_transformer_transformer_layers_0_self_attn_in_proj_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(135872)))]; + tensor mimi_decoder_transformer_transformer_layers_0_self_attn_out_proj_weight = const()[name = tensor("mimi_decoder_transformer_transformer_layers_0_self_attn_out_proj_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(3281664)))]; + tensor mimi_decoder_transformer_transformer_layers_0_layer_scale_1_scale = const()[name = tensor("mimi_decoder_transformer_transformer_layers_0_layer_scale_1_scale"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(4330304)))]; + tensor mimi_decoder_transformer_transformer_layers_0_norm2_bias = const()[name = tensor("mimi_decoder_transformer_transformer_layers_0_norm2_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(4332416)))]; + tensor mimi_decoder_transformer_transformer_layers_0_norm2_weight = const()[name = tensor("mimi_decoder_transformer_transformer_layers_0_norm2_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(4334528)))]; + tensor mimi_decoder_transformer_transformer_layers_0_linear1_weight = const()[name = tensor("mimi_decoder_transformer_transformer_layers_0_linear1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(4336640)))]; + tensor mimi_decoder_transformer_transformer_layers_0_linear2_weight = const()[name = tensor("mimi_decoder_transformer_transformer_layers_0_linear2_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8531008)))]; + tensor mimi_decoder_transformer_transformer_layers_0_layer_scale_2_scale = const()[name = tensor("mimi_decoder_transformer_transformer_layers_0_layer_scale_2_scale"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(12725376)))]; + tensor mimi_decoder_transformer_transformer_layers_1_norm1_bias = const()[name = tensor("mimi_decoder_transformer_transformer_layers_1_norm1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(12727488)))]; + tensor mimi_decoder_transformer_transformer_layers_1_norm1_weight = const()[name = tensor("mimi_decoder_transformer_transformer_layers_1_norm1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(12729600)))]; + tensor mimi_decoder_transformer_transformer_layers_1_self_attn_in_proj_weight = const()[name = tensor("mimi_decoder_transformer_transformer_layers_1_self_attn_in_proj_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(12731712)))]; + tensor mimi_decoder_transformer_transformer_layers_1_self_attn_out_proj_weight = const()[name = tensor("mimi_decoder_transformer_transformer_layers_1_self_attn_out_proj_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15877504)))]; + tensor mimi_decoder_transformer_transformer_layers_1_layer_scale_1_scale = const()[name = tensor("mimi_decoder_transformer_transformer_layers_1_layer_scale_1_scale"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(16926144)))]; + tensor mimi_decoder_transformer_transformer_layers_1_norm2_bias = const()[name = tensor("mimi_decoder_transformer_transformer_layers_1_norm2_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(16928256)))]; + tensor mimi_decoder_transformer_transformer_layers_1_norm2_weight = const()[name = tensor("mimi_decoder_transformer_transformer_layers_1_norm2_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(16930368)))]; + tensor mimi_decoder_transformer_transformer_layers_1_linear1_weight = const()[name = tensor("mimi_decoder_transformer_transformer_layers_1_linear1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(16932480)))]; + tensor mimi_decoder_transformer_transformer_layers_1_linear2_weight = const()[name = tensor("mimi_decoder_transformer_transformer_layers_1_linear2_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21126848)))]; + tensor mimi_decoder_transformer_transformer_layers_1_layer_scale_2_scale = const()[name = tensor("mimi_decoder_transformer_transformer_layers_1_layer_scale_2_scale"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(25321216)))]; + tensor mimi_decoder_model_0_conv_bias = const()[name = tensor("mimi_decoder_model_0_conv_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(25323328)))]; + tensor mimi_decoder_model_0_conv_weight = const()[name = tensor("mimi_decoder_model_0_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(25325440)))]; + tensor mimi_decoder_model_2_convtr_bias = const()[name = tensor("mimi_decoder_model_2_convtr_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(32665536)))]; + tensor mimi_decoder_model_2_convtr_weight = const()[name = tensor("mimi_decoder_model_2_convtr_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(32666624)))]; + tensor mimi_decoder_model_3_block_1_conv_bias = const()[name = tensor("mimi_decoder_model_3_block_1_conv_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(38958144)))]; + tensor mimi_decoder_model_3_block_1_conv_weight = const()[name = tensor("mimi_decoder_model_3_block_1_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(38958720)))]; + tensor mimi_decoder_model_3_block_3_conv_bias = const()[name = tensor("mimi_decoder_model_3_block_3_conv_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(39352000)))]; + tensor mimi_decoder_model_3_block_3_conv_weight = const()[name = tensor("mimi_decoder_model_3_block_3_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(39353088)))]; + tensor mimi_decoder_model_5_convtr_bias = const()[name = tensor("mimi_decoder_model_5_convtr_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(39484224)))]; + tensor mimi_decoder_model_5_convtr_weight = const()[name = tensor("mimi_decoder_model_5_convtr_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(39484800)))]; + tensor mimi_decoder_model_6_block_1_conv_bias = const()[name = tensor("mimi_decoder_model_6_block_1_conv_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(40795584)))]; + tensor mimi_decoder_model_6_block_1_conv_weight = const()[name = tensor("mimi_decoder_model_6_block_1_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(40795904)))]; + tensor mimi_decoder_model_6_block_3_conv_bias = const()[name = tensor("mimi_decoder_model_6_block_3_conv_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(40894272)))]; + tensor mimi_decoder_model_6_block_3_conv_weight = const()[name = tensor("mimi_decoder_model_6_block_3_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(40894848)))]; + tensor mimi_decoder_model_8_convtr_bias = const()[name = tensor("mimi_decoder_model_8_convtr_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(40927680)))]; + tensor mimi_decoder_model_8_convtr_weight = const()[name = tensor("mimi_decoder_model_8_convtr_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(40928000)))]; + tensor mimi_decoder_model_9_block_1_conv_bias = const()[name = tensor("mimi_decoder_model_9_block_1_conv_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(41190208)))]; + tensor mimi_decoder_model_9_block_1_conv_weight = const()[name = tensor("mimi_decoder_model_9_block_1_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(41190400)))]; + tensor mimi_decoder_model_9_block_3_conv_bias = const()[name = tensor("mimi_decoder_model_9_block_3_conv_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(41215040)))]; + tensor mimi_decoder_model_9_block_3_conv_weight = const()[name = tensor("mimi_decoder_model_9_block_3_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(41215360)))]; + tensor mimi_decoder_model_11_conv_bias = const()[name = tensor("mimi_decoder_model_11_conv_bias"), val = tensor([0x1.28p-11])]; + tensor mimi_decoder_model_11_conv_weight = const()[name = tensor("mimi_decoder_model_11_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(41223616)))]; + tensor var_38 = mul(x = latent, y = emb_std)[name = tensor("op_38")]; + tensor denorm = add(x = var_38, y = emb_mean)[name = tensor("denorm")]; + tensor input_1_axes_0 = const()[name = tensor("input_1_axes_0"), val = tensor([-1])]; + tensor input_1 = expand_dims(axes = input_1_axes_0, x = denorm)[name = tensor("input_1")]; + tensor x_1_pad_type_0 = const()[name = tensor("x_1_pad_type_0"), val = tensor("valid")]; + tensor x_1_strides_0 = const()[name = tensor("x_1_strides_0"), val = tensor([1])]; + tensor x_1_pad_0 = const()[name = tensor("x_1_pad_0"), val = tensor([0, 0])]; + tensor x_1_dilations_0 = const()[name = tensor("x_1_dilations_0"), val = tensor([1])]; + tensor x_1_groups_0 = const()[name = tensor("x_1_groups_0"), val = tensor(1)]; + tensor x_1 = conv(dilations = x_1_dilations_0, groups = x_1_groups_0, pad = x_1_pad_0, pad_type = x_1_pad_type_0, strides = x_1_strides_0, weight = mimi_quantizer_output_proj_weight, x = input_1)[name = tensor("x_1")]; + tensor var_62 = const()[name = tensor("op_62"), val = tensor(-1)]; + tensor y_1_pad_type_0 = const()[name = tensor("y_1_pad_type_0"), val = tensor("valid")]; + tensor y_1_strides_0 = const()[name = tensor("y_1_strides_0"), val = tensor([16])]; + tensor y_1_groups_0 = const()[name = tensor("y_1_groups_0"), val = tensor(512)]; + tensor y_1_pad_0 = const()[name = tensor("y_1_pad_0"), val = tensor([0, 0])]; + tensor y_1_dilations_0 = const()[name = tensor("y_1_dilations_0"), val = tensor([1])]; + tensor y_1_has_output_shape_output_shape_0 = const()[name = tensor("y_1_has_output_shape_output_shape_0"), val = tensor([1, 512, 32])]; + tensor y_1_has_output_shape = conv_transpose(dilations = y_1_dilations_0, groups = y_1_groups_0, output_shape = y_1_has_output_shape_output_shape_0, pad = y_1_pad_0, pad_type = y_1_pad_type_0, strides = y_1_strides_0, weight = mimi_upsample_convtr_convtr_weight, x = x_1)[name = tensor("y_1_has_output_shape")]; + tensor var_72_begin_0 = const()[name = tensor("op_72_begin_0"), val = tensor([0, 0, 0])]; + tensor var_72_end_0 = const()[name = tensor("op_72_end_0"), val = tensor([1, 512, 16])]; + tensor var_72_end_mask_0 = const()[name = tensor("op_72_end_mask_0"), val = tensor([true, true, false])]; + tensor var_72 = slice_by_index(begin = var_72_begin_0, end = var_72_end_0, end_mask = var_72_end_mask_0, x = y_1_has_output_shape)[name = tensor("op_72")]; + tensor var_73 = add(x = var_72, y = upsample_partial)[name = tensor("op_73")]; + tensor var_74_begin_0 = const()[name = tensor("op_74_begin_0"), val = tensor([0, 0, 16])]; + tensor var_74_end_0 = const()[name = tensor("op_74_end_0"), val = tensor([1, 512, 32])]; + tensor var_74_end_mask_0 = const()[name = tensor("op_74_end_mask_0"), val = tensor([true, true, true])]; + tensor var_74 = slice_by_index(begin = var_74_begin_0, end = var_74_end_0, end_mask = var_74_end_mask_0, x = y_1_has_output_shape)[name = tensor("op_74")]; + tensor y_3_interleave_0 = const()[name = tensor("y_3_interleave_0"), val = tensor(false)]; + tensor y_3 = concat(axis = var_62, interleave = y_3_interleave_0, values = (var_73, var_74))[name = tensor("y_3")]; + tensor var_77_begin_0 = const()[name = tensor("op_77_begin_0"), val = tensor([0, 0, 16])]; + tensor var_77_end_0 = const()[name = tensor("op_77_end_0"), val = tensor([1, 512, 32])]; + tensor var_77_end_mask_0 = const()[name = tensor("op_77_end_mask_0"), val = tensor([true, true, true])]; + tensor var_77 = slice_by_index(begin = var_77_begin_0, end = var_77_end_0, end_mask = var_77_end_mask_0, x = y_3)[name = tensor("op_77")]; + tensor x_3_begin_0 = const()[name = tensor("x_3_begin_0"), val = tensor([0, 0, 0])]; + tensor x_3_end_0 = const()[name = tensor("x_3_end_0"), val = tensor([1, 512, 16])]; + tensor x_3_end_mask_0 = const()[name = tensor("x_3_end_mask_0"), val = tensor([true, true, false])]; + tensor x_3 = slice_by_index(begin = x_3_begin_0, end = x_3_end_0, end_mask = x_3_end_mask_0, x = y_3)[name = tensor("x_3")]; + tensor var_86 = const()[name = tensor("op_86"), val = tensor(0)]; + tensor var_91 = const()[name = tensor("op_91"), val = tensor(-1)]; + tensor var_100 = const()[name = tensor("op_100"), val = tensor(-0x1.ff933cp+127)]; + tensor var_102 = const()[name = tensor("op_102"), val = tensor(0x1.4f8b58p-17)]; + tensor input_3_perm_0 = const()[name = tensor("input_3_perm_0"), val = tensor([0, 2, 1])]; + tensor query_1_axes_0 = const()[name = tensor("query_1_axes_0"), val = tensor([-1])]; + tensor input_3 = transpose(perm = input_3_perm_0, x = x_3)[name = tensor("transpose_19")]; + tensor query_1 = layer_norm(axes = query_1_axes_0, beta = mimi_decoder_transformer_transformer_layers_0_norm1_bias, epsilon = var_102, gamma = mimi_decoder_transformer_transformer_layers_0_norm1_weight, x = input_3)[name = tensor("query_1")]; + tensor linear_0_bias_0 = const()[name = tensor("linear_0_bias_0"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(41224448)))]; + tensor projected_1 = linear(bias = linear_0_bias_0, weight = mimi_decoder_transformer_transformer_layers_0_self_attn_in_proj_weight, x = query_1)[name = tensor("linear_0")]; + tensor var_130 = const()[name = tensor("op_130"), val = tensor([1, 16, 3, 8, 64])]; + tensor packed_1 = reshape(shape = var_130, x = projected_1)[name = tensor("packed_1")]; + tensor var_132_split_sizes_0 = const()[name = tensor("op_132_split_sizes_0"), val = tensor([1, 1, 1])]; + tensor var_132_axis_0 = const()[name = tensor("op_132_axis_0"), val = tensor(2)]; + tensor var_132_0, tensor var_132_1, tensor var_132_2 = split(axis = var_132_axis_0, split_sizes = var_132_split_sizes_0, x = packed_1)[name = tensor("op_132")]; + tensor squeeze_0_axes_0 = const()[name = tensor("squeeze_0_axes_0"), val = tensor([2])]; + tensor squeeze_0 = squeeze(axes = squeeze_0_axes_0, x = var_132_0)[name = tensor("squeeze_0")]; + tensor squeeze_1_axes_0 = const()[name = tensor("squeeze_1_axes_0"), val = tensor([2])]; + tensor squeeze_1 = squeeze(axes = squeeze_1_axes_0, x = var_132_1)[name = tensor("squeeze_1")]; + tensor squeeze_2_axes_0 = const()[name = tensor("squeeze_2_axes_0"), val = tensor([2])]; + tensor squeeze_2 = squeeze(axes = squeeze_2_axes_0, x = var_132_2)[name = tensor("squeeze_2")]; + tensor offset_3_begin_0 = const()[name = tensor("offset_3_begin_0"), val = tensor([0])]; + tensor offset_3_end_0 = const()[name = tensor("offset_3_end_0"), val = tensor([1])]; + tensor offset_3_end_mask_0 = const()[name = tensor("offset_3_end_mask_0"), val = tensor([false])]; + tensor offset_3_squeeze_mask_0 = const()[name = tensor("offset_3_squeeze_mask_0"), val = tensor([true])]; + tensor offset_3 = slice_by_index(begin = offset_3_begin_0, end = offset_3_end_0, end_mask = offset_3_end_mask_0, squeeze_mask = offset_3_squeeze_mask_0, x = attn0_offset)[name = tensor("offset_3")]; + tensor freqs_1 = const()[name = tensor("freqs_1"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(41230656)))]; + tensor ts_1_promoted = const()[name = tensor("ts_1_promoted"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(41230848)))]; + tensor ts_3 = add(x = ts_1_promoted, y = offset_3)[name = tensor("ts_3")]; + tensor var_148 = const()[name = tensor("op_148"), val = tensor([-1, 1, 1])]; + tensor ts_5 = reshape(shape = var_148, x = ts_3)[name = tensor("ts_5")]; + tensor var_150 = const()[name = tensor("op_150"), val = tensor([1, 16, 8, 32, 2])]; + tensor q_3 = reshape(shape = var_150, x = squeeze_0)[name = tensor("q_3")]; + tensor var_152 = const()[name = tensor("op_152"), val = tensor([1, 16, 8, 32, 2])]; + tensor k_3 = reshape(shape = var_152, x = squeeze_1)[name = tensor("k_3")]; + tensor var_154_begin_0 = const()[name = tensor("op_154_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_154_end_0 = const()[name = tensor("op_154_end_0"), val = tensor([1, 16, 8, 32, 1])]; + tensor var_154_end_mask_0 = const()[name = tensor("op_154_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_154_squeeze_mask_0 = const()[name = tensor("op_154_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_154 = slice_by_index(begin = var_154_begin_0, end = var_154_end_0, end_mask = var_154_end_mask_0, squeeze_mask = var_154_squeeze_mask_0, x = q_3)[name = tensor("op_154")]; + tensor var_156_begin_0 = const()[name = tensor("op_156_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_156_end_0 = const()[name = tensor("op_156_end_0"), val = tensor([1, 16, 8, 32, 2])]; + tensor var_156_end_mask_0 = const()[name = tensor("op_156_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_156_squeeze_mask_0 = const()[name = tensor("op_156_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_156 = slice_by_index(begin = var_156_begin_0, end = var_156_end_0, end_mask = var_156_end_mask_0, squeeze_mask = var_156_squeeze_mask_0, x = q_3)[name = tensor("op_156")]; + tensor var_158_begin_0 = const()[name = tensor("op_158_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_158_end_0 = const()[name = tensor("op_158_end_0"), val = tensor([1, 16, 8, 32, 1])]; + tensor var_158_end_mask_0 = const()[name = tensor("op_158_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_158_squeeze_mask_0 = const()[name = tensor("op_158_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_158 = slice_by_index(begin = var_158_begin_0, end = var_158_end_0, end_mask = var_158_end_mask_0, squeeze_mask = var_158_squeeze_mask_0, x = k_3)[name = tensor("op_158")]; + tensor var_160_begin_0 = const()[name = tensor("op_160_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_160_end_0 = const()[name = tensor("op_160_end_0"), val = tensor([1, 16, 8, 32, 2])]; + tensor var_160_end_mask_0 = const()[name = tensor("op_160_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_160_squeeze_mask_0 = const()[name = tensor("op_160_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_160 = slice_by_index(begin = var_160_begin_0, end = var_160_end_0, end_mask = var_160_end_mask_0, squeeze_mask = var_160_squeeze_mask_0, x = k_3)[name = tensor("op_160")]; + tensor var_162 = mul(x = freqs_1, y = ts_5)[name = tensor("op_162")]; + tensor rotr_1 = cos(x = var_162)[name = tensor("rotr_1")]; + tensor roti_1 = sin(x = var_162)[name = tensor("roti_1")]; + tensor var_166 = mul(x = var_154, y = rotr_1)[name = tensor("op_166")]; + tensor var_167 = mul(x = var_156, y = roti_1)[name = tensor("op_167")]; + tensor qor_1 = sub(x = var_166, y = var_167)[name = tensor("qor_1")]; + tensor var_169 = mul(x = var_154, y = roti_1)[name = tensor("op_169")]; + tensor var_170 = mul(x = var_156, y = rotr_1)[name = tensor("op_170")]; + tensor qoi_1 = add(x = var_169, y = var_170)[name = tensor("qoi_1")]; + tensor var_172 = mul(x = var_158, y = rotr_1)[name = tensor("op_172")]; + tensor var_173 = mul(x = var_160, y = roti_1)[name = tensor("op_173")]; + tensor kor_1 = sub(x = var_172, y = var_173)[name = tensor("kor_1")]; + tensor var_175 = mul(x = var_158, y = roti_1)[name = tensor("op_175")]; + tensor var_176 = mul(x = var_160, y = rotr_1)[name = tensor("op_176")]; + tensor koi_1 = add(x = var_175, y = var_176)[name = tensor("koi_1")]; + tensor qo_1_axis_0 = const()[name = tensor("qo_1_axis_0"), val = tensor(-1)]; + tensor qo_1 = stack(axis = qo_1_axis_0, values = (qor_1, qoi_1))[name = tensor("qo_1")]; + tensor ko_1_axis_0 = const()[name = tensor("ko_1_axis_0"), val = tensor(-1)]; + tensor ko_1 = stack(axis = ko_1_axis_0, values = (kor_1, koi_1))[name = tensor("ko_1")]; + tensor var_186 = const()[name = tensor("op_186"), val = tensor([1, 16, 8, 64])]; + tensor q_5 = reshape(shape = var_186, x = qo_1)[name = tensor("q_5")]; + tensor var_188 = const()[name = tensor("op_188"), val = tensor([1, 16, 8, 64])]; + tensor k_5 = reshape(shape = var_188, x = ko_1)[name = tensor("k_5")]; + tensor capacity_1 = const()[name = tensor("capacity_1"), val = tensor([256])]; + tensor var_193_dtype_0 = const()[name = tensor("op_193_dtype_0"), val = tensor("int32")]; + tensor var_194 = const()[name = tensor("op_194"), val = tensor([1, 1])]; + tensor var_193 = cast(dtype = var_193_dtype_0, x = attn0_offset)[name = tensor("cast_49")]; + tensor write_base_1 = reshape(shape = var_194, x = var_193)[name = tensor("write_base_1")]; + tensor write_range_1 = const()[name = tensor("write_range_1"), val = tensor([[0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15]])]; + tensor abs_idx_1 = add(x = write_base_1, y = write_range_1)[name = tensor("abs_idx_1")]; + tensor wrapped_1_div = floor_div(x = abs_idx_1, y = capacity_1)[name = tensor("wrapped_1_div")]; + tensor wrapped_1_div_scaled = mul(x = wrapped_1_div, y = capacity_1)[name = tensor("wrapped_1_div_scaled")]; + tensor wrapped_1 = sub(x = abs_idx_1, y = wrapped_1_div_scaled)[name = tensor("wrapped_1")]; + tensor var_201 = const()[name = tensor("op_201"), val = tensor([1, 16, 1, 1])]; + tensor var_202 = reshape(shape = var_201, x = wrapped_1)[name = tensor("op_202")]; + tensor write_indexes_1_reps_0 = const()[name = tensor("write_indexes_1_reps_0"), val = tensor([1, 1, 8, 64])]; + tensor write_indexes_1 = tile(reps = write_indexes_1_reps_0, x = var_202)[name = tensor("write_indexes_1")]; + tensor var_205_begin_0 = const()[name = tensor("op_205_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_205_end_0 = const()[name = tensor("op_205_end_0"), val = tensor([1, 1, 256, 8, 64])]; + tensor var_205_end_mask_0 = const()[name = tensor("op_205_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_205_squeeze_mask_0 = const()[name = tensor("op_205_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor var_205 = slice_by_index(begin = var_205_begin_0, end = var_205_end_0, end_mask = var_205_end_mask_0, squeeze_mask = var_205_squeeze_mask_0, x = attn0_cache)[name = tensor("op_205")]; + tensor new_k_cache_1_axis_0 = const()[name = tensor("new_k_cache_1_axis_0"), val = tensor(1)]; + tensor new_k_cache_1_mode_0 = const()[name = tensor("new_k_cache_1_mode_0"), val = tensor("update")]; + tensor new_k_cache_1_validate_indices_0 = const()[name = tensor("new_k_cache_1_validate_indices_0"), val = tensor(false)]; + tensor new_k_cache_1 = scatter_along_axis(axis = new_k_cache_1_axis_0, data = var_205, indices = write_indexes_1, mode = new_k_cache_1_mode_0, updates = k_5, validate_indices = new_k_cache_1_validate_indices_0)[name = tensor("new_k_cache_1")]; + tensor var_207_begin_0 = const()[name = tensor("op_207_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor var_207_end_0 = const()[name = tensor("op_207_end_0"), val = tensor([2, 1, 256, 8, 64])]; + tensor var_207_end_mask_0 = const()[name = tensor("op_207_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_207_squeeze_mask_0 = const()[name = tensor("op_207_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor var_207 = slice_by_index(begin = var_207_begin_0, end = var_207_end_0, end_mask = var_207_end_mask_0, squeeze_mask = var_207_squeeze_mask_0, x = attn0_cache)[name = tensor("op_207")]; + tensor new_v_cache_1_axis_0 = const()[name = tensor("new_v_cache_1_axis_0"), val = tensor(1)]; + tensor new_v_cache_1_mode_0 = const()[name = tensor("new_v_cache_1_mode_0"), val = tensor("update")]; + tensor new_v_cache_1_validate_indices_0 = const()[name = tensor("new_v_cache_1_validate_indices_0"), val = tensor(false)]; + tensor new_v_cache_1 = scatter_along_axis(axis = new_v_cache_1_axis_0, data = var_207, indices = write_indexes_1, mode = new_v_cache_1_mode_0, updates = squeeze_2, validate_indices = new_v_cache_1_validate_indices_0)[name = tensor("new_v_cache_1")]; + tensor var_210_axis_0 = const()[name = tensor("op_210_axis_0"), val = tensor(0)]; + tensor var_210 = stack(axis = var_210_axis_0, values = (new_k_cache_1, new_v_cache_1))[name = tensor("op_210")]; + tensor var_211 = not_equal(x = new_k_cache_1, y = new_k_cache_1)[name = tensor("op_211")]; + tensor var_212 = const()[name = tensor("op_212"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(41230976)))]; + tensor new_k_cache_3 = select(a = var_212, b = new_k_cache_1, cond = var_211)[name = tensor("new_k_cache_3")]; + tensor var_214 = not_equal(x = new_v_cache_1, y = new_v_cache_1)[name = tensor("op_214")]; + tensor new_v_cache_3 = select(a = var_212, b = new_v_cache_1, cond = var_214)[name = tensor("new_v_cache_3")]; + tensor var_219 = const()[name = tensor("op_219"), val = tensor([0, 2, 1, 3])]; + tensor var_221 = const()[name = tensor("op_221"), val = tensor([1, 1])]; + tensor var_222 = reshape(shape = var_221, x = attn0_offset)[name = tensor("op_222")]; + tensor var_224_promoted = const()[name = tensor("op_224_promoted"), val = tensor([0x1.ep+3])]; + tensor var_225 = add(x = var_222, y = var_224_promoted)[name = tensor("op_225")]; + tensor last_pos_1_dtype_0 = const()[name = tensor("last_pos_1_dtype_0"), val = tensor("int32")]; + tensor slot_idx_1 = const()[name = tensor("slot_idx_1"), val = tensor([[0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19, 20, 21, 22, 23, 24, 25, 26, 27, 28, 29, 30, 31, 32, 33, 34, 35, 36, 37, 38, 39, 40, 41, 42, 43, 44, 45, 46, 47, 48, 49, 50, 51, 52, 53, 54, 55, 56, 57, 58, 59, 60, 61, 62, 63, 64, 65, 66, 67, 68, 69, 70, 71, 72, 73, 74, 75, 76, 77, 78, 79, 80, 81, 82, 83, 84, 85, 86, 87, 88, 89, 90, 91, 92, 93, 94, 95, 96, 97, 98, 99, 100, 101, 102, 103, 104, 105, 106, 107, 108, 109, 110, 111, 112, 113, 114, 115, 116, 117, 118, 119, 120, 121, 122, 123, 124, 125, 126, 127, 128, 129, 130, 131, 132, 133, 134, 135, 136, 137, 138, 139, 140, 141, 142, 143, 144, 145, 146, 147, 148, 149, 150, 151, 152, 153, 154, 155, 156, 157, 158, 159, 160, 161, 162, 163, 164, 165, 166, 167, 168, 169, 170, 171, 172, 173, 174, 175, 176, 177, 178, 179, 180, 181, 182, 183, 184, 185, 186, 187, 188, 189, 190, 191, 192, 193, 194, 195, 196, 197, 198, 199, 200, 201, 202, 203, 204, 205, 206, 207, 208, 209, 210, 211, 212, 213, 214, 215, 216, 217, 218, 219, 220, 221, 222, 223, 224, 225, 226, 227, 228, 229, 230, 231, 232, 233, 234, 235, 236, 237, 238, 239, 240, 241, 242, 243, 244, 245, 246, 247, 248, 249, 250, 251, 252, 253, 254, 255]])]; + tensor last_pos_1 = cast(dtype = last_pos_1_dtype_0, x = var_225)[name = tensor("cast_48")]; + tensor diff_1 = sub(x = last_pos_1, y = slot_idx_1)[name = tensor("diff_1")]; + tensor var_231_div = floor_div(x = diff_1, y = capacity_1)[name = tensor("op_231_div")]; + tensor var_231_div_scaled = mul(x = var_231_div, y = capacity_1)[name = tensor("op_231_div_scaled")]; + tensor var_231 = sub(x = diff_1, y = var_231_div_scaled)[name = tensor("op_231")]; + tensor pos_k_1 = sub(x = last_pos_1, y = var_231)[name = tensor("pos_k_1")]; + tensor var_237_promoted = const()[name = tensor("op_237_promoted"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(41755328)))]; + tensor pos_q_1 = add(x = var_222, y = var_237_promoted)[name = tensor("pos_q_1")]; + tensor var_241_axes_0 = const()[name = tensor("op_241_axes_0"), val = tensor([2])]; + tensor var_241 = expand_dims(axes = var_241_axes_0, x = pos_q_1)[name = tensor("op_241")]; + tensor var_243_axes_0 = const()[name = tensor("op_243_axes_0"), val = tensor([1])]; + tensor var_243 = expand_dims(axes = var_243_axes_0, x = pos_k_1)[name = tensor("op_243")]; + tensor var_244_promoted_dtype_0 = const()[name = tensor("op_244_promoted_dtype_0"), val = tensor("fp32")]; + tensor var_244_promoted = cast(dtype = var_244_promoted_dtype_0, x = var_243)[name = tensor("cast_47")]; + tensor delta_1 = sub(x = var_241, y = var_244_promoted)[name = tensor("delta_1")]; + tensor valid_1 = greater_equal(x = var_243, y = var_86)[name = tensor("valid_1")]; + tensor var_253 = const()[name = tensor("op_253"), val = tensor([1, 1, 1])]; + tensor var_254 = reshape(shape = var_253, x = attn0_offset)[name = tensor("op_254")]; + tensor var_256_promoted = const()[name = tensor("op_256_promoted"), val = tensor([0x1.ep+3])]; + tensor var_257 = add(x = var_254, y = var_256_promoted)[name = tensor("op_257")]; + tensor var_258 = less_equal(x = var_244_promoted, y = var_257)[name = tensor("op_258")]; + tensor valid_3 = logical_and(x = valid_1, y = var_258)[name = tensor("valid_3")]; + tensor var_86_promoted = const()[name = tensor("op_86_promoted"), val = tensor(0x0p+0)]; + tensor var_260 = greater_equal(x = delta_1, y = var_86_promoted)[name = tensor("op_260")]; + tensor attn_mask_1 = logical_and(x = valid_3, y = var_260)[name = tensor("attn_mask_1")]; + tensor var_98_promoted = const()[name = tensor("op_98_promoted"), val = tensor(0x1.f4p+7)]; + tensor var_262 = less(x = delta_1, y = var_98_promoted)[name = tensor("op_262")]; + tensor attn_mask_3 = logical_and(x = attn_mask_1, y = var_262)[name = tensor("attn_mask_3")]; + tensor attn_mask_5_axes_0 = const()[name = tensor("attn_mask_5_axes_0"), val = tensor([1])]; + tensor attn_mask_5 = expand_dims(axes = attn_mask_5_axes_0, x = attn_mask_3)[name = tensor("attn_mask_5")]; + tensor var_267_transpose_x_0 = const()[name = tensor("op_267_transpose_x_0"), val = tensor(false)]; + tensor var_267_transpose_y_0 = const()[name = tensor("op_267_transpose_y_0"), val = tensor(false)]; + tensor transpose_6_perm_0 = const()[name = tensor("transpose_6_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_7_perm_0 = const()[name = tensor("transpose_7_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_7 = transpose(perm = transpose_7_perm_0, x = new_k_cache_3)[name = tensor("transpose_16")]; + tensor transpose_6 = transpose(perm = transpose_6_perm_0, x = q_5)[name = tensor("transpose_17")]; + tensor var_267 = matmul(transpose_x = var_267_transpose_x_0, transpose_y = var_267_transpose_y_0, x = transpose_6, y = transpose_7)[name = tensor("op_267")]; + tensor var_268 = const()[name = tensor("op_268"), val = tensor(0x1p-3)]; + tensor attn_1 = mul(x = var_267, y = var_268)[name = tensor("attn_1")]; + tensor var_270 = logical_not(x = attn_mask_5)[name = tensor("op_270")]; + tensor attn_3 = select(a = var_100, b = attn_1, cond = var_270)[name = tensor("attn_3")]; + tensor attn_5 = softmax(axis = var_91, x = attn_3)[name = tensor("attn_5")]; + tensor x_5_transpose_x_0 = const()[name = tensor("x_5_transpose_x_0"), val = tensor(false)]; + tensor x_5_transpose_y_0 = const()[name = tensor("x_5_transpose_y_0"), val = tensor(false)]; + tensor v_attn_1 = transpose(perm = var_219, x = new_v_cache_3)[name = tensor("transpose_18")]; + tensor x_5 = matmul(transpose_x = x_5_transpose_x_0, transpose_y = x_5_transpose_y_0, x = attn_5, y = v_attn_1)[name = tensor("x_5")]; + tensor var_274_perm_0 = const()[name = tensor("op_274_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_275 = const()[name = tensor("op_275"), val = tensor([1, 16, 512])]; + tensor var_274 = transpose(perm = var_274_perm_0, x = x_5)[name = tensor("transpose_15")]; + tensor input_5 = reshape(shape = var_275, x = var_274)[name = tensor("input_5")]; + tensor linear_1_bias_0 = const()[name = tensor("linear_1_bias_0"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(41755456)))]; + tensor x_7 = linear(bias = linear_1_bias_0, weight = mimi_decoder_transformer_transformer_layers_0_self_attn_out_proj_weight, x = input_5)[name = tensor("linear_1")]; + tensor var_284 = mul(x = mimi_decoder_transformer_transformer_layers_0_layer_scale_1_scale, y = x_7)[name = tensor("op_284")]; + tensor input_7 = add(x = input_3, y = var_284)[name = tensor("input_7")]; + tensor input_9_axes_0 = const()[name = tensor("input_9_axes_0"), val = tensor([-1])]; + tensor input_9 = layer_norm(axes = input_9_axes_0, beta = mimi_decoder_transformer_transformer_layers_0_norm2_bias, epsilon = var_102, gamma = mimi_decoder_transformer_transformer_layers_0_norm2_weight, x = input_7)[name = tensor("input_9")]; + tensor linear_2_bias_0 = const()[name = tensor("linear_2_bias_0"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(41757568)))]; + tensor var_291 = linear(bias = linear_2_bias_0, weight = mimi_decoder_transformer_transformer_layers_0_linear1_weight, x = input_9)[name = tensor("linear_2")]; + tensor input_11_mode_0 = const()[name = tensor("input_11_mode_0"), val = tensor("EXACT")]; + tensor input_11 = gelu(mode = input_11_mode_0, x = var_291)[name = tensor("input_11")]; + tensor x_9 = linear(bias = linear_1_bias_0, weight = mimi_decoder_transformer_transformer_layers_0_linear2_weight, x = input_11)[name = tensor("linear_3")]; + tensor var_297 = mul(x = mimi_decoder_transformer_transformer_layers_0_layer_scale_2_scale, y = x_9)[name = tensor("op_297")]; + tensor input_13 = add(x = input_7, y = var_297)[name = tensor("input_13")]; + tensor query_axes_0 = const()[name = tensor("query_axes_0"), val = tensor([-1])]; + tensor query = layer_norm(axes = query_axes_0, beta = mimi_decoder_transformer_transformer_layers_1_norm1_bias, epsilon = var_102, gamma = mimi_decoder_transformer_transformer_layers_1_norm1_weight, x = input_13)[name = tensor("query")]; + tensor projected = linear(bias = linear_0_bias_0, weight = mimi_decoder_transformer_transformer_layers_1_self_attn_in_proj_weight, x = query)[name = tensor("linear_4")]; + tensor var_320 = const()[name = tensor("op_320"), val = tensor([1, 16, 3, 8, 64])]; + tensor packed = reshape(shape = var_320, x = projected)[name = tensor("packed")]; + tensor var_322_split_sizes_0 = const()[name = tensor("op_322_split_sizes_0"), val = tensor([1, 1, 1])]; + tensor var_322_axis_0 = const()[name = tensor("op_322_axis_0"), val = tensor(2)]; + tensor var_322_0, tensor var_322_1, tensor var_322_2 = split(axis = var_322_axis_0, split_sizes = var_322_split_sizes_0, x = packed)[name = tensor("op_322")]; + tensor squeeze_3_axes_0 = const()[name = tensor("squeeze_3_axes_0"), val = tensor([2])]; + tensor squeeze_3 = squeeze(axes = squeeze_3_axes_0, x = var_322_0)[name = tensor("squeeze_3")]; + tensor squeeze_4_axes_0 = const()[name = tensor("squeeze_4_axes_0"), val = tensor([2])]; + tensor squeeze_4 = squeeze(axes = squeeze_4_axes_0, x = var_322_1)[name = tensor("squeeze_4")]; + tensor squeeze_5_axes_0 = const()[name = tensor("squeeze_5_axes_0"), val = tensor([2])]; + tensor squeeze_5 = squeeze(axes = squeeze_5_axes_0, x = var_322_2)[name = tensor("squeeze_5")]; + tensor offset_begin_0 = const()[name = tensor("offset_begin_0"), val = tensor([0])]; + tensor offset_end_0 = const()[name = tensor("offset_end_0"), val = tensor([1])]; + tensor offset_end_mask_0 = const()[name = tensor("offset_end_mask_0"), val = tensor([false])]; + tensor offset_squeeze_mask_0 = const()[name = tensor("offset_squeeze_mask_0"), val = tensor([true])]; + tensor offset = slice_by_index(begin = offset_begin_0, end = offset_end_0, end_mask = offset_end_mask_0, squeeze_mask = offset_squeeze_mask_0, x = attn1_offset)[name = tensor("offset")]; + tensor freqs = const()[name = tensor("freqs"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(41765824)))]; + tensor ts_7_promoted = const()[name = tensor("ts_7_promoted"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(41766016)))]; + tensor ts_9 = add(x = ts_7_promoted, y = offset)[name = tensor("ts_9")]; + tensor var_338 = const()[name = tensor("op_338"), val = tensor([-1, 1, 1])]; + tensor ts = reshape(shape = var_338, x = ts_9)[name = tensor("ts")]; + tensor var_340 = const()[name = tensor("op_340"), val = tensor([1, 16, 8, 32, 2])]; + tensor q_9 = reshape(shape = var_340, x = squeeze_3)[name = tensor("q_9")]; + tensor var_342 = const()[name = tensor("op_342"), val = tensor([1, 16, 8, 32, 2])]; + tensor k_9 = reshape(shape = var_342, x = squeeze_4)[name = tensor("k_9")]; + tensor var_344_begin_0 = const()[name = tensor("op_344_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_344_end_0 = const()[name = tensor("op_344_end_0"), val = tensor([1, 16, 8, 32, 1])]; + tensor var_344_end_mask_0 = const()[name = tensor("op_344_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_344_squeeze_mask_0 = const()[name = tensor("op_344_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_344 = slice_by_index(begin = var_344_begin_0, end = var_344_end_0, end_mask = var_344_end_mask_0, squeeze_mask = var_344_squeeze_mask_0, x = q_9)[name = tensor("op_344")]; + tensor var_346_begin_0 = const()[name = tensor("op_346_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_346_end_0 = const()[name = tensor("op_346_end_0"), val = tensor([1, 16, 8, 32, 2])]; + tensor var_346_end_mask_0 = const()[name = tensor("op_346_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_346_squeeze_mask_0 = const()[name = tensor("op_346_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_346 = slice_by_index(begin = var_346_begin_0, end = var_346_end_0, end_mask = var_346_end_mask_0, squeeze_mask = var_346_squeeze_mask_0, x = q_9)[name = tensor("op_346")]; + tensor var_348_begin_0 = const()[name = tensor("op_348_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_348_end_0 = const()[name = tensor("op_348_end_0"), val = tensor([1, 16, 8, 32, 1])]; + tensor var_348_end_mask_0 = const()[name = tensor("op_348_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_348_squeeze_mask_0 = const()[name = tensor("op_348_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_348 = slice_by_index(begin = var_348_begin_0, end = var_348_end_0, end_mask = var_348_end_mask_0, squeeze_mask = var_348_squeeze_mask_0, x = k_9)[name = tensor("op_348")]; + tensor var_350_begin_0 = const()[name = tensor("op_350_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_350_end_0 = const()[name = tensor("op_350_end_0"), val = tensor([1, 16, 8, 32, 2])]; + tensor var_350_end_mask_0 = const()[name = tensor("op_350_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_350_squeeze_mask_0 = const()[name = tensor("op_350_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_350 = slice_by_index(begin = var_350_begin_0, end = var_350_end_0, end_mask = var_350_end_mask_0, squeeze_mask = var_350_squeeze_mask_0, x = k_9)[name = tensor("op_350")]; + tensor var_352 = mul(x = freqs, y = ts)[name = tensor("op_352")]; + tensor rotr = cos(x = var_352)[name = tensor("rotr")]; + tensor roti = sin(x = var_352)[name = tensor("roti")]; + tensor var_356 = mul(x = var_344, y = rotr)[name = tensor("op_356")]; + tensor var_357 = mul(x = var_346, y = roti)[name = tensor("op_357")]; + tensor qor_5 = sub(x = var_356, y = var_357)[name = tensor("qor_5")]; + tensor var_359 = mul(x = var_344, y = roti)[name = tensor("op_359")]; + tensor var_360 = mul(x = var_346, y = rotr)[name = tensor("op_360")]; + tensor qoi_5 = add(x = var_359, y = var_360)[name = tensor("qoi_5")]; + tensor var_362 = mul(x = var_348, y = rotr)[name = tensor("op_362")]; + tensor var_363 = mul(x = var_350, y = roti)[name = tensor("op_363")]; + tensor kor_5 = sub(x = var_362, y = var_363)[name = tensor("kor_5")]; + tensor var_365 = mul(x = var_348, y = roti)[name = tensor("op_365")]; + tensor var_366 = mul(x = var_350, y = rotr)[name = tensor("op_366")]; + tensor koi_5 = add(x = var_365, y = var_366)[name = tensor("koi_5")]; + tensor qo_axis_0 = const()[name = tensor("qo_axis_0"), val = tensor(-1)]; + tensor qo = stack(axis = qo_axis_0, values = (qor_5, qoi_5))[name = tensor("qo")]; + tensor ko_axis_0 = const()[name = tensor("ko_axis_0"), val = tensor(-1)]; + tensor ko = stack(axis = ko_axis_0, values = (kor_5, koi_5))[name = tensor("ko")]; + tensor var_376 = const()[name = tensor("op_376"), val = tensor([1, 16, 8, 64])]; + tensor q = reshape(shape = var_376, x = qo)[name = tensor("q")]; + tensor var_378 = const()[name = tensor("op_378"), val = tensor([1, 16, 8, 64])]; + tensor k = reshape(shape = var_378, x = ko)[name = tensor("k")]; + tensor capacity = const()[name = tensor("capacity"), val = tensor([256])]; + tensor var_383_dtype_0 = const()[name = tensor("op_383_dtype_0"), val = tensor("int32")]; + tensor var_384 = const()[name = tensor("op_384"), val = tensor([1, 1])]; + tensor var_383 = cast(dtype = var_383_dtype_0, x = attn1_offset)[name = tensor("cast_46")]; + tensor write_base = reshape(shape = var_384, x = var_383)[name = tensor("write_base")]; + tensor write_range = const()[name = tensor("write_range"), val = tensor([[0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15]])]; + tensor abs_idx = add(x = write_base, y = write_range)[name = tensor("abs_idx")]; + tensor wrapped_div = floor_div(x = abs_idx, y = capacity)[name = tensor("wrapped_div")]; + tensor wrapped_div_scaled = mul(x = wrapped_div, y = capacity)[name = tensor("wrapped_div_scaled")]; + tensor wrapped = sub(x = abs_idx, y = wrapped_div_scaled)[name = tensor("wrapped")]; + tensor var_391 = const()[name = tensor("op_391"), val = tensor([1, 16, 1, 1])]; + tensor var_392 = reshape(shape = var_391, x = wrapped)[name = tensor("op_392")]; + tensor write_indexes_reps_0 = const()[name = tensor("write_indexes_reps_0"), val = tensor([1, 1, 8, 64])]; + tensor write_indexes = tile(reps = write_indexes_reps_0, x = var_392)[name = tensor("write_indexes")]; + tensor var_395_begin_0 = const()[name = tensor("op_395_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_395_end_0 = const()[name = tensor("op_395_end_0"), val = tensor([1, 1, 256, 8, 64])]; + tensor var_395_end_mask_0 = const()[name = tensor("op_395_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_395_squeeze_mask_0 = const()[name = tensor("op_395_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor var_395 = slice_by_index(begin = var_395_begin_0, end = var_395_end_0, end_mask = var_395_end_mask_0, squeeze_mask = var_395_squeeze_mask_0, x = attn1_cache)[name = tensor("op_395")]; + tensor new_k_cache_5_axis_0 = const()[name = tensor("new_k_cache_5_axis_0"), val = tensor(1)]; + tensor new_k_cache_5_mode_0 = const()[name = tensor("new_k_cache_5_mode_0"), val = tensor("update")]; + tensor new_k_cache_5_validate_indices_0 = const()[name = tensor("new_k_cache_5_validate_indices_0"), val = tensor(false)]; + tensor new_k_cache_5 = scatter_along_axis(axis = new_k_cache_5_axis_0, data = var_395, indices = write_indexes, mode = new_k_cache_5_mode_0, updates = k, validate_indices = new_k_cache_5_validate_indices_0)[name = tensor("new_k_cache_5")]; + tensor var_397_begin_0 = const()[name = tensor("op_397_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor var_397_end_0 = const()[name = tensor("op_397_end_0"), val = tensor([2, 1, 256, 8, 64])]; + tensor var_397_end_mask_0 = const()[name = tensor("op_397_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_397_squeeze_mask_0 = const()[name = tensor("op_397_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor var_397 = slice_by_index(begin = var_397_begin_0, end = var_397_end_0, end_mask = var_397_end_mask_0, squeeze_mask = var_397_squeeze_mask_0, x = attn1_cache)[name = tensor("op_397")]; + tensor new_v_cache_5_axis_0 = const()[name = tensor("new_v_cache_5_axis_0"), val = tensor(1)]; + tensor new_v_cache_5_mode_0 = const()[name = tensor("new_v_cache_5_mode_0"), val = tensor("update")]; + tensor new_v_cache_5_validate_indices_0 = const()[name = tensor("new_v_cache_5_validate_indices_0"), val = tensor(false)]; + tensor new_v_cache_5 = scatter_along_axis(axis = new_v_cache_5_axis_0, data = var_397, indices = write_indexes, mode = new_v_cache_5_mode_0, updates = squeeze_5, validate_indices = new_v_cache_5_validate_indices_0)[name = tensor("new_v_cache_5")]; + tensor var_400_axis_0 = const()[name = tensor("op_400_axis_0"), val = tensor(0)]; + tensor var_400 = stack(axis = var_400_axis_0, values = (new_k_cache_5, new_v_cache_5))[name = tensor("op_400")]; + tensor var_401 = not_equal(x = new_k_cache_5, y = new_k_cache_5)[name = tensor("op_401")]; + tensor new_k_cache = select(a = var_212, b = new_k_cache_5, cond = var_401)[name = tensor("new_k_cache")]; + tensor var_404 = not_equal(x = new_v_cache_5, y = new_v_cache_5)[name = tensor("op_404")]; + tensor new_v_cache = select(a = var_212, b = new_v_cache_5, cond = var_404)[name = tensor("new_v_cache")]; + tensor var_409 = const()[name = tensor("op_409"), val = tensor([0, 2, 1, 3])]; + tensor var_411 = const()[name = tensor("op_411"), val = tensor([1, 1])]; + tensor var_412 = reshape(shape = var_411, x = attn1_offset)[name = tensor("op_412")]; + tensor var_414_promoted = const()[name = tensor("op_414_promoted"), val = tensor([0x1.ep+3])]; + tensor var_415 = add(x = var_412, y = var_414_promoted)[name = tensor("op_415")]; + tensor last_pos_dtype_0 = const()[name = tensor("last_pos_dtype_0"), val = tensor("int32")]; + tensor last_pos = cast(dtype = last_pos_dtype_0, x = var_415)[name = tensor("cast_45")]; + tensor diff = sub(x = last_pos, y = slot_idx_1)[name = tensor("diff")]; + tensor var_421_div = floor_div(x = diff, y = capacity)[name = tensor("op_421_div")]; + tensor var_421_div_scaled = mul(x = var_421_div, y = capacity)[name = tensor("op_421_div_scaled")]; + tensor var_421 = sub(x = diff, y = var_421_div_scaled)[name = tensor("op_421")]; + tensor pos_k = sub(x = last_pos, y = var_421)[name = tensor("pos_k")]; + tensor var_427_promoted = const()[name = tensor("op_427_promoted"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(41766144)))]; + tensor pos_q = add(x = var_412, y = var_427_promoted)[name = tensor("pos_q")]; + tensor var_431_axes_0 = const()[name = tensor("op_431_axes_0"), val = tensor([2])]; + tensor var_431 = expand_dims(axes = var_431_axes_0, x = pos_q)[name = tensor("op_431")]; + tensor var_433_axes_0 = const()[name = tensor("op_433_axes_0"), val = tensor([1])]; + tensor var_433 = expand_dims(axes = var_433_axes_0, x = pos_k)[name = tensor("op_433")]; + tensor var_434_promoted_dtype_0 = const()[name = tensor("op_434_promoted_dtype_0"), val = tensor("fp32")]; + tensor var_434_promoted = cast(dtype = var_434_promoted_dtype_0, x = var_433)[name = tensor("cast_44")]; + tensor delta = sub(x = var_431, y = var_434_promoted)[name = tensor("delta")]; + tensor valid_5 = greater_equal(x = var_433, y = var_86)[name = tensor("valid_5")]; + tensor var_443 = const()[name = tensor("op_443"), val = tensor([1, 1, 1])]; + tensor var_444 = reshape(shape = var_443, x = attn1_offset)[name = tensor("op_444")]; + tensor var_446_promoted = const()[name = tensor("op_446_promoted"), val = tensor([0x1.ep+3])]; + tensor var_447 = add(x = var_444, y = var_446_promoted)[name = tensor("op_447")]; + tensor var_448 = less_equal(x = var_434_promoted, y = var_447)[name = tensor("op_448")]; + tensor valid = logical_and(x = valid_5, y = var_448)[name = tensor("valid")]; + tensor var_86_promoted_1 = const()[name = tensor("op_86_promoted_1"), val = tensor(0x0p+0)]; + tensor var_450 = greater_equal(x = delta, y = var_86_promoted_1)[name = tensor("op_450")]; + tensor attn_mask_7 = logical_and(x = valid, y = var_450)[name = tensor("attn_mask_7")]; + tensor var_98_promoted_1 = const()[name = tensor("op_98_promoted_1"), val = tensor(0x1.f4p+7)]; + tensor var_452 = less(x = delta, y = var_98_promoted_1)[name = tensor("op_452")]; + tensor attn_mask_9 = logical_and(x = attn_mask_7, y = var_452)[name = tensor("attn_mask_9")]; + tensor attn_mask_axes_0 = const()[name = tensor("attn_mask_axes_0"), val = tensor([1])]; + tensor attn_mask = expand_dims(axes = attn_mask_axes_0, x = attn_mask_9)[name = tensor("attn_mask")]; + tensor var_457_transpose_x_0 = const()[name = tensor("op_457_transpose_x_0"), val = tensor(false)]; + tensor var_457_transpose_y_0 = const()[name = tensor("op_457_transpose_y_0"), val = tensor(false)]; + tensor transpose_8_perm_0 = const()[name = tensor("transpose_8_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_9_perm_0 = const()[name = tensor("transpose_9_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_9 = transpose(perm = transpose_9_perm_0, x = new_k_cache)[name = tensor("transpose_12")]; + tensor transpose_8 = transpose(perm = transpose_8_perm_0, x = q)[name = tensor("transpose_13")]; + tensor var_457 = matmul(transpose_x = var_457_transpose_x_0, transpose_y = var_457_transpose_y_0, x = transpose_8, y = transpose_9)[name = tensor("op_457")]; + tensor var_458 = const()[name = tensor("op_458"), val = tensor(0x1p-3)]; + tensor attn_7 = mul(x = var_457, y = var_458)[name = tensor("attn_7")]; + tensor var_460 = logical_not(x = attn_mask)[name = tensor("op_460")]; + tensor attn_9 = select(a = var_100, b = attn_7, cond = var_460)[name = tensor("attn_9")]; + tensor attn = softmax(axis = var_91, x = attn_9)[name = tensor("attn")]; + tensor x_11_transpose_x_0 = const()[name = tensor("x_11_transpose_x_0"), val = tensor(false)]; + tensor x_11_transpose_y_0 = const()[name = tensor("x_11_transpose_y_0"), val = tensor(false)]; + tensor v_attn = transpose(perm = var_409, x = new_v_cache)[name = tensor("transpose_14")]; + tensor x_11 = matmul(transpose_x = x_11_transpose_x_0, transpose_y = x_11_transpose_y_0, x = attn, y = v_attn)[name = tensor("x_11")]; + tensor var_464_perm_0 = const()[name = tensor("op_464_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_465 = const()[name = tensor("op_465"), val = tensor([1, 16, 512])]; + tensor var_464 = transpose(perm = var_464_perm_0, x = x_11)[name = tensor("transpose_11")]; + tensor input_15 = reshape(shape = var_465, x = var_464)[name = tensor("input_15")]; + tensor x_13 = linear(bias = linear_1_bias_0, weight = mimi_decoder_transformer_transformer_layers_1_self_attn_out_proj_weight, x = input_15)[name = tensor("linear_5")]; + tensor var_474 = mul(x = mimi_decoder_transformer_transformer_layers_1_layer_scale_1_scale, y = x_13)[name = tensor("op_474")]; + tensor input_17 = add(x = input_13, y = var_474)[name = tensor("input_17")]; + tensor input_19_axes_0 = const()[name = tensor("input_19_axes_0"), val = tensor([-1])]; + tensor input_19 = layer_norm(axes = input_19_axes_0, beta = mimi_decoder_transformer_transformer_layers_1_norm2_bias, epsilon = var_102, gamma = mimi_decoder_transformer_transformer_layers_1_norm2_weight, x = input_17)[name = tensor("input_19")]; + tensor var_481 = linear(bias = linear_2_bias_0, weight = mimi_decoder_transformer_transformer_layers_1_linear1_weight, x = input_19)[name = tensor("linear_6")]; + tensor input_21_mode_0 = const()[name = tensor("input_21_mode_0"), val = tensor("EXACT")]; + tensor input_21 = gelu(mode = input_21_mode_0, x = var_481)[name = tensor("input_21")]; + tensor x_15 = linear(bias = linear_1_bias_0, weight = mimi_decoder_transformer_transformer_layers_1_linear2_weight, x = input_21)[name = tensor("linear_7")]; + tensor var_487 = mul(x = mimi_decoder_transformer_transformer_layers_1_layer_scale_2_scale, y = x_15)[name = tensor("op_487")]; + tensor z = add(x = input_17, y = var_487)[name = tensor("z")]; + tensor x_17_perm_0 = const()[name = tensor("x_17_perm_0"), val = tensor([0, 2, 1])]; + tensor var_507 = const()[name = tensor("op_507"), val = tensor(0x1p+0)]; + tensor var_508 = const()[name = tensor("op_508"), val = tensor(-1)]; + tensor input_23_interleave_0 = const()[name = tensor("input_23_interleave_0"), val = tensor(false)]; + tensor x_17 = transpose(perm = x_17_perm_0, x = z)[name = tensor("transpose_10")]; + tensor input_23 = concat(axis = var_508, interleave = input_23_interleave_0, values = (conv0_prev, x_17))[name = tensor("input_23")]; + tensor input_25_pad_type_0 = const()[name = tensor("input_25_pad_type_0"), val = tensor("valid")]; + tensor input_25_strides_0 = const()[name = tensor("input_25_strides_0"), val = tensor([1])]; + tensor input_25_pad_0 = const()[name = tensor("input_25_pad_0"), val = tensor([0, 0])]; + tensor input_25_dilations_0 = const()[name = tensor("input_25_dilations_0"), val = tensor([1])]; + tensor input_25_groups_0 = const()[name = tensor("input_25_groups_0"), val = tensor(1)]; + tensor input_25 = conv(bias = mimi_decoder_model_0_conv_bias, dilations = input_25_dilations_0, groups = input_25_groups_0, pad = input_25_pad_0, pad_type = input_25_pad_type_0, strides = input_25_strides_0, weight = mimi_decoder_model_0_conv_weight, x = input_23)[name = tensor("input_25")]; + tensor var_542_begin_0 = const()[name = tensor("op_542_begin_0"), val = tensor([0, 0, 16])]; + tensor var_542_end_0 = const()[name = tensor("op_542_end_0"), val = tensor([1, 512, 22])]; + tensor var_542_end_mask_0 = const()[name = tensor("op_542_end_mask_0"), val = tensor([true, true, true])]; + tensor var_542 = slice_by_index(begin = var_542_begin_0, end = var_542_end_0, end_mask = var_542_end_mask_0, x = input_23)[name = tensor("op_542")]; + tensor input_27 = elu(alpha = var_507, x = input_25)[name = tensor("input_27")]; + tensor y_5_pad_type_0 = const()[name = tensor("y_5_pad_type_0"), val = tensor("valid")]; + tensor y_5_strides_0 = const()[name = tensor("y_5_strides_0"), val = tensor([6])]; + tensor y_5_pad_0 = const()[name = tensor("y_5_pad_0"), val = tensor([0, 0])]; + tensor y_5_dilations_0 = const()[name = tensor("y_5_dilations_0"), val = tensor([1])]; + tensor y_5_groups_0 = const()[name = tensor("y_5_groups_0"), val = tensor(1)]; + tensor y_5_has_output_shape_output_shape_0 = const()[name = tensor("y_5_has_output_shape_output_shape_0"), val = tensor([1, 256, 102])]; + tensor y_5_has_output_shape = conv_transpose(bias = mimi_decoder_model_2_convtr_bias, dilations = y_5_dilations_0, groups = y_5_groups_0, output_shape = y_5_has_output_shape_output_shape_0, pad = y_5_pad_0, pad_type = y_5_pad_type_0, strides = y_5_strides_0, weight = mimi_decoder_model_2_convtr_weight, x = input_27)[name = tensor("y_5_has_output_shape")]; + tensor var_557_begin_0 = const()[name = tensor("op_557_begin_0"), val = tensor([0, 0, 0])]; + tensor var_557_end_0 = const()[name = tensor("op_557_end_0"), val = tensor([1, 256, 6])]; + tensor var_557_end_mask_0 = const()[name = tensor("op_557_end_mask_0"), val = tensor([true, true, false])]; + tensor var_557 = slice_by_index(begin = var_557_begin_0, end = var_557_end_0, end_mask = var_557_end_mask_0, x = y_5_has_output_shape)[name = tensor("op_557")]; + tensor var_558 = add(x = var_557, y = convtr0_partial)[name = tensor("op_558")]; + tensor var_559_begin_0 = const()[name = tensor("op_559_begin_0"), val = tensor([0, 0, 6])]; + tensor var_559_end_0 = const()[name = tensor("op_559_end_0"), val = tensor([1, 256, 102])]; + tensor var_559_end_mask_0 = const()[name = tensor("op_559_end_mask_0"), val = tensor([true, true, true])]; + tensor var_559 = slice_by_index(begin = var_559_begin_0, end = var_559_end_0, end_mask = var_559_end_mask_0, x = y_5_has_output_shape)[name = tensor("op_559")]; + tensor y_7_interleave_0 = const()[name = tensor("y_7_interleave_0"), val = tensor(false)]; + tensor y_7 = concat(axis = var_508, interleave = y_7_interleave_0, values = (var_558, var_559))[name = tensor("y_7")]; + tensor new_partial_1_begin_0 = const()[name = tensor("new_partial_1_begin_0"), val = tensor([0, 0, 96])]; + tensor new_partial_1_end_0 = const()[name = tensor("new_partial_1_end_0"), val = tensor([1, 256, 102])]; + tensor new_partial_1_end_mask_0 = const()[name = tensor("new_partial_1_end_mask_0"), val = tensor([true, true, true])]; + tensor new_partial_1 = slice_by_index(begin = new_partial_1_begin_0, end = new_partial_1_end_0, end_mask = new_partial_1_end_mask_0, x = y_7)[name = tensor("new_partial_1")]; + tensor var_564 = const()[name = tensor("op_564"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(41766272)))]; + tensor var_565 = sub(x = new_partial_1, y = var_564)[name = tensor("op_565")]; + tensor input_29_begin_0 = const()[name = tensor("input_29_begin_0"), val = tensor([0, 0, 0])]; + tensor input_29_end_0 = const()[name = tensor("input_29_end_0"), val = tensor([1, 256, 96])]; + tensor input_29_end_mask_0 = const()[name = tensor("input_29_end_mask_0"), val = tensor([true, true, false])]; + tensor input_29 = slice_by_index(begin = input_29_begin_0, end = input_29_end_0, end_mask = input_29_end_mask_0, x = y_7)[name = tensor("input_29")]; + tensor x_19 = elu(alpha = var_507, x = input_29)[name = tensor("x_19")]; + tensor input_31_interleave_0 = const()[name = tensor("input_31_interleave_0"), val = tensor(false)]; + tensor input_31 = concat(axis = var_508, interleave = input_31_interleave_0, values = (res0_conv0_prev, x_19))[name = tensor("input_31")]; + tensor input_33_pad_type_0 = const()[name = tensor("input_33_pad_type_0"), val = tensor("valid")]; + tensor input_33_strides_0 = const()[name = tensor("input_33_strides_0"), val = tensor([1])]; + tensor input_33_pad_0 = const()[name = tensor("input_33_pad_0"), val = tensor([0, 0])]; + tensor input_33_dilations_0 = const()[name = tensor("input_33_dilations_0"), val = tensor([1])]; + tensor input_33_groups_0 = const()[name = tensor("input_33_groups_0"), val = tensor(1)]; + tensor input_33 = conv(bias = mimi_decoder_model_3_block_1_conv_bias, dilations = input_33_dilations_0, groups = input_33_groups_0, pad = input_33_pad_0, pad_type = input_33_pad_type_0, strides = input_33_strides_0, weight = mimi_decoder_model_3_block_1_conv_weight, x = input_31)[name = tensor("input_33")]; + tensor var_585_begin_0 = const()[name = tensor("op_585_begin_0"), val = tensor([0, 0, 96])]; + tensor var_585_end_0 = const()[name = tensor("op_585_end_0"), val = tensor([1, 256, 98])]; + tensor var_585_end_mask_0 = const()[name = tensor("op_585_end_mask_0"), val = tensor([true, true, true])]; + tensor var_585 = slice_by_index(begin = var_585_begin_0, end = var_585_end_0, end_mask = var_585_end_mask_0, x = input_31)[name = tensor("op_585")]; + tensor x_21 = elu(alpha = var_507, x = input_33)[name = tensor("x_21")]; + tensor v_5_pad_type_0 = const()[name = tensor("v_5_pad_type_0"), val = tensor("valid")]; + tensor v_5_strides_0 = const()[name = tensor("v_5_strides_0"), val = tensor([1])]; + tensor v_5_pad_0 = const()[name = tensor("v_5_pad_0"), val = tensor([0, 0])]; + tensor v_5_dilations_0 = const()[name = tensor("v_5_dilations_0"), val = tensor([1])]; + tensor v_5_groups_0 = const()[name = tensor("v_5_groups_0"), val = tensor(1)]; + tensor v_5 = conv(bias = mimi_decoder_model_3_block_3_conv_bias, dilations = v_5_dilations_0, groups = v_5_groups_0, pad = v_5_pad_0, pad_type = v_5_pad_type_0, strides = v_5_strides_0, weight = mimi_decoder_model_3_block_3_conv_weight, x = x_21)[name = tensor("v_5")]; + tensor input_35 = add(x = input_29, y = v_5)[name = tensor("input_35")]; + tensor input_37 = elu(alpha = var_507, x = input_35)[name = tensor("input_37")]; + tensor y_9_pad_type_0 = const()[name = tensor("y_9_pad_type_0"), val = tensor("valid")]; + tensor y_9_strides_0 = const()[name = tensor("y_9_strides_0"), val = tensor([5])]; + tensor y_9_pad_0 = const()[name = tensor("y_9_pad_0"), val = tensor([0, 0])]; + tensor y_9_dilations_0 = const()[name = tensor("y_9_dilations_0"), val = tensor([1])]; + tensor y_9_groups_0 = const()[name = tensor("y_9_groups_0"), val = tensor(1)]; + tensor y_9_has_output_shape_output_shape_0 = const()[name = tensor("y_9_has_output_shape_output_shape_0"), val = tensor([1, 128, 485])]; + tensor y_9_has_output_shape = conv_transpose(bias = mimi_decoder_model_5_convtr_bias, dilations = y_9_dilations_0, groups = y_9_groups_0, output_shape = y_9_has_output_shape_output_shape_0, pad = y_9_pad_0, pad_type = y_9_pad_type_0, strides = y_9_strides_0, weight = mimi_decoder_model_5_convtr_weight, x = input_37)[name = tensor("y_9_has_output_shape")]; + tensor var_613_begin_0 = const()[name = tensor("op_613_begin_0"), val = tensor([0, 0, 0])]; + tensor var_613_end_0 = const()[name = tensor("op_613_end_0"), val = tensor([1, 128, 5])]; + tensor var_613_end_mask_0 = const()[name = tensor("op_613_end_mask_0"), val = tensor([true, true, false])]; + tensor var_613 = slice_by_index(begin = var_613_begin_0, end = var_613_end_0, end_mask = var_613_end_mask_0, x = y_9_has_output_shape)[name = tensor("op_613")]; + tensor var_614 = add(x = var_613, y = convtr1_partial)[name = tensor("op_614")]; + tensor var_615_begin_0 = const()[name = tensor("op_615_begin_0"), val = tensor([0, 0, 5])]; + tensor var_615_end_0 = const()[name = tensor("op_615_end_0"), val = tensor([1, 128, 485])]; + tensor var_615_end_mask_0 = const()[name = tensor("op_615_end_mask_0"), val = tensor([true, true, true])]; + tensor var_615 = slice_by_index(begin = var_615_begin_0, end = var_615_end_0, end_mask = var_615_end_mask_0, x = y_9_has_output_shape)[name = tensor("op_615")]; + tensor y_11_interleave_0 = const()[name = tensor("y_11_interleave_0"), val = tensor(false)]; + tensor y_11 = concat(axis = var_508, interleave = y_11_interleave_0, values = (var_614, var_615))[name = tensor("y_11")]; + tensor new_partial_3_begin_0 = const()[name = tensor("new_partial_3_begin_0"), val = tensor([0, 0, 480])]; + tensor new_partial_3_end_0 = const()[name = tensor("new_partial_3_end_0"), val = tensor([1, 128, 485])]; + tensor new_partial_3_end_mask_0 = const()[name = tensor("new_partial_3_end_mask_0"), val = tensor([true, true, true])]; + tensor new_partial_3 = slice_by_index(begin = new_partial_3_begin_0, end = new_partial_3_end_0, end_mask = new_partial_3_end_mask_0, x = y_11)[name = tensor("new_partial_3")]; + tensor var_620 = const()[name = tensor("op_620"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(41767360)))]; + tensor var_621 = sub(x = new_partial_3, y = var_620)[name = tensor("op_621")]; + tensor input_39_begin_0 = const()[name = tensor("input_39_begin_0"), val = tensor([0, 0, 0])]; + tensor input_39_end_0 = const()[name = tensor("input_39_end_0"), val = tensor([1, 128, 480])]; + tensor input_39_end_mask_0 = const()[name = tensor("input_39_end_mask_0"), val = tensor([true, true, false])]; + tensor input_39 = slice_by_index(begin = input_39_begin_0, end = input_39_end_0, end_mask = input_39_end_mask_0, x = y_11)[name = tensor("input_39")]; + tensor x_23 = elu(alpha = var_507, x = input_39)[name = tensor("x_23")]; + tensor input_41_interleave_0 = const()[name = tensor("input_41_interleave_0"), val = tensor(false)]; + tensor input_41 = concat(axis = var_508, interleave = input_41_interleave_0, values = (res1_conv0_prev, x_23))[name = tensor("input_41")]; + tensor input_43_pad_type_0 = const()[name = tensor("input_43_pad_type_0"), val = tensor("valid")]; + tensor input_43_strides_0 = const()[name = tensor("input_43_strides_0"), val = tensor([1])]; + tensor input_43_pad_0 = const()[name = tensor("input_43_pad_0"), val = tensor([0, 0])]; + tensor input_43_dilations_0 = const()[name = tensor("input_43_dilations_0"), val = tensor([1])]; + tensor input_43_groups_0 = const()[name = tensor("input_43_groups_0"), val = tensor(1)]; + tensor input_43 = conv(bias = mimi_decoder_model_6_block_1_conv_bias, dilations = input_43_dilations_0, groups = input_43_groups_0, pad = input_43_pad_0, pad_type = input_43_pad_type_0, strides = input_43_strides_0, weight = mimi_decoder_model_6_block_1_conv_weight, x = input_41)[name = tensor("input_43")]; + tensor var_641_begin_0 = const()[name = tensor("op_641_begin_0"), val = tensor([0, 0, 480])]; + tensor var_641_end_0 = const()[name = tensor("op_641_end_0"), val = tensor([1, 128, 482])]; + tensor var_641_end_mask_0 = const()[name = tensor("op_641_end_mask_0"), val = tensor([true, true, true])]; + tensor var_641 = slice_by_index(begin = var_641_begin_0, end = var_641_end_0, end_mask = var_641_end_mask_0, x = input_41)[name = tensor("op_641")]; + tensor x_25 = elu(alpha = var_507, x = input_43)[name = tensor("x_25")]; + tensor v_7_pad_type_0 = const()[name = tensor("v_7_pad_type_0"), val = tensor("valid")]; + tensor v_7_strides_0 = const()[name = tensor("v_7_strides_0"), val = tensor([1])]; + tensor v_7_pad_0 = const()[name = tensor("v_7_pad_0"), val = tensor([0, 0])]; + tensor v_7_dilations_0 = const()[name = tensor("v_7_dilations_0"), val = tensor([1])]; + tensor v_7_groups_0 = const()[name = tensor("v_7_groups_0"), val = tensor(1)]; + tensor v_7 = conv(bias = mimi_decoder_model_6_block_3_conv_bias, dilations = v_7_dilations_0, groups = v_7_groups_0, pad = v_7_pad_0, pad_type = v_7_pad_type_0, strides = v_7_strides_0, weight = mimi_decoder_model_6_block_3_conv_weight, x = x_25)[name = tensor("v_7")]; + tensor input_45 = add(x = input_39, y = v_7)[name = tensor("input_45")]; + tensor input_47 = elu(alpha = var_507, x = input_45)[name = tensor("input_47")]; + tensor y_13_pad_type_0 = const()[name = tensor("y_13_pad_type_0"), val = tensor("valid")]; + tensor y_13_strides_0 = const()[name = tensor("y_13_strides_0"), val = tensor([4])]; + tensor y_13_pad_0 = const()[name = tensor("y_13_pad_0"), val = tensor([0, 0])]; + tensor y_13_dilations_0 = const()[name = tensor("y_13_dilations_0"), val = tensor([1])]; + tensor y_13_groups_0 = const()[name = tensor("y_13_groups_0"), val = tensor(1)]; + tensor y_13_has_output_shape_output_shape_0 = const()[name = tensor("y_13_has_output_shape_output_shape_0"), val = tensor([1, 64, 1924])]; + tensor y_13_has_output_shape = conv_transpose(bias = mimi_decoder_model_8_convtr_bias, dilations = y_13_dilations_0, groups = y_13_groups_0, output_shape = y_13_has_output_shape_output_shape_0, pad = y_13_pad_0, pad_type = y_13_pad_type_0, strides = y_13_strides_0, weight = mimi_decoder_model_8_convtr_weight, x = input_47)[name = tensor("y_13_has_output_shape")]; + tensor var_669_begin_0 = const()[name = tensor("op_669_begin_0"), val = tensor([0, 0, 0])]; + tensor var_669_end_0 = const()[name = tensor("op_669_end_0"), val = tensor([1, 64, 4])]; + tensor var_669_end_mask_0 = const()[name = tensor("op_669_end_mask_0"), val = tensor([true, true, false])]; + tensor var_669 = slice_by_index(begin = var_669_begin_0, end = var_669_end_0, end_mask = var_669_end_mask_0, x = y_13_has_output_shape)[name = tensor("op_669")]; + tensor var_670 = add(x = var_669, y = convtr2_partial)[name = tensor("op_670")]; + tensor var_671_begin_0 = const()[name = tensor("op_671_begin_0"), val = tensor([0, 0, 4])]; + tensor var_671_end_0 = const()[name = tensor("op_671_end_0"), val = tensor([1, 64, 1924])]; + tensor var_671_end_mask_0 = const()[name = tensor("op_671_end_mask_0"), val = tensor([true, true, true])]; + tensor var_671 = slice_by_index(begin = var_671_begin_0, end = var_671_end_0, end_mask = var_671_end_mask_0, x = y_13_has_output_shape)[name = tensor("op_671")]; + tensor y_interleave_0 = const()[name = tensor("y_interleave_0"), val = tensor(false)]; + tensor y = concat(axis = var_508, interleave = y_interleave_0, values = (var_670, var_671))[name = tensor("y")]; + tensor new_partial_begin_0 = const()[name = tensor("new_partial_begin_0"), val = tensor([0, 0, 1920])]; + tensor new_partial_end_0 = const()[name = tensor("new_partial_end_0"), val = tensor([1, 64, 1924])]; + tensor new_partial_end_mask_0 = const()[name = tensor("new_partial_end_mask_0"), val = tensor([true, true, true])]; + tensor new_partial = slice_by_index(begin = new_partial_begin_0, end = new_partial_end_0, end_mask = new_partial_end_mask_0, x = y)[name = tensor("new_partial")]; + tensor var_676 = const()[name = tensor("op_676"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(41767936)))]; + tensor var_677 = sub(x = new_partial, y = var_676)[name = tensor("op_677")]; + tensor input_49_begin_0 = const()[name = tensor("input_49_begin_0"), val = tensor([0, 0, 0])]; + tensor input_49_end_0 = const()[name = tensor("input_49_end_0"), val = tensor([1, 64, 1920])]; + tensor input_49_end_mask_0 = const()[name = tensor("input_49_end_mask_0"), val = tensor([true, true, false])]; + tensor input_49 = slice_by_index(begin = input_49_begin_0, end = input_49_end_0, end_mask = input_49_end_mask_0, x = y)[name = tensor("input_49")]; + tensor x_27 = elu(alpha = var_507, x = input_49)[name = tensor("x_27")]; + tensor input_51_interleave_0 = const()[name = tensor("input_51_interleave_0"), val = tensor(false)]; + tensor input_51 = concat(axis = var_508, interleave = input_51_interleave_0, values = (res2_conv0_prev, x_27))[name = tensor("input_51")]; + tensor input_53_pad_type_0 = const()[name = tensor("input_53_pad_type_0"), val = tensor("valid")]; + tensor input_53_strides_0 = const()[name = tensor("input_53_strides_0"), val = tensor([1])]; + tensor input_53_pad_0 = const()[name = tensor("input_53_pad_0"), val = tensor([0, 0])]; + tensor input_53_dilations_0 = const()[name = tensor("input_53_dilations_0"), val = tensor([1])]; + tensor input_53_groups_0 = const()[name = tensor("input_53_groups_0"), val = tensor(1)]; + tensor input_53 = conv(bias = mimi_decoder_model_9_block_1_conv_bias, dilations = input_53_dilations_0, groups = input_53_groups_0, pad = input_53_pad_0, pad_type = input_53_pad_type_0, strides = input_53_strides_0, weight = mimi_decoder_model_9_block_1_conv_weight, x = input_51)[name = tensor("input_53")]; + tensor var_697_begin_0 = const()[name = tensor("op_697_begin_0"), val = tensor([0, 0, 1920])]; + tensor var_697_end_0 = const()[name = tensor("op_697_end_0"), val = tensor([1, 64, 1922])]; + tensor var_697_end_mask_0 = const()[name = tensor("op_697_end_mask_0"), val = tensor([true, true, true])]; + tensor var_697 = slice_by_index(begin = var_697_begin_0, end = var_697_end_0, end_mask = var_697_end_mask_0, x = input_51)[name = tensor("op_697")]; + tensor x_29 = elu(alpha = var_507, x = input_53)[name = tensor("x_29")]; + tensor v_pad_type_0 = const()[name = tensor("v_pad_type_0"), val = tensor("valid")]; + tensor v_strides_0 = const()[name = tensor("v_strides_0"), val = tensor([1])]; + tensor v_pad_0 = const()[name = tensor("v_pad_0"), val = tensor([0, 0])]; + tensor v_dilations_0 = const()[name = tensor("v_dilations_0"), val = tensor([1])]; + tensor v_groups_0 = const()[name = tensor("v_groups_0"), val = tensor(1)]; + tensor v = conv(bias = mimi_decoder_model_9_block_3_conv_bias, dilations = v_dilations_0, groups = v_groups_0, pad = v_pad_0, pad_type = v_pad_type_0, strides = v_strides_0, weight = mimi_decoder_model_9_block_3_conv_weight, x = x_29)[name = tensor("v")]; + tensor input_55 = add(x = input_49, y = v)[name = tensor("input_55")]; + tensor x = elu(alpha = var_507, x = input_55)[name = tensor("x")]; + tensor input_interleave_0 = const()[name = tensor("input_interleave_0"), val = tensor(false)]; + tensor input = concat(axis = var_508, interleave = input_interleave_0, values = (conv_final_prev, x))[name = tensor("input")]; + tensor var_724_pad_type_0 = const()[name = tensor("op_724_pad_type_0"), val = tensor("valid")]; + tensor var_724_strides_0 = const()[name = tensor("op_724_strides_0"), val = tensor([1])]; + tensor var_724_pad_0 = const()[name = tensor("op_724_pad_0"), val = tensor([0, 0])]; + tensor var_724_dilations_0 = const()[name = tensor("op_724_dilations_0"), val = tensor([1])]; + tensor var_724_groups_0 = const()[name = tensor("op_724_groups_0"), val = tensor(1)]; + tensor var_724 = conv(bias = mimi_decoder_model_11_conv_bias, dilations = var_724_dilations_0, groups = var_724_groups_0, pad = var_724_pad_0, pad_type = var_724_pad_type_0, strides = var_724_strides_0, weight = mimi_decoder_model_11_conv_weight, x = input)[name = tensor("op_724")]; + tensor var_725_begin_0 = const()[name = tensor("op_725_begin_0"), val = tensor([0, 0, 1920])]; + tensor var_725_end_0 = const()[name = tensor("op_725_end_0"), val = tensor([1, 64, 1922])]; + tensor var_725_end_mask_0 = const()[name = tensor("op_725_end_mask_0"), val = tensor([true, true, true])]; + tensor var_725 = slice_by_index(begin = var_725_begin_0, end = var_725_end_0, end_mask = var_725_end_mask_0, x = input)[name = tensor("op_725")]; + tensor var_740_promoted = const()[name = tensor("op_740_promoted"), val = tensor(0x1p+4)]; + tensor var_741 = add(x = attn0_offset, y = var_740_promoted)[name = tensor("op_741")]; + tensor var_743_promoted = const()[name = tensor("op_743_promoted"), val = tensor(0x1p+4)]; + tensor var_744 = add(x = attn1_offset, y = var_743_promoted)[name = tensor("op_744")]; + tensor conv0_first_tmp = identity(x = conv0_first)[name = tensor("conv0_first_tmp")]; + tensor res0_conv0_first_tmp = identity(x = res0_conv0_first)[name = tensor("res0_conv0_first_tmp")]; + tensor res0_conv1_prev_tmp = identity(x = res0_conv1_prev)[name = tensor("res0_conv1_prev_tmp")]; + tensor res0_conv1_first_tmp = identity(x = res0_conv1_first)[name = tensor("res0_conv1_first_tmp")]; + tensor res1_conv0_first_tmp = identity(x = res1_conv0_first)[name = tensor("res1_conv0_first_tmp")]; + tensor res1_conv1_prev_tmp = identity(x = res1_conv1_prev)[name = tensor("res1_conv1_prev_tmp")]; + tensor res1_conv1_first_tmp = identity(x = res1_conv1_first)[name = tensor("res1_conv1_first_tmp")]; + tensor res2_conv0_first_tmp = identity(x = res2_conv0_first)[name = tensor("res2_conv0_first_tmp")]; + tensor res2_conv1_prev_tmp = identity(x = res2_conv1_prev)[name = tensor("res2_conv1_prev_tmp")]; + tensor res2_conv1_first_tmp = identity(x = res2_conv1_first)[name = tensor("res2_conv1_first_tmp")]; + tensor conv_final_first_tmp = identity(x = conv_final_first)[name = tensor("conv_final_first_tmp")]; + } -> (var_724, var_77, var_210, var_741, var_400, var_744, var_542, conv0_first, var_565, var_585, res0_conv0_first, res0_conv1_prev, res0_conv1_first, var_621, var_641, res1_conv0_first, res1_conv1_prev, res1_conv1_first, var_677, var_697, res2_conv0_first, res2_conv1_prev, res2_conv1_first, var_725, conv_final_first); +} \ No newline at end of file diff --git a/v2.1/portuguese/mimi_decoder.mlmodelc/weights/weight.bin b/v2.1/portuguese/mimi_decoder.mlmodelc/weights/weight.bin new file mode 100644 index 0000000000000000000000000000000000000000..1370ffd7f5e8989e5db9d220d47f098f2b1235d8 --- /dev/null +++ b/v2.1/portuguese/mimi_decoder.mlmodelc/weights/weight.bin @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:1d5e99ea6732f93fdf15a439fb22d32172021a27cc779fd32b77f79f06a286b0 +size 41768256 diff --git a/v2.1/portuguese/mimi_decoder.mlpackage/Data/com.apple.CoreML/model.mlmodel b/v2.1/portuguese/mimi_decoder.mlpackage/Data/com.apple.CoreML/model.mlmodel new file mode 100644 index 0000000000000000000000000000000000000000..8c92021f7e3a2fa48b9d23e81b066b5730f0d362 --- /dev/null +++ b/v2.1/portuguese/mimi_decoder.mlpackage/Data/com.apple.CoreML/model.mlmodel @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:ab84f4d8c6260efbde7119b2fcdd0cb5750c39f3198a2af93609701ce52156b4 +size 93270 diff --git a/v2.1/portuguese/mimi_decoder.mlpackage/Data/com.apple.CoreML/weights/weight.bin b/v2.1/portuguese/mimi_decoder.mlpackage/Data/com.apple.CoreML/weights/weight.bin new file mode 100644 index 0000000000000000000000000000000000000000..1370ffd7f5e8989e5db9d220d47f098f2b1235d8 --- /dev/null +++ b/v2.1/portuguese/mimi_decoder.mlpackage/Data/com.apple.CoreML/weights/weight.bin @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:1d5e99ea6732f93fdf15a439fb22d32172021a27cc779fd32b77f79f06a286b0 +size 41768256 diff --git a/v2.1/portuguese/mimi_decoder.mlpackage/Manifest.json b/v2.1/portuguese/mimi_decoder.mlpackage/Manifest.json new file mode 100644 index 0000000000000000000000000000000000000000..e64e3108bfd45262c8463bd08abe3643ca9293bb --- /dev/null +++ b/v2.1/portuguese/mimi_decoder.mlpackage/Manifest.json @@ -0,0 +1,18 @@ +{ + "fileFormatVersion": "1.0.0", + "itemInfoEntries": { + "11AFB38D-3937-450E-882C-A9D58D91490B": { + "author": "com.apple.CoreML", + "description": "CoreML Model Specification", + "name": "model.mlmodel", + "path": "com.apple.CoreML/model.mlmodel" + }, + "F3CCFF6B-7AC5-4F04-BB50-8B56A38427F3": { + "author": "com.apple.CoreML", + "description": "CoreML Model Weights", + "name": "weights", + "path": "com.apple.CoreML/weights" + } + }, + "rootModelIdentifier": "11AFB38D-3937-450E-882C-A9D58D91490B" +} diff --git a/v2.1/portuguese_24l/cond_prefill.mlmodelc/analytics/coremldata.bin b/v2.1/portuguese_24l/cond_prefill.mlmodelc/analytics/coremldata.bin new file mode 100644 index 0000000000000000000000000000000000000000..5d9f8b780c7cd9679bd2f082c048bcf963eedb37 --- /dev/null +++ b/v2.1/portuguese_24l/cond_prefill.mlmodelc/analytics/coremldata.bin @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:3b94b59259b2df1a7104eb8216d55d8dad559db9869febc67b8d23f43d6d8b3d +size 243 diff --git a/v2.1/portuguese_24l/cond_prefill.mlmodelc/coremldata.bin b/v2.1/portuguese_24l/cond_prefill.mlmodelc/coremldata.bin new file mode 100644 index 0000000000000000000000000000000000000000..f2467fde537f1a38166f68e8f684a9835e1334d2 --- /dev/null +++ b/v2.1/portuguese_24l/cond_prefill.mlmodelc/coremldata.bin @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:01395552022603517ff29e34ff21ffda9de2336b33b22bd38ce1f43a2037492a +size 3519 diff --git a/v2.1/portuguese_24l/cond_prefill.mlmodelc/model.mil b/v2.1/portuguese_24l/cond_prefill.mlmodelc/model.mil new file mode 100644 index 0000000000000000000000000000000000000000..7028cd94630c33bb5143ac20b6b86ca016eb1d25 --- /dev/null +++ b/v2.1/portuguese_24l/cond_prefill.mlmodelc/model.mil @@ -0,0 +1,5431 @@ +program(1.0) +[buildInfo = dict, tensor>({{"coremlc-component-MIL", "3520.4.1"}, {"coremlc-version", "3520.5.1"}, {"coremltools-component-torch", "2.12.0"}, {"coremltools-source-dialect", "TorchScript"}, {"coremltools-version", "9.0"}})] +{ + func main(tensor cache0, tensor cache1, tensor cache10, tensor cache11, tensor cache12, tensor cache13, tensor cache14, tensor cache15, tensor cache16, tensor cache17, tensor cache18, tensor cache19, tensor cache2, tensor cache20, tensor cache21, tensor cache22, tensor cache23, tensor cache3, tensor cache4, tensor cache5, tensor cache6, tensor cache7, tensor cache8, tensor cache9, tensor conditioning, tensor position0, tensor position1, tensor position10, tensor position11, tensor position12, tensor position13, tensor position14, tensor position15, tensor position16, tensor position17, tensor position18, tensor position19, tensor position2, tensor position20, tensor position21, tensor position22, tensor position23, tensor position3, tensor position4, tensor position5, tensor position6, tensor position7, tensor position8, tensor position9, tensor valid_len) { + tensor input_3_axes_0 = const()[name = tensor("input_3_axes_0"), val = tensor([-1])]; + tensor conditioning_to_fp16_dtype_0 = const()[name = tensor("conditioning_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor norm0_1_weight_to_fp16 = const()[name = tensor("norm0_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(64)))]; + tensor norm0_1_bias_to_fp16 = const()[name = tensor("norm0_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(2176)))]; + tensor var_192_to_fp16 = const()[name = tensor("op_192_to_fp16"), val = tensor(0x1.5p-17)]; + tensor conditioning_to_fp16 = cast(dtype = conditioning_to_fp16_dtype_0, x = conditioning)[name = tensor("cast_663")]; + tensor input_3_cast_fp16 = layer_norm(axes = input_3_axes_0, beta = norm0_1_bias_to_fp16, epsilon = var_192_to_fp16, gamma = norm0_1_weight_to_fp16, x = conditioning_to_fp16)[name = tensor("input_3_cast_fp16")]; + tensor attn0_in_proj_weight_to_fp16 = const()[name = tensor("attn0_in_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(4288)))]; + tensor linear_0_bias_0_to_fp16 = const()[name = tensor("linear_0_bias_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(6295808)))]; + tensor linear_0_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = attn0_in_proj_weight_to_fp16, x = input_3_cast_fp16)[name = tensor("linear_0_cast_fp16")]; + tensor var_206 = const()[name = tensor("op_206"), val = tensor([1, 256, 3, 16, 64])]; + tensor qkv_1_cast_fp16 = reshape(shape = var_206, x = linear_0_cast_fp16)[name = tensor("qkv_1_cast_fp16")]; + tensor q_1_begin_0 = const()[name = tensor("q_1_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor q_1_end_0 = const()[name = tensor("q_1_end_0"), val = tensor([1, 256, 1, 16, 64])]; + tensor q_1_end_mask_0 = const()[name = tensor("q_1_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor q_1_squeeze_mask_0 = const()[name = tensor("q_1_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor q_1_cast_fp16 = slice_by_index(begin = q_1_begin_0, end = q_1_end_0, end_mask = q_1_end_mask_0, squeeze_mask = q_1_squeeze_mask_0, x = qkv_1_cast_fp16)[name = tensor("q_1_cast_fp16")]; + tensor k_1_begin_0 = const()[name = tensor("k_1_begin_0"), val = tensor([0, 0, 1, 0, 0])]; + tensor k_1_end_0 = const()[name = tensor("k_1_end_0"), val = tensor([1, 256, 2, 16, 64])]; + tensor k_1_end_mask_0 = const()[name = tensor("k_1_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor k_1_squeeze_mask_0 = const()[name = tensor("k_1_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor k_1_cast_fp16 = slice_by_index(begin = k_1_begin_0, end = k_1_end_0, end_mask = k_1_end_mask_0, squeeze_mask = k_1_squeeze_mask_0, x = qkv_1_cast_fp16)[name = tensor("k_1_cast_fp16")]; + tensor v_1_begin_0 = const()[name = tensor("v_1_begin_0"), val = tensor([0, 0, 2, 0, 0])]; + tensor v_1_end_0 = const()[name = tensor("v_1_end_0"), val = tensor([1, 256, 3, 16, 64])]; + tensor v_1_end_mask_0 = const()[name = tensor("v_1_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor v_1_squeeze_mask_0 = const()[name = tensor("v_1_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor v_1_cast_fp16 = slice_by_index(begin = v_1_begin_0, end = v_1_end_0, end_mask = v_1_end_mask_0, squeeze_mask = v_1_squeeze_mask_0, x = qkv_1_cast_fp16)[name = tensor("v_1_cast_fp16")]; + tensor var_264 = const()[name = tensor("op_264"), val = tensor([1, 1])]; + tensor position0_to_fp16_dtype_0 = const()[name = tensor("position0_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor position0_to_fp16 = cast(dtype = position0_to_fp16_dtype_0, x = position0)[name = tensor("cast_662")]; + tensor var_265_cast_fp16 = reshape(shape = var_264, x = position0_to_fp16)[name = tensor("op_265_cast_fp16")]; + tensor ts_1_promoted_to_fp16 = const()[name = tensor("ts_1_promoted_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(6302016)))]; + tensor ts_3_cast_fp16 = add(x = ts_1_promoted_to_fp16, y = var_265_cast_fp16)[name = tensor("ts_3_cast_fp16")]; + tensor var_272 = const()[name = tensor("op_272"), val = tensor([1, 256, 1, 1])]; + tensor ts_5_cast_fp16 = reshape(shape = var_272, x = ts_3_cast_fp16)[name = tensor("ts_5_cast_fp16")]; + tensor var_279 = const()[name = tensor("op_279"), val = tensor([1, 256, 16, 32, 2])]; + tensor q_complex_1_cast_fp16 = reshape(shape = var_279, x = q_1_cast_fp16)[name = tensor("q_complex_1_cast_fp16")]; + tensor var_286 = const()[name = tensor("op_286"), val = tensor([1, 256, 16, 32, 2])]; + tensor k_complex_1_cast_fp16 = reshape(shape = var_286, x = k_1_cast_fp16)[name = tensor("k_complex_1_cast_fp16")]; + tensor var_290_begin_0 = const()[name = tensor("op_290_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_290_end_0 = const()[name = tensor("op_290_end_0"), val = tensor([1, 256, 16, 32, 1])]; + tensor var_290_end_mask_0 = const()[name = tensor("op_290_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_290_squeeze_mask_0 = const()[name = tensor("op_290_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_290_cast_fp16 = slice_by_index(begin = var_290_begin_0, end = var_290_end_0, end_mask = var_290_end_mask_0, squeeze_mask = var_290_squeeze_mask_0, x = q_complex_1_cast_fp16)[name = tensor("op_290_cast_fp16")]; + tensor var_298_begin_0 = const()[name = tensor("op_298_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_298_end_0 = const()[name = tensor("op_298_end_0"), val = tensor([1, 256, 16, 32, 2])]; + tensor var_298_end_mask_0 = const()[name = tensor("op_298_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_298_squeeze_mask_0 = const()[name = tensor("op_298_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_298_cast_fp16 = slice_by_index(begin = var_298_begin_0, end = var_298_end_0, end_mask = var_298_end_mask_0, squeeze_mask = var_298_squeeze_mask_0, x = q_complex_1_cast_fp16)[name = tensor("op_298_cast_fp16")]; + tensor var_306_begin_0 = const()[name = tensor("op_306_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_306_end_0 = const()[name = tensor("op_306_end_0"), val = tensor([1, 256, 16, 32, 1])]; + tensor var_306_end_mask_0 = const()[name = tensor("op_306_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_306_squeeze_mask_0 = const()[name = tensor("op_306_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_306_cast_fp16 = slice_by_index(begin = var_306_begin_0, end = var_306_end_0, end_mask = var_306_end_mask_0, squeeze_mask = var_306_squeeze_mask_0, x = k_complex_1_cast_fp16)[name = tensor("op_306_cast_fp16")]; + tensor var_314_begin_0 = const()[name = tensor("op_314_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_314_end_0 = const()[name = tensor("op_314_end_0"), val = tensor([1, 256, 16, 32, 2])]; + tensor var_314_end_mask_0 = const()[name = tensor("op_314_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_314_squeeze_mask_0 = const()[name = tensor("op_314_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_314_cast_fp16 = slice_by_index(begin = var_314_begin_0, end = var_314_end_0, end_mask = var_314_end_mask_0, squeeze_mask = var_314_squeeze_mask_0, x = k_complex_1_cast_fp16)[name = tensor("op_314_cast_fp16")]; + tensor freqs_1_to_fp16 = const()[name = tensor("freqs_1_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(6302592)))]; + tensor var_320_cast_fp16 = mul(x = freqs_1_to_fp16, y = ts_5_cast_fp16)[name = tensor("op_320_cast_fp16")]; + tensor rotr_1_cast_fp16 = cos(x = var_320_cast_fp16)[name = tensor("rotr_1_cast_fp16")]; + tensor roti_1_cast_fp16 = sin(x = var_320_cast_fp16)[name = tensor("roti_1_cast_fp16")]; + tensor var_324_cast_fp16 = mul(x = var_290_cast_fp16, y = rotr_1_cast_fp16)[name = tensor("op_324_cast_fp16")]; + tensor var_325_cast_fp16 = mul(x = var_298_cast_fp16, y = roti_1_cast_fp16)[name = tensor("op_325_cast_fp16")]; + tensor qor_1_cast_fp16 = sub(x = var_324_cast_fp16, y = var_325_cast_fp16)[name = tensor("qor_1_cast_fp16")]; + tensor var_328_cast_fp16 = mul(x = var_290_cast_fp16, y = roti_1_cast_fp16)[name = tensor("op_328_cast_fp16")]; + tensor var_329_cast_fp16 = mul(x = var_298_cast_fp16, y = rotr_1_cast_fp16)[name = tensor("op_329_cast_fp16")]; + tensor qoi_1_cast_fp16 = add(x = var_328_cast_fp16, y = var_329_cast_fp16)[name = tensor("qoi_1_cast_fp16")]; + tensor var_332_cast_fp16 = mul(x = var_306_cast_fp16, y = rotr_1_cast_fp16)[name = tensor("op_332_cast_fp16")]; + tensor var_333_cast_fp16 = mul(x = var_314_cast_fp16, y = roti_1_cast_fp16)[name = tensor("op_333_cast_fp16")]; + tensor kor_1_cast_fp16 = sub(x = var_332_cast_fp16, y = var_333_cast_fp16)[name = tensor("kor_1_cast_fp16")]; + tensor var_336_cast_fp16 = mul(x = var_306_cast_fp16, y = roti_1_cast_fp16)[name = tensor("op_336_cast_fp16")]; + tensor var_337_cast_fp16 = mul(x = var_314_cast_fp16, y = rotr_1_cast_fp16)[name = tensor("op_337_cast_fp16")]; + tensor koi_1_cast_fp16 = add(x = var_336_cast_fp16, y = var_337_cast_fp16)[name = tensor("koi_1_cast_fp16")]; + tensor qo_1_axis_0 = const()[name = tensor("qo_1_axis_0"), val = tensor(-1)]; + tensor qo_1_cast_fp16 = stack(axis = qo_1_axis_0, values = (qor_1_cast_fp16, qoi_1_cast_fp16))[name = tensor("qo_1_cast_fp16")]; + tensor ko_1_axis_0 = const()[name = tensor("ko_1_axis_0"), val = tensor(-1)]; + tensor ko_1_cast_fp16 = stack(axis = ko_1_axis_0, values = (kor_1_cast_fp16, koi_1_cast_fp16))[name = tensor("ko_1_cast_fp16")]; + tensor var_370 = const()[name = tensor("op_370"), val = tensor([1, 256, 16, 64])]; + tensor q_3_cast_fp16 = reshape(shape = var_370, x = qo_1_cast_fp16)[name = tensor("q_3_cast_fp16")]; + tensor var_376 = const()[name = tensor("op_376"), val = tensor([1, 256, 16, 64])]; + tensor k_3_cast_fp16 = reshape(shape = var_376, x = ko_1_cast_fp16)[name = tensor("k_3_cast_fp16")]; + tensor var_393_promoted_to_fp16 = const()[name = tensor("op_393_promoted_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(6302720)))]; + tensor write_indices_float_1_cast_fp16 = add(x = var_265_cast_fp16, y = var_393_promoted_to_fp16)[name = tensor("write_indices_float_1_cast_fp16")]; + tensor _inversed_397_y_0_to_fp16 = const()[name = tensor("_inversed_397_y_0_to_fp16"), val = tensor(0x1p-9)]; + tensor _inversed_397_cast_fp16 = mul(x = write_indices_float_1_cast_fp16, y = _inversed_397_y_0_to_fp16)[name = tensor("_inversed_397_cast_fp16")]; + tensor var_398_cast_fp16 = floor(x = _inversed_397_cast_fp16)[name = tensor("op_398_cast_fp16")]; + tensor var_399_to_fp16 = const()[name = tensor("op_399_to_fp16"), val = tensor(0x1p+9)]; + tensor var_400_cast_fp16 = mul(x = var_398_cast_fp16, y = var_399_to_fp16)[name = tensor("op_400_cast_fp16")]; + tensor write_indices_float_3_cast_fp16 = sub(x = write_indices_float_1_cast_fp16, y = var_400_cast_fp16)[name = tensor("write_indices_float_3_cast_fp16")]; + tensor var_409 = const()[name = tensor("op_409"), val = tensor([1, 1])]; + tensor valid_len_to_fp16_dtype_0 = const()[name = tensor("valid_len_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor valid_len_to_fp16 = cast(dtype = valid_len_to_fp16_dtype_0, x = valid_len)[name = tensor("cast_661")]; + tensor var_410_cast_fp16 = reshape(shape = var_409, x = valid_len_to_fp16)[name = tensor("op_410_cast_fp16")]; + tensor token_valid_1_cast_fp16 = less(x = var_393_promoted_to_fp16, y = var_410_cast_fp16)[name = tensor("token_valid_1_cast_fp16")]; + tensor var_418_value_0_to_fp16 = const()[name = tensor("op_418_value_0_to_fp16"), val = tensor(0x1.ffp+8)]; + tensor var_418_cast_fp16 = fill_like(ref_tensor = write_indices_float_3_cast_fp16, value = var_418_value_0_to_fp16)[name = tensor("op_418_cast_fp16")]; + tensor write_indices_float_5_cast_fp16 = select(a = write_indices_float_3_cast_fp16, b = var_418_cast_fp16, cond = token_valid_1_cast_fp16)[name = tensor("write_indices_float_5_cast_fp16")]; + tensor var_424_dtype_0 = const()[name = tensor("op_424_dtype_0"), val = tensor("int32")]; + tensor var_429 = const()[name = tensor("op_429"), val = tensor([1, 256, 1, 1])]; + tensor write_indices_float_5_cast_fp16_to_int32 = cast(dtype = var_424_dtype_0, x = write_indices_float_5_cast_fp16)[name = tensor("cast_660")]; + tensor var_430 = reshape(shape = var_429, x = write_indices_float_5_cast_fp16_to_int32)[name = tensor("op_430")]; + tensor write_indices_1_reps_0 = const()[name = tensor("write_indices_1_reps_0"), val = tensor([1, 1, 16, 64])]; + tensor write_indices_1 = tile(reps = write_indices_1_reps_0, x = var_430)[name = tensor("write_indices_1")]; + tensor var_440_begin_0 = const()[name = tensor("op_440_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_440_end_0 = const()[name = tensor("op_440_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor var_440_end_mask_0 = const()[name = tensor("op_440_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_440_squeeze_mask_0 = const()[name = tensor("op_440_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor cache0_to_fp16_dtype_0 = const()[name = tensor("cache0_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor cache0_to_fp16 = cast(dtype = cache0_to_fp16_dtype_0, x = cache0)[name = tensor("cast_659")]; + tensor var_440_cast_fp16 = slice_by_index(begin = var_440_begin_0, end = var_440_end_0, end_mask = var_440_end_mask_0, squeeze_mask = var_440_squeeze_mask_0, x = cache0_to_fp16)[name = tensor("op_440_cast_fp16")]; + tensor var_442_axis_0 = const()[name = tensor("op_442_axis_0"), val = tensor(1)]; + tensor var_442_mode_0 = const()[name = tensor("op_442_mode_0"), val = tensor("update")]; + tensor var_442_validate_indices_0 = const()[name = tensor("op_442_validate_indices_0"), val = tensor(false)]; + tensor var_442_cast_fp16 = scatter_along_axis(axis = var_442_axis_0, data = var_440_cast_fp16, indices = write_indices_1, mode = var_442_mode_0, updates = k_3_cast_fp16, validate_indices = var_442_validate_indices_0)[name = tensor("op_442_cast_fp16")]; + tensor concat_2 = const()[name = tensor("concat_2"), val = tensor([0, 0, 0, 0, 0])]; + tensor concat_3 = const()[name = tensor("concat_3"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_1_internal_tensor_assign_1_stride_0 = const()[name = tensor("new_cache_1_internal_tensor_assign_1_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_1_internal_tensor_assign_1_begin_mask_0 = const()[name = tensor("new_cache_1_internal_tensor_assign_1_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_1_internal_tensor_assign_1_end_mask_0 = const()[name = tensor("new_cache_1_internal_tensor_assign_1_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_1_internal_tensor_assign_1_squeeze_mask_0 = const()[name = tensor("new_cache_1_internal_tensor_assign_1_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_46 = const()[name = tensor("shape_46"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_0 = const()[name = tensor("reduce_prod_0"), val = tensor(1048576)]; + tensor range_1d_0_start_0 = const()[name = tensor("range_1d_0_start_0"), val = tensor(0)]; + tensor range_1d_0_step_0 = const()[name = tensor("range_1d_0_step_0"), val = tensor(1)]; + tensor range_1d_0 = range_1d(end = reduce_prod_0, start = range_1d_0_start_0, step = range_1d_0_step_0)[name = tensor("range_1d_0")]; + tensor reshape_0 = reshape(shape = shape_46, x = range_1d_0)[name = tensor("reshape_0")]; + tensor slice_by_index_0 = slice_by_index(begin = concat_2, begin_mask = new_cache_1_internal_tensor_assign_1_begin_mask_0, end = concat_3, end_mask = new_cache_1_internal_tensor_assign_1_end_mask_0, squeeze_mask = new_cache_1_internal_tensor_assign_1_squeeze_mask_0, stride = new_cache_1_internal_tensor_assign_1_stride_0, x = reshape_0)[name = tensor("slice_by_index_0")]; + tensor reshape_1_shape_0 = const()[name = tensor("reshape_1_shape_0"), val = tensor([-1])]; + tensor reshape_1 = reshape(shape = reshape_1_shape_0, x = slice_by_index_0)[name = tensor("reshape_1")]; + tensor reshape_2_shape_0 = const()[name = tensor("reshape_2_shape_0"), val = tensor([-1])]; + tensor reshape_2_cast_fp16 = reshape(shape = reshape_2_shape_0, x = var_442_cast_fp16)[name = tensor("reshape_2_cast_fp16")]; + tensor reshape_3_shape_0 = const()[name = tensor("reshape_3_shape_0"), val = tensor([-1])]; + tensor reshape_3_cast_fp16 = reshape(shape = reshape_3_shape_0, x = cache0_to_fp16)[name = tensor("reshape_3_cast_fp16")]; + tensor scatter_0_mode_0 = const()[name = tensor("scatter_0_mode_0"), val = tensor("update")]; + tensor scatter_0_axis_0 = const()[name = tensor("scatter_0_axis_0"), val = tensor(0)]; + tensor scatter_0_validate_indices_0 = const()[name = tensor("scatter_0_validate_indices_0"), val = tensor(false)]; + tensor scatter_0_cast_fp16 = scatter(axis = scatter_0_axis_0, data = reshape_3_cast_fp16, indices = reshape_1, mode = scatter_0_mode_0, updates = reshape_2_cast_fp16, validate_indices = scatter_0_validate_indices_0)[name = tensor("scatter_0_cast_fp16")]; + tensor reshape_4_cast_fp16 = reshape(shape = shape_46, x = scatter_0_cast_fp16)[name = tensor("reshape_4_cast_fp16")]; + tensor var_450_begin_0 = const()[name = tensor("op_450_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor var_450_end_0 = const()[name = tensor("op_450_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor var_450_end_mask_0 = const()[name = tensor("op_450_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_450_squeeze_mask_0 = const()[name = tensor("op_450_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor var_450_cast_fp16 = slice_by_index(begin = var_450_begin_0, end = var_450_end_0, end_mask = var_450_end_mask_0, squeeze_mask = var_450_squeeze_mask_0, x = reshape_4_cast_fp16)[name = tensor("op_450_cast_fp16")]; + tensor var_452_axis_0 = const()[name = tensor("op_452_axis_0"), val = tensor(1)]; + tensor var_452_mode_0 = const()[name = tensor("op_452_mode_0"), val = tensor("update")]; + tensor var_452_validate_indices_0 = const()[name = tensor("op_452_validate_indices_0"), val = tensor(false)]; + tensor var_452_cast_fp16 = scatter_along_axis(axis = var_452_axis_0, data = var_450_cast_fp16, indices = write_indices_1, mode = var_452_mode_0, updates = v_1_cast_fp16, validate_indices = var_452_validate_indices_0)[name = tensor("op_452_cast_fp16")]; + tensor concat_4 = const()[name = tensor("concat_4"), val = tensor([1, 0, 0, 0, 0])]; + tensor concat_5 = const()[name = tensor("concat_5"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_1_internal_tensor_assign_2_stride_0 = const()[name = tensor("new_cache_1_internal_tensor_assign_2_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_1_internal_tensor_assign_2_begin_mask_0 = const()[name = tensor("new_cache_1_internal_tensor_assign_2_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_1_internal_tensor_assign_2_end_mask_0 = const()[name = tensor("new_cache_1_internal_tensor_assign_2_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_1_internal_tensor_assign_2_squeeze_mask_0 = const()[name = tensor("new_cache_1_internal_tensor_assign_2_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_47 = const()[name = tensor("shape_47"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_1 = const()[name = tensor("reduce_prod_1"), val = tensor(1048576)]; + tensor range_1d_1_start_0 = const()[name = tensor("range_1d_1_start_0"), val = tensor(0)]; + tensor range_1d_1_step_0 = const()[name = tensor("range_1d_1_step_0"), val = tensor(1)]; + tensor range_1d_1 = range_1d(end = reduce_prod_1, start = range_1d_1_start_0, step = range_1d_1_step_0)[name = tensor("range_1d_1")]; + tensor reshape_5 = reshape(shape = shape_47, x = range_1d_1)[name = tensor("reshape_5")]; + tensor slice_by_index_1 = slice_by_index(begin = concat_4, begin_mask = new_cache_1_internal_tensor_assign_2_begin_mask_0, end = concat_5, end_mask = new_cache_1_internal_tensor_assign_2_end_mask_0, squeeze_mask = new_cache_1_internal_tensor_assign_2_squeeze_mask_0, stride = new_cache_1_internal_tensor_assign_2_stride_0, x = reshape_5)[name = tensor("slice_by_index_1")]; + tensor reshape_6_shape_0 = const()[name = tensor("reshape_6_shape_0"), val = tensor([-1])]; + tensor reshape_6 = reshape(shape = reshape_6_shape_0, x = slice_by_index_1)[name = tensor("reshape_6")]; + tensor reshape_7_shape_0 = const()[name = tensor("reshape_7_shape_0"), val = tensor([-1])]; + tensor reshape_7_cast_fp16 = reshape(shape = reshape_7_shape_0, x = var_452_cast_fp16)[name = tensor("reshape_7_cast_fp16")]; + tensor reshape_8_shape_0 = const()[name = tensor("reshape_8_shape_0"), val = tensor([-1])]; + tensor reshape_8_cast_fp16 = reshape(shape = reshape_8_shape_0, x = reshape_4_cast_fp16)[name = tensor("reshape_8_cast_fp16")]; + tensor scatter_1_mode_0 = const()[name = tensor("scatter_1_mode_0"), val = tensor("update")]; + tensor scatter_1_axis_0 = const()[name = tensor("scatter_1_axis_0"), val = tensor(0)]; + tensor scatter_1_validate_indices_0 = const()[name = tensor("scatter_1_validate_indices_0"), val = tensor(false)]; + tensor scatter_1_cast_fp16 = scatter(axis = scatter_1_axis_0, data = reshape_8_cast_fp16, indices = reshape_6, mode = scatter_1_mode_0, updates = reshape_7_cast_fp16, validate_indices = scatter_1_validate_indices_0)[name = tensor("scatter_1_cast_fp16")]; + tensor reshape_9_cast_fp16 = reshape(shape = shape_47, x = scatter_1_cast_fp16)[name = tensor("reshape_9_cast_fp16")]; + tensor reshape_9_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("reshape_9_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor keys_1_begin_0 = const()[name = tensor("keys_1_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor keys_1_end_0 = const()[name = tensor("keys_1_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor keys_1_end_mask_0 = const()[name = tensor("keys_1_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor keys_1_squeeze_mask_0 = const()[name = tensor("keys_1_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor keys_1_cast_fp16 = slice_by_index(begin = keys_1_begin_0, end = keys_1_end_0, end_mask = keys_1_end_mask_0, squeeze_mask = keys_1_squeeze_mask_0, x = reshape_9_cast_fp16)[name = tensor("keys_1_cast_fp16")]; + tensor values_1_begin_0 = const()[name = tensor("values_1_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor values_1_end_0 = const()[name = tensor("values_1_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor values_1_end_mask_0 = const()[name = tensor("values_1_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor values_1_squeeze_mask_0 = const()[name = tensor("values_1_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor values_1_cast_fp16 = slice_by_index(begin = values_1_begin_0, end = values_1_end_0, end_mask = values_1_end_mask_0, squeeze_mask = values_1_squeeze_mask_0, x = reshape_9_cast_fp16)[name = tensor("values_1_cast_fp16")]; + tensor var_464_cast_fp16 = not_equal(x = keys_1_cast_fp16, y = keys_1_cast_fp16)[name = tensor("op_464_cast_fp16")]; + tensor var_470_to_fp16 = const()[name = tensor("op_470_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(6303296)))]; + tensor keys_3_cast_fp16 = select(a = var_470_to_fp16, b = keys_1_cast_fp16, cond = var_464_cast_fp16)[name = tensor("keys_3_cast_fp16")]; + tensor var_472_cast_fp16 = not_equal(x = values_1_cast_fp16, y = values_1_cast_fp16)[name = tensor("op_472_cast_fp16")]; + tensor values_3_cast_fp16 = select(a = var_470_to_fp16, b = values_1_cast_fp16, cond = var_472_cast_fp16)[name = tensor("values_3_cast_fp16")]; + tensor var_496 = const()[name = tensor("op_496"), val = tensor([0, 2, 1, 3])]; + tensor var_512 = const()[name = tensor("op_512"), val = tensor([1, 1, 1])]; + tensor var_513_cast_fp16 = reshape(shape = var_512, x = position0_to_fp16)[name = tensor("op_513_cast_fp16")]; + tensor q_offsets_1_promoted_to_fp16 = const()[name = tensor("q_offsets_1_promoted_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(7351936)))]; + tensor q_positions_1_cast_fp16 = add(x = var_513_cast_fp16, y = q_offsets_1_promoted_to_fp16)[name = tensor("q_positions_1_cast_fp16")]; + tensor var_535 = const()[name = tensor("op_535"), val = tensor([1, 1, 1])]; + tensor var_536_cast_fp16 = reshape(shape = var_535, x = valid_len_to_fp16)[name = tensor("op_536_cast_fp16")]; + tensor valid_end_1_cast_fp16 = add(x = var_513_cast_fp16, y = var_536_cast_fp16)[name = tensor("valid_end_1_cast_fp16")]; + tensor k_positions_1_promoted_to_fp16 = const()[name = tensor("k_positions_1_promoted_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(7352512)))]; + tensor valid_mask_1_cast_fp16 = less(x = k_positions_1_promoted_to_fp16, y = valid_end_1_cast_fp16)[name = tensor("valid_mask_1_cast_fp16")]; + tensor causal_mask_1_cast_fp16 = less_equal(x = k_positions_1_promoted_to_fp16, y = q_positions_1_cast_fp16)[name = tensor("causal_mask_1_cast_fp16")]; + tensor var_541 = logical_and(x = valid_mask_1_cast_fp16, y = causal_mask_1_cast_fp16)[name = tensor("op_541")]; + tensor attn_mask_1_axes_0 = const()[name = tensor("attn_mask_1_axes_0"), val = tensor([1])]; + tensor attn_mask_1 = expand_dims(axes = attn_mask_1_axes_0, x = var_541)[name = tensor("attn_mask_1")]; + tensor var_555_transpose_x_0 = const()[name = tensor("op_555_transpose_x_0"), val = tensor(false)]; + tensor var_555_transpose_y_0 = const()[name = tensor("op_555_transpose_y_0"), val = tensor(false)]; + tensor transpose_69_perm_0 = const()[name = tensor("transpose_69_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_70_perm_0 = const()[name = tensor("transpose_70_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_70 = transpose(perm = transpose_70_perm_0, x = keys_3_cast_fp16)[name = tensor("transpose_204")]; + tensor transpose_69 = transpose(perm = transpose_69_perm_0, x = q_3_cast_fp16)[name = tensor("transpose_205")]; + tensor var_555_cast_fp16 = matmul(transpose_x = var_555_transpose_x_0, transpose_y = var_555_transpose_y_0, x = transpose_69, y = transpose_70)[name = tensor("op_555_cast_fp16")]; + tensor var_549_to_fp16 = const()[name = tensor("op_549_to_fp16"), val = tensor([0x1p-3])]; + tensor attn_weights_1_cast_fp16 = mul(x = var_555_cast_fp16, y = var_549_to_fp16)[name = tensor("attn_weights_1_cast_fp16")]; + tensor var_557 = logical_not(x = attn_mask_1)[name = tensor("op_557")]; + tensor var_558_to_fp16 = const()[name = tensor("op_558_to_fp16"), val = tensor(-inf)]; + tensor attn_weights_3_cast_fp16 = select(a = var_558_to_fp16, b = attn_weights_1_cast_fp16, cond = var_557)[name = tensor("attn_weights_3_cast_fp16")]; + tensor var_560 = const()[name = tensor("op_560"), val = tensor(-1)]; + tensor attn_weights_5_cast_fp16 = softmax(axis = var_560, x = attn_weights_3_cast_fp16)[name = tensor("attn_weights_5_cast_fp16")]; + tensor attn_output_1_transpose_x_0 = const()[name = tensor("attn_output_1_transpose_x_0"), val = tensor(false)]; + tensor attn_output_1_transpose_y_0 = const()[name = tensor("attn_output_1_transpose_y_0"), val = tensor(false)]; + tensor values_5_cast_fp16 = transpose(perm = var_496, x = values_3_cast_fp16)[name = tensor("transpose_206")]; + tensor attn_output_1_cast_fp16 = matmul(transpose_x = attn_output_1_transpose_x_0, transpose_y = attn_output_1_transpose_y_0, x = attn_weights_5_cast_fp16, y = values_5_cast_fp16)[name = tensor("attn_output_1_cast_fp16")]; + tensor var_568 = const()[name = tensor("op_568"), val = tensor([0, 2, 1, 3])]; + tensor var_573 = const()[name = tensor("op_573"), val = tensor([1, 256, 1024])]; + tensor var_569_cast_fp16 = transpose(perm = var_568, x = attn_output_1_cast_fp16)[name = tensor("transpose_203")]; + tensor input_5_cast_fp16 = reshape(shape = var_573, x = var_569_cast_fp16)[name = tensor("input_5_cast_fp16")]; + tensor attn0_out_proj_weight_to_fp16 = const()[name = tensor("attn0_out_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(7353600)))]; + tensor linear_1_bias_0_to_fp16 = const()[name = tensor("linear_1_bias_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9450816)))]; + tensor linear_1_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = attn0_out_proj_weight_to_fp16, x = input_5_cast_fp16)[name = tensor("linear_1_cast_fp16")]; + tensor var_582_cast_fp16 = add(x = position0_to_fp16, y = valid_len_to_fp16)[name = tensor("op_582_cast_fp16")]; + tensor var_582_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_582_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor input_7_cast_fp16 = add(x = conditioning_to_fp16, y = linear_1_cast_fp16)[name = tensor("input_7_cast_fp16")]; + tensor input_9_axes_0 = const()[name = tensor("input_9_axes_0"), val = tensor([-1])]; + tensor norm0_2_weight_to_fp16 = const()[name = tensor("norm0_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9452928)))]; + tensor norm0_2_bias_to_fp16 = const()[name = tensor("norm0_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9455040)))]; + tensor var_586_to_fp16 = const()[name = tensor("op_586_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_9_cast_fp16 = layer_norm(axes = input_9_axes_0, beta = norm0_2_bias_to_fp16, epsilon = var_586_to_fp16, gamma = norm0_2_weight_to_fp16, x = input_7_cast_fp16)[name = tensor("input_9_cast_fp16")]; + tensor linear0_1_weight_to_fp16 = const()[name = tensor("linear0_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9457152)))]; + tensor linear_2_bias_0_to_fp16 = const()[name = tensor("linear_2_bias_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(17845824)))]; + tensor linear_2_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = linear0_1_weight_to_fp16, x = input_9_cast_fp16)[name = tensor("linear_2_cast_fp16")]; + tensor input_11_mode_0 = const()[name = tensor("input_11_mode_0"), val = tensor("EXACT")]; + tensor input_11_cast_fp16 = gelu(mode = input_11_mode_0, x = linear_2_cast_fp16)[name = tensor("input_11_cast_fp16")]; + tensor linear0_2_weight_to_fp16 = const()[name = tensor("linear0_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(17854080)))]; + tensor linear_3_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = linear0_2_weight_to_fp16, x = input_11_cast_fp16)[name = tensor("linear_3_cast_fp16")]; + tensor input_13_cast_fp16 = add(x = input_7_cast_fp16, y = linear_3_cast_fp16)[name = tensor("input_13_cast_fp16")]; + tensor input_15_axes_0 = const()[name = tensor("input_15_axes_0"), val = tensor([-1])]; + tensor norm1_1_weight_to_fp16 = const()[name = tensor("norm1_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(26242752)))]; + tensor norm1_1_bias_to_fp16 = const()[name = tensor("norm1_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(26244864)))]; + tensor var_603_to_fp16 = const()[name = tensor("op_603_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_15_cast_fp16 = layer_norm(axes = input_15_axes_0, beta = norm1_1_bias_to_fp16, epsilon = var_603_to_fp16, gamma = norm1_1_weight_to_fp16, x = input_13_cast_fp16)[name = tensor("input_15_cast_fp16")]; + tensor attn1_in_proj_weight_to_fp16 = const()[name = tensor("attn1_in_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(26246976)))]; + tensor linear_4_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = attn1_in_proj_weight_to_fp16, x = input_15_cast_fp16)[name = tensor("linear_4_cast_fp16")]; + tensor var_617 = const()[name = tensor("op_617"), val = tensor([1, 256, 3, 16, 64])]; + tensor qkv_3_cast_fp16 = reshape(shape = var_617, x = linear_4_cast_fp16)[name = tensor("qkv_3_cast_fp16")]; + tensor q_7_begin_0 = const()[name = tensor("q_7_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor q_7_end_0 = const()[name = tensor("q_7_end_0"), val = tensor([1, 256, 1, 16, 64])]; + tensor q_7_end_mask_0 = const()[name = tensor("q_7_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor q_7_squeeze_mask_0 = const()[name = tensor("q_7_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor q_7_cast_fp16 = slice_by_index(begin = q_7_begin_0, end = q_7_end_0, end_mask = q_7_end_mask_0, squeeze_mask = q_7_squeeze_mask_0, x = qkv_3_cast_fp16)[name = tensor("q_7_cast_fp16")]; + tensor k_5_begin_0 = const()[name = tensor("k_5_begin_0"), val = tensor([0, 0, 1, 0, 0])]; + tensor k_5_end_0 = const()[name = tensor("k_5_end_0"), val = tensor([1, 256, 2, 16, 64])]; + tensor k_5_end_mask_0 = const()[name = tensor("k_5_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor k_5_squeeze_mask_0 = const()[name = tensor("k_5_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor k_5_cast_fp16 = slice_by_index(begin = k_5_begin_0, end = k_5_end_0, end_mask = k_5_end_mask_0, squeeze_mask = k_5_squeeze_mask_0, x = qkv_3_cast_fp16)[name = tensor("k_5_cast_fp16")]; + tensor v_3_begin_0 = const()[name = tensor("v_3_begin_0"), val = tensor([0, 0, 2, 0, 0])]; + tensor v_3_end_0 = const()[name = tensor("v_3_end_0"), val = tensor([1, 256, 3, 16, 64])]; + tensor v_3_end_mask_0 = const()[name = tensor("v_3_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor v_3_squeeze_mask_0 = const()[name = tensor("v_3_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor v_3_cast_fp16 = slice_by_index(begin = v_3_begin_0, end = v_3_end_0, end_mask = v_3_end_mask_0, squeeze_mask = v_3_squeeze_mask_0, x = qkv_3_cast_fp16)[name = tensor("v_3_cast_fp16")]; + tensor var_675 = const()[name = tensor("op_675"), val = tensor([1, 1])]; + tensor position1_to_fp16_dtype_0 = const()[name = tensor("position1_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor position1_to_fp16 = cast(dtype = position1_to_fp16_dtype_0, x = position1)[name = tensor("cast_656")]; + tensor var_676_cast_fp16 = reshape(shape = var_675, x = position1_to_fp16)[name = tensor("op_676_cast_fp16")]; + tensor ts_9_cast_fp16 = add(x = ts_1_promoted_to_fp16, y = var_676_cast_fp16)[name = tensor("ts_9_cast_fp16")]; + tensor var_683 = const()[name = tensor("op_683"), val = tensor([1, 256, 1, 1])]; + tensor ts_11_cast_fp16 = reshape(shape = var_683, x = ts_9_cast_fp16)[name = tensor("ts_11_cast_fp16")]; + tensor var_690 = const()[name = tensor("op_690"), val = tensor([1, 256, 16, 32, 2])]; + tensor q_complex_3_cast_fp16 = reshape(shape = var_690, x = q_7_cast_fp16)[name = tensor("q_complex_3_cast_fp16")]; + tensor var_697 = const()[name = tensor("op_697"), val = tensor([1, 256, 16, 32, 2])]; + tensor k_complex_3_cast_fp16 = reshape(shape = var_697, x = k_5_cast_fp16)[name = tensor("k_complex_3_cast_fp16")]; + tensor var_701_begin_0 = const()[name = tensor("op_701_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_701_end_0 = const()[name = tensor("op_701_end_0"), val = tensor([1, 256, 16, 32, 1])]; + tensor var_701_end_mask_0 = const()[name = tensor("op_701_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_701_squeeze_mask_0 = const()[name = tensor("op_701_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_701_cast_fp16 = slice_by_index(begin = var_701_begin_0, end = var_701_end_0, end_mask = var_701_end_mask_0, squeeze_mask = var_701_squeeze_mask_0, x = q_complex_3_cast_fp16)[name = tensor("op_701_cast_fp16")]; + tensor var_709_begin_0 = const()[name = tensor("op_709_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_709_end_0 = const()[name = tensor("op_709_end_0"), val = tensor([1, 256, 16, 32, 2])]; + tensor var_709_end_mask_0 = const()[name = tensor("op_709_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_709_squeeze_mask_0 = const()[name = tensor("op_709_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_709_cast_fp16 = slice_by_index(begin = var_709_begin_0, end = var_709_end_0, end_mask = var_709_end_mask_0, squeeze_mask = var_709_squeeze_mask_0, x = q_complex_3_cast_fp16)[name = tensor("op_709_cast_fp16")]; + tensor var_717_begin_0 = const()[name = tensor("op_717_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_717_end_0 = const()[name = tensor("op_717_end_0"), val = tensor([1, 256, 16, 32, 1])]; + tensor var_717_end_mask_0 = const()[name = tensor("op_717_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_717_squeeze_mask_0 = const()[name = tensor("op_717_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_717_cast_fp16 = slice_by_index(begin = var_717_begin_0, end = var_717_end_0, end_mask = var_717_end_mask_0, squeeze_mask = var_717_squeeze_mask_0, x = k_complex_3_cast_fp16)[name = tensor("op_717_cast_fp16")]; + tensor var_725_begin_0 = const()[name = tensor("op_725_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_725_end_0 = const()[name = tensor("op_725_end_0"), val = tensor([1, 256, 16, 32, 2])]; + tensor var_725_end_mask_0 = const()[name = tensor("op_725_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_725_squeeze_mask_0 = const()[name = tensor("op_725_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_725_cast_fp16 = slice_by_index(begin = var_725_begin_0, end = var_725_end_0, end_mask = var_725_end_mask_0, squeeze_mask = var_725_squeeze_mask_0, x = k_complex_3_cast_fp16)[name = tensor("op_725_cast_fp16")]; + tensor freqs_3_to_fp16 = const()[name = tensor("freqs_3_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(32538496)))]; + tensor var_731_cast_fp16 = mul(x = freqs_3_to_fp16, y = ts_11_cast_fp16)[name = tensor("op_731_cast_fp16")]; + tensor rotr_3_cast_fp16 = cos(x = var_731_cast_fp16)[name = tensor("rotr_3_cast_fp16")]; + tensor roti_3_cast_fp16 = sin(x = var_731_cast_fp16)[name = tensor("roti_3_cast_fp16")]; + tensor var_735_cast_fp16 = mul(x = var_701_cast_fp16, y = rotr_3_cast_fp16)[name = tensor("op_735_cast_fp16")]; + tensor var_736_cast_fp16 = mul(x = var_709_cast_fp16, y = roti_3_cast_fp16)[name = tensor("op_736_cast_fp16")]; + tensor qor_5_cast_fp16 = sub(x = var_735_cast_fp16, y = var_736_cast_fp16)[name = tensor("qor_5_cast_fp16")]; + tensor var_739_cast_fp16 = mul(x = var_701_cast_fp16, y = roti_3_cast_fp16)[name = tensor("op_739_cast_fp16")]; + tensor var_740_cast_fp16 = mul(x = var_709_cast_fp16, y = rotr_3_cast_fp16)[name = tensor("op_740_cast_fp16")]; + tensor qoi_5_cast_fp16 = add(x = var_739_cast_fp16, y = var_740_cast_fp16)[name = tensor("qoi_5_cast_fp16")]; + tensor var_743_cast_fp16 = mul(x = var_717_cast_fp16, y = rotr_3_cast_fp16)[name = tensor("op_743_cast_fp16")]; + tensor var_744_cast_fp16 = mul(x = var_725_cast_fp16, y = roti_3_cast_fp16)[name = tensor("op_744_cast_fp16")]; + tensor kor_5_cast_fp16 = sub(x = var_743_cast_fp16, y = var_744_cast_fp16)[name = tensor("kor_5_cast_fp16")]; + tensor var_747_cast_fp16 = mul(x = var_717_cast_fp16, y = roti_3_cast_fp16)[name = tensor("op_747_cast_fp16")]; + tensor var_748_cast_fp16 = mul(x = var_725_cast_fp16, y = rotr_3_cast_fp16)[name = tensor("op_748_cast_fp16")]; + tensor koi_5_cast_fp16 = add(x = var_747_cast_fp16, y = var_748_cast_fp16)[name = tensor("koi_5_cast_fp16")]; + tensor qo_3_axis_0 = const()[name = tensor("qo_3_axis_0"), val = tensor(-1)]; + tensor qo_3_cast_fp16 = stack(axis = qo_3_axis_0, values = (qor_5_cast_fp16, qoi_5_cast_fp16))[name = tensor("qo_3_cast_fp16")]; + tensor ko_3_axis_0 = const()[name = tensor("ko_3_axis_0"), val = tensor(-1)]; + tensor ko_3_cast_fp16 = stack(axis = ko_3_axis_0, values = (kor_5_cast_fp16, koi_5_cast_fp16))[name = tensor("ko_3_cast_fp16")]; + tensor var_781 = const()[name = tensor("op_781"), val = tensor([1, 256, 16, 64])]; + tensor q_9_cast_fp16 = reshape(shape = var_781, x = qo_3_cast_fp16)[name = tensor("q_9_cast_fp16")]; + tensor var_787 = const()[name = tensor("op_787"), val = tensor([1, 256, 16, 64])]; + tensor k_7_cast_fp16 = reshape(shape = var_787, x = ko_3_cast_fp16)[name = tensor("k_7_cast_fp16")]; + tensor write_indices_float_7_cast_fp16 = add(x = var_676_cast_fp16, y = var_393_promoted_to_fp16)[name = tensor("write_indices_float_7_cast_fp16")]; + tensor _inversed_808_y_0_to_fp16 = const()[name = tensor("_inversed_808_y_0_to_fp16"), val = tensor(0x1p-9)]; + tensor _inversed_808_cast_fp16 = mul(x = write_indices_float_7_cast_fp16, y = _inversed_808_y_0_to_fp16)[name = tensor("_inversed_808_cast_fp16")]; + tensor var_809_cast_fp16 = floor(x = _inversed_808_cast_fp16)[name = tensor("op_809_cast_fp16")]; + tensor var_810_to_fp16 = const()[name = tensor("op_810_to_fp16"), val = tensor(0x1p+9)]; + tensor var_811_cast_fp16 = mul(x = var_809_cast_fp16, y = var_810_to_fp16)[name = tensor("op_811_cast_fp16")]; + tensor write_indices_float_9_cast_fp16 = sub(x = write_indices_float_7_cast_fp16, y = var_811_cast_fp16)[name = tensor("write_indices_float_9_cast_fp16")]; + tensor var_829_value_0_to_fp16 = const()[name = tensor("op_829_value_0_to_fp16"), val = tensor(0x1.ffp+8)]; + tensor var_829_cast_fp16 = fill_like(ref_tensor = write_indices_float_9_cast_fp16, value = var_829_value_0_to_fp16)[name = tensor("op_829_cast_fp16")]; + tensor write_indices_float_11_cast_fp16 = select(a = write_indices_float_9_cast_fp16, b = var_829_cast_fp16, cond = token_valid_1_cast_fp16)[name = tensor("write_indices_float_11_cast_fp16")]; + tensor var_835_dtype_0 = const()[name = tensor("op_835_dtype_0"), val = tensor("int32")]; + tensor var_840 = const()[name = tensor("op_840"), val = tensor([1, 256, 1, 1])]; + tensor write_indices_float_11_cast_fp16_to_int32 = cast(dtype = var_835_dtype_0, x = write_indices_float_11_cast_fp16)[name = tensor("cast_655")]; + tensor var_841 = reshape(shape = var_840, x = write_indices_float_11_cast_fp16_to_int32)[name = tensor("op_841")]; + tensor write_indices_3_reps_0 = const()[name = tensor("write_indices_3_reps_0"), val = tensor([1, 1, 16, 64])]; + tensor write_indices_3 = tile(reps = write_indices_3_reps_0, x = var_841)[name = tensor("write_indices_3")]; + tensor var_851_begin_0 = const()[name = tensor("op_851_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_851_end_0 = const()[name = tensor("op_851_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor var_851_end_mask_0 = const()[name = tensor("op_851_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_851_squeeze_mask_0 = const()[name = tensor("op_851_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor cache1_to_fp16_dtype_0 = const()[name = tensor("cache1_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor cache1_to_fp16 = cast(dtype = cache1_to_fp16_dtype_0, x = cache1)[name = tensor("cast_654")]; + tensor var_851_cast_fp16 = slice_by_index(begin = var_851_begin_0, end = var_851_end_0, end_mask = var_851_end_mask_0, squeeze_mask = var_851_squeeze_mask_0, x = cache1_to_fp16)[name = tensor("op_851_cast_fp16")]; + tensor var_853_axis_0 = const()[name = tensor("op_853_axis_0"), val = tensor(1)]; + tensor var_853_mode_0 = const()[name = tensor("op_853_mode_0"), val = tensor("update")]; + tensor var_853_validate_indices_0 = const()[name = tensor("op_853_validate_indices_0"), val = tensor(false)]; + tensor var_853_cast_fp16 = scatter_along_axis(axis = var_853_axis_0, data = var_851_cast_fp16, indices = write_indices_3, mode = var_853_mode_0, updates = k_7_cast_fp16, validate_indices = var_853_validate_indices_0)[name = tensor("op_853_cast_fp16")]; + tensor concat_10 = const()[name = tensor("concat_10"), val = tensor([0, 0, 0, 0, 0])]; + tensor concat_11 = const()[name = tensor("concat_11"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_3_internal_tensor_assign_1_stride_0 = const()[name = tensor("new_cache_3_internal_tensor_assign_1_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_3_internal_tensor_assign_1_begin_mask_0 = const()[name = tensor("new_cache_3_internal_tensor_assign_1_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_3_internal_tensor_assign_1_end_mask_0 = const()[name = tensor("new_cache_3_internal_tensor_assign_1_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_3_internal_tensor_assign_1_squeeze_mask_0 = const()[name = tensor("new_cache_3_internal_tensor_assign_1_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_48 = const()[name = tensor("shape_48"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_2 = const()[name = tensor("reduce_prod_2"), val = tensor(1048576)]; + tensor range_1d_2_start_0 = const()[name = tensor("range_1d_2_start_0"), val = tensor(0)]; + tensor range_1d_2_step_0 = const()[name = tensor("range_1d_2_step_0"), val = tensor(1)]; + tensor range_1d_2 = range_1d(end = reduce_prod_2, start = range_1d_2_start_0, step = range_1d_2_step_0)[name = tensor("range_1d_2")]; + tensor reshape_10 = reshape(shape = shape_48, x = range_1d_2)[name = tensor("reshape_10")]; + tensor slice_by_index_2 = slice_by_index(begin = concat_10, begin_mask = new_cache_3_internal_tensor_assign_1_begin_mask_0, end = concat_11, end_mask = new_cache_3_internal_tensor_assign_1_end_mask_0, squeeze_mask = new_cache_3_internal_tensor_assign_1_squeeze_mask_0, stride = new_cache_3_internal_tensor_assign_1_stride_0, x = reshape_10)[name = tensor("slice_by_index_2")]; + tensor reshape_11_shape_0 = const()[name = tensor("reshape_11_shape_0"), val = tensor([-1])]; + tensor reshape_11 = reshape(shape = reshape_11_shape_0, x = slice_by_index_2)[name = tensor("reshape_11")]; + tensor reshape_12_shape_0 = const()[name = tensor("reshape_12_shape_0"), val = tensor([-1])]; + tensor reshape_12_cast_fp16 = reshape(shape = reshape_12_shape_0, x = var_853_cast_fp16)[name = tensor("reshape_12_cast_fp16")]; + tensor reshape_13_shape_0 = const()[name = tensor("reshape_13_shape_0"), val = tensor([-1])]; + tensor reshape_13_cast_fp16 = reshape(shape = reshape_13_shape_0, x = cache1_to_fp16)[name = tensor("reshape_13_cast_fp16")]; + tensor scatter_2_mode_0 = const()[name = tensor("scatter_2_mode_0"), val = tensor("update")]; + tensor scatter_2_axis_0 = const()[name = tensor("scatter_2_axis_0"), val = tensor(0)]; + tensor scatter_2_validate_indices_0 = const()[name = tensor("scatter_2_validate_indices_0"), val = tensor(false)]; + tensor scatter_2_cast_fp16 = scatter(axis = scatter_2_axis_0, data = reshape_13_cast_fp16, indices = reshape_11, mode = scatter_2_mode_0, updates = reshape_12_cast_fp16, validate_indices = scatter_2_validate_indices_0)[name = tensor("scatter_2_cast_fp16")]; + tensor reshape_14_cast_fp16 = reshape(shape = shape_48, x = scatter_2_cast_fp16)[name = tensor("reshape_14_cast_fp16")]; + tensor var_861_begin_0 = const()[name = tensor("op_861_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor var_861_end_0 = const()[name = tensor("op_861_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor var_861_end_mask_0 = const()[name = tensor("op_861_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_861_squeeze_mask_0 = const()[name = tensor("op_861_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor var_861_cast_fp16 = slice_by_index(begin = var_861_begin_0, end = var_861_end_0, end_mask = var_861_end_mask_0, squeeze_mask = var_861_squeeze_mask_0, x = reshape_14_cast_fp16)[name = tensor("op_861_cast_fp16")]; + tensor var_863_axis_0 = const()[name = tensor("op_863_axis_0"), val = tensor(1)]; + tensor var_863_mode_0 = const()[name = tensor("op_863_mode_0"), val = tensor("update")]; + tensor var_863_validate_indices_0 = const()[name = tensor("op_863_validate_indices_0"), val = tensor(false)]; + tensor var_863_cast_fp16 = scatter_along_axis(axis = var_863_axis_0, data = var_861_cast_fp16, indices = write_indices_3, mode = var_863_mode_0, updates = v_3_cast_fp16, validate_indices = var_863_validate_indices_0)[name = tensor("op_863_cast_fp16")]; + tensor concat_12 = const()[name = tensor("concat_12"), val = tensor([1, 0, 0, 0, 0])]; + tensor concat_13 = const()[name = tensor("concat_13"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_3_internal_tensor_assign_2_stride_0 = const()[name = tensor("new_cache_3_internal_tensor_assign_2_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_3_internal_tensor_assign_2_begin_mask_0 = const()[name = tensor("new_cache_3_internal_tensor_assign_2_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_3_internal_tensor_assign_2_end_mask_0 = const()[name = tensor("new_cache_3_internal_tensor_assign_2_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_3_internal_tensor_assign_2_squeeze_mask_0 = const()[name = tensor("new_cache_3_internal_tensor_assign_2_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_49 = const()[name = tensor("shape_49"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_3 = const()[name = tensor("reduce_prod_3"), val = tensor(1048576)]; + tensor range_1d_3_start_0 = const()[name = tensor("range_1d_3_start_0"), val = tensor(0)]; + tensor range_1d_3_step_0 = const()[name = tensor("range_1d_3_step_0"), val = tensor(1)]; + tensor range_1d_3 = range_1d(end = reduce_prod_3, start = range_1d_3_start_0, step = range_1d_3_step_0)[name = tensor("range_1d_3")]; + tensor reshape_15 = reshape(shape = shape_49, x = range_1d_3)[name = tensor("reshape_15")]; + tensor slice_by_index_3 = slice_by_index(begin = concat_12, begin_mask = new_cache_3_internal_tensor_assign_2_begin_mask_0, end = concat_13, end_mask = new_cache_3_internal_tensor_assign_2_end_mask_0, squeeze_mask = new_cache_3_internal_tensor_assign_2_squeeze_mask_0, stride = new_cache_3_internal_tensor_assign_2_stride_0, x = reshape_15)[name = tensor("slice_by_index_3")]; + tensor reshape_16_shape_0 = const()[name = tensor("reshape_16_shape_0"), val = tensor([-1])]; + tensor reshape_16 = reshape(shape = reshape_16_shape_0, x = slice_by_index_3)[name = tensor("reshape_16")]; + tensor reshape_17_shape_0 = const()[name = tensor("reshape_17_shape_0"), val = tensor([-1])]; + tensor reshape_17_cast_fp16 = reshape(shape = reshape_17_shape_0, x = var_863_cast_fp16)[name = tensor("reshape_17_cast_fp16")]; + tensor reshape_18_shape_0 = const()[name = tensor("reshape_18_shape_0"), val = tensor([-1])]; + tensor reshape_18_cast_fp16 = reshape(shape = reshape_18_shape_0, x = reshape_14_cast_fp16)[name = tensor("reshape_18_cast_fp16")]; + tensor scatter_3_mode_0 = const()[name = tensor("scatter_3_mode_0"), val = tensor("update")]; + tensor scatter_3_axis_0 = const()[name = tensor("scatter_3_axis_0"), val = tensor(0)]; + tensor scatter_3_validate_indices_0 = const()[name = tensor("scatter_3_validate_indices_0"), val = tensor(false)]; + tensor scatter_3_cast_fp16 = scatter(axis = scatter_3_axis_0, data = reshape_18_cast_fp16, indices = reshape_16, mode = scatter_3_mode_0, updates = reshape_17_cast_fp16, validate_indices = scatter_3_validate_indices_0)[name = tensor("scatter_3_cast_fp16")]; + tensor reshape_19_cast_fp16 = reshape(shape = shape_49, x = scatter_3_cast_fp16)[name = tensor("reshape_19_cast_fp16")]; + tensor reshape_19_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("reshape_19_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor keys_7_begin_0 = const()[name = tensor("keys_7_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor keys_7_end_0 = const()[name = tensor("keys_7_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor keys_7_end_mask_0 = const()[name = tensor("keys_7_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor keys_7_squeeze_mask_0 = const()[name = tensor("keys_7_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor keys_7_cast_fp16 = slice_by_index(begin = keys_7_begin_0, end = keys_7_end_0, end_mask = keys_7_end_mask_0, squeeze_mask = keys_7_squeeze_mask_0, x = reshape_19_cast_fp16)[name = tensor("keys_7_cast_fp16")]; + tensor values_7_begin_0 = const()[name = tensor("values_7_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor values_7_end_0 = const()[name = tensor("values_7_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor values_7_end_mask_0 = const()[name = tensor("values_7_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor values_7_squeeze_mask_0 = const()[name = tensor("values_7_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor values_7_cast_fp16 = slice_by_index(begin = values_7_begin_0, end = values_7_end_0, end_mask = values_7_end_mask_0, squeeze_mask = values_7_squeeze_mask_0, x = reshape_19_cast_fp16)[name = tensor("values_7_cast_fp16")]; + tensor var_875_cast_fp16 = not_equal(x = keys_7_cast_fp16, y = keys_7_cast_fp16)[name = tensor("op_875_cast_fp16")]; + tensor keys_9_cast_fp16 = select(a = var_470_to_fp16, b = keys_7_cast_fp16, cond = var_875_cast_fp16)[name = tensor("keys_9_cast_fp16")]; + tensor var_883_cast_fp16 = not_equal(x = values_7_cast_fp16, y = values_7_cast_fp16)[name = tensor("op_883_cast_fp16")]; + tensor values_9_cast_fp16 = select(a = var_470_to_fp16, b = values_7_cast_fp16, cond = var_883_cast_fp16)[name = tensor("values_9_cast_fp16")]; + tensor var_907 = const()[name = tensor("op_907"), val = tensor([0, 2, 1, 3])]; + tensor var_923 = const()[name = tensor("op_923"), val = tensor([1, 1, 1])]; + tensor var_924_cast_fp16 = reshape(shape = var_923, x = position1_to_fp16)[name = tensor("op_924_cast_fp16")]; + tensor q_positions_3_cast_fp16 = add(x = var_924_cast_fp16, y = q_offsets_1_promoted_to_fp16)[name = tensor("q_positions_3_cast_fp16")]; + tensor valid_end_3_cast_fp16 = add(x = var_924_cast_fp16, y = var_536_cast_fp16)[name = tensor("valid_end_3_cast_fp16")]; + tensor valid_mask_3_cast_fp16 = less(x = k_positions_1_promoted_to_fp16, y = valid_end_3_cast_fp16)[name = tensor("valid_mask_3_cast_fp16")]; + tensor causal_mask_3_cast_fp16 = less_equal(x = k_positions_1_promoted_to_fp16, y = q_positions_3_cast_fp16)[name = tensor("causal_mask_3_cast_fp16")]; + tensor var_952 = logical_and(x = valid_mask_3_cast_fp16, y = causal_mask_3_cast_fp16)[name = tensor("op_952")]; + tensor attn_mask_3_axes_0 = const()[name = tensor("attn_mask_3_axes_0"), val = tensor([1])]; + tensor attn_mask_3 = expand_dims(axes = attn_mask_3_axes_0, x = var_952)[name = tensor("attn_mask_3")]; + tensor var_966_transpose_x_0 = const()[name = tensor("op_966_transpose_x_0"), val = tensor(false)]; + tensor var_966_transpose_y_0 = const()[name = tensor("op_966_transpose_y_0"), val = tensor(false)]; + tensor transpose_71_perm_0 = const()[name = tensor("transpose_71_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_72_perm_0 = const()[name = tensor("transpose_72_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_72 = transpose(perm = transpose_72_perm_0, x = keys_9_cast_fp16)[name = tensor("transpose_200")]; + tensor transpose_71 = transpose(perm = transpose_71_perm_0, x = q_9_cast_fp16)[name = tensor("transpose_201")]; + tensor var_966_cast_fp16 = matmul(transpose_x = var_966_transpose_x_0, transpose_y = var_966_transpose_y_0, x = transpose_71, y = transpose_72)[name = tensor("op_966_cast_fp16")]; + tensor var_960_to_fp16 = const()[name = tensor("op_960_to_fp16"), val = tensor([0x1p-3])]; + tensor attn_weights_7_cast_fp16 = mul(x = var_966_cast_fp16, y = var_960_to_fp16)[name = tensor("attn_weights_7_cast_fp16")]; + tensor var_968 = logical_not(x = attn_mask_3)[name = tensor("op_968")]; + tensor var_969_to_fp16 = const()[name = tensor("op_969_to_fp16"), val = tensor(-inf)]; + tensor attn_weights_9_cast_fp16 = select(a = var_969_to_fp16, b = attn_weights_7_cast_fp16, cond = var_968)[name = tensor("attn_weights_9_cast_fp16")]; + tensor var_971 = const()[name = tensor("op_971"), val = tensor(-1)]; + tensor attn_weights_11_cast_fp16 = softmax(axis = var_971, x = attn_weights_9_cast_fp16)[name = tensor("attn_weights_11_cast_fp16")]; + tensor attn_output_3_transpose_x_0 = const()[name = tensor("attn_output_3_transpose_x_0"), val = tensor(false)]; + tensor attn_output_3_transpose_y_0 = const()[name = tensor("attn_output_3_transpose_y_0"), val = tensor(false)]; + tensor values_11_cast_fp16 = transpose(perm = var_907, x = values_9_cast_fp16)[name = tensor("transpose_202")]; + tensor attn_output_3_cast_fp16 = matmul(transpose_x = attn_output_3_transpose_x_0, transpose_y = attn_output_3_transpose_y_0, x = attn_weights_11_cast_fp16, y = values_11_cast_fp16)[name = tensor("attn_output_3_cast_fp16")]; + tensor var_979 = const()[name = tensor("op_979"), val = tensor([0, 2, 1, 3])]; + tensor var_984 = const()[name = tensor("op_984"), val = tensor([1, 256, 1024])]; + tensor var_980_cast_fp16 = transpose(perm = var_979, x = attn_output_3_cast_fp16)[name = tensor("transpose_199")]; + tensor input_17_cast_fp16 = reshape(shape = var_984, x = var_980_cast_fp16)[name = tensor("input_17_cast_fp16")]; + tensor attn1_out_proj_weight_to_fp16 = const()[name = tensor("attn1_out_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(32538624)))]; + tensor linear_5_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = attn1_out_proj_weight_to_fp16, x = input_17_cast_fp16)[name = tensor("linear_5_cast_fp16")]; + tensor var_993_cast_fp16 = add(x = position1_to_fp16, y = valid_len_to_fp16)[name = tensor("op_993_cast_fp16")]; + tensor var_993_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_993_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor input_19_cast_fp16 = add(x = input_13_cast_fp16, y = linear_5_cast_fp16)[name = tensor("input_19_cast_fp16")]; + tensor input_21_axes_0 = const()[name = tensor("input_21_axes_0"), val = tensor([-1])]; + tensor norm1_2_weight_to_fp16 = const()[name = tensor("norm1_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(34635840)))]; + tensor norm1_2_bias_to_fp16 = const()[name = tensor("norm1_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(34637952)))]; + tensor var_997_to_fp16 = const()[name = tensor("op_997_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_21_cast_fp16 = layer_norm(axes = input_21_axes_0, beta = norm1_2_bias_to_fp16, epsilon = var_997_to_fp16, gamma = norm1_2_weight_to_fp16, x = input_19_cast_fp16)[name = tensor("input_21_cast_fp16")]; + tensor linear1_1_weight_to_fp16 = const()[name = tensor("linear1_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(34640064)))]; + tensor linear_6_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = linear1_1_weight_to_fp16, x = input_21_cast_fp16)[name = tensor("linear_6_cast_fp16")]; + tensor input_23_mode_0 = const()[name = tensor("input_23_mode_0"), val = tensor("EXACT")]; + tensor input_23_cast_fp16 = gelu(mode = input_23_mode_0, x = linear_6_cast_fp16)[name = tensor("input_23_cast_fp16")]; + tensor linear1_2_weight_to_fp16 = const()[name = tensor("linear1_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(43028736)))]; + tensor linear_7_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = linear1_2_weight_to_fp16, x = input_23_cast_fp16)[name = tensor("linear_7_cast_fp16")]; + tensor input_25_cast_fp16 = add(x = input_19_cast_fp16, y = linear_7_cast_fp16)[name = tensor("input_25_cast_fp16")]; + tensor input_27_axes_0 = const()[name = tensor("input_27_axes_0"), val = tensor([-1])]; + tensor norm2_1_weight_to_fp16 = const()[name = tensor("norm2_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(51417408)))]; + tensor norm2_1_bias_to_fp16 = const()[name = tensor("norm2_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(51419520)))]; + tensor var_1014_to_fp16 = const()[name = tensor("op_1014_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_27_cast_fp16 = layer_norm(axes = input_27_axes_0, beta = norm2_1_bias_to_fp16, epsilon = var_1014_to_fp16, gamma = norm2_1_weight_to_fp16, x = input_25_cast_fp16)[name = tensor("input_27_cast_fp16")]; + tensor attn2_in_proj_weight_to_fp16 = const()[name = tensor("attn2_in_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(51421632)))]; + tensor linear_8_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = attn2_in_proj_weight_to_fp16, x = input_27_cast_fp16)[name = tensor("linear_8_cast_fp16")]; + tensor var_1028 = const()[name = tensor("op_1028"), val = tensor([1, 256, 3, 16, 64])]; + tensor qkv_5_cast_fp16 = reshape(shape = var_1028, x = linear_8_cast_fp16)[name = tensor("qkv_5_cast_fp16")]; + tensor q_13_begin_0 = const()[name = tensor("q_13_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor q_13_end_0 = const()[name = tensor("q_13_end_0"), val = tensor([1, 256, 1, 16, 64])]; + tensor q_13_end_mask_0 = const()[name = tensor("q_13_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor q_13_squeeze_mask_0 = const()[name = tensor("q_13_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor q_13_cast_fp16 = slice_by_index(begin = q_13_begin_0, end = q_13_end_0, end_mask = q_13_end_mask_0, squeeze_mask = q_13_squeeze_mask_0, x = qkv_5_cast_fp16)[name = tensor("q_13_cast_fp16")]; + tensor k_9_begin_0 = const()[name = tensor("k_9_begin_0"), val = tensor([0, 0, 1, 0, 0])]; + tensor k_9_end_0 = const()[name = tensor("k_9_end_0"), val = tensor([1, 256, 2, 16, 64])]; + tensor k_9_end_mask_0 = const()[name = tensor("k_9_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor k_9_squeeze_mask_0 = const()[name = tensor("k_9_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor k_9_cast_fp16 = slice_by_index(begin = k_9_begin_0, end = k_9_end_0, end_mask = k_9_end_mask_0, squeeze_mask = k_9_squeeze_mask_0, x = qkv_5_cast_fp16)[name = tensor("k_9_cast_fp16")]; + tensor v_5_begin_0 = const()[name = tensor("v_5_begin_0"), val = tensor([0, 0, 2, 0, 0])]; + tensor v_5_end_0 = const()[name = tensor("v_5_end_0"), val = tensor([1, 256, 3, 16, 64])]; + tensor v_5_end_mask_0 = const()[name = tensor("v_5_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor v_5_squeeze_mask_0 = const()[name = tensor("v_5_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor v_5_cast_fp16 = slice_by_index(begin = v_5_begin_0, end = v_5_end_0, end_mask = v_5_end_mask_0, squeeze_mask = v_5_squeeze_mask_0, x = qkv_5_cast_fp16)[name = tensor("v_5_cast_fp16")]; + tensor var_1086 = const()[name = tensor("op_1086"), val = tensor([1, 1])]; + tensor position2_to_fp16_dtype_0 = const()[name = tensor("position2_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor position2_to_fp16 = cast(dtype = position2_to_fp16_dtype_0, x = position2)[name = tensor("cast_651")]; + tensor var_1087_cast_fp16 = reshape(shape = var_1086, x = position2_to_fp16)[name = tensor("op_1087_cast_fp16")]; + tensor ts_15_cast_fp16 = add(x = ts_1_promoted_to_fp16, y = var_1087_cast_fp16)[name = tensor("ts_15_cast_fp16")]; + tensor var_1094 = const()[name = tensor("op_1094"), val = tensor([1, 256, 1, 1])]; + tensor ts_17_cast_fp16 = reshape(shape = var_1094, x = ts_15_cast_fp16)[name = tensor("ts_17_cast_fp16")]; + tensor var_1101 = const()[name = tensor("op_1101"), val = tensor([1, 256, 16, 32, 2])]; + tensor q_complex_5_cast_fp16 = reshape(shape = var_1101, x = q_13_cast_fp16)[name = tensor("q_complex_5_cast_fp16")]; + tensor var_1108 = const()[name = tensor("op_1108"), val = tensor([1, 256, 16, 32, 2])]; + tensor k_complex_5_cast_fp16 = reshape(shape = var_1108, x = k_9_cast_fp16)[name = tensor("k_complex_5_cast_fp16")]; + tensor var_1112_begin_0 = const()[name = tensor("op_1112_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_1112_end_0 = const()[name = tensor("op_1112_end_0"), val = tensor([1, 256, 16, 32, 1])]; + tensor var_1112_end_mask_0 = const()[name = tensor("op_1112_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_1112_squeeze_mask_0 = const()[name = tensor("op_1112_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_1112_cast_fp16 = slice_by_index(begin = var_1112_begin_0, end = var_1112_end_0, end_mask = var_1112_end_mask_0, squeeze_mask = var_1112_squeeze_mask_0, x = q_complex_5_cast_fp16)[name = tensor("op_1112_cast_fp16")]; + tensor var_1120_begin_0 = const()[name = tensor("op_1120_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_1120_end_0 = const()[name = tensor("op_1120_end_0"), val = tensor([1, 256, 16, 32, 2])]; + tensor var_1120_end_mask_0 = const()[name = tensor("op_1120_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_1120_squeeze_mask_0 = const()[name = tensor("op_1120_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_1120_cast_fp16 = slice_by_index(begin = var_1120_begin_0, end = var_1120_end_0, end_mask = var_1120_end_mask_0, squeeze_mask = var_1120_squeeze_mask_0, x = q_complex_5_cast_fp16)[name = tensor("op_1120_cast_fp16")]; + tensor var_1128_begin_0 = const()[name = tensor("op_1128_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_1128_end_0 = const()[name = tensor("op_1128_end_0"), val = tensor([1, 256, 16, 32, 1])]; + tensor var_1128_end_mask_0 = const()[name = tensor("op_1128_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_1128_squeeze_mask_0 = const()[name = tensor("op_1128_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_1128_cast_fp16 = slice_by_index(begin = var_1128_begin_0, end = var_1128_end_0, end_mask = var_1128_end_mask_0, squeeze_mask = var_1128_squeeze_mask_0, x = k_complex_5_cast_fp16)[name = tensor("op_1128_cast_fp16")]; + tensor var_1136_begin_0 = const()[name = tensor("op_1136_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_1136_end_0 = const()[name = tensor("op_1136_end_0"), val = tensor([1, 256, 16, 32, 2])]; + tensor var_1136_end_mask_0 = const()[name = tensor("op_1136_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_1136_squeeze_mask_0 = const()[name = tensor("op_1136_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_1136_cast_fp16 = slice_by_index(begin = var_1136_begin_0, end = var_1136_end_0, end_mask = var_1136_end_mask_0, squeeze_mask = var_1136_squeeze_mask_0, x = k_complex_5_cast_fp16)[name = tensor("op_1136_cast_fp16")]; + tensor freqs_5_to_fp16 = const()[name = tensor("freqs_5_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(57713152)))]; + tensor var_1142_cast_fp16 = mul(x = freqs_5_to_fp16, y = ts_17_cast_fp16)[name = tensor("op_1142_cast_fp16")]; + tensor rotr_5_cast_fp16 = cos(x = var_1142_cast_fp16)[name = tensor("rotr_5_cast_fp16")]; + tensor roti_5_cast_fp16 = sin(x = var_1142_cast_fp16)[name = tensor("roti_5_cast_fp16")]; + tensor var_1146_cast_fp16 = mul(x = var_1112_cast_fp16, y = rotr_5_cast_fp16)[name = tensor("op_1146_cast_fp16")]; + tensor var_1147_cast_fp16 = mul(x = var_1120_cast_fp16, y = roti_5_cast_fp16)[name = tensor("op_1147_cast_fp16")]; + tensor qor_9_cast_fp16 = sub(x = var_1146_cast_fp16, y = var_1147_cast_fp16)[name = tensor("qor_9_cast_fp16")]; + tensor var_1150_cast_fp16 = mul(x = var_1112_cast_fp16, y = roti_5_cast_fp16)[name = tensor("op_1150_cast_fp16")]; + tensor var_1151_cast_fp16 = mul(x = var_1120_cast_fp16, y = rotr_5_cast_fp16)[name = tensor("op_1151_cast_fp16")]; + tensor qoi_9_cast_fp16 = add(x = var_1150_cast_fp16, y = var_1151_cast_fp16)[name = tensor("qoi_9_cast_fp16")]; + tensor var_1154_cast_fp16 = mul(x = var_1128_cast_fp16, y = rotr_5_cast_fp16)[name = tensor("op_1154_cast_fp16")]; + tensor var_1155_cast_fp16 = mul(x = var_1136_cast_fp16, y = roti_5_cast_fp16)[name = tensor("op_1155_cast_fp16")]; + tensor kor_9_cast_fp16 = sub(x = var_1154_cast_fp16, y = var_1155_cast_fp16)[name = tensor("kor_9_cast_fp16")]; + tensor var_1158_cast_fp16 = mul(x = var_1128_cast_fp16, y = roti_5_cast_fp16)[name = tensor("op_1158_cast_fp16")]; + tensor var_1159_cast_fp16 = mul(x = var_1136_cast_fp16, y = rotr_5_cast_fp16)[name = tensor("op_1159_cast_fp16")]; + tensor koi_9_cast_fp16 = add(x = var_1158_cast_fp16, y = var_1159_cast_fp16)[name = tensor("koi_9_cast_fp16")]; + tensor qo_5_axis_0 = const()[name = tensor("qo_5_axis_0"), val = tensor(-1)]; + tensor qo_5_cast_fp16 = stack(axis = qo_5_axis_0, values = (qor_9_cast_fp16, qoi_9_cast_fp16))[name = tensor("qo_5_cast_fp16")]; + tensor ko_5_axis_0 = const()[name = tensor("ko_5_axis_0"), val = tensor(-1)]; + tensor ko_5_cast_fp16 = stack(axis = ko_5_axis_0, values = (kor_9_cast_fp16, koi_9_cast_fp16))[name = tensor("ko_5_cast_fp16")]; + tensor var_1192 = const()[name = tensor("op_1192"), val = tensor([1, 256, 16, 64])]; + tensor q_15_cast_fp16 = reshape(shape = var_1192, x = qo_5_cast_fp16)[name = tensor("q_15_cast_fp16")]; + tensor var_1198 = const()[name = tensor("op_1198"), val = tensor([1, 256, 16, 64])]; + tensor k_11_cast_fp16 = reshape(shape = var_1198, x = ko_5_cast_fp16)[name = tensor("k_11_cast_fp16")]; + tensor write_indices_float_13_cast_fp16 = add(x = var_1087_cast_fp16, y = var_393_promoted_to_fp16)[name = tensor("write_indices_float_13_cast_fp16")]; + tensor _inversed_1219_y_0_to_fp16 = const()[name = tensor("_inversed_1219_y_0_to_fp16"), val = tensor(0x1p-9)]; + tensor _inversed_1219_cast_fp16 = mul(x = write_indices_float_13_cast_fp16, y = _inversed_1219_y_0_to_fp16)[name = tensor("_inversed_1219_cast_fp16")]; + tensor var_1220_cast_fp16 = floor(x = _inversed_1219_cast_fp16)[name = tensor("op_1220_cast_fp16")]; + tensor var_1221_to_fp16 = const()[name = tensor("op_1221_to_fp16"), val = tensor(0x1p+9)]; + tensor var_1222_cast_fp16 = mul(x = var_1220_cast_fp16, y = var_1221_to_fp16)[name = tensor("op_1222_cast_fp16")]; + tensor write_indices_float_15_cast_fp16 = sub(x = write_indices_float_13_cast_fp16, y = var_1222_cast_fp16)[name = tensor("write_indices_float_15_cast_fp16")]; + tensor var_1240_value_0_to_fp16 = const()[name = tensor("op_1240_value_0_to_fp16"), val = tensor(0x1.ffp+8)]; + tensor var_1240_cast_fp16 = fill_like(ref_tensor = write_indices_float_15_cast_fp16, value = var_1240_value_0_to_fp16)[name = tensor("op_1240_cast_fp16")]; + tensor write_indices_float_17_cast_fp16 = select(a = write_indices_float_15_cast_fp16, b = var_1240_cast_fp16, cond = token_valid_1_cast_fp16)[name = tensor("write_indices_float_17_cast_fp16")]; + tensor var_1246_dtype_0 = const()[name = tensor("op_1246_dtype_0"), val = tensor("int32")]; + tensor var_1251 = const()[name = tensor("op_1251"), val = tensor([1, 256, 1, 1])]; + tensor write_indices_float_17_cast_fp16_to_int32 = cast(dtype = var_1246_dtype_0, x = write_indices_float_17_cast_fp16)[name = tensor("cast_650")]; + tensor var_1252 = reshape(shape = var_1251, x = write_indices_float_17_cast_fp16_to_int32)[name = tensor("op_1252")]; + tensor write_indices_5_reps_0 = const()[name = tensor("write_indices_5_reps_0"), val = tensor([1, 1, 16, 64])]; + tensor write_indices_5 = tile(reps = write_indices_5_reps_0, x = var_1252)[name = tensor("write_indices_5")]; + tensor var_1262_begin_0 = const()[name = tensor("op_1262_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_1262_end_0 = const()[name = tensor("op_1262_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor var_1262_end_mask_0 = const()[name = tensor("op_1262_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_1262_squeeze_mask_0 = const()[name = tensor("op_1262_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor cache2_to_fp16_dtype_0 = const()[name = tensor("cache2_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor cache2_to_fp16 = cast(dtype = cache2_to_fp16_dtype_0, x = cache2)[name = tensor("cast_649")]; + tensor var_1262_cast_fp16 = slice_by_index(begin = var_1262_begin_0, end = var_1262_end_0, end_mask = var_1262_end_mask_0, squeeze_mask = var_1262_squeeze_mask_0, x = cache2_to_fp16)[name = tensor("op_1262_cast_fp16")]; + tensor var_1264_axis_0 = const()[name = tensor("op_1264_axis_0"), val = tensor(1)]; + tensor var_1264_mode_0 = const()[name = tensor("op_1264_mode_0"), val = tensor("update")]; + tensor var_1264_validate_indices_0 = const()[name = tensor("op_1264_validate_indices_0"), val = tensor(false)]; + tensor var_1264_cast_fp16 = scatter_along_axis(axis = var_1264_axis_0, data = var_1262_cast_fp16, indices = write_indices_5, mode = var_1264_mode_0, updates = k_11_cast_fp16, validate_indices = var_1264_validate_indices_0)[name = tensor("op_1264_cast_fp16")]; + tensor concat_18 = const()[name = tensor("concat_18"), val = tensor([0, 0, 0, 0, 0])]; + tensor concat_19 = const()[name = tensor("concat_19"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_5_internal_tensor_assign_1_stride_0 = const()[name = tensor("new_cache_5_internal_tensor_assign_1_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_5_internal_tensor_assign_1_begin_mask_0 = const()[name = tensor("new_cache_5_internal_tensor_assign_1_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_5_internal_tensor_assign_1_end_mask_0 = const()[name = tensor("new_cache_5_internal_tensor_assign_1_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_5_internal_tensor_assign_1_squeeze_mask_0 = const()[name = tensor("new_cache_5_internal_tensor_assign_1_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_50 = const()[name = tensor("shape_50"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_4 = const()[name = tensor("reduce_prod_4"), val = tensor(1048576)]; + tensor range_1d_4_start_0 = const()[name = tensor("range_1d_4_start_0"), val = tensor(0)]; + tensor range_1d_4_step_0 = const()[name = tensor("range_1d_4_step_0"), val = tensor(1)]; + tensor range_1d_4 = range_1d(end = reduce_prod_4, start = range_1d_4_start_0, step = range_1d_4_step_0)[name = tensor("range_1d_4")]; + tensor reshape_20 = reshape(shape = shape_50, x = range_1d_4)[name = tensor("reshape_20")]; + tensor slice_by_index_4 = slice_by_index(begin = concat_18, begin_mask = new_cache_5_internal_tensor_assign_1_begin_mask_0, end = concat_19, end_mask = new_cache_5_internal_tensor_assign_1_end_mask_0, squeeze_mask = new_cache_5_internal_tensor_assign_1_squeeze_mask_0, stride = new_cache_5_internal_tensor_assign_1_stride_0, x = reshape_20)[name = tensor("slice_by_index_4")]; + tensor reshape_21_shape_0 = const()[name = tensor("reshape_21_shape_0"), val = tensor([-1])]; + tensor reshape_21 = reshape(shape = reshape_21_shape_0, x = slice_by_index_4)[name = tensor("reshape_21")]; + tensor reshape_22_shape_0 = const()[name = tensor("reshape_22_shape_0"), val = tensor([-1])]; + tensor reshape_22_cast_fp16 = reshape(shape = reshape_22_shape_0, x = var_1264_cast_fp16)[name = tensor("reshape_22_cast_fp16")]; + tensor reshape_23_shape_0 = const()[name = tensor("reshape_23_shape_0"), val = tensor([-1])]; + tensor reshape_23_cast_fp16 = reshape(shape = reshape_23_shape_0, x = cache2_to_fp16)[name = tensor("reshape_23_cast_fp16")]; + tensor scatter_4_mode_0 = const()[name = tensor("scatter_4_mode_0"), val = tensor("update")]; + tensor scatter_4_axis_0 = const()[name = tensor("scatter_4_axis_0"), val = tensor(0)]; + tensor scatter_4_validate_indices_0 = const()[name = tensor("scatter_4_validate_indices_0"), val = tensor(false)]; + tensor scatter_4_cast_fp16 = scatter(axis = scatter_4_axis_0, data = reshape_23_cast_fp16, indices = reshape_21, mode = scatter_4_mode_0, updates = reshape_22_cast_fp16, validate_indices = scatter_4_validate_indices_0)[name = tensor("scatter_4_cast_fp16")]; + tensor reshape_24_cast_fp16 = reshape(shape = shape_50, x = scatter_4_cast_fp16)[name = tensor("reshape_24_cast_fp16")]; + tensor var_1272_begin_0 = const()[name = tensor("op_1272_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor var_1272_end_0 = const()[name = tensor("op_1272_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor var_1272_end_mask_0 = const()[name = tensor("op_1272_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_1272_squeeze_mask_0 = const()[name = tensor("op_1272_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor var_1272_cast_fp16 = slice_by_index(begin = var_1272_begin_0, end = var_1272_end_0, end_mask = var_1272_end_mask_0, squeeze_mask = var_1272_squeeze_mask_0, x = reshape_24_cast_fp16)[name = tensor("op_1272_cast_fp16")]; + tensor var_1274_axis_0 = const()[name = tensor("op_1274_axis_0"), val = tensor(1)]; + tensor var_1274_mode_0 = const()[name = tensor("op_1274_mode_0"), val = tensor("update")]; + tensor var_1274_validate_indices_0 = const()[name = tensor("op_1274_validate_indices_0"), val = tensor(false)]; + tensor var_1274_cast_fp16 = scatter_along_axis(axis = var_1274_axis_0, data = var_1272_cast_fp16, indices = write_indices_5, mode = var_1274_mode_0, updates = v_5_cast_fp16, validate_indices = var_1274_validate_indices_0)[name = tensor("op_1274_cast_fp16")]; + tensor concat_20 = const()[name = tensor("concat_20"), val = tensor([1, 0, 0, 0, 0])]; + tensor concat_21 = const()[name = tensor("concat_21"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_5_internal_tensor_assign_2_stride_0 = const()[name = tensor("new_cache_5_internal_tensor_assign_2_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_5_internal_tensor_assign_2_begin_mask_0 = const()[name = tensor("new_cache_5_internal_tensor_assign_2_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_5_internal_tensor_assign_2_end_mask_0 = const()[name = tensor("new_cache_5_internal_tensor_assign_2_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_5_internal_tensor_assign_2_squeeze_mask_0 = const()[name = tensor("new_cache_5_internal_tensor_assign_2_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_51 = const()[name = tensor("shape_51"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_5 = const()[name = tensor("reduce_prod_5"), val = tensor(1048576)]; + tensor range_1d_5_start_0 = const()[name = tensor("range_1d_5_start_0"), val = tensor(0)]; + tensor range_1d_5_step_0 = const()[name = tensor("range_1d_5_step_0"), val = tensor(1)]; + tensor range_1d_5 = range_1d(end = reduce_prod_5, start = range_1d_5_start_0, step = range_1d_5_step_0)[name = tensor("range_1d_5")]; + tensor reshape_25 = reshape(shape = shape_51, x = range_1d_5)[name = tensor("reshape_25")]; + tensor slice_by_index_5 = slice_by_index(begin = concat_20, begin_mask = new_cache_5_internal_tensor_assign_2_begin_mask_0, end = concat_21, end_mask = new_cache_5_internal_tensor_assign_2_end_mask_0, squeeze_mask = new_cache_5_internal_tensor_assign_2_squeeze_mask_0, stride = new_cache_5_internal_tensor_assign_2_stride_0, x = reshape_25)[name = tensor("slice_by_index_5")]; + tensor reshape_26_shape_0 = const()[name = tensor("reshape_26_shape_0"), val = tensor([-1])]; + tensor reshape_26 = reshape(shape = reshape_26_shape_0, x = slice_by_index_5)[name = tensor("reshape_26")]; + tensor reshape_27_shape_0 = const()[name = tensor("reshape_27_shape_0"), val = tensor([-1])]; + tensor reshape_27_cast_fp16 = reshape(shape = reshape_27_shape_0, x = var_1274_cast_fp16)[name = tensor("reshape_27_cast_fp16")]; + tensor reshape_28_shape_0 = const()[name = tensor("reshape_28_shape_0"), val = tensor([-1])]; + tensor reshape_28_cast_fp16 = reshape(shape = reshape_28_shape_0, x = reshape_24_cast_fp16)[name = tensor("reshape_28_cast_fp16")]; + tensor scatter_5_mode_0 = const()[name = tensor("scatter_5_mode_0"), val = tensor("update")]; + tensor scatter_5_axis_0 = const()[name = tensor("scatter_5_axis_0"), val = tensor(0)]; + tensor scatter_5_validate_indices_0 = const()[name = tensor("scatter_5_validate_indices_0"), val = tensor(false)]; + tensor scatter_5_cast_fp16 = scatter(axis = scatter_5_axis_0, data = reshape_28_cast_fp16, indices = reshape_26, mode = scatter_5_mode_0, updates = reshape_27_cast_fp16, validate_indices = scatter_5_validate_indices_0)[name = tensor("scatter_5_cast_fp16")]; + tensor reshape_29_cast_fp16 = reshape(shape = shape_51, x = scatter_5_cast_fp16)[name = tensor("reshape_29_cast_fp16")]; + tensor reshape_29_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("reshape_29_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor keys_13_begin_0 = const()[name = tensor("keys_13_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor keys_13_end_0 = const()[name = tensor("keys_13_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor keys_13_end_mask_0 = const()[name = tensor("keys_13_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor keys_13_squeeze_mask_0 = const()[name = tensor("keys_13_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor keys_13_cast_fp16 = slice_by_index(begin = keys_13_begin_0, end = keys_13_end_0, end_mask = keys_13_end_mask_0, squeeze_mask = keys_13_squeeze_mask_0, x = reshape_29_cast_fp16)[name = tensor("keys_13_cast_fp16")]; + tensor values_13_begin_0 = const()[name = tensor("values_13_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor values_13_end_0 = const()[name = tensor("values_13_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor values_13_end_mask_0 = const()[name = tensor("values_13_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor values_13_squeeze_mask_0 = const()[name = tensor("values_13_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor values_13_cast_fp16 = slice_by_index(begin = values_13_begin_0, end = values_13_end_0, end_mask = values_13_end_mask_0, squeeze_mask = values_13_squeeze_mask_0, x = reshape_29_cast_fp16)[name = tensor("values_13_cast_fp16")]; + tensor var_1286_cast_fp16 = not_equal(x = keys_13_cast_fp16, y = keys_13_cast_fp16)[name = tensor("op_1286_cast_fp16")]; + tensor keys_15_cast_fp16 = select(a = var_470_to_fp16, b = keys_13_cast_fp16, cond = var_1286_cast_fp16)[name = tensor("keys_15_cast_fp16")]; + tensor var_1294_cast_fp16 = not_equal(x = values_13_cast_fp16, y = values_13_cast_fp16)[name = tensor("op_1294_cast_fp16")]; + tensor values_15_cast_fp16 = select(a = var_470_to_fp16, b = values_13_cast_fp16, cond = var_1294_cast_fp16)[name = tensor("values_15_cast_fp16")]; + tensor var_1318 = const()[name = tensor("op_1318"), val = tensor([0, 2, 1, 3])]; + tensor var_1334 = const()[name = tensor("op_1334"), val = tensor([1, 1, 1])]; + tensor var_1335_cast_fp16 = reshape(shape = var_1334, x = position2_to_fp16)[name = tensor("op_1335_cast_fp16")]; + tensor q_positions_5_cast_fp16 = add(x = var_1335_cast_fp16, y = q_offsets_1_promoted_to_fp16)[name = tensor("q_positions_5_cast_fp16")]; + tensor valid_end_5_cast_fp16 = add(x = var_1335_cast_fp16, y = var_536_cast_fp16)[name = tensor("valid_end_5_cast_fp16")]; + tensor valid_mask_5_cast_fp16 = less(x = k_positions_1_promoted_to_fp16, y = valid_end_5_cast_fp16)[name = tensor("valid_mask_5_cast_fp16")]; + tensor causal_mask_5_cast_fp16 = less_equal(x = k_positions_1_promoted_to_fp16, y = q_positions_5_cast_fp16)[name = tensor("causal_mask_5_cast_fp16")]; + tensor var_1363 = logical_and(x = valid_mask_5_cast_fp16, y = causal_mask_5_cast_fp16)[name = tensor("op_1363")]; + tensor attn_mask_5_axes_0 = const()[name = tensor("attn_mask_5_axes_0"), val = tensor([1])]; + tensor attn_mask_5 = expand_dims(axes = attn_mask_5_axes_0, x = var_1363)[name = tensor("attn_mask_5")]; + tensor var_1377_transpose_x_0 = const()[name = tensor("op_1377_transpose_x_0"), val = tensor(false)]; + tensor var_1377_transpose_y_0 = const()[name = tensor("op_1377_transpose_y_0"), val = tensor(false)]; + tensor transpose_73_perm_0 = const()[name = tensor("transpose_73_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_74_perm_0 = const()[name = tensor("transpose_74_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_74 = transpose(perm = transpose_74_perm_0, x = keys_15_cast_fp16)[name = tensor("transpose_196")]; + tensor transpose_73 = transpose(perm = transpose_73_perm_0, x = q_15_cast_fp16)[name = tensor("transpose_197")]; + tensor var_1377_cast_fp16 = matmul(transpose_x = var_1377_transpose_x_0, transpose_y = var_1377_transpose_y_0, x = transpose_73, y = transpose_74)[name = tensor("op_1377_cast_fp16")]; + tensor var_1371_to_fp16 = const()[name = tensor("op_1371_to_fp16"), val = tensor([0x1p-3])]; + tensor attn_weights_13_cast_fp16 = mul(x = var_1377_cast_fp16, y = var_1371_to_fp16)[name = tensor("attn_weights_13_cast_fp16")]; + tensor var_1379 = logical_not(x = attn_mask_5)[name = tensor("op_1379")]; + tensor var_1380_to_fp16 = const()[name = tensor("op_1380_to_fp16"), val = tensor(-inf)]; + tensor attn_weights_15_cast_fp16 = select(a = var_1380_to_fp16, b = attn_weights_13_cast_fp16, cond = var_1379)[name = tensor("attn_weights_15_cast_fp16")]; + tensor var_1382 = const()[name = tensor("op_1382"), val = tensor(-1)]; + tensor attn_weights_17_cast_fp16 = softmax(axis = var_1382, x = attn_weights_15_cast_fp16)[name = tensor("attn_weights_17_cast_fp16")]; + tensor attn_output_5_transpose_x_0 = const()[name = tensor("attn_output_5_transpose_x_0"), val = tensor(false)]; + tensor attn_output_5_transpose_y_0 = const()[name = tensor("attn_output_5_transpose_y_0"), val = tensor(false)]; + tensor values_17_cast_fp16 = transpose(perm = var_1318, x = values_15_cast_fp16)[name = tensor("transpose_198")]; + tensor attn_output_5_cast_fp16 = matmul(transpose_x = attn_output_5_transpose_x_0, transpose_y = attn_output_5_transpose_y_0, x = attn_weights_17_cast_fp16, y = values_17_cast_fp16)[name = tensor("attn_output_5_cast_fp16")]; + tensor var_1390 = const()[name = tensor("op_1390"), val = tensor([0, 2, 1, 3])]; + tensor var_1395 = const()[name = tensor("op_1395"), val = tensor([1, 256, 1024])]; + tensor var_1391_cast_fp16 = transpose(perm = var_1390, x = attn_output_5_cast_fp16)[name = tensor("transpose_195")]; + tensor input_29_cast_fp16 = reshape(shape = var_1395, x = var_1391_cast_fp16)[name = tensor("input_29_cast_fp16")]; + tensor attn2_out_proj_weight_to_fp16 = const()[name = tensor("attn2_out_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(57713280)))]; + tensor linear_9_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = attn2_out_proj_weight_to_fp16, x = input_29_cast_fp16)[name = tensor("linear_9_cast_fp16")]; + tensor var_1404_cast_fp16 = add(x = position2_to_fp16, y = valid_len_to_fp16)[name = tensor("op_1404_cast_fp16")]; + tensor var_1404_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_1404_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor input_31_cast_fp16 = add(x = input_25_cast_fp16, y = linear_9_cast_fp16)[name = tensor("input_31_cast_fp16")]; + tensor input_33_axes_0 = const()[name = tensor("input_33_axes_0"), val = tensor([-1])]; + tensor norm2_2_weight_to_fp16 = const()[name = tensor("norm2_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(59810496)))]; + tensor norm2_2_bias_to_fp16 = const()[name = tensor("norm2_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(59812608)))]; + tensor var_1408_to_fp16 = const()[name = tensor("op_1408_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_33_cast_fp16 = layer_norm(axes = input_33_axes_0, beta = norm2_2_bias_to_fp16, epsilon = var_1408_to_fp16, gamma = norm2_2_weight_to_fp16, x = input_31_cast_fp16)[name = tensor("input_33_cast_fp16")]; + tensor linear2_1_weight_to_fp16 = const()[name = tensor("linear2_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(59814720)))]; + tensor linear_10_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = linear2_1_weight_to_fp16, x = input_33_cast_fp16)[name = tensor("linear_10_cast_fp16")]; + tensor input_35_mode_0 = const()[name = tensor("input_35_mode_0"), val = tensor("EXACT")]; + tensor input_35_cast_fp16 = gelu(mode = input_35_mode_0, x = linear_10_cast_fp16)[name = tensor("input_35_cast_fp16")]; + tensor linear2_2_weight_to_fp16 = const()[name = tensor("linear2_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(68203392)))]; + tensor linear_11_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = linear2_2_weight_to_fp16, x = input_35_cast_fp16)[name = tensor("linear_11_cast_fp16")]; + tensor input_37_cast_fp16 = add(x = input_31_cast_fp16, y = linear_11_cast_fp16)[name = tensor("input_37_cast_fp16")]; + tensor input_39_axes_0 = const()[name = tensor("input_39_axes_0"), val = tensor([-1])]; + tensor norm3_1_weight_to_fp16 = const()[name = tensor("norm3_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(76592064)))]; + tensor norm3_1_bias_to_fp16 = const()[name = tensor("norm3_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(76594176)))]; + tensor var_1425_to_fp16 = const()[name = tensor("op_1425_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_39_cast_fp16 = layer_norm(axes = input_39_axes_0, beta = norm3_1_bias_to_fp16, epsilon = var_1425_to_fp16, gamma = norm3_1_weight_to_fp16, x = input_37_cast_fp16)[name = tensor("input_39_cast_fp16")]; + tensor attn3_in_proj_weight_to_fp16 = const()[name = tensor("attn3_in_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(76596288)))]; + tensor linear_12_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = attn3_in_proj_weight_to_fp16, x = input_39_cast_fp16)[name = tensor("linear_12_cast_fp16")]; + tensor var_1439 = const()[name = tensor("op_1439"), val = tensor([1, 256, 3, 16, 64])]; + tensor qkv_7_cast_fp16 = reshape(shape = var_1439, x = linear_12_cast_fp16)[name = tensor("qkv_7_cast_fp16")]; + tensor q_19_begin_0 = const()[name = tensor("q_19_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor q_19_end_0 = const()[name = tensor("q_19_end_0"), val = tensor([1, 256, 1, 16, 64])]; + tensor q_19_end_mask_0 = const()[name = tensor("q_19_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor q_19_squeeze_mask_0 = const()[name = tensor("q_19_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor q_19_cast_fp16 = slice_by_index(begin = q_19_begin_0, end = q_19_end_0, end_mask = q_19_end_mask_0, squeeze_mask = q_19_squeeze_mask_0, x = qkv_7_cast_fp16)[name = tensor("q_19_cast_fp16")]; + tensor k_13_begin_0 = const()[name = tensor("k_13_begin_0"), val = tensor([0, 0, 1, 0, 0])]; + tensor k_13_end_0 = const()[name = tensor("k_13_end_0"), val = tensor([1, 256, 2, 16, 64])]; + tensor k_13_end_mask_0 = const()[name = tensor("k_13_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor k_13_squeeze_mask_0 = const()[name = tensor("k_13_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor k_13_cast_fp16 = slice_by_index(begin = k_13_begin_0, end = k_13_end_0, end_mask = k_13_end_mask_0, squeeze_mask = k_13_squeeze_mask_0, x = qkv_7_cast_fp16)[name = tensor("k_13_cast_fp16")]; + tensor v_7_begin_0 = const()[name = tensor("v_7_begin_0"), val = tensor([0, 0, 2, 0, 0])]; + tensor v_7_end_0 = const()[name = tensor("v_7_end_0"), val = tensor([1, 256, 3, 16, 64])]; + tensor v_7_end_mask_0 = const()[name = tensor("v_7_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor v_7_squeeze_mask_0 = const()[name = tensor("v_7_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor v_7_cast_fp16 = slice_by_index(begin = v_7_begin_0, end = v_7_end_0, end_mask = v_7_end_mask_0, squeeze_mask = v_7_squeeze_mask_0, x = qkv_7_cast_fp16)[name = tensor("v_7_cast_fp16")]; + tensor var_1497 = const()[name = tensor("op_1497"), val = tensor([1, 1])]; + tensor position3_to_fp16_dtype_0 = const()[name = tensor("position3_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor position3_to_fp16 = cast(dtype = position3_to_fp16_dtype_0, x = position3)[name = tensor("cast_646")]; + tensor var_1498_cast_fp16 = reshape(shape = var_1497, x = position3_to_fp16)[name = tensor("op_1498_cast_fp16")]; + tensor ts_21_cast_fp16 = add(x = ts_1_promoted_to_fp16, y = var_1498_cast_fp16)[name = tensor("ts_21_cast_fp16")]; + tensor var_1505 = const()[name = tensor("op_1505"), val = tensor([1, 256, 1, 1])]; + tensor ts_23_cast_fp16 = reshape(shape = var_1505, x = ts_21_cast_fp16)[name = tensor("ts_23_cast_fp16")]; + tensor var_1512 = const()[name = tensor("op_1512"), val = tensor([1, 256, 16, 32, 2])]; + tensor q_complex_7_cast_fp16 = reshape(shape = var_1512, x = q_19_cast_fp16)[name = tensor("q_complex_7_cast_fp16")]; + tensor var_1519 = const()[name = tensor("op_1519"), val = tensor([1, 256, 16, 32, 2])]; + tensor k_complex_7_cast_fp16 = reshape(shape = var_1519, x = k_13_cast_fp16)[name = tensor("k_complex_7_cast_fp16")]; + tensor var_1523_begin_0 = const()[name = tensor("op_1523_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_1523_end_0 = const()[name = tensor("op_1523_end_0"), val = tensor([1, 256, 16, 32, 1])]; + tensor var_1523_end_mask_0 = const()[name = tensor("op_1523_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_1523_squeeze_mask_0 = const()[name = tensor("op_1523_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_1523_cast_fp16 = slice_by_index(begin = var_1523_begin_0, end = var_1523_end_0, end_mask = var_1523_end_mask_0, squeeze_mask = var_1523_squeeze_mask_0, x = q_complex_7_cast_fp16)[name = tensor("op_1523_cast_fp16")]; + tensor var_1531_begin_0 = const()[name = tensor("op_1531_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_1531_end_0 = const()[name = tensor("op_1531_end_0"), val = tensor([1, 256, 16, 32, 2])]; + tensor var_1531_end_mask_0 = const()[name = tensor("op_1531_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_1531_squeeze_mask_0 = const()[name = tensor("op_1531_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_1531_cast_fp16 = slice_by_index(begin = var_1531_begin_0, end = var_1531_end_0, end_mask = var_1531_end_mask_0, squeeze_mask = var_1531_squeeze_mask_0, x = q_complex_7_cast_fp16)[name = tensor("op_1531_cast_fp16")]; + tensor var_1539_begin_0 = const()[name = tensor("op_1539_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_1539_end_0 = const()[name = tensor("op_1539_end_0"), val = tensor([1, 256, 16, 32, 1])]; + tensor var_1539_end_mask_0 = const()[name = tensor("op_1539_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_1539_squeeze_mask_0 = const()[name = tensor("op_1539_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_1539_cast_fp16 = slice_by_index(begin = var_1539_begin_0, end = var_1539_end_0, end_mask = var_1539_end_mask_0, squeeze_mask = var_1539_squeeze_mask_0, x = k_complex_7_cast_fp16)[name = tensor("op_1539_cast_fp16")]; + tensor var_1547_begin_0 = const()[name = tensor("op_1547_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_1547_end_0 = const()[name = tensor("op_1547_end_0"), val = tensor([1, 256, 16, 32, 2])]; + tensor var_1547_end_mask_0 = const()[name = tensor("op_1547_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_1547_squeeze_mask_0 = const()[name = tensor("op_1547_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_1547_cast_fp16 = slice_by_index(begin = var_1547_begin_0, end = var_1547_end_0, end_mask = var_1547_end_mask_0, squeeze_mask = var_1547_squeeze_mask_0, x = k_complex_7_cast_fp16)[name = tensor("op_1547_cast_fp16")]; + tensor freqs_7_to_fp16 = const()[name = tensor("freqs_7_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(82887808)))]; + tensor var_1553_cast_fp16 = mul(x = freqs_7_to_fp16, y = ts_23_cast_fp16)[name = tensor("op_1553_cast_fp16")]; + tensor rotr_7_cast_fp16 = cos(x = var_1553_cast_fp16)[name = tensor("rotr_7_cast_fp16")]; + tensor roti_7_cast_fp16 = sin(x = var_1553_cast_fp16)[name = tensor("roti_7_cast_fp16")]; + tensor var_1557_cast_fp16 = mul(x = var_1523_cast_fp16, y = rotr_7_cast_fp16)[name = tensor("op_1557_cast_fp16")]; + tensor var_1558_cast_fp16 = mul(x = var_1531_cast_fp16, y = roti_7_cast_fp16)[name = tensor("op_1558_cast_fp16")]; + tensor qor_13_cast_fp16 = sub(x = var_1557_cast_fp16, y = var_1558_cast_fp16)[name = tensor("qor_13_cast_fp16")]; + tensor var_1561_cast_fp16 = mul(x = var_1523_cast_fp16, y = roti_7_cast_fp16)[name = tensor("op_1561_cast_fp16")]; + tensor var_1562_cast_fp16 = mul(x = var_1531_cast_fp16, y = rotr_7_cast_fp16)[name = tensor("op_1562_cast_fp16")]; + tensor qoi_13_cast_fp16 = add(x = var_1561_cast_fp16, y = var_1562_cast_fp16)[name = tensor("qoi_13_cast_fp16")]; + tensor var_1565_cast_fp16 = mul(x = var_1539_cast_fp16, y = rotr_7_cast_fp16)[name = tensor("op_1565_cast_fp16")]; + tensor var_1566_cast_fp16 = mul(x = var_1547_cast_fp16, y = roti_7_cast_fp16)[name = tensor("op_1566_cast_fp16")]; + tensor kor_13_cast_fp16 = sub(x = var_1565_cast_fp16, y = var_1566_cast_fp16)[name = tensor("kor_13_cast_fp16")]; + tensor var_1569_cast_fp16 = mul(x = var_1539_cast_fp16, y = roti_7_cast_fp16)[name = tensor("op_1569_cast_fp16")]; + tensor var_1570_cast_fp16 = mul(x = var_1547_cast_fp16, y = rotr_7_cast_fp16)[name = tensor("op_1570_cast_fp16")]; + tensor koi_13_cast_fp16 = add(x = var_1569_cast_fp16, y = var_1570_cast_fp16)[name = tensor("koi_13_cast_fp16")]; + tensor qo_7_axis_0 = const()[name = tensor("qo_7_axis_0"), val = tensor(-1)]; + tensor qo_7_cast_fp16 = stack(axis = qo_7_axis_0, values = (qor_13_cast_fp16, qoi_13_cast_fp16))[name = tensor("qo_7_cast_fp16")]; + tensor ko_7_axis_0 = const()[name = tensor("ko_7_axis_0"), val = tensor(-1)]; + tensor ko_7_cast_fp16 = stack(axis = ko_7_axis_0, values = (kor_13_cast_fp16, koi_13_cast_fp16))[name = tensor("ko_7_cast_fp16")]; + tensor var_1603 = const()[name = tensor("op_1603"), val = tensor([1, 256, 16, 64])]; + tensor q_21_cast_fp16 = reshape(shape = var_1603, x = qo_7_cast_fp16)[name = tensor("q_21_cast_fp16")]; + tensor var_1609 = const()[name = tensor("op_1609"), val = tensor([1, 256, 16, 64])]; + tensor k_15_cast_fp16 = reshape(shape = var_1609, x = ko_7_cast_fp16)[name = tensor("k_15_cast_fp16")]; + tensor write_indices_float_19_cast_fp16 = add(x = var_1498_cast_fp16, y = var_393_promoted_to_fp16)[name = tensor("write_indices_float_19_cast_fp16")]; + tensor _inversed_1630_y_0_to_fp16 = const()[name = tensor("_inversed_1630_y_0_to_fp16"), val = tensor(0x1p-9)]; + tensor _inversed_1630_cast_fp16 = mul(x = write_indices_float_19_cast_fp16, y = _inversed_1630_y_0_to_fp16)[name = tensor("_inversed_1630_cast_fp16")]; + tensor var_1631_cast_fp16 = floor(x = _inversed_1630_cast_fp16)[name = tensor("op_1631_cast_fp16")]; + tensor var_1632_to_fp16 = const()[name = tensor("op_1632_to_fp16"), val = tensor(0x1p+9)]; + tensor var_1633_cast_fp16 = mul(x = var_1631_cast_fp16, y = var_1632_to_fp16)[name = tensor("op_1633_cast_fp16")]; + tensor write_indices_float_21_cast_fp16 = sub(x = write_indices_float_19_cast_fp16, y = var_1633_cast_fp16)[name = tensor("write_indices_float_21_cast_fp16")]; + tensor var_1651_value_0_to_fp16 = const()[name = tensor("op_1651_value_0_to_fp16"), val = tensor(0x1.ffp+8)]; + tensor var_1651_cast_fp16 = fill_like(ref_tensor = write_indices_float_21_cast_fp16, value = var_1651_value_0_to_fp16)[name = tensor("op_1651_cast_fp16")]; + tensor write_indices_float_23_cast_fp16 = select(a = write_indices_float_21_cast_fp16, b = var_1651_cast_fp16, cond = token_valid_1_cast_fp16)[name = tensor("write_indices_float_23_cast_fp16")]; + tensor var_1657_dtype_0 = const()[name = tensor("op_1657_dtype_0"), val = tensor("int32")]; + tensor var_1662 = const()[name = tensor("op_1662"), val = tensor([1, 256, 1, 1])]; + tensor write_indices_float_23_cast_fp16_to_int32 = cast(dtype = var_1657_dtype_0, x = write_indices_float_23_cast_fp16)[name = tensor("cast_645")]; + tensor var_1663 = reshape(shape = var_1662, x = write_indices_float_23_cast_fp16_to_int32)[name = tensor("op_1663")]; + tensor write_indices_7_reps_0 = const()[name = tensor("write_indices_7_reps_0"), val = tensor([1, 1, 16, 64])]; + tensor write_indices_7 = tile(reps = write_indices_7_reps_0, x = var_1663)[name = tensor("write_indices_7")]; + tensor var_1673_begin_0 = const()[name = tensor("op_1673_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_1673_end_0 = const()[name = tensor("op_1673_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor var_1673_end_mask_0 = const()[name = tensor("op_1673_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_1673_squeeze_mask_0 = const()[name = tensor("op_1673_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor cache3_to_fp16_dtype_0 = const()[name = tensor("cache3_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor cache3_to_fp16 = cast(dtype = cache3_to_fp16_dtype_0, x = cache3)[name = tensor("cast_644")]; + tensor var_1673_cast_fp16 = slice_by_index(begin = var_1673_begin_0, end = var_1673_end_0, end_mask = var_1673_end_mask_0, squeeze_mask = var_1673_squeeze_mask_0, x = cache3_to_fp16)[name = tensor("op_1673_cast_fp16")]; + tensor var_1675_axis_0 = const()[name = tensor("op_1675_axis_0"), val = tensor(1)]; + tensor var_1675_mode_0 = const()[name = tensor("op_1675_mode_0"), val = tensor("update")]; + tensor var_1675_validate_indices_0 = const()[name = tensor("op_1675_validate_indices_0"), val = tensor(false)]; + tensor var_1675_cast_fp16 = scatter_along_axis(axis = var_1675_axis_0, data = var_1673_cast_fp16, indices = write_indices_7, mode = var_1675_mode_0, updates = k_15_cast_fp16, validate_indices = var_1675_validate_indices_0)[name = tensor("op_1675_cast_fp16")]; + tensor concat_26 = const()[name = tensor("concat_26"), val = tensor([0, 0, 0, 0, 0])]; + tensor concat_27 = const()[name = tensor("concat_27"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_7_internal_tensor_assign_1_stride_0 = const()[name = tensor("new_cache_7_internal_tensor_assign_1_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_7_internal_tensor_assign_1_begin_mask_0 = const()[name = tensor("new_cache_7_internal_tensor_assign_1_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_7_internal_tensor_assign_1_end_mask_0 = const()[name = tensor("new_cache_7_internal_tensor_assign_1_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_7_internal_tensor_assign_1_squeeze_mask_0 = const()[name = tensor("new_cache_7_internal_tensor_assign_1_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_52 = const()[name = tensor("shape_52"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_6 = const()[name = tensor("reduce_prod_6"), val = tensor(1048576)]; + tensor range_1d_6_start_0 = const()[name = tensor("range_1d_6_start_0"), val = tensor(0)]; + tensor range_1d_6_step_0 = const()[name = tensor("range_1d_6_step_0"), val = tensor(1)]; + tensor range_1d_6 = range_1d(end = reduce_prod_6, start = range_1d_6_start_0, step = range_1d_6_step_0)[name = tensor("range_1d_6")]; + tensor reshape_30 = reshape(shape = shape_52, x = range_1d_6)[name = tensor("reshape_30")]; + tensor slice_by_index_6 = slice_by_index(begin = concat_26, begin_mask = new_cache_7_internal_tensor_assign_1_begin_mask_0, end = concat_27, end_mask = new_cache_7_internal_tensor_assign_1_end_mask_0, squeeze_mask = new_cache_7_internal_tensor_assign_1_squeeze_mask_0, stride = new_cache_7_internal_tensor_assign_1_stride_0, x = reshape_30)[name = tensor("slice_by_index_6")]; + tensor reshape_31_shape_0 = const()[name = tensor("reshape_31_shape_0"), val = tensor([-1])]; + tensor reshape_31 = reshape(shape = reshape_31_shape_0, x = slice_by_index_6)[name = tensor("reshape_31")]; + tensor reshape_32_shape_0 = const()[name = tensor("reshape_32_shape_0"), val = tensor([-1])]; + tensor reshape_32_cast_fp16 = reshape(shape = reshape_32_shape_0, x = var_1675_cast_fp16)[name = tensor("reshape_32_cast_fp16")]; + tensor reshape_33_shape_0 = const()[name = tensor("reshape_33_shape_0"), val = tensor([-1])]; + tensor reshape_33_cast_fp16 = reshape(shape = reshape_33_shape_0, x = cache3_to_fp16)[name = tensor("reshape_33_cast_fp16")]; + tensor scatter_6_mode_0 = const()[name = tensor("scatter_6_mode_0"), val = tensor("update")]; + tensor scatter_6_axis_0 = const()[name = tensor("scatter_6_axis_0"), val = tensor(0)]; + tensor scatter_6_validate_indices_0 = const()[name = tensor("scatter_6_validate_indices_0"), val = tensor(false)]; + tensor scatter_6_cast_fp16 = scatter(axis = scatter_6_axis_0, data = reshape_33_cast_fp16, indices = reshape_31, mode = scatter_6_mode_0, updates = reshape_32_cast_fp16, validate_indices = scatter_6_validate_indices_0)[name = tensor("scatter_6_cast_fp16")]; + tensor reshape_34_cast_fp16 = reshape(shape = shape_52, x = scatter_6_cast_fp16)[name = tensor("reshape_34_cast_fp16")]; + tensor var_1683_begin_0 = const()[name = tensor("op_1683_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor var_1683_end_0 = const()[name = tensor("op_1683_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor var_1683_end_mask_0 = const()[name = tensor("op_1683_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_1683_squeeze_mask_0 = const()[name = tensor("op_1683_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor var_1683_cast_fp16 = slice_by_index(begin = var_1683_begin_0, end = var_1683_end_0, end_mask = var_1683_end_mask_0, squeeze_mask = var_1683_squeeze_mask_0, x = reshape_34_cast_fp16)[name = tensor("op_1683_cast_fp16")]; + tensor var_1685_axis_0 = const()[name = tensor("op_1685_axis_0"), val = tensor(1)]; + tensor var_1685_mode_0 = const()[name = tensor("op_1685_mode_0"), val = tensor("update")]; + tensor var_1685_validate_indices_0 = const()[name = tensor("op_1685_validate_indices_0"), val = tensor(false)]; + tensor var_1685_cast_fp16 = scatter_along_axis(axis = var_1685_axis_0, data = var_1683_cast_fp16, indices = write_indices_7, mode = var_1685_mode_0, updates = v_7_cast_fp16, validate_indices = var_1685_validate_indices_0)[name = tensor("op_1685_cast_fp16")]; + tensor concat_28 = const()[name = tensor("concat_28"), val = tensor([1, 0, 0, 0, 0])]; + tensor concat_29 = const()[name = tensor("concat_29"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_7_internal_tensor_assign_2_stride_0 = const()[name = tensor("new_cache_7_internal_tensor_assign_2_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_7_internal_tensor_assign_2_begin_mask_0 = const()[name = tensor("new_cache_7_internal_tensor_assign_2_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_7_internal_tensor_assign_2_end_mask_0 = const()[name = tensor("new_cache_7_internal_tensor_assign_2_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_7_internal_tensor_assign_2_squeeze_mask_0 = const()[name = tensor("new_cache_7_internal_tensor_assign_2_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_53 = const()[name = tensor("shape_53"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_7 = const()[name = tensor("reduce_prod_7"), val = tensor(1048576)]; + tensor range_1d_7_start_0 = const()[name = tensor("range_1d_7_start_0"), val = tensor(0)]; + tensor range_1d_7_step_0 = const()[name = tensor("range_1d_7_step_0"), val = tensor(1)]; + tensor range_1d_7 = range_1d(end = reduce_prod_7, start = range_1d_7_start_0, step = range_1d_7_step_0)[name = tensor("range_1d_7")]; + tensor reshape_35 = reshape(shape = shape_53, x = range_1d_7)[name = tensor("reshape_35")]; + tensor slice_by_index_7 = slice_by_index(begin = concat_28, begin_mask = new_cache_7_internal_tensor_assign_2_begin_mask_0, end = concat_29, end_mask = new_cache_7_internal_tensor_assign_2_end_mask_0, squeeze_mask = new_cache_7_internal_tensor_assign_2_squeeze_mask_0, stride = new_cache_7_internal_tensor_assign_2_stride_0, x = reshape_35)[name = tensor("slice_by_index_7")]; + tensor reshape_36_shape_0 = const()[name = tensor("reshape_36_shape_0"), val = tensor([-1])]; + tensor reshape_36 = reshape(shape = reshape_36_shape_0, x = slice_by_index_7)[name = tensor("reshape_36")]; + tensor reshape_37_shape_0 = const()[name = tensor("reshape_37_shape_0"), val = tensor([-1])]; + tensor reshape_37_cast_fp16 = reshape(shape = reshape_37_shape_0, x = var_1685_cast_fp16)[name = tensor("reshape_37_cast_fp16")]; + tensor reshape_38_shape_0 = const()[name = tensor("reshape_38_shape_0"), val = tensor([-1])]; + tensor reshape_38_cast_fp16 = reshape(shape = reshape_38_shape_0, x = reshape_34_cast_fp16)[name = tensor("reshape_38_cast_fp16")]; + tensor scatter_7_mode_0 = const()[name = tensor("scatter_7_mode_0"), val = tensor("update")]; + tensor scatter_7_axis_0 = const()[name = tensor("scatter_7_axis_0"), val = tensor(0)]; + tensor scatter_7_validate_indices_0 = const()[name = tensor("scatter_7_validate_indices_0"), val = tensor(false)]; + tensor scatter_7_cast_fp16 = scatter(axis = scatter_7_axis_0, data = reshape_38_cast_fp16, indices = reshape_36, mode = scatter_7_mode_0, updates = reshape_37_cast_fp16, validate_indices = scatter_7_validate_indices_0)[name = tensor("scatter_7_cast_fp16")]; + tensor reshape_39_cast_fp16 = reshape(shape = shape_53, x = scatter_7_cast_fp16)[name = tensor("reshape_39_cast_fp16")]; + tensor reshape_39_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("reshape_39_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor keys_19_begin_0 = const()[name = tensor("keys_19_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor keys_19_end_0 = const()[name = tensor("keys_19_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor keys_19_end_mask_0 = const()[name = tensor("keys_19_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor keys_19_squeeze_mask_0 = const()[name = tensor("keys_19_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor keys_19_cast_fp16 = slice_by_index(begin = keys_19_begin_0, end = keys_19_end_0, end_mask = keys_19_end_mask_0, squeeze_mask = keys_19_squeeze_mask_0, x = reshape_39_cast_fp16)[name = tensor("keys_19_cast_fp16")]; + tensor values_19_begin_0 = const()[name = tensor("values_19_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor values_19_end_0 = const()[name = tensor("values_19_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor values_19_end_mask_0 = const()[name = tensor("values_19_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor values_19_squeeze_mask_0 = const()[name = tensor("values_19_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor values_19_cast_fp16 = slice_by_index(begin = values_19_begin_0, end = values_19_end_0, end_mask = values_19_end_mask_0, squeeze_mask = values_19_squeeze_mask_0, x = reshape_39_cast_fp16)[name = tensor("values_19_cast_fp16")]; + tensor var_1697_cast_fp16 = not_equal(x = keys_19_cast_fp16, y = keys_19_cast_fp16)[name = tensor("op_1697_cast_fp16")]; + tensor keys_21_cast_fp16 = select(a = var_470_to_fp16, b = keys_19_cast_fp16, cond = var_1697_cast_fp16)[name = tensor("keys_21_cast_fp16")]; + tensor var_1705_cast_fp16 = not_equal(x = values_19_cast_fp16, y = values_19_cast_fp16)[name = tensor("op_1705_cast_fp16")]; + tensor values_21_cast_fp16 = select(a = var_470_to_fp16, b = values_19_cast_fp16, cond = var_1705_cast_fp16)[name = tensor("values_21_cast_fp16")]; + tensor var_1729 = const()[name = tensor("op_1729"), val = tensor([0, 2, 1, 3])]; + tensor var_1745 = const()[name = tensor("op_1745"), val = tensor([1, 1, 1])]; + tensor var_1746_cast_fp16 = reshape(shape = var_1745, x = position3_to_fp16)[name = tensor("op_1746_cast_fp16")]; + tensor q_positions_7_cast_fp16 = add(x = var_1746_cast_fp16, y = q_offsets_1_promoted_to_fp16)[name = tensor("q_positions_7_cast_fp16")]; + tensor valid_end_7_cast_fp16 = add(x = var_1746_cast_fp16, y = var_536_cast_fp16)[name = tensor("valid_end_7_cast_fp16")]; + tensor valid_mask_7_cast_fp16 = less(x = k_positions_1_promoted_to_fp16, y = valid_end_7_cast_fp16)[name = tensor("valid_mask_7_cast_fp16")]; + tensor causal_mask_7_cast_fp16 = less_equal(x = k_positions_1_promoted_to_fp16, y = q_positions_7_cast_fp16)[name = tensor("causal_mask_7_cast_fp16")]; + tensor var_1774 = logical_and(x = valid_mask_7_cast_fp16, y = causal_mask_7_cast_fp16)[name = tensor("op_1774")]; + tensor attn_mask_7_axes_0 = const()[name = tensor("attn_mask_7_axes_0"), val = tensor([1])]; + tensor attn_mask_7 = expand_dims(axes = attn_mask_7_axes_0, x = var_1774)[name = tensor("attn_mask_7")]; + tensor var_1788_transpose_x_0 = const()[name = tensor("op_1788_transpose_x_0"), val = tensor(false)]; + tensor var_1788_transpose_y_0 = const()[name = tensor("op_1788_transpose_y_0"), val = tensor(false)]; + tensor transpose_75_perm_0 = const()[name = tensor("transpose_75_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_76_perm_0 = const()[name = tensor("transpose_76_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_76 = transpose(perm = transpose_76_perm_0, x = keys_21_cast_fp16)[name = tensor("transpose_192")]; + tensor transpose_75 = transpose(perm = transpose_75_perm_0, x = q_21_cast_fp16)[name = tensor("transpose_193")]; + tensor var_1788_cast_fp16 = matmul(transpose_x = var_1788_transpose_x_0, transpose_y = var_1788_transpose_y_0, x = transpose_75, y = transpose_76)[name = tensor("op_1788_cast_fp16")]; + tensor var_1782_to_fp16 = const()[name = tensor("op_1782_to_fp16"), val = tensor([0x1p-3])]; + tensor attn_weights_19_cast_fp16 = mul(x = var_1788_cast_fp16, y = var_1782_to_fp16)[name = tensor("attn_weights_19_cast_fp16")]; + tensor var_1790 = logical_not(x = attn_mask_7)[name = tensor("op_1790")]; + tensor var_1791_to_fp16 = const()[name = tensor("op_1791_to_fp16"), val = tensor(-inf)]; + tensor attn_weights_21_cast_fp16 = select(a = var_1791_to_fp16, b = attn_weights_19_cast_fp16, cond = var_1790)[name = tensor("attn_weights_21_cast_fp16")]; + tensor var_1793 = const()[name = tensor("op_1793"), val = tensor(-1)]; + tensor attn_weights_23_cast_fp16 = softmax(axis = var_1793, x = attn_weights_21_cast_fp16)[name = tensor("attn_weights_23_cast_fp16")]; + tensor attn_output_7_transpose_x_0 = const()[name = tensor("attn_output_7_transpose_x_0"), val = tensor(false)]; + tensor attn_output_7_transpose_y_0 = const()[name = tensor("attn_output_7_transpose_y_0"), val = tensor(false)]; + tensor values_23_cast_fp16 = transpose(perm = var_1729, x = values_21_cast_fp16)[name = tensor("transpose_194")]; + tensor attn_output_7_cast_fp16 = matmul(transpose_x = attn_output_7_transpose_x_0, transpose_y = attn_output_7_transpose_y_0, x = attn_weights_23_cast_fp16, y = values_23_cast_fp16)[name = tensor("attn_output_7_cast_fp16")]; + tensor var_1801 = const()[name = tensor("op_1801"), val = tensor([0, 2, 1, 3])]; + tensor var_1806 = const()[name = tensor("op_1806"), val = tensor([1, 256, 1024])]; + tensor var_1802_cast_fp16 = transpose(perm = var_1801, x = attn_output_7_cast_fp16)[name = tensor("transpose_191")]; + tensor input_41_cast_fp16 = reshape(shape = var_1806, x = var_1802_cast_fp16)[name = tensor("input_41_cast_fp16")]; + tensor attn3_out_proj_weight_to_fp16 = const()[name = tensor("attn3_out_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(82887936)))]; + tensor linear_13_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = attn3_out_proj_weight_to_fp16, x = input_41_cast_fp16)[name = tensor("linear_13_cast_fp16")]; + tensor var_1815_cast_fp16 = add(x = position3_to_fp16, y = valid_len_to_fp16)[name = tensor("op_1815_cast_fp16")]; + tensor var_1815_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_1815_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor input_43_cast_fp16 = add(x = input_37_cast_fp16, y = linear_13_cast_fp16)[name = tensor("input_43_cast_fp16")]; + tensor input_45_axes_0 = const()[name = tensor("input_45_axes_0"), val = tensor([-1])]; + tensor norm3_2_weight_to_fp16 = const()[name = tensor("norm3_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(84985152)))]; + tensor norm3_2_bias_to_fp16 = const()[name = tensor("norm3_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(84987264)))]; + tensor var_1819_to_fp16 = const()[name = tensor("op_1819_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_45_cast_fp16 = layer_norm(axes = input_45_axes_0, beta = norm3_2_bias_to_fp16, epsilon = var_1819_to_fp16, gamma = norm3_2_weight_to_fp16, x = input_43_cast_fp16)[name = tensor("input_45_cast_fp16")]; + tensor linear3_1_weight_to_fp16 = const()[name = tensor("linear3_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(84989376)))]; + tensor linear_14_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = linear3_1_weight_to_fp16, x = input_45_cast_fp16)[name = tensor("linear_14_cast_fp16")]; + tensor input_47_mode_0 = const()[name = tensor("input_47_mode_0"), val = tensor("EXACT")]; + tensor input_47_cast_fp16 = gelu(mode = input_47_mode_0, x = linear_14_cast_fp16)[name = tensor("input_47_cast_fp16")]; + tensor linear3_2_weight_to_fp16 = const()[name = tensor("linear3_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(93378048)))]; + tensor linear_15_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = linear3_2_weight_to_fp16, x = input_47_cast_fp16)[name = tensor("linear_15_cast_fp16")]; + tensor input_49_cast_fp16 = add(x = input_43_cast_fp16, y = linear_15_cast_fp16)[name = tensor("input_49_cast_fp16")]; + tensor input_51_axes_0 = const()[name = tensor("input_51_axes_0"), val = tensor([-1])]; + tensor norm4_1_weight_to_fp16 = const()[name = tensor("norm4_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(101766720)))]; + tensor norm4_1_bias_to_fp16 = const()[name = tensor("norm4_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(101768832)))]; + tensor var_1836_to_fp16 = const()[name = tensor("op_1836_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_51_cast_fp16 = layer_norm(axes = input_51_axes_0, beta = norm4_1_bias_to_fp16, epsilon = var_1836_to_fp16, gamma = norm4_1_weight_to_fp16, x = input_49_cast_fp16)[name = tensor("input_51_cast_fp16")]; + tensor attn4_in_proj_weight_to_fp16 = const()[name = tensor("attn4_in_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(101770944)))]; + tensor linear_16_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = attn4_in_proj_weight_to_fp16, x = input_51_cast_fp16)[name = tensor("linear_16_cast_fp16")]; + tensor var_1850 = const()[name = tensor("op_1850"), val = tensor([1, 256, 3, 16, 64])]; + tensor qkv_9_cast_fp16 = reshape(shape = var_1850, x = linear_16_cast_fp16)[name = tensor("qkv_9_cast_fp16")]; + tensor q_25_begin_0 = const()[name = tensor("q_25_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor q_25_end_0 = const()[name = tensor("q_25_end_0"), val = tensor([1, 256, 1, 16, 64])]; + tensor q_25_end_mask_0 = const()[name = tensor("q_25_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor q_25_squeeze_mask_0 = const()[name = tensor("q_25_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor q_25_cast_fp16 = slice_by_index(begin = q_25_begin_0, end = q_25_end_0, end_mask = q_25_end_mask_0, squeeze_mask = q_25_squeeze_mask_0, x = qkv_9_cast_fp16)[name = tensor("q_25_cast_fp16")]; + tensor k_17_begin_0 = const()[name = tensor("k_17_begin_0"), val = tensor([0, 0, 1, 0, 0])]; + tensor k_17_end_0 = const()[name = tensor("k_17_end_0"), val = tensor([1, 256, 2, 16, 64])]; + tensor k_17_end_mask_0 = const()[name = tensor("k_17_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor k_17_squeeze_mask_0 = const()[name = tensor("k_17_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor k_17_cast_fp16 = slice_by_index(begin = k_17_begin_0, end = k_17_end_0, end_mask = k_17_end_mask_0, squeeze_mask = k_17_squeeze_mask_0, x = qkv_9_cast_fp16)[name = tensor("k_17_cast_fp16")]; + tensor v_9_begin_0 = const()[name = tensor("v_9_begin_0"), val = tensor([0, 0, 2, 0, 0])]; + tensor v_9_end_0 = const()[name = tensor("v_9_end_0"), val = tensor([1, 256, 3, 16, 64])]; + tensor v_9_end_mask_0 = const()[name = tensor("v_9_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor v_9_squeeze_mask_0 = const()[name = tensor("v_9_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor v_9_cast_fp16 = slice_by_index(begin = v_9_begin_0, end = v_9_end_0, end_mask = v_9_end_mask_0, squeeze_mask = v_9_squeeze_mask_0, x = qkv_9_cast_fp16)[name = tensor("v_9_cast_fp16")]; + tensor var_1908 = const()[name = tensor("op_1908"), val = tensor([1, 1])]; + tensor position4_to_fp16_dtype_0 = const()[name = tensor("position4_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor position4_to_fp16 = cast(dtype = position4_to_fp16_dtype_0, x = position4)[name = tensor("cast_641")]; + tensor var_1909_cast_fp16 = reshape(shape = var_1908, x = position4_to_fp16)[name = tensor("op_1909_cast_fp16")]; + tensor ts_27_cast_fp16 = add(x = ts_1_promoted_to_fp16, y = var_1909_cast_fp16)[name = tensor("ts_27_cast_fp16")]; + tensor var_1916 = const()[name = tensor("op_1916"), val = tensor([1, 256, 1, 1])]; + tensor ts_29_cast_fp16 = reshape(shape = var_1916, x = ts_27_cast_fp16)[name = tensor("ts_29_cast_fp16")]; + tensor var_1923 = const()[name = tensor("op_1923"), val = tensor([1, 256, 16, 32, 2])]; + tensor q_complex_9_cast_fp16 = reshape(shape = var_1923, x = q_25_cast_fp16)[name = tensor("q_complex_9_cast_fp16")]; + tensor var_1930 = const()[name = tensor("op_1930"), val = tensor([1, 256, 16, 32, 2])]; + tensor k_complex_9_cast_fp16 = reshape(shape = var_1930, x = k_17_cast_fp16)[name = tensor("k_complex_9_cast_fp16")]; + tensor var_1934_begin_0 = const()[name = tensor("op_1934_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_1934_end_0 = const()[name = tensor("op_1934_end_0"), val = tensor([1, 256, 16, 32, 1])]; + tensor var_1934_end_mask_0 = const()[name = tensor("op_1934_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_1934_squeeze_mask_0 = const()[name = tensor("op_1934_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_1934_cast_fp16 = slice_by_index(begin = var_1934_begin_0, end = var_1934_end_0, end_mask = var_1934_end_mask_0, squeeze_mask = var_1934_squeeze_mask_0, x = q_complex_9_cast_fp16)[name = tensor("op_1934_cast_fp16")]; + tensor var_1942_begin_0 = const()[name = tensor("op_1942_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_1942_end_0 = const()[name = tensor("op_1942_end_0"), val = tensor([1, 256, 16, 32, 2])]; + tensor var_1942_end_mask_0 = const()[name = tensor("op_1942_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_1942_squeeze_mask_0 = const()[name = tensor("op_1942_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_1942_cast_fp16 = slice_by_index(begin = var_1942_begin_0, end = var_1942_end_0, end_mask = var_1942_end_mask_0, squeeze_mask = var_1942_squeeze_mask_0, x = q_complex_9_cast_fp16)[name = tensor("op_1942_cast_fp16")]; + tensor var_1950_begin_0 = const()[name = tensor("op_1950_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_1950_end_0 = const()[name = tensor("op_1950_end_0"), val = tensor([1, 256, 16, 32, 1])]; + tensor var_1950_end_mask_0 = const()[name = tensor("op_1950_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_1950_squeeze_mask_0 = const()[name = tensor("op_1950_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_1950_cast_fp16 = slice_by_index(begin = var_1950_begin_0, end = var_1950_end_0, end_mask = var_1950_end_mask_0, squeeze_mask = var_1950_squeeze_mask_0, x = k_complex_9_cast_fp16)[name = tensor("op_1950_cast_fp16")]; + tensor var_1958_begin_0 = const()[name = tensor("op_1958_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_1958_end_0 = const()[name = tensor("op_1958_end_0"), val = tensor([1, 256, 16, 32, 2])]; + tensor var_1958_end_mask_0 = const()[name = tensor("op_1958_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_1958_squeeze_mask_0 = const()[name = tensor("op_1958_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_1958_cast_fp16 = slice_by_index(begin = var_1958_begin_0, end = var_1958_end_0, end_mask = var_1958_end_mask_0, squeeze_mask = var_1958_squeeze_mask_0, x = k_complex_9_cast_fp16)[name = tensor("op_1958_cast_fp16")]; + tensor freqs_9_to_fp16 = const()[name = tensor("freqs_9_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(108062464)))]; + tensor var_1964_cast_fp16 = mul(x = freqs_9_to_fp16, y = ts_29_cast_fp16)[name = tensor("op_1964_cast_fp16")]; + tensor rotr_9_cast_fp16 = cos(x = var_1964_cast_fp16)[name = tensor("rotr_9_cast_fp16")]; + tensor roti_9_cast_fp16 = sin(x = var_1964_cast_fp16)[name = tensor("roti_9_cast_fp16")]; + tensor var_1968_cast_fp16 = mul(x = var_1934_cast_fp16, y = rotr_9_cast_fp16)[name = tensor("op_1968_cast_fp16")]; + tensor var_1969_cast_fp16 = mul(x = var_1942_cast_fp16, y = roti_9_cast_fp16)[name = tensor("op_1969_cast_fp16")]; + tensor qor_17_cast_fp16 = sub(x = var_1968_cast_fp16, y = var_1969_cast_fp16)[name = tensor("qor_17_cast_fp16")]; + tensor var_1972_cast_fp16 = mul(x = var_1934_cast_fp16, y = roti_9_cast_fp16)[name = tensor("op_1972_cast_fp16")]; + tensor var_1973_cast_fp16 = mul(x = var_1942_cast_fp16, y = rotr_9_cast_fp16)[name = tensor("op_1973_cast_fp16")]; + tensor qoi_17_cast_fp16 = add(x = var_1972_cast_fp16, y = var_1973_cast_fp16)[name = tensor("qoi_17_cast_fp16")]; + tensor var_1976_cast_fp16 = mul(x = var_1950_cast_fp16, y = rotr_9_cast_fp16)[name = tensor("op_1976_cast_fp16")]; + tensor var_1977_cast_fp16 = mul(x = var_1958_cast_fp16, y = roti_9_cast_fp16)[name = tensor("op_1977_cast_fp16")]; + tensor kor_17_cast_fp16 = sub(x = var_1976_cast_fp16, y = var_1977_cast_fp16)[name = tensor("kor_17_cast_fp16")]; + tensor var_1980_cast_fp16 = mul(x = var_1950_cast_fp16, y = roti_9_cast_fp16)[name = tensor("op_1980_cast_fp16")]; + tensor var_1981_cast_fp16 = mul(x = var_1958_cast_fp16, y = rotr_9_cast_fp16)[name = tensor("op_1981_cast_fp16")]; + tensor koi_17_cast_fp16 = add(x = var_1980_cast_fp16, y = var_1981_cast_fp16)[name = tensor("koi_17_cast_fp16")]; + tensor qo_9_axis_0 = const()[name = tensor("qo_9_axis_0"), val = tensor(-1)]; + tensor qo_9_cast_fp16 = stack(axis = qo_9_axis_0, values = (qor_17_cast_fp16, qoi_17_cast_fp16))[name = tensor("qo_9_cast_fp16")]; + tensor ko_9_axis_0 = const()[name = tensor("ko_9_axis_0"), val = tensor(-1)]; + tensor ko_9_cast_fp16 = stack(axis = ko_9_axis_0, values = (kor_17_cast_fp16, koi_17_cast_fp16))[name = tensor("ko_9_cast_fp16")]; + tensor var_2014 = const()[name = tensor("op_2014"), val = tensor([1, 256, 16, 64])]; + tensor q_27_cast_fp16 = reshape(shape = var_2014, x = qo_9_cast_fp16)[name = tensor("q_27_cast_fp16")]; + tensor var_2020 = const()[name = tensor("op_2020"), val = tensor([1, 256, 16, 64])]; + tensor k_19_cast_fp16 = reshape(shape = var_2020, x = ko_9_cast_fp16)[name = tensor("k_19_cast_fp16")]; + tensor write_indices_float_25_cast_fp16 = add(x = var_1909_cast_fp16, y = var_393_promoted_to_fp16)[name = tensor("write_indices_float_25_cast_fp16")]; + tensor _inversed_2041_y_0_to_fp16 = const()[name = tensor("_inversed_2041_y_0_to_fp16"), val = tensor(0x1p-9)]; + tensor _inversed_2041_cast_fp16 = mul(x = write_indices_float_25_cast_fp16, y = _inversed_2041_y_0_to_fp16)[name = tensor("_inversed_2041_cast_fp16")]; + tensor var_2042_cast_fp16 = floor(x = _inversed_2041_cast_fp16)[name = tensor("op_2042_cast_fp16")]; + tensor var_2043_to_fp16 = const()[name = tensor("op_2043_to_fp16"), val = tensor(0x1p+9)]; + tensor var_2044_cast_fp16 = mul(x = var_2042_cast_fp16, y = var_2043_to_fp16)[name = tensor("op_2044_cast_fp16")]; + tensor write_indices_float_27_cast_fp16 = sub(x = write_indices_float_25_cast_fp16, y = var_2044_cast_fp16)[name = tensor("write_indices_float_27_cast_fp16")]; + tensor var_2062_value_0_to_fp16 = const()[name = tensor("op_2062_value_0_to_fp16"), val = tensor(0x1.ffp+8)]; + tensor var_2062_cast_fp16 = fill_like(ref_tensor = write_indices_float_27_cast_fp16, value = var_2062_value_0_to_fp16)[name = tensor("op_2062_cast_fp16")]; + tensor write_indices_float_29_cast_fp16 = select(a = write_indices_float_27_cast_fp16, b = var_2062_cast_fp16, cond = token_valid_1_cast_fp16)[name = tensor("write_indices_float_29_cast_fp16")]; + tensor var_2068_dtype_0 = const()[name = tensor("op_2068_dtype_0"), val = tensor("int32")]; + tensor var_2073 = const()[name = tensor("op_2073"), val = tensor([1, 256, 1, 1])]; + tensor write_indices_float_29_cast_fp16_to_int32 = cast(dtype = var_2068_dtype_0, x = write_indices_float_29_cast_fp16)[name = tensor("cast_640")]; + tensor var_2074 = reshape(shape = var_2073, x = write_indices_float_29_cast_fp16_to_int32)[name = tensor("op_2074")]; + tensor write_indices_9_reps_0 = const()[name = tensor("write_indices_9_reps_0"), val = tensor([1, 1, 16, 64])]; + tensor write_indices_9 = tile(reps = write_indices_9_reps_0, x = var_2074)[name = tensor("write_indices_9")]; + tensor var_2084_begin_0 = const()[name = tensor("op_2084_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_2084_end_0 = const()[name = tensor("op_2084_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor var_2084_end_mask_0 = const()[name = tensor("op_2084_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_2084_squeeze_mask_0 = const()[name = tensor("op_2084_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor cache4_to_fp16_dtype_0 = const()[name = tensor("cache4_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor cache4_to_fp16 = cast(dtype = cache4_to_fp16_dtype_0, x = cache4)[name = tensor("cast_639")]; + tensor var_2084_cast_fp16 = slice_by_index(begin = var_2084_begin_0, end = var_2084_end_0, end_mask = var_2084_end_mask_0, squeeze_mask = var_2084_squeeze_mask_0, x = cache4_to_fp16)[name = tensor("op_2084_cast_fp16")]; + tensor var_2086_axis_0 = const()[name = tensor("op_2086_axis_0"), val = tensor(1)]; + tensor var_2086_mode_0 = const()[name = tensor("op_2086_mode_0"), val = tensor("update")]; + tensor var_2086_validate_indices_0 = const()[name = tensor("op_2086_validate_indices_0"), val = tensor(false)]; + tensor var_2086_cast_fp16 = scatter_along_axis(axis = var_2086_axis_0, data = var_2084_cast_fp16, indices = write_indices_9, mode = var_2086_mode_0, updates = k_19_cast_fp16, validate_indices = var_2086_validate_indices_0)[name = tensor("op_2086_cast_fp16")]; + tensor concat_34 = const()[name = tensor("concat_34"), val = tensor([0, 0, 0, 0, 0])]; + tensor concat_35 = const()[name = tensor("concat_35"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_9_internal_tensor_assign_1_stride_0 = const()[name = tensor("new_cache_9_internal_tensor_assign_1_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_9_internal_tensor_assign_1_begin_mask_0 = const()[name = tensor("new_cache_9_internal_tensor_assign_1_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_9_internal_tensor_assign_1_end_mask_0 = const()[name = tensor("new_cache_9_internal_tensor_assign_1_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_9_internal_tensor_assign_1_squeeze_mask_0 = const()[name = tensor("new_cache_9_internal_tensor_assign_1_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_54 = const()[name = tensor("shape_54"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_8 = const()[name = tensor("reduce_prod_8"), val = tensor(1048576)]; + tensor range_1d_8_start_0 = const()[name = tensor("range_1d_8_start_0"), val = tensor(0)]; + tensor range_1d_8_step_0 = const()[name = tensor("range_1d_8_step_0"), val = tensor(1)]; + tensor range_1d_8 = range_1d(end = reduce_prod_8, start = range_1d_8_start_0, step = range_1d_8_step_0)[name = tensor("range_1d_8")]; + tensor reshape_40 = reshape(shape = shape_54, x = range_1d_8)[name = tensor("reshape_40")]; + tensor slice_by_index_8 = slice_by_index(begin = concat_34, begin_mask = new_cache_9_internal_tensor_assign_1_begin_mask_0, end = concat_35, end_mask = new_cache_9_internal_tensor_assign_1_end_mask_0, squeeze_mask = new_cache_9_internal_tensor_assign_1_squeeze_mask_0, stride = new_cache_9_internal_tensor_assign_1_stride_0, x = reshape_40)[name = tensor("slice_by_index_8")]; + tensor reshape_41_shape_0 = const()[name = tensor("reshape_41_shape_0"), val = tensor([-1])]; + tensor reshape_41 = reshape(shape = reshape_41_shape_0, x = slice_by_index_8)[name = tensor("reshape_41")]; + tensor reshape_42_shape_0 = const()[name = tensor("reshape_42_shape_0"), val = tensor([-1])]; + tensor reshape_42_cast_fp16 = reshape(shape = reshape_42_shape_0, x = var_2086_cast_fp16)[name = tensor("reshape_42_cast_fp16")]; + tensor reshape_43_shape_0 = const()[name = tensor("reshape_43_shape_0"), val = tensor([-1])]; + tensor reshape_43_cast_fp16 = reshape(shape = reshape_43_shape_0, x = cache4_to_fp16)[name = tensor("reshape_43_cast_fp16")]; + tensor scatter_8_mode_0 = const()[name = tensor("scatter_8_mode_0"), val = tensor("update")]; + tensor scatter_8_axis_0 = const()[name = tensor("scatter_8_axis_0"), val = tensor(0)]; + tensor scatter_8_validate_indices_0 = const()[name = tensor("scatter_8_validate_indices_0"), val = tensor(false)]; + tensor scatter_8_cast_fp16 = scatter(axis = scatter_8_axis_0, data = reshape_43_cast_fp16, indices = reshape_41, mode = scatter_8_mode_0, updates = reshape_42_cast_fp16, validate_indices = scatter_8_validate_indices_0)[name = tensor("scatter_8_cast_fp16")]; + tensor reshape_44_cast_fp16 = reshape(shape = shape_54, x = scatter_8_cast_fp16)[name = tensor("reshape_44_cast_fp16")]; + tensor var_2094_begin_0 = const()[name = tensor("op_2094_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor var_2094_end_0 = const()[name = tensor("op_2094_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor var_2094_end_mask_0 = const()[name = tensor("op_2094_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_2094_squeeze_mask_0 = const()[name = tensor("op_2094_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor var_2094_cast_fp16 = slice_by_index(begin = var_2094_begin_0, end = var_2094_end_0, end_mask = var_2094_end_mask_0, squeeze_mask = var_2094_squeeze_mask_0, x = reshape_44_cast_fp16)[name = tensor("op_2094_cast_fp16")]; + tensor var_2096_axis_0 = const()[name = tensor("op_2096_axis_0"), val = tensor(1)]; + tensor var_2096_mode_0 = const()[name = tensor("op_2096_mode_0"), val = tensor("update")]; + tensor var_2096_validate_indices_0 = const()[name = tensor("op_2096_validate_indices_0"), val = tensor(false)]; + tensor var_2096_cast_fp16 = scatter_along_axis(axis = var_2096_axis_0, data = var_2094_cast_fp16, indices = write_indices_9, mode = var_2096_mode_0, updates = v_9_cast_fp16, validate_indices = var_2096_validate_indices_0)[name = tensor("op_2096_cast_fp16")]; + tensor concat_36 = const()[name = tensor("concat_36"), val = tensor([1, 0, 0, 0, 0])]; + tensor concat_37 = const()[name = tensor("concat_37"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_9_internal_tensor_assign_2_stride_0 = const()[name = tensor("new_cache_9_internal_tensor_assign_2_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_9_internal_tensor_assign_2_begin_mask_0 = const()[name = tensor("new_cache_9_internal_tensor_assign_2_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_9_internal_tensor_assign_2_end_mask_0 = const()[name = tensor("new_cache_9_internal_tensor_assign_2_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_9_internal_tensor_assign_2_squeeze_mask_0 = const()[name = tensor("new_cache_9_internal_tensor_assign_2_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_55 = const()[name = tensor("shape_55"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_9 = const()[name = tensor("reduce_prod_9"), val = tensor(1048576)]; + tensor range_1d_9_start_0 = const()[name = tensor("range_1d_9_start_0"), val = tensor(0)]; + tensor range_1d_9_step_0 = const()[name = tensor("range_1d_9_step_0"), val = tensor(1)]; + tensor range_1d_9 = range_1d(end = reduce_prod_9, start = range_1d_9_start_0, step = range_1d_9_step_0)[name = tensor("range_1d_9")]; + tensor reshape_45 = reshape(shape = shape_55, x = range_1d_9)[name = tensor("reshape_45")]; + tensor slice_by_index_9 = slice_by_index(begin = concat_36, begin_mask = new_cache_9_internal_tensor_assign_2_begin_mask_0, end = concat_37, end_mask = new_cache_9_internal_tensor_assign_2_end_mask_0, squeeze_mask = new_cache_9_internal_tensor_assign_2_squeeze_mask_0, stride = new_cache_9_internal_tensor_assign_2_stride_0, x = reshape_45)[name = tensor("slice_by_index_9")]; + tensor reshape_46_shape_0 = const()[name = tensor("reshape_46_shape_0"), val = tensor([-1])]; + tensor reshape_46 = reshape(shape = reshape_46_shape_0, x = slice_by_index_9)[name = tensor("reshape_46")]; + tensor reshape_47_shape_0 = const()[name = tensor("reshape_47_shape_0"), val = tensor([-1])]; + tensor reshape_47_cast_fp16 = reshape(shape = reshape_47_shape_0, x = var_2096_cast_fp16)[name = tensor("reshape_47_cast_fp16")]; + tensor reshape_48_shape_0 = const()[name = tensor("reshape_48_shape_0"), val = tensor([-1])]; + tensor reshape_48_cast_fp16 = reshape(shape = reshape_48_shape_0, x = reshape_44_cast_fp16)[name = tensor("reshape_48_cast_fp16")]; + tensor scatter_9_mode_0 = const()[name = tensor("scatter_9_mode_0"), val = tensor("update")]; + tensor scatter_9_axis_0 = const()[name = tensor("scatter_9_axis_0"), val = tensor(0)]; + tensor scatter_9_validate_indices_0 = const()[name = tensor("scatter_9_validate_indices_0"), val = tensor(false)]; + tensor scatter_9_cast_fp16 = scatter(axis = scatter_9_axis_0, data = reshape_48_cast_fp16, indices = reshape_46, mode = scatter_9_mode_0, updates = reshape_47_cast_fp16, validate_indices = scatter_9_validate_indices_0)[name = tensor("scatter_9_cast_fp16")]; + tensor reshape_49_cast_fp16 = reshape(shape = shape_55, x = scatter_9_cast_fp16)[name = tensor("reshape_49_cast_fp16")]; + tensor reshape_49_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("reshape_49_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor keys_25_begin_0 = const()[name = tensor("keys_25_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor keys_25_end_0 = const()[name = tensor("keys_25_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor keys_25_end_mask_0 = const()[name = tensor("keys_25_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor keys_25_squeeze_mask_0 = const()[name = tensor("keys_25_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor keys_25_cast_fp16 = slice_by_index(begin = keys_25_begin_0, end = keys_25_end_0, end_mask = keys_25_end_mask_0, squeeze_mask = keys_25_squeeze_mask_0, x = reshape_49_cast_fp16)[name = tensor("keys_25_cast_fp16")]; + tensor values_25_begin_0 = const()[name = tensor("values_25_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor values_25_end_0 = const()[name = tensor("values_25_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor values_25_end_mask_0 = const()[name = tensor("values_25_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor values_25_squeeze_mask_0 = const()[name = tensor("values_25_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor values_25_cast_fp16 = slice_by_index(begin = values_25_begin_0, end = values_25_end_0, end_mask = values_25_end_mask_0, squeeze_mask = values_25_squeeze_mask_0, x = reshape_49_cast_fp16)[name = tensor("values_25_cast_fp16")]; + tensor var_2108_cast_fp16 = not_equal(x = keys_25_cast_fp16, y = keys_25_cast_fp16)[name = tensor("op_2108_cast_fp16")]; + tensor keys_27_cast_fp16 = select(a = var_470_to_fp16, b = keys_25_cast_fp16, cond = var_2108_cast_fp16)[name = tensor("keys_27_cast_fp16")]; + tensor var_2116_cast_fp16 = not_equal(x = values_25_cast_fp16, y = values_25_cast_fp16)[name = tensor("op_2116_cast_fp16")]; + tensor values_27_cast_fp16 = select(a = var_470_to_fp16, b = values_25_cast_fp16, cond = var_2116_cast_fp16)[name = tensor("values_27_cast_fp16")]; + tensor var_2140 = const()[name = tensor("op_2140"), val = tensor([0, 2, 1, 3])]; + tensor var_2156 = const()[name = tensor("op_2156"), val = tensor([1, 1, 1])]; + tensor var_2157_cast_fp16 = reshape(shape = var_2156, x = position4_to_fp16)[name = tensor("op_2157_cast_fp16")]; + tensor q_positions_9_cast_fp16 = add(x = var_2157_cast_fp16, y = q_offsets_1_promoted_to_fp16)[name = tensor("q_positions_9_cast_fp16")]; + tensor valid_end_9_cast_fp16 = add(x = var_2157_cast_fp16, y = var_536_cast_fp16)[name = tensor("valid_end_9_cast_fp16")]; + tensor valid_mask_9_cast_fp16 = less(x = k_positions_1_promoted_to_fp16, y = valid_end_9_cast_fp16)[name = tensor("valid_mask_9_cast_fp16")]; + tensor causal_mask_9_cast_fp16 = less_equal(x = k_positions_1_promoted_to_fp16, y = q_positions_9_cast_fp16)[name = tensor("causal_mask_9_cast_fp16")]; + tensor var_2185 = logical_and(x = valid_mask_9_cast_fp16, y = causal_mask_9_cast_fp16)[name = tensor("op_2185")]; + tensor attn_mask_9_axes_0 = const()[name = tensor("attn_mask_9_axes_0"), val = tensor([1])]; + tensor attn_mask_9 = expand_dims(axes = attn_mask_9_axes_0, x = var_2185)[name = tensor("attn_mask_9")]; + tensor var_2199_transpose_x_0 = const()[name = tensor("op_2199_transpose_x_0"), val = tensor(false)]; + tensor var_2199_transpose_y_0 = const()[name = tensor("op_2199_transpose_y_0"), val = tensor(false)]; + tensor transpose_77_perm_0 = const()[name = tensor("transpose_77_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_78_perm_0 = const()[name = tensor("transpose_78_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_78 = transpose(perm = transpose_78_perm_0, x = keys_27_cast_fp16)[name = tensor("transpose_188")]; + tensor transpose_77 = transpose(perm = transpose_77_perm_0, x = q_27_cast_fp16)[name = tensor("transpose_189")]; + tensor var_2199_cast_fp16 = matmul(transpose_x = var_2199_transpose_x_0, transpose_y = var_2199_transpose_y_0, x = transpose_77, y = transpose_78)[name = tensor("op_2199_cast_fp16")]; + tensor var_2193_to_fp16 = const()[name = tensor("op_2193_to_fp16"), val = tensor([0x1p-3])]; + tensor attn_weights_25_cast_fp16 = mul(x = var_2199_cast_fp16, y = var_2193_to_fp16)[name = tensor("attn_weights_25_cast_fp16")]; + tensor var_2201 = logical_not(x = attn_mask_9)[name = tensor("op_2201")]; + tensor var_2202_to_fp16 = const()[name = tensor("op_2202_to_fp16"), val = tensor(-inf)]; + tensor attn_weights_27_cast_fp16 = select(a = var_2202_to_fp16, b = attn_weights_25_cast_fp16, cond = var_2201)[name = tensor("attn_weights_27_cast_fp16")]; + tensor var_2204 = const()[name = tensor("op_2204"), val = tensor(-1)]; + tensor attn_weights_29_cast_fp16 = softmax(axis = var_2204, x = attn_weights_27_cast_fp16)[name = tensor("attn_weights_29_cast_fp16")]; + tensor attn_output_9_transpose_x_0 = const()[name = tensor("attn_output_9_transpose_x_0"), val = tensor(false)]; + tensor attn_output_9_transpose_y_0 = const()[name = tensor("attn_output_9_transpose_y_0"), val = tensor(false)]; + tensor values_29_cast_fp16 = transpose(perm = var_2140, x = values_27_cast_fp16)[name = tensor("transpose_190")]; + tensor attn_output_9_cast_fp16 = matmul(transpose_x = attn_output_9_transpose_x_0, transpose_y = attn_output_9_transpose_y_0, x = attn_weights_29_cast_fp16, y = values_29_cast_fp16)[name = tensor("attn_output_9_cast_fp16")]; + tensor var_2212 = const()[name = tensor("op_2212"), val = tensor([0, 2, 1, 3])]; + tensor var_2217 = const()[name = tensor("op_2217"), val = tensor([1, 256, 1024])]; + tensor var_2213_cast_fp16 = transpose(perm = var_2212, x = attn_output_9_cast_fp16)[name = tensor("transpose_187")]; + tensor input_53_cast_fp16 = reshape(shape = var_2217, x = var_2213_cast_fp16)[name = tensor("input_53_cast_fp16")]; + tensor attn4_out_proj_weight_to_fp16 = const()[name = tensor("attn4_out_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(108062592)))]; + tensor linear_17_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = attn4_out_proj_weight_to_fp16, x = input_53_cast_fp16)[name = tensor("linear_17_cast_fp16")]; + tensor var_2226_cast_fp16 = add(x = position4_to_fp16, y = valid_len_to_fp16)[name = tensor("op_2226_cast_fp16")]; + tensor var_2226_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_2226_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor input_55_cast_fp16 = add(x = input_49_cast_fp16, y = linear_17_cast_fp16)[name = tensor("input_55_cast_fp16")]; + tensor input_57_axes_0 = const()[name = tensor("input_57_axes_0"), val = tensor([-1])]; + tensor norm4_2_weight_to_fp16 = const()[name = tensor("norm4_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(110159808)))]; + tensor norm4_2_bias_to_fp16 = const()[name = tensor("norm4_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(110161920)))]; + tensor var_2230_to_fp16 = const()[name = tensor("op_2230_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_57_cast_fp16 = layer_norm(axes = input_57_axes_0, beta = norm4_2_bias_to_fp16, epsilon = var_2230_to_fp16, gamma = norm4_2_weight_to_fp16, x = input_55_cast_fp16)[name = tensor("input_57_cast_fp16")]; + tensor linear4_1_weight_to_fp16 = const()[name = tensor("linear4_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(110164032)))]; + tensor linear_18_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = linear4_1_weight_to_fp16, x = input_57_cast_fp16)[name = tensor("linear_18_cast_fp16")]; + tensor input_59_mode_0 = const()[name = tensor("input_59_mode_0"), val = tensor("EXACT")]; + tensor input_59_cast_fp16 = gelu(mode = input_59_mode_0, x = linear_18_cast_fp16)[name = tensor("input_59_cast_fp16")]; + tensor linear4_2_weight_to_fp16 = const()[name = tensor("linear4_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(118552704)))]; + tensor linear_19_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = linear4_2_weight_to_fp16, x = input_59_cast_fp16)[name = tensor("linear_19_cast_fp16")]; + tensor input_61_cast_fp16 = add(x = input_55_cast_fp16, y = linear_19_cast_fp16)[name = tensor("input_61_cast_fp16")]; + tensor input_63_axes_0 = const()[name = tensor("input_63_axes_0"), val = tensor([-1])]; + tensor norm5_1_weight_to_fp16 = const()[name = tensor("norm5_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(126941376)))]; + tensor norm5_1_bias_to_fp16 = const()[name = tensor("norm5_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(126943488)))]; + tensor var_2247_to_fp16 = const()[name = tensor("op_2247_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_63_cast_fp16 = layer_norm(axes = input_63_axes_0, beta = norm5_1_bias_to_fp16, epsilon = var_2247_to_fp16, gamma = norm5_1_weight_to_fp16, x = input_61_cast_fp16)[name = tensor("input_63_cast_fp16")]; + tensor attn5_in_proj_weight_to_fp16 = const()[name = tensor("attn5_in_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(126945600)))]; + tensor linear_20_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = attn5_in_proj_weight_to_fp16, x = input_63_cast_fp16)[name = tensor("linear_20_cast_fp16")]; + tensor var_2261 = const()[name = tensor("op_2261"), val = tensor([1, 256, 3, 16, 64])]; + tensor qkv_11_cast_fp16 = reshape(shape = var_2261, x = linear_20_cast_fp16)[name = tensor("qkv_11_cast_fp16")]; + tensor q_31_begin_0 = const()[name = tensor("q_31_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor q_31_end_0 = const()[name = tensor("q_31_end_0"), val = tensor([1, 256, 1, 16, 64])]; + tensor q_31_end_mask_0 = const()[name = tensor("q_31_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor q_31_squeeze_mask_0 = const()[name = tensor("q_31_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor q_31_cast_fp16 = slice_by_index(begin = q_31_begin_0, end = q_31_end_0, end_mask = q_31_end_mask_0, squeeze_mask = q_31_squeeze_mask_0, x = qkv_11_cast_fp16)[name = tensor("q_31_cast_fp16")]; + tensor k_21_begin_0 = const()[name = tensor("k_21_begin_0"), val = tensor([0, 0, 1, 0, 0])]; + tensor k_21_end_0 = const()[name = tensor("k_21_end_0"), val = tensor([1, 256, 2, 16, 64])]; + tensor k_21_end_mask_0 = const()[name = tensor("k_21_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor k_21_squeeze_mask_0 = const()[name = tensor("k_21_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor k_21_cast_fp16 = slice_by_index(begin = k_21_begin_0, end = k_21_end_0, end_mask = k_21_end_mask_0, squeeze_mask = k_21_squeeze_mask_0, x = qkv_11_cast_fp16)[name = tensor("k_21_cast_fp16")]; + tensor v_11_begin_0 = const()[name = tensor("v_11_begin_0"), val = tensor([0, 0, 2, 0, 0])]; + tensor v_11_end_0 = const()[name = tensor("v_11_end_0"), val = tensor([1, 256, 3, 16, 64])]; + tensor v_11_end_mask_0 = const()[name = tensor("v_11_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor v_11_squeeze_mask_0 = const()[name = tensor("v_11_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor v_11_cast_fp16 = slice_by_index(begin = v_11_begin_0, end = v_11_end_0, end_mask = v_11_end_mask_0, squeeze_mask = v_11_squeeze_mask_0, x = qkv_11_cast_fp16)[name = tensor("v_11_cast_fp16")]; + tensor var_2319 = const()[name = tensor("op_2319"), val = tensor([1, 1])]; + tensor position5_to_fp16_dtype_0 = const()[name = tensor("position5_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor position5_to_fp16 = cast(dtype = position5_to_fp16_dtype_0, x = position5)[name = tensor("cast_636")]; + tensor var_2320_cast_fp16 = reshape(shape = var_2319, x = position5_to_fp16)[name = tensor("op_2320_cast_fp16")]; + tensor ts_33_cast_fp16 = add(x = ts_1_promoted_to_fp16, y = var_2320_cast_fp16)[name = tensor("ts_33_cast_fp16")]; + tensor var_2327 = const()[name = tensor("op_2327"), val = tensor([1, 256, 1, 1])]; + tensor ts_35_cast_fp16 = reshape(shape = var_2327, x = ts_33_cast_fp16)[name = tensor("ts_35_cast_fp16")]; + tensor var_2334 = const()[name = tensor("op_2334"), val = tensor([1, 256, 16, 32, 2])]; + tensor q_complex_11_cast_fp16 = reshape(shape = var_2334, x = q_31_cast_fp16)[name = tensor("q_complex_11_cast_fp16")]; + tensor var_2341 = const()[name = tensor("op_2341"), val = tensor([1, 256, 16, 32, 2])]; + tensor k_complex_11_cast_fp16 = reshape(shape = var_2341, x = k_21_cast_fp16)[name = tensor("k_complex_11_cast_fp16")]; + tensor var_2345_begin_0 = const()[name = tensor("op_2345_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_2345_end_0 = const()[name = tensor("op_2345_end_0"), val = tensor([1, 256, 16, 32, 1])]; + tensor var_2345_end_mask_0 = const()[name = tensor("op_2345_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_2345_squeeze_mask_0 = const()[name = tensor("op_2345_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_2345_cast_fp16 = slice_by_index(begin = var_2345_begin_0, end = var_2345_end_0, end_mask = var_2345_end_mask_0, squeeze_mask = var_2345_squeeze_mask_0, x = q_complex_11_cast_fp16)[name = tensor("op_2345_cast_fp16")]; + tensor var_2353_begin_0 = const()[name = tensor("op_2353_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_2353_end_0 = const()[name = tensor("op_2353_end_0"), val = tensor([1, 256, 16, 32, 2])]; + tensor var_2353_end_mask_0 = const()[name = tensor("op_2353_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_2353_squeeze_mask_0 = const()[name = tensor("op_2353_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_2353_cast_fp16 = slice_by_index(begin = var_2353_begin_0, end = var_2353_end_0, end_mask = var_2353_end_mask_0, squeeze_mask = var_2353_squeeze_mask_0, x = q_complex_11_cast_fp16)[name = tensor("op_2353_cast_fp16")]; + tensor var_2361_begin_0 = const()[name = tensor("op_2361_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_2361_end_0 = const()[name = tensor("op_2361_end_0"), val = tensor([1, 256, 16, 32, 1])]; + tensor var_2361_end_mask_0 = const()[name = tensor("op_2361_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_2361_squeeze_mask_0 = const()[name = tensor("op_2361_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_2361_cast_fp16 = slice_by_index(begin = var_2361_begin_0, end = var_2361_end_0, end_mask = var_2361_end_mask_0, squeeze_mask = var_2361_squeeze_mask_0, x = k_complex_11_cast_fp16)[name = tensor("op_2361_cast_fp16")]; + tensor var_2369_begin_0 = const()[name = tensor("op_2369_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_2369_end_0 = const()[name = tensor("op_2369_end_0"), val = tensor([1, 256, 16, 32, 2])]; + tensor var_2369_end_mask_0 = const()[name = tensor("op_2369_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_2369_squeeze_mask_0 = const()[name = tensor("op_2369_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_2369_cast_fp16 = slice_by_index(begin = var_2369_begin_0, end = var_2369_end_0, end_mask = var_2369_end_mask_0, squeeze_mask = var_2369_squeeze_mask_0, x = k_complex_11_cast_fp16)[name = tensor("op_2369_cast_fp16")]; + tensor freqs_11_to_fp16 = const()[name = tensor("freqs_11_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(133237120)))]; + tensor var_2375_cast_fp16 = mul(x = freqs_11_to_fp16, y = ts_35_cast_fp16)[name = tensor("op_2375_cast_fp16")]; + tensor rotr_11_cast_fp16 = cos(x = var_2375_cast_fp16)[name = tensor("rotr_11_cast_fp16")]; + tensor roti_11_cast_fp16 = sin(x = var_2375_cast_fp16)[name = tensor("roti_11_cast_fp16")]; + tensor var_2379_cast_fp16 = mul(x = var_2345_cast_fp16, y = rotr_11_cast_fp16)[name = tensor("op_2379_cast_fp16")]; + tensor var_2380_cast_fp16 = mul(x = var_2353_cast_fp16, y = roti_11_cast_fp16)[name = tensor("op_2380_cast_fp16")]; + tensor qor_21_cast_fp16 = sub(x = var_2379_cast_fp16, y = var_2380_cast_fp16)[name = tensor("qor_21_cast_fp16")]; + tensor var_2383_cast_fp16 = mul(x = var_2345_cast_fp16, y = roti_11_cast_fp16)[name = tensor("op_2383_cast_fp16")]; + tensor var_2384_cast_fp16 = mul(x = var_2353_cast_fp16, y = rotr_11_cast_fp16)[name = tensor("op_2384_cast_fp16")]; + tensor qoi_21_cast_fp16 = add(x = var_2383_cast_fp16, y = var_2384_cast_fp16)[name = tensor("qoi_21_cast_fp16")]; + tensor var_2387_cast_fp16 = mul(x = var_2361_cast_fp16, y = rotr_11_cast_fp16)[name = tensor("op_2387_cast_fp16")]; + tensor var_2388_cast_fp16 = mul(x = var_2369_cast_fp16, y = roti_11_cast_fp16)[name = tensor("op_2388_cast_fp16")]; + tensor kor_21_cast_fp16 = sub(x = var_2387_cast_fp16, y = var_2388_cast_fp16)[name = tensor("kor_21_cast_fp16")]; + tensor var_2391_cast_fp16 = mul(x = var_2361_cast_fp16, y = roti_11_cast_fp16)[name = tensor("op_2391_cast_fp16")]; + tensor var_2392_cast_fp16 = mul(x = var_2369_cast_fp16, y = rotr_11_cast_fp16)[name = tensor("op_2392_cast_fp16")]; + tensor koi_21_cast_fp16 = add(x = var_2391_cast_fp16, y = var_2392_cast_fp16)[name = tensor("koi_21_cast_fp16")]; + tensor qo_11_axis_0 = const()[name = tensor("qo_11_axis_0"), val = tensor(-1)]; + tensor qo_11_cast_fp16 = stack(axis = qo_11_axis_0, values = (qor_21_cast_fp16, qoi_21_cast_fp16))[name = tensor("qo_11_cast_fp16")]; + tensor ko_11_axis_0 = const()[name = tensor("ko_11_axis_0"), val = tensor(-1)]; + tensor ko_11_cast_fp16 = stack(axis = ko_11_axis_0, values = (kor_21_cast_fp16, koi_21_cast_fp16))[name = tensor("ko_11_cast_fp16")]; + tensor var_2425 = const()[name = tensor("op_2425"), val = tensor([1, 256, 16, 64])]; + tensor q_33_cast_fp16 = reshape(shape = var_2425, x = qo_11_cast_fp16)[name = tensor("q_33_cast_fp16")]; + tensor var_2431 = const()[name = tensor("op_2431"), val = tensor([1, 256, 16, 64])]; + tensor k_23_cast_fp16 = reshape(shape = var_2431, x = ko_11_cast_fp16)[name = tensor("k_23_cast_fp16")]; + tensor write_indices_float_31_cast_fp16 = add(x = var_2320_cast_fp16, y = var_393_promoted_to_fp16)[name = tensor("write_indices_float_31_cast_fp16")]; + tensor _inversed_2452_y_0_to_fp16 = const()[name = tensor("_inversed_2452_y_0_to_fp16"), val = tensor(0x1p-9)]; + tensor _inversed_2452_cast_fp16 = mul(x = write_indices_float_31_cast_fp16, y = _inversed_2452_y_0_to_fp16)[name = tensor("_inversed_2452_cast_fp16")]; + tensor var_2453_cast_fp16 = floor(x = _inversed_2452_cast_fp16)[name = tensor("op_2453_cast_fp16")]; + tensor var_2454_to_fp16 = const()[name = tensor("op_2454_to_fp16"), val = tensor(0x1p+9)]; + tensor var_2455_cast_fp16 = mul(x = var_2453_cast_fp16, y = var_2454_to_fp16)[name = tensor("op_2455_cast_fp16")]; + tensor write_indices_float_33_cast_fp16 = sub(x = write_indices_float_31_cast_fp16, y = var_2455_cast_fp16)[name = tensor("write_indices_float_33_cast_fp16")]; + tensor var_2473_value_0_to_fp16 = const()[name = tensor("op_2473_value_0_to_fp16"), val = tensor(0x1.ffp+8)]; + tensor var_2473_cast_fp16 = fill_like(ref_tensor = write_indices_float_33_cast_fp16, value = var_2473_value_0_to_fp16)[name = tensor("op_2473_cast_fp16")]; + tensor write_indices_float_35_cast_fp16 = select(a = write_indices_float_33_cast_fp16, b = var_2473_cast_fp16, cond = token_valid_1_cast_fp16)[name = tensor("write_indices_float_35_cast_fp16")]; + tensor var_2479_dtype_0 = const()[name = tensor("op_2479_dtype_0"), val = tensor("int32")]; + tensor var_2484 = const()[name = tensor("op_2484"), val = tensor([1, 256, 1, 1])]; + tensor write_indices_float_35_cast_fp16_to_int32 = cast(dtype = var_2479_dtype_0, x = write_indices_float_35_cast_fp16)[name = tensor("cast_635")]; + tensor var_2485 = reshape(shape = var_2484, x = write_indices_float_35_cast_fp16_to_int32)[name = tensor("op_2485")]; + tensor write_indices_11_reps_0 = const()[name = tensor("write_indices_11_reps_0"), val = tensor([1, 1, 16, 64])]; + tensor write_indices_11 = tile(reps = write_indices_11_reps_0, x = var_2485)[name = tensor("write_indices_11")]; + tensor var_2495_begin_0 = const()[name = tensor("op_2495_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_2495_end_0 = const()[name = tensor("op_2495_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor var_2495_end_mask_0 = const()[name = tensor("op_2495_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_2495_squeeze_mask_0 = const()[name = tensor("op_2495_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor cache5_to_fp16_dtype_0 = const()[name = tensor("cache5_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor cache5_to_fp16 = cast(dtype = cache5_to_fp16_dtype_0, x = cache5)[name = tensor("cast_634")]; + tensor var_2495_cast_fp16 = slice_by_index(begin = var_2495_begin_0, end = var_2495_end_0, end_mask = var_2495_end_mask_0, squeeze_mask = var_2495_squeeze_mask_0, x = cache5_to_fp16)[name = tensor("op_2495_cast_fp16")]; + tensor var_2497_axis_0 = const()[name = tensor("op_2497_axis_0"), val = tensor(1)]; + tensor var_2497_mode_0 = const()[name = tensor("op_2497_mode_0"), val = tensor("update")]; + tensor var_2497_validate_indices_0 = const()[name = tensor("op_2497_validate_indices_0"), val = tensor(false)]; + tensor var_2497_cast_fp16 = scatter_along_axis(axis = var_2497_axis_0, data = var_2495_cast_fp16, indices = write_indices_11, mode = var_2497_mode_0, updates = k_23_cast_fp16, validate_indices = var_2497_validate_indices_0)[name = tensor("op_2497_cast_fp16")]; + tensor concat_42 = const()[name = tensor("concat_42"), val = tensor([0, 0, 0, 0, 0])]; + tensor concat_43 = const()[name = tensor("concat_43"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_11_internal_tensor_assign_1_stride_0 = const()[name = tensor("new_cache_11_internal_tensor_assign_1_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_11_internal_tensor_assign_1_begin_mask_0 = const()[name = tensor("new_cache_11_internal_tensor_assign_1_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_11_internal_tensor_assign_1_end_mask_0 = const()[name = tensor("new_cache_11_internal_tensor_assign_1_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_11_internal_tensor_assign_1_squeeze_mask_0 = const()[name = tensor("new_cache_11_internal_tensor_assign_1_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_56 = const()[name = tensor("shape_56"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_10 = const()[name = tensor("reduce_prod_10"), val = tensor(1048576)]; + tensor range_1d_10_start_0 = const()[name = tensor("range_1d_10_start_0"), val = tensor(0)]; + tensor range_1d_10_step_0 = const()[name = tensor("range_1d_10_step_0"), val = tensor(1)]; + tensor range_1d_10 = range_1d(end = reduce_prod_10, start = range_1d_10_start_0, step = range_1d_10_step_0)[name = tensor("range_1d_10")]; + tensor reshape_50 = reshape(shape = shape_56, x = range_1d_10)[name = tensor("reshape_50")]; + tensor slice_by_index_10 = slice_by_index(begin = concat_42, begin_mask = new_cache_11_internal_tensor_assign_1_begin_mask_0, end = concat_43, end_mask = new_cache_11_internal_tensor_assign_1_end_mask_0, squeeze_mask = new_cache_11_internal_tensor_assign_1_squeeze_mask_0, stride = new_cache_11_internal_tensor_assign_1_stride_0, x = reshape_50)[name = tensor("slice_by_index_10")]; + tensor reshape_51_shape_0 = const()[name = tensor("reshape_51_shape_0"), val = tensor([-1])]; + tensor reshape_51 = reshape(shape = reshape_51_shape_0, x = slice_by_index_10)[name = tensor("reshape_51")]; + tensor reshape_52_shape_0 = const()[name = tensor("reshape_52_shape_0"), val = tensor([-1])]; + tensor reshape_52_cast_fp16 = reshape(shape = reshape_52_shape_0, x = var_2497_cast_fp16)[name = tensor("reshape_52_cast_fp16")]; + tensor reshape_53_shape_0 = const()[name = tensor("reshape_53_shape_0"), val = tensor([-1])]; + tensor reshape_53_cast_fp16 = reshape(shape = reshape_53_shape_0, x = cache5_to_fp16)[name = tensor("reshape_53_cast_fp16")]; + tensor scatter_10_mode_0 = const()[name = tensor("scatter_10_mode_0"), val = tensor("update")]; + tensor scatter_10_axis_0 = const()[name = tensor("scatter_10_axis_0"), val = tensor(0)]; + tensor scatter_10_validate_indices_0 = const()[name = tensor("scatter_10_validate_indices_0"), val = tensor(false)]; + tensor scatter_10_cast_fp16 = scatter(axis = scatter_10_axis_0, data = reshape_53_cast_fp16, indices = reshape_51, mode = scatter_10_mode_0, updates = reshape_52_cast_fp16, validate_indices = scatter_10_validate_indices_0)[name = tensor("scatter_10_cast_fp16")]; + tensor reshape_54_cast_fp16 = reshape(shape = shape_56, x = scatter_10_cast_fp16)[name = tensor("reshape_54_cast_fp16")]; + tensor var_2505_begin_0 = const()[name = tensor("op_2505_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor var_2505_end_0 = const()[name = tensor("op_2505_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor var_2505_end_mask_0 = const()[name = tensor("op_2505_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_2505_squeeze_mask_0 = const()[name = tensor("op_2505_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor var_2505_cast_fp16 = slice_by_index(begin = var_2505_begin_0, end = var_2505_end_0, end_mask = var_2505_end_mask_0, squeeze_mask = var_2505_squeeze_mask_0, x = reshape_54_cast_fp16)[name = tensor("op_2505_cast_fp16")]; + tensor var_2507_axis_0 = const()[name = tensor("op_2507_axis_0"), val = tensor(1)]; + tensor var_2507_mode_0 = const()[name = tensor("op_2507_mode_0"), val = tensor("update")]; + tensor var_2507_validate_indices_0 = const()[name = tensor("op_2507_validate_indices_0"), val = tensor(false)]; + tensor var_2507_cast_fp16 = scatter_along_axis(axis = var_2507_axis_0, data = var_2505_cast_fp16, indices = write_indices_11, mode = var_2507_mode_0, updates = v_11_cast_fp16, validate_indices = var_2507_validate_indices_0)[name = tensor("op_2507_cast_fp16")]; + tensor concat_44 = const()[name = tensor("concat_44"), val = tensor([1, 0, 0, 0, 0])]; + tensor concat_45 = const()[name = tensor("concat_45"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_11_internal_tensor_assign_2_stride_0 = const()[name = tensor("new_cache_11_internal_tensor_assign_2_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_11_internal_tensor_assign_2_begin_mask_0 = const()[name = tensor("new_cache_11_internal_tensor_assign_2_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_11_internal_tensor_assign_2_end_mask_0 = const()[name = tensor("new_cache_11_internal_tensor_assign_2_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_11_internal_tensor_assign_2_squeeze_mask_0 = const()[name = tensor("new_cache_11_internal_tensor_assign_2_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_57 = const()[name = tensor("shape_57"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_11 = const()[name = tensor("reduce_prod_11"), val = tensor(1048576)]; + tensor range_1d_11_start_0 = const()[name = tensor("range_1d_11_start_0"), val = tensor(0)]; + tensor range_1d_11_step_0 = const()[name = tensor("range_1d_11_step_0"), val = tensor(1)]; + tensor range_1d_11 = range_1d(end = reduce_prod_11, start = range_1d_11_start_0, step = range_1d_11_step_0)[name = tensor("range_1d_11")]; + tensor reshape_55 = reshape(shape = shape_57, x = range_1d_11)[name = tensor("reshape_55")]; + tensor slice_by_index_11 = slice_by_index(begin = concat_44, begin_mask = new_cache_11_internal_tensor_assign_2_begin_mask_0, end = concat_45, end_mask = new_cache_11_internal_tensor_assign_2_end_mask_0, squeeze_mask = new_cache_11_internal_tensor_assign_2_squeeze_mask_0, stride = new_cache_11_internal_tensor_assign_2_stride_0, x = reshape_55)[name = tensor("slice_by_index_11")]; + tensor reshape_56_shape_0 = const()[name = tensor("reshape_56_shape_0"), val = tensor([-1])]; + tensor reshape_56 = reshape(shape = reshape_56_shape_0, x = slice_by_index_11)[name = tensor("reshape_56")]; + tensor reshape_57_shape_0 = const()[name = tensor("reshape_57_shape_0"), val = tensor([-1])]; + tensor reshape_57_cast_fp16 = reshape(shape = reshape_57_shape_0, x = var_2507_cast_fp16)[name = tensor("reshape_57_cast_fp16")]; + tensor reshape_58_shape_0 = const()[name = tensor("reshape_58_shape_0"), val = tensor([-1])]; + tensor reshape_58_cast_fp16 = reshape(shape = reshape_58_shape_0, x = reshape_54_cast_fp16)[name = tensor("reshape_58_cast_fp16")]; + tensor scatter_11_mode_0 = const()[name = tensor("scatter_11_mode_0"), val = tensor("update")]; + tensor scatter_11_axis_0 = const()[name = tensor("scatter_11_axis_0"), val = tensor(0)]; + tensor scatter_11_validate_indices_0 = const()[name = tensor("scatter_11_validate_indices_0"), val = tensor(false)]; + tensor scatter_11_cast_fp16 = scatter(axis = scatter_11_axis_0, data = reshape_58_cast_fp16, indices = reshape_56, mode = scatter_11_mode_0, updates = reshape_57_cast_fp16, validate_indices = scatter_11_validate_indices_0)[name = tensor("scatter_11_cast_fp16")]; + tensor reshape_59_cast_fp16 = reshape(shape = shape_57, x = scatter_11_cast_fp16)[name = tensor("reshape_59_cast_fp16")]; + tensor reshape_59_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("reshape_59_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor keys_31_begin_0 = const()[name = tensor("keys_31_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor keys_31_end_0 = const()[name = tensor("keys_31_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor keys_31_end_mask_0 = const()[name = tensor("keys_31_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor keys_31_squeeze_mask_0 = const()[name = tensor("keys_31_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor keys_31_cast_fp16 = slice_by_index(begin = keys_31_begin_0, end = keys_31_end_0, end_mask = keys_31_end_mask_0, squeeze_mask = keys_31_squeeze_mask_0, x = reshape_59_cast_fp16)[name = tensor("keys_31_cast_fp16")]; + tensor values_31_begin_0 = const()[name = tensor("values_31_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor values_31_end_0 = const()[name = tensor("values_31_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor values_31_end_mask_0 = const()[name = tensor("values_31_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor values_31_squeeze_mask_0 = const()[name = tensor("values_31_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor values_31_cast_fp16 = slice_by_index(begin = values_31_begin_0, end = values_31_end_0, end_mask = values_31_end_mask_0, squeeze_mask = values_31_squeeze_mask_0, x = reshape_59_cast_fp16)[name = tensor("values_31_cast_fp16")]; + tensor var_2519_cast_fp16 = not_equal(x = keys_31_cast_fp16, y = keys_31_cast_fp16)[name = tensor("op_2519_cast_fp16")]; + tensor keys_33_cast_fp16 = select(a = var_470_to_fp16, b = keys_31_cast_fp16, cond = var_2519_cast_fp16)[name = tensor("keys_33_cast_fp16")]; + tensor var_2527_cast_fp16 = not_equal(x = values_31_cast_fp16, y = values_31_cast_fp16)[name = tensor("op_2527_cast_fp16")]; + tensor values_33_cast_fp16 = select(a = var_470_to_fp16, b = values_31_cast_fp16, cond = var_2527_cast_fp16)[name = tensor("values_33_cast_fp16")]; + tensor var_2551 = const()[name = tensor("op_2551"), val = tensor([0, 2, 1, 3])]; + tensor var_2567 = const()[name = tensor("op_2567"), val = tensor([1, 1, 1])]; + tensor var_2568_cast_fp16 = reshape(shape = var_2567, x = position5_to_fp16)[name = tensor("op_2568_cast_fp16")]; + tensor q_positions_11_cast_fp16 = add(x = var_2568_cast_fp16, y = q_offsets_1_promoted_to_fp16)[name = tensor("q_positions_11_cast_fp16")]; + tensor valid_end_11_cast_fp16 = add(x = var_2568_cast_fp16, y = var_536_cast_fp16)[name = tensor("valid_end_11_cast_fp16")]; + tensor valid_mask_11_cast_fp16 = less(x = k_positions_1_promoted_to_fp16, y = valid_end_11_cast_fp16)[name = tensor("valid_mask_11_cast_fp16")]; + tensor causal_mask_11_cast_fp16 = less_equal(x = k_positions_1_promoted_to_fp16, y = q_positions_11_cast_fp16)[name = tensor("causal_mask_11_cast_fp16")]; + tensor var_2596 = logical_and(x = valid_mask_11_cast_fp16, y = causal_mask_11_cast_fp16)[name = tensor("op_2596")]; + tensor attn_mask_11_axes_0 = const()[name = tensor("attn_mask_11_axes_0"), val = tensor([1])]; + tensor attn_mask_11 = expand_dims(axes = attn_mask_11_axes_0, x = var_2596)[name = tensor("attn_mask_11")]; + tensor var_2610_transpose_x_0 = const()[name = tensor("op_2610_transpose_x_0"), val = tensor(false)]; + tensor var_2610_transpose_y_0 = const()[name = tensor("op_2610_transpose_y_0"), val = tensor(false)]; + tensor transpose_79_perm_0 = const()[name = tensor("transpose_79_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_80_perm_0 = const()[name = tensor("transpose_80_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_80 = transpose(perm = transpose_80_perm_0, x = keys_33_cast_fp16)[name = tensor("transpose_184")]; + tensor transpose_79 = transpose(perm = transpose_79_perm_0, x = q_33_cast_fp16)[name = tensor("transpose_185")]; + tensor var_2610_cast_fp16 = matmul(transpose_x = var_2610_transpose_x_0, transpose_y = var_2610_transpose_y_0, x = transpose_79, y = transpose_80)[name = tensor("op_2610_cast_fp16")]; + tensor var_2604_to_fp16 = const()[name = tensor("op_2604_to_fp16"), val = tensor([0x1p-3])]; + tensor attn_weights_31_cast_fp16 = mul(x = var_2610_cast_fp16, y = var_2604_to_fp16)[name = tensor("attn_weights_31_cast_fp16")]; + tensor var_2612 = logical_not(x = attn_mask_11)[name = tensor("op_2612")]; + tensor var_2613_to_fp16 = const()[name = tensor("op_2613_to_fp16"), val = tensor(-inf)]; + tensor attn_weights_33_cast_fp16 = select(a = var_2613_to_fp16, b = attn_weights_31_cast_fp16, cond = var_2612)[name = tensor("attn_weights_33_cast_fp16")]; + tensor var_2615 = const()[name = tensor("op_2615"), val = tensor(-1)]; + tensor attn_weights_35_cast_fp16 = softmax(axis = var_2615, x = attn_weights_33_cast_fp16)[name = tensor("attn_weights_35_cast_fp16")]; + tensor attn_output_11_transpose_x_0 = const()[name = tensor("attn_output_11_transpose_x_0"), val = tensor(false)]; + tensor attn_output_11_transpose_y_0 = const()[name = tensor("attn_output_11_transpose_y_0"), val = tensor(false)]; + tensor values_35_cast_fp16 = transpose(perm = var_2551, x = values_33_cast_fp16)[name = tensor("transpose_186")]; + tensor attn_output_11_cast_fp16 = matmul(transpose_x = attn_output_11_transpose_x_0, transpose_y = attn_output_11_transpose_y_0, x = attn_weights_35_cast_fp16, y = values_35_cast_fp16)[name = tensor("attn_output_11_cast_fp16")]; + tensor var_2623 = const()[name = tensor("op_2623"), val = tensor([0, 2, 1, 3])]; + tensor var_2628 = const()[name = tensor("op_2628"), val = tensor([1, 256, 1024])]; + tensor var_2624_cast_fp16 = transpose(perm = var_2623, x = attn_output_11_cast_fp16)[name = tensor("transpose_183")]; + tensor input_65_cast_fp16 = reshape(shape = var_2628, x = var_2624_cast_fp16)[name = tensor("input_65_cast_fp16")]; + tensor attn5_out_proj_weight_to_fp16 = const()[name = tensor("attn5_out_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(133237248)))]; + tensor linear_21_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = attn5_out_proj_weight_to_fp16, x = input_65_cast_fp16)[name = tensor("linear_21_cast_fp16")]; + tensor var_2637_cast_fp16 = add(x = position5_to_fp16, y = valid_len_to_fp16)[name = tensor("op_2637_cast_fp16")]; + tensor var_2637_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_2637_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor input_67_cast_fp16 = add(x = input_61_cast_fp16, y = linear_21_cast_fp16)[name = tensor("input_67_cast_fp16")]; + tensor input_69_axes_0 = const()[name = tensor("input_69_axes_0"), val = tensor([-1])]; + tensor norm5_2_weight_to_fp16 = const()[name = tensor("norm5_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(135334464)))]; + tensor norm5_2_bias_to_fp16 = const()[name = tensor("norm5_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(135336576)))]; + tensor var_2641_to_fp16 = const()[name = tensor("op_2641_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_69_cast_fp16 = layer_norm(axes = input_69_axes_0, beta = norm5_2_bias_to_fp16, epsilon = var_2641_to_fp16, gamma = norm5_2_weight_to_fp16, x = input_67_cast_fp16)[name = tensor("input_69_cast_fp16")]; + tensor linear5_1_weight_to_fp16 = const()[name = tensor("linear5_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(135338688)))]; + tensor linear_22_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = linear5_1_weight_to_fp16, x = input_69_cast_fp16)[name = tensor("linear_22_cast_fp16")]; + tensor input_71_mode_0 = const()[name = tensor("input_71_mode_0"), val = tensor("EXACT")]; + tensor input_71_cast_fp16 = gelu(mode = input_71_mode_0, x = linear_22_cast_fp16)[name = tensor("input_71_cast_fp16")]; + tensor linear5_2_weight_to_fp16 = const()[name = tensor("linear5_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(143727360)))]; + tensor linear_23_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = linear5_2_weight_to_fp16, x = input_71_cast_fp16)[name = tensor("linear_23_cast_fp16")]; + tensor input_73_cast_fp16 = add(x = input_67_cast_fp16, y = linear_23_cast_fp16)[name = tensor("input_73_cast_fp16")]; + tensor input_75_axes_0 = const()[name = tensor("input_75_axes_0"), val = tensor([-1])]; + tensor norm6_1_weight_to_fp16 = const()[name = tensor("norm6_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(152116032)))]; + tensor norm6_1_bias_to_fp16 = const()[name = tensor("norm6_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(152118144)))]; + tensor var_2658_to_fp16 = const()[name = tensor("op_2658_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_75_cast_fp16 = layer_norm(axes = input_75_axes_0, beta = norm6_1_bias_to_fp16, epsilon = var_2658_to_fp16, gamma = norm6_1_weight_to_fp16, x = input_73_cast_fp16)[name = tensor("input_75_cast_fp16")]; + tensor attn6_in_proj_weight_to_fp16 = const()[name = tensor("attn6_in_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(152120256)))]; + tensor linear_24_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = attn6_in_proj_weight_to_fp16, x = input_75_cast_fp16)[name = tensor("linear_24_cast_fp16")]; + tensor var_2672 = const()[name = tensor("op_2672"), val = tensor([1, 256, 3, 16, 64])]; + tensor qkv_13_cast_fp16 = reshape(shape = var_2672, x = linear_24_cast_fp16)[name = tensor("qkv_13_cast_fp16")]; + tensor q_37_begin_0 = const()[name = tensor("q_37_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor q_37_end_0 = const()[name = tensor("q_37_end_0"), val = tensor([1, 256, 1, 16, 64])]; + tensor q_37_end_mask_0 = const()[name = tensor("q_37_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor q_37_squeeze_mask_0 = const()[name = tensor("q_37_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor q_37_cast_fp16 = slice_by_index(begin = q_37_begin_0, end = q_37_end_0, end_mask = q_37_end_mask_0, squeeze_mask = q_37_squeeze_mask_0, x = qkv_13_cast_fp16)[name = tensor("q_37_cast_fp16")]; + tensor k_25_begin_0 = const()[name = tensor("k_25_begin_0"), val = tensor([0, 0, 1, 0, 0])]; + tensor k_25_end_0 = const()[name = tensor("k_25_end_0"), val = tensor([1, 256, 2, 16, 64])]; + tensor k_25_end_mask_0 = const()[name = tensor("k_25_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor k_25_squeeze_mask_0 = const()[name = tensor("k_25_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor k_25_cast_fp16 = slice_by_index(begin = k_25_begin_0, end = k_25_end_0, end_mask = k_25_end_mask_0, squeeze_mask = k_25_squeeze_mask_0, x = qkv_13_cast_fp16)[name = tensor("k_25_cast_fp16")]; + tensor v_13_begin_0 = const()[name = tensor("v_13_begin_0"), val = tensor([0, 0, 2, 0, 0])]; + tensor v_13_end_0 = const()[name = tensor("v_13_end_0"), val = tensor([1, 256, 3, 16, 64])]; + tensor v_13_end_mask_0 = const()[name = tensor("v_13_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor v_13_squeeze_mask_0 = const()[name = tensor("v_13_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor v_13_cast_fp16 = slice_by_index(begin = v_13_begin_0, end = v_13_end_0, end_mask = v_13_end_mask_0, squeeze_mask = v_13_squeeze_mask_0, x = qkv_13_cast_fp16)[name = tensor("v_13_cast_fp16")]; + tensor var_2730 = const()[name = tensor("op_2730"), val = tensor([1, 1])]; + tensor position6_to_fp16_dtype_0 = const()[name = tensor("position6_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor position6_to_fp16 = cast(dtype = position6_to_fp16_dtype_0, x = position6)[name = tensor("cast_631")]; + tensor var_2731_cast_fp16 = reshape(shape = var_2730, x = position6_to_fp16)[name = tensor("op_2731_cast_fp16")]; + tensor ts_39_cast_fp16 = add(x = ts_1_promoted_to_fp16, y = var_2731_cast_fp16)[name = tensor("ts_39_cast_fp16")]; + tensor var_2738 = const()[name = tensor("op_2738"), val = tensor([1, 256, 1, 1])]; + tensor ts_41_cast_fp16 = reshape(shape = var_2738, x = ts_39_cast_fp16)[name = tensor("ts_41_cast_fp16")]; + tensor var_2745 = const()[name = tensor("op_2745"), val = tensor([1, 256, 16, 32, 2])]; + tensor q_complex_13_cast_fp16 = reshape(shape = var_2745, x = q_37_cast_fp16)[name = tensor("q_complex_13_cast_fp16")]; + tensor var_2752 = const()[name = tensor("op_2752"), val = tensor([1, 256, 16, 32, 2])]; + tensor k_complex_13_cast_fp16 = reshape(shape = var_2752, x = k_25_cast_fp16)[name = tensor("k_complex_13_cast_fp16")]; + tensor var_2756_begin_0 = const()[name = tensor("op_2756_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_2756_end_0 = const()[name = tensor("op_2756_end_0"), val = tensor([1, 256, 16, 32, 1])]; + tensor var_2756_end_mask_0 = const()[name = tensor("op_2756_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_2756_squeeze_mask_0 = const()[name = tensor("op_2756_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_2756_cast_fp16 = slice_by_index(begin = var_2756_begin_0, end = var_2756_end_0, end_mask = var_2756_end_mask_0, squeeze_mask = var_2756_squeeze_mask_0, x = q_complex_13_cast_fp16)[name = tensor("op_2756_cast_fp16")]; + tensor var_2764_begin_0 = const()[name = tensor("op_2764_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_2764_end_0 = const()[name = tensor("op_2764_end_0"), val = tensor([1, 256, 16, 32, 2])]; + tensor var_2764_end_mask_0 = const()[name = tensor("op_2764_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_2764_squeeze_mask_0 = const()[name = tensor("op_2764_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_2764_cast_fp16 = slice_by_index(begin = var_2764_begin_0, end = var_2764_end_0, end_mask = var_2764_end_mask_0, squeeze_mask = var_2764_squeeze_mask_0, x = q_complex_13_cast_fp16)[name = tensor("op_2764_cast_fp16")]; + tensor var_2772_begin_0 = const()[name = tensor("op_2772_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_2772_end_0 = const()[name = tensor("op_2772_end_0"), val = tensor([1, 256, 16, 32, 1])]; + tensor var_2772_end_mask_0 = const()[name = tensor("op_2772_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_2772_squeeze_mask_0 = const()[name = tensor("op_2772_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_2772_cast_fp16 = slice_by_index(begin = var_2772_begin_0, end = var_2772_end_0, end_mask = var_2772_end_mask_0, squeeze_mask = var_2772_squeeze_mask_0, x = k_complex_13_cast_fp16)[name = tensor("op_2772_cast_fp16")]; + tensor var_2780_begin_0 = const()[name = tensor("op_2780_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_2780_end_0 = const()[name = tensor("op_2780_end_0"), val = tensor([1, 256, 16, 32, 2])]; + tensor var_2780_end_mask_0 = const()[name = tensor("op_2780_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_2780_squeeze_mask_0 = const()[name = tensor("op_2780_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_2780_cast_fp16 = slice_by_index(begin = var_2780_begin_0, end = var_2780_end_0, end_mask = var_2780_end_mask_0, squeeze_mask = var_2780_squeeze_mask_0, x = k_complex_13_cast_fp16)[name = tensor("op_2780_cast_fp16")]; + tensor freqs_13_to_fp16 = const()[name = tensor("freqs_13_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(158411776)))]; + tensor var_2786_cast_fp16 = mul(x = freqs_13_to_fp16, y = ts_41_cast_fp16)[name = tensor("op_2786_cast_fp16")]; + tensor rotr_13_cast_fp16 = cos(x = var_2786_cast_fp16)[name = tensor("rotr_13_cast_fp16")]; + tensor roti_13_cast_fp16 = sin(x = var_2786_cast_fp16)[name = tensor("roti_13_cast_fp16")]; + tensor var_2790_cast_fp16 = mul(x = var_2756_cast_fp16, y = rotr_13_cast_fp16)[name = tensor("op_2790_cast_fp16")]; + tensor var_2791_cast_fp16 = mul(x = var_2764_cast_fp16, y = roti_13_cast_fp16)[name = tensor("op_2791_cast_fp16")]; + tensor qor_25_cast_fp16 = sub(x = var_2790_cast_fp16, y = var_2791_cast_fp16)[name = tensor("qor_25_cast_fp16")]; + tensor var_2794_cast_fp16 = mul(x = var_2756_cast_fp16, y = roti_13_cast_fp16)[name = tensor("op_2794_cast_fp16")]; + tensor var_2795_cast_fp16 = mul(x = var_2764_cast_fp16, y = rotr_13_cast_fp16)[name = tensor("op_2795_cast_fp16")]; + tensor qoi_25_cast_fp16 = add(x = var_2794_cast_fp16, y = var_2795_cast_fp16)[name = tensor("qoi_25_cast_fp16")]; + tensor var_2798_cast_fp16 = mul(x = var_2772_cast_fp16, y = rotr_13_cast_fp16)[name = tensor("op_2798_cast_fp16")]; + tensor var_2799_cast_fp16 = mul(x = var_2780_cast_fp16, y = roti_13_cast_fp16)[name = tensor("op_2799_cast_fp16")]; + tensor kor_25_cast_fp16 = sub(x = var_2798_cast_fp16, y = var_2799_cast_fp16)[name = tensor("kor_25_cast_fp16")]; + tensor var_2802_cast_fp16 = mul(x = var_2772_cast_fp16, y = roti_13_cast_fp16)[name = tensor("op_2802_cast_fp16")]; + tensor var_2803_cast_fp16 = mul(x = var_2780_cast_fp16, y = rotr_13_cast_fp16)[name = tensor("op_2803_cast_fp16")]; + tensor koi_25_cast_fp16 = add(x = var_2802_cast_fp16, y = var_2803_cast_fp16)[name = tensor("koi_25_cast_fp16")]; + tensor qo_13_axis_0 = const()[name = tensor("qo_13_axis_0"), val = tensor(-1)]; + tensor qo_13_cast_fp16 = stack(axis = qo_13_axis_0, values = (qor_25_cast_fp16, qoi_25_cast_fp16))[name = tensor("qo_13_cast_fp16")]; + tensor ko_13_axis_0 = const()[name = tensor("ko_13_axis_0"), val = tensor(-1)]; + tensor ko_13_cast_fp16 = stack(axis = ko_13_axis_0, values = (kor_25_cast_fp16, koi_25_cast_fp16))[name = tensor("ko_13_cast_fp16")]; + tensor var_2836 = const()[name = tensor("op_2836"), val = tensor([1, 256, 16, 64])]; + tensor q_39_cast_fp16 = reshape(shape = var_2836, x = qo_13_cast_fp16)[name = tensor("q_39_cast_fp16")]; + tensor var_2842 = const()[name = tensor("op_2842"), val = tensor([1, 256, 16, 64])]; + tensor k_27_cast_fp16 = reshape(shape = var_2842, x = ko_13_cast_fp16)[name = tensor("k_27_cast_fp16")]; + tensor write_indices_float_37_cast_fp16 = add(x = var_2731_cast_fp16, y = var_393_promoted_to_fp16)[name = tensor("write_indices_float_37_cast_fp16")]; + tensor _inversed_2863_y_0_to_fp16 = const()[name = tensor("_inversed_2863_y_0_to_fp16"), val = tensor(0x1p-9)]; + tensor _inversed_2863_cast_fp16 = mul(x = write_indices_float_37_cast_fp16, y = _inversed_2863_y_0_to_fp16)[name = tensor("_inversed_2863_cast_fp16")]; + tensor var_2864_cast_fp16 = floor(x = _inversed_2863_cast_fp16)[name = tensor("op_2864_cast_fp16")]; + tensor var_2865_to_fp16 = const()[name = tensor("op_2865_to_fp16"), val = tensor(0x1p+9)]; + tensor var_2866_cast_fp16 = mul(x = var_2864_cast_fp16, y = var_2865_to_fp16)[name = tensor("op_2866_cast_fp16")]; + tensor write_indices_float_39_cast_fp16 = sub(x = write_indices_float_37_cast_fp16, y = var_2866_cast_fp16)[name = tensor("write_indices_float_39_cast_fp16")]; + tensor var_2884_value_0_to_fp16 = const()[name = tensor("op_2884_value_0_to_fp16"), val = tensor(0x1.ffp+8)]; + tensor var_2884_cast_fp16 = fill_like(ref_tensor = write_indices_float_39_cast_fp16, value = var_2884_value_0_to_fp16)[name = tensor("op_2884_cast_fp16")]; + tensor write_indices_float_41_cast_fp16 = select(a = write_indices_float_39_cast_fp16, b = var_2884_cast_fp16, cond = token_valid_1_cast_fp16)[name = tensor("write_indices_float_41_cast_fp16")]; + tensor var_2890_dtype_0 = const()[name = tensor("op_2890_dtype_0"), val = tensor("int32")]; + tensor var_2895 = const()[name = tensor("op_2895"), val = tensor([1, 256, 1, 1])]; + tensor write_indices_float_41_cast_fp16_to_int32 = cast(dtype = var_2890_dtype_0, x = write_indices_float_41_cast_fp16)[name = tensor("cast_630")]; + tensor var_2896 = reshape(shape = var_2895, x = write_indices_float_41_cast_fp16_to_int32)[name = tensor("op_2896")]; + tensor write_indices_13_reps_0 = const()[name = tensor("write_indices_13_reps_0"), val = tensor([1, 1, 16, 64])]; + tensor write_indices_13 = tile(reps = write_indices_13_reps_0, x = var_2896)[name = tensor("write_indices_13")]; + tensor var_2906_begin_0 = const()[name = tensor("op_2906_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_2906_end_0 = const()[name = tensor("op_2906_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor var_2906_end_mask_0 = const()[name = tensor("op_2906_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_2906_squeeze_mask_0 = const()[name = tensor("op_2906_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor cache6_to_fp16_dtype_0 = const()[name = tensor("cache6_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor cache6_to_fp16 = cast(dtype = cache6_to_fp16_dtype_0, x = cache6)[name = tensor("cast_629")]; + tensor var_2906_cast_fp16 = slice_by_index(begin = var_2906_begin_0, end = var_2906_end_0, end_mask = var_2906_end_mask_0, squeeze_mask = var_2906_squeeze_mask_0, x = cache6_to_fp16)[name = tensor("op_2906_cast_fp16")]; + tensor var_2908_axis_0 = const()[name = tensor("op_2908_axis_0"), val = tensor(1)]; + tensor var_2908_mode_0 = const()[name = tensor("op_2908_mode_0"), val = tensor("update")]; + tensor var_2908_validate_indices_0 = const()[name = tensor("op_2908_validate_indices_0"), val = tensor(false)]; + tensor var_2908_cast_fp16 = scatter_along_axis(axis = var_2908_axis_0, data = var_2906_cast_fp16, indices = write_indices_13, mode = var_2908_mode_0, updates = k_27_cast_fp16, validate_indices = var_2908_validate_indices_0)[name = tensor("op_2908_cast_fp16")]; + tensor concat_50 = const()[name = tensor("concat_50"), val = tensor([0, 0, 0, 0, 0])]; + tensor concat_51 = const()[name = tensor("concat_51"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_13_internal_tensor_assign_1_stride_0 = const()[name = tensor("new_cache_13_internal_tensor_assign_1_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_13_internal_tensor_assign_1_begin_mask_0 = const()[name = tensor("new_cache_13_internal_tensor_assign_1_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_13_internal_tensor_assign_1_end_mask_0 = const()[name = tensor("new_cache_13_internal_tensor_assign_1_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_13_internal_tensor_assign_1_squeeze_mask_0 = const()[name = tensor("new_cache_13_internal_tensor_assign_1_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_58 = const()[name = tensor("shape_58"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_12 = const()[name = tensor("reduce_prod_12"), val = tensor(1048576)]; + tensor range_1d_12_start_0 = const()[name = tensor("range_1d_12_start_0"), val = tensor(0)]; + tensor range_1d_12_step_0 = const()[name = tensor("range_1d_12_step_0"), val = tensor(1)]; + tensor range_1d_12 = range_1d(end = reduce_prod_12, start = range_1d_12_start_0, step = range_1d_12_step_0)[name = tensor("range_1d_12")]; + tensor reshape_60 = reshape(shape = shape_58, x = range_1d_12)[name = tensor("reshape_60")]; + tensor slice_by_index_12 = slice_by_index(begin = concat_50, begin_mask = new_cache_13_internal_tensor_assign_1_begin_mask_0, end = concat_51, end_mask = new_cache_13_internal_tensor_assign_1_end_mask_0, squeeze_mask = new_cache_13_internal_tensor_assign_1_squeeze_mask_0, stride = new_cache_13_internal_tensor_assign_1_stride_0, x = reshape_60)[name = tensor("slice_by_index_12")]; + tensor reshape_61_shape_0 = const()[name = tensor("reshape_61_shape_0"), val = tensor([-1])]; + tensor reshape_61 = reshape(shape = reshape_61_shape_0, x = slice_by_index_12)[name = tensor("reshape_61")]; + tensor reshape_62_shape_0 = const()[name = tensor("reshape_62_shape_0"), val = tensor([-1])]; + tensor reshape_62_cast_fp16 = reshape(shape = reshape_62_shape_0, x = var_2908_cast_fp16)[name = tensor("reshape_62_cast_fp16")]; + tensor reshape_63_shape_0 = const()[name = tensor("reshape_63_shape_0"), val = tensor([-1])]; + tensor reshape_63_cast_fp16 = reshape(shape = reshape_63_shape_0, x = cache6_to_fp16)[name = tensor("reshape_63_cast_fp16")]; + tensor scatter_12_mode_0 = const()[name = tensor("scatter_12_mode_0"), val = tensor("update")]; + tensor scatter_12_axis_0 = const()[name = tensor("scatter_12_axis_0"), val = tensor(0)]; + tensor scatter_12_validate_indices_0 = const()[name = tensor("scatter_12_validate_indices_0"), val = tensor(false)]; + tensor scatter_12_cast_fp16 = scatter(axis = scatter_12_axis_0, data = reshape_63_cast_fp16, indices = reshape_61, mode = scatter_12_mode_0, updates = reshape_62_cast_fp16, validate_indices = scatter_12_validate_indices_0)[name = tensor("scatter_12_cast_fp16")]; + tensor reshape_64_cast_fp16 = reshape(shape = shape_58, x = scatter_12_cast_fp16)[name = tensor("reshape_64_cast_fp16")]; + tensor var_2916_begin_0 = const()[name = tensor("op_2916_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor var_2916_end_0 = const()[name = tensor("op_2916_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor var_2916_end_mask_0 = const()[name = tensor("op_2916_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_2916_squeeze_mask_0 = const()[name = tensor("op_2916_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor var_2916_cast_fp16 = slice_by_index(begin = var_2916_begin_0, end = var_2916_end_0, end_mask = var_2916_end_mask_0, squeeze_mask = var_2916_squeeze_mask_0, x = reshape_64_cast_fp16)[name = tensor("op_2916_cast_fp16")]; + tensor var_2918_axis_0 = const()[name = tensor("op_2918_axis_0"), val = tensor(1)]; + tensor var_2918_mode_0 = const()[name = tensor("op_2918_mode_0"), val = tensor("update")]; + tensor var_2918_validate_indices_0 = const()[name = tensor("op_2918_validate_indices_0"), val = tensor(false)]; + tensor var_2918_cast_fp16 = scatter_along_axis(axis = var_2918_axis_0, data = var_2916_cast_fp16, indices = write_indices_13, mode = var_2918_mode_0, updates = v_13_cast_fp16, validate_indices = var_2918_validate_indices_0)[name = tensor("op_2918_cast_fp16")]; + tensor concat_52 = const()[name = tensor("concat_52"), val = tensor([1, 0, 0, 0, 0])]; + tensor concat_53 = const()[name = tensor("concat_53"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_13_internal_tensor_assign_2_stride_0 = const()[name = tensor("new_cache_13_internal_tensor_assign_2_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_13_internal_tensor_assign_2_begin_mask_0 = const()[name = tensor("new_cache_13_internal_tensor_assign_2_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_13_internal_tensor_assign_2_end_mask_0 = const()[name = tensor("new_cache_13_internal_tensor_assign_2_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_13_internal_tensor_assign_2_squeeze_mask_0 = const()[name = tensor("new_cache_13_internal_tensor_assign_2_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_59 = const()[name = tensor("shape_59"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_13 = const()[name = tensor("reduce_prod_13"), val = tensor(1048576)]; + tensor range_1d_13_start_0 = const()[name = tensor("range_1d_13_start_0"), val = tensor(0)]; + tensor range_1d_13_step_0 = const()[name = tensor("range_1d_13_step_0"), val = tensor(1)]; + tensor range_1d_13 = range_1d(end = reduce_prod_13, start = range_1d_13_start_0, step = range_1d_13_step_0)[name = tensor("range_1d_13")]; + tensor reshape_65 = reshape(shape = shape_59, x = range_1d_13)[name = tensor("reshape_65")]; + tensor slice_by_index_13 = slice_by_index(begin = concat_52, begin_mask = new_cache_13_internal_tensor_assign_2_begin_mask_0, end = concat_53, end_mask = new_cache_13_internal_tensor_assign_2_end_mask_0, squeeze_mask = new_cache_13_internal_tensor_assign_2_squeeze_mask_0, stride = new_cache_13_internal_tensor_assign_2_stride_0, x = reshape_65)[name = tensor("slice_by_index_13")]; + tensor reshape_66_shape_0 = const()[name = tensor("reshape_66_shape_0"), val = tensor([-1])]; + tensor reshape_66 = reshape(shape = reshape_66_shape_0, x = slice_by_index_13)[name = tensor("reshape_66")]; + tensor reshape_67_shape_0 = const()[name = tensor("reshape_67_shape_0"), val = tensor([-1])]; + tensor reshape_67_cast_fp16 = reshape(shape = reshape_67_shape_0, x = var_2918_cast_fp16)[name = tensor("reshape_67_cast_fp16")]; + tensor reshape_68_shape_0 = const()[name = tensor("reshape_68_shape_0"), val = tensor([-1])]; + tensor reshape_68_cast_fp16 = reshape(shape = reshape_68_shape_0, x = reshape_64_cast_fp16)[name = tensor("reshape_68_cast_fp16")]; + tensor scatter_13_mode_0 = const()[name = tensor("scatter_13_mode_0"), val = tensor("update")]; + tensor scatter_13_axis_0 = const()[name = tensor("scatter_13_axis_0"), val = tensor(0)]; + tensor scatter_13_validate_indices_0 = const()[name = tensor("scatter_13_validate_indices_0"), val = tensor(false)]; + tensor scatter_13_cast_fp16 = scatter(axis = scatter_13_axis_0, data = reshape_68_cast_fp16, indices = reshape_66, mode = scatter_13_mode_0, updates = reshape_67_cast_fp16, validate_indices = scatter_13_validate_indices_0)[name = tensor("scatter_13_cast_fp16")]; + tensor reshape_69_cast_fp16 = reshape(shape = shape_59, x = scatter_13_cast_fp16)[name = tensor("reshape_69_cast_fp16")]; + tensor reshape_69_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("reshape_69_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor keys_37_begin_0 = const()[name = tensor("keys_37_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor keys_37_end_0 = const()[name = tensor("keys_37_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor keys_37_end_mask_0 = const()[name = tensor("keys_37_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor keys_37_squeeze_mask_0 = const()[name = tensor("keys_37_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor keys_37_cast_fp16 = slice_by_index(begin = keys_37_begin_0, end = keys_37_end_0, end_mask = keys_37_end_mask_0, squeeze_mask = keys_37_squeeze_mask_0, x = reshape_69_cast_fp16)[name = tensor("keys_37_cast_fp16")]; + tensor values_37_begin_0 = const()[name = tensor("values_37_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor values_37_end_0 = const()[name = tensor("values_37_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor values_37_end_mask_0 = const()[name = tensor("values_37_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor values_37_squeeze_mask_0 = const()[name = tensor("values_37_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor values_37_cast_fp16 = slice_by_index(begin = values_37_begin_0, end = values_37_end_0, end_mask = values_37_end_mask_0, squeeze_mask = values_37_squeeze_mask_0, x = reshape_69_cast_fp16)[name = tensor("values_37_cast_fp16")]; + tensor var_2930_cast_fp16 = not_equal(x = keys_37_cast_fp16, y = keys_37_cast_fp16)[name = tensor("op_2930_cast_fp16")]; + tensor keys_39_cast_fp16 = select(a = var_470_to_fp16, b = keys_37_cast_fp16, cond = var_2930_cast_fp16)[name = tensor("keys_39_cast_fp16")]; + tensor var_2938_cast_fp16 = not_equal(x = values_37_cast_fp16, y = values_37_cast_fp16)[name = tensor("op_2938_cast_fp16")]; + tensor values_39_cast_fp16 = select(a = var_470_to_fp16, b = values_37_cast_fp16, cond = var_2938_cast_fp16)[name = tensor("values_39_cast_fp16")]; + tensor var_2962 = const()[name = tensor("op_2962"), val = tensor([0, 2, 1, 3])]; + tensor var_2978 = const()[name = tensor("op_2978"), val = tensor([1, 1, 1])]; + tensor var_2979_cast_fp16 = reshape(shape = var_2978, x = position6_to_fp16)[name = tensor("op_2979_cast_fp16")]; + tensor q_positions_13_cast_fp16 = add(x = var_2979_cast_fp16, y = q_offsets_1_promoted_to_fp16)[name = tensor("q_positions_13_cast_fp16")]; + tensor valid_end_13_cast_fp16 = add(x = var_2979_cast_fp16, y = var_536_cast_fp16)[name = tensor("valid_end_13_cast_fp16")]; + tensor valid_mask_13_cast_fp16 = less(x = k_positions_1_promoted_to_fp16, y = valid_end_13_cast_fp16)[name = tensor("valid_mask_13_cast_fp16")]; + tensor causal_mask_13_cast_fp16 = less_equal(x = k_positions_1_promoted_to_fp16, y = q_positions_13_cast_fp16)[name = tensor("causal_mask_13_cast_fp16")]; + tensor var_3007 = logical_and(x = valid_mask_13_cast_fp16, y = causal_mask_13_cast_fp16)[name = tensor("op_3007")]; + tensor attn_mask_13_axes_0 = const()[name = tensor("attn_mask_13_axes_0"), val = tensor([1])]; + tensor attn_mask_13 = expand_dims(axes = attn_mask_13_axes_0, x = var_3007)[name = tensor("attn_mask_13")]; + tensor var_3021_transpose_x_0 = const()[name = tensor("op_3021_transpose_x_0"), val = tensor(false)]; + tensor var_3021_transpose_y_0 = const()[name = tensor("op_3021_transpose_y_0"), val = tensor(false)]; + tensor transpose_81_perm_0 = const()[name = tensor("transpose_81_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_82_perm_0 = const()[name = tensor("transpose_82_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_82 = transpose(perm = transpose_82_perm_0, x = keys_39_cast_fp16)[name = tensor("transpose_180")]; + tensor transpose_81 = transpose(perm = transpose_81_perm_0, x = q_39_cast_fp16)[name = tensor("transpose_181")]; + tensor var_3021_cast_fp16 = matmul(transpose_x = var_3021_transpose_x_0, transpose_y = var_3021_transpose_y_0, x = transpose_81, y = transpose_82)[name = tensor("op_3021_cast_fp16")]; + tensor var_3015_to_fp16 = const()[name = tensor("op_3015_to_fp16"), val = tensor([0x1p-3])]; + tensor attn_weights_37_cast_fp16 = mul(x = var_3021_cast_fp16, y = var_3015_to_fp16)[name = tensor("attn_weights_37_cast_fp16")]; + tensor var_3023 = logical_not(x = attn_mask_13)[name = tensor("op_3023")]; + tensor var_3024_to_fp16 = const()[name = tensor("op_3024_to_fp16"), val = tensor(-inf)]; + tensor attn_weights_39_cast_fp16 = select(a = var_3024_to_fp16, b = attn_weights_37_cast_fp16, cond = var_3023)[name = tensor("attn_weights_39_cast_fp16")]; + tensor var_3026 = const()[name = tensor("op_3026"), val = tensor(-1)]; + tensor attn_weights_41_cast_fp16 = softmax(axis = var_3026, x = attn_weights_39_cast_fp16)[name = tensor("attn_weights_41_cast_fp16")]; + tensor attn_output_13_transpose_x_0 = const()[name = tensor("attn_output_13_transpose_x_0"), val = tensor(false)]; + tensor attn_output_13_transpose_y_0 = const()[name = tensor("attn_output_13_transpose_y_0"), val = tensor(false)]; + tensor values_41_cast_fp16 = transpose(perm = var_2962, x = values_39_cast_fp16)[name = tensor("transpose_182")]; + tensor attn_output_13_cast_fp16 = matmul(transpose_x = attn_output_13_transpose_x_0, transpose_y = attn_output_13_transpose_y_0, x = attn_weights_41_cast_fp16, y = values_41_cast_fp16)[name = tensor("attn_output_13_cast_fp16")]; + tensor var_3034 = const()[name = tensor("op_3034"), val = tensor([0, 2, 1, 3])]; + tensor var_3039 = const()[name = tensor("op_3039"), val = tensor([1, 256, 1024])]; + tensor var_3035_cast_fp16 = transpose(perm = var_3034, x = attn_output_13_cast_fp16)[name = tensor("transpose_179")]; + tensor input_77_cast_fp16 = reshape(shape = var_3039, x = var_3035_cast_fp16)[name = tensor("input_77_cast_fp16")]; + tensor attn6_out_proj_weight_to_fp16 = const()[name = tensor("attn6_out_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(158411904)))]; + tensor linear_25_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = attn6_out_proj_weight_to_fp16, x = input_77_cast_fp16)[name = tensor("linear_25_cast_fp16")]; + tensor var_3048_cast_fp16 = add(x = position6_to_fp16, y = valid_len_to_fp16)[name = tensor("op_3048_cast_fp16")]; + tensor var_3048_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_3048_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor input_79_cast_fp16 = add(x = input_73_cast_fp16, y = linear_25_cast_fp16)[name = tensor("input_79_cast_fp16")]; + tensor input_81_axes_0 = const()[name = tensor("input_81_axes_0"), val = tensor([-1])]; + tensor norm6_2_weight_to_fp16 = const()[name = tensor("norm6_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(160509120)))]; + tensor norm6_2_bias_to_fp16 = const()[name = tensor("norm6_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(160511232)))]; + tensor var_3052_to_fp16 = const()[name = tensor("op_3052_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_81_cast_fp16 = layer_norm(axes = input_81_axes_0, beta = norm6_2_bias_to_fp16, epsilon = var_3052_to_fp16, gamma = norm6_2_weight_to_fp16, x = input_79_cast_fp16)[name = tensor("input_81_cast_fp16")]; + tensor linear6_1_weight_to_fp16 = const()[name = tensor("linear6_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(160513344)))]; + tensor linear_26_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = linear6_1_weight_to_fp16, x = input_81_cast_fp16)[name = tensor("linear_26_cast_fp16")]; + tensor input_83_mode_0 = const()[name = tensor("input_83_mode_0"), val = tensor("EXACT")]; + tensor input_83_cast_fp16 = gelu(mode = input_83_mode_0, x = linear_26_cast_fp16)[name = tensor("input_83_cast_fp16")]; + tensor linear6_2_weight_to_fp16 = const()[name = tensor("linear6_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(168902016)))]; + tensor linear_27_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = linear6_2_weight_to_fp16, x = input_83_cast_fp16)[name = tensor("linear_27_cast_fp16")]; + tensor input_85_cast_fp16 = add(x = input_79_cast_fp16, y = linear_27_cast_fp16)[name = tensor("input_85_cast_fp16")]; + tensor input_87_axes_0 = const()[name = tensor("input_87_axes_0"), val = tensor([-1])]; + tensor norm7_1_weight_to_fp16 = const()[name = tensor("norm7_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(177290688)))]; + tensor norm7_1_bias_to_fp16 = const()[name = tensor("norm7_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(177292800)))]; + tensor var_3069_to_fp16 = const()[name = tensor("op_3069_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_87_cast_fp16 = layer_norm(axes = input_87_axes_0, beta = norm7_1_bias_to_fp16, epsilon = var_3069_to_fp16, gamma = norm7_1_weight_to_fp16, x = input_85_cast_fp16)[name = tensor("input_87_cast_fp16")]; + tensor attn7_in_proj_weight_to_fp16 = const()[name = tensor("attn7_in_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(177294912)))]; + tensor linear_28_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = attn7_in_proj_weight_to_fp16, x = input_87_cast_fp16)[name = tensor("linear_28_cast_fp16")]; + tensor var_3083 = const()[name = tensor("op_3083"), val = tensor([1, 256, 3, 16, 64])]; + tensor qkv_15_cast_fp16 = reshape(shape = var_3083, x = linear_28_cast_fp16)[name = tensor("qkv_15_cast_fp16")]; + tensor q_43_begin_0 = const()[name = tensor("q_43_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor q_43_end_0 = const()[name = tensor("q_43_end_0"), val = tensor([1, 256, 1, 16, 64])]; + tensor q_43_end_mask_0 = const()[name = tensor("q_43_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor q_43_squeeze_mask_0 = const()[name = tensor("q_43_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor q_43_cast_fp16 = slice_by_index(begin = q_43_begin_0, end = q_43_end_0, end_mask = q_43_end_mask_0, squeeze_mask = q_43_squeeze_mask_0, x = qkv_15_cast_fp16)[name = tensor("q_43_cast_fp16")]; + tensor k_29_begin_0 = const()[name = tensor("k_29_begin_0"), val = tensor([0, 0, 1, 0, 0])]; + tensor k_29_end_0 = const()[name = tensor("k_29_end_0"), val = tensor([1, 256, 2, 16, 64])]; + tensor k_29_end_mask_0 = const()[name = tensor("k_29_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor k_29_squeeze_mask_0 = const()[name = tensor("k_29_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor k_29_cast_fp16 = slice_by_index(begin = k_29_begin_0, end = k_29_end_0, end_mask = k_29_end_mask_0, squeeze_mask = k_29_squeeze_mask_0, x = qkv_15_cast_fp16)[name = tensor("k_29_cast_fp16")]; + tensor v_15_begin_0 = const()[name = tensor("v_15_begin_0"), val = tensor([0, 0, 2, 0, 0])]; + tensor v_15_end_0 = const()[name = tensor("v_15_end_0"), val = tensor([1, 256, 3, 16, 64])]; + tensor v_15_end_mask_0 = const()[name = tensor("v_15_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor v_15_squeeze_mask_0 = const()[name = tensor("v_15_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor v_15_cast_fp16 = slice_by_index(begin = v_15_begin_0, end = v_15_end_0, end_mask = v_15_end_mask_0, squeeze_mask = v_15_squeeze_mask_0, x = qkv_15_cast_fp16)[name = tensor("v_15_cast_fp16")]; + tensor var_3141 = const()[name = tensor("op_3141"), val = tensor([1, 1])]; + tensor position7_to_fp16_dtype_0 = const()[name = tensor("position7_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor position7_to_fp16 = cast(dtype = position7_to_fp16_dtype_0, x = position7)[name = tensor("cast_626")]; + tensor var_3142_cast_fp16 = reshape(shape = var_3141, x = position7_to_fp16)[name = tensor("op_3142_cast_fp16")]; + tensor ts_45_cast_fp16 = add(x = ts_1_promoted_to_fp16, y = var_3142_cast_fp16)[name = tensor("ts_45_cast_fp16")]; + tensor var_3149 = const()[name = tensor("op_3149"), val = tensor([1, 256, 1, 1])]; + tensor ts_47_cast_fp16 = reshape(shape = var_3149, x = ts_45_cast_fp16)[name = tensor("ts_47_cast_fp16")]; + tensor var_3156 = const()[name = tensor("op_3156"), val = tensor([1, 256, 16, 32, 2])]; + tensor q_complex_15_cast_fp16 = reshape(shape = var_3156, x = q_43_cast_fp16)[name = tensor("q_complex_15_cast_fp16")]; + tensor var_3163 = const()[name = tensor("op_3163"), val = tensor([1, 256, 16, 32, 2])]; + tensor k_complex_15_cast_fp16 = reshape(shape = var_3163, x = k_29_cast_fp16)[name = tensor("k_complex_15_cast_fp16")]; + tensor var_3167_begin_0 = const()[name = tensor("op_3167_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_3167_end_0 = const()[name = tensor("op_3167_end_0"), val = tensor([1, 256, 16, 32, 1])]; + tensor var_3167_end_mask_0 = const()[name = tensor("op_3167_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_3167_squeeze_mask_0 = const()[name = tensor("op_3167_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_3167_cast_fp16 = slice_by_index(begin = var_3167_begin_0, end = var_3167_end_0, end_mask = var_3167_end_mask_0, squeeze_mask = var_3167_squeeze_mask_0, x = q_complex_15_cast_fp16)[name = tensor("op_3167_cast_fp16")]; + tensor var_3175_begin_0 = const()[name = tensor("op_3175_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_3175_end_0 = const()[name = tensor("op_3175_end_0"), val = tensor([1, 256, 16, 32, 2])]; + tensor var_3175_end_mask_0 = const()[name = tensor("op_3175_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_3175_squeeze_mask_0 = const()[name = tensor("op_3175_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_3175_cast_fp16 = slice_by_index(begin = var_3175_begin_0, end = var_3175_end_0, end_mask = var_3175_end_mask_0, squeeze_mask = var_3175_squeeze_mask_0, x = q_complex_15_cast_fp16)[name = tensor("op_3175_cast_fp16")]; + tensor var_3183_begin_0 = const()[name = tensor("op_3183_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_3183_end_0 = const()[name = tensor("op_3183_end_0"), val = tensor([1, 256, 16, 32, 1])]; + tensor var_3183_end_mask_0 = const()[name = tensor("op_3183_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_3183_squeeze_mask_0 = const()[name = tensor("op_3183_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_3183_cast_fp16 = slice_by_index(begin = var_3183_begin_0, end = var_3183_end_0, end_mask = var_3183_end_mask_0, squeeze_mask = var_3183_squeeze_mask_0, x = k_complex_15_cast_fp16)[name = tensor("op_3183_cast_fp16")]; + tensor var_3191_begin_0 = const()[name = tensor("op_3191_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_3191_end_0 = const()[name = tensor("op_3191_end_0"), val = tensor([1, 256, 16, 32, 2])]; + tensor var_3191_end_mask_0 = const()[name = tensor("op_3191_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_3191_squeeze_mask_0 = const()[name = tensor("op_3191_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_3191_cast_fp16 = slice_by_index(begin = var_3191_begin_0, end = var_3191_end_0, end_mask = var_3191_end_mask_0, squeeze_mask = var_3191_squeeze_mask_0, x = k_complex_15_cast_fp16)[name = tensor("op_3191_cast_fp16")]; + tensor freqs_15_to_fp16 = const()[name = tensor("freqs_15_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(183586432)))]; + tensor var_3197_cast_fp16 = mul(x = freqs_15_to_fp16, y = ts_47_cast_fp16)[name = tensor("op_3197_cast_fp16")]; + tensor rotr_15_cast_fp16 = cos(x = var_3197_cast_fp16)[name = tensor("rotr_15_cast_fp16")]; + tensor roti_15_cast_fp16 = sin(x = var_3197_cast_fp16)[name = tensor("roti_15_cast_fp16")]; + tensor var_3201_cast_fp16 = mul(x = var_3167_cast_fp16, y = rotr_15_cast_fp16)[name = tensor("op_3201_cast_fp16")]; + tensor var_3202_cast_fp16 = mul(x = var_3175_cast_fp16, y = roti_15_cast_fp16)[name = tensor("op_3202_cast_fp16")]; + tensor qor_29_cast_fp16 = sub(x = var_3201_cast_fp16, y = var_3202_cast_fp16)[name = tensor("qor_29_cast_fp16")]; + tensor var_3205_cast_fp16 = mul(x = var_3167_cast_fp16, y = roti_15_cast_fp16)[name = tensor("op_3205_cast_fp16")]; + tensor var_3206_cast_fp16 = mul(x = var_3175_cast_fp16, y = rotr_15_cast_fp16)[name = tensor("op_3206_cast_fp16")]; + tensor qoi_29_cast_fp16 = add(x = var_3205_cast_fp16, y = var_3206_cast_fp16)[name = tensor("qoi_29_cast_fp16")]; + tensor var_3209_cast_fp16 = mul(x = var_3183_cast_fp16, y = rotr_15_cast_fp16)[name = tensor("op_3209_cast_fp16")]; + tensor var_3210_cast_fp16 = mul(x = var_3191_cast_fp16, y = roti_15_cast_fp16)[name = tensor("op_3210_cast_fp16")]; + tensor kor_29_cast_fp16 = sub(x = var_3209_cast_fp16, y = var_3210_cast_fp16)[name = tensor("kor_29_cast_fp16")]; + tensor var_3213_cast_fp16 = mul(x = var_3183_cast_fp16, y = roti_15_cast_fp16)[name = tensor("op_3213_cast_fp16")]; + tensor var_3214_cast_fp16 = mul(x = var_3191_cast_fp16, y = rotr_15_cast_fp16)[name = tensor("op_3214_cast_fp16")]; + tensor koi_29_cast_fp16 = add(x = var_3213_cast_fp16, y = var_3214_cast_fp16)[name = tensor("koi_29_cast_fp16")]; + tensor qo_15_axis_0 = const()[name = tensor("qo_15_axis_0"), val = tensor(-1)]; + tensor qo_15_cast_fp16 = stack(axis = qo_15_axis_0, values = (qor_29_cast_fp16, qoi_29_cast_fp16))[name = tensor("qo_15_cast_fp16")]; + tensor ko_15_axis_0 = const()[name = tensor("ko_15_axis_0"), val = tensor(-1)]; + tensor ko_15_cast_fp16 = stack(axis = ko_15_axis_0, values = (kor_29_cast_fp16, koi_29_cast_fp16))[name = tensor("ko_15_cast_fp16")]; + tensor var_3247 = const()[name = tensor("op_3247"), val = tensor([1, 256, 16, 64])]; + tensor q_45_cast_fp16 = reshape(shape = var_3247, x = qo_15_cast_fp16)[name = tensor("q_45_cast_fp16")]; + tensor var_3253 = const()[name = tensor("op_3253"), val = tensor([1, 256, 16, 64])]; + tensor k_31_cast_fp16 = reshape(shape = var_3253, x = ko_15_cast_fp16)[name = tensor("k_31_cast_fp16")]; + tensor write_indices_float_43_cast_fp16 = add(x = var_3142_cast_fp16, y = var_393_promoted_to_fp16)[name = tensor("write_indices_float_43_cast_fp16")]; + tensor _inversed_3274_y_0_to_fp16 = const()[name = tensor("_inversed_3274_y_0_to_fp16"), val = tensor(0x1p-9)]; + tensor _inversed_3274_cast_fp16 = mul(x = write_indices_float_43_cast_fp16, y = _inversed_3274_y_0_to_fp16)[name = tensor("_inversed_3274_cast_fp16")]; + tensor var_3275_cast_fp16 = floor(x = _inversed_3274_cast_fp16)[name = tensor("op_3275_cast_fp16")]; + tensor var_3276_to_fp16 = const()[name = tensor("op_3276_to_fp16"), val = tensor(0x1p+9)]; + tensor var_3277_cast_fp16 = mul(x = var_3275_cast_fp16, y = var_3276_to_fp16)[name = tensor("op_3277_cast_fp16")]; + tensor write_indices_float_45_cast_fp16 = sub(x = write_indices_float_43_cast_fp16, y = var_3277_cast_fp16)[name = tensor("write_indices_float_45_cast_fp16")]; + tensor var_3295_value_0_to_fp16 = const()[name = tensor("op_3295_value_0_to_fp16"), val = tensor(0x1.ffp+8)]; + tensor var_3295_cast_fp16 = fill_like(ref_tensor = write_indices_float_45_cast_fp16, value = var_3295_value_0_to_fp16)[name = tensor("op_3295_cast_fp16")]; + tensor write_indices_float_47_cast_fp16 = select(a = write_indices_float_45_cast_fp16, b = var_3295_cast_fp16, cond = token_valid_1_cast_fp16)[name = tensor("write_indices_float_47_cast_fp16")]; + tensor var_3301_dtype_0 = const()[name = tensor("op_3301_dtype_0"), val = tensor("int32")]; + tensor var_3306 = const()[name = tensor("op_3306"), val = tensor([1, 256, 1, 1])]; + tensor write_indices_float_47_cast_fp16_to_int32 = cast(dtype = var_3301_dtype_0, x = write_indices_float_47_cast_fp16)[name = tensor("cast_625")]; + tensor var_3307 = reshape(shape = var_3306, x = write_indices_float_47_cast_fp16_to_int32)[name = tensor("op_3307")]; + tensor write_indices_15_reps_0 = const()[name = tensor("write_indices_15_reps_0"), val = tensor([1, 1, 16, 64])]; + tensor write_indices_15 = tile(reps = write_indices_15_reps_0, x = var_3307)[name = tensor("write_indices_15")]; + tensor var_3317_begin_0 = const()[name = tensor("op_3317_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_3317_end_0 = const()[name = tensor("op_3317_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor var_3317_end_mask_0 = const()[name = tensor("op_3317_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_3317_squeeze_mask_0 = const()[name = tensor("op_3317_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor cache7_to_fp16_dtype_0 = const()[name = tensor("cache7_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor cache7_to_fp16 = cast(dtype = cache7_to_fp16_dtype_0, x = cache7)[name = tensor("cast_624")]; + tensor var_3317_cast_fp16 = slice_by_index(begin = var_3317_begin_0, end = var_3317_end_0, end_mask = var_3317_end_mask_0, squeeze_mask = var_3317_squeeze_mask_0, x = cache7_to_fp16)[name = tensor("op_3317_cast_fp16")]; + tensor var_3319_axis_0 = const()[name = tensor("op_3319_axis_0"), val = tensor(1)]; + tensor var_3319_mode_0 = const()[name = tensor("op_3319_mode_0"), val = tensor("update")]; + tensor var_3319_validate_indices_0 = const()[name = tensor("op_3319_validate_indices_0"), val = tensor(false)]; + tensor var_3319_cast_fp16 = scatter_along_axis(axis = var_3319_axis_0, data = var_3317_cast_fp16, indices = write_indices_15, mode = var_3319_mode_0, updates = k_31_cast_fp16, validate_indices = var_3319_validate_indices_0)[name = tensor("op_3319_cast_fp16")]; + tensor concat_58 = const()[name = tensor("concat_58"), val = tensor([0, 0, 0, 0, 0])]; + tensor concat_59 = const()[name = tensor("concat_59"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_15_internal_tensor_assign_1_stride_0 = const()[name = tensor("new_cache_15_internal_tensor_assign_1_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_15_internal_tensor_assign_1_begin_mask_0 = const()[name = tensor("new_cache_15_internal_tensor_assign_1_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_15_internal_tensor_assign_1_end_mask_0 = const()[name = tensor("new_cache_15_internal_tensor_assign_1_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_15_internal_tensor_assign_1_squeeze_mask_0 = const()[name = tensor("new_cache_15_internal_tensor_assign_1_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_60 = const()[name = tensor("shape_60"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_14 = const()[name = tensor("reduce_prod_14"), val = tensor(1048576)]; + tensor range_1d_14_start_0 = const()[name = tensor("range_1d_14_start_0"), val = tensor(0)]; + tensor range_1d_14_step_0 = const()[name = tensor("range_1d_14_step_0"), val = tensor(1)]; + tensor range_1d_14 = range_1d(end = reduce_prod_14, start = range_1d_14_start_0, step = range_1d_14_step_0)[name = tensor("range_1d_14")]; + tensor reshape_70 = reshape(shape = shape_60, x = range_1d_14)[name = tensor("reshape_70")]; + tensor slice_by_index_14 = slice_by_index(begin = concat_58, begin_mask = new_cache_15_internal_tensor_assign_1_begin_mask_0, end = concat_59, end_mask = new_cache_15_internal_tensor_assign_1_end_mask_0, squeeze_mask = new_cache_15_internal_tensor_assign_1_squeeze_mask_0, stride = new_cache_15_internal_tensor_assign_1_stride_0, x = reshape_70)[name = tensor("slice_by_index_14")]; + tensor reshape_71_shape_0 = const()[name = tensor("reshape_71_shape_0"), val = tensor([-1])]; + tensor reshape_71 = reshape(shape = reshape_71_shape_0, x = slice_by_index_14)[name = tensor("reshape_71")]; + tensor reshape_72_shape_0 = const()[name = tensor("reshape_72_shape_0"), val = tensor([-1])]; + tensor reshape_72_cast_fp16 = reshape(shape = reshape_72_shape_0, x = var_3319_cast_fp16)[name = tensor("reshape_72_cast_fp16")]; + tensor reshape_73_shape_0 = const()[name = tensor("reshape_73_shape_0"), val = tensor([-1])]; + tensor reshape_73_cast_fp16 = reshape(shape = reshape_73_shape_0, x = cache7_to_fp16)[name = tensor("reshape_73_cast_fp16")]; + tensor scatter_14_mode_0 = const()[name = tensor("scatter_14_mode_0"), val = tensor("update")]; + tensor scatter_14_axis_0 = const()[name = tensor("scatter_14_axis_0"), val = tensor(0)]; + tensor scatter_14_validate_indices_0 = const()[name = tensor("scatter_14_validate_indices_0"), val = tensor(false)]; + tensor scatter_14_cast_fp16 = scatter(axis = scatter_14_axis_0, data = reshape_73_cast_fp16, indices = reshape_71, mode = scatter_14_mode_0, updates = reshape_72_cast_fp16, validate_indices = scatter_14_validate_indices_0)[name = tensor("scatter_14_cast_fp16")]; + tensor reshape_74_cast_fp16 = reshape(shape = shape_60, x = scatter_14_cast_fp16)[name = tensor("reshape_74_cast_fp16")]; + tensor var_3327_begin_0 = const()[name = tensor("op_3327_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor var_3327_end_0 = const()[name = tensor("op_3327_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor var_3327_end_mask_0 = const()[name = tensor("op_3327_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_3327_squeeze_mask_0 = const()[name = tensor("op_3327_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor var_3327_cast_fp16 = slice_by_index(begin = var_3327_begin_0, end = var_3327_end_0, end_mask = var_3327_end_mask_0, squeeze_mask = var_3327_squeeze_mask_0, x = reshape_74_cast_fp16)[name = tensor("op_3327_cast_fp16")]; + tensor var_3329_axis_0 = const()[name = tensor("op_3329_axis_0"), val = tensor(1)]; + tensor var_3329_mode_0 = const()[name = tensor("op_3329_mode_0"), val = tensor("update")]; + tensor var_3329_validate_indices_0 = const()[name = tensor("op_3329_validate_indices_0"), val = tensor(false)]; + tensor var_3329_cast_fp16 = scatter_along_axis(axis = var_3329_axis_0, data = var_3327_cast_fp16, indices = write_indices_15, mode = var_3329_mode_0, updates = v_15_cast_fp16, validate_indices = var_3329_validate_indices_0)[name = tensor("op_3329_cast_fp16")]; + tensor concat_60 = const()[name = tensor("concat_60"), val = tensor([1, 0, 0, 0, 0])]; + tensor concat_61 = const()[name = tensor("concat_61"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_15_internal_tensor_assign_2_stride_0 = const()[name = tensor("new_cache_15_internal_tensor_assign_2_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_15_internal_tensor_assign_2_begin_mask_0 = const()[name = tensor("new_cache_15_internal_tensor_assign_2_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_15_internal_tensor_assign_2_end_mask_0 = const()[name = tensor("new_cache_15_internal_tensor_assign_2_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_15_internal_tensor_assign_2_squeeze_mask_0 = const()[name = tensor("new_cache_15_internal_tensor_assign_2_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_61 = const()[name = tensor("shape_61"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_15 = const()[name = tensor("reduce_prod_15"), val = tensor(1048576)]; + tensor range_1d_15_start_0 = const()[name = tensor("range_1d_15_start_0"), val = tensor(0)]; + tensor range_1d_15_step_0 = const()[name = tensor("range_1d_15_step_0"), val = tensor(1)]; + tensor range_1d_15 = range_1d(end = reduce_prod_15, start = range_1d_15_start_0, step = range_1d_15_step_0)[name = tensor("range_1d_15")]; + tensor reshape_75 = reshape(shape = shape_61, x = range_1d_15)[name = tensor("reshape_75")]; + tensor slice_by_index_15 = slice_by_index(begin = concat_60, begin_mask = new_cache_15_internal_tensor_assign_2_begin_mask_0, end = concat_61, end_mask = new_cache_15_internal_tensor_assign_2_end_mask_0, squeeze_mask = new_cache_15_internal_tensor_assign_2_squeeze_mask_0, stride = new_cache_15_internal_tensor_assign_2_stride_0, x = reshape_75)[name = tensor("slice_by_index_15")]; + tensor reshape_76_shape_0 = const()[name = tensor("reshape_76_shape_0"), val = tensor([-1])]; + tensor reshape_76 = reshape(shape = reshape_76_shape_0, x = slice_by_index_15)[name = tensor("reshape_76")]; + tensor reshape_77_shape_0 = const()[name = tensor("reshape_77_shape_0"), val = tensor([-1])]; + tensor reshape_77_cast_fp16 = reshape(shape = reshape_77_shape_0, x = var_3329_cast_fp16)[name = tensor("reshape_77_cast_fp16")]; + tensor reshape_78_shape_0 = const()[name = tensor("reshape_78_shape_0"), val = tensor([-1])]; + tensor reshape_78_cast_fp16 = reshape(shape = reshape_78_shape_0, x = reshape_74_cast_fp16)[name = tensor("reshape_78_cast_fp16")]; + tensor scatter_15_mode_0 = const()[name = tensor("scatter_15_mode_0"), val = tensor("update")]; + tensor scatter_15_axis_0 = const()[name = tensor("scatter_15_axis_0"), val = tensor(0)]; + tensor scatter_15_validate_indices_0 = const()[name = tensor("scatter_15_validate_indices_0"), val = tensor(false)]; + tensor scatter_15_cast_fp16 = scatter(axis = scatter_15_axis_0, data = reshape_78_cast_fp16, indices = reshape_76, mode = scatter_15_mode_0, updates = reshape_77_cast_fp16, validate_indices = scatter_15_validate_indices_0)[name = tensor("scatter_15_cast_fp16")]; + tensor reshape_79_cast_fp16 = reshape(shape = shape_61, x = scatter_15_cast_fp16)[name = tensor("reshape_79_cast_fp16")]; + tensor reshape_79_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("reshape_79_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor keys_43_begin_0 = const()[name = tensor("keys_43_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor keys_43_end_0 = const()[name = tensor("keys_43_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor keys_43_end_mask_0 = const()[name = tensor("keys_43_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor keys_43_squeeze_mask_0 = const()[name = tensor("keys_43_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor keys_43_cast_fp16 = slice_by_index(begin = keys_43_begin_0, end = keys_43_end_0, end_mask = keys_43_end_mask_0, squeeze_mask = keys_43_squeeze_mask_0, x = reshape_79_cast_fp16)[name = tensor("keys_43_cast_fp16")]; + tensor values_43_begin_0 = const()[name = tensor("values_43_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor values_43_end_0 = const()[name = tensor("values_43_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor values_43_end_mask_0 = const()[name = tensor("values_43_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor values_43_squeeze_mask_0 = const()[name = tensor("values_43_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor values_43_cast_fp16 = slice_by_index(begin = values_43_begin_0, end = values_43_end_0, end_mask = values_43_end_mask_0, squeeze_mask = values_43_squeeze_mask_0, x = reshape_79_cast_fp16)[name = tensor("values_43_cast_fp16")]; + tensor var_3341_cast_fp16 = not_equal(x = keys_43_cast_fp16, y = keys_43_cast_fp16)[name = tensor("op_3341_cast_fp16")]; + tensor keys_45_cast_fp16 = select(a = var_470_to_fp16, b = keys_43_cast_fp16, cond = var_3341_cast_fp16)[name = tensor("keys_45_cast_fp16")]; + tensor var_3349_cast_fp16 = not_equal(x = values_43_cast_fp16, y = values_43_cast_fp16)[name = tensor("op_3349_cast_fp16")]; + tensor values_45_cast_fp16 = select(a = var_470_to_fp16, b = values_43_cast_fp16, cond = var_3349_cast_fp16)[name = tensor("values_45_cast_fp16")]; + tensor var_3373 = const()[name = tensor("op_3373"), val = tensor([0, 2, 1, 3])]; + tensor var_3389 = const()[name = tensor("op_3389"), val = tensor([1, 1, 1])]; + tensor var_3390_cast_fp16 = reshape(shape = var_3389, x = position7_to_fp16)[name = tensor("op_3390_cast_fp16")]; + tensor q_positions_15_cast_fp16 = add(x = var_3390_cast_fp16, y = q_offsets_1_promoted_to_fp16)[name = tensor("q_positions_15_cast_fp16")]; + tensor valid_end_15_cast_fp16 = add(x = var_3390_cast_fp16, y = var_536_cast_fp16)[name = tensor("valid_end_15_cast_fp16")]; + tensor valid_mask_15_cast_fp16 = less(x = k_positions_1_promoted_to_fp16, y = valid_end_15_cast_fp16)[name = tensor("valid_mask_15_cast_fp16")]; + tensor causal_mask_15_cast_fp16 = less_equal(x = k_positions_1_promoted_to_fp16, y = q_positions_15_cast_fp16)[name = tensor("causal_mask_15_cast_fp16")]; + tensor var_3418 = logical_and(x = valid_mask_15_cast_fp16, y = causal_mask_15_cast_fp16)[name = tensor("op_3418")]; + tensor attn_mask_15_axes_0 = const()[name = tensor("attn_mask_15_axes_0"), val = tensor([1])]; + tensor attn_mask_15 = expand_dims(axes = attn_mask_15_axes_0, x = var_3418)[name = tensor("attn_mask_15")]; + tensor var_3432_transpose_x_0 = const()[name = tensor("op_3432_transpose_x_0"), val = tensor(false)]; + tensor var_3432_transpose_y_0 = const()[name = tensor("op_3432_transpose_y_0"), val = tensor(false)]; + tensor transpose_83_perm_0 = const()[name = tensor("transpose_83_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_84_perm_0 = const()[name = tensor("transpose_84_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_84 = transpose(perm = transpose_84_perm_0, x = keys_45_cast_fp16)[name = tensor("transpose_176")]; + tensor transpose_83 = transpose(perm = transpose_83_perm_0, x = q_45_cast_fp16)[name = tensor("transpose_177")]; + tensor var_3432_cast_fp16 = matmul(transpose_x = var_3432_transpose_x_0, transpose_y = var_3432_transpose_y_0, x = transpose_83, y = transpose_84)[name = tensor("op_3432_cast_fp16")]; + tensor var_3426_to_fp16 = const()[name = tensor("op_3426_to_fp16"), val = tensor([0x1p-3])]; + tensor attn_weights_43_cast_fp16 = mul(x = var_3432_cast_fp16, y = var_3426_to_fp16)[name = tensor("attn_weights_43_cast_fp16")]; + tensor var_3434 = logical_not(x = attn_mask_15)[name = tensor("op_3434")]; + tensor var_3435_to_fp16 = const()[name = tensor("op_3435_to_fp16"), val = tensor(-inf)]; + tensor attn_weights_45_cast_fp16 = select(a = var_3435_to_fp16, b = attn_weights_43_cast_fp16, cond = var_3434)[name = tensor("attn_weights_45_cast_fp16")]; + tensor var_3437 = const()[name = tensor("op_3437"), val = tensor(-1)]; + tensor attn_weights_47_cast_fp16 = softmax(axis = var_3437, x = attn_weights_45_cast_fp16)[name = tensor("attn_weights_47_cast_fp16")]; + tensor attn_output_15_transpose_x_0 = const()[name = tensor("attn_output_15_transpose_x_0"), val = tensor(false)]; + tensor attn_output_15_transpose_y_0 = const()[name = tensor("attn_output_15_transpose_y_0"), val = tensor(false)]; + tensor values_47_cast_fp16 = transpose(perm = var_3373, x = values_45_cast_fp16)[name = tensor("transpose_178")]; + tensor attn_output_15_cast_fp16 = matmul(transpose_x = attn_output_15_transpose_x_0, transpose_y = attn_output_15_transpose_y_0, x = attn_weights_47_cast_fp16, y = values_47_cast_fp16)[name = tensor("attn_output_15_cast_fp16")]; + tensor var_3445 = const()[name = tensor("op_3445"), val = tensor([0, 2, 1, 3])]; + tensor var_3450 = const()[name = tensor("op_3450"), val = tensor([1, 256, 1024])]; + tensor var_3446_cast_fp16 = transpose(perm = var_3445, x = attn_output_15_cast_fp16)[name = tensor("transpose_175")]; + tensor input_89_cast_fp16 = reshape(shape = var_3450, x = var_3446_cast_fp16)[name = tensor("input_89_cast_fp16")]; + tensor attn7_out_proj_weight_to_fp16 = const()[name = tensor("attn7_out_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(183586560)))]; + tensor linear_29_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = attn7_out_proj_weight_to_fp16, x = input_89_cast_fp16)[name = tensor("linear_29_cast_fp16")]; + tensor var_3459_cast_fp16 = add(x = position7_to_fp16, y = valid_len_to_fp16)[name = tensor("op_3459_cast_fp16")]; + tensor var_3459_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_3459_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor input_91_cast_fp16 = add(x = input_85_cast_fp16, y = linear_29_cast_fp16)[name = tensor("input_91_cast_fp16")]; + tensor input_93_axes_0 = const()[name = tensor("input_93_axes_0"), val = tensor([-1])]; + tensor norm7_2_weight_to_fp16 = const()[name = tensor("norm7_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(185683776)))]; + tensor norm7_2_bias_to_fp16 = const()[name = tensor("norm7_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(185685888)))]; + tensor var_3463_to_fp16 = const()[name = tensor("op_3463_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_93_cast_fp16 = layer_norm(axes = input_93_axes_0, beta = norm7_2_bias_to_fp16, epsilon = var_3463_to_fp16, gamma = norm7_2_weight_to_fp16, x = input_91_cast_fp16)[name = tensor("input_93_cast_fp16")]; + tensor linear7_1_weight_to_fp16 = const()[name = tensor("linear7_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(185688000)))]; + tensor linear_30_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = linear7_1_weight_to_fp16, x = input_93_cast_fp16)[name = tensor("linear_30_cast_fp16")]; + tensor input_95_mode_0 = const()[name = tensor("input_95_mode_0"), val = tensor("EXACT")]; + tensor input_95_cast_fp16 = gelu(mode = input_95_mode_0, x = linear_30_cast_fp16)[name = tensor("input_95_cast_fp16")]; + tensor linear7_2_weight_to_fp16 = const()[name = tensor("linear7_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(194076672)))]; + tensor linear_31_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = linear7_2_weight_to_fp16, x = input_95_cast_fp16)[name = tensor("linear_31_cast_fp16")]; + tensor input_97_cast_fp16 = add(x = input_91_cast_fp16, y = linear_31_cast_fp16)[name = tensor("input_97_cast_fp16")]; + tensor input_99_axes_0 = const()[name = tensor("input_99_axes_0"), val = tensor([-1])]; + tensor norm8_1_weight_to_fp16 = const()[name = tensor("norm8_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(202465344)))]; + tensor norm8_1_bias_to_fp16 = const()[name = tensor("norm8_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(202467456)))]; + tensor var_3480_to_fp16 = const()[name = tensor("op_3480_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_99_cast_fp16 = layer_norm(axes = input_99_axes_0, beta = norm8_1_bias_to_fp16, epsilon = var_3480_to_fp16, gamma = norm8_1_weight_to_fp16, x = input_97_cast_fp16)[name = tensor("input_99_cast_fp16")]; + tensor attn8_in_proj_weight_to_fp16 = const()[name = tensor("attn8_in_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(202469568)))]; + tensor linear_32_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = attn8_in_proj_weight_to_fp16, x = input_99_cast_fp16)[name = tensor("linear_32_cast_fp16")]; + tensor var_3494 = const()[name = tensor("op_3494"), val = tensor([1, 256, 3, 16, 64])]; + tensor qkv_17_cast_fp16 = reshape(shape = var_3494, x = linear_32_cast_fp16)[name = tensor("qkv_17_cast_fp16")]; + tensor q_49_begin_0 = const()[name = tensor("q_49_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor q_49_end_0 = const()[name = tensor("q_49_end_0"), val = tensor([1, 256, 1, 16, 64])]; + tensor q_49_end_mask_0 = const()[name = tensor("q_49_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor q_49_squeeze_mask_0 = const()[name = tensor("q_49_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor q_49_cast_fp16 = slice_by_index(begin = q_49_begin_0, end = q_49_end_0, end_mask = q_49_end_mask_0, squeeze_mask = q_49_squeeze_mask_0, x = qkv_17_cast_fp16)[name = tensor("q_49_cast_fp16")]; + tensor k_33_begin_0 = const()[name = tensor("k_33_begin_0"), val = tensor([0, 0, 1, 0, 0])]; + tensor k_33_end_0 = const()[name = tensor("k_33_end_0"), val = tensor([1, 256, 2, 16, 64])]; + tensor k_33_end_mask_0 = const()[name = tensor("k_33_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor k_33_squeeze_mask_0 = const()[name = tensor("k_33_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor k_33_cast_fp16 = slice_by_index(begin = k_33_begin_0, end = k_33_end_0, end_mask = k_33_end_mask_0, squeeze_mask = k_33_squeeze_mask_0, x = qkv_17_cast_fp16)[name = tensor("k_33_cast_fp16")]; + tensor v_17_begin_0 = const()[name = tensor("v_17_begin_0"), val = tensor([0, 0, 2, 0, 0])]; + tensor v_17_end_0 = const()[name = tensor("v_17_end_0"), val = tensor([1, 256, 3, 16, 64])]; + tensor v_17_end_mask_0 = const()[name = tensor("v_17_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor v_17_squeeze_mask_0 = const()[name = tensor("v_17_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor v_17_cast_fp16 = slice_by_index(begin = v_17_begin_0, end = v_17_end_0, end_mask = v_17_end_mask_0, squeeze_mask = v_17_squeeze_mask_0, x = qkv_17_cast_fp16)[name = tensor("v_17_cast_fp16")]; + tensor var_3552 = const()[name = tensor("op_3552"), val = tensor([1, 1])]; + tensor position8_to_fp16_dtype_0 = const()[name = tensor("position8_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor position8_to_fp16 = cast(dtype = position8_to_fp16_dtype_0, x = position8)[name = tensor("cast_621")]; + tensor var_3553_cast_fp16 = reshape(shape = var_3552, x = position8_to_fp16)[name = tensor("op_3553_cast_fp16")]; + tensor ts_51_cast_fp16 = add(x = ts_1_promoted_to_fp16, y = var_3553_cast_fp16)[name = tensor("ts_51_cast_fp16")]; + tensor var_3560 = const()[name = tensor("op_3560"), val = tensor([1, 256, 1, 1])]; + tensor ts_53_cast_fp16 = reshape(shape = var_3560, x = ts_51_cast_fp16)[name = tensor("ts_53_cast_fp16")]; + tensor var_3567 = const()[name = tensor("op_3567"), val = tensor([1, 256, 16, 32, 2])]; + tensor q_complex_17_cast_fp16 = reshape(shape = var_3567, x = q_49_cast_fp16)[name = tensor("q_complex_17_cast_fp16")]; + tensor var_3574 = const()[name = tensor("op_3574"), val = tensor([1, 256, 16, 32, 2])]; + tensor k_complex_17_cast_fp16 = reshape(shape = var_3574, x = k_33_cast_fp16)[name = tensor("k_complex_17_cast_fp16")]; + tensor var_3578_begin_0 = const()[name = tensor("op_3578_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_3578_end_0 = const()[name = tensor("op_3578_end_0"), val = tensor([1, 256, 16, 32, 1])]; + tensor var_3578_end_mask_0 = const()[name = tensor("op_3578_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_3578_squeeze_mask_0 = const()[name = tensor("op_3578_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_3578_cast_fp16 = slice_by_index(begin = var_3578_begin_0, end = var_3578_end_0, end_mask = var_3578_end_mask_0, squeeze_mask = var_3578_squeeze_mask_0, x = q_complex_17_cast_fp16)[name = tensor("op_3578_cast_fp16")]; + tensor var_3586_begin_0 = const()[name = tensor("op_3586_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_3586_end_0 = const()[name = tensor("op_3586_end_0"), val = tensor([1, 256, 16, 32, 2])]; + tensor var_3586_end_mask_0 = const()[name = tensor("op_3586_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_3586_squeeze_mask_0 = const()[name = tensor("op_3586_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_3586_cast_fp16 = slice_by_index(begin = var_3586_begin_0, end = var_3586_end_0, end_mask = var_3586_end_mask_0, squeeze_mask = var_3586_squeeze_mask_0, x = q_complex_17_cast_fp16)[name = tensor("op_3586_cast_fp16")]; + tensor var_3594_begin_0 = const()[name = tensor("op_3594_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_3594_end_0 = const()[name = tensor("op_3594_end_0"), val = tensor([1, 256, 16, 32, 1])]; + tensor var_3594_end_mask_0 = const()[name = tensor("op_3594_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_3594_squeeze_mask_0 = const()[name = tensor("op_3594_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_3594_cast_fp16 = slice_by_index(begin = var_3594_begin_0, end = var_3594_end_0, end_mask = var_3594_end_mask_0, squeeze_mask = var_3594_squeeze_mask_0, x = k_complex_17_cast_fp16)[name = tensor("op_3594_cast_fp16")]; + tensor var_3602_begin_0 = const()[name = tensor("op_3602_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_3602_end_0 = const()[name = tensor("op_3602_end_0"), val = tensor([1, 256, 16, 32, 2])]; + tensor var_3602_end_mask_0 = const()[name = tensor("op_3602_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_3602_squeeze_mask_0 = const()[name = tensor("op_3602_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_3602_cast_fp16 = slice_by_index(begin = var_3602_begin_0, end = var_3602_end_0, end_mask = var_3602_end_mask_0, squeeze_mask = var_3602_squeeze_mask_0, x = k_complex_17_cast_fp16)[name = tensor("op_3602_cast_fp16")]; + tensor freqs_17_to_fp16 = const()[name = tensor("freqs_17_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(208761088)))]; + tensor var_3608_cast_fp16 = mul(x = freqs_17_to_fp16, y = ts_53_cast_fp16)[name = tensor("op_3608_cast_fp16")]; + tensor rotr_17_cast_fp16 = cos(x = var_3608_cast_fp16)[name = tensor("rotr_17_cast_fp16")]; + tensor roti_17_cast_fp16 = sin(x = var_3608_cast_fp16)[name = tensor("roti_17_cast_fp16")]; + tensor var_3612_cast_fp16 = mul(x = var_3578_cast_fp16, y = rotr_17_cast_fp16)[name = tensor("op_3612_cast_fp16")]; + tensor var_3613_cast_fp16 = mul(x = var_3586_cast_fp16, y = roti_17_cast_fp16)[name = tensor("op_3613_cast_fp16")]; + tensor qor_33_cast_fp16 = sub(x = var_3612_cast_fp16, y = var_3613_cast_fp16)[name = tensor("qor_33_cast_fp16")]; + tensor var_3616_cast_fp16 = mul(x = var_3578_cast_fp16, y = roti_17_cast_fp16)[name = tensor("op_3616_cast_fp16")]; + tensor var_3617_cast_fp16 = mul(x = var_3586_cast_fp16, y = rotr_17_cast_fp16)[name = tensor("op_3617_cast_fp16")]; + tensor qoi_33_cast_fp16 = add(x = var_3616_cast_fp16, y = var_3617_cast_fp16)[name = tensor("qoi_33_cast_fp16")]; + tensor var_3620_cast_fp16 = mul(x = var_3594_cast_fp16, y = rotr_17_cast_fp16)[name = tensor("op_3620_cast_fp16")]; + tensor var_3621_cast_fp16 = mul(x = var_3602_cast_fp16, y = roti_17_cast_fp16)[name = tensor("op_3621_cast_fp16")]; + tensor kor_33_cast_fp16 = sub(x = var_3620_cast_fp16, y = var_3621_cast_fp16)[name = tensor("kor_33_cast_fp16")]; + tensor var_3624_cast_fp16 = mul(x = var_3594_cast_fp16, y = roti_17_cast_fp16)[name = tensor("op_3624_cast_fp16")]; + tensor var_3625_cast_fp16 = mul(x = var_3602_cast_fp16, y = rotr_17_cast_fp16)[name = tensor("op_3625_cast_fp16")]; + tensor koi_33_cast_fp16 = add(x = var_3624_cast_fp16, y = var_3625_cast_fp16)[name = tensor("koi_33_cast_fp16")]; + tensor qo_17_axis_0 = const()[name = tensor("qo_17_axis_0"), val = tensor(-1)]; + tensor qo_17_cast_fp16 = stack(axis = qo_17_axis_0, values = (qor_33_cast_fp16, qoi_33_cast_fp16))[name = tensor("qo_17_cast_fp16")]; + tensor ko_17_axis_0 = const()[name = tensor("ko_17_axis_0"), val = tensor(-1)]; + tensor ko_17_cast_fp16 = stack(axis = ko_17_axis_0, values = (kor_33_cast_fp16, koi_33_cast_fp16))[name = tensor("ko_17_cast_fp16")]; + tensor var_3658 = const()[name = tensor("op_3658"), val = tensor([1, 256, 16, 64])]; + tensor q_51_cast_fp16 = reshape(shape = var_3658, x = qo_17_cast_fp16)[name = tensor("q_51_cast_fp16")]; + tensor var_3664 = const()[name = tensor("op_3664"), val = tensor([1, 256, 16, 64])]; + tensor k_35_cast_fp16 = reshape(shape = var_3664, x = ko_17_cast_fp16)[name = tensor("k_35_cast_fp16")]; + tensor write_indices_float_49_cast_fp16 = add(x = var_3553_cast_fp16, y = var_393_promoted_to_fp16)[name = tensor("write_indices_float_49_cast_fp16")]; + tensor _inversed_3685_y_0_to_fp16 = const()[name = tensor("_inversed_3685_y_0_to_fp16"), val = tensor(0x1p-9)]; + tensor _inversed_3685_cast_fp16 = mul(x = write_indices_float_49_cast_fp16, y = _inversed_3685_y_0_to_fp16)[name = tensor("_inversed_3685_cast_fp16")]; + tensor var_3686_cast_fp16 = floor(x = _inversed_3685_cast_fp16)[name = tensor("op_3686_cast_fp16")]; + tensor var_3687_to_fp16 = const()[name = tensor("op_3687_to_fp16"), val = tensor(0x1p+9)]; + tensor var_3688_cast_fp16 = mul(x = var_3686_cast_fp16, y = var_3687_to_fp16)[name = tensor("op_3688_cast_fp16")]; + tensor write_indices_float_51_cast_fp16 = sub(x = write_indices_float_49_cast_fp16, y = var_3688_cast_fp16)[name = tensor("write_indices_float_51_cast_fp16")]; + tensor var_3706_value_0_to_fp16 = const()[name = tensor("op_3706_value_0_to_fp16"), val = tensor(0x1.ffp+8)]; + tensor var_3706_cast_fp16 = fill_like(ref_tensor = write_indices_float_51_cast_fp16, value = var_3706_value_0_to_fp16)[name = tensor("op_3706_cast_fp16")]; + tensor write_indices_float_53_cast_fp16 = select(a = write_indices_float_51_cast_fp16, b = var_3706_cast_fp16, cond = token_valid_1_cast_fp16)[name = tensor("write_indices_float_53_cast_fp16")]; + tensor var_3712_dtype_0 = const()[name = tensor("op_3712_dtype_0"), val = tensor("int32")]; + tensor var_3717 = const()[name = tensor("op_3717"), val = tensor([1, 256, 1, 1])]; + tensor write_indices_float_53_cast_fp16_to_int32 = cast(dtype = var_3712_dtype_0, x = write_indices_float_53_cast_fp16)[name = tensor("cast_620")]; + tensor var_3718 = reshape(shape = var_3717, x = write_indices_float_53_cast_fp16_to_int32)[name = tensor("op_3718")]; + tensor write_indices_17_reps_0 = const()[name = tensor("write_indices_17_reps_0"), val = tensor([1, 1, 16, 64])]; + tensor write_indices_17 = tile(reps = write_indices_17_reps_0, x = var_3718)[name = tensor("write_indices_17")]; + tensor var_3728_begin_0 = const()[name = tensor("op_3728_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_3728_end_0 = const()[name = tensor("op_3728_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor var_3728_end_mask_0 = const()[name = tensor("op_3728_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_3728_squeeze_mask_0 = const()[name = tensor("op_3728_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor cache8_to_fp16_dtype_0 = const()[name = tensor("cache8_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor cache8_to_fp16 = cast(dtype = cache8_to_fp16_dtype_0, x = cache8)[name = tensor("cast_619")]; + tensor var_3728_cast_fp16 = slice_by_index(begin = var_3728_begin_0, end = var_3728_end_0, end_mask = var_3728_end_mask_0, squeeze_mask = var_3728_squeeze_mask_0, x = cache8_to_fp16)[name = tensor("op_3728_cast_fp16")]; + tensor var_3730_axis_0 = const()[name = tensor("op_3730_axis_0"), val = tensor(1)]; + tensor var_3730_mode_0 = const()[name = tensor("op_3730_mode_0"), val = tensor("update")]; + tensor var_3730_validate_indices_0 = const()[name = tensor("op_3730_validate_indices_0"), val = tensor(false)]; + tensor var_3730_cast_fp16 = scatter_along_axis(axis = var_3730_axis_0, data = var_3728_cast_fp16, indices = write_indices_17, mode = var_3730_mode_0, updates = k_35_cast_fp16, validate_indices = var_3730_validate_indices_0)[name = tensor("op_3730_cast_fp16")]; + tensor concat_66 = const()[name = tensor("concat_66"), val = tensor([0, 0, 0, 0, 0])]; + tensor concat_67 = const()[name = tensor("concat_67"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_17_internal_tensor_assign_1_stride_0 = const()[name = tensor("new_cache_17_internal_tensor_assign_1_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_17_internal_tensor_assign_1_begin_mask_0 = const()[name = tensor("new_cache_17_internal_tensor_assign_1_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_17_internal_tensor_assign_1_end_mask_0 = const()[name = tensor("new_cache_17_internal_tensor_assign_1_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_17_internal_tensor_assign_1_squeeze_mask_0 = const()[name = tensor("new_cache_17_internal_tensor_assign_1_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_62 = const()[name = tensor("shape_62"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_16 = const()[name = tensor("reduce_prod_16"), val = tensor(1048576)]; + tensor range_1d_16_start_0 = const()[name = tensor("range_1d_16_start_0"), val = tensor(0)]; + tensor range_1d_16_step_0 = const()[name = tensor("range_1d_16_step_0"), val = tensor(1)]; + tensor range_1d_16 = range_1d(end = reduce_prod_16, start = range_1d_16_start_0, step = range_1d_16_step_0)[name = tensor("range_1d_16")]; + tensor reshape_80 = reshape(shape = shape_62, x = range_1d_16)[name = tensor("reshape_80")]; + tensor slice_by_index_16 = slice_by_index(begin = concat_66, begin_mask = new_cache_17_internal_tensor_assign_1_begin_mask_0, end = concat_67, end_mask = new_cache_17_internal_tensor_assign_1_end_mask_0, squeeze_mask = new_cache_17_internal_tensor_assign_1_squeeze_mask_0, stride = new_cache_17_internal_tensor_assign_1_stride_0, x = reshape_80)[name = tensor("slice_by_index_16")]; + tensor reshape_81_shape_0 = const()[name = tensor("reshape_81_shape_0"), val = tensor([-1])]; + tensor reshape_81 = reshape(shape = reshape_81_shape_0, x = slice_by_index_16)[name = tensor("reshape_81")]; + tensor reshape_82_shape_0 = const()[name = tensor("reshape_82_shape_0"), val = tensor([-1])]; + tensor reshape_82_cast_fp16 = reshape(shape = reshape_82_shape_0, x = var_3730_cast_fp16)[name = tensor("reshape_82_cast_fp16")]; + tensor reshape_83_shape_0 = const()[name = tensor("reshape_83_shape_0"), val = tensor([-1])]; + tensor reshape_83_cast_fp16 = reshape(shape = reshape_83_shape_0, x = cache8_to_fp16)[name = tensor("reshape_83_cast_fp16")]; + tensor scatter_16_mode_0 = const()[name = tensor("scatter_16_mode_0"), val = tensor("update")]; + tensor scatter_16_axis_0 = const()[name = tensor("scatter_16_axis_0"), val = tensor(0)]; + tensor scatter_16_validate_indices_0 = const()[name = tensor("scatter_16_validate_indices_0"), val = tensor(false)]; + tensor scatter_16_cast_fp16 = scatter(axis = scatter_16_axis_0, data = reshape_83_cast_fp16, indices = reshape_81, mode = scatter_16_mode_0, updates = reshape_82_cast_fp16, validate_indices = scatter_16_validate_indices_0)[name = tensor("scatter_16_cast_fp16")]; + tensor reshape_84_cast_fp16 = reshape(shape = shape_62, x = scatter_16_cast_fp16)[name = tensor("reshape_84_cast_fp16")]; + tensor var_3738_begin_0 = const()[name = tensor("op_3738_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor var_3738_end_0 = const()[name = tensor("op_3738_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor var_3738_end_mask_0 = const()[name = tensor("op_3738_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_3738_squeeze_mask_0 = const()[name = tensor("op_3738_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor var_3738_cast_fp16 = slice_by_index(begin = var_3738_begin_0, end = var_3738_end_0, end_mask = var_3738_end_mask_0, squeeze_mask = var_3738_squeeze_mask_0, x = reshape_84_cast_fp16)[name = tensor("op_3738_cast_fp16")]; + tensor var_3740_axis_0 = const()[name = tensor("op_3740_axis_0"), val = tensor(1)]; + tensor var_3740_mode_0 = const()[name = tensor("op_3740_mode_0"), val = tensor("update")]; + tensor var_3740_validate_indices_0 = const()[name = tensor("op_3740_validate_indices_0"), val = tensor(false)]; + tensor var_3740_cast_fp16 = scatter_along_axis(axis = var_3740_axis_0, data = var_3738_cast_fp16, indices = write_indices_17, mode = var_3740_mode_0, updates = v_17_cast_fp16, validate_indices = var_3740_validate_indices_0)[name = tensor("op_3740_cast_fp16")]; + tensor concat_68 = const()[name = tensor("concat_68"), val = tensor([1, 0, 0, 0, 0])]; + tensor concat_69 = const()[name = tensor("concat_69"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_17_internal_tensor_assign_2_stride_0 = const()[name = tensor("new_cache_17_internal_tensor_assign_2_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_17_internal_tensor_assign_2_begin_mask_0 = const()[name = tensor("new_cache_17_internal_tensor_assign_2_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_17_internal_tensor_assign_2_end_mask_0 = const()[name = tensor("new_cache_17_internal_tensor_assign_2_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_17_internal_tensor_assign_2_squeeze_mask_0 = const()[name = tensor("new_cache_17_internal_tensor_assign_2_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_63 = const()[name = tensor("shape_63"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_17 = const()[name = tensor("reduce_prod_17"), val = tensor(1048576)]; + tensor range_1d_17_start_0 = const()[name = tensor("range_1d_17_start_0"), val = tensor(0)]; + tensor range_1d_17_step_0 = const()[name = tensor("range_1d_17_step_0"), val = tensor(1)]; + tensor range_1d_17 = range_1d(end = reduce_prod_17, start = range_1d_17_start_0, step = range_1d_17_step_0)[name = tensor("range_1d_17")]; + tensor reshape_85 = reshape(shape = shape_63, x = range_1d_17)[name = tensor("reshape_85")]; + tensor slice_by_index_17 = slice_by_index(begin = concat_68, begin_mask = new_cache_17_internal_tensor_assign_2_begin_mask_0, end = concat_69, end_mask = new_cache_17_internal_tensor_assign_2_end_mask_0, squeeze_mask = new_cache_17_internal_tensor_assign_2_squeeze_mask_0, stride = new_cache_17_internal_tensor_assign_2_stride_0, x = reshape_85)[name = tensor("slice_by_index_17")]; + tensor reshape_86_shape_0 = const()[name = tensor("reshape_86_shape_0"), val = tensor([-1])]; + tensor reshape_86 = reshape(shape = reshape_86_shape_0, x = slice_by_index_17)[name = tensor("reshape_86")]; + tensor reshape_87_shape_0 = const()[name = tensor("reshape_87_shape_0"), val = tensor([-1])]; + tensor reshape_87_cast_fp16 = reshape(shape = reshape_87_shape_0, x = var_3740_cast_fp16)[name = tensor("reshape_87_cast_fp16")]; + tensor reshape_88_shape_0 = const()[name = tensor("reshape_88_shape_0"), val = tensor([-1])]; + tensor reshape_88_cast_fp16 = reshape(shape = reshape_88_shape_0, x = reshape_84_cast_fp16)[name = tensor("reshape_88_cast_fp16")]; + tensor scatter_17_mode_0 = const()[name = tensor("scatter_17_mode_0"), val = tensor("update")]; + tensor scatter_17_axis_0 = const()[name = tensor("scatter_17_axis_0"), val = tensor(0)]; + tensor scatter_17_validate_indices_0 = const()[name = tensor("scatter_17_validate_indices_0"), val = tensor(false)]; + tensor scatter_17_cast_fp16 = scatter(axis = scatter_17_axis_0, data = reshape_88_cast_fp16, indices = reshape_86, mode = scatter_17_mode_0, updates = reshape_87_cast_fp16, validate_indices = scatter_17_validate_indices_0)[name = tensor("scatter_17_cast_fp16")]; + tensor reshape_89_cast_fp16 = reshape(shape = shape_63, x = scatter_17_cast_fp16)[name = tensor("reshape_89_cast_fp16")]; + tensor reshape_89_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("reshape_89_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor keys_49_begin_0 = const()[name = tensor("keys_49_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor keys_49_end_0 = const()[name = tensor("keys_49_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor keys_49_end_mask_0 = const()[name = tensor("keys_49_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor keys_49_squeeze_mask_0 = const()[name = tensor("keys_49_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor keys_49_cast_fp16 = slice_by_index(begin = keys_49_begin_0, end = keys_49_end_0, end_mask = keys_49_end_mask_0, squeeze_mask = keys_49_squeeze_mask_0, x = reshape_89_cast_fp16)[name = tensor("keys_49_cast_fp16")]; + tensor values_49_begin_0 = const()[name = tensor("values_49_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor values_49_end_0 = const()[name = tensor("values_49_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor values_49_end_mask_0 = const()[name = tensor("values_49_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor values_49_squeeze_mask_0 = const()[name = tensor("values_49_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor values_49_cast_fp16 = slice_by_index(begin = values_49_begin_0, end = values_49_end_0, end_mask = values_49_end_mask_0, squeeze_mask = values_49_squeeze_mask_0, x = reshape_89_cast_fp16)[name = tensor("values_49_cast_fp16")]; + tensor var_3752_cast_fp16 = not_equal(x = keys_49_cast_fp16, y = keys_49_cast_fp16)[name = tensor("op_3752_cast_fp16")]; + tensor keys_51_cast_fp16 = select(a = var_470_to_fp16, b = keys_49_cast_fp16, cond = var_3752_cast_fp16)[name = tensor("keys_51_cast_fp16")]; + tensor var_3760_cast_fp16 = not_equal(x = values_49_cast_fp16, y = values_49_cast_fp16)[name = tensor("op_3760_cast_fp16")]; + tensor values_51_cast_fp16 = select(a = var_470_to_fp16, b = values_49_cast_fp16, cond = var_3760_cast_fp16)[name = tensor("values_51_cast_fp16")]; + tensor var_3784 = const()[name = tensor("op_3784"), val = tensor([0, 2, 1, 3])]; + tensor var_3800 = const()[name = tensor("op_3800"), val = tensor([1, 1, 1])]; + tensor var_3801_cast_fp16 = reshape(shape = var_3800, x = position8_to_fp16)[name = tensor("op_3801_cast_fp16")]; + tensor q_positions_17_cast_fp16 = add(x = var_3801_cast_fp16, y = q_offsets_1_promoted_to_fp16)[name = tensor("q_positions_17_cast_fp16")]; + tensor valid_end_17_cast_fp16 = add(x = var_3801_cast_fp16, y = var_536_cast_fp16)[name = tensor("valid_end_17_cast_fp16")]; + tensor valid_mask_17_cast_fp16 = less(x = k_positions_1_promoted_to_fp16, y = valid_end_17_cast_fp16)[name = tensor("valid_mask_17_cast_fp16")]; + tensor causal_mask_17_cast_fp16 = less_equal(x = k_positions_1_promoted_to_fp16, y = q_positions_17_cast_fp16)[name = tensor("causal_mask_17_cast_fp16")]; + tensor var_3829 = logical_and(x = valid_mask_17_cast_fp16, y = causal_mask_17_cast_fp16)[name = tensor("op_3829")]; + tensor attn_mask_17_axes_0 = const()[name = tensor("attn_mask_17_axes_0"), val = tensor([1])]; + tensor attn_mask_17 = expand_dims(axes = attn_mask_17_axes_0, x = var_3829)[name = tensor("attn_mask_17")]; + tensor var_3843_transpose_x_0 = const()[name = tensor("op_3843_transpose_x_0"), val = tensor(false)]; + tensor var_3843_transpose_y_0 = const()[name = tensor("op_3843_transpose_y_0"), val = tensor(false)]; + tensor transpose_85_perm_0 = const()[name = tensor("transpose_85_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_86_perm_0 = const()[name = tensor("transpose_86_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_86 = transpose(perm = transpose_86_perm_0, x = keys_51_cast_fp16)[name = tensor("transpose_172")]; + tensor transpose_85 = transpose(perm = transpose_85_perm_0, x = q_51_cast_fp16)[name = tensor("transpose_173")]; + tensor var_3843_cast_fp16 = matmul(transpose_x = var_3843_transpose_x_0, transpose_y = var_3843_transpose_y_0, x = transpose_85, y = transpose_86)[name = tensor("op_3843_cast_fp16")]; + tensor var_3837_to_fp16 = const()[name = tensor("op_3837_to_fp16"), val = tensor([0x1p-3])]; + tensor attn_weights_49_cast_fp16 = mul(x = var_3843_cast_fp16, y = var_3837_to_fp16)[name = tensor("attn_weights_49_cast_fp16")]; + tensor var_3845 = logical_not(x = attn_mask_17)[name = tensor("op_3845")]; + tensor var_3846_to_fp16 = const()[name = tensor("op_3846_to_fp16"), val = tensor(-inf)]; + tensor attn_weights_51_cast_fp16 = select(a = var_3846_to_fp16, b = attn_weights_49_cast_fp16, cond = var_3845)[name = tensor("attn_weights_51_cast_fp16")]; + tensor var_3848 = const()[name = tensor("op_3848"), val = tensor(-1)]; + tensor attn_weights_53_cast_fp16 = softmax(axis = var_3848, x = attn_weights_51_cast_fp16)[name = tensor("attn_weights_53_cast_fp16")]; + tensor attn_output_17_transpose_x_0 = const()[name = tensor("attn_output_17_transpose_x_0"), val = tensor(false)]; + tensor attn_output_17_transpose_y_0 = const()[name = tensor("attn_output_17_transpose_y_0"), val = tensor(false)]; + tensor values_53_cast_fp16 = transpose(perm = var_3784, x = values_51_cast_fp16)[name = tensor("transpose_174")]; + tensor attn_output_17_cast_fp16 = matmul(transpose_x = attn_output_17_transpose_x_0, transpose_y = attn_output_17_transpose_y_0, x = attn_weights_53_cast_fp16, y = values_53_cast_fp16)[name = tensor("attn_output_17_cast_fp16")]; + tensor var_3856 = const()[name = tensor("op_3856"), val = tensor([0, 2, 1, 3])]; + tensor var_3861 = const()[name = tensor("op_3861"), val = tensor([1, 256, 1024])]; + tensor var_3857_cast_fp16 = transpose(perm = var_3856, x = attn_output_17_cast_fp16)[name = tensor("transpose_171")]; + tensor input_101_cast_fp16 = reshape(shape = var_3861, x = var_3857_cast_fp16)[name = tensor("input_101_cast_fp16")]; + tensor attn8_out_proj_weight_to_fp16 = const()[name = tensor("attn8_out_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(208761216)))]; + tensor linear_33_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = attn8_out_proj_weight_to_fp16, x = input_101_cast_fp16)[name = tensor("linear_33_cast_fp16")]; + tensor var_3870_cast_fp16 = add(x = position8_to_fp16, y = valid_len_to_fp16)[name = tensor("op_3870_cast_fp16")]; + tensor var_3870_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_3870_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor input_103_cast_fp16 = add(x = input_97_cast_fp16, y = linear_33_cast_fp16)[name = tensor("input_103_cast_fp16")]; + tensor input_105_axes_0 = const()[name = tensor("input_105_axes_0"), val = tensor([-1])]; + tensor norm8_2_weight_to_fp16 = const()[name = tensor("norm8_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(210858432)))]; + tensor norm8_2_bias_to_fp16 = const()[name = tensor("norm8_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(210860544)))]; + tensor var_3874_to_fp16 = const()[name = tensor("op_3874_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_105_cast_fp16 = layer_norm(axes = input_105_axes_0, beta = norm8_2_bias_to_fp16, epsilon = var_3874_to_fp16, gamma = norm8_2_weight_to_fp16, x = input_103_cast_fp16)[name = tensor("input_105_cast_fp16")]; + tensor linear8_1_weight_to_fp16 = const()[name = tensor("linear8_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(210862656)))]; + tensor linear_34_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = linear8_1_weight_to_fp16, x = input_105_cast_fp16)[name = tensor("linear_34_cast_fp16")]; + tensor input_107_mode_0 = const()[name = tensor("input_107_mode_0"), val = tensor("EXACT")]; + tensor input_107_cast_fp16 = gelu(mode = input_107_mode_0, x = linear_34_cast_fp16)[name = tensor("input_107_cast_fp16")]; + tensor linear8_2_weight_to_fp16 = const()[name = tensor("linear8_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(219251328)))]; + tensor linear_35_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = linear8_2_weight_to_fp16, x = input_107_cast_fp16)[name = tensor("linear_35_cast_fp16")]; + tensor input_109_cast_fp16 = add(x = input_103_cast_fp16, y = linear_35_cast_fp16)[name = tensor("input_109_cast_fp16")]; + tensor input_111_axes_0 = const()[name = tensor("input_111_axes_0"), val = tensor([-1])]; + tensor norm9_1_weight_to_fp16 = const()[name = tensor("norm9_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(227640000)))]; + tensor norm9_1_bias_to_fp16 = const()[name = tensor("norm9_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(227642112)))]; + tensor var_3891_to_fp16 = const()[name = tensor("op_3891_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_111_cast_fp16 = layer_norm(axes = input_111_axes_0, beta = norm9_1_bias_to_fp16, epsilon = var_3891_to_fp16, gamma = norm9_1_weight_to_fp16, x = input_109_cast_fp16)[name = tensor("input_111_cast_fp16")]; + tensor attn9_in_proj_weight_to_fp16 = const()[name = tensor("attn9_in_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(227644224)))]; + tensor linear_36_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = attn9_in_proj_weight_to_fp16, x = input_111_cast_fp16)[name = tensor("linear_36_cast_fp16")]; + tensor var_3905 = const()[name = tensor("op_3905"), val = tensor([1, 256, 3, 16, 64])]; + tensor qkv_19_cast_fp16 = reshape(shape = var_3905, x = linear_36_cast_fp16)[name = tensor("qkv_19_cast_fp16")]; + tensor q_55_begin_0 = const()[name = tensor("q_55_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor q_55_end_0 = const()[name = tensor("q_55_end_0"), val = tensor([1, 256, 1, 16, 64])]; + tensor q_55_end_mask_0 = const()[name = tensor("q_55_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor q_55_squeeze_mask_0 = const()[name = tensor("q_55_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor q_55_cast_fp16 = slice_by_index(begin = q_55_begin_0, end = q_55_end_0, end_mask = q_55_end_mask_0, squeeze_mask = q_55_squeeze_mask_0, x = qkv_19_cast_fp16)[name = tensor("q_55_cast_fp16")]; + tensor k_37_begin_0 = const()[name = tensor("k_37_begin_0"), val = tensor([0, 0, 1, 0, 0])]; + tensor k_37_end_0 = const()[name = tensor("k_37_end_0"), val = tensor([1, 256, 2, 16, 64])]; + tensor k_37_end_mask_0 = const()[name = tensor("k_37_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor k_37_squeeze_mask_0 = const()[name = tensor("k_37_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor k_37_cast_fp16 = slice_by_index(begin = k_37_begin_0, end = k_37_end_0, end_mask = k_37_end_mask_0, squeeze_mask = k_37_squeeze_mask_0, x = qkv_19_cast_fp16)[name = tensor("k_37_cast_fp16")]; + tensor v_19_begin_0 = const()[name = tensor("v_19_begin_0"), val = tensor([0, 0, 2, 0, 0])]; + tensor v_19_end_0 = const()[name = tensor("v_19_end_0"), val = tensor([1, 256, 3, 16, 64])]; + tensor v_19_end_mask_0 = const()[name = tensor("v_19_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor v_19_squeeze_mask_0 = const()[name = tensor("v_19_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor v_19_cast_fp16 = slice_by_index(begin = v_19_begin_0, end = v_19_end_0, end_mask = v_19_end_mask_0, squeeze_mask = v_19_squeeze_mask_0, x = qkv_19_cast_fp16)[name = tensor("v_19_cast_fp16")]; + tensor var_3963 = const()[name = tensor("op_3963"), val = tensor([1, 1])]; + tensor position9_to_fp16_dtype_0 = const()[name = tensor("position9_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor position9_to_fp16 = cast(dtype = position9_to_fp16_dtype_0, x = position9)[name = tensor("cast_616")]; + tensor var_3964_cast_fp16 = reshape(shape = var_3963, x = position9_to_fp16)[name = tensor("op_3964_cast_fp16")]; + tensor ts_57_cast_fp16 = add(x = ts_1_promoted_to_fp16, y = var_3964_cast_fp16)[name = tensor("ts_57_cast_fp16")]; + tensor var_3971 = const()[name = tensor("op_3971"), val = tensor([1, 256, 1, 1])]; + tensor ts_59_cast_fp16 = reshape(shape = var_3971, x = ts_57_cast_fp16)[name = tensor("ts_59_cast_fp16")]; + tensor var_3978 = const()[name = tensor("op_3978"), val = tensor([1, 256, 16, 32, 2])]; + tensor q_complex_19_cast_fp16 = reshape(shape = var_3978, x = q_55_cast_fp16)[name = tensor("q_complex_19_cast_fp16")]; + tensor var_3985 = const()[name = tensor("op_3985"), val = tensor([1, 256, 16, 32, 2])]; + tensor k_complex_19_cast_fp16 = reshape(shape = var_3985, x = k_37_cast_fp16)[name = tensor("k_complex_19_cast_fp16")]; + tensor var_3989_begin_0 = const()[name = tensor("op_3989_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_3989_end_0 = const()[name = tensor("op_3989_end_0"), val = tensor([1, 256, 16, 32, 1])]; + tensor var_3989_end_mask_0 = const()[name = tensor("op_3989_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_3989_squeeze_mask_0 = const()[name = tensor("op_3989_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_3989_cast_fp16 = slice_by_index(begin = var_3989_begin_0, end = var_3989_end_0, end_mask = var_3989_end_mask_0, squeeze_mask = var_3989_squeeze_mask_0, x = q_complex_19_cast_fp16)[name = tensor("op_3989_cast_fp16")]; + tensor var_3997_begin_0 = const()[name = tensor("op_3997_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_3997_end_0 = const()[name = tensor("op_3997_end_0"), val = tensor([1, 256, 16, 32, 2])]; + tensor var_3997_end_mask_0 = const()[name = tensor("op_3997_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_3997_squeeze_mask_0 = const()[name = tensor("op_3997_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_3997_cast_fp16 = slice_by_index(begin = var_3997_begin_0, end = var_3997_end_0, end_mask = var_3997_end_mask_0, squeeze_mask = var_3997_squeeze_mask_0, x = q_complex_19_cast_fp16)[name = tensor("op_3997_cast_fp16")]; + tensor var_4005_begin_0 = const()[name = tensor("op_4005_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_4005_end_0 = const()[name = tensor("op_4005_end_0"), val = tensor([1, 256, 16, 32, 1])]; + tensor var_4005_end_mask_0 = const()[name = tensor("op_4005_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_4005_squeeze_mask_0 = const()[name = tensor("op_4005_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_4005_cast_fp16 = slice_by_index(begin = var_4005_begin_0, end = var_4005_end_0, end_mask = var_4005_end_mask_0, squeeze_mask = var_4005_squeeze_mask_0, x = k_complex_19_cast_fp16)[name = tensor("op_4005_cast_fp16")]; + tensor var_4013_begin_0 = const()[name = tensor("op_4013_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_4013_end_0 = const()[name = tensor("op_4013_end_0"), val = tensor([1, 256, 16, 32, 2])]; + tensor var_4013_end_mask_0 = const()[name = tensor("op_4013_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_4013_squeeze_mask_0 = const()[name = tensor("op_4013_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_4013_cast_fp16 = slice_by_index(begin = var_4013_begin_0, end = var_4013_end_0, end_mask = var_4013_end_mask_0, squeeze_mask = var_4013_squeeze_mask_0, x = k_complex_19_cast_fp16)[name = tensor("op_4013_cast_fp16")]; + tensor freqs_19_to_fp16 = const()[name = tensor("freqs_19_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(233935744)))]; + tensor var_4019_cast_fp16 = mul(x = freqs_19_to_fp16, y = ts_59_cast_fp16)[name = tensor("op_4019_cast_fp16")]; + tensor rotr_19_cast_fp16 = cos(x = var_4019_cast_fp16)[name = tensor("rotr_19_cast_fp16")]; + tensor roti_19_cast_fp16 = sin(x = var_4019_cast_fp16)[name = tensor("roti_19_cast_fp16")]; + tensor var_4023_cast_fp16 = mul(x = var_3989_cast_fp16, y = rotr_19_cast_fp16)[name = tensor("op_4023_cast_fp16")]; + tensor var_4024_cast_fp16 = mul(x = var_3997_cast_fp16, y = roti_19_cast_fp16)[name = tensor("op_4024_cast_fp16")]; + tensor qor_37_cast_fp16 = sub(x = var_4023_cast_fp16, y = var_4024_cast_fp16)[name = tensor("qor_37_cast_fp16")]; + tensor var_4027_cast_fp16 = mul(x = var_3989_cast_fp16, y = roti_19_cast_fp16)[name = tensor("op_4027_cast_fp16")]; + tensor var_4028_cast_fp16 = mul(x = var_3997_cast_fp16, y = rotr_19_cast_fp16)[name = tensor("op_4028_cast_fp16")]; + tensor qoi_37_cast_fp16 = add(x = var_4027_cast_fp16, y = var_4028_cast_fp16)[name = tensor("qoi_37_cast_fp16")]; + tensor var_4031_cast_fp16 = mul(x = var_4005_cast_fp16, y = rotr_19_cast_fp16)[name = tensor("op_4031_cast_fp16")]; + tensor var_4032_cast_fp16 = mul(x = var_4013_cast_fp16, y = roti_19_cast_fp16)[name = tensor("op_4032_cast_fp16")]; + tensor kor_37_cast_fp16 = sub(x = var_4031_cast_fp16, y = var_4032_cast_fp16)[name = tensor("kor_37_cast_fp16")]; + tensor var_4035_cast_fp16 = mul(x = var_4005_cast_fp16, y = roti_19_cast_fp16)[name = tensor("op_4035_cast_fp16")]; + tensor var_4036_cast_fp16 = mul(x = var_4013_cast_fp16, y = rotr_19_cast_fp16)[name = tensor("op_4036_cast_fp16")]; + tensor koi_37_cast_fp16 = add(x = var_4035_cast_fp16, y = var_4036_cast_fp16)[name = tensor("koi_37_cast_fp16")]; + tensor qo_19_axis_0 = const()[name = tensor("qo_19_axis_0"), val = tensor(-1)]; + tensor qo_19_cast_fp16 = stack(axis = qo_19_axis_0, values = (qor_37_cast_fp16, qoi_37_cast_fp16))[name = tensor("qo_19_cast_fp16")]; + tensor ko_19_axis_0 = const()[name = tensor("ko_19_axis_0"), val = tensor(-1)]; + tensor ko_19_cast_fp16 = stack(axis = ko_19_axis_0, values = (kor_37_cast_fp16, koi_37_cast_fp16))[name = tensor("ko_19_cast_fp16")]; + tensor var_4069 = const()[name = tensor("op_4069"), val = tensor([1, 256, 16, 64])]; + tensor q_57_cast_fp16 = reshape(shape = var_4069, x = qo_19_cast_fp16)[name = tensor("q_57_cast_fp16")]; + tensor var_4075 = const()[name = tensor("op_4075"), val = tensor([1, 256, 16, 64])]; + tensor k_39_cast_fp16 = reshape(shape = var_4075, x = ko_19_cast_fp16)[name = tensor("k_39_cast_fp16")]; + tensor write_indices_float_55_cast_fp16 = add(x = var_3964_cast_fp16, y = var_393_promoted_to_fp16)[name = tensor("write_indices_float_55_cast_fp16")]; + tensor _inversed_4096_y_0_to_fp16 = const()[name = tensor("_inversed_4096_y_0_to_fp16"), val = tensor(0x1p-9)]; + tensor _inversed_4096_cast_fp16 = mul(x = write_indices_float_55_cast_fp16, y = _inversed_4096_y_0_to_fp16)[name = tensor("_inversed_4096_cast_fp16")]; + tensor var_4097_cast_fp16 = floor(x = _inversed_4096_cast_fp16)[name = tensor("op_4097_cast_fp16")]; + tensor var_4098_to_fp16 = const()[name = tensor("op_4098_to_fp16"), val = tensor(0x1p+9)]; + tensor var_4099_cast_fp16 = mul(x = var_4097_cast_fp16, y = var_4098_to_fp16)[name = tensor("op_4099_cast_fp16")]; + tensor write_indices_float_57_cast_fp16 = sub(x = write_indices_float_55_cast_fp16, y = var_4099_cast_fp16)[name = tensor("write_indices_float_57_cast_fp16")]; + tensor var_4117_value_0_to_fp16 = const()[name = tensor("op_4117_value_0_to_fp16"), val = tensor(0x1.ffp+8)]; + tensor var_4117_cast_fp16 = fill_like(ref_tensor = write_indices_float_57_cast_fp16, value = var_4117_value_0_to_fp16)[name = tensor("op_4117_cast_fp16")]; + tensor write_indices_float_59_cast_fp16 = select(a = write_indices_float_57_cast_fp16, b = var_4117_cast_fp16, cond = token_valid_1_cast_fp16)[name = tensor("write_indices_float_59_cast_fp16")]; + tensor var_4123_dtype_0 = const()[name = tensor("op_4123_dtype_0"), val = tensor("int32")]; + tensor var_4128 = const()[name = tensor("op_4128"), val = tensor([1, 256, 1, 1])]; + tensor write_indices_float_59_cast_fp16_to_int32 = cast(dtype = var_4123_dtype_0, x = write_indices_float_59_cast_fp16)[name = tensor("cast_615")]; + tensor var_4129 = reshape(shape = var_4128, x = write_indices_float_59_cast_fp16_to_int32)[name = tensor("op_4129")]; + tensor write_indices_19_reps_0 = const()[name = tensor("write_indices_19_reps_0"), val = tensor([1, 1, 16, 64])]; + tensor write_indices_19 = tile(reps = write_indices_19_reps_0, x = var_4129)[name = tensor("write_indices_19")]; + tensor var_4139_begin_0 = const()[name = tensor("op_4139_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_4139_end_0 = const()[name = tensor("op_4139_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor var_4139_end_mask_0 = const()[name = tensor("op_4139_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_4139_squeeze_mask_0 = const()[name = tensor("op_4139_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor cache9_to_fp16_dtype_0 = const()[name = tensor("cache9_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor cache9_to_fp16 = cast(dtype = cache9_to_fp16_dtype_0, x = cache9)[name = tensor("cast_614")]; + tensor var_4139_cast_fp16 = slice_by_index(begin = var_4139_begin_0, end = var_4139_end_0, end_mask = var_4139_end_mask_0, squeeze_mask = var_4139_squeeze_mask_0, x = cache9_to_fp16)[name = tensor("op_4139_cast_fp16")]; + tensor var_4141_axis_0 = const()[name = tensor("op_4141_axis_0"), val = tensor(1)]; + tensor var_4141_mode_0 = const()[name = tensor("op_4141_mode_0"), val = tensor("update")]; + tensor var_4141_validate_indices_0 = const()[name = tensor("op_4141_validate_indices_0"), val = tensor(false)]; + tensor var_4141_cast_fp16 = scatter_along_axis(axis = var_4141_axis_0, data = var_4139_cast_fp16, indices = write_indices_19, mode = var_4141_mode_0, updates = k_39_cast_fp16, validate_indices = var_4141_validate_indices_0)[name = tensor("op_4141_cast_fp16")]; + tensor concat_74 = const()[name = tensor("concat_74"), val = tensor([0, 0, 0, 0, 0])]; + tensor concat_75 = const()[name = tensor("concat_75"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_19_internal_tensor_assign_1_stride_0 = const()[name = tensor("new_cache_19_internal_tensor_assign_1_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_19_internal_tensor_assign_1_begin_mask_0 = const()[name = tensor("new_cache_19_internal_tensor_assign_1_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_19_internal_tensor_assign_1_end_mask_0 = const()[name = tensor("new_cache_19_internal_tensor_assign_1_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_19_internal_tensor_assign_1_squeeze_mask_0 = const()[name = tensor("new_cache_19_internal_tensor_assign_1_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_64 = const()[name = tensor("shape_64"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_18 = const()[name = tensor("reduce_prod_18"), val = tensor(1048576)]; + tensor range_1d_18_start_0 = const()[name = tensor("range_1d_18_start_0"), val = tensor(0)]; + tensor range_1d_18_step_0 = const()[name = tensor("range_1d_18_step_0"), val = tensor(1)]; + tensor range_1d_18 = range_1d(end = reduce_prod_18, start = range_1d_18_start_0, step = range_1d_18_step_0)[name = tensor("range_1d_18")]; + tensor reshape_90 = reshape(shape = shape_64, x = range_1d_18)[name = tensor("reshape_90")]; + tensor slice_by_index_18 = slice_by_index(begin = concat_74, begin_mask = new_cache_19_internal_tensor_assign_1_begin_mask_0, end = concat_75, end_mask = new_cache_19_internal_tensor_assign_1_end_mask_0, squeeze_mask = new_cache_19_internal_tensor_assign_1_squeeze_mask_0, stride = new_cache_19_internal_tensor_assign_1_stride_0, x = reshape_90)[name = tensor("slice_by_index_18")]; + tensor reshape_91_shape_0 = const()[name = tensor("reshape_91_shape_0"), val = tensor([-1])]; + tensor reshape_91 = reshape(shape = reshape_91_shape_0, x = slice_by_index_18)[name = tensor("reshape_91")]; + tensor reshape_92_shape_0 = const()[name = tensor("reshape_92_shape_0"), val = tensor([-1])]; + tensor reshape_92_cast_fp16 = reshape(shape = reshape_92_shape_0, x = var_4141_cast_fp16)[name = tensor("reshape_92_cast_fp16")]; + tensor reshape_93_shape_0 = const()[name = tensor("reshape_93_shape_0"), val = tensor([-1])]; + tensor reshape_93_cast_fp16 = reshape(shape = reshape_93_shape_0, x = cache9_to_fp16)[name = tensor("reshape_93_cast_fp16")]; + tensor scatter_18_mode_0 = const()[name = tensor("scatter_18_mode_0"), val = tensor("update")]; + tensor scatter_18_axis_0 = const()[name = tensor("scatter_18_axis_0"), val = tensor(0)]; + tensor scatter_18_validate_indices_0 = const()[name = tensor("scatter_18_validate_indices_0"), val = tensor(false)]; + tensor scatter_18_cast_fp16 = scatter(axis = scatter_18_axis_0, data = reshape_93_cast_fp16, indices = reshape_91, mode = scatter_18_mode_0, updates = reshape_92_cast_fp16, validate_indices = scatter_18_validate_indices_0)[name = tensor("scatter_18_cast_fp16")]; + tensor reshape_94_cast_fp16 = reshape(shape = shape_64, x = scatter_18_cast_fp16)[name = tensor("reshape_94_cast_fp16")]; + tensor var_4149_begin_0 = const()[name = tensor("op_4149_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor var_4149_end_0 = const()[name = tensor("op_4149_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor var_4149_end_mask_0 = const()[name = tensor("op_4149_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_4149_squeeze_mask_0 = const()[name = tensor("op_4149_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor var_4149_cast_fp16 = slice_by_index(begin = var_4149_begin_0, end = var_4149_end_0, end_mask = var_4149_end_mask_0, squeeze_mask = var_4149_squeeze_mask_0, x = reshape_94_cast_fp16)[name = tensor("op_4149_cast_fp16")]; + tensor var_4151_axis_0 = const()[name = tensor("op_4151_axis_0"), val = tensor(1)]; + tensor var_4151_mode_0 = const()[name = tensor("op_4151_mode_0"), val = tensor("update")]; + tensor var_4151_validate_indices_0 = const()[name = tensor("op_4151_validate_indices_0"), val = tensor(false)]; + tensor var_4151_cast_fp16 = scatter_along_axis(axis = var_4151_axis_0, data = var_4149_cast_fp16, indices = write_indices_19, mode = var_4151_mode_0, updates = v_19_cast_fp16, validate_indices = var_4151_validate_indices_0)[name = tensor("op_4151_cast_fp16")]; + tensor concat_76 = const()[name = tensor("concat_76"), val = tensor([1, 0, 0, 0, 0])]; + tensor concat_77 = const()[name = tensor("concat_77"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_19_internal_tensor_assign_2_stride_0 = const()[name = tensor("new_cache_19_internal_tensor_assign_2_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_19_internal_tensor_assign_2_begin_mask_0 = const()[name = tensor("new_cache_19_internal_tensor_assign_2_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_19_internal_tensor_assign_2_end_mask_0 = const()[name = tensor("new_cache_19_internal_tensor_assign_2_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_19_internal_tensor_assign_2_squeeze_mask_0 = const()[name = tensor("new_cache_19_internal_tensor_assign_2_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_65 = const()[name = tensor("shape_65"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_19 = const()[name = tensor("reduce_prod_19"), val = tensor(1048576)]; + tensor range_1d_19_start_0 = const()[name = tensor("range_1d_19_start_0"), val = tensor(0)]; + tensor range_1d_19_step_0 = const()[name = tensor("range_1d_19_step_0"), val = tensor(1)]; + tensor range_1d_19 = range_1d(end = reduce_prod_19, start = range_1d_19_start_0, step = range_1d_19_step_0)[name = tensor("range_1d_19")]; + tensor reshape_95 = reshape(shape = shape_65, x = range_1d_19)[name = tensor("reshape_95")]; + tensor slice_by_index_19 = slice_by_index(begin = concat_76, begin_mask = new_cache_19_internal_tensor_assign_2_begin_mask_0, end = concat_77, end_mask = new_cache_19_internal_tensor_assign_2_end_mask_0, squeeze_mask = new_cache_19_internal_tensor_assign_2_squeeze_mask_0, stride = new_cache_19_internal_tensor_assign_2_stride_0, x = reshape_95)[name = tensor("slice_by_index_19")]; + tensor reshape_96_shape_0 = const()[name = tensor("reshape_96_shape_0"), val = tensor([-1])]; + tensor reshape_96 = reshape(shape = reshape_96_shape_0, x = slice_by_index_19)[name = tensor("reshape_96")]; + tensor reshape_97_shape_0 = const()[name = tensor("reshape_97_shape_0"), val = tensor([-1])]; + tensor reshape_97_cast_fp16 = reshape(shape = reshape_97_shape_0, x = var_4151_cast_fp16)[name = tensor("reshape_97_cast_fp16")]; + tensor reshape_98_shape_0 = const()[name = tensor("reshape_98_shape_0"), val = tensor([-1])]; + tensor reshape_98_cast_fp16 = reshape(shape = reshape_98_shape_0, x = reshape_94_cast_fp16)[name = tensor("reshape_98_cast_fp16")]; + tensor scatter_19_mode_0 = const()[name = tensor("scatter_19_mode_0"), val = tensor("update")]; + tensor scatter_19_axis_0 = const()[name = tensor("scatter_19_axis_0"), val = tensor(0)]; + tensor scatter_19_validate_indices_0 = const()[name = tensor("scatter_19_validate_indices_0"), val = tensor(false)]; + tensor scatter_19_cast_fp16 = scatter(axis = scatter_19_axis_0, data = reshape_98_cast_fp16, indices = reshape_96, mode = scatter_19_mode_0, updates = reshape_97_cast_fp16, validate_indices = scatter_19_validate_indices_0)[name = tensor("scatter_19_cast_fp16")]; + tensor reshape_99_cast_fp16 = reshape(shape = shape_65, x = scatter_19_cast_fp16)[name = tensor("reshape_99_cast_fp16")]; + tensor reshape_99_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("reshape_99_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor keys_55_begin_0 = const()[name = tensor("keys_55_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor keys_55_end_0 = const()[name = tensor("keys_55_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor keys_55_end_mask_0 = const()[name = tensor("keys_55_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor keys_55_squeeze_mask_0 = const()[name = tensor("keys_55_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor keys_55_cast_fp16 = slice_by_index(begin = keys_55_begin_0, end = keys_55_end_0, end_mask = keys_55_end_mask_0, squeeze_mask = keys_55_squeeze_mask_0, x = reshape_99_cast_fp16)[name = tensor("keys_55_cast_fp16")]; + tensor values_55_begin_0 = const()[name = tensor("values_55_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor values_55_end_0 = const()[name = tensor("values_55_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor values_55_end_mask_0 = const()[name = tensor("values_55_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor values_55_squeeze_mask_0 = const()[name = tensor("values_55_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor values_55_cast_fp16 = slice_by_index(begin = values_55_begin_0, end = values_55_end_0, end_mask = values_55_end_mask_0, squeeze_mask = values_55_squeeze_mask_0, x = reshape_99_cast_fp16)[name = tensor("values_55_cast_fp16")]; + tensor var_4163_cast_fp16 = not_equal(x = keys_55_cast_fp16, y = keys_55_cast_fp16)[name = tensor("op_4163_cast_fp16")]; + tensor keys_57_cast_fp16 = select(a = var_470_to_fp16, b = keys_55_cast_fp16, cond = var_4163_cast_fp16)[name = tensor("keys_57_cast_fp16")]; + tensor var_4171_cast_fp16 = not_equal(x = values_55_cast_fp16, y = values_55_cast_fp16)[name = tensor("op_4171_cast_fp16")]; + tensor values_57_cast_fp16 = select(a = var_470_to_fp16, b = values_55_cast_fp16, cond = var_4171_cast_fp16)[name = tensor("values_57_cast_fp16")]; + tensor var_4195 = const()[name = tensor("op_4195"), val = tensor([0, 2, 1, 3])]; + tensor var_4211 = const()[name = tensor("op_4211"), val = tensor([1, 1, 1])]; + tensor var_4212_cast_fp16 = reshape(shape = var_4211, x = position9_to_fp16)[name = tensor("op_4212_cast_fp16")]; + tensor q_positions_19_cast_fp16 = add(x = var_4212_cast_fp16, y = q_offsets_1_promoted_to_fp16)[name = tensor("q_positions_19_cast_fp16")]; + tensor valid_end_19_cast_fp16 = add(x = var_4212_cast_fp16, y = var_536_cast_fp16)[name = tensor("valid_end_19_cast_fp16")]; + tensor valid_mask_19_cast_fp16 = less(x = k_positions_1_promoted_to_fp16, y = valid_end_19_cast_fp16)[name = tensor("valid_mask_19_cast_fp16")]; + tensor causal_mask_19_cast_fp16 = less_equal(x = k_positions_1_promoted_to_fp16, y = q_positions_19_cast_fp16)[name = tensor("causal_mask_19_cast_fp16")]; + tensor var_4240 = logical_and(x = valid_mask_19_cast_fp16, y = causal_mask_19_cast_fp16)[name = tensor("op_4240")]; + tensor attn_mask_19_axes_0 = const()[name = tensor("attn_mask_19_axes_0"), val = tensor([1])]; + tensor attn_mask_19 = expand_dims(axes = attn_mask_19_axes_0, x = var_4240)[name = tensor("attn_mask_19")]; + tensor var_4254_transpose_x_0 = const()[name = tensor("op_4254_transpose_x_0"), val = tensor(false)]; + tensor var_4254_transpose_y_0 = const()[name = tensor("op_4254_transpose_y_0"), val = tensor(false)]; + tensor transpose_87_perm_0 = const()[name = tensor("transpose_87_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_88_perm_0 = const()[name = tensor("transpose_88_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_88 = transpose(perm = transpose_88_perm_0, x = keys_57_cast_fp16)[name = tensor("transpose_168")]; + tensor transpose_87 = transpose(perm = transpose_87_perm_0, x = q_57_cast_fp16)[name = tensor("transpose_169")]; + tensor var_4254_cast_fp16 = matmul(transpose_x = var_4254_transpose_x_0, transpose_y = var_4254_transpose_y_0, x = transpose_87, y = transpose_88)[name = tensor("op_4254_cast_fp16")]; + tensor var_4248_to_fp16 = const()[name = tensor("op_4248_to_fp16"), val = tensor([0x1p-3])]; + tensor attn_weights_55_cast_fp16 = mul(x = var_4254_cast_fp16, y = var_4248_to_fp16)[name = tensor("attn_weights_55_cast_fp16")]; + tensor var_4256 = logical_not(x = attn_mask_19)[name = tensor("op_4256")]; + tensor var_4257_to_fp16 = const()[name = tensor("op_4257_to_fp16"), val = tensor(-inf)]; + tensor attn_weights_57_cast_fp16 = select(a = var_4257_to_fp16, b = attn_weights_55_cast_fp16, cond = var_4256)[name = tensor("attn_weights_57_cast_fp16")]; + tensor var_4259 = const()[name = tensor("op_4259"), val = tensor(-1)]; + tensor attn_weights_59_cast_fp16 = softmax(axis = var_4259, x = attn_weights_57_cast_fp16)[name = tensor("attn_weights_59_cast_fp16")]; + tensor attn_output_19_transpose_x_0 = const()[name = tensor("attn_output_19_transpose_x_0"), val = tensor(false)]; + tensor attn_output_19_transpose_y_0 = const()[name = tensor("attn_output_19_transpose_y_0"), val = tensor(false)]; + tensor values_59_cast_fp16 = transpose(perm = var_4195, x = values_57_cast_fp16)[name = tensor("transpose_170")]; + tensor attn_output_19_cast_fp16 = matmul(transpose_x = attn_output_19_transpose_x_0, transpose_y = attn_output_19_transpose_y_0, x = attn_weights_59_cast_fp16, y = values_59_cast_fp16)[name = tensor("attn_output_19_cast_fp16")]; + tensor var_4267 = const()[name = tensor("op_4267"), val = tensor([0, 2, 1, 3])]; + tensor var_4272 = const()[name = tensor("op_4272"), val = tensor([1, 256, 1024])]; + tensor var_4268_cast_fp16 = transpose(perm = var_4267, x = attn_output_19_cast_fp16)[name = tensor("transpose_167")]; + tensor input_113_cast_fp16 = reshape(shape = var_4272, x = var_4268_cast_fp16)[name = tensor("input_113_cast_fp16")]; + tensor attn9_out_proj_weight_to_fp16 = const()[name = tensor("attn9_out_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(233935872)))]; + tensor linear_37_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = attn9_out_proj_weight_to_fp16, x = input_113_cast_fp16)[name = tensor("linear_37_cast_fp16")]; + tensor var_4281_cast_fp16 = add(x = position9_to_fp16, y = valid_len_to_fp16)[name = tensor("op_4281_cast_fp16")]; + tensor var_4281_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_4281_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor input_115_cast_fp16 = add(x = input_109_cast_fp16, y = linear_37_cast_fp16)[name = tensor("input_115_cast_fp16")]; + tensor input_117_axes_0 = const()[name = tensor("input_117_axes_0"), val = tensor([-1])]; + tensor norm9_2_weight_to_fp16 = const()[name = tensor("norm9_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(236033088)))]; + tensor norm9_2_bias_to_fp16 = const()[name = tensor("norm9_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(236035200)))]; + tensor var_4285_to_fp16 = const()[name = tensor("op_4285_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_117_cast_fp16 = layer_norm(axes = input_117_axes_0, beta = norm9_2_bias_to_fp16, epsilon = var_4285_to_fp16, gamma = norm9_2_weight_to_fp16, x = input_115_cast_fp16)[name = tensor("input_117_cast_fp16")]; + tensor linear9_1_weight_to_fp16 = const()[name = tensor("linear9_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(236037312)))]; + tensor linear_38_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = linear9_1_weight_to_fp16, x = input_117_cast_fp16)[name = tensor("linear_38_cast_fp16")]; + tensor input_119_mode_0 = const()[name = tensor("input_119_mode_0"), val = tensor("EXACT")]; + tensor input_119_cast_fp16 = gelu(mode = input_119_mode_0, x = linear_38_cast_fp16)[name = tensor("input_119_cast_fp16")]; + tensor linear9_2_weight_to_fp16 = const()[name = tensor("linear9_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(244425984)))]; + tensor linear_39_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = linear9_2_weight_to_fp16, x = input_119_cast_fp16)[name = tensor("linear_39_cast_fp16")]; + tensor input_121_cast_fp16 = add(x = input_115_cast_fp16, y = linear_39_cast_fp16)[name = tensor("input_121_cast_fp16")]; + tensor input_123_axes_0 = const()[name = tensor("input_123_axes_0"), val = tensor([-1])]; + tensor norm10_1_weight_to_fp16 = const()[name = tensor("norm10_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(252814656)))]; + tensor norm10_1_bias_to_fp16 = const()[name = tensor("norm10_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(252816768)))]; + tensor var_4302_to_fp16 = const()[name = tensor("op_4302_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_123_cast_fp16 = layer_norm(axes = input_123_axes_0, beta = norm10_1_bias_to_fp16, epsilon = var_4302_to_fp16, gamma = norm10_1_weight_to_fp16, x = input_121_cast_fp16)[name = tensor("input_123_cast_fp16")]; + tensor attn10_in_proj_weight_to_fp16 = const()[name = tensor("attn10_in_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(252818880)))]; + tensor linear_40_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = attn10_in_proj_weight_to_fp16, x = input_123_cast_fp16)[name = tensor("linear_40_cast_fp16")]; + tensor var_4316 = const()[name = tensor("op_4316"), val = tensor([1, 256, 3, 16, 64])]; + tensor qkv_21_cast_fp16 = reshape(shape = var_4316, x = linear_40_cast_fp16)[name = tensor("qkv_21_cast_fp16")]; + tensor q_61_begin_0 = const()[name = tensor("q_61_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor q_61_end_0 = const()[name = tensor("q_61_end_0"), val = tensor([1, 256, 1, 16, 64])]; + tensor q_61_end_mask_0 = const()[name = tensor("q_61_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor q_61_squeeze_mask_0 = const()[name = tensor("q_61_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor q_61_cast_fp16 = slice_by_index(begin = q_61_begin_0, end = q_61_end_0, end_mask = q_61_end_mask_0, squeeze_mask = q_61_squeeze_mask_0, x = qkv_21_cast_fp16)[name = tensor("q_61_cast_fp16")]; + tensor k_41_begin_0 = const()[name = tensor("k_41_begin_0"), val = tensor([0, 0, 1, 0, 0])]; + tensor k_41_end_0 = const()[name = tensor("k_41_end_0"), val = tensor([1, 256, 2, 16, 64])]; + tensor k_41_end_mask_0 = const()[name = tensor("k_41_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor k_41_squeeze_mask_0 = const()[name = tensor("k_41_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor k_41_cast_fp16 = slice_by_index(begin = k_41_begin_0, end = k_41_end_0, end_mask = k_41_end_mask_0, squeeze_mask = k_41_squeeze_mask_0, x = qkv_21_cast_fp16)[name = tensor("k_41_cast_fp16")]; + tensor v_21_begin_0 = const()[name = tensor("v_21_begin_0"), val = tensor([0, 0, 2, 0, 0])]; + tensor v_21_end_0 = const()[name = tensor("v_21_end_0"), val = tensor([1, 256, 3, 16, 64])]; + tensor v_21_end_mask_0 = const()[name = tensor("v_21_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor v_21_squeeze_mask_0 = const()[name = tensor("v_21_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor v_21_cast_fp16 = slice_by_index(begin = v_21_begin_0, end = v_21_end_0, end_mask = v_21_end_mask_0, squeeze_mask = v_21_squeeze_mask_0, x = qkv_21_cast_fp16)[name = tensor("v_21_cast_fp16")]; + tensor var_4374 = const()[name = tensor("op_4374"), val = tensor([1, 1])]; + tensor position10_to_fp16_dtype_0 = const()[name = tensor("position10_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor position10_to_fp16 = cast(dtype = position10_to_fp16_dtype_0, x = position10)[name = tensor("cast_611")]; + tensor var_4375_cast_fp16 = reshape(shape = var_4374, x = position10_to_fp16)[name = tensor("op_4375_cast_fp16")]; + tensor ts_63_cast_fp16 = add(x = ts_1_promoted_to_fp16, y = var_4375_cast_fp16)[name = tensor("ts_63_cast_fp16")]; + tensor var_4382 = const()[name = tensor("op_4382"), val = tensor([1, 256, 1, 1])]; + tensor ts_65_cast_fp16 = reshape(shape = var_4382, x = ts_63_cast_fp16)[name = tensor("ts_65_cast_fp16")]; + tensor var_4389 = const()[name = tensor("op_4389"), val = tensor([1, 256, 16, 32, 2])]; + tensor q_complex_21_cast_fp16 = reshape(shape = var_4389, x = q_61_cast_fp16)[name = tensor("q_complex_21_cast_fp16")]; + tensor var_4396 = const()[name = tensor("op_4396"), val = tensor([1, 256, 16, 32, 2])]; + tensor k_complex_21_cast_fp16 = reshape(shape = var_4396, x = k_41_cast_fp16)[name = tensor("k_complex_21_cast_fp16")]; + tensor var_4400_begin_0 = const()[name = tensor("op_4400_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_4400_end_0 = const()[name = tensor("op_4400_end_0"), val = tensor([1, 256, 16, 32, 1])]; + tensor var_4400_end_mask_0 = const()[name = tensor("op_4400_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_4400_squeeze_mask_0 = const()[name = tensor("op_4400_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_4400_cast_fp16 = slice_by_index(begin = var_4400_begin_0, end = var_4400_end_0, end_mask = var_4400_end_mask_0, squeeze_mask = var_4400_squeeze_mask_0, x = q_complex_21_cast_fp16)[name = tensor("op_4400_cast_fp16")]; + tensor var_4408_begin_0 = const()[name = tensor("op_4408_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_4408_end_0 = const()[name = tensor("op_4408_end_0"), val = tensor([1, 256, 16, 32, 2])]; + tensor var_4408_end_mask_0 = const()[name = tensor("op_4408_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_4408_squeeze_mask_0 = const()[name = tensor("op_4408_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_4408_cast_fp16 = slice_by_index(begin = var_4408_begin_0, end = var_4408_end_0, end_mask = var_4408_end_mask_0, squeeze_mask = var_4408_squeeze_mask_0, x = q_complex_21_cast_fp16)[name = tensor("op_4408_cast_fp16")]; + tensor var_4416_begin_0 = const()[name = tensor("op_4416_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_4416_end_0 = const()[name = tensor("op_4416_end_0"), val = tensor([1, 256, 16, 32, 1])]; + tensor var_4416_end_mask_0 = const()[name = tensor("op_4416_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_4416_squeeze_mask_0 = const()[name = tensor("op_4416_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_4416_cast_fp16 = slice_by_index(begin = var_4416_begin_0, end = var_4416_end_0, end_mask = var_4416_end_mask_0, squeeze_mask = var_4416_squeeze_mask_0, x = k_complex_21_cast_fp16)[name = tensor("op_4416_cast_fp16")]; + tensor var_4424_begin_0 = const()[name = tensor("op_4424_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_4424_end_0 = const()[name = tensor("op_4424_end_0"), val = tensor([1, 256, 16, 32, 2])]; + tensor var_4424_end_mask_0 = const()[name = tensor("op_4424_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_4424_squeeze_mask_0 = const()[name = tensor("op_4424_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_4424_cast_fp16 = slice_by_index(begin = var_4424_begin_0, end = var_4424_end_0, end_mask = var_4424_end_mask_0, squeeze_mask = var_4424_squeeze_mask_0, x = k_complex_21_cast_fp16)[name = tensor("op_4424_cast_fp16")]; + tensor freqs_21_to_fp16 = const()[name = tensor("freqs_21_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(259110400)))]; + tensor var_4430_cast_fp16 = mul(x = freqs_21_to_fp16, y = ts_65_cast_fp16)[name = tensor("op_4430_cast_fp16")]; + tensor rotr_21_cast_fp16 = cos(x = var_4430_cast_fp16)[name = tensor("rotr_21_cast_fp16")]; + tensor roti_21_cast_fp16 = sin(x = var_4430_cast_fp16)[name = tensor("roti_21_cast_fp16")]; + tensor var_4434_cast_fp16 = mul(x = var_4400_cast_fp16, y = rotr_21_cast_fp16)[name = tensor("op_4434_cast_fp16")]; + tensor var_4435_cast_fp16 = mul(x = var_4408_cast_fp16, y = roti_21_cast_fp16)[name = tensor("op_4435_cast_fp16")]; + tensor qor_41_cast_fp16 = sub(x = var_4434_cast_fp16, y = var_4435_cast_fp16)[name = tensor("qor_41_cast_fp16")]; + tensor var_4438_cast_fp16 = mul(x = var_4400_cast_fp16, y = roti_21_cast_fp16)[name = tensor("op_4438_cast_fp16")]; + tensor var_4439_cast_fp16 = mul(x = var_4408_cast_fp16, y = rotr_21_cast_fp16)[name = tensor("op_4439_cast_fp16")]; + tensor qoi_41_cast_fp16 = add(x = var_4438_cast_fp16, y = var_4439_cast_fp16)[name = tensor("qoi_41_cast_fp16")]; + tensor var_4442_cast_fp16 = mul(x = var_4416_cast_fp16, y = rotr_21_cast_fp16)[name = tensor("op_4442_cast_fp16")]; + tensor var_4443_cast_fp16 = mul(x = var_4424_cast_fp16, y = roti_21_cast_fp16)[name = tensor("op_4443_cast_fp16")]; + tensor kor_41_cast_fp16 = sub(x = var_4442_cast_fp16, y = var_4443_cast_fp16)[name = tensor("kor_41_cast_fp16")]; + tensor var_4446_cast_fp16 = mul(x = var_4416_cast_fp16, y = roti_21_cast_fp16)[name = tensor("op_4446_cast_fp16")]; + tensor var_4447_cast_fp16 = mul(x = var_4424_cast_fp16, y = rotr_21_cast_fp16)[name = tensor("op_4447_cast_fp16")]; + tensor koi_41_cast_fp16 = add(x = var_4446_cast_fp16, y = var_4447_cast_fp16)[name = tensor("koi_41_cast_fp16")]; + tensor qo_21_axis_0 = const()[name = tensor("qo_21_axis_0"), val = tensor(-1)]; + tensor qo_21_cast_fp16 = stack(axis = qo_21_axis_0, values = (qor_41_cast_fp16, qoi_41_cast_fp16))[name = tensor("qo_21_cast_fp16")]; + tensor ko_21_axis_0 = const()[name = tensor("ko_21_axis_0"), val = tensor(-1)]; + tensor ko_21_cast_fp16 = stack(axis = ko_21_axis_0, values = (kor_41_cast_fp16, koi_41_cast_fp16))[name = tensor("ko_21_cast_fp16")]; + tensor var_4480 = const()[name = tensor("op_4480"), val = tensor([1, 256, 16, 64])]; + tensor q_63_cast_fp16 = reshape(shape = var_4480, x = qo_21_cast_fp16)[name = tensor("q_63_cast_fp16")]; + tensor var_4486 = const()[name = tensor("op_4486"), val = tensor([1, 256, 16, 64])]; + tensor k_43_cast_fp16 = reshape(shape = var_4486, x = ko_21_cast_fp16)[name = tensor("k_43_cast_fp16")]; + tensor write_indices_float_61_cast_fp16 = add(x = var_4375_cast_fp16, y = var_393_promoted_to_fp16)[name = tensor("write_indices_float_61_cast_fp16")]; + tensor _inversed_4507_y_0_to_fp16 = const()[name = tensor("_inversed_4507_y_0_to_fp16"), val = tensor(0x1p-9)]; + tensor _inversed_4507_cast_fp16 = mul(x = write_indices_float_61_cast_fp16, y = _inversed_4507_y_0_to_fp16)[name = tensor("_inversed_4507_cast_fp16")]; + tensor var_4508_cast_fp16 = floor(x = _inversed_4507_cast_fp16)[name = tensor("op_4508_cast_fp16")]; + tensor var_4509_to_fp16 = const()[name = tensor("op_4509_to_fp16"), val = tensor(0x1p+9)]; + tensor var_4510_cast_fp16 = mul(x = var_4508_cast_fp16, y = var_4509_to_fp16)[name = tensor("op_4510_cast_fp16")]; + tensor write_indices_float_63_cast_fp16 = sub(x = write_indices_float_61_cast_fp16, y = var_4510_cast_fp16)[name = tensor("write_indices_float_63_cast_fp16")]; + tensor var_4528_value_0_to_fp16 = const()[name = tensor("op_4528_value_0_to_fp16"), val = tensor(0x1.ffp+8)]; + tensor var_4528_cast_fp16 = fill_like(ref_tensor = write_indices_float_63_cast_fp16, value = var_4528_value_0_to_fp16)[name = tensor("op_4528_cast_fp16")]; + tensor write_indices_float_65_cast_fp16 = select(a = write_indices_float_63_cast_fp16, b = var_4528_cast_fp16, cond = token_valid_1_cast_fp16)[name = tensor("write_indices_float_65_cast_fp16")]; + tensor var_4534_dtype_0 = const()[name = tensor("op_4534_dtype_0"), val = tensor("int32")]; + tensor var_4539 = const()[name = tensor("op_4539"), val = tensor([1, 256, 1, 1])]; + tensor write_indices_float_65_cast_fp16_to_int32 = cast(dtype = var_4534_dtype_0, x = write_indices_float_65_cast_fp16)[name = tensor("cast_610")]; + tensor var_4540 = reshape(shape = var_4539, x = write_indices_float_65_cast_fp16_to_int32)[name = tensor("op_4540")]; + tensor write_indices_21_reps_0 = const()[name = tensor("write_indices_21_reps_0"), val = tensor([1, 1, 16, 64])]; + tensor write_indices_21 = tile(reps = write_indices_21_reps_0, x = var_4540)[name = tensor("write_indices_21")]; + tensor var_4550_begin_0 = const()[name = tensor("op_4550_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_4550_end_0 = const()[name = tensor("op_4550_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor var_4550_end_mask_0 = const()[name = tensor("op_4550_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_4550_squeeze_mask_0 = const()[name = tensor("op_4550_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor cache10_to_fp16_dtype_0 = const()[name = tensor("cache10_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor cache10_to_fp16 = cast(dtype = cache10_to_fp16_dtype_0, x = cache10)[name = tensor("cast_609")]; + tensor var_4550_cast_fp16 = slice_by_index(begin = var_4550_begin_0, end = var_4550_end_0, end_mask = var_4550_end_mask_0, squeeze_mask = var_4550_squeeze_mask_0, x = cache10_to_fp16)[name = tensor("op_4550_cast_fp16")]; + tensor var_4552_axis_0 = const()[name = tensor("op_4552_axis_0"), val = tensor(1)]; + tensor var_4552_mode_0 = const()[name = tensor("op_4552_mode_0"), val = tensor("update")]; + tensor var_4552_validate_indices_0 = const()[name = tensor("op_4552_validate_indices_0"), val = tensor(false)]; + tensor var_4552_cast_fp16 = scatter_along_axis(axis = var_4552_axis_0, data = var_4550_cast_fp16, indices = write_indices_21, mode = var_4552_mode_0, updates = k_43_cast_fp16, validate_indices = var_4552_validate_indices_0)[name = tensor("op_4552_cast_fp16")]; + tensor concat_82 = const()[name = tensor("concat_82"), val = tensor([0, 0, 0, 0, 0])]; + tensor concat_83 = const()[name = tensor("concat_83"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_21_internal_tensor_assign_1_stride_0 = const()[name = tensor("new_cache_21_internal_tensor_assign_1_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_21_internal_tensor_assign_1_begin_mask_0 = const()[name = tensor("new_cache_21_internal_tensor_assign_1_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_21_internal_tensor_assign_1_end_mask_0 = const()[name = tensor("new_cache_21_internal_tensor_assign_1_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_21_internal_tensor_assign_1_squeeze_mask_0 = const()[name = tensor("new_cache_21_internal_tensor_assign_1_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_66 = const()[name = tensor("shape_66"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_20 = const()[name = tensor("reduce_prod_20"), val = tensor(1048576)]; + tensor range_1d_20_start_0 = const()[name = tensor("range_1d_20_start_0"), val = tensor(0)]; + tensor range_1d_20_step_0 = const()[name = tensor("range_1d_20_step_0"), val = tensor(1)]; + tensor range_1d_20 = range_1d(end = reduce_prod_20, start = range_1d_20_start_0, step = range_1d_20_step_0)[name = tensor("range_1d_20")]; + tensor reshape_100 = reshape(shape = shape_66, x = range_1d_20)[name = tensor("reshape_100")]; + tensor slice_by_index_20 = slice_by_index(begin = concat_82, begin_mask = new_cache_21_internal_tensor_assign_1_begin_mask_0, end = concat_83, end_mask = new_cache_21_internal_tensor_assign_1_end_mask_0, squeeze_mask = new_cache_21_internal_tensor_assign_1_squeeze_mask_0, stride = new_cache_21_internal_tensor_assign_1_stride_0, x = reshape_100)[name = tensor("slice_by_index_20")]; + tensor reshape_101_shape_0 = const()[name = tensor("reshape_101_shape_0"), val = tensor([-1])]; + tensor reshape_101 = reshape(shape = reshape_101_shape_0, x = slice_by_index_20)[name = tensor("reshape_101")]; + tensor reshape_102_shape_0 = const()[name = tensor("reshape_102_shape_0"), val = tensor([-1])]; + tensor reshape_102_cast_fp16 = reshape(shape = reshape_102_shape_0, x = var_4552_cast_fp16)[name = tensor("reshape_102_cast_fp16")]; + tensor reshape_103_shape_0 = const()[name = tensor("reshape_103_shape_0"), val = tensor([-1])]; + tensor reshape_103_cast_fp16 = reshape(shape = reshape_103_shape_0, x = cache10_to_fp16)[name = tensor("reshape_103_cast_fp16")]; + tensor scatter_20_mode_0 = const()[name = tensor("scatter_20_mode_0"), val = tensor("update")]; + tensor scatter_20_axis_0 = const()[name = tensor("scatter_20_axis_0"), val = tensor(0)]; + tensor scatter_20_validate_indices_0 = const()[name = tensor("scatter_20_validate_indices_0"), val = tensor(false)]; + tensor scatter_20_cast_fp16 = scatter(axis = scatter_20_axis_0, data = reshape_103_cast_fp16, indices = reshape_101, mode = scatter_20_mode_0, updates = reshape_102_cast_fp16, validate_indices = scatter_20_validate_indices_0)[name = tensor("scatter_20_cast_fp16")]; + tensor reshape_104_cast_fp16 = reshape(shape = shape_66, x = scatter_20_cast_fp16)[name = tensor("reshape_104_cast_fp16")]; + tensor var_4560_begin_0 = const()[name = tensor("op_4560_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor var_4560_end_0 = const()[name = tensor("op_4560_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor var_4560_end_mask_0 = const()[name = tensor("op_4560_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_4560_squeeze_mask_0 = const()[name = tensor("op_4560_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor var_4560_cast_fp16 = slice_by_index(begin = var_4560_begin_0, end = var_4560_end_0, end_mask = var_4560_end_mask_0, squeeze_mask = var_4560_squeeze_mask_0, x = reshape_104_cast_fp16)[name = tensor("op_4560_cast_fp16")]; + tensor var_4562_axis_0 = const()[name = tensor("op_4562_axis_0"), val = tensor(1)]; + tensor var_4562_mode_0 = const()[name = tensor("op_4562_mode_0"), val = tensor("update")]; + tensor var_4562_validate_indices_0 = const()[name = tensor("op_4562_validate_indices_0"), val = tensor(false)]; + tensor var_4562_cast_fp16 = scatter_along_axis(axis = var_4562_axis_0, data = var_4560_cast_fp16, indices = write_indices_21, mode = var_4562_mode_0, updates = v_21_cast_fp16, validate_indices = var_4562_validate_indices_0)[name = tensor("op_4562_cast_fp16")]; + tensor concat_84 = const()[name = tensor("concat_84"), val = tensor([1, 0, 0, 0, 0])]; + tensor concat_85 = const()[name = tensor("concat_85"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_21_internal_tensor_assign_2_stride_0 = const()[name = tensor("new_cache_21_internal_tensor_assign_2_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_21_internal_tensor_assign_2_begin_mask_0 = const()[name = tensor("new_cache_21_internal_tensor_assign_2_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_21_internal_tensor_assign_2_end_mask_0 = const()[name = tensor("new_cache_21_internal_tensor_assign_2_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_21_internal_tensor_assign_2_squeeze_mask_0 = const()[name = tensor("new_cache_21_internal_tensor_assign_2_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_67 = const()[name = tensor("shape_67"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_21 = const()[name = tensor("reduce_prod_21"), val = tensor(1048576)]; + tensor range_1d_21_start_0 = const()[name = tensor("range_1d_21_start_0"), val = tensor(0)]; + tensor range_1d_21_step_0 = const()[name = tensor("range_1d_21_step_0"), val = tensor(1)]; + tensor range_1d_21 = range_1d(end = reduce_prod_21, start = range_1d_21_start_0, step = range_1d_21_step_0)[name = tensor("range_1d_21")]; + tensor reshape_105 = reshape(shape = shape_67, x = range_1d_21)[name = tensor("reshape_105")]; + tensor slice_by_index_21 = slice_by_index(begin = concat_84, begin_mask = new_cache_21_internal_tensor_assign_2_begin_mask_0, end = concat_85, end_mask = new_cache_21_internal_tensor_assign_2_end_mask_0, squeeze_mask = new_cache_21_internal_tensor_assign_2_squeeze_mask_0, stride = new_cache_21_internal_tensor_assign_2_stride_0, x = reshape_105)[name = tensor("slice_by_index_21")]; + tensor reshape_106_shape_0 = const()[name = tensor("reshape_106_shape_0"), val = tensor([-1])]; + tensor reshape_106 = reshape(shape = reshape_106_shape_0, x = slice_by_index_21)[name = tensor("reshape_106")]; + tensor reshape_107_shape_0 = const()[name = tensor("reshape_107_shape_0"), val = tensor([-1])]; + tensor reshape_107_cast_fp16 = reshape(shape = reshape_107_shape_0, x = var_4562_cast_fp16)[name = tensor("reshape_107_cast_fp16")]; + tensor reshape_108_shape_0 = const()[name = tensor("reshape_108_shape_0"), val = tensor([-1])]; + tensor reshape_108_cast_fp16 = reshape(shape = reshape_108_shape_0, x = reshape_104_cast_fp16)[name = tensor("reshape_108_cast_fp16")]; + tensor scatter_21_mode_0 = const()[name = tensor("scatter_21_mode_0"), val = tensor("update")]; + tensor scatter_21_axis_0 = const()[name = tensor("scatter_21_axis_0"), val = tensor(0)]; + tensor scatter_21_validate_indices_0 = const()[name = tensor("scatter_21_validate_indices_0"), val = tensor(false)]; + tensor scatter_21_cast_fp16 = scatter(axis = scatter_21_axis_0, data = reshape_108_cast_fp16, indices = reshape_106, mode = scatter_21_mode_0, updates = reshape_107_cast_fp16, validate_indices = scatter_21_validate_indices_0)[name = tensor("scatter_21_cast_fp16")]; + tensor reshape_109_cast_fp16 = reshape(shape = shape_67, x = scatter_21_cast_fp16)[name = tensor("reshape_109_cast_fp16")]; + tensor reshape_109_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("reshape_109_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor keys_61_begin_0 = const()[name = tensor("keys_61_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor keys_61_end_0 = const()[name = tensor("keys_61_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor keys_61_end_mask_0 = const()[name = tensor("keys_61_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor keys_61_squeeze_mask_0 = const()[name = tensor("keys_61_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor keys_61_cast_fp16 = slice_by_index(begin = keys_61_begin_0, end = keys_61_end_0, end_mask = keys_61_end_mask_0, squeeze_mask = keys_61_squeeze_mask_0, x = reshape_109_cast_fp16)[name = tensor("keys_61_cast_fp16")]; + tensor values_61_begin_0 = const()[name = tensor("values_61_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor values_61_end_0 = const()[name = tensor("values_61_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor values_61_end_mask_0 = const()[name = tensor("values_61_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor values_61_squeeze_mask_0 = const()[name = tensor("values_61_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor values_61_cast_fp16 = slice_by_index(begin = values_61_begin_0, end = values_61_end_0, end_mask = values_61_end_mask_0, squeeze_mask = values_61_squeeze_mask_0, x = reshape_109_cast_fp16)[name = tensor("values_61_cast_fp16")]; + tensor var_4574_cast_fp16 = not_equal(x = keys_61_cast_fp16, y = keys_61_cast_fp16)[name = tensor("op_4574_cast_fp16")]; + tensor keys_63_cast_fp16 = select(a = var_470_to_fp16, b = keys_61_cast_fp16, cond = var_4574_cast_fp16)[name = tensor("keys_63_cast_fp16")]; + tensor var_4582_cast_fp16 = not_equal(x = values_61_cast_fp16, y = values_61_cast_fp16)[name = tensor("op_4582_cast_fp16")]; + tensor values_63_cast_fp16 = select(a = var_470_to_fp16, b = values_61_cast_fp16, cond = var_4582_cast_fp16)[name = tensor("values_63_cast_fp16")]; + tensor var_4606 = const()[name = tensor("op_4606"), val = tensor([0, 2, 1, 3])]; + tensor var_4622 = const()[name = tensor("op_4622"), val = tensor([1, 1, 1])]; + tensor var_4623_cast_fp16 = reshape(shape = var_4622, x = position10_to_fp16)[name = tensor("op_4623_cast_fp16")]; + tensor q_positions_21_cast_fp16 = add(x = var_4623_cast_fp16, y = q_offsets_1_promoted_to_fp16)[name = tensor("q_positions_21_cast_fp16")]; + tensor valid_end_21_cast_fp16 = add(x = var_4623_cast_fp16, y = var_536_cast_fp16)[name = tensor("valid_end_21_cast_fp16")]; + tensor valid_mask_21_cast_fp16 = less(x = k_positions_1_promoted_to_fp16, y = valid_end_21_cast_fp16)[name = tensor("valid_mask_21_cast_fp16")]; + tensor causal_mask_21_cast_fp16 = less_equal(x = k_positions_1_promoted_to_fp16, y = q_positions_21_cast_fp16)[name = tensor("causal_mask_21_cast_fp16")]; + tensor var_4651 = logical_and(x = valid_mask_21_cast_fp16, y = causal_mask_21_cast_fp16)[name = tensor("op_4651")]; + tensor attn_mask_21_axes_0 = const()[name = tensor("attn_mask_21_axes_0"), val = tensor([1])]; + tensor attn_mask_21 = expand_dims(axes = attn_mask_21_axes_0, x = var_4651)[name = tensor("attn_mask_21")]; + tensor var_4665_transpose_x_0 = const()[name = tensor("op_4665_transpose_x_0"), val = tensor(false)]; + tensor var_4665_transpose_y_0 = const()[name = tensor("op_4665_transpose_y_0"), val = tensor(false)]; + tensor transpose_89_perm_0 = const()[name = tensor("transpose_89_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_90_perm_0 = const()[name = tensor("transpose_90_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_90 = transpose(perm = transpose_90_perm_0, x = keys_63_cast_fp16)[name = tensor("transpose_164")]; + tensor transpose_89 = transpose(perm = transpose_89_perm_0, x = q_63_cast_fp16)[name = tensor("transpose_165")]; + tensor var_4665_cast_fp16 = matmul(transpose_x = var_4665_transpose_x_0, transpose_y = var_4665_transpose_y_0, x = transpose_89, y = transpose_90)[name = tensor("op_4665_cast_fp16")]; + tensor var_4659_to_fp16 = const()[name = tensor("op_4659_to_fp16"), val = tensor([0x1p-3])]; + tensor attn_weights_61_cast_fp16 = mul(x = var_4665_cast_fp16, y = var_4659_to_fp16)[name = tensor("attn_weights_61_cast_fp16")]; + tensor var_4667 = logical_not(x = attn_mask_21)[name = tensor("op_4667")]; + tensor var_4668_to_fp16 = const()[name = tensor("op_4668_to_fp16"), val = tensor(-inf)]; + tensor attn_weights_63_cast_fp16 = select(a = var_4668_to_fp16, b = attn_weights_61_cast_fp16, cond = var_4667)[name = tensor("attn_weights_63_cast_fp16")]; + tensor var_4670 = const()[name = tensor("op_4670"), val = tensor(-1)]; + tensor attn_weights_65_cast_fp16 = softmax(axis = var_4670, x = attn_weights_63_cast_fp16)[name = tensor("attn_weights_65_cast_fp16")]; + tensor attn_output_21_transpose_x_0 = const()[name = tensor("attn_output_21_transpose_x_0"), val = tensor(false)]; + tensor attn_output_21_transpose_y_0 = const()[name = tensor("attn_output_21_transpose_y_0"), val = tensor(false)]; + tensor values_65_cast_fp16 = transpose(perm = var_4606, x = values_63_cast_fp16)[name = tensor("transpose_166")]; + tensor attn_output_21_cast_fp16 = matmul(transpose_x = attn_output_21_transpose_x_0, transpose_y = attn_output_21_transpose_y_0, x = attn_weights_65_cast_fp16, y = values_65_cast_fp16)[name = tensor("attn_output_21_cast_fp16")]; + tensor var_4678 = const()[name = tensor("op_4678"), val = tensor([0, 2, 1, 3])]; + tensor var_4683 = const()[name = tensor("op_4683"), val = tensor([1, 256, 1024])]; + tensor var_4679_cast_fp16 = transpose(perm = var_4678, x = attn_output_21_cast_fp16)[name = tensor("transpose_163")]; + tensor input_125_cast_fp16 = reshape(shape = var_4683, x = var_4679_cast_fp16)[name = tensor("input_125_cast_fp16")]; + tensor attn10_out_proj_weight_to_fp16 = const()[name = tensor("attn10_out_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(259110528)))]; + tensor linear_41_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = attn10_out_proj_weight_to_fp16, x = input_125_cast_fp16)[name = tensor("linear_41_cast_fp16")]; + tensor var_4692_cast_fp16 = add(x = position10_to_fp16, y = valid_len_to_fp16)[name = tensor("op_4692_cast_fp16")]; + tensor var_4692_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_4692_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor input_127_cast_fp16 = add(x = input_121_cast_fp16, y = linear_41_cast_fp16)[name = tensor("input_127_cast_fp16")]; + tensor input_129_axes_0 = const()[name = tensor("input_129_axes_0"), val = tensor([-1])]; + tensor norm10_2_weight_to_fp16 = const()[name = tensor("norm10_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(261207744)))]; + tensor norm10_2_bias_to_fp16 = const()[name = tensor("norm10_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(261209856)))]; + tensor var_4696_to_fp16 = const()[name = tensor("op_4696_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_129_cast_fp16 = layer_norm(axes = input_129_axes_0, beta = norm10_2_bias_to_fp16, epsilon = var_4696_to_fp16, gamma = norm10_2_weight_to_fp16, x = input_127_cast_fp16)[name = tensor("input_129_cast_fp16")]; + tensor linear10_1_weight_to_fp16 = const()[name = tensor("linear10_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(261211968)))]; + tensor linear_42_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = linear10_1_weight_to_fp16, x = input_129_cast_fp16)[name = tensor("linear_42_cast_fp16")]; + tensor input_131_mode_0 = const()[name = tensor("input_131_mode_0"), val = tensor("EXACT")]; + tensor input_131_cast_fp16 = gelu(mode = input_131_mode_0, x = linear_42_cast_fp16)[name = tensor("input_131_cast_fp16")]; + tensor linear10_2_weight_to_fp16 = const()[name = tensor("linear10_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(269600640)))]; + tensor linear_43_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = linear10_2_weight_to_fp16, x = input_131_cast_fp16)[name = tensor("linear_43_cast_fp16")]; + tensor input_133_cast_fp16 = add(x = input_127_cast_fp16, y = linear_43_cast_fp16)[name = tensor("input_133_cast_fp16")]; + tensor input_135_axes_0 = const()[name = tensor("input_135_axes_0"), val = tensor([-1])]; + tensor norm11_1_weight_to_fp16 = const()[name = tensor("norm11_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(277989312)))]; + tensor norm11_1_bias_to_fp16 = const()[name = tensor("norm11_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(277991424)))]; + tensor var_4713_to_fp16 = const()[name = tensor("op_4713_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_135_cast_fp16 = layer_norm(axes = input_135_axes_0, beta = norm11_1_bias_to_fp16, epsilon = var_4713_to_fp16, gamma = norm11_1_weight_to_fp16, x = input_133_cast_fp16)[name = tensor("input_135_cast_fp16")]; + tensor attn11_in_proj_weight_to_fp16 = const()[name = tensor("attn11_in_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(277993536)))]; + tensor linear_44_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = attn11_in_proj_weight_to_fp16, x = input_135_cast_fp16)[name = tensor("linear_44_cast_fp16")]; + tensor var_4727 = const()[name = tensor("op_4727"), val = tensor([1, 256, 3, 16, 64])]; + tensor qkv_23_cast_fp16 = reshape(shape = var_4727, x = linear_44_cast_fp16)[name = tensor("qkv_23_cast_fp16")]; + tensor q_67_begin_0 = const()[name = tensor("q_67_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor q_67_end_0 = const()[name = tensor("q_67_end_0"), val = tensor([1, 256, 1, 16, 64])]; + tensor q_67_end_mask_0 = const()[name = tensor("q_67_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor q_67_squeeze_mask_0 = const()[name = tensor("q_67_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor q_67_cast_fp16 = slice_by_index(begin = q_67_begin_0, end = q_67_end_0, end_mask = q_67_end_mask_0, squeeze_mask = q_67_squeeze_mask_0, x = qkv_23_cast_fp16)[name = tensor("q_67_cast_fp16")]; + tensor k_45_begin_0 = const()[name = tensor("k_45_begin_0"), val = tensor([0, 0, 1, 0, 0])]; + tensor k_45_end_0 = const()[name = tensor("k_45_end_0"), val = tensor([1, 256, 2, 16, 64])]; + tensor k_45_end_mask_0 = const()[name = tensor("k_45_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor k_45_squeeze_mask_0 = const()[name = tensor("k_45_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor k_45_cast_fp16 = slice_by_index(begin = k_45_begin_0, end = k_45_end_0, end_mask = k_45_end_mask_0, squeeze_mask = k_45_squeeze_mask_0, x = qkv_23_cast_fp16)[name = tensor("k_45_cast_fp16")]; + tensor v_23_begin_0 = const()[name = tensor("v_23_begin_0"), val = tensor([0, 0, 2, 0, 0])]; + tensor v_23_end_0 = const()[name = tensor("v_23_end_0"), val = tensor([1, 256, 3, 16, 64])]; + tensor v_23_end_mask_0 = const()[name = tensor("v_23_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor v_23_squeeze_mask_0 = const()[name = tensor("v_23_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor v_23_cast_fp16 = slice_by_index(begin = v_23_begin_0, end = v_23_end_0, end_mask = v_23_end_mask_0, squeeze_mask = v_23_squeeze_mask_0, x = qkv_23_cast_fp16)[name = tensor("v_23_cast_fp16")]; + tensor var_4785 = const()[name = tensor("op_4785"), val = tensor([1, 1])]; + tensor position11_to_fp16_dtype_0 = const()[name = tensor("position11_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor position11_to_fp16 = cast(dtype = position11_to_fp16_dtype_0, x = position11)[name = tensor("cast_606")]; + tensor var_4786_cast_fp16 = reshape(shape = var_4785, x = position11_to_fp16)[name = tensor("op_4786_cast_fp16")]; + tensor ts_69_cast_fp16 = add(x = ts_1_promoted_to_fp16, y = var_4786_cast_fp16)[name = tensor("ts_69_cast_fp16")]; + tensor var_4793 = const()[name = tensor("op_4793"), val = tensor([1, 256, 1, 1])]; + tensor ts_71_cast_fp16 = reshape(shape = var_4793, x = ts_69_cast_fp16)[name = tensor("ts_71_cast_fp16")]; + tensor var_4800 = const()[name = tensor("op_4800"), val = tensor([1, 256, 16, 32, 2])]; + tensor q_complex_23_cast_fp16 = reshape(shape = var_4800, x = q_67_cast_fp16)[name = tensor("q_complex_23_cast_fp16")]; + tensor var_4807 = const()[name = tensor("op_4807"), val = tensor([1, 256, 16, 32, 2])]; + tensor k_complex_23_cast_fp16 = reshape(shape = var_4807, x = k_45_cast_fp16)[name = tensor("k_complex_23_cast_fp16")]; + tensor var_4811_begin_0 = const()[name = tensor("op_4811_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_4811_end_0 = const()[name = tensor("op_4811_end_0"), val = tensor([1, 256, 16, 32, 1])]; + tensor var_4811_end_mask_0 = const()[name = tensor("op_4811_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_4811_squeeze_mask_0 = const()[name = tensor("op_4811_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_4811_cast_fp16 = slice_by_index(begin = var_4811_begin_0, end = var_4811_end_0, end_mask = var_4811_end_mask_0, squeeze_mask = var_4811_squeeze_mask_0, x = q_complex_23_cast_fp16)[name = tensor("op_4811_cast_fp16")]; + tensor var_4819_begin_0 = const()[name = tensor("op_4819_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_4819_end_0 = const()[name = tensor("op_4819_end_0"), val = tensor([1, 256, 16, 32, 2])]; + tensor var_4819_end_mask_0 = const()[name = tensor("op_4819_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_4819_squeeze_mask_0 = const()[name = tensor("op_4819_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_4819_cast_fp16 = slice_by_index(begin = var_4819_begin_0, end = var_4819_end_0, end_mask = var_4819_end_mask_0, squeeze_mask = var_4819_squeeze_mask_0, x = q_complex_23_cast_fp16)[name = tensor("op_4819_cast_fp16")]; + tensor var_4827_begin_0 = const()[name = tensor("op_4827_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_4827_end_0 = const()[name = tensor("op_4827_end_0"), val = tensor([1, 256, 16, 32, 1])]; + tensor var_4827_end_mask_0 = const()[name = tensor("op_4827_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_4827_squeeze_mask_0 = const()[name = tensor("op_4827_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_4827_cast_fp16 = slice_by_index(begin = var_4827_begin_0, end = var_4827_end_0, end_mask = var_4827_end_mask_0, squeeze_mask = var_4827_squeeze_mask_0, x = k_complex_23_cast_fp16)[name = tensor("op_4827_cast_fp16")]; + tensor var_4835_begin_0 = const()[name = tensor("op_4835_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_4835_end_0 = const()[name = tensor("op_4835_end_0"), val = tensor([1, 256, 16, 32, 2])]; + tensor var_4835_end_mask_0 = const()[name = tensor("op_4835_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_4835_squeeze_mask_0 = const()[name = tensor("op_4835_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_4835_cast_fp16 = slice_by_index(begin = var_4835_begin_0, end = var_4835_end_0, end_mask = var_4835_end_mask_0, squeeze_mask = var_4835_squeeze_mask_0, x = k_complex_23_cast_fp16)[name = tensor("op_4835_cast_fp16")]; + tensor freqs_23_to_fp16 = const()[name = tensor("freqs_23_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(284285056)))]; + tensor var_4841_cast_fp16 = mul(x = freqs_23_to_fp16, y = ts_71_cast_fp16)[name = tensor("op_4841_cast_fp16")]; + tensor rotr_23_cast_fp16 = cos(x = var_4841_cast_fp16)[name = tensor("rotr_23_cast_fp16")]; + tensor roti_23_cast_fp16 = sin(x = var_4841_cast_fp16)[name = tensor("roti_23_cast_fp16")]; + tensor var_4845_cast_fp16 = mul(x = var_4811_cast_fp16, y = rotr_23_cast_fp16)[name = tensor("op_4845_cast_fp16")]; + tensor var_4846_cast_fp16 = mul(x = var_4819_cast_fp16, y = roti_23_cast_fp16)[name = tensor("op_4846_cast_fp16")]; + tensor qor_45_cast_fp16 = sub(x = var_4845_cast_fp16, y = var_4846_cast_fp16)[name = tensor("qor_45_cast_fp16")]; + tensor var_4849_cast_fp16 = mul(x = var_4811_cast_fp16, y = roti_23_cast_fp16)[name = tensor("op_4849_cast_fp16")]; + tensor var_4850_cast_fp16 = mul(x = var_4819_cast_fp16, y = rotr_23_cast_fp16)[name = tensor("op_4850_cast_fp16")]; + tensor qoi_45_cast_fp16 = add(x = var_4849_cast_fp16, y = var_4850_cast_fp16)[name = tensor("qoi_45_cast_fp16")]; + tensor var_4853_cast_fp16 = mul(x = var_4827_cast_fp16, y = rotr_23_cast_fp16)[name = tensor("op_4853_cast_fp16")]; + tensor var_4854_cast_fp16 = mul(x = var_4835_cast_fp16, y = roti_23_cast_fp16)[name = tensor("op_4854_cast_fp16")]; + tensor kor_45_cast_fp16 = sub(x = var_4853_cast_fp16, y = var_4854_cast_fp16)[name = tensor("kor_45_cast_fp16")]; + tensor var_4857_cast_fp16 = mul(x = var_4827_cast_fp16, y = roti_23_cast_fp16)[name = tensor("op_4857_cast_fp16")]; + tensor var_4858_cast_fp16 = mul(x = var_4835_cast_fp16, y = rotr_23_cast_fp16)[name = tensor("op_4858_cast_fp16")]; + tensor koi_45_cast_fp16 = add(x = var_4857_cast_fp16, y = var_4858_cast_fp16)[name = tensor("koi_45_cast_fp16")]; + tensor qo_23_axis_0 = const()[name = tensor("qo_23_axis_0"), val = tensor(-1)]; + tensor qo_23_cast_fp16 = stack(axis = qo_23_axis_0, values = (qor_45_cast_fp16, qoi_45_cast_fp16))[name = tensor("qo_23_cast_fp16")]; + tensor ko_23_axis_0 = const()[name = tensor("ko_23_axis_0"), val = tensor(-1)]; + tensor ko_23_cast_fp16 = stack(axis = ko_23_axis_0, values = (kor_45_cast_fp16, koi_45_cast_fp16))[name = tensor("ko_23_cast_fp16")]; + tensor var_4891 = const()[name = tensor("op_4891"), val = tensor([1, 256, 16, 64])]; + tensor q_69_cast_fp16 = reshape(shape = var_4891, x = qo_23_cast_fp16)[name = tensor("q_69_cast_fp16")]; + tensor var_4897 = const()[name = tensor("op_4897"), val = tensor([1, 256, 16, 64])]; + tensor k_47_cast_fp16 = reshape(shape = var_4897, x = ko_23_cast_fp16)[name = tensor("k_47_cast_fp16")]; + tensor write_indices_float_67_cast_fp16 = add(x = var_4786_cast_fp16, y = var_393_promoted_to_fp16)[name = tensor("write_indices_float_67_cast_fp16")]; + tensor _inversed_4918_y_0_to_fp16 = const()[name = tensor("_inversed_4918_y_0_to_fp16"), val = tensor(0x1p-9)]; + tensor _inversed_4918_cast_fp16 = mul(x = write_indices_float_67_cast_fp16, y = _inversed_4918_y_0_to_fp16)[name = tensor("_inversed_4918_cast_fp16")]; + tensor var_4919_cast_fp16 = floor(x = _inversed_4918_cast_fp16)[name = tensor("op_4919_cast_fp16")]; + tensor var_4920_to_fp16 = const()[name = tensor("op_4920_to_fp16"), val = tensor(0x1p+9)]; + tensor var_4921_cast_fp16 = mul(x = var_4919_cast_fp16, y = var_4920_to_fp16)[name = tensor("op_4921_cast_fp16")]; + tensor write_indices_float_69_cast_fp16 = sub(x = write_indices_float_67_cast_fp16, y = var_4921_cast_fp16)[name = tensor("write_indices_float_69_cast_fp16")]; + tensor var_4939_value_0_to_fp16 = const()[name = tensor("op_4939_value_0_to_fp16"), val = tensor(0x1.ffp+8)]; + tensor var_4939_cast_fp16 = fill_like(ref_tensor = write_indices_float_69_cast_fp16, value = var_4939_value_0_to_fp16)[name = tensor("op_4939_cast_fp16")]; + tensor write_indices_float_71_cast_fp16 = select(a = write_indices_float_69_cast_fp16, b = var_4939_cast_fp16, cond = token_valid_1_cast_fp16)[name = tensor("write_indices_float_71_cast_fp16")]; + tensor var_4945_dtype_0 = const()[name = tensor("op_4945_dtype_0"), val = tensor("int32")]; + tensor var_4950 = const()[name = tensor("op_4950"), val = tensor([1, 256, 1, 1])]; + tensor write_indices_float_71_cast_fp16_to_int32 = cast(dtype = var_4945_dtype_0, x = write_indices_float_71_cast_fp16)[name = tensor("cast_605")]; + tensor var_4951 = reshape(shape = var_4950, x = write_indices_float_71_cast_fp16_to_int32)[name = tensor("op_4951")]; + tensor write_indices_23_reps_0 = const()[name = tensor("write_indices_23_reps_0"), val = tensor([1, 1, 16, 64])]; + tensor write_indices_23 = tile(reps = write_indices_23_reps_0, x = var_4951)[name = tensor("write_indices_23")]; + tensor var_4961_begin_0 = const()[name = tensor("op_4961_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_4961_end_0 = const()[name = tensor("op_4961_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor var_4961_end_mask_0 = const()[name = tensor("op_4961_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_4961_squeeze_mask_0 = const()[name = tensor("op_4961_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor cache11_to_fp16_dtype_0 = const()[name = tensor("cache11_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor cache11_to_fp16 = cast(dtype = cache11_to_fp16_dtype_0, x = cache11)[name = tensor("cast_604")]; + tensor var_4961_cast_fp16 = slice_by_index(begin = var_4961_begin_0, end = var_4961_end_0, end_mask = var_4961_end_mask_0, squeeze_mask = var_4961_squeeze_mask_0, x = cache11_to_fp16)[name = tensor("op_4961_cast_fp16")]; + tensor var_4963_axis_0 = const()[name = tensor("op_4963_axis_0"), val = tensor(1)]; + tensor var_4963_mode_0 = const()[name = tensor("op_4963_mode_0"), val = tensor("update")]; + tensor var_4963_validate_indices_0 = const()[name = tensor("op_4963_validate_indices_0"), val = tensor(false)]; + tensor var_4963_cast_fp16 = scatter_along_axis(axis = var_4963_axis_0, data = var_4961_cast_fp16, indices = write_indices_23, mode = var_4963_mode_0, updates = k_47_cast_fp16, validate_indices = var_4963_validate_indices_0)[name = tensor("op_4963_cast_fp16")]; + tensor concat_90 = const()[name = tensor("concat_90"), val = tensor([0, 0, 0, 0, 0])]; + tensor concat_91 = const()[name = tensor("concat_91"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_23_internal_tensor_assign_1_stride_0 = const()[name = tensor("new_cache_23_internal_tensor_assign_1_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_23_internal_tensor_assign_1_begin_mask_0 = const()[name = tensor("new_cache_23_internal_tensor_assign_1_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_23_internal_tensor_assign_1_end_mask_0 = const()[name = tensor("new_cache_23_internal_tensor_assign_1_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_23_internal_tensor_assign_1_squeeze_mask_0 = const()[name = tensor("new_cache_23_internal_tensor_assign_1_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_68 = const()[name = tensor("shape_68"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_22 = const()[name = tensor("reduce_prod_22"), val = tensor(1048576)]; + tensor range_1d_22_start_0 = const()[name = tensor("range_1d_22_start_0"), val = tensor(0)]; + tensor range_1d_22_step_0 = const()[name = tensor("range_1d_22_step_0"), val = tensor(1)]; + tensor range_1d_22 = range_1d(end = reduce_prod_22, start = range_1d_22_start_0, step = range_1d_22_step_0)[name = tensor("range_1d_22")]; + tensor reshape_110 = reshape(shape = shape_68, x = range_1d_22)[name = tensor("reshape_110")]; + tensor slice_by_index_22 = slice_by_index(begin = concat_90, begin_mask = new_cache_23_internal_tensor_assign_1_begin_mask_0, end = concat_91, end_mask = new_cache_23_internal_tensor_assign_1_end_mask_0, squeeze_mask = new_cache_23_internal_tensor_assign_1_squeeze_mask_0, stride = new_cache_23_internal_tensor_assign_1_stride_0, x = reshape_110)[name = tensor("slice_by_index_22")]; + tensor reshape_111_shape_0 = const()[name = tensor("reshape_111_shape_0"), val = tensor([-1])]; + tensor reshape_111 = reshape(shape = reshape_111_shape_0, x = slice_by_index_22)[name = tensor("reshape_111")]; + tensor reshape_112_shape_0 = const()[name = tensor("reshape_112_shape_0"), val = tensor([-1])]; + tensor reshape_112_cast_fp16 = reshape(shape = reshape_112_shape_0, x = var_4963_cast_fp16)[name = tensor("reshape_112_cast_fp16")]; + tensor reshape_113_shape_0 = const()[name = tensor("reshape_113_shape_0"), val = tensor([-1])]; + tensor reshape_113_cast_fp16 = reshape(shape = reshape_113_shape_0, x = cache11_to_fp16)[name = tensor("reshape_113_cast_fp16")]; + tensor scatter_22_mode_0 = const()[name = tensor("scatter_22_mode_0"), val = tensor("update")]; + tensor scatter_22_axis_0 = const()[name = tensor("scatter_22_axis_0"), val = tensor(0)]; + tensor scatter_22_validate_indices_0 = const()[name = tensor("scatter_22_validate_indices_0"), val = tensor(false)]; + tensor scatter_22_cast_fp16 = scatter(axis = scatter_22_axis_0, data = reshape_113_cast_fp16, indices = reshape_111, mode = scatter_22_mode_0, updates = reshape_112_cast_fp16, validate_indices = scatter_22_validate_indices_0)[name = tensor("scatter_22_cast_fp16")]; + tensor reshape_114_cast_fp16 = reshape(shape = shape_68, x = scatter_22_cast_fp16)[name = tensor("reshape_114_cast_fp16")]; + tensor var_4971_begin_0 = const()[name = tensor("op_4971_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor var_4971_end_0 = const()[name = tensor("op_4971_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor var_4971_end_mask_0 = const()[name = tensor("op_4971_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_4971_squeeze_mask_0 = const()[name = tensor("op_4971_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor var_4971_cast_fp16 = slice_by_index(begin = var_4971_begin_0, end = var_4971_end_0, end_mask = var_4971_end_mask_0, squeeze_mask = var_4971_squeeze_mask_0, x = reshape_114_cast_fp16)[name = tensor("op_4971_cast_fp16")]; + tensor var_4973_axis_0 = const()[name = tensor("op_4973_axis_0"), val = tensor(1)]; + tensor var_4973_mode_0 = const()[name = tensor("op_4973_mode_0"), val = tensor("update")]; + tensor var_4973_validate_indices_0 = const()[name = tensor("op_4973_validate_indices_0"), val = tensor(false)]; + tensor var_4973_cast_fp16 = scatter_along_axis(axis = var_4973_axis_0, data = var_4971_cast_fp16, indices = write_indices_23, mode = var_4973_mode_0, updates = v_23_cast_fp16, validate_indices = var_4973_validate_indices_0)[name = tensor("op_4973_cast_fp16")]; + tensor concat_92 = const()[name = tensor("concat_92"), val = tensor([1, 0, 0, 0, 0])]; + tensor concat_93 = const()[name = tensor("concat_93"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_23_internal_tensor_assign_2_stride_0 = const()[name = tensor("new_cache_23_internal_tensor_assign_2_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_23_internal_tensor_assign_2_begin_mask_0 = const()[name = tensor("new_cache_23_internal_tensor_assign_2_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_23_internal_tensor_assign_2_end_mask_0 = const()[name = tensor("new_cache_23_internal_tensor_assign_2_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_23_internal_tensor_assign_2_squeeze_mask_0 = const()[name = tensor("new_cache_23_internal_tensor_assign_2_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_69 = const()[name = tensor("shape_69"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_23 = const()[name = tensor("reduce_prod_23"), val = tensor(1048576)]; + tensor range_1d_23_start_0 = const()[name = tensor("range_1d_23_start_0"), val = tensor(0)]; + tensor range_1d_23_step_0 = const()[name = tensor("range_1d_23_step_0"), val = tensor(1)]; + tensor range_1d_23 = range_1d(end = reduce_prod_23, start = range_1d_23_start_0, step = range_1d_23_step_0)[name = tensor("range_1d_23")]; + tensor reshape_115 = reshape(shape = shape_69, x = range_1d_23)[name = tensor("reshape_115")]; + tensor slice_by_index_23 = slice_by_index(begin = concat_92, begin_mask = new_cache_23_internal_tensor_assign_2_begin_mask_0, end = concat_93, end_mask = new_cache_23_internal_tensor_assign_2_end_mask_0, squeeze_mask = new_cache_23_internal_tensor_assign_2_squeeze_mask_0, stride = new_cache_23_internal_tensor_assign_2_stride_0, x = reshape_115)[name = tensor("slice_by_index_23")]; + tensor reshape_116_shape_0 = const()[name = tensor("reshape_116_shape_0"), val = tensor([-1])]; + tensor reshape_116 = reshape(shape = reshape_116_shape_0, x = slice_by_index_23)[name = tensor("reshape_116")]; + tensor reshape_117_shape_0 = const()[name = tensor("reshape_117_shape_0"), val = tensor([-1])]; + tensor reshape_117_cast_fp16 = reshape(shape = reshape_117_shape_0, x = var_4973_cast_fp16)[name = tensor("reshape_117_cast_fp16")]; + tensor reshape_118_shape_0 = const()[name = tensor("reshape_118_shape_0"), val = tensor([-1])]; + tensor reshape_118_cast_fp16 = reshape(shape = reshape_118_shape_0, x = reshape_114_cast_fp16)[name = tensor("reshape_118_cast_fp16")]; + tensor scatter_23_mode_0 = const()[name = tensor("scatter_23_mode_0"), val = tensor("update")]; + tensor scatter_23_axis_0 = const()[name = tensor("scatter_23_axis_0"), val = tensor(0)]; + tensor scatter_23_validate_indices_0 = const()[name = tensor("scatter_23_validate_indices_0"), val = tensor(false)]; + tensor scatter_23_cast_fp16 = scatter(axis = scatter_23_axis_0, data = reshape_118_cast_fp16, indices = reshape_116, mode = scatter_23_mode_0, updates = reshape_117_cast_fp16, validate_indices = scatter_23_validate_indices_0)[name = tensor("scatter_23_cast_fp16")]; + tensor reshape_119_cast_fp16 = reshape(shape = shape_69, x = scatter_23_cast_fp16)[name = tensor("reshape_119_cast_fp16")]; + tensor reshape_119_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("reshape_119_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor keys_67_begin_0 = const()[name = tensor("keys_67_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor keys_67_end_0 = const()[name = tensor("keys_67_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor keys_67_end_mask_0 = const()[name = tensor("keys_67_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor keys_67_squeeze_mask_0 = const()[name = tensor("keys_67_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor keys_67_cast_fp16 = slice_by_index(begin = keys_67_begin_0, end = keys_67_end_0, end_mask = keys_67_end_mask_0, squeeze_mask = keys_67_squeeze_mask_0, x = reshape_119_cast_fp16)[name = tensor("keys_67_cast_fp16")]; + tensor values_67_begin_0 = const()[name = tensor("values_67_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor values_67_end_0 = const()[name = tensor("values_67_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor values_67_end_mask_0 = const()[name = tensor("values_67_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor values_67_squeeze_mask_0 = const()[name = tensor("values_67_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor values_67_cast_fp16 = slice_by_index(begin = values_67_begin_0, end = values_67_end_0, end_mask = values_67_end_mask_0, squeeze_mask = values_67_squeeze_mask_0, x = reshape_119_cast_fp16)[name = tensor("values_67_cast_fp16")]; + tensor var_4985_cast_fp16 = not_equal(x = keys_67_cast_fp16, y = keys_67_cast_fp16)[name = tensor("op_4985_cast_fp16")]; + tensor keys_69_cast_fp16 = select(a = var_470_to_fp16, b = keys_67_cast_fp16, cond = var_4985_cast_fp16)[name = tensor("keys_69_cast_fp16")]; + tensor var_4993_cast_fp16 = not_equal(x = values_67_cast_fp16, y = values_67_cast_fp16)[name = tensor("op_4993_cast_fp16")]; + tensor values_69_cast_fp16 = select(a = var_470_to_fp16, b = values_67_cast_fp16, cond = var_4993_cast_fp16)[name = tensor("values_69_cast_fp16")]; + tensor var_5017 = const()[name = tensor("op_5017"), val = tensor([0, 2, 1, 3])]; + tensor var_5033 = const()[name = tensor("op_5033"), val = tensor([1, 1, 1])]; + tensor var_5034_cast_fp16 = reshape(shape = var_5033, x = position11_to_fp16)[name = tensor("op_5034_cast_fp16")]; + tensor q_positions_23_cast_fp16 = add(x = var_5034_cast_fp16, y = q_offsets_1_promoted_to_fp16)[name = tensor("q_positions_23_cast_fp16")]; + tensor valid_end_23_cast_fp16 = add(x = var_5034_cast_fp16, y = var_536_cast_fp16)[name = tensor("valid_end_23_cast_fp16")]; + tensor valid_mask_23_cast_fp16 = less(x = k_positions_1_promoted_to_fp16, y = valid_end_23_cast_fp16)[name = tensor("valid_mask_23_cast_fp16")]; + tensor causal_mask_23_cast_fp16 = less_equal(x = k_positions_1_promoted_to_fp16, y = q_positions_23_cast_fp16)[name = tensor("causal_mask_23_cast_fp16")]; + tensor var_5062 = logical_and(x = valid_mask_23_cast_fp16, y = causal_mask_23_cast_fp16)[name = tensor("op_5062")]; + tensor attn_mask_23_axes_0 = const()[name = tensor("attn_mask_23_axes_0"), val = tensor([1])]; + tensor attn_mask_23 = expand_dims(axes = attn_mask_23_axes_0, x = var_5062)[name = tensor("attn_mask_23")]; + tensor var_5076_transpose_x_0 = const()[name = tensor("op_5076_transpose_x_0"), val = tensor(false)]; + tensor var_5076_transpose_y_0 = const()[name = tensor("op_5076_transpose_y_0"), val = tensor(false)]; + tensor transpose_91_perm_0 = const()[name = tensor("transpose_91_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_92_perm_0 = const()[name = tensor("transpose_92_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_92 = transpose(perm = transpose_92_perm_0, x = keys_69_cast_fp16)[name = tensor("transpose_160")]; + tensor transpose_91 = transpose(perm = transpose_91_perm_0, x = q_69_cast_fp16)[name = tensor("transpose_161")]; + tensor var_5076_cast_fp16 = matmul(transpose_x = var_5076_transpose_x_0, transpose_y = var_5076_transpose_y_0, x = transpose_91, y = transpose_92)[name = tensor("op_5076_cast_fp16")]; + tensor var_5070_to_fp16 = const()[name = tensor("op_5070_to_fp16"), val = tensor([0x1p-3])]; + tensor attn_weights_67_cast_fp16 = mul(x = var_5076_cast_fp16, y = var_5070_to_fp16)[name = tensor("attn_weights_67_cast_fp16")]; + tensor var_5078 = logical_not(x = attn_mask_23)[name = tensor("op_5078")]; + tensor var_5079_to_fp16 = const()[name = tensor("op_5079_to_fp16"), val = tensor(-inf)]; + tensor attn_weights_69_cast_fp16 = select(a = var_5079_to_fp16, b = attn_weights_67_cast_fp16, cond = var_5078)[name = tensor("attn_weights_69_cast_fp16")]; + tensor var_5081 = const()[name = tensor("op_5081"), val = tensor(-1)]; + tensor attn_weights_71_cast_fp16 = softmax(axis = var_5081, x = attn_weights_69_cast_fp16)[name = tensor("attn_weights_71_cast_fp16")]; + tensor attn_output_23_transpose_x_0 = const()[name = tensor("attn_output_23_transpose_x_0"), val = tensor(false)]; + tensor attn_output_23_transpose_y_0 = const()[name = tensor("attn_output_23_transpose_y_0"), val = tensor(false)]; + tensor values_71_cast_fp16 = transpose(perm = var_5017, x = values_69_cast_fp16)[name = tensor("transpose_162")]; + tensor attn_output_23_cast_fp16 = matmul(transpose_x = attn_output_23_transpose_x_0, transpose_y = attn_output_23_transpose_y_0, x = attn_weights_71_cast_fp16, y = values_71_cast_fp16)[name = tensor("attn_output_23_cast_fp16")]; + tensor var_5089 = const()[name = tensor("op_5089"), val = tensor([0, 2, 1, 3])]; + tensor var_5094 = const()[name = tensor("op_5094"), val = tensor([1, 256, 1024])]; + tensor var_5090_cast_fp16 = transpose(perm = var_5089, x = attn_output_23_cast_fp16)[name = tensor("transpose_159")]; + tensor input_137_cast_fp16 = reshape(shape = var_5094, x = var_5090_cast_fp16)[name = tensor("input_137_cast_fp16")]; + tensor attn11_out_proj_weight_to_fp16 = const()[name = tensor("attn11_out_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(284285184)))]; + tensor linear_45_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = attn11_out_proj_weight_to_fp16, x = input_137_cast_fp16)[name = tensor("linear_45_cast_fp16")]; + tensor var_5103_cast_fp16 = add(x = position11_to_fp16, y = valid_len_to_fp16)[name = tensor("op_5103_cast_fp16")]; + tensor var_5103_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_5103_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor input_139_cast_fp16 = add(x = input_133_cast_fp16, y = linear_45_cast_fp16)[name = tensor("input_139_cast_fp16")]; + tensor input_141_axes_0 = const()[name = tensor("input_141_axes_0"), val = tensor([-1])]; + tensor norm11_2_weight_to_fp16 = const()[name = tensor("norm11_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(286382400)))]; + tensor norm11_2_bias_to_fp16 = const()[name = tensor("norm11_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(286384512)))]; + tensor var_5107_to_fp16 = const()[name = tensor("op_5107_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_141_cast_fp16 = layer_norm(axes = input_141_axes_0, beta = norm11_2_bias_to_fp16, epsilon = var_5107_to_fp16, gamma = norm11_2_weight_to_fp16, x = input_139_cast_fp16)[name = tensor("input_141_cast_fp16")]; + tensor linear11_1_weight_to_fp16 = const()[name = tensor("linear11_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(286386624)))]; + tensor linear_46_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = linear11_1_weight_to_fp16, x = input_141_cast_fp16)[name = tensor("linear_46_cast_fp16")]; + tensor input_143_mode_0 = const()[name = tensor("input_143_mode_0"), val = tensor("EXACT")]; + tensor input_143_cast_fp16 = gelu(mode = input_143_mode_0, x = linear_46_cast_fp16)[name = tensor("input_143_cast_fp16")]; + tensor linear11_2_weight_to_fp16 = const()[name = tensor("linear11_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(294775296)))]; + tensor linear_47_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = linear11_2_weight_to_fp16, x = input_143_cast_fp16)[name = tensor("linear_47_cast_fp16")]; + tensor input_145_cast_fp16 = add(x = input_139_cast_fp16, y = linear_47_cast_fp16)[name = tensor("input_145_cast_fp16")]; + tensor input_147_axes_0 = const()[name = tensor("input_147_axes_0"), val = tensor([-1])]; + tensor norm12_1_weight_to_fp16 = const()[name = tensor("norm12_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(303163968)))]; + tensor norm12_1_bias_to_fp16 = const()[name = tensor("norm12_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(303166080)))]; + tensor var_5124_to_fp16 = const()[name = tensor("op_5124_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_147_cast_fp16 = layer_norm(axes = input_147_axes_0, beta = norm12_1_bias_to_fp16, epsilon = var_5124_to_fp16, gamma = norm12_1_weight_to_fp16, x = input_145_cast_fp16)[name = tensor("input_147_cast_fp16")]; + tensor attn12_in_proj_weight_to_fp16 = const()[name = tensor("attn12_in_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(303168192)))]; + tensor linear_48_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = attn12_in_proj_weight_to_fp16, x = input_147_cast_fp16)[name = tensor("linear_48_cast_fp16")]; + tensor var_5138 = const()[name = tensor("op_5138"), val = tensor([1, 256, 3, 16, 64])]; + tensor qkv_25_cast_fp16 = reshape(shape = var_5138, x = linear_48_cast_fp16)[name = tensor("qkv_25_cast_fp16")]; + tensor q_73_begin_0 = const()[name = tensor("q_73_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor q_73_end_0 = const()[name = tensor("q_73_end_0"), val = tensor([1, 256, 1, 16, 64])]; + tensor q_73_end_mask_0 = const()[name = tensor("q_73_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor q_73_squeeze_mask_0 = const()[name = tensor("q_73_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor q_73_cast_fp16 = slice_by_index(begin = q_73_begin_0, end = q_73_end_0, end_mask = q_73_end_mask_0, squeeze_mask = q_73_squeeze_mask_0, x = qkv_25_cast_fp16)[name = tensor("q_73_cast_fp16")]; + tensor k_49_begin_0 = const()[name = tensor("k_49_begin_0"), val = tensor([0, 0, 1, 0, 0])]; + tensor k_49_end_0 = const()[name = tensor("k_49_end_0"), val = tensor([1, 256, 2, 16, 64])]; + tensor k_49_end_mask_0 = const()[name = tensor("k_49_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor k_49_squeeze_mask_0 = const()[name = tensor("k_49_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor k_49_cast_fp16 = slice_by_index(begin = k_49_begin_0, end = k_49_end_0, end_mask = k_49_end_mask_0, squeeze_mask = k_49_squeeze_mask_0, x = qkv_25_cast_fp16)[name = tensor("k_49_cast_fp16")]; + tensor v_25_begin_0 = const()[name = tensor("v_25_begin_0"), val = tensor([0, 0, 2, 0, 0])]; + tensor v_25_end_0 = const()[name = tensor("v_25_end_0"), val = tensor([1, 256, 3, 16, 64])]; + tensor v_25_end_mask_0 = const()[name = tensor("v_25_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor v_25_squeeze_mask_0 = const()[name = tensor("v_25_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor v_25_cast_fp16 = slice_by_index(begin = v_25_begin_0, end = v_25_end_0, end_mask = v_25_end_mask_0, squeeze_mask = v_25_squeeze_mask_0, x = qkv_25_cast_fp16)[name = tensor("v_25_cast_fp16")]; + tensor var_5196 = const()[name = tensor("op_5196"), val = tensor([1, 1])]; + tensor position12_to_fp16_dtype_0 = const()[name = tensor("position12_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor position12_to_fp16 = cast(dtype = position12_to_fp16_dtype_0, x = position12)[name = tensor("cast_601")]; + tensor var_5197_cast_fp16 = reshape(shape = var_5196, x = position12_to_fp16)[name = tensor("op_5197_cast_fp16")]; + tensor ts_75_cast_fp16 = add(x = ts_1_promoted_to_fp16, y = var_5197_cast_fp16)[name = tensor("ts_75_cast_fp16")]; + tensor var_5204 = const()[name = tensor("op_5204"), val = tensor([1, 256, 1, 1])]; + tensor ts_77_cast_fp16 = reshape(shape = var_5204, x = ts_75_cast_fp16)[name = tensor("ts_77_cast_fp16")]; + tensor var_5211 = const()[name = tensor("op_5211"), val = tensor([1, 256, 16, 32, 2])]; + tensor q_complex_25_cast_fp16 = reshape(shape = var_5211, x = q_73_cast_fp16)[name = tensor("q_complex_25_cast_fp16")]; + tensor var_5218 = const()[name = tensor("op_5218"), val = tensor([1, 256, 16, 32, 2])]; + tensor k_complex_25_cast_fp16 = reshape(shape = var_5218, x = k_49_cast_fp16)[name = tensor("k_complex_25_cast_fp16")]; + tensor var_5222_begin_0 = const()[name = tensor("op_5222_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_5222_end_0 = const()[name = tensor("op_5222_end_0"), val = tensor([1, 256, 16, 32, 1])]; + tensor var_5222_end_mask_0 = const()[name = tensor("op_5222_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_5222_squeeze_mask_0 = const()[name = tensor("op_5222_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_5222_cast_fp16 = slice_by_index(begin = var_5222_begin_0, end = var_5222_end_0, end_mask = var_5222_end_mask_0, squeeze_mask = var_5222_squeeze_mask_0, x = q_complex_25_cast_fp16)[name = tensor("op_5222_cast_fp16")]; + tensor var_5230_begin_0 = const()[name = tensor("op_5230_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_5230_end_0 = const()[name = tensor("op_5230_end_0"), val = tensor([1, 256, 16, 32, 2])]; + tensor var_5230_end_mask_0 = const()[name = tensor("op_5230_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_5230_squeeze_mask_0 = const()[name = tensor("op_5230_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_5230_cast_fp16 = slice_by_index(begin = var_5230_begin_0, end = var_5230_end_0, end_mask = var_5230_end_mask_0, squeeze_mask = var_5230_squeeze_mask_0, x = q_complex_25_cast_fp16)[name = tensor("op_5230_cast_fp16")]; + tensor var_5238_begin_0 = const()[name = tensor("op_5238_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_5238_end_0 = const()[name = tensor("op_5238_end_0"), val = tensor([1, 256, 16, 32, 1])]; + tensor var_5238_end_mask_0 = const()[name = tensor("op_5238_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_5238_squeeze_mask_0 = const()[name = tensor("op_5238_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_5238_cast_fp16 = slice_by_index(begin = var_5238_begin_0, end = var_5238_end_0, end_mask = var_5238_end_mask_0, squeeze_mask = var_5238_squeeze_mask_0, x = k_complex_25_cast_fp16)[name = tensor("op_5238_cast_fp16")]; + tensor var_5246_begin_0 = const()[name = tensor("op_5246_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_5246_end_0 = const()[name = tensor("op_5246_end_0"), val = tensor([1, 256, 16, 32, 2])]; + tensor var_5246_end_mask_0 = const()[name = tensor("op_5246_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_5246_squeeze_mask_0 = const()[name = tensor("op_5246_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_5246_cast_fp16 = slice_by_index(begin = var_5246_begin_0, end = var_5246_end_0, end_mask = var_5246_end_mask_0, squeeze_mask = var_5246_squeeze_mask_0, x = k_complex_25_cast_fp16)[name = tensor("op_5246_cast_fp16")]; + tensor freqs_25_to_fp16 = const()[name = tensor("freqs_25_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(309459712)))]; + tensor var_5252_cast_fp16 = mul(x = freqs_25_to_fp16, y = ts_77_cast_fp16)[name = tensor("op_5252_cast_fp16")]; + tensor rotr_25_cast_fp16 = cos(x = var_5252_cast_fp16)[name = tensor("rotr_25_cast_fp16")]; + tensor roti_25_cast_fp16 = sin(x = var_5252_cast_fp16)[name = tensor("roti_25_cast_fp16")]; + tensor var_5256_cast_fp16 = mul(x = var_5222_cast_fp16, y = rotr_25_cast_fp16)[name = tensor("op_5256_cast_fp16")]; + tensor var_5257_cast_fp16 = mul(x = var_5230_cast_fp16, y = roti_25_cast_fp16)[name = tensor("op_5257_cast_fp16")]; + tensor qor_49_cast_fp16 = sub(x = var_5256_cast_fp16, y = var_5257_cast_fp16)[name = tensor("qor_49_cast_fp16")]; + tensor var_5260_cast_fp16 = mul(x = var_5222_cast_fp16, y = roti_25_cast_fp16)[name = tensor("op_5260_cast_fp16")]; + tensor var_5261_cast_fp16 = mul(x = var_5230_cast_fp16, y = rotr_25_cast_fp16)[name = tensor("op_5261_cast_fp16")]; + tensor qoi_49_cast_fp16 = add(x = var_5260_cast_fp16, y = var_5261_cast_fp16)[name = tensor("qoi_49_cast_fp16")]; + tensor var_5264_cast_fp16 = mul(x = var_5238_cast_fp16, y = rotr_25_cast_fp16)[name = tensor("op_5264_cast_fp16")]; + tensor var_5265_cast_fp16 = mul(x = var_5246_cast_fp16, y = roti_25_cast_fp16)[name = tensor("op_5265_cast_fp16")]; + tensor kor_49_cast_fp16 = sub(x = var_5264_cast_fp16, y = var_5265_cast_fp16)[name = tensor("kor_49_cast_fp16")]; + tensor var_5268_cast_fp16 = mul(x = var_5238_cast_fp16, y = roti_25_cast_fp16)[name = tensor("op_5268_cast_fp16")]; + tensor var_5269_cast_fp16 = mul(x = var_5246_cast_fp16, y = rotr_25_cast_fp16)[name = tensor("op_5269_cast_fp16")]; + tensor koi_49_cast_fp16 = add(x = var_5268_cast_fp16, y = var_5269_cast_fp16)[name = tensor("koi_49_cast_fp16")]; + tensor qo_25_axis_0 = const()[name = tensor("qo_25_axis_0"), val = tensor(-1)]; + tensor qo_25_cast_fp16 = stack(axis = qo_25_axis_0, values = (qor_49_cast_fp16, qoi_49_cast_fp16))[name = tensor("qo_25_cast_fp16")]; + tensor ko_25_axis_0 = const()[name = tensor("ko_25_axis_0"), val = tensor(-1)]; + tensor ko_25_cast_fp16 = stack(axis = ko_25_axis_0, values = (kor_49_cast_fp16, koi_49_cast_fp16))[name = tensor("ko_25_cast_fp16")]; + tensor var_5302 = const()[name = tensor("op_5302"), val = tensor([1, 256, 16, 64])]; + tensor q_75_cast_fp16 = reshape(shape = var_5302, x = qo_25_cast_fp16)[name = tensor("q_75_cast_fp16")]; + tensor var_5308 = const()[name = tensor("op_5308"), val = tensor([1, 256, 16, 64])]; + tensor k_51_cast_fp16 = reshape(shape = var_5308, x = ko_25_cast_fp16)[name = tensor("k_51_cast_fp16")]; + tensor write_indices_float_73_cast_fp16 = add(x = var_5197_cast_fp16, y = var_393_promoted_to_fp16)[name = tensor("write_indices_float_73_cast_fp16")]; + tensor _inversed_5329_y_0_to_fp16 = const()[name = tensor("_inversed_5329_y_0_to_fp16"), val = tensor(0x1p-9)]; + tensor _inversed_5329_cast_fp16 = mul(x = write_indices_float_73_cast_fp16, y = _inversed_5329_y_0_to_fp16)[name = tensor("_inversed_5329_cast_fp16")]; + tensor var_5330_cast_fp16 = floor(x = _inversed_5329_cast_fp16)[name = tensor("op_5330_cast_fp16")]; + tensor var_5331_to_fp16 = const()[name = tensor("op_5331_to_fp16"), val = tensor(0x1p+9)]; + tensor var_5332_cast_fp16 = mul(x = var_5330_cast_fp16, y = var_5331_to_fp16)[name = tensor("op_5332_cast_fp16")]; + tensor write_indices_float_75_cast_fp16 = sub(x = write_indices_float_73_cast_fp16, y = var_5332_cast_fp16)[name = tensor("write_indices_float_75_cast_fp16")]; + tensor var_5350_value_0_to_fp16 = const()[name = tensor("op_5350_value_0_to_fp16"), val = tensor(0x1.ffp+8)]; + tensor var_5350_cast_fp16 = fill_like(ref_tensor = write_indices_float_75_cast_fp16, value = var_5350_value_0_to_fp16)[name = tensor("op_5350_cast_fp16")]; + tensor write_indices_float_77_cast_fp16 = select(a = write_indices_float_75_cast_fp16, b = var_5350_cast_fp16, cond = token_valid_1_cast_fp16)[name = tensor("write_indices_float_77_cast_fp16")]; + tensor var_5356_dtype_0 = const()[name = tensor("op_5356_dtype_0"), val = tensor("int32")]; + tensor var_5361 = const()[name = tensor("op_5361"), val = tensor([1, 256, 1, 1])]; + tensor write_indices_float_77_cast_fp16_to_int32 = cast(dtype = var_5356_dtype_0, x = write_indices_float_77_cast_fp16)[name = tensor("cast_600")]; + tensor var_5362 = reshape(shape = var_5361, x = write_indices_float_77_cast_fp16_to_int32)[name = tensor("op_5362")]; + tensor write_indices_25_reps_0 = const()[name = tensor("write_indices_25_reps_0"), val = tensor([1, 1, 16, 64])]; + tensor write_indices_25 = tile(reps = write_indices_25_reps_0, x = var_5362)[name = tensor("write_indices_25")]; + tensor var_5372_begin_0 = const()[name = tensor("op_5372_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_5372_end_0 = const()[name = tensor("op_5372_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor var_5372_end_mask_0 = const()[name = tensor("op_5372_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_5372_squeeze_mask_0 = const()[name = tensor("op_5372_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor cache12_to_fp16_dtype_0 = const()[name = tensor("cache12_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor cache12_to_fp16 = cast(dtype = cache12_to_fp16_dtype_0, x = cache12)[name = tensor("cast_599")]; + tensor var_5372_cast_fp16 = slice_by_index(begin = var_5372_begin_0, end = var_5372_end_0, end_mask = var_5372_end_mask_0, squeeze_mask = var_5372_squeeze_mask_0, x = cache12_to_fp16)[name = tensor("op_5372_cast_fp16")]; + tensor var_5374_axis_0 = const()[name = tensor("op_5374_axis_0"), val = tensor(1)]; + tensor var_5374_mode_0 = const()[name = tensor("op_5374_mode_0"), val = tensor("update")]; + tensor var_5374_validate_indices_0 = const()[name = tensor("op_5374_validate_indices_0"), val = tensor(false)]; + tensor var_5374_cast_fp16 = scatter_along_axis(axis = var_5374_axis_0, data = var_5372_cast_fp16, indices = write_indices_25, mode = var_5374_mode_0, updates = k_51_cast_fp16, validate_indices = var_5374_validate_indices_0)[name = tensor("op_5374_cast_fp16")]; + tensor concat_98 = const()[name = tensor("concat_98"), val = tensor([0, 0, 0, 0, 0])]; + tensor concat_99 = const()[name = tensor("concat_99"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_25_internal_tensor_assign_1_stride_0 = const()[name = tensor("new_cache_25_internal_tensor_assign_1_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_25_internal_tensor_assign_1_begin_mask_0 = const()[name = tensor("new_cache_25_internal_tensor_assign_1_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_25_internal_tensor_assign_1_end_mask_0 = const()[name = tensor("new_cache_25_internal_tensor_assign_1_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_25_internal_tensor_assign_1_squeeze_mask_0 = const()[name = tensor("new_cache_25_internal_tensor_assign_1_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_70 = const()[name = tensor("shape_70"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_24 = const()[name = tensor("reduce_prod_24"), val = tensor(1048576)]; + tensor range_1d_24_start_0 = const()[name = tensor("range_1d_24_start_0"), val = tensor(0)]; + tensor range_1d_24_step_0 = const()[name = tensor("range_1d_24_step_0"), val = tensor(1)]; + tensor range_1d_24 = range_1d(end = reduce_prod_24, start = range_1d_24_start_0, step = range_1d_24_step_0)[name = tensor("range_1d_24")]; + tensor reshape_120 = reshape(shape = shape_70, x = range_1d_24)[name = tensor("reshape_120")]; + tensor slice_by_index_24 = slice_by_index(begin = concat_98, begin_mask = new_cache_25_internal_tensor_assign_1_begin_mask_0, end = concat_99, end_mask = new_cache_25_internal_tensor_assign_1_end_mask_0, squeeze_mask = new_cache_25_internal_tensor_assign_1_squeeze_mask_0, stride = new_cache_25_internal_tensor_assign_1_stride_0, x = reshape_120)[name = tensor("slice_by_index_24")]; + tensor reshape_121_shape_0 = const()[name = tensor("reshape_121_shape_0"), val = tensor([-1])]; + tensor reshape_121 = reshape(shape = reshape_121_shape_0, x = slice_by_index_24)[name = tensor("reshape_121")]; + tensor reshape_122_shape_0 = const()[name = tensor("reshape_122_shape_0"), val = tensor([-1])]; + tensor reshape_122_cast_fp16 = reshape(shape = reshape_122_shape_0, x = var_5374_cast_fp16)[name = tensor("reshape_122_cast_fp16")]; + tensor reshape_123_shape_0 = const()[name = tensor("reshape_123_shape_0"), val = tensor([-1])]; + tensor reshape_123_cast_fp16 = reshape(shape = reshape_123_shape_0, x = cache12_to_fp16)[name = tensor("reshape_123_cast_fp16")]; + tensor scatter_24_mode_0 = const()[name = tensor("scatter_24_mode_0"), val = tensor("update")]; + tensor scatter_24_axis_0 = const()[name = tensor("scatter_24_axis_0"), val = tensor(0)]; + tensor scatter_24_validate_indices_0 = const()[name = tensor("scatter_24_validate_indices_0"), val = tensor(false)]; + tensor scatter_24_cast_fp16 = scatter(axis = scatter_24_axis_0, data = reshape_123_cast_fp16, indices = reshape_121, mode = scatter_24_mode_0, updates = reshape_122_cast_fp16, validate_indices = scatter_24_validate_indices_0)[name = tensor("scatter_24_cast_fp16")]; + tensor reshape_124_cast_fp16 = reshape(shape = shape_70, x = scatter_24_cast_fp16)[name = tensor("reshape_124_cast_fp16")]; + tensor var_5382_begin_0 = const()[name = tensor("op_5382_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor var_5382_end_0 = const()[name = tensor("op_5382_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor var_5382_end_mask_0 = const()[name = tensor("op_5382_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_5382_squeeze_mask_0 = const()[name = tensor("op_5382_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor var_5382_cast_fp16 = slice_by_index(begin = var_5382_begin_0, end = var_5382_end_0, end_mask = var_5382_end_mask_0, squeeze_mask = var_5382_squeeze_mask_0, x = reshape_124_cast_fp16)[name = tensor("op_5382_cast_fp16")]; + tensor var_5384_axis_0 = const()[name = tensor("op_5384_axis_0"), val = tensor(1)]; + tensor var_5384_mode_0 = const()[name = tensor("op_5384_mode_0"), val = tensor("update")]; + tensor var_5384_validate_indices_0 = const()[name = tensor("op_5384_validate_indices_0"), val = tensor(false)]; + tensor var_5384_cast_fp16 = scatter_along_axis(axis = var_5384_axis_0, data = var_5382_cast_fp16, indices = write_indices_25, mode = var_5384_mode_0, updates = v_25_cast_fp16, validate_indices = var_5384_validate_indices_0)[name = tensor("op_5384_cast_fp16")]; + tensor concat_100 = const()[name = tensor("concat_100"), val = tensor([1, 0, 0, 0, 0])]; + tensor concat_101 = const()[name = tensor("concat_101"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_25_internal_tensor_assign_2_stride_0 = const()[name = tensor("new_cache_25_internal_tensor_assign_2_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_25_internal_tensor_assign_2_begin_mask_0 = const()[name = tensor("new_cache_25_internal_tensor_assign_2_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_25_internal_tensor_assign_2_end_mask_0 = const()[name = tensor("new_cache_25_internal_tensor_assign_2_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_25_internal_tensor_assign_2_squeeze_mask_0 = const()[name = tensor("new_cache_25_internal_tensor_assign_2_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_71 = const()[name = tensor("shape_71"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_25 = const()[name = tensor("reduce_prod_25"), val = tensor(1048576)]; + tensor range_1d_25_start_0 = const()[name = tensor("range_1d_25_start_0"), val = tensor(0)]; + tensor range_1d_25_step_0 = const()[name = tensor("range_1d_25_step_0"), val = tensor(1)]; + tensor range_1d_25 = range_1d(end = reduce_prod_25, start = range_1d_25_start_0, step = range_1d_25_step_0)[name = tensor("range_1d_25")]; + tensor reshape_125 = reshape(shape = shape_71, x = range_1d_25)[name = tensor("reshape_125")]; + tensor slice_by_index_25 = slice_by_index(begin = concat_100, begin_mask = new_cache_25_internal_tensor_assign_2_begin_mask_0, end = concat_101, end_mask = new_cache_25_internal_tensor_assign_2_end_mask_0, squeeze_mask = new_cache_25_internal_tensor_assign_2_squeeze_mask_0, stride = new_cache_25_internal_tensor_assign_2_stride_0, x = reshape_125)[name = tensor("slice_by_index_25")]; + tensor reshape_126_shape_0 = const()[name = tensor("reshape_126_shape_0"), val = tensor([-1])]; + tensor reshape_126 = reshape(shape = reshape_126_shape_0, x = slice_by_index_25)[name = tensor("reshape_126")]; + tensor reshape_127_shape_0 = const()[name = tensor("reshape_127_shape_0"), val = tensor([-1])]; + tensor reshape_127_cast_fp16 = reshape(shape = reshape_127_shape_0, x = var_5384_cast_fp16)[name = tensor("reshape_127_cast_fp16")]; + tensor reshape_128_shape_0 = const()[name = tensor("reshape_128_shape_0"), val = tensor([-1])]; + tensor reshape_128_cast_fp16 = reshape(shape = reshape_128_shape_0, x = reshape_124_cast_fp16)[name = tensor("reshape_128_cast_fp16")]; + tensor scatter_25_mode_0 = const()[name = tensor("scatter_25_mode_0"), val = tensor("update")]; + tensor scatter_25_axis_0 = const()[name = tensor("scatter_25_axis_0"), val = tensor(0)]; + tensor scatter_25_validate_indices_0 = const()[name = tensor("scatter_25_validate_indices_0"), val = tensor(false)]; + tensor scatter_25_cast_fp16 = scatter(axis = scatter_25_axis_0, data = reshape_128_cast_fp16, indices = reshape_126, mode = scatter_25_mode_0, updates = reshape_127_cast_fp16, validate_indices = scatter_25_validate_indices_0)[name = tensor("scatter_25_cast_fp16")]; + tensor reshape_129_cast_fp16 = reshape(shape = shape_71, x = scatter_25_cast_fp16)[name = tensor("reshape_129_cast_fp16")]; + tensor reshape_129_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("reshape_129_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor keys_73_begin_0 = const()[name = tensor("keys_73_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor keys_73_end_0 = const()[name = tensor("keys_73_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor keys_73_end_mask_0 = const()[name = tensor("keys_73_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor keys_73_squeeze_mask_0 = const()[name = tensor("keys_73_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor keys_73_cast_fp16 = slice_by_index(begin = keys_73_begin_0, end = keys_73_end_0, end_mask = keys_73_end_mask_0, squeeze_mask = keys_73_squeeze_mask_0, x = reshape_129_cast_fp16)[name = tensor("keys_73_cast_fp16")]; + tensor values_73_begin_0 = const()[name = tensor("values_73_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor values_73_end_0 = const()[name = tensor("values_73_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor values_73_end_mask_0 = const()[name = tensor("values_73_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor values_73_squeeze_mask_0 = const()[name = tensor("values_73_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor values_73_cast_fp16 = slice_by_index(begin = values_73_begin_0, end = values_73_end_0, end_mask = values_73_end_mask_0, squeeze_mask = values_73_squeeze_mask_0, x = reshape_129_cast_fp16)[name = tensor("values_73_cast_fp16")]; + tensor var_5396_cast_fp16 = not_equal(x = keys_73_cast_fp16, y = keys_73_cast_fp16)[name = tensor("op_5396_cast_fp16")]; + tensor keys_75_cast_fp16 = select(a = var_470_to_fp16, b = keys_73_cast_fp16, cond = var_5396_cast_fp16)[name = tensor("keys_75_cast_fp16")]; + tensor var_5404_cast_fp16 = not_equal(x = values_73_cast_fp16, y = values_73_cast_fp16)[name = tensor("op_5404_cast_fp16")]; + tensor values_75_cast_fp16 = select(a = var_470_to_fp16, b = values_73_cast_fp16, cond = var_5404_cast_fp16)[name = tensor("values_75_cast_fp16")]; + tensor var_5428 = const()[name = tensor("op_5428"), val = tensor([0, 2, 1, 3])]; + tensor var_5444 = const()[name = tensor("op_5444"), val = tensor([1, 1, 1])]; + tensor var_5445_cast_fp16 = reshape(shape = var_5444, x = position12_to_fp16)[name = tensor("op_5445_cast_fp16")]; + tensor q_positions_25_cast_fp16 = add(x = var_5445_cast_fp16, y = q_offsets_1_promoted_to_fp16)[name = tensor("q_positions_25_cast_fp16")]; + tensor valid_end_25_cast_fp16 = add(x = var_5445_cast_fp16, y = var_536_cast_fp16)[name = tensor("valid_end_25_cast_fp16")]; + tensor valid_mask_25_cast_fp16 = less(x = k_positions_1_promoted_to_fp16, y = valid_end_25_cast_fp16)[name = tensor("valid_mask_25_cast_fp16")]; + tensor causal_mask_25_cast_fp16 = less_equal(x = k_positions_1_promoted_to_fp16, y = q_positions_25_cast_fp16)[name = tensor("causal_mask_25_cast_fp16")]; + tensor var_5473 = logical_and(x = valid_mask_25_cast_fp16, y = causal_mask_25_cast_fp16)[name = tensor("op_5473")]; + tensor attn_mask_25_axes_0 = const()[name = tensor("attn_mask_25_axes_0"), val = tensor([1])]; + tensor attn_mask_25 = expand_dims(axes = attn_mask_25_axes_0, x = var_5473)[name = tensor("attn_mask_25")]; + tensor var_5487_transpose_x_0 = const()[name = tensor("op_5487_transpose_x_0"), val = tensor(false)]; + tensor var_5487_transpose_y_0 = const()[name = tensor("op_5487_transpose_y_0"), val = tensor(false)]; + tensor transpose_93_perm_0 = const()[name = tensor("transpose_93_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_94_perm_0 = const()[name = tensor("transpose_94_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_94 = transpose(perm = transpose_94_perm_0, x = keys_75_cast_fp16)[name = tensor("transpose_156")]; + tensor transpose_93 = transpose(perm = transpose_93_perm_0, x = q_75_cast_fp16)[name = tensor("transpose_157")]; + tensor var_5487_cast_fp16 = matmul(transpose_x = var_5487_transpose_x_0, transpose_y = var_5487_transpose_y_0, x = transpose_93, y = transpose_94)[name = tensor("op_5487_cast_fp16")]; + tensor var_5481_to_fp16 = const()[name = tensor("op_5481_to_fp16"), val = tensor([0x1p-3])]; + tensor attn_weights_73_cast_fp16 = mul(x = var_5487_cast_fp16, y = var_5481_to_fp16)[name = tensor("attn_weights_73_cast_fp16")]; + tensor var_5489 = logical_not(x = attn_mask_25)[name = tensor("op_5489")]; + tensor var_5490_to_fp16 = const()[name = tensor("op_5490_to_fp16"), val = tensor(-inf)]; + tensor attn_weights_75_cast_fp16 = select(a = var_5490_to_fp16, b = attn_weights_73_cast_fp16, cond = var_5489)[name = tensor("attn_weights_75_cast_fp16")]; + tensor var_5492 = const()[name = tensor("op_5492"), val = tensor(-1)]; + tensor attn_weights_77_cast_fp16 = softmax(axis = var_5492, x = attn_weights_75_cast_fp16)[name = tensor("attn_weights_77_cast_fp16")]; + tensor attn_output_25_transpose_x_0 = const()[name = tensor("attn_output_25_transpose_x_0"), val = tensor(false)]; + tensor attn_output_25_transpose_y_0 = const()[name = tensor("attn_output_25_transpose_y_0"), val = tensor(false)]; + tensor values_77_cast_fp16 = transpose(perm = var_5428, x = values_75_cast_fp16)[name = tensor("transpose_158")]; + tensor attn_output_25_cast_fp16 = matmul(transpose_x = attn_output_25_transpose_x_0, transpose_y = attn_output_25_transpose_y_0, x = attn_weights_77_cast_fp16, y = values_77_cast_fp16)[name = tensor("attn_output_25_cast_fp16")]; + tensor var_5500 = const()[name = tensor("op_5500"), val = tensor([0, 2, 1, 3])]; + tensor var_5505 = const()[name = tensor("op_5505"), val = tensor([1, 256, 1024])]; + tensor var_5501_cast_fp16 = transpose(perm = var_5500, x = attn_output_25_cast_fp16)[name = tensor("transpose_155")]; + tensor input_149_cast_fp16 = reshape(shape = var_5505, x = var_5501_cast_fp16)[name = tensor("input_149_cast_fp16")]; + tensor attn12_out_proj_weight_to_fp16 = const()[name = tensor("attn12_out_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(309459840)))]; + tensor linear_49_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = attn12_out_proj_weight_to_fp16, x = input_149_cast_fp16)[name = tensor("linear_49_cast_fp16")]; + tensor var_5514_cast_fp16 = add(x = position12_to_fp16, y = valid_len_to_fp16)[name = tensor("op_5514_cast_fp16")]; + tensor var_5514_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_5514_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor input_151_cast_fp16 = add(x = input_145_cast_fp16, y = linear_49_cast_fp16)[name = tensor("input_151_cast_fp16")]; + tensor input_153_axes_0 = const()[name = tensor("input_153_axes_0"), val = tensor([-1])]; + tensor norm12_2_weight_to_fp16 = const()[name = tensor("norm12_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(311557056)))]; + tensor norm12_2_bias_to_fp16 = const()[name = tensor("norm12_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(311559168)))]; + tensor var_5518_to_fp16 = const()[name = tensor("op_5518_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_153_cast_fp16 = layer_norm(axes = input_153_axes_0, beta = norm12_2_bias_to_fp16, epsilon = var_5518_to_fp16, gamma = norm12_2_weight_to_fp16, x = input_151_cast_fp16)[name = tensor("input_153_cast_fp16")]; + tensor linear12_1_weight_to_fp16 = const()[name = tensor("linear12_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(311561280)))]; + tensor linear_50_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = linear12_1_weight_to_fp16, x = input_153_cast_fp16)[name = tensor("linear_50_cast_fp16")]; + tensor input_155_mode_0 = const()[name = tensor("input_155_mode_0"), val = tensor("EXACT")]; + tensor input_155_cast_fp16 = gelu(mode = input_155_mode_0, x = linear_50_cast_fp16)[name = tensor("input_155_cast_fp16")]; + tensor linear12_2_weight_to_fp16 = const()[name = tensor("linear12_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(319949952)))]; + tensor linear_51_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = linear12_2_weight_to_fp16, x = input_155_cast_fp16)[name = tensor("linear_51_cast_fp16")]; + tensor input_157_cast_fp16 = add(x = input_151_cast_fp16, y = linear_51_cast_fp16)[name = tensor("input_157_cast_fp16")]; + tensor input_159_axes_0 = const()[name = tensor("input_159_axes_0"), val = tensor([-1])]; + tensor norm13_1_weight_to_fp16 = const()[name = tensor("norm13_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(328338624)))]; + tensor norm13_1_bias_to_fp16 = const()[name = tensor("norm13_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(328340736)))]; + tensor var_5535_to_fp16 = const()[name = tensor("op_5535_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_159_cast_fp16 = layer_norm(axes = input_159_axes_0, beta = norm13_1_bias_to_fp16, epsilon = var_5535_to_fp16, gamma = norm13_1_weight_to_fp16, x = input_157_cast_fp16)[name = tensor("input_159_cast_fp16")]; + tensor attn13_in_proj_weight_to_fp16 = const()[name = tensor("attn13_in_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(328342848)))]; + tensor linear_52_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = attn13_in_proj_weight_to_fp16, x = input_159_cast_fp16)[name = tensor("linear_52_cast_fp16")]; + tensor var_5549 = const()[name = tensor("op_5549"), val = tensor([1, 256, 3, 16, 64])]; + tensor qkv_27_cast_fp16 = reshape(shape = var_5549, x = linear_52_cast_fp16)[name = tensor("qkv_27_cast_fp16")]; + tensor q_79_begin_0 = const()[name = tensor("q_79_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor q_79_end_0 = const()[name = tensor("q_79_end_0"), val = tensor([1, 256, 1, 16, 64])]; + tensor q_79_end_mask_0 = const()[name = tensor("q_79_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor q_79_squeeze_mask_0 = const()[name = tensor("q_79_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor q_79_cast_fp16 = slice_by_index(begin = q_79_begin_0, end = q_79_end_0, end_mask = q_79_end_mask_0, squeeze_mask = q_79_squeeze_mask_0, x = qkv_27_cast_fp16)[name = tensor("q_79_cast_fp16")]; + tensor k_53_begin_0 = const()[name = tensor("k_53_begin_0"), val = tensor([0, 0, 1, 0, 0])]; + tensor k_53_end_0 = const()[name = tensor("k_53_end_0"), val = tensor([1, 256, 2, 16, 64])]; + tensor k_53_end_mask_0 = const()[name = tensor("k_53_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor k_53_squeeze_mask_0 = const()[name = tensor("k_53_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor k_53_cast_fp16 = slice_by_index(begin = k_53_begin_0, end = k_53_end_0, end_mask = k_53_end_mask_0, squeeze_mask = k_53_squeeze_mask_0, x = qkv_27_cast_fp16)[name = tensor("k_53_cast_fp16")]; + tensor v_27_begin_0 = const()[name = tensor("v_27_begin_0"), val = tensor([0, 0, 2, 0, 0])]; + tensor v_27_end_0 = const()[name = tensor("v_27_end_0"), val = tensor([1, 256, 3, 16, 64])]; + tensor v_27_end_mask_0 = const()[name = tensor("v_27_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor v_27_squeeze_mask_0 = const()[name = tensor("v_27_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor v_27_cast_fp16 = slice_by_index(begin = v_27_begin_0, end = v_27_end_0, end_mask = v_27_end_mask_0, squeeze_mask = v_27_squeeze_mask_0, x = qkv_27_cast_fp16)[name = tensor("v_27_cast_fp16")]; + tensor var_5607 = const()[name = tensor("op_5607"), val = tensor([1, 1])]; + tensor position13_to_fp16_dtype_0 = const()[name = tensor("position13_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor position13_to_fp16 = cast(dtype = position13_to_fp16_dtype_0, x = position13)[name = tensor("cast_596")]; + tensor var_5608_cast_fp16 = reshape(shape = var_5607, x = position13_to_fp16)[name = tensor("op_5608_cast_fp16")]; + tensor ts_81_cast_fp16 = add(x = ts_1_promoted_to_fp16, y = var_5608_cast_fp16)[name = tensor("ts_81_cast_fp16")]; + tensor var_5615 = const()[name = tensor("op_5615"), val = tensor([1, 256, 1, 1])]; + tensor ts_83_cast_fp16 = reshape(shape = var_5615, x = ts_81_cast_fp16)[name = tensor("ts_83_cast_fp16")]; + tensor var_5622 = const()[name = tensor("op_5622"), val = tensor([1, 256, 16, 32, 2])]; + tensor q_complex_27_cast_fp16 = reshape(shape = var_5622, x = q_79_cast_fp16)[name = tensor("q_complex_27_cast_fp16")]; + tensor var_5629 = const()[name = tensor("op_5629"), val = tensor([1, 256, 16, 32, 2])]; + tensor k_complex_27_cast_fp16 = reshape(shape = var_5629, x = k_53_cast_fp16)[name = tensor("k_complex_27_cast_fp16")]; + tensor var_5633_begin_0 = const()[name = tensor("op_5633_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_5633_end_0 = const()[name = tensor("op_5633_end_0"), val = tensor([1, 256, 16, 32, 1])]; + tensor var_5633_end_mask_0 = const()[name = tensor("op_5633_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_5633_squeeze_mask_0 = const()[name = tensor("op_5633_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_5633_cast_fp16 = slice_by_index(begin = var_5633_begin_0, end = var_5633_end_0, end_mask = var_5633_end_mask_0, squeeze_mask = var_5633_squeeze_mask_0, x = q_complex_27_cast_fp16)[name = tensor("op_5633_cast_fp16")]; + tensor var_5641_begin_0 = const()[name = tensor("op_5641_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_5641_end_0 = const()[name = tensor("op_5641_end_0"), val = tensor([1, 256, 16, 32, 2])]; + tensor var_5641_end_mask_0 = const()[name = tensor("op_5641_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_5641_squeeze_mask_0 = const()[name = tensor("op_5641_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_5641_cast_fp16 = slice_by_index(begin = var_5641_begin_0, end = var_5641_end_0, end_mask = var_5641_end_mask_0, squeeze_mask = var_5641_squeeze_mask_0, x = q_complex_27_cast_fp16)[name = tensor("op_5641_cast_fp16")]; + tensor var_5649_begin_0 = const()[name = tensor("op_5649_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_5649_end_0 = const()[name = tensor("op_5649_end_0"), val = tensor([1, 256, 16, 32, 1])]; + tensor var_5649_end_mask_0 = const()[name = tensor("op_5649_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_5649_squeeze_mask_0 = const()[name = tensor("op_5649_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_5649_cast_fp16 = slice_by_index(begin = var_5649_begin_0, end = var_5649_end_0, end_mask = var_5649_end_mask_0, squeeze_mask = var_5649_squeeze_mask_0, x = k_complex_27_cast_fp16)[name = tensor("op_5649_cast_fp16")]; + tensor var_5657_begin_0 = const()[name = tensor("op_5657_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_5657_end_0 = const()[name = tensor("op_5657_end_0"), val = tensor([1, 256, 16, 32, 2])]; + tensor var_5657_end_mask_0 = const()[name = tensor("op_5657_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_5657_squeeze_mask_0 = const()[name = tensor("op_5657_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_5657_cast_fp16 = slice_by_index(begin = var_5657_begin_0, end = var_5657_end_0, end_mask = var_5657_end_mask_0, squeeze_mask = var_5657_squeeze_mask_0, x = k_complex_27_cast_fp16)[name = tensor("op_5657_cast_fp16")]; + tensor freqs_27_to_fp16 = const()[name = tensor("freqs_27_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(334634368)))]; + tensor var_5663_cast_fp16 = mul(x = freqs_27_to_fp16, y = ts_83_cast_fp16)[name = tensor("op_5663_cast_fp16")]; + tensor rotr_27_cast_fp16 = cos(x = var_5663_cast_fp16)[name = tensor("rotr_27_cast_fp16")]; + tensor roti_27_cast_fp16 = sin(x = var_5663_cast_fp16)[name = tensor("roti_27_cast_fp16")]; + tensor var_5667_cast_fp16 = mul(x = var_5633_cast_fp16, y = rotr_27_cast_fp16)[name = tensor("op_5667_cast_fp16")]; + tensor var_5668_cast_fp16 = mul(x = var_5641_cast_fp16, y = roti_27_cast_fp16)[name = tensor("op_5668_cast_fp16")]; + tensor qor_53_cast_fp16 = sub(x = var_5667_cast_fp16, y = var_5668_cast_fp16)[name = tensor("qor_53_cast_fp16")]; + tensor var_5671_cast_fp16 = mul(x = var_5633_cast_fp16, y = roti_27_cast_fp16)[name = tensor("op_5671_cast_fp16")]; + tensor var_5672_cast_fp16 = mul(x = var_5641_cast_fp16, y = rotr_27_cast_fp16)[name = tensor("op_5672_cast_fp16")]; + tensor qoi_53_cast_fp16 = add(x = var_5671_cast_fp16, y = var_5672_cast_fp16)[name = tensor("qoi_53_cast_fp16")]; + tensor var_5675_cast_fp16 = mul(x = var_5649_cast_fp16, y = rotr_27_cast_fp16)[name = tensor("op_5675_cast_fp16")]; + tensor var_5676_cast_fp16 = mul(x = var_5657_cast_fp16, y = roti_27_cast_fp16)[name = tensor("op_5676_cast_fp16")]; + tensor kor_53_cast_fp16 = sub(x = var_5675_cast_fp16, y = var_5676_cast_fp16)[name = tensor("kor_53_cast_fp16")]; + tensor var_5679_cast_fp16 = mul(x = var_5649_cast_fp16, y = roti_27_cast_fp16)[name = tensor("op_5679_cast_fp16")]; + tensor var_5680_cast_fp16 = mul(x = var_5657_cast_fp16, y = rotr_27_cast_fp16)[name = tensor("op_5680_cast_fp16")]; + tensor koi_53_cast_fp16 = add(x = var_5679_cast_fp16, y = var_5680_cast_fp16)[name = tensor("koi_53_cast_fp16")]; + tensor qo_27_axis_0 = const()[name = tensor("qo_27_axis_0"), val = tensor(-1)]; + tensor qo_27_cast_fp16 = stack(axis = qo_27_axis_0, values = (qor_53_cast_fp16, qoi_53_cast_fp16))[name = tensor("qo_27_cast_fp16")]; + tensor ko_27_axis_0 = const()[name = tensor("ko_27_axis_0"), val = tensor(-1)]; + tensor ko_27_cast_fp16 = stack(axis = ko_27_axis_0, values = (kor_53_cast_fp16, koi_53_cast_fp16))[name = tensor("ko_27_cast_fp16")]; + tensor var_5713 = const()[name = tensor("op_5713"), val = tensor([1, 256, 16, 64])]; + tensor q_81_cast_fp16 = reshape(shape = var_5713, x = qo_27_cast_fp16)[name = tensor("q_81_cast_fp16")]; + tensor var_5719 = const()[name = tensor("op_5719"), val = tensor([1, 256, 16, 64])]; + tensor k_55_cast_fp16 = reshape(shape = var_5719, x = ko_27_cast_fp16)[name = tensor("k_55_cast_fp16")]; + tensor write_indices_float_79_cast_fp16 = add(x = var_5608_cast_fp16, y = var_393_promoted_to_fp16)[name = tensor("write_indices_float_79_cast_fp16")]; + tensor _inversed_5740_y_0_to_fp16 = const()[name = tensor("_inversed_5740_y_0_to_fp16"), val = tensor(0x1p-9)]; + tensor _inversed_5740_cast_fp16 = mul(x = write_indices_float_79_cast_fp16, y = _inversed_5740_y_0_to_fp16)[name = tensor("_inversed_5740_cast_fp16")]; + tensor var_5741_cast_fp16 = floor(x = _inversed_5740_cast_fp16)[name = tensor("op_5741_cast_fp16")]; + tensor var_5742_to_fp16 = const()[name = tensor("op_5742_to_fp16"), val = tensor(0x1p+9)]; + tensor var_5743_cast_fp16 = mul(x = var_5741_cast_fp16, y = var_5742_to_fp16)[name = tensor("op_5743_cast_fp16")]; + tensor write_indices_float_81_cast_fp16 = sub(x = write_indices_float_79_cast_fp16, y = var_5743_cast_fp16)[name = tensor("write_indices_float_81_cast_fp16")]; + tensor var_5761_value_0_to_fp16 = const()[name = tensor("op_5761_value_0_to_fp16"), val = tensor(0x1.ffp+8)]; + tensor var_5761_cast_fp16 = fill_like(ref_tensor = write_indices_float_81_cast_fp16, value = var_5761_value_0_to_fp16)[name = tensor("op_5761_cast_fp16")]; + tensor write_indices_float_83_cast_fp16 = select(a = write_indices_float_81_cast_fp16, b = var_5761_cast_fp16, cond = token_valid_1_cast_fp16)[name = tensor("write_indices_float_83_cast_fp16")]; + tensor var_5767_dtype_0 = const()[name = tensor("op_5767_dtype_0"), val = tensor("int32")]; + tensor var_5772 = const()[name = tensor("op_5772"), val = tensor([1, 256, 1, 1])]; + tensor write_indices_float_83_cast_fp16_to_int32 = cast(dtype = var_5767_dtype_0, x = write_indices_float_83_cast_fp16)[name = tensor("cast_595")]; + tensor var_5773 = reshape(shape = var_5772, x = write_indices_float_83_cast_fp16_to_int32)[name = tensor("op_5773")]; + tensor write_indices_27_reps_0 = const()[name = tensor("write_indices_27_reps_0"), val = tensor([1, 1, 16, 64])]; + tensor write_indices_27 = tile(reps = write_indices_27_reps_0, x = var_5773)[name = tensor("write_indices_27")]; + tensor var_5783_begin_0 = const()[name = tensor("op_5783_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_5783_end_0 = const()[name = tensor("op_5783_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor var_5783_end_mask_0 = const()[name = tensor("op_5783_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_5783_squeeze_mask_0 = const()[name = tensor("op_5783_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor cache13_to_fp16_dtype_0 = const()[name = tensor("cache13_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor cache13_to_fp16 = cast(dtype = cache13_to_fp16_dtype_0, x = cache13)[name = tensor("cast_594")]; + tensor var_5783_cast_fp16 = slice_by_index(begin = var_5783_begin_0, end = var_5783_end_0, end_mask = var_5783_end_mask_0, squeeze_mask = var_5783_squeeze_mask_0, x = cache13_to_fp16)[name = tensor("op_5783_cast_fp16")]; + tensor var_5785_axis_0 = const()[name = tensor("op_5785_axis_0"), val = tensor(1)]; + tensor var_5785_mode_0 = const()[name = tensor("op_5785_mode_0"), val = tensor("update")]; + tensor var_5785_validate_indices_0 = const()[name = tensor("op_5785_validate_indices_0"), val = tensor(false)]; + tensor var_5785_cast_fp16 = scatter_along_axis(axis = var_5785_axis_0, data = var_5783_cast_fp16, indices = write_indices_27, mode = var_5785_mode_0, updates = k_55_cast_fp16, validate_indices = var_5785_validate_indices_0)[name = tensor("op_5785_cast_fp16")]; + tensor concat_106 = const()[name = tensor("concat_106"), val = tensor([0, 0, 0, 0, 0])]; + tensor concat_107 = const()[name = tensor("concat_107"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_27_internal_tensor_assign_1_stride_0 = const()[name = tensor("new_cache_27_internal_tensor_assign_1_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_27_internal_tensor_assign_1_begin_mask_0 = const()[name = tensor("new_cache_27_internal_tensor_assign_1_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_27_internal_tensor_assign_1_end_mask_0 = const()[name = tensor("new_cache_27_internal_tensor_assign_1_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_27_internal_tensor_assign_1_squeeze_mask_0 = const()[name = tensor("new_cache_27_internal_tensor_assign_1_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_72 = const()[name = tensor("shape_72"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_26 = const()[name = tensor("reduce_prod_26"), val = tensor(1048576)]; + tensor range_1d_26_start_0 = const()[name = tensor("range_1d_26_start_0"), val = tensor(0)]; + tensor range_1d_26_step_0 = const()[name = tensor("range_1d_26_step_0"), val = tensor(1)]; + tensor range_1d_26 = range_1d(end = reduce_prod_26, start = range_1d_26_start_0, step = range_1d_26_step_0)[name = tensor("range_1d_26")]; + tensor reshape_130 = reshape(shape = shape_72, x = range_1d_26)[name = tensor("reshape_130")]; + tensor slice_by_index_26 = slice_by_index(begin = concat_106, begin_mask = new_cache_27_internal_tensor_assign_1_begin_mask_0, end = concat_107, end_mask = new_cache_27_internal_tensor_assign_1_end_mask_0, squeeze_mask = new_cache_27_internal_tensor_assign_1_squeeze_mask_0, stride = new_cache_27_internal_tensor_assign_1_stride_0, x = reshape_130)[name = tensor("slice_by_index_26")]; + tensor reshape_131_shape_0 = const()[name = tensor("reshape_131_shape_0"), val = tensor([-1])]; + tensor reshape_131 = reshape(shape = reshape_131_shape_0, x = slice_by_index_26)[name = tensor("reshape_131")]; + tensor reshape_132_shape_0 = const()[name = tensor("reshape_132_shape_0"), val = tensor([-1])]; + tensor reshape_132_cast_fp16 = reshape(shape = reshape_132_shape_0, x = var_5785_cast_fp16)[name = tensor("reshape_132_cast_fp16")]; + tensor reshape_133_shape_0 = const()[name = tensor("reshape_133_shape_0"), val = tensor([-1])]; + tensor reshape_133_cast_fp16 = reshape(shape = reshape_133_shape_0, x = cache13_to_fp16)[name = tensor("reshape_133_cast_fp16")]; + tensor scatter_26_mode_0 = const()[name = tensor("scatter_26_mode_0"), val = tensor("update")]; + tensor scatter_26_axis_0 = const()[name = tensor("scatter_26_axis_0"), val = tensor(0)]; + tensor scatter_26_validate_indices_0 = const()[name = tensor("scatter_26_validate_indices_0"), val = tensor(false)]; + tensor scatter_26_cast_fp16 = scatter(axis = scatter_26_axis_0, data = reshape_133_cast_fp16, indices = reshape_131, mode = scatter_26_mode_0, updates = reshape_132_cast_fp16, validate_indices = scatter_26_validate_indices_0)[name = tensor("scatter_26_cast_fp16")]; + tensor reshape_134_cast_fp16 = reshape(shape = shape_72, x = scatter_26_cast_fp16)[name = tensor("reshape_134_cast_fp16")]; + tensor var_5793_begin_0 = const()[name = tensor("op_5793_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor var_5793_end_0 = const()[name = tensor("op_5793_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor var_5793_end_mask_0 = const()[name = tensor("op_5793_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_5793_squeeze_mask_0 = const()[name = tensor("op_5793_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor var_5793_cast_fp16 = slice_by_index(begin = var_5793_begin_0, end = var_5793_end_0, end_mask = var_5793_end_mask_0, squeeze_mask = var_5793_squeeze_mask_0, x = reshape_134_cast_fp16)[name = tensor("op_5793_cast_fp16")]; + tensor var_5795_axis_0 = const()[name = tensor("op_5795_axis_0"), val = tensor(1)]; + tensor var_5795_mode_0 = const()[name = tensor("op_5795_mode_0"), val = tensor("update")]; + tensor var_5795_validate_indices_0 = const()[name = tensor("op_5795_validate_indices_0"), val = tensor(false)]; + tensor var_5795_cast_fp16 = scatter_along_axis(axis = var_5795_axis_0, data = var_5793_cast_fp16, indices = write_indices_27, mode = var_5795_mode_0, updates = v_27_cast_fp16, validate_indices = var_5795_validate_indices_0)[name = tensor("op_5795_cast_fp16")]; + tensor concat_108 = const()[name = tensor("concat_108"), val = tensor([1, 0, 0, 0, 0])]; + tensor concat_109 = const()[name = tensor("concat_109"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_27_internal_tensor_assign_2_stride_0 = const()[name = tensor("new_cache_27_internal_tensor_assign_2_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_27_internal_tensor_assign_2_begin_mask_0 = const()[name = tensor("new_cache_27_internal_tensor_assign_2_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_27_internal_tensor_assign_2_end_mask_0 = const()[name = tensor("new_cache_27_internal_tensor_assign_2_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_27_internal_tensor_assign_2_squeeze_mask_0 = const()[name = tensor("new_cache_27_internal_tensor_assign_2_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_73 = const()[name = tensor("shape_73"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_27 = const()[name = tensor("reduce_prod_27"), val = tensor(1048576)]; + tensor range_1d_27_start_0 = const()[name = tensor("range_1d_27_start_0"), val = tensor(0)]; + tensor range_1d_27_step_0 = const()[name = tensor("range_1d_27_step_0"), val = tensor(1)]; + tensor range_1d_27 = range_1d(end = reduce_prod_27, start = range_1d_27_start_0, step = range_1d_27_step_0)[name = tensor("range_1d_27")]; + tensor reshape_135 = reshape(shape = shape_73, x = range_1d_27)[name = tensor("reshape_135")]; + tensor slice_by_index_27 = slice_by_index(begin = concat_108, begin_mask = new_cache_27_internal_tensor_assign_2_begin_mask_0, end = concat_109, end_mask = new_cache_27_internal_tensor_assign_2_end_mask_0, squeeze_mask = new_cache_27_internal_tensor_assign_2_squeeze_mask_0, stride = new_cache_27_internal_tensor_assign_2_stride_0, x = reshape_135)[name = tensor("slice_by_index_27")]; + tensor reshape_136_shape_0 = const()[name = tensor("reshape_136_shape_0"), val = tensor([-1])]; + tensor reshape_136 = reshape(shape = reshape_136_shape_0, x = slice_by_index_27)[name = tensor("reshape_136")]; + tensor reshape_137_shape_0 = const()[name = tensor("reshape_137_shape_0"), val = tensor([-1])]; + tensor reshape_137_cast_fp16 = reshape(shape = reshape_137_shape_0, x = var_5795_cast_fp16)[name = tensor("reshape_137_cast_fp16")]; + tensor reshape_138_shape_0 = const()[name = tensor("reshape_138_shape_0"), val = tensor([-1])]; + tensor reshape_138_cast_fp16 = reshape(shape = reshape_138_shape_0, x = reshape_134_cast_fp16)[name = tensor("reshape_138_cast_fp16")]; + tensor scatter_27_mode_0 = const()[name = tensor("scatter_27_mode_0"), val = tensor("update")]; + tensor scatter_27_axis_0 = const()[name = tensor("scatter_27_axis_0"), val = tensor(0)]; + tensor scatter_27_validate_indices_0 = const()[name = tensor("scatter_27_validate_indices_0"), val = tensor(false)]; + tensor scatter_27_cast_fp16 = scatter(axis = scatter_27_axis_0, data = reshape_138_cast_fp16, indices = reshape_136, mode = scatter_27_mode_0, updates = reshape_137_cast_fp16, validate_indices = scatter_27_validate_indices_0)[name = tensor("scatter_27_cast_fp16")]; + tensor reshape_139_cast_fp16 = reshape(shape = shape_73, x = scatter_27_cast_fp16)[name = tensor("reshape_139_cast_fp16")]; + tensor reshape_139_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("reshape_139_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor keys_79_begin_0 = const()[name = tensor("keys_79_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor keys_79_end_0 = const()[name = tensor("keys_79_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor keys_79_end_mask_0 = const()[name = tensor("keys_79_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor keys_79_squeeze_mask_0 = const()[name = tensor("keys_79_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor keys_79_cast_fp16 = slice_by_index(begin = keys_79_begin_0, end = keys_79_end_0, end_mask = keys_79_end_mask_0, squeeze_mask = keys_79_squeeze_mask_0, x = reshape_139_cast_fp16)[name = tensor("keys_79_cast_fp16")]; + tensor values_79_begin_0 = const()[name = tensor("values_79_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor values_79_end_0 = const()[name = tensor("values_79_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor values_79_end_mask_0 = const()[name = tensor("values_79_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor values_79_squeeze_mask_0 = const()[name = tensor("values_79_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor values_79_cast_fp16 = slice_by_index(begin = values_79_begin_0, end = values_79_end_0, end_mask = values_79_end_mask_0, squeeze_mask = values_79_squeeze_mask_0, x = reshape_139_cast_fp16)[name = tensor("values_79_cast_fp16")]; + tensor var_5807_cast_fp16 = not_equal(x = keys_79_cast_fp16, y = keys_79_cast_fp16)[name = tensor("op_5807_cast_fp16")]; + tensor keys_81_cast_fp16 = select(a = var_470_to_fp16, b = keys_79_cast_fp16, cond = var_5807_cast_fp16)[name = tensor("keys_81_cast_fp16")]; + tensor var_5815_cast_fp16 = not_equal(x = values_79_cast_fp16, y = values_79_cast_fp16)[name = tensor("op_5815_cast_fp16")]; + tensor values_81_cast_fp16 = select(a = var_470_to_fp16, b = values_79_cast_fp16, cond = var_5815_cast_fp16)[name = tensor("values_81_cast_fp16")]; + tensor var_5839 = const()[name = tensor("op_5839"), val = tensor([0, 2, 1, 3])]; + tensor var_5855 = const()[name = tensor("op_5855"), val = tensor([1, 1, 1])]; + tensor var_5856_cast_fp16 = reshape(shape = var_5855, x = position13_to_fp16)[name = tensor("op_5856_cast_fp16")]; + tensor q_positions_27_cast_fp16 = add(x = var_5856_cast_fp16, y = q_offsets_1_promoted_to_fp16)[name = tensor("q_positions_27_cast_fp16")]; + tensor valid_end_27_cast_fp16 = add(x = var_5856_cast_fp16, y = var_536_cast_fp16)[name = tensor("valid_end_27_cast_fp16")]; + tensor valid_mask_27_cast_fp16 = less(x = k_positions_1_promoted_to_fp16, y = valid_end_27_cast_fp16)[name = tensor("valid_mask_27_cast_fp16")]; + tensor causal_mask_27_cast_fp16 = less_equal(x = k_positions_1_promoted_to_fp16, y = q_positions_27_cast_fp16)[name = tensor("causal_mask_27_cast_fp16")]; + tensor var_5884 = logical_and(x = valid_mask_27_cast_fp16, y = causal_mask_27_cast_fp16)[name = tensor("op_5884")]; + tensor attn_mask_27_axes_0 = const()[name = tensor("attn_mask_27_axes_0"), val = tensor([1])]; + tensor attn_mask_27 = expand_dims(axes = attn_mask_27_axes_0, x = var_5884)[name = tensor("attn_mask_27")]; + tensor var_5898_transpose_x_0 = const()[name = tensor("op_5898_transpose_x_0"), val = tensor(false)]; + tensor var_5898_transpose_y_0 = const()[name = tensor("op_5898_transpose_y_0"), val = tensor(false)]; + tensor transpose_95_perm_0 = const()[name = tensor("transpose_95_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_96_perm_0 = const()[name = tensor("transpose_96_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_96 = transpose(perm = transpose_96_perm_0, x = keys_81_cast_fp16)[name = tensor("transpose_152")]; + tensor transpose_95 = transpose(perm = transpose_95_perm_0, x = q_81_cast_fp16)[name = tensor("transpose_153")]; + tensor var_5898_cast_fp16 = matmul(transpose_x = var_5898_transpose_x_0, transpose_y = var_5898_transpose_y_0, x = transpose_95, y = transpose_96)[name = tensor("op_5898_cast_fp16")]; + tensor var_5892_to_fp16 = const()[name = tensor("op_5892_to_fp16"), val = tensor([0x1p-3])]; + tensor attn_weights_79_cast_fp16 = mul(x = var_5898_cast_fp16, y = var_5892_to_fp16)[name = tensor("attn_weights_79_cast_fp16")]; + tensor var_5900 = logical_not(x = attn_mask_27)[name = tensor("op_5900")]; + tensor var_5901_to_fp16 = const()[name = tensor("op_5901_to_fp16"), val = tensor(-inf)]; + tensor attn_weights_81_cast_fp16 = select(a = var_5901_to_fp16, b = attn_weights_79_cast_fp16, cond = var_5900)[name = tensor("attn_weights_81_cast_fp16")]; + tensor var_5903 = const()[name = tensor("op_5903"), val = tensor(-1)]; + tensor attn_weights_83_cast_fp16 = softmax(axis = var_5903, x = attn_weights_81_cast_fp16)[name = tensor("attn_weights_83_cast_fp16")]; + tensor attn_output_27_transpose_x_0 = const()[name = tensor("attn_output_27_transpose_x_0"), val = tensor(false)]; + tensor attn_output_27_transpose_y_0 = const()[name = tensor("attn_output_27_transpose_y_0"), val = tensor(false)]; + tensor values_83_cast_fp16 = transpose(perm = var_5839, x = values_81_cast_fp16)[name = tensor("transpose_154")]; + tensor attn_output_27_cast_fp16 = matmul(transpose_x = attn_output_27_transpose_x_0, transpose_y = attn_output_27_transpose_y_0, x = attn_weights_83_cast_fp16, y = values_83_cast_fp16)[name = tensor("attn_output_27_cast_fp16")]; + tensor var_5911 = const()[name = tensor("op_5911"), val = tensor([0, 2, 1, 3])]; + tensor var_5916 = const()[name = tensor("op_5916"), val = tensor([1, 256, 1024])]; + tensor var_5912_cast_fp16 = transpose(perm = var_5911, x = attn_output_27_cast_fp16)[name = tensor("transpose_151")]; + tensor input_161_cast_fp16 = reshape(shape = var_5916, x = var_5912_cast_fp16)[name = tensor("input_161_cast_fp16")]; + tensor attn13_out_proj_weight_to_fp16 = const()[name = tensor("attn13_out_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(334634496)))]; + tensor linear_53_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = attn13_out_proj_weight_to_fp16, x = input_161_cast_fp16)[name = tensor("linear_53_cast_fp16")]; + tensor var_5925_cast_fp16 = add(x = position13_to_fp16, y = valid_len_to_fp16)[name = tensor("op_5925_cast_fp16")]; + tensor var_5925_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_5925_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor input_163_cast_fp16 = add(x = input_157_cast_fp16, y = linear_53_cast_fp16)[name = tensor("input_163_cast_fp16")]; + tensor input_165_axes_0 = const()[name = tensor("input_165_axes_0"), val = tensor([-1])]; + tensor norm13_2_weight_to_fp16 = const()[name = tensor("norm13_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(336731712)))]; + tensor norm13_2_bias_to_fp16 = const()[name = tensor("norm13_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(336733824)))]; + tensor var_5929_to_fp16 = const()[name = tensor("op_5929_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_165_cast_fp16 = layer_norm(axes = input_165_axes_0, beta = norm13_2_bias_to_fp16, epsilon = var_5929_to_fp16, gamma = norm13_2_weight_to_fp16, x = input_163_cast_fp16)[name = tensor("input_165_cast_fp16")]; + tensor linear13_1_weight_to_fp16 = const()[name = tensor("linear13_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(336735936)))]; + tensor linear_54_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = linear13_1_weight_to_fp16, x = input_165_cast_fp16)[name = tensor("linear_54_cast_fp16")]; + tensor input_167_mode_0 = const()[name = tensor("input_167_mode_0"), val = tensor("EXACT")]; + tensor input_167_cast_fp16 = gelu(mode = input_167_mode_0, x = linear_54_cast_fp16)[name = tensor("input_167_cast_fp16")]; + tensor linear13_2_weight_to_fp16 = const()[name = tensor("linear13_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(345124608)))]; + tensor linear_55_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = linear13_2_weight_to_fp16, x = input_167_cast_fp16)[name = tensor("linear_55_cast_fp16")]; + tensor input_169_cast_fp16 = add(x = input_163_cast_fp16, y = linear_55_cast_fp16)[name = tensor("input_169_cast_fp16")]; + tensor input_171_axes_0 = const()[name = tensor("input_171_axes_0"), val = tensor([-1])]; + tensor norm14_1_weight_to_fp16 = const()[name = tensor("norm14_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(353513280)))]; + tensor norm14_1_bias_to_fp16 = const()[name = tensor("norm14_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(353515392)))]; + tensor var_5946_to_fp16 = const()[name = tensor("op_5946_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_171_cast_fp16 = layer_norm(axes = input_171_axes_0, beta = norm14_1_bias_to_fp16, epsilon = var_5946_to_fp16, gamma = norm14_1_weight_to_fp16, x = input_169_cast_fp16)[name = tensor("input_171_cast_fp16")]; + tensor attn14_in_proj_weight_to_fp16 = const()[name = tensor("attn14_in_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(353517504)))]; + tensor linear_56_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = attn14_in_proj_weight_to_fp16, x = input_171_cast_fp16)[name = tensor("linear_56_cast_fp16")]; + tensor var_5960 = const()[name = tensor("op_5960"), val = tensor([1, 256, 3, 16, 64])]; + tensor qkv_29_cast_fp16 = reshape(shape = var_5960, x = linear_56_cast_fp16)[name = tensor("qkv_29_cast_fp16")]; + tensor q_85_begin_0 = const()[name = tensor("q_85_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor q_85_end_0 = const()[name = tensor("q_85_end_0"), val = tensor([1, 256, 1, 16, 64])]; + tensor q_85_end_mask_0 = const()[name = tensor("q_85_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor q_85_squeeze_mask_0 = const()[name = tensor("q_85_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor q_85_cast_fp16 = slice_by_index(begin = q_85_begin_0, end = q_85_end_0, end_mask = q_85_end_mask_0, squeeze_mask = q_85_squeeze_mask_0, x = qkv_29_cast_fp16)[name = tensor("q_85_cast_fp16")]; + tensor k_57_begin_0 = const()[name = tensor("k_57_begin_0"), val = tensor([0, 0, 1, 0, 0])]; + tensor k_57_end_0 = const()[name = tensor("k_57_end_0"), val = tensor([1, 256, 2, 16, 64])]; + tensor k_57_end_mask_0 = const()[name = tensor("k_57_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor k_57_squeeze_mask_0 = const()[name = tensor("k_57_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor k_57_cast_fp16 = slice_by_index(begin = k_57_begin_0, end = k_57_end_0, end_mask = k_57_end_mask_0, squeeze_mask = k_57_squeeze_mask_0, x = qkv_29_cast_fp16)[name = tensor("k_57_cast_fp16")]; + tensor v_29_begin_0 = const()[name = tensor("v_29_begin_0"), val = tensor([0, 0, 2, 0, 0])]; + tensor v_29_end_0 = const()[name = tensor("v_29_end_0"), val = tensor([1, 256, 3, 16, 64])]; + tensor v_29_end_mask_0 = const()[name = tensor("v_29_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor v_29_squeeze_mask_0 = const()[name = tensor("v_29_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor v_29_cast_fp16 = slice_by_index(begin = v_29_begin_0, end = v_29_end_0, end_mask = v_29_end_mask_0, squeeze_mask = v_29_squeeze_mask_0, x = qkv_29_cast_fp16)[name = tensor("v_29_cast_fp16")]; + tensor var_6018 = const()[name = tensor("op_6018"), val = tensor([1, 1])]; + tensor position14_to_fp16_dtype_0 = const()[name = tensor("position14_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor position14_to_fp16 = cast(dtype = position14_to_fp16_dtype_0, x = position14)[name = tensor("cast_591")]; + tensor var_6019_cast_fp16 = reshape(shape = var_6018, x = position14_to_fp16)[name = tensor("op_6019_cast_fp16")]; + tensor ts_87_cast_fp16 = add(x = ts_1_promoted_to_fp16, y = var_6019_cast_fp16)[name = tensor("ts_87_cast_fp16")]; + tensor var_6026 = const()[name = tensor("op_6026"), val = tensor([1, 256, 1, 1])]; + tensor ts_89_cast_fp16 = reshape(shape = var_6026, x = ts_87_cast_fp16)[name = tensor("ts_89_cast_fp16")]; + tensor var_6033 = const()[name = tensor("op_6033"), val = tensor([1, 256, 16, 32, 2])]; + tensor q_complex_29_cast_fp16 = reshape(shape = var_6033, x = q_85_cast_fp16)[name = tensor("q_complex_29_cast_fp16")]; + tensor var_6040 = const()[name = tensor("op_6040"), val = tensor([1, 256, 16, 32, 2])]; + tensor k_complex_29_cast_fp16 = reshape(shape = var_6040, x = k_57_cast_fp16)[name = tensor("k_complex_29_cast_fp16")]; + tensor var_6044_begin_0 = const()[name = tensor("op_6044_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_6044_end_0 = const()[name = tensor("op_6044_end_0"), val = tensor([1, 256, 16, 32, 1])]; + tensor var_6044_end_mask_0 = const()[name = tensor("op_6044_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_6044_squeeze_mask_0 = const()[name = tensor("op_6044_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_6044_cast_fp16 = slice_by_index(begin = var_6044_begin_0, end = var_6044_end_0, end_mask = var_6044_end_mask_0, squeeze_mask = var_6044_squeeze_mask_0, x = q_complex_29_cast_fp16)[name = tensor("op_6044_cast_fp16")]; + tensor var_6052_begin_0 = const()[name = tensor("op_6052_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_6052_end_0 = const()[name = tensor("op_6052_end_0"), val = tensor([1, 256, 16, 32, 2])]; + tensor var_6052_end_mask_0 = const()[name = tensor("op_6052_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_6052_squeeze_mask_0 = const()[name = tensor("op_6052_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_6052_cast_fp16 = slice_by_index(begin = var_6052_begin_0, end = var_6052_end_0, end_mask = var_6052_end_mask_0, squeeze_mask = var_6052_squeeze_mask_0, x = q_complex_29_cast_fp16)[name = tensor("op_6052_cast_fp16")]; + tensor var_6060_begin_0 = const()[name = tensor("op_6060_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_6060_end_0 = const()[name = tensor("op_6060_end_0"), val = tensor([1, 256, 16, 32, 1])]; + tensor var_6060_end_mask_0 = const()[name = tensor("op_6060_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_6060_squeeze_mask_0 = const()[name = tensor("op_6060_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_6060_cast_fp16 = slice_by_index(begin = var_6060_begin_0, end = var_6060_end_0, end_mask = var_6060_end_mask_0, squeeze_mask = var_6060_squeeze_mask_0, x = k_complex_29_cast_fp16)[name = tensor("op_6060_cast_fp16")]; + tensor var_6068_begin_0 = const()[name = tensor("op_6068_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_6068_end_0 = const()[name = tensor("op_6068_end_0"), val = tensor([1, 256, 16, 32, 2])]; + tensor var_6068_end_mask_0 = const()[name = tensor("op_6068_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_6068_squeeze_mask_0 = const()[name = tensor("op_6068_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_6068_cast_fp16 = slice_by_index(begin = var_6068_begin_0, end = var_6068_end_0, end_mask = var_6068_end_mask_0, squeeze_mask = var_6068_squeeze_mask_0, x = k_complex_29_cast_fp16)[name = tensor("op_6068_cast_fp16")]; + tensor freqs_29_to_fp16 = const()[name = tensor("freqs_29_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(359809024)))]; + tensor var_6074_cast_fp16 = mul(x = freqs_29_to_fp16, y = ts_89_cast_fp16)[name = tensor("op_6074_cast_fp16")]; + tensor rotr_29_cast_fp16 = cos(x = var_6074_cast_fp16)[name = tensor("rotr_29_cast_fp16")]; + tensor roti_29_cast_fp16 = sin(x = var_6074_cast_fp16)[name = tensor("roti_29_cast_fp16")]; + tensor var_6078_cast_fp16 = mul(x = var_6044_cast_fp16, y = rotr_29_cast_fp16)[name = tensor("op_6078_cast_fp16")]; + tensor var_6079_cast_fp16 = mul(x = var_6052_cast_fp16, y = roti_29_cast_fp16)[name = tensor("op_6079_cast_fp16")]; + tensor qor_57_cast_fp16 = sub(x = var_6078_cast_fp16, y = var_6079_cast_fp16)[name = tensor("qor_57_cast_fp16")]; + tensor var_6082_cast_fp16 = mul(x = var_6044_cast_fp16, y = roti_29_cast_fp16)[name = tensor("op_6082_cast_fp16")]; + tensor var_6083_cast_fp16 = mul(x = var_6052_cast_fp16, y = rotr_29_cast_fp16)[name = tensor("op_6083_cast_fp16")]; + tensor qoi_57_cast_fp16 = add(x = var_6082_cast_fp16, y = var_6083_cast_fp16)[name = tensor("qoi_57_cast_fp16")]; + tensor var_6086_cast_fp16 = mul(x = var_6060_cast_fp16, y = rotr_29_cast_fp16)[name = tensor("op_6086_cast_fp16")]; + tensor var_6087_cast_fp16 = mul(x = var_6068_cast_fp16, y = roti_29_cast_fp16)[name = tensor("op_6087_cast_fp16")]; + tensor kor_57_cast_fp16 = sub(x = var_6086_cast_fp16, y = var_6087_cast_fp16)[name = tensor("kor_57_cast_fp16")]; + tensor var_6090_cast_fp16 = mul(x = var_6060_cast_fp16, y = roti_29_cast_fp16)[name = tensor("op_6090_cast_fp16")]; + tensor var_6091_cast_fp16 = mul(x = var_6068_cast_fp16, y = rotr_29_cast_fp16)[name = tensor("op_6091_cast_fp16")]; + tensor koi_57_cast_fp16 = add(x = var_6090_cast_fp16, y = var_6091_cast_fp16)[name = tensor("koi_57_cast_fp16")]; + tensor qo_29_axis_0 = const()[name = tensor("qo_29_axis_0"), val = tensor(-1)]; + tensor qo_29_cast_fp16 = stack(axis = qo_29_axis_0, values = (qor_57_cast_fp16, qoi_57_cast_fp16))[name = tensor("qo_29_cast_fp16")]; + tensor ko_29_axis_0 = const()[name = tensor("ko_29_axis_0"), val = tensor(-1)]; + tensor ko_29_cast_fp16 = stack(axis = ko_29_axis_0, values = (kor_57_cast_fp16, koi_57_cast_fp16))[name = tensor("ko_29_cast_fp16")]; + tensor var_6124 = const()[name = tensor("op_6124"), val = tensor([1, 256, 16, 64])]; + tensor q_87_cast_fp16 = reshape(shape = var_6124, x = qo_29_cast_fp16)[name = tensor("q_87_cast_fp16")]; + tensor var_6130 = const()[name = tensor("op_6130"), val = tensor([1, 256, 16, 64])]; + tensor k_59_cast_fp16 = reshape(shape = var_6130, x = ko_29_cast_fp16)[name = tensor("k_59_cast_fp16")]; + tensor write_indices_float_85_cast_fp16 = add(x = var_6019_cast_fp16, y = var_393_promoted_to_fp16)[name = tensor("write_indices_float_85_cast_fp16")]; + tensor _inversed_6151_y_0_to_fp16 = const()[name = tensor("_inversed_6151_y_0_to_fp16"), val = tensor(0x1p-9)]; + tensor _inversed_6151_cast_fp16 = mul(x = write_indices_float_85_cast_fp16, y = _inversed_6151_y_0_to_fp16)[name = tensor("_inversed_6151_cast_fp16")]; + tensor var_6152_cast_fp16 = floor(x = _inversed_6151_cast_fp16)[name = tensor("op_6152_cast_fp16")]; + tensor var_6153_to_fp16 = const()[name = tensor("op_6153_to_fp16"), val = tensor(0x1p+9)]; + tensor var_6154_cast_fp16 = mul(x = var_6152_cast_fp16, y = var_6153_to_fp16)[name = tensor("op_6154_cast_fp16")]; + tensor write_indices_float_87_cast_fp16 = sub(x = write_indices_float_85_cast_fp16, y = var_6154_cast_fp16)[name = tensor("write_indices_float_87_cast_fp16")]; + tensor var_6172_value_0_to_fp16 = const()[name = tensor("op_6172_value_0_to_fp16"), val = tensor(0x1.ffp+8)]; + tensor var_6172_cast_fp16 = fill_like(ref_tensor = write_indices_float_87_cast_fp16, value = var_6172_value_0_to_fp16)[name = tensor("op_6172_cast_fp16")]; + tensor write_indices_float_89_cast_fp16 = select(a = write_indices_float_87_cast_fp16, b = var_6172_cast_fp16, cond = token_valid_1_cast_fp16)[name = tensor("write_indices_float_89_cast_fp16")]; + tensor var_6178_dtype_0 = const()[name = tensor("op_6178_dtype_0"), val = tensor("int32")]; + tensor var_6183 = const()[name = tensor("op_6183"), val = tensor([1, 256, 1, 1])]; + tensor write_indices_float_89_cast_fp16_to_int32 = cast(dtype = var_6178_dtype_0, x = write_indices_float_89_cast_fp16)[name = tensor("cast_590")]; + tensor var_6184 = reshape(shape = var_6183, x = write_indices_float_89_cast_fp16_to_int32)[name = tensor("op_6184")]; + tensor write_indices_29_reps_0 = const()[name = tensor("write_indices_29_reps_0"), val = tensor([1, 1, 16, 64])]; + tensor write_indices_29 = tile(reps = write_indices_29_reps_0, x = var_6184)[name = tensor("write_indices_29")]; + tensor var_6194_begin_0 = const()[name = tensor("op_6194_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_6194_end_0 = const()[name = tensor("op_6194_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor var_6194_end_mask_0 = const()[name = tensor("op_6194_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_6194_squeeze_mask_0 = const()[name = tensor("op_6194_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor cache14_to_fp16_dtype_0 = const()[name = tensor("cache14_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor cache14_to_fp16 = cast(dtype = cache14_to_fp16_dtype_0, x = cache14)[name = tensor("cast_589")]; + tensor var_6194_cast_fp16 = slice_by_index(begin = var_6194_begin_0, end = var_6194_end_0, end_mask = var_6194_end_mask_0, squeeze_mask = var_6194_squeeze_mask_0, x = cache14_to_fp16)[name = tensor("op_6194_cast_fp16")]; + tensor var_6196_axis_0 = const()[name = tensor("op_6196_axis_0"), val = tensor(1)]; + tensor var_6196_mode_0 = const()[name = tensor("op_6196_mode_0"), val = tensor("update")]; + tensor var_6196_validate_indices_0 = const()[name = tensor("op_6196_validate_indices_0"), val = tensor(false)]; + tensor var_6196_cast_fp16 = scatter_along_axis(axis = var_6196_axis_0, data = var_6194_cast_fp16, indices = write_indices_29, mode = var_6196_mode_0, updates = k_59_cast_fp16, validate_indices = var_6196_validate_indices_0)[name = tensor("op_6196_cast_fp16")]; + tensor concat_114 = const()[name = tensor("concat_114"), val = tensor([0, 0, 0, 0, 0])]; + tensor concat_115 = const()[name = tensor("concat_115"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_29_internal_tensor_assign_1_stride_0 = const()[name = tensor("new_cache_29_internal_tensor_assign_1_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_29_internal_tensor_assign_1_begin_mask_0 = const()[name = tensor("new_cache_29_internal_tensor_assign_1_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_29_internal_tensor_assign_1_end_mask_0 = const()[name = tensor("new_cache_29_internal_tensor_assign_1_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_29_internal_tensor_assign_1_squeeze_mask_0 = const()[name = tensor("new_cache_29_internal_tensor_assign_1_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_74 = const()[name = tensor("shape_74"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_28 = const()[name = tensor("reduce_prod_28"), val = tensor(1048576)]; + tensor range_1d_28_start_0 = const()[name = tensor("range_1d_28_start_0"), val = tensor(0)]; + tensor range_1d_28_step_0 = const()[name = tensor("range_1d_28_step_0"), val = tensor(1)]; + tensor range_1d_28 = range_1d(end = reduce_prod_28, start = range_1d_28_start_0, step = range_1d_28_step_0)[name = tensor("range_1d_28")]; + tensor reshape_140 = reshape(shape = shape_74, x = range_1d_28)[name = tensor("reshape_140")]; + tensor slice_by_index_28 = slice_by_index(begin = concat_114, begin_mask = new_cache_29_internal_tensor_assign_1_begin_mask_0, end = concat_115, end_mask = new_cache_29_internal_tensor_assign_1_end_mask_0, squeeze_mask = new_cache_29_internal_tensor_assign_1_squeeze_mask_0, stride = new_cache_29_internal_tensor_assign_1_stride_0, x = reshape_140)[name = tensor("slice_by_index_28")]; + tensor reshape_141_shape_0 = const()[name = tensor("reshape_141_shape_0"), val = tensor([-1])]; + tensor reshape_141 = reshape(shape = reshape_141_shape_0, x = slice_by_index_28)[name = tensor("reshape_141")]; + tensor reshape_142_shape_0 = const()[name = tensor("reshape_142_shape_0"), val = tensor([-1])]; + tensor reshape_142_cast_fp16 = reshape(shape = reshape_142_shape_0, x = var_6196_cast_fp16)[name = tensor("reshape_142_cast_fp16")]; + tensor reshape_143_shape_0 = const()[name = tensor("reshape_143_shape_0"), val = tensor([-1])]; + tensor reshape_143_cast_fp16 = reshape(shape = reshape_143_shape_0, x = cache14_to_fp16)[name = tensor("reshape_143_cast_fp16")]; + tensor scatter_28_mode_0 = const()[name = tensor("scatter_28_mode_0"), val = tensor("update")]; + tensor scatter_28_axis_0 = const()[name = tensor("scatter_28_axis_0"), val = tensor(0)]; + tensor scatter_28_validate_indices_0 = const()[name = tensor("scatter_28_validate_indices_0"), val = tensor(false)]; + tensor scatter_28_cast_fp16 = scatter(axis = scatter_28_axis_0, data = reshape_143_cast_fp16, indices = reshape_141, mode = scatter_28_mode_0, updates = reshape_142_cast_fp16, validate_indices = scatter_28_validate_indices_0)[name = tensor("scatter_28_cast_fp16")]; + tensor reshape_144_cast_fp16 = reshape(shape = shape_74, x = scatter_28_cast_fp16)[name = tensor("reshape_144_cast_fp16")]; + tensor var_6204_begin_0 = const()[name = tensor("op_6204_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor var_6204_end_0 = const()[name = tensor("op_6204_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor var_6204_end_mask_0 = const()[name = tensor("op_6204_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_6204_squeeze_mask_0 = const()[name = tensor("op_6204_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor var_6204_cast_fp16 = slice_by_index(begin = var_6204_begin_0, end = var_6204_end_0, end_mask = var_6204_end_mask_0, squeeze_mask = var_6204_squeeze_mask_0, x = reshape_144_cast_fp16)[name = tensor("op_6204_cast_fp16")]; + tensor var_6206_axis_0 = const()[name = tensor("op_6206_axis_0"), val = tensor(1)]; + tensor var_6206_mode_0 = const()[name = tensor("op_6206_mode_0"), val = tensor("update")]; + tensor var_6206_validate_indices_0 = const()[name = tensor("op_6206_validate_indices_0"), val = tensor(false)]; + tensor var_6206_cast_fp16 = scatter_along_axis(axis = var_6206_axis_0, data = var_6204_cast_fp16, indices = write_indices_29, mode = var_6206_mode_0, updates = v_29_cast_fp16, validate_indices = var_6206_validate_indices_0)[name = tensor("op_6206_cast_fp16")]; + tensor concat_116 = const()[name = tensor("concat_116"), val = tensor([1, 0, 0, 0, 0])]; + tensor concat_117 = const()[name = tensor("concat_117"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_29_internal_tensor_assign_2_stride_0 = const()[name = tensor("new_cache_29_internal_tensor_assign_2_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_29_internal_tensor_assign_2_begin_mask_0 = const()[name = tensor("new_cache_29_internal_tensor_assign_2_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_29_internal_tensor_assign_2_end_mask_0 = const()[name = tensor("new_cache_29_internal_tensor_assign_2_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_29_internal_tensor_assign_2_squeeze_mask_0 = const()[name = tensor("new_cache_29_internal_tensor_assign_2_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_75 = const()[name = tensor("shape_75"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_29 = const()[name = tensor("reduce_prod_29"), val = tensor(1048576)]; + tensor range_1d_29_start_0 = const()[name = tensor("range_1d_29_start_0"), val = tensor(0)]; + tensor range_1d_29_step_0 = const()[name = tensor("range_1d_29_step_0"), val = tensor(1)]; + tensor range_1d_29 = range_1d(end = reduce_prod_29, start = range_1d_29_start_0, step = range_1d_29_step_0)[name = tensor("range_1d_29")]; + tensor reshape_145 = reshape(shape = shape_75, x = range_1d_29)[name = tensor("reshape_145")]; + tensor slice_by_index_29 = slice_by_index(begin = concat_116, begin_mask = new_cache_29_internal_tensor_assign_2_begin_mask_0, end = concat_117, end_mask = new_cache_29_internal_tensor_assign_2_end_mask_0, squeeze_mask = new_cache_29_internal_tensor_assign_2_squeeze_mask_0, stride = new_cache_29_internal_tensor_assign_2_stride_0, x = reshape_145)[name = tensor("slice_by_index_29")]; + tensor reshape_146_shape_0 = const()[name = tensor("reshape_146_shape_0"), val = tensor([-1])]; + tensor reshape_146 = reshape(shape = reshape_146_shape_0, x = slice_by_index_29)[name = tensor("reshape_146")]; + tensor reshape_147_shape_0 = const()[name = tensor("reshape_147_shape_0"), val = tensor([-1])]; + tensor reshape_147_cast_fp16 = reshape(shape = reshape_147_shape_0, x = var_6206_cast_fp16)[name = tensor("reshape_147_cast_fp16")]; + tensor reshape_148_shape_0 = const()[name = tensor("reshape_148_shape_0"), val = tensor([-1])]; + tensor reshape_148_cast_fp16 = reshape(shape = reshape_148_shape_0, x = reshape_144_cast_fp16)[name = tensor("reshape_148_cast_fp16")]; + tensor scatter_29_mode_0 = const()[name = tensor("scatter_29_mode_0"), val = tensor("update")]; + tensor scatter_29_axis_0 = const()[name = tensor("scatter_29_axis_0"), val = tensor(0)]; + tensor scatter_29_validate_indices_0 = const()[name = tensor("scatter_29_validate_indices_0"), val = tensor(false)]; + tensor scatter_29_cast_fp16 = scatter(axis = scatter_29_axis_0, data = reshape_148_cast_fp16, indices = reshape_146, mode = scatter_29_mode_0, updates = reshape_147_cast_fp16, validate_indices = scatter_29_validate_indices_0)[name = tensor("scatter_29_cast_fp16")]; + tensor reshape_149_cast_fp16 = reshape(shape = shape_75, x = scatter_29_cast_fp16)[name = tensor("reshape_149_cast_fp16")]; + tensor reshape_149_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("reshape_149_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor keys_85_begin_0 = const()[name = tensor("keys_85_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor keys_85_end_0 = const()[name = tensor("keys_85_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor keys_85_end_mask_0 = const()[name = tensor("keys_85_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor keys_85_squeeze_mask_0 = const()[name = tensor("keys_85_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor keys_85_cast_fp16 = slice_by_index(begin = keys_85_begin_0, end = keys_85_end_0, end_mask = keys_85_end_mask_0, squeeze_mask = keys_85_squeeze_mask_0, x = reshape_149_cast_fp16)[name = tensor("keys_85_cast_fp16")]; + tensor values_85_begin_0 = const()[name = tensor("values_85_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor values_85_end_0 = const()[name = tensor("values_85_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor values_85_end_mask_0 = const()[name = tensor("values_85_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor values_85_squeeze_mask_0 = const()[name = tensor("values_85_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor values_85_cast_fp16 = slice_by_index(begin = values_85_begin_0, end = values_85_end_0, end_mask = values_85_end_mask_0, squeeze_mask = values_85_squeeze_mask_0, x = reshape_149_cast_fp16)[name = tensor("values_85_cast_fp16")]; + tensor var_6218_cast_fp16 = not_equal(x = keys_85_cast_fp16, y = keys_85_cast_fp16)[name = tensor("op_6218_cast_fp16")]; + tensor keys_87_cast_fp16 = select(a = var_470_to_fp16, b = keys_85_cast_fp16, cond = var_6218_cast_fp16)[name = tensor("keys_87_cast_fp16")]; + tensor var_6226_cast_fp16 = not_equal(x = values_85_cast_fp16, y = values_85_cast_fp16)[name = tensor("op_6226_cast_fp16")]; + tensor values_87_cast_fp16 = select(a = var_470_to_fp16, b = values_85_cast_fp16, cond = var_6226_cast_fp16)[name = tensor("values_87_cast_fp16")]; + tensor var_6250 = const()[name = tensor("op_6250"), val = tensor([0, 2, 1, 3])]; + tensor var_6266 = const()[name = tensor("op_6266"), val = tensor([1, 1, 1])]; + tensor var_6267_cast_fp16 = reshape(shape = var_6266, x = position14_to_fp16)[name = tensor("op_6267_cast_fp16")]; + tensor q_positions_29_cast_fp16 = add(x = var_6267_cast_fp16, y = q_offsets_1_promoted_to_fp16)[name = tensor("q_positions_29_cast_fp16")]; + tensor valid_end_29_cast_fp16 = add(x = var_6267_cast_fp16, y = var_536_cast_fp16)[name = tensor("valid_end_29_cast_fp16")]; + tensor valid_mask_29_cast_fp16 = less(x = k_positions_1_promoted_to_fp16, y = valid_end_29_cast_fp16)[name = tensor("valid_mask_29_cast_fp16")]; + tensor causal_mask_29_cast_fp16 = less_equal(x = k_positions_1_promoted_to_fp16, y = q_positions_29_cast_fp16)[name = tensor("causal_mask_29_cast_fp16")]; + tensor var_6295 = logical_and(x = valid_mask_29_cast_fp16, y = causal_mask_29_cast_fp16)[name = tensor("op_6295")]; + tensor attn_mask_29_axes_0 = const()[name = tensor("attn_mask_29_axes_0"), val = tensor([1])]; + tensor attn_mask_29 = expand_dims(axes = attn_mask_29_axes_0, x = var_6295)[name = tensor("attn_mask_29")]; + tensor var_6309_transpose_x_0 = const()[name = tensor("op_6309_transpose_x_0"), val = tensor(false)]; + tensor var_6309_transpose_y_0 = const()[name = tensor("op_6309_transpose_y_0"), val = tensor(false)]; + tensor transpose_97_perm_0 = const()[name = tensor("transpose_97_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_98_perm_0 = const()[name = tensor("transpose_98_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_98 = transpose(perm = transpose_98_perm_0, x = keys_87_cast_fp16)[name = tensor("transpose_148")]; + tensor transpose_97 = transpose(perm = transpose_97_perm_0, x = q_87_cast_fp16)[name = tensor("transpose_149")]; + tensor var_6309_cast_fp16 = matmul(transpose_x = var_6309_transpose_x_0, transpose_y = var_6309_transpose_y_0, x = transpose_97, y = transpose_98)[name = tensor("op_6309_cast_fp16")]; + tensor var_6303_to_fp16 = const()[name = tensor("op_6303_to_fp16"), val = tensor([0x1p-3])]; + tensor attn_weights_85_cast_fp16 = mul(x = var_6309_cast_fp16, y = var_6303_to_fp16)[name = tensor("attn_weights_85_cast_fp16")]; + tensor var_6311 = logical_not(x = attn_mask_29)[name = tensor("op_6311")]; + tensor var_6312_to_fp16 = const()[name = tensor("op_6312_to_fp16"), val = tensor(-inf)]; + tensor attn_weights_87_cast_fp16 = select(a = var_6312_to_fp16, b = attn_weights_85_cast_fp16, cond = var_6311)[name = tensor("attn_weights_87_cast_fp16")]; + tensor var_6314 = const()[name = tensor("op_6314"), val = tensor(-1)]; + tensor attn_weights_89_cast_fp16 = softmax(axis = var_6314, x = attn_weights_87_cast_fp16)[name = tensor("attn_weights_89_cast_fp16")]; + tensor attn_output_29_transpose_x_0 = const()[name = tensor("attn_output_29_transpose_x_0"), val = tensor(false)]; + tensor attn_output_29_transpose_y_0 = const()[name = tensor("attn_output_29_transpose_y_0"), val = tensor(false)]; + tensor values_89_cast_fp16 = transpose(perm = var_6250, x = values_87_cast_fp16)[name = tensor("transpose_150")]; + tensor attn_output_29_cast_fp16 = matmul(transpose_x = attn_output_29_transpose_x_0, transpose_y = attn_output_29_transpose_y_0, x = attn_weights_89_cast_fp16, y = values_89_cast_fp16)[name = tensor("attn_output_29_cast_fp16")]; + tensor var_6322 = const()[name = tensor("op_6322"), val = tensor([0, 2, 1, 3])]; + tensor var_6327 = const()[name = tensor("op_6327"), val = tensor([1, 256, 1024])]; + tensor var_6323_cast_fp16 = transpose(perm = var_6322, x = attn_output_29_cast_fp16)[name = tensor("transpose_147")]; + tensor input_173_cast_fp16 = reshape(shape = var_6327, x = var_6323_cast_fp16)[name = tensor("input_173_cast_fp16")]; + tensor attn14_out_proj_weight_to_fp16 = const()[name = tensor("attn14_out_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(359809152)))]; + tensor linear_57_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = attn14_out_proj_weight_to_fp16, x = input_173_cast_fp16)[name = tensor("linear_57_cast_fp16")]; + tensor var_6336_cast_fp16 = add(x = position14_to_fp16, y = valid_len_to_fp16)[name = tensor("op_6336_cast_fp16")]; + tensor var_6336_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_6336_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor input_175_cast_fp16 = add(x = input_169_cast_fp16, y = linear_57_cast_fp16)[name = tensor("input_175_cast_fp16")]; + tensor input_177_axes_0 = const()[name = tensor("input_177_axes_0"), val = tensor([-1])]; + tensor norm14_2_weight_to_fp16 = const()[name = tensor("norm14_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(361906368)))]; + tensor norm14_2_bias_to_fp16 = const()[name = tensor("norm14_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(361908480)))]; + tensor var_6340_to_fp16 = const()[name = tensor("op_6340_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_177_cast_fp16 = layer_norm(axes = input_177_axes_0, beta = norm14_2_bias_to_fp16, epsilon = var_6340_to_fp16, gamma = norm14_2_weight_to_fp16, x = input_175_cast_fp16)[name = tensor("input_177_cast_fp16")]; + tensor linear14_1_weight_to_fp16 = const()[name = tensor("linear14_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(361910592)))]; + tensor linear_58_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = linear14_1_weight_to_fp16, x = input_177_cast_fp16)[name = tensor("linear_58_cast_fp16")]; + tensor input_179_mode_0 = const()[name = tensor("input_179_mode_0"), val = tensor("EXACT")]; + tensor input_179_cast_fp16 = gelu(mode = input_179_mode_0, x = linear_58_cast_fp16)[name = tensor("input_179_cast_fp16")]; + tensor linear14_2_weight_to_fp16 = const()[name = tensor("linear14_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(370299264)))]; + tensor linear_59_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = linear14_2_weight_to_fp16, x = input_179_cast_fp16)[name = tensor("linear_59_cast_fp16")]; + tensor input_181_cast_fp16 = add(x = input_175_cast_fp16, y = linear_59_cast_fp16)[name = tensor("input_181_cast_fp16")]; + tensor input_183_axes_0 = const()[name = tensor("input_183_axes_0"), val = tensor([-1])]; + tensor norm15_1_weight_to_fp16 = const()[name = tensor("norm15_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(378687936)))]; + tensor norm15_1_bias_to_fp16 = const()[name = tensor("norm15_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(378690048)))]; + tensor var_6357_to_fp16 = const()[name = tensor("op_6357_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_183_cast_fp16 = layer_norm(axes = input_183_axes_0, beta = norm15_1_bias_to_fp16, epsilon = var_6357_to_fp16, gamma = norm15_1_weight_to_fp16, x = input_181_cast_fp16)[name = tensor("input_183_cast_fp16")]; + tensor attn15_in_proj_weight_to_fp16 = const()[name = tensor("attn15_in_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(378692160)))]; + tensor linear_60_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = attn15_in_proj_weight_to_fp16, x = input_183_cast_fp16)[name = tensor("linear_60_cast_fp16")]; + tensor var_6371 = const()[name = tensor("op_6371"), val = tensor([1, 256, 3, 16, 64])]; + tensor qkv_31_cast_fp16 = reshape(shape = var_6371, x = linear_60_cast_fp16)[name = tensor("qkv_31_cast_fp16")]; + tensor q_91_begin_0 = const()[name = tensor("q_91_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor q_91_end_0 = const()[name = tensor("q_91_end_0"), val = tensor([1, 256, 1, 16, 64])]; + tensor q_91_end_mask_0 = const()[name = tensor("q_91_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor q_91_squeeze_mask_0 = const()[name = tensor("q_91_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor q_91_cast_fp16 = slice_by_index(begin = q_91_begin_0, end = q_91_end_0, end_mask = q_91_end_mask_0, squeeze_mask = q_91_squeeze_mask_0, x = qkv_31_cast_fp16)[name = tensor("q_91_cast_fp16")]; + tensor k_61_begin_0 = const()[name = tensor("k_61_begin_0"), val = tensor([0, 0, 1, 0, 0])]; + tensor k_61_end_0 = const()[name = tensor("k_61_end_0"), val = tensor([1, 256, 2, 16, 64])]; + tensor k_61_end_mask_0 = const()[name = tensor("k_61_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor k_61_squeeze_mask_0 = const()[name = tensor("k_61_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor k_61_cast_fp16 = slice_by_index(begin = k_61_begin_0, end = k_61_end_0, end_mask = k_61_end_mask_0, squeeze_mask = k_61_squeeze_mask_0, x = qkv_31_cast_fp16)[name = tensor("k_61_cast_fp16")]; + tensor v_31_begin_0 = const()[name = tensor("v_31_begin_0"), val = tensor([0, 0, 2, 0, 0])]; + tensor v_31_end_0 = const()[name = tensor("v_31_end_0"), val = tensor([1, 256, 3, 16, 64])]; + tensor v_31_end_mask_0 = const()[name = tensor("v_31_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor v_31_squeeze_mask_0 = const()[name = tensor("v_31_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor v_31_cast_fp16 = slice_by_index(begin = v_31_begin_0, end = v_31_end_0, end_mask = v_31_end_mask_0, squeeze_mask = v_31_squeeze_mask_0, x = qkv_31_cast_fp16)[name = tensor("v_31_cast_fp16")]; + tensor var_6429 = const()[name = tensor("op_6429"), val = tensor([1, 1])]; + tensor position15_to_fp16_dtype_0 = const()[name = tensor("position15_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor position15_to_fp16 = cast(dtype = position15_to_fp16_dtype_0, x = position15)[name = tensor("cast_586")]; + tensor var_6430_cast_fp16 = reshape(shape = var_6429, x = position15_to_fp16)[name = tensor("op_6430_cast_fp16")]; + tensor ts_93_cast_fp16 = add(x = ts_1_promoted_to_fp16, y = var_6430_cast_fp16)[name = tensor("ts_93_cast_fp16")]; + tensor var_6437 = const()[name = tensor("op_6437"), val = tensor([1, 256, 1, 1])]; + tensor ts_95_cast_fp16 = reshape(shape = var_6437, x = ts_93_cast_fp16)[name = tensor("ts_95_cast_fp16")]; + tensor var_6444 = const()[name = tensor("op_6444"), val = tensor([1, 256, 16, 32, 2])]; + tensor q_complex_31_cast_fp16 = reshape(shape = var_6444, x = q_91_cast_fp16)[name = tensor("q_complex_31_cast_fp16")]; + tensor var_6451 = const()[name = tensor("op_6451"), val = tensor([1, 256, 16, 32, 2])]; + tensor k_complex_31_cast_fp16 = reshape(shape = var_6451, x = k_61_cast_fp16)[name = tensor("k_complex_31_cast_fp16")]; + tensor var_6455_begin_0 = const()[name = tensor("op_6455_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_6455_end_0 = const()[name = tensor("op_6455_end_0"), val = tensor([1, 256, 16, 32, 1])]; + tensor var_6455_end_mask_0 = const()[name = tensor("op_6455_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_6455_squeeze_mask_0 = const()[name = tensor("op_6455_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_6455_cast_fp16 = slice_by_index(begin = var_6455_begin_0, end = var_6455_end_0, end_mask = var_6455_end_mask_0, squeeze_mask = var_6455_squeeze_mask_0, x = q_complex_31_cast_fp16)[name = tensor("op_6455_cast_fp16")]; + tensor var_6463_begin_0 = const()[name = tensor("op_6463_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_6463_end_0 = const()[name = tensor("op_6463_end_0"), val = tensor([1, 256, 16, 32, 2])]; + tensor var_6463_end_mask_0 = const()[name = tensor("op_6463_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_6463_squeeze_mask_0 = const()[name = tensor("op_6463_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_6463_cast_fp16 = slice_by_index(begin = var_6463_begin_0, end = var_6463_end_0, end_mask = var_6463_end_mask_0, squeeze_mask = var_6463_squeeze_mask_0, x = q_complex_31_cast_fp16)[name = tensor("op_6463_cast_fp16")]; + tensor var_6471_begin_0 = const()[name = tensor("op_6471_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_6471_end_0 = const()[name = tensor("op_6471_end_0"), val = tensor([1, 256, 16, 32, 1])]; + tensor var_6471_end_mask_0 = const()[name = tensor("op_6471_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_6471_squeeze_mask_0 = const()[name = tensor("op_6471_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_6471_cast_fp16 = slice_by_index(begin = var_6471_begin_0, end = var_6471_end_0, end_mask = var_6471_end_mask_0, squeeze_mask = var_6471_squeeze_mask_0, x = k_complex_31_cast_fp16)[name = tensor("op_6471_cast_fp16")]; + tensor var_6479_begin_0 = const()[name = tensor("op_6479_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_6479_end_0 = const()[name = tensor("op_6479_end_0"), val = tensor([1, 256, 16, 32, 2])]; + tensor var_6479_end_mask_0 = const()[name = tensor("op_6479_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_6479_squeeze_mask_0 = const()[name = tensor("op_6479_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_6479_cast_fp16 = slice_by_index(begin = var_6479_begin_0, end = var_6479_end_0, end_mask = var_6479_end_mask_0, squeeze_mask = var_6479_squeeze_mask_0, x = k_complex_31_cast_fp16)[name = tensor("op_6479_cast_fp16")]; + tensor freqs_31_to_fp16 = const()[name = tensor("freqs_31_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(384983680)))]; + tensor var_6485_cast_fp16 = mul(x = freqs_31_to_fp16, y = ts_95_cast_fp16)[name = tensor("op_6485_cast_fp16")]; + tensor rotr_31_cast_fp16 = cos(x = var_6485_cast_fp16)[name = tensor("rotr_31_cast_fp16")]; + tensor roti_31_cast_fp16 = sin(x = var_6485_cast_fp16)[name = tensor("roti_31_cast_fp16")]; + tensor var_6489_cast_fp16 = mul(x = var_6455_cast_fp16, y = rotr_31_cast_fp16)[name = tensor("op_6489_cast_fp16")]; + tensor var_6490_cast_fp16 = mul(x = var_6463_cast_fp16, y = roti_31_cast_fp16)[name = tensor("op_6490_cast_fp16")]; + tensor qor_61_cast_fp16 = sub(x = var_6489_cast_fp16, y = var_6490_cast_fp16)[name = tensor("qor_61_cast_fp16")]; + tensor var_6493_cast_fp16 = mul(x = var_6455_cast_fp16, y = roti_31_cast_fp16)[name = tensor("op_6493_cast_fp16")]; + tensor var_6494_cast_fp16 = mul(x = var_6463_cast_fp16, y = rotr_31_cast_fp16)[name = tensor("op_6494_cast_fp16")]; + tensor qoi_61_cast_fp16 = add(x = var_6493_cast_fp16, y = var_6494_cast_fp16)[name = tensor("qoi_61_cast_fp16")]; + tensor var_6497_cast_fp16 = mul(x = var_6471_cast_fp16, y = rotr_31_cast_fp16)[name = tensor("op_6497_cast_fp16")]; + tensor var_6498_cast_fp16 = mul(x = var_6479_cast_fp16, y = roti_31_cast_fp16)[name = tensor("op_6498_cast_fp16")]; + tensor kor_61_cast_fp16 = sub(x = var_6497_cast_fp16, y = var_6498_cast_fp16)[name = tensor("kor_61_cast_fp16")]; + tensor var_6501_cast_fp16 = mul(x = var_6471_cast_fp16, y = roti_31_cast_fp16)[name = tensor("op_6501_cast_fp16")]; + tensor var_6502_cast_fp16 = mul(x = var_6479_cast_fp16, y = rotr_31_cast_fp16)[name = tensor("op_6502_cast_fp16")]; + tensor koi_61_cast_fp16 = add(x = var_6501_cast_fp16, y = var_6502_cast_fp16)[name = tensor("koi_61_cast_fp16")]; + tensor qo_31_axis_0 = const()[name = tensor("qo_31_axis_0"), val = tensor(-1)]; + tensor qo_31_cast_fp16 = stack(axis = qo_31_axis_0, values = (qor_61_cast_fp16, qoi_61_cast_fp16))[name = tensor("qo_31_cast_fp16")]; + tensor ko_31_axis_0 = const()[name = tensor("ko_31_axis_0"), val = tensor(-1)]; + tensor ko_31_cast_fp16 = stack(axis = ko_31_axis_0, values = (kor_61_cast_fp16, koi_61_cast_fp16))[name = tensor("ko_31_cast_fp16")]; + tensor var_6535 = const()[name = tensor("op_6535"), val = tensor([1, 256, 16, 64])]; + tensor q_93_cast_fp16 = reshape(shape = var_6535, x = qo_31_cast_fp16)[name = tensor("q_93_cast_fp16")]; + tensor var_6541 = const()[name = tensor("op_6541"), val = tensor([1, 256, 16, 64])]; + tensor k_63_cast_fp16 = reshape(shape = var_6541, x = ko_31_cast_fp16)[name = tensor("k_63_cast_fp16")]; + tensor write_indices_float_91_cast_fp16 = add(x = var_6430_cast_fp16, y = var_393_promoted_to_fp16)[name = tensor("write_indices_float_91_cast_fp16")]; + tensor _inversed_6562_y_0_to_fp16 = const()[name = tensor("_inversed_6562_y_0_to_fp16"), val = tensor(0x1p-9)]; + tensor _inversed_6562_cast_fp16 = mul(x = write_indices_float_91_cast_fp16, y = _inversed_6562_y_0_to_fp16)[name = tensor("_inversed_6562_cast_fp16")]; + tensor var_6563_cast_fp16 = floor(x = _inversed_6562_cast_fp16)[name = tensor("op_6563_cast_fp16")]; + tensor var_6564_to_fp16 = const()[name = tensor("op_6564_to_fp16"), val = tensor(0x1p+9)]; + tensor var_6565_cast_fp16 = mul(x = var_6563_cast_fp16, y = var_6564_to_fp16)[name = tensor("op_6565_cast_fp16")]; + tensor write_indices_float_93_cast_fp16 = sub(x = write_indices_float_91_cast_fp16, y = var_6565_cast_fp16)[name = tensor("write_indices_float_93_cast_fp16")]; + tensor var_6583_value_0_to_fp16 = const()[name = tensor("op_6583_value_0_to_fp16"), val = tensor(0x1.ffp+8)]; + tensor var_6583_cast_fp16 = fill_like(ref_tensor = write_indices_float_93_cast_fp16, value = var_6583_value_0_to_fp16)[name = tensor("op_6583_cast_fp16")]; + tensor write_indices_float_95_cast_fp16 = select(a = write_indices_float_93_cast_fp16, b = var_6583_cast_fp16, cond = token_valid_1_cast_fp16)[name = tensor("write_indices_float_95_cast_fp16")]; + tensor var_6589_dtype_0 = const()[name = tensor("op_6589_dtype_0"), val = tensor("int32")]; + tensor var_6594 = const()[name = tensor("op_6594"), val = tensor([1, 256, 1, 1])]; + tensor write_indices_float_95_cast_fp16_to_int32 = cast(dtype = var_6589_dtype_0, x = write_indices_float_95_cast_fp16)[name = tensor("cast_585")]; + tensor var_6595 = reshape(shape = var_6594, x = write_indices_float_95_cast_fp16_to_int32)[name = tensor("op_6595")]; + tensor write_indices_31_reps_0 = const()[name = tensor("write_indices_31_reps_0"), val = tensor([1, 1, 16, 64])]; + tensor write_indices_31 = tile(reps = write_indices_31_reps_0, x = var_6595)[name = tensor("write_indices_31")]; + tensor var_6605_begin_0 = const()[name = tensor("op_6605_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_6605_end_0 = const()[name = tensor("op_6605_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor var_6605_end_mask_0 = const()[name = tensor("op_6605_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_6605_squeeze_mask_0 = const()[name = tensor("op_6605_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor cache15_to_fp16_dtype_0 = const()[name = tensor("cache15_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor cache15_to_fp16 = cast(dtype = cache15_to_fp16_dtype_0, x = cache15)[name = tensor("cast_584")]; + tensor var_6605_cast_fp16 = slice_by_index(begin = var_6605_begin_0, end = var_6605_end_0, end_mask = var_6605_end_mask_0, squeeze_mask = var_6605_squeeze_mask_0, x = cache15_to_fp16)[name = tensor("op_6605_cast_fp16")]; + tensor var_6607_axis_0 = const()[name = tensor("op_6607_axis_0"), val = tensor(1)]; + tensor var_6607_mode_0 = const()[name = tensor("op_6607_mode_0"), val = tensor("update")]; + tensor var_6607_validate_indices_0 = const()[name = tensor("op_6607_validate_indices_0"), val = tensor(false)]; + tensor var_6607_cast_fp16 = scatter_along_axis(axis = var_6607_axis_0, data = var_6605_cast_fp16, indices = write_indices_31, mode = var_6607_mode_0, updates = k_63_cast_fp16, validate_indices = var_6607_validate_indices_0)[name = tensor("op_6607_cast_fp16")]; + tensor concat_122 = const()[name = tensor("concat_122"), val = tensor([0, 0, 0, 0, 0])]; + tensor concat_123 = const()[name = tensor("concat_123"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_31_internal_tensor_assign_1_stride_0 = const()[name = tensor("new_cache_31_internal_tensor_assign_1_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_31_internal_tensor_assign_1_begin_mask_0 = const()[name = tensor("new_cache_31_internal_tensor_assign_1_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_31_internal_tensor_assign_1_end_mask_0 = const()[name = tensor("new_cache_31_internal_tensor_assign_1_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_31_internal_tensor_assign_1_squeeze_mask_0 = const()[name = tensor("new_cache_31_internal_tensor_assign_1_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_76 = const()[name = tensor("shape_76"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_30 = const()[name = tensor("reduce_prod_30"), val = tensor(1048576)]; + tensor range_1d_30_start_0 = const()[name = tensor("range_1d_30_start_0"), val = tensor(0)]; + tensor range_1d_30_step_0 = const()[name = tensor("range_1d_30_step_0"), val = tensor(1)]; + tensor range_1d_30 = range_1d(end = reduce_prod_30, start = range_1d_30_start_0, step = range_1d_30_step_0)[name = tensor("range_1d_30")]; + tensor reshape_150 = reshape(shape = shape_76, x = range_1d_30)[name = tensor("reshape_150")]; + tensor slice_by_index_30 = slice_by_index(begin = concat_122, begin_mask = new_cache_31_internal_tensor_assign_1_begin_mask_0, end = concat_123, end_mask = new_cache_31_internal_tensor_assign_1_end_mask_0, squeeze_mask = new_cache_31_internal_tensor_assign_1_squeeze_mask_0, stride = new_cache_31_internal_tensor_assign_1_stride_0, x = reshape_150)[name = tensor("slice_by_index_30")]; + tensor reshape_151_shape_0 = const()[name = tensor("reshape_151_shape_0"), val = tensor([-1])]; + tensor reshape_151 = reshape(shape = reshape_151_shape_0, x = slice_by_index_30)[name = tensor("reshape_151")]; + tensor reshape_152_shape_0 = const()[name = tensor("reshape_152_shape_0"), val = tensor([-1])]; + tensor reshape_152_cast_fp16 = reshape(shape = reshape_152_shape_0, x = var_6607_cast_fp16)[name = tensor("reshape_152_cast_fp16")]; + tensor reshape_153_shape_0 = const()[name = tensor("reshape_153_shape_0"), val = tensor([-1])]; + tensor reshape_153_cast_fp16 = reshape(shape = reshape_153_shape_0, x = cache15_to_fp16)[name = tensor("reshape_153_cast_fp16")]; + tensor scatter_30_mode_0 = const()[name = tensor("scatter_30_mode_0"), val = tensor("update")]; + tensor scatter_30_axis_0 = const()[name = tensor("scatter_30_axis_0"), val = tensor(0)]; + tensor scatter_30_validate_indices_0 = const()[name = tensor("scatter_30_validate_indices_0"), val = tensor(false)]; + tensor scatter_30_cast_fp16 = scatter(axis = scatter_30_axis_0, data = reshape_153_cast_fp16, indices = reshape_151, mode = scatter_30_mode_0, updates = reshape_152_cast_fp16, validate_indices = scatter_30_validate_indices_0)[name = tensor("scatter_30_cast_fp16")]; + tensor reshape_154_cast_fp16 = reshape(shape = shape_76, x = scatter_30_cast_fp16)[name = tensor("reshape_154_cast_fp16")]; + tensor var_6615_begin_0 = const()[name = tensor("op_6615_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor var_6615_end_0 = const()[name = tensor("op_6615_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor var_6615_end_mask_0 = const()[name = tensor("op_6615_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_6615_squeeze_mask_0 = const()[name = tensor("op_6615_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor var_6615_cast_fp16 = slice_by_index(begin = var_6615_begin_0, end = var_6615_end_0, end_mask = var_6615_end_mask_0, squeeze_mask = var_6615_squeeze_mask_0, x = reshape_154_cast_fp16)[name = tensor("op_6615_cast_fp16")]; + tensor var_6617_axis_0 = const()[name = tensor("op_6617_axis_0"), val = tensor(1)]; + tensor var_6617_mode_0 = const()[name = tensor("op_6617_mode_0"), val = tensor("update")]; + tensor var_6617_validate_indices_0 = const()[name = tensor("op_6617_validate_indices_0"), val = tensor(false)]; + tensor var_6617_cast_fp16 = scatter_along_axis(axis = var_6617_axis_0, data = var_6615_cast_fp16, indices = write_indices_31, mode = var_6617_mode_0, updates = v_31_cast_fp16, validate_indices = var_6617_validate_indices_0)[name = tensor("op_6617_cast_fp16")]; + tensor concat_124 = const()[name = tensor("concat_124"), val = tensor([1, 0, 0, 0, 0])]; + tensor concat_125 = const()[name = tensor("concat_125"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_31_internal_tensor_assign_2_stride_0 = const()[name = tensor("new_cache_31_internal_tensor_assign_2_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_31_internal_tensor_assign_2_begin_mask_0 = const()[name = tensor("new_cache_31_internal_tensor_assign_2_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_31_internal_tensor_assign_2_end_mask_0 = const()[name = tensor("new_cache_31_internal_tensor_assign_2_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_31_internal_tensor_assign_2_squeeze_mask_0 = const()[name = tensor("new_cache_31_internal_tensor_assign_2_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_77 = const()[name = tensor("shape_77"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_31 = const()[name = tensor("reduce_prod_31"), val = tensor(1048576)]; + tensor range_1d_31_start_0 = const()[name = tensor("range_1d_31_start_0"), val = tensor(0)]; + tensor range_1d_31_step_0 = const()[name = tensor("range_1d_31_step_0"), val = tensor(1)]; + tensor range_1d_31 = range_1d(end = reduce_prod_31, start = range_1d_31_start_0, step = range_1d_31_step_0)[name = tensor("range_1d_31")]; + tensor reshape_155 = reshape(shape = shape_77, x = range_1d_31)[name = tensor("reshape_155")]; + tensor slice_by_index_31 = slice_by_index(begin = concat_124, begin_mask = new_cache_31_internal_tensor_assign_2_begin_mask_0, end = concat_125, end_mask = new_cache_31_internal_tensor_assign_2_end_mask_0, squeeze_mask = new_cache_31_internal_tensor_assign_2_squeeze_mask_0, stride = new_cache_31_internal_tensor_assign_2_stride_0, x = reshape_155)[name = tensor("slice_by_index_31")]; + tensor reshape_156_shape_0 = const()[name = tensor("reshape_156_shape_0"), val = tensor([-1])]; + tensor reshape_156 = reshape(shape = reshape_156_shape_0, x = slice_by_index_31)[name = tensor("reshape_156")]; + tensor reshape_157_shape_0 = const()[name = tensor("reshape_157_shape_0"), val = tensor([-1])]; + tensor reshape_157_cast_fp16 = reshape(shape = reshape_157_shape_0, x = var_6617_cast_fp16)[name = tensor("reshape_157_cast_fp16")]; + tensor reshape_158_shape_0 = const()[name = tensor("reshape_158_shape_0"), val = tensor([-1])]; + tensor reshape_158_cast_fp16 = reshape(shape = reshape_158_shape_0, x = reshape_154_cast_fp16)[name = tensor("reshape_158_cast_fp16")]; + tensor scatter_31_mode_0 = const()[name = tensor("scatter_31_mode_0"), val = tensor("update")]; + tensor scatter_31_axis_0 = const()[name = tensor("scatter_31_axis_0"), val = tensor(0)]; + tensor scatter_31_validate_indices_0 = const()[name = tensor("scatter_31_validate_indices_0"), val = tensor(false)]; + tensor scatter_31_cast_fp16 = scatter(axis = scatter_31_axis_0, data = reshape_158_cast_fp16, indices = reshape_156, mode = scatter_31_mode_0, updates = reshape_157_cast_fp16, validate_indices = scatter_31_validate_indices_0)[name = tensor("scatter_31_cast_fp16")]; + tensor reshape_159_cast_fp16 = reshape(shape = shape_77, x = scatter_31_cast_fp16)[name = tensor("reshape_159_cast_fp16")]; + tensor reshape_159_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("reshape_159_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor keys_91_begin_0 = const()[name = tensor("keys_91_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor keys_91_end_0 = const()[name = tensor("keys_91_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor keys_91_end_mask_0 = const()[name = tensor("keys_91_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor keys_91_squeeze_mask_0 = const()[name = tensor("keys_91_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor keys_91_cast_fp16 = slice_by_index(begin = keys_91_begin_0, end = keys_91_end_0, end_mask = keys_91_end_mask_0, squeeze_mask = keys_91_squeeze_mask_0, x = reshape_159_cast_fp16)[name = tensor("keys_91_cast_fp16")]; + tensor values_91_begin_0 = const()[name = tensor("values_91_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor values_91_end_0 = const()[name = tensor("values_91_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor values_91_end_mask_0 = const()[name = tensor("values_91_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor values_91_squeeze_mask_0 = const()[name = tensor("values_91_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor values_91_cast_fp16 = slice_by_index(begin = values_91_begin_0, end = values_91_end_0, end_mask = values_91_end_mask_0, squeeze_mask = values_91_squeeze_mask_0, x = reshape_159_cast_fp16)[name = tensor("values_91_cast_fp16")]; + tensor var_6629_cast_fp16 = not_equal(x = keys_91_cast_fp16, y = keys_91_cast_fp16)[name = tensor("op_6629_cast_fp16")]; + tensor keys_93_cast_fp16 = select(a = var_470_to_fp16, b = keys_91_cast_fp16, cond = var_6629_cast_fp16)[name = tensor("keys_93_cast_fp16")]; + tensor var_6637_cast_fp16 = not_equal(x = values_91_cast_fp16, y = values_91_cast_fp16)[name = tensor("op_6637_cast_fp16")]; + tensor values_93_cast_fp16 = select(a = var_470_to_fp16, b = values_91_cast_fp16, cond = var_6637_cast_fp16)[name = tensor("values_93_cast_fp16")]; + tensor var_6661 = const()[name = tensor("op_6661"), val = tensor([0, 2, 1, 3])]; + tensor var_6677 = const()[name = tensor("op_6677"), val = tensor([1, 1, 1])]; + tensor var_6678_cast_fp16 = reshape(shape = var_6677, x = position15_to_fp16)[name = tensor("op_6678_cast_fp16")]; + tensor q_positions_31_cast_fp16 = add(x = var_6678_cast_fp16, y = q_offsets_1_promoted_to_fp16)[name = tensor("q_positions_31_cast_fp16")]; + tensor valid_end_31_cast_fp16 = add(x = var_6678_cast_fp16, y = var_536_cast_fp16)[name = tensor("valid_end_31_cast_fp16")]; + tensor valid_mask_31_cast_fp16 = less(x = k_positions_1_promoted_to_fp16, y = valid_end_31_cast_fp16)[name = tensor("valid_mask_31_cast_fp16")]; + tensor causal_mask_31_cast_fp16 = less_equal(x = k_positions_1_promoted_to_fp16, y = q_positions_31_cast_fp16)[name = tensor("causal_mask_31_cast_fp16")]; + tensor var_6706 = logical_and(x = valid_mask_31_cast_fp16, y = causal_mask_31_cast_fp16)[name = tensor("op_6706")]; + tensor attn_mask_31_axes_0 = const()[name = tensor("attn_mask_31_axes_0"), val = tensor([1])]; + tensor attn_mask_31 = expand_dims(axes = attn_mask_31_axes_0, x = var_6706)[name = tensor("attn_mask_31")]; + tensor var_6720_transpose_x_0 = const()[name = tensor("op_6720_transpose_x_0"), val = tensor(false)]; + tensor var_6720_transpose_y_0 = const()[name = tensor("op_6720_transpose_y_0"), val = tensor(false)]; + tensor transpose_99_perm_0 = const()[name = tensor("transpose_99_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_100_perm_0 = const()[name = tensor("transpose_100_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_100 = transpose(perm = transpose_100_perm_0, x = keys_93_cast_fp16)[name = tensor("transpose_144")]; + tensor transpose_99 = transpose(perm = transpose_99_perm_0, x = q_93_cast_fp16)[name = tensor("transpose_145")]; + tensor var_6720_cast_fp16 = matmul(transpose_x = var_6720_transpose_x_0, transpose_y = var_6720_transpose_y_0, x = transpose_99, y = transpose_100)[name = tensor("op_6720_cast_fp16")]; + tensor var_6714_to_fp16 = const()[name = tensor("op_6714_to_fp16"), val = tensor([0x1p-3])]; + tensor attn_weights_91_cast_fp16 = mul(x = var_6720_cast_fp16, y = var_6714_to_fp16)[name = tensor("attn_weights_91_cast_fp16")]; + tensor var_6722 = logical_not(x = attn_mask_31)[name = tensor("op_6722")]; + tensor var_6723_to_fp16 = const()[name = tensor("op_6723_to_fp16"), val = tensor(-inf)]; + tensor attn_weights_93_cast_fp16 = select(a = var_6723_to_fp16, b = attn_weights_91_cast_fp16, cond = var_6722)[name = tensor("attn_weights_93_cast_fp16")]; + tensor var_6725 = const()[name = tensor("op_6725"), val = tensor(-1)]; + tensor attn_weights_95_cast_fp16 = softmax(axis = var_6725, x = attn_weights_93_cast_fp16)[name = tensor("attn_weights_95_cast_fp16")]; + tensor attn_output_31_transpose_x_0 = const()[name = tensor("attn_output_31_transpose_x_0"), val = tensor(false)]; + tensor attn_output_31_transpose_y_0 = const()[name = tensor("attn_output_31_transpose_y_0"), val = tensor(false)]; + tensor values_95_cast_fp16 = transpose(perm = var_6661, x = values_93_cast_fp16)[name = tensor("transpose_146")]; + tensor attn_output_31_cast_fp16 = matmul(transpose_x = attn_output_31_transpose_x_0, transpose_y = attn_output_31_transpose_y_0, x = attn_weights_95_cast_fp16, y = values_95_cast_fp16)[name = tensor("attn_output_31_cast_fp16")]; + tensor var_6733 = const()[name = tensor("op_6733"), val = tensor([0, 2, 1, 3])]; + tensor var_6738 = const()[name = tensor("op_6738"), val = tensor([1, 256, 1024])]; + tensor var_6734_cast_fp16 = transpose(perm = var_6733, x = attn_output_31_cast_fp16)[name = tensor("transpose_143")]; + tensor input_185_cast_fp16 = reshape(shape = var_6738, x = var_6734_cast_fp16)[name = tensor("input_185_cast_fp16")]; + tensor attn15_out_proj_weight_to_fp16 = const()[name = tensor("attn15_out_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(384983808)))]; + tensor linear_61_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = attn15_out_proj_weight_to_fp16, x = input_185_cast_fp16)[name = tensor("linear_61_cast_fp16")]; + tensor var_6747_cast_fp16 = add(x = position15_to_fp16, y = valid_len_to_fp16)[name = tensor("op_6747_cast_fp16")]; + tensor var_6747_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_6747_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor input_187_cast_fp16 = add(x = input_181_cast_fp16, y = linear_61_cast_fp16)[name = tensor("input_187_cast_fp16")]; + tensor input_189_axes_0 = const()[name = tensor("input_189_axes_0"), val = tensor([-1])]; + tensor norm15_2_weight_to_fp16 = const()[name = tensor("norm15_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(387081024)))]; + tensor norm15_2_bias_to_fp16 = const()[name = tensor("norm15_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(387083136)))]; + tensor var_6751_to_fp16 = const()[name = tensor("op_6751_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_189_cast_fp16 = layer_norm(axes = input_189_axes_0, beta = norm15_2_bias_to_fp16, epsilon = var_6751_to_fp16, gamma = norm15_2_weight_to_fp16, x = input_187_cast_fp16)[name = tensor("input_189_cast_fp16")]; + tensor linear15_1_weight_to_fp16 = const()[name = tensor("linear15_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(387085248)))]; + tensor linear_62_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = linear15_1_weight_to_fp16, x = input_189_cast_fp16)[name = tensor("linear_62_cast_fp16")]; + tensor input_191_mode_0 = const()[name = tensor("input_191_mode_0"), val = tensor("EXACT")]; + tensor input_191_cast_fp16 = gelu(mode = input_191_mode_0, x = linear_62_cast_fp16)[name = tensor("input_191_cast_fp16")]; + tensor linear15_2_weight_to_fp16 = const()[name = tensor("linear15_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(395473920)))]; + tensor linear_63_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = linear15_2_weight_to_fp16, x = input_191_cast_fp16)[name = tensor("linear_63_cast_fp16")]; + tensor input_193_cast_fp16 = add(x = input_187_cast_fp16, y = linear_63_cast_fp16)[name = tensor("input_193_cast_fp16")]; + tensor input_195_axes_0 = const()[name = tensor("input_195_axes_0"), val = tensor([-1])]; + tensor norm16_1_weight_to_fp16 = const()[name = tensor("norm16_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(403862592)))]; + tensor norm16_1_bias_to_fp16 = const()[name = tensor("norm16_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(403864704)))]; + tensor var_6768_to_fp16 = const()[name = tensor("op_6768_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_195_cast_fp16 = layer_norm(axes = input_195_axes_0, beta = norm16_1_bias_to_fp16, epsilon = var_6768_to_fp16, gamma = norm16_1_weight_to_fp16, x = input_193_cast_fp16)[name = tensor("input_195_cast_fp16")]; + tensor attn16_in_proj_weight_to_fp16 = const()[name = tensor("attn16_in_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(403866816)))]; + tensor linear_64_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = attn16_in_proj_weight_to_fp16, x = input_195_cast_fp16)[name = tensor("linear_64_cast_fp16")]; + tensor var_6782 = const()[name = tensor("op_6782"), val = tensor([1, 256, 3, 16, 64])]; + tensor qkv_33_cast_fp16 = reshape(shape = var_6782, x = linear_64_cast_fp16)[name = tensor("qkv_33_cast_fp16")]; + tensor q_97_begin_0 = const()[name = tensor("q_97_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor q_97_end_0 = const()[name = tensor("q_97_end_0"), val = tensor([1, 256, 1, 16, 64])]; + tensor q_97_end_mask_0 = const()[name = tensor("q_97_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor q_97_squeeze_mask_0 = const()[name = tensor("q_97_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor q_97_cast_fp16 = slice_by_index(begin = q_97_begin_0, end = q_97_end_0, end_mask = q_97_end_mask_0, squeeze_mask = q_97_squeeze_mask_0, x = qkv_33_cast_fp16)[name = tensor("q_97_cast_fp16")]; + tensor k_65_begin_0 = const()[name = tensor("k_65_begin_0"), val = tensor([0, 0, 1, 0, 0])]; + tensor k_65_end_0 = const()[name = tensor("k_65_end_0"), val = tensor([1, 256, 2, 16, 64])]; + tensor k_65_end_mask_0 = const()[name = tensor("k_65_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor k_65_squeeze_mask_0 = const()[name = tensor("k_65_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor k_65_cast_fp16 = slice_by_index(begin = k_65_begin_0, end = k_65_end_0, end_mask = k_65_end_mask_0, squeeze_mask = k_65_squeeze_mask_0, x = qkv_33_cast_fp16)[name = tensor("k_65_cast_fp16")]; + tensor v_33_begin_0 = const()[name = tensor("v_33_begin_0"), val = tensor([0, 0, 2, 0, 0])]; + tensor v_33_end_0 = const()[name = tensor("v_33_end_0"), val = tensor([1, 256, 3, 16, 64])]; + tensor v_33_end_mask_0 = const()[name = tensor("v_33_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor v_33_squeeze_mask_0 = const()[name = tensor("v_33_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor v_33_cast_fp16 = slice_by_index(begin = v_33_begin_0, end = v_33_end_0, end_mask = v_33_end_mask_0, squeeze_mask = v_33_squeeze_mask_0, x = qkv_33_cast_fp16)[name = tensor("v_33_cast_fp16")]; + tensor var_6840 = const()[name = tensor("op_6840"), val = tensor([1, 1])]; + tensor position16_to_fp16_dtype_0 = const()[name = tensor("position16_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor position16_to_fp16 = cast(dtype = position16_to_fp16_dtype_0, x = position16)[name = tensor("cast_581")]; + tensor var_6841_cast_fp16 = reshape(shape = var_6840, x = position16_to_fp16)[name = tensor("op_6841_cast_fp16")]; + tensor ts_99_cast_fp16 = add(x = ts_1_promoted_to_fp16, y = var_6841_cast_fp16)[name = tensor("ts_99_cast_fp16")]; + tensor var_6848 = const()[name = tensor("op_6848"), val = tensor([1, 256, 1, 1])]; + tensor ts_101_cast_fp16 = reshape(shape = var_6848, x = ts_99_cast_fp16)[name = tensor("ts_101_cast_fp16")]; + tensor var_6855 = const()[name = tensor("op_6855"), val = tensor([1, 256, 16, 32, 2])]; + tensor q_complex_33_cast_fp16 = reshape(shape = var_6855, x = q_97_cast_fp16)[name = tensor("q_complex_33_cast_fp16")]; + tensor var_6862 = const()[name = tensor("op_6862"), val = tensor([1, 256, 16, 32, 2])]; + tensor k_complex_33_cast_fp16 = reshape(shape = var_6862, x = k_65_cast_fp16)[name = tensor("k_complex_33_cast_fp16")]; + tensor var_6866_begin_0 = const()[name = tensor("op_6866_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_6866_end_0 = const()[name = tensor("op_6866_end_0"), val = tensor([1, 256, 16, 32, 1])]; + tensor var_6866_end_mask_0 = const()[name = tensor("op_6866_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_6866_squeeze_mask_0 = const()[name = tensor("op_6866_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_6866_cast_fp16 = slice_by_index(begin = var_6866_begin_0, end = var_6866_end_0, end_mask = var_6866_end_mask_0, squeeze_mask = var_6866_squeeze_mask_0, x = q_complex_33_cast_fp16)[name = tensor("op_6866_cast_fp16")]; + tensor var_6874_begin_0 = const()[name = tensor("op_6874_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_6874_end_0 = const()[name = tensor("op_6874_end_0"), val = tensor([1, 256, 16, 32, 2])]; + tensor var_6874_end_mask_0 = const()[name = tensor("op_6874_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_6874_squeeze_mask_0 = const()[name = tensor("op_6874_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_6874_cast_fp16 = slice_by_index(begin = var_6874_begin_0, end = var_6874_end_0, end_mask = var_6874_end_mask_0, squeeze_mask = var_6874_squeeze_mask_0, x = q_complex_33_cast_fp16)[name = tensor("op_6874_cast_fp16")]; + tensor var_6882_begin_0 = const()[name = tensor("op_6882_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_6882_end_0 = const()[name = tensor("op_6882_end_0"), val = tensor([1, 256, 16, 32, 1])]; + tensor var_6882_end_mask_0 = const()[name = tensor("op_6882_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_6882_squeeze_mask_0 = const()[name = tensor("op_6882_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_6882_cast_fp16 = slice_by_index(begin = var_6882_begin_0, end = var_6882_end_0, end_mask = var_6882_end_mask_0, squeeze_mask = var_6882_squeeze_mask_0, x = k_complex_33_cast_fp16)[name = tensor("op_6882_cast_fp16")]; + tensor var_6890_begin_0 = const()[name = tensor("op_6890_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_6890_end_0 = const()[name = tensor("op_6890_end_0"), val = tensor([1, 256, 16, 32, 2])]; + tensor var_6890_end_mask_0 = const()[name = tensor("op_6890_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_6890_squeeze_mask_0 = const()[name = tensor("op_6890_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_6890_cast_fp16 = slice_by_index(begin = var_6890_begin_0, end = var_6890_end_0, end_mask = var_6890_end_mask_0, squeeze_mask = var_6890_squeeze_mask_0, x = k_complex_33_cast_fp16)[name = tensor("op_6890_cast_fp16")]; + tensor freqs_33_to_fp16 = const()[name = tensor("freqs_33_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(410158336)))]; + tensor var_6896_cast_fp16 = mul(x = freqs_33_to_fp16, y = ts_101_cast_fp16)[name = tensor("op_6896_cast_fp16")]; + tensor rotr_33_cast_fp16 = cos(x = var_6896_cast_fp16)[name = tensor("rotr_33_cast_fp16")]; + tensor roti_33_cast_fp16 = sin(x = var_6896_cast_fp16)[name = tensor("roti_33_cast_fp16")]; + tensor var_6900_cast_fp16 = mul(x = var_6866_cast_fp16, y = rotr_33_cast_fp16)[name = tensor("op_6900_cast_fp16")]; + tensor var_6901_cast_fp16 = mul(x = var_6874_cast_fp16, y = roti_33_cast_fp16)[name = tensor("op_6901_cast_fp16")]; + tensor qor_65_cast_fp16 = sub(x = var_6900_cast_fp16, y = var_6901_cast_fp16)[name = tensor("qor_65_cast_fp16")]; + tensor var_6904_cast_fp16 = mul(x = var_6866_cast_fp16, y = roti_33_cast_fp16)[name = tensor("op_6904_cast_fp16")]; + tensor var_6905_cast_fp16 = mul(x = var_6874_cast_fp16, y = rotr_33_cast_fp16)[name = tensor("op_6905_cast_fp16")]; + tensor qoi_65_cast_fp16 = add(x = var_6904_cast_fp16, y = var_6905_cast_fp16)[name = tensor("qoi_65_cast_fp16")]; + tensor var_6908_cast_fp16 = mul(x = var_6882_cast_fp16, y = rotr_33_cast_fp16)[name = tensor("op_6908_cast_fp16")]; + tensor var_6909_cast_fp16 = mul(x = var_6890_cast_fp16, y = roti_33_cast_fp16)[name = tensor("op_6909_cast_fp16")]; + tensor kor_65_cast_fp16 = sub(x = var_6908_cast_fp16, y = var_6909_cast_fp16)[name = tensor("kor_65_cast_fp16")]; + tensor var_6912_cast_fp16 = mul(x = var_6882_cast_fp16, y = roti_33_cast_fp16)[name = tensor("op_6912_cast_fp16")]; + tensor var_6913_cast_fp16 = mul(x = var_6890_cast_fp16, y = rotr_33_cast_fp16)[name = tensor("op_6913_cast_fp16")]; + tensor koi_65_cast_fp16 = add(x = var_6912_cast_fp16, y = var_6913_cast_fp16)[name = tensor("koi_65_cast_fp16")]; + tensor qo_33_axis_0 = const()[name = tensor("qo_33_axis_0"), val = tensor(-1)]; + tensor qo_33_cast_fp16 = stack(axis = qo_33_axis_0, values = (qor_65_cast_fp16, qoi_65_cast_fp16))[name = tensor("qo_33_cast_fp16")]; + tensor ko_33_axis_0 = const()[name = tensor("ko_33_axis_0"), val = tensor(-1)]; + tensor ko_33_cast_fp16 = stack(axis = ko_33_axis_0, values = (kor_65_cast_fp16, koi_65_cast_fp16))[name = tensor("ko_33_cast_fp16")]; + tensor var_6946 = const()[name = tensor("op_6946"), val = tensor([1, 256, 16, 64])]; + tensor q_99_cast_fp16 = reshape(shape = var_6946, x = qo_33_cast_fp16)[name = tensor("q_99_cast_fp16")]; + tensor var_6952 = const()[name = tensor("op_6952"), val = tensor([1, 256, 16, 64])]; + tensor k_67_cast_fp16 = reshape(shape = var_6952, x = ko_33_cast_fp16)[name = tensor("k_67_cast_fp16")]; + tensor write_indices_float_97_cast_fp16 = add(x = var_6841_cast_fp16, y = var_393_promoted_to_fp16)[name = tensor("write_indices_float_97_cast_fp16")]; + tensor _inversed_6973_y_0_to_fp16 = const()[name = tensor("_inversed_6973_y_0_to_fp16"), val = tensor(0x1p-9)]; + tensor _inversed_6973_cast_fp16 = mul(x = write_indices_float_97_cast_fp16, y = _inversed_6973_y_0_to_fp16)[name = tensor("_inversed_6973_cast_fp16")]; + tensor var_6974_cast_fp16 = floor(x = _inversed_6973_cast_fp16)[name = tensor("op_6974_cast_fp16")]; + tensor var_6975_to_fp16 = const()[name = tensor("op_6975_to_fp16"), val = tensor(0x1p+9)]; + tensor var_6976_cast_fp16 = mul(x = var_6974_cast_fp16, y = var_6975_to_fp16)[name = tensor("op_6976_cast_fp16")]; + tensor write_indices_float_99_cast_fp16 = sub(x = write_indices_float_97_cast_fp16, y = var_6976_cast_fp16)[name = tensor("write_indices_float_99_cast_fp16")]; + tensor var_6994_value_0_to_fp16 = const()[name = tensor("op_6994_value_0_to_fp16"), val = tensor(0x1.ffp+8)]; + tensor var_6994_cast_fp16 = fill_like(ref_tensor = write_indices_float_99_cast_fp16, value = var_6994_value_0_to_fp16)[name = tensor("op_6994_cast_fp16")]; + tensor write_indices_float_101_cast_fp16 = select(a = write_indices_float_99_cast_fp16, b = var_6994_cast_fp16, cond = token_valid_1_cast_fp16)[name = tensor("write_indices_float_101_cast_fp16")]; + tensor var_7000_dtype_0 = const()[name = tensor("op_7000_dtype_0"), val = tensor("int32")]; + tensor var_7005 = const()[name = tensor("op_7005"), val = tensor([1, 256, 1, 1])]; + tensor write_indices_float_101_cast_fp16_to_int32 = cast(dtype = var_7000_dtype_0, x = write_indices_float_101_cast_fp16)[name = tensor("cast_580")]; + tensor var_7006 = reshape(shape = var_7005, x = write_indices_float_101_cast_fp16_to_int32)[name = tensor("op_7006")]; + tensor write_indices_33_reps_0 = const()[name = tensor("write_indices_33_reps_0"), val = tensor([1, 1, 16, 64])]; + tensor write_indices_33 = tile(reps = write_indices_33_reps_0, x = var_7006)[name = tensor("write_indices_33")]; + tensor var_7016_begin_0 = const()[name = tensor("op_7016_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_7016_end_0 = const()[name = tensor("op_7016_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor var_7016_end_mask_0 = const()[name = tensor("op_7016_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_7016_squeeze_mask_0 = const()[name = tensor("op_7016_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor cache16_to_fp16_dtype_0 = const()[name = tensor("cache16_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor cache16_to_fp16 = cast(dtype = cache16_to_fp16_dtype_0, x = cache16)[name = tensor("cast_579")]; + tensor var_7016_cast_fp16 = slice_by_index(begin = var_7016_begin_0, end = var_7016_end_0, end_mask = var_7016_end_mask_0, squeeze_mask = var_7016_squeeze_mask_0, x = cache16_to_fp16)[name = tensor("op_7016_cast_fp16")]; + tensor var_7018_axis_0 = const()[name = tensor("op_7018_axis_0"), val = tensor(1)]; + tensor var_7018_mode_0 = const()[name = tensor("op_7018_mode_0"), val = tensor("update")]; + tensor var_7018_validate_indices_0 = const()[name = tensor("op_7018_validate_indices_0"), val = tensor(false)]; + tensor var_7018_cast_fp16 = scatter_along_axis(axis = var_7018_axis_0, data = var_7016_cast_fp16, indices = write_indices_33, mode = var_7018_mode_0, updates = k_67_cast_fp16, validate_indices = var_7018_validate_indices_0)[name = tensor("op_7018_cast_fp16")]; + tensor concat_130 = const()[name = tensor("concat_130"), val = tensor([0, 0, 0, 0, 0])]; + tensor concat_131 = const()[name = tensor("concat_131"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_33_internal_tensor_assign_1_stride_0 = const()[name = tensor("new_cache_33_internal_tensor_assign_1_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_33_internal_tensor_assign_1_begin_mask_0 = const()[name = tensor("new_cache_33_internal_tensor_assign_1_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_33_internal_tensor_assign_1_end_mask_0 = const()[name = tensor("new_cache_33_internal_tensor_assign_1_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_33_internal_tensor_assign_1_squeeze_mask_0 = const()[name = tensor("new_cache_33_internal_tensor_assign_1_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_78 = const()[name = tensor("shape_78"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_32 = const()[name = tensor("reduce_prod_32"), val = tensor(1048576)]; + tensor range_1d_32_start_0 = const()[name = tensor("range_1d_32_start_0"), val = tensor(0)]; + tensor range_1d_32_step_0 = const()[name = tensor("range_1d_32_step_0"), val = tensor(1)]; + tensor range_1d_32 = range_1d(end = reduce_prod_32, start = range_1d_32_start_0, step = range_1d_32_step_0)[name = tensor("range_1d_32")]; + tensor reshape_160 = reshape(shape = shape_78, x = range_1d_32)[name = tensor("reshape_160")]; + tensor slice_by_index_32 = slice_by_index(begin = concat_130, begin_mask = new_cache_33_internal_tensor_assign_1_begin_mask_0, end = concat_131, end_mask = new_cache_33_internal_tensor_assign_1_end_mask_0, squeeze_mask = new_cache_33_internal_tensor_assign_1_squeeze_mask_0, stride = new_cache_33_internal_tensor_assign_1_stride_0, x = reshape_160)[name = tensor("slice_by_index_32")]; + tensor reshape_161_shape_0 = const()[name = tensor("reshape_161_shape_0"), val = tensor([-1])]; + tensor reshape_161 = reshape(shape = reshape_161_shape_0, x = slice_by_index_32)[name = tensor("reshape_161")]; + tensor reshape_162_shape_0 = const()[name = tensor("reshape_162_shape_0"), val = tensor([-1])]; + tensor reshape_162_cast_fp16 = reshape(shape = reshape_162_shape_0, x = var_7018_cast_fp16)[name = tensor("reshape_162_cast_fp16")]; + tensor reshape_163_shape_0 = const()[name = tensor("reshape_163_shape_0"), val = tensor([-1])]; + tensor reshape_163_cast_fp16 = reshape(shape = reshape_163_shape_0, x = cache16_to_fp16)[name = tensor("reshape_163_cast_fp16")]; + tensor scatter_32_mode_0 = const()[name = tensor("scatter_32_mode_0"), val = tensor("update")]; + tensor scatter_32_axis_0 = const()[name = tensor("scatter_32_axis_0"), val = tensor(0)]; + tensor scatter_32_validate_indices_0 = const()[name = tensor("scatter_32_validate_indices_0"), val = tensor(false)]; + tensor scatter_32_cast_fp16 = scatter(axis = scatter_32_axis_0, data = reshape_163_cast_fp16, indices = reshape_161, mode = scatter_32_mode_0, updates = reshape_162_cast_fp16, validate_indices = scatter_32_validate_indices_0)[name = tensor("scatter_32_cast_fp16")]; + tensor reshape_164_cast_fp16 = reshape(shape = shape_78, x = scatter_32_cast_fp16)[name = tensor("reshape_164_cast_fp16")]; + tensor var_7026_begin_0 = const()[name = tensor("op_7026_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor var_7026_end_0 = const()[name = tensor("op_7026_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor var_7026_end_mask_0 = const()[name = tensor("op_7026_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_7026_squeeze_mask_0 = const()[name = tensor("op_7026_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor var_7026_cast_fp16 = slice_by_index(begin = var_7026_begin_0, end = var_7026_end_0, end_mask = var_7026_end_mask_0, squeeze_mask = var_7026_squeeze_mask_0, x = reshape_164_cast_fp16)[name = tensor("op_7026_cast_fp16")]; + tensor var_7028_axis_0 = const()[name = tensor("op_7028_axis_0"), val = tensor(1)]; + tensor var_7028_mode_0 = const()[name = tensor("op_7028_mode_0"), val = tensor("update")]; + tensor var_7028_validate_indices_0 = const()[name = tensor("op_7028_validate_indices_0"), val = tensor(false)]; + tensor var_7028_cast_fp16 = scatter_along_axis(axis = var_7028_axis_0, data = var_7026_cast_fp16, indices = write_indices_33, mode = var_7028_mode_0, updates = v_33_cast_fp16, validate_indices = var_7028_validate_indices_0)[name = tensor("op_7028_cast_fp16")]; + tensor concat_132 = const()[name = tensor("concat_132"), val = tensor([1, 0, 0, 0, 0])]; + tensor concat_133 = const()[name = tensor("concat_133"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_33_internal_tensor_assign_2_stride_0 = const()[name = tensor("new_cache_33_internal_tensor_assign_2_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_33_internal_tensor_assign_2_begin_mask_0 = const()[name = tensor("new_cache_33_internal_tensor_assign_2_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_33_internal_tensor_assign_2_end_mask_0 = const()[name = tensor("new_cache_33_internal_tensor_assign_2_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_33_internal_tensor_assign_2_squeeze_mask_0 = const()[name = tensor("new_cache_33_internal_tensor_assign_2_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_79 = const()[name = tensor("shape_79"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_33 = const()[name = tensor("reduce_prod_33"), val = tensor(1048576)]; + tensor range_1d_33_start_0 = const()[name = tensor("range_1d_33_start_0"), val = tensor(0)]; + tensor range_1d_33_step_0 = const()[name = tensor("range_1d_33_step_0"), val = tensor(1)]; + tensor range_1d_33 = range_1d(end = reduce_prod_33, start = range_1d_33_start_0, step = range_1d_33_step_0)[name = tensor("range_1d_33")]; + tensor reshape_165 = reshape(shape = shape_79, x = range_1d_33)[name = tensor("reshape_165")]; + tensor slice_by_index_33 = slice_by_index(begin = concat_132, begin_mask = new_cache_33_internal_tensor_assign_2_begin_mask_0, end = concat_133, end_mask = new_cache_33_internal_tensor_assign_2_end_mask_0, squeeze_mask = new_cache_33_internal_tensor_assign_2_squeeze_mask_0, stride = new_cache_33_internal_tensor_assign_2_stride_0, x = reshape_165)[name = tensor("slice_by_index_33")]; + tensor reshape_166_shape_0 = const()[name = tensor("reshape_166_shape_0"), val = tensor([-1])]; + tensor reshape_166 = reshape(shape = reshape_166_shape_0, x = slice_by_index_33)[name = tensor("reshape_166")]; + tensor reshape_167_shape_0 = const()[name = tensor("reshape_167_shape_0"), val = tensor([-1])]; + tensor reshape_167_cast_fp16 = reshape(shape = reshape_167_shape_0, x = var_7028_cast_fp16)[name = tensor("reshape_167_cast_fp16")]; + tensor reshape_168_shape_0 = const()[name = tensor("reshape_168_shape_0"), val = tensor([-1])]; + tensor reshape_168_cast_fp16 = reshape(shape = reshape_168_shape_0, x = reshape_164_cast_fp16)[name = tensor("reshape_168_cast_fp16")]; + tensor scatter_33_mode_0 = const()[name = tensor("scatter_33_mode_0"), val = tensor("update")]; + tensor scatter_33_axis_0 = const()[name = tensor("scatter_33_axis_0"), val = tensor(0)]; + tensor scatter_33_validate_indices_0 = const()[name = tensor("scatter_33_validate_indices_0"), val = tensor(false)]; + tensor scatter_33_cast_fp16 = scatter(axis = scatter_33_axis_0, data = reshape_168_cast_fp16, indices = reshape_166, mode = scatter_33_mode_0, updates = reshape_167_cast_fp16, validate_indices = scatter_33_validate_indices_0)[name = tensor("scatter_33_cast_fp16")]; + tensor reshape_169_cast_fp16 = reshape(shape = shape_79, x = scatter_33_cast_fp16)[name = tensor("reshape_169_cast_fp16")]; + tensor reshape_169_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("reshape_169_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor keys_97_begin_0 = const()[name = tensor("keys_97_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor keys_97_end_0 = const()[name = tensor("keys_97_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor keys_97_end_mask_0 = const()[name = tensor("keys_97_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor keys_97_squeeze_mask_0 = const()[name = tensor("keys_97_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor keys_97_cast_fp16 = slice_by_index(begin = keys_97_begin_0, end = keys_97_end_0, end_mask = keys_97_end_mask_0, squeeze_mask = keys_97_squeeze_mask_0, x = reshape_169_cast_fp16)[name = tensor("keys_97_cast_fp16")]; + tensor values_97_begin_0 = const()[name = tensor("values_97_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor values_97_end_0 = const()[name = tensor("values_97_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor values_97_end_mask_0 = const()[name = tensor("values_97_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor values_97_squeeze_mask_0 = const()[name = tensor("values_97_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor values_97_cast_fp16 = slice_by_index(begin = values_97_begin_0, end = values_97_end_0, end_mask = values_97_end_mask_0, squeeze_mask = values_97_squeeze_mask_0, x = reshape_169_cast_fp16)[name = tensor("values_97_cast_fp16")]; + tensor var_7040_cast_fp16 = not_equal(x = keys_97_cast_fp16, y = keys_97_cast_fp16)[name = tensor("op_7040_cast_fp16")]; + tensor keys_99_cast_fp16 = select(a = var_470_to_fp16, b = keys_97_cast_fp16, cond = var_7040_cast_fp16)[name = tensor("keys_99_cast_fp16")]; + tensor var_7048_cast_fp16 = not_equal(x = values_97_cast_fp16, y = values_97_cast_fp16)[name = tensor("op_7048_cast_fp16")]; + tensor values_99_cast_fp16 = select(a = var_470_to_fp16, b = values_97_cast_fp16, cond = var_7048_cast_fp16)[name = tensor("values_99_cast_fp16")]; + tensor var_7072 = const()[name = tensor("op_7072"), val = tensor([0, 2, 1, 3])]; + tensor var_7088 = const()[name = tensor("op_7088"), val = tensor([1, 1, 1])]; + tensor var_7089_cast_fp16 = reshape(shape = var_7088, x = position16_to_fp16)[name = tensor("op_7089_cast_fp16")]; + tensor q_positions_33_cast_fp16 = add(x = var_7089_cast_fp16, y = q_offsets_1_promoted_to_fp16)[name = tensor("q_positions_33_cast_fp16")]; + tensor valid_end_33_cast_fp16 = add(x = var_7089_cast_fp16, y = var_536_cast_fp16)[name = tensor("valid_end_33_cast_fp16")]; + tensor valid_mask_33_cast_fp16 = less(x = k_positions_1_promoted_to_fp16, y = valid_end_33_cast_fp16)[name = tensor("valid_mask_33_cast_fp16")]; + tensor causal_mask_33_cast_fp16 = less_equal(x = k_positions_1_promoted_to_fp16, y = q_positions_33_cast_fp16)[name = tensor("causal_mask_33_cast_fp16")]; + tensor var_7117 = logical_and(x = valid_mask_33_cast_fp16, y = causal_mask_33_cast_fp16)[name = tensor("op_7117")]; + tensor attn_mask_33_axes_0 = const()[name = tensor("attn_mask_33_axes_0"), val = tensor([1])]; + tensor attn_mask_33 = expand_dims(axes = attn_mask_33_axes_0, x = var_7117)[name = tensor("attn_mask_33")]; + tensor var_7131_transpose_x_0 = const()[name = tensor("op_7131_transpose_x_0"), val = tensor(false)]; + tensor var_7131_transpose_y_0 = const()[name = tensor("op_7131_transpose_y_0"), val = tensor(false)]; + tensor transpose_101_perm_0 = const()[name = tensor("transpose_101_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_102_perm_0 = const()[name = tensor("transpose_102_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_102 = transpose(perm = transpose_102_perm_0, x = keys_99_cast_fp16)[name = tensor("transpose_140")]; + tensor transpose_101 = transpose(perm = transpose_101_perm_0, x = q_99_cast_fp16)[name = tensor("transpose_141")]; + tensor var_7131_cast_fp16 = matmul(transpose_x = var_7131_transpose_x_0, transpose_y = var_7131_transpose_y_0, x = transpose_101, y = transpose_102)[name = tensor("op_7131_cast_fp16")]; + tensor var_7125_to_fp16 = const()[name = tensor("op_7125_to_fp16"), val = tensor([0x1p-3])]; + tensor attn_weights_97_cast_fp16 = mul(x = var_7131_cast_fp16, y = var_7125_to_fp16)[name = tensor("attn_weights_97_cast_fp16")]; + tensor var_7133 = logical_not(x = attn_mask_33)[name = tensor("op_7133")]; + tensor var_7134_to_fp16 = const()[name = tensor("op_7134_to_fp16"), val = tensor(-inf)]; + tensor attn_weights_99_cast_fp16 = select(a = var_7134_to_fp16, b = attn_weights_97_cast_fp16, cond = var_7133)[name = tensor("attn_weights_99_cast_fp16")]; + tensor var_7136 = const()[name = tensor("op_7136"), val = tensor(-1)]; + tensor attn_weights_101_cast_fp16 = softmax(axis = var_7136, x = attn_weights_99_cast_fp16)[name = tensor("attn_weights_101_cast_fp16")]; + tensor attn_output_33_transpose_x_0 = const()[name = tensor("attn_output_33_transpose_x_0"), val = tensor(false)]; + tensor attn_output_33_transpose_y_0 = const()[name = tensor("attn_output_33_transpose_y_0"), val = tensor(false)]; + tensor values_101_cast_fp16 = transpose(perm = var_7072, x = values_99_cast_fp16)[name = tensor("transpose_142")]; + tensor attn_output_33_cast_fp16 = matmul(transpose_x = attn_output_33_transpose_x_0, transpose_y = attn_output_33_transpose_y_0, x = attn_weights_101_cast_fp16, y = values_101_cast_fp16)[name = tensor("attn_output_33_cast_fp16")]; + tensor var_7144 = const()[name = tensor("op_7144"), val = tensor([0, 2, 1, 3])]; + tensor var_7149 = const()[name = tensor("op_7149"), val = tensor([1, 256, 1024])]; + tensor var_7145_cast_fp16 = transpose(perm = var_7144, x = attn_output_33_cast_fp16)[name = tensor("transpose_139")]; + tensor input_197_cast_fp16 = reshape(shape = var_7149, x = var_7145_cast_fp16)[name = tensor("input_197_cast_fp16")]; + tensor attn16_out_proj_weight_to_fp16 = const()[name = tensor("attn16_out_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(410158464)))]; + tensor linear_65_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = attn16_out_proj_weight_to_fp16, x = input_197_cast_fp16)[name = tensor("linear_65_cast_fp16")]; + tensor var_7158_cast_fp16 = add(x = position16_to_fp16, y = valid_len_to_fp16)[name = tensor("op_7158_cast_fp16")]; + tensor var_7158_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_7158_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor input_199_cast_fp16 = add(x = input_193_cast_fp16, y = linear_65_cast_fp16)[name = tensor("input_199_cast_fp16")]; + tensor input_201_axes_0 = const()[name = tensor("input_201_axes_0"), val = tensor([-1])]; + tensor norm16_2_weight_to_fp16 = const()[name = tensor("norm16_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(412255680)))]; + tensor norm16_2_bias_to_fp16 = const()[name = tensor("norm16_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(412257792)))]; + tensor var_7162_to_fp16 = const()[name = tensor("op_7162_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_201_cast_fp16 = layer_norm(axes = input_201_axes_0, beta = norm16_2_bias_to_fp16, epsilon = var_7162_to_fp16, gamma = norm16_2_weight_to_fp16, x = input_199_cast_fp16)[name = tensor("input_201_cast_fp16")]; + tensor linear16_1_weight_to_fp16 = const()[name = tensor("linear16_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(412259904)))]; + tensor linear_66_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = linear16_1_weight_to_fp16, x = input_201_cast_fp16)[name = tensor("linear_66_cast_fp16")]; + tensor input_203_mode_0 = const()[name = tensor("input_203_mode_0"), val = tensor("EXACT")]; + tensor input_203_cast_fp16 = gelu(mode = input_203_mode_0, x = linear_66_cast_fp16)[name = tensor("input_203_cast_fp16")]; + tensor linear16_2_weight_to_fp16 = const()[name = tensor("linear16_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(420648576)))]; + tensor linear_67_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = linear16_2_weight_to_fp16, x = input_203_cast_fp16)[name = tensor("linear_67_cast_fp16")]; + tensor input_205_cast_fp16 = add(x = input_199_cast_fp16, y = linear_67_cast_fp16)[name = tensor("input_205_cast_fp16")]; + tensor input_207_axes_0 = const()[name = tensor("input_207_axes_0"), val = tensor([-1])]; + tensor norm17_1_weight_to_fp16 = const()[name = tensor("norm17_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(429037248)))]; + tensor norm17_1_bias_to_fp16 = const()[name = tensor("norm17_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(429039360)))]; + tensor var_7179_to_fp16 = const()[name = tensor("op_7179_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_207_cast_fp16 = layer_norm(axes = input_207_axes_0, beta = norm17_1_bias_to_fp16, epsilon = var_7179_to_fp16, gamma = norm17_1_weight_to_fp16, x = input_205_cast_fp16)[name = tensor("input_207_cast_fp16")]; + tensor attn17_in_proj_weight_to_fp16 = const()[name = tensor("attn17_in_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(429041472)))]; + tensor linear_68_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = attn17_in_proj_weight_to_fp16, x = input_207_cast_fp16)[name = tensor("linear_68_cast_fp16")]; + tensor var_7193 = const()[name = tensor("op_7193"), val = tensor([1, 256, 3, 16, 64])]; + tensor qkv_35_cast_fp16 = reshape(shape = var_7193, x = linear_68_cast_fp16)[name = tensor("qkv_35_cast_fp16")]; + tensor q_103_begin_0 = const()[name = tensor("q_103_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor q_103_end_0 = const()[name = tensor("q_103_end_0"), val = tensor([1, 256, 1, 16, 64])]; + tensor q_103_end_mask_0 = const()[name = tensor("q_103_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor q_103_squeeze_mask_0 = const()[name = tensor("q_103_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor q_103_cast_fp16 = slice_by_index(begin = q_103_begin_0, end = q_103_end_0, end_mask = q_103_end_mask_0, squeeze_mask = q_103_squeeze_mask_0, x = qkv_35_cast_fp16)[name = tensor("q_103_cast_fp16")]; + tensor k_69_begin_0 = const()[name = tensor("k_69_begin_0"), val = tensor([0, 0, 1, 0, 0])]; + tensor k_69_end_0 = const()[name = tensor("k_69_end_0"), val = tensor([1, 256, 2, 16, 64])]; + tensor k_69_end_mask_0 = const()[name = tensor("k_69_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor k_69_squeeze_mask_0 = const()[name = tensor("k_69_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor k_69_cast_fp16 = slice_by_index(begin = k_69_begin_0, end = k_69_end_0, end_mask = k_69_end_mask_0, squeeze_mask = k_69_squeeze_mask_0, x = qkv_35_cast_fp16)[name = tensor("k_69_cast_fp16")]; + tensor v_35_begin_0 = const()[name = tensor("v_35_begin_0"), val = tensor([0, 0, 2, 0, 0])]; + tensor v_35_end_0 = const()[name = tensor("v_35_end_0"), val = tensor([1, 256, 3, 16, 64])]; + tensor v_35_end_mask_0 = const()[name = tensor("v_35_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor v_35_squeeze_mask_0 = const()[name = tensor("v_35_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor v_35_cast_fp16 = slice_by_index(begin = v_35_begin_0, end = v_35_end_0, end_mask = v_35_end_mask_0, squeeze_mask = v_35_squeeze_mask_0, x = qkv_35_cast_fp16)[name = tensor("v_35_cast_fp16")]; + tensor var_7251 = const()[name = tensor("op_7251"), val = tensor([1, 1])]; + tensor position17_to_fp16_dtype_0 = const()[name = tensor("position17_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor position17_to_fp16 = cast(dtype = position17_to_fp16_dtype_0, x = position17)[name = tensor("cast_576")]; + tensor var_7252_cast_fp16 = reshape(shape = var_7251, x = position17_to_fp16)[name = tensor("op_7252_cast_fp16")]; + tensor ts_105_cast_fp16 = add(x = ts_1_promoted_to_fp16, y = var_7252_cast_fp16)[name = tensor("ts_105_cast_fp16")]; + tensor var_7259 = const()[name = tensor("op_7259"), val = tensor([1, 256, 1, 1])]; + tensor ts_107_cast_fp16 = reshape(shape = var_7259, x = ts_105_cast_fp16)[name = tensor("ts_107_cast_fp16")]; + tensor var_7266 = const()[name = tensor("op_7266"), val = tensor([1, 256, 16, 32, 2])]; + tensor q_complex_35_cast_fp16 = reshape(shape = var_7266, x = q_103_cast_fp16)[name = tensor("q_complex_35_cast_fp16")]; + tensor var_7273 = const()[name = tensor("op_7273"), val = tensor([1, 256, 16, 32, 2])]; + tensor k_complex_35_cast_fp16 = reshape(shape = var_7273, x = k_69_cast_fp16)[name = tensor("k_complex_35_cast_fp16")]; + tensor var_7277_begin_0 = const()[name = tensor("op_7277_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_7277_end_0 = const()[name = tensor("op_7277_end_0"), val = tensor([1, 256, 16, 32, 1])]; + tensor var_7277_end_mask_0 = const()[name = tensor("op_7277_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_7277_squeeze_mask_0 = const()[name = tensor("op_7277_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_7277_cast_fp16 = slice_by_index(begin = var_7277_begin_0, end = var_7277_end_0, end_mask = var_7277_end_mask_0, squeeze_mask = var_7277_squeeze_mask_0, x = q_complex_35_cast_fp16)[name = tensor("op_7277_cast_fp16")]; + tensor var_7285_begin_0 = const()[name = tensor("op_7285_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_7285_end_0 = const()[name = tensor("op_7285_end_0"), val = tensor([1, 256, 16, 32, 2])]; + tensor var_7285_end_mask_0 = const()[name = tensor("op_7285_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_7285_squeeze_mask_0 = const()[name = tensor("op_7285_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_7285_cast_fp16 = slice_by_index(begin = var_7285_begin_0, end = var_7285_end_0, end_mask = var_7285_end_mask_0, squeeze_mask = var_7285_squeeze_mask_0, x = q_complex_35_cast_fp16)[name = tensor("op_7285_cast_fp16")]; + tensor var_7293_begin_0 = const()[name = tensor("op_7293_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_7293_end_0 = const()[name = tensor("op_7293_end_0"), val = tensor([1, 256, 16, 32, 1])]; + tensor var_7293_end_mask_0 = const()[name = tensor("op_7293_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_7293_squeeze_mask_0 = const()[name = tensor("op_7293_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_7293_cast_fp16 = slice_by_index(begin = var_7293_begin_0, end = var_7293_end_0, end_mask = var_7293_end_mask_0, squeeze_mask = var_7293_squeeze_mask_0, x = k_complex_35_cast_fp16)[name = tensor("op_7293_cast_fp16")]; + tensor var_7301_begin_0 = const()[name = tensor("op_7301_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_7301_end_0 = const()[name = tensor("op_7301_end_0"), val = tensor([1, 256, 16, 32, 2])]; + tensor var_7301_end_mask_0 = const()[name = tensor("op_7301_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_7301_squeeze_mask_0 = const()[name = tensor("op_7301_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_7301_cast_fp16 = slice_by_index(begin = var_7301_begin_0, end = var_7301_end_0, end_mask = var_7301_end_mask_0, squeeze_mask = var_7301_squeeze_mask_0, x = k_complex_35_cast_fp16)[name = tensor("op_7301_cast_fp16")]; + tensor freqs_35_to_fp16 = const()[name = tensor("freqs_35_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(435332992)))]; + tensor var_7307_cast_fp16 = mul(x = freqs_35_to_fp16, y = ts_107_cast_fp16)[name = tensor("op_7307_cast_fp16")]; + tensor rotr_35_cast_fp16 = cos(x = var_7307_cast_fp16)[name = tensor("rotr_35_cast_fp16")]; + tensor roti_35_cast_fp16 = sin(x = var_7307_cast_fp16)[name = tensor("roti_35_cast_fp16")]; + tensor var_7311_cast_fp16 = mul(x = var_7277_cast_fp16, y = rotr_35_cast_fp16)[name = tensor("op_7311_cast_fp16")]; + tensor var_7312_cast_fp16 = mul(x = var_7285_cast_fp16, y = roti_35_cast_fp16)[name = tensor("op_7312_cast_fp16")]; + tensor qor_69_cast_fp16 = sub(x = var_7311_cast_fp16, y = var_7312_cast_fp16)[name = tensor("qor_69_cast_fp16")]; + tensor var_7315_cast_fp16 = mul(x = var_7277_cast_fp16, y = roti_35_cast_fp16)[name = tensor("op_7315_cast_fp16")]; + tensor var_7316_cast_fp16 = mul(x = var_7285_cast_fp16, y = rotr_35_cast_fp16)[name = tensor("op_7316_cast_fp16")]; + tensor qoi_69_cast_fp16 = add(x = var_7315_cast_fp16, y = var_7316_cast_fp16)[name = tensor("qoi_69_cast_fp16")]; + tensor var_7319_cast_fp16 = mul(x = var_7293_cast_fp16, y = rotr_35_cast_fp16)[name = tensor("op_7319_cast_fp16")]; + tensor var_7320_cast_fp16 = mul(x = var_7301_cast_fp16, y = roti_35_cast_fp16)[name = tensor("op_7320_cast_fp16")]; + tensor kor_69_cast_fp16 = sub(x = var_7319_cast_fp16, y = var_7320_cast_fp16)[name = tensor("kor_69_cast_fp16")]; + tensor var_7323_cast_fp16 = mul(x = var_7293_cast_fp16, y = roti_35_cast_fp16)[name = tensor("op_7323_cast_fp16")]; + tensor var_7324_cast_fp16 = mul(x = var_7301_cast_fp16, y = rotr_35_cast_fp16)[name = tensor("op_7324_cast_fp16")]; + tensor koi_69_cast_fp16 = add(x = var_7323_cast_fp16, y = var_7324_cast_fp16)[name = tensor("koi_69_cast_fp16")]; + tensor qo_35_axis_0 = const()[name = tensor("qo_35_axis_0"), val = tensor(-1)]; + tensor qo_35_cast_fp16 = stack(axis = qo_35_axis_0, values = (qor_69_cast_fp16, qoi_69_cast_fp16))[name = tensor("qo_35_cast_fp16")]; + tensor ko_35_axis_0 = const()[name = tensor("ko_35_axis_0"), val = tensor(-1)]; + tensor ko_35_cast_fp16 = stack(axis = ko_35_axis_0, values = (kor_69_cast_fp16, koi_69_cast_fp16))[name = tensor("ko_35_cast_fp16")]; + tensor var_7357 = const()[name = tensor("op_7357"), val = tensor([1, 256, 16, 64])]; + tensor q_105_cast_fp16 = reshape(shape = var_7357, x = qo_35_cast_fp16)[name = tensor("q_105_cast_fp16")]; + tensor var_7363 = const()[name = tensor("op_7363"), val = tensor([1, 256, 16, 64])]; + tensor k_71_cast_fp16 = reshape(shape = var_7363, x = ko_35_cast_fp16)[name = tensor("k_71_cast_fp16")]; + tensor write_indices_float_103_cast_fp16 = add(x = var_7252_cast_fp16, y = var_393_promoted_to_fp16)[name = tensor("write_indices_float_103_cast_fp16")]; + tensor _inversed_7384_y_0_to_fp16 = const()[name = tensor("_inversed_7384_y_0_to_fp16"), val = tensor(0x1p-9)]; + tensor _inversed_7384_cast_fp16 = mul(x = write_indices_float_103_cast_fp16, y = _inversed_7384_y_0_to_fp16)[name = tensor("_inversed_7384_cast_fp16")]; + tensor var_7385_cast_fp16 = floor(x = _inversed_7384_cast_fp16)[name = tensor("op_7385_cast_fp16")]; + tensor var_7386_to_fp16 = const()[name = tensor("op_7386_to_fp16"), val = tensor(0x1p+9)]; + tensor var_7387_cast_fp16 = mul(x = var_7385_cast_fp16, y = var_7386_to_fp16)[name = tensor("op_7387_cast_fp16")]; + tensor write_indices_float_105_cast_fp16 = sub(x = write_indices_float_103_cast_fp16, y = var_7387_cast_fp16)[name = tensor("write_indices_float_105_cast_fp16")]; + tensor var_7405_value_0_to_fp16 = const()[name = tensor("op_7405_value_0_to_fp16"), val = tensor(0x1.ffp+8)]; + tensor var_7405_cast_fp16 = fill_like(ref_tensor = write_indices_float_105_cast_fp16, value = var_7405_value_0_to_fp16)[name = tensor("op_7405_cast_fp16")]; + tensor write_indices_float_107_cast_fp16 = select(a = write_indices_float_105_cast_fp16, b = var_7405_cast_fp16, cond = token_valid_1_cast_fp16)[name = tensor("write_indices_float_107_cast_fp16")]; + tensor var_7411_dtype_0 = const()[name = tensor("op_7411_dtype_0"), val = tensor("int32")]; + tensor var_7416 = const()[name = tensor("op_7416"), val = tensor([1, 256, 1, 1])]; + tensor write_indices_float_107_cast_fp16_to_int32 = cast(dtype = var_7411_dtype_0, x = write_indices_float_107_cast_fp16)[name = tensor("cast_575")]; + tensor var_7417 = reshape(shape = var_7416, x = write_indices_float_107_cast_fp16_to_int32)[name = tensor("op_7417")]; + tensor write_indices_35_reps_0 = const()[name = tensor("write_indices_35_reps_0"), val = tensor([1, 1, 16, 64])]; + tensor write_indices_35 = tile(reps = write_indices_35_reps_0, x = var_7417)[name = tensor("write_indices_35")]; + tensor var_7427_begin_0 = const()[name = tensor("op_7427_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_7427_end_0 = const()[name = tensor("op_7427_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor var_7427_end_mask_0 = const()[name = tensor("op_7427_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_7427_squeeze_mask_0 = const()[name = tensor("op_7427_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor cache17_to_fp16_dtype_0 = const()[name = tensor("cache17_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor cache17_to_fp16 = cast(dtype = cache17_to_fp16_dtype_0, x = cache17)[name = tensor("cast_574")]; + tensor var_7427_cast_fp16 = slice_by_index(begin = var_7427_begin_0, end = var_7427_end_0, end_mask = var_7427_end_mask_0, squeeze_mask = var_7427_squeeze_mask_0, x = cache17_to_fp16)[name = tensor("op_7427_cast_fp16")]; + tensor var_7429_axis_0 = const()[name = tensor("op_7429_axis_0"), val = tensor(1)]; + tensor var_7429_mode_0 = const()[name = tensor("op_7429_mode_0"), val = tensor("update")]; + tensor var_7429_validate_indices_0 = const()[name = tensor("op_7429_validate_indices_0"), val = tensor(false)]; + tensor var_7429_cast_fp16 = scatter_along_axis(axis = var_7429_axis_0, data = var_7427_cast_fp16, indices = write_indices_35, mode = var_7429_mode_0, updates = k_71_cast_fp16, validate_indices = var_7429_validate_indices_0)[name = tensor("op_7429_cast_fp16")]; + tensor concat_138 = const()[name = tensor("concat_138"), val = tensor([0, 0, 0, 0, 0])]; + tensor concat_139 = const()[name = tensor("concat_139"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_35_internal_tensor_assign_1_stride_0 = const()[name = tensor("new_cache_35_internal_tensor_assign_1_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_35_internal_tensor_assign_1_begin_mask_0 = const()[name = tensor("new_cache_35_internal_tensor_assign_1_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_35_internal_tensor_assign_1_end_mask_0 = const()[name = tensor("new_cache_35_internal_tensor_assign_1_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_35_internal_tensor_assign_1_squeeze_mask_0 = const()[name = tensor("new_cache_35_internal_tensor_assign_1_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_80 = const()[name = tensor("shape_80"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_34 = const()[name = tensor("reduce_prod_34"), val = tensor(1048576)]; + tensor range_1d_34_start_0 = const()[name = tensor("range_1d_34_start_0"), val = tensor(0)]; + tensor range_1d_34_step_0 = const()[name = tensor("range_1d_34_step_0"), val = tensor(1)]; + tensor range_1d_34 = range_1d(end = reduce_prod_34, start = range_1d_34_start_0, step = range_1d_34_step_0)[name = tensor("range_1d_34")]; + tensor reshape_170 = reshape(shape = shape_80, x = range_1d_34)[name = tensor("reshape_170")]; + tensor slice_by_index_34 = slice_by_index(begin = concat_138, begin_mask = new_cache_35_internal_tensor_assign_1_begin_mask_0, end = concat_139, end_mask = new_cache_35_internal_tensor_assign_1_end_mask_0, squeeze_mask = new_cache_35_internal_tensor_assign_1_squeeze_mask_0, stride = new_cache_35_internal_tensor_assign_1_stride_0, x = reshape_170)[name = tensor("slice_by_index_34")]; + tensor reshape_171_shape_0 = const()[name = tensor("reshape_171_shape_0"), val = tensor([-1])]; + tensor reshape_171 = reshape(shape = reshape_171_shape_0, x = slice_by_index_34)[name = tensor("reshape_171")]; + tensor reshape_172_shape_0 = const()[name = tensor("reshape_172_shape_0"), val = tensor([-1])]; + tensor reshape_172_cast_fp16 = reshape(shape = reshape_172_shape_0, x = var_7429_cast_fp16)[name = tensor("reshape_172_cast_fp16")]; + tensor reshape_173_shape_0 = const()[name = tensor("reshape_173_shape_0"), val = tensor([-1])]; + tensor reshape_173_cast_fp16 = reshape(shape = reshape_173_shape_0, x = cache17_to_fp16)[name = tensor("reshape_173_cast_fp16")]; + tensor scatter_34_mode_0 = const()[name = tensor("scatter_34_mode_0"), val = tensor("update")]; + tensor scatter_34_axis_0 = const()[name = tensor("scatter_34_axis_0"), val = tensor(0)]; + tensor scatter_34_validate_indices_0 = const()[name = tensor("scatter_34_validate_indices_0"), val = tensor(false)]; + tensor scatter_34_cast_fp16 = scatter(axis = scatter_34_axis_0, data = reshape_173_cast_fp16, indices = reshape_171, mode = scatter_34_mode_0, updates = reshape_172_cast_fp16, validate_indices = scatter_34_validate_indices_0)[name = tensor("scatter_34_cast_fp16")]; + tensor reshape_174_cast_fp16 = reshape(shape = shape_80, x = scatter_34_cast_fp16)[name = tensor("reshape_174_cast_fp16")]; + tensor var_7437_begin_0 = const()[name = tensor("op_7437_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor var_7437_end_0 = const()[name = tensor("op_7437_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor var_7437_end_mask_0 = const()[name = tensor("op_7437_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_7437_squeeze_mask_0 = const()[name = tensor("op_7437_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor var_7437_cast_fp16 = slice_by_index(begin = var_7437_begin_0, end = var_7437_end_0, end_mask = var_7437_end_mask_0, squeeze_mask = var_7437_squeeze_mask_0, x = reshape_174_cast_fp16)[name = tensor("op_7437_cast_fp16")]; + tensor var_7439_axis_0 = const()[name = tensor("op_7439_axis_0"), val = tensor(1)]; + tensor var_7439_mode_0 = const()[name = tensor("op_7439_mode_0"), val = tensor("update")]; + tensor var_7439_validate_indices_0 = const()[name = tensor("op_7439_validate_indices_0"), val = tensor(false)]; + tensor var_7439_cast_fp16 = scatter_along_axis(axis = var_7439_axis_0, data = var_7437_cast_fp16, indices = write_indices_35, mode = var_7439_mode_0, updates = v_35_cast_fp16, validate_indices = var_7439_validate_indices_0)[name = tensor("op_7439_cast_fp16")]; + tensor concat_140 = const()[name = tensor("concat_140"), val = tensor([1, 0, 0, 0, 0])]; + tensor concat_141 = const()[name = tensor("concat_141"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_35_internal_tensor_assign_2_stride_0 = const()[name = tensor("new_cache_35_internal_tensor_assign_2_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_35_internal_tensor_assign_2_begin_mask_0 = const()[name = tensor("new_cache_35_internal_tensor_assign_2_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_35_internal_tensor_assign_2_end_mask_0 = const()[name = tensor("new_cache_35_internal_tensor_assign_2_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_35_internal_tensor_assign_2_squeeze_mask_0 = const()[name = tensor("new_cache_35_internal_tensor_assign_2_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_81 = const()[name = tensor("shape_81"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_35 = const()[name = tensor("reduce_prod_35"), val = tensor(1048576)]; + tensor range_1d_35_start_0 = const()[name = tensor("range_1d_35_start_0"), val = tensor(0)]; + tensor range_1d_35_step_0 = const()[name = tensor("range_1d_35_step_0"), val = tensor(1)]; + tensor range_1d_35 = range_1d(end = reduce_prod_35, start = range_1d_35_start_0, step = range_1d_35_step_0)[name = tensor("range_1d_35")]; + tensor reshape_175 = reshape(shape = shape_81, x = range_1d_35)[name = tensor("reshape_175")]; + tensor slice_by_index_35 = slice_by_index(begin = concat_140, begin_mask = new_cache_35_internal_tensor_assign_2_begin_mask_0, end = concat_141, end_mask = new_cache_35_internal_tensor_assign_2_end_mask_0, squeeze_mask = new_cache_35_internal_tensor_assign_2_squeeze_mask_0, stride = new_cache_35_internal_tensor_assign_2_stride_0, x = reshape_175)[name = tensor("slice_by_index_35")]; + tensor reshape_176_shape_0 = const()[name = tensor("reshape_176_shape_0"), val = tensor([-1])]; + tensor reshape_176 = reshape(shape = reshape_176_shape_0, x = slice_by_index_35)[name = tensor("reshape_176")]; + tensor reshape_177_shape_0 = const()[name = tensor("reshape_177_shape_0"), val = tensor([-1])]; + tensor reshape_177_cast_fp16 = reshape(shape = reshape_177_shape_0, x = var_7439_cast_fp16)[name = tensor("reshape_177_cast_fp16")]; + tensor reshape_178_shape_0 = const()[name = tensor("reshape_178_shape_0"), val = tensor([-1])]; + tensor reshape_178_cast_fp16 = reshape(shape = reshape_178_shape_0, x = reshape_174_cast_fp16)[name = tensor("reshape_178_cast_fp16")]; + tensor scatter_35_mode_0 = const()[name = tensor("scatter_35_mode_0"), val = tensor("update")]; + tensor scatter_35_axis_0 = const()[name = tensor("scatter_35_axis_0"), val = tensor(0)]; + tensor scatter_35_validate_indices_0 = const()[name = tensor("scatter_35_validate_indices_0"), val = tensor(false)]; + tensor scatter_35_cast_fp16 = scatter(axis = scatter_35_axis_0, data = reshape_178_cast_fp16, indices = reshape_176, mode = scatter_35_mode_0, updates = reshape_177_cast_fp16, validate_indices = scatter_35_validate_indices_0)[name = tensor("scatter_35_cast_fp16")]; + tensor reshape_179_cast_fp16 = reshape(shape = shape_81, x = scatter_35_cast_fp16)[name = tensor("reshape_179_cast_fp16")]; + tensor reshape_179_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("reshape_179_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor keys_103_begin_0 = const()[name = tensor("keys_103_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor keys_103_end_0 = const()[name = tensor("keys_103_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor keys_103_end_mask_0 = const()[name = tensor("keys_103_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor keys_103_squeeze_mask_0 = const()[name = tensor("keys_103_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor keys_103_cast_fp16 = slice_by_index(begin = keys_103_begin_0, end = keys_103_end_0, end_mask = keys_103_end_mask_0, squeeze_mask = keys_103_squeeze_mask_0, x = reshape_179_cast_fp16)[name = tensor("keys_103_cast_fp16")]; + tensor values_103_begin_0 = const()[name = tensor("values_103_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor values_103_end_0 = const()[name = tensor("values_103_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor values_103_end_mask_0 = const()[name = tensor("values_103_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor values_103_squeeze_mask_0 = const()[name = tensor("values_103_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor values_103_cast_fp16 = slice_by_index(begin = values_103_begin_0, end = values_103_end_0, end_mask = values_103_end_mask_0, squeeze_mask = values_103_squeeze_mask_0, x = reshape_179_cast_fp16)[name = tensor("values_103_cast_fp16")]; + tensor var_7451_cast_fp16 = not_equal(x = keys_103_cast_fp16, y = keys_103_cast_fp16)[name = tensor("op_7451_cast_fp16")]; + tensor keys_105_cast_fp16 = select(a = var_470_to_fp16, b = keys_103_cast_fp16, cond = var_7451_cast_fp16)[name = tensor("keys_105_cast_fp16")]; + tensor var_7459_cast_fp16 = not_equal(x = values_103_cast_fp16, y = values_103_cast_fp16)[name = tensor("op_7459_cast_fp16")]; + tensor values_105_cast_fp16 = select(a = var_470_to_fp16, b = values_103_cast_fp16, cond = var_7459_cast_fp16)[name = tensor("values_105_cast_fp16")]; + tensor var_7483 = const()[name = tensor("op_7483"), val = tensor([0, 2, 1, 3])]; + tensor var_7499 = const()[name = tensor("op_7499"), val = tensor([1, 1, 1])]; + tensor var_7500_cast_fp16 = reshape(shape = var_7499, x = position17_to_fp16)[name = tensor("op_7500_cast_fp16")]; + tensor q_positions_35_cast_fp16 = add(x = var_7500_cast_fp16, y = q_offsets_1_promoted_to_fp16)[name = tensor("q_positions_35_cast_fp16")]; + tensor valid_end_35_cast_fp16 = add(x = var_7500_cast_fp16, y = var_536_cast_fp16)[name = tensor("valid_end_35_cast_fp16")]; + tensor valid_mask_35_cast_fp16 = less(x = k_positions_1_promoted_to_fp16, y = valid_end_35_cast_fp16)[name = tensor("valid_mask_35_cast_fp16")]; + tensor causal_mask_35_cast_fp16 = less_equal(x = k_positions_1_promoted_to_fp16, y = q_positions_35_cast_fp16)[name = tensor("causal_mask_35_cast_fp16")]; + tensor var_7528 = logical_and(x = valid_mask_35_cast_fp16, y = causal_mask_35_cast_fp16)[name = tensor("op_7528")]; + tensor attn_mask_35_axes_0 = const()[name = tensor("attn_mask_35_axes_0"), val = tensor([1])]; + tensor attn_mask_35 = expand_dims(axes = attn_mask_35_axes_0, x = var_7528)[name = tensor("attn_mask_35")]; + tensor var_7542_transpose_x_0 = const()[name = tensor("op_7542_transpose_x_0"), val = tensor(false)]; + tensor var_7542_transpose_y_0 = const()[name = tensor("op_7542_transpose_y_0"), val = tensor(false)]; + tensor transpose_103_perm_0 = const()[name = tensor("transpose_103_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_104_perm_0 = const()[name = tensor("transpose_104_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_104 = transpose(perm = transpose_104_perm_0, x = keys_105_cast_fp16)[name = tensor("transpose_136")]; + tensor transpose_103 = transpose(perm = transpose_103_perm_0, x = q_105_cast_fp16)[name = tensor("transpose_137")]; + tensor var_7542_cast_fp16 = matmul(transpose_x = var_7542_transpose_x_0, transpose_y = var_7542_transpose_y_0, x = transpose_103, y = transpose_104)[name = tensor("op_7542_cast_fp16")]; + tensor var_7536_to_fp16 = const()[name = tensor("op_7536_to_fp16"), val = tensor([0x1p-3])]; + tensor attn_weights_103_cast_fp16 = mul(x = var_7542_cast_fp16, y = var_7536_to_fp16)[name = tensor("attn_weights_103_cast_fp16")]; + tensor var_7544 = logical_not(x = attn_mask_35)[name = tensor("op_7544")]; + tensor var_7545_to_fp16 = const()[name = tensor("op_7545_to_fp16"), val = tensor(-inf)]; + tensor attn_weights_105_cast_fp16 = select(a = var_7545_to_fp16, b = attn_weights_103_cast_fp16, cond = var_7544)[name = tensor("attn_weights_105_cast_fp16")]; + tensor var_7547 = const()[name = tensor("op_7547"), val = tensor(-1)]; + tensor attn_weights_107_cast_fp16 = softmax(axis = var_7547, x = attn_weights_105_cast_fp16)[name = tensor("attn_weights_107_cast_fp16")]; + tensor attn_output_35_transpose_x_0 = const()[name = tensor("attn_output_35_transpose_x_0"), val = tensor(false)]; + tensor attn_output_35_transpose_y_0 = const()[name = tensor("attn_output_35_transpose_y_0"), val = tensor(false)]; + tensor values_107_cast_fp16 = transpose(perm = var_7483, x = values_105_cast_fp16)[name = tensor("transpose_138")]; + tensor attn_output_35_cast_fp16 = matmul(transpose_x = attn_output_35_transpose_x_0, transpose_y = attn_output_35_transpose_y_0, x = attn_weights_107_cast_fp16, y = values_107_cast_fp16)[name = tensor("attn_output_35_cast_fp16")]; + tensor var_7555 = const()[name = tensor("op_7555"), val = tensor([0, 2, 1, 3])]; + tensor var_7560 = const()[name = tensor("op_7560"), val = tensor([1, 256, 1024])]; + tensor var_7556_cast_fp16 = transpose(perm = var_7555, x = attn_output_35_cast_fp16)[name = tensor("transpose_135")]; + tensor input_209_cast_fp16 = reshape(shape = var_7560, x = var_7556_cast_fp16)[name = tensor("input_209_cast_fp16")]; + tensor attn17_out_proj_weight_to_fp16 = const()[name = tensor("attn17_out_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(435333120)))]; + tensor linear_69_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = attn17_out_proj_weight_to_fp16, x = input_209_cast_fp16)[name = tensor("linear_69_cast_fp16")]; + tensor var_7569_cast_fp16 = add(x = position17_to_fp16, y = valid_len_to_fp16)[name = tensor("op_7569_cast_fp16")]; + tensor var_7569_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_7569_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor input_211_cast_fp16 = add(x = input_205_cast_fp16, y = linear_69_cast_fp16)[name = tensor("input_211_cast_fp16")]; + tensor input_213_axes_0 = const()[name = tensor("input_213_axes_0"), val = tensor([-1])]; + tensor norm17_2_weight_to_fp16 = const()[name = tensor("norm17_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(437430336)))]; + tensor norm17_2_bias_to_fp16 = const()[name = tensor("norm17_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(437432448)))]; + tensor var_7573_to_fp16 = const()[name = tensor("op_7573_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_213_cast_fp16 = layer_norm(axes = input_213_axes_0, beta = norm17_2_bias_to_fp16, epsilon = var_7573_to_fp16, gamma = norm17_2_weight_to_fp16, x = input_211_cast_fp16)[name = tensor("input_213_cast_fp16")]; + tensor linear17_1_weight_to_fp16 = const()[name = tensor("linear17_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(437434560)))]; + tensor linear_70_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = linear17_1_weight_to_fp16, x = input_213_cast_fp16)[name = tensor("linear_70_cast_fp16")]; + tensor input_215_mode_0 = const()[name = tensor("input_215_mode_0"), val = tensor("EXACT")]; + tensor input_215_cast_fp16 = gelu(mode = input_215_mode_0, x = linear_70_cast_fp16)[name = tensor("input_215_cast_fp16")]; + tensor linear17_2_weight_to_fp16 = const()[name = tensor("linear17_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(445823232)))]; + tensor linear_71_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = linear17_2_weight_to_fp16, x = input_215_cast_fp16)[name = tensor("linear_71_cast_fp16")]; + tensor input_217_cast_fp16 = add(x = input_211_cast_fp16, y = linear_71_cast_fp16)[name = tensor("input_217_cast_fp16")]; + tensor input_219_axes_0 = const()[name = tensor("input_219_axes_0"), val = tensor([-1])]; + tensor norm18_1_weight_to_fp16 = const()[name = tensor("norm18_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(454211904)))]; + tensor norm18_1_bias_to_fp16 = const()[name = tensor("norm18_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(454214016)))]; + tensor var_7590_to_fp16 = const()[name = tensor("op_7590_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_219_cast_fp16 = layer_norm(axes = input_219_axes_0, beta = norm18_1_bias_to_fp16, epsilon = var_7590_to_fp16, gamma = norm18_1_weight_to_fp16, x = input_217_cast_fp16)[name = tensor("input_219_cast_fp16")]; + tensor attn18_in_proj_weight_to_fp16 = const()[name = tensor("attn18_in_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(454216128)))]; + tensor linear_72_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = attn18_in_proj_weight_to_fp16, x = input_219_cast_fp16)[name = tensor("linear_72_cast_fp16")]; + tensor var_7604 = const()[name = tensor("op_7604"), val = tensor([1, 256, 3, 16, 64])]; + tensor qkv_37_cast_fp16 = reshape(shape = var_7604, x = linear_72_cast_fp16)[name = tensor("qkv_37_cast_fp16")]; + tensor q_109_begin_0 = const()[name = tensor("q_109_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor q_109_end_0 = const()[name = tensor("q_109_end_0"), val = tensor([1, 256, 1, 16, 64])]; + tensor q_109_end_mask_0 = const()[name = tensor("q_109_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor q_109_squeeze_mask_0 = const()[name = tensor("q_109_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor q_109_cast_fp16 = slice_by_index(begin = q_109_begin_0, end = q_109_end_0, end_mask = q_109_end_mask_0, squeeze_mask = q_109_squeeze_mask_0, x = qkv_37_cast_fp16)[name = tensor("q_109_cast_fp16")]; + tensor k_73_begin_0 = const()[name = tensor("k_73_begin_0"), val = tensor([0, 0, 1, 0, 0])]; + tensor k_73_end_0 = const()[name = tensor("k_73_end_0"), val = tensor([1, 256, 2, 16, 64])]; + tensor k_73_end_mask_0 = const()[name = tensor("k_73_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor k_73_squeeze_mask_0 = const()[name = tensor("k_73_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor k_73_cast_fp16 = slice_by_index(begin = k_73_begin_0, end = k_73_end_0, end_mask = k_73_end_mask_0, squeeze_mask = k_73_squeeze_mask_0, x = qkv_37_cast_fp16)[name = tensor("k_73_cast_fp16")]; + tensor v_37_begin_0 = const()[name = tensor("v_37_begin_0"), val = tensor([0, 0, 2, 0, 0])]; + tensor v_37_end_0 = const()[name = tensor("v_37_end_0"), val = tensor([1, 256, 3, 16, 64])]; + tensor v_37_end_mask_0 = const()[name = tensor("v_37_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor v_37_squeeze_mask_0 = const()[name = tensor("v_37_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor v_37_cast_fp16 = slice_by_index(begin = v_37_begin_0, end = v_37_end_0, end_mask = v_37_end_mask_0, squeeze_mask = v_37_squeeze_mask_0, x = qkv_37_cast_fp16)[name = tensor("v_37_cast_fp16")]; + tensor var_7662 = const()[name = tensor("op_7662"), val = tensor([1, 1])]; + tensor position18_to_fp16_dtype_0 = const()[name = tensor("position18_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor position18_to_fp16 = cast(dtype = position18_to_fp16_dtype_0, x = position18)[name = tensor("cast_571")]; + tensor var_7663_cast_fp16 = reshape(shape = var_7662, x = position18_to_fp16)[name = tensor("op_7663_cast_fp16")]; + tensor ts_111_cast_fp16 = add(x = ts_1_promoted_to_fp16, y = var_7663_cast_fp16)[name = tensor("ts_111_cast_fp16")]; + tensor var_7670 = const()[name = tensor("op_7670"), val = tensor([1, 256, 1, 1])]; + tensor ts_113_cast_fp16 = reshape(shape = var_7670, x = ts_111_cast_fp16)[name = tensor("ts_113_cast_fp16")]; + tensor var_7677 = const()[name = tensor("op_7677"), val = tensor([1, 256, 16, 32, 2])]; + tensor q_complex_37_cast_fp16 = reshape(shape = var_7677, x = q_109_cast_fp16)[name = tensor("q_complex_37_cast_fp16")]; + tensor var_7684 = const()[name = tensor("op_7684"), val = tensor([1, 256, 16, 32, 2])]; + tensor k_complex_37_cast_fp16 = reshape(shape = var_7684, x = k_73_cast_fp16)[name = tensor("k_complex_37_cast_fp16")]; + tensor var_7688_begin_0 = const()[name = tensor("op_7688_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_7688_end_0 = const()[name = tensor("op_7688_end_0"), val = tensor([1, 256, 16, 32, 1])]; + tensor var_7688_end_mask_0 = const()[name = tensor("op_7688_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_7688_squeeze_mask_0 = const()[name = tensor("op_7688_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_7688_cast_fp16 = slice_by_index(begin = var_7688_begin_0, end = var_7688_end_0, end_mask = var_7688_end_mask_0, squeeze_mask = var_7688_squeeze_mask_0, x = q_complex_37_cast_fp16)[name = tensor("op_7688_cast_fp16")]; + tensor var_7696_begin_0 = const()[name = tensor("op_7696_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_7696_end_0 = const()[name = tensor("op_7696_end_0"), val = tensor([1, 256, 16, 32, 2])]; + tensor var_7696_end_mask_0 = const()[name = tensor("op_7696_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_7696_squeeze_mask_0 = const()[name = tensor("op_7696_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_7696_cast_fp16 = slice_by_index(begin = var_7696_begin_0, end = var_7696_end_0, end_mask = var_7696_end_mask_0, squeeze_mask = var_7696_squeeze_mask_0, x = q_complex_37_cast_fp16)[name = tensor("op_7696_cast_fp16")]; + tensor var_7704_begin_0 = const()[name = tensor("op_7704_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_7704_end_0 = const()[name = tensor("op_7704_end_0"), val = tensor([1, 256, 16, 32, 1])]; + tensor var_7704_end_mask_0 = const()[name = tensor("op_7704_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_7704_squeeze_mask_0 = const()[name = tensor("op_7704_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_7704_cast_fp16 = slice_by_index(begin = var_7704_begin_0, end = var_7704_end_0, end_mask = var_7704_end_mask_0, squeeze_mask = var_7704_squeeze_mask_0, x = k_complex_37_cast_fp16)[name = tensor("op_7704_cast_fp16")]; + tensor var_7712_begin_0 = const()[name = tensor("op_7712_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_7712_end_0 = const()[name = tensor("op_7712_end_0"), val = tensor([1, 256, 16, 32, 2])]; + tensor var_7712_end_mask_0 = const()[name = tensor("op_7712_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_7712_squeeze_mask_0 = const()[name = tensor("op_7712_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_7712_cast_fp16 = slice_by_index(begin = var_7712_begin_0, end = var_7712_end_0, end_mask = var_7712_end_mask_0, squeeze_mask = var_7712_squeeze_mask_0, x = k_complex_37_cast_fp16)[name = tensor("op_7712_cast_fp16")]; + tensor freqs_37_to_fp16 = const()[name = tensor("freqs_37_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(460507648)))]; + tensor var_7718_cast_fp16 = mul(x = freqs_37_to_fp16, y = ts_113_cast_fp16)[name = tensor("op_7718_cast_fp16")]; + tensor rotr_37_cast_fp16 = cos(x = var_7718_cast_fp16)[name = tensor("rotr_37_cast_fp16")]; + tensor roti_37_cast_fp16 = sin(x = var_7718_cast_fp16)[name = tensor("roti_37_cast_fp16")]; + tensor var_7722_cast_fp16 = mul(x = var_7688_cast_fp16, y = rotr_37_cast_fp16)[name = tensor("op_7722_cast_fp16")]; + tensor var_7723_cast_fp16 = mul(x = var_7696_cast_fp16, y = roti_37_cast_fp16)[name = tensor("op_7723_cast_fp16")]; + tensor qor_73_cast_fp16 = sub(x = var_7722_cast_fp16, y = var_7723_cast_fp16)[name = tensor("qor_73_cast_fp16")]; + tensor var_7726_cast_fp16 = mul(x = var_7688_cast_fp16, y = roti_37_cast_fp16)[name = tensor("op_7726_cast_fp16")]; + tensor var_7727_cast_fp16 = mul(x = var_7696_cast_fp16, y = rotr_37_cast_fp16)[name = tensor("op_7727_cast_fp16")]; + tensor qoi_73_cast_fp16 = add(x = var_7726_cast_fp16, y = var_7727_cast_fp16)[name = tensor("qoi_73_cast_fp16")]; + tensor var_7730_cast_fp16 = mul(x = var_7704_cast_fp16, y = rotr_37_cast_fp16)[name = tensor("op_7730_cast_fp16")]; + tensor var_7731_cast_fp16 = mul(x = var_7712_cast_fp16, y = roti_37_cast_fp16)[name = tensor("op_7731_cast_fp16")]; + tensor kor_73_cast_fp16 = sub(x = var_7730_cast_fp16, y = var_7731_cast_fp16)[name = tensor("kor_73_cast_fp16")]; + tensor var_7734_cast_fp16 = mul(x = var_7704_cast_fp16, y = roti_37_cast_fp16)[name = tensor("op_7734_cast_fp16")]; + tensor var_7735_cast_fp16 = mul(x = var_7712_cast_fp16, y = rotr_37_cast_fp16)[name = tensor("op_7735_cast_fp16")]; + tensor koi_73_cast_fp16 = add(x = var_7734_cast_fp16, y = var_7735_cast_fp16)[name = tensor("koi_73_cast_fp16")]; + tensor qo_37_axis_0 = const()[name = tensor("qo_37_axis_0"), val = tensor(-1)]; + tensor qo_37_cast_fp16 = stack(axis = qo_37_axis_0, values = (qor_73_cast_fp16, qoi_73_cast_fp16))[name = tensor("qo_37_cast_fp16")]; + tensor ko_37_axis_0 = const()[name = tensor("ko_37_axis_0"), val = tensor(-1)]; + tensor ko_37_cast_fp16 = stack(axis = ko_37_axis_0, values = (kor_73_cast_fp16, koi_73_cast_fp16))[name = tensor("ko_37_cast_fp16")]; + tensor var_7768 = const()[name = tensor("op_7768"), val = tensor([1, 256, 16, 64])]; + tensor q_111_cast_fp16 = reshape(shape = var_7768, x = qo_37_cast_fp16)[name = tensor("q_111_cast_fp16")]; + tensor var_7774 = const()[name = tensor("op_7774"), val = tensor([1, 256, 16, 64])]; + tensor k_75_cast_fp16 = reshape(shape = var_7774, x = ko_37_cast_fp16)[name = tensor("k_75_cast_fp16")]; + tensor write_indices_float_109_cast_fp16 = add(x = var_7663_cast_fp16, y = var_393_promoted_to_fp16)[name = tensor("write_indices_float_109_cast_fp16")]; + tensor _inversed_7795_y_0_to_fp16 = const()[name = tensor("_inversed_7795_y_0_to_fp16"), val = tensor(0x1p-9)]; + tensor _inversed_7795_cast_fp16 = mul(x = write_indices_float_109_cast_fp16, y = _inversed_7795_y_0_to_fp16)[name = tensor("_inversed_7795_cast_fp16")]; + tensor var_7796_cast_fp16 = floor(x = _inversed_7795_cast_fp16)[name = tensor("op_7796_cast_fp16")]; + tensor var_7797_to_fp16 = const()[name = tensor("op_7797_to_fp16"), val = tensor(0x1p+9)]; + tensor var_7798_cast_fp16 = mul(x = var_7796_cast_fp16, y = var_7797_to_fp16)[name = tensor("op_7798_cast_fp16")]; + tensor write_indices_float_111_cast_fp16 = sub(x = write_indices_float_109_cast_fp16, y = var_7798_cast_fp16)[name = tensor("write_indices_float_111_cast_fp16")]; + tensor var_7816_value_0_to_fp16 = const()[name = tensor("op_7816_value_0_to_fp16"), val = tensor(0x1.ffp+8)]; + tensor var_7816_cast_fp16 = fill_like(ref_tensor = write_indices_float_111_cast_fp16, value = var_7816_value_0_to_fp16)[name = tensor("op_7816_cast_fp16")]; + tensor write_indices_float_113_cast_fp16 = select(a = write_indices_float_111_cast_fp16, b = var_7816_cast_fp16, cond = token_valid_1_cast_fp16)[name = tensor("write_indices_float_113_cast_fp16")]; + tensor var_7822_dtype_0 = const()[name = tensor("op_7822_dtype_0"), val = tensor("int32")]; + tensor var_7827 = const()[name = tensor("op_7827"), val = tensor([1, 256, 1, 1])]; + tensor write_indices_float_113_cast_fp16_to_int32 = cast(dtype = var_7822_dtype_0, x = write_indices_float_113_cast_fp16)[name = tensor("cast_570")]; + tensor var_7828 = reshape(shape = var_7827, x = write_indices_float_113_cast_fp16_to_int32)[name = tensor("op_7828")]; + tensor write_indices_37_reps_0 = const()[name = tensor("write_indices_37_reps_0"), val = tensor([1, 1, 16, 64])]; + tensor write_indices_37 = tile(reps = write_indices_37_reps_0, x = var_7828)[name = tensor("write_indices_37")]; + tensor var_7838_begin_0 = const()[name = tensor("op_7838_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_7838_end_0 = const()[name = tensor("op_7838_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor var_7838_end_mask_0 = const()[name = tensor("op_7838_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_7838_squeeze_mask_0 = const()[name = tensor("op_7838_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor cache18_to_fp16_dtype_0 = const()[name = tensor("cache18_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor cache18_to_fp16 = cast(dtype = cache18_to_fp16_dtype_0, x = cache18)[name = tensor("cast_569")]; + tensor var_7838_cast_fp16 = slice_by_index(begin = var_7838_begin_0, end = var_7838_end_0, end_mask = var_7838_end_mask_0, squeeze_mask = var_7838_squeeze_mask_0, x = cache18_to_fp16)[name = tensor("op_7838_cast_fp16")]; + tensor var_7840_axis_0 = const()[name = tensor("op_7840_axis_0"), val = tensor(1)]; + tensor var_7840_mode_0 = const()[name = tensor("op_7840_mode_0"), val = tensor("update")]; + tensor var_7840_validate_indices_0 = const()[name = tensor("op_7840_validate_indices_0"), val = tensor(false)]; + tensor var_7840_cast_fp16 = scatter_along_axis(axis = var_7840_axis_0, data = var_7838_cast_fp16, indices = write_indices_37, mode = var_7840_mode_0, updates = k_75_cast_fp16, validate_indices = var_7840_validate_indices_0)[name = tensor("op_7840_cast_fp16")]; + tensor concat_146 = const()[name = tensor("concat_146"), val = tensor([0, 0, 0, 0, 0])]; + tensor concat_147 = const()[name = tensor("concat_147"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_37_internal_tensor_assign_1_stride_0 = const()[name = tensor("new_cache_37_internal_tensor_assign_1_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_37_internal_tensor_assign_1_begin_mask_0 = const()[name = tensor("new_cache_37_internal_tensor_assign_1_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_37_internal_tensor_assign_1_end_mask_0 = const()[name = tensor("new_cache_37_internal_tensor_assign_1_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_37_internal_tensor_assign_1_squeeze_mask_0 = const()[name = tensor("new_cache_37_internal_tensor_assign_1_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_82 = const()[name = tensor("shape_82"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_36 = const()[name = tensor("reduce_prod_36"), val = tensor(1048576)]; + tensor range_1d_36_start_0 = const()[name = tensor("range_1d_36_start_0"), val = tensor(0)]; + tensor range_1d_36_step_0 = const()[name = tensor("range_1d_36_step_0"), val = tensor(1)]; + tensor range_1d_36 = range_1d(end = reduce_prod_36, start = range_1d_36_start_0, step = range_1d_36_step_0)[name = tensor("range_1d_36")]; + tensor reshape_180 = reshape(shape = shape_82, x = range_1d_36)[name = tensor("reshape_180")]; + tensor slice_by_index_36 = slice_by_index(begin = concat_146, begin_mask = new_cache_37_internal_tensor_assign_1_begin_mask_0, end = concat_147, end_mask = new_cache_37_internal_tensor_assign_1_end_mask_0, squeeze_mask = new_cache_37_internal_tensor_assign_1_squeeze_mask_0, stride = new_cache_37_internal_tensor_assign_1_stride_0, x = reshape_180)[name = tensor("slice_by_index_36")]; + tensor reshape_181_shape_0 = const()[name = tensor("reshape_181_shape_0"), val = tensor([-1])]; + tensor reshape_181 = reshape(shape = reshape_181_shape_0, x = slice_by_index_36)[name = tensor("reshape_181")]; + tensor reshape_182_shape_0 = const()[name = tensor("reshape_182_shape_0"), val = tensor([-1])]; + tensor reshape_182_cast_fp16 = reshape(shape = reshape_182_shape_0, x = var_7840_cast_fp16)[name = tensor("reshape_182_cast_fp16")]; + tensor reshape_183_shape_0 = const()[name = tensor("reshape_183_shape_0"), val = tensor([-1])]; + tensor reshape_183_cast_fp16 = reshape(shape = reshape_183_shape_0, x = cache18_to_fp16)[name = tensor("reshape_183_cast_fp16")]; + tensor scatter_36_mode_0 = const()[name = tensor("scatter_36_mode_0"), val = tensor("update")]; + tensor scatter_36_axis_0 = const()[name = tensor("scatter_36_axis_0"), val = tensor(0)]; + tensor scatter_36_validate_indices_0 = const()[name = tensor("scatter_36_validate_indices_0"), val = tensor(false)]; + tensor scatter_36_cast_fp16 = scatter(axis = scatter_36_axis_0, data = reshape_183_cast_fp16, indices = reshape_181, mode = scatter_36_mode_0, updates = reshape_182_cast_fp16, validate_indices = scatter_36_validate_indices_0)[name = tensor("scatter_36_cast_fp16")]; + tensor reshape_184_cast_fp16 = reshape(shape = shape_82, x = scatter_36_cast_fp16)[name = tensor("reshape_184_cast_fp16")]; + tensor var_7848_begin_0 = const()[name = tensor("op_7848_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor var_7848_end_0 = const()[name = tensor("op_7848_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor var_7848_end_mask_0 = const()[name = tensor("op_7848_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_7848_squeeze_mask_0 = const()[name = tensor("op_7848_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor var_7848_cast_fp16 = slice_by_index(begin = var_7848_begin_0, end = var_7848_end_0, end_mask = var_7848_end_mask_0, squeeze_mask = var_7848_squeeze_mask_0, x = reshape_184_cast_fp16)[name = tensor("op_7848_cast_fp16")]; + tensor var_7850_axis_0 = const()[name = tensor("op_7850_axis_0"), val = tensor(1)]; + tensor var_7850_mode_0 = const()[name = tensor("op_7850_mode_0"), val = tensor("update")]; + tensor var_7850_validate_indices_0 = const()[name = tensor("op_7850_validate_indices_0"), val = tensor(false)]; + tensor var_7850_cast_fp16 = scatter_along_axis(axis = var_7850_axis_0, data = var_7848_cast_fp16, indices = write_indices_37, mode = var_7850_mode_0, updates = v_37_cast_fp16, validate_indices = var_7850_validate_indices_0)[name = tensor("op_7850_cast_fp16")]; + tensor concat_148 = const()[name = tensor("concat_148"), val = tensor([1, 0, 0, 0, 0])]; + tensor concat_149 = const()[name = tensor("concat_149"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_37_internal_tensor_assign_2_stride_0 = const()[name = tensor("new_cache_37_internal_tensor_assign_2_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_37_internal_tensor_assign_2_begin_mask_0 = const()[name = tensor("new_cache_37_internal_tensor_assign_2_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_37_internal_tensor_assign_2_end_mask_0 = const()[name = tensor("new_cache_37_internal_tensor_assign_2_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_37_internal_tensor_assign_2_squeeze_mask_0 = const()[name = tensor("new_cache_37_internal_tensor_assign_2_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_83 = const()[name = tensor("shape_83"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_37 = const()[name = tensor("reduce_prod_37"), val = tensor(1048576)]; + tensor range_1d_37_start_0 = const()[name = tensor("range_1d_37_start_0"), val = tensor(0)]; + tensor range_1d_37_step_0 = const()[name = tensor("range_1d_37_step_0"), val = tensor(1)]; + tensor range_1d_37 = range_1d(end = reduce_prod_37, start = range_1d_37_start_0, step = range_1d_37_step_0)[name = tensor("range_1d_37")]; + tensor reshape_185 = reshape(shape = shape_83, x = range_1d_37)[name = tensor("reshape_185")]; + tensor slice_by_index_37 = slice_by_index(begin = concat_148, begin_mask = new_cache_37_internal_tensor_assign_2_begin_mask_0, end = concat_149, end_mask = new_cache_37_internal_tensor_assign_2_end_mask_0, squeeze_mask = new_cache_37_internal_tensor_assign_2_squeeze_mask_0, stride = new_cache_37_internal_tensor_assign_2_stride_0, x = reshape_185)[name = tensor("slice_by_index_37")]; + tensor reshape_186_shape_0 = const()[name = tensor("reshape_186_shape_0"), val = tensor([-1])]; + tensor reshape_186 = reshape(shape = reshape_186_shape_0, x = slice_by_index_37)[name = tensor("reshape_186")]; + tensor reshape_187_shape_0 = const()[name = tensor("reshape_187_shape_0"), val = tensor([-1])]; + tensor reshape_187_cast_fp16 = reshape(shape = reshape_187_shape_0, x = var_7850_cast_fp16)[name = tensor("reshape_187_cast_fp16")]; + tensor reshape_188_shape_0 = const()[name = tensor("reshape_188_shape_0"), val = tensor([-1])]; + tensor reshape_188_cast_fp16 = reshape(shape = reshape_188_shape_0, x = reshape_184_cast_fp16)[name = tensor("reshape_188_cast_fp16")]; + tensor scatter_37_mode_0 = const()[name = tensor("scatter_37_mode_0"), val = tensor("update")]; + tensor scatter_37_axis_0 = const()[name = tensor("scatter_37_axis_0"), val = tensor(0)]; + tensor scatter_37_validate_indices_0 = const()[name = tensor("scatter_37_validate_indices_0"), val = tensor(false)]; + tensor scatter_37_cast_fp16 = scatter(axis = scatter_37_axis_0, data = reshape_188_cast_fp16, indices = reshape_186, mode = scatter_37_mode_0, updates = reshape_187_cast_fp16, validate_indices = scatter_37_validate_indices_0)[name = tensor("scatter_37_cast_fp16")]; + tensor reshape_189_cast_fp16 = reshape(shape = shape_83, x = scatter_37_cast_fp16)[name = tensor("reshape_189_cast_fp16")]; + tensor reshape_189_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("reshape_189_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor keys_109_begin_0 = const()[name = tensor("keys_109_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor keys_109_end_0 = const()[name = tensor("keys_109_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor keys_109_end_mask_0 = const()[name = tensor("keys_109_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor keys_109_squeeze_mask_0 = const()[name = tensor("keys_109_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor keys_109_cast_fp16 = slice_by_index(begin = keys_109_begin_0, end = keys_109_end_0, end_mask = keys_109_end_mask_0, squeeze_mask = keys_109_squeeze_mask_0, x = reshape_189_cast_fp16)[name = tensor("keys_109_cast_fp16")]; + tensor values_109_begin_0 = const()[name = tensor("values_109_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor values_109_end_0 = const()[name = tensor("values_109_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor values_109_end_mask_0 = const()[name = tensor("values_109_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor values_109_squeeze_mask_0 = const()[name = tensor("values_109_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor values_109_cast_fp16 = slice_by_index(begin = values_109_begin_0, end = values_109_end_0, end_mask = values_109_end_mask_0, squeeze_mask = values_109_squeeze_mask_0, x = reshape_189_cast_fp16)[name = tensor("values_109_cast_fp16")]; + tensor var_7862_cast_fp16 = not_equal(x = keys_109_cast_fp16, y = keys_109_cast_fp16)[name = tensor("op_7862_cast_fp16")]; + tensor keys_111_cast_fp16 = select(a = var_470_to_fp16, b = keys_109_cast_fp16, cond = var_7862_cast_fp16)[name = tensor("keys_111_cast_fp16")]; + tensor var_7870_cast_fp16 = not_equal(x = values_109_cast_fp16, y = values_109_cast_fp16)[name = tensor("op_7870_cast_fp16")]; + tensor values_111_cast_fp16 = select(a = var_470_to_fp16, b = values_109_cast_fp16, cond = var_7870_cast_fp16)[name = tensor("values_111_cast_fp16")]; + tensor var_7894 = const()[name = tensor("op_7894"), val = tensor([0, 2, 1, 3])]; + tensor var_7910 = const()[name = tensor("op_7910"), val = tensor([1, 1, 1])]; + tensor var_7911_cast_fp16 = reshape(shape = var_7910, x = position18_to_fp16)[name = tensor("op_7911_cast_fp16")]; + tensor q_positions_37_cast_fp16 = add(x = var_7911_cast_fp16, y = q_offsets_1_promoted_to_fp16)[name = tensor("q_positions_37_cast_fp16")]; + tensor valid_end_37_cast_fp16 = add(x = var_7911_cast_fp16, y = var_536_cast_fp16)[name = tensor("valid_end_37_cast_fp16")]; + tensor valid_mask_37_cast_fp16 = less(x = k_positions_1_promoted_to_fp16, y = valid_end_37_cast_fp16)[name = tensor("valid_mask_37_cast_fp16")]; + tensor causal_mask_37_cast_fp16 = less_equal(x = k_positions_1_promoted_to_fp16, y = q_positions_37_cast_fp16)[name = tensor("causal_mask_37_cast_fp16")]; + tensor var_7939 = logical_and(x = valid_mask_37_cast_fp16, y = causal_mask_37_cast_fp16)[name = tensor("op_7939")]; + tensor attn_mask_37_axes_0 = const()[name = tensor("attn_mask_37_axes_0"), val = tensor([1])]; + tensor attn_mask_37 = expand_dims(axes = attn_mask_37_axes_0, x = var_7939)[name = tensor("attn_mask_37")]; + tensor var_7953_transpose_x_0 = const()[name = tensor("op_7953_transpose_x_0"), val = tensor(false)]; + tensor var_7953_transpose_y_0 = const()[name = tensor("op_7953_transpose_y_0"), val = tensor(false)]; + tensor transpose_105_perm_0 = const()[name = tensor("transpose_105_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_106_perm_0 = const()[name = tensor("transpose_106_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_106 = transpose(perm = transpose_106_perm_0, x = keys_111_cast_fp16)[name = tensor("transpose_132")]; + tensor transpose_105 = transpose(perm = transpose_105_perm_0, x = q_111_cast_fp16)[name = tensor("transpose_133")]; + tensor var_7953_cast_fp16 = matmul(transpose_x = var_7953_transpose_x_0, transpose_y = var_7953_transpose_y_0, x = transpose_105, y = transpose_106)[name = tensor("op_7953_cast_fp16")]; + tensor var_7947_to_fp16 = const()[name = tensor("op_7947_to_fp16"), val = tensor([0x1p-3])]; + tensor attn_weights_109_cast_fp16 = mul(x = var_7953_cast_fp16, y = var_7947_to_fp16)[name = tensor("attn_weights_109_cast_fp16")]; + tensor var_7955 = logical_not(x = attn_mask_37)[name = tensor("op_7955")]; + tensor var_7956_to_fp16 = const()[name = tensor("op_7956_to_fp16"), val = tensor(-inf)]; + tensor attn_weights_111_cast_fp16 = select(a = var_7956_to_fp16, b = attn_weights_109_cast_fp16, cond = var_7955)[name = tensor("attn_weights_111_cast_fp16")]; + tensor var_7958 = const()[name = tensor("op_7958"), val = tensor(-1)]; + tensor attn_weights_113_cast_fp16 = softmax(axis = var_7958, x = attn_weights_111_cast_fp16)[name = tensor("attn_weights_113_cast_fp16")]; + tensor attn_output_37_transpose_x_0 = const()[name = tensor("attn_output_37_transpose_x_0"), val = tensor(false)]; + tensor attn_output_37_transpose_y_0 = const()[name = tensor("attn_output_37_transpose_y_0"), val = tensor(false)]; + tensor values_113_cast_fp16 = transpose(perm = var_7894, x = values_111_cast_fp16)[name = tensor("transpose_134")]; + tensor attn_output_37_cast_fp16 = matmul(transpose_x = attn_output_37_transpose_x_0, transpose_y = attn_output_37_transpose_y_0, x = attn_weights_113_cast_fp16, y = values_113_cast_fp16)[name = tensor("attn_output_37_cast_fp16")]; + tensor var_7966 = const()[name = tensor("op_7966"), val = tensor([0, 2, 1, 3])]; + tensor var_7971 = const()[name = tensor("op_7971"), val = tensor([1, 256, 1024])]; + tensor var_7967_cast_fp16 = transpose(perm = var_7966, x = attn_output_37_cast_fp16)[name = tensor("transpose_131")]; + tensor input_221_cast_fp16 = reshape(shape = var_7971, x = var_7967_cast_fp16)[name = tensor("input_221_cast_fp16")]; + tensor attn18_out_proj_weight_to_fp16 = const()[name = tensor("attn18_out_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(460507776)))]; + tensor linear_73_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = attn18_out_proj_weight_to_fp16, x = input_221_cast_fp16)[name = tensor("linear_73_cast_fp16")]; + tensor var_7980_cast_fp16 = add(x = position18_to_fp16, y = valid_len_to_fp16)[name = tensor("op_7980_cast_fp16")]; + tensor var_7980_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_7980_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor input_223_cast_fp16 = add(x = input_217_cast_fp16, y = linear_73_cast_fp16)[name = tensor("input_223_cast_fp16")]; + tensor input_225_axes_0 = const()[name = tensor("input_225_axes_0"), val = tensor([-1])]; + tensor norm18_2_weight_to_fp16 = const()[name = tensor("norm18_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(462604992)))]; + tensor norm18_2_bias_to_fp16 = const()[name = tensor("norm18_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(462607104)))]; + tensor var_7984_to_fp16 = const()[name = tensor("op_7984_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_225_cast_fp16 = layer_norm(axes = input_225_axes_0, beta = norm18_2_bias_to_fp16, epsilon = var_7984_to_fp16, gamma = norm18_2_weight_to_fp16, x = input_223_cast_fp16)[name = tensor("input_225_cast_fp16")]; + tensor linear18_1_weight_to_fp16 = const()[name = tensor("linear18_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(462609216)))]; + tensor linear_74_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = linear18_1_weight_to_fp16, x = input_225_cast_fp16)[name = tensor("linear_74_cast_fp16")]; + tensor input_227_mode_0 = const()[name = tensor("input_227_mode_0"), val = tensor("EXACT")]; + tensor input_227_cast_fp16 = gelu(mode = input_227_mode_0, x = linear_74_cast_fp16)[name = tensor("input_227_cast_fp16")]; + tensor linear18_2_weight_to_fp16 = const()[name = tensor("linear18_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(470997888)))]; + tensor linear_75_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = linear18_2_weight_to_fp16, x = input_227_cast_fp16)[name = tensor("linear_75_cast_fp16")]; + tensor input_229_cast_fp16 = add(x = input_223_cast_fp16, y = linear_75_cast_fp16)[name = tensor("input_229_cast_fp16")]; + tensor input_231_axes_0 = const()[name = tensor("input_231_axes_0"), val = tensor([-1])]; + tensor norm19_1_weight_to_fp16 = const()[name = tensor("norm19_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(479386560)))]; + tensor norm19_1_bias_to_fp16 = const()[name = tensor("norm19_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(479388672)))]; + tensor var_8001_to_fp16 = const()[name = tensor("op_8001_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_231_cast_fp16 = layer_norm(axes = input_231_axes_0, beta = norm19_1_bias_to_fp16, epsilon = var_8001_to_fp16, gamma = norm19_1_weight_to_fp16, x = input_229_cast_fp16)[name = tensor("input_231_cast_fp16")]; + tensor attn19_in_proj_weight_to_fp16 = const()[name = tensor("attn19_in_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(479390784)))]; + tensor linear_76_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = attn19_in_proj_weight_to_fp16, x = input_231_cast_fp16)[name = tensor("linear_76_cast_fp16")]; + tensor var_8015 = const()[name = tensor("op_8015"), val = tensor([1, 256, 3, 16, 64])]; + tensor qkv_39_cast_fp16 = reshape(shape = var_8015, x = linear_76_cast_fp16)[name = tensor("qkv_39_cast_fp16")]; + tensor q_115_begin_0 = const()[name = tensor("q_115_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor q_115_end_0 = const()[name = tensor("q_115_end_0"), val = tensor([1, 256, 1, 16, 64])]; + tensor q_115_end_mask_0 = const()[name = tensor("q_115_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor q_115_squeeze_mask_0 = const()[name = tensor("q_115_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor q_115_cast_fp16 = slice_by_index(begin = q_115_begin_0, end = q_115_end_0, end_mask = q_115_end_mask_0, squeeze_mask = q_115_squeeze_mask_0, x = qkv_39_cast_fp16)[name = tensor("q_115_cast_fp16")]; + tensor k_77_begin_0 = const()[name = tensor("k_77_begin_0"), val = tensor([0, 0, 1, 0, 0])]; + tensor k_77_end_0 = const()[name = tensor("k_77_end_0"), val = tensor([1, 256, 2, 16, 64])]; + tensor k_77_end_mask_0 = const()[name = tensor("k_77_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor k_77_squeeze_mask_0 = const()[name = tensor("k_77_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor k_77_cast_fp16 = slice_by_index(begin = k_77_begin_0, end = k_77_end_0, end_mask = k_77_end_mask_0, squeeze_mask = k_77_squeeze_mask_0, x = qkv_39_cast_fp16)[name = tensor("k_77_cast_fp16")]; + tensor v_39_begin_0 = const()[name = tensor("v_39_begin_0"), val = tensor([0, 0, 2, 0, 0])]; + tensor v_39_end_0 = const()[name = tensor("v_39_end_0"), val = tensor([1, 256, 3, 16, 64])]; + tensor v_39_end_mask_0 = const()[name = tensor("v_39_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor v_39_squeeze_mask_0 = const()[name = tensor("v_39_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor v_39_cast_fp16 = slice_by_index(begin = v_39_begin_0, end = v_39_end_0, end_mask = v_39_end_mask_0, squeeze_mask = v_39_squeeze_mask_0, x = qkv_39_cast_fp16)[name = tensor("v_39_cast_fp16")]; + tensor var_8073 = const()[name = tensor("op_8073"), val = tensor([1, 1])]; + tensor position19_to_fp16_dtype_0 = const()[name = tensor("position19_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor position19_to_fp16 = cast(dtype = position19_to_fp16_dtype_0, x = position19)[name = tensor("cast_566")]; + tensor var_8074_cast_fp16 = reshape(shape = var_8073, x = position19_to_fp16)[name = tensor("op_8074_cast_fp16")]; + tensor ts_117_cast_fp16 = add(x = ts_1_promoted_to_fp16, y = var_8074_cast_fp16)[name = tensor("ts_117_cast_fp16")]; + tensor var_8081 = const()[name = tensor("op_8081"), val = tensor([1, 256, 1, 1])]; + tensor ts_119_cast_fp16 = reshape(shape = var_8081, x = ts_117_cast_fp16)[name = tensor("ts_119_cast_fp16")]; + tensor var_8088 = const()[name = tensor("op_8088"), val = tensor([1, 256, 16, 32, 2])]; + tensor q_complex_39_cast_fp16 = reshape(shape = var_8088, x = q_115_cast_fp16)[name = tensor("q_complex_39_cast_fp16")]; + tensor var_8095 = const()[name = tensor("op_8095"), val = tensor([1, 256, 16, 32, 2])]; + tensor k_complex_39_cast_fp16 = reshape(shape = var_8095, x = k_77_cast_fp16)[name = tensor("k_complex_39_cast_fp16")]; + tensor var_8099_begin_0 = const()[name = tensor("op_8099_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_8099_end_0 = const()[name = tensor("op_8099_end_0"), val = tensor([1, 256, 16, 32, 1])]; + tensor var_8099_end_mask_0 = const()[name = tensor("op_8099_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_8099_squeeze_mask_0 = const()[name = tensor("op_8099_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_8099_cast_fp16 = slice_by_index(begin = var_8099_begin_0, end = var_8099_end_0, end_mask = var_8099_end_mask_0, squeeze_mask = var_8099_squeeze_mask_0, x = q_complex_39_cast_fp16)[name = tensor("op_8099_cast_fp16")]; + tensor var_8107_begin_0 = const()[name = tensor("op_8107_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_8107_end_0 = const()[name = tensor("op_8107_end_0"), val = tensor([1, 256, 16, 32, 2])]; + tensor var_8107_end_mask_0 = const()[name = tensor("op_8107_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_8107_squeeze_mask_0 = const()[name = tensor("op_8107_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_8107_cast_fp16 = slice_by_index(begin = var_8107_begin_0, end = var_8107_end_0, end_mask = var_8107_end_mask_0, squeeze_mask = var_8107_squeeze_mask_0, x = q_complex_39_cast_fp16)[name = tensor("op_8107_cast_fp16")]; + tensor var_8115_begin_0 = const()[name = tensor("op_8115_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_8115_end_0 = const()[name = tensor("op_8115_end_0"), val = tensor([1, 256, 16, 32, 1])]; + tensor var_8115_end_mask_0 = const()[name = tensor("op_8115_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_8115_squeeze_mask_0 = const()[name = tensor("op_8115_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_8115_cast_fp16 = slice_by_index(begin = var_8115_begin_0, end = var_8115_end_0, end_mask = var_8115_end_mask_0, squeeze_mask = var_8115_squeeze_mask_0, x = k_complex_39_cast_fp16)[name = tensor("op_8115_cast_fp16")]; + tensor var_8123_begin_0 = const()[name = tensor("op_8123_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_8123_end_0 = const()[name = tensor("op_8123_end_0"), val = tensor([1, 256, 16, 32, 2])]; + tensor var_8123_end_mask_0 = const()[name = tensor("op_8123_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_8123_squeeze_mask_0 = const()[name = tensor("op_8123_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_8123_cast_fp16 = slice_by_index(begin = var_8123_begin_0, end = var_8123_end_0, end_mask = var_8123_end_mask_0, squeeze_mask = var_8123_squeeze_mask_0, x = k_complex_39_cast_fp16)[name = tensor("op_8123_cast_fp16")]; + tensor freqs_39_to_fp16 = const()[name = tensor("freqs_39_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(485682304)))]; + tensor var_8129_cast_fp16 = mul(x = freqs_39_to_fp16, y = ts_119_cast_fp16)[name = tensor("op_8129_cast_fp16")]; + tensor rotr_39_cast_fp16 = cos(x = var_8129_cast_fp16)[name = tensor("rotr_39_cast_fp16")]; + tensor roti_39_cast_fp16 = sin(x = var_8129_cast_fp16)[name = tensor("roti_39_cast_fp16")]; + tensor var_8133_cast_fp16 = mul(x = var_8099_cast_fp16, y = rotr_39_cast_fp16)[name = tensor("op_8133_cast_fp16")]; + tensor var_8134_cast_fp16 = mul(x = var_8107_cast_fp16, y = roti_39_cast_fp16)[name = tensor("op_8134_cast_fp16")]; + tensor qor_77_cast_fp16 = sub(x = var_8133_cast_fp16, y = var_8134_cast_fp16)[name = tensor("qor_77_cast_fp16")]; + tensor var_8137_cast_fp16 = mul(x = var_8099_cast_fp16, y = roti_39_cast_fp16)[name = tensor("op_8137_cast_fp16")]; + tensor var_8138_cast_fp16 = mul(x = var_8107_cast_fp16, y = rotr_39_cast_fp16)[name = tensor("op_8138_cast_fp16")]; + tensor qoi_77_cast_fp16 = add(x = var_8137_cast_fp16, y = var_8138_cast_fp16)[name = tensor("qoi_77_cast_fp16")]; + tensor var_8141_cast_fp16 = mul(x = var_8115_cast_fp16, y = rotr_39_cast_fp16)[name = tensor("op_8141_cast_fp16")]; + tensor var_8142_cast_fp16 = mul(x = var_8123_cast_fp16, y = roti_39_cast_fp16)[name = tensor("op_8142_cast_fp16")]; + tensor kor_77_cast_fp16 = sub(x = var_8141_cast_fp16, y = var_8142_cast_fp16)[name = tensor("kor_77_cast_fp16")]; + tensor var_8145_cast_fp16 = mul(x = var_8115_cast_fp16, y = roti_39_cast_fp16)[name = tensor("op_8145_cast_fp16")]; + tensor var_8146_cast_fp16 = mul(x = var_8123_cast_fp16, y = rotr_39_cast_fp16)[name = tensor("op_8146_cast_fp16")]; + tensor koi_77_cast_fp16 = add(x = var_8145_cast_fp16, y = var_8146_cast_fp16)[name = tensor("koi_77_cast_fp16")]; + tensor qo_39_axis_0 = const()[name = tensor("qo_39_axis_0"), val = tensor(-1)]; + tensor qo_39_cast_fp16 = stack(axis = qo_39_axis_0, values = (qor_77_cast_fp16, qoi_77_cast_fp16))[name = tensor("qo_39_cast_fp16")]; + tensor ko_39_axis_0 = const()[name = tensor("ko_39_axis_0"), val = tensor(-1)]; + tensor ko_39_cast_fp16 = stack(axis = ko_39_axis_0, values = (kor_77_cast_fp16, koi_77_cast_fp16))[name = tensor("ko_39_cast_fp16")]; + tensor var_8179 = const()[name = tensor("op_8179"), val = tensor([1, 256, 16, 64])]; + tensor q_117_cast_fp16 = reshape(shape = var_8179, x = qo_39_cast_fp16)[name = tensor("q_117_cast_fp16")]; + tensor var_8185 = const()[name = tensor("op_8185"), val = tensor([1, 256, 16, 64])]; + tensor k_79_cast_fp16 = reshape(shape = var_8185, x = ko_39_cast_fp16)[name = tensor("k_79_cast_fp16")]; + tensor write_indices_float_115_cast_fp16 = add(x = var_8074_cast_fp16, y = var_393_promoted_to_fp16)[name = tensor("write_indices_float_115_cast_fp16")]; + tensor _inversed_8206_y_0_to_fp16 = const()[name = tensor("_inversed_8206_y_0_to_fp16"), val = tensor(0x1p-9)]; + tensor _inversed_8206_cast_fp16 = mul(x = write_indices_float_115_cast_fp16, y = _inversed_8206_y_0_to_fp16)[name = tensor("_inversed_8206_cast_fp16")]; + tensor var_8207_cast_fp16 = floor(x = _inversed_8206_cast_fp16)[name = tensor("op_8207_cast_fp16")]; + tensor var_8208_to_fp16 = const()[name = tensor("op_8208_to_fp16"), val = tensor(0x1p+9)]; + tensor var_8209_cast_fp16 = mul(x = var_8207_cast_fp16, y = var_8208_to_fp16)[name = tensor("op_8209_cast_fp16")]; + tensor write_indices_float_117_cast_fp16 = sub(x = write_indices_float_115_cast_fp16, y = var_8209_cast_fp16)[name = tensor("write_indices_float_117_cast_fp16")]; + tensor var_8227_value_0_to_fp16 = const()[name = tensor("op_8227_value_0_to_fp16"), val = tensor(0x1.ffp+8)]; + tensor var_8227_cast_fp16 = fill_like(ref_tensor = write_indices_float_117_cast_fp16, value = var_8227_value_0_to_fp16)[name = tensor("op_8227_cast_fp16")]; + tensor write_indices_float_119_cast_fp16 = select(a = write_indices_float_117_cast_fp16, b = var_8227_cast_fp16, cond = token_valid_1_cast_fp16)[name = tensor("write_indices_float_119_cast_fp16")]; + tensor var_8233_dtype_0 = const()[name = tensor("op_8233_dtype_0"), val = tensor("int32")]; + tensor var_8238 = const()[name = tensor("op_8238"), val = tensor([1, 256, 1, 1])]; + tensor write_indices_float_119_cast_fp16_to_int32 = cast(dtype = var_8233_dtype_0, x = write_indices_float_119_cast_fp16)[name = tensor("cast_565")]; + tensor var_8239 = reshape(shape = var_8238, x = write_indices_float_119_cast_fp16_to_int32)[name = tensor("op_8239")]; + tensor write_indices_39_reps_0 = const()[name = tensor("write_indices_39_reps_0"), val = tensor([1, 1, 16, 64])]; + tensor write_indices_39 = tile(reps = write_indices_39_reps_0, x = var_8239)[name = tensor("write_indices_39")]; + tensor var_8249_begin_0 = const()[name = tensor("op_8249_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_8249_end_0 = const()[name = tensor("op_8249_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor var_8249_end_mask_0 = const()[name = tensor("op_8249_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_8249_squeeze_mask_0 = const()[name = tensor("op_8249_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor cache19_to_fp16_dtype_0 = const()[name = tensor("cache19_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor cache19_to_fp16 = cast(dtype = cache19_to_fp16_dtype_0, x = cache19)[name = tensor("cast_564")]; + tensor var_8249_cast_fp16 = slice_by_index(begin = var_8249_begin_0, end = var_8249_end_0, end_mask = var_8249_end_mask_0, squeeze_mask = var_8249_squeeze_mask_0, x = cache19_to_fp16)[name = tensor("op_8249_cast_fp16")]; + tensor var_8251_axis_0 = const()[name = tensor("op_8251_axis_0"), val = tensor(1)]; + tensor var_8251_mode_0 = const()[name = tensor("op_8251_mode_0"), val = tensor("update")]; + tensor var_8251_validate_indices_0 = const()[name = tensor("op_8251_validate_indices_0"), val = tensor(false)]; + tensor var_8251_cast_fp16 = scatter_along_axis(axis = var_8251_axis_0, data = var_8249_cast_fp16, indices = write_indices_39, mode = var_8251_mode_0, updates = k_79_cast_fp16, validate_indices = var_8251_validate_indices_0)[name = tensor("op_8251_cast_fp16")]; + tensor concat_154 = const()[name = tensor("concat_154"), val = tensor([0, 0, 0, 0, 0])]; + tensor concat_155 = const()[name = tensor("concat_155"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_39_internal_tensor_assign_1_stride_0 = const()[name = tensor("new_cache_39_internal_tensor_assign_1_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_39_internal_tensor_assign_1_begin_mask_0 = const()[name = tensor("new_cache_39_internal_tensor_assign_1_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_39_internal_tensor_assign_1_end_mask_0 = const()[name = tensor("new_cache_39_internal_tensor_assign_1_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_39_internal_tensor_assign_1_squeeze_mask_0 = const()[name = tensor("new_cache_39_internal_tensor_assign_1_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_84 = const()[name = tensor("shape_84"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_38 = const()[name = tensor("reduce_prod_38"), val = tensor(1048576)]; + tensor range_1d_38_start_0 = const()[name = tensor("range_1d_38_start_0"), val = tensor(0)]; + tensor range_1d_38_step_0 = const()[name = tensor("range_1d_38_step_0"), val = tensor(1)]; + tensor range_1d_38 = range_1d(end = reduce_prod_38, start = range_1d_38_start_0, step = range_1d_38_step_0)[name = tensor("range_1d_38")]; + tensor reshape_190 = reshape(shape = shape_84, x = range_1d_38)[name = tensor("reshape_190")]; + tensor slice_by_index_38 = slice_by_index(begin = concat_154, begin_mask = new_cache_39_internal_tensor_assign_1_begin_mask_0, end = concat_155, end_mask = new_cache_39_internal_tensor_assign_1_end_mask_0, squeeze_mask = new_cache_39_internal_tensor_assign_1_squeeze_mask_0, stride = new_cache_39_internal_tensor_assign_1_stride_0, x = reshape_190)[name = tensor("slice_by_index_38")]; + tensor reshape_191_shape_0 = const()[name = tensor("reshape_191_shape_0"), val = tensor([-1])]; + tensor reshape_191 = reshape(shape = reshape_191_shape_0, x = slice_by_index_38)[name = tensor("reshape_191")]; + tensor reshape_192_shape_0 = const()[name = tensor("reshape_192_shape_0"), val = tensor([-1])]; + tensor reshape_192_cast_fp16 = reshape(shape = reshape_192_shape_0, x = var_8251_cast_fp16)[name = tensor("reshape_192_cast_fp16")]; + tensor reshape_193_shape_0 = const()[name = tensor("reshape_193_shape_0"), val = tensor([-1])]; + tensor reshape_193_cast_fp16 = reshape(shape = reshape_193_shape_0, x = cache19_to_fp16)[name = tensor("reshape_193_cast_fp16")]; + tensor scatter_38_mode_0 = const()[name = tensor("scatter_38_mode_0"), val = tensor("update")]; + tensor scatter_38_axis_0 = const()[name = tensor("scatter_38_axis_0"), val = tensor(0)]; + tensor scatter_38_validate_indices_0 = const()[name = tensor("scatter_38_validate_indices_0"), val = tensor(false)]; + tensor scatter_38_cast_fp16 = scatter(axis = scatter_38_axis_0, data = reshape_193_cast_fp16, indices = reshape_191, mode = scatter_38_mode_0, updates = reshape_192_cast_fp16, validate_indices = scatter_38_validate_indices_0)[name = tensor("scatter_38_cast_fp16")]; + tensor reshape_194_cast_fp16 = reshape(shape = shape_84, x = scatter_38_cast_fp16)[name = tensor("reshape_194_cast_fp16")]; + tensor var_8259_begin_0 = const()[name = tensor("op_8259_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor var_8259_end_0 = const()[name = tensor("op_8259_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor var_8259_end_mask_0 = const()[name = tensor("op_8259_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_8259_squeeze_mask_0 = const()[name = tensor("op_8259_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor var_8259_cast_fp16 = slice_by_index(begin = var_8259_begin_0, end = var_8259_end_0, end_mask = var_8259_end_mask_0, squeeze_mask = var_8259_squeeze_mask_0, x = reshape_194_cast_fp16)[name = tensor("op_8259_cast_fp16")]; + tensor var_8261_axis_0 = const()[name = tensor("op_8261_axis_0"), val = tensor(1)]; + tensor var_8261_mode_0 = const()[name = tensor("op_8261_mode_0"), val = tensor("update")]; + tensor var_8261_validate_indices_0 = const()[name = tensor("op_8261_validate_indices_0"), val = tensor(false)]; + tensor var_8261_cast_fp16 = scatter_along_axis(axis = var_8261_axis_0, data = var_8259_cast_fp16, indices = write_indices_39, mode = var_8261_mode_0, updates = v_39_cast_fp16, validate_indices = var_8261_validate_indices_0)[name = tensor("op_8261_cast_fp16")]; + tensor concat_156 = const()[name = tensor("concat_156"), val = tensor([1, 0, 0, 0, 0])]; + tensor concat_157 = const()[name = tensor("concat_157"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_39_internal_tensor_assign_2_stride_0 = const()[name = tensor("new_cache_39_internal_tensor_assign_2_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_39_internal_tensor_assign_2_begin_mask_0 = const()[name = tensor("new_cache_39_internal_tensor_assign_2_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_39_internal_tensor_assign_2_end_mask_0 = const()[name = tensor("new_cache_39_internal_tensor_assign_2_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_39_internal_tensor_assign_2_squeeze_mask_0 = const()[name = tensor("new_cache_39_internal_tensor_assign_2_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_85 = const()[name = tensor("shape_85"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_39 = const()[name = tensor("reduce_prod_39"), val = tensor(1048576)]; + tensor range_1d_39_start_0 = const()[name = tensor("range_1d_39_start_0"), val = tensor(0)]; + tensor range_1d_39_step_0 = const()[name = tensor("range_1d_39_step_0"), val = tensor(1)]; + tensor range_1d_39 = range_1d(end = reduce_prod_39, start = range_1d_39_start_0, step = range_1d_39_step_0)[name = tensor("range_1d_39")]; + tensor reshape_195 = reshape(shape = shape_85, x = range_1d_39)[name = tensor("reshape_195")]; + tensor slice_by_index_39 = slice_by_index(begin = concat_156, begin_mask = new_cache_39_internal_tensor_assign_2_begin_mask_0, end = concat_157, end_mask = new_cache_39_internal_tensor_assign_2_end_mask_0, squeeze_mask = new_cache_39_internal_tensor_assign_2_squeeze_mask_0, stride = new_cache_39_internal_tensor_assign_2_stride_0, x = reshape_195)[name = tensor("slice_by_index_39")]; + tensor reshape_196_shape_0 = const()[name = tensor("reshape_196_shape_0"), val = tensor([-1])]; + tensor reshape_196 = reshape(shape = reshape_196_shape_0, x = slice_by_index_39)[name = tensor("reshape_196")]; + tensor reshape_197_shape_0 = const()[name = tensor("reshape_197_shape_0"), val = tensor([-1])]; + tensor reshape_197_cast_fp16 = reshape(shape = reshape_197_shape_0, x = var_8261_cast_fp16)[name = tensor("reshape_197_cast_fp16")]; + tensor reshape_198_shape_0 = const()[name = tensor("reshape_198_shape_0"), val = tensor([-1])]; + tensor reshape_198_cast_fp16 = reshape(shape = reshape_198_shape_0, x = reshape_194_cast_fp16)[name = tensor("reshape_198_cast_fp16")]; + tensor scatter_39_mode_0 = const()[name = tensor("scatter_39_mode_0"), val = tensor("update")]; + tensor scatter_39_axis_0 = const()[name = tensor("scatter_39_axis_0"), val = tensor(0)]; + tensor scatter_39_validate_indices_0 = const()[name = tensor("scatter_39_validate_indices_0"), val = tensor(false)]; + tensor scatter_39_cast_fp16 = scatter(axis = scatter_39_axis_0, data = reshape_198_cast_fp16, indices = reshape_196, mode = scatter_39_mode_0, updates = reshape_197_cast_fp16, validate_indices = scatter_39_validate_indices_0)[name = tensor("scatter_39_cast_fp16")]; + tensor reshape_199_cast_fp16 = reshape(shape = shape_85, x = scatter_39_cast_fp16)[name = tensor("reshape_199_cast_fp16")]; + tensor reshape_199_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("reshape_199_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor keys_115_begin_0 = const()[name = tensor("keys_115_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor keys_115_end_0 = const()[name = tensor("keys_115_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor keys_115_end_mask_0 = const()[name = tensor("keys_115_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor keys_115_squeeze_mask_0 = const()[name = tensor("keys_115_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor keys_115_cast_fp16 = slice_by_index(begin = keys_115_begin_0, end = keys_115_end_0, end_mask = keys_115_end_mask_0, squeeze_mask = keys_115_squeeze_mask_0, x = reshape_199_cast_fp16)[name = tensor("keys_115_cast_fp16")]; + tensor values_115_begin_0 = const()[name = tensor("values_115_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor values_115_end_0 = const()[name = tensor("values_115_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor values_115_end_mask_0 = const()[name = tensor("values_115_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor values_115_squeeze_mask_0 = const()[name = tensor("values_115_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor values_115_cast_fp16 = slice_by_index(begin = values_115_begin_0, end = values_115_end_0, end_mask = values_115_end_mask_0, squeeze_mask = values_115_squeeze_mask_0, x = reshape_199_cast_fp16)[name = tensor("values_115_cast_fp16")]; + tensor var_8273_cast_fp16 = not_equal(x = keys_115_cast_fp16, y = keys_115_cast_fp16)[name = tensor("op_8273_cast_fp16")]; + tensor keys_117_cast_fp16 = select(a = var_470_to_fp16, b = keys_115_cast_fp16, cond = var_8273_cast_fp16)[name = tensor("keys_117_cast_fp16")]; + tensor var_8281_cast_fp16 = not_equal(x = values_115_cast_fp16, y = values_115_cast_fp16)[name = tensor("op_8281_cast_fp16")]; + tensor values_117_cast_fp16 = select(a = var_470_to_fp16, b = values_115_cast_fp16, cond = var_8281_cast_fp16)[name = tensor("values_117_cast_fp16")]; + tensor var_8305 = const()[name = tensor("op_8305"), val = tensor([0, 2, 1, 3])]; + tensor var_8321 = const()[name = tensor("op_8321"), val = tensor([1, 1, 1])]; + tensor var_8322_cast_fp16 = reshape(shape = var_8321, x = position19_to_fp16)[name = tensor("op_8322_cast_fp16")]; + tensor q_positions_39_cast_fp16 = add(x = var_8322_cast_fp16, y = q_offsets_1_promoted_to_fp16)[name = tensor("q_positions_39_cast_fp16")]; + tensor valid_end_39_cast_fp16 = add(x = var_8322_cast_fp16, y = var_536_cast_fp16)[name = tensor("valid_end_39_cast_fp16")]; + tensor valid_mask_39_cast_fp16 = less(x = k_positions_1_promoted_to_fp16, y = valid_end_39_cast_fp16)[name = tensor("valid_mask_39_cast_fp16")]; + tensor causal_mask_39_cast_fp16 = less_equal(x = k_positions_1_promoted_to_fp16, y = q_positions_39_cast_fp16)[name = tensor("causal_mask_39_cast_fp16")]; + tensor var_8350 = logical_and(x = valid_mask_39_cast_fp16, y = causal_mask_39_cast_fp16)[name = tensor("op_8350")]; + tensor attn_mask_39_axes_0 = const()[name = tensor("attn_mask_39_axes_0"), val = tensor([1])]; + tensor attn_mask_39 = expand_dims(axes = attn_mask_39_axes_0, x = var_8350)[name = tensor("attn_mask_39")]; + tensor var_8364_transpose_x_0 = const()[name = tensor("op_8364_transpose_x_0"), val = tensor(false)]; + tensor var_8364_transpose_y_0 = const()[name = tensor("op_8364_transpose_y_0"), val = tensor(false)]; + tensor transpose_107_perm_0 = const()[name = tensor("transpose_107_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_108_perm_0 = const()[name = tensor("transpose_108_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_108 = transpose(perm = transpose_108_perm_0, x = keys_117_cast_fp16)[name = tensor("transpose_128")]; + tensor transpose_107 = transpose(perm = transpose_107_perm_0, x = q_117_cast_fp16)[name = tensor("transpose_129")]; + tensor var_8364_cast_fp16 = matmul(transpose_x = var_8364_transpose_x_0, transpose_y = var_8364_transpose_y_0, x = transpose_107, y = transpose_108)[name = tensor("op_8364_cast_fp16")]; + tensor var_8358_to_fp16 = const()[name = tensor("op_8358_to_fp16"), val = tensor([0x1p-3])]; + tensor attn_weights_115_cast_fp16 = mul(x = var_8364_cast_fp16, y = var_8358_to_fp16)[name = tensor("attn_weights_115_cast_fp16")]; + tensor var_8366 = logical_not(x = attn_mask_39)[name = tensor("op_8366")]; + tensor var_8367_to_fp16 = const()[name = tensor("op_8367_to_fp16"), val = tensor(-inf)]; + tensor attn_weights_117_cast_fp16 = select(a = var_8367_to_fp16, b = attn_weights_115_cast_fp16, cond = var_8366)[name = tensor("attn_weights_117_cast_fp16")]; + tensor var_8369 = const()[name = tensor("op_8369"), val = tensor(-1)]; + tensor attn_weights_119_cast_fp16 = softmax(axis = var_8369, x = attn_weights_117_cast_fp16)[name = tensor("attn_weights_119_cast_fp16")]; + tensor attn_output_39_transpose_x_0 = const()[name = tensor("attn_output_39_transpose_x_0"), val = tensor(false)]; + tensor attn_output_39_transpose_y_0 = const()[name = tensor("attn_output_39_transpose_y_0"), val = tensor(false)]; + tensor values_119_cast_fp16 = transpose(perm = var_8305, x = values_117_cast_fp16)[name = tensor("transpose_130")]; + tensor attn_output_39_cast_fp16 = matmul(transpose_x = attn_output_39_transpose_x_0, transpose_y = attn_output_39_transpose_y_0, x = attn_weights_119_cast_fp16, y = values_119_cast_fp16)[name = tensor("attn_output_39_cast_fp16")]; + tensor var_8377 = const()[name = tensor("op_8377"), val = tensor([0, 2, 1, 3])]; + tensor var_8382 = const()[name = tensor("op_8382"), val = tensor([1, 256, 1024])]; + tensor var_8378_cast_fp16 = transpose(perm = var_8377, x = attn_output_39_cast_fp16)[name = tensor("transpose_127")]; + tensor input_233_cast_fp16 = reshape(shape = var_8382, x = var_8378_cast_fp16)[name = tensor("input_233_cast_fp16")]; + tensor attn19_out_proj_weight_to_fp16 = const()[name = tensor("attn19_out_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(485682432)))]; + tensor linear_77_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = attn19_out_proj_weight_to_fp16, x = input_233_cast_fp16)[name = tensor("linear_77_cast_fp16")]; + tensor var_8391_cast_fp16 = add(x = position19_to_fp16, y = valid_len_to_fp16)[name = tensor("op_8391_cast_fp16")]; + tensor var_8391_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_8391_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor input_235_cast_fp16 = add(x = input_229_cast_fp16, y = linear_77_cast_fp16)[name = tensor("input_235_cast_fp16")]; + tensor input_237_axes_0 = const()[name = tensor("input_237_axes_0"), val = tensor([-1])]; + tensor norm19_2_weight_to_fp16 = const()[name = tensor("norm19_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(487779648)))]; + tensor norm19_2_bias_to_fp16 = const()[name = tensor("norm19_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(487781760)))]; + tensor var_8395_to_fp16 = const()[name = tensor("op_8395_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_237_cast_fp16 = layer_norm(axes = input_237_axes_0, beta = norm19_2_bias_to_fp16, epsilon = var_8395_to_fp16, gamma = norm19_2_weight_to_fp16, x = input_235_cast_fp16)[name = tensor("input_237_cast_fp16")]; + tensor linear19_1_weight_to_fp16 = const()[name = tensor("linear19_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(487783872)))]; + tensor linear_78_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = linear19_1_weight_to_fp16, x = input_237_cast_fp16)[name = tensor("linear_78_cast_fp16")]; + tensor input_239_mode_0 = const()[name = tensor("input_239_mode_0"), val = tensor("EXACT")]; + tensor input_239_cast_fp16 = gelu(mode = input_239_mode_0, x = linear_78_cast_fp16)[name = tensor("input_239_cast_fp16")]; + tensor linear19_2_weight_to_fp16 = const()[name = tensor("linear19_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(496172544)))]; + tensor linear_79_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = linear19_2_weight_to_fp16, x = input_239_cast_fp16)[name = tensor("linear_79_cast_fp16")]; + tensor input_241_cast_fp16 = add(x = input_235_cast_fp16, y = linear_79_cast_fp16)[name = tensor("input_241_cast_fp16")]; + tensor input_243_axes_0 = const()[name = tensor("input_243_axes_0"), val = tensor([-1])]; + tensor norm20_1_weight_to_fp16 = const()[name = tensor("norm20_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(504561216)))]; + tensor norm20_1_bias_to_fp16 = const()[name = tensor("norm20_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(504563328)))]; + tensor var_8412_to_fp16 = const()[name = tensor("op_8412_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_243_cast_fp16 = layer_norm(axes = input_243_axes_0, beta = norm20_1_bias_to_fp16, epsilon = var_8412_to_fp16, gamma = norm20_1_weight_to_fp16, x = input_241_cast_fp16)[name = tensor("input_243_cast_fp16")]; + tensor attn20_in_proj_weight_to_fp16 = const()[name = tensor("attn20_in_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(504565440)))]; + tensor linear_80_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = attn20_in_proj_weight_to_fp16, x = input_243_cast_fp16)[name = tensor("linear_80_cast_fp16")]; + tensor var_8426 = const()[name = tensor("op_8426"), val = tensor([1, 256, 3, 16, 64])]; + tensor qkv_41_cast_fp16 = reshape(shape = var_8426, x = linear_80_cast_fp16)[name = tensor("qkv_41_cast_fp16")]; + tensor q_121_begin_0 = const()[name = tensor("q_121_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor q_121_end_0 = const()[name = tensor("q_121_end_0"), val = tensor([1, 256, 1, 16, 64])]; + tensor q_121_end_mask_0 = const()[name = tensor("q_121_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor q_121_squeeze_mask_0 = const()[name = tensor("q_121_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor q_121_cast_fp16 = slice_by_index(begin = q_121_begin_0, end = q_121_end_0, end_mask = q_121_end_mask_0, squeeze_mask = q_121_squeeze_mask_0, x = qkv_41_cast_fp16)[name = tensor("q_121_cast_fp16")]; + tensor k_81_begin_0 = const()[name = tensor("k_81_begin_0"), val = tensor([0, 0, 1, 0, 0])]; + tensor k_81_end_0 = const()[name = tensor("k_81_end_0"), val = tensor([1, 256, 2, 16, 64])]; + tensor k_81_end_mask_0 = const()[name = tensor("k_81_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor k_81_squeeze_mask_0 = const()[name = tensor("k_81_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor k_81_cast_fp16 = slice_by_index(begin = k_81_begin_0, end = k_81_end_0, end_mask = k_81_end_mask_0, squeeze_mask = k_81_squeeze_mask_0, x = qkv_41_cast_fp16)[name = tensor("k_81_cast_fp16")]; + tensor v_41_begin_0 = const()[name = tensor("v_41_begin_0"), val = tensor([0, 0, 2, 0, 0])]; + tensor v_41_end_0 = const()[name = tensor("v_41_end_0"), val = tensor([1, 256, 3, 16, 64])]; + tensor v_41_end_mask_0 = const()[name = tensor("v_41_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor v_41_squeeze_mask_0 = const()[name = tensor("v_41_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor v_41_cast_fp16 = slice_by_index(begin = v_41_begin_0, end = v_41_end_0, end_mask = v_41_end_mask_0, squeeze_mask = v_41_squeeze_mask_0, x = qkv_41_cast_fp16)[name = tensor("v_41_cast_fp16")]; + tensor var_8484 = const()[name = tensor("op_8484"), val = tensor([1, 1])]; + tensor position20_to_fp16_dtype_0 = const()[name = tensor("position20_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor position20_to_fp16 = cast(dtype = position20_to_fp16_dtype_0, x = position20)[name = tensor("cast_561")]; + tensor var_8485_cast_fp16 = reshape(shape = var_8484, x = position20_to_fp16)[name = tensor("op_8485_cast_fp16")]; + tensor ts_123_cast_fp16 = add(x = ts_1_promoted_to_fp16, y = var_8485_cast_fp16)[name = tensor("ts_123_cast_fp16")]; + tensor var_8492 = const()[name = tensor("op_8492"), val = tensor([1, 256, 1, 1])]; + tensor ts_125_cast_fp16 = reshape(shape = var_8492, x = ts_123_cast_fp16)[name = tensor("ts_125_cast_fp16")]; + tensor var_8499 = const()[name = tensor("op_8499"), val = tensor([1, 256, 16, 32, 2])]; + tensor q_complex_41_cast_fp16 = reshape(shape = var_8499, x = q_121_cast_fp16)[name = tensor("q_complex_41_cast_fp16")]; + tensor var_8506 = const()[name = tensor("op_8506"), val = tensor([1, 256, 16, 32, 2])]; + tensor k_complex_41_cast_fp16 = reshape(shape = var_8506, x = k_81_cast_fp16)[name = tensor("k_complex_41_cast_fp16")]; + tensor var_8510_begin_0 = const()[name = tensor("op_8510_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_8510_end_0 = const()[name = tensor("op_8510_end_0"), val = tensor([1, 256, 16, 32, 1])]; + tensor var_8510_end_mask_0 = const()[name = tensor("op_8510_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_8510_squeeze_mask_0 = const()[name = tensor("op_8510_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_8510_cast_fp16 = slice_by_index(begin = var_8510_begin_0, end = var_8510_end_0, end_mask = var_8510_end_mask_0, squeeze_mask = var_8510_squeeze_mask_0, x = q_complex_41_cast_fp16)[name = tensor("op_8510_cast_fp16")]; + tensor var_8518_begin_0 = const()[name = tensor("op_8518_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_8518_end_0 = const()[name = tensor("op_8518_end_0"), val = tensor([1, 256, 16, 32, 2])]; + tensor var_8518_end_mask_0 = const()[name = tensor("op_8518_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_8518_squeeze_mask_0 = const()[name = tensor("op_8518_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_8518_cast_fp16 = slice_by_index(begin = var_8518_begin_0, end = var_8518_end_0, end_mask = var_8518_end_mask_0, squeeze_mask = var_8518_squeeze_mask_0, x = q_complex_41_cast_fp16)[name = tensor("op_8518_cast_fp16")]; + tensor var_8526_begin_0 = const()[name = tensor("op_8526_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_8526_end_0 = const()[name = tensor("op_8526_end_0"), val = tensor([1, 256, 16, 32, 1])]; + tensor var_8526_end_mask_0 = const()[name = tensor("op_8526_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_8526_squeeze_mask_0 = const()[name = tensor("op_8526_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_8526_cast_fp16 = slice_by_index(begin = var_8526_begin_0, end = var_8526_end_0, end_mask = var_8526_end_mask_0, squeeze_mask = var_8526_squeeze_mask_0, x = k_complex_41_cast_fp16)[name = tensor("op_8526_cast_fp16")]; + tensor var_8534_begin_0 = const()[name = tensor("op_8534_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_8534_end_0 = const()[name = tensor("op_8534_end_0"), val = tensor([1, 256, 16, 32, 2])]; + tensor var_8534_end_mask_0 = const()[name = tensor("op_8534_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_8534_squeeze_mask_0 = const()[name = tensor("op_8534_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_8534_cast_fp16 = slice_by_index(begin = var_8534_begin_0, end = var_8534_end_0, end_mask = var_8534_end_mask_0, squeeze_mask = var_8534_squeeze_mask_0, x = k_complex_41_cast_fp16)[name = tensor("op_8534_cast_fp16")]; + tensor freqs_41_to_fp16 = const()[name = tensor("freqs_41_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(510856960)))]; + tensor var_8540_cast_fp16 = mul(x = freqs_41_to_fp16, y = ts_125_cast_fp16)[name = tensor("op_8540_cast_fp16")]; + tensor rotr_41_cast_fp16 = cos(x = var_8540_cast_fp16)[name = tensor("rotr_41_cast_fp16")]; + tensor roti_41_cast_fp16 = sin(x = var_8540_cast_fp16)[name = tensor("roti_41_cast_fp16")]; + tensor var_8544_cast_fp16 = mul(x = var_8510_cast_fp16, y = rotr_41_cast_fp16)[name = tensor("op_8544_cast_fp16")]; + tensor var_8545_cast_fp16 = mul(x = var_8518_cast_fp16, y = roti_41_cast_fp16)[name = tensor("op_8545_cast_fp16")]; + tensor qor_81_cast_fp16 = sub(x = var_8544_cast_fp16, y = var_8545_cast_fp16)[name = tensor("qor_81_cast_fp16")]; + tensor var_8548_cast_fp16 = mul(x = var_8510_cast_fp16, y = roti_41_cast_fp16)[name = tensor("op_8548_cast_fp16")]; + tensor var_8549_cast_fp16 = mul(x = var_8518_cast_fp16, y = rotr_41_cast_fp16)[name = tensor("op_8549_cast_fp16")]; + tensor qoi_81_cast_fp16 = add(x = var_8548_cast_fp16, y = var_8549_cast_fp16)[name = tensor("qoi_81_cast_fp16")]; + tensor var_8552_cast_fp16 = mul(x = var_8526_cast_fp16, y = rotr_41_cast_fp16)[name = tensor("op_8552_cast_fp16")]; + tensor var_8553_cast_fp16 = mul(x = var_8534_cast_fp16, y = roti_41_cast_fp16)[name = tensor("op_8553_cast_fp16")]; + tensor kor_81_cast_fp16 = sub(x = var_8552_cast_fp16, y = var_8553_cast_fp16)[name = tensor("kor_81_cast_fp16")]; + tensor var_8556_cast_fp16 = mul(x = var_8526_cast_fp16, y = roti_41_cast_fp16)[name = tensor("op_8556_cast_fp16")]; + tensor var_8557_cast_fp16 = mul(x = var_8534_cast_fp16, y = rotr_41_cast_fp16)[name = tensor("op_8557_cast_fp16")]; + tensor koi_81_cast_fp16 = add(x = var_8556_cast_fp16, y = var_8557_cast_fp16)[name = tensor("koi_81_cast_fp16")]; + tensor qo_41_axis_0 = const()[name = tensor("qo_41_axis_0"), val = tensor(-1)]; + tensor qo_41_cast_fp16 = stack(axis = qo_41_axis_0, values = (qor_81_cast_fp16, qoi_81_cast_fp16))[name = tensor("qo_41_cast_fp16")]; + tensor ko_41_axis_0 = const()[name = tensor("ko_41_axis_0"), val = tensor(-1)]; + tensor ko_41_cast_fp16 = stack(axis = ko_41_axis_0, values = (kor_81_cast_fp16, koi_81_cast_fp16))[name = tensor("ko_41_cast_fp16")]; + tensor var_8590 = const()[name = tensor("op_8590"), val = tensor([1, 256, 16, 64])]; + tensor q_123_cast_fp16 = reshape(shape = var_8590, x = qo_41_cast_fp16)[name = tensor("q_123_cast_fp16")]; + tensor var_8596 = const()[name = tensor("op_8596"), val = tensor([1, 256, 16, 64])]; + tensor k_83_cast_fp16 = reshape(shape = var_8596, x = ko_41_cast_fp16)[name = tensor("k_83_cast_fp16")]; + tensor write_indices_float_121_cast_fp16 = add(x = var_8485_cast_fp16, y = var_393_promoted_to_fp16)[name = tensor("write_indices_float_121_cast_fp16")]; + tensor _inversed_8617_y_0_to_fp16 = const()[name = tensor("_inversed_8617_y_0_to_fp16"), val = tensor(0x1p-9)]; + tensor _inversed_8617_cast_fp16 = mul(x = write_indices_float_121_cast_fp16, y = _inversed_8617_y_0_to_fp16)[name = tensor("_inversed_8617_cast_fp16")]; + tensor var_8618_cast_fp16 = floor(x = _inversed_8617_cast_fp16)[name = tensor("op_8618_cast_fp16")]; + tensor var_8619_to_fp16 = const()[name = tensor("op_8619_to_fp16"), val = tensor(0x1p+9)]; + tensor var_8620_cast_fp16 = mul(x = var_8618_cast_fp16, y = var_8619_to_fp16)[name = tensor("op_8620_cast_fp16")]; + tensor write_indices_float_123_cast_fp16 = sub(x = write_indices_float_121_cast_fp16, y = var_8620_cast_fp16)[name = tensor("write_indices_float_123_cast_fp16")]; + tensor var_8638_value_0_to_fp16 = const()[name = tensor("op_8638_value_0_to_fp16"), val = tensor(0x1.ffp+8)]; + tensor var_8638_cast_fp16 = fill_like(ref_tensor = write_indices_float_123_cast_fp16, value = var_8638_value_0_to_fp16)[name = tensor("op_8638_cast_fp16")]; + tensor write_indices_float_125_cast_fp16 = select(a = write_indices_float_123_cast_fp16, b = var_8638_cast_fp16, cond = token_valid_1_cast_fp16)[name = tensor("write_indices_float_125_cast_fp16")]; + tensor var_8644_dtype_0 = const()[name = tensor("op_8644_dtype_0"), val = tensor("int32")]; + tensor var_8649 = const()[name = tensor("op_8649"), val = tensor([1, 256, 1, 1])]; + tensor write_indices_float_125_cast_fp16_to_int32 = cast(dtype = var_8644_dtype_0, x = write_indices_float_125_cast_fp16)[name = tensor("cast_560")]; + tensor var_8650 = reshape(shape = var_8649, x = write_indices_float_125_cast_fp16_to_int32)[name = tensor("op_8650")]; + tensor write_indices_41_reps_0 = const()[name = tensor("write_indices_41_reps_0"), val = tensor([1, 1, 16, 64])]; + tensor write_indices_41 = tile(reps = write_indices_41_reps_0, x = var_8650)[name = tensor("write_indices_41")]; + tensor var_8660_begin_0 = const()[name = tensor("op_8660_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_8660_end_0 = const()[name = tensor("op_8660_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor var_8660_end_mask_0 = const()[name = tensor("op_8660_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_8660_squeeze_mask_0 = const()[name = tensor("op_8660_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor cache20_to_fp16_dtype_0 = const()[name = tensor("cache20_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor cache20_to_fp16 = cast(dtype = cache20_to_fp16_dtype_0, x = cache20)[name = tensor("cast_559")]; + tensor var_8660_cast_fp16 = slice_by_index(begin = var_8660_begin_0, end = var_8660_end_0, end_mask = var_8660_end_mask_0, squeeze_mask = var_8660_squeeze_mask_0, x = cache20_to_fp16)[name = tensor("op_8660_cast_fp16")]; + tensor var_8662_axis_0 = const()[name = tensor("op_8662_axis_0"), val = tensor(1)]; + tensor var_8662_mode_0 = const()[name = tensor("op_8662_mode_0"), val = tensor("update")]; + tensor var_8662_validate_indices_0 = const()[name = tensor("op_8662_validate_indices_0"), val = tensor(false)]; + tensor var_8662_cast_fp16 = scatter_along_axis(axis = var_8662_axis_0, data = var_8660_cast_fp16, indices = write_indices_41, mode = var_8662_mode_0, updates = k_83_cast_fp16, validate_indices = var_8662_validate_indices_0)[name = tensor("op_8662_cast_fp16")]; + tensor concat_162 = const()[name = tensor("concat_162"), val = tensor([0, 0, 0, 0, 0])]; + tensor concat_163 = const()[name = tensor("concat_163"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_41_internal_tensor_assign_1_stride_0 = const()[name = tensor("new_cache_41_internal_tensor_assign_1_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_41_internal_tensor_assign_1_begin_mask_0 = const()[name = tensor("new_cache_41_internal_tensor_assign_1_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_41_internal_tensor_assign_1_end_mask_0 = const()[name = tensor("new_cache_41_internal_tensor_assign_1_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_41_internal_tensor_assign_1_squeeze_mask_0 = const()[name = tensor("new_cache_41_internal_tensor_assign_1_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_86 = const()[name = tensor("shape_86"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_40 = const()[name = tensor("reduce_prod_40"), val = tensor(1048576)]; + tensor range_1d_40_start_0 = const()[name = tensor("range_1d_40_start_0"), val = tensor(0)]; + tensor range_1d_40_step_0 = const()[name = tensor("range_1d_40_step_0"), val = tensor(1)]; + tensor range_1d_40 = range_1d(end = reduce_prod_40, start = range_1d_40_start_0, step = range_1d_40_step_0)[name = tensor("range_1d_40")]; + tensor reshape_200 = reshape(shape = shape_86, x = range_1d_40)[name = tensor("reshape_200")]; + tensor slice_by_index_40 = slice_by_index(begin = concat_162, begin_mask = new_cache_41_internal_tensor_assign_1_begin_mask_0, end = concat_163, end_mask = new_cache_41_internal_tensor_assign_1_end_mask_0, squeeze_mask = new_cache_41_internal_tensor_assign_1_squeeze_mask_0, stride = new_cache_41_internal_tensor_assign_1_stride_0, x = reshape_200)[name = tensor("slice_by_index_40")]; + tensor reshape_201_shape_0 = const()[name = tensor("reshape_201_shape_0"), val = tensor([-1])]; + tensor reshape_201 = reshape(shape = reshape_201_shape_0, x = slice_by_index_40)[name = tensor("reshape_201")]; + tensor reshape_202_shape_0 = const()[name = tensor("reshape_202_shape_0"), val = tensor([-1])]; + tensor reshape_202_cast_fp16 = reshape(shape = reshape_202_shape_0, x = var_8662_cast_fp16)[name = tensor("reshape_202_cast_fp16")]; + tensor reshape_203_shape_0 = const()[name = tensor("reshape_203_shape_0"), val = tensor([-1])]; + tensor reshape_203_cast_fp16 = reshape(shape = reshape_203_shape_0, x = cache20_to_fp16)[name = tensor("reshape_203_cast_fp16")]; + tensor scatter_40_mode_0 = const()[name = tensor("scatter_40_mode_0"), val = tensor("update")]; + tensor scatter_40_axis_0 = const()[name = tensor("scatter_40_axis_0"), val = tensor(0)]; + tensor scatter_40_validate_indices_0 = const()[name = tensor("scatter_40_validate_indices_0"), val = tensor(false)]; + tensor scatter_40_cast_fp16 = scatter(axis = scatter_40_axis_0, data = reshape_203_cast_fp16, indices = reshape_201, mode = scatter_40_mode_0, updates = reshape_202_cast_fp16, validate_indices = scatter_40_validate_indices_0)[name = tensor("scatter_40_cast_fp16")]; + tensor reshape_204_cast_fp16 = reshape(shape = shape_86, x = scatter_40_cast_fp16)[name = tensor("reshape_204_cast_fp16")]; + tensor var_8670_begin_0 = const()[name = tensor("op_8670_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor var_8670_end_0 = const()[name = tensor("op_8670_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor var_8670_end_mask_0 = const()[name = tensor("op_8670_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_8670_squeeze_mask_0 = const()[name = tensor("op_8670_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor var_8670_cast_fp16 = slice_by_index(begin = var_8670_begin_0, end = var_8670_end_0, end_mask = var_8670_end_mask_0, squeeze_mask = var_8670_squeeze_mask_0, x = reshape_204_cast_fp16)[name = tensor("op_8670_cast_fp16")]; + tensor var_8672_axis_0 = const()[name = tensor("op_8672_axis_0"), val = tensor(1)]; + tensor var_8672_mode_0 = const()[name = tensor("op_8672_mode_0"), val = tensor("update")]; + tensor var_8672_validate_indices_0 = const()[name = tensor("op_8672_validate_indices_0"), val = tensor(false)]; + tensor var_8672_cast_fp16 = scatter_along_axis(axis = var_8672_axis_0, data = var_8670_cast_fp16, indices = write_indices_41, mode = var_8672_mode_0, updates = v_41_cast_fp16, validate_indices = var_8672_validate_indices_0)[name = tensor("op_8672_cast_fp16")]; + tensor concat_164 = const()[name = tensor("concat_164"), val = tensor([1, 0, 0, 0, 0])]; + tensor concat_165 = const()[name = tensor("concat_165"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_41_internal_tensor_assign_2_stride_0 = const()[name = tensor("new_cache_41_internal_tensor_assign_2_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_41_internal_tensor_assign_2_begin_mask_0 = const()[name = tensor("new_cache_41_internal_tensor_assign_2_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_41_internal_tensor_assign_2_end_mask_0 = const()[name = tensor("new_cache_41_internal_tensor_assign_2_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_41_internal_tensor_assign_2_squeeze_mask_0 = const()[name = tensor("new_cache_41_internal_tensor_assign_2_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_87 = const()[name = tensor("shape_87"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_41 = const()[name = tensor("reduce_prod_41"), val = tensor(1048576)]; + tensor range_1d_41_start_0 = const()[name = tensor("range_1d_41_start_0"), val = tensor(0)]; + tensor range_1d_41_step_0 = const()[name = tensor("range_1d_41_step_0"), val = tensor(1)]; + tensor range_1d_41 = range_1d(end = reduce_prod_41, start = range_1d_41_start_0, step = range_1d_41_step_0)[name = tensor("range_1d_41")]; + tensor reshape_205 = reshape(shape = shape_87, x = range_1d_41)[name = tensor("reshape_205")]; + tensor slice_by_index_41 = slice_by_index(begin = concat_164, begin_mask = new_cache_41_internal_tensor_assign_2_begin_mask_0, end = concat_165, end_mask = new_cache_41_internal_tensor_assign_2_end_mask_0, squeeze_mask = new_cache_41_internal_tensor_assign_2_squeeze_mask_0, stride = new_cache_41_internal_tensor_assign_2_stride_0, x = reshape_205)[name = tensor("slice_by_index_41")]; + tensor reshape_206_shape_0 = const()[name = tensor("reshape_206_shape_0"), val = tensor([-1])]; + tensor reshape_206 = reshape(shape = reshape_206_shape_0, x = slice_by_index_41)[name = tensor("reshape_206")]; + tensor reshape_207_shape_0 = const()[name = tensor("reshape_207_shape_0"), val = tensor([-1])]; + tensor reshape_207_cast_fp16 = reshape(shape = reshape_207_shape_0, x = var_8672_cast_fp16)[name = tensor("reshape_207_cast_fp16")]; + tensor reshape_208_shape_0 = const()[name = tensor("reshape_208_shape_0"), val = tensor([-1])]; + tensor reshape_208_cast_fp16 = reshape(shape = reshape_208_shape_0, x = reshape_204_cast_fp16)[name = tensor("reshape_208_cast_fp16")]; + tensor scatter_41_mode_0 = const()[name = tensor("scatter_41_mode_0"), val = tensor("update")]; + tensor scatter_41_axis_0 = const()[name = tensor("scatter_41_axis_0"), val = tensor(0)]; + tensor scatter_41_validate_indices_0 = const()[name = tensor("scatter_41_validate_indices_0"), val = tensor(false)]; + tensor scatter_41_cast_fp16 = scatter(axis = scatter_41_axis_0, data = reshape_208_cast_fp16, indices = reshape_206, mode = scatter_41_mode_0, updates = reshape_207_cast_fp16, validate_indices = scatter_41_validate_indices_0)[name = tensor("scatter_41_cast_fp16")]; + tensor reshape_209_cast_fp16 = reshape(shape = shape_87, x = scatter_41_cast_fp16)[name = tensor("reshape_209_cast_fp16")]; + tensor reshape_209_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("reshape_209_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor keys_121_begin_0 = const()[name = tensor("keys_121_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor keys_121_end_0 = const()[name = tensor("keys_121_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor keys_121_end_mask_0 = const()[name = tensor("keys_121_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor keys_121_squeeze_mask_0 = const()[name = tensor("keys_121_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor keys_121_cast_fp16 = slice_by_index(begin = keys_121_begin_0, end = keys_121_end_0, end_mask = keys_121_end_mask_0, squeeze_mask = keys_121_squeeze_mask_0, x = reshape_209_cast_fp16)[name = tensor("keys_121_cast_fp16")]; + tensor values_121_begin_0 = const()[name = tensor("values_121_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor values_121_end_0 = const()[name = tensor("values_121_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor values_121_end_mask_0 = const()[name = tensor("values_121_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor values_121_squeeze_mask_0 = const()[name = tensor("values_121_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor values_121_cast_fp16 = slice_by_index(begin = values_121_begin_0, end = values_121_end_0, end_mask = values_121_end_mask_0, squeeze_mask = values_121_squeeze_mask_0, x = reshape_209_cast_fp16)[name = tensor("values_121_cast_fp16")]; + tensor var_8684_cast_fp16 = not_equal(x = keys_121_cast_fp16, y = keys_121_cast_fp16)[name = tensor("op_8684_cast_fp16")]; + tensor keys_123_cast_fp16 = select(a = var_470_to_fp16, b = keys_121_cast_fp16, cond = var_8684_cast_fp16)[name = tensor("keys_123_cast_fp16")]; + tensor var_8692_cast_fp16 = not_equal(x = values_121_cast_fp16, y = values_121_cast_fp16)[name = tensor("op_8692_cast_fp16")]; + tensor values_123_cast_fp16 = select(a = var_470_to_fp16, b = values_121_cast_fp16, cond = var_8692_cast_fp16)[name = tensor("values_123_cast_fp16")]; + tensor var_8716 = const()[name = tensor("op_8716"), val = tensor([0, 2, 1, 3])]; + tensor var_8732 = const()[name = tensor("op_8732"), val = tensor([1, 1, 1])]; + tensor var_8733_cast_fp16 = reshape(shape = var_8732, x = position20_to_fp16)[name = tensor("op_8733_cast_fp16")]; + tensor q_positions_41_cast_fp16 = add(x = var_8733_cast_fp16, y = q_offsets_1_promoted_to_fp16)[name = tensor("q_positions_41_cast_fp16")]; + tensor valid_end_41_cast_fp16 = add(x = var_8733_cast_fp16, y = var_536_cast_fp16)[name = tensor("valid_end_41_cast_fp16")]; + tensor valid_mask_41_cast_fp16 = less(x = k_positions_1_promoted_to_fp16, y = valid_end_41_cast_fp16)[name = tensor("valid_mask_41_cast_fp16")]; + tensor causal_mask_41_cast_fp16 = less_equal(x = k_positions_1_promoted_to_fp16, y = q_positions_41_cast_fp16)[name = tensor("causal_mask_41_cast_fp16")]; + tensor var_8761 = logical_and(x = valid_mask_41_cast_fp16, y = causal_mask_41_cast_fp16)[name = tensor("op_8761")]; + tensor attn_mask_41_axes_0 = const()[name = tensor("attn_mask_41_axes_0"), val = tensor([1])]; + tensor attn_mask_41 = expand_dims(axes = attn_mask_41_axes_0, x = var_8761)[name = tensor("attn_mask_41")]; + tensor var_8775_transpose_x_0 = const()[name = tensor("op_8775_transpose_x_0"), val = tensor(false)]; + tensor var_8775_transpose_y_0 = const()[name = tensor("op_8775_transpose_y_0"), val = tensor(false)]; + tensor transpose_109_perm_0 = const()[name = tensor("transpose_109_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_110_perm_0 = const()[name = tensor("transpose_110_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_110 = transpose(perm = transpose_110_perm_0, x = keys_123_cast_fp16)[name = tensor("transpose_124")]; + tensor transpose_109 = transpose(perm = transpose_109_perm_0, x = q_123_cast_fp16)[name = tensor("transpose_125")]; + tensor var_8775_cast_fp16 = matmul(transpose_x = var_8775_transpose_x_0, transpose_y = var_8775_transpose_y_0, x = transpose_109, y = transpose_110)[name = tensor("op_8775_cast_fp16")]; + tensor var_8769_to_fp16 = const()[name = tensor("op_8769_to_fp16"), val = tensor([0x1p-3])]; + tensor attn_weights_121_cast_fp16 = mul(x = var_8775_cast_fp16, y = var_8769_to_fp16)[name = tensor("attn_weights_121_cast_fp16")]; + tensor var_8777 = logical_not(x = attn_mask_41)[name = tensor("op_8777")]; + tensor var_8778_to_fp16 = const()[name = tensor("op_8778_to_fp16"), val = tensor(-inf)]; + tensor attn_weights_123_cast_fp16 = select(a = var_8778_to_fp16, b = attn_weights_121_cast_fp16, cond = var_8777)[name = tensor("attn_weights_123_cast_fp16")]; + tensor var_8780 = const()[name = tensor("op_8780"), val = tensor(-1)]; + tensor attn_weights_125_cast_fp16 = softmax(axis = var_8780, x = attn_weights_123_cast_fp16)[name = tensor("attn_weights_125_cast_fp16")]; + tensor attn_output_41_transpose_x_0 = const()[name = tensor("attn_output_41_transpose_x_0"), val = tensor(false)]; + tensor attn_output_41_transpose_y_0 = const()[name = tensor("attn_output_41_transpose_y_0"), val = tensor(false)]; + tensor values_125_cast_fp16 = transpose(perm = var_8716, x = values_123_cast_fp16)[name = tensor("transpose_126")]; + tensor attn_output_41_cast_fp16 = matmul(transpose_x = attn_output_41_transpose_x_0, transpose_y = attn_output_41_transpose_y_0, x = attn_weights_125_cast_fp16, y = values_125_cast_fp16)[name = tensor("attn_output_41_cast_fp16")]; + tensor var_8788 = const()[name = tensor("op_8788"), val = tensor([0, 2, 1, 3])]; + tensor var_8793 = const()[name = tensor("op_8793"), val = tensor([1, 256, 1024])]; + tensor var_8789_cast_fp16 = transpose(perm = var_8788, x = attn_output_41_cast_fp16)[name = tensor("transpose_123")]; + tensor input_245_cast_fp16 = reshape(shape = var_8793, x = var_8789_cast_fp16)[name = tensor("input_245_cast_fp16")]; + tensor attn20_out_proj_weight_to_fp16 = const()[name = tensor("attn20_out_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(510857088)))]; + tensor linear_81_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = attn20_out_proj_weight_to_fp16, x = input_245_cast_fp16)[name = tensor("linear_81_cast_fp16")]; + tensor var_8802_cast_fp16 = add(x = position20_to_fp16, y = valid_len_to_fp16)[name = tensor("op_8802_cast_fp16")]; + tensor var_8802_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_8802_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor input_247_cast_fp16 = add(x = input_241_cast_fp16, y = linear_81_cast_fp16)[name = tensor("input_247_cast_fp16")]; + tensor input_249_axes_0 = const()[name = tensor("input_249_axes_0"), val = tensor([-1])]; + tensor norm20_2_weight_to_fp16 = const()[name = tensor("norm20_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(512954304)))]; + tensor norm20_2_bias_to_fp16 = const()[name = tensor("norm20_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(512956416)))]; + tensor var_8806_to_fp16 = const()[name = tensor("op_8806_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_249_cast_fp16 = layer_norm(axes = input_249_axes_0, beta = norm20_2_bias_to_fp16, epsilon = var_8806_to_fp16, gamma = norm20_2_weight_to_fp16, x = input_247_cast_fp16)[name = tensor("input_249_cast_fp16")]; + tensor linear20_1_weight_to_fp16 = const()[name = tensor("linear20_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(512958528)))]; + tensor linear_82_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = linear20_1_weight_to_fp16, x = input_249_cast_fp16)[name = tensor("linear_82_cast_fp16")]; + tensor input_251_mode_0 = const()[name = tensor("input_251_mode_0"), val = tensor("EXACT")]; + tensor input_251_cast_fp16 = gelu(mode = input_251_mode_0, x = linear_82_cast_fp16)[name = tensor("input_251_cast_fp16")]; + tensor linear20_2_weight_to_fp16 = const()[name = tensor("linear20_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(521347200)))]; + tensor linear_83_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = linear20_2_weight_to_fp16, x = input_251_cast_fp16)[name = tensor("linear_83_cast_fp16")]; + tensor input_253_cast_fp16 = add(x = input_247_cast_fp16, y = linear_83_cast_fp16)[name = tensor("input_253_cast_fp16")]; + tensor input_255_axes_0 = const()[name = tensor("input_255_axes_0"), val = tensor([-1])]; + tensor norm21_1_weight_to_fp16 = const()[name = tensor("norm21_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(529735872)))]; + tensor norm21_1_bias_to_fp16 = const()[name = tensor("norm21_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(529737984)))]; + tensor var_8823_to_fp16 = const()[name = tensor("op_8823_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_255_cast_fp16 = layer_norm(axes = input_255_axes_0, beta = norm21_1_bias_to_fp16, epsilon = var_8823_to_fp16, gamma = norm21_1_weight_to_fp16, x = input_253_cast_fp16)[name = tensor("input_255_cast_fp16")]; + tensor attn21_in_proj_weight_to_fp16 = const()[name = tensor("attn21_in_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(529740096)))]; + tensor linear_84_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = attn21_in_proj_weight_to_fp16, x = input_255_cast_fp16)[name = tensor("linear_84_cast_fp16")]; + tensor var_8837 = const()[name = tensor("op_8837"), val = tensor([1, 256, 3, 16, 64])]; + tensor qkv_43_cast_fp16 = reshape(shape = var_8837, x = linear_84_cast_fp16)[name = tensor("qkv_43_cast_fp16")]; + tensor q_127_begin_0 = const()[name = tensor("q_127_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor q_127_end_0 = const()[name = tensor("q_127_end_0"), val = tensor([1, 256, 1, 16, 64])]; + tensor q_127_end_mask_0 = const()[name = tensor("q_127_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor q_127_squeeze_mask_0 = const()[name = tensor("q_127_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor q_127_cast_fp16 = slice_by_index(begin = q_127_begin_0, end = q_127_end_0, end_mask = q_127_end_mask_0, squeeze_mask = q_127_squeeze_mask_0, x = qkv_43_cast_fp16)[name = tensor("q_127_cast_fp16")]; + tensor k_85_begin_0 = const()[name = tensor("k_85_begin_0"), val = tensor([0, 0, 1, 0, 0])]; + tensor k_85_end_0 = const()[name = tensor("k_85_end_0"), val = tensor([1, 256, 2, 16, 64])]; + tensor k_85_end_mask_0 = const()[name = tensor("k_85_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor k_85_squeeze_mask_0 = const()[name = tensor("k_85_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor k_85_cast_fp16 = slice_by_index(begin = k_85_begin_0, end = k_85_end_0, end_mask = k_85_end_mask_0, squeeze_mask = k_85_squeeze_mask_0, x = qkv_43_cast_fp16)[name = tensor("k_85_cast_fp16")]; + tensor v_43_begin_0 = const()[name = tensor("v_43_begin_0"), val = tensor([0, 0, 2, 0, 0])]; + tensor v_43_end_0 = const()[name = tensor("v_43_end_0"), val = tensor([1, 256, 3, 16, 64])]; + tensor v_43_end_mask_0 = const()[name = tensor("v_43_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor v_43_squeeze_mask_0 = const()[name = tensor("v_43_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor v_43_cast_fp16 = slice_by_index(begin = v_43_begin_0, end = v_43_end_0, end_mask = v_43_end_mask_0, squeeze_mask = v_43_squeeze_mask_0, x = qkv_43_cast_fp16)[name = tensor("v_43_cast_fp16")]; + tensor var_8895 = const()[name = tensor("op_8895"), val = tensor([1, 1])]; + tensor position21_to_fp16_dtype_0 = const()[name = tensor("position21_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor position21_to_fp16 = cast(dtype = position21_to_fp16_dtype_0, x = position21)[name = tensor("cast_556")]; + tensor var_8896_cast_fp16 = reshape(shape = var_8895, x = position21_to_fp16)[name = tensor("op_8896_cast_fp16")]; + tensor ts_129_cast_fp16 = add(x = ts_1_promoted_to_fp16, y = var_8896_cast_fp16)[name = tensor("ts_129_cast_fp16")]; + tensor var_8903 = const()[name = tensor("op_8903"), val = tensor([1, 256, 1, 1])]; + tensor ts_131_cast_fp16 = reshape(shape = var_8903, x = ts_129_cast_fp16)[name = tensor("ts_131_cast_fp16")]; + tensor var_8910 = const()[name = tensor("op_8910"), val = tensor([1, 256, 16, 32, 2])]; + tensor q_complex_43_cast_fp16 = reshape(shape = var_8910, x = q_127_cast_fp16)[name = tensor("q_complex_43_cast_fp16")]; + tensor var_8917 = const()[name = tensor("op_8917"), val = tensor([1, 256, 16, 32, 2])]; + tensor k_complex_43_cast_fp16 = reshape(shape = var_8917, x = k_85_cast_fp16)[name = tensor("k_complex_43_cast_fp16")]; + tensor var_8921_begin_0 = const()[name = tensor("op_8921_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_8921_end_0 = const()[name = tensor("op_8921_end_0"), val = tensor([1, 256, 16, 32, 1])]; + tensor var_8921_end_mask_0 = const()[name = tensor("op_8921_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_8921_squeeze_mask_0 = const()[name = tensor("op_8921_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_8921_cast_fp16 = slice_by_index(begin = var_8921_begin_0, end = var_8921_end_0, end_mask = var_8921_end_mask_0, squeeze_mask = var_8921_squeeze_mask_0, x = q_complex_43_cast_fp16)[name = tensor("op_8921_cast_fp16")]; + tensor var_8929_begin_0 = const()[name = tensor("op_8929_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_8929_end_0 = const()[name = tensor("op_8929_end_0"), val = tensor([1, 256, 16, 32, 2])]; + tensor var_8929_end_mask_0 = const()[name = tensor("op_8929_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_8929_squeeze_mask_0 = const()[name = tensor("op_8929_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_8929_cast_fp16 = slice_by_index(begin = var_8929_begin_0, end = var_8929_end_0, end_mask = var_8929_end_mask_0, squeeze_mask = var_8929_squeeze_mask_0, x = q_complex_43_cast_fp16)[name = tensor("op_8929_cast_fp16")]; + tensor var_8937_begin_0 = const()[name = tensor("op_8937_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_8937_end_0 = const()[name = tensor("op_8937_end_0"), val = tensor([1, 256, 16, 32, 1])]; + tensor var_8937_end_mask_0 = const()[name = tensor("op_8937_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_8937_squeeze_mask_0 = const()[name = tensor("op_8937_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_8937_cast_fp16 = slice_by_index(begin = var_8937_begin_0, end = var_8937_end_0, end_mask = var_8937_end_mask_0, squeeze_mask = var_8937_squeeze_mask_0, x = k_complex_43_cast_fp16)[name = tensor("op_8937_cast_fp16")]; + tensor var_8945_begin_0 = const()[name = tensor("op_8945_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_8945_end_0 = const()[name = tensor("op_8945_end_0"), val = tensor([1, 256, 16, 32, 2])]; + tensor var_8945_end_mask_0 = const()[name = tensor("op_8945_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_8945_squeeze_mask_0 = const()[name = tensor("op_8945_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_8945_cast_fp16 = slice_by_index(begin = var_8945_begin_0, end = var_8945_end_0, end_mask = var_8945_end_mask_0, squeeze_mask = var_8945_squeeze_mask_0, x = k_complex_43_cast_fp16)[name = tensor("op_8945_cast_fp16")]; + tensor freqs_43_to_fp16 = const()[name = tensor("freqs_43_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(536031616)))]; + tensor var_8951_cast_fp16 = mul(x = freqs_43_to_fp16, y = ts_131_cast_fp16)[name = tensor("op_8951_cast_fp16")]; + tensor rotr_43_cast_fp16 = cos(x = var_8951_cast_fp16)[name = tensor("rotr_43_cast_fp16")]; + tensor roti_43_cast_fp16 = sin(x = var_8951_cast_fp16)[name = tensor("roti_43_cast_fp16")]; + tensor var_8955_cast_fp16 = mul(x = var_8921_cast_fp16, y = rotr_43_cast_fp16)[name = tensor("op_8955_cast_fp16")]; + tensor var_8956_cast_fp16 = mul(x = var_8929_cast_fp16, y = roti_43_cast_fp16)[name = tensor("op_8956_cast_fp16")]; + tensor qor_85_cast_fp16 = sub(x = var_8955_cast_fp16, y = var_8956_cast_fp16)[name = tensor("qor_85_cast_fp16")]; + tensor var_8959_cast_fp16 = mul(x = var_8921_cast_fp16, y = roti_43_cast_fp16)[name = tensor("op_8959_cast_fp16")]; + tensor var_8960_cast_fp16 = mul(x = var_8929_cast_fp16, y = rotr_43_cast_fp16)[name = tensor("op_8960_cast_fp16")]; + tensor qoi_85_cast_fp16 = add(x = var_8959_cast_fp16, y = var_8960_cast_fp16)[name = tensor("qoi_85_cast_fp16")]; + tensor var_8963_cast_fp16 = mul(x = var_8937_cast_fp16, y = rotr_43_cast_fp16)[name = tensor("op_8963_cast_fp16")]; + tensor var_8964_cast_fp16 = mul(x = var_8945_cast_fp16, y = roti_43_cast_fp16)[name = tensor("op_8964_cast_fp16")]; + tensor kor_85_cast_fp16 = sub(x = var_8963_cast_fp16, y = var_8964_cast_fp16)[name = tensor("kor_85_cast_fp16")]; + tensor var_8967_cast_fp16 = mul(x = var_8937_cast_fp16, y = roti_43_cast_fp16)[name = tensor("op_8967_cast_fp16")]; + tensor var_8968_cast_fp16 = mul(x = var_8945_cast_fp16, y = rotr_43_cast_fp16)[name = tensor("op_8968_cast_fp16")]; + tensor koi_85_cast_fp16 = add(x = var_8967_cast_fp16, y = var_8968_cast_fp16)[name = tensor("koi_85_cast_fp16")]; + tensor qo_43_axis_0 = const()[name = tensor("qo_43_axis_0"), val = tensor(-1)]; + tensor qo_43_cast_fp16 = stack(axis = qo_43_axis_0, values = (qor_85_cast_fp16, qoi_85_cast_fp16))[name = tensor("qo_43_cast_fp16")]; + tensor ko_43_axis_0 = const()[name = tensor("ko_43_axis_0"), val = tensor(-1)]; + tensor ko_43_cast_fp16 = stack(axis = ko_43_axis_0, values = (kor_85_cast_fp16, koi_85_cast_fp16))[name = tensor("ko_43_cast_fp16")]; + tensor var_9001 = const()[name = tensor("op_9001"), val = tensor([1, 256, 16, 64])]; + tensor q_129_cast_fp16 = reshape(shape = var_9001, x = qo_43_cast_fp16)[name = tensor("q_129_cast_fp16")]; + tensor var_9007 = const()[name = tensor("op_9007"), val = tensor([1, 256, 16, 64])]; + tensor k_87_cast_fp16 = reshape(shape = var_9007, x = ko_43_cast_fp16)[name = tensor("k_87_cast_fp16")]; + tensor write_indices_float_127_cast_fp16 = add(x = var_8896_cast_fp16, y = var_393_promoted_to_fp16)[name = tensor("write_indices_float_127_cast_fp16")]; + tensor _inversed_9028_y_0_to_fp16 = const()[name = tensor("_inversed_9028_y_0_to_fp16"), val = tensor(0x1p-9)]; + tensor _inversed_9028_cast_fp16 = mul(x = write_indices_float_127_cast_fp16, y = _inversed_9028_y_0_to_fp16)[name = tensor("_inversed_9028_cast_fp16")]; + tensor var_9029_cast_fp16 = floor(x = _inversed_9028_cast_fp16)[name = tensor("op_9029_cast_fp16")]; + tensor var_9030_to_fp16 = const()[name = tensor("op_9030_to_fp16"), val = tensor(0x1p+9)]; + tensor var_9031_cast_fp16 = mul(x = var_9029_cast_fp16, y = var_9030_to_fp16)[name = tensor("op_9031_cast_fp16")]; + tensor write_indices_float_129_cast_fp16 = sub(x = write_indices_float_127_cast_fp16, y = var_9031_cast_fp16)[name = tensor("write_indices_float_129_cast_fp16")]; + tensor var_9049_value_0_to_fp16 = const()[name = tensor("op_9049_value_0_to_fp16"), val = tensor(0x1.ffp+8)]; + tensor var_9049_cast_fp16 = fill_like(ref_tensor = write_indices_float_129_cast_fp16, value = var_9049_value_0_to_fp16)[name = tensor("op_9049_cast_fp16")]; + tensor write_indices_float_131_cast_fp16 = select(a = write_indices_float_129_cast_fp16, b = var_9049_cast_fp16, cond = token_valid_1_cast_fp16)[name = tensor("write_indices_float_131_cast_fp16")]; + tensor var_9055_dtype_0 = const()[name = tensor("op_9055_dtype_0"), val = tensor("int32")]; + tensor var_9060 = const()[name = tensor("op_9060"), val = tensor([1, 256, 1, 1])]; + tensor write_indices_float_131_cast_fp16_to_int32 = cast(dtype = var_9055_dtype_0, x = write_indices_float_131_cast_fp16)[name = tensor("cast_555")]; + tensor var_9061 = reshape(shape = var_9060, x = write_indices_float_131_cast_fp16_to_int32)[name = tensor("op_9061")]; + tensor write_indices_43_reps_0 = const()[name = tensor("write_indices_43_reps_0"), val = tensor([1, 1, 16, 64])]; + tensor write_indices_43 = tile(reps = write_indices_43_reps_0, x = var_9061)[name = tensor("write_indices_43")]; + tensor var_9071_begin_0 = const()[name = tensor("op_9071_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_9071_end_0 = const()[name = tensor("op_9071_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor var_9071_end_mask_0 = const()[name = tensor("op_9071_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_9071_squeeze_mask_0 = const()[name = tensor("op_9071_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor cache21_to_fp16_dtype_0 = const()[name = tensor("cache21_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor cache21_to_fp16 = cast(dtype = cache21_to_fp16_dtype_0, x = cache21)[name = tensor("cast_554")]; + tensor var_9071_cast_fp16 = slice_by_index(begin = var_9071_begin_0, end = var_9071_end_0, end_mask = var_9071_end_mask_0, squeeze_mask = var_9071_squeeze_mask_0, x = cache21_to_fp16)[name = tensor("op_9071_cast_fp16")]; + tensor var_9073_axis_0 = const()[name = tensor("op_9073_axis_0"), val = tensor(1)]; + tensor var_9073_mode_0 = const()[name = tensor("op_9073_mode_0"), val = tensor("update")]; + tensor var_9073_validate_indices_0 = const()[name = tensor("op_9073_validate_indices_0"), val = tensor(false)]; + tensor var_9073_cast_fp16 = scatter_along_axis(axis = var_9073_axis_0, data = var_9071_cast_fp16, indices = write_indices_43, mode = var_9073_mode_0, updates = k_87_cast_fp16, validate_indices = var_9073_validate_indices_0)[name = tensor("op_9073_cast_fp16")]; + tensor concat_170 = const()[name = tensor("concat_170"), val = tensor([0, 0, 0, 0, 0])]; + tensor concat_171 = const()[name = tensor("concat_171"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_43_internal_tensor_assign_1_stride_0 = const()[name = tensor("new_cache_43_internal_tensor_assign_1_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_43_internal_tensor_assign_1_begin_mask_0 = const()[name = tensor("new_cache_43_internal_tensor_assign_1_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_43_internal_tensor_assign_1_end_mask_0 = const()[name = tensor("new_cache_43_internal_tensor_assign_1_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_43_internal_tensor_assign_1_squeeze_mask_0 = const()[name = tensor("new_cache_43_internal_tensor_assign_1_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_88 = const()[name = tensor("shape_88"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_42 = const()[name = tensor("reduce_prod_42"), val = tensor(1048576)]; + tensor range_1d_42_start_0 = const()[name = tensor("range_1d_42_start_0"), val = tensor(0)]; + tensor range_1d_42_step_0 = const()[name = tensor("range_1d_42_step_0"), val = tensor(1)]; + tensor range_1d_42 = range_1d(end = reduce_prod_42, start = range_1d_42_start_0, step = range_1d_42_step_0)[name = tensor("range_1d_42")]; + tensor reshape_210 = reshape(shape = shape_88, x = range_1d_42)[name = tensor("reshape_210")]; + tensor slice_by_index_42 = slice_by_index(begin = concat_170, begin_mask = new_cache_43_internal_tensor_assign_1_begin_mask_0, end = concat_171, end_mask = new_cache_43_internal_tensor_assign_1_end_mask_0, squeeze_mask = new_cache_43_internal_tensor_assign_1_squeeze_mask_0, stride = new_cache_43_internal_tensor_assign_1_stride_0, x = reshape_210)[name = tensor("slice_by_index_42")]; + tensor reshape_211_shape_0 = const()[name = tensor("reshape_211_shape_0"), val = tensor([-1])]; + tensor reshape_211 = reshape(shape = reshape_211_shape_0, x = slice_by_index_42)[name = tensor("reshape_211")]; + tensor reshape_212_shape_0 = const()[name = tensor("reshape_212_shape_0"), val = tensor([-1])]; + tensor reshape_212_cast_fp16 = reshape(shape = reshape_212_shape_0, x = var_9073_cast_fp16)[name = tensor("reshape_212_cast_fp16")]; + tensor reshape_213_shape_0 = const()[name = tensor("reshape_213_shape_0"), val = tensor([-1])]; + tensor reshape_213_cast_fp16 = reshape(shape = reshape_213_shape_0, x = cache21_to_fp16)[name = tensor("reshape_213_cast_fp16")]; + tensor scatter_42_mode_0 = const()[name = tensor("scatter_42_mode_0"), val = tensor("update")]; + tensor scatter_42_axis_0 = const()[name = tensor("scatter_42_axis_0"), val = tensor(0)]; + tensor scatter_42_validate_indices_0 = const()[name = tensor("scatter_42_validate_indices_0"), val = tensor(false)]; + tensor scatter_42_cast_fp16 = scatter(axis = scatter_42_axis_0, data = reshape_213_cast_fp16, indices = reshape_211, mode = scatter_42_mode_0, updates = reshape_212_cast_fp16, validate_indices = scatter_42_validate_indices_0)[name = tensor("scatter_42_cast_fp16")]; + tensor reshape_214_cast_fp16 = reshape(shape = shape_88, x = scatter_42_cast_fp16)[name = tensor("reshape_214_cast_fp16")]; + tensor var_9081_begin_0 = const()[name = tensor("op_9081_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor var_9081_end_0 = const()[name = tensor("op_9081_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor var_9081_end_mask_0 = const()[name = tensor("op_9081_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_9081_squeeze_mask_0 = const()[name = tensor("op_9081_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor var_9081_cast_fp16 = slice_by_index(begin = var_9081_begin_0, end = var_9081_end_0, end_mask = var_9081_end_mask_0, squeeze_mask = var_9081_squeeze_mask_0, x = reshape_214_cast_fp16)[name = tensor("op_9081_cast_fp16")]; + tensor var_9083_axis_0 = const()[name = tensor("op_9083_axis_0"), val = tensor(1)]; + tensor var_9083_mode_0 = const()[name = tensor("op_9083_mode_0"), val = tensor("update")]; + tensor var_9083_validate_indices_0 = const()[name = tensor("op_9083_validate_indices_0"), val = tensor(false)]; + tensor var_9083_cast_fp16 = scatter_along_axis(axis = var_9083_axis_0, data = var_9081_cast_fp16, indices = write_indices_43, mode = var_9083_mode_0, updates = v_43_cast_fp16, validate_indices = var_9083_validate_indices_0)[name = tensor("op_9083_cast_fp16")]; + tensor concat_172 = const()[name = tensor("concat_172"), val = tensor([1, 0, 0, 0, 0])]; + tensor concat_173 = const()[name = tensor("concat_173"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_43_internal_tensor_assign_2_stride_0 = const()[name = tensor("new_cache_43_internal_tensor_assign_2_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_43_internal_tensor_assign_2_begin_mask_0 = const()[name = tensor("new_cache_43_internal_tensor_assign_2_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_43_internal_tensor_assign_2_end_mask_0 = const()[name = tensor("new_cache_43_internal_tensor_assign_2_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_43_internal_tensor_assign_2_squeeze_mask_0 = const()[name = tensor("new_cache_43_internal_tensor_assign_2_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_89 = const()[name = tensor("shape_89"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_43 = const()[name = tensor("reduce_prod_43"), val = tensor(1048576)]; + tensor range_1d_43_start_0 = const()[name = tensor("range_1d_43_start_0"), val = tensor(0)]; + tensor range_1d_43_step_0 = const()[name = tensor("range_1d_43_step_0"), val = tensor(1)]; + tensor range_1d_43 = range_1d(end = reduce_prod_43, start = range_1d_43_start_0, step = range_1d_43_step_0)[name = tensor("range_1d_43")]; + tensor reshape_215 = reshape(shape = shape_89, x = range_1d_43)[name = tensor("reshape_215")]; + tensor slice_by_index_43 = slice_by_index(begin = concat_172, begin_mask = new_cache_43_internal_tensor_assign_2_begin_mask_0, end = concat_173, end_mask = new_cache_43_internal_tensor_assign_2_end_mask_0, squeeze_mask = new_cache_43_internal_tensor_assign_2_squeeze_mask_0, stride = new_cache_43_internal_tensor_assign_2_stride_0, x = reshape_215)[name = tensor("slice_by_index_43")]; + tensor reshape_216_shape_0 = const()[name = tensor("reshape_216_shape_0"), val = tensor([-1])]; + tensor reshape_216 = reshape(shape = reshape_216_shape_0, x = slice_by_index_43)[name = tensor("reshape_216")]; + tensor reshape_217_shape_0 = const()[name = tensor("reshape_217_shape_0"), val = tensor([-1])]; + tensor reshape_217_cast_fp16 = reshape(shape = reshape_217_shape_0, x = var_9083_cast_fp16)[name = tensor("reshape_217_cast_fp16")]; + tensor reshape_218_shape_0 = const()[name = tensor("reshape_218_shape_0"), val = tensor([-1])]; + tensor reshape_218_cast_fp16 = reshape(shape = reshape_218_shape_0, x = reshape_214_cast_fp16)[name = tensor("reshape_218_cast_fp16")]; + tensor scatter_43_mode_0 = const()[name = tensor("scatter_43_mode_0"), val = tensor("update")]; + tensor scatter_43_axis_0 = const()[name = tensor("scatter_43_axis_0"), val = tensor(0)]; + tensor scatter_43_validate_indices_0 = const()[name = tensor("scatter_43_validate_indices_0"), val = tensor(false)]; + tensor scatter_43_cast_fp16 = scatter(axis = scatter_43_axis_0, data = reshape_218_cast_fp16, indices = reshape_216, mode = scatter_43_mode_0, updates = reshape_217_cast_fp16, validate_indices = scatter_43_validate_indices_0)[name = tensor("scatter_43_cast_fp16")]; + tensor reshape_219_cast_fp16 = reshape(shape = shape_89, x = scatter_43_cast_fp16)[name = tensor("reshape_219_cast_fp16")]; + tensor reshape_219_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("reshape_219_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor keys_127_begin_0 = const()[name = tensor("keys_127_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor keys_127_end_0 = const()[name = tensor("keys_127_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor keys_127_end_mask_0 = const()[name = tensor("keys_127_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor keys_127_squeeze_mask_0 = const()[name = tensor("keys_127_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor keys_127_cast_fp16 = slice_by_index(begin = keys_127_begin_0, end = keys_127_end_0, end_mask = keys_127_end_mask_0, squeeze_mask = keys_127_squeeze_mask_0, x = reshape_219_cast_fp16)[name = tensor("keys_127_cast_fp16")]; + tensor values_127_begin_0 = const()[name = tensor("values_127_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor values_127_end_0 = const()[name = tensor("values_127_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor values_127_end_mask_0 = const()[name = tensor("values_127_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor values_127_squeeze_mask_0 = const()[name = tensor("values_127_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor values_127_cast_fp16 = slice_by_index(begin = values_127_begin_0, end = values_127_end_0, end_mask = values_127_end_mask_0, squeeze_mask = values_127_squeeze_mask_0, x = reshape_219_cast_fp16)[name = tensor("values_127_cast_fp16")]; + tensor var_9095_cast_fp16 = not_equal(x = keys_127_cast_fp16, y = keys_127_cast_fp16)[name = tensor("op_9095_cast_fp16")]; + tensor keys_129_cast_fp16 = select(a = var_470_to_fp16, b = keys_127_cast_fp16, cond = var_9095_cast_fp16)[name = tensor("keys_129_cast_fp16")]; + tensor var_9103_cast_fp16 = not_equal(x = values_127_cast_fp16, y = values_127_cast_fp16)[name = tensor("op_9103_cast_fp16")]; + tensor values_129_cast_fp16 = select(a = var_470_to_fp16, b = values_127_cast_fp16, cond = var_9103_cast_fp16)[name = tensor("values_129_cast_fp16")]; + tensor var_9127 = const()[name = tensor("op_9127"), val = tensor([0, 2, 1, 3])]; + tensor var_9143 = const()[name = tensor("op_9143"), val = tensor([1, 1, 1])]; + tensor var_9144_cast_fp16 = reshape(shape = var_9143, x = position21_to_fp16)[name = tensor("op_9144_cast_fp16")]; + tensor q_positions_43_cast_fp16 = add(x = var_9144_cast_fp16, y = q_offsets_1_promoted_to_fp16)[name = tensor("q_positions_43_cast_fp16")]; + tensor valid_end_43_cast_fp16 = add(x = var_9144_cast_fp16, y = var_536_cast_fp16)[name = tensor("valid_end_43_cast_fp16")]; + tensor valid_mask_43_cast_fp16 = less(x = k_positions_1_promoted_to_fp16, y = valid_end_43_cast_fp16)[name = tensor("valid_mask_43_cast_fp16")]; + tensor causal_mask_43_cast_fp16 = less_equal(x = k_positions_1_promoted_to_fp16, y = q_positions_43_cast_fp16)[name = tensor("causal_mask_43_cast_fp16")]; + tensor var_9172 = logical_and(x = valid_mask_43_cast_fp16, y = causal_mask_43_cast_fp16)[name = tensor("op_9172")]; + tensor attn_mask_43_axes_0 = const()[name = tensor("attn_mask_43_axes_0"), val = tensor([1])]; + tensor attn_mask_43 = expand_dims(axes = attn_mask_43_axes_0, x = var_9172)[name = tensor("attn_mask_43")]; + tensor var_9186_transpose_x_0 = const()[name = tensor("op_9186_transpose_x_0"), val = tensor(false)]; + tensor var_9186_transpose_y_0 = const()[name = tensor("op_9186_transpose_y_0"), val = tensor(false)]; + tensor transpose_111_perm_0 = const()[name = tensor("transpose_111_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_112_perm_0 = const()[name = tensor("transpose_112_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_112 = transpose(perm = transpose_112_perm_0, x = keys_129_cast_fp16)[name = tensor("transpose_120")]; + tensor transpose_111 = transpose(perm = transpose_111_perm_0, x = q_129_cast_fp16)[name = tensor("transpose_121")]; + tensor var_9186_cast_fp16 = matmul(transpose_x = var_9186_transpose_x_0, transpose_y = var_9186_transpose_y_0, x = transpose_111, y = transpose_112)[name = tensor("op_9186_cast_fp16")]; + tensor var_9180_to_fp16 = const()[name = tensor("op_9180_to_fp16"), val = tensor([0x1p-3])]; + tensor attn_weights_127_cast_fp16 = mul(x = var_9186_cast_fp16, y = var_9180_to_fp16)[name = tensor("attn_weights_127_cast_fp16")]; + tensor var_9188 = logical_not(x = attn_mask_43)[name = tensor("op_9188")]; + tensor var_9189_to_fp16 = const()[name = tensor("op_9189_to_fp16"), val = tensor(-inf)]; + tensor attn_weights_129_cast_fp16 = select(a = var_9189_to_fp16, b = attn_weights_127_cast_fp16, cond = var_9188)[name = tensor("attn_weights_129_cast_fp16")]; + tensor var_9191 = const()[name = tensor("op_9191"), val = tensor(-1)]; + tensor attn_weights_131_cast_fp16 = softmax(axis = var_9191, x = attn_weights_129_cast_fp16)[name = tensor("attn_weights_131_cast_fp16")]; + tensor attn_output_43_transpose_x_0 = const()[name = tensor("attn_output_43_transpose_x_0"), val = tensor(false)]; + tensor attn_output_43_transpose_y_0 = const()[name = tensor("attn_output_43_transpose_y_0"), val = tensor(false)]; + tensor values_131_cast_fp16 = transpose(perm = var_9127, x = values_129_cast_fp16)[name = tensor("transpose_122")]; + tensor attn_output_43_cast_fp16 = matmul(transpose_x = attn_output_43_transpose_x_0, transpose_y = attn_output_43_transpose_y_0, x = attn_weights_131_cast_fp16, y = values_131_cast_fp16)[name = tensor("attn_output_43_cast_fp16")]; + tensor var_9199 = const()[name = tensor("op_9199"), val = tensor([0, 2, 1, 3])]; + tensor var_9204 = const()[name = tensor("op_9204"), val = tensor([1, 256, 1024])]; + tensor var_9200_cast_fp16 = transpose(perm = var_9199, x = attn_output_43_cast_fp16)[name = tensor("transpose_119")]; + tensor input_257_cast_fp16 = reshape(shape = var_9204, x = var_9200_cast_fp16)[name = tensor("input_257_cast_fp16")]; + tensor attn21_out_proj_weight_to_fp16 = const()[name = tensor("attn21_out_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(536031744)))]; + tensor linear_85_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = attn21_out_proj_weight_to_fp16, x = input_257_cast_fp16)[name = tensor("linear_85_cast_fp16")]; + tensor var_9213_cast_fp16 = add(x = position21_to_fp16, y = valid_len_to_fp16)[name = tensor("op_9213_cast_fp16")]; + tensor var_9213_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_9213_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor input_259_cast_fp16 = add(x = input_253_cast_fp16, y = linear_85_cast_fp16)[name = tensor("input_259_cast_fp16")]; + tensor input_261_axes_0 = const()[name = tensor("input_261_axes_0"), val = tensor([-1])]; + tensor norm21_2_weight_to_fp16 = const()[name = tensor("norm21_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(538128960)))]; + tensor norm21_2_bias_to_fp16 = const()[name = tensor("norm21_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(538131072)))]; + tensor var_9217_to_fp16 = const()[name = tensor("op_9217_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_261_cast_fp16 = layer_norm(axes = input_261_axes_0, beta = norm21_2_bias_to_fp16, epsilon = var_9217_to_fp16, gamma = norm21_2_weight_to_fp16, x = input_259_cast_fp16)[name = tensor("input_261_cast_fp16")]; + tensor linear21_1_weight_to_fp16 = const()[name = tensor("linear21_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(538133184)))]; + tensor linear_86_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = linear21_1_weight_to_fp16, x = input_261_cast_fp16)[name = tensor("linear_86_cast_fp16")]; + tensor input_263_mode_0 = const()[name = tensor("input_263_mode_0"), val = tensor("EXACT")]; + tensor input_263_cast_fp16 = gelu(mode = input_263_mode_0, x = linear_86_cast_fp16)[name = tensor("input_263_cast_fp16")]; + tensor linear21_2_weight_to_fp16 = const()[name = tensor("linear21_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(546521856)))]; + tensor linear_87_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = linear21_2_weight_to_fp16, x = input_263_cast_fp16)[name = tensor("linear_87_cast_fp16")]; + tensor input_265_cast_fp16 = add(x = input_259_cast_fp16, y = linear_87_cast_fp16)[name = tensor("input_265_cast_fp16")]; + tensor input_267_axes_0 = const()[name = tensor("input_267_axes_0"), val = tensor([-1])]; + tensor norm22_1_weight_to_fp16 = const()[name = tensor("norm22_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(554910528)))]; + tensor norm22_1_bias_to_fp16 = const()[name = tensor("norm22_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(554912640)))]; + tensor var_9234_to_fp16 = const()[name = tensor("op_9234_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_267_cast_fp16 = layer_norm(axes = input_267_axes_0, beta = norm22_1_bias_to_fp16, epsilon = var_9234_to_fp16, gamma = norm22_1_weight_to_fp16, x = input_265_cast_fp16)[name = tensor("input_267_cast_fp16")]; + tensor attn22_in_proj_weight_to_fp16 = const()[name = tensor("attn22_in_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(554914752)))]; + tensor linear_88_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = attn22_in_proj_weight_to_fp16, x = input_267_cast_fp16)[name = tensor("linear_88_cast_fp16")]; + tensor var_9248 = const()[name = tensor("op_9248"), val = tensor([1, 256, 3, 16, 64])]; + tensor qkv_45_cast_fp16 = reshape(shape = var_9248, x = linear_88_cast_fp16)[name = tensor("qkv_45_cast_fp16")]; + tensor q_133_begin_0 = const()[name = tensor("q_133_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor q_133_end_0 = const()[name = tensor("q_133_end_0"), val = tensor([1, 256, 1, 16, 64])]; + tensor q_133_end_mask_0 = const()[name = tensor("q_133_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor q_133_squeeze_mask_0 = const()[name = tensor("q_133_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor q_133_cast_fp16 = slice_by_index(begin = q_133_begin_0, end = q_133_end_0, end_mask = q_133_end_mask_0, squeeze_mask = q_133_squeeze_mask_0, x = qkv_45_cast_fp16)[name = tensor("q_133_cast_fp16")]; + tensor k_89_begin_0 = const()[name = tensor("k_89_begin_0"), val = tensor([0, 0, 1, 0, 0])]; + tensor k_89_end_0 = const()[name = tensor("k_89_end_0"), val = tensor([1, 256, 2, 16, 64])]; + tensor k_89_end_mask_0 = const()[name = tensor("k_89_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor k_89_squeeze_mask_0 = const()[name = tensor("k_89_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor k_89_cast_fp16 = slice_by_index(begin = k_89_begin_0, end = k_89_end_0, end_mask = k_89_end_mask_0, squeeze_mask = k_89_squeeze_mask_0, x = qkv_45_cast_fp16)[name = tensor("k_89_cast_fp16")]; + tensor v_45_begin_0 = const()[name = tensor("v_45_begin_0"), val = tensor([0, 0, 2, 0, 0])]; + tensor v_45_end_0 = const()[name = tensor("v_45_end_0"), val = tensor([1, 256, 3, 16, 64])]; + tensor v_45_end_mask_0 = const()[name = tensor("v_45_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor v_45_squeeze_mask_0 = const()[name = tensor("v_45_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor v_45_cast_fp16 = slice_by_index(begin = v_45_begin_0, end = v_45_end_0, end_mask = v_45_end_mask_0, squeeze_mask = v_45_squeeze_mask_0, x = qkv_45_cast_fp16)[name = tensor("v_45_cast_fp16")]; + tensor var_9306 = const()[name = tensor("op_9306"), val = tensor([1, 1])]; + tensor position22_to_fp16_dtype_0 = const()[name = tensor("position22_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor position22_to_fp16 = cast(dtype = position22_to_fp16_dtype_0, x = position22)[name = tensor("cast_551")]; + tensor var_9307_cast_fp16 = reshape(shape = var_9306, x = position22_to_fp16)[name = tensor("op_9307_cast_fp16")]; + tensor ts_135_cast_fp16 = add(x = ts_1_promoted_to_fp16, y = var_9307_cast_fp16)[name = tensor("ts_135_cast_fp16")]; + tensor var_9314 = const()[name = tensor("op_9314"), val = tensor([1, 256, 1, 1])]; + tensor ts_137_cast_fp16 = reshape(shape = var_9314, x = ts_135_cast_fp16)[name = tensor("ts_137_cast_fp16")]; + tensor var_9321 = const()[name = tensor("op_9321"), val = tensor([1, 256, 16, 32, 2])]; + tensor q_complex_45_cast_fp16 = reshape(shape = var_9321, x = q_133_cast_fp16)[name = tensor("q_complex_45_cast_fp16")]; + tensor var_9328 = const()[name = tensor("op_9328"), val = tensor([1, 256, 16, 32, 2])]; + tensor k_complex_45_cast_fp16 = reshape(shape = var_9328, x = k_89_cast_fp16)[name = tensor("k_complex_45_cast_fp16")]; + tensor var_9332_begin_0 = const()[name = tensor("op_9332_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_9332_end_0 = const()[name = tensor("op_9332_end_0"), val = tensor([1, 256, 16, 32, 1])]; + tensor var_9332_end_mask_0 = const()[name = tensor("op_9332_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_9332_squeeze_mask_0 = const()[name = tensor("op_9332_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_9332_cast_fp16 = slice_by_index(begin = var_9332_begin_0, end = var_9332_end_0, end_mask = var_9332_end_mask_0, squeeze_mask = var_9332_squeeze_mask_0, x = q_complex_45_cast_fp16)[name = tensor("op_9332_cast_fp16")]; + tensor var_9340_begin_0 = const()[name = tensor("op_9340_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_9340_end_0 = const()[name = tensor("op_9340_end_0"), val = tensor([1, 256, 16, 32, 2])]; + tensor var_9340_end_mask_0 = const()[name = tensor("op_9340_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_9340_squeeze_mask_0 = const()[name = tensor("op_9340_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_9340_cast_fp16 = slice_by_index(begin = var_9340_begin_0, end = var_9340_end_0, end_mask = var_9340_end_mask_0, squeeze_mask = var_9340_squeeze_mask_0, x = q_complex_45_cast_fp16)[name = tensor("op_9340_cast_fp16")]; + tensor var_9348_begin_0 = const()[name = tensor("op_9348_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_9348_end_0 = const()[name = tensor("op_9348_end_0"), val = tensor([1, 256, 16, 32, 1])]; + tensor var_9348_end_mask_0 = const()[name = tensor("op_9348_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_9348_squeeze_mask_0 = const()[name = tensor("op_9348_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_9348_cast_fp16 = slice_by_index(begin = var_9348_begin_0, end = var_9348_end_0, end_mask = var_9348_end_mask_0, squeeze_mask = var_9348_squeeze_mask_0, x = k_complex_45_cast_fp16)[name = tensor("op_9348_cast_fp16")]; + tensor var_9356_begin_0 = const()[name = tensor("op_9356_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_9356_end_0 = const()[name = tensor("op_9356_end_0"), val = tensor([1, 256, 16, 32, 2])]; + tensor var_9356_end_mask_0 = const()[name = tensor("op_9356_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_9356_squeeze_mask_0 = const()[name = tensor("op_9356_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_9356_cast_fp16 = slice_by_index(begin = var_9356_begin_0, end = var_9356_end_0, end_mask = var_9356_end_mask_0, squeeze_mask = var_9356_squeeze_mask_0, x = k_complex_45_cast_fp16)[name = tensor("op_9356_cast_fp16")]; + tensor freqs_45_to_fp16 = const()[name = tensor("freqs_45_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(561206272)))]; + tensor var_9362_cast_fp16 = mul(x = freqs_45_to_fp16, y = ts_137_cast_fp16)[name = tensor("op_9362_cast_fp16")]; + tensor rotr_45_cast_fp16 = cos(x = var_9362_cast_fp16)[name = tensor("rotr_45_cast_fp16")]; + tensor roti_45_cast_fp16 = sin(x = var_9362_cast_fp16)[name = tensor("roti_45_cast_fp16")]; + tensor var_9366_cast_fp16 = mul(x = var_9332_cast_fp16, y = rotr_45_cast_fp16)[name = tensor("op_9366_cast_fp16")]; + tensor var_9367_cast_fp16 = mul(x = var_9340_cast_fp16, y = roti_45_cast_fp16)[name = tensor("op_9367_cast_fp16")]; + tensor qor_89_cast_fp16 = sub(x = var_9366_cast_fp16, y = var_9367_cast_fp16)[name = tensor("qor_89_cast_fp16")]; + tensor var_9370_cast_fp16 = mul(x = var_9332_cast_fp16, y = roti_45_cast_fp16)[name = tensor("op_9370_cast_fp16")]; + tensor var_9371_cast_fp16 = mul(x = var_9340_cast_fp16, y = rotr_45_cast_fp16)[name = tensor("op_9371_cast_fp16")]; + tensor qoi_89_cast_fp16 = add(x = var_9370_cast_fp16, y = var_9371_cast_fp16)[name = tensor("qoi_89_cast_fp16")]; + tensor var_9374_cast_fp16 = mul(x = var_9348_cast_fp16, y = rotr_45_cast_fp16)[name = tensor("op_9374_cast_fp16")]; + tensor var_9375_cast_fp16 = mul(x = var_9356_cast_fp16, y = roti_45_cast_fp16)[name = tensor("op_9375_cast_fp16")]; + tensor kor_89_cast_fp16 = sub(x = var_9374_cast_fp16, y = var_9375_cast_fp16)[name = tensor("kor_89_cast_fp16")]; + tensor var_9378_cast_fp16 = mul(x = var_9348_cast_fp16, y = roti_45_cast_fp16)[name = tensor("op_9378_cast_fp16")]; + tensor var_9379_cast_fp16 = mul(x = var_9356_cast_fp16, y = rotr_45_cast_fp16)[name = tensor("op_9379_cast_fp16")]; + tensor koi_89_cast_fp16 = add(x = var_9378_cast_fp16, y = var_9379_cast_fp16)[name = tensor("koi_89_cast_fp16")]; + tensor qo_45_axis_0 = const()[name = tensor("qo_45_axis_0"), val = tensor(-1)]; + tensor qo_45_cast_fp16 = stack(axis = qo_45_axis_0, values = (qor_89_cast_fp16, qoi_89_cast_fp16))[name = tensor("qo_45_cast_fp16")]; + tensor ko_45_axis_0 = const()[name = tensor("ko_45_axis_0"), val = tensor(-1)]; + tensor ko_45_cast_fp16 = stack(axis = ko_45_axis_0, values = (kor_89_cast_fp16, koi_89_cast_fp16))[name = tensor("ko_45_cast_fp16")]; + tensor var_9412 = const()[name = tensor("op_9412"), val = tensor([1, 256, 16, 64])]; + tensor q_135_cast_fp16 = reshape(shape = var_9412, x = qo_45_cast_fp16)[name = tensor("q_135_cast_fp16")]; + tensor var_9418 = const()[name = tensor("op_9418"), val = tensor([1, 256, 16, 64])]; + tensor k_91_cast_fp16 = reshape(shape = var_9418, x = ko_45_cast_fp16)[name = tensor("k_91_cast_fp16")]; + tensor write_indices_float_133_cast_fp16 = add(x = var_9307_cast_fp16, y = var_393_promoted_to_fp16)[name = tensor("write_indices_float_133_cast_fp16")]; + tensor _inversed_9439_y_0_to_fp16 = const()[name = tensor("_inversed_9439_y_0_to_fp16"), val = tensor(0x1p-9)]; + tensor _inversed_9439_cast_fp16 = mul(x = write_indices_float_133_cast_fp16, y = _inversed_9439_y_0_to_fp16)[name = tensor("_inversed_9439_cast_fp16")]; + tensor var_9440_cast_fp16 = floor(x = _inversed_9439_cast_fp16)[name = tensor("op_9440_cast_fp16")]; + tensor var_9441_to_fp16 = const()[name = tensor("op_9441_to_fp16"), val = tensor(0x1p+9)]; + tensor var_9442_cast_fp16 = mul(x = var_9440_cast_fp16, y = var_9441_to_fp16)[name = tensor("op_9442_cast_fp16")]; + tensor write_indices_float_135_cast_fp16 = sub(x = write_indices_float_133_cast_fp16, y = var_9442_cast_fp16)[name = tensor("write_indices_float_135_cast_fp16")]; + tensor var_9460_value_0_to_fp16 = const()[name = tensor("op_9460_value_0_to_fp16"), val = tensor(0x1.ffp+8)]; + tensor var_9460_cast_fp16 = fill_like(ref_tensor = write_indices_float_135_cast_fp16, value = var_9460_value_0_to_fp16)[name = tensor("op_9460_cast_fp16")]; + tensor write_indices_float_137_cast_fp16 = select(a = write_indices_float_135_cast_fp16, b = var_9460_cast_fp16, cond = token_valid_1_cast_fp16)[name = tensor("write_indices_float_137_cast_fp16")]; + tensor var_9466_dtype_0 = const()[name = tensor("op_9466_dtype_0"), val = tensor("int32")]; + tensor var_9471 = const()[name = tensor("op_9471"), val = tensor([1, 256, 1, 1])]; + tensor write_indices_float_137_cast_fp16_to_int32 = cast(dtype = var_9466_dtype_0, x = write_indices_float_137_cast_fp16)[name = tensor("cast_550")]; + tensor var_9472 = reshape(shape = var_9471, x = write_indices_float_137_cast_fp16_to_int32)[name = tensor("op_9472")]; + tensor write_indices_45_reps_0 = const()[name = tensor("write_indices_45_reps_0"), val = tensor([1, 1, 16, 64])]; + tensor write_indices_45 = tile(reps = write_indices_45_reps_0, x = var_9472)[name = tensor("write_indices_45")]; + tensor var_9482_begin_0 = const()[name = tensor("op_9482_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_9482_end_0 = const()[name = tensor("op_9482_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor var_9482_end_mask_0 = const()[name = tensor("op_9482_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_9482_squeeze_mask_0 = const()[name = tensor("op_9482_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor cache22_to_fp16_dtype_0 = const()[name = tensor("cache22_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor cache22_to_fp16 = cast(dtype = cache22_to_fp16_dtype_0, x = cache22)[name = tensor("cast_549")]; + tensor var_9482_cast_fp16 = slice_by_index(begin = var_9482_begin_0, end = var_9482_end_0, end_mask = var_9482_end_mask_0, squeeze_mask = var_9482_squeeze_mask_0, x = cache22_to_fp16)[name = tensor("op_9482_cast_fp16")]; + tensor var_9484_axis_0 = const()[name = tensor("op_9484_axis_0"), val = tensor(1)]; + tensor var_9484_mode_0 = const()[name = tensor("op_9484_mode_0"), val = tensor("update")]; + tensor var_9484_validate_indices_0 = const()[name = tensor("op_9484_validate_indices_0"), val = tensor(false)]; + tensor var_9484_cast_fp16 = scatter_along_axis(axis = var_9484_axis_0, data = var_9482_cast_fp16, indices = write_indices_45, mode = var_9484_mode_0, updates = k_91_cast_fp16, validate_indices = var_9484_validate_indices_0)[name = tensor("op_9484_cast_fp16")]; + tensor concat_178 = const()[name = tensor("concat_178"), val = tensor([0, 0, 0, 0, 0])]; + tensor concat_179 = const()[name = tensor("concat_179"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_45_internal_tensor_assign_1_stride_0 = const()[name = tensor("new_cache_45_internal_tensor_assign_1_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_45_internal_tensor_assign_1_begin_mask_0 = const()[name = tensor("new_cache_45_internal_tensor_assign_1_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_45_internal_tensor_assign_1_end_mask_0 = const()[name = tensor("new_cache_45_internal_tensor_assign_1_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_45_internal_tensor_assign_1_squeeze_mask_0 = const()[name = tensor("new_cache_45_internal_tensor_assign_1_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_90 = const()[name = tensor("shape_90"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_44 = const()[name = tensor("reduce_prod_44"), val = tensor(1048576)]; + tensor range_1d_44_start_0 = const()[name = tensor("range_1d_44_start_0"), val = tensor(0)]; + tensor range_1d_44_step_0 = const()[name = tensor("range_1d_44_step_0"), val = tensor(1)]; + tensor range_1d_44 = range_1d(end = reduce_prod_44, start = range_1d_44_start_0, step = range_1d_44_step_0)[name = tensor("range_1d_44")]; + tensor reshape_220 = reshape(shape = shape_90, x = range_1d_44)[name = tensor("reshape_220")]; + tensor slice_by_index_44 = slice_by_index(begin = concat_178, begin_mask = new_cache_45_internal_tensor_assign_1_begin_mask_0, end = concat_179, end_mask = new_cache_45_internal_tensor_assign_1_end_mask_0, squeeze_mask = new_cache_45_internal_tensor_assign_1_squeeze_mask_0, stride = new_cache_45_internal_tensor_assign_1_stride_0, x = reshape_220)[name = tensor("slice_by_index_44")]; + tensor reshape_221_shape_0 = const()[name = tensor("reshape_221_shape_0"), val = tensor([-1])]; + tensor reshape_221 = reshape(shape = reshape_221_shape_0, x = slice_by_index_44)[name = tensor("reshape_221")]; + tensor reshape_222_shape_0 = const()[name = tensor("reshape_222_shape_0"), val = tensor([-1])]; + tensor reshape_222_cast_fp16 = reshape(shape = reshape_222_shape_0, x = var_9484_cast_fp16)[name = tensor("reshape_222_cast_fp16")]; + tensor reshape_223_shape_0 = const()[name = tensor("reshape_223_shape_0"), val = tensor([-1])]; + tensor reshape_223_cast_fp16 = reshape(shape = reshape_223_shape_0, x = cache22_to_fp16)[name = tensor("reshape_223_cast_fp16")]; + tensor scatter_44_mode_0 = const()[name = tensor("scatter_44_mode_0"), val = tensor("update")]; + tensor scatter_44_axis_0 = const()[name = tensor("scatter_44_axis_0"), val = tensor(0)]; + tensor scatter_44_validate_indices_0 = const()[name = tensor("scatter_44_validate_indices_0"), val = tensor(false)]; + tensor scatter_44_cast_fp16 = scatter(axis = scatter_44_axis_0, data = reshape_223_cast_fp16, indices = reshape_221, mode = scatter_44_mode_0, updates = reshape_222_cast_fp16, validate_indices = scatter_44_validate_indices_0)[name = tensor("scatter_44_cast_fp16")]; + tensor reshape_224_cast_fp16 = reshape(shape = shape_90, x = scatter_44_cast_fp16)[name = tensor("reshape_224_cast_fp16")]; + tensor var_9492_begin_0 = const()[name = tensor("op_9492_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor var_9492_end_0 = const()[name = tensor("op_9492_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor var_9492_end_mask_0 = const()[name = tensor("op_9492_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_9492_squeeze_mask_0 = const()[name = tensor("op_9492_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor var_9492_cast_fp16 = slice_by_index(begin = var_9492_begin_0, end = var_9492_end_0, end_mask = var_9492_end_mask_0, squeeze_mask = var_9492_squeeze_mask_0, x = reshape_224_cast_fp16)[name = tensor("op_9492_cast_fp16")]; + tensor var_9494_axis_0 = const()[name = tensor("op_9494_axis_0"), val = tensor(1)]; + tensor var_9494_mode_0 = const()[name = tensor("op_9494_mode_0"), val = tensor("update")]; + tensor var_9494_validate_indices_0 = const()[name = tensor("op_9494_validate_indices_0"), val = tensor(false)]; + tensor var_9494_cast_fp16 = scatter_along_axis(axis = var_9494_axis_0, data = var_9492_cast_fp16, indices = write_indices_45, mode = var_9494_mode_0, updates = v_45_cast_fp16, validate_indices = var_9494_validate_indices_0)[name = tensor("op_9494_cast_fp16")]; + tensor concat_180 = const()[name = tensor("concat_180"), val = tensor([1, 0, 0, 0, 0])]; + tensor concat_181 = const()[name = tensor("concat_181"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_45_internal_tensor_assign_2_stride_0 = const()[name = tensor("new_cache_45_internal_tensor_assign_2_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_45_internal_tensor_assign_2_begin_mask_0 = const()[name = tensor("new_cache_45_internal_tensor_assign_2_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_45_internal_tensor_assign_2_end_mask_0 = const()[name = tensor("new_cache_45_internal_tensor_assign_2_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_45_internal_tensor_assign_2_squeeze_mask_0 = const()[name = tensor("new_cache_45_internal_tensor_assign_2_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_91 = const()[name = tensor("shape_91"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_45 = const()[name = tensor("reduce_prod_45"), val = tensor(1048576)]; + tensor range_1d_45_start_0 = const()[name = tensor("range_1d_45_start_0"), val = tensor(0)]; + tensor range_1d_45_step_0 = const()[name = tensor("range_1d_45_step_0"), val = tensor(1)]; + tensor range_1d_45 = range_1d(end = reduce_prod_45, start = range_1d_45_start_0, step = range_1d_45_step_0)[name = tensor("range_1d_45")]; + tensor reshape_225 = reshape(shape = shape_91, x = range_1d_45)[name = tensor("reshape_225")]; + tensor slice_by_index_45 = slice_by_index(begin = concat_180, begin_mask = new_cache_45_internal_tensor_assign_2_begin_mask_0, end = concat_181, end_mask = new_cache_45_internal_tensor_assign_2_end_mask_0, squeeze_mask = new_cache_45_internal_tensor_assign_2_squeeze_mask_0, stride = new_cache_45_internal_tensor_assign_2_stride_0, x = reshape_225)[name = tensor("slice_by_index_45")]; + tensor reshape_226_shape_0 = const()[name = tensor("reshape_226_shape_0"), val = tensor([-1])]; + tensor reshape_226 = reshape(shape = reshape_226_shape_0, x = slice_by_index_45)[name = tensor("reshape_226")]; + tensor reshape_227_shape_0 = const()[name = tensor("reshape_227_shape_0"), val = tensor([-1])]; + tensor reshape_227_cast_fp16 = reshape(shape = reshape_227_shape_0, x = var_9494_cast_fp16)[name = tensor("reshape_227_cast_fp16")]; + tensor reshape_228_shape_0 = const()[name = tensor("reshape_228_shape_0"), val = tensor([-1])]; + tensor reshape_228_cast_fp16 = reshape(shape = reshape_228_shape_0, x = reshape_224_cast_fp16)[name = tensor("reshape_228_cast_fp16")]; + tensor scatter_45_mode_0 = const()[name = tensor("scatter_45_mode_0"), val = tensor("update")]; + tensor scatter_45_axis_0 = const()[name = tensor("scatter_45_axis_0"), val = tensor(0)]; + tensor scatter_45_validate_indices_0 = const()[name = tensor("scatter_45_validate_indices_0"), val = tensor(false)]; + tensor scatter_45_cast_fp16 = scatter(axis = scatter_45_axis_0, data = reshape_228_cast_fp16, indices = reshape_226, mode = scatter_45_mode_0, updates = reshape_227_cast_fp16, validate_indices = scatter_45_validate_indices_0)[name = tensor("scatter_45_cast_fp16")]; + tensor reshape_229_cast_fp16 = reshape(shape = shape_91, x = scatter_45_cast_fp16)[name = tensor("reshape_229_cast_fp16")]; + tensor reshape_229_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("reshape_229_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor keys_133_begin_0 = const()[name = tensor("keys_133_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor keys_133_end_0 = const()[name = tensor("keys_133_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor keys_133_end_mask_0 = const()[name = tensor("keys_133_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor keys_133_squeeze_mask_0 = const()[name = tensor("keys_133_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor keys_133_cast_fp16 = slice_by_index(begin = keys_133_begin_0, end = keys_133_end_0, end_mask = keys_133_end_mask_0, squeeze_mask = keys_133_squeeze_mask_0, x = reshape_229_cast_fp16)[name = tensor("keys_133_cast_fp16")]; + tensor values_133_begin_0 = const()[name = tensor("values_133_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor values_133_end_0 = const()[name = tensor("values_133_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor values_133_end_mask_0 = const()[name = tensor("values_133_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor values_133_squeeze_mask_0 = const()[name = tensor("values_133_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor values_133_cast_fp16 = slice_by_index(begin = values_133_begin_0, end = values_133_end_0, end_mask = values_133_end_mask_0, squeeze_mask = values_133_squeeze_mask_0, x = reshape_229_cast_fp16)[name = tensor("values_133_cast_fp16")]; + tensor var_9506_cast_fp16 = not_equal(x = keys_133_cast_fp16, y = keys_133_cast_fp16)[name = tensor("op_9506_cast_fp16")]; + tensor keys_135_cast_fp16 = select(a = var_470_to_fp16, b = keys_133_cast_fp16, cond = var_9506_cast_fp16)[name = tensor("keys_135_cast_fp16")]; + tensor var_9514_cast_fp16 = not_equal(x = values_133_cast_fp16, y = values_133_cast_fp16)[name = tensor("op_9514_cast_fp16")]; + tensor values_135_cast_fp16 = select(a = var_470_to_fp16, b = values_133_cast_fp16, cond = var_9514_cast_fp16)[name = tensor("values_135_cast_fp16")]; + tensor var_9538 = const()[name = tensor("op_9538"), val = tensor([0, 2, 1, 3])]; + tensor var_9554 = const()[name = tensor("op_9554"), val = tensor([1, 1, 1])]; + tensor var_9555_cast_fp16 = reshape(shape = var_9554, x = position22_to_fp16)[name = tensor("op_9555_cast_fp16")]; + tensor q_positions_45_cast_fp16 = add(x = var_9555_cast_fp16, y = q_offsets_1_promoted_to_fp16)[name = tensor("q_positions_45_cast_fp16")]; + tensor valid_end_45_cast_fp16 = add(x = var_9555_cast_fp16, y = var_536_cast_fp16)[name = tensor("valid_end_45_cast_fp16")]; + tensor valid_mask_45_cast_fp16 = less(x = k_positions_1_promoted_to_fp16, y = valid_end_45_cast_fp16)[name = tensor("valid_mask_45_cast_fp16")]; + tensor causal_mask_45_cast_fp16 = less_equal(x = k_positions_1_promoted_to_fp16, y = q_positions_45_cast_fp16)[name = tensor("causal_mask_45_cast_fp16")]; + tensor var_9583 = logical_and(x = valid_mask_45_cast_fp16, y = causal_mask_45_cast_fp16)[name = tensor("op_9583")]; + tensor attn_mask_45_axes_0 = const()[name = tensor("attn_mask_45_axes_0"), val = tensor([1])]; + tensor attn_mask_45 = expand_dims(axes = attn_mask_45_axes_0, x = var_9583)[name = tensor("attn_mask_45")]; + tensor var_9597_transpose_x_0 = const()[name = tensor("op_9597_transpose_x_0"), val = tensor(false)]; + tensor var_9597_transpose_y_0 = const()[name = tensor("op_9597_transpose_y_0"), val = tensor(false)]; + tensor transpose_113_perm_0 = const()[name = tensor("transpose_113_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_114_perm_0 = const()[name = tensor("transpose_114_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_114 = transpose(perm = transpose_114_perm_0, x = keys_135_cast_fp16)[name = tensor("transpose_116")]; + tensor transpose_113 = transpose(perm = transpose_113_perm_0, x = q_135_cast_fp16)[name = tensor("transpose_117")]; + tensor var_9597_cast_fp16 = matmul(transpose_x = var_9597_transpose_x_0, transpose_y = var_9597_transpose_y_0, x = transpose_113, y = transpose_114)[name = tensor("op_9597_cast_fp16")]; + tensor var_9591_to_fp16 = const()[name = tensor("op_9591_to_fp16"), val = tensor([0x1p-3])]; + tensor attn_weights_133_cast_fp16 = mul(x = var_9597_cast_fp16, y = var_9591_to_fp16)[name = tensor("attn_weights_133_cast_fp16")]; + tensor var_9599 = logical_not(x = attn_mask_45)[name = tensor("op_9599")]; + tensor var_9600_to_fp16 = const()[name = tensor("op_9600_to_fp16"), val = tensor(-inf)]; + tensor attn_weights_135_cast_fp16 = select(a = var_9600_to_fp16, b = attn_weights_133_cast_fp16, cond = var_9599)[name = tensor("attn_weights_135_cast_fp16")]; + tensor var_9602 = const()[name = tensor("op_9602"), val = tensor(-1)]; + tensor attn_weights_137_cast_fp16 = softmax(axis = var_9602, x = attn_weights_135_cast_fp16)[name = tensor("attn_weights_137_cast_fp16")]; + tensor attn_output_45_transpose_x_0 = const()[name = tensor("attn_output_45_transpose_x_0"), val = tensor(false)]; + tensor attn_output_45_transpose_y_0 = const()[name = tensor("attn_output_45_transpose_y_0"), val = tensor(false)]; + tensor values_137_cast_fp16 = transpose(perm = var_9538, x = values_135_cast_fp16)[name = tensor("transpose_118")]; + tensor attn_output_45_cast_fp16 = matmul(transpose_x = attn_output_45_transpose_x_0, transpose_y = attn_output_45_transpose_y_0, x = attn_weights_137_cast_fp16, y = values_137_cast_fp16)[name = tensor("attn_output_45_cast_fp16")]; + tensor var_9610 = const()[name = tensor("op_9610"), val = tensor([0, 2, 1, 3])]; + tensor var_9615 = const()[name = tensor("op_9615"), val = tensor([1, 256, 1024])]; + tensor var_9611_cast_fp16 = transpose(perm = var_9610, x = attn_output_45_cast_fp16)[name = tensor("transpose_115")]; + tensor input_269_cast_fp16 = reshape(shape = var_9615, x = var_9611_cast_fp16)[name = tensor("input_269_cast_fp16")]; + tensor attn22_out_proj_weight_to_fp16 = const()[name = tensor("attn22_out_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(561206400)))]; + tensor linear_89_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = attn22_out_proj_weight_to_fp16, x = input_269_cast_fp16)[name = tensor("linear_89_cast_fp16")]; + tensor var_9624_cast_fp16 = add(x = position22_to_fp16, y = valid_len_to_fp16)[name = tensor("op_9624_cast_fp16")]; + tensor var_9624_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_9624_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor input_271_cast_fp16 = add(x = input_265_cast_fp16, y = linear_89_cast_fp16)[name = tensor("input_271_cast_fp16")]; + tensor input_273_axes_0 = const()[name = tensor("input_273_axes_0"), val = tensor([-1])]; + tensor norm22_2_weight_to_fp16 = const()[name = tensor("norm22_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(563303616)))]; + tensor norm22_2_bias_to_fp16 = const()[name = tensor("norm22_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(563305728)))]; + tensor var_9628_to_fp16 = const()[name = tensor("op_9628_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_273_cast_fp16 = layer_norm(axes = input_273_axes_0, beta = norm22_2_bias_to_fp16, epsilon = var_9628_to_fp16, gamma = norm22_2_weight_to_fp16, x = input_271_cast_fp16)[name = tensor("input_273_cast_fp16")]; + tensor linear22_1_weight_to_fp16 = const()[name = tensor("linear22_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(563307840)))]; + tensor linear_90_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = linear22_1_weight_to_fp16, x = input_273_cast_fp16)[name = tensor("linear_90_cast_fp16")]; + tensor input_275_mode_0 = const()[name = tensor("input_275_mode_0"), val = tensor("EXACT")]; + tensor input_275_cast_fp16 = gelu(mode = input_275_mode_0, x = linear_90_cast_fp16)[name = tensor("input_275_cast_fp16")]; + tensor linear22_2_weight_to_fp16 = const()[name = tensor("linear22_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(571696512)))]; + tensor linear_91_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = linear22_2_weight_to_fp16, x = input_275_cast_fp16)[name = tensor("linear_91_cast_fp16")]; + tensor input_277_cast_fp16 = add(x = input_271_cast_fp16, y = linear_91_cast_fp16)[name = tensor("input_277_cast_fp16")]; + tensor input_279_axes_0 = const()[name = tensor("input_279_axes_0"), val = tensor([-1])]; + tensor norm23_1_weight_to_fp16 = const()[name = tensor("norm23_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(580085184)))]; + tensor norm23_1_bias_to_fp16 = const()[name = tensor("norm23_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(580087296)))]; + tensor var_9645_to_fp16 = const()[name = tensor("op_9645_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_279_cast_fp16 = layer_norm(axes = input_279_axes_0, beta = norm23_1_bias_to_fp16, epsilon = var_9645_to_fp16, gamma = norm23_1_weight_to_fp16, x = input_277_cast_fp16)[name = tensor("input_279_cast_fp16")]; + tensor attn23_in_proj_weight_to_fp16 = const()[name = tensor("attn23_in_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(580089408)))]; + tensor linear_92_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = attn23_in_proj_weight_to_fp16, x = input_279_cast_fp16)[name = tensor("linear_92_cast_fp16")]; + tensor var_9659 = const()[name = tensor("op_9659"), val = tensor([1, 256, 3, 16, 64])]; + tensor qkv_cast_fp16 = reshape(shape = var_9659, x = linear_92_cast_fp16)[name = tensor("qkv_cast_fp16")]; + tensor k_93_begin_0 = const()[name = tensor("k_93_begin_0"), val = tensor([0, 0, 1, 0, 0])]; + tensor k_93_end_0 = const()[name = tensor("k_93_end_0"), val = tensor([1, 256, 2, 16, 64])]; + tensor k_93_end_mask_0 = const()[name = tensor("k_93_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor k_93_squeeze_mask_0 = const()[name = tensor("k_93_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor k_93_cast_fp16 = slice_by_index(begin = k_93_begin_0, end = k_93_end_0, end_mask = k_93_end_mask_0, squeeze_mask = k_93_squeeze_mask_0, x = qkv_cast_fp16)[name = tensor("k_93_cast_fp16")]; + tensor v_begin_0 = const()[name = tensor("v_begin_0"), val = tensor([0, 0, 2, 0, 0])]; + tensor v_end_0 = const()[name = tensor("v_end_0"), val = tensor([1, 256, 3, 16, 64])]; + tensor v_end_mask_0 = const()[name = tensor("v_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor v_squeeze_mask_0 = const()[name = tensor("v_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor v_cast_fp16 = slice_by_index(begin = v_begin_0, end = v_end_0, end_mask = v_end_mask_0, squeeze_mask = v_squeeze_mask_0, x = qkv_cast_fp16)[name = tensor("v_cast_fp16")]; + tensor var_9704 = const()[name = tensor("op_9704"), val = tensor([1, 1])]; + tensor position23_to_fp16_dtype_0 = const()[name = tensor("position23_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor position23_to_fp16 = cast(dtype = position23_to_fp16_dtype_0, x = position23)[name = tensor("cast_546")]; + tensor var_9705_cast_fp16 = reshape(shape = var_9704, x = position23_to_fp16)[name = tensor("op_9705_cast_fp16")]; + tensor ts_141_cast_fp16 = add(x = ts_1_promoted_to_fp16, y = var_9705_cast_fp16)[name = tensor("ts_141_cast_fp16")]; + tensor var_9712 = const()[name = tensor("op_9712"), val = tensor([1, 256, 1, 1])]; + tensor ts_cast_fp16 = reshape(shape = var_9712, x = ts_141_cast_fp16)[name = tensor("ts_cast_fp16")]; + tensor var_9719 = const()[name = tensor("op_9719"), val = tensor([1, 256, 16, 32, 2])]; + tensor k_complex_cast_fp16 = reshape(shape = var_9719, x = k_93_cast_fp16)[name = tensor("k_complex_cast_fp16")]; + tensor var_9723_begin_0 = const()[name = tensor("op_9723_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_9723_end_0 = const()[name = tensor("op_9723_end_0"), val = tensor([1, 256, 16, 32, 1])]; + tensor var_9723_end_mask_0 = const()[name = tensor("op_9723_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_9723_squeeze_mask_0 = const()[name = tensor("op_9723_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_9723_cast_fp16 = slice_by_index(begin = var_9723_begin_0, end = var_9723_end_0, end_mask = var_9723_end_mask_0, squeeze_mask = var_9723_squeeze_mask_0, x = k_complex_cast_fp16)[name = tensor("op_9723_cast_fp16")]; + tensor var_9731_begin_0 = const()[name = tensor("op_9731_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_9731_end_0 = const()[name = tensor("op_9731_end_0"), val = tensor([1, 256, 16, 32, 2])]; + tensor var_9731_end_mask_0 = const()[name = tensor("op_9731_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_9731_squeeze_mask_0 = const()[name = tensor("op_9731_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_9731_cast_fp16 = slice_by_index(begin = var_9731_begin_0, end = var_9731_end_0, end_mask = var_9731_end_mask_0, squeeze_mask = var_9731_squeeze_mask_0, x = k_complex_cast_fp16)[name = tensor("op_9731_cast_fp16")]; + tensor freqs_to_fp16 = const()[name = tensor("freqs_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(586380928)))]; + tensor var_9737_cast_fp16 = mul(x = freqs_to_fp16, y = ts_cast_fp16)[name = tensor("op_9737_cast_fp16")]; + tensor rotr_cast_fp16 = cos(x = var_9737_cast_fp16)[name = tensor("rotr_cast_fp16")]; + tensor roti_cast_fp16 = sin(x = var_9737_cast_fp16)[name = tensor("roti_cast_fp16")]; + tensor var_9741_cast_fp16 = mul(x = var_9723_cast_fp16, y = rotr_cast_fp16)[name = tensor("op_9741_cast_fp16")]; + tensor var_9742_cast_fp16 = mul(x = var_9731_cast_fp16, y = roti_cast_fp16)[name = tensor("op_9742_cast_fp16")]; + tensor kor_93_cast_fp16 = sub(x = var_9741_cast_fp16, y = var_9742_cast_fp16)[name = tensor("kor_93_cast_fp16")]; + tensor var_9745_cast_fp16 = mul(x = var_9723_cast_fp16, y = roti_cast_fp16)[name = tensor("op_9745_cast_fp16")]; + tensor var_9746_cast_fp16 = mul(x = var_9731_cast_fp16, y = rotr_cast_fp16)[name = tensor("op_9746_cast_fp16")]; + tensor koi_93_cast_fp16 = add(x = var_9745_cast_fp16, y = var_9746_cast_fp16)[name = tensor("koi_93_cast_fp16")]; + tensor ko_axis_0 = const()[name = tensor("ko_axis_0"), val = tensor(-1)]; + tensor ko_cast_fp16 = stack(axis = ko_axis_0, values = (kor_93_cast_fp16, koi_93_cast_fp16))[name = tensor("ko_cast_fp16")]; + tensor var_9766 = const()[name = tensor("op_9766"), val = tensor([1, 256, 16, 64])]; + tensor k_cast_fp16 = reshape(shape = var_9766, x = ko_cast_fp16)[name = tensor("k_cast_fp16")]; + tensor write_indices_float_139_cast_fp16 = add(x = var_9705_cast_fp16, y = var_393_promoted_to_fp16)[name = tensor("write_indices_float_139_cast_fp16")]; + tensor _inversed_9787_y_0_to_fp16 = const()[name = tensor("_inversed_9787_y_0_to_fp16"), val = tensor(0x1p-9)]; + tensor _inversed_9787_cast_fp16 = mul(x = write_indices_float_139_cast_fp16, y = _inversed_9787_y_0_to_fp16)[name = tensor("_inversed_9787_cast_fp16")]; + tensor var_9788_cast_fp16 = floor(x = _inversed_9787_cast_fp16)[name = tensor("op_9788_cast_fp16")]; + tensor var_9789_to_fp16 = const()[name = tensor("op_9789_to_fp16"), val = tensor(0x1p+9)]; + tensor var_9790_cast_fp16 = mul(x = var_9788_cast_fp16, y = var_9789_to_fp16)[name = tensor("op_9790_cast_fp16")]; + tensor write_indices_float_141_cast_fp16 = sub(x = write_indices_float_139_cast_fp16, y = var_9790_cast_fp16)[name = tensor("write_indices_float_141_cast_fp16")]; + tensor var_9808_value_0_to_fp16 = const()[name = tensor("op_9808_value_0_to_fp16"), val = tensor(0x1.ffp+8)]; + tensor var_9808_cast_fp16 = fill_like(ref_tensor = write_indices_float_141_cast_fp16, value = var_9808_value_0_to_fp16)[name = tensor("op_9808_cast_fp16")]; + tensor write_indices_float_cast_fp16 = select(a = write_indices_float_141_cast_fp16, b = var_9808_cast_fp16, cond = token_valid_1_cast_fp16)[name = tensor("write_indices_float_cast_fp16")]; + tensor var_9814_dtype_0 = const()[name = tensor("op_9814_dtype_0"), val = tensor("int32")]; + tensor var_9819 = const()[name = tensor("op_9819"), val = tensor([1, 256, 1, 1])]; + tensor write_indices_float_cast_fp16_to_int32 = cast(dtype = var_9814_dtype_0, x = write_indices_float_cast_fp16)[name = tensor("cast_545")]; + tensor var_9820 = reshape(shape = var_9819, x = write_indices_float_cast_fp16_to_int32)[name = tensor("op_9820")]; + tensor write_indices_reps_0 = const()[name = tensor("write_indices_reps_0"), val = tensor([1, 1, 16, 64])]; + tensor write_indices = tile(reps = write_indices_reps_0, x = var_9820)[name = tensor("write_indices")]; + tensor var_9830_begin_0 = const()[name = tensor("op_9830_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_9830_end_0 = const()[name = tensor("op_9830_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor var_9830_end_mask_0 = const()[name = tensor("op_9830_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_9830_squeeze_mask_0 = const()[name = tensor("op_9830_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor cache23_to_fp16_dtype_0 = const()[name = tensor("cache23_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor cache23_to_fp16 = cast(dtype = cache23_to_fp16_dtype_0, x = cache23)[name = tensor("cast_544")]; + tensor var_9830_cast_fp16 = slice_by_index(begin = var_9830_begin_0, end = var_9830_end_0, end_mask = var_9830_end_mask_0, squeeze_mask = var_9830_squeeze_mask_0, x = cache23_to_fp16)[name = tensor("op_9830_cast_fp16")]; + tensor var_9832_axis_0 = const()[name = tensor("op_9832_axis_0"), val = tensor(1)]; + tensor var_9832_mode_0 = const()[name = tensor("op_9832_mode_0"), val = tensor("update")]; + tensor var_9832_validate_indices_0 = const()[name = tensor("op_9832_validate_indices_0"), val = tensor(false)]; + tensor var_9832_cast_fp16 = scatter_along_axis(axis = var_9832_axis_0, data = var_9830_cast_fp16, indices = write_indices, mode = var_9832_mode_0, updates = k_cast_fp16, validate_indices = var_9832_validate_indices_0)[name = tensor("op_9832_cast_fp16")]; + tensor concat_186 = const()[name = tensor("concat_186"), val = tensor([0, 0, 0, 0, 0])]; + tensor concat_187 = const()[name = tensor("concat_187"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_internal_tensor_assign_1_stride_0 = const()[name = tensor("new_cache_internal_tensor_assign_1_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_internal_tensor_assign_1_begin_mask_0 = const()[name = tensor("new_cache_internal_tensor_assign_1_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_internal_tensor_assign_1_end_mask_0 = const()[name = tensor("new_cache_internal_tensor_assign_1_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_internal_tensor_assign_1_squeeze_mask_0 = const()[name = tensor("new_cache_internal_tensor_assign_1_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_92 = const()[name = tensor("shape_92"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_46 = const()[name = tensor("reduce_prod_46"), val = tensor(1048576)]; + tensor range_1d_46_start_0 = const()[name = tensor("range_1d_46_start_0"), val = tensor(0)]; + tensor range_1d_46_step_0 = const()[name = tensor("range_1d_46_step_0"), val = tensor(1)]; + tensor range_1d_46 = range_1d(end = reduce_prod_46, start = range_1d_46_start_0, step = range_1d_46_step_0)[name = tensor("range_1d_46")]; + tensor reshape_230 = reshape(shape = shape_92, x = range_1d_46)[name = tensor("reshape_230")]; + tensor slice_by_index_46 = slice_by_index(begin = concat_186, begin_mask = new_cache_internal_tensor_assign_1_begin_mask_0, end = concat_187, end_mask = new_cache_internal_tensor_assign_1_end_mask_0, squeeze_mask = new_cache_internal_tensor_assign_1_squeeze_mask_0, stride = new_cache_internal_tensor_assign_1_stride_0, x = reshape_230)[name = tensor("slice_by_index_46")]; + tensor reshape_231_shape_0 = const()[name = tensor("reshape_231_shape_0"), val = tensor([-1])]; + tensor reshape_231 = reshape(shape = reshape_231_shape_0, x = slice_by_index_46)[name = tensor("reshape_231")]; + tensor reshape_232_shape_0 = const()[name = tensor("reshape_232_shape_0"), val = tensor([-1])]; + tensor reshape_232_cast_fp16 = reshape(shape = reshape_232_shape_0, x = var_9832_cast_fp16)[name = tensor("reshape_232_cast_fp16")]; + tensor reshape_233_shape_0 = const()[name = tensor("reshape_233_shape_0"), val = tensor([-1])]; + tensor reshape_233_cast_fp16 = reshape(shape = reshape_233_shape_0, x = cache23_to_fp16)[name = tensor("reshape_233_cast_fp16")]; + tensor scatter_46_mode_0 = const()[name = tensor("scatter_46_mode_0"), val = tensor("update")]; + tensor scatter_46_axis_0 = const()[name = tensor("scatter_46_axis_0"), val = tensor(0)]; + tensor scatter_46_validate_indices_0 = const()[name = tensor("scatter_46_validate_indices_0"), val = tensor(false)]; + tensor scatter_46_cast_fp16 = scatter(axis = scatter_46_axis_0, data = reshape_233_cast_fp16, indices = reshape_231, mode = scatter_46_mode_0, updates = reshape_232_cast_fp16, validate_indices = scatter_46_validate_indices_0)[name = tensor("scatter_46_cast_fp16")]; + tensor reshape_234_cast_fp16 = reshape(shape = shape_92, x = scatter_46_cast_fp16)[name = tensor("reshape_234_cast_fp16")]; + tensor var_9840_begin_0 = const()[name = tensor("op_9840_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor var_9840_end_0 = const()[name = tensor("op_9840_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor var_9840_end_mask_0 = const()[name = tensor("op_9840_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_9840_squeeze_mask_0 = const()[name = tensor("op_9840_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor var_9840_cast_fp16 = slice_by_index(begin = var_9840_begin_0, end = var_9840_end_0, end_mask = var_9840_end_mask_0, squeeze_mask = var_9840_squeeze_mask_0, x = reshape_234_cast_fp16)[name = tensor("op_9840_cast_fp16")]; + tensor var_9842_axis_0 = const()[name = tensor("op_9842_axis_0"), val = tensor(1)]; + tensor var_9842_mode_0 = const()[name = tensor("op_9842_mode_0"), val = tensor("update")]; + tensor var_9842_validate_indices_0 = const()[name = tensor("op_9842_validate_indices_0"), val = tensor(false)]; + tensor var_9842_cast_fp16 = scatter_along_axis(axis = var_9842_axis_0, data = var_9840_cast_fp16, indices = write_indices, mode = var_9842_mode_0, updates = v_cast_fp16, validate_indices = var_9842_validate_indices_0)[name = tensor("op_9842_cast_fp16")]; + tensor concat_188 = const()[name = tensor("concat_188"), val = tensor([1, 0, 0, 0, 0])]; + tensor concat_189 = const()[name = tensor("concat_189"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_internal_tensor_assign_2_stride_0 = const()[name = tensor("new_cache_internal_tensor_assign_2_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_internal_tensor_assign_2_begin_mask_0 = const()[name = tensor("new_cache_internal_tensor_assign_2_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_internal_tensor_assign_2_end_mask_0 = const()[name = tensor("new_cache_internal_tensor_assign_2_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_internal_tensor_assign_2_squeeze_mask_0 = const()[name = tensor("new_cache_internal_tensor_assign_2_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_93 = const()[name = tensor("shape_93"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_47 = const()[name = tensor("reduce_prod_47"), val = tensor(1048576)]; + tensor range_1d_47_start_0 = const()[name = tensor("range_1d_47_start_0"), val = tensor(0)]; + tensor range_1d_47_step_0 = const()[name = tensor("range_1d_47_step_0"), val = tensor(1)]; + tensor range_1d_47 = range_1d(end = reduce_prod_47, start = range_1d_47_start_0, step = range_1d_47_step_0)[name = tensor("range_1d_47")]; + tensor reshape_235 = reshape(shape = shape_93, x = range_1d_47)[name = tensor("reshape_235")]; + tensor slice_by_index_47 = slice_by_index(begin = concat_188, begin_mask = new_cache_internal_tensor_assign_2_begin_mask_0, end = concat_189, end_mask = new_cache_internal_tensor_assign_2_end_mask_0, squeeze_mask = new_cache_internal_tensor_assign_2_squeeze_mask_0, stride = new_cache_internal_tensor_assign_2_stride_0, x = reshape_235)[name = tensor("slice_by_index_47")]; + tensor reshape_236_shape_0 = const()[name = tensor("reshape_236_shape_0"), val = tensor([-1])]; + tensor reshape_236 = reshape(shape = reshape_236_shape_0, x = slice_by_index_47)[name = tensor("reshape_236")]; + tensor reshape_237_shape_0 = const()[name = tensor("reshape_237_shape_0"), val = tensor([-1])]; + tensor reshape_237_cast_fp16 = reshape(shape = reshape_237_shape_0, x = var_9842_cast_fp16)[name = tensor("reshape_237_cast_fp16")]; + tensor reshape_238_shape_0 = const()[name = tensor("reshape_238_shape_0"), val = tensor([-1])]; + tensor reshape_238_cast_fp16 = reshape(shape = reshape_238_shape_0, x = reshape_234_cast_fp16)[name = tensor("reshape_238_cast_fp16")]; + tensor scatter_47_mode_0 = const()[name = tensor("scatter_47_mode_0"), val = tensor("update")]; + tensor scatter_47_axis_0 = const()[name = tensor("scatter_47_axis_0"), val = tensor(0)]; + tensor scatter_47_validate_indices_0 = const()[name = tensor("scatter_47_validate_indices_0"), val = tensor(false)]; + tensor scatter_47_cast_fp16 = scatter(axis = scatter_47_axis_0, data = reshape_238_cast_fp16, indices = reshape_236, mode = scatter_47_mode_0, updates = reshape_237_cast_fp16, validate_indices = scatter_47_validate_indices_0)[name = tensor("scatter_47_cast_fp16")]; + tensor reshape_239_cast_fp16 = reshape(shape = shape_93, x = scatter_47_cast_fp16)[name = tensor("reshape_239_cast_fp16")]; + tensor reshape_239_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("reshape_239_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor var_9852_cast_fp16 = add(x = position23_to_fp16, y = valid_len_to_fp16)[name = tensor("op_9852_cast_fp16")]; + tensor var_9852_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_9852_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor var_9852 = cast(dtype = var_9852_cast_fp16_to_fp32_dtype_0, x = var_9852_cast_fp16)[name = tensor("cast_542")]; + tensor new_cache_internal_tensor_assign_2 = cast(dtype = reshape_239_cast_fp16_to_fp32_dtype_0, x = reshape_239_cast_fp16)[name = tensor("cast_543")]; + tensor var_9624 = cast(dtype = var_9624_cast_fp16_to_fp32_dtype_0, x = var_9624_cast_fp16)[name = tensor("cast_547")]; + tensor new_cache_45_internal_tensor_assign_2 = cast(dtype = reshape_229_cast_fp16_to_fp32_dtype_0, x = reshape_229_cast_fp16)[name = tensor("cast_548")]; + tensor var_9213 = cast(dtype = var_9213_cast_fp16_to_fp32_dtype_0, x = var_9213_cast_fp16)[name = tensor("cast_552")]; + tensor new_cache_43_internal_tensor_assign_2 = cast(dtype = reshape_219_cast_fp16_to_fp32_dtype_0, x = reshape_219_cast_fp16)[name = tensor("cast_553")]; + tensor var_8802 = cast(dtype = var_8802_cast_fp16_to_fp32_dtype_0, x = var_8802_cast_fp16)[name = tensor("cast_557")]; + tensor new_cache_41_internal_tensor_assign_2 = cast(dtype = reshape_209_cast_fp16_to_fp32_dtype_0, x = reshape_209_cast_fp16)[name = tensor("cast_558")]; + tensor var_8391 = cast(dtype = var_8391_cast_fp16_to_fp32_dtype_0, x = var_8391_cast_fp16)[name = tensor("cast_562")]; + tensor new_cache_39_internal_tensor_assign_2 = cast(dtype = reshape_199_cast_fp16_to_fp32_dtype_0, x = reshape_199_cast_fp16)[name = tensor("cast_563")]; + tensor var_7980 = cast(dtype = var_7980_cast_fp16_to_fp32_dtype_0, x = var_7980_cast_fp16)[name = tensor("cast_567")]; + tensor new_cache_37_internal_tensor_assign_2 = cast(dtype = reshape_189_cast_fp16_to_fp32_dtype_0, x = reshape_189_cast_fp16)[name = tensor("cast_568")]; + tensor var_7569 = cast(dtype = var_7569_cast_fp16_to_fp32_dtype_0, x = var_7569_cast_fp16)[name = tensor("cast_572")]; + tensor new_cache_35_internal_tensor_assign_2 = cast(dtype = reshape_179_cast_fp16_to_fp32_dtype_0, x = reshape_179_cast_fp16)[name = tensor("cast_573")]; + tensor var_7158 = cast(dtype = var_7158_cast_fp16_to_fp32_dtype_0, x = var_7158_cast_fp16)[name = tensor("cast_577")]; + tensor new_cache_33_internal_tensor_assign_2 = cast(dtype = reshape_169_cast_fp16_to_fp32_dtype_0, x = reshape_169_cast_fp16)[name = tensor("cast_578")]; + tensor var_6747 = cast(dtype = var_6747_cast_fp16_to_fp32_dtype_0, x = var_6747_cast_fp16)[name = tensor("cast_582")]; + tensor new_cache_31_internal_tensor_assign_2 = cast(dtype = reshape_159_cast_fp16_to_fp32_dtype_0, x = reshape_159_cast_fp16)[name = tensor("cast_583")]; + tensor var_6336 = cast(dtype = var_6336_cast_fp16_to_fp32_dtype_0, x = var_6336_cast_fp16)[name = tensor("cast_587")]; + tensor new_cache_29_internal_tensor_assign_2 = cast(dtype = reshape_149_cast_fp16_to_fp32_dtype_0, x = reshape_149_cast_fp16)[name = tensor("cast_588")]; + tensor var_5925 = cast(dtype = var_5925_cast_fp16_to_fp32_dtype_0, x = var_5925_cast_fp16)[name = tensor("cast_592")]; + tensor new_cache_27_internal_tensor_assign_2 = cast(dtype = reshape_139_cast_fp16_to_fp32_dtype_0, x = reshape_139_cast_fp16)[name = tensor("cast_593")]; + tensor var_5514 = cast(dtype = var_5514_cast_fp16_to_fp32_dtype_0, x = var_5514_cast_fp16)[name = tensor("cast_597")]; + tensor new_cache_25_internal_tensor_assign_2 = cast(dtype = reshape_129_cast_fp16_to_fp32_dtype_0, x = reshape_129_cast_fp16)[name = tensor("cast_598")]; + tensor var_5103 = cast(dtype = var_5103_cast_fp16_to_fp32_dtype_0, x = var_5103_cast_fp16)[name = tensor("cast_602")]; + tensor new_cache_23_internal_tensor_assign_2 = cast(dtype = reshape_119_cast_fp16_to_fp32_dtype_0, x = reshape_119_cast_fp16)[name = tensor("cast_603")]; + tensor var_4692 = cast(dtype = var_4692_cast_fp16_to_fp32_dtype_0, x = var_4692_cast_fp16)[name = tensor("cast_607")]; + tensor new_cache_21_internal_tensor_assign_2 = cast(dtype = reshape_109_cast_fp16_to_fp32_dtype_0, x = reshape_109_cast_fp16)[name = tensor("cast_608")]; + tensor var_4281 = cast(dtype = var_4281_cast_fp16_to_fp32_dtype_0, x = var_4281_cast_fp16)[name = tensor("cast_612")]; + tensor new_cache_19_internal_tensor_assign_2 = cast(dtype = reshape_99_cast_fp16_to_fp32_dtype_0, x = reshape_99_cast_fp16)[name = tensor("cast_613")]; + tensor var_3870 = cast(dtype = var_3870_cast_fp16_to_fp32_dtype_0, x = var_3870_cast_fp16)[name = tensor("cast_617")]; + tensor new_cache_17_internal_tensor_assign_2 = cast(dtype = reshape_89_cast_fp16_to_fp32_dtype_0, x = reshape_89_cast_fp16)[name = tensor("cast_618")]; + tensor var_3459 = cast(dtype = var_3459_cast_fp16_to_fp32_dtype_0, x = var_3459_cast_fp16)[name = tensor("cast_622")]; + tensor new_cache_15_internal_tensor_assign_2 = cast(dtype = reshape_79_cast_fp16_to_fp32_dtype_0, x = reshape_79_cast_fp16)[name = tensor("cast_623")]; + tensor var_3048 = cast(dtype = var_3048_cast_fp16_to_fp32_dtype_0, x = var_3048_cast_fp16)[name = tensor("cast_627")]; + tensor new_cache_13_internal_tensor_assign_2 = cast(dtype = reshape_69_cast_fp16_to_fp32_dtype_0, x = reshape_69_cast_fp16)[name = tensor("cast_628")]; + tensor var_2637 = cast(dtype = var_2637_cast_fp16_to_fp32_dtype_0, x = var_2637_cast_fp16)[name = tensor("cast_632")]; + tensor new_cache_11_internal_tensor_assign_2 = cast(dtype = reshape_59_cast_fp16_to_fp32_dtype_0, x = reshape_59_cast_fp16)[name = tensor("cast_633")]; + tensor var_2226 = cast(dtype = var_2226_cast_fp16_to_fp32_dtype_0, x = var_2226_cast_fp16)[name = tensor("cast_637")]; + tensor new_cache_9_internal_tensor_assign_2 = cast(dtype = reshape_49_cast_fp16_to_fp32_dtype_0, x = reshape_49_cast_fp16)[name = tensor("cast_638")]; + tensor var_1815 = cast(dtype = var_1815_cast_fp16_to_fp32_dtype_0, x = var_1815_cast_fp16)[name = tensor("cast_642")]; + tensor new_cache_7_internal_tensor_assign_2 = cast(dtype = reshape_39_cast_fp16_to_fp32_dtype_0, x = reshape_39_cast_fp16)[name = tensor("cast_643")]; + tensor var_1404 = cast(dtype = var_1404_cast_fp16_to_fp32_dtype_0, x = var_1404_cast_fp16)[name = tensor("cast_647")]; + tensor new_cache_5_internal_tensor_assign_2 = cast(dtype = reshape_29_cast_fp16_to_fp32_dtype_0, x = reshape_29_cast_fp16)[name = tensor("cast_648")]; + tensor var_993 = cast(dtype = var_993_cast_fp16_to_fp32_dtype_0, x = var_993_cast_fp16)[name = tensor("cast_652")]; + tensor new_cache_3_internal_tensor_assign_2 = cast(dtype = reshape_19_cast_fp16_to_fp32_dtype_0, x = reshape_19_cast_fp16)[name = tensor("cast_653")]; + tensor var_582 = cast(dtype = var_582_cast_fp16_to_fp32_dtype_0, x = var_582_cast_fp16)[name = tensor("cast_657")]; + tensor new_cache_1_internal_tensor_assign_2 = cast(dtype = reshape_9_cast_fp16_to_fp32_dtype_0, x = reshape_9_cast_fp16)[name = tensor("cast_658")]; + } -> (new_cache_1_internal_tensor_assign_2, var_582, new_cache_3_internal_tensor_assign_2, var_993, new_cache_5_internal_tensor_assign_2, var_1404, new_cache_7_internal_tensor_assign_2, var_1815, new_cache_9_internal_tensor_assign_2, var_2226, new_cache_11_internal_tensor_assign_2, var_2637, new_cache_13_internal_tensor_assign_2, var_3048, new_cache_15_internal_tensor_assign_2, var_3459, new_cache_17_internal_tensor_assign_2, var_3870, new_cache_19_internal_tensor_assign_2, var_4281, new_cache_21_internal_tensor_assign_2, var_4692, new_cache_23_internal_tensor_assign_2, var_5103, new_cache_25_internal_tensor_assign_2, var_5514, new_cache_27_internal_tensor_assign_2, var_5925, new_cache_29_internal_tensor_assign_2, var_6336, new_cache_31_internal_tensor_assign_2, var_6747, new_cache_33_internal_tensor_assign_2, var_7158, new_cache_35_internal_tensor_assign_2, var_7569, new_cache_37_internal_tensor_assign_2, var_7980, new_cache_39_internal_tensor_assign_2, var_8391, new_cache_41_internal_tensor_assign_2, var_8802, new_cache_43_internal_tensor_assign_2, var_9213, new_cache_45_internal_tensor_assign_2, var_9624, new_cache_internal_tensor_assign_2, var_9852); +} \ No newline at end of file diff --git a/v2.1/portuguese_24l/cond_prefill.mlmodelc/weights/weight.bin b/v2.1/portuguese_24l/cond_prefill.mlmodelc/weights/weight.bin new file mode 100644 index 0000000000000000000000000000000000000000..da4138e6ac7b6a1e8f1b4364f62754dc8297c947 --- /dev/null +++ b/v2.1/portuguese_24l/cond_prefill.mlmodelc/weights/weight.bin @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:a4bf30d44c991212c42dd0e5b88d0fc5e0bbaad79bf609b0625e20126ef4b6df +size 586381056 diff --git a/v2.1/portuguese_24l/cond_prefill.mlpackage/Data/com.apple.CoreML/model.mlmodel b/v2.1/portuguese_24l/cond_prefill.mlpackage/Data/com.apple.CoreML/model.mlmodel new file mode 100644 index 0000000000000000000000000000000000000000..f41ec34fc03e676f761a0d79b365634055001515 --- /dev/null +++ b/v2.1/portuguese_24l/cond_prefill.mlpackage/Data/com.apple.CoreML/model.mlmodel @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:4dc1a8c47b2e2f7edf2cc02981e541729fd7d3f1fed0cbd48b093ed1ee830e7c +size 842494 diff --git a/v2.1/portuguese_24l/cond_prefill.mlpackage/Data/com.apple.CoreML/weights/weight.bin b/v2.1/portuguese_24l/cond_prefill.mlpackage/Data/com.apple.CoreML/weights/weight.bin new file mode 100644 index 0000000000000000000000000000000000000000..da4138e6ac7b6a1e8f1b4364f62754dc8297c947 --- /dev/null +++ b/v2.1/portuguese_24l/cond_prefill.mlpackage/Data/com.apple.CoreML/weights/weight.bin @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:a4bf30d44c991212c42dd0e5b88d0fc5e0bbaad79bf609b0625e20126ef4b6df +size 586381056 diff --git a/v2.1/portuguese_24l/cond_prefill.mlpackage/Manifest.json b/v2.1/portuguese_24l/cond_prefill.mlpackage/Manifest.json new file mode 100644 index 0000000000000000000000000000000000000000..966286088c8cade16baf8cf11ac66b5feb369063 --- /dev/null +++ b/v2.1/portuguese_24l/cond_prefill.mlpackage/Manifest.json @@ -0,0 +1,18 @@ +{ + "fileFormatVersion": "1.0.0", + "itemInfoEntries": { + "ACB20C56-A7D3-4821-B0F0-508804F84400": { + "author": "com.apple.CoreML", + "description": "CoreML Model Weights", + "name": "weights", + "path": "com.apple.CoreML/weights" + }, + "F9044B83-05D4-4150-A891-A42B178FC0BA": { + "author": "com.apple.CoreML", + "description": "CoreML Model Specification", + "name": "model.mlmodel", + "path": "com.apple.CoreML/model.mlmodel" + } + }, + "rootModelIdentifier": "F9044B83-05D4-4150-A891-A42B178FC0BA" +} diff --git a/v2.1/portuguese_24l/constants/bos_emb.npy b/v2.1/portuguese_24l/constants/bos_emb.npy new file mode 100644 index 0000000000000000000000000000000000000000..9a6f3a6ade71dc6a5e720ea9b3b334e177a9907d --- /dev/null +++ b/v2.1/portuguese_24l/constants/bos_emb.npy @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:76f5195970f2502e5453ac213487a37a2decd2ca98cb3878e731401fc79fa105 +size 256 diff --git a/v2.1/portuguese_24l/constants/emb_mean.npy b/v2.1/portuguese_24l/constants/emb_mean.npy new file mode 100644 index 0000000000000000000000000000000000000000..f3e81b1fd3c115acf06fc2f8f16c98ab076786de --- /dev/null +++ b/v2.1/portuguese_24l/constants/emb_mean.npy @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:6e8f8d7687fbfb92550df4980660eb445bff75ad6a72017524a15c532f38c0c7 +size 256 diff --git a/v2.1/portuguese_24l/constants/emb_std.npy b/v2.1/portuguese_24l/constants/emb_std.npy new file mode 100644 index 0000000000000000000000000000000000000000..73d8b42712f888ece58633a0648eb11b89f00518 --- /dev/null +++ b/v2.1/portuguese_24l/constants/emb_std.npy @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:f15a323c90a6e6914ac259aa3e5fb32be19a6b40818293c84330aefd60d59bb8 +size 256 diff --git a/v2.1/portuguese_24l/constants/mimi_init_state.npz b/v2.1/portuguese_24l/constants/mimi_init_state.npz new file mode 100644 index 0000000000000000000000000000000000000000..3b52f383a86ef0439f9993506e4529c4c1234828 --- /dev/null +++ b/v2.1/portuguese_24l/constants/mimi_init_state.npz @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:515588e10e9c14a6b009f1cfda9fcd592542b1640ad856d5fc69e50024702a39 +size 1083108 diff --git a/v2.1/portuguese_24l/constants/quantizer_weight.npy b/v2.1/portuguese_24l/constants/quantizer_weight.npy new file mode 100644 index 0000000000000000000000000000000000000000..a6d7797c101669ae052c8f4fa42bf751b61d3694 --- /dev/null +++ b/v2.1/portuguese_24l/constants/quantizer_weight.npy @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:f44dcaa3aae563c566c19291ba101b751765d96e1ed4cb3ded8f8fc655d23cee +size 65664 diff --git a/v2.1/portuguese_24l/constants/text_embed_table.npy b/v2.1/portuguese_24l/constants/text_embed_table.npy new file mode 100644 index 0000000000000000000000000000000000000000..e142d27c947d8186b4eee3a7556d7777f90c1310 --- /dev/null +++ b/v2.1/portuguese_24l/constants/text_embed_table.npy @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:647a70f0fd85b4618d784d5bb08f4542fc48a51e29ead61d59f6431006a6b0a5 +size 16388224 diff --git a/v2.1/portuguese_24l/constants_bin/alba.safetensors b/v2.1/portuguese_24l/constants_bin/alba.safetensors new file mode 100644 index 0000000000000000000000000000000000000000..0771072c4b7c64b1b60e017b38bda3bd5dbd0483 --- /dev/null +++ b/v2.1/portuguese_24l/constants_bin/alba.safetensors @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:e3a6206c2f25b3b5a550563e4e163f7bf5e97a4ac671cf0aad5ff04d5179f9be +size 24777760 diff --git a/v2.1/portuguese_24l/constants_bin/anna.safetensors b/v2.1/portuguese_24l/constants_bin/anna.safetensors new file mode 100644 index 0000000000000000000000000000000000000000..9a32d27dd980969bfed6ecdd5f25cb4f3cd8d074 --- /dev/null +++ b/v2.1/portuguese_24l/constants_bin/anna.safetensors @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:c3ad8d9b8213411bbb76ac0adb6cecca5778716489b618b14e5a239280a357da +size 31265824 diff --git a/v2.1/portuguese_24l/constants_bin/azelma.safetensors b/v2.1/portuguese_24l/constants_bin/azelma.safetensors new file mode 100644 index 0000000000000000000000000000000000000000..7743367b4e319544efa601a9d14515b50993a263 --- /dev/null +++ b/v2.1/portuguese_24l/constants_bin/azelma.safetensors @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:2bd4dd2fdcae7790bace547f9e5f4dfc12f30a9150b0ab1d1e9fa7c2aa273e17 +size 31855648 diff --git a/v2.1/portuguese_24l/constants_bin/bill_boerst.safetensors b/v2.1/portuguese_24l/constants_bin/bill_boerst.safetensors new file mode 100644 index 0000000000000000000000000000000000000000..276c7d03a496dcda982edf245575efed66c8ae47 --- /dev/null +++ b/v2.1/portuguese_24l/constants_bin/bill_boerst.safetensors @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:c73db388d50dd61017b27e07f674667ebcb5143745792dec2785fda668aa88cd +size 26940448 diff --git a/v2.1/portuguese_24l/constants_bin/bos_before_voice.bin b/v2.1/portuguese_24l/constants_bin/bos_before_voice.bin new file mode 100644 index 0000000000000000000000000000000000000000..20827eb3e892a771c827c30f4a5157f7215a9c80 --- /dev/null +++ b/v2.1/portuguese_24l/constants_bin/bos_before_voice.bin @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:230f76bb4eadd75b409fcfbe063432a5f6d02054e0f9b3b5bcc74d476785d8ae +size 4096 diff --git a/v2.1/portuguese_24l/constants_bin/bos_emb.bin b/v2.1/portuguese_24l/constants_bin/bos_emb.bin new file mode 100644 index 0000000000000000000000000000000000000000..8d28c04d03a54bb32d512ca4d9f824517b8176fd --- /dev/null +++ b/v2.1/portuguese_24l/constants_bin/bos_emb.bin @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:f246d94d64d1393735e26f51f3ed3f5057cf821cd5602f263f308f83889a78f2 +size 128 diff --git a/v2.1/portuguese_24l/constants_bin/caro_davy.safetensors b/v2.1/portuguese_24l/constants_bin/caro_davy.safetensors new file mode 100644 index 0000000000000000000000000000000000000000..b6b77336a52c21ce3d41d8d96490c74eb6f4c0e3 --- /dev/null +++ b/v2.1/portuguese_24l/constants_bin/caro_davy.safetensors @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:da23995bcf5d2241dfdbb1e01ae7354c0a5a9c26faa67933dfd80907193c11d2 +size 21042200 diff --git a/v2.1/portuguese_24l/constants_bin/charles.safetensors b/v2.1/portuguese_24l/constants_bin/charles.safetensors new file mode 100644 index 0000000000000000000000000000000000000000..a005b5a828c3bfd33eac50aa9c4339e82293bf6a --- /dev/null +++ b/v2.1/portuguese_24l/constants_bin/charles.safetensors @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:778a6deb4048eef6fdedc332bd44db5f96724757a6f32b6530b945ff5c34ecde +size 24777760 diff --git a/v2.1/portuguese_24l/constants_bin/cosette.safetensors b/v2.1/portuguese_24l/constants_bin/cosette.safetensors new file mode 100644 index 0000000000000000000000000000000000000000..94ce99b246698be1d98cd2db50678cf8bd5e0d62 --- /dev/null +++ b/v2.1/portuguese_24l/constants_bin/cosette.safetensors @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:ee83d39e63d24158cb04be187af7a0d8672ee70bed18fd49bf052c1d1e79cd9d +size 24777760 diff --git a/v2.1/portuguese_24l/constants_bin/eponine.safetensors b/v2.1/portuguese_24l/constants_bin/eponine.safetensors new file mode 100644 index 0000000000000000000000000000000000000000..df009c158783eb45145062d516cd7ef6626f12f8 --- /dev/null +++ b/v2.1/portuguese_24l/constants_bin/eponine.safetensors @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:3e46900c93f7c7c4c2e0907f767218ea4448aaf10883b72fcc4f17cbb5a2ca1e +size 27726880 diff --git a/v2.1/portuguese_24l/constants_bin/estelle.safetensors b/v2.1/portuguese_24l/constants_bin/estelle.safetensors new file mode 100644 index 0000000000000000000000000000000000000000..ef2816d6099681cffa4206ce243598f3222fb7b5 --- /dev/null +++ b/v2.1/portuguese_24l/constants_bin/estelle.safetensors @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:b832a48fe2dcb2c4a3d45509c2b5923b23b2cae552197ac66ccc6c4058169a73 +size 33035296 diff --git a/v2.1/portuguese_24l/constants_bin/eve.safetensors b/v2.1/portuguese_24l/constants_bin/eve.safetensors new file mode 100644 index 0000000000000000000000000000000000000000..d98e6e788c199f91b626c8a1caa21780285bc816 --- /dev/null +++ b/v2.1/portuguese_24l/constants_bin/eve.safetensors @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:05300e07f0abc233b74273fc2d621bdbe87543bce8a8d08d6d96d1566feac595 +size 26154016 diff --git a/v2.1/portuguese_24l/constants_bin/fantine.safetensors b/v2.1/portuguese_24l/constants_bin/fantine.safetensors new file mode 100644 index 0000000000000000000000000000000000000000..d3350d526999ae902959779079f27e07ee5ec455 --- /dev/null +++ b/v2.1/portuguese_24l/constants_bin/fantine.safetensors @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:eb49a4df2c7089473fc75f47c0da36ac1669a05eac954838ab3ade70ff1a3218 +size 26154016 diff --git a/v2.1/portuguese_24l/constants_bin/george.safetensors b/v2.1/portuguese_24l/constants_bin/george.safetensors new file mode 100644 index 0000000000000000000000000000000000000000..f3dc58681b9c167002e67e0725c026331a9570c1 --- /dev/null +++ b/v2.1/portuguese_24l/constants_bin/george.safetensors @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:02ab4c9b107b107e6482b55bc8e552a66ef6f7b38a3c8413493539f03a025ba2 +size 24974368 diff --git a/v2.1/portuguese_24l/constants_bin/giovanni.safetensors b/v2.1/portuguese_24l/constants_bin/giovanni.safetensors new file mode 100644 index 0000000000000000000000000000000000000000..0af4e4f9f156b38f7e1102b78ea1f210ba69aaf0 --- /dev/null +++ b/v2.1/portuguese_24l/constants_bin/giovanni.safetensors @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:d89146f53c5beb0298710ba99b9ee7bd487a49446a8b03ddcc517c6f0a081044 +size 18486272 diff --git a/v2.1/portuguese_24l/constants_bin/jane.safetensors b/v2.1/portuguese_24l/constants_bin/jane.safetensors new file mode 100644 index 0000000000000000000000000000000000000000..cbfe6f69fb9797a7afce2410cec793d0ad2cc07a --- /dev/null +++ b/v2.1/portuguese_24l/constants_bin/jane.safetensors @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:845071b854f727c36d67e2afa099cea221d5a7ed1e2d2730b329422cb97e92ac +size 29496352 diff --git a/v2.1/portuguese_24l/constants_bin/javert.safetensors b/v2.1/portuguese_24l/constants_bin/javert.safetensors new file mode 100644 index 0000000000000000000000000000000000000000..60938977c2bdd467a427d73de00647a96380ca8c --- /dev/null +++ b/v2.1/portuguese_24l/constants_bin/javert.safetensors @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:918450696f2511dde79ac6a79dfec88b19aa62f353d67be7694e7edf7c25e30c +size 24777760 diff --git a/v2.1/portuguese_24l/constants_bin/jean.safetensors b/v2.1/portuguese_24l/constants_bin/jean.safetensors new file mode 100644 index 0000000000000000000000000000000000000000..754c311300cf494ecc374170b7d906c7ff3f0029 --- /dev/null +++ b/v2.1/portuguese_24l/constants_bin/jean.safetensors @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:1c94f7fbc94951814c9eab8b893539e8be6ad815bd157c08dbcb2ceebc6faba7 +size 24777760 diff --git a/v2.1/portuguese_24l/constants_bin/juergen.safetensors b/v2.1/portuguese_24l/constants_bin/juergen.safetensors new file mode 100644 index 0000000000000000000000000000000000000000..f3919b51f2011df877c664b1660dd7ab152666a1 --- /dev/null +++ b/v2.1/portuguese_24l/constants_bin/juergen.safetensors @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:dd9eb3db0428fb5080fe471b08bdaa84bc94a7fc2cce04a5467e5df5ad231a1b +size 24974368 diff --git a/v2.1/portuguese_24l/constants_bin/lola.safetensors b/v2.1/portuguese_24l/constants_bin/lola.safetensors new file mode 100644 index 0000000000000000000000000000000000000000..14b52a46e1321181c6ff8d4b497a1c055d0e05cb --- /dev/null +++ b/v2.1/portuguese_24l/constants_bin/lola.safetensors @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:fc06fdc643c81f47a756749bd71530300cab9c26c77f3d42d0c10e5511c2bfdf +size 23794712 diff --git a/v2.1/portuguese_24l/constants_bin/marius.safetensors b/v2.1/portuguese_24l/constants_bin/marius.safetensors new file mode 100644 index 0000000000000000000000000000000000000000..1c1a6ddaa4072447fd60a117a16620827fd0e753 --- /dev/null +++ b/v2.1/portuguese_24l/constants_bin/marius.safetensors @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:c2cc362dfdd4b1a3248ffc8fe7e2d053cd47c83ddda96d23b63e4eca48829399 +size 24777760 diff --git a/v2.1/portuguese_24l/constants_bin/mary.safetensors b/v2.1/portuguese_24l/constants_bin/mary.safetensors new file mode 100644 index 0000000000000000000000000000000000000000..1701e82f666d3dce494fbe48a1a876be88e4c746 --- /dev/null +++ b/v2.1/portuguese_24l/constants_bin/mary.safetensors @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:1236428e293f88c7a1c48d6b91572b5476672d1b4b4526f4a806231921f6bdcd +size 24777760 diff --git a/v2.1/portuguese_24l/constants_bin/michael.safetensors b/v2.1/portuguese_24l/constants_bin/michael.safetensors new file mode 100644 index 0000000000000000000000000000000000000000..1814a847484038ae91658f317dd46ece4b162227 --- /dev/null +++ b/v2.1/portuguese_24l/constants_bin/michael.safetensors @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:37c3a8cd707f8695e528a090ac92cef49eef887ebb177d9a4ecac1202590315f +size 29103136 diff --git a/v2.1/portuguese_24l/constants_bin/paul.safetensors b/v2.1/portuguese_24l/constants_bin/paul.safetensors new file mode 100644 index 0000000000000000000000000000000000000000..6e9e1ca43a68ab94ce0426a4baa676c960cb74d2 --- /dev/null +++ b/v2.1/portuguese_24l/constants_bin/paul.safetensors @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:2b5975480eadd41af531ab42336e2f5f7ff3a0e5970e866d32540169096c31f9 +size 27923488 diff --git a/v2.1/portuguese_24l/constants_bin/peter_yearsley.safetensors b/v2.1/portuguese_24l/constants_bin/peter_yearsley.safetensors new file mode 100644 index 0000000000000000000000000000000000000000..c860bdd994f3235f0b27754c6cccb85914c75487 --- /dev/null +++ b/v2.1/portuguese_24l/constants_bin/peter_yearsley.safetensors @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:732a601337adccf2d16df2f569d1f821129fd2556e99586377a0b550be9dcbe7 +size 14947320 diff --git a/v2.1/portuguese_24l/constants_bin/rafael.safetensors b/v2.1/portuguese_24l/constants_bin/rafael.safetensors new file mode 100644 index 0000000000000000000000000000000000000000..f97f697ed55073a51643f18e1928de7f42e31ac3 --- /dev/null +++ b/v2.1/portuguese_24l/constants_bin/rafael.safetensors @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:f51d1b95e3510c1a7b05efe81693cd52cdcde261464179cde8410a14b98ee5e6 +size 24777760 diff --git a/v2.1/portuguese_24l/constants_bin/stuart_bell.safetensors b/v2.1/portuguese_24l/constants_bin/stuart_bell.safetensors new file mode 100644 index 0000000000000000000000000000000000000000..49f7e9a7b6920b0d3fe760643f257881524fde34 --- /dev/null +++ b/v2.1/portuguese_24l/constants_bin/stuart_bell.safetensors @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:7747f0f4eb873e8a392d57cd0b2403b9e463511f6c0f912ece118f3856045e36 +size 21042200 diff --git a/v2.1/portuguese_24l/constants_bin/text_embed_table.bin b/v2.1/portuguese_24l/constants_bin/text_embed_table.bin new file mode 100644 index 0000000000000000000000000000000000000000..7742e7854fd575397a68666031cb7f03b78300eb --- /dev/null +++ b/v2.1/portuguese_24l/constants_bin/text_embed_table.bin @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:75ca1c05a9ced9a5d5246b01155299c3b881ac529c4a7cc64938a40ba7866539 +size 16388096 diff --git a/v2.1/portuguese_24l/constants_bin/tokenizer.model b/v2.1/portuguese_24l/constants_bin/tokenizer.model new file mode 100644 index 0000000000000000000000000000000000000000..1b7208270cda6aa8e2fa350d24d798eb6dee35ad --- /dev/null +++ b/v2.1/portuguese_24l/constants_bin/tokenizer.model @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:3aa51309c55f114771c156aaeb86f6fc325991364aa3c38af74aecf1cbd0fade +size 60995 diff --git a/v2.1/portuguese_24l/constants_bin/vera.safetensors b/v2.1/portuguese_24l/constants_bin/vera.safetensors new file mode 100644 index 0000000000000000000000000000000000000000..2a66f637c2b0ea10a4da8999eef3b17e00f0470b --- /dev/null +++ b/v2.1/portuguese_24l/constants_bin/vera.safetensors @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:b6505407c63dd50a334a11f86450ee767e59e055cbdacd5d25d248f5cd63bb4d +size 26940448 diff --git a/v2.1/portuguese_24l/flow_decoder_fused.mlmodelc/analytics/coremldata.bin b/v2.1/portuguese_24l/flow_decoder_fused.mlmodelc/analytics/coremldata.bin new file mode 100644 index 0000000000000000000000000000000000000000..0e8222bf46d581754347f6508c10894aebd15c39 --- /dev/null +++ b/v2.1/portuguese_24l/flow_decoder_fused.mlmodelc/analytics/coremldata.bin @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:1d8147b08eecfde7928aa4879175419df0acdeed4d2f34d11f88c0552a1e9a45 +size 243 diff --git a/v2.1/portuguese_24l/flow_decoder_fused.mlmodelc/coremldata.bin b/v2.1/portuguese_24l/flow_decoder_fused.mlmodelc/coremldata.bin new file mode 100644 index 0000000000000000000000000000000000000000..48d17a84f8eff292be995d39150618a51b091326 --- /dev/null +++ b/v2.1/portuguese_24l/flow_decoder_fused.mlmodelc/coremldata.bin @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:28cfa0328e31d406f42b645e7a86f29c216315287a58e1626f80584fc4bc6d17 +size 413 diff --git a/v2.1/portuguese_24l/flow_decoder_fused.mlmodelc/model.mil b/v2.1/portuguese_24l/flow_decoder_fused.mlmodelc/model.mil new file mode 100644 index 0000000000000000000000000000000000000000..c64a1f34555e7bbf9c8d4502d29f9ef52dd6b0ec --- /dev/null +++ b/v2.1/portuguese_24l/flow_decoder_fused.mlmodelc/model.mil @@ -0,0 +1,1915 @@ +program(1.0) +[buildInfo = dict, tensor>({{"coremlc-component-MIL", "3520.4.1"}, {"coremlc-version", "3520.5.1"}, {"coremltools-component-torch", "2.12.0"}, {"coremltools-source-dialect", "TorchScript"}, {"coremltools-version", "9.0"}})] +{ + func main(tensor latent_init, tensor transformer_out) { + tensor latent_init_to_fp16_dtype_0 = const()[name = tensor("latent_init_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor flow_net_input_proj_weight_to_fp16 = const()[name = tensor("flow_net_input_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(64)))]; + tensor flow_net_input_proj_bias_to_fp16 = const()[name = tensor("flow_net_input_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(32896)))]; + tensor latent_init_to_fp16 = cast(dtype = latent_init_to_fp16_dtype_0, x = latent_init)[name = tensor("cast_146")]; + tensor linear_0_cast_fp16 = linear(bias = flow_net_input_proj_bias_to_fp16, weight = flow_net_input_proj_weight_to_fp16, x = latent_init_to_fp16)[name = tensor("linear_0_cast_fp16")]; + tensor input_3_to_fp16 = const()[name = tensor("input_3_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(33984)))]; + tensor input_5_cast_fp16 = silu(x = input_3_to_fp16)[name = tensor("input_5_cast_fp16")]; + tensor flow_net_time_embed_0_mlp_2_weight_to_fp16 = const()[name = tensor("flow_net_time_embed_0_mlp_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(35072)))]; + tensor flow_net_time_embed_0_mlp_2_bias_to_fp16 = const()[name = tensor("flow_net_time_embed_0_mlp_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(559424)))]; + tensor linear_2_cast_fp16 = linear(bias = flow_net_time_embed_0_mlp_2_bias_to_fp16, weight = flow_net_time_embed_0_mlp_2_weight_to_fp16, x = input_5_cast_fp16)[name = tensor("linear_2_cast_fp16")]; + tensor reduce_mean_0_axes_0 = const()[name = tensor("reduce_mean_0_axes_0"), val = tensor([-1])]; + tensor reduce_mean_0_keep_dims_0 = const()[name = tensor("reduce_mean_0_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_0_cast_fp16 = reduce_mean(axes = reduce_mean_0_axes_0, keep_dims = reduce_mean_0_keep_dims_0, x = linear_2_cast_fp16)[name = tensor("reduce_mean_0_cast_fp16")]; + tensor sub_0_cast_fp16 = sub(x = linear_2_cast_fp16, y = reduce_mean_0_cast_fp16)[name = tensor("sub_0_cast_fp16")]; + tensor square_0_cast_fp16 = square(x = sub_0_cast_fp16)[name = tensor("square_0_cast_fp16")]; + tensor reduce_mean_1_axes_0 = const()[name = tensor("reduce_mean_1_axes_0"), val = tensor([-1])]; + tensor reduce_mean_1_keep_dims_0 = const()[name = tensor("reduce_mean_1_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_1_cast_fp16 = reduce_mean(axes = reduce_mean_1_axes_0, keep_dims = reduce_mean_1_keep_dims_0, x = square_0_cast_fp16)[name = tensor("reduce_mean_1_cast_fp16")]; + tensor real_div_0_to_fp16 = const()[name = tensor("real_div_0_to_fp16"), val = tensor(0x1.008p+0)]; + tensor mul_0_cast_fp16 = mul(x = reduce_mean_1_cast_fp16, y = real_div_0_to_fp16)[name = tensor("mul_0_cast_fp16")]; + tensor var_70_to_fp16 = const()[name = tensor("op_70_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_1_cast_fp16 = add(x = mul_0_cast_fp16, y = var_70_to_fp16)[name = tensor("var_1_cast_fp16")]; + tensor var_73_epsilon_0 = const()[name = tensor("op_73_epsilon_0"), val = tensor(0x1.197998p-40)]; + tensor var_73_cast_fp16 = rsqrt(epsilon = var_73_epsilon_0, x = var_1_cast_fp16)[name = tensor("op_73_cast_fp16")]; + tensor const_3_to_fp16 = const()[name = tensor("const_3_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(560512)))]; + tensor var_74_cast_fp16 = mul(x = const_3_to_fp16, y = var_73_cast_fp16)[name = tensor("op_74_cast_fp16")]; + tensor var_75_cast_fp16 = mul(x = linear_2_cast_fp16, y = var_74_cast_fp16)[name = tensor("op_75_cast_fp16")]; + tensor input_9_to_fp16 = const()[name = tensor("input_9_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(561600)))]; + tensor input_11_cast_fp16 = silu(x = input_9_to_fp16)[name = tensor("input_11_cast_fp16")]; + tensor flow_net_time_embed_1_mlp_2_weight_to_fp16 = const()[name = tensor("flow_net_time_embed_1_mlp_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(562688)))]; + tensor flow_net_time_embed_1_mlp_2_bias_to_fp16 = const()[name = tensor("flow_net_time_embed_1_mlp_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1087040)))]; + tensor linear_4_cast_fp16 = linear(bias = flow_net_time_embed_1_mlp_2_bias_to_fp16, weight = flow_net_time_embed_1_mlp_2_weight_to_fp16, x = input_11_cast_fp16)[name = tensor("linear_4_cast_fp16")]; + tensor reduce_mean_2_axes_0 = const()[name = tensor("reduce_mean_2_axes_0"), val = tensor([-1])]; + tensor reduce_mean_2_keep_dims_0 = const()[name = tensor("reduce_mean_2_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_2_cast_fp16 = reduce_mean(axes = reduce_mean_2_axes_0, keep_dims = reduce_mean_2_keep_dims_0, x = linear_4_cast_fp16)[name = tensor("reduce_mean_2_cast_fp16")]; + tensor sub_2_cast_fp16 = sub(x = linear_4_cast_fp16, y = reduce_mean_2_cast_fp16)[name = tensor("sub_2_cast_fp16")]; + tensor square_1_cast_fp16 = square(x = sub_2_cast_fp16)[name = tensor("square_1_cast_fp16")]; + tensor reduce_mean_3_axes_0 = const()[name = tensor("reduce_mean_3_axes_0"), val = tensor([-1])]; + tensor reduce_mean_3_keep_dims_0 = const()[name = tensor("reduce_mean_3_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_3_cast_fp16 = reduce_mean(axes = reduce_mean_3_axes_0, keep_dims = reduce_mean_3_keep_dims_0, x = square_1_cast_fp16)[name = tensor("reduce_mean_3_cast_fp16")]; + tensor real_div_1_to_fp16 = const()[name = tensor("real_div_1_to_fp16"), val = tensor(0x1.008p+0)]; + tensor mul_1_cast_fp16 = mul(x = reduce_mean_3_cast_fp16, y = real_div_1_to_fp16)[name = tensor("mul_1_cast_fp16")]; + tensor var_110_to_fp16 = const()[name = tensor("op_110_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_3_cast_fp16 = add(x = mul_1_cast_fp16, y = var_110_to_fp16)[name = tensor("var_3_cast_fp16")]; + tensor var_113_epsilon_0 = const()[name = tensor("op_113_epsilon_0"), val = tensor(0x1.197998p-40)]; + tensor var_113_cast_fp16 = rsqrt(epsilon = var_113_epsilon_0, x = var_3_cast_fp16)[name = tensor("op_113_cast_fp16")]; + tensor const_5_to_fp16 = const()[name = tensor("const_5_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1088128)))]; + tensor var_114_cast_fp16 = mul(x = const_5_to_fp16, y = var_113_cast_fp16)[name = tensor("op_114_cast_fp16")]; + tensor var_115_cast_fp16 = mul(x = linear_4_cast_fp16, y = var_114_cast_fp16)[name = tensor("op_115_cast_fp16")]; + tensor var_127_cast_fp16 = add(x = var_75_cast_fp16, y = var_115_cast_fp16)[name = tensor("op_127_cast_fp16")]; + tensor _inversed_t_combined_1_y_0_to_fp16 = const()[name = tensor("_inversed_t_combined_1_y_0_to_fp16"), val = tensor(0x1p-1)]; + tensor _inversed_t_combined_1_cast_fp16 = mul(x = var_127_cast_fp16, y = _inversed_t_combined_1_y_0_to_fp16)[name = tensor("_inversed_t_combined_1_cast_fp16")]; + tensor transformer_out_to_fp16_dtype_0 = const()[name = tensor("transformer_out_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor flow_net_cond_embed_weight_to_fp16 = const()[name = tensor("flow_net_cond_embed_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1089216)))]; + tensor flow_net_cond_embed_bias_to_fp16 = const()[name = tensor("flow_net_cond_embed_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(2137856)))]; + tensor transformer_out_to_fp16 = cast(dtype = transformer_out_to_fp16_dtype_0, x = transformer_out)[name = tensor("cast_145")]; + tensor linear_5_cast_fp16 = linear(bias = flow_net_cond_embed_bias_to_fp16, weight = flow_net_cond_embed_weight_to_fp16, x = transformer_out_to_fp16)[name = tensor("linear_5_cast_fp16")]; + tensor input_13_cast_fp16 = add(x = _inversed_t_combined_1_cast_fp16, y = linear_5_cast_fp16)[name = tensor("input_13_cast_fp16")]; + tensor input_15_cast_fp16 = silu(x = input_13_cast_fp16)[name = tensor("input_15_cast_fp16")]; + tensor flow_net_res_blocks_0_adaLN_modulation_1_weight_to_fp16 = const()[name = tensor("flow_net_res_blocks_0_adaLN_modulation_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(2138944)))]; + tensor flow_net_res_blocks_0_adaLN_modulation_1_bias_to_fp16 = const()[name = tensor("flow_net_res_blocks_0_adaLN_modulation_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(3711872)))]; + tensor linear_6_cast_fp16 = linear(bias = flow_net_res_blocks_0_adaLN_modulation_1_bias_to_fp16, weight = flow_net_res_blocks_0_adaLN_modulation_1_weight_to_fp16, x = input_15_cast_fp16)[name = tensor("linear_6_cast_fp16")]; + tensor var_142_split_sizes_0 = const()[name = tensor("op_142_split_sizes_0"), val = tensor([512, 512, 512])]; + tensor var_142_axis_0 = const()[name = tensor("op_142_axis_0"), val = tensor(-1)]; + tensor var_142_cast_fp16_0, tensor var_142_cast_fp16_1, tensor var_142_cast_fp16_2 = split(axis = var_142_axis_0, split_sizes = var_142_split_sizes_0, x = linear_6_cast_fp16)[name = tensor("op_142_cast_fp16")]; + tensor mean_1_axes_0 = const()[name = tensor("mean_1_axes_0"), val = tensor([-1])]; + tensor mean_1_keep_dims_0 = const()[name = tensor("mean_1_keep_dims_0"), val = tensor(true)]; + tensor mean_1_cast_fp16 = reduce_mean(axes = mean_1_axes_0, keep_dims = mean_1_keep_dims_0, x = linear_0_cast_fp16)[name = tensor("mean_1_cast_fp16")]; + tensor sub_4_cast_fp16 = sub(x = linear_0_cast_fp16, y = mean_1_cast_fp16)[name = tensor("sub_4_cast_fp16")]; + tensor square_2_cast_fp16 = square(x = sub_4_cast_fp16)[name = tensor("square_2_cast_fp16")]; + tensor reduce_mean_5_axes_0 = const()[name = tensor("reduce_mean_5_axes_0"), val = tensor([-1])]; + tensor reduce_mean_5_keep_dims_0 = const()[name = tensor("reduce_mean_5_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_5_cast_fp16 = reduce_mean(axes = reduce_mean_5_axes_0, keep_dims = reduce_mean_5_keep_dims_0, x = square_2_cast_fp16)[name = tensor("reduce_mean_5_cast_fp16")]; + tensor var_152_to_fp16 = const()[name = tensor("op_152_to_fp16"), val = tensor(0x1.1p-20)]; + tensor var_153_cast_fp16 = add(x = reduce_mean_5_cast_fp16, y = var_152_to_fp16)[name = tensor("op_153_cast_fp16")]; + tensor var_154_cast_fp16 = sqrt(x = var_153_cast_fp16)[name = tensor("op_154_cast_fp16")]; + tensor x_7_cast_fp16 = real_div(x = sub_4_cast_fp16, y = var_154_cast_fp16)[name = tensor("x_7_cast_fp16")]; + tensor flow_net_res_blocks_0_in_ln_weight_to_fp16 = const()[name = tensor("flow_net_res_blocks_0_in_ln_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(3715008)))]; + tensor var_156_cast_fp16 = mul(x = x_7_cast_fp16, y = flow_net_res_blocks_0_in_ln_weight_to_fp16)[name = tensor("op_156_cast_fp16")]; + tensor flow_net_res_blocks_0_in_ln_bias_to_fp16 = const()[name = tensor("flow_net_res_blocks_0_in_ln_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(3716096)))]; + tensor x_9_cast_fp16 = add(x = var_156_cast_fp16, y = flow_net_res_blocks_0_in_ln_bias_to_fp16)[name = tensor("x_9_cast_fp16")]; + tensor var_158_promoted_to_fp16 = const()[name = tensor("op_158_promoted_to_fp16"), val = tensor(0x1p+0)]; + tensor var_159_cast_fp16 = add(x = var_142_cast_fp16_1, y = var_158_promoted_to_fp16)[name = tensor("op_159_cast_fp16")]; + tensor var_160_cast_fp16 = mul(x = x_9_cast_fp16, y = var_159_cast_fp16)[name = tensor("op_160_cast_fp16")]; + tensor input_17_cast_fp16 = add(x = var_160_cast_fp16, y = var_142_cast_fp16_0)[name = tensor("input_17_cast_fp16")]; + tensor flow_net_res_blocks_0_mlp_0_weight_to_fp16 = const()[name = tensor("flow_net_res_blocks_0_mlp_0_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(3717184)))]; + tensor flow_net_res_blocks_0_mlp_0_bias_to_fp16 = const()[name = tensor("flow_net_res_blocks_0_mlp_0_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(4241536)))]; + tensor linear_7_cast_fp16 = linear(bias = flow_net_res_blocks_0_mlp_0_bias_to_fp16, weight = flow_net_res_blocks_0_mlp_0_weight_to_fp16, x = input_17_cast_fp16)[name = tensor("linear_7_cast_fp16")]; + tensor input_21_cast_fp16 = silu(x = linear_7_cast_fp16)[name = tensor("input_21_cast_fp16")]; + tensor flow_net_res_blocks_0_mlp_2_weight_to_fp16 = const()[name = tensor("flow_net_res_blocks_0_mlp_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(4242624)))]; + tensor flow_net_res_blocks_0_mlp_2_bias_to_fp16 = const()[name = tensor("flow_net_res_blocks_0_mlp_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(4766976)))]; + tensor linear_8_cast_fp16 = linear(bias = flow_net_res_blocks_0_mlp_2_bias_to_fp16, weight = flow_net_res_blocks_0_mlp_2_weight_to_fp16, x = input_21_cast_fp16)[name = tensor("linear_8_cast_fp16")]; + tensor var_171_cast_fp16 = mul(x = var_142_cast_fp16_2, y = linear_8_cast_fp16)[name = tensor("op_171_cast_fp16")]; + tensor x_11_cast_fp16 = add(x = linear_0_cast_fp16, y = var_171_cast_fp16)[name = tensor("x_11_cast_fp16")]; + tensor flow_net_res_blocks_1_adaLN_modulation_1_weight_to_fp16 = const()[name = tensor("flow_net_res_blocks_1_adaLN_modulation_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(4768064)))]; + tensor flow_net_res_blocks_1_adaLN_modulation_1_bias_to_fp16 = const()[name = tensor("flow_net_res_blocks_1_adaLN_modulation_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(6340992)))]; + tensor linear_9_cast_fp16 = linear(bias = flow_net_res_blocks_1_adaLN_modulation_1_bias_to_fp16, weight = flow_net_res_blocks_1_adaLN_modulation_1_weight_to_fp16, x = input_15_cast_fp16)[name = tensor("linear_9_cast_fp16")]; + tensor var_181_split_sizes_0 = const()[name = tensor("op_181_split_sizes_0"), val = tensor([512, 512, 512])]; + tensor var_181_axis_0 = const()[name = tensor("op_181_axis_0"), val = tensor(-1)]; + tensor var_181_cast_fp16_0, tensor var_181_cast_fp16_1, tensor var_181_cast_fp16_2 = split(axis = var_181_axis_0, split_sizes = var_181_split_sizes_0, x = linear_9_cast_fp16)[name = tensor("op_181_cast_fp16")]; + tensor mean_3_axes_0 = const()[name = tensor("mean_3_axes_0"), val = tensor([-1])]; + tensor mean_3_keep_dims_0 = const()[name = tensor("mean_3_keep_dims_0"), val = tensor(true)]; + tensor mean_3_cast_fp16 = reduce_mean(axes = mean_3_axes_0, keep_dims = mean_3_keep_dims_0, x = x_11_cast_fp16)[name = tensor("mean_3_cast_fp16")]; + tensor sub_5_cast_fp16 = sub(x = x_11_cast_fp16, y = mean_3_cast_fp16)[name = tensor("sub_5_cast_fp16")]; + tensor square_3_cast_fp16 = square(x = sub_5_cast_fp16)[name = tensor("square_3_cast_fp16")]; + tensor reduce_mean_7_axes_0 = const()[name = tensor("reduce_mean_7_axes_0"), val = tensor([-1])]; + tensor reduce_mean_7_keep_dims_0 = const()[name = tensor("reduce_mean_7_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_7_cast_fp16 = reduce_mean(axes = reduce_mean_7_axes_0, keep_dims = reduce_mean_7_keep_dims_0, x = square_3_cast_fp16)[name = tensor("reduce_mean_7_cast_fp16")]; + tensor var_191_to_fp16 = const()[name = tensor("op_191_to_fp16"), val = tensor(0x1.1p-20)]; + tensor var_192_cast_fp16 = add(x = reduce_mean_7_cast_fp16, y = var_191_to_fp16)[name = tensor("op_192_cast_fp16")]; + tensor var_193_cast_fp16 = sqrt(x = var_192_cast_fp16)[name = tensor("op_193_cast_fp16")]; + tensor x_13_cast_fp16 = real_div(x = sub_5_cast_fp16, y = var_193_cast_fp16)[name = tensor("x_13_cast_fp16")]; + tensor flow_net_res_blocks_1_in_ln_weight_to_fp16 = const()[name = tensor("flow_net_res_blocks_1_in_ln_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(6344128)))]; + tensor var_195_cast_fp16 = mul(x = x_13_cast_fp16, y = flow_net_res_blocks_1_in_ln_weight_to_fp16)[name = tensor("op_195_cast_fp16")]; + tensor flow_net_res_blocks_1_in_ln_bias_to_fp16 = const()[name = tensor("flow_net_res_blocks_1_in_ln_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(6345216)))]; + tensor x_15_cast_fp16 = add(x = var_195_cast_fp16, y = flow_net_res_blocks_1_in_ln_bias_to_fp16)[name = tensor("x_15_cast_fp16")]; + tensor var_197_promoted_to_fp16 = const()[name = tensor("op_197_promoted_to_fp16"), val = tensor(0x1p+0)]; + tensor var_198_cast_fp16 = add(x = var_181_cast_fp16_1, y = var_197_promoted_to_fp16)[name = tensor("op_198_cast_fp16")]; + tensor var_199_cast_fp16 = mul(x = x_15_cast_fp16, y = var_198_cast_fp16)[name = tensor("op_199_cast_fp16")]; + tensor input_25_cast_fp16 = add(x = var_199_cast_fp16, y = var_181_cast_fp16_0)[name = tensor("input_25_cast_fp16")]; + tensor flow_net_res_blocks_1_mlp_0_weight_to_fp16 = const()[name = tensor("flow_net_res_blocks_1_mlp_0_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(6346304)))]; + tensor flow_net_res_blocks_1_mlp_0_bias_to_fp16 = const()[name = tensor("flow_net_res_blocks_1_mlp_0_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(6870656)))]; + tensor linear_10_cast_fp16 = linear(bias = flow_net_res_blocks_1_mlp_0_bias_to_fp16, weight = flow_net_res_blocks_1_mlp_0_weight_to_fp16, x = input_25_cast_fp16)[name = tensor("linear_10_cast_fp16")]; + tensor input_29_cast_fp16 = silu(x = linear_10_cast_fp16)[name = tensor("input_29_cast_fp16")]; + tensor flow_net_res_blocks_1_mlp_2_weight_to_fp16 = const()[name = tensor("flow_net_res_blocks_1_mlp_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(6871744)))]; + tensor flow_net_res_blocks_1_mlp_2_bias_to_fp16 = const()[name = tensor("flow_net_res_blocks_1_mlp_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(7396096)))]; + tensor linear_11_cast_fp16 = linear(bias = flow_net_res_blocks_1_mlp_2_bias_to_fp16, weight = flow_net_res_blocks_1_mlp_2_weight_to_fp16, x = input_29_cast_fp16)[name = tensor("linear_11_cast_fp16")]; + tensor var_210_cast_fp16 = mul(x = var_181_cast_fp16_2, y = linear_11_cast_fp16)[name = tensor("op_210_cast_fp16")]; + tensor x_17_cast_fp16 = add(x = x_11_cast_fp16, y = var_210_cast_fp16)[name = tensor("x_17_cast_fp16")]; + tensor flow_net_res_blocks_2_adaLN_modulation_1_weight_to_fp16 = const()[name = tensor("flow_net_res_blocks_2_adaLN_modulation_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(7397184)))]; + tensor flow_net_res_blocks_2_adaLN_modulation_1_bias_to_fp16 = const()[name = tensor("flow_net_res_blocks_2_adaLN_modulation_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8970112)))]; + tensor linear_12_cast_fp16 = linear(bias = flow_net_res_blocks_2_adaLN_modulation_1_bias_to_fp16, weight = flow_net_res_blocks_2_adaLN_modulation_1_weight_to_fp16, x = input_15_cast_fp16)[name = tensor("linear_12_cast_fp16")]; + tensor var_220_split_sizes_0 = const()[name = tensor("op_220_split_sizes_0"), val = tensor([512, 512, 512])]; + tensor var_220_axis_0 = const()[name = tensor("op_220_axis_0"), val = tensor(-1)]; + tensor var_220_cast_fp16_0, tensor var_220_cast_fp16_1, tensor var_220_cast_fp16_2 = split(axis = var_220_axis_0, split_sizes = var_220_split_sizes_0, x = linear_12_cast_fp16)[name = tensor("op_220_cast_fp16")]; + tensor mean_5_axes_0 = const()[name = tensor("mean_5_axes_0"), val = tensor([-1])]; + tensor mean_5_keep_dims_0 = const()[name = tensor("mean_5_keep_dims_0"), val = tensor(true)]; + tensor mean_5_cast_fp16 = reduce_mean(axes = mean_5_axes_0, keep_dims = mean_5_keep_dims_0, x = x_17_cast_fp16)[name = tensor("mean_5_cast_fp16")]; + tensor sub_6_cast_fp16 = sub(x = x_17_cast_fp16, y = mean_5_cast_fp16)[name = tensor("sub_6_cast_fp16")]; + tensor square_4_cast_fp16 = square(x = sub_6_cast_fp16)[name = tensor("square_4_cast_fp16")]; + tensor reduce_mean_9_axes_0 = const()[name = tensor("reduce_mean_9_axes_0"), val = tensor([-1])]; + tensor reduce_mean_9_keep_dims_0 = const()[name = tensor("reduce_mean_9_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_9_cast_fp16 = reduce_mean(axes = reduce_mean_9_axes_0, keep_dims = reduce_mean_9_keep_dims_0, x = square_4_cast_fp16)[name = tensor("reduce_mean_9_cast_fp16")]; + tensor var_230_to_fp16 = const()[name = tensor("op_230_to_fp16"), val = tensor(0x1.1p-20)]; + tensor var_231_cast_fp16 = add(x = reduce_mean_9_cast_fp16, y = var_230_to_fp16)[name = tensor("op_231_cast_fp16")]; + tensor var_232_cast_fp16 = sqrt(x = var_231_cast_fp16)[name = tensor("op_232_cast_fp16")]; + tensor x_19_cast_fp16 = real_div(x = sub_6_cast_fp16, y = var_232_cast_fp16)[name = tensor("x_19_cast_fp16")]; + tensor flow_net_res_blocks_2_in_ln_weight_to_fp16 = const()[name = tensor("flow_net_res_blocks_2_in_ln_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8973248)))]; + tensor var_234_cast_fp16 = mul(x = x_19_cast_fp16, y = flow_net_res_blocks_2_in_ln_weight_to_fp16)[name = tensor("op_234_cast_fp16")]; + tensor flow_net_res_blocks_2_in_ln_bias_to_fp16 = const()[name = tensor("flow_net_res_blocks_2_in_ln_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8974336)))]; + tensor x_21_cast_fp16 = add(x = var_234_cast_fp16, y = flow_net_res_blocks_2_in_ln_bias_to_fp16)[name = tensor("x_21_cast_fp16")]; + tensor var_236_promoted_to_fp16 = const()[name = tensor("op_236_promoted_to_fp16"), val = tensor(0x1p+0)]; + tensor var_237_cast_fp16 = add(x = var_220_cast_fp16_1, y = var_236_promoted_to_fp16)[name = tensor("op_237_cast_fp16")]; + tensor var_238_cast_fp16 = mul(x = x_21_cast_fp16, y = var_237_cast_fp16)[name = tensor("op_238_cast_fp16")]; + tensor input_33_cast_fp16 = add(x = var_238_cast_fp16, y = var_220_cast_fp16_0)[name = tensor("input_33_cast_fp16")]; + tensor flow_net_res_blocks_2_mlp_0_weight_to_fp16 = const()[name = tensor("flow_net_res_blocks_2_mlp_0_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8975424)))]; + tensor flow_net_res_blocks_2_mlp_0_bias_to_fp16 = const()[name = tensor("flow_net_res_blocks_2_mlp_0_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9499776)))]; + tensor linear_13_cast_fp16 = linear(bias = flow_net_res_blocks_2_mlp_0_bias_to_fp16, weight = flow_net_res_blocks_2_mlp_0_weight_to_fp16, x = input_33_cast_fp16)[name = tensor("linear_13_cast_fp16")]; + tensor input_37_cast_fp16 = silu(x = linear_13_cast_fp16)[name = tensor("input_37_cast_fp16")]; + tensor flow_net_res_blocks_2_mlp_2_weight_to_fp16 = const()[name = tensor("flow_net_res_blocks_2_mlp_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9500864)))]; + tensor flow_net_res_blocks_2_mlp_2_bias_to_fp16 = const()[name = tensor("flow_net_res_blocks_2_mlp_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(10025216)))]; + tensor linear_14_cast_fp16 = linear(bias = flow_net_res_blocks_2_mlp_2_bias_to_fp16, weight = flow_net_res_blocks_2_mlp_2_weight_to_fp16, x = input_37_cast_fp16)[name = tensor("linear_14_cast_fp16")]; + tensor var_249_cast_fp16 = mul(x = var_220_cast_fp16_2, y = linear_14_cast_fp16)[name = tensor("op_249_cast_fp16")]; + tensor x_23_cast_fp16 = add(x = x_17_cast_fp16, y = var_249_cast_fp16)[name = tensor("x_23_cast_fp16")]; + tensor flow_net_res_blocks_3_adaLN_modulation_1_weight_to_fp16 = const()[name = tensor("flow_net_res_blocks_3_adaLN_modulation_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(10026304)))]; + tensor flow_net_res_blocks_3_adaLN_modulation_1_bias_to_fp16 = const()[name = tensor("flow_net_res_blocks_3_adaLN_modulation_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(11599232)))]; + tensor linear_15_cast_fp16 = linear(bias = flow_net_res_blocks_3_adaLN_modulation_1_bias_to_fp16, weight = flow_net_res_blocks_3_adaLN_modulation_1_weight_to_fp16, x = input_15_cast_fp16)[name = tensor("linear_15_cast_fp16")]; + tensor var_259_split_sizes_0 = const()[name = tensor("op_259_split_sizes_0"), val = tensor([512, 512, 512])]; + tensor var_259_axis_0 = const()[name = tensor("op_259_axis_0"), val = tensor(-1)]; + tensor var_259_cast_fp16_0, tensor var_259_cast_fp16_1, tensor var_259_cast_fp16_2 = split(axis = var_259_axis_0, split_sizes = var_259_split_sizes_0, x = linear_15_cast_fp16)[name = tensor("op_259_cast_fp16")]; + tensor mean_7_axes_0 = const()[name = tensor("mean_7_axes_0"), val = tensor([-1])]; + tensor mean_7_keep_dims_0 = const()[name = tensor("mean_7_keep_dims_0"), val = tensor(true)]; + tensor mean_7_cast_fp16 = reduce_mean(axes = mean_7_axes_0, keep_dims = mean_7_keep_dims_0, x = x_23_cast_fp16)[name = tensor("mean_7_cast_fp16")]; + tensor sub_7_cast_fp16 = sub(x = x_23_cast_fp16, y = mean_7_cast_fp16)[name = tensor("sub_7_cast_fp16")]; + tensor square_5_cast_fp16 = square(x = sub_7_cast_fp16)[name = tensor("square_5_cast_fp16")]; + tensor reduce_mean_11_axes_0 = const()[name = tensor("reduce_mean_11_axes_0"), val = tensor([-1])]; + tensor reduce_mean_11_keep_dims_0 = const()[name = tensor("reduce_mean_11_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_11_cast_fp16 = reduce_mean(axes = reduce_mean_11_axes_0, keep_dims = reduce_mean_11_keep_dims_0, x = square_5_cast_fp16)[name = tensor("reduce_mean_11_cast_fp16")]; + tensor var_269_to_fp16 = const()[name = tensor("op_269_to_fp16"), val = tensor(0x1.1p-20)]; + tensor var_270_cast_fp16 = add(x = reduce_mean_11_cast_fp16, y = var_269_to_fp16)[name = tensor("op_270_cast_fp16")]; + tensor var_271_cast_fp16 = sqrt(x = var_270_cast_fp16)[name = tensor("op_271_cast_fp16")]; + tensor x_25_cast_fp16 = real_div(x = sub_7_cast_fp16, y = var_271_cast_fp16)[name = tensor("x_25_cast_fp16")]; + tensor flow_net_res_blocks_3_in_ln_weight_to_fp16 = const()[name = tensor("flow_net_res_blocks_3_in_ln_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(11602368)))]; + tensor var_273_cast_fp16 = mul(x = x_25_cast_fp16, y = flow_net_res_blocks_3_in_ln_weight_to_fp16)[name = tensor("op_273_cast_fp16")]; + tensor flow_net_res_blocks_3_in_ln_bias_to_fp16 = const()[name = tensor("flow_net_res_blocks_3_in_ln_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(11603456)))]; + tensor x_27_cast_fp16 = add(x = var_273_cast_fp16, y = flow_net_res_blocks_3_in_ln_bias_to_fp16)[name = tensor("x_27_cast_fp16")]; + tensor var_275_promoted_to_fp16 = const()[name = tensor("op_275_promoted_to_fp16"), val = tensor(0x1p+0)]; + tensor var_276_cast_fp16 = add(x = var_259_cast_fp16_1, y = var_275_promoted_to_fp16)[name = tensor("op_276_cast_fp16")]; + tensor var_277_cast_fp16 = mul(x = x_27_cast_fp16, y = var_276_cast_fp16)[name = tensor("op_277_cast_fp16")]; + tensor input_41_cast_fp16 = add(x = var_277_cast_fp16, y = var_259_cast_fp16_0)[name = tensor("input_41_cast_fp16")]; + tensor flow_net_res_blocks_3_mlp_0_weight_to_fp16 = const()[name = tensor("flow_net_res_blocks_3_mlp_0_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(11604544)))]; + tensor flow_net_res_blocks_3_mlp_0_bias_to_fp16 = const()[name = tensor("flow_net_res_blocks_3_mlp_0_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(12128896)))]; + tensor linear_16_cast_fp16 = linear(bias = flow_net_res_blocks_3_mlp_0_bias_to_fp16, weight = flow_net_res_blocks_3_mlp_0_weight_to_fp16, x = input_41_cast_fp16)[name = tensor("linear_16_cast_fp16")]; + tensor input_45_cast_fp16 = silu(x = linear_16_cast_fp16)[name = tensor("input_45_cast_fp16")]; + tensor flow_net_res_blocks_3_mlp_2_weight_to_fp16 = const()[name = tensor("flow_net_res_blocks_3_mlp_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(12129984)))]; + tensor flow_net_res_blocks_3_mlp_2_bias_to_fp16 = const()[name = tensor("flow_net_res_blocks_3_mlp_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(12654336)))]; + tensor linear_17_cast_fp16 = linear(bias = flow_net_res_blocks_3_mlp_2_bias_to_fp16, weight = flow_net_res_blocks_3_mlp_2_weight_to_fp16, x = input_45_cast_fp16)[name = tensor("linear_17_cast_fp16")]; + tensor var_288_cast_fp16 = mul(x = var_259_cast_fp16_2, y = linear_17_cast_fp16)[name = tensor("op_288_cast_fp16")]; + tensor x_29_cast_fp16 = add(x = x_23_cast_fp16, y = var_288_cast_fp16)[name = tensor("x_29_cast_fp16")]; + tensor flow_net_res_blocks_4_adaLN_modulation_1_weight_to_fp16 = const()[name = tensor("flow_net_res_blocks_4_adaLN_modulation_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(12655424)))]; + tensor flow_net_res_blocks_4_adaLN_modulation_1_bias_to_fp16 = const()[name = tensor("flow_net_res_blocks_4_adaLN_modulation_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(14228352)))]; + tensor linear_18_cast_fp16 = linear(bias = flow_net_res_blocks_4_adaLN_modulation_1_bias_to_fp16, weight = flow_net_res_blocks_4_adaLN_modulation_1_weight_to_fp16, x = input_15_cast_fp16)[name = tensor("linear_18_cast_fp16")]; + tensor var_298_split_sizes_0 = const()[name = tensor("op_298_split_sizes_0"), val = tensor([512, 512, 512])]; + tensor var_298_axis_0 = const()[name = tensor("op_298_axis_0"), val = tensor(-1)]; + tensor var_298_cast_fp16_0, tensor var_298_cast_fp16_1, tensor var_298_cast_fp16_2 = split(axis = var_298_axis_0, split_sizes = var_298_split_sizes_0, x = linear_18_cast_fp16)[name = tensor("op_298_cast_fp16")]; + tensor mean_9_axes_0 = const()[name = tensor("mean_9_axes_0"), val = tensor([-1])]; + tensor mean_9_keep_dims_0 = const()[name = tensor("mean_9_keep_dims_0"), val = tensor(true)]; + tensor mean_9_cast_fp16 = reduce_mean(axes = mean_9_axes_0, keep_dims = mean_9_keep_dims_0, x = x_29_cast_fp16)[name = tensor("mean_9_cast_fp16")]; + tensor sub_8_cast_fp16 = sub(x = x_29_cast_fp16, y = mean_9_cast_fp16)[name = tensor("sub_8_cast_fp16")]; + tensor square_6_cast_fp16 = square(x = sub_8_cast_fp16)[name = tensor("square_6_cast_fp16")]; + tensor reduce_mean_13_axes_0 = const()[name = tensor("reduce_mean_13_axes_0"), val = tensor([-1])]; + tensor reduce_mean_13_keep_dims_0 = const()[name = tensor("reduce_mean_13_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_13_cast_fp16 = reduce_mean(axes = reduce_mean_13_axes_0, keep_dims = reduce_mean_13_keep_dims_0, x = square_6_cast_fp16)[name = tensor("reduce_mean_13_cast_fp16")]; + tensor var_308_to_fp16 = const()[name = tensor("op_308_to_fp16"), val = tensor(0x1.1p-20)]; + tensor var_309_cast_fp16 = add(x = reduce_mean_13_cast_fp16, y = var_308_to_fp16)[name = tensor("op_309_cast_fp16")]; + tensor var_310_cast_fp16 = sqrt(x = var_309_cast_fp16)[name = tensor("op_310_cast_fp16")]; + tensor x_31_cast_fp16 = real_div(x = sub_8_cast_fp16, y = var_310_cast_fp16)[name = tensor("x_31_cast_fp16")]; + tensor flow_net_res_blocks_4_in_ln_weight_to_fp16 = const()[name = tensor("flow_net_res_blocks_4_in_ln_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(14231488)))]; + tensor var_312_cast_fp16 = mul(x = x_31_cast_fp16, y = flow_net_res_blocks_4_in_ln_weight_to_fp16)[name = tensor("op_312_cast_fp16")]; + tensor flow_net_res_blocks_4_in_ln_bias_to_fp16 = const()[name = tensor("flow_net_res_blocks_4_in_ln_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(14232576)))]; + tensor x_33_cast_fp16 = add(x = var_312_cast_fp16, y = flow_net_res_blocks_4_in_ln_bias_to_fp16)[name = tensor("x_33_cast_fp16")]; + tensor var_314_promoted_to_fp16 = const()[name = tensor("op_314_promoted_to_fp16"), val = tensor(0x1p+0)]; + tensor var_315_cast_fp16 = add(x = var_298_cast_fp16_1, y = var_314_promoted_to_fp16)[name = tensor("op_315_cast_fp16")]; + tensor var_316_cast_fp16 = mul(x = x_33_cast_fp16, y = var_315_cast_fp16)[name = tensor("op_316_cast_fp16")]; + tensor input_49_cast_fp16 = add(x = var_316_cast_fp16, y = var_298_cast_fp16_0)[name = tensor("input_49_cast_fp16")]; + tensor flow_net_res_blocks_4_mlp_0_weight_to_fp16 = const()[name = tensor("flow_net_res_blocks_4_mlp_0_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(14233664)))]; + tensor flow_net_res_blocks_4_mlp_0_bias_to_fp16 = const()[name = tensor("flow_net_res_blocks_4_mlp_0_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(14758016)))]; + tensor linear_19_cast_fp16 = linear(bias = flow_net_res_blocks_4_mlp_0_bias_to_fp16, weight = flow_net_res_blocks_4_mlp_0_weight_to_fp16, x = input_49_cast_fp16)[name = tensor("linear_19_cast_fp16")]; + tensor input_53_cast_fp16 = silu(x = linear_19_cast_fp16)[name = tensor("input_53_cast_fp16")]; + tensor flow_net_res_blocks_4_mlp_2_weight_to_fp16 = const()[name = tensor("flow_net_res_blocks_4_mlp_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(14759104)))]; + tensor flow_net_res_blocks_4_mlp_2_bias_to_fp16 = const()[name = tensor("flow_net_res_blocks_4_mlp_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15283456)))]; + tensor linear_20_cast_fp16 = linear(bias = flow_net_res_blocks_4_mlp_2_bias_to_fp16, weight = flow_net_res_blocks_4_mlp_2_weight_to_fp16, x = input_53_cast_fp16)[name = tensor("linear_20_cast_fp16")]; + tensor var_327_cast_fp16 = mul(x = var_298_cast_fp16_2, y = linear_20_cast_fp16)[name = tensor("op_327_cast_fp16")]; + tensor x_35_cast_fp16 = add(x = x_29_cast_fp16, y = var_327_cast_fp16)[name = tensor("x_35_cast_fp16")]; + tensor flow_net_res_blocks_5_adaLN_modulation_1_weight_to_fp16 = const()[name = tensor("flow_net_res_blocks_5_adaLN_modulation_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15284544)))]; + tensor flow_net_res_blocks_5_adaLN_modulation_1_bias_to_fp16 = const()[name = tensor("flow_net_res_blocks_5_adaLN_modulation_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(16857472)))]; + tensor linear_21_cast_fp16 = linear(bias = flow_net_res_blocks_5_adaLN_modulation_1_bias_to_fp16, weight = flow_net_res_blocks_5_adaLN_modulation_1_weight_to_fp16, x = input_15_cast_fp16)[name = tensor("linear_21_cast_fp16")]; + tensor var_337_split_sizes_0 = const()[name = tensor("op_337_split_sizes_0"), val = tensor([512, 512, 512])]; + tensor var_337_axis_0 = const()[name = tensor("op_337_axis_0"), val = tensor(-1)]; + tensor var_337_cast_fp16_0, tensor var_337_cast_fp16_1, tensor var_337_cast_fp16_2 = split(axis = var_337_axis_0, split_sizes = var_337_split_sizes_0, x = linear_21_cast_fp16)[name = tensor("op_337_cast_fp16")]; + tensor mean_11_axes_0 = const()[name = tensor("mean_11_axes_0"), val = tensor([-1])]; + tensor mean_11_keep_dims_0 = const()[name = tensor("mean_11_keep_dims_0"), val = tensor(true)]; + tensor mean_11_cast_fp16 = reduce_mean(axes = mean_11_axes_0, keep_dims = mean_11_keep_dims_0, x = x_35_cast_fp16)[name = tensor("mean_11_cast_fp16")]; + tensor sub_9_cast_fp16 = sub(x = x_35_cast_fp16, y = mean_11_cast_fp16)[name = tensor("sub_9_cast_fp16")]; + tensor square_7_cast_fp16 = square(x = sub_9_cast_fp16)[name = tensor("square_7_cast_fp16")]; + tensor reduce_mean_15_axes_0 = const()[name = tensor("reduce_mean_15_axes_0"), val = tensor([-1])]; + tensor reduce_mean_15_keep_dims_0 = const()[name = tensor("reduce_mean_15_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_15_cast_fp16 = reduce_mean(axes = reduce_mean_15_axes_0, keep_dims = reduce_mean_15_keep_dims_0, x = square_7_cast_fp16)[name = tensor("reduce_mean_15_cast_fp16")]; + tensor var_347_to_fp16 = const()[name = tensor("op_347_to_fp16"), val = tensor(0x1.1p-20)]; + tensor var_348_cast_fp16 = add(x = reduce_mean_15_cast_fp16, y = var_347_to_fp16)[name = tensor("op_348_cast_fp16")]; + tensor var_349_cast_fp16 = sqrt(x = var_348_cast_fp16)[name = tensor("op_349_cast_fp16")]; + tensor x_37_cast_fp16 = real_div(x = sub_9_cast_fp16, y = var_349_cast_fp16)[name = tensor("x_37_cast_fp16")]; + tensor flow_net_res_blocks_5_in_ln_weight_to_fp16 = const()[name = tensor("flow_net_res_blocks_5_in_ln_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(16860608)))]; + tensor var_351_cast_fp16 = mul(x = x_37_cast_fp16, y = flow_net_res_blocks_5_in_ln_weight_to_fp16)[name = tensor("op_351_cast_fp16")]; + tensor flow_net_res_blocks_5_in_ln_bias_to_fp16 = const()[name = tensor("flow_net_res_blocks_5_in_ln_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(16861696)))]; + tensor x_39_cast_fp16 = add(x = var_351_cast_fp16, y = flow_net_res_blocks_5_in_ln_bias_to_fp16)[name = tensor("x_39_cast_fp16")]; + tensor var_353_promoted_to_fp16 = const()[name = tensor("op_353_promoted_to_fp16"), val = tensor(0x1p+0)]; + tensor var_354_cast_fp16 = add(x = var_337_cast_fp16_1, y = var_353_promoted_to_fp16)[name = tensor("op_354_cast_fp16")]; + tensor var_355_cast_fp16 = mul(x = x_39_cast_fp16, y = var_354_cast_fp16)[name = tensor("op_355_cast_fp16")]; + tensor input_57_cast_fp16 = add(x = var_355_cast_fp16, y = var_337_cast_fp16_0)[name = tensor("input_57_cast_fp16")]; + tensor flow_net_res_blocks_5_mlp_0_weight_to_fp16 = const()[name = tensor("flow_net_res_blocks_5_mlp_0_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(16862784)))]; + tensor flow_net_res_blocks_5_mlp_0_bias_to_fp16 = const()[name = tensor("flow_net_res_blocks_5_mlp_0_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(17387136)))]; + tensor linear_22_cast_fp16 = linear(bias = flow_net_res_blocks_5_mlp_0_bias_to_fp16, weight = flow_net_res_blocks_5_mlp_0_weight_to_fp16, x = input_57_cast_fp16)[name = tensor("linear_22_cast_fp16")]; + tensor input_61_cast_fp16 = silu(x = linear_22_cast_fp16)[name = tensor("input_61_cast_fp16")]; + tensor flow_net_res_blocks_5_mlp_2_weight_to_fp16 = const()[name = tensor("flow_net_res_blocks_5_mlp_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(17388224)))]; + tensor flow_net_res_blocks_5_mlp_2_bias_to_fp16 = const()[name = tensor("flow_net_res_blocks_5_mlp_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(17912576)))]; + tensor linear_23_cast_fp16 = linear(bias = flow_net_res_blocks_5_mlp_2_bias_to_fp16, weight = flow_net_res_blocks_5_mlp_2_weight_to_fp16, x = input_61_cast_fp16)[name = tensor("linear_23_cast_fp16")]; + tensor var_366_cast_fp16 = mul(x = var_337_cast_fp16_2, y = linear_23_cast_fp16)[name = tensor("op_366_cast_fp16")]; + tensor x_41_cast_fp16 = add(x = x_35_cast_fp16, y = var_366_cast_fp16)[name = tensor("x_41_cast_fp16")]; + tensor flow_net_final_layer_adaLN_modulation_1_weight_to_fp16 = const()[name = tensor("flow_net_final_layer_adaLN_modulation_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(17913664)))]; + tensor flow_net_final_layer_adaLN_modulation_1_bias_to_fp16 = const()[name = tensor("flow_net_final_layer_adaLN_modulation_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(18962304)))]; + tensor linear_24_cast_fp16 = linear(bias = flow_net_final_layer_adaLN_modulation_1_bias_to_fp16, weight = flow_net_final_layer_adaLN_modulation_1_weight_to_fp16, x = input_15_cast_fp16)[name = tensor("linear_24_cast_fp16")]; + tensor var_375_split_sizes_0 = const()[name = tensor("op_375_split_sizes_0"), val = tensor([512, 512])]; + tensor var_375_axis_0 = const()[name = tensor("op_375_axis_0"), val = tensor(-1)]; + tensor var_375_cast_fp16_0, tensor var_375_cast_fp16_1 = split(axis = var_375_axis_0, split_sizes = var_375_split_sizes_0, x = linear_24_cast_fp16)[name = tensor("op_375_cast_fp16")]; + tensor mean_13_axes_0 = const()[name = tensor("mean_13_axes_0"), val = tensor([-1])]; + tensor mean_13_keep_dims_0 = const()[name = tensor("mean_13_keep_dims_0"), val = tensor(true)]; + tensor mean_13_cast_fp16 = reduce_mean(axes = mean_13_axes_0, keep_dims = mean_13_keep_dims_0, x = x_41_cast_fp16)[name = tensor("mean_13_cast_fp16")]; + tensor sub_10_cast_fp16 = sub(x = x_41_cast_fp16, y = mean_13_cast_fp16)[name = tensor("sub_10_cast_fp16")]; + tensor square_8_cast_fp16 = square(x = sub_10_cast_fp16)[name = tensor("square_8_cast_fp16")]; + tensor reduce_mean_17_axes_0 = const()[name = tensor("reduce_mean_17_axes_0"), val = tensor([-1])]; + tensor reduce_mean_17_keep_dims_0 = const()[name = tensor("reduce_mean_17_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_17_cast_fp16 = reduce_mean(axes = reduce_mean_17_axes_0, keep_dims = reduce_mean_17_keep_dims_0, x = square_8_cast_fp16)[name = tensor("reduce_mean_17_cast_fp16")]; + tensor var_382_to_fp16 = const()[name = tensor("op_382_to_fp16"), val = tensor(0x1.1p-20)]; + tensor var_383_cast_fp16 = add(x = reduce_mean_17_cast_fp16, y = var_382_to_fp16)[name = tensor("op_383_cast_fp16")]; + tensor var_384_cast_fp16 = sqrt(x = var_383_cast_fp16)[name = tensor("op_384_cast_fp16")]; + tensor x_43_cast_fp16 = real_div(x = sub_10_cast_fp16, y = var_384_cast_fp16)[name = tensor("x_43_cast_fp16")]; + tensor var_386_promoted_to_fp16 = const()[name = tensor("op_386_promoted_to_fp16"), val = tensor(0x1p+0)]; + tensor var_387_cast_fp16 = add(x = var_375_cast_fp16_1, y = var_386_promoted_to_fp16)[name = tensor("op_387_cast_fp16")]; + tensor var_388_cast_fp16 = mul(x = x_43_cast_fp16, y = var_387_cast_fp16)[name = tensor("op_388_cast_fp16")]; + tensor input_65_cast_fp16 = add(x = var_388_cast_fp16, y = var_375_cast_fp16_0)[name = tensor("input_65_cast_fp16")]; + tensor flow_net_final_layer_linear_weight_to_fp16 = const()[name = tensor("flow_net_final_layer_linear_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(18964416)))]; + tensor flow_net_final_layer_linear_bias_to_fp16 = const()[name = tensor("flow_net_final_layer_linear_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(18997248)))]; + tensor linear_25_cast_fp16 = linear(bias = flow_net_final_layer_linear_bias_to_fp16, weight = flow_net_final_layer_linear_weight_to_fp16, x = input_65_cast_fp16)[name = tensor("linear_25_cast_fp16")]; + tensor var_399_to_fp16 = const()[name = tensor("op_399_to_fp16"), val = tensor(0x1p-3)]; + tensor var_400_cast_fp16 = mul(x = linear_25_cast_fp16, y = var_399_to_fp16)[name = tensor("op_400_cast_fp16")]; + tensor input_67_cast_fp16 = add(x = latent_init_to_fp16, y = var_400_cast_fp16)[name = tensor("input_67_cast_fp16")]; + tensor linear_26_cast_fp16 = linear(bias = flow_net_input_proj_bias_to_fp16, weight = flow_net_input_proj_weight_to_fp16, x = input_67_cast_fp16)[name = tensor("linear_26_cast_fp16")]; + tensor input_71_to_fp16 = const()[name = tensor("input_71_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(18997376)))]; + tensor input_73_cast_fp16 = silu(x = input_71_to_fp16)[name = tensor("input_73_cast_fp16")]; + tensor linear_28_cast_fp16 = linear(bias = flow_net_time_embed_0_mlp_2_bias_to_fp16, weight = flow_net_time_embed_0_mlp_2_weight_to_fp16, x = input_73_cast_fp16)[name = tensor("linear_28_cast_fp16")]; + tensor reduce_mean_18_axes_0 = const()[name = tensor("reduce_mean_18_axes_0"), val = tensor([-1])]; + tensor reduce_mean_18_keep_dims_0 = const()[name = tensor("reduce_mean_18_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_18_cast_fp16 = reduce_mean(axes = reduce_mean_18_axes_0, keep_dims = reduce_mean_18_keep_dims_0, x = linear_28_cast_fp16)[name = tensor("reduce_mean_18_cast_fp16")]; + tensor sub_11_cast_fp16 = sub(x = linear_28_cast_fp16, y = reduce_mean_18_cast_fp16)[name = tensor("sub_11_cast_fp16")]; + tensor square_9_cast_fp16 = square(x = sub_11_cast_fp16)[name = tensor("square_9_cast_fp16")]; + tensor reduce_mean_19_axes_0 = const()[name = tensor("reduce_mean_19_axes_0"), val = tensor([-1])]; + tensor reduce_mean_19_keep_dims_0 = const()[name = tensor("reduce_mean_19_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_19_cast_fp16 = reduce_mean(axes = reduce_mean_19_axes_0, keep_dims = reduce_mean_19_keep_dims_0, x = square_9_cast_fp16)[name = tensor("reduce_mean_19_cast_fp16")]; + tensor real_div_2_to_fp16 = const()[name = tensor("real_div_2_to_fp16"), val = tensor(0x1.008p+0)]; + tensor mul_2_cast_fp16 = mul(x = reduce_mean_19_cast_fp16, y = real_div_2_to_fp16)[name = tensor("mul_2_cast_fp16")]; + tensor var_466_to_fp16 = const()[name = tensor("op_466_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_19_cast_fp16 = add(x = mul_2_cast_fp16, y = var_466_to_fp16)[name = tensor("var_19_cast_fp16")]; + tensor var_469_epsilon_0 = const()[name = tensor("op_469_epsilon_0"), val = tensor(0x1.197998p-40)]; + tensor var_469_cast_fp16 = rsqrt(epsilon = var_469_epsilon_0, x = var_19_cast_fp16)[name = tensor("op_469_cast_fp16")]; + tensor var_470_cast_fp16 = mul(x = const_3_to_fp16, y = var_469_cast_fp16)[name = tensor("op_470_cast_fp16")]; + tensor var_471_cast_fp16 = mul(x = linear_28_cast_fp16, y = var_470_cast_fp16)[name = tensor("op_471_cast_fp16")]; + tensor input_77_to_fp16 = const()[name = tensor("input_77_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(18998464)))]; + tensor input_79_cast_fp16 = silu(x = input_77_to_fp16)[name = tensor("input_79_cast_fp16")]; + tensor linear_30_cast_fp16 = linear(bias = flow_net_time_embed_1_mlp_2_bias_to_fp16, weight = flow_net_time_embed_1_mlp_2_weight_to_fp16, x = input_79_cast_fp16)[name = tensor("linear_30_cast_fp16")]; + tensor reduce_mean_20_axes_0 = const()[name = tensor("reduce_mean_20_axes_0"), val = tensor([-1])]; + tensor reduce_mean_20_keep_dims_0 = const()[name = tensor("reduce_mean_20_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_20_cast_fp16 = reduce_mean(axes = reduce_mean_20_axes_0, keep_dims = reduce_mean_20_keep_dims_0, x = linear_30_cast_fp16)[name = tensor("reduce_mean_20_cast_fp16")]; + tensor sub_13_cast_fp16 = sub(x = linear_30_cast_fp16, y = reduce_mean_20_cast_fp16)[name = tensor("sub_13_cast_fp16")]; + tensor square_10_cast_fp16 = square(x = sub_13_cast_fp16)[name = tensor("square_10_cast_fp16")]; + tensor reduce_mean_21_axes_0 = const()[name = tensor("reduce_mean_21_axes_0"), val = tensor([-1])]; + tensor reduce_mean_21_keep_dims_0 = const()[name = tensor("reduce_mean_21_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_21_cast_fp16 = reduce_mean(axes = reduce_mean_21_axes_0, keep_dims = reduce_mean_21_keep_dims_0, x = square_10_cast_fp16)[name = tensor("reduce_mean_21_cast_fp16")]; + tensor real_div_3_to_fp16 = const()[name = tensor("real_div_3_to_fp16"), val = tensor(0x1.008p+0)]; + tensor mul_3_cast_fp16 = mul(x = reduce_mean_21_cast_fp16, y = real_div_3_to_fp16)[name = tensor("mul_3_cast_fp16")]; + tensor var_503_to_fp16 = const()[name = tensor("op_503_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_21_cast_fp16 = add(x = mul_3_cast_fp16, y = var_503_to_fp16)[name = tensor("var_21_cast_fp16")]; + tensor var_506_epsilon_0 = const()[name = tensor("op_506_epsilon_0"), val = tensor(0x1.197998p-40)]; + tensor var_506_cast_fp16 = rsqrt(epsilon = var_506_epsilon_0, x = var_21_cast_fp16)[name = tensor("op_506_cast_fp16")]; + tensor var_507_cast_fp16 = mul(x = const_5_to_fp16, y = var_506_cast_fp16)[name = tensor("op_507_cast_fp16")]; + tensor var_508_cast_fp16 = mul(x = linear_30_cast_fp16, y = var_507_cast_fp16)[name = tensor("op_508_cast_fp16")]; + tensor var_520_cast_fp16 = add(x = var_471_cast_fp16, y = var_508_cast_fp16)[name = tensor("op_520_cast_fp16")]; + tensor _inversed_t_combined_3_y_0_to_fp16 = const()[name = tensor("_inversed_t_combined_3_y_0_to_fp16"), val = tensor(0x1p-1)]; + tensor _inversed_t_combined_3_cast_fp16 = mul(x = var_520_cast_fp16, y = _inversed_t_combined_3_y_0_to_fp16)[name = tensor("_inversed_t_combined_3_cast_fp16")]; + tensor input_81_cast_fp16 = add(x = _inversed_t_combined_3_cast_fp16, y = linear_5_cast_fp16)[name = tensor("input_81_cast_fp16")]; + tensor input_83_cast_fp16 = silu(x = input_81_cast_fp16)[name = tensor("input_83_cast_fp16")]; + tensor linear_32_cast_fp16 = linear(bias = flow_net_res_blocks_0_adaLN_modulation_1_bias_to_fp16, weight = flow_net_res_blocks_0_adaLN_modulation_1_weight_to_fp16, x = input_83_cast_fp16)[name = tensor("linear_32_cast_fp16")]; + tensor var_535_split_sizes_0 = const()[name = tensor("op_535_split_sizes_0"), val = tensor([512, 512, 512])]; + tensor var_535_axis_0 = const()[name = tensor("op_535_axis_0"), val = tensor(-1)]; + tensor var_535_cast_fp16_0, tensor var_535_cast_fp16_1, tensor var_535_cast_fp16_2 = split(axis = var_535_axis_0, split_sizes = var_535_split_sizes_0, x = linear_32_cast_fp16)[name = tensor("op_535_cast_fp16")]; + tensor mean_15_axes_0 = const()[name = tensor("mean_15_axes_0"), val = tensor([-1])]; + tensor mean_15_keep_dims_0 = const()[name = tensor("mean_15_keep_dims_0"), val = tensor(true)]; + tensor mean_15_cast_fp16 = reduce_mean(axes = mean_15_axes_0, keep_dims = mean_15_keep_dims_0, x = linear_26_cast_fp16)[name = tensor("mean_15_cast_fp16")]; + tensor sub_15_cast_fp16 = sub(x = linear_26_cast_fp16, y = mean_15_cast_fp16)[name = tensor("sub_15_cast_fp16")]; + tensor square_11_cast_fp16 = square(x = sub_15_cast_fp16)[name = tensor("square_11_cast_fp16")]; + tensor reduce_mean_23_axes_0 = const()[name = tensor("reduce_mean_23_axes_0"), val = tensor([-1])]; + tensor reduce_mean_23_keep_dims_0 = const()[name = tensor("reduce_mean_23_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_23_cast_fp16 = reduce_mean(axes = reduce_mean_23_axes_0, keep_dims = reduce_mean_23_keep_dims_0, x = square_11_cast_fp16)[name = tensor("reduce_mean_23_cast_fp16")]; + tensor var_545_to_fp16 = const()[name = tensor("op_545_to_fp16"), val = tensor(0x1.1p-20)]; + tensor var_546_cast_fp16 = add(x = reduce_mean_23_cast_fp16, y = var_545_to_fp16)[name = tensor("op_546_cast_fp16")]; + tensor var_547_cast_fp16 = sqrt(x = var_546_cast_fp16)[name = tensor("op_547_cast_fp16")]; + tensor x_51_cast_fp16 = real_div(x = sub_15_cast_fp16, y = var_547_cast_fp16)[name = tensor("x_51_cast_fp16")]; + tensor var_549_cast_fp16 = mul(x = x_51_cast_fp16, y = flow_net_res_blocks_0_in_ln_weight_to_fp16)[name = tensor("op_549_cast_fp16")]; + tensor x_53_cast_fp16 = add(x = var_549_cast_fp16, y = flow_net_res_blocks_0_in_ln_bias_to_fp16)[name = tensor("x_53_cast_fp16")]; + tensor var_551_promoted_to_fp16 = const()[name = tensor("op_551_promoted_to_fp16"), val = tensor(0x1p+0)]; + tensor var_552_cast_fp16 = add(x = var_535_cast_fp16_1, y = var_551_promoted_to_fp16)[name = tensor("op_552_cast_fp16")]; + tensor var_553_cast_fp16 = mul(x = x_53_cast_fp16, y = var_552_cast_fp16)[name = tensor("op_553_cast_fp16")]; + tensor input_85_cast_fp16 = add(x = var_553_cast_fp16, y = var_535_cast_fp16_0)[name = tensor("input_85_cast_fp16")]; + tensor linear_33_cast_fp16 = linear(bias = flow_net_res_blocks_0_mlp_0_bias_to_fp16, weight = flow_net_res_blocks_0_mlp_0_weight_to_fp16, x = input_85_cast_fp16)[name = tensor("linear_33_cast_fp16")]; + tensor input_89_cast_fp16 = silu(x = linear_33_cast_fp16)[name = tensor("input_89_cast_fp16")]; + tensor linear_34_cast_fp16 = linear(bias = flow_net_res_blocks_0_mlp_2_bias_to_fp16, weight = flow_net_res_blocks_0_mlp_2_weight_to_fp16, x = input_89_cast_fp16)[name = tensor("linear_34_cast_fp16")]; + tensor var_564_cast_fp16 = mul(x = var_535_cast_fp16_2, y = linear_34_cast_fp16)[name = tensor("op_564_cast_fp16")]; + tensor x_55_cast_fp16 = add(x = linear_26_cast_fp16, y = var_564_cast_fp16)[name = tensor("x_55_cast_fp16")]; + tensor linear_35_cast_fp16 = linear(bias = flow_net_res_blocks_1_adaLN_modulation_1_bias_to_fp16, weight = flow_net_res_blocks_1_adaLN_modulation_1_weight_to_fp16, x = input_83_cast_fp16)[name = tensor("linear_35_cast_fp16")]; + tensor var_574_split_sizes_0 = const()[name = tensor("op_574_split_sizes_0"), val = tensor([512, 512, 512])]; + tensor var_574_axis_0 = const()[name = tensor("op_574_axis_0"), val = tensor(-1)]; + tensor var_574_cast_fp16_0, tensor var_574_cast_fp16_1, tensor var_574_cast_fp16_2 = split(axis = var_574_axis_0, split_sizes = var_574_split_sizes_0, x = linear_35_cast_fp16)[name = tensor("op_574_cast_fp16")]; + tensor mean_17_axes_0 = const()[name = tensor("mean_17_axes_0"), val = tensor([-1])]; + tensor mean_17_keep_dims_0 = const()[name = tensor("mean_17_keep_dims_0"), val = tensor(true)]; + tensor mean_17_cast_fp16 = reduce_mean(axes = mean_17_axes_0, keep_dims = mean_17_keep_dims_0, x = x_55_cast_fp16)[name = tensor("mean_17_cast_fp16")]; + tensor sub_16_cast_fp16 = sub(x = x_55_cast_fp16, y = mean_17_cast_fp16)[name = tensor("sub_16_cast_fp16")]; + tensor square_12_cast_fp16 = square(x = sub_16_cast_fp16)[name = tensor("square_12_cast_fp16")]; + tensor reduce_mean_25_axes_0 = const()[name = tensor("reduce_mean_25_axes_0"), val = tensor([-1])]; + tensor reduce_mean_25_keep_dims_0 = const()[name = tensor("reduce_mean_25_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_25_cast_fp16 = reduce_mean(axes = reduce_mean_25_axes_0, keep_dims = reduce_mean_25_keep_dims_0, x = square_12_cast_fp16)[name = tensor("reduce_mean_25_cast_fp16")]; + tensor var_584_to_fp16 = const()[name = tensor("op_584_to_fp16"), val = tensor(0x1.1p-20)]; + tensor var_585_cast_fp16 = add(x = reduce_mean_25_cast_fp16, y = var_584_to_fp16)[name = tensor("op_585_cast_fp16")]; + tensor var_586_cast_fp16 = sqrt(x = var_585_cast_fp16)[name = tensor("op_586_cast_fp16")]; + tensor x_57_cast_fp16 = real_div(x = sub_16_cast_fp16, y = var_586_cast_fp16)[name = tensor("x_57_cast_fp16")]; + tensor var_588_cast_fp16 = mul(x = x_57_cast_fp16, y = flow_net_res_blocks_1_in_ln_weight_to_fp16)[name = tensor("op_588_cast_fp16")]; + tensor x_59_cast_fp16 = add(x = var_588_cast_fp16, y = flow_net_res_blocks_1_in_ln_bias_to_fp16)[name = tensor("x_59_cast_fp16")]; + tensor var_590_promoted_to_fp16 = const()[name = tensor("op_590_promoted_to_fp16"), val = tensor(0x1p+0)]; + tensor var_591_cast_fp16 = add(x = var_574_cast_fp16_1, y = var_590_promoted_to_fp16)[name = tensor("op_591_cast_fp16")]; + tensor var_592_cast_fp16 = mul(x = x_59_cast_fp16, y = var_591_cast_fp16)[name = tensor("op_592_cast_fp16")]; + tensor input_93_cast_fp16 = add(x = var_592_cast_fp16, y = var_574_cast_fp16_0)[name = tensor("input_93_cast_fp16")]; + tensor linear_36_cast_fp16 = linear(bias = flow_net_res_blocks_1_mlp_0_bias_to_fp16, weight = flow_net_res_blocks_1_mlp_0_weight_to_fp16, x = input_93_cast_fp16)[name = tensor("linear_36_cast_fp16")]; + tensor input_97_cast_fp16 = silu(x = linear_36_cast_fp16)[name = tensor("input_97_cast_fp16")]; + tensor linear_37_cast_fp16 = linear(bias = flow_net_res_blocks_1_mlp_2_bias_to_fp16, weight = flow_net_res_blocks_1_mlp_2_weight_to_fp16, x = input_97_cast_fp16)[name = tensor("linear_37_cast_fp16")]; + tensor var_603_cast_fp16 = mul(x = var_574_cast_fp16_2, y = linear_37_cast_fp16)[name = tensor("op_603_cast_fp16")]; + tensor x_61_cast_fp16 = add(x = x_55_cast_fp16, y = var_603_cast_fp16)[name = tensor("x_61_cast_fp16")]; + tensor linear_38_cast_fp16 = linear(bias = flow_net_res_blocks_2_adaLN_modulation_1_bias_to_fp16, weight = flow_net_res_blocks_2_adaLN_modulation_1_weight_to_fp16, x = input_83_cast_fp16)[name = tensor("linear_38_cast_fp16")]; + tensor var_613_split_sizes_0 = const()[name = tensor("op_613_split_sizes_0"), val = tensor([512, 512, 512])]; + tensor var_613_axis_0 = const()[name = tensor("op_613_axis_0"), val = tensor(-1)]; + tensor var_613_cast_fp16_0, tensor var_613_cast_fp16_1, tensor var_613_cast_fp16_2 = split(axis = var_613_axis_0, split_sizes = var_613_split_sizes_0, x = linear_38_cast_fp16)[name = tensor("op_613_cast_fp16")]; + tensor mean_19_axes_0 = const()[name = tensor("mean_19_axes_0"), val = tensor([-1])]; + tensor mean_19_keep_dims_0 = const()[name = tensor("mean_19_keep_dims_0"), val = tensor(true)]; + tensor mean_19_cast_fp16 = reduce_mean(axes = mean_19_axes_0, keep_dims = mean_19_keep_dims_0, x = x_61_cast_fp16)[name = tensor("mean_19_cast_fp16")]; + tensor sub_17_cast_fp16 = sub(x = x_61_cast_fp16, y = mean_19_cast_fp16)[name = tensor("sub_17_cast_fp16")]; + tensor square_13_cast_fp16 = square(x = sub_17_cast_fp16)[name = tensor("square_13_cast_fp16")]; + tensor reduce_mean_27_axes_0 = const()[name = tensor("reduce_mean_27_axes_0"), val = tensor([-1])]; + tensor reduce_mean_27_keep_dims_0 = const()[name = tensor("reduce_mean_27_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_27_cast_fp16 = reduce_mean(axes = reduce_mean_27_axes_0, keep_dims = reduce_mean_27_keep_dims_0, x = square_13_cast_fp16)[name = tensor("reduce_mean_27_cast_fp16")]; + tensor var_623_to_fp16 = const()[name = tensor("op_623_to_fp16"), val = tensor(0x1.1p-20)]; + tensor var_624_cast_fp16 = add(x = reduce_mean_27_cast_fp16, y = var_623_to_fp16)[name = tensor("op_624_cast_fp16")]; + tensor var_625_cast_fp16 = sqrt(x = var_624_cast_fp16)[name = tensor("op_625_cast_fp16")]; + tensor x_63_cast_fp16 = real_div(x = sub_17_cast_fp16, y = var_625_cast_fp16)[name = tensor("x_63_cast_fp16")]; + tensor var_627_cast_fp16 = mul(x = x_63_cast_fp16, y = flow_net_res_blocks_2_in_ln_weight_to_fp16)[name = tensor("op_627_cast_fp16")]; + tensor x_65_cast_fp16 = add(x = var_627_cast_fp16, y = flow_net_res_blocks_2_in_ln_bias_to_fp16)[name = tensor("x_65_cast_fp16")]; + tensor var_629_promoted_to_fp16 = const()[name = tensor("op_629_promoted_to_fp16"), val = tensor(0x1p+0)]; + tensor var_630_cast_fp16 = add(x = var_613_cast_fp16_1, y = var_629_promoted_to_fp16)[name = tensor("op_630_cast_fp16")]; + tensor var_631_cast_fp16 = mul(x = x_65_cast_fp16, y = var_630_cast_fp16)[name = tensor("op_631_cast_fp16")]; + tensor input_101_cast_fp16 = add(x = var_631_cast_fp16, y = var_613_cast_fp16_0)[name = tensor("input_101_cast_fp16")]; + tensor linear_39_cast_fp16 = linear(bias = flow_net_res_blocks_2_mlp_0_bias_to_fp16, weight = flow_net_res_blocks_2_mlp_0_weight_to_fp16, x = input_101_cast_fp16)[name = tensor("linear_39_cast_fp16")]; + tensor input_105_cast_fp16 = silu(x = linear_39_cast_fp16)[name = tensor("input_105_cast_fp16")]; + tensor linear_40_cast_fp16 = linear(bias = flow_net_res_blocks_2_mlp_2_bias_to_fp16, weight = flow_net_res_blocks_2_mlp_2_weight_to_fp16, x = input_105_cast_fp16)[name = tensor("linear_40_cast_fp16")]; + tensor var_642_cast_fp16 = mul(x = var_613_cast_fp16_2, y = linear_40_cast_fp16)[name = tensor("op_642_cast_fp16")]; + tensor x_67_cast_fp16 = add(x = x_61_cast_fp16, y = var_642_cast_fp16)[name = tensor("x_67_cast_fp16")]; + tensor linear_41_cast_fp16 = linear(bias = flow_net_res_blocks_3_adaLN_modulation_1_bias_to_fp16, weight = flow_net_res_blocks_3_adaLN_modulation_1_weight_to_fp16, x = input_83_cast_fp16)[name = tensor("linear_41_cast_fp16")]; + tensor var_652_split_sizes_0 = const()[name = tensor("op_652_split_sizes_0"), val = tensor([512, 512, 512])]; + tensor var_652_axis_0 = const()[name = tensor("op_652_axis_0"), val = tensor(-1)]; + tensor var_652_cast_fp16_0, tensor var_652_cast_fp16_1, tensor var_652_cast_fp16_2 = split(axis = var_652_axis_0, split_sizes = var_652_split_sizes_0, x = linear_41_cast_fp16)[name = tensor("op_652_cast_fp16")]; + tensor mean_21_axes_0 = const()[name = tensor("mean_21_axes_0"), val = tensor([-1])]; + tensor mean_21_keep_dims_0 = const()[name = tensor("mean_21_keep_dims_0"), val = tensor(true)]; + tensor mean_21_cast_fp16 = reduce_mean(axes = mean_21_axes_0, keep_dims = mean_21_keep_dims_0, x = x_67_cast_fp16)[name = tensor("mean_21_cast_fp16")]; + tensor sub_18_cast_fp16 = sub(x = x_67_cast_fp16, y = mean_21_cast_fp16)[name = tensor("sub_18_cast_fp16")]; + tensor square_14_cast_fp16 = square(x = sub_18_cast_fp16)[name = tensor("square_14_cast_fp16")]; + tensor reduce_mean_29_axes_0 = const()[name = tensor("reduce_mean_29_axes_0"), val = tensor([-1])]; + tensor reduce_mean_29_keep_dims_0 = const()[name = tensor("reduce_mean_29_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_29_cast_fp16 = reduce_mean(axes = reduce_mean_29_axes_0, keep_dims = reduce_mean_29_keep_dims_0, x = square_14_cast_fp16)[name = tensor("reduce_mean_29_cast_fp16")]; + tensor var_662_to_fp16 = const()[name = tensor("op_662_to_fp16"), val = tensor(0x1.1p-20)]; + tensor var_663_cast_fp16 = add(x = reduce_mean_29_cast_fp16, y = var_662_to_fp16)[name = tensor("op_663_cast_fp16")]; + tensor var_664_cast_fp16 = sqrt(x = var_663_cast_fp16)[name = tensor("op_664_cast_fp16")]; + tensor x_69_cast_fp16 = real_div(x = sub_18_cast_fp16, y = var_664_cast_fp16)[name = tensor("x_69_cast_fp16")]; + tensor var_666_cast_fp16 = mul(x = x_69_cast_fp16, y = flow_net_res_blocks_3_in_ln_weight_to_fp16)[name = tensor("op_666_cast_fp16")]; + tensor x_71_cast_fp16 = add(x = var_666_cast_fp16, y = flow_net_res_blocks_3_in_ln_bias_to_fp16)[name = tensor("x_71_cast_fp16")]; + tensor var_668_promoted_to_fp16 = const()[name = tensor("op_668_promoted_to_fp16"), val = tensor(0x1p+0)]; + tensor var_669_cast_fp16 = add(x = var_652_cast_fp16_1, y = var_668_promoted_to_fp16)[name = tensor("op_669_cast_fp16")]; + tensor var_670_cast_fp16 = mul(x = x_71_cast_fp16, y = var_669_cast_fp16)[name = tensor("op_670_cast_fp16")]; + tensor input_109_cast_fp16 = add(x = var_670_cast_fp16, y = var_652_cast_fp16_0)[name = tensor("input_109_cast_fp16")]; + tensor linear_42_cast_fp16 = linear(bias = flow_net_res_blocks_3_mlp_0_bias_to_fp16, weight = flow_net_res_blocks_3_mlp_0_weight_to_fp16, x = input_109_cast_fp16)[name = tensor("linear_42_cast_fp16")]; + tensor input_113_cast_fp16 = silu(x = linear_42_cast_fp16)[name = tensor("input_113_cast_fp16")]; + tensor linear_43_cast_fp16 = linear(bias = flow_net_res_blocks_3_mlp_2_bias_to_fp16, weight = flow_net_res_blocks_3_mlp_2_weight_to_fp16, x = input_113_cast_fp16)[name = tensor("linear_43_cast_fp16")]; + tensor var_681_cast_fp16 = mul(x = var_652_cast_fp16_2, y = linear_43_cast_fp16)[name = tensor("op_681_cast_fp16")]; + tensor x_73_cast_fp16 = add(x = x_67_cast_fp16, y = var_681_cast_fp16)[name = tensor("x_73_cast_fp16")]; + tensor linear_44_cast_fp16 = linear(bias = flow_net_res_blocks_4_adaLN_modulation_1_bias_to_fp16, weight = flow_net_res_blocks_4_adaLN_modulation_1_weight_to_fp16, x = input_83_cast_fp16)[name = tensor("linear_44_cast_fp16")]; + tensor var_691_split_sizes_0 = const()[name = tensor("op_691_split_sizes_0"), val = tensor([512, 512, 512])]; + tensor var_691_axis_0 = const()[name = tensor("op_691_axis_0"), val = tensor(-1)]; + tensor var_691_cast_fp16_0, tensor var_691_cast_fp16_1, tensor var_691_cast_fp16_2 = split(axis = var_691_axis_0, split_sizes = var_691_split_sizes_0, x = linear_44_cast_fp16)[name = tensor("op_691_cast_fp16")]; + tensor mean_23_axes_0 = const()[name = tensor("mean_23_axes_0"), val = tensor([-1])]; + tensor mean_23_keep_dims_0 = const()[name = tensor("mean_23_keep_dims_0"), val = tensor(true)]; + tensor mean_23_cast_fp16 = reduce_mean(axes = mean_23_axes_0, keep_dims = mean_23_keep_dims_0, x = x_73_cast_fp16)[name = tensor("mean_23_cast_fp16")]; + tensor sub_19_cast_fp16 = sub(x = x_73_cast_fp16, y = mean_23_cast_fp16)[name = tensor("sub_19_cast_fp16")]; + tensor square_15_cast_fp16 = square(x = sub_19_cast_fp16)[name = tensor("square_15_cast_fp16")]; + tensor reduce_mean_31_axes_0 = const()[name = tensor("reduce_mean_31_axes_0"), val = tensor([-1])]; + tensor reduce_mean_31_keep_dims_0 = const()[name = tensor("reduce_mean_31_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_31_cast_fp16 = reduce_mean(axes = reduce_mean_31_axes_0, keep_dims = reduce_mean_31_keep_dims_0, x = square_15_cast_fp16)[name = tensor("reduce_mean_31_cast_fp16")]; + tensor var_701_to_fp16 = const()[name = tensor("op_701_to_fp16"), val = tensor(0x1.1p-20)]; + tensor var_702_cast_fp16 = add(x = reduce_mean_31_cast_fp16, y = var_701_to_fp16)[name = tensor("op_702_cast_fp16")]; + tensor var_703_cast_fp16 = sqrt(x = var_702_cast_fp16)[name = tensor("op_703_cast_fp16")]; + tensor x_75_cast_fp16 = real_div(x = sub_19_cast_fp16, y = var_703_cast_fp16)[name = tensor("x_75_cast_fp16")]; + tensor var_705_cast_fp16 = mul(x = x_75_cast_fp16, y = flow_net_res_blocks_4_in_ln_weight_to_fp16)[name = tensor("op_705_cast_fp16")]; + tensor x_77_cast_fp16 = add(x = var_705_cast_fp16, y = flow_net_res_blocks_4_in_ln_bias_to_fp16)[name = tensor("x_77_cast_fp16")]; + tensor var_707_promoted_to_fp16 = const()[name = tensor("op_707_promoted_to_fp16"), val = tensor(0x1p+0)]; + tensor var_708_cast_fp16 = add(x = var_691_cast_fp16_1, y = var_707_promoted_to_fp16)[name = tensor("op_708_cast_fp16")]; + tensor var_709_cast_fp16 = mul(x = x_77_cast_fp16, y = var_708_cast_fp16)[name = tensor("op_709_cast_fp16")]; + tensor input_117_cast_fp16 = add(x = var_709_cast_fp16, y = var_691_cast_fp16_0)[name = tensor("input_117_cast_fp16")]; + tensor linear_45_cast_fp16 = linear(bias = flow_net_res_blocks_4_mlp_0_bias_to_fp16, weight = flow_net_res_blocks_4_mlp_0_weight_to_fp16, x = input_117_cast_fp16)[name = tensor("linear_45_cast_fp16")]; + tensor input_121_cast_fp16 = silu(x = linear_45_cast_fp16)[name = tensor("input_121_cast_fp16")]; + tensor linear_46_cast_fp16 = linear(bias = flow_net_res_blocks_4_mlp_2_bias_to_fp16, weight = flow_net_res_blocks_4_mlp_2_weight_to_fp16, x = input_121_cast_fp16)[name = tensor("linear_46_cast_fp16")]; + tensor var_720_cast_fp16 = mul(x = var_691_cast_fp16_2, y = linear_46_cast_fp16)[name = tensor("op_720_cast_fp16")]; + tensor x_79_cast_fp16 = add(x = x_73_cast_fp16, y = var_720_cast_fp16)[name = tensor("x_79_cast_fp16")]; + tensor linear_47_cast_fp16 = linear(bias = flow_net_res_blocks_5_adaLN_modulation_1_bias_to_fp16, weight = flow_net_res_blocks_5_adaLN_modulation_1_weight_to_fp16, x = input_83_cast_fp16)[name = tensor("linear_47_cast_fp16")]; + tensor var_730_split_sizes_0 = const()[name = tensor("op_730_split_sizes_0"), val = tensor([512, 512, 512])]; + tensor var_730_axis_0 = const()[name = tensor("op_730_axis_0"), val = tensor(-1)]; + tensor var_730_cast_fp16_0, tensor var_730_cast_fp16_1, tensor var_730_cast_fp16_2 = split(axis = var_730_axis_0, split_sizes = var_730_split_sizes_0, x = linear_47_cast_fp16)[name = tensor("op_730_cast_fp16")]; + tensor mean_25_axes_0 = const()[name = tensor("mean_25_axes_0"), val = tensor([-1])]; + tensor mean_25_keep_dims_0 = const()[name = tensor("mean_25_keep_dims_0"), val = tensor(true)]; + tensor mean_25_cast_fp16 = reduce_mean(axes = mean_25_axes_0, keep_dims = mean_25_keep_dims_0, x = x_79_cast_fp16)[name = tensor("mean_25_cast_fp16")]; + tensor sub_20_cast_fp16 = sub(x = x_79_cast_fp16, y = mean_25_cast_fp16)[name = tensor("sub_20_cast_fp16")]; + tensor square_16_cast_fp16 = square(x = sub_20_cast_fp16)[name = tensor("square_16_cast_fp16")]; + tensor reduce_mean_33_axes_0 = const()[name = tensor("reduce_mean_33_axes_0"), val = tensor([-1])]; + tensor reduce_mean_33_keep_dims_0 = const()[name = tensor("reduce_mean_33_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_33_cast_fp16 = reduce_mean(axes = reduce_mean_33_axes_0, keep_dims = reduce_mean_33_keep_dims_0, x = square_16_cast_fp16)[name = tensor("reduce_mean_33_cast_fp16")]; + tensor var_740_to_fp16 = const()[name = tensor("op_740_to_fp16"), val = tensor(0x1.1p-20)]; + tensor var_741_cast_fp16 = add(x = reduce_mean_33_cast_fp16, y = var_740_to_fp16)[name = tensor("op_741_cast_fp16")]; + tensor var_742_cast_fp16 = sqrt(x = var_741_cast_fp16)[name = tensor("op_742_cast_fp16")]; + tensor x_81_cast_fp16 = real_div(x = sub_20_cast_fp16, y = var_742_cast_fp16)[name = tensor("x_81_cast_fp16")]; + tensor var_744_cast_fp16 = mul(x = x_81_cast_fp16, y = flow_net_res_blocks_5_in_ln_weight_to_fp16)[name = tensor("op_744_cast_fp16")]; + tensor x_83_cast_fp16 = add(x = var_744_cast_fp16, y = flow_net_res_blocks_5_in_ln_bias_to_fp16)[name = tensor("x_83_cast_fp16")]; + tensor var_746_promoted_to_fp16 = const()[name = tensor("op_746_promoted_to_fp16"), val = tensor(0x1p+0)]; + tensor var_747_cast_fp16 = add(x = var_730_cast_fp16_1, y = var_746_promoted_to_fp16)[name = tensor("op_747_cast_fp16")]; + tensor var_748_cast_fp16 = mul(x = x_83_cast_fp16, y = var_747_cast_fp16)[name = tensor("op_748_cast_fp16")]; + tensor input_125_cast_fp16 = add(x = var_748_cast_fp16, y = var_730_cast_fp16_0)[name = tensor("input_125_cast_fp16")]; + tensor linear_48_cast_fp16 = linear(bias = flow_net_res_blocks_5_mlp_0_bias_to_fp16, weight = flow_net_res_blocks_5_mlp_0_weight_to_fp16, x = input_125_cast_fp16)[name = tensor("linear_48_cast_fp16")]; + tensor input_129_cast_fp16 = silu(x = linear_48_cast_fp16)[name = tensor("input_129_cast_fp16")]; + tensor linear_49_cast_fp16 = linear(bias = flow_net_res_blocks_5_mlp_2_bias_to_fp16, weight = flow_net_res_blocks_5_mlp_2_weight_to_fp16, x = input_129_cast_fp16)[name = tensor("linear_49_cast_fp16")]; + tensor var_759_cast_fp16 = mul(x = var_730_cast_fp16_2, y = linear_49_cast_fp16)[name = tensor("op_759_cast_fp16")]; + tensor x_85_cast_fp16 = add(x = x_79_cast_fp16, y = var_759_cast_fp16)[name = tensor("x_85_cast_fp16")]; + tensor linear_50_cast_fp16 = linear(bias = flow_net_final_layer_adaLN_modulation_1_bias_to_fp16, weight = flow_net_final_layer_adaLN_modulation_1_weight_to_fp16, x = input_83_cast_fp16)[name = tensor("linear_50_cast_fp16")]; + tensor var_768_split_sizes_0 = const()[name = tensor("op_768_split_sizes_0"), val = tensor([512, 512])]; + tensor var_768_axis_0 = const()[name = tensor("op_768_axis_0"), val = tensor(-1)]; + tensor var_768_cast_fp16_0, tensor var_768_cast_fp16_1 = split(axis = var_768_axis_0, split_sizes = var_768_split_sizes_0, x = linear_50_cast_fp16)[name = tensor("op_768_cast_fp16")]; + tensor mean_27_axes_0 = const()[name = tensor("mean_27_axes_0"), val = tensor([-1])]; + tensor mean_27_keep_dims_0 = const()[name = tensor("mean_27_keep_dims_0"), val = tensor(true)]; + tensor mean_27_cast_fp16 = reduce_mean(axes = mean_27_axes_0, keep_dims = mean_27_keep_dims_0, x = x_85_cast_fp16)[name = tensor("mean_27_cast_fp16")]; + tensor sub_21_cast_fp16 = sub(x = x_85_cast_fp16, y = mean_27_cast_fp16)[name = tensor("sub_21_cast_fp16")]; + tensor square_17_cast_fp16 = square(x = sub_21_cast_fp16)[name = tensor("square_17_cast_fp16")]; + tensor reduce_mean_35_axes_0 = const()[name = tensor("reduce_mean_35_axes_0"), val = tensor([-1])]; + tensor reduce_mean_35_keep_dims_0 = const()[name = tensor("reduce_mean_35_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_35_cast_fp16 = reduce_mean(axes = reduce_mean_35_axes_0, keep_dims = reduce_mean_35_keep_dims_0, x = square_17_cast_fp16)[name = tensor("reduce_mean_35_cast_fp16")]; + tensor var_775_to_fp16 = const()[name = tensor("op_775_to_fp16"), val = tensor(0x1.1p-20)]; + tensor var_776_cast_fp16 = add(x = reduce_mean_35_cast_fp16, y = var_775_to_fp16)[name = tensor("op_776_cast_fp16")]; + tensor var_777_cast_fp16 = sqrt(x = var_776_cast_fp16)[name = tensor("op_777_cast_fp16")]; + tensor x_87_cast_fp16 = real_div(x = sub_21_cast_fp16, y = var_777_cast_fp16)[name = tensor("x_87_cast_fp16")]; + tensor var_779_promoted_to_fp16 = const()[name = tensor("op_779_promoted_to_fp16"), val = tensor(0x1p+0)]; + tensor var_780_cast_fp16 = add(x = var_768_cast_fp16_1, y = var_779_promoted_to_fp16)[name = tensor("op_780_cast_fp16")]; + tensor var_781_cast_fp16 = mul(x = x_87_cast_fp16, y = var_780_cast_fp16)[name = tensor("op_781_cast_fp16")]; + tensor input_133_cast_fp16 = add(x = var_781_cast_fp16, y = var_768_cast_fp16_0)[name = tensor("input_133_cast_fp16")]; + tensor linear_51_cast_fp16 = linear(bias = flow_net_final_layer_linear_bias_to_fp16, weight = flow_net_final_layer_linear_weight_to_fp16, x = input_133_cast_fp16)[name = tensor("linear_51_cast_fp16")]; + tensor var_792_to_fp16 = const()[name = tensor("op_792_to_fp16"), val = tensor(0x1p-3)]; + tensor var_793_cast_fp16 = mul(x = linear_51_cast_fp16, y = var_792_to_fp16)[name = tensor("op_793_cast_fp16")]; + tensor input_135_cast_fp16 = add(x = input_67_cast_fp16, y = var_793_cast_fp16)[name = tensor("input_135_cast_fp16")]; + tensor linear_52_cast_fp16 = linear(bias = flow_net_input_proj_bias_to_fp16, weight = flow_net_input_proj_weight_to_fp16, x = input_135_cast_fp16)[name = tensor("linear_52_cast_fp16")]; + tensor input_139_to_fp16 = const()[name = tensor("input_139_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(18999552)))]; + tensor input_141_cast_fp16 = silu(x = input_139_to_fp16)[name = tensor("input_141_cast_fp16")]; + tensor linear_54_cast_fp16 = linear(bias = flow_net_time_embed_0_mlp_2_bias_to_fp16, weight = flow_net_time_embed_0_mlp_2_weight_to_fp16, x = input_141_cast_fp16)[name = tensor("linear_54_cast_fp16")]; + tensor reduce_mean_36_axes_0 = const()[name = tensor("reduce_mean_36_axes_0"), val = tensor([-1])]; + tensor reduce_mean_36_keep_dims_0 = const()[name = tensor("reduce_mean_36_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_36_cast_fp16 = reduce_mean(axes = reduce_mean_36_axes_0, keep_dims = reduce_mean_36_keep_dims_0, x = linear_54_cast_fp16)[name = tensor("reduce_mean_36_cast_fp16")]; + tensor sub_22_cast_fp16 = sub(x = linear_54_cast_fp16, y = reduce_mean_36_cast_fp16)[name = tensor("sub_22_cast_fp16")]; + tensor square_18_cast_fp16 = square(x = sub_22_cast_fp16)[name = tensor("square_18_cast_fp16")]; + tensor reduce_mean_37_axes_0 = const()[name = tensor("reduce_mean_37_axes_0"), val = tensor([-1])]; + tensor reduce_mean_37_keep_dims_0 = const()[name = tensor("reduce_mean_37_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_37_cast_fp16 = reduce_mean(axes = reduce_mean_37_axes_0, keep_dims = reduce_mean_37_keep_dims_0, x = square_18_cast_fp16)[name = tensor("reduce_mean_37_cast_fp16")]; + tensor real_div_4_to_fp16 = const()[name = tensor("real_div_4_to_fp16"), val = tensor(0x1.008p+0)]; + tensor mul_4_cast_fp16 = mul(x = reduce_mean_37_cast_fp16, y = real_div_4_to_fp16)[name = tensor("mul_4_cast_fp16")]; + tensor var_859_to_fp16 = const()[name = tensor("op_859_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_37_cast_fp16 = add(x = mul_4_cast_fp16, y = var_859_to_fp16)[name = tensor("var_37_cast_fp16")]; + tensor var_862_epsilon_0 = const()[name = tensor("op_862_epsilon_0"), val = tensor(0x1.197998p-40)]; + tensor var_862_cast_fp16 = rsqrt(epsilon = var_862_epsilon_0, x = var_37_cast_fp16)[name = tensor("op_862_cast_fp16")]; + tensor var_863_cast_fp16 = mul(x = const_3_to_fp16, y = var_862_cast_fp16)[name = tensor("op_863_cast_fp16")]; + tensor var_864_cast_fp16 = mul(x = linear_54_cast_fp16, y = var_863_cast_fp16)[name = tensor("op_864_cast_fp16")]; + tensor input_145_to_fp16 = const()[name = tensor("input_145_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(19000640)))]; + tensor input_147_cast_fp16 = silu(x = input_145_to_fp16)[name = tensor("input_147_cast_fp16")]; + tensor linear_56_cast_fp16 = linear(bias = flow_net_time_embed_1_mlp_2_bias_to_fp16, weight = flow_net_time_embed_1_mlp_2_weight_to_fp16, x = input_147_cast_fp16)[name = tensor("linear_56_cast_fp16")]; + tensor reduce_mean_38_axes_0 = const()[name = tensor("reduce_mean_38_axes_0"), val = tensor([-1])]; + tensor reduce_mean_38_keep_dims_0 = const()[name = tensor("reduce_mean_38_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_38_cast_fp16 = reduce_mean(axes = reduce_mean_38_axes_0, keep_dims = reduce_mean_38_keep_dims_0, x = linear_56_cast_fp16)[name = tensor("reduce_mean_38_cast_fp16")]; + tensor sub_24_cast_fp16 = sub(x = linear_56_cast_fp16, y = reduce_mean_38_cast_fp16)[name = tensor("sub_24_cast_fp16")]; + tensor square_19_cast_fp16 = square(x = sub_24_cast_fp16)[name = tensor("square_19_cast_fp16")]; + tensor reduce_mean_39_axes_0 = const()[name = tensor("reduce_mean_39_axes_0"), val = tensor([-1])]; + tensor reduce_mean_39_keep_dims_0 = const()[name = tensor("reduce_mean_39_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_39_cast_fp16 = reduce_mean(axes = reduce_mean_39_axes_0, keep_dims = reduce_mean_39_keep_dims_0, x = square_19_cast_fp16)[name = tensor("reduce_mean_39_cast_fp16")]; + tensor real_div_5_to_fp16 = const()[name = tensor("real_div_5_to_fp16"), val = tensor(0x1.008p+0)]; + tensor mul_5_cast_fp16 = mul(x = reduce_mean_39_cast_fp16, y = real_div_5_to_fp16)[name = tensor("mul_5_cast_fp16")]; + tensor var_896_to_fp16 = const()[name = tensor("op_896_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_39_cast_fp16 = add(x = mul_5_cast_fp16, y = var_896_to_fp16)[name = tensor("var_39_cast_fp16")]; + tensor var_899_epsilon_0 = const()[name = tensor("op_899_epsilon_0"), val = tensor(0x1.197998p-40)]; + tensor var_899_cast_fp16 = rsqrt(epsilon = var_899_epsilon_0, x = var_39_cast_fp16)[name = tensor("op_899_cast_fp16")]; + tensor var_900_cast_fp16 = mul(x = const_5_to_fp16, y = var_899_cast_fp16)[name = tensor("op_900_cast_fp16")]; + tensor var_901_cast_fp16 = mul(x = linear_56_cast_fp16, y = var_900_cast_fp16)[name = tensor("op_901_cast_fp16")]; + tensor var_913_cast_fp16 = add(x = var_864_cast_fp16, y = var_901_cast_fp16)[name = tensor("op_913_cast_fp16")]; + tensor _inversed_t_combined_5_y_0_to_fp16 = const()[name = tensor("_inversed_t_combined_5_y_0_to_fp16"), val = tensor(0x1p-1)]; + tensor _inversed_t_combined_5_cast_fp16 = mul(x = var_913_cast_fp16, y = _inversed_t_combined_5_y_0_to_fp16)[name = tensor("_inversed_t_combined_5_cast_fp16")]; + tensor input_149_cast_fp16 = add(x = _inversed_t_combined_5_cast_fp16, y = linear_5_cast_fp16)[name = tensor("input_149_cast_fp16")]; + tensor input_151_cast_fp16 = silu(x = input_149_cast_fp16)[name = tensor("input_151_cast_fp16")]; + tensor linear_58_cast_fp16 = linear(bias = flow_net_res_blocks_0_adaLN_modulation_1_bias_to_fp16, weight = flow_net_res_blocks_0_adaLN_modulation_1_weight_to_fp16, x = input_151_cast_fp16)[name = tensor("linear_58_cast_fp16")]; + tensor var_928_split_sizes_0 = const()[name = tensor("op_928_split_sizes_0"), val = tensor([512, 512, 512])]; + tensor var_928_axis_0 = const()[name = tensor("op_928_axis_0"), val = tensor(-1)]; + tensor var_928_cast_fp16_0, tensor var_928_cast_fp16_1, tensor var_928_cast_fp16_2 = split(axis = var_928_axis_0, split_sizes = var_928_split_sizes_0, x = linear_58_cast_fp16)[name = tensor("op_928_cast_fp16")]; + tensor mean_29_axes_0 = const()[name = tensor("mean_29_axes_0"), val = tensor([-1])]; + tensor mean_29_keep_dims_0 = const()[name = tensor("mean_29_keep_dims_0"), val = tensor(true)]; + tensor mean_29_cast_fp16 = reduce_mean(axes = mean_29_axes_0, keep_dims = mean_29_keep_dims_0, x = linear_52_cast_fp16)[name = tensor("mean_29_cast_fp16")]; + tensor sub_26_cast_fp16 = sub(x = linear_52_cast_fp16, y = mean_29_cast_fp16)[name = tensor("sub_26_cast_fp16")]; + tensor square_20_cast_fp16 = square(x = sub_26_cast_fp16)[name = tensor("square_20_cast_fp16")]; + tensor reduce_mean_41_axes_0 = const()[name = tensor("reduce_mean_41_axes_0"), val = tensor([-1])]; + tensor reduce_mean_41_keep_dims_0 = const()[name = tensor("reduce_mean_41_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_41_cast_fp16 = reduce_mean(axes = reduce_mean_41_axes_0, keep_dims = reduce_mean_41_keep_dims_0, x = square_20_cast_fp16)[name = tensor("reduce_mean_41_cast_fp16")]; + tensor var_938_to_fp16 = const()[name = tensor("op_938_to_fp16"), val = tensor(0x1.1p-20)]; + tensor var_939_cast_fp16 = add(x = reduce_mean_41_cast_fp16, y = var_938_to_fp16)[name = tensor("op_939_cast_fp16")]; + tensor var_940_cast_fp16 = sqrt(x = var_939_cast_fp16)[name = tensor("op_940_cast_fp16")]; + tensor x_95_cast_fp16 = real_div(x = sub_26_cast_fp16, y = var_940_cast_fp16)[name = tensor("x_95_cast_fp16")]; + tensor var_942_cast_fp16 = mul(x = x_95_cast_fp16, y = flow_net_res_blocks_0_in_ln_weight_to_fp16)[name = tensor("op_942_cast_fp16")]; + tensor x_97_cast_fp16 = add(x = var_942_cast_fp16, y = flow_net_res_blocks_0_in_ln_bias_to_fp16)[name = tensor("x_97_cast_fp16")]; + tensor var_944_promoted_to_fp16 = const()[name = tensor("op_944_promoted_to_fp16"), val = tensor(0x1p+0)]; + tensor var_945_cast_fp16 = add(x = var_928_cast_fp16_1, y = var_944_promoted_to_fp16)[name = tensor("op_945_cast_fp16")]; + tensor var_946_cast_fp16 = mul(x = x_97_cast_fp16, y = var_945_cast_fp16)[name = tensor("op_946_cast_fp16")]; + tensor input_153_cast_fp16 = add(x = var_946_cast_fp16, y = var_928_cast_fp16_0)[name = tensor("input_153_cast_fp16")]; + tensor linear_59_cast_fp16 = linear(bias = flow_net_res_blocks_0_mlp_0_bias_to_fp16, weight = flow_net_res_blocks_0_mlp_0_weight_to_fp16, x = input_153_cast_fp16)[name = tensor("linear_59_cast_fp16")]; + tensor input_157_cast_fp16 = silu(x = linear_59_cast_fp16)[name = tensor("input_157_cast_fp16")]; + tensor linear_60_cast_fp16 = linear(bias = flow_net_res_blocks_0_mlp_2_bias_to_fp16, weight = flow_net_res_blocks_0_mlp_2_weight_to_fp16, x = input_157_cast_fp16)[name = tensor("linear_60_cast_fp16")]; + tensor var_957_cast_fp16 = mul(x = var_928_cast_fp16_2, y = linear_60_cast_fp16)[name = tensor("op_957_cast_fp16")]; + tensor x_99_cast_fp16 = add(x = linear_52_cast_fp16, y = var_957_cast_fp16)[name = tensor("x_99_cast_fp16")]; + tensor linear_61_cast_fp16 = linear(bias = flow_net_res_blocks_1_adaLN_modulation_1_bias_to_fp16, weight = flow_net_res_blocks_1_adaLN_modulation_1_weight_to_fp16, x = input_151_cast_fp16)[name = tensor("linear_61_cast_fp16")]; + tensor var_967_split_sizes_0 = const()[name = tensor("op_967_split_sizes_0"), val = tensor([512, 512, 512])]; + tensor var_967_axis_0 = const()[name = tensor("op_967_axis_0"), val = tensor(-1)]; + tensor var_967_cast_fp16_0, tensor var_967_cast_fp16_1, tensor var_967_cast_fp16_2 = split(axis = var_967_axis_0, split_sizes = var_967_split_sizes_0, x = linear_61_cast_fp16)[name = tensor("op_967_cast_fp16")]; + tensor mean_31_axes_0 = const()[name = tensor("mean_31_axes_0"), val = tensor([-1])]; + tensor mean_31_keep_dims_0 = const()[name = tensor("mean_31_keep_dims_0"), val = tensor(true)]; + tensor mean_31_cast_fp16 = reduce_mean(axes = mean_31_axes_0, keep_dims = mean_31_keep_dims_0, x = x_99_cast_fp16)[name = tensor("mean_31_cast_fp16")]; + tensor sub_27_cast_fp16 = sub(x = x_99_cast_fp16, y = mean_31_cast_fp16)[name = tensor("sub_27_cast_fp16")]; + tensor square_21_cast_fp16 = square(x = sub_27_cast_fp16)[name = tensor("square_21_cast_fp16")]; + tensor reduce_mean_43_axes_0 = const()[name = tensor("reduce_mean_43_axes_0"), val = tensor([-1])]; + tensor reduce_mean_43_keep_dims_0 = const()[name = tensor("reduce_mean_43_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_43_cast_fp16 = reduce_mean(axes = reduce_mean_43_axes_0, keep_dims = reduce_mean_43_keep_dims_0, x = square_21_cast_fp16)[name = tensor("reduce_mean_43_cast_fp16")]; + tensor var_977_to_fp16 = const()[name = tensor("op_977_to_fp16"), val = tensor(0x1.1p-20)]; + tensor var_978_cast_fp16 = add(x = reduce_mean_43_cast_fp16, y = var_977_to_fp16)[name = tensor("op_978_cast_fp16")]; + tensor var_979_cast_fp16 = sqrt(x = var_978_cast_fp16)[name = tensor("op_979_cast_fp16")]; + tensor x_101_cast_fp16 = real_div(x = sub_27_cast_fp16, y = var_979_cast_fp16)[name = tensor("x_101_cast_fp16")]; + tensor var_981_cast_fp16 = mul(x = x_101_cast_fp16, y = flow_net_res_blocks_1_in_ln_weight_to_fp16)[name = tensor("op_981_cast_fp16")]; + tensor x_103_cast_fp16 = add(x = var_981_cast_fp16, y = flow_net_res_blocks_1_in_ln_bias_to_fp16)[name = tensor("x_103_cast_fp16")]; + tensor var_983_promoted_to_fp16 = const()[name = tensor("op_983_promoted_to_fp16"), val = tensor(0x1p+0)]; + tensor var_984_cast_fp16 = add(x = var_967_cast_fp16_1, y = var_983_promoted_to_fp16)[name = tensor("op_984_cast_fp16")]; + tensor var_985_cast_fp16 = mul(x = x_103_cast_fp16, y = var_984_cast_fp16)[name = tensor("op_985_cast_fp16")]; + tensor input_161_cast_fp16 = add(x = var_985_cast_fp16, y = var_967_cast_fp16_0)[name = tensor("input_161_cast_fp16")]; + tensor linear_62_cast_fp16 = linear(bias = flow_net_res_blocks_1_mlp_0_bias_to_fp16, weight = flow_net_res_blocks_1_mlp_0_weight_to_fp16, x = input_161_cast_fp16)[name = tensor("linear_62_cast_fp16")]; + tensor input_165_cast_fp16 = silu(x = linear_62_cast_fp16)[name = tensor("input_165_cast_fp16")]; + tensor linear_63_cast_fp16 = linear(bias = flow_net_res_blocks_1_mlp_2_bias_to_fp16, weight = flow_net_res_blocks_1_mlp_2_weight_to_fp16, x = input_165_cast_fp16)[name = tensor("linear_63_cast_fp16")]; + tensor var_996_cast_fp16 = mul(x = var_967_cast_fp16_2, y = linear_63_cast_fp16)[name = tensor("op_996_cast_fp16")]; + tensor x_105_cast_fp16 = add(x = x_99_cast_fp16, y = var_996_cast_fp16)[name = tensor("x_105_cast_fp16")]; + tensor linear_64_cast_fp16 = linear(bias = flow_net_res_blocks_2_adaLN_modulation_1_bias_to_fp16, weight = flow_net_res_blocks_2_adaLN_modulation_1_weight_to_fp16, x = input_151_cast_fp16)[name = tensor("linear_64_cast_fp16")]; + tensor var_1006_split_sizes_0 = const()[name = tensor("op_1006_split_sizes_0"), val = tensor([512, 512, 512])]; + tensor var_1006_axis_0 = const()[name = tensor("op_1006_axis_0"), val = tensor(-1)]; + tensor var_1006_cast_fp16_0, tensor var_1006_cast_fp16_1, tensor var_1006_cast_fp16_2 = split(axis = var_1006_axis_0, split_sizes = var_1006_split_sizes_0, x = linear_64_cast_fp16)[name = tensor("op_1006_cast_fp16")]; + tensor mean_33_axes_0 = const()[name = tensor("mean_33_axes_0"), val = tensor([-1])]; + tensor mean_33_keep_dims_0 = const()[name = tensor("mean_33_keep_dims_0"), val = tensor(true)]; + tensor mean_33_cast_fp16 = reduce_mean(axes = mean_33_axes_0, keep_dims = mean_33_keep_dims_0, x = x_105_cast_fp16)[name = tensor("mean_33_cast_fp16")]; + tensor sub_28_cast_fp16 = sub(x = x_105_cast_fp16, y = mean_33_cast_fp16)[name = tensor("sub_28_cast_fp16")]; + tensor square_22_cast_fp16 = square(x = sub_28_cast_fp16)[name = tensor("square_22_cast_fp16")]; + tensor reduce_mean_45_axes_0 = const()[name = tensor("reduce_mean_45_axes_0"), val = tensor([-1])]; + tensor reduce_mean_45_keep_dims_0 = const()[name = tensor("reduce_mean_45_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_45_cast_fp16 = reduce_mean(axes = reduce_mean_45_axes_0, keep_dims = reduce_mean_45_keep_dims_0, x = square_22_cast_fp16)[name = tensor("reduce_mean_45_cast_fp16")]; + tensor var_1016_to_fp16 = const()[name = tensor("op_1016_to_fp16"), val = tensor(0x1.1p-20)]; + tensor var_1017_cast_fp16 = add(x = reduce_mean_45_cast_fp16, y = var_1016_to_fp16)[name = tensor("op_1017_cast_fp16")]; + tensor var_1018_cast_fp16 = sqrt(x = var_1017_cast_fp16)[name = tensor("op_1018_cast_fp16")]; + tensor x_107_cast_fp16 = real_div(x = sub_28_cast_fp16, y = var_1018_cast_fp16)[name = tensor("x_107_cast_fp16")]; + tensor var_1020_cast_fp16 = mul(x = x_107_cast_fp16, y = flow_net_res_blocks_2_in_ln_weight_to_fp16)[name = tensor("op_1020_cast_fp16")]; + tensor x_109_cast_fp16 = add(x = var_1020_cast_fp16, y = flow_net_res_blocks_2_in_ln_bias_to_fp16)[name = tensor("x_109_cast_fp16")]; + tensor var_1022_promoted_to_fp16 = const()[name = tensor("op_1022_promoted_to_fp16"), val = tensor(0x1p+0)]; + tensor var_1023_cast_fp16 = add(x = var_1006_cast_fp16_1, y = var_1022_promoted_to_fp16)[name = tensor("op_1023_cast_fp16")]; + tensor var_1024_cast_fp16 = mul(x = x_109_cast_fp16, y = var_1023_cast_fp16)[name = tensor("op_1024_cast_fp16")]; + tensor input_169_cast_fp16 = add(x = var_1024_cast_fp16, y = var_1006_cast_fp16_0)[name = tensor("input_169_cast_fp16")]; + tensor linear_65_cast_fp16 = linear(bias = flow_net_res_blocks_2_mlp_0_bias_to_fp16, weight = flow_net_res_blocks_2_mlp_0_weight_to_fp16, x = input_169_cast_fp16)[name = tensor("linear_65_cast_fp16")]; + tensor input_173_cast_fp16 = silu(x = linear_65_cast_fp16)[name = tensor("input_173_cast_fp16")]; + tensor linear_66_cast_fp16 = linear(bias = flow_net_res_blocks_2_mlp_2_bias_to_fp16, weight = flow_net_res_blocks_2_mlp_2_weight_to_fp16, x = input_173_cast_fp16)[name = tensor("linear_66_cast_fp16")]; + tensor var_1035_cast_fp16 = mul(x = var_1006_cast_fp16_2, y = linear_66_cast_fp16)[name = tensor("op_1035_cast_fp16")]; + tensor x_111_cast_fp16 = add(x = x_105_cast_fp16, y = var_1035_cast_fp16)[name = tensor("x_111_cast_fp16")]; + tensor linear_67_cast_fp16 = linear(bias = flow_net_res_blocks_3_adaLN_modulation_1_bias_to_fp16, weight = flow_net_res_blocks_3_adaLN_modulation_1_weight_to_fp16, x = input_151_cast_fp16)[name = tensor("linear_67_cast_fp16")]; + tensor var_1045_split_sizes_0 = const()[name = tensor("op_1045_split_sizes_0"), val = tensor([512, 512, 512])]; + tensor var_1045_axis_0 = const()[name = tensor("op_1045_axis_0"), val = tensor(-1)]; + tensor var_1045_cast_fp16_0, tensor var_1045_cast_fp16_1, tensor var_1045_cast_fp16_2 = split(axis = var_1045_axis_0, split_sizes = var_1045_split_sizes_0, x = linear_67_cast_fp16)[name = tensor("op_1045_cast_fp16")]; + tensor mean_35_axes_0 = const()[name = tensor("mean_35_axes_0"), val = tensor([-1])]; + tensor mean_35_keep_dims_0 = const()[name = tensor("mean_35_keep_dims_0"), val = tensor(true)]; + tensor mean_35_cast_fp16 = reduce_mean(axes = mean_35_axes_0, keep_dims = mean_35_keep_dims_0, x = x_111_cast_fp16)[name = tensor("mean_35_cast_fp16")]; + tensor sub_29_cast_fp16 = sub(x = x_111_cast_fp16, y = mean_35_cast_fp16)[name = tensor("sub_29_cast_fp16")]; + tensor square_23_cast_fp16 = square(x = sub_29_cast_fp16)[name = tensor("square_23_cast_fp16")]; + tensor reduce_mean_47_axes_0 = const()[name = tensor("reduce_mean_47_axes_0"), val = tensor([-1])]; + tensor reduce_mean_47_keep_dims_0 = const()[name = tensor("reduce_mean_47_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_47_cast_fp16 = reduce_mean(axes = reduce_mean_47_axes_0, keep_dims = reduce_mean_47_keep_dims_0, x = square_23_cast_fp16)[name = tensor("reduce_mean_47_cast_fp16")]; + tensor var_1055_to_fp16 = const()[name = tensor("op_1055_to_fp16"), val = tensor(0x1.1p-20)]; + tensor var_1056_cast_fp16 = add(x = reduce_mean_47_cast_fp16, y = var_1055_to_fp16)[name = tensor("op_1056_cast_fp16")]; + tensor var_1057_cast_fp16 = sqrt(x = var_1056_cast_fp16)[name = tensor("op_1057_cast_fp16")]; + tensor x_113_cast_fp16 = real_div(x = sub_29_cast_fp16, y = var_1057_cast_fp16)[name = tensor("x_113_cast_fp16")]; + tensor var_1059_cast_fp16 = mul(x = x_113_cast_fp16, y = flow_net_res_blocks_3_in_ln_weight_to_fp16)[name = tensor("op_1059_cast_fp16")]; + tensor x_115_cast_fp16 = add(x = var_1059_cast_fp16, y = flow_net_res_blocks_3_in_ln_bias_to_fp16)[name = tensor("x_115_cast_fp16")]; + tensor var_1061_promoted_to_fp16 = const()[name = tensor("op_1061_promoted_to_fp16"), val = tensor(0x1p+0)]; + tensor var_1062_cast_fp16 = add(x = var_1045_cast_fp16_1, y = var_1061_promoted_to_fp16)[name = tensor("op_1062_cast_fp16")]; + tensor var_1063_cast_fp16 = mul(x = x_115_cast_fp16, y = var_1062_cast_fp16)[name = tensor("op_1063_cast_fp16")]; + tensor input_177_cast_fp16 = add(x = var_1063_cast_fp16, y = var_1045_cast_fp16_0)[name = tensor("input_177_cast_fp16")]; + tensor linear_68_cast_fp16 = linear(bias = flow_net_res_blocks_3_mlp_0_bias_to_fp16, weight = flow_net_res_blocks_3_mlp_0_weight_to_fp16, x = input_177_cast_fp16)[name = tensor("linear_68_cast_fp16")]; + tensor input_181_cast_fp16 = silu(x = linear_68_cast_fp16)[name = tensor("input_181_cast_fp16")]; + tensor linear_69_cast_fp16 = linear(bias = flow_net_res_blocks_3_mlp_2_bias_to_fp16, weight = flow_net_res_blocks_3_mlp_2_weight_to_fp16, x = input_181_cast_fp16)[name = tensor("linear_69_cast_fp16")]; + tensor var_1074_cast_fp16 = mul(x = var_1045_cast_fp16_2, y = linear_69_cast_fp16)[name = tensor("op_1074_cast_fp16")]; + tensor x_117_cast_fp16 = add(x = x_111_cast_fp16, y = var_1074_cast_fp16)[name = tensor("x_117_cast_fp16")]; + tensor linear_70_cast_fp16 = linear(bias = flow_net_res_blocks_4_adaLN_modulation_1_bias_to_fp16, weight = flow_net_res_blocks_4_adaLN_modulation_1_weight_to_fp16, x = input_151_cast_fp16)[name = tensor("linear_70_cast_fp16")]; + tensor var_1084_split_sizes_0 = const()[name = tensor("op_1084_split_sizes_0"), val = tensor([512, 512, 512])]; + tensor var_1084_axis_0 = const()[name = tensor("op_1084_axis_0"), val = tensor(-1)]; + tensor var_1084_cast_fp16_0, tensor var_1084_cast_fp16_1, tensor var_1084_cast_fp16_2 = split(axis = var_1084_axis_0, split_sizes = var_1084_split_sizes_0, x = linear_70_cast_fp16)[name = tensor("op_1084_cast_fp16")]; + tensor mean_37_axes_0 = const()[name = tensor("mean_37_axes_0"), val = tensor([-1])]; + tensor mean_37_keep_dims_0 = const()[name = tensor("mean_37_keep_dims_0"), val = tensor(true)]; + tensor mean_37_cast_fp16 = reduce_mean(axes = mean_37_axes_0, keep_dims = mean_37_keep_dims_0, x = x_117_cast_fp16)[name = tensor("mean_37_cast_fp16")]; + tensor sub_30_cast_fp16 = sub(x = x_117_cast_fp16, y = mean_37_cast_fp16)[name = tensor("sub_30_cast_fp16")]; + tensor square_24_cast_fp16 = square(x = sub_30_cast_fp16)[name = tensor("square_24_cast_fp16")]; + tensor reduce_mean_49_axes_0 = const()[name = tensor("reduce_mean_49_axes_0"), val = tensor([-1])]; + tensor reduce_mean_49_keep_dims_0 = const()[name = tensor("reduce_mean_49_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_49_cast_fp16 = reduce_mean(axes = reduce_mean_49_axes_0, keep_dims = reduce_mean_49_keep_dims_0, x = square_24_cast_fp16)[name = tensor("reduce_mean_49_cast_fp16")]; + tensor var_1094_to_fp16 = const()[name = tensor("op_1094_to_fp16"), val = tensor(0x1.1p-20)]; + tensor var_1095_cast_fp16 = add(x = reduce_mean_49_cast_fp16, y = var_1094_to_fp16)[name = tensor("op_1095_cast_fp16")]; + tensor var_1096_cast_fp16 = sqrt(x = var_1095_cast_fp16)[name = tensor("op_1096_cast_fp16")]; + tensor x_119_cast_fp16 = real_div(x = sub_30_cast_fp16, y = var_1096_cast_fp16)[name = tensor("x_119_cast_fp16")]; + tensor var_1098_cast_fp16 = mul(x = x_119_cast_fp16, y = flow_net_res_blocks_4_in_ln_weight_to_fp16)[name = tensor("op_1098_cast_fp16")]; + tensor x_121_cast_fp16 = add(x = var_1098_cast_fp16, y = flow_net_res_blocks_4_in_ln_bias_to_fp16)[name = tensor("x_121_cast_fp16")]; + tensor var_1100_promoted_to_fp16 = const()[name = tensor("op_1100_promoted_to_fp16"), val = tensor(0x1p+0)]; + tensor var_1101_cast_fp16 = add(x = var_1084_cast_fp16_1, y = var_1100_promoted_to_fp16)[name = tensor("op_1101_cast_fp16")]; + tensor var_1102_cast_fp16 = mul(x = x_121_cast_fp16, y = var_1101_cast_fp16)[name = tensor("op_1102_cast_fp16")]; + tensor input_185_cast_fp16 = add(x = var_1102_cast_fp16, y = var_1084_cast_fp16_0)[name = tensor("input_185_cast_fp16")]; + tensor linear_71_cast_fp16 = linear(bias = flow_net_res_blocks_4_mlp_0_bias_to_fp16, weight = flow_net_res_blocks_4_mlp_0_weight_to_fp16, x = input_185_cast_fp16)[name = tensor("linear_71_cast_fp16")]; + tensor input_189_cast_fp16 = silu(x = linear_71_cast_fp16)[name = tensor("input_189_cast_fp16")]; + tensor linear_72_cast_fp16 = linear(bias = flow_net_res_blocks_4_mlp_2_bias_to_fp16, weight = flow_net_res_blocks_4_mlp_2_weight_to_fp16, x = input_189_cast_fp16)[name = tensor("linear_72_cast_fp16")]; + tensor var_1113_cast_fp16 = mul(x = var_1084_cast_fp16_2, y = linear_72_cast_fp16)[name = tensor("op_1113_cast_fp16")]; + tensor x_123_cast_fp16 = add(x = x_117_cast_fp16, y = var_1113_cast_fp16)[name = tensor("x_123_cast_fp16")]; + tensor linear_73_cast_fp16 = linear(bias = flow_net_res_blocks_5_adaLN_modulation_1_bias_to_fp16, weight = flow_net_res_blocks_5_adaLN_modulation_1_weight_to_fp16, x = input_151_cast_fp16)[name = tensor("linear_73_cast_fp16")]; + tensor var_1123_split_sizes_0 = const()[name = tensor("op_1123_split_sizes_0"), val = tensor([512, 512, 512])]; + tensor var_1123_axis_0 = const()[name = tensor("op_1123_axis_0"), val = tensor(-1)]; + tensor var_1123_cast_fp16_0, tensor var_1123_cast_fp16_1, tensor var_1123_cast_fp16_2 = split(axis = var_1123_axis_0, split_sizes = var_1123_split_sizes_0, x = linear_73_cast_fp16)[name = tensor("op_1123_cast_fp16")]; + tensor mean_39_axes_0 = const()[name = tensor("mean_39_axes_0"), val = tensor([-1])]; + tensor mean_39_keep_dims_0 = const()[name = tensor("mean_39_keep_dims_0"), val = tensor(true)]; + tensor mean_39_cast_fp16 = reduce_mean(axes = mean_39_axes_0, keep_dims = mean_39_keep_dims_0, x = x_123_cast_fp16)[name = tensor("mean_39_cast_fp16")]; + tensor sub_31_cast_fp16 = sub(x = x_123_cast_fp16, y = mean_39_cast_fp16)[name = tensor("sub_31_cast_fp16")]; + tensor square_25_cast_fp16 = square(x = sub_31_cast_fp16)[name = tensor("square_25_cast_fp16")]; + tensor reduce_mean_51_axes_0 = const()[name = tensor("reduce_mean_51_axes_0"), val = tensor([-1])]; + tensor reduce_mean_51_keep_dims_0 = const()[name = tensor("reduce_mean_51_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_51_cast_fp16 = reduce_mean(axes = reduce_mean_51_axes_0, keep_dims = reduce_mean_51_keep_dims_0, x = square_25_cast_fp16)[name = tensor("reduce_mean_51_cast_fp16")]; + tensor var_1133_to_fp16 = const()[name = tensor("op_1133_to_fp16"), val = tensor(0x1.1p-20)]; + tensor var_1134_cast_fp16 = add(x = reduce_mean_51_cast_fp16, y = var_1133_to_fp16)[name = tensor("op_1134_cast_fp16")]; + tensor var_1135_cast_fp16 = sqrt(x = var_1134_cast_fp16)[name = tensor("op_1135_cast_fp16")]; + tensor x_125_cast_fp16 = real_div(x = sub_31_cast_fp16, y = var_1135_cast_fp16)[name = tensor("x_125_cast_fp16")]; + tensor var_1137_cast_fp16 = mul(x = x_125_cast_fp16, y = flow_net_res_blocks_5_in_ln_weight_to_fp16)[name = tensor("op_1137_cast_fp16")]; + tensor x_127_cast_fp16 = add(x = var_1137_cast_fp16, y = flow_net_res_blocks_5_in_ln_bias_to_fp16)[name = tensor("x_127_cast_fp16")]; + tensor var_1139_promoted_to_fp16 = const()[name = tensor("op_1139_promoted_to_fp16"), val = tensor(0x1p+0)]; + tensor var_1140_cast_fp16 = add(x = var_1123_cast_fp16_1, y = var_1139_promoted_to_fp16)[name = tensor("op_1140_cast_fp16")]; + tensor var_1141_cast_fp16 = mul(x = x_127_cast_fp16, y = var_1140_cast_fp16)[name = tensor("op_1141_cast_fp16")]; + tensor input_193_cast_fp16 = add(x = var_1141_cast_fp16, y = var_1123_cast_fp16_0)[name = tensor("input_193_cast_fp16")]; + tensor linear_74_cast_fp16 = linear(bias = flow_net_res_blocks_5_mlp_0_bias_to_fp16, weight = flow_net_res_blocks_5_mlp_0_weight_to_fp16, x = input_193_cast_fp16)[name = tensor("linear_74_cast_fp16")]; + tensor input_197_cast_fp16 = silu(x = linear_74_cast_fp16)[name = tensor("input_197_cast_fp16")]; + tensor linear_75_cast_fp16 = linear(bias = flow_net_res_blocks_5_mlp_2_bias_to_fp16, weight = flow_net_res_blocks_5_mlp_2_weight_to_fp16, x = input_197_cast_fp16)[name = tensor("linear_75_cast_fp16")]; + tensor var_1152_cast_fp16 = mul(x = var_1123_cast_fp16_2, y = linear_75_cast_fp16)[name = tensor("op_1152_cast_fp16")]; + tensor x_129_cast_fp16 = add(x = x_123_cast_fp16, y = var_1152_cast_fp16)[name = tensor("x_129_cast_fp16")]; + tensor linear_76_cast_fp16 = linear(bias = flow_net_final_layer_adaLN_modulation_1_bias_to_fp16, weight = flow_net_final_layer_adaLN_modulation_1_weight_to_fp16, x = input_151_cast_fp16)[name = tensor("linear_76_cast_fp16")]; + tensor var_1161_split_sizes_0 = const()[name = tensor("op_1161_split_sizes_0"), val = tensor([512, 512])]; + tensor var_1161_axis_0 = const()[name = tensor("op_1161_axis_0"), val = tensor(-1)]; + tensor var_1161_cast_fp16_0, tensor var_1161_cast_fp16_1 = split(axis = var_1161_axis_0, split_sizes = var_1161_split_sizes_0, x = linear_76_cast_fp16)[name = tensor("op_1161_cast_fp16")]; + tensor mean_41_axes_0 = const()[name = tensor("mean_41_axes_0"), val = tensor([-1])]; + tensor mean_41_keep_dims_0 = const()[name = tensor("mean_41_keep_dims_0"), val = tensor(true)]; + tensor mean_41_cast_fp16 = reduce_mean(axes = mean_41_axes_0, keep_dims = mean_41_keep_dims_0, x = x_129_cast_fp16)[name = tensor("mean_41_cast_fp16")]; + tensor sub_32_cast_fp16 = sub(x = x_129_cast_fp16, y = mean_41_cast_fp16)[name = tensor("sub_32_cast_fp16")]; + tensor square_26_cast_fp16 = square(x = sub_32_cast_fp16)[name = tensor("square_26_cast_fp16")]; + tensor reduce_mean_53_axes_0 = const()[name = tensor("reduce_mean_53_axes_0"), val = tensor([-1])]; + tensor reduce_mean_53_keep_dims_0 = const()[name = tensor("reduce_mean_53_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_53_cast_fp16 = reduce_mean(axes = reduce_mean_53_axes_0, keep_dims = reduce_mean_53_keep_dims_0, x = square_26_cast_fp16)[name = tensor("reduce_mean_53_cast_fp16")]; + tensor var_1168_to_fp16 = const()[name = tensor("op_1168_to_fp16"), val = tensor(0x1.1p-20)]; + tensor var_1169_cast_fp16 = add(x = reduce_mean_53_cast_fp16, y = var_1168_to_fp16)[name = tensor("op_1169_cast_fp16")]; + tensor var_1170_cast_fp16 = sqrt(x = var_1169_cast_fp16)[name = tensor("op_1170_cast_fp16")]; + tensor x_131_cast_fp16 = real_div(x = sub_32_cast_fp16, y = var_1170_cast_fp16)[name = tensor("x_131_cast_fp16")]; + tensor var_1172_promoted_to_fp16 = const()[name = tensor("op_1172_promoted_to_fp16"), val = tensor(0x1p+0)]; + tensor var_1173_cast_fp16 = add(x = var_1161_cast_fp16_1, y = var_1172_promoted_to_fp16)[name = tensor("op_1173_cast_fp16")]; + tensor var_1174_cast_fp16 = mul(x = x_131_cast_fp16, y = var_1173_cast_fp16)[name = tensor("op_1174_cast_fp16")]; + tensor input_201_cast_fp16 = add(x = var_1174_cast_fp16, y = var_1161_cast_fp16_0)[name = tensor("input_201_cast_fp16")]; + tensor linear_77_cast_fp16 = linear(bias = flow_net_final_layer_linear_bias_to_fp16, weight = flow_net_final_layer_linear_weight_to_fp16, x = input_201_cast_fp16)[name = tensor("linear_77_cast_fp16")]; + tensor var_1185_to_fp16 = const()[name = tensor("op_1185_to_fp16"), val = tensor(0x1p-3)]; + tensor var_1186_cast_fp16 = mul(x = linear_77_cast_fp16, y = var_1185_to_fp16)[name = tensor("op_1186_cast_fp16")]; + tensor input_203_cast_fp16 = add(x = input_135_cast_fp16, y = var_1186_cast_fp16)[name = tensor("input_203_cast_fp16")]; + tensor linear_78_cast_fp16 = linear(bias = flow_net_input_proj_bias_to_fp16, weight = flow_net_input_proj_weight_to_fp16, x = input_203_cast_fp16)[name = tensor("linear_78_cast_fp16")]; + tensor input_207_to_fp16 = const()[name = tensor("input_207_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(19001728)))]; + tensor input_209_cast_fp16 = silu(x = input_207_to_fp16)[name = tensor("input_209_cast_fp16")]; + tensor linear_80_cast_fp16 = linear(bias = flow_net_time_embed_0_mlp_2_bias_to_fp16, weight = flow_net_time_embed_0_mlp_2_weight_to_fp16, x = input_209_cast_fp16)[name = tensor("linear_80_cast_fp16")]; + tensor reduce_mean_54_axes_0 = const()[name = tensor("reduce_mean_54_axes_0"), val = tensor([-1])]; + tensor reduce_mean_54_keep_dims_0 = const()[name = tensor("reduce_mean_54_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_54_cast_fp16 = reduce_mean(axes = reduce_mean_54_axes_0, keep_dims = reduce_mean_54_keep_dims_0, x = linear_80_cast_fp16)[name = tensor("reduce_mean_54_cast_fp16")]; + tensor sub_33_cast_fp16 = sub(x = linear_80_cast_fp16, y = reduce_mean_54_cast_fp16)[name = tensor("sub_33_cast_fp16")]; + tensor square_27_cast_fp16 = square(x = sub_33_cast_fp16)[name = tensor("square_27_cast_fp16")]; + tensor reduce_mean_55_axes_0 = const()[name = tensor("reduce_mean_55_axes_0"), val = tensor([-1])]; + tensor reduce_mean_55_keep_dims_0 = const()[name = tensor("reduce_mean_55_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_55_cast_fp16 = reduce_mean(axes = reduce_mean_55_axes_0, keep_dims = reduce_mean_55_keep_dims_0, x = square_27_cast_fp16)[name = tensor("reduce_mean_55_cast_fp16")]; + tensor real_div_6_to_fp16 = const()[name = tensor("real_div_6_to_fp16"), val = tensor(0x1.008p+0)]; + tensor mul_6_cast_fp16 = mul(x = reduce_mean_55_cast_fp16, y = real_div_6_to_fp16)[name = tensor("mul_6_cast_fp16")]; + tensor var_1252_to_fp16 = const()[name = tensor("op_1252_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_55_cast_fp16 = add(x = mul_6_cast_fp16, y = var_1252_to_fp16)[name = tensor("var_55_cast_fp16")]; + tensor var_1255_epsilon_0 = const()[name = tensor("op_1255_epsilon_0"), val = tensor(0x1.197998p-40)]; + tensor var_1255_cast_fp16 = rsqrt(epsilon = var_1255_epsilon_0, x = var_55_cast_fp16)[name = tensor("op_1255_cast_fp16")]; + tensor var_1256_cast_fp16 = mul(x = const_3_to_fp16, y = var_1255_cast_fp16)[name = tensor("op_1256_cast_fp16")]; + tensor var_1257_cast_fp16 = mul(x = linear_80_cast_fp16, y = var_1256_cast_fp16)[name = tensor("op_1257_cast_fp16")]; + tensor input_213_to_fp16 = const()[name = tensor("input_213_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(19002816)))]; + tensor input_215_cast_fp16 = silu(x = input_213_to_fp16)[name = tensor("input_215_cast_fp16")]; + tensor linear_82_cast_fp16 = linear(bias = flow_net_time_embed_1_mlp_2_bias_to_fp16, weight = flow_net_time_embed_1_mlp_2_weight_to_fp16, x = input_215_cast_fp16)[name = tensor("linear_82_cast_fp16")]; + tensor reduce_mean_56_axes_0 = const()[name = tensor("reduce_mean_56_axes_0"), val = tensor([-1])]; + tensor reduce_mean_56_keep_dims_0 = const()[name = tensor("reduce_mean_56_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_56_cast_fp16 = reduce_mean(axes = reduce_mean_56_axes_0, keep_dims = reduce_mean_56_keep_dims_0, x = linear_82_cast_fp16)[name = tensor("reduce_mean_56_cast_fp16")]; + tensor sub_35_cast_fp16 = sub(x = linear_82_cast_fp16, y = reduce_mean_56_cast_fp16)[name = tensor("sub_35_cast_fp16")]; + tensor square_28_cast_fp16 = square(x = sub_35_cast_fp16)[name = tensor("square_28_cast_fp16")]; + tensor reduce_mean_57_axes_0 = const()[name = tensor("reduce_mean_57_axes_0"), val = tensor([-1])]; + tensor reduce_mean_57_keep_dims_0 = const()[name = tensor("reduce_mean_57_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_57_cast_fp16 = reduce_mean(axes = reduce_mean_57_axes_0, keep_dims = reduce_mean_57_keep_dims_0, x = square_28_cast_fp16)[name = tensor("reduce_mean_57_cast_fp16")]; + tensor real_div_7_to_fp16 = const()[name = tensor("real_div_7_to_fp16"), val = tensor(0x1.008p+0)]; + tensor mul_7_cast_fp16 = mul(x = reduce_mean_57_cast_fp16, y = real_div_7_to_fp16)[name = tensor("mul_7_cast_fp16")]; + tensor var_1289_to_fp16 = const()[name = tensor("op_1289_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_57_cast_fp16 = add(x = mul_7_cast_fp16, y = var_1289_to_fp16)[name = tensor("var_57_cast_fp16")]; + tensor var_1292_epsilon_0 = const()[name = tensor("op_1292_epsilon_0"), val = tensor(0x1.197998p-40)]; + tensor var_1292_cast_fp16 = rsqrt(epsilon = var_1292_epsilon_0, x = var_57_cast_fp16)[name = tensor("op_1292_cast_fp16")]; + tensor var_1293_cast_fp16 = mul(x = const_5_to_fp16, y = var_1292_cast_fp16)[name = tensor("op_1293_cast_fp16")]; + tensor var_1294_cast_fp16 = mul(x = linear_82_cast_fp16, y = var_1293_cast_fp16)[name = tensor("op_1294_cast_fp16")]; + tensor var_1306_cast_fp16 = add(x = var_1257_cast_fp16, y = var_1294_cast_fp16)[name = tensor("op_1306_cast_fp16")]; + tensor _inversed_t_combined_7_y_0_to_fp16 = const()[name = tensor("_inversed_t_combined_7_y_0_to_fp16"), val = tensor(0x1p-1)]; + tensor _inversed_t_combined_7_cast_fp16 = mul(x = var_1306_cast_fp16, y = _inversed_t_combined_7_y_0_to_fp16)[name = tensor("_inversed_t_combined_7_cast_fp16")]; + tensor input_217_cast_fp16 = add(x = _inversed_t_combined_7_cast_fp16, y = linear_5_cast_fp16)[name = tensor("input_217_cast_fp16")]; + tensor input_219_cast_fp16 = silu(x = input_217_cast_fp16)[name = tensor("input_219_cast_fp16")]; + tensor linear_84_cast_fp16 = linear(bias = flow_net_res_blocks_0_adaLN_modulation_1_bias_to_fp16, weight = flow_net_res_blocks_0_adaLN_modulation_1_weight_to_fp16, x = input_219_cast_fp16)[name = tensor("linear_84_cast_fp16")]; + tensor var_1321_split_sizes_0 = const()[name = tensor("op_1321_split_sizes_0"), val = tensor([512, 512, 512])]; + tensor var_1321_axis_0 = const()[name = tensor("op_1321_axis_0"), val = tensor(-1)]; + tensor var_1321_cast_fp16_0, tensor var_1321_cast_fp16_1, tensor var_1321_cast_fp16_2 = split(axis = var_1321_axis_0, split_sizes = var_1321_split_sizes_0, x = linear_84_cast_fp16)[name = tensor("op_1321_cast_fp16")]; + tensor mean_43_axes_0 = const()[name = tensor("mean_43_axes_0"), val = tensor([-1])]; + tensor mean_43_keep_dims_0 = const()[name = tensor("mean_43_keep_dims_0"), val = tensor(true)]; + tensor mean_43_cast_fp16 = reduce_mean(axes = mean_43_axes_0, keep_dims = mean_43_keep_dims_0, x = linear_78_cast_fp16)[name = tensor("mean_43_cast_fp16")]; + tensor sub_37_cast_fp16 = sub(x = linear_78_cast_fp16, y = mean_43_cast_fp16)[name = tensor("sub_37_cast_fp16")]; + tensor square_29_cast_fp16 = square(x = sub_37_cast_fp16)[name = tensor("square_29_cast_fp16")]; + tensor reduce_mean_59_axes_0 = const()[name = tensor("reduce_mean_59_axes_0"), val = tensor([-1])]; + tensor reduce_mean_59_keep_dims_0 = const()[name = tensor("reduce_mean_59_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_59_cast_fp16 = reduce_mean(axes = reduce_mean_59_axes_0, keep_dims = reduce_mean_59_keep_dims_0, x = square_29_cast_fp16)[name = tensor("reduce_mean_59_cast_fp16")]; + tensor var_1331_to_fp16 = const()[name = tensor("op_1331_to_fp16"), val = tensor(0x1.1p-20)]; + tensor var_1332_cast_fp16 = add(x = reduce_mean_59_cast_fp16, y = var_1331_to_fp16)[name = tensor("op_1332_cast_fp16")]; + tensor var_1333_cast_fp16 = sqrt(x = var_1332_cast_fp16)[name = tensor("op_1333_cast_fp16")]; + tensor x_139_cast_fp16 = real_div(x = sub_37_cast_fp16, y = var_1333_cast_fp16)[name = tensor("x_139_cast_fp16")]; + tensor var_1335_cast_fp16 = mul(x = x_139_cast_fp16, y = flow_net_res_blocks_0_in_ln_weight_to_fp16)[name = tensor("op_1335_cast_fp16")]; + tensor x_141_cast_fp16 = add(x = var_1335_cast_fp16, y = flow_net_res_blocks_0_in_ln_bias_to_fp16)[name = tensor("x_141_cast_fp16")]; + tensor var_1337_promoted_to_fp16 = const()[name = tensor("op_1337_promoted_to_fp16"), val = tensor(0x1p+0)]; + tensor var_1338_cast_fp16 = add(x = var_1321_cast_fp16_1, y = var_1337_promoted_to_fp16)[name = tensor("op_1338_cast_fp16")]; + tensor var_1339_cast_fp16 = mul(x = x_141_cast_fp16, y = var_1338_cast_fp16)[name = tensor("op_1339_cast_fp16")]; + tensor input_221_cast_fp16 = add(x = var_1339_cast_fp16, y = var_1321_cast_fp16_0)[name = tensor("input_221_cast_fp16")]; + tensor linear_85_cast_fp16 = linear(bias = flow_net_res_blocks_0_mlp_0_bias_to_fp16, weight = flow_net_res_blocks_0_mlp_0_weight_to_fp16, x = input_221_cast_fp16)[name = tensor("linear_85_cast_fp16")]; + tensor input_225_cast_fp16 = silu(x = linear_85_cast_fp16)[name = tensor("input_225_cast_fp16")]; + tensor linear_86_cast_fp16 = linear(bias = flow_net_res_blocks_0_mlp_2_bias_to_fp16, weight = flow_net_res_blocks_0_mlp_2_weight_to_fp16, x = input_225_cast_fp16)[name = tensor("linear_86_cast_fp16")]; + tensor var_1350_cast_fp16 = mul(x = var_1321_cast_fp16_2, y = linear_86_cast_fp16)[name = tensor("op_1350_cast_fp16")]; + tensor x_143_cast_fp16 = add(x = linear_78_cast_fp16, y = var_1350_cast_fp16)[name = tensor("x_143_cast_fp16")]; + tensor linear_87_cast_fp16 = linear(bias = flow_net_res_blocks_1_adaLN_modulation_1_bias_to_fp16, weight = flow_net_res_blocks_1_adaLN_modulation_1_weight_to_fp16, x = input_219_cast_fp16)[name = tensor("linear_87_cast_fp16")]; + tensor var_1360_split_sizes_0 = const()[name = tensor("op_1360_split_sizes_0"), val = tensor([512, 512, 512])]; + tensor var_1360_axis_0 = const()[name = tensor("op_1360_axis_0"), val = tensor(-1)]; + tensor var_1360_cast_fp16_0, tensor var_1360_cast_fp16_1, tensor var_1360_cast_fp16_2 = split(axis = var_1360_axis_0, split_sizes = var_1360_split_sizes_0, x = linear_87_cast_fp16)[name = tensor("op_1360_cast_fp16")]; + tensor mean_45_axes_0 = const()[name = tensor("mean_45_axes_0"), val = tensor([-1])]; + tensor mean_45_keep_dims_0 = const()[name = tensor("mean_45_keep_dims_0"), val = tensor(true)]; + tensor mean_45_cast_fp16 = reduce_mean(axes = mean_45_axes_0, keep_dims = mean_45_keep_dims_0, x = x_143_cast_fp16)[name = tensor("mean_45_cast_fp16")]; + tensor sub_38_cast_fp16 = sub(x = x_143_cast_fp16, y = mean_45_cast_fp16)[name = tensor("sub_38_cast_fp16")]; + tensor square_30_cast_fp16 = square(x = sub_38_cast_fp16)[name = tensor("square_30_cast_fp16")]; + tensor reduce_mean_61_axes_0 = const()[name = tensor("reduce_mean_61_axes_0"), val = tensor([-1])]; + tensor reduce_mean_61_keep_dims_0 = const()[name = tensor("reduce_mean_61_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_61_cast_fp16 = reduce_mean(axes = reduce_mean_61_axes_0, keep_dims = reduce_mean_61_keep_dims_0, x = square_30_cast_fp16)[name = tensor("reduce_mean_61_cast_fp16")]; + tensor var_1370_to_fp16 = const()[name = tensor("op_1370_to_fp16"), val = tensor(0x1.1p-20)]; + tensor var_1371_cast_fp16 = add(x = reduce_mean_61_cast_fp16, y = var_1370_to_fp16)[name = tensor("op_1371_cast_fp16")]; + tensor var_1372_cast_fp16 = sqrt(x = var_1371_cast_fp16)[name = tensor("op_1372_cast_fp16")]; + tensor x_145_cast_fp16 = real_div(x = sub_38_cast_fp16, y = var_1372_cast_fp16)[name = tensor("x_145_cast_fp16")]; + tensor var_1374_cast_fp16 = mul(x = x_145_cast_fp16, y = flow_net_res_blocks_1_in_ln_weight_to_fp16)[name = tensor("op_1374_cast_fp16")]; + tensor x_147_cast_fp16 = add(x = var_1374_cast_fp16, y = flow_net_res_blocks_1_in_ln_bias_to_fp16)[name = tensor("x_147_cast_fp16")]; + tensor var_1376_promoted_to_fp16 = const()[name = tensor("op_1376_promoted_to_fp16"), val = tensor(0x1p+0)]; + tensor var_1377_cast_fp16 = add(x = var_1360_cast_fp16_1, y = var_1376_promoted_to_fp16)[name = tensor("op_1377_cast_fp16")]; + tensor var_1378_cast_fp16 = mul(x = x_147_cast_fp16, y = var_1377_cast_fp16)[name = tensor("op_1378_cast_fp16")]; + tensor input_229_cast_fp16 = add(x = var_1378_cast_fp16, y = var_1360_cast_fp16_0)[name = tensor("input_229_cast_fp16")]; + tensor linear_88_cast_fp16 = linear(bias = flow_net_res_blocks_1_mlp_0_bias_to_fp16, weight = flow_net_res_blocks_1_mlp_0_weight_to_fp16, x = input_229_cast_fp16)[name = tensor("linear_88_cast_fp16")]; + tensor input_233_cast_fp16 = silu(x = linear_88_cast_fp16)[name = tensor("input_233_cast_fp16")]; + tensor linear_89_cast_fp16 = linear(bias = flow_net_res_blocks_1_mlp_2_bias_to_fp16, weight = flow_net_res_blocks_1_mlp_2_weight_to_fp16, x = input_233_cast_fp16)[name = tensor("linear_89_cast_fp16")]; + tensor var_1389_cast_fp16 = mul(x = var_1360_cast_fp16_2, y = linear_89_cast_fp16)[name = tensor("op_1389_cast_fp16")]; + tensor x_149_cast_fp16 = add(x = x_143_cast_fp16, y = var_1389_cast_fp16)[name = tensor("x_149_cast_fp16")]; + tensor linear_90_cast_fp16 = linear(bias = flow_net_res_blocks_2_adaLN_modulation_1_bias_to_fp16, weight = flow_net_res_blocks_2_adaLN_modulation_1_weight_to_fp16, x = input_219_cast_fp16)[name = tensor("linear_90_cast_fp16")]; + tensor var_1399_split_sizes_0 = const()[name = tensor("op_1399_split_sizes_0"), val = tensor([512, 512, 512])]; + tensor var_1399_axis_0 = const()[name = tensor("op_1399_axis_0"), val = tensor(-1)]; + tensor var_1399_cast_fp16_0, tensor var_1399_cast_fp16_1, tensor var_1399_cast_fp16_2 = split(axis = var_1399_axis_0, split_sizes = var_1399_split_sizes_0, x = linear_90_cast_fp16)[name = tensor("op_1399_cast_fp16")]; + tensor mean_47_axes_0 = const()[name = tensor("mean_47_axes_0"), val = tensor([-1])]; + tensor mean_47_keep_dims_0 = const()[name = tensor("mean_47_keep_dims_0"), val = tensor(true)]; + tensor mean_47_cast_fp16 = reduce_mean(axes = mean_47_axes_0, keep_dims = mean_47_keep_dims_0, x = x_149_cast_fp16)[name = tensor("mean_47_cast_fp16")]; + tensor sub_39_cast_fp16 = sub(x = x_149_cast_fp16, y = mean_47_cast_fp16)[name = tensor("sub_39_cast_fp16")]; + tensor square_31_cast_fp16 = square(x = sub_39_cast_fp16)[name = tensor("square_31_cast_fp16")]; + tensor reduce_mean_63_axes_0 = const()[name = tensor("reduce_mean_63_axes_0"), val = tensor([-1])]; + tensor reduce_mean_63_keep_dims_0 = const()[name = tensor("reduce_mean_63_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_63_cast_fp16 = reduce_mean(axes = reduce_mean_63_axes_0, keep_dims = reduce_mean_63_keep_dims_0, x = square_31_cast_fp16)[name = tensor("reduce_mean_63_cast_fp16")]; + tensor var_1409_to_fp16 = const()[name = tensor("op_1409_to_fp16"), val = tensor(0x1.1p-20)]; + tensor var_1410_cast_fp16 = add(x = reduce_mean_63_cast_fp16, y = var_1409_to_fp16)[name = tensor("op_1410_cast_fp16")]; + tensor var_1411_cast_fp16 = sqrt(x = var_1410_cast_fp16)[name = tensor("op_1411_cast_fp16")]; + tensor x_151_cast_fp16 = real_div(x = sub_39_cast_fp16, y = var_1411_cast_fp16)[name = tensor("x_151_cast_fp16")]; + tensor var_1413_cast_fp16 = mul(x = x_151_cast_fp16, y = flow_net_res_blocks_2_in_ln_weight_to_fp16)[name = tensor("op_1413_cast_fp16")]; + tensor x_153_cast_fp16 = add(x = var_1413_cast_fp16, y = flow_net_res_blocks_2_in_ln_bias_to_fp16)[name = tensor("x_153_cast_fp16")]; + tensor var_1415_promoted_to_fp16 = const()[name = tensor("op_1415_promoted_to_fp16"), val = tensor(0x1p+0)]; + tensor var_1416_cast_fp16 = add(x = var_1399_cast_fp16_1, y = var_1415_promoted_to_fp16)[name = tensor("op_1416_cast_fp16")]; + tensor var_1417_cast_fp16 = mul(x = x_153_cast_fp16, y = var_1416_cast_fp16)[name = tensor("op_1417_cast_fp16")]; + tensor input_237_cast_fp16 = add(x = var_1417_cast_fp16, y = var_1399_cast_fp16_0)[name = tensor("input_237_cast_fp16")]; + tensor linear_91_cast_fp16 = linear(bias = flow_net_res_blocks_2_mlp_0_bias_to_fp16, weight = flow_net_res_blocks_2_mlp_0_weight_to_fp16, x = input_237_cast_fp16)[name = tensor("linear_91_cast_fp16")]; + tensor input_241_cast_fp16 = silu(x = linear_91_cast_fp16)[name = tensor("input_241_cast_fp16")]; + tensor linear_92_cast_fp16 = linear(bias = flow_net_res_blocks_2_mlp_2_bias_to_fp16, weight = flow_net_res_blocks_2_mlp_2_weight_to_fp16, x = input_241_cast_fp16)[name = tensor("linear_92_cast_fp16")]; + tensor var_1428_cast_fp16 = mul(x = var_1399_cast_fp16_2, y = linear_92_cast_fp16)[name = tensor("op_1428_cast_fp16")]; + tensor x_155_cast_fp16 = add(x = x_149_cast_fp16, y = var_1428_cast_fp16)[name = tensor("x_155_cast_fp16")]; + tensor linear_93_cast_fp16 = linear(bias = flow_net_res_blocks_3_adaLN_modulation_1_bias_to_fp16, weight = flow_net_res_blocks_3_adaLN_modulation_1_weight_to_fp16, x = input_219_cast_fp16)[name = tensor("linear_93_cast_fp16")]; + tensor var_1438_split_sizes_0 = const()[name = tensor("op_1438_split_sizes_0"), val = tensor([512, 512, 512])]; + tensor var_1438_axis_0 = const()[name = tensor("op_1438_axis_0"), val = tensor(-1)]; + tensor var_1438_cast_fp16_0, tensor var_1438_cast_fp16_1, tensor var_1438_cast_fp16_2 = split(axis = var_1438_axis_0, split_sizes = var_1438_split_sizes_0, x = linear_93_cast_fp16)[name = tensor("op_1438_cast_fp16")]; + tensor mean_49_axes_0 = const()[name = tensor("mean_49_axes_0"), val = tensor([-1])]; + tensor mean_49_keep_dims_0 = const()[name = tensor("mean_49_keep_dims_0"), val = tensor(true)]; + tensor mean_49_cast_fp16 = reduce_mean(axes = mean_49_axes_0, keep_dims = mean_49_keep_dims_0, x = x_155_cast_fp16)[name = tensor("mean_49_cast_fp16")]; + tensor sub_40_cast_fp16 = sub(x = x_155_cast_fp16, y = mean_49_cast_fp16)[name = tensor("sub_40_cast_fp16")]; + tensor square_32_cast_fp16 = square(x = sub_40_cast_fp16)[name = tensor("square_32_cast_fp16")]; + tensor reduce_mean_65_axes_0 = const()[name = tensor("reduce_mean_65_axes_0"), val = tensor([-1])]; + tensor reduce_mean_65_keep_dims_0 = const()[name = tensor("reduce_mean_65_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_65_cast_fp16 = reduce_mean(axes = reduce_mean_65_axes_0, keep_dims = reduce_mean_65_keep_dims_0, x = square_32_cast_fp16)[name = tensor("reduce_mean_65_cast_fp16")]; + tensor var_1448_to_fp16 = const()[name = tensor("op_1448_to_fp16"), val = tensor(0x1.1p-20)]; + tensor var_1449_cast_fp16 = add(x = reduce_mean_65_cast_fp16, y = var_1448_to_fp16)[name = tensor("op_1449_cast_fp16")]; + tensor var_1450_cast_fp16 = sqrt(x = var_1449_cast_fp16)[name = tensor("op_1450_cast_fp16")]; + tensor x_157_cast_fp16 = real_div(x = sub_40_cast_fp16, y = var_1450_cast_fp16)[name = tensor("x_157_cast_fp16")]; + tensor var_1452_cast_fp16 = mul(x = x_157_cast_fp16, y = flow_net_res_blocks_3_in_ln_weight_to_fp16)[name = tensor("op_1452_cast_fp16")]; + tensor x_159_cast_fp16 = add(x = var_1452_cast_fp16, y = flow_net_res_blocks_3_in_ln_bias_to_fp16)[name = tensor("x_159_cast_fp16")]; + tensor var_1454_promoted_to_fp16 = const()[name = tensor("op_1454_promoted_to_fp16"), val = tensor(0x1p+0)]; + tensor var_1455_cast_fp16 = add(x = var_1438_cast_fp16_1, y = var_1454_promoted_to_fp16)[name = tensor("op_1455_cast_fp16")]; + tensor var_1456_cast_fp16 = mul(x = x_159_cast_fp16, y = var_1455_cast_fp16)[name = tensor("op_1456_cast_fp16")]; + tensor input_245_cast_fp16 = add(x = var_1456_cast_fp16, y = var_1438_cast_fp16_0)[name = tensor("input_245_cast_fp16")]; + tensor linear_94_cast_fp16 = linear(bias = flow_net_res_blocks_3_mlp_0_bias_to_fp16, weight = flow_net_res_blocks_3_mlp_0_weight_to_fp16, x = input_245_cast_fp16)[name = tensor("linear_94_cast_fp16")]; + tensor input_249_cast_fp16 = silu(x = linear_94_cast_fp16)[name = tensor("input_249_cast_fp16")]; + tensor linear_95_cast_fp16 = linear(bias = flow_net_res_blocks_3_mlp_2_bias_to_fp16, weight = flow_net_res_blocks_3_mlp_2_weight_to_fp16, x = input_249_cast_fp16)[name = tensor("linear_95_cast_fp16")]; + tensor var_1467_cast_fp16 = mul(x = var_1438_cast_fp16_2, y = linear_95_cast_fp16)[name = tensor("op_1467_cast_fp16")]; + tensor x_161_cast_fp16 = add(x = x_155_cast_fp16, y = var_1467_cast_fp16)[name = tensor("x_161_cast_fp16")]; + tensor linear_96_cast_fp16 = linear(bias = flow_net_res_blocks_4_adaLN_modulation_1_bias_to_fp16, weight = flow_net_res_blocks_4_adaLN_modulation_1_weight_to_fp16, x = input_219_cast_fp16)[name = tensor("linear_96_cast_fp16")]; + tensor var_1477_split_sizes_0 = const()[name = tensor("op_1477_split_sizes_0"), val = tensor([512, 512, 512])]; + tensor var_1477_axis_0 = const()[name = tensor("op_1477_axis_0"), val = tensor(-1)]; + tensor var_1477_cast_fp16_0, tensor var_1477_cast_fp16_1, tensor var_1477_cast_fp16_2 = split(axis = var_1477_axis_0, split_sizes = var_1477_split_sizes_0, x = linear_96_cast_fp16)[name = tensor("op_1477_cast_fp16")]; + tensor mean_51_axes_0 = const()[name = tensor("mean_51_axes_0"), val = tensor([-1])]; + tensor mean_51_keep_dims_0 = const()[name = tensor("mean_51_keep_dims_0"), val = tensor(true)]; + tensor mean_51_cast_fp16 = reduce_mean(axes = mean_51_axes_0, keep_dims = mean_51_keep_dims_0, x = x_161_cast_fp16)[name = tensor("mean_51_cast_fp16")]; + tensor sub_41_cast_fp16 = sub(x = x_161_cast_fp16, y = mean_51_cast_fp16)[name = tensor("sub_41_cast_fp16")]; + tensor square_33_cast_fp16 = square(x = sub_41_cast_fp16)[name = tensor("square_33_cast_fp16")]; + tensor reduce_mean_67_axes_0 = const()[name = tensor("reduce_mean_67_axes_0"), val = tensor([-1])]; + tensor reduce_mean_67_keep_dims_0 = const()[name = tensor("reduce_mean_67_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_67_cast_fp16 = reduce_mean(axes = reduce_mean_67_axes_0, keep_dims = reduce_mean_67_keep_dims_0, x = square_33_cast_fp16)[name = tensor("reduce_mean_67_cast_fp16")]; + tensor var_1487_to_fp16 = const()[name = tensor("op_1487_to_fp16"), val = tensor(0x1.1p-20)]; + tensor var_1488_cast_fp16 = add(x = reduce_mean_67_cast_fp16, y = var_1487_to_fp16)[name = tensor("op_1488_cast_fp16")]; + tensor var_1489_cast_fp16 = sqrt(x = var_1488_cast_fp16)[name = tensor("op_1489_cast_fp16")]; + tensor x_163_cast_fp16 = real_div(x = sub_41_cast_fp16, y = var_1489_cast_fp16)[name = tensor("x_163_cast_fp16")]; + tensor var_1491_cast_fp16 = mul(x = x_163_cast_fp16, y = flow_net_res_blocks_4_in_ln_weight_to_fp16)[name = tensor("op_1491_cast_fp16")]; + tensor x_165_cast_fp16 = add(x = var_1491_cast_fp16, y = flow_net_res_blocks_4_in_ln_bias_to_fp16)[name = tensor("x_165_cast_fp16")]; + tensor var_1493_promoted_to_fp16 = const()[name = tensor("op_1493_promoted_to_fp16"), val = tensor(0x1p+0)]; + tensor var_1494_cast_fp16 = add(x = var_1477_cast_fp16_1, y = var_1493_promoted_to_fp16)[name = tensor("op_1494_cast_fp16")]; + tensor var_1495_cast_fp16 = mul(x = x_165_cast_fp16, y = var_1494_cast_fp16)[name = tensor("op_1495_cast_fp16")]; + tensor input_253_cast_fp16 = add(x = var_1495_cast_fp16, y = var_1477_cast_fp16_0)[name = tensor("input_253_cast_fp16")]; + tensor linear_97_cast_fp16 = linear(bias = flow_net_res_blocks_4_mlp_0_bias_to_fp16, weight = flow_net_res_blocks_4_mlp_0_weight_to_fp16, x = input_253_cast_fp16)[name = tensor("linear_97_cast_fp16")]; + tensor input_257_cast_fp16 = silu(x = linear_97_cast_fp16)[name = tensor("input_257_cast_fp16")]; + tensor linear_98_cast_fp16 = linear(bias = flow_net_res_blocks_4_mlp_2_bias_to_fp16, weight = flow_net_res_blocks_4_mlp_2_weight_to_fp16, x = input_257_cast_fp16)[name = tensor("linear_98_cast_fp16")]; + tensor var_1506_cast_fp16 = mul(x = var_1477_cast_fp16_2, y = linear_98_cast_fp16)[name = tensor("op_1506_cast_fp16")]; + tensor x_167_cast_fp16 = add(x = x_161_cast_fp16, y = var_1506_cast_fp16)[name = tensor("x_167_cast_fp16")]; + tensor linear_99_cast_fp16 = linear(bias = flow_net_res_blocks_5_adaLN_modulation_1_bias_to_fp16, weight = flow_net_res_blocks_5_adaLN_modulation_1_weight_to_fp16, x = input_219_cast_fp16)[name = tensor("linear_99_cast_fp16")]; + tensor var_1516_split_sizes_0 = const()[name = tensor("op_1516_split_sizes_0"), val = tensor([512, 512, 512])]; + tensor var_1516_axis_0 = const()[name = tensor("op_1516_axis_0"), val = tensor(-1)]; + tensor var_1516_cast_fp16_0, tensor var_1516_cast_fp16_1, tensor var_1516_cast_fp16_2 = split(axis = var_1516_axis_0, split_sizes = var_1516_split_sizes_0, x = linear_99_cast_fp16)[name = tensor("op_1516_cast_fp16")]; + tensor mean_53_axes_0 = const()[name = tensor("mean_53_axes_0"), val = tensor([-1])]; + tensor mean_53_keep_dims_0 = const()[name = tensor("mean_53_keep_dims_0"), val = tensor(true)]; + tensor mean_53_cast_fp16 = reduce_mean(axes = mean_53_axes_0, keep_dims = mean_53_keep_dims_0, x = x_167_cast_fp16)[name = tensor("mean_53_cast_fp16")]; + tensor sub_42_cast_fp16 = sub(x = x_167_cast_fp16, y = mean_53_cast_fp16)[name = tensor("sub_42_cast_fp16")]; + tensor square_34_cast_fp16 = square(x = sub_42_cast_fp16)[name = tensor("square_34_cast_fp16")]; + tensor reduce_mean_69_axes_0 = const()[name = tensor("reduce_mean_69_axes_0"), val = tensor([-1])]; + tensor reduce_mean_69_keep_dims_0 = const()[name = tensor("reduce_mean_69_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_69_cast_fp16 = reduce_mean(axes = reduce_mean_69_axes_0, keep_dims = reduce_mean_69_keep_dims_0, x = square_34_cast_fp16)[name = tensor("reduce_mean_69_cast_fp16")]; + tensor var_1526_to_fp16 = const()[name = tensor("op_1526_to_fp16"), val = tensor(0x1.1p-20)]; + tensor var_1527_cast_fp16 = add(x = reduce_mean_69_cast_fp16, y = var_1526_to_fp16)[name = tensor("op_1527_cast_fp16")]; + tensor var_1528_cast_fp16 = sqrt(x = var_1527_cast_fp16)[name = tensor("op_1528_cast_fp16")]; + tensor x_169_cast_fp16 = real_div(x = sub_42_cast_fp16, y = var_1528_cast_fp16)[name = tensor("x_169_cast_fp16")]; + tensor var_1530_cast_fp16 = mul(x = x_169_cast_fp16, y = flow_net_res_blocks_5_in_ln_weight_to_fp16)[name = tensor("op_1530_cast_fp16")]; + tensor x_171_cast_fp16 = add(x = var_1530_cast_fp16, y = flow_net_res_blocks_5_in_ln_bias_to_fp16)[name = tensor("x_171_cast_fp16")]; + tensor var_1532_promoted_to_fp16 = const()[name = tensor("op_1532_promoted_to_fp16"), val = tensor(0x1p+0)]; + tensor var_1533_cast_fp16 = add(x = var_1516_cast_fp16_1, y = var_1532_promoted_to_fp16)[name = tensor("op_1533_cast_fp16")]; + tensor var_1534_cast_fp16 = mul(x = x_171_cast_fp16, y = var_1533_cast_fp16)[name = tensor("op_1534_cast_fp16")]; + tensor input_261_cast_fp16 = add(x = var_1534_cast_fp16, y = var_1516_cast_fp16_0)[name = tensor("input_261_cast_fp16")]; + tensor linear_100_cast_fp16 = linear(bias = flow_net_res_blocks_5_mlp_0_bias_to_fp16, weight = flow_net_res_blocks_5_mlp_0_weight_to_fp16, x = input_261_cast_fp16)[name = tensor("linear_100_cast_fp16")]; + tensor input_265_cast_fp16 = silu(x = linear_100_cast_fp16)[name = tensor("input_265_cast_fp16")]; + tensor linear_101_cast_fp16 = linear(bias = flow_net_res_blocks_5_mlp_2_bias_to_fp16, weight = flow_net_res_blocks_5_mlp_2_weight_to_fp16, x = input_265_cast_fp16)[name = tensor("linear_101_cast_fp16")]; + tensor var_1545_cast_fp16 = mul(x = var_1516_cast_fp16_2, y = linear_101_cast_fp16)[name = tensor("op_1545_cast_fp16")]; + tensor x_173_cast_fp16 = add(x = x_167_cast_fp16, y = var_1545_cast_fp16)[name = tensor("x_173_cast_fp16")]; + tensor linear_102_cast_fp16 = linear(bias = flow_net_final_layer_adaLN_modulation_1_bias_to_fp16, weight = flow_net_final_layer_adaLN_modulation_1_weight_to_fp16, x = input_219_cast_fp16)[name = tensor("linear_102_cast_fp16")]; + tensor var_1554_split_sizes_0 = const()[name = tensor("op_1554_split_sizes_0"), val = tensor([512, 512])]; + tensor var_1554_axis_0 = const()[name = tensor("op_1554_axis_0"), val = tensor(-1)]; + tensor var_1554_cast_fp16_0, tensor var_1554_cast_fp16_1 = split(axis = var_1554_axis_0, split_sizes = var_1554_split_sizes_0, x = linear_102_cast_fp16)[name = tensor("op_1554_cast_fp16")]; + tensor mean_55_axes_0 = const()[name = tensor("mean_55_axes_0"), val = tensor([-1])]; + tensor mean_55_keep_dims_0 = const()[name = tensor("mean_55_keep_dims_0"), val = tensor(true)]; + tensor mean_55_cast_fp16 = reduce_mean(axes = mean_55_axes_0, keep_dims = mean_55_keep_dims_0, x = x_173_cast_fp16)[name = tensor("mean_55_cast_fp16")]; + tensor sub_43_cast_fp16 = sub(x = x_173_cast_fp16, y = mean_55_cast_fp16)[name = tensor("sub_43_cast_fp16")]; + tensor square_35_cast_fp16 = square(x = sub_43_cast_fp16)[name = tensor("square_35_cast_fp16")]; + tensor reduce_mean_71_axes_0 = const()[name = tensor("reduce_mean_71_axes_0"), val = tensor([-1])]; + tensor reduce_mean_71_keep_dims_0 = const()[name = tensor("reduce_mean_71_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_71_cast_fp16 = reduce_mean(axes = reduce_mean_71_axes_0, keep_dims = reduce_mean_71_keep_dims_0, x = square_35_cast_fp16)[name = tensor("reduce_mean_71_cast_fp16")]; + tensor var_1561_to_fp16 = const()[name = tensor("op_1561_to_fp16"), val = tensor(0x1.1p-20)]; + tensor var_1562_cast_fp16 = add(x = reduce_mean_71_cast_fp16, y = var_1561_to_fp16)[name = tensor("op_1562_cast_fp16")]; + tensor var_1563_cast_fp16 = sqrt(x = var_1562_cast_fp16)[name = tensor("op_1563_cast_fp16")]; + tensor x_175_cast_fp16 = real_div(x = sub_43_cast_fp16, y = var_1563_cast_fp16)[name = tensor("x_175_cast_fp16")]; + tensor var_1565_promoted_to_fp16 = const()[name = tensor("op_1565_promoted_to_fp16"), val = tensor(0x1p+0)]; + tensor var_1566_cast_fp16 = add(x = var_1554_cast_fp16_1, y = var_1565_promoted_to_fp16)[name = tensor("op_1566_cast_fp16")]; + tensor var_1567_cast_fp16 = mul(x = x_175_cast_fp16, y = var_1566_cast_fp16)[name = tensor("op_1567_cast_fp16")]; + tensor input_269_cast_fp16 = add(x = var_1567_cast_fp16, y = var_1554_cast_fp16_0)[name = tensor("input_269_cast_fp16")]; + tensor linear_103_cast_fp16 = linear(bias = flow_net_final_layer_linear_bias_to_fp16, weight = flow_net_final_layer_linear_weight_to_fp16, x = input_269_cast_fp16)[name = tensor("linear_103_cast_fp16")]; + tensor var_1578_to_fp16 = const()[name = tensor("op_1578_to_fp16"), val = tensor(0x1p-3)]; + tensor var_1579_cast_fp16 = mul(x = linear_103_cast_fp16, y = var_1578_to_fp16)[name = tensor("op_1579_cast_fp16")]; + tensor input_271_cast_fp16 = add(x = input_203_cast_fp16, y = var_1579_cast_fp16)[name = tensor("input_271_cast_fp16")]; + tensor linear_104_cast_fp16 = linear(bias = flow_net_input_proj_bias_to_fp16, weight = flow_net_input_proj_weight_to_fp16, x = input_271_cast_fp16)[name = tensor("linear_104_cast_fp16")]; + tensor input_275_to_fp16 = const()[name = tensor("input_275_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(19003904)))]; + tensor input_277_cast_fp16 = silu(x = input_275_to_fp16)[name = tensor("input_277_cast_fp16")]; + tensor linear_106_cast_fp16 = linear(bias = flow_net_time_embed_0_mlp_2_bias_to_fp16, weight = flow_net_time_embed_0_mlp_2_weight_to_fp16, x = input_277_cast_fp16)[name = tensor("linear_106_cast_fp16")]; + tensor reduce_mean_72_axes_0 = const()[name = tensor("reduce_mean_72_axes_0"), val = tensor([-1])]; + tensor reduce_mean_72_keep_dims_0 = const()[name = tensor("reduce_mean_72_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_72_cast_fp16 = reduce_mean(axes = reduce_mean_72_axes_0, keep_dims = reduce_mean_72_keep_dims_0, x = linear_106_cast_fp16)[name = tensor("reduce_mean_72_cast_fp16")]; + tensor sub_44_cast_fp16 = sub(x = linear_106_cast_fp16, y = reduce_mean_72_cast_fp16)[name = tensor("sub_44_cast_fp16")]; + tensor square_36_cast_fp16 = square(x = sub_44_cast_fp16)[name = tensor("square_36_cast_fp16")]; + tensor reduce_mean_73_axes_0 = const()[name = tensor("reduce_mean_73_axes_0"), val = tensor([-1])]; + tensor reduce_mean_73_keep_dims_0 = const()[name = tensor("reduce_mean_73_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_73_cast_fp16 = reduce_mean(axes = reduce_mean_73_axes_0, keep_dims = reduce_mean_73_keep_dims_0, x = square_36_cast_fp16)[name = tensor("reduce_mean_73_cast_fp16")]; + tensor real_div_8_to_fp16 = const()[name = tensor("real_div_8_to_fp16"), val = tensor(0x1.008p+0)]; + tensor mul_8_cast_fp16 = mul(x = reduce_mean_73_cast_fp16, y = real_div_8_to_fp16)[name = tensor("mul_8_cast_fp16")]; + tensor var_1645_to_fp16 = const()[name = tensor("op_1645_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_73_cast_fp16_0 = add(x = mul_8_cast_fp16, y = var_1645_to_fp16)[name = tensor("var_73_cast_fp16")]; + tensor var_1648_epsilon_0 = const()[name = tensor("op_1648_epsilon_0"), val = tensor(0x1.197998p-40)]; + tensor var_1648_cast_fp16 = rsqrt(epsilon = var_1648_epsilon_0, x = var_73_cast_fp16_0)[name = tensor("op_1648_cast_fp16")]; + tensor var_1649_cast_fp16 = mul(x = const_3_to_fp16, y = var_1648_cast_fp16)[name = tensor("op_1649_cast_fp16")]; + tensor var_1650_cast_fp16 = mul(x = linear_106_cast_fp16, y = var_1649_cast_fp16)[name = tensor("op_1650_cast_fp16")]; + tensor input_281_to_fp16 = const()[name = tensor("input_281_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(19004992)))]; + tensor input_283_cast_fp16 = silu(x = input_281_to_fp16)[name = tensor("input_283_cast_fp16")]; + tensor linear_108_cast_fp16 = linear(bias = flow_net_time_embed_1_mlp_2_bias_to_fp16, weight = flow_net_time_embed_1_mlp_2_weight_to_fp16, x = input_283_cast_fp16)[name = tensor("linear_108_cast_fp16")]; + tensor reduce_mean_74_axes_0 = const()[name = tensor("reduce_mean_74_axes_0"), val = tensor([-1])]; + tensor reduce_mean_74_keep_dims_0 = const()[name = tensor("reduce_mean_74_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_74_cast_fp16 = reduce_mean(axes = reduce_mean_74_axes_0, keep_dims = reduce_mean_74_keep_dims_0, x = linear_108_cast_fp16)[name = tensor("reduce_mean_74_cast_fp16")]; + tensor sub_46_cast_fp16 = sub(x = linear_108_cast_fp16, y = reduce_mean_74_cast_fp16)[name = tensor("sub_46_cast_fp16")]; + tensor square_37_cast_fp16 = square(x = sub_46_cast_fp16)[name = tensor("square_37_cast_fp16")]; + tensor reduce_mean_75_axes_0 = const()[name = tensor("reduce_mean_75_axes_0"), val = tensor([-1])]; + tensor reduce_mean_75_keep_dims_0 = const()[name = tensor("reduce_mean_75_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_75_cast_fp16 = reduce_mean(axes = reduce_mean_75_axes_0, keep_dims = reduce_mean_75_keep_dims_0, x = square_37_cast_fp16)[name = tensor("reduce_mean_75_cast_fp16")]; + tensor real_div_9_to_fp16 = const()[name = tensor("real_div_9_to_fp16"), val = tensor(0x1.008p+0)]; + tensor mul_9_cast_fp16 = mul(x = reduce_mean_75_cast_fp16, y = real_div_9_to_fp16)[name = tensor("mul_9_cast_fp16")]; + tensor var_1682_to_fp16 = const()[name = tensor("op_1682_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_75_cast_fp16_0 = add(x = mul_9_cast_fp16, y = var_1682_to_fp16)[name = tensor("var_75_cast_fp16")]; + tensor var_1685_epsilon_0 = const()[name = tensor("op_1685_epsilon_0"), val = tensor(0x1.197998p-40)]; + tensor var_1685_cast_fp16 = rsqrt(epsilon = var_1685_epsilon_0, x = var_75_cast_fp16_0)[name = tensor("op_1685_cast_fp16")]; + tensor var_1686_cast_fp16 = mul(x = const_5_to_fp16, y = var_1685_cast_fp16)[name = tensor("op_1686_cast_fp16")]; + tensor var_1687_cast_fp16 = mul(x = linear_108_cast_fp16, y = var_1686_cast_fp16)[name = tensor("op_1687_cast_fp16")]; + tensor var_1699_cast_fp16 = add(x = var_1650_cast_fp16, y = var_1687_cast_fp16)[name = tensor("op_1699_cast_fp16")]; + tensor _inversed_t_combined_9_y_0_to_fp16 = const()[name = tensor("_inversed_t_combined_9_y_0_to_fp16"), val = tensor(0x1p-1)]; + tensor _inversed_t_combined_9_cast_fp16 = mul(x = var_1699_cast_fp16, y = _inversed_t_combined_9_y_0_to_fp16)[name = tensor("_inversed_t_combined_9_cast_fp16")]; + tensor input_285_cast_fp16 = add(x = _inversed_t_combined_9_cast_fp16, y = linear_5_cast_fp16)[name = tensor("input_285_cast_fp16")]; + tensor input_287_cast_fp16 = silu(x = input_285_cast_fp16)[name = tensor("input_287_cast_fp16")]; + tensor linear_110_cast_fp16 = linear(bias = flow_net_res_blocks_0_adaLN_modulation_1_bias_to_fp16, weight = flow_net_res_blocks_0_adaLN_modulation_1_weight_to_fp16, x = input_287_cast_fp16)[name = tensor("linear_110_cast_fp16")]; + tensor var_1714_split_sizes_0 = const()[name = tensor("op_1714_split_sizes_0"), val = tensor([512, 512, 512])]; + tensor var_1714_axis_0 = const()[name = tensor("op_1714_axis_0"), val = tensor(-1)]; + tensor var_1714_cast_fp16_0, tensor var_1714_cast_fp16_1, tensor var_1714_cast_fp16_2 = split(axis = var_1714_axis_0, split_sizes = var_1714_split_sizes_0, x = linear_110_cast_fp16)[name = tensor("op_1714_cast_fp16")]; + tensor mean_57_axes_0 = const()[name = tensor("mean_57_axes_0"), val = tensor([-1])]; + tensor mean_57_keep_dims_0 = const()[name = tensor("mean_57_keep_dims_0"), val = tensor(true)]; + tensor mean_57_cast_fp16 = reduce_mean(axes = mean_57_axes_0, keep_dims = mean_57_keep_dims_0, x = linear_104_cast_fp16)[name = tensor("mean_57_cast_fp16")]; + tensor sub_48_cast_fp16 = sub(x = linear_104_cast_fp16, y = mean_57_cast_fp16)[name = tensor("sub_48_cast_fp16")]; + tensor square_38_cast_fp16 = square(x = sub_48_cast_fp16)[name = tensor("square_38_cast_fp16")]; + tensor reduce_mean_77_axes_0 = const()[name = tensor("reduce_mean_77_axes_0"), val = tensor([-1])]; + tensor reduce_mean_77_keep_dims_0 = const()[name = tensor("reduce_mean_77_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_77_cast_fp16 = reduce_mean(axes = reduce_mean_77_axes_0, keep_dims = reduce_mean_77_keep_dims_0, x = square_38_cast_fp16)[name = tensor("reduce_mean_77_cast_fp16")]; + tensor var_1724_to_fp16 = const()[name = tensor("op_1724_to_fp16"), val = tensor(0x1.1p-20)]; + tensor var_1725_cast_fp16 = add(x = reduce_mean_77_cast_fp16, y = var_1724_to_fp16)[name = tensor("op_1725_cast_fp16")]; + tensor var_1726_cast_fp16 = sqrt(x = var_1725_cast_fp16)[name = tensor("op_1726_cast_fp16")]; + tensor x_183_cast_fp16 = real_div(x = sub_48_cast_fp16, y = var_1726_cast_fp16)[name = tensor("x_183_cast_fp16")]; + tensor var_1728_cast_fp16 = mul(x = x_183_cast_fp16, y = flow_net_res_blocks_0_in_ln_weight_to_fp16)[name = tensor("op_1728_cast_fp16")]; + tensor x_185_cast_fp16 = add(x = var_1728_cast_fp16, y = flow_net_res_blocks_0_in_ln_bias_to_fp16)[name = tensor("x_185_cast_fp16")]; + tensor var_1730_promoted_to_fp16 = const()[name = tensor("op_1730_promoted_to_fp16"), val = tensor(0x1p+0)]; + tensor var_1731_cast_fp16 = add(x = var_1714_cast_fp16_1, y = var_1730_promoted_to_fp16)[name = tensor("op_1731_cast_fp16")]; + tensor var_1732_cast_fp16 = mul(x = x_185_cast_fp16, y = var_1731_cast_fp16)[name = tensor("op_1732_cast_fp16")]; + tensor input_289_cast_fp16 = add(x = var_1732_cast_fp16, y = var_1714_cast_fp16_0)[name = tensor("input_289_cast_fp16")]; + tensor linear_111_cast_fp16 = linear(bias = flow_net_res_blocks_0_mlp_0_bias_to_fp16, weight = flow_net_res_blocks_0_mlp_0_weight_to_fp16, x = input_289_cast_fp16)[name = tensor("linear_111_cast_fp16")]; + tensor input_293_cast_fp16 = silu(x = linear_111_cast_fp16)[name = tensor("input_293_cast_fp16")]; + tensor linear_112_cast_fp16 = linear(bias = flow_net_res_blocks_0_mlp_2_bias_to_fp16, weight = flow_net_res_blocks_0_mlp_2_weight_to_fp16, x = input_293_cast_fp16)[name = tensor("linear_112_cast_fp16")]; + tensor var_1743_cast_fp16 = mul(x = var_1714_cast_fp16_2, y = linear_112_cast_fp16)[name = tensor("op_1743_cast_fp16")]; + tensor x_187_cast_fp16 = add(x = linear_104_cast_fp16, y = var_1743_cast_fp16)[name = tensor("x_187_cast_fp16")]; + tensor linear_113_cast_fp16 = linear(bias = flow_net_res_blocks_1_adaLN_modulation_1_bias_to_fp16, weight = flow_net_res_blocks_1_adaLN_modulation_1_weight_to_fp16, x = input_287_cast_fp16)[name = tensor("linear_113_cast_fp16")]; + tensor var_1753_split_sizes_0 = const()[name = tensor("op_1753_split_sizes_0"), val = tensor([512, 512, 512])]; + tensor var_1753_axis_0 = const()[name = tensor("op_1753_axis_0"), val = tensor(-1)]; + tensor var_1753_cast_fp16_0, tensor var_1753_cast_fp16_1, tensor var_1753_cast_fp16_2 = split(axis = var_1753_axis_0, split_sizes = var_1753_split_sizes_0, x = linear_113_cast_fp16)[name = tensor("op_1753_cast_fp16")]; + tensor mean_59_axes_0 = const()[name = tensor("mean_59_axes_0"), val = tensor([-1])]; + tensor mean_59_keep_dims_0 = const()[name = tensor("mean_59_keep_dims_0"), val = tensor(true)]; + tensor mean_59_cast_fp16 = reduce_mean(axes = mean_59_axes_0, keep_dims = mean_59_keep_dims_0, x = x_187_cast_fp16)[name = tensor("mean_59_cast_fp16")]; + tensor sub_49_cast_fp16 = sub(x = x_187_cast_fp16, y = mean_59_cast_fp16)[name = tensor("sub_49_cast_fp16")]; + tensor square_39_cast_fp16 = square(x = sub_49_cast_fp16)[name = tensor("square_39_cast_fp16")]; + tensor reduce_mean_79_axes_0 = const()[name = tensor("reduce_mean_79_axes_0"), val = tensor([-1])]; + tensor reduce_mean_79_keep_dims_0 = const()[name = tensor("reduce_mean_79_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_79_cast_fp16 = reduce_mean(axes = reduce_mean_79_axes_0, keep_dims = reduce_mean_79_keep_dims_0, x = square_39_cast_fp16)[name = tensor("reduce_mean_79_cast_fp16")]; + tensor var_1763_to_fp16 = const()[name = tensor("op_1763_to_fp16"), val = tensor(0x1.1p-20)]; + tensor var_1764_cast_fp16 = add(x = reduce_mean_79_cast_fp16, y = var_1763_to_fp16)[name = tensor("op_1764_cast_fp16")]; + tensor var_1765_cast_fp16 = sqrt(x = var_1764_cast_fp16)[name = tensor("op_1765_cast_fp16")]; + tensor x_189_cast_fp16 = real_div(x = sub_49_cast_fp16, y = var_1765_cast_fp16)[name = tensor("x_189_cast_fp16")]; + tensor var_1767_cast_fp16 = mul(x = x_189_cast_fp16, y = flow_net_res_blocks_1_in_ln_weight_to_fp16)[name = tensor("op_1767_cast_fp16")]; + tensor x_191_cast_fp16 = add(x = var_1767_cast_fp16, y = flow_net_res_blocks_1_in_ln_bias_to_fp16)[name = tensor("x_191_cast_fp16")]; + tensor var_1769_promoted_to_fp16 = const()[name = tensor("op_1769_promoted_to_fp16"), val = tensor(0x1p+0)]; + tensor var_1770_cast_fp16 = add(x = var_1753_cast_fp16_1, y = var_1769_promoted_to_fp16)[name = tensor("op_1770_cast_fp16")]; + tensor var_1771_cast_fp16 = mul(x = x_191_cast_fp16, y = var_1770_cast_fp16)[name = tensor("op_1771_cast_fp16")]; + tensor input_297_cast_fp16 = add(x = var_1771_cast_fp16, y = var_1753_cast_fp16_0)[name = tensor("input_297_cast_fp16")]; + tensor linear_114_cast_fp16 = linear(bias = flow_net_res_blocks_1_mlp_0_bias_to_fp16, weight = flow_net_res_blocks_1_mlp_0_weight_to_fp16, x = input_297_cast_fp16)[name = tensor("linear_114_cast_fp16")]; + tensor input_301_cast_fp16 = silu(x = linear_114_cast_fp16)[name = tensor("input_301_cast_fp16")]; + tensor linear_115_cast_fp16 = linear(bias = flow_net_res_blocks_1_mlp_2_bias_to_fp16, weight = flow_net_res_blocks_1_mlp_2_weight_to_fp16, x = input_301_cast_fp16)[name = tensor("linear_115_cast_fp16")]; + tensor var_1782_cast_fp16 = mul(x = var_1753_cast_fp16_2, y = linear_115_cast_fp16)[name = tensor("op_1782_cast_fp16")]; + tensor x_193_cast_fp16 = add(x = x_187_cast_fp16, y = var_1782_cast_fp16)[name = tensor("x_193_cast_fp16")]; + tensor linear_116_cast_fp16 = linear(bias = flow_net_res_blocks_2_adaLN_modulation_1_bias_to_fp16, weight = flow_net_res_blocks_2_adaLN_modulation_1_weight_to_fp16, x = input_287_cast_fp16)[name = tensor("linear_116_cast_fp16")]; + tensor var_1792_split_sizes_0 = const()[name = tensor("op_1792_split_sizes_0"), val = tensor([512, 512, 512])]; + tensor var_1792_axis_0 = const()[name = tensor("op_1792_axis_0"), val = tensor(-1)]; + tensor var_1792_cast_fp16_0, tensor var_1792_cast_fp16_1, tensor var_1792_cast_fp16_2 = split(axis = var_1792_axis_0, split_sizes = var_1792_split_sizes_0, x = linear_116_cast_fp16)[name = tensor("op_1792_cast_fp16")]; + tensor mean_61_axes_0 = const()[name = tensor("mean_61_axes_0"), val = tensor([-1])]; + tensor mean_61_keep_dims_0 = const()[name = tensor("mean_61_keep_dims_0"), val = tensor(true)]; + tensor mean_61_cast_fp16 = reduce_mean(axes = mean_61_axes_0, keep_dims = mean_61_keep_dims_0, x = x_193_cast_fp16)[name = tensor("mean_61_cast_fp16")]; + tensor sub_50_cast_fp16 = sub(x = x_193_cast_fp16, y = mean_61_cast_fp16)[name = tensor("sub_50_cast_fp16")]; + tensor square_40_cast_fp16 = square(x = sub_50_cast_fp16)[name = tensor("square_40_cast_fp16")]; + tensor reduce_mean_81_axes_0 = const()[name = tensor("reduce_mean_81_axes_0"), val = tensor([-1])]; + tensor reduce_mean_81_keep_dims_0 = const()[name = tensor("reduce_mean_81_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_81_cast_fp16 = reduce_mean(axes = reduce_mean_81_axes_0, keep_dims = reduce_mean_81_keep_dims_0, x = square_40_cast_fp16)[name = tensor("reduce_mean_81_cast_fp16")]; + tensor var_1802_to_fp16 = const()[name = tensor("op_1802_to_fp16"), val = tensor(0x1.1p-20)]; + tensor var_1803_cast_fp16 = add(x = reduce_mean_81_cast_fp16, y = var_1802_to_fp16)[name = tensor("op_1803_cast_fp16")]; + tensor var_1804_cast_fp16 = sqrt(x = var_1803_cast_fp16)[name = tensor("op_1804_cast_fp16")]; + tensor x_195_cast_fp16 = real_div(x = sub_50_cast_fp16, y = var_1804_cast_fp16)[name = tensor("x_195_cast_fp16")]; + tensor var_1806_cast_fp16 = mul(x = x_195_cast_fp16, y = flow_net_res_blocks_2_in_ln_weight_to_fp16)[name = tensor("op_1806_cast_fp16")]; + tensor x_197_cast_fp16 = add(x = var_1806_cast_fp16, y = flow_net_res_blocks_2_in_ln_bias_to_fp16)[name = tensor("x_197_cast_fp16")]; + tensor var_1808_promoted_to_fp16 = const()[name = tensor("op_1808_promoted_to_fp16"), val = tensor(0x1p+0)]; + tensor var_1809_cast_fp16 = add(x = var_1792_cast_fp16_1, y = var_1808_promoted_to_fp16)[name = tensor("op_1809_cast_fp16")]; + tensor var_1810_cast_fp16 = mul(x = x_197_cast_fp16, y = var_1809_cast_fp16)[name = tensor("op_1810_cast_fp16")]; + tensor input_305_cast_fp16 = add(x = var_1810_cast_fp16, y = var_1792_cast_fp16_0)[name = tensor("input_305_cast_fp16")]; + tensor linear_117_cast_fp16 = linear(bias = flow_net_res_blocks_2_mlp_0_bias_to_fp16, weight = flow_net_res_blocks_2_mlp_0_weight_to_fp16, x = input_305_cast_fp16)[name = tensor("linear_117_cast_fp16")]; + tensor input_309_cast_fp16 = silu(x = linear_117_cast_fp16)[name = tensor("input_309_cast_fp16")]; + tensor linear_118_cast_fp16 = linear(bias = flow_net_res_blocks_2_mlp_2_bias_to_fp16, weight = flow_net_res_blocks_2_mlp_2_weight_to_fp16, x = input_309_cast_fp16)[name = tensor("linear_118_cast_fp16")]; + tensor var_1821_cast_fp16 = mul(x = var_1792_cast_fp16_2, y = linear_118_cast_fp16)[name = tensor("op_1821_cast_fp16")]; + tensor x_199_cast_fp16 = add(x = x_193_cast_fp16, y = var_1821_cast_fp16)[name = tensor("x_199_cast_fp16")]; + tensor linear_119_cast_fp16 = linear(bias = flow_net_res_blocks_3_adaLN_modulation_1_bias_to_fp16, weight = flow_net_res_blocks_3_adaLN_modulation_1_weight_to_fp16, x = input_287_cast_fp16)[name = tensor("linear_119_cast_fp16")]; + tensor var_1831_split_sizes_0 = const()[name = tensor("op_1831_split_sizes_0"), val = tensor([512, 512, 512])]; + tensor var_1831_axis_0 = const()[name = tensor("op_1831_axis_0"), val = tensor(-1)]; + tensor var_1831_cast_fp16_0, tensor var_1831_cast_fp16_1, tensor var_1831_cast_fp16_2 = split(axis = var_1831_axis_0, split_sizes = var_1831_split_sizes_0, x = linear_119_cast_fp16)[name = tensor("op_1831_cast_fp16")]; + tensor mean_63_axes_0 = const()[name = tensor("mean_63_axes_0"), val = tensor([-1])]; + tensor mean_63_keep_dims_0 = const()[name = tensor("mean_63_keep_dims_0"), val = tensor(true)]; + tensor mean_63_cast_fp16 = reduce_mean(axes = mean_63_axes_0, keep_dims = mean_63_keep_dims_0, x = x_199_cast_fp16)[name = tensor("mean_63_cast_fp16")]; + tensor sub_51_cast_fp16 = sub(x = x_199_cast_fp16, y = mean_63_cast_fp16)[name = tensor("sub_51_cast_fp16")]; + tensor square_41_cast_fp16 = square(x = sub_51_cast_fp16)[name = tensor("square_41_cast_fp16")]; + tensor reduce_mean_83_axes_0 = const()[name = tensor("reduce_mean_83_axes_0"), val = tensor([-1])]; + tensor reduce_mean_83_keep_dims_0 = const()[name = tensor("reduce_mean_83_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_83_cast_fp16 = reduce_mean(axes = reduce_mean_83_axes_0, keep_dims = reduce_mean_83_keep_dims_0, x = square_41_cast_fp16)[name = tensor("reduce_mean_83_cast_fp16")]; + tensor var_1841_to_fp16 = const()[name = tensor("op_1841_to_fp16"), val = tensor(0x1.1p-20)]; + tensor var_1842_cast_fp16 = add(x = reduce_mean_83_cast_fp16, y = var_1841_to_fp16)[name = tensor("op_1842_cast_fp16")]; + tensor var_1843_cast_fp16 = sqrt(x = var_1842_cast_fp16)[name = tensor("op_1843_cast_fp16")]; + tensor x_201_cast_fp16 = real_div(x = sub_51_cast_fp16, y = var_1843_cast_fp16)[name = tensor("x_201_cast_fp16")]; + tensor var_1845_cast_fp16 = mul(x = x_201_cast_fp16, y = flow_net_res_blocks_3_in_ln_weight_to_fp16)[name = tensor("op_1845_cast_fp16")]; + tensor x_203_cast_fp16 = add(x = var_1845_cast_fp16, y = flow_net_res_blocks_3_in_ln_bias_to_fp16)[name = tensor("x_203_cast_fp16")]; + tensor var_1847_promoted_to_fp16 = const()[name = tensor("op_1847_promoted_to_fp16"), val = tensor(0x1p+0)]; + tensor var_1848_cast_fp16 = add(x = var_1831_cast_fp16_1, y = var_1847_promoted_to_fp16)[name = tensor("op_1848_cast_fp16")]; + tensor var_1849_cast_fp16 = mul(x = x_203_cast_fp16, y = var_1848_cast_fp16)[name = tensor("op_1849_cast_fp16")]; + tensor input_313_cast_fp16 = add(x = var_1849_cast_fp16, y = var_1831_cast_fp16_0)[name = tensor("input_313_cast_fp16")]; + tensor linear_120_cast_fp16 = linear(bias = flow_net_res_blocks_3_mlp_0_bias_to_fp16, weight = flow_net_res_blocks_3_mlp_0_weight_to_fp16, x = input_313_cast_fp16)[name = tensor("linear_120_cast_fp16")]; + tensor input_317_cast_fp16 = silu(x = linear_120_cast_fp16)[name = tensor("input_317_cast_fp16")]; + tensor linear_121_cast_fp16 = linear(bias = flow_net_res_blocks_3_mlp_2_bias_to_fp16, weight = flow_net_res_blocks_3_mlp_2_weight_to_fp16, x = input_317_cast_fp16)[name = tensor("linear_121_cast_fp16")]; + tensor var_1860_cast_fp16 = mul(x = var_1831_cast_fp16_2, y = linear_121_cast_fp16)[name = tensor("op_1860_cast_fp16")]; + tensor x_205_cast_fp16 = add(x = x_199_cast_fp16, y = var_1860_cast_fp16)[name = tensor("x_205_cast_fp16")]; + tensor linear_122_cast_fp16 = linear(bias = flow_net_res_blocks_4_adaLN_modulation_1_bias_to_fp16, weight = flow_net_res_blocks_4_adaLN_modulation_1_weight_to_fp16, x = input_287_cast_fp16)[name = tensor("linear_122_cast_fp16")]; + tensor var_1870_split_sizes_0 = const()[name = tensor("op_1870_split_sizes_0"), val = tensor([512, 512, 512])]; + tensor var_1870_axis_0 = const()[name = tensor("op_1870_axis_0"), val = tensor(-1)]; + tensor var_1870_cast_fp16_0, tensor var_1870_cast_fp16_1, tensor var_1870_cast_fp16_2 = split(axis = var_1870_axis_0, split_sizes = var_1870_split_sizes_0, x = linear_122_cast_fp16)[name = tensor("op_1870_cast_fp16")]; + tensor mean_65_axes_0 = const()[name = tensor("mean_65_axes_0"), val = tensor([-1])]; + tensor mean_65_keep_dims_0 = const()[name = tensor("mean_65_keep_dims_0"), val = tensor(true)]; + tensor mean_65_cast_fp16 = reduce_mean(axes = mean_65_axes_0, keep_dims = mean_65_keep_dims_0, x = x_205_cast_fp16)[name = tensor("mean_65_cast_fp16")]; + tensor sub_52_cast_fp16 = sub(x = x_205_cast_fp16, y = mean_65_cast_fp16)[name = tensor("sub_52_cast_fp16")]; + tensor square_42_cast_fp16 = square(x = sub_52_cast_fp16)[name = tensor("square_42_cast_fp16")]; + tensor reduce_mean_85_axes_0 = const()[name = tensor("reduce_mean_85_axes_0"), val = tensor([-1])]; + tensor reduce_mean_85_keep_dims_0 = const()[name = tensor("reduce_mean_85_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_85_cast_fp16 = reduce_mean(axes = reduce_mean_85_axes_0, keep_dims = reduce_mean_85_keep_dims_0, x = square_42_cast_fp16)[name = tensor("reduce_mean_85_cast_fp16")]; + tensor var_1880_to_fp16 = const()[name = tensor("op_1880_to_fp16"), val = tensor(0x1.1p-20)]; + tensor var_1881_cast_fp16 = add(x = reduce_mean_85_cast_fp16, y = var_1880_to_fp16)[name = tensor("op_1881_cast_fp16")]; + tensor var_1882_cast_fp16 = sqrt(x = var_1881_cast_fp16)[name = tensor("op_1882_cast_fp16")]; + tensor x_207_cast_fp16 = real_div(x = sub_52_cast_fp16, y = var_1882_cast_fp16)[name = tensor("x_207_cast_fp16")]; + tensor var_1884_cast_fp16 = mul(x = x_207_cast_fp16, y = flow_net_res_blocks_4_in_ln_weight_to_fp16)[name = tensor("op_1884_cast_fp16")]; + tensor x_209_cast_fp16 = add(x = var_1884_cast_fp16, y = flow_net_res_blocks_4_in_ln_bias_to_fp16)[name = tensor("x_209_cast_fp16")]; + tensor var_1886_promoted_to_fp16 = const()[name = tensor("op_1886_promoted_to_fp16"), val = tensor(0x1p+0)]; + tensor var_1887_cast_fp16 = add(x = var_1870_cast_fp16_1, y = var_1886_promoted_to_fp16)[name = tensor("op_1887_cast_fp16")]; + tensor var_1888_cast_fp16 = mul(x = x_209_cast_fp16, y = var_1887_cast_fp16)[name = tensor("op_1888_cast_fp16")]; + tensor input_321_cast_fp16 = add(x = var_1888_cast_fp16, y = var_1870_cast_fp16_0)[name = tensor("input_321_cast_fp16")]; + tensor linear_123_cast_fp16 = linear(bias = flow_net_res_blocks_4_mlp_0_bias_to_fp16, weight = flow_net_res_blocks_4_mlp_0_weight_to_fp16, x = input_321_cast_fp16)[name = tensor("linear_123_cast_fp16")]; + tensor input_325_cast_fp16 = silu(x = linear_123_cast_fp16)[name = tensor("input_325_cast_fp16")]; + tensor linear_124_cast_fp16 = linear(bias = flow_net_res_blocks_4_mlp_2_bias_to_fp16, weight = flow_net_res_blocks_4_mlp_2_weight_to_fp16, x = input_325_cast_fp16)[name = tensor("linear_124_cast_fp16")]; + tensor var_1899_cast_fp16 = mul(x = var_1870_cast_fp16_2, y = linear_124_cast_fp16)[name = tensor("op_1899_cast_fp16")]; + tensor x_211_cast_fp16 = add(x = x_205_cast_fp16, y = var_1899_cast_fp16)[name = tensor("x_211_cast_fp16")]; + tensor linear_125_cast_fp16 = linear(bias = flow_net_res_blocks_5_adaLN_modulation_1_bias_to_fp16, weight = flow_net_res_blocks_5_adaLN_modulation_1_weight_to_fp16, x = input_287_cast_fp16)[name = tensor("linear_125_cast_fp16")]; + tensor var_1909_split_sizes_0 = const()[name = tensor("op_1909_split_sizes_0"), val = tensor([512, 512, 512])]; + tensor var_1909_axis_0 = const()[name = tensor("op_1909_axis_0"), val = tensor(-1)]; + tensor var_1909_cast_fp16_0, tensor var_1909_cast_fp16_1, tensor var_1909_cast_fp16_2 = split(axis = var_1909_axis_0, split_sizes = var_1909_split_sizes_0, x = linear_125_cast_fp16)[name = tensor("op_1909_cast_fp16")]; + tensor mean_67_axes_0 = const()[name = tensor("mean_67_axes_0"), val = tensor([-1])]; + tensor mean_67_keep_dims_0 = const()[name = tensor("mean_67_keep_dims_0"), val = tensor(true)]; + tensor mean_67_cast_fp16 = reduce_mean(axes = mean_67_axes_0, keep_dims = mean_67_keep_dims_0, x = x_211_cast_fp16)[name = tensor("mean_67_cast_fp16")]; + tensor sub_53_cast_fp16 = sub(x = x_211_cast_fp16, y = mean_67_cast_fp16)[name = tensor("sub_53_cast_fp16")]; + tensor square_43_cast_fp16 = square(x = sub_53_cast_fp16)[name = tensor("square_43_cast_fp16")]; + tensor reduce_mean_87_axes_0 = const()[name = tensor("reduce_mean_87_axes_0"), val = tensor([-1])]; + tensor reduce_mean_87_keep_dims_0 = const()[name = tensor("reduce_mean_87_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_87_cast_fp16 = reduce_mean(axes = reduce_mean_87_axes_0, keep_dims = reduce_mean_87_keep_dims_0, x = square_43_cast_fp16)[name = tensor("reduce_mean_87_cast_fp16")]; + tensor var_1919_to_fp16 = const()[name = tensor("op_1919_to_fp16"), val = tensor(0x1.1p-20)]; + tensor var_1920_cast_fp16 = add(x = reduce_mean_87_cast_fp16, y = var_1919_to_fp16)[name = tensor("op_1920_cast_fp16")]; + tensor var_1921_cast_fp16 = sqrt(x = var_1920_cast_fp16)[name = tensor("op_1921_cast_fp16")]; + tensor x_213_cast_fp16 = real_div(x = sub_53_cast_fp16, y = var_1921_cast_fp16)[name = tensor("x_213_cast_fp16")]; + tensor var_1923_cast_fp16 = mul(x = x_213_cast_fp16, y = flow_net_res_blocks_5_in_ln_weight_to_fp16)[name = tensor("op_1923_cast_fp16")]; + tensor x_215_cast_fp16 = add(x = var_1923_cast_fp16, y = flow_net_res_blocks_5_in_ln_bias_to_fp16)[name = tensor("x_215_cast_fp16")]; + tensor var_1925_promoted_to_fp16 = const()[name = tensor("op_1925_promoted_to_fp16"), val = tensor(0x1p+0)]; + tensor var_1926_cast_fp16 = add(x = var_1909_cast_fp16_1, y = var_1925_promoted_to_fp16)[name = tensor("op_1926_cast_fp16")]; + tensor var_1927_cast_fp16 = mul(x = x_215_cast_fp16, y = var_1926_cast_fp16)[name = tensor("op_1927_cast_fp16")]; + tensor input_329_cast_fp16 = add(x = var_1927_cast_fp16, y = var_1909_cast_fp16_0)[name = tensor("input_329_cast_fp16")]; + tensor linear_126_cast_fp16 = linear(bias = flow_net_res_blocks_5_mlp_0_bias_to_fp16, weight = flow_net_res_blocks_5_mlp_0_weight_to_fp16, x = input_329_cast_fp16)[name = tensor("linear_126_cast_fp16")]; + tensor input_333_cast_fp16 = silu(x = linear_126_cast_fp16)[name = tensor("input_333_cast_fp16")]; + tensor linear_127_cast_fp16 = linear(bias = flow_net_res_blocks_5_mlp_2_bias_to_fp16, weight = flow_net_res_blocks_5_mlp_2_weight_to_fp16, x = input_333_cast_fp16)[name = tensor("linear_127_cast_fp16")]; + tensor var_1938_cast_fp16 = mul(x = var_1909_cast_fp16_2, y = linear_127_cast_fp16)[name = tensor("op_1938_cast_fp16")]; + tensor x_217_cast_fp16 = add(x = x_211_cast_fp16, y = var_1938_cast_fp16)[name = tensor("x_217_cast_fp16")]; + tensor linear_128_cast_fp16 = linear(bias = flow_net_final_layer_adaLN_modulation_1_bias_to_fp16, weight = flow_net_final_layer_adaLN_modulation_1_weight_to_fp16, x = input_287_cast_fp16)[name = tensor("linear_128_cast_fp16")]; + tensor var_1947_split_sizes_0 = const()[name = tensor("op_1947_split_sizes_0"), val = tensor([512, 512])]; + tensor var_1947_axis_0 = const()[name = tensor("op_1947_axis_0"), val = tensor(-1)]; + tensor var_1947_cast_fp16_0, tensor var_1947_cast_fp16_1 = split(axis = var_1947_axis_0, split_sizes = var_1947_split_sizes_0, x = linear_128_cast_fp16)[name = tensor("op_1947_cast_fp16")]; + tensor mean_69_axes_0 = const()[name = tensor("mean_69_axes_0"), val = tensor([-1])]; + tensor mean_69_keep_dims_0 = const()[name = tensor("mean_69_keep_dims_0"), val = tensor(true)]; + tensor mean_69_cast_fp16 = reduce_mean(axes = mean_69_axes_0, keep_dims = mean_69_keep_dims_0, x = x_217_cast_fp16)[name = tensor("mean_69_cast_fp16")]; + tensor sub_54_cast_fp16 = sub(x = x_217_cast_fp16, y = mean_69_cast_fp16)[name = tensor("sub_54_cast_fp16")]; + tensor square_44_cast_fp16 = square(x = sub_54_cast_fp16)[name = tensor("square_44_cast_fp16")]; + tensor reduce_mean_89_axes_0 = const()[name = tensor("reduce_mean_89_axes_0"), val = tensor([-1])]; + tensor reduce_mean_89_keep_dims_0 = const()[name = tensor("reduce_mean_89_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_89_cast_fp16 = reduce_mean(axes = reduce_mean_89_axes_0, keep_dims = reduce_mean_89_keep_dims_0, x = square_44_cast_fp16)[name = tensor("reduce_mean_89_cast_fp16")]; + tensor var_1954_to_fp16 = const()[name = tensor("op_1954_to_fp16"), val = tensor(0x1.1p-20)]; + tensor var_1955_cast_fp16 = add(x = reduce_mean_89_cast_fp16, y = var_1954_to_fp16)[name = tensor("op_1955_cast_fp16")]; + tensor var_1956_cast_fp16 = sqrt(x = var_1955_cast_fp16)[name = tensor("op_1956_cast_fp16")]; + tensor x_219_cast_fp16 = real_div(x = sub_54_cast_fp16, y = var_1956_cast_fp16)[name = tensor("x_219_cast_fp16")]; + tensor var_1958_promoted_to_fp16 = const()[name = tensor("op_1958_promoted_to_fp16"), val = tensor(0x1p+0)]; + tensor var_1959_cast_fp16 = add(x = var_1947_cast_fp16_1, y = var_1958_promoted_to_fp16)[name = tensor("op_1959_cast_fp16")]; + tensor var_1960_cast_fp16 = mul(x = x_219_cast_fp16, y = var_1959_cast_fp16)[name = tensor("op_1960_cast_fp16")]; + tensor input_337_cast_fp16 = add(x = var_1960_cast_fp16, y = var_1947_cast_fp16_0)[name = tensor("input_337_cast_fp16")]; + tensor linear_129_cast_fp16 = linear(bias = flow_net_final_layer_linear_bias_to_fp16, weight = flow_net_final_layer_linear_weight_to_fp16, x = input_337_cast_fp16)[name = tensor("linear_129_cast_fp16")]; + tensor var_1971_to_fp16 = const()[name = tensor("op_1971_to_fp16"), val = tensor(0x1p-3)]; + tensor var_1972_cast_fp16 = mul(x = linear_129_cast_fp16, y = var_1971_to_fp16)[name = tensor("op_1972_cast_fp16")]; + tensor input_339_cast_fp16 = add(x = input_271_cast_fp16, y = var_1972_cast_fp16)[name = tensor("input_339_cast_fp16")]; + tensor linear_130_cast_fp16 = linear(bias = flow_net_input_proj_bias_to_fp16, weight = flow_net_input_proj_weight_to_fp16, x = input_339_cast_fp16)[name = tensor("linear_130_cast_fp16")]; + tensor input_343_to_fp16 = const()[name = tensor("input_343_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(19006080)))]; + tensor input_345_cast_fp16 = silu(x = input_343_to_fp16)[name = tensor("input_345_cast_fp16")]; + tensor linear_132_cast_fp16 = linear(bias = flow_net_time_embed_0_mlp_2_bias_to_fp16, weight = flow_net_time_embed_0_mlp_2_weight_to_fp16, x = input_345_cast_fp16)[name = tensor("linear_132_cast_fp16")]; + tensor reduce_mean_90_axes_0 = const()[name = tensor("reduce_mean_90_axes_0"), val = tensor([-1])]; + tensor reduce_mean_90_keep_dims_0 = const()[name = tensor("reduce_mean_90_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_90_cast_fp16 = reduce_mean(axes = reduce_mean_90_axes_0, keep_dims = reduce_mean_90_keep_dims_0, x = linear_132_cast_fp16)[name = tensor("reduce_mean_90_cast_fp16")]; + tensor sub_55_cast_fp16 = sub(x = linear_132_cast_fp16, y = reduce_mean_90_cast_fp16)[name = tensor("sub_55_cast_fp16")]; + tensor square_45_cast_fp16 = square(x = sub_55_cast_fp16)[name = tensor("square_45_cast_fp16")]; + tensor reduce_mean_91_axes_0 = const()[name = tensor("reduce_mean_91_axes_0"), val = tensor([-1])]; + tensor reduce_mean_91_keep_dims_0 = const()[name = tensor("reduce_mean_91_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_91_cast_fp16 = reduce_mean(axes = reduce_mean_91_axes_0, keep_dims = reduce_mean_91_keep_dims_0, x = square_45_cast_fp16)[name = tensor("reduce_mean_91_cast_fp16")]; + tensor real_div_10_to_fp16 = const()[name = tensor("real_div_10_to_fp16"), val = tensor(0x1.008p+0)]; + tensor mul_10_cast_fp16 = mul(x = reduce_mean_91_cast_fp16, y = real_div_10_to_fp16)[name = tensor("mul_10_cast_fp16")]; + tensor var_2038_to_fp16 = const()[name = tensor("op_2038_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_91_cast_fp16 = add(x = mul_10_cast_fp16, y = var_2038_to_fp16)[name = tensor("var_91_cast_fp16")]; + tensor var_2041_epsilon_0 = const()[name = tensor("op_2041_epsilon_0"), val = tensor(0x1.197998p-40)]; + tensor var_2041_cast_fp16 = rsqrt(epsilon = var_2041_epsilon_0, x = var_91_cast_fp16)[name = tensor("op_2041_cast_fp16")]; + tensor var_2042_cast_fp16 = mul(x = const_3_to_fp16, y = var_2041_cast_fp16)[name = tensor("op_2042_cast_fp16")]; + tensor var_2043_cast_fp16 = mul(x = linear_132_cast_fp16, y = var_2042_cast_fp16)[name = tensor("op_2043_cast_fp16")]; + tensor input_349_to_fp16 = const()[name = tensor("input_349_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(19007168)))]; + tensor input_351_cast_fp16 = silu(x = input_349_to_fp16)[name = tensor("input_351_cast_fp16")]; + tensor linear_134_cast_fp16 = linear(bias = flow_net_time_embed_1_mlp_2_bias_to_fp16, weight = flow_net_time_embed_1_mlp_2_weight_to_fp16, x = input_351_cast_fp16)[name = tensor("linear_134_cast_fp16")]; + tensor reduce_mean_92_axes_0 = const()[name = tensor("reduce_mean_92_axes_0"), val = tensor([-1])]; + tensor reduce_mean_92_keep_dims_0 = const()[name = tensor("reduce_mean_92_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_92_cast_fp16 = reduce_mean(axes = reduce_mean_92_axes_0, keep_dims = reduce_mean_92_keep_dims_0, x = linear_134_cast_fp16)[name = tensor("reduce_mean_92_cast_fp16")]; + tensor sub_57_cast_fp16 = sub(x = linear_134_cast_fp16, y = reduce_mean_92_cast_fp16)[name = tensor("sub_57_cast_fp16")]; + tensor square_46_cast_fp16 = square(x = sub_57_cast_fp16)[name = tensor("square_46_cast_fp16")]; + tensor reduce_mean_93_axes_0 = const()[name = tensor("reduce_mean_93_axes_0"), val = tensor([-1])]; + tensor reduce_mean_93_keep_dims_0 = const()[name = tensor("reduce_mean_93_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_93_cast_fp16 = reduce_mean(axes = reduce_mean_93_axes_0, keep_dims = reduce_mean_93_keep_dims_0, x = square_46_cast_fp16)[name = tensor("reduce_mean_93_cast_fp16")]; + tensor real_div_11_to_fp16 = const()[name = tensor("real_div_11_to_fp16"), val = tensor(0x1.008p+0)]; + tensor mul_11_cast_fp16 = mul(x = reduce_mean_93_cast_fp16, y = real_div_11_to_fp16)[name = tensor("mul_11_cast_fp16")]; + tensor var_2075_to_fp16 = const()[name = tensor("op_2075_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_93_cast_fp16 = add(x = mul_11_cast_fp16, y = var_2075_to_fp16)[name = tensor("var_93_cast_fp16")]; + tensor var_2078_epsilon_0 = const()[name = tensor("op_2078_epsilon_0"), val = tensor(0x1.197998p-40)]; + tensor var_2078_cast_fp16 = rsqrt(epsilon = var_2078_epsilon_0, x = var_93_cast_fp16)[name = tensor("op_2078_cast_fp16")]; + tensor var_2079_cast_fp16 = mul(x = const_5_to_fp16, y = var_2078_cast_fp16)[name = tensor("op_2079_cast_fp16")]; + tensor var_2080_cast_fp16 = mul(x = linear_134_cast_fp16, y = var_2079_cast_fp16)[name = tensor("op_2080_cast_fp16")]; + tensor var_2092_cast_fp16 = add(x = var_2043_cast_fp16, y = var_2080_cast_fp16)[name = tensor("op_2092_cast_fp16")]; + tensor _inversed_t_combined_11_y_0_to_fp16 = const()[name = tensor("_inversed_t_combined_11_y_0_to_fp16"), val = tensor(0x1p-1)]; + tensor _inversed_t_combined_11_cast_fp16 = mul(x = var_2092_cast_fp16, y = _inversed_t_combined_11_y_0_to_fp16)[name = tensor("_inversed_t_combined_11_cast_fp16")]; + tensor input_353_cast_fp16 = add(x = _inversed_t_combined_11_cast_fp16, y = linear_5_cast_fp16)[name = tensor("input_353_cast_fp16")]; + tensor input_355_cast_fp16 = silu(x = input_353_cast_fp16)[name = tensor("input_355_cast_fp16")]; + tensor linear_136_cast_fp16 = linear(bias = flow_net_res_blocks_0_adaLN_modulation_1_bias_to_fp16, weight = flow_net_res_blocks_0_adaLN_modulation_1_weight_to_fp16, x = input_355_cast_fp16)[name = tensor("linear_136_cast_fp16")]; + tensor var_2107_split_sizes_0 = const()[name = tensor("op_2107_split_sizes_0"), val = tensor([512, 512, 512])]; + tensor var_2107_axis_0 = const()[name = tensor("op_2107_axis_0"), val = tensor(-1)]; + tensor var_2107_cast_fp16_0, tensor var_2107_cast_fp16_1, tensor var_2107_cast_fp16_2 = split(axis = var_2107_axis_0, split_sizes = var_2107_split_sizes_0, x = linear_136_cast_fp16)[name = tensor("op_2107_cast_fp16")]; + tensor mean_71_axes_0 = const()[name = tensor("mean_71_axes_0"), val = tensor([-1])]; + tensor mean_71_keep_dims_0 = const()[name = tensor("mean_71_keep_dims_0"), val = tensor(true)]; + tensor mean_71_cast_fp16 = reduce_mean(axes = mean_71_axes_0, keep_dims = mean_71_keep_dims_0, x = linear_130_cast_fp16)[name = tensor("mean_71_cast_fp16")]; + tensor sub_59_cast_fp16 = sub(x = linear_130_cast_fp16, y = mean_71_cast_fp16)[name = tensor("sub_59_cast_fp16")]; + tensor square_47_cast_fp16 = square(x = sub_59_cast_fp16)[name = tensor("square_47_cast_fp16")]; + tensor reduce_mean_95_axes_0 = const()[name = tensor("reduce_mean_95_axes_0"), val = tensor([-1])]; + tensor reduce_mean_95_keep_dims_0 = const()[name = tensor("reduce_mean_95_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_95_cast_fp16 = reduce_mean(axes = reduce_mean_95_axes_0, keep_dims = reduce_mean_95_keep_dims_0, x = square_47_cast_fp16)[name = tensor("reduce_mean_95_cast_fp16")]; + tensor var_2117_to_fp16 = const()[name = tensor("op_2117_to_fp16"), val = tensor(0x1.1p-20)]; + tensor var_2118_cast_fp16 = add(x = reduce_mean_95_cast_fp16, y = var_2117_to_fp16)[name = tensor("op_2118_cast_fp16")]; + tensor var_2119_cast_fp16 = sqrt(x = var_2118_cast_fp16)[name = tensor("op_2119_cast_fp16")]; + tensor x_227_cast_fp16 = real_div(x = sub_59_cast_fp16, y = var_2119_cast_fp16)[name = tensor("x_227_cast_fp16")]; + tensor var_2121_cast_fp16 = mul(x = x_227_cast_fp16, y = flow_net_res_blocks_0_in_ln_weight_to_fp16)[name = tensor("op_2121_cast_fp16")]; + tensor x_229_cast_fp16 = add(x = var_2121_cast_fp16, y = flow_net_res_blocks_0_in_ln_bias_to_fp16)[name = tensor("x_229_cast_fp16")]; + tensor var_2123_promoted_to_fp16 = const()[name = tensor("op_2123_promoted_to_fp16"), val = tensor(0x1p+0)]; + tensor var_2124_cast_fp16 = add(x = var_2107_cast_fp16_1, y = var_2123_promoted_to_fp16)[name = tensor("op_2124_cast_fp16")]; + tensor var_2125_cast_fp16 = mul(x = x_229_cast_fp16, y = var_2124_cast_fp16)[name = tensor("op_2125_cast_fp16")]; + tensor input_357_cast_fp16 = add(x = var_2125_cast_fp16, y = var_2107_cast_fp16_0)[name = tensor("input_357_cast_fp16")]; + tensor linear_137_cast_fp16 = linear(bias = flow_net_res_blocks_0_mlp_0_bias_to_fp16, weight = flow_net_res_blocks_0_mlp_0_weight_to_fp16, x = input_357_cast_fp16)[name = tensor("linear_137_cast_fp16")]; + tensor input_361_cast_fp16 = silu(x = linear_137_cast_fp16)[name = tensor("input_361_cast_fp16")]; + tensor linear_138_cast_fp16 = linear(bias = flow_net_res_blocks_0_mlp_2_bias_to_fp16, weight = flow_net_res_blocks_0_mlp_2_weight_to_fp16, x = input_361_cast_fp16)[name = tensor("linear_138_cast_fp16")]; + tensor var_2136_cast_fp16 = mul(x = var_2107_cast_fp16_2, y = linear_138_cast_fp16)[name = tensor("op_2136_cast_fp16")]; + tensor x_231_cast_fp16 = add(x = linear_130_cast_fp16, y = var_2136_cast_fp16)[name = tensor("x_231_cast_fp16")]; + tensor linear_139_cast_fp16 = linear(bias = flow_net_res_blocks_1_adaLN_modulation_1_bias_to_fp16, weight = flow_net_res_blocks_1_adaLN_modulation_1_weight_to_fp16, x = input_355_cast_fp16)[name = tensor("linear_139_cast_fp16")]; + tensor var_2146_split_sizes_0 = const()[name = tensor("op_2146_split_sizes_0"), val = tensor([512, 512, 512])]; + tensor var_2146_axis_0 = const()[name = tensor("op_2146_axis_0"), val = tensor(-1)]; + tensor var_2146_cast_fp16_0, tensor var_2146_cast_fp16_1, tensor var_2146_cast_fp16_2 = split(axis = var_2146_axis_0, split_sizes = var_2146_split_sizes_0, x = linear_139_cast_fp16)[name = tensor("op_2146_cast_fp16")]; + tensor mean_73_axes_0 = const()[name = tensor("mean_73_axes_0"), val = tensor([-1])]; + tensor mean_73_keep_dims_0 = const()[name = tensor("mean_73_keep_dims_0"), val = tensor(true)]; + tensor mean_73_cast_fp16 = reduce_mean(axes = mean_73_axes_0, keep_dims = mean_73_keep_dims_0, x = x_231_cast_fp16)[name = tensor("mean_73_cast_fp16")]; + tensor sub_60_cast_fp16 = sub(x = x_231_cast_fp16, y = mean_73_cast_fp16)[name = tensor("sub_60_cast_fp16")]; + tensor square_48_cast_fp16 = square(x = sub_60_cast_fp16)[name = tensor("square_48_cast_fp16")]; + tensor reduce_mean_97_axes_0 = const()[name = tensor("reduce_mean_97_axes_0"), val = tensor([-1])]; + tensor reduce_mean_97_keep_dims_0 = const()[name = tensor("reduce_mean_97_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_97_cast_fp16 = reduce_mean(axes = reduce_mean_97_axes_0, keep_dims = reduce_mean_97_keep_dims_0, x = square_48_cast_fp16)[name = tensor("reduce_mean_97_cast_fp16")]; + tensor var_2156_to_fp16 = const()[name = tensor("op_2156_to_fp16"), val = tensor(0x1.1p-20)]; + tensor var_2157_cast_fp16 = add(x = reduce_mean_97_cast_fp16, y = var_2156_to_fp16)[name = tensor("op_2157_cast_fp16")]; + tensor var_2158_cast_fp16 = sqrt(x = var_2157_cast_fp16)[name = tensor("op_2158_cast_fp16")]; + tensor x_233_cast_fp16 = real_div(x = sub_60_cast_fp16, y = var_2158_cast_fp16)[name = tensor("x_233_cast_fp16")]; + tensor var_2160_cast_fp16 = mul(x = x_233_cast_fp16, y = flow_net_res_blocks_1_in_ln_weight_to_fp16)[name = tensor("op_2160_cast_fp16")]; + tensor x_235_cast_fp16 = add(x = var_2160_cast_fp16, y = flow_net_res_blocks_1_in_ln_bias_to_fp16)[name = tensor("x_235_cast_fp16")]; + tensor var_2162_promoted_to_fp16 = const()[name = tensor("op_2162_promoted_to_fp16"), val = tensor(0x1p+0)]; + tensor var_2163_cast_fp16 = add(x = var_2146_cast_fp16_1, y = var_2162_promoted_to_fp16)[name = tensor("op_2163_cast_fp16")]; + tensor var_2164_cast_fp16 = mul(x = x_235_cast_fp16, y = var_2163_cast_fp16)[name = tensor("op_2164_cast_fp16")]; + tensor input_365_cast_fp16 = add(x = var_2164_cast_fp16, y = var_2146_cast_fp16_0)[name = tensor("input_365_cast_fp16")]; + tensor linear_140_cast_fp16 = linear(bias = flow_net_res_blocks_1_mlp_0_bias_to_fp16, weight = flow_net_res_blocks_1_mlp_0_weight_to_fp16, x = input_365_cast_fp16)[name = tensor("linear_140_cast_fp16")]; + tensor input_369_cast_fp16 = silu(x = linear_140_cast_fp16)[name = tensor("input_369_cast_fp16")]; + tensor linear_141_cast_fp16 = linear(bias = flow_net_res_blocks_1_mlp_2_bias_to_fp16, weight = flow_net_res_blocks_1_mlp_2_weight_to_fp16, x = input_369_cast_fp16)[name = tensor("linear_141_cast_fp16")]; + tensor var_2175_cast_fp16 = mul(x = var_2146_cast_fp16_2, y = linear_141_cast_fp16)[name = tensor("op_2175_cast_fp16")]; + tensor x_237_cast_fp16 = add(x = x_231_cast_fp16, y = var_2175_cast_fp16)[name = tensor("x_237_cast_fp16")]; + tensor linear_142_cast_fp16 = linear(bias = flow_net_res_blocks_2_adaLN_modulation_1_bias_to_fp16, weight = flow_net_res_blocks_2_adaLN_modulation_1_weight_to_fp16, x = input_355_cast_fp16)[name = tensor("linear_142_cast_fp16")]; + tensor var_2185_split_sizes_0 = const()[name = tensor("op_2185_split_sizes_0"), val = tensor([512, 512, 512])]; + tensor var_2185_axis_0 = const()[name = tensor("op_2185_axis_0"), val = tensor(-1)]; + tensor var_2185_cast_fp16_0, tensor var_2185_cast_fp16_1, tensor var_2185_cast_fp16_2 = split(axis = var_2185_axis_0, split_sizes = var_2185_split_sizes_0, x = linear_142_cast_fp16)[name = tensor("op_2185_cast_fp16")]; + tensor mean_75_axes_0 = const()[name = tensor("mean_75_axes_0"), val = tensor([-1])]; + tensor mean_75_keep_dims_0 = const()[name = tensor("mean_75_keep_dims_0"), val = tensor(true)]; + tensor mean_75_cast_fp16 = reduce_mean(axes = mean_75_axes_0, keep_dims = mean_75_keep_dims_0, x = x_237_cast_fp16)[name = tensor("mean_75_cast_fp16")]; + tensor sub_61_cast_fp16 = sub(x = x_237_cast_fp16, y = mean_75_cast_fp16)[name = tensor("sub_61_cast_fp16")]; + tensor square_49_cast_fp16 = square(x = sub_61_cast_fp16)[name = tensor("square_49_cast_fp16")]; + tensor reduce_mean_99_axes_0 = const()[name = tensor("reduce_mean_99_axes_0"), val = tensor([-1])]; + tensor reduce_mean_99_keep_dims_0 = const()[name = tensor("reduce_mean_99_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_99_cast_fp16 = reduce_mean(axes = reduce_mean_99_axes_0, keep_dims = reduce_mean_99_keep_dims_0, x = square_49_cast_fp16)[name = tensor("reduce_mean_99_cast_fp16")]; + tensor var_2195_to_fp16 = const()[name = tensor("op_2195_to_fp16"), val = tensor(0x1.1p-20)]; + tensor var_2196_cast_fp16 = add(x = reduce_mean_99_cast_fp16, y = var_2195_to_fp16)[name = tensor("op_2196_cast_fp16")]; + tensor var_2197_cast_fp16 = sqrt(x = var_2196_cast_fp16)[name = tensor("op_2197_cast_fp16")]; + tensor x_239_cast_fp16 = real_div(x = sub_61_cast_fp16, y = var_2197_cast_fp16)[name = tensor("x_239_cast_fp16")]; + tensor var_2199_cast_fp16 = mul(x = x_239_cast_fp16, y = flow_net_res_blocks_2_in_ln_weight_to_fp16)[name = tensor("op_2199_cast_fp16")]; + tensor x_241_cast_fp16 = add(x = var_2199_cast_fp16, y = flow_net_res_blocks_2_in_ln_bias_to_fp16)[name = tensor("x_241_cast_fp16")]; + tensor var_2201_promoted_to_fp16 = const()[name = tensor("op_2201_promoted_to_fp16"), val = tensor(0x1p+0)]; + tensor var_2202_cast_fp16 = add(x = var_2185_cast_fp16_1, y = var_2201_promoted_to_fp16)[name = tensor("op_2202_cast_fp16")]; + tensor var_2203_cast_fp16 = mul(x = x_241_cast_fp16, y = var_2202_cast_fp16)[name = tensor("op_2203_cast_fp16")]; + tensor input_373_cast_fp16 = add(x = var_2203_cast_fp16, y = var_2185_cast_fp16_0)[name = tensor("input_373_cast_fp16")]; + tensor linear_143_cast_fp16 = linear(bias = flow_net_res_blocks_2_mlp_0_bias_to_fp16, weight = flow_net_res_blocks_2_mlp_0_weight_to_fp16, x = input_373_cast_fp16)[name = tensor("linear_143_cast_fp16")]; + tensor input_377_cast_fp16 = silu(x = linear_143_cast_fp16)[name = tensor("input_377_cast_fp16")]; + tensor linear_144_cast_fp16 = linear(bias = flow_net_res_blocks_2_mlp_2_bias_to_fp16, weight = flow_net_res_blocks_2_mlp_2_weight_to_fp16, x = input_377_cast_fp16)[name = tensor("linear_144_cast_fp16")]; + tensor var_2214_cast_fp16 = mul(x = var_2185_cast_fp16_2, y = linear_144_cast_fp16)[name = tensor("op_2214_cast_fp16")]; + tensor x_243_cast_fp16 = add(x = x_237_cast_fp16, y = var_2214_cast_fp16)[name = tensor("x_243_cast_fp16")]; + tensor linear_145_cast_fp16 = linear(bias = flow_net_res_blocks_3_adaLN_modulation_1_bias_to_fp16, weight = flow_net_res_blocks_3_adaLN_modulation_1_weight_to_fp16, x = input_355_cast_fp16)[name = tensor("linear_145_cast_fp16")]; + tensor var_2224_split_sizes_0 = const()[name = tensor("op_2224_split_sizes_0"), val = tensor([512, 512, 512])]; + tensor var_2224_axis_0 = const()[name = tensor("op_2224_axis_0"), val = tensor(-1)]; + tensor var_2224_cast_fp16_0, tensor var_2224_cast_fp16_1, tensor var_2224_cast_fp16_2 = split(axis = var_2224_axis_0, split_sizes = var_2224_split_sizes_0, x = linear_145_cast_fp16)[name = tensor("op_2224_cast_fp16")]; + tensor mean_77_axes_0 = const()[name = tensor("mean_77_axes_0"), val = tensor([-1])]; + tensor mean_77_keep_dims_0 = const()[name = tensor("mean_77_keep_dims_0"), val = tensor(true)]; + tensor mean_77_cast_fp16 = reduce_mean(axes = mean_77_axes_0, keep_dims = mean_77_keep_dims_0, x = x_243_cast_fp16)[name = tensor("mean_77_cast_fp16")]; + tensor sub_62_cast_fp16 = sub(x = x_243_cast_fp16, y = mean_77_cast_fp16)[name = tensor("sub_62_cast_fp16")]; + tensor square_50_cast_fp16 = square(x = sub_62_cast_fp16)[name = tensor("square_50_cast_fp16")]; + tensor reduce_mean_101_axes_0 = const()[name = tensor("reduce_mean_101_axes_0"), val = tensor([-1])]; + tensor reduce_mean_101_keep_dims_0 = const()[name = tensor("reduce_mean_101_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_101_cast_fp16 = reduce_mean(axes = reduce_mean_101_axes_0, keep_dims = reduce_mean_101_keep_dims_0, x = square_50_cast_fp16)[name = tensor("reduce_mean_101_cast_fp16")]; + tensor var_2234_to_fp16 = const()[name = tensor("op_2234_to_fp16"), val = tensor(0x1.1p-20)]; + tensor var_2235_cast_fp16 = add(x = reduce_mean_101_cast_fp16, y = var_2234_to_fp16)[name = tensor("op_2235_cast_fp16")]; + tensor var_2236_cast_fp16 = sqrt(x = var_2235_cast_fp16)[name = tensor("op_2236_cast_fp16")]; + tensor x_245_cast_fp16 = real_div(x = sub_62_cast_fp16, y = var_2236_cast_fp16)[name = tensor("x_245_cast_fp16")]; + tensor var_2238_cast_fp16 = mul(x = x_245_cast_fp16, y = flow_net_res_blocks_3_in_ln_weight_to_fp16)[name = tensor("op_2238_cast_fp16")]; + tensor x_247_cast_fp16 = add(x = var_2238_cast_fp16, y = flow_net_res_blocks_3_in_ln_bias_to_fp16)[name = tensor("x_247_cast_fp16")]; + tensor var_2240_promoted_to_fp16 = const()[name = tensor("op_2240_promoted_to_fp16"), val = tensor(0x1p+0)]; + tensor var_2241_cast_fp16 = add(x = var_2224_cast_fp16_1, y = var_2240_promoted_to_fp16)[name = tensor("op_2241_cast_fp16")]; + tensor var_2242_cast_fp16 = mul(x = x_247_cast_fp16, y = var_2241_cast_fp16)[name = tensor("op_2242_cast_fp16")]; + tensor input_381_cast_fp16 = add(x = var_2242_cast_fp16, y = var_2224_cast_fp16_0)[name = tensor("input_381_cast_fp16")]; + tensor linear_146_cast_fp16 = linear(bias = flow_net_res_blocks_3_mlp_0_bias_to_fp16, weight = flow_net_res_blocks_3_mlp_0_weight_to_fp16, x = input_381_cast_fp16)[name = tensor("linear_146_cast_fp16")]; + tensor input_385_cast_fp16 = silu(x = linear_146_cast_fp16)[name = tensor("input_385_cast_fp16")]; + tensor linear_147_cast_fp16 = linear(bias = flow_net_res_blocks_3_mlp_2_bias_to_fp16, weight = flow_net_res_blocks_3_mlp_2_weight_to_fp16, x = input_385_cast_fp16)[name = tensor("linear_147_cast_fp16")]; + tensor var_2253_cast_fp16 = mul(x = var_2224_cast_fp16_2, y = linear_147_cast_fp16)[name = tensor("op_2253_cast_fp16")]; + tensor x_249_cast_fp16 = add(x = x_243_cast_fp16, y = var_2253_cast_fp16)[name = tensor("x_249_cast_fp16")]; + tensor linear_148_cast_fp16 = linear(bias = flow_net_res_blocks_4_adaLN_modulation_1_bias_to_fp16, weight = flow_net_res_blocks_4_adaLN_modulation_1_weight_to_fp16, x = input_355_cast_fp16)[name = tensor("linear_148_cast_fp16")]; + tensor var_2263_split_sizes_0 = const()[name = tensor("op_2263_split_sizes_0"), val = tensor([512, 512, 512])]; + tensor var_2263_axis_0 = const()[name = tensor("op_2263_axis_0"), val = tensor(-1)]; + tensor var_2263_cast_fp16_0, tensor var_2263_cast_fp16_1, tensor var_2263_cast_fp16_2 = split(axis = var_2263_axis_0, split_sizes = var_2263_split_sizes_0, x = linear_148_cast_fp16)[name = tensor("op_2263_cast_fp16")]; + tensor mean_79_axes_0 = const()[name = tensor("mean_79_axes_0"), val = tensor([-1])]; + tensor mean_79_keep_dims_0 = const()[name = tensor("mean_79_keep_dims_0"), val = tensor(true)]; + tensor mean_79_cast_fp16 = reduce_mean(axes = mean_79_axes_0, keep_dims = mean_79_keep_dims_0, x = x_249_cast_fp16)[name = tensor("mean_79_cast_fp16")]; + tensor sub_63_cast_fp16 = sub(x = x_249_cast_fp16, y = mean_79_cast_fp16)[name = tensor("sub_63_cast_fp16")]; + tensor square_51_cast_fp16 = square(x = sub_63_cast_fp16)[name = tensor("square_51_cast_fp16")]; + tensor reduce_mean_103_axes_0 = const()[name = tensor("reduce_mean_103_axes_0"), val = tensor([-1])]; + tensor reduce_mean_103_keep_dims_0 = const()[name = tensor("reduce_mean_103_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_103_cast_fp16 = reduce_mean(axes = reduce_mean_103_axes_0, keep_dims = reduce_mean_103_keep_dims_0, x = square_51_cast_fp16)[name = tensor("reduce_mean_103_cast_fp16")]; + tensor var_2273_to_fp16 = const()[name = tensor("op_2273_to_fp16"), val = tensor(0x1.1p-20)]; + tensor var_2274_cast_fp16 = add(x = reduce_mean_103_cast_fp16, y = var_2273_to_fp16)[name = tensor("op_2274_cast_fp16")]; + tensor var_2275_cast_fp16 = sqrt(x = var_2274_cast_fp16)[name = tensor("op_2275_cast_fp16")]; + tensor x_251_cast_fp16 = real_div(x = sub_63_cast_fp16, y = var_2275_cast_fp16)[name = tensor("x_251_cast_fp16")]; + tensor var_2277_cast_fp16 = mul(x = x_251_cast_fp16, y = flow_net_res_blocks_4_in_ln_weight_to_fp16)[name = tensor("op_2277_cast_fp16")]; + tensor x_253_cast_fp16 = add(x = var_2277_cast_fp16, y = flow_net_res_blocks_4_in_ln_bias_to_fp16)[name = tensor("x_253_cast_fp16")]; + tensor var_2279_promoted_to_fp16 = const()[name = tensor("op_2279_promoted_to_fp16"), val = tensor(0x1p+0)]; + tensor var_2280_cast_fp16 = add(x = var_2263_cast_fp16_1, y = var_2279_promoted_to_fp16)[name = tensor("op_2280_cast_fp16")]; + tensor var_2281_cast_fp16 = mul(x = x_253_cast_fp16, y = var_2280_cast_fp16)[name = tensor("op_2281_cast_fp16")]; + tensor input_389_cast_fp16 = add(x = var_2281_cast_fp16, y = var_2263_cast_fp16_0)[name = tensor("input_389_cast_fp16")]; + tensor linear_149_cast_fp16 = linear(bias = flow_net_res_blocks_4_mlp_0_bias_to_fp16, weight = flow_net_res_blocks_4_mlp_0_weight_to_fp16, x = input_389_cast_fp16)[name = tensor("linear_149_cast_fp16")]; + tensor input_393_cast_fp16 = silu(x = linear_149_cast_fp16)[name = tensor("input_393_cast_fp16")]; + tensor linear_150_cast_fp16 = linear(bias = flow_net_res_blocks_4_mlp_2_bias_to_fp16, weight = flow_net_res_blocks_4_mlp_2_weight_to_fp16, x = input_393_cast_fp16)[name = tensor("linear_150_cast_fp16")]; + tensor var_2292_cast_fp16 = mul(x = var_2263_cast_fp16_2, y = linear_150_cast_fp16)[name = tensor("op_2292_cast_fp16")]; + tensor x_255_cast_fp16 = add(x = x_249_cast_fp16, y = var_2292_cast_fp16)[name = tensor("x_255_cast_fp16")]; + tensor linear_151_cast_fp16 = linear(bias = flow_net_res_blocks_5_adaLN_modulation_1_bias_to_fp16, weight = flow_net_res_blocks_5_adaLN_modulation_1_weight_to_fp16, x = input_355_cast_fp16)[name = tensor("linear_151_cast_fp16")]; + tensor var_2302_split_sizes_0 = const()[name = tensor("op_2302_split_sizes_0"), val = tensor([512, 512, 512])]; + tensor var_2302_axis_0 = const()[name = tensor("op_2302_axis_0"), val = tensor(-1)]; + tensor var_2302_cast_fp16_0, tensor var_2302_cast_fp16_1, tensor var_2302_cast_fp16_2 = split(axis = var_2302_axis_0, split_sizes = var_2302_split_sizes_0, x = linear_151_cast_fp16)[name = tensor("op_2302_cast_fp16")]; + tensor mean_81_axes_0 = const()[name = tensor("mean_81_axes_0"), val = tensor([-1])]; + tensor mean_81_keep_dims_0 = const()[name = tensor("mean_81_keep_dims_0"), val = tensor(true)]; + tensor mean_81_cast_fp16 = reduce_mean(axes = mean_81_axes_0, keep_dims = mean_81_keep_dims_0, x = x_255_cast_fp16)[name = tensor("mean_81_cast_fp16")]; + tensor sub_64_cast_fp16 = sub(x = x_255_cast_fp16, y = mean_81_cast_fp16)[name = tensor("sub_64_cast_fp16")]; + tensor square_52_cast_fp16 = square(x = sub_64_cast_fp16)[name = tensor("square_52_cast_fp16")]; + tensor reduce_mean_105_axes_0 = const()[name = tensor("reduce_mean_105_axes_0"), val = tensor([-1])]; + tensor reduce_mean_105_keep_dims_0 = const()[name = tensor("reduce_mean_105_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_105_cast_fp16 = reduce_mean(axes = reduce_mean_105_axes_0, keep_dims = reduce_mean_105_keep_dims_0, x = square_52_cast_fp16)[name = tensor("reduce_mean_105_cast_fp16")]; + tensor var_2312_to_fp16 = const()[name = tensor("op_2312_to_fp16"), val = tensor(0x1.1p-20)]; + tensor var_2313_cast_fp16 = add(x = reduce_mean_105_cast_fp16, y = var_2312_to_fp16)[name = tensor("op_2313_cast_fp16")]; + tensor var_2314_cast_fp16 = sqrt(x = var_2313_cast_fp16)[name = tensor("op_2314_cast_fp16")]; + tensor x_257_cast_fp16 = real_div(x = sub_64_cast_fp16, y = var_2314_cast_fp16)[name = tensor("x_257_cast_fp16")]; + tensor var_2316_cast_fp16 = mul(x = x_257_cast_fp16, y = flow_net_res_blocks_5_in_ln_weight_to_fp16)[name = tensor("op_2316_cast_fp16")]; + tensor x_259_cast_fp16 = add(x = var_2316_cast_fp16, y = flow_net_res_blocks_5_in_ln_bias_to_fp16)[name = tensor("x_259_cast_fp16")]; + tensor var_2318_promoted_to_fp16 = const()[name = tensor("op_2318_promoted_to_fp16"), val = tensor(0x1p+0)]; + tensor var_2319_cast_fp16 = add(x = var_2302_cast_fp16_1, y = var_2318_promoted_to_fp16)[name = tensor("op_2319_cast_fp16")]; + tensor var_2320_cast_fp16 = mul(x = x_259_cast_fp16, y = var_2319_cast_fp16)[name = tensor("op_2320_cast_fp16")]; + tensor input_397_cast_fp16 = add(x = var_2320_cast_fp16, y = var_2302_cast_fp16_0)[name = tensor("input_397_cast_fp16")]; + tensor linear_152_cast_fp16 = linear(bias = flow_net_res_blocks_5_mlp_0_bias_to_fp16, weight = flow_net_res_blocks_5_mlp_0_weight_to_fp16, x = input_397_cast_fp16)[name = tensor("linear_152_cast_fp16")]; + tensor input_401_cast_fp16 = silu(x = linear_152_cast_fp16)[name = tensor("input_401_cast_fp16")]; + tensor linear_153_cast_fp16 = linear(bias = flow_net_res_blocks_5_mlp_2_bias_to_fp16, weight = flow_net_res_blocks_5_mlp_2_weight_to_fp16, x = input_401_cast_fp16)[name = tensor("linear_153_cast_fp16")]; + tensor var_2331_cast_fp16 = mul(x = var_2302_cast_fp16_2, y = linear_153_cast_fp16)[name = tensor("op_2331_cast_fp16")]; + tensor x_261_cast_fp16 = add(x = x_255_cast_fp16, y = var_2331_cast_fp16)[name = tensor("x_261_cast_fp16")]; + tensor linear_154_cast_fp16 = linear(bias = flow_net_final_layer_adaLN_modulation_1_bias_to_fp16, weight = flow_net_final_layer_adaLN_modulation_1_weight_to_fp16, x = input_355_cast_fp16)[name = tensor("linear_154_cast_fp16")]; + tensor var_2340_split_sizes_0 = const()[name = tensor("op_2340_split_sizes_0"), val = tensor([512, 512])]; + tensor var_2340_axis_0 = const()[name = tensor("op_2340_axis_0"), val = tensor(-1)]; + tensor var_2340_cast_fp16_0, tensor var_2340_cast_fp16_1 = split(axis = var_2340_axis_0, split_sizes = var_2340_split_sizes_0, x = linear_154_cast_fp16)[name = tensor("op_2340_cast_fp16")]; + tensor mean_83_axes_0 = const()[name = tensor("mean_83_axes_0"), val = tensor([-1])]; + tensor mean_83_keep_dims_0 = const()[name = tensor("mean_83_keep_dims_0"), val = tensor(true)]; + tensor mean_83_cast_fp16 = reduce_mean(axes = mean_83_axes_0, keep_dims = mean_83_keep_dims_0, x = x_261_cast_fp16)[name = tensor("mean_83_cast_fp16")]; + tensor sub_65_cast_fp16 = sub(x = x_261_cast_fp16, y = mean_83_cast_fp16)[name = tensor("sub_65_cast_fp16")]; + tensor square_53_cast_fp16 = square(x = sub_65_cast_fp16)[name = tensor("square_53_cast_fp16")]; + tensor reduce_mean_107_axes_0 = const()[name = tensor("reduce_mean_107_axes_0"), val = tensor([-1])]; + tensor reduce_mean_107_keep_dims_0 = const()[name = tensor("reduce_mean_107_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_107_cast_fp16 = reduce_mean(axes = reduce_mean_107_axes_0, keep_dims = reduce_mean_107_keep_dims_0, x = square_53_cast_fp16)[name = tensor("reduce_mean_107_cast_fp16")]; + tensor var_2347_to_fp16 = const()[name = tensor("op_2347_to_fp16"), val = tensor(0x1.1p-20)]; + tensor var_2348_cast_fp16 = add(x = reduce_mean_107_cast_fp16, y = var_2347_to_fp16)[name = tensor("op_2348_cast_fp16")]; + tensor var_2349_cast_fp16 = sqrt(x = var_2348_cast_fp16)[name = tensor("op_2349_cast_fp16")]; + tensor x_263_cast_fp16 = real_div(x = sub_65_cast_fp16, y = var_2349_cast_fp16)[name = tensor("x_263_cast_fp16")]; + tensor var_2351_promoted_to_fp16 = const()[name = tensor("op_2351_promoted_to_fp16"), val = tensor(0x1p+0)]; + tensor var_2352_cast_fp16 = add(x = var_2340_cast_fp16_1, y = var_2351_promoted_to_fp16)[name = tensor("op_2352_cast_fp16")]; + tensor var_2353_cast_fp16 = mul(x = x_263_cast_fp16, y = var_2352_cast_fp16)[name = tensor("op_2353_cast_fp16")]; + tensor input_405_cast_fp16 = add(x = var_2353_cast_fp16, y = var_2340_cast_fp16_0)[name = tensor("input_405_cast_fp16")]; + tensor linear_155_cast_fp16 = linear(bias = flow_net_final_layer_linear_bias_to_fp16, weight = flow_net_final_layer_linear_weight_to_fp16, x = input_405_cast_fp16)[name = tensor("linear_155_cast_fp16")]; + tensor var_2364_to_fp16 = const()[name = tensor("op_2364_to_fp16"), val = tensor(0x1p-3)]; + tensor var_2365_cast_fp16 = mul(x = linear_155_cast_fp16, y = var_2364_to_fp16)[name = tensor("op_2365_cast_fp16")]; + tensor input_407_cast_fp16 = add(x = input_339_cast_fp16, y = var_2365_cast_fp16)[name = tensor("input_407_cast_fp16")]; + tensor linear_156_cast_fp16 = linear(bias = flow_net_input_proj_bias_to_fp16, weight = flow_net_input_proj_weight_to_fp16, x = input_407_cast_fp16)[name = tensor("linear_156_cast_fp16")]; + tensor input_411_to_fp16 = const()[name = tensor("input_411_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(19008256)))]; + tensor input_413_cast_fp16 = silu(x = input_411_to_fp16)[name = tensor("input_413_cast_fp16")]; + tensor linear_158_cast_fp16 = linear(bias = flow_net_time_embed_0_mlp_2_bias_to_fp16, weight = flow_net_time_embed_0_mlp_2_weight_to_fp16, x = input_413_cast_fp16)[name = tensor("linear_158_cast_fp16")]; + tensor reduce_mean_108_axes_0 = const()[name = tensor("reduce_mean_108_axes_0"), val = tensor([-1])]; + tensor reduce_mean_108_keep_dims_0 = const()[name = tensor("reduce_mean_108_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_108_cast_fp16 = reduce_mean(axes = reduce_mean_108_axes_0, keep_dims = reduce_mean_108_keep_dims_0, x = linear_158_cast_fp16)[name = tensor("reduce_mean_108_cast_fp16")]; + tensor sub_66_cast_fp16 = sub(x = linear_158_cast_fp16, y = reduce_mean_108_cast_fp16)[name = tensor("sub_66_cast_fp16")]; + tensor square_54_cast_fp16 = square(x = sub_66_cast_fp16)[name = tensor("square_54_cast_fp16")]; + tensor reduce_mean_109_axes_0 = const()[name = tensor("reduce_mean_109_axes_0"), val = tensor([-1])]; + tensor reduce_mean_109_keep_dims_0 = const()[name = tensor("reduce_mean_109_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_109_cast_fp16 = reduce_mean(axes = reduce_mean_109_axes_0, keep_dims = reduce_mean_109_keep_dims_0, x = square_54_cast_fp16)[name = tensor("reduce_mean_109_cast_fp16")]; + tensor real_div_12_to_fp16 = const()[name = tensor("real_div_12_to_fp16"), val = tensor(0x1.008p+0)]; + tensor mul_12_cast_fp16 = mul(x = reduce_mean_109_cast_fp16, y = real_div_12_to_fp16)[name = tensor("mul_12_cast_fp16")]; + tensor var_2431_to_fp16 = const()[name = tensor("op_2431_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_109_cast_fp16 = add(x = mul_12_cast_fp16, y = var_2431_to_fp16)[name = tensor("var_109_cast_fp16")]; + tensor var_2434_epsilon_0 = const()[name = tensor("op_2434_epsilon_0"), val = tensor(0x1.197998p-40)]; + tensor var_2434_cast_fp16 = rsqrt(epsilon = var_2434_epsilon_0, x = var_109_cast_fp16)[name = tensor("op_2434_cast_fp16")]; + tensor var_2435_cast_fp16 = mul(x = const_3_to_fp16, y = var_2434_cast_fp16)[name = tensor("op_2435_cast_fp16")]; + tensor var_2436_cast_fp16 = mul(x = linear_158_cast_fp16, y = var_2435_cast_fp16)[name = tensor("op_2436_cast_fp16")]; + tensor input_417_to_fp16 = const()[name = tensor("input_417_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(19009344)))]; + tensor input_419_cast_fp16 = silu(x = input_417_to_fp16)[name = tensor("input_419_cast_fp16")]; + tensor linear_160_cast_fp16 = linear(bias = flow_net_time_embed_1_mlp_2_bias_to_fp16, weight = flow_net_time_embed_1_mlp_2_weight_to_fp16, x = input_419_cast_fp16)[name = tensor("linear_160_cast_fp16")]; + tensor reduce_mean_110_axes_0 = const()[name = tensor("reduce_mean_110_axes_0"), val = tensor([-1])]; + tensor reduce_mean_110_keep_dims_0 = const()[name = tensor("reduce_mean_110_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_110_cast_fp16 = reduce_mean(axes = reduce_mean_110_axes_0, keep_dims = reduce_mean_110_keep_dims_0, x = linear_160_cast_fp16)[name = tensor("reduce_mean_110_cast_fp16")]; + tensor sub_68_cast_fp16 = sub(x = linear_160_cast_fp16, y = reduce_mean_110_cast_fp16)[name = tensor("sub_68_cast_fp16")]; + tensor square_55_cast_fp16 = square(x = sub_68_cast_fp16)[name = tensor("square_55_cast_fp16")]; + tensor reduce_mean_111_axes_0 = const()[name = tensor("reduce_mean_111_axes_0"), val = tensor([-1])]; + tensor reduce_mean_111_keep_dims_0 = const()[name = tensor("reduce_mean_111_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_111_cast_fp16 = reduce_mean(axes = reduce_mean_111_axes_0, keep_dims = reduce_mean_111_keep_dims_0, x = square_55_cast_fp16)[name = tensor("reduce_mean_111_cast_fp16")]; + tensor real_div_13_to_fp16 = const()[name = tensor("real_div_13_to_fp16"), val = tensor(0x1.008p+0)]; + tensor mul_13_cast_fp16 = mul(x = reduce_mean_111_cast_fp16, y = real_div_13_to_fp16)[name = tensor("mul_13_cast_fp16")]; + tensor var_2468_to_fp16 = const()[name = tensor("op_2468_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_111_cast_fp16 = add(x = mul_13_cast_fp16, y = var_2468_to_fp16)[name = tensor("var_111_cast_fp16")]; + tensor var_2471_epsilon_0 = const()[name = tensor("op_2471_epsilon_0"), val = tensor(0x1.197998p-40)]; + tensor var_2471_cast_fp16 = rsqrt(epsilon = var_2471_epsilon_0, x = var_111_cast_fp16)[name = tensor("op_2471_cast_fp16")]; + tensor var_2472_cast_fp16 = mul(x = const_5_to_fp16, y = var_2471_cast_fp16)[name = tensor("op_2472_cast_fp16")]; + tensor var_2473_cast_fp16 = mul(x = linear_160_cast_fp16, y = var_2472_cast_fp16)[name = tensor("op_2473_cast_fp16")]; + tensor var_2485_cast_fp16 = add(x = var_2436_cast_fp16, y = var_2473_cast_fp16)[name = tensor("op_2485_cast_fp16")]; + tensor _inversed_t_combined_13_y_0_to_fp16 = const()[name = tensor("_inversed_t_combined_13_y_0_to_fp16"), val = tensor(0x1p-1)]; + tensor _inversed_t_combined_13_cast_fp16 = mul(x = var_2485_cast_fp16, y = _inversed_t_combined_13_y_0_to_fp16)[name = tensor("_inversed_t_combined_13_cast_fp16")]; + tensor input_421_cast_fp16 = add(x = _inversed_t_combined_13_cast_fp16, y = linear_5_cast_fp16)[name = tensor("input_421_cast_fp16")]; + tensor input_423_cast_fp16 = silu(x = input_421_cast_fp16)[name = tensor("input_423_cast_fp16")]; + tensor linear_162_cast_fp16 = linear(bias = flow_net_res_blocks_0_adaLN_modulation_1_bias_to_fp16, weight = flow_net_res_blocks_0_adaLN_modulation_1_weight_to_fp16, x = input_423_cast_fp16)[name = tensor("linear_162_cast_fp16")]; + tensor var_2500_split_sizes_0 = const()[name = tensor("op_2500_split_sizes_0"), val = tensor([512, 512, 512])]; + tensor var_2500_axis_0 = const()[name = tensor("op_2500_axis_0"), val = tensor(-1)]; + tensor var_2500_cast_fp16_0, tensor var_2500_cast_fp16_1, tensor var_2500_cast_fp16_2 = split(axis = var_2500_axis_0, split_sizes = var_2500_split_sizes_0, x = linear_162_cast_fp16)[name = tensor("op_2500_cast_fp16")]; + tensor mean_85_axes_0 = const()[name = tensor("mean_85_axes_0"), val = tensor([-1])]; + tensor mean_85_keep_dims_0 = const()[name = tensor("mean_85_keep_dims_0"), val = tensor(true)]; + tensor mean_85_cast_fp16 = reduce_mean(axes = mean_85_axes_0, keep_dims = mean_85_keep_dims_0, x = linear_156_cast_fp16)[name = tensor("mean_85_cast_fp16")]; + tensor sub_70_cast_fp16 = sub(x = linear_156_cast_fp16, y = mean_85_cast_fp16)[name = tensor("sub_70_cast_fp16")]; + tensor square_56_cast_fp16 = square(x = sub_70_cast_fp16)[name = tensor("square_56_cast_fp16")]; + tensor reduce_mean_113_axes_0 = const()[name = tensor("reduce_mean_113_axes_0"), val = tensor([-1])]; + tensor reduce_mean_113_keep_dims_0 = const()[name = tensor("reduce_mean_113_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_113_cast_fp16 = reduce_mean(axes = reduce_mean_113_axes_0, keep_dims = reduce_mean_113_keep_dims_0, x = square_56_cast_fp16)[name = tensor("reduce_mean_113_cast_fp16")]; + tensor var_2510_to_fp16 = const()[name = tensor("op_2510_to_fp16"), val = tensor(0x1.1p-20)]; + tensor var_2511_cast_fp16 = add(x = reduce_mean_113_cast_fp16, y = var_2510_to_fp16)[name = tensor("op_2511_cast_fp16")]; + tensor var_2512_cast_fp16 = sqrt(x = var_2511_cast_fp16)[name = tensor("op_2512_cast_fp16")]; + tensor x_271_cast_fp16 = real_div(x = sub_70_cast_fp16, y = var_2512_cast_fp16)[name = tensor("x_271_cast_fp16")]; + tensor var_2514_cast_fp16 = mul(x = x_271_cast_fp16, y = flow_net_res_blocks_0_in_ln_weight_to_fp16)[name = tensor("op_2514_cast_fp16")]; + tensor x_273_cast_fp16 = add(x = var_2514_cast_fp16, y = flow_net_res_blocks_0_in_ln_bias_to_fp16)[name = tensor("x_273_cast_fp16")]; + tensor var_2516_promoted_to_fp16 = const()[name = tensor("op_2516_promoted_to_fp16"), val = tensor(0x1p+0)]; + tensor var_2517_cast_fp16 = add(x = var_2500_cast_fp16_1, y = var_2516_promoted_to_fp16)[name = tensor("op_2517_cast_fp16")]; + tensor var_2518_cast_fp16 = mul(x = x_273_cast_fp16, y = var_2517_cast_fp16)[name = tensor("op_2518_cast_fp16")]; + tensor input_425_cast_fp16 = add(x = var_2518_cast_fp16, y = var_2500_cast_fp16_0)[name = tensor("input_425_cast_fp16")]; + tensor linear_163_cast_fp16 = linear(bias = flow_net_res_blocks_0_mlp_0_bias_to_fp16, weight = flow_net_res_blocks_0_mlp_0_weight_to_fp16, x = input_425_cast_fp16)[name = tensor("linear_163_cast_fp16")]; + tensor input_429_cast_fp16 = silu(x = linear_163_cast_fp16)[name = tensor("input_429_cast_fp16")]; + tensor linear_164_cast_fp16 = linear(bias = flow_net_res_blocks_0_mlp_2_bias_to_fp16, weight = flow_net_res_blocks_0_mlp_2_weight_to_fp16, x = input_429_cast_fp16)[name = tensor("linear_164_cast_fp16")]; + tensor var_2529_cast_fp16 = mul(x = var_2500_cast_fp16_2, y = linear_164_cast_fp16)[name = tensor("op_2529_cast_fp16")]; + tensor x_275_cast_fp16 = add(x = linear_156_cast_fp16, y = var_2529_cast_fp16)[name = tensor("x_275_cast_fp16")]; + tensor linear_165_cast_fp16 = linear(bias = flow_net_res_blocks_1_adaLN_modulation_1_bias_to_fp16, weight = flow_net_res_blocks_1_adaLN_modulation_1_weight_to_fp16, x = input_423_cast_fp16)[name = tensor("linear_165_cast_fp16")]; + tensor var_2539_split_sizes_0 = const()[name = tensor("op_2539_split_sizes_0"), val = tensor([512, 512, 512])]; + tensor var_2539_axis_0 = const()[name = tensor("op_2539_axis_0"), val = tensor(-1)]; + tensor var_2539_cast_fp16_0, tensor var_2539_cast_fp16_1, tensor var_2539_cast_fp16_2 = split(axis = var_2539_axis_0, split_sizes = var_2539_split_sizes_0, x = linear_165_cast_fp16)[name = tensor("op_2539_cast_fp16")]; + tensor mean_87_axes_0 = const()[name = tensor("mean_87_axes_0"), val = tensor([-1])]; + tensor mean_87_keep_dims_0 = const()[name = tensor("mean_87_keep_dims_0"), val = tensor(true)]; + tensor mean_87_cast_fp16 = reduce_mean(axes = mean_87_axes_0, keep_dims = mean_87_keep_dims_0, x = x_275_cast_fp16)[name = tensor("mean_87_cast_fp16")]; + tensor sub_71_cast_fp16 = sub(x = x_275_cast_fp16, y = mean_87_cast_fp16)[name = tensor("sub_71_cast_fp16")]; + tensor square_57_cast_fp16 = square(x = sub_71_cast_fp16)[name = tensor("square_57_cast_fp16")]; + tensor reduce_mean_115_axes_0 = const()[name = tensor("reduce_mean_115_axes_0"), val = tensor([-1])]; + tensor reduce_mean_115_keep_dims_0 = const()[name = tensor("reduce_mean_115_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_115_cast_fp16 = reduce_mean(axes = reduce_mean_115_axes_0, keep_dims = reduce_mean_115_keep_dims_0, x = square_57_cast_fp16)[name = tensor("reduce_mean_115_cast_fp16")]; + tensor var_2549_to_fp16 = const()[name = tensor("op_2549_to_fp16"), val = tensor(0x1.1p-20)]; + tensor var_2550_cast_fp16 = add(x = reduce_mean_115_cast_fp16, y = var_2549_to_fp16)[name = tensor("op_2550_cast_fp16")]; + tensor var_2551_cast_fp16 = sqrt(x = var_2550_cast_fp16)[name = tensor("op_2551_cast_fp16")]; + tensor x_277_cast_fp16 = real_div(x = sub_71_cast_fp16, y = var_2551_cast_fp16)[name = tensor("x_277_cast_fp16")]; + tensor var_2553_cast_fp16 = mul(x = x_277_cast_fp16, y = flow_net_res_blocks_1_in_ln_weight_to_fp16)[name = tensor("op_2553_cast_fp16")]; + tensor x_279_cast_fp16 = add(x = var_2553_cast_fp16, y = flow_net_res_blocks_1_in_ln_bias_to_fp16)[name = tensor("x_279_cast_fp16")]; + tensor var_2555_promoted_to_fp16 = const()[name = tensor("op_2555_promoted_to_fp16"), val = tensor(0x1p+0)]; + tensor var_2556_cast_fp16 = add(x = var_2539_cast_fp16_1, y = var_2555_promoted_to_fp16)[name = tensor("op_2556_cast_fp16")]; + tensor var_2557_cast_fp16 = mul(x = x_279_cast_fp16, y = var_2556_cast_fp16)[name = tensor("op_2557_cast_fp16")]; + tensor input_433_cast_fp16 = add(x = var_2557_cast_fp16, y = var_2539_cast_fp16_0)[name = tensor("input_433_cast_fp16")]; + tensor linear_166_cast_fp16 = linear(bias = flow_net_res_blocks_1_mlp_0_bias_to_fp16, weight = flow_net_res_blocks_1_mlp_0_weight_to_fp16, x = input_433_cast_fp16)[name = tensor("linear_166_cast_fp16")]; + tensor input_437_cast_fp16 = silu(x = linear_166_cast_fp16)[name = tensor("input_437_cast_fp16")]; + tensor linear_167_cast_fp16 = linear(bias = flow_net_res_blocks_1_mlp_2_bias_to_fp16, weight = flow_net_res_blocks_1_mlp_2_weight_to_fp16, x = input_437_cast_fp16)[name = tensor("linear_167_cast_fp16")]; + tensor var_2568_cast_fp16 = mul(x = var_2539_cast_fp16_2, y = linear_167_cast_fp16)[name = tensor("op_2568_cast_fp16")]; + tensor x_281_cast_fp16 = add(x = x_275_cast_fp16, y = var_2568_cast_fp16)[name = tensor("x_281_cast_fp16")]; + tensor linear_168_cast_fp16 = linear(bias = flow_net_res_blocks_2_adaLN_modulation_1_bias_to_fp16, weight = flow_net_res_blocks_2_adaLN_modulation_1_weight_to_fp16, x = input_423_cast_fp16)[name = tensor("linear_168_cast_fp16")]; + tensor var_2578_split_sizes_0 = const()[name = tensor("op_2578_split_sizes_0"), val = tensor([512, 512, 512])]; + tensor var_2578_axis_0 = const()[name = tensor("op_2578_axis_0"), val = tensor(-1)]; + tensor var_2578_cast_fp16_0, tensor var_2578_cast_fp16_1, tensor var_2578_cast_fp16_2 = split(axis = var_2578_axis_0, split_sizes = var_2578_split_sizes_0, x = linear_168_cast_fp16)[name = tensor("op_2578_cast_fp16")]; + tensor mean_89_axes_0 = const()[name = tensor("mean_89_axes_0"), val = tensor([-1])]; + tensor mean_89_keep_dims_0 = const()[name = tensor("mean_89_keep_dims_0"), val = tensor(true)]; + tensor mean_89_cast_fp16 = reduce_mean(axes = mean_89_axes_0, keep_dims = mean_89_keep_dims_0, x = x_281_cast_fp16)[name = tensor("mean_89_cast_fp16")]; + tensor sub_72_cast_fp16 = sub(x = x_281_cast_fp16, y = mean_89_cast_fp16)[name = tensor("sub_72_cast_fp16")]; + tensor square_58_cast_fp16 = square(x = sub_72_cast_fp16)[name = tensor("square_58_cast_fp16")]; + tensor reduce_mean_117_axes_0 = const()[name = tensor("reduce_mean_117_axes_0"), val = tensor([-1])]; + tensor reduce_mean_117_keep_dims_0 = const()[name = tensor("reduce_mean_117_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_117_cast_fp16 = reduce_mean(axes = reduce_mean_117_axes_0, keep_dims = reduce_mean_117_keep_dims_0, x = square_58_cast_fp16)[name = tensor("reduce_mean_117_cast_fp16")]; + tensor var_2588_to_fp16 = const()[name = tensor("op_2588_to_fp16"), val = tensor(0x1.1p-20)]; + tensor var_2589_cast_fp16 = add(x = reduce_mean_117_cast_fp16, y = var_2588_to_fp16)[name = tensor("op_2589_cast_fp16")]; + tensor var_2590_cast_fp16 = sqrt(x = var_2589_cast_fp16)[name = tensor("op_2590_cast_fp16")]; + tensor x_283_cast_fp16 = real_div(x = sub_72_cast_fp16, y = var_2590_cast_fp16)[name = tensor("x_283_cast_fp16")]; + tensor var_2592_cast_fp16 = mul(x = x_283_cast_fp16, y = flow_net_res_blocks_2_in_ln_weight_to_fp16)[name = tensor("op_2592_cast_fp16")]; + tensor x_285_cast_fp16 = add(x = var_2592_cast_fp16, y = flow_net_res_blocks_2_in_ln_bias_to_fp16)[name = tensor("x_285_cast_fp16")]; + tensor var_2594_promoted_to_fp16 = const()[name = tensor("op_2594_promoted_to_fp16"), val = tensor(0x1p+0)]; + tensor var_2595_cast_fp16 = add(x = var_2578_cast_fp16_1, y = var_2594_promoted_to_fp16)[name = tensor("op_2595_cast_fp16")]; + tensor var_2596_cast_fp16 = mul(x = x_285_cast_fp16, y = var_2595_cast_fp16)[name = tensor("op_2596_cast_fp16")]; + tensor input_441_cast_fp16 = add(x = var_2596_cast_fp16, y = var_2578_cast_fp16_0)[name = tensor("input_441_cast_fp16")]; + tensor linear_169_cast_fp16 = linear(bias = flow_net_res_blocks_2_mlp_0_bias_to_fp16, weight = flow_net_res_blocks_2_mlp_0_weight_to_fp16, x = input_441_cast_fp16)[name = tensor("linear_169_cast_fp16")]; + tensor input_445_cast_fp16 = silu(x = linear_169_cast_fp16)[name = tensor("input_445_cast_fp16")]; + tensor linear_170_cast_fp16 = linear(bias = flow_net_res_blocks_2_mlp_2_bias_to_fp16, weight = flow_net_res_blocks_2_mlp_2_weight_to_fp16, x = input_445_cast_fp16)[name = tensor("linear_170_cast_fp16")]; + tensor var_2607_cast_fp16 = mul(x = var_2578_cast_fp16_2, y = linear_170_cast_fp16)[name = tensor("op_2607_cast_fp16")]; + tensor x_287_cast_fp16 = add(x = x_281_cast_fp16, y = var_2607_cast_fp16)[name = tensor("x_287_cast_fp16")]; + tensor linear_171_cast_fp16 = linear(bias = flow_net_res_blocks_3_adaLN_modulation_1_bias_to_fp16, weight = flow_net_res_blocks_3_adaLN_modulation_1_weight_to_fp16, x = input_423_cast_fp16)[name = tensor("linear_171_cast_fp16")]; + tensor var_2617_split_sizes_0 = const()[name = tensor("op_2617_split_sizes_0"), val = tensor([512, 512, 512])]; + tensor var_2617_axis_0 = const()[name = tensor("op_2617_axis_0"), val = tensor(-1)]; + tensor var_2617_cast_fp16_0, tensor var_2617_cast_fp16_1, tensor var_2617_cast_fp16_2 = split(axis = var_2617_axis_0, split_sizes = var_2617_split_sizes_0, x = linear_171_cast_fp16)[name = tensor("op_2617_cast_fp16")]; + tensor mean_91_axes_0 = const()[name = tensor("mean_91_axes_0"), val = tensor([-1])]; + tensor mean_91_keep_dims_0 = const()[name = tensor("mean_91_keep_dims_0"), val = tensor(true)]; + tensor mean_91_cast_fp16 = reduce_mean(axes = mean_91_axes_0, keep_dims = mean_91_keep_dims_0, x = x_287_cast_fp16)[name = tensor("mean_91_cast_fp16")]; + tensor sub_73_cast_fp16 = sub(x = x_287_cast_fp16, y = mean_91_cast_fp16)[name = tensor("sub_73_cast_fp16")]; + tensor square_59_cast_fp16 = square(x = sub_73_cast_fp16)[name = tensor("square_59_cast_fp16")]; + tensor reduce_mean_119_axes_0 = const()[name = tensor("reduce_mean_119_axes_0"), val = tensor([-1])]; + tensor reduce_mean_119_keep_dims_0 = const()[name = tensor("reduce_mean_119_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_119_cast_fp16 = reduce_mean(axes = reduce_mean_119_axes_0, keep_dims = reduce_mean_119_keep_dims_0, x = square_59_cast_fp16)[name = tensor("reduce_mean_119_cast_fp16")]; + tensor var_2627_to_fp16 = const()[name = tensor("op_2627_to_fp16"), val = tensor(0x1.1p-20)]; + tensor var_2628_cast_fp16 = add(x = reduce_mean_119_cast_fp16, y = var_2627_to_fp16)[name = tensor("op_2628_cast_fp16")]; + tensor var_2629_cast_fp16 = sqrt(x = var_2628_cast_fp16)[name = tensor("op_2629_cast_fp16")]; + tensor x_289_cast_fp16 = real_div(x = sub_73_cast_fp16, y = var_2629_cast_fp16)[name = tensor("x_289_cast_fp16")]; + tensor var_2631_cast_fp16 = mul(x = x_289_cast_fp16, y = flow_net_res_blocks_3_in_ln_weight_to_fp16)[name = tensor("op_2631_cast_fp16")]; + tensor x_291_cast_fp16 = add(x = var_2631_cast_fp16, y = flow_net_res_blocks_3_in_ln_bias_to_fp16)[name = tensor("x_291_cast_fp16")]; + tensor var_2633_promoted_to_fp16 = const()[name = tensor("op_2633_promoted_to_fp16"), val = tensor(0x1p+0)]; + tensor var_2634_cast_fp16 = add(x = var_2617_cast_fp16_1, y = var_2633_promoted_to_fp16)[name = tensor("op_2634_cast_fp16")]; + tensor var_2635_cast_fp16 = mul(x = x_291_cast_fp16, y = var_2634_cast_fp16)[name = tensor("op_2635_cast_fp16")]; + tensor input_449_cast_fp16 = add(x = var_2635_cast_fp16, y = var_2617_cast_fp16_0)[name = tensor("input_449_cast_fp16")]; + tensor linear_172_cast_fp16 = linear(bias = flow_net_res_blocks_3_mlp_0_bias_to_fp16, weight = flow_net_res_blocks_3_mlp_0_weight_to_fp16, x = input_449_cast_fp16)[name = tensor("linear_172_cast_fp16")]; + tensor input_453_cast_fp16 = silu(x = linear_172_cast_fp16)[name = tensor("input_453_cast_fp16")]; + tensor linear_173_cast_fp16 = linear(bias = flow_net_res_blocks_3_mlp_2_bias_to_fp16, weight = flow_net_res_blocks_3_mlp_2_weight_to_fp16, x = input_453_cast_fp16)[name = tensor("linear_173_cast_fp16")]; + tensor var_2646_cast_fp16 = mul(x = var_2617_cast_fp16_2, y = linear_173_cast_fp16)[name = tensor("op_2646_cast_fp16")]; + tensor x_293_cast_fp16 = add(x = x_287_cast_fp16, y = var_2646_cast_fp16)[name = tensor("x_293_cast_fp16")]; + tensor linear_174_cast_fp16 = linear(bias = flow_net_res_blocks_4_adaLN_modulation_1_bias_to_fp16, weight = flow_net_res_blocks_4_adaLN_modulation_1_weight_to_fp16, x = input_423_cast_fp16)[name = tensor("linear_174_cast_fp16")]; + tensor var_2656_split_sizes_0 = const()[name = tensor("op_2656_split_sizes_0"), val = tensor([512, 512, 512])]; + tensor var_2656_axis_0 = const()[name = tensor("op_2656_axis_0"), val = tensor(-1)]; + tensor var_2656_cast_fp16_0, tensor var_2656_cast_fp16_1, tensor var_2656_cast_fp16_2 = split(axis = var_2656_axis_0, split_sizes = var_2656_split_sizes_0, x = linear_174_cast_fp16)[name = tensor("op_2656_cast_fp16")]; + tensor mean_93_axes_0 = const()[name = tensor("mean_93_axes_0"), val = tensor([-1])]; + tensor mean_93_keep_dims_0 = const()[name = tensor("mean_93_keep_dims_0"), val = tensor(true)]; + tensor mean_93_cast_fp16 = reduce_mean(axes = mean_93_axes_0, keep_dims = mean_93_keep_dims_0, x = x_293_cast_fp16)[name = tensor("mean_93_cast_fp16")]; + tensor sub_74_cast_fp16 = sub(x = x_293_cast_fp16, y = mean_93_cast_fp16)[name = tensor("sub_74_cast_fp16")]; + tensor square_60_cast_fp16 = square(x = sub_74_cast_fp16)[name = tensor("square_60_cast_fp16")]; + tensor reduce_mean_121_axes_0 = const()[name = tensor("reduce_mean_121_axes_0"), val = tensor([-1])]; + tensor reduce_mean_121_keep_dims_0 = const()[name = tensor("reduce_mean_121_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_121_cast_fp16 = reduce_mean(axes = reduce_mean_121_axes_0, keep_dims = reduce_mean_121_keep_dims_0, x = square_60_cast_fp16)[name = tensor("reduce_mean_121_cast_fp16")]; + tensor var_2666_to_fp16 = const()[name = tensor("op_2666_to_fp16"), val = tensor(0x1.1p-20)]; + tensor var_2667_cast_fp16 = add(x = reduce_mean_121_cast_fp16, y = var_2666_to_fp16)[name = tensor("op_2667_cast_fp16")]; + tensor var_2668_cast_fp16 = sqrt(x = var_2667_cast_fp16)[name = tensor("op_2668_cast_fp16")]; + tensor x_295_cast_fp16 = real_div(x = sub_74_cast_fp16, y = var_2668_cast_fp16)[name = tensor("x_295_cast_fp16")]; + tensor var_2670_cast_fp16 = mul(x = x_295_cast_fp16, y = flow_net_res_blocks_4_in_ln_weight_to_fp16)[name = tensor("op_2670_cast_fp16")]; + tensor x_297_cast_fp16 = add(x = var_2670_cast_fp16, y = flow_net_res_blocks_4_in_ln_bias_to_fp16)[name = tensor("x_297_cast_fp16")]; + tensor var_2672_promoted_to_fp16 = const()[name = tensor("op_2672_promoted_to_fp16"), val = tensor(0x1p+0)]; + tensor var_2673_cast_fp16 = add(x = var_2656_cast_fp16_1, y = var_2672_promoted_to_fp16)[name = tensor("op_2673_cast_fp16")]; + tensor var_2674_cast_fp16 = mul(x = x_297_cast_fp16, y = var_2673_cast_fp16)[name = tensor("op_2674_cast_fp16")]; + tensor input_457_cast_fp16 = add(x = var_2674_cast_fp16, y = var_2656_cast_fp16_0)[name = tensor("input_457_cast_fp16")]; + tensor linear_175_cast_fp16 = linear(bias = flow_net_res_blocks_4_mlp_0_bias_to_fp16, weight = flow_net_res_blocks_4_mlp_0_weight_to_fp16, x = input_457_cast_fp16)[name = tensor("linear_175_cast_fp16")]; + tensor input_461_cast_fp16 = silu(x = linear_175_cast_fp16)[name = tensor("input_461_cast_fp16")]; + tensor linear_176_cast_fp16 = linear(bias = flow_net_res_blocks_4_mlp_2_bias_to_fp16, weight = flow_net_res_blocks_4_mlp_2_weight_to_fp16, x = input_461_cast_fp16)[name = tensor("linear_176_cast_fp16")]; + tensor var_2685_cast_fp16 = mul(x = var_2656_cast_fp16_2, y = linear_176_cast_fp16)[name = tensor("op_2685_cast_fp16")]; + tensor x_299_cast_fp16 = add(x = x_293_cast_fp16, y = var_2685_cast_fp16)[name = tensor("x_299_cast_fp16")]; + tensor linear_177_cast_fp16 = linear(bias = flow_net_res_blocks_5_adaLN_modulation_1_bias_to_fp16, weight = flow_net_res_blocks_5_adaLN_modulation_1_weight_to_fp16, x = input_423_cast_fp16)[name = tensor("linear_177_cast_fp16")]; + tensor var_2695_split_sizes_0 = const()[name = tensor("op_2695_split_sizes_0"), val = tensor([512, 512, 512])]; + tensor var_2695_axis_0 = const()[name = tensor("op_2695_axis_0"), val = tensor(-1)]; + tensor var_2695_cast_fp16_0, tensor var_2695_cast_fp16_1, tensor var_2695_cast_fp16_2 = split(axis = var_2695_axis_0, split_sizes = var_2695_split_sizes_0, x = linear_177_cast_fp16)[name = tensor("op_2695_cast_fp16")]; + tensor mean_95_axes_0 = const()[name = tensor("mean_95_axes_0"), val = tensor([-1])]; + tensor mean_95_keep_dims_0 = const()[name = tensor("mean_95_keep_dims_0"), val = tensor(true)]; + tensor mean_95_cast_fp16 = reduce_mean(axes = mean_95_axes_0, keep_dims = mean_95_keep_dims_0, x = x_299_cast_fp16)[name = tensor("mean_95_cast_fp16")]; + tensor sub_75_cast_fp16 = sub(x = x_299_cast_fp16, y = mean_95_cast_fp16)[name = tensor("sub_75_cast_fp16")]; + tensor square_61_cast_fp16 = square(x = sub_75_cast_fp16)[name = tensor("square_61_cast_fp16")]; + tensor reduce_mean_123_axes_0 = const()[name = tensor("reduce_mean_123_axes_0"), val = tensor([-1])]; + tensor reduce_mean_123_keep_dims_0 = const()[name = tensor("reduce_mean_123_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_123_cast_fp16 = reduce_mean(axes = reduce_mean_123_axes_0, keep_dims = reduce_mean_123_keep_dims_0, x = square_61_cast_fp16)[name = tensor("reduce_mean_123_cast_fp16")]; + tensor var_2705_to_fp16 = const()[name = tensor("op_2705_to_fp16"), val = tensor(0x1.1p-20)]; + tensor var_2706_cast_fp16 = add(x = reduce_mean_123_cast_fp16, y = var_2705_to_fp16)[name = tensor("op_2706_cast_fp16")]; + tensor var_2707_cast_fp16 = sqrt(x = var_2706_cast_fp16)[name = tensor("op_2707_cast_fp16")]; + tensor x_301_cast_fp16 = real_div(x = sub_75_cast_fp16, y = var_2707_cast_fp16)[name = tensor("x_301_cast_fp16")]; + tensor var_2709_cast_fp16 = mul(x = x_301_cast_fp16, y = flow_net_res_blocks_5_in_ln_weight_to_fp16)[name = tensor("op_2709_cast_fp16")]; + tensor x_303_cast_fp16 = add(x = var_2709_cast_fp16, y = flow_net_res_blocks_5_in_ln_bias_to_fp16)[name = tensor("x_303_cast_fp16")]; + tensor var_2711_promoted_to_fp16 = const()[name = tensor("op_2711_promoted_to_fp16"), val = tensor(0x1p+0)]; + tensor var_2712_cast_fp16 = add(x = var_2695_cast_fp16_1, y = var_2711_promoted_to_fp16)[name = tensor("op_2712_cast_fp16")]; + tensor var_2713_cast_fp16 = mul(x = x_303_cast_fp16, y = var_2712_cast_fp16)[name = tensor("op_2713_cast_fp16")]; + tensor input_465_cast_fp16 = add(x = var_2713_cast_fp16, y = var_2695_cast_fp16_0)[name = tensor("input_465_cast_fp16")]; + tensor linear_178_cast_fp16 = linear(bias = flow_net_res_blocks_5_mlp_0_bias_to_fp16, weight = flow_net_res_blocks_5_mlp_0_weight_to_fp16, x = input_465_cast_fp16)[name = tensor("linear_178_cast_fp16")]; + tensor input_469_cast_fp16 = silu(x = linear_178_cast_fp16)[name = tensor("input_469_cast_fp16")]; + tensor linear_179_cast_fp16 = linear(bias = flow_net_res_blocks_5_mlp_2_bias_to_fp16, weight = flow_net_res_blocks_5_mlp_2_weight_to_fp16, x = input_469_cast_fp16)[name = tensor("linear_179_cast_fp16")]; + tensor var_2724_cast_fp16 = mul(x = var_2695_cast_fp16_2, y = linear_179_cast_fp16)[name = tensor("op_2724_cast_fp16")]; + tensor x_305_cast_fp16 = add(x = x_299_cast_fp16, y = var_2724_cast_fp16)[name = tensor("x_305_cast_fp16")]; + tensor linear_180_cast_fp16 = linear(bias = flow_net_final_layer_adaLN_modulation_1_bias_to_fp16, weight = flow_net_final_layer_adaLN_modulation_1_weight_to_fp16, x = input_423_cast_fp16)[name = tensor("linear_180_cast_fp16")]; + tensor var_2733_split_sizes_0 = const()[name = tensor("op_2733_split_sizes_0"), val = tensor([512, 512])]; + tensor var_2733_axis_0 = const()[name = tensor("op_2733_axis_0"), val = tensor(-1)]; + tensor var_2733_cast_fp16_0, tensor var_2733_cast_fp16_1 = split(axis = var_2733_axis_0, split_sizes = var_2733_split_sizes_0, x = linear_180_cast_fp16)[name = tensor("op_2733_cast_fp16")]; + tensor mean_97_axes_0 = const()[name = tensor("mean_97_axes_0"), val = tensor([-1])]; + tensor mean_97_keep_dims_0 = const()[name = tensor("mean_97_keep_dims_0"), val = tensor(true)]; + tensor mean_97_cast_fp16 = reduce_mean(axes = mean_97_axes_0, keep_dims = mean_97_keep_dims_0, x = x_305_cast_fp16)[name = tensor("mean_97_cast_fp16")]; + tensor sub_76_cast_fp16 = sub(x = x_305_cast_fp16, y = mean_97_cast_fp16)[name = tensor("sub_76_cast_fp16")]; + tensor square_62_cast_fp16 = square(x = sub_76_cast_fp16)[name = tensor("square_62_cast_fp16")]; + tensor reduce_mean_125_axes_0 = const()[name = tensor("reduce_mean_125_axes_0"), val = tensor([-1])]; + tensor reduce_mean_125_keep_dims_0 = const()[name = tensor("reduce_mean_125_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_125_cast_fp16 = reduce_mean(axes = reduce_mean_125_axes_0, keep_dims = reduce_mean_125_keep_dims_0, x = square_62_cast_fp16)[name = tensor("reduce_mean_125_cast_fp16")]; + tensor var_2740_to_fp16 = const()[name = tensor("op_2740_to_fp16"), val = tensor(0x1.1p-20)]; + tensor var_2741_cast_fp16 = add(x = reduce_mean_125_cast_fp16, y = var_2740_to_fp16)[name = tensor("op_2741_cast_fp16")]; + tensor var_2742_cast_fp16 = sqrt(x = var_2741_cast_fp16)[name = tensor("op_2742_cast_fp16")]; + tensor x_307_cast_fp16 = real_div(x = sub_76_cast_fp16, y = var_2742_cast_fp16)[name = tensor("x_307_cast_fp16")]; + tensor var_2744_promoted_to_fp16 = const()[name = tensor("op_2744_promoted_to_fp16"), val = tensor(0x1p+0)]; + tensor var_2745_cast_fp16 = add(x = var_2733_cast_fp16_1, y = var_2744_promoted_to_fp16)[name = tensor("op_2745_cast_fp16")]; + tensor var_2746_cast_fp16 = mul(x = x_307_cast_fp16, y = var_2745_cast_fp16)[name = tensor("op_2746_cast_fp16")]; + tensor input_473_cast_fp16 = add(x = var_2746_cast_fp16, y = var_2733_cast_fp16_0)[name = tensor("input_473_cast_fp16")]; + tensor linear_181_cast_fp16 = linear(bias = flow_net_final_layer_linear_bias_to_fp16, weight = flow_net_final_layer_linear_weight_to_fp16, x = input_473_cast_fp16)[name = tensor("linear_181_cast_fp16")]; + tensor var_2757_to_fp16 = const()[name = tensor("op_2757_to_fp16"), val = tensor(0x1p-3)]; + tensor var_2758_cast_fp16 = mul(x = linear_181_cast_fp16, y = var_2757_to_fp16)[name = tensor("op_2758_cast_fp16")]; + tensor input_475_cast_fp16 = add(x = input_407_cast_fp16, y = var_2758_cast_fp16)[name = tensor("input_475_cast_fp16")]; + tensor linear_182_cast_fp16 = linear(bias = flow_net_input_proj_bias_to_fp16, weight = flow_net_input_proj_weight_to_fp16, x = input_475_cast_fp16)[name = tensor("linear_182_cast_fp16")]; + tensor input_479_to_fp16 = const()[name = tensor("input_479_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(19010432)))]; + tensor input_481_cast_fp16 = silu(x = input_479_to_fp16)[name = tensor("input_481_cast_fp16")]; + tensor linear_184_cast_fp16 = linear(bias = flow_net_time_embed_0_mlp_2_bias_to_fp16, weight = flow_net_time_embed_0_mlp_2_weight_to_fp16, x = input_481_cast_fp16)[name = tensor("linear_184_cast_fp16")]; + tensor reduce_mean_126_axes_0 = const()[name = tensor("reduce_mean_126_axes_0"), val = tensor([-1])]; + tensor reduce_mean_126_keep_dims_0 = const()[name = tensor("reduce_mean_126_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_126_cast_fp16 = reduce_mean(axes = reduce_mean_126_axes_0, keep_dims = reduce_mean_126_keep_dims_0, x = linear_184_cast_fp16)[name = tensor("reduce_mean_126_cast_fp16")]; + tensor sub_77_cast_fp16 = sub(x = linear_184_cast_fp16, y = reduce_mean_126_cast_fp16)[name = tensor("sub_77_cast_fp16")]; + tensor square_63_cast_fp16 = square(x = sub_77_cast_fp16)[name = tensor("square_63_cast_fp16")]; + tensor reduce_mean_127_axes_0 = const()[name = tensor("reduce_mean_127_axes_0"), val = tensor([-1])]; + tensor reduce_mean_127_keep_dims_0 = const()[name = tensor("reduce_mean_127_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_127_cast_fp16 = reduce_mean(axes = reduce_mean_127_axes_0, keep_dims = reduce_mean_127_keep_dims_0, x = square_63_cast_fp16)[name = tensor("reduce_mean_127_cast_fp16")]; + tensor real_div_14_to_fp16 = const()[name = tensor("real_div_14_to_fp16"), val = tensor(0x1.008p+0)]; + tensor mul_14_cast_fp16 = mul(x = reduce_mean_127_cast_fp16, y = real_div_14_to_fp16)[name = tensor("mul_14_cast_fp16")]; + tensor var_2824_to_fp16 = const()[name = tensor("op_2824_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_127_cast_fp16_0 = add(x = mul_14_cast_fp16, y = var_2824_to_fp16)[name = tensor("var_127_cast_fp16")]; + tensor var_2827_epsilon_0 = const()[name = tensor("op_2827_epsilon_0"), val = tensor(0x1.197998p-40)]; + tensor var_2827_cast_fp16 = rsqrt(epsilon = var_2827_epsilon_0, x = var_127_cast_fp16_0)[name = tensor("op_2827_cast_fp16")]; + tensor var_2828_cast_fp16 = mul(x = const_3_to_fp16, y = var_2827_cast_fp16)[name = tensor("op_2828_cast_fp16")]; + tensor var_2829_cast_fp16 = mul(x = linear_184_cast_fp16, y = var_2828_cast_fp16)[name = tensor("op_2829_cast_fp16")]; + tensor input_485_to_fp16 = const()[name = tensor("input_485_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(19011520)))]; + tensor input_487_cast_fp16 = silu(x = input_485_to_fp16)[name = tensor("input_487_cast_fp16")]; + tensor linear_186_cast_fp16 = linear(bias = flow_net_time_embed_1_mlp_2_bias_to_fp16, weight = flow_net_time_embed_1_mlp_2_weight_to_fp16, x = input_487_cast_fp16)[name = tensor("linear_186_cast_fp16")]; + tensor reduce_mean_128_axes_0 = const()[name = tensor("reduce_mean_128_axes_0"), val = tensor([-1])]; + tensor reduce_mean_128_keep_dims_0 = const()[name = tensor("reduce_mean_128_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_128_cast_fp16 = reduce_mean(axes = reduce_mean_128_axes_0, keep_dims = reduce_mean_128_keep_dims_0, x = linear_186_cast_fp16)[name = tensor("reduce_mean_128_cast_fp16")]; + tensor sub_79_cast_fp16 = sub(x = linear_186_cast_fp16, y = reduce_mean_128_cast_fp16)[name = tensor("sub_79_cast_fp16")]; + tensor square_64_cast_fp16 = square(x = sub_79_cast_fp16)[name = tensor("square_64_cast_fp16")]; + tensor reduce_mean_129_axes_0 = const()[name = tensor("reduce_mean_129_axes_0"), val = tensor([-1])]; + tensor reduce_mean_129_keep_dims_0 = const()[name = tensor("reduce_mean_129_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_129_cast_fp16 = reduce_mean(axes = reduce_mean_129_axes_0, keep_dims = reduce_mean_129_keep_dims_0, x = square_64_cast_fp16)[name = tensor("reduce_mean_129_cast_fp16")]; + tensor real_div_15_to_fp16 = const()[name = tensor("real_div_15_to_fp16"), val = tensor(0x1.008p+0)]; + tensor mul_15_cast_fp16 = mul(x = reduce_mean_129_cast_fp16, y = real_div_15_to_fp16)[name = tensor("mul_15_cast_fp16")]; + tensor var_2851_to_fp16 = const()[name = tensor("op_2851_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_129_cast_fp16 = add(x = mul_15_cast_fp16, y = var_2851_to_fp16)[name = tensor("var_129_cast_fp16")]; + tensor var_2854_epsilon_0 = const()[name = tensor("op_2854_epsilon_0"), val = tensor(0x1.197998p-40)]; + tensor var_2854_cast_fp16 = rsqrt(epsilon = var_2854_epsilon_0, x = var_129_cast_fp16)[name = tensor("op_2854_cast_fp16")]; + tensor var_2855_cast_fp16 = mul(x = const_5_to_fp16, y = var_2854_cast_fp16)[name = tensor("op_2855_cast_fp16")]; + tensor var_2856_cast_fp16 = mul(x = linear_186_cast_fp16, y = var_2855_cast_fp16)[name = tensor("op_2856_cast_fp16")]; + tensor var_2858_cast_fp16 = add(x = var_2829_cast_fp16, y = var_2856_cast_fp16)[name = tensor("op_2858_cast_fp16")]; + tensor _inversed_t_combined_y_0_to_fp16 = const()[name = tensor("_inversed_t_combined_y_0_to_fp16"), val = tensor(0x1p-1)]; + tensor _inversed_t_combined_cast_fp16 = mul(x = var_2858_cast_fp16, y = _inversed_t_combined_y_0_to_fp16)[name = tensor("_inversed_t_combined_cast_fp16")]; + tensor input_489_cast_fp16 = add(x = _inversed_t_combined_cast_fp16, y = linear_5_cast_fp16)[name = tensor("input_489_cast_fp16")]; + tensor input_491_cast_fp16 = silu(x = input_489_cast_fp16)[name = tensor("input_491_cast_fp16")]; + tensor linear_188_cast_fp16 = linear(bias = flow_net_res_blocks_0_adaLN_modulation_1_bias_to_fp16, weight = flow_net_res_blocks_0_adaLN_modulation_1_weight_to_fp16, x = input_491_cast_fp16)[name = tensor("linear_188_cast_fp16")]; + tensor var_2873_split_sizes_0 = const()[name = tensor("op_2873_split_sizes_0"), val = tensor([512, 512, 512])]; + tensor var_2873_axis_0 = const()[name = tensor("op_2873_axis_0"), val = tensor(-1)]; + tensor var_2873_cast_fp16_0, tensor var_2873_cast_fp16_1, tensor var_2873_cast_fp16_2 = split(axis = var_2873_axis_0, split_sizes = var_2873_split_sizes_0, x = linear_188_cast_fp16)[name = tensor("op_2873_cast_fp16")]; + tensor mean_99_axes_0 = const()[name = tensor("mean_99_axes_0"), val = tensor([-1])]; + tensor mean_99_keep_dims_0 = const()[name = tensor("mean_99_keep_dims_0"), val = tensor(true)]; + tensor mean_99_cast_fp16 = reduce_mean(axes = mean_99_axes_0, keep_dims = mean_99_keep_dims_0, x = linear_182_cast_fp16)[name = tensor("mean_99_cast_fp16")]; + tensor sub_81_cast_fp16 = sub(x = linear_182_cast_fp16, y = mean_99_cast_fp16)[name = tensor("sub_81_cast_fp16")]; + tensor square_65_cast_fp16 = square(x = sub_81_cast_fp16)[name = tensor("square_65_cast_fp16")]; + tensor reduce_mean_131_axes_0 = const()[name = tensor("reduce_mean_131_axes_0"), val = tensor([-1])]; + tensor reduce_mean_131_keep_dims_0 = const()[name = tensor("reduce_mean_131_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_131_cast_fp16 = reduce_mean(axes = reduce_mean_131_axes_0, keep_dims = reduce_mean_131_keep_dims_0, x = square_65_cast_fp16)[name = tensor("reduce_mean_131_cast_fp16")]; + tensor var_2883_to_fp16 = const()[name = tensor("op_2883_to_fp16"), val = tensor(0x1.1p-20)]; + tensor var_2884_cast_fp16 = add(x = reduce_mean_131_cast_fp16, y = var_2883_to_fp16)[name = tensor("op_2884_cast_fp16")]; + tensor var_2885_cast_fp16 = sqrt(x = var_2884_cast_fp16)[name = tensor("op_2885_cast_fp16")]; + tensor x_315_cast_fp16 = real_div(x = sub_81_cast_fp16, y = var_2885_cast_fp16)[name = tensor("x_315_cast_fp16")]; + tensor var_2887_cast_fp16 = mul(x = x_315_cast_fp16, y = flow_net_res_blocks_0_in_ln_weight_to_fp16)[name = tensor("op_2887_cast_fp16")]; + tensor x_317_cast_fp16 = add(x = var_2887_cast_fp16, y = flow_net_res_blocks_0_in_ln_bias_to_fp16)[name = tensor("x_317_cast_fp16")]; + tensor var_2889_promoted_to_fp16 = const()[name = tensor("op_2889_promoted_to_fp16"), val = tensor(0x1p+0)]; + tensor var_2890_cast_fp16 = add(x = var_2873_cast_fp16_1, y = var_2889_promoted_to_fp16)[name = tensor("op_2890_cast_fp16")]; + tensor var_2891_cast_fp16 = mul(x = x_317_cast_fp16, y = var_2890_cast_fp16)[name = tensor("op_2891_cast_fp16")]; + tensor input_493_cast_fp16 = add(x = var_2891_cast_fp16, y = var_2873_cast_fp16_0)[name = tensor("input_493_cast_fp16")]; + tensor linear_189_cast_fp16 = linear(bias = flow_net_res_blocks_0_mlp_0_bias_to_fp16, weight = flow_net_res_blocks_0_mlp_0_weight_to_fp16, x = input_493_cast_fp16)[name = tensor("linear_189_cast_fp16")]; + tensor input_497_cast_fp16 = silu(x = linear_189_cast_fp16)[name = tensor("input_497_cast_fp16")]; + tensor linear_190_cast_fp16 = linear(bias = flow_net_res_blocks_0_mlp_2_bias_to_fp16, weight = flow_net_res_blocks_0_mlp_2_weight_to_fp16, x = input_497_cast_fp16)[name = tensor("linear_190_cast_fp16")]; + tensor var_2902_cast_fp16 = mul(x = var_2873_cast_fp16_2, y = linear_190_cast_fp16)[name = tensor("op_2902_cast_fp16")]; + tensor x_319_cast_fp16 = add(x = linear_182_cast_fp16, y = var_2902_cast_fp16)[name = tensor("x_319_cast_fp16")]; + tensor linear_191_cast_fp16 = linear(bias = flow_net_res_blocks_1_adaLN_modulation_1_bias_to_fp16, weight = flow_net_res_blocks_1_adaLN_modulation_1_weight_to_fp16, x = input_491_cast_fp16)[name = tensor("linear_191_cast_fp16")]; + tensor var_2912_split_sizes_0 = const()[name = tensor("op_2912_split_sizes_0"), val = tensor([512, 512, 512])]; + tensor var_2912_axis_0 = const()[name = tensor("op_2912_axis_0"), val = tensor(-1)]; + tensor var_2912_cast_fp16_0, tensor var_2912_cast_fp16_1, tensor var_2912_cast_fp16_2 = split(axis = var_2912_axis_0, split_sizes = var_2912_split_sizes_0, x = linear_191_cast_fp16)[name = tensor("op_2912_cast_fp16")]; + tensor mean_101_axes_0 = const()[name = tensor("mean_101_axes_0"), val = tensor([-1])]; + tensor mean_101_keep_dims_0 = const()[name = tensor("mean_101_keep_dims_0"), val = tensor(true)]; + tensor mean_101_cast_fp16 = reduce_mean(axes = mean_101_axes_0, keep_dims = mean_101_keep_dims_0, x = x_319_cast_fp16)[name = tensor("mean_101_cast_fp16")]; + tensor sub_82_cast_fp16 = sub(x = x_319_cast_fp16, y = mean_101_cast_fp16)[name = tensor("sub_82_cast_fp16")]; + tensor square_66_cast_fp16 = square(x = sub_82_cast_fp16)[name = tensor("square_66_cast_fp16")]; + tensor reduce_mean_133_axes_0 = const()[name = tensor("reduce_mean_133_axes_0"), val = tensor([-1])]; + tensor reduce_mean_133_keep_dims_0 = const()[name = tensor("reduce_mean_133_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_133_cast_fp16 = reduce_mean(axes = reduce_mean_133_axes_0, keep_dims = reduce_mean_133_keep_dims_0, x = square_66_cast_fp16)[name = tensor("reduce_mean_133_cast_fp16")]; + tensor var_2922_to_fp16 = const()[name = tensor("op_2922_to_fp16"), val = tensor(0x1.1p-20)]; + tensor var_2923_cast_fp16 = add(x = reduce_mean_133_cast_fp16, y = var_2922_to_fp16)[name = tensor("op_2923_cast_fp16")]; + tensor var_2924_cast_fp16 = sqrt(x = var_2923_cast_fp16)[name = tensor("op_2924_cast_fp16")]; + tensor x_321_cast_fp16 = real_div(x = sub_82_cast_fp16, y = var_2924_cast_fp16)[name = tensor("x_321_cast_fp16")]; + tensor var_2926_cast_fp16 = mul(x = x_321_cast_fp16, y = flow_net_res_blocks_1_in_ln_weight_to_fp16)[name = tensor("op_2926_cast_fp16")]; + tensor x_323_cast_fp16 = add(x = var_2926_cast_fp16, y = flow_net_res_blocks_1_in_ln_bias_to_fp16)[name = tensor("x_323_cast_fp16")]; + tensor var_2928_promoted_to_fp16 = const()[name = tensor("op_2928_promoted_to_fp16"), val = tensor(0x1p+0)]; + tensor var_2929_cast_fp16 = add(x = var_2912_cast_fp16_1, y = var_2928_promoted_to_fp16)[name = tensor("op_2929_cast_fp16")]; + tensor var_2930_cast_fp16 = mul(x = x_323_cast_fp16, y = var_2929_cast_fp16)[name = tensor("op_2930_cast_fp16")]; + tensor input_501_cast_fp16 = add(x = var_2930_cast_fp16, y = var_2912_cast_fp16_0)[name = tensor("input_501_cast_fp16")]; + tensor linear_192_cast_fp16 = linear(bias = flow_net_res_blocks_1_mlp_0_bias_to_fp16, weight = flow_net_res_blocks_1_mlp_0_weight_to_fp16, x = input_501_cast_fp16)[name = tensor("linear_192_cast_fp16")]; + tensor input_505_cast_fp16 = silu(x = linear_192_cast_fp16)[name = tensor("input_505_cast_fp16")]; + tensor linear_193_cast_fp16 = linear(bias = flow_net_res_blocks_1_mlp_2_bias_to_fp16, weight = flow_net_res_blocks_1_mlp_2_weight_to_fp16, x = input_505_cast_fp16)[name = tensor("linear_193_cast_fp16")]; + tensor var_2941_cast_fp16 = mul(x = var_2912_cast_fp16_2, y = linear_193_cast_fp16)[name = tensor("op_2941_cast_fp16")]; + tensor x_325_cast_fp16 = add(x = x_319_cast_fp16, y = var_2941_cast_fp16)[name = tensor("x_325_cast_fp16")]; + tensor linear_194_cast_fp16 = linear(bias = flow_net_res_blocks_2_adaLN_modulation_1_bias_to_fp16, weight = flow_net_res_blocks_2_adaLN_modulation_1_weight_to_fp16, x = input_491_cast_fp16)[name = tensor("linear_194_cast_fp16")]; + tensor var_2951_split_sizes_0 = const()[name = tensor("op_2951_split_sizes_0"), val = tensor([512, 512, 512])]; + tensor var_2951_axis_0 = const()[name = tensor("op_2951_axis_0"), val = tensor(-1)]; + tensor var_2951_cast_fp16_0, tensor var_2951_cast_fp16_1, tensor var_2951_cast_fp16_2 = split(axis = var_2951_axis_0, split_sizes = var_2951_split_sizes_0, x = linear_194_cast_fp16)[name = tensor("op_2951_cast_fp16")]; + tensor mean_103_axes_0 = const()[name = tensor("mean_103_axes_0"), val = tensor([-1])]; + tensor mean_103_keep_dims_0 = const()[name = tensor("mean_103_keep_dims_0"), val = tensor(true)]; + tensor mean_103_cast_fp16 = reduce_mean(axes = mean_103_axes_0, keep_dims = mean_103_keep_dims_0, x = x_325_cast_fp16)[name = tensor("mean_103_cast_fp16")]; + tensor sub_83_cast_fp16 = sub(x = x_325_cast_fp16, y = mean_103_cast_fp16)[name = tensor("sub_83_cast_fp16")]; + tensor square_67_cast_fp16 = square(x = sub_83_cast_fp16)[name = tensor("square_67_cast_fp16")]; + tensor reduce_mean_135_axes_0 = const()[name = tensor("reduce_mean_135_axes_0"), val = tensor([-1])]; + tensor reduce_mean_135_keep_dims_0 = const()[name = tensor("reduce_mean_135_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_135_cast_fp16 = reduce_mean(axes = reduce_mean_135_axes_0, keep_dims = reduce_mean_135_keep_dims_0, x = square_67_cast_fp16)[name = tensor("reduce_mean_135_cast_fp16")]; + tensor var_2961_to_fp16 = const()[name = tensor("op_2961_to_fp16"), val = tensor(0x1.1p-20)]; + tensor var_2962_cast_fp16 = add(x = reduce_mean_135_cast_fp16, y = var_2961_to_fp16)[name = tensor("op_2962_cast_fp16")]; + tensor var_2963_cast_fp16 = sqrt(x = var_2962_cast_fp16)[name = tensor("op_2963_cast_fp16")]; + tensor x_327_cast_fp16 = real_div(x = sub_83_cast_fp16, y = var_2963_cast_fp16)[name = tensor("x_327_cast_fp16")]; + tensor var_2965_cast_fp16 = mul(x = x_327_cast_fp16, y = flow_net_res_blocks_2_in_ln_weight_to_fp16)[name = tensor("op_2965_cast_fp16")]; + tensor x_329_cast_fp16 = add(x = var_2965_cast_fp16, y = flow_net_res_blocks_2_in_ln_bias_to_fp16)[name = tensor("x_329_cast_fp16")]; + tensor var_2967_promoted_to_fp16 = const()[name = tensor("op_2967_promoted_to_fp16"), val = tensor(0x1p+0)]; + tensor var_2968_cast_fp16 = add(x = var_2951_cast_fp16_1, y = var_2967_promoted_to_fp16)[name = tensor("op_2968_cast_fp16")]; + tensor var_2969_cast_fp16 = mul(x = x_329_cast_fp16, y = var_2968_cast_fp16)[name = tensor("op_2969_cast_fp16")]; + tensor input_509_cast_fp16 = add(x = var_2969_cast_fp16, y = var_2951_cast_fp16_0)[name = tensor("input_509_cast_fp16")]; + tensor linear_195_cast_fp16 = linear(bias = flow_net_res_blocks_2_mlp_0_bias_to_fp16, weight = flow_net_res_blocks_2_mlp_0_weight_to_fp16, x = input_509_cast_fp16)[name = tensor("linear_195_cast_fp16")]; + tensor input_513_cast_fp16 = silu(x = linear_195_cast_fp16)[name = tensor("input_513_cast_fp16")]; + tensor linear_196_cast_fp16 = linear(bias = flow_net_res_blocks_2_mlp_2_bias_to_fp16, weight = flow_net_res_blocks_2_mlp_2_weight_to_fp16, x = input_513_cast_fp16)[name = tensor("linear_196_cast_fp16")]; + tensor var_2980_cast_fp16 = mul(x = var_2951_cast_fp16_2, y = linear_196_cast_fp16)[name = tensor("op_2980_cast_fp16")]; + tensor x_331_cast_fp16 = add(x = x_325_cast_fp16, y = var_2980_cast_fp16)[name = tensor("x_331_cast_fp16")]; + tensor linear_197_cast_fp16 = linear(bias = flow_net_res_blocks_3_adaLN_modulation_1_bias_to_fp16, weight = flow_net_res_blocks_3_adaLN_modulation_1_weight_to_fp16, x = input_491_cast_fp16)[name = tensor("linear_197_cast_fp16")]; + tensor var_2990_split_sizes_0 = const()[name = tensor("op_2990_split_sizes_0"), val = tensor([512, 512, 512])]; + tensor var_2990_axis_0 = const()[name = tensor("op_2990_axis_0"), val = tensor(-1)]; + tensor var_2990_cast_fp16_0, tensor var_2990_cast_fp16_1, tensor var_2990_cast_fp16_2 = split(axis = var_2990_axis_0, split_sizes = var_2990_split_sizes_0, x = linear_197_cast_fp16)[name = tensor("op_2990_cast_fp16")]; + tensor mean_105_axes_0 = const()[name = tensor("mean_105_axes_0"), val = tensor([-1])]; + tensor mean_105_keep_dims_0 = const()[name = tensor("mean_105_keep_dims_0"), val = tensor(true)]; + tensor mean_105_cast_fp16 = reduce_mean(axes = mean_105_axes_0, keep_dims = mean_105_keep_dims_0, x = x_331_cast_fp16)[name = tensor("mean_105_cast_fp16")]; + tensor sub_84_cast_fp16 = sub(x = x_331_cast_fp16, y = mean_105_cast_fp16)[name = tensor("sub_84_cast_fp16")]; + tensor square_68_cast_fp16 = square(x = sub_84_cast_fp16)[name = tensor("square_68_cast_fp16")]; + tensor reduce_mean_137_axes_0 = const()[name = tensor("reduce_mean_137_axes_0"), val = tensor([-1])]; + tensor reduce_mean_137_keep_dims_0 = const()[name = tensor("reduce_mean_137_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_137_cast_fp16 = reduce_mean(axes = reduce_mean_137_axes_0, keep_dims = reduce_mean_137_keep_dims_0, x = square_68_cast_fp16)[name = tensor("reduce_mean_137_cast_fp16")]; + tensor var_3000_to_fp16 = const()[name = tensor("op_3000_to_fp16"), val = tensor(0x1.1p-20)]; + tensor var_3001_cast_fp16 = add(x = reduce_mean_137_cast_fp16, y = var_3000_to_fp16)[name = tensor("op_3001_cast_fp16")]; + tensor var_3002_cast_fp16 = sqrt(x = var_3001_cast_fp16)[name = tensor("op_3002_cast_fp16")]; + tensor x_333_cast_fp16 = real_div(x = sub_84_cast_fp16, y = var_3002_cast_fp16)[name = tensor("x_333_cast_fp16")]; + tensor var_3004_cast_fp16 = mul(x = x_333_cast_fp16, y = flow_net_res_blocks_3_in_ln_weight_to_fp16)[name = tensor("op_3004_cast_fp16")]; + tensor x_335_cast_fp16 = add(x = var_3004_cast_fp16, y = flow_net_res_blocks_3_in_ln_bias_to_fp16)[name = tensor("x_335_cast_fp16")]; + tensor var_3006_promoted_to_fp16 = const()[name = tensor("op_3006_promoted_to_fp16"), val = tensor(0x1p+0)]; + tensor var_3007_cast_fp16 = add(x = var_2990_cast_fp16_1, y = var_3006_promoted_to_fp16)[name = tensor("op_3007_cast_fp16")]; + tensor var_3008_cast_fp16 = mul(x = x_335_cast_fp16, y = var_3007_cast_fp16)[name = tensor("op_3008_cast_fp16")]; + tensor input_517_cast_fp16 = add(x = var_3008_cast_fp16, y = var_2990_cast_fp16_0)[name = tensor("input_517_cast_fp16")]; + tensor linear_198_cast_fp16 = linear(bias = flow_net_res_blocks_3_mlp_0_bias_to_fp16, weight = flow_net_res_blocks_3_mlp_0_weight_to_fp16, x = input_517_cast_fp16)[name = tensor("linear_198_cast_fp16")]; + tensor input_521_cast_fp16 = silu(x = linear_198_cast_fp16)[name = tensor("input_521_cast_fp16")]; + tensor linear_199_cast_fp16 = linear(bias = flow_net_res_blocks_3_mlp_2_bias_to_fp16, weight = flow_net_res_blocks_3_mlp_2_weight_to_fp16, x = input_521_cast_fp16)[name = tensor("linear_199_cast_fp16")]; + tensor var_3019_cast_fp16 = mul(x = var_2990_cast_fp16_2, y = linear_199_cast_fp16)[name = tensor("op_3019_cast_fp16")]; + tensor x_337_cast_fp16 = add(x = x_331_cast_fp16, y = var_3019_cast_fp16)[name = tensor("x_337_cast_fp16")]; + tensor linear_200_cast_fp16 = linear(bias = flow_net_res_blocks_4_adaLN_modulation_1_bias_to_fp16, weight = flow_net_res_blocks_4_adaLN_modulation_1_weight_to_fp16, x = input_491_cast_fp16)[name = tensor("linear_200_cast_fp16")]; + tensor var_3029_split_sizes_0 = const()[name = tensor("op_3029_split_sizes_0"), val = tensor([512, 512, 512])]; + tensor var_3029_axis_0 = const()[name = tensor("op_3029_axis_0"), val = tensor(-1)]; + tensor var_3029_cast_fp16_0, tensor var_3029_cast_fp16_1, tensor var_3029_cast_fp16_2 = split(axis = var_3029_axis_0, split_sizes = var_3029_split_sizes_0, x = linear_200_cast_fp16)[name = tensor("op_3029_cast_fp16")]; + tensor mean_107_axes_0 = const()[name = tensor("mean_107_axes_0"), val = tensor([-1])]; + tensor mean_107_keep_dims_0 = const()[name = tensor("mean_107_keep_dims_0"), val = tensor(true)]; + tensor mean_107_cast_fp16 = reduce_mean(axes = mean_107_axes_0, keep_dims = mean_107_keep_dims_0, x = x_337_cast_fp16)[name = tensor("mean_107_cast_fp16")]; + tensor sub_85_cast_fp16 = sub(x = x_337_cast_fp16, y = mean_107_cast_fp16)[name = tensor("sub_85_cast_fp16")]; + tensor square_69_cast_fp16 = square(x = sub_85_cast_fp16)[name = tensor("square_69_cast_fp16")]; + tensor reduce_mean_139_axes_0 = const()[name = tensor("reduce_mean_139_axes_0"), val = tensor([-1])]; + tensor reduce_mean_139_keep_dims_0 = const()[name = tensor("reduce_mean_139_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_139_cast_fp16 = reduce_mean(axes = reduce_mean_139_axes_0, keep_dims = reduce_mean_139_keep_dims_0, x = square_69_cast_fp16)[name = tensor("reduce_mean_139_cast_fp16")]; + tensor var_3039_to_fp16 = const()[name = tensor("op_3039_to_fp16"), val = tensor(0x1.1p-20)]; + tensor var_3040_cast_fp16 = add(x = reduce_mean_139_cast_fp16, y = var_3039_to_fp16)[name = tensor("op_3040_cast_fp16")]; + tensor var_3041_cast_fp16 = sqrt(x = var_3040_cast_fp16)[name = tensor("op_3041_cast_fp16")]; + tensor x_339_cast_fp16 = real_div(x = sub_85_cast_fp16, y = var_3041_cast_fp16)[name = tensor("x_339_cast_fp16")]; + tensor var_3043_cast_fp16 = mul(x = x_339_cast_fp16, y = flow_net_res_blocks_4_in_ln_weight_to_fp16)[name = tensor("op_3043_cast_fp16")]; + tensor x_341_cast_fp16 = add(x = var_3043_cast_fp16, y = flow_net_res_blocks_4_in_ln_bias_to_fp16)[name = tensor("x_341_cast_fp16")]; + tensor var_3045_promoted_to_fp16 = const()[name = tensor("op_3045_promoted_to_fp16"), val = tensor(0x1p+0)]; + tensor var_3046_cast_fp16 = add(x = var_3029_cast_fp16_1, y = var_3045_promoted_to_fp16)[name = tensor("op_3046_cast_fp16")]; + tensor var_3047_cast_fp16 = mul(x = x_341_cast_fp16, y = var_3046_cast_fp16)[name = tensor("op_3047_cast_fp16")]; + tensor input_525_cast_fp16 = add(x = var_3047_cast_fp16, y = var_3029_cast_fp16_0)[name = tensor("input_525_cast_fp16")]; + tensor linear_201_cast_fp16 = linear(bias = flow_net_res_blocks_4_mlp_0_bias_to_fp16, weight = flow_net_res_blocks_4_mlp_0_weight_to_fp16, x = input_525_cast_fp16)[name = tensor("linear_201_cast_fp16")]; + tensor input_529_cast_fp16 = silu(x = linear_201_cast_fp16)[name = tensor("input_529_cast_fp16")]; + tensor linear_202_cast_fp16 = linear(bias = flow_net_res_blocks_4_mlp_2_bias_to_fp16, weight = flow_net_res_blocks_4_mlp_2_weight_to_fp16, x = input_529_cast_fp16)[name = tensor("linear_202_cast_fp16")]; + tensor var_3058_cast_fp16 = mul(x = var_3029_cast_fp16_2, y = linear_202_cast_fp16)[name = tensor("op_3058_cast_fp16")]; + tensor x_343_cast_fp16 = add(x = x_337_cast_fp16, y = var_3058_cast_fp16)[name = tensor("x_343_cast_fp16")]; + tensor linear_203_cast_fp16 = linear(bias = flow_net_res_blocks_5_adaLN_modulation_1_bias_to_fp16, weight = flow_net_res_blocks_5_adaLN_modulation_1_weight_to_fp16, x = input_491_cast_fp16)[name = tensor("linear_203_cast_fp16")]; + tensor var_3068_split_sizes_0 = const()[name = tensor("op_3068_split_sizes_0"), val = tensor([512, 512, 512])]; + tensor var_3068_axis_0 = const()[name = tensor("op_3068_axis_0"), val = tensor(-1)]; + tensor var_3068_cast_fp16_0, tensor var_3068_cast_fp16_1, tensor var_3068_cast_fp16_2 = split(axis = var_3068_axis_0, split_sizes = var_3068_split_sizes_0, x = linear_203_cast_fp16)[name = tensor("op_3068_cast_fp16")]; + tensor mean_109_axes_0 = const()[name = tensor("mean_109_axes_0"), val = tensor([-1])]; + tensor mean_109_keep_dims_0 = const()[name = tensor("mean_109_keep_dims_0"), val = tensor(true)]; + tensor mean_109_cast_fp16 = reduce_mean(axes = mean_109_axes_0, keep_dims = mean_109_keep_dims_0, x = x_343_cast_fp16)[name = tensor("mean_109_cast_fp16")]; + tensor sub_86_cast_fp16 = sub(x = x_343_cast_fp16, y = mean_109_cast_fp16)[name = tensor("sub_86_cast_fp16")]; + tensor square_70_cast_fp16 = square(x = sub_86_cast_fp16)[name = tensor("square_70_cast_fp16")]; + tensor reduce_mean_141_axes_0 = const()[name = tensor("reduce_mean_141_axes_0"), val = tensor([-1])]; + tensor reduce_mean_141_keep_dims_0 = const()[name = tensor("reduce_mean_141_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_141_cast_fp16 = reduce_mean(axes = reduce_mean_141_axes_0, keep_dims = reduce_mean_141_keep_dims_0, x = square_70_cast_fp16)[name = tensor("reduce_mean_141_cast_fp16")]; + tensor var_3078_to_fp16 = const()[name = tensor("op_3078_to_fp16"), val = tensor(0x1.1p-20)]; + tensor var_3079_cast_fp16 = add(x = reduce_mean_141_cast_fp16, y = var_3078_to_fp16)[name = tensor("op_3079_cast_fp16")]; + tensor var_3080_cast_fp16 = sqrt(x = var_3079_cast_fp16)[name = tensor("op_3080_cast_fp16")]; + tensor x_345_cast_fp16 = real_div(x = sub_86_cast_fp16, y = var_3080_cast_fp16)[name = tensor("x_345_cast_fp16")]; + tensor var_3082_cast_fp16 = mul(x = x_345_cast_fp16, y = flow_net_res_blocks_5_in_ln_weight_to_fp16)[name = tensor("op_3082_cast_fp16")]; + tensor x_347_cast_fp16 = add(x = var_3082_cast_fp16, y = flow_net_res_blocks_5_in_ln_bias_to_fp16)[name = tensor("x_347_cast_fp16")]; + tensor var_3084_promoted_to_fp16 = const()[name = tensor("op_3084_promoted_to_fp16"), val = tensor(0x1p+0)]; + tensor var_3085_cast_fp16 = add(x = var_3068_cast_fp16_1, y = var_3084_promoted_to_fp16)[name = tensor("op_3085_cast_fp16")]; + tensor var_3086_cast_fp16 = mul(x = x_347_cast_fp16, y = var_3085_cast_fp16)[name = tensor("op_3086_cast_fp16")]; + tensor input_533_cast_fp16 = add(x = var_3086_cast_fp16, y = var_3068_cast_fp16_0)[name = tensor("input_533_cast_fp16")]; + tensor linear_204_cast_fp16 = linear(bias = flow_net_res_blocks_5_mlp_0_bias_to_fp16, weight = flow_net_res_blocks_5_mlp_0_weight_to_fp16, x = input_533_cast_fp16)[name = tensor("linear_204_cast_fp16")]; + tensor input_537_cast_fp16 = silu(x = linear_204_cast_fp16)[name = tensor("input_537_cast_fp16")]; + tensor linear_205_cast_fp16 = linear(bias = flow_net_res_blocks_5_mlp_2_bias_to_fp16, weight = flow_net_res_blocks_5_mlp_2_weight_to_fp16, x = input_537_cast_fp16)[name = tensor("linear_205_cast_fp16")]; + tensor var_3097_cast_fp16 = mul(x = var_3068_cast_fp16_2, y = linear_205_cast_fp16)[name = tensor("op_3097_cast_fp16")]; + tensor x_349_cast_fp16 = add(x = x_343_cast_fp16, y = var_3097_cast_fp16)[name = tensor("x_349_cast_fp16")]; + tensor linear_206_cast_fp16 = linear(bias = flow_net_final_layer_adaLN_modulation_1_bias_to_fp16, weight = flow_net_final_layer_adaLN_modulation_1_weight_to_fp16, x = input_491_cast_fp16)[name = tensor("linear_206_cast_fp16")]; + tensor var_3106_split_sizes_0 = const()[name = tensor("op_3106_split_sizes_0"), val = tensor([512, 512])]; + tensor var_3106_axis_0 = const()[name = tensor("op_3106_axis_0"), val = tensor(-1)]; + tensor var_3106_cast_fp16_0, tensor var_3106_cast_fp16_1 = split(axis = var_3106_axis_0, split_sizes = var_3106_split_sizes_0, x = linear_206_cast_fp16)[name = tensor("op_3106_cast_fp16")]; + tensor mean_axes_0 = const()[name = tensor("mean_axes_0"), val = tensor([-1])]; + tensor mean_keep_dims_0 = const()[name = tensor("mean_keep_dims_0"), val = tensor(true)]; + tensor mean_cast_fp16 = reduce_mean(axes = mean_axes_0, keep_dims = mean_keep_dims_0, x = x_349_cast_fp16)[name = tensor("mean_cast_fp16")]; + tensor sub_87_cast_fp16 = sub(x = x_349_cast_fp16, y = mean_cast_fp16)[name = tensor("sub_87_cast_fp16")]; + tensor square_71_cast_fp16 = square(x = sub_87_cast_fp16)[name = tensor("square_71_cast_fp16")]; + tensor reduce_mean_143_axes_0 = const()[name = tensor("reduce_mean_143_axes_0"), val = tensor([-1])]; + tensor reduce_mean_143_keep_dims_0 = const()[name = tensor("reduce_mean_143_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_143_cast_fp16 = reduce_mean(axes = reduce_mean_143_axes_0, keep_dims = reduce_mean_143_keep_dims_0, x = square_71_cast_fp16)[name = tensor("reduce_mean_143_cast_fp16")]; + tensor var_3113_to_fp16 = const()[name = tensor("op_3113_to_fp16"), val = tensor(0x1.1p-20)]; + tensor var_3114_cast_fp16 = add(x = reduce_mean_143_cast_fp16, y = var_3113_to_fp16)[name = tensor("op_3114_cast_fp16")]; + tensor var_3115_cast_fp16 = sqrt(x = var_3114_cast_fp16)[name = tensor("op_3115_cast_fp16")]; + tensor x_cast_fp16 = real_div(x = sub_87_cast_fp16, y = var_3115_cast_fp16)[name = tensor("x_cast_fp16")]; + tensor var_3117_promoted_to_fp16 = const()[name = tensor("op_3117_promoted_to_fp16"), val = tensor(0x1p+0)]; + tensor var_3118_cast_fp16 = add(x = var_3106_cast_fp16_1, y = var_3117_promoted_to_fp16)[name = tensor("op_3118_cast_fp16")]; + tensor var_3119_cast_fp16 = mul(x = x_cast_fp16, y = var_3118_cast_fp16)[name = tensor("op_3119_cast_fp16")]; + tensor input_cast_fp16 = add(x = var_3119_cast_fp16, y = var_3106_cast_fp16_0)[name = tensor("input_cast_fp16")]; + tensor linear_207_cast_fp16 = linear(bias = flow_net_final_layer_linear_bias_to_fp16, weight = flow_net_final_layer_linear_weight_to_fp16, x = input_cast_fp16)[name = tensor("linear_207_cast_fp16")]; + tensor var_3124_to_fp16 = const()[name = tensor("op_3124_to_fp16"), val = tensor(0x1p-3)]; + tensor var_3125_cast_fp16 = mul(x = linear_207_cast_fp16, y = var_3124_to_fp16)[name = tensor("op_3125_cast_fp16")]; + tensor var_3127_cast_fp16 = add(x = input_475_cast_fp16, y = var_3125_cast_fp16)[name = tensor("op_3127_cast_fp16")]; + tensor var_3127_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_3127_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor latent_final = cast(dtype = var_3127_cast_fp16_to_fp32_dtype_0, x = var_3127_cast_fp16)[name = tensor("cast_144")]; + } -> (latent_final); +} \ No newline at end of file diff --git a/v2.1/portuguese_24l/flow_decoder_fused.mlmodelc/weights/weight.bin b/v2.1/portuguese_24l/flow_decoder_fused.mlmodelc/weights/weight.bin new file mode 100644 index 0000000000000000000000000000000000000000..cea8cca7f8dee90a55f1d0ca46903c8b578db21d --- /dev/null +++ b/v2.1/portuguese_24l/flow_decoder_fused.mlmodelc/weights/weight.bin @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:a55edc1a818e94cf3f3b00b89be7ddae87aa064b6131986b92f6625e4d5753f6 +size 19012608 diff --git a/v2.1/portuguese_24l/flow_decoder_fused.mlpackage/Data/com.apple.CoreML/model.mlmodel b/v2.1/portuguese_24l/flow_decoder_fused.mlpackage/Data/com.apple.CoreML/model.mlmodel new file mode 100644 index 0000000000000000000000000000000000000000..3a160fed741602c26e615c1bdd9f1346b7b46690 --- /dev/null +++ b/v2.1/portuguese_24l/flow_decoder_fused.mlpackage/Data/com.apple.CoreML/model.mlmodel @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:068c34cc36c7e3587a19363c7714447f8b99975af3352a48929b066d092da66c +size 301733 diff --git a/v2.1/portuguese_24l/flow_decoder_fused.mlpackage/Data/com.apple.CoreML/weights/weight.bin b/v2.1/portuguese_24l/flow_decoder_fused.mlpackage/Data/com.apple.CoreML/weights/weight.bin new file mode 100644 index 0000000000000000000000000000000000000000..cea8cca7f8dee90a55f1d0ca46903c8b578db21d --- /dev/null +++ b/v2.1/portuguese_24l/flow_decoder_fused.mlpackage/Data/com.apple.CoreML/weights/weight.bin @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:a55edc1a818e94cf3f3b00b89be7ddae87aa064b6131986b92f6625e4d5753f6 +size 19012608 diff --git a/v2.1/portuguese_24l/flow_decoder_fused.mlpackage/Manifest.json b/v2.1/portuguese_24l/flow_decoder_fused.mlpackage/Manifest.json new file mode 100644 index 0000000000000000000000000000000000000000..03e35bcd5ed3e725860a61acd8ed4012fe227f1f --- /dev/null +++ b/v2.1/portuguese_24l/flow_decoder_fused.mlpackage/Manifest.json @@ -0,0 +1,18 @@ +{ + "fileFormatVersion": "1.0.0", + "itemInfoEntries": { + "00878162-40AA-48E0-B24D-4DCADF9FD3C1": { + "author": "com.apple.CoreML", + "description": "CoreML Model Weights", + "name": "weights", + "path": "com.apple.CoreML/weights" + }, + "D62DDF8F-9BF8-4FBB-8BFB-55D90DB306C6": { + "author": "com.apple.CoreML", + "description": "CoreML Model Specification", + "name": "model.mlmodel", + "path": "com.apple.CoreML/model.mlmodel" + } + }, + "rootModelIdentifier": "D62DDF8F-9BF8-4FBB-8BFB-55D90DB306C6" +} diff --git a/v2.1/portuguese_24l/flowlm_step.mlmodelc/analytics/coremldata.bin b/v2.1/portuguese_24l/flowlm_step.mlmodelc/analytics/coremldata.bin new file mode 100644 index 0000000000000000000000000000000000000000..e857288a26752500c195f378889942456f683729 --- /dev/null +++ b/v2.1/portuguese_24l/flowlm_step.mlmodelc/analytics/coremldata.bin @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:c21a9e18235fd3b4f83abbf130c741ac8dd607600b7f4d56c8588e69a8aa2403 +size 243 diff --git a/v2.1/portuguese_24l/flowlm_step.mlmodelc/coremldata.bin b/v2.1/portuguese_24l/flowlm_step.mlmodelc/coremldata.bin new file mode 100644 index 0000000000000000000000000000000000000000..680da724cb575a72ee38801553e60d9051e924ba --- /dev/null +++ b/v2.1/portuguese_24l/flowlm_step.mlmodelc/coremldata.bin @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:ddf33d4eef08cea65c750745837fd09610220b0425189940c023279971b271c5 +size 3559 diff --git a/v2.1/portuguese_24l/flowlm_step.mlmodelc/model.mil b/v2.1/portuguese_24l/flowlm_step.mlmodelc/model.mil new file mode 100644 index 0000000000000000000000000000000000000000..d12a998f3d2be5e7361b2155805638b166ac5338 --- /dev/null +++ b/v2.1/portuguese_24l/flowlm_step.mlmodelc/model.mil @@ -0,0 +1,5337 @@ +program(1.0) +[buildInfo = dict, tensor>({{"coremlc-component-MIL", "3520.4.1"}, {"coremlc-version", "3520.5.1"}, {"coremltools-component-torch", "2.12.0"}, {"coremltools-source-dialect", "TorchScript"}, {"coremltools-version", "9.0"}})] +{ + func main(tensor bos_emb, tensor cache0, tensor cache1, tensor cache10, tensor cache11, tensor cache12, tensor cache13, tensor cache14, tensor cache15, tensor cache16, tensor cache17, tensor cache18, tensor cache19, tensor cache2, tensor cache20, tensor cache21, tensor cache22, tensor cache23, tensor cache3, tensor cache4, tensor cache5, tensor cache6, tensor cache7, tensor cache8, tensor cache9, tensor position0, tensor position1, tensor position10, tensor position11, tensor position12, tensor position13, tensor position14, tensor position15, tensor position16, tensor position17, tensor position18, tensor position19, tensor position2, tensor position20, tensor position21, tensor position22, tensor position23, tensor position3, tensor position4, tensor position5, tensor position6, tensor position7, tensor position8, tensor position9, tensor sequence) { + tensor sequence_to_fp16_dtype_0 = const()[name = tensor("sequence_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor sequence_to_fp16 = cast(dtype = sequence_to_fp16_dtype_0, x = sequence)[name = tensor("cast_555")]; + tensor var_198_cast_fp16 = not_equal(x = sequence_to_fp16, y = sequence_to_fp16)[name = tensor("op_198_cast_fp16")]; + tensor expand_dims_0_axes_0 = const()[name = tensor("expand_dims_0_axes_0"), val = tensor([0, 1])]; + tensor bos_emb_to_fp16_dtype_0 = const()[name = tensor("bos_emb_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor bos_emb_to_fp16 = cast(dtype = bos_emb_to_fp16_dtype_0, x = bos_emb)[name = tensor("cast_554")]; + tensor expand_dims_0_cast_fp16 = expand_dims(axes = expand_dims_0_axes_0, x = bos_emb_to_fp16)[name = tensor("expand_dims_0_cast_fp16")]; + tensor input_1_cast_fp16 = select(a = expand_dims_0_cast_fp16, b = sequence_to_fp16, cond = var_198_cast_fp16)[name = tensor("input_1_cast_fp16")]; + tensor input_linear_weight_to_fp16 = const()[name = tensor("input_linear_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(64)))]; + tensor linear_0_bias_0_to_fp16 = const()[name = tensor("linear_0_bias_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(65664)))]; + tensor linear_0_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = input_linear_weight_to_fp16, x = input_1_cast_fp16)[name = tensor("linear_0_cast_fp16")]; + tensor input_5_axes_0 = const()[name = tensor("input_5_axes_0"), val = tensor([-1])]; + tensor norm0_1_weight_to_fp16 = const()[name = tensor("norm0_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(67776)))]; + tensor norm0_1_bias_to_fp16 = const()[name = tensor("norm0_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(69888)))]; + tensor var_204_to_fp16 = const()[name = tensor("op_204_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_5_cast_fp16 = layer_norm(axes = input_5_axes_0, beta = norm0_1_bias_to_fp16, epsilon = var_204_to_fp16, gamma = norm0_1_weight_to_fp16, x = linear_0_cast_fp16)[name = tensor("input_5_cast_fp16")]; + tensor attn0_in_proj_weight_to_fp16 = const()[name = tensor("attn0_in_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(72000)))]; + tensor linear_1_bias_0_to_fp16 = const()[name = tensor("linear_1_bias_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(6363520)))]; + tensor linear_1_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = attn0_in_proj_weight_to_fp16, x = input_5_cast_fp16)[name = tensor("linear_1_cast_fp16")]; + tensor var_218 = const()[name = tensor("op_218"), val = tensor([1, 1, 3, 16, 64])]; + tensor qkv_1_cast_fp16 = reshape(shape = var_218, x = linear_1_cast_fp16)[name = tensor("qkv_1_cast_fp16")]; + tensor q_1_begin_0 = const()[name = tensor("q_1_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor q_1_end_0 = const()[name = tensor("q_1_end_0"), val = tensor([1, 1, 1, 16, 64])]; + tensor q_1_end_mask_0 = const()[name = tensor("q_1_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor q_1_squeeze_mask_0 = const()[name = tensor("q_1_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor q_1_cast_fp16 = slice_by_index(begin = q_1_begin_0, end = q_1_end_0, end_mask = q_1_end_mask_0, squeeze_mask = q_1_squeeze_mask_0, x = qkv_1_cast_fp16)[name = tensor("q_1_cast_fp16")]; + tensor k_1_begin_0 = const()[name = tensor("k_1_begin_0"), val = tensor([0, 0, 1, 0, 0])]; + tensor k_1_end_0 = const()[name = tensor("k_1_end_0"), val = tensor([1, 1, 2, 16, 64])]; + tensor k_1_end_mask_0 = const()[name = tensor("k_1_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor k_1_squeeze_mask_0 = const()[name = tensor("k_1_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor k_1_cast_fp16 = slice_by_index(begin = k_1_begin_0, end = k_1_end_0, end_mask = k_1_end_mask_0, squeeze_mask = k_1_squeeze_mask_0, x = qkv_1_cast_fp16)[name = tensor("k_1_cast_fp16")]; + tensor v_1_begin_0 = const()[name = tensor("v_1_begin_0"), val = tensor([0, 0, 2, 0, 0])]; + tensor v_1_end_0 = const()[name = tensor("v_1_end_0"), val = tensor([1, 1, 3, 16, 64])]; + tensor v_1_end_mask_0 = const()[name = tensor("v_1_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor v_1_squeeze_mask_0 = const()[name = tensor("v_1_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor v_1_cast_fp16 = slice_by_index(begin = v_1_begin_0, end = v_1_end_0, end_mask = v_1_end_mask_0, squeeze_mask = v_1_squeeze_mask_0, x = qkv_1_cast_fp16)[name = tensor("v_1_cast_fp16")]; + tensor var_284 = const()[name = tensor("op_284"), val = tensor([1, 1, 1, 1])]; + tensor position0_to_fp16_dtype_0 = const()[name = tensor("position0_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor position0_to_fp16 = cast(dtype = position0_to_fp16_dtype_0, x = position0)[name = tensor("cast_553")]; + tensor ts_5_cast_fp16 = reshape(shape = var_284, x = position0_to_fp16)[name = tensor("ts_5_cast_fp16")]; + tensor var_291 = const()[name = tensor("op_291"), val = tensor([1, 1, 16, 32, 2])]; + tensor q_complex_1_cast_fp16 = reshape(shape = var_291, x = q_1_cast_fp16)[name = tensor("q_complex_1_cast_fp16")]; + tensor var_298 = const()[name = tensor("op_298"), val = tensor([1, 1, 16, 32, 2])]; + tensor k_complex_1_cast_fp16 = reshape(shape = var_298, x = k_1_cast_fp16)[name = tensor("k_complex_1_cast_fp16")]; + tensor var_302_begin_0 = const()[name = tensor("op_302_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_302_end_0 = const()[name = tensor("op_302_end_0"), val = tensor([1, 1, 16, 32, 1])]; + tensor var_302_end_mask_0 = const()[name = tensor("op_302_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_302_squeeze_mask_0 = const()[name = tensor("op_302_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_302_cast_fp16 = slice_by_index(begin = var_302_begin_0, end = var_302_end_0, end_mask = var_302_end_mask_0, squeeze_mask = var_302_squeeze_mask_0, x = q_complex_1_cast_fp16)[name = tensor("op_302_cast_fp16")]; + tensor var_310_begin_0 = const()[name = tensor("op_310_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_310_end_0 = const()[name = tensor("op_310_end_0"), val = tensor([1, 1, 16, 32, 2])]; + tensor var_310_end_mask_0 = const()[name = tensor("op_310_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_310_squeeze_mask_0 = const()[name = tensor("op_310_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_310_cast_fp16 = slice_by_index(begin = var_310_begin_0, end = var_310_end_0, end_mask = var_310_end_mask_0, squeeze_mask = var_310_squeeze_mask_0, x = q_complex_1_cast_fp16)[name = tensor("op_310_cast_fp16")]; + tensor var_318_begin_0 = const()[name = tensor("op_318_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_318_end_0 = const()[name = tensor("op_318_end_0"), val = tensor([1, 1, 16, 32, 1])]; + tensor var_318_end_mask_0 = const()[name = tensor("op_318_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_318_squeeze_mask_0 = const()[name = tensor("op_318_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_318_cast_fp16 = slice_by_index(begin = var_318_begin_0, end = var_318_end_0, end_mask = var_318_end_mask_0, squeeze_mask = var_318_squeeze_mask_0, x = k_complex_1_cast_fp16)[name = tensor("op_318_cast_fp16")]; + tensor var_326_begin_0 = const()[name = tensor("op_326_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_326_end_0 = const()[name = tensor("op_326_end_0"), val = tensor([1, 1, 16, 32, 2])]; + tensor var_326_end_mask_0 = const()[name = tensor("op_326_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_326_squeeze_mask_0 = const()[name = tensor("op_326_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_326_cast_fp16 = slice_by_index(begin = var_326_begin_0, end = var_326_end_0, end_mask = var_326_end_mask_0, squeeze_mask = var_326_squeeze_mask_0, x = k_complex_1_cast_fp16)[name = tensor("op_326_cast_fp16")]; + tensor freqs_1_to_fp16 = const()[name = tensor("freqs_1_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(6369728)))]; + tensor var_332_cast_fp16 = mul(x = freqs_1_to_fp16, y = ts_5_cast_fp16)[name = tensor("op_332_cast_fp16")]; + tensor rotr_1_cast_fp16 = cos(x = var_332_cast_fp16)[name = tensor("rotr_1_cast_fp16")]; + tensor roti_1_cast_fp16 = sin(x = var_332_cast_fp16)[name = tensor("roti_1_cast_fp16")]; + tensor var_336_cast_fp16 = mul(x = var_302_cast_fp16, y = rotr_1_cast_fp16)[name = tensor("op_336_cast_fp16")]; + tensor var_337_cast_fp16 = mul(x = var_310_cast_fp16, y = roti_1_cast_fp16)[name = tensor("op_337_cast_fp16")]; + tensor qor_1_cast_fp16 = sub(x = var_336_cast_fp16, y = var_337_cast_fp16)[name = tensor("qor_1_cast_fp16")]; + tensor var_340_cast_fp16 = mul(x = var_302_cast_fp16, y = roti_1_cast_fp16)[name = tensor("op_340_cast_fp16")]; + tensor var_341_cast_fp16 = mul(x = var_310_cast_fp16, y = rotr_1_cast_fp16)[name = tensor("op_341_cast_fp16")]; + tensor qoi_1_cast_fp16 = add(x = var_340_cast_fp16, y = var_341_cast_fp16)[name = tensor("qoi_1_cast_fp16")]; + tensor var_344_cast_fp16 = mul(x = var_318_cast_fp16, y = rotr_1_cast_fp16)[name = tensor("op_344_cast_fp16")]; + tensor var_345_cast_fp16 = mul(x = var_326_cast_fp16, y = roti_1_cast_fp16)[name = tensor("op_345_cast_fp16")]; + tensor kor_1_cast_fp16 = sub(x = var_344_cast_fp16, y = var_345_cast_fp16)[name = tensor("kor_1_cast_fp16")]; + tensor var_348_cast_fp16 = mul(x = var_318_cast_fp16, y = roti_1_cast_fp16)[name = tensor("op_348_cast_fp16")]; + tensor var_349_cast_fp16 = mul(x = var_326_cast_fp16, y = rotr_1_cast_fp16)[name = tensor("op_349_cast_fp16")]; + tensor koi_1_cast_fp16 = add(x = var_348_cast_fp16, y = var_349_cast_fp16)[name = tensor("koi_1_cast_fp16")]; + tensor qo_1_axis_0 = const()[name = tensor("qo_1_axis_0"), val = tensor(-1)]; + tensor qo_1_cast_fp16 = stack(axis = qo_1_axis_0, values = (qor_1_cast_fp16, qoi_1_cast_fp16))[name = tensor("qo_1_cast_fp16")]; + tensor ko_1_axis_0 = const()[name = tensor("ko_1_axis_0"), val = tensor(-1)]; + tensor ko_1_cast_fp16 = stack(axis = ko_1_axis_0, values = (kor_1_cast_fp16, koi_1_cast_fp16))[name = tensor("ko_1_cast_fp16")]; + tensor var_382 = const()[name = tensor("op_382"), val = tensor([1, 1, 16, 64])]; + tensor q_3_cast_fp16 = reshape(shape = var_382, x = qo_1_cast_fp16)[name = tensor("q_3_cast_fp16")]; + tensor var_388 = const()[name = tensor("op_388"), val = tensor([1, 1, 16, 64])]; + tensor k_3_cast_fp16 = reshape(shape = var_388, x = ko_1_cast_fp16)[name = tensor("k_3_cast_fp16")]; + tensor _inversed_413_y_0_to_fp16 = const()[name = tensor("_inversed_413_y_0_to_fp16"), val = tensor(0x1p-9)]; + tensor _inversed_413_cast_fp16 = mul(x = ts_5_cast_fp16, y = _inversed_413_y_0_to_fp16)[name = tensor("_inversed_413_cast_fp16")]; + tensor var_414_cast_fp16 = floor(x = _inversed_413_cast_fp16)[name = tensor("op_414_cast_fp16")]; + tensor var_415_to_fp16 = const()[name = tensor("op_415_to_fp16"), val = tensor(0x1p+9)]; + tensor var_416_cast_fp16 = mul(x = var_414_cast_fp16, y = var_415_to_fp16)[name = tensor("op_416_cast_fp16")]; + tensor write_indices_float_3_cast_fp16 = sub(x = ts_5_cast_fp16, y = var_416_cast_fp16)[name = tensor("write_indices_float_3_cast_fp16")]; + tensor var_423_dtype_0 = const()[name = tensor("op_423_dtype_0"), val = tensor("int32")]; + tensor write_indices_1_reps_0 = const()[name = tensor("write_indices_1_reps_0"), val = tensor([1, 1, 16, 64])]; + tensor write_indices_float_3_cast_fp16_to_int32 = cast(dtype = var_423_dtype_0, x = write_indices_float_3_cast_fp16)[name = tensor("cast_552")]; + tensor write_indices_1 = tile(reps = write_indices_1_reps_0, x = write_indices_float_3_cast_fp16_to_int32)[name = tensor("write_indices_1")]; + tensor var_433_begin_0 = const()[name = tensor("op_433_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_433_end_0 = const()[name = tensor("op_433_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor var_433_end_mask_0 = const()[name = tensor("op_433_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_433_squeeze_mask_0 = const()[name = tensor("op_433_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor cache0_to_fp16_dtype_0 = const()[name = tensor("cache0_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor cache0_to_fp16 = cast(dtype = cache0_to_fp16_dtype_0, x = cache0)[name = tensor("cast_551")]; + tensor var_433_cast_fp16 = slice_by_index(begin = var_433_begin_0, end = var_433_end_0, end_mask = var_433_end_mask_0, squeeze_mask = var_433_squeeze_mask_0, x = cache0_to_fp16)[name = tensor("op_433_cast_fp16")]; + tensor var_435_axis_0 = const()[name = tensor("op_435_axis_0"), val = tensor(1)]; + tensor var_435_mode_0 = const()[name = tensor("op_435_mode_0"), val = tensor("update")]; + tensor var_435_validate_indices_0 = const()[name = tensor("op_435_validate_indices_0"), val = tensor(false)]; + tensor var_435_cast_fp16 = scatter_along_axis(axis = var_435_axis_0, data = var_433_cast_fp16, indices = write_indices_1, mode = var_435_mode_0, updates = k_3_cast_fp16, validate_indices = var_435_validate_indices_0)[name = tensor("op_435_cast_fp16")]; + tensor concat_2 = const()[name = tensor("concat_2"), val = tensor([0, 0, 0, 0, 0])]; + tensor concat_3 = const()[name = tensor("concat_3"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_1_internal_tensor_assign_1_stride_0 = const()[name = tensor("new_cache_1_internal_tensor_assign_1_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_1_internal_tensor_assign_1_begin_mask_0 = const()[name = tensor("new_cache_1_internal_tensor_assign_1_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_1_internal_tensor_assign_1_end_mask_0 = const()[name = tensor("new_cache_1_internal_tensor_assign_1_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_1_internal_tensor_assign_1_squeeze_mask_0 = const()[name = tensor("new_cache_1_internal_tensor_assign_1_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_48 = const()[name = tensor("shape_48"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_0 = const()[name = tensor("reduce_prod_0"), val = tensor(1048576)]; + tensor range_1d_0_start_0 = const()[name = tensor("range_1d_0_start_0"), val = tensor(0)]; + tensor range_1d_0_step_0 = const()[name = tensor("range_1d_0_step_0"), val = tensor(1)]; + tensor range_1d_0 = range_1d(end = reduce_prod_0, start = range_1d_0_start_0, step = range_1d_0_step_0)[name = tensor("range_1d_0")]; + tensor reshape_0 = reshape(shape = shape_48, x = range_1d_0)[name = tensor("reshape_0")]; + tensor slice_by_index_0 = slice_by_index(begin = concat_2, begin_mask = new_cache_1_internal_tensor_assign_1_begin_mask_0, end = concat_3, end_mask = new_cache_1_internal_tensor_assign_1_end_mask_0, squeeze_mask = new_cache_1_internal_tensor_assign_1_squeeze_mask_0, stride = new_cache_1_internal_tensor_assign_1_stride_0, x = reshape_0)[name = tensor("slice_by_index_0")]; + tensor reshape_1_shape_0 = const()[name = tensor("reshape_1_shape_0"), val = tensor([-1])]; + tensor reshape_1 = reshape(shape = reshape_1_shape_0, x = slice_by_index_0)[name = tensor("reshape_1")]; + tensor reshape_2_shape_0 = const()[name = tensor("reshape_2_shape_0"), val = tensor([-1])]; + tensor reshape_2_cast_fp16 = reshape(shape = reshape_2_shape_0, x = var_435_cast_fp16)[name = tensor("reshape_2_cast_fp16")]; + tensor reshape_3_shape_0 = const()[name = tensor("reshape_3_shape_0"), val = tensor([-1])]; + tensor reshape_3_cast_fp16 = reshape(shape = reshape_3_shape_0, x = cache0_to_fp16)[name = tensor("reshape_3_cast_fp16")]; + tensor scatter_0_mode_0 = const()[name = tensor("scatter_0_mode_0"), val = tensor("update")]; + tensor scatter_0_axis_0 = const()[name = tensor("scatter_0_axis_0"), val = tensor(0)]; + tensor scatter_0_validate_indices_0 = const()[name = tensor("scatter_0_validate_indices_0"), val = tensor(false)]; + tensor scatter_0_cast_fp16 = scatter(axis = scatter_0_axis_0, data = reshape_3_cast_fp16, indices = reshape_1, mode = scatter_0_mode_0, updates = reshape_2_cast_fp16, validate_indices = scatter_0_validate_indices_0)[name = tensor("scatter_0_cast_fp16")]; + tensor reshape_4_cast_fp16 = reshape(shape = shape_48, x = scatter_0_cast_fp16)[name = tensor("reshape_4_cast_fp16")]; + tensor var_443_begin_0 = const()[name = tensor("op_443_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor var_443_end_0 = const()[name = tensor("op_443_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor var_443_end_mask_0 = const()[name = tensor("op_443_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_443_squeeze_mask_0 = const()[name = tensor("op_443_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor var_443_cast_fp16 = slice_by_index(begin = var_443_begin_0, end = var_443_end_0, end_mask = var_443_end_mask_0, squeeze_mask = var_443_squeeze_mask_0, x = reshape_4_cast_fp16)[name = tensor("op_443_cast_fp16")]; + tensor var_445_axis_0 = const()[name = tensor("op_445_axis_0"), val = tensor(1)]; + tensor var_445_mode_0 = const()[name = tensor("op_445_mode_0"), val = tensor("update")]; + tensor var_445_validate_indices_0 = const()[name = tensor("op_445_validate_indices_0"), val = tensor(false)]; + tensor var_445_cast_fp16 = scatter_along_axis(axis = var_445_axis_0, data = var_443_cast_fp16, indices = write_indices_1, mode = var_445_mode_0, updates = v_1_cast_fp16, validate_indices = var_445_validate_indices_0)[name = tensor("op_445_cast_fp16")]; + tensor concat_4 = const()[name = tensor("concat_4"), val = tensor([1, 0, 0, 0, 0])]; + tensor concat_5 = const()[name = tensor("concat_5"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_1_internal_tensor_assign_2_stride_0 = const()[name = tensor("new_cache_1_internal_tensor_assign_2_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_1_internal_tensor_assign_2_begin_mask_0 = const()[name = tensor("new_cache_1_internal_tensor_assign_2_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_1_internal_tensor_assign_2_end_mask_0 = const()[name = tensor("new_cache_1_internal_tensor_assign_2_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_1_internal_tensor_assign_2_squeeze_mask_0 = const()[name = tensor("new_cache_1_internal_tensor_assign_2_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_49 = const()[name = tensor("shape_49"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_1 = const()[name = tensor("reduce_prod_1"), val = tensor(1048576)]; + tensor range_1d_1_start_0 = const()[name = tensor("range_1d_1_start_0"), val = tensor(0)]; + tensor range_1d_1_step_0 = const()[name = tensor("range_1d_1_step_0"), val = tensor(1)]; + tensor range_1d_1 = range_1d(end = reduce_prod_1, start = range_1d_1_start_0, step = range_1d_1_step_0)[name = tensor("range_1d_1")]; + tensor reshape_5 = reshape(shape = shape_49, x = range_1d_1)[name = tensor("reshape_5")]; + tensor slice_by_index_1 = slice_by_index(begin = concat_4, begin_mask = new_cache_1_internal_tensor_assign_2_begin_mask_0, end = concat_5, end_mask = new_cache_1_internal_tensor_assign_2_end_mask_0, squeeze_mask = new_cache_1_internal_tensor_assign_2_squeeze_mask_0, stride = new_cache_1_internal_tensor_assign_2_stride_0, x = reshape_5)[name = tensor("slice_by_index_1")]; + tensor reshape_6_shape_0 = const()[name = tensor("reshape_6_shape_0"), val = tensor([-1])]; + tensor reshape_6 = reshape(shape = reshape_6_shape_0, x = slice_by_index_1)[name = tensor("reshape_6")]; + tensor reshape_7_shape_0 = const()[name = tensor("reshape_7_shape_0"), val = tensor([-1])]; + tensor reshape_7_cast_fp16 = reshape(shape = reshape_7_shape_0, x = var_445_cast_fp16)[name = tensor("reshape_7_cast_fp16")]; + tensor reshape_8_shape_0 = const()[name = tensor("reshape_8_shape_0"), val = tensor([-1])]; + tensor reshape_8_cast_fp16 = reshape(shape = reshape_8_shape_0, x = reshape_4_cast_fp16)[name = tensor("reshape_8_cast_fp16")]; + tensor scatter_1_mode_0 = const()[name = tensor("scatter_1_mode_0"), val = tensor("update")]; + tensor scatter_1_axis_0 = const()[name = tensor("scatter_1_axis_0"), val = tensor(0)]; + tensor scatter_1_validate_indices_0 = const()[name = tensor("scatter_1_validate_indices_0"), val = tensor(false)]; + tensor scatter_1_cast_fp16 = scatter(axis = scatter_1_axis_0, data = reshape_8_cast_fp16, indices = reshape_6, mode = scatter_1_mode_0, updates = reshape_7_cast_fp16, validate_indices = scatter_1_validate_indices_0)[name = tensor("scatter_1_cast_fp16")]; + tensor reshape_9_cast_fp16 = reshape(shape = shape_49, x = scatter_1_cast_fp16)[name = tensor("reshape_9_cast_fp16")]; + tensor reshape_9_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("reshape_9_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor keys_1_begin_0 = const()[name = tensor("keys_1_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor keys_1_end_0 = const()[name = tensor("keys_1_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor keys_1_end_mask_0 = const()[name = tensor("keys_1_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor keys_1_squeeze_mask_0 = const()[name = tensor("keys_1_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor keys_1_cast_fp16 = slice_by_index(begin = keys_1_begin_0, end = keys_1_end_0, end_mask = keys_1_end_mask_0, squeeze_mask = keys_1_squeeze_mask_0, x = reshape_9_cast_fp16)[name = tensor("keys_1_cast_fp16")]; + tensor values_1_begin_0 = const()[name = tensor("values_1_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor values_1_end_0 = const()[name = tensor("values_1_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor values_1_end_mask_0 = const()[name = tensor("values_1_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor values_1_squeeze_mask_0 = const()[name = tensor("values_1_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor values_1_cast_fp16 = slice_by_index(begin = values_1_begin_0, end = values_1_end_0, end_mask = values_1_end_mask_0, squeeze_mask = values_1_squeeze_mask_0, x = reshape_9_cast_fp16)[name = tensor("values_1_cast_fp16")]; + tensor var_457_cast_fp16 = not_equal(x = keys_1_cast_fp16, y = keys_1_cast_fp16)[name = tensor("op_457_cast_fp16")]; + tensor var_463_to_fp16 = const()[name = tensor("op_463_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(6369856)))]; + tensor keys_3_cast_fp16 = select(a = var_463_to_fp16, b = keys_1_cast_fp16, cond = var_457_cast_fp16)[name = tensor("keys_3_cast_fp16")]; + tensor var_465_cast_fp16 = not_equal(x = values_1_cast_fp16, y = values_1_cast_fp16)[name = tensor("op_465_cast_fp16")]; + tensor values_3_cast_fp16 = select(a = var_463_to_fp16, b = values_1_cast_fp16, cond = var_465_cast_fp16)[name = tensor("values_3_cast_fp16")]; + tensor var_489 = const()[name = tensor("op_489"), val = tensor([0, 2, 1, 3])]; + tensor var_505 = const()[name = tensor("op_505"), val = tensor([1, 1, 1])]; + tensor var_506_cast_fp16 = reshape(shape = var_505, x = position0_to_fp16)[name = tensor("op_506_cast_fp16")]; + tensor var_526_to_fp16 = const()[name = tensor("op_526_to_fp16"), val = tensor(0x1p+0)]; + tensor valid_len_1_cast_fp16 = add(x = var_506_cast_fp16, y = var_526_to_fp16)[name = tensor("valid_len_1_cast_fp16")]; + tensor k_positions_1_promoted_to_fp16 = const()[name = tensor("k_positions_1_promoted_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(7418496)))]; + tensor valid_mask_1_cast_fp16 = less(x = k_positions_1_promoted_to_fp16, y = valid_len_1_cast_fp16)[name = tensor("valid_mask_1_cast_fp16")]; + tensor causal_mask_1_cast_fp16 = less_equal(x = k_positions_1_promoted_to_fp16, y = var_506_cast_fp16)[name = tensor("causal_mask_1_cast_fp16")]; + tensor attn_mask_1 = logical_and(x = valid_mask_1_cast_fp16, y = causal_mask_1_cast_fp16)[name = tensor("attn_mask_1")]; + tensor attn_mask_3_axes_0 = const()[name = tensor("attn_mask_3_axes_0"), val = tensor([1])]; + tensor attn_mask_3 = expand_dims(axes = attn_mask_3_axes_0, x = attn_mask_1)[name = tensor("attn_mask_3")]; + tensor var_544_transpose_x_0 = const()[name = tensor("op_544_transpose_x_0"), val = tensor(false)]; + tensor var_544_transpose_y_0 = const()[name = tensor("op_544_transpose_y_0"), val = tensor(false)]; + tensor transpose_72_perm_0 = const()[name = tensor("transpose_72_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_73_perm_0 = const()[name = tensor("transpose_73_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_73 = transpose(perm = transpose_73_perm_0, x = keys_3_cast_fp16)[name = tensor("transpose_213")]; + tensor transpose_72 = transpose(perm = transpose_72_perm_0, x = q_3_cast_fp16)[name = tensor("transpose_214")]; + tensor var_544_cast_fp16 = matmul(transpose_x = var_544_transpose_x_0, transpose_y = var_544_transpose_y_0, x = transpose_72, y = transpose_73)[name = tensor("op_544_cast_fp16")]; + tensor var_538_to_fp16 = const()[name = tensor("op_538_to_fp16"), val = tensor([0x1p-3])]; + tensor attn_weights_1_cast_fp16 = mul(x = var_544_cast_fp16, y = var_538_to_fp16)[name = tensor("attn_weights_1_cast_fp16")]; + tensor var_546 = logical_not(x = attn_mask_3)[name = tensor("op_546")]; + tensor var_547_to_fp16 = const()[name = tensor("op_547_to_fp16"), val = tensor(-inf)]; + tensor attn_weights_3_cast_fp16 = select(a = var_547_to_fp16, b = attn_weights_1_cast_fp16, cond = var_546)[name = tensor("attn_weights_3_cast_fp16")]; + tensor var_549 = const()[name = tensor("op_549"), val = tensor(-1)]; + tensor attn_weights_5_cast_fp16 = softmax(axis = var_549, x = attn_weights_3_cast_fp16)[name = tensor("attn_weights_5_cast_fp16")]; + tensor attn_output_1_transpose_x_0 = const()[name = tensor("attn_output_1_transpose_x_0"), val = tensor(false)]; + tensor attn_output_1_transpose_y_0 = const()[name = tensor("attn_output_1_transpose_y_0"), val = tensor(false)]; + tensor values_5_cast_fp16 = transpose(perm = var_489, x = values_3_cast_fp16)[name = tensor("transpose_215")]; + tensor attn_output_1_cast_fp16 = matmul(transpose_x = attn_output_1_transpose_x_0, transpose_y = attn_output_1_transpose_y_0, x = attn_weights_5_cast_fp16, y = values_5_cast_fp16)[name = tensor("attn_output_1_cast_fp16")]; + tensor var_557 = const()[name = tensor("op_557"), val = tensor([0, 2, 1, 3])]; + tensor var_562 = const()[name = tensor("op_562"), val = tensor([1, 1, 1024])]; + tensor var_558_cast_fp16 = transpose(perm = var_557, x = attn_output_1_cast_fp16)[name = tensor("transpose_212")]; + tensor input_7_cast_fp16 = reshape(shape = var_562, x = var_558_cast_fp16)[name = tensor("input_7_cast_fp16")]; + tensor attn0_out_proj_weight_to_fp16 = const()[name = tensor("attn0_out_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(7419584)))]; + tensor linear_2_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = attn0_out_proj_weight_to_fp16, x = input_7_cast_fp16)[name = tensor("linear_2_cast_fp16")]; + tensor var_568_to_fp16 = const()[name = tensor("op_568_to_fp16"), val = tensor(0x1p+0)]; + tensor var_569_cast_fp16 = add(x = position0_to_fp16, y = var_568_to_fp16)[name = tensor("op_569_cast_fp16")]; + tensor var_569_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_569_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor input_9_cast_fp16 = add(x = linear_0_cast_fp16, y = linear_2_cast_fp16)[name = tensor("input_9_cast_fp16")]; + tensor input_11_axes_0 = const()[name = tensor("input_11_axes_0"), val = tensor([-1])]; + tensor norm0_2_weight_to_fp16 = const()[name = tensor("norm0_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9516800)))]; + tensor norm0_2_bias_to_fp16 = const()[name = tensor("norm0_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9518912)))]; + tensor var_573_to_fp16 = const()[name = tensor("op_573_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_11_cast_fp16 = layer_norm(axes = input_11_axes_0, beta = norm0_2_bias_to_fp16, epsilon = var_573_to_fp16, gamma = norm0_2_weight_to_fp16, x = input_9_cast_fp16)[name = tensor("input_11_cast_fp16")]; + tensor linear0_1_weight_to_fp16 = const()[name = tensor("linear0_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9521024)))]; + tensor linear_3_bias_0_to_fp16 = const()[name = tensor("linear_3_bias_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(17909696)))]; + tensor linear_3_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = linear0_1_weight_to_fp16, x = input_11_cast_fp16)[name = tensor("linear_3_cast_fp16")]; + tensor input_13_mode_0 = const()[name = tensor("input_13_mode_0"), val = tensor("EXACT")]; + tensor input_13_cast_fp16 = gelu(mode = input_13_mode_0, x = linear_3_cast_fp16)[name = tensor("input_13_cast_fp16")]; + tensor linear0_2_weight_to_fp16 = const()[name = tensor("linear0_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(17917952)))]; + tensor linear_4_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = linear0_2_weight_to_fp16, x = input_13_cast_fp16)[name = tensor("linear_4_cast_fp16")]; + tensor input_15_cast_fp16 = add(x = input_9_cast_fp16, y = linear_4_cast_fp16)[name = tensor("input_15_cast_fp16")]; + tensor input_17_axes_0 = const()[name = tensor("input_17_axes_0"), val = tensor([-1])]; + tensor norm1_1_weight_to_fp16 = const()[name = tensor("norm1_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(26306624)))]; + tensor norm1_1_bias_to_fp16 = const()[name = tensor("norm1_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(26308736)))]; + tensor var_590_to_fp16 = const()[name = tensor("op_590_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_17_cast_fp16 = layer_norm(axes = input_17_axes_0, beta = norm1_1_bias_to_fp16, epsilon = var_590_to_fp16, gamma = norm1_1_weight_to_fp16, x = input_15_cast_fp16)[name = tensor("input_17_cast_fp16")]; + tensor attn1_in_proj_weight_to_fp16 = const()[name = tensor("attn1_in_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(26310848)))]; + tensor linear_5_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = attn1_in_proj_weight_to_fp16, x = input_17_cast_fp16)[name = tensor("linear_5_cast_fp16")]; + tensor var_604 = const()[name = tensor("op_604"), val = tensor([1, 1, 3, 16, 64])]; + tensor qkv_3_cast_fp16 = reshape(shape = var_604, x = linear_5_cast_fp16)[name = tensor("qkv_3_cast_fp16")]; + tensor q_7_begin_0 = const()[name = tensor("q_7_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor q_7_end_0 = const()[name = tensor("q_7_end_0"), val = tensor([1, 1, 1, 16, 64])]; + tensor q_7_end_mask_0 = const()[name = tensor("q_7_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor q_7_squeeze_mask_0 = const()[name = tensor("q_7_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor q_7_cast_fp16 = slice_by_index(begin = q_7_begin_0, end = q_7_end_0, end_mask = q_7_end_mask_0, squeeze_mask = q_7_squeeze_mask_0, x = qkv_3_cast_fp16)[name = tensor("q_7_cast_fp16")]; + tensor k_5_begin_0 = const()[name = tensor("k_5_begin_0"), val = tensor([0, 0, 1, 0, 0])]; + tensor k_5_end_0 = const()[name = tensor("k_5_end_0"), val = tensor([1, 1, 2, 16, 64])]; + tensor k_5_end_mask_0 = const()[name = tensor("k_5_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor k_5_squeeze_mask_0 = const()[name = tensor("k_5_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor k_5_cast_fp16 = slice_by_index(begin = k_5_begin_0, end = k_5_end_0, end_mask = k_5_end_mask_0, squeeze_mask = k_5_squeeze_mask_0, x = qkv_3_cast_fp16)[name = tensor("k_5_cast_fp16")]; + tensor v_3_begin_0 = const()[name = tensor("v_3_begin_0"), val = tensor([0, 0, 2, 0, 0])]; + tensor v_3_end_0 = const()[name = tensor("v_3_end_0"), val = tensor([1, 1, 3, 16, 64])]; + tensor v_3_end_mask_0 = const()[name = tensor("v_3_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor v_3_squeeze_mask_0 = const()[name = tensor("v_3_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor v_3_cast_fp16 = slice_by_index(begin = v_3_begin_0, end = v_3_end_0, end_mask = v_3_end_mask_0, squeeze_mask = v_3_squeeze_mask_0, x = qkv_3_cast_fp16)[name = tensor("v_3_cast_fp16")]; + tensor var_670 = const()[name = tensor("op_670"), val = tensor([1, 1, 1, 1])]; + tensor position1_to_fp16_dtype_0 = const()[name = tensor("position1_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor position1_to_fp16 = cast(dtype = position1_to_fp16_dtype_0, x = position1)[name = tensor("cast_548")]; + tensor ts_11_cast_fp16 = reshape(shape = var_670, x = position1_to_fp16)[name = tensor("ts_11_cast_fp16")]; + tensor var_677 = const()[name = tensor("op_677"), val = tensor([1, 1, 16, 32, 2])]; + tensor q_complex_3_cast_fp16 = reshape(shape = var_677, x = q_7_cast_fp16)[name = tensor("q_complex_3_cast_fp16")]; + tensor var_684 = const()[name = tensor("op_684"), val = tensor([1, 1, 16, 32, 2])]; + tensor k_complex_3_cast_fp16 = reshape(shape = var_684, x = k_5_cast_fp16)[name = tensor("k_complex_3_cast_fp16")]; + tensor var_688_begin_0 = const()[name = tensor("op_688_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_688_end_0 = const()[name = tensor("op_688_end_0"), val = tensor([1, 1, 16, 32, 1])]; + tensor var_688_end_mask_0 = const()[name = tensor("op_688_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_688_squeeze_mask_0 = const()[name = tensor("op_688_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_688_cast_fp16 = slice_by_index(begin = var_688_begin_0, end = var_688_end_0, end_mask = var_688_end_mask_0, squeeze_mask = var_688_squeeze_mask_0, x = q_complex_3_cast_fp16)[name = tensor("op_688_cast_fp16")]; + tensor var_696_begin_0 = const()[name = tensor("op_696_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_696_end_0 = const()[name = tensor("op_696_end_0"), val = tensor([1, 1, 16, 32, 2])]; + tensor var_696_end_mask_0 = const()[name = tensor("op_696_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_696_squeeze_mask_0 = const()[name = tensor("op_696_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_696_cast_fp16 = slice_by_index(begin = var_696_begin_0, end = var_696_end_0, end_mask = var_696_end_mask_0, squeeze_mask = var_696_squeeze_mask_0, x = q_complex_3_cast_fp16)[name = tensor("op_696_cast_fp16")]; + tensor var_704_begin_0 = const()[name = tensor("op_704_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_704_end_0 = const()[name = tensor("op_704_end_0"), val = tensor([1, 1, 16, 32, 1])]; + tensor var_704_end_mask_0 = const()[name = tensor("op_704_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_704_squeeze_mask_0 = const()[name = tensor("op_704_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_704_cast_fp16 = slice_by_index(begin = var_704_begin_0, end = var_704_end_0, end_mask = var_704_end_mask_0, squeeze_mask = var_704_squeeze_mask_0, x = k_complex_3_cast_fp16)[name = tensor("op_704_cast_fp16")]; + tensor var_712_begin_0 = const()[name = tensor("op_712_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_712_end_0 = const()[name = tensor("op_712_end_0"), val = tensor([1, 1, 16, 32, 2])]; + tensor var_712_end_mask_0 = const()[name = tensor("op_712_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_712_squeeze_mask_0 = const()[name = tensor("op_712_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_712_cast_fp16 = slice_by_index(begin = var_712_begin_0, end = var_712_end_0, end_mask = var_712_end_mask_0, squeeze_mask = var_712_squeeze_mask_0, x = k_complex_3_cast_fp16)[name = tensor("op_712_cast_fp16")]; + tensor freqs_3_to_fp16 = const()[name = tensor("freqs_3_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(32602368)))]; + tensor var_718_cast_fp16 = mul(x = freqs_3_to_fp16, y = ts_11_cast_fp16)[name = tensor("op_718_cast_fp16")]; + tensor rotr_3_cast_fp16 = cos(x = var_718_cast_fp16)[name = tensor("rotr_3_cast_fp16")]; + tensor roti_3_cast_fp16 = sin(x = var_718_cast_fp16)[name = tensor("roti_3_cast_fp16")]; + tensor var_722_cast_fp16 = mul(x = var_688_cast_fp16, y = rotr_3_cast_fp16)[name = tensor("op_722_cast_fp16")]; + tensor var_723_cast_fp16 = mul(x = var_696_cast_fp16, y = roti_3_cast_fp16)[name = tensor("op_723_cast_fp16")]; + tensor qor_5_cast_fp16 = sub(x = var_722_cast_fp16, y = var_723_cast_fp16)[name = tensor("qor_5_cast_fp16")]; + tensor var_726_cast_fp16 = mul(x = var_688_cast_fp16, y = roti_3_cast_fp16)[name = tensor("op_726_cast_fp16")]; + tensor var_727_cast_fp16 = mul(x = var_696_cast_fp16, y = rotr_3_cast_fp16)[name = tensor("op_727_cast_fp16")]; + tensor qoi_5_cast_fp16 = add(x = var_726_cast_fp16, y = var_727_cast_fp16)[name = tensor("qoi_5_cast_fp16")]; + tensor var_730_cast_fp16 = mul(x = var_704_cast_fp16, y = rotr_3_cast_fp16)[name = tensor("op_730_cast_fp16")]; + tensor var_731_cast_fp16 = mul(x = var_712_cast_fp16, y = roti_3_cast_fp16)[name = tensor("op_731_cast_fp16")]; + tensor kor_5_cast_fp16 = sub(x = var_730_cast_fp16, y = var_731_cast_fp16)[name = tensor("kor_5_cast_fp16")]; + tensor var_734_cast_fp16 = mul(x = var_704_cast_fp16, y = roti_3_cast_fp16)[name = tensor("op_734_cast_fp16")]; + tensor var_735_cast_fp16 = mul(x = var_712_cast_fp16, y = rotr_3_cast_fp16)[name = tensor("op_735_cast_fp16")]; + tensor koi_5_cast_fp16 = add(x = var_734_cast_fp16, y = var_735_cast_fp16)[name = tensor("koi_5_cast_fp16")]; + tensor qo_3_axis_0 = const()[name = tensor("qo_3_axis_0"), val = tensor(-1)]; + tensor qo_3_cast_fp16 = stack(axis = qo_3_axis_0, values = (qor_5_cast_fp16, qoi_5_cast_fp16))[name = tensor("qo_3_cast_fp16")]; + tensor ko_3_axis_0 = const()[name = tensor("ko_3_axis_0"), val = tensor(-1)]; + tensor ko_3_cast_fp16 = stack(axis = ko_3_axis_0, values = (kor_5_cast_fp16, koi_5_cast_fp16))[name = tensor("ko_3_cast_fp16")]; + tensor var_768 = const()[name = tensor("op_768"), val = tensor([1, 1, 16, 64])]; + tensor q_9_cast_fp16 = reshape(shape = var_768, x = qo_3_cast_fp16)[name = tensor("q_9_cast_fp16")]; + tensor var_774 = const()[name = tensor("op_774"), val = tensor([1, 1, 16, 64])]; + tensor k_7_cast_fp16 = reshape(shape = var_774, x = ko_3_cast_fp16)[name = tensor("k_7_cast_fp16")]; + tensor _inversed_799_y_0_to_fp16 = const()[name = tensor("_inversed_799_y_0_to_fp16"), val = tensor(0x1p-9)]; + tensor _inversed_799_cast_fp16 = mul(x = ts_11_cast_fp16, y = _inversed_799_y_0_to_fp16)[name = tensor("_inversed_799_cast_fp16")]; + tensor var_800_cast_fp16 = floor(x = _inversed_799_cast_fp16)[name = tensor("op_800_cast_fp16")]; + tensor var_801_to_fp16 = const()[name = tensor("op_801_to_fp16"), val = tensor(0x1p+9)]; + tensor var_802_cast_fp16 = mul(x = var_800_cast_fp16, y = var_801_to_fp16)[name = tensor("op_802_cast_fp16")]; + tensor write_indices_float_7_cast_fp16 = sub(x = ts_11_cast_fp16, y = var_802_cast_fp16)[name = tensor("write_indices_float_7_cast_fp16")]; + tensor var_809_dtype_0 = const()[name = tensor("op_809_dtype_0"), val = tensor("int32")]; + tensor write_indices_3_reps_0 = const()[name = tensor("write_indices_3_reps_0"), val = tensor([1, 1, 16, 64])]; + tensor write_indices_float_7_cast_fp16_to_int32 = cast(dtype = var_809_dtype_0, x = write_indices_float_7_cast_fp16)[name = tensor("cast_547")]; + tensor write_indices_3 = tile(reps = write_indices_3_reps_0, x = write_indices_float_7_cast_fp16_to_int32)[name = tensor("write_indices_3")]; + tensor var_819_begin_0 = const()[name = tensor("op_819_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_819_end_0 = const()[name = tensor("op_819_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor var_819_end_mask_0 = const()[name = tensor("op_819_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_819_squeeze_mask_0 = const()[name = tensor("op_819_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor cache1_to_fp16_dtype_0 = const()[name = tensor("cache1_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor cache1_to_fp16 = cast(dtype = cache1_to_fp16_dtype_0, x = cache1)[name = tensor("cast_546")]; + tensor var_819_cast_fp16 = slice_by_index(begin = var_819_begin_0, end = var_819_end_0, end_mask = var_819_end_mask_0, squeeze_mask = var_819_squeeze_mask_0, x = cache1_to_fp16)[name = tensor("op_819_cast_fp16")]; + tensor var_821_axis_0 = const()[name = tensor("op_821_axis_0"), val = tensor(1)]; + tensor var_821_mode_0 = const()[name = tensor("op_821_mode_0"), val = tensor("update")]; + tensor var_821_validate_indices_0 = const()[name = tensor("op_821_validate_indices_0"), val = tensor(false)]; + tensor var_821_cast_fp16 = scatter_along_axis(axis = var_821_axis_0, data = var_819_cast_fp16, indices = write_indices_3, mode = var_821_mode_0, updates = k_7_cast_fp16, validate_indices = var_821_validate_indices_0)[name = tensor("op_821_cast_fp16")]; + tensor concat_9 = const()[name = tensor("concat_9"), val = tensor([0, 0, 0, 0, 0])]; + tensor concat_10 = const()[name = tensor("concat_10"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_3_internal_tensor_assign_1_stride_0 = const()[name = tensor("new_cache_3_internal_tensor_assign_1_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_3_internal_tensor_assign_1_begin_mask_0 = const()[name = tensor("new_cache_3_internal_tensor_assign_1_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_3_internal_tensor_assign_1_end_mask_0 = const()[name = tensor("new_cache_3_internal_tensor_assign_1_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_3_internal_tensor_assign_1_squeeze_mask_0 = const()[name = tensor("new_cache_3_internal_tensor_assign_1_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_50 = const()[name = tensor("shape_50"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_2 = const()[name = tensor("reduce_prod_2"), val = tensor(1048576)]; + tensor range_1d_2_start_0 = const()[name = tensor("range_1d_2_start_0"), val = tensor(0)]; + tensor range_1d_2_step_0 = const()[name = tensor("range_1d_2_step_0"), val = tensor(1)]; + tensor range_1d_2 = range_1d(end = reduce_prod_2, start = range_1d_2_start_0, step = range_1d_2_step_0)[name = tensor("range_1d_2")]; + tensor reshape_10 = reshape(shape = shape_50, x = range_1d_2)[name = tensor("reshape_10")]; + tensor slice_by_index_2 = slice_by_index(begin = concat_9, begin_mask = new_cache_3_internal_tensor_assign_1_begin_mask_0, end = concat_10, end_mask = new_cache_3_internal_tensor_assign_1_end_mask_0, squeeze_mask = new_cache_3_internal_tensor_assign_1_squeeze_mask_0, stride = new_cache_3_internal_tensor_assign_1_stride_0, x = reshape_10)[name = tensor("slice_by_index_2")]; + tensor reshape_11_shape_0 = const()[name = tensor("reshape_11_shape_0"), val = tensor([-1])]; + tensor reshape_11 = reshape(shape = reshape_11_shape_0, x = slice_by_index_2)[name = tensor("reshape_11")]; + tensor reshape_12_shape_0 = const()[name = tensor("reshape_12_shape_0"), val = tensor([-1])]; + tensor reshape_12_cast_fp16 = reshape(shape = reshape_12_shape_0, x = var_821_cast_fp16)[name = tensor("reshape_12_cast_fp16")]; + tensor reshape_13_shape_0 = const()[name = tensor("reshape_13_shape_0"), val = tensor([-1])]; + tensor reshape_13_cast_fp16 = reshape(shape = reshape_13_shape_0, x = cache1_to_fp16)[name = tensor("reshape_13_cast_fp16")]; + tensor scatter_2_mode_0 = const()[name = tensor("scatter_2_mode_0"), val = tensor("update")]; + tensor scatter_2_axis_0 = const()[name = tensor("scatter_2_axis_0"), val = tensor(0)]; + tensor scatter_2_validate_indices_0 = const()[name = tensor("scatter_2_validate_indices_0"), val = tensor(false)]; + tensor scatter_2_cast_fp16 = scatter(axis = scatter_2_axis_0, data = reshape_13_cast_fp16, indices = reshape_11, mode = scatter_2_mode_0, updates = reshape_12_cast_fp16, validate_indices = scatter_2_validate_indices_0)[name = tensor("scatter_2_cast_fp16")]; + tensor reshape_14_cast_fp16 = reshape(shape = shape_50, x = scatter_2_cast_fp16)[name = tensor("reshape_14_cast_fp16")]; + tensor var_829_begin_0 = const()[name = tensor("op_829_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor var_829_end_0 = const()[name = tensor("op_829_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor var_829_end_mask_0 = const()[name = tensor("op_829_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_829_squeeze_mask_0 = const()[name = tensor("op_829_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor var_829_cast_fp16 = slice_by_index(begin = var_829_begin_0, end = var_829_end_0, end_mask = var_829_end_mask_0, squeeze_mask = var_829_squeeze_mask_0, x = reshape_14_cast_fp16)[name = tensor("op_829_cast_fp16")]; + tensor var_831_axis_0 = const()[name = tensor("op_831_axis_0"), val = tensor(1)]; + tensor var_831_mode_0 = const()[name = tensor("op_831_mode_0"), val = tensor("update")]; + tensor var_831_validate_indices_0 = const()[name = tensor("op_831_validate_indices_0"), val = tensor(false)]; + tensor var_831_cast_fp16 = scatter_along_axis(axis = var_831_axis_0, data = var_829_cast_fp16, indices = write_indices_3, mode = var_831_mode_0, updates = v_3_cast_fp16, validate_indices = var_831_validate_indices_0)[name = tensor("op_831_cast_fp16")]; + tensor concat_11 = const()[name = tensor("concat_11"), val = tensor([1, 0, 0, 0, 0])]; + tensor concat_12 = const()[name = tensor("concat_12"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_3_internal_tensor_assign_2_stride_0 = const()[name = tensor("new_cache_3_internal_tensor_assign_2_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_3_internal_tensor_assign_2_begin_mask_0 = const()[name = tensor("new_cache_3_internal_tensor_assign_2_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_3_internal_tensor_assign_2_end_mask_0 = const()[name = tensor("new_cache_3_internal_tensor_assign_2_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_3_internal_tensor_assign_2_squeeze_mask_0 = const()[name = tensor("new_cache_3_internal_tensor_assign_2_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_51 = const()[name = tensor("shape_51"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_3 = const()[name = tensor("reduce_prod_3"), val = tensor(1048576)]; + tensor range_1d_3_start_0 = const()[name = tensor("range_1d_3_start_0"), val = tensor(0)]; + tensor range_1d_3_step_0 = const()[name = tensor("range_1d_3_step_0"), val = tensor(1)]; + tensor range_1d_3 = range_1d(end = reduce_prod_3, start = range_1d_3_start_0, step = range_1d_3_step_0)[name = tensor("range_1d_3")]; + tensor reshape_15 = reshape(shape = shape_51, x = range_1d_3)[name = tensor("reshape_15")]; + tensor slice_by_index_3 = slice_by_index(begin = concat_11, begin_mask = new_cache_3_internal_tensor_assign_2_begin_mask_0, end = concat_12, end_mask = new_cache_3_internal_tensor_assign_2_end_mask_0, squeeze_mask = new_cache_3_internal_tensor_assign_2_squeeze_mask_0, stride = new_cache_3_internal_tensor_assign_2_stride_0, x = reshape_15)[name = tensor("slice_by_index_3")]; + tensor reshape_16_shape_0 = const()[name = tensor("reshape_16_shape_0"), val = tensor([-1])]; + tensor reshape_16 = reshape(shape = reshape_16_shape_0, x = slice_by_index_3)[name = tensor("reshape_16")]; + tensor reshape_17_shape_0 = const()[name = tensor("reshape_17_shape_0"), val = tensor([-1])]; + tensor reshape_17_cast_fp16 = reshape(shape = reshape_17_shape_0, x = var_831_cast_fp16)[name = tensor("reshape_17_cast_fp16")]; + tensor reshape_18_shape_0 = const()[name = tensor("reshape_18_shape_0"), val = tensor([-1])]; + tensor reshape_18_cast_fp16 = reshape(shape = reshape_18_shape_0, x = reshape_14_cast_fp16)[name = tensor("reshape_18_cast_fp16")]; + tensor scatter_3_mode_0 = const()[name = tensor("scatter_3_mode_0"), val = tensor("update")]; + tensor scatter_3_axis_0 = const()[name = tensor("scatter_3_axis_0"), val = tensor(0)]; + tensor scatter_3_validate_indices_0 = const()[name = tensor("scatter_3_validate_indices_0"), val = tensor(false)]; + tensor scatter_3_cast_fp16 = scatter(axis = scatter_3_axis_0, data = reshape_18_cast_fp16, indices = reshape_16, mode = scatter_3_mode_0, updates = reshape_17_cast_fp16, validate_indices = scatter_3_validate_indices_0)[name = tensor("scatter_3_cast_fp16")]; + tensor reshape_19_cast_fp16 = reshape(shape = shape_51, x = scatter_3_cast_fp16)[name = tensor("reshape_19_cast_fp16")]; + tensor reshape_19_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("reshape_19_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor keys_7_begin_0 = const()[name = tensor("keys_7_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor keys_7_end_0 = const()[name = tensor("keys_7_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor keys_7_end_mask_0 = const()[name = tensor("keys_7_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor keys_7_squeeze_mask_0 = const()[name = tensor("keys_7_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor keys_7_cast_fp16 = slice_by_index(begin = keys_7_begin_0, end = keys_7_end_0, end_mask = keys_7_end_mask_0, squeeze_mask = keys_7_squeeze_mask_0, x = reshape_19_cast_fp16)[name = tensor("keys_7_cast_fp16")]; + tensor values_7_begin_0 = const()[name = tensor("values_7_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor values_7_end_0 = const()[name = tensor("values_7_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor values_7_end_mask_0 = const()[name = tensor("values_7_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor values_7_squeeze_mask_0 = const()[name = tensor("values_7_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor values_7_cast_fp16 = slice_by_index(begin = values_7_begin_0, end = values_7_end_0, end_mask = values_7_end_mask_0, squeeze_mask = values_7_squeeze_mask_0, x = reshape_19_cast_fp16)[name = tensor("values_7_cast_fp16")]; + tensor var_843_cast_fp16 = not_equal(x = keys_7_cast_fp16, y = keys_7_cast_fp16)[name = tensor("op_843_cast_fp16")]; + tensor keys_9_cast_fp16 = select(a = var_463_to_fp16, b = keys_7_cast_fp16, cond = var_843_cast_fp16)[name = tensor("keys_9_cast_fp16")]; + tensor var_851_cast_fp16 = not_equal(x = values_7_cast_fp16, y = values_7_cast_fp16)[name = tensor("op_851_cast_fp16")]; + tensor values_9_cast_fp16 = select(a = var_463_to_fp16, b = values_7_cast_fp16, cond = var_851_cast_fp16)[name = tensor("values_9_cast_fp16")]; + tensor var_875 = const()[name = tensor("op_875"), val = tensor([0, 2, 1, 3])]; + tensor var_891 = const()[name = tensor("op_891"), val = tensor([1, 1, 1])]; + tensor var_892_cast_fp16 = reshape(shape = var_891, x = position1_to_fp16)[name = tensor("op_892_cast_fp16")]; + tensor var_912_to_fp16 = const()[name = tensor("op_912_to_fp16"), val = tensor(0x1p+0)]; + tensor valid_len_3_cast_fp16 = add(x = var_892_cast_fp16, y = var_912_to_fp16)[name = tensor("valid_len_3_cast_fp16")]; + tensor valid_mask_3_cast_fp16 = less(x = k_positions_1_promoted_to_fp16, y = valid_len_3_cast_fp16)[name = tensor("valid_mask_3_cast_fp16")]; + tensor causal_mask_3_cast_fp16 = less_equal(x = k_positions_1_promoted_to_fp16, y = var_892_cast_fp16)[name = tensor("causal_mask_3_cast_fp16")]; + tensor attn_mask_5 = logical_and(x = valid_mask_3_cast_fp16, y = causal_mask_3_cast_fp16)[name = tensor("attn_mask_5")]; + tensor attn_mask_7_axes_0 = const()[name = tensor("attn_mask_7_axes_0"), val = tensor([1])]; + tensor attn_mask_7 = expand_dims(axes = attn_mask_7_axes_0, x = attn_mask_5)[name = tensor("attn_mask_7")]; + tensor var_930_transpose_x_0 = const()[name = tensor("op_930_transpose_x_0"), val = tensor(false)]; + tensor var_930_transpose_y_0 = const()[name = tensor("op_930_transpose_y_0"), val = tensor(false)]; + tensor transpose_74_perm_0 = const()[name = tensor("transpose_74_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_75_perm_0 = const()[name = tensor("transpose_75_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_75 = transpose(perm = transpose_75_perm_0, x = keys_9_cast_fp16)[name = tensor("transpose_209")]; + tensor transpose_74 = transpose(perm = transpose_74_perm_0, x = q_9_cast_fp16)[name = tensor("transpose_210")]; + tensor var_930_cast_fp16 = matmul(transpose_x = var_930_transpose_x_0, transpose_y = var_930_transpose_y_0, x = transpose_74, y = transpose_75)[name = tensor("op_930_cast_fp16")]; + tensor var_924_to_fp16 = const()[name = tensor("op_924_to_fp16"), val = tensor([0x1p-3])]; + tensor attn_weights_7_cast_fp16 = mul(x = var_930_cast_fp16, y = var_924_to_fp16)[name = tensor("attn_weights_7_cast_fp16")]; + tensor var_932 = logical_not(x = attn_mask_7)[name = tensor("op_932")]; + tensor var_933_to_fp16 = const()[name = tensor("op_933_to_fp16"), val = tensor(-inf)]; + tensor attn_weights_9_cast_fp16 = select(a = var_933_to_fp16, b = attn_weights_7_cast_fp16, cond = var_932)[name = tensor("attn_weights_9_cast_fp16")]; + tensor var_935 = const()[name = tensor("op_935"), val = tensor(-1)]; + tensor attn_weights_11_cast_fp16 = softmax(axis = var_935, x = attn_weights_9_cast_fp16)[name = tensor("attn_weights_11_cast_fp16")]; + tensor attn_output_3_transpose_x_0 = const()[name = tensor("attn_output_3_transpose_x_0"), val = tensor(false)]; + tensor attn_output_3_transpose_y_0 = const()[name = tensor("attn_output_3_transpose_y_0"), val = tensor(false)]; + tensor values_11_cast_fp16 = transpose(perm = var_875, x = values_9_cast_fp16)[name = tensor("transpose_211")]; + tensor attn_output_3_cast_fp16 = matmul(transpose_x = attn_output_3_transpose_x_0, transpose_y = attn_output_3_transpose_y_0, x = attn_weights_11_cast_fp16, y = values_11_cast_fp16)[name = tensor("attn_output_3_cast_fp16")]; + tensor var_943 = const()[name = tensor("op_943"), val = tensor([0, 2, 1, 3])]; + tensor var_948 = const()[name = tensor("op_948"), val = tensor([1, 1, 1024])]; + tensor var_944_cast_fp16 = transpose(perm = var_943, x = attn_output_3_cast_fp16)[name = tensor("transpose_208")]; + tensor input_19_cast_fp16 = reshape(shape = var_948, x = var_944_cast_fp16)[name = tensor("input_19_cast_fp16")]; + tensor attn1_out_proj_weight_to_fp16 = const()[name = tensor("attn1_out_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(32602496)))]; + tensor linear_6_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = attn1_out_proj_weight_to_fp16, x = input_19_cast_fp16)[name = tensor("linear_6_cast_fp16")]; + tensor var_954_to_fp16 = const()[name = tensor("op_954_to_fp16"), val = tensor(0x1p+0)]; + tensor var_955_cast_fp16 = add(x = position1_to_fp16, y = var_954_to_fp16)[name = tensor("op_955_cast_fp16")]; + tensor var_955_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_955_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor input_21_cast_fp16 = add(x = input_15_cast_fp16, y = linear_6_cast_fp16)[name = tensor("input_21_cast_fp16")]; + tensor input_23_axes_0 = const()[name = tensor("input_23_axes_0"), val = tensor([-1])]; + tensor norm1_2_weight_to_fp16 = const()[name = tensor("norm1_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(34699712)))]; + tensor norm1_2_bias_to_fp16 = const()[name = tensor("norm1_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(34701824)))]; + tensor var_959_to_fp16 = const()[name = tensor("op_959_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_23_cast_fp16 = layer_norm(axes = input_23_axes_0, beta = norm1_2_bias_to_fp16, epsilon = var_959_to_fp16, gamma = norm1_2_weight_to_fp16, x = input_21_cast_fp16)[name = tensor("input_23_cast_fp16")]; + tensor linear1_1_weight_to_fp16 = const()[name = tensor("linear1_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(34703936)))]; + tensor linear_7_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = linear1_1_weight_to_fp16, x = input_23_cast_fp16)[name = tensor("linear_7_cast_fp16")]; + tensor input_25_mode_0 = const()[name = tensor("input_25_mode_0"), val = tensor("EXACT")]; + tensor input_25_cast_fp16 = gelu(mode = input_25_mode_0, x = linear_7_cast_fp16)[name = tensor("input_25_cast_fp16")]; + tensor linear1_2_weight_to_fp16 = const()[name = tensor("linear1_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(43092608)))]; + tensor linear_8_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = linear1_2_weight_to_fp16, x = input_25_cast_fp16)[name = tensor("linear_8_cast_fp16")]; + tensor input_27_cast_fp16 = add(x = input_21_cast_fp16, y = linear_8_cast_fp16)[name = tensor("input_27_cast_fp16")]; + tensor input_29_axes_0 = const()[name = tensor("input_29_axes_0"), val = tensor([-1])]; + tensor norm2_1_weight_to_fp16 = const()[name = tensor("norm2_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(51481280)))]; + tensor norm2_1_bias_to_fp16 = const()[name = tensor("norm2_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(51483392)))]; + tensor var_976_to_fp16 = const()[name = tensor("op_976_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_29_cast_fp16 = layer_norm(axes = input_29_axes_0, beta = norm2_1_bias_to_fp16, epsilon = var_976_to_fp16, gamma = norm2_1_weight_to_fp16, x = input_27_cast_fp16)[name = tensor("input_29_cast_fp16")]; + tensor attn2_in_proj_weight_to_fp16 = const()[name = tensor("attn2_in_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(51485504)))]; + tensor linear_9_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = attn2_in_proj_weight_to_fp16, x = input_29_cast_fp16)[name = tensor("linear_9_cast_fp16")]; + tensor var_990 = const()[name = tensor("op_990"), val = tensor([1, 1, 3, 16, 64])]; + tensor qkv_5_cast_fp16 = reshape(shape = var_990, x = linear_9_cast_fp16)[name = tensor("qkv_5_cast_fp16")]; + tensor q_13_begin_0 = const()[name = tensor("q_13_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor q_13_end_0 = const()[name = tensor("q_13_end_0"), val = tensor([1, 1, 1, 16, 64])]; + tensor q_13_end_mask_0 = const()[name = tensor("q_13_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor q_13_squeeze_mask_0 = const()[name = tensor("q_13_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor q_13_cast_fp16 = slice_by_index(begin = q_13_begin_0, end = q_13_end_0, end_mask = q_13_end_mask_0, squeeze_mask = q_13_squeeze_mask_0, x = qkv_5_cast_fp16)[name = tensor("q_13_cast_fp16")]; + tensor k_9_begin_0 = const()[name = tensor("k_9_begin_0"), val = tensor([0, 0, 1, 0, 0])]; + tensor k_9_end_0 = const()[name = tensor("k_9_end_0"), val = tensor([1, 1, 2, 16, 64])]; + tensor k_9_end_mask_0 = const()[name = tensor("k_9_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor k_9_squeeze_mask_0 = const()[name = tensor("k_9_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor k_9_cast_fp16 = slice_by_index(begin = k_9_begin_0, end = k_9_end_0, end_mask = k_9_end_mask_0, squeeze_mask = k_9_squeeze_mask_0, x = qkv_5_cast_fp16)[name = tensor("k_9_cast_fp16")]; + tensor v_5_begin_0 = const()[name = tensor("v_5_begin_0"), val = tensor([0, 0, 2, 0, 0])]; + tensor v_5_end_0 = const()[name = tensor("v_5_end_0"), val = tensor([1, 1, 3, 16, 64])]; + tensor v_5_end_mask_0 = const()[name = tensor("v_5_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor v_5_squeeze_mask_0 = const()[name = tensor("v_5_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor v_5_cast_fp16 = slice_by_index(begin = v_5_begin_0, end = v_5_end_0, end_mask = v_5_end_mask_0, squeeze_mask = v_5_squeeze_mask_0, x = qkv_5_cast_fp16)[name = tensor("v_5_cast_fp16")]; + tensor var_1056 = const()[name = tensor("op_1056"), val = tensor([1, 1, 1, 1])]; + tensor position2_to_fp16_dtype_0 = const()[name = tensor("position2_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor position2_to_fp16 = cast(dtype = position2_to_fp16_dtype_0, x = position2)[name = tensor("cast_543")]; + tensor ts_17_cast_fp16 = reshape(shape = var_1056, x = position2_to_fp16)[name = tensor("ts_17_cast_fp16")]; + tensor var_1063 = const()[name = tensor("op_1063"), val = tensor([1, 1, 16, 32, 2])]; + tensor q_complex_5_cast_fp16 = reshape(shape = var_1063, x = q_13_cast_fp16)[name = tensor("q_complex_5_cast_fp16")]; + tensor var_1070 = const()[name = tensor("op_1070"), val = tensor([1, 1, 16, 32, 2])]; + tensor k_complex_5_cast_fp16 = reshape(shape = var_1070, x = k_9_cast_fp16)[name = tensor("k_complex_5_cast_fp16")]; + tensor var_1074_begin_0 = const()[name = tensor("op_1074_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_1074_end_0 = const()[name = tensor("op_1074_end_0"), val = tensor([1, 1, 16, 32, 1])]; + tensor var_1074_end_mask_0 = const()[name = tensor("op_1074_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_1074_squeeze_mask_0 = const()[name = tensor("op_1074_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_1074_cast_fp16 = slice_by_index(begin = var_1074_begin_0, end = var_1074_end_0, end_mask = var_1074_end_mask_0, squeeze_mask = var_1074_squeeze_mask_0, x = q_complex_5_cast_fp16)[name = tensor("op_1074_cast_fp16")]; + tensor var_1082_begin_0 = const()[name = tensor("op_1082_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_1082_end_0 = const()[name = tensor("op_1082_end_0"), val = tensor([1, 1, 16, 32, 2])]; + tensor var_1082_end_mask_0 = const()[name = tensor("op_1082_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_1082_squeeze_mask_0 = const()[name = tensor("op_1082_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_1082_cast_fp16 = slice_by_index(begin = var_1082_begin_0, end = var_1082_end_0, end_mask = var_1082_end_mask_0, squeeze_mask = var_1082_squeeze_mask_0, x = q_complex_5_cast_fp16)[name = tensor("op_1082_cast_fp16")]; + tensor var_1090_begin_0 = const()[name = tensor("op_1090_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_1090_end_0 = const()[name = tensor("op_1090_end_0"), val = tensor([1, 1, 16, 32, 1])]; + tensor var_1090_end_mask_0 = const()[name = tensor("op_1090_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_1090_squeeze_mask_0 = const()[name = tensor("op_1090_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_1090_cast_fp16 = slice_by_index(begin = var_1090_begin_0, end = var_1090_end_0, end_mask = var_1090_end_mask_0, squeeze_mask = var_1090_squeeze_mask_0, x = k_complex_5_cast_fp16)[name = tensor("op_1090_cast_fp16")]; + tensor var_1098_begin_0 = const()[name = tensor("op_1098_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_1098_end_0 = const()[name = tensor("op_1098_end_0"), val = tensor([1, 1, 16, 32, 2])]; + tensor var_1098_end_mask_0 = const()[name = tensor("op_1098_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_1098_squeeze_mask_0 = const()[name = tensor("op_1098_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_1098_cast_fp16 = slice_by_index(begin = var_1098_begin_0, end = var_1098_end_0, end_mask = var_1098_end_mask_0, squeeze_mask = var_1098_squeeze_mask_0, x = k_complex_5_cast_fp16)[name = tensor("op_1098_cast_fp16")]; + tensor freqs_5_to_fp16 = const()[name = tensor("freqs_5_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(57777024)))]; + tensor var_1104_cast_fp16 = mul(x = freqs_5_to_fp16, y = ts_17_cast_fp16)[name = tensor("op_1104_cast_fp16")]; + tensor rotr_5_cast_fp16 = cos(x = var_1104_cast_fp16)[name = tensor("rotr_5_cast_fp16")]; + tensor roti_5_cast_fp16 = sin(x = var_1104_cast_fp16)[name = tensor("roti_5_cast_fp16")]; + tensor var_1108_cast_fp16 = mul(x = var_1074_cast_fp16, y = rotr_5_cast_fp16)[name = tensor("op_1108_cast_fp16")]; + tensor var_1109_cast_fp16 = mul(x = var_1082_cast_fp16, y = roti_5_cast_fp16)[name = tensor("op_1109_cast_fp16")]; + tensor qor_9_cast_fp16 = sub(x = var_1108_cast_fp16, y = var_1109_cast_fp16)[name = tensor("qor_9_cast_fp16")]; + tensor var_1112_cast_fp16 = mul(x = var_1074_cast_fp16, y = roti_5_cast_fp16)[name = tensor("op_1112_cast_fp16")]; + tensor var_1113_cast_fp16 = mul(x = var_1082_cast_fp16, y = rotr_5_cast_fp16)[name = tensor("op_1113_cast_fp16")]; + tensor qoi_9_cast_fp16 = add(x = var_1112_cast_fp16, y = var_1113_cast_fp16)[name = tensor("qoi_9_cast_fp16")]; + tensor var_1116_cast_fp16 = mul(x = var_1090_cast_fp16, y = rotr_5_cast_fp16)[name = tensor("op_1116_cast_fp16")]; + tensor var_1117_cast_fp16 = mul(x = var_1098_cast_fp16, y = roti_5_cast_fp16)[name = tensor("op_1117_cast_fp16")]; + tensor kor_9_cast_fp16 = sub(x = var_1116_cast_fp16, y = var_1117_cast_fp16)[name = tensor("kor_9_cast_fp16")]; + tensor var_1120_cast_fp16 = mul(x = var_1090_cast_fp16, y = roti_5_cast_fp16)[name = tensor("op_1120_cast_fp16")]; + tensor var_1121_cast_fp16 = mul(x = var_1098_cast_fp16, y = rotr_5_cast_fp16)[name = tensor("op_1121_cast_fp16")]; + tensor koi_9_cast_fp16 = add(x = var_1120_cast_fp16, y = var_1121_cast_fp16)[name = tensor("koi_9_cast_fp16")]; + tensor qo_5_axis_0 = const()[name = tensor("qo_5_axis_0"), val = tensor(-1)]; + tensor qo_5_cast_fp16 = stack(axis = qo_5_axis_0, values = (qor_9_cast_fp16, qoi_9_cast_fp16))[name = tensor("qo_5_cast_fp16")]; + tensor ko_5_axis_0 = const()[name = tensor("ko_5_axis_0"), val = tensor(-1)]; + tensor ko_5_cast_fp16 = stack(axis = ko_5_axis_0, values = (kor_9_cast_fp16, koi_9_cast_fp16))[name = tensor("ko_5_cast_fp16")]; + tensor var_1154 = const()[name = tensor("op_1154"), val = tensor([1, 1, 16, 64])]; + tensor q_15_cast_fp16 = reshape(shape = var_1154, x = qo_5_cast_fp16)[name = tensor("q_15_cast_fp16")]; + tensor var_1160 = const()[name = tensor("op_1160"), val = tensor([1, 1, 16, 64])]; + tensor k_11_cast_fp16 = reshape(shape = var_1160, x = ko_5_cast_fp16)[name = tensor("k_11_cast_fp16")]; + tensor _inversed_1185_y_0_to_fp16 = const()[name = tensor("_inversed_1185_y_0_to_fp16"), val = tensor(0x1p-9)]; + tensor _inversed_1185_cast_fp16 = mul(x = ts_17_cast_fp16, y = _inversed_1185_y_0_to_fp16)[name = tensor("_inversed_1185_cast_fp16")]; + tensor var_1186_cast_fp16 = floor(x = _inversed_1185_cast_fp16)[name = tensor("op_1186_cast_fp16")]; + tensor var_1187_to_fp16 = const()[name = tensor("op_1187_to_fp16"), val = tensor(0x1p+9)]; + tensor var_1188_cast_fp16 = mul(x = var_1186_cast_fp16, y = var_1187_to_fp16)[name = tensor("op_1188_cast_fp16")]; + tensor write_indices_float_11_cast_fp16 = sub(x = ts_17_cast_fp16, y = var_1188_cast_fp16)[name = tensor("write_indices_float_11_cast_fp16")]; + tensor var_1195_dtype_0 = const()[name = tensor("op_1195_dtype_0"), val = tensor("int32")]; + tensor write_indices_5_reps_0 = const()[name = tensor("write_indices_5_reps_0"), val = tensor([1, 1, 16, 64])]; + tensor write_indices_float_11_cast_fp16_to_int32 = cast(dtype = var_1195_dtype_0, x = write_indices_float_11_cast_fp16)[name = tensor("cast_542")]; + tensor write_indices_5 = tile(reps = write_indices_5_reps_0, x = write_indices_float_11_cast_fp16_to_int32)[name = tensor("write_indices_5")]; + tensor var_1205_begin_0 = const()[name = tensor("op_1205_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_1205_end_0 = const()[name = tensor("op_1205_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor var_1205_end_mask_0 = const()[name = tensor("op_1205_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_1205_squeeze_mask_0 = const()[name = tensor("op_1205_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor cache2_to_fp16_dtype_0 = const()[name = tensor("cache2_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor cache2_to_fp16 = cast(dtype = cache2_to_fp16_dtype_0, x = cache2)[name = tensor("cast_541")]; + tensor var_1205_cast_fp16 = slice_by_index(begin = var_1205_begin_0, end = var_1205_end_0, end_mask = var_1205_end_mask_0, squeeze_mask = var_1205_squeeze_mask_0, x = cache2_to_fp16)[name = tensor("op_1205_cast_fp16")]; + tensor var_1207_axis_0 = const()[name = tensor("op_1207_axis_0"), val = tensor(1)]; + tensor var_1207_mode_0 = const()[name = tensor("op_1207_mode_0"), val = tensor("update")]; + tensor var_1207_validate_indices_0 = const()[name = tensor("op_1207_validate_indices_0"), val = tensor(false)]; + tensor var_1207_cast_fp16 = scatter_along_axis(axis = var_1207_axis_0, data = var_1205_cast_fp16, indices = write_indices_5, mode = var_1207_mode_0, updates = k_11_cast_fp16, validate_indices = var_1207_validate_indices_0)[name = tensor("op_1207_cast_fp16")]; + tensor concat_16 = const()[name = tensor("concat_16"), val = tensor([0, 0, 0, 0, 0])]; + tensor concat_17 = const()[name = tensor("concat_17"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_5_internal_tensor_assign_1_stride_0 = const()[name = tensor("new_cache_5_internal_tensor_assign_1_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_5_internal_tensor_assign_1_begin_mask_0 = const()[name = tensor("new_cache_5_internal_tensor_assign_1_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_5_internal_tensor_assign_1_end_mask_0 = const()[name = tensor("new_cache_5_internal_tensor_assign_1_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_5_internal_tensor_assign_1_squeeze_mask_0 = const()[name = tensor("new_cache_5_internal_tensor_assign_1_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_52 = const()[name = tensor("shape_52"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_4 = const()[name = tensor("reduce_prod_4"), val = tensor(1048576)]; + tensor range_1d_4_start_0 = const()[name = tensor("range_1d_4_start_0"), val = tensor(0)]; + tensor range_1d_4_step_0 = const()[name = tensor("range_1d_4_step_0"), val = tensor(1)]; + tensor range_1d_4 = range_1d(end = reduce_prod_4, start = range_1d_4_start_0, step = range_1d_4_step_0)[name = tensor("range_1d_4")]; + tensor reshape_20 = reshape(shape = shape_52, x = range_1d_4)[name = tensor("reshape_20")]; + tensor slice_by_index_4 = slice_by_index(begin = concat_16, begin_mask = new_cache_5_internal_tensor_assign_1_begin_mask_0, end = concat_17, end_mask = new_cache_5_internal_tensor_assign_1_end_mask_0, squeeze_mask = new_cache_5_internal_tensor_assign_1_squeeze_mask_0, stride = new_cache_5_internal_tensor_assign_1_stride_0, x = reshape_20)[name = tensor("slice_by_index_4")]; + tensor reshape_21_shape_0 = const()[name = tensor("reshape_21_shape_0"), val = tensor([-1])]; + tensor reshape_21 = reshape(shape = reshape_21_shape_0, x = slice_by_index_4)[name = tensor("reshape_21")]; + tensor reshape_22_shape_0 = const()[name = tensor("reshape_22_shape_0"), val = tensor([-1])]; + tensor reshape_22_cast_fp16 = reshape(shape = reshape_22_shape_0, x = var_1207_cast_fp16)[name = tensor("reshape_22_cast_fp16")]; + tensor reshape_23_shape_0 = const()[name = tensor("reshape_23_shape_0"), val = tensor([-1])]; + tensor reshape_23_cast_fp16 = reshape(shape = reshape_23_shape_0, x = cache2_to_fp16)[name = tensor("reshape_23_cast_fp16")]; + tensor scatter_4_mode_0 = const()[name = tensor("scatter_4_mode_0"), val = tensor("update")]; + tensor scatter_4_axis_0 = const()[name = tensor("scatter_4_axis_0"), val = tensor(0)]; + tensor scatter_4_validate_indices_0 = const()[name = tensor("scatter_4_validate_indices_0"), val = tensor(false)]; + tensor scatter_4_cast_fp16 = scatter(axis = scatter_4_axis_0, data = reshape_23_cast_fp16, indices = reshape_21, mode = scatter_4_mode_0, updates = reshape_22_cast_fp16, validate_indices = scatter_4_validate_indices_0)[name = tensor("scatter_4_cast_fp16")]; + tensor reshape_24_cast_fp16 = reshape(shape = shape_52, x = scatter_4_cast_fp16)[name = tensor("reshape_24_cast_fp16")]; + tensor var_1215_begin_0 = const()[name = tensor("op_1215_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor var_1215_end_0 = const()[name = tensor("op_1215_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor var_1215_end_mask_0 = const()[name = tensor("op_1215_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_1215_squeeze_mask_0 = const()[name = tensor("op_1215_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor var_1215_cast_fp16 = slice_by_index(begin = var_1215_begin_0, end = var_1215_end_0, end_mask = var_1215_end_mask_0, squeeze_mask = var_1215_squeeze_mask_0, x = reshape_24_cast_fp16)[name = tensor("op_1215_cast_fp16")]; + tensor var_1217_axis_0 = const()[name = tensor("op_1217_axis_0"), val = tensor(1)]; + tensor var_1217_mode_0 = const()[name = tensor("op_1217_mode_0"), val = tensor("update")]; + tensor var_1217_validate_indices_0 = const()[name = tensor("op_1217_validate_indices_0"), val = tensor(false)]; + tensor var_1217_cast_fp16 = scatter_along_axis(axis = var_1217_axis_0, data = var_1215_cast_fp16, indices = write_indices_5, mode = var_1217_mode_0, updates = v_5_cast_fp16, validate_indices = var_1217_validate_indices_0)[name = tensor("op_1217_cast_fp16")]; + tensor concat_18 = const()[name = tensor("concat_18"), val = tensor([1, 0, 0, 0, 0])]; + tensor concat_19 = const()[name = tensor("concat_19"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_5_internal_tensor_assign_2_stride_0 = const()[name = tensor("new_cache_5_internal_tensor_assign_2_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_5_internal_tensor_assign_2_begin_mask_0 = const()[name = tensor("new_cache_5_internal_tensor_assign_2_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_5_internal_tensor_assign_2_end_mask_0 = const()[name = tensor("new_cache_5_internal_tensor_assign_2_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_5_internal_tensor_assign_2_squeeze_mask_0 = const()[name = tensor("new_cache_5_internal_tensor_assign_2_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_53 = const()[name = tensor("shape_53"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_5 = const()[name = tensor("reduce_prod_5"), val = tensor(1048576)]; + tensor range_1d_5_start_0 = const()[name = tensor("range_1d_5_start_0"), val = tensor(0)]; + tensor range_1d_5_step_0 = const()[name = tensor("range_1d_5_step_0"), val = tensor(1)]; + tensor range_1d_5 = range_1d(end = reduce_prod_5, start = range_1d_5_start_0, step = range_1d_5_step_0)[name = tensor("range_1d_5")]; + tensor reshape_25 = reshape(shape = shape_53, x = range_1d_5)[name = tensor("reshape_25")]; + tensor slice_by_index_5 = slice_by_index(begin = concat_18, begin_mask = new_cache_5_internal_tensor_assign_2_begin_mask_0, end = concat_19, end_mask = new_cache_5_internal_tensor_assign_2_end_mask_0, squeeze_mask = new_cache_5_internal_tensor_assign_2_squeeze_mask_0, stride = new_cache_5_internal_tensor_assign_2_stride_0, x = reshape_25)[name = tensor("slice_by_index_5")]; + tensor reshape_26_shape_0 = const()[name = tensor("reshape_26_shape_0"), val = tensor([-1])]; + tensor reshape_26 = reshape(shape = reshape_26_shape_0, x = slice_by_index_5)[name = tensor("reshape_26")]; + tensor reshape_27_shape_0 = const()[name = tensor("reshape_27_shape_0"), val = tensor([-1])]; + tensor reshape_27_cast_fp16 = reshape(shape = reshape_27_shape_0, x = var_1217_cast_fp16)[name = tensor("reshape_27_cast_fp16")]; + tensor reshape_28_shape_0 = const()[name = tensor("reshape_28_shape_0"), val = tensor([-1])]; + tensor reshape_28_cast_fp16 = reshape(shape = reshape_28_shape_0, x = reshape_24_cast_fp16)[name = tensor("reshape_28_cast_fp16")]; + tensor scatter_5_mode_0 = const()[name = tensor("scatter_5_mode_0"), val = tensor("update")]; + tensor scatter_5_axis_0 = const()[name = tensor("scatter_5_axis_0"), val = tensor(0)]; + tensor scatter_5_validate_indices_0 = const()[name = tensor("scatter_5_validate_indices_0"), val = tensor(false)]; + tensor scatter_5_cast_fp16 = scatter(axis = scatter_5_axis_0, data = reshape_28_cast_fp16, indices = reshape_26, mode = scatter_5_mode_0, updates = reshape_27_cast_fp16, validate_indices = scatter_5_validate_indices_0)[name = tensor("scatter_5_cast_fp16")]; + tensor reshape_29_cast_fp16 = reshape(shape = shape_53, x = scatter_5_cast_fp16)[name = tensor("reshape_29_cast_fp16")]; + tensor reshape_29_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("reshape_29_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor keys_13_begin_0 = const()[name = tensor("keys_13_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor keys_13_end_0 = const()[name = tensor("keys_13_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor keys_13_end_mask_0 = const()[name = tensor("keys_13_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor keys_13_squeeze_mask_0 = const()[name = tensor("keys_13_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor keys_13_cast_fp16 = slice_by_index(begin = keys_13_begin_0, end = keys_13_end_0, end_mask = keys_13_end_mask_0, squeeze_mask = keys_13_squeeze_mask_0, x = reshape_29_cast_fp16)[name = tensor("keys_13_cast_fp16")]; + tensor values_13_begin_0 = const()[name = tensor("values_13_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor values_13_end_0 = const()[name = tensor("values_13_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor values_13_end_mask_0 = const()[name = tensor("values_13_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor values_13_squeeze_mask_0 = const()[name = tensor("values_13_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor values_13_cast_fp16 = slice_by_index(begin = values_13_begin_0, end = values_13_end_0, end_mask = values_13_end_mask_0, squeeze_mask = values_13_squeeze_mask_0, x = reshape_29_cast_fp16)[name = tensor("values_13_cast_fp16")]; + tensor var_1229_cast_fp16 = not_equal(x = keys_13_cast_fp16, y = keys_13_cast_fp16)[name = tensor("op_1229_cast_fp16")]; + tensor keys_15_cast_fp16 = select(a = var_463_to_fp16, b = keys_13_cast_fp16, cond = var_1229_cast_fp16)[name = tensor("keys_15_cast_fp16")]; + tensor var_1237_cast_fp16 = not_equal(x = values_13_cast_fp16, y = values_13_cast_fp16)[name = tensor("op_1237_cast_fp16")]; + tensor values_15_cast_fp16 = select(a = var_463_to_fp16, b = values_13_cast_fp16, cond = var_1237_cast_fp16)[name = tensor("values_15_cast_fp16")]; + tensor var_1261 = const()[name = tensor("op_1261"), val = tensor([0, 2, 1, 3])]; + tensor var_1277 = const()[name = tensor("op_1277"), val = tensor([1, 1, 1])]; + tensor var_1278_cast_fp16 = reshape(shape = var_1277, x = position2_to_fp16)[name = tensor("op_1278_cast_fp16")]; + tensor var_1298_to_fp16 = const()[name = tensor("op_1298_to_fp16"), val = tensor(0x1p+0)]; + tensor valid_len_5_cast_fp16 = add(x = var_1278_cast_fp16, y = var_1298_to_fp16)[name = tensor("valid_len_5_cast_fp16")]; + tensor valid_mask_5_cast_fp16 = less(x = k_positions_1_promoted_to_fp16, y = valid_len_5_cast_fp16)[name = tensor("valid_mask_5_cast_fp16")]; + tensor causal_mask_5_cast_fp16 = less_equal(x = k_positions_1_promoted_to_fp16, y = var_1278_cast_fp16)[name = tensor("causal_mask_5_cast_fp16")]; + tensor attn_mask_9 = logical_and(x = valid_mask_5_cast_fp16, y = causal_mask_5_cast_fp16)[name = tensor("attn_mask_9")]; + tensor attn_mask_11_axes_0 = const()[name = tensor("attn_mask_11_axes_0"), val = tensor([1])]; + tensor attn_mask_11 = expand_dims(axes = attn_mask_11_axes_0, x = attn_mask_9)[name = tensor("attn_mask_11")]; + tensor var_1316_transpose_x_0 = const()[name = tensor("op_1316_transpose_x_0"), val = tensor(false)]; + tensor var_1316_transpose_y_0 = const()[name = tensor("op_1316_transpose_y_0"), val = tensor(false)]; + tensor transpose_76_perm_0 = const()[name = tensor("transpose_76_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_77_perm_0 = const()[name = tensor("transpose_77_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_77 = transpose(perm = transpose_77_perm_0, x = keys_15_cast_fp16)[name = tensor("transpose_205")]; + tensor transpose_76 = transpose(perm = transpose_76_perm_0, x = q_15_cast_fp16)[name = tensor("transpose_206")]; + tensor var_1316_cast_fp16 = matmul(transpose_x = var_1316_transpose_x_0, transpose_y = var_1316_transpose_y_0, x = transpose_76, y = transpose_77)[name = tensor("op_1316_cast_fp16")]; + tensor var_1310_to_fp16 = const()[name = tensor("op_1310_to_fp16"), val = tensor([0x1p-3])]; + tensor attn_weights_13_cast_fp16 = mul(x = var_1316_cast_fp16, y = var_1310_to_fp16)[name = tensor("attn_weights_13_cast_fp16")]; + tensor var_1318 = logical_not(x = attn_mask_11)[name = tensor("op_1318")]; + tensor var_1319_to_fp16 = const()[name = tensor("op_1319_to_fp16"), val = tensor(-inf)]; + tensor attn_weights_15_cast_fp16 = select(a = var_1319_to_fp16, b = attn_weights_13_cast_fp16, cond = var_1318)[name = tensor("attn_weights_15_cast_fp16")]; + tensor var_1321 = const()[name = tensor("op_1321"), val = tensor(-1)]; + tensor attn_weights_17_cast_fp16 = softmax(axis = var_1321, x = attn_weights_15_cast_fp16)[name = tensor("attn_weights_17_cast_fp16")]; + tensor attn_output_5_transpose_x_0 = const()[name = tensor("attn_output_5_transpose_x_0"), val = tensor(false)]; + tensor attn_output_5_transpose_y_0 = const()[name = tensor("attn_output_5_transpose_y_0"), val = tensor(false)]; + tensor values_17_cast_fp16 = transpose(perm = var_1261, x = values_15_cast_fp16)[name = tensor("transpose_207")]; + tensor attn_output_5_cast_fp16 = matmul(transpose_x = attn_output_5_transpose_x_0, transpose_y = attn_output_5_transpose_y_0, x = attn_weights_17_cast_fp16, y = values_17_cast_fp16)[name = tensor("attn_output_5_cast_fp16")]; + tensor var_1329 = const()[name = tensor("op_1329"), val = tensor([0, 2, 1, 3])]; + tensor var_1334 = const()[name = tensor("op_1334"), val = tensor([1, 1, 1024])]; + tensor var_1330_cast_fp16 = transpose(perm = var_1329, x = attn_output_5_cast_fp16)[name = tensor("transpose_204")]; + tensor input_31_cast_fp16 = reshape(shape = var_1334, x = var_1330_cast_fp16)[name = tensor("input_31_cast_fp16")]; + tensor attn2_out_proj_weight_to_fp16 = const()[name = tensor("attn2_out_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(57777152)))]; + tensor linear_10_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = attn2_out_proj_weight_to_fp16, x = input_31_cast_fp16)[name = tensor("linear_10_cast_fp16")]; + tensor var_1340_to_fp16 = const()[name = tensor("op_1340_to_fp16"), val = tensor(0x1p+0)]; + tensor var_1341_cast_fp16 = add(x = position2_to_fp16, y = var_1340_to_fp16)[name = tensor("op_1341_cast_fp16")]; + tensor var_1341_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_1341_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor input_33_cast_fp16 = add(x = input_27_cast_fp16, y = linear_10_cast_fp16)[name = tensor("input_33_cast_fp16")]; + tensor input_35_axes_0 = const()[name = tensor("input_35_axes_0"), val = tensor([-1])]; + tensor norm2_2_weight_to_fp16 = const()[name = tensor("norm2_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(59874368)))]; + tensor norm2_2_bias_to_fp16 = const()[name = tensor("norm2_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(59876480)))]; + tensor var_1345_to_fp16 = const()[name = tensor("op_1345_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_35_cast_fp16 = layer_norm(axes = input_35_axes_0, beta = norm2_2_bias_to_fp16, epsilon = var_1345_to_fp16, gamma = norm2_2_weight_to_fp16, x = input_33_cast_fp16)[name = tensor("input_35_cast_fp16")]; + tensor linear2_1_weight_to_fp16 = const()[name = tensor("linear2_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(59878592)))]; + tensor linear_11_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = linear2_1_weight_to_fp16, x = input_35_cast_fp16)[name = tensor("linear_11_cast_fp16")]; + tensor input_37_mode_0 = const()[name = tensor("input_37_mode_0"), val = tensor("EXACT")]; + tensor input_37_cast_fp16 = gelu(mode = input_37_mode_0, x = linear_11_cast_fp16)[name = tensor("input_37_cast_fp16")]; + tensor linear2_2_weight_to_fp16 = const()[name = tensor("linear2_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(68267264)))]; + tensor linear_12_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = linear2_2_weight_to_fp16, x = input_37_cast_fp16)[name = tensor("linear_12_cast_fp16")]; + tensor input_39_cast_fp16 = add(x = input_33_cast_fp16, y = linear_12_cast_fp16)[name = tensor("input_39_cast_fp16")]; + tensor input_41_axes_0 = const()[name = tensor("input_41_axes_0"), val = tensor([-1])]; + tensor norm3_1_weight_to_fp16 = const()[name = tensor("norm3_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(76655936)))]; + tensor norm3_1_bias_to_fp16 = const()[name = tensor("norm3_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(76658048)))]; + tensor var_1362_to_fp16 = const()[name = tensor("op_1362_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_41_cast_fp16 = layer_norm(axes = input_41_axes_0, beta = norm3_1_bias_to_fp16, epsilon = var_1362_to_fp16, gamma = norm3_1_weight_to_fp16, x = input_39_cast_fp16)[name = tensor("input_41_cast_fp16")]; + tensor attn3_in_proj_weight_to_fp16 = const()[name = tensor("attn3_in_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(76660160)))]; + tensor linear_13_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = attn3_in_proj_weight_to_fp16, x = input_41_cast_fp16)[name = tensor("linear_13_cast_fp16")]; + tensor var_1376 = const()[name = tensor("op_1376"), val = tensor([1, 1, 3, 16, 64])]; + tensor qkv_7_cast_fp16 = reshape(shape = var_1376, x = linear_13_cast_fp16)[name = tensor("qkv_7_cast_fp16")]; + tensor q_19_begin_0 = const()[name = tensor("q_19_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor q_19_end_0 = const()[name = tensor("q_19_end_0"), val = tensor([1, 1, 1, 16, 64])]; + tensor q_19_end_mask_0 = const()[name = tensor("q_19_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor q_19_squeeze_mask_0 = const()[name = tensor("q_19_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor q_19_cast_fp16 = slice_by_index(begin = q_19_begin_0, end = q_19_end_0, end_mask = q_19_end_mask_0, squeeze_mask = q_19_squeeze_mask_0, x = qkv_7_cast_fp16)[name = tensor("q_19_cast_fp16")]; + tensor k_13_begin_0 = const()[name = tensor("k_13_begin_0"), val = tensor([0, 0, 1, 0, 0])]; + tensor k_13_end_0 = const()[name = tensor("k_13_end_0"), val = tensor([1, 1, 2, 16, 64])]; + tensor k_13_end_mask_0 = const()[name = tensor("k_13_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor k_13_squeeze_mask_0 = const()[name = tensor("k_13_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor k_13_cast_fp16 = slice_by_index(begin = k_13_begin_0, end = k_13_end_0, end_mask = k_13_end_mask_0, squeeze_mask = k_13_squeeze_mask_0, x = qkv_7_cast_fp16)[name = tensor("k_13_cast_fp16")]; + tensor v_7_begin_0 = const()[name = tensor("v_7_begin_0"), val = tensor([0, 0, 2, 0, 0])]; + tensor v_7_end_0 = const()[name = tensor("v_7_end_0"), val = tensor([1, 1, 3, 16, 64])]; + tensor v_7_end_mask_0 = const()[name = tensor("v_7_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor v_7_squeeze_mask_0 = const()[name = tensor("v_7_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor v_7_cast_fp16 = slice_by_index(begin = v_7_begin_0, end = v_7_end_0, end_mask = v_7_end_mask_0, squeeze_mask = v_7_squeeze_mask_0, x = qkv_7_cast_fp16)[name = tensor("v_7_cast_fp16")]; + tensor var_1442 = const()[name = tensor("op_1442"), val = tensor([1, 1, 1, 1])]; + tensor position3_to_fp16_dtype_0 = const()[name = tensor("position3_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor position3_to_fp16 = cast(dtype = position3_to_fp16_dtype_0, x = position3)[name = tensor("cast_538")]; + tensor ts_23_cast_fp16 = reshape(shape = var_1442, x = position3_to_fp16)[name = tensor("ts_23_cast_fp16")]; + tensor var_1449 = const()[name = tensor("op_1449"), val = tensor([1, 1, 16, 32, 2])]; + tensor q_complex_7_cast_fp16 = reshape(shape = var_1449, x = q_19_cast_fp16)[name = tensor("q_complex_7_cast_fp16")]; + tensor var_1456 = const()[name = tensor("op_1456"), val = tensor([1, 1, 16, 32, 2])]; + tensor k_complex_7_cast_fp16 = reshape(shape = var_1456, x = k_13_cast_fp16)[name = tensor("k_complex_7_cast_fp16")]; + tensor var_1460_begin_0 = const()[name = tensor("op_1460_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_1460_end_0 = const()[name = tensor("op_1460_end_0"), val = tensor([1, 1, 16, 32, 1])]; + tensor var_1460_end_mask_0 = const()[name = tensor("op_1460_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_1460_squeeze_mask_0 = const()[name = tensor("op_1460_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_1460_cast_fp16 = slice_by_index(begin = var_1460_begin_0, end = var_1460_end_0, end_mask = var_1460_end_mask_0, squeeze_mask = var_1460_squeeze_mask_0, x = q_complex_7_cast_fp16)[name = tensor("op_1460_cast_fp16")]; + tensor var_1468_begin_0 = const()[name = tensor("op_1468_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_1468_end_0 = const()[name = tensor("op_1468_end_0"), val = tensor([1, 1, 16, 32, 2])]; + tensor var_1468_end_mask_0 = const()[name = tensor("op_1468_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_1468_squeeze_mask_0 = const()[name = tensor("op_1468_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_1468_cast_fp16 = slice_by_index(begin = var_1468_begin_0, end = var_1468_end_0, end_mask = var_1468_end_mask_0, squeeze_mask = var_1468_squeeze_mask_0, x = q_complex_7_cast_fp16)[name = tensor("op_1468_cast_fp16")]; + tensor var_1476_begin_0 = const()[name = tensor("op_1476_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_1476_end_0 = const()[name = tensor("op_1476_end_0"), val = tensor([1, 1, 16, 32, 1])]; + tensor var_1476_end_mask_0 = const()[name = tensor("op_1476_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_1476_squeeze_mask_0 = const()[name = tensor("op_1476_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_1476_cast_fp16 = slice_by_index(begin = var_1476_begin_0, end = var_1476_end_0, end_mask = var_1476_end_mask_0, squeeze_mask = var_1476_squeeze_mask_0, x = k_complex_7_cast_fp16)[name = tensor("op_1476_cast_fp16")]; + tensor var_1484_begin_0 = const()[name = tensor("op_1484_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_1484_end_0 = const()[name = tensor("op_1484_end_0"), val = tensor([1, 1, 16, 32, 2])]; + tensor var_1484_end_mask_0 = const()[name = tensor("op_1484_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_1484_squeeze_mask_0 = const()[name = tensor("op_1484_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_1484_cast_fp16 = slice_by_index(begin = var_1484_begin_0, end = var_1484_end_0, end_mask = var_1484_end_mask_0, squeeze_mask = var_1484_squeeze_mask_0, x = k_complex_7_cast_fp16)[name = tensor("op_1484_cast_fp16")]; + tensor freqs_7_to_fp16 = const()[name = tensor("freqs_7_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(82951680)))]; + tensor var_1490_cast_fp16 = mul(x = freqs_7_to_fp16, y = ts_23_cast_fp16)[name = tensor("op_1490_cast_fp16")]; + tensor rotr_7_cast_fp16 = cos(x = var_1490_cast_fp16)[name = tensor("rotr_7_cast_fp16")]; + tensor roti_7_cast_fp16 = sin(x = var_1490_cast_fp16)[name = tensor("roti_7_cast_fp16")]; + tensor var_1494_cast_fp16 = mul(x = var_1460_cast_fp16, y = rotr_7_cast_fp16)[name = tensor("op_1494_cast_fp16")]; + tensor var_1495_cast_fp16 = mul(x = var_1468_cast_fp16, y = roti_7_cast_fp16)[name = tensor("op_1495_cast_fp16")]; + tensor qor_13_cast_fp16 = sub(x = var_1494_cast_fp16, y = var_1495_cast_fp16)[name = tensor("qor_13_cast_fp16")]; + tensor var_1498_cast_fp16 = mul(x = var_1460_cast_fp16, y = roti_7_cast_fp16)[name = tensor("op_1498_cast_fp16")]; + tensor var_1499_cast_fp16 = mul(x = var_1468_cast_fp16, y = rotr_7_cast_fp16)[name = tensor("op_1499_cast_fp16")]; + tensor qoi_13_cast_fp16 = add(x = var_1498_cast_fp16, y = var_1499_cast_fp16)[name = tensor("qoi_13_cast_fp16")]; + tensor var_1502_cast_fp16 = mul(x = var_1476_cast_fp16, y = rotr_7_cast_fp16)[name = tensor("op_1502_cast_fp16")]; + tensor var_1503_cast_fp16 = mul(x = var_1484_cast_fp16, y = roti_7_cast_fp16)[name = tensor("op_1503_cast_fp16")]; + tensor kor_13_cast_fp16 = sub(x = var_1502_cast_fp16, y = var_1503_cast_fp16)[name = tensor("kor_13_cast_fp16")]; + tensor var_1506_cast_fp16 = mul(x = var_1476_cast_fp16, y = roti_7_cast_fp16)[name = tensor("op_1506_cast_fp16")]; + tensor var_1507_cast_fp16 = mul(x = var_1484_cast_fp16, y = rotr_7_cast_fp16)[name = tensor("op_1507_cast_fp16")]; + tensor koi_13_cast_fp16 = add(x = var_1506_cast_fp16, y = var_1507_cast_fp16)[name = tensor("koi_13_cast_fp16")]; + tensor qo_7_axis_0 = const()[name = tensor("qo_7_axis_0"), val = tensor(-1)]; + tensor qo_7_cast_fp16 = stack(axis = qo_7_axis_0, values = (qor_13_cast_fp16, qoi_13_cast_fp16))[name = tensor("qo_7_cast_fp16")]; + tensor ko_7_axis_0 = const()[name = tensor("ko_7_axis_0"), val = tensor(-1)]; + tensor ko_7_cast_fp16 = stack(axis = ko_7_axis_0, values = (kor_13_cast_fp16, koi_13_cast_fp16))[name = tensor("ko_7_cast_fp16")]; + tensor var_1540 = const()[name = tensor("op_1540"), val = tensor([1, 1, 16, 64])]; + tensor q_21_cast_fp16 = reshape(shape = var_1540, x = qo_7_cast_fp16)[name = tensor("q_21_cast_fp16")]; + tensor var_1546 = const()[name = tensor("op_1546"), val = tensor([1, 1, 16, 64])]; + tensor k_15_cast_fp16 = reshape(shape = var_1546, x = ko_7_cast_fp16)[name = tensor("k_15_cast_fp16")]; + tensor _inversed_1571_y_0_to_fp16 = const()[name = tensor("_inversed_1571_y_0_to_fp16"), val = tensor(0x1p-9)]; + tensor _inversed_1571_cast_fp16 = mul(x = ts_23_cast_fp16, y = _inversed_1571_y_0_to_fp16)[name = tensor("_inversed_1571_cast_fp16")]; + tensor var_1572_cast_fp16 = floor(x = _inversed_1571_cast_fp16)[name = tensor("op_1572_cast_fp16")]; + tensor var_1573_to_fp16 = const()[name = tensor("op_1573_to_fp16"), val = tensor(0x1p+9)]; + tensor var_1574_cast_fp16 = mul(x = var_1572_cast_fp16, y = var_1573_to_fp16)[name = tensor("op_1574_cast_fp16")]; + tensor write_indices_float_15_cast_fp16 = sub(x = ts_23_cast_fp16, y = var_1574_cast_fp16)[name = tensor("write_indices_float_15_cast_fp16")]; + tensor var_1581_dtype_0 = const()[name = tensor("op_1581_dtype_0"), val = tensor("int32")]; + tensor write_indices_7_reps_0 = const()[name = tensor("write_indices_7_reps_0"), val = tensor([1, 1, 16, 64])]; + tensor write_indices_float_15_cast_fp16_to_int32 = cast(dtype = var_1581_dtype_0, x = write_indices_float_15_cast_fp16)[name = tensor("cast_537")]; + tensor write_indices_7 = tile(reps = write_indices_7_reps_0, x = write_indices_float_15_cast_fp16_to_int32)[name = tensor("write_indices_7")]; + tensor var_1591_begin_0 = const()[name = tensor("op_1591_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_1591_end_0 = const()[name = tensor("op_1591_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor var_1591_end_mask_0 = const()[name = tensor("op_1591_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_1591_squeeze_mask_0 = const()[name = tensor("op_1591_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor cache3_to_fp16_dtype_0 = const()[name = tensor("cache3_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor cache3_to_fp16 = cast(dtype = cache3_to_fp16_dtype_0, x = cache3)[name = tensor("cast_536")]; + tensor var_1591_cast_fp16 = slice_by_index(begin = var_1591_begin_0, end = var_1591_end_0, end_mask = var_1591_end_mask_0, squeeze_mask = var_1591_squeeze_mask_0, x = cache3_to_fp16)[name = tensor("op_1591_cast_fp16")]; + tensor var_1593_axis_0 = const()[name = tensor("op_1593_axis_0"), val = tensor(1)]; + tensor var_1593_mode_0 = const()[name = tensor("op_1593_mode_0"), val = tensor("update")]; + tensor var_1593_validate_indices_0 = const()[name = tensor("op_1593_validate_indices_0"), val = tensor(false)]; + tensor var_1593_cast_fp16 = scatter_along_axis(axis = var_1593_axis_0, data = var_1591_cast_fp16, indices = write_indices_7, mode = var_1593_mode_0, updates = k_15_cast_fp16, validate_indices = var_1593_validate_indices_0)[name = tensor("op_1593_cast_fp16")]; + tensor concat_23 = const()[name = tensor("concat_23"), val = tensor([0, 0, 0, 0, 0])]; + tensor concat_24 = const()[name = tensor("concat_24"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_7_internal_tensor_assign_1_stride_0 = const()[name = tensor("new_cache_7_internal_tensor_assign_1_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_7_internal_tensor_assign_1_begin_mask_0 = const()[name = tensor("new_cache_7_internal_tensor_assign_1_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_7_internal_tensor_assign_1_end_mask_0 = const()[name = tensor("new_cache_7_internal_tensor_assign_1_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_7_internal_tensor_assign_1_squeeze_mask_0 = const()[name = tensor("new_cache_7_internal_tensor_assign_1_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_54 = const()[name = tensor("shape_54"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_6 = const()[name = tensor("reduce_prod_6"), val = tensor(1048576)]; + tensor range_1d_6_start_0 = const()[name = tensor("range_1d_6_start_0"), val = tensor(0)]; + tensor range_1d_6_step_0 = const()[name = tensor("range_1d_6_step_0"), val = tensor(1)]; + tensor range_1d_6 = range_1d(end = reduce_prod_6, start = range_1d_6_start_0, step = range_1d_6_step_0)[name = tensor("range_1d_6")]; + tensor reshape_30 = reshape(shape = shape_54, x = range_1d_6)[name = tensor("reshape_30")]; + tensor slice_by_index_6 = slice_by_index(begin = concat_23, begin_mask = new_cache_7_internal_tensor_assign_1_begin_mask_0, end = concat_24, end_mask = new_cache_7_internal_tensor_assign_1_end_mask_0, squeeze_mask = new_cache_7_internal_tensor_assign_1_squeeze_mask_0, stride = new_cache_7_internal_tensor_assign_1_stride_0, x = reshape_30)[name = tensor("slice_by_index_6")]; + tensor reshape_31_shape_0 = const()[name = tensor("reshape_31_shape_0"), val = tensor([-1])]; + tensor reshape_31 = reshape(shape = reshape_31_shape_0, x = slice_by_index_6)[name = tensor("reshape_31")]; + tensor reshape_32_shape_0 = const()[name = tensor("reshape_32_shape_0"), val = tensor([-1])]; + tensor reshape_32_cast_fp16 = reshape(shape = reshape_32_shape_0, x = var_1593_cast_fp16)[name = tensor("reshape_32_cast_fp16")]; + tensor reshape_33_shape_0 = const()[name = tensor("reshape_33_shape_0"), val = tensor([-1])]; + tensor reshape_33_cast_fp16 = reshape(shape = reshape_33_shape_0, x = cache3_to_fp16)[name = tensor("reshape_33_cast_fp16")]; + tensor scatter_6_mode_0 = const()[name = tensor("scatter_6_mode_0"), val = tensor("update")]; + tensor scatter_6_axis_0 = const()[name = tensor("scatter_6_axis_0"), val = tensor(0)]; + tensor scatter_6_validate_indices_0 = const()[name = tensor("scatter_6_validate_indices_0"), val = tensor(false)]; + tensor scatter_6_cast_fp16 = scatter(axis = scatter_6_axis_0, data = reshape_33_cast_fp16, indices = reshape_31, mode = scatter_6_mode_0, updates = reshape_32_cast_fp16, validate_indices = scatter_6_validate_indices_0)[name = tensor("scatter_6_cast_fp16")]; + tensor reshape_34_cast_fp16 = reshape(shape = shape_54, x = scatter_6_cast_fp16)[name = tensor("reshape_34_cast_fp16")]; + tensor var_1601_begin_0 = const()[name = tensor("op_1601_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor var_1601_end_0 = const()[name = tensor("op_1601_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor var_1601_end_mask_0 = const()[name = tensor("op_1601_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_1601_squeeze_mask_0 = const()[name = tensor("op_1601_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor var_1601_cast_fp16 = slice_by_index(begin = var_1601_begin_0, end = var_1601_end_0, end_mask = var_1601_end_mask_0, squeeze_mask = var_1601_squeeze_mask_0, x = reshape_34_cast_fp16)[name = tensor("op_1601_cast_fp16")]; + tensor var_1603_axis_0 = const()[name = tensor("op_1603_axis_0"), val = tensor(1)]; + tensor var_1603_mode_0 = const()[name = tensor("op_1603_mode_0"), val = tensor("update")]; + tensor var_1603_validate_indices_0 = const()[name = tensor("op_1603_validate_indices_0"), val = tensor(false)]; + tensor var_1603_cast_fp16 = scatter_along_axis(axis = var_1603_axis_0, data = var_1601_cast_fp16, indices = write_indices_7, mode = var_1603_mode_0, updates = v_7_cast_fp16, validate_indices = var_1603_validate_indices_0)[name = tensor("op_1603_cast_fp16")]; + tensor concat_25 = const()[name = tensor("concat_25"), val = tensor([1, 0, 0, 0, 0])]; + tensor concat_26 = const()[name = tensor("concat_26"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_7_internal_tensor_assign_2_stride_0 = const()[name = tensor("new_cache_7_internal_tensor_assign_2_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_7_internal_tensor_assign_2_begin_mask_0 = const()[name = tensor("new_cache_7_internal_tensor_assign_2_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_7_internal_tensor_assign_2_end_mask_0 = const()[name = tensor("new_cache_7_internal_tensor_assign_2_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_7_internal_tensor_assign_2_squeeze_mask_0 = const()[name = tensor("new_cache_7_internal_tensor_assign_2_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_55 = const()[name = tensor("shape_55"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_7 = const()[name = tensor("reduce_prod_7"), val = tensor(1048576)]; + tensor range_1d_7_start_0 = const()[name = tensor("range_1d_7_start_0"), val = tensor(0)]; + tensor range_1d_7_step_0 = const()[name = tensor("range_1d_7_step_0"), val = tensor(1)]; + tensor range_1d_7 = range_1d(end = reduce_prod_7, start = range_1d_7_start_0, step = range_1d_7_step_0)[name = tensor("range_1d_7")]; + tensor reshape_35 = reshape(shape = shape_55, x = range_1d_7)[name = tensor("reshape_35")]; + tensor slice_by_index_7 = slice_by_index(begin = concat_25, begin_mask = new_cache_7_internal_tensor_assign_2_begin_mask_0, end = concat_26, end_mask = new_cache_7_internal_tensor_assign_2_end_mask_0, squeeze_mask = new_cache_7_internal_tensor_assign_2_squeeze_mask_0, stride = new_cache_7_internal_tensor_assign_2_stride_0, x = reshape_35)[name = tensor("slice_by_index_7")]; + tensor reshape_36_shape_0 = const()[name = tensor("reshape_36_shape_0"), val = tensor([-1])]; + tensor reshape_36 = reshape(shape = reshape_36_shape_0, x = slice_by_index_7)[name = tensor("reshape_36")]; + tensor reshape_37_shape_0 = const()[name = tensor("reshape_37_shape_0"), val = tensor([-1])]; + tensor reshape_37_cast_fp16 = reshape(shape = reshape_37_shape_0, x = var_1603_cast_fp16)[name = tensor("reshape_37_cast_fp16")]; + tensor reshape_38_shape_0 = const()[name = tensor("reshape_38_shape_0"), val = tensor([-1])]; + tensor reshape_38_cast_fp16 = reshape(shape = reshape_38_shape_0, x = reshape_34_cast_fp16)[name = tensor("reshape_38_cast_fp16")]; + tensor scatter_7_mode_0 = const()[name = tensor("scatter_7_mode_0"), val = tensor("update")]; + tensor scatter_7_axis_0 = const()[name = tensor("scatter_7_axis_0"), val = tensor(0)]; + tensor scatter_7_validate_indices_0 = const()[name = tensor("scatter_7_validate_indices_0"), val = tensor(false)]; + tensor scatter_7_cast_fp16 = scatter(axis = scatter_7_axis_0, data = reshape_38_cast_fp16, indices = reshape_36, mode = scatter_7_mode_0, updates = reshape_37_cast_fp16, validate_indices = scatter_7_validate_indices_0)[name = tensor("scatter_7_cast_fp16")]; + tensor reshape_39_cast_fp16 = reshape(shape = shape_55, x = scatter_7_cast_fp16)[name = tensor("reshape_39_cast_fp16")]; + tensor reshape_39_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("reshape_39_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor keys_19_begin_0 = const()[name = tensor("keys_19_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor keys_19_end_0 = const()[name = tensor("keys_19_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor keys_19_end_mask_0 = const()[name = tensor("keys_19_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor keys_19_squeeze_mask_0 = const()[name = tensor("keys_19_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor keys_19_cast_fp16 = slice_by_index(begin = keys_19_begin_0, end = keys_19_end_0, end_mask = keys_19_end_mask_0, squeeze_mask = keys_19_squeeze_mask_0, x = reshape_39_cast_fp16)[name = tensor("keys_19_cast_fp16")]; + tensor values_19_begin_0 = const()[name = tensor("values_19_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor values_19_end_0 = const()[name = tensor("values_19_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor values_19_end_mask_0 = const()[name = tensor("values_19_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor values_19_squeeze_mask_0 = const()[name = tensor("values_19_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor values_19_cast_fp16 = slice_by_index(begin = values_19_begin_0, end = values_19_end_0, end_mask = values_19_end_mask_0, squeeze_mask = values_19_squeeze_mask_0, x = reshape_39_cast_fp16)[name = tensor("values_19_cast_fp16")]; + tensor var_1615_cast_fp16 = not_equal(x = keys_19_cast_fp16, y = keys_19_cast_fp16)[name = tensor("op_1615_cast_fp16")]; + tensor keys_21_cast_fp16 = select(a = var_463_to_fp16, b = keys_19_cast_fp16, cond = var_1615_cast_fp16)[name = tensor("keys_21_cast_fp16")]; + tensor var_1623_cast_fp16 = not_equal(x = values_19_cast_fp16, y = values_19_cast_fp16)[name = tensor("op_1623_cast_fp16")]; + tensor values_21_cast_fp16 = select(a = var_463_to_fp16, b = values_19_cast_fp16, cond = var_1623_cast_fp16)[name = tensor("values_21_cast_fp16")]; + tensor var_1647 = const()[name = tensor("op_1647"), val = tensor([0, 2, 1, 3])]; + tensor var_1663 = const()[name = tensor("op_1663"), val = tensor([1, 1, 1])]; + tensor var_1664_cast_fp16 = reshape(shape = var_1663, x = position3_to_fp16)[name = tensor("op_1664_cast_fp16")]; + tensor var_1684_to_fp16 = const()[name = tensor("op_1684_to_fp16"), val = tensor(0x1p+0)]; + tensor valid_len_7_cast_fp16 = add(x = var_1664_cast_fp16, y = var_1684_to_fp16)[name = tensor("valid_len_7_cast_fp16")]; + tensor valid_mask_7_cast_fp16 = less(x = k_positions_1_promoted_to_fp16, y = valid_len_7_cast_fp16)[name = tensor("valid_mask_7_cast_fp16")]; + tensor causal_mask_7_cast_fp16 = less_equal(x = k_positions_1_promoted_to_fp16, y = var_1664_cast_fp16)[name = tensor("causal_mask_7_cast_fp16")]; + tensor attn_mask_13 = logical_and(x = valid_mask_7_cast_fp16, y = causal_mask_7_cast_fp16)[name = tensor("attn_mask_13")]; + tensor attn_mask_15_axes_0 = const()[name = tensor("attn_mask_15_axes_0"), val = tensor([1])]; + tensor attn_mask_15 = expand_dims(axes = attn_mask_15_axes_0, x = attn_mask_13)[name = tensor("attn_mask_15")]; + tensor var_1702_transpose_x_0 = const()[name = tensor("op_1702_transpose_x_0"), val = tensor(false)]; + tensor var_1702_transpose_y_0 = const()[name = tensor("op_1702_transpose_y_0"), val = tensor(false)]; + tensor transpose_78_perm_0 = const()[name = tensor("transpose_78_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_79_perm_0 = const()[name = tensor("transpose_79_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_79 = transpose(perm = transpose_79_perm_0, x = keys_21_cast_fp16)[name = tensor("transpose_201")]; + tensor transpose_78 = transpose(perm = transpose_78_perm_0, x = q_21_cast_fp16)[name = tensor("transpose_202")]; + tensor var_1702_cast_fp16 = matmul(transpose_x = var_1702_transpose_x_0, transpose_y = var_1702_transpose_y_0, x = transpose_78, y = transpose_79)[name = tensor("op_1702_cast_fp16")]; + tensor var_1696_to_fp16 = const()[name = tensor("op_1696_to_fp16"), val = tensor([0x1p-3])]; + tensor attn_weights_19_cast_fp16 = mul(x = var_1702_cast_fp16, y = var_1696_to_fp16)[name = tensor("attn_weights_19_cast_fp16")]; + tensor var_1704 = logical_not(x = attn_mask_15)[name = tensor("op_1704")]; + tensor var_1705_to_fp16 = const()[name = tensor("op_1705_to_fp16"), val = tensor(-inf)]; + tensor attn_weights_21_cast_fp16 = select(a = var_1705_to_fp16, b = attn_weights_19_cast_fp16, cond = var_1704)[name = tensor("attn_weights_21_cast_fp16")]; + tensor var_1707 = const()[name = tensor("op_1707"), val = tensor(-1)]; + tensor attn_weights_23_cast_fp16 = softmax(axis = var_1707, x = attn_weights_21_cast_fp16)[name = tensor("attn_weights_23_cast_fp16")]; + tensor attn_output_7_transpose_x_0 = const()[name = tensor("attn_output_7_transpose_x_0"), val = tensor(false)]; + tensor attn_output_7_transpose_y_0 = const()[name = tensor("attn_output_7_transpose_y_0"), val = tensor(false)]; + tensor values_23_cast_fp16 = transpose(perm = var_1647, x = values_21_cast_fp16)[name = tensor("transpose_203")]; + tensor attn_output_7_cast_fp16 = matmul(transpose_x = attn_output_7_transpose_x_0, transpose_y = attn_output_7_transpose_y_0, x = attn_weights_23_cast_fp16, y = values_23_cast_fp16)[name = tensor("attn_output_7_cast_fp16")]; + tensor var_1715 = const()[name = tensor("op_1715"), val = tensor([0, 2, 1, 3])]; + tensor var_1720 = const()[name = tensor("op_1720"), val = tensor([1, 1, 1024])]; + tensor var_1716_cast_fp16 = transpose(perm = var_1715, x = attn_output_7_cast_fp16)[name = tensor("transpose_200")]; + tensor input_43_cast_fp16 = reshape(shape = var_1720, x = var_1716_cast_fp16)[name = tensor("input_43_cast_fp16")]; + tensor attn3_out_proj_weight_to_fp16 = const()[name = tensor("attn3_out_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(82951808)))]; + tensor linear_14_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = attn3_out_proj_weight_to_fp16, x = input_43_cast_fp16)[name = tensor("linear_14_cast_fp16")]; + tensor var_1726_to_fp16 = const()[name = tensor("op_1726_to_fp16"), val = tensor(0x1p+0)]; + tensor var_1727_cast_fp16 = add(x = position3_to_fp16, y = var_1726_to_fp16)[name = tensor("op_1727_cast_fp16")]; + tensor var_1727_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_1727_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor input_45_cast_fp16 = add(x = input_39_cast_fp16, y = linear_14_cast_fp16)[name = tensor("input_45_cast_fp16")]; + tensor input_47_axes_0 = const()[name = tensor("input_47_axes_0"), val = tensor([-1])]; + tensor norm3_2_weight_to_fp16 = const()[name = tensor("norm3_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(85049024)))]; + tensor norm3_2_bias_to_fp16 = const()[name = tensor("norm3_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(85051136)))]; + tensor var_1731_to_fp16 = const()[name = tensor("op_1731_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_47_cast_fp16 = layer_norm(axes = input_47_axes_0, beta = norm3_2_bias_to_fp16, epsilon = var_1731_to_fp16, gamma = norm3_2_weight_to_fp16, x = input_45_cast_fp16)[name = tensor("input_47_cast_fp16")]; + tensor linear3_1_weight_to_fp16 = const()[name = tensor("linear3_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(85053248)))]; + tensor linear_15_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = linear3_1_weight_to_fp16, x = input_47_cast_fp16)[name = tensor("linear_15_cast_fp16")]; + tensor input_49_mode_0 = const()[name = tensor("input_49_mode_0"), val = tensor("EXACT")]; + tensor input_49_cast_fp16 = gelu(mode = input_49_mode_0, x = linear_15_cast_fp16)[name = tensor("input_49_cast_fp16")]; + tensor linear3_2_weight_to_fp16 = const()[name = tensor("linear3_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(93441920)))]; + tensor linear_16_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = linear3_2_weight_to_fp16, x = input_49_cast_fp16)[name = tensor("linear_16_cast_fp16")]; + tensor input_51_cast_fp16 = add(x = input_45_cast_fp16, y = linear_16_cast_fp16)[name = tensor("input_51_cast_fp16")]; + tensor input_53_axes_0 = const()[name = tensor("input_53_axes_0"), val = tensor([-1])]; + tensor norm4_1_weight_to_fp16 = const()[name = tensor("norm4_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(101830592)))]; + tensor norm4_1_bias_to_fp16 = const()[name = tensor("norm4_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(101832704)))]; + tensor var_1748_to_fp16 = const()[name = tensor("op_1748_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_53_cast_fp16 = layer_norm(axes = input_53_axes_0, beta = norm4_1_bias_to_fp16, epsilon = var_1748_to_fp16, gamma = norm4_1_weight_to_fp16, x = input_51_cast_fp16)[name = tensor("input_53_cast_fp16")]; + tensor attn4_in_proj_weight_to_fp16 = const()[name = tensor("attn4_in_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(101834816)))]; + tensor linear_17_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = attn4_in_proj_weight_to_fp16, x = input_53_cast_fp16)[name = tensor("linear_17_cast_fp16")]; + tensor var_1762 = const()[name = tensor("op_1762"), val = tensor([1, 1, 3, 16, 64])]; + tensor qkv_9_cast_fp16 = reshape(shape = var_1762, x = linear_17_cast_fp16)[name = tensor("qkv_9_cast_fp16")]; + tensor q_25_begin_0 = const()[name = tensor("q_25_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor q_25_end_0 = const()[name = tensor("q_25_end_0"), val = tensor([1, 1, 1, 16, 64])]; + tensor q_25_end_mask_0 = const()[name = tensor("q_25_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor q_25_squeeze_mask_0 = const()[name = tensor("q_25_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor q_25_cast_fp16 = slice_by_index(begin = q_25_begin_0, end = q_25_end_0, end_mask = q_25_end_mask_0, squeeze_mask = q_25_squeeze_mask_0, x = qkv_9_cast_fp16)[name = tensor("q_25_cast_fp16")]; + tensor k_17_begin_0 = const()[name = tensor("k_17_begin_0"), val = tensor([0, 0, 1, 0, 0])]; + tensor k_17_end_0 = const()[name = tensor("k_17_end_0"), val = tensor([1, 1, 2, 16, 64])]; + tensor k_17_end_mask_0 = const()[name = tensor("k_17_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor k_17_squeeze_mask_0 = const()[name = tensor("k_17_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor k_17_cast_fp16 = slice_by_index(begin = k_17_begin_0, end = k_17_end_0, end_mask = k_17_end_mask_0, squeeze_mask = k_17_squeeze_mask_0, x = qkv_9_cast_fp16)[name = tensor("k_17_cast_fp16")]; + tensor v_9_begin_0 = const()[name = tensor("v_9_begin_0"), val = tensor([0, 0, 2, 0, 0])]; + tensor v_9_end_0 = const()[name = tensor("v_9_end_0"), val = tensor([1, 1, 3, 16, 64])]; + tensor v_9_end_mask_0 = const()[name = tensor("v_9_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor v_9_squeeze_mask_0 = const()[name = tensor("v_9_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor v_9_cast_fp16 = slice_by_index(begin = v_9_begin_0, end = v_9_end_0, end_mask = v_9_end_mask_0, squeeze_mask = v_9_squeeze_mask_0, x = qkv_9_cast_fp16)[name = tensor("v_9_cast_fp16")]; + tensor var_1828 = const()[name = tensor("op_1828"), val = tensor([1, 1, 1, 1])]; + tensor position4_to_fp16_dtype_0 = const()[name = tensor("position4_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor position4_to_fp16 = cast(dtype = position4_to_fp16_dtype_0, x = position4)[name = tensor("cast_533")]; + tensor ts_29_cast_fp16 = reshape(shape = var_1828, x = position4_to_fp16)[name = tensor("ts_29_cast_fp16")]; + tensor var_1835 = const()[name = tensor("op_1835"), val = tensor([1, 1, 16, 32, 2])]; + tensor q_complex_9_cast_fp16 = reshape(shape = var_1835, x = q_25_cast_fp16)[name = tensor("q_complex_9_cast_fp16")]; + tensor var_1842 = const()[name = tensor("op_1842"), val = tensor([1, 1, 16, 32, 2])]; + tensor k_complex_9_cast_fp16 = reshape(shape = var_1842, x = k_17_cast_fp16)[name = tensor("k_complex_9_cast_fp16")]; + tensor var_1846_begin_0 = const()[name = tensor("op_1846_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_1846_end_0 = const()[name = tensor("op_1846_end_0"), val = tensor([1, 1, 16, 32, 1])]; + tensor var_1846_end_mask_0 = const()[name = tensor("op_1846_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_1846_squeeze_mask_0 = const()[name = tensor("op_1846_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_1846_cast_fp16 = slice_by_index(begin = var_1846_begin_0, end = var_1846_end_0, end_mask = var_1846_end_mask_0, squeeze_mask = var_1846_squeeze_mask_0, x = q_complex_9_cast_fp16)[name = tensor("op_1846_cast_fp16")]; + tensor var_1854_begin_0 = const()[name = tensor("op_1854_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_1854_end_0 = const()[name = tensor("op_1854_end_0"), val = tensor([1, 1, 16, 32, 2])]; + tensor var_1854_end_mask_0 = const()[name = tensor("op_1854_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_1854_squeeze_mask_0 = const()[name = tensor("op_1854_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_1854_cast_fp16 = slice_by_index(begin = var_1854_begin_0, end = var_1854_end_0, end_mask = var_1854_end_mask_0, squeeze_mask = var_1854_squeeze_mask_0, x = q_complex_9_cast_fp16)[name = tensor("op_1854_cast_fp16")]; + tensor var_1862_begin_0 = const()[name = tensor("op_1862_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_1862_end_0 = const()[name = tensor("op_1862_end_0"), val = tensor([1, 1, 16, 32, 1])]; + tensor var_1862_end_mask_0 = const()[name = tensor("op_1862_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_1862_squeeze_mask_0 = const()[name = tensor("op_1862_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_1862_cast_fp16 = slice_by_index(begin = var_1862_begin_0, end = var_1862_end_0, end_mask = var_1862_end_mask_0, squeeze_mask = var_1862_squeeze_mask_0, x = k_complex_9_cast_fp16)[name = tensor("op_1862_cast_fp16")]; + tensor var_1870_begin_0 = const()[name = tensor("op_1870_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_1870_end_0 = const()[name = tensor("op_1870_end_0"), val = tensor([1, 1, 16, 32, 2])]; + tensor var_1870_end_mask_0 = const()[name = tensor("op_1870_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_1870_squeeze_mask_0 = const()[name = tensor("op_1870_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_1870_cast_fp16 = slice_by_index(begin = var_1870_begin_0, end = var_1870_end_0, end_mask = var_1870_end_mask_0, squeeze_mask = var_1870_squeeze_mask_0, x = k_complex_9_cast_fp16)[name = tensor("op_1870_cast_fp16")]; + tensor freqs_9_to_fp16 = const()[name = tensor("freqs_9_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(108126336)))]; + tensor var_1876_cast_fp16 = mul(x = freqs_9_to_fp16, y = ts_29_cast_fp16)[name = tensor("op_1876_cast_fp16")]; + tensor rotr_9_cast_fp16 = cos(x = var_1876_cast_fp16)[name = tensor("rotr_9_cast_fp16")]; + tensor roti_9_cast_fp16 = sin(x = var_1876_cast_fp16)[name = tensor("roti_9_cast_fp16")]; + tensor var_1880_cast_fp16 = mul(x = var_1846_cast_fp16, y = rotr_9_cast_fp16)[name = tensor("op_1880_cast_fp16")]; + tensor var_1881_cast_fp16 = mul(x = var_1854_cast_fp16, y = roti_9_cast_fp16)[name = tensor("op_1881_cast_fp16")]; + tensor qor_17_cast_fp16 = sub(x = var_1880_cast_fp16, y = var_1881_cast_fp16)[name = tensor("qor_17_cast_fp16")]; + tensor var_1884_cast_fp16 = mul(x = var_1846_cast_fp16, y = roti_9_cast_fp16)[name = tensor("op_1884_cast_fp16")]; + tensor var_1885_cast_fp16 = mul(x = var_1854_cast_fp16, y = rotr_9_cast_fp16)[name = tensor("op_1885_cast_fp16")]; + tensor qoi_17_cast_fp16 = add(x = var_1884_cast_fp16, y = var_1885_cast_fp16)[name = tensor("qoi_17_cast_fp16")]; + tensor var_1888_cast_fp16 = mul(x = var_1862_cast_fp16, y = rotr_9_cast_fp16)[name = tensor("op_1888_cast_fp16")]; + tensor var_1889_cast_fp16 = mul(x = var_1870_cast_fp16, y = roti_9_cast_fp16)[name = tensor("op_1889_cast_fp16")]; + tensor kor_17_cast_fp16 = sub(x = var_1888_cast_fp16, y = var_1889_cast_fp16)[name = tensor("kor_17_cast_fp16")]; + tensor var_1892_cast_fp16 = mul(x = var_1862_cast_fp16, y = roti_9_cast_fp16)[name = tensor("op_1892_cast_fp16")]; + tensor var_1893_cast_fp16 = mul(x = var_1870_cast_fp16, y = rotr_9_cast_fp16)[name = tensor("op_1893_cast_fp16")]; + tensor koi_17_cast_fp16 = add(x = var_1892_cast_fp16, y = var_1893_cast_fp16)[name = tensor("koi_17_cast_fp16")]; + tensor qo_9_axis_0 = const()[name = tensor("qo_9_axis_0"), val = tensor(-1)]; + tensor qo_9_cast_fp16 = stack(axis = qo_9_axis_0, values = (qor_17_cast_fp16, qoi_17_cast_fp16))[name = tensor("qo_9_cast_fp16")]; + tensor ko_9_axis_0 = const()[name = tensor("ko_9_axis_0"), val = tensor(-1)]; + tensor ko_9_cast_fp16 = stack(axis = ko_9_axis_0, values = (kor_17_cast_fp16, koi_17_cast_fp16))[name = tensor("ko_9_cast_fp16")]; + tensor var_1926 = const()[name = tensor("op_1926"), val = tensor([1, 1, 16, 64])]; + tensor q_27_cast_fp16 = reshape(shape = var_1926, x = qo_9_cast_fp16)[name = tensor("q_27_cast_fp16")]; + tensor var_1932 = const()[name = tensor("op_1932"), val = tensor([1, 1, 16, 64])]; + tensor k_19_cast_fp16 = reshape(shape = var_1932, x = ko_9_cast_fp16)[name = tensor("k_19_cast_fp16")]; + tensor _inversed_1957_y_0_to_fp16 = const()[name = tensor("_inversed_1957_y_0_to_fp16"), val = tensor(0x1p-9)]; + tensor _inversed_1957_cast_fp16 = mul(x = ts_29_cast_fp16, y = _inversed_1957_y_0_to_fp16)[name = tensor("_inversed_1957_cast_fp16")]; + tensor var_1958_cast_fp16 = floor(x = _inversed_1957_cast_fp16)[name = tensor("op_1958_cast_fp16")]; + tensor var_1959_to_fp16 = const()[name = tensor("op_1959_to_fp16"), val = tensor(0x1p+9)]; + tensor var_1960_cast_fp16 = mul(x = var_1958_cast_fp16, y = var_1959_to_fp16)[name = tensor("op_1960_cast_fp16")]; + tensor write_indices_float_19_cast_fp16 = sub(x = ts_29_cast_fp16, y = var_1960_cast_fp16)[name = tensor("write_indices_float_19_cast_fp16")]; + tensor var_1967_dtype_0 = const()[name = tensor("op_1967_dtype_0"), val = tensor("int32")]; + tensor write_indices_9_reps_0 = const()[name = tensor("write_indices_9_reps_0"), val = tensor([1, 1, 16, 64])]; + tensor write_indices_float_19_cast_fp16_to_int32 = cast(dtype = var_1967_dtype_0, x = write_indices_float_19_cast_fp16)[name = tensor("cast_532")]; + tensor write_indices_9 = tile(reps = write_indices_9_reps_0, x = write_indices_float_19_cast_fp16_to_int32)[name = tensor("write_indices_9")]; + tensor var_1977_begin_0 = const()[name = tensor("op_1977_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_1977_end_0 = const()[name = tensor("op_1977_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor var_1977_end_mask_0 = const()[name = tensor("op_1977_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_1977_squeeze_mask_0 = const()[name = tensor("op_1977_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor cache4_to_fp16_dtype_0 = const()[name = tensor("cache4_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor cache4_to_fp16 = cast(dtype = cache4_to_fp16_dtype_0, x = cache4)[name = tensor("cast_531")]; + tensor var_1977_cast_fp16 = slice_by_index(begin = var_1977_begin_0, end = var_1977_end_0, end_mask = var_1977_end_mask_0, squeeze_mask = var_1977_squeeze_mask_0, x = cache4_to_fp16)[name = tensor("op_1977_cast_fp16")]; + tensor var_1979_axis_0 = const()[name = tensor("op_1979_axis_0"), val = tensor(1)]; + tensor var_1979_mode_0 = const()[name = tensor("op_1979_mode_0"), val = tensor("update")]; + tensor var_1979_validate_indices_0 = const()[name = tensor("op_1979_validate_indices_0"), val = tensor(false)]; + tensor var_1979_cast_fp16 = scatter_along_axis(axis = var_1979_axis_0, data = var_1977_cast_fp16, indices = write_indices_9, mode = var_1979_mode_0, updates = k_19_cast_fp16, validate_indices = var_1979_validate_indices_0)[name = tensor("op_1979_cast_fp16")]; + tensor concat_30 = const()[name = tensor("concat_30"), val = tensor([0, 0, 0, 0, 0])]; + tensor concat_31 = const()[name = tensor("concat_31"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_9_internal_tensor_assign_1_stride_0 = const()[name = tensor("new_cache_9_internal_tensor_assign_1_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_9_internal_tensor_assign_1_begin_mask_0 = const()[name = tensor("new_cache_9_internal_tensor_assign_1_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_9_internal_tensor_assign_1_end_mask_0 = const()[name = tensor("new_cache_9_internal_tensor_assign_1_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_9_internal_tensor_assign_1_squeeze_mask_0 = const()[name = tensor("new_cache_9_internal_tensor_assign_1_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_56 = const()[name = tensor("shape_56"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_8 = const()[name = tensor("reduce_prod_8"), val = tensor(1048576)]; + tensor range_1d_8_start_0 = const()[name = tensor("range_1d_8_start_0"), val = tensor(0)]; + tensor range_1d_8_step_0 = const()[name = tensor("range_1d_8_step_0"), val = tensor(1)]; + tensor range_1d_8 = range_1d(end = reduce_prod_8, start = range_1d_8_start_0, step = range_1d_8_step_0)[name = tensor("range_1d_8")]; + tensor reshape_40 = reshape(shape = shape_56, x = range_1d_8)[name = tensor("reshape_40")]; + tensor slice_by_index_8 = slice_by_index(begin = concat_30, begin_mask = new_cache_9_internal_tensor_assign_1_begin_mask_0, end = concat_31, end_mask = new_cache_9_internal_tensor_assign_1_end_mask_0, squeeze_mask = new_cache_9_internal_tensor_assign_1_squeeze_mask_0, stride = new_cache_9_internal_tensor_assign_1_stride_0, x = reshape_40)[name = tensor("slice_by_index_8")]; + tensor reshape_41_shape_0 = const()[name = tensor("reshape_41_shape_0"), val = tensor([-1])]; + tensor reshape_41 = reshape(shape = reshape_41_shape_0, x = slice_by_index_8)[name = tensor("reshape_41")]; + tensor reshape_42_shape_0 = const()[name = tensor("reshape_42_shape_0"), val = tensor([-1])]; + tensor reshape_42_cast_fp16 = reshape(shape = reshape_42_shape_0, x = var_1979_cast_fp16)[name = tensor("reshape_42_cast_fp16")]; + tensor reshape_43_shape_0 = const()[name = tensor("reshape_43_shape_0"), val = tensor([-1])]; + tensor reshape_43_cast_fp16 = reshape(shape = reshape_43_shape_0, x = cache4_to_fp16)[name = tensor("reshape_43_cast_fp16")]; + tensor scatter_8_mode_0 = const()[name = tensor("scatter_8_mode_0"), val = tensor("update")]; + tensor scatter_8_axis_0 = const()[name = tensor("scatter_8_axis_0"), val = tensor(0)]; + tensor scatter_8_validate_indices_0 = const()[name = tensor("scatter_8_validate_indices_0"), val = tensor(false)]; + tensor scatter_8_cast_fp16 = scatter(axis = scatter_8_axis_0, data = reshape_43_cast_fp16, indices = reshape_41, mode = scatter_8_mode_0, updates = reshape_42_cast_fp16, validate_indices = scatter_8_validate_indices_0)[name = tensor("scatter_8_cast_fp16")]; + tensor reshape_44_cast_fp16 = reshape(shape = shape_56, x = scatter_8_cast_fp16)[name = tensor("reshape_44_cast_fp16")]; + tensor var_1987_begin_0 = const()[name = tensor("op_1987_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor var_1987_end_0 = const()[name = tensor("op_1987_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor var_1987_end_mask_0 = const()[name = tensor("op_1987_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_1987_squeeze_mask_0 = const()[name = tensor("op_1987_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor var_1987_cast_fp16 = slice_by_index(begin = var_1987_begin_0, end = var_1987_end_0, end_mask = var_1987_end_mask_0, squeeze_mask = var_1987_squeeze_mask_0, x = reshape_44_cast_fp16)[name = tensor("op_1987_cast_fp16")]; + tensor var_1989_axis_0 = const()[name = tensor("op_1989_axis_0"), val = tensor(1)]; + tensor var_1989_mode_0 = const()[name = tensor("op_1989_mode_0"), val = tensor("update")]; + tensor var_1989_validate_indices_0 = const()[name = tensor("op_1989_validate_indices_0"), val = tensor(false)]; + tensor var_1989_cast_fp16 = scatter_along_axis(axis = var_1989_axis_0, data = var_1987_cast_fp16, indices = write_indices_9, mode = var_1989_mode_0, updates = v_9_cast_fp16, validate_indices = var_1989_validate_indices_0)[name = tensor("op_1989_cast_fp16")]; + tensor concat_32 = const()[name = tensor("concat_32"), val = tensor([1, 0, 0, 0, 0])]; + tensor concat_33 = const()[name = tensor("concat_33"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_9_internal_tensor_assign_2_stride_0 = const()[name = tensor("new_cache_9_internal_tensor_assign_2_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_9_internal_tensor_assign_2_begin_mask_0 = const()[name = tensor("new_cache_9_internal_tensor_assign_2_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_9_internal_tensor_assign_2_end_mask_0 = const()[name = tensor("new_cache_9_internal_tensor_assign_2_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_9_internal_tensor_assign_2_squeeze_mask_0 = const()[name = tensor("new_cache_9_internal_tensor_assign_2_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_57 = const()[name = tensor("shape_57"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_9 = const()[name = tensor("reduce_prod_9"), val = tensor(1048576)]; + tensor range_1d_9_start_0 = const()[name = tensor("range_1d_9_start_0"), val = tensor(0)]; + tensor range_1d_9_step_0 = const()[name = tensor("range_1d_9_step_0"), val = tensor(1)]; + tensor range_1d_9 = range_1d(end = reduce_prod_9, start = range_1d_9_start_0, step = range_1d_9_step_0)[name = tensor("range_1d_9")]; + tensor reshape_45 = reshape(shape = shape_57, x = range_1d_9)[name = tensor("reshape_45")]; + tensor slice_by_index_9 = slice_by_index(begin = concat_32, begin_mask = new_cache_9_internal_tensor_assign_2_begin_mask_0, end = concat_33, end_mask = new_cache_9_internal_tensor_assign_2_end_mask_0, squeeze_mask = new_cache_9_internal_tensor_assign_2_squeeze_mask_0, stride = new_cache_9_internal_tensor_assign_2_stride_0, x = reshape_45)[name = tensor("slice_by_index_9")]; + tensor reshape_46_shape_0 = const()[name = tensor("reshape_46_shape_0"), val = tensor([-1])]; + tensor reshape_46 = reshape(shape = reshape_46_shape_0, x = slice_by_index_9)[name = tensor("reshape_46")]; + tensor reshape_47_shape_0 = const()[name = tensor("reshape_47_shape_0"), val = tensor([-1])]; + tensor reshape_47_cast_fp16 = reshape(shape = reshape_47_shape_0, x = var_1989_cast_fp16)[name = tensor("reshape_47_cast_fp16")]; + tensor reshape_48_shape_0 = const()[name = tensor("reshape_48_shape_0"), val = tensor([-1])]; + tensor reshape_48_cast_fp16 = reshape(shape = reshape_48_shape_0, x = reshape_44_cast_fp16)[name = tensor("reshape_48_cast_fp16")]; + tensor scatter_9_mode_0 = const()[name = tensor("scatter_9_mode_0"), val = tensor("update")]; + tensor scatter_9_axis_0 = const()[name = tensor("scatter_9_axis_0"), val = tensor(0)]; + tensor scatter_9_validate_indices_0 = const()[name = tensor("scatter_9_validate_indices_0"), val = tensor(false)]; + tensor scatter_9_cast_fp16 = scatter(axis = scatter_9_axis_0, data = reshape_48_cast_fp16, indices = reshape_46, mode = scatter_9_mode_0, updates = reshape_47_cast_fp16, validate_indices = scatter_9_validate_indices_0)[name = tensor("scatter_9_cast_fp16")]; + tensor reshape_49_cast_fp16 = reshape(shape = shape_57, x = scatter_9_cast_fp16)[name = tensor("reshape_49_cast_fp16")]; + tensor reshape_49_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("reshape_49_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor keys_25_begin_0 = const()[name = tensor("keys_25_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor keys_25_end_0 = const()[name = tensor("keys_25_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor keys_25_end_mask_0 = const()[name = tensor("keys_25_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor keys_25_squeeze_mask_0 = const()[name = tensor("keys_25_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor keys_25_cast_fp16 = slice_by_index(begin = keys_25_begin_0, end = keys_25_end_0, end_mask = keys_25_end_mask_0, squeeze_mask = keys_25_squeeze_mask_0, x = reshape_49_cast_fp16)[name = tensor("keys_25_cast_fp16")]; + tensor values_25_begin_0 = const()[name = tensor("values_25_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor values_25_end_0 = const()[name = tensor("values_25_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor values_25_end_mask_0 = const()[name = tensor("values_25_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor values_25_squeeze_mask_0 = const()[name = tensor("values_25_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor values_25_cast_fp16 = slice_by_index(begin = values_25_begin_0, end = values_25_end_0, end_mask = values_25_end_mask_0, squeeze_mask = values_25_squeeze_mask_0, x = reshape_49_cast_fp16)[name = tensor("values_25_cast_fp16")]; + tensor var_2001_cast_fp16 = not_equal(x = keys_25_cast_fp16, y = keys_25_cast_fp16)[name = tensor("op_2001_cast_fp16")]; + tensor keys_27_cast_fp16 = select(a = var_463_to_fp16, b = keys_25_cast_fp16, cond = var_2001_cast_fp16)[name = tensor("keys_27_cast_fp16")]; + tensor var_2009_cast_fp16 = not_equal(x = values_25_cast_fp16, y = values_25_cast_fp16)[name = tensor("op_2009_cast_fp16")]; + tensor values_27_cast_fp16 = select(a = var_463_to_fp16, b = values_25_cast_fp16, cond = var_2009_cast_fp16)[name = tensor("values_27_cast_fp16")]; + tensor var_2033 = const()[name = tensor("op_2033"), val = tensor([0, 2, 1, 3])]; + tensor var_2049 = const()[name = tensor("op_2049"), val = tensor([1, 1, 1])]; + tensor var_2050_cast_fp16 = reshape(shape = var_2049, x = position4_to_fp16)[name = tensor("op_2050_cast_fp16")]; + tensor var_2070_to_fp16 = const()[name = tensor("op_2070_to_fp16"), val = tensor(0x1p+0)]; + tensor valid_len_9_cast_fp16 = add(x = var_2050_cast_fp16, y = var_2070_to_fp16)[name = tensor("valid_len_9_cast_fp16")]; + tensor valid_mask_9_cast_fp16 = less(x = k_positions_1_promoted_to_fp16, y = valid_len_9_cast_fp16)[name = tensor("valid_mask_9_cast_fp16")]; + tensor causal_mask_9_cast_fp16 = less_equal(x = k_positions_1_promoted_to_fp16, y = var_2050_cast_fp16)[name = tensor("causal_mask_9_cast_fp16")]; + tensor attn_mask_17 = logical_and(x = valid_mask_9_cast_fp16, y = causal_mask_9_cast_fp16)[name = tensor("attn_mask_17")]; + tensor attn_mask_19_axes_0 = const()[name = tensor("attn_mask_19_axes_0"), val = tensor([1])]; + tensor attn_mask_19 = expand_dims(axes = attn_mask_19_axes_0, x = attn_mask_17)[name = tensor("attn_mask_19")]; + tensor var_2088_transpose_x_0 = const()[name = tensor("op_2088_transpose_x_0"), val = tensor(false)]; + tensor var_2088_transpose_y_0 = const()[name = tensor("op_2088_transpose_y_0"), val = tensor(false)]; + tensor transpose_80_perm_0 = const()[name = tensor("transpose_80_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_81_perm_0 = const()[name = tensor("transpose_81_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_81 = transpose(perm = transpose_81_perm_0, x = keys_27_cast_fp16)[name = tensor("transpose_197")]; + tensor transpose_80 = transpose(perm = transpose_80_perm_0, x = q_27_cast_fp16)[name = tensor("transpose_198")]; + tensor var_2088_cast_fp16 = matmul(transpose_x = var_2088_transpose_x_0, transpose_y = var_2088_transpose_y_0, x = transpose_80, y = transpose_81)[name = tensor("op_2088_cast_fp16")]; + tensor var_2082_to_fp16 = const()[name = tensor("op_2082_to_fp16"), val = tensor([0x1p-3])]; + tensor attn_weights_25_cast_fp16 = mul(x = var_2088_cast_fp16, y = var_2082_to_fp16)[name = tensor("attn_weights_25_cast_fp16")]; + tensor var_2090 = logical_not(x = attn_mask_19)[name = tensor("op_2090")]; + tensor var_2091_to_fp16 = const()[name = tensor("op_2091_to_fp16"), val = tensor(-inf)]; + tensor attn_weights_27_cast_fp16 = select(a = var_2091_to_fp16, b = attn_weights_25_cast_fp16, cond = var_2090)[name = tensor("attn_weights_27_cast_fp16")]; + tensor var_2093 = const()[name = tensor("op_2093"), val = tensor(-1)]; + tensor attn_weights_29_cast_fp16 = softmax(axis = var_2093, x = attn_weights_27_cast_fp16)[name = tensor("attn_weights_29_cast_fp16")]; + tensor attn_output_9_transpose_x_0 = const()[name = tensor("attn_output_9_transpose_x_0"), val = tensor(false)]; + tensor attn_output_9_transpose_y_0 = const()[name = tensor("attn_output_9_transpose_y_0"), val = tensor(false)]; + tensor values_29_cast_fp16 = transpose(perm = var_2033, x = values_27_cast_fp16)[name = tensor("transpose_199")]; + tensor attn_output_9_cast_fp16 = matmul(transpose_x = attn_output_9_transpose_x_0, transpose_y = attn_output_9_transpose_y_0, x = attn_weights_29_cast_fp16, y = values_29_cast_fp16)[name = tensor("attn_output_9_cast_fp16")]; + tensor var_2101 = const()[name = tensor("op_2101"), val = tensor([0, 2, 1, 3])]; + tensor var_2106 = const()[name = tensor("op_2106"), val = tensor([1, 1, 1024])]; + tensor var_2102_cast_fp16 = transpose(perm = var_2101, x = attn_output_9_cast_fp16)[name = tensor("transpose_196")]; + tensor input_55_cast_fp16 = reshape(shape = var_2106, x = var_2102_cast_fp16)[name = tensor("input_55_cast_fp16")]; + tensor attn4_out_proj_weight_to_fp16 = const()[name = tensor("attn4_out_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(108126464)))]; + tensor linear_18_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = attn4_out_proj_weight_to_fp16, x = input_55_cast_fp16)[name = tensor("linear_18_cast_fp16")]; + tensor var_2112_to_fp16 = const()[name = tensor("op_2112_to_fp16"), val = tensor(0x1p+0)]; + tensor var_2113_cast_fp16 = add(x = position4_to_fp16, y = var_2112_to_fp16)[name = tensor("op_2113_cast_fp16")]; + tensor var_2113_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_2113_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor input_57_cast_fp16 = add(x = input_51_cast_fp16, y = linear_18_cast_fp16)[name = tensor("input_57_cast_fp16")]; + tensor input_59_axes_0 = const()[name = tensor("input_59_axes_0"), val = tensor([-1])]; + tensor norm4_2_weight_to_fp16 = const()[name = tensor("norm4_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(110223680)))]; + tensor norm4_2_bias_to_fp16 = const()[name = tensor("norm4_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(110225792)))]; + tensor var_2117_to_fp16 = const()[name = tensor("op_2117_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_59_cast_fp16 = layer_norm(axes = input_59_axes_0, beta = norm4_2_bias_to_fp16, epsilon = var_2117_to_fp16, gamma = norm4_2_weight_to_fp16, x = input_57_cast_fp16)[name = tensor("input_59_cast_fp16")]; + tensor linear4_1_weight_to_fp16 = const()[name = tensor("linear4_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(110227904)))]; + tensor linear_19_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = linear4_1_weight_to_fp16, x = input_59_cast_fp16)[name = tensor("linear_19_cast_fp16")]; + tensor input_61_mode_0 = const()[name = tensor("input_61_mode_0"), val = tensor("EXACT")]; + tensor input_61_cast_fp16 = gelu(mode = input_61_mode_0, x = linear_19_cast_fp16)[name = tensor("input_61_cast_fp16")]; + tensor linear4_2_weight_to_fp16 = const()[name = tensor("linear4_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(118616576)))]; + tensor linear_20_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = linear4_2_weight_to_fp16, x = input_61_cast_fp16)[name = tensor("linear_20_cast_fp16")]; + tensor input_63_cast_fp16 = add(x = input_57_cast_fp16, y = linear_20_cast_fp16)[name = tensor("input_63_cast_fp16")]; + tensor input_65_axes_0 = const()[name = tensor("input_65_axes_0"), val = tensor([-1])]; + tensor norm5_1_weight_to_fp16 = const()[name = tensor("norm5_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(127005248)))]; + tensor norm5_1_bias_to_fp16 = const()[name = tensor("norm5_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(127007360)))]; + tensor var_2134_to_fp16 = const()[name = tensor("op_2134_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_65_cast_fp16 = layer_norm(axes = input_65_axes_0, beta = norm5_1_bias_to_fp16, epsilon = var_2134_to_fp16, gamma = norm5_1_weight_to_fp16, x = input_63_cast_fp16)[name = tensor("input_65_cast_fp16")]; + tensor attn5_in_proj_weight_to_fp16 = const()[name = tensor("attn5_in_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(127009472)))]; + tensor linear_21_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = attn5_in_proj_weight_to_fp16, x = input_65_cast_fp16)[name = tensor("linear_21_cast_fp16")]; + tensor var_2148 = const()[name = tensor("op_2148"), val = tensor([1, 1, 3, 16, 64])]; + tensor qkv_11_cast_fp16 = reshape(shape = var_2148, x = linear_21_cast_fp16)[name = tensor("qkv_11_cast_fp16")]; + tensor q_31_begin_0 = const()[name = tensor("q_31_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor q_31_end_0 = const()[name = tensor("q_31_end_0"), val = tensor([1, 1, 1, 16, 64])]; + tensor q_31_end_mask_0 = const()[name = tensor("q_31_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor q_31_squeeze_mask_0 = const()[name = tensor("q_31_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor q_31_cast_fp16 = slice_by_index(begin = q_31_begin_0, end = q_31_end_0, end_mask = q_31_end_mask_0, squeeze_mask = q_31_squeeze_mask_0, x = qkv_11_cast_fp16)[name = tensor("q_31_cast_fp16")]; + tensor k_21_begin_0 = const()[name = tensor("k_21_begin_0"), val = tensor([0, 0, 1, 0, 0])]; + tensor k_21_end_0 = const()[name = tensor("k_21_end_0"), val = tensor([1, 1, 2, 16, 64])]; + tensor k_21_end_mask_0 = const()[name = tensor("k_21_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor k_21_squeeze_mask_0 = const()[name = tensor("k_21_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor k_21_cast_fp16 = slice_by_index(begin = k_21_begin_0, end = k_21_end_0, end_mask = k_21_end_mask_0, squeeze_mask = k_21_squeeze_mask_0, x = qkv_11_cast_fp16)[name = tensor("k_21_cast_fp16")]; + tensor v_11_begin_0 = const()[name = tensor("v_11_begin_0"), val = tensor([0, 0, 2, 0, 0])]; + tensor v_11_end_0 = const()[name = tensor("v_11_end_0"), val = tensor([1, 1, 3, 16, 64])]; + tensor v_11_end_mask_0 = const()[name = tensor("v_11_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor v_11_squeeze_mask_0 = const()[name = tensor("v_11_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor v_11_cast_fp16 = slice_by_index(begin = v_11_begin_0, end = v_11_end_0, end_mask = v_11_end_mask_0, squeeze_mask = v_11_squeeze_mask_0, x = qkv_11_cast_fp16)[name = tensor("v_11_cast_fp16")]; + tensor var_2214 = const()[name = tensor("op_2214"), val = tensor([1, 1, 1, 1])]; + tensor position5_to_fp16_dtype_0 = const()[name = tensor("position5_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor position5_to_fp16 = cast(dtype = position5_to_fp16_dtype_0, x = position5)[name = tensor("cast_528")]; + tensor ts_35_cast_fp16 = reshape(shape = var_2214, x = position5_to_fp16)[name = tensor("ts_35_cast_fp16")]; + tensor var_2221 = const()[name = tensor("op_2221"), val = tensor([1, 1, 16, 32, 2])]; + tensor q_complex_11_cast_fp16 = reshape(shape = var_2221, x = q_31_cast_fp16)[name = tensor("q_complex_11_cast_fp16")]; + tensor var_2228 = const()[name = tensor("op_2228"), val = tensor([1, 1, 16, 32, 2])]; + tensor k_complex_11_cast_fp16 = reshape(shape = var_2228, x = k_21_cast_fp16)[name = tensor("k_complex_11_cast_fp16")]; + tensor var_2232_begin_0 = const()[name = tensor("op_2232_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_2232_end_0 = const()[name = tensor("op_2232_end_0"), val = tensor([1, 1, 16, 32, 1])]; + tensor var_2232_end_mask_0 = const()[name = tensor("op_2232_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_2232_squeeze_mask_0 = const()[name = tensor("op_2232_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_2232_cast_fp16 = slice_by_index(begin = var_2232_begin_0, end = var_2232_end_0, end_mask = var_2232_end_mask_0, squeeze_mask = var_2232_squeeze_mask_0, x = q_complex_11_cast_fp16)[name = tensor("op_2232_cast_fp16")]; + tensor var_2240_begin_0 = const()[name = tensor("op_2240_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_2240_end_0 = const()[name = tensor("op_2240_end_0"), val = tensor([1, 1, 16, 32, 2])]; + tensor var_2240_end_mask_0 = const()[name = tensor("op_2240_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_2240_squeeze_mask_0 = const()[name = tensor("op_2240_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_2240_cast_fp16 = slice_by_index(begin = var_2240_begin_0, end = var_2240_end_0, end_mask = var_2240_end_mask_0, squeeze_mask = var_2240_squeeze_mask_0, x = q_complex_11_cast_fp16)[name = tensor("op_2240_cast_fp16")]; + tensor var_2248_begin_0 = const()[name = tensor("op_2248_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_2248_end_0 = const()[name = tensor("op_2248_end_0"), val = tensor([1, 1, 16, 32, 1])]; + tensor var_2248_end_mask_0 = const()[name = tensor("op_2248_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_2248_squeeze_mask_0 = const()[name = tensor("op_2248_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_2248_cast_fp16 = slice_by_index(begin = var_2248_begin_0, end = var_2248_end_0, end_mask = var_2248_end_mask_0, squeeze_mask = var_2248_squeeze_mask_0, x = k_complex_11_cast_fp16)[name = tensor("op_2248_cast_fp16")]; + tensor var_2256_begin_0 = const()[name = tensor("op_2256_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_2256_end_0 = const()[name = tensor("op_2256_end_0"), val = tensor([1, 1, 16, 32, 2])]; + tensor var_2256_end_mask_0 = const()[name = tensor("op_2256_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_2256_squeeze_mask_0 = const()[name = tensor("op_2256_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_2256_cast_fp16 = slice_by_index(begin = var_2256_begin_0, end = var_2256_end_0, end_mask = var_2256_end_mask_0, squeeze_mask = var_2256_squeeze_mask_0, x = k_complex_11_cast_fp16)[name = tensor("op_2256_cast_fp16")]; + tensor freqs_11_to_fp16 = const()[name = tensor("freqs_11_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(133300992)))]; + tensor var_2262_cast_fp16 = mul(x = freqs_11_to_fp16, y = ts_35_cast_fp16)[name = tensor("op_2262_cast_fp16")]; + tensor rotr_11_cast_fp16 = cos(x = var_2262_cast_fp16)[name = tensor("rotr_11_cast_fp16")]; + tensor roti_11_cast_fp16 = sin(x = var_2262_cast_fp16)[name = tensor("roti_11_cast_fp16")]; + tensor var_2266_cast_fp16 = mul(x = var_2232_cast_fp16, y = rotr_11_cast_fp16)[name = tensor("op_2266_cast_fp16")]; + tensor var_2267_cast_fp16 = mul(x = var_2240_cast_fp16, y = roti_11_cast_fp16)[name = tensor("op_2267_cast_fp16")]; + tensor qor_21_cast_fp16 = sub(x = var_2266_cast_fp16, y = var_2267_cast_fp16)[name = tensor("qor_21_cast_fp16")]; + tensor var_2270_cast_fp16 = mul(x = var_2232_cast_fp16, y = roti_11_cast_fp16)[name = tensor("op_2270_cast_fp16")]; + tensor var_2271_cast_fp16 = mul(x = var_2240_cast_fp16, y = rotr_11_cast_fp16)[name = tensor("op_2271_cast_fp16")]; + tensor qoi_21_cast_fp16 = add(x = var_2270_cast_fp16, y = var_2271_cast_fp16)[name = tensor("qoi_21_cast_fp16")]; + tensor var_2274_cast_fp16 = mul(x = var_2248_cast_fp16, y = rotr_11_cast_fp16)[name = tensor("op_2274_cast_fp16")]; + tensor var_2275_cast_fp16 = mul(x = var_2256_cast_fp16, y = roti_11_cast_fp16)[name = tensor("op_2275_cast_fp16")]; + tensor kor_21_cast_fp16 = sub(x = var_2274_cast_fp16, y = var_2275_cast_fp16)[name = tensor("kor_21_cast_fp16")]; + tensor var_2278_cast_fp16 = mul(x = var_2248_cast_fp16, y = roti_11_cast_fp16)[name = tensor("op_2278_cast_fp16")]; + tensor var_2279_cast_fp16 = mul(x = var_2256_cast_fp16, y = rotr_11_cast_fp16)[name = tensor("op_2279_cast_fp16")]; + tensor koi_21_cast_fp16 = add(x = var_2278_cast_fp16, y = var_2279_cast_fp16)[name = tensor("koi_21_cast_fp16")]; + tensor qo_11_axis_0 = const()[name = tensor("qo_11_axis_0"), val = tensor(-1)]; + tensor qo_11_cast_fp16 = stack(axis = qo_11_axis_0, values = (qor_21_cast_fp16, qoi_21_cast_fp16))[name = tensor("qo_11_cast_fp16")]; + tensor ko_11_axis_0 = const()[name = tensor("ko_11_axis_0"), val = tensor(-1)]; + tensor ko_11_cast_fp16 = stack(axis = ko_11_axis_0, values = (kor_21_cast_fp16, koi_21_cast_fp16))[name = tensor("ko_11_cast_fp16")]; + tensor var_2312 = const()[name = tensor("op_2312"), val = tensor([1, 1, 16, 64])]; + tensor q_33_cast_fp16 = reshape(shape = var_2312, x = qo_11_cast_fp16)[name = tensor("q_33_cast_fp16")]; + tensor var_2318 = const()[name = tensor("op_2318"), val = tensor([1, 1, 16, 64])]; + tensor k_23_cast_fp16 = reshape(shape = var_2318, x = ko_11_cast_fp16)[name = tensor("k_23_cast_fp16")]; + tensor _inversed_2343_y_0_to_fp16 = const()[name = tensor("_inversed_2343_y_0_to_fp16"), val = tensor(0x1p-9)]; + tensor _inversed_2343_cast_fp16 = mul(x = ts_35_cast_fp16, y = _inversed_2343_y_0_to_fp16)[name = tensor("_inversed_2343_cast_fp16")]; + tensor var_2344_cast_fp16 = floor(x = _inversed_2343_cast_fp16)[name = tensor("op_2344_cast_fp16")]; + tensor var_2345_to_fp16 = const()[name = tensor("op_2345_to_fp16"), val = tensor(0x1p+9)]; + tensor var_2346_cast_fp16 = mul(x = var_2344_cast_fp16, y = var_2345_to_fp16)[name = tensor("op_2346_cast_fp16")]; + tensor write_indices_float_23_cast_fp16 = sub(x = ts_35_cast_fp16, y = var_2346_cast_fp16)[name = tensor("write_indices_float_23_cast_fp16")]; + tensor var_2353_dtype_0 = const()[name = tensor("op_2353_dtype_0"), val = tensor("int32")]; + tensor write_indices_11_reps_0 = const()[name = tensor("write_indices_11_reps_0"), val = tensor([1, 1, 16, 64])]; + tensor write_indices_float_23_cast_fp16_to_int32 = cast(dtype = var_2353_dtype_0, x = write_indices_float_23_cast_fp16)[name = tensor("cast_527")]; + tensor write_indices_11 = tile(reps = write_indices_11_reps_0, x = write_indices_float_23_cast_fp16_to_int32)[name = tensor("write_indices_11")]; + tensor var_2363_begin_0 = const()[name = tensor("op_2363_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_2363_end_0 = const()[name = tensor("op_2363_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor var_2363_end_mask_0 = const()[name = tensor("op_2363_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_2363_squeeze_mask_0 = const()[name = tensor("op_2363_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor cache5_to_fp16_dtype_0 = const()[name = tensor("cache5_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor cache5_to_fp16 = cast(dtype = cache5_to_fp16_dtype_0, x = cache5)[name = tensor("cast_526")]; + tensor var_2363_cast_fp16 = slice_by_index(begin = var_2363_begin_0, end = var_2363_end_0, end_mask = var_2363_end_mask_0, squeeze_mask = var_2363_squeeze_mask_0, x = cache5_to_fp16)[name = tensor("op_2363_cast_fp16")]; + tensor var_2365_axis_0 = const()[name = tensor("op_2365_axis_0"), val = tensor(1)]; + tensor var_2365_mode_0 = const()[name = tensor("op_2365_mode_0"), val = tensor("update")]; + tensor var_2365_validate_indices_0 = const()[name = tensor("op_2365_validate_indices_0"), val = tensor(false)]; + tensor var_2365_cast_fp16 = scatter_along_axis(axis = var_2365_axis_0, data = var_2363_cast_fp16, indices = write_indices_11, mode = var_2365_mode_0, updates = k_23_cast_fp16, validate_indices = var_2365_validate_indices_0)[name = tensor("op_2365_cast_fp16")]; + tensor concat_37 = const()[name = tensor("concat_37"), val = tensor([0, 0, 0, 0, 0])]; + tensor concat_38 = const()[name = tensor("concat_38"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_11_internal_tensor_assign_1_stride_0 = const()[name = tensor("new_cache_11_internal_tensor_assign_1_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_11_internal_tensor_assign_1_begin_mask_0 = const()[name = tensor("new_cache_11_internal_tensor_assign_1_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_11_internal_tensor_assign_1_end_mask_0 = const()[name = tensor("new_cache_11_internal_tensor_assign_1_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_11_internal_tensor_assign_1_squeeze_mask_0 = const()[name = tensor("new_cache_11_internal_tensor_assign_1_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_58 = const()[name = tensor("shape_58"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_10 = const()[name = tensor("reduce_prod_10"), val = tensor(1048576)]; + tensor range_1d_10_start_0 = const()[name = tensor("range_1d_10_start_0"), val = tensor(0)]; + tensor range_1d_10_step_0 = const()[name = tensor("range_1d_10_step_0"), val = tensor(1)]; + tensor range_1d_10 = range_1d(end = reduce_prod_10, start = range_1d_10_start_0, step = range_1d_10_step_0)[name = tensor("range_1d_10")]; + tensor reshape_50 = reshape(shape = shape_58, x = range_1d_10)[name = tensor("reshape_50")]; + tensor slice_by_index_10 = slice_by_index(begin = concat_37, begin_mask = new_cache_11_internal_tensor_assign_1_begin_mask_0, end = concat_38, end_mask = new_cache_11_internal_tensor_assign_1_end_mask_0, squeeze_mask = new_cache_11_internal_tensor_assign_1_squeeze_mask_0, stride = new_cache_11_internal_tensor_assign_1_stride_0, x = reshape_50)[name = tensor("slice_by_index_10")]; + tensor reshape_51_shape_0 = const()[name = tensor("reshape_51_shape_0"), val = tensor([-1])]; + tensor reshape_51 = reshape(shape = reshape_51_shape_0, x = slice_by_index_10)[name = tensor("reshape_51")]; + tensor reshape_52_shape_0 = const()[name = tensor("reshape_52_shape_0"), val = tensor([-1])]; + tensor reshape_52_cast_fp16 = reshape(shape = reshape_52_shape_0, x = var_2365_cast_fp16)[name = tensor("reshape_52_cast_fp16")]; + tensor reshape_53_shape_0 = const()[name = tensor("reshape_53_shape_0"), val = tensor([-1])]; + tensor reshape_53_cast_fp16 = reshape(shape = reshape_53_shape_0, x = cache5_to_fp16)[name = tensor("reshape_53_cast_fp16")]; + tensor scatter_10_mode_0 = const()[name = tensor("scatter_10_mode_0"), val = tensor("update")]; + tensor scatter_10_axis_0 = const()[name = tensor("scatter_10_axis_0"), val = tensor(0)]; + tensor scatter_10_validate_indices_0 = const()[name = tensor("scatter_10_validate_indices_0"), val = tensor(false)]; + tensor scatter_10_cast_fp16 = scatter(axis = scatter_10_axis_0, data = reshape_53_cast_fp16, indices = reshape_51, mode = scatter_10_mode_0, updates = reshape_52_cast_fp16, validate_indices = scatter_10_validate_indices_0)[name = tensor("scatter_10_cast_fp16")]; + tensor reshape_54_cast_fp16 = reshape(shape = shape_58, x = scatter_10_cast_fp16)[name = tensor("reshape_54_cast_fp16")]; + tensor var_2373_begin_0 = const()[name = tensor("op_2373_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor var_2373_end_0 = const()[name = tensor("op_2373_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor var_2373_end_mask_0 = const()[name = tensor("op_2373_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_2373_squeeze_mask_0 = const()[name = tensor("op_2373_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor var_2373_cast_fp16 = slice_by_index(begin = var_2373_begin_0, end = var_2373_end_0, end_mask = var_2373_end_mask_0, squeeze_mask = var_2373_squeeze_mask_0, x = reshape_54_cast_fp16)[name = tensor("op_2373_cast_fp16")]; + tensor var_2375_axis_0 = const()[name = tensor("op_2375_axis_0"), val = tensor(1)]; + tensor var_2375_mode_0 = const()[name = tensor("op_2375_mode_0"), val = tensor("update")]; + tensor var_2375_validate_indices_0 = const()[name = tensor("op_2375_validate_indices_0"), val = tensor(false)]; + tensor var_2375_cast_fp16 = scatter_along_axis(axis = var_2375_axis_0, data = var_2373_cast_fp16, indices = write_indices_11, mode = var_2375_mode_0, updates = v_11_cast_fp16, validate_indices = var_2375_validate_indices_0)[name = tensor("op_2375_cast_fp16")]; + tensor concat_39 = const()[name = tensor("concat_39"), val = tensor([1, 0, 0, 0, 0])]; + tensor concat_40 = const()[name = tensor("concat_40"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_11_internal_tensor_assign_2_stride_0 = const()[name = tensor("new_cache_11_internal_tensor_assign_2_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_11_internal_tensor_assign_2_begin_mask_0 = const()[name = tensor("new_cache_11_internal_tensor_assign_2_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_11_internal_tensor_assign_2_end_mask_0 = const()[name = tensor("new_cache_11_internal_tensor_assign_2_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_11_internal_tensor_assign_2_squeeze_mask_0 = const()[name = tensor("new_cache_11_internal_tensor_assign_2_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_59 = const()[name = tensor("shape_59"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_11 = const()[name = tensor("reduce_prod_11"), val = tensor(1048576)]; + tensor range_1d_11_start_0 = const()[name = tensor("range_1d_11_start_0"), val = tensor(0)]; + tensor range_1d_11_step_0 = const()[name = tensor("range_1d_11_step_0"), val = tensor(1)]; + tensor range_1d_11 = range_1d(end = reduce_prod_11, start = range_1d_11_start_0, step = range_1d_11_step_0)[name = tensor("range_1d_11")]; + tensor reshape_55 = reshape(shape = shape_59, x = range_1d_11)[name = tensor("reshape_55")]; + tensor slice_by_index_11 = slice_by_index(begin = concat_39, begin_mask = new_cache_11_internal_tensor_assign_2_begin_mask_0, end = concat_40, end_mask = new_cache_11_internal_tensor_assign_2_end_mask_0, squeeze_mask = new_cache_11_internal_tensor_assign_2_squeeze_mask_0, stride = new_cache_11_internal_tensor_assign_2_stride_0, x = reshape_55)[name = tensor("slice_by_index_11")]; + tensor reshape_56_shape_0 = const()[name = tensor("reshape_56_shape_0"), val = tensor([-1])]; + tensor reshape_56 = reshape(shape = reshape_56_shape_0, x = slice_by_index_11)[name = tensor("reshape_56")]; + tensor reshape_57_shape_0 = const()[name = tensor("reshape_57_shape_0"), val = tensor([-1])]; + tensor reshape_57_cast_fp16 = reshape(shape = reshape_57_shape_0, x = var_2375_cast_fp16)[name = tensor("reshape_57_cast_fp16")]; + tensor reshape_58_shape_0 = const()[name = tensor("reshape_58_shape_0"), val = tensor([-1])]; + tensor reshape_58_cast_fp16 = reshape(shape = reshape_58_shape_0, x = reshape_54_cast_fp16)[name = tensor("reshape_58_cast_fp16")]; + tensor scatter_11_mode_0 = const()[name = tensor("scatter_11_mode_0"), val = tensor("update")]; + tensor scatter_11_axis_0 = const()[name = tensor("scatter_11_axis_0"), val = tensor(0)]; + tensor scatter_11_validate_indices_0 = const()[name = tensor("scatter_11_validate_indices_0"), val = tensor(false)]; + tensor scatter_11_cast_fp16 = scatter(axis = scatter_11_axis_0, data = reshape_58_cast_fp16, indices = reshape_56, mode = scatter_11_mode_0, updates = reshape_57_cast_fp16, validate_indices = scatter_11_validate_indices_0)[name = tensor("scatter_11_cast_fp16")]; + tensor reshape_59_cast_fp16 = reshape(shape = shape_59, x = scatter_11_cast_fp16)[name = tensor("reshape_59_cast_fp16")]; + tensor reshape_59_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("reshape_59_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor keys_31_begin_0 = const()[name = tensor("keys_31_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor keys_31_end_0 = const()[name = tensor("keys_31_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor keys_31_end_mask_0 = const()[name = tensor("keys_31_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor keys_31_squeeze_mask_0 = const()[name = tensor("keys_31_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor keys_31_cast_fp16 = slice_by_index(begin = keys_31_begin_0, end = keys_31_end_0, end_mask = keys_31_end_mask_0, squeeze_mask = keys_31_squeeze_mask_0, x = reshape_59_cast_fp16)[name = tensor("keys_31_cast_fp16")]; + tensor values_31_begin_0 = const()[name = tensor("values_31_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor values_31_end_0 = const()[name = tensor("values_31_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor values_31_end_mask_0 = const()[name = tensor("values_31_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor values_31_squeeze_mask_0 = const()[name = tensor("values_31_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor values_31_cast_fp16 = slice_by_index(begin = values_31_begin_0, end = values_31_end_0, end_mask = values_31_end_mask_0, squeeze_mask = values_31_squeeze_mask_0, x = reshape_59_cast_fp16)[name = tensor("values_31_cast_fp16")]; + tensor var_2387_cast_fp16 = not_equal(x = keys_31_cast_fp16, y = keys_31_cast_fp16)[name = tensor("op_2387_cast_fp16")]; + tensor keys_33_cast_fp16 = select(a = var_463_to_fp16, b = keys_31_cast_fp16, cond = var_2387_cast_fp16)[name = tensor("keys_33_cast_fp16")]; + tensor var_2395_cast_fp16 = not_equal(x = values_31_cast_fp16, y = values_31_cast_fp16)[name = tensor("op_2395_cast_fp16")]; + tensor values_33_cast_fp16 = select(a = var_463_to_fp16, b = values_31_cast_fp16, cond = var_2395_cast_fp16)[name = tensor("values_33_cast_fp16")]; + tensor var_2419 = const()[name = tensor("op_2419"), val = tensor([0, 2, 1, 3])]; + tensor var_2435 = const()[name = tensor("op_2435"), val = tensor([1, 1, 1])]; + tensor var_2436_cast_fp16 = reshape(shape = var_2435, x = position5_to_fp16)[name = tensor("op_2436_cast_fp16")]; + tensor var_2456_to_fp16 = const()[name = tensor("op_2456_to_fp16"), val = tensor(0x1p+0)]; + tensor valid_len_11_cast_fp16 = add(x = var_2436_cast_fp16, y = var_2456_to_fp16)[name = tensor("valid_len_11_cast_fp16")]; + tensor valid_mask_11_cast_fp16 = less(x = k_positions_1_promoted_to_fp16, y = valid_len_11_cast_fp16)[name = tensor("valid_mask_11_cast_fp16")]; + tensor causal_mask_11_cast_fp16 = less_equal(x = k_positions_1_promoted_to_fp16, y = var_2436_cast_fp16)[name = tensor("causal_mask_11_cast_fp16")]; + tensor attn_mask_21 = logical_and(x = valid_mask_11_cast_fp16, y = causal_mask_11_cast_fp16)[name = tensor("attn_mask_21")]; + tensor attn_mask_23_axes_0 = const()[name = tensor("attn_mask_23_axes_0"), val = tensor([1])]; + tensor attn_mask_23 = expand_dims(axes = attn_mask_23_axes_0, x = attn_mask_21)[name = tensor("attn_mask_23")]; + tensor var_2474_transpose_x_0 = const()[name = tensor("op_2474_transpose_x_0"), val = tensor(false)]; + tensor var_2474_transpose_y_0 = const()[name = tensor("op_2474_transpose_y_0"), val = tensor(false)]; + tensor transpose_82_perm_0 = const()[name = tensor("transpose_82_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_83_perm_0 = const()[name = tensor("transpose_83_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_83 = transpose(perm = transpose_83_perm_0, x = keys_33_cast_fp16)[name = tensor("transpose_193")]; + tensor transpose_82 = transpose(perm = transpose_82_perm_0, x = q_33_cast_fp16)[name = tensor("transpose_194")]; + tensor var_2474_cast_fp16 = matmul(transpose_x = var_2474_transpose_x_0, transpose_y = var_2474_transpose_y_0, x = transpose_82, y = transpose_83)[name = tensor("op_2474_cast_fp16")]; + tensor var_2468_to_fp16 = const()[name = tensor("op_2468_to_fp16"), val = tensor([0x1p-3])]; + tensor attn_weights_31_cast_fp16 = mul(x = var_2474_cast_fp16, y = var_2468_to_fp16)[name = tensor("attn_weights_31_cast_fp16")]; + tensor var_2476 = logical_not(x = attn_mask_23)[name = tensor("op_2476")]; + tensor var_2477_to_fp16 = const()[name = tensor("op_2477_to_fp16"), val = tensor(-inf)]; + tensor attn_weights_33_cast_fp16 = select(a = var_2477_to_fp16, b = attn_weights_31_cast_fp16, cond = var_2476)[name = tensor("attn_weights_33_cast_fp16")]; + tensor var_2479 = const()[name = tensor("op_2479"), val = tensor(-1)]; + tensor attn_weights_35_cast_fp16 = softmax(axis = var_2479, x = attn_weights_33_cast_fp16)[name = tensor("attn_weights_35_cast_fp16")]; + tensor attn_output_11_transpose_x_0 = const()[name = tensor("attn_output_11_transpose_x_0"), val = tensor(false)]; + tensor attn_output_11_transpose_y_0 = const()[name = tensor("attn_output_11_transpose_y_0"), val = tensor(false)]; + tensor values_35_cast_fp16 = transpose(perm = var_2419, x = values_33_cast_fp16)[name = tensor("transpose_195")]; + tensor attn_output_11_cast_fp16 = matmul(transpose_x = attn_output_11_transpose_x_0, transpose_y = attn_output_11_transpose_y_0, x = attn_weights_35_cast_fp16, y = values_35_cast_fp16)[name = tensor("attn_output_11_cast_fp16")]; + tensor var_2487 = const()[name = tensor("op_2487"), val = tensor([0, 2, 1, 3])]; + tensor var_2492 = const()[name = tensor("op_2492"), val = tensor([1, 1, 1024])]; + tensor var_2488_cast_fp16 = transpose(perm = var_2487, x = attn_output_11_cast_fp16)[name = tensor("transpose_192")]; + tensor input_67_cast_fp16 = reshape(shape = var_2492, x = var_2488_cast_fp16)[name = tensor("input_67_cast_fp16")]; + tensor attn5_out_proj_weight_to_fp16 = const()[name = tensor("attn5_out_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(133301120)))]; + tensor linear_22_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = attn5_out_proj_weight_to_fp16, x = input_67_cast_fp16)[name = tensor("linear_22_cast_fp16")]; + tensor var_2498_to_fp16 = const()[name = tensor("op_2498_to_fp16"), val = tensor(0x1p+0)]; + tensor var_2499_cast_fp16 = add(x = position5_to_fp16, y = var_2498_to_fp16)[name = tensor("op_2499_cast_fp16")]; + tensor var_2499_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_2499_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor input_69_cast_fp16 = add(x = input_63_cast_fp16, y = linear_22_cast_fp16)[name = tensor("input_69_cast_fp16")]; + tensor input_71_axes_0 = const()[name = tensor("input_71_axes_0"), val = tensor([-1])]; + tensor norm5_2_weight_to_fp16 = const()[name = tensor("norm5_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(135398336)))]; + tensor norm5_2_bias_to_fp16 = const()[name = tensor("norm5_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(135400448)))]; + tensor var_2503_to_fp16 = const()[name = tensor("op_2503_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_71_cast_fp16 = layer_norm(axes = input_71_axes_0, beta = norm5_2_bias_to_fp16, epsilon = var_2503_to_fp16, gamma = norm5_2_weight_to_fp16, x = input_69_cast_fp16)[name = tensor("input_71_cast_fp16")]; + tensor linear5_1_weight_to_fp16 = const()[name = tensor("linear5_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(135402560)))]; + tensor linear_23_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = linear5_1_weight_to_fp16, x = input_71_cast_fp16)[name = tensor("linear_23_cast_fp16")]; + tensor input_73_mode_0 = const()[name = tensor("input_73_mode_0"), val = tensor("EXACT")]; + tensor input_73_cast_fp16 = gelu(mode = input_73_mode_0, x = linear_23_cast_fp16)[name = tensor("input_73_cast_fp16")]; + tensor linear5_2_weight_to_fp16 = const()[name = tensor("linear5_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(143791232)))]; + tensor linear_24_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = linear5_2_weight_to_fp16, x = input_73_cast_fp16)[name = tensor("linear_24_cast_fp16")]; + tensor input_75_cast_fp16 = add(x = input_69_cast_fp16, y = linear_24_cast_fp16)[name = tensor("input_75_cast_fp16")]; + tensor input_77_axes_0 = const()[name = tensor("input_77_axes_0"), val = tensor([-1])]; + tensor norm6_1_weight_to_fp16 = const()[name = tensor("norm6_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(152179904)))]; + tensor norm6_1_bias_to_fp16 = const()[name = tensor("norm6_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(152182016)))]; + tensor var_2520_to_fp16 = const()[name = tensor("op_2520_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_77_cast_fp16 = layer_norm(axes = input_77_axes_0, beta = norm6_1_bias_to_fp16, epsilon = var_2520_to_fp16, gamma = norm6_1_weight_to_fp16, x = input_75_cast_fp16)[name = tensor("input_77_cast_fp16")]; + tensor attn6_in_proj_weight_to_fp16 = const()[name = tensor("attn6_in_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(152184128)))]; + tensor linear_25_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = attn6_in_proj_weight_to_fp16, x = input_77_cast_fp16)[name = tensor("linear_25_cast_fp16")]; + tensor var_2534 = const()[name = tensor("op_2534"), val = tensor([1, 1, 3, 16, 64])]; + tensor qkv_13_cast_fp16 = reshape(shape = var_2534, x = linear_25_cast_fp16)[name = tensor("qkv_13_cast_fp16")]; + tensor q_37_begin_0 = const()[name = tensor("q_37_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor q_37_end_0 = const()[name = tensor("q_37_end_0"), val = tensor([1, 1, 1, 16, 64])]; + tensor q_37_end_mask_0 = const()[name = tensor("q_37_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor q_37_squeeze_mask_0 = const()[name = tensor("q_37_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor q_37_cast_fp16 = slice_by_index(begin = q_37_begin_0, end = q_37_end_0, end_mask = q_37_end_mask_0, squeeze_mask = q_37_squeeze_mask_0, x = qkv_13_cast_fp16)[name = tensor("q_37_cast_fp16")]; + tensor k_25_begin_0 = const()[name = tensor("k_25_begin_0"), val = tensor([0, 0, 1, 0, 0])]; + tensor k_25_end_0 = const()[name = tensor("k_25_end_0"), val = tensor([1, 1, 2, 16, 64])]; + tensor k_25_end_mask_0 = const()[name = tensor("k_25_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor k_25_squeeze_mask_0 = const()[name = tensor("k_25_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor k_25_cast_fp16 = slice_by_index(begin = k_25_begin_0, end = k_25_end_0, end_mask = k_25_end_mask_0, squeeze_mask = k_25_squeeze_mask_0, x = qkv_13_cast_fp16)[name = tensor("k_25_cast_fp16")]; + tensor v_13_begin_0 = const()[name = tensor("v_13_begin_0"), val = tensor([0, 0, 2, 0, 0])]; + tensor v_13_end_0 = const()[name = tensor("v_13_end_0"), val = tensor([1, 1, 3, 16, 64])]; + tensor v_13_end_mask_0 = const()[name = tensor("v_13_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor v_13_squeeze_mask_0 = const()[name = tensor("v_13_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor v_13_cast_fp16 = slice_by_index(begin = v_13_begin_0, end = v_13_end_0, end_mask = v_13_end_mask_0, squeeze_mask = v_13_squeeze_mask_0, x = qkv_13_cast_fp16)[name = tensor("v_13_cast_fp16")]; + tensor var_2600 = const()[name = tensor("op_2600"), val = tensor([1, 1, 1, 1])]; + tensor position6_to_fp16_dtype_0 = const()[name = tensor("position6_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor position6_to_fp16 = cast(dtype = position6_to_fp16_dtype_0, x = position6)[name = tensor("cast_523")]; + tensor ts_41_cast_fp16 = reshape(shape = var_2600, x = position6_to_fp16)[name = tensor("ts_41_cast_fp16")]; + tensor var_2607 = const()[name = tensor("op_2607"), val = tensor([1, 1, 16, 32, 2])]; + tensor q_complex_13_cast_fp16 = reshape(shape = var_2607, x = q_37_cast_fp16)[name = tensor("q_complex_13_cast_fp16")]; + tensor var_2614 = const()[name = tensor("op_2614"), val = tensor([1, 1, 16, 32, 2])]; + tensor k_complex_13_cast_fp16 = reshape(shape = var_2614, x = k_25_cast_fp16)[name = tensor("k_complex_13_cast_fp16")]; + tensor var_2618_begin_0 = const()[name = tensor("op_2618_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_2618_end_0 = const()[name = tensor("op_2618_end_0"), val = tensor([1, 1, 16, 32, 1])]; + tensor var_2618_end_mask_0 = const()[name = tensor("op_2618_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_2618_squeeze_mask_0 = const()[name = tensor("op_2618_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_2618_cast_fp16 = slice_by_index(begin = var_2618_begin_0, end = var_2618_end_0, end_mask = var_2618_end_mask_0, squeeze_mask = var_2618_squeeze_mask_0, x = q_complex_13_cast_fp16)[name = tensor("op_2618_cast_fp16")]; + tensor var_2626_begin_0 = const()[name = tensor("op_2626_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_2626_end_0 = const()[name = tensor("op_2626_end_0"), val = tensor([1, 1, 16, 32, 2])]; + tensor var_2626_end_mask_0 = const()[name = tensor("op_2626_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_2626_squeeze_mask_0 = const()[name = tensor("op_2626_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_2626_cast_fp16 = slice_by_index(begin = var_2626_begin_0, end = var_2626_end_0, end_mask = var_2626_end_mask_0, squeeze_mask = var_2626_squeeze_mask_0, x = q_complex_13_cast_fp16)[name = tensor("op_2626_cast_fp16")]; + tensor var_2634_begin_0 = const()[name = tensor("op_2634_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_2634_end_0 = const()[name = tensor("op_2634_end_0"), val = tensor([1, 1, 16, 32, 1])]; + tensor var_2634_end_mask_0 = const()[name = tensor("op_2634_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_2634_squeeze_mask_0 = const()[name = tensor("op_2634_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_2634_cast_fp16 = slice_by_index(begin = var_2634_begin_0, end = var_2634_end_0, end_mask = var_2634_end_mask_0, squeeze_mask = var_2634_squeeze_mask_0, x = k_complex_13_cast_fp16)[name = tensor("op_2634_cast_fp16")]; + tensor var_2642_begin_0 = const()[name = tensor("op_2642_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_2642_end_0 = const()[name = tensor("op_2642_end_0"), val = tensor([1, 1, 16, 32, 2])]; + tensor var_2642_end_mask_0 = const()[name = tensor("op_2642_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_2642_squeeze_mask_0 = const()[name = tensor("op_2642_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_2642_cast_fp16 = slice_by_index(begin = var_2642_begin_0, end = var_2642_end_0, end_mask = var_2642_end_mask_0, squeeze_mask = var_2642_squeeze_mask_0, x = k_complex_13_cast_fp16)[name = tensor("op_2642_cast_fp16")]; + tensor freqs_13_to_fp16 = const()[name = tensor("freqs_13_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(158475648)))]; + tensor var_2648_cast_fp16 = mul(x = freqs_13_to_fp16, y = ts_41_cast_fp16)[name = tensor("op_2648_cast_fp16")]; + tensor rotr_13_cast_fp16 = cos(x = var_2648_cast_fp16)[name = tensor("rotr_13_cast_fp16")]; + tensor roti_13_cast_fp16 = sin(x = var_2648_cast_fp16)[name = tensor("roti_13_cast_fp16")]; + tensor var_2652_cast_fp16 = mul(x = var_2618_cast_fp16, y = rotr_13_cast_fp16)[name = tensor("op_2652_cast_fp16")]; + tensor var_2653_cast_fp16 = mul(x = var_2626_cast_fp16, y = roti_13_cast_fp16)[name = tensor("op_2653_cast_fp16")]; + tensor qor_25_cast_fp16 = sub(x = var_2652_cast_fp16, y = var_2653_cast_fp16)[name = tensor("qor_25_cast_fp16")]; + tensor var_2656_cast_fp16 = mul(x = var_2618_cast_fp16, y = roti_13_cast_fp16)[name = tensor("op_2656_cast_fp16")]; + tensor var_2657_cast_fp16 = mul(x = var_2626_cast_fp16, y = rotr_13_cast_fp16)[name = tensor("op_2657_cast_fp16")]; + tensor qoi_25_cast_fp16 = add(x = var_2656_cast_fp16, y = var_2657_cast_fp16)[name = tensor("qoi_25_cast_fp16")]; + tensor var_2660_cast_fp16 = mul(x = var_2634_cast_fp16, y = rotr_13_cast_fp16)[name = tensor("op_2660_cast_fp16")]; + tensor var_2661_cast_fp16 = mul(x = var_2642_cast_fp16, y = roti_13_cast_fp16)[name = tensor("op_2661_cast_fp16")]; + tensor kor_25_cast_fp16 = sub(x = var_2660_cast_fp16, y = var_2661_cast_fp16)[name = tensor("kor_25_cast_fp16")]; + tensor var_2664_cast_fp16 = mul(x = var_2634_cast_fp16, y = roti_13_cast_fp16)[name = tensor("op_2664_cast_fp16")]; + tensor var_2665_cast_fp16 = mul(x = var_2642_cast_fp16, y = rotr_13_cast_fp16)[name = tensor("op_2665_cast_fp16")]; + tensor koi_25_cast_fp16 = add(x = var_2664_cast_fp16, y = var_2665_cast_fp16)[name = tensor("koi_25_cast_fp16")]; + tensor qo_13_axis_0 = const()[name = tensor("qo_13_axis_0"), val = tensor(-1)]; + tensor qo_13_cast_fp16 = stack(axis = qo_13_axis_0, values = (qor_25_cast_fp16, qoi_25_cast_fp16))[name = tensor("qo_13_cast_fp16")]; + tensor ko_13_axis_0 = const()[name = tensor("ko_13_axis_0"), val = tensor(-1)]; + tensor ko_13_cast_fp16 = stack(axis = ko_13_axis_0, values = (kor_25_cast_fp16, koi_25_cast_fp16))[name = tensor("ko_13_cast_fp16")]; + tensor var_2698 = const()[name = tensor("op_2698"), val = tensor([1, 1, 16, 64])]; + tensor q_39_cast_fp16 = reshape(shape = var_2698, x = qo_13_cast_fp16)[name = tensor("q_39_cast_fp16")]; + tensor var_2704 = const()[name = tensor("op_2704"), val = tensor([1, 1, 16, 64])]; + tensor k_27_cast_fp16 = reshape(shape = var_2704, x = ko_13_cast_fp16)[name = tensor("k_27_cast_fp16")]; + tensor _inversed_2729_y_0_to_fp16 = const()[name = tensor("_inversed_2729_y_0_to_fp16"), val = tensor(0x1p-9)]; + tensor _inversed_2729_cast_fp16 = mul(x = ts_41_cast_fp16, y = _inversed_2729_y_0_to_fp16)[name = tensor("_inversed_2729_cast_fp16")]; + tensor var_2730_cast_fp16 = floor(x = _inversed_2729_cast_fp16)[name = tensor("op_2730_cast_fp16")]; + tensor var_2731_to_fp16 = const()[name = tensor("op_2731_to_fp16"), val = tensor(0x1p+9)]; + tensor var_2732_cast_fp16 = mul(x = var_2730_cast_fp16, y = var_2731_to_fp16)[name = tensor("op_2732_cast_fp16")]; + tensor write_indices_float_27_cast_fp16 = sub(x = ts_41_cast_fp16, y = var_2732_cast_fp16)[name = tensor("write_indices_float_27_cast_fp16")]; + tensor var_2739_dtype_0 = const()[name = tensor("op_2739_dtype_0"), val = tensor("int32")]; + tensor write_indices_13_reps_0 = const()[name = tensor("write_indices_13_reps_0"), val = tensor([1, 1, 16, 64])]; + tensor write_indices_float_27_cast_fp16_to_int32 = cast(dtype = var_2739_dtype_0, x = write_indices_float_27_cast_fp16)[name = tensor("cast_522")]; + tensor write_indices_13 = tile(reps = write_indices_13_reps_0, x = write_indices_float_27_cast_fp16_to_int32)[name = tensor("write_indices_13")]; + tensor var_2749_begin_0 = const()[name = tensor("op_2749_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_2749_end_0 = const()[name = tensor("op_2749_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor var_2749_end_mask_0 = const()[name = tensor("op_2749_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_2749_squeeze_mask_0 = const()[name = tensor("op_2749_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor cache6_to_fp16_dtype_0 = const()[name = tensor("cache6_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor cache6_to_fp16 = cast(dtype = cache6_to_fp16_dtype_0, x = cache6)[name = tensor("cast_521")]; + tensor var_2749_cast_fp16 = slice_by_index(begin = var_2749_begin_0, end = var_2749_end_0, end_mask = var_2749_end_mask_0, squeeze_mask = var_2749_squeeze_mask_0, x = cache6_to_fp16)[name = tensor("op_2749_cast_fp16")]; + tensor var_2751_axis_0 = const()[name = tensor("op_2751_axis_0"), val = tensor(1)]; + tensor var_2751_mode_0 = const()[name = tensor("op_2751_mode_0"), val = tensor("update")]; + tensor var_2751_validate_indices_0 = const()[name = tensor("op_2751_validate_indices_0"), val = tensor(false)]; + tensor var_2751_cast_fp16 = scatter_along_axis(axis = var_2751_axis_0, data = var_2749_cast_fp16, indices = write_indices_13, mode = var_2751_mode_0, updates = k_27_cast_fp16, validate_indices = var_2751_validate_indices_0)[name = tensor("op_2751_cast_fp16")]; + tensor concat_44 = const()[name = tensor("concat_44"), val = tensor([0, 0, 0, 0, 0])]; + tensor concat_45 = const()[name = tensor("concat_45"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_13_internal_tensor_assign_1_stride_0 = const()[name = tensor("new_cache_13_internal_tensor_assign_1_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_13_internal_tensor_assign_1_begin_mask_0 = const()[name = tensor("new_cache_13_internal_tensor_assign_1_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_13_internal_tensor_assign_1_end_mask_0 = const()[name = tensor("new_cache_13_internal_tensor_assign_1_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_13_internal_tensor_assign_1_squeeze_mask_0 = const()[name = tensor("new_cache_13_internal_tensor_assign_1_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_60 = const()[name = tensor("shape_60"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_12 = const()[name = tensor("reduce_prod_12"), val = tensor(1048576)]; + tensor range_1d_12_start_0 = const()[name = tensor("range_1d_12_start_0"), val = tensor(0)]; + tensor range_1d_12_step_0 = const()[name = tensor("range_1d_12_step_0"), val = tensor(1)]; + tensor range_1d_12 = range_1d(end = reduce_prod_12, start = range_1d_12_start_0, step = range_1d_12_step_0)[name = tensor("range_1d_12")]; + tensor reshape_60 = reshape(shape = shape_60, x = range_1d_12)[name = tensor("reshape_60")]; + tensor slice_by_index_12 = slice_by_index(begin = concat_44, begin_mask = new_cache_13_internal_tensor_assign_1_begin_mask_0, end = concat_45, end_mask = new_cache_13_internal_tensor_assign_1_end_mask_0, squeeze_mask = new_cache_13_internal_tensor_assign_1_squeeze_mask_0, stride = new_cache_13_internal_tensor_assign_1_stride_0, x = reshape_60)[name = tensor("slice_by_index_12")]; + tensor reshape_61_shape_0 = const()[name = tensor("reshape_61_shape_0"), val = tensor([-1])]; + tensor reshape_61 = reshape(shape = reshape_61_shape_0, x = slice_by_index_12)[name = tensor("reshape_61")]; + tensor reshape_62_shape_0 = const()[name = tensor("reshape_62_shape_0"), val = tensor([-1])]; + tensor reshape_62_cast_fp16 = reshape(shape = reshape_62_shape_0, x = var_2751_cast_fp16)[name = tensor("reshape_62_cast_fp16")]; + tensor reshape_63_shape_0 = const()[name = tensor("reshape_63_shape_0"), val = tensor([-1])]; + tensor reshape_63_cast_fp16 = reshape(shape = reshape_63_shape_0, x = cache6_to_fp16)[name = tensor("reshape_63_cast_fp16")]; + tensor scatter_12_mode_0 = const()[name = tensor("scatter_12_mode_0"), val = tensor("update")]; + tensor scatter_12_axis_0 = const()[name = tensor("scatter_12_axis_0"), val = tensor(0)]; + tensor scatter_12_validate_indices_0 = const()[name = tensor("scatter_12_validate_indices_0"), val = tensor(false)]; + tensor scatter_12_cast_fp16 = scatter(axis = scatter_12_axis_0, data = reshape_63_cast_fp16, indices = reshape_61, mode = scatter_12_mode_0, updates = reshape_62_cast_fp16, validate_indices = scatter_12_validate_indices_0)[name = tensor("scatter_12_cast_fp16")]; + tensor reshape_64_cast_fp16 = reshape(shape = shape_60, x = scatter_12_cast_fp16)[name = tensor("reshape_64_cast_fp16")]; + tensor var_2759_begin_0 = const()[name = tensor("op_2759_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor var_2759_end_0 = const()[name = tensor("op_2759_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor var_2759_end_mask_0 = const()[name = tensor("op_2759_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_2759_squeeze_mask_0 = const()[name = tensor("op_2759_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor var_2759_cast_fp16 = slice_by_index(begin = var_2759_begin_0, end = var_2759_end_0, end_mask = var_2759_end_mask_0, squeeze_mask = var_2759_squeeze_mask_0, x = reshape_64_cast_fp16)[name = tensor("op_2759_cast_fp16")]; + tensor var_2761_axis_0 = const()[name = tensor("op_2761_axis_0"), val = tensor(1)]; + tensor var_2761_mode_0 = const()[name = tensor("op_2761_mode_0"), val = tensor("update")]; + tensor var_2761_validate_indices_0 = const()[name = tensor("op_2761_validate_indices_0"), val = tensor(false)]; + tensor var_2761_cast_fp16 = scatter_along_axis(axis = var_2761_axis_0, data = var_2759_cast_fp16, indices = write_indices_13, mode = var_2761_mode_0, updates = v_13_cast_fp16, validate_indices = var_2761_validate_indices_0)[name = tensor("op_2761_cast_fp16")]; + tensor concat_46 = const()[name = tensor("concat_46"), val = tensor([1, 0, 0, 0, 0])]; + tensor concat_47 = const()[name = tensor("concat_47"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_13_internal_tensor_assign_2_stride_0 = const()[name = tensor("new_cache_13_internal_tensor_assign_2_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_13_internal_tensor_assign_2_begin_mask_0 = const()[name = tensor("new_cache_13_internal_tensor_assign_2_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_13_internal_tensor_assign_2_end_mask_0 = const()[name = tensor("new_cache_13_internal_tensor_assign_2_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_13_internal_tensor_assign_2_squeeze_mask_0 = const()[name = tensor("new_cache_13_internal_tensor_assign_2_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_61 = const()[name = tensor("shape_61"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_13 = const()[name = tensor("reduce_prod_13"), val = tensor(1048576)]; + tensor range_1d_13_start_0 = const()[name = tensor("range_1d_13_start_0"), val = tensor(0)]; + tensor range_1d_13_step_0 = const()[name = tensor("range_1d_13_step_0"), val = tensor(1)]; + tensor range_1d_13 = range_1d(end = reduce_prod_13, start = range_1d_13_start_0, step = range_1d_13_step_0)[name = tensor("range_1d_13")]; + tensor reshape_65 = reshape(shape = shape_61, x = range_1d_13)[name = tensor("reshape_65")]; + tensor slice_by_index_13 = slice_by_index(begin = concat_46, begin_mask = new_cache_13_internal_tensor_assign_2_begin_mask_0, end = concat_47, end_mask = new_cache_13_internal_tensor_assign_2_end_mask_0, squeeze_mask = new_cache_13_internal_tensor_assign_2_squeeze_mask_0, stride = new_cache_13_internal_tensor_assign_2_stride_0, x = reshape_65)[name = tensor("slice_by_index_13")]; + tensor reshape_66_shape_0 = const()[name = tensor("reshape_66_shape_0"), val = tensor([-1])]; + tensor reshape_66 = reshape(shape = reshape_66_shape_0, x = slice_by_index_13)[name = tensor("reshape_66")]; + tensor reshape_67_shape_0 = const()[name = tensor("reshape_67_shape_0"), val = tensor([-1])]; + tensor reshape_67_cast_fp16 = reshape(shape = reshape_67_shape_0, x = var_2761_cast_fp16)[name = tensor("reshape_67_cast_fp16")]; + tensor reshape_68_shape_0 = const()[name = tensor("reshape_68_shape_0"), val = tensor([-1])]; + tensor reshape_68_cast_fp16 = reshape(shape = reshape_68_shape_0, x = reshape_64_cast_fp16)[name = tensor("reshape_68_cast_fp16")]; + tensor scatter_13_mode_0 = const()[name = tensor("scatter_13_mode_0"), val = tensor("update")]; + tensor scatter_13_axis_0 = const()[name = tensor("scatter_13_axis_0"), val = tensor(0)]; + tensor scatter_13_validate_indices_0 = const()[name = tensor("scatter_13_validate_indices_0"), val = tensor(false)]; + tensor scatter_13_cast_fp16 = scatter(axis = scatter_13_axis_0, data = reshape_68_cast_fp16, indices = reshape_66, mode = scatter_13_mode_0, updates = reshape_67_cast_fp16, validate_indices = scatter_13_validate_indices_0)[name = tensor("scatter_13_cast_fp16")]; + tensor reshape_69_cast_fp16 = reshape(shape = shape_61, x = scatter_13_cast_fp16)[name = tensor("reshape_69_cast_fp16")]; + tensor reshape_69_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("reshape_69_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor keys_37_begin_0 = const()[name = tensor("keys_37_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor keys_37_end_0 = const()[name = tensor("keys_37_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor keys_37_end_mask_0 = const()[name = tensor("keys_37_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor keys_37_squeeze_mask_0 = const()[name = tensor("keys_37_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor keys_37_cast_fp16 = slice_by_index(begin = keys_37_begin_0, end = keys_37_end_0, end_mask = keys_37_end_mask_0, squeeze_mask = keys_37_squeeze_mask_0, x = reshape_69_cast_fp16)[name = tensor("keys_37_cast_fp16")]; + tensor values_37_begin_0 = const()[name = tensor("values_37_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor values_37_end_0 = const()[name = tensor("values_37_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor values_37_end_mask_0 = const()[name = tensor("values_37_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor values_37_squeeze_mask_0 = const()[name = tensor("values_37_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor values_37_cast_fp16 = slice_by_index(begin = values_37_begin_0, end = values_37_end_0, end_mask = values_37_end_mask_0, squeeze_mask = values_37_squeeze_mask_0, x = reshape_69_cast_fp16)[name = tensor("values_37_cast_fp16")]; + tensor var_2773_cast_fp16 = not_equal(x = keys_37_cast_fp16, y = keys_37_cast_fp16)[name = tensor("op_2773_cast_fp16")]; + tensor keys_39_cast_fp16 = select(a = var_463_to_fp16, b = keys_37_cast_fp16, cond = var_2773_cast_fp16)[name = tensor("keys_39_cast_fp16")]; + tensor var_2781_cast_fp16 = not_equal(x = values_37_cast_fp16, y = values_37_cast_fp16)[name = tensor("op_2781_cast_fp16")]; + tensor values_39_cast_fp16 = select(a = var_463_to_fp16, b = values_37_cast_fp16, cond = var_2781_cast_fp16)[name = tensor("values_39_cast_fp16")]; + tensor var_2805 = const()[name = tensor("op_2805"), val = tensor([0, 2, 1, 3])]; + tensor var_2821 = const()[name = tensor("op_2821"), val = tensor([1, 1, 1])]; + tensor var_2822_cast_fp16 = reshape(shape = var_2821, x = position6_to_fp16)[name = tensor("op_2822_cast_fp16")]; + tensor var_2842_to_fp16 = const()[name = tensor("op_2842_to_fp16"), val = tensor(0x1p+0)]; + tensor valid_len_13_cast_fp16 = add(x = var_2822_cast_fp16, y = var_2842_to_fp16)[name = tensor("valid_len_13_cast_fp16")]; + tensor valid_mask_13_cast_fp16 = less(x = k_positions_1_promoted_to_fp16, y = valid_len_13_cast_fp16)[name = tensor("valid_mask_13_cast_fp16")]; + tensor causal_mask_13_cast_fp16 = less_equal(x = k_positions_1_promoted_to_fp16, y = var_2822_cast_fp16)[name = tensor("causal_mask_13_cast_fp16")]; + tensor attn_mask_25 = logical_and(x = valid_mask_13_cast_fp16, y = causal_mask_13_cast_fp16)[name = tensor("attn_mask_25")]; + tensor attn_mask_27_axes_0 = const()[name = tensor("attn_mask_27_axes_0"), val = tensor([1])]; + tensor attn_mask_27 = expand_dims(axes = attn_mask_27_axes_0, x = attn_mask_25)[name = tensor("attn_mask_27")]; + tensor var_2860_transpose_x_0 = const()[name = tensor("op_2860_transpose_x_0"), val = tensor(false)]; + tensor var_2860_transpose_y_0 = const()[name = tensor("op_2860_transpose_y_0"), val = tensor(false)]; + tensor transpose_84_perm_0 = const()[name = tensor("transpose_84_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_85_perm_0 = const()[name = tensor("transpose_85_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_85 = transpose(perm = transpose_85_perm_0, x = keys_39_cast_fp16)[name = tensor("transpose_189")]; + tensor transpose_84 = transpose(perm = transpose_84_perm_0, x = q_39_cast_fp16)[name = tensor("transpose_190")]; + tensor var_2860_cast_fp16 = matmul(transpose_x = var_2860_transpose_x_0, transpose_y = var_2860_transpose_y_0, x = transpose_84, y = transpose_85)[name = tensor("op_2860_cast_fp16")]; + tensor var_2854_to_fp16 = const()[name = tensor("op_2854_to_fp16"), val = tensor([0x1p-3])]; + tensor attn_weights_37_cast_fp16 = mul(x = var_2860_cast_fp16, y = var_2854_to_fp16)[name = tensor("attn_weights_37_cast_fp16")]; + tensor var_2862 = logical_not(x = attn_mask_27)[name = tensor("op_2862")]; + tensor var_2863_to_fp16 = const()[name = tensor("op_2863_to_fp16"), val = tensor(-inf)]; + tensor attn_weights_39_cast_fp16 = select(a = var_2863_to_fp16, b = attn_weights_37_cast_fp16, cond = var_2862)[name = tensor("attn_weights_39_cast_fp16")]; + tensor var_2865 = const()[name = tensor("op_2865"), val = tensor(-1)]; + tensor attn_weights_41_cast_fp16 = softmax(axis = var_2865, x = attn_weights_39_cast_fp16)[name = tensor("attn_weights_41_cast_fp16")]; + tensor attn_output_13_transpose_x_0 = const()[name = tensor("attn_output_13_transpose_x_0"), val = tensor(false)]; + tensor attn_output_13_transpose_y_0 = const()[name = tensor("attn_output_13_transpose_y_0"), val = tensor(false)]; + tensor values_41_cast_fp16 = transpose(perm = var_2805, x = values_39_cast_fp16)[name = tensor("transpose_191")]; + tensor attn_output_13_cast_fp16 = matmul(transpose_x = attn_output_13_transpose_x_0, transpose_y = attn_output_13_transpose_y_0, x = attn_weights_41_cast_fp16, y = values_41_cast_fp16)[name = tensor("attn_output_13_cast_fp16")]; + tensor var_2873 = const()[name = tensor("op_2873"), val = tensor([0, 2, 1, 3])]; + tensor var_2878 = const()[name = tensor("op_2878"), val = tensor([1, 1, 1024])]; + tensor var_2874_cast_fp16 = transpose(perm = var_2873, x = attn_output_13_cast_fp16)[name = tensor("transpose_188")]; + tensor input_79_cast_fp16 = reshape(shape = var_2878, x = var_2874_cast_fp16)[name = tensor("input_79_cast_fp16")]; + tensor attn6_out_proj_weight_to_fp16 = const()[name = tensor("attn6_out_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(158475776)))]; + tensor linear_26_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = attn6_out_proj_weight_to_fp16, x = input_79_cast_fp16)[name = tensor("linear_26_cast_fp16")]; + tensor var_2884_to_fp16 = const()[name = tensor("op_2884_to_fp16"), val = tensor(0x1p+0)]; + tensor var_2885_cast_fp16 = add(x = position6_to_fp16, y = var_2884_to_fp16)[name = tensor("op_2885_cast_fp16")]; + tensor var_2885_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_2885_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor input_81_cast_fp16 = add(x = input_75_cast_fp16, y = linear_26_cast_fp16)[name = tensor("input_81_cast_fp16")]; + tensor input_83_axes_0 = const()[name = tensor("input_83_axes_0"), val = tensor([-1])]; + tensor norm6_2_weight_to_fp16 = const()[name = tensor("norm6_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(160572992)))]; + tensor norm6_2_bias_to_fp16 = const()[name = tensor("norm6_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(160575104)))]; + tensor var_2889_to_fp16 = const()[name = tensor("op_2889_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_83_cast_fp16 = layer_norm(axes = input_83_axes_0, beta = norm6_2_bias_to_fp16, epsilon = var_2889_to_fp16, gamma = norm6_2_weight_to_fp16, x = input_81_cast_fp16)[name = tensor("input_83_cast_fp16")]; + tensor linear6_1_weight_to_fp16 = const()[name = tensor("linear6_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(160577216)))]; + tensor linear_27_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = linear6_1_weight_to_fp16, x = input_83_cast_fp16)[name = tensor("linear_27_cast_fp16")]; + tensor input_85_mode_0 = const()[name = tensor("input_85_mode_0"), val = tensor("EXACT")]; + tensor input_85_cast_fp16 = gelu(mode = input_85_mode_0, x = linear_27_cast_fp16)[name = tensor("input_85_cast_fp16")]; + tensor linear6_2_weight_to_fp16 = const()[name = tensor("linear6_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(168965888)))]; + tensor linear_28_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = linear6_2_weight_to_fp16, x = input_85_cast_fp16)[name = tensor("linear_28_cast_fp16")]; + tensor input_87_cast_fp16 = add(x = input_81_cast_fp16, y = linear_28_cast_fp16)[name = tensor("input_87_cast_fp16")]; + tensor input_89_axes_0 = const()[name = tensor("input_89_axes_0"), val = tensor([-1])]; + tensor norm7_1_weight_to_fp16 = const()[name = tensor("norm7_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(177354560)))]; + tensor norm7_1_bias_to_fp16 = const()[name = tensor("norm7_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(177356672)))]; + tensor var_2906_to_fp16 = const()[name = tensor("op_2906_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_89_cast_fp16 = layer_norm(axes = input_89_axes_0, beta = norm7_1_bias_to_fp16, epsilon = var_2906_to_fp16, gamma = norm7_1_weight_to_fp16, x = input_87_cast_fp16)[name = tensor("input_89_cast_fp16")]; + tensor attn7_in_proj_weight_to_fp16 = const()[name = tensor("attn7_in_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(177358784)))]; + tensor linear_29_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = attn7_in_proj_weight_to_fp16, x = input_89_cast_fp16)[name = tensor("linear_29_cast_fp16")]; + tensor var_2920 = const()[name = tensor("op_2920"), val = tensor([1, 1, 3, 16, 64])]; + tensor qkv_15_cast_fp16 = reshape(shape = var_2920, x = linear_29_cast_fp16)[name = tensor("qkv_15_cast_fp16")]; + tensor q_43_begin_0 = const()[name = tensor("q_43_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor q_43_end_0 = const()[name = tensor("q_43_end_0"), val = tensor([1, 1, 1, 16, 64])]; + tensor q_43_end_mask_0 = const()[name = tensor("q_43_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor q_43_squeeze_mask_0 = const()[name = tensor("q_43_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor q_43_cast_fp16 = slice_by_index(begin = q_43_begin_0, end = q_43_end_0, end_mask = q_43_end_mask_0, squeeze_mask = q_43_squeeze_mask_0, x = qkv_15_cast_fp16)[name = tensor("q_43_cast_fp16")]; + tensor k_29_begin_0 = const()[name = tensor("k_29_begin_0"), val = tensor([0, 0, 1, 0, 0])]; + tensor k_29_end_0 = const()[name = tensor("k_29_end_0"), val = tensor([1, 1, 2, 16, 64])]; + tensor k_29_end_mask_0 = const()[name = tensor("k_29_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor k_29_squeeze_mask_0 = const()[name = tensor("k_29_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor k_29_cast_fp16 = slice_by_index(begin = k_29_begin_0, end = k_29_end_0, end_mask = k_29_end_mask_0, squeeze_mask = k_29_squeeze_mask_0, x = qkv_15_cast_fp16)[name = tensor("k_29_cast_fp16")]; + tensor v_15_begin_0 = const()[name = tensor("v_15_begin_0"), val = tensor([0, 0, 2, 0, 0])]; + tensor v_15_end_0 = const()[name = tensor("v_15_end_0"), val = tensor([1, 1, 3, 16, 64])]; + tensor v_15_end_mask_0 = const()[name = tensor("v_15_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor v_15_squeeze_mask_0 = const()[name = tensor("v_15_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor v_15_cast_fp16 = slice_by_index(begin = v_15_begin_0, end = v_15_end_0, end_mask = v_15_end_mask_0, squeeze_mask = v_15_squeeze_mask_0, x = qkv_15_cast_fp16)[name = tensor("v_15_cast_fp16")]; + tensor var_2986 = const()[name = tensor("op_2986"), val = tensor([1, 1, 1, 1])]; + tensor position7_to_fp16_dtype_0 = const()[name = tensor("position7_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor position7_to_fp16 = cast(dtype = position7_to_fp16_dtype_0, x = position7)[name = tensor("cast_518")]; + tensor ts_47_cast_fp16 = reshape(shape = var_2986, x = position7_to_fp16)[name = tensor("ts_47_cast_fp16")]; + tensor var_2993 = const()[name = tensor("op_2993"), val = tensor([1, 1, 16, 32, 2])]; + tensor q_complex_15_cast_fp16 = reshape(shape = var_2993, x = q_43_cast_fp16)[name = tensor("q_complex_15_cast_fp16")]; + tensor var_3000 = const()[name = tensor("op_3000"), val = tensor([1, 1, 16, 32, 2])]; + tensor k_complex_15_cast_fp16 = reshape(shape = var_3000, x = k_29_cast_fp16)[name = tensor("k_complex_15_cast_fp16")]; + tensor var_3004_begin_0 = const()[name = tensor("op_3004_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_3004_end_0 = const()[name = tensor("op_3004_end_0"), val = tensor([1, 1, 16, 32, 1])]; + tensor var_3004_end_mask_0 = const()[name = tensor("op_3004_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_3004_squeeze_mask_0 = const()[name = tensor("op_3004_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_3004_cast_fp16 = slice_by_index(begin = var_3004_begin_0, end = var_3004_end_0, end_mask = var_3004_end_mask_0, squeeze_mask = var_3004_squeeze_mask_0, x = q_complex_15_cast_fp16)[name = tensor("op_3004_cast_fp16")]; + tensor var_3012_begin_0 = const()[name = tensor("op_3012_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_3012_end_0 = const()[name = tensor("op_3012_end_0"), val = tensor([1, 1, 16, 32, 2])]; + tensor var_3012_end_mask_0 = const()[name = tensor("op_3012_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_3012_squeeze_mask_0 = const()[name = tensor("op_3012_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_3012_cast_fp16 = slice_by_index(begin = var_3012_begin_0, end = var_3012_end_0, end_mask = var_3012_end_mask_0, squeeze_mask = var_3012_squeeze_mask_0, x = q_complex_15_cast_fp16)[name = tensor("op_3012_cast_fp16")]; + tensor var_3020_begin_0 = const()[name = tensor("op_3020_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_3020_end_0 = const()[name = tensor("op_3020_end_0"), val = tensor([1, 1, 16, 32, 1])]; + tensor var_3020_end_mask_0 = const()[name = tensor("op_3020_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_3020_squeeze_mask_0 = const()[name = tensor("op_3020_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_3020_cast_fp16 = slice_by_index(begin = var_3020_begin_0, end = var_3020_end_0, end_mask = var_3020_end_mask_0, squeeze_mask = var_3020_squeeze_mask_0, x = k_complex_15_cast_fp16)[name = tensor("op_3020_cast_fp16")]; + tensor var_3028_begin_0 = const()[name = tensor("op_3028_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_3028_end_0 = const()[name = tensor("op_3028_end_0"), val = tensor([1, 1, 16, 32, 2])]; + tensor var_3028_end_mask_0 = const()[name = tensor("op_3028_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_3028_squeeze_mask_0 = const()[name = tensor("op_3028_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_3028_cast_fp16 = slice_by_index(begin = var_3028_begin_0, end = var_3028_end_0, end_mask = var_3028_end_mask_0, squeeze_mask = var_3028_squeeze_mask_0, x = k_complex_15_cast_fp16)[name = tensor("op_3028_cast_fp16")]; + tensor freqs_15_to_fp16 = const()[name = tensor("freqs_15_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(183650304)))]; + tensor var_3034_cast_fp16 = mul(x = freqs_15_to_fp16, y = ts_47_cast_fp16)[name = tensor("op_3034_cast_fp16")]; + tensor rotr_15_cast_fp16 = cos(x = var_3034_cast_fp16)[name = tensor("rotr_15_cast_fp16")]; + tensor roti_15_cast_fp16 = sin(x = var_3034_cast_fp16)[name = tensor("roti_15_cast_fp16")]; + tensor var_3038_cast_fp16 = mul(x = var_3004_cast_fp16, y = rotr_15_cast_fp16)[name = tensor("op_3038_cast_fp16")]; + tensor var_3039_cast_fp16 = mul(x = var_3012_cast_fp16, y = roti_15_cast_fp16)[name = tensor("op_3039_cast_fp16")]; + tensor qor_29_cast_fp16 = sub(x = var_3038_cast_fp16, y = var_3039_cast_fp16)[name = tensor("qor_29_cast_fp16")]; + tensor var_3042_cast_fp16 = mul(x = var_3004_cast_fp16, y = roti_15_cast_fp16)[name = tensor("op_3042_cast_fp16")]; + tensor var_3043_cast_fp16 = mul(x = var_3012_cast_fp16, y = rotr_15_cast_fp16)[name = tensor("op_3043_cast_fp16")]; + tensor qoi_29_cast_fp16 = add(x = var_3042_cast_fp16, y = var_3043_cast_fp16)[name = tensor("qoi_29_cast_fp16")]; + tensor var_3046_cast_fp16 = mul(x = var_3020_cast_fp16, y = rotr_15_cast_fp16)[name = tensor("op_3046_cast_fp16")]; + tensor var_3047_cast_fp16 = mul(x = var_3028_cast_fp16, y = roti_15_cast_fp16)[name = tensor("op_3047_cast_fp16")]; + tensor kor_29_cast_fp16 = sub(x = var_3046_cast_fp16, y = var_3047_cast_fp16)[name = tensor("kor_29_cast_fp16")]; + tensor var_3050_cast_fp16 = mul(x = var_3020_cast_fp16, y = roti_15_cast_fp16)[name = tensor("op_3050_cast_fp16")]; + tensor var_3051_cast_fp16 = mul(x = var_3028_cast_fp16, y = rotr_15_cast_fp16)[name = tensor("op_3051_cast_fp16")]; + tensor koi_29_cast_fp16 = add(x = var_3050_cast_fp16, y = var_3051_cast_fp16)[name = tensor("koi_29_cast_fp16")]; + tensor qo_15_axis_0 = const()[name = tensor("qo_15_axis_0"), val = tensor(-1)]; + tensor qo_15_cast_fp16 = stack(axis = qo_15_axis_0, values = (qor_29_cast_fp16, qoi_29_cast_fp16))[name = tensor("qo_15_cast_fp16")]; + tensor ko_15_axis_0 = const()[name = tensor("ko_15_axis_0"), val = tensor(-1)]; + tensor ko_15_cast_fp16 = stack(axis = ko_15_axis_0, values = (kor_29_cast_fp16, koi_29_cast_fp16))[name = tensor("ko_15_cast_fp16")]; + tensor var_3084 = const()[name = tensor("op_3084"), val = tensor([1, 1, 16, 64])]; + tensor q_45_cast_fp16 = reshape(shape = var_3084, x = qo_15_cast_fp16)[name = tensor("q_45_cast_fp16")]; + tensor var_3090 = const()[name = tensor("op_3090"), val = tensor([1, 1, 16, 64])]; + tensor k_31_cast_fp16 = reshape(shape = var_3090, x = ko_15_cast_fp16)[name = tensor("k_31_cast_fp16")]; + tensor _inversed_3115_y_0_to_fp16 = const()[name = tensor("_inversed_3115_y_0_to_fp16"), val = tensor(0x1p-9)]; + tensor _inversed_3115_cast_fp16 = mul(x = ts_47_cast_fp16, y = _inversed_3115_y_0_to_fp16)[name = tensor("_inversed_3115_cast_fp16")]; + tensor var_3116_cast_fp16 = floor(x = _inversed_3115_cast_fp16)[name = tensor("op_3116_cast_fp16")]; + tensor var_3117_to_fp16 = const()[name = tensor("op_3117_to_fp16"), val = tensor(0x1p+9)]; + tensor var_3118_cast_fp16 = mul(x = var_3116_cast_fp16, y = var_3117_to_fp16)[name = tensor("op_3118_cast_fp16")]; + tensor write_indices_float_31_cast_fp16 = sub(x = ts_47_cast_fp16, y = var_3118_cast_fp16)[name = tensor("write_indices_float_31_cast_fp16")]; + tensor var_3125_dtype_0 = const()[name = tensor("op_3125_dtype_0"), val = tensor("int32")]; + tensor write_indices_15_reps_0 = const()[name = tensor("write_indices_15_reps_0"), val = tensor([1, 1, 16, 64])]; + tensor write_indices_float_31_cast_fp16_to_int32 = cast(dtype = var_3125_dtype_0, x = write_indices_float_31_cast_fp16)[name = tensor("cast_517")]; + tensor write_indices_15 = tile(reps = write_indices_15_reps_0, x = write_indices_float_31_cast_fp16_to_int32)[name = tensor("write_indices_15")]; + tensor var_3135_begin_0 = const()[name = tensor("op_3135_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_3135_end_0 = const()[name = tensor("op_3135_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor var_3135_end_mask_0 = const()[name = tensor("op_3135_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_3135_squeeze_mask_0 = const()[name = tensor("op_3135_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor cache7_to_fp16_dtype_0 = const()[name = tensor("cache7_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor cache7_to_fp16 = cast(dtype = cache7_to_fp16_dtype_0, x = cache7)[name = tensor("cast_516")]; + tensor var_3135_cast_fp16 = slice_by_index(begin = var_3135_begin_0, end = var_3135_end_0, end_mask = var_3135_end_mask_0, squeeze_mask = var_3135_squeeze_mask_0, x = cache7_to_fp16)[name = tensor("op_3135_cast_fp16")]; + tensor var_3137_axis_0 = const()[name = tensor("op_3137_axis_0"), val = tensor(1)]; + tensor var_3137_mode_0 = const()[name = tensor("op_3137_mode_0"), val = tensor("update")]; + tensor var_3137_validate_indices_0 = const()[name = tensor("op_3137_validate_indices_0"), val = tensor(false)]; + tensor var_3137_cast_fp16 = scatter_along_axis(axis = var_3137_axis_0, data = var_3135_cast_fp16, indices = write_indices_15, mode = var_3137_mode_0, updates = k_31_cast_fp16, validate_indices = var_3137_validate_indices_0)[name = tensor("op_3137_cast_fp16")]; + tensor concat_51 = const()[name = tensor("concat_51"), val = tensor([0, 0, 0, 0, 0])]; + tensor concat_52 = const()[name = tensor("concat_52"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_15_internal_tensor_assign_1_stride_0 = const()[name = tensor("new_cache_15_internal_tensor_assign_1_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_15_internal_tensor_assign_1_begin_mask_0 = const()[name = tensor("new_cache_15_internal_tensor_assign_1_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_15_internal_tensor_assign_1_end_mask_0 = const()[name = tensor("new_cache_15_internal_tensor_assign_1_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_15_internal_tensor_assign_1_squeeze_mask_0 = const()[name = tensor("new_cache_15_internal_tensor_assign_1_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_62 = const()[name = tensor("shape_62"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_14 = const()[name = tensor("reduce_prod_14"), val = tensor(1048576)]; + tensor range_1d_14_start_0 = const()[name = tensor("range_1d_14_start_0"), val = tensor(0)]; + tensor range_1d_14_step_0 = const()[name = tensor("range_1d_14_step_0"), val = tensor(1)]; + tensor range_1d_14 = range_1d(end = reduce_prod_14, start = range_1d_14_start_0, step = range_1d_14_step_0)[name = tensor("range_1d_14")]; + tensor reshape_70 = reshape(shape = shape_62, x = range_1d_14)[name = tensor("reshape_70")]; + tensor slice_by_index_14 = slice_by_index(begin = concat_51, begin_mask = new_cache_15_internal_tensor_assign_1_begin_mask_0, end = concat_52, end_mask = new_cache_15_internal_tensor_assign_1_end_mask_0, squeeze_mask = new_cache_15_internal_tensor_assign_1_squeeze_mask_0, stride = new_cache_15_internal_tensor_assign_1_stride_0, x = reshape_70)[name = tensor("slice_by_index_14")]; + tensor reshape_71_shape_0 = const()[name = tensor("reshape_71_shape_0"), val = tensor([-1])]; + tensor reshape_71 = reshape(shape = reshape_71_shape_0, x = slice_by_index_14)[name = tensor("reshape_71")]; + tensor reshape_72_shape_0 = const()[name = tensor("reshape_72_shape_0"), val = tensor([-1])]; + tensor reshape_72_cast_fp16 = reshape(shape = reshape_72_shape_0, x = var_3137_cast_fp16)[name = tensor("reshape_72_cast_fp16")]; + tensor reshape_73_shape_0 = const()[name = tensor("reshape_73_shape_0"), val = tensor([-1])]; + tensor reshape_73_cast_fp16 = reshape(shape = reshape_73_shape_0, x = cache7_to_fp16)[name = tensor("reshape_73_cast_fp16")]; + tensor scatter_14_mode_0 = const()[name = tensor("scatter_14_mode_0"), val = tensor("update")]; + tensor scatter_14_axis_0 = const()[name = tensor("scatter_14_axis_0"), val = tensor(0)]; + tensor scatter_14_validate_indices_0 = const()[name = tensor("scatter_14_validate_indices_0"), val = tensor(false)]; + tensor scatter_14_cast_fp16 = scatter(axis = scatter_14_axis_0, data = reshape_73_cast_fp16, indices = reshape_71, mode = scatter_14_mode_0, updates = reshape_72_cast_fp16, validate_indices = scatter_14_validate_indices_0)[name = tensor("scatter_14_cast_fp16")]; + tensor reshape_74_cast_fp16 = reshape(shape = shape_62, x = scatter_14_cast_fp16)[name = tensor("reshape_74_cast_fp16")]; + tensor var_3145_begin_0 = const()[name = tensor("op_3145_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor var_3145_end_0 = const()[name = tensor("op_3145_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor var_3145_end_mask_0 = const()[name = tensor("op_3145_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_3145_squeeze_mask_0 = const()[name = tensor("op_3145_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor var_3145_cast_fp16 = slice_by_index(begin = var_3145_begin_0, end = var_3145_end_0, end_mask = var_3145_end_mask_0, squeeze_mask = var_3145_squeeze_mask_0, x = reshape_74_cast_fp16)[name = tensor("op_3145_cast_fp16")]; + tensor var_3147_axis_0 = const()[name = tensor("op_3147_axis_0"), val = tensor(1)]; + tensor var_3147_mode_0 = const()[name = tensor("op_3147_mode_0"), val = tensor("update")]; + tensor var_3147_validate_indices_0 = const()[name = tensor("op_3147_validate_indices_0"), val = tensor(false)]; + tensor var_3147_cast_fp16 = scatter_along_axis(axis = var_3147_axis_0, data = var_3145_cast_fp16, indices = write_indices_15, mode = var_3147_mode_0, updates = v_15_cast_fp16, validate_indices = var_3147_validate_indices_0)[name = tensor("op_3147_cast_fp16")]; + tensor concat_53 = const()[name = tensor("concat_53"), val = tensor([1, 0, 0, 0, 0])]; + tensor concat_54 = const()[name = tensor("concat_54"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_15_internal_tensor_assign_2_stride_0 = const()[name = tensor("new_cache_15_internal_tensor_assign_2_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_15_internal_tensor_assign_2_begin_mask_0 = const()[name = tensor("new_cache_15_internal_tensor_assign_2_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_15_internal_tensor_assign_2_end_mask_0 = const()[name = tensor("new_cache_15_internal_tensor_assign_2_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_15_internal_tensor_assign_2_squeeze_mask_0 = const()[name = tensor("new_cache_15_internal_tensor_assign_2_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_63 = const()[name = tensor("shape_63"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_15 = const()[name = tensor("reduce_prod_15"), val = tensor(1048576)]; + tensor range_1d_15_start_0 = const()[name = tensor("range_1d_15_start_0"), val = tensor(0)]; + tensor range_1d_15_step_0 = const()[name = tensor("range_1d_15_step_0"), val = tensor(1)]; + tensor range_1d_15 = range_1d(end = reduce_prod_15, start = range_1d_15_start_0, step = range_1d_15_step_0)[name = tensor("range_1d_15")]; + tensor reshape_75 = reshape(shape = shape_63, x = range_1d_15)[name = tensor("reshape_75")]; + tensor slice_by_index_15 = slice_by_index(begin = concat_53, begin_mask = new_cache_15_internal_tensor_assign_2_begin_mask_0, end = concat_54, end_mask = new_cache_15_internal_tensor_assign_2_end_mask_0, squeeze_mask = new_cache_15_internal_tensor_assign_2_squeeze_mask_0, stride = new_cache_15_internal_tensor_assign_2_stride_0, x = reshape_75)[name = tensor("slice_by_index_15")]; + tensor reshape_76_shape_0 = const()[name = tensor("reshape_76_shape_0"), val = tensor([-1])]; + tensor reshape_76 = reshape(shape = reshape_76_shape_0, x = slice_by_index_15)[name = tensor("reshape_76")]; + tensor reshape_77_shape_0 = const()[name = tensor("reshape_77_shape_0"), val = tensor([-1])]; + tensor reshape_77_cast_fp16 = reshape(shape = reshape_77_shape_0, x = var_3147_cast_fp16)[name = tensor("reshape_77_cast_fp16")]; + tensor reshape_78_shape_0 = const()[name = tensor("reshape_78_shape_0"), val = tensor([-1])]; + tensor reshape_78_cast_fp16 = reshape(shape = reshape_78_shape_0, x = reshape_74_cast_fp16)[name = tensor("reshape_78_cast_fp16")]; + tensor scatter_15_mode_0 = const()[name = tensor("scatter_15_mode_0"), val = tensor("update")]; + tensor scatter_15_axis_0 = const()[name = tensor("scatter_15_axis_0"), val = tensor(0)]; + tensor scatter_15_validate_indices_0 = const()[name = tensor("scatter_15_validate_indices_0"), val = tensor(false)]; + tensor scatter_15_cast_fp16 = scatter(axis = scatter_15_axis_0, data = reshape_78_cast_fp16, indices = reshape_76, mode = scatter_15_mode_0, updates = reshape_77_cast_fp16, validate_indices = scatter_15_validate_indices_0)[name = tensor("scatter_15_cast_fp16")]; + tensor reshape_79_cast_fp16 = reshape(shape = shape_63, x = scatter_15_cast_fp16)[name = tensor("reshape_79_cast_fp16")]; + tensor reshape_79_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("reshape_79_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor keys_43_begin_0 = const()[name = tensor("keys_43_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor keys_43_end_0 = const()[name = tensor("keys_43_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor keys_43_end_mask_0 = const()[name = tensor("keys_43_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor keys_43_squeeze_mask_0 = const()[name = tensor("keys_43_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor keys_43_cast_fp16 = slice_by_index(begin = keys_43_begin_0, end = keys_43_end_0, end_mask = keys_43_end_mask_0, squeeze_mask = keys_43_squeeze_mask_0, x = reshape_79_cast_fp16)[name = tensor("keys_43_cast_fp16")]; + tensor values_43_begin_0 = const()[name = tensor("values_43_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor values_43_end_0 = const()[name = tensor("values_43_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor values_43_end_mask_0 = const()[name = tensor("values_43_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor values_43_squeeze_mask_0 = const()[name = tensor("values_43_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor values_43_cast_fp16 = slice_by_index(begin = values_43_begin_0, end = values_43_end_0, end_mask = values_43_end_mask_0, squeeze_mask = values_43_squeeze_mask_0, x = reshape_79_cast_fp16)[name = tensor("values_43_cast_fp16")]; + tensor var_3159_cast_fp16 = not_equal(x = keys_43_cast_fp16, y = keys_43_cast_fp16)[name = tensor("op_3159_cast_fp16")]; + tensor keys_45_cast_fp16 = select(a = var_463_to_fp16, b = keys_43_cast_fp16, cond = var_3159_cast_fp16)[name = tensor("keys_45_cast_fp16")]; + tensor var_3167_cast_fp16 = not_equal(x = values_43_cast_fp16, y = values_43_cast_fp16)[name = tensor("op_3167_cast_fp16")]; + tensor values_45_cast_fp16 = select(a = var_463_to_fp16, b = values_43_cast_fp16, cond = var_3167_cast_fp16)[name = tensor("values_45_cast_fp16")]; + tensor var_3191 = const()[name = tensor("op_3191"), val = tensor([0, 2, 1, 3])]; + tensor var_3207 = const()[name = tensor("op_3207"), val = tensor([1, 1, 1])]; + tensor var_3208_cast_fp16 = reshape(shape = var_3207, x = position7_to_fp16)[name = tensor("op_3208_cast_fp16")]; + tensor var_3228_to_fp16 = const()[name = tensor("op_3228_to_fp16"), val = tensor(0x1p+0)]; + tensor valid_len_15_cast_fp16 = add(x = var_3208_cast_fp16, y = var_3228_to_fp16)[name = tensor("valid_len_15_cast_fp16")]; + tensor valid_mask_15_cast_fp16 = less(x = k_positions_1_promoted_to_fp16, y = valid_len_15_cast_fp16)[name = tensor("valid_mask_15_cast_fp16")]; + tensor causal_mask_15_cast_fp16 = less_equal(x = k_positions_1_promoted_to_fp16, y = var_3208_cast_fp16)[name = tensor("causal_mask_15_cast_fp16")]; + tensor attn_mask_29 = logical_and(x = valid_mask_15_cast_fp16, y = causal_mask_15_cast_fp16)[name = tensor("attn_mask_29")]; + tensor attn_mask_31_axes_0 = const()[name = tensor("attn_mask_31_axes_0"), val = tensor([1])]; + tensor attn_mask_31 = expand_dims(axes = attn_mask_31_axes_0, x = attn_mask_29)[name = tensor("attn_mask_31")]; + tensor var_3246_transpose_x_0 = const()[name = tensor("op_3246_transpose_x_0"), val = tensor(false)]; + tensor var_3246_transpose_y_0 = const()[name = tensor("op_3246_transpose_y_0"), val = tensor(false)]; + tensor transpose_86_perm_0 = const()[name = tensor("transpose_86_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_87_perm_0 = const()[name = tensor("transpose_87_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_87 = transpose(perm = transpose_87_perm_0, x = keys_45_cast_fp16)[name = tensor("transpose_185")]; + tensor transpose_86 = transpose(perm = transpose_86_perm_0, x = q_45_cast_fp16)[name = tensor("transpose_186")]; + tensor var_3246_cast_fp16 = matmul(transpose_x = var_3246_transpose_x_0, transpose_y = var_3246_transpose_y_0, x = transpose_86, y = transpose_87)[name = tensor("op_3246_cast_fp16")]; + tensor var_3240_to_fp16 = const()[name = tensor("op_3240_to_fp16"), val = tensor([0x1p-3])]; + tensor attn_weights_43_cast_fp16 = mul(x = var_3246_cast_fp16, y = var_3240_to_fp16)[name = tensor("attn_weights_43_cast_fp16")]; + tensor var_3248 = logical_not(x = attn_mask_31)[name = tensor("op_3248")]; + tensor var_3249_to_fp16 = const()[name = tensor("op_3249_to_fp16"), val = tensor(-inf)]; + tensor attn_weights_45_cast_fp16 = select(a = var_3249_to_fp16, b = attn_weights_43_cast_fp16, cond = var_3248)[name = tensor("attn_weights_45_cast_fp16")]; + tensor var_3251 = const()[name = tensor("op_3251"), val = tensor(-1)]; + tensor attn_weights_47_cast_fp16 = softmax(axis = var_3251, x = attn_weights_45_cast_fp16)[name = tensor("attn_weights_47_cast_fp16")]; + tensor attn_output_15_transpose_x_0 = const()[name = tensor("attn_output_15_transpose_x_0"), val = tensor(false)]; + tensor attn_output_15_transpose_y_0 = const()[name = tensor("attn_output_15_transpose_y_0"), val = tensor(false)]; + tensor values_47_cast_fp16 = transpose(perm = var_3191, x = values_45_cast_fp16)[name = tensor("transpose_187")]; + tensor attn_output_15_cast_fp16 = matmul(transpose_x = attn_output_15_transpose_x_0, transpose_y = attn_output_15_transpose_y_0, x = attn_weights_47_cast_fp16, y = values_47_cast_fp16)[name = tensor("attn_output_15_cast_fp16")]; + tensor var_3259 = const()[name = tensor("op_3259"), val = tensor([0, 2, 1, 3])]; + tensor var_3264 = const()[name = tensor("op_3264"), val = tensor([1, 1, 1024])]; + tensor var_3260_cast_fp16 = transpose(perm = var_3259, x = attn_output_15_cast_fp16)[name = tensor("transpose_184")]; + tensor input_91_cast_fp16 = reshape(shape = var_3264, x = var_3260_cast_fp16)[name = tensor("input_91_cast_fp16")]; + tensor attn7_out_proj_weight_to_fp16 = const()[name = tensor("attn7_out_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(183650432)))]; + tensor linear_30_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = attn7_out_proj_weight_to_fp16, x = input_91_cast_fp16)[name = tensor("linear_30_cast_fp16")]; + tensor var_3270_to_fp16 = const()[name = tensor("op_3270_to_fp16"), val = tensor(0x1p+0)]; + tensor var_3271_cast_fp16 = add(x = position7_to_fp16, y = var_3270_to_fp16)[name = tensor("op_3271_cast_fp16")]; + tensor var_3271_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_3271_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor input_93_cast_fp16 = add(x = input_87_cast_fp16, y = linear_30_cast_fp16)[name = tensor("input_93_cast_fp16")]; + tensor input_95_axes_0 = const()[name = tensor("input_95_axes_0"), val = tensor([-1])]; + tensor norm7_2_weight_to_fp16 = const()[name = tensor("norm7_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(185747648)))]; + tensor norm7_2_bias_to_fp16 = const()[name = tensor("norm7_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(185749760)))]; + tensor var_3275_to_fp16 = const()[name = tensor("op_3275_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_95_cast_fp16 = layer_norm(axes = input_95_axes_0, beta = norm7_2_bias_to_fp16, epsilon = var_3275_to_fp16, gamma = norm7_2_weight_to_fp16, x = input_93_cast_fp16)[name = tensor("input_95_cast_fp16")]; + tensor linear7_1_weight_to_fp16 = const()[name = tensor("linear7_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(185751872)))]; + tensor linear_31_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = linear7_1_weight_to_fp16, x = input_95_cast_fp16)[name = tensor("linear_31_cast_fp16")]; + tensor input_97_mode_0 = const()[name = tensor("input_97_mode_0"), val = tensor("EXACT")]; + tensor input_97_cast_fp16 = gelu(mode = input_97_mode_0, x = linear_31_cast_fp16)[name = tensor("input_97_cast_fp16")]; + tensor linear7_2_weight_to_fp16 = const()[name = tensor("linear7_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(194140544)))]; + tensor linear_32_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = linear7_2_weight_to_fp16, x = input_97_cast_fp16)[name = tensor("linear_32_cast_fp16")]; + tensor input_99_cast_fp16 = add(x = input_93_cast_fp16, y = linear_32_cast_fp16)[name = tensor("input_99_cast_fp16")]; + tensor input_101_axes_0 = const()[name = tensor("input_101_axes_0"), val = tensor([-1])]; + tensor norm8_1_weight_to_fp16 = const()[name = tensor("norm8_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(202529216)))]; + tensor norm8_1_bias_to_fp16 = const()[name = tensor("norm8_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(202531328)))]; + tensor var_3292_to_fp16 = const()[name = tensor("op_3292_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_101_cast_fp16 = layer_norm(axes = input_101_axes_0, beta = norm8_1_bias_to_fp16, epsilon = var_3292_to_fp16, gamma = norm8_1_weight_to_fp16, x = input_99_cast_fp16)[name = tensor("input_101_cast_fp16")]; + tensor attn8_in_proj_weight_to_fp16 = const()[name = tensor("attn8_in_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(202533440)))]; + tensor linear_33_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = attn8_in_proj_weight_to_fp16, x = input_101_cast_fp16)[name = tensor("linear_33_cast_fp16")]; + tensor var_3306 = const()[name = tensor("op_3306"), val = tensor([1, 1, 3, 16, 64])]; + tensor qkv_17_cast_fp16 = reshape(shape = var_3306, x = linear_33_cast_fp16)[name = tensor("qkv_17_cast_fp16")]; + tensor q_49_begin_0 = const()[name = tensor("q_49_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor q_49_end_0 = const()[name = tensor("q_49_end_0"), val = tensor([1, 1, 1, 16, 64])]; + tensor q_49_end_mask_0 = const()[name = tensor("q_49_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor q_49_squeeze_mask_0 = const()[name = tensor("q_49_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor q_49_cast_fp16 = slice_by_index(begin = q_49_begin_0, end = q_49_end_0, end_mask = q_49_end_mask_0, squeeze_mask = q_49_squeeze_mask_0, x = qkv_17_cast_fp16)[name = tensor("q_49_cast_fp16")]; + tensor k_33_begin_0 = const()[name = tensor("k_33_begin_0"), val = tensor([0, 0, 1, 0, 0])]; + tensor k_33_end_0 = const()[name = tensor("k_33_end_0"), val = tensor([1, 1, 2, 16, 64])]; + tensor k_33_end_mask_0 = const()[name = tensor("k_33_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor k_33_squeeze_mask_0 = const()[name = tensor("k_33_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor k_33_cast_fp16 = slice_by_index(begin = k_33_begin_0, end = k_33_end_0, end_mask = k_33_end_mask_0, squeeze_mask = k_33_squeeze_mask_0, x = qkv_17_cast_fp16)[name = tensor("k_33_cast_fp16")]; + tensor v_17_begin_0 = const()[name = tensor("v_17_begin_0"), val = tensor([0, 0, 2, 0, 0])]; + tensor v_17_end_0 = const()[name = tensor("v_17_end_0"), val = tensor([1, 1, 3, 16, 64])]; + tensor v_17_end_mask_0 = const()[name = tensor("v_17_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor v_17_squeeze_mask_0 = const()[name = tensor("v_17_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor v_17_cast_fp16 = slice_by_index(begin = v_17_begin_0, end = v_17_end_0, end_mask = v_17_end_mask_0, squeeze_mask = v_17_squeeze_mask_0, x = qkv_17_cast_fp16)[name = tensor("v_17_cast_fp16")]; + tensor var_3372 = const()[name = tensor("op_3372"), val = tensor([1, 1, 1, 1])]; + tensor position8_to_fp16_dtype_0 = const()[name = tensor("position8_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor position8_to_fp16 = cast(dtype = position8_to_fp16_dtype_0, x = position8)[name = tensor("cast_513")]; + tensor ts_53_cast_fp16 = reshape(shape = var_3372, x = position8_to_fp16)[name = tensor("ts_53_cast_fp16")]; + tensor var_3379 = const()[name = tensor("op_3379"), val = tensor([1, 1, 16, 32, 2])]; + tensor q_complex_17_cast_fp16 = reshape(shape = var_3379, x = q_49_cast_fp16)[name = tensor("q_complex_17_cast_fp16")]; + tensor var_3386 = const()[name = tensor("op_3386"), val = tensor([1, 1, 16, 32, 2])]; + tensor k_complex_17_cast_fp16 = reshape(shape = var_3386, x = k_33_cast_fp16)[name = tensor("k_complex_17_cast_fp16")]; + tensor var_3390_begin_0 = const()[name = tensor("op_3390_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_3390_end_0 = const()[name = tensor("op_3390_end_0"), val = tensor([1, 1, 16, 32, 1])]; + tensor var_3390_end_mask_0 = const()[name = tensor("op_3390_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_3390_squeeze_mask_0 = const()[name = tensor("op_3390_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_3390_cast_fp16 = slice_by_index(begin = var_3390_begin_0, end = var_3390_end_0, end_mask = var_3390_end_mask_0, squeeze_mask = var_3390_squeeze_mask_0, x = q_complex_17_cast_fp16)[name = tensor("op_3390_cast_fp16")]; + tensor var_3398_begin_0 = const()[name = tensor("op_3398_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_3398_end_0 = const()[name = tensor("op_3398_end_0"), val = tensor([1, 1, 16, 32, 2])]; + tensor var_3398_end_mask_0 = const()[name = tensor("op_3398_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_3398_squeeze_mask_0 = const()[name = tensor("op_3398_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_3398_cast_fp16 = slice_by_index(begin = var_3398_begin_0, end = var_3398_end_0, end_mask = var_3398_end_mask_0, squeeze_mask = var_3398_squeeze_mask_0, x = q_complex_17_cast_fp16)[name = tensor("op_3398_cast_fp16")]; + tensor var_3406_begin_0 = const()[name = tensor("op_3406_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_3406_end_0 = const()[name = tensor("op_3406_end_0"), val = tensor([1, 1, 16, 32, 1])]; + tensor var_3406_end_mask_0 = const()[name = tensor("op_3406_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_3406_squeeze_mask_0 = const()[name = tensor("op_3406_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_3406_cast_fp16 = slice_by_index(begin = var_3406_begin_0, end = var_3406_end_0, end_mask = var_3406_end_mask_0, squeeze_mask = var_3406_squeeze_mask_0, x = k_complex_17_cast_fp16)[name = tensor("op_3406_cast_fp16")]; + tensor var_3414_begin_0 = const()[name = tensor("op_3414_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_3414_end_0 = const()[name = tensor("op_3414_end_0"), val = tensor([1, 1, 16, 32, 2])]; + tensor var_3414_end_mask_0 = const()[name = tensor("op_3414_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_3414_squeeze_mask_0 = const()[name = tensor("op_3414_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_3414_cast_fp16 = slice_by_index(begin = var_3414_begin_0, end = var_3414_end_0, end_mask = var_3414_end_mask_0, squeeze_mask = var_3414_squeeze_mask_0, x = k_complex_17_cast_fp16)[name = tensor("op_3414_cast_fp16")]; + tensor freqs_17_to_fp16 = const()[name = tensor("freqs_17_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(208824960)))]; + tensor var_3420_cast_fp16 = mul(x = freqs_17_to_fp16, y = ts_53_cast_fp16)[name = tensor("op_3420_cast_fp16")]; + tensor rotr_17_cast_fp16 = cos(x = var_3420_cast_fp16)[name = tensor("rotr_17_cast_fp16")]; + tensor roti_17_cast_fp16 = sin(x = var_3420_cast_fp16)[name = tensor("roti_17_cast_fp16")]; + tensor var_3424_cast_fp16 = mul(x = var_3390_cast_fp16, y = rotr_17_cast_fp16)[name = tensor("op_3424_cast_fp16")]; + tensor var_3425_cast_fp16 = mul(x = var_3398_cast_fp16, y = roti_17_cast_fp16)[name = tensor("op_3425_cast_fp16")]; + tensor qor_33_cast_fp16 = sub(x = var_3424_cast_fp16, y = var_3425_cast_fp16)[name = tensor("qor_33_cast_fp16")]; + tensor var_3428_cast_fp16 = mul(x = var_3390_cast_fp16, y = roti_17_cast_fp16)[name = tensor("op_3428_cast_fp16")]; + tensor var_3429_cast_fp16 = mul(x = var_3398_cast_fp16, y = rotr_17_cast_fp16)[name = tensor("op_3429_cast_fp16")]; + tensor qoi_33_cast_fp16 = add(x = var_3428_cast_fp16, y = var_3429_cast_fp16)[name = tensor("qoi_33_cast_fp16")]; + tensor var_3432_cast_fp16 = mul(x = var_3406_cast_fp16, y = rotr_17_cast_fp16)[name = tensor("op_3432_cast_fp16")]; + tensor var_3433_cast_fp16 = mul(x = var_3414_cast_fp16, y = roti_17_cast_fp16)[name = tensor("op_3433_cast_fp16")]; + tensor kor_33_cast_fp16 = sub(x = var_3432_cast_fp16, y = var_3433_cast_fp16)[name = tensor("kor_33_cast_fp16")]; + tensor var_3436_cast_fp16 = mul(x = var_3406_cast_fp16, y = roti_17_cast_fp16)[name = tensor("op_3436_cast_fp16")]; + tensor var_3437_cast_fp16 = mul(x = var_3414_cast_fp16, y = rotr_17_cast_fp16)[name = tensor("op_3437_cast_fp16")]; + tensor koi_33_cast_fp16 = add(x = var_3436_cast_fp16, y = var_3437_cast_fp16)[name = tensor("koi_33_cast_fp16")]; + tensor qo_17_axis_0 = const()[name = tensor("qo_17_axis_0"), val = tensor(-1)]; + tensor qo_17_cast_fp16 = stack(axis = qo_17_axis_0, values = (qor_33_cast_fp16, qoi_33_cast_fp16))[name = tensor("qo_17_cast_fp16")]; + tensor ko_17_axis_0 = const()[name = tensor("ko_17_axis_0"), val = tensor(-1)]; + tensor ko_17_cast_fp16 = stack(axis = ko_17_axis_0, values = (kor_33_cast_fp16, koi_33_cast_fp16))[name = tensor("ko_17_cast_fp16")]; + tensor var_3470 = const()[name = tensor("op_3470"), val = tensor([1, 1, 16, 64])]; + tensor q_51_cast_fp16 = reshape(shape = var_3470, x = qo_17_cast_fp16)[name = tensor("q_51_cast_fp16")]; + tensor var_3476 = const()[name = tensor("op_3476"), val = tensor([1, 1, 16, 64])]; + tensor k_35_cast_fp16 = reshape(shape = var_3476, x = ko_17_cast_fp16)[name = tensor("k_35_cast_fp16")]; + tensor _inversed_3501_y_0_to_fp16 = const()[name = tensor("_inversed_3501_y_0_to_fp16"), val = tensor(0x1p-9)]; + tensor _inversed_3501_cast_fp16 = mul(x = ts_53_cast_fp16, y = _inversed_3501_y_0_to_fp16)[name = tensor("_inversed_3501_cast_fp16")]; + tensor var_3502_cast_fp16 = floor(x = _inversed_3501_cast_fp16)[name = tensor("op_3502_cast_fp16")]; + tensor var_3503_to_fp16 = const()[name = tensor("op_3503_to_fp16"), val = tensor(0x1p+9)]; + tensor var_3504_cast_fp16 = mul(x = var_3502_cast_fp16, y = var_3503_to_fp16)[name = tensor("op_3504_cast_fp16")]; + tensor write_indices_float_35_cast_fp16 = sub(x = ts_53_cast_fp16, y = var_3504_cast_fp16)[name = tensor("write_indices_float_35_cast_fp16")]; + tensor var_3511_dtype_0 = const()[name = tensor("op_3511_dtype_0"), val = tensor("int32")]; + tensor write_indices_17_reps_0 = const()[name = tensor("write_indices_17_reps_0"), val = tensor([1, 1, 16, 64])]; + tensor write_indices_float_35_cast_fp16_to_int32 = cast(dtype = var_3511_dtype_0, x = write_indices_float_35_cast_fp16)[name = tensor("cast_512")]; + tensor write_indices_17 = tile(reps = write_indices_17_reps_0, x = write_indices_float_35_cast_fp16_to_int32)[name = tensor("write_indices_17")]; + tensor var_3521_begin_0 = const()[name = tensor("op_3521_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_3521_end_0 = const()[name = tensor("op_3521_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor var_3521_end_mask_0 = const()[name = tensor("op_3521_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_3521_squeeze_mask_0 = const()[name = tensor("op_3521_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor cache8_to_fp16_dtype_0 = const()[name = tensor("cache8_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor cache8_to_fp16 = cast(dtype = cache8_to_fp16_dtype_0, x = cache8)[name = tensor("cast_511")]; + tensor var_3521_cast_fp16 = slice_by_index(begin = var_3521_begin_0, end = var_3521_end_0, end_mask = var_3521_end_mask_0, squeeze_mask = var_3521_squeeze_mask_0, x = cache8_to_fp16)[name = tensor("op_3521_cast_fp16")]; + tensor var_3523_axis_0 = const()[name = tensor("op_3523_axis_0"), val = tensor(1)]; + tensor var_3523_mode_0 = const()[name = tensor("op_3523_mode_0"), val = tensor("update")]; + tensor var_3523_validate_indices_0 = const()[name = tensor("op_3523_validate_indices_0"), val = tensor(false)]; + tensor var_3523_cast_fp16 = scatter_along_axis(axis = var_3523_axis_0, data = var_3521_cast_fp16, indices = write_indices_17, mode = var_3523_mode_0, updates = k_35_cast_fp16, validate_indices = var_3523_validate_indices_0)[name = tensor("op_3523_cast_fp16")]; + tensor concat_58 = const()[name = tensor("concat_58"), val = tensor([0, 0, 0, 0, 0])]; + tensor concat_59 = const()[name = tensor("concat_59"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_17_internal_tensor_assign_1_stride_0 = const()[name = tensor("new_cache_17_internal_tensor_assign_1_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_17_internal_tensor_assign_1_begin_mask_0 = const()[name = tensor("new_cache_17_internal_tensor_assign_1_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_17_internal_tensor_assign_1_end_mask_0 = const()[name = tensor("new_cache_17_internal_tensor_assign_1_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_17_internal_tensor_assign_1_squeeze_mask_0 = const()[name = tensor("new_cache_17_internal_tensor_assign_1_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_64 = const()[name = tensor("shape_64"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_16 = const()[name = tensor("reduce_prod_16"), val = tensor(1048576)]; + tensor range_1d_16_start_0 = const()[name = tensor("range_1d_16_start_0"), val = tensor(0)]; + tensor range_1d_16_step_0 = const()[name = tensor("range_1d_16_step_0"), val = tensor(1)]; + tensor range_1d_16 = range_1d(end = reduce_prod_16, start = range_1d_16_start_0, step = range_1d_16_step_0)[name = tensor("range_1d_16")]; + tensor reshape_80 = reshape(shape = shape_64, x = range_1d_16)[name = tensor("reshape_80")]; + tensor slice_by_index_16 = slice_by_index(begin = concat_58, begin_mask = new_cache_17_internal_tensor_assign_1_begin_mask_0, end = concat_59, end_mask = new_cache_17_internal_tensor_assign_1_end_mask_0, squeeze_mask = new_cache_17_internal_tensor_assign_1_squeeze_mask_0, stride = new_cache_17_internal_tensor_assign_1_stride_0, x = reshape_80)[name = tensor("slice_by_index_16")]; + tensor reshape_81_shape_0 = const()[name = tensor("reshape_81_shape_0"), val = tensor([-1])]; + tensor reshape_81 = reshape(shape = reshape_81_shape_0, x = slice_by_index_16)[name = tensor("reshape_81")]; + tensor reshape_82_shape_0 = const()[name = tensor("reshape_82_shape_0"), val = tensor([-1])]; + tensor reshape_82_cast_fp16 = reshape(shape = reshape_82_shape_0, x = var_3523_cast_fp16)[name = tensor("reshape_82_cast_fp16")]; + tensor reshape_83_shape_0 = const()[name = tensor("reshape_83_shape_0"), val = tensor([-1])]; + tensor reshape_83_cast_fp16 = reshape(shape = reshape_83_shape_0, x = cache8_to_fp16)[name = tensor("reshape_83_cast_fp16")]; + tensor scatter_16_mode_0 = const()[name = tensor("scatter_16_mode_0"), val = tensor("update")]; + tensor scatter_16_axis_0 = const()[name = tensor("scatter_16_axis_0"), val = tensor(0)]; + tensor scatter_16_validate_indices_0 = const()[name = tensor("scatter_16_validate_indices_0"), val = tensor(false)]; + tensor scatter_16_cast_fp16 = scatter(axis = scatter_16_axis_0, data = reshape_83_cast_fp16, indices = reshape_81, mode = scatter_16_mode_0, updates = reshape_82_cast_fp16, validate_indices = scatter_16_validate_indices_0)[name = tensor("scatter_16_cast_fp16")]; + tensor reshape_84_cast_fp16 = reshape(shape = shape_64, x = scatter_16_cast_fp16)[name = tensor("reshape_84_cast_fp16")]; + tensor var_3531_begin_0 = const()[name = tensor("op_3531_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor var_3531_end_0 = const()[name = tensor("op_3531_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor var_3531_end_mask_0 = const()[name = tensor("op_3531_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_3531_squeeze_mask_0 = const()[name = tensor("op_3531_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor var_3531_cast_fp16 = slice_by_index(begin = var_3531_begin_0, end = var_3531_end_0, end_mask = var_3531_end_mask_0, squeeze_mask = var_3531_squeeze_mask_0, x = reshape_84_cast_fp16)[name = tensor("op_3531_cast_fp16")]; + tensor var_3533_axis_0 = const()[name = tensor("op_3533_axis_0"), val = tensor(1)]; + tensor var_3533_mode_0 = const()[name = tensor("op_3533_mode_0"), val = tensor("update")]; + tensor var_3533_validate_indices_0 = const()[name = tensor("op_3533_validate_indices_0"), val = tensor(false)]; + tensor var_3533_cast_fp16 = scatter_along_axis(axis = var_3533_axis_0, data = var_3531_cast_fp16, indices = write_indices_17, mode = var_3533_mode_0, updates = v_17_cast_fp16, validate_indices = var_3533_validate_indices_0)[name = tensor("op_3533_cast_fp16")]; + tensor concat_60 = const()[name = tensor("concat_60"), val = tensor([1, 0, 0, 0, 0])]; + tensor concat_61 = const()[name = tensor("concat_61"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_17_internal_tensor_assign_2_stride_0 = const()[name = tensor("new_cache_17_internal_tensor_assign_2_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_17_internal_tensor_assign_2_begin_mask_0 = const()[name = tensor("new_cache_17_internal_tensor_assign_2_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_17_internal_tensor_assign_2_end_mask_0 = const()[name = tensor("new_cache_17_internal_tensor_assign_2_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_17_internal_tensor_assign_2_squeeze_mask_0 = const()[name = tensor("new_cache_17_internal_tensor_assign_2_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_65 = const()[name = tensor("shape_65"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_17 = const()[name = tensor("reduce_prod_17"), val = tensor(1048576)]; + tensor range_1d_17_start_0 = const()[name = tensor("range_1d_17_start_0"), val = tensor(0)]; + tensor range_1d_17_step_0 = const()[name = tensor("range_1d_17_step_0"), val = tensor(1)]; + tensor range_1d_17 = range_1d(end = reduce_prod_17, start = range_1d_17_start_0, step = range_1d_17_step_0)[name = tensor("range_1d_17")]; + tensor reshape_85 = reshape(shape = shape_65, x = range_1d_17)[name = tensor("reshape_85")]; + tensor slice_by_index_17 = slice_by_index(begin = concat_60, begin_mask = new_cache_17_internal_tensor_assign_2_begin_mask_0, end = concat_61, end_mask = new_cache_17_internal_tensor_assign_2_end_mask_0, squeeze_mask = new_cache_17_internal_tensor_assign_2_squeeze_mask_0, stride = new_cache_17_internal_tensor_assign_2_stride_0, x = reshape_85)[name = tensor("slice_by_index_17")]; + tensor reshape_86_shape_0 = const()[name = tensor("reshape_86_shape_0"), val = tensor([-1])]; + tensor reshape_86 = reshape(shape = reshape_86_shape_0, x = slice_by_index_17)[name = tensor("reshape_86")]; + tensor reshape_87_shape_0 = const()[name = tensor("reshape_87_shape_0"), val = tensor([-1])]; + tensor reshape_87_cast_fp16 = reshape(shape = reshape_87_shape_0, x = var_3533_cast_fp16)[name = tensor("reshape_87_cast_fp16")]; + tensor reshape_88_shape_0 = const()[name = tensor("reshape_88_shape_0"), val = tensor([-1])]; + tensor reshape_88_cast_fp16 = reshape(shape = reshape_88_shape_0, x = reshape_84_cast_fp16)[name = tensor("reshape_88_cast_fp16")]; + tensor scatter_17_mode_0 = const()[name = tensor("scatter_17_mode_0"), val = tensor("update")]; + tensor scatter_17_axis_0 = const()[name = tensor("scatter_17_axis_0"), val = tensor(0)]; + tensor scatter_17_validate_indices_0 = const()[name = tensor("scatter_17_validate_indices_0"), val = tensor(false)]; + tensor scatter_17_cast_fp16 = scatter(axis = scatter_17_axis_0, data = reshape_88_cast_fp16, indices = reshape_86, mode = scatter_17_mode_0, updates = reshape_87_cast_fp16, validate_indices = scatter_17_validate_indices_0)[name = tensor("scatter_17_cast_fp16")]; + tensor reshape_89_cast_fp16 = reshape(shape = shape_65, x = scatter_17_cast_fp16)[name = tensor("reshape_89_cast_fp16")]; + tensor reshape_89_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("reshape_89_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor keys_49_begin_0 = const()[name = tensor("keys_49_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor keys_49_end_0 = const()[name = tensor("keys_49_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor keys_49_end_mask_0 = const()[name = tensor("keys_49_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor keys_49_squeeze_mask_0 = const()[name = tensor("keys_49_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor keys_49_cast_fp16 = slice_by_index(begin = keys_49_begin_0, end = keys_49_end_0, end_mask = keys_49_end_mask_0, squeeze_mask = keys_49_squeeze_mask_0, x = reshape_89_cast_fp16)[name = tensor("keys_49_cast_fp16")]; + tensor values_49_begin_0 = const()[name = tensor("values_49_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor values_49_end_0 = const()[name = tensor("values_49_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor values_49_end_mask_0 = const()[name = tensor("values_49_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor values_49_squeeze_mask_0 = const()[name = tensor("values_49_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor values_49_cast_fp16 = slice_by_index(begin = values_49_begin_0, end = values_49_end_0, end_mask = values_49_end_mask_0, squeeze_mask = values_49_squeeze_mask_0, x = reshape_89_cast_fp16)[name = tensor("values_49_cast_fp16")]; + tensor var_3545_cast_fp16 = not_equal(x = keys_49_cast_fp16, y = keys_49_cast_fp16)[name = tensor("op_3545_cast_fp16")]; + tensor keys_51_cast_fp16 = select(a = var_463_to_fp16, b = keys_49_cast_fp16, cond = var_3545_cast_fp16)[name = tensor("keys_51_cast_fp16")]; + tensor var_3553_cast_fp16 = not_equal(x = values_49_cast_fp16, y = values_49_cast_fp16)[name = tensor("op_3553_cast_fp16")]; + tensor values_51_cast_fp16 = select(a = var_463_to_fp16, b = values_49_cast_fp16, cond = var_3553_cast_fp16)[name = tensor("values_51_cast_fp16")]; + tensor var_3577 = const()[name = tensor("op_3577"), val = tensor([0, 2, 1, 3])]; + tensor var_3593 = const()[name = tensor("op_3593"), val = tensor([1, 1, 1])]; + tensor var_3594_cast_fp16 = reshape(shape = var_3593, x = position8_to_fp16)[name = tensor("op_3594_cast_fp16")]; + tensor var_3614_to_fp16 = const()[name = tensor("op_3614_to_fp16"), val = tensor(0x1p+0)]; + tensor valid_len_17_cast_fp16 = add(x = var_3594_cast_fp16, y = var_3614_to_fp16)[name = tensor("valid_len_17_cast_fp16")]; + tensor valid_mask_17_cast_fp16 = less(x = k_positions_1_promoted_to_fp16, y = valid_len_17_cast_fp16)[name = tensor("valid_mask_17_cast_fp16")]; + tensor causal_mask_17_cast_fp16 = less_equal(x = k_positions_1_promoted_to_fp16, y = var_3594_cast_fp16)[name = tensor("causal_mask_17_cast_fp16")]; + tensor attn_mask_33 = logical_and(x = valid_mask_17_cast_fp16, y = causal_mask_17_cast_fp16)[name = tensor("attn_mask_33")]; + tensor attn_mask_35_axes_0 = const()[name = tensor("attn_mask_35_axes_0"), val = tensor([1])]; + tensor attn_mask_35 = expand_dims(axes = attn_mask_35_axes_0, x = attn_mask_33)[name = tensor("attn_mask_35")]; + tensor var_3632_transpose_x_0 = const()[name = tensor("op_3632_transpose_x_0"), val = tensor(false)]; + tensor var_3632_transpose_y_0 = const()[name = tensor("op_3632_transpose_y_0"), val = tensor(false)]; + tensor transpose_88_perm_0 = const()[name = tensor("transpose_88_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_89_perm_0 = const()[name = tensor("transpose_89_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_89 = transpose(perm = transpose_89_perm_0, x = keys_51_cast_fp16)[name = tensor("transpose_181")]; + tensor transpose_88 = transpose(perm = transpose_88_perm_0, x = q_51_cast_fp16)[name = tensor("transpose_182")]; + tensor var_3632_cast_fp16 = matmul(transpose_x = var_3632_transpose_x_0, transpose_y = var_3632_transpose_y_0, x = transpose_88, y = transpose_89)[name = tensor("op_3632_cast_fp16")]; + tensor var_3626_to_fp16 = const()[name = tensor("op_3626_to_fp16"), val = tensor([0x1p-3])]; + tensor attn_weights_49_cast_fp16 = mul(x = var_3632_cast_fp16, y = var_3626_to_fp16)[name = tensor("attn_weights_49_cast_fp16")]; + tensor var_3634 = logical_not(x = attn_mask_35)[name = tensor("op_3634")]; + tensor var_3635_to_fp16 = const()[name = tensor("op_3635_to_fp16"), val = tensor(-inf)]; + tensor attn_weights_51_cast_fp16 = select(a = var_3635_to_fp16, b = attn_weights_49_cast_fp16, cond = var_3634)[name = tensor("attn_weights_51_cast_fp16")]; + tensor var_3637 = const()[name = tensor("op_3637"), val = tensor(-1)]; + tensor attn_weights_53_cast_fp16 = softmax(axis = var_3637, x = attn_weights_51_cast_fp16)[name = tensor("attn_weights_53_cast_fp16")]; + tensor attn_output_17_transpose_x_0 = const()[name = tensor("attn_output_17_transpose_x_0"), val = tensor(false)]; + tensor attn_output_17_transpose_y_0 = const()[name = tensor("attn_output_17_transpose_y_0"), val = tensor(false)]; + tensor values_53_cast_fp16 = transpose(perm = var_3577, x = values_51_cast_fp16)[name = tensor("transpose_183")]; + tensor attn_output_17_cast_fp16 = matmul(transpose_x = attn_output_17_transpose_x_0, transpose_y = attn_output_17_transpose_y_0, x = attn_weights_53_cast_fp16, y = values_53_cast_fp16)[name = tensor("attn_output_17_cast_fp16")]; + tensor var_3645 = const()[name = tensor("op_3645"), val = tensor([0, 2, 1, 3])]; + tensor var_3650 = const()[name = tensor("op_3650"), val = tensor([1, 1, 1024])]; + tensor var_3646_cast_fp16 = transpose(perm = var_3645, x = attn_output_17_cast_fp16)[name = tensor("transpose_180")]; + tensor input_103_cast_fp16 = reshape(shape = var_3650, x = var_3646_cast_fp16)[name = tensor("input_103_cast_fp16")]; + tensor attn8_out_proj_weight_to_fp16 = const()[name = tensor("attn8_out_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(208825088)))]; + tensor linear_34_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = attn8_out_proj_weight_to_fp16, x = input_103_cast_fp16)[name = tensor("linear_34_cast_fp16")]; + tensor var_3656_to_fp16 = const()[name = tensor("op_3656_to_fp16"), val = tensor(0x1p+0)]; + tensor var_3657_cast_fp16 = add(x = position8_to_fp16, y = var_3656_to_fp16)[name = tensor("op_3657_cast_fp16")]; + tensor var_3657_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_3657_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor input_105_cast_fp16 = add(x = input_99_cast_fp16, y = linear_34_cast_fp16)[name = tensor("input_105_cast_fp16")]; + tensor input_107_axes_0 = const()[name = tensor("input_107_axes_0"), val = tensor([-1])]; + tensor norm8_2_weight_to_fp16 = const()[name = tensor("norm8_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(210922304)))]; + tensor norm8_2_bias_to_fp16 = const()[name = tensor("norm8_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(210924416)))]; + tensor var_3661_to_fp16 = const()[name = tensor("op_3661_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_107_cast_fp16 = layer_norm(axes = input_107_axes_0, beta = norm8_2_bias_to_fp16, epsilon = var_3661_to_fp16, gamma = norm8_2_weight_to_fp16, x = input_105_cast_fp16)[name = tensor("input_107_cast_fp16")]; + tensor linear8_1_weight_to_fp16 = const()[name = tensor("linear8_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(210926528)))]; + tensor linear_35_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = linear8_1_weight_to_fp16, x = input_107_cast_fp16)[name = tensor("linear_35_cast_fp16")]; + tensor input_109_mode_0 = const()[name = tensor("input_109_mode_0"), val = tensor("EXACT")]; + tensor input_109_cast_fp16 = gelu(mode = input_109_mode_0, x = linear_35_cast_fp16)[name = tensor("input_109_cast_fp16")]; + tensor linear8_2_weight_to_fp16 = const()[name = tensor("linear8_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(219315200)))]; + tensor linear_36_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = linear8_2_weight_to_fp16, x = input_109_cast_fp16)[name = tensor("linear_36_cast_fp16")]; + tensor input_111_cast_fp16 = add(x = input_105_cast_fp16, y = linear_36_cast_fp16)[name = tensor("input_111_cast_fp16")]; + tensor input_113_axes_0 = const()[name = tensor("input_113_axes_0"), val = tensor([-1])]; + tensor norm9_1_weight_to_fp16 = const()[name = tensor("norm9_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(227703872)))]; + tensor norm9_1_bias_to_fp16 = const()[name = tensor("norm9_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(227705984)))]; + tensor var_3678_to_fp16 = const()[name = tensor("op_3678_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_113_cast_fp16 = layer_norm(axes = input_113_axes_0, beta = norm9_1_bias_to_fp16, epsilon = var_3678_to_fp16, gamma = norm9_1_weight_to_fp16, x = input_111_cast_fp16)[name = tensor("input_113_cast_fp16")]; + tensor attn9_in_proj_weight_to_fp16 = const()[name = tensor("attn9_in_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(227708096)))]; + tensor linear_37_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = attn9_in_proj_weight_to_fp16, x = input_113_cast_fp16)[name = tensor("linear_37_cast_fp16")]; + tensor var_3692 = const()[name = tensor("op_3692"), val = tensor([1, 1, 3, 16, 64])]; + tensor qkv_19_cast_fp16 = reshape(shape = var_3692, x = linear_37_cast_fp16)[name = tensor("qkv_19_cast_fp16")]; + tensor q_55_begin_0 = const()[name = tensor("q_55_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor q_55_end_0 = const()[name = tensor("q_55_end_0"), val = tensor([1, 1, 1, 16, 64])]; + tensor q_55_end_mask_0 = const()[name = tensor("q_55_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor q_55_squeeze_mask_0 = const()[name = tensor("q_55_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor q_55_cast_fp16 = slice_by_index(begin = q_55_begin_0, end = q_55_end_0, end_mask = q_55_end_mask_0, squeeze_mask = q_55_squeeze_mask_0, x = qkv_19_cast_fp16)[name = tensor("q_55_cast_fp16")]; + tensor k_37_begin_0 = const()[name = tensor("k_37_begin_0"), val = tensor([0, 0, 1, 0, 0])]; + tensor k_37_end_0 = const()[name = tensor("k_37_end_0"), val = tensor([1, 1, 2, 16, 64])]; + tensor k_37_end_mask_0 = const()[name = tensor("k_37_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor k_37_squeeze_mask_0 = const()[name = tensor("k_37_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor k_37_cast_fp16 = slice_by_index(begin = k_37_begin_0, end = k_37_end_0, end_mask = k_37_end_mask_0, squeeze_mask = k_37_squeeze_mask_0, x = qkv_19_cast_fp16)[name = tensor("k_37_cast_fp16")]; + tensor v_19_begin_0 = const()[name = tensor("v_19_begin_0"), val = tensor([0, 0, 2, 0, 0])]; + tensor v_19_end_0 = const()[name = tensor("v_19_end_0"), val = tensor([1, 1, 3, 16, 64])]; + tensor v_19_end_mask_0 = const()[name = tensor("v_19_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor v_19_squeeze_mask_0 = const()[name = tensor("v_19_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor v_19_cast_fp16 = slice_by_index(begin = v_19_begin_0, end = v_19_end_0, end_mask = v_19_end_mask_0, squeeze_mask = v_19_squeeze_mask_0, x = qkv_19_cast_fp16)[name = tensor("v_19_cast_fp16")]; + tensor var_3758 = const()[name = tensor("op_3758"), val = tensor([1, 1, 1, 1])]; + tensor position9_to_fp16_dtype_0 = const()[name = tensor("position9_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor position9_to_fp16 = cast(dtype = position9_to_fp16_dtype_0, x = position9)[name = tensor("cast_508")]; + tensor ts_59_cast_fp16 = reshape(shape = var_3758, x = position9_to_fp16)[name = tensor("ts_59_cast_fp16")]; + tensor var_3765 = const()[name = tensor("op_3765"), val = tensor([1, 1, 16, 32, 2])]; + tensor q_complex_19_cast_fp16 = reshape(shape = var_3765, x = q_55_cast_fp16)[name = tensor("q_complex_19_cast_fp16")]; + tensor var_3772 = const()[name = tensor("op_3772"), val = tensor([1, 1, 16, 32, 2])]; + tensor k_complex_19_cast_fp16 = reshape(shape = var_3772, x = k_37_cast_fp16)[name = tensor("k_complex_19_cast_fp16")]; + tensor var_3776_begin_0 = const()[name = tensor("op_3776_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_3776_end_0 = const()[name = tensor("op_3776_end_0"), val = tensor([1, 1, 16, 32, 1])]; + tensor var_3776_end_mask_0 = const()[name = tensor("op_3776_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_3776_squeeze_mask_0 = const()[name = tensor("op_3776_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_3776_cast_fp16 = slice_by_index(begin = var_3776_begin_0, end = var_3776_end_0, end_mask = var_3776_end_mask_0, squeeze_mask = var_3776_squeeze_mask_0, x = q_complex_19_cast_fp16)[name = tensor("op_3776_cast_fp16")]; + tensor var_3784_begin_0 = const()[name = tensor("op_3784_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_3784_end_0 = const()[name = tensor("op_3784_end_0"), val = tensor([1, 1, 16, 32, 2])]; + tensor var_3784_end_mask_0 = const()[name = tensor("op_3784_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_3784_squeeze_mask_0 = const()[name = tensor("op_3784_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_3784_cast_fp16 = slice_by_index(begin = var_3784_begin_0, end = var_3784_end_0, end_mask = var_3784_end_mask_0, squeeze_mask = var_3784_squeeze_mask_0, x = q_complex_19_cast_fp16)[name = tensor("op_3784_cast_fp16")]; + tensor var_3792_begin_0 = const()[name = tensor("op_3792_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_3792_end_0 = const()[name = tensor("op_3792_end_0"), val = tensor([1, 1, 16, 32, 1])]; + tensor var_3792_end_mask_0 = const()[name = tensor("op_3792_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_3792_squeeze_mask_0 = const()[name = tensor("op_3792_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_3792_cast_fp16 = slice_by_index(begin = var_3792_begin_0, end = var_3792_end_0, end_mask = var_3792_end_mask_0, squeeze_mask = var_3792_squeeze_mask_0, x = k_complex_19_cast_fp16)[name = tensor("op_3792_cast_fp16")]; + tensor var_3800_begin_0 = const()[name = tensor("op_3800_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_3800_end_0 = const()[name = tensor("op_3800_end_0"), val = tensor([1, 1, 16, 32, 2])]; + tensor var_3800_end_mask_0 = const()[name = tensor("op_3800_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_3800_squeeze_mask_0 = const()[name = tensor("op_3800_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_3800_cast_fp16 = slice_by_index(begin = var_3800_begin_0, end = var_3800_end_0, end_mask = var_3800_end_mask_0, squeeze_mask = var_3800_squeeze_mask_0, x = k_complex_19_cast_fp16)[name = tensor("op_3800_cast_fp16")]; + tensor freqs_19_to_fp16 = const()[name = tensor("freqs_19_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(233999616)))]; + tensor var_3806_cast_fp16 = mul(x = freqs_19_to_fp16, y = ts_59_cast_fp16)[name = tensor("op_3806_cast_fp16")]; + tensor rotr_19_cast_fp16 = cos(x = var_3806_cast_fp16)[name = tensor("rotr_19_cast_fp16")]; + tensor roti_19_cast_fp16 = sin(x = var_3806_cast_fp16)[name = tensor("roti_19_cast_fp16")]; + tensor var_3810_cast_fp16 = mul(x = var_3776_cast_fp16, y = rotr_19_cast_fp16)[name = tensor("op_3810_cast_fp16")]; + tensor var_3811_cast_fp16 = mul(x = var_3784_cast_fp16, y = roti_19_cast_fp16)[name = tensor("op_3811_cast_fp16")]; + tensor qor_37_cast_fp16 = sub(x = var_3810_cast_fp16, y = var_3811_cast_fp16)[name = tensor("qor_37_cast_fp16")]; + tensor var_3814_cast_fp16 = mul(x = var_3776_cast_fp16, y = roti_19_cast_fp16)[name = tensor("op_3814_cast_fp16")]; + tensor var_3815_cast_fp16 = mul(x = var_3784_cast_fp16, y = rotr_19_cast_fp16)[name = tensor("op_3815_cast_fp16")]; + tensor qoi_37_cast_fp16 = add(x = var_3814_cast_fp16, y = var_3815_cast_fp16)[name = tensor("qoi_37_cast_fp16")]; + tensor var_3818_cast_fp16 = mul(x = var_3792_cast_fp16, y = rotr_19_cast_fp16)[name = tensor("op_3818_cast_fp16")]; + tensor var_3819_cast_fp16 = mul(x = var_3800_cast_fp16, y = roti_19_cast_fp16)[name = tensor("op_3819_cast_fp16")]; + tensor kor_37_cast_fp16 = sub(x = var_3818_cast_fp16, y = var_3819_cast_fp16)[name = tensor("kor_37_cast_fp16")]; + tensor var_3822_cast_fp16 = mul(x = var_3792_cast_fp16, y = roti_19_cast_fp16)[name = tensor("op_3822_cast_fp16")]; + tensor var_3823_cast_fp16 = mul(x = var_3800_cast_fp16, y = rotr_19_cast_fp16)[name = tensor("op_3823_cast_fp16")]; + tensor koi_37_cast_fp16 = add(x = var_3822_cast_fp16, y = var_3823_cast_fp16)[name = tensor("koi_37_cast_fp16")]; + tensor qo_19_axis_0 = const()[name = tensor("qo_19_axis_0"), val = tensor(-1)]; + tensor qo_19_cast_fp16 = stack(axis = qo_19_axis_0, values = (qor_37_cast_fp16, qoi_37_cast_fp16))[name = tensor("qo_19_cast_fp16")]; + tensor ko_19_axis_0 = const()[name = tensor("ko_19_axis_0"), val = tensor(-1)]; + tensor ko_19_cast_fp16 = stack(axis = ko_19_axis_0, values = (kor_37_cast_fp16, koi_37_cast_fp16))[name = tensor("ko_19_cast_fp16")]; + tensor var_3856 = const()[name = tensor("op_3856"), val = tensor([1, 1, 16, 64])]; + tensor q_57_cast_fp16 = reshape(shape = var_3856, x = qo_19_cast_fp16)[name = tensor("q_57_cast_fp16")]; + tensor var_3862 = const()[name = tensor("op_3862"), val = tensor([1, 1, 16, 64])]; + tensor k_39_cast_fp16 = reshape(shape = var_3862, x = ko_19_cast_fp16)[name = tensor("k_39_cast_fp16")]; + tensor _inversed_3887_y_0_to_fp16 = const()[name = tensor("_inversed_3887_y_0_to_fp16"), val = tensor(0x1p-9)]; + tensor _inversed_3887_cast_fp16 = mul(x = ts_59_cast_fp16, y = _inversed_3887_y_0_to_fp16)[name = tensor("_inversed_3887_cast_fp16")]; + tensor var_3888_cast_fp16 = floor(x = _inversed_3887_cast_fp16)[name = tensor("op_3888_cast_fp16")]; + tensor var_3889_to_fp16 = const()[name = tensor("op_3889_to_fp16"), val = tensor(0x1p+9)]; + tensor var_3890_cast_fp16 = mul(x = var_3888_cast_fp16, y = var_3889_to_fp16)[name = tensor("op_3890_cast_fp16")]; + tensor write_indices_float_39_cast_fp16 = sub(x = ts_59_cast_fp16, y = var_3890_cast_fp16)[name = tensor("write_indices_float_39_cast_fp16")]; + tensor var_3897_dtype_0 = const()[name = tensor("op_3897_dtype_0"), val = tensor("int32")]; + tensor write_indices_19_reps_0 = const()[name = tensor("write_indices_19_reps_0"), val = tensor([1, 1, 16, 64])]; + tensor write_indices_float_39_cast_fp16_to_int32 = cast(dtype = var_3897_dtype_0, x = write_indices_float_39_cast_fp16)[name = tensor("cast_507")]; + tensor write_indices_19 = tile(reps = write_indices_19_reps_0, x = write_indices_float_39_cast_fp16_to_int32)[name = tensor("write_indices_19")]; + tensor var_3907_begin_0 = const()[name = tensor("op_3907_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_3907_end_0 = const()[name = tensor("op_3907_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor var_3907_end_mask_0 = const()[name = tensor("op_3907_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_3907_squeeze_mask_0 = const()[name = tensor("op_3907_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor cache9_to_fp16_dtype_0 = const()[name = tensor("cache9_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor cache9_to_fp16 = cast(dtype = cache9_to_fp16_dtype_0, x = cache9)[name = tensor("cast_506")]; + tensor var_3907_cast_fp16 = slice_by_index(begin = var_3907_begin_0, end = var_3907_end_0, end_mask = var_3907_end_mask_0, squeeze_mask = var_3907_squeeze_mask_0, x = cache9_to_fp16)[name = tensor("op_3907_cast_fp16")]; + tensor var_3909_axis_0 = const()[name = tensor("op_3909_axis_0"), val = tensor(1)]; + tensor var_3909_mode_0 = const()[name = tensor("op_3909_mode_0"), val = tensor("update")]; + tensor var_3909_validate_indices_0 = const()[name = tensor("op_3909_validate_indices_0"), val = tensor(false)]; + tensor var_3909_cast_fp16 = scatter_along_axis(axis = var_3909_axis_0, data = var_3907_cast_fp16, indices = write_indices_19, mode = var_3909_mode_0, updates = k_39_cast_fp16, validate_indices = var_3909_validate_indices_0)[name = tensor("op_3909_cast_fp16")]; + tensor concat_65 = const()[name = tensor("concat_65"), val = tensor([0, 0, 0, 0, 0])]; + tensor concat_66 = const()[name = tensor("concat_66"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_19_internal_tensor_assign_1_stride_0 = const()[name = tensor("new_cache_19_internal_tensor_assign_1_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_19_internal_tensor_assign_1_begin_mask_0 = const()[name = tensor("new_cache_19_internal_tensor_assign_1_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_19_internal_tensor_assign_1_end_mask_0 = const()[name = tensor("new_cache_19_internal_tensor_assign_1_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_19_internal_tensor_assign_1_squeeze_mask_0 = const()[name = tensor("new_cache_19_internal_tensor_assign_1_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_66 = const()[name = tensor("shape_66"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_18 = const()[name = tensor("reduce_prod_18"), val = tensor(1048576)]; + tensor range_1d_18_start_0 = const()[name = tensor("range_1d_18_start_0"), val = tensor(0)]; + tensor range_1d_18_step_0 = const()[name = tensor("range_1d_18_step_0"), val = tensor(1)]; + tensor range_1d_18 = range_1d(end = reduce_prod_18, start = range_1d_18_start_0, step = range_1d_18_step_0)[name = tensor("range_1d_18")]; + tensor reshape_90 = reshape(shape = shape_66, x = range_1d_18)[name = tensor("reshape_90")]; + tensor slice_by_index_18 = slice_by_index(begin = concat_65, begin_mask = new_cache_19_internal_tensor_assign_1_begin_mask_0, end = concat_66, end_mask = new_cache_19_internal_tensor_assign_1_end_mask_0, squeeze_mask = new_cache_19_internal_tensor_assign_1_squeeze_mask_0, stride = new_cache_19_internal_tensor_assign_1_stride_0, x = reshape_90)[name = tensor("slice_by_index_18")]; + tensor reshape_91_shape_0 = const()[name = tensor("reshape_91_shape_0"), val = tensor([-1])]; + tensor reshape_91 = reshape(shape = reshape_91_shape_0, x = slice_by_index_18)[name = tensor("reshape_91")]; + tensor reshape_92_shape_0 = const()[name = tensor("reshape_92_shape_0"), val = tensor([-1])]; + tensor reshape_92_cast_fp16 = reshape(shape = reshape_92_shape_0, x = var_3909_cast_fp16)[name = tensor("reshape_92_cast_fp16")]; + tensor reshape_93_shape_0 = const()[name = tensor("reshape_93_shape_0"), val = tensor([-1])]; + tensor reshape_93_cast_fp16 = reshape(shape = reshape_93_shape_0, x = cache9_to_fp16)[name = tensor("reshape_93_cast_fp16")]; + tensor scatter_18_mode_0 = const()[name = tensor("scatter_18_mode_0"), val = tensor("update")]; + tensor scatter_18_axis_0 = const()[name = tensor("scatter_18_axis_0"), val = tensor(0)]; + tensor scatter_18_validate_indices_0 = const()[name = tensor("scatter_18_validate_indices_0"), val = tensor(false)]; + tensor scatter_18_cast_fp16 = scatter(axis = scatter_18_axis_0, data = reshape_93_cast_fp16, indices = reshape_91, mode = scatter_18_mode_0, updates = reshape_92_cast_fp16, validate_indices = scatter_18_validate_indices_0)[name = tensor("scatter_18_cast_fp16")]; + tensor reshape_94_cast_fp16 = reshape(shape = shape_66, x = scatter_18_cast_fp16)[name = tensor("reshape_94_cast_fp16")]; + tensor var_3917_begin_0 = const()[name = tensor("op_3917_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor var_3917_end_0 = const()[name = tensor("op_3917_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor var_3917_end_mask_0 = const()[name = tensor("op_3917_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_3917_squeeze_mask_0 = const()[name = tensor("op_3917_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor var_3917_cast_fp16 = slice_by_index(begin = var_3917_begin_0, end = var_3917_end_0, end_mask = var_3917_end_mask_0, squeeze_mask = var_3917_squeeze_mask_0, x = reshape_94_cast_fp16)[name = tensor("op_3917_cast_fp16")]; + tensor var_3919_axis_0 = const()[name = tensor("op_3919_axis_0"), val = tensor(1)]; + tensor var_3919_mode_0 = const()[name = tensor("op_3919_mode_0"), val = tensor("update")]; + tensor var_3919_validate_indices_0 = const()[name = tensor("op_3919_validate_indices_0"), val = tensor(false)]; + tensor var_3919_cast_fp16 = scatter_along_axis(axis = var_3919_axis_0, data = var_3917_cast_fp16, indices = write_indices_19, mode = var_3919_mode_0, updates = v_19_cast_fp16, validate_indices = var_3919_validate_indices_0)[name = tensor("op_3919_cast_fp16")]; + tensor concat_67 = const()[name = tensor("concat_67"), val = tensor([1, 0, 0, 0, 0])]; + tensor concat_68 = const()[name = tensor("concat_68"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_19_internal_tensor_assign_2_stride_0 = const()[name = tensor("new_cache_19_internal_tensor_assign_2_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_19_internal_tensor_assign_2_begin_mask_0 = const()[name = tensor("new_cache_19_internal_tensor_assign_2_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_19_internal_tensor_assign_2_end_mask_0 = const()[name = tensor("new_cache_19_internal_tensor_assign_2_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_19_internal_tensor_assign_2_squeeze_mask_0 = const()[name = tensor("new_cache_19_internal_tensor_assign_2_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_67 = const()[name = tensor("shape_67"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_19 = const()[name = tensor("reduce_prod_19"), val = tensor(1048576)]; + tensor range_1d_19_start_0 = const()[name = tensor("range_1d_19_start_0"), val = tensor(0)]; + tensor range_1d_19_step_0 = const()[name = tensor("range_1d_19_step_0"), val = tensor(1)]; + tensor range_1d_19 = range_1d(end = reduce_prod_19, start = range_1d_19_start_0, step = range_1d_19_step_0)[name = tensor("range_1d_19")]; + tensor reshape_95 = reshape(shape = shape_67, x = range_1d_19)[name = tensor("reshape_95")]; + tensor slice_by_index_19 = slice_by_index(begin = concat_67, begin_mask = new_cache_19_internal_tensor_assign_2_begin_mask_0, end = concat_68, end_mask = new_cache_19_internal_tensor_assign_2_end_mask_0, squeeze_mask = new_cache_19_internal_tensor_assign_2_squeeze_mask_0, stride = new_cache_19_internal_tensor_assign_2_stride_0, x = reshape_95)[name = tensor("slice_by_index_19")]; + tensor reshape_96_shape_0 = const()[name = tensor("reshape_96_shape_0"), val = tensor([-1])]; + tensor reshape_96 = reshape(shape = reshape_96_shape_0, x = slice_by_index_19)[name = tensor("reshape_96")]; + tensor reshape_97_shape_0 = const()[name = tensor("reshape_97_shape_0"), val = tensor([-1])]; + tensor reshape_97_cast_fp16 = reshape(shape = reshape_97_shape_0, x = var_3919_cast_fp16)[name = tensor("reshape_97_cast_fp16")]; + tensor reshape_98_shape_0 = const()[name = tensor("reshape_98_shape_0"), val = tensor([-1])]; + tensor reshape_98_cast_fp16 = reshape(shape = reshape_98_shape_0, x = reshape_94_cast_fp16)[name = tensor("reshape_98_cast_fp16")]; + tensor scatter_19_mode_0 = const()[name = tensor("scatter_19_mode_0"), val = tensor("update")]; + tensor scatter_19_axis_0 = const()[name = tensor("scatter_19_axis_0"), val = tensor(0)]; + tensor scatter_19_validate_indices_0 = const()[name = tensor("scatter_19_validate_indices_0"), val = tensor(false)]; + tensor scatter_19_cast_fp16 = scatter(axis = scatter_19_axis_0, data = reshape_98_cast_fp16, indices = reshape_96, mode = scatter_19_mode_0, updates = reshape_97_cast_fp16, validate_indices = scatter_19_validate_indices_0)[name = tensor("scatter_19_cast_fp16")]; + tensor reshape_99_cast_fp16 = reshape(shape = shape_67, x = scatter_19_cast_fp16)[name = tensor("reshape_99_cast_fp16")]; + tensor reshape_99_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("reshape_99_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor keys_55_begin_0 = const()[name = tensor("keys_55_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor keys_55_end_0 = const()[name = tensor("keys_55_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor keys_55_end_mask_0 = const()[name = tensor("keys_55_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor keys_55_squeeze_mask_0 = const()[name = tensor("keys_55_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor keys_55_cast_fp16 = slice_by_index(begin = keys_55_begin_0, end = keys_55_end_0, end_mask = keys_55_end_mask_0, squeeze_mask = keys_55_squeeze_mask_0, x = reshape_99_cast_fp16)[name = tensor("keys_55_cast_fp16")]; + tensor values_55_begin_0 = const()[name = tensor("values_55_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor values_55_end_0 = const()[name = tensor("values_55_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor values_55_end_mask_0 = const()[name = tensor("values_55_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor values_55_squeeze_mask_0 = const()[name = tensor("values_55_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor values_55_cast_fp16 = slice_by_index(begin = values_55_begin_0, end = values_55_end_0, end_mask = values_55_end_mask_0, squeeze_mask = values_55_squeeze_mask_0, x = reshape_99_cast_fp16)[name = tensor("values_55_cast_fp16")]; + tensor var_3931_cast_fp16 = not_equal(x = keys_55_cast_fp16, y = keys_55_cast_fp16)[name = tensor("op_3931_cast_fp16")]; + tensor keys_57_cast_fp16 = select(a = var_463_to_fp16, b = keys_55_cast_fp16, cond = var_3931_cast_fp16)[name = tensor("keys_57_cast_fp16")]; + tensor var_3939_cast_fp16 = not_equal(x = values_55_cast_fp16, y = values_55_cast_fp16)[name = tensor("op_3939_cast_fp16")]; + tensor values_57_cast_fp16 = select(a = var_463_to_fp16, b = values_55_cast_fp16, cond = var_3939_cast_fp16)[name = tensor("values_57_cast_fp16")]; + tensor var_3963 = const()[name = tensor("op_3963"), val = tensor([0, 2, 1, 3])]; + tensor var_3979 = const()[name = tensor("op_3979"), val = tensor([1, 1, 1])]; + tensor var_3980_cast_fp16 = reshape(shape = var_3979, x = position9_to_fp16)[name = tensor("op_3980_cast_fp16")]; + tensor var_4000_to_fp16 = const()[name = tensor("op_4000_to_fp16"), val = tensor(0x1p+0)]; + tensor valid_len_19_cast_fp16 = add(x = var_3980_cast_fp16, y = var_4000_to_fp16)[name = tensor("valid_len_19_cast_fp16")]; + tensor valid_mask_19_cast_fp16 = less(x = k_positions_1_promoted_to_fp16, y = valid_len_19_cast_fp16)[name = tensor("valid_mask_19_cast_fp16")]; + tensor causal_mask_19_cast_fp16 = less_equal(x = k_positions_1_promoted_to_fp16, y = var_3980_cast_fp16)[name = tensor("causal_mask_19_cast_fp16")]; + tensor attn_mask_37 = logical_and(x = valid_mask_19_cast_fp16, y = causal_mask_19_cast_fp16)[name = tensor("attn_mask_37")]; + tensor attn_mask_39_axes_0 = const()[name = tensor("attn_mask_39_axes_0"), val = tensor([1])]; + tensor attn_mask_39 = expand_dims(axes = attn_mask_39_axes_0, x = attn_mask_37)[name = tensor("attn_mask_39")]; + tensor var_4018_transpose_x_0 = const()[name = tensor("op_4018_transpose_x_0"), val = tensor(false)]; + tensor var_4018_transpose_y_0 = const()[name = tensor("op_4018_transpose_y_0"), val = tensor(false)]; + tensor transpose_90_perm_0 = const()[name = tensor("transpose_90_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_91_perm_0 = const()[name = tensor("transpose_91_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_91 = transpose(perm = transpose_91_perm_0, x = keys_57_cast_fp16)[name = tensor("transpose_177")]; + tensor transpose_90 = transpose(perm = transpose_90_perm_0, x = q_57_cast_fp16)[name = tensor("transpose_178")]; + tensor var_4018_cast_fp16 = matmul(transpose_x = var_4018_transpose_x_0, transpose_y = var_4018_transpose_y_0, x = transpose_90, y = transpose_91)[name = tensor("op_4018_cast_fp16")]; + tensor var_4012_to_fp16 = const()[name = tensor("op_4012_to_fp16"), val = tensor([0x1p-3])]; + tensor attn_weights_55_cast_fp16 = mul(x = var_4018_cast_fp16, y = var_4012_to_fp16)[name = tensor("attn_weights_55_cast_fp16")]; + tensor var_4020 = logical_not(x = attn_mask_39)[name = tensor("op_4020")]; + tensor var_4021_to_fp16 = const()[name = tensor("op_4021_to_fp16"), val = tensor(-inf)]; + tensor attn_weights_57_cast_fp16 = select(a = var_4021_to_fp16, b = attn_weights_55_cast_fp16, cond = var_4020)[name = tensor("attn_weights_57_cast_fp16")]; + tensor var_4023 = const()[name = tensor("op_4023"), val = tensor(-1)]; + tensor attn_weights_59_cast_fp16 = softmax(axis = var_4023, x = attn_weights_57_cast_fp16)[name = tensor("attn_weights_59_cast_fp16")]; + tensor attn_output_19_transpose_x_0 = const()[name = tensor("attn_output_19_transpose_x_0"), val = tensor(false)]; + tensor attn_output_19_transpose_y_0 = const()[name = tensor("attn_output_19_transpose_y_0"), val = tensor(false)]; + tensor values_59_cast_fp16 = transpose(perm = var_3963, x = values_57_cast_fp16)[name = tensor("transpose_179")]; + tensor attn_output_19_cast_fp16 = matmul(transpose_x = attn_output_19_transpose_x_0, transpose_y = attn_output_19_transpose_y_0, x = attn_weights_59_cast_fp16, y = values_59_cast_fp16)[name = tensor("attn_output_19_cast_fp16")]; + tensor var_4031 = const()[name = tensor("op_4031"), val = tensor([0, 2, 1, 3])]; + tensor var_4036 = const()[name = tensor("op_4036"), val = tensor([1, 1, 1024])]; + tensor var_4032_cast_fp16 = transpose(perm = var_4031, x = attn_output_19_cast_fp16)[name = tensor("transpose_176")]; + tensor input_115_cast_fp16 = reshape(shape = var_4036, x = var_4032_cast_fp16)[name = tensor("input_115_cast_fp16")]; + tensor attn9_out_proj_weight_to_fp16 = const()[name = tensor("attn9_out_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(233999744)))]; + tensor linear_38_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = attn9_out_proj_weight_to_fp16, x = input_115_cast_fp16)[name = tensor("linear_38_cast_fp16")]; + tensor var_4042_to_fp16 = const()[name = tensor("op_4042_to_fp16"), val = tensor(0x1p+0)]; + tensor var_4043_cast_fp16 = add(x = position9_to_fp16, y = var_4042_to_fp16)[name = tensor("op_4043_cast_fp16")]; + tensor var_4043_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_4043_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor input_117_cast_fp16 = add(x = input_111_cast_fp16, y = linear_38_cast_fp16)[name = tensor("input_117_cast_fp16")]; + tensor input_119_axes_0 = const()[name = tensor("input_119_axes_0"), val = tensor([-1])]; + tensor norm9_2_weight_to_fp16 = const()[name = tensor("norm9_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(236096960)))]; + tensor norm9_2_bias_to_fp16 = const()[name = tensor("norm9_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(236099072)))]; + tensor var_4047_to_fp16 = const()[name = tensor("op_4047_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_119_cast_fp16 = layer_norm(axes = input_119_axes_0, beta = norm9_2_bias_to_fp16, epsilon = var_4047_to_fp16, gamma = norm9_2_weight_to_fp16, x = input_117_cast_fp16)[name = tensor("input_119_cast_fp16")]; + tensor linear9_1_weight_to_fp16 = const()[name = tensor("linear9_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(236101184)))]; + tensor linear_39_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = linear9_1_weight_to_fp16, x = input_119_cast_fp16)[name = tensor("linear_39_cast_fp16")]; + tensor input_121_mode_0 = const()[name = tensor("input_121_mode_0"), val = tensor("EXACT")]; + tensor input_121_cast_fp16 = gelu(mode = input_121_mode_0, x = linear_39_cast_fp16)[name = tensor("input_121_cast_fp16")]; + tensor linear9_2_weight_to_fp16 = const()[name = tensor("linear9_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(244489856)))]; + tensor linear_40_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = linear9_2_weight_to_fp16, x = input_121_cast_fp16)[name = tensor("linear_40_cast_fp16")]; + tensor input_123_cast_fp16 = add(x = input_117_cast_fp16, y = linear_40_cast_fp16)[name = tensor("input_123_cast_fp16")]; + tensor input_125_axes_0 = const()[name = tensor("input_125_axes_0"), val = tensor([-1])]; + tensor norm10_1_weight_to_fp16 = const()[name = tensor("norm10_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(252878528)))]; + tensor norm10_1_bias_to_fp16 = const()[name = tensor("norm10_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(252880640)))]; + tensor var_4064_to_fp16 = const()[name = tensor("op_4064_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_125_cast_fp16 = layer_norm(axes = input_125_axes_0, beta = norm10_1_bias_to_fp16, epsilon = var_4064_to_fp16, gamma = norm10_1_weight_to_fp16, x = input_123_cast_fp16)[name = tensor("input_125_cast_fp16")]; + tensor attn10_in_proj_weight_to_fp16 = const()[name = tensor("attn10_in_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(252882752)))]; + tensor linear_41_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = attn10_in_proj_weight_to_fp16, x = input_125_cast_fp16)[name = tensor("linear_41_cast_fp16")]; + tensor var_4078 = const()[name = tensor("op_4078"), val = tensor([1, 1, 3, 16, 64])]; + tensor qkv_21_cast_fp16 = reshape(shape = var_4078, x = linear_41_cast_fp16)[name = tensor("qkv_21_cast_fp16")]; + tensor q_61_begin_0 = const()[name = tensor("q_61_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor q_61_end_0 = const()[name = tensor("q_61_end_0"), val = tensor([1, 1, 1, 16, 64])]; + tensor q_61_end_mask_0 = const()[name = tensor("q_61_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor q_61_squeeze_mask_0 = const()[name = tensor("q_61_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor q_61_cast_fp16 = slice_by_index(begin = q_61_begin_0, end = q_61_end_0, end_mask = q_61_end_mask_0, squeeze_mask = q_61_squeeze_mask_0, x = qkv_21_cast_fp16)[name = tensor("q_61_cast_fp16")]; + tensor k_41_begin_0 = const()[name = tensor("k_41_begin_0"), val = tensor([0, 0, 1, 0, 0])]; + tensor k_41_end_0 = const()[name = tensor("k_41_end_0"), val = tensor([1, 1, 2, 16, 64])]; + tensor k_41_end_mask_0 = const()[name = tensor("k_41_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor k_41_squeeze_mask_0 = const()[name = tensor("k_41_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor k_41_cast_fp16 = slice_by_index(begin = k_41_begin_0, end = k_41_end_0, end_mask = k_41_end_mask_0, squeeze_mask = k_41_squeeze_mask_0, x = qkv_21_cast_fp16)[name = tensor("k_41_cast_fp16")]; + tensor v_21_begin_0 = const()[name = tensor("v_21_begin_0"), val = tensor([0, 0, 2, 0, 0])]; + tensor v_21_end_0 = const()[name = tensor("v_21_end_0"), val = tensor([1, 1, 3, 16, 64])]; + tensor v_21_end_mask_0 = const()[name = tensor("v_21_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor v_21_squeeze_mask_0 = const()[name = tensor("v_21_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor v_21_cast_fp16 = slice_by_index(begin = v_21_begin_0, end = v_21_end_0, end_mask = v_21_end_mask_0, squeeze_mask = v_21_squeeze_mask_0, x = qkv_21_cast_fp16)[name = tensor("v_21_cast_fp16")]; + tensor var_4144 = const()[name = tensor("op_4144"), val = tensor([1, 1, 1, 1])]; + tensor position10_to_fp16_dtype_0 = const()[name = tensor("position10_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor position10_to_fp16 = cast(dtype = position10_to_fp16_dtype_0, x = position10)[name = tensor("cast_503")]; + tensor ts_65_cast_fp16 = reshape(shape = var_4144, x = position10_to_fp16)[name = tensor("ts_65_cast_fp16")]; + tensor var_4151 = const()[name = tensor("op_4151"), val = tensor([1, 1, 16, 32, 2])]; + tensor q_complex_21_cast_fp16 = reshape(shape = var_4151, x = q_61_cast_fp16)[name = tensor("q_complex_21_cast_fp16")]; + tensor var_4158 = const()[name = tensor("op_4158"), val = tensor([1, 1, 16, 32, 2])]; + tensor k_complex_21_cast_fp16 = reshape(shape = var_4158, x = k_41_cast_fp16)[name = tensor("k_complex_21_cast_fp16")]; + tensor var_4162_begin_0 = const()[name = tensor("op_4162_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_4162_end_0 = const()[name = tensor("op_4162_end_0"), val = tensor([1, 1, 16, 32, 1])]; + tensor var_4162_end_mask_0 = const()[name = tensor("op_4162_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_4162_squeeze_mask_0 = const()[name = tensor("op_4162_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_4162_cast_fp16 = slice_by_index(begin = var_4162_begin_0, end = var_4162_end_0, end_mask = var_4162_end_mask_0, squeeze_mask = var_4162_squeeze_mask_0, x = q_complex_21_cast_fp16)[name = tensor("op_4162_cast_fp16")]; + tensor var_4170_begin_0 = const()[name = tensor("op_4170_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_4170_end_0 = const()[name = tensor("op_4170_end_0"), val = tensor([1, 1, 16, 32, 2])]; + tensor var_4170_end_mask_0 = const()[name = tensor("op_4170_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_4170_squeeze_mask_0 = const()[name = tensor("op_4170_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_4170_cast_fp16 = slice_by_index(begin = var_4170_begin_0, end = var_4170_end_0, end_mask = var_4170_end_mask_0, squeeze_mask = var_4170_squeeze_mask_0, x = q_complex_21_cast_fp16)[name = tensor("op_4170_cast_fp16")]; + tensor var_4178_begin_0 = const()[name = tensor("op_4178_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_4178_end_0 = const()[name = tensor("op_4178_end_0"), val = tensor([1, 1, 16, 32, 1])]; + tensor var_4178_end_mask_0 = const()[name = tensor("op_4178_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_4178_squeeze_mask_0 = const()[name = tensor("op_4178_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_4178_cast_fp16 = slice_by_index(begin = var_4178_begin_0, end = var_4178_end_0, end_mask = var_4178_end_mask_0, squeeze_mask = var_4178_squeeze_mask_0, x = k_complex_21_cast_fp16)[name = tensor("op_4178_cast_fp16")]; + tensor var_4186_begin_0 = const()[name = tensor("op_4186_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_4186_end_0 = const()[name = tensor("op_4186_end_0"), val = tensor([1, 1, 16, 32, 2])]; + tensor var_4186_end_mask_0 = const()[name = tensor("op_4186_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_4186_squeeze_mask_0 = const()[name = tensor("op_4186_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_4186_cast_fp16 = slice_by_index(begin = var_4186_begin_0, end = var_4186_end_0, end_mask = var_4186_end_mask_0, squeeze_mask = var_4186_squeeze_mask_0, x = k_complex_21_cast_fp16)[name = tensor("op_4186_cast_fp16")]; + tensor freqs_21_to_fp16 = const()[name = tensor("freqs_21_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(259174272)))]; + tensor var_4192_cast_fp16 = mul(x = freqs_21_to_fp16, y = ts_65_cast_fp16)[name = tensor("op_4192_cast_fp16")]; + tensor rotr_21_cast_fp16 = cos(x = var_4192_cast_fp16)[name = tensor("rotr_21_cast_fp16")]; + tensor roti_21_cast_fp16 = sin(x = var_4192_cast_fp16)[name = tensor("roti_21_cast_fp16")]; + tensor var_4196_cast_fp16 = mul(x = var_4162_cast_fp16, y = rotr_21_cast_fp16)[name = tensor("op_4196_cast_fp16")]; + tensor var_4197_cast_fp16 = mul(x = var_4170_cast_fp16, y = roti_21_cast_fp16)[name = tensor("op_4197_cast_fp16")]; + tensor qor_41_cast_fp16 = sub(x = var_4196_cast_fp16, y = var_4197_cast_fp16)[name = tensor("qor_41_cast_fp16")]; + tensor var_4200_cast_fp16 = mul(x = var_4162_cast_fp16, y = roti_21_cast_fp16)[name = tensor("op_4200_cast_fp16")]; + tensor var_4201_cast_fp16 = mul(x = var_4170_cast_fp16, y = rotr_21_cast_fp16)[name = tensor("op_4201_cast_fp16")]; + tensor qoi_41_cast_fp16 = add(x = var_4200_cast_fp16, y = var_4201_cast_fp16)[name = tensor("qoi_41_cast_fp16")]; + tensor var_4204_cast_fp16 = mul(x = var_4178_cast_fp16, y = rotr_21_cast_fp16)[name = tensor("op_4204_cast_fp16")]; + tensor var_4205_cast_fp16 = mul(x = var_4186_cast_fp16, y = roti_21_cast_fp16)[name = tensor("op_4205_cast_fp16")]; + tensor kor_41_cast_fp16 = sub(x = var_4204_cast_fp16, y = var_4205_cast_fp16)[name = tensor("kor_41_cast_fp16")]; + tensor var_4208_cast_fp16 = mul(x = var_4178_cast_fp16, y = roti_21_cast_fp16)[name = tensor("op_4208_cast_fp16")]; + tensor var_4209_cast_fp16 = mul(x = var_4186_cast_fp16, y = rotr_21_cast_fp16)[name = tensor("op_4209_cast_fp16")]; + tensor koi_41_cast_fp16 = add(x = var_4208_cast_fp16, y = var_4209_cast_fp16)[name = tensor("koi_41_cast_fp16")]; + tensor qo_21_axis_0 = const()[name = tensor("qo_21_axis_0"), val = tensor(-1)]; + tensor qo_21_cast_fp16 = stack(axis = qo_21_axis_0, values = (qor_41_cast_fp16, qoi_41_cast_fp16))[name = tensor("qo_21_cast_fp16")]; + tensor ko_21_axis_0 = const()[name = tensor("ko_21_axis_0"), val = tensor(-1)]; + tensor ko_21_cast_fp16 = stack(axis = ko_21_axis_0, values = (kor_41_cast_fp16, koi_41_cast_fp16))[name = tensor("ko_21_cast_fp16")]; + tensor var_4242 = const()[name = tensor("op_4242"), val = tensor([1, 1, 16, 64])]; + tensor q_63_cast_fp16 = reshape(shape = var_4242, x = qo_21_cast_fp16)[name = tensor("q_63_cast_fp16")]; + tensor var_4248 = const()[name = tensor("op_4248"), val = tensor([1, 1, 16, 64])]; + tensor k_43_cast_fp16 = reshape(shape = var_4248, x = ko_21_cast_fp16)[name = tensor("k_43_cast_fp16")]; + tensor _inversed_4273_y_0_to_fp16 = const()[name = tensor("_inversed_4273_y_0_to_fp16"), val = tensor(0x1p-9)]; + tensor _inversed_4273_cast_fp16 = mul(x = ts_65_cast_fp16, y = _inversed_4273_y_0_to_fp16)[name = tensor("_inversed_4273_cast_fp16")]; + tensor var_4274_cast_fp16 = floor(x = _inversed_4273_cast_fp16)[name = tensor("op_4274_cast_fp16")]; + tensor var_4275_to_fp16 = const()[name = tensor("op_4275_to_fp16"), val = tensor(0x1p+9)]; + tensor var_4276_cast_fp16 = mul(x = var_4274_cast_fp16, y = var_4275_to_fp16)[name = tensor("op_4276_cast_fp16")]; + tensor write_indices_float_43_cast_fp16 = sub(x = ts_65_cast_fp16, y = var_4276_cast_fp16)[name = tensor("write_indices_float_43_cast_fp16")]; + tensor var_4283_dtype_0 = const()[name = tensor("op_4283_dtype_0"), val = tensor("int32")]; + tensor write_indices_21_reps_0 = const()[name = tensor("write_indices_21_reps_0"), val = tensor([1, 1, 16, 64])]; + tensor write_indices_float_43_cast_fp16_to_int32 = cast(dtype = var_4283_dtype_0, x = write_indices_float_43_cast_fp16)[name = tensor("cast_502")]; + tensor write_indices_21 = tile(reps = write_indices_21_reps_0, x = write_indices_float_43_cast_fp16_to_int32)[name = tensor("write_indices_21")]; + tensor var_4293_begin_0 = const()[name = tensor("op_4293_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_4293_end_0 = const()[name = tensor("op_4293_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor var_4293_end_mask_0 = const()[name = tensor("op_4293_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_4293_squeeze_mask_0 = const()[name = tensor("op_4293_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor cache10_to_fp16_dtype_0 = const()[name = tensor("cache10_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor cache10_to_fp16 = cast(dtype = cache10_to_fp16_dtype_0, x = cache10)[name = tensor("cast_501")]; + tensor var_4293_cast_fp16 = slice_by_index(begin = var_4293_begin_0, end = var_4293_end_0, end_mask = var_4293_end_mask_0, squeeze_mask = var_4293_squeeze_mask_0, x = cache10_to_fp16)[name = tensor("op_4293_cast_fp16")]; + tensor var_4295_axis_0 = const()[name = tensor("op_4295_axis_0"), val = tensor(1)]; + tensor var_4295_mode_0 = const()[name = tensor("op_4295_mode_0"), val = tensor("update")]; + tensor var_4295_validate_indices_0 = const()[name = tensor("op_4295_validate_indices_0"), val = tensor(false)]; + tensor var_4295_cast_fp16 = scatter_along_axis(axis = var_4295_axis_0, data = var_4293_cast_fp16, indices = write_indices_21, mode = var_4295_mode_0, updates = k_43_cast_fp16, validate_indices = var_4295_validate_indices_0)[name = tensor("op_4295_cast_fp16")]; + tensor concat_72 = const()[name = tensor("concat_72"), val = tensor([0, 0, 0, 0, 0])]; + tensor concat_73 = const()[name = tensor("concat_73"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_21_internal_tensor_assign_1_stride_0 = const()[name = tensor("new_cache_21_internal_tensor_assign_1_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_21_internal_tensor_assign_1_begin_mask_0 = const()[name = tensor("new_cache_21_internal_tensor_assign_1_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_21_internal_tensor_assign_1_end_mask_0 = const()[name = tensor("new_cache_21_internal_tensor_assign_1_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_21_internal_tensor_assign_1_squeeze_mask_0 = const()[name = tensor("new_cache_21_internal_tensor_assign_1_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_68 = const()[name = tensor("shape_68"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_20 = const()[name = tensor("reduce_prod_20"), val = tensor(1048576)]; + tensor range_1d_20_start_0 = const()[name = tensor("range_1d_20_start_0"), val = tensor(0)]; + tensor range_1d_20_step_0 = const()[name = tensor("range_1d_20_step_0"), val = tensor(1)]; + tensor range_1d_20 = range_1d(end = reduce_prod_20, start = range_1d_20_start_0, step = range_1d_20_step_0)[name = tensor("range_1d_20")]; + tensor reshape_100 = reshape(shape = shape_68, x = range_1d_20)[name = tensor("reshape_100")]; + tensor slice_by_index_20 = slice_by_index(begin = concat_72, begin_mask = new_cache_21_internal_tensor_assign_1_begin_mask_0, end = concat_73, end_mask = new_cache_21_internal_tensor_assign_1_end_mask_0, squeeze_mask = new_cache_21_internal_tensor_assign_1_squeeze_mask_0, stride = new_cache_21_internal_tensor_assign_1_stride_0, x = reshape_100)[name = tensor("slice_by_index_20")]; + tensor reshape_101_shape_0 = const()[name = tensor("reshape_101_shape_0"), val = tensor([-1])]; + tensor reshape_101 = reshape(shape = reshape_101_shape_0, x = slice_by_index_20)[name = tensor("reshape_101")]; + tensor reshape_102_shape_0 = const()[name = tensor("reshape_102_shape_0"), val = tensor([-1])]; + tensor reshape_102_cast_fp16 = reshape(shape = reshape_102_shape_0, x = var_4295_cast_fp16)[name = tensor("reshape_102_cast_fp16")]; + tensor reshape_103_shape_0 = const()[name = tensor("reshape_103_shape_0"), val = tensor([-1])]; + tensor reshape_103_cast_fp16 = reshape(shape = reshape_103_shape_0, x = cache10_to_fp16)[name = tensor("reshape_103_cast_fp16")]; + tensor scatter_20_mode_0 = const()[name = tensor("scatter_20_mode_0"), val = tensor("update")]; + tensor scatter_20_axis_0 = const()[name = tensor("scatter_20_axis_0"), val = tensor(0)]; + tensor scatter_20_validate_indices_0 = const()[name = tensor("scatter_20_validate_indices_0"), val = tensor(false)]; + tensor scatter_20_cast_fp16 = scatter(axis = scatter_20_axis_0, data = reshape_103_cast_fp16, indices = reshape_101, mode = scatter_20_mode_0, updates = reshape_102_cast_fp16, validate_indices = scatter_20_validate_indices_0)[name = tensor("scatter_20_cast_fp16")]; + tensor reshape_104_cast_fp16 = reshape(shape = shape_68, x = scatter_20_cast_fp16)[name = tensor("reshape_104_cast_fp16")]; + tensor var_4303_begin_0 = const()[name = tensor("op_4303_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor var_4303_end_0 = const()[name = tensor("op_4303_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor var_4303_end_mask_0 = const()[name = tensor("op_4303_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_4303_squeeze_mask_0 = const()[name = tensor("op_4303_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor var_4303_cast_fp16 = slice_by_index(begin = var_4303_begin_0, end = var_4303_end_0, end_mask = var_4303_end_mask_0, squeeze_mask = var_4303_squeeze_mask_0, x = reshape_104_cast_fp16)[name = tensor("op_4303_cast_fp16")]; + tensor var_4305_axis_0 = const()[name = tensor("op_4305_axis_0"), val = tensor(1)]; + tensor var_4305_mode_0 = const()[name = tensor("op_4305_mode_0"), val = tensor("update")]; + tensor var_4305_validate_indices_0 = const()[name = tensor("op_4305_validate_indices_0"), val = tensor(false)]; + tensor var_4305_cast_fp16 = scatter_along_axis(axis = var_4305_axis_0, data = var_4303_cast_fp16, indices = write_indices_21, mode = var_4305_mode_0, updates = v_21_cast_fp16, validate_indices = var_4305_validate_indices_0)[name = tensor("op_4305_cast_fp16")]; + tensor concat_74 = const()[name = tensor("concat_74"), val = tensor([1, 0, 0, 0, 0])]; + tensor concat_75 = const()[name = tensor("concat_75"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_21_internal_tensor_assign_2_stride_0 = const()[name = tensor("new_cache_21_internal_tensor_assign_2_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_21_internal_tensor_assign_2_begin_mask_0 = const()[name = tensor("new_cache_21_internal_tensor_assign_2_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_21_internal_tensor_assign_2_end_mask_0 = const()[name = tensor("new_cache_21_internal_tensor_assign_2_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_21_internal_tensor_assign_2_squeeze_mask_0 = const()[name = tensor("new_cache_21_internal_tensor_assign_2_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_69 = const()[name = tensor("shape_69"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_21 = const()[name = tensor("reduce_prod_21"), val = tensor(1048576)]; + tensor range_1d_21_start_0 = const()[name = tensor("range_1d_21_start_0"), val = tensor(0)]; + tensor range_1d_21_step_0 = const()[name = tensor("range_1d_21_step_0"), val = tensor(1)]; + tensor range_1d_21 = range_1d(end = reduce_prod_21, start = range_1d_21_start_0, step = range_1d_21_step_0)[name = tensor("range_1d_21")]; + tensor reshape_105 = reshape(shape = shape_69, x = range_1d_21)[name = tensor("reshape_105")]; + tensor slice_by_index_21 = slice_by_index(begin = concat_74, begin_mask = new_cache_21_internal_tensor_assign_2_begin_mask_0, end = concat_75, end_mask = new_cache_21_internal_tensor_assign_2_end_mask_0, squeeze_mask = new_cache_21_internal_tensor_assign_2_squeeze_mask_0, stride = new_cache_21_internal_tensor_assign_2_stride_0, x = reshape_105)[name = tensor("slice_by_index_21")]; + tensor reshape_106_shape_0 = const()[name = tensor("reshape_106_shape_0"), val = tensor([-1])]; + tensor reshape_106 = reshape(shape = reshape_106_shape_0, x = slice_by_index_21)[name = tensor("reshape_106")]; + tensor reshape_107_shape_0 = const()[name = tensor("reshape_107_shape_0"), val = tensor([-1])]; + tensor reshape_107_cast_fp16 = reshape(shape = reshape_107_shape_0, x = var_4305_cast_fp16)[name = tensor("reshape_107_cast_fp16")]; + tensor reshape_108_shape_0 = const()[name = tensor("reshape_108_shape_0"), val = tensor([-1])]; + tensor reshape_108_cast_fp16 = reshape(shape = reshape_108_shape_0, x = reshape_104_cast_fp16)[name = tensor("reshape_108_cast_fp16")]; + tensor scatter_21_mode_0 = const()[name = tensor("scatter_21_mode_0"), val = tensor("update")]; + tensor scatter_21_axis_0 = const()[name = tensor("scatter_21_axis_0"), val = tensor(0)]; + tensor scatter_21_validate_indices_0 = const()[name = tensor("scatter_21_validate_indices_0"), val = tensor(false)]; + tensor scatter_21_cast_fp16 = scatter(axis = scatter_21_axis_0, data = reshape_108_cast_fp16, indices = reshape_106, mode = scatter_21_mode_0, updates = reshape_107_cast_fp16, validate_indices = scatter_21_validate_indices_0)[name = tensor("scatter_21_cast_fp16")]; + tensor reshape_109_cast_fp16 = reshape(shape = shape_69, x = scatter_21_cast_fp16)[name = tensor("reshape_109_cast_fp16")]; + tensor reshape_109_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("reshape_109_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor keys_61_begin_0 = const()[name = tensor("keys_61_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor keys_61_end_0 = const()[name = tensor("keys_61_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor keys_61_end_mask_0 = const()[name = tensor("keys_61_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor keys_61_squeeze_mask_0 = const()[name = tensor("keys_61_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor keys_61_cast_fp16 = slice_by_index(begin = keys_61_begin_0, end = keys_61_end_0, end_mask = keys_61_end_mask_0, squeeze_mask = keys_61_squeeze_mask_0, x = reshape_109_cast_fp16)[name = tensor("keys_61_cast_fp16")]; + tensor values_61_begin_0 = const()[name = tensor("values_61_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor values_61_end_0 = const()[name = tensor("values_61_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor values_61_end_mask_0 = const()[name = tensor("values_61_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor values_61_squeeze_mask_0 = const()[name = tensor("values_61_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor values_61_cast_fp16 = slice_by_index(begin = values_61_begin_0, end = values_61_end_0, end_mask = values_61_end_mask_0, squeeze_mask = values_61_squeeze_mask_0, x = reshape_109_cast_fp16)[name = tensor("values_61_cast_fp16")]; + tensor var_4317_cast_fp16 = not_equal(x = keys_61_cast_fp16, y = keys_61_cast_fp16)[name = tensor("op_4317_cast_fp16")]; + tensor keys_63_cast_fp16 = select(a = var_463_to_fp16, b = keys_61_cast_fp16, cond = var_4317_cast_fp16)[name = tensor("keys_63_cast_fp16")]; + tensor var_4325_cast_fp16 = not_equal(x = values_61_cast_fp16, y = values_61_cast_fp16)[name = tensor("op_4325_cast_fp16")]; + tensor values_63_cast_fp16 = select(a = var_463_to_fp16, b = values_61_cast_fp16, cond = var_4325_cast_fp16)[name = tensor("values_63_cast_fp16")]; + tensor var_4349 = const()[name = tensor("op_4349"), val = tensor([0, 2, 1, 3])]; + tensor var_4365 = const()[name = tensor("op_4365"), val = tensor([1, 1, 1])]; + tensor var_4366_cast_fp16 = reshape(shape = var_4365, x = position10_to_fp16)[name = tensor("op_4366_cast_fp16")]; + tensor var_4386_to_fp16 = const()[name = tensor("op_4386_to_fp16"), val = tensor(0x1p+0)]; + tensor valid_len_21_cast_fp16 = add(x = var_4366_cast_fp16, y = var_4386_to_fp16)[name = tensor("valid_len_21_cast_fp16")]; + tensor valid_mask_21_cast_fp16 = less(x = k_positions_1_promoted_to_fp16, y = valid_len_21_cast_fp16)[name = tensor("valid_mask_21_cast_fp16")]; + tensor causal_mask_21_cast_fp16 = less_equal(x = k_positions_1_promoted_to_fp16, y = var_4366_cast_fp16)[name = tensor("causal_mask_21_cast_fp16")]; + tensor attn_mask_41 = logical_and(x = valid_mask_21_cast_fp16, y = causal_mask_21_cast_fp16)[name = tensor("attn_mask_41")]; + tensor attn_mask_43_axes_0 = const()[name = tensor("attn_mask_43_axes_0"), val = tensor([1])]; + tensor attn_mask_43 = expand_dims(axes = attn_mask_43_axes_0, x = attn_mask_41)[name = tensor("attn_mask_43")]; + tensor var_4404_transpose_x_0 = const()[name = tensor("op_4404_transpose_x_0"), val = tensor(false)]; + tensor var_4404_transpose_y_0 = const()[name = tensor("op_4404_transpose_y_0"), val = tensor(false)]; + tensor transpose_92_perm_0 = const()[name = tensor("transpose_92_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_93_perm_0 = const()[name = tensor("transpose_93_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_93 = transpose(perm = transpose_93_perm_0, x = keys_63_cast_fp16)[name = tensor("transpose_173")]; + tensor transpose_92 = transpose(perm = transpose_92_perm_0, x = q_63_cast_fp16)[name = tensor("transpose_174")]; + tensor var_4404_cast_fp16 = matmul(transpose_x = var_4404_transpose_x_0, transpose_y = var_4404_transpose_y_0, x = transpose_92, y = transpose_93)[name = tensor("op_4404_cast_fp16")]; + tensor var_4398_to_fp16 = const()[name = tensor("op_4398_to_fp16"), val = tensor([0x1p-3])]; + tensor attn_weights_61_cast_fp16 = mul(x = var_4404_cast_fp16, y = var_4398_to_fp16)[name = tensor("attn_weights_61_cast_fp16")]; + tensor var_4406 = logical_not(x = attn_mask_43)[name = tensor("op_4406")]; + tensor var_4407_to_fp16 = const()[name = tensor("op_4407_to_fp16"), val = tensor(-inf)]; + tensor attn_weights_63_cast_fp16 = select(a = var_4407_to_fp16, b = attn_weights_61_cast_fp16, cond = var_4406)[name = tensor("attn_weights_63_cast_fp16")]; + tensor var_4409 = const()[name = tensor("op_4409"), val = tensor(-1)]; + tensor attn_weights_65_cast_fp16 = softmax(axis = var_4409, x = attn_weights_63_cast_fp16)[name = tensor("attn_weights_65_cast_fp16")]; + tensor attn_output_21_transpose_x_0 = const()[name = tensor("attn_output_21_transpose_x_0"), val = tensor(false)]; + tensor attn_output_21_transpose_y_0 = const()[name = tensor("attn_output_21_transpose_y_0"), val = tensor(false)]; + tensor values_65_cast_fp16 = transpose(perm = var_4349, x = values_63_cast_fp16)[name = tensor("transpose_175")]; + tensor attn_output_21_cast_fp16 = matmul(transpose_x = attn_output_21_transpose_x_0, transpose_y = attn_output_21_transpose_y_0, x = attn_weights_65_cast_fp16, y = values_65_cast_fp16)[name = tensor("attn_output_21_cast_fp16")]; + tensor var_4417 = const()[name = tensor("op_4417"), val = tensor([0, 2, 1, 3])]; + tensor var_4422 = const()[name = tensor("op_4422"), val = tensor([1, 1, 1024])]; + tensor var_4418_cast_fp16 = transpose(perm = var_4417, x = attn_output_21_cast_fp16)[name = tensor("transpose_172")]; + tensor input_127_cast_fp16 = reshape(shape = var_4422, x = var_4418_cast_fp16)[name = tensor("input_127_cast_fp16")]; + tensor attn10_out_proj_weight_to_fp16 = const()[name = tensor("attn10_out_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(259174400)))]; + tensor linear_42_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = attn10_out_proj_weight_to_fp16, x = input_127_cast_fp16)[name = tensor("linear_42_cast_fp16")]; + tensor var_4428_to_fp16 = const()[name = tensor("op_4428_to_fp16"), val = tensor(0x1p+0)]; + tensor var_4429_cast_fp16 = add(x = position10_to_fp16, y = var_4428_to_fp16)[name = tensor("op_4429_cast_fp16")]; + tensor var_4429_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_4429_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor input_129_cast_fp16 = add(x = input_123_cast_fp16, y = linear_42_cast_fp16)[name = tensor("input_129_cast_fp16")]; + tensor input_131_axes_0 = const()[name = tensor("input_131_axes_0"), val = tensor([-1])]; + tensor norm10_2_weight_to_fp16 = const()[name = tensor("norm10_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(261271616)))]; + tensor norm10_2_bias_to_fp16 = const()[name = tensor("norm10_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(261273728)))]; + tensor var_4433_to_fp16 = const()[name = tensor("op_4433_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_131_cast_fp16 = layer_norm(axes = input_131_axes_0, beta = norm10_2_bias_to_fp16, epsilon = var_4433_to_fp16, gamma = norm10_2_weight_to_fp16, x = input_129_cast_fp16)[name = tensor("input_131_cast_fp16")]; + tensor linear10_1_weight_to_fp16 = const()[name = tensor("linear10_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(261275840)))]; + tensor linear_43_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = linear10_1_weight_to_fp16, x = input_131_cast_fp16)[name = tensor("linear_43_cast_fp16")]; + tensor input_133_mode_0 = const()[name = tensor("input_133_mode_0"), val = tensor("EXACT")]; + tensor input_133_cast_fp16 = gelu(mode = input_133_mode_0, x = linear_43_cast_fp16)[name = tensor("input_133_cast_fp16")]; + tensor linear10_2_weight_to_fp16 = const()[name = tensor("linear10_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(269664512)))]; + tensor linear_44_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = linear10_2_weight_to_fp16, x = input_133_cast_fp16)[name = tensor("linear_44_cast_fp16")]; + tensor input_135_cast_fp16 = add(x = input_129_cast_fp16, y = linear_44_cast_fp16)[name = tensor("input_135_cast_fp16")]; + tensor input_137_axes_0 = const()[name = tensor("input_137_axes_0"), val = tensor([-1])]; + tensor norm11_1_weight_to_fp16 = const()[name = tensor("norm11_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(278053184)))]; + tensor norm11_1_bias_to_fp16 = const()[name = tensor("norm11_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(278055296)))]; + tensor var_4450_to_fp16 = const()[name = tensor("op_4450_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_137_cast_fp16 = layer_norm(axes = input_137_axes_0, beta = norm11_1_bias_to_fp16, epsilon = var_4450_to_fp16, gamma = norm11_1_weight_to_fp16, x = input_135_cast_fp16)[name = tensor("input_137_cast_fp16")]; + tensor attn11_in_proj_weight_to_fp16 = const()[name = tensor("attn11_in_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(278057408)))]; + tensor linear_45_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = attn11_in_proj_weight_to_fp16, x = input_137_cast_fp16)[name = tensor("linear_45_cast_fp16")]; + tensor var_4464 = const()[name = tensor("op_4464"), val = tensor([1, 1, 3, 16, 64])]; + tensor qkv_23_cast_fp16 = reshape(shape = var_4464, x = linear_45_cast_fp16)[name = tensor("qkv_23_cast_fp16")]; + tensor q_67_begin_0 = const()[name = tensor("q_67_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor q_67_end_0 = const()[name = tensor("q_67_end_0"), val = tensor([1, 1, 1, 16, 64])]; + tensor q_67_end_mask_0 = const()[name = tensor("q_67_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor q_67_squeeze_mask_0 = const()[name = tensor("q_67_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor q_67_cast_fp16 = slice_by_index(begin = q_67_begin_0, end = q_67_end_0, end_mask = q_67_end_mask_0, squeeze_mask = q_67_squeeze_mask_0, x = qkv_23_cast_fp16)[name = tensor("q_67_cast_fp16")]; + tensor k_45_begin_0 = const()[name = tensor("k_45_begin_0"), val = tensor([0, 0, 1, 0, 0])]; + tensor k_45_end_0 = const()[name = tensor("k_45_end_0"), val = tensor([1, 1, 2, 16, 64])]; + tensor k_45_end_mask_0 = const()[name = tensor("k_45_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor k_45_squeeze_mask_0 = const()[name = tensor("k_45_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor k_45_cast_fp16 = slice_by_index(begin = k_45_begin_0, end = k_45_end_0, end_mask = k_45_end_mask_0, squeeze_mask = k_45_squeeze_mask_0, x = qkv_23_cast_fp16)[name = tensor("k_45_cast_fp16")]; + tensor v_23_begin_0 = const()[name = tensor("v_23_begin_0"), val = tensor([0, 0, 2, 0, 0])]; + tensor v_23_end_0 = const()[name = tensor("v_23_end_0"), val = tensor([1, 1, 3, 16, 64])]; + tensor v_23_end_mask_0 = const()[name = tensor("v_23_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor v_23_squeeze_mask_0 = const()[name = tensor("v_23_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor v_23_cast_fp16 = slice_by_index(begin = v_23_begin_0, end = v_23_end_0, end_mask = v_23_end_mask_0, squeeze_mask = v_23_squeeze_mask_0, x = qkv_23_cast_fp16)[name = tensor("v_23_cast_fp16")]; + tensor var_4530 = const()[name = tensor("op_4530"), val = tensor([1, 1, 1, 1])]; + tensor position11_to_fp16_dtype_0 = const()[name = tensor("position11_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor position11_to_fp16 = cast(dtype = position11_to_fp16_dtype_0, x = position11)[name = tensor("cast_498")]; + tensor ts_71_cast_fp16 = reshape(shape = var_4530, x = position11_to_fp16)[name = tensor("ts_71_cast_fp16")]; + tensor var_4537 = const()[name = tensor("op_4537"), val = tensor([1, 1, 16, 32, 2])]; + tensor q_complex_23_cast_fp16 = reshape(shape = var_4537, x = q_67_cast_fp16)[name = tensor("q_complex_23_cast_fp16")]; + tensor var_4544 = const()[name = tensor("op_4544"), val = tensor([1, 1, 16, 32, 2])]; + tensor k_complex_23_cast_fp16 = reshape(shape = var_4544, x = k_45_cast_fp16)[name = tensor("k_complex_23_cast_fp16")]; + tensor var_4548_begin_0 = const()[name = tensor("op_4548_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_4548_end_0 = const()[name = tensor("op_4548_end_0"), val = tensor([1, 1, 16, 32, 1])]; + tensor var_4548_end_mask_0 = const()[name = tensor("op_4548_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_4548_squeeze_mask_0 = const()[name = tensor("op_4548_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_4548_cast_fp16 = slice_by_index(begin = var_4548_begin_0, end = var_4548_end_0, end_mask = var_4548_end_mask_0, squeeze_mask = var_4548_squeeze_mask_0, x = q_complex_23_cast_fp16)[name = tensor("op_4548_cast_fp16")]; + tensor var_4556_begin_0 = const()[name = tensor("op_4556_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_4556_end_0 = const()[name = tensor("op_4556_end_0"), val = tensor([1, 1, 16, 32, 2])]; + tensor var_4556_end_mask_0 = const()[name = tensor("op_4556_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_4556_squeeze_mask_0 = const()[name = tensor("op_4556_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_4556_cast_fp16 = slice_by_index(begin = var_4556_begin_0, end = var_4556_end_0, end_mask = var_4556_end_mask_0, squeeze_mask = var_4556_squeeze_mask_0, x = q_complex_23_cast_fp16)[name = tensor("op_4556_cast_fp16")]; + tensor var_4564_begin_0 = const()[name = tensor("op_4564_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_4564_end_0 = const()[name = tensor("op_4564_end_0"), val = tensor([1, 1, 16, 32, 1])]; + tensor var_4564_end_mask_0 = const()[name = tensor("op_4564_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_4564_squeeze_mask_0 = const()[name = tensor("op_4564_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_4564_cast_fp16 = slice_by_index(begin = var_4564_begin_0, end = var_4564_end_0, end_mask = var_4564_end_mask_0, squeeze_mask = var_4564_squeeze_mask_0, x = k_complex_23_cast_fp16)[name = tensor("op_4564_cast_fp16")]; + tensor var_4572_begin_0 = const()[name = tensor("op_4572_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_4572_end_0 = const()[name = tensor("op_4572_end_0"), val = tensor([1, 1, 16, 32, 2])]; + tensor var_4572_end_mask_0 = const()[name = tensor("op_4572_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_4572_squeeze_mask_0 = const()[name = tensor("op_4572_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_4572_cast_fp16 = slice_by_index(begin = var_4572_begin_0, end = var_4572_end_0, end_mask = var_4572_end_mask_0, squeeze_mask = var_4572_squeeze_mask_0, x = k_complex_23_cast_fp16)[name = tensor("op_4572_cast_fp16")]; + tensor freqs_23_to_fp16 = const()[name = tensor("freqs_23_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(284348928)))]; + tensor var_4578_cast_fp16 = mul(x = freqs_23_to_fp16, y = ts_71_cast_fp16)[name = tensor("op_4578_cast_fp16")]; + tensor rotr_23_cast_fp16 = cos(x = var_4578_cast_fp16)[name = tensor("rotr_23_cast_fp16")]; + tensor roti_23_cast_fp16 = sin(x = var_4578_cast_fp16)[name = tensor("roti_23_cast_fp16")]; + tensor var_4582_cast_fp16 = mul(x = var_4548_cast_fp16, y = rotr_23_cast_fp16)[name = tensor("op_4582_cast_fp16")]; + tensor var_4583_cast_fp16 = mul(x = var_4556_cast_fp16, y = roti_23_cast_fp16)[name = tensor("op_4583_cast_fp16")]; + tensor qor_45_cast_fp16 = sub(x = var_4582_cast_fp16, y = var_4583_cast_fp16)[name = tensor("qor_45_cast_fp16")]; + tensor var_4586_cast_fp16 = mul(x = var_4548_cast_fp16, y = roti_23_cast_fp16)[name = tensor("op_4586_cast_fp16")]; + tensor var_4587_cast_fp16 = mul(x = var_4556_cast_fp16, y = rotr_23_cast_fp16)[name = tensor("op_4587_cast_fp16")]; + tensor qoi_45_cast_fp16 = add(x = var_4586_cast_fp16, y = var_4587_cast_fp16)[name = tensor("qoi_45_cast_fp16")]; + tensor var_4590_cast_fp16 = mul(x = var_4564_cast_fp16, y = rotr_23_cast_fp16)[name = tensor("op_4590_cast_fp16")]; + tensor var_4591_cast_fp16 = mul(x = var_4572_cast_fp16, y = roti_23_cast_fp16)[name = tensor("op_4591_cast_fp16")]; + tensor kor_45_cast_fp16 = sub(x = var_4590_cast_fp16, y = var_4591_cast_fp16)[name = tensor("kor_45_cast_fp16")]; + tensor var_4594_cast_fp16 = mul(x = var_4564_cast_fp16, y = roti_23_cast_fp16)[name = tensor("op_4594_cast_fp16")]; + tensor var_4595_cast_fp16 = mul(x = var_4572_cast_fp16, y = rotr_23_cast_fp16)[name = tensor("op_4595_cast_fp16")]; + tensor koi_45_cast_fp16 = add(x = var_4594_cast_fp16, y = var_4595_cast_fp16)[name = tensor("koi_45_cast_fp16")]; + tensor qo_23_axis_0 = const()[name = tensor("qo_23_axis_0"), val = tensor(-1)]; + tensor qo_23_cast_fp16 = stack(axis = qo_23_axis_0, values = (qor_45_cast_fp16, qoi_45_cast_fp16))[name = tensor("qo_23_cast_fp16")]; + tensor ko_23_axis_0 = const()[name = tensor("ko_23_axis_0"), val = tensor(-1)]; + tensor ko_23_cast_fp16 = stack(axis = ko_23_axis_0, values = (kor_45_cast_fp16, koi_45_cast_fp16))[name = tensor("ko_23_cast_fp16")]; + tensor var_4628 = const()[name = tensor("op_4628"), val = tensor([1, 1, 16, 64])]; + tensor q_69_cast_fp16 = reshape(shape = var_4628, x = qo_23_cast_fp16)[name = tensor("q_69_cast_fp16")]; + tensor var_4634 = const()[name = tensor("op_4634"), val = tensor([1, 1, 16, 64])]; + tensor k_47_cast_fp16 = reshape(shape = var_4634, x = ko_23_cast_fp16)[name = tensor("k_47_cast_fp16")]; + tensor _inversed_4659_y_0_to_fp16 = const()[name = tensor("_inversed_4659_y_0_to_fp16"), val = tensor(0x1p-9)]; + tensor _inversed_4659_cast_fp16 = mul(x = ts_71_cast_fp16, y = _inversed_4659_y_0_to_fp16)[name = tensor("_inversed_4659_cast_fp16")]; + tensor var_4660_cast_fp16 = floor(x = _inversed_4659_cast_fp16)[name = tensor("op_4660_cast_fp16")]; + tensor var_4661_to_fp16 = const()[name = tensor("op_4661_to_fp16"), val = tensor(0x1p+9)]; + tensor var_4662_cast_fp16 = mul(x = var_4660_cast_fp16, y = var_4661_to_fp16)[name = tensor("op_4662_cast_fp16")]; + tensor write_indices_float_47_cast_fp16 = sub(x = ts_71_cast_fp16, y = var_4662_cast_fp16)[name = tensor("write_indices_float_47_cast_fp16")]; + tensor var_4669_dtype_0 = const()[name = tensor("op_4669_dtype_0"), val = tensor("int32")]; + tensor write_indices_23_reps_0 = const()[name = tensor("write_indices_23_reps_0"), val = tensor([1, 1, 16, 64])]; + tensor write_indices_float_47_cast_fp16_to_int32 = cast(dtype = var_4669_dtype_0, x = write_indices_float_47_cast_fp16)[name = tensor("cast_497")]; + tensor write_indices_23 = tile(reps = write_indices_23_reps_0, x = write_indices_float_47_cast_fp16_to_int32)[name = tensor("write_indices_23")]; + tensor var_4679_begin_0 = const()[name = tensor("op_4679_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_4679_end_0 = const()[name = tensor("op_4679_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor var_4679_end_mask_0 = const()[name = tensor("op_4679_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_4679_squeeze_mask_0 = const()[name = tensor("op_4679_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor cache11_to_fp16_dtype_0 = const()[name = tensor("cache11_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor cache11_to_fp16 = cast(dtype = cache11_to_fp16_dtype_0, x = cache11)[name = tensor("cast_496")]; + tensor var_4679_cast_fp16 = slice_by_index(begin = var_4679_begin_0, end = var_4679_end_0, end_mask = var_4679_end_mask_0, squeeze_mask = var_4679_squeeze_mask_0, x = cache11_to_fp16)[name = tensor("op_4679_cast_fp16")]; + tensor var_4681_axis_0 = const()[name = tensor("op_4681_axis_0"), val = tensor(1)]; + tensor var_4681_mode_0 = const()[name = tensor("op_4681_mode_0"), val = tensor("update")]; + tensor var_4681_validate_indices_0 = const()[name = tensor("op_4681_validate_indices_0"), val = tensor(false)]; + tensor var_4681_cast_fp16 = scatter_along_axis(axis = var_4681_axis_0, data = var_4679_cast_fp16, indices = write_indices_23, mode = var_4681_mode_0, updates = k_47_cast_fp16, validate_indices = var_4681_validate_indices_0)[name = tensor("op_4681_cast_fp16")]; + tensor concat_79 = const()[name = tensor("concat_79"), val = tensor([0, 0, 0, 0, 0])]; + tensor concat_80 = const()[name = tensor("concat_80"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_23_internal_tensor_assign_1_stride_0 = const()[name = tensor("new_cache_23_internal_tensor_assign_1_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_23_internal_tensor_assign_1_begin_mask_0 = const()[name = tensor("new_cache_23_internal_tensor_assign_1_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_23_internal_tensor_assign_1_end_mask_0 = const()[name = tensor("new_cache_23_internal_tensor_assign_1_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_23_internal_tensor_assign_1_squeeze_mask_0 = const()[name = tensor("new_cache_23_internal_tensor_assign_1_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_70 = const()[name = tensor("shape_70"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_22 = const()[name = tensor("reduce_prod_22"), val = tensor(1048576)]; + tensor range_1d_22_start_0 = const()[name = tensor("range_1d_22_start_0"), val = tensor(0)]; + tensor range_1d_22_step_0 = const()[name = tensor("range_1d_22_step_0"), val = tensor(1)]; + tensor range_1d_22 = range_1d(end = reduce_prod_22, start = range_1d_22_start_0, step = range_1d_22_step_0)[name = tensor("range_1d_22")]; + tensor reshape_110 = reshape(shape = shape_70, x = range_1d_22)[name = tensor("reshape_110")]; + tensor slice_by_index_22 = slice_by_index(begin = concat_79, begin_mask = new_cache_23_internal_tensor_assign_1_begin_mask_0, end = concat_80, end_mask = new_cache_23_internal_tensor_assign_1_end_mask_0, squeeze_mask = new_cache_23_internal_tensor_assign_1_squeeze_mask_0, stride = new_cache_23_internal_tensor_assign_1_stride_0, x = reshape_110)[name = tensor("slice_by_index_22")]; + tensor reshape_111_shape_0 = const()[name = tensor("reshape_111_shape_0"), val = tensor([-1])]; + tensor reshape_111 = reshape(shape = reshape_111_shape_0, x = slice_by_index_22)[name = tensor("reshape_111")]; + tensor reshape_112_shape_0 = const()[name = tensor("reshape_112_shape_0"), val = tensor([-1])]; + tensor reshape_112_cast_fp16 = reshape(shape = reshape_112_shape_0, x = var_4681_cast_fp16)[name = tensor("reshape_112_cast_fp16")]; + tensor reshape_113_shape_0 = const()[name = tensor("reshape_113_shape_0"), val = tensor([-1])]; + tensor reshape_113_cast_fp16 = reshape(shape = reshape_113_shape_0, x = cache11_to_fp16)[name = tensor("reshape_113_cast_fp16")]; + tensor scatter_22_mode_0 = const()[name = tensor("scatter_22_mode_0"), val = tensor("update")]; + tensor scatter_22_axis_0 = const()[name = tensor("scatter_22_axis_0"), val = tensor(0)]; + tensor scatter_22_validate_indices_0 = const()[name = tensor("scatter_22_validate_indices_0"), val = tensor(false)]; + tensor scatter_22_cast_fp16 = scatter(axis = scatter_22_axis_0, data = reshape_113_cast_fp16, indices = reshape_111, mode = scatter_22_mode_0, updates = reshape_112_cast_fp16, validate_indices = scatter_22_validate_indices_0)[name = tensor("scatter_22_cast_fp16")]; + tensor reshape_114_cast_fp16 = reshape(shape = shape_70, x = scatter_22_cast_fp16)[name = tensor("reshape_114_cast_fp16")]; + tensor var_4689_begin_0 = const()[name = tensor("op_4689_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor var_4689_end_0 = const()[name = tensor("op_4689_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor var_4689_end_mask_0 = const()[name = tensor("op_4689_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_4689_squeeze_mask_0 = const()[name = tensor("op_4689_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor var_4689_cast_fp16 = slice_by_index(begin = var_4689_begin_0, end = var_4689_end_0, end_mask = var_4689_end_mask_0, squeeze_mask = var_4689_squeeze_mask_0, x = reshape_114_cast_fp16)[name = tensor("op_4689_cast_fp16")]; + tensor var_4691_axis_0 = const()[name = tensor("op_4691_axis_0"), val = tensor(1)]; + tensor var_4691_mode_0 = const()[name = tensor("op_4691_mode_0"), val = tensor("update")]; + tensor var_4691_validate_indices_0 = const()[name = tensor("op_4691_validate_indices_0"), val = tensor(false)]; + tensor var_4691_cast_fp16 = scatter_along_axis(axis = var_4691_axis_0, data = var_4689_cast_fp16, indices = write_indices_23, mode = var_4691_mode_0, updates = v_23_cast_fp16, validate_indices = var_4691_validate_indices_0)[name = tensor("op_4691_cast_fp16")]; + tensor concat_81 = const()[name = tensor("concat_81"), val = tensor([1, 0, 0, 0, 0])]; + tensor concat_82 = const()[name = tensor("concat_82"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_23_internal_tensor_assign_2_stride_0 = const()[name = tensor("new_cache_23_internal_tensor_assign_2_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_23_internal_tensor_assign_2_begin_mask_0 = const()[name = tensor("new_cache_23_internal_tensor_assign_2_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_23_internal_tensor_assign_2_end_mask_0 = const()[name = tensor("new_cache_23_internal_tensor_assign_2_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_23_internal_tensor_assign_2_squeeze_mask_0 = const()[name = tensor("new_cache_23_internal_tensor_assign_2_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_71 = const()[name = tensor("shape_71"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_23 = const()[name = tensor("reduce_prod_23"), val = tensor(1048576)]; + tensor range_1d_23_start_0 = const()[name = tensor("range_1d_23_start_0"), val = tensor(0)]; + tensor range_1d_23_step_0 = const()[name = tensor("range_1d_23_step_0"), val = tensor(1)]; + tensor range_1d_23 = range_1d(end = reduce_prod_23, start = range_1d_23_start_0, step = range_1d_23_step_0)[name = tensor("range_1d_23")]; + tensor reshape_115 = reshape(shape = shape_71, x = range_1d_23)[name = tensor("reshape_115")]; + tensor slice_by_index_23 = slice_by_index(begin = concat_81, begin_mask = new_cache_23_internal_tensor_assign_2_begin_mask_0, end = concat_82, end_mask = new_cache_23_internal_tensor_assign_2_end_mask_0, squeeze_mask = new_cache_23_internal_tensor_assign_2_squeeze_mask_0, stride = new_cache_23_internal_tensor_assign_2_stride_0, x = reshape_115)[name = tensor("slice_by_index_23")]; + tensor reshape_116_shape_0 = const()[name = tensor("reshape_116_shape_0"), val = tensor([-1])]; + tensor reshape_116 = reshape(shape = reshape_116_shape_0, x = slice_by_index_23)[name = tensor("reshape_116")]; + tensor reshape_117_shape_0 = const()[name = tensor("reshape_117_shape_0"), val = tensor([-1])]; + tensor reshape_117_cast_fp16 = reshape(shape = reshape_117_shape_0, x = var_4691_cast_fp16)[name = tensor("reshape_117_cast_fp16")]; + tensor reshape_118_shape_0 = const()[name = tensor("reshape_118_shape_0"), val = tensor([-1])]; + tensor reshape_118_cast_fp16 = reshape(shape = reshape_118_shape_0, x = reshape_114_cast_fp16)[name = tensor("reshape_118_cast_fp16")]; + tensor scatter_23_mode_0 = const()[name = tensor("scatter_23_mode_0"), val = tensor("update")]; + tensor scatter_23_axis_0 = const()[name = tensor("scatter_23_axis_0"), val = tensor(0)]; + tensor scatter_23_validate_indices_0 = const()[name = tensor("scatter_23_validate_indices_0"), val = tensor(false)]; + tensor scatter_23_cast_fp16 = scatter(axis = scatter_23_axis_0, data = reshape_118_cast_fp16, indices = reshape_116, mode = scatter_23_mode_0, updates = reshape_117_cast_fp16, validate_indices = scatter_23_validate_indices_0)[name = tensor("scatter_23_cast_fp16")]; + tensor reshape_119_cast_fp16 = reshape(shape = shape_71, x = scatter_23_cast_fp16)[name = tensor("reshape_119_cast_fp16")]; + tensor reshape_119_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("reshape_119_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor keys_67_begin_0 = const()[name = tensor("keys_67_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor keys_67_end_0 = const()[name = tensor("keys_67_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor keys_67_end_mask_0 = const()[name = tensor("keys_67_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor keys_67_squeeze_mask_0 = const()[name = tensor("keys_67_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor keys_67_cast_fp16 = slice_by_index(begin = keys_67_begin_0, end = keys_67_end_0, end_mask = keys_67_end_mask_0, squeeze_mask = keys_67_squeeze_mask_0, x = reshape_119_cast_fp16)[name = tensor("keys_67_cast_fp16")]; + tensor values_67_begin_0 = const()[name = tensor("values_67_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor values_67_end_0 = const()[name = tensor("values_67_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor values_67_end_mask_0 = const()[name = tensor("values_67_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor values_67_squeeze_mask_0 = const()[name = tensor("values_67_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor values_67_cast_fp16 = slice_by_index(begin = values_67_begin_0, end = values_67_end_0, end_mask = values_67_end_mask_0, squeeze_mask = values_67_squeeze_mask_0, x = reshape_119_cast_fp16)[name = tensor("values_67_cast_fp16")]; + tensor var_4703_cast_fp16 = not_equal(x = keys_67_cast_fp16, y = keys_67_cast_fp16)[name = tensor("op_4703_cast_fp16")]; + tensor keys_69_cast_fp16 = select(a = var_463_to_fp16, b = keys_67_cast_fp16, cond = var_4703_cast_fp16)[name = tensor("keys_69_cast_fp16")]; + tensor var_4711_cast_fp16 = not_equal(x = values_67_cast_fp16, y = values_67_cast_fp16)[name = tensor("op_4711_cast_fp16")]; + tensor values_69_cast_fp16 = select(a = var_463_to_fp16, b = values_67_cast_fp16, cond = var_4711_cast_fp16)[name = tensor("values_69_cast_fp16")]; + tensor var_4735 = const()[name = tensor("op_4735"), val = tensor([0, 2, 1, 3])]; + tensor var_4751 = const()[name = tensor("op_4751"), val = tensor([1, 1, 1])]; + tensor var_4752_cast_fp16 = reshape(shape = var_4751, x = position11_to_fp16)[name = tensor("op_4752_cast_fp16")]; + tensor var_4772_to_fp16 = const()[name = tensor("op_4772_to_fp16"), val = tensor(0x1p+0)]; + tensor valid_len_23_cast_fp16 = add(x = var_4752_cast_fp16, y = var_4772_to_fp16)[name = tensor("valid_len_23_cast_fp16")]; + tensor valid_mask_23_cast_fp16 = less(x = k_positions_1_promoted_to_fp16, y = valid_len_23_cast_fp16)[name = tensor("valid_mask_23_cast_fp16")]; + tensor causal_mask_23_cast_fp16 = less_equal(x = k_positions_1_promoted_to_fp16, y = var_4752_cast_fp16)[name = tensor("causal_mask_23_cast_fp16")]; + tensor attn_mask_45 = logical_and(x = valid_mask_23_cast_fp16, y = causal_mask_23_cast_fp16)[name = tensor("attn_mask_45")]; + tensor attn_mask_47_axes_0 = const()[name = tensor("attn_mask_47_axes_0"), val = tensor([1])]; + tensor attn_mask_47 = expand_dims(axes = attn_mask_47_axes_0, x = attn_mask_45)[name = tensor("attn_mask_47")]; + tensor var_4790_transpose_x_0 = const()[name = tensor("op_4790_transpose_x_0"), val = tensor(false)]; + tensor var_4790_transpose_y_0 = const()[name = tensor("op_4790_transpose_y_0"), val = tensor(false)]; + tensor transpose_94_perm_0 = const()[name = tensor("transpose_94_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_95_perm_0 = const()[name = tensor("transpose_95_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_95 = transpose(perm = transpose_95_perm_0, x = keys_69_cast_fp16)[name = tensor("transpose_169")]; + tensor transpose_94 = transpose(perm = transpose_94_perm_0, x = q_69_cast_fp16)[name = tensor("transpose_170")]; + tensor var_4790_cast_fp16 = matmul(transpose_x = var_4790_transpose_x_0, transpose_y = var_4790_transpose_y_0, x = transpose_94, y = transpose_95)[name = tensor("op_4790_cast_fp16")]; + tensor var_4784_to_fp16 = const()[name = tensor("op_4784_to_fp16"), val = tensor([0x1p-3])]; + tensor attn_weights_67_cast_fp16 = mul(x = var_4790_cast_fp16, y = var_4784_to_fp16)[name = tensor("attn_weights_67_cast_fp16")]; + tensor var_4792 = logical_not(x = attn_mask_47)[name = tensor("op_4792")]; + tensor var_4793_to_fp16 = const()[name = tensor("op_4793_to_fp16"), val = tensor(-inf)]; + tensor attn_weights_69_cast_fp16 = select(a = var_4793_to_fp16, b = attn_weights_67_cast_fp16, cond = var_4792)[name = tensor("attn_weights_69_cast_fp16")]; + tensor var_4795 = const()[name = tensor("op_4795"), val = tensor(-1)]; + tensor attn_weights_71_cast_fp16 = softmax(axis = var_4795, x = attn_weights_69_cast_fp16)[name = tensor("attn_weights_71_cast_fp16")]; + tensor attn_output_23_transpose_x_0 = const()[name = tensor("attn_output_23_transpose_x_0"), val = tensor(false)]; + tensor attn_output_23_transpose_y_0 = const()[name = tensor("attn_output_23_transpose_y_0"), val = tensor(false)]; + tensor values_71_cast_fp16 = transpose(perm = var_4735, x = values_69_cast_fp16)[name = tensor("transpose_171")]; + tensor attn_output_23_cast_fp16 = matmul(transpose_x = attn_output_23_transpose_x_0, transpose_y = attn_output_23_transpose_y_0, x = attn_weights_71_cast_fp16, y = values_71_cast_fp16)[name = tensor("attn_output_23_cast_fp16")]; + tensor var_4803 = const()[name = tensor("op_4803"), val = tensor([0, 2, 1, 3])]; + tensor var_4808 = const()[name = tensor("op_4808"), val = tensor([1, 1, 1024])]; + tensor var_4804_cast_fp16 = transpose(perm = var_4803, x = attn_output_23_cast_fp16)[name = tensor("transpose_168")]; + tensor input_139_cast_fp16 = reshape(shape = var_4808, x = var_4804_cast_fp16)[name = tensor("input_139_cast_fp16")]; + tensor attn11_out_proj_weight_to_fp16 = const()[name = tensor("attn11_out_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(284349056)))]; + tensor linear_46_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = attn11_out_proj_weight_to_fp16, x = input_139_cast_fp16)[name = tensor("linear_46_cast_fp16")]; + tensor var_4814_to_fp16 = const()[name = tensor("op_4814_to_fp16"), val = tensor(0x1p+0)]; + tensor var_4815_cast_fp16 = add(x = position11_to_fp16, y = var_4814_to_fp16)[name = tensor("op_4815_cast_fp16")]; + tensor var_4815_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_4815_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor input_141_cast_fp16 = add(x = input_135_cast_fp16, y = linear_46_cast_fp16)[name = tensor("input_141_cast_fp16")]; + tensor input_143_axes_0 = const()[name = tensor("input_143_axes_0"), val = tensor([-1])]; + tensor norm11_2_weight_to_fp16 = const()[name = tensor("norm11_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(286446272)))]; + tensor norm11_2_bias_to_fp16 = const()[name = tensor("norm11_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(286448384)))]; + tensor var_4819_to_fp16 = const()[name = tensor("op_4819_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_143_cast_fp16 = layer_norm(axes = input_143_axes_0, beta = norm11_2_bias_to_fp16, epsilon = var_4819_to_fp16, gamma = norm11_2_weight_to_fp16, x = input_141_cast_fp16)[name = tensor("input_143_cast_fp16")]; + tensor linear11_1_weight_to_fp16 = const()[name = tensor("linear11_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(286450496)))]; + tensor linear_47_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = linear11_1_weight_to_fp16, x = input_143_cast_fp16)[name = tensor("linear_47_cast_fp16")]; + tensor input_145_mode_0 = const()[name = tensor("input_145_mode_0"), val = tensor("EXACT")]; + tensor input_145_cast_fp16 = gelu(mode = input_145_mode_0, x = linear_47_cast_fp16)[name = tensor("input_145_cast_fp16")]; + tensor linear11_2_weight_to_fp16 = const()[name = tensor("linear11_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(294839168)))]; + tensor linear_48_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = linear11_2_weight_to_fp16, x = input_145_cast_fp16)[name = tensor("linear_48_cast_fp16")]; + tensor input_147_cast_fp16 = add(x = input_141_cast_fp16, y = linear_48_cast_fp16)[name = tensor("input_147_cast_fp16")]; + tensor input_149_axes_0 = const()[name = tensor("input_149_axes_0"), val = tensor([-1])]; + tensor norm12_1_weight_to_fp16 = const()[name = tensor("norm12_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(303227840)))]; + tensor norm12_1_bias_to_fp16 = const()[name = tensor("norm12_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(303229952)))]; + tensor var_4836_to_fp16 = const()[name = tensor("op_4836_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_149_cast_fp16 = layer_norm(axes = input_149_axes_0, beta = norm12_1_bias_to_fp16, epsilon = var_4836_to_fp16, gamma = norm12_1_weight_to_fp16, x = input_147_cast_fp16)[name = tensor("input_149_cast_fp16")]; + tensor attn12_in_proj_weight_to_fp16 = const()[name = tensor("attn12_in_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(303232064)))]; + tensor linear_49_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = attn12_in_proj_weight_to_fp16, x = input_149_cast_fp16)[name = tensor("linear_49_cast_fp16")]; + tensor var_4850 = const()[name = tensor("op_4850"), val = tensor([1, 1, 3, 16, 64])]; + tensor qkv_25_cast_fp16 = reshape(shape = var_4850, x = linear_49_cast_fp16)[name = tensor("qkv_25_cast_fp16")]; + tensor q_73_begin_0 = const()[name = tensor("q_73_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor q_73_end_0 = const()[name = tensor("q_73_end_0"), val = tensor([1, 1, 1, 16, 64])]; + tensor q_73_end_mask_0 = const()[name = tensor("q_73_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor q_73_squeeze_mask_0 = const()[name = tensor("q_73_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor q_73_cast_fp16 = slice_by_index(begin = q_73_begin_0, end = q_73_end_0, end_mask = q_73_end_mask_0, squeeze_mask = q_73_squeeze_mask_0, x = qkv_25_cast_fp16)[name = tensor("q_73_cast_fp16")]; + tensor k_49_begin_0 = const()[name = tensor("k_49_begin_0"), val = tensor([0, 0, 1, 0, 0])]; + tensor k_49_end_0 = const()[name = tensor("k_49_end_0"), val = tensor([1, 1, 2, 16, 64])]; + tensor k_49_end_mask_0 = const()[name = tensor("k_49_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor k_49_squeeze_mask_0 = const()[name = tensor("k_49_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor k_49_cast_fp16 = slice_by_index(begin = k_49_begin_0, end = k_49_end_0, end_mask = k_49_end_mask_0, squeeze_mask = k_49_squeeze_mask_0, x = qkv_25_cast_fp16)[name = tensor("k_49_cast_fp16")]; + tensor v_25_begin_0 = const()[name = tensor("v_25_begin_0"), val = tensor([0, 0, 2, 0, 0])]; + tensor v_25_end_0 = const()[name = tensor("v_25_end_0"), val = tensor([1, 1, 3, 16, 64])]; + tensor v_25_end_mask_0 = const()[name = tensor("v_25_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor v_25_squeeze_mask_0 = const()[name = tensor("v_25_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor v_25_cast_fp16 = slice_by_index(begin = v_25_begin_0, end = v_25_end_0, end_mask = v_25_end_mask_0, squeeze_mask = v_25_squeeze_mask_0, x = qkv_25_cast_fp16)[name = tensor("v_25_cast_fp16")]; + tensor var_4916 = const()[name = tensor("op_4916"), val = tensor([1, 1, 1, 1])]; + tensor position12_to_fp16_dtype_0 = const()[name = tensor("position12_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor position12_to_fp16 = cast(dtype = position12_to_fp16_dtype_0, x = position12)[name = tensor("cast_493")]; + tensor ts_77_cast_fp16 = reshape(shape = var_4916, x = position12_to_fp16)[name = tensor("ts_77_cast_fp16")]; + tensor var_4923 = const()[name = tensor("op_4923"), val = tensor([1, 1, 16, 32, 2])]; + tensor q_complex_25_cast_fp16 = reshape(shape = var_4923, x = q_73_cast_fp16)[name = tensor("q_complex_25_cast_fp16")]; + tensor var_4930 = const()[name = tensor("op_4930"), val = tensor([1, 1, 16, 32, 2])]; + tensor k_complex_25_cast_fp16 = reshape(shape = var_4930, x = k_49_cast_fp16)[name = tensor("k_complex_25_cast_fp16")]; + tensor var_4934_begin_0 = const()[name = tensor("op_4934_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_4934_end_0 = const()[name = tensor("op_4934_end_0"), val = tensor([1, 1, 16, 32, 1])]; + tensor var_4934_end_mask_0 = const()[name = tensor("op_4934_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_4934_squeeze_mask_0 = const()[name = tensor("op_4934_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_4934_cast_fp16 = slice_by_index(begin = var_4934_begin_0, end = var_4934_end_0, end_mask = var_4934_end_mask_0, squeeze_mask = var_4934_squeeze_mask_0, x = q_complex_25_cast_fp16)[name = tensor("op_4934_cast_fp16")]; + tensor var_4942_begin_0 = const()[name = tensor("op_4942_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_4942_end_0 = const()[name = tensor("op_4942_end_0"), val = tensor([1, 1, 16, 32, 2])]; + tensor var_4942_end_mask_0 = const()[name = tensor("op_4942_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_4942_squeeze_mask_0 = const()[name = tensor("op_4942_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_4942_cast_fp16 = slice_by_index(begin = var_4942_begin_0, end = var_4942_end_0, end_mask = var_4942_end_mask_0, squeeze_mask = var_4942_squeeze_mask_0, x = q_complex_25_cast_fp16)[name = tensor("op_4942_cast_fp16")]; + tensor var_4950_begin_0 = const()[name = tensor("op_4950_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_4950_end_0 = const()[name = tensor("op_4950_end_0"), val = tensor([1, 1, 16, 32, 1])]; + tensor var_4950_end_mask_0 = const()[name = tensor("op_4950_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_4950_squeeze_mask_0 = const()[name = tensor("op_4950_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_4950_cast_fp16 = slice_by_index(begin = var_4950_begin_0, end = var_4950_end_0, end_mask = var_4950_end_mask_0, squeeze_mask = var_4950_squeeze_mask_0, x = k_complex_25_cast_fp16)[name = tensor("op_4950_cast_fp16")]; + tensor var_4958_begin_0 = const()[name = tensor("op_4958_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_4958_end_0 = const()[name = tensor("op_4958_end_0"), val = tensor([1, 1, 16, 32, 2])]; + tensor var_4958_end_mask_0 = const()[name = tensor("op_4958_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_4958_squeeze_mask_0 = const()[name = tensor("op_4958_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_4958_cast_fp16 = slice_by_index(begin = var_4958_begin_0, end = var_4958_end_0, end_mask = var_4958_end_mask_0, squeeze_mask = var_4958_squeeze_mask_0, x = k_complex_25_cast_fp16)[name = tensor("op_4958_cast_fp16")]; + tensor freqs_25_to_fp16 = const()[name = tensor("freqs_25_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(309523584)))]; + tensor var_4964_cast_fp16 = mul(x = freqs_25_to_fp16, y = ts_77_cast_fp16)[name = tensor("op_4964_cast_fp16")]; + tensor rotr_25_cast_fp16 = cos(x = var_4964_cast_fp16)[name = tensor("rotr_25_cast_fp16")]; + tensor roti_25_cast_fp16 = sin(x = var_4964_cast_fp16)[name = tensor("roti_25_cast_fp16")]; + tensor var_4968_cast_fp16 = mul(x = var_4934_cast_fp16, y = rotr_25_cast_fp16)[name = tensor("op_4968_cast_fp16")]; + tensor var_4969_cast_fp16 = mul(x = var_4942_cast_fp16, y = roti_25_cast_fp16)[name = tensor("op_4969_cast_fp16")]; + tensor qor_49_cast_fp16 = sub(x = var_4968_cast_fp16, y = var_4969_cast_fp16)[name = tensor("qor_49_cast_fp16")]; + tensor var_4972_cast_fp16 = mul(x = var_4934_cast_fp16, y = roti_25_cast_fp16)[name = tensor("op_4972_cast_fp16")]; + tensor var_4973_cast_fp16 = mul(x = var_4942_cast_fp16, y = rotr_25_cast_fp16)[name = tensor("op_4973_cast_fp16")]; + tensor qoi_49_cast_fp16 = add(x = var_4972_cast_fp16, y = var_4973_cast_fp16)[name = tensor("qoi_49_cast_fp16")]; + tensor var_4976_cast_fp16 = mul(x = var_4950_cast_fp16, y = rotr_25_cast_fp16)[name = tensor("op_4976_cast_fp16")]; + tensor var_4977_cast_fp16 = mul(x = var_4958_cast_fp16, y = roti_25_cast_fp16)[name = tensor("op_4977_cast_fp16")]; + tensor kor_49_cast_fp16 = sub(x = var_4976_cast_fp16, y = var_4977_cast_fp16)[name = tensor("kor_49_cast_fp16")]; + tensor var_4980_cast_fp16 = mul(x = var_4950_cast_fp16, y = roti_25_cast_fp16)[name = tensor("op_4980_cast_fp16")]; + tensor var_4981_cast_fp16 = mul(x = var_4958_cast_fp16, y = rotr_25_cast_fp16)[name = tensor("op_4981_cast_fp16")]; + tensor koi_49_cast_fp16 = add(x = var_4980_cast_fp16, y = var_4981_cast_fp16)[name = tensor("koi_49_cast_fp16")]; + tensor qo_25_axis_0 = const()[name = tensor("qo_25_axis_0"), val = tensor(-1)]; + tensor qo_25_cast_fp16 = stack(axis = qo_25_axis_0, values = (qor_49_cast_fp16, qoi_49_cast_fp16))[name = tensor("qo_25_cast_fp16")]; + tensor ko_25_axis_0 = const()[name = tensor("ko_25_axis_0"), val = tensor(-1)]; + tensor ko_25_cast_fp16 = stack(axis = ko_25_axis_0, values = (kor_49_cast_fp16, koi_49_cast_fp16))[name = tensor("ko_25_cast_fp16")]; + tensor var_5014 = const()[name = tensor("op_5014"), val = tensor([1, 1, 16, 64])]; + tensor q_75_cast_fp16 = reshape(shape = var_5014, x = qo_25_cast_fp16)[name = tensor("q_75_cast_fp16")]; + tensor var_5020 = const()[name = tensor("op_5020"), val = tensor([1, 1, 16, 64])]; + tensor k_51_cast_fp16 = reshape(shape = var_5020, x = ko_25_cast_fp16)[name = tensor("k_51_cast_fp16")]; + tensor _inversed_5045_y_0_to_fp16 = const()[name = tensor("_inversed_5045_y_0_to_fp16"), val = tensor(0x1p-9)]; + tensor _inversed_5045_cast_fp16 = mul(x = ts_77_cast_fp16, y = _inversed_5045_y_0_to_fp16)[name = tensor("_inversed_5045_cast_fp16")]; + tensor var_5046_cast_fp16 = floor(x = _inversed_5045_cast_fp16)[name = tensor("op_5046_cast_fp16")]; + tensor var_5047_to_fp16 = const()[name = tensor("op_5047_to_fp16"), val = tensor(0x1p+9)]; + tensor var_5048_cast_fp16 = mul(x = var_5046_cast_fp16, y = var_5047_to_fp16)[name = tensor("op_5048_cast_fp16")]; + tensor write_indices_float_51_cast_fp16 = sub(x = ts_77_cast_fp16, y = var_5048_cast_fp16)[name = tensor("write_indices_float_51_cast_fp16")]; + tensor var_5055_dtype_0 = const()[name = tensor("op_5055_dtype_0"), val = tensor("int32")]; + tensor write_indices_25_reps_0 = const()[name = tensor("write_indices_25_reps_0"), val = tensor([1, 1, 16, 64])]; + tensor write_indices_float_51_cast_fp16_to_int32 = cast(dtype = var_5055_dtype_0, x = write_indices_float_51_cast_fp16)[name = tensor("cast_492")]; + tensor write_indices_25 = tile(reps = write_indices_25_reps_0, x = write_indices_float_51_cast_fp16_to_int32)[name = tensor("write_indices_25")]; + tensor var_5065_begin_0 = const()[name = tensor("op_5065_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_5065_end_0 = const()[name = tensor("op_5065_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor var_5065_end_mask_0 = const()[name = tensor("op_5065_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_5065_squeeze_mask_0 = const()[name = tensor("op_5065_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor cache12_to_fp16_dtype_0 = const()[name = tensor("cache12_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor cache12_to_fp16 = cast(dtype = cache12_to_fp16_dtype_0, x = cache12)[name = tensor("cast_491")]; + tensor var_5065_cast_fp16 = slice_by_index(begin = var_5065_begin_0, end = var_5065_end_0, end_mask = var_5065_end_mask_0, squeeze_mask = var_5065_squeeze_mask_0, x = cache12_to_fp16)[name = tensor("op_5065_cast_fp16")]; + tensor var_5067_axis_0 = const()[name = tensor("op_5067_axis_0"), val = tensor(1)]; + tensor var_5067_mode_0 = const()[name = tensor("op_5067_mode_0"), val = tensor("update")]; + tensor var_5067_validate_indices_0 = const()[name = tensor("op_5067_validate_indices_0"), val = tensor(false)]; + tensor var_5067_cast_fp16 = scatter_along_axis(axis = var_5067_axis_0, data = var_5065_cast_fp16, indices = write_indices_25, mode = var_5067_mode_0, updates = k_51_cast_fp16, validate_indices = var_5067_validate_indices_0)[name = tensor("op_5067_cast_fp16")]; + tensor concat_86 = const()[name = tensor("concat_86"), val = tensor([0, 0, 0, 0, 0])]; + tensor concat_87 = const()[name = tensor("concat_87"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_25_internal_tensor_assign_1_stride_0 = const()[name = tensor("new_cache_25_internal_tensor_assign_1_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_25_internal_tensor_assign_1_begin_mask_0 = const()[name = tensor("new_cache_25_internal_tensor_assign_1_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_25_internal_tensor_assign_1_end_mask_0 = const()[name = tensor("new_cache_25_internal_tensor_assign_1_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_25_internal_tensor_assign_1_squeeze_mask_0 = const()[name = tensor("new_cache_25_internal_tensor_assign_1_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_72 = const()[name = tensor("shape_72"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_24 = const()[name = tensor("reduce_prod_24"), val = tensor(1048576)]; + tensor range_1d_24_start_0 = const()[name = tensor("range_1d_24_start_0"), val = tensor(0)]; + tensor range_1d_24_step_0 = const()[name = tensor("range_1d_24_step_0"), val = tensor(1)]; + tensor range_1d_24 = range_1d(end = reduce_prod_24, start = range_1d_24_start_0, step = range_1d_24_step_0)[name = tensor("range_1d_24")]; + tensor reshape_120 = reshape(shape = shape_72, x = range_1d_24)[name = tensor("reshape_120")]; + tensor slice_by_index_24 = slice_by_index(begin = concat_86, begin_mask = new_cache_25_internal_tensor_assign_1_begin_mask_0, end = concat_87, end_mask = new_cache_25_internal_tensor_assign_1_end_mask_0, squeeze_mask = new_cache_25_internal_tensor_assign_1_squeeze_mask_0, stride = new_cache_25_internal_tensor_assign_1_stride_0, x = reshape_120)[name = tensor("slice_by_index_24")]; + tensor reshape_121_shape_0 = const()[name = tensor("reshape_121_shape_0"), val = tensor([-1])]; + tensor reshape_121 = reshape(shape = reshape_121_shape_0, x = slice_by_index_24)[name = tensor("reshape_121")]; + tensor reshape_122_shape_0 = const()[name = tensor("reshape_122_shape_0"), val = tensor([-1])]; + tensor reshape_122_cast_fp16 = reshape(shape = reshape_122_shape_0, x = var_5067_cast_fp16)[name = tensor("reshape_122_cast_fp16")]; + tensor reshape_123_shape_0 = const()[name = tensor("reshape_123_shape_0"), val = tensor([-1])]; + tensor reshape_123_cast_fp16 = reshape(shape = reshape_123_shape_0, x = cache12_to_fp16)[name = tensor("reshape_123_cast_fp16")]; + tensor scatter_24_mode_0 = const()[name = tensor("scatter_24_mode_0"), val = tensor("update")]; + tensor scatter_24_axis_0 = const()[name = tensor("scatter_24_axis_0"), val = tensor(0)]; + tensor scatter_24_validate_indices_0 = const()[name = tensor("scatter_24_validate_indices_0"), val = tensor(false)]; + tensor scatter_24_cast_fp16 = scatter(axis = scatter_24_axis_0, data = reshape_123_cast_fp16, indices = reshape_121, mode = scatter_24_mode_0, updates = reshape_122_cast_fp16, validate_indices = scatter_24_validate_indices_0)[name = tensor("scatter_24_cast_fp16")]; + tensor reshape_124_cast_fp16 = reshape(shape = shape_72, x = scatter_24_cast_fp16)[name = tensor("reshape_124_cast_fp16")]; + tensor var_5075_begin_0 = const()[name = tensor("op_5075_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor var_5075_end_0 = const()[name = tensor("op_5075_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor var_5075_end_mask_0 = const()[name = tensor("op_5075_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_5075_squeeze_mask_0 = const()[name = tensor("op_5075_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor var_5075_cast_fp16 = slice_by_index(begin = var_5075_begin_0, end = var_5075_end_0, end_mask = var_5075_end_mask_0, squeeze_mask = var_5075_squeeze_mask_0, x = reshape_124_cast_fp16)[name = tensor("op_5075_cast_fp16")]; + tensor var_5077_axis_0 = const()[name = tensor("op_5077_axis_0"), val = tensor(1)]; + tensor var_5077_mode_0 = const()[name = tensor("op_5077_mode_0"), val = tensor("update")]; + tensor var_5077_validate_indices_0 = const()[name = tensor("op_5077_validate_indices_0"), val = tensor(false)]; + tensor var_5077_cast_fp16 = scatter_along_axis(axis = var_5077_axis_0, data = var_5075_cast_fp16, indices = write_indices_25, mode = var_5077_mode_0, updates = v_25_cast_fp16, validate_indices = var_5077_validate_indices_0)[name = tensor("op_5077_cast_fp16")]; + tensor concat_88 = const()[name = tensor("concat_88"), val = tensor([1, 0, 0, 0, 0])]; + tensor concat_89 = const()[name = tensor("concat_89"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_25_internal_tensor_assign_2_stride_0 = const()[name = tensor("new_cache_25_internal_tensor_assign_2_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_25_internal_tensor_assign_2_begin_mask_0 = const()[name = tensor("new_cache_25_internal_tensor_assign_2_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_25_internal_tensor_assign_2_end_mask_0 = const()[name = tensor("new_cache_25_internal_tensor_assign_2_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_25_internal_tensor_assign_2_squeeze_mask_0 = const()[name = tensor("new_cache_25_internal_tensor_assign_2_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_73 = const()[name = tensor("shape_73"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_25 = const()[name = tensor("reduce_prod_25"), val = tensor(1048576)]; + tensor range_1d_25_start_0 = const()[name = tensor("range_1d_25_start_0"), val = tensor(0)]; + tensor range_1d_25_step_0 = const()[name = tensor("range_1d_25_step_0"), val = tensor(1)]; + tensor range_1d_25 = range_1d(end = reduce_prod_25, start = range_1d_25_start_0, step = range_1d_25_step_0)[name = tensor("range_1d_25")]; + tensor reshape_125 = reshape(shape = shape_73, x = range_1d_25)[name = tensor("reshape_125")]; + tensor slice_by_index_25 = slice_by_index(begin = concat_88, begin_mask = new_cache_25_internal_tensor_assign_2_begin_mask_0, end = concat_89, end_mask = new_cache_25_internal_tensor_assign_2_end_mask_0, squeeze_mask = new_cache_25_internal_tensor_assign_2_squeeze_mask_0, stride = new_cache_25_internal_tensor_assign_2_stride_0, x = reshape_125)[name = tensor("slice_by_index_25")]; + tensor reshape_126_shape_0 = const()[name = tensor("reshape_126_shape_0"), val = tensor([-1])]; + tensor reshape_126 = reshape(shape = reshape_126_shape_0, x = slice_by_index_25)[name = tensor("reshape_126")]; + tensor reshape_127_shape_0 = const()[name = tensor("reshape_127_shape_0"), val = tensor([-1])]; + tensor reshape_127_cast_fp16 = reshape(shape = reshape_127_shape_0, x = var_5077_cast_fp16)[name = tensor("reshape_127_cast_fp16")]; + tensor reshape_128_shape_0 = const()[name = tensor("reshape_128_shape_0"), val = tensor([-1])]; + tensor reshape_128_cast_fp16 = reshape(shape = reshape_128_shape_0, x = reshape_124_cast_fp16)[name = tensor("reshape_128_cast_fp16")]; + tensor scatter_25_mode_0 = const()[name = tensor("scatter_25_mode_0"), val = tensor("update")]; + tensor scatter_25_axis_0 = const()[name = tensor("scatter_25_axis_0"), val = tensor(0)]; + tensor scatter_25_validate_indices_0 = const()[name = tensor("scatter_25_validate_indices_0"), val = tensor(false)]; + tensor scatter_25_cast_fp16 = scatter(axis = scatter_25_axis_0, data = reshape_128_cast_fp16, indices = reshape_126, mode = scatter_25_mode_0, updates = reshape_127_cast_fp16, validate_indices = scatter_25_validate_indices_0)[name = tensor("scatter_25_cast_fp16")]; + tensor reshape_129_cast_fp16 = reshape(shape = shape_73, x = scatter_25_cast_fp16)[name = tensor("reshape_129_cast_fp16")]; + tensor reshape_129_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("reshape_129_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor keys_73_begin_0 = const()[name = tensor("keys_73_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor keys_73_end_0 = const()[name = tensor("keys_73_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor keys_73_end_mask_0 = const()[name = tensor("keys_73_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor keys_73_squeeze_mask_0 = const()[name = tensor("keys_73_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor keys_73_cast_fp16 = slice_by_index(begin = keys_73_begin_0, end = keys_73_end_0, end_mask = keys_73_end_mask_0, squeeze_mask = keys_73_squeeze_mask_0, x = reshape_129_cast_fp16)[name = tensor("keys_73_cast_fp16")]; + tensor values_73_begin_0 = const()[name = tensor("values_73_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor values_73_end_0 = const()[name = tensor("values_73_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor values_73_end_mask_0 = const()[name = tensor("values_73_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor values_73_squeeze_mask_0 = const()[name = tensor("values_73_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor values_73_cast_fp16 = slice_by_index(begin = values_73_begin_0, end = values_73_end_0, end_mask = values_73_end_mask_0, squeeze_mask = values_73_squeeze_mask_0, x = reshape_129_cast_fp16)[name = tensor("values_73_cast_fp16")]; + tensor var_5089_cast_fp16 = not_equal(x = keys_73_cast_fp16, y = keys_73_cast_fp16)[name = tensor("op_5089_cast_fp16")]; + tensor keys_75_cast_fp16 = select(a = var_463_to_fp16, b = keys_73_cast_fp16, cond = var_5089_cast_fp16)[name = tensor("keys_75_cast_fp16")]; + tensor var_5097_cast_fp16 = not_equal(x = values_73_cast_fp16, y = values_73_cast_fp16)[name = tensor("op_5097_cast_fp16")]; + tensor values_75_cast_fp16 = select(a = var_463_to_fp16, b = values_73_cast_fp16, cond = var_5097_cast_fp16)[name = tensor("values_75_cast_fp16")]; + tensor var_5121 = const()[name = tensor("op_5121"), val = tensor([0, 2, 1, 3])]; + tensor var_5137 = const()[name = tensor("op_5137"), val = tensor([1, 1, 1])]; + tensor var_5138_cast_fp16 = reshape(shape = var_5137, x = position12_to_fp16)[name = tensor("op_5138_cast_fp16")]; + tensor var_5158_to_fp16 = const()[name = tensor("op_5158_to_fp16"), val = tensor(0x1p+0)]; + tensor valid_len_25_cast_fp16 = add(x = var_5138_cast_fp16, y = var_5158_to_fp16)[name = tensor("valid_len_25_cast_fp16")]; + tensor valid_mask_25_cast_fp16 = less(x = k_positions_1_promoted_to_fp16, y = valid_len_25_cast_fp16)[name = tensor("valid_mask_25_cast_fp16")]; + tensor causal_mask_25_cast_fp16 = less_equal(x = k_positions_1_promoted_to_fp16, y = var_5138_cast_fp16)[name = tensor("causal_mask_25_cast_fp16")]; + tensor attn_mask_49 = logical_and(x = valid_mask_25_cast_fp16, y = causal_mask_25_cast_fp16)[name = tensor("attn_mask_49")]; + tensor attn_mask_51_axes_0 = const()[name = tensor("attn_mask_51_axes_0"), val = tensor([1])]; + tensor attn_mask_51 = expand_dims(axes = attn_mask_51_axes_0, x = attn_mask_49)[name = tensor("attn_mask_51")]; + tensor var_5176_transpose_x_0 = const()[name = tensor("op_5176_transpose_x_0"), val = tensor(false)]; + tensor var_5176_transpose_y_0 = const()[name = tensor("op_5176_transpose_y_0"), val = tensor(false)]; + tensor transpose_96_perm_0 = const()[name = tensor("transpose_96_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_97_perm_0 = const()[name = tensor("transpose_97_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_97 = transpose(perm = transpose_97_perm_0, x = keys_75_cast_fp16)[name = tensor("transpose_165")]; + tensor transpose_96 = transpose(perm = transpose_96_perm_0, x = q_75_cast_fp16)[name = tensor("transpose_166")]; + tensor var_5176_cast_fp16 = matmul(transpose_x = var_5176_transpose_x_0, transpose_y = var_5176_transpose_y_0, x = transpose_96, y = transpose_97)[name = tensor("op_5176_cast_fp16")]; + tensor var_5170_to_fp16 = const()[name = tensor("op_5170_to_fp16"), val = tensor([0x1p-3])]; + tensor attn_weights_73_cast_fp16 = mul(x = var_5176_cast_fp16, y = var_5170_to_fp16)[name = tensor("attn_weights_73_cast_fp16")]; + tensor var_5178 = logical_not(x = attn_mask_51)[name = tensor("op_5178")]; + tensor var_5179_to_fp16 = const()[name = tensor("op_5179_to_fp16"), val = tensor(-inf)]; + tensor attn_weights_75_cast_fp16 = select(a = var_5179_to_fp16, b = attn_weights_73_cast_fp16, cond = var_5178)[name = tensor("attn_weights_75_cast_fp16")]; + tensor var_5181 = const()[name = tensor("op_5181"), val = tensor(-1)]; + tensor attn_weights_77_cast_fp16 = softmax(axis = var_5181, x = attn_weights_75_cast_fp16)[name = tensor("attn_weights_77_cast_fp16")]; + tensor attn_output_25_transpose_x_0 = const()[name = tensor("attn_output_25_transpose_x_0"), val = tensor(false)]; + tensor attn_output_25_transpose_y_0 = const()[name = tensor("attn_output_25_transpose_y_0"), val = tensor(false)]; + tensor values_77_cast_fp16 = transpose(perm = var_5121, x = values_75_cast_fp16)[name = tensor("transpose_167")]; + tensor attn_output_25_cast_fp16 = matmul(transpose_x = attn_output_25_transpose_x_0, transpose_y = attn_output_25_transpose_y_0, x = attn_weights_77_cast_fp16, y = values_77_cast_fp16)[name = tensor("attn_output_25_cast_fp16")]; + tensor var_5189 = const()[name = tensor("op_5189"), val = tensor([0, 2, 1, 3])]; + tensor var_5194 = const()[name = tensor("op_5194"), val = tensor([1, 1, 1024])]; + tensor var_5190_cast_fp16 = transpose(perm = var_5189, x = attn_output_25_cast_fp16)[name = tensor("transpose_164")]; + tensor input_151_cast_fp16 = reshape(shape = var_5194, x = var_5190_cast_fp16)[name = tensor("input_151_cast_fp16")]; + tensor attn12_out_proj_weight_to_fp16 = const()[name = tensor("attn12_out_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(309523712)))]; + tensor linear_50_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = attn12_out_proj_weight_to_fp16, x = input_151_cast_fp16)[name = tensor("linear_50_cast_fp16")]; + tensor var_5200_to_fp16 = const()[name = tensor("op_5200_to_fp16"), val = tensor(0x1p+0)]; + tensor var_5201_cast_fp16 = add(x = position12_to_fp16, y = var_5200_to_fp16)[name = tensor("op_5201_cast_fp16")]; + tensor var_5201_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_5201_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor input_153_cast_fp16 = add(x = input_147_cast_fp16, y = linear_50_cast_fp16)[name = tensor("input_153_cast_fp16")]; + tensor input_155_axes_0 = const()[name = tensor("input_155_axes_0"), val = tensor([-1])]; + tensor norm12_2_weight_to_fp16 = const()[name = tensor("norm12_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(311620928)))]; + tensor norm12_2_bias_to_fp16 = const()[name = tensor("norm12_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(311623040)))]; + tensor var_5205_to_fp16 = const()[name = tensor("op_5205_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_155_cast_fp16 = layer_norm(axes = input_155_axes_0, beta = norm12_2_bias_to_fp16, epsilon = var_5205_to_fp16, gamma = norm12_2_weight_to_fp16, x = input_153_cast_fp16)[name = tensor("input_155_cast_fp16")]; + tensor linear12_1_weight_to_fp16 = const()[name = tensor("linear12_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(311625152)))]; + tensor linear_51_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = linear12_1_weight_to_fp16, x = input_155_cast_fp16)[name = tensor("linear_51_cast_fp16")]; + tensor input_157_mode_0 = const()[name = tensor("input_157_mode_0"), val = tensor("EXACT")]; + tensor input_157_cast_fp16 = gelu(mode = input_157_mode_0, x = linear_51_cast_fp16)[name = tensor("input_157_cast_fp16")]; + tensor linear12_2_weight_to_fp16 = const()[name = tensor("linear12_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(320013824)))]; + tensor linear_52_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = linear12_2_weight_to_fp16, x = input_157_cast_fp16)[name = tensor("linear_52_cast_fp16")]; + tensor input_159_cast_fp16 = add(x = input_153_cast_fp16, y = linear_52_cast_fp16)[name = tensor("input_159_cast_fp16")]; + tensor input_161_axes_0 = const()[name = tensor("input_161_axes_0"), val = tensor([-1])]; + tensor norm13_1_weight_to_fp16 = const()[name = tensor("norm13_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(328402496)))]; + tensor norm13_1_bias_to_fp16 = const()[name = tensor("norm13_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(328404608)))]; + tensor var_5222_to_fp16 = const()[name = tensor("op_5222_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_161_cast_fp16 = layer_norm(axes = input_161_axes_0, beta = norm13_1_bias_to_fp16, epsilon = var_5222_to_fp16, gamma = norm13_1_weight_to_fp16, x = input_159_cast_fp16)[name = tensor("input_161_cast_fp16")]; + tensor attn13_in_proj_weight_to_fp16 = const()[name = tensor("attn13_in_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(328406720)))]; + tensor linear_53_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = attn13_in_proj_weight_to_fp16, x = input_161_cast_fp16)[name = tensor("linear_53_cast_fp16")]; + tensor var_5236 = const()[name = tensor("op_5236"), val = tensor([1, 1, 3, 16, 64])]; + tensor qkv_27_cast_fp16 = reshape(shape = var_5236, x = linear_53_cast_fp16)[name = tensor("qkv_27_cast_fp16")]; + tensor q_79_begin_0 = const()[name = tensor("q_79_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor q_79_end_0 = const()[name = tensor("q_79_end_0"), val = tensor([1, 1, 1, 16, 64])]; + tensor q_79_end_mask_0 = const()[name = tensor("q_79_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor q_79_squeeze_mask_0 = const()[name = tensor("q_79_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor q_79_cast_fp16 = slice_by_index(begin = q_79_begin_0, end = q_79_end_0, end_mask = q_79_end_mask_0, squeeze_mask = q_79_squeeze_mask_0, x = qkv_27_cast_fp16)[name = tensor("q_79_cast_fp16")]; + tensor k_53_begin_0 = const()[name = tensor("k_53_begin_0"), val = tensor([0, 0, 1, 0, 0])]; + tensor k_53_end_0 = const()[name = tensor("k_53_end_0"), val = tensor([1, 1, 2, 16, 64])]; + tensor k_53_end_mask_0 = const()[name = tensor("k_53_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor k_53_squeeze_mask_0 = const()[name = tensor("k_53_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor k_53_cast_fp16 = slice_by_index(begin = k_53_begin_0, end = k_53_end_0, end_mask = k_53_end_mask_0, squeeze_mask = k_53_squeeze_mask_0, x = qkv_27_cast_fp16)[name = tensor("k_53_cast_fp16")]; + tensor v_27_begin_0 = const()[name = tensor("v_27_begin_0"), val = tensor([0, 0, 2, 0, 0])]; + tensor v_27_end_0 = const()[name = tensor("v_27_end_0"), val = tensor([1, 1, 3, 16, 64])]; + tensor v_27_end_mask_0 = const()[name = tensor("v_27_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor v_27_squeeze_mask_0 = const()[name = tensor("v_27_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor v_27_cast_fp16 = slice_by_index(begin = v_27_begin_0, end = v_27_end_0, end_mask = v_27_end_mask_0, squeeze_mask = v_27_squeeze_mask_0, x = qkv_27_cast_fp16)[name = tensor("v_27_cast_fp16")]; + tensor var_5302 = const()[name = tensor("op_5302"), val = tensor([1, 1, 1, 1])]; + tensor position13_to_fp16_dtype_0 = const()[name = tensor("position13_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor position13_to_fp16 = cast(dtype = position13_to_fp16_dtype_0, x = position13)[name = tensor("cast_488")]; + tensor ts_83_cast_fp16 = reshape(shape = var_5302, x = position13_to_fp16)[name = tensor("ts_83_cast_fp16")]; + tensor var_5309 = const()[name = tensor("op_5309"), val = tensor([1, 1, 16, 32, 2])]; + tensor q_complex_27_cast_fp16 = reshape(shape = var_5309, x = q_79_cast_fp16)[name = tensor("q_complex_27_cast_fp16")]; + tensor var_5316 = const()[name = tensor("op_5316"), val = tensor([1, 1, 16, 32, 2])]; + tensor k_complex_27_cast_fp16 = reshape(shape = var_5316, x = k_53_cast_fp16)[name = tensor("k_complex_27_cast_fp16")]; + tensor var_5320_begin_0 = const()[name = tensor("op_5320_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_5320_end_0 = const()[name = tensor("op_5320_end_0"), val = tensor([1, 1, 16, 32, 1])]; + tensor var_5320_end_mask_0 = const()[name = tensor("op_5320_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_5320_squeeze_mask_0 = const()[name = tensor("op_5320_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_5320_cast_fp16 = slice_by_index(begin = var_5320_begin_0, end = var_5320_end_0, end_mask = var_5320_end_mask_0, squeeze_mask = var_5320_squeeze_mask_0, x = q_complex_27_cast_fp16)[name = tensor("op_5320_cast_fp16")]; + tensor var_5328_begin_0 = const()[name = tensor("op_5328_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_5328_end_0 = const()[name = tensor("op_5328_end_0"), val = tensor([1, 1, 16, 32, 2])]; + tensor var_5328_end_mask_0 = const()[name = tensor("op_5328_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_5328_squeeze_mask_0 = const()[name = tensor("op_5328_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_5328_cast_fp16 = slice_by_index(begin = var_5328_begin_0, end = var_5328_end_0, end_mask = var_5328_end_mask_0, squeeze_mask = var_5328_squeeze_mask_0, x = q_complex_27_cast_fp16)[name = tensor("op_5328_cast_fp16")]; + tensor var_5336_begin_0 = const()[name = tensor("op_5336_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_5336_end_0 = const()[name = tensor("op_5336_end_0"), val = tensor([1, 1, 16, 32, 1])]; + tensor var_5336_end_mask_0 = const()[name = tensor("op_5336_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_5336_squeeze_mask_0 = const()[name = tensor("op_5336_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_5336_cast_fp16 = slice_by_index(begin = var_5336_begin_0, end = var_5336_end_0, end_mask = var_5336_end_mask_0, squeeze_mask = var_5336_squeeze_mask_0, x = k_complex_27_cast_fp16)[name = tensor("op_5336_cast_fp16")]; + tensor var_5344_begin_0 = const()[name = tensor("op_5344_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_5344_end_0 = const()[name = tensor("op_5344_end_0"), val = tensor([1, 1, 16, 32, 2])]; + tensor var_5344_end_mask_0 = const()[name = tensor("op_5344_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_5344_squeeze_mask_0 = const()[name = tensor("op_5344_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_5344_cast_fp16 = slice_by_index(begin = var_5344_begin_0, end = var_5344_end_0, end_mask = var_5344_end_mask_0, squeeze_mask = var_5344_squeeze_mask_0, x = k_complex_27_cast_fp16)[name = tensor("op_5344_cast_fp16")]; + tensor freqs_27_to_fp16 = const()[name = tensor("freqs_27_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(334698240)))]; + tensor var_5350_cast_fp16 = mul(x = freqs_27_to_fp16, y = ts_83_cast_fp16)[name = tensor("op_5350_cast_fp16")]; + tensor rotr_27_cast_fp16 = cos(x = var_5350_cast_fp16)[name = tensor("rotr_27_cast_fp16")]; + tensor roti_27_cast_fp16 = sin(x = var_5350_cast_fp16)[name = tensor("roti_27_cast_fp16")]; + tensor var_5354_cast_fp16 = mul(x = var_5320_cast_fp16, y = rotr_27_cast_fp16)[name = tensor("op_5354_cast_fp16")]; + tensor var_5355_cast_fp16 = mul(x = var_5328_cast_fp16, y = roti_27_cast_fp16)[name = tensor("op_5355_cast_fp16")]; + tensor qor_53_cast_fp16 = sub(x = var_5354_cast_fp16, y = var_5355_cast_fp16)[name = tensor("qor_53_cast_fp16")]; + tensor var_5358_cast_fp16 = mul(x = var_5320_cast_fp16, y = roti_27_cast_fp16)[name = tensor("op_5358_cast_fp16")]; + tensor var_5359_cast_fp16 = mul(x = var_5328_cast_fp16, y = rotr_27_cast_fp16)[name = tensor("op_5359_cast_fp16")]; + tensor qoi_53_cast_fp16 = add(x = var_5358_cast_fp16, y = var_5359_cast_fp16)[name = tensor("qoi_53_cast_fp16")]; + tensor var_5362_cast_fp16 = mul(x = var_5336_cast_fp16, y = rotr_27_cast_fp16)[name = tensor("op_5362_cast_fp16")]; + tensor var_5363_cast_fp16 = mul(x = var_5344_cast_fp16, y = roti_27_cast_fp16)[name = tensor("op_5363_cast_fp16")]; + tensor kor_53_cast_fp16 = sub(x = var_5362_cast_fp16, y = var_5363_cast_fp16)[name = tensor("kor_53_cast_fp16")]; + tensor var_5366_cast_fp16 = mul(x = var_5336_cast_fp16, y = roti_27_cast_fp16)[name = tensor("op_5366_cast_fp16")]; + tensor var_5367_cast_fp16 = mul(x = var_5344_cast_fp16, y = rotr_27_cast_fp16)[name = tensor("op_5367_cast_fp16")]; + tensor koi_53_cast_fp16 = add(x = var_5366_cast_fp16, y = var_5367_cast_fp16)[name = tensor("koi_53_cast_fp16")]; + tensor qo_27_axis_0 = const()[name = tensor("qo_27_axis_0"), val = tensor(-1)]; + tensor qo_27_cast_fp16 = stack(axis = qo_27_axis_0, values = (qor_53_cast_fp16, qoi_53_cast_fp16))[name = tensor("qo_27_cast_fp16")]; + tensor ko_27_axis_0 = const()[name = tensor("ko_27_axis_0"), val = tensor(-1)]; + tensor ko_27_cast_fp16 = stack(axis = ko_27_axis_0, values = (kor_53_cast_fp16, koi_53_cast_fp16))[name = tensor("ko_27_cast_fp16")]; + tensor var_5400 = const()[name = tensor("op_5400"), val = tensor([1, 1, 16, 64])]; + tensor q_81_cast_fp16 = reshape(shape = var_5400, x = qo_27_cast_fp16)[name = tensor("q_81_cast_fp16")]; + tensor var_5406 = const()[name = tensor("op_5406"), val = tensor([1, 1, 16, 64])]; + tensor k_55_cast_fp16 = reshape(shape = var_5406, x = ko_27_cast_fp16)[name = tensor("k_55_cast_fp16")]; + tensor _inversed_5431_y_0_to_fp16 = const()[name = tensor("_inversed_5431_y_0_to_fp16"), val = tensor(0x1p-9)]; + tensor _inversed_5431_cast_fp16 = mul(x = ts_83_cast_fp16, y = _inversed_5431_y_0_to_fp16)[name = tensor("_inversed_5431_cast_fp16")]; + tensor var_5432_cast_fp16 = floor(x = _inversed_5431_cast_fp16)[name = tensor("op_5432_cast_fp16")]; + tensor var_5433_to_fp16 = const()[name = tensor("op_5433_to_fp16"), val = tensor(0x1p+9)]; + tensor var_5434_cast_fp16 = mul(x = var_5432_cast_fp16, y = var_5433_to_fp16)[name = tensor("op_5434_cast_fp16")]; + tensor write_indices_float_55_cast_fp16 = sub(x = ts_83_cast_fp16, y = var_5434_cast_fp16)[name = tensor("write_indices_float_55_cast_fp16")]; + tensor var_5441_dtype_0 = const()[name = tensor("op_5441_dtype_0"), val = tensor("int32")]; + tensor write_indices_27_reps_0 = const()[name = tensor("write_indices_27_reps_0"), val = tensor([1, 1, 16, 64])]; + tensor write_indices_float_55_cast_fp16_to_int32 = cast(dtype = var_5441_dtype_0, x = write_indices_float_55_cast_fp16)[name = tensor("cast_487")]; + tensor write_indices_27 = tile(reps = write_indices_27_reps_0, x = write_indices_float_55_cast_fp16_to_int32)[name = tensor("write_indices_27")]; + tensor var_5451_begin_0 = const()[name = tensor("op_5451_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_5451_end_0 = const()[name = tensor("op_5451_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor var_5451_end_mask_0 = const()[name = tensor("op_5451_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_5451_squeeze_mask_0 = const()[name = tensor("op_5451_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor cache13_to_fp16_dtype_0 = const()[name = tensor("cache13_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor cache13_to_fp16 = cast(dtype = cache13_to_fp16_dtype_0, x = cache13)[name = tensor("cast_486")]; + tensor var_5451_cast_fp16 = slice_by_index(begin = var_5451_begin_0, end = var_5451_end_0, end_mask = var_5451_end_mask_0, squeeze_mask = var_5451_squeeze_mask_0, x = cache13_to_fp16)[name = tensor("op_5451_cast_fp16")]; + tensor var_5453_axis_0 = const()[name = tensor("op_5453_axis_0"), val = tensor(1)]; + tensor var_5453_mode_0 = const()[name = tensor("op_5453_mode_0"), val = tensor("update")]; + tensor var_5453_validate_indices_0 = const()[name = tensor("op_5453_validate_indices_0"), val = tensor(false)]; + tensor var_5453_cast_fp16 = scatter_along_axis(axis = var_5453_axis_0, data = var_5451_cast_fp16, indices = write_indices_27, mode = var_5453_mode_0, updates = k_55_cast_fp16, validate_indices = var_5453_validate_indices_0)[name = tensor("op_5453_cast_fp16")]; + tensor concat_93 = const()[name = tensor("concat_93"), val = tensor([0, 0, 0, 0, 0])]; + tensor concat_94 = const()[name = tensor("concat_94"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_27_internal_tensor_assign_1_stride_0 = const()[name = tensor("new_cache_27_internal_tensor_assign_1_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_27_internal_tensor_assign_1_begin_mask_0 = const()[name = tensor("new_cache_27_internal_tensor_assign_1_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_27_internal_tensor_assign_1_end_mask_0 = const()[name = tensor("new_cache_27_internal_tensor_assign_1_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_27_internal_tensor_assign_1_squeeze_mask_0 = const()[name = tensor("new_cache_27_internal_tensor_assign_1_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_74 = const()[name = tensor("shape_74"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_26 = const()[name = tensor("reduce_prod_26"), val = tensor(1048576)]; + tensor range_1d_26_start_0 = const()[name = tensor("range_1d_26_start_0"), val = tensor(0)]; + tensor range_1d_26_step_0 = const()[name = tensor("range_1d_26_step_0"), val = tensor(1)]; + tensor range_1d_26 = range_1d(end = reduce_prod_26, start = range_1d_26_start_0, step = range_1d_26_step_0)[name = tensor("range_1d_26")]; + tensor reshape_130 = reshape(shape = shape_74, x = range_1d_26)[name = tensor("reshape_130")]; + tensor slice_by_index_26 = slice_by_index(begin = concat_93, begin_mask = new_cache_27_internal_tensor_assign_1_begin_mask_0, end = concat_94, end_mask = new_cache_27_internal_tensor_assign_1_end_mask_0, squeeze_mask = new_cache_27_internal_tensor_assign_1_squeeze_mask_0, stride = new_cache_27_internal_tensor_assign_1_stride_0, x = reshape_130)[name = tensor("slice_by_index_26")]; + tensor reshape_131_shape_0 = const()[name = tensor("reshape_131_shape_0"), val = tensor([-1])]; + tensor reshape_131 = reshape(shape = reshape_131_shape_0, x = slice_by_index_26)[name = tensor("reshape_131")]; + tensor reshape_132_shape_0 = const()[name = tensor("reshape_132_shape_0"), val = tensor([-1])]; + tensor reshape_132_cast_fp16 = reshape(shape = reshape_132_shape_0, x = var_5453_cast_fp16)[name = tensor("reshape_132_cast_fp16")]; + tensor reshape_133_shape_0 = const()[name = tensor("reshape_133_shape_0"), val = tensor([-1])]; + tensor reshape_133_cast_fp16 = reshape(shape = reshape_133_shape_0, x = cache13_to_fp16)[name = tensor("reshape_133_cast_fp16")]; + tensor scatter_26_mode_0 = const()[name = tensor("scatter_26_mode_0"), val = tensor("update")]; + tensor scatter_26_axis_0 = const()[name = tensor("scatter_26_axis_0"), val = tensor(0)]; + tensor scatter_26_validate_indices_0 = const()[name = tensor("scatter_26_validate_indices_0"), val = tensor(false)]; + tensor scatter_26_cast_fp16 = scatter(axis = scatter_26_axis_0, data = reshape_133_cast_fp16, indices = reshape_131, mode = scatter_26_mode_0, updates = reshape_132_cast_fp16, validate_indices = scatter_26_validate_indices_0)[name = tensor("scatter_26_cast_fp16")]; + tensor reshape_134_cast_fp16 = reshape(shape = shape_74, x = scatter_26_cast_fp16)[name = tensor("reshape_134_cast_fp16")]; + tensor var_5461_begin_0 = const()[name = tensor("op_5461_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor var_5461_end_0 = const()[name = tensor("op_5461_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor var_5461_end_mask_0 = const()[name = tensor("op_5461_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_5461_squeeze_mask_0 = const()[name = tensor("op_5461_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor var_5461_cast_fp16 = slice_by_index(begin = var_5461_begin_0, end = var_5461_end_0, end_mask = var_5461_end_mask_0, squeeze_mask = var_5461_squeeze_mask_0, x = reshape_134_cast_fp16)[name = tensor("op_5461_cast_fp16")]; + tensor var_5463_axis_0 = const()[name = tensor("op_5463_axis_0"), val = tensor(1)]; + tensor var_5463_mode_0 = const()[name = tensor("op_5463_mode_0"), val = tensor("update")]; + tensor var_5463_validate_indices_0 = const()[name = tensor("op_5463_validate_indices_0"), val = tensor(false)]; + tensor var_5463_cast_fp16 = scatter_along_axis(axis = var_5463_axis_0, data = var_5461_cast_fp16, indices = write_indices_27, mode = var_5463_mode_0, updates = v_27_cast_fp16, validate_indices = var_5463_validate_indices_0)[name = tensor("op_5463_cast_fp16")]; + tensor concat_95 = const()[name = tensor("concat_95"), val = tensor([1, 0, 0, 0, 0])]; + tensor concat_96 = const()[name = tensor("concat_96"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_27_internal_tensor_assign_2_stride_0 = const()[name = tensor("new_cache_27_internal_tensor_assign_2_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_27_internal_tensor_assign_2_begin_mask_0 = const()[name = tensor("new_cache_27_internal_tensor_assign_2_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_27_internal_tensor_assign_2_end_mask_0 = const()[name = tensor("new_cache_27_internal_tensor_assign_2_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_27_internal_tensor_assign_2_squeeze_mask_0 = const()[name = tensor("new_cache_27_internal_tensor_assign_2_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_75 = const()[name = tensor("shape_75"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_27 = const()[name = tensor("reduce_prod_27"), val = tensor(1048576)]; + tensor range_1d_27_start_0 = const()[name = tensor("range_1d_27_start_0"), val = tensor(0)]; + tensor range_1d_27_step_0 = const()[name = tensor("range_1d_27_step_0"), val = tensor(1)]; + tensor range_1d_27 = range_1d(end = reduce_prod_27, start = range_1d_27_start_0, step = range_1d_27_step_0)[name = tensor("range_1d_27")]; + tensor reshape_135 = reshape(shape = shape_75, x = range_1d_27)[name = tensor("reshape_135")]; + tensor slice_by_index_27 = slice_by_index(begin = concat_95, begin_mask = new_cache_27_internal_tensor_assign_2_begin_mask_0, end = concat_96, end_mask = new_cache_27_internal_tensor_assign_2_end_mask_0, squeeze_mask = new_cache_27_internal_tensor_assign_2_squeeze_mask_0, stride = new_cache_27_internal_tensor_assign_2_stride_0, x = reshape_135)[name = tensor("slice_by_index_27")]; + tensor reshape_136_shape_0 = const()[name = tensor("reshape_136_shape_0"), val = tensor([-1])]; + tensor reshape_136 = reshape(shape = reshape_136_shape_0, x = slice_by_index_27)[name = tensor("reshape_136")]; + tensor reshape_137_shape_0 = const()[name = tensor("reshape_137_shape_0"), val = tensor([-1])]; + tensor reshape_137_cast_fp16 = reshape(shape = reshape_137_shape_0, x = var_5463_cast_fp16)[name = tensor("reshape_137_cast_fp16")]; + tensor reshape_138_shape_0 = const()[name = tensor("reshape_138_shape_0"), val = tensor([-1])]; + tensor reshape_138_cast_fp16 = reshape(shape = reshape_138_shape_0, x = reshape_134_cast_fp16)[name = tensor("reshape_138_cast_fp16")]; + tensor scatter_27_mode_0 = const()[name = tensor("scatter_27_mode_0"), val = tensor("update")]; + tensor scatter_27_axis_0 = const()[name = tensor("scatter_27_axis_0"), val = tensor(0)]; + tensor scatter_27_validate_indices_0 = const()[name = tensor("scatter_27_validate_indices_0"), val = tensor(false)]; + tensor scatter_27_cast_fp16 = scatter(axis = scatter_27_axis_0, data = reshape_138_cast_fp16, indices = reshape_136, mode = scatter_27_mode_0, updates = reshape_137_cast_fp16, validate_indices = scatter_27_validate_indices_0)[name = tensor("scatter_27_cast_fp16")]; + tensor reshape_139_cast_fp16 = reshape(shape = shape_75, x = scatter_27_cast_fp16)[name = tensor("reshape_139_cast_fp16")]; + tensor reshape_139_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("reshape_139_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor keys_79_begin_0 = const()[name = tensor("keys_79_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor keys_79_end_0 = const()[name = tensor("keys_79_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor keys_79_end_mask_0 = const()[name = tensor("keys_79_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor keys_79_squeeze_mask_0 = const()[name = tensor("keys_79_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor keys_79_cast_fp16 = slice_by_index(begin = keys_79_begin_0, end = keys_79_end_0, end_mask = keys_79_end_mask_0, squeeze_mask = keys_79_squeeze_mask_0, x = reshape_139_cast_fp16)[name = tensor("keys_79_cast_fp16")]; + tensor values_79_begin_0 = const()[name = tensor("values_79_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor values_79_end_0 = const()[name = tensor("values_79_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor values_79_end_mask_0 = const()[name = tensor("values_79_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor values_79_squeeze_mask_0 = const()[name = tensor("values_79_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor values_79_cast_fp16 = slice_by_index(begin = values_79_begin_0, end = values_79_end_0, end_mask = values_79_end_mask_0, squeeze_mask = values_79_squeeze_mask_0, x = reshape_139_cast_fp16)[name = tensor("values_79_cast_fp16")]; + tensor var_5475_cast_fp16 = not_equal(x = keys_79_cast_fp16, y = keys_79_cast_fp16)[name = tensor("op_5475_cast_fp16")]; + tensor keys_81_cast_fp16 = select(a = var_463_to_fp16, b = keys_79_cast_fp16, cond = var_5475_cast_fp16)[name = tensor("keys_81_cast_fp16")]; + tensor var_5483_cast_fp16 = not_equal(x = values_79_cast_fp16, y = values_79_cast_fp16)[name = tensor("op_5483_cast_fp16")]; + tensor values_81_cast_fp16 = select(a = var_463_to_fp16, b = values_79_cast_fp16, cond = var_5483_cast_fp16)[name = tensor("values_81_cast_fp16")]; + tensor var_5507 = const()[name = tensor("op_5507"), val = tensor([0, 2, 1, 3])]; + tensor var_5523 = const()[name = tensor("op_5523"), val = tensor([1, 1, 1])]; + tensor var_5524_cast_fp16 = reshape(shape = var_5523, x = position13_to_fp16)[name = tensor("op_5524_cast_fp16")]; + tensor var_5544_to_fp16 = const()[name = tensor("op_5544_to_fp16"), val = tensor(0x1p+0)]; + tensor valid_len_27_cast_fp16 = add(x = var_5524_cast_fp16, y = var_5544_to_fp16)[name = tensor("valid_len_27_cast_fp16")]; + tensor valid_mask_27_cast_fp16 = less(x = k_positions_1_promoted_to_fp16, y = valid_len_27_cast_fp16)[name = tensor("valid_mask_27_cast_fp16")]; + tensor causal_mask_27_cast_fp16 = less_equal(x = k_positions_1_promoted_to_fp16, y = var_5524_cast_fp16)[name = tensor("causal_mask_27_cast_fp16")]; + tensor attn_mask_53 = logical_and(x = valid_mask_27_cast_fp16, y = causal_mask_27_cast_fp16)[name = tensor("attn_mask_53")]; + tensor attn_mask_55_axes_0 = const()[name = tensor("attn_mask_55_axes_0"), val = tensor([1])]; + tensor attn_mask_55 = expand_dims(axes = attn_mask_55_axes_0, x = attn_mask_53)[name = tensor("attn_mask_55")]; + tensor var_5562_transpose_x_0 = const()[name = tensor("op_5562_transpose_x_0"), val = tensor(false)]; + tensor var_5562_transpose_y_0 = const()[name = tensor("op_5562_transpose_y_0"), val = tensor(false)]; + tensor transpose_98_perm_0 = const()[name = tensor("transpose_98_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_99_perm_0 = const()[name = tensor("transpose_99_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_99 = transpose(perm = transpose_99_perm_0, x = keys_81_cast_fp16)[name = tensor("transpose_161")]; + tensor transpose_98 = transpose(perm = transpose_98_perm_0, x = q_81_cast_fp16)[name = tensor("transpose_162")]; + tensor var_5562_cast_fp16 = matmul(transpose_x = var_5562_transpose_x_0, transpose_y = var_5562_transpose_y_0, x = transpose_98, y = transpose_99)[name = tensor("op_5562_cast_fp16")]; + tensor var_5556_to_fp16 = const()[name = tensor("op_5556_to_fp16"), val = tensor([0x1p-3])]; + tensor attn_weights_79_cast_fp16 = mul(x = var_5562_cast_fp16, y = var_5556_to_fp16)[name = tensor("attn_weights_79_cast_fp16")]; + tensor var_5564 = logical_not(x = attn_mask_55)[name = tensor("op_5564")]; + tensor var_5565_to_fp16 = const()[name = tensor("op_5565_to_fp16"), val = tensor(-inf)]; + tensor attn_weights_81_cast_fp16 = select(a = var_5565_to_fp16, b = attn_weights_79_cast_fp16, cond = var_5564)[name = tensor("attn_weights_81_cast_fp16")]; + tensor var_5567 = const()[name = tensor("op_5567"), val = tensor(-1)]; + tensor attn_weights_83_cast_fp16 = softmax(axis = var_5567, x = attn_weights_81_cast_fp16)[name = tensor("attn_weights_83_cast_fp16")]; + tensor attn_output_27_transpose_x_0 = const()[name = tensor("attn_output_27_transpose_x_0"), val = tensor(false)]; + tensor attn_output_27_transpose_y_0 = const()[name = tensor("attn_output_27_transpose_y_0"), val = tensor(false)]; + tensor values_83_cast_fp16 = transpose(perm = var_5507, x = values_81_cast_fp16)[name = tensor("transpose_163")]; + tensor attn_output_27_cast_fp16 = matmul(transpose_x = attn_output_27_transpose_x_0, transpose_y = attn_output_27_transpose_y_0, x = attn_weights_83_cast_fp16, y = values_83_cast_fp16)[name = tensor("attn_output_27_cast_fp16")]; + tensor var_5575 = const()[name = tensor("op_5575"), val = tensor([0, 2, 1, 3])]; + tensor var_5580 = const()[name = tensor("op_5580"), val = tensor([1, 1, 1024])]; + tensor var_5576_cast_fp16 = transpose(perm = var_5575, x = attn_output_27_cast_fp16)[name = tensor("transpose_160")]; + tensor input_163_cast_fp16 = reshape(shape = var_5580, x = var_5576_cast_fp16)[name = tensor("input_163_cast_fp16")]; + tensor attn13_out_proj_weight_to_fp16 = const()[name = tensor("attn13_out_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(334698368)))]; + tensor linear_54_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = attn13_out_proj_weight_to_fp16, x = input_163_cast_fp16)[name = tensor("linear_54_cast_fp16")]; + tensor var_5586_to_fp16 = const()[name = tensor("op_5586_to_fp16"), val = tensor(0x1p+0)]; + tensor var_5587_cast_fp16 = add(x = position13_to_fp16, y = var_5586_to_fp16)[name = tensor("op_5587_cast_fp16")]; + tensor var_5587_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_5587_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor input_165_cast_fp16 = add(x = input_159_cast_fp16, y = linear_54_cast_fp16)[name = tensor("input_165_cast_fp16")]; + tensor input_167_axes_0 = const()[name = tensor("input_167_axes_0"), val = tensor([-1])]; + tensor norm13_2_weight_to_fp16 = const()[name = tensor("norm13_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(336795584)))]; + tensor norm13_2_bias_to_fp16 = const()[name = tensor("norm13_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(336797696)))]; + tensor var_5591_to_fp16 = const()[name = tensor("op_5591_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_167_cast_fp16 = layer_norm(axes = input_167_axes_0, beta = norm13_2_bias_to_fp16, epsilon = var_5591_to_fp16, gamma = norm13_2_weight_to_fp16, x = input_165_cast_fp16)[name = tensor("input_167_cast_fp16")]; + tensor linear13_1_weight_to_fp16 = const()[name = tensor("linear13_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(336799808)))]; + tensor linear_55_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = linear13_1_weight_to_fp16, x = input_167_cast_fp16)[name = tensor("linear_55_cast_fp16")]; + tensor input_169_mode_0 = const()[name = tensor("input_169_mode_0"), val = tensor("EXACT")]; + tensor input_169_cast_fp16 = gelu(mode = input_169_mode_0, x = linear_55_cast_fp16)[name = tensor("input_169_cast_fp16")]; + tensor linear13_2_weight_to_fp16 = const()[name = tensor("linear13_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(345188480)))]; + tensor linear_56_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = linear13_2_weight_to_fp16, x = input_169_cast_fp16)[name = tensor("linear_56_cast_fp16")]; + tensor input_171_cast_fp16 = add(x = input_165_cast_fp16, y = linear_56_cast_fp16)[name = tensor("input_171_cast_fp16")]; + tensor input_173_axes_0 = const()[name = tensor("input_173_axes_0"), val = tensor([-1])]; + tensor norm14_1_weight_to_fp16 = const()[name = tensor("norm14_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(353577152)))]; + tensor norm14_1_bias_to_fp16 = const()[name = tensor("norm14_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(353579264)))]; + tensor var_5608_to_fp16 = const()[name = tensor("op_5608_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_173_cast_fp16 = layer_norm(axes = input_173_axes_0, beta = norm14_1_bias_to_fp16, epsilon = var_5608_to_fp16, gamma = norm14_1_weight_to_fp16, x = input_171_cast_fp16)[name = tensor("input_173_cast_fp16")]; + tensor attn14_in_proj_weight_to_fp16 = const()[name = tensor("attn14_in_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(353581376)))]; + tensor linear_57_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = attn14_in_proj_weight_to_fp16, x = input_173_cast_fp16)[name = tensor("linear_57_cast_fp16")]; + tensor var_5622 = const()[name = tensor("op_5622"), val = tensor([1, 1, 3, 16, 64])]; + tensor qkv_29_cast_fp16 = reshape(shape = var_5622, x = linear_57_cast_fp16)[name = tensor("qkv_29_cast_fp16")]; + tensor q_85_begin_0 = const()[name = tensor("q_85_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor q_85_end_0 = const()[name = tensor("q_85_end_0"), val = tensor([1, 1, 1, 16, 64])]; + tensor q_85_end_mask_0 = const()[name = tensor("q_85_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor q_85_squeeze_mask_0 = const()[name = tensor("q_85_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor q_85_cast_fp16 = slice_by_index(begin = q_85_begin_0, end = q_85_end_0, end_mask = q_85_end_mask_0, squeeze_mask = q_85_squeeze_mask_0, x = qkv_29_cast_fp16)[name = tensor("q_85_cast_fp16")]; + tensor k_57_begin_0 = const()[name = tensor("k_57_begin_0"), val = tensor([0, 0, 1, 0, 0])]; + tensor k_57_end_0 = const()[name = tensor("k_57_end_0"), val = tensor([1, 1, 2, 16, 64])]; + tensor k_57_end_mask_0 = const()[name = tensor("k_57_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor k_57_squeeze_mask_0 = const()[name = tensor("k_57_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor k_57_cast_fp16 = slice_by_index(begin = k_57_begin_0, end = k_57_end_0, end_mask = k_57_end_mask_0, squeeze_mask = k_57_squeeze_mask_0, x = qkv_29_cast_fp16)[name = tensor("k_57_cast_fp16")]; + tensor v_29_begin_0 = const()[name = tensor("v_29_begin_0"), val = tensor([0, 0, 2, 0, 0])]; + tensor v_29_end_0 = const()[name = tensor("v_29_end_0"), val = tensor([1, 1, 3, 16, 64])]; + tensor v_29_end_mask_0 = const()[name = tensor("v_29_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor v_29_squeeze_mask_0 = const()[name = tensor("v_29_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor v_29_cast_fp16 = slice_by_index(begin = v_29_begin_0, end = v_29_end_0, end_mask = v_29_end_mask_0, squeeze_mask = v_29_squeeze_mask_0, x = qkv_29_cast_fp16)[name = tensor("v_29_cast_fp16")]; + tensor var_5688 = const()[name = tensor("op_5688"), val = tensor([1, 1, 1, 1])]; + tensor position14_to_fp16_dtype_0 = const()[name = tensor("position14_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor position14_to_fp16 = cast(dtype = position14_to_fp16_dtype_0, x = position14)[name = tensor("cast_483")]; + tensor ts_89_cast_fp16 = reshape(shape = var_5688, x = position14_to_fp16)[name = tensor("ts_89_cast_fp16")]; + tensor var_5695 = const()[name = tensor("op_5695"), val = tensor([1, 1, 16, 32, 2])]; + tensor q_complex_29_cast_fp16 = reshape(shape = var_5695, x = q_85_cast_fp16)[name = tensor("q_complex_29_cast_fp16")]; + tensor var_5702 = const()[name = tensor("op_5702"), val = tensor([1, 1, 16, 32, 2])]; + tensor k_complex_29_cast_fp16 = reshape(shape = var_5702, x = k_57_cast_fp16)[name = tensor("k_complex_29_cast_fp16")]; + tensor var_5706_begin_0 = const()[name = tensor("op_5706_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_5706_end_0 = const()[name = tensor("op_5706_end_0"), val = tensor([1, 1, 16, 32, 1])]; + tensor var_5706_end_mask_0 = const()[name = tensor("op_5706_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_5706_squeeze_mask_0 = const()[name = tensor("op_5706_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_5706_cast_fp16 = slice_by_index(begin = var_5706_begin_0, end = var_5706_end_0, end_mask = var_5706_end_mask_0, squeeze_mask = var_5706_squeeze_mask_0, x = q_complex_29_cast_fp16)[name = tensor("op_5706_cast_fp16")]; + tensor var_5714_begin_0 = const()[name = tensor("op_5714_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_5714_end_0 = const()[name = tensor("op_5714_end_0"), val = tensor([1, 1, 16, 32, 2])]; + tensor var_5714_end_mask_0 = const()[name = tensor("op_5714_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_5714_squeeze_mask_0 = const()[name = tensor("op_5714_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_5714_cast_fp16 = slice_by_index(begin = var_5714_begin_0, end = var_5714_end_0, end_mask = var_5714_end_mask_0, squeeze_mask = var_5714_squeeze_mask_0, x = q_complex_29_cast_fp16)[name = tensor("op_5714_cast_fp16")]; + tensor var_5722_begin_0 = const()[name = tensor("op_5722_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_5722_end_0 = const()[name = tensor("op_5722_end_0"), val = tensor([1, 1, 16, 32, 1])]; + tensor var_5722_end_mask_0 = const()[name = tensor("op_5722_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_5722_squeeze_mask_0 = const()[name = tensor("op_5722_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_5722_cast_fp16 = slice_by_index(begin = var_5722_begin_0, end = var_5722_end_0, end_mask = var_5722_end_mask_0, squeeze_mask = var_5722_squeeze_mask_0, x = k_complex_29_cast_fp16)[name = tensor("op_5722_cast_fp16")]; + tensor var_5730_begin_0 = const()[name = tensor("op_5730_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_5730_end_0 = const()[name = tensor("op_5730_end_0"), val = tensor([1, 1, 16, 32, 2])]; + tensor var_5730_end_mask_0 = const()[name = tensor("op_5730_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_5730_squeeze_mask_0 = const()[name = tensor("op_5730_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_5730_cast_fp16 = slice_by_index(begin = var_5730_begin_0, end = var_5730_end_0, end_mask = var_5730_end_mask_0, squeeze_mask = var_5730_squeeze_mask_0, x = k_complex_29_cast_fp16)[name = tensor("op_5730_cast_fp16")]; + tensor freqs_29_to_fp16 = const()[name = tensor("freqs_29_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(359872896)))]; + tensor var_5736_cast_fp16 = mul(x = freqs_29_to_fp16, y = ts_89_cast_fp16)[name = tensor("op_5736_cast_fp16")]; + tensor rotr_29_cast_fp16 = cos(x = var_5736_cast_fp16)[name = tensor("rotr_29_cast_fp16")]; + tensor roti_29_cast_fp16 = sin(x = var_5736_cast_fp16)[name = tensor("roti_29_cast_fp16")]; + tensor var_5740_cast_fp16 = mul(x = var_5706_cast_fp16, y = rotr_29_cast_fp16)[name = tensor("op_5740_cast_fp16")]; + tensor var_5741_cast_fp16 = mul(x = var_5714_cast_fp16, y = roti_29_cast_fp16)[name = tensor("op_5741_cast_fp16")]; + tensor qor_57_cast_fp16 = sub(x = var_5740_cast_fp16, y = var_5741_cast_fp16)[name = tensor("qor_57_cast_fp16")]; + tensor var_5744_cast_fp16 = mul(x = var_5706_cast_fp16, y = roti_29_cast_fp16)[name = tensor("op_5744_cast_fp16")]; + tensor var_5745_cast_fp16 = mul(x = var_5714_cast_fp16, y = rotr_29_cast_fp16)[name = tensor("op_5745_cast_fp16")]; + tensor qoi_57_cast_fp16 = add(x = var_5744_cast_fp16, y = var_5745_cast_fp16)[name = tensor("qoi_57_cast_fp16")]; + tensor var_5748_cast_fp16 = mul(x = var_5722_cast_fp16, y = rotr_29_cast_fp16)[name = tensor("op_5748_cast_fp16")]; + tensor var_5749_cast_fp16 = mul(x = var_5730_cast_fp16, y = roti_29_cast_fp16)[name = tensor("op_5749_cast_fp16")]; + tensor kor_57_cast_fp16 = sub(x = var_5748_cast_fp16, y = var_5749_cast_fp16)[name = tensor("kor_57_cast_fp16")]; + tensor var_5752_cast_fp16 = mul(x = var_5722_cast_fp16, y = roti_29_cast_fp16)[name = tensor("op_5752_cast_fp16")]; + tensor var_5753_cast_fp16 = mul(x = var_5730_cast_fp16, y = rotr_29_cast_fp16)[name = tensor("op_5753_cast_fp16")]; + tensor koi_57_cast_fp16 = add(x = var_5752_cast_fp16, y = var_5753_cast_fp16)[name = tensor("koi_57_cast_fp16")]; + tensor qo_29_axis_0 = const()[name = tensor("qo_29_axis_0"), val = tensor(-1)]; + tensor qo_29_cast_fp16 = stack(axis = qo_29_axis_0, values = (qor_57_cast_fp16, qoi_57_cast_fp16))[name = tensor("qo_29_cast_fp16")]; + tensor ko_29_axis_0 = const()[name = tensor("ko_29_axis_0"), val = tensor(-1)]; + tensor ko_29_cast_fp16 = stack(axis = ko_29_axis_0, values = (kor_57_cast_fp16, koi_57_cast_fp16))[name = tensor("ko_29_cast_fp16")]; + tensor var_5786 = const()[name = tensor("op_5786"), val = tensor([1, 1, 16, 64])]; + tensor q_87_cast_fp16 = reshape(shape = var_5786, x = qo_29_cast_fp16)[name = tensor("q_87_cast_fp16")]; + tensor var_5792 = const()[name = tensor("op_5792"), val = tensor([1, 1, 16, 64])]; + tensor k_59_cast_fp16 = reshape(shape = var_5792, x = ko_29_cast_fp16)[name = tensor("k_59_cast_fp16")]; + tensor _inversed_5817_y_0_to_fp16 = const()[name = tensor("_inversed_5817_y_0_to_fp16"), val = tensor(0x1p-9)]; + tensor _inversed_5817_cast_fp16 = mul(x = ts_89_cast_fp16, y = _inversed_5817_y_0_to_fp16)[name = tensor("_inversed_5817_cast_fp16")]; + tensor var_5818_cast_fp16 = floor(x = _inversed_5817_cast_fp16)[name = tensor("op_5818_cast_fp16")]; + tensor var_5819_to_fp16 = const()[name = tensor("op_5819_to_fp16"), val = tensor(0x1p+9)]; + tensor var_5820_cast_fp16 = mul(x = var_5818_cast_fp16, y = var_5819_to_fp16)[name = tensor("op_5820_cast_fp16")]; + tensor write_indices_float_59_cast_fp16 = sub(x = ts_89_cast_fp16, y = var_5820_cast_fp16)[name = tensor("write_indices_float_59_cast_fp16")]; + tensor var_5827_dtype_0 = const()[name = tensor("op_5827_dtype_0"), val = tensor("int32")]; + tensor write_indices_29_reps_0 = const()[name = tensor("write_indices_29_reps_0"), val = tensor([1, 1, 16, 64])]; + tensor write_indices_float_59_cast_fp16_to_int32 = cast(dtype = var_5827_dtype_0, x = write_indices_float_59_cast_fp16)[name = tensor("cast_482")]; + tensor write_indices_29 = tile(reps = write_indices_29_reps_0, x = write_indices_float_59_cast_fp16_to_int32)[name = tensor("write_indices_29")]; + tensor var_5837_begin_0 = const()[name = tensor("op_5837_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_5837_end_0 = const()[name = tensor("op_5837_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor var_5837_end_mask_0 = const()[name = tensor("op_5837_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_5837_squeeze_mask_0 = const()[name = tensor("op_5837_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor cache14_to_fp16_dtype_0 = const()[name = tensor("cache14_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor cache14_to_fp16 = cast(dtype = cache14_to_fp16_dtype_0, x = cache14)[name = tensor("cast_481")]; + tensor var_5837_cast_fp16 = slice_by_index(begin = var_5837_begin_0, end = var_5837_end_0, end_mask = var_5837_end_mask_0, squeeze_mask = var_5837_squeeze_mask_0, x = cache14_to_fp16)[name = tensor("op_5837_cast_fp16")]; + tensor var_5839_axis_0 = const()[name = tensor("op_5839_axis_0"), val = tensor(1)]; + tensor var_5839_mode_0 = const()[name = tensor("op_5839_mode_0"), val = tensor("update")]; + tensor var_5839_validate_indices_0 = const()[name = tensor("op_5839_validate_indices_0"), val = tensor(false)]; + tensor var_5839_cast_fp16 = scatter_along_axis(axis = var_5839_axis_0, data = var_5837_cast_fp16, indices = write_indices_29, mode = var_5839_mode_0, updates = k_59_cast_fp16, validate_indices = var_5839_validate_indices_0)[name = tensor("op_5839_cast_fp16")]; + tensor concat_100 = const()[name = tensor("concat_100"), val = tensor([0, 0, 0, 0, 0])]; + tensor concat_101 = const()[name = tensor("concat_101"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_29_internal_tensor_assign_1_stride_0 = const()[name = tensor("new_cache_29_internal_tensor_assign_1_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_29_internal_tensor_assign_1_begin_mask_0 = const()[name = tensor("new_cache_29_internal_tensor_assign_1_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_29_internal_tensor_assign_1_end_mask_0 = const()[name = tensor("new_cache_29_internal_tensor_assign_1_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_29_internal_tensor_assign_1_squeeze_mask_0 = const()[name = tensor("new_cache_29_internal_tensor_assign_1_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_76 = const()[name = tensor("shape_76"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_28 = const()[name = tensor("reduce_prod_28"), val = tensor(1048576)]; + tensor range_1d_28_start_0 = const()[name = tensor("range_1d_28_start_0"), val = tensor(0)]; + tensor range_1d_28_step_0 = const()[name = tensor("range_1d_28_step_0"), val = tensor(1)]; + tensor range_1d_28 = range_1d(end = reduce_prod_28, start = range_1d_28_start_0, step = range_1d_28_step_0)[name = tensor("range_1d_28")]; + tensor reshape_140 = reshape(shape = shape_76, x = range_1d_28)[name = tensor("reshape_140")]; + tensor slice_by_index_28 = slice_by_index(begin = concat_100, begin_mask = new_cache_29_internal_tensor_assign_1_begin_mask_0, end = concat_101, end_mask = new_cache_29_internal_tensor_assign_1_end_mask_0, squeeze_mask = new_cache_29_internal_tensor_assign_1_squeeze_mask_0, stride = new_cache_29_internal_tensor_assign_1_stride_0, x = reshape_140)[name = tensor("slice_by_index_28")]; + tensor reshape_141_shape_0 = const()[name = tensor("reshape_141_shape_0"), val = tensor([-1])]; + tensor reshape_141 = reshape(shape = reshape_141_shape_0, x = slice_by_index_28)[name = tensor("reshape_141")]; + tensor reshape_142_shape_0 = const()[name = tensor("reshape_142_shape_0"), val = tensor([-1])]; + tensor reshape_142_cast_fp16 = reshape(shape = reshape_142_shape_0, x = var_5839_cast_fp16)[name = tensor("reshape_142_cast_fp16")]; + tensor reshape_143_shape_0 = const()[name = tensor("reshape_143_shape_0"), val = tensor([-1])]; + tensor reshape_143_cast_fp16 = reshape(shape = reshape_143_shape_0, x = cache14_to_fp16)[name = tensor("reshape_143_cast_fp16")]; + tensor scatter_28_mode_0 = const()[name = tensor("scatter_28_mode_0"), val = tensor("update")]; + tensor scatter_28_axis_0 = const()[name = tensor("scatter_28_axis_0"), val = tensor(0)]; + tensor scatter_28_validate_indices_0 = const()[name = tensor("scatter_28_validate_indices_0"), val = tensor(false)]; + tensor scatter_28_cast_fp16 = scatter(axis = scatter_28_axis_0, data = reshape_143_cast_fp16, indices = reshape_141, mode = scatter_28_mode_0, updates = reshape_142_cast_fp16, validate_indices = scatter_28_validate_indices_0)[name = tensor("scatter_28_cast_fp16")]; + tensor reshape_144_cast_fp16 = reshape(shape = shape_76, x = scatter_28_cast_fp16)[name = tensor("reshape_144_cast_fp16")]; + tensor var_5847_begin_0 = const()[name = tensor("op_5847_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor var_5847_end_0 = const()[name = tensor("op_5847_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor var_5847_end_mask_0 = const()[name = tensor("op_5847_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_5847_squeeze_mask_0 = const()[name = tensor("op_5847_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor var_5847_cast_fp16 = slice_by_index(begin = var_5847_begin_0, end = var_5847_end_0, end_mask = var_5847_end_mask_0, squeeze_mask = var_5847_squeeze_mask_0, x = reshape_144_cast_fp16)[name = tensor("op_5847_cast_fp16")]; + tensor var_5849_axis_0 = const()[name = tensor("op_5849_axis_0"), val = tensor(1)]; + tensor var_5849_mode_0 = const()[name = tensor("op_5849_mode_0"), val = tensor("update")]; + tensor var_5849_validate_indices_0 = const()[name = tensor("op_5849_validate_indices_0"), val = tensor(false)]; + tensor var_5849_cast_fp16 = scatter_along_axis(axis = var_5849_axis_0, data = var_5847_cast_fp16, indices = write_indices_29, mode = var_5849_mode_0, updates = v_29_cast_fp16, validate_indices = var_5849_validate_indices_0)[name = tensor("op_5849_cast_fp16")]; + tensor concat_102 = const()[name = tensor("concat_102"), val = tensor([1, 0, 0, 0, 0])]; + tensor concat_103 = const()[name = tensor("concat_103"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_29_internal_tensor_assign_2_stride_0 = const()[name = tensor("new_cache_29_internal_tensor_assign_2_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_29_internal_tensor_assign_2_begin_mask_0 = const()[name = tensor("new_cache_29_internal_tensor_assign_2_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_29_internal_tensor_assign_2_end_mask_0 = const()[name = tensor("new_cache_29_internal_tensor_assign_2_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_29_internal_tensor_assign_2_squeeze_mask_0 = const()[name = tensor("new_cache_29_internal_tensor_assign_2_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_77 = const()[name = tensor("shape_77"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_29 = const()[name = tensor("reduce_prod_29"), val = tensor(1048576)]; + tensor range_1d_29_start_0 = const()[name = tensor("range_1d_29_start_0"), val = tensor(0)]; + tensor range_1d_29_step_0 = const()[name = tensor("range_1d_29_step_0"), val = tensor(1)]; + tensor range_1d_29 = range_1d(end = reduce_prod_29, start = range_1d_29_start_0, step = range_1d_29_step_0)[name = tensor("range_1d_29")]; + tensor reshape_145 = reshape(shape = shape_77, x = range_1d_29)[name = tensor("reshape_145")]; + tensor slice_by_index_29 = slice_by_index(begin = concat_102, begin_mask = new_cache_29_internal_tensor_assign_2_begin_mask_0, end = concat_103, end_mask = new_cache_29_internal_tensor_assign_2_end_mask_0, squeeze_mask = new_cache_29_internal_tensor_assign_2_squeeze_mask_0, stride = new_cache_29_internal_tensor_assign_2_stride_0, x = reshape_145)[name = tensor("slice_by_index_29")]; + tensor reshape_146_shape_0 = const()[name = tensor("reshape_146_shape_0"), val = tensor([-1])]; + tensor reshape_146 = reshape(shape = reshape_146_shape_0, x = slice_by_index_29)[name = tensor("reshape_146")]; + tensor reshape_147_shape_0 = const()[name = tensor("reshape_147_shape_0"), val = tensor([-1])]; + tensor reshape_147_cast_fp16 = reshape(shape = reshape_147_shape_0, x = var_5849_cast_fp16)[name = tensor("reshape_147_cast_fp16")]; + tensor reshape_148_shape_0 = const()[name = tensor("reshape_148_shape_0"), val = tensor([-1])]; + tensor reshape_148_cast_fp16 = reshape(shape = reshape_148_shape_0, x = reshape_144_cast_fp16)[name = tensor("reshape_148_cast_fp16")]; + tensor scatter_29_mode_0 = const()[name = tensor("scatter_29_mode_0"), val = tensor("update")]; + tensor scatter_29_axis_0 = const()[name = tensor("scatter_29_axis_0"), val = tensor(0)]; + tensor scatter_29_validate_indices_0 = const()[name = tensor("scatter_29_validate_indices_0"), val = tensor(false)]; + tensor scatter_29_cast_fp16 = scatter(axis = scatter_29_axis_0, data = reshape_148_cast_fp16, indices = reshape_146, mode = scatter_29_mode_0, updates = reshape_147_cast_fp16, validate_indices = scatter_29_validate_indices_0)[name = tensor("scatter_29_cast_fp16")]; + tensor reshape_149_cast_fp16 = reshape(shape = shape_77, x = scatter_29_cast_fp16)[name = tensor("reshape_149_cast_fp16")]; + tensor reshape_149_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("reshape_149_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor keys_85_begin_0 = const()[name = tensor("keys_85_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor keys_85_end_0 = const()[name = tensor("keys_85_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor keys_85_end_mask_0 = const()[name = tensor("keys_85_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor keys_85_squeeze_mask_0 = const()[name = tensor("keys_85_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor keys_85_cast_fp16 = slice_by_index(begin = keys_85_begin_0, end = keys_85_end_0, end_mask = keys_85_end_mask_0, squeeze_mask = keys_85_squeeze_mask_0, x = reshape_149_cast_fp16)[name = tensor("keys_85_cast_fp16")]; + tensor values_85_begin_0 = const()[name = tensor("values_85_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor values_85_end_0 = const()[name = tensor("values_85_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor values_85_end_mask_0 = const()[name = tensor("values_85_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor values_85_squeeze_mask_0 = const()[name = tensor("values_85_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor values_85_cast_fp16 = slice_by_index(begin = values_85_begin_0, end = values_85_end_0, end_mask = values_85_end_mask_0, squeeze_mask = values_85_squeeze_mask_0, x = reshape_149_cast_fp16)[name = tensor("values_85_cast_fp16")]; + tensor var_5861_cast_fp16 = not_equal(x = keys_85_cast_fp16, y = keys_85_cast_fp16)[name = tensor("op_5861_cast_fp16")]; + tensor keys_87_cast_fp16 = select(a = var_463_to_fp16, b = keys_85_cast_fp16, cond = var_5861_cast_fp16)[name = tensor("keys_87_cast_fp16")]; + tensor var_5869_cast_fp16 = not_equal(x = values_85_cast_fp16, y = values_85_cast_fp16)[name = tensor("op_5869_cast_fp16")]; + tensor values_87_cast_fp16 = select(a = var_463_to_fp16, b = values_85_cast_fp16, cond = var_5869_cast_fp16)[name = tensor("values_87_cast_fp16")]; + tensor var_5893 = const()[name = tensor("op_5893"), val = tensor([0, 2, 1, 3])]; + tensor var_5909 = const()[name = tensor("op_5909"), val = tensor([1, 1, 1])]; + tensor var_5910_cast_fp16 = reshape(shape = var_5909, x = position14_to_fp16)[name = tensor("op_5910_cast_fp16")]; + tensor var_5930_to_fp16 = const()[name = tensor("op_5930_to_fp16"), val = tensor(0x1p+0)]; + tensor valid_len_29_cast_fp16 = add(x = var_5910_cast_fp16, y = var_5930_to_fp16)[name = tensor("valid_len_29_cast_fp16")]; + tensor valid_mask_29_cast_fp16 = less(x = k_positions_1_promoted_to_fp16, y = valid_len_29_cast_fp16)[name = tensor("valid_mask_29_cast_fp16")]; + tensor causal_mask_29_cast_fp16 = less_equal(x = k_positions_1_promoted_to_fp16, y = var_5910_cast_fp16)[name = tensor("causal_mask_29_cast_fp16")]; + tensor attn_mask_57 = logical_and(x = valid_mask_29_cast_fp16, y = causal_mask_29_cast_fp16)[name = tensor("attn_mask_57")]; + tensor attn_mask_59_axes_0 = const()[name = tensor("attn_mask_59_axes_0"), val = tensor([1])]; + tensor attn_mask_59 = expand_dims(axes = attn_mask_59_axes_0, x = attn_mask_57)[name = tensor("attn_mask_59")]; + tensor var_5948_transpose_x_0 = const()[name = tensor("op_5948_transpose_x_0"), val = tensor(false)]; + tensor var_5948_transpose_y_0 = const()[name = tensor("op_5948_transpose_y_0"), val = tensor(false)]; + tensor transpose_100_perm_0 = const()[name = tensor("transpose_100_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_101_perm_0 = const()[name = tensor("transpose_101_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_101 = transpose(perm = transpose_101_perm_0, x = keys_87_cast_fp16)[name = tensor("transpose_157")]; + tensor transpose_100 = transpose(perm = transpose_100_perm_0, x = q_87_cast_fp16)[name = tensor("transpose_158")]; + tensor var_5948_cast_fp16 = matmul(transpose_x = var_5948_transpose_x_0, transpose_y = var_5948_transpose_y_0, x = transpose_100, y = transpose_101)[name = tensor("op_5948_cast_fp16")]; + tensor var_5942_to_fp16 = const()[name = tensor("op_5942_to_fp16"), val = tensor([0x1p-3])]; + tensor attn_weights_85_cast_fp16 = mul(x = var_5948_cast_fp16, y = var_5942_to_fp16)[name = tensor("attn_weights_85_cast_fp16")]; + tensor var_5950 = logical_not(x = attn_mask_59)[name = tensor("op_5950")]; + tensor var_5951_to_fp16 = const()[name = tensor("op_5951_to_fp16"), val = tensor(-inf)]; + tensor attn_weights_87_cast_fp16 = select(a = var_5951_to_fp16, b = attn_weights_85_cast_fp16, cond = var_5950)[name = tensor("attn_weights_87_cast_fp16")]; + tensor var_5953 = const()[name = tensor("op_5953"), val = tensor(-1)]; + tensor attn_weights_89_cast_fp16 = softmax(axis = var_5953, x = attn_weights_87_cast_fp16)[name = tensor("attn_weights_89_cast_fp16")]; + tensor attn_output_29_transpose_x_0 = const()[name = tensor("attn_output_29_transpose_x_0"), val = tensor(false)]; + tensor attn_output_29_transpose_y_0 = const()[name = tensor("attn_output_29_transpose_y_0"), val = tensor(false)]; + tensor values_89_cast_fp16 = transpose(perm = var_5893, x = values_87_cast_fp16)[name = tensor("transpose_159")]; + tensor attn_output_29_cast_fp16 = matmul(transpose_x = attn_output_29_transpose_x_0, transpose_y = attn_output_29_transpose_y_0, x = attn_weights_89_cast_fp16, y = values_89_cast_fp16)[name = tensor("attn_output_29_cast_fp16")]; + tensor var_5961 = const()[name = tensor("op_5961"), val = tensor([0, 2, 1, 3])]; + tensor var_5966 = const()[name = tensor("op_5966"), val = tensor([1, 1, 1024])]; + tensor var_5962_cast_fp16 = transpose(perm = var_5961, x = attn_output_29_cast_fp16)[name = tensor("transpose_156")]; + tensor input_175_cast_fp16 = reshape(shape = var_5966, x = var_5962_cast_fp16)[name = tensor("input_175_cast_fp16")]; + tensor attn14_out_proj_weight_to_fp16 = const()[name = tensor("attn14_out_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(359873024)))]; + tensor linear_58_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = attn14_out_proj_weight_to_fp16, x = input_175_cast_fp16)[name = tensor("linear_58_cast_fp16")]; + tensor var_5972_to_fp16 = const()[name = tensor("op_5972_to_fp16"), val = tensor(0x1p+0)]; + tensor var_5973_cast_fp16 = add(x = position14_to_fp16, y = var_5972_to_fp16)[name = tensor("op_5973_cast_fp16")]; + tensor var_5973_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_5973_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor input_177_cast_fp16 = add(x = input_171_cast_fp16, y = linear_58_cast_fp16)[name = tensor("input_177_cast_fp16")]; + tensor input_179_axes_0 = const()[name = tensor("input_179_axes_0"), val = tensor([-1])]; + tensor norm14_2_weight_to_fp16 = const()[name = tensor("norm14_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(361970240)))]; + tensor norm14_2_bias_to_fp16 = const()[name = tensor("norm14_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(361972352)))]; + tensor var_5977_to_fp16 = const()[name = tensor("op_5977_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_179_cast_fp16 = layer_norm(axes = input_179_axes_0, beta = norm14_2_bias_to_fp16, epsilon = var_5977_to_fp16, gamma = norm14_2_weight_to_fp16, x = input_177_cast_fp16)[name = tensor("input_179_cast_fp16")]; + tensor linear14_1_weight_to_fp16 = const()[name = tensor("linear14_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(361974464)))]; + tensor linear_59_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = linear14_1_weight_to_fp16, x = input_179_cast_fp16)[name = tensor("linear_59_cast_fp16")]; + tensor input_181_mode_0 = const()[name = tensor("input_181_mode_0"), val = tensor("EXACT")]; + tensor input_181_cast_fp16 = gelu(mode = input_181_mode_0, x = linear_59_cast_fp16)[name = tensor("input_181_cast_fp16")]; + tensor linear14_2_weight_to_fp16 = const()[name = tensor("linear14_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(370363136)))]; + tensor linear_60_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = linear14_2_weight_to_fp16, x = input_181_cast_fp16)[name = tensor("linear_60_cast_fp16")]; + tensor input_183_cast_fp16 = add(x = input_177_cast_fp16, y = linear_60_cast_fp16)[name = tensor("input_183_cast_fp16")]; + tensor input_185_axes_0 = const()[name = tensor("input_185_axes_0"), val = tensor([-1])]; + tensor norm15_1_weight_to_fp16 = const()[name = tensor("norm15_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(378751808)))]; + tensor norm15_1_bias_to_fp16 = const()[name = tensor("norm15_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(378753920)))]; + tensor var_5994_to_fp16 = const()[name = tensor("op_5994_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_185_cast_fp16 = layer_norm(axes = input_185_axes_0, beta = norm15_1_bias_to_fp16, epsilon = var_5994_to_fp16, gamma = norm15_1_weight_to_fp16, x = input_183_cast_fp16)[name = tensor("input_185_cast_fp16")]; + tensor attn15_in_proj_weight_to_fp16 = const()[name = tensor("attn15_in_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(378756032)))]; + tensor linear_61_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = attn15_in_proj_weight_to_fp16, x = input_185_cast_fp16)[name = tensor("linear_61_cast_fp16")]; + tensor var_6008 = const()[name = tensor("op_6008"), val = tensor([1, 1, 3, 16, 64])]; + tensor qkv_31_cast_fp16 = reshape(shape = var_6008, x = linear_61_cast_fp16)[name = tensor("qkv_31_cast_fp16")]; + tensor q_91_begin_0 = const()[name = tensor("q_91_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor q_91_end_0 = const()[name = tensor("q_91_end_0"), val = tensor([1, 1, 1, 16, 64])]; + tensor q_91_end_mask_0 = const()[name = tensor("q_91_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor q_91_squeeze_mask_0 = const()[name = tensor("q_91_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor q_91_cast_fp16 = slice_by_index(begin = q_91_begin_0, end = q_91_end_0, end_mask = q_91_end_mask_0, squeeze_mask = q_91_squeeze_mask_0, x = qkv_31_cast_fp16)[name = tensor("q_91_cast_fp16")]; + tensor k_61_begin_0 = const()[name = tensor("k_61_begin_0"), val = tensor([0, 0, 1, 0, 0])]; + tensor k_61_end_0 = const()[name = tensor("k_61_end_0"), val = tensor([1, 1, 2, 16, 64])]; + tensor k_61_end_mask_0 = const()[name = tensor("k_61_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor k_61_squeeze_mask_0 = const()[name = tensor("k_61_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor k_61_cast_fp16 = slice_by_index(begin = k_61_begin_0, end = k_61_end_0, end_mask = k_61_end_mask_0, squeeze_mask = k_61_squeeze_mask_0, x = qkv_31_cast_fp16)[name = tensor("k_61_cast_fp16")]; + tensor v_31_begin_0 = const()[name = tensor("v_31_begin_0"), val = tensor([0, 0, 2, 0, 0])]; + tensor v_31_end_0 = const()[name = tensor("v_31_end_0"), val = tensor([1, 1, 3, 16, 64])]; + tensor v_31_end_mask_0 = const()[name = tensor("v_31_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor v_31_squeeze_mask_0 = const()[name = tensor("v_31_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor v_31_cast_fp16 = slice_by_index(begin = v_31_begin_0, end = v_31_end_0, end_mask = v_31_end_mask_0, squeeze_mask = v_31_squeeze_mask_0, x = qkv_31_cast_fp16)[name = tensor("v_31_cast_fp16")]; + tensor var_6074 = const()[name = tensor("op_6074"), val = tensor([1, 1, 1, 1])]; + tensor position15_to_fp16_dtype_0 = const()[name = tensor("position15_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor position15_to_fp16 = cast(dtype = position15_to_fp16_dtype_0, x = position15)[name = tensor("cast_478")]; + tensor ts_95_cast_fp16 = reshape(shape = var_6074, x = position15_to_fp16)[name = tensor("ts_95_cast_fp16")]; + tensor var_6081 = const()[name = tensor("op_6081"), val = tensor([1, 1, 16, 32, 2])]; + tensor q_complex_31_cast_fp16 = reshape(shape = var_6081, x = q_91_cast_fp16)[name = tensor("q_complex_31_cast_fp16")]; + tensor var_6088 = const()[name = tensor("op_6088"), val = tensor([1, 1, 16, 32, 2])]; + tensor k_complex_31_cast_fp16 = reshape(shape = var_6088, x = k_61_cast_fp16)[name = tensor("k_complex_31_cast_fp16")]; + tensor var_6092_begin_0 = const()[name = tensor("op_6092_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_6092_end_0 = const()[name = tensor("op_6092_end_0"), val = tensor([1, 1, 16, 32, 1])]; + tensor var_6092_end_mask_0 = const()[name = tensor("op_6092_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_6092_squeeze_mask_0 = const()[name = tensor("op_6092_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_6092_cast_fp16 = slice_by_index(begin = var_6092_begin_0, end = var_6092_end_0, end_mask = var_6092_end_mask_0, squeeze_mask = var_6092_squeeze_mask_0, x = q_complex_31_cast_fp16)[name = tensor("op_6092_cast_fp16")]; + tensor var_6100_begin_0 = const()[name = tensor("op_6100_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_6100_end_0 = const()[name = tensor("op_6100_end_0"), val = tensor([1, 1, 16, 32, 2])]; + tensor var_6100_end_mask_0 = const()[name = tensor("op_6100_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_6100_squeeze_mask_0 = const()[name = tensor("op_6100_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_6100_cast_fp16 = slice_by_index(begin = var_6100_begin_0, end = var_6100_end_0, end_mask = var_6100_end_mask_0, squeeze_mask = var_6100_squeeze_mask_0, x = q_complex_31_cast_fp16)[name = tensor("op_6100_cast_fp16")]; + tensor var_6108_begin_0 = const()[name = tensor("op_6108_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_6108_end_0 = const()[name = tensor("op_6108_end_0"), val = tensor([1, 1, 16, 32, 1])]; + tensor var_6108_end_mask_0 = const()[name = tensor("op_6108_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_6108_squeeze_mask_0 = const()[name = tensor("op_6108_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_6108_cast_fp16 = slice_by_index(begin = var_6108_begin_0, end = var_6108_end_0, end_mask = var_6108_end_mask_0, squeeze_mask = var_6108_squeeze_mask_0, x = k_complex_31_cast_fp16)[name = tensor("op_6108_cast_fp16")]; + tensor var_6116_begin_0 = const()[name = tensor("op_6116_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_6116_end_0 = const()[name = tensor("op_6116_end_0"), val = tensor([1, 1, 16, 32, 2])]; + tensor var_6116_end_mask_0 = const()[name = tensor("op_6116_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_6116_squeeze_mask_0 = const()[name = tensor("op_6116_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_6116_cast_fp16 = slice_by_index(begin = var_6116_begin_0, end = var_6116_end_0, end_mask = var_6116_end_mask_0, squeeze_mask = var_6116_squeeze_mask_0, x = k_complex_31_cast_fp16)[name = tensor("op_6116_cast_fp16")]; + tensor freqs_31_to_fp16 = const()[name = tensor("freqs_31_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(385047552)))]; + tensor var_6122_cast_fp16 = mul(x = freqs_31_to_fp16, y = ts_95_cast_fp16)[name = tensor("op_6122_cast_fp16")]; + tensor rotr_31_cast_fp16 = cos(x = var_6122_cast_fp16)[name = tensor("rotr_31_cast_fp16")]; + tensor roti_31_cast_fp16 = sin(x = var_6122_cast_fp16)[name = tensor("roti_31_cast_fp16")]; + tensor var_6126_cast_fp16 = mul(x = var_6092_cast_fp16, y = rotr_31_cast_fp16)[name = tensor("op_6126_cast_fp16")]; + tensor var_6127_cast_fp16 = mul(x = var_6100_cast_fp16, y = roti_31_cast_fp16)[name = tensor("op_6127_cast_fp16")]; + tensor qor_61_cast_fp16 = sub(x = var_6126_cast_fp16, y = var_6127_cast_fp16)[name = tensor("qor_61_cast_fp16")]; + tensor var_6130_cast_fp16 = mul(x = var_6092_cast_fp16, y = roti_31_cast_fp16)[name = tensor("op_6130_cast_fp16")]; + tensor var_6131_cast_fp16 = mul(x = var_6100_cast_fp16, y = rotr_31_cast_fp16)[name = tensor("op_6131_cast_fp16")]; + tensor qoi_61_cast_fp16 = add(x = var_6130_cast_fp16, y = var_6131_cast_fp16)[name = tensor("qoi_61_cast_fp16")]; + tensor var_6134_cast_fp16 = mul(x = var_6108_cast_fp16, y = rotr_31_cast_fp16)[name = tensor("op_6134_cast_fp16")]; + tensor var_6135_cast_fp16 = mul(x = var_6116_cast_fp16, y = roti_31_cast_fp16)[name = tensor("op_6135_cast_fp16")]; + tensor kor_61_cast_fp16 = sub(x = var_6134_cast_fp16, y = var_6135_cast_fp16)[name = tensor("kor_61_cast_fp16")]; + tensor var_6138_cast_fp16 = mul(x = var_6108_cast_fp16, y = roti_31_cast_fp16)[name = tensor("op_6138_cast_fp16")]; + tensor var_6139_cast_fp16 = mul(x = var_6116_cast_fp16, y = rotr_31_cast_fp16)[name = tensor("op_6139_cast_fp16")]; + tensor koi_61_cast_fp16 = add(x = var_6138_cast_fp16, y = var_6139_cast_fp16)[name = tensor("koi_61_cast_fp16")]; + tensor qo_31_axis_0 = const()[name = tensor("qo_31_axis_0"), val = tensor(-1)]; + tensor qo_31_cast_fp16 = stack(axis = qo_31_axis_0, values = (qor_61_cast_fp16, qoi_61_cast_fp16))[name = tensor("qo_31_cast_fp16")]; + tensor ko_31_axis_0 = const()[name = tensor("ko_31_axis_0"), val = tensor(-1)]; + tensor ko_31_cast_fp16 = stack(axis = ko_31_axis_0, values = (kor_61_cast_fp16, koi_61_cast_fp16))[name = tensor("ko_31_cast_fp16")]; + tensor var_6172 = const()[name = tensor("op_6172"), val = tensor([1, 1, 16, 64])]; + tensor q_93_cast_fp16 = reshape(shape = var_6172, x = qo_31_cast_fp16)[name = tensor("q_93_cast_fp16")]; + tensor var_6178 = const()[name = tensor("op_6178"), val = tensor([1, 1, 16, 64])]; + tensor k_63_cast_fp16 = reshape(shape = var_6178, x = ko_31_cast_fp16)[name = tensor("k_63_cast_fp16")]; + tensor _inversed_6203_y_0_to_fp16 = const()[name = tensor("_inversed_6203_y_0_to_fp16"), val = tensor(0x1p-9)]; + tensor _inversed_6203_cast_fp16 = mul(x = ts_95_cast_fp16, y = _inversed_6203_y_0_to_fp16)[name = tensor("_inversed_6203_cast_fp16")]; + tensor var_6204_cast_fp16 = floor(x = _inversed_6203_cast_fp16)[name = tensor("op_6204_cast_fp16")]; + tensor var_6205_to_fp16 = const()[name = tensor("op_6205_to_fp16"), val = tensor(0x1p+9)]; + tensor var_6206_cast_fp16 = mul(x = var_6204_cast_fp16, y = var_6205_to_fp16)[name = tensor("op_6206_cast_fp16")]; + tensor write_indices_float_63_cast_fp16 = sub(x = ts_95_cast_fp16, y = var_6206_cast_fp16)[name = tensor("write_indices_float_63_cast_fp16")]; + tensor var_6213_dtype_0 = const()[name = tensor("op_6213_dtype_0"), val = tensor("int32")]; + tensor write_indices_31_reps_0 = const()[name = tensor("write_indices_31_reps_0"), val = tensor([1, 1, 16, 64])]; + tensor write_indices_float_63_cast_fp16_to_int32 = cast(dtype = var_6213_dtype_0, x = write_indices_float_63_cast_fp16)[name = tensor("cast_477")]; + tensor write_indices_31 = tile(reps = write_indices_31_reps_0, x = write_indices_float_63_cast_fp16_to_int32)[name = tensor("write_indices_31")]; + tensor var_6223_begin_0 = const()[name = tensor("op_6223_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_6223_end_0 = const()[name = tensor("op_6223_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor var_6223_end_mask_0 = const()[name = tensor("op_6223_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_6223_squeeze_mask_0 = const()[name = tensor("op_6223_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor cache15_to_fp16_dtype_0 = const()[name = tensor("cache15_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor cache15_to_fp16 = cast(dtype = cache15_to_fp16_dtype_0, x = cache15)[name = tensor("cast_476")]; + tensor var_6223_cast_fp16 = slice_by_index(begin = var_6223_begin_0, end = var_6223_end_0, end_mask = var_6223_end_mask_0, squeeze_mask = var_6223_squeeze_mask_0, x = cache15_to_fp16)[name = tensor("op_6223_cast_fp16")]; + tensor var_6225_axis_0 = const()[name = tensor("op_6225_axis_0"), val = tensor(1)]; + tensor var_6225_mode_0 = const()[name = tensor("op_6225_mode_0"), val = tensor("update")]; + tensor var_6225_validate_indices_0 = const()[name = tensor("op_6225_validate_indices_0"), val = tensor(false)]; + tensor var_6225_cast_fp16 = scatter_along_axis(axis = var_6225_axis_0, data = var_6223_cast_fp16, indices = write_indices_31, mode = var_6225_mode_0, updates = k_63_cast_fp16, validate_indices = var_6225_validate_indices_0)[name = tensor("op_6225_cast_fp16")]; + tensor concat_107 = const()[name = tensor("concat_107"), val = tensor([0, 0, 0, 0, 0])]; + tensor concat_108 = const()[name = tensor("concat_108"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_31_internal_tensor_assign_1_stride_0 = const()[name = tensor("new_cache_31_internal_tensor_assign_1_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_31_internal_tensor_assign_1_begin_mask_0 = const()[name = tensor("new_cache_31_internal_tensor_assign_1_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_31_internal_tensor_assign_1_end_mask_0 = const()[name = tensor("new_cache_31_internal_tensor_assign_1_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_31_internal_tensor_assign_1_squeeze_mask_0 = const()[name = tensor("new_cache_31_internal_tensor_assign_1_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_78 = const()[name = tensor("shape_78"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_30 = const()[name = tensor("reduce_prod_30"), val = tensor(1048576)]; + tensor range_1d_30_start_0 = const()[name = tensor("range_1d_30_start_0"), val = tensor(0)]; + tensor range_1d_30_step_0 = const()[name = tensor("range_1d_30_step_0"), val = tensor(1)]; + tensor range_1d_30 = range_1d(end = reduce_prod_30, start = range_1d_30_start_0, step = range_1d_30_step_0)[name = tensor("range_1d_30")]; + tensor reshape_150 = reshape(shape = shape_78, x = range_1d_30)[name = tensor("reshape_150")]; + tensor slice_by_index_30 = slice_by_index(begin = concat_107, begin_mask = new_cache_31_internal_tensor_assign_1_begin_mask_0, end = concat_108, end_mask = new_cache_31_internal_tensor_assign_1_end_mask_0, squeeze_mask = new_cache_31_internal_tensor_assign_1_squeeze_mask_0, stride = new_cache_31_internal_tensor_assign_1_stride_0, x = reshape_150)[name = tensor("slice_by_index_30")]; + tensor reshape_151_shape_0 = const()[name = tensor("reshape_151_shape_0"), val = tensor([-1])]; + tensor reshape_151 = reshape(shape = reshape_151_shape_0, x = slice_by_index_30)[name = tensor("reshape_151")]; + tensor reshape_152_shape_0 = const()[name = tensor("reshape_152_shape_0"), val = tensor([-1])]; + tensor reshape_152_cast_fp16 = reshape(shape = reshape_152_shape_0, x = var_6225_cast_fp16)[name = tensor("reshape_152_cast_fp16")]; + tensor reshape_153_shape_0 = const()[name = tensor("reshape_153_shape_0"), val = tensor([-1])]; + tensor reshape_153_cast_fp16 = reshape(shape = reshape_153_shape_0, x = cache15_to_fp16)[name = tensor("reshape_153_cast_fp16")]; + tensor scatter_30_mode_0 = const()[name = tensor("scatter_30_mode_0"), val = tensor("update")]; + tensor scatter_30_axis_0 = const()[name = tensor("scatter_30_axis_0"), val = tensor(0)]; + tensor scatter_30_validate_indices_0 = const()[name = tensor("scatter_30_validate_indices_0"), val = tensor(false)]; + tensor scatter_30_cast_fp16 = scatter(axis = scatter_30_axis_0, data = reshape_153_cast_fp16, indices = reshape_151, mode = scatter_30_mode_0, updates = reshape_152_cast_fp16, validate_indices = scatter_30_validate_indices_0)[name = tensor("scatter_30_cast_fp16")]; + tensor reshape_154_cast_fp16 = reshape(shape = shape_78, x = scatter_30_cast_fp16)[name = tensor("reshape_154_cast_fp16")]; + tensor var_6233_begin_0 = const()[name = tensor("op_6233_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor var_6233_end_0 = const()[name = tensor("op_6233_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor var_6233_end_mask_0 = const()[name = tensor("op_6233_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_6233_squeeze_mask_0 = const()[name = tensor("op_6233_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor var_6233_cast_fp16 = slice_by_index(begin = var_6233_begin_0, end = var_6233_end_0, end_mask = var_6233_end_mask_0, squeeze_mask = var_6233_squeeze_mask_0, x = reshape_154_cast_fp16)[name = tensor("op_6233_cast_fp16")]; + tensor var_6235_axis_0 = const()[name = tensor("op_6235_axis_0"), val = tensor(1)]; + tensor var_6235_mode_0 = const()[name = tensor("op_6235_mode_0"), val = tensor("update")]; + tensor var_6235_validate_indices_0 = const()[name = tensor("op_6235_validate_indices_0"), val = tensor(false)]; + tensor var_6235_cast_fp16 = scatter_along_axis(axis = var_6235_axis_0, data = var_6233_cast_fp16, indices = write_indices_31, mode = var_6235_mode_0, updates = v_31_cast_fp16, validate_indices = var_6235_validate_indices_0)[name = tensor("op_6235_cast_fp16")]; + tensor concat_109 = const()[name = tensor("concat_109"), val = tensor([1, 0, 0, 0, 0])]; + tensor concat_110 = const()[name = tensor("concat_110"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_31_internal_tensor_assign_2_stride_0 = const()[name = tensor("new_cache_31_internal_tensor_assign_2_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_31_internal_tensor_assign_2_begin_mask_0 = const()[name = tensor("new_cache_31_internal_tensor_assign_2_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_31_internal_tensor_assign_2_end_mask_0 = const()[name = tensor("new_cache_31_internal_tensor_assign_2_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_31_internal_tensor_assign_2_squeeze_mask_0 = const()[name = tensor("new_cache_31_internal_tensor_assign_2_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_79 = const()[name = tensor("shape_79"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_31 = const()[name = tensor("reduce_prod_31"), val = tensor(1048576)]; + tensor range_1d_31_start_0 = const()[name = tensor("range_1d_31_start_0"), val = tensor(0)]; + tensor range_1d_31_step_0 = const()[name = tensor("range_1d_31_step_0"), val = tensor(1)]; + tensor range_1d_31 = range_1d(end = reduce_prod_31, start = range_1d_31_start_0, step = range_1d_31_step_0)[name = tensor("range_1d_31")]; + tensor reshape_155 = reshape(shape = shape_79, x = range_1d_31)[name = tensor("reshape_155")]; + tensor slice_by_index_31 = slice_by_index(begin = concat_109, begin_mask = new_cache_31_internal_tensor_assign_2_begin_mask_0, end = concat_110, end_mask = new_cache_31_internal_tensor_assign_2_end_mask_0, squeeze_mask = new_cache_31_internal_tensor_assign_2_squeeze_mask_0, stride = new_cache_31_internal_tensor_assign_2_stride_0, x = reshape_155)[name = tensor("slice_by_index_31")]; + tensor reshape_156_shape_0 = const()[name = tensor("reshape_156_shape_0"), val = tensor([-1])]; + tensor reshape_156 = reshape(shape = reshape_156_shape_0, x = slice_by_index_31)[name = tensor("reshape_156")]; + tensor reshape_157_shape_0 = const()[name = tensor("reshape_157_shape_0"), val = tensor([-1])]; + tensor reshape_157_cast_fp16 = reshape(shape = reshape_157_shape_0, x = var_6235_cast_fp16)[name = tensor("reshape_157_cast_fp16")]; + tensor reshape_158_shape_0 = const()[name = tensor("reshape_158_shape_0"), val = tensor([-1])]; + tensor reshape_158_cast_fp16 = reshape(shape = reshape_158_shape_0, x = reshape_154_cast_fp16)[name = tensor("reshape_158_cast_fp16")]; + tensor scatter_31_mode_0 = const()[name = tensor("scatter_31_mode_0"), val = tensor("update")]; + tensor scatter_31_axis_0 = const()[name = tensor("scatter_31_axis_0"), val = tensor(0)]; + tensor scatter_31_validate_indices_0 = const()[name = tensor("scatter_31_validate_indices_0"), val = tensor(false)]; + tensor scatter_31_cast_fp16 = scatter(axis = scatter_31_axis_0, data = reshape_158_cast_fp16, indices = reshape_156, mode = scatter_31_mode_0, updates = reshape_157_cast_fp16, validate_indices = scatter_31_validate_indices_0)[name = tensor("scatter_31_cast_fp16")]; + tensor reshape_159_cast_fp16 = reshape(shape = shape_79, x = scatter_31_cast_fp16)[name = tensor("reshape_159_cast_fp16")]; + tensor reshape_159_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("reshape_159_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor keys_91_begin_0 = const()[name = tensor("keys_91_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor keys_91_end_0 = const()[name = tensor("keys_91_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor keys_91_end_mask_0 = const()[name = tensor("keys_91_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor keys_91_squeeze_mask_0 = const()[name = tensor("keys_91_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor keys_91_cast_fp16 = slice_by_index(begin = keys_91_begin_0, end = keys_91_end_0, end_mask = keys_91_end_mask_0, squeeze_mask = keys_91_squeeze_mask_0, x = reshape_159_cast_fp16)[name = tensor("keys_91_cast_fp16")]; + tensor values_91_begin_0 = const()[name = tensor("values_91_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor values_91_end_0 = const()[name = tensor("values_91_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor values_91_end_mask_0 = const()[name = tensor("values_91_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor values_91_squeeze_mask_0 = const()[name = tensor("values_91_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor values_91_cast_fp16 = slice_by_index(begin = values_91_begin_0, end = values_91_end_0, end_mask = values_91_end_mask_0, squeeze_mask = values_91_squeeze_mask_0, x = reshape_159_cast_fp16)[name = tensor("values_91_cast_fp16")]; + tensor var_6247_cast_fp16 = not_equal(x = keys_91_cast_fp16, y = keys_91_cast_fp16)[name = tensor("op_6247_cast_fp16")]; + tensor keys_93_cast_fp16 = select(a = var_463_to_fp16, b = keys_91_cast_fp16, cond = var_6247_cast_fp16)[name = tensor("keys_93_cast_fp16")]; + tensor var_6255_cast_fp16 = not_equal(x = values_91_cast_fp16, y = values_91_cast_fp16)[name = tensor("op_6255_cast_fp16")]; + tensor values_93_cast_fp16 = select(a = var_463_to_fp16, b = values_91_cast_fp16, cond = var_6255_cast_fp16)[name = tensor("values_93_cast_fp16")]; + tensor var_6279 = const()[name = tensor("op_6279"), val = tensor([0, 2, 1, 3])]; + tensor var_6295 = const()[name = tensor("op_6295"), val = tensor([1, 1, 1])]; + tensor var_6296_cast_fp16 = reshape(shape = var_6295, x = position15_to_fp16)[name = tensor("op_6296_cast_fp16")]; + tensor var_6316_to_fp16 = const()[name = tensor("op_6316_to_fp16"), val = tensor(0x1p+0)]; + tensor valid_len_31_cast_fp16 = add(x = var_6296_cast_fp16, y = var_6316_to_fp16)[name = tensor("valid_len_31_cast_fp16")]; + tensor valid_mask_31_cast_fp16 = less(x = k_positions_1_promoted_to_fp16, y = valid_len_31_cast_fp16)[name = tensor("valid_mask_31_cast_fp16")]; + tensor causal_mask_31_cast_fp16 = less_equal(x = k_positions_1_promoted_to_fp16, y = var_6296_cast_fp16)[name = tensor("causal_mask_31_cast_fp16")]; + tensor attn_mask_61 = logical_and(x = valid_mask_31_cast_fp16, y = causal_mask_31_cast_fp16)[name = tensor("attn_mask_61")]; + tensor attn_mask_63_axes_0 = const()[name = tensor("attn_mask_63_axes_0"), val = tensor([1])]; + tensor attn_mask_63 = expand_dims(axes = attn_mask_63_axes_0, x = attn_mask_61)[name = tensor("attn_mask_63")]; + tensor var_6334_transpose_x_0 = const()[name = tensor("op_6334_transpose_x_0"), val = tensor(false)]; + tensor var_6334_transpose_y_0 = const()[name = tensor("op_6334_transpose_y_0"), val = tensor(false)]; + tensor transpose_102_perm_0 = const()[name = tensor("transpose_102_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_103_perm_0 = const()[name = tensor("transpose_103_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_103 = transpose(perm = transpose_103_perm_0, x = keys_93_cast_fp16)[name = tensor("transpose_153")]; + tensor transpose_102 = transpose(perm = transpose_102_perm_0, x = q_93_cast_fp16)[name = tensor("transpose_154")]; + tensor var_6334_cast_fp16 = matmul(transpose_x = var_6334_transpose_x_0, transpose_y = var_6334_transpose_y_0, x = transpose_102, y = transpose_103)[name = tensor("op_6334_cast_fp16")]; + tensor var_6328_to_fp16 = const()[name = tensor("op_6328_to_fp16"), val = tensor([0x1p-3])]; + tensor attn_weights_91_cast_fp16 = mul(x = var_6334_cast_fp16, y = var_6328_to_fp16)[name = tensor("attn_weights_91_cast_fp16")]; + tensor var_6336 = logical_not(x = attn_mask_63)[name = tensor("op_6336")]; + tensor var_6337_to_fp16 = const()[name = tensor("op_6337_to_fp16"), val = tensor(-inf)]; + tensor attn_weights_93_cast_fp16 = select(a = var_6337_to_fp16, b = attn_weights_91_cast_fp16, cond = var_6336)[name = tensor("attn_weights_93_cast_fp16")]; + tensor var_6339 = const()[name = tensor("op_6339"), val = tensor(-1)]; + tensor attn_weights_95_cast_fp16 = softmax(axis = var_6339, x = attn_weights_93_cast_fp16)[name = tensor("attn_weights_95_cast_fp16")]; + tensor attn_output_31_transpose_x_0 = const()[name = tensor("attn_output_31_transpose_x_0"), val = tensor(false)]; + tensor attn_output_31_transpose_y_0 = const()[name = tensor("attn_output_31_transpose_y_0"), val = tensor(false)]; + tensor values_95_cast_fp16 = transpose(perm = var_6279, x = values_93_cast_fp16)[name = tensor("transpose_155")]; + tensor attn_output_31_cast_fp16 = matmul(transpose_x = attn_output_31_transpose_x_0, transpose_y = attn_output_31_transpose_y_0, x = attn_weights_95_cast_fp16, y = values_95_cast_fp16)[name = tensor("attn_output_31_cast_fp16")]; + tensor var_6347 = const()[name = tensor("op_6347"), val = tensor([0, 2, 1, 3])]; + tensor var_6352 = const()[name = tensor("op_6352"), val = tensor([1, 1, 1024])]; + tensor var_6348_cast_fp16 = transpose(perm = var_6347, x = attn_output_31_cast_fp16)[name = tensor("transpose_152")]; + tensor input_187_cast_fp16 = reshape(shape = var_6352, x = var_6348_cast_fp16)[name = tensor("input_187_cast_fp16")]; + tensor attn15_out_proj_weight_to_fp16 = const()[name = tensor("attn15_out_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(385047680)))]; + tensor linear_62_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = attn15_out_proj_weight_to_fp16, x = input_187_cast_fp16)[name = tensor("linear_62_cast_fp16")]; + tensor var_6358_to_fp16 = const()[name = tensor("op_6358_to_fp16"), val = tensor(0x1p+0)]; + tensor var_6359_cast_fp16 = add(x = position15_to_fp16, y = var_6358_to_fp16)[name = tensor("op_6359_cast_fp16")]; + tensor var_6359_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_6359_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor input_189_cast_fp16 = add(x = input_183_cast_fp16, y = linear_62_cast_fp16)[name = tensor("input_189_cast_fp16")]; + tensor input_191_axes_0 = const()[name = tensor("input_191_axes_0"), val = tensor([-1])]; + tensor norm15_2_weight_to_fp16 = const()[name = tensor("norm15_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(387144896)))]; + tensor norm15_2_bias_to_fp16 = const()[name = tensor("norm15_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(387147008)))]; + tensor var_6363_to_fp16 = const()[name = tensor("op_6363_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_191_cast_fp16 = layer_norm(axes = input_191_axes_0, beta = norm15_2_bias_to_fp16, epsilon = var_6363_to_fp16, gamma = norm15_2_weight_to_fp16, x = input_189_cast_fp16)[name = tensor("input_191_cast_fp16")]; + tensor linear15_1_weight_to_fp16 = const()[name = tensor("linear15_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(387149120)))]; + tensor linear_63_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = linear15_1_weight_to_fp16, x = input_191_cast_fp16)[name = tensor("linear_63_cast_fp16")]; + tensor input_193_mode_0 = const()[name = tensor("input_193_mode_0"), val = tensor("EXACT")]; + tensor input_193_cast_fp16 = gelu(mode = input_193_mode_0, x = linear_63_cast_fp16)[name = tensor("input_193_cast_fp16")]; + tensor linear15_2_weight_to_fp16 = const()[name = tensor("linear15_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(395537792)))]; + tensor linear_64_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = linear15_2_weight_to_fp16, x = input_193_cast_fp16)[name = tensor("linear_64_cast_fp16")]; + tensor input_195_cast_fp16 = add(x = input_189_cast_fp16, y = linear_64_cast_fp16)[name = tensor("input_195_cast_fp16")]; + tensor input_197_axes_0 = const()[name = tensor("input_197_axes_0"), val = tensor([-1])]; + tensor norm16_1_weight_to_fp16 = const()[name = tensor("norm16_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(403926464)))]; + tensor norm16_1_bias_to_fp16 = const()[name = tensor("norm16_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(403928576)))]; + tensor var_6380_to_fp16 = const()[name = tensor("op_6380_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_197_cast_fp16 = layer_norm(axes = input_197_axes_0, beta = norm16_1_bias_to_fp16, epsilon = var_6380_to_fp16, gamma = norm16_1_weight_to_fp16, x = input_195_cast_fp16)[name = tensor("input_197_cast_fp16")]; + tensor attn16_in_proj_weight_to_fp16 = const()[name = tensor("attn16_in_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(403930688)))]; + tensor linear_65_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = attn16_in_proj_weight_to_fp16, x = input_197_cast_fp16)[name = tensor("linear_65_cast_fp16")]; + tensor var_6394 = const()[name = tensor("op_6394"), val = tensor([1, 1, 3, 16, 64])]; + tensor qkv_33_cast_fp16 = reshape(shape = var_6394, x = linear_65_cast_fp16)[name = tensor("qkv_33_cast_fp16")]; + tensor q_97_begin_0 = const()[name = tensor("q_97_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor q_97_end_0 = const()[name = tensor("q_97_end_0"), val = tensor([1, 1, 1, 16, 64])]; + tensor q_97_end_mask_0 = const()[name = tensor("q_97_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor q_97_squeeze_mask_0 = const()[name = tensor("q_97_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor q_97_cast_fp16 = slice_by_index(begin = q_97_begin_0, end = q_97_end_0, end_mask = q_97_end_mask_0, squeeze_mask = q_97_squeeze_mask_0, x = qkv_33_cast_fp16)[name = tensor("q_97_cast_fp16")]; + tensor k_65_begin_0 = const()[name = tensor("k_65_begin_0"), val = tensor([0, 0, 1, 0, 0])]; + tensor k_65_end_0 = const()[name = tensor("k_65_end_0"), val = tensor([1, 1, 2, 16, 64])]; + tensor k_65_end_mask_0 = const()[name = tensor("k_65_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor k_65_squeeze_mask_0 = const()[name = tensor("k_65_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor k_65_cast_fp16 = slice_by_index(begin = k_65_begin_0, end = k_65_end_0, end_mask = k_65_end_mask_0, squeeze_mask = k_65_squeeze_mask_0, x = qkv_33_cast_fp16)[name = tensor("k_65_cast_fp16")]; + tensor v_33_begin_0 = const()[name = tensor("v_33_begin_0"), val = tensor([0, 0, 2, 0, 0])]; + tensor v_33_end_0 = const()[name = tensor("v_33_end_0"), val = tensor([1, 1, 3, 16, 64])]; + tensor v_33_end_mask_0 = const()[name = tensor("v_33_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor v_33_squeeze_mask_0 = const()[name = tensor("v_33_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor v_33_cast_fp16 = slice_by_index(begin = v_33_begin_0, end = v_33_end_0, end_mask = v_33_end_mask_0, squeeze_mask = v_33_squeeze_mask_0, x = qkv_33_cast_fp16)[name = tensor("v_33_cast_fp16")]; + tensor var_6460 = const()[name = tensor("op_6460"), val = tensor([1, 1, 1, 1])]; + tensor position16_to_fp16_dtype_0 = const()[name = tensor("position16_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor position16_to_fp16 = cast(dtype = position16_to_fp16_dtype_0, x = position16)[name = tensor("cast_473")]; + tensor ts_101_cast_fp16 = reshape(shape = var_6460, x = position16_to_fp16)[name = tensor("ts_101_cast_fp16")]; + tensor var_6467 = const()[name = tensor("op_6467"), val = tensor([1, 1, 16, 32, 2])]; + tensor q_complex_33_cast_fp16 = reshape(shape = var_6467, x = q_97_cast_fp16)[name = tensor("q_complex_33_cast_fp16")]; + tensor var_6474 = const()[name = tensor("op_6474"), val = tensor([1, 1, 16, 32, 2])]; + tensor k_complex_33_cast_fp16 = reshape(shape = var_6474, x = k_65_cast_fp16)[name = tensor("k_complex_33_cast_fp16")]; + tensor var_6478_begin_0 = const()[name = tensor("op_6478_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_6478_end_0 = const()[name = tensor("op_6478_end_0"), val = tensor([1, 1, 16, 32, 1])]; + tensor var_6478_end_mask_0 = const()[name = tensor("op_6478_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_6478_squeeze_mask_0 = const()[name = tensor("op_6478_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_6478_cast_fp16 = slice_by_index(begin = var_6478_begin_0, end = var_6478_end_0, end_mask = var_6478_end_mask_0, squeeze_mask = var_6478_squeeze_mask_0, x = q_complex_33_cast_fp16)[name = tensor("op_6478_cast_fp16")]; + tensor var_6486_begin_0 = const()[name = tensor("op_6486_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_6486_end_0 = const()[name = tensor("op_6486_end_0"), val = tensor([1, 1, 16, 32, 2])]; + tensor var_6486_end_mask_0 = const()[name = tensor("op_6486_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_6486_squeeze_mask_0 = const()[name = tensor("op_6486_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_6486_cast_fp16 = slice_by_index(begin = var_6486_begin_0, end = var_6486_end_0, end_mask = var_6486_end_mask_0, squeeze_mask = var_6486_squeeze_mask_0, x = q_complex_33_cast_fp16)[name = tensor("op_6486_cast_fp16")]; + tensor var_6494_begin_0 = const()[name = tensor("op_6494_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_6494_end_0 = const()[name = tensor("op_6494_end_0"), val = tensor([1, 1, 16, 32, 1])]; + tensor var_6494_end_mask_0 = const()[name = tensor("op_6494_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_6494_squeeze_mask_0 = const()[name = tensor("op_6494_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_6494_cast_fp16 = slice_by_index(begin = var_6494_begin_0, end = var_6494_end_0, end_mask = var_6494_end_mask_0, squeeze_mask = var_6494_squeeze_mask_0, x = k_complex_33_cast_fp16)[name = tensor("op_6494_cast_fp16")]; + tensor var_6502_begin_0 = const()[name = tensor("op_6502_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_6502_end_0 = const()[name = tensor("op_6502_end_0"), val = tensor([1, 1, 16, 32, 2])]; + tensor var_6502_end_mask_0 = const()[name = tensor("op_6502_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_6502_squeeze_mask_0 = const()[name = tensor("op_6502_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_6502_cast_fp16 = slice_by_index(begin = var_6502_begin_0, end = var_6502_end_0, end_mask = var_6502_end_mask_0, squeeze_mask = var_6502_squeeze_mask_0, x = k_complex_33_cast_fp16)[name = tensor("op_6502_cast_fp16")]; + tensor freqs_33_to_fp16 = const()[name = tensor("freqs_33_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(410222208)))]; + tensor var_6508_cast_fp16 = mul(x = freqs_33_to_fp16, y = ts_101_cast_fp16)[name = tensor("op_6508_cast_fp16")]; + tensor rotr_33_cast_fp16 = cos(x = var_6508_cast_fp16)[name = tensor("rotr_33_cast_fp16")]; + tensor roti_33_cast_fp16 = sin(x = var_6508_cast_fp16)[name = tensor("roti_33_cast_fp16")]; + tensor var_6512_cast_fp16 = mul(x = var_6478_cast_fp16, y = rotr_33_cast_fp16)[name = tensor("op_6512_cast_fp16")]; + tensor var_6513_cast_fp16 = mul(x = var_6486_cast_fp16, y = roti_33_cast_fp16)[name = tensor("op_6513_cast_fp16")]; + tensor qor_65_cast_fp16 = sub(x = var_6512_cast_fp16, y = var_6513_cast_fp16)[name = tensor("qor_65_cast_fp16")]; + tensor var_6516_cast_fp16 = mul(x = var_6478_cast_fp16, y = roti_33_cast_fp16)[name = tensor("op_6516_cast_fp16")]; + tensor var_6517_cast_fp16 = mul(x = var_6486_cast_fp16, y = rotr_33_cast_fp16)[name = tensor("op_6517_cast_fp16")]; + tensor qoi_65_cast_fp16 = add(x = var_6516_cast_fp16, y = var_6517_cast_fp16)[name = tensor("qoi_65_cast_fp16")]; + tensor var_6520_cast_fp16 = mul(x = var_6494_cast_fp16, y = rotr_33_cast_fp16)[name = tensor("op_6520_cast_fp16")]; + tensor var_6521_cast_fp16 = mul(x = var_6502_cast_fp16, y = roti_33_cast_fp16)[name = tensor("op_6521_cast_fp16")]; + tensor kor_65_cast_fp16 = sub(x = var_6520_cast_fp16, y = var_6521_cast_fp16)[name = tensor("kor_65_cast_fp16")]; + tensor var_6524_cast_fp16 = mul(x = var_6494_cast_fp16, y = roti_33_cast_fp16)[name = tensor("op_6524_cast_fp16")]; + tensor var_6525_cast_fp16 = mul(x = var_6502_cast_fp16, y = rotr_33_cast_fp16)[name = tensor("op_6525_cast_fp16")]; + tensor koi_65_cast_fp16 = add(x = var_6524_cast_fp16, y = var_6525_cast_fp16)[name = tensor("koi_65_cast_fp16")]; + tensor qo_33_axis_0 = const()[name = tensor("qo_33_axis_0"), val = tensor(-1)]; + tensor qo_33_cast_fp16 = stack(axis = qo_33_axis_0, values = (qor_65_cast_fp16, qoi_65_cast_fp16))[name = tensor("qo_33_cast_fp16")]; + tensor ko_33_axis_0 = const()[name = tensor("ko_33_axis_0"), val = tensor(-1)]; + tensor ko_33_cast_fp16 = stack(axis = ko_33_axis_0, values = (kor_65_cast_fp16, koi_65_cast_fp16))[name = tensor("ko_33_cast_fp16")]; + tensor var_6558 = const()[name = tensor("op_6558"), val = tensor([1, 1, 16, 64])]; + tensor q_99_cast_fp16 = reshape(shape = var_6558, x = qo_33_cast_fp16)[name = tensor("q_99_cast_fp16")]; + tensor var_6564 = const()[name = tensor("op_6564"), val = tensor([1, 1, 16, 64])]; + tensor k_67_cast_fp16 = reshape(shape = var_6564, x = ko_33_cast_fp16)[name = tensor("k_67_cast_fp16")]; + tensor _inversed_6589_y_0_to_fp16 = const()[name = tensor("_inversed_6589_y_0_to_fp16"), val = tensor(0x1p-9)]; + tensor _inversed_6589_cast_fp16 = mul(x = ts_101_cast_fp16, y = _inversed_6589_y_0_to_fp16)[name = tensor("_inversed_6589_cast_fp16")]; + tensor var_6590_cast_fp16 = floor(x = _inversed_6589_cast_fp16)[name = tensor("op_6590_cast_fp16")]; + tensor var_6591_to_fp16 = const()[name = tensor("op_6591_to_fp16"), val = tensor(0x1p+9)]; + tensor var_6592_cast_fp16 = mul(x = var_6590_cast_fp16, y = var_6591_to_fp16)[name = tensor("op_6592_cast_fp16")]; + tensor write_indices_float_67_cast_fp16 = sub(x = ts_101_cast_fp16, y = var_6592_cast_fp16)[name = tensor("write_indices_float_67_cast_fp16")]; + tensor var_6599_dtype_0 = const()[name = tensor("op_6599_dtype_0"), val = tensor("int32")]; + tensor write_indices_33_reps_0 = const()[name = tensor("write_indices_33_reps_0"), val = tensor([1, 1, 16, 64])]; + tensor write_indices_float_67_cast_fp16_to_int32 = cast(dtype = var_6599_dtype_0, x = write_indices_float_67_cast_fp16)[name = tensor("cast_472")]; + tensor write_indices_33 = tile(reps = write_indices_33_reps_0, x = write_indices_float_67_cast_fp16_to_int32)[name = tensor("write_indices_33")]; + tensor var_6609_begin_0 = const()[name = tensor("op_6609_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_6609_end_0 = const()[name = tensor("op_6609_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor var_6609_end_mask_0 = const()[name = tensor("op_6609_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_6609_squeeze_mask_0 = const()[name = tensor("op_6609_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor cache16_to_fp16_dtype_0 = const()[name = tensor("cache16_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor cache16_to_fp16 = cast(dtype = cache16_to_fp16_dtype_0, x = cache16)[name = tensor("cast_471")]; + tensor var_6609_cast_fp16 = slice_by_index(begin = var_6609_begin_0, end = var_6609_end_0, end_mask = var_6609_end_mask_0, squeeze_mask = var_6609_squeeze_mask_0, x = cache16_to_fp16)[name = tensor("op_6609_cast_fp16")]; + tensor var_6611_axis_0 = const()[name = tensor("op_6611_axis_0"), val = tensor(1)]; + tensor var_6611_mode_0 = const()[name = tensor("op_6611_mode_0"), val = tensor("update")]; + tensor var_6611_validate_indices_0 = const()[name = tensor("op_6611_validate_indices_0"), val = tensor(false)]; + tensor var_6611_cast_fp16 = scatter_along_axis(axis = var_6611_axis_0, data = var_6609_cast_fp16, indices = write_indices_33, mode = var_6611_mode_0, updates = k_67_cast_fp16, validate_indices = var_6611_validate_indices_0)[name = tensor("op_6611_cast_fp16")]; + tensor concat_114 = const()[name = tensor("concat_114"), val = tensor([0, 0, 0, 0, 0])]; + tensor concat_115 = const()[name = tensor("concat_115"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_33_internal_tensor_assign_1_stride_0 = const()[name = tensor("new_cache_33_internal_tensor_assign_1_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_33_internal_tensor_assign_1_begin_mask_0 = const()[name = tensor("new_cache_33_internal_tensor_assign_1_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_33_internal_tensor_assign_1_end_mask_0 = const()[name = tensor("new_cache_33_internal_tensor_assign_1_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_33_internal_tensor_assign_1_squeeze_mask_0 = const()[name = tensor("new_cache_33_internal_tensor_assign_1_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_80 = const()[name = tensor("shape_80"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_32 = const()[name = tensor("reduce_prod_32"), val = tensor(1048576)]; + tensor range_1d_32_start_0 = const()[name = tensor("range_1d_32_start_0"), val = tensor(0)]; + tensor range_1d_32_step_0 = const()[name = tensor("range_1d_32_step_0"), val = tensor(1)]; + tensor range_1d_32 = range_1d(end = reduce_prod_32, start = range_1d_32_start_0, step = range_1d_32_step_0)[name = tensor("range_1d_32")]; + tensor reshape_160 = reshape(shape = shape_80, x = range_1d_32)[name = tensor("reshape_160")]; + tensor slice_by_index_32 = slice_by_index(begin = concat_114, begin_mask = new_cache_33_internal_tensor_assign_1_begin_mask_0, end = concat_115, end_mask = new_cache_33_internal_tensor_assign_1_end_mask_0, squeeze_mask = new_cache_33_internal_tensor_assign_1_squeeze_mask_0, stride = new_cache_33_internal_tensor_assign_1_stride_0, x = reshape_160)[name = tensor("slice_by_index_32")]; + tensor reshape_161_shape_0 = const()[name = tensor("reshape_161_shape_0"), val = tensor([-1])]; + tensor reshape_161 = reshape(shape = reshape_161_shape_0, x = slice_by_index_32)[name = tensor("reshape_161")]; + tensor reshape_162_shape_0 = const()[name = tensor("reshape_162_shape_0"), val = tensor([-1])]; + tensor reshape_162_cast_fp16 = reshape(shape = reshape_162_shape_0, x = var_6611_cast_fp16)[name = tensor("reshape_162_cast_fp16")]; + tensor reshape_163_shape_0 = const()[name = tensor("reshape_163_shape_0"), val = tensor([-1])]; + tensor reshape_163_cast_fp16 = reshape(shape = reshape_163_shape_0, x = cache16_to_fp16)[name = tensor("reshape_163_cast_fp16")]; + tensor scatter_32_mode_0 = const()[name = tensor("scatter_32_mode_0"), val = tensor("update")]; + tensor scatter_32_axis_0 = const()[name = tensor("scatter_32_axis_0"), val = tensor(0)]; + tensor scatter_32_validate_indices_0 = const()[name = tensor("scatter_32_validate_indices_0"), val = tensor(false)]; + tensor scatter_32_cast_fp16 = scatter(axis = scatter_32_axis_0, data = reshape_163_cast_fp16, indices = reshape_161, mode = scatter_32_mode_0, updates = reshape_162_cast_fp16, validate_indices = scatter_32_validate_indices_0)[name = tensor("scatter_32_cast_fp16")]; + tensor reshape_164_cast_fp16 = reshape(shape = shape_80, x = scatter_32_cast_fp16)[name = tensor("reshape_164_cast_fp16")]; + tensor var_6619_begin_0 = const()[name = tensor("op_6619_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor var_6619_end_0 = const()[name = tensor("op_6619_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor var_6619_end_mask_0 = const()[name = tensor("op_6619_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_6619_squeeze_mask_0 = const()[name = tensor("op_6619_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor var_6619_cast_fp16 = slice_by_index(begin = var_6619_begin_0, end = var_6619_end_0, end_mask = var_6619_end_mask_0, squeeze_mask = var_6619_squeeze_mask_0, x = reshape_164_cast_fp16)[name = tensor("op_6619_cast_fp16")]; + tensor var_6621_axis_0 = const()[name = tensor("op_6621_axis_0"), val = tensor(1)]; + tensor var_6621_mode_0 = const()[name = tensor("op_6621_mode_0"), val = tensor("update")]; + tensor var_6621_validate_indices_0 = const()[name = tensor("op_6621_validate_indices_0"), val = tensor(false)]; + tensor var_6621_cast_fp16 = scatter_along_axis(axis = var_6621_axis_0, data = var_6619_cast_fp16, indices = write_indices_33, mode = var_6621_mode_0, updates = v_33_cast_fp16, validate_indices = var_6621_validate_indices_0)[name = tensor("op_6621_cast_fp16")]; + tensor concat_116 = const()[name = tensor("concat_116"), val = tensor([1, 0, 0, 0, 0])]; + tensor concat_117 = const()[name = tensor("concat_117"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_33_internal_tensor_assign_2_stride_0 = const()[name = tensor("new_cache_33_internal_tensor_assign_2_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_33_internal_tensor_assign_2_begin_mask_0 = const()[name = tensor("new_cache_33_internal_tensor_assign_2_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_33_internal_tensor_assign_2_end_mask_0 = const()[name = tensor("new_cache_33_internal_tensor_assign_2_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_33_internal_tensor_assign_2_squeeze_mask_0 = const()[name = tensor("new_cache_33_internal_tensor_assign_2_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_81 = const()[name = tensor("shape_81"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_33 = const()[name = tensor("reduce_prod_33"), val = tensor(1048576)]; + tensor range_1d_33_start_0 = const()[name = tensor("range_1d_33_start_0"), val = tensor(0)]; + tensor range_1d_33_step_0 = const()[name = tensor("range_1d_33_step_0"), val = tensor(1)]; + tensor range_1d_33 = range_1d(end = reduce_prod_33, start = range_1d_33_start_0, step = range_1d_33_step_0)[name = tensor("range_1d_33")]; + tensor reshape_165 = reshape(shape = shape_81, x = range_1d_33)[name = tensor("reshape_165")]; + tensor slice_by_index_33 = slice_by_index(begin = concat_116, begin_mask = new_cache_33_internal_tensor_assign_2_begin_mask_0, end = concat_117, end_mask = new_cache_33_internal_tensor_assign_2_end_mask_0, squeeze_mask = new_cache_33_internal_tensor_assign_2_squeeze_mask_0, stride = new_cache_33_internal_tensor_assign_2_stride_0, x = reshape_165)[name = tensor("slice_by_index_33")]; + tensor reshape_166_shape_0 = const()[name = tensor("reshape_166_shape_0"), val = tensor([-1])]; + tensor reshape_166 = reshape(shape = reshape_166_shape_0, x = slice_by_index_33)[name = tensor("reshape_166")]; + tensor reshape_167_shape_0 = const()[name = tensor("reshape_167_shape_0"), val = tensor([-1])]; + tensor reshape_167_cast_fp16 = reshape(shape = reshape_167_shape_0, x = var_6621_cast_fp16)[name = tensor("reshape_167_cast_fp16")]; + tensor reshape_168_shape_0 = const()[name = tensor("reshape_168_shape_0"), val = tensor([-1])]; + tensor reshape_168_cast_fp16 = reshape(shape = reshape_168_shape_0, x = reshape_164_cast_fp16)[name = tensor("reshape_168_cast_fp16")]; + tensor scatter_33_mode_0 = const()[name = tensor("scatter_33_mode_0"), val = tensor("update")]; + tensor scatter_33_axis_0 = const()[name = tensor("scatter_33_axis_0"), val = tensor(0)]; + tensor scatter_33_validate_indices_0 = const()[name = tensor("scatter_33_validate_indices_0"), val = tensor(false)]; + tensor scatter_33_cast_fp16 = scatter(axis = scatter_33_axis_0, data = reshape_168_cast_fp16, indices = reshape_166, mode = scatter_33_mode_0, updates = reshape_167_cast_fp16, validate_indices = scatter_33_validate_indices_0)[name = tensor("scatter_33_cast_fp16")]; + tensor reshape_169_cast_fp16 = reshape(shape = shape_81, x = scatter_33_cast_fp16)[name = tensor("reshape_169_cast_fp16")]; + tensor reshape_169_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("reshape_169_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor keys_97_begin_0 = const()[name = tensor("keys_97_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor keys_97_end_0 = const()[name = tensor("keys_97_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor keys_97_end_mask_0 = const()[name = tensor("keys_97_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor keys_97_squeeze_mask_0 = const()[name = tensor("keys_97_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor keys_97_cast_fp16 = slice_by_index(begin = keys_97_begin_0, end = keys_97_end_0, end_mask = keys_97_end_mask_0, squeeze_mask = keys_97_squeeze_mask_0, x = reshape_169_cast_fp16)[name = tensor("keys_97_cast_fp16")]; + tensor values_97_begin_0 = const()[name = tensor("values_97_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor values_97_end_0 = const()[name = tensor("values_97_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor values_97_end_mask_0 = const()[name = tensor("values_97_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor values_97_squeeze_mask_0 = const()[name = tensor("values_97_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor values_97_cast_fp16 = slice_by_index(begin = values_97_begin_0, end = values_97_end_0, end_mask = values_97_end_mask_0, squeeze_mask = values_97_squeeze_mask_0, x = reshape_169_cast_fp16)[name = tensor("values_97_cast_fp16")]; + tensor var_6633_cast_fp16 = not_equal(x = keys_97_cast_fp16, y = keys_97_cast_fp16)[name = tensor("op_6633_cast_fp16")]; + tensor keys_99_cast_fp16 = select(a = var_463_to_fp16, b = keys_97_cast_fp16, cond = var_6633_cast_fp16)[name = tensor("keys_99_cast_fp16")]; + tensor var_6641_cast_fp16 = not_equal(x = values_97_cast_fp16, y = values_97_cast_fp16)[name = tensor("op_6641_cast_fp16")]; + tensor values_99_cast_fp16 = select(a = var_463_to_fp16, b = values_97_cast_fp16, cond = var_6641_cast_fp16)[name = tensor("values_99_cast_fp16")]; + tensor var_6665 = const()[name = tensor("op_6665"), val = tensor([0, 2, 1, 3])]; + tensor var_6681 = const()[name = tensor("op_6681"), val = tensor([1, 1, 1])]; + tensor var_6682_cast_fp16 = reshape(shape = var_6681, x = position16_to_fp16)[name = tensor("op_6682_cast_fp16")]; + tensor var_6702_to_fp16 = const()[name = tensor("op_6702_to_fp16"), val = tensor(0x1p+0)]; + tensor valid_len_33_cast_fp16 = add(x = var_6682_cast_fp16, y = var_6702_to_fp16)[name = tensor("valid_len_33_cast_fp16")]; + tensor valid_mask_33_cast_fp16 = less(x = k_positions_1_promoted_to_fp16, y = valid_len_33_cast_fp16)[name = tensor("valid_mask_33_cast_fp16")]; + tensor causal_mask_33_cast_fp16 = less_equal(x = k_positions_1_promoted_to_fp16, y = var_6682_cast_fp16)[name = tensor("causal_mask_33_cast_fp16")]; + tensor attn_mask_65 = logical_and(x = valid_mask_33_cast_fp16, y = causal_mask_33_cast_fp16)[name = tensor("attn_mask_65")]; + tensor attn_mask_67_axes_0 = const()[name = tensor("attn_mask_67_axes_0"), val = tensor([1])]; + tensor attn_mask_67 = expand_dims(axes = attn_mask_67_axes_0, x = attn_mask_65)[name = tensor("attn_mask_67")]; + tensor var_6720_transpose_x_0 = const()[name = tensor("op_6720_transpose_x_0"), val = tensor(false)]; + tensor var_6720_transpose_y_0 = const()[name = tensor("op_6720_transpose_y_0"), val = tensor(false)]; + tensor transpose_104_perm_0 = const()[name = tensor("transpose_104_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_105_perm_0 = const()[name = tensor("transpose_105_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_105 = transpose(perm = transpose_105_perm_0, x = keys_99_cast_fp16)[name = tensor("transpose_149")]; + tensor transpose_104 = transpose(perm = transpose_104_perm_0, x = q_99_cast_fp16)[name = tensor("transpose_150")]; + tensor var_6720_cast_fp16 = matmul(transpose_x = var_6720_transpose_x_0, transpose_y = var_6720_transpose_y_0, x = transpose_104, y = transpose_105)[name = tensor("op_6720_cast_fp16")]; + tensor var_6714_to_fp16 = const()[name = tensor("op_6714_to_fp16"), val = tensor([0x1p-3])]; + tensor attn_weights_97_cast_fp16 = mul(x = var_6720_cast_fp16, y = var_6714_to_fp16)[name = tensor("attn_weights_97_cast_fp16")]; + tensor var_6722 = logical_not(x = attn_mask_67)[name = tensor("op_6722")]; + tensor var_6723_to_fp16 = const()[name = tensor("op_6723_to_fp16"), val = tensor(-inf)]; + tensor attn_weights_99_cast_fp16 = select(a = var_6723_to_fp16, b = attn_weights_97_cast_fp16, cond = var_6722)[name = tensor("attn_weights_99_cast_fp16")]; + tensor var_6725 = const()[name = tensor("op_6725"), val = tensor(-1)]; + tensor attn_weights_101_cast_fp16 = softmax(axis = var_6725, x = attn_weights_99_cast_fp16)[name = tensor("attn_weights_101_cast_fp16")]; + tensor attn_output_33_transpose_x_0 = const()[name = tensor("attn_output_33_transpose_x_0"), val = tensor(false)]; + tensor attn_output_33_transpose_y_0 = const()[name = tensor("attn_output_33_transpose_y_0"), val = tensor(false)]; + tensor values_101_cast_fp16 = transpose(perm = var_6665, x = values_99_cast_fp16)[name = tensor("transpose_151")]; + tensor attn_output_33_cast_fp16 = matmul(transpose_x = attn_output_33_transpose_x_0, transpose_y = attn_output_33_transpose_y_0, x = attn_weights_101_cast_fp16, y = values_101_cast_fp16)[name = tensor("attn_output_33_cast_fp16")]; + tensor var_6733 = const()[name = tensor("op_6733"), val = tensor([0, 2, 1, 3])]; + tensor var_6738 = const()[name = tensor("op_6738"), val = tensor([1, 1, 1024])]; + tensor var_6734_cast_fp16 = transpose(perm = var_6733, x = attn_output_33_cast_fp16)[name = tensor("transpose_148")]; + tensor input_199_cast_fp16 = reshape(shape = var_6738, x = var_6734_cast_fp16)[name = tensor("input_199_cast_fp16")]; + tensor attn16_out_proj_weight_to_fp16 = const()[name = tensor("attn16_out_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(410222336)))]; + tensor linear_66_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = attn16_out_proj_weight_to_fp16, x = input_199_cast_fp16)[name = tensor("linear_66_cast_fp16")]; + tensor var_6744_to_fp16 = const()[name = tensor("op_6744_to_fp16"), val = tensor(0x1p+0)]; + tensor var_6745_cast_fp16 = add(x = position16_to_fp16, y = var_6744_to_fp16)[name = tensor("op_6745_cast_fp16")]; + tensor var_6745_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_6745_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor input_201_cast_fp16 = add(x = input_195_cast_fp16, y = linear_66_cast_fp16)[name = tensor("input_201_cast_fp16")]; + tensor input_203_axes_0 = const()[name = tensor("input_203_axes_0"), val = tensor([-1])]; + tensor norm16_2_weight_to_fp16 = const()[name = tensor("norm16_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(412319552)))]; + tensor norm16_2_bias_to_fp16 = const()[name = tensor("norm16_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(412321664)))]; + tensor var_6749_to_fp16 = const()[name = tensor("op_6749_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_203_cast_fp16 = layer_norm(axes = input_203_axes_0, beta = norm16_2_bias_to_fp16, epsilon = var_6749_to_fp16, gamma = norm16_2_weight_to_fp16, x = input_201_cast_fp16)[name = tensor("input_203_cast_fp16")]; + tensor linear16_1_weight_to_fp16 = const()[name = tensor("linear16_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(412323776)))]; + tensor linear_67_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = linear16_1_weight_to_fp16, x = input_203_cast_fp16)[name = tensor("linear_67_cast_fp16")]; + tensor input_205_mode_0 = const()[name = tensor("input_205_mode_0"), val = tensor("EXACT")]; + tensor input_205_cast_fp16 = gelu(mode = input_205_mode_0, x = linear_67_cast_fp16)[name = tensor("input_205_cast_fp16")]; + tensor linear16_2_weight_to_fp16 = const()[name = tensor("linear16_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(420712448)))]; + tensor linear_68_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = linear16_2_weight_to_fp16, x = input_205_cast_fp16)[name = tensor("linear_68_cast_fp16")]; + tensor input_207_cast_fp16 = add(x = input_201_cast_fp16, y = linear_68_cast_fp16)[name = tensor("input_207_cast_fp16")]; + tensor input_209_axes_0 = const()[name = tensor("input_209_axes_0"), val = tensor([-1])]; + tensor norm17_1_weight_to_fp16 = const()[name = tensor("norm17_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(429101120)))]; + tensor norm17_1_bias_to_fp16 = const()[name = tensor("norm17_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(429103232)))]; + tensor var_6766_to_fp16 = const()[name = tensor("op_6766_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_209_cast_fp16 = layer_norm(axes = input_209_axes_0, beta = norm17_1_bias_to_fp16, epsilon = var_6766_to_fp16, gamma = norm17_1_weight_to_fp16, x = input_207_cast_fp16)[name = tensor("input_209_cast_fp16")]; + tensor attn17_in_proj_weight_to_fp16 = const()[name = tensor("attn17_in_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(429105344)))]; + tensor linear_69_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = attn17_in_proj_weight_to_fp16, x = input_209_cast_fp16)[name = tensor("linear_69_cast_fp16")]; + tensor var_6780 = const()[name = tensor("op_6780"), val = tensor([1, 1, 3, 16, 64])]; + tensor qkv_35_cast_fp16 = reshape(shape = var_6780, x = linear_69_cast_fp16)[name = tensor("qkv_35_cast_fp16")]; + tensor q_103_begin_0 = const()[name = tensor("q_103_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor q_103_end_0 = const()[name = tensor("q_103_end_0"), val = tensor([1, 1, 1, 16, 64])]; + tensor q_103_end_mask_0 = const()[name = tensor("q_103_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor q_103_squeeze_mask_0 = const()[name = tensor("q_103_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor q_103_cast_fp16 = slice_by_index(begin = q_103_begin_0, end = q_103_end_0, end_mask = q_103_end_mask_0, squeeze_mask = q_103_squeeze_mask_0, x = qkv_35_cast_fp16)[name = tensor("q_103_cast_fp16")]; + tensor k_69_begin_0 = const()[name = tensor("k_69_begin_0"), val = tensor([0, 0, 1, 0, 0])]; + tensor k_69_end_0 = const()[name = tensor("k_69_end_0"), val = tensor([1, 1, 2, 16, 64])]; + tensor k_69_end_mask_0 = const()[name = tensor("k_69_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor k_69_squeeze_mask_0 = const()[name = tensor("k_69_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor k_69_cast_fp16 = slice_by_index(begin = k_69_begin_0, end = k_69_end_0, end_mask = k_69_end_mask_0, squeeze_mask = k_69_squeeze_mask_0, x = qkv_35_cast_fp16)[name = tensor("k_69_cast_fp16")]; + tensor v_35_begin_0 = const()[name = tensor("v_35_begin_0"), val = tensor([0, 0, 2, 0, 0])]; + tensor v_35_end_0 = const()[name = tensor("v_35_end_0"), val = tensor([1, 1, 3, 16, 64])]; + tensor v_35_end_mask_0 = const()[name = tensor("v_35_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor v_35_squeeze_mask_0 = const()[name = tensor("v_35_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor v_35_cast_fp16 = slice_by_index(begin = v_35_begin_0, end = v_35_end_0, end_mask = v_35_end_mask_0, squeeze_mask = v_35_squeeze_mask_0, x = qkv_35_cast_fp16)[name = tensor("v_35_cast_fp16")]; + tensor var_6846 = const()[name = tensor("op_6846"), val = tensor([1, 1, 1, 1])]; + tensor position17_to_fp16_dtype_0 = const()[name = tensor("position17_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor position17_to_fp16 = cast(dtype = position17_to_fp16_dtype_0, x = position17)[name = tensor("cast_468")]; + tensor ts_107_cast_fp16 = reshape(shape = var_6846, x = position17_to_fp16)[name = tensor("ts_107_cast_fp16")]; + tensor var_6853 = const()[name = tensor("op_6853"), val = tensor([1, 1, 16, 32, 2])]; + tensor q_complex_35_cast_fp16 = reshape(shape = var_6853, x = q_103_cast_fp16)[name = tensor("q_complex_35_cast_fp16")]; + tensor var_6860 = const()[name = tensor("op_6860"), val = tensor([1, 1, 16, 32, 2])]; + tensor k_complex_35_cast_fp16 = reshape(shape = var_6860, x = k_69_cast_fp16)[name = tensor("k_complex_35_cast_fp16")]; + tensor var_6864_begin_0 = const()[name = tensor("op_6864_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_6864_end_0 = const()[name = tensor("op_6864_end_0"), val = tensor([1, 1, 16, 32, 1])]; + tensor var_6864_end_mask_0 = const()[name = tensor("op_6864_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_6864_squeeze_mask_0 = const()[name = tensor("op_6864_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_6864_cast_fp16 = slice_by_index(begin = var_6864_begin_0, end = var_6864_end_0, end_mask = var_6864_end_mask_0, squeeze_mask = var_6864_squeeze_mask_0, x = q_complex_35_cast_fp16)[name = tensor("op_6864_cast_fp16")]; + tensor var_6872_begin_0 = const()[name = tensor("op_6872_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_6872_end_0 = const()[name = tensor("op_6872_end_0"), val = tensor([1, 1, 16, 32, 2])]; + tensor var_6872_end_mask_0 = const()[name = tensor("op_6872_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_6872_squeeze_mask_0 = const()[name = tensor("op_6872_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_6872_cast_fp16 = slice_by_index(begin = var_6872_begin_0, end = var_6872_end_0, end_mask = var_6872_end_mask_0, squeeze_mask = var_6872_squeeze_mask_0, x = q_complex_35_cast_fp16)[name = tensor("op_6872_cast_fp16")]; + tensor var_6880_begin_0 = const()[name = tensor("op_6880_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_6880_end_0 = const()[name = tensor("op_6880_end_0"), val = tensor([1, 1, 16, 32, 1])]; + tensor var_6880_end_mask_0 = const()[name = tensor("op_6880_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_6880_squeeze_mask_0 = const()[name = tensor("op_6880_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_6880_cast_fp16 = slice_by_index(begin = var_6880_begin_0, end = var_6880_end_0, end_mask = var_6880_end_mask_0, squeeze_mask = var_6880_squeeze_mask_0, x = k_complex_35_cast_fp16)[name = tensor("op_6880_cast_fp16")]; + tensor var_6888_begin_0 = const()[name = tensor("op_6888_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_6888_end_0 = const()[name = tensor("op_6888_end_0"), val = tensor([1, 1, 16, 32, 2])]; + tensor var_6888_end_mask_0 = const()[name = tensor("op_6888_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_6888_squeeze_mask_0 = const()[name = tensor("op_6888_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_6888_cast_fp16 = slice_by_index(begin = var_6888_begin_0, end = var_6888_end_0, end_mask = var_6888_end_mask_0, squeeze_mask = var_6888_squeeze_mask_0, x = k_complex_35_cast_fp16)[name = tensor("op_6888_cast_fp16")]; + tensor freqs_35_to_fp16 = const()[name = tensor("freqs_35_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(435396864)))]; + tensor var_6894_cast_fp16 = mul(x = freqs_35_to_fp16, y = ts_107_cast_fp16)[name = tensor("op_6894_cast_fp16")]; + tensor rotr_35_cast_fp16 = cos(x = var_6894_cast_fp16)[name = tensor("rotr_35_cast_fp16")]; + tensor roti_35_cast_fp16 = sin(x = var_6894_cast_fp16)[name = tensor("roti_35_cast_fp16")]; + tensor var_6898_cast_fp16 = mul(x = var_6864_cast_fp16, y = rotr_35_cast_fp16)[name = tensor("op_6898_cast_fp16")]; + tensor var_6899_cast_fp16 = mul(x = var_6872_cast_fp16, y = roti_35_cast_fp16)[name = tensor("op_6899_cast_fp16")]; + tensor qor_69_cast_fp16 = sub(x = var_6898_cast_fp16, y = var_6899_cast_fp16)[name = tensor("qor_69_cast_fp16")]; + tensor var_6902_cast_fp16 = mul(x = var_6864_cast_fp16, y = roti_35_cast_fp16)[name = tensor("op_6902_cast_fp16")]; + tensor var_6903_cast_fp16 = mul(x = var_6872_cast_fp16, y = rotr_35_cast_fp16)[name = tensor("op_6903_cast_fp16")]; + tensor qoi_69_cast_fp16 = add(x = var_6902_cast_fp16, y = var_6903_cast_fp16)[name = tensor("qoi_69_cast_fp16")]; + tensor var_6906_cast_fp16 = mul(x = var_6880_cast_fp16, y = rotr_35_cast_fp16)[name = tensor("op_6906_cast_fp16")]; + tensor var_6907_cast_fp16 = mul(x = var_6888_cast_fp16, y = roti_35_cast_fp16)[name = tensor("op_6907_cast_fp16")]; + tensor kor_69_cast_fp16 = sub(x = var_6906_cast_fp16, y = var_6907_cast_fp16)[name = tensor("kor_69_cast_fp16")]; + tensor var_6910_cast_fp16 = mul(x = var_6880_cast_fp16, y = roti_35_cast_fp16)[name = tensor("op_6910_cast_fp16")]; + tensor var_6911_cast_fp16 = mul(x = var_6888_cast_fp16, y = rotr_35_cast_fp16)[name = tensor("op_6911_cast_fp16")]; + tensor koi_69_cast_fp16 = add(x = var_6910_cast_fp16, y = var_6911_cast_fp16)[name = tensor("koi_69_cast_fp16")]; + tensor qo_35_axis_0 = const()[name = tensor("qo_35_axis_0"), val = tensor(-1)]; + tensor qo_35_cast_fp16 = stack(axis = qo_35_axis_0, values = (qor_69_cast_fp16, qoi_69_cast_fp16))[name = tensor("qo_35_cast_fp16")]; + tensor ko_35_axis_0 = const()[name = tensor("ko_35_axis_0"), val = tensor(-1)]; + tensor ko_35_cast_fp16 = stack(axis = ko_35_axis_0, values = (kor_69_cast_fp16, koi_69_cast_fp16))[name = tensor("ko_35_cast_fp16")]; + tensor var_6944 = const()[name = tensor("op_6944"), val = tensor([1, 1, 16, 64])]; + tensor q_105_cast_fp16 = reshape(shape = var_6944, x = qo_35_cast_fp16)[name = tensor("q_105_cast_fp16")]; + tensor var_6950 = const()[name = tensor("op_6950"), val = tensor([1, 1, 16, 64])]; + tensor k_71_cast_fp16 = reshape(shape = var_6950, x = ko_35_cast_fp16)[name = tensor("k_71_cast_fp16")]; + tensor _inversed_6975_y_0_to_fp16 = const()[name = tensor("_inversed_6975_y_0_to_fp16"), val = tensor(0x1p-9)]; + tensor _inversed_6975_cast_fp16 = mul(x = ts_107_cast_fp16, y = _inversed_6975_y_0_to_fp16)[name = tensor("_inversed_6975_cast_fp16")]; + tensor var_6976_cast_fp16 = floor(x = _inversed_6975_cast_fp16)[name = tensor("op_6976_cast_fp16")]; + tensor var_6977_to_fp16 = const()[name = tensor("op_6977_to_fp16"), val = tensor(0x1p+9)]; + tensor var_6978_cast_fp16 = mul(x = var_6976_cast_fp16, y = var_6977_to_fp16)[name = tensor("op_6978_cast_fp16")]; + tensor write_indices_float_71_cast_fp16 = sub(x = ts_107_cast_fp16, y = var_6978_cast_fp16)[name = tensor("write_indices_float_71_cast_fp16")]; + tensor var_6985_dtype_0 = const()[name = tensor("op_6985_dtype_0"), val = tensor("int32")]; + tensor write_indices_35_reps_0 = const()[name = tensor("write_indices_35_reps_0"), val = tensor([1, 1, 16, 64])]; + tensor write_indices_float_71_cast_fp16_to_int32 = cast(dtype = var_6985_dtype_0, x = write_indices_float_71_cast_fp16)[name = tensor("cast_467")]; + tensor write_indices_35 = tile(reps = write_indices_35_reps_0, x = write_indices_float_71_cast_fp16_to_int32)[name = tensor("write_indices_35")]; + tensor var_6995_begin_0 = const()[name = tensor("op_6995_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_6995_end_0 = const()[name = tensor("op_6995_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor var_6995_end_mask_0 = const()[name = tensor("op_6995_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_6995_squeeze_mask_0 = const()[name = tensor("op_6995_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor cache17_to_fp16_dtype_0 = const()[name = tensor("cache17_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor cache17_to_fp16 = cast(dtype = cache17_to_fp16_dtype_0, x = cache17)[name = tensor("cast_466")]; + tensor var_6995_cast_fp16 = slice_by_index(begin = var_6995_begin_0, end = var_6995_end_0, end_mask = var_6995_end_mask_0, squeeze_mask = var_6995_squeeze_mask_0, x = cache17_to_fp16)[name = tensor("op_6995_cast_fp16")]; + tensor var_6997_axis_0 = const()[name = tensor("op_6997_axis_0"), val = tensor(1)]; + tensor var_6997_mode_0 = const()[name = tensor("op_6997_mode_0"), val = tensor("update")]; + tensor var_6997_validate_indices_0 = const()[name = tensor("op_6997_validate_indices_0"), val = tensor(false)]; + tensor var_6997_cast_fp16 = scatter_along_axis(axis = var_6997_axis_0, data = var_6995_cast_fp16, indices = write_indices_35, mode = var_6997_mode_0, updates = k_71_cast_fp16, validate_indices = var_6997_validate_indices_0)[name = tensor("op_6997_cast_fp16")]; + tensor concat_121 = const()[name = tensor("concat_121"), val = tensor([0, 0, 0, 0, 0])]; + tensor concat_122 = const()[name = tensor("concat_122"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_35_internal_tensor_assign_1_stride_0 = const()[name = tensor("new_cache_35_internal_tensor_assign_1_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_35_internal_tensor_assign_1_begin_mask_0 = const()[name = tensor("new_cache_35_internal_tensor_assign_1_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_35_internal_tensor_assign_1_end_mask_0 = const()[name = tensor("new_cache_35_internal_tensor_assign_1_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_35_internal_tensor_assign_1_squeeze_mask_0 = const()[name = tensor("new_cache_35_internal_tensor_assign_1_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_82 = const()[name = tensor("shape_82"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_34 = const()[name = tensor("reduce_prod_34"), val = tensor(1048576)]; + tensor range_1d_34_start_0 = const()[name = tensor("range_1d_34_start_0"), val = tensor(0)]; + tensor range_1d_34_step_0 = const()[name = tensor("range_1d_34_step_0"), val = tensor(1)]; + tensor range_1d_34 = range_1d(end = reduce_prod_34, start = range_1d_34_start_0, step = range_1d_34_step_0)[name = tensor("range_1d_34")]; + tensor reshape_170 = reshape(shape = shape_82, x = range_1d_34)[name = tensor("reshape_170")]; + tensor slice_by_index_34 = slice_by_index(begin = concat_121, begin_mask = new_cache_35_internal_tensor_assign_1_begin_mask_0, end = concat_122, end_mask = new_cache_35_internal_tensor_assign_1_end_mask_0, squeeze_mask = new_cache_35_internal_tensor_assign_1_squeeze_mask_0, stride = new_cache_35_internal_tensor_assign_1_stride_0, x = reshape_170)[name = tensor("slice_by_index_34")]; + tensor reshape_171_shape_0 = const()[name = tensor("reshape_171_shape_0"), val = tensor([-1])]; + tensor reshape_171 = reshape(shape = reshape_171_shape_0, x = slice_by_index_34)[name = tensor("reshape_171")]; + tensor reshape_172_shape_0 = const()[name = tensor("reshape_172_shape_0"), val = tensor([-1])]; + tensor reshape_172_cast_fp16 = reshape(shape = reshape_172_shape_0, x = var_6997_cast_fp16)[name = tensor("reshape_172_cast_fp16")]; + tensor reshape_173_shape_0 = const()[name = tensor("reshape_173_shape_0"), val = tensor([-1])]; + tensor reshape_173_cast_fp16 = reshape(shape = reshape_173_shape_0, x = cache17_to_fp16)[name = tensor("reshape_173_cast_fp16")]; + tensor scatter_34_mode_0 = const()[name = tensor("scatter_34_mode_0"), val = tensor("update")]; + tensor scatter_34_axis_0 = const()[name = tensor("scatter_34_axis_0"), val = tensor(0)]; + tensor scatter_34_validate_indices_0 = const()[name = tensor("scatter_34_validate_indices_0"), val = tensor(false)]; + tensor scatter_34_cast_fp16 = scatter(axis = scatter_34_axis_0, data = reshape_173_cast_fp16, indices = reshape_171, mode = scatter_34_mode_0, updates = reshape_172_cast_fp16, validate_indices = scatter_34_validate_indices_0)[name = tensor("scatter_34_cast_fp16")]; + tensor reshape_174_cast_fp16 = reshape(shape = shape_82, x = scatter_34_cast_fp16)[name = tensor("reshape_174_cast_fp16")]; + tensor var_7005_begin_0 = const()[name = tensor("op_7005_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor var_7005_end_0 = const()[name = tensor("op_7005_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor var_7005_end_mask_0 = const()[name = tensor("op_7005_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_7005_squeeze_mask_0 = const()[name = tensor("op_7005_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor var_7005_cast_fp16 = slice_by_index(begin = var_7005_begin_0, end = var_7005_end_0, end_mask = var_7005_end_mask_0, squeeze_mask = var_7005_squeeze_mask_0, x = reshape_174_cast_fp16)[name = tensor("op_7005_cast_fp16")]; + tensor var_7007_axis_0 = const()[name = tensor("op_7007_axis_0"), val = tensor(1)]; + tensor var_7007_mode_0 = const()[name = tensor("op_7007_mode_0"), val = tensor("update")]; + tensor var_7007_validate_indices_0 = const()[name = tensor("op_7007_validate_indices_0"), val = tensor(false)]; + tensor var_7007_cast_fp16 = scatter_along_axis(axis = var_7007_axis_0, data = var_7005_cast_fp16, indices = write_indices_35, mode = var_7007_mode_0, updates = v_35_cast_fp16, validate_indices = var_7007_validate_indices_0)[name = tensor("op_7007_cast_fp16")]; + tensor concat_123 = const()[name = tensor("concat_123"), val = tensor([1, 0, 0, 0, 0])]; + tensor concat_124 = const()[name = tensor("concat_124"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_35_internal_tensor_assign_2_stride_0 = const()[name = tensor("new_cache_35_internal_tensor_assign_2_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_35_internal_tensor_assign_2_begin_mask_0 = const()[name = tensor("new_cache_35_internal_tensor_assign_2_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_35_internal_tensor_assign_2_end_mask_0 = const()[name = tensor("new_cache_35_internal_tensor_assign_2_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_35_internal_tensor_assign_2_squeeze_mask_0 = const()[name = tensor("new_cache_35_internal_tensor_assign_2_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_83 = const()[name = tensor("shape_83"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_35 = const()[name = tensor("reduce_prod_35"), val = tensor(1048576)]; + tensor range_1d_35_start_0 = const()[name = tensor("range_1d_35_start_0"), val = tensor(0)]; + tensor range_1d_35_step_0 = const()[name = tensor("range_1d_35_step_0"), val = tensor(1)]; + tensor range_1d_35 = range_1d(end = reduce_prod_35, start = range_1d_35_start_0, step = range_1d_35_step_0)[name = tensor("range_1d_35")]; + tensor reshape_175 = reshape(shape = shape_83, x = range_1d_35)[name = tensor("reshape_175")]; + tensor slice_by_index_35 = slice_by_index(begin = concat_123, begin_mask = new_cache_35_internal_tensor_assign_2_begin_mask_0, end = concat_124, end_mask = new_cache_35_internal_tensor_assign_2_end_mask_0, squeeze_mask = new_cache_35_internal_tensor_assign_2_squeeze_mask_0, stride = new_cache_35_internal_tensor_assign_2_stride_0, x = reshape_175)[name = tensor("slice_by_index_35")]; + tensor reshape_176_shape_0 = const()[name = tensor("reshape_176_shape_0"), val = tensor([-1])]; + tensor reshape_176 = reshape(shape = reshape_176_shape_0, x = slice_by_index_35)[name = tensor("reshape_176")]; + tensor reshape_177_shape_0 = const()[name = tensor("reshape_177_shape_0"), val = tensor([-1])]; + tensor reshape_177_cast_fp16 = reshape(shape = reshape_177_shape_0, x = var_7007_cast_fp16)[name = tensor("reshape_177_cast_fp16")]; + tensor reshape_178_shape_0 = const()[name = tensor("reshape_178_shape_0"), val = tensor([-1])]; + tensor reshape_178_cast_fp16 = reshape(shape = reshape_178_shape_0, x = reshape_174_cast_fp16)[name = tensor("reshape_178_cast_fp16")]; + tensor scatter_35_mode_0 = const()[name = tensor("scatter_35_mode_0"), val = tensor("update")]; + tensor scatter_35_axis_0 = const()[name = tensor("scatter_35_axis_0"), val = tensor(0)]; + tensor scatter_35_validate_indices_0 = const()[name = tensor("scatter_35_validate_indices_0"), val = tensor(false)]; + tensor scatter_35_cast_fp16 = scatter(axis = scatter_35_axis_0, data = reshape_178_cast_fp16, indices = reshape_176, mode = scatter_35_mode_0, updates = reshape_177_cast_fp16, validate_indices = scatter_35_validate_indices_0)[name = tensor("scatter_35_cast_fp16")]; + tensor reshape_179_cast_fp16 = reshape(shape = shape_83, x = scatter_35_cast_fp16)[name = tensor("reshape_179_cast_fp16")]; + tensor reshape_179_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("reshape_179_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor keys_103_begin_0 = const()[name = tensor("keys_103_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor keys_103_end_0 = const()[name = tensor("keys_103_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor keys_103_end_mask_0 = const()[name = tensor("keys_103_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor keys_103_squeeze_mask_0 = const()[name = tensor("keys_103_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor keys_103_cast_fp16 = slice_by_index(begin = keys_103_begin_0, end = keys_103_end_0, end_mask = keys_103_end_mask_0, squeeze_mask = keys_103_squeeze_mask_0, x = reshape_179_cast_fp16)[name = tensor("keys_103_cast_fp16")]; + tensor values_103_begin_0 = const()[name = tensor("values_103_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor values_103_end_0 = const()[name = tensor("values_103_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor values_103_end_mask_0 = const()[name = tensor("values_103_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor values_103_squeeze_mask_0 = const()[name = tensor("values_103_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor values_103_cast_fp16 = slice_by_index(begin = values_103_begin_0, end = values_103_end_0, end_mask = values_103_end_mask_0, squeeze_mask = values_103_squeeze_mask_0, x = reshape_179_cast_fp16)[name = tensor("values_103_cast_fp16")]; + tensor var_7019_cast_fp16 = not_equal(x = keys_103_cast_fp16, y = keys_103_cast_fp16)[name = tensor("op_7019_cast_fp16")]; + tensor keys_105_cast_fp16 = select(a = var_463_to_fp16, b = keys_103_cast_fp16, cond = var_7019_cast_fp16)[name = tensor("keys_105_cast_fp16")]; + tensor var_7027_cast_fp16 = not_equal(x = values_103_cast_fp16, y = values_103_cast_fp16)[name = tensor("op_7027_cast_fp16")]; + tensor values_105_cast_fp16 = select(a = var_463_to_fp16, b = values_103_cast_fp16, cond = var_7027_cast_fp16)[name = tensor("values_105_cast_fp16")]; + tensor var_7051 = const()[name = tensor("op_7051"), val = tensor([0, 2, 1, 3])]; + tensor var_7067 = const()[name = tensor("op_7067"), val = tensor([1, 1, 1])]; + tensor var_7068_cast_fp16 = reshape(shape = var_7067, x = position17_to_fp16)[name = tensor("op_7068_cast_fp16")]; + tensor var_7088_to_fp16 = const()[name = tensor("op_7088_to_fp16"), val = tensor(0x1p+0)]; + tensor valid_len_35_cast_fp16 = add(x = var_7068_cast_fp16, y = var_7088_to_fp16)[name = tensor("valid_len_35_cast_fp16")]; + tensor valid_mask_35_cast_fp16 = less(x = k_positions_1_promoted_to_fp16, y = valid_len_35_cast_fp16)[name = tensor("valid_mask_35_cast_fp16")]; + tensor causal_mask_35_cast_fp16 = less_equal(x = k_positions_1_promoted_to_fp16, y = var_7068_cast_fp16)[name = tensor("causal_mask_35_cast_fp16")]; + tensor attn_mask_69 = logical_and(x = valid_mask_35_cast_fp16, y = causal_mask_35_cast_fp16)[name = tensor("attn_mask_69")]; + tensor attn_mask_71_axes_0 = const()[name = tensor("attn_mask_71_axes_0"), val = tensor([1])]; + tensor attn_mask_71 = expand_dims(axes = attn_mask_71_axes_0, x = attn_mask_69)[name = tensor("attn_mask_71")]; + tensor var_7106_transpose_x_0 = const()[name = tensor("op_7106_transpose_x_0"), val = tensor(false)]; + tensor var_7106_transpose_y_0 = const()[name = tensor("op_7106_transpose_y_0"), val = tensor(false)]; + tensor transpose_106_perm_0 = const()[name = tensor("transpose_106_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_107_perm_0 = const()[name = tensor("transpose_107_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_107 = transpose(perm = transpose_107_perm_0, x = keys_105_cast_fp16)[name = tensor("transpose_145")]; + tensor transpose_106 = transpose(perm = transpose_106_perm_0, x = q_105_cast_fp16)[name = tensor("transpose_146")]; + tensor var_7106_cast_fp16 = matmul(transpose_x = var_7106_transpose_x_0, transpose_y = var_7106_transpose_y_0, x = transpose_106, y = transpose_107)[name = tensor("op_7106_cast_fp16")]; + tensor var_7100_to_fp16 = const()[name = tensor("op_7100_to_fp16"), val = tensor([0x1p-3])]; + tensor attn_weights_103_cast_fp16 = mul(x = var_7106_cast_fp16, y = var_7100_to_fp16)[name = tensor("attn_weights_103_cast_fp16")]; + tensor var_7108 = logical_not(x = attn_mask_71)[name = tensor("op_7108")]; + tensor var_7109_to_fp16 = const()[name = tensor("op_7109_to_fp16"), val = tensor(-inf)]; + tensor attn_weights_105_cast_fp16 = select(a = var_7109_to_fp16, b = attn_weights_103_cast_fp16, cond = var_7108)[name = tensor("attn_weights_105_cast_fp16")]; + tensor var_7111 = const()[name = tensor("op_7111"), val = tensor(-1)]; + tensor attn_weights_107_cast_fp16 = softmax(axis = var_7111, x = attn_weights_105_cast_fp16)[name = tensor("attn_weights_107_cast_fp16")]; + tensor attn_output_35_transpose_x_0 = const()[name = tensor("attn_output_35_transpose_x_0"), val = tensor(false)]; + tensor attn_output_35_transpose_y_0 = const()[name = tensor("attn_output_35_transpose_y_0"), val = tensor(false)]; + tensor values_107_cast_fp16 = transpose(perm = var_7051, x = values_105_cast_fp16)[name = tensor("transpose_147")]; + tensor attn_output_35_cast_fp16 = matmul(transpose_x = attn_output_35_transpose_x_0, transpose_y = attn_output_35_transpose_y_0, x = attn_weights_107_cast_fp16, y = values_107_cast_fp16)[name = tensor("attn_output_35_cast_fp16")]; + tensor var_7119 = const()[name = tensor("op_7119"), val = tensor([0, 2, 1, 3])]; + tensor var_7124 = const()[name = tensor("op_7124"), val = tensor([1, 1, 1024])]; + tensor var_7120_cast_fp16 = transpose(perm = var_7119, x = attn_output_35_cast_fp16)[name = tensor("transpose_144")]; + tensor input_211_cast_fp16 = reshape(shape = var_7124, x = var_7120_cast_fp16)[name = tensor("input_211_cast_fp16")]; + tensor attn17_out_proj_weight_to_fp16 = const()[name = tensor("attn17_out_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(435396992)))]; + tensor linear_70_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = attn17_out_proj_weight_to_fp16, x = input_211_cast_fp16)[name = tensor("linear_70_cast_fp16")]; + tensor var_7130_to_fp16 = const()[name = tensor("op_7130_to_fp16"), val = tensor(0x1p+0)]; + tensor var_7131_cast_fp16 = add(x = position17_to_fp16, y = var_7130_to_fp16)[name = tensor("op_7131_cast_fp16")]; + tensor var_7131_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_7131_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor input_213_cast_fp16 = add(x = input_207_cast_fp16, y = linear_70_cast_fp16)[name = tensor("input_213_cast_fp16")]; + tensor input_215_axes_0 = const()[name = tensor("input_215_axes_0"), val = tensor([-1])]; + tensor norm17_2_weight_to_fp16 = const()[name = tensor("norm17_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(437494208)))]; + tensor norm17_2_bias_to_fp16 = const()[name = tensor("norm17_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(437496320)))]; + tensor var_7135_to_fp16 = const()[name = tensor("op_7135_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_215_cast_fp16 = layer_norm(axes = input_215_axes_0, beta = norm17_2_bias_to_fp16, epsilon = var_7135_to_fp16, gamma = norm17_2_weight_to_fp16, x = input_213_cast_fp16)[name = tensor("input_215_cast_fp16")]; + tensor linear17_1_weight_to_fp16 = const()[name = tensor("linear17_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(437498432)))]; + tensor linear_71_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = linear17_1_weight_to_fp16, x = input_215_cast_fp16)[name = tensor("linear_71_cast_fp16")]; + tensor input_217_mode_0 = const()[name = tensor("input_217_mode_0"), val = tensor("EXACT")]; + tensor input_217_cast_fp16 = gelu(mode = input_217_mode_0, x = linear_71_cast_fp16)[name = tensor("input_217_cast_fp16")]; + tensor linear17_2_weight_to_fp16 = const()[name = tensor("linear17_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(445887104)))]; + tensor linear_72_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = linear17_2_weight_to_fp16, x = input_217_cast_fp16)[name = tensor("linear_72_cast_fp16")]; + tensor input_219_cast_fp16 = add(x = input_213_cast_fp16, y = linear_72_cast_fp16)[name = tensor("input_219_cast_fp16")]; + tensor input_221_axes_0 = const()[name = tensor("input_221_axes_0"), val = tensor([-1])]; + tensor norm18_1_weight_to_fp16 = const()[name = tensor("norm18_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(454275776)))]; + tensor norm18_1_bias_to_fp16 = const()[name = tensor("norm18_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(454277888)))]; + tensor var_7152_to_fp16 = const()[name = tensor("op_7152_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_221_cast_fp16 = layer_norm(axes = input_221_axes_0, beta = norm18_1_bias_to_fp16, epsilon = var_7152_to_fp16, gamma = norm18_1_weight_to_fp16, x = input_219_cast_fp16)[name = tensor("input_221_cast_fp16")]; + tensor attn18_in_proj_weight_to_fp16 = const()[name = tensor("attn18_in_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(454280000)))]; + tensor linear_73_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = attn18_in_proj_weight_to_fp16, x = input_221_cast_fp16)[name = tensor("linear_73_cast_fp16")]; + tensor var_7166 = const()[name = tensor("op_7166"), val = tensor([1, 1, 3, 16, 64])]; + tensor qkv_37_cast_fp16 = reshape(shape = var_7166, x = linear_73_cast_fp16)[name = tensor("qkv_37_cast_fp16")]; + tensor q_109_begin_0 = const()[name = tensor("q_109_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor q_109_end_0 = const()[name = tensor("q_109_end_0"), val = tensor([1, 1, 1, 16, 64])]; + tensor q_109_end_mask_0 = const()[name = tensor("q_109_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor q_109_squeeze_mask_0 = const()[name = tensor("q_109_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor q_109_cast_fp16 = slice_by_index(begin = q_109_begin_0, end = q_109_end_0, end_mask = q_109_end_mask_0, squeeze_mask = q_109_squeeze_mask_0, x = qkv_37_cast_fp16)[name = tensor("q_109_cast_fp16")]; + tensor k_73_begin_0 = const()[name = tensor("k_73_begin_0"), val = tensor([0, 0, 1, 0, 0])]; + tensor k_73_end_0 = const()[name = tensor("k_73_end_0"), val = tensor([1, 1, 2, 16, 64])]; + tensor k_73_end_mask_0 = const()[name = tensor("k_73_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor k_73_squeeze_mask_0 = const()[name = tensor("k_73_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor k_73_cast_fp16 = slice_by_index(begin = k_73_begin_0, end = k_73_end_0, end_mask = k_73_end_mask_0, squeeze_mask = k_73_squeeze_mask_0, x = qkv_37_cast_fp16)[name = tensor("k_73_cast_fp16")]; + tensor v_37_begin_0 = const()[name = tensor("v_37_begin_0"), val = tensor([0, 0, 2, 0, 0])]; + tensor v_37_end_0 = const()[name = tensor("v_37_end_0"), val = tensor([1, 1, 3, 16, 64])]; + tensor v_37_end_mask_0 = const()[name = tensor("v_37_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor v_37_squeeze_mask_0 = const()[name = tensor("v_37_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor v_37_cast_fp16 = slice_by_index(begin = v_37_begin_0, end = v_37_end_0, end_mask = v_37_end_mask_0, squeeze_mask = v_37_squeeze_mask_0, x = qkv_37_cast_fp16)[name = tensor("v_37_cast_fp16")]; + tensor var_7232 = const()[name = tensor("op_7232"), val = tensor([1, 1, 1, 1])]; + tensor position18_to_fp16_dtype_0 = const()[name = tensor("position18_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor position18_to_fp16 = cast(dtype = position18_to_fp16_dtype_0, x = position18)[name = tensor("cast_463")]; + tensor ts_113_cast_fp16 = reshape(shape = var_7232, x = position18_to_fp16)[name = tensor("ts_113_cast_fp16")]; + tensor var_7239 = const()[name = tensor("op_7239"), val = tensor([1, 1, 16, 32, 2])]; + tensor q_complex_37_cast_fp16 = reshape(shape = var_7239, x = q_109_cast_fp16)[name = tensor("q_complex_37_cast_fp16")]; + tensor var_7246 = const()[name = tensor("op_7246"), val = tensor([1, 1, 16, 32, 2])]; + tensor k_complex_37_cast_fp16 = reshape(shape = var_7246, x = k_73_cast_fp16)[name = tensor("k_complex_37_cast_fp16")]; + tensor var_7250_begin_0 = const()[name = tensor("op_7250_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_7250_end_0 = const()[name = tensor("op_7250_end_0"), val = tensor([1, 1, 16, 32, 1])]; + tensor var_7250_end_mask_0 = const()[name = tensor("op_7250_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_7250_squeeze_mask_0 = const()[name = tensor("op_7250_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_7250_cast_fp16 = slice_by_index(begin = var_7250_begin_0, end = var_7250_end_0, end_mask = var_7250_end_mask_0, squeeze_mask = var_7250_squeeze_mask_0, x = q_complex_37_cast_fp16)[name = tensor("op_7250_cast_fp16")]; + tensor var_7258_begin_0 = const()[name = tensor("op_7258_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_7258_end_0 = const()[name = tensor("op_7258_end_0"), val = tensor([1, 1, 16, 32, 2])]; + tensor var_7258_end_mask_0 = const()[name = tensor("op_7258_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_7258_squeeze_mask_0 = const()[name = tensor("op_7258_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_7258_cast_fp16 = slice_by_index(begin = var_7258_begin_0, end = var_7258_end_0, end_mask = var_7258_end_mask_0, squeeze_mask = var_7258_squeeze_mask_0, x = q_complex_37_cast_fp16)[name = tensor("op_7258_cast_fp16")]; + tensor var_7266_begin_0 = const()[name = tensor("op_7266_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_7266_end_0 = const()[name = tensor("op_7266_end_0"), val = tensor([1, 1, 16, 32, 1])]; + tensor var_7266_end_mask_0 = const()[name = tensor("op_7266_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_7266_squeeze_mask_0 = const()[name = tensor("op_7266_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_7266_cast_fp16 = slice_by_index(begin = var_7266_begin_0, end = var_7266_end_0, end_mask = var_7266_end_mask_0, squeeze_mask = var_7266_squeeze_mask_0, x = k_complex_37_cast_fp16)[name = tensor("op_7266_cast_fp16")]; + tensor var_7274_begin_0 = const()[name = tensor("op_7274_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_7274_end_0 = const()[name = tensor("op_7274_end_0"), val = tensor([1, 1, 16, 32, 2])]; + tensor var_7274_end_mask_0 = const()[name = tensor("op_7274_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_7274_squeeze_mask_0 = const()[name = tensor("op_7274_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_7274_cast_fp16 = slice_by_index(begin = var_7274_begin_0, end = var_7274_end_0, end_mask = var_7274_end_mask_0, squeeze_mask = var_7274_squeeze_mask_0, x = k_complex_37_cast_fp16)[name = tensor("op_7274_cast_fp16")]; + tensor freqs_37_to_fp16 = const()[name = tensor("freqs_37_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(460571520)))]; + tensor var_7280_cast_fp16 = mul(x = freqs_37_to_fp16, y = ts_113_cast_fp16)[name = tensor("op_7280_cast_fp16")]; + tensor rotr_37_cast_fp16 = cos(x = var_7280_cast_fp16)[name = tensor("rotr_37_cast_fp16")]; + tensor roti_37_cast_fp16 = sin(x = var_7280_cast_fp16)[name = tensor("roti_37_cast_fp16")]; + tensor var_7284_cast_fp16 = mul(x = var_7250_cast_fp16, y = rotr_37_cast_fp16)[name = tensor("op_7284_cast_fp16")]; + tensor var_7285_cast_fp16 = mul(x = var_7258_cast_fp16, y = roti_37_cast_fp16)[name = tensor("op_7285_cast_fp16")]; + tensor qor_73_cast_fp16 = sub(x = var_7284_cast_fp16, y = var_7285_cast_fp16)[name = tensor("qor_73_cast_fp16")]; + tensor var_7288_cast_fp16 = mul(x = var_7250_cast_fp16, y = roti_37_cast_fp16)[name = tensor("op_7288_cast_fp16")]; + tensor var_7289_cast_fp16 = mul(x = var_7258_cast_fp16, y = rotr_37_cast_fp16)[name = tensor("op_7289_cast_fp16")]; + tensor qoi_73_cast_fp16 = add(x = var_7288_cast_fp16, y = var_7289_cast_fp16)[name = tensor("qoi_73_cast_fp16")]; + tensor var_7292_cast_fp16 = mul(x = var_7266_cast_fp16, y = rotr_37_cast_fp16)[name = tensor("op_7292_cast_fp16")]; + tensor var_7293_cast_fp16 = mul(x = var_7274_cast_fp16, y = roti_37_cast_fp16)[name = tensor("op_7293_cast_fp16")]; + tensor kor_73_cast_fp16 = sub(x = var_7292_cast_fp16, y = var_7293_cast_fp16)[name = tensor("kor_73_cast_fp16")]; + tensor var_7296_cast_fp16 = mul(x = var_7266_cast_fp16, y = roti_37_cast_fp16)[name = tensor("op_7296_cast_fp16")]; + tensor var_7297_cast_fp16 = mul(x = var_7274_cast_fp16, y = rotr_37_cast_fp16)[name = tensor("op_7297_cast_fp16")]; + tensor koi_73_cast_fp16 = add(x = var_7296_cast_fp16, y = var_7297_cast_fp16)[name = tensor("koi_73_cast_fp16")]; + tensor qo_37_axis_0 = const()[name = tensor("qo_37_axis_0"), val = tensor(-1)]; + tensor qo_37_cast_fp16 = stack(axis = qo_37_axis_0, values = (qor_73_cast_fp16, qoi_73_cast_fp16))[name = tensor("qo_37_cast_fp16")]; + tensor ko_37_axis_0 = const()[name = tensor("ko_37_axis_0"), val = tensor(-1)]; + tensor ko_37_cast_fp16 = stack(axis = ko_37_axis_0, values = (kor_73_cast_fp16, koi_73_cast_fp16))[name = tensor("ko_37_cast_fp16")]; + tensor var_7330 = const()[name = tensor("op_7330"), val = tensor([1, 1, 16, 64])]; + tensor q_111_cast_fp16 = reshape(shape = var_7330, x = qo_37_cast_fp16)[name = tensor("q_111_cast_fp16")]; + tensor var_7336 = const()[name = tensor("op_7336"), val = tensor([1, 1, 16, 64])]; + tensor k_75_cast_fp16 = reshape(shape = var_7336, x = ko_37_cast_fp16)[name = tensor("k_75_cast_fp16")]; + tensor _inversed_7361_y_0_to_fp16 = const()[name = tensor("_inversed_7361_y_0_to_fp16"), val = tensor(0x1p-9)]; + tensor _inversed_7361_cast_fp16 = mul(x = ts_113_cast_fp16, y = _inversed_7361_y_0_to_fp16)[name = tensor("_inversed_7361_cast_fp16")]; + tensor var_7362_cast_fp16 = floor(x = _inversed_7361_cast_fp16)[name = tensor("op_7362_cast_fp16")]; + tensor var_7363_to_fp16 = const()[name = tensor("op_7363_to_fp16"), val = tensor(0x1p+9)]; + tensor var_7364_cast_fp16 = mul(x = var_7362_cast_fp16, y = var_7363_to_fp16)[name = tensor("op_7364_cast_fp16")]; + tensor write_indices_float_75_cast_fp16 = sub(x = ts_113_cast_fp16, y = var_7364_cast_fp16)[name = tensor("write_indices_float_75_cast_fp16")]; + tensor var_7371_dtype_0 = const()[name = tensor("op_7371_dtype_0"), val = tensor("int32")]; + tensor write_indices_37_reps_0 = const()[name = tensor("write_indices_37_reps_0"), val = tensor([1, 1, 16, 64])]; + tensor write_indices_float_75_cast_fp16_to_int32 = cast(dtype = var_7371_dtype_0, x = write_indices_float_75_cast_fp16)[name = tensor("cast_462")]; + tensor write_indices_37 = tile(reps = write_indices_37_reps_0, x = write_indices_float_75_cast_fp16_to_int32)[name = tensor("write_indices_37")]; + tensor var_7381_begin_0 = const()[name = tensor("op_7381_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_7381_end_0 = const()[name = tensor("op_7381_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor var_7381_end_mask_0 = const()[name = tensor("op_7381_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_7381_squeeze_mask_0 = const()[name = tensor("op_7381_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor cache18_to_fp16_dtype_0 = const()[name = tensor("cache18_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor cache18_to_fp16 = cast(dtype = cache18_to_fp16_dtype_0, x = cache18)[name = tensor("cast_461")]; + tensor var_7381_cast_fp16 = slice_by_index(begin = var_7381_begin_0, end = var_7381_end_0, end_mask = var_7381_end_mask_0, squeeze_mask = var_7381_squeeze_mask_0, x = cache18_to_fp16)[name = tensor("op_7381_cast_fp16")]; + tensor var_7383_axis_0 = const()[name = tensor("op_7383_axis_0"), val = tensor(1)]; + tensor var_7383_mode_0 = const()[name = tensor("op_7383_mode_0"), val = tensor("update")]; + tensor var_7383_validate_indices_0 = const()[name = tensor("op_7383_validate_indices_0"), val = tensor(false)]; + tensor var_7383_cast_fp16 = scatter_along_axis(axis = var_7383_axis_0, data = var_7381_cast_fp16, indices = write_indices_37, mode = var_7383_mode_0, updates = k_75_cast_fp16, validate_indices = var_7383_validate_indices_0)[name = tensor("op_7383_cast_fp16")]; + tensor concat_128 = const()[name = tensor("concat_128"), val = tensor([0, 0, 0, 0, 0])]; + tensor concat_129 = const()[name = tensor("concat_129"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_37_internal_tensor_assign_1_stride_0 = const()[name = tensor("new_cache_37_internal_tensor_assign_1_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_37_internal_tensor_assign_1_begin_mask_0 = const()[name = tensor("new_cache_37_internal_tensor_assign_1_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_37_internal_tensor_assign_1_end_mask_0 = const()[name = tensor("new_cache_37_internal_tensor_assign_1_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_37_internal_tensor_assign_1_squeeze_mask_0 = const()[name = tensor("new_cache_37_internal_tensor_assign_1_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_84 = const()[name = tensor("shape_84"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_36 = const()[name = tensor("reduce_prod_36"), val = tensor(1048576)]; + tensor range_1d_36_start_0 = const()[name = tensor("range_1d_36_start_0"), val = tensor(0)]; + tensor range_1d_36_step_0 = const()[name = tensor("range_1d_36_step_0"), val = tensor(1)]; + tensor range_1d_36 = range_1d(end = reduce_prod_36, start = range_1d_36_start_0, step = range_1d_36_step_0)[name = tensor("range_1d_36")]; + tensor reshape_180 = reshape(shape = shape_84, x = range_1d_36)[name = tensor("reshape_180")]; + tensor slice_by_index_36 = slice_by_index(begin = concat_128, begin_mask = new_cache_37_internal_tensor_assign_1_begin_mask_0, end = concat_129, end_mask = new_cache_37_internal_tensor_assign_1_end_mask_0, squeeze_mask = new_cache_37_internal_tensor_assign_1_squeeze_mask_0, stride = new_cache_37_internal_tensor_assign_1_stride_0, x = reshape_180)[name = tensor("slice_by_index_36")]; + tensor reshape_181_shape_0 = const()[name = tensor("reshape_181_shape_0"), val = tensor([-1])]; + tensor reshape_181 = reshape(shape = reshape_181_shape_0, x = slice_by_index_36)[name = tensor("reshape_181")]; + tensor reshape_182_shape_0 = const()[name = tensor("reshape_182_shape_0"), val = tensor([-1])]; + tensor reshape_182_cast_fp16 = reshape(shape = reshape_182_shape_0, x = var_7383_cast_fp16)[name = tensor("reshape_182_cast_fp16")]; + tensor reshape_183_shape_0 = const()[name = tensor("reshape_183_shape_0"), val = tensor([-1])]; + tensor reshape_183_cast_fp16 = reshape(shape = reshape_183_shape_0, x = cache18_to_fp16)[name = tensor("reshape_183_cast_fp16")]; + tensor scatter_36_mode_0 = const()[name = tensor("scatter_36_mode_0"), val = tensor("update")]; + tensor scatter_36_axis_0 = const()[name = tensor("scatter_36_axis_0"), val = tensor(0)]; + tensor scatter_36_validate_indices_0 = const()[name = tensor("scatter_36_validate_indices_0"), val = tensor(false)]; + tensor scatter_36_cast_fp16 = scatter(axis = scatter_36_axis_0, data = reshape_183_cast_fp16, indices = reshape_181, mode = scatter_36_mode_0, updates = reshape_182_cast_fp16, validate_indices = scatter_36_validate_indices_0)[name = tensor("scatter_36_cast_fp16")]; + tensor reshape_184_cast_fp16 = reshape(shape = shape_84, x = scatter_36_cast_fp16)[name = tensor("reshape_184_cast_fp16")]; + tensor var_7391_begin_0 = const()[name = tensor("op_7391_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor var_7391_end_0 = const()[name = tensor("op_7391_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor var_7391_end_mask_0 = const()[name = tensor("op_7391_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_7391_squeeze_mask_0 = const()[name = tensor("op_7391_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor var_7391_cast_fp16 = slice_by_index(begin = var_7391_begin_0, end = var_7391_end_0, end_mask = var_7391_end_mask_0, squeeze_mask = var_7391_squeeze_mask_0, x = reshape_184_cast_fp16)[name = tensor("op_7391_cast_fp16")]; + tensor var_7393_axis_0 = const()[name = tensor("op_7393_axis_0"), val = tensor(1)]; + tensor var_7393_mode_0 = const()[name = tensor("op_7393_mode_0"), val = tensor("update")]; + tensor var_7393_validate_indices_0 = const()[name = tensor("op_7393_validate_indices_0"), val = tensor(false)]; + tensor var_7393_cast_fp16 = scatter_along_axis(axis = var_7393_axis_0, data = var_7391_cast_fp16, indices = write_indices_37, mode = var_7393_mode_0, updates = v_37_cast_fp16, validate_indices = var_7393_validate_indices_0)[name = tensor("op_7393_cast_fp16")]; + tensor concat_130 = const()[name = tensor("concat_130"), val = tensor([1, 0, 0, 0, 0])]; + tensor concat_131 = const()[name = tensor("concat_131"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_37_internal_tensor_assign_2_stride_0 = const()[name = tensor("new_cache_37_internal_tensor_assign_2_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_37_internal_tensor_assign_2_begin_mask_0 = const()[name = tensor("new_cache_37_internal_tensor_assign_2_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_37_internal_tensor_assign_2_end_mask_0 = const()[name = tensor("new_cache_37_internal_tensor_assign_2_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_37_internal_tensor_assign_2_squeeze_mask_0 = const()[name = tensor("new_cache_37_internal_tensor_assign_2_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_85 = const()[name = tensor("shape_85"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_37 = const()[name = tensor("reduce_prod_37"), val = tensor(1048576)]; + tensor range_1d_37_start_0 = const()[name = tensor("range_1d_37_start_0"), val = tensor(0)]; + tensor range_1d_37_step_0 = const()[name = tensor("range_1d_37_step_0"), val = tensor(1)]; + tensor range_1d_37 = range_1d(end = reduce_prod_37, start = range_1d_37_start_0, step = range_1d_37_step_0)[name = tensor("range_1d_37")]; + tensor reshape_185 = reshape(shape = shape_85, x = range_1d_37)[name = tensor("reshape_185")]; + tensor slice_by_index_37 = slice_by_index(begin = concat_130, begin_mask = new_cache_37_internal_tensor_assign_2_begin_mask_0, end = concat_131, end_mask = new_cache_37_internal_tensor_assign_2_end_mask_0, squeeze_mask = new_cache_37_internal_tensor_assign_2_squeeze_mask_0, stride = new_cache_37_internal_tensor_assign_2_stride_0, x = reshape_185)[name = tensor("slice_by_index_37")]; + tensor reshape_186_shape_0 = const()[name = tensor("reshape_186_shape_0"), val = tensor([-1])]; + tensor reshape_186 = reshape(shape = reshape_186_shape_0, x = slice_by_index_37)[name = tensor("reshape_186")]; + tensor reshape_187_shape_0 = const()[name = tensor("reshape_187_shape_0"), val = tensor([-1])]; + tensor reshape_187_cast_fp16 = reshape(shape = reshape_187_shape_0, x = var_7393_cast_fp16)[name = tensor("reshape_187_cast_fp16")]; + tensor reshape_188_shape_0 = const()[name = tensor("reshape_188_shape_0"), val = tensor([-1])]; + tensor reshape_188_cast_fp16 = reshape(shape = reshape_188_shape_0, x = reshape_184_cast_fp16)[name = tensor("reshape_188_cast_fp16")]; + tensor scatter_37_mode_0 = const()[name = tensor("scatter_37_mode_0"), val = tensor("update")]; + tensor scatter_37_axis_0 = const()[name = tensor("scatter_37_axis_0"), val = tensor(0)]; + tensor scatter_37_validate_indices_0 = const()[name = tensor("scatter_37_validate_indices_0"), val = tensor(false)]; + tensor scatter_37_cast_fp16 = scatter(axis = scatter_37_axis_0, data = reshape_188_cast_fp16, indices = reshape_186, mode = scatter_37_mode_0, updates = reshape_187_cast_fp16, validate_indices = scatter_37_validate_indices_0)[name = tensor("scatter_37_cast_fp16")]; + tensor reshape_189_cast_fp16 = reshape(shape = shape_85, x = scatter_37_cast_fp16)[name = tensor("reshape_189_cast_fp16")]; + tensor reshape_189_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("reshape_189_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor keys_109_begin_0 = const()[name = tensor("keys_109_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor keys_109_end_0 = const()[name = tensor("keys_109_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor keys_109_end_mask_0 = const()[name = tensor("keys_109_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor keys_109_squeeze_mask_0 = const()[name = tensor("keys_109_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor keys_109_cast_fp16 = slice_by_index(begin = keys_109_begin_0, end = keys_109_end_0, end_mask = keys_109_end_mask_0, squeeze_mask = keys_109_squeeze_mask_0, x = reshape_189_cast_fp16)[name = tensor("keys_109_cast_fp16")]; + tensor values_109_begin_0 = const()[name = tensor("values_109_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor values_109_end_0 = const()[name = tensor("values_109_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor values_109_end_mask_0 = const()[name = tensor("values_109_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor values_109_squeeze_mask_0 = const()[name = tensor("values_109_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor values_109_cast_fp16 = slice_by_index(begin = values_109_begin_0, end = values_109_end_0, end_mask = values_109_end_mask_0, squeeze_mask = values_109_squeeze_mask_0, x = reshape_189_cast_fp16)[name = tensor("values_109_cast_fp16")]; + tensor var_7405_cast_fp16 = not_equal(x = keys_109_cast_fp16, y = keys_109_cast_fp16)[name = tensor("op_7405_cast_fp16")]; + tensor keys_111_cast_fp16 = select(a = var_463_to_fp16, b = keys_109_cast_fp16, cond = var_7405_cast_fp16)[name = tensor("keys_111_cast_fp16")]; + tensor var_7413_cast_fp16 = not_equal(x = values_109_cast_fp16, y = values_109_cast_fp16)[name = tensor("op_7413_cast_fp16")]; + tensor values_111_cast_fp16 = select(a = var_463_to_fp16, b = values_109_cast_fp16, cond = var_7413_cast_fp16)[name = tensor("values_111_cast_fp16")]; + tensor var_7437 = const()[name = tensor("op_7437"), val = tensor([0, 2, 1, 3])]; + tensor var_7453 = const()[name = tensor("op_7453"), val = tensor([1, 1, 1])]; + tensor var_7454_cast_fp16 = reshape(shape = var_7453, x = position18_to_fp16)[name = tensor("op_7454_cast_fp16")]; + tensor var_7474_to_fp16 = const()[name = tensor("op_7474_to_fp16"), val = tensor(0x1p+0)]; + tensor valid_len_37_cast_fp16 = add(x = var_7454_cast_fp16, y = var_7474_to_fp16)[name = tensor("valid_len_37_cast_fp16")]; + tensor valid_mask_37_cast_fp16 = less(x = k_positions_1_promoted_to_fp16, y = valid_len_37_cast_fp16)[name = tensor("valid_mask_37_cast_fp16")]; + tensor causal_mask_37_cast_fp16 = less_equal(x = k_positions_1_promoted_to_fp16, y = var_7454_cast_fp16)[name = tensor("causal_mask_37_cast_fp16")]; + tensor attn_mask_73 = logical_and(x = valid_mask_37_cast_fp16, y = causal_mask_37_cast_fp16)[name = tensor("attn_mask_73")]; + tensor attn_mask_75_axes_0 = const()[name = tensor("attn_mask_75_axes_0"), val = tensor([1])]; + tensor attn_mask_75 = expand_dims(axes = attn_mask_75_axes_0, x = attn_mask_73)[name = tensor("attn_mask_75")]; + tensor var_7492_transpose_x_0 = const()[name = tensor("op_7492_transpose_x_0"), val = tensor(false)]; + tensor var_7492_transpose_y_0 = const()[name = tensor("op_7492_transpose_y_0"), val = tensor(false)]; + tensor transpose_108_perm_0 = const()[name = tensor("transpose_108_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_109_perm_0 = const()[name = tensor("transpose_109_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_109 = transpose(perm = transpose_109_perm_0, x = keys_111_cast_fp16)[name = tensor("transpose_141")]; + tensor transpose_108 = transpose(perm = transpose_108_perm_0, x = q_111_cast_fp16)[name = tensor("transpose_142")]; + tensor var_7492_cast_fp16 = matmul(transpose_x = var_7492_transpose_x_0, transpose_y = var_7492_transpose_y_0, x = transpose_108, y = transpose_109)[name = tensor("op_7492_cast_fp16")]; + tensor var_7486_to_fp16 = const()[name = tensor("op_7486_to_fp16"), val = tensor([0x1p-3])]; + tensor attn_weights_109_cast_fp16 = mul(x = var_7492_cast_fp16, y = var_7486_to_fp16)[name = tensor("attn_weights_109_cast_fp16")]; + tensor var_7494 = logical_not(x = attn_mask_75)[name = tensor("op_7494")]; + tensor var_7495_to_fp16 = const()[name = tensor("op_7495_to_fp16"), val = tensor(-inf)]; + tensor attn_weights_111_cast_fp16 = select(a = var_7495_to_fp16, b = attn_weights_109_cast_fp16, cond = var_7494)[name = tensor("attn_weights_111_cast_fp16")]; + tensor var_7497 = const()[name = tensor("op_7497"), val = tensor(-1)]; + tensor attn_weights_113_cast_fp16 = softmax(axis = var_7497, x = attn_weights_111_cast_fp16)[name = tensor("attn_weights_113_cast_fp16")]; + tensor attn_output_37_transpose_x_0 = const()[name = tensor("attn_output_37_transpose_x_0"), val = tensor(false)]; + tensor attn_output_37_transpose_y_0 = const()[name = tensor("attn_output_37_transpose_y_0"), val = tensor(false)]; + tensor values_113_cast_fp16 = transpose(perm = var_7437, x = values_111_cast_fp16)[name = tensor("transpose_143")]; + tensor attn_output_37_cast_fp16 = matmul(transpose_x = attn_output_37_transpose_x_0, transpose_y = attn_output_37_transpose_y_0, x = attn_weights_113_cast_fp16, y = values_113_cast_fp16)[name = tensor("attn_output_37_cast_fp16")]; + tensor var_7505 = const()[name = tensor("op_7505"), val = tensor([0, 2, 1, 3])]; + tensor var_7510 = const()[name = tensor("op_7510"), val = tensor([1, 1, 1024])]; + tensor var_7506_cast_fp16 = transpose(perm = var_7505, x = attn_output_37_cast_fp16)[name = tensor("transpose_140")]; + tensor input_223_cast_fp16 = reshape(shape = var_7510, x = var_7506_cast_fp16)[name = tensor("input_223_cast_fp16")]; + tensor attn18_out_proj_weight_to_fp16 = const()[name = tensor("attn18_out_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(460571648)))]; + tensor linear_74_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = attn18_out_proj_weight_to_fp16, x = input_223_cast_fp16)[name = tensor("linear_74_cast_fp16")]; + tensor var_7516_to_fp16 = const()[name = tensor("op_7516_to_fp16"), val = tensor(0x1p+0)]; + tensor var_7517_cast_fp16 = add(x = position18_to_fp16, y = var_7516_to_fp16)[name = tensor("op_7517_cast_fp16")]; + tensor var_7517_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_7517_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor input_225_cast_fp16 = add(x = input_219_cast_fp16, y = linear_74_cast_fp16)[name = tensor("input_225_cast_fp16")]; + tensor input_227_axes_0 = const()[name = tensor("input_227_axes_0"), val = tensor([-1])]; + tensor norm18_2_weight_to_fp16 = const()[name = tensor("norm18_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(462668864)))]; + tensor norm18_2_bias_to_fp16 = const()[name = tensor("norm18_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(462670976)))]; + tensor var_7521_to_fp16 = const()[name = tensor("op_7521_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_227_cast_fp16 = layer_norm(axes = input_227_axes_0, beta = norm18_2_bias_to_fp16, epsilon = var_7521_to_fp16, gamma = norm18_2_weight_to_fp16, x = input_225_cast_fp16)[name = tensor("input_227_cast_fp16")]; + tensor linear18_1_weight_to_fp16 = const()[name = tensor("linear18_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(462673088)))]; + tensor linear_75_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = linear18_1_weight_to_fp16, x = input_227_cast_fp16)[name = tensor("linear_75_cast_fp16")]; + tensor input_229_mode_0 = const()[name = tensor("input_229_mode_0"), val = tensor("EXACT")]; + tensor input_229_cast_fp16 = gelu(mode = input_229_mode_0, x = linear_75_cast_fp16)[name = tensor("input_229_cast_fp16")]; + tensor linear18_2_weight_to_fp16 = const()[name = tensor("linear18_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(471061760)))]; + tensor linear_76_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = linear18_2_weight_to_fp16, x = input_229_cast_fp16)[name = tensor("linear_76_cast_fp16")]; + tensor input_231_cast_fp16 = add(x = input_225_cast_fp16, y = linear_76_cast_fp16)[name = tensor("input_231_cast_fp16")]; + tensor input_233_axes_0 = const()[name = tensor("input_233_axes_0"), val = tensor([-1])]; + tensor norm19_1_weight_to_fp16 = const()[name = tensor("norm19_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(479450432)))]; + tensor norm19_1_bias_to_fp16 = const()[name = tensor("norm19_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(479452544)))]; + tensor var_7538_to_fp16 = const()[name = tensor("op_7538_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_233_cast_fp16 = layer_norm(axes = input_233_axes_0, beta = norm19_1_bias_to_fp16, epsilon = var_7538_to_fp16, gamma = norm19_1_weight_to_fp16, x = input_231_cast_fp16)[name = tensor("input_233_cast_fp16")]; + tensor attn19_in_proj_weight_to_fp16 = const()[name = tensor("attn19_in_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(479454656)))]; + tensor linear_77_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = attn19_in_proj_weight_to_fp16, x = input_233_cast_fp16)[name = tensor("linear_77_cast_fp16")]; + tensor var_7552 = const()[name = tensor("op_7552"), val = tensor([1, 1, 3, 16, 64])]; + tensor qkv_39_cast_fp16 = reshape(shape = var_7552, x = linear_77_cast_fp16)[name = tensor("qkv_39_cast_fp16")]; + tensor q_115_begin_0 = const()[name = tensor("q_115_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor q_115_end_0 = const()[name = tensor("q_115_end_0"), val = tensor([1, 1, 1, 16, 64])]; + tensor q_115_end_mask_0 = const()[name = tensor("q_115_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor q_115_squeeze_mask_0 = const()[name = tensor("q_115_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor q_115_cast_fp16 = slice_by_index(begin = q_115_begin_0, end = q_115_end_0, end_mask = q_115_end_mask_0, squeeze_mask = q_115_squeeze_mask_0, x = qkv_39_cast_fp16)[name = tensor("q_115_cast_fp16")]; + tensor k_77_begin_0 = const()[name = tensor("k_77_begin_0"), val = tensor([0, 0, 1, 0, 0])]; + tensor k_77_end_0 = const()[name = tensor("k_77_end_0"), val = tensor([1, 1, 2, 16, 64])]; + tensor k_77_end_mask_0 = const()[name = tensor("k_77_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor k_77_squeeze_mask_0 = const()[name = tensor("k_77_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor k_77_cast_fp16 = slice_by_index(begin = k_77_begin_0, end = k_77_end_0, end_mask = k_77_end_mask_0, squeeze_mask = k_77_squeeze_mask_0, x = qkv_39_cast_fp16)[name = tensor("k_77_cast_fp16")]; + tensor v_39_begin_0 = const()[name = tensor("v_39_begin_0"), val = tensor([0, 0, 2, 0, 0])]; + tensor v_39_end_0 = const()[name = tensor("v_39_end_0"), val = tensor([1, 1, 3, 16, 64])]; + tensor v_39_end_mask_0 = const()[name = tensor("v_39_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor v_39_squeeze_mask_0 = const()[name = tensor("v_39_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor v_39_cast_fp16 = slice_by_index(begin = v_39_begin_0, end = v_39_end_0, end_mask = v_39_end_mask_0, squeeze_mask = v_39_squeeze_mask_0, x = qkv_39_cast_fp16)[name = tensor("v_39_cast_fp16")]; + tensor var_7618 = const()[name = tensor("op_7618"), val = tensor([1, 1, 1, 1])]; + tensor position19_to_fp16_dtype_0 = const()[name = tensor("position19_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor position19_to_fp16 = cast(dtype = position19_to_fp16_dtype_0, x = position19)[name = tensor("cast_458")]; + tensor ts_119_cast_fp16 = reshape(shape = var_7618, x = position19_to_fp16)[name = tensor("ts_119_cast_fp16")]; + tensor var_7625 = const()[name = tensor("op_7625"), val = tensor([1, 1, 16, 32, 2])]; + tensor q_complex_39_cast_fp16 = reshape(shape = var_7625, x = q_115_cast_fp16)[name = tensor("q_complex_39_cast_fp16")]; + tensor var_7632 = const()[name = tensor("op_7632"), val = tensor([1, 1, 16, 32, 2])]; + tensor k_complex_39_cast_fp16 = reshape(shape = var_7632, x = k_77_cast_fp16)[name = tensor("k_complex_39_cast_fp16")]; + tensor var_7636_begin_0 = const()[name = tensor("op_7636_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_7636_end_0 = const()[name = tensor("op_7636_end_0"), val = tensor([1, 1, 16, 32, 1])]; + tensor var_7636_end_mask_0 = const()[name = tensor("op_7636_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_7636_squeeze_mask_0 = const()[name = tensor("op_7636_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_7636_cast_fp16 = slice_by_index(begin = var_7636_begin_0, end = var_7636_end_0, end_mask = var_7636_end_mask_0, squeeze_mask = var_7636_squeeze_mask_0, x = q_complex_39_cast_fp16)[name = tensor("op_7636_cast_fp16")]; + tensor var_7644_begin_0 = const()[name = tensor("op_7644_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_7644_end_0 = const()[name = tensor("op_7644_end_0"), val = tensor([1, 1, 16, 32, 2])]; + tensor var_7644_end_mask_0 = const()[name = tensor("op_7644_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_7644_squeeze_mask_0 = const()[name = tensor("op_7644_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_7644_cast_fp16 = slice_by_index(begin = var_7644_begin_0, end = var_7644_end_0, end_mask = var_7644_end_mask_0, squeeze_mask = var_7644_squeeze_mask_0, x = q_complex_39_cast_fp16)[name = tensor("op_7644_cast_fp16")]; + tensor var_7652_begin_0 = const()[name = tensor("op_7652_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_7652_end_0 = const()[name = tensor("op_7652_end_0"), val = tensor([1, 1, 16, 32, 1])]; + tensor var_7652_end_mask_0 = const()[name = tensor("op_7652_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_7652_squeeze_mask_0 = const()[name = tensor("op_7652_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_7652_cast_fp16 = slice_by_index(begin = var_7652_begin_0, end = var_7652_end_0, end_mask = var_7652_end_mask_0, squeeze_mask = var_7652_squeeze_mask_0, x = k_complex_39_cast_fp16)[name = tensor("op_7652_cast_fp16")]; + tensor var_7660_begin_0 = const()[name = tensor("op_7660_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_7660_end_0 = const()[name = tensor("op_7660_end_0"), val = tensor([1, 1, 16, 32, 2])]; + tensor var_7660_end_mask_0 = const()[name = tensor("op_7660_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_7660_squeeze_mask_0 = const()[name = tensor("op_7660_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_7660_cast_fp16 = slice_by_index(begin = var_7660_begin_0, end = var_7660_end_0, end_mask = var_7660_end_mask_0, squeeze_mask = var_7660_squeeze_mask_0, x = k_complex_39_cast_fp16)[name = tensor("op_7660_cast_fp16")]; + tensor freqs_39_to_fp16 = const()[name = tensor("freqs_39_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(485746176)))]; + tensor var_7666_cast_fp16 = mul(x = freqs_39_to_fp16, y = ts_119_cast_fp16)[name = tensor("op_7666_cast_fp16")]; + tensor rotr_39_cast_fp16 = cos(x = var_7666_cast_fp16)[name = tensor("rotr_39_cast_fp16")]; + tensor roti_39_cast_fp16 = sin(x = var_7666_cast_fp16)[name = tensor("roti_39_cast_fp16")]; + tensor var_7670_cast_fp16 = mul(x = var_7636_cast_fp16, y = rotr_39_cast_fp16)[name = tensor("op_7670_cast_fp16")]; + tensor var_7671_cast_fp16 = mul(x = var_7644_cast_fp16, y = roti_39_cast_fp16)[name = tensor("op_7671_cast_fp16")]; + tensor qor_77_cast_fp16 = sub(x = var_7670_cast_fp16, y = var_7671_cast_fp16)[name = tensor("qor_77_cast_fp16")]; + tensor var_7674_cast_fp16 = mul(x = var_7636_cast_fp16, y = roti_39_cast_fp16)[name = tensor("op_7674_cast_fp16")]; + tensor var_7675_cast_fp16 = mul(x = var_7644_cast_fp16, y = rotr_39_cast_fp16)[name = tensor("op_7675_cast_fp16")]; + tensor qoi_77_cast_fp16 = add(x = var_7674_cast_fp16, y = var_7675_cast_fp16)[name = tensor("qoi_77_cast_fp16")]; + tensor var_7678_cast_fp16 = mul(x = var_7652_cast_fp16, y = rotr_39_cast_fp16)[name = tensor("op_7678_cast_fp16")]; + tensor var_7679_cast_fp16 = mul(x = var_7660_cast_fp16, y = roti_39_cast_fp16)[name = tensor("op_7679_cast_fp16")]; + tensor kor_77_cast_fp16 = sub(x = var_7678_cast_fp16, y = var_7679_cast_fp16)[name = tensor("kor_77_cast_fp16")]; + tensor var_7682_cast_fp16 = mul(x = var_7652_cast_fp16, y = roti_39_cast_fp16)[name = tensor("op_7682_cast_fp16")]; + tensor var_7683_cast_fp16 = mul(x = var_7660_cast_fp16, y = rotr_39_cast_fp16)[name = tensor("op_7683_cast_fp16")]; + tensor koi_77_cast_fp16 = add(x = var_7682_cast_fp16, y = var_7683_cast_fp16)[name = tensor("koi_77_cast_fp16")]; + tensor qo_39_axis_0 = const()[name = tensor("qo_39_axis_0"), val = tensor(-1)]; + tensor qo_39_cast_fp16 = stack(axis = qo_39_axis_0, values = (qor_77_cast_fp16, qoi_77_cast_fp16))[name = tensor("qo_39_cast_fp16")]; + tensor ko_39_axis_0 = const()[name = tensor("ko_39_axis_0"), val = tensor(-1)]; + tensor ko_39_cast_fp16 = stack(axis = ko_39_axis_0, values = (kor_77_cast_fp16, koi_77_cast_fp16))[name = tensor("ko_39_cast_fp16")]; + tensor var_7716 = const()[name = tensor("op_7716"), val = tensor([1, 1, 16, 64])]; + tensor q_117_cast_fp16 = reshape(shape = var_7716, x = qo_39_cast_fp16)[name = tensor("q_117_cast_fp16")]; + tensor var_7722 = const()[name = tensor("op_7722"), val = tensor([1, 1, 16, 64])]; + tensor k_79_cast_fp16 = reshape(shape = var_7722, x = ko_39_cast_fp16)[name = tensor("k_79_cast_fp16")]; + tensor _inversed_7747_y_0_to_fp16 = const()[name = tensor("_inversed_7747_y_0_to_fp16"), val = tensor(0x1p-9)]; + tensor _inversed_7747_cast_fp16 = mul(x = ts_119_cast_fp16, y = _inversed_7747_y_0_to_fp16)[name = tensor("_inversed_7747_cast_fp16")]; + tensor var_7748_cast_fp16 = floor(x = _inversed_7747_cast_fp16)[name = tensor("op_7748_cast_fp16")]; + tensor var_7749_to_fp16 = const()[name = tensor("op_7749_to_fp16"), val = tensor(0x1p+9)]; + tensor var_7750_cast_fp16 = mul(x = var_7748_cast_fp16, y = var_7749_to_fp16)[name = tensor("op_7750_cast_fp16")]; + tensor write_indices_float_79_cast_fp16 = sub(x = ts_119_cast_fp16, y = var_7750_cast_fp16)[name = tensor("write_indices_float_79_cast_fp16")]; + tensor var_7757_dtype_0 = const()[name = tensor("op_7757_dtype_0"), val = tensor("int32")]; + tensor write_indices_39_reps_0 = const()[name = tensor("write_indices_39_reps_0"), val = tensor([1, 1, 16, 64])]; + tensor write_indices_float_79_cast_fp16_to_int32 = cast(dtype = var_7757_dtype_0, x = write_indices_float_79_cast_fp16)[name = tensor("cast_457")]; + tensor write_indices_39 = tile(reps = write_indices_39_reps_0, x = write_indices_float_79_cast_fp16_to_int32)[name = tensor("write_indices_39")]; + tensor var_7767_begin_0 = const()[name = tensor("op_7767_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_7767_end_0 = const()[name = tensor("op_7767_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor var_7767_end_mask_0 = const()[name = tensor("op_7767_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_7767_squeeze_mask_0 = const()[name = tensor("op_7767_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor cache19_to_fp16_dtype_0 = const()[name = tensor("cache19_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor cache19_to_fp16 = cast(dtype = cache19_to_fp16_dtype_0, x = cache19)[name = tensor("cast_456")]; + tensor var_7767_cast_fp16 = slice_by_index(begin = var_7767_begin_0, end = var_7767_end_0, end_mask = var_7767_end_mask_0, squeeze_mask = var_7767_squeeze_mask_0, x = cache19_to_fp16)[name = tensor("op_7767_cast_fp16")]; + tensor var_7769_axis_0 = const()[name = tensor("op_7769_axis_0"), val = tensor(1)]; + tensor var_7769_mode_0 = const()[name = tensor("op_7769_mode_0"), val = tensor("update")]; + tensor var_7769_validate_indices_0 = const()[name = tensor("op_7769_validate_indices_0"), val = tensor(false)]; + tensor var_7769_cast_fp16 = scatter_along_axis(axis = var_7769_axis_0, data = var_7767_cast_fp16, indices = write_indices_39, mode = var_7769_mode_0, updates = k_79_cast_fp16, validate_indices = var_7769_validate_indices_0)[name = tensor("op_7769_cast_fp16")]; + tensor concat_135 = const()[name = tensor("concat_135"), val = tensor([0, 0, 0, 0, 0])]; + tensor concat_136 = const()[name = tensor("concat_136"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_39_internal_tensor_assign_1_stride_0 = const()[name = tensor("new_cache_39_internal_tensor_assign_1_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_39_internal_tensor_assign_1_begin_mask_0 = const()[name = tensor("new_cache_39_internal_tensor_assign_1_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_39_internal_tensor_assign_1_end_mask_0 = const()[name = tensor("new_cache_39_internal_tensor_assign_1_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_39_internal_tensor_assign_1_squeeze_mask_0 = const()[name = tensor("new_cache_39_internal_tensor_assign_1_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_86 = const()[name = tensor("shape_86"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_38 = const()[name = tensor("reduce_prod_38"), val = tensor(1048576)]; + tensor range_1d_38_start_0 = const()[name = tensor("range_1d_38_start_0"), val = tensor(0)]; + tensor range_1d_38_step_0 = const()[name = tensor("range_1d_38_step_0"), val = tensor(1)]; + tensor range_1d_38 = range_1d(end = reduce_prod_38, start = range_1d_38_start_0, step = range_1d_38_step_0)[name = tensor("range_1d_38")]; + tensor reshape_190 = reshape(shape = shape_86, x = range_1d_38)[name = tensor("reshape_190")]; + tensor slice_by_index_38 = slice_by_index(begin = concat_135, begin_mask = new_cache_39_internal_tensor_assign_1_begin_mask_0, end = concat_136, end_mask = new_cache_39_internal_tensor_assign_1_end_mask_0, squeeze_mask = new_cache_39_internal_tensor_assign_1_squeeze_mask_0, stride = new_cache_39_internal_tensor_assign_1_stride_0, x = reshape_190)[name = tensor("slice_by_index_38")]; + tensor reshape_191_shape_0 = const()[name = tensor("reshape_191_shape_0"), val = tensor([-1])]; + tensor reshape_191 = reshape(shape = reshape_191_shape_0, x = slice_by_index_38)[name = tensor("reshape_191")]; + tensor reshape_192_shape_0 = const()[name = tensor("reshape_192_shape_0"), val = tensor([-1])]; + tensor reshape_192_cast_fp16 = reshape(shape = reshape_192_shape_0, x = var_7769_cast_fp16)[name = tensor("reshape_192_cast_fp16")]; + tensor reshape_193_shape_0 = const()[name = tensor("reshape_193_shape_0"), val = tensor([-1])]; + tensor reshape_193_cast_fp16 = reshape(shape = reshape_193_shape_0, x = cache19_to_fp16)[name = tensor("reshape_193_cast_fp16")]; + tensor scatter_38_mode_0 = const()[name = tensor("scatter_38_mode_0"), val = tensor("update")]; + tensor scatter_38_axis_0 = const()[name = tensor("scatter_38_axis_0"), val = tensor(0)]; + tensor scatter_38_validate_indices_0 = const()[name = tensor("scatter_38_validate_indices_0"), val = tensor(false)]; + tensor scatter_38_cast_fp16 = scatter(axis = scatter_38_axis_0, data = reshape_193_cast_fp16, indices = reshape_191, mode = scatter_38_mode_0, updates = reshape_192_cast_fp16, validate_indices = scatter_38_validate_indices_0)[name = tensor("scatter_38_cast_fp16")]; + tensor reshape_194_cast_fp16 = reshape(shape = shape_86, x = scatter_38_cast_fp16)[name = tensor("reshape_194_cast_fp16")]; + tensor var_7777_begin_0 = const()[name = tensor("op_7777_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor var_7777_end_0 = const()[name = tensor("op_7777_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor var_7777_end_mask_0 = const()[name = tensor("op_7777_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_7777_squeeze_mask_0 = const()[name = tensor("op_7777_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor var_7777_cast_fp16 = slice_by_index(begin = var_7777_begin_0, end = var_7777_end_0, end_mask = var_7777_end_mask_0, squeeze_mask = var_7777_squeeze_mask_0, x = reshape_194_cast_fp16)[name = tensor("op_7777_cast_fp16")]; + tensor var_7779_axis_0 = const()[name = tensor("op_7779_axis_0"), val = tensor(1)]; + tensor var_7779_mode_0 = const()[name = tensor("op_7779_mode_0"), val = tensor("update")]; + tensor var_7779_validate_indices_0 = const()[name = tensor("op_7779_validate_indices_0"), val = tensor(false)]; + tensor var_7779_cast_fp16 = scatter_along_axis(axis = var_7779_axis_0, data = var_7777_cast_fp16, indices = write_indices_39, mode = var_7779_mode_0, updates = v_39_cast_fp16, validate_indices = var_7779_validate_indices_0)[name = tensor("op_7779_cast_fp16")]; + tensor concat_137 = const()[name = tensor("concat_137"), val = tensor([1, 0, 0, 0, 0])]; + tensor concat_138 = const()[name = tensor("concat_138"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_39_internal_tensor_assign_2_stride_0 = const()[name = tensor("new_cache_39_internal_tensor_assign_2_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_39_internal_tensor_assign_2_begin_mask_0 = const()[name = tensor("new_cache_39_internal_tensor_assign_2_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_39_internal_tensor_assign_2_end_mask_0 = const()[name = tensor("new_cache_39_internal_tensor_assign_2_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_39_internal_tensor_assign_2_squeeze_mask_0 = const()[name = tensor("new_cache_39_internal_tensor_assign_2_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_87 = const()[name = tensor("shape_87"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_39 = const()[name = tensor("reduce_prod_39"), val = tensor(1048576)]; + tensor range_1d_39_start_0 = const()[name = tensor("range_1d_39_start_0"), val = tensor(0)]; + tensor range_1d_39_step_0 = const()[name = tensor("range_1d_39_step_0"), val = tensor(1)]; + tensor range_1d_39 = range_1d(end = reduce_prod_39, start = range_1d_39_start_0, step = range_1d_39_step_0)[name = tensor("range_1d_39")]; + tensor reshape_195 = reshape(shape = shape_87, x = range_1d_39)[name = tensor("reshape_195")]; + tensor slice_by_index_39 = slice_by_index(begin = concat_137, begin_mask = new_cache_39_internal_tensor_assign_2_begin_mask_0, end = concat_138, end_mask = new_cache_39_internal_tensor_assign_2_end_mask_0, squeeze_mask = new_cache_39_internal_tensor_assign_2_squeeze_mask_0, stride = new_cache_39_internal_tensor_assign_2_stride_0, x = reshape_195)[name = tensor("slice_by_index_39")]; + tensor reshape_196_shape_0 = const()[name = tensor("reshape_196_shape_0"), val = tensor([-1])]; + tensor reshape_196 = reshape(shape = reshape_196_shape_0, x = slice_by_index_39)[name = tensor("reshape_196")]; + tensor reshape_197_shape_0 = const()[name = tensor("reshape_197_shape_0"), val = tensor([-1])]; + tensor reshape_197_cast_fp16 = reshape(shape = reshape_197_shape_0, x = var_7779_cast_fp16)[name = tensor("reshape_197_cast_fp16")]; + tensor reshape_198_shape_0 = const()[name = tensor("reshape_198_shape_0"), val = tensor([-1])]; + tensor reshape_198_cast_fp16 = reshape(shape = reshape_198_shape_0, x = reshape_194_cast_fp16)[name = tensor("reshape_198_cast_fp16")]; + tensor scatter_39_mode_0 = const()[name = tensor("scatter_39_mode_0"), val = tensor("update")]; + tensor scatter_39_axis_0 = const()[name = tensor("scatter_39_axis_0"), val = tensor(0)]; + tensor scatter_39_validate_indices_0 = const()[name = tensor("scatter_39_validate_indices_0"), val = tensor(false)]; + tensor scatter_39_cast_fp16 = scatter(axis = scatter_39_axis_0, data = reshape_198_cast_fp16, indices = reshape_196, mode = scatter_39_mode_0, updates = reshape_197_cast_fp16, validate_indices = scatter_39_validate_indices_0)[name = tensor("scatter_39_cast_fp16")]; + tensor reshape_199_cast_fp16 = reshape(shape = shape_87, x = scatter_39_cast_fp16)[name = tensor("reshape_199_cast_fp16")]; + tensor reshape_199_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("reshape_199_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor keys_115_begin_0 = const()[name = tensor("keys_115_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor keys_115_end_0 = const()[name = tensor("keys_115_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor keys_115_end_mask_0 = const()[name = tensor("keys_115_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor keys_115_squeeze_mask_0 = const()[name = tensor("keys_115_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor keys_115_cast_fp16 = slice_by_index(begin = keys_115_begin_0, end = keys_115_end_0, end_mask = keys_115_end_mask_0, squeeze_mask = keys_115_squeeze_mask_0, x = reshape_199_cast_fp16)[name = tensor("keys_115_cast_fp16")]; + tensor values_115_begin_0 = const()[name = tensor("values_115_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor values_115_end_0 = const()[name = tensor("values_115_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor values_115_end_mask_0 = const()[name = tensor("values_115_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor values_115_squeeze_mask_0 = const()[name = tensor("values_115_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor values_115_cast_fp16 = slice_by_index(begin = values_115_begin_0, end = values_115_end_0, end_mask = values_115_end_mask_0, squeeze_mask = values_115_squeeze_mask_0, x = reshape_199_cast_fp16)[name = tensor("values_115_cast_fp16")]; + tensor var_7791_cast_fp16 = not_equal(x = keys_115_cast_fp16, y = keys_115_cast_fp16)[name = tensor("op_7791_cast_fp16")]; + tensor keys_117_cast_fp16 = select(a = var_463_to_fp16, b = keys_115_cast_fp16, cond = var_7791_cast_fp16)[name = tensor("keys_117_cast_fp16")]; + tensor var_7799_cast_fp16 = not_equal(x = values_115_cast_fp16, y = values_115_cast_fp16)[name = tensor("op_7799_cast_fp16")]; + tensor values_117_cast_fp16 = select(a = var_463_to_fp16, b = values_115_cast_fp16, cond = var_7799_cast_fp16)[name = tensor("values_117_cast_fp16")]; + tensor var_7823 = const()[name = tensor("op_7823"), val = tensor([0, 2, 1, 3])]; + tensor var_7839 = const()[name = tensor("op_7839"), val = tensor([1, 1, 1])]; + tensor var_7840_cast_fp16 = reshape(shape = var_7839, x = position19_to_fp16)[name = tensor("op_7840_cast_fp16")]; + tensor var_7860_to_fp16 = const()[name = tensor("op_7860_to_fp16"), val = tensor(0x1p+0)]; + tensor valid_len_39_cast_fp16 = add(x = var_7840_cast_fp16, y = var_7860_to_fp16)[name = tensor("valid_len_39_cast_fp16")]; + tensor valid_mask_39_cast_fp16 = less(x = k_positions_1_promoted_to_fp16, y = valid_len_39_cast_fp16)[name = tensor("valid_mask_39_cast_fp16")]; + tensor causal_mask_39_cast_fp16 = less_equal(x = k_positions_1_promoted_to_fp16, y = var_7840_cast_fp16)[name = tensor("causal_mask_39_cast_fp16")]; + tensor attn_mask_77 = logical_and(x = valid_mask_39_cast_fp16, y = causal_mask_39_cast_fp16)[name = tensor("attn_mask_77")]; + tensor attn_mask_79_axes_0 = const()[name = tensor("attn_mask_79_axes_0"), val = tensor([1])]; + tensor attn_mask_79 = expand_dims(axes = attn_mask_79_axes_0, x = attn_mask_77)[name = tensor("attn_mask_79")]; + tensor var_7878_transpose_x_0 = const()[name = tensor("op_7878_transpose_x_0"), val = tensor(false)]; + tensor var_7878_transpose_y_0 = const()[name = tensor("op_7878_transpose_y_0"), val = tensor(false)]; + tensor transpose_110_perm_0 = const()[name = tensor("transpose_110_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_111_perm_0 = const()[name = tensor("transpose_111_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_111 = transpose(perm = transpose_111_perm_0, x = keys_117_cast_fp16)[name = tensor("transpose_137")]; + tensor transpose_110 = transpose(perm = transpose_110_perm_0, x = q_117_cast_fp16)[name = tensor("transpose_138")]; + tensor var_7878_cast_fp16 = matmul(transpose_x = var_7878_transpose_x_0, transpose_y = var_7878_transpose_y_0, x = transpose_110, y = transpose_111)[name = tensor("op_7878_cast_fp16")]; + tensor var_7872_to_fp16 = const()[name = tensor("op_7872_to_fp16"), val = tensor([0x1p-3])]; + tensor attn_weights_115_cast_fp16 = mul(x = var_7878_cast_fp16, y = var_7872_to_fp16)[name = tensor("attn_weights_115_cast_fp16")]; + tensor var_7880 = logical_not(x = attn_mask_79)[name = tensor("op_7880")]; + tensor var_7881_to_fp16 = const()[name = tensor("op_7881_to_fp16"), val = tensor(-inf)]; + tensor attn_weights_117_cast_fp16 = select(a = var_7881_to_fp16, b = attn_weights_115_cast_fp16, cond = var_7880)[name = tensor("attn_weights_117_cast_fp16")]; + tensor var_7883 = const()[name = tensor("op_7883"), val = tensor(-1)]; + tensor attn_weights_119_cast_fp16 = softmax(axis = var_7883, x = attn_weights_117_cast_fp16)[name = tensor("attn_weights_119_cast_fp16")]; + tensor attn_output_39_transpose_x_0 = const()[name = tensor("attn_output_39_transpose_x_0"), val = tensor(false)]; + tensor attn_output_39_transpose_y_0 = const()[name = tensor("attn_output_39_transpose_y_0"), val = tensor(false)]; + tensor values_119_cast_fp16 = transpose(perm = var_7823, x = values_117_cast_fp16)[name = tensor("transpose_139")]; + tensor attn_output_39_cast_fp16 = matmul(transpose_x = attn_output_39_transpose_x_0, transpose_y = attn_output_39_transpose_y_0, x = attn_weights_119_cast_fp16, y = values_119_cast_fp16)[name = tensor("attn_output_39_cast_fp16")]; + tensor var_7891 = const()[name = tensor("op_7891"), val = tensor([0, 2, 1, 3])]; + tensor var_7896 = const()[name = tensor("op_7896"), val = tensor([1, 1, 1024])]; + tensor var_7892_cast_fp16 = transpose(perm = var_7891, x = attn_output_39_cast_fp16)[name = tensor("transpose_136")]; + tensor input_235_cast_fp16 = reshape(shape = var_7896, x = var_7892_cast_fp16)[name = tensor("input_235_cast_fp16")]; + tensor attn19_out_proj_weight_to_fp16 = const()[name = tensor("attn19_out_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(485746304)))]; + tensor linear_78_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = attn19_out_proj_weight_to_fp16, x = input_235_cast_fp16)[name = tensor("linear_78_cast_fp16")]; + tensor var_7902_to_fp16 = const()[name = tensor("op_7902_to_fp16"), val = tensor(0x1p+0)]; + tensor var_7903_cast_fp16 = add(x = position19_to_fp16, y = var_7902_to_fp16)[name = tensor("op_7903_cast_fp16")]; + tensor var_7903_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_7903_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor input_237_cast_fp16 = add(x = input_231_cast_fp16, y = linear_78_cast_fp16)[name = tensor("input_237_cast_fp16")]; + tensor input_239_axes_0 = const()[name = tensor("input_239_axes_0"), val = tensor([-1])]; + tensor norm19_2_weight_to_fp16 = const()[name = tensor("norm19_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(487843520)))]; + tensor norm19_2_bias_to_fp16 = const()[name = tensor("norm19_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(487845632)))]; + tensor var_7907_to_fp16 = const()[name = tensor("op_7907_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_239_cast_fp16 = layer_norm(axes = input_239_axes_0, beta = norm19_2_bias_to_fp16, epsilon = var_7907_to_fp16, gamma = norm19_2_weight_to_fp16, x = input_237_cast_fp16)[name = tensor("input_239_cast_fp16")]; + tensor linear19_1_weight_to_fp16 = const()[name = tensor("linear19_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(487847744)))]; + tensor linear_79_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = linear19_1_weight_to_fp16, x = input_239_cast_fp16)[name = tensor("linear_79_cast_fp16")]; + tensor input_241_mode_0 = const()[name = tensor("input_241_mode_0"), val = tensor("EXACT")]; + tensor input_241_cast_fp16 = gelu(mode = input_241_mode_0, x = linear_79_cast_fp16)[name = tensor("input_241_cast_fp16")]; + tensor linear19_2_weight_to_fp16 = const()[name = tensor("linear19_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(496236416)))]; + tensor linear_80_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = linear19_2_weight_to_fp16, x = input_241_cast_fp16)[name = tensor("linear_80_cast_fp16")]; + tensor input_243_cast_fp16 = add(x = input_237_cast_fp16, y = linear_80_cast_fp16)[name = tensor("input_243_cast_fp16")]; + tensor input_245_axes_0 = const()[name = tensor("input_245_axes_0"), val = tensor([-1])]; + tensor norm20_1_weight_to_fp16 = const()[name = tensor("norm20_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(504625088)))]; + tensor norm20_1_bias_to_fp16 = const()[name = tensor("norm20_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(504627200)))]; + tensor var_7924_to_fp16 = const()[name = tensor("op_7924_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_245_cast_fp16 = layer_norm(axes = input_245_axes_0, beta = norm20_1_bias_to_fp16, epsilon = var_7924_to_fp16, gamma = norm20_1_weight_to_fp16, x = input_243_cast_fp16)[name = tensor("input_245_cast_fp16")]; + tensor attn20_in_proj_weight_to_fp16 = const()[name = tensor("attn20_in_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(504629312)))]; + tensor linear_81_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = attn20_in_proj_weight_to_fp16, x = input_245_cast_fp16)[name = tensor("linear_81_cast_fp16")]; + tensor var_7938 = const()[name = tensor("op_7938"), val = tensor([1, 1, 3, 16, 64])]; + tensor qkv_41_cast_fp16 = reshape(shape = var_7938, x = linear_81_cast_fp16)[name = tensor("qkv_41_cast_fp16")]; + tensor q_121_begin_0 = const()[name = tensor("q_121_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor q_121_end_0 = const()[name = tensor("q_121_end_0"), val = tensor([1, 1, 1, 16, 64])]; + tensor q_121_end_mask_0 = const()[name = tensor("q_121_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor q_121_squeeze_mask_0 = const()[name = tensor("q_121_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor q_121_cast_fp16 = slice_by_index(begin = q_121_begin_0, end = q_121_end_0, end_mask = q_121_end_mask_0, squeeze_mask = q_121_squeeze_mask_0, x = qkv_41_cast_fp16)[name = tensor("q_121_cast_fp16")]; + tensor k_81_begin_0 = const()[name = tensor("k_81_begin_0"), val = tensor([0, 0, 1, 0, 0])]; + tensor k_81_end_0 = const()[name = tensor("k_81_end_0"), val = tensor([1, 1, 2, 16, 64])]; + tensor k_81_end_mask_0 = const()[name = tensor("k_81_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor k_81_squeeze_mask_0 = const()[name = tensor("k_81_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor k_81_cast_fp16 = slice_by_index(begin = k_81_begin_0, end = k_81_end_0, end_mask = k_81_end_mask_0, squeeze_mask = k_81_squeeze_mask_0, x = qkv_41_cast_fp16)[name = tensor("k_81_cast_fp16")]; + tensor v_41_begin_0 = const()[name = tensor("v_41_begin_0"), val = tensor([0, 0, 2, 0, 0])]; + tensor v_41_end_0 = const()[name = tensor("v_41_end_0"), val = tensor([1, 1, 3, 16, 64])]; + tensor v_41_end_mask_0 = const()[name = tensor("v_41_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor v_41_squeeze_mask_0 = const()[name = tensor("v_41_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor v_41_cast_fp16 = slice_by_index(begin = v_41_begin_0, end = v_41_end_0, end_mask = v_41_end_mask_0, squeeze_mask = v_41_squeeze_mask_0, x = qkv_41_cast_fp16)[name = tensor("v_41_cast_fp16")]; + tensor var_8004 = const()[name = tensor("op_8004"), val = tensor([1, 1, 1, 1])]; + tensor position20_to_fp16_dtype_0 = const()[name = tensor("position20_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor position20_to_fp16 = cast(dtype = position20_to_fp16_dtype_0, x = position20)[name = tensor("cast_453")]; + tensor ts_125_cast_fp16 = reshape(shape = var_8004, x = position20_to_fp16)[name = tensor("ts_125_cast_fp16")]; + tensor var_8011 = const()[name = tensor("op_8011"), val = tensor([1, 1, 16, 32, 2])]; + tensor q_complex_41_cast_fp16 = reshape(shape = var_8011, x = q_121_cast_fp16)[name = tensor("q_complex_41_cast_fp16")]; + tensor var_8018 = const()[name = tensor("op_8018"), val = tensor([1, 1, 16, 32, 2])]; + tensor k_complex_41_cast_fp16 = reshape(shape = var_8018, x = k_81_cast_fp16)[name = tensor("k_complex_41_cast_fp16")]; + tensor var_8022_begin_0 = const()[name = tensor("op_8022_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_8022_end_0 = const()[name = tensor("op_8022_end_0"), val = tensor([1, 1, 16, 32, 1])]; + tensor var_8022_end_mask_0 = const()[name = tensor("op_8022_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_8022_squeeze_mask_0 = const()[name = tensor("op_8022_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_8022_cast_fp16 = slice_by_index(begin = var_8022_begin_0, end = var_8022_end_0, end_mask = var_8022_end_mask_0, squeeze_mask = var_8022_squeeze_mask_0, x = q_complex_41_cast_fp16)[name = tensor("op_8022_cast_fp16")]; + tensor var_8030_begin_0 = const()[name = tensor("op_8030_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_8030_end_0 = const()[name = tensor("op_8030_end_0"), val = tensor([1, 1, 16, 32, 2])]; + tensor var_8030_end_mask_0 = const()[name = tensor("op_8030_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_8030_squeeze_mask_0 = const()[name = tensor("op_8030_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_8030_cast_fp16 = slice_by_index(begin = var_8030_begin_0, end = var_8030_end_0, end_mask = var_8030_end_mask_0, squeeze_mask = var_8030_squeeze_mask_0, x = q_complex_41_cast_fp16)[name = tensor("op_8030_cast_fp16")]; + tensor var_8038_begin_0 = const()[name = tensor("op_8038_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_8038_end_0 = const()[name = tensor("op_8038_end_0"), val = tensor([1, 1, 16, 32, 1])]; + tensor var_8038_end_mask_0 = const()[name = tensor("op_8038_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_8038_squeeze_mask_0 = const()[name = tensor("op_8038_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_8038_cast_fp16 = slice_by_index(begin = var_8038_begin_0, end = var_8038_end_0, end_mask = var_8038_end_mask_0, squeeze_mask = var_8038_squeeze_mask_0, x = k_complex_41_cast_fp16)[name = tensor("op_8038_cast_fp16")]; + tensor var_8046_begin_0 = const()[name = tensor("op_8046_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_8046_end_0 = const()[name = tensor("op_8046_end_0"), val = tensor([1, 1, 16, 32, 2])]; + tensor var_8046_end_mask_0 = const()[name = tensor("op_8046_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_8046_squeeze_mask_0 = const()[name = tensor("op_8046_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_8046_cast_fp16 = slice_by_index(begin = var_8046_begin_0, end = var_8046_end_0, end_mask = var_8046_end_mask_0, squeeze_mask = var_8046_squeeze_mask_0, x = k_complex_41_cast_fp16)[name = tensor("op_8046_cast_fp16")]; + tensor freqs_41_to_fp16 = const()[name = tensor("freqs_41_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(510920832)))]; + tensor var_8052_cast_fp16 = mul(x = freqs_41_to_fp16, y = ts_125_cast_fp16)[name = tensor("op_8052_cast_fp16")]; + tensor rotr_41_cast_fp16 = cos(x = var_8052_cast_fp16)[name = tensor("rotr_41_cast_fp16")]; + tensor roti_41_cast_fp16 = sin(x = var_8052_cast_fp16)[name = tensor("roti_41_cast_fp16")]; + tensor var_8056_cast_fp16 = mul(x = var_8022_cast_fp16, y = rotr_41_cast_fp16)[name = tensor("op_8056_cast_fp16")]; + tensor var_8057_cast_fp16 = mul(x = var_8030_cast_fp16, y = roti_41_cast_fp16)[name = tensor("op_8057_cast_fp16")]; + tensor qor_81_cast_fp16 = sub(x = var_8056_cast_fp16, y = var_8057_cast_fp16)[name = tensor("qor_81_cast_fp16")]; + tensor var_8060_cast_fp16 = mul(x = var_8022_cast_fp16, y = roti_41_cast_fp16)[name = tensor("op_8060_cast_fp16")]; + tensor var_8061_cast_fp16 = mul(x = var_8030_cast_fp16, y = rotr_41_cast_fp16)[name = tensor("op_8061_cast_fp16")]; + tensor qoi_81_cast_fp16 = add(x = var_8060_cast_fp16, y = var_8061_cast_fp16)[name = tensor("qoi_81_cast_fp16")]; + tensor var_8064_cast_fp16 = mul(x = var_8038_cast_fp16, y = rotr_41_cast_fp16)[name = tensor("op_8064_cast_fp16")]; + tensor var_8065_cast_fp16 = mul(x = var_8046_cast_fp16, y = roti_41_cast_fp16)[name = tensor("op_8065_cast_fp16")]; + tensor kor_81_cast_fp16 = sub(x = var_8064_cast_fp16, y = var_8065_cast_fp16)[name = tensor("kor_81_cast_fp16")]; + tensor var_8068_cast_fp16 = mul(x = var_8038_cast_fp16, y = roti_41_cast_fp16)[name = tensor("op_8068_cast_fp16")]; + tensor var_8069_cast_fp16 = mul(x = var_8046_cast_fp16, y = rotr_41_cast_fp16)[name = tensor("op_8069_cast_fp16")]; + tensor koi_81_cast_fp16 = add(x = var_8068_cast_fp16, y = var_8069_cast_fp16)[name = tensor("koi_81_cast_fp16")]; + tensor qo_41_axis_0 = const()[name = tensor("qo_41_axis_0"), val = tensor(-1)]; + tensor qo_41_cast_fp16 = stack(axis = qo_41_axis_0, values = (qor_81_cast_fp16, qoi_81_cast_fp16))[name = tensor("qo_41_cast_fp16")]; + tensor ko_41_axis_0 = const()[name = tensor("ko_41_axis_0"), val = tensor(-1)]; + tensor ko_41_cast_fp16 = stack(axis = ko_41_axis_0, values = (kor_81_cast_fp16, koi_81_cast_fp16))[name = tensor("ko_41_cast_fp16")]; + tensor var_8102 = const()[name = tensor("op_8102"), val = tensor([1, 1, 16, 64])]; + tensor q_123_cast_fp16 = reshape(shape = var_8102, x = qo_41_cast_fp16)[name = tensor("q_123_cast_fp16")]; + tensor var_8108 = const()[name = tensor("op_8108"), val = tensor([1, 1, 16, 64])]; + tensor k_83_cast_fp16 = reshape(shape = var_8108, x = ko_41_cast_fp16)[name = tensor("k_83_cast_fp16")]; + tensor _inversed_8133_y_0_to_fp16 = const()[name = tensor("_inversed_8133_y_0_to_fp16"), val = tensor(0x1p-9)]; + tensor _inversed_8133_cast_fp16 = mul(x = ts_125_cast_fp16, y = _inversed_8133_y_0_to_fp16)[name = tensor("_inversed_8133_cast_fp16")]; + tensor var_8134_cast_fp16 = floor(x = _inversed_8133_cast_fp16)[name = tensor("op_8134_cast_fp16")]; + tensor var_8135_to_fp16 = const()[name = tensor("op_8135_to_fp16"), val = tensor(0x1p+9)]; + tensor var_8136_cast_fp16 = mul(x = var_8134_cast_fp16, y = var_8135_to_fp16)[name = tensor("op_8136_cast_fp16")]; + tensor write_indices_float_83_cast_fp16 = sub(x = ts_125_cast_fp16, y = var_8136_cast_fp16)[name = tensor("write_indices_float_83_cast_fp16")]; + tensor var_8143_dtype_0 = const()[name = tensor("op_8143_dtype_0"), val = tensor("int32")]; + tensor write_indices_41_reps_0 = const()[name = tensor("write_indices_41_reps_0"), val = tensor([1, 1, 16, 64])]; + tensor write_indices_float_83_cast_fp16_to_int32 = cast(dtype = var_8143_dtype_0, x = write_indices_float_83_cast_fp16)[name = tensor("cast_452")]; + tensor write_indices_41 = tile(reps = write_indices_41_reps_0, x = write_indices_float_83_cast_fp16_to_int32)[name = tensor("write_indices_41")]; + tensor var_8153_begin_0 = const()[name = tensor("op_8153_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_8153_end_0 = const()[name = tensor("op_8153_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor var_8153_end_mask_0 = const()[name = tensor("op_8153_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_8153_squeeze_mask_0 = const()[name = tensor("op_8153_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor cache20_to_fp16_dtype_0 = const()[name = tensor("cache20_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor cache20_to_fp16 = cast(dtype = cache20_to_fp16_dtype_0, x = cache20)[name = tensor("cast_451")]; + tensor var_8153_cast_fp16 = slice_by_index(begin = var_8153_begin_0, end = var_8153_end_0, end_mask = var_8153_end_mask_0, squeeze_mask = var_8153_squeeze_mask_0, x = cache20_to_fp16)[name = tensor("op_8153_cast_fp16")]; + tensor var_8155_axis_0 = const()[name = tensor("op_8155_axis_0"), val = tensor(1)]; + tensor var_8155_mode_0 = const()[name = tensor("op_8155_mode_0"), val = tensor("update")]; + tensor var_8155_validate_indices_0 = const()[name = tensor("op_8155_validate_indices_0"), val = tensor(false)]; + tensor var_8155_cast_fp16 = scatter_along_axis(axis = var_8155_axis_0, data = var_8153_cast_fp16, indices = write_indices_41, mode = var_8155_mode_0, updates = k_83_cast_fp16, validate_indices = var_8155_validate_indices_0)[name = tensor("op_8155_cast_fp16")]; + tensor concat_142 = const()[name = tensor("concat_142"), val = tensor([0, 0, 0, 0, 0])]; + tensor concat_143 = const()[name = tensor("concat_143"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_41_internal_tensor_assign_1_stride_0 = const()[name = tensor("new_cache_41_internal_tensor_assign_1_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_41_internal_tensor_assign_1_begin_mask_0 = const()[name = tensor("new_cache_41_internal_tensor_assign_1_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_41_internal_tensor_assign_1_end_mask_0 = const()[name = tensor("new_cache_41_internal_tensor_assign_1_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_41_internal_tensor_assign_1_squeeze_mask_0 = const()[name = tensor("new_cache_41_internal_tensor_assign_1_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_88 = const()[name = tensor("shape_88"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_40 = const()[name = tensor("reduce_prod_40"), val = tensor(1048576)]; + tensor range_1d_40_start_0 = const()[name = tensor("range_1d_40_start_0"), val = tensor(0)]; + tensor range_1d_40_step_0 = const()[name = tensor("range_1d_40_step_0"), val = tensor(1)]; + tensor range_1d_40 = range_1d(end = reduce_prod_40, start = range_1d_40_start_0, step = range_1d_40_step_0)[name = tensor("range_1d_40")]; + tensor reshape_200 = reshape(shape = shape_88, x = range_1d_40)[name = tensor("reshape_200")]; + tensor slice_by_index_40 = slice_by_index(begin = concat_142, begin_mask = new_cache_41_internal_tensor_assign_1_begin_mask_0, end = concat_143, end_mask = new_cache_41_internal_tensor_assign_1_end_mask_0, squeeze_mask = new_cache_41_internal_tensor_assign_1_squeeze_mask_0, stride = new_cache_41_internal_tensor_assign_1_stride_0, x = reshape_200)[name = tensor("slice_by_index_40")]; + tensor reshape_201_shape_0 = const()[name = tensor("reshape_201_shape_0"), val = tensor([-1])]; + tensor reshape_201 = reshape(shape = reshape_201_shape_0, x = slice_by_index_40)[name = tensor("reshape_201")]; + tensor reshape_202_shape_0 = const()[name = tensor("reshape_202_shape_0"), val = tensor([-1])]; + tensor reshape_202_cast_fp16 = reshape(shape = reshape_202_shape_0, x = var_8155_cast_fp16)[name = tensor("reshape_202_cast_fp16")]; + tensor reshape_203_shape_0 = const()[name = tensor("reshape_203_shape_0"), val = tensor([-1])]; + tensor reshape_203_cast_fp16 = reshape(shape = reshape_203_shape_0, x = cache20_to_fp16)[name = tensor("reshape_203_cast_fp16")]; + tensor scatter_40_mode_0 = const()[name = tensor("scatter_40_mode_0"), val = tensor("update")]; + tensor scatter_40_axis_0 = const()[name = tensor("scatter_40_axis_0"), val = tensor(0)]; + tensor scatter_40_validate_indices_0 = const()[name = tensor("scatter_40_validate_indices_0"), val = tensor(false)]; + tensor scatter_40_cast_fp16 = scatter(axis = scatter_40_axis_0, data = reshape_203_cast_fp16, indices = reshape_201, mode = scatter_40_mode_0, updates = reshape_202_cast_fp16, validate_indices = scatter_40_validate_indices_0)[name = tensor("scatter_40_cast_fp16")]; + tensor reshape_204_cast_fp16 = reshape(shape = shape_88, x = scatter_40_cast_fp16)[name = tensor("reshape_204_cast_fp16")]; + tensor var_8163_begin_0 = const()[name = tensor("op_8163_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor var_8163_end_0 = const()[name = tensor("op_8163_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor var_8163_end_mask_0 = const()[name = tensor("op_8163_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_8163_squeeze_mask_0 = const()[name = tensor("op_8163_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor var_8163_cast_fp16 = slice_by_index(begin = var_8163_begin_0, end = var_8163_end_0, end_mask = var_8163_end_mask_0, squeeze_mask = var_8163_squeeze_mask_0, x = reshape_204_cast_fp16)[name = tensor("op_8163_cast_fp16")]; + tensor var_8165_axis_0 = const()[name = tensor("op_8165_axis_0"), val = tensor(1)]; + tensor var_8165_mode_0 = const()[name = tensor("op_8165_mode_0"), val = tensor("update")]; + tensor var_8165_validate_indices_0 = const()[name = tensor("op_8165_validate_indices_0"), val = tensor(false)]; + tensor var_8165_cast_fp16 = scatter_along_axis(axis = var_8165_axis_0, data = var_8163_cast_fp16, indices = write_indices_41, mode = var_8165_mode_0, updates = v_41_cast_fp16, validate_indices = var_8165_validate_indices_0)[name = tensor("op_8165_cast_fp16")]; + tensor concat_144 = const()[name = tensor("concat_144"), val = tensor([1, 0, 0, 0, 0])]; + tensor concat_145 = const()[name = tensor("concat_145"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_41_internal_tensor_assign_2_stride_0 = const()[name = tensor("new_cache_41_internal_tensor_assign_2_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_41_internal_tensor_assign_2_begin_mask_0 = const()[name = tensor("new_cache_41_internal_tensor_assign_2_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_41_internal_tensor_assign_2_end_mask_0 = const()[name = tensor("new_cache_41_internal_tensor_assign_2_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_41_internal_tensor_assign_2_squeeze_mask_0 = const()[name = tensor("new_cache_41_internal_tensor_assign_2_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_89 = const()[name = tensor("shape_89"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_41 = const()[name = tensor("reduce_prod_41"), val = tensor(1048576)]; + tensor range_1d_41_start_0 = const()[name = tensor("range_1d_41_start_0"), val = tensor(0)]; + tensor range_1d_41_step_0 = const()[name = tensor("range_1d_41_step_0"), val = tensor(1)]; + tensor range_1d_41 = range_1d(end = reduce_prod_41, start = range_1d_41_start_0, step = range_1d_41_step_0)[name = tensor("range_1d_41")]; + tensor reshape_205 = reshape(shape = shape_89, x = range_1d_41)[name = tensor("reshape_205")]; + tensor slice_by_index_41 = slice_by_index(begin = concat_144, begin_mask = new_cache_41_internal_tensor_assign_2_begin_mask_0, end = concat_145, end_mask = new_cache_41_internal_tensor_assign_2_end_mask_0, squeeze_mask = new_cache_41_internal_tensor_assign_2_squeeze_mask_0, stride = new_cache_41_internal_tensor_assign_2_stride_0, x = reshape_205)[name = tensor("slice_by_index_41")]; + tensor reshape_206_shape_0 = const()[name = tensor("reshape_206_shape_0"), val = tensor([-1])]; + tensor reshape_206 = reshape(shape = reshape_206_shape_0, x = slice_by_index_41)[name = tensor("reshape_206")]; + tensor reshape_207_shape_0 = const()[name = tensor("reshape_207_shape_0"), val = tensor([-1])]; + tensor reshape_207_cast_fp16 = reshape(shape = reshape_207_shape_0, x = var_8165_cast_fp16)[name = tensor("reshape_207_cast_fp16")]; + tensor reshape_208_shape_0 = const()[name = tensor("reshape_208_shape_0"), val = tensor([-1])]; + tensor reshape_208_cast_fp16 = reshape(shape = reshape_208_shape_0, x = reshape_204_cast_fp16)[name = tensor("reshape_208_cast_fp16")]; + tensor scatter_41_mode_0 = const()[name = tensor("scatter_41_mode_0"), val = tensor("update")]; + tensor scatter_41_axis_0 = const()[name = tensor("scatter_41_axis_0"), val = tensor(0)]; + tensor scatter_41_validate_indices_0 = const()[name = tensor("scatter_41_validate_indices_0"), val = tensor(false)]; + tensor scatter_41_cast_fp16 = scatter(axis = scatter_41_axis_0, data = reshape_208_cast_fp16, indices = reshape_206, mode = scatter_41_mode_0, updates = reshape_207_cast_fp16, validate_indices = scatter_41_validate_indices_0)[name = tensor("scatter_41_cast_fp16")]; + tensor reshape_209_cast_fp16 = reshape(shape = shape_89, x = scatter_41_cast_fp16)[name = tensor("reshape_209_cast_fp16")]; + tensor reshape_209_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("reshape_209_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor keys_121_begin_0 = const()[name = tensor("keys_121_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor keys_121_end_0 = const()[name = tensor("keys_121_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor keys_121_end_mask_0 = const()[name = tensor("keys_121_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor keys_121_squeeze_mask_0 = const()[name = tensor("keys_121_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor keys_121_cast_fp16 = slice_by_index(begin = keys_121_begin_0, end = keys_121_end_0, end_mask = keys_121_end_mask_0, squeeze_mask = keys_121_squeeze_mask_0, x = reshape_209_cast_fp16)[name = tensor("keys_121_cast_fp16")]; + tensor values_121_begin_0 = const()[name = tensor("values_121_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor values_121_end_0 = const()[name = tensor("values_121_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor values_121_end_mask_0 = const()[name = tensor("values_121_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor values_121_squeeze_mask_0 = const()[name = tensor("values_121_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor values_121_cast_fp16 = slice_by_index(begin = values_121_begin_0, end = values_121_end_0, end_mask = values_121_end_mask_0, squeeze_mask = values_121_squeeze_mask_0, x = reshape_209_cast_fp16)[name = tensor("values_121_cast_fp16")]; + tensor var_8177_cast_fp16 = not_equal(x = keys_121_cast_fp16, y = keys_121_cast_fp16)[name = tensor("op_8177_cast_fp16")]; + tensor keys_123_cast_fp16 = select(a = var_463_to_fp16, b = keys_121_cast_fp16, cond = var_8177_cast_fp16)[name = tensor("keys_123_cast_fp16")]; + tensor var_8185_cast_fp16 = not_equal(x = values_121_cast_fp16, y = values_121_cast_fp16)[name = tensor("op_8185_cast_fp16")]; + tensor values_123_cast_fp16 = select(a = var_463_to_fp16, b = values_121_cast_fp16, cond = var_8185_cast_fp16)[name = tensor("values_123_cast_fp16")]; + tensor var_8209 = const()[name = tensor("op_8209"), val = tensor([0, 2, 1, 3])]; + tensor var_8225 = const()[name = tensor("op_8225"), val = tensor([1, 1, 1])]; + tensor var_8226_cast_fp16 = reshape(shape = var_8225, x = position20_to_fp16)[name = tensor("op_8226_cast_fp16")]; + tensor var_8246_to_fp16 = const()[name = tensor("op_8246_to_fp16"), val = tensor(0x1p+0)]; + tensor valid_len_41_cast_fp16 = add(x = var_8226_cast_fp16, y = var_8246_to_fp16)[name = tensor("valid_len_41_cast_fp16")]; + tensor valid_mask_41_cast_fp16 = less(x = k_positions_1_promoted_to_fp16, y = valid_len_41_cast_fp16)[name = tensor("valid_mask_41_cast_fp16")]; + tensor causal_mask_41_cast_fp16 = less_equal(x = k_positions_1_promoted_to_fp16, y = var_8226_cast_fp16)[name = tensor("causal_mask_41_cast_fp16")]; + tensor attn_mask_81 = logical_and(x = valid_mask_41_cast_fp16, y = causal_mask_41_cast_fp16)[name = tensor("attn_mask_81")]; + tensor attn_mask_83_axes_0 = const()[name = tensor("attn_mask_83_axes_0"), val = tensor([1])]; + tensor attn_mask_83 = expand_dims(axes = attn_mask_83_axes_0, x = attn_mask_81)[name = tensor("attn_mask_83")]; + tensor var_8264_transpose_x_0 = const()[name = tensor("op_8264_transpose_x_0"), val = tensor(false)]; + tensor var_8264_transpose_y_0 = const()[name = tensor("op_8264_transpose_y_0"), val = tensor(false)]; + tensor transpose_112_perm_0 = const()[name = tensor("transpose_112_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_113_perm_0 = const()[name = tensor("transpose_113_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_113 = transpose(perm = transpose_113_perm_0, x = keys_123_cast_fp16)[name = tensor("transpose_133")]; + tensor transpose_112 = transpose(perm = transpose_112_perm_0, x = q_123_cast_fp16)[name = tensor("transpose_134")]; + tensor var_8264_cast_fp16 = matmul(transpose_x = var_8264_transpose_x_0, transpose_y = var_8264_transpose_y_0, x = transpose_112, y = transpose_113)[name = tensor("op_8264_cast_fp16")]; + tensor var_8258_to_fp16 = const()[name = tensor("op_8258_to_fp16"), val = tensor([0x1p-3])]; + tensor attn_weights_121_cast_fp16 = mul(x = var_8264_cast_fp16, y = var_8258_to_fp16)[name = tensor("attn_weights_121_cast_fp16")]; + tensor var_8266 = logical_not(x = attn_mask_83)[name = tensor("op_8266")]; + tensor var_8267_to_fp16 = const()[name = tensor("op_8267_to_fp16"), val = tensor(-inf)]; + tensor attn_weights_123_cast_fp16 = select(a = var_8267_to_fp16, b = attn_weights_121_cast_fp16, cond = var_8266)[name = tensor("attn_weights_123_cast_fp16")]; + tensor var_8269 = const()[name = tensor("op_8269"), val = tensor(-1)]; + tensor attn_weights_125_cast_fp16 = softmax(axis = var_8269, x = attn_weights_123_cast_fp16)[name = tensor("attn_weights_125_cast_fp16")]; + tensor attn_output_41_transpose_x_0 = const()[name = tensor("attn_output_41_transpose_x_0"), val = tensor(false)]; + tensor attn_output_41_transpose_y_0 = const()[name = tensor("attn_output_41_transpose_y_0"), val = tensor(false)]; + tensor values_125_cast_fp16 = transpose(perm = var_8209, x = values_123_cast_fp16)[name = tensor("transpose_135")]; + tensor attn_output_41_cast_fp16 = matmul(transpose_x = attn_output_41_transpose_x_0, transpose_y = attn_output_41_transpose_y_0, x = attn_weights_125_cast_fp16, y = values_125_cast_fp16)[name = tensor("attn_output_41_cast_fp16")]; + tensor var_8277 = const()[name = tensor("op_8277"), val = tensor([0, 2, 1, 3])]; + tensor var_8282 = const()[name = tensor("op_8282"), val = tensor([1, 1, 1024])]; + tensor var_8278_cast_fp16 = transpose(perm = var_8277, x = attn_output_41_cast_fp16)[name = tensor("transpose_132")]; + tensor input_247_cast_fp16 = reshape(shape = var_8282, x = var_8278_cast_fp16)[name = tensor("input_247_cast_fp16")]; + tensor attn20_out_proj_weight_to_fp16 = const()[name = tensor("attn20_out_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(510920960)))]; + tensor linear_82_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = attn20_out_proj_weight_to_fp16, x = input_247_cast_fp16)[name = tensor("linear_82_cast_fp16")]; + tensor var_8288_to_fp16 = const()[name = tensor("op_8288_to_fp16"), val = tensor(0x1p+0)]; + tensor var_8289_cast_fp16 = add(x = position20_to_fp16, y = var_8288_to_fp16)[name = tensor("op_8289_cast_fp16")]; + tensor var_8289_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_8289_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor input_249_cast_fp16 = add(x = input_243_cast_fp16, y = linear_82_cast_fp16)[name = tensor("input_249_cast_fp16")]; + tensor input_251_axes_0 = const()[name = tensor("input_251_axes_0"), val = tensor([-1])]; + tensor norm20_2_weight_to_fp16 = const()[name = tensor("norm20_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(513018176)))]; + tensor norm20_2_bias_to_fp16 = const()[name = tensor("norm20_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(513020288)))]; + tensor var_8293_to_fp16 = const()[name = tensor("op_8293_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_251_cast_fp16 = layer_norm(axes = input_251_axes_0, beta = norm20_2_bias_to_fp16, epsilon = var_8293_to_fp16, gamma = norm20_2_weight_to_fp16, x = input_249_cast_fp16)[name = tensor("input_251_cast_fp16")]; + tensor linear20_1_weight_to_fp16 = const()[name = tensor("linear20_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(513022400)))]; + tensor linear_83_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = linear20_1_weight_to_fp16, x = input_251_cast_fp16)[name = tensor("linear_83_cast_fp16")]; + tensor input_253_mode_0 = const()[name = tensor("input_253_mode_0"), val = tensor("EXACT")]; + tensor input_253_cast_fp16 = gelu(mode = input_253_mode_0, x = linear_83_cast_fp16)[name = tensor("input_253_cast_fp16")]; + tensor linear20_2_weight_to_fp16 = const()[name = tensor("linear20_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(521411072)))]; + tensor linear_84_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = linear20_2_weight_to_fp16, x = input_253_cast_fp16)[name = tensor("linear_84_cast_fp16")]; + tensor input_255_cast_fp16 = add(x = input_249_cast_fp16, y = linear_84_cast_fp16)[name = tensor("input_255_cast_fp16")]; + tensor input_257_axes_0 = const()[name = tensor("input_257_axes_0"), val = tensor([-1])]; + tensor norm21_1_weight_to_fp16 = const()[name = tensor("norm21_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(529799744)))]; + tensor norm21_1_bias_to_fp16 = const()[name = tensor("norm21_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(529801856)))]; + tensor var_8310_to_fp16 = const()[name = tensor("op_8310_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_257_cast_fp16 = layer_norm(axes = input_257_axes_0, beta = norm21_1_bias_to_fp16, epsilon = var_8310_to_fp16, gamma = norm21_1_weight_to_fp16, x = input_255_cast_fp16)[name = tensor("input_257_cast_fp16")]; + tensor attn21_in_proj_weight_to_fp16 = const()[name = tensor("attn21_in_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(529803968)))]; + tensor linear_85_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = attn21_in_proj_weight_to_fp16, x = input_257_cast_fp16)[name = tensor("linear_85_cast_fp16")]; + tensor var_8324 = const()[name = tensor("op_8324"), val = tensor([1, 1, 3, 16, 64])]; + tensor qkv_43_cast_fp16 = reshape(shape = var_8324, x = linear_85_cast_fp16)[name = tensor("qkv_43_cast_fp16")]; + tensor q_127_begin_0 = const()[name = tensor("q_127_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor q_127_end_0 = const()[name = tensor("q_127_end_0"), val = tensor([1, 1, 1, 16, 64])]; + tensor q_127_end_mask_0 = const()[name = tensor("q_127_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor q_127_squeeze_mask_0 = const()[name = tensor("q_127_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor q_127_cast_fp16 = slice_by_index(begin = q_127_begin_0, end = q_127_end_0, end_mask = q_127_end_mask_0, squeeze_mask = q_127_squeeze_mask_0, x = qkv_43_cast_fp16)[name = tensor("q_127_cast_fp16")]; + tensor k_85_begin_0 = const()[name = tensor("k_85_begin_0"), val = tensor([0, 0, 1, 0, 0])]; + tensor k_85_end_0 = const()[name = tensor("k_85_end_0"), val = tensor([1, 1, 2, 16, 64])]; + tensor k_85_end_mask_0 = const()[name = tensor("k_85_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor k_85_squeeze_mask_0 = const()[name = tensor("k_85_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor k_85_cast_fp16 = slice_by_index(begin = k_85_begin_0, end = k_85_end_0, end_mask = k_85_end_mask_0, squeeze_mask = k_85_squeeze_mask_0, x = qkv_43_cast_fp16)[name = tensor("k_85_cast_fp16")]; + tensor v_43_begin_0 = const()[name = tensor("v_43_begin_0"), val = tensor([0, 0, 2, 0, 0])]; + tensor v_43_end_0 = const()[name = tensor("v_43_end_0"), val = tensor([1, 1, 3, 16, 64])]; + tensor v_43_end_mask_0 = const()[name = tensor("v_43_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor v_43_squeeze_mask_0 = const()[name = tensor("v_43_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor v_43_cast_fp16 = slice_by_index(begin = v_43_begin_0, end = v_43_end_0, end_mask = v_43_end_mask_0, squeeze_mask = v_43_squeeze_mask_0, x = qkv_43_cast_fp16)[name = tensor("v_43_cast_fp16")]; + tensor var_8390 = const()[name = tensor("op_8390"), val = tensor([1, 1, 1, 1])]; + tensor position21_to_fp16_dtype_0 = const()[name = tensor("position21_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor position21_to_fp16 = cast(dtype = position21_to_fp16_dtype_0, x = position21)[name = tensor("cast_448")]; + tensor ts_131_cast_fp16 = reshape(shape = var_8390, x = position21_to_fp16)[name = tensor("ts_131_cast_fp16")]; + tensor var_8397 = const()[name = tensor("op_8397"), val = tensor([1, 1, 16, 32, 2])]; + tensor q_complex_43_cast_fp16 = reshape(shape = var_8397, x = q_127_cast_fp16)[name = tensor("q_complex_43_cast_fp16")]; + tensor var_8404 = const()[name = tensor("op_8404"), val = tensor([1, 1, 16, 32, 2])]; + tensor k_complex_43_cast_fp16 = reshape(shape = var_8404, x = k_85_cast_fp16)[name = tensor("k_complex_43_cast_fp16")]; + tensor var_8408_begin_0 = const()[name = tensor("op_8408_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_8408_end_0 = const()[name = tensor("op_8408_end_0"), val = tensor([1, 1, 16, 32, 1])]; + tensor var_8408_end_mask_0 = const()[name = tensor("op_8408_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_8408_squeeze_mask_0 = const()[name = tensor("op_8408_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_8408_cast_fp16 = slice_by_index(begin = var_8408_begin_0, end = var_8408_end_0, end_mask = var_8408_end_mask_0, squeeze_mask = var_8408_squeeze_mask_0, x = q_complex_43_cast_fp16)[name = tensor("op_8408_cast_fp16")]; + tensor var_8416_begin_0 = const()[name = tensor("op_8416_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_8416_end_0 = const()[name = tensor("op_8416_end_0"), val = tensor([1, 1, 16, 32, 2])]; + tensor var_8416_end_mask_0 = const()[name = tensor("op_8416_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_8416_squeeze_mask_0 = const()[name = tensor("op_8416_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_8416_cast_fp16 = slice_by_index(begin = var_8416_begin_0, end = var_8416_end_0, end_mask = var_8416_end_mask_0, squeeze_mask = var_8416_squeeze_mask_0, x = q_complex_43_cast_fp16)[name = tensor("op_8416_cast_fp16")]; + tensor var_8424_begin_0 = const()[name = tensor("op_8424_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_8424_end_0 = const()[name = tensor("op_8424_end_0"), val = tensor([1, 1, 16, 32, 1])]; + tensor var_8424_end_mask_0 = const()[name = tensor("op_8424_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_8424_squeeze_mask_0 = const()[name = tensor("op_8424_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_8424_cast_fp16 = slice_by_index(begin = var_8424_begin_0, end = var_8424_end_0, end_mask = var_8424_end_mask_0, squeeze_mask = var_8424_squeeze_mask_0, x = k_complex_43_cast_fp16)[name = tensor("op_8424_cast_fp16")]; + tensor var_8432_begin_0 = const()[name = tensor("op_8432_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_8432_end_0 = const()[name = tensor("op_8432_end_0"), val = tensor([1, 1, 16, 32, 2])]; + tensor var_8432_end_mask_0 = const()[name = tensor("op_8432_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_8432_squeeze_mask_0 = const()[name = tensor("op_8432_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_8432_cast_fp16 = slice_by_index(begin = var_8432_begin_0, end = var_8432_end_0, end_mask = var_8432_end_mask_0, squeeze_mask = var_8432_squeeze_mask_0, x = k_complex_43_cast_fp16)[name = tensor("op_8432_cast_fp16")]; + tensor freqs_43_to_fp16 = const()[name = tensor("freqs_43_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(536095488)))]; + tensor var_8438_cast_fp16 = mul(x = freqs_43_to_fp16, y = ts_131_cast_fp16)[name = tensor("op_8438_cast_fp16")]; + tensor rotr_43_cast_fp16 = cos(x = var_8438_cast_fp16)[name = tensor("rotr_43_cast_fp16")]; + tensor roti_43_cast_fp16 = sin(x = var_8438_cast_fp16)[name = tensor("roti_43_cast_fp16")]; + tensor var_8442_cast_fp16 = mul(x = var_8408_cast_fp16, y = rotr_43_cast_fp16)[name = tensor("op_8442_cast_fp16")]; + tensor var_8443_cast_fp16 = mul(x = var_8416_cast_fp16, y = roti_43_cast_fp16)[name = tensor("op_8443_cast_fp16")]; + tensor qor_85_cast_fp16 = sub(x = var_8442_cast_fp16, y = var_8443_cast_fp16)[name = tensor("qor_85_cast_fp16")]; + tensor var_8446_cast_fp16 = mul(x = var_8408_cast_fp16, y = roti_43_cast_fp16)[name = tensor("op_8446_cast_fp16")]; + tensor var_8447_cast_fp16 = mul(x = var_8416_cast_fp16, y = rotr_43_cast_fp16)[name = tensor("op_8447_cast_fp16")]; + tensor qoi_85_cast_fp16 = add(x = var_8446_cast_fp16, y = var_8447_cast_fp16)[name = tensor("qoi_85_cast_fp16")]; + tensor var_8450_cast_fp16 = mul(x = var_8424_cast_fp16, y = rotr_43_cast_fp16)[name = tensor("op_8450_cast_fp16")]; + tensor var_8451_cast_fp16 = mul(x = var_8432_cast_fp16, y = roti_43_cast_fp16)[name = tensor("op_8451_cast_fp16")]; + tensor kor_85_cast_fp16 = sub(x = var_8450_cast_fp16, y = var_8451_cast_fp16)[name = tensor("kor_85_cast_fp16")]; + tensor var_8454_cast_fp16 = mul(x = var_8424_cast_fp16, y = roti_43_cast_fp16)[name = tensor("op_8454_cast_fp16")]; + tensor var_8455_cast_fp16 = mul(x = var_8432_cast_fp16, y = rotr_43_cast_fp16)[name = tensor("op_8455_cast_fp16")]; + tensor koi_85_cast_fp16 = add(x = var_8454_cast_fp16, y = var_8455_cast_fp16)[name = tensor("koi_85_cast_fp16")]; + tensor qo_43_axis_0 = const()[name = tensor("qo_43_axis_0"), val = tensor(-1)]; + tensor qo_43_cast_fp16 = stack(axis = qo_43_axis_0, values = (qor_85_cast_fp16, qoi_85_cast_fp16))[name = tensor("qo_43_cast_fp16")]; + tensor ko_43_axis_0 = const()[name = tensor("ko_43_axis_0"), val = tensor(-1)]; + tensor ko_43_cast_fp16 = stack(axis = ko_43_axis_0, values = (kor_85_cast_fp16, koi_85_cast_fp16))[name = tensor("ko_43_cast_fp16")]; + tensor var_8488 = const()[name = tensor("op_8488"), val = tensor([1, 1, 16, 64])]; + tensor q_129_cast_fp16 = reshape(shape = var_8488, x = qo_43_cast_fp16)[name = tensor("q_129_cast_fp16")]; + tensor var_8494 = const()[name = tensor("op_8494"), val = tensor([1, 1, 16, 64])]; + tensor k_87_cast_fp16 = reshape(shape = var_8494, x = ko_43_cast_fp16)[name = tensor("k_87_cast_fp16")]; + tensor _inversed_8519_y_0_to_fp16 = const()[name = tensor("_inversed_8519_y_0_to_fp16"), val = tensor(0x1p-9)]; + tensor _inversed_8519_cast_fp16 = mul(x = ts_131_cast_fp16, y = _inversed_8519_y_0_to_fp16)[name = tensor("_inversed_8519_cast_fp16")]; + tensor var_8520_cast_fp16 = floor(x = _inversed_8519_cast_fp16)[name = tensor("op_8520_cast_fp16")]; + tensor var_8521_to_fp16 = const()[name = tensor("op_8521_to_fp16"), val = tensor(0x1p+9)]; + tensor var_8522_cast_fp16 = mul(x = var_8520_cast_fp16, y = var_8521_to_fp16)[name = tensor("op_8522_cast_fp16")]; + tensor write_indices_float_87_cast_fp16 = sub(x = ts_131_cast_fp16, y = var_8522_cast_fp16)[name = tensor("write_indices_float_87_cast_fp16")]; + tensor var_8529_dtype_0 = const()[name = tensor("op_8529_dtype_0"), val = tensor("int32")]; + tensor write_indices_43_reps_0 = const()[name = tensor("write_indices_43_reps_0"), val = tensor([1, 1, 16, 64])]; + tensor write_indices_float_87_cast_fp16_to_int32 = cast(dtype = var_8529_dtype_0, x = write_indices_float_87_cast_fp16)[name = tensor("cast_447")]; + tensor write_indices_43 = tile(reps = write_indices_43_reps_0, x = write_indices_float_87_cast_fp16_to_int32)[name = tensor("write_indices_43")]; + tensor var_8539_begin_0 = const()[name = tensor("op_8539_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_8539_end_0 = const()[name = tensor("op_8539_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor var_8539_end_mask_0 = const()[name = tensor("op_8539_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_8539_squeeze_mask_0 = const()[name = tensor("op_8539_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor cache21_to_fp16_dtype_0 = const()[name = tensor("cache21_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor cache21_to_fp16 = cast(dtype = cache21_to_fp16_dtype_0, x = cache21)[name = tensor("cast_446")]; + tensor var_8539_cast_fp16 = slice_by_index(begin = var_8539_begin_0, end = var_8539_end_0, end_mask = var_8539_end_mask_0, squeeze_mask = var_8539_squeeze_mask_0, x = cache21_to_fp16)[name = tensor("op_8539_cast_fp16")]; + tensor var_8541_axis_0 = const()[name = tensor("op_8541_axis_0"), val = tensor(1)]; + tensor var_8541_mode_0 = const()[name = tensor("op_8541_mode_0"), val = tensor("update")]; + tensor var_8541_validate_indices_0 = const()[name = tensor("op_8541_validate_indices_0"), val = tensor(false)]; + tensor var_8541_cast_fp16 = scatter_along_axis(axis = var_8541_axis_0, data = var_8539_cast_fp16, indices = write_indices_43, mode = var_8541_mode_0, updates = k_87_cast_fp16, validate_indices = var_8541_validate_indices_0)[name = tensor("op_8541_cast_fp16")]; + tensor concat_149 = const()[name = tensor("concat_149"), val = tensor([0, 0, 0, 0, 0])]; + tensor concat_150 = const()[name = tensor("concat_150"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_43_internal_tensor_assign_1_stride_0 = const()[name = tensor("new_cache_43_internal_tensor_assign_1_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_43_internal_tensor_assign_1_begin_mask_0 = const()[name = tensor("new_cache_43_internal_tensor_assign_1_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_43_internal_tensor_assign_1_end_mask_0 = const()[name = tensor("new_cache_43_internal_tensor_assign_1_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_43_internal_tensor_assign_1_squeeze_mask_0 = const()[name = tensor("new_cache_43_internal_tensor_assign_1_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_90 = const()[name = tensor("shape_90"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_42 = const()[name = tensor("reduce_prod_42"), val = tensor(1048576)]; + tensor range_1d_42_start_0 = const()[name = tensor("range_1d_42_start_0"), val = tensor(0)]; + tensor range_1d_42_step_0 = const()[name = tensor("range_1d_42_step_0"), val = tensor(1)]; + tensor range_1d_42 = range_1d(end = reduce_prod_42, start = range_1d_42_start_0, step = range_1d_42_step_0)[name = tensor("range_1d_42")]; + tensor reshape_210 = reshape(shape = shape_90, x = range_1d_42)[name = tensor("reshape_210")]; + tensor slice_by_index_42 = slice_by_index(begin = concat_149, begin_mask = new_cache_43_internal_tensor_assign_1_begin_mask_0, end = concat_150, end_mask = new_cache_43_internal_tensor_assign_1_end_mask_0, squeeze_mask = new_cache_43_internal_tensor_assign_1_squeeze_mask_0, stride = new_cache_43_internal_tensor_assign_1_stride_0, x = reshape_210)[name = tensor("slice_by_index_42")]; + tensor reshape_211_shape_0 = const()[name = tensor("reshape_211_shape_0"), val = tensor([-1])]; + tensor reshape_211 = reshape(shape = reshape_211_shape_0, x = slice_by_index_42)[name = tensor("reshape_211")]; + tensor reshape_212_shape_0 = const()[name = tensor("reshape_212_shape_0"), val = tensor([-1])]; + tensor reshape_212_cast_fp16 = reshape(shape = reshape_212_shape_0, x = var_8541_cast_fp16)[name = tensor("reshape_212_cast_fp16")]; + tensor reshape_213_shape_0 = const()[name = tensor("reshape_213_shape_0"), val = tensor([-1])]; + tensor reshape_213_cast_fp16 = reshape(shape = reshape_213_shape_0, x = cache21_to_fp16)[name = tensor("reshape_213_cast_fp16")]; + tensor scatter_42_mode_0 = const()[name = tensor("scatter_42_mode_0"), val = tensor("update")]; + tensor scatter_42_axis_0 = const()[name = tensor("scatter_42_axis_0"), val = tensor(0)]; + tensor scatter_42_validate_indices_0 = const()[name = tensor("scatter_42_validate_indices_0"), val = tensor(false)]; + tensor scatter_42_cast_fp16 = scatter(axis = scatter_42_axis_0, data = reshape_213_cast_fp16, indices = reshape_211, mode = scatter_42_mode_0, updates = reshape_212_cast_fp16, validate_indices = scatter_42_validate_indices_0)[name = tensor("scatter_42_cast_fp16")]; + tensor reshape_214_cast_fp16 = reshape(shape = shape_90, x = scatter_42_cast_fp16)[name = tensor("reshape_214_cast_fp16")]; + tensor var_8549_begin_0 = const()[name = tensor("op_8549_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor var_8549_end_0 = const()[name = tensor("op_8549_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor var_8549_end_mask_0 = const()[name = tensor("op_8549_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_8549_squeeze_mask_0 = const()[name = tensor("op_8549_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor var_8549_cast_fp16 = slice_by_index(begin = var_8549_begin_0, end = var_8549_end_0, end_mask = var_8549_end_mask_0, squeeze_mask = var_8549_squeeze_mask_0, x = reshape_214_cast_fp16)[name = tensor("op_8549_cast_fp16")]; + tensor var_8551_axis_0 = const()[name = tensor("op_8551_axis_0"), val = tensor(1)]; + tensor var_8551_mode_0 = const()[name = tensor("op_8551_mode_0"), val = tensor("update")]; + tensor var_8551_validate_indices_0 = const()[name = tensor("op_8551_validate_indices_0"), val = tensor(false)]; + tensor var_8551_cast_fp16 = scatter_along_axis(axis = var_8551_axis_0, data = var_8549_cast_fp16, indices = write_indices_43, mode = var_8551_mode_0, updates = v_43_cast_fp16, validate_indices = var_8551_validate_indices_0)[name = tensor("op_8551_cast_fp16")]; + tensor concat_151 = const()[name = tensor("concat_151"), val = tensor([1, 0, 0, 0, 0])]; + tensor concat_152 = const()[name = tensor("concat_152"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_43_internal_tensor_assign_2_stride_0 = const()[name = tensor("new_cache_43_internal_tensor_assign_2_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_43_internal_tensor_assign_2_begin_mask_0 = const()[name = tensor("new_cache_43_internal_tensor_assign_2_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_43_internal_tensor_assign_2_end_mask_0 = const()[name = tensor("new_cache_43_internal_tensor_assign_2_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_43_internal_tensor_assign_2_squeeze_mask_0 = const()[name = tensor("new_cache_43_internal_tensor_assign_2_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_91 = const()[name = tensor("shape_91"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_43 = const()[name = tensor("reduce_prod_43"), val = tensor(1048576)]; + tensor range_1d_43_start_0 = const()[name = tensor("range_1d_43_start_0"), val = tensor(0)]; + tensor range_1d_43_step_0 = const()[name = tensor("range_1d_43_step_0"), val = tensor(1)]; + tensor range_1d_43 = range_1d(end = reduce_prod_43, start = range_1d_43_start_0, step = range_1d_43_step_0)[name = tensor("range_1d_43")]; + tensor reshape_215 = reshape(shape = shape_91, x = range_1d_43)[name = tensor("reshape_215")]; + tensor slice_by_index_43 = slice_by_index(begin = concat_151, begin_mask = new_cache_43_internal_tensor_assign_2_begin_mask_0, end = concat_152, end_mask = new_cache_43_internal_tensor_assign_2_end_mask_0, squeeze_mask = new_cache_43_internal_tensor_assign_2_squeeze_mask_0, stride = new_cache_43_internal_tensor_assign_2_stride_0, x = reshape_215)[name = tensor("slice_by_index_43")]; + tensor reshape_216_shape_0 = const()[name = tensor("reshape_216_shape_0"), val = tensor([-1])]; + tensor reshape_216 = reshape(shape = reshape_216_shape_0, x = slice_by_index_43)[name = tensor("reshape_216")]; + tensor reshape_217_shape_0 = const()[name = tensor("reshape_217_shape_0"), val = tensor([-1])]; + tensor reshape_217_cast_fp16 = reshape(shape = reshape_217_shape_0, x = var_8551_cast_fp16)[name = tensor("reshape_217_cast_fp16")]; + tensor reshape_218_shape_0 = const()[name = tensor("reshape_218_shape_0"), val = tensor([-1])]; + tensor reshape_218_cast_fp16 = reshape(shape = reshape_218_shape_0, x = reshape_214_cast_fp16)[name = tensor("reshape_218_cast_fp16")]; + tensor scatter_43_mode_0 = const()[name = tensor("scatter_43_mode_0"), val = tensor("update")]; + tensor scatter_43_axis_0 = const()[name = tensor("scatter_43_axis_0"), val = tensor(0)]; + tensor scatter_43_validate_indices_0 = const()[name = tensor("scatter_43_validate_indices_0"), val = tensor(false)]; + tensor scatter_43_cast_fp16 = scatter(axis = scatter_43_axis_0, data = reshape_218_cast_fp16, indices = reshape_216, mode = scatter_43_mode_0, updates = reshape_217_cast_fp16, validate_indices = scatter_43_validate_indices_0)[name = tensor("scatter_43_cast_fp16")]; + tensor reshape_219_cast_fp16 = reshape(shape = shape_91, x = scatter_43_cast_fp16)[name = tensor("reshape_219_cast_fp16")]; + tensor reshape_219_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("reshape_219_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor keys_127_begin_0 = const()[name = tensor("keys_127_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor keys_127_end_0 = const()[name = tensor("keys_127_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor keys_127_end_mask_0 = const()[name = tensor("keys_127_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor keys_127_squeeze_mask_0 = const()[name = tensor("keys_127_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor keys_127_cast_fp16 = slice_by_index(begin = keys_127_begin_0, end = keys_127_end_0, end_mask = keys_127_end_mask_0, squeeze_mask = keys_127_squeeze_mask_0, x = reshape_219_cast_fp16)[name = tensor("keys_127_cast_fp16")]; + tensor values_127_begin_0 = const()[name = tensor("values_127_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor values_127_end_0 = const()[name = tensor("values_127_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor values_127_end_mask_0 = const()[name = tensor("values_127_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor values_127_squeeze_mask_0 = const()[name = tensor("values_127_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor values_127_cast_fp16 = slice_by_index(begin = values_127_begin_0, end = values_127_end_0, end_mask = values_127_end_mask_0, squeeze_mask = values_127_squeeze_mask_0, x = reshape_219_cast_fp16)[name = tensor("values_127_cast_fp16")]; + tensor var_8563_cast_fp16 = not_equal(x = keys_127_cast_fp16, y = keys_127_cast_fp16)[name = tensor("op_8563_cast_fp16")]; + tensor keys_129_cast_fp16 = select(a = var_463_to_fp16, b = keys_127_cast_fp16, cond = var_8563_cast_fp16)[name = tensor("keys_129_cast_fp16")]; + tensor var_8571_cast_fp16 = not_equal(x = values_127_cast_fp16, y = values_127_cast_fp16)[name = tensor("op_8571_cast_fp16")]; + tensor values_129_cast_fp16 = select(a = var_463_to_fp16, b = values_127_cast_fp16, cond = var_8571_cast_fp16)[name = tensor("values_129_cast_fp16")]; + tensor var_8595 = const()[name = tensor("op_8595"), val = tensor([0, 2, 1, 3])]; + tensor var_8611 = const()[name = tensor("op_8611"), val = tensor([1, 1, 1])]; + tensor var_8612_cast_fp16 = reshape(shape = var_8611, x = position21_to_fp16)[name = tensor("op_8612_cast_fp16")]; + tensor var_8632_to_fp16 = const()[name = tensor("op_8632_to_fp16"), val = tensor(0x1p+0)]; + tensor valid_len_43_cast_fp16 = add(x = var_8612_cast_fp16, y = var_8632_to_fp16)[name = tensor("valid_len_43_cast_fp16")]; + tensor valid_mask_43_cast_fp16 = less(x = k_positions_1_promoted_to_fp16, y = valid_len_43_cast_fp16)[name = tensor("valid_mask_43_cast_fp16")]; + tensor causal_mask_43_cast_fp16 = less_equal(x = k_positions_1_promoted_to_fp16, y = var_8612_cast_fp16)[name = tensor("causal_mask_43_cast_fp16")]; + tensor attn_mask_85 = logical_and(x = valid_mask_43_cast_fp16, y = causal_mask_43_cast_fp16)[name = tensor("attn_mask_85")]; + tensor attn_mask_87_axes_0 = const()[name = tensor("attn_mask_87_axes_0"), val = tensor([1])]; + tensor attn_mask_87 = expand_dims(axes = attn_mask_87_axes_0, x = attn_mask_85)[name = tensor("attn_mask_87")]; + tensor var_8650_transpose_x_0 = const()[name = tensor("op_8650_transpose_x_0"), val = tensor(false)]; + tensor var_8650_transpose_y_0 = const()[name = tensor("op_8650_transpose_y_0"), val = tensor(false)]; + tensor transpose_114_perm_0 = const()[name = tensor("transpose_114_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_115_perm_0 = const()[name = tensor("transpose_115_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_115 = transpose(perm = transpose_115_perm_0, x = keys_129_cast_fp16)[name = tensor("transpose_129")]; + tensor transpose_114 = transpose(perm = transpose_114_perm_0, x = q_129_cast_fp16)[name = tensor("transpose_130")]; + tensor var_8650_cast_fp16 = matmul(transpose_x = var_8650_transpose_x_0, transpose_y = var_8650_transpose_y_0, x = transpose_114, y = transpose_115)[name = tensor("op_8650_cast_fp16")]; + tensor var_8644_to_fp16 = const()[name = tensor("op_8644_to_fp16"), val = tensor([0x1p-3])]; + tensor attn_weights_127_cast_fp16 = mul(x = var_8650_cast_fp16, y = var_8644_to_fp16)[name = tensor("attn_weights_127_cast_fp16")]; + tensor var_8652 = logical_not(x = attn_mask_87)[name = tensor("op_8652")]; + tensor var_8653_to_fp16 = const()[name = tensor("op_8653_to_fp16"), val = tensor(-inf)]; + tensor attn_weights_129_cast_fp16 = select(a = var_8653_to_fp16, b = attn_weights_127_cast_fp16, cond = var_8652)[name = tensor("attn_weights_129_cast_fp16")]; + tensor var_8655 = const()[name = tensor("op_8655"), val = tensor(-1)]; + tensor attn_weights_131_cast_fp16 = softmax(axis = var_8655, x = attn_weights_129_cast_fp16)[name = tensor("attn_weights_131_cast_fp16")]; + tensor attn_output_43_transpose_x_0 = const()[name = tensor("attn_output_43_transpose_x_0"), val = tensor(false)]; + tensor attn_output_43_transpose_y_0 = const()[name = tensor("attn_output_43_transpose_y_0"), val = tensor(false)]; + tensor values_131_cast_fp16 = transpose(perm = var_8595, x = values_129_cast_fp16)[name = tensor("transpose_131")]; + tensor attn_output_43_cast_fp16 = matmul(transpose_x = attn_output_43_transpose_x_0, transpose_y = attn_output_43_transpose_y_0, x = attn_weights_131_cast_fp16, y = values_131_cast_fp16)[name = tensor("attn_output_43_cast_fp16")]; + tensor var_8663 = const()[name = tensor("op_8663"), val = tensor([0, 2, 1, 3])]; + tensor var_8668 = const()[name = tensor("op_8668"), val = tensor([1, 1, 1024])]; + tensor var_8664_cast_fp16 = transpose(perm = var_8663, x = attn_output_43_cast_fp16)[name = tensor("transpose_128")]; + tensor input_259_cast_fp16 = reshape(shape = var_8668, x = var_8664_cast_fp16)[name = tensor("input_259_cast_fp16")]; + tensor attn21_out_proj_weight_to_fp16 = const()[name = tensor("attn21_out_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(536095616)))]; + tensor linear_86_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = attn21_out_proj_weight_to_fp16, x = input_259_cast_fp16)[name = tensor("linear_86_cast_fp16")]; + tensor var_8674_to_fp16 = const()[name = tensor("op_8674_to_fp16"), val = tensor(0x1p+0)]; + tensor var_8675_cast_fp16 = add(x = position21_to_fp16, y = var_8674_to_fp16)[name = tensor("op_8675_cast_fp16")]; + tensor var_8675_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_8675_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor input_261_cast_fp16 = add(x = input_255_cast_fp16, y = linear_86_cast_fp16)[name = tensor("input_261_cast_fp16")]; + tensor input_263_axes_0 = const()[name = tensor("input_263_axes_0"), val = tensor([-1])]; + tensor norm21_2_weight_to_fp16 = const()[name = tensor("norm21_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(538192832)))]; + tensor norm21_2_bias_to_fp16 = const()[name = tensor("norm21_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(538194944)))]; + tensor var_8679_to_fp16 = const()[name = tensor("op_8679_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_263_cast_fp16 = layer_norm(axes = input_263_axes_0, beta = norm21_2_bias_to_fp16, epsilon = var_8679_to_fp16, gamma = norm21_2_weight_to_fp16, x = input_261_cast_fp16)[name = tensor("input_263_cast_fp16")]; + tensor linear21_1_weight_to_fp16 = const()[name = tensor("linear21_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(538197056)))]; + tensor linear_87_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = linear21_1_weight_to_fp16, x = input_263_cast_fp16)[name = tensor("linear_87_cast_fp16")]; + tensor input_265_mode_0 = const()[name = tensor("input_265_mode_0"), val = tensor("EXACT")]; + tensor input_265_cast_fp16 = gelu(mode = input_265_mode_0, x = linear_87_cast_fp16)[name = tensor("input_265_cast_fp16")]; + tensor linear21_2_weight_to_fp16 = const()[name = tensor("linear21_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(546585728)))]; + tensor linear_88_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = linear21_2_weight_to_fp16, x = input_265_cast_fp16)[name = tensor("linear_88_cast_fp16")]; + tensor input_267_cast_fp16 = add(x = input_261_cast_fp16, y = linear_88_cast_fp16)[name = tensor("input_267_cast_fp16")]; + tensor input_269_axes_0 = const()[name = tensor("input_269_axes_0"), val = tensor([-1])]; + tensor norm22_1_weight_to_fp16 = const()[name = tensor("norm22_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(554974400)))]; + tensor norm22_1_bias_to_fp16 = const()[name = tensor("norm22_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(554976512)))]; + tensor var_8696_to_fp16 = const()[name = tensor("op_8696_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_269_cast_fp16 = layer_norm(axes = input_269_axes_0, beta = norm22_1_bias_to_fp16, epsilon = var_8696_to_fp16, gamma = norm22_1_weight_to_fp16, x = input_267_cast_fp16)[name = tensor("input_269_cast_fp16")]; + tensor attn22_in_proj_weight_to_fp16 = const()[name = tensor("attn22_in_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(554978624)))]; + tensor linear_89_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = attn22_in_proj_weight_to_fp16, x = input_269_cast_fp16)[name = tensor("linear_89_cast_fp16")]; + tensor var_8710 = const()[name = tensor("op_8710"), val = tensor([1, 1, 3, 16, 64])]; + tensor qkv_45_cast_fp16 = reshape(shape = var_8710, x = linear_89_cast_fp16)[name = tensor("qkv_45_cast_fp16")]; + tensor q_133_begin_0 = const()[name = tensor("q_133_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor q_133_end_0 = const()[name = tensor("q_133_end_0"), val = tensor([1, 1, 1, 16, 64])]; + tensor q_133_end_mask_0 = const()[name = tensor("q_133_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor q_133_squeeze_mask_0 = const()[name = tensor("q_133_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor q_133_cast_fp16 = slice_by_index(begin = q_133_begin_0, end = q_133_end_0, end_mask = q_133_end_mask_0, squeeze_mask = q_133_squeeze_mask_0, x = qkv_45_cast_fp16)[name = tensor("q_133_cast_fp16")]; + tensor k_89_begin_0 = const()[name = tensor("k_89_begin_0"), val = tensor([0, 0, 1, 0, 0])]; + tensor k_89_end_0 = const()[name = tensor("k_89_end_0"), val = tensor([1, 1, 2, 16, 64])]; + tensor k_89_end_mask_0 = const()[name = tensor("k_89_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor k_89_squeeze_mask_0 = const()[name = tensor("k_89_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor k_89_cast_fp16 = slice_by_index(begin = k_89_begin_0, end = k_89_end_0, end_mask = k_89_end_mask_0, squeeze_mask = k_89_squeeze_mask_0, x = qkv_45_cast_fp16)[name = tensor("k_89_cast_fp16")]; + tensor v_45_begin_0 = const()[name = tensor("v_45_begin_0"), val = tensor([0, 0, 2, 0, 0])]; + tensor v_45_end_0 = const()[name = tensor("v_45_end_0"), val = tensor([1, 1, 3, 16, 64])]; + tensor v_45_end_mask_0 = const()[name = tensor("v_45_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor v_45_squeeze_mask_0 = const()[name = tensor("v_45_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor v_45_cast_fp16 = slice_by_index(begin = v_45_begin_0, end = v_45_end_0, end_mask = v_45_end_mask_0, squeeze_mask = v_45_squeeze_mask_0, x = qkv_45_cast_fp16)[name = tensor("v_45_cast_fp16")]; + tensor var_8776 = const()[name = tensor("op_8776"), val = tensor([1, 1, 1, 1])]; + tensor position22_to_fp16_dtype_0 = const()[name = tensor("position22_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor position22_to_fp16 = cast(dtype = position22_to_fp16_dtype_0, x = position22)[name = tensor("cast_443")]; + tensor ts_137_cast_fp16 = reshape(shape = var_8776, x = position22_to_fp16)[name = tensor("ts_137_cast_fp16")]; + tensor var_8783 = const()[name = tensor("op_8783"), val = tensor([1, 1, 16, 32, 2])]; + tensor q_complex_45_cast_fp16 = reshape(shape = var_8783, x = q_133_cast_fp16)[name = tensor("q_complex_45_cast_fp16")]; + tensor var_8790 = const()[name = tensor("op_8790"), val = tensor([1, 1, 16, 32, 2])]; + tensor k_complex_45_cast_fp16 = reshape(shape = var_8790, x = k_89_cast_fp16)[name = tensor("k_complex_45_cast_fp16")]; + tensor var_8794_begin_0 = const()[name = tensor("op_8794_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_8794_end_0 = const()[name = tensor("op_8794_end_0"), val = tensor([1, 1, 16, 32, 1])]; + tensor var_8794_end_mask_0 = const()[name = tensor("op_8794_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_8794_squeeze_mask_0 = const()[name = tensor("op_8794_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_8794_cast_fp16 = slice_by_index(begin = var_8794_begin_0, end = var_8794_end_0, end_mask = var_8794_end_mask_0, squeeze_mask = var_8794_squeeze_mask_0, x = q_complex_45_cast_fp16)[name = tensor("op_8794_cast_fp16")]; + tensor var_8802_begin_0 = const()[name = tensor("op_8802_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_8802_end_0 = const()[name = tensor("op_8802_end_0"), val = tensor([1, 1, 16, 32, 2])]; + tensor var_8802_end_mask_0 = const()[name = tensor("op_8802_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_8802_squeeze_mask_0 = const()[name = tensor("op_8802_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_8802_cast_fp16 = slice_by_index(begin = var_8802_begin_0, end = var_8802_end_0, end_mask = var_8802_end_mask_0, squeeze_mask = var_8802_squeeze_mask_0, x = q_complex_45_cast_fp16)[name = tensor("op_8802_cast_fp16")]; + tensor var_8810_begin_0 = const()[name = tensor("op_8810_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_8810_end_0 = const()[name = tensor("op_8810_end_0"), val = tensor([1, 1, 16, 32, 1])]; + tensor var_8810_end_mask_0 = const()[name = tensor("op_8810_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_8810_squeeze_mask_0 = const()[name = tensor("op_8810_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_8810_cast_fp16 = slice_by_index(begin = var_8810_begin_0, end = var_8810_end_0, end_mask = var_8810_end_mask_0, squeeze_mask = var_8810_squeeze_mask_0, x = k_complex_45_cast_fp16)[name = tensor("op_8810_cast_fp16")]; + tensor var_8818_begin_0 = const()[name = tensor("op_8818_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_8818_end_0 = const()[name = tensor("op_8818_end_0"), val = tensor([1, 1, 16, 32, 2])]; + tensor var_8818_end_mask_0 = const()[name = tensor("op_8818_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_8818_squeeze_mask_0 = const()[name = tensor("op_8818_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_8818_cast_fp16 = slice_by_index(begin = var_8818_begin_0, end = var_8818_end_0, end_mask = var_8818_end_mask_0, squeeze_mask = var_8818_squeeze_mask_0, x = k_complex_45_cast_fp16)[name = tensor("op_8818_cast_fp16")]; + tensor freqs_45_to_fp16 = const()[name = tensor("freqs_45_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(561270144)))]; + tensor var_8824_cast_fp16 = mul(x = freqs_45_to_fp16, y = ts_137_cast_fp16)[name = tensor("op_8824_cast_fp16")]; + tensor rotr_45_cast_fp16 = cos(x = var_8824_cast_fp16)[name = tensor("rotr_45_cast_fp16")]; + tensor roti_45_cast_fp16 = sin(x = var_8824_cast_fp16)[name = tensor("roti_45_cast_fp16")]; + tensor var_8828_cast_fp16 = mul(x = var_8794_cast_fp16, y = rotr_45_cast_fp16)[name = tensor("op_8828_cast_fp16")]; + tensor var_8829_cast_fp16 = mul(x = var_8802_cast_fp16, y = roti_45_cast_fp16)[name = tensor("op_8829_cast_fp16")]; + tensor qor_89_cast_fp16 = sub(x = var_8828_cast_fp16, y = var_8829_cast_fp16)[name = tensor("qor_89_cast_fp16")]; + tensor var_8832_cast_fp16 = mul(x = var_8794_cast_fp16, y = roti_45_cast_fp16)[name = tensor("op_8832_cast_fp16")]; + tensor var_8833_cast_fp16 = mul(x = var_8802_cast_fp16, y = rotr_45_cast_fp16)[name = tensor("op_8833_cast_fp16")]; + tensor qoi_89_cast_fp16 = add(x = var_8832_cast_fp16, y = var_8833_cast_fp16)[name = tensor("qoi_89_cast_fp16")]; + tensor var_8836_cast_fp16 = mul(x = var_8810_cast_fp16, y = rotr_45_cast_fp16)[name = tensor("op_8836_cast_fp16")]; + tensor var_8837_cast_fp16 = mul(x = var_8818_cast_fp16, y = roti_45_cast_fp16)[name = tensor("op_8837_cast_fp16")]; + tensor kor_89_cast_fp16 = sub(x = var_8836_cast_fp16, y = var_8837_cast_fp16)[name = tensor("kor_89_cast_fp16")]; + tensor var_8840_cast_fp16 = mul(x = var_8810_cast_fp16, y = roti_45_cast_fp16)[name = tensor("op_8840_cast_fp16")]; + tensor var_8841_cast_fp16 = mul(x = var_8818_cast_fp16, y = rotr_45_cast_fp16)[name = tensor("op_8841_cast_fp16")]; + tensor koi_89_cast_fp16 = add(x = var_8840_cast_fp16, y = var_8841_cast_fp16)[name = tensor("koi_89_cast_fp16")]; + tensor qo_45_axis_0 = const()[name = tensor("qo_45_axis_0"), val = tensor(-1)]; + tensor qo_45_cast_fp16 = stack(axis = qo_45_axis_0, values = (qor_89_cast_fp16, qoi_89_cast_fp16))[name = tensor("qo_45_cast_fp16")]; + tensor ko_45_axis_0 = const()[name = tensor("ko_45_axis_0"), val = tensor(-1)]; + tensor ko_45_cast_fp16 = stack(axis = ko_45_axis_0, values = (kor_89_cast_fp16, koi_89_cast_fp16))[name = tensor("ko_45_cast_fp16")]; + tensor var_8874 = const()[name = tensor("op_8874"), val = tensor([1, 1, 16, 64])]; + tensor q_135_cast_fp16 = reshape(shape = var_8874, x = qo_45_cast_fp16)[name = tensor("q_135_cast_fp16")]; + tensor var_8880 = const()[name = tensor("op_8880"), val = tensor([1, 1, 16, 64])]; + tensor k_91_cast_fp16 = reshape(shape = var_8880, x = ko_45_cast_fp16)[name = tensor("k_91_cast_fp16")]; + tensor _inversed_8905_y_0_to_fp16 = const()[name = tensor("_inversed_8905_y_0_to_fp16"), val = tensor(0x1p-9)]; + tensor _inversed_8905_cast_fp16 = mul(x = ts_137_cast_fp16, y = _inversed_8905_y_0_to_fp16)[name = tensor("_inversed_8905_cast_fp16")]; + tensor var_8906_cast_fp16 = floor(x = _inversed_8905_cast_fp16)[name = tensor("op_8906_cast_fp16")]; + tensor var_8907_to_fp16 = const()[name = tensor("op_8907_to_fp16"), val = tensor(0x1p+9)]; + tensor var_8908_cast_fp16 = mul(x = var_8906_cast_fp16, y = var_8907_to_fp16)[name = tensor("op_8908_cast_fp16")]; + tensor write_indices_float_91_cast_fp16 = sub(x = ts_137_cast_fp16, y = var_8908_cast_fp16)[name = tensor("write_indices_float_91_cast_fp16")]; + tensor var_8915_dtype_0 = const()[name = tensor("op_8915_dtype_0"), val = tensor("int32")]; + tensor write_indices_45_reps_0 = const()[name = tensor("write_indices_45_reps_0"), val = tensor([1, 1, 16, 64])]; + tensor write_indices_float_91_cast_fp16_to_int32 = cast(dtype = var_8915_dtype_0, x = write_indices_float_91_cast_fp16)[name = tensor("cast_442")]; + tensor write_indices_45 = tile(reps = write_indices_45_reps_0, x = write_indices_float_91_cast_fp16_to_int32)[name = tensor("write_indices_45")]; + tensor var_8925_begin_0 = const()[name = tensor("op_8925_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_8925_end_0 = const()[name = tensor("op_8925_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor var_8925_end_mask_0 = const()[name = tensor("op_8925_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_8925_squeeze_mask_0 = const()[name = tensor("op_8925_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor cache22_to_fp16_dtype_0 = const()[name = tensor("cache22_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor cache22_to_fp16 = cast(dtype = cache22_to_fp16_dtype_0, x = cache22)[name = tensor("cast_441")]; + tensor var_8925_cast_fp16 = slice_by_index(begin = var_8925_begin_0, end = var_8925_end_0, end_mask = var_8925_end_mask_0, squeeze_mask = var_8925_squeeze_mask_0, x = cache22_to_fp16)[name = tensor("op_8925_cast_fp16")]; + tensor var_8927_axis_0 = const()[name = tensor("op_8927_axis_0"), val = tensor(1)]; + tensor var_8927_mode_0 = const()[name = tensor("op_8927_mode_0"), val = tensor("update")]; + tensor var_8927_validate_indices_0 = const()[name = tensor("op_8927_validate_indices_0"), val = tensor(false)]; + tensor var_8927_cast_fp16 = scatter_along_axis(axis = var_8927_axis_0, data = var_8925_cast_fp16, indices = write_indices_45, mode = var_8927_mode_0, updates = k_91_cast_fp16, validate_indices = var_8927_validate_indices_0)[name = tensor("op_8927_cast_fp16")]; + tensor concat_156 = const()[name = tensor("concat_156"), val = tensor([0, 0, 0, 0, 0])]; + tensor concat_157 = const()[name = tensor("concat_157"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_45_internal_tensor_assign_1_stride_0 = const()[name = tensor("new_cache_45_internal_tensor_assign_1_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_45_internal_tensor_assign_1_begin_mask_0 = const()[name = tensor("new_cache_45_internal_tensor_assign_1_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_45_internal_tensor_assign_1_end_mask_0 = const()[name = tensor("new_cache_45_internal_tensor_assign_1_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_45_internal_tensor_assign_1_squeeze_mask_0 = const()[name = tensor("new_cache_45_internal_tensor_assign_1_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_92 = const()[name = tensor("shape_92"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_44 = const()[name = tensor("reduce_prod_44"), val = tensor(1048576)]; + tensor range_1d_44_start_0 = const()[name = tensor("range_1d_44_start_0"), val = tensor(0)]; + tensor range_1d_44_step_0 = const()[name = tensor("range_1d_44_step_0"), val = tensor(1)]; + tensor range_1d_44 = range_1d(end = reduce_prod_44, start = range_1d_44_start_0, step = range_1d_44_step_0)[name = tensor("range_1d_44")]; + tensor reshape_220 = reshape(shape = shape_92, x = range_1d_44)[name = tensor("reshape_220")]; + tensor slice_by_index_44 = slice_by_index(begin = concat_156, begin_mask = new_cache_45_internal_tensor_assign_1_begin_mask_0, end = concat_157, end_mask = new_cache_45_internal_tensor_assign_1_end_mask_0, squeeze_mask = new_cache_45_internal_tensor_assign_1_squeeze_mask_0, stride = new_cache_45_internal_tensor_assign_1_stride_0, x = reshape_220)[name = tensor("slice_by_index_44")]; + tensor reshape_221_shape_0 = const()[name = tensor("reshape_221_shape_0"), val = tensor([-1])]; + tensor reshape_221 = reshape(shape = reshape_221_shape_0, x = slice_by_index_44)[name = tensor("reshape_221")]; + tensor reshape_222_shape_0 = const()[name = tensor("reshape_222_shape_0"), val = tensor([-1])]; + tensor reshape_222_cast_fp16 = reshape(shape = reshape_222_shape_0, x = var_8927_cast_fp16)[name = tensor("reshape_222_cast_fp16")]; + tensor reshape_223_shape_0 = const()[name = tensor("reshape_223_shape_0"), val = tensor([-1])]; + tensor reshape_223_cast_fp16 = reshape(shape = reshape_223_shape_0, x = cache22_to_fp16)[name = tensor("reshape_223_cast_fp16")]; + tensor scatter_44_mode_0 = const()[name = tensor("scatter_44_mode_0"), val = tensor("update")]; + tensor scatter_44_axis_0 = const()[name = tensor("scatter_44_axis_0"), val = tensor(0)]; + tensor scatter_44_validate_indices_0 = const()[name = tensor("scatter_44_validate_indices_0"), val = tensor(false)]; + tensor scatter_44_cast_fp16 = scatter(axis = scatter_44_axis_0, data = reshape_223_cast_fp16, indices = reshape_221, mode = scatter_44_mode_0, updates = reshape_222_cast_fp16, validate_indices = scatter_44_validate_indices_0)[name = tensor("scatter_44_cast_fp16")]; + tensor reshape_224_cast_fp16 = reshape(shape = shape_92, x = scatter_44_cast_fp16)[name = tensor("reshape_224_cast_fp16")]; + tensor var_8935_begin_0 = const()[name = tensor("op_8935_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor var_8935_end_0 = const()[name = tensor("op_8935_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor var_8935_end_mask_0 = const()[name = tensor("op_8935_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_8935_squeeze_mask_0 = const()[name = tensor("op_8935_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor var_8935_cast_fp16 = slice_by_index(begin = var_8935_begin_0, end = var_8935_end_0, end_mask = var_8935_end_mask_0, squeeze_mask = var_8935_squeeze_mask_0, x = reshape_224_cast_fp16)[name = tensor("op_8935_cast_fp16")]; + tensor var_8937_axis_0 = const()[name = tensor("op_8937_axis_0"), val = tensor(1)]; + tensor var_8937_mode_0 = const()[name = tensor("op_8937_mode_0"), val = tensor("update")]; + tensor var_8937_validate_indices_0 = const()[name = tensor("op_8937_validate_indices_0"), val = tensor(false)]; + tensor var_8937_cast_fp16 = scatter_along_axis(axis = var_8937_axis_0, data = var_8935_cast_fp16, indices = write_indices_45, mode = var_8937_mode_0, updates = v_45_cast_fp16, validate_indices = var_8937_validate_indices_0)[name = tensor("op_8937_cast_fp16")]; + tensor concat_158 = const()[name = tensor("concat_158"), val = tensor([1, 0, 0, 0, 0])]; + tensor concat_159 = const()[name = tensor("concat_159"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_45_internal_tensor_assign_2_stride_0 = const()[name = tensor("new_cache_45_internal_tensor_assign_2_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_45_internal_tensor_assign_2_begin_mask_0 = const()[name = tensor("new_cache_45_internal_tensor_assign_2_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_45_internal_tensor_assign_2_end_mask_0 = const()[name = tensor("new_cache_45_internal_tensor_assign_2_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_45_internal_tensor_assign_2_squeeze_mask_0 = const()[name = tensor("new_cache_45_internal_tensor_assign_2_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_93 = const()[name = tensor("shape_93"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_45 = const()[name = tensor("reduce_prod_45"), val = tensor(1048576)]; + tensor range_1d_45_start_0 = const()[name = tensor("range_1d_45_start_0"), val = tensor(0)]; + tensor range_1d_45_step_0 = const()[name = tensor("range_1d_45_step_0"), val = tensor(1)]; + tensor range_1d_45 = range_1d(end = reduce_prod_45, start = range_1d_45_start_0, step = range_1d_45_step_0)[name = tensor("range_1d_45")]; + tensor reshape_225 = reshape(shape = shape_93, x = range_1d_45)[name = tensor("reshape_225")]; + tensor slice_by_index_45 = slice_by_index(begin = concat_158, begin_mask = new_cache_45_internal_tensor_assign_2_begin_mask_0, end = concat_159, end_mask = new_cache_45_internal_tensor_assign_2_end_mask_0, squeeze_mask = new_cache_45_internal_tensor_assign_2_squeeze_mask_0, stride = new_cache_45_internal_tensor_assign_2_stride_0, x = reshape_225)[name = tensor("slice_by_index_45")]; + tensor reshape_226_shape_0 = const()[name = tensor("reshape_226_shape_0"), val = tensor([-1])]; + tensor reshape_226 = reshape(shape = reshape_226_shape_0, x = slice_by_index_45)[name = tensor("reshape_226")]; + tensor reshape_227_shape_0 = const()[name = tensor("reshape_227_shape_0"), val = tensor([-1])]; + tensor reshape_227_cast_fp16 = reshape(shape = reshape_227_shape_0, x = var_8937_cast_fp16)[name = tensor("reshape_227_cast_fp16")]; + tensor reshape_228_shape_0 = const()[name = tensor("reshape_228_shape_0"), val = tensor([-1])]; + tensor reshape_228_cast_fp16 = reshape(shape = reshape_228_shape_0, x = reshape_224_cast_fp16)[name = tensor("reshape_228_cast_fp16")]; + tensor scatter_45_mode_0 = const()[name = tensor("scatter_45_mode_0"), val = tensor("update")]; + tensor scatter_45_axis_0 = const()[name = tensor("scatter_45_axis_0"), val = tensor(0)]; + tensor scatter_45_validate_indices_0 = const()[name = tensor("scatter_45_validate_indices_0"), val = tensor(false)]; + tensor scatter_45_cast_fp16 = scatter(axis = scatter_45_axis_0, data = reshape_228_cast_fp16, indices = reshape_226, mode = scatter_45_mode_0, updates = reshape_227_cast_fp16, validate_indices = scatter_45_validate_indices_0)[name = tensor("scatter_45_cast_fp16")]; + tensor reshape_229_cast_fp16 = reshape(shape = shape_93, x = scatter_45_cast_fp16)[name = tensor("reshape_229_cast_fp16")]; + tensor reshape_229_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("reshape_229_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor keys_133_begin_0 = const()[name = tensor("keys_133_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor keys_133_end_0 = const()[name = tensor("keys_133_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor keys_133_end_mask_0 = const()[name = tensor("keys_133_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor keys_133_squeeze_mask_0 = const()[name = tensor("keys_133_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor keys_133_cast_fp16 = slice_by_index(begin = keys_133_begin_0, end = keys_133_end_0, end_mask = keys_133_end_mask_0, squeeze_mask = keys_133_squeeze_mask_0, x = reshape_229_cast_fp16)[name = tensor("keys_133_cast_fp16")]; + tensor values_133_begin_0 = const()[name = tensor("values_133_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor values_133_end_0 = const()[name = tensor("values_133_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor values_133_end_mask_0 = const()[name = tensor("values_133_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor values_133_squeeze_mask_0 = const()[name = tensor("values_133_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor values_133_cast_fp16 = slice_by_index(begin = values_133_begin_0, end = values_133_end_0, end_mask = values_133_end_mask_0, squeeze_mask = values_133_squeeze_mask_0, x = reshape_229_cast_fp16)[name = tensor("values_133_cast_fp16")]; + tensor var_8949_cast_fp16 = not_equal(x = keys_133_cast_fp16, y = keys_133_cast_fp16)[name = tensor("op_8949_cast_fp16")]; + tensor keys_135_cast_fp16 = select(a = var_463_to_fp16, b = keys_133_cast_fp16, cond = var_8949_cast_fp16)[name = tensor("keys_135_cast_fp16")]; + tensor var_8957_cast_fp16 = not_equal(x = values_133_cast_fp16, y = values_133_cast_fp16)[name = tensor("op_8957_cast_fp16")]; + tensor values_135_cast_fp16 = select(a = var_463_to_fp16, b = values_133_cast_fp16, cond = var_8957_cast_fp16)[name = tensor("values_135_cast_fp16")]; + tensor var_8981 = const()[name = tensor("op_8981"), val = tensor([0, 2, 1, 3])]; + tensor var_8997 = const()[name = tensor("op_8997"), val = tensor([1, 1, 1])]; + tensor var_8998_cast_fp16 = reshape(shape = var_8997, x = position22_to_fp16)[name = tensor("op_8998_cast_fp16")]; + tensor var_9018_to_fp16 = const()[name = tensor("op_9018_to_fp16"), val = tensor(0x1p+0)]; + tensor valid_len_45_cast_fp16 = add(x = var_8998_cast_fp16, y = var_9018_to_fp16)[name = tensor("valid_len_45_cast_fp16")]; + tensor valid_mask_45_cast_fp16 = less(x = k_positions_1_promoted_to_fp16, y = valid_len_45_cast_fp16)[name = tensor("valid_mask_45_cast_fp16")]; + tensor causal_mask_45_cast_fp16 = less_equal(x = k_positions_1_promoted_to_fp16, y = var_8998_cast_fp16)[name = tensor("causal_mask_45_cast_fp16")]; + tensor attn_mask_89 = logical_and(x = valid_mask_45_cast_fp16, y = causal_mask_45_cast_fp16)[name = tensor("attn_mask_89")]; + tensor attn_mask_91_axes_0 = const()[name = tensor("attn_mask_91_axes_0"), val = tensor([1])]; + tensor attn_mask_91 = expand_dims(axes = attn_mask_91_axes_0, x = attn_mask_89)[name = tensor("attn_mask_91")]; + tensor var_9036_transpose_x_0 = const()[name = tensor("op_9036_transpose_x_0"), val = tensor(false)]; + tensor var_9036_transpose_y_0 = const()[name = tensor("op_9036_transpose_y_0"), val = tensor(false)]; + tensor transpose_116_perm_0 = const()[name = tensor("transpose_116_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_117_perm_0 = const()[name = tensor("transpose_117_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_117 = transpose(perm = transpose_117_perm_0, x = keys_135_cast_fp16)[name = tensor("transpose_125")]; + tensor transpose_116 = transpose(perm = transpose_116_perm_0, x = q_135_cast_fp16)[name = tensor("transpose_126")]; + tensor var_9036_cast_fp16 = matmul(transpose_x = var_9036_transpose_x_0, transpose_y = var_9036_transpose_y_0, x = transpose_116, y = transpose_117)[name = tensor("op_9036_cast_fp16")]; + tensor var_9030_to_fp16 = const()[name = tensor("op_9030_to_fp16"), val = tensor([0x1p-3])]; + tensor attn_weights_133_cast_fp16 = mul(x = var_9036_cast_fp16, y = var_9030_to_fp16)[name = tensor("attn_weights_133_cast_fp16")]; + tensor var_9038 = logical_not(x = attn_mask_91)[name = tensor("op_9038")]; + tensor var_9039_to_fp16 = const()[name = tensor("op_9039_to_fp16"), val = tensor(-inf)]; + tensor attn_weights_135_cast_fp16 = select(a = var_9039_to_fp16, b = attn_weights_133_cast_fp16, cond = var_9038)[name = tensor("attn_weights_135_cast_fp16")]; + tensor var_9041 = const()[name = tensor("op_9041"), val = tensor(-1)]; + tensor attn_weights_137_cast_fp16 = softmax(axis = var_9041, x = attn_weights_135_cast_fp16)[name = tensor("attn_weights_137_cast_fp16")]; + tensor attn_output_45_transpose_x_0 = const()[name = tensor("attn_output_45_transpose_x_0"), val = tensor(false)]; + tensor attn_output_45_transpose_y_0 = const()[name = tensor("attn_output_45_transpose_y_0"), val = tensor(false)]; + tensor values_137_cast_fp16 = transpose(perm = var_8981, x = values_135_cast_fp16)[name = tensor("transpose_127")]; + tensor attn_output_45_cast_fp16 = matmul(transpose_x = attn_output_45_transpose_x_0, transpose_y = attn_output_45_transpose_y_0, x = attn_weights_137_cast_fp16, y = values_137_cast_fp16)[name = tensor("attn_output_45_cast_fp16")]; + tensor var_9049 = const()[name = tensor("op_9049"), val = tensor([0, 2, 1, 3])]; + tensor var_9054 = const()[name = tensor("op_9054"), val = tensor([1, 1, 1024])]; + tensor var_9050_cast_fp16 = transpose(perm = var_9049, x = attn_output_45_cast_fp16)[name = tensor("transpose_124")]; + tensor input_271_cast_fp16 = reshape(shape = var_9054, x = var_9050_cast_fp16)[name = tensor("input_271_cast_fp16")]; + tensor attn22_out_proj_weight_to_fp16 = const()[name = tensor("attn22_out_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(561270272)))]; + tensor linear_90_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = attn22_out_proj_weight_to_fp16, x = input_271_cast_fp16)[name = tensor("linear_90_cast_fp16")]; + tensor var_9060_to_fp16 = const()[name = tensor("op_9060_to_fp16"), val = tensor(0x1p+0)]; + tensor var_9061_cast_fp16 = add(x = position22_to_fp16, y = var_9060_to_fp16)[name = tensor("op_9061_cast_fp16")]; + tensor var_9061_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_9061_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor input_273_cast_fp16 = add(x = input_267_cast_fp16, y = linear_90_cast_fp16)[name = tensor("input_273_cast_fp16")]; + tensor input_275_axes_0 = const()[name = tensor("input_275_axes_0"), val = tensor([-1])]; + tensor norm22_2_weight_to_fp16 = const()[name = tensor("norm22_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(563367488)))]; + tensor norm22_2_bias_to_fp16 = const()[name = tensor("norm22_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(563369600)))]; + tensor var_9065_to_fp16 = const()[name = tensor("op_9065_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_275_cast_fp16 = layer_norm(axes = input_275_axes_0, beta = norm22_2_bias_to_fp16, epsilon = var_9065_to_fp16, gamma = norm22_2_weight_to_fp16, x = input_273_cast_fp16)[name = tensor("input_275_cast_fp16")]; + tensor linear22_1_weight_to_fp16 = const()[name = tensor("linear22_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(563371712)))]; + tensor linear_91_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = linear22_1_weight_to_fp16, x = input_275_cast_fp16)[name = tensor("linear_91_cast_fp16")]; + tensor input_277_mode_0 = const()[name = tensor("input_277_mode_0"), val = tensor("EXACT")]; + tensor input_277_cast_fp16 = gelu(mode = input_277_mode_0, x = linear_91_cast_fp16)[name = tensor("input_277_cast_fp16")]; + tensor linear22_2_weight_to_fp16 = const()[name = tensor("linear22_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(571760384)))]; + tensor linear_92_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = linear22_2_weight_to_fp16, x = input_277_cast_fp16)[name = tensor("linear_92_cast_fp16")]; + tensor input_279_cast_fp16 = add(x = input_273_cast_fp16, y = linear_92_cast_fp16)[name = tensor("input_279_cast_fp16")]; + tensor input_281_axes_0 = const()[name = tensor("input_281_axes_0"), val = tensor([-1])]; + tensor norm23_1_weight_to_fp16 = const()[name = tensor("norm23_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(580149056)))]; + tensor norm23_1_bias_to_fp16 = const()[name = tensor("norm23_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(580151168)))]; + tensor var_9082_to_fp16 = const()[name = tensor("op_9082_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_281_cast_fp16 = layer_norm(axes = input_281_axes_0, beta = norm23_1_bias_to_fp16, epsilon = var_9082_to_fp16, gamma = norm23_1_weight_to_fp16, x = input_279_cast_fp16)[name = tensor("input_281_cast_fp16")]; + tensor attn23_in_proj_weight_to_fp16 = const()[name = tensor("attn23_in_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(580153280)))]; + tensor linear_93_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = attn23_in_proj_weight_to_fp16, x = input_281_cast_fp16)[name = tensor("linear_93_cast_fp16")]; + tensor var_9096 = const()[name = tensor("op_9096"), val = tensor([1, 1, 3, 16, 64])]; + tensor qkv_cast_fp16 = reshape(shape = var_9096, x = linear_93_cast_fp16)[name = tensor("qkv_cast_fp16")]; + tensor q_139_begin_0 = const()[name = tensor("q_139_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor q_139_end_0 = const()[name = tensor("q_139_end_0"), val = tensor([1, 1, 1, 16, 64])]; + tensor q_139_end_mask_0 = const()[name = tensor("q_139_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor q_139_squeeze_mask_0 = const()[name = tensor("q_139_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor q_139_cast_fp16 = slice_by_index(begin = q_139_begin_0, end = q_139_end_0, end_mask = q_139_end_mask_0, squeeze_mask = q_139_squeeze_mask_0, x = qkv_cast_fp16)[name = tensor("q_139_cast_fp16")]; + tensor k_93_begin_0 = const()[name = tensor("k_93_begin_0"), val = tensor([0, 0, 1, 0, 0])]; + tensor k_93_end_0 = const()[name = tensor("k_93_end_0"), val = tensor([1, 1, 2, 16, 64])]; + tensor k_93_end_mask_0 = const()[name = tensor("k_93_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor k_93_squeeze_mask_0 = const()[name = tensor("k_93_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor k_93_cast_fp16 = slice_by_index(begin = k_93_begin_0, end = k_93_end_0, end_mask = k_93_end_mask_0, squeeze_mask = k_93_squeeze_mask_0, x = qkv_cast_fp16)[name = tensor("k_93_cast_fp16")]; + tensor v_begin_0 = const()[name = tensor("v_begin_0"), val = tensor([0, 0, 2, 0, 0])]; + tensor v_end_0 = const()[name = tensor("v_end_0"), val = tensor([1, 1, 3, 16, 64])]; + tensor v_end_mask_0 = const()[name = tensor("v_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor v_squeeze_mask_0 = const()[name = tensor("v_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor v_cast_fp16 = slice_by_index(begin = v_begin_0, end = v_end_0, end_mask = v_end_mask_0, squeeze_mask = v_squeeze_mask_0, x = qkv_cast_fp16)[name = tensor("v_cast_fp16")]; + tensor var_9162 = const()[name = tensor("op_9162"), val = tensor([1, 1, 1, 1])]; + tensor position23_to_fp16_dtype_0 = const()[name = tensor("position23_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor position23_to_fp16 = cast(dtype = position23_to_fp16_dtype_0, x = position23)[name = tensor("cast_438")]; + tensor ts_cast_fp16 = reshape(shape = var_9162, x = position23_to_fp16)[name = tensor("ts_cast_fp16")]; + tensor var_9169 = const()[name = tensor("op_9169"), val = tensor([1, 1, 16, 32, 2])]; + tensor q_complex_cast_fp16 = reshape(shape = var_9169, x = q_139_cast_fp16)[name = tensor("q_complex_cast_fp16")]; + tensor var_9176 = const()[name = tensor("op_9176"), val = tensor([1, 1, 16, 32, 2])]; + tensor k_complex_cast_fp16 = reshape(shape = var_9176, x = k_93_cast_fp16)[name = tensor("k_complex_cast_fp16")]; + tensor var_9180_begin_0 = const()[name = tensor("op_9180_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_9180_end_0 = const()[name = tensor("op_9180_end_0"), val = tensor([1, 1, 16, 32, 1])]; + tensor var_9180_end_mask_0 = const()[name = tensor("op_9180_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_9180_squeeze_mask_0 = const()[name = tensor("op_9180_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_9180_cast_fp16 = slice_by_index(begin = var_9180_begin_0, end = var_9180_end_0, end_mask = var_9180_end_mask_0, squeeze_mask = var_9180_squeeze_mask_0, x = q_complex_cast_fp16)[name = tensor("op_9180_cast_fp16")]; + tensor var_9188_begin_0 = const()[name = tensor("op_9188_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_9188_end_0 = const()[name = tensor("op_9188_end_0"), val = tensor([1, 1, 16, 32, 2])]; + tensor var_9188_end_mask_0 = const()[name = tensor("op_9188_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_9188_squeeze_mask_0 = const()[name = tensor("op_9188_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_9188_cast_fp16 = slice_by_index(begin = var_9188_begin_0, end = var_9188_end_0, end_mask = var_9188_end_mask_0, squeeze_mask = var_9188_squeeze_mask_0, x = q_complex_cast_fp16)[name = tensor("op_9188_cast_fp16")]; + tensor var_9196_begin_0 = const()[name = tensor("op_9196_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_9196_end_0 = const()[name = tensor("op_9196_end_0"), val = tensor([1, 1, 16, 32, 1])]; + tensor var_9196_end_mask_0 = const()[name = tensor("op_9196_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_9196_squeeze_mask_0 = const()[name = tensor("op_9196_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_9196_cast_fp16 = slice_by_index(begin = var_9196_begin_0, end = var_9196_end_0, end_mask = var_9196_end_mask_0, squeeze_mask = var_9196_squeeze_mask_0, x = k_complex_cast_fp16)[name = tensor("op_9196_cast_fp16")]; + tensor var_9204_begin_0 = const()[name = tensor("op_9204_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_9204_end_0 = const()[name = tensor("op_9204_end_0"), val = tensor([1, 1, 16, 32, 2])]; + tensor var_9204_end_mask_0 = const()[name = tensor("op_9204_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_9204_squeeze_mask_0 = const()[name = tensor("op_9204_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_9204_cast_fp16 = slice_by_index(begin = var_9204_begin_0, end = var_9204_end_0, end_mask = var_9204_end_mask_0, squeeze_mask = var_9204_squeeze_mask_0, x = k_complex_cast_fp16)[name = tensor("op_9204_cast_fp16")]; + tensor freqs_to_fp16 = const()[name = tensor("freqs_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(586444800)))]; + tensor var_9210_cast_fp16 = mul(x = freqs_to_fp16, y = ts_cast_fp16)[name = tensor("op_9210_cast_fp16")]; + tensor rotr_cast_fp16 = cos(x = var_9210_cast_fp16)[name = tensor("rotr_cast_fp16")]; + tensor roti_cast_fp16 = sin(x = var_9210_cast_fp16)[name = tensor("roti_cast_fp16")]; + tensor var_9214_cast_fp16 = mul(x = var_9180_cast_fp16, y = rotr_cast_fp16)[name = tensor("op_9214_cast_fp16")]; + tensor var_9215_cast_fp16 = mul(x = var_9188_cast_fp16, y = roti_cast_fp16)[name = tensor("op_9215_cast_fp16")]; + tensor qor_93_cast_fp16 = sub(x = var_9214_cast_fp16, y = var_9215_cast_fp16)[name = tensor("qor_93_cast_fp16")]; + tensor var_9218_cast_fp16 = mul(x = var_9180_cast_fp16, y = roti_cast_fp16)[name = tensor("op_9218_cast_fp16")]; + tensor var_9219_cast_fp16 = mul(x = var_9188_cast_fp16, y = rotr_cast_fp16)[name = tensor("op_9219_cast_fp16")]; + tensor qoi_93_cast_fp16 = add(x = var_9218_cast_fp16, y = var_9219_cast_fp16)[name = tensor("qoi_93_cast_fp16")]; + tensor var_9222_cast_fp16 = mul(x = var_9196_cast_fp16, y = rotr_cast_fp16)[name = tensor("op_9222_cast_fp16")]; + tensor var_9223_cast_fp16 = mul(x = var_9204_cast_fp16, y = roti_cast_fp16)[name = tensor("op_9223_cast_fp16")]; + tensor kor_93_cast_fp16 = sub(x = var_9222_cast_fp16, y = var_9223_cast_fp16)[name = tensor("kor_93_cast_fp16")]; + tensor var_9226_cast_fp16 = mul(x = var_9196_cast_fp16, y = roti_cast_fp16)[name = tensor("op_9226_cast_fp16")]; + tensor var_9227_cast_fp16 = mul(x = var_9204_cast_fp16, y = rotr_cast_fp16)[name = tensor("op_9227_cast_fp16")]; + tensor koi_93_cast_fp16 = add(x = var_9226_cast_fp16, y = var_9227_cast_fp16)[name = tensor("koi_93_cast_fp16")]; + tensor qo_axis_0 = const()[name = tensor("qo_axis_0"), val = tensor(-1)]; + tensor qo_cast_fp16 = stack(axis = qo_axis_0, values = (qor_93_cast_fp16, qoi_93_cast_fp16))[name = tensor("qo_cast_fp16")]; + tensor ko_axis_0 = const()[name = tensor("ko_axis_0"), val = tensor(-1)]; + tensor ko_cast_fp16 = stack(axis = ko_axis_0, values = (kor_93_cast_fp16, koi_93_cast_fp16))[name = tensor("ko_cast_fp16")]; + tensor var_9260 = const()[name = tensor("op_9260"), val = tensor([1, 1, 16, 64])]; + tensor q_141_cast_fp16 = reshape(shape = var_9260, x = qo_cast_fp16)[name = tensor("q_141_cast_fp16")]; + tensor var_9266 = const()[name = tensor("op_9266"), val = tensor([1, 1, 16, 64])]; + tensor k_cast_fp16 = reshape(shape = var_9266, x = ko_cast_fp16)[name = tensor("k_cast_fp16")]; + tensor _inversed_9291_y_0_to_fp16 = const()[name = tensor("_inversed_9291_y_0_to_fp16"), val = tensor(0x1p-9)]; + tensor _inversed_9291_cast_fp16 = mul(x = ts_cast_fp16, y = _inversed_9291_y_0_to_fp16)[name = tensor("_inversed_9291_cast_fp16")]; + tensor var_9292_cast_fp16 = floor(x = _inversed_9291_cast_fp16)[name = tensor("op_9292_cast_fp16")]; + tensor var_9293_to_fp16 = const()[name = tensor("op_9293_to_fp16"), val = tensor(0x1p+9)]; + tensor var_9294_cast_fp16 = mul(x = var_9292_cast_fp16, y = var_9293_to_fp16)[name = tensor("op_9294_cast_fp16")]; + tensor write_indices_float_cast_fp16 = sub(x = ts_cast_fp16, y = var_9294_cast_fp16)[name = tensor("write_indices_float_cast_fp16")]; + tensor var_9301_dtype_0 = const()[name = tensor("op_9301_dtype_0"), val = tensor("int32")]; + tensor write_indices_reps_0 = const()[name = tensor("write_indices_reps_0"), val = tensor([1, 1, 16, 64])]; + tensor write_indices_float_cast_fp16_to_int32 = cast(dtype = var_9301_dtype_0, x = write_indices_float_cast_fp16)[name = tensor("cast_437")]; + tensor write_indices = tile(reps = write_indices_reps_0, x = write_indices_float_cast_fp16_to_int32)[name = tensor("write_indices")]; + tensor var_9311_begin_0 = const()[name = tensor("op_9311_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_9311_end_0 = const()[name = tensor("op_9311_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor var_9311_end_mask_0 = const()[name = tensor("op_9311_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_9311_squeeze_mask_0 = const()[name = tensor("op_9311_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor cache23_to_fp16_dtype_0 = const()[name = tensor("cache23_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor cache23_to_fp16 = cast(dtype = cache23_to_fp16_dtype_0, x = cache23)[name = tensor("cast_436")]; + tensor var_9311_cast_fp16 = slice_by_index(begin = var_9311_begin_0, end = var_9311_end_0, end_mask = var_9311_end_mask_0, squeeze_mask = var_9311_squeeze_mask_0, x = cache23_to_fp16)[name = tensor("op_9311_cast_fp16")]; + tensor var_9313_axis_0 = const()[name = tensor("op_9313_axis_0"), val = tensor(1)]; + tensor var_9313_mode_0 = const()[name = tensor("op_9313_mode_0"), val = tensor("update")]; + tensor var_9313_validate_indices_0 = const()[name = tensor("op_9313_validate_indices_0"), val = tensor(false)]; + tensor var_9313_cast_fp16 = scatter_along_axis(axis = var_9313_axis_0, data = var_9311_cast_fp16, indices = write_indices, mode = var_9313_mode_0, updates = k_cast_fp16, validate_indices = var_9313_validate_indices_0)[name = tensor("op_9313_cast_fp16")]; + tensor concat_163 = const()[name = tensor("concat_163"), val = tensor([0, 0, 0, 0, 0])]; + tensor concat_164 = const()[name = tensor("concat_164"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_internal_tensor_assign_1_stride_0 = const()[name = tensor("new_cache_internal_tensor_assign_1_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_internal_tensor_assign_1_begin_mask_0 = const()[name = tensor("new_cache_internal_tensor_assign_1_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_internal_tensor_assign_1_end_mask_0 = const()[name = tensor("new_cache_internal_tensor_assign_1_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_internal_tensor_assign_1_squeeze_mask_0 = const()[name = tensor("new_cache_internal_tensor_assign_1_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_94 = const()[name = tensor("shape_94"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_46 = const()[name = tensor("reduce_prod_46"), val = tensor(1048576)]; + tensor range_1d_46_start_0 = const()[name = tensor("range_1d_46_start_0"), val = tensor(0)]; + tensor range_1d_46_step_0 = const()[name = tensor("range_1d_46_step_0"), val = tensor(1)]; + tensor range_1d_46 = range_1d(end = reduce_prod_46, start = range_1d_46_start_0, step = range_1d_46_step_0)[name = tensor("range_1d_46")]; + tensor reshape_230 = reshape(shape = shape_94, x = range_1d_46)[name = tensor("reshape_230")]; + tensor slice_by_index_46 = slice_by_index(begin = concat_163, begin_mask = new_cache_internal_tensor_assign_1_begin_mask_0, end = concat_164, end_mask = new_cache_internal_tensor_assign_1_end_mask_0, squeeze_mask = new_cache_internal_tensor_assign_1_squeeze_mask_0, stride = new_cache_internal_tensor_assign_1_stride_0, x = reshape_230)[name = tensor("slice_by_index_46")]; + tensor reshape_231_shape_0 = const()[name = tensor("reshape_231_shape_0"), val = tensor([-1])]; + tensor reshape_231 = reshape(shape = reshape_231_shape_0, x = slice_by_index_46)[name = tensor("reshape_231")]; + tensor reshape_232_shape_0 = const()[name = tensor("reshape_232_shape_0"), val = tensor([-1])]; + tensor reshape_232_cast_fp16 = reshape(shape = reshape_232_shape_0, x = var_9313_cast_fp16)[name = tensor("reshape_232_cast_fp16")]; + tensor reshape_233_shape_0 = const()[name = tensor("reshape_233_shape_0"), val = tensor([-1])]; + tensor reshape_233_cast_fp16 = reshape(shape = reshape_233_shape_0, x = cache23_to_fp16)[name = tensor("reshape_233_cast_fp16")]; + tensor scatter_46_mode_0 = const()[name = tensor("scatter_46_mode_0"), val = tensor("update")]; + tensor scatter_46_axis_0 = const()[name = tensor("scatter_46_axis_0"), val = tensor(0)]; + tensor scatter_46_validate_indices_0 = const()[name = tensor("scatter_46_validate_indices_0"), val = tensor(false)]; + tensor scatter_46_cast_fp16 = scatter(axis = scatter_46_axis_0, data = reshape_233_cast_fp16, indices = reshape_231, mode = scatter_46_mode_0, updates = reshape_232_cast_fp16, validate_indices = scatter_46_validate_indices_0)[name = tensor("scatter_46_cast_fp16")]; + tensor reshape_234_cast_fp16 = reshape(shape = shape_94, x = scatter_46_cast_fp16)[name = tensor("reshape_234_cast_fp16")]; + tensor var_9321_begin_0 = const()[name = tensor("op_9321_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor var_9321_end_0 = const()[name = tensor("op_9321_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor var_9321_end_mask_0 = const()[name = tensor("op_9321_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_9321_squeeze_mask_0 = const()[name = tensor("op_9321_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor var_9321_cast_fp16 = slice_by_index(begin = var_9321_begin_0, end = var_9321_end_0, end_mask = var_9321_end_mask_0, squeeze_mask = var_9321_squeeze_mask_0, x = reshape_234_cast_fp16)[name = tensor("op_9321_cast_fp16")]; + tensor var_9323_axis_0 = const()[name = tensor("op_9323_axis_0"), val = tensor(1)]; + tensor var_9323_mode_0 = const()[name = tensor("op_9323_mode_0"), val = tensor("update")]; + tensor var_9323_validate_indices_0 = const()[name = tensor("op_9323_validate_indices_0"), val = tensor(false)]; + tensor var_9323_cast_fp16 = scatter_along_axis(axis = var_9323_axis_0, data = var_9321_cast_fp16, indices = write_indices, mode = var_9323_mode_0, updates = v_cast_fp16, validate_indices = var_9323_validate_indices_0)[name = tensor("op_9323_cast_fp16")]; + tensor concat_165 = const()[name = tensor("concat_165"), val = tensor([1, 0, 0, 0, 0])]; + tensor concat_166 = const()[name = tensor("concat_166"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_internal_tensor_assign_2_stride_0 = const()[name = tensor("new_cache_internal_tensor_assign_2_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_internal_tensor_assign_2_begin_mask_0 = const()[name = tensor("new_cache_internal_tensor_assign_2_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_internal_tensor_assign_2_end_mask_0 = const()[name = tensor("new_cache_internal_tensor_assign_2_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_internal_tensor_assign_2_squeeze_mask_0 = const()[name = tensor("new_cache_internal_tensor_assign_2_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_95 = const()[name = tensor("shape_95"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_47 = const()[name = tensor("reduce_prod_47"), val = tensor(1048576)]; + tensor range_1d_47_start_0 = const()[name = tensor("range_1d_47_start_0"), val = tensor(0)]; + tensor range_1d_47_step_0 = const()[name = tensor("range_1d_47_step_0"), val = tensor(1)]; + tensor range_1d_47 = range_1d(end = reduce_prod_47, start = range_1d_47_start_0, step = range_1d_47_step_0)[name = tensor("range_1d_47")]; + tensor reshape_235 = reshape(shape = shape_95, x = range_1d_47)[name = tensor("reshape_235")]; + tensor slice_by_index_47 = slice_by_index(begin = concat_165, begin_mask = new_cache_internal_tensor_assign_2_begin_mask_0, end = concat_166, end_mask = new_cache_internal_tensor_assign_2_end_mask_0, squeeze_mask = new_cache_internal_tensor_assign_2_squeeze_mask_0, stride = new_cache_internal_tensor_assign_2_stride_0, x = reshape_235)[name = tensor("slice_by_index_47")]; + tensor reshape_236_shape_0 = const()[name = tensor("reshape_236_shape_0"), val = tensor([-1])]; + tensor reshape_236 = reshape(shape = reshape_236_shape_0, x = slice_by_index_47)[name = tensor("reshape_236")]; + tensor reshape_237_shape_0 = const()[name = tensor("reshape_237_shape_0"), val = tensor([-1])]; + tensor reshape_237_cast_fp16 = reshape(shape = reshape_237_shape_0, x = var_9323_cast_fp16)[name = tensor("reshape_237_cast_fp16")]; + tensor reshape_238_shape_0 = const()[name = tensor("reshape_238_shape_0"), val = tensor([-1])]; + tensor reshape_238_cast_fp16 = reshape(shape = reshape_238_shape_0, x = reshape_234_cast_fp16)[name = tensor("reshape_238_cast_fp16")]; + tensor scatter_47_mode_0 = const()[name = tensor("scatter_47_mode_0"), val = tensor("update")]; + tensor scatter_47_axis_0 = const()[name = tensor("scatter_47_axis_0"), val = tensor(0)]; + tensor scatter_47_validate_indices_0 = const()[name = tensor("scatter_47_validate_indices_0"), val = tensor(false)]; + tensor scatter_47_cast_fp16 = scatter(axis = scatter_47_axis_0, data = reshape_238_cast_fp16, indices = reshape_236, mode = scatter_47_mode_0, updates = reshape_237_cast_fp16, validate_indices = scatter_47_validate_indices_0)[name = tensor("scatter_47_cast_fp16")]; + tensor reshape_239_cast_fp16 = reshape(shape = shape_95, x = scatter_47_cast_fp16)[name = tensor("reshape_239_cast_fp16")]; + tensor reshape_239_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("reshape_239_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor keys_139_begin_0 = const()[name = tensor("keys_139_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor keys_139_end_0 = const()[name = tensor("keys_139_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor keys_139_end_mask_0 = const()[name = tensor("keys_139_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor keys_139_squeeze_mask_0 = const()[name = tensor("keys_139_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor keys_139_cast_fp16 = slice_by_index(begin = keys_139_begin_0, end = keys_139_end_0, end_mask = keys_139_end_mask_0, squeeze_mask = keys_139_squeeze_mask_0, x = reshape_239_cast_fp16)[name = tensor("keys_139_cast_fp16")]; + tensor values_139_begin_0 = const()[name = tensor("values_139_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor values_139_end_0 = const()[name = tensor("values_139_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor values_139_end_mask_0 = const()[name = tensor("values_139_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor values_139_squeeze_mask_0 = const()[name = tensor("values_139_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor values_139_cast_fp16 = slice_by_index(begin = values_139_begin_0, end = values_139_end_0, end_mask = values_139_end_mask_0, squeeze_mask = values_139_squeeze_mask_0, x = reshape_239_cast_fp16)[name = tensor("values_139_cast_fp16")]; + tensor var_9335_cast_fp16 = not_equal(x = keys_139_cast_fp16, y = keys_139_cast_fp16)[name = tensor("op_9335_cast_fp16")]; + tensor keys_141_cast_fp16 = select(a = var_463_to_fp16, b = keys_139_cast_fp16, cond = var_9335_cast_fp16)[name = tensor("keys_141_cast_fp16")]; + tensor var_9343_cast_fp16 = not_equal(x = values_139_cast_fp16, y = values_139_cast_fp16)[name = tensor("op_9343_cast_fp16")]; + tensor values_141_cast_fp16 = select(a = var_463_to_fp16, b = values_139_cast_fp16, cond = var_9343_cast_fp16)[name = tensor("values_141_cast_fp16")]; + tensor var_9367 = const()[name = tensor("op_9367"), val = tensor([0, 2, 1, 3])]; + tensor var_9383 = const()[name = tensor("op_9383"), val = tensor([1, 1, 1])]; + tensor var_9384_cast_fp16 = reshape(shape = var_9383, x = position23_to_fp16)[name = tensor("op_9384_cast_fp16")]; + tensor var_9404_to_fp16 = const()[name = tensor("op_9404_to_fp16"), val = tensor(0x1p+0)]; + tensor valid_len_cast_fp16 = add(x = var_9384_cast_fp16, y = var_9404_to_fp16)[name = tensor("valid_len_cast_fp16")]; + tensor valid_mask_cast_fp16 = less(x = k_positions_1_promoted_to_fp16, y = valid_len_cast_fp16)[name = tensor("valid_mask_cast_fp16")]; + tensor causal_mask_cast_fp16 = less_equal(x = k_positions_1_promoted_to_fp16, y = var_9384_cast_fp16)[name = tensor("causal_mask_cast_fp16")]; + tensor attn_mask_93 = logical_and(x = valid_mask_cast_fp16, y = causal_mask_cast_fp16)[name = tensor("attn_mask_93")]; + tensor attn_mask_axes_0 = const()[name = tensor("attn_mask_axes_0"), val = tensor([1])]; + tensor attn_mask = expand_dims(axes = attn_mask_axes_0, x = attn_mask_93)[name = tensor("attn_mask")]; + tensor var_9422_transpose_x_0 = const()[name = tensor("op_9422_transpose_x_0"), val = tensor(false)]; + tensor var_9422_transpose_y_0 = const()[name = tensor("op_9422_transpose_y_0"), val = tensor(false)]; + tensor transpose_118_perm_0 = const()[name = tensor("transpose_118_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_119_perm_0 = const()[name = tensor("transpose_119_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_119 = transpose(perm = transpose_119_perm_0, x = keys_141_cast_fp16)[name = tensor("transpose_121")]; + tensor transpose_118 = transpose(perm = transpose_118_perm_0, x = q_141_cast_fp16)[name = tensor("transpose_122")]; + tensor var_9422_cast_fp16 = matmul(transpose_x = var_9422_transpose_x_0, transpose_y = var_9422_transpose_y_0, x = transpose_118, y = transpose_119)[name = tensor("op_9422_cast_fp16")]; + tensor var_9416_to_fp16 = const()[name = tensor("op_9416_to_fp16"), val = tensor([0x1p-3])]; + tensor attn_weights_139_cast_fp16 = mul(x = var_9422_cast_fp16, y = var_9416_to_fp16)[name = tensor("attn_weights_139_cast_fp16")]; + tensor var_9424 = logical_not(x = attn_mask)[name = tensor("op_9424")]; + tensor var_9425_to_fp16 = const()[name = tensor("op_9425_to_fp16"), val = tensor(-inf)]; + tensor attn_weights_141_cast_fp16 = select(a = var_9425_to_fp16, b = attn_weights_139_cast_fp16, cond = var_9424)[name = tensor("attn_weights_141_cast_fp16")]; + tensor var_9427 = const()[name = tensor("op_9427"), val = tensor(-1)]; + tensor attn_weights_cast_fp16 = softmax(axis = var_9427, x = attn_weights_141_cast_fp16)[name = tensor("attn_weights_cast_fp16")]; + tensor attn_output_transpose_x_0 = const()[name = tensor("attn_output_transpose_x_0"), val = tensor(false)]; + tensor attn_output_transpose_y_0 = const()[name = tensor("attn_output_transpose_y_0"), val = tensor(false)]; + tensor values_cast_fp16 = transpose(perm = var_9367, x = values_141_cast_fp16)[name = tensor("transpose_123")]; + tensor attn_output_cast_fp16 = matmul(transpose_x = attn_output_transpose_x_0, transpose_y = attn_output_transpose_y_0, x = attn_weights_cast_fp16, y = values_cast_fp16)[name = tensor("attn_output_cast_fp16")]; + tensor var_9435 = const()[name = tensor("op_9435"), val = tensor([0, 2, 1, 3])]; + tensor var_9440 = const()[name = tensor("op_9440"), val = tensor([1, 1, 1024])]; + tensor var_9436_cast_fp16 = transpose(perm = var_9435, x = attn_output_cast_fp16)[name = tensor("transpose_120")]; + tensor input_283_cast_fp16 = reshape(shape = var_9440, x = var_9436_cast_fp16)[name = tensor("input_283_cast_fp16")]; + tensor attn23_out_proj_weight_to_fp16 = const()[name = tensor("attn23_out_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(586444928)))]; + tensor linear_94_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = attn23_out_proj_weight_to_fp16, x = input_283_cast_fp16)[name = tensor("linear_94_cast_fp16")]; + tensor var_9446_to_fp16 = const()[name = tensor("op_9446_to_fp16"), val = tensor(0x1p+0)]; + tensor var_9447_cast_fp16 = add(x = position23_to_fp16, y = var_9446_to_fp16)[name = tensor("op_9447_cast_fp16")]; + tensor var_9447_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_9447_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor input_285_cast_fp16 = add(x = input_279_cast_fp16, y = linear_94_cast_fp16)[name = tensor("input_285_cast_fp16")]; + tensor input_287_axes_0 = const()[name = tensor("input_287_axes_0"), val = tensor([-1])]; + tensor norm23_2_weight_to_fp16 = const()[name = tensor("norm23_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(588542144)))]; + tensor norm23_2_bias_to_fp16 = const()[name = tensor("norm23_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(588544256)))]; + tensor var_9451_to_fp16 = const()[name = tensor("op_9451_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_287_cast_fp16 = layer_norm(axes = input_287_axes_0, beta = norm23_2_bias_to_fp16, epsilon = var_9451_to_fp16, gamma = norm23_2_weight_to_fp16, x = input_285_cast_fp16)[name = tensor("input_287_cast_fp16")]; + tensor linear23_1_weight_to_fp16 = const()[name = tensor("linear23_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(588546368)))]; + tensor linear_95_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = linear23_1_weight_to_fp16, x = input_287_cast_fp16)[name = tensor("linear_95_cast_fp16")]; + tensor input_289_mode_0 = const()[name = tensor("input_289_mode_0"), val = tensor("EXACT")]; + tensor input_289_cast_fp16 = gelu(mode = input_289_mode_0, x = linear_95_cast_fp16)[name = tensor("input_289_cast_fp16")]; + tensor linear23_2_weight_to_fp16 = const()[name = tensor("linear23_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(596935040)))]; + tensor linear_96_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = linear23_2_weight_to_fp16, x = input_289_cast_fp16)[name = tensor("linear_96_cast_fp16")]; + tensor input_291_cast_fp16 = add(x = input_285_cast_fp16, y = linear_96_cast_fp16)[name = tensor("input_291_cast_fp16")]; + tensor input_axes_0 = const()[name = tensor("input_axes_0"), val = tensor([-1])]; + tensor out_norm_weight_to_fp16 = const()[name = tensor("out_norm_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(605323712)))]; + tensor out_norm_bias_to_fp16 = const()[name = tensor("out_norm_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(605325824)))]; + tensor var_9468_to_fp16 = const()[name = tensor("op_9468_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_cast_fp16 = layer_norm(axes = input_axes_0, beta = out_norm_bias_to_fp16, epsilon = var_9468_to_fp16, gamma = out_norm_weight_to_fp16, x = input_291_cast_fp16)[name = tensor("input_cast_fp16")]; + tensor input_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("input_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor out_eos_weight_to_fp16 = const()[name = tensor("out_eos_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(605327936)))]; + tensor out_eos_bias_to_fp16 = const()[name = tensor("out_eos_bias_to_fp16"), val = tensor([-0x1.5ap-3])]; + tensor linear_97_cast_fp16 = linear(bias = out_eos_bias_to_fp16, weight = out_eos_weight_to_fp16, x = input_cast_fp16)[name = tensor("linear_97_cast_fp16")]; + tensor linear_97_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("linear_97_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor var_9476 = cast(dtype = linear_97_cast_fp16_to_fp32_dtype_0, x = linear_97_cast_fp16)[name = tensor("cast_432")]; + tensor input = cast(dtype = input_cast_fp16_to_fp32_dtype_0, x = input_cast_fp16)[name = tensor("cast_433")]; + tensor var_9447 = cast(dtype = var_9447_cast_fp16_to_fp32_dtype_0, x = var_9447_cast_fp16)[name = tensor("cast_434")]; + tensor new_cache_internal_tensor_assign_2 = cast(dtype = reshape_239_cast_fp16_to_fp32_dtype_0, x = reshape_239_cast_fp16)[name = tensor("cast_435")]; + tensor var_9061 = cast(dtype = var_9061_cast_fp16_to_fp32_dtype_0, x = var_9061_cast_fp16)[name = tensor("cast_439")]; + tensor new_cache_45_internal_tensor_assign_2 = cast(dtype = reshape_229_cast_fp16_to_fp32_dtype_0, x = reshape_229_cast_fp16)[name = tensor("cast_440")]; + tensor var_8675 = cast(dtype = var_8675_cast_fp16_to_fp32_dtype_0, x = var_8675_cast_fp16)[name = tensor("cast_444")]; + tensor new_cache_43_internal_tensor_assign_2 = cast(dtype = reshape_219_cast_fp16_to_fp32_dtype_0, x = reshape_219_cast_fp16)[name = tensor("cast_445")]; + tensor var_8289 = cast(dtype = var_8289_cast_fp16_to_fp32_dtype_0, x = var_8289_cast_fp16)[name = tensor("cast_449")]; + tensor new_cache_41_internal_tensor_assign_2 = cast(dtype = reshape_209_cast_fp16_to_fp32_dtype_0, x = reshape_209_cast_fp16)[name = tensor("cast_450")]; + tensor var_7903 = cast(dtype = var_7903_cast_fp16_to_fp32_dtype_0, x = var_7903_cast_fp16)[name = tensor("cast_454")]; + tensor new_cache_39_internal_tensor_assign_2 = cast(dtype = reshape_199_cast_fp16_to_fp32_dtype_0, x = reshape_199_cast_fp16)[name = tensor("cast_455")]; + tensor var_7517 = cast(dtype = var_7517_cast_fp16_to_fp32_dtype_0, x = var_7517_cast_fp16)[name = tensor("cast_459")]; + tensor new_cache_37_internal_tensor_assign_2 = cast(dtype = reshape_189_cast_fp16_to_fp32_dtype_0, x = reshape_189_cast_fp16)[name = tensor("cast_460")]; + tensor var_7131 = cast(dtype = var_7131_cast_fp16_to_fp32_dtype_0, x = var_7131_cast_fp16)[name = tensor("cast_464")]; + tensor new_cache_35_internal_tensor_assign_2 = cast(dtype = reshape_179_cast_fp16_to_fp32_dtype_0, x = reshape_179_cast_fp16)[name = tensor("cast_465")]; + tensor var_6745 = cast(dtype = var_6745_cast_fp16_to_fp32_dtype_0, x = var_6745_cast_fp16)[name = tensor("cast_469")]; + tensor new_cache_33_internal_tensor_assign_2 = cast(dtype = reshape_169_cast_fp16_to_fp32_dtype_0, x = reshape_169_cast_fp16)[name = tensor("cast_470")]; + tensor var_6359 = cast(dtype = var_6359_cast_fp16_to_fp32_dtype_0, x = var_6359_cast_fp16)[name = tensor("cast_474")]; + tensor new_cache_31_internal_tensor_assign_2 = cast(dtype = reshape_159_cast_fp16_to_fp32_dtype_0, x = reshape_159_cast_fp16)[name = tensor("cast_475")]; + tensor var_5973 = cast(dtype = var_5973_cast_fp16_to_fp32_dtype_0, x = var_5973_cast_fp16)[name = tensor("cast_479")]; + tensor new_cache_29_internal_tensor_assign_2 = cast(dtype = reshape_149_cast_fp16_to_fp32_dtype_0, x = reshape_149_cast_fp16)[name = tensor("cast_480")]; + tensor var_5587 = cast(dtype = var_5587_cast_fp16_to_fp32_dtype_0, x = var_5587_cast_fp16)[name = tensor("cast_484")]; + tensor new_cache_27_internal_tensor_assign_2 = cast(dtype = reshape_139_cast_fp16_to_fp32_dtype_0, x = reshape_139_cast_fp16)[name = tensor("cast_485")]; + tensor var_5201 = cast(dtype = var_5201_cast_fp16_to_fp32_dtype_0, x = var_5201_cast_fp16)[name = tensor("cast_489")]; + tensor new_cache_25_internal_tensor_assign_2 = cast(dtype = reshape_129_cast_fp16_to_fp32_dtype_0, x = reshape_129_cast_fp16)[name = tensor("cast_490")]; + tensor var_4815 = cast(dtype = var_4815_cast_fp16_to_fp32_dtype_0, x = var_4815_cast_fp16)[name = tensor("cast_494")]; + tensor new_cache_23_internal_tensor_assign_2 = cast(dtype = reshape_119_cast_fp16_to_fp32_dtype_0, x = reshape_119_cast_fp16)[name = tensor("cast_495")]; + tensor var_4429 = cast(dtype = var_4429_cast_fp16_to_fp32_dtype_0, x = var_4429_cast_fp16)[name = tensor("cast_499")]; + tensor new_cache_21_internal_tensor_assign_2 = cast(dtype = reshape_109_cast_fp16_to_fp32_dtype_0, x = reshape_109_cast_fp16)[name = tensor("cast_500")]; + tensor var_4043 = cast(dtype = var_4043_cast_fp16_to_fp32_dtype_0, x = var_4043_cast_fp16)[name = tensor("cast_504")]; + tensor new_cache_19_internal_tensor_assign_2 = cast(dtype = reshape_99_cast_fp16_to_fp32_dtype_0, x = reshape_99_cast_fp16)[name = tensor("cast_505")]; + tensor var_3657 = cast(dtype = var_3657_cast_fp16_to_fp32_dtype_0, x = var_3657_cast_fp16)[name = tensor("cast_509")]; + tensor new_cache_17_internal_tensor_assign_2 = cast(dtype = reshape_89_cast_fp16_to_fp32_dtype_0, x = reshape_89_cast_fp16)[name = tensor("cast_510")]; + tensor var_3271 = cast(dtype = var_3271_cast_fp16_to_fp32_dtype_0, x = var_3271_cast_fp16)[name = tensor("cast_514")]; + tensor new_cache_15_internal_tensor_assign_2 = cast(dtype = reshape_79_cast_fp16_to_fp32_dtype_0, x = reshape_79_cast_fp16)[name = tensor("cast_515")]; + tensor var_2885 = cast(dtype = var_2885_cast_fp16_to_fp32_dtype_0, x = var_2885_cast_fp16)[name = tensor("cast_519")]; + tensor new_cache_13_internal_tensor_assign_2 = cast(dtype = reshape_69_cast_fp16_to_fp32_dtype_0, x = reshape_69_cast_fp16)[name = tensor("cast_520")]; + tensor var_2499 = cast(dtype = var_2499_cast_fp16_to_fp32_dtype_0, x = var_2499_cast_fp16)[name = tensor("cast_524")]; + tensor new_cache_11_internal_tensor_assign_2 = cast(dtype = reshape_59_cast_fp16_to_fp32_dtype_0, x = reshape_59_cast_fp16)[name = tensor("cast_525")]; + tensor var_2113 = cast(dtype = var_2113_cast_fp16_to_fp32_dtype_0, x = var_2113_cast_fp16)[name = tensor("cast_529")]; + tensor new_cache_9_internal_tensor_assign_2 = cast(dtype = reshape_49_cast_fp16_to_fp32_dtype_0, x = reshape_49_cast_fp16)[name = tensor("cast_530")]; + tensor var_1727 = cast(dtype = var_1727_cast_fp16_to_fp32_dtype_0, x = var_1727_cast_fp16)[name = tensor("cast_534")]; + tensor new_cache_7_internal_tensor_assign_2 = cast(dtype = reshape_39_cast_fp16_to_fp32_dtype_0, x = reshape_39_cast_fp16)[name = tensor("cast_535")]; + tensor var_1341 = cast(dtype = var_1341_cast_fp16_to_fp32_dtype_0, x = var_1341_cast_fp16)[name = tensor("cast_539")]; + tensor new_cache_5_internal_tensor_assign_2 = cast(dtype = reshape_29_cast_fp16_to_fp32_dtype_0, x = reshape_29_cast_fp16)[name = tensor("cast_540")]; + tensor var_955 = cast(dtype = var_955_cast_fp16_to_fp32_dtype_0, x = var_955_cast_fp16)[name = tensor("cast_544")]; + tensor new_cache_3_internal_tensor_assign_2 = cast(dtype = reshape_19_cast_fp16_to_fp32_dtype_0, x = reshape_19_cast_fp16)[name = tensor("cast_545")]; + tensor var_569 = cast(dtype = var_569_cast_fp16_to_fp32_dtype_0, x = var_569_cast_fp16)[name = tensor("cast_549")]; + tensor new_cache_1_internal_tensor_assign_2 = cast(dtype = reshape_9_cast_fp16_to_fp32_dtype_0, x = reshape_9_cast_fp16)[name = tensor("cast_550")]; + } -> (input, var_9476, new_cache_1_internal_tensor_assign_2, var_569, new_cache_3_internal_tensor_assign_2, var_955, new_cache_5_internal_tensor_assign_2, var_1341, new_cache_7_internal_tensor_assign_2, var_1727, new_cache_9_internal_tensor_assign_2, var_2113, new_cache_11_internal_tensor_assign_2, var_2499, new_cache_13_internal_tensor_assign_2, var_2885, new_cache_15_internal_tensor_assign_2, var_3271, new_cache_17_internal_tensor_assign_2, var_3657, new_cache_19_internal_tensor_assign_2, var_4043, new_cache_21_internal_tensor_assign_2, var_4429, new_cache_23_internal_tensor_assign_2, var_4815, new_cache_25_internal_tensor_assign_2, var_5201, new_cache_27_internal_tensor_assign_2, var_5587, new_cache_29_internal_tensor_assign_2, var_5973, new_cache_31_internal_tensor_assign_2, var_6359, new_cache_33_internal_tensor_assign_2, var_6745, new_cache_35_internal_tensor_assign_2, var_7131, new_cache_37_internal_tensor_assign_2, var_7517, new_cache_39_internal_tensor_assign_2, var_7903, new_cache_41_internal_tensor_assign_2, var_8289, new_cache_43_internal_tensor_assign_2, var_8675, new_cache_45_internal_tensor_assign_2, var_9061, new_cache_internal_tensor_assign_2, var_9447); +} \ No newline at end of file diff --git a/v2.1/portuguese_24l/flowlm_step.mlmodelc/weights/weight.bin b/v2.1/portuguese_24l/flowlm_step.mlmodelc/weights/weight.bin new file mode 100644 index 0000000000000000000000000000000000000000..1fce112bb0cbcb7b9c5cc863931b249d8cfd1308 --- /dev/null +++ b/v2.1/portuguese_24l/flowlm_step.mlmodelc/weights/weight.bin @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:7fccf4acb735831cdfefe210f9fb110e199000366abcac4113310cd5ffd76387 +size 605330048 diff --git a/v2.1/portuguese_24l/flowlm_step.mlpackage/Data/com.apple.CoreML/model.mlmodel b/v2.1/portuguese_24l/flowlm_step.mlpackage/Data/com.apple.CoreML/model.mlmodel new file mode 100644 index 0000000000000000000000000000000000000000..9aeebfc429cdf15170df0d12b59db4d176bc2d1e --- /dev/null +++ b/v2.1/portuguese_24l/flowlm_step.mlpackage/Data/com.apple.CoreML/model.mlmodel @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:0a19012917b7405221baff5a80a3b600f97561c3ec0bf84701640a3721e2828a +size 825108 diff --git a/v2.1/portuguese_24l/flowlm_step.mlpackage/Data/com.apple.CoreML/weights/weight.bin b/v2.1/portuguese_24l/flowlm_step.mlpackage/Data/com.apple.CoreML/weights/weight.bin new file mode 100644 index 0000000000000000000000000000000000000000..1fce112bb0cbcb7b9c5cc863931b249d8cfd1308 --- /dev/null +++ b/v2.1/portuguese_24l/flowlm_step.mlpackage/Data/com.apple.CoreML/weights/weight.bin @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:7fccf4acb735831cdfefe210f9fb110e199000366abcac4113310cd5ffd76387 +size 605330048 diff --git a/v2.1/portuguese_24l/flowlm_step.mlpackage/Manifest.json b/v2.1/portuguese_24l/flowlm_step.mlpackage/Manifest.json new file mode 100644 index 0000000000000000000000000000000000000000..acb7c9c06105961648fa7a564fa2a8efd7034543 --- /dev/null +++ b/v2.1/portuguese_24l/flowlm_step.mlpackage/Manifest.json @@ -0,0 +1,18 @@ +{ + "fileFormatVersion": "1.0.0", + "itemInfoEntries": { + "C15406EC-E62A-4C74-8F03-ACE9DCFD2C7C": { + "author": "com.apple.CoreML", + "description": "CoreML Model Specification", + "name": "model.mlmodel", + "path": "com.apple.CoreML/model.mlmodel" + }, + "D7E2FCE1-2677-4A65-96DB-EECA75A1BBCD": { + "author": "com.apple.CoreML", + "description": "CoreML Model Weights", + "name": "weights", + "path": "com.apple.CoreML/weights" + } + }, + "rootModelIdentifier": "C15406EC-E62A-4C74-8F03-ACE9DCFD2C7C" +} diff --git a/v2.1/portuguese_24l/flowlm_stepv2.mlmodelc/analytics/coremldata.bin b/v2.1/portuguese_24l/flowlm_stepv2.mlmodelc/analytics/coremldata.bin new file mode 100644 index 0000000000000000000000000000000000000000..1e8f378abf45f534c5b5fcda8f2020e91bfb9a8a --- /dev/null +++ b/v2.1/portuguese_24l/flowlm_stepv2.mlmodelc/analytics/coremldata.bin @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:c909e128b59fa4582e08a6c920d5750b36d17d62493ba9b96149788276d36ee1 +size 243 diff --git a/v2.1/portuguese_24l/flowlm_stepv2.mlmodelc/coremldata.bin b/v2.1/portuguese_24l/flowlm_stepv2.mlmodelc/coremldata.bin new file mode 100644 index 0000000000000000000000000000000000000000..30710404ff45f11b9bb9b0d702a07c4700b145a8 --- /dev/null +++ b/v2.1/portuguese_24l/flowlm_stepv2.mlmodelc/coremldata.bin @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:0080f802803ce4be9dae5c42c52f9f56884c78f96187a2661b4819f9218be149 +size 3561 diff --git a/v2.1/portuguese_24l/flowlm_stepv2.mlmodelc/metadata.json b/v2.1/portuguese_24l/flowlm_stepv2.mlmodelc/metadata.json new file mode 100644 index 0000000000000000000000000000000000000000..629da9a69819375413a6c11a7789eb2ca4f04e28 --- /dev/null +++ b/v2.1/portuguese_24l/flowlm_stepv2.mlmodelc/metadata.json @@ -0,0 +1,1068 @@ +[ + { + "metadataOutputVersion" : "3.0", + "storagePrecision" : "Mixed (Float16, Int8)", + "outputSchema" : [ + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1 × 1 × 1024)", + "shortDescription" : "", + "shape" : "[1, 1, 1024]", + "name" : "input", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1 × 1 × 1)", + "shortDescription" : "", + "shape" : "[1, 1, 1]", + "name" : "var_10268", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 2 × 1 × 512 × 16 × 64)", + "shortDescription" : "", + "shape" : "[2, 1, 512, 16, 64]", + "name" : "new_cache_1_internal_tensor_assign_2", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1)", + "shortDescription" : "", + "shape" : "[1]", + "name" : "var_602", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 2 × 1 × 512 × 16 × 64)", + "shortDescription" : "", + "shape" : "[2, 1, 512, 16, 64]", + "name" : "new_cache_3_internal_tensor_assign_2", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1)", + "shortDescription" : "", + "shape" : "[1]", + "name" : "var_1021", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 2 × 1 × 512 × 16 × 64)", + "shortDescription" : "", + "shape" : "[2, 1, 512, 16, 64]", + "name" : "new_cache_5_internal_tensor_assign_2", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1)", + "shortDescription" : "", + "shape" : "[1]", + "name" : "var_1440", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 2 × 1 × 512 × 16 × 64)", + "shortDescription" : "", + "shape" : "[2, 1, 512, 16, 64]", + "name" : "new_cache_7_internal_tensor_assign_2", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1)", + "shortDescription" : "", + "shape" : "[1]", + "name" : "var_1859", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 2 × 1 × 512 × 16 × 64)", + "shortDescription" : "", + "shape" : "[2, 1, 512, 16, 64]", + "name" : "new_cache_9_internal_tensor_assign_2", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1)", + "shortDescription" : "", + "shape" : "[1]", + "name" : "var_2278", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 2 × 1 × 512 × 16 × 64)", + "shortDescription" : "", + "shape" : "[2, 1, 512, 16, 64]", + "name" : "new_cache_11_internal_tensor_assign_2", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1)", + "shortDescription" : "", + "shape" : "[1]", + "name" : "var_2697", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 2 × 1 × 512 × 16 × 64)", + "shortDescription" : "", + "shape" : "[2, 1, 512, 16, 64]", + "name" : "new_cache_13_internal_tensor_assign_2", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1)", + "shortDescription" : "", + "shape" : "[1]", + "name" : "var_3116", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 2 × 1 × 512 × 16 × 64)", + "shortDescription" : "", + "shape" : "[2, 1, 512, 16, 64]", + "name" : "new_cache_15_internal_tensor_assign_2", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1)", + "shortDescription" : "", + "shape" : "[1]", + "name" : "var_3535", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 2 × 1 × 512 × 16 × 64)", + "shortDescription" : "", + "shape" : "[2, 1, 512, 16, 64]", + "name" : "new_cache_17_internal_tensor_assign_2", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1)", + "shortDescription" : "", + "shape" : "[1]", + "name" : "var_3954", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 2 × 1 × 512 × 16 × 64)", + "shortDescription" : "", + "shape" : "[2, 1, 512, 16, 64]", + "name" : "new_cache_19_internal_tensor_assign_2", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1)", + "shortDescription" : "", + "shape" : "[1]", + "name" : "var_4373", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 2 × 1 × 512 × 16 × 64)", + "shortDescription" : "", + "shape" : "[2, 1, 512, 16, 64]", + "name" : "new_cache_21_internal_tensor_assign_2", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1)", + "shortDescription" : "", + "shape" : "[1]", + "name" : "var_4792", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 2 × 1 × 512 × 16 × 64)", + "shortDescription" : "", + "shape" : "[2, 1, 512, 16, 64]", + "name" : "new_cache_23_internal_tensor_assign_2", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1)", + "shortDescription" : "", + "shape" : "[1]", + "name" : "var_5211", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 2 × 1 × 512 × 16 × 64)", + "shortDescription" : "", + "shape" : "[2, 1, 512, 16, 64]", + "name" : "new_cache_25_internal_tensor_assign_2", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1)", + "shortDescription" : "", + "shape" : "[1]", + "name" : "var_5630", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 2 × 1 × 512 × 16 × 64)", + "shortDescription" : "", + "shape" : "[2, 1, 512, 16, 64]", + "name" : "new_cache_27_internal_tensor_assign_2", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1)", + "shortDescription" : "", + "shape" : "[1]", + "name" : "var_6049", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 2 × 1 × 512 × 16 × 64)", + "shortDescription" : "", + "shape" : "[2, 1, 512, 16, 64]", + "name" : "new_cache_29_internal_tensor_assign_2", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1)", + "shortDescription" : "", + "shape" : "[1]", + "name" : "var_6468", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 2 × 1 × 512 × 16 × 64)", + "shortDescription" : "", + "shape" : "[2, 1, 512, 16, 64]", + "name" : "new_cache_31_internal_tensor_assign_2", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1)", + "shortDescription" : "", + "shape" : "[1]", + "name" : "var_6887", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 2 × 1 × 512 × 16 × 64)", + "shortDescription" : "", + "shape" : "[2, 1, 512, 16, 64]", + "name" : "new_cache_33_internal_tensor_assign_2", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1)", + "shortDescription" : "", + "shape" : "[1]", + "name" : "var_7306", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 2 × 1 × 512 × 16 × 64)", + "shortDescription" : "", + "shape" : "[2, 1, 512, 16, 64]", + "name" : "new_cache_35_internal_tensor_assign_2", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1)", + "shortDescription" : "", + "shape" : "[1]", + "name" : "var_7725", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 2 × 1 × 512 × 16 × 64)", + "shortDescription" : "", + "shape" : "[2, 1, 512, 16, 64]", + "name" : "new_cache_37_internal_tensor_assign_2", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1)", + "shortDescription" : "", + "shape" : "[1]", + "name" : "var_8144", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 2 × 1 × 512 × 16 × 64)", + "shortDescription" : "", + "shape" : "[2, 1, 512, 16, 64]", + "name" : "new_cache_39_internal_tensor_assign_2", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1)", + "shortDescription" : "", + "shape" : "[1]", + "name" : "var_8563", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 2 × 1 × 512 × 16 × 64)", + "shortDescription" : "", + "shape" : "[2, 1, 512, 16, 64]", + "name" : "new_cache_41_internal_tensor_assign_2", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1)", + "shortDescription" : "", + "shape" : "[1]", + "name" : "var_8982", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 2 × 1 × 512 × 16 × 64)", + "shortDescription" : "", + "shape" : "[2, 1, 512, 16, 64]", + "name" : "new_cache_43_internal_tensor_assign_2", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1)", + "shortDescription" : "", + "shape" : "[1]", + "name" : "var_9401", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 2 × 1 × 512 × 16 × 64)", + "shortDescription" : "", + "shape" : "[2, 1, 512, 16, 64]", + "name" : "new_cache_45_internal_tensor_assign_2", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1)", + "shortDescription" : "", + "shape" : "[1]", + "name" : "var_9820", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 2 × 1 × 512 × 16 × 64)", + "shortDescription" : "", + "shape" : "[2, 1, 512, 16, 64]", + "name" : "new_cache_internal_tensor_assign_2", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1)", + "shortDescription" : "", + "shape" : "[1]", + "name" : "var_10239", + "type" : "MultiArray" + } + ], + "modelParameters" : [ + + ], + "specificationVersion" : 8, + "mlProgramOperationTypeHistogram" : { + "Ios17.cos" : 24, + "Range1d" : 48, + "Ios17.notEqual" : 49, + "Ios17.floor" : 24, + "Ios17.reshape" : 432, + "Ios17.logicalAnd" : 24, + "Ios16.softmax" : 24, + "Ios17.scatter" : 48, + "Ios17.matmul" : 48, + "Ios17.sin" : 24, + "Ios17.transpose" : 96, + "Select" : 73, + "Ios17.expandDims" : 25, + "Ios17.add" : 144, + "Tile" : 24, + "Ios17.sliceByIndex" : 312, + "Ios17.lessEqual" : 24, + "Ios17.scatterAlongAxis" : 48, + "Ios17.logicalNot" : 24, + "Ios17.layerNorm" : 49, + "Ios16.constexprAffineDequantize" : 96, + "Ios17.cast" : 124, + "Ios17.sub" : 72, + "Ios17.less" : 24, + "Ios16.gelu" : 24, + "Ios17.linear" : 98, + "Stack" : 48, + "Ios17.mul" : 288 + }, + "computePrecision" : "Mixed (Float16, Float32, Int32)", + "isUpdatable" : "0", + "stateSchema" : [ + + ], + "availability" : { + "macOS" : "14.0", + "tvOS" : "17.0", + "visionOS" : "1.0", + "watchOS" : "10.0", + "iOS" : "17.0", + "macCatalyst" : "17.0" + }, + "modelType" : { + "name" : "MLModelType_mlProgram" + }, + "userDefinedMetadata" : { + "com.github.apple.coremltools.conversion_date" : "2026-04-27", + "com.github.apple.coremltools.source" : "torch==2.9.1", + "com.github.apple.coremltools.version" : "9.0", + "com.github.apple.coremltools.source_dialect" : "TorchScript" + }, + "inputSchema" : [ + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1 × 1 × 32)", + "shortDescription" : "", + "shape" : "[1, 1, 32]", + "name" : "sequence", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 32)", + "shortDescription" : "", + "shape" : "[32]", + "name" : "bos_emb", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 2 × 1 × 512 × 16 × 64)", + "shortDescription" : "", + "shape" : "[2, 1, 512, 16, 64]", + "name" : "cache0", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1)", + "shortDescription" : "", + "shape" : "[1]", + "name" : "position0", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 2 × 1 × 512 × 16 × 64)", + "shortDescription" : "", + "shape" : "[2, 1, 512, 16, 64]", + "name" : "cache1", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1)", + "shortDescription" : "", + "shape" : "[1]", + "name" : "position1", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 2 × 1 × 512 × 16 × 64)", + "shortDescription" : "", + "shape" : "[2, 1, 512, 16, 64]", + "name" : "cache2", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1)", + "shortDescription" : "", + "shape" : "[1]", + "name" : "position2", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 2 × 1 × 512 × 16 × 64)", + "shortDescription" : "", + "shape" : "[2, 1, 512, 16, 64]", + "name" : "cache3", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1)", + "shortDescription" : "", + "shape" : "[1]", + "name" : "position3", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 2 × 1 × 512 × 16 × 64)", + "shortDescription" : "", + "shape" : "[2, 1, 512, 16, 64]", + "name" : "cache4", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1)", + "shortDescription" : "", + "shape" : "[1]", + "name" : "position4", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 2 × 1 × 512 × 16 × 64)", + "shortDescription" : "", + "shape" : "[2, 1, 512, 16, 64]", + "name" : "cache5", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1)", + "shortDescription" : "", + "shape" : "[1]", + "name" : "position5", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 2 × 1 × 512 × 16 × 64)", + "shortDescription" : "", + "shape" : "[2, 1, 512, 16, 64]", + "name" : "cache6", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1)", + "shortDescription" : "", + "shape" : "[1]", + "name" : "position6", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 2 × 1 × 512 × 16 × 64)", + "shortDescription" : "", + "shape" : "[2, 1, 512, 16, 64]", + "name" : "cache7", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1)", + "shortDescription" : "", + "shape" : "[1]", + "name" : "position7", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 2 × 1 × 512 × 16 × 64)", + "shortDescription" : "", + "shape" : "[2, 1, 512, 16, 64]", + "name" : "cache8", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1)", + "shortDescription" : "", + "shape" : "[1]", + "name" : "position8", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 2 × 1 × 512 × 16 × 64)", + "shortDescription" : "", + "shape" : "[2, 1, 512, 16, 64]", + "name" : "cache9", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1)", + "shortDescription" : "", + "shape" : "[1]", + "name" : "position9", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 2 × 1 × 512 × 16 × 64)", + "shortDescription" : "", + "shape" : "[2, 1, 512, 16, 64]", + "name" : "cache10", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1)", + "shortDescription" : "", + "shape" : "[1]", + "name" : "position10", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 2 × 1 × 512 × 16 × 64)", + "shortDescription" : "", + "shape" : "[2, 1, 512, 16, 64]", + "name" : "cache11", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1)", + "shortDescription" : "", + "shape" : "[1]", + "name" : "position11", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 2 × 1 × 512 × 16 × 64)", + "shortDescription" : "", + "shape" : "[2, 1, 512, 16, 64]", + "name" : "cache12", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1)", + "shortDescription" : "", + "shape" : "[1]", + "name" : "position12", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 2 × 1 × 512 × 16 × 64)", + "shortDescription" : "", + "shape" : "[2, 1, 512, 16, 64]", + "name" : "cache13", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1)", + "shortDescription" : "", + "shape" : "[1]", + "name" : "position13", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 2 × 1 × 512 × 16 × 64)", + "shortDescription" : "", + "shape" : "[2, 1, 512, 16, 64]", + "name" : "cache14", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1)", + "shortDescription" : "", + "shape" : "[1]", + "name" : "position14", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 2 × 1 × 512 × 16 × 64)", + "shortDescription" : "", + "shape" : "[2, 1, 512, 16, 64]", + "name" : "cache15", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1)", + "shortDescription" : "", + "shape" : "[1]", + "name" : "position15", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 2 × 1 × 512 × 16 × 64)", + "shortDescription" : "", + "shape" : "[2, 1, 512, 16, 64]", + "name" : "cache16", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1)", + "shortDescription" : "", + "shape" : "[1]", + "name" : "position16", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 2 × 1 × 512 × 16 × 64)", + "shortDescription" : "", + "shape" : "[2, 1, 512, 16, 64]", + "name" : "cache17", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1)", + "shortDescription" : "", + "shape" : "[1]", + "name" : "position17", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 2 × 1 × 512 × 16 × 64)", + "shortDescription" : "", + "shape" : "[2, 1, 512, 16, 64]", + "name" : "cache18", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1)", + "shortDescription" : "", + "shape" : "[1]", + "name" : "position18", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 2 × 1 × 512 × 16 × 64)", + "shortDescription" : "", + "shape" : "[2, 1, 512, 16, 64]", + "name" : "cache19", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1)", + "shortDescription" : "", + "shape" : "[1]", + "name" : "position19", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 2 × 1 × 512 × 16 × 64)", + "shortDescription" : "", + "shape" : "[2, 1, 512, 16, 64]", + "name" : "cache20", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1)", + "shortDescription" : "", + "shape" : "[1]", + "name" : "position20", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 2 × 1 × 512 × 16 × 64)", + "shortDescription" : "", + "shape" : "[2, 1, 512, 16, 64]", + "name" : "cache21", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1)", + "shortDescription" : "", + "shape" : "[1]", + "name" : "position21", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 2 × 1 × 512 × 16 × 64)", + "shortDescription" : "", + "shape" : "[2, 1, 512, 16, 64]", + "name" : "cache22", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1)", + "shortDescription" : "", + "shape" : "[1]", + "name" : "position22", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 2 × 1 × 512 × 16 × 64)", + "shortDescription" : "", + "shape" : "[2, 1, 512, 16, 64]", + "name" : "cache23", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1)", + "shortDescription" : "", + "shape" : "[1]", + "name" : "position23", + "type" : "MultiArray" + } + ], + "generatedClassName" : "flowlm_stepv2", + "method" : "predict" + } +] \ No newline at end of file diff --git a/v2.1/portuguese_24l/flowlm_stepv2.mlmodelc/model.mil b/v2.1/portuguese_24l/flowlm_stepv2.mlmodelc/model.mil new file mode 100644 index 0000000000000000000000000000000000000000..14271b1a5cd3a0427a28dd1ce46789958cbd95d1 --- /dev/null +++ b/v2.1/portuguese_24l/flowlm_stepv2.mlmodelc/model.mil @@ -0,0 +1,5337 @@ +program(1.0) +[buildInfo = dict, tensor>({{"coremlc-component-MIL", "3520.4.1"}, {"coremlc-version", "3520.5.1"}, {"coremltools-component-torch", "2.9.1"}, {"coremltools-source-dialect", "TorchScript"}, {"coremltools-version", "9.0"}})] +{ + func main(tensor bos_emb, tensor cache0, tensor cache1, tensor cache10, tensor cache11, tensor cache12, tensor cache13, tensor cache14, tensor cache15, tensor cache16, tensor cache17, tensor cache18, tensor cache19, tensor cache2, tensor cache20, tensor cache21, tensor cache22, tensor cache23, tensor cache3, tensor cache4, tensor cache5, tensor cache6, tensor cache7, tensor cache8, tensor cache9, tensor position0, tensor position1, tensor position10, tensor position11, tensor position12, tensor position13, tensor position14, tensor position15, tensor position16, tensor position17, tensor position18, tensor position19, tensor position2, tensor position20, tensor position21, tensor position22, tensor position23, tensor position3, tensor position4, tensor position5, tensor position6, tensor position7, tensor position8, tensor position9, tensor sequence) { + tensor attn0_in_proj_weight_cast_fp16 = constexpr_affine_dequantize()[axis = tensor(0), name = tensor("attn0_in_proj_weight_cast_fp16"), quantized_data = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(64))), scale = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(3148992))), zero_point = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(3145856)))]; + tensor attn0_out_proj_weight_cast_fp16 = constexpr_affine_dequantize()[axis = tensor(0), name = tensor("attn0_out_proj_weight_cast_fp16"), quantized_data = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(3155200))), scale = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(4204928))), zero_point = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(4203840)))]; + tensor linear0_1_weight_cast_fp16 = constexpr_affine_dequantize()[axis = tensor(0), name = tensor("linear0_1_weight_cast_fp16"), quantized_data = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(4207040))), scale = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8405568))), zero_point = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8401408)))]; + tensor linear0_2_weight_cast_fp16 = constexpr_affine_dequantize()[axis = tensor(0), name = tensor("linear0_2_weight_cast_fp16"), quantized_data = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8413824))), scale = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(12608192))), zero_point = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(4203840)))]; + tensor attn1_in_proj_weight_cast_fp16 = constexpr_affine_dequantize()[axis = tensor(0), name = tensor("attn1_in_proj_weight_cast_fp16"), quantized_data = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(12610304))), scale = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15756096))), zero_point = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(3145856)))]; + tensor attn1_out_proj_weight_cast_fp16 = constexpr_affine_dequantize()[axis = tensor(0), name = tensor("attn1_out_proj_weight_cast_fp16"), quantized_data = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15762304))), scale = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(16810944))), zero_point = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(4203840)))]; + tensor linear1_1_weight_cast_fp16 = constexpr_affine_dequantize()[axis = tensor(0), name = tensor("linear1_1_weight_cast_fp16"), quantized_data = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(16813056))), scale = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21007424))), zero_point = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8401408)))]; + tensor linear1_2_weight_cast_fp16 = constexpr_affine_dequantize()[axis = tensor(0), name = tensor("linear1_2_weight_cast_fp16"), quantized_data = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21015680))), scale = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(25210048))), zero_point = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(4203840)))]; + tensor attn2_in_proj_weight_cast_fp16 = constexpr_affine_dequantize()[axis = tensor(0), name = tensor("attn2_in_proj_weight_cast_fp16"), quantized_data = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(25212160))), scale = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28357952))), zero_point = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(3145856)))]; + tensor attn2_out_proj_weight_cast_fp16 = constexpr_affine_dequantize()[axis = tensor(0), name = tensor("attn2_out_proj_weight_cast_fp16"), quantized_data = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28364160))), scale = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(29412800))), zero_point = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(4203840)))]; + tensor linear2_1_weight_cast_fp16 = constexpr_affine_dequantize()[axis = tensor(0), name = tensor("linear2_1_weight_cast_fp16"), quantized_data = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(29414912))), scale = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(33609280))), zero_point = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8401408)))]; + tensor linear2_2_weight_cast_fp16 = constexpr_affine_dequantize()[axis = tensor(0), name = tensor("linear2_2_weight_cast_fp16"), quantized_data = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(33617536))), scale = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(37811904))), zero_point = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(4203840)))]; + tensor attn3_in_proj_weight_cast_fp16 = constexpr_affine_dequantize()[axis = tensor(0), name = tensor("attn3_in_proj_weight_cast_fp16"), quantized_data = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(37814016))), scale = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(40959808))), zero_point = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(3145856)))]; + tensor attn3_out_proj_weight_cast_fp16 = constexpr_affine_dequantize()[axis = tensor(0), name = tensor("attn3_out_proj_weight_cast_fp16"), quantized_data = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(40966016))), scale = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(42014656))), zero_point = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(4203840)))]; + tensor linear3_1_weight_cast_fp16 = constexpr_affine_dequantize()[axis = tensor(0), name = tensor("linear3_1_weight_cast_fp16"), quantized_data = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(42016768))), scale = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(46211136))), zero_point = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8401408)))]; + tensor linear3_2_weight_cast_fp16 = constexpr_affine_dequantize()[axis = tensor(0), name = tensor("linear3_2_weight_cast_fp16"), quantized_data = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(46219392))), scale = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(50413760))), zero_point = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(4203840)))]; + tensor attn4_in_proj_weight_cast_fp16 = constexpr_affine_dequantize()[axis = tensor(0), name = tensor("attn4_in_proj_weight_cast_fp16"), quantized_data = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(50415872))), scale = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(53561664))), zero_point = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(3145856)))]; + tensor attn4_out_proj_weight_cast_fp16 = constexpr_affine_dequantize()[axis = tensor(0), name = tensor("attn4_out_proj_weight_cast_fp16"), quantized_data = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(53567872))), scale = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(54616512))), zero_point = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(4203840)))]; + tensor linear4_1_weight_cast_fp16 = constexpr_affine_dequantize()[axis = tensor(0), name = tensor("linear4_1_weight_cast_fp16"), quantized_data = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(54618624))), scale = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(58812992))), zero_point = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8401408)))]; + tensor linear4_2_weight_cast_fp16 = constexpr_affine_dequantize()[axis = tensor(0), name = tensor("linear4_2_weight_cast_fp16"), quantized_data = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(58821248))), scale = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(63015616))), zero_point = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(4203840)))]; + tensor attn5_in_proj_weight_cast_fp16 = constexpr_affine_dequantize()[axis = tensor(0), name = tensor("attn5_in_proj_weight_cast_fp16"), quantized_data = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(63017728))), scale = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(66163520))), zero_point = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(3145856)))]; + tensor attn5_out_proj_weight_cast_fp16 = constexpr_affine_dequantize()[axis = tensor(0), name = tensor("attn5_out_proj_weight_cast_fp16"), quantized_data = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(66169728))), scale = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(67218368))), zero_point = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(4203840)))]; + tensor linear5_1_weight_cast_fp16 = constexpr_affine_dequantize()[axis = tensor(0), name = tensor("linear5_1_weight_cast_fp16"), quantized_data = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(67220480))), scale = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(71414848))), zero_point = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8401408)))]; + tensor linear5_2_weight_cast_fp16 = constexpr_affine_dequantize()[axis = tensor(0), name = tensor("linear5_2_weight_cast_fp16"), quantized_data = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(71423104))), scale = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(75617472))), zero_point = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(4203840)))]; + tensor attn6_in_proj_weight_cast_fp16 = constexpr_affine_dequantize()[axis = tensor(0), name = tensor("attn6_in_proj_weight_cast_fp16"), quantized_data = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(75619584))), scale = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(78765376))), zero_point = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(3145856)))]; + tensor attn6_out_proj_weight_cast_fp16 = constexpr_affine_dequantize()[axis = tensor(0), name = tensor("attn6_out_proj_weight_cast_fp16"), quantized_data = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(78771584))), scale = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(79820224))), zero_point = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(4203840)))]; + tensor linear6_1_weight_cast_fp16 = constexpr_affine_dequantize()[axis = tensor(0), name = tensor("linear6_1_weight_cast_fp16"), quantized_data = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(79822336))), scale = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(84016704))), zero_point = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8401408)))]; + tensor linear6_2_weight_cast_fp16 = constexpr_affine_dequantize()[axis = tensor(0), name = tensor("linear6_2_weight_cast_fp16"), quantized_data = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(84024960))), scale = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(88219328))), zero_point = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(4203840)))]; + tensor attn7_in_proj_weight_cast_fp16 = constexpr_affine_dequantize()[axis = tensor(0), name = tensor("attn7_in_proj_weight_cast_fp16"), quantized_data = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(88221440))), scale = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(91367232))), zero_point = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(3145856)))]; + tensor attn7_out_proj_weight_cast_fp16 = constexpr_affine_dequantize()[axis = tensor(0), name = tensor("attn7_out_proj_weight_cast_fp16"), quantized_data = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(91373440))), scale = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(92422080))), zero_point = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(4203840)))]; + tensor linear7_1_weight_cast_fp16 = constexpr_affine_dequantize()[axis = tensor(0), name = tensor("linear7_1_weight_cast_fp16"), quantized_data = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(92424192))), scale = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(96618560))), zero_point = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8401408)))]; + tensor linear7_2_weight_cast_fp16 = constexpr_affine_dequantize()[axis = tensor(0), name = tensor("linear7_2_weight_cast_fp16"), quantized_data = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(96626816))), scale = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(100821184))), zero_point = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(4203840)))]; + tensor attn8_in_proj_weight_cast_fp16 = constexpr_affine_dequantize()[axis = tensor(0), name = tensor("attn8_in_proj_weight_cast_fp16"), quantized_data = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(100823296))), scale = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(103969088))), zero_point = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(3145856)))]; + tensor attn8_out_proj_weight_cast_fp16 = constexpr_affine_dequantize()[axis = tensor(0), name = tensor("attn8_out_proj_weight_cast_fp16"), quantized_data = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(103975296))), scale = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(105023936))), zero_point = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(4203840)))]; + tensor linear8_1_weight_cast_fp16 = constexpr_affine_dequantize()[axis = tensor(0), name = tensor("linear8_1_weight_cast_fp16"), quantized_data = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(105026048))), scale = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(109220416))), zero_point = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8401408)))]; + tensor linear8_2_weight_cast_fp16 = constexpr_affine_dequantize()[axis = tensor(0), name = tensor("linear8_2_weight_cast_fp16"), quantized_data = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(109228672))), scale = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(113423040))), zero_point = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(4203840)))]; + tensor attn9_in_proj_weight_cast_fp16 = constexpr_affine_dequantize()[axis = tensor(0), name = tensor("attn9_in_proj_weight_cast_fp16"), quantized_data = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(113425152))), scale = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(116570944))), zero_point = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(3145856)))]; + tensor attn9_out_proj_weight_cast_fp16 = constexpr_affine_dequantize()[axis = tensor(0), name = tensor("attn9_out_proj_weight_cast_fp16"), quantized_data = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(116577152))), scale = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(117625792))), zero_point = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(4203840)))]; + tensor linear9_1_weight_cast_fp16 = constexpr_affine_dequantize()[axis = tensor(0), name = tensor("linear9_1_weight_cast_fp16"), quantized_data = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(117627904))), scale = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(121822272))), zero_point = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8401408)))]; + tensor linear9_2_weight_cast_fp16 = constexpr_affine_dequantize()[axis = tensor(0), name = tensor("linear9_2_weight_cast_fp16"), quantized_data = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(121830528))), scale = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(126024896))), zero_point = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(4203840)))]; + tensor attn10_in_proj_weight_cast_fp16 = constexpr_affine_dequantize()[axis = tensor(0), name = tensor("attn10_in_proj_weight_cast_fp16"), quantized_data = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(126027008))), scale = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(129172800))), zero_point = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(3145856)))]; + tensor attn10_out_proj_weight_cast_fp16 = constexpr_affine_dequantize()[axis = tensor(0), name = tensor("attn10_out_proj_weight_cast_fp16"), quantized_data = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(129179008))), scale = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(130227648))), zero_point = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(4203840)))]; + tensor linear10_1_weight_cast_fp16 = constexpr_affine_dequantize()[axis = tensor(0), name = tensor("linear10_1_weight_cast_fp16"), quantized_data = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(130229760))), scale = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(134424128))), zero_point = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8401408)))]; + tensor linear10_2_weight_cast_fp16 = constexpr_affine_dequantize()[axis = tensor(0), name = tensor("linear10_2_weight_cast_fp16"), quantized_data = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(134432384))), scale = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(138626752))), zero_point = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(4203840)))]; + tensor attn11_in_proj_weight_cast_fp16 = constexpr_affine_dequantize()[axis = tensor(0), name = tensor("attn11_in_proj_weight_cast_fp16"), quantized_data = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(138628864))), scale = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(141774656))), zero_point = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(3145856)))]; + tensor attn11_out_proj_weight_cast_fp16 = constexpr_affine_dequantize()[axis = tensor(0), name = tensor("attn11_out_proj_weight_cast_fp16"), quantized_data = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(141780864))), scale = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(142829504))), zero_point = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(4203840)))]; + tensor linear11_1_weight_cast_fp16 = constexpr_affine_dequantize()[axis = tensor(0), name = tensor("linear11_1_weight_cast_fp16"), quantized_data = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(142831616))), scale = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(147025984))), zero_point = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8401408)))]; + tensor linear11_2_weight_cast_fp16 = constexpr_affine_dequantize()[axis = tensor(0), name = tensor("linear11_2_weight_cast_fp16"), quantized_data = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(147034240))), scale = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(151228608))), zero_point = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(4203840)))]; + tensor attn12_in_proj_weight_cast_fp16 = constexpr_affine_dequantize()[axis = tensor(0), name = tensor("attn12_in_proj_weight_cast_fp16"), quantized_data = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(151230720))), scale = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(154376512))), zero_point = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(3145856)))]; + tensor attn12_out_proj_weight_cast_fp16 = constexpr_affine_dequantize()[axis = tensor(0), name = tensor("attn12_out_proj_weight_cast_fp16"), quantized_data = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(154382720))), scale = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(155431360))), zero_point = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(4203840)))]; + tensor linear12_1_weight_cast_fp16 = constexpr_affine_dequantize()[axis = tensor(0), name = tensor("linear12_1_weight_cast_fp16"), quantized_data = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(155433472))), scale = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(159627840))), zero_point = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8401408)))]; + tensor linear12_2_weight_cast_fp16 = constexpr_affine_dequantize()[axis = tensor(0), name = tensor("linear12_2_weight_cast_fp16"), quantized_data = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(159636096))), scale = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(163830464))), zero_point = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(4203840)))]; + tensor attn13_in_proj_weight_cast_fp16 = constexpr_affine_dequantize()[axis = tensor(0), name = tensor("attn13_in_proj_weight_cast_fp16"), quantized_data = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(163832576))), scale = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(166978368))), zero_point = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(3145856)))]; + tensor attn13_out_proj_weight_cast_fp16 = constexpr_affine_dequantize()[axis = tensor(0), name = tensor("attn13_out_proj_weight_cast_fp16"), quantized_data = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(166984576))), scale = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(168033216))), zero_point = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(4203840)))]; + tensor linear13_1_weight_cast_fp16 = constexpr_affine_dequantize()[axis = tensor(0), name = tensor("linear13_1_weight_cast_fp16"), quantized_data = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(168035328))), scale = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(172229696))), zero_point = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8401408)))]; + tensor linear13_2_weight_cast_fp16 = constexpr_affine_dequantize()[axis = tensor(0), name = tensor("linear13_2_weight_cast_fp16"), quantized_data = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(172237952))), scale = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(176432320))), zero_point = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(4203840)))]; + tensor attn14_in_proj_weight_cast_fp16 = constexpr_affine_dequantize()[axis = tensor(0), name = tensor("attn14_in_proj_weight_cast_fp16"), quantized_data = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(176434432))), scale = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(179580224))), zero_point = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(3145856)))]; + tensor attn14_out_proj_weight_cast_fp16 = constexpr_affine_dequantize()[axis = tensor(0), name = tensor("attn14_out_proj_weight_cast_fp16"), quantized_data = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(179586432))), scale = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(180635072))), zero_point = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(4203840)))]; + tensor linear14_1_weight_cast_fp16 = constexpr_affine_dequantize()[axis = tensor(0), name = tensor("linear14_1_weight_cast_fp16"), quantized_data = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(180637184))), scale = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(184831552))), zero_point = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8401408)))]; + tensor linear14_2_weight_cast_fp16 = constexpr_affine_dequantize()[axis = tensor(0), name = tensor("linear14_2_weight_cast_fp16"), quantized_data = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(184839808))), scale = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(189034176))), zero_point = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(4203840)))]; + tensor attn15_in_proj_weight_cast_fp16 = constexpr_affine_dequantize()[axis = tensor(0), name = tensor("attn15_in_proj_weight_cast_fp16"), quantized_data = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(189036288))), scale = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(192182080))), zero_point = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(3145856)))]; + tensor attn15_out_proj_weight_cast_fp16 = constexpr_affine_dequantize()[axis = tensor(0), name = tensor("attn15_out_proj_weight_cast_fp16"), quantized_data = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(192188288))), scale = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(193236928))), zero_point = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(4203840)))]; + tensor linear15_1_weight_cast_fp16 = constexpr_affine_dequantize()[axis = tensor(0), name = tensor("linear15_1_weight_cast_fp16"), quantized_data = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(193239040))), scale = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(197433408))), zero_point = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8401408)))]; + tensor linear15_2_weight_cast_fp16 = constexpr_affine_dequantize()[axis = tensor(0), name = tensor("linear15_2_weight_cast_fp16"), quantized_data = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(197441664))), scale = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(201636032))), zero_point = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(4203840)))]; + tensor attn16_in_proj_weight_cast_fp16 = constexpr_affine_dequantize()[axis = tensor(0), name = tensor("attn16_in_proj_weight_cast_fp16"), quantized_data = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(201638144))), scale = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(204783936))), zero_point = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(3145856)))]; + tensor attn16_out_proj_weight_cast_fp16 = constexpr_affine_dequantize()[axis = tensor(0), name = tensor("attn16_out_proj_weight_cast_fp16"), quantized_data = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(204790144))), scale = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(205838784))), zero_point = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(4203840)))]; + tensor linear16_1_weight_cast_fp16 = constexpr_affine_dequantize()[axis = tensor(0), name = tensor("linear16_1_weight_cast_fp16"), quantized_data = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(205840896))), scale = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(210035264))), zero_point = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8401408)))]; + tensor linear16_2_weight_cast_fp16 = constexpr_affine_dequantize()[axis = tensor(0), name = tensor("linear16_2_weight_cast_fp16"), quantized_data = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(210043520))), scale = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(214237888))), zero_point = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(4203840)))]; + tensor attn17_in_proj_weight_cast_fp16 = constexpr_affine_dequantize()[axis = tensor(0), name = tensor("attn17_in_proj_weight_cast_fp16"), quantized_data = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(214240000))), scale = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(217385792))), zero_point = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(3145856)))]; + tensor attn17_out_proj_weight_cast_fp16 = constexpr_affine_dequantize()[axis = tensor(0), name = tensor("attn17_out_proj_weight_cast_fp16"), quantized_data = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(217392000))), scale = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(218440640))), zero_point = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(4203840)))]; + tensor linear17_1_weight_cast_fp16 = constexpr_affine_dequantize()[axis = tensor(0), name = tensor("linear17_1_weight_cast_fp16"), quantized_data = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(218442752))), scale = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(222637120))), zero_point = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8401408)))]; + tensor linear17_2_weight_cast_fp16 = constexpr_affine_dequantize()[axis = tensor(0), name = tensor("linear17_2_weight_cast_fp16"), quantized_data = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(222645376))), scale = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(226839744))), zero_point = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(4203840)))]; + tensor attn18_in_proj_weight_cast_fp16 = constexpr_affine_dequantize()[axis = tensor(0), name = tensor("attn18_in_proj_weight_cast_fp16"), quantized_data = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(226841856))), scale = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(229987648))), zero_point = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(3145856)))]; + tensor attn18_out_proj_weight_cast_fp16 = constexpr_affine_dequantize()[axis = tensor(0), name = tensor("attn18_out_proj_weight_cast_fp16"), quantized_data = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(229993856))), scale = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(231042496))), zero_point = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(4203840)))]; + tensor linear18_1_weight_cast_fp16 = constexpr_affine_dequantize()[axis = tensor(0), name = tensor("linear18_1_weight_cast_fp16"), quantized_data = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(231044608))), scale = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(235238976))), zero_point = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8401408)))]; + tensor linear18_2_weight_cast_fp16 = constexpr_affine_dequantize()[axis = tensor(0), name = tensor("linear18_2_weight_cast_fp16"), quantized_data = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(235247232))), scale = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(239441600))), zero_point = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(4203840)))]; + tensor attn19_in_proj_weight_cast_fp16 = constexpr_affine_dequantize()[axis = tensor(0), name = tensor("attn19_in_proj_weight_cast_fp16"), quantized_data = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(239443712))), scale = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(242589504))), zero_point = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(3145856)))]; + tensor attn19_out_proj_weight_cast_fp16 = constexpr_affine_dequantize()[axis = tensor(0), name = tensor("attn19_out_proj_weight_cast_fp16"), quantized_data = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(242595712))), scale = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(243644352))), zero_point = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(4203840)))]; + tensor linear19_1_weight_cast_fp16 = constexpr_affine_dequantize()[axis = tensor(0), name = tensor("linear19_1_weight_cast_fp16"), quantized_data = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(243646464))), scale = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(247840832))), zero_point = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8401408)))]; + tensor linear19_2_weight_cast_fp16 = constexpr_affine_dequantize()[axis = tensor(0), name = tensor("linear19_2_weight_cast_fp16"), quantized_data = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(247849088))), scale = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(252043456))), zero_point = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(4203840)))]; + tensor attn20_in_proj_weight_cast_fp16 = constexpr_affine_dequantize()[axis = tensor(0), name = tensor("attn20_in_proj_weight_cast_fp16"), quantized_data = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(252045568))), scale = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(255191360))), zero_point = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(3145856)))]; + tensor attn20_out_proj_weight_cast_fp16 = constexpr_affine_dequantize()[axis = tensor(0), name = tensor("attn20_out_proj_weight_cast_fp16"), quantized_data = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(255197568))), scale = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(256246208))), zero_point = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(4203840)))]; + tensor linear20_1_weight_cast_fp16 = constexpr_affine_dequantize()[axis = tensor(0), name = tensor("linear20_1_weight_cast_fp16"), quantized_data = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(256248320))), scale = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(260442688))), zero_point = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8401408)))]; + tensor linear20_2_weight_cast_fp16 = constexpr_affine_dequantize()[axis = tensor(0), name = tensor("linear20_2_weight_cast_fp16"), quantized_data = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(260450944))), scale = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(264645312))), zero_point = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(4203840)))]; + tensor attn21_in_proj_weight_cast_fp16 = constexpr_affine_dequantize()[axis = tensor(0), name = tensor("attn21_in_proj_weight_cast_fp16"), quantized_data = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(264647424))), scale = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(267793216))), zero_point = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(3145856)))]; + tensor attn21_out_proj_weight_cast_fp16 = constexpr_affine_dequantize()[axis = tensor(0), name = tensor("attn21_out_proj_weight_cast_fp16"), quantized_data = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(267799424))), scale = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(268848064))), zero_point = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(4203840)))]; + tensor linear21_1_weight_cast_fp16 = constexpr_affine_dequantize()[axis = tensor(0), name = tensor("linear21_1_weight_cast_fp16"), quantized_data = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(268850176))), scale = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(273044544))), zero_point = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8401408)))]; + tensor linear21_2_weight_cast_fp16 = constexpr_affine_dequantize()[axis = tensor(0), name = tensor("linear21_2_weight_cast_fp16"), quantized_data = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(273052800))), scale = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(277247168))), zero_point = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(4203840)))]; + tensor attn22_in_proj_weight_cast_fp16 = constexpr_affine_dequantize()[axis = tensor(0), name = tensor("attn22_in_proj_weight_cast_fp16"), quantized_data = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(277249280))), scale = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(280395072))), zero_point = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(3145856)))]; + tensor attn22_out_proj_weight_cast_fp16 = constexpr_affine_dequantize()[axis = tensor(0), name = tensor("attn22_out_proj_weight_cast_fp16"), quantized_data = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(280401280))), scale = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(281449920))), zero_point = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(4203840)))]; + tensor linear22_1_weight_cast_fp16 = constexpr_affine_dequantize()[axis = tensor(0), name = tensor("linear22_1_weight_cast_fp16"), quantized_data = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(281452032))), scale = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(285646400))), zero_point = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8401408)))]; + tensor linear22_2_weight_cast_fp16 = constexpr_affine_dequantize()[axis = tensor(0), name = tensor("linear22_2_weight_cast_fp16"), quantized_data = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(285654656))), scale = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(289849024))), zero_point = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(4203840)))]; + tensor attn23_in_proj_weight_cast_fp16 = constexpr_affine_dequantize()[axis = tensor(0), name = tensor("attn23_in_proj_weight_cast_fp16"), quantized_data = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(289851136))), scale = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(292996928))), zero_point = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(3145856)))]; + tensor attn23_out_proj_weight_cast_fp16 = constexpr_affine_dequantize()[axis = tensor(0), name = tensor("attn23_out_proj_weight_cast_fp16"), quantized_data = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(293003136))), scale = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(294051776))), zero_point = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(4203840)))]; + tensor linear23_1_weight_cast_fp16 = constexpr_affine_dequantize()[axis = tensor(0), name = tensor("linear23_1_weight_cast_fp16"), quantized_data = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(294053888))), scale = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(298248256))), zero_point = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8401408)))]; + tensor linear23_2_weight_cast_fp16 = constexpr_affine_dequantize()[axis = tensor(0), name = tensor("linear23_2_weight_cast_fp16"), quantized_data = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(298256512))), scale = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(302450880))), zero_point = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(4203840)))]; + tensor sequence_to_fp16_dtype_0 = const()[name = tensor("sequence_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor sequence_to_fp16 = cast(dtype = sequence_to_fp16_dtype_0, x = sequence)[name = tensor("cast_555")]; + tensor var_198_cast_fp16 = not_equal(x = sequence_to_fp16, y = sequence_to_fp16)[name = tensor("op_198_cast_fp16")]; + tensor expand_dims_0_axes_0 = const()[name = tensor("expand_dims_0_axes_0"), val = tensor([0, 1])]; + tensor bos_emb_to_fp16_dtype_0 = const()[name = tensor("bos_emb_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor bos_emb_to_fp16 = cast(dtype = bos_emb_to_fp16_dtype_0, x = bos_emb)[name = tensor("cast_554")]; + tensor expand_dims_0_cast_fp16 = expand_dims(axes = expand_dims_0_axes_0, x = bos_emb_to_fp16)[name = tensor("expand_dims_0_cast_fp16")]; + tensor input_1_cast_fp16 = select(a = expand_dims_0_cast_fp16, b = sequence_to_fp16, cond = var_198_cast_fp16)[name = tensor("input_1_cast_fp16")]; + tensor input_linear_weight_to_fp16 = const()[name = tensor("input_linear_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(302452992)))]; + tensor linear_0_bias_0_to_fp16 = const()[name = tensor("linear_0_bias_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(302518592)))]; + tensor linear_0_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = input_linear_weight_to_fp16, x = input_1_cast_fp16)[name = tensor("linear_0_cast_fp16")]; + tensor x_1_axes_0 = const()[name = tensor("x_1_axes_0"), val = tensor([-1])]; + tensor norm0_1_weight_to_fp16 = const()[name = tensor("norm0_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(302520704)))]; + tensor norm0_1_bias_to_fp16 = const()[name = tensor("norm0_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(302522816)))]; + tensor var_204_to_fp16 = const()[name = tensor("op_204_to_fp16"), val = tensor(0x1.5p-17)]; + tensor x_1_cast_fp16 = layer_norm(axes = x_1_axes_0, beta = norm0_1_bias_to_fp16, epsilon = var_204_to_fp16, gamma = norm0_1_weight_to_fp16, x = linear_0_cast_fp16)[name = tensor("x_1_cast_fp16")]; + tensor linear_1_bias_0_to_fp16 = const()[name = tensor("linear_1_bias_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(302524928)))]; + tensor linear_1_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = attn0_in_proj_weight_cast_fp16, x = x_1_cast_fp16)[name = tensor("linear_1_cast_fp16")]; + tensor var_240 = const()[name = tensor("op_240"), val = tensor([1, 1, 3, 16, 64])]; + tensor qkv_1_cast_fp16 = reshape(shape = var_240, x = linear_1_cast_fp16)[name = tensor("qkv_1_cast_fp16")]; + tensor q_1_begin_0 = const()[name = tensor("q_1_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor q_1_end_0 = const()[name = tensor("q_1_end_0"), val = tensor([1, 1, 1, 16, 64])]; + tensor q_1_end_mask_0 = const()[name = tensor("q_1_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor q_1_squeeze_mask_0 = const()[name = tensor("q_1_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor q_1_cast_fp16 = slice_by_index(begin = q_1_begin_0, end = q_1_end_0, end_mask = q_1_end_mask_0, squeeze_mask = q_1_squeeze_mask_0, x = qkv_1_cast_fp16)[name = tensor("q_1_cast_fp16")]; + tensor k_1_begin_0 = const()[name = tensor("k_1_begin_0"), val = tensor([0, 0, 1, 0, 0])]; + tensor k_1_end_0 = const()[name = tensor("k_1_end_0"), val = tensor([1, 1, 2, 16, 64])]; + tensor k_1_end_mask_0 = const()[name = tensor("k_1_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor k_1_squeeze_mask_0 = const()[name = tensor("k_1_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor k_1_cast_fp16 = slice_by_index(begin = k_1_begin_0, end = k_1_end_0, end_mask = k_1_end_mask_0, squeeze_mask = k_1_squeeze_mask_0, x = qkv_1_cast_fp16)[name = tensor("k_1_cast_fp16")]; + tensor v_1_begin_0 = const()[name = tensor("v_1_begin_0"), val = tensor([0, 0, 2, 0, 0])]; + tensor v_1_end_0 = const()[name = tensor("v_1_end_0"), val = tensor([1, 1, 3, 16, 64])]; + tensor v_1_end_mask_0 = const()[name = tensor("v_1_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor v_1_squeeze_mask_0 = const()[name = tensor("v_1_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor v_1_cast_fp16 = slice_by_index(begin = v_1_begin_0, end = v_1_end_0, end_mask = v_1_end_mask_0, squeeze_mask = v_1_squeeze_mask_0, x = qkv_1_cast_fp16)[name = tensor("v_1_cast_fp16")]; + tensor var_344 = const()[name = tensor("op_344"), val = tensor([1, 1, 1, 1])]; + tensor position0_to_fp16_dtype_0 = const()[name = tensor("position0_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor position0_to_fp16 = cast(dtype = position0_to_fp16_dtype_0, x = position0)[name = tensor("cast_553")]; + tensor ts_5_cast_fp16 = reshape(shape = var_344, x = position0_to_fp16)[name = tensor("ts_5_cast_fp16")]; + tensor var_348 = const()[name = tensor("op_348"), val = tensor([1, 1, 16, 32, 2])]; + tensor q_complex_1_cast_fp16 = reshape(shape = var_348, x = q_1_cast_fp16)[name = tensor("q_complex_1_cast_fp16")]; + tensor var_352 = const()[name = tensor("op_352"), val = tensor([1, 1, 16, 32, 2])]; + tensor k_complex_1_cast_fp16 = reshape(shape = var_352, x = k_1_cast_fp16)[name = tensor("k_complex_1_cast_fp16")]; + tensor var_356_begin_0 = const()[name = tensor("op_356_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_356_end_0 = const()[name = tensor("op_356_end_0"), val = tensor([1, 1, 16, 32, 1])]; + tensor var_356_end_mask_0 = const()[name = tensor("op_356_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_356_squeeze_mask_0 = const()[name = tensor("op_356_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_356_cast_fp16 = slice_by_index(begin = var_356_begin_0, end = var_356_end_0, end_mask = var_356_end_mask_0, squeeze_mask = var_356_squeeze_mask_0, x = q_complex_1_cast_fp16)[name = tensor("op_356_cast_fp16")]; + tensor var_364_begin_0 = const()[name = tensor("op_364_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_364_end_0 = const()[name = tensor("op_364_end_0"), val = tensor([1, 1, 16, 32, 2])]; + tensor var_364_end_mask_0 = const()[name = tensor("op_364_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_364_squeeze_mask_0 = const()[name = tensor("op_364_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_364_cast_fp16 = slice_by_index(begin = var_364_begin_0, end = var_364_end_0, end_mask = var_364_end_mask_0, squeeze_mask = var_364_squeeze_mask_0, x = q_complex_1_cast_fp16)[name = tensor("op_364_cast_fp16")]; + tensor var_372_begin_0 = const()[name = tensor("op_372_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_372_end_0 = const()[name = tensor("op_372_end_0"), val = tensor([1, 1, 16, 32, 1])]; + tensor var_372_end_mask_0 = const()[name = tensor("op_372_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_372_squeeze_mask_0 = const()[name = tensor("op_372_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_372_cast_fp16 = slice_by_index(begin = var_372_begin_0, end = var_372_end_0, end_mask = var_372_end_mask_0, squeeze_mask = var_372_squeeze_mask_0, x = k_complex_1_cast_fp16)[name = tensor("op_372_cast_fp16")]; + tensor var_380_begin_0 = const()[name = tensor("op_380_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_380_end_0 = const()[name = tensor("op_380_end_0"), val = tensor([1, 1, 16, 32, 2])]; + tensor var_380_end_mask_0 = const()[name = tensor("op_380_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_380_squeeze_mask_0 = const()[name = tensor("op_380_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_380_cast_fp16 = slice_by_index(begin = var_380_begin_0, end = var_380_end_0, end_mask = var_380_end_mask_0, squeeze_mask = var_380_squeeze_mask_0, x = k_complex_1_cast_fp16)[name = tensor("op_380_cast_fp16")]; + tensor freqs_1_to_fp16 = const()[name = tensor("freqs_1_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(302531136)))]; + tensor var_386_cast_fp16 = mul(x = freqs_1_to_fp16, y = ts_5_cast_fp16)[name = tensor("op_386_cast_fp16")]; + tensor rotr_1_cast_fp16 = cos(x = var_386_cast_fp16)[name = tensor("rotr_1_cast_fp16")]; + tensor roti_1_cast_fp16 = sin(x = var_386_cast_fp16)[name = tensor("roti_1_cast_fp16")]; + tensor var_390_cast_fp16 = mul(x = var_356_cast_fp16, y = rotr_1_cast_fp16)[name = tensor("op_390_cast_fp16")]; + tensor var_391_cast_fp16 = mul(x = var_364_cast_fp16, y = roti_1_cast_fp16)[name = tensor("op_391_cast_fp16")]; + tensor qor_1_cast_fp16 = sub(x = var_390_cast_fp16, y = var_391_cast_fp16)[name = tensor("qor_1_cast_fp16")]; + tensor var_394_cast_fp16 = mul(x = var_356_cast_fp16, y = roti_1_cast_fp16)[name = tensor("op_394_cast_fp16")]; + tensor var_395_cast_fp16 = mul(x = var_364_cast_fp16, y = rotr_1_cast_fp16)[name = tensor("op_395_cast_fp16")]; + tensor qoi_1_cast_fp16 = add(x = var_394_cast_fp16, y = var_395_cast_fp16)[name = tensor("qoi_1_cast_fp16")]; + tensor var_398_cast_fp16 = mul(x = var_372_cast_fp16, y = rotr_1_cast_fp16)[name = tensor("op_398_cast_fp16")]; + tensor var_399_cast_fp16 = mul(x = var_380_cast_fp16, y = roti_1_cast_fp16)[name = tensor("op_399_cast_fp16")]; + tensor kor_1_cast_fp16 = sub(x = var_398_cast_fp16, y = var_399_cast_fp16)[name = tensor("kor_1_cast_fp16")]; + tensor var_402_cast_fp16 = mul(x = var_372_cast_fp16, y = roti_1_cast_fp16)[name = tensor("op_402_cast_fp16")]; + tensor var_403_cast_fp16 = mul(x = var_380_cast_fp16, y = rotr_1_cast_fp16)[name = tensor("op_403_cast_fp16")]; + tensor koi_1_cast_fp16 = add(x = var_402_cast_fp16, y = var_403_cast_fp16)[name = tensor("koi_1_cast_fp16")]; + tensor qo_1_axis_0 = const()[name = tensor("qo_1_axis_0"), val = tensor(-1)]; + tensor qo_1_cast_fp16 = stack(axis = qo_1_axis_0, values = (qor_1_cast_fp16, qoi_1_cast_fp16))[name = tensor("qo_1_cast_fp16")]; + tensor ko_1_axis_0 = const()[name = tensor("ko_1_axis_0"), val = tensor(-1)]; + tensor ko_1_cast_fp16 = stack(axis = ko_1_axis_0, values = (kor_1_cast_fp16, koi_1_cast_fp16))[name = tensor("ko_1_cast_fp16")]; + tensor var_432 = const()[name = tensor("op_432"), val = tensor([1, 1, 16, 64])]; + tensor q_3_cast_fp16 = reshape(shape = var_432, x = qo_1_cast_fp16)[name = tensor("q_3_cast_fp16")]; + tensor var_434 = const()[name = tensor("op_434"), val = tensor([1, 1, 16, 64])]; + tensor k_3_cast_fp16 = reshape(shape = var_434, x = ko_1_cast_fp16)[name = tensor("k_3_cast_fp16")]; + tensor _inversed_456_y_0_to_fp16 = const()[name = tensor("_inversed_456_y_0_to_fp16"), val = tensor(0x1p-9)]; + tensor _inversed_456_cast_fp16 = mul(x = ts_5_cast_fp16, y = _inversed_456_y_0_to_fp16)[name = tensor("_inversed_456_cast_fp16")]; + tensor var_457_cast_fp16 = floor(x = _inversed_456_cast_fp16)[name = tensor("op_457_cast_fp16")]; + tensor var_458_to_fp16 = const()[name = tensor("op_458_to_fp16"), val = tensor(0x1p+9)]; + tensor var_459_cast_fp16 = mul(x = var_457_cast_fp16, y = var_458_to_fp16)[name = tensor("op_459_cast_fp16")]; + tensor write_indices_float_3_cast_fp16 = sub(x = ts_5_cast_fp16, y = var_459_cast_fp16)[name = tensor("write_indices_float_3_cast_fp16")]; + tensor var_466_dtype_0 = const()[name = tensor("op_466_dtype_0"), val = tensor("int32")]; + tensor write_indices_1_reps_0 = const()[name = tensor("write_indices_1_reps_0"), val = tensor([1, 1, 16, 64])]; + tensor write_indices_float_3_cast_fp16_to_int32 = cast(dtype = var_466_dtype_0, x = write_indices_float_3_cast_fp16)[name = tensor("cast_552")]; + tensor write_indices_1 = tile(reps = write_indices_1_reps_0, x = write_indices_float_3_cast_fp16_to_int32)[name = tensor("write_indices_1")]; + tensor var_474_begin_0 = const()[name = tensor("op_474_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_474_end_0 = const()[name = tensor("op_474_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor var_474_end_mask_0 = const()[name = tensor("op_474_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_474_squeeze_mask_0 = const()[name = tensor("op_474_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor cache0_to_fp16_dtype_0 = const()[name = tensor("cache0_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor cache0_to_fp16 = cast(dtype = cache0_to_fp16_dtype_0, x = cache0)[name = tensor("cast_551")]; + tensor var_474_cast_fp16 = slice_by_index(begin = var_474_begin_0, end = var_474_end_0, end_mask = var_474_end_mask_0, squeeze_mask = var_474_squeeze_mask_0, x = cache0_to_fp16)[name = tensor("op_474_cast_fp16")]; + tensor var_476_axis_0 = const()[name = tensor("op_476_axis_0"), val = tensor(1)]; + tensor var_476_mode_0 = const()[name = tensor("op_476_mode_0"), val = tensor("update")]; + tensor var_476_validate_indices_0 = const()[name = tensor("op_476_validate_indices_0"), val = tensor(false)]; + tensor var_476_cast_fp16 = scatter_along_axis(axis = var_476_axis_0, data = var_474_cast_fp16, indices = write_indices_1, mode = var_476_mode_0, updates = k_3_cast_fp16, validate_indices = var_476_validate_indices_0)[name = tensor("op_476_cast_fp16")]; + tensor concat_2 = const()[name = tensor("concat_2"), val = tensor([0, 0, 0, 0, 0])]; + tensor concat_3 = const()[name = tensor("concat_3"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_1_internal_tensor_assign_1_stride_0 = const()[name = tensor("new_cache_1_internal_tensor_assign_1_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_1_internal_tensor_assign_1_begin_mask_0 = const()[name = tensor("new_cache_1_internal_tensor_assign_1_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_1_internal_tensor_assign_1_end_mask_0 = const()[name = tensor("new_cache_1_internal_tensor_assign_1_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_1_internal_tensor_assign_1_squeeze_mask_0 = const()[name = tensor("new_cache_1_internal_tensor_assign_1_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_48 = const()[name = tensor("shape_48"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_0 = const()[name = tensor("reduce_prod_0"), val = tensor(1048576)]; + tensor range_1d_0_start_0 = const()[name = tensor("range_1d_0_start_0"), val = tensor(0)]; + tensor range_1d_0_step_0 = const()[name = tensor("range_1d_0_step_0"), val = tensor(1)]; + tensor range_1d_0 = range_1d(end = reduce_prod_0, start = range_1d_0_start_0, step = range_1d_0_step_0)[name = tensor("range_1d_0")]; + tensor reshape_0 = reshape(shape = shape_48, x = range_1d_0)[name = tensor("reshape_0")]; + tensor slice_by_index_0 = slice_by_index(begin = concat_2, begin_mask = new_cache_1_internal_tensor_assign_1_begin_mask_0, end = concat_3, end_mask = new_cache_1_internal_tensor_assign_1_end_mask_0, squeeze_mask = new_cache_1_internal_tensor_assign_1_squeeze_mask_0, stride = new_cache_1_internal_tensor_assign_1_stride_0, x = reshape_0)[name = tensor("slice_by_index_0")]; + tensor reshape_1_shape_0 = const()[name = tensor("reshape_1_shape_0"), val = tensor([-1])]; + tensor reshape_1 = reshape(shape = reshape_1_shape_0, x = slice_by_index_0)[name = tensor("reshape_1")]; + tensor reshape_2_shape_0 = const()[name = tensor("reshape_2_shape_0"), val = tensor([-1])]; + tensor reshape_2_cast_fp16 = reshape(shape = reshape_2_shape_0, x = var_476_cast_fp16)[name = tensor("reshape_2_cast_fp16")]; + tensor reshape_3_shape_0 = const()[name = tensor("reshape_3_shape_0"), val = tensor([-1])]; + tensor reshape_3_cast_fp16 = reshape(shape = reshape_3_shape_0, x = cache0_to_fp16)[name = tensor("reshape_3_cast_fp16")]; + tensor scatter_0_mode_0 = const()[name = tensor("scatter_0_mode_0"), val = tensor("update")]; + tensor scatter_0_axis_0 = const()[name = tensor("scatter_0_axis_0"), val = tensor(0)]; + tensor scatter_0_validate_indices_0 = const()[name = tensor("scatter_0_validate_indices_0"), val = tensor(false)]; + tensor scatter_0_cast_fp16 = scatter(axis = scatter_0_axis_0, data = reshape_3_cast_fp16, indices = reshape_1, mode = scatter_0_mode_0, updates = reshape_2_cast_fp16, validate_indices = scatter_0_validate_indices_0)[name = tensor("scatter_0_cast_fp16")]; + tensor reshape_4_cast_fp16 = reshape(shape = shape_48, x = scatter_0_cast_fp16)[name = tensor("reshape_4_cast_fp16")]; + tensor var_484_begin_0 = const()[name = tensor("op_484_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor var_484_end_0 = const()[name = tensor("op_484_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor var_484_end_mask_0 = const()[name = tensor("op_484_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_484_squeeze_mask_0 = const()[name = tensor("op_484_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor var_484_cast_fp16 = slice_by_index(begin = var_484_begin_0, end = var_484_end_0, end_mask = var_484_end_mask_0, squeeze_mask = var_484_squeeze_mask_0, x = reshape_4_cast_fp16)[name = tensor("op_484_cast_fp16")]; + tensor var_486_axis_0 = const()[name = tensor("op_486_axis_0"), val = tensor(1)]; + tensor var_486_mode_0 = const()[name = tensor("op_486_mode_0"), val = tensor("update")]; + tensor var_486_validate_indices_0 = const()[name = tensor("op_486_validate_indices_0"), val = tensor(false)]; + tensor var_486_cast_fp16 = scatter_along_axis(axis = var_486_axis_0, data = var_484_cast_fp16, indices = write_indices_1, mode = var_486_mode_0, updates = v_1_cast_fp16, validate_indices = var_486_validate_indices_0)[name = tensor("op_486_cast_fp16")]; + tensor concat_4 = const()[name = tensor("concat_4"), val = tensor([1, 0, 0, 0, 0])]; + tensor concat_5 = const()[name = tensor("concat_5"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_1_internal_tensor_assign_2_stride_0 = const()[name = tensor("new_cache_1_internal_tensor_assign_2_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_1_internal_tensor_assign_2_begin_mask_0 = const()[name = tensor("new_cache_1_internal_tensor_assign_2_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_1_internal_tensor_assign_2_end_mask_0 = const()[name = tensor("new_cache_1_internal_tensor_assign_2_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_1_internal_tensor_assign_2_squeeze_mask_0 = const()[name = tensor("new_cache_1_internal_tensor_assign_2_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_49 = const()[name = tensor("shape_49"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_1 = const()[name = tensor("reduce_prod_1"), val = tensor(1048576)]; + tensor range_1d_1_start_0 = const()[name = tensor("range_1d_1_start_0"), val = tensor(0)]; + tensor range_1d_1_step_0 = const()[name = tensor("range_1d_1_step_0"), val = tensor(1)]; + tensor range_1d_1 = range_1d(end = reduce_prod_1, start = range_1d_1_start_0, step = range_1d_1_step_0)[name = tensor("range_1d_1")]; + tensor reshape_5 = reshape(shape = shape_49, x = range_1d_1)[name = tensor("reshape_5")]; + tensor slice_by_index_1 = slice_by_index(begin = concat_4, begin_mask = new_cache_1_internal_tensor_assign_2_begin_mask_0, end = concat_5, end_mask = new_cache_1_internal_tensor_assign_2_end_mask_0, squeeze_mask = new_cache_1_internal_tensor_assign_2_squeeze_mask_0, stride = new_cache_1_internal_tensor_assign_2_stride_0, x = reshape_5)[name = tensor("slice_by_index_1")]; + tensor reshape_6_shape_0 = const()[name = tensor("reshape_6_shape_0"), val = tensor([-1])]; + tensor reshape_6 = reshape(shape = reshape_6_shape_0, x = slice_by_index_1)[name = tensor("reshape_6")]; + tensor reshape_7_shape_0 = const()[name = tensor("reshape_7_shape_0"), val = tensor([-1])]; + tensor reshape_7_cast_fp16 = reshape(shape = reshape_7_shape_0, x = var_486_cast_fp16)[name = tensor("reshape_7_cast_fp16")]; + tensor reshape_8_shape_0 = const()[name = tensor("reshape_8_shape_0"), val = tensor([-1])]; + tensor reshape_8_cast_fp16 = reshape(shape = reshape_8_shape_0, x = reshape_4_cast_fp16)[name = tensor("reshape_8_cast_fp16")]; + tensor scatter_1_mode_0 = const()[name = tensor("scatter_1_mode_0"), val = tensor("update")]; + tensor scatter_1_axis_0 = const()[name = tensor("scatter_1_axis_0"), val = tensor(0)]; + tensor scatter_1_validate_indices_0 = const()[name = tensor("scatter_1_validate_indices_0"), val = tensor(false)]; + tensor scatter_1_cast_fp16 = scatter(axis = scatter_1_axis_0, data = reshape_8_cast_fp16, indices = reshape_6, mode = scatter_1_mode_0, updates = reshape_7_cast_fp16, validate_indices = scatter_1_validate_indices_0)[name = tensor("scatter_1_cast_fp16")]; + tensor reshape_9_cast_fp16 = reshape(shape = shape_49, x = scatter_1_cast_fp16)[name = tensor("reshape_9_cast_fp16")]; + tensor reshape_9_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("reshape_9_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor keys_1_begin_0 = const()[name = tensor("keys_1_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor keys_1_end_0 = const()[name = tensor("keys_1_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor keys_1_end_mask_0 = const()[name = tensor("keys_1_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor keys_1_squeeze_mask_0 = const()[name = tensor("keys_1_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor keys_1_cast_fp16 = slice_by_index(begin = keys_1_begin_0, end = keys_1_end_0, end_mask = keys_1_end_mask_0, squeeze_mask = keys_1_squeeze_mask_0, x = reshape_9_cast_fp16)[name = tensor("keys_1_cast_fp16")]; + tensor values_1_begin_0 = const()[name = tensor("values_1_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor values_1_end_0 = const()[name = tensor("values_1_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor values_1_end_mask_0 = const()[name = tensor("values_1_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor values_1_squeeze_mask_0 = const()[name = tensor("values_1_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor values_1_cast_fp16 = slice_by_index(begin = values_1_begin_0, end = values_1_end_0, end_mask = values_1_end_mask_0, squeeze_mask = values_1_squeeze_mask_0, x = reshape_9_cast_fp16)[name = tensor("values_1_cast_fp16")]; + tensor var_498_cast_fp16 = not_equal(x = keys_1_cast_fp16, y = keys_1_cast_fp16)[name = tensor("op_498_cast_fp16")]; + tensor var_504_to_fp16 = const()[name = tensor("op_504_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(302531264)))]; + tensor keys_3_cast_fp16 = select(a = var_504_to_fp16, b = keys_1_cast_fp16, cond = var_498_cast_fp16)[name = tensor("keys_3_cast_fp16")]; + tensor var_506_cast_fp16 = not_equal(x = values_1_cast_fp16, y = values_1_cast_fp16)[name = tensor("op_506_cast_fp16")]; + tensor values_3_cast_fp16 = select(a = var_504_to_fp16, b = values_1_cast_fp16, cond = var_506_cast_fp16)[name = tensor("values_3_cast_fp16")]; + tensor var_530 = const()[name = tensor("op_530"), val = tensor([0, 2, 1, 3])]; + tensor var_543 = const()[name = tensor("op_543"), val = tensor([1, 1, 1])]; + tensor var_544_cast_fp16 = reshape(shape = var_543, x = position0_to_fp16)[name = tensor("op_544_cast_fp16")]; + tensor var_561_to_fp16 = const()[name = tensor("op_561_to_fp16"), val = tensor(0x1p+0)]; + tensor valid_len_1_cast_fp16 = add(x = var_544_cast_fp16, y = var_561_to_fp16)[name = tensor("valid_len_1_cast_fp16")]; + tensor k_positions_1_promoted_to_fp16 = const()[name = tensor("k_positions_1_promoted_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(303579904)))]; + tensor valid_mask_1_cast_fp16 = less(x = k_positions_1_promoted_to_fp16, y = valid_len_1_cast_fp16)[name = tensor("valid_mask_1_cast_fp16")]; + tensor causal_mask_1_cast_fp16 = less_equal(x = k_positions_1_promoted_to_fp16, y = var_544_cast_fp16)[name = tensor("causal_mask_1_cast_fp16")]; + tensor attn_mask_1 = logical_and(x = valid_mask_1_cast_fp16, y = causal_mask_1_cast_fp16)[name = tensor("attn_mask_1")]; + tensor attn_mask_3_axes_0 = const()[name = tensor("attn_mask_3_axes_0"), val = tensor([1])]; + tensor attn_mask_3 = expand_dims(axes = attn_mask_3_axes_0, x = attn_mask_1)[name = tensor("attn_mask_3")]; + tensor var_579_transpose_x_0 = const()[name = tensor("op_579_transpose_x_0"), val = tensor(false)]; + tensor var_579_transpose_y_0 = const()[name = tensor("op_579_transpose_y_0"), val = tensor(false)]; + tensor transpose_72_perm_0 = const()[name = tensor("transpose_72_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_73_perm_0 = const()[name = tensor("transpose_73_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_73 = transpose(perm = transpose_73_perm_0, x = keys_3_cast_fp16)[name = tensor("transpose_213")]; + tensor transpose_72 = transpose(perm = transpose_72_perm_0, x = q_3_cast_fp16)[name = tensor("transpose_214")]; + tensor var_579_cast_fp16 = matmul(transpose_x = var_579_transpose_x_0, transpose_y = var_579_transpose_y_0, x = transpose_72, y = transpose_73)[name = tensor("op_579_cast_fp16")]; + tensor var_573_to_fp16 = const()[name = tensor("op_573_to_fp16"), val = tensor([0x1p-3])]; + tensor attn_weights_1_cast_fp16 = mul(x = var_579_cast_fp16, y = var_573_to_fp16)[name = tensor("attn_weights_1_cast_fp16")]; + tensor var_581 = logical_not(x = attn_mask_3)[name = tensor("op_581")]; + tensor var_582_to_fp16 = const()[name = tensor("op_582_to_fp16"), val = tensor(-inf)]; + tensor attn_weights_3_cast_fp16 = select(a = var_582_to_fp16, b = attn_weights_1_cast_fp16, cond = var_581)[name = tensor("attn_weights_3_cast_fp16")]; + tensor var_584 = const()[name = tensor("op_584"), val = tensor(-1)]; + tensor attn_weights_5_cast_fp16 = softmax(axis = var_584, x = attn_weights_3_cast_fp16)[name = tensor("attn_weights_5_cast_fp16")]; + tensor attn_output_1_transpose_x_0 = const()[name = tensor("attn_output_1_transpose_x_0"), val = tensor(false)]; + tensor attn_output_1_transpose_y_0 = const()[name = tensor("attn_output_1_transpose_y_0"), val = tensor(false)]; + tensor values_5_cast_fp16 = transpose(perm = var_530, x = values_3_cast_fp16)[name = tensor("transpose_215")]; + tensor attn_output_1_cast_fp16 = matmul(transpose_x = attn_output_1_transpose_x_0, transpose_y = attn_output_1_transpose_y_0, x = attn_weights_5_cast_fp16, y = values_5_cast_fp16)[name = tensor("attn_output_1_cast_fp16")]; + tensor var_592 = const()[name = tensor("op_592"), val = tensor([0, 2, 1, 3])]; + tensor var_595 = const()[name = tensor("op_595"), val = tensor([1, 1, 1024])]; + tensor var_593_cast_fp16 = transpose(perm = var_592, x = attn_output_1_cast_fp16)[name = tensor("transpose_212")]; + tensor input_5_cast_fp16 = reshape(shape = var_595, x = var_593_cast_fp16)[name = tensor("input_5_cast_fp16")]; + tensor linear_2_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = attn0_out_proj_weight_cast_fp16, x = input_5_cast_fp16)[name = tensor("linear_2_cast_fp16")]; + tensor var_601_to_fp16 = const()[name = tensor("op_601_to_fp16"), val = tensor(0x1p+0)]; + tensor var_602_cast_fp16 = add(x = position0_to_fp16, y = var_601_to_fp16)[name = tensor("op_602_cast_fp16")]; + tensor var_602_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_602_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor input_7_cast_fp16 = add(x = linear_0_cast_fp16, y = linear_2_cast_fp16)[name = tensor("input_7_cast_fp16")]; + tensor input_9_axes_0 = const()[name = tensor("input_9_axes_0"), val = tensor([-1])]; + tensor norm0_2_weight_to_fp16 = const()[name = tensor("norm0_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(303580992)))]; + tensor norm0_2_bias_to_fp16 = const()[name = tensor("norm0_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(303583104)))]; + tensor var_606_to_fp16 = const()[name = tensor("op_606_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_9_cast_fp16 = layer_norm(axes = input_9_axes_0, beta = norm0_2_bias_to_fp16, epsilon = var_606_to_fp16, gamma = norm0_2_weight_to_fp16, x = input_7_cast_fp16)[name = tensor("input_9_cast_fp16")]; + tensor linear_3_bias_0_to_fp16 = const()[name = tensor("linear_3_bias_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(303585216)))]; + tensor linear_3_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = linear0_1_weight_cast_fp16, x = input_9_cast_fp16)[name = tensor("linear_3_cast_fp16")]; + tensor input_11_mode_0 = const()[name = tensor("input_11_mode_0"), val = tensor("EXACT")]; + tensor input_11_cast_fp16 = gelu(mode = input_11_mode_0, x = linear_3_cast_fp16)[name = tensor("input_11_cast_fp16")]; + tensor linear_4_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = linear0_2_weight_cast_fp16, x = input_11_cast_fp16)[name = tensor("linear_4_cast_fp16")]; + tensor input_13_cast_fp16 = add(x = input_7_cast_fp16, y = linear_4_cast_fp16)[name = tensor("input_13_cast_fp16")]; + tensor x_3_axes_0 = const()[name = tensor("x_3_axes_0"), val = tensor([-1])]; + tensor norm1_1_weight_to_fp16 = const()[name = tensor("norm1_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(303593472)))]; + tensor norm1_1_bias_to_fp16 = const()[name = tensor("norm1_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(303595584)))]; + tensor var_623_to_fp16 = const()[name = tensor("op_623_to_fp16"), val = tensor(0x1.5p-17)]; + tensor x_3_cast_fp16 = layer_norm(axes = x_3_axes_0, beta = norm1_1_bias_to_fp16, epsilon = var_623_to_fp16, gamma = norm1_1_weight_to_fp16, x = input_13_cast_fp16)[name = tensor("x_3_cast_fp16")]; + tensor linear_5_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = attn1_in_proj_weight_cast_fp16, x = x_3_cast_fp16)[name = tensor("linear_5_cast_fp16")]; + tensor var_659 = const()[name = tensor("op_659"), val = tensor([1, 1, 3, 16, 64])]; + tensor qkv_3_cast_fp16 = reshape(shape = var_659, x = linear_5_cast_fp16)[name = tensor("qkv_3_cast_fp16")]; + tensor q_7_begin_0 = const()[name = tensor("q_7_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor q_7_end_0 = const()[name = tensor("q_7_end_0"), val = tensor([1, 1, 1, 16, 64])]; + tensor q_7_end_mask_0 = const()[name = tensor("q_7_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor q_7_squeeze_mask_0 = const()[name = tensor("q_7_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor q_7_cast_fp16 = slice_by_index(begin = q_7_begin_0, end = q_7_end_0, end_mask = q_7_end_mask_0, squeeze_mask = q_7_squeeze_mask_0, x = qkv_3_cast_fp16)[name = tensor("q_7_cast_fp16")]; + tensor k_5_begin_0 = const()[name = tensor("k_5_begin_0"), val = tensor([0, 0, 1, 0, 0])]; + tensor k_5_end_0 = const()[name = tensor("k_5_end_0"), val = tensor([1, 1, 2, 16, 64])]; + tensor k_5_end_mask_0 = const()[name = tensor("k_5_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor k_5_squeeze_mask_0 = const()[name = tensor("k_5_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor k_5_cast_fp16 = slice_by_index(begin = k_5_begin_0, end = k_5_end_0, end_mask = k_5_end_mask_0, squeeze_mask = k_5_squeeze_mask_0, x = qkv_3_cast_fp16)[name = tensor("k_5_cast_fp16")]; + tensor v_3_begin_0 = const()[name = tensor("v_3_begin_0"), val = tensor([0, 0, 2, 0, 0])]; + tensor v_3_end_0 = const()[name = tensor("v_3_end_0"), val = tensor([1, 1, 3, 16, 64])]; + tensor v_3_end_mask_0 = const()[name = tensor("v_3_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor v_3_squeeze_mask_0 = const()[name = tensor("v_3_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor v_3_cast_fp16 = slice_by_index(begin = v_3_begin_0, end = v_3_end_0, end_mask = v_3_end_mask_0, squeeze_mask = v_3_squeeze_mask_0, x = qkv_3_cast_fp16)[name = tensor("v_3_cast_fp16")]; + tensor var_763 = const()[name = tensor("op_763"), val = tensor([1, 1, 1, 1])]; + tensor position1_to_fp16_dtype_0 = const()[name = tensor("position1_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor position1_to_fp16 = cast(dtype = position1_to_fp16_dtype_0, x = position1)[name = tensor("cast_548")]; + tensor ts_11_cast_fp16 = reshape(shape = var_763, x = position1_to_fp16)[name = tensor("ts_11_cast_fp16")]; + tensor var_767 = const()[name = tensor("op_767"), val = tensor([1, 1, 16, 32, 2])]; + tensor q_complex_3_cast_fp16 = reshape(shape = var_767, x = q_7_cast_fp16)[name = tensor("q_complex_3_cast_fp16")]; + tensor var_771 = const()[name = tensor("op_771"), val = tensor([1, 1, 16, 32, 2])]; + tensor k_complex_3_cast_fp16 = reshape(shape = var_771, x = k_5_cast_fp16)[name = tensor("k_complex_3_cast_fp16")]; + tensor var_775_begin_0 = const()[name = tensor("op_775_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_775_end_0 = const()[name = tensor("op_775_end_0"), val = tensor([1, 1, 16, 32, 1])]; + tensor var_775_end_mask_0 = const()[name = tensor("op_775_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_775_squeeze_mask_0 = const()[name = tensor("op_775_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_775_cast_fp16 = slice_by_index(begin = var_775_begin_0, end = var_775_end_0, end_mask = var_775_end_mask_0, squeeze_mask = var_775_squeeze_mask_0, x = q_complex_3_cast_fp16)[name = tensor("op_775_cast_fp16")]; + tensor var_783_begin_0 = const()[name = tensor("op_783_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_783_end_0 = const()[name = tensor("op_783_end_0"), val = tensor([1, 1, 16, 32, 2])]; + tensor var_783_end_mask_0 = const()[name = tensor("op_783_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_783_squeeze_mask_0 = const()[name = tensor("op_783_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_783_cast_fp16 = slice_by_index(begin = var_783_begin_0, end = var_783_end_0, end_mask = var_783_end_mask_0, squeeze_mask = var_783_squeeze_mask_0, x = q_complex_3_cast_fp16)[name = tensor("op_783_cast_fp16")]; + tensor var_791_begin_0 = const()[name = tensor("op_791_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_791_end_0 = const()[name = tensor("op_791_end_0"), val = tensor([1, 1, 16, 32, 1])]; + tensor var_791_end_mask_0 = const()[name = tensor("op_791_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_791_squeeze_mask_0 = const()[name = tensor("op_791_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_791_cast_fp16 = slice_by_index(begin = var_791_begin_0, end = var_791_end_0, end_mask = var_791_end_mask_0, squeeze_mask = var_791_squeeze_mask_0, x = k_complex_3_cast_fp16)[name = tensor("op_791_cast_fp16")]; + tensor var_799_begin_0 = const()[name = tensor("op_799_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_799_end_0 = const()[name = tensor("op_799_end_0"), val = tensor([1, 1, 16, 32, 2])]; + tensor var_799_end_mask_0 = const()[name = tensor("op_799_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_799_squeeze_mask_0 = const()[name = tensor("op_799_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_799_cast_fp16 = slice_by_index(begin = var_799_begin_0, end = var_799_end_0, end_mask = var_799_end_mask_0, squeeze_mask = var_799_squeeze_mask_0, x = k_complex_3_cast_fp16)[name = tensor("op_799_cast_fp16")]; + tensor freqs_3_to_fp16 = const()[name = tensor("freqs_3_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(303597696)))]; + tensor var_805_cast_fp16 = mul(x = freqs_3_to_fp16, y = ts_11_cast_fp16)[name = tensor("op_805_cast_fp16")]; + tensor rotr_3_cast_fp16 = cos(x = var_805_cast_fp16)[name = tensor("rotr_3_cast_fp16")]; + tensor roti_3_cast_fp16 = sin(x = var_805_cast_fp16)[name = tensor("roti_3_cast_fp16")]; + tensor var_809_cast_fp16 = mul(x = var_775_cast_fp16, y = rotr_3_cast_fp16)[name = tensor("op_809_cast_fp16")]; + tensor var_810_cast_fp16 = mul(x = var_783_cast_fp16, y = roti_3_cast_fp16)[name = tensor("op_810_cast_fp16")]; + tensor qor_5_cast_fp16 = sub(x = var_809_cast_fp16, y = var_810_cast_fp16)[name = tensor("qor_5_cast_fp16")]; + tensor var_813_cast_fp16 = mul(x = var_775_cast_fp16, y = roti_3_cast_fp16)[name = tensor("op_813_cast_fp16")]; + tensor var_814_cast_fp16 = mul(x = var_783_cast_fp16, y = rotr_3_cast_fp16)[name = tensor("op_814_cast_fp16")]; + tensor qoi_5_cast_fp16 = add(x = var_813_cast_fp16, y = var_814_cast_fp16)[name = tensor("qoi_5_cast_fp16")]; + tensor var_817_cast_fp16 = mul(x = var_791_cast_fp16, y = rotr_3_cast_fp16)[name = tensor("op_817_cast_fp16")]; + tensor var_818_cast_fp16 = mul(x = var_799_cast_fp16, y = roti_3_cast_fp16)[name = tensor("op_818_cast_fp16")]; + tensor kor_5_cast_fp16 = sub(x = var_817_cast_fp16, y = var_818_cast_fp16)[name = tensor("kor_5_cast_fp16")]; + tensor var_821_cast_fp16 = mul(x = var_791_cast_fp16, y = roti_3_cast_fp16)[name = tensor("op_821_cast_fp16")]; + tensor var_822_cast_fp16 = mul(x = var_799_cast_fp16, y = rotr_3_cast_fp16)[name = tensor("op_822_cast_fp16")]; + tensor koi_5_cast_fp16 = add(x = var_821_cast_fp16, y = var_822_cast_fp16)[name = tensor("koi_5_cast_fp16")]; + tensor qo_3_axis_0 = const()[name = tensor("qo_3_axis_0"), val = tensor(-1)]; + tensor qo_3_cast_fp16 = stack(axis = qo_3_axis_0, values = (qor_5_cast_fp16, qoi_5_cast_fp16))[name = tensor("qo_3_cast_fp16")]; + tensor ko_3_axis_0 = const()[name = tensor("ko_3_axis_0"), val = tensor(-1)]; + tensor ko_3_cast_fp16 = stack(axis = ko_3_axis_0, values = (kor_5_cast_fp16, koi_5_cast_fp16))[name = tensor("ko_3_cast_fp16")]; + tensor var_851 = const()[name = tensor("op_851"), val = tensor([1, 1, 16, 64])]; + tensor q_9_cast_fp16 = reshape(shape = var_851, x = qo_3_cast_fp16)[name = tensor("q_9_cast_fp16")]; + tensor var_853 = const()[name = tensor("op_853"), val = tensor([1, 1, 16, 64])]; + tensor k_7_cast_fp16 = reshape(shape = var_853, x = ko_3_cast_fp16)[name = tensor("k_7_cast_fp16")]; + tensor _inversed_875_y_0_to_fp16 = const()[name = tensor("_inversed_875_y_0_to_fp16"), val = tensor(0x1p-9)]; + tensor _inversed_875_cast_fp16 = mul(x = ts_11_cast_fp16, y = _inversed_875_y_0_to_fp16)[name = tensor("_inversed_875_cast_fp16")]; + tensor var_876_cast_fp16 = floor(x = _inversed_875_cast_fp16)[name = tensor("op_876_cast_fp16")]; + tensor var_877_to_fp16 = const()[name = tensor("op_877_to_fp16"), val = tensor(0x1p+9)]; + tensor var_878_cast_fp16 = mul(x = var_876_cast_fp16, y = var_877_to_fp16)[name = tensor("op_878_cast_fp16")]; + tensor write_indices_float_7_cast_fp16 = sub(x = ts_11_cast_fp16, y = var_878_cast_fp16)[name = tensor("write_indices_float_7_cast_fp16")]; + tensor var_885_dtype_0 = const()[name = tensor("op_885_dtype_0"), val = tensor("int32")]; + tensor write_indices_3_reps_0 = const()[name = tensor("write_indices_3_reps_0"), val = tensor([1, 1, 16, 64])]; + tensor write_indices_float_7_cast_fp16_to_int32 = cast(dtype = var_885_dtype_0, x = write_indices_float_7_cast_fp16)[name = tensor("cast_547")]; + tensor write_indices_3 = tile(reps = write_indices_3_reps_0, x = write_indices_float_7_cast_fp16_to_int32)[name = tensor("write_indices_3")]; + tensor var_893_begin_0 = const()[name = tensor("op_893_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_893_end_0 = const()[name = tensor("op_893_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor var_893_end_mask_0 = const()[name = tensor("op_893_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_893_squeeze_mask_0 = const()[name = tensor("op_893_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor cache1_to_fp16_dtype_0 = const()[name = tensor("cache1_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor cache1_to_fp16 = cast(dtype = cache1_to_fp16_dtype_0, x = cache1)[name = tensor("cast_546")]; + tensor var_893_cast_fp16 = slice_by_index(begin = var_893_begin_0, end = var_893_end_0, end_mask = var_893_end_mask_0, squeeze_mask = var_893_squeeze_mask_0, x = cache1_to_fp16)[name = tensor("op_893_cast_fp16")]; + tensor var_895_axis_0 = const()[name = tensor("op_895_axis_0"), val = tensor(1)]; + tensor var_895_mode_0 = const()[name = tensor("op_895_mode_0"), val = tensor("update")]; + tensor var_895_validate_indices_0 = const()[name = tensor("op_895_validate_indices_0"), val = tensor(false)]; + tensor var_895_cast_fp16 = scatter_along_axis(axis = var_895_axis_0, data = var_893_cast_fp16, indices = write_indices_3, mode = var_895_mode_0, updates = k_7_cast_fp16, validate_indices = var_895_validate_indices_0)[name = tensor("op_895_cast_fp16")]; + tensor concat_9 = const()[name = tensor("concat_9"), val = tensor([0, 0, 0, 0, 0])]; + tensor concat_10 = const()[name = tensor("concat_10"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_3_internal_tensor_assign_1_stride_0 = const()[name = tensor("new_cache_3_internal_tensor_assign_1_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_3_internal_tensor_assign_1_begin_mask_0 = const()[name = tensor("new_cache_3_internal_tensor_assign_1_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_3_internal_tensor_assign_1_end_mask_0 = const()[name = tensor("new_cache_3_internal_tensor_assign_1_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_3_internal_tensor_assign_1_squeeze_mask_0 = const()[name = tensor("new_cache_3_internal_tensor_assign_1_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_50 = const()[name = tensor("shape_50"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_2 = const()[name = tensor("reduce_prod_2"), val = tensor(1048576)]; + tensor range_1d_2_start_0 = const()[name = tensor("range_1d_2_start_0"), val = tensor(0)]; + tensor range_1d_2_step_0 = const()[name = tensor("range_1d_2_step_0"), val = tensor(1)]; + tensor range_1d_2 = range_1d(end = reduce_prod_2, start = range_1d_2_start_0, step = range_1d_2_step_0)[name = tensor("range_1d_2")]; + tensor reshape_10 = reshape(shape = shape_50, x = range_1d_2)[name = tensor("reshape_10")]; + tensor slice_by_index_2 = slice_by_index(begin = concat_9, begin_mask = new_cache_3_internal_tensor_assign_1_begin_mask_0, end = concat_10, end_mask = new_cache_3_internal_tensor_assign_1_end_mask_0, squeeze_mask = new_cache_3_internal_tensor_assign_1_squeeze_mask_0, stride = new_cache_3_internal_tensor_assign_1_stride_0, x = reshape_10)[name = tensor("slice_by_index_2")]; + tensor reshape_11_shape_0 = const()[name = tensor("reshape_11_shape_0"), val = tensor([-1])]; + tensor reshape_11 = reshape(shape = reshape_11_shape_0, x = slice_by_index_2)[name = tensor("reshape_11")]; + tensor reshape_12_shape_0 = const()[name = tensor("reshape_12_shape_0"), val = tensor([-1])]; + tensor reshape_12_cast_fp16 = reshape(shape = reshape_12_shape_0, x = var_895_cast_fp16)[name = tensor("reshape_12_cast_fp16")]; + tensor reshape_13_shape_0 = const()[name = tensor("reshape_13_shape_0"), val = tensor([-1])]; + tensor reshape_13_cast_fp16 = reshape(shape = reshape_13_shape_0, x = cache1_to_fp16)[name = tensor("reshape_13_cast_fp16")]; + tensor scatter_2_mode_0 = const()[name = tensor("scatter_2_mode_0"), val = tensor("update")]; + tensor scatter_2_axis_0 = const()[name = tensor("scatter_2_axis_0"), val = tensor(0)]; + tensor scatter_2_validate_indices_0 = const()[name = tensor("scatter_2_validate_indices_0"), val = tensor(false)]; + tensor scatter_2_cast_fp16 = scatter(axis = scatter_2_axis_0, data = reshape_13_cast_fp16, indices = reshape_11, mode = scatter_2_mode_0, updates = reshape_12_cast_fp16, validate_indices = scatter_2_validate_indices_0)[name = tensor("scatter_2_cast_fp16")]; + tensor reshape_14_cast_fp16 = reshape(shape = shape_50, x = scatter_2_cast_fp16)[name = tensor("reshape_14_cast_fp16")]; + tensor var_903_begin_0 = const()[name = tensor("op_903_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor var_903_end_0 = const()[name = tensor("op_903_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor var_903_end_mask_0 = const()[name = tensor("op_903_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_903_squeeze_mask_0 = const()[name = tensor("op_903_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor var_903_cast_fp16 = slice_by_index(begin = var_903_begin_0, end = var_903_end_0, end_mask = var_903_end_mask_0, squeeze_mask = var_903_squeeze_mask_0, x = reshape_14_cast_fp16)[name = tensor("op_903_cast_fp16")]; + tensor var_905_axis_0 = const()[name = tensor("op_905_axis_0"), val = tensor(1)]; + tensor var_905_mode_0 = const()[name = tensor("op_905_mode_0"), val = tensor("update")]; + tensor var_905_validate_indices_0 = const()[name = tensor("op_905_validate_indices_0"), val = tensor(false)]; + tensor var_905_cast_fp16 = scatter_along_axis(axis = var_905_axis_0, data = var_903_cast_fp16, indices = write_indices_3, mode = var_905_mode_0, updates = v_3_cast_fp16, validate_indices = var_905_validate_indices_0)[name = tensor("op_905_cast_fp16")]; + tensor concat_11 = const()[name = tensor("concat_11"), val = tensor([1, 0, 0, 0, 0])]; + tensor concat_12 = const()[name = tensor("concat_12"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_3_internal_tensor_assign_2_stride_0 = const()[name = tensor("new_cache_3_internal_tensor_assign_2_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_3_internal_tensor_assign_2_begin_mask_0 = const()[name = tensor("new_cache_3_internal_tensor_assign_2_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_3_internal_tensor_assign_2_end_mask_0 = const()[name = tensor("new_cache_3_internal_tensor_assign_2_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_3_internal_tensor_assign_2_squeeze_mask_0 = const()[name = tensor("new_cache_3_internal_tensor_assign_2_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_51 = const()[name = tensor("shape_51"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_3 = const()[name = tensor("reduce_prod_3"), val = tensor(1048576)]; + tensor range_1d_3_start_0 = const()[name = tensor("range_1d_3_start_0"), val = tensor(0)]; + tensor range_1d_3_step_0 = const()[name = tensor("range_1d_3_step_0"), val = tensor(1)]; + tensor range_1d_3 = range_1d(end = reduce_prod_3, start = range_1d_3_start_0, step = range_1d_3_step_0)[name = tensor("range_1d_3")]; + tensor reshape_15 = reshape(shape = shape_51, x = range_1d_3)[name = tensor("reshape_15")]; + tensor slice_by_index_3 = slice_by_index(begin = concat_11, begin_mask = new_cache_3_internal_tensor_assign_2_begin_mask_0, end = concat_12, end_mask = new_cache_3_internal_tensor_assign_2_end_mask_0, squeeze_mask = new_cache_3_internal_tensor_assign_2_squeeze_mask_0, stride = new_cache_3_internal_tensor_assign_2_stride_0, x = reshape_15)[name = tensor("slice_by_index_3")]; + tensor reshape_16_shape_0 = const()[name = tensor("reshape_16_shape_0"), val = tensor([-1])]; + tensor reshape_16 = reshape(shape = reshape_16_shape_0, x = slice_by_index_3)[name = tensor("reshape_16")]; + tensor reshape_17_shape_0 = const()[name = tensor("reshape_17_shape_0"), val = tensor([-1])]; + tensor reshape_17_cast_fp16 = reshape(shape = reshape_17_shape_0, x = var_905_cast_fp16)[name = tensor("reshape_17_cast_fp16")]; + tensor reshape_18_shape_0 = const()[name = tensor("reshape_18_shape_0"), val = tensor([-1])]; + tensor reshape_18_cast_fp16 = reshape(shape = reshape_18_shape_0, x = reshape_14_cast_fp16)[name = tensor("reshape_18_cast_fp16")]; + tensor scatter_3_mode_0 = const()[name = tensor("scatter_3_mode_0"), val = tensor("update")]; + tensor scatter_3_axis_0 = const()[name = tensor("scatter_3_axis_0"), val = tensor(0)]; + tensor scatter_3_validate_indices_0 = const()[name = tensor("scatter_3_validate_indices_0"), val = tensor(false)]; + tensor scatter_3_cast_fp16 = scatter(axis = scatter_3_axis_0, data = reshape_18_cast_fp16, indices = reshape_16, mode = scatter_3_mode_0, updates = reshape_17_cast_fp16, validate_indices = scatter_3_validate_indices_0)[name = tensor("scatter_3_cast_fp16")]; + tensor reshape_19_cast_fp16 = reshape(shape = shape_51, x = scatter_3_cast_fp16)[name = tensor("reshape_19_cast_fp16")]; + tensor reshape_19_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("reshape_19_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor keys_7_begin_0 = const()[name = tensor("keys_7_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor keys_7_end_0 = const()[name = tensor("keys_7_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor keys_7_end_mask_0 = const()[name = tensor("keys_7_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor keys_7_squeeze_mask_0 = const()[name = tensor("keys_7_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor keys_7_cast_fp16 = slice_by_index(begin = keys_7_begin_0, end = keys_7_end_0, end_mask = keys_7_end_mask_0, squeeze_mask = keys_7_squeeze_mask_0, x = reshape_19_cast_fp16)[name = tensor("keys_7_cast_fp16")]; + tensor values_7_begin_0 = const()[name = tensor("values_7_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor values_7_end_0 = const()[name = tensor("values_7_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor values_7_end_mask_0 = const()[name = tensor("values_7_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor values_7_squeeze_mask_0 = const()[name = tensor("values_7_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor values_7_cast_fp16 = slice_by_index(begin = values_7_begin_0, end = values_7_end_0, end_mask = values_7_end_mask_0, squeeze_mask = values_7_squeeze_mask_0, x = reshape_19_cast_fp16)[name = tensor("values_7_cast_fp16")]; + tensor var_917_cast_fp16 = not_equal(x = keys_7_cast_fp16, y = keys_7_cast_fp16)[name = tensor("op_917_cast_fp16")]; + tensor keys_9_cast_fp16 = select(a = var_504_to_fp16, b = keys_7_cast_fp16, cond = var_917_cast_fp16)[name = tensor("keys_9_cast_fp16")]; + tensor var_925_cast_fp16 = not_equal(x = values_7_cast_fp16, y = values_7_cast_fp16)[name = tensor("op_925_cast_fp16")]; + tensor values_9_cast_fp16 = select(a = var_504_to_fp16, b = values_7_cast_fp16, cond = var_925_cast_fp16)[name = tensor("values_9_cast_fp16")]; + tensor var_949 = const()[name = tensor("op_949"), val = tensor([0, 2, 1, 3])]; + tensor var_962 = const()[name = tensor("op_962"), val = tensor([1, 1, 1])]; + tensor var_963_cast_fp16 = reshape(shape = var_962, x = position1_to_fp16)[name = tensor("op_963_cast_fp16")]; + tensor var_980_to_fp16 = const()[name = tensor("op_980_to_fp16"), val = tensor(0x1p+0)]; + tensor valid_len_3_cast_fp16 = add(x = var_963_cast_fp16, y = var_980_to_fp16)[name = tensor("valid_len_3_cast_fp16")]; + tensor valid_mask_3_cast_fp16 = less(x = k_positions_1_promoted_to_fp16, y = valid_len_3_cast_fp16)[name = tensor("valid_mask_3_cast_fp16")]; + tensor causal_mask_3_cast_fp16 = less_equal(x = k_positions_1_promoted_to_fp16, y = var_963_cast_fp16)[name = tensor("causal_mask_3_cast_fp16")]; + tensor attn_mask_5 = logical_and(x = valid_mask_3_cast_fp16, y = causal_mask_3_cast_fp16)[name = tensor("attn_mask_5")]; + tensor attn_mask_7_axes_0 = const()[name = tensor("attn_mask_7_axes_0"), val = tensor([1])]; + tensor attn_mask_7 = expand_dims(axes = attn_mask_7_axes_0, x = attn_mask_5)[name = tensor("attn_mask_7")]; + tensor var_998_transpose_x_0 = const()[name = tensor("op_998_transpose_x_0"), val = tensor(false)]; + tensor var_998_transpose_y_0 = const()[name = tensor("op_998_transpose_y_0"), val = tensor(false)]; + tensor transpose_74_perm_0 = const()[name = tensor("transpose_74_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_75_perm_0 = const()[name = tensor("transpose_75_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_75 = transpose(perm = transpose_75_perm_0, x = keys_9_cast_fp16)[name = tensor("transpose_209")]; + tensor transpose_74 = transpose(perm = transpose_74_perm_0, x = q_9_cast_fp16)[name = tensor("transpose_210")]; + tensor var_998_cast_fp16 = matmul(transpose_x = var_998_transpose_x_0, transpose_y = var_998_transpose_y_0, x = transpose_74, y = transpose_75)[name = tensor("op_998_cast_fp16")]; + tensor var_992_to_fp16 = const()[name = tensor("op_992_to_fp16"), val = tensor([0x1p-3])]; + tensor attn_weights_7_cast_fp16 = mul(x = var_998_cast_fp16, y = var_992_to_fp16)[name = tensor("attn_weights_7_cast_fp16")]; + tensor var_1000 = logical_not(x = attn_mask_7)[name = tensor("op_1000")]; + tensor var_1001_to_fp16 = const()[name = tensor("op_1001_to_fp16"), val = tensor(-inf)]; + tensor attn_weights_9_cast_fp16 = select(a = var_1001_to_fp16, b = attn_weights_7_cast_fp16, cond = var_1000)[name = tensor("attn_weights_9_cast_fp16")]; + tensor var_1003 = const()[name = tensor("op_1003"), val = tensor(-1)]; + tensor attn_weights_11_cast_fp16 = softmax(axis = var_1003, x = attn_weights_9_cast_fp16)[name = tensor("attn_weights_11_cast_fp16")]; + tensor attn_output_3_transpose_x_0 = const()[name = tensor("attn_output_3_transpose_x_0"), val = tensor(false)]; + tensor attn_output_3_transpose_y_0 = const()[name = tensor("attn_output_3_transpose_y_0"), val = tensor(false)]; + tensor values_11_cast_fp16 = transpose(perm = var_949, x = values_9_cast_fp16)[name = tensor("transpose_211")]; + tensor attn_output_3_cast_fp16 = matmul(transpose_x = attn_output_3_transpose_x_0, transpose_y = attn_output_3_transpose_y_0, x = attn_weights_11_cast_fp16, y = values_11_cast_fp16)[name = tensor("attn_output_3_cast_fp16")]; + tensor var_1011 = const()[name = tensor("op_1011"), val = tensor([0, 2, 1, 3])]; + tensor var_1014 = const()[name = tensor("op_1014"), val = tensor([1, 1, 1024])]; + tensor var_1012_cast_fp16 = transpose(perm = var_1011, x = attn_output_3_cast_fp16)[name = tensor("transpose_208")]; + tensor input_15_cast_fp16 = reshape(shape = var_1014, x = var_1012_cast_fp16)[name = tensor("input_15_cast_fp16")]; + tensor linear_6_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = attn1_out_proj_weight_cast_fp16, x = input_15_cast_fp16)[name = tensor("linear_6_cast_fp16")]; + tensor var_1020_to_fp16 = const()[name = tensor("op_1020_to_fp16"), val = tensor(0x1p+0)]; + tensor var_1021_cast_fp16 = add(x = position1_to_fp16, y = var_1020_to_fp16)[name = tensor("op_1021_cast_fp16")]; + tensor var_1021_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_1021_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor input_17_cast_fp16 = add(x = input_13_cast_fp16, y = linear_6_cast_fp16)[name = tensor("input_17_cast_fp16")]; + tensor input_19_axes_0 = const()[name = tensor("input_19_axes_0"), val = tensor([-1])]; + tensor norm1_2_weight_to_fp16 = const()[name = tensor("norm1_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(303597824)))]; + tensor norm1_2_bias_to_fp16 = const()[name = tensor("norm1_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(303599936)))]; + tensor var_1025_to_fp16 = const()[name = tensor("op_1025_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_19_cast_fp16 = layer_norm(axes = input_19_axes_0, beta = norm1_2_bias_to_fp16, epsilon = var_1025_to_fp16, gamma = norm1_2_weight_to_fp16, x = input_17_cast_fp16)[name = tensor("input_19_cast_fp16")]; + tensor linear_7_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = linear1_1_weight_cast_fp16, x = input_19_cast_fp16)[name = tensor("linear_7_cast_fp16")]; + tensor input_21_mode_0 = const()[name = tensor("input_21_mode_0"), val = tensor("EXACT")]; + tensor input_21_cast_fp16 = gelu(mode = input_21_mode_0, x = linear_7_cast_fp16)[name = tensor("input_21_cast_fp16")]; + tensor linear_8_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = linear1_2_weight_cast_fp16, x = input_21_cast_fp16)[name = tensor("linear_8_cast_fp16")]; + tensor input_23_cast_fp16 = add(x = input_17_cast_fp16, y = linear_8_cast_fp16)[name = tensor("input_23_cast_fp16")]; + tensor x_5_axes_0 = const()[name = tensor("x_5_axes_0"), val = tensor([-1])]; + tensor norm2_1_weight_to_fp16 = const()[name = tensor("norm2_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(303602048)))]; + tensor norm2_1_bias_to_fp16 = const()[name = tensor("norm2_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(303604160)))]; + tensor var_1042_to_fp16 = const()[name = tensor("op_1042_to_fp16"), val = tensor(0x1.5p-17)]; + tensor x_5_cast_fp16 = layer_norm(axes = x_5_axes_0, beta = norm2_1_bias_to_fp16, epsilon = var_1042_to_fp16, gamma = norm2_1_weight_to_fp16, x = input_23_cast_fp16)[name = tensor("x_5_cast_fp16")]; + tensor linear_9_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = attn2_in_proj_weight_cast_fp16, x = x_5_cast_fp16)[name = tensor("linear_9_cast_fp16")]; + tensor var_1078 = const()[name = tensor("op_1078"), val = tensor([1, 1, 3, 16, 64])]; + tensor qkv_5_cast_fp16 = reshape(shape = var_1078, x = linear_9_cast_fp16)[name = tensor("qkv_5_cast_fp16")]; + tensor q_13_begin_0 = const()[name = tensor("q_13_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor q_13_end_0 = const()[name = tensor("q_13_end_0"), val = tensor([1, 1, 1, 16, 64])]; + tensor q_13_end_mask_0 = const()[name = tensor("q_13_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor q_13_squeeze_mask_0 = const()[name = tensor("q_13_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor q_13_cast_fp16 = slice_by_index(begin = q_13_begin_0, end = q_13_end_0, end_mask = q_13_end_mask_0, squeeze_mask = q_13_squeeze_mask_0, x = qkv_5_cast_fp16)[name = tensor("q_13_cast_fp16")]; + tensor k_9_begin_0 = const()[name = tensor("k_9_begin_0"), val = tensor([0, 0, 1, 0, 0])]; + tensor k_9_end_0 = const()[name = tensor("k_9_end_0"), val = tensor([1, 1, 2, 16, 64])]; + tensor k_9_end_mask_0 = const()[name = tensor("k_9_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor k_9_squeeze_mask_0 = const()[name = tensor("k_9_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor k_9_cast_fp16 = slice_by_index(begin = k_9_begin_0, end = k_9_end_0, end_mask = k_9_end_mask_0, squeeze_mask = k_9_squeeze_mask_0, x = qkv_5_cast_fp16)[name = tensor("k_9_cast_fp16")]; + tensor v_5_begin_0 = const()[name = tensor("v_5_begin_0"), val = tensor([0, 0, 2, 0, 0])]; + tensor v_5_end_0 = const()[name = tensor("v_5_end_0"), val = tensor([1, 1, 3, 16, 64])]; + tensor v_5_end_mask_0 = const()[name = tensor("v_5_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor v_5_squeeze_mask_0 = const()[name = tensor("v_5_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor v_5_cast_fp16 = slice_by_index(begin = v_5_begin_0, end = v_5_end_0, end_mask = v_5_end_mask_0, squeeze_mask = v_5_squeeze_mask_0, x = qkv_5_cast_fp16)[name = tensor("v_5_cast_fp16")]; + tensor var_1182 = const()[name = tensor("op_1182"), val = tensor([1, 1, 1, 1])]; + tensor position2_to_fp16_dtype_0 = const()[name = tensor("position2_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor position2_to_fp16 = cast(dtype = position2_to_fp16_dtype_0, x = position2)[name = tensor("cast_543")]; + tensor ts_17_cast_fp16 = reshape(shape = var_1182, x = position2_to_fp16)[name = tensor("ts_17_cast_fp16")]; + tensor var_1186 = const()[name = tensor("op_1186"), val = tensor([1, 1, 16, 32, 2])]; + tensor q_complex_5_cast_fp16 = reshape(shape = var_1186, x = q_13_cast_fp16)[name = tensor("q_complex_5_cast_fp16")]; + tensor var_1190 = const()[name = tensor("op_1190"), val = tensor([1, 1, 16, 32, 2])]; + tensor k_complex_5_cast_fp16 = reshape(shape = var_1190, x = k_9_cast_fp16)[name = tensor("k_complex_5_cast_fp16")]; + tensor var_1194_begin_0 = const()[name = tensor("op_1194_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_1194_end_0 = const()[name = tensor("op_1194_end_0"), val = tensor([1, 1, 16, 32, 1])]; + tensor var_1194_end_mask_0 = const()[name = tensor("op_1194_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_1194_squeeze_mask_0 = const()[name = tensor("op_1194_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_1194_cast_fp16 = slice_by_index(begin = var_1194_begin_0, end = var_1194_end_0, end_mask = var_1194_end_mask_0, squeeze_mask = var_1194_squeeze_mask_0, x = q_complex_5_cast_fp16)[name = tensor("op_1194_cast_fp16")]; + tensor var_1202_begin_0 = const()[name = tensor("op_1202_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_1202_end_0 = const()[name = tensor("op_1202_end_0"), val = tensor([1, 1, 16, 32, 2])]; + tensor var_1202_end_mask_0 = const()[name = tensor("op_1202_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_1202_squeeze_mask_0 = const()[name = tensor("op_1202_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_1202_cast_fp16 = slice_by_index(begin = var_1202_begin_0, end = var_1202_end_0, end_mask = var_1202_end_mask_0, squeeze_mask = var_1202_squeeze_mask_0, x = q_complex_5_cast_fp16)[name = tensor("op_1202_cast_fp16")]; + tensor var_1210_begin_0 = const()[name = tensor("op_1210_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_1210_end_0 = const()[name = tensor("op_1210_end_0"), val = tensor([1, 1, 16, 32, 1])]; + tensor var_1210_end_mask_0 = const()[name = tensor("op_1210_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_1210_squeeze_mask_0 = const()[name = tensor("op_1210_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_1210_cast_fp16 = slice_by_index(begin = var_1210_begin_0, end = var_1210_end_0, end_mask = var_1210_end_mask_0, squeeze_mask = var_1210_squeeze_mask_0, x = k_complex_5_cast_fp16)[name = tensor("op_1210_cast_fp16")]; + tensor var_1218_begin_0 = const()[name = tensor("op_1218_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_1218_end_0 = const()[name = tensor("op_1218_end_0"), val = tensor([1, 1, 16, 32, 2])]; + tensor var_1218_end_mask_0 = const()[name = tensor("op_1218_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_1218_squeeze_mask_0 = const()[name = tensor("op_1218_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_1218_cast_fp16 = slice_by_index(begin = var_1218_begin_0, end = var_1218_end_0, end_mask = var_1218_end_mask_0, squeeze_mask = var_1218_squeeze_mask_0, x = k_complex_5_cast_fp16)[name = tensor("op_1218_cast_fp16")]; + tensor freqs_5_to_fp16 = const()[name = tensor("freqs_5_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(303606272)))]; + tensor var_1224_cast_fp16 = mul(x = freqs_5_to_fp16, y = ts_17_cast_fp16)[name = tensor("op_1224_cast_fp16")]; + tensor rotr_5_cast_fp16 = cos(x = var_1224_cast_fp16)[name = tensor("rotr_5_cast_fp16")]; + tensor roti_5_cast_fp16 = sin(x = var_1224_cast_fp16)[name = tensor("roti_5_cast_fp16")]; + tensor var_1228_cast_fp16 = mul(x = var_1194_cast_fp16, y = rotr_5_cast_fp16)[name = tensor("op_1228_cast_fp16")]; + tensor var_1229_cast_fp16 = mul(x = var_1202_cast_fp16, y = roti_5_cast_fp16)[name = tensor("op_1229_cast_fp16")]; + tensor qor_9_cast_fp16 = sub(x = var_1228_cast_fp16, y = var_1229_cast_fp16)[name = tensor("qor_9_cast_fp16")]; + tensor var_1232_cast_fp16 = mul(x = var_1194_cast_fp16, y = roti_5_cast_fp16)[name = tensor("op_1232_cast_fp16")]; + tensor var_1233_cast_fp16 = mul(x = var_1202_cast_fp16, y = rotr_5_cast_fp16)[name = tensor("op_1233_cast_fp16")]; + tensor qoi_9_cast_fp16 = add(x = var_1232_cast_fp16, y = var_1233_cast_fp16)[name = tensor("qoi_9_cast_fp16")]; + tensor var_1236_cast_fp16 = mul(x = var_1210_cast_fp16, y = rotr_5_cast_fp16)[name = tensor("op_1236_cast_fp16")]; + tensor var_1237_cast_fp16 = mul(x = var_1218_cast_fp16, y = roti_5_cast_fp16)[name = tensor("op_1237_cast_fp16")]; + tensor kor_9_cast_fp16 = sub(x = var_1236_cast_fp16, y = var_1237_cast_fp16)[name = tensor("kor_9_cast_fp16")]; + tensor var_1240_cast_fp16 = mul(x = var_1210_cast_fp16, y = roti_5_cast_fp16)[name = tensor("op_1240_cast_fp16")]; + tensor var_1241_cast_fp16 = mul(x = var_1218_cast_fp16, y = rotr_5_cast_fp16)[name = tensor("op_1241_cast_fp16")]; + tensor koi_9_cast_fp16 = add(x = var_1240_cast_fp16, y = var_1241_cast_fp16)[name = tensor("koi_9_cast_fp16")]; + tensor qo_5_axis_0 = const()[name = tensor("qo_5_axis_0"), val = tensor(-1)]; + tensor qo_5_cast_fp16 = stack(axis = qo_5_axis_0, values = (qor_9_cast_fp16, qoi_9_cast_fp16))[name = tensor("qo_5_cast_fp16")]; + tensor ko_5_axis_0 = const()[name = tensor("ko_5_axis_0"), val = tensor(-1)]; + tensor ko_5_cast_fp16 = stack(axis = ko_5_axis_0, values = (kor_9_cast_fp16, koi_9_cast_fp16))[name = tensor("ko_5_cast_fp16")]; + tensor var_1270 = const()[name = tensor("op_1270"), val = tensor([1, 1, 16, 64])]; + tensor q_15_cast_fp16 = reshape(shape = var_1270, x = qo_5_cast_fp16)[name = tensor("q_15_cast_fp16")]; + tensor var_1272 = const()[name = tensor("op_1272"), val = tensor([1, 1, 16, 64])]; + tensor k_11_cast_fp16 = reshape(shape = var_1272, x = ko_5_cast_fp16)[name = tensor("k_11_cast_fp16")]; + tensor _inversed_1294_y_0_to_fp16 = const()[name = tensor("_inversed_1294_y_0_to_fp16"), val = tensor(0x1p-9)]; + tensor _inversed_1294_cast_fp16 = mul(x = ts_17_cast_fp16, y = _inversed_1294_y_0_to_fp16)[name = tensor("_inversed_1294_cast_fp16")]; + tensor var_1295_cast_fp16 = floor(x = _inversed_1294_cast_fp16)[name = tensor("op_1295_cast_fp16")]; + tensor var_1296_to_fp16 = const()[name = tensor("op_1296_to_fp16"), val = tensor(0x1p+9)]; + tensor var_1297_cast_fp16 = mul(x = var_1295_cast_fp16, y = var_1296_to_fp16)[name = tensor("op_1297_cast_fp16")]; + tensor write_indices_float_11_cast_fp16 = sub(x = ts_17_cast_fp16, y = var_1297_cast_fp16)[name = tensor("write_indices_float_11_cast_fp16")]; + tensor var_1304_dtype_0 = const()[name = tensor("op_1304_dtype_0"), val = tensor("int32")]; + tensor write_indices_5_reps_0 = const()[name = tensor("write_indices_5_reps_0"), val = tensor([1, 1, 16, 64])]; + tensor write_indices_float_11_cast_fp16_to_int32 = cast(dtype = var_1304_dtype_0, x = write_indices_float_11_cast_fp16)[name = tensor("cast_542")]; + tensor write_indices_5 = tile(reps = write_indices_5_reps_0, x = write_indices_float_11_cast_fp16_to_int32)[name = tensor("write_indices_5")]; + tensor var_1312_begin_0 = const()[name = tensor("op_1312_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_1312_end_0 = const()[name = tensor("op_1312_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor var_1312_end_mask_0 = const()[name = tensor("op_1312_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_1312_squeeze_mask_0 = const()[name = tensor("op_1312_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor cache2_to_fp16_dtype_0 = const()[name = tensor("cache2_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor cache2_to_fp16 = cast(dtype = cache2_to_fp16_dtype_0, x = cache2)[name = tensor("cast_541")]; + tensor var_1312_cast_fp16 = slice_by_index(begin = var_1312_begin_0, end = var_1312_end_0, end_mask = var_1312_end_mask_0, squeeze_mask = var_1312_squeeze_mask_0, x = cache2_to_fp16)[name = tensor("op_1312_cast_fp16")]; + tensor var_1314_axis_0 = const()[name = tensor("op_1314_axis_0"), val = tensor(1)]; + tensor var_1314_mode_0 = const()[name = tensor("op_1314_mode_0"), val = tensor("update")]; + tensor var_1314_validate_indices_0 = const()[name = tensor("op_1314_validate_indices_0"), val = tensor(false)]; + tensor var_1314_cast_fp16 = scatter_along_axis(axis = var_1314_axis_0, data = var_1312_cast_fp16, indices = write_indices_5, mode = var_1314_mode_0, updates = k_11_cast_fp16, validate_indices = var_1314_validate_indices_0)[name = tensor("op_1314_cast_fp16")]; + tensor concat_16 = const()[name = tensor("concat_16"), val = tensor([0, 0, 0, 0, 0])]; + tensor concat_17 = const()[name = tensor("concat_17"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_5_internal_tensor_assign_1_stride_0 = const()[name = tensor("new_cache_5_internal_tensor_assign_1_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_5_internal_tensor_assign_1_begin_mask_0 = const()[name = tensor("new_cache_5_internal_tensor_assign_1_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_5_internal_tensor_assign_1_end_mask_0 = const()[name = tensor("new_cache_5_internal_tensor_assign_1_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_5_internal_tensor_assign_1_squeeze_mask_0 = const()[name = tensor("new_cache_5_internal_tensor_assign_1_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_52 = const()[name = tensor("shape_52"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_4 = const()[name = tensor("reduce_prod_4"), val = tensor(1048576)]; + tensor range_1d_4_start_0 = const()[name = tensor("range_1d_4_start_0"), val = tensor(0)]; + tensor range_1d_4_step_0 = const()[name = tensor("range_1d_4_step_0"), val = tensor(1)]; + tensor range_1d_4 = range_1d(end = reduce_prod_4, start = range_1d_4_start_0, step = range_1d_4_step_0)[name = tensor("range_1d_4")]; + tensor reshape_20 = reshape(shape = shape_52, x = range_1d_4)[name = tensor("reshape_20")]; + tensor slice_by_index_4 = slice_by_index(begin = concat_16, begin_mask = new_cache_5_internal_tensor_assign_1_begin_mask_0, end = concat_17, end_mask = new_cache_5_internal_tensor_assign_1_end_mask_0, squeeze_mask = new_cache_5_internal_tensor_assign_1_squeeze_mask_0, stride = new_cache_5_internal_tensor_assign_1_stride_0, x = reshape_20)[name = tensor("slice_by_index_4")]; + tensor reshape_21_shape_0 = const()[name = tensor("reshape_21_shape_0"), val = tensor([-1])]; + tensor reshape_21 = reshape(shape = reshape_21_shape_0, x = slice_by_index_4)[name = tensor("reshape_21")]; + tensor reshape_22_shape_0 = const()[name = tensor("reshape_22_shape_0"), val = tensor([-1])]; + tensor reshape_22_cast_fp16 = reshape(shape = reshape_22_shape_0, x = var_1314_cast_fp16)[name = tensor("reshape_22_cast_fp16")]; + tensor reshape_23_shape_0 = const()[name = tensor("reshape_23_shape_0"), val = tensor([-1])]; + tensor reshape_23_cast_fp16 = reshape(shape = reshape_23_shape_0, x = cache2_to_fp16)[name = tensor("reshape_23_cast_fp16")]; + tensor scatter_4_mode_0 = const()[name = tensor("scatter_4_mode_0"), val = tensor("update")]; + tensor scatter_4_axis_0 = const()[name = tensor("scatter_4_axis_0"), val = tensor(0)]; + tensor scatter_4_validate_indices_0 = const()[name = tensor("scatter_4_validate_indices_0"), val = tensor(false)]; + tensor scatter_4_cast_fp16 = scatter(axis = scatter_4_axis_0, data = reshape_23_cast_fp16, indices = reshape_21, mode = scatter_4_mode_0, updates = reshape_22_cast_fp16, validate_indices = scatter_4_validate_indices_0)[name = tensor("scatter_4_cast_fp16")]; + tensor reshape_24_cast_fp16 = reshape(shape = shape_52, x = scatter_4_cast_fp16)[name = tensor("reshape_24_cast_fp16")]; + tensor var_1322_begin_0 = const()[name = tensor("op_1322_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor var_1322_end_0 = const()[name = tensor("op_1322_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor var_1322_end_mask_0 = const()[name = tensor("op_1322_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_1322_squeeze_mask_0 = const()[name = tensor("op_1322_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor var_1322_cast_fp16 = slice_by_index(begin = var_1322_begin_0, end = var_1322_end_0, end_mask = var_1322_end_mask_0, squeeze_mask = var_1322_squeeze_mask_0, x = reshape_24_cast_fp16)[name = tensor("op_1322_cast_fp16")]; + tensor var_1324_axis_0 = const()[name = tensor("op_1324_axis_0"), val = tensor(1)]; + tensor var_1324_mode_0 = const()[name = tensor("op_1324_mode_0"), val = tensor("update")]; + tensor var_1324_validate_indices_0 = const()[name = tensor("op_1324_validate_indices_0"), val = tensor(false)]; + tensor var_1324_cast_fp16 = scatter_along_axis(axis = var_1324_axis_0, data = var_1322_cast_fp16, indices = write_indices_5, mode = var_1324_mode_0, updates = v_5_cast_fp16, validate_indices = var_1324_validate_indices_0)[name = tensor("op_1324_cast_fp16")]; + tensor concat_18 = const()[name = tensor("concat_18"), val = tensor([1, 0, 0, 0, 0])]; + tensor concat_19 = const()[name = tensor("concat_19"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_5_internal_tensor_assign_2_stride_0 = const()[name = tensor("new_cache_5_internal_tensor_assign_2_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_5_internal_tensor_assign_2_begin_mask_0 = const()[name = tensor("new_cache_5_internal_tensor_assign_2_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_5_internal_tensor_assign_2_end_mask_0 = const()[name = tensor("new_cache_5_internal_tensor_assign_2_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_5_internal_tensor_assign_2_squeeze_mask_0 = const()[name = tensor("new_cache_5_internal_tensor_assign_2_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_53 = const()[name = tensor("shape_53"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_5 = const()[name = tensor("reduce_prod_5"), val = tensor(1048576)]; + tensor range_1d_5_start_0 = const()[name = tensor("range_1d_5_start_0"), val = tensor(0)]; + tensor range_1d_5_step_0 = const()[name = tensor("range_1d_5_step_0"), val = tensor(1)]; + tensor range_1d_5 = range_1d(end = reduce_prod_5, start = range_1d_5_start_0, step = range_1d_5_step_0)[name = tensor("range_1d_5")]; + tensor reshape_25 = reshape(shape = shape_53, x = range_1d_5)[name = tensor("reshape_25")]; + tensor slice_by_index_5 = slice_by_index(begin = concat_18, begin_mask = new_cache_5_internal_tensor_assign_2_begin_mask_0, end = concat_19, end_mask = new_cache_5_internal_tensor_assign_2_end_mask_0, squeeze_mask = new_cache_5_internal_tensor_assign_2_squeeze_mask_0, stride = new_cache_5_internal_tensor_assign_2_stride_0, x = reshape_25)[name = tensor("slice_by_index_5")]; + tensor reshape_26_shape_0 = const()[name = tensor("reshape_26_shape_0"), val = tensor([-1])]; + tensor reshape_26 = reshape(shape = reshape_26_shape_0, x = slice_by_index_5)[name = tensor("reshape_26")]; + tensor reshape_27_shape_0 = const()[name = tensor("reshape_27_shape_0"), val = tensor([-1])]; + tensor reshape_27_cast_fp16 = reshape(shape = reshape_27_shape_0, x = var_1324_cast_fp16)[name = tensor("reshape_27_cast_fp16")]; + tensor reshape_28_shape_0 = const()[name = tensor("reshape_28_shape_0"), val = tensor([-1])]; + tensor reshape_28_cast_fp16 = reshape(shape = reshape_28_shape_0, x = reshape_24_cast_fp16)[name = tensor("reshape_28_cast_fp16")]; + tensor scatter_5_mode_0 = const()[name = tensor("scatter_5_mode_0"), val = tensor("update")]; + tensor scatter_5_axis_0 = const()[name = tensor("scatter_5_axis_0"), val = tensor(0)]; + tensor scatter_5_validate_indices_0 = const()[name = tensor("scatter_5_validate_indices_0"), val = tensor(false)]; + tensor scatter_5_cast_fp16 = scatter(axis = scatter_5_axis_0, data = reshape_28_cast_fp16, indices = reshape_26, mode = scatter_5_mode_0, updates = reshape_27_cast_fp16, validate_indices = scatter_5_validate_indices_0)[name = tensor("scatter_5_cast_fp16")]; + tensor reshape_29_cast_fp16 = reshape(shape = shape_53, x = scatter_5_cast_fp16)[name = tensor("reshape_29_cast_fp16")]; + tensor reshape_29_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("reshape_29_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor keys_13_begin_0 = const()[name = tensor("keys_13_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor keys_13_end_0 = const()[name = tensor("keys_13_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor keys_13_end_mask_0 = const()[name = tensor("keys_13_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor keys_13_squeeze_mask_0 = const()[name = tensor("keys_13_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor keys_13_cast_fp16 = slice_by_index(begin = keys_13_begin_0, end = keys_13_end_0, end_mask = keys_13_end_mask_0, squeeze_mask = keys_13_squeeze_mask_0, x = reshape_29_cast_fp16)[name = tensor("keys_13_cast_fp16")]; + tensor values_13_begin_0 = const()[name = tensor("values_13_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor values_13_end_0 = const()[name = tensor("values_13_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor values_13_end_mask_0 = const()[name = tensor("values_13_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor values_13_squeeze_mask_0 = const()[name = tensor("values_13_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor values_13_cast_fp16 = slice_by_index(begin = values_13_begin_0, end = values_13_end_0, end_mask = values_13_end_mask_0, squeeze_mask = values_13_squeeze_mask_0, x = reshape_29_cast_fp16)[name = tensor("values_13_cast_fp16")]; + tensor var_1336_cast_fp16 = not_equal(x = keys_13_cast_fp16, y = keys_13_cast_fp16)[name = tensor("op_1336_cast_fp16")]; + tensor keys_15_cast_fp16 = select(a = var_504_to_fp16, b = keys_13_cast_fp16, cond = var_1336_cast_fp16)[name = tensor("keys_15_cast_fp16")]; + tensor var_1344_cast_fp16 = not_equal(x = values_13_cast_fp16, y = values_13_cast_fp16)[name = tensor("op_1344_cast_fp16")]; + tensor values_15_cast_fp16 = select(a = var_504_to_fp16, b = values_13_cast_fp16, cond = var_1344_cast_fp16)[name = tensor("values_15_cast_fp16")]; + tensor var_1368 = const()[name = tensor("op_1368"), val = tensor([0, 2, 1, 3])]; + tensor var_1381 = const()[name = tensor("op_1381"), val = tensor([1, 1, 1])]; + tensor var_1382_cast_fp16 = reshape(shape = var_1381, x = position2_to_fp16)[name = tensor("op_1382_cast_fp16")]; + tensor var_1399_to_fp16 = const()[name = tensor("op_1399_to_fp16"), val = tensor(0x1p+0)]; + tensor valid_len_5_cast_fp16 = add(x = var_1382_cast_fp16, y = var_1399_to_fp16)[name = tensor("valid_len_5_cast_fp16")]; + tensor valid_mask_5_cast_fp16 = less(x = k_positions_1_promoted_to_fp16, y = valid_len_5_cast_fp16)[name = tensor("valid_mask_5_cast_fp16")]; + tensor causal_mask_5_cast_fp16 = less_equal(x = k_positions_1_promoted_to_fp16, y = var_1382_cast_fp16)[name = tensor("causal_mask_5_cast_fp16")]; + tensor attn_mask_9 = logical_and(x = valid_mask_5_cast_fp16, y = causal_mask_5_cast_fp16)[name = tensor("attn_mask_9")]; + tensor attn_mask_11_axes_0 = const()[name = tensor("attn_mask_11_axes_0"), val = tensor([1])]; + tensor attn_mask_11 = expand_dims(axes = attn_mask_11_axes_0, x = attn_mask_9)[name = tensor("attn_mask_11")]; + tensor var_1417_transpose_x_0 = const()[name = tensor("op_1417_transpose_x_0"), val = tensor(false)]; + tensor var_1417_transpose_y_0 = const()[name = tensor("op_1417_transpose_y_0"), val = tensor(false)]; + tensor transpose_76_perm_0 = const()[name = tensor("transpose_76_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_77_perm_0 = const()[name = tensor("transpose_77_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_77 = transpose(perm = transpose_77_perm_0, x = keys_15_cast_fp16)[name = tensor("transpose_205")]; + tensor transpose_76 = transpose(perm = transpose_76_perm_0, x = q_15_cast_fp16)[name = tensor("transpose_206")]; + tensor var_1417_cast_fp16 = matmul(transpose_x = var_1417_transpose_x_0, transpose_y = var_1417_transpose_y_0, x = transpose_76, y = transpose_77)[name = tensor("op_1417_cast_fp16")]; + tensor var_1411_to_fp16 = const()[name = tensor("op_1411_to_fp16"), val = tensor([0x1p-3])]; + tensor attn_weights_13_cast_fp16 = mul(x = var_1417_cast_fp16, y = var_1411_to_fp16)[name = tensor("attn_weights_13_cast_fp16")]; + tensor var_1419 = logical_not(x = attn_mask_11)[name = tensor("op_1419")]; + tensor var_1420_to_fp16 = const()[name = tensor("op_1420_to_fp16"), val = tensor(-inf)]; + tensor attn_weights_15_cast_fp16 = select(a = var_1420_to_fp16, b = attn_weights_13_cast_fp16, cond = var_1419)[name = tensor("attn_weights_15_cast_fp16")]; + tensor var_1422 = const()[name = tensor("op_1422"), val = tensor(-1)]; + tensor attn_weights_17_cast_fp16 = softmax(axis = var_1422, x = attn_weights_15_cast_fp16)[name = tensor("attn_weights_17_cast_fp16")]; + tensor attn_output_5_transpose_x_0 = const()[name = tensor("attn_output_5_transpose_x_0"), val = tensor(false)]; + tensor attn_output_5_transpose_y_0 = const()[name = tensor("attn_output_5_transpose_y_0"), val = tensor(false)]; + tensor values_17_cast_fp16 = transpose(perm = var_1368, x = values_15_cast_fp16)[name = tensor("transpose_207")]; + tensor attn_output_5_cast_fp16 = matmul(transpose_x = attn_output_5_transpose_x_0, transpose_y = attn_output_5_transpose_y_0, x = attn_weights_17_cast_fp16, y = values_17_cast_fp16)[name = tensor("attn_output_5_cast_fp16")]; + tensor var_1430 = const()[name = tensor("op_1430"), val = tensor([0, 2, 1, 3])]; + tensor var_1433 = const()[name = tensor("op_1433"), val = tensor([1, 1, 1024])]; + tensor var_1431_cast_fp16 = transpose(perm = var_1430, x = attn_output_5_cast_fp16)[name = tensor("transpose_204")]; + tensor input_25_cast_fp16 = reshape(shape = var_1433, x = var_1431_cast_fp16)[name = tensor("input_25_cast_fp16")]; + tensor linear_10_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = attn2_out_proj_weight_cast_fp16, x = input_25_cast_fp16)[name = tensor("linear_10_cast_fp16")]; + tensor var_1439_to_fp16 = const()[name = tensor("op_1439_to_fp16"), val = tensor(0x1p+0)]; + tensor var_1440_cast_fp16 = add(x = position2_to_fp16, y = var_1439_to_fp16)[name = tensor("op_1440_cast_fp16")]; + tensor var_1440_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_1440_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor input_27_cast_fp16 = add(x = input_23_cast_fp16, y = linear_10_cast_fp16)[name = tensor("input_27_cast_fp16")]; + tensor input_29_axes_0 = const()[name = tensor("input_29_axes_0"), val = tensor([-1])]; + tensor norm2_2_weight_to_fp16 = const()[name = tensor("norm2_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(303606400)))]; + tensor norm2_2_bias_to_fp16 = const()[name = tensor("norm2_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(303608512)))]; + tensor var_1444_to_fp16 = const()[name = tensor("op_1444_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_29_cast_fp16 = layer_norm(axes = input_29_axes_0, beta = norm2_2_bias_to_fp16, epsilon = var_1444_to_fp16, gamma = norm2_2_weight_to_fp16, x = input_27_cast_fp16)[name = tensor("input_29_cast_fp16")]; + tensor linear_11_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = linear2_1_weight_cast_fp16, x = input_29_cast_fp16)[name = tensor("linear_11_cast_fp16")]; + tensor input_31_mode_0 = const()[name = tensor("input_31_mode_0"), val = tensor("EXACT")]; + tensor input_31_cast_fp16 = gelu(mode = input_31_mode_0, x = linear_11_cast_fp16)[name = tensor("input_31_cast_fp16")]; + tensor linear_12_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = linear2_2_weight_cast_fp16, x = input_31_cast_fp16)[name = tensor("linear_12_cast_fp16")]; + tensor input_33_cast_fp16 = add(x = input_27_cast_fp16, y = linear_12_cast_fp16)[name = tensor("input_33_cast_fp16")]; + tensor x_7_axes_0 = const()[name = tensor("x_7_axes_0"), val = tensor([-1])]; + tensor norm3_1_weight_to_fp16 = const()[name = tensor("norm3_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(303610624)))]; + tensor norm3_1_bias_to_fp16 = const()[name = tensor("norm3_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(303612736)))]; + tensor var_1461_to_fp16 = const()[name = tensor("op_1461_to_fp16"), val = tensor(0x1.5p-17)]; + tensor x_7_cast_fp16 = layer_norm(axes = x_7_axes_0, beta = norm3_1_bias_to_fp16, epsilon = var_1461_to_fp16, gamma = norm3_1_weight_to_fp16, x = input_33_cast_fp16)[name = tensor("x_7_cast_fp16")]; + tensor linear_13_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = attn3_in_proj_weight_cast_fp16, x = x_7_cast_fp16)[name = tensor("linear_13_cast_fp16")]; + tensor var_1497 = const()[name = tensor("op_1497"), val = tensor([1, 1, 3, 16, 64])]; + tensor qkv_7_cast_fp16 = reshape(shape = var_1497, x = linear_13_cast_fp16)[name = tensor("qkv_7_cast_fp16")]; + tensor q_19_begin_0 = const()[name = tensor("q_19_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor q_19_end_0 = const()[name = tensor("q_19_end_0"), val = tensor([1, 1, 1, 16, 64])]; + tensor q_19_end_mask_0 = const()[name = tensor("q_19_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor q_19_squeeze_mask_0 = const()[name = tensor("q_19_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor q_19_cast_fp16 = slice_by_index(begin = q_19_begin_0, end = q_19_end_0, end_mask = q_19_end_mask_0, squeeze_mask = q_19_squeeze_mask_0, x = qkv_7_cast_fp16)[name = tensor("q_19_cast_fp16")]; + tensor k_13_begin_0 = const()[name = tensor("k_13_begin_0"), val = tensor([0, 0, 1, 0, 0])]; + tensor k_13_end_0 = const()[name = tensor("k_13_end_0"), val = tensor([1, 1, 2, 16, 64])]; + tensor k_13_end_mask_0 = const()[name = tensor("k_13_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor k_13_squeeze_mask_0 = const()[name = tensor("k_13_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor k_13_cast_fp16 = slice_by_index(begin = k_13_begin_0, end = k_13_end_0, end_mask = k_13_end_mask_0, squeeze_mask = k_13_squeeze_mask_0, x = qkv_7_cast_fp16)[name = tensor("k_13_cast_fp16")]; + tensor v_7_begin_0 = const()[name = tensor("v_7_begin_0"), val = tensor([0, 0, 2, 0, 0])]; + tensor v_7_end_0 = const()[name = tensor("v_7_end_0"), val = tensor([1, 1, 3, 16, 64])]; + tensor v_7_end_mask_0 = const()[name = tensor("v_7_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor v_7_squeeze_mask_0 = const()[name = tensor("v_7_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor v_7_cast_fp16 = slice_by_index(begin = v_7_begin_0, end = v_7_end_0, end_mask = v_7_end_mask_0, squeeze_mask = v_7_squeeze_mask_0, x = qkv_7_cast_fp16)[name = tensor("v_7_cast_fp16")]; + tensor var_1601 = const()[name = tensor("op_1601"), val = tensor([1, 1, 1, 1])]; + tensor position3_to_fp16_dtype_0 = const()[name = tensor("position3_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor position3_to_fp16 = cast(dtype = position3_to_fp16_dtype_0, x = position3)[name = tensor("cast_538")]; + tensor ts_23_cast_fp16 = reshape(shape = var_1601, x = position3_to_fp16)[name = tensor("ts_23_cast_fp16")]; + tensor var_1605 = const()[name = tensor("op_1605"), val = tensor([1, 1, 16, 32, 2])]; + tensor q_complex_7_cast_fp16 = reshape(shape = var_1605, x = q_19_cast_fp16)[name = tensor("q_complex_7_cast_fp16")]; + tensor var_1609 = const()[name = tensor("op_1609"), val = tensor([1, 1, 16, 32, 2])]; + tensor k_complex_7_cast_fp16 = reshape(shape = var_1609, x = k_13_cast_fp16)[name = tensor("k_complex_7_cast_fp16")]; + tensor var_1613_begin_0 = const()[name = tensor("op_1613_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_1613_end_0 = const()[name = tensor("op_1613_end_0"), val = tensor([1, 1, 16, 32, 1])]; + tensor var_1613_end_mask_0 = const()[name = tensor("op_1613_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_1613_squeeze_mask_0 = const()[name = tensor("op_1613_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_1613_cast_fp16 = slice_by_index(begin = var_1613_begin_0, end = var_1613_end_0, end_mask = var_1613_end_mask_0, squeeze_mask = var_1613_squeeze_mask_0, x = q_complex_7_cast_fp16)[name = tensor("op_1613_cast_fp16")]; + tensor var_1621_begin_0 = const()[name = tensor("op_1621_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_1621_end_0 = const()[name = tensor("op_1621_end_0"), val = tensor([1, 1, 16, 32, 2])]; + tensor var_1621_end_mask_0 = const()[name = tensor("op_1621_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_1621_squeeze_mask_0 = const()[name = tensor("op_1621_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_1621_cast_fp16 = slice_by_index(begin = var_1621_begin_0, end = var_1621_end_0, end_mask = var_1621_end_mask_0, squeeze_mask = var_1621_squeeze_mask_0, x = q_complex_7_cast_fp16)[name = tensor("op_1621_cast_fp16")]; + tensor var_1629_begin_0 = const()[name = tensor("op_1629_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_1629_end_0 = const()[name = tensor("op_1629_end_0"), val = tensor([1, 1, 16, 32, 1])]; + tensor var_1629_end_mask_0 = const()[name = tensor("op_1629_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_1629_squeeze_mask_0 = const()[name = tensor("op_1629_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_1629_cast_fp16 = slice_by_index(begin = var_1629_begin_0, end = var_1629_end_0, end_mask = var_1629_end_mask_0, squeeze_mask = var_1629_squeeze_mask_0, x = k_complex_7_cast_fp16)[name = tensor("op_1629_cast_fp16")]; + tensor var_1637_begin_0 = const()[name = tensor("op_1637_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_1637_end_0 = const()[name = tensor("op_1637_end_0"), val = tensor([1, 1, 16, 32, 2])]; + tensor var_1637_end_mask_0 = const()[name = tensor("op_1637_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_1637_squeeze_mask_0 = const()[name = tensor("op_1637_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_1637_cast_fp16 = slice_by_index(begin = var_1637_begin_0, end = var_1637_end_0, end_mask = var_1637_end_mask_0, squeeze_mask = var_1637_squeeze_mask_0, x = k_complex_7_cast_fp16)[name = tensor("op_1637_cast_fp16")]; + tensor freqs_7_to_fp16 = const()[name = tensor("freqs_7_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(303614848)))]; + tensor var_1643_cast_fp16 = mul(x = freqs_7_to_fp16, y = ts_23_cast_fp16)[name = tensor("op_1643_cast_fp16")]; + tensor rotr_7_cast_fp16 = cos(x = var_1643_cast_fp16)[name = tensor("rotr_7_cast_fp16")]; + tensor roti_7_cast_fp16 = sin(x = var_1643_cast_fp16)[name = tensor("roti_7_cast_fp16")]; + tensor var_1647_cast_fp16 = mul(x = var_1613_cast_fp16, y = rotr_7_cast_fp16)[name = tensor("op_1647_cast_fp16")]; + tensor var_1648_cast_fp16 = mul(x = var_1621_cast_fp16, y = roti_7_cast_fp16)[name = tensor("op_1648_cast_fp16")]; + tensor qor_13_cast_fp16 = sub(x = var_1647_cast_fp16, y = var_1648_cast_fp16)[name = tensor("qor_13_cast_fp16")]; + tensor var_1651_cast_fp16 = mul(x = var_1613_cast_fp16, y = roti_7_cast_fp16)[name = tensor("op_1651_cast_fp16")]; + tensor var_1652_cast_fp16 = mul(x = var_1621_cast_fp16, y = rotr_7_cast_fp16)[name = tensor("op_1652_cast_fp16")]; + tensor qoi_13_cast_fp16 = add(x = var_1651_cast_fp16, y = var_1652_cast_fp16)[name = tensor("qoi_13_cast_fp16")]; + tensor var_1655_cast_fp16 = mul(x = var_1629_cast_fp16, y = rotr_7_cast_fp16)[name = tensor("op_1655_cast_fp16")]; + tensor var_1656_cast_fp16 = mul(x = var_1637_cast_fp16, y = roti_7_cast_fp16)[name = tensor("op_1656_cast_fp16")]; + tensor kor_13_cast_fp16 = sub(x = var_1655_cast_fp16, y = var_1656_cast_fp16)[name = tensor("kor_13_cast_fp16")]; + tensor var_1659_cast_fp16 = mul(x = var_1629_cast_fp16, y = roti_7_cast_fp16)[name = tensor("op_1659_cast_fp16")]; + tensor var_1660_cast_fp16 = mul(x = var_1637_cast_fp16, y = rotr_7_cast_fp16)[name = tensor("op_1660_cast_fp16")]; + tensor koi_13_cast_fp16 = add(x = var_1659_cast_fp16, y = var_1660_cast_fp16)[name = tensor("koi_13_cast_fp16")]; + tensor qo_7_axis_0 = const()[name = tensor("qo_7_axis_0"), val = tensor(-1)]; + tensor qo_7_cast_fp16 = stack(axis = qo_7_axis_0, values = (qor_13_cast_fp16, qoi_13_cast_fp16))[name = tensor("qo_7_cast_fp16")]; + tensor ko_7_axis_0 = const()[name = tensor("ko_7_axis_0"), val = tensor(-1)]; + tensor ko_7_cast_fp16 = stack(axis = ko_7_axis_0, values = (kor_13_cast_fp16, koi_13_cast_fp16))[name = tensor("ko_7_cast_fp16")]; + tensor var_1689 = const()[name = tensor("op_1689"), val = tensor([1, 1, 16, 64])]; + tensor q_21_cast_fp16 = reshape(shape = var_1689, x = qo_7_cast_fp16)[name = tensor("q_21_cast_fp16")]; + tensor var_1691 = const()[name = tensor("op_1691"), val = tensor([1, 1, 16, 64])]; + tensor k_15_cast_fp16 = reshape(shape = var_1691, x = ko_7_cast_fp16)[name = tensor("k_15_cast_fp16")]; + tensor _inversed_1713_y_0_to_fp16 = const()[name = tensor("_inversed_1713_y_0_to_fp16"), val = tensor(0x1p-9)]; + tensor _inversed_1713_cast_fp16 = mul(x = ts_23_cast_fp16, y = _inversed_1713_y_0_to_fp16)[name = tensor("_inversed_1713_cast_fp16")]; + tensor var_1714_cast_fp16 = floor(x = _inversed_1713_cast_fp16)[name = tensor("op_1714_cast_fp16")]; + tensor var_1715_to_fp16 = const()[name = tensor("op_1715_to_fp16"), val = tensor(0x1p+9)]; + tensor var_1716_cast_fp16 = mul(x = var_1714_cast_fp16, y = var_1715_to_fp16)[name = tensor("op_1716_cast_fp16")]; + tensor write_indices_float_15_cast_fp16 = sub(x = ts_23_cast_fp16, y = var_1716_cast_fp16)[name = tensor("write_indices_float_15_cast_fp16")]; + tensor var_1723_dtype_0 = const()[name = tensor("op_1723_dtype_0"), val = tensor("int32")]; + tensor write_indices_7_reps_0 = const()[name = tensor("write_indices_7_reps_0"), val = tensor([1, 1, 16, 64])]; + tensor write_indices_float_15_cast_fp16_to_int32 = cast(dtype = var_1723_dtype_0, x = write_indices_float_15_cast_fp16)[name = tensor("cast_537")]; + tensor write_indices_7 = tile(reps = write_indices_7_reps_0, x = write_indices_float_15_cast_fp16_to_int32)[name = tensor("write_indices_7")]; + tensor var_1731_begin_0 = const()[name = tensor("op_1731_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_1731_end_0 = const()[name = tensor("op_1731_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor var_1731_end_mask_0 = const()[name = tensor("op_1731_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_1731_squeeze_mask_0 = const()[name = tensor("op_1731_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor cache3_to_fp16_dtype_0 = const()[name = tensor("cache3_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor cache3_to_fp16 = cast(dtype = cache3_to_fp16_dtype_0, x = cache3)[name = tensor("cast_536")]; + tensor var_1731_cast_fp16 = slice_by_index(begin = var_1731_begin_0, end = var_1731_end_0, end_mask = var_1731_end_mask_0, squeeze_mask = var_1731_squeeze_mask_0, x = cache3_to_fp16)[name = tensor("op_1731_cast_fp16")]; + tensor var_1733_axis_0 = const()[name = tensor("op_1733_axis_0"), val = tensor(1)]; + tensor var_1733_mode_0 = const()[name = tensor("op_1733_mode_0"), val = tensor("update")]; + tensor var_1733_validate_indices_0 = const()[name = tensor("op_1733_validate_indices_0"), val = tensor(false)]; + tensor var_1733_cast_fp16 = scatter_along_axis(axis = var_1733_axis_0, data = var_1731_cast_fp16, indices = write_indices_7, mode = var_1733_mode_0, updates = k_15_cast_fp16, validate_indices = var_1733_validate_indices_0)[name = tensor("op_1733_cast_fp16")]; + tensor concat_23 = const()[name = tensor("concat_23"), val = tensor([0, 0, 0, 0, 0])]; + tensor concat_24 = const()[name = tensor("concat_24"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_7_internal_tensor_assign_1_stride_0 = const()[name = tensor("new_cache_7_internal_tensor_assign_1_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_7_internal_tensor_assign_1_begin_mask_0 = const()[name = tensor("new_cache_7_internal_tensor_assign_1_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_7_internal_tensor_assign_1_end_mask_0 = const()[name = tensor("new_cache_7_internal_tensor_assign_1_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_7_internal_tensor_assign_1_squeeze_mask_0 = const()[name = tensor("new_cache_7_internal_tensor_assign_1_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_54 = const()[name = tensor("shape_54"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_6 = const()[name = tensor("reduce_prod_6"), val = tensor(1048576)]; + tensor range_1d_6_start_0 = const()[name = tensor("range_1d_6_start_0"), val = tensor(0)]; + tensor range_1d_6_step_0 = const()[name = tensor("range_1d_6_step_0"), val = tensor(1)]; + tensor range_1d_6 = range_1d(end = reduce_prod_6, start = range_1d_6_start_0, step = range_1d_6_step_0)[name = tensor("range_1d_6")]; + tensor reshape_30 = reshape(shape = shape_54, x = range_1d_6)[name = tensor("reshape_30")]; + tensor slice_by_index_6 = slice_by_index(begin = concat_23, begin_mask = new_cache_7_internal_tensor_assign_1_begin_mask_0, end = concat_24, end_mask = new_cache_7_internal_tensor_assign_1_end_mask_0, squeeze_mask = new_cache_7_internal_tensor_assign_1_squeeze_mask_0, stride = new_cache_7_internal_tensor_assign_1_stride_0, x = reshape_30)[name = tensor("slice_by_index_6")]; + tensor reshape_31_shape_0 = const()[name = tensor("reshape_31_shape_0"), val = tensor([-1])]; + tensor reshape_31 = reshape(shape = reshape_31_shape_0, x = slice_by_index_6)[name = tensor("reshape_31")]; + tensor reshape_32_shape_0 = const()[name = tensor("reshape_32_shape_0"), val = tensor([-1])]; + tensor reshape_32_cast_fp16 = reshape(shape = reshape_32_shape_0, x = var_1733_cast_fp16)[name = tensor("reshape_32_cast_fp16")]; + tensor reshape_33_shape_0 = const()[name = tensor("reshape_33_shape_0"), val = tensor([-1])]; + tensor reshape_33_cast_fp16 = reshape(shape = reshape_33_shape_0, x = cache3_to_fp16)[name = tensor("reshape_33_cast_fp16")]; + tensor scatter_6_mode_0 = const()[name = tensor("scatter_6_mode_0"), val = tensor("update")]; + tensor scatter_6_axis_0 = const()[name = tensor("scatter_6_axis_0"), val = tensor(0)]; + tensor scatter_6_validate_indices_0 = const()[name = tensor("scatter_6_validate_indices_0"), val = tensor(false)]; + tensor scatter_6_cast_fp16 = scatter(axis = scatter_6_axis_0, data = reshape_33_cast_fp16, indices = reshape_31, mode = scatter_6_mode_0, updates = reshape_32_cast_fp16, validate_indices = scatter_6_validate_indices_0)[name = tensor("scatter_6_cast_fp16")]; + tensor reshape_34_cast_fp16 = reshape(shape = shape_54, x = scatter_6_cast_fp16)[name = tensor("reshape_34_cast_fp16")]; + tensor var_1741_begin_0 = const()[name = tensor("op_1741_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor var_1741_end_0 = const()[name = tensor("op_1741_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor var_1741_end_mask_0 = const()[name = tensor("op_1741_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_1741_squeeze_mask_0 = const()[name = tensor("op_1741_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor var_1741_cast_fp16 = slice_by_index(begin = var_1741_begin_0, end = var_1741_end_0, end_mask = var_1741_end_mask_0, squeeze_mask = var_1741_squeeze_mask_0, x = reshape_34_cast_fp16)[name = tensor("op_1741_cast_fp16")]; + tensor var_1743_axis_0 = const()[name = tensor("op_1743_axis_0"), val = tensor(1)]; + tensor var_1743_mode_0 = const()[name = tensor("op_1743_mode_0"), val = tensor("update")]; + tensor var_1743_validate_indices_0 = const()[name = tensor("op_1743_validate_indices_0"), val = tensor(false)]; + tensor var_1743_cast_fp16 = scatter_along_axis(axis = var_1743_axis_0, data = var_1741_cast_fp16, indices = write_indices_7, mode = var_1743_mode_0, updates = v_7_cast_fp16, validate_indices = var_1743_validate_indices_0)[name = tensor("op_1743_cast_fp16")]; + tensor concat_25 = const()[name = tensor("concat_25"), val = tensor([1, 0, 0, 0, 0])]; + tensor concat_26 = const()[name = tensor("concat_26"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_7_internal_tensor_assign_2_stride_0 = const()[name = tensor("new_cache_7_internal_tensor_assign_2_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_7_internal_tensor_assign_2_begin_mask_0 = const()[name = tensor("new_cache_7_internal_tensor_assign_2_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_7_internal_tensor_assign_2_end_mask_0 = const()[name = tensor("new_cache_7_internal_tensor_assign_2_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_7_internal_tensor_assign_2_squeeze_mask_0 = const()[name = tensor("new_cache_7_internal_tensor_assign_2_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_55 = const()[name = tensor("shape_55"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_7 = const()[name = tensor("reduce_prod_7"), val = tensor(1048576)]; + tensor range_1d_7_start_0 = const()[name = tensor("range_1d_7_start_0"), val = tensor(0)]; + tensor range_1d_7_step_0 = const()[name = tensor("range_1d_7_step_0"), val = tensor(1)]; + tensor range_1d_7 = range_1d(end = reduce_prod_7, start = range_1d_7_start_0, step = range_1d_7_step_0)[name = tensor("range_1d_7")]; + tensor reshape_35 = reshape(shape = shape_55, x = range_1d_7)[name = tensor("reshape_35")]; + tensor slice_by_index_7 = slice_by_index(begin = concat_25, begin_mask = new_cache_7_internal_tensor_assign_2_begin_mask_0, end = concat_26, end_mask = new_cache_7_internal_tensor_assign_2_end_mask_0, squeeze_mask = new_cache_7_internal_tensor_assign_2_squeeze_mask_0, stride = new_cache_7_internal_tensor_assign_2_stride_0, x = reshape_35)[name = tensor("slice_by_index_7")]; + tensor reshape_36_shape_0 = const()[name = tensor("reshape_36_shape_0"), val = tensor([-1])]; + tensor reshape_36 = reshape(shape = reshape_36_shape_0, x = slice_by_index_7)[name = tensor("reshape_36")]; + tensor reshape_37_shape_0 = const()[name = tensor("reshape_37_shape_0"), val = tensor([-1])]; + tensor reshape_37_cast_fp16 = reshape(shape = reshape_37_shape_0, x = var_1743_cast_fp16)[name = tensor("reshape_37_cast_fp16")]; + tensor reshape_38_shape_0 = const()[name = tensor("reshape_38_shape_0"), val = tensor([-1])]; + tensor reshape_38_cast_fp16 = reshape(shape = reshape_38_shape_0, x = reshape_34_cast_fp16)[name = tensor("reshape_38_cast_fp16")]; + tensor scatter_7_mode_0 = const()[name = tensor("scatter_7_mode_0"), val = tensor("update")]; + tensor scatter_7_axis_0 = const()[name = tensor("scatter_7_axis_0"), val = tensor(0)]; + tensor scatter_7_validate_indices_0 = const()[name = tensor("scatter_7_validate_indices_0"), val = tensor(false)]; + tensor scatter_7_cast_fp16 = scatter(axis = scatter_7_axis_0, data = reshape_38_cast_fp16, indices = reshape_36, mode = scatter_7_mode_0, updates = reshape_37_cast_fp16, validate_indices = scatter_7_validate_indices_0)[name = tensor("scatter_7_cast_fp16")]; + tensor reshape_39_cast_fp16 = reshape(shape = shape_55, x = scatter_7_cast_fp16)[name = tensor("reshape_39_cast_fp16")]; + tensor reshape_39_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("reshape_39_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor keys_19_begin_0 = const()[name = tensor("keys_19_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor keys_19_end_0 = const()[name = tensor("keys_19_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor keys_19_end_mask_0 = const()[name = tensor("keys_19_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor keys_19_squeeze_mask_0 = const()[name = tensor("keys_19_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor keys_19_cast_fp16 = slice_by_index(begin = keys_19_begin_0, end = keys_19_end_0, end_mask = keys_19_end_mask_0, squeeze_mask = keys_19_squeeze_mask_0, x = reshape_39_cast_fp16)[name = tensor("keys_19_cast_fp16")]; + tensor values_19_begin_0 = const()[name = tensor("values_19_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor values_19_end_0 = const()[name = tensor("values_19_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor values_19_end_mask_0 = const()[name = tensor("values_19_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor values_19_squeeze_mask_0 = const()[name = tensor("values_19_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor values_19_cast_fp16 = slice_by_index(begin = values_19_begin_0, end = values_19_end_0, end_mask = values_19_end_mask_0, squeeze_mask = values_19_squeeze_mask_0, x = reshape_39_cast_fp16)[name = tensor("values_19_cast_fp16")]; + tensor var_1755_cast_fp16 = not_equal(x = keys_19_cast_fp16, y = keys_19_cast_fp16)[name = tensor("op_1755_cast_fp16")]; + tensor keys_21_cast_fp16 = select(a = var_504_to_fp16, b = keys_19_cast_fp16, cond = var_1755_cast_fp16)[name = tensor("keys_21_cast_fp16")]; + tensor var_1763_cast_fp16 = not_equal(x = values_19_cast_fp16, y = values_19_cast_fp16)[name = tensor("op_1763_cast_fp16")]; + tensor values_21_cast_fp16 = select(a = var_504_to_fp16, b = values_19_cast_fp16, cond = var_1763_cast_fp16)[name = tensor("values_21_cast_fp16")]; + tensor var_1787 = const()[name = tensor("op_1787"), val = tensor([0, 2, 1, 3])]; + tensor var_1800 = const()[name = tensor("op_1800"), val = tensor([1, 1, 1])]; + tensor var_1801_cast_fp16 = reshape(shape = var_1800, x = position3_to_fp16)[name = tensor("op_1801_cast_fp16")]; + tensor var_1818_to_fp16 = const()[name = tensor("op_1818_to_fp16"), val = tensor(0x1p+0)]; + tensor valid_len_7_cast_fp16 = add(x = var_1801_cast_fp16, y = var_1818_to_fp16)[name = tensor("valid_len_7_cast_fp16")]; + tensor valid_mask_7_cast_fp16 = less(x = k_positions_1_promoted_to_fp16, y = valid_len_7_cast_fp16)[name = tensor("valid_mask_7_cast_fp16")]; + tensor causal_mask_7_cast_fp16 = less_equal(x = k_positions_1_promoted_to_fp16, y = var_1801_cast_fp16)[name = tensor("causal_mask_7_cast_fp16")]; + tensor attn_mask_13 = logical_and(x = valid_mask_7_cast_fp16, y = causal_mask_7_cast_fp16)[name = tensor("attn_mask_13")]; + tensor attn_mask_15_axes_0 = const()[name = tensor("attn_mask_15_axes_0"), val = tensor([1])]; + tensor attn_mask_15 = expand_dims(axes = attn_mask_15_axes_0, x = attn_mask_13)[name = tensor("attn_mask_15")]; + tensor var_1836_transpose_x_0 = const()[name = tensor("op_1836_transpose_x_0"), val = tensor(false)]; + tensor var_1836_transpose_y_0 = const()[name = tensor("op_1836_transpose_y_0"), val = tensor(false)]; + tensor transpose_78_perm_0 = const()[name = tensor("transpose_78_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_79_perm_0 = const()[name = tensor("transpose_79_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_79 = transpose(perm = transpose_79_perm_0, x = keys_21_cast_fp16)[name = tensor("transpose_201")]; + tensor transpose_78 = transpose(perm = transpose_78_perm_0, x = q_21_cast_fp16)[name = tensor("transpose_202")]; + tensor var_1836_cast_fp16 = matmul(transpose_x = var_1836_transpose_x_0, transpose_y = var_1836_transpose_y_0, x = transpose_78, y = transpose_79)[name = tensor("op_1836_cast_fp16")]; + tensor var_1830_to_fp16 = const()[name = tensor("op_1830_to_fp16"), val = tensor([0x1p-3])]; + tensor attn_weights_19_cast_fp16 = mul(x = var_1836_cast_fp16, y = var_1830_to_fp16)[name = tensor("attn_weights_19_cast_fp16")]; + tensor var_1838 = logical_not(x = attn_mask_15)[name = tensor("op_1838")]; + tensor var_1839_to_fp16 = const()[name = tensor("op_1839_to_fp16"), val = tensor(-inf)]; + tensor attn_weights_21_cast_fp16 = select(a = var_1839_to_fp16, b = attn_weights_19_cast_fp16, cond = var_1838)[name = tensor("attn_weights_21_cast_fp16")]; + tensor var_1841 = const()[name = tensor("op_1841"), val = tensor(-1)]; + tensor attn_weights_23_cast_fp16 = softmax(axis = var_1841, x = attn_weights_21_cast_fp16)[name = tensor("attn_weights_23_cast_fp16")]; + tensor attn_output_7_transpose_x_0 = const()[name = tensor("attn_output_7_transpose_x_0"), val = tensor(false)]; + tensor attn_output_7_transpose_y_0 = const()[name = tensor("attn_output_7_transpose_y_0"), val = tensor(false)]; + tensor values_23_cast_fp16 = transpose(perm = var_1787, x = values_21_cast_fp16)[name = tensor("transpose_203")]; + tensor attn_output_7_cast_fp16 = matmul(transpose_x = attn_output_7_transpose_x_0, transpose_y = attn_output_7_transpose_y_0, x = attn_weights_23_cast_fp16, y = values_23_cast_fp16)[name = tensor("attn_output_7_cast_fp16")]; + tensor var_1849 = const()[name = tensor("op_1849"), val = tensor([0, 2, 1, 3])]; + tensor var_1852 = const()[name = tensor("op_1852"), val = tensor([1, 1, 1024])]; + tensor var_1850_cast_fp16 = transpose(perm = var_1849, x = attn_output_7_cast_fp16)[name = tensor("transpose_200")]; + tensor input_35_cast_fp16 = reshape(shape = var_1852, x = var_1850_cast_fp16)[name = tensor("input_35_cast_fp16")]; + tensor linear_14_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = attn3_out_proj_weight_cast_fp16, x = input_35_cast_fp16)[name = tensor("linear_14_cast_fp16")]; + tensor var_1858_to_fp16 = const()[name = tensor("op_1858_to_fp16"), val = tensor(0x1p+0)]; + tensor var_1859_cast_fp16 = add(x = position3_to_fp16, y = var_1858_to_fp16)[name = tensor("op_1859_cast_fp16")]; + tensor var_1859_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_1859_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor input_37_cast_fp16 = add(x = input_33_cast_fp16, y = linear_14_cast_fp16)[name = tensor("input_37_cast_fp16")]; + tensor input_39_axes_0 = const()[name = tensor("input_39_axes_0"), val = tensor([-1])]; + tensor norm3_2_weight_to_fp16 = const()[name = tensor("norm3_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(303614976)))]; + tensor norm3_2_bias_to_fp16 = const()[name = tensor("norm3_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(303617088)))]; + tensor var_1863_to_fp16 = const()[name = tensor("op_1863_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_39_cast_fp16 = layer_norm(axes = input_39_axes_0, beta = norm3_2_bias_to_fp16, epsilon = var_1863_to_fp16, gamma = norm3_2_weight_to_fp16, x = input_37_cast_fp16)[name = tensor("input_39_cast_fp16")]; + tensor linear_15_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = linear3_1_weight_cast_fp16, x = input_39_cast_fp16)[name = tensor("linear_15_cast_fp16")]; + tensor input_41_mode_0 = const()[name = tensor("input_41_mode_0"), val = tensor("EXACT")]; + tensor input_41_cast_fp16 = gelu(mode = input_41_mode_0, x = linear_15_cast_fp16)[name = tensor("input_41_cast_fp16")]; + tensor linear_16_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = linear3_2_weight_cast_fp16, x = input_41_cast_fp16)[name = tensor("linear_16_cast_fp16")]; + tensor input_43_cast_fp16 = add(x = input_37_cast_fp16, y = linear_16_cast_fp16)[name = tensor("input_43_cast_fp16")]; + tensor x_9_axes_0 = const()[name = tensor("x_9_axes_0"), val = tensor([-1])]; + tensor norm4_1_weight_to_fp16 = const()[name = tensor("norm4_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(303619200)))]; + tensor norm4_1_bias_to_fp16 = const()[name = tensor("norm4_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(303621312)))]; + tensor var_1880_to_fp16 = const()[name = tensor("op_1880_to_fp16"), val = tensor(0x1.5p-17)]; + tensor x_9_cast_fp16 = layer_norm(axes = x_9_axes_0, beta = norm4_1_bias_to_fp16, epsilon = var_1880_to_fp16, gamma = norm4_1_weight_to_fp16, x = input_43_cast_fp16)[name = tensor("x_9_cast_fp16")]; + tensor linear_17_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = attn4_in_proj_weight_cast_fp16, x = x_9_cast_fp16)[name = tensor("linear_17_cast_fp16")]; + tensor var_1916 = const()[name = tensor("op_1916"), val = tensor([1, 1, 3, 16, 64])]; + tensor qkv_9_cast_fp16 = reshape(shape = var_1916, x = linear_17_cast_fp16)[name = tensor("qkv_9_cast_fp16")]; + tensor q_25_begin_0 = const()[name = tensor("q_25_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor q_25_end_0 = const()[name = tensor("q_25_end_0"), val = tensor([1, 1, 1, 16, 64])]; + tensor q_25_end_mask_0 = const()[name = tensor("q_25_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor q_25_squeeze_mask_0 = const()[name = tensor("q_25_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor q_25_cast_fp16 = slice_by_index(begin = q_25_begin_0, end = q_25_end_0, end_mask = q_25_end_mask_0, squeeze_mask = q_25_squeeze_mask_0, x = qkv_9_cast_fp16)[name = tensor("q_25_cast_fp16")]; + tensor k_17_begin_0 = const()[name = tensor("k_17_begin_0"), val = tensor([0, 0, 1, 0, 0])]; + tensor k_17_end_0 = const()[name = tensor("k_17_end_0"), val = tensor([1, 1, 2, 16, 64])]; + tensor k_17_end_mask_0 = const()[name = tensor("k_17_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor k_17_squeeze_mask_0 = const()[name = tensor("k_17_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor k_17_cast_fp16 = slice_by_index(begin = k_17_begin_0, end = k_17_end_0, end_mask = k_17_end_mask_0, squeeze_mask = k_17_squeeze_mask_0, x = qkv_9_cast_fp16)[name = tensor("k_17_cast_fp16")]; + tensor v_9_begin_0 = const()[name = tensor("v_9_begin_0"), val = tensor([0, 0, 2, 0, 0])]; + tensor v_9_end_0 = const()[name = tensor("v_9_end_0"), val = tensor([1, 1, 3, 16, 64])]; + tensor v_9_end_mask_0 = const()[name = tensor("v_9_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor v_9_squeeze_mask_0 = const()[name = tensor("v_9_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor v_9_cast_fp16 = slice_by_index(begin = v_9_begin_0, end = v_9_end_0, end_mask = v_9_end_mask_0, squeeze_mask = v_9_squeeze_mask_0, x = qkv_9_cast_fp16)[name = tensor("v_9_cast_fp16")]; + tensor var_2020 = const()[name = tensor("op_2020"), val = tensor([1, 1, 1, 1])]; + tensor position4_to_fp16_dtype_0 = const()[name = tensor("position4_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor position4_to_fp16 = cast(dtype = position4_to_fp16_dtype_0, x = position4)[name = tensor("cast_533")]; + tensor ts_29_cast_fp16 = reshape(shape = var_2020, x = position4_to_fp16)[name = tensor("ts_29_cast_fp16")]; + tensor var_2024 = const()[name = tensor("op_2024"), val = tensor([1, 1, 16, 32, 2])]; + tensor q_complex_9_cast_fp16 = reshape(shape = var_2024, x = q_25_cast_fp16)[name = tensor("q_complex_9_cast_fp16")]; + tensor var_2028 = const()[name = tensor("op_2028"), val = tensor([1, 1, 16, 32, 2])]; + tensor k_complex_9_cast_fp16 = reshape(shape = var_2028, x = k_17_cast_fp16)[name = tensor("k_complex_9_cast_fp16")]; + tensor var_2032_begin_0 = const()[name = tensor("op_2032_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_2032_end_0 = const()[name = tensor("op_2032_end_0"), val = tensor([1, 1, 16, 32, 1])]; + tensor var_2032_end_mask_0 = const()[name = tensor("op_2032_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_2032_squeeze_mask_0 = const()[name = tensor("op_2032_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_2032_cast_fp16 = slice_by_index(begin = var_2032_begin_0, end = var_2032_end_0, end_mask = var_2032_end_mask_0, squeeze_mask = var_2032_squeeze_mask_0, x = q_complex_9_cast_fp16)[name = tensor("op_2032_cast_fp16")]; + tensor var_2040_begin_0 = const()[name = tensor("op_2040_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_2040_end_0 = const()[name = tensor("op_2040_end_0"), val = tensor([1, 1, 16, 32, 2])]; + tensor var_2040_end_mask_0 = const()[name = tensor("op_2040_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_2040_squeeze_mask_0 = const()[name = tensor("op_2040_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_2040_cast_fp16 = slice_by_index(begin = var_2040_begin_0, end = var_2040_end_0, end_mask = var_2040_end_mask_0, squeeze_mask = var_2040_squeeze_mask_0, x = q_complex_9_cast_fp16)[name = tensor("op_2040_cast_fp16")]; + tensor var_2048_begin_0 = const()[name = tensor("op_2048_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_2048_end_0 = const()[name = tensor("op_2048_end_0"), val = tensor([1, 1, 16, 32, 1])]; + tensor var_2048_end_mask_0 = const()[name = tensor("op_2048_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_2048_squeeze_mask_0 = const()[name = tensor("op_2048_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_2048_cast_fp16 = slice_by_index(begin = var_2048_begin_0, end = var_2048_end_0, end_mask = var_2048_end_mask_0, squeeze_mask = var_2048_squeeze_mask_0, x = k_complex_9_cast_fp16)[name = tensor("op_2048_cast_fp16")]; + tensor var_2056_begin_0 = const()[name = tensor("op_2056_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_2056_end_0 = const()[name = tensor("op_2056_end_0"), val = tensor([1, 1, 16, 32, 2])]; + tensor var_2056_end_mask_0 = const()[name = tensor("op_2056_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_2056_squeeze_mask_0 = const()[name = tensor("op_2056_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_2056_cast_fp16 = slice_by_index(begin = var_2056_begin_0, end = var_2056_end_0, end_mask = var_2056_end_mask_0, squeeze_mask = var_2056_squeeze_mask_0, x = k_complex_9_cast_fp16)[name = tensor("op_2056_cast_fp16")]; + tensor freqs_9_to_fp16 = const()[name = tensor("freqs_9_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(303623424)))]; + tensor var_2062_cast_fp16 = mul(x = freqs_9_to_fp16, y = ts_29_cast_fp16)[name = tensor("op_2062_cast_fp16")]; + tensor rotr_9_cast_fp16 = cos(x = var_2062_cast_fp16)[name = tensor("rotr_9_cast_fp16")]; + tensor roti_9_cast_fp16 = sin(x = var_2062_cast_fp16)[name = tensor("roti_9_cast_fp16")]; + tensor var_2066_cast_fp16 = mul(x = var_2032_cast_fp16, y = rotr_9_cast_fp16)[name = tensor("op_2066_cast_fp16")]; + tensor var_2067_cast_fp16 = mul(x = var_2040_cast_fp16, y = roti_9_cast_fp16)[name = tensor("op_2067_cast_fp16")]; + tensor qor_17_cast_fp16 = sub(x = var_2066_cast_fp16, y = var_2067_cast_fp16)[name = tensor("qor_17_cast_fp16")]; + tensor var_2070_cast_fp16 = mul(x = var_2032_cast_fp16, y = roti_9_cast_fp16)[name = tensor("op_2070_cast_fp16")]; + tensor var_2071_cast_fp16 = mul(x = var_2040_cast_fp16, y = rotr_9_cast_fp16)[name = tensor("op_2071_cast_fp16")]; + tensor qoi_17_cast_fp16 = add(x = var_2070_cast_fp16, y = var_2071_cast_fp16)[name = tensor("qoi_17_cast_fp16")]; + tensor var_2074_cast_fp16 = mul(x = var_2048_cast_fp16, y = rotr_9_cast_fp16)[name = tensor("op_2074_cast_fp16")]; + tensor var_2075_cast_fp16 = mul(x = var_2056_cast_fp16, y = roti_9_cast_fp16)[name = tensor("op_2075_cast_fp16")]; + tensor kor_17_cast_fp16 = sub(x = var_2074_cast_fp16, y = var_2075_cast_fp16)[name = tensor("kor_17_cast_fp16")]; + tensor var_2078_cast_fp16 = mul(x = var_2048_cast_fp16, y = roti_9_cast_fp16)[name = tensor("op_2078_cast_fp16")]; + tensor var_2079_cast_fp16 = mul(x = var_2056_cast_fp16, y = rotr_9_cast_fp16)[name = tensor("op_2079_cast_fp16")]; + tensor koi_17_cast_fp16 = add(x = var_2078_cast_fp16, y = var_2079_cast_fp16)[name = tensor("koi_17_cast_fp16")]; + tensor qo_9_axis_0 = const()[name = tensor("qo_9_axis_0"), val = tensor(-1)]; + tensor qo_9_cast_fp16 = stack(axis = qo_9_axis_0, values = (qor_17_cast_fp16, qoi_17_cast_fp16))[name = tensor("qo_9_cast_fp16")]; + tensor ko_9_axis_0 = const()[name = tensor("ko_9_axis_0"), val = tensor(-1)]; + tensor ko_9_cast_fp16 = stack(axis = ko_9_axis_0, values = (kor_17_cast_fp16, koi_17_cast_fp16))[name = tensor("ko_9_cast_fp16")]; + tensor var_2108 = const()[name = tensor("op_2108"), val = tensor([1, 1, 16, 64])]; + tensor q_27_cast_fp16 = reshape(shape = var_2108, x = qo_9_cast_fp16)[name = tensor("q_27_cast_fp16")]; + tensor var_2110 = const()[name = tensor("op_2110"), val = tensor([1, 1, 16, 64])]; + tensor k_19_cast_fp16 = reshape(shape = var_2110, x = ko_9_cast_fp16)[name = tensor("k_19_cast_fp16")]; + tensor _inversed_2132_y_0_to_fp16 = const()[name = tensor("_inversed_2132_y_0_to_fp16"), val = tensor(0x1p-9)]; + tensor _inversed_2132_cast_fp16 = mul(x = ts_29_cast_fp16, y = _inversed_2132_y_0_to_fp16)[name = tensor("_inversed_2132_cast_fp16")]; + tensor var_2133_cast_fp16 = floor(x = _inversed_2132_cast_fp16)[name = tensor("op_2133_cast_fp16")]; + tensor var_2134_to_fp16 = const()[name = tensor("op_2134_to_fp16"), val = tensor(0x1p+9)]; + tensor var_2135_cast_fp16 = mul(x = var_2133_cast_fp16, y = var_2134_to_fp16)[name = tensor("op_2135_cast_fp16")]; + tensor write_indices_float_19_cast_fp16 = sub(x = ts_29_cast_fp16, y = var_2135_cast_fp16)[name = tensor("write_indices_float_19_cast_fp16")]; + tensor var_2142_dtype_0 = const()[name = tensor("op_2142_dtype_0"), val = tensor("int32")]; + tensor write_indices_9_reps_0 = const()[name = tensor("write_indices_9_reps_0"), val = tensor([1, 1, 16, 64])]; + tensor write_indices_float_19_cast_fp16_to_int32 = cast(dtype = var_2142_dtype_0, x = write_indices_float_19_cast_fp16)[name = tensor("cast_532")]; + tensor write_indices_9 = tile(reps = write_indices_9_reps_0, x = write_indices_float_19_cast_fp16_to_int32)[name = tensor("write_indices_9")]; + tensor var_2150_begin_0 = const()[name = tensor("op_2150_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_2150_end_0 = const()[name = tensor("op_2150_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor var_2150_end_mask_0 = const()[name = tensor("op_2150_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_2150_squeeze_mask_0 = const()[name = tensor("op_2150_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor cache4_to_fp16_dtype_0 = const()[name = tensor("cache4_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor cache4_to_fp16 = cast(dtype = cache4_to_fp16_dtype_0, x = cache4)[name = tensor("cast_531")]; + tensor var_2150_cast_fp16 = slice_by_index(begin = var_2150_begin_0, end = var_2150_end_0, end_mask = var_2150_end_mask_0, squeeze_mask = var_2150_squeeze_mask_0, x = cache4_to_fp16)[name = tensor("op_2150_cast_fp16")]; + tensor var_2152_axis_0 = const()[name = tensor("op_2152_axis_0"), val = tensor(1)]; + tensor var_2152_mode_0 = const()[name = tensor("op_2152_mode_0"), val = tensor("update")]; + tensor var_2152_validate_indices_0 = const()[name = tensor("op_2152_validate_indices_0"), val = tensor(false)]; + tensor var_2152_cast_fp16 = scatter_along_axis(axis = var_2152_axis_0, data = var_2150_cast_fp16, indices = write_indices_9, mode = var_2152_mode_0, updates = k_19_cast_fp16, validate_indices = var_2152_validate_indices_0)[name = tensor("op_2152_cast_fp16")]; + tensor concat_30 = const()[name = tensor("concat_30"), val = tensor([0, 0, 0, 0, 0])]; + tensor concat_31 = const()[name = tensor("concat_31"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_9_internal_tensor_assign_1_stride_0 = const()[name = tensor("new_cache_9_internal_tensor_assign_1_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_9_internal_tensor_assign_1_begin_mask_0 = const()[name = tensor("new_cache_9_internal_tensor_assign_1_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_9_internal_tensor_assign_1_end_mask_0 = const()[name = tensor("new_cache_9_internal_tensor_assign_1_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_9_internal_tensor_assign_1_squeeze_mask_0 = const()[name = tensor("new_cache_9_internal_tensor_assign_1_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_56 = const()[name = tensor("shape_56"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_8 = const()[name = tensor("reduce_prod_8"), val = tensor(1048576)]; + tensor range_1d_8_start_0 = const()[name = tensor("range_1d_8_start_0"), val = tensor(0)]; + tensor range_1d_8_step_0 = const()[name = tensor("range_1d_8_step_0"), val = tensor(1)]; + tensor range_1d_8 = range_1d(end = reduce_prod_8, start = range_1d_8_start_0, step = range_1d_8_step_0)[name = tensor("range_1d_8")]; + tensor reshape_40 = reshape(shape = shape_56, x = range_1d_8)[name = tensor("reshape_40")]; + tensor slice_by_index_8 = slice_by_index(begin = concat_30, begin_mask = new_cache_9_internal_tensor_assign_1_begin_mask_0, end = concat_31, end_mask = new_cache_9_internal_tensor_assign_1_end_mask_0, squeeze_mask = new_cache_9_internal_tensor_assign_1_squeeze_mask_0, stride = new_cache_9_internal_tensor_assign_1_stride_0, x = reshape_40)[name = tensor("slice_by_index_8")]; + tensor reshape_41_shape_0 = const()[name = tensor("reshape_41_shape_0"), val = tensor([-1])]; + tensor reshape_41 = reshape(shape = reshape_41_shape_0, x = slice_by_index_8)[name = tensor("reshape_41")]; + tensor reshape_42_shape_0 = const()[name = tensor("reshape_42_shape_0"), val = tensor([-1])]; + tensor reshape_42_cast_fp16 = reshape(shape = reshape_42_shape_0, x = var_2152_cast_fp16)[name = tensor("reshape_42_cast_fp16")]; + tensor reshape_43_shape_0 = const()[name = tensor("reshape_43_shape_0"), val = tensor([-1])]; + tensor reshape_43_cast_fp16 = reshape(shape = reshape_43_shape_0, x = cache4_to_fp16)[name = tensor("reshape_43_cast_fp16")]; + tensor scatter_8_mode_0 = const()[name = tensor("scatter_8_mode_0"), val = tensor("update")]; + tensor scatter_8_axis_0 = const()[name = tensor("scatter_8_axis_0"), val = tensor(0)]; + tensor scatter_8_validate_indices_0 = const()[name = tensor("scatter_8_validate_indices_0"), val = tensor(false)]; + tensor scatter_8_cast_fp16 = scatter(axis = scatter_8_axis_0, data = reshape_43_cast_fp16, indices = reshape_41, mode = scatter_8_mode_0, updates = reshape_42_cast_fp16, validate_indices = scatter_8_validate_indices_0)[name = tensor("scatter_8_cast_fp16")]; + tensor reshape_44_cast_fp16 = reshape(shape = shape_56, x = scatter_8_cast_fp16)[name = tensor("reshape_44_cast_fp16")]; + tensor var_2160_begin_0 = const()[name = tensor("op_2160_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor var_2160_end_0 = const()[name = tensor("op_2160_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor var_2160_end_mask_0 = const()[name = tensor("op_2160_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_2160_squeeze_mask_0 = const()[name = tensor("op_2160_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor var_2160_cast_fp16 = slice_by_index(begin = var_2160_begin_0, end = var_2160_end_0, end_mask = var_2160_end_mask_0, squeeze_mask = var_2160_squeeze_mask_0, x = reshape_44_cast_fp16)[name = tensor("op_2160_cast_fp16")]; + tensor var_2162_axis_0 = const()[name = tensor("op_2162_axis_0"), val = tensor(1)]; + tensor var_2162_mode_0 = const()[name = tensor("op_2162_mode_0"), val = tensor("update")]; + tensor var_2162_validate_indices_0 = const()[name = tensor("op_2162_validate_indices_0"), val = tensor(false)]; + tensor var_2162_cast_fp16 = scatter_along_axis(axis = var_2162_axis_0, data = var_2160_cast_fp16, indices = write_indices_9, mode = var_2162_mode_0, updates = v_9_cast_fp16, validate_indices = var_2162_validate_indices_0)[name = tensor("op_2162_cast_fp16")]; + tensor concat_32 = const()[name = tensor("concat_32"), val = tensor([1, 0, 0, 0, 0])]; + tensor concat_33 = const()[name = tensor("concat_33"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_9_internal_tensor_assign_2_stride_0 = const()[name = tensor("new_cache_9_internal_tensor_assign_2_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_9_internal_tensor_assign_2_begin_mask_0 = const()[name = tensor("new_cache_9_internal_tensor_assign_2_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_9_internal_tensor_assign_2_end_mask_0 = const()[name = tensor("new_cache_9_internal_tensor_assign_2_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_9_internal_tensor_assign_2_squeeze_mask_0 = const()[name = tensor("new_cache_9_internal_tensor_assign_2_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_57 = const()[name = tensor("shape_57"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_9 = const()[name = tensor("reduce_prod_9"), val = tensor(1048576)]; + tensor range_1d_9_start_0 = const()[name = tensor("range_1d_9_start_0"), val = tensor(0)]; + tensor range_1d_9_step_0 = const()[name = tensor("range_1d_9_step_0"), val = tensor(1)]; + tensor range_1d_9 = range_1d(end = reduce_prod_9, start = range_1d_9_start_0, step = range_1d_9_step_0)[name = tensor("range_1d_9")]; + tensor reshape_45 = reshape(shape = shape_57, x = range_1d_9)[name = tensor("reshape_45")]; + tensor slice_by_index_9 = slice_by_index(begin = concat_32, begin_mask = new_cache_9_internal_tensor_assign_2_begin_mask_0, end = concat_33, end_mask = new_cache_9_internal_tensor_assign_2_end_mask_0, squeeze_mask = new_cache_9_internal_tensor_assign_2_squeeze_mask_0, stride = new_cache_9_internal_tensor_assign_2_stride_0, x = reshape_45)[name = tensor("slice_by_index_9")]; + tensor reshape_46_shape_0 = const()[name = tensor("reshape_46_shape_0"), val = tensor([-1])]; + tensor reshape_46 = reshape(shape = reshape_46_shape_0, x = slice_by_index_9)[name = tensor("reshape_46")]; + tensor reshape_47_shape_0 = const()[name = tensor("reshape_47_shape_0"), val = tensor([-1])]; + tensor reshape_47_cast_fp16 = reshape(shape = reshape_47_shape_0, x = var_2162_cast_fp16)[name = tensor("reshape_47_cast_fp16")]; + tensor reshape_48_shape_0 = const()[name = tensor("reshape_48_shape_0"), val = tensor([-1])]; + tensor reshape_48_cast_fp16 = reshape(shape = reshape_48_shape_0, x = reshape_44_cast_fp16)[name = tensor("reshape_48_cast_fp16")]; + tensor scatter_9_mode_0 = const()[name = tensor("scatter_9_mode_0"), val = tensor("update")]; + tensor scatter_9_axis_0 = const()[name = tensor("scatter_9_axis_0"), val = tensor(0)]; + tensor scatter_9_validate_indices_0 = const()[name = tensor("scatter_9_validate_indices_0"), val = tensor(false)]; + tensor scatter_9_cast_fp16 = scatter(axis = scatter_9_axis_0, data = reshape_48_cast_fp16, indices = reshape_46, mode = scatter_9_mode_0, updates = reshape_47_cast_fp16, validate_indices = scatter_9_validate_indices_0)[name = tensor("scatter_9_cast_fp16")]; + tensor reshape_49_cast_fp16 = reshape(shape = shape_57, x = scatter_9_cast_fp16)[name = tensor("reshape_49_cast_fp16")]; + tensor reshape_49_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("reshape_49_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor keys_25_begin_0 = const()[name = tensor("keys_25_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor keys_25_end_0 = const()[name = tensor("keys_25_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor keys_25_end_mask_0 = const()[name = tensor("keys_25_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor keys_25_squeeze_mask_0 = const()[name = tensor("keys_25_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor keys_25_cast_fp16 = slice_by_index(begin = keys_25_begin_0, end = keys_25_end_0, end_mask = keys_25_end_mask_0, squeeze_mask = keys_25_squeeze_mask_0, x = reshape_49_cast_fp16)[name = tensor("keys_25_cast_fp16")]; + tensor values_25_begin_0 = const()[name = tensor("values_25_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor values_25_end_0 = const()[name = tensor("values_25_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor values_25_end_mask_0 = const()[name = tensor("values_25_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor values_25_squeeze_mask_0 = const()[name = tensor("values_25_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor values_25_cast_fp16 = slice_by_index(begin = values_25_begin_0, end = values_25_end_0, end_mask = values_25_end_mask_0, squeeze_mask = values_25_squeeze_mask_0, x = reshape_49_cast_fp16)[name = tensor("values_25_cast_fp16")]; + tensor var_2174_cast_fp16 = not_equal(x = keys_25_cast_fp16, y = keys_25_cast_fp16)[name = tensor("op_2174_cast_fp16")]; + tensor keys_27_cast_fp16 = select(a = var_504_to_fp16, b = keys_25_cast_fp16, cond = var_2174_cast_fp16)[name = tensor("keys_27_cast_fp16")]; + tensor var_2182_cast_fp16 = not_equal(x = values_25_cast_fp16, y = values_25_cast_fp16)[name = tensor("op_2182_cast_fp16")]; + tensor values_27_cast_fp16 = select(a = var_504_to_fp16, b = values_25_cast_fp16, cond = var_2182_cast_fp16)[name = tensor("values_27_cast_fp16")]; + tensor var_2206 = const()[name = tensor("op_2206"), val = tensor([0, 2, 1, 3])]; + tensor var_2219 = const()[name = tensor("op_2219"), val = tensor([1, 1, 1])]; + tensor var_2220_cast_fp16 = reshape(shape = var_2219, x = position4_to_fp16)[name = tensor("op_2220_cast_fp16")]; + tensor var_2237_to_fp16 = const()[name = tensor("op_2237_to_fp16"), val = tensor(0x1p+0)]; + tensor valid_len_9_cast_fp16 = add(x = var_2220_cast_fp16, y = var_2237_to_fp16)[name = tensor("valid_len_9_cast_fp16")]; + tensor valid_mask_9_cast_fp16 = less(x = k_positions_1_promoted_to_fp16, y = valid_len_9_cast_fp16)[name = tensor("valid_mask_9_cast_fp16")]; + tensor causal_mask_9_cast_fp16 = less_equal(x = k_positions_1_promoted_to_fp16, y = var_2220_cast_fp16)[name = tensor("causal_mask_9_cast_fp16")]; + tensor attn_mask_17 = logical_and(x = valid_mask_9_cast_fp16, y = causal_mask_9_cast_fp16)[name = tensor("attn_mask_17")]; + tensor attn_mask_19_axes_0 = const()[name = tensor("attn_mask_19_axes_0"), val = tensor([1])]; + tensor attn_mask_19 = expand_dims(axes = attn_mask_19_axes_0, x = attn_mask_17)[name = tensor("attn_mask_19")]; + tensor var_2255_transpose_x_0 = const()[name = tensor("op_2255_transpose_x_0"), val = tensor(false)]; + tensor var_2255_transpose_y_0 = const()[name = tensor("op_2255_transpose_y_0"), val = tensor(false)]; + tensor transpose_80_perm_0 = const()[name = tensor("transpose_80_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_81_perm_0 = const()[name = tensor("transpose_81_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_81 = transpose(perm = transpose_81_perm_0, x = keys_27_cast_fp16)[name = tensor("transpose_197")]; + tensor transpose_80 = transpose(perm = transpose_80_perm_0, x = q_27_cast_fp16)[name = tensor("transpose_198")]; + tensor var_2255_cast_fp16 = matmul(transpose_x = var_2255_transpose_x_0, transpose_y = var_2255_transpose_y_0, x = transpose_80, y = transpose_81)[name = tensor("op_2255_cast_fp16")]; + tensor var_2249_to_fp16 = const()[name = tensor("op_2249_to_fp16"), val = tensor([0x1p-3])]; + tensor attn_weights_25_cast_fp16 = mul(x = var_2255_cast_fp16, y = var_2249_to_fp16)[name = tensor("attn_weights_25_cast_fp16")]; + tensor var_2257 = logical_not(x = attn_mask_19)[name = tensor("op_2257")]; + tensor var_2258_to_fp16 = const()[name = tensor("op_2258_to_fp16"), val = tensor(-inf)]; + tensor attn_weights_27_cast_fp16 = select(a = var_2258_to_fp16, b = attn_weights_25_cast_fp16, cond = var_2257)[name = tensor("attn_weights_27_cast_fp16")]; + tensor var_2260 = const()[name = tensor("op_2260"), val = tensor(-1)]; + tensor attn_weights_29_cast_fp16 = softmax(axis = var_2260, x = attn_weights_27_cast_fp16)[name = tensor("attn_weights_29_cast_fp16")]; + tensor attn_output_9_transpose_x_0 = const()[name = tensor("attn_output_9_transpose_x_0"), val = tensor(false)]; + tensor attn_output_9_transpose_y_0 = const()[name = tensor("attn_output_9_transpose_y_0"), val = tensor(false)]; + tensor values_29_cast_fp16 = transpose(perm = var_2206, x = values_27_cast_fp16)[name = tensor("transpose_199")]; + tensor attn_output_9_cast_fp16 = matmul(transpose_x = attn_output_9_transpose_x_0, transpose_y = attn_output_9_transpose_y_0, x = attn_weights_29_cast_fp16, y = values_29_cast_fp16)[name = tensor("attn_output_9_cast_fp16")]; + tensor var_2268 = const()[name = tensor("op_2268"), val = tensor([0, 2, 1, 3])]; + tensor var_2271 = const()[name = tensor("op_2271"), val = tensor([1, 1, 1024])]; + tensor var_2269_cast_fp16 = transpose(perm = var_2268, x = attn_output_9_cast_fp16)[name = tensor("transpose_196")]; + tensor input_45_cast_fp16 = reshape(shape = var_2271, x = var_2269_cast_fp16)[name = tensor("input_45_cast_fp16")]; + tensor linear_18_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = attn4_out_proj_weight_cast_fp16, x = input_45_cast_fp16)[name = tensor("linear_18_cast_fp16")]; + tensor var_2277_to_fp16 = const()[name = tensor("op_2277_to_fp16"), val = tensor(0x1p+0)]; + tensor var_2278_cast_fp16 = add(x = position4_to_fp16, y = var_2277_to_fp16)[name = tensor("op_2278_cast_fp16")]; + tensor var_2278_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_2278_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor input_47_cast_fp16 = add(x = input_43_cast_fp16, y = linear_18_cast_fp16)[name = tensor("input_47_cast_fp16")]; + tensor input_49_axes_0 = const()[name = tensor("input_49_axes_0"), val = tensor([-1])]; + tensor norm4_2_weight_to_fp16 = const()[name = tensor("norm4_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(303623552)))]; + tensor norm4_2_bias_to_fp16 = const()[name = tensor("norm4_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(303625664)))]; + tensor var_2282_to_fp16 = const()[name = tensor("op_2282_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_49_cast_fp16 = layer_norm(axes = input_49_axes_0, beta = norm4_2_bias_to_fp16, epsilon = var_2282_to_fp16, gamma = norm4_2_weight_to_fp16, x = input_47_cast_fp16)[name = tensor("input_49_cast_fp16")]; + tensor linear_19_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = linear4_1_weight_cast_fp16, x = input_49_cast_fp16)[name = tensor("linear_19_cast_fp16")]; + tensor input_51_mode_0 = const()[name = tensor("input_51_mode_0"), val = tensor("EXACT")]; + tensor input_51_cast_fp16 = gelu(mode = input_51_mode_0, x = linear_19_cast_fp16)[name = tensor("input_51_cast_fp16")]; + tensor linear_20_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = linear4_2_weight_cast_fp16, x = input_51_cast_fp16)[name = tensor("linear_20_cast_fp16")]; + tensor input_53_cast_fp16 = add(x = input_47_cast_fp16, y = linear_20_cast_fp16)[name = tensor("input_53_cast_fp16")]; + tensor x_11_axes_0 = const()[name = tensor("x_11_axes_0"), val = tensor([-1])]; + tensor norm5_1_weight_to_fp16 = const()[name = tensor("norm5_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(303627776)))]; + tensor norm5_1_bias_to_fp16 = const()[name = tensor("norm5_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(303629888)))]; + tensor var_2299_to_fp16 = const()[name = tensor("op_2299_to_fp16"), val = tensor(0x1.5p-17)]; + tensor x_11_cast_fp16 = layer_norm(axes = x_11_axes_0, beta = norm5_1_bias_to_fp16, epsilon = var_2299_to_fp16, gamma = norm5_1_weight_to_fp16, x = input_53_cast_fp16)[name = tensor("x_11_cast_fp16")]; + tensor linear_21_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = attn5_in_proj_weight_cast_fp16, x = x_11_cast_fp16)[name = tensor("linear_21_cast_fp16")]; + tensor var_2335 = const()[name = tensor("op_2335"), val = tensor([1, 1, 3, 16, 64])]; + tensor qkv_11_cast_fp16 = reshape(shape = var_2335, x = linear_21_cast_fp16)[name = tensor("qkv_11_cast_fp16")]; + tensor q_31_begin_0 = const()[name = tensor("q_31_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor q_31_end_0 = const()[name = tensor("q_31_end_0"), val = tensor([1, 1, 1, 16, 64])]; + tensor q_31_end_mask_0 = const()[name = tensor("q_31_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor q_31_squeeze_mask_0 = const()[name = tensor("q_31_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor q_31_cast_fp16 = slice_by_index(begin = q_31_begin_0, end = q_31_end_0, end_mask = q_31_end_mask_0, squeeze_mask = q_31_squeeze_mask_0, x = qkv_11_cast_fp16)[name = tensor("q_31_cast_fp16")]; + tensor k_21_begin_0 = const()[name = tensor("k_21_begin_0"), val = tensor([0, 0, 1, 0, 0])]; + tensor k_21_end_0 = const()[name = tensor("k_21_end_0"), val = tensor([1, 1, 2, 16, 64])]; + tensor k_21_end_mask_0 = const()[name = tensor("k_21_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor k_21_squeeze_mask_0 = const()[name = tensor("k_21_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor k_21_cast_fp16 = slice_by_index(begin = k_21_begin_0, end = k_21_end_0, end_mask = k_21_end_mask_0, squeeze_mask = k_21_squeeze_mask_0, x = qkv_11_cast_fp16)[name = tensor("k_21_cast_fp16")]; + tensor v_11_begin_0 = const()[name = tensor("v_11_begin_0"), val = tensor([0, 0, 2, 0, 0])]; + tensor v_11_end_0 = const()[name = tensor("v_11_end_0"), val = tensor([1, 1, 3, 16, 64])]; + tensor v_11_end_mask_0 = const()[name = tensor("v_11_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor v_11_squeeze_mask_0 = const()[name = tensor("v_11_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor v_11_cast_fp16 = slice_by_index(begin = v_11_begin_0, end = v_11_end_0, end_mask = v_11_end_mask_0, squeeze_mask = v_11_squeeze_mask_0, x = qkv_11_cast_fp16)[name = tensor("v_11_cast_fp16")]; + tensor var_2439 = const()[name = tensor("op_2439"), val = tensor([1, 1, 1, 1])]; + tensor position5_to_fp16_dtype_0 = const()[name = tensor("position5_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor position5_to_fp16 = cast(dtype = position5_to_fp16_dtype_0, x = position5)[name = tensor("cast_528")]; + tensor ts_35_cast_fp16 = reshape(shape = var_2439, x = position5_to_fp16)[name = tensor("ts_35_cast_fp16")]; + tensor var_2443 = const()[name = tensor("op_2443"), val = tensor([1, 1, 16, 32, 2])]; + tensor q_complex_11_cast_fp16 = reshape(shape = var_2443, x = q_31_cast_fp16)[name = tensor("q_complex_11_cast_fp16")]; + tensor var_2447 = const()[name = tensor("op_2447"), val = tensor([1, 1, 16, 32, 2])]; + tensor k_complex_11_cast_fp16 = reshape(shape = var_2447, x = k_21_cast_fp16)[name = tensor("k_complex_11_cast_fp16")]; + tensor var_2451_begin_0 = const()[name = tensor("op_2451_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_2451_end_0 = const()[name = tensor("op_2451_end_0"), val = tensor([1, 1, 16, 32, 1])]; + tensor var_2451_end_mask_0 = const()[name = tensor("op_2451_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_2451_squeeze_mask_0 = const()[name = tensor("op_2451_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_2451_cast_fp16 = slice_by_index(begin = var_2451_begin_0, end = var_2451_end_0, end_mask = var_2451_end_mask_0, squeeze_mask = var_2451_squeeze_mask_0, x = q_complex_11_cast_fp16)[name = tensor("op_2451_cast_fp16")]; + tensor var_2459_begin_0 = const()[name = tensor("op_2459_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_2459_end_0 = const()[name = tensor("op_2459_end_0"), val = tensor([1, 1, 16, 32, 2])]; + tensor var_2459_end_mask_0 = const()[name = tensor("op_2459_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_2459_squeeze_mask_0 = const()[name = tensor("op_2459_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_2459_cast_fp16 = slice_by_index(begin = var_2459_begin_0, end = var_2459_end_0, end_mask = var_2459_end_mask_0, squeeze_mask = var_2459_squeeze_mask_0, x = q_complex_11_cast_fp16)[name = tensor("op_2459_cast_fp16")]; + tensor var_2467_begin_0 = const()[name = tensor("op_2467_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_2467_end_0 = const()[name = tensor("op_2467_end_0"), val = tensor([1, 1, 16, 32, 1])]; + tensor var_2467_end_mask_0 = const()[name = tensor("op_2467_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_2467_squeeze_mask_0 = const()[name = tensor("op_2467_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_2467_cast_fp16 = slice_by_index(begin = var_2467_begin_0, end = var_2467_end_0, end_mask = var_2467_end_mask_0, squeeze_mask = var_2467_squeeze_mask_0, x = k_complex_11_cast_fp16)[name = tensor("op_2467_cast_fp16")]; + tensor var_2475_begin_0 = const()[name = tensor("op_2475_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_2475_end_0 = const()[name = tensor("op_2475_end_0"), val = tensor([1, 1, 16, 32, 2])]; + tensor var_2475_end_mask_0 = const()[name = tensor("op_2475_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_2475_squeeze_mask_0 = const()[name = tensor("op_2475_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_2475_cast_fp16 = slice_by_index(begin = var_2475_begin_0, end = var_2475_end_0, end_mask = var_2475_end_mask_0, squeeze_mask = var_2475_squeeze_mask_0, x = k_complex_11_cast_fp16)[name = tensor("op_2475_cast_fp16")]; + tensor freqs_11_to_fp16 = const()[name = tensor("freqs_11_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(303632000)))]; + tensor var_2481_cast_fp16 = mul(x = freqs_11_to_fp16, y = ts_35_cast_fp16)[name = tensor("op_2481_cast_fp16")]; + tensor rotr_11_cast_fp16 = cos(x = var_2481_cast_fp16)[name = tensor("rotr_11_cast_fp16")]; + tensor roti_11_cast_fp16 = sin(x = var_2481_cast_fp16)[name = tensor("roti_11_cast_fp16")]; + tensor var_2485_cast_fp16 = mul(x = var_2451_cast_fp16, y = rotr_11_cast_fp16)[name = tensor("op_2485_cast_fp16")]; + tensor var_2486_cast_fp16 = mul(x = var_2459_cast_fp16, y = roti_11_cast_fp16)[name = tensor("op_2486_cast_fp16")]; + tensor qor_21_cast_fp16 = sub(x = var_2485_cast_fp16, y = var_2486_cast_fp16)[name = tensor("qor_21_cast_fp16")]; + tensor var_2489_cast_fp16 = mul(x = var_2451_cast_fp16, y = roti_11_cast_fp16)[name = tensor("op_2489_cast_fp16")]; + tensor var_2490_cast_fp16 = mul(x = var_2459_cast_fp16, y = rotr_11_cast_fp16)[name = tensor("op_2490_cast_fp16")]; + tensor qoi_21_cast_fp16 = add(x = var_2489_cast_fp16, y = var_2490_cast_fp16)[name = tensor("qoi_21_cast_fp16")]; + tensor var_2493_cast_fp16 = mul(x = var_2467_cast_fp16, y = rotr_11_cast_fp16)[name = tensor("op_2493_cast_fp16")]; + tensor var_2494_cast_fp16 = mul(x = var_2475_cast_fp16, y = roti_11_cast_fp16)[name = tensor("op_2494_cast_fp16")]; + tensor kor_21_cast_fp16 = sub(x = var_2493_cast_fp16, y = var_2494_cast_fp16)[name = tensor("kor_21_cast_fp16")]; + tensor var_2497_cast_fp16 = mul(x = var_2467_cast_fp16, y = roti_11_cast_fp16)[name = tensor("op_2497_cast_fp16")]; + tensor var_2498_cast_fp16 = mul(x = var_2475_cast_fp16, y = rotr_11_cast_fp16)[name = tensor("op_2498_cast_fp16")]; + tensor koi_21_cast_fp16 = add(x = var_2497_cast_fp16, y = var_2498_cast_fp16)[name = tensor("koi_21_cast_fp16")]; + tensor qo_11_axis_0 = const()[name = tensor("qo_11_axis_0"), val = tensor(-1)]; + tensor qo_11_cast_fp16 = stack(axis = qo_11_axis_0, values = (qor_21_cast_fp16, qoi_21_cast_fp16))[name = tensor("qo_11_cast_fp16")]; + tensor ko_11_axis_0 = const()[name = tensor("ko_11_axis_0"), val = tensor(-1)]; + tensor ko_11_cast_fp16 = stack(axis = ko_11_axis_0, values = (kor_21_cast_fp16, koi_21_cast_fp16))[name = tensor("ko_11_cast_fp16")]; + tensor var_2527 = const()[name = tensor("op_2527"), val = tensor([1, 1, 16, 64])]; + tensor q_33_cast_fp16 = reshape(shape = var_2527, x = qo_11_cast_fp16)[name = tensor("q_33_cast_fp16")]; + tensor var_2529 = const()[name = tensor("op_2529"), val = tensor([1, 1, 16, 64])]; + tensor k_23_cast_fp16 = reshape(shape = var_2529, x = ko_11_cast_fp16)[name = tensor("k_23_cast_fp16")]; + tensor _inversed_2551_y_0_to_fp16 = const()[name = tensor("_inversed_2551_y_0_to_fp16"), val = tensor(0x1p-9)]; + tensor _inversed_2551_cast_fp16 = mul(x = ts_35_cast_fp16, y = _inversed_2551_y_0_to_fp16)[name = tensor("_inversed_2551_cast_fp16")]; + tensor var_2552_cast_fp16 = floor(x = _inversed_2551_cast_fp16)[name = tensor("op_2552_cast_fp16")]; + tensor var_2553_to_fp16 = const()[name = tensor("op_2553_to_fp16"), val = tensor(0x1p+9)]; + tensor var_2554_cast_fp16 = mul(x = var_2552_cast_fp16, y = var_2553_to_fp16)[name = tensor("op_2554_cast_fp16")]; + tensor write_indices_float_23_cast_fp16 = sub(x = ts_35_cast_fp16, y = var_2554_cast_fp16)[name = tensor("write_indices_float_23_cast_fp16")]; + tensor var_2561_dtype_0 = const()[name = tensor("op_2561_dtype_0"), val = tensor("int32")]; + tensor write_indices_11_reps_0 = const()[name = tensor("write_indices_11_reps_0"), val = tensor([1, 1, 16, 64])]; + tensor write_indices_float_23_cast_fp16_to_int32 = cast(dtype = var_2561_dtype_0, x = write_indices_float_23_cast_fp16)[name = tensor("cast_527")]; + tensor write_indices_11 = tile(reps = write_indices_11_reps_0, x = write_indices_float_23_cast_fp16_to_int32)[name = tensor("write_indices_11")]; + tensor var_2569_begin_0 = const()[name = tensor("op_2569_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_2569_end_0 = const()[name = tensor("op_2569_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor var_2569_end_mask_0 = const()[name = tensor("op_2569_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_2569_squeeze_mask_0 = const()[name = tensor("op_2569_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor cache5_to_fp16_dtype_0 = const()[name = tensor("cache5_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor cache5_to_fp16 = cast(dtype = cache5_to_fp16_dtype_0, x = cache5)[name = tensor("cast_526")]; + tensor var_2569_cast_fp16 = slice_by_index(begin = var_2569_begin_0, end = var_2569_end_0, end_mask = var_2569_end_mask_0, squeeze_mask = var_2569_squeeze_mask_0, x = cache5_to_fp16)[name = tensor("op_2569_cast_fp16")]; + tensor var_2571_axis_0 = const()[name = tensor("op_2571_axis_0"), val = tensor(1)]; + tensor var_2571_mode_0 = const()[name = tensor("op_2571_mode_0"), val = tensor("update")]; + tensor var_2571_validate_indices_0 = const()[name = tensor("op_2571_validate_indices_0"), val = tensor(false)]; + tensor var_2571_cast_fp16 = scatter_along_axis(axis = var_2571_axis_0, data = var_2569_cast_fp16, indices = write_indices_11, mode = var_2571_mode_0, updates = k_23_cast_fp16, validate_indices = var_2571_validate_indices_0)[name = tensor("op_2571_cast_fp16")]; + tensor concat_37 = const()[name = tensor("concat_37"), val = tensor([0, 0, 0, 0, 0])]; + tensor concat_38 = const()[name = tensor("concat_38"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_11_internal_tensor_assign_1_stride_0 = const()[name = tensor("new_cache_11_internal_tensor_assign_1_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_11_internal_tensor_assign_1_begin_mask_0 = const()[name = tensor("new_cache_11_internal_tensor_assign_1_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_11_internal_tensor_assign_1_end_mask_0 = const()[name = tensor("new_cache_11_internal_tensor_assign_1_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_11_internal_tensor_assign_1_squeeze_mask_0 = const()[name = tensor("new_cache_11_internal_tensor_assign_1_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_58 = const()[name = tensor("shape_58"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_10 = const()[name = tensor("reduce_prod_10"), val = tensor(1048576)]; + tensor range_1d_10_start_0 = const()[name = tensor("range_1d_10_start_0"), val = tensor(0)]; + tensor range_1d_10_step_0 = const()[name = tensor("range_1d_10_step_0"), val = tensor(1)]; + tensor range_1d_10 = range_1d(end = reduce_prod_10, start = range_1d_10_start_0, step = range_1d_10_step_0)[name = tensor("range_1d_10")]; + tensor reshape_50 = reshape(shape = shape_58, x = range_1d_10)[name = tensor("reshape_50")]; + tensor slice_by_index_10 = slice_by_index(begin = concat_37, begin_mask = new_cache_11_internal_tensor_assign_1_begin_mask_0, end = concat_38, end_mask = new_cache_11_internal_tensor_assign_1_end_mask_0, squeeze_mask = new_cache_11_internal_tensor_assign_1_squeeze_mask_0, stride = new_cache_11_internal_tensor_assign_1_stride_0, x = reshape_50)[name = tensor("slice_by_index_10")]; + tensor reshape_51_shape_0 = const()[name = tensor("reshape_51_shape_0"), val = tensor([-1])]; + tensor reshape_51 = reshape(shape = reshape_51_shape_0, x = slice_by_index_10)[name = tensor("reshape_51")]; + tensor reshape_52_shape_0 = const()[name = tensor("reshape_52_shape_0"), val = tensor([-1])]; + tensor reshape_52_cast_fp16 = reshape(shape = reshape_52_shape_0, x = var_2571_cast_fp16)[name = tensor("reshape_52_cast_fp16")]; + tensor reshape_53_shape_0 = const()[name = tensor("reshape_53_shape_0"), val = tensor([-1])]; + tensor reshape_53_cast_fp16 = reshape(shape = reshape_53_shape_0, x = cache5_to_fp16)[name = tensor("reshape_53_cast_fp16")]; + tensor scatter_10_mode_0 = const()[name = tensor("scatter_10_mode_0"), val = tensor("update")]; + tensor scatter_10_axis_0 = const()[name = tensor("scatter_10_axis_0"), val = tensor(0)]; + tensor scatter_10_validate_indices_0 = const()[name = tensor("scatter_10_validate_indices_0"), val = tensor(false)]; + tensor scatter_10_cast_fp16 = scatter(axis = scatter_10_axis_0, data = reshape_53_cast_fp16, indices = reshape_51, mode = scatter_10_mode_0, updates = reshape_52_cast_fp16, validate_indices = scatter_10_validate_indices_0)[name = tensor("scatter_10_cast_fp16")]; + tensor reshape_54_cast_fp16 = reshape(shape = shape_58, x = scatter_10_cast_fp16)[name = tensor("reshape_54_cast_fp16")]; + tensor var_2579_begin_0 = const()[name = tensor("op_2579_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor var_2579_end_0 = const()[name = tensor("op_2579_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor var_2579_end_mask_0 = const()[name = tensor("op_2579_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_2579_squeeze_mask_0 = const()[name = tensor("op_2579_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor var_2579_cast_fp16 = slice_by_index(begin = var_2579_begin_0, end = var_2579_end_0, end_mask = var_2579_end_mask_0, squeeze_mask = var_2579_squeeze_mask_0, x = reshape_54_cast_fp16)[name = tensor("op_2579_cast_fp16")]; + tensor var_2581_axis_0 = const()[name = tensor("op_2581_axis_0"), val = tensor(1)]; + tensor var_2581_mode_0 = const()[name = tensor("op_2581_mode_0"), val = tensor("update")]; + tensor var_2581_validate_indices_0 = const()[name = tensor("op_2581_validate_indices_0"), val = tensor(false)]; + tensor var_2581_cast_fp16 = scatter_along_axis(axis = var_2581_axis_0, data = var_2579_cast_fp16, indices = write_indices_11, mode = var_2581_mode_0, updates = v_11_cast_fp16, validate_indices = var_2581_validate_indices_0)[name = tensor("op_2581_cast_fp16")]; + tensor concat_39 = const()[name = tensor("concat_39"), val = tensor([1, 0, 0, 0, 0])]; + tensor concat_40 = const()[name = tensor("concat_40"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_11_internal_tensor_assign_2_stride_0 = const()[name = tensor("new_cache_11_internal_tensor_assign_2_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_11_internal_tensor_assign_2_begin_mask_0 = const()[name = tensor("new_cache_11_internal_tensor_assign_2_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_11_internal_tensor_assign_2_end_mask_0 = const()[name = tensor("new_cache_11_internal_tensor_assign_2_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_11_internal_tensor_assign_2_squeeze_mask_0 = const()[name = tensor("new_cache_11_internal_tensor_assign_2_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_59 = const()[name = tensor("shape_59"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_11 = const()[name = tensor("reduce_prod_11"), val = tensor(1048576)]; + tensor range_1d_11_start_0 = const()[name = tensor("range_1d_11_start_0"), val = tensor(0)]; + tensor range_1d_11_step_0 = const()[name = tensor("range_1d_11_step_0"), val = tensor(1)]; + tensor range_1d_11 = range_1d(end = reduce_prod_11, start = range_1d_11_start_0, step = range_1d_11_step_0)[name = tensor("range_1d_11")]; + tensor reshape_55 = reshape(shape = shape_59, x = range_1d_11)[name = tensor("reshape_55")]; + tensor slice_by_index_11 = slice_by_index(begin = concat_39, begin_mask = new_cache_11_internal_tensor_assign_2_begin_mask_0, end = concat_40, end_mask = new_cache_11_internal_tensor_assign_2_end_mask_0, squeeze_mask = new_cache_11_internal_tensor_assign_2_squeeze_mask_0, stride = new_cache_11_internal_tensor_assign_2_stride_0, x = reshape_55)[name = tensor("slice_by_index_11")]; + tensor reshape_56_shape_0 = const()[name = tensor("reshape_56_shape_0"), val = tensor([-1])]; + tensor reshape_56 = reshape(shape = reshape_56_shape_0, x = slice_by_index_11)[name = tensor("reshape_56")]; + tensor reshape_57_shape_0 = const()[name = tensor("reshape_57_shape_0"), val = tensor([-1])]; + tensor reshape_57_cast_fp16 = reshape(shape = reshape_57_shape_0, x = var_2581_cast_fp16)[name = tensor("reshape_57_cast_fp16")]; + tensor reshape_58_shape_0 = const()[name = tensor("reshape_58_shape_0"), val = tensor([-1])]; + tensor reshape_58_cast_fp16 = reshape(shape = reshape_58_shape_0, x = reshape_54_cast_fp16)[name = tensor("reshape_58_cast_fp16")]; + tensor scatter_11_mode_0 = const()[name = tensor("scatter_11_mode_0"), val = tensor("update")]; + tensor scatter_11_axis_0 = const()[name = tensor("scatter_11_axis_0"), val = tensor(0)]; + tensor scatter_11_validate_indices_0 = const()[name = tensor("scatter_11_validate_indices_0"), val = tensor(false)]; + tensor scatter_11_cast_fp16 = scatter(axis = scatter_11_axis_0, data = reshape_58_cast_fp16, indices = reshape_56, mode = scatter_11_mode_0, updates = reshape_57_cast_fp16, validate_indices = scatter_11_validate_indices_0)[name = tensor("scatter_11_cast_fp16")]; + tensor reshape_59_cast_fp16 = reshape(shape = shape_59, x = scatter_11_cast_fp16)[name = tensor("reshape_59_cast_fp16")]; + tensor reshape_59_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("reshape_59_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor keys_31_begin_0 = const()[name = tensor("keys_31_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor keys_31_end_0 = const()[name = tensor("keys_31_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor keys_31_end_mask_0 = const()[name = tensor("keys_31_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor keys_31_squeeze_mask_0 = const()[name = tensor("keys_31_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor keys_31_cast_fp16 = slice_by_index(begin = keys_31_begin_0, end = keys_31_end_0, end_mask = keys_31_end_mask_0, squeeze_mask = keys_31_squeeze_mask_0, x = reshape_59_cast_fp16)[name = tensor("keys_31_cast_fp16")]; + tensor values_31_begin_0 = const()[name = tensor("values_31_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor values_31_end_0 = const()[name = tensor("values_31_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor values_31_end_mask_0 = const()[name = tensor("values_31_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor values_31_squeeze_mask_0 = const()[name = tensor("values_31_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor values_31_cast_fp16 = slice_by_index(begin = values_31_begin_0, end = values_31_end_0, end_mask = values_31_end_mask_0, squeeze_mask = values_31_squeeze_mask_0, x = reshape_59_cast_fp16)[name = tensor("values_31_cast_fp16")]; + tensor var_2593_cast_fp16 = not_equal(x = keys_31_cast_fp16, y = keys_31_cast_fp16)[name = tensor("op_2593_cast_fp16")]; + tensor keys_33_cast_fp16 = select(a = var_504_to_fp16, b = keys_31_cast_fp16, cond = var_2593_cast_fp16)[name = tensor("keys_33_cast_fp16")]; + tensor var_2601_cast_fp16 = not_equal(x = values_31_cast_fp16, y = values_31_cast_fp16)[name = tensor("op_2601_cast_fp16")]; + tensor values_33_cast_fp16 = select(a = var_504_to_fp16, b = values_31_cast_fp16, cond = var_2601_cast_fp16)[name = tensor("values_33_cast_fp16")]; + tensor var_2625 = const()[name = tensor("op_2625"), val = tensor([0, 2, 1, 3])]; + tensor var_2638 = const()[name = tensor("op_2638"), val = tensor([1, 1, 1])]; + tensor var_2639_cast_fp16 = reshape(shape = var_2638, x = position5_to_fp16)[name = tensor("op_2639_cast_fp16")]; + tensor var_2656_to_fp16 = const()[name = tensor("op_2656_to_fp16"), val = tensor(0x1p+0)]; + tensor valid_len_11_cast_fp16 = add(x = var_2639_cast_fp16, y = var_2656_to_fp16)[name = tensor("valid_len_11_cast_fp16")]; + tensor valid_mask_11_cast_fp16 = less(x = k_positions_1_promoted_to_fp16, y = valid_len_11_cast_fp16)[name = tensor("valid_mask_11_cast_fp16")]; + tensor causal_mask_11_cast_fp16 = less_equal(x = k_positions_1_promoted_to_fp16, y = var_2639_cast_fp16)[name = tensor("causal_mask_11_cast_fp16")]; + tensor attn_mask_21 = logical_and(x = valid_mask_11_cast_fp16, y = causal_mask_11_cast_fp16)[name = tensor("attn_mask_21")]; + tensor attn_mask_23_axes_0 = const()[name = tensor("attn_mask_23_axes_0"), val = tensor([1])]; + tensor attn_mask_23 = expand_dims(axes = attn_mask_23_axes_0, x = attn_mask_21)[name = tensor("attn_mask_23")]; + tensor var_2674_transpose_x_0 = const()[name = tensor("op_2674_transpose_x_0"), val = tensor(false)]; + tensor var_2674_transpose_y_0 = const()[name = tensor("op_2674_transpose_y_0"), val = tensor(false)]; + tensor transpose_82_perm_0 = const()[name = tensor("transpose_82_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_83_perm_0 = const()[name = tensor("transpose_83_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_83 = transpose(perm = transpose_83_perm_0, x = keys_33_cast_fp16)[name = tensor("transpose_193")]; + tensor transpose_82 = transpose(perm = transpose_82_perm_0, x = q_33_cast_fp16)[name = tensor("transpose_194")]; + tensor var_2674_cast_fp16 = matmul(transpose_x = var_2674_transpose_x_0, transpose_y = var_2674_transpose_y_0, x = transpose_82, y = transpose_83)[name = tensor("op_2674_cast_fp16")]; + tensor var_2668_to_fp16 = const()[name = tensor("op_2668_to_fp16"), val = tensor([0x1p-3])]; + tensor attn_weights_31_cast_fp16 = mul(x = var_2674_cast_fp16, y = var_2668_to_fp16)[name = tensor("attn_weights_31_cast_fp16")]; + tensor var_2676 = logical_not(x = attn_mask_23)[name = tensor("op_2676")]; + tensor var_2677_to_fp16 = const()[name = tensor("op_2677_to_fp16"), val = tensor(-inf)]; + tensor attn_weights_33_cast_fp16 = select(a = var_2677_to_fp16, b = attn_weights_31_cast_fp16, cond = var_2676)[name = tensor("attn_weights_33_cast_fp16")]; + tensor var_2679 = const()[name = tensor("op_2679"), val = tensor(-1)]; + tensor attn_weights_35_cast_fp16 = softmax(axis = var_2679, x = attn_weights_33_cast_fp16)[name = tensor("attn_weights_35_cast_fp16")]; + tensor attn_output_11_transpose_x_0 = const()[name = tensor("attn_output_11_transpose_x_0"), val = tensor(false)]; + tensor attn_output_11_transpose_y_0 = const()[name = tensor("attn_output_11_transpose_y_0"), val = tensor(false)]; + tensor values_35_cast_fp16 = transpose(perm = var_2625, x = values_33_cast_fp16)[name = tensor("transpose_195")]; + tensor attn_output_11_cast_fp16 = matmul(transpose_x = attn_output_11_transpose_x_0, transpose_y = attn_output_11_transpose_y_0, x = attn_weights_35_cast_fp16, y = values_35_cast_fp16)[name = tensor("attn_output_11_cast_fp16")]; + tensor var_2687 = const()[name = tensor("op_2687"), val = tensor([0, 2, 1, 3])]; + tensor var_2690 = const()[name = tensor("op_2690"), val = tensor([1, 1, 1024])]; + tensor var_2688_cast_fp16 = transpose(perm = var_2687, x = attn_output_11_cast_fp16)[name = tensor("transpose_192")]; + tensor input_55_cast_fp16 = reshape(shape = var_2690, x = var_2688_cast_fp16)[name = tensor("input_55_cast_fp16")]; + tensor linear_22_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = attn5_out_proj_weight_cast_fp16, x = input_55_cast_fp16)[name = tensor("linear_22_cast_fp16")]; + tensor var_2696_to_fp16 = const()[name = tensor("op_2696_to_fp16"), val = tensor(0x1p+0)]; + tensor var_2697_cast_fp16 = add(x = position5_to_fp16, y = var_2696_to_fp16)[name = tensor("op_2697_cast_fp16")]; + tensor var_2697_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_2697_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor input_57_cast_fp16 = add(x = input_53_cast_fp16, y = linear_22_cast_fp16)[name = tensor("input_57_cast_fp16")]; + tensor input_59_axes_0 = const()[name = tensor("input_59_axes_0"), val = tensor([-1])]; + tensor norm5_2_weight_to_fp16 = const()[name = tensor("norm5_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(303632128)))]; + tensor norm5_2_bias_to_fp16 = const()[name = tensor("norm5_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(303634240)))]; + tensor var_2701_to_fp16 = const()[name = tensor("op_2701_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_59_cast_fp16 = layer_norm(axes = input_59_axes_0, beta = norm5_2_bias_to_fp16, epsilon = var_2701_to_fp16, gamma = norm5_2_weight_to_fp16, x = input_57_cast_fp16)[name = tensor("input_59_cast_fp16")]; + tensor linear_23_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = linear5_1_weight_cast_fp16, x = input_59_cast_fp16)[name = tensor("linear_23_cast_fp16")]; + tensor input_61_mode_0 = const()[name = tensor("input_61_mode_0"), val = tensor("EXACT")]; + tensor input_61_cast_fp16 = gelu(mode = input_61_mode_0, x = linear_23_cast_fp16)[name = tensor("input_61_cast_fp16")]; + tensor linear_24_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = linear5_2_weight_cast_fp16, x = input_61_cast_fp16)[name = tensor("linear_24_cast_fp16")]; + tensor input_63_cast_fp16 = add(x = input_57_cast_fp16, y = linear_24_cast_fp16)[name = tensor("input_63_cast_fp16")]; + tensor x_13_axes_0 = const()[name = tensor("x_13_axes_0"), val = tensor([-1])]; + tensor norm6_1_weight_to_fp16 = const()[name = tensor("norm6_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(303636352)))]; + tensor norm6_1_bias_to_fp16 = const()[name = tensor("norm6_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(303638464)))]; + tensor var_2718_to_fp16 = const()[name = tensor("op_2718_to_fp16"), val = tensor(0x1.5p-17)]; + tensor x_13_cast_fp16 = layer_norm(axes = x_13_axes_0, beta = norm6_1_bias_to_fp16, epsilon = var_2718_to_fp16, gamma = norm6_1_weight_to_fp16, x = input_63_cast_fp16)[name = tensor("x_13_cast_fp16")]; + tensor linear_25_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = attn6_in_proj_weight_cast_fp16, x = x_13_cast_fp16)[name = tensor("linear_25_cast_fp16")]; + tensor var_2754 = const()[name = tensor("op_2754"), val = tensor([1, 1, 3, 16, 64])]; + tensor qkv_13_cast_fp16 = reshape(shape = var_2754, x = linear_25_cast_fp16)[name = tensor("qkv_13_cast_fp16")]; + tensor q_37_begin_0 = const()[name = tensor("q_37_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor q_37_end_0 = const()[name = tensor("q_37_end_0"), val = tensor([1, 1, 1, 16, 64])]; + tensor q_37_end_mask_0 = const()[name = tensor("q_37_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor q_37_squeeze_mask_0 = const()[name = tensor("q_37_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor q_37_cast_fp16 = slice_by_index(begin = q_37_begin_0, end = q_37_end_0, end_mask = q_37_end_mask_0, squeeze_mask = q_37_squeeze_mask_0, x = qkv_13_cast_fp16)[name = tensor("q_37_cast_fp16")]; + tensor k_25_begin_0 = const()[name = tensor("k_25_begin_0"), val = tensor([0, 0, 1, 0, 0])]; + tensor k_25_end_0 = const()[name = tensor("k_25_end_0"), val = tensor([1, 1, 2, 16, 64])]; + tensor k_25_end_mask_0 = const()[name = tensor("k_25_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor k_25_squeeze_mask_0 = const()[name = tensor("k_25_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor k_25_cast_fp16 = slice_by_index(begin = k_25_begin_0, end = k_25_end_0, end_mask = k_25_end_mask_0, squeeze_mask = k_25_squeeze_mask_0, x = qkv_13_cast_fp16)[name = tensor("k_25_cast_fp16")]; + tensor v_13_begin_0 = const()[name = tensor("v_13_begin_0"), val = tensor([0, 0, 2, 0, 0])]; + tensor v_13_end_0 = const()[name = tensor("v_13_end_0"), val = tensor([1, 1, 3, 16, 64])]; + tensor v_13_end_mask_0 = const()[name = tensor("v_13_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor v_13_squeeze_mask_0 = const()[name = tensor("v_13_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor v_13_cast_fp16 = slice_by_index(begin = v_13_begin_0, end = v_13_end_0, end_mask = v_13_end_mask_0, squeeze_mask = v_13_squeeze_mask_0, x = qkv_13_cast_fp16)[name = tensor("v_13_cast_fp16")]; + tensor var_2858 = const()[name = tensor("op_2858"), val = tensor([1, 1, 1, 1])]; + tensor position6_to_fp16_dtype_0 = const()[name = tensor("position6_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor position6_to_fp16 = cast(dtype = position6_to_fp16_dtype_0, x = position6)[name = tensor("cast_523")]; + tensor ts_41_cast_fp16 = reshape(shape = var_2858, x = position6_to_fp16)[name = tensor("ts_41_cast_fp16")]; + tensor var_2862 = const()[name = tensor("op_2862"), val = tensor([1, 1, 16, 32, 2])]; + tensor q_complex_13_cast_fp16 = reshape(shape = var_2862, x = q_37_cast_fp16)[name = tensor("q_complex_13_cast_fp16")]; + tensor var_2866 = const()[name = tensor("op_2866"), val = tensor([1, 1, 16, 32, 2])]; + tensor k_complex_13_cast_fp16 = reshape(shape = var_2866, x = k_25_cast_fp16)[name = tensor("k_complex_13_cast_fp16")]; + tensor var_2870_begin_0 = const()[name = tensor("op_2870_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_2870_end_0 = const()[name = tensor("op_2870_end_0"), val = tensor([1, 1, 16, 32, 1])]; + tensor var_2870_end_mask_0 = const()[name = tensor("op_2870_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_2870_squeeze_mask_0 = const()[name = tensor("op_2870_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_2870_cast_fp16 = slice_by_index(begin = var_2870_begin_0, end = var_2870_end_0, end_mask = var_2870_end_mask_0, squeeze_mask = var_2870_squeeze_mask_0, x = q_complex_13_cast_fp16)[name = tensor("op_2870_cast_fp16")]; + tensor var_2878_begin_0 = const()[name = tensor("op_2878_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_2878_end_0 = const()[name = tensor("op_2878_end_0"), val = tensor([1, 1, 16, 32, 2])]; + tensor var_2878_end_mask_0 = const()[name = tensor("op_2878_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_2878_squeeze_mask_0 = const()[name = tensor("op_2878_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_2878_cast_fp16 = slice_by_index(begin = var_2878_begin_0, end = var_2878_end_0, end_mask = var_2878_end_mask_0, squeeze_mask = var_2878_squeeze_mask_0, x = q_complex_13_cast_fp16)[name = tensor("op_2878_cast_fp16")]; + tensor var_2886_begin_0 = const()[name = tensor("op_2886_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_2886_end_0 = const()[name = tensor("op_2886_end_0"), val = tensor([1, 1, 16, 32, 1])]; + tensor var_2886_end_mask_0 = const()[name = tensor("op_2886_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_2886_squeeze_mask_0 = const()[name = tensor("op_2886_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_2886_cast_fp16 = slice_by_index(begin = var_2886_begin_0, end = var_2886_end_0, end_mask = var_2886_end_mask_0, squeeze_mask = var_2886_squeeze_mask_0, x = k_complex_13_cast_fp16)[name = tensor("op_2886_cast_fp16")]; + tensor var_2894_begin_0 = const()[name = tensor("op_2894_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_2894_end_0 = const()[name = tensor("op_2894_end_0"), val = tensor([1, 1, 16, 32, 2])]; + tensor var_2894_end_mask_0 = const()[name = tensor("op_2894_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_2894_squeeze_mask_0 = const()[name = tensor("op_2894_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_2894_cast_fp16 = slice_by_index(begin = var_2894_begin_0, end = var_2894_end_0, end_mask = var_2894_end_mask_0, squeeze_mask = var_2894_squeeze_mask_0, x = k_complex_13_cast_fp16)[name = tensor("op_2894_cast_fp16")]; + tensor freqs_13_to_fp16 = const()[name = tensor("freqs_13_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(303640576)))]; + tensor var_2900_cast_fp16 = mul(x = freqs_13_to_fp16, y = ts_41_cast_fp16)[name = tensor("op_2900_cast_fp16")]; + tensor rotr_13_cast_fp16 = cos(x = var_2900_cast_fp16)[name = tensor("rotr_13_cast_fp16")]; + tensor roti_13_cast_fp16 = sin(x = var_2900_cast_fp16)[name = tensor("roti_13_cast_fp16")]; + tensor var_2904_cast_fp16 = mul(x = var_2870_cast_fp16, y = rotr_13_cast_fp16)[name = tensor("op_2904_cast_fp16")]; + tensor var_2905_cast_fp16 = mul(x = var_2878_cast_fp16, y = roti_13_cast_fp16)[name = tensor("op_2905_cast_fp16")]; + tensor qor_25_cast_fp16 = sub(x = var_2904_cast_fp16, y = var_2905_cast_fp16)[name = tensor("qor_25_cast_fp16")]; + tensor var_2908_cast_fp16 = mul(x = var_2870_cast_fp16, y = roti_13_cast_fp16)[name = tensor("op_2908_cast_fp16")]; + tensor var_2909_cast_fp16 = mul(x = var_2878_cast_fp16, y = rotr_13_cast_fp16)[name = tensor("op_2909_cast_fp16")]; + tensor qoi_25_cast_fp16 = add(x = var_2908_cast_fp16, y = var_2909_cast_fp16)[name = tensor("qoi_25_cast_fp16")]; + tensor var_2912_cast_fp16 = mul(x = var_2886_cast_fp16, y = rotr_13_cast_fp16)[name = tensor("op_2912_cast_fp16")]; + tensor var_2913_cast_fp16 = mul(x = var_2894_cast_fp16, y = roti_13_cast_fp16)[name = tensor("op_2913_cast_fp16")]; + tensor kor_25_cast_fp16 = sub(x = var_2912_cast_fp16, y = var_2913_cast_fp16)[name = tensor("kor_25_cast_fp16")]; + tensor var_2916_cast_fp16 = mul(x = var_2886_cast_fp16, y = roti_13_cast_fp16)[name = tensor("op_2916_cast_fp16")]; + tensor var_2917_cast_fp16 = mul(x = var_2894_cast_fp16, y = rotr_13_cast_fp16)[name = tensor("op_2917_cast_fp16")]; + tensor koi_25_cast_fp16 = add(x = var_2916_cast_fp16, y = var_2917_cast_fp16)[name = tensor("koi_25_cast_fp16")]; + tensor qo_13_axis_0 = const()[name = tensor("qo_13_axis_0"), val = tensor(-1)]; + tensor qo_13_cast_fp16 = stack(axis = qo_13_axis_0, values = (qor_25_cast_fp16, qoi_25_cast_fp16))[name = tensor("qo_13_cast_fp16")]; + tensor ko_13_axis_0 = const()[name = tensor("ko_13_axis_0"), val = tensor(-1)]; + tensor ko_13_cast_fp16 = stack(axis = ko_13_axis_0, values = (kor_25_cast_fp16, koi_25_cast_fp16))[name = tensor("ko_13_cast_fp16")]; + tensor var_2946 = const()[name = tensor("op_2946"), val = tensor([1, 1, 16, 64])]; + tensor q_39_cast_fp16 = reshape(shape = var_2946, x = qo_13_cast_fp16)[name = tensor("q_39_cast_fp16")]; + tensor var_2948 = const()[name = tensor("op_2948"), val = tensor([1, 1, 16, 64])]; + tensor k_27_cast_fp16 = reshape(shape = var_2948, x = ko_13_cast_fp16)[name = tensor("k_27_cast_fp16")]; + tensor _inversed_2970_y_0_to_fp16 = const()[name = tensor("_inversed_2970_y_0_to_fp16"), val = tensor(0x1p-9)]; + tensor _inversed_2970_cast_fp16 = mul(x = ts_41_cast_fp16, y = _inversed_2970_y_0_to_fp16)[name = tensor("_inversed_2970_cast_fp16")]; + tensor var_2971_cast_fp16 = floor(x = _inversed_2970_cast_fp16)[name = tensor("op_2971_cast_fp16")]; + tensor var_2972_to_fp16 = const()[name = tensor("op_2972_to_fp16"), val = tensor(0x1p+9)]; + tensor var_2973_cast_fp16 = mul(x = var_2971_cast_fp16, y = var_2972_to_fp16)[name = tensor("op_2973_cast_fp16")]; + tensor write_indices_float_27_cast_fp16 = sub(x = ts_41_cast_fp16, y = var_2973_cast_fp16)[name = tensor("write_indices_float_27_cast_fp16")]; + tensor var_2980_dtype_0 = const()[name = tensor("op_2980_dtype_0"), val = tensor("int32")]; + tensor write_indices_13_reps_0 = const()[name = tensor("write_indices_13_reps_0"), val = tensor([1, 1, 16, 64])]; + tensor write_indices_float_27_cast_fp16_to_int32 = cast(dtype = var_2980_dtype_0, x = write_indices_float_27_cast_fp16)[name = tensor("cast_522")]; + tensor write_indices_13 = tile(reps = write_indices_13_reps_0, x = write_indices_float_27_cast_fp16_to_int32)[name = tensor("write_indices_13")]; + tensor var_2988_begin_0 = const()[name = tensor("op_2988_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_2988_end_0 = const()[name = tensor("op_2988_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor var_2988_end_mask_0 = const()[name = tensor("op_2988_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_2988_squeeze_mask_0 = const()[name = tensor("op_2988_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor cache6_to_fp16_dtype_0 = const()[name = tensor("cache6_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor cache6_to_fp16 = cast(dtype = cache6_to_fp16_dtype_0, x = cache6)[name = tensor("cast_521")]; + tensor var_2988_cast_fp16 = slice_by_index(begin = var_2988_begin_0, end = var_2988_end_0, end_mask = var_2988_end_mask_0, squeeze_mask = var_2988_squeeze_mask_0, x = cache6_to_fp16)[name = tensor("op_2988_cast_fp16")]; + tensor var_2990_axis_0 = const()[name = tensor("op_2990_axis_0"), val = tensor(1)]; + tensor var_2990_mode_0 = const()[name = tensor("op_2990_mode_0"), val = tensor("update")]; + tensor var_2990_validate_indices_0 = const()[name = tensor("op_2990_validate_indices_0"), val = tensor(false)]; + tensor var_2990_cast_fp16 = scatter_along_axis(axis = var_2990_axis_0, data = var_2988_cast_fp16, indices = write_indices_13, mode = var_2990_mode_0, updates = k_27_cast_fp16, validate_indices = var_2990_validate_indices_0)[name = tensor("op_2990_cast_fp16")]; + tensor concat_44 = const()[name = tensor("concat_44"), val = tensor([0, 0, 0, 0, 0])]; + tensor concat_45 = const()[name = tensor("concat_45"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_13_internal_tensor_assign_1_stride_0 = const()[name = tensor("new_cache_13_internal_tensor_assign_1_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_13_internal_tensor_assign_1_begin_mask_0 = const()[name = tensor("new_cache_13_internal_tensor_assign_1_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_13_internal_tensor_assign_1_end_mask_0 = const()[name = tensor("new_cache_13_internal_tensor_assign_1_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_13_internal_tensor_assign_1_squeeze_mask_0 = const()[name = tensor("new_cache_13_internal_tensor_assign_1_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_60 = const()[name = tensor("shape_60"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_12 = const()[name = tensor("reduce_prod_12"), val = tensor(1048576)]; + tensor range_1d_12_start_0 = const()[name = tensor("range_1d_12_start_0"), val = tensor(0)]; + tensor range_1d_12_step_0 = const()[name = tensor("range_1d_12_step_0"), val = tensor(1)]; + tensor range_1d_12 = range_1d(end = reduce_prod_12, start = range_1d_12_start_0, step = range_1d_12_step_0)[name = tensor("range_1d_12")]; + tensor reshape_60 = reshape(shape = shape_60, x = range_1d_12)[name = tensor("reshape_60")]; + tensor slice_by_index_12 = slice_by_index(begin = concat_44, begin_mask = new_cache_13_internal_tensor_assign_1_begin_mask_0, end = concat_45, end_mask = new_cache_13_internal_tensor_assign_1_end_mask_0, squeeze_mask = new_cache_13_internal_tensor_assign_1_squeeze_mask_0, stride = new_cache_13_internal_tensor_assign_1_stride_0, x = reshape_60)[name = tensor("slice_by_index_12")]; + tensor reshape_61_shape_0 = const()[name = tensor("reshape_61_shape_0"), val = tensor([-1])]; + tensor reshape_61 = reshape(shape = reshape_61_shape_0, x = slice_by_index_12)[name = tensor("reshape_61")]; + tensor reshape_62_shape_0 = const()[name = tensor("reshape_62_shape_0"), val = tensor([-1])]; + tensor reshape_62_cast_fp16 = reshape(shape = reshape_62_shape_0, x = var_2990_cast_fp16)[name = tensor("reshape_62_cast_fp16")]; + tensor reshape_63_shape_0 = const()[name = tensor("reshape_63_shape_0"), val = tensor([-1])]; + tensor reshape_63_cast_fp16 = reshape(shape = reshape_63_shape_0, x = cache6_to_fp16)[name = tensor("reshape_63_cast_fp16")]; + tensor scatter_12_mode_0 = const()[name = tensor("scatter_12_mode_0"), val = tensor("update")]; + tensor scatter_12_axis_0 = const()[name = tensor("scatter_12_axis_0"), val = tensor(0)]; + tensor scatter_12_validate_indices_0 = const()[name = tensor("scatter_12_validate_indices_0"), val = tensor(false)]; + tensor scatter_12_cast_fp16 = scatter(axis = scatter_12_axis_0, data = reshape_63_cast_fp16, indices = reshape_61, mode = scatter_12_mode_0, updates = reshape_62_cast_fp16, validate_indices = scatter_12_validate_indices_0)[name = tensor("scatter_12_cast_fp16")]; + tensor reshape_64_cast_fp16 = reshape(shape = shape_60, x = scatter_12_cast_fp16)[name = tensor("reshape_64_cast_fp16")]; + tensor var_2998_begin_0 = const()[name = tensor("op_2998_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor var_2998_end_0 = const()[name = tensor("op_2998_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor var_2998_end_mask_0 = const()[name = tensor("op_2998_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_2998_squeeze_mask_0 = const()[name = tensor("op_2998_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor var_2998_cast_fp16 = slice_by_index(begin = var_2998_begin_0, end = var_2998_end_0, end_mask = var_2998_end_mask_0, squeeze_mask = var_2998_squeeze_mask_0, x = reshape_64_cast_fp16)[name = tensor("op_2998_cast_fp16")]; + tensor var_3000_axis_0 = const()[name = tensor("op_3000_axis_0"), val = tensor(1)]; + tensor var_3000_mode_0 = const()[name = tensor("op_3000_mode_0"), val = tensor("update")]; + tensor var_3000_validate_indices_0 = const()[name = tensor("op_3000_validate_indices_0"), val = tensor(false)]; + tensor var_3000_cast_fp16 = scatter_along_axis(axis = var_3000_axis_0, data = var_2998_cast_fp16, indices = write_indices_13, mode = var_3000_mode_0, updates = v_13_cast_fp16, validate_indices = var_3000_validate_indices_0)[name = tensor("op_3000_cast_fp16")]; + tensor concat_46 = const()[name = tensor("concat_46"), val = tensor([1, 0, 0, 0, 0])]; + tensor concat_47 = const()[name = tensor("concat_47"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_13_internal_tensor_assign_2_stride_0 = const()[name = tensor("new_cache_13_internal_tensor_assign_2_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_13_internal_tensor_assign_2_begin_mask_0 = const()[name = tensor("new_cache_13_internal_tensor_assign_2_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_13_internal_tensor_assign_2_end_mask_0 = const()[name = tensor("new_cache_13_internal_tensor_assign_2_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_13_internal_tensor_assign_2_squeeze_mask_0 = const()[name = tensor("new_cache_13_internal_tensor_assign_2_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_61 = const()[name = tensor("shape_61"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_13 = const()[name = tensor("reduce_prod_13"), val = tensor(1048576)]; + tensor range_1d_13_start_0 = const()[name = tensor("range_1d_13_start_0"), val = tensor(0)]; + tensor range_1d_13_step_0 = const()[name = tensor("range_1d_13_step_0"), val = tensor(1)]; + tensor range_1d_13 = range_1d(end = reduce_prod_13, start = range_1d_13_start_0, step = range_1d_13_step_0)[name = tensor("range_1d_13")]; + tensor reshape_65 = reshape(shape = shape_61, x = range_1d_13)[name = tensor("reshape_65")]; + tensor slice_by_index_13 = slice_by_index(begin = concat_46, begin_mask = new_cache_13_internal_tensor_assign_2_begin_mask_0, end = concat_47, end_mask = new_cache_13_internal_tensor_assign_2_end_mask_0, squeeze_mask = new_cache_13_internal_tensor_assign_2_squeeze_mask_0, stride = new_cache_13_internal_tensor_assign_2_stride_0, x = reshape_65)[name = tensor("slice_by_index_13")]; + tensor reshape_66_shape_0 = const()[name = tensor("reshape_66_shape_0"), val = tensor([-1])]; + tensor reshape_66 = reshape(shape = reshape_66_shape_0, x = slice_by_index_13)[name = tensor("reshape_66")]; + tensor reshape_67_shape_0 = const()[name = tensor("reshape_67_shape_0"), val = tensor([-1])]; + tensor reshape_67_cast_fp16 = reshape(shape = reshape_67_shape_0, x = var_3000_cast_fp16)[name = tensor("reshape_67_cast_fp16")]; + tensor reshape_68_shape_0 = const()[name = tensor("reshape_68_shape_0"), val = tensor([-1])]; + tensor reshape_68_cast_fp16 = reshape(shape = reshape_68_shape_0, x = reshape_64_cast_fp16)[name = tensor("reshape_68_cast_fp16")]; + tensor scatter_13_mode_0 = const()[name = tensor("scatter_13_mode_0"), val = tensor("update")]; + tensor scatter_13_axis_0 = const()[name = tensor("scatter_13_axis_0"), val = tensor(0)]; + tensor scatter_13_validate_indices_0 = const()[name = tensor("scatter_13_validate_indices_0"), val = tensor(false)]; + tensor scatter_13_cast_fp16 = scatter(axis = scatter_13_axis_0, data = reshape_68_cast_fp16, indices = reshape_66, mode = scatter_13_mode_0, updates = reshape_67_cast_fp16, validate_indices = scatter_13_validate_indices_0)[name = tensor("scatter_13_cast_fp16")]; + tensor reshape_69_cast_fp16 = reshape(shape = shape_61, x = scatter_13_cast_fp16)[name = tensor("reshape_69_cast_fp16")]; + tensor reshape_69_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("reshape_69_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor keys_37_begin_0 = const()[name = tensor("keys_37_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor keys_37_end_0 = const()[name = tensor("keys_37_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor keys_37_end_mask_0 = const()[name = tensor("keys_37_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor keys_37_squeeze_mask_0 = const()[name = tensor("keys_37_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor keys_37_cast_fp16 = slice_by_index(begin = keys_37_begin_0, end = keys_37_end_0, end_mask = keys_37_end_mask_0, squeeze_mask = keys_37_squeeze_mask_0, x = reshape_69_cast_fp16)[name = tensor("keys_37_cast_fp16")]; + tensor values_37_begin_0 = const()[name = tensor("values_37_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor values_37_end_0 = const()[name = tensor("values_37_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor values_37_end_mask_0 = const()[name = tensor("values_37_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor values_37_squeeze_mask_0 = const()[name = tensor("values_37_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor values_37_cast_fp16 = slice_by_index(begin = values_37_begin_0, end = values_37_end_0, end_mask = values_37_end_mask_0, squeeze_mask = values_37_squeeze_mask_0, x = reshape_69_cast_fp16)[name = tensor("values_37_cast_fp16")]; + tensor var_3012_cast_fp16 = not_equal(x = keys_37_cast_fp16, y = keys_37_cast_fp16)[name = tensor("op_3012_cast_fp16")]; + tensor keys_39_cast_fp16 = select(a = var_504_to_fp16, b = keys_37_cast_fp16, cond = var_3012_cast_fp16)[name = tensor("keys_39_cast_fp16")]; + tensor var_3020_cast_fp16 = not_equal(x = values_37_cast_fp16, y = values_37_cast_fp16)[name = tensor("op_3020_cast_fp16")]; + tensor values_39_cast_fp16 = select(a = var_504_to_fp16, b = values_37_cast_fp16, cond = var_3020_cast_fp16)[name = tensor("values_39_cast_fp16")]; + tensor var_3044 = const()[name = tensor("op_3044"), val = tensor([0, 2, 1, 3])]; + tensor var_3057 = const()[name = tensor("op_3057"), val = tensor([1, 1, 1])]; + tensor var_3058_cast_fp16 = reshape(shape = var_3057, x = position6_to_fp16)[name = tensor("op_3058_cast_fp16")]; + tensor var_3075_to_fp16 = const()[name = tensor("op_3075_to_fp16"), val = tensor(0x1p+0)]; + tensor valid_len_13_cast_fp16 = add(x = var_3058_cast_fp16, y = var_3075_to_fp16)[name = tensor("valid_len_13_cast_fp16")]; + tensor valid_mask_13_cast_fp16 = less(x = k_positions_1_promoted_to_fp16, y = valid_len_13_cast_fp16)[name = tensor("valid_mask_13_cast_fp16")]; + tensor causal_mask_13_cast_fp16 = less_equal(x = k_positions_1_promoted_to_fp16, y = var_3058_cast_fp16)[name = tensor("causal_mask_13_cast_fp16")]; + tensor attn_mask_25 = logical_and(x = valid_mask_13_cast_fp16, y = causal_mask_13_cast_fp16)[name = tensor("attn_mask_25")]; + tensor attn_mask_27_axes_0 = const()[name = tensor("attn_mask_27_axes_0"), val = tensor([1])]; + tensor attn_mask_27 = expand_dims(axes = attn_mask_27_axes_0, x = attn_mask_25)[name = tensor("attn_mask_27")]; + tensor var_3093_transpose_x_0 = const()[name = tensor("op_3093_transpose_x_0"), val = tensor(false)]; + tensor var_3093_transpose_y_0 = const()[name = tensor("op_3093_transpose_y_0"), val = tensor(false)]; + tensor transpose_84_perm_0 = const()[name = tensor("transpose_84_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_85_perm_0 = const()[name = tensor("transpose_85_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_85 = transpose(perm = transpose_85_perm_0, x = keys_39_cast_fp16)[name = tensor("transpose_189")]; + tensor transpose_84 = transpose(perm = transpose_84_perm_0, x = q_39_cast_fp16)[name = tensor("transpose_190")]; + tensor var_3093_cast_fp16 = matmul(transpose_x = var_3093_transpose_x_0, transpose_y = var_3093_transpose_y_0, x = transpose_84, y = transpose_85)[name = tensor("op_3093_cast_fp16")]; + tensor var_3087_to_fp16 = const()[name = tensor("op_3087_to_fp16"), val = tensor([0x1p-3])]; + tensor attn_weights_37_cast_fp16 = mul(x = var_3093_cast_fp16, y = var_3087_to_fp16)[name = tensor("attn_weights_37_cast_fp16")]; + tensor var_3095 = logical_not(x = attn_mask_27)[name = tensor("op_3095")]; + tensor var_3096_to_fp16 = const()[name = tensor("op_3096_to_fp16"), val = tensor(-inf)]; + tensor attn_weights_39_cast_fp16 = select(a = var_3096_to_fp16, b = attn_weights_37_cast_fp16, cond = var_3095)[name = tensor("attn_weights_39_cast_fp16")]; + tensor var_3098 = const()[name = tensor("op_3098"), val = tensor(-1)]; + tensor attn_weights_41_cast_fp16 = softmax(axis = var_3098, x = attn_weights_39_cast_fp16)[name = tensor("attn_weights_41_cast_fp16")]; + tensor attn_output_13_transpose_x_0 = const()[name = tensor("attn_output_13_transpose_x_0"), val = tensor(false)]; + tensor attn_output_13_transpose_y_0 = const()[name = tensor("attn_output_13_transpose_y_0"), val = tensor(false)]; + tensor values_41_cast_fp16 = transpose(perm = var_3044, x = values_39_cast_fp16)[name = tensor("transpose_191")]; + tensor attn_output_13_cast_fp16 = matmul(transpose_x = attn_output_13_transpose_x_0, transpose_y = attn_output_13_transpose_y_0, x = attn_weights_41_cast_fp16, y = values_41_cast_fp16)[name = tensor("attn_output_13_cast_fp16")]; + tensor var_3106 = const()[name = tensor("op_3106"), val = tensor([0, 2, 1, 3])]; + tensor var_3109 = const()[name = tensor("op_3109"), val = tensor([1, 1, 1024])]; + tensor var_3107_cast_fp16 = transpose(perm = var_3106, x = attn_output_13_cast_fp16)[name = tensor("transpose_188")]; + tensor input_65_cast_fp16 = reshape(shape = var_3109, x = var_3107_cast_fp16)[name = tensor("input_65_cast_fp16")]; + tensor linear_26_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = attn6_out_proj_weight_cast_fp16, x = input_65_cast_fp16)[name = tensor("linear_26_cast_fp16")]; + tensor var_3115_to_fp16 = const()[name = tensor("op_3115_to_fp16"), val = tensor(0x1p+0)]; + tensor var_3116_cast_fp16 = add(x = position6_to_fp16, y = var_3115_to_fp16)[name = tensor("op_3116_cast_fp16")]; + tensor var_3116_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_3116_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor input_67_cast_fp16 = add(x = input_63_cast_fp16, y = linear_26_cast_fp16)[name = tensor("input_67_cast_fp16")]; + tensor input_69_axes_0 = const()[name = tensor("input_69_axes_0"), val = tensor([-1])]; + tensor norm6_2_weight_to_fp16 = const()[name = tensor("norm6_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(303640704)))]; + tensor norm6_2_bias_to_fp16 = const()[name = tensor("norm6_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(303642816)))]; + tensor var_3120_to_fp16 = const()[name = tensor("op_3120_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_69_cast_fp16 = layer_norm(axes = input_69_axes_0, beta = norm6_2_bias_to_fp16, epsilon = var_3120_to_fp16, gamma = norm6_2_weight_to_fp16, x = input_67_cast_fp16)[name = tensor("input_69_cast_fp16")]; + tensor linear_27_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = linear6_1_weight_cast_fp16, x = input_69_cast_fp16)[name = tensor("linear_27_cast_fp16")]; + tensor input_71_mode_0 = const()[name = tensor("input_71_mode_0"), val = tensor("EXACT")]; + tensor input_71_cast_fp16 = gelu(mode = input_71_mode_0, x = linear_27_cast_fp16)[name = tensor("input_71_cast_fp16")]; + tensor linear_28_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = linear6_2_weight_cast_fp16, x = input_71_cast_fp16)[name = tensor("linear_28_cast_fp16")]; + tensor input_73_cast_fp16 = add(x = input_67_cast_fp16, y = linear_28_cast_fp16)[name = tensor("input_73_cast_fp16")]; + tensor x_15_axes_0 = const()[name = tensor("x_15_axes_0"), val = tensor([-1])]; + tensor norm7_1_weight_to_fp16 = const()[name = tensor("norm7_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(303644928)))]; + tensor norm7_1_bias_to_fp16 = const()[name = tensor("norm7_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(303647040)))]; + tensor var_3137_to_fp16 = const()[name = tensor("op_3137_to_fp16"), val = tensor(0x1.5p-17)]; + tensor x_15_cast_fp16 = layer_norm(axes = x_15_axes_0, beta = norm7_1_bias_to_fp16, epsilon = var_3137_to_fp16, gamma = norm7_1_weight_to_fp16, x = input_73_cast_fp16)[name = tensor("x_15_cast_fp16")]; + tensor linear_29_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = attn7_in_proj_weight_cast_fp16, x = x_15_cast_fp16)[name = tensor("linear_29_cast_fp16")]; + tensor var_3173 = const()[name = tensor("op_3173"), val = tensor([1, 1, 3, 16, 64])]; + tensor qkv_15_cast_fp16 = reshape(shape = var_3173, x = linear_29_cast_fp16)[name = tensor("qkv_15_cast_fp16")]; + tensor q_43_begin_0 = const()[name = tensor("q_43_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor q_43_end_0 = const()[name = tensor("q_43_end_0"), val = tensor([1, 1, 1, 16, 64])]; + tensor q_43_end_mask_0 = const()[name = tensor("q_43_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor q_43_squeeze_mask_0 = const()[name = tensor("q_43_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor q_43_cast_fp16 = slice_by_index(begin = q_43_begin_0, end = q_43_end_0, end_mask = q_43_end_mask_0, squeeze_mask = q_43_squeeze_mask_0, x = qkv_15_cast_fp16)[name = tensor("q_43_cast_fp16")]; + tensor k_29_begin_0 = const()[name = tensor("k_29_begin_0"), val = tensor([0, 0, 1, 0, 0])]; + tensor k_29_end_0 = const()[name = tensor("k_29_end_0"), val = tensor([1, 1, 2, 16, 64])]; + tensor k_29_end_mask_0 = const()[name = tensor("k_29_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor k_29_squeeze_mask_0 = const()[name = tensor("k_29_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor k_29_cast_fp16 = slice_by_index(begin = k_29_begin_0, end = k_29_end_0, end_mask = k_29_end_mask_0, squeeze_mask = k_29_squeeze_mask_0, x = qkv_15_cast_fp16)[name = tensor("k_29_cast_fp16")]; + tensor v_15_begin_0 = const()[name = tensor("v_15_begin_0"), val = tensor([0, 0, 2, 0, 0])]; + tensor v_15_end_0 = const()[name = tensor("v_15_end_0"), val = tensor([1, 1, 3, 16, 64])]; + tensor v_15_end_mask_0 = const()[name = tensor("v_15_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor v_15_squeeze_mask_0 = const()[name = tensor("v_15_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor v_15_cast_fp16 = slice_by_index(begin = v_15_begin_0, end = v_15_end_0, end_mask = v_15_end_mask_0, squeeze_mask = v_15_squeeze_mask_0, x = qkv_15_cast_fp16)[name = tensor("v_15_cast_fp16")]; + tensor var_3277 = const()[name = tensor("op_3277"), val = tensor([1, 1, 1, 1])]; + tensor position7_to_fp16_dtype_0 = const()[name = tensor("position7_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor position7_to_fp16 = cast(dtype = position7_to_fp16_dtype_0, x = position7)[name = tensor("cast_518")]; + tensor ts_47_cast_fp16 = reshape(shape = var_3277, x = position7_to_fp16)[name = tensor("ts_47_cast_fp16")]; + tensor var_3281 = const()[name = tensor("op_3281"), val = tensor([1, 1, 16, 32, 2])]; + tensor q_complex_15_cast_fp16 = reshape(shape = var_3281, x = q_43_cast_fp16)[name = tensor("q_complex_15_cast_fp16")]; + tensor var_3285 = const()[name = tensor("op_3285"), val = tensor([1, 1, 16, 32, 2])]; + tensor k_complex_15_cast_fp16 = reshape(shape = var_3285, x = k_29_cast_fp16)[name = tensor("k_complex_15_cast_fp16")]; + tensor var_3289_begin_0 = const()[name = tensor("op_3289_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_3289_end_0 = const()[name = tensor("op_3289_end_0"), val = tensor([1, 1, 16, 32, 1])]; + tensor var_3289_end_mask_0 = const()[name = tensor("op_3289_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_3289_squeeze_mask_0 = const()[name = tensor("op_3289_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_3289_cast_fp16 = slice_by_index(begin = var_3289_begin_0, end = var_3289_end_0, end_mask = var_3289_end_mask_0, squeeze_mask = var_3289_squeeze_mask_0, x = q_complex_15_cast_fp16)[name = tensor("op_3289_cast_fp16")]; + tensor var_3297_begin_0 = const()[name = tensor("op_3297_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_3297_end_0 = const()[name = tensor("op_3297_end_0"), val = tensor([1, 1, 16, 32, 2])]; + tensor var_3297_end_mask_0 = const()[name = tensor("op_3297_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_3297_squeeze_mask_0 = const()[name = tensor("op_3297_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_3297_cast_fp16 = slice_by_index(begin = var_3297_begin_0, end = var_3297_end_0, end_mask = var_3297_end_mask_0, squeeze_mask = var_3297_squeeze_mask_0, x = q_complex_15_cast_fp16)[name = tensor("op_3297_cast_fp16")]; + tensor var_3305_begin_0 = const()[name = tensor("op_3305_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_3305_end_0 = const()[name = tensor("op_3305_end_0"), val = tensor([1, 1, 16, 32, 1])]; + tensor var_3305_end_mask_0 = const()[name = tensor("op_3305_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_3305_squeeze_mask_0 = const()[name = tensor("op_3305_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_3305_cast_fp16 = slice_by_index(begin = var_3305_begin_0, end = var_3305_end_0, end_mask = var_3305_end_mask_0, squeeze_mask = var_3305_squeeze_mask_0, x = k_complex_15_cast_fp16)[name = tensor("op_3305_cast_fp16")]; + tensor var_3313_begin_0 = const()[name = tensor("op_3313_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_3313_end_0 = const()[name = tensor("op_3313_end_0"), val = tensor([1, 1, 16, 32, 2])]; + tensor var_3313_end_mask_0 = const()[name = tensor("op_3313_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_3313_squeeze_mask_0 = const()[name = tensor("op_3313_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_3313_cast_fp16 = slice_by_index(begin = var_3313_begin_0, end = var_3313_end_0, end_mask = var_3313_end_mask_0, squeeze_mask = var_3313_squeeze_mask_0, x = k_complex_15_cast_fp16)[name = tensor("op_3313_cast_fp16")]; + tensor freqs_15_to_fp16 = const()[name = tensor("freqs_15_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(303649152)))]; + tensor var_3319_cast_fp16 = mul(x = freqs_15_to_fp16, y = ts_47_cast_fp16)[name = tensor("op_3319_cast_fp16")]; + tensor rotr_15_cast_fp16 = cos(x = var_3319_cast_fp16)[name = tensor("rotr_15_cast_fp16")]; + tensor roti_15_cast_fp16 = sin(x = var_3319_cast_fp16)[name = tensor("roti_15_cast_fp16")]; + tensor var_3323_cast_fp16 = mul(x = var_3289_cast_fp16, y = rotr_15_cast_fp16)[name = tensor("op_3323_cast_fp16")]; + tensor var_3324_cast_fp16 = mul(x = var_3297_cast_fp16, y = roti_15_cast_fp16)[name = tensor("op_3324_cast_fp16")]; + tensor qor_29_cast_fp16 = sub(x = var_3323_cast_fp16, y = var_3324_cast_fp16)[name = tensor("qor_29_cast_fp16")]; + tensor var_3327_cast_fp16 = mul(x = var_3289_cast_fp16, y = roti_15_cast_fp16)[name = tensor("op_3327_cast_fp16")]; + tensor var_3328_cast_fp16 = mul(x = var_3297_cast_fp16, y = rotr_15_cast_fp16)[name = tensor("op_3328_cast_fp16")]; + tensor qoi_29_cast_fp16 = add(x = var_3327_cast_fp16, y = var_3328_cast_fp16)[name = tensor("qoi_29_cast_fp16")]; + tensor var_3331_cast_fp16 = mul(x = var_3305_cast_fp16, y = rotr_15_cast_fp16)[name = tensor("op_3331_cast_fp16")]; + tensor var_3332_cast_fp16 = mul(x = var_3313_cast_fp16, y = roti_15_cast_fp16)[name = tensor("op_3332_cast_fp16")]; + tensor kor_29_cast_fp16 = sub(x = var_3331_cast_fp16, y = var_3332_cast_fp16)[name = tensor("kor_29_cast_fp16")]; + tensor var_3335_cast_fp16 = mul(x = var_3305_cast_fp16, y = roti_15_cast_fp16)[name = tensor("op_3335_cast_fp16")]; + tensor var_3336_cast_fp16 = mul(x = var_3313_cast_fp16, y = rotr_15_cast_fp16)[name = tensor("op_3336_cast_fp16")]; + tensor koi_29_cast_fp16 = add(x = var_3335_cast_fp16, y = var_3336_cast_fp16)[name = tensor("koi_29_cast_fp16")]; + tensor qo_15_axis_0 = const()[name = tensor("qo_15_axis_0"), val = tensor(-1)]; + tensor qo_15_cast_fp16 = stack(axis = qo_15_axis_0, values = (qor_29_cast_fp16, qoi_29_cast_fp16))[name = tensor("qo_15_cast_fp16")]; + tensor ko_15_axis_0 = const()[name = tensor("ko_15_axis_0"), val = tensor(-1)]; + tensor ko_15_cast_fp16 = stack(axis = ko_15_axis_0, values = (kor_29_cast_fp16, koi_29_cast_fp16))[name = tensor("ko_15_cast_fp16")]; + tensor var_3365 = const()[name = tensor("op_3365"), val = tensor([1, 1, 16, 64])]; + tensor q_45_cast_fp16 = reshape(shape = var_3365, x = qo_15_cast_fp16)[name = tensor("q_45_cast_fp16")]; + tensor var_3367 = const()[name = tensor("op_3367"), val = tensor([1, 1, 16, 64])]; + tensor k_31_cast_fp16 = reshape(shape = var_3367, x = ko_15_cast_fp16)[name = tensor("k_31_cast_fp16")]; + tensor _inversed_3389_y_0_to_fp16 = const()[name = tensor("_inversed_3389_y_0_to_fp16"), val = tensor(0x1p-9)]; + tensor _inversed_3389_cast_fp16 = mul(x = ts_47_cast_fp16, y = _inversed_3389_y_0_to_fp16)[name = tensor("_inversed_3389_cast_fp16")]; + tensor var_3390_cast_fp16 = floor(x = _inversed_3389_cast_fp16)[name = tensor("op_3390_cast_fp16")]; + tensor var_3391_to_fp16 = const()[name = tensor("op_3391_to_fp16"), val = tensor(0x1p+9)]; + tensor var_3392_cast_fp16 = mul(x = var_3390_cast_fp16, y = var_3391_to_fp16)[name = tensor("op_3392_cast_fp16")]; + tensor write_indices_float_31_cast_fp16 = sub(x = ts_47_cast_fp16, y = var_3392_cast_fp16)[name = tensor("write_indices_float_31_cast_fp16")]; + tensor var_3399_dtype_0 = const()[name = tensor("op_3399_dtype_0"), val = tensor("int32")]; + tensor write_indices_15_reps_0 = const()[name = tensor("write_indices_15_reps_0"), val = tensor([1, 1, 16, 64])]; + tensor write_indices_float_31_cast_fp16_to_int32 = cast(dtype = var_3399_dtype_0, x = write_indices_float_31_cast_fp16)[name = tensor("cast_517")]; + tensor write_indices_15 = tile(reps = write_indices_15_reps_0, x = write_indices_float_31_cast_fp16_to_int32)[name = tensor("write_indices_15")]; + tensor var_3407_begin_0 = const()[name = tensor("op_3407_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_3407_end_0 = const()[name = tensor("op_3407_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor var_3407_end_mask_0 = const()[name = tensor("op_3407_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_3407_squeeze_mask_0 = const()[name = tensor("op_3407_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor cache7_to_fp16_dtype_0 = const()[name = tensor("cache7_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor cache7_to_fp16 = cast(dtype = cache7_to_fp16_dtype_0, x = cache7)[name = tensor("cast_516")]; + tensor var_3407_cast_fp16 = slice_by_index(begin = var_3407_begin_0, end = var_3407_end_0, end_mask = var_3407_end_mask_0, squeeze_mask = var_3407_squeeze_mask_0, x = cache7_to_fp16)[name = tensor("op_3407_cast_fp16")]; + tensor var_3409_axis_0 = const()[name = tensor("op_3409_axis_0"), val = tensor(1)]; + tensor var_3409_mode_0 = const()[name = tensor("op_3409_mode_0"), val = tensor("update")]; + tensor var_3409_validate_indices_0 = const()[name = tensor("op_3409_validate_indices_0"), val = tensor(false)]; + tensor var_3409_cast_fp16 = scatter_along_axis(axis = var_3409_axis_0, data = var_3407_cast_fp16, indices = write_indices_15, mode = var_3409_mode_0, updates = k_31_cast_fp16, validate_indices = var_3409_validate_indices_0)[name = tensor("op_3409_cast_fp16")]; + tensor concat_51 = const()[name = tensor("concat_51"), val = tensor([0, 0, 0, 0, 0])]; + tensor concat_52 = const()[name = tensor("concat_52"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_15_internal_tensor_assign_1_stride_0 = const()[name = tensor("new_cache_15_internal_tensor_assign_1_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_15_internal_tensor_assign_1_begin_mask_0 = const()[name = tensor("new_cache_15_internal_tensor_assign_1_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_15_internal_tensor_assign_1_end_mask_0 = const()[name = tensor("new_cache_15_internal_tensor_assign_1_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_15_internal_tensor_assign_1_squeeze_mask_0 = const()[name = tensor("new_cache_15_internal_tensor_assign_1_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_62 = const()[name = tensor("shape_62"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_14 = const()[name = tensor("reduce_prod_14"), val = tensor(1048576)]; + tensor range_1d_14_start_0 = const()[name = tensor("range_1d_14_start_0"), val = tensor(0)]; + tensor range_1d_14_step_0 = const()[name = tensor("range_1d_14_step_0"), val = tensor(1)]; + tensor range_1d_14 = range_1d(end = reduce_prod_14, start = range_1d_14_start_0, step = range_1d_14_step_0)[name = tensor("range_1d_14")]; + tensor reshape_70 = reshape(shape = shape_62, x = range_1d_14)[name = tensor("reshape_70")]; + tensor slice_by_index_14 = slice_by_index(begin = concat_51, begin_mask = new_cache_15_internal_tensor_assign_1_begin_mask_0, end = concat_52, end_mask = new_cache_15_internal_tensor_assign_1_end_mask_0, squeeze_mask = new_cache_15_internal_tensor_assign_1_squeeze_mask_0, stride = new_cache_15_internal_tensor_assign_1_stride_0, x = reshape_70)[name = tensor("slice_by_index_14")]; + tensor reshape_71_shape_0 = const()[name = tensor("reshape_71_shape_0"), val = tensor([-1])]; + tensor reshape_71 = reshape(shape = reshape_71_shape_0, x = slice_by_index_14)[name = tensor("reshape_71")]; + tensor reshape_72_shape_0 = const()[name = tensor("reshape_72_shape_0"), val = tensor([-1])]; + tensor reshape_72_cast_fp16 = reshape(shape = reshape_72_shape_0, x = var_3409_cast_fp16)[name = tensor("reshape_72_cast_fp16")]; + tensor reshape_73_shape_0 = const()[name = tensor("reshape_73_shape_0"), val = tensor([-1])]; + tensor reshape_73_cast_fp16 = reshape(shape = reshape_73_shape_0, x = cache7_to_fp16)[name = tensor("reshape_73_cast_fp16")]; + tensor scatter_14_mode_0 = const()[name = tensor("scatter_14_mode_0"), val = tensor("update")]; + tensor scatter_14_axis_0 = const()[name = tensor("scatter_14_axis_0"), val = tensor(0)]; + tensor scatter_14_validate_indices_0 = const()[name = tensor("scatter_14_validate_indices_0"), val = tensor(false)]; + tensor scatter_14_cast_fp16 = scatter(axis = scatter_14_axis_0, data = reshape_73_cast_fp16, indices = reshape_71, mode = scatter_14_mode_0, updates = reshape_72_cast_fp16, validate_indices = scatter_14_validate_indices_0)[name = tensor("scatter_14_cast_fp16")]; + tensor reshape_74_cast_fp16 = reshape(shape = shape_62, x = scatter_14_cast_fp16)[name = tensor("reshape_74_cast_fp16")]; + tensor var_3417_begin_0 = const()[name = tensor("op_3417_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor var_3417_end_0 = const()[name = tensor("op_3417_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor var_3417_end_mask_0 = const()[name = tensor("op_3417_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_3417_squeeze_mask_0 = const()[name = tensor("op_3417_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor var_3417_cast_fp16 = slice_by_index(begin = var_3417_begin_0, end = var_3417_end_0, end_mask = var_3417_end_mask_0, squeeze_mask = var_3417_squeeze_mask_0, x = reshape_74_cast_fp16)[name = tensor("op_3417_cast_fp16")]; + tensor var_3419_axis_0 = const()[name = tensor("op_3419_axis_0"), val = tensor(1)]; + tensor var_3419_mode_0 = const()[name = tensor("op_3419_mode_0"), val = tensor("update")]; + tensor var_3419_validate_indices_0 = const()[name = tensor("op_3419_validate_indices_0"), val = tensor(false)]; + tensor var_3419_cast_fp16 = scatter_along_axis(axis = var_3419_axis_0, data = var_3417_cast_fp16, indices = write_indices_15, mode = var_3419_mode_0, updates = v_15_cast_fp16, validate_indices = var_3419_validate_indices_0)[name = tensor("op_3419_cast_fp16")]; + tensor concat_53 = const()[name = tensor("concat_53"), val = tensor([1, 0, 0, 0, 0])]; + tensor concat_54 = const()[name = tensor("concat_54"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_15_internal_tensor_assign_2_stride_0 = const()[name = tensor("new_cache_15_internal_tensor_assign_2_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_15_internal_tensor_assign_2_begin_mask_0 = const()[name = tensor("new_cache_15_internal_tensor_assign_2_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_15_internal_tensor_assign_2_end_mask_0 = const()[name = tensor("new_cache_15_internal_tensor_assign_2_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_15_internal_tensor_assign_2_squeeze_mask_0 = const()[name = tensor("new_cache_15_internal_tensor_assign_2_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_63 = const()[name = tensor("shape_63"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_15 = const()[name = tensor("reduce_prod_15"), val = tensor(1048576)]; + tensor range_1d_15_start_0 = const()[name = tensor("range_1d_15_start_0"), val = tensor(0)]; + tensor range_1d_15_step_0 = const()[name = tensor("range_1d_15_step_0"), val = tensor(1)]; + tensor range_1d_15 = range_1d(end = reduce_prod_15, start = range_1d_15_start_0, step = range_1d_15_step_0)[name = tensor("range_1d_15")]; + tensor reshape_75 = reshape(shape = shape_63, x = range_1d_15)[name = tensor("reshape_75")]; + tensor slice_by_index_15 = slice_by_index(begin = concat_53, begin_mask = new_cache_15_internal_tensor_assign_2_begin_mask_0, end = concat_54, end_mask = new_cache_15_internal_tensor_assign_2_end_mask_0, squeeze_mask = new_cache_15_internal_tensor_assign_2_squeeze_mask_0, stride = new_cache_15_internal_tensor_assign_2_stride_0, x = reshape_75)[name = tensor("slice_by_index_15")]; + tensor reshape_76_shape_0 = const()[name = tensor("reshape_76_shape_0"), val = tensor([-1])]; + tensor reshape_76 = reshape(shape = reshape_76_shape_0, x = slice_by_index_15)[name = tensor("reshape_76")]; + tensor reshape_77_shape_0 = const()[name = tensor("reshape_77_shape_0"), val = tensor([-1])]; + tensor reshape_77_cast_fp16 = reshape(shape = reshape_77_shape_0, x = var_3419_cast_fp16)[name = tensor("reshape_77_cast_fp16")]; + tensor reshape_78_shape_0 = const()[name = tensor("reshape_78_shape_0"), val = tensor([-1])]; + tensor reshape_78_cast_fp16 = reshape(shape = reshape_78_shape_0, x = reshape_74_cast_fp16)[name = tensor("reshape_78_cast_fp16")]; + tensor scatter_15_mode_0 = const()[name = tensor("scatter_15_mode_0"), val = tensor("update")]; + tensor scatter_15_axis_0 = const()[name = tensor("scatter_15_axis_0"), val = tensor(0)]; + tensor scatter_15_validate_indices_0 = const()[name = tensor("scatter_15_validate_indices_0"), val = tensor(false)]; + tensor scatter_15_cast_fp16 = scatter(axis = scatter_15_axis_0, data = reshape_78_cast_fp16, indices = reshape_76, mode = scatter_15_mode_0, updates = reshape_77_cast_fp16, validate_indices = scatter_15_validate_indices_0)[name = tensor("scatter_15_cast_fp16")]; + tensor reshape_79_cast_fp16 = reshape(shape = shape_63, x = scatter_15_cast_fp16)[name = tensor("reshape_79_cast_fp16")]; + tensor reshape_79_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("reshape_79_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor keys_43_begin_0 = const()[name = tensor("keys_43_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor keys_43_end_0 = const()[name = tensor("keys_43_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor keys_43_end_mask_0 = const()[name = tensor("keys_43_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor keys_43_squeeze_mask_0 = const()[name = tensor("keys_43_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor keys_43_cast_fp16 = slice_by_index(begin = keys_43_begin_0, end = keys_43_end_0, end_mask = keys_43_end_mask_0, squeeze_mask = keys_43_squeeze_mask_0, x = reshape_79_cast_fp16)[name = tensor("keys_43_cast_fp16")]; + tensor values_43_begin_0 = const()[name = tensor("values_43_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor values_43_end_0 = const()[name = tensor("values_43_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor values_43_end_mask_0 = const()[name = tensor("values_43_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor values_43_squeeze_mask_0 = const()[name = tensor("values_43_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor values_43_cast_fp16 = slice_by_index(begin = values_43_begin_0, end = values_43_end_0, end_mask = values_43_end_mask_0, squeeze_mask = values_43_squeeze_mask_0, x = reshape_79_cast_fp16)[name = tensor("values_43_cast_fp16")]; + tensor var_3431_cast_fp16 = not_equal(x = keys_43_cast_fp16, y = keys_43_cast_fp16)[name = tensor("op_3431_cast_fp16")]; + tensor keys_45_cast_fp16 = select(a = var_504_to_fp16, b = keys_43_cast_fp16, cond = var_3431_cast_fp16)[name = tensor("keys_45_cast_fp16")]; + tensor var_3439_cast_fp16 = not_equal(x = values_43_cast_fp16, y = values_43_cast_fp16)[name = tensor("op_3439_cast_fp16")]; + tensor values_45_cast_fp16 = select(a = var_504_to_fp16, b = values_43_cast_fp16, cond = var_3439_cast_fp16)[name = tensor("values_45_cast_fp16")]; + tensor var_3463 = const()[name = tensor("op_3463"), val = tensor([0, 2, 1, 3])]; + tensor var_3476 = const()[name = tensor("op_3476"), val = tensor([1, 1, 1])]; + tensor var_3477_cast_fp16 = reshape(shape = var_3476, x = position7_to_fp16)[name = tensor("op_3477_cast_fp16")]; + tensor var_3494_to_fp16 = const()[name = tensor("op_3494_to_fp16"), val = tensor(0x1p+0)]; + tensor valid_len_15_cast_fp16 = add(x = var_3477_cast_fp16, y = var_3494_to_fp16)[name = tensor("valid_len_15_cast_fp16")]; + tensor valid_mask_15_cast_fp16 = less(x = k_positions_1_promoted_to_fp16, y = valid_len_15_cast_fp16)[name = tensor("valid_mask_15_cast_fp16")]; + tensor causal_mask_15_cast_fp16 = less_equal(x = k_positions_1_promoted_to_fp16, y = var_3477_cast_fp16)[name = tensor("causal_mask_15_cast_fp16")]; + tensor attn_mask_29 = logical_and(x = valid_mask_15_cast_fp16, y = causal_mask_15_cast_fp16)[name = tensor("attn_mask_29")]; + tensor attn_mask_31_axes_0 = const()[name = tensor("attn_mask_31_axes_0"), val = tensor([1])]; + tensor attn_mask_31 = expand_dims(axes = attn_mask_31_axes_0, x = attn_mask_29)[name = tensor("attn_mask_31")]; + tensor var_3512_transpose_x_0 = const()[name = tensor("op_3512_transpose_x_0"), val = tensor(false)]; + tensor var_3512_transpose_y_0 = const()[name = tensor("op_3512_transpose_y_0"), val = tensor(false)]; + tensor transpose_86_perm_0 = const()[name = tensor("transpose_86_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_87_perm_0 = const()[name = tensor("transpose_87_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_87 = transpose(perm = transpose_87_perm_0, x = keys_45_cast_fp16)[name = tensor("transpose_185")]; + tensor transpose_86 = transpose(perm = transpose_86_perm_0, x = q_45_cast_fp16)[name = tensor("transpose_186")]; + tensor var_3512_cast_fp16 = matmul(transpose_x = var_3512_transpose_x_0, transpose_y = var_3512_transpose_y_0, x = transpose_86, y = transpose_87)[name = tensor("op_3512_cast_fp16")]; + tensor var_3506_to_fp16 = const()[name = tensor("op_3506_to_fp16"), val = tensor([0x1p-3])]; + tensor attn_weights_43_cast_fp16 = mul(x = var_3512_cast_fp16, y = var_3506_to_fp16)[name = tensor("attn_weights_43_cast_fp16")]; + tensor var_3514 = logical_not(x = attn_mask_31)[name = tensor("op_3514")]; + tensor var_3515_to_fp16 = const()[name = tensor("op_3515_to_fp16"), val = tensor(-inf)]; + tensor attn_weights_45_cast_fp16 = select(a = var_3515_to_fp16, b = attn_weights_43_cast_fp16, cond = var_3514)[name = tensor("attn_weights_45_cast_fp16")]; + tensor var_3517 = const()[name = tensor("op_3517"), val = tensor(-1)]; + tensor attn_weights_47_cast_fp16 = softmax(axis = var_3517, x = attn_weights_45_cast_fp16)[name = tensor("attn_weights_47_cast_fp16")]; + tensor attn_output_15_transpose_x_0 = const()[name = tensor("attn_output_15_transpose_x_0"), val = tensor(false)]; + tensor attn_output_15_transpose_y_0 = const()[name = tensor("attn_output_15_transpose_y_0"), val = tensor(false)]; + tensor values_47_cast_fp16 = transpose(perm = var_3463, x = values_45_cast_fp16)[name = tensor("transpose_187")]; + tensor attn_output_15_cast_fp16 = matmul(transpose_x = attn_output_15_transpose_x_0, transpose_y = attn_output_15_transpose_y_0, x = attn_weights_47_cast_fp16, y = values_47_cast_fp16)[name = tensor("attn_output_15_cast_fp16")]; + tensor var_3525 = const()[name = tensor("op_3525"), val = tensor([0, 2, 1, 3])]; + tensor var_3528 = const()[name = tensor("op_3528"), val = tensor([1, 1, 1024])]; + tensor var_3526_cast_fp16 = transpose(perm = var_3525, x = attn_output_15_cast_fp16)[name = tensor("transpose_184")]; + tensor input_75_cast_fp16 = reshape(shape = var_3528, x = var_3526_cast_fp16)[name = tensor("input_75_cast_fp16")]; + tensor linear_30_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = attn7_out_proj_weight_cast_fp16, x = input_75_cast_fp16)[name = tensor("linear_30_cast_fp16")]; + tensor var_3534_to_fp16 = const()[name = tensor("op_3534_to_fp16"), val = tensor(0x1p+0)]; + tensor var_3535_cast_fp16 = add(x = position7_to_fp16, y = var_3534_to_fp16)[name = tensor("op_3535_cast_fp16")]; + tensor var_3535_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_3535_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor input_77_cast_fp16 = add(x = input_73_cast_fp16, y = linear_30_cast_fp16)[name = tensor("input_77_cast_fp16")]; + tensor input_79_axes_0 = const()[name = tensor("input_79_axes_0"), val = tensor([-1])]; + tensor norm7_2_weight_to_fp16 = const()[name = tensor("norm7_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(303649280)))]; + tensor norm7_2_bias_to_fp16 = const()[name = tensor("norm7_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(303651392)))]; + tensor var_3539_to_fp16 = const()[name = tensor("op_3539_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_79_cast_fp16 = layer_norm(axes = input_79_axes_0, beta = norm7_2_bias_to_fp16, epsilon = var_3539_to_fp16, gamma = norm7_2_weight_to_fp16, x = input_77_cast_fp16)[name = tensor("input_79_cast_fp16")]; + tensor linear_31_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = linear7_1_weight_cast_fp16, x = input_79_cast_fp16)[name = tensor("linear_31_cast_fp16")]; + tensor input_81_mode_0 = const()[name = tensor("input_81_mode_0"), val = tensor("EXACT")]; + tensor input_81_cast_fp16 = gelu(mode = input_81_mode_0, x = linear_31_cast_fp16)[name = tensor("input_81_cast_fp16")]; + tensor linear_32_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = linear7_2_weight_cast_fp16, x = input_81_cast_fp16)[name = tensor("linear_32_cast_fp16")]; + tensor input_83_cast_fp16 = add(x = input_77_cast_fp16, y = linear_32_cast_fp16)[name = tensor("input_83_cast_fp16")]; + tensor x_17_axes_0 = const()[name = tensor("x_17_axes_0"), val = tensor([-1])]; + tensor norm8_1_weight_to_fp16 = const()[name = tensor("norm8_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(303653504)))]; + tensor norm8_1_bias_to_fp16 = const()[name = tensor("norm8_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(303655616)))]; + tensor var_3556_to_fp16 = const()[name = tensor("op_3556_to_fp16"), val = tensor(0x1.5p-17)]; + tensor x_17_cast_fp16 = layer_norm(axes = x_17_axes_0, beta = norm8_1_bias_to_fp16, epsilon = var_3556_to_fp16, gamma = norm8_1_weight_to_fp16, x = input_83_cast_fp16)[name = tensor("x_17_cast_fp16")]; + tensor linear_33_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = attn8_in_proj_weight_cast_fp16, x = x_17_cast_fp16)[name = tensor("linear_33_cast_fp16")]; + tensor var_3592 = const()[name = tensor("op_3592"), val = tensor([1, 1, 3, 16, 64])]; + tensor qkv_17_cast_fp16 = reshape(shape = var_3592, x = linear_33_cast_fp16)[name = tensor("qkv_17_cast_fp16")]; + tensor q_49_begin_0 = const()[name = tensor("q_49_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor q_49_end_0 = const()[name = tensor("q_49_end_0"), val = tensor([1, 1, 1, 16, 64])]; + tensor q_49_end_mask_0 = const()[name = tensor("q_49_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor q_49_squeeze_mask_0 = const()[name = tensor("q_49_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor q_49_cast_fp16 = slice_by_index(begin = q_49_begin_0, end = q_49_end_0, end_mask = q_49_end_mask_0, squeeze_mask = q_49_squeeze_mask_0, x = qkv_17_cast_fp16)[name = tensor("q_49_cast_fp16")]; + tensor k_33_begin_0 = const()[name = tensor("k_33_begin_0"), val = tensor([0, 0, 1, 0, 0])]; + tensor k_33_end_0 = const()[name = tensor("k_33_end_0"), val = tensor([1, 1, 2, 16, 64])]; + tensor k_33_end_mask_0 = const()[name = tensor("k_33_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor k_33_squeeze_mask_0 = const()[name = tensor("k_33_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor k_33_cast_fp16 = slice_by_index(begin = k_33_begin_0, end = k_33_end_0, end_mask = k_33_end_mask_0, squeeze_mask = k_33_squeeze_mask_0, x = qkv_17_cast_fp16)[name = tensor("k_33_cast_fp16")]; + tensor v_17_begin_0 = const()[name = tensor("v_17_begin_0"), val = tensor([0, 0, 2, 0, 0])]; + tensor v_17_end_0 = const()[name = tensor("v_17_end_0"), val = tensor([1, 1, 3, 16, 64])]; + tensor v_17_end_mask_0 = const()[name = tensor("v_17_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor v_17_squeeze_mask_0 = const()[name = tensor("v_17_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor v_17_cast_fp16 = slice_by_index(begin = v_17_begin_0, end = v_17_end_0, end_mask = v_17_end_mask_0, squeeze_mask = v_17_squeeze_mask_0, x = qkv_17_cast_fp16)[name = tensor("v_17_cast_fp16")]; + tensor var_3696 = const()[name = tensor("op_3696"), val = tensor([1, 1, 1, 1])]; + tensor position8_to_fp16_dtype_0 = const()[name = tensor("position8_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor position8_to_fp16 = cast(dtype = position8_to_fp16_dtype_0, x = position8)[name = tensor("cast_513")]; + tensor ts_53_cast_fp16 = reshape(shape = var_3696, x = position8_to_fp16)[name = tensor("ts_53_cast_fp16")]; + tensor var_3700 = const()[name = tensor("op_3700"), val = tensor([1, 1, 16, 32, 2])]; + tensor q_complex_17_cast_fp16 = reshape(shape = var_3700, x = q_49_cast_fp16)[name = tensor("q_complex_17_cast_fp16")]; + tensor var_3704 = const()[name = tensor("op_3704"), val = tensor([1, 1, 16, 32, 2])]; + tensor k_complex_17_cast_fp16 = reshape(shape = var_3704, x = k_33_cast_fp16)[name = tensor("k_complex_17_cast_fp16")]; + tensor var_3708_begin_0 = const()[name = tensor("op_3708_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_3708_end_0 = const()[name = tensor("op_3708_end_0"), val = tensor([1, 1, 16, 32, 1])]; + tensor var_3708_end_mask_0 = const()[name = tensor("op_3708_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_3708_squeeze_mask_0 = const()[name = tensor("op_3708_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_3708_cast_fp16 = slice_by_index(begin = var_3708_begin_0, end = var_3708_end_0, end_mask = var_3708_end_mask_0, squeeze_mask = var_3708_squeeze_mask_0, x = q_complex_17_cast_fp16)[name = tensor("op_3708_cast_fp16")]; + tensor var_3716_begin_0 = const()[name = tensor("op_3716_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_3716_end_0 = const()[name = tensor("op_3716_end_0"), val = tensor([1, 1, 16, 32, 2])]; + tensor var_3716_end_mask_0 = const()[name = tensor("op_3716_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_3716_squeeze_mask_0 = const()[name = tensor("op_3716_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_3716_cast_fp16 = slice_by_index(begin = var_3716_begin_0, end = var_3716_end_0, end_mask = var_3716_end_mask_0, squeeze_mask = var_3716_squeeze_mask_0, x = q_complex_17_cast_fp16)[name = tensor("op_3716_cast_fp16")]; + tensor var_3724_begin_0 = const()[name = tensor("op_3724_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_3724_end_0 = const()[name = tensor("op_3724_end_0"), val = tensor([1, 1, 16, 32, 1])]; + tensor var_3724_end_mask_0 = const()[name = tensor("op_3724_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_3724_squeeze_mask_0 = const()[name = tensor("op_3724_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_3724_cast_fp16 = slice_by_index(begin = var_3724_begin_0, end = var_3724_end_0, end_mask = var_3724_end_mask_0, squeeze_mask = var_3724_squeeze_mask_0, x = k_complex_17_cast_fp16)[name = tensor("op_3724_cast_fp16")]; + tensor var_3732_begin_0 = const()[name = tensor("op_3732_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_3732_end_0 = const()[name = tensor("op_3732_end_0"), val = tensor([1, 1, 16, 32, 2])]; + tensor var_3732_end_mask_0 = const()[name = tensor("op_3732_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_3732_squeeze_mask_0 = const()[name = tensor("op_3732_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_3732_cast_fp16 = slice_by_index(begin = var_3732_begin_0, end = var_3732_end_0, end_mask = var_3732_end_mask_0, squeeze_mask = var_3732_squeeze_mask_0, x = k_complex_17_cast_fp16)[name = tensor("op_3732_cast_fp16")]; + tensor freqs_17_to_fp16 = const()[name = tensor("freqs_17_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(303657728)))]; + tensor var_3738_cast_fp16 = mul(x = freqs_17_to_fp16, y = ts_53_cast_fp16)[name = tensor("op_3738_cast_fp16")]; + tensor rotr_17_cast_fp16 = cos(x = var_3738_cast_fp16)[name = tensor("rotr_17_cast_fp16")]; + tensor roti_17_cast_fp16 = sin(x = var_3738_cast_fp16)[name = tensor("roti_17_cast_fp16")]; + tensor var_3742_cast_fp16 = mul(x = var_3708_cast_fp16, y = rotr_17_cast_fp16)[name = tensor("op_3742_cast_fp16")]; + tensor var_3743_cast_fp16 = mul(x = var_3716_cast_fp16, y = roti_17_cast_fp16)[name = tensor("op_3743_cast_fp16")]; + tensor qor_33_cast_fp16 = sub(x = var_3742_cast_fp16, y = var_3743_cast_fp16)[name = tensor("qor_33_cast_fp16")]; + tensor var_3746_cast_fp16 = mul(x = var_3708_cast_fp16, y = roti_17_cast_fp16)[name = tensor("op_3746_cast_fp16")]; + tensor var_3747_cast_fp16 = mul(x = var_3716_cast_fp16, y = rotr_17_cast_fp16)[name = tensor("op_3747_cast_fp16")]; + tensor qoi_33_cast_fp16 = add(x = var_3746_cast_fp16, y = var_3747_cast_fp16)[name = tensor("qoi_33_cast_fp16")]; + tensor var_3750_cast_fp16 = mul(x = var_3724_cast_fp16, y = rotr_17_cast_fp16)[name = tensor("op_3750_cast_fp16")]; + tensor var_3751_cast_fp16 = mul(x = var_3732_cast_fp16, y = roti_17_cast_fp16)[name = tensor("op_3751_cast_fp16")]; + tensor kor_33_cast_fp16 = sub(x = var_3750_cast_fp16, y = var_3751_cast_fp16)[name = tensor("kor_33_cast_fp16")]; + tensor var_3754_cast_fp16 = mul(x = var_3724_cast_fp16, y = roti_17_cast_fp16)[name = tensor("op_3754_cast_fp16")]; + tensor var_3755_cast_fp16 = mul(x = var_3732_cast_fp16, y = rotr_17_cast_fp16)[name = tensor("op_3755_cast_fp16")]; + tensor koi_33_cast_fp16 = add(x = var_3754_cast_fp16, y = var_3755_cast_fp16)[name = tensor("koi_33_cast_fp16")]; + tensor qo_17_axis_0 = const()[name = tensor("qo_17_axis_0"), val = tensor(-1)]; + tensor qo_17_cast_fp16 = stack(axis = qo_17_axis_0, values = (qor_33_cast_fp16, qoi_33_cast_fp16))[name = tensor("qo_17_cast_fp16")]; + tensor ko_17_axis_0 = const()[name = tensor("ko_17_axis_0"), val = tensor(-1)]; + tensor ko_17_cast_fp16 = stack(axis = ko_17_axis_0, values = (kor_33_cast_fp16, koi_33_cast_fp16))[name = tensor("ko_17_cast_fp16")]; + tensor var_3784 = const()[name = tensor("op_3784"), val = tensor([1, 1, 16, 64])]; + tensor q_51_cast_fp16 = reshape(shape = var_3784, x = qo_17_cast_fp16)[name = tensor("q_51_cast_fp16")]; + tensor var_3786 = const()[name = tensor("op_3786"), val = tensor([1, 1, 16, 64])]; + tensor k_35_cast_fp16 = reshape(shape = var_3786, x = ko_17_cast_fp16)[name = tensor("k_35_cast_fp16")]; + tensor _inversed_3808_y_0_to_fp16 = const()[name = tensor("_inversed_3808_y_0_to_fp16"), val = tensor(0x1p-9)]; + tensor _inversed_3808_cast_fp16 = mul(x = ts_53_cast_fp16, y = _inversed_3808_y_0_to_fp16)[name = tensor("_inversed_3808_cast_fp16")]; + tensor var_3809_cast_fp16 = floor(x = _inversed_3808_cast_fp16)[name = tensor("op_3809_cast_fp16")]; + tensor var_3810_to_fp16 = const()[name = tensor("op_3810_to_fp16"), val = tensor(0x1p+9)]; + tensor var_3811_cast_fp16 = mul(x = var_3809_cast_fp16, y = var_3810_to_fp16)[name = tensor("op_3811_cast_fp16")]; + tensor write_indices_float_35_cast_fp16 = sub(x = ts_53_cast_fp16, y = var_3811_cast_fp16)[name = tensor("write_indices_float_35_cast_fp16")]; + tensor var_3818_dtype_0 = const()[name = tensor("op_3818_dtype_0"), val = tensor("int32")]; + tensor write_indices_17_reps_0 = const()[name = tensor("write_indices_17_reps_0"), val = tensor([1, 1, 16, 64])]; + tensor write_indices_float_35_cast_fp16_to_int32 = cast(dtype = var_3818_dtype_0, x = write_indices_float_35_cast_fp16)[name = tensor("cast_512")]; + tensor write_indices_17 = tile(reps = write_indices_17_reps_0, x = write_indices_float_35_cast_fp16_to_int32)[name = tensor("write_indices_17")]; + tensor var_3826_begin_0 = const()[name = tensor("op_3826_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_3826_end_0 = const()[name = tensor("op_3826_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor var_3826_end_mask_0 = const()[name = tensor("op_3826_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_3826_squeeze_mask_0 = const()[name = tensor("op_3826_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor cache8_to_fp16_dtype_0 = const()[name = tensor("cache8_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor cache8_to_fp16 = cast(dtype = cache8_to_fp16_dtype_0, x = cache8)[name = tensor("cast_511")]; + tensor var_3826_cast_fp16 = slice_by_index(begin = var_3826_begin_0, end = var_3826_end_0, end_mask = var_3826_end_mask_0, squeeze_mask = var_3826_squeeze_mask_0, x = cache8_to_fp16)[name = tensor("op_3826_cast_fp16")]; + tensor var_3828_axis_0 = const()[name = tensor("op_3828_axis_0"), val = tensor(1)]; + tensor var_3828_mode_0 = const()[name = tensor("op_3828_mode_0"), val = tensor("update")]; + tensor var_3828_validate_indices_0 = const()[name = tensor("op_3828_validate_indices_0"), val = tensor(false)]; + tensor var_3828_cast_fp16 = scatter_along_axis(axis = var_3828_axis_0, data = var_3826_cast_fp16, indices = write_indices_17, mode = var_3828_mode_0, updates = k_35_cast_fp16, validate_indices = var_3828_validate_indices_0)[name = tensor("op_3828_cast_fp16")]; + tensor concat_58 = const()[name = tensor("concat_58"), val = tensor([0, 0, 0, 0, 0])]; + tensor concat_59 = const()[name = tensor("concat_59"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_17_internal_tensor_assign_1_stride_0 = const()[name = tensor("new_cache_17_internal_tensor_assign_1_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_17_internal_tensor_assign_1_begin_mask_0 = const()[name = tensor("new_cache_17_internal_tensor_assign_1_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_17_internal_tensor_assign_1_end_mask_0 = const()[name = tensor("new_cache_17_internal_tensor_assign_1_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_17_internal_tensor_assign_1_squeeze_mask_0 = const()[name = tensor("new_cache_17_internal_tensor_assign_1_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_64 = const()[name = tensor("shape_64"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_16 = const()[name = tensor("reduce_prod_16"), val = tensor(1048576)]; + tensor range_1d_16_start_0 = const()[name = tensor("range_1d_16_start_0"), val = tensor(0)]; + tensor range_1d_16_step_0 = const()[name = tensor("range_1d_16_step_0"), val = tensor(1)]; + tensor range_1d_16 = range_1d(end = reduce_prod_16, start = range_1d_16_start_0, step = range_1d_16_step_0)[name = tensor("range_1d_16")]; + tensor reshape_80 = reshape(shape = shape_64, x = range_1d_16)[name = tensor("reshape_80")]; + tensor slice_by_index_16 = slice_by_index(begin = concat_58, begin_mask = new_cache_17_internal_tensor_assign_1_begin_mask_0, end = concat_59, end_mask = new_cache_17_internal_tensor_assign_1_end_mask_0, squeeze_mask = new_cache_17_internal_tensor_assign_1_squeeze_mask_0, stride = new_cache_17_internal_tensor_assign_1_stride_0, x = reshape_80)[name = tensor("slice_by_index_16")]; + tensor reshape_81_shape_0 = const()[name = tensor("reshape_81_shape_0"), val = tensor([-1])]; + tensor reshape_81 = reshape(shape = reshape_81_shape_0, x = slice_by_index_16)[name = tensor("reshape_81")]; + tensor reshape_82_shape_0 = const()[name = tensor("reshape_82_shape_0"), val = tensor([-1])]; + tensor reshape_82_cast_fp16 = reshape(shape = reshape_82_shape_0, x = var_3828_cast_fp16)[name = tensor("reshape_82_cast_fp16")]; + tensor reshape_83_shape_0 = const()[name = tensor("reshape_83_shape_0"), val = tensor([-1])]; + tensor reshape_83_cast_fp16 = reshape(shape = reshape_83_shape_0, x = cache8_to_fp16)[name = tensor("reshape_83_cast_fp16")]; + tensor scatter_16_mode_0 = const()[name = tensor("scatter_16_mode_0"), val = tensor("update")]; + tensor scatter_16_axis_0 = const()[name = tensor("scatter_16_axis_0"), val = tensor(0)]; + tensor scatter_16_validate_indices_0 = const()[name = tensor("scatter_16_validate_indices_0"), val = tensor(false)]; + tensor scatter_16_cast_fp16 = scatter(axis = scatter_16_axis_0, data = reshape_83_cast_fp16, indices = reshape_81, mode = scatter_16_mode_0, updates = reshape_82_cast_fp16, validate_indices = scatter_16_validate_indices_0)[name = tensor("scatter_16_cast_fp16")]; + tensor reshape_84_cast_fp16 = reshape(shape = shape_64, x = scatter_16_cast_fp16)[name = tensor("reshape_84_cast_fp16")]; + tensor var_3836_begin_0 = const()[name = tensor("op_3836_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor var_3836_end_0 = const()[name = tensor("op_3836_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor var_3836_end_mask_0 = const()[name = tensor("op_3836_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_3836_squeeze_mask_0 = const()[name = tensor("op_3836_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor var_3836_cast_fp16 = slice_by_index(begin = var_3836_begin_0, end = var_3836_end_0, end_mask = var_3836_end_mask_0, squeeze_mask = var_3836_squeeze_mask_0, x = reshape_84_cast_fp16)[name = tensor("op_3836_cast_fp16")]; + tensor var_3838_axis_0 = const()[name = tensor("op_3838_axis_0"), val = tensor(1)]; + tensor var_3838_mode_0 = const()[name = tensor("op_3838_mode_0"), val = tensor("update")]; + tensor var_3838_validate_indices_0 = const()[name = tensor("op_3838_validate_indices_0"), val = tensor(false)]; + tensor var_3838_cast_fp16 = scatter_along_axis(axis = var_3838_axis_0, data = var_3836_cast_fp16, indices = write_indices_17, mode = var_3838_mode_0, updates = v_17_cast_fp16, validate_indices = var_3838_validate_indices_0)[name = tensor("op_3838_cast_fp16")]; + tensor concat_60 = const()[name = tensor("concat_60"), val = tensor([1, 0, 0, 0, 0])]; + tensor concat_61 = const()[name = tensor("concat_61"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_17_internal_tensor_assign_2_stride_0 = const()[name = tensor("new_cache_17_internal_tensor_assign_2_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_17_internal_tensor_assign_2_begin_mask_0 = const()[name = tensor("new_cache_17_internal_tensor_assign_2_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_17_internal_tensor_assign_2_end_mask_0 = const()[name = tensor("new_cache_17_internal_tensor_assign_2_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_17_internal_tensor_assign_2_squeeze_mask_0 = const()[name = tensor("new_cache_17_internal_tensor_assign_2_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_65 = const()[name = tensor("shape_65"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_17 = const()[name = tensor("reduce_prod_17"), val = tensor(1048576)]; + tensor range_1d_17_start_0 = const()[name = tensor("range_1d_17_start_0"), val = tensor(0)]; + tensor range_1d_17_step_0 = const()[name = tensor("range_1d_17_step_0"), val = tensor(1)]; + tensor range_1d_17 = range_1d(end = reduce_prod_17, start = range_1d_17_start_0, step = range_1d_17_step_0)[name = tensor("range_1d_17")]; + tensor reshape_85 = reshape(shape = shape_65, x = range_1d_17)[name = tensor("reshape_85")]; + tensor slice_by_index_17 = slice_by_index(begin = concat_60, begin_mask = new_cache_17_internal_tensor_assign_2_begin_mask_0, end = concat_61, end_mask = new_cache_17_internal_tensor_assign_2_end_mask_0, squeeze_mask = new_cache_17_internal_tensor_assign_2_squeeze_mask_0, stride = new_cache_17_internal_tensor_assign_2_stride_0, x = reshape_85)[name = tensor("slice_by_index_17")]; + tensor reshape_86_shape_0 = const()[name = tensor("reshape_86_shape_0"), val = tensor([-1])]; + tensor reshape_86 = reshape(shape = reshape_86_shape_0, x = slice_by_index_17)[name = tensor("reshape_86")]; + tensor reshape_87_shape_0 = const()[name = tensor("reshape_87_shape_0"), val = tensor([-1])]; + tensor reshape_87_cast_fp16 = reshape(shape = reshape_87_shape_0, x = var_3838_cast_fp16)[name = tensor("reshape_87_cast_fp16")]; + tensor reshape_88_shape_0 = const()[name = tensor("reshape_88_shape_0"), val = tensor([-1])]; + tensor reshape_88_cast_fp16 = reshape(shape = reshape_88_shape_0, x = reshape_84_cast_fp16)[name = tensor("reshape_88_cast_fp16")]; + tensor scatter_17_mode_0 = const()[name = tensor("scatter_17_mode_0"), val = tensor("update")]; + tensor scatter_17_axis_0 = const()[name = tensor("scatter_17_axis_0"), val = tensor(0)]; + tensor scatter_17_validate_indices_0 = const()[name = tensor("scatter_17_validate_indices_0"), val = tensor(false)]; + tensor scatter_17_cast_fp16 = scatter(axis = scatter_17_axis_0, data = reshape_88_cast_fp16, indices = reshape_86, mode = scatter_17_mode_0, updates = reshape_87_cast_fp16, validate_indices = scatter_17_validate_indices_0)[name = tensor("scatter_17_cast_fp16")]; + tensor reshape_89_cast_fp16 = reshape(shape = shape_65, x = scatter_17_cast_fp16)[name = tensor("reshape_89_cast_fp16")]; + tensor reshape_89_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("reshape_89_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor keys_49_begin_0 = const()[name = tensor("keys_49_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor keys_49_end_0 = const()[name = tensor("keys_49_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor keys_49_end_mask_0 = const()[name = tensor("keys_49_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor keys_49_squeeze_mask_0 = const()[name = tensor("keys_49_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor keys_49_cast_fp16 = slice_by_index(begin = keys_49_begin_0, end = keys_49_end_0, end_mask = keys_49_end_mask_0, squeeze_mask = keys_49_squeeze_mask_0, x = reshape_89_cast_fp16)[name = tensor("keys_49_cast_fp16")]; + tensor values_49_begin_0 = const()[name = tensor("values_49_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor values_49_end_0 = const()[name = tensor("values_49_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor values_49_end_mask_0 = const()[name = tensor("values_49_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor values_49_squeeze_mask_0 = const()[name = tensor("values_49_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor values_49_cast_fp16 = slice_by_index(begin = values_49_begin_0, end = values_49_end_0, end_mask = values_49_end_mask_0, squeeze_mask = values_49_squeeze_mask_0, x = reshape_89_cast_fp16)[name = tensor("values_49_cast_fp16")]; + tensor var_3850_cast_fp16 = not_equal(x = keys_49_cast_fp16, y = keys_49_cast_fp16)[name = tensor("op_3850_cast_fp16")]; + tensor keys_51_cast_fp16 = select(a = var_504_to_fp16, b = keys_49_cast_fp16, cond = var_3850_cast_fp16)[name = tensor("keys_51_cast_fp16")]; + tensor var_3858_cast_fp16 = not_equal(x = values_49_cast_fp16, y = values_49_cast_fp16)[name = tensor("op_3858_cast_fp16")]; + tensor values_51_cast_fp16 = select(a = var_504_to_fp16, b = values_49_cast_fp16, cond = var_3858_cast_fp16)[name = tensor("values_51_cast_fp16")]; + tensor var_3882 = const()[name = tensor("op_3882"), val = tensor([0, 2, 1, 3])]; + tensor var_3895 = const()[name = tensor("op_3895"), val = tensor([1, 1, 1])]; + tensor var_3896_cast_fp16 = reshape(shape = var_3895, x = position8_to_fp16)[name = tensor("op_3896_cast_fp16")]; + tensor var_3913_to_fp16 = const()[name = tensor("op_3913_to_fp16"), val = tensor(0x1p+0)]; + tensor valid_len_17_cast_fp16 = add(x = var_3896_cast_fp16, y = var_3913_to_fp16)[name = tensor("valid_len_17_cast_fp16")]; + tensor valid_mask_17_cast_fp16 = less(x = k_positions_1_promoted_to_fp16, y = valid_len_17_cast_fp16)[name = tensor("valid_mask_17_cast_fp16")]; + tensor causal_mask_17_cast_fp16 = less_equal(x = k_positions_1_promoted_to_fp16, y = var_3896_cast_fp16)[name = tensor("causal_mask_17_cast_fp16")]; + tensor attn_mask_33 = logical_and(x = valid_mask_17_cast_fp16, y = causal_mask_17_cast_fp16)[name = tensor("attn_mask_33")]; + tensor attn_mask_35_axes_0 = const()[name = tensor("attn_mask_35_axes_0"), val = tensor([1])]; + tensor attn_mask_35 = expand_dims(axes = attn_mask_35_axes_0, x = attn_mask_33)[name = tensor("attn_mask_35")]; + tensor var_3931_transpose_x_0 = const()[name = tensor("op_3931_transpose_x_0"), val = tensor(false)]; + tensor var_3931_transpose_y_0 = const()[name = tensor("op_3931_transpose_y_0"), val = tensor(false)]; + tensor transpose_88_perm_0 = const()[name = tensor("transpose_88_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_89_perm_0 = const()[name = tensor("transpose_89_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_89 = transpose(perm = transpose_89_perm_0, x = keys_51_cast_fp16)[name = tensor("transpose_181")]; + tensor transpose_88 = transpose(perm = transpose_88_perm_0, x = q_51_cast_fp16)[name = tensor("transpose_182")]; + tensor var_3931_cast_fp16 = matmul(transpose_x = var_3931_transpose_x_0, transpose_y = var_3931_transpose_y_0, x = transpose_88, y = transpose_89)[name = tensor("op_3931_cast_fp16")]; + tensor var_3925_to_fp16 = const()[name = tensor("op_3925_to_fp16"), val = tensor([0x1p-3])]; + tensor attn_weights_49_cast_fp16 = mul(x = var_3931_cast_fp16, y = var_3925_to_fp16)[name = tensor("attn_weights_49_cast_fp16")]; + tensor var_3933 = logical_not(x = attn_mask_35)[name = tensor("op_3933")]; + tensor var_3934_to_fp16 = const()[name = tensor("op_3934_to_fp16"), val = tensor(-inf)]; + tensor attn_weights_51_cast_fp16 = select(a = var_3934_to_fp16, b = attn_weights_49_cast_fp16, cond = var_3933)[name = tensor("attn_weights_51_cast_fp16")]; + tensor var_3936 = const()[name = tensor("op_3936"), val = tensor(-1)]; + tensor attn_weights_53_cast_fp16 = softmax(axis = var_3936, x = attn_weights_51_cast_fp16)[name = tensor("attn_weights_53_cast_fp16")]; + tensor attn_output_17_transpose_x_0 = const()[name = tensor("attn_output_17_transpose_x_0"), val = tensor(false)]; + tensor attn_output_17_transpose_y_0 = const()[name = tensor("attn_output_17_transpose_y_0"), val = tensor(false)]; + tensor values_53_cast_fp16 = transpose(perm = var_3882, x = values_51_cast_fp16)[name = tensor("transpose_183")]; + tensor attn_output_17_cast_fp16 = matmul(transpose_x = attn_output_17_transpose_x_0, transpose_y = attn_output_17_transpose_y_0, x = attn_weights_53_cast_fp16, y = values_53_cast_fp16)[name = tensor("attn_output_17_cast_fp16")]; + tensor var_3944 = const()[name = tensor("op_3944"), val = tensor([0, 2, 1, 3])]; + tensor var_3947 = const()[name = tensor("op_3947"), val = tensor([1, 1, 1024])]; + tensor var_3945_cast_fp16 = transpose(perm = var_3944, x = attn_output_17_cast_fp16)[name = tensor("transpose_180")]; + tensor input_85_cast_fp16 = reshape(shape = var_3947, x = var_3945_cast_fp16)[name = tensor("input_85_cast_fp16")]; + tensor linear_34_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = attn8_out_proj_weight_cast_fp16, x = input_85_cast_fp16)[name = tensor("linear_34_cast_fp16")]; + tensor var_3953_to_fp16 = const()[name = tensor("op_3953_to_fp16"), val = tensor(0x1p+0)]; + tensor var_3954_cast_fp16 = add(x = position8_to_fp16, y = var_3953_to_fp16)[name = tensor("op_3954_cast_fp16")]; + tensor var_3954_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_3954_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor input_87_cast_fp16 = add(x = input_83_cast_fp16, y = linear_34_cast_fp16)[name = tensor("input_87_cast_fp16")]; + tensor input_89_axes_0 = const()[name = tensor("input_89_axes_0"), val = tensor([-1])]; + tensor norm8_2_weight_to_fp16 = const()[name = tensor("norm8_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(303657856)))]; + tensor norm8_2_bias_to_fp16 = const()[name = tensor("norm8_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(303659968)))]; + tensor var_3958_to_fp16 = const()[name = tensor("op_3958_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_89_cast_fp16 = layer_norm(axes = input_89_axes_0, beta = norm8_2_bias_to_fp16, epsilon = var_3958_to_fp16, gamma = norm8_2_weight_to_fp16, x = input_87_cast_fp16)[name = tensor("input_89_cast_fp16")]; + tensor linear_35_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = linear8_1_weight_cast_fp16, x = input_89_cast_fp16)[name = tensor("linear_35_cast_fp16")]; + tensor input_91_mode_0 = const()[name = tensor("input_91_mode_0"), val = tensor("EXACT")]; + tensor input_91_cast_fp16 = gelu(mode = input_91_mode_0, x = linear_35_cast_fp16)[name = tensor("input_91_cast_fp16")]; + tensor linear_36_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = linear8_2_weight_cast_fp16, x = input_91_cast_fp16)[name = tensor("linear_36_cast_fp16")]; + tensor input_93_cast_fp16 = add(x = input_87_cast_fp16, y = linear_36_cast_fp16)[name = tensor("input_93_cast_fp16")]; + tensor x_19_axes_0 = const()[name = tensor("x_19_axes_0"), val = tensor([-1])]; + tensor norm9_1_weight_to_fp16 = const()[name = tensor("norm9_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(303662080)))]; + tensor norm9_1_bias_to_fp16 = const()[name = tensor("norm9_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(303664192)))]; + tensor var_3975_to_fp16 = const()[name = tensor("op_3975_to_fp16"), val = tensor(0x1.5p-17)]; + tensor x_19_cast_fp16 = layer_norm(axes = x_19_axes_0, beta = norm9_1_bias_to_fp16, epsilon = var_3975_to_fp16, gamma = norm9_1_weight_to_fp16, x = input_93_cast_fp16)[name = tensor("x_19_cast_fp16")]; + tensor linear_37_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = attn9_in_proj_weight_cast_fp16, x = x_19_cast_fp16)[name = tensor("linear_37_cast_fp16")]; + tensor var_4011 = const()[name = tensor("op_4011"), val = tensor([1, 1, 3, 16, 64])]; + tensor qkv_19_cast_fp16 = reshape(shape = var_4011, x = linear_37_cast_fp16)[name = tensor("qkv_19_cast_fp16")]; + tensor q_55_begin_0 = const()[name = tensor("q_55_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor q_55_end_0 = const()[name = tensor("q_55_end_0"), val = tensor([1, 1, 1, 16, 64])]; + tensor q_55_end_mask_0 = const()[name = tensor("q_55_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor q_55_squeeze_mask_0 = const()[name = tensor("q_55_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor q_55_cast_fp16 = slice_by_index(begin = q_55_begin_0, end = q_55_end_0, end_mask = q_55_end_mask_0, squeeze_mask = q_55_squeeze_mask_0, x = qkv_19_cast_fp16)[name = tensor("q_55_cast_fp16")]; + tensor k_37_begin_0 = const()[name = tensor("k_37_begin_0"), val = tensor([0, 0, 1, 0, 0])]; + tensor k_37_end_0 = const()[name = tensor("k_37_end_0"), val = tensor([1, 1, 2, 16, 64])]; + tensor k_37_end_mask_0 = const()[name = tensor("k_37_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor k_37_squeeze_mask_0 = const()[name = tensor("k_37_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor k_37_cast_fp16 = slice_by_index(begin = k_37_begin_0, end = k_37_end_0, end_mask = k_37_end_mask_0, squeeze_mask = k_37_squeeze_mask_0, x = qkv_19_cast_fp16)[name = tensor("k_37_cast_fp16")]; + tensor v_19_begin_0 = const()[name = tensor("v_19_begin_0"), val = tensor([0, 0, 2, 0, 0])]; + tensor v_19_end_0 = const()[name = tensor("v_19_end_0"), val = tensor([1, 1, 3, 16, 64])]; + tensor v_19_end_mask_0 = const()[name = tensor("v_19_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor v_19_squeeze_mask_0 = const()[name = tensor("v_19_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor v_19_cast_fp16 = slice_by_index(begin = v_19_begin_0, end = v_19_end_0, end_mask = v_19_end_mask_0, squeeze_mask = v_19_squeeze_mask_0, x = qkv_19_cast_fp16)[name = tensor("v_19_cast_fp16")]; + tensor var_4115 = const()[name = tensor("op_4115"), val = tensor([1, 1, 1, 1])]; + tensor position9_to_fp16_dtype_0 = const()[name = tensor("position9_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor position9_to_fp16 = cast(dtype = position9_to_fp16_dtype_0, x = position9)[name = tensor("cast_508")]; + tensor ts_59_cast_fp16 = reshape(shape = var_4115, x = position9_to_fp16)[name = tensor("ts_59_cast_fp16")]; + tensor var_4119 = const()[name = tensor("op_4119"), val = tensor([1, 1, 16, 32, 2])]; + tensor q_complex_19_cast_fp16 = reshape(shape = var_4119, x = q_55_cast_fp16)[name = tensor("q_complex_19_cast_fp16")]; + tensor var_4123 = const()[name = tensor("op_4123"), val = tensor([1, 1, 16, 32, 2])]; + tensor k_complex_19_cast_fp16 = reshape(shape = var_4123, x = k_37_cast_fp16)[name = tensor("k_complex_19_cast_fp16")]; + tensor var_4127_begin_0 = const()[name = tensor("op_4127_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_4127_end_0 = const()[name = tensor("op_4127_end_0"), val = tensor([1, 1, 16, 32, 1])]; + tensor var_4127_end_mask_0 = const()[name = tensor("op_4127_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_4127_squeeze_mask_0 = const()[name = tensor("op_4127_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_4127_cast_fp16 = slice_by_index(begin = var_4127_begin_0, end = var_4127_end_0, end_mask = var_4127_end_mask_0, squeeze_mask = var_4127_squeeze_mask_0, x = q_complex_19_cast_fp16)[name = tensor("op_4127_cast_fp16")]; + tensor var_4135_begin_0 = const()[name = tensor("op_4135_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_4135_end_0 = const()[name = tensor("op_4135_end_0"), val = tensor([1, 1, 16, 32, 2])]; + tensor var_4135_end_mask_0 = const()[name = tensor("op_4135_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_4135_squeeze_mask_0 = const()[name = tensor("op_4135_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_4135_cast_fp16 = slice_by_index(begin = var_4135_begin_0, end = var_4135_end_0, end_mask = var_4135_end_mask_0, squeeze_mask = var_4135_squeeze_mask_0, x = q_complex_19_cast_fp16)[name = tensor("op_4135_cast_fp16")]; + tensor var_4143_begin_0 = const()[name = tensor("op_4143_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_4143_end_0 = const()[name = tensor("op_4143_end_0"), val = tensor([1, 1, 16, 32, 1])]; + tensor var_4143_end_mask_0 = const()[name = tensor("op_4143_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_4143_squeeze_mask_0 = const()[name = tensor("op_4143_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_4143_cast_fp16 = slice_by_index(begin = var_4143_begin_0, end = var_4143_end_0, end_mask = var_4143_end_mask_0, squeeze_mask = var_4143_squeeze_mask_0, x = k_complex_19_cast_fp16)[name = tensor("op_4143_cast_fp16")]; + tensor var_4151_begin_0 = const()[name = tensor("op_4151_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_4151_end_0 = const()[name = tensor("op_4151_end_0"), val = tensor([1, 1, 16, 32, 2])]; + tensor var_4151_end_mask_0 = const()[name = tensor("op_4151_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_4151_squeeze_mask_0 = const()[name = tensor("op_4151_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_4151_cast_fp16 = slice_by_index(begin = var_4151_begin_0, end = var_4151_end_0, end_mask = var_4151_end_mask_0, squeeze_mask = var_4151_squeeze_mask_0, x = k_complex_19_cast_fp16)[name = tensor("op_4151_cast_fp16")]; + tensor freqs_19_to_fp16 = const()[name = tensor("freqs_19_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(303666304)))]; + tensor var_4157_cast_fp16 = mul(x = freqs_19_to_fp16, y = ts_59_cast_fp16)[name = tensor("op_4157_cast_fp16")]; + tensor rotr_19_cast_fp16 = cos(x = var_4157_cast_fp16)[name = tensor("rotr_19_cast_fp16")]; + tensor roti_19_cast_fp16 = sin(x = var_4157_cast_fp16)[name = tensor("roti_19_cast_fp16")]; + tensor var_4161_cast_fp16 = mul(x = var_4127_cast_fp16, y = rotr_19_cast_fp16)[name = tensor("op_4161_cast_fp16")]; + tensor var_4162_cast_fp16 = mul(x = var_4135_cast_fp16, y = roti_19_cast_fp16)[name = tensor("op_4162_cast_fp16")]; + tensor qor_37_cast_fp16 = sub(x = var_4161_cast_fp16, y = var_4162_cast_fp16)[name = tensor("qor_37_cast_fp16")]; + tensor var_4165_cast_fp16 = mul(x = var_4127_cast_fp16, y = roti_19_cast_fp16)[name = tensor("op_4165_cast_fp16")]; + tensor var_4166_cast_fp16 = mul(x = var_4135_cast_fp16, y = rotr_19_cast_fp16)[name = tensor("op_4166_cast_fp16")]; + tensor qoi_37_cast_fp16 = add(x = var_4165_cast_fp16, y = var_4166_cast_fp16)[name = tensor("qoi_37_cast_fp16")]; + tensor var_4169_cast_fp16 = mul(x = var_4143_cast_fp16, y = rotr_19_cast_fp16)[name = tensor("op_4169_cast_fp16")]; + tensor var_4170_cast_fp16 = mul(x = var_4151_cast_fp16, y = roti_19_cast_fp16)[name = tensor("op_4170_cast_fp16")]; + tensor kor_37_cast_fp16 = sub(x = var_4169_cast_fp16, y = var_4170_cast_fp16)[name = tensor("kor_37_cast_fp16")]; + tensor var_4173_cast_fp16 = mul(x = var_4143_cast_fp16, y = roti_19_cast_fp16)[name = tensor("op_4173_cast_fp16")]; + tensor var_4174_cast_fp16 = mul(x = var_4151_cast_fp16, y = rotr_19_cast_fp16)[name = tensor("op_4174_cast_fp16")]; + tensor koi_37_cast_fp16 = add(x = var_4173_cast_fp16, y = var_4174_cast_fp16)[name = tensor("koi_37_cast_fp16")]; + tensor qo_19_axis_0 = const()[name = tensor("qo_19_axis_0"), val = tensor(-1)]; + tensor qo_19_cast_fp16 = stack(axis = qo_19_axis_0, values = (qor_37_cast_fp16, qoi_37_cast_fp16))[name = tensor("qo_19_cast_fp16")]; + tensor ko_19_axis_0 = const()[name = tensor("ko_19_axis_0"), val = tensor(-1)]; + tensor ko_19_cast_fp16 = stack(axis = ko_19_axis_0, values = (kor_37_cast_fp16, koi_37_cast_fp16))[name = tensor("ko_19_cast_fp16")]; + tensor var_4203 = const()[name = tensor("op_4203"), val = tensor([1, 1, 16, 64])]; + tensor q_57_cast_fp16 = reshape(shape = var_4203, x = qo_19_cast_fp16)[name = tensor("q_57_cast_fp16")]; + tensor var_4205 = const()[name = tensor("op_4205"), val = tensor([1, 1, 16, 64])]; + tensor k_39_cast_fp16 = reshape(shape = var_4205, x = ko_19_cast_fp16)[name = tensor("k_39_cast_fp16")]; + tensor _inversed_4227_y_0_to_fp16 = const()[name = tensor("_inversed_4227_y_0_to_fp16"), val = tensor(0x1p-9)]; + tensor _inversed_4227_cast_fp16 = mul(x = ts_59_cast_fp16, y = _inversed_4227_y_0_to_fp16)[name = tensor("_inversed_4227_cast_fp16")]; + tensor var_4228_cast_fp16 = floor(x = _inversed_4227_cast_fp16)[name = tensor("op_4228_cast_fp16")]; + tensor var_4229_to_fp16 = const()[name = tensor("op_4229_to_fp16"), val = tensor(0x1p+9)]; + tensor var_4230_cast_fp16 = mul(x = var_4228_cast_fp16, y = var_4229_to_fp16)[name = tensor("op_4230_cast_fp16")]; + tensor write_indices_float_39_cast_fp16 = sub(x = ts_59_cast_fp16, y = var_4230_cast_fp16)[name = tensor("write_indices_float_39_cast_fp16")]; + tensor var_4237_dtype_0 = const()[name = tensor("op_4237_dtype_0"), val = tensor("int32")]; + tensor write_indices_19_reps_0 = const()[name = tensor("write_indices_19_reps_0"), val = tensor([1, 1, 16, 64])]; + tensor write_indices_float_39_cast_fp16_to_int32 = cast(dtype = var_4237_dtype_0, x = write_indices_float_39_cast_fp16)[name = tensor("cast_507")]; + tensor write_indices_19 = tile(reps = write_indices_19_reps_0, x = write_indices_float_39_cast_fp16_to_int32)[name = tensor("write_indices_19")]; + tensor var_4245_begin_0 = const()[name = tensor("op_4245_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_4245_end_0 = const()[name = tensor("op_4245_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor var_4245_end_mask_0 = const()[name = tensor("op_4245_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_4245_squeeze_mask_0 = const()[name = tensor("op_4245_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor cache9_to_fp16_dtype_0 = const()[name = tensor("cache9_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor cache9_to_fp16 = cast(dtype = cache9_to_fp16_dtype_0, x = cache9)[name = tensor("cast_506")]; + tensor var_4245_cast_fp16 = slice_by_index(begin = var_4245_begin_0, end = var_4245_end_0, end_mask = var_4245_end_mask_0, squeeze_mask = var_4245_squeeze_mask_0, x = cache9_to_fp16)[name = tensor("op_4245_cast_fp16")]; + tensor var_4247_axis_0 = const()[name = tensor("op_4247_axis_0"), val = tensor(1)]; + tensor var_4247_mode_0 = const()[name = tensor("op_4247_mode_0"), val = tensor("update")]; + tensor var_4247_validate_indices_0 = const()[name = tensor("op_4247_validate_indices_0"), val = tensor(false)]; + tensor var_4247_cast_fp16 = scatter_along_axis(axis = var_4247_axis_0, data = var_4245_cast_fp16, indices = write_indices_19, mode = var_4247_mode_0, updates = k_39_cast_fp16, validate_indices = var_4247_validate_indices_0)[name = tensor("op_4247_cast_fp16")]; + tensor concat_65 = const()[name = tensor("concat_65"), val = tensor([0, 0, 0, 0, 0])]; + tensor concat_66 = const()[name = tensor("concat_66"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_19_internal_tensor_assign_1_stride_0 = const()[name = tensor("new_cache_19_internal_tensor_assign_1_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_19_internal_tensor_assign_1_begin_mask_0 = const()[name = tensor("new_cache_19_internal_tensor_assign_1_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_19_internal_tensor_assign_1_end_mask_0 = const()[name = tensor("new_cache_19_internal_tensor_assign_1_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_19_internal_tensor_assign_1_squeeze_mask_0 = const()[name = tensor("new_cache_19_internal_tensor_assign_1_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_66 = const()[name = tensor("shape_66"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_18 = const()[name = tensor("reduce_prod_18"), val = tensor(1048576)]; + tensor range_1d_18_start_0 = const()[name = tensor("range_1d_18_start_0"), val = tensor(0)]; + tensor range_1d_18_step_0 = const()[name = tensor("range_1d_18_step_0"), val = tensor(1)]; + tensor range_1d_18 = range_1d(end = reduce_prod_18, start = range_1d_18_start_0, step = range_1d_18_step_0)[name = tensor("range_1d_18")]; + tensor reshape_90 = reshape(shape = shape_66, x = range_1d_18)[name = tensor("reshape_90")]; + tensor slice_by_index_18 = slice_by_index(begin = concat_65, begin_mask = new_cache_19_internal_tensor_assign_1_begin_mask_0, end = concat_66, end_mask = new_cache_19_internal_tensor_assign_1_end_mask_0, squeeze_mask = new_cache_19_internal_tensor_assign_1_squeeze_mask_0, stride = new_cache_19_internal_tensor_assign_1_stride_0, x = reshape_90)[name = tensor("slice_by_index_18")]; + tensor reshape_91_shape_0 = const()[name = tensor("reshape_91_shape_0"), val = tensor([-1])]; + tensor reshape_91 = reshape(shape = reshape_91_shape_0, x = slice_by_index_18)[name = tensor("reshape_91")]; + tensor reshape_92_shape_0 = const()[name = tensor("reshape_92_shape_0"), val = tensor([-1])]; + tensor reshape_92_cast_fp16 = reshape(shape = reshape_92_shape_0, x = var_4247_cast_fp16)[name = tensor("reshape_92_cast_fp16")]; + tensor reshape_93_shape_0 = const()[name = tensor("reshape_93_shape_0"), val = tensor([-1])]; + tensor reshape_93_cast_fp16 = reshape(shape = reshape_93_shape_0, x = cache9_to_fp16)[name = tensor("reshape_93_cast_fp16")]; + tensor scatter_18_mode_0 = const()[name = tensor("scatter_18_mode_0"), val = tensor("update")]; + tensor scatter_18_axis_0 = const()[name = tensor("scatter_18_axis_0"), val = tensor(0)]; + tensor scatter_18_validate_indices_0 = const()[name = tensor("scatter_18_validate_indices_0"), val = tensor(false)]; + tensor scatter_18_cast_fp16 = scatter(axis = scatter_18_axis_0, data = reshape_93_cast_fp16, indices = reshape_91, mode = scatter_18_mode_0, updates = reshape_92_cast_fp16, validate_indices = scatter_18_validate_indices_0)[name = tensor("scatter_18_cast_fp16")]; + tensor reshape_94_cast_fp16 = reshape(shape = shape_66, x = scatter_18_cast_fp16)[name = tensor("reshape_94_cast_fp16")]; + tensor var_4255_begin_0 = const()[name = tensor("op_4255_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor var_4255_end_0 = const()[name = tensor("op_4255_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor var_4255_end_mask_0 = const()[name = tensor("op_4255_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_4255_squeeze_mask_0 = const()[name = tensor("op_4255_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor var_4255_cast_fp16 = slice_by_index(begin = var_4255_begin_0, end = var_4255_end_0, end_mask = var_4255_end_mask_0, squeeze_mask = var_4255_squeeze_mask_0, x = reshape_94_cast_fp16)[name = tensor("op_4255_cast_fp16")]; + tensor var_4257_axis_0 = const()[name = tensor("op_4257_axis_0"), val = tensor(1)]; + tensor var_4257_mode_0 = const()[name = tensor("op_4257_mode_0"), val = tensor("update")]; + tensor var_4257_validate_indices_0 = const()[name = tensor("op_4257_validate_indices_0"), val = tensor(false)]; + tensor var_4257_cast_fp16 = scatter_along_axis(axis = var_4257_axis_0, data = var_4255_cast_fp16, indices = write_indices_19, mode = var_4257_mode_0, updates = v_19_cast_fp16, validate_indices = var_4257_validate_indices_0)[name = tensor("op_4257_cast_fp16")]; + tensor concat_67 = const()[name = tensor("concat_67"), val = tensor([1, 0, 0, 0, 0])]; + tensor concat_68 = const()[name = tensor("concat_68"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_19_internal_tensor_assign_2_stride_0 = const()[name = tensor("new_cache_19_internal_tensor_assign_2_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_19_internal_tensor_assign_2_begin_mask_0 = const()[name = tensor("new_cache_19_internal_tensor_assign_2_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_19_internal_tensor_assign_2_end_mask_0 = const()[name = tensor("new_cache_19_internal_tensor_assign_2_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_19_internal_tensor_assign_2_squeeze_mask_0 = const()[name = tensor("new_cache_19_internal_tensor_assign_2_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_67 = const()[name = tensor("shape_67"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_19 = const()[name = tensor("reduce_prod_19"), val = tensor(1048576)]; + tensor range_1d_19_start_0 = const()[name = tensor("range_1d_19_start_0"), val = tensor(0)]; + tensor range_1d_19_step_0 = const()[name = tensor("range_1d_19_step_0"), val = tensor(1)]; + tensor range_1d_19 = range_1d(end = reduce_prod_19, start = range_1d_19_start_0, step = range_1d_19_step_0)[name = tensor("range_1d_19")]; + tensor reshape_95 = reshape(shape = shape_67, x = range_1d_19)[name = tensor("reshape_95")]; + tensor slice_by_index_19 = slice_by_index(begin = concat_67, begin_mask = new_cache_19_internal_tensor_assign_2_begin_mask_0, end = concat_68, end_mask = new_cache_19_internal_tensor_assign_2_end_mask_0, squeeze_mask = new_cache_19_internal_tensor_assign_2_squeeze_mask_0, stride = new_cache_19_internal_tensor_assign_2_stride_0, x = reshape_95)[name = tensor("slice_by_index_19")]; + tensor reshape_96_shape_0 = const()[name = tensor("reshape_96_shape_0"), val = tensor([-1])]; + tensor reshape_96 = reshape(shape = reshape_96_shape_0, x = slice_by_index_19)[name = tensor("reshape_96")]; + tensor reshape_97_shape_0 = const()[name = tensor("reshape_97_shape_0"), val = tensor([-1])]; + tensor reshape_97_cast_fp16 = reshape(shape = reshape_97_shape_0, x = var_4257_cast_fp16)[name = tensor("reshape_97_cast_fp16")]; + tensor reshape_98_shape_0 = const()[name = tensor("reshape_98_shape_0"), val = tensor([-1])]; + tensor reshape_98_cast_fp16 = reshape(shape = reshape_98_shape_0, x = reshape_94_cast_fp16)[name = tensor("reshape_98_cast_fp16")]; + tensor scatter_19_mode_0 = const()[name = tensor("scatter_19_mode_0"), val = tensor("update")]; + tensor scatter_19_axis_0 = const()[name = tensor("scatter_19_axis_0"), val = tensor(0)]; + tensor scatter_19_validate_indices_0 = const()[name = tensor("scatter_19_validate_indices_0"), val = tensor(false)]; + tensor scatter_19_cast_fp16 = scatter(axis = scatter_19_axis_0, data = reshape_98_cast_fp16, indices = reshape_96, mode = scatter_19_mode_0, updates = reshape_97_cast_fp16, validate_indices = scatter_19_validate_indices_0)[name = tensor("scatter_19_cast_fp16")]; + tensor reshape_99_cast_fp16 = reshape(shape = shape_67, x = scatter_19_cast_fp16)[name = tensor("reshape_99_cast_fp16")]; + tensor reshape_99_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("reshape_99_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor keys_55_begin_0 = const()[name = tensor("keys_55_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor keys_55_end_0 = const()[name = tensor("keys_55_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor keys_55_end_mask_0 = const()[name = tensor("keys_55_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor keys_55_squeeze_mask_0 = const()[name = tensor("keys_55_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor keys_55_cast_fp16 = slice_by_index(begin = keys_55_begin_0, end = keys_55_end_0, end_mask = keys_55_end_mask_0, squeeze_mask = keys_55_squeeze_mask_0, x = reshape_99_cast_fp16)[name = tensor("keys_55_cast_fp16")]; + tensor values_55_begin_0 = const()[name = tensor("values_55_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor values_55_end_0 = const()[name = tensor("values_55_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor values_55_end_mask_0 = const()[name = tensor("values_55_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor values_55_squeeze_mask_0 = const()[name = tensor("values_55_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor values_55_cast_fp16 = slice_by_index(begin = values_55_begin_0, end = values_55_end_0, end_mask = values_55_end_mask_0, squeeze_mask = values_55_squeeze_mask_0, x = reshape_99_cast_fp16)[name = tensor("values_55_cast_fp16")]; + tensor var_4269_cast_fp16 = not_equal(x = keys_55_cast_fp16, y = keys_55_cast_fp16)[name = tensor("op_4269_cast_fp16")]; + tensor keys_57_cast_fp16 = select(a = var_504_to_fp16, b = keys_55_cast_fp16, cond = var_4269_cast_fp16)[name = tensor("keys_57_cast_fp16")]; + tensor var_4277_cast_fp16 = not_equal(x = values_55_cast_fp16, y = values_55_cast_fp16)[name = tensor("op_4277_cast_fp16")]; + tensor values_57_cast_fp16 = select(a = var_504_to_fp16, b = values_55_cast_fp16, cond = var_4277_cast_fp16)[name = tensor("values_57_cast_fp16")]; + tensor var_4301 = const()[name = tensor("op_4301"), val = tensor([0, 2, 1, 3])]; + tensor var_4314 = const()[name = tensor("op_4314"), val = tensor([1, 1, 1])]; + tensor var_4315_cast_fp16 = reshape(shape = var_4314, x = position9_to_fp16)[name = tensor("op_4315_cast_fp16")]; + tensor var_4332_to_fp16 = const()[name = tensor("op_4332_to_fp16"), val = tensor(0x1p+0)]; + tensor valid_len_19_cast_fp16 = add(x = var_4315_cast_fp16, y = var_4332_to_fp16)[name = tensor("valid_len_19_cast_fp16")]; + tensor valid_mask_19_cast_fp16 = less(x = k_positions_1_promoted_to_fp16, y = valid_len_19_cast_fp16)[name = tensor("valid_mask_19_cast_fp16")]; + tensor causal_mask_19_cast_fp16 = less_equal(x = k_positions_1_promoted_to_fp16, y = var_4315_cast_fp16)[name = tensor("causal_mask_19_cast_fp16")]; + tensor attn_mask_37 = logical_and(x = valid_mask_19_cast_fp16, y = causal_mask_19_cast_fp16)[name = tensor("attn_mask_37")]; + tensor attn_mask_39_axes_0 = const()[name = tensor("attn_mask_39_axes_0"), val = tensor([1])]; + tensor attn_mask_39 = expand_dims(axes = attn_mask_39_axes_0, x = attn_mask_37)[name = tensor("attn_mask_39")]; + tensor var_4350_transpose_x_0 = const()[name = tensor("op_4350_transpose_x_0"), val = tensor(false)]; + tensor var_4350_transpose_y_0 = const()[name = tensor("op_4350_transpose_y_0"), val = tensor(false)]; + tensor transpose_90_perm_0 = const()[name = tensor("transpose_90_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_91_perm_0 = const()[name = tensor("transpose_91_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_91 = transpose(perm = transpose_91_perm_0, x = keys_57_cast_fp16)[name = tensor("transpose_177")]; + tensor transpose_90 = transpose(perm = transpose_90_perm_0, x = q_57_cast_fp16)[name = tensor("transpose_178")]; + tensor var_4350_cast_fp16 = matmul(transpose_x = var_4350_transpose_x_0, transpose_y = var_4350_transpose_y_0, x = transpose_90, y = transpose_91)[name = tensor("op_4350_cast_fp16")]; + tensor var_4344_to_fp16 = const()[name = tensor("op_4344_to_fp16"), val = tensor([0x1p-3])]; + tensor attn_weights_55_cast_fp16 = mul(x = var_4350_cast_fp16, y = var_4344_to_fp16)[name = tensor("attn_weights_55_cast_fp16")]; + tensor var_4352 = logical_not(x = attn_mask_39)[name = tensor("op_4352")]; + tensor var_4353_to_fp16 = const()[name = tensor("op_4353_to_fp16"), val = tensor(-inf)]; + tensor attn_weights_57_cast_fp16 = select(a = var_4353_to_fp16, b = attn_weights_55_cast_fp16, cond = var_4352)[name = tensor("attn_weights_57_cast_fp16")]; + tensor var_4355 = const()[name = tensor("op_4355"), val = tensor(-1)]; + tensor attn_weights_59_cast_fp16 = softmax(axis = var_4355, x = attn_weights_57_cast_fp16)[name = tensor("attn_weights_59_cast_fp16")]; + tensor attn_output_19_transpose_x_0 = const()[name = tensor("attn_output_19_transpose_x_0"), val = tensor(false)]; + tensor attn_output_19_transpose_y_0 = const()[name = tensor("attn_output_19_transpose_y_0"), val = tensor(false)]; + tensor values_59_cast_fp16 = transpose(perm = var_4301, x = values_57_cast_fp16)[name = tensor("transpose_179")]; + tensor attn_output_19_cast_fp16 = matmul(transpose_x = attn_output_19_transpose_x_0, transpose_y = attn_output_19_transpose_y_0, x = attn_weights_59_cast_fp16, y = values_59_cast_fp16)[name = tensor("attn_output_19_cast_fp16")]; + tensor var_4363 = const()[name = tensor("op_4363"), val = tensor([0, 2, 1, 3])]; + tensor var_4366 = const()[name = tensor("op_4366"), val = tensor([1, 1, 1024])]; + tensor var_4364_cast_fp16 = transpose(perm = var_4363, x = attn_output_19_cast_fp16)[name = tensor("transpose_176")]; + tensor input_95_cast_fp16 = reshape(shape = var_4366, x = var_4364_cast_fp16)[name = tensor("input_95_cast_fp16")]; + tensor linear_38_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = attn9_out_proj_weight_cast_fp16, x = input_95_cast_fp16)[name = tensor("linear_38_cast_fp16")]; + tensor var_4372_to_fp16 = const()[name = tensor("op_4372_to_fp16"), val = tensor(0x1p+0)]; + tensor var_4373_cast_fp16 = add(x = position9_to_fp16, y = var_4372_to_fp16)[name = tensor("op_4373_cast_fp16")]; + tensor var_4373_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_4373_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor input_97_cast_fp16 = add(x = input_93_cast_fp16, y = linear_38_cast_fp16)[name = tensor("input_97_cast_fp16")]; + tensor input_99_axes_0 = const()[name = tensor("input_99_axes_0"), val = tensor([-1])]; + tensor norm9_2_weight_to_fp16 = const()[name = tensor("norm9_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(303666432)))]; + tensor norm9_2_bias_to_fp16 = const()[name = tensor("norm9_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(303668544)))]; + tensor var_4377_to_fp16 = const()[name = tensor("op_4377_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_99_cast_fp16 = layer_norm(axes = input_99_axes_0, beta = norm9_2_bias_to_fp16, epsilon = var_4377_to_fp16, gamma = norm9_2_weight_to_fp16, x = input_97_cast_fp16)[name = tensor("input_99_cast_fp16")]; + tensor linear_39_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = linear9_1_weight_cast_fp16, x = input_99_cast_fp16)[name = tensor("linear_39_cast_fp16")]; + tensor input_101_mode_0 = const()[name = tensor("input_101_mode_0"), val = tensor("EXACT")]; + tensor input_101_cast_fp16 = gelu(mode = input_101_mode_0, x = linear_39_cast_fp16)[name = tensor("input_101_cast_fp16")]; + tensor linear_40_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = linear9_2_weight_cast_fp16, x = input_101_cast_fp16)[name = tensor("linear_40_cast_fp16")]; + tensor input_103_cast_fp16 = add(x = input_97_cast_fp16, y = linear_40_cast_fp16)[name = tensor("input_103_cast_fp16")]; + tensor x_21_axes_0 = const()[name = tensor("x_21_axes_0"), val = tensor([-1])]; + tensor norm10_1_weight_to_fp16 = const()[name = tensor("norm10_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(303670656)))]; + tensor norm10_1_bias_to_fp16 = const()[name = tensor("norm10_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(303672768)))]; + tensor var_4394_to_fp16 = const()[name = tensor("op_4394_to_fp16"), val = tensor(0x1.5p-17)]; + tensor x_21_cast_fp16 = layer_norm(axes = x_21_axes_0, beta = norm10_1_bias_to_fp16, epsilon = var_4394_to_fp16, gamma = norm10_1_weight_to_fp16, x = input_103_cast_fp16)[name = tensor("x_21_cast_fp16")]; + tensor linear_41_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = attn10_in_proj_weight_cast_fp16, x = x_21_cast_fp16)[name = tensor("linear_41_cast_fp16")]; + tensor var_4430 = const()[name = tensor("op_4430"), val = tensor([1, 1, 3, 16, 64])]; + tensor qkv_21_cast_fp16 = reshape(shape = var_4430, x = linear_41_cast_fp16)[name = tensor("qkv_21_cast_fp16")]; + tensor q_61_begin_0 = const()[name = tensor("q_61_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor q_61_end_0 = const()[name = tensor("q_61_end_0"), val = tensor([1, 1, 1, 16, 64])]; + tensor q_61_end_mask_0 = const()[name = tensor("q_61_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor q_61_squeeze_mask_0 = const()[name = tensor("q_61_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor q_61_cast_fp16 = slice_by_index(begin = q_61_begin_0, end = q_61_end_0, end_mask = q_61_end_mask_0, squeeze_mask = q_61_squeeze_mask_0, x = qkv_21_cast_fp16)[name = tensor("q_61_cast_fp16")]; + tensor k_41_begin_0 = const()[name = tensor("k_41_begin_0"), val = tensor([0, 0, 1, 0, 0])]; + tensor k_41_end_0 = const()[name = tensor("k_41_end_0"), val = tensor([1, 1, 2, 16, 64])]; + tensor k_41_end_mask_0 = const()[name = tensor("k_41_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor k_41_squeeze_mask_0 = const()[name = tensor("k_41_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor k_41_cast_fp16 = slice_by_index(begin = k_41_begin_0, end = k_41_end_0, end_mask = k_41_end_mask_0, squeeze_mask = k_41_squeeze_mask_0, x = qkv_21_cast_fp16)[name = tensor("k_41_cast_fp16")]; + tensor v_21_begin_0 = const()[name = tensor("v_21_begin_0"), val = tensor([0, 0, 2, 0, 0])]; + tensor v_21_end_0 = const()[name = tensor("v_21_end_0"), val = tensor([1, 1, 3, 16, 64])]; + tensor v_21_end_mask_0 = const()[name = tensor("v_21_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor v_21_squeeze_mask_0 = const()[name = tensor("v_21_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor v_21_cast_fp16 = slice_by_index(begin = v_21_begin_0, end = v_21_end_0, end_mask = v_21_end_mask_0, squeeze_mask = v_21_squeeze_mask_0, x = qkv_21_cast_fp16)[name = tensor("v_21_cast_fp16")]; + tensor var_4534 = const()[name = tensor("op_4534"), val = tensor([1, 1, 1, 1])]; + tensor position10_to_fp16_dtype_0 = const()[name = tensor("position10_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor position10_to_fp16 = cast(dtype = position10_to_fp16_dtype_0, x = position10)[name = tensor("cast_503")]; + tensor ts_65_cast_fp16 = reshape(shape = var_4534, x = position10_to_fp16)[name = tensor("ts_65_cast_fp16")]; + tensor var_4538 = const()[name = tensor("op_4538"), val = tensor([1, 1, 16, 32, 2])]; + tensor q_complex_21_cast_fp16 = reshape(shape = var_4538, x = q_61_cast_fp16)[name = tensor("q_complex_21_cast_fp16")]; + tensor var_4542 = const()[name = tensor("op_4542"), val = tensor([1, 1, 16, 32, 2])]; + tensor k_complex_21_cast_fp16 = reshape(shape = var_4542, x = k_41_cast_fp16)[name = tensor("k_complex_21_cast_fp16")]; + tensor var_4546_begin_0 = const()[name = tensor("op_4546_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_4546_end_0 = const()[name = tensor("op_4546_end_0"), val = tensor([1, 1, 16, 32, 1])]; + tensor var_4546_end_mask_0 = const()[name = tensor("op_4546_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_4546_squeeze_mask_0 = const()[name = tensor("op_4546_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_4546_cast_fp16 = slice_by_index(begin = var_4546_begin_0, end = var_4546_end_0, end_mask = var_4546_end_mask_0, squeeze_mask = var_4546_squeeze_mask_0, x = q_complex_21_cast_fp16)[name = tensor("op_4546_cast_fp16")]; + tensor var_4554_begin_0 = const()[name = tensor("op_4554_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_4554_end_0 = const()[name = tensor("op_4554_end_0"), val = tensor([1, 1, 16, 32, 2])]; + tensor var_4554_end_mask_0 = const()[name = tensor("op_4554_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_4554_squeeze_mask_0 = const()[name = tensor("op_4554_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_4554_cast_fp16 = slice_by_index(begin = var_4554_begin_0, end = var_4554_end_0, end_mask = var_4554_end_mask_0, squeeze_mask = var_4554_squeeze_mask_0, x = q_complex_21_cast_fp16)[name = tensor("op_4554_cast_fp16")]; + tensor var_4562_begin_0 = const()[name = tensor("op_4562_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_4562_end_0 = const()[name = tensor("op_4562_end_0"), val = tensor([1, 1, 16, 32, 1])]; + tensor var_4562_end_mask_0 = const()[name = tensor("op_4562_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_4562_squeeze_mask_0 = const()[name = tensor("op_4562_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_4562_cast_fp16 = slice_by_index(begin = var_4562_begin_0, end = var_4562_end_0, end_mask = var_4562_end_mask_0, squeeze_mask = var_4562_squeeze_mask_0, x = k_complex_21_cast_fp16)[name = tensor("op_4562_cast_fp16")]; + tensor var_4570_begin_0 = const()[name = tensor("op_4570_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_4570_end_0 = const()[name = tensor("op_4570_end_0"), val = tensor([1, 1, 16, 32, 2])]; + tensor var_4570_end_mask_0 = const()[name = tensor("op_4570_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_4570_squeeze_mask_0 = const()[name = tensor("op_4570_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_4570_cast_fp16 = slice_by_index(begin = var_4570_begin_0, end = var_4570_end_0, end_mask = var_4570_end_mask_0, squeeze_mask = var_4570_squeeze_mask_0, x = k_complex_21_cast_fp16)[name = tensor("op_4570_cast_fp16")]; + tensor freqs_21_to_fp16 = const()[name = tensor("freqs_21_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(303674880)))]; + tensor var_4576_cast_fp16 = mul(x = freqs_21_to_fp16, y = ts_65_cast_fp16)[name = tensor("op_4576_cast_fp16")]; + tensor rotr_21_cast_fp16 = cos(x = var_4576_cast_fp16)[name = tensor("rotr_21_cast_fp16")]; + tensor roti_21_cast_fp16 = sin(x = var_4576_cast_fp16)[name = tensor("roti_21_cast_fp16")]; + tensor var_4580_cast_fp16 = mul(x = var_4546_cast_fp16, y = rotr_21_cast_fp16)[name = tensor("op_4580_cast_fp16")]; + tensor var_4581_cast_fp16 = mul(x = var_4554_cast_fp16, y = roti_21_cast_fp16)[name = tensor("op_4581_cast_fp16")]; + tensor qor_41_cast_fp16 = sub(x = var_4580_cast_fp16, y = var_4581_cast_fp16)[name = tensor("qor_41_cast_fp16")]; + tensor var_4584_cast_fp16 = mul(x = var_4546_cast_fp16, y = roti_21_cast_fp16)[name = tensor("op_4584_cast_fp16")]; + tensor var_4585_cast_fp16 = mul(x = var_4554_cast_fp16, y = rotr_21_cast_fp16)[name = tensor("op_4585_cast_fp16")]; + tensor qoi_41_cast_fp16 = add(x = var_4584_cast_fp16, y = var_4585_cast_fp16)[name = tensor("qoi_41_cast_fp16")]; + tensor var_4588_cast_fp16 = mul(x = var_4562_cast_fp16, y = rotr_21_cast_fp16)[name = tensor("op_4588_cast_fp16")]; + tensor var_4589_cast_fp16 = mul(x = var_4570_cast_fp16, y = roti_21_cast_fp16)[name = tensor("op_4589_cast_fp16")]; + tensor kor_41_cast_fp16 = sub(x = var_4588_cast_fp16, y = var_4589_cast_fp16)[name = tensor("kor_41_cast_fp16")]; + tensor var_4592_cast_fp16 = mul(x = var_4562_cast_fp16, y = roti_21_cast_fp16)[name = tensor("op_4592_cast_fp16")]; + tensor var_4593_cast_fp16 = mul(x = var_4570_cast_fp16, y = rotr_21_cast_fp16)[name = tensor("op_4593_cast_fp16")]; + tensor koi_41_cast_fp16 = add(x = var_4592_cast_fp16, y = var_4593_cast_fp16)[name = tensor("koi_41_cast_fp16")]; + tensor qo_21_axis_0 = const()[name = tensor("qo_21_axis_0"), val = tensor(-1)]; + tensor qo_21_cast_fp16 = stack(axis = qo_21_axis_0, values = (qor_41_cast_fp16, qoi_41_cast_fp16))[name = tensor("qo_21_cast_fp16")]; + tensor ko_21_axis_0 = const()[name = tensor("ko_21_axis_0"), val = tensor(-1)]; + tensor ko_21_cast_fp16 = stack(axis = ko_21_axis_0, values = (kor_41_cast_fp16, koi_41_cast_fp16))[name = tensor("ko_21_cast_fp16")]; + tensor var_4622 = const()[name = tensor("op_4622"), val = tensor([1, 1, 16, 64])]; + tensor q_63_cast_fp16 = reshape(shape = var_4622, x = qo_21_cast_fp16)[name = tensor("q_63_cast_fp16")]; + tensor var_4624 = const()[name = tensor("op_4624"), val = tensor([1, 1, 16, 64])]; + tensor k_43_cast_fp16 = reshape(shape = var_4624, x = ko_21_cast_fp16)[name = tensor("k_43_cast_fp16")]; + tensor _inversed_4646_y_0_to_fp16 = const()[name = tensor("_inversed_4646_y_0_to_fp16"), val = tensor(0x1p-9)]; + tensor _inversed_4646_cast_fp16 = mul(x = ts_65_cast_fp16, y = _inversed_4646_y_0_to_fp16)[name = tensor("_inversed_4646_cast_fp16")]; + tensor var_4647_cast_fp16 = floor(x = _inversed_4646_cast_fp16)[name = tensor("op_4647_cast_fp16")]; + tensor var_4648_to_fp16 = const()[name = tensor("op_4648_to_fp16"), val = tensor(0x1p+9)]; + tensor var_4649_cast_fp16 = mul(x = var_4647_cast_fp16, y = var_4648_to_fp16)[name = tensor("op_4649_cast_fp16")]; + tensor write_indices_float_43_cast_fp16 = sub(x = ts_65_cast_fp16, y = var_4649_cast_fp16)[name = tensor("write_indices_float_43_cast_fp16")]; + tensor var_4656_dtype_0 = const()[name = tensor("op_4656_dtype_0"), val = tensor("int32")]; + tensor write_indices_21_reps_0 = const()[name = tensor("write_indices_21_reps_0"), val = tensor([1, 1, 16, 64])]; + tensor write_indices_float_43_cast_fp16_to_int32 = cast(dtype = var_4656_dtype_0, x = write_indices_float_43_cast_fp16)[name = tensor("cast_502")]; + tensor write_indices_21 = tile(reps = write_indices_21_reps_0, x = write_indices_float_43_cast_fp16_to_int32)[name = tensor("write_indices_21")]; + tensor var_4664_begin_0 = const()[name = tensor("op_4664_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_4664_end_0 = const()[name = tensor("op_4664_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor var_4664_end_mask_0 = const()[name = tensor("op_4664_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_4664_squeeze_mask_0 = const()[name = tensor("op_4664_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor cache10_to_fp16_dtype_0 = const()[name = tensor("cache10_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor cache10_to_fp16 = cast(dtype = cache10_to_fp16_dtype_0, x = cache10)[name = tensor("cast_501")]; + tensor var_4664_cast_fp16 = slice_by_index(begin = var_4664_begin_0, end = var_4664_end_0, end_mask = var_4664_end_mask_0, squeeze_mask = var_4664_squeeze_mask_0, x = cache10_to_fp16)[name = tensor("op_4664_cast_fp16")]; + tensor var_4666_axis_0 = const()[name = tensor("op_4666_axis_0"), val = tensor(1)]; + tensor var_4666_mode_0 = const()[name = tensor("op_4666_mode_0"), val = tensor("update")]; + tensor var_4666_validate_indices_0 = const()[name = tensor("op_4666_validate_indices_0"), val = tensor(false)]; + tensor var_4666_cast_fp16 = scatter_along_axis(axis = var_4666_axis_0, data = var_4664_cast_fp16, indices = write_indices_21, mode = var_4666_mode_0, updates = k_43_cast_fp16, validate_indices = var_4666_validate_indices_0)[name = tensor("op_4666_cast_fp16")]; + tensor concat_72 = const()[name = tensor("concat_72"), val = tensor([0, 0, 0, 0, 0])]; + tensor concat_73 = const()[name = tensor("concat_73"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_21_internal_tensor_assign_1_stride_0 = const()[name = tensor("new_cache_21_internal_tensor_assign_1_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_21_internal_tensor_assign_1_begin_mask_0 = const()[name = tensor("new_cache_21_internal_tensor_assign_1_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_21_internal_tensor_assign_1_end_mask_0 = const()[name = tensor("new_cache_21_internal_tensor_assign_1_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_21_internal_tensor_assign_1_squeeze_mask_0 = const()[name = tensor("new_cache_21_internal_tensor_assign_1_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_68 = const()[name = tensor("shape_68"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_20 = const()[name = tensor("reduce_prod_20"), val = tensor(1048576)]; + tensor range_1d_20_start_0 = const()[name = tensor("range_1d_20_start_0"), val = tensor(0)]; + tensor range_1d_20_step_0 = const()[name = tensor("range_1d_20_step_0"), val = tensor(1)]; + tensor range_1d_20 = range_1d(end = reduce_prod_20, start = range_1d_20_start_0, step = range_1d_20_step_0)[name = tensor("range_1d_20")]; + tensor reshape_100 = reshape(shape = shape_68, x = range_1d_20)[name = tensor("reshape_100")]; + tensor slice_by_index_20 = slice_by_index(begin = concat_72, begin_mask = new_cache_21_internal_tensor_assign_1_begin_mask_0, end = concat_73, end_mask = new_cache_21_internal_tensor_assign_1_end_mask_0, squeeze_mask = new_cache_21_internal_tensor_assign_1_squeeze_mask_0, stride = new_cache_21_internal_tensor_assign_1_stride_0, x = reshape_100)[name = tensor("slice_by_index_20")]; + tensor reshape_101_shape_0 = const()[name = tensor("reshape_101_shape_0"), val = tensor([-1])]; + tensor reshape_101 = reshape(shape = reshape_101_shape_0, x = slice_by_index_20)[name = tensor("reshape_101")]; + tensor reshape_102_shape_0 = const()[name = tensor("reshape_102_shape_0"), val = tensor([-1])]; + tensor reshape_102_cast_fp16 = reshape(shape = reshape_102_shape_0, x = var_4666_cast_fp16)[name = tensor("reshape_102_cast_fp16")]; + tensor reshape_103_shape_0 = const()[name = tensor("reshape_103_shape_0"), val = tensor([-1])]; + tensor reshape_103_cast_fp16 = reshape(shape = reshape_103_shape_0, x = cache10_to_fp16)[name = tensor("reshape_103_cast_fp16")]; + tensor scatter_20_mode_0 = const()[name = tensor("scatter_20_mode_0"), val = tensor("update")]; + tensor scatter_20_axis_0 = const()[name = tensor("scatter_20_axis_0"), val = tensor(0)]; + tensor scatter_20_validate_indices_0 = const()[name = tensor("scatter_20_validate_indices_0"), val = tensor(false)]; + tensor scatter_20_cast_fp16 = scatter(axis = scatter_20_axis_0, data = reshape_103_cast_fp16, indices = reshape_101, mode = scatter_20_mode_0, updates = reshape_102_cast_fp16, validate_indices = scatter_20_validate_indices_0)[name = tensor("scatter_20_cast_fp16")]; + tensor reshape_104_cast_fp16 = reshape(shape = shape_68, x = scatter_20_cast_fp16)[name = tensor("reshape_104_cast_fp16")]; + tensor var_4674_begin_0 = const()[name = tensor("op_4674_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor var_4674_end_0 = const()[name = tensor("op_4674_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor var_4674_end_mask_0 = const()[name = tensor("op_4674_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_4674_squeeze_mask_0 = const()[name = tensor("op_4674_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor var_4674_cast_fp16 = slice_by_index(begin = var_4674_begin_0, end = var_4674_end_0, end_mask = var_4674_end_mask_0, squeeze_mask = var_4674_squeeze_mask_0, x = reshape_104_cast_fp16)[name = tensor("op_4674_cast_fp16")]; + tensor var_4676_axis_0 = const()[name = tensor("op_4676_axis_0"), val = tensor(1)]; + tensor var_4676_mode_0 = const()[name = tensor("op_4676_mode_0"), val = tensor("update")]; + tensor var_4676_validate_indices_0 = const()[name = tensor("op_4676_validate_indices_0"), val = tensor(false)]; + tensor var_4676_cast_fp16 = scatter_along_axis(axis = var_4676_axis_0, data = var_4674_cast_fp16, indices = write_indices_21, mode = var_4676_mode_0, updates = v_21_cast_fp16, validate_indices = var_4676_validate_indices_0)[name = tensor("op_4676_cast_fp16")]; + tensor concat_74 = const()[name = tensor("concat_74"), val = tensor([1, 0, 0, 0, 0])]; + tensor concat_75 = const()[name = tensor("concat_75"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_21_internal_tensor_assign_2_stride_0 = const()[name = tensor("new_cache_21_internal_tensor_assign_2_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_21_internal_tensor_assign_2_begin_mask_0 = const()[name = tensor("new_cache_21_internal_tensor_assign_2_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_21_internal_tensor_assign_2_end_mask_0 = const()[name = tensor("new_cache_21_internal_tensor_assign_2_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_21_internal_tensor_assign_2_squeeze_mask_0 = const()[name = tensor("new_cache_21_internal_tensor_assign_2_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_69 = const()[name = tensor("shape_69"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_21 = const()[name = tensor("reduce_prod_21"), val = tensor(1048576)]; + tensor range_1d_21_start_0 = const()[name = tensor("range_1d_21_start_0"), val = tensor(0)]; + tensor range_1d_21_step_0 = const()[name = tensor("range_1d_21_step_0"), val = tensor(1)]; + tensor range_1d_21 = range_1d(end = reduce_prod_21, start = range_1d_21_start_0, step = range_1d_21_step_0)[name = tensor("range_1d_21")]; + tensor reshape_105 = reshape(shape = shape_69, x = range_1d_21)[name = tensor("reshape_105")]; + tensor slice_by_index_21 = slice_by_index(begin = concat_74, begin_mask = new_cache_21_internal_tensor_assign_2_begin_mask_0, end = concat_75, end_mask = new_cache_21_internal_tensor_assign_2_end_mask_0, squeeze_mask = new_cache_21_internal_tensor_assign_2_squeeze_mask_0, stride = new_cache_21_internal_tensor_assign_2_stride_0, x = reshape_105)[name = tensor("slice_by_index_21")]; + tensor reshape_106_shape_0 = const()[name = tensor("reshape_106_shape_0"), val = tensor([-1])]; + tensor reshape_106 = reshape(shape = reshape_106_shape_0, x = slice_by_index_21)[name = tensor("reshape_106")]; + tensor reshape_107_shape_0 = const()[name = tensor("reshape_107_shape_0"), val = tensor([-1])]; + tensor reshape_107_cast_fp16 = reshape(shape = reshape_107_shape_0, x = var_4676_cast_fp16)[name = tensor("reshape_107_cast_fp16")]; + tensor reshape_108_shape_0 = const()[name = tensor("reshape_108_shape_0"), val = tensor([-1])]; + tensor reshape_108_cast_fp16 = reshape(shape = reshape_108_shape_0, x = reshape_104_cast_fp16)[name = tensor("reshape_108_cast_fp16")]; + tensor scatter_21_mode_0 = const()[name = tensor("scatter_21_mode_0"), val = tensor("update")]; + tensor scatter_21_axis_0 = const()[name = tensor("scatter_21_axis_0"), val = tensor(0)]; + tensor scatter_21_validate_indices_0 = const()[name = tensor("scatter_21_validate_indices_0"), val = tensor(false)]; + tensor scatter_21_cast_fp16 = scatter(axis = scatter_21_axis_0, data = reshape_108_cast_fp16, indices = reshape_106, mode = scatter_21_mode_0, updates = reshape_107_cast_fp16, validate_indices = scatter_21_validate_indices_0)[name = tensor("scatter_21_cast_fp16")]; + tensor reshape_109_cast_fp16 = reshape(shape = shape_69, x = scatter_21_cast_fp16)[name = tensor("reshape_109_cast_fp16")]; + tensor reshape_109_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("reshape_109_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor keys_61_begin_0 = const()[name = tensor("keys_61_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor keys_61_end_0 = const()[name = tensor("keys_61_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor keys_61_end_mask_0 = const()[name = tensor("keys_61_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor keys_61_squeeze_mask_0 = const()[name = tensor("keys_61_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor keys_61_cast_fp16 = slice_by_index(begin = keys_61_begin_0, end = keys_61_end_0, end_mask = keys_61_end_mask_0, squeeze_mask = keys_61_squeeze_mask_0, x = reshape_109_cast_fp16)[name = tensor("keys_61_cast_fp16")]; + tensor values_61_begin_0 = const()[name = tensor("values_61_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor values_61_end_0 = const()[name = tensor("values_61_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor values_61_end_mask_0 = const()[name = tensor("values_61_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor values_61_squeeze_mask_0 = const()[name = tensor("values_61_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor values_61_cast_fp16 = slice_by_index(begin = values_61_begin_0, end = values_61_end_0, end_mask = values_61_end_mask_0, squeeze_mask = values_61_squeeze_mask_0, x = reshape_109_cast_fp16)[name = tensor("values_61_cast_fp16")]; + tensor var_4688_cast_fp16 = not_equal(x = keys_61_cast_fp16, y = keys_61_cast_fp16)[name = tensor("op_4688_cast_fp16")]; + tensor keys_63_cast_fp16 = select(a = var_504_to_fp16, b = keys_61_cast_fp16, cond = var_4688_cast_fp16)[name = tensor("keys_63_cast_fp16")]; + tensor var_4696_cast_fp16 = not_equal(x = values_61_cast_fp16, y = values_61_cast_fp16)[name = tensor("op_4696_cast_fp16")]; + tensor values_63_cast_fp16 = select(a = var_504_to_fp16, b = values_61_cast_fp16, cond = var_4696_cast_fp16)[name = tensor("values_63_cast_fp16")]; + tensor var_4720 = const()[name = tensor("op_4720"), val = tensor([0, 2, 1, 3])]; + tensor var_4733 = const()[name = tensor("op_4733"), val = tensor([1, 1, 1])]; + tensor var_4734_cast_fp16 = reshape(shape = var_4733, x = position10_to_fp16)[name = tensor("op_4734_cast_fp16")]; + tensor var_4751_to_fp16 = const()[name = tensor("op_4751_to_fp16"), val = tensor(0x1p+0)]; + tensor valid_len_21_cast_fp16 = add(x = var_4734_cast_fp16, y = var_4751_to_fp16)[name = tensor("valid_len_21_cast_fp16")]; + tensor valid_mask_21_cast_fp16 = less(x = k_positions_1_promoted_to_fp16, y = valid_len_21_cast_fp16)[name = tensor("valid_mask_21_cast_fp16")]; + tensor causal_mask_21_cast_fp16 = less_equal(x = k_positions_1_promoted_to_fp16, y = var_4734_cast_fp16)[name = tensor("causal_mask_21_cast_fp16")]; + tensor attn_mask_41 = logical_and(x = valid_mask_21_cast_fp16, y = causal_mask_21_cast_fp16)[name = tensor("attn_mask_41")]; + tensor attn_mask_43_axes_0 = const()[name = tensor("attn_mask_43_axes_0"), val = tensor([1])]; + tensor attn_mask_43 = expand_dims(axes = attn_mask_43_axes_0, x = attn_mask_41)[name = tensor("attn_mask_43")]; + tensor var_4769_transpose_x_0 = const()[name = tensor("op_4769_transpose_x_0"), val = tensor(false)]; + tensor var_4769_transpose_y_0 = const()[name = tensor("op_4769_transpose_y_0"), val = tensor(false)]; + tensor transpose_92_perm_0 = const()[name = tensor("transpose_92_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_93_perm_0 = const()[name = tensor("transpose_93_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_93 = transpose(perm = transpose_93_perm_0, x = keys_63_cast_fp16)[name = tensor("transpose_173")]; + tensor transpose_92 = transpose(perm = transpose_92_perm_0, x = q_63_cast_fp16)[name = tensor("transpose_174")]; + tensor var_4769_cast_fp16 = matmul(transpose_x = var_4769_transpose_x_0, transpose_y = var_4769_transpose_y_0, x = transpose_92, y = transpose_93)[name = tensor("op_4769_cast_fp16")]; + tensor var_4763_to_fp16 = const()[name = tensor("op_4763_to_fp16"), val = tensor([0x1p-3])]; + tensor attn_weights_61_cast_fp16 = mul(x = var_4769_cast_fp16, y = var_4763_to_fp16)[name = tensor("attn_weights_61_cast_fp16")]; + tensor var_4771 = logical_not(x = attn_mask_43)[name = tensor("op_4771")]; + tensor var_4772_to_fp16 = const()[name = tensor("op_4772_to_fp16"), val = tensor(-inf)]; + tensor attn_weights_63_cast_fp16 = select(a = var_4772_to_fp16, b = attn_weights_61_cast_fp16, cond = var_4771)[name = tensor("attn_weights_63_cast_fp16")]; + tensor var_4774 = const()[name = tensor("op_4774"), val = tensor(-1)]; + tensor attn_weights_65_cast_fp16 = softmax(axis = var_4774, x = attn_weights_63_cast_fp16)[name = tensor("attn_weights_65_cast_fp16")]; + tensor attn_output_21_transpose_x_0 = const()[name = tensor("attn_output_21_transpose_x_0"), val = tensor(false)]; + tensor attn_output_21_transpose_y_0 = const()[name = tensor("attn_output_21_transpose_y_0"), val = tensor(false)]; + tensor values_65_cast_fp16 = transpose(perm = var_4720, x = values_63_cast_fp16)[name = tensor("transpose_175")]; + tensor attn_output_21_cast_fp16 = matmul(transpose_x = attn_output_21_transpose_x_0, transpose_y = attn_output_21_transpose_y_0, x = attn_weights_65_cast_fp16, y = values_65_cast_fp16)[name = tensor("attn_output_21_cast_fp16")]; + tensor var_4782 = const()[name = tensor("op_4782"), val = tensor([0, 2, 1, 3])]; + tensor var_4785 = const()[name = tensor("op_4785"), val = tensor([1, 1, 1024])]; + tensor var_4783_cast_fp16 = transpose(perm = var_4782, x = attn_output_21_cast_fp16)[name = tensor("transpose_172")]; + tensor input_105_cast_fp16 = reshape(shape = var_4785, x = var_4783_cast_fp16)[name = tensor("input_105_cast_fp16")]; + tensor linear_42_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = attn10_out_proj_weight_cast_fp16, x = input_105_cast_fp16)[name = tensor("linear_42_cast_fp16")]; + tensor var_4791_to_fp16 = const()[name = tensor("op_4791_to_fp16"), val = tensor(0x1p+0)]; + tensor var_4792_cast_fp16 = add(x = position10_to_fp16, y = var_4791_to_fp16)[name = tensor("op_4792_cast_fp16")]; + tensor var_4792_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_4792_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor input_107_cast_fp16 = add(x = input_103_cast_fp16, y = linear_42_cast_fp16)[name = tensor("input_107_cast_fp16")]; + tensor input_109_axes_0 = const()[name = tensor("input_109_axes_0"), val = tensor([-1])]; + tensor norm10_2_weight_to_fp16 = const()[name = tensor("norm10_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(303675008)))]; + tensor norm10_2_bias_to_fp16 = const()[name = tensor("norm10_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(303677120)))]; + tensor var_4796_to_fp16 = const()[name = tensor("op_4796_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_109_cast_fp16 = layer_norm(axes = input_109_axes_0, beta = norm10_2_bias_to_fp16, epsilon = var_4796_to_fp16, gamma = norm10_2_weight_to_fp16, x = input_107_cast_fp16)[name = tensor("input_109_cast_fp16")]; + tensor linear_43_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = linear10_1_weight_cast_fp16, x = input_109_cast_fp16)[name = tensor("linear_43_cast_fp16")]; + tensor input_111_mode_0 = const()[name = tensor("input_111_mode_0"), val = tensor("EXACT")]; + tensor input_111_cast_fp16 = gelu(mode = input_111_mode_0, x = linear_43_cast_fp16)[name = tensor("input_111_cast_fp16")]; + tensor linear_44_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = linear10_2_weight_cast_fp16, x = input_111_cast_fp16)[name = tensor("linear_44_cast_fp16")]; + tensor input_113_cast_fp16 = add(x = input_107_cast_fp16, y = linear_44_cast_fp16)[name = tensor("input_113_cast_fp16")]; + tensor x_23_axes_0 = const()[name = tensor("x_23_axes_0"), val = tensor([-1])]; + tensor norm11_1_weight_to_fp16 = const()[name = tensor("norm11_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(303679232)))]; + tensor norm11_1_bias_to_fp16 = const()[name = tensor("norm11_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(303681344)))]; + tensor var_4813_to_fp16 = const()[name = tensor("op_4813_to_fp16"), val = tensor(0x1.5p-17)]; + tensor x_23_cast_fp16 = layer_norm(axes = x_23_axes_0, beta = norm11_1_bias_to_fp16, epsilon = var_4813_to_fp16, gamma = norm11_1_weight_to_fp16, x = input_113_cast_fp16)[name = tensor("x_23_cast_fp16")]; + tensor linear_45_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = attn11_in_proj_weight_cast_fp16, x = x_23_cast_fp16)[name = tensor("linear_45_cast_fp16")]; + tensor var_4849 = const()[name = tensor("op_4849"), val = tensor([1, 1, 3, 16, 64])]; + tensor qkv_23_cast_fp16 = reshape(shape = var_4849, x = linear_45_cast_fp16)[name = tensor("qkv_23_cast_fp16")]; + tensor q_67_begin_0 = const()[name = tensor("q_67_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor q_67_end_0 = const()[name = tensor("q_67_end_0"), val = tensor([1, 1, 1, 16, 64])]; + tensor q_67_end_mask_0 = const()[name = tensor("q_67_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor q_67_squeeze_mask_0 = const()[name = tensor("q_67_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor q_67_cast_fp16 = slice_by_index(begin = q_67_begin_0, end = q_67_end_0, end_mask = q_67_end_mask_0, squeeze_mask = q_67_squeeze_mask_0, x = qkv_23_cast_fp16)[name = tensor("q_67_cast_fp16")]; + tensor k_45_begin_0 = const()[name = tensor("k_45_begin_0"), val = tensor([0, 0, 1, 0, 0])]; + tensor k_45_end_0 = const()[name = tensor("k_45_end_0"), val = tensor([1, 1, 2, 16, 64])]; + tensor k_45_end_mask_0 = const()[name = tensor("k_45_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor k_45_squeeze_mask_0 = const()[name = tensor("k_45_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor k_45_cast_fp16 = slice_by_index(begin = k_45_begin_0, end = k_45_end_0, end_mask = k_45_end_mask_0, squeeze_mask = k_45_squeeze_mask_0, x = qkv_23_cast_fp16)[name = tensor("k_45_cast_fp16")]; + tensor v_23_begin_0 = const()[name = tensor("v_23_begin_0"), val = tensor([0, 0, 2, 0, 0])]; + tensor v_23_end_0 = const()[name = tensor("v_23_end_0"), val = tensor([1, 1, 3, 16, 64])]; + tensor v_23_end_mask_0 = const()[name = tensor("v_23_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor v_23_squeeze_mask_0 = const()[name = tensor("v_23_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor v_23_cast_fp16 = slice_by_index(begin = v_23_begin_0, end = v_23_end_0, end_mask = v_23_end_mask_0, squeeze_mask = v_23_squeeze_mask_0, x = qkv_23_cast_fp16)[name = tensor("v_23_cast_fp16")]; + tensor var_4953 = const()[name = tensor("op_4953"), val = tensor([1, 1, 1, 1])]; + tensor position11_to_fp16_dtype_0 = const()[name = tensor("position11_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor position11_to_fp16 = cast(dtype = position11_to_fp16_dtype_0, x = position11)[name = tensor("cast_498")]; + tensor ts_71_cast_fp16 = reshape(shape = var_4953, x = position11_to_fp16)[name = tensor("ts_71_cast_fp16")]; + tensor var_4957 = const()[name = tensor("op_4957"), val = tensor([1, 1, 16, 32, 2])]; + tensor q_complex_23_cast_fp16 = reshape(shape = var_4957, x = q_67_cast_fp16)[name = tensor("q_complex_23_cast_fp16")]; + tensor var_4961 = const()[name = tensor("op_4961"), val = tensor([1, 1, 16, 32, 2])]; + tensor k_complex_23_cast_fp16 = reshape(shape = var_4961, x = k_45_cast_fp16)[name = tensor("k_complex_23_cast_fp16")]; + tensor var_4965_begin_0 = const()[name = tensor("op_4965_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_4965_end_0 = const()[name = tensor("op_4965_end_0"), val = tensor([1, 1, 16, 32, 1])]; + tensor var_4965_end_mask_0 = const()[name = tensor("op_4965_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_4965_squeeze_mask_0 = const()[name = tensor("op_4965_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_4965_cast_fp16 = slice_by_index(begin = var_4965_begin_0, end = var_4965_end_0, end_mask = var_4965_end_mask_0, squeeze_mask = var_4965_squeeze_mask_0, x = q_complex_23_cast_fp16)[name = tensor("op_4965_cast_fp16")]; + tensor var_4973_begin_0 = const()[name = tensor("op_4973_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_4973_end_0 = const()[name = tensor("op_4973_end_0"), val = tensor([1, 1, 16, 32, 2])]; + tensor var_4973_end_mask_0 = const()[name = tensor("op_4973_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_4973_squeeze_mask_0 = const()[name = tensor("op_4973_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_4973_cast_fp16 = slice_by_index(begin = var_4973_begin_0, end = var_4973_end_0, end_mask = var_4973_end_mask_0, squeeze_mask = var_4973_squeeze_mask_0, x = q_complex_23_cast_fp16)[name = tensor("op_4973_cast_fp16")]; + tensor var_4981_begin_0 = const()[name = tensor("op_4981_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_4981_end_0 = const()[name = tensor("op_4981_end_0"), val = tensor([1, 1, 16, 32, 1])]; + tensor var_4981_end_mask_0 = const()[name = tensor("op_4981_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_4981_squeeze_mask_0 = const()[name = tensor("op_4981_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_4981_cast_fp16 = slice_by_index(begin = var_4981_begin_0, end = var_4981_end_0, end_mask = var_4981_end_mask_0, squeeze_mask = var_4981_squeeze_mask_0, x = k_complex_23_cast_fp16)[name = tensor("op_4981_cast_fp16")]; + tensor var_4989_begin_0 = const()[name = tensor("op_4989_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_4989_end_0 = const()[name = tensor("op_4989_end_0"), val = tensor([1, 1, 16, 32, 2])]; + tensor var_4989_end_mask_0 = const()[name = tensor("op_4989_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_4989_squeeze_mask_0 = const()[name = tensor("op_4989_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_4989_cast_fp16 = slice_by_index(begin = var_4989_begin_0, end = var_4989_end_0, end_mask = var_4989_end_mask_0, squeeze_mask = var_4989_squeeze_mask_0, x = k_complex_23_cast_fp16)[name = tensor("op_4989_cast_fp16")]; + tensor freqs_23_to_fp16 = const()[name = tensor("freqs_23_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(303683456)))]; + tensor var_4995_cast_fp16 = mul(x = freqs_23_to_fp16, y = ts_71_cast_fp16)[name = tensor("op_4995_cast_fp16")]; + tensor rotr_23_cast_fp16 = cos(x = var_4995_cast_fp16)[name = tensor("rotr_23_cast_fp16")]; + tensor roti_23_cast_fp16 = sin(x = var_4995_cast_fp16)[name = tensor("roti_23_cast_fp16")]; + tensor var_4999_cast_fp16 = mul(x = var_4965_cast_fp16, y = rotr_23_cast_fp16)[name = tensor("op_4999_cast_fp16")]; + tensor var_5000_cast_fp16 = mul(x = var_4973_cast_fp16, y = roti_23_cast_fp16)[name = tensor("op_5000_cast_fp16")]; + tensor qor_45_cast_fp16 = sub(x = var_4999_cast_fp16, y = var_5000_cast_fp16)[name = tensor("qor_45_cast_fp16")]; + tensor var_5003_cast_fp16 = mul(x = var_4965_cast_fp16, y = roti_23_cast_fp16)[name = tensor("op_5003_cast_fp16")]; + tensor var_5004_cast_fp16 = mul(x = var_4973_cast_fp16, y = rotr_23_cast_fp16)[name = tensor("op_5004_cast_fp16")]; + tensor qoi_45_cast_fp16 = add(x = var_5003_cast_fp16, y = var_5004_cast_fp16)[name = tensor("qoi_45_cast_fp16")]; + tensor var_5007_cast_fp16 = mul(x = var_4981_cast_fp16, y = rotr_23_cast_fp16)[name = tensor("op_5007_cast_fp16")]; + tensor var_5008_cast_fp16 = mul(x = var_4989_cast_fp16, y = roti_23_cast_fp16)[name = tensor("op_5008_cast_fp16")]; + tensor kor_45_cast_fp16 = sub(x = var_5007_cast_fp16, y = var_5008_cast_fp16)[name = tensor("kor_45_cast_fp16")]; + tensor var_5011_cast_fp16 = mul(x = var_4981_cast_fp16, y = roti_23_cast_fp16)[name = tensor("op_5011_cast_fp16")]; + tensor var_5012_cast_fp16 = mul(x = var_4989_cast_fp16, y = rotr_23_cast_fp16)[name = tensor("op_5012_cast_fp16")]; + tensor koi_45_cast_fp16 = add(x = var_5011_cast_fp16, y = var_5012_cast_fp16)[name = tensor("koi_45_cast_fp16")]; + tensor qo_23_axis_0 = const()[name = tensor("qo_23_axis_0"), val = tensor(-1)]; + tensor qo_23_cast_fp16 = stack(axis = qo_23_axis_0, values = (qor_45_cast_fp16, qoi_45_cast_fp16))[name = tensor("qo_23_cast_fp16")]; + tensor ko_23_axis_0 = const()[name = tensor("ko_23_axis_0"), val = tensor(-1)]; + tensor ko_23_cast_fp16 = stack(axis = ko_23_axis_0, values = (kor_45_cast_fp16, koi_45_cast_fp16))[name = tensor("ko_23_cast_fp16")]; + tensor var_5041 = const()[name = tensor("op_5041"), val = tensor([1, 1, 16, 64])]; + tensor q_69_cast_fp16 = reshape(shape = var_5041, x = qo_23_cast_fp16)[name = tensor("q_69_cast_fp16")]; + tensor var_5043 = const()[name = tensor("op_5043"), val = tensor([1, 1, 16, 64])]; + tensor k_47_cast_fp16 = reshape(shape = var_5043, x = ko_23_cast_fp16)[name = tensor("k_47_cast_fp16")]; + tensor _inversed_5065_y_0_to_fp16 = const()[name = tensor("_inversed_5065_y_0_to_fp16"), val = tensor(0x1p-9)]; + tensor _inversed_5065_cast_fp16 = mul(x = ts_71_cast_fp16, y = _inversed_5065_y_0_to_fp16)[name = tensor("_inversed_5065_cast_fp16")]; + tensor var_5066_cast_fp16 = floor(x = _inversed_5065_cast_fp16)[name = tensor("op_5066_cast_fp16")]; + tensor var_5067_to_fp16 = const()[name = tensor("op_5067_to_fp16"), val = tensor(0x1p+9)]; + tensor var_5068_cast_fp16 = mul(x = var_5066_cast_fp16, y = var_5067_to_fp16)[name = tensor("op_5068_cast_fp16")]; + tensor write_indices_float_47_cast_fp16 = sub(x = ts_71_cast_fp16, y = var_5068_cast_fp16)[name = tensor("write_indices_float_47_cast_fp16")]; + tensor var_5075_dtype_0 = const()[name = tensor("op_5075_dtype_0"), val = tensor("int32")]; + tensor write_indices_23_reps_0 = const()[name = tensor("write_indices_23_reps_0"), val = tensor([1, 1, 16, 64])]; + tensor write_indices_float_47_cast_fp16_to_int32 = cast(dtype = var_5075_dtype_0, x = write_indices_float_47_cast_fp16)[name = tensor("cast_497")]; + tensor write_indices_23 = tile(reps = write_indices_23_reps_0, x = write_indices_float_47_cast_fp16_to_int32)[name = tensor("write_indices_23")]; + tensor var_5083_begin_0 = const()[name = tensor("op_5083_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_5083_end_0 = const()[name = tensor("op_5083_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor var_5083_end_mask_0 = const()[name = tensor("op_5083_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_5083_squeeze_mask_0 = const()[name = tensor("op_5083_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor cache11_to_fp16_dtype_0 = const()[name = tensor("cache11_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor cache11_to_fp16 = cast(dtype = cache11_to_fp16_dtype_0, x = cache11)[name = tensor("cast_496")]; + tensor var_5083_cast_fp16 = slice_by_index(begin = var_5083_begin_0, end = var_5083_end_0, end_mask = var_5083_end_mask_0, squeeze_mask = var_5083_squeeze_mask_0, x = cache11_to_fp16)[name = tensor("op_5083_cast_fp16")]; + tensor var_5085_axis_0 = const()[name = tensor("op_5085_axis_0"), val = tensor(1)]; + tensor var_5085_mode_0 = const()[name = tensor("op_5085_mode_0"), val = tensor("update")]; + tensor var_5085_validate_indices_0 = const()[name = tensor("op_5085_validate_indices_0"), val = tensor(false)]; + tensor var_5085_cast_fp16 = scatter_along_axis(axis = var_5085_axis_0, data = var_5083_cast_fp16, indices = write_indices_23, mode = var_5085_mode_0, updates = k_47_cast_fp16, validate_indices = var_5085_validate_indices_0)[name = tensor("op_5085_cast_fp16")]; + tensor concat_79 = const()[name = tensor("concat_79"), val = tensor([0, 0, 0, 0, 0])]; + tensor concat_80 = const()[name = tensor("concat_80"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_23_internal_tensor_assign_1_stride_0 = const()[name = tensor("new_cache_23_internal_tensor_assign_1_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_23_internal_tensor_assign_1_begin_mask_0 = const()[name = tensor("new_cache_23_internal_tensor_assign_1_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_23_internal_tensor_assign_1_end_mask_0 = const()[name = tensor("new_cache_23_internal_tensor_assign_1_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_23_internal_tensor_assign_1_squeeze_mask_0 = const()[name = tensor("new_cache_23_internal_tensor_assign_1_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_70 = const()[name = tensor("shape_70"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_22 = const()[name = tensor("reduce_prod_22"), val = tensor(1048576)]; + tensor range_1d_22_start_0 = const()[name = tensor("range_1d_22_start_0"), val = tensor(0)]; + tensor range_1d_22_step_0 = const()[name = tensor("range_1d_22_step_0"), val = tensor(1)]; + tensor range_1d_22 = range_1d(end = reduce_prod_22, start = range_1d_22_start_0, step = range_1d_22_step_0)[name = tensor("range_1d_22")]; + tensor reshape_110 = reshape(shape = shape_70, x = range_1d_22)[name = tensor("reshape_110")]; + tensor slice_by_index_22 = slice_by_index(begin = concat_79, begin_mask = new_cache_23_internal_tensor_assign_1_begin_mask_0, end = concat_80, end_mask = new_cache_23_internal_tensor_assign_1_end_mask_0, squeeze_mask = new_cache_23_internal_tensor_assign_1_squeeze_mask_0, stride = new_cache_23_internal_tensor_assign_1_stride_0, x = reshape_110)[name = tensor("slice_by_index_22")]; + tensor reshape_111_shape_0 = const()[name = tensor("reshape_111_shape_0"), val = tensor([-1])]; + tensor reshape_111 = reshape(shape = reshape_111_shape_0, x = slice_by_index_22)[name = tensor("reshape_111")]; + tensor reshape_112_shape_0 = const()[name = tensor("reshape_112_shape_0"), val = tensor([-1])]; + tensor reshape_112_cast_fp16 = reshape(shape = reshape_112_shape_0, x = var_5085_cast_fp16)[name = tensor("reshape_112_cast_fp16")]; + tensor reshape_113_shape_0 = const()[name = tensor("reshape_113_shape_0"), val = tensor([-1])]; + tensor reshape_113_cast_fp16 = reshape(shape = reshape_113_shape_0, x = cache11_to_fp16)[name = tensor("reshape_113_cast_fp16")]; + tensor scatter_22_mode_0 = const()[name = tensor("scatter_22_mode_0"), val = tensor("update")]; + tensor scatter_22_axis_0 = const()[name = tensor("scatter_22_axis_0"), val = tensor(0)]; + tensor scatter_22_validate_indices_0 = const()[name = tensor("scatter_22_validate_indices_0"), val = tensor(false)]; + tensor scatter_22_cast_fp16 = scatter(axis = scatter_22_axis_0, data = reshape_113_cast_fp16, indices = reshape_111, mode = scatter_22_mode_0, updates = reshape_112_cast_fp16, validate_indices = scatter_22_validate_indices_0)[name = tensor("scatter_22_cast_fp16")]; + tensor reshape_114_cast_fp16 = reshape(shape = shape_70, x = scatter_22_cast_fp16)[name = tensor("reshape_114_cast_fp16")]; + tensor var_5093_begin_0 = const()[name = tensor("op_5093_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor var_5093_end_0 = const()[name = tensor("op_5093_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor var_5093_end_mask_0 = const()[name = tensor("op_5093_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_5093_squeeze_mask_0 = const()[name = tensor("op_5093_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor var_5093_cast_fp16 = slice_by_index(begin = var_5093_begin_0, end = var_5093_end_0, end_mask = var_5093_end_mask_0, squeeze_mask = var_5093_squeeze_mask_0, x = reshape_114_cast_fp16)[name = tensor("op_5093_cast_fp16")]; + tensor var_5095_axis_0 = const()[name = tensor("op_5095_axis_0"), val = tensor(1)]; + tensor var_5095_mode_0 = const()[name = tensor("op_5095_mode_0"), val = tensor("update")]; + tensor var_5095_validate_indices_0 = const()[name = tensor("op_5095_validate_indices_0"), val = tensor(false)]; + tensor var_5095_cast_fp16 = scatter_along_axis(axis = var_5095_axis_0, data = var_5093_cast_fp16, indices = write_indices_23, mode = var_5095_mode_0, updates = v_23_cast_fp16, validate_indices = var_5095_validate_indices_0)[name = tensor("op_5095_cast_fp16")]; + tensor concat_81 = const()[name = tensor("concat_81"), val = tensor([1, 0, 0, 0, 0])]; + tensor concat_82 = const()[name = tensor("concat_82"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_23_internal_tensor_assign_2_stride_0 = const()[name = tensor("new_cache_23_internal_tensor_assign_2_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_23_internal_tensor_assign_2_begin_mask_0 = const()[name = tensor("new_cache_23_internal_tensor_assign_2_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_23_internal_tensor_assign_2_end_mask_0 = const()[name = tensor("new_cache_23_internal_tensor_assign_2_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_23_internal_tensor_assign_2_squeeze_mask_0 = const()[name = tensor("new_cache_23_internal_tensor_assign_2_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_71 = const()[name = tensor("shape_71"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_23 = const()[name = tensor("reduce_prod_23"), val = tensor(1048576)]; + tensor range_1d_23_start_0 = const()[name = tensor("range_1d_23_start_0"), val = tensor(0)]; + tensor range_1d_23_step_0 = const()[name = tensor("range_1d_23_step_0"), val = tensor(1)]; + tensor range_1d_23 = range_1d(end = reduce_prod_23, start = range_1d_23_start_0, step = range_1d_23_step_0)[name = tensor("range_1d_23")]; + tensor reshape_115 = reshape(shape = shape_71, x = range_1d_23)[name = tensor("reshape_115")]; + tensor slice_by_index_23 = slice_by_index(begin = concat_81, begin_mask = new_cache_23_internal_tensor_assign_2_begin_mask_0, end = concat_82, end_mask = new_cache_23_internal_tensor_assign_2_end_mask_0, squeeze_mask = new_cache_23_internal_tensor_assign_2_squeeze_mask_0, stride = new_cache_23_internal_tensor_assign_2_stride_0, x = reshape_115)[name = tensor("slice_by_index_23")]; + tensor reshape_116_shape_0 = const()[name = tensor("reshape_116_shape_0"), val = tensor([-1])]; + tensor reshape_116 = reshape(shape = reshape_116_shape_0, x = slice_by_index_23)[name = tensor("reshape_116")]; + tensor reshape_117_shape_0 = const()[name = tensor("reshape_117_shape_0"), val = tensor([-1])]; + tensor reshape_117_cast_fp16 = reshape(shape = reshape_117_shape_0, x = var_5095_cast_fp16)[name = tensor("reshape_117_cast_fp16")]; + tensor reshape_118_shape_0 = const()[name = tensor("reshape_118_shape_0"), val = tensor([-1])]; + tensor reshape_118_cast_fp16 = reshape(shape = reshape_118_shape_0, x = reshape_114_cast_fp16)[name = tensor("reshape_118_cast_fp16")]; + tensor scatter_23_mode_0 = const()[name = tensor("scatter_23_mode_0"), val = tensor("update")]; + tensor scatter_23_axis_0 = const()[name = tensor("scatter_23_axis_0"), val = tensor(0)]; + tensor scatter_23_validate_indices_0 = const()[name = tensor("scatter_23_validate_indices_0"), val = tensor(false)]; + tensor scatter_23_cast_fp16 = scatter(axis = scatter_23_axis_0, data = reshape_118_cast_fp16, indices = reshape_116, mode = scatter_23_mode_0, updates = reshape_117_cast_fp16, validate_indices = scatter_23_validate_indices_0)[name = tensor("scatter_23_cast_fp16")]; + tensor reshape_119_cast_fp16 = reshape(shape = shape_71, x = scatter_23_cast_fp16)[name = tensor("reshape_119_cast_fp16")]; + tensor reshape_119_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("reshape_119_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor keys_67_begin_0 = const()[name = tensor("keys_67_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor keys_67_end_0 = const()[name = tensor("keys_67_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor keys_67_end_mask_0 = const()[name = tensor("keys_67_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor keys_67_squeeze_mask_0 = const()[name = tensor("keys_67_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor keys_67_cast_fp16 = slice_by_index(begin = keys_67_begin_0, end = keys_67_end_0, end_mask = keys_67_end_mask_0, squeeze_mask = keys_67_squeeze_mask_0, x = reshape_119_cast_fp16)[name = tensor("keys_67_cast_fp16")]; + tensor values_67_begin_0 = const()[name = tensor("values_67_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor values_67_end_0 = const()[name = tensor("values_67_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor values_67_end_mask_0 = const()[name = tensor("values_67_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor values_67_squeeze_mask_0 = const()[name = tensor("values_67_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor values_67_cast_fp16 = slice_by_index(begin = values_67_begin_0, end = values_67_end_0, end_mask = values_67_end_mask_0, squeeze_mask = values_67_squeeze_mask_0, x = reshape_119_cast_fp16)[name = tensor("values_67_cast_fp16")]; + tensor var_5107_cast_fp16 = not_equal(x = keys_67_cast_fp16, y = keys_67_cast_fp16)[name = tensor("op_5107_cast_fp16")]; + tensor keys_69_cast_fp16 = select(a = var_504_to_fp16, b = keys_67_cast_fp16, cond = var_5107_cast_fp16)[name = tensor("keys_69_cast_fp16")]; + tensor var_5115_cast_fp16 = not_equal(x = values_67_cast_fp16, y = values_67_cast_fp16)[name = tensor("op_5115_cast_fp16")]; + tensor values_69_cast_fp16 = select(a = var_504_to_fp16, b = values_67_cast_fp16, cond = var_5115_cast_fp16)[name = tensor("values_69_cast_fp16")]; + tensor var_5139 = const()[name = tensor("op_5139"), val = tensor([0, 2, 1, 3])]; + tensor var_5152 = const()[name = tensor("op_5152"), val = tensor([1, 1, 1])]; + tensor var_5153_cast_fp16 = reshape(shape = var_5152, x = position11_to_fp16)[name = tensor("op_5153_cast_fp16")]; + tensor var_5170_to_fp16 = const()[name = tensor("op_5170_to_fp16"), val = tensor(0x1p+0)]; + tensor valid_len_23_cast_fp16 = add(x = var_5153_cast_fp16, y = var_5170_to_fp16)[name = tensor("valid_len_23_cast_fp16")]; + tensor valid_mask_23_cast_fp16 = less(x = k_positions_1_promoted_to_fp16, y = valid_len_23_cast_fp16)[name = tensor("valid_mask_23_cast_fp16")]; + tensor causal_mask_23_cast_fp16 = less_equal(x = k_positions_1_promoted_to_fp16, y = var_5153_cast_fp16)[name = tensor("causal_mask_23_cast_fp16")]; + tensor attn_mask_45 = logical_and(x = valid_mask_23_cast_fp16, y = causal_mask_23_cast_fp16)[name = tensor("attn_mask_45")]; + tensor attn_mask_47_axes_0 = const()[name = tensor("attn_mask_47_axes_0"), val = tensor([1])]; + tensor attn_mask_47 = expand_dims(axes = attn_mask_47_axes_0, x = attn_mask_45)[name = tensor("attn_mask_47")]; + tensor var_5188_transpose_x_0 = const()[name = tensor("op_5188_transpose_x_0"), val = tensor(false)]; + tensor var_5188_transpose_y_0 = const()[name = tensor("op_5188_transpose_y_0"), val = tensor(false)]; + tensor transpose_94_perm_0 = const()[name = tensor("transpose_94_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_95_perm_0 = const()[name = tensor("transpose_95_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_95 = transpose(perm = transpose_95_perm_0, x = keys_69_cast_fp16)[name = tensor("transpose_169")]; + tensor transpose_94 = transpose(perm = transpose_94_perm_0, x = q_69_cast_fp16)[name = tensor("transpose_170")]; + tensor var_5188_cast_fp16 = matmul(transpose_x = var_5188_transpose_x_0, transpose_y = var_5188_transpose_y_0, x = transpose_94, y = transpose_95)[name = tensor("op_5188_cast_fp16")]; + tensor var_5182_to_fp16 = const()[name = tensor("op_5182_to_fp16"), val = tensor([0x1p-3])]; + tensor attn_weights_67_cast_fp16 = mul(x = var_5188_cast_fp16, y = var_5182_to_fp16)[name = tensor("attn_weights_67_cast_fp16")]; + tensor var_5190 = logical_not(x = attn_mask_47)[name = tensor("op_5190")]; + tensor var_5191_to_fp16 = const()[name = tensor("op_5191_to_fp16"), val = tensor(-inf)]; + tensor attn_weights_69_cast_fp16 = select(a = var_5191_to_fp16, b = attn_weights_67_cast_fp16, cond = var_5190)[name = tensor("attn_weights_69_cast_fp16")]; + tensor var_5193 = const()[name = tensor("op_5193"), val = tensor(-1)]; + tensor attn_weights_71_cast_fp16 = softmax(axis = var_5193, x = attn_weights_69_cast_fp16)[name = tensor("attn_weights_71_cast_fp16")]; + tensor attn_output_23_transpose_x_0 = const()[name = tensor("attn_output_23_transpose_x_0"), val = tensor(false)]; + tensor attn_output_23_transpose_y_0 = const()[name = tensor("attn_output_23_transpose_y_0"), val = tensor(false)]; + tensor values_71_cast_fp16 = transpose(perm = var_5139, x = values_69_cast_fp16)[name = tensor("transpose_171")]; + tensor attn_output_23_cast_fp16 = matmul(transpose_x = attn_output_23_transpose_x_0, transpose_y = attn_output_23_transpose_y_0, x = attn_weights_71_cast_fp16, y = values_71_cast_fp16)[name = tensor("attn_output_23_cast_fp16")]; + tensor var_5201 = const()[name = tensor("op_5201"), val = tensor([0, 2, 1, 3])]; + tensor var_5204 = const()[name = tensor("op_5204"), val = tensor([1, 1, 1024])]; + tensor var_5202_cast_fp16 = transpose(perm = var_5201, x = attn_output_23_cast_fp16)[name = tensor("transpose_168")]; + tensor input_115_cast_fp16 = reshape(shape = var_5204, x = var_5202_cast_fp16)[name = tensor("input_115_cast_fp16")]; + tensor linear_46_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = attn11_out_proj_weight_cast_fp16, x = input_115_cast_fp16)[name = tensor("linear_46_cast_fp16")]; + tensor var_5210_to_fp16 = const()[name = tensor("op_5210_to_fp16"), val = tensor(0x1p+0)]; + tensor var_5211_cast_fp16 = add(x = position11_to_fp16, y = var_5210_to_fp16)[name = tensor("op_5211_cast_fp16")]; + tensor var_5211_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_5211_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor input_117_cast_fp16 = add(x = input_113_cast_fp16, y = linear_46_cast_fp16)[name = tensor("input_117_cast_fp16")]; + tensor input_119_axes_0 = const()[name = tensor("input_119_axes_0"), val = tensor([-1])]; + tensor norm11_2_weight_to_fp16 = const()[name = tensor("norm11_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(303683584)))]; + tensor norm11_2_bias_to_fp16 = const()[name = tensor("norm11_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(303685696)))]; + tensor var_5215_to_fp16 = const()[name = tensor("op_5215_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_119_cast_fp16 = layer_norm(axes = input_119_axes_0, beta = norm11_2_bias_to_fp16, epsilon = var_5215_to_fp16, gamma = norm11_2_weight_to_fp16, x = input_117_cast_fp16)[name = tensor("input_119_cast_fp16")]; + tensor linear_47_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = linear11_1_weight_cast_fp16, x = input_119_cast_fp16)[name = tensor("linear_47_cast_fp16")]; + tensor input_121_mode_0 = const()[name = tensor("input_121_mode_0"), val = tensor("EXACT")]; + tensor input_121_cast_fp16 = gelu(mode = input_121_mode_0, x = linear_47_cast_fp16)[name = tensor("input_121_cast_fp16")]; + tensor linear_48_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = linear11_2_weight_cast_fp16, x = input_121_cast_fp16)[name = tensor("linear_48_cast_fp16")]; + tensor input_123_cast_fp16 = add(x = input_117_cast_fp16, y = linear_48_cast_fp16)[name = tensor("input_123_cast_fp16")]; + tensor x_25_axes_0 = const()[name = tensor("x_25_axes_0"), val = tensor([-1])]; + tensor norm12_1_weight_to_fp16 = const()[name = tensor("norm12_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(303687808)))]; + tensor norm12_1_bias_to_fp16 = const()[name = tensor("norm12_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(303689920)))]; + tensor var_5232_to_fp16 = const()[name = tensor("op_5232_to_fp16"), val = tensor(0x1.5p-17)]; + tensor x_25_cast_fp16 = layer_norm(axes = x_25_axes_0, beta = norm12_1_bias_to_fp16, epsilon = var_5232_to_fp16, gamma = norm12_1_weight_to_fp16, x = input_123_cast_fp16)[name = tensor("x_25_cast_fp16")]; + tensor linear_49_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = attn12_in_proj_weight_cast_fp16, x = x_25_cast_fp16)[name = tensor("linear_49_cast_fp16")]; + tensor var_5268 = const()[name = tensor("op_5268"), val = tensor([1, 1, 3, 16, 64])]; + tensor qkv_25_cast_fp16 = reshape(shape = var_5268, x = linear_49_cast_fp16)[name = tensor("qkv_25_cast_fp16")]; + tensor q_73_begin_0 = const()[name = tensor("q_73_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor q_73_end_0 = const()[name = tensor("q_73_end_0"), val = tensor([1, 1, 1, 16, 64])]; + tensor q_73_end_mask_0 = const()[name = tensor("q_73_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor q_73_squeeze_mask_0 = const()[name = tensor("q_73_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor q_73_cast_fp16 = slice_by_index(begin = q_73_begin_0, end = q_73_end_0, end_mask = q_73_end_mask_0, squeeze_mask = q_73_squeeze_mask_0, x = qkv_25_cast_fp16)[name = tensor("q_73_cast_fp16")]; + tensor k_49_begin_0 = const()[name = tensor("k_49_begin_0"), val = tensor([0, 0, 1, 0, 0])]; + tensor k_49_end_0 = const()[name = tensor("k_49_end_0"), val = tensor([1, 1, 2, 16, 64])]; + tensor k_49_end_mask_0 = const()[name = tensor("k_49_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor k_49_squeeze_mask_0 = const()[name = tensor("k_49_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor k_49_cast_fp16 = slice_by_index(begin = k_49_begin_0, end = k_49_end_0, end_mask = k_49_end_mask_0, squeeze_mask = k_49_squeeze_mask_0, x = qkv_25_cast_fp16)[name = tensor("k_49_cast_fp16")]; + tensor v_25_begin_0 = const()[name = tensor("v_25_begin_0"), val = tensor([0, 0, 2, 0, 0])]; + tensor v_25_end_0 = const()[name = tensor("v_25_end_0"), val = tensor([1, 1, 3, 16, 64])]; + tensor v_25_end_mask_0 = const()[name = tensor("v_25_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor v_25_squeeze_mask_0 = const()[name = tensor("v_25_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor v_25_cast_fp16 = slice_by_index(begin = v_25_begin_0, end = v_25_end_0, end_mask = v_25_end_mask_0, squeeze_mask = v_25_squeeze_mask_0, x = qkv_25_cast_fp16)[name = tensor("v_25_cast_fp16")]; + tensor var_5372 = const()[name = tensor("op_5372"), val = tensor([1, 1, 1, 1])]; + tensor position12_to_fp16_dtype_0 = const()[name = tensor("position12_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor position12_to_fp16 = cast(dtype = position12_to_fp16_dtype_0, x = position12)[name = tensor("cast_493")]; + tensor ts_77_cast_fp16 = reshape(shape = var_5372, x = position12_to_fp16)[name = tensor("ts_77_cast_fp16")]; + tensor var_5376 = const()[name = tensor("op_5376"), val = tensor([1, 1, 16, 32, 2])]; + tensor q_complex_25_cast_fp16 = reshape(shape = var_5376, x = q_73_cast_fp16)[name = tensor("q_complex_25_cast_fp16")]; + tensor var_5380 = const()[name = tensor("op_5380"), val = tensor([1, 1, 16, 32, 2])]; + tensor k_complex_25_cast_fp16 = reshape(shape = var_5380, x = k_49_cast_fp16)[name = tensor("k_complex_25_cast_fp16")]; + tensor var_5384_begin_0 = const()[name = tensor("op_5384_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_5384_end_0 = const()[name = tensor("op_5384_end_0"), val = tensor([1, 1, 16, 32, 1])]; + tensor var_5384_end_mask_0 = const()[name = tensor("op_5384_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_5384_squeeze_mask_0 = const()[name = tensor("op_5384_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_5384_cast_fp16 = slice_by_index(begin = var_5384_begin_0, end = var_5384_end_0, end_mask = var_5384_end_mask_0, squeeze_mask = var_5384_squeeze_mask_0, x = q_complex_25_cast_fp16)[name = tensor("op_5384_cast_fp16")]; + tensor var_5392_begin_0 = const()[name = tensor("op_5392_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_5392_end_0 = const()[name = tensor("op_5392_end_0"), val = tensor([1, 1, 16, 32, 2])]; + tensor var_5392_end_mask_0 = const()[name = tensor("op_5392_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_5392_squeeze_mask_0 = const()[name = tensor("op_5392_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_5392_cast_fp16 = slice_by_index(begin = var_5392_begin_0, end = var_5392_end_0, end_mask = var_5392_end_mask_0, squeeze_mask = var_5392_squeeze_mask_0, x = q_complex_25_cast_fp16)[name = tensor("op_5392_cast_fp16")]; + tensor var_5400_begin_0 = const()[name = tensor("op_5400_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_5400_end_0 = const()[name = tensor("op_5400_end_0"), val = tensor([1, 1, 16, 32, 1])]; + tensor var_5400_end_mask_0 = const()[name = tensor("op_5400_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_5400_squeeze_mask_0 = const()[name = tensor("op_5400_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_5400_cast_fp16 = slice_by_index(begin = var_5400_begin_0, end = var_5400_end_0, end_mask = var_5400_end_mask_0, squeeze_mask = var_5400_squeeze_mask_0, x = k_complex_25_cast_fp16)[name = tensor("op_5400_cast_fp16")]; + tensor var_5408_begin_0 = const()[name = tensor("op_5408_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_5408_end_0 = const()[name = tensor("op_5408_end_0"), val = tensor([1, 1, 16, 32, 2])]; + tensor var_5408_end_mask_0 = const()[name = tensor("op_5408_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_5408_squeeze_mask_0 = const()[name = tensor("op_5408_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_5408_cast_fp16 = slice_by_index(begin = var_5408_begin_0, end = var_5408_end_0, end_mask = var_5408_end_mask_0, squeeze_mask = var_5408_squeeze_mask_0, x = k_complex_25_cast_fp16)[name = tensor("op_5408_cast_fp16")]; + tensor freqs_25_to_fp16 = const()[name = tensor("freqs_25_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(303692032)))]; + tensor var_5414_cast_fp16 = mul(x = freqs_25_to_fp16, y = ts_77_cast_fp16)[name = tensor("op_5414_cast_fp16")]; + tensor rotr_25_cast_fp16 = cos(x = var_5414_cast_fp16)[name = tensor("rotr_25_cast_fp16")]; + tensor roti_25_cast_fp16 = sin(x = var_5414_cast_fp16)[name = tensor("roti_25_cast_fp16")]; + tensor var_5418_cast_fp16 = mul(x = var_5384_cast_fp16, y = rotr_25_cast_fp16)[name = tensor("op_5418_cast_fp16")]; + tensor var_5419_cast_fp16 = mul(x = var_5392_cast_fp16, y = roti_25_cast_fp16)[name = tensor("op_5419_cast_fp16")]; + tensor qor_49_cast_fp16 = sub(x = var_5418_cast_fp16, y = var_5419_cast_fp16)[name = tensor("qor_49_cast_fp16")]; + tensor var_5422_cast_fp16 = mul(x = var_5384_cast_fp16, y = roti_25_cast_fp16)[name = tensor("op_5422_cast_fp16")]; + tensor var_5423_cast_fp16 = mul(x = var_5392_cast_fp16, y = rotr_25_cast_fp16)[name = tensor("op_5423_cast_fp16")]; + tensor qoi_49_cast_fp16 = add(x = var_5422_cast_fp16, y = var_5423_cast_fp16)[name = tensor("qoi_49_cast_fp16")]; + tensor var_5426_cast_fp16 = mul(x = var_5400_cast_fp16, y = rotr_25_cast_fp16)[name = tensor("op_5426_cast_fp16")]; + tensor var_5427_cast_fp16 = mul(x = var_5408_cast_fp16, y = roti_25_cast_fp16)[name = tensor("op_5427_cast_fp16")]; + tensor kor_49_cast_fp16 = sub(x = var_5426_cast_fp16, y = var_5427_cast_fp16)[name = tensor("kor_49_cast_fp16")]; + tensor var_5430_cast_fp16 = mul(x = var_5400_cast_fp16, y = roti_25_cast_fp16)[name = tensor("op_5430_cast_fp16")]; + tensor var_5431_cast_fp16 = mul(x = var_5408_cast_fp16, y = rotr_25_cast_fp16)[name = tensor("op_5431_cast_fp16")]; + tensor koi_49_cast_fp16 = add(x = var_5430_cast_fp16, y = var_5431_cast_fp16)[name = tensor("koi_49_cast_fp16")]; + tensor qo_25_axis_0 = const()[name = tensor("qo_25_axis_0"), val = tensor(-1)]; + tensor qo_25_cast_fp16 = stack(axis = qo_25_axis_0, values = (qor_49_cast_fp16, qoi_49_cast_fp16))[name = tensor("qo_25_cast_fp16")]; + tensor ko_25_axis_0 = const()[name = tensor("ko_25_axis_0"), val = tensor(-1)]; + tensor ko_25_cast_fp16 = stack(axis = ko_25_axis_0, values = (kor_49_cast_fp16, koi_49_cast_fp16))[name = tensor("ko_25_cast_fp16")]; + tensor var_5460 = const()[name = tensor("op_5460"), val = tensor([1, 1, 16, 64])]; + tensor q_75_cast_fp16 = reshape(shape = var_5460, x = qo_25_cast_fp16)[name = tensor("q_75_cast_fp16")]; + tensor var_5462 = const()[name = tensor("op_5462"), val = tensor([1, 1, 16, 64])]; + tensor k_51_cast_fp16 = reshape(shape = var_5462, x = ko_25_cast_fp16)[name = tensor("k_51_cast_fp16")]; + tensor _inversed_5484_y_0_to_fp16 = const()[name = tensor("_inversed_5484_y_0_to_fp16"), val = tensor(0x1p-9)]; + tensor _inversed_5484_cast_fp16 = mul(x = ts_77_cast_fp16, y = _inversed_5484_y_0_to_fp16)[name = tensor("_inversed_5484_cast_fp16")]; + tensor var_5485_cast_fp16 = floor(x = _inversed_5484_cast_fp16)[name = tensor("op_5485_cast_fp16")]; + tensor var_5486_to_fp16 = const()[name = tensor("op_5486_to_fp16"), val = tensor(0x1p+9)]; + tensor var_5487_cast_fp16 = mul(x = var_5485_cast_fp16, y = var_5486_to_fp16)[name = tensor("op_5487_cast_fp16")]; + tensor write_indices_float_51_cast_fp16 = sub(x = ts_77_cast_fp16, y = var_5487_cast_fp16)[name = tensor("write_indices_float_51_cast_fp16")]; + tensor var_5494_dtype_0 = const()[name = tensor("op_5494_dtype_0"), val = tensor("int32")]; + tensor write_indices_25_reps_0 = const()[name = tensor("write_indices_25_reps_0"), val = tensor([1, 1, 16, 64])]; + tensor write_indices_float_51_cast_fp16_to_int32 = cast(dtype = var_5494_dtype_0, x = write_indices_float_51_cast_fp16)[name = tensor("cast_492")]; + tensor write_indices_25 = tile(reps = write_indices_25_reps_0, x = write_indices_float_51_cast_fp16_to_int32)[name = tensor("write_indices_25")]; + tensor var_5502_begin_0 = const()[name = tensor("op_5502_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_5502_end_0 = const()[name = tensor("op_5502_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor var_5502_end_mask_0 = const()[name = tensor("op_5502_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_5502_squeeze_mask_0 = const()[name = tensor("op_5502_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor cache12_to_fp16_dtype_0 = const()[name = tensor("cache12_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor cache12_to_fp16 = cast(dtype = cache12_to_fp16_dtype_0, x = cache12)[name = tensor("cast_491")]; + tensor var_5502_cast_fp16 = slice_by_index(begin = var_5502_begin_0, end = var_5502_end_0, end_mask = var_5502_end_mask_0, squeeze_mask = var_5502_squeeze_mask_0, x = cache12_to_fp16)[name = tensor("op_5502_cast_fp16")]; + tensor var_5504_axis_0 = const()[name = tensor("op_5504_axis_0"), val = tensor(1)]; + tensor var_5504_mode_0 = const()[name = tensor("op_5504_mode_0"), val = tensor("update")]; + tensor var_5504_validate_indices_0 = const()[name = tensor("op_5504_validate_indices_0"), val = tensor(false)]; + tensor var_5504_cast_fp16 = scatter_along_axis(axis = var_5504_axis_0, data = var_5502_cast_fp16, indices = write_indices_25, mode = var_5504_mode_0, updates = k_51_cast_fp16, validate_indices = var_5504_validate_indices_0)[name = tensor("op_5504_cast_fp16")]; + tensor concat_86 = const()[name = tensor("concat_86"), val = tensor([0, 0, 0, 0, 0])]; + tensor concat_87 = const()[name = tensor("concat_87"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_25_internal_tensor_assign_1_stride_0 = const()[name = tensor("new_cache_25_internal_tensor_assign_1_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_25_internal_tensor_assign_1_begin_mask_0 = const()[name = tensor("new_cache_25_internal_tensor_assign_1_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_25_internal_tensor_assign_1_end_mask_0 = const()[name = tensor("new_cache_25_internal_tensor_assign_1_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_25_internal_tensor_assign_1_squeeze_mask_0 = const()[name = tensor("new_cache_25_internal_tensor_assign_1_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_72 = const()[name = tensor("shape_72"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_24 = const()[name = tensor("reduce_prod_24"), val = tensor(1048576)]; + tensor range_1d_24_start_0 = const()[name = tensor("range_1d_24_start_0"), val = tensor(0)]; + tensor range_1d_24_step_0 = const()[name = tensor("range_1d_24_step_0"), val = tensor(1)]; + tensor range_1d_24 = range_1d(end = reduce_prod_24, start = range_1d_24_start_0, step = range_1d_24_step_0)[name = tensor("range_1d_24")]; + tensor reshape_120 = reshape(shape = shape_72, x = range_1d_24)[name = tensor("reshape_120")]; + tensor slice_by_index_24 = slice_by_index(begin = concat_86, begin_mask = new_cache_25_internal_tensor_assign_1_begin_mask_0, end = concat_87, end_mask = new_cache_25_internal_tensor_assign_1_end_mask_0, squeeze_mask = new_cache_25_internal_tensor_assign_1_squeeze_mask_0, stride = new_cache_25_internal_tensor_assign_1_stride_0, x = reshape_120)[name = tensor("slice_by_index_24")]; + tensor reshape_121_shape_0 = const()[name = tensor("reshape_121_shape_0"), val = tensor([-1])]; + tensor reshape_121 = reshape(shape = reshape_121_shape_0, x = slice_by_index_24)[name = tensor("reshape_121")]; + tensor reshape_122_shape_0 = const()[name = tensor("reshape_122_shape_0"), val = tensor([-1])]; + tensor reshape_122_cast_fp16 = reshape(shape = reshape_122_shape_0, x = var_5504_cast_fp16)[name = tensor("reshape_122_cast_fp16")]; + tensor reshape_123_shape_0 = const()[name = tensor("reshape_123_shape_0"), val = tensor([-1])]; + tensor reshape_123_cast_fp16 = reshape(shape = reshape_123_shape_0, x = cache12_to_fp16)[name = tensor("reshape_123_cast_fp16")]; + tensor scatter_24_mode_0 = const()[name = tensor("scatter_24_mode_0"), val = tensor("update")]; + tensor scatter_24_axis_0 = const()[name = tensor("scatter_24_axis_0"), val = tensor(0)]; + tensor scatter_24_validate_indices_0 = const()[name = tensor("scatter_24_validate_indices_0"), val = tensor(false)]; + tensor scatter_24_cast_fp16 = scatter(axis = scatter_24_axis_0, data = reshape_123_cast_fp16, indices = reshape_121, mode = scatter_24_mode_0, updates = reshape_122_cast_fp16, validate_indices = scatter_24_validate_indices_0)[name = tensor("scatter_24_cast_fp16")]; + tensor reshape_124_cast_fp16 = reshape(shape = shape_72, x = scatter_24_cast_fp16)[name = tensor("reshape_124_cast_fp16")]; + tensor var_5512_begin_0 = const()[name = tensor("op_5512_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor var_5512_end_0 = const()[name = tensor("op_5512_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor var_5512_end_mask_0 = const()[name = tensor("op_5512_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_5512_squeeze_mask_0 = const()[name = tensor("op_5512_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor var_5512_cast_fp16 = slice_by_index(begin = var_5512_begin_0, end = var_5512_end_0, end_mask = var_5512_end_mask_0, squeeze_mask = var_5512_squeeze_mask_0, x = reshape_124_cast_fp16)[name = tensor("op_5512_cast_fp16")]; + tensor var_5514_axis_0 = const()[name = tensor("op_5514_axis_0"), val = tensor(1)]; + tensor var_5514_mode_0 = const()[name = tensor("op_5514_mode_0"), val = tensor("update")]; + tensor var_5514_validate_indices_0 = const()[name = tensor("op_5514_validate_indices_0"), val = tensor(false)]; + tensor var_5514_cast_fp16 = scatter_along_axis(axis = var_5514_axis_0, data = var_5512_cast_fp16, indices = write_indices_25, mode = var_5514_mode_0, updates = v_25_cast_fp16, validate_indices = var_5514_validate_indices_0)[name = tensor("op_5514_cast_fp16")]; + tensor concat_88 = const()[name = tensor("concat_88"), val = tensor([1, 0, 0, 0, 0])]; + tensor concat_89 = const()[name = tensor("concat_89"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_25_internal_tensor_assign_2_stride_0 = const()[name = tensor("new_cache_25_internal_tensor_assign_2_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_25_internal_tensor_assign_2_begin_mask_0 = const()[name = tensor("new_cache_25_internal_tensor_assign_2_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_25_internal_tensor_assign_2_end_mask_0 = const()[name = tensor("new_cache_25_internal_tensor_assign_2_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_25_internal_tensor_assign_2_squeeze_mask_0 = const()[name = tensor("new_cache_25_internal_tensor_assign_2_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_73 = const()[name = tensor("shape_73"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_25 = const()[name = tensor("reduce_prod_25"), val = tensor(1048576)]; + tensor range_1d_25_start_0 = const()[name = tensor("range_1d_25_start_0"), val = tensor(0)]; + tensor range_1d_25_step_0 = const()[name = tensor("range_1d_25_step_0"), val = tensor(1)]; + tensor range_1d_25 = range_1d(end = reduce_prod_25, start = range_1d_25_start_0, step = range_1d_25_step_0)[name = tensor("range_1d_25")]; + tensor reshape_125 = reshape(shape = shape_73, x = range_1d_25)[name = tensor("reshape_125")]; + tensor slice_by_index_25 = slice_by_index(begin = concat_88, begin_mask = new_cache_25_internal_tensor_assign_2_begin_mask_0, end = concat_89, end_mask = new_cache_25_internal_tensor_assign_2_end_mask_0, squeeze_mask = new_cache_25_internal_tensor_assign_2_squeeze_mask_0, stride = new_cache_25_internal_tensor_assign_2_stride_0, x = reshape_125)[name = tensor("slice_by_index_25")]; + tensor reshape_126_shape_0 = const()[name = tensor("reshape_126_shape_0"), val = tensor([-1])]; + tensor reshape_126 = reshape(shape = reshape_126_shape_0, x = slice_by_index_25)[name = tensor("reshape_126")]; + tensor reshape_127_shape_0 = const()[name = tensor("reshape_127_shape_0"), val = tensor([-1])]; + tensor reshape_127_cast_fp16 = reshape(shape = reshape_127_shape_0, x = var_5514_cast_fp16)[name = tensor("reshape_127_cast_fp16")]; + tensor reshape_128_shape_0 = const()[name = tensor("reshape_128_shape_0"), val = tensor([-1])]; + tensor reshape_128_cast_fp16 = reshape(shape = reshape_128_shape_0, x = reshape_124_cast_fp16)[name = tensor("reshape_128_cast_fp16")]; + tensor scatter_25_mode_0 = const()[name = tensor("scatter_25_mode_0"), val = tensor("update")]; + tensor scatter_25_axis_0 = const()[name = tensor("scatter_25_axis_0"), val = tensor(0)]; + tensor scatter_25_validate_indices_0 = const()[name = tensor("scatter_25_validate_indices_0"), val = tensor(false)]; + tensor scatter_25_cast_fp16 = scatter(axis = scatter_25_axis_0, data = reshape_128_cast_fp16, indices = reshape_126, mode = scatter_25_mode_0, updates = reshape_127_cast_fp16, validate_indices = scatter_25_validate_indices_0)[name = tensor("scatter_25_cast_fp16")]; + tensor reshape_129_cast_fp16 = reshape(shape = shape_73, x = scatter_25_cast_fp16)[name = tensor("reshape_129_cast_fp16")]; + tensor reshape_129_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("reshape_129_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor keys_73_begin_0 = const()[name = tensor("keys_73_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor keys_73_end_0 = const()[name = tensor("keys_73_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor keys_73_end_mask_0 = const()[name = tensor("keys_73_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor keys_73_squeeze_mask_0 = const()[name = tensor("keys_73_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor keys_73_cast_fp16 = slice_by_index(begin = keys_73_begin_0, end = keys_73_end_0, end_mask = keys_73_end_mask_0, squeeze_mask = keys_73_squeeze_mask_0, x = reshape_129_cast_fp16)[name = tensor("keys_73_cast_fp16")]; + tensor values_73_begin_0 = const()[name = tensor("values_73_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor values_73_end_0 = const()[name = tensor("values_73_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor values_73_end_mask_0 = const()[name = tensor("values_73_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor values_73_squeeze_mask_0 = const()[name = tensor("values_73_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor values_73_cast_fp16 = slice_by_index(begin = values_73_begin_0, end = values_73_end_0, end_mask = values_73_end_mask_0, squeeze_mask = values_73_squeeze_mask_0, x = reshape_129_cast_fp16)[name = tensor("values_73_cast_fp16")]; + tensor var_5526_cast_fp16 = not_equal(x = keys_73_cast_fp16, y = keys_73_cast_fp16)[name = tensor("op_5526_cast_fp16")]; + tensor keys_75_cast_fp16 = select(a = var_504_to_fp16, b = keys_73_cast_fp16, cond = var_5526_cast_fp16)[name = tensor("keys_75_cast_fp16")]; + tensor var_5534_cast_fp16 = not_equal(x = values_73_cast_fp16, y = values_73_cast_fp16)[name = tensor("op_5534_cast_fp16")]; + tensor values_75_cast_fp16 = select(a = var_504_to_fp16, b = values_73_cast_fp16, cond = var_5534_cast_fp16)[name = tensor("values_75_cast_fp16")]; + tensor var_5558 = const()[name = tensor("op_5558"), val = tensor([0, 2, 1, 3])]; + tensor var_5571 = const()[name = tensor("op_5571"), val = tensor([1, 1, 1])]; + tensor var_5572_cast_fp16 = reshape(shape = var_5571, x = position12_to_fp16)[name = tensor("op_5572_cast_fp16")]; + tensor var_5589_to_fp16 = const()[name = tensor("op_5589_to_fp16"), val = tensor(0x1p+0)]; + tensor valid_len_25_cast_fp16 = add(x = var_5572_cast_fp16, y = var_5589_to_fp16)[name = tensor("valid_len_25_cast_fp16")]; + tensor valid_mask_25_cast_fp16 = less(x = k_positions_1_promoted_to_fp16, y = valid_len_25_cast_fp16)[name = tensor("valid_mask_25_cast_fp16")]; + tensor causal_mask_25_cast_fp16 = less_equal(x = k_positions_1_promoted_to_fp16, y = var_5572_cast_fp16)[name = tensor("causal_mask_25_cast_fp16")]; + tensor attn_mask_49 = logical_and(x = valid_mask_25_cast_fp16, y = causal_mask_25_cast_fp16)[name = tensor("attn_mask_49")]; + tensor attn_mask_51_axes_0 = const()[name = tensor("attn_mask_51_axes_0"), val = tensor([1])]; + tensor attn_mask_51 = expand_dims(axes = attn_mask_51_axes_0, x = attn_mask_49)[name = tensor("attn_mask_51")]; + tensor var_5607_transpose_x_0 = const()[name = tensor("op_5607_transpose_x_0"), val = tensor(false)]; + tensor var_5607_transpose_y_0 = const()[name = tensor("op_5607_transpose_y_0"), val = tensor(false)]; + tensor transpose_96_perm_0 = const()[name = tensor("transpose_96_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_97_perm_0 = const()[name = tensor("transpose_97_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_97 = transpose(perm = transpose_97_perm_0, x = keys_75_cast_fp16)[name = tensor("transpose_165")]; + tensor transpose_96 = transpose(perm = transpose_96_perm_0, x = q_75_cast_fp16)[name = tensor("transpose_166")]; + tensor var_5607_cast_fp16 = matmul(transpose_x = var_5607_transpose_x_0, transpose_y = var_5607_transpose_y_0, x = transpose_96, y = transpose_97)[name = tensor("op_5607_cast_fp16")]; + tensor var_5601_to_fp16 = const()[name = tensor("op_5601_to_fp16"), val = tensor([0x1p-3])]; + tensor attn_weights_73_cast_fp16 = mul(x = var_5607_cast_fp16, y = var_5601_to_fp16)[name = tensor("attn_weights_73_cast_fp16")]; + tensor var_5609 = logical_not(x = attn_mask_51)[name = tensor("op_5609")]; + tensor var_5610_to_fp16 = const()[name = tensor("op_5610_to_fp16"), val = tensor(-inf)]; + tensor attn_weights_75_cast_fp16 = select(a = var_5610_to_fp16, b = attn_weights_73_cast_fp16, cond = var_5609)[name = tensor("attn_weights_75_cast_fp16")]; + tensor var_5612 = const()[name = tensor("op_5612"), val = tensor(-1)]; + tensor attn_weights_77_cast_fp16 = softmax(axis = var_5612, x = attn_weights_75_cast_fp16)[name = tensor("attn_weights_77_cast_fp16")]; + tensor attn_output_25_transpose_x_0 = const()[name = tensor("attn_output_25_transpose_x_0"), val = tensor(false)]; + tensor attn_output_25_transpose_y_0 = const()[name = tensor("attn_output_25_transpose_y_0"), val = tensor(false)]; + tensor values_77_cast_fp16 = transpose(perm = var_5558, x = values_75_cast_fp16)[name = tensor("transpose_167")]; + tensor attn_output_25_cast_fp16 = matmul(transpose_x = attn_output_25_transpose_x_0, transpose_y = attn_output_25_transpose_y_0, x = attn_weights_77_cast_fp16, y = values_77_cast_fp16)[name = tensor("attn_output_25_cast_fp16")]; + tensor var_5620 = const()[name = tensor("op_5620"), val = tensor([0, 2, 1, 3])]; + tensor var_5623 = const()[name = tensor("op_5623"), val = tensor([1, 1, 1024])]; + tensor var_5621_cast_fp16 = transpose(perm = var_5620, x = attn_output_25_cast_fp16)[name = tensor("transpose_164")]; + tensor input_125_cast_fp16 = reshape(shape = var_5623, x = var_5621_cast_fp16)[name = tensor("input_125_cast_fp16")]; + tensor linear_50_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = attn12_out_proj_weight_cast_fp16, x = input_125_cast_fp16)[name = tensor("linear_50_cast_fp16")]; + tensor var_5629_to_fp16 = const()[name = tensor("op_5629_to_fp16"), val = tensor(0x1p+0)]; + tensor var_5630_cast_fp16 = add(x = position12_to_fp16, y = var_5629_to_fp16)[name = tensor("op_5630_cast_fp16")]; + tensor var_5630_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_5630_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor input_127_cast_fp16 = add(x = input_123_cast_fp16, y = linear_50_cast_fp16)[name = tensor("input_127_cast_fp16")]; + tensor input_129_axes_0 = const()[name = tensor("input_129_axes_0"), val = tensor([-1])]; + tensor norm12_2_weight_to_fp16 = const()[name = tensor("norm12_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(303692160)))]; + tensor norm12_2_bias_to_fp16 = const()[name = tensor("norm12_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(303694272)))]; + tensor var_5634_to_fp16 = const()[name = tensor("op_5634_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_129_cast_fp16 = layer_norm(axes = input_129_axes_0, beta = norm12_2_bias_to_fp16, epsilon = var_5634_to_fp16, gamma = norm12_2_weight_to_fp16, x = input_127_cast_fp16)[name = tensor("input_129_cast_fp16")]; + tensor linear_51_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = linear12_1_weight_cast_fp16, x = input_129_cast_fp16)[name = tensor("linear_51_cast_fp16")]; + tensor input_131_mode_0 = const()[name = tensor("input_131_mode_0"), val = tensor("EXACT")]; + tensor input_131_cast_fp16 = gelu(mode = input_131_mode_0, x = linear_51_cast_fp16)[name = tensor("input_131_cast_fp16")]; + tensor linear_52_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = linear12_2_weight_cast_fp16, x = input_131_cast_fp16)[name = tensor("linear_52_cast_fp16")]; + tensor input_133_cast_fp16 = add(x = input_127_cast_fp16, y = linear_52_cast_fp16)[name = tensor("input_133_cast_fp16")]; + tensor x_27_axes_0 = const()[name = tensor("x_27_axes_0"), val = tensor([-1])]; + tensor norm13_1_weight_to_fp16 = const()[name = tensor("norm13_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(303696384)))]; + tensor norm13_1_bias_to_fp16 = const()[name = tensor("norm13_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(303698496)))]; + tensor var_5651_to_fp16 = const()[name = tensor("op_5651_to_fp16"), val = tensor(0x1.5p-17)]; + tensor x_27_cast_fp16 = layer_norm(axes = x_27_axes_0, beta = norm13_1_bias_to_fp16, epsilon = var_5651_to_fp16, gamma = norm13_1_weight_to_fp16, x = input_133_cast_fp16)[name = tensor("x_27_cast_fp16")]; + tensor linear_53_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = attn13_in_proj_weight_cast_fp16, x = x_27_cast_fp16)[name = tensor("linear_53_cast_fp16")]; + tensor var_5687 = const()[name = tensor("op_5687"), val = tensor([1, 1, 3, 16, 64])]; + tensor qkv_27_cast_fp16 = reshape(shape = var_5687, x = linear_53_cast_fp16)[name = tensor("qkv_27_cast_fp16")]; + tensor q_79_begin_0 = const()[name = tensor("q_79_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor q_79_end_0 = const()[name = tensor("q_79_end_0"), val = tensor([1, 1, 1, 16, 64])]; + tensor q_79_end_mask_0 = const()[name = tensor("q_79_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor q_79_squeeze_mask_0 = const()[name = tensor("q_79_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor q_79_cast_fp16 = slice_by_index(begin = q_79_begin_0, end = q_79_end_0, end_mask = q_79_end_mask_0, squeeze_mask = q_79_squeeze_mask_0, x = qkv_27_cast_fp16)[name = tensor("q_79_cast_fp16")]; + tensor k_53_begin_0 = const()[name = tensor("k_53_begin_0"), val = tensor([0, 0, 1, 0, 0])]; + tensor k_53_end_0 = const()[name = tensor("k_53_end_0"), val = tensor([1, 1, 2, 16, 64])]; + tensor k_53_end_mask_0 = const()[name = tensor("k_53_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor k_53_squeeze_mask_0 = const()[name = tensor("k_53_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor k_53_cast_fp16 = slice_by_index(begin = k_53_begin_0, end = k_53_end_0, end_mask = k_53_end_mask_0, squeeze_mask = k_53_squeeze_mask_0, x = qkv_27_cast_fp16)[name = tensor("k_53_cast_fp16")]; + tensor v_27_begin_0 = const()[name = tensor("v_27_begin_0"), val = tensor([0, 0, 2, 0, 0])]; + tensor v_27_end_0 = const()[name = tensor("v_27_end_0"), val = tensor([1, 1, 3, 16, 64])]; + tensor v_27_end_mask_0 = const()[name = tensor("v_27_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor v_27_squeeze_mask_0 = const()[name = tensor("v_27_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor v_27_cast_fp16 = slice_by_index(begin = v_27_begin_0, end = v_27_end_0, end_mask = v_27_end_mask_0, squeeze_mask = v_27_squeeze_mask_0, x = qkv_27_cast_fp16)[name = tensor("v_27_cast_fp16")]; + tensor var_5791 = const()[name = tensor("op_5791"), val = tensor([1, 1, 1, 1])]; + tensor position13_to_fp16_dtype_0 = const()[name = tensor("position13_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor position13_to_fp16 = cast(dtype = position13_to_fp16_dtype_0, x = position13)[name = tensor("cast_488")]; + tensor ts_83_cast_fp16 = reshape(shape = var_5791, x = position13_to_fp16)[name = tensor("ts_83_cast_fp16")]; + tensor var_5795 = const()[name = tensor("op_5795"), val = tensor([1, 1, 16, 32, 2])]; + tensor q_complex_27_cast_fp16 = reshape(shape = var_5795, x = q_79_cast_fp16)[name = tensor("q_complex_27_cast_fp16")]; + tensor var_5799 = const()[name = tensor("op_5799"), val = tensor([1, 1, 16, 32, 2])]; + tensor k_complex_27_cast_fp16 = reshape(shape = var_5799, x = k_53_cast_fp16)[name = tensor("k_complex_27_cast_fp16")]; + tensor var_5803_begin_0 = const()[name = tensor("op_5803_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_5803_end_0 = const()[name = tensor("op_5803_end_0"), val = tensor([1, 1, 16, 32, 1])]; + tensor var_5803_end_mask_0 = const()[name = tensor("op_5803_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_5803_squeeze_mask_0 = const()[name = tensor("op_5803_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_5803_cast_fp16 = slice_by_index(begin = var_5803_begin_0, end = var_5803_end_0, end_mask = var_5803_end_mask_0, squeeze_mask = var_5803_squeeze_mask_0, x = q_complex_27_cast_fp16)[name = tensor("op_5803_cast_fp16")]; + tensor var_5811_begin_0 = const()[name = tensor("op_5811_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_5811_end_0 = const()[name = tensor("op_5811_end_0"), val = tensor([1, 1, 16, 32, 2])]; + tensor var_5811_end_mask_0 = const()[name = tensor("op_5811_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_5811_squeeze_mask_0 = const()[name = tensor("op_5811_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_5811_cast_fp16 = slice_by_index(begin = var_5811_begin_0, end = var_5811_end_0, end_mask = var_5811_end_mask_0, squeeze_mask = var_5811_squeeze_mask_0, x = q_complex_27_cast_fp16)[name = tensor("op_5811_cast_fp16")]; + tensor var_5819_begin_0 = const()[name = tensor("op_5819_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_5819_end_0 = const()[name = tensor("op_5819_end_0"), val = tensor([1, 1, 16, 32, 1])]; + tensor var_5819_end_mask_0 = const()[name = tensor("op_5819_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_5819_squeeze_mask_0 = const()[name = tensor("op_5819_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_5819_cast_fp16 = slice_by_index(begin = var_5819_begin_0, end = var_5819_end_0, end_mask = var_5819_end_mask_0, squeeze_mask = var_5819_squeeze_mask_0, x = k_complex_27_cast_fp16)[name = tensor("op_5819_cast_fp16")]; + tensor var_5827_begin_0 = const()[name = tensor("op_5827_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_5827_end_0 = const()[name = tensor("op_5827_end_0"), val = tensor([1, 1, 16, 32, 2])]; + tensor var_5827_end_mask_0 = const()[name = tensor("op_5827_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_5827_squeeze_mask_0 = const()[name = tensor("op_5827_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_5827_cast_fp16 = slice_by_index(begin = var_5827_begin_0, end = var_5827_end_0, end_mask = var_5827_end_mask_0, squeeze_mask = var_5827_squeeze_mask_0, x = k_complex_27_cast_fp16)[name = tensor("op_5827_cast_fp16")]; + tensor freqs_27_to_fp16 = const()[name = tensor("freqs_27_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(303700608)))]; + tensor var_5833_cast_fp16 = mul(x = freqs_27_to_fp16, y = ts_83_cast_fp16)[name = tensor("op_5833_cast_fp16")]; + tensor rotr_27_cast_fp16 = cos(x = var_5833_cast_fp16)[name = tensor("rotr_27_cast_fp16")]; + tensor roti_27_cast_fp16 = sin(x = var_5833_cast_fp16)[name = tensor("roti_27_cast_fp16")]; + tensor var_5837_cast_fp16 = mul(x = var_5803_cast_fp16, y = rotr_27_cast_fp16)[name = tensor("op_5837_cast_fp16")]; + tensor var_5838_cast_fp16 = mul(x = var_5811_cast_fp16, y = roti_27_cast_fp16)[name = tensor("op_5838_cast_fp16")]; + tensor qor_53_cast_fp16 = sub(x = var_5837_cast_fp16, y = var_5838_cast_fp16)[name = tensor("qor_53_cast_fp16")]; + tensor var_5841_cast_fp16 = mul(x = var_5803_cast_fp16, y = roti_27_cast_fp16)[name = tensor("op_5841_cast_fp16")]; + tensor var_5842_cast_fp16 = mul(x = var_5811_cast_fp16, y = rotr_27_cast_fp16)[name = tensor("op_5842_cast_fp16")]; + tensor qoi_53_cast_fp16 = add(x = var_5841_cast_fp16, y = var_5842_cast_fp16)[name = tensor("qoi_53_cast_fp16")]; + tensor var_5845_cast_fp16 = mul(x = var_5819_cast_fp16, y = rotr_27_cast_fp16)[name = tensor("op_5845_cast_fp16")]; + tensor var_5846_cast_fp16 = mul(x = var_5827_cast_fp16, y = roti_27_cast_fp16)[name = tensor("op_5846_cast_fp16")]; + tensor kor_53_cast_fp16 = sub(x = var_5845_cast_fp16, y = var_5846_cast_fp16)[name = tensor("kor_53_cast_fp16")]; + tensor var_5849_cast_fp16 = mul(x = var_5819_cast_fp16, y = roti_27_cast_fp16)[name = tensor("op_5849_cast_fp16")]; + tensor var_5850_cast_fp16 = mul(x = var_5827_cast_fp16, y = rotr_27_cast_fp16)[name = tensor("op_5850_cast_fp16")]; + tensor koi_53_cast_fp16 = add(x = var_5849_cast_fp16, y = var_5850_cast_fp16)[name = tensor("koi_53_cast_fp16")]; + tensor qo_27_axis_0 = const()[name = tensor("qo_27_axis_0"), val = tensor(-1)]; + tensor qo_27_cast_fp16 = stack(axis = qo_27_axis_0, values = (qor_53_cast_fp16, qoi_53_cast_fp16))[name = tensor("qo_27_cast_fp16")]; + tensor ko_27_axis_0 = const()[name = tensor("ko_27_axis_0"), val = tensor(-1)]; + tensor ko_27_cast_fp16 = stack(axis = ko_27_axis_0, values = (kor_53_cast_fp16, koi_53_cast_fp16))[name = tensor("ko_27_cast_fp16")]; + tensor var_5879 = const()[name = tensor("op_5879"), val = tensor([1, 1, 16, 64])]; + tensor q_81_cast_fp16 = reshape(shape = var_5879, x = qo_27_cast_fp16)[name = tensor("q_81_cast_fp16")]; + tensor var_5881 = const()[name = tensor("op_5881"), val = tensor([1, 1, 16, 64])]; + tensor k_55_cast_fp16 = reshape(shape = var_5881, x = ko_27_cast_fp16)[name = tensor("k_55_cast_fp16")]; + tensor _inversed_5903_y_0_to_fp16 = const()[name = tensor("_inversed_5903_y_0_to_fp16"), val = tensor(0x1p-9)]; + tensor _inversed_5903_cast_fp16 = mul(x = ts_83_cast_fp16, y = _inversed_5903_y_0_to_fp16)[name = tensor("_inversed_5903_cast_fp16")]; + tensor var_5904_cast_fp16 = floor(x = _inversed_5903_cast_fp16)[name = tensor("op_5904_cast_fp16")]; + tensor var_5905_to_fp16 = const()[name = tensor("op_5905_to_fp16"), val = tensor(0x1p+9)]; + tensor var_5906_cast_fp16 = mul(x = var_5904_cast_fp16, y = var_5905_to_fp16)[name = tensor("op_5906_cast_fp16")]; + tensor write_indices_float_55_cast_fp16 = sub(x = ts_83_cast_fp16, y = var_5906_cast_fp16)[name = tensor("write_indices_float_55_cast_fp16")]; + tensor var_5913_dtype_0 = const()[name = tensor("op_5913_dtype_0"), val = tensor("int32")]; + tensor write_indices_27_reps_0 = const()[name = tensor("write_indices_27_reps_0"), val = tensor([1, 1, 16, 64])]; + tensor write_indices_float_55_cast_fp16_to_int32 = cast(dtype = var_5913_dtype_0, x = write_indices_float_55_cast_fp16)[name = tensor("cast_487")]; + tensor write_indices_27 = tile(reps = write_indices_27_reps_0, x = write_indices_float_55_cast_fp16_to_int32)[name = tensor("write_indices_27")]; + tensor var_5921_begin_0 = const()[name = tensor("op_5921_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_5921_end_0 = const()[name = tensor("op_5921_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor var_5921_end_mask_0 = const()[name = tensor("op_5921_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_5921_squeeze_mask_0 = const()[name = tensor("op_5921_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor cache13_to_fp16_dtype_0 = const()[name = tensor("cache13_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor cache13_to_fp16 = cast(dtype = cache13_to_fp16_dtype_0, x = cache13)[name = tensor("cast_486")]; + tensor var_5921_cast_fp16 = slice_by_index(begin = var_5921_begin_0, end = var_5921_end_0, end_mask = var_5921_end_mask_0, squeeze_mask = var_5921_squeeze_mask_0, x = cache13_to_fp16)[name = tensor("op_5921_cast_fp16")]; + tensor var_5923_axis_0 = const()[name = tensor("op_5923_axis_0"), val = tensor(1)]; + tensor var_5923_mode_0 = const()[name = tensor("op_5923_mode_0"), val = tensor("update")]; + tensor var_5923_validate_indices_0 = const()[name = tensor("op_5923_validate_indices_0"), val = tensor(false)]; + tensor var_5923_cast_fp16 = scatter_along_axis(axis = var_5923_axis_0, data = var_5921_cast_fp16, indices = write_indices_27, mode = var_5923_mode_0, updates = k_55_cast_fp16, validate_indices = var_5923_validate_indices_0)[name = tensor("op_5923_cast_fp16")]; + tensor concat_93 = const()[name = tensor("concat_93"), val = tensor([0, 0, 0, 0, 0])]; + tensor concat_94 = const()[name = tensor("concat_94"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_27_internal_tensor_assign_1_stride_0 = const()[name = tensor("new_cache_27_internal_tensor_assign_1_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_27_internal_tensor_assign_1_begin_mask_0 = const()[name = tensor("new_cache_27_internal_tensor_assign_1_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_27_internal_tensor_assign_1_end_mask_0 = const()[name = tensor("new_cache_27_internal_tensor_assign_1_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_27_internal_tensor_assign_1_squeeze_mask_0 = const()[name = tensor("new_cache_27_internal_tensor_assign_1_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_74 = const()[name = tensor("shape_74"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_26 = const()[name = tensor("reduce_prod_26"), val = tensor(1048576)]; + tensor range_1d_26_start_0 = const()[name = tensor("range_1d_26_start_0"), val = tensor(0)]; + tensor range_1d_26_step_0 = const()[name = tensor("range_1d_26_step_0"), val = tensor(1)]; + tensor range_1d_26 = range_1d(end = reduce_prod_26, start = range_1d_26_start_0, step = range_1d_26_step_0)[name = tensor("range_1d_26")]; + tensor reshape_130 = reshape(shape = shape_74, x = range_1d_26)[name = tensor("reshape_130")]; + tensor slice_by_index_26 = slice_by_index(begin = concat_93, begin_mask = new_cache_27_internal_tensor_assign_1_begin_mask_0, end = concat_94, end_mask = new_cache_27_internal_tensor_assign_1_end_mask_0, squeeze_mask = new_cache_27_internal_tensor_assign_1_squeeze_mask_0, stride = new_cache_27_internal_tensor_assign_1_stride_0, x = reshape_130)[name = tensor("slice_by_index_26")]; + tensor reshape_131_shape_0 = const()[name = tensor("reshape_131_shape_0"), val = tensor([-1])]; + tensor reshape_131 = reshape(shape = reshape_131_shape_0, x = slice_by_index_26)[name = tensor("reshape_131")]; + tensor reshape_132_shape_0 = const()[name = tensor("reshape_132_shape_0"), val = tensor([-1])]; + tensor reshape_132_cast_fp16 = reshape(shape = reshape_132_shape_0, x = var_5923_cast_fp16)[name = tensor("reshape_132_cast_fp16")]; + tensor reshape_133_shape_0 = const()[name = tensor("reshape_133_shape_0"), val = tensor([-1])]; + tensor reshape_133_cast_fp16 = reshape(shape = reshape_133_shape_0, x = cache13_to_fp16)[name = tensor("reshape_133_cast_fp16")]; + tensor scatter_26_mode_0 = const()[name = tensor("scatter_26_mode_0"), val = tensor("update")]; + tensor scatter_26_axis_0 = const()[name = tensor("scatter_26_axis_0"), val = tensor(0)]; + tensor scatter_26_validate_indices_0 = const()[name = tensor("scatter_26_validate_indices_0"), val = tensor(false)]; + tensor scatter_26_cast_fp16 = scatter(axis = scatter_26_axis_0, data = reshape_133_cast_fp16, indices = reshape_131, mode = scatter_26_mode_0, updates = reshape_132_cast_fp16, validate_indices = scatter_26_validate_indices_0)[name = tensor("scatter_26_cast_fp16")]; + tensor reshape_134_cast_fp16 = reshape(shape = shape_74, x = scatter_26_cast_fp16)[name = tensor("reshape_134_cast_fp16")]; + tensor var_5931_begin_0 = const()[name = tensor("op_5931_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor var_5931_end_0 = const()[name = tensor("op_5931_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor var_5931_end_mask_0 = const()[name = tensor("op_5931_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_5931_squeeze_mask_0 = const()[name = tensor("op_5931_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor var_5931_cast_fp16 = slice_by_index(begin = var_5931_begin_0, end = var_5931_end_0, end_mask = var_5931_end_mask_0, squeeze_mask = var_5931_squeeze_mask_0, x = reshape_134_cast_fp16)[name = tensor("op_5931_cast_fp16")]; + tensor var_5933_axis_0 = const()[name = tensor("op_5933_axis_0"), val = tensor(1)]; + tensor var_5933_mode_0 = const()[name = tensor("op_5933_mode_0"), val = tensor("update")]; + tensor var_5933_validate_indices_0 = const()[name = tensor("op_5933_validate_indices_0"), val = tensor(false)]; + tensor var_5933_cast_fp16 = scatter_along_axis(axis = var_5933_axis_0, data = var_5931_cast_fp16, indices = write_indices_27, mode = var_5933_mode_0, updates = v_27_cast_fp16, validate_indices = var_5933_validate_indices_0)[name = tensor("op_5933_cast_fp16")]; + tensor concat_95 = const()[name = tensor("concat_95"), val = tensor([1, 0, 0, 0, 0])]; + tensor concat_96 = const()[name = tensor("concat_96"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_27_internal_tensor_assign_2_stride_0 = const()[name = tensor("new_cache_27_internal_tensor_assign_2_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_27_internal_tensor_assign_2_begin_mask_0 = const()[name = tensor("new_cache_27_internal_tensor_assign_2_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_27_internal_tensor_assign_2_end_mask_0 = const()[name = tensor("new_cache_27_internal_tensor_assign_2_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_27_internal_tensor_assign_2_squeeze_mask_0 = const()[name = tensor("new_cache_27_internal_tensor_assign_2_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_75 = const()[name = tensor("shape_75"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_27 = const()[name = tensor("reduce_prod_27"), val = tensor(1048576)]; + tensor range_1d_27_start_0 = const()[name = tensor("range_1d_27_start_0"), val = tensor(0)]; + tensor range_1d_27_step_0 = const()[name = tensor("range_1d_27_step_0"), val = tensor(1)]; + tensor range_1d_27 = range_1d(end = reduce_prod_27, start = range_1d_27_start_0, step = range_1d_27_step_0)[name = tensor("range_1d_27")]; + tensor reshape_135 = reshape(shape = shape_75, x = range_1d_27)[name = tensor("reshape_135")]; + tensor slice_by_index_27 = slice_by_index(begin = concat_95, begin_mask = new_cache_27_internal_tensor_assign_2_begin_mask_0, end = concat_96, end_mask = new_cache_27_internal_tensor_assign_2_end_mask_0, squeeze_mask = new_cache_27_internal_tensor_assign_2_squeeze_mask_0, stride = new_cache_27_internal_tensor_assign_2_stride_0, x = reshape_135)[name = tensor("slice_by_index_27")]; + tensor reshape_136_shape_0 = const()[name = tensor("reshape_136_shape_0"), val = tensor([-1])]; + tensor reshape_136 = reshape(shape = reshape_136_shape_0, x = slice_by_index_27)[name = tensor("reshape_136")]; + tensor reshape_137_shape_0 = const()[name = tensor("reshape_137_shape_0"), val = tensor([-1])]; + tensor reshape_137_cast_fp16 = reshape(shape = reshape_137_shape_0, x = var_5933_cast_fp16)[name = tensor("reshape_137_cast_fp16")]; + tensor reshape_138_shape_0 = const()[name = tensor("reshape_138_shape_0"), val = tensor([-1])]; + tensor reshape_138_cast_fp16 = reshape(shape = reshape_138_shape_0, x = reshape_134_cast_fp16)[name = tensor("reshape_138_cast_fp16")]; + tensor scatter_27_mode_0 = const()[name = tensor("scatter_27_mode_0"), val = tensor("update")]; + tensor scatter_27_axis_0 = const()[name = tensor("scatter_27_axis_0"), val = tensor(0)]; + tensor scatter_27_validate_indices_0 = const()[name = tensor("scatter_27_validate_indices_0"), val = tensor(false)]; + tensor scatter_27_cast_fp16 = scatter(axis = scatter_27_axis_0, data = reshape_138_cast_fp16, indices = reshape_136, mode = scatter_27_mode_0, updates = reshape_137_cast_fp16, validate_indices = scatter_27_validate_indices_0)[name = tensor("scatter_27_cast_fp16")]; + tensor reshape_139_cast_fp16 = reshape(shape = shape_75, x = scatter_27_cast_fp16)[name = tensor("reshape_139_cast_fp16")]; + tensor reshape_139_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("reshape_139_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor keys_79_begin_0 = const()[name = tensor("keys_79_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor keys_79_end_0 = const()[name = tensor("keys_79_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor keys_79_end_mask_0 = const()[name = tensor("keys_79_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor keys_79_squeeze_mask_0 = const()[name = tensor("keys_79_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor keys_79_cast_fp16 = slice_by_index(begin = keys_79_begin_0, end = keys_79_end_0, end_mask = keys_79_end_mask_0, squeeze_mask = keys_79_squeeze_mask_0, x = reshape_139_cast_fp16)[name = tensor("keys_79_cast_fp16")]; + tensor values_79_begin_0 = const()[name = tensor("values_79_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor values_79_end_0 = const()[name = tensor("values_79_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor values_79_end_mask_0 = const()[name = tensor("values_79_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor values_79_squeeze_mask_0 = const()[name = tensor("values_79_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor values_79_cast_fp16 = slice_by_index(begin = values_79_begin_0, end = values_79_end_0, end_mask = values_79_end_mask_0, squeeze_mask = values_79_squeeze_mask_0, x = reshape_139_cast_fp16)[name = tensor("values_79_cast_fp16")]; + tensor var_5945_cast_fp16 = not_equal(x = keys_79_cast_fp16, y = keys_79_cast_fp16)[name = tensor("op_5945_cast_fp16")]; + tensor keys_81_cast_fp16 = select(a = var_504_to_fp16, b = keys_79_cast_fp16, cond = var_5945_cast_fp16)[name = tensor("keys_81_cast_fp16")]; + tensor var_5953_cast_fp16 = not_equal(x = values_79_cast_fp16, y = values_79_cast_fp16)[name = tensor("op_5953_cast_fp16")]; + tensor values_81_cast_fp16 = select(a = var_504_to_fp16, b = values_79_cast_fp16, cond = var_5953_cast_fp16)[name = tensor("values_81_cast_fp16")]; + tensor var_5977 = const()[name = tensor("op_5977"), val = tensor([0, 2, 1, 3])]; + tensor var_5990 = const()[name = tensor("op_5990"), val = tensor([1, 1, 1])]; + tensor var_5991_cast_fp16 = reshape(shape = var_5990, x = position13_to_fp16)[name = tensor("op_5991_cast_fp16")]; + tensor var_6008_to_fp16 = const()[name = tensor("op_6008_to_fp16"), val = tensor(0x1p+0)]; + tensor valid_len_27_cast_fp16 = add(x = var_5991_cast_fp16, y = var_6008_to_fp16)[name = tensor("valid_len_27_cast_fp16")]; + tensor valid_mask_27_cast_fp16 = less(x = k_positions_1_promoted_to_fp16, y = valid_len_27_cast_fp16)[name = tensor("valid_mask_27_cast_fp16")]; + tensor causal_mask_27_cast_fp16 = less_equal(x = k_positions_1_promoted_to_fp16, y = var_5991_cast_fp16)[name = tensor("causal_mask_27_cast_fp16")]; + tensor attn_mask_53 = logical_and(x = valid_mask_27_cast_fp16, y = causal_mask_27_cast_fp16)[name = tensor("attn_mask_53")]; + tensor attn_mask_55_axes_0 = const()[name = tensor("attn_mask_55_axes_0"), val = tensor([1])]; + tensor attn_mask_55 = expand_dims(axes = attn_mask_55_axes_0, x = attn_mask_53)[name = tensor("attn_mask_55")]; + tensor var_6026_transpose_x_0 = const()[name = tensor("op_6026_transpose_x_0"), val = tensor(false)]; + tensor var_6026_transpose_y_0 = const()[name = tensor("op_6026_transpose_y_0"), val = tensor(false)]; + tensor transpose_98_perm_0 = const()[name = tensor("transpose_98_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_99_perm_0 = const()[name = tensor("transpose_99_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_99 = transpose(perm = transpose_99_perm_0, x = keys_81_cast_fp16)[name = tensor("transpose_161")]; + tensor transpose_98 = transpose(perm = transpose_98_perm_0, x = q_81_cast_fp16)[name = tensor("transpose_162")]; + tensor var_6026_cast_fp16 = matmul(transpose_x = var_6026_transpose_x_0, transpose_y = var_6026_transpose_y_0, x = transpose_98, y = transpose_99)[name = tensor("op_6026_cast_fp16")]; + tensor var_6020_to_fp16 = const()[name = tensor("op_6020_to_fp16"), val = tensor([0x1p-3])]; + tensor attn_weights_79_cast_fp16 = mul(x = var_6026_cast_fp16, y = var_6020_to_fp16)[name = tensor("attn_weights_79_cast_fp16")]; + tensor var_6028 = logical_not(x = attn_mask_55)[name = tensor("op_6028")]; + tensor var_6029_to_fp16 = const()[name = tensor("op_6029_to_fp16"), val = tensor(-inf)]; + tensor attn_weights_81_cast_fp16 = select(a = var_6029_to_fp16, b = attn_weights_79_cast_fp16, cond = var_6028)[name = tensor("attn_weights_81_cast_fp16")]; + tensor var_6031 = const()[name = tensor("op_6031"), val = tensor(-1)]; + tensor attn_weights_83_cast_fp16 = softmax(axis = var_6031, x = attn_weights_81_cast_fp16)[name = tensor("attn_weights_83_cast_fp16")]; + tensor attn_output_27_transpose_x_0 = const()[name = tensor("attn_output_27_transpose_x_0"), val = tensor(false)]; + tensor attn_output_27_transpose_y_0 = const()[name = tensor("attn_output_27_transpose_y_0"), val = tensor(false)]; + tensor values_83_cast_fp16 = transpose(perm = var_5977, x = values_81_cast_fp16)[name = tensor("transpose_163")]; + tensor attn_output_27_cast_fp16 = matmul(transpose_x = attn_output_27_transpose_x_0, transpose_y = attn_output_27_transpose_y_0, x = attn_weights_83_cast_fp16, y = values_83_cast_fp16)[name = tensor("attn_output_27_cast_fp16")]; + tensor var_6039 = const()[name = tensor("op_6039"), val = tensor([0, 2, 1, 3])]; + tensor var_6042 = const()[name = tensor("op_6042"), val = tensor([1, 1, 1024])]; + tensor var_6040_cast_fp16 = transpose(perm = var_6039, x = attn_output_27_cast_fp16)[name = tensor("transpose_160")]; + tensor input_135_cast_fp16 = reshape(shape = var_6042, x = var_6040_cast_fp16)[name = tensor("input_135_cast_fp16")]; + tensor linear_54_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = attn13_out_proj_weight_cast_fp16, x = input_135_cast_fp16)[name = tensor("linear_54_cast_fp16")]; + tensor var_6048_to_fp16 = const()[name = tensor("op_6048_to_fp16"), val = tensor(0x1p+0)]; + tensor var_6049_cast_fp16 = add(x = position13_to_fp16, y = var_6048_to_fp16)[name = tensor("op_6049_cast_fp16")]; + tensor var_6049_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_6049_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor input_137_cast_fp16 = add(x = input_133_cast_fp16, y = linear_54_cast_fp16)[name = tensor("input_137_cast_fp16")]; + tensor input_139_axes_0 = const()[name = tensor("input_139_axes_0"), val = tensor([-1])]; + tensor norm13_2_weight_to_fp16 = const()[name = tensor("norm13_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(303700736)))]; + tensor norm13_2_bias_to_fp16 = const()[name = tensor("norm13_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(303702848)))]; + tensor var_6053_to_fp16 = const()[name = tensor("op_6053_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_139_cast_fp16 = layer_norm(axes = input_139_axes_0, beta = norm13_2_bias_to_fp16, epsilon = var_6053_to_fp16, gamma = norm13_2_weight_to_fp16, x = input_137_cast_fp16)[name = tensor("input_139_cast_fp16")]; + tensor linear_55_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = linear13_1_weight_cast_fp16, x = input_139_cast_fp16)[name = tensor("linear_55_cast_fp16")]; + tensor input_141_mode_0 = const()[name = tensor("input_141_mode_0"), val = tensor("EXACT")]; + tensor input_141_cast_fp16 = gelu(mode = input_141_mode_0, x = linear_55_cast_fp16)[name = tensor("input_141_cast_fp16")]; + tensor linear_56_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = linear13_2_weight_cast_fp16, x = input_141_cast_fp16)[name = tensor("linear_56_cast_fp16")]; + tensor input_143_cast_fp16 = add(x = input_137_cast_fp16, y = linear_56_cast_fp16)[name = tensor("input_143_cast_fp16")]; + tensor x_29_axes_0 = const()[name = tensor("x_29_axes_0"), val = tensor([-1])]; + tensor norm14_1_weight_to_fp16 = const()[name = tensor("norm14_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(303704960)))]; + tensor norm14_1_bias_to_fp16 = const()[name = tensor("norm14_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(303707072)))]; + tensor var_6070_to_fp16 = const()[name = tensor("op_6070_to_fp16"), val = tensor(0x1.5p-17)]; + tensor x_29_cast_fp16 = layer_norm(axes = x_29_axes_0, beta = norm14_1_bias_to_fp16, epsilon = var_6070_to_fp16, gamma = norm14_1_weight_to_fp16, x = input_143_cast_fp16)[name = tensor("x_29_cast_fp16")]; + tensor linear_57_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = attn14_in_proj_weight_cast_fp16, x = x_29_cast_fp16)[name = tensor("linear_57_cast_fp16")]; + tensor var_6106 = const()[name = tensor("op_6106"), val = tensor([1, 1, 3, 16, 64])]; + tensor qkv_29_cast_fp16 = reshape(shape = var_6106, x = linear_57_cast_fp16)[name = tensor("qkv_29_cast_fp16")]; + tensor q_85_begin_0 = const()[name = tensor("q_85_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor q_85_end_0 = const()[name = tensor("q_85_end_0"), val = tensor([1, 1, 1, 16, 64])]; + tensor q_85_end_mask_0 = const()[name = tensor("q_85_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor q_85_squeeze_mask_0 = const()[name = tensor("q_85_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor q_85_cast_fp16 = slice_by_index(begin = q_85_begin_0, end = q_85_end_0, end_mask = q_85_end_mask_0, squeeze_mask = q_85_squeeze_mask_0, x = qkv_29_cast_fp16)[name = tensor("q_85_cast_fp16")]; + tensor k_57_begin_0 = const()[name = tensor("k_57_begin_0"), val = tensor([0, 0, 1, 0, 0])]; + tensor k_57_end_0 = const()[name = tensor("k_57_end_0"), val = tensor([1, 1, 2, 16, 64])]; + tensor k_57_end_mask_0 = const()[name = tensor("k_57_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor k_57_squeeze_mask_0 = const()[name = tensor("k_57_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor k_57_cast_fp16 = slice_by_index(begin = k_57_begin_0, end = k_57_end_0, end_mask = k_57_end_mask_0, squeeze_mask = k_57_squeeze_mask_0, x = qkv_29_cast_fp16)[name = tensor("k_57_cast_fp16")]; + tensor v_29_begin_0 = const()[name = tensor("v_29_begin_0"), val = tensor([0, 0, 2, 0, 0])]; + tensor v_29_end_0 = const()[name = tensor("v_29_end_0"), val = tensor([1, 1, 3, 16, 64])]; + tensor v_29_end_mask_0 = const()[name = tensor("v_29_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor v_29_squeeze_mask_0 = const()[name = tensor("v_29_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor v_29_cast_fp16 = slice_by_index(begin = v_29_begin_0, end = v_29_end_0, end_mask = v_29_end_mask_0, squeeze_mask = v_29_squeeze_mask_0, x = qkv_29_cast_fp16)[name = tensor("v_29_cast_fp16")]; + tensor var_6210 = const()[name = tensor("op_6210"), val = tensor([1, 1, 1, 1])]; + tensor position14_to_fp16_dtype_0 = const()[name = tensor("position14_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor position14_to_fp16 = cast(dtype = position14_to_fp16_dtype_0, x = position14)[name = tensor("cast_483")]; + tensor ts_89_cast_fp16 = reshape(shape = var_6210, x = position14_to_fp16)[name = tensor("ts_89_cast_fp16")]; + tensor var_6214 = const()[name = tensor("op_6214"), val = tensor([1, 1, 16, 32, 2])]; + tensor q_complex_29_cast_fp16 = reshape(shape = var_6214, x = q_85_cast_fp16)[name = tensor("q_complex_29_cast_fp16")]; + tensor var_6218 = const()[name = tensor("op_6218"), val = tensor([1, 1, 16, 32, 2])]; + tensor k_complex_29_cast_fp16 = reshape(shape = var_6218, x = k_57_cast_fp16)[name = tensor("k_complex_29_cast_fp16")]; + tensor var_6222_begin_0 = const()[name = tensor("op_6222_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_6222_end_0 = const()[name = tensor("op_6222_end_0"), val = tensor([1, 1, 16, 32, 1])]; + tensor var_6222_end_mask_0 = const()[name = tensor("op_6222_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_6222_squeeze_mask_0 = const()[name = tensor("op_6222_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_6222_cast_fp16 = slice_by_index(begin = var_6222_begin_0, end = var_6222_end_0, end_mask = var_6222_end_mask_0, squeeze_mask = var_6222_squeeze_mask_0, x = q_complex_29_cast_fp16)[name = tensor("op_6222_cast_fp16")]; + tensor var_6230_begin_0 = const()[name = tensor("op_6230_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_6230_end_0 = const()[name = tensor("op_6230_end_0"), val = tensor([1, 1, 16, 32, 2])]; + tensor var_6230_end_mask_0 = const()[name = tensor("op_6230_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_6230_squeeze_mask_0 = const()[name = tensor("op_6230_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_6230_cast_fp16 = slice_by_index(begin = var_6230_begin_0, end = var_6230_end_0, end_mask = var_6230_end_mask_0, squeeze_mask = var_6230_squeeze_mask_0, x = q_complex_29_cast_fp16)[name = tensor("op_6230_cast_fp16")]; + tensor var_6238_begin_0 = const()[name = tensor("op_6238_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_6238_end_0 = const()[name = tensor("op_6238_end_0"), val = tensor([1, 1, 16, 32, 1])]; + tensor var_6238_end_mask_0 = const()[name = tensor("op_6238_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_6238_squeeze_mask_0 = const()[name = tensor("op_6238_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_6238_cast_fp16 = slice_by_index(begin = var_6238_begin_0, end = var_6238_end_0, end_mask = var_6238_end_mask_0, squeeze_mask = var_6238_squeeze_mask_0, x = k_complex_29_cast_fp16)[name = tensor("op_6238_cast_fp16")]; + tensor var_6246_begin_0 = const()[name = tensor("op_6246_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_6246_end_0 = const()[name = tensor("op_6246_end_0"), val = tensor([1, 1, 16, 32, 2])]; + tensor var_6246_end_mask_0 = const()[name = tensor("op_6246_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_6246_squeeze_mask_0 = const()[name = tensor("op_6246_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_6246_cast_fp16 = slice_by_index(begin = var_6246_begin_0, end = var_6246_end_0, end_mask = var_6246_end_mask_0, squeeze_mask = var_6246_squeeze_mask_0, x = k_complex_29_cast_fp16)[name = tensor("op_6246_cast_fp16")]; + tensor freqs_29_to_fp16 = const()[name = tensor("freqs_29_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(303709184)))]; + tensor var_6252_cast_fp16 = mul(x = freqs_29_to_fp16, y = ts_89_cast_fp16)[name = tensor("op_6252_cast_fp16")]; + tensor rotr_29_cast_fp16 = cos(x = var_6252_cast_fp16)[name = tensor("rotr_29_cast_fp16")]; + tensor roti_29_cast_fp16 = sin(x = var_6252_cast_fp16)[name = tensor("roti_29_cast_fp16")]; + tensor var_6256_cast_fp16 = mul(x = var_6222_cast_fp16, y = rotr_29_cast_fp16)[name = tensor("op_6256_cast_fp16")]; + tensor var_6257_cast_fp16 = mul(x = var_6230_cast_fp16, y = roti_29_cast_fp16)[name = tensor("op_6257_cast_fp16")]; + tensor qor_57_cast_fp16 = sub(x = var_6256_cast_fp16, y = var_6257_cast_fp16)[name = tensor("qor_57_cast_fp16")]; + tensor var_6260_cast_fp16 = mul(x = var_6222_cast_fp16, y = roti_29_cast_fp16)[name = tensor("op_6260_cast_fp16")]; + tensor var_6261_cast_fp16 = mul(x = var_6230_cast_fp16, y = rotr_29_cast_fp16)[name = tensor("op_6261_cast_fp16")]; + tensor qoi_57_cast_fp16 = add(x = var_6260_cast_fp16, y = var_6261_cast_fp16)[name = tensor("qoi_57_cast_fp16")]; + tensor var_6264_cast_fp16 = mul(x = var_6238_cast_fp16, y = rotr_29_cast_fp16)[name = tensor("op_6264_cast_fp16")]; + tensor var_6265_cast_fp16 = mul(x = var_6246_cast_fp16, y = roti_29_cast_fp16)[name = tensor("op_6265_cast_fp16")]; + tensor kor_57_cast_fp16 = sub(x = var_6264_cast_fp16, y = var_6265_cast_fp16)[name = tensor("kor_57_cast_fp16")]; + tensor var_6268_cast_fp16 = mul(x = var_6238_cast_fp16, y = roti_29_cast_fp16)[name = tensor("op_6268_cast_fp16")]; + tensor var_6269_cast_fp16 = mul(x = var_6246_cast_fp16, y = rotr_29_cast_fp16)[name = tensor("op_6269_cast_fp16")]; + tensor koi_57_cast_fp16 = add(x = var_6268_cast_fp16, y = var_6269_cast_fp16)[name = tensor("koi_57_cast_fp16")]; + tensor qo_29_axis_0 = const()[name = tensor("qo_29_axis_0"), val = tensor(-1)]; + tensor qo_29_cast_fp16 = stack(axis = qo_29_axis_0, values = (qor_57_cast_fp16, qoi_57_cast_fp16))[name = tensor("qo_29_cast_fp16")]; + tensor ko_29_axis_0 = const()[name = tensor("ko_29_axis_0"), val = tensor(-1)]; + tensor ko_29_cast_fp16 = stack(axis = ko_29_axis_0, values = (kor_57_cast_fp16, koi_57_cast_fp16))[name = tensor("ko_29_cast_fp16")]; + tensor var_6298 = const()[name = tensor("op_6298"), val = tensor([1, 1, 16, 64])]; + tensor q_87_cast_fp16 = reshape(shape = var_6298, x = qo_29_cast_fp16)[name = tensor("q_87_cast_fp16")]; + tensor var_6300 = const()[name = tensor("op_6300"), val = tensor([1, 1, 16, 64])]; + tensor k_59_cast_fp16 = reshape(shape = var_6300, x = ko_29_cast_fp16)[name = tensor("k_59_cast_fp16")]; + tensor _inversed_6322_y_0_to_fp16 = const()[name = tensor("_inversed_6322_y_0_to_fp16"), val = tensor(0x1p-9)]; + tensor _inversed_6322_cast_fp16 = mul(x = ts_89_cast_fp16, y = _inversed_6322_y_0_to_fp16)[name = tensor("_inversed_6322_cast_fp16")]; + tensor var_6323_cast_fp16 = floor(x = _inversed_6322_cast_fp16)[name = tensor("op_6323_cast_fp16")]; + tensor var_6324_to_fp16 = const()[name = tensor("op_6324_to_fp16"), val = tensor(0x1p+9)]; + tensor var_6325_cast_fp16 = mul(x = var_6323_cast_fp16, y = var_6324_to_fp16)[name = tensor("op_6325_cast_fp16")]; + tensor write_indices_float_59_cast_fp16 = sub(x = ts_89_cast_fp16, y = var_6325_cast_fp16)[name = tensor("write_indices_float_59_cast_fp16")]; + tensor var_6332_dtype_0 = const()[name = tensor("op_6332_dtype_0"), val = tensor("int32")]; + tensor write_indices_29_reps_0 = const()[name = tensor("write_indices_29_reps_0"), val = tensor([1, 1, 16, 64])]; + tensor write_indices_float_59_cast_fp16_to_int32 = cast(dtype = var_6332_dtype_0, x = write_indices_float_59_cast_fp16)[name = tensor("cast_482")]; + tensor write_indices_29 = tile(reps = write_indices_29_reps_0, x = write_indices_float_59_cast_fp16_to_int32)[name = tensor("write_indices_29")]; + tensor var_6340_begin_0 = const()[name = tensor("op_6340_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_6340_end_0 = const()[name = tensor("op_6340_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor var_6340_end_mask_0 = const()[name = tensor("op_6340_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_6340_squeeze_mask_0 = const()[name = tensor("op_6340_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor cache14_to_fp16_dtype_0 = const()[name = tensor("cache14_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor cache14_to_fp16 = cast(dtype = cache14_to_fp16_dtype_0, x = cache14)[name = tensor("cast_481")]; + tensor var_6340_cast_fp16 = slice_by_index(begin = var_6340_begin_0, end = var_6340_end_0, end_mask = var_6340_end_mask_0, squeeze_mask = var_6340_squeeze_mask_0, x = cache14_to_fp16)[name = tensor("op_6340_cast_fp16")]; + tensor var_6342_axis_0 = const()[name = tensor("op_6342_axis_0"), val = tensor(1)]; + tensor var_6342_mode_0 = const()[name = tensor("op_6342_mode_0"), val = tensor("update")]; + tensor var_6342_validate_indices_0 = const()[name = tensor("op_6342_validate_indices_0"), val = tensor(false)]; + tensor var_6342_cast_fp16 = scatter_along_axis(axis = var_6342_axis_0, data = var_6340_cast_fp16, indices = write_indices_29, mode = var_6342_mode_0, updates = k_59_cast_fp16, validate_indices = var_6342_validate_indices_0)[name = tensor("op_6342_cast_fp16")]; + tensor concat_100 = const()[name = tensor("concat_100"), val = tensor([0, 0, 0, 0, 0])]; + tensor concat_101 = const()[name = tensor("concat_101"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_29_internal_tensor_assign_1_stride_0 = const()[name = tensor("new_cache_29_internal_tensor_assign_1_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_29_internal_tensor_assign_1_begin_mask_0 = const()[name = tensor("new_cache_29_internal_tensor_assign_1_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_29_internal_tensor_assign_1_end_mask_0 = const()[name = tensor("new_cache_29_internal_tensor_assign_1_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_29_internal_tensor_assign_1_squeeze_mask_0 = const()[name = tensor("new_cache_29_internal_tensor_assign_1_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_76 = const()[name = tensor("shape_76"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_28 = const()[name = tensor("reduce_prod_28"), val = tensor(1048576)]; + tensor range_1d_28_start_0 = const()[name = tensor("range_1d_28_start_0"), val = tensor(0)]; + tensor range_1d_28_step_0 = const()[name = tensor("range_1d_28_step_0"), val = tensor(1)]; + tensor range_1d_28 = range_1d(end = reduce_prod_28, start = range_1d_28_start_0, step = range_1d_28_step_0)[name = tensor("range_1d_28")]; + tensor reshape_140 = reshape(shape = shape_76, x = range_1d_28)[name = tensor("reshape_140")]; + tensor slice_by_index_28 = slice_by_index(begin = concat_100, begin_mask = new_cache_29_internal_tensor_assign_1_begin_mask_0, end = concat_101, end_mask = new_cache_29_internal_tensor_assign_1_end_mask_0, squeeze_mask = new_cache_29_internal_tensor_assign_1_squeeze_mask_0, stride = new_cache_29_internal_tensor_assign_1_stride_0, x = reshape_140)[name = tensor("slice_by_index_28")]; + tensor reshape_141_shape_0 = const()[name = tensor("reshape_141_shape_0"), val = tensor([-1])]; + tensor reshape_141 = reshape(shape = reshape_141_shape_0, x = slice_by_index_28)[name = tensor("reshape_141")]; + tensor reshape_142_shape_0 = const()[name = tensor("reshape_142_shape_0"), val = tensor([-1])]; + tensor reshape_142_cast_fp16 = reshape(shape = reshape_142_shape_0, x = var_6342_cast_fp16)[name = tensor("reshape_142_cast_fp16")]; + tensor reshape_143_shape_0 = const()[name = tensor("reshape_143_shape_0"), val = tensor([-1])]; + tensor reshape_143_cast_fp16 = reshape(shape = reshape_143_shape_0, x = cache14_to_fp16)[name = tensor("reshape_143_cast_fp16")]; + tensor scatter_28_mode_0 = const()[name = tensor("scatter_28_mode_0"), val = tensor("update")]; + tensor scatter_28_axis_0 = const()[name = tensor("scatter_28_axis_0"), val = tensor(0)]; + tensor scatter_28_validate_indices_0 = const()[name = tensor("scatter_28_validate_indices_0"), val = tensor(false)]; + tensor scatter_28_cast_fp16 = scatter(axis = scatter_28_axis_0, data = reshape_143_cast_fp16, indices = reshape_141, mode = scatter_28_mode_0, updates = reshape_142_cast_fp16, validate_indices = scatter_28_validate_indices_0)[name = tensor("scatter_28_cast_fp16")]; + tensor reshape_144_cast_fp16 = reshape(shape = shape_76, x = scatter_28_cast_fp16)[name = tensor("reshape_144_cast_fp16")]; + tensor var_6350_begin_0 = const()[name = tensor("op_6350_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor var_6350_end_0 = const()[name = tensor("op_6350_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor var_6350_end_mask_0 = const()[name = tensor("op_6350_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_6350_squeeze_mask_0 = const()[name = tensor("op_6350_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor var_6350_cast_fp16 = slice_by_index(begin = var_6350_begin_0, end = var_6350_end_0, end_mask = var_6350_end_mask_0, squeeze_mask = var_6350_squeeze_mask_0, x = reshape_144_cast_fp16)[name = tensor("op_6350_cast_fp16")]; + tensor var_6352_axis_0 = const()[name = tensor("op_6352_axis_0"), val = tensor(1)]; + tensor var_6352_mode_0 = const()[name = tensor("op_6352_mode_0"), val = tensor("update")]; + tensor var_6352_validate_indices_0 = const()[name = tensor("op_6352_validate_indices_0"), val = tensor(false)]; + tensor var_6352_cast_fp16 = scatter_along_axis(axis = var_6352_axis_0, data = var_6350_cast_fp16, indices = write_indices_29, mode = var_6352_mode_0, updates = v_29_cast_fp16, validate_indices = var_6352_validate_indices_0)[name = tensor("op_6352_cast_fp16")]; + tensor concat_102 = const()[name = tensor("concat_102"), val = tensor([1, 0, 0, 0, 0])]; + tensor concat_103 = const()[name = tensor("concat_103"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_29_internal_tensor_assign_2_stride_0 = const()[name = tensor("new_cache_29_internal_tensor_assign_2_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_29_internal_tensor_assign_2_begin_mask_0 = const()[name = tensor("new_cache_29_internal_tensor_assign_2_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_29_internal_tensor_assign_2_end_mask_0 = const()[name = tensor("new_cache_29_internal_tensor_assign_2_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_29_internal_tensor_assign_2_squeeze_mask_0 = const()[name = tensor("new_cache_29_internal_tensor_assign_2_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_77 = const()[name = tensor("shape_77"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_29 = const()[name = tensor("reduce_prod_29"), val = tensor(1048576)]; + tensor range_1d_29_start_0 = const()[name = tensor("range_1d_29_start_0"), val = tensor(0)]; + tensor range_1d_29_step_0 = const()[name = tensor("range_1d_29_step_0"), val = tensor(1)]; + tensor range_1d_29 = range_1d(end = reduce_prod_29, start = range_1d_29_start_0, step = range_1d_29_step_0)[name = tensor("range_1d_29")]; + tensor reshape_145 = reshape(shape = shape_77, x = range_1d_29)[name = tensor("reshape_145")]; + tensor slice_by_index_29 = slice_by_index(begin = concat_102, begin_mask = new_cache_29_internal_tensor_assign_2_begin_mask_0, end = concat_103, end_mask = new_cache_29_internal_tensor_assign_2_end_mask_0, squeeze_mask = new_cache_29_internal_tensor_assign_2_squeeze_mask_0, stride = new_cache_29_internal_tensor_assign_2_stride_0, x = reshape_145)[name = tensor("slice_by_index_29")]; + tensor reshape_146_shape_0 = const()[name = tensor("reshape_146_shape_0"), val = tensor([-1])]; + tensor reshape_146 = reshape(shape = reshape_146_shape_0, x = slice_by_index_29)[name = tensor("reshape_146")]; + tensor reshape_147_shape_0 = const()[name = tensor("reshape_147_shape_0"), val = tensor([-1])]; + tensor reshape_147_cast_fp16 = reshape(shape = reshape_147_shape_0, x = var_6352_cast_fp16)[name = tensor("reshape_147_cast_fp16")]; + tensor reshape_148_shape_0 = const()[name = tensor("reshape_148_shape_0"), val = tensor([-1])]; + tensor reshape_148_cast_fp16 = reshape(shape = reshape_148_shape_0, x = reshape_144_cast_fp16)[name = tensor("reshape_148_cast_fp16")]; + tensor scatter_29_mode_0 = const()[name = tensor("scatter_29_mode_0"), val = tensor("update")]; + tensor scatter_29_axis_0 = const()[name = tensor("scatter_29_axis_0"), val = tensor(0)]; + tensor scatter_29_validate_indices_0 = const()[name = tensor("scatter_29_validate_indices_0"), val = tensor(false)]; + tensor scatter_29_cast_fp16 = scatter(axis = scatter_29_axis_0, data = reshape_148_cast_fp16, indices = reshape_146, mode = scatter_29_mode_0, updates = reshape_147_cast_fp16, validate_indices = scatter_29_validate_indices_0)[name = tensor("scatter_29_cast_fp16")]; + tensor reshape_149_cast_fp16 = reshape(shape = shape_77, x = scatter_29_cast_fp16)[name = tensor("reshape_149_cast_fp16")]; + tensor reshape_149_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("reshape_149_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor keys_85_begin_0 = const()[name = tensor("keys_85_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor keys_85_end_0 = const()[name = tensor("keys_85_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor keys_85_end_mask_0 = const()[name = tensor("keys_85_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor keys_85_squeeze_mask_0 = const()[name = tensor("keys_85_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor keys_85_cast_fp16 = slice_by_index(begin = keys_85_begin_0, end = keys_85_end_0, end_mask = keys_85_end_mask_0, squeeze_mask = keys_85_squeeze_mask_0, x = reshape_149_cast_fp16)[name = tensor("keys_85_cast_fp16")]; + tensor values_85_begin_0 = const()[name = tensor("values_85_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor values_85_end_0 = const()[name = tensor("values_85_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor values_85_end_mask_0 = const()[name = tensor("values_85_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor values_85_squeeze_mask_0 = const()[name = tensor("values_85_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor values_85_cast_fp16 = slice_by_index(begin = values_85_begin_0, end = values_85_end_0, end_mask = values_85_end_mask_0, squeeze_mask = values_85_squeeze_mask_0, x = reshape_149_cast_fp16)[name = tensor("values_85_cast_fp16")]; + tensor var_6364_cast_fp16 = not_equal(x = keys_85_cast_fp16, y = keys_85_cast_fp16)[name = tensor("op_6364_cast_fp16")]; + tensor keys_87_cast_fp16 = select(a = var_504_to_fp16, b = keys_85_cast_fp16, cond = var_6364_cast_fp16)[name = tensor("keys_87_cast_fp16")]; + tensor var_6372_cast_fp16 = not_equal(x = values_85_cast_fp16, y = values_85_cast_fp16)[name = tensor("op_6372_cast_fp16")]; + tensor values_87_cast_fp16 = select(a = var_504_to_fp16, b = values_85_cast_fp16, cond = var_6372_cast_fp16)[name = tensor("values_87_cast_fp16")]; + tensor var_6396 = const()[name = tensor("op_6396"), val = tensor([0, 2, 1, 3])]; + tensor var_6409 = const()[name = tensor("op_6409"), val = tensor([1, 1, 1])]; + tensor var_6410_cast_fp16 = reshape(shape = var_6409, x = position14_to_fp16)[name = tensor("op_6410_cast_fp16")]; + tensor var_6427_to_fp16 = const()[name = tensor("op_6427_to_fp16"), val = tensor(0x1p+0)]; + tensor valid_len_29_cast_fp16 = add(x = var_6410_cast_fp16, y = var_6427_to_fp16)[name = tensor("valid_len_29_cast_fp16")]; + tensor valid_mask_29_cast_fp16 = less(x = k_positions_1_promoted_to_fp16, y = valid_len_29_cast_fp16)[name = tensor("valid_mask_29_cast_fp16")]; + tensor causal_mask_29_cast_fp16 = less_equal(x = k_positions_1_promoted_to_fp16, y = var_6410_cast_fp16)[name = tensor("causal_mask_29_cast_fp16")]; + tensor attn_mask_57 = logical_and(x = valid_mask_29_cast_fp16, y = causal_mask_29_cast_fp16)[name = tensor("attn_mask_57")]; + tensor attn_mask_59_axes_0 = const()[name = tensor("attn_mask_59_axes_0"), val = tensor([1])]; + tensor attn_mask_59 = expand_dims(axes = attn_mask_59_axes_0, x = attn_mask_57)[name = tensor("attn_mask_59")]; + tensor var_6445_transpose_x_0 = const()[name = tensor("op_6445_transpose_x_0"), val = tensor(false)]; + tensor var_6445_transpose_y_0 = const()[name = tensor("op_6445_transpose_y_0"), val = tensor(false)]; + tensor transpose_100_perm_0 = const()[name = tensor("transpose_100_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_101_perm_0 = const()[name = tensor("transpose_101_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_101 = transpose(perm = transpose_101_perm_0, x = keys_87_cast_fp16)[name = tensor("transpose_157")]; + tensor transpose_100 = transpose(perm = transpose_100_perm_0, x = q_87_cast_fp16)[name = tensor("transpose_158")]; + tensor var_6445_cast_fp16 = matmul(transpose_x = var_6445_transpose_x_0, transpose_y = var_6445_transpose_y_0, x = transpose_100, y = transpose_101)[name = tensor("op_6445_cast_fp16")]; + tensor var_6439_to_fp16 = const()[name = tensor("op_6439_to_fp16"), val = tensor([0x1p-3])]; + tensor attn_weights_85_cast_fp16 = mul(x = var_6445_cast_fp16, y = var_6439_to_fp16)[name = tensor("attn_weights_85_cast_fp16")]; + tensor var_6447 = logical_not(x = attn_mask_59)[name = tensor("op_6447")]; + tensor var_6448_to_fp16 = const()[name = tensor("op_6448_to_fp16"), val = tensor(-inf)]; + tensor attn_weights_87_cast_fp16 = select(a = var_6448_to_fp16, b = attn_weights_85_cast_fp16, cond = var_6447)[name = tensor("attn_weights_87_cast_fp16")]; + tensor var_6450 = const()[name = tensor("op_6450"), val = tensor(-1)]; + tensor attn_weights_89_cast_fp16 = softmax(axis = var_6450, x = attn_weights_87_cast_fp16)[name = tensor("attn_weights_89_cast_fp16")]; + tensor attn_output_29_transpose_x_0 = const()[name = tensor("attn_output_29_transpose_x_0"), val = tensor(false)]; + tensor attn_output_29_transpose_y_0 = const()[name = tensor("attn_output_29_transpose_y_0"), val = tensor(false)]; + tensor values_89_cast_fp16 = transpose(perm = var_6396, x = values_87_cast_fp16)[name = tensor("transpose_159")]; + tensor attn_output_29_cast_fp16 = matmul(transpose_x = attn_output_29_transpose_x_0, transpose_y = attn_output_29_transpose_y_0, x = attn_weights_89_cast_fp16, y = values_89_cast_fp16)[name = tensor("attn_output_29_cast_fp16")]; + tensor var_6458 = const()[name = tensor("op_6458"), val = tensor([0, 2, 1, 3])]; + tensor var_6461 = const()[name = tensor("op_6461"), val = tensor([1, 1, 1024])]; + tensor var_6459_cast_fp16 = transpose(perm = var_6458, x = attn_output_29_cast_fp16)[name = tensor("transpose_156")]; + tensor input_145_cast_fp16 = reshape(shape = var_6461, x = var_6459_cast_fp16)[name = tensor("input_145_cast_fp16")]; + tensor linear_58_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = attn14_out_proj_weight_cast_fp16, x = input_145_cast_fp16)[name = tensor("linear_58_cast_fp16")]; + tensor var_6467_to_fp16 = const()[name = tensor("op_6467_to_fp16"), val = tensor(0x1p+0)]; + tensor var_6468_cast_fp16 = add(x = position14_to_fp16, y = var_6467_to_fp16)[name = tensor("op_6468_cast_fp16")]; + tensor var_6468_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_6468_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor input_147_cast_fp16 = add(x = input_143_cast_fp16, y = linear_58_cast_fp16)[name = tensor("input_147_cast_fp16")]; + tensor input_149_axes_0 = const()[name = tensor("input_149_axes_0"), val = tensor([-1])]; + tensor norm14_2_weight_to_fp16 = const()[name = tensor("norm14_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(303709312)))]; + tensor norm14_2_bias_to_fp16 = const()[name = tensor("norm14_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(303711424)))]; + tensor var_6472_to_fp16 = const()[name = tensor("op_6472_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_149_cast_fp16 = layer_norm(axes = input_149_axes_0, beta = norm14_2_bias_to_fp16, epsilon = var_6472_to_fp16, gamma = norm14_2_weight_to_fp16, x = input_147_cast_fp16)[name = tensor("input_149_cast_fp16")]; + tensor linear_59_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = linear14_1_weight_cast_fp16, x = input_149_cast_fp16)[name = tensor("linear_59_cast_fp16")]; + tensor input_151_mode_0 = const()[name = tensor("input_151_mode_0"), val = tensor("EXACT")]; + tensor input_151_cast_fp16 = gelu(mode = input_151_mode_0, x = linear_59_cast_fp16)[name = tensor("input_151_cast_fp16")]; + tensor linear_60_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = linear14_2_weight_cast_fp16, x = input_151_cast_fp16)[name = tensor("linear_60_cast_fp16")]; + tensor input_153_cast_fp16 = add(x = input_147_cast_fp16, y = linear_60_cast_fp16)[name = tensor("input_153_cast_fp16")]; + tensor x_31_axes_0 = const()[name = tensor("x_31_axes_0"), val = tensor([-1])]; + tensor norm15_1_weight_to_fp16 = const()[name = tensor("norm15_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(303713536)))]; + tensor norm15_1_bias_to_fp16 = const()[name = tensor("norm15_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(303715648)))]; + tensor var_6489_to_fp16 = const()[name = tensor("op_6489_to_fp16"), val = tensor(0x1.5p-17)]; + tensor x_31_cast_fp16 = layer_norm(axes = x_31_axes_0, beta = norm15_1_bias_to_fp16, epsilon = var_6489_to_fp16, gamma = norm15_1_weight_to_fp16, x = input_153_cast_fp16)[name = tensor("x_31_cast_fp16")]; + tensor linear_61_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = attn15_in_proj_weight_cast_fp16, x = x_31_cast_fp16)[name = tensor("linear_61_cast_fp16")]; + tensor var_6525 = const()[name = tensor("op_6525"), val = tensor([1, 1, 3, 16, 64])]; + tensor qkv_31_cast_fp16 = reshape(shape = var_6525, x = linear_61_cast_fp16)[name = tensor("qkv_31_cast_fp16")]; + tensor q_91_begin_0 = const()[name = tensor("q_91_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor q_91_end_0 = const()[name = tensor("q_91_end_0"), val = tensor([1, 1, 1, 16, 64])]; + tensor q_91_end_mask_0 = const()[name = tensor("q_91_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor q_91_squeeze_mask_0 = const()[name = tensor("q_91_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor q_91_cast_fp16 = slice_by_index(begin = q_91_begin_0, end = q_91_end_0, end_mask = q_91_end_mask_0, squeeze_mask = q_91_squeeze_mask_0, x = qkv_31_cast_fp16)[name = tensor("q_91_cast_fp16")]; + tensor k_61_begin_0 = const()[name = tensor("k_61_begin_0"), val = tensor([0, 0, 1, 0, 0])]; + tensor k_61_end_0 = const()[name = tensor("k_61_end_0"), val = tensor([1, 1, 2, 16, 64])]; + tensor k_61_end_mask_0 = const()[name = tensor("k_61_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor k_61_squeeze_mask_0 = const()[name = tensor("k_61_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor k_61_cast_fp16 = slice_by_index(begin = k_61_begin_0, end = k_61_end_0, end_mask = k_61_end_mask_0, squeeze_mask = k_61_squeeze_mask_0, x = qkv_31_cast_fp16)[name = tensor("k_61_cast_fp16")]; + tensor v_31_begin_0 = const()[name = tensor("v_31_begin_0"), val = tensor([0, 0, 2, 0, 0])]; + tensor v_31_end_0 = const()[name = tensor("v_31_end_0"), val = tensor([1, 1, 3, 16, 64])]; + tensor v_31_end_mask_0 = const()[name = tensor("v_31_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor v_31_squeeze_mask_0 = const()[name = tensor("v_31_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor v_31_cast_fp16 = slice_by_index(begin = v_31_begin_0, end = v_31_end_0, end_mask = v_31_end_mask_0, squeeze_mask = v_31_squeeze_mask_0, x = qkv_31_cast_fp16)[name = tensor("v_31_cast_fp16")]; + tensor var_6629 = const()[name = tensor("op_6629"), val = tensor([1, 1, 1, 1])]; + tensor position15_to_fp16_dtype_0 = const()[name = tensor("position15_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor position15_to_fp16 = cast(dtype = position15_to_fp16_dtype_0, x = position15)[name = tensor("cast_478")]; + tensor ts_95_cast_fp16 = reshape(shape = var_6629, x = position15_to_fp16)[name = tensor("ts_95_cast_fp16")]; + tensor var_6633 = const()[name = tensor("op_6633"), val = tensor([1, 1, 16, 32, 2])]; + tensor q_complex_31_cast_fp16 = reshape(shape = var_6633, x = q_91_cast_fp16)[name = tensor("q_complex_31_cast_fp16")]; + tensor var_6637 = const()[name = tensor("op_6637"), val = tensor([1, 1, 16, 32, 2])]; + tensor k_complex_31_cast_fp16 = reshape(shape = var_6637, x = k_61_cast_fp16)[name = tensor("k_complex_31_cast_fp16")]; + tensor var_6641_begin_0 = const()[name = tensor("op_6641_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_6641_end_0 = const()[name = tensor("op_6641_end_0"), val = tensor([1, 1, 16, 32, 1])]; + tensor var_6641_end_mask_0 = const()[name = tensor("op_6641_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_6641_squeeze_mask_0 = const()[name = tensor("op_6641_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_6641_cast_fp16 = slice_by_index(begin = var_6641_begin_0, end = var_6641_end_0, end_mask = var_6641_end_mask_0, squeeze_mask = var_6641_squeeze_mask_0, x = q_complex_31_cast_fp16)[name = tensor("op_6641_cast_fp16")]; + tensor var_6649_begin_0 = const()[name = tensor("op_6649_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_6649_end_0 = const()[name = tensor("op_6649_end_0"), val = tensor([1, 1, 16, 32, 2])]; + tensor var_6649_end_mask_0 = const()[name = tensor("op_6649_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_6649_squeeze_mask_0 = const()[name = tensor("op_6649_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_6649_cast_fp16 = slice_by_index(begin = var_6649_begin_0, end = var_6649_end_0, end_mask = var_6649_end_mask_0, squeeze_mask = var_6649_squeeze_mask_0, x = q_complex_31_cast_fp16)[name = tensor("op_6649_cast_fp16")]; + tensor var_6657_begin_0 = const()[name = tensor("op_6657_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_6657_end_0 = const()[name = tensor("op_6657_end_0"), val = tensor([1, 1, 16, 32, 1])]; + tensor var_6657_end_mask_0 = const()[name = tensor("op_6657_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_6657_squeeze_mask_0 = const()[name = tensor("op_6657_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_6657_cast_fp16 = slice_by_index(begin = var_6657_begin_0, end = var_6657_end_0, end_mask = var_6657_end_mask_0, squeeze_mask = var_6657_squeeze_mask_0, x = k_complex_31_cast_fp16)[name = tensor("op_6657_cast_fp16")]; + tensor var_6665_begin_0 = const()[name = tensor("op_6665_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_6665_end_0 = const()[name = tensor("op_6665_end_0"), val = tensor([1, 1, 16, 32, 2])]; + tensor var_6665_end_mask_0 = const()[name = tensor("op_6665_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_6665_squeeze_mask_0 = const()[name = tensor("op_6665_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_6665_cast_fp16 = slice_by_index(begin = var_6665_begin_0, end = var_6665_end_0, end_mask = var_6665_end_mask_0, squeeze_mask = var_6665_squeeze_mask_0, x = k_complex_31_cast_fp16)[name = tensor("op_6665_cast_fp16")]; + tensor freqs_31_to_fp16 = const()[name = tensor("freqs_31_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(303717760)))]; + tensor var_6671_cast_fp16 = mul(x = freqs_31_to_fp16, y = ts_95_cast_fp16)[name = tensor("op_6671_cast_fp16")]; + tensor rotr_31_cast_fp16 = cos(x = var_6671_cast_fp16)[name = tensor("rotr_31_cast_fp16")]; + tensor roti_31_cast_fp16 = sin(x = var_6671_cast_fp16)[name = tensor("roti_31_cast_fp16")]; + tensor var_6675_cast_fp16 = mul(x = var_6641_cast_fp16, y = rotr_31_cast_fp16)[name = tensor("op_6675_cast_fp16")]; + tensor var_6676_cast_fp16 = mul(x = var_6649_cast_fp16, y = roti_31_cast_fp16)[name = tensor("op_6676_cast_fp16")]; + tensor qor_61_cast_fp16 = sub(x = var_6675_cast_fp16, y = var_6676_cast_fp16)[name = tensor("qor_61_cast_fp16")]; + tensor var_6679_cast_fp16 = mul(x = var_6641_cast_fp16, y = roti_31_cast_fp16)[name = tensor("op_6679_cast_fp16")]; + tensor var_6680_cast_fp16 = mul(x = var_6649_cast_fp16, y = rotr_31_cast_fp16)[name = tensor("op_6680_cast_fp16")]; + tensor qoi_61_cast_fp16 = add(x = var_6679_cast_fp16, y = var_6680_cast_fp16)[name = tensor("qoi_61_cast_fp16")]; + tensor var_6683_cast_fp16 = mul(x = var_6657_cast_fp16, y = rotr_31_cast_fp16)[name = tensor("op_6683_cast_fp16")]; + tensor var_6684_cast_fp16 = mul(x = var_6665_cast_fp16, y = roti_31_cast_fp16)[name = tensor("op_6684_cast_fp16")]; + tensor kor_61_cast_fp16 = sub(x = var_6683_cast_fp16, y = var_6684_cast_fp16)[name = tensor("kor_61_cast_fp16")]; + tensor var_6687_cast_fp16 = mul(x = var_6657_cast_fp16, y = roti_31_cast_fp16)[name = tensor("op_6687_cast_fp16")]; + tensor var_6688_cast_fp16 = mul(x = var_6665_cast_fp16, y = rotr_31_cast_fp16)[name = tensor("op_6688_cast_fp16")]; + tensor koi_61_cast_fp16 = add(x = var_6687_cast_fp16, y = var_6688_cast_fp16)[name = tensor("koi_61_cast_fp16")]; + tensor qo_31_axis_0 = const()[name = tensor("qo_31_axis_0"), val = tensor(-1)]; + tensor qo_31_cast_fp16 = stack(axis = qo_31_axis_0, values = (qor_61_cast_fp16, qoi_61_cast_fp16))[name = tensor("qo_31_cast_fp16")]; + tensor ko_31_axis_0 = const()[name = tensor("ko_31_axis_0"), val = tensor(-1)]; + tensor ko_31_cast_fp16 = stack(axis = ko_31_axis_0, values = (kor_61_cast_fp16, koi_61_cast_fp16))[name = tensor("ko_31_cast_fp16")]; + tensor var_6717 = const()[name = tensor("op_6717"), val = tensor([1, 1, 16, 64])]; + tensor q_93_cast_fp16 = reshape(shape = var_6717, x = qo_31_cast_fp16)[name = tensor("q_93_cast_fp16")]; + tensor var_6719 = const()[name = tensor("op_6719"), val = tensor([1, 1, 16, 64])]; + tensor k_63_cast_fp16 = reshape(shape = var_6719, x = ko_31_cast_fp16)[name = tensor("k_63_cast_fp16")]; + tensor _inversed_6741_y_0_to_fp16 = const()[name = tensor("_inversed_6741_y_0_to_fp16"), val = tensor(0x1p-9)]; + tensor _inversed_6741_cast_fp16 = mul(x = ts_95_cast_fp16, y = _inversed_6741_y_0_to_fp16)[name = tensor("_inversed_6741_cast_fp16")]; + tensor var_6742_cast_fp16 = floor(x = _inversed_6741_cast_fp16)[name = tensor("op_6742_cast_fp16")]; + tensor var_6743_to_fp16 = const()[name = tensor("op_6743_to_fp16"), val = tensor(0x1p+9)]; + tensor var_6744_cast_fp16 = mul(x = var_6742_cast_fp16, y = var_6743_to_fp16)[name = tensor("op_6744_cast_fp16")]; + tensor write_indices_float_63_cast_fp16 = sub(x = ts_95_cast_fp16, y = var_6744_cast_fp16)[name = tensor("write_indices_float_63_cast_fp16")]; + tensor var_6751_dtype_0 = const()[name = tensor("op_6751_dtype_0"), val = tensor("int32")]; + tensor write_indices_31_reps_0 = const()[name = tensor("write_indices_31_reps_0"), val = tensor([1, 1, 16, 64])]; + tensor write_indices_float_63_cast_fp16_to_int32 = cast(dtype = var_6751_dtype_0, x = write_indices_float_63_cast_fp16)[name = tensor("cast_477")]; + tensor write_indices_31 = tile(reps = write_indices_31_reps_0, x = write_indices_float_63_cast_fp16_to_int32)[name = tensor("write_indices_31")]; + tensor var_6759_begin_0 = const()[name = tensor("op_6759_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_6759_end_0 = const()[name = tensor("op_6759_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor var_6759_end_mask_0 = const()[name = tensor("op_6759_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_6759_squeeze_mask_0 = const()[name = tensor("op_6759_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor cache15_to_fp16_dtype_0 = const()[name = tensor("cache15_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor cache15_to_fp16 = cast(dtype = cache15_to_fp16_dtype_0, x = cache15)[name = tensor("cast_476")]; + tensor var_6759_cast_fp16 = slice_by_index(begin = var_6759_begin_0, end = var_6759_end_0, end_mask = var_6759_end_mask_0, squeeze_mask = var_6759_squeeze_mask_0, x = cache15_to_fp16)[name = tensor("op_6759_cast_fp16")]; + tensor var_6761_axis_0 = const()[name = tensor("op_6761_axis_0"), val = tensor(1)]; + tensor var_6761_mode_0 = const()[name = tensor("op_6761_mode_0"), val = tensor("update")]; + tensor var_6761_validate_indices_0 = const()[name = tensor("op_6761_validate_indices_0"), val = tensor(false)]; + tensor var_6761_cast_fp16 = scatter_along_axis(axis = var_6761_axis_0, data = var_6759_cast_fp16, indices = write_indices_31, mode = var_6761_mode_0, updates = k_63_cast_fp16, validate_indices = var_6761_validate_indices_0)[name = tensor("op_6761_cast_fp16")]; + tensor concat_107 = const()[name = tensor("concat_107"), val = tensor([0, 0, 0, 0, 0])]; + tensor concat_108 = const()[name = tensor("concat_108"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_31_internal_tensor_assign_1_stride_0 = const()[name = tensor("new_cache_31_internal_tensor_assign_1_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_31_internal_tensor_assign_1_begin_mask_0 = const()[name = tensor("new_cache_31_internal_tensor_assign_1_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_31_internal_tensor_assign_1_end_mask_0 = const()[name = tensor("new_cache_31_internal_tensor_assign_1_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_31_internal_tensor_assign_1_squeeze_mask_0 = const()[name = tensor("new_cache_31_internal_tensor_assign_1_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_78 = const()[name = tensor("shape_78"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_30 = const()[name = tensor("reduce_prod_30"), val = tensor(1048576)]; + tensor range_1d_30_start_0 = const()[name = tensor("range_1d_30_start_0"), val = tensor(0)]; + tensor range_1d_30_step_0 = const()[name = tensor("range_1d_30_step_0"), val = tensor(1)]; + tensor range_1d_30 = range_1d(end = reduce_prod_30, start = range_1d_30_start_0, step = range_1d_30_step_0)[name = tensor("range_1d_30")]; + tensor reshape_150 = reshape(shape = shape_78, x = range_1d_30)[name = tensor("reshape_150")]; + tensor slice_by_index_30 = slice_by_index(begin = concat_107, begin_mask = new_cache_31_internal_tensor_assign_1_begin_mask_0, end = concat_108, end_mask = new_cache_31_internal_tensor_assign_1_end_mask_0, squeeze_mask = new_cache_31_internal_tensor_assign_1_squeeze_mask_0, stride = new_cache_31_internal_tensor_assign_1_stride_0, x = reshape_150)[name = tensor("slice_by_index_30")]; + tensor reshape_151_shape_0 = const()[name = tensor("reshape_151_shape_0"), val = tensor([-1])]; + tensor reshape_151 = reshape(shape = reshape_151_shape_0, x = slice_by_index_30)[name = tensor("reshape_151")]; + tensor reshape_152_shape_0 = const()[name = tensor("reshape_152_shape_0"), val = tensor([-1])]; + tensor reshape_152_cast_fp16 = reshape(shape = reshape_152_shape_0, x = var_6761_cast_fp16)[name = tensor("reshape_152_cast_fp16")]; + tensor reshape_153_shape_0 = const()[name = tensor("reshape_153_shape_0"), val = tensor([-1])]; + tensor reshape_153_cast_fp16 = reshape(shape = reshape_153_shape_0, x = cache15_to_fp16)[name = tensor("reshape_153_cast_fp16")]; + tensor scatter_30_mode_0 = const()[name = tensor("scatter_30_mode_0"), val = tensor("update")]; + tensor scatter_30_axis_0 = const()[name = tensor("scatter_30_axis_0"), val = tensor(0)]; + tensor scatter_30_validate_indices_0 = const()[name = tensor("scatter_30_validate_indices_0"), val = tensor(false)]; + tensor scatter_30_cast_fp16 = scatter(axis = scatter_30_axis_0, data = reshape_153_cast_fp16, indices = reshape_151, mode = scatter_30_mode_0, updates = reshape_152_cast_fp16, validate_indices = scatter_30_validate_indices_0)[name = tensor("scatter_30_cast_fp16")]; + tensor reshape_154_cast_fp16 = reshape(shape = shape_78, x = scatter_30_cast_fp16)[name = tensor("reshape_154_cast_fp16")]; + tensor var_6769_begin_0 = const()[name = tensor("op_6769_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor var_6769_end_0 = const()[name = tensor("op_6769_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor var_6769_end_mask_0 = const()[name = tensor("op_6769_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_6769_squeeze_mask_0 = const()[name = tensor("op_6769_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor var_6769_cast_fp16 = slice_by_index(begin = var_6769_begin_0, end = var_6769_end_0, end_mask = var_6769_end_mask_0, squeeze_mask = var_6769_squeeze_mask_0, x = reshape_154_cast_fp16)[name = tensor("op_6769_cast_fp16")]; + tensor var_6771_axis_0 = const()[name = tensor("op_6771_axis_0"), val = tensor(1)]; + tensor var_6771_mode_0 = const()[name = tensor("op_6771_mode_0"), val = tensor("update")]; + tensor var_6771_validate_indices_0 = const()[name = tensor("op_6771_validate_indices_0"), val = tensor(false)]; + tensor var_6771_cast_fp16 = scatter_along_axis(axis = var_6771_axis_0, data = var_6769_cast_fp16, indices = write_indices_31, mode = var_6771_mode_0, updates = v_31_cast_fp16, validate_indices = var_6771_validate_indices_0)[name = tensor("op_6771_cast_fp16")]; + tensor concat_109 = const()[name = tensor("concat_109"), val = tensor([1, 0, 0, 0, 0])]; + tensor concat_110 = const()[name = tensor("concat_110"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_31_internal_tensor_assign_2_stride_0 = const()[name = tensor("new_cache_31_internal_tensor_assign_2_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_31_internal_tensor_assign_2_begin_mask_0 = const()[name = tensor("new_cache_31_internal_tensor_assign_2_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_31_internal_tensor_assign_2_end_mask_0 = const()[name = tensor("new_cache_31_internal_tensor_assign_2_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_31_internal_tensor_assign_2_squeeze_mask_0 = const()[name = tensor("new_cache_31_internal_tensor_assign_2_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_79 = const()[name = tensor("shape_79"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_31 = const()[name = tensor("reduce_prod_31"), val = tensor(1048576)]; + tensor range_1d_31_start_0 = const()[name = tensor("range_1d_31_start_0"), val = tensor(0)]; + tensor range_1d_31_step_0 = const()[name = tensor("range_1d_31_step_0"), val = tensor(1)]; + tensor range_1d_31 = range_1d(end = reduce_prod_31, start = range_1d_31_start_0, step = range_1d_31_step_0)[name = tensor("range_1d_31")]; + tensor reshape_155 = reshape(shape = shape_79, x = range_1d_31)[name = tensor("reshape_155")]; + tensor slice_by_index_31 = slice_by_index(begin = concat_109, begin_mask = new_cache_31_internal_tensor_assign_2_begin_mask_0, end = concat_110, end_mask = new_cache_31_internal_tensor_assign_2_end_mask_0, squeeze_mask = new_cache_31_internal_tensor_assign_2_squeeze_mask_0, stride = new_cache_31_internal_tensor_assign_2_stride_0, x = reshape_155)[name = tensor("slice_by_index_31")]; + tensor reshape_156_shape_0 = const()[name = tensor("reshape_156_shape_0"), val = tensor([-1])]; + tensor reshape_156 = reshape(shape = reshape_156_shape_0, x = slice_by_index_31)[name = tensor("reshape_156")]; + tensor reshape_157_shape_0 = const()[name = tensor("reshape_157_shape_0"), val = tensor([-1])]; + tensor reshape_157_cast_fp16 = reshape(shape = reshape_157_shape_0, x = var_6771_cast_fp16)[name = tensor("reshape_157_cast_fp16")]; + tensor reshape_158_shape_0 = const()[name = tensor("reshape_158_shape_0"), val = tensor([-1])]; + tensor reshape_158_cast_fp16 = reshape(shape = reshape_158_shape_0, x = reshape_154_cast_fp16)[name = tensor("reshape_158_cast_fp16")]; + tensor scatter_31_mode_0 = const()[name = tensor("scatter_31_mode_0"), val = tensor("update")]; + tensor scatter_31_axis_0 = const()[name = tensor("scatter_31_axis_0"), val = tensor(0)]; + tensor scatter_31_validate_indices_0 = const()[name = tensor("scatter_31_validate_indices_0"), val = tensor(false)]; + tensor scatter_31_cast_fp16 = scatter(axis = scatter_31_axis_0, data = reshape_158_cast_fp16, indices = reshape_156, mode = scatter_31_mode_0, updates = reshape_157_cast_fp16, validate_indices = scatter_31_validate_indices_0)[name = tensor("scatter_31_cast_fp16")]; + tensor reshape_159_cast_fp16 = reshape(shape = shape_79, x = scatter_31_cast_fp16)[name = tensor("reshape_159_cast_fp16")]; + tensor reshape_159_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("reshape_159_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor keys_91_begin_0 = const()[name = tensor("keys_91_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor keys_91_end_0 = const()[name = tensor("keys_91_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor keys_91_end_mask_0 = const()[name = tensor("keys_91_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor keys_91_squeeze_mask_0 = const()[name = tensor("keys_91_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor keys_91_cast_fp16 = slice_by_index(begin = keys_91_begin_0, end = keys_91_end_0, end_mask = keys_91_end_mask_0, squeeze_mask = keys_91_squeeze_mask_0, x = reshape_159_cast_fp16)[name = tensor("keys_91_cast_fp16")]; + tensor values_91_begin_0 = const()[name = tensor("values_91_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor values_91_end_0 = const()[name = tensor("values_91_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor values_91_end_mask_0 = const()[name = tensor("values_91_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor values_91_squeeze_mask_0 = const()[name = tensor("values_91_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor values_91_cast_fp16 = slice_by_index(begin = values_91_begin_0, end = values_91_end_0, end_mask = values_91_end_mask_0, squeeze_mask = values_91_squeeze_mask_0, x = reshape_159_cast_fp16)[name = tensor("values_91_cast_fp16")]; + tensor var_6783_cast_fp16 = not_equal(x = keys_91_cast_fp16, y = keys_91_cast_fp16)[name = tensor("op_6783_cast_fp16")]; + tensor keys_93_cast_fp16 = select(a = var_504_to_fp16, b = keys_91_cast_fp16, cond = var_6783_cast_fp16)[name = tensor("keys_93_cast_fp16")]; + tensor var_6791_cast_fp16 = not_equal(x = values_91_cast_fp16, y = values_91_cast_fp16)[name = tensor("op_6791_cast_fp16")]; + tensor values_93_cast_fp16 = select(a = var_504_to_fp16, b = values_91_cast_fp16, cond = var_6791_cast_fp16)[name = tensor("values_93_cast_fp16")]; + tensor var_6815 = const()[name = tensor("op_6815"), val = tensor([0, 2, 1, 3])]; + tensor var_6828 = const()[name = tensor("op_6828"), val = tensor([1, 1, 1])]; + tensor var_6829_cast_fp16 = reshape(shape = var_6828, x = position15_to_fp16)[name = tensor("op_6829_cast_fp16")]; + tensor var_6846_to_fp16 = const()[name = tensor("op_6846_to_fp16"), val = tensor(0x1p+0)]; + tensor valid_len_31_cast_fp16 = add(x = var_6829_cast_fp16, y = var_6846_to_fp16)[name = tensor("valid_len_31_cast_fp16")]; + tensor valid_mask_31_cast_fp16 = less(x = k_positions_1_promoted_to_fp16, y = valid_len_31_cast_fp16)[name = tensor("valid_mask_31_cast_fp16")]; + tensor causal_mask_31_cast_fp16 = less_equal(x = k_positions_1_promoted_to_fp16, y = var_6829_cast_fp16)[name = tensor("causal_mask_31_cast_fp16")]; + tensor attn_mask_61 = logical_and(x = valid_mask_31_cast_fp16, y = causal_mask_31_cast_fp16)[name = tensor("attn_mask_61")]; + tensor attn_mask_63_axes_0 = const()[name = tensor("attn_mask_63_axes_0"), val = tensor([1])]; + tensor attn_mask_63 = expand_dims(axes = attn_mask_63_axes_0, x = attn_mask_61)[name = tensor("attn_mask_63")]; + tensor var_6864_transpose_x_0 = const()[name = tensor("op_6864_transpose_x_0"), val = tensor(false)]; + tensor var_6864_transpose_y_0 = const()[name = tensor("op_6864_transpose_y_0"), val = tensor(false)]; + tensor transpose_102_perm_0 = const()[name = tensor("transpose_102_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_103_perm_0 = const()[name = tensor("transpose_103_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_103 = transpose(perm = transpose_103_perm_0, x = keys_93_cast_fp16)[name = tensor("transpose_153")]; + tensor transpose_102 = transpose(perm = transpose_102_perm_0, x = q_93_cast_fp16)[name = tensor("transpose_154")]; + tensor var_6864_cast_fp16 = matmul(transpose_x = var_6864_transpose_x_0, transpose_y = var_6864_transpose_y_0, x = transpose_102, y = transpose_103)[name = tensor("op_6864_cast_fp16")]; + tensor var_6858_to_fp16 = const()[name = tensor("op_6858_to_fp16"), val = tensor([0x1p-3])]; + tensor attn_weights_91_cast_fp16 = mul(x = var_6864_cast_fp16, y = var_6858_to_fp16)[name = tensor("attn_weights_91_cast_fp16")]; + tensor var_6866 = logical_not(x = attn_mask_63)[name = tensor("op_6866")]; + tensor var_6867_to_fp16 = const()[name = tensor("op_6867_to_fp16"), val = tensor(-inf)]; + tensor attn_weights_93_cast_fp16 = select(a = var_6867_to_fp16, b = attn_weights_91_cast_fp16, cond = var_6866)[name = tensor("attn_weights_93_cast_fp16")]; + tensor var_6869 = const()[name = tensor("op_6869"), val = tensor(-1)]; + tensor attn_weights_95_cast_fp16 = softmax(axis = var_6869, x = attn_weights_93_cast_fp16)[name = tensor("attn_weights_95_cast_fp16")]; + tensor attn_output_31_transpose_x_0 = const()[name = tensor("attn_output_31_transpose_x_0"), val = tensor(false)]; + tensor attn_output_31_transpose_y_0 = const()[name = tensor("attn_output_31_transpose_y_0"), val = tensor(false)]; + tensor values_95_cast_fp16 = transpose(perm = var_6815, x = values_93_cast_fp16)[name = tensor("transpose_155")]; + tensor attn_output_31_cast_fp16 = matmul(transpose_x = attn_output_31_transpose_x_0, transpose_y = attn_output_31_transpose_y_0, x = attn_weights_95_cast_fp16, y = values_95_cast_fp16)[name = tensor("attn_output_31_cast_fp16")]; + tensor var_6877 = const()[name = tensor("op_6877"), val = tensor([0, 2, 1, 3])]; + tensor var_6880 = const()[name = tensor("op_6880"), val = tensor([1, 1, 1024])]; + tensor var_6878_cast_fp16 = transpose(perm = var_6877, x = attn_output_31_cast_fp16)[name = tensor("transpose_152")]; + tensor input_155_cast_fp16 = reshape(shape = var_6880, x = var_6878_cast_fp16)[name = tensor("input_155_cast_fp16")]; + tensor linear_62_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = attn15_out_proj_weight_cast_fp16, x = input_155_cast_fp16)[name = tensor("linear_62_cast_fp16")]; + tensor var_6886_to_fp16 = const()[name = tensor("op_6886_to_fp16"), val = tensor(0x1p+0)]; + tensor var_6887_cast_fp16 = add(x = position15_to_fp16, y = var_6886_to_fp16)[name = tensor("op_6887_cast_fp16")]; + tensor var_6887_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_6887_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor input_157_cast_fp16 = add(x = input_153_cast_fp16, y = linear_62_cast_fp16)[name = tensor("input_157_cast_fp16")]; + tensor input_159_axes_0 = const()[name = tensor("input_159_axes_0"), val = tensor([-1])]; + tensor norm15_2_weight_to_fp16 = const()[name = tensor("norm15_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(303717888)))]; + tensor norm15_2_bias_to_fp16 = const()[name = tensor("norm15_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(303720000)))]; + tensor var_6891_to_fp16 = const()[name = tensor("op_6891_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_159_cast_fp16 = layer_norm(axes = input_159_axes_0, beta = norm15_2_bias_to_fp16, epsilon = var_6891_to_fp16, gamma = norm15_2_weight_to_fp16, x = input_157_cast_fp16)[name = tensor("input_159_cast_fp16")]; + tensor linear_63_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = linear15_1_weight_cast_fp16, x = input_159_cast_fp16)[name = tensor("linear_63_cast_fp16")]; + tensor input_161_mode_0 = const()[name = tensor("input_161_mode_0"), val = tensor("EXACT")]; + tensor input_161_cast_fp16 = gelu(mode = input_161_mode_0, x = linear_63_cast_fp16)[name = tensor("input_161_cast_fp16")]; + tensor linear_64_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = linear15_2_weight_cast_fp16, x = input_161_cast_fp16)[name = tensor("linear_64_cast_fp16")]; + tensor input_163_cast_fp16 = add(x = input_157_cast_fp16, y = linear_64_cast_fp16)[name = tensor("input_163_cast_fp16")]; + tensor x_33_axes_0 = const()[name = tensor("x_33_axes_0"), val = tensor([-1])]; + tensor norm16_1_weight_to_fp16 = const()[name = tensor("norm16_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(303722112)))]; + tensor norm16_1_bias_to_fp16 = const()[name = tensor("norm16_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(303724224)))]; + tensor var_6908_to_fp16 = const()[name = tensor("op_6908_to_fp16"), val = tensor(0x1.5p-17)]; + tensor x_33_cast_fp16 = layer_norm(axes = x_33_axes_0, beta = norm16_1_bias_to_fp16, epsilon = var_6908_to_fp16, gamma = norm16_1_weight_to_fp16, x = input_163_cast_fp16)[name = tensor("x_33_cast_fp16")]; + tensor linear_65_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = attn16_in_proj_weight_cast_fp16, x = x_33_cast_fp16)[name = tensor("linear_65_cast_fp16")]; + tensor var_6944 = const()[name = tensor("op_6944"), val = tensor([1, 1, 3, 16, 64])]; + tensor qkv_33_cast_fp16 = reshape(shape = var_6944, x = linear_65_cast_fp16)[name = tensor("qkv_33_cast_fp16")]; + tensor q_97_begin_0 = const()[name = tensor("q_97_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor q_97_end_0 = const()[name = tensor("q_97_end_0"), val = tensor([1, 1, 1, 16, 64])]; + tensor q_97_end_mask_0 = const()[name = tensor("q_97_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor q_97_squeeze_mask_0 = const()[name = tensor("q_97_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor q_97_cast_fp16 = slice_by_index(begin = q_97_begin_0, end = q_97_end_0, end_mask = q_97_end_mask_0, squeeze_mask = q_97_squeeze_mask_0, x = qkv_33_cast_fp16)[name = tensor("q_97_cast_fp16")]; + tensor k_65_begin_0 = const()[name = tensor("k_65_begin_0"), val = tensor([0, 0, 1, 0, 0])]; + tensor k_65_end_0 = const()[name = tensor("k_65_end_0"), val = tensor([1, 1, 2, 16, 64])]; + tensor k_65_end_mask_0 = const()[name = tensor("k_65_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor k_65_squeeze_mask_0 = const()[name = tensor("k_65_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor k_65_cast_fp16 = slice_by_index(begin = k_65_begin_0, end = k_65_end_0, end_mask = k_65_end_mask_0, squeeze_mask = k_65_squeeze_mask_0, x = qkv_33_cast_fp16)[name = tensor("k_65_cast_fp16")]; + tensor v_33_begin_0 = const()[name = tensor("v_33_begin_0"), val = tensor([0, 0, 2, 0, 0])]; + tensor v_33_end_0 = const()[name = tensor("v_33_end_0"), val = tensor([1, 1, 3, 16, 64])]; + tensor v_33_end_mask_0 = const()[name = tensor("v_33_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor v_33_squeeze_mask_0 = const()[name = tensor("v_33_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor v_33_cast_fp16 = slice_by_index(begin = v_33_begin_0, end = v_33_end_0, end_mask = v_33_end_mask_0, squeeze_mask = v_33_squeeze_mask_0, x = qkv_33_cast_fp16)[name = tensor("v_33_cast_fp16")]; + tensor var_7048 = const()[name = tensor("op_7048"), val = tensor([1, 1, 1, 1])]; + tensor position16_to_fp16_dtype_0 = const()[name = tensor("position16_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor position16_to_fp16 = cast(dtype = position16_to_fp16_dtype_0, x = position16)[name = tensor("cast_473")]; + tensor ts_101_cast_fp16 = reshape(shape = var_7048, x = position16_to_fp16)[name = tensor("ts_101_cast_fp16")]; + tensor var_7052 = const()[name = tensor("op_7052"), val = tensor([1, 1, 16, 32, 2])]; + tensor q_complex_33_cast_fp16 = reshape(shape = var_7052, x = q_97_cast_fp16)[name = tensor("q_complex_33_cast_fp16")]; + tensor var_7056 = const()[name = tensor("op_7056"), val = tensor([1, 1, 16, 32, 2])]; + tensor k_complex_33_cast_fp16 = reshape(shape = var_7056, x = k_65_cast_fp16)[name = tensor("k_complex_33_cast_fp16")]; + tensor var_7060_begin_0 = const()[name = tensor("op_7060_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_7060_end_0 = const()[name = tensor("op_7060_end_0"), val = tensor([1, 1, 16, 32, 1])]; + tensor var_7060_end_mask_0 = const()[name = tensor("op_7060_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_7060_squeeze_mask_0 = const()[name = tensor("op_7060_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_7060_cast_fp16 = slice_by_index(begin = var_7060_begin_0, end = var_7060_end_0, end_mask = var_7060_end_mask_0, squeeze_mask = var_7060_squeeze_mask_0, x = q_complex_33_cast_fp16)[name = tensor("op_7060_cast_fp16")]; + tensor var_7068_begin_0 = const()[name = tensor("op_7068_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_7068_end_0 = const()[name = tensor("op_7068_end_0"), val = tensor([1, 1, 16, 32, 2])]; + tensor var_7068_end_mask_0 = const()[name = tensor("op_7068_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_7068_squeeze_mask_0 = const()[name = tensor("op_7068_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_7068_cast_fp16 = slice_by_index(begin = var_7068_begin_0, end = var_7068_end_0, end_mask = var_7068_end_mask_0, squeeze_mask = var_7068_squeeze_mask_0, x = q_complex_33_cast_fp16)[name = tensor("op_7068_cast_fp16")]; + tensor var_7076_begin_0 = const()[name = tensor("op_7076_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_7076_end_0 = const()[name = tensor("op_7076_end_0"), val = tensor([1, 1, 16, 32, 1])]; + tensor var_7076_end_mask_0 = const()[name = tensor("op_7076_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_7076_squeeze_mask_0 = const()[name = tensor("op_7076_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_7076_cast_fp16 = slice_by_index(begin = var_7076_begin_0, end = var_7076_end_0, end_mask = var_7076_end_mask_0, squeeze_mask = var_7076_squeeze_mask_0, x = k_complex_33_cast_fp16)[name = tensor("op_7076_cast_fp16")]; + tensor var_7084_begin_0 = const()[name = tensor("op_7084_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_7084_end_0 = const()[name = tensor("op_7084_end_0"), val = tensor([1, 1, 16, 32, 2])]; + tensor var_7084_end_mask_0 = const()[name = tensor("op_7084_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_7084_squeeze_mask_0 = const()[name = tensor("op_7084_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_7084_cast_fp16 = slice_by_index(begin = var_7084_begin_0, end = var_7084_end_0, end_mask = var_7084_end_mask_0, squeeze_mask = var_7084_squeeze_mask_0, x = k_complex_33_cast_fp16)[name = tensor("op_7084_cast_fp16")]; + tensor freqs_33_to_fp16 = const()[name = tensor("freqs_33_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(303726336)))]; + tensor var_7090_cast_fp16 = mul(x = freqs_33_to_fp16, y = ts_101_cast_fp16)[name = tensor("op_7090_cast_fp16")]; + tensor rotr_33_cast_fp16 = cos(x = var_7090_cast_fp16)[name = tensor("rotr_33_cast_fp16")]; + tensor roti_33_cast_fp16 = sin(x = var_7090_cast_fp16)[name = tensor("roti_33_cast_fp16")]; + tensor var_7094_cast_fp16 = mul(x = var_7060_cast_fp16, y = rotr_33_cast_fp16)[name = tensor("op_7094_cast_fp16")]; + tensor var_7095_cast_fp16 = mul(x = var_7068_cast_fp16, y = roti_33_cast_fp16)[name = tensor("op_7095_cast_fp16")]; + tensor qor_65_cast_fp16 = sub(x = var_7094_cast_fp16, y = var_7095_cast_fp16)[name = tensor("qor_65_cast_fp16")]; + tensor var_7098_cast_fp16 = mul(x = var_7060_cast_fp16, y = roti_33_cast_fp16)[name = tensor("op_7098_cast_fp16")]; + tensor var_7099_cast_fp16 = mul(x = var_7068_cast_fp16, y = rotr_33_cast_fp16)[name = tensor("op_7099_cast_fp16")]; + tensor qoi_65_cast_fp16 = add(x = var_7098_cast_fp16, y = var_7099_cast_fp16)[name = tensor("qoi_65_cast_fp16")]; + tensor var_7102_cast_fp16 = mul(x = var_7076_cast_fp16, y = rotr_33_cast_fp16)[name = tensor("op_7102_cast_fp16")]; + tensor var_7103_cast_fp16 = mul(x = var_7084_cast_fp16, y = roti_33_cast_fp16)[name = tensor("op_7103_cast_fp16")]; + tensor kor_65_cast_fp16 = sub(x = var_7102_cast_fp16, y = var_7103_cast_fp16)[name = tensor("kor_65_cast_fp16")]; + tensor var_7106_cast_fp16 = mul(x = var_7076_cast_fp16, y = roti_33_cast_fp16)[name = tensor("op_7106_cast_fp16")]; + tensor var_7107_cast_fp16 = mul(x = var_7084_cast_fp16, y = rotr_33_cast_fp16)[name = tensor("op_7107_cast_fp16")]; + tensor koi_65_cast_fp16 = add(x = var_7106_cast_fp16, y = var_7107_cast_fp16)[name = tensor("koi_65_cast_fp16")]; + tensor qo_33_axis_0 = const()[name = tensor("qo_33_axis_0"), val = tensor(-1)]; + tensor qo_33_cast_fp16 = stack(axis = qo_33_axis_0, values = (qor_65_cast_fp16, qoi_65_cast_fp16))[name = tensor("qo_33_cast_fp16")]; + tensor ko_33_axis_0 = const()[name = tensor("ko_33_axis_0"), val = tensor(-1)]; + tensor ko_33_cast_fp16 = stack(axis = ko_33_axis_0, values = (kor_65_cast_fp16, koi_65_cast_fp16))[name = tensor("ko_33_cast_fp16")]; + tensor var_7136 = const()[name = tensor("op_7136"), val = tensor([1, 1, 16, 64])]; + tensor q_99_cast_fp16 = reshape(shape = var_7136, x = qo_33_cast_fp16)[name = tensor("q_99_cast_fp16")]; + tensor var_7138 = const()[name = tensor("op_7138"), val = tensor([1, 1, 16, 64])]; + tensor k_67_cast_fp16 = reshape(shape = var_7138, x = ko_33_cast_fp16)[name = tensor("k_67_cast_fp16")]; + tensor _inversed_7160_y_0_to_fp16 = const()[name = tensor("_inversed_7160_y_0_to_fp16"), val = tensor(0x1p-9)]; + tensor _inversed_7160_cast_fp16 = mul(x = ts_101_cast_fp16, y = _inversed_7160_y_0_to_fp16)[name = tensor("_inversed_7160_cast_fp16")]; + tensor var_7161_cast_fp16 = floor(x = _inversed_7160_cast_fp16)[name = tensor("op_7161_cast_fp16")]; + tensor var_7162_to_fp16 = const()[name = tensor("op_7162_to_fp16"), val = tensor(0x1p+9)]; + tensor var_7163_cast_fp16 = mul(x = var_7161_cast_fp16, y = var_7162_to_fp16)[name = tensor("op_7163_cast_fp16")]; + tensor write_indices_float_67_cast_fp16 = sub(x = ts_101_cast_fp16, y = var_7163_cast_fp16)[name = tensor("write_indices_float_67_cast_fp16")]; + tensor var_7170_dtype_0 = const()[name = tensor("op_7170_dtype_0"), val = tensor("int32")]; + tensor write_indices_33_reps_0 = const()[name = tensor("write_indices_33_reps_0"), val = tensor([1, 1, 16, 64])]; + tensor write_indices_float_67_cast_fp16_to_int32 = cast(dtype = var_7170_dtype_0, x = write_indices_float_67_cast_fp16)[name = tensor("cast_472")]; + tensor write_indices_33 = tile(reps = write_indices_33_reps_0, x = write_indices_float_67_cast_fp16_to_int32)[name = tensor("write_indices_33")]; + tensor var_7178_begin_0 = const()[name = tensor("op_7178_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_7178_end_0 = const()[name = tensor("op_7178_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor var_7178_end_mask_0 = const()[name = tensor("op_7178_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_7178_squeeze_mask_0 = const()[name = tensor("op_7178_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor cache16_to_fp16_dtype_0 = const()[name = tensor("cache16_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor cache16_to_fp16 = cast(dtype = cache16_to_fp16_dtype_0, x = cache16)[name = tensor("cast_471")]; + tensor var_7178_cast_fp16 = slice_by_index(begin = var_7178_begin_0, end = var_7178_end_0, end_mask = var_7178_end_mask_0, squeeze_mask = var_7178_squeeze_mask_0, x = cache16_to_fp16)[name = tensor("op_7178_cast_fp16")]; + tensor var_7180_axis_0 = const()[name = tensor("op_7180_axis_0"), val = tensor(1)]; + tensor var_7180_mode_0 = const()[name = tensor("op_7180_mode_0"), val = tensor("update")]; + tensor var_7180_validate_indices_0 = const()[name = tensor("op_7180_validate_indices_0"), val = tensor(false)]; + tensor var_7180_cast_fp16 = scatter_along_axis(axis = var_7180_axis_0, data = var_7178_cast_fp16, indices = write_indices_33, mode = var_7180_mode_0, updates = k_67_cast_fp16, validate_indices = var_7180_validate_indices_0)[name = tensor("op_7180_cast_fp16")]; + tensor concat_114 = const()[name = tensor("concat_114"), val = tensor([0, 0, 0, 0, 0])]; + tensor concat_115 = const()[name = tensor("concat_115"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_33_internal_tensor_assign_1_stride_0 = const()[name = tensor("new_cache_33_internal_tensor_assign_1_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_33_internal_tensor_assign_1_begin_mask_0 = const()[name = tensor("new_cache_33_internal_tensor_assign_1_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_33_internal_tensor_assign_1_end_mask_0 = const()[name = tensor("new_cache_33_internal_tensor_assign_1_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_33_internal_tensor_assign_1_squeeze_mask_0 = const()[name = tensor("new_cache_33_internal_tensor_assign_1_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_80 = const()[name = tensor("shape_80"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_32 = const()[name = tensor("reduce_prod_32"), val = tensor(1048576)]; + tensor range_1d_32_start_0 = const()[name = tensor("range_1d_32_start_0"), val = tensor(0)]; + tensor range_1d_32_step_0 = const()[name = tensor("range_1d_32_step_0"), val = tensor(1)]; + tensor range_1d_32 = range_1d(end = reduce_prod_32, start = range_1d_32_start_0, step = range_1d_32_step_0)[name = tensor("range_1d_32")]; + tensor reshape_160 = reshape(shape = shape_80, x = range_1d_32)[name = tensor("reshape_160")]; + tensor slice_by_index_32 = slice_by_index(begin = concat_114, begin_mask = new_cache_33_internal_tensor_assign_1_begin_mask_0, end = concat_115, end_mask = new_cache_33_internal_tensor_assign_1_end_mask_0, squeeze_mask = new_cache_33_internal_tensor_assign_1_squeeze_mask_0, stride = new_cache_33_internal_tensor_assign_1_stride_0, x = reshape_160)[name = tensor("slice_by_index_32")]; + tensor reshape_161_shape_0 = const()[name = tensor("reshape_161_shape_0"), val = tensor([-1])]; + tensor reshape_161 = reshape(shape = reshape_161_shape_0, x = slice_by_index_32)[name = tensor("reshape_161")]; + tensor reshape_162_shape_0 = const()[name = tensor("reshape_162_shape_0"), val = tensor([-1])]; + tensor reshape_162_cast_fp16 = reshape(shape = reshape_162_shape_0, x = var_7180_cast_fp16)[name = tensor("reshape_162_cast_fp16")]; + tensor reshape_163_shape_0 = const()[name = tensor("reshape_163_shape_0"), val = tensor([-1])]; + tensor reshape_163_cast_fp16 = reshape(shape = reshape_163_shape_0, x = cache16_to_fp16)[name = tensor("reshape_163_cast_fp16")]; + tensor scatter_32_mode_0 = const()[name = tensor("scatter_32_mode_0"), val = tensor("update")]; + tensor scatter_32_axis_0 = const()[name = tensor("scatter_32_axis_0"), val = tensor(0)]; + tensor scatter_32_validate_indices_0 = const()[name = tensor("scatter_32_validate_indices_0"), val = tensor(false)]; + tensor scatter_32_cast_fp16 = scatter(axis = scatter_32_axis_0, data = reshape_163_cast_fp16, indices = reshape_161, mode = scatter_32_mode_0, updates = reshape_162_cast_fp16, validate_indices = scatter_32_validate_indices_0)[name = tensor("scatter_32_cast_fp16")]; + tensor reshape_164_cast_fp16 = reshape(shape = shape_80, x = scatter_32_cast_fp16)[name = tensor("reshape_164_cast_fp16")]; + tensor var_7188_begin_0 = const()[name = tensor("op_7188_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor var_7188_end_0 = const()[name = tensor("op_7188_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor var_7188_end_mask_0 = const()[name = tensor("op_7188_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_7188_squeeze_mask_0 = const()[name = tensor("op_7188_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor var_7188_cast_fp16 = slice_by_index(begin = var_7188_begin_0, end = var_7188_end_0, end_mask = var_7188_end_mask_0, squeeze_mask = var_7188_squeeze_mask_0, x = reshape_164_cast_fp16)[name = tensor("op_7188_cast_fp16")]; + tensor var_7190_axis_0 = const()[name = tensor("op_7190_axis_0"), val = tensor(1)]; + tensor var_7190_mode_0 = const()[name = tensor("op_7190_mode_0"), val = tensor("update")]; + tensor var_7190_validate_indices_0 = const()[name = tensor("op_7190_validate_indices_0"), val = tensor(false)]; + tensor var_7190_cast_fp16 = scatter_along_axis(axis = var_7190_axis_0, data = var_7188_cast_fp16, indices = write_indices_33, mode = var_7190_mode_0, updates = v_33_cast_fp16, validate_indices = var_7190_validate_indices_0)[name = tensor("op_7190_cast_fp16")]; + tensor concat_116 = const()[name = tensor("concat_116"), val = tensor([1, 0, 0, 0, 0])]; + tensor concat_117 = const()[name = tensor("concat_117"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_33_internal_tensor_assign_2_stride_0 = const()[name = tensor("new_cache_33_internal_tensor_assign_2_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_33_internal_tensor_assign_2_begin_mask_0 = const()[name = tensor("new_cache_33_internal_tensor_assign_2_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_33_internal_tensor_assign_2_end_mask_0 = const()[name = tensor("new_cache_33_internal_tensor_assign_2_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_33_internal_tensor_assign_2_squeeze_mask_0 = const()[name = tensor("new_cache_33_internal_tensor_assign_2_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_81 = const()[name = tensor("shape_81"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_33 = const()[name = tensor("reduce_prod_33"), val = tensor(1048576)]; + tensor range_1d_33_start_0 = const()[name = tensor("range_1d_33_start_0"), val = tensor(0)]; + tensor range_1d_33_step_0 = const()[name = tensor("range_1d_33_step_0"), val = tensor(1)]; + tensor range_1d_33 = range_1d(end = reduce_prod_33, start = range_1d_33_start_0, step = range_1d_33_step_0)[name = tensor("range_1d_33")]; + tensor reshape_165 = reshape(shape = shape_81, x = range_1d_33)[name = tensor("reshape_165")]; + tensor slice_by_index_33 = slice_by_index(begin = concat_116, begin_mask = new_cache_33_internal_tensor_assign_2_begin_mask_0, end = concat_117, end_mask = new_cache_33_internal_tensor_assign_2_end_mask_0, squeeze_mask = new_cache_33_internal_tensor_assign_2_squeeze_mask_0, stride = new_cache_33_internal_tensor_assign_2_stride_0, x = reshape_165)[name = tensor("slice_by_index_33")]; + tensor reshape_166_shape_0 = const()[name = tensor("reshape_166_shape_0"), val = tensor([-1])]; + tensor reshape_166 = reshape(shape = reshape_166_shape_0, x = slice_by_index_33)[name = tensor("reshape_166")]; + tensor reshape_167_shape_0 = const()[name = tensor("reshape_167_shape_0"), val = tensor([-1])]; + tensor reshape_167_cast_fp16 = reshape(shape = reshape_167_shape_0, x = var_7190_cast_fp16)[name = tensor("reshape_167_cast_fp16")]; + tensor reshape_168_shape_0 = const()[name = tensor("reshape_168_shape_0"), val = tensor([-1])]; + tensor reshape_168_cast_fp16 = reshape(shape = reshape_168_shape_0, x = reshape_164_cast_fp16)[name = tensor("reshape_168_cast_fp16")]; + tensor scatter_33_mode_0 = const()[name = tensor("scatter_33_mode_0"), val = tensor("update")]; + tensor scatter_33_axis_0 = const()[name = tensor("scatter_33_axis_0"), val = tensor(0)]; + tensor scatter_33_validate_indices_0 = const()[name = tensor("scatter_33_validate_indices_0"), val = tensor(false)]; + tensor scatter_33_cast_fp16 = scatter(axis = scatter_33_axis_0, data = reshape_168_cast_fp16, indices = reshape_166, mode = scatter_33_mode_0, updates = reshape_167_cast_fp16, validate_indices = scatter_33_validate_indices_0)[name = tensor("scatter_33_cast_fp16")]; + tensor reshape_169_cast_fp16 = reshape(shape = shape_81, x = scatter_33_cast_fp16)[name = tensor("reshape_169_cast_fp16")]; + tensor reshape_169_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("reshape_169_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor keys_97_begin_0 = const()[name = tensor("keys_97_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor keys_97_end_0 = const()[name = tensor("keys_97_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor keys_97_end_mask_0 = const()[name = tensor("keys_97_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor keys_97_squeeze_mask_0 = const()[name = tensor("keys_97_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor keys_97_cast_fp16 = slice_by_index(begin = keys_97_begin_0, end = keys_97_end_0, end_mask = keys_97_end_mask_0, squeeze_mask = keys_97_squeeze_mask_0, x = reshape_169_cast_fp16)[name = tensor("keys_97_cast_fp16")]; + tensor values_97_begin_0 = const()[name = tensor("values_97_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor values_97_end_0 = const()[name = tensor("values_97_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor values_97_end_mask_0 = const()[name = tensor("values_97_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor values_97_squeeze_mask_0 = const()[name = tensor("values_97_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor values_97_cast_fp16 = slice_by_index(begin = values_97_begin_0, end = values_97_end_0, end_mask = values_97_end_mask_0, squeeze_mask = values_97_squeeze_mask_0, x = reshape_169_cast_fp16)[name = tensor("values_97_cast_fp16")]; + tensor var_7202_cast_fp16 = not_equal(x = keys_97_cast_fp16, y = keys_97_cast_fp16)[name = tensor("op_7202_cast_fp16")]; + tensor keys_99_cast_fp16 = select(a = var_504_to_fp16, b = keys_97_cast_fp16, cond = var_7202_cast_fp16)[name = tensor("keys_99_cast_fp16")]; + tensor var_7210_cast_fp16 = not_equal(x = values_97_cast_fp16, y = values_97_cast_fp16)[name = tensor("op_7210_cast_fp16")]; + tensor values_99_cast_fp16 = select(a = var_504_to_fp16, b = values_97_cast_fp16, cond = var_7210_cast_fp16)[name = tensor("values_99_cast_fp16")]; + tensor var_7234 = const()[name = tensor("op_7234"), val = tensor([0, 2, 1, 3])]; + tensor var_7247 = const()[name = tensor("op_7247"), val = tensor([1, 1, 1])]; + tensor var_7248_cast_fp16 = reshape(shape = var_7247, x = position16_to_fp16)[name = tensor("op_7248_cast_fp16")]; + tensor var_7265_to_fp16 = const()[name = tensor("op_7265_to_fp16"), val = tensor(0x1p+0)]; + tensor valid_len_33_cast_fp16 = add(x = var_7248_cast_fp16, y = var_7265_to_fp16)[name = tensor("valid_len_33_cast_fp16")]; + tensor valid_mask_33_cast_fp16 = less(x = k_positions_1_promoted_to_fp16, y = valid_len_33_cast_fp16)[name = tensor("valid_mask_33_cast_fp16")]; + tensor causal_mask_33_cast_fp16 = less_equal(x = k_positions_1_promoted_to_fp16, y = var_7248_cast_fp16)[name = tensor("causal_mask_33_cast_fp16")]; + tensor attn_mask_65 = logical_and(x = valid_mask_33_cast_fp16, y = causal_mask_33_cast_fp16)[name = tensor("attn_mask_65")]; + tensor attn_mask_67_axes_0 = const()[name = tensor("attn_mask_67_axes_0"), val = tensor([1])]; + tensor attn_mask_67 = expand_dims(axes = attn_mask_67_axes_0, x = attn_mask_65)[name = tensor("attn_mask_67")]; + tensor var_7283_transpose_x_0 = const()[name = tensor("op_7283_transpose_x_0"), val = tensor(false)]; + tensor var_7283_transpose_y_0 = const()[name = tensor("op_7283_transpose_y_0"), val = tensor(false)]; + tensor transpose_104_perm_0 = const()[name = tensor("transpose_104_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_105_perm_0 = const()[name = tensor("transpose_105_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_105 = transpose(perm = transpose_105_perm_0, x = keys_99_cast_fp16)[name = tensor("transpose_149")]; + tensor transpose_104 = transpose(perm = transpose_104_perm_0, x = q_99_cast_fp16)[name = tensor("transpose_150")]; + tensor var_7283_cast_fp16 = matmul(transpose_x = var_7283_transpose_x_0, transpose_y = var_7283_transpose_y_0, x = transpose_104, y = transpose_105)[name = tensor("op_7283_cast_fp16")]; + tensor var_7277_to_fp16 = const()[name = tensor("op_7277_to_fp16"), val = tensor([0x1p-3])]; + tensor attn_weights_97_cast_fp16 = mul(x = var_7283_cast_fp16, y = var_7277_to_fp16)[name = tensor("attn_weights_97_cast_fp16")]; + tensor var_7285 = logical_not(x = attn_mask_67)[name = tensor("op_7285")]; + tensor var_7286_to_fp16 = const()[name = tensor("op_7286_to_fp16"), val = tensor(-inf)]; + tensor attn_weights_99_cast_fp16 = select(a = var_7286_to_fp16, b = attn_weights_97_cast_fp16, cond = var_7285)[name = tensor("attn_weights_99_cast_fp16")]; + tensor var_7288 = const()[name = tensor("op_7288"), val = tensor(-1)]; + tensor attn_weights_101_cast_fp16 = softmax(axis = var_7288, x = attn_weights_99_cast_fp16)[name = tensor("attn_weights_101_cast_fp16")]; + tensor attn_output_33_transpose_x_0 = const()[name = tensor("attn_output_33_transpose_x_0"), val = tensor(false)]; + tensor attn_output_33_transpose_y_0 = const()[name = tensor("attn_output_33_transpose_y_0"), val = tensor(false)]; + tensor values_101_cast_fp16 = transpose(perm = var_7234, x = values_99_cast_fp16)[name = tensor("transpose_151")]; + tensor attn_output_33_cast_fp16 = matmul(transpose_x = attn_output_33_transpose_x_0, transpose_y = attn_output_33_transpose_y_0, x = attn_weights_101_cast_fp16, y = values_101_cast_fp16)[name = tensor("attn_output_33_cast_fp16")]; + tensor var_7296 = const()[name = tensor("op_7296"), val = tensor([0, 2, 1, 3])]; + tensor var_7299 = const()[name = tensor("op_7299"), val = tensor([1, 1, 1024])]; + tensor var_7297_cast_fp16 = transpose(perm = var_7296, x = attn_output_33_cast_fp16)[name = tensor("transpose_148")]; + tensor input_165_cast_fp16 = reshape(shape = var_7299, x = var_7297_cast_fp16)[name = tensor("input_165_cast_fp16")]; + tensor linear_66_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = attn16_out_proj_weight_cast_fp16, x = input_165_cast_fp16)[name = tensor("linear_66_cast_fp16")]; + tensor var_7305_to_fp16 = const()[name = tensor("op_7305_to_fp16"), val = tensor(0x1p+0)]; + tensor var_7306_cast_fp16 = add(x = position16_to_fp16, y = var_7305_to_fp16)[name = tensor("op_7306_cast_fp16")]; + tensor var_7306_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_7306_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor input_167_cast_fp16 = add(x = input_163_cast_fp16, y = linear_66_cast_fp16)[name = tensor("input_167_cast_fp16")]; + tensor input_169_axes_0 = const()[name = tensor("input_169_axes_0"), val = tensor([-1])]; + tensor norm16_2_weight_to_fp16 = const()[name = tensor("norm16_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(303726464)))]; + tensor norm16_2_bias_to_fp16 = const()[name = tensor("norm16_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(303728576)))]; + tensor var_7310_to_fp16 = const()[name = tensor("op_7310_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_169_cast_fp16 = layer_norm(axes = input_169_axes_0, beta = norm16_2_bias_to_fp16, epsilon = var_7310_to_fp16, gamma = norm16_2_weight_to_fp16, x = input_167_cast_fp16)[name = tensor("input_169_cast_fp16")]; + tensor linear_67_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = linear16_1_weight_cast_fp16, x = input_169_cast_fp16)[name = tensor("linear_67_cast_fp16")]; + tensor input_171_mode_0 = const()[name = tensor("input_171_mode_0"), val = tensor("EXACT")]; + tensor input_171_cast_fp16 = gelu(mode = input_171_mode_0, x = linear_67_cast_fp16)[name = tensor("input_171_cast_fp16")]; + tensor linear_68_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = linear16_2_weight_cast_fp16, x = input_171_cast_fp16)[name = tensor("linear_68_cast_fp16")]; + tensor input_173_cast_fp16 = add(x = input_167_cast_fp16, y = linear_68_cast_fp16)[name = tensor("input_173_cast_fp16")]; + tensor x_35_axes_0 = const()[name = tensor("x_35_axes_0"), val = tensor([-1])]; + tensor norm17_1_weight_to_fp16 = const()[name = tensor("norm17_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(303730688)))]; + tensor norm17_1_bias_to_fp16 = const()[name = tensor("norm17_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(303732800)))]; + tensor var_7327_to_fp16 = const()[name = tensor("op_7327_to_fp16"), val = tensor(0x1.5p-17)]; + tensor x_35_cast_fp16 = layer_norm(axes = x_35_axes_0, beta = norm17_1_bias_to_fp16, epsilon = var_7327_to_fp16, gamma = norm17_1_weight_to_fp16, x = input_173_cast_fp16)[name = tensor("x_35_cast_fp16")]; + tensor linear_69_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = attn17_in_proj_weight_cast_fp16, x = x_35_cast_fp16)[name = tensor("linear_69_cast_fp16")]; + tensor var_7363 = const()[name = tensor("op_7363"), val = tensor([1, 1, 3, 16, 64])]; + tensor qkv_35_cast_fp16 = reshape(shape = var_7363, x = linear_69_cast_fp16)[name = tensor("qkv_35_cast_fp16")]; + tensor q_103_begin_0 = const()[name = tensor("q_103_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor q_103_end_0 = const()[name = tensor("q_103_end_0"), val = tensor([1, 1, 1, 16, 64])]; + tensor q_103_end_mask_0 = const()[name = tensor("q_103_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor q_103_squeeze_mask_0 = const()[name = tensor("q_103_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor q_103_cast_fp16 = slice_by_index(begin = q_103_begin_0, end = q_103_end_0, end_mask = q_103_end_mask_0, squeeze_mask = q_103_squeeze_mask_0, x = qkv_35_cast_fp16)[name = tensor("q_103_cast_fp16")]; + tensor k_69_begin_0 = const()[name = tensor("k_69_begin_0"), val = tensor([0, 0, 1, 0, 0])]; + tensor k_69_end_0 = const()[name = tensor("k_69_end_0"), val = tensor([1, 1, 2, 16, 64])]; + tensor k_69_end_mask_0 = const()[name = tensor("k_69_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor k_69_squeeze_mask_0 = const()[name = tensor("k_69_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor k_69_cast_fp16 = slice_by_index(begin = k_69_begin_0, end = k_69_end_0, end_mask = k_69_end_mask_0, squeeze_mask = k_69_squeeze_mask_0, x = qkv_35_cast_fp16)[name = tensor("k_69_cast_fp16")]; + tensor v_35_begin_0 = const()[name = tensor("v_35_begin_0"), val = tensor([0, 0, 2, 0, 0])]; + tensor v_35_end_0 = const()[name = tensor("v_35_end_0"), val = tensor([1, 1, 3, 16, 64])]; + tensor v_35_end_mask_0 = const()[name = tensor("v_35_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor v_35_squeeze_mask_0 = const()[name = tensor("v_35_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor v_35_cast_fp16 = slice_by_index(begin = v_35_begin_0, end = v_35_end_0, end_mask = v_35_end_mask_0, squeeze_mask = v_35_squeeze_mask_0, x = qkv_35_cast_fp16)[name = tensor("v_35_cast_fp16")]; + tensor var_7467 = const()[name = tensor("op_7467"), val = tensor([1, 1, 1, 1])]; + tensor position17_to_fp16_dtype_0 = const()[name = tensor("position17_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor position17_to_fp16 = cast(dtype = position17_to_fp16_dtype_0, x = position17)[name = tensor("cast_468")]; + tensor ts_107_cast_fp16 = reshape(shape = var_7467, x = position17_to_fp16)[name = tensor("ts_107_cast_fp16")]; + tensor var_7471 = const()[name = tensor("op_7471"), val = tensor([1, 1, 16, 32, 2])]; + tensor q_complex_35_cast_fp16 = reshape(shape = var_7471, x = q_103_cast_fp16)[name = tensor("q_complex_35_cast_fp16")]; + tensor var_7475 = const()[name = tensor("op_7475"), val = tensor([1, 1, 16, 32, 2])]; + tensor k_complex_35_cast_fp16 = reshape(shape = var_7475, x = k_69_cast_fp16)[name = tensor("k_complex_35_cast_fp16")]; + tensor var_7479_begin_0 = const()[name = tensor("op_7479_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_7479_end_0 = const()[name = tensor("op_7479_end_0"), val = tensor([1, 1, 16, 32, 1])]; + tensor var_7479_end_mask_0 = const()[name = tensor("op_7479_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_7479_squeeze_mask_0 = const()[name = tensor("op_7479_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_7479_cast_fp16 = slice_by_index(begin = var_7479_begin_0, end = var_7479_end_0, end_mask = var_7479_end_mask_0, squeeze_mask = var_7479_squeeze_mask_0, x = q_complex_35_cast_fp16)[name = tensor("op_7479_cast_fp16")]; + tensor var_7487_begin_0 = const()[name = tensor("op_7487_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_7487_end_0 = const()[name = tensor("op_7487_end_0"), val = tensor([1, 1, 16, 32, 2])]; + tensor var_7487_end_mask_0 = const()[name = tensor("op_7487_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_7487_squeeze_mask_0 = const()[name = tensor("op_7487_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_7487_cast_fp16 = slice_by_index(begin = var_7487_begin_0, end = var_7487_end_0, end_mask = var_7487_end_mask_0, squeeze_mask = var_7487_squeeze_mask_0, x = q_complex_35_cast_fp16)[name = tensor("op_7487_cast_fp16")]; + tensor var_7495_begin_0 = const()[name = tensor("op_7495_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_7495_end_0 = const()[name = tensor("op_7495_end_0"), val = tensor([1, 1, 16, 32, 1])]; + tensor var_7495_end_mask_0 = const()[name = tensor("op_7495_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_7495_squeeze_mask_0 = const()[name = tensor("op_7495_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_7495_cast_fp16 = slice_by_index(begin = var_7495_begin_0, end = var_7495_end_0, end_mask = var_7495_end_mask_0, squeeze_mask = var_7495_squeeze_mask_0, x = k_complex_35_cast_fp16)[name = tensor("op_7495_cast_fp16")]; + tensor var_7503_begin_0 = const()[name = tensor("op_7503_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_7503_end_0 = const()[name = tensor("op_7503_end_0"), val = tensor([1, 1, 16, 32, 2])]; + tensor var_7503_end_mask_0 = const()[name = tensor("op_7503_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_7503_squeeze_mask_0 = const()[name = tensor("op_7503_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_7503_cast_fp16 = slice_by_index(begin = var_7503_begin_0, end = var_7503_end_0, end_mask = var_7503_end_mask_0, squeeze_mask = var_7503_squeeze_mask_0, x = k_complex_35_cast_fp16)[name = tensor("op_7503_cast_fp16")]; + tensor freqs_35_to_fp16 = const()[name = tensor("freqs_35_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(303734912)))]; + tensor var_7509_cast_fp16 = mul(x = freqs_35_to_fp16, y = ts_107_cast_fp16)[name = tensor("op_7509_cast_fp16")]; + tensor rotr_35_cast_fp16 = cos(x = var_7509_cast_fp16)[name = tensor("rotr_35_cast_fp16")]; + tensor roti_35_cast_fp16 = sin(x = var_7509_cast_fp16)[name = tensor("roti_35_cast_fp16")]; + tensor var_7513_cast_fp16 = mul(x = var_7479_cast_fp16, y = rotr_35_cast_fp16)[name = tensor("op_7513_cast_fp16")]; + tensor var_7514_cast_fp16 = mul(x = var_7487_cast_fp16, y = roti_35_cast_fp16)[name = tensor("op_7514_cast_fp16")]; + tensor qor_69_cast_fp16 = sub(x = var_7513_cast_fp16, y = var_7514_cast_fp16)[name = tensor("qor_69_cast_fp16")]; + tensor var_7517_cast_fp16 = mul(x = var_7479_cast_fp16, y = roti_35_cast_fp16)[name = tensor("op_7517_cast_fp16")]; + tensor var_7518_cast_fp16 = mul(x = var_7487_cast_fp16, y = rotr_35_cast_fp16)[name = tensor("op_7518_cast_fp16")]; + tensor qoi_69_cast_fp16 = add(x = var_7517_cast_fp16, y = var_7518_cast_fp16)[name = tensor("qoi_69_cast_fp16")]; + tensor var_7521_cast_fp16 = mul(x = var_7495_cast_fp16, y = rotr_35_cast_fp16)[name = tensor("op_7521_cast_fp16")]; + tensor var_7522_cast_fp16 = mul(x = var_7503_cast_fp16, y = roti_35_cast_fp16)[name = tensor("op_7522_cast_fp16")]; + tensor kor_69_cast_fp16 = sub(x = var_7521_cast_fp16, y = var_7522_cast_fp16)[name = tensor("kor_69_cast_fp16")]; + tensor var_7525_cast_fp16 = mul(x = var_7495_cast_fp16, y = roti_35_cast_fp16)[name = tensor("op_7525_cast_fp16")]; + tensor var_7526_cast_fp16 = mul(x = var_7503_cast_fp16, y = rotr_35_cast_fp16)[name = tensor("op_7526_cast_fp16")]; + tensor koi_69_cast_fp16 = add(x = var_7525_cast_fp16, y = var_7526_cast_fp16)[name = tensor("koi_69_cast_fp16")]; + tensor qo_35_axis_0 = const()[name = tensor("qo_35_axis_0"), val = tensor(-1)]; + tensor qo_35_cast_fp16 = stack(axis = qo_35_axis_0, values = (qor_69_cast_fp16, qoi_69_cast_fp16))[name = tensor("qo_35_cast_fp16")]; + tensor ko_35_axis_0 = const()[name = tensor("ko_35_axis_0"), val = tensor(-1)]; + tensor ko_35_cast_fp16 = stack(axis = ko_35_axis_0, values = (kor_69_cast_fp16, koi_69_cast_fp16))[name = tensor("ko_35_cast_fp16")]; + tensor var_7555 = const()[name = tensor("op_7555"), val = tensor([1, 1, 16, 64])]; + tensor q_105_cast_fp16 = reshape(shape = var_7555, x = qo_35_cast_fp16)[name = tensor("q_105_cast_fp16")]; + tensor var_7557 = const()[name = tensor("op_7557"), val = tensor([1, 1, 16, 64])]; + tensor k_71_cast_fp16 = reshape(shape = var_7557, x = ko_35_cast_fp16)[name = tensor("k_71_cast_fp16")]; + tensor _inversed_7579_y_0_to_fp16 = const()[name = tensor("_inversed_7579_y_0_to_fp16"), val = tensor(0x1p-9)]; + tensor _inversed_7579_cast_fp16 = mul(x = ts_107_cast_fp16, y = _inversed_7579_y_0_to_fp16)[name = tensor("_inversed_7579_cast_fp16")]; + tensor var_7580_cast_fp16 = floor(x = _inversed_7579_cast_fp16)[name = tensor("op_7580_cast_fp16")]; + tensor var_7581_to_fp16 = const()[name = tensor("op_7581_to_fp16"), val = tensor(0x1p+9)]; + tensor var_7582_cast_fp16 = mul(x = var_7580_cast_fp16, y = var_7581_to_fp16)[name = tensor("op_7582_cast_fp16")]; + tensor write_indices_float_71_cast_fp16 = sub(x = ts_107_cast_fp16, y = var_7582_cast_fp16)[name = tensor("write_indices_float_71_cast_fp16")]; + tensor var_7589_dtype_0 = const()[name = tensor("op_7589_dtype_0"), val = tensor("int32")]; + tensor write_indices_35_reps_0 = const()[name = tensor("write_indices_35_reps_0"), val = tensor([1, 1, 16, 64])]; + tensor write_indices_float_71_cast_fp16_to_int32 = cast(dtype = var_7589_dtype_0, x = write_indices_float_71_cast_fp16)[name = tensor("cast_467")]; + tensor write_indices_35 = tile(reps = write_indices_35_reps_0, x = write_indices_float_71_cast_fp16_to_int32)[name = tensor("write_indices_35")]; + tensor var_7597_begin_0 = const()[name = tensor("op_7597_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_7597_end_0 = const()[name = tensor("op_7597_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor var_7597_end_mask_0 = const()[name = tensor("op_7597_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_7597_squeeze_mask_0 = const()[name = tensor("op_7597_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor cache17_to_fp16_dtype_0 = const()[name = tensor("cache17_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor cache17_to_fp16 = cast(dtype = cache17_to_fp16_dtype_0, x = cache17)[name = tensor("cast_466")]; + tensor var_7597_cast_fp16 = slice_by_index(begin = var_7597_begin_0, end = var_7597_end_0, end_mask = var_7597_end_mask_0, squeeze_mask = var_7597_squeeze_mask_0, x = cache17_to_fp16)[name = tensor("op_7597_cast_fp16")]; + tensor var_7599_axis_0 = const()[name = tensor("op_7599_axis_0"), val = tensor(1)]; + tensor var_7599_mode_0 = const()[name = tensor("op_7599_mode_0"), val = tensor("update")]; + tensor var_7599_validate_indices_0 = const()[name = tensor("op_7599_validate_indices_0"), val = tensor(false)]; + tensor var_7599_cast_fp16 = scatter_along_axis(axis = var_7599_axis_0, data = var_7597_cast_fp16, indices = write_indices_35, mode = var_7599_mode_0, updates = k_71_cast_fp16, validate_indices = var_7599_validate_indices_0)[name = tensor("op_7599_cast_fp16")]; + tensor concat_121 = const()[name = tensor("concat_121"), val = tensor([0, 0, 0, 0, 0])]; + tensor concat_122 = const()[name = tensor("concat_122"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_35_internal_tensor_assign_1_stride_0 = const()[name = tensor("new_cache_35_internal_tensor_assign_1_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_35_internal_tensor_assign_1_begin_mask_0 = const()[name = tensor("new_cache_35_internal_tensor_assign_1_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_35_internal_tensor_assign_1_end_mask_0 = const()[name = tensor("new_cache_35_internal_tensor_assign_1_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_35_internal_tensor_assign_1_squeeze_mask_0 = const()[name = tensor("new_cache_35_internal_tensor_assign_1_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_82 = const()[name = tensor("shape_82"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_34 = const()[name = tensor("reduce_prod_34"), val = tensor(1048576)]; + tensor range_1d_34_start_0 = const()[name = tensor("range_1d_34_start_0"), val = tensor(0)]; + tensor range_1d_34_step_0 = const()[name = tensor("range_1d_34_step_0"), val = tensor(1)]; + tensor range_1d_34 = range_1d(end = reduce_prod_34, start = range_1d_34_start_0, step = range_1d_34_step_0)[name = tensor("range_1d_34")]; + tensor reshape_170 = reshape(shape = shape_82, x = range_1d_34)[name = tensor("reshape_170")]; + tensor slice_by_index_34 = slice_by_index(begin = concat_121, begin_mask = new_cache_35_internal_tensor_assign_1_begin_mask_0, end = concat_122, end_mask = new_cache_35_internal_tensor_assign_1_end_mask_0, squeeze_mask = new_cache_35_internal_tensor_assign_1_squeeze_mask_0, stride = new_cache_35_internal_tensor_assign_1_stride_0, x = reshape_170)[name = tensor("slice_by_index_34")]; + tensor reshape_171_shape_0 = const()[name = tensor("reshape_171_shape_0"), val = tensor([-1])]; + tensor reshape_171 = reshape(shape = reshape_171_shape_0, x = slice_by_index_34)[name = tensor("reshape_171")]; + tensor reshape_172_shape_0 = const()[name = tensor("reshape_172_shape_0"), val = tensor([-1])]; + tensor reshape_172_cast_fp16 = reshape(shape = reshape_172_shape_0, x = var_7599_cast_fp16)[name = tensor("reshape_172_cast_fp16")]; + tensor reshape_173_shape_0 = const()[name = tensor("reshape_173_shape_0"), val = tensor([-1])]; + tensor reshape_173_cast_fp16 = reshape(shape = reshape_173_shape_0, x = cache17_to_fp16)[name = tensor("reshape_173_cast_fp16")]; + tensor scatter_34_mode_0 = const()[name = tensor("scatter_34_mode_0"), val = tensor("update")]; + tensor scatter_34_axis_0 = const()[name = tensor("scatter_34_axis_0"), val = tensor(0)]; + tensor scatter_34_validate_indices_0 = const()[name = tensor("scatter_34_validate_indices_0"), val = tensor(false)]; + tensor scatter_34_cast_fp16 = scatter(axis = scatter_34_axis_0, data = reshape_173_cast_fp16, indices = reshape_171, mode = scatter_34_mode_0, updates = reshape_172_cast_fp16, validate_indices = scatter_34_validate_indices_0)[name = tensor("scatter_34_cast_fp16")]; + tensor reshape_174_cast_fp16 = reshape(shape = shape_82, x = scatter_34_cast_fp16)[name = tensor("reshape_174_cast_fp16")]; + tensor var_7607_begin_0 = const()[name = tensor("op_7607_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor var_7607_end_0 = const()[name = tensor("op_7607_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor var_7607_end_mask_0 = const()[name = tensor("op_7607_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_7607_squeeze_mask_0 = const()[name = tensor("op_7607_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor var_7607_cast_fp16 = slice_by_index(begin = var_7607_begin_0, end = var_7607_end_0, end_mask = var_7607_end_mask_0, squeeze_mask = var_7607_squeeze_mask_0, x = reshape_174_cast_fp16)[name = tensor("op_7607_cast_fp16")]; + tensor var_7609_axis_0 = const()[name = tensor("op_7609_axis_0"), val = tensor(1)]; + tensor var_7609_mode_0 = const()[name = tensor("op_7609_mode_0"), val = tensor("update")]; + tensor var_7609_validate_indices_0 = const()[name = tensor("op_7609_validate_indices_0"), val = tensor(false)]; + tensor var_7609_cast_fp16 = scatter_along_axis(axis = var_7609_axis_0, data = var_7607_cast_fp16, indices = write_indices_35, mode = var_7609_mode_0, updates = v_35_cast_fp16, validate_indices = var_7609_validate_indices_0)[name = tensor("op_7609_cast_fp16")]; + tensor concat_123 = const()[name = tensor("concat_123"), val = tensor([1, 0, 0, 0, 0])]; + tensor concat_124 = const()[name = tensor("concat_124"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_35_internal_tensor_assign_2_stride_0 = const()[name = tensor("new_cache_35_internal_tensor_assign_2_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_35_internal_tensor_assign_2_begin_mask_0 = const()[name = tensor("new_cache_35_internal_tensor_assign_2_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_35_internal_tensor_assign_2_end_mask_0 = const()[name = tensor("new_cache_35_internal_tensor_assign_2_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_35_internal_tensor_assign_2_squeeze_mask_0 = const()[name = tensor("new_cache_35_internal_tensor_assign_2_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_83 = const()[name = tensor("shape_83"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_35 = const()[name = tensor("reduce_prod_35"), val = tensor(1048576)]; + tensor range_1d_35_start_0 = const()[name = tensor("range_1d_35_start_0"), val = tensor(0)]; + tensor range_1d_35_step_0 = const()[name = tensor("range_1d_35_step_0"), val = tensor(1)]; + tensor range_1d_35 = range_1d(end = reduce_prod_35, start = range_1d_35_start_0, step = range_1d_35_step_0)[name = tensor("range_1d_35")]; + tensor reshape_175 = reshape(shape = shape_83, x = range_1d_35)[name = tensor("reshape_175")]; + tensor slice_by_index_35 = slice_by_index(begin = concat_123, begin_mask = new_cache_35_internal_tensor_assign_2_begin_mask_0, end = concat_124, end_mask = new_cache_35_internal_tensor_assign_2_end_mask_0, squeeze_mask = new_cache_35_internal_tensor_assign_2_squeeze_mask_0, stride = new_cache_35_internal_tensor_assign_2_stride_0, x = reshape_175)[name = tensor("slice_by_index_35")]; + tensor reshape_176_shape_0 = const()[name = tensor("reshape_176_shape_0"), val = tensor([-1])]; + tensor reshape_176 = reshape(shape = reshape_176_shape_0, x = slice_by_index_35)[name = tensor("reshape_176")]; + tensor reshape_177_shape_0 = const()[name = tensor("reshape_177_shape_0"), val = tensor([-1])]; + tensor reshape_177_cast_fp16 = reshape(shape = reshape_177_shape_0, x = var_7609_cast_fp16)[name = tensor("reshape_177_cast_fp16")]; + tensor reshape_178_shape_0 = const()[name = tensor("reshape_178_shape_0"), val = tensor([-1])]; + tensor reshape_178_cast_fp16 = reshape(shape = reshape_178_shape_0, x = reshape_174_cast_fp16)[name = tensor("reshape_178_cast_fp16")]; + tensor scatter_35_mode_0 = const()[name = tensor("scatter_35_mode_0"), val = tensor("update")]; + tensor scatter_35_axis_0 = const()[name = tensor("scatter_35_axis_0"), val = tensor(0)]; + tensor scatter_35_validate_indices_0 = const()[name = tensor("scatter_35_validate_indices_0"), val = tensor(false)]; + tensor scatter_35_cast_fp16 = scatter(axis = scatter_35_axis_0, data = reshape_178_cast_fp16, indices = reshape_176, mode = scatter_35_mode_0, updates = reshape_177_cast_fp16, validate_indices = scatter_35_validate_indices_0)[name = tensor("scatter_35_cast_fp16")]; + tensor reshape_179_cast_fp16 = reshape(shape = shape_83, x = scatter_35_cast_fp16)[name = tensor("reshape_179_cast_fp16")]; + tensor reshape_179_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("reshape_179_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor keys_103_begin_0 = const()[name = tensor("keys_103_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor keys_103_end_0 = const()[name = tensor("keys_103_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor keys_103_end_mask_0 = const()[name = tensor("keys_103_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor keys_103_squeeze_mask_0 = const()[name = tensor("keys_103_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor keys_103_cast_fp16 = slice_by_index(begin = keys_103_begin_0, end = keys_103_end_0, end_mask = keys_103_end_mask_0, squeeze_mask = keys_103_squeeze_mask_0, x = reshape_179_cast_fp16)[name = tensor("keys_103_cast_fp16")]; + tensor values_103_begin_0 = const()[name = tensor("values_103_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor values_103_end_0 = const()[name = tensor("values_103_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor values_103_end_mask_0 = const()[name = tensor("values_103_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor values_103_squeeze_mask_0 = const()[name = tensor("values_103_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor values_103_cast_fp16 = slice_by_index(begin = values_103_begin_0, end = values_103_end_0, end_mask = values_103_end_mask_0, squeeze_mask = values_103_squeeze_mask_0, x = reshape_179_cast_fp16)[name = tensor("values_103_cast_fp16")]; + tensor var_7621_cast_fp16 = not_equal(x = keys_103_cast_fp16, y = keys_103_cast_fp16)[name = tensor("op_7621_cast_fp16")]; + tensor keys_105_cast_fp16 = select(a = var_504_to_fp16, b = keys_103_cast_fp16, cond = var_7621_cast_fp16)[name = tensor("keys_105_cast_fp16")]; + tensor var_7629_cast_fp16 = not_equal(x = values_103_cast_fp16, y = values_103_cast_fp16)[name = tensor("op_7629_cast_fp16")]; + tensor values_105_cast_fp16 = select(a = var_504_to_fp16, b = values_103_cast_fp16, cond = var_7629_cast_fp16)[name = tensor("values_105_cast_fp16")]; + tensor var_7653 = const()[name = tensor("op_7653"), val = tensor([0, 2, 1, 3])]; + tensor var_7666 = const()[name = tensor("op_7666"), val = tensor([1, 1, 1])]; + tensor var_7667_cast_fp16 = reshape(shape = var_7666, x = position17_to_fp16)[name = tensor("op_7667_cast_fp16")]; + tensor var_7684_to_fp16 = const()[name = tensor("op_7684_to_fp16"), val = tensor(0x1p+0)]; + tensor valid_len_35_cast_fp16 = add(x = var_7667_cast_fp16, y = var_7684_to_fp16)[name = tensor("valid_len_35_cast_fp16")]; + tensor valid_mask_35_cast_fp16 = less(x = k_positions_1_promoted_to_fp16, y = valid_len_35_cast_fp16)[name = tensor("valid_mask_35_cast_fp16")]; + tensor causal_mask_35_cast_fp16 = less_equal(x = k_positions_1_promoted_to_fp16, y = var_7667_cast_fp16)[name = tensor("causal_mask_35_cast_fp16")]; + tensor attn_mask_69 = logical_and(x = valid_mask_35_cast_fp16, y = causal_mask_35_cast_fp16)[name = tensor("attn_mask_69")]; + tensor attn_mask_71_axes_0 = const()[name = tensor("attn_mask_71_axes_0"), val = tensor([1])]; + tensor attn_mask_71 = expand_dims(axes = attn_mask_71_axes_0, x = attn_mask_69)[name = tensor("attn_mask_71")]; + tensor var_7702_transpose_x_0 = const()[name = tensor("op_7702_transpose_x_0"), val = tensor(false)]; + tensor var_7702_transpose_y_0 = const()[name = tensor("op_7702_transpose_y_0"), val = tensor(false)]; + tensor transpose_106_perm_0 = const()[name = tensor("transpose_106_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_107_perm_0 = const()[name = tensor("transpose_107_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_107 = transpose(perm = transpose_107_perm_0, x = keys_105_cast_fp16)[name = tensor("transpose_145")]; + tensor transpose_106 = transpose(perm = transpose_106_perm_0, x = q_105_cast_fp16)[name = tensor("transpose_146")]; + tensor var_7702_cast_fp16 = matmul(transpose_x = var_7702_transpose_x_0, transpose_y = var_7702_transpose_y_0, x = transpose_106, y = transpose_107)[name = tensor("op_7702_cast_fp16")]; + tensor var_7696_to_fp16 = const()[name = tensor("op_7696_to_fp16"), val = tensor([0x1p-3])]; + tensor attn_weights_103_cast_fp16 = mul(x = var_7702_cast_fp16, y = var_7696_to_fp16)[name = tensor("attn_weights_103_cast_fp16")]; + tensor var_7704 = logical_not(x = attn_mask_71)[name = tensor("op_7704")]; + tensor var_7705_to_fp16 = const()[name = tensor("op_7705_to_fp16"), val = tensor(-inf)]; + tensor attn_weights_105_cast_fp16 = select(a = var_7705_to_fp16, b = attn_weights_103_cast_fp16, cond = var_7704)[name = tensor("attn_weights_105_cast_fp16")]; + tensor var_7707 = const()[name = tensor("op_7707"), val = tensor(-1)]; + tensor attn_weights_107_cast_fp16 = softmax(axis = var_7707, x = attn_weights_105_cast_fp16)[name = tensor("attn_weights_107_cast_fp16")]; + tensor attn_output_35_transpose_x_0 = const()[name = tensor("attn_output_35_transpose_x_0"), val = tensor(false)]; + tensor attn_output_35_transpose_y_0 = const()[name = tensor("attn_output_35_transpose_y_0"), val = tensor(false)]; + tensor values_107_cast_fp16 = transpose(perm = var_7653, x = values_105_cast_fp16)[name = tensor("transpose_147")]; + tensor attn_output_35_cast_fp16 = matmul(transpose_x = attn_output_35_transpose_x_0, transpose_y = attn_output_35_transpose_y_0, x = attn_weights_107_cast_fp16, y = values_107_cast_fp16)[name = tensor("attn_output_35_cast_fp16")]; + tensor var_7715 = const()[name = tensor("op_7715"), val = tensor([0, 2, 1, 3])]; + tensor var_7718 = const()[name = tensor("op_7718"), val = tensor([1, 1, 1024])]; + tensor var_7716_cast_fp16 = transpose(perm = var_7715, x = attn_output_35_cast_fp16)[name = tensor("transpose_144")]; + tensor input_175_cast_fp16 = reshape(shape = var_7718, x = var_7716_cast_fp16)[name = tensor("input_175_cast_fp16")]; + tensor linear_70_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = attn17_out_proj_weight_cast_fp16, x = input_175_cast_fp16)[name = tensor("linear_70_cast_fp16")]; + tensor var_7724_to_fp16 = const()[name = tensor("op_7724_to_fp16"), val = tensor(0x1p+0)]; + tensor var_7725_cast_fp16 = add(x = position17_to_fp16, y = var_7724_to_fp16)[name = tensor("op_7725_cast_fp16")]; + tensor var_7725_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_7725_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor input_177_cast_fp16 = add(x = input_173_cast_fp16, y = linear_70_cast_fp16)[name = tensor("input_177_cast_fp16")]; + tensor input_179_axes_0 = const()[name = tensor("input_179_axes_0"), val = tensor([-1])]; + tensor norm17_2_weight_to_fp16 = const()[name = tensor("norm17_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(303735040)))]; + tensor norm17_2_bias_to_fp16 = const()[name = tensor("norm17_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(303737152)))]; + tensor var_7729_to_fp16 = const()[name = tensor("op_7729_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_179_cast_fp16 = layer_norm(axes = input_179_axes_0, beta = norm17_2_bias_to_fp16, epsilon = var_7729_to_fp16, gamma = norm17_2_weight_to_fp16, x = input_177_cast_fp16)[name = tensor("input_179_cast_fp16")]; + tensor linear_71_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = linear17_1_weight_cast_fp16, x = input_179_cast_fp16)[name = tensor("linear_71_cast_fp16")]; + tensor input_181_mode_0 = const()[name = tensor("input_181_mode_0"), val = tensor("EXACT")]; + tensor input_181_cast_fp16 = gelu(mode = input_181_mode_0, x = linear_71_cast_fp16)[name = tensor("input_181_cast_fp16")]; + tensor linear_72_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = linear17_2_weight_cast_fp16, x = input_181_cast_fp16)[name = tensor("linear_72_cast_fp16")]; + tensor input_183_cast_fp16 = add(x = input_177_cast_fp16, y = linear_72_cast_fp16)[name = tensor("input_183_cast_fp16")]; + tensor x_37_axes_0 = const()[name = tensor("x_37_axes_0"), val = tensor([-1])]; + tensor norm18_1_weight_to_fp16 = const()[name = tensor("norm18_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(303739264)))]; + tensor norm18_1_bias_to_fp16 = const()[name = tensor("norm18_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(303741376)))]; + tensor var_7746_to_fp16 = const()[name = tensor("op_7746_to_fp16"), val = tensor(0x1.5p-17)]; + tensor x_37_cast_fp16 = layer_norm(axes = x_37_axes_0, beta = norm18_1_bias_to_fp16, epsilon = var_7746_to_fp16, gamma = norm18_1_weight_to_fp16, x = input_183_cast_fp16)[name = tensor("x_37_cast_fp16")]; + tensor linear_73_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = attn18_in_proj_weight_cast_fp16, x = x_37_cast_fp16)[name = tensor("linear_73_cast_fp16")]; + tensor var_7782 = const()[name = tensor("op_7782"), val = tensor([1, 1, 3, 16, 64])]; + tensor qkv_37_cast_fp16 = reshape(shape = var_7782, x = linear_73_cast_fp16)[name = tensor("qkv_37_cast_fp16")]; + tensor q_109_begin_0 = const()[name = tensor("q_109_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor q_109_end_0 = const()[name = tensor("q_109_end_0"), val = tensor([1, 1, 1, 16, 64])]; + tensor q_109_end_mask_0 = const()[name = tensor("q_109_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor q_109_squeeze_mask_0 = const()[name = tensor("q_109_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor q_109_cast_fp16 = slice_by_index(begin = q_109_begin_0, end = q_109_end_0, end_mask = q_109_end_mask_0, squeeze_mask = q_109_squeeze_mask_0, x = qkv_37_cast_fp16)[name = tensor("q_109_cast_fp16")]; + tensor k_73_begin_0 = const()[name = tensor("k_73_begin_0"), val = tensor([0, 0, 1, 0, 0])]; + tensor k_73_end_0 = const()[name = tensor("k_73_end_0"), val = tensor([1, 1, 2, 16, 64])]; + tensor k_73_end_mask_0 = const()[name = tensor("k_73_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor k_73_squeeze_mask_0 = const()[name = tensor("k_73_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor k_73_cast_fp16 = slice_by_index(begin = k_73_begin_0, end = k_73_end_0, end_mask = k_73_end_mask_0, squeeze_mask = k_73_squeeze_mask_0, x = qkv_37_cast_fp16)[name = tensor("k_73_cast_fp16")]; + tensor v_37_begin_0 = const()[name = tensor("v_37_begin_0"), val = tensor([0, 0, 2, 0, 0])]; + tensor v_37_end_0 = const()[name = tensor("v_37_end_0"), val = tensor([1, 1, 3, 16, 64])]; + tensor v_37_end_mask_0 = const()[name = tensor("v_37_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor v_37_squeeze_mask_0 = const()[name = tensor("v_37_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor v_37_cast_fp16 = slice_by_index(begin = v_37_begin_0, end = v_37_end_0, end_mask = v_37_end_mask_0, squeeze_mask = v_37_squeeze_mask_0, x = qkv_37_cast_fp16)[name = tensor("v_37_cast_fp16")]; + tensor var_7886 = const()[name = tensor("op_7886"), val = tensor([1, 1, 1, 1])]; + tensor position18_to_fp16_dtype_0 = const()[name = tensor("position18_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor position18_to_fp16 = cast(dtype = position18_to_fp16_dtype_0, x = position18)[name = tensor("cast_463")]; + tensor ts_113_cast_fp16 = reshape(shape = var_7886, x = position18_to_fp16)[name = tensor("ts_113_cast_fp16")]; + tensor var_7890 = const()[name = tensor("op_7890"), val = tensor([1, 1, 16, 32, 2])]; + tensor q_complex_37_cast_fp16 = reshape(shape = var_7890, x = q_109_cast_fp16)[name = tensor("q_complex_37_cast_fp16")]; + tensor var_7894 = const()[name = tensor("op_7894"), val = tensor([1, 1, 16, 32, 2])]; + tensor k_complex_37_cast_fp16 = reshape(shape = var_7894, x = k_73_cast_fp16)[name = tensor("k_complex_37_cast_fp16")]; + tensor var_7898_begin_0 = const()[name = tensor("op_7898_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_7898_end_0 = const()[name = tensor("op_7898_end_0"), val = tensor([1, 1, 16, 32, 1])]; + tensor var_7898_end_mask_0 = const()[name = tensor("op_7898_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_7898_squeeze_mask_0 = const()[name = tensor("op_7898_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_7898_cast_fp16 = slice_by_index(begin = var_7898_begin_0, end = var_7898_end_0, end_mask = var_7898_end_mask_0, squeeze_mask = var_7898_squeeze_mask_0, x = q_complex_37_cast_fp16)[name = tensor("op_7898_cast_fp16")]; + tensor var_7906_begin_0 = const()[name = tensor("op_7906_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_7906_end_0 = const()[name = tensor("op_7906_end_0"), val = tensor([1, 1, 16, 32, 2])]; + tensor var_7906_end_mask_0 = const()[name = tensor("op_7906_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_7906_squeeze_mask_0 = const()[name = tensor("op_7906_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_7906_cast_fp16 = slice_by_index(begin = var_7906_begin_0, end = var_7906_end_0, end_mask = var_7906_end_mask_0, squeeze_mask = var_7906_squeeze_mask_0, x = q_complex_37_cast_fp16)[name = tensor("op_7906_cast_fp16")]; + tensor var_7914_begin_0 = const()[name = tensor("op_7914_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_7914_end_0 = const()[name = tensor("op_7914_end_0"), val = tensor([1, 1, 16, 32, 1])]; + tensor var_7914_end_mask_0 = const()[name = tensor("op_7914_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_7914_squeeze_mask_0 = const()[name = tensor("op_7914_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_7914_cast_fp16 = slice_by_index(begin = var_7914_begin_0, end = var_7914_end_0, end_mask = var_7914_end_mask_0, squeeze_mask = var_7914_squeeze_mask_0, x = k_complex_37_cast_fp16)[name = tensor("op_7914_cast_fp16")]; + tensor var_7922_begin_0 = const()[name = tensor("op_7922_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_7922_end_0 = const()[name = tensor("op_7922_end_0"), val = tensor([1, 1, 16, 32, 2])]; + tensor var_7922_end_mask_0 = const()[name = tensor("op_7922_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_7922_squeeze_mask_0 = const()[name = tensor("op_7922_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_7922_cast_fp16 = slice_by_index(begin = var_7922_begin_0, end = var_7922_end_0, end_mask = var_7922_end_mask_0, squeeze_mask = var_7922_squeeze_mask_0, x = k_complex_37_cast_fp16)[name = tensor("op_7922_cast_fp16")]; + tensor freqs_37_to_fp16 = const()[name = tensor("freqs_37_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(303743488)))]; + tensor var_7928_cast_fp16 = mul(x = freqs_37_to_fp16, y = ts_113_cast_fp16)[name = tensor("op_7928_cast_fp16")]; + tensor rotr_37_cast_fp16 = cos(x = var_7928_cast_fp16)[name = tensor("rotr_37_cast_fp16")]; + tensor roti_37_cast_fp16 = sin(x = var_7928_cast_fp16)[name = tensor("roti_37_cast_fp16")]; + tensor var_7932_cast_fp16 = mul(x = var_7898_cast_fp16, y = rotr_37_cast_fp16)[name = tensor("op_7932_cast_fp16")]; + tensor var_7933_cast_fp16 = mul(x = var_7906_cast_fp16, y = roti_37_cast_fp16)[name = tensor("op_7933_cast_fp16")]; + tensor qor_73_cast_fp16 = sub(x = var_7932_cast_fp16, y = var_7933_cast_fp16)[name = tensor("qor_73_cast_fp16")]; + tensor var_7936_cast_fp16 = mul(x = var_7898_cast_fp16, y = roti_37_cast_fp16)[name = tensor("op_7936_cast_fp16")]; + tensor var_7937_cast_fp16 = mul(x = var_7906_cast_fp16, y = rotr_37_cast_fp16)[name = tensor("op_7937_cast_fp16")]; + tensor qoi_73_cast_fp16 = add(x = var_7936_cast_fp16, y = var_7937_cast_fp16)[name = tensor("qoi_73_cast_fp16")]; + tensor var_7940_cast_fp16 = mul(x = var_7914_cast_fp16, y = rotr_37_cast_fp16)[name = tensor("op_7940_cast_fp16")]; + tensor var_7941_cast_fp16 = mul(x = var_7922_cast_fp16, y = roti_37_cast_fp16)[name = tensor("op_7941_cast_fp16")]; + tensor kor_73_cast_fp16 = sub(x = var_7940_cast_fp16, y = var_7941_cast_fp16)[name = tensor("kor_73_cast_fp16")]; + tensor var_7944_cast_fp16 = mul(x = var_7914_cast_fp16, y = roti_37_cast_fp16)[name = tensor("op_7944_cast_fp16")]; + tensor var_7945_cast_fp16 = mul(x = var_7922_cast_fp16, y = rotr_37_cast_fp16)[name = tensor("op_7945_cast_fp16")]; + tensor koi_73_cast_fp16 = add(x = var_7944_cast_fp16, y = var_7945_cast_fp16)[name = tensor("koi_73_cast_fp16")]; + tensor qo_37_axis_0 = const()[name = tensor("qo_37_axis_0"), val = tensor(-1)]; + tensor qo_37_cast_fp16 = stack(axis = qo_37_axis_0, values = (qor_73_cast_fp16, qoi_73_cast_fp16))[name = tensor("qo_37_cast_fp16")]; + tensor ko_37_axis_0 = const()[name = tensor("ko_37_axis_0"), val = tensor(-1)]; + tensor ko_37_cast_fp16 = stack(axis = ko_37_axis_0, values = (kor_73_cast_fp16, koi_73_cast_fp16))[name = tensor("ko_37_cast_fp16")]; + tensor var_7974 = const()[name = tensor("op_7974"), val = tensor([1, 1, 16, 64])]; + tensor q_111_cast_fp16 = reshape(shape = var_7974, x = qo_37_cast_fp16)[name = tensor("q_111_cast_fp16")]; + tensor var_7976 = const()[name = tensor("op_7976"), val = tensor([1, 1, 16, 64])]; + tensor k_75_cast_fp16 = reshape(shape = var_7976, x = ko_37_cast_fp16)[name = tensor("k_75_cast_fp16")]; + tensor _inversed_7998_y_0_to_fp16 = const()[name = tensor("_inversed_7998_y_0_to_fp16"), val = tensor(0x1p-9)]; + tensor _inversed_7998_cast_fp16 = mul(x = ts_113_cast_fp16, y = _inversed_7998_y_0_to_fp16)[name = tensor("_inversed_7998_cast_fp16")]; + tensor var_7999_cast_fp16 = floor(x = _inversed_7998_cast_fp16)[name = tensor("op_7999_cast_fp16")]; + tensor var_8000_to_fp16 = const()[name = tensor("op_8000_to_fp16"), val = tensor(0x1p+9)]; + tensor var_8001_cast_fp16 = mul(x = var_7999_cast_fp16, y = var_8000_to_fp16)[name = tensor("op_8001_cast_fp16")]; + tensor write_indices_float_75_cast_fp16 = sub(x = ts_113_cast_fp16, y = var_8001_cast_fp16)[name = tensor("write_indices_float_75_cast_fp16")]; + tensor var_8008_dtype_0 = const()[name = tensor("op_8008_dtype_0"), val = tensor("int32")]; + tensor write_indices_37_reps_0 = const()[name = tensor("write_indices_37_reps_0"), val = tensor([1, 1, 16, 64])]; + tensor write_indices_float_75_cast_fp16_to_int32 = cast(dtype = var_8008_dtype_0, x = write_indices_float_75_cast_fp16)[name = tensor("cast_462")]; + tensor write_indices_37 = tile(reps = write_indices_37_reps_0, x = write_indices_float_75_cast_fp16_to_int32)[name = tensor("write_indices_37")]; + tensor var_8016_begin_0 = const()[name = tensor("op_8016_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_8016_end_0 = const()[name = tensor("op_8016_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor var_8016_end_mask_0 = const()[name = tensor("op_8016_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_8016_squeeze_mask_0 = const()[name = tensor("op_8016_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor cache18_to_fp16_dtype_0 = const()[name = tensor("cache18_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor cache18_to_fp16 = cast(dtype = cache18_to_fp16_dtype_0, x = cache18)[name = tensor("cast_461")]; + tensor var_8016_cast_fp16 = slice_by_index(begin = var_8016_begin_0, end = var_8016_end_0, end_mask = var_8016_end_mask_0, squeeze_mask = var_8016_squeeze_mask_0, x = cache18_to_fp16)[name = tensor("op_8016_cast_fp16")]; + tensor var_8018_axis_0 = const()[name = tensor("op_8018_axis_0"), val = tensor(1)]; + tensor var_8018_mode_0 = const()[name = tensor("op_8018_mode_0"), val = tensor("update")]; + tensor var_8018_validate_indices_0 = const()[name = tensor("op_8018_validate_indices_0"), val = tensor(false)]; + tensor var_8018_cast_fp16 = scatter_along_axis(axis = var_8018_axis_0, data = var_8016_cast_fp16, indices = write_indices_37, mode = var_8018_mode_0, updates = k_75_cast_fp16, validate_indices = var_8018_validate_indices_0)[name = tensor("op_8018_cast_fp16")]; + tensor concat_128 = const()[name = tensor("concat_128"), val = tensor([0, 0, 0, 0, 0])]; + tensor concat_129 = const()[name = tensor("concat_129"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_37_internal_tensor_assign_1_stride_0 = const()[name = tensor("new_cache_37_internal_tensor_assign_1_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_37_internal_tensor_assign_1_begin_mask_0 = const()[name = tensor("new_cache_37_internal_tensor_assign_1_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_37_internal_tensor_assign_1_end_mask_0 = const()[name = tensor("new_cache_37_internal_tensor_assign_1_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_37_internal_tensor_assign_1_squeeze_mask_0 = const()[name = tensor("new_cache_37_internal_tensor_assign_1_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_84 = const()[name = tensor("shape_84"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_36 = const()[name = tensor("reduce_prod_36"), val = tensor(1048576)]; + tensor range_1d_36_start_0 = const()[name = tensor("range_1d_36_start_0"), val = tensor(0)]; + tensor range_1d_36_step_0 = const()[name = tensor("range_1d_36_step_0"), val = tensor(1)]; + tensor range_1d_36 = range_1d(end = reduce_prod_36, start = range_1d_36_start_0, step = range_1d_36_step_0)[name = tensor("range_1d_36")]; + tensor reshape_180 = reshape(shape = shape_84, x = range_1d_36)[name = tensor("reshape_180")]; + tensor slice_by_index_36 = slice_by_index(begin = concat_128, begin_mask = new_cache_37_internal_tensor_assign_1_begin_mask_0, end = concat_129, end_mask = new_cache_37_internal_tensor_assign_1_end_mask_0, squeeze_mask = new_cache_37_internal_tensor_assign_1_squeeze_mask_0, stride = new_cache_37_internal_tensor_assign_1_stride_0, x = reshape_180)[name = tensor("slice_by_index_36")]; + tensor reshape_181_shape_0 = const()[name = tensor("reshape_181_shape_0"), val = tensor([-1])]; + tensor reshape_181 = reshape(shape = reshape_181_shape_0, x = slice_by_index_36)[name = tensor("reshape_181")]; + tensor reshape_182_shape_0 = const()[name = tensor("reshape_182_shape_0"), val = tensor([-1])]; + tensor reshape_182_cast_fp16 = reshape(shape = reshape_182_shape_0, x = var_8018_cast_fp16)[name = tensor("reshape_182_cast_fp16")]; + tensor reshape_183_shape_0 = const()[name = tensor("reshape_183_shape_0"), val = tensor([-1])]; + tensor reshape_183_cast_fp16 = reshape(shape = reshape_183_shape_0, x = cache18_to_fp16)[name = tensor("reshape_183_cast_fp16")]; + tensor scatter_36_mode_0 = const()[name = tensor("scatter_36_mode_0"), val = tensor("update")]; + tensor scatter_36_axis_0 = const()[name = tensor("scatter_36_axis_0"), val = tensor(0)]; + tensor scatter_36_validate_indices_0 = const()[name = tensor("scatter_36_validate_indices_0"), val = tensor(false)]; + tensor scatter_36_cast_fp16 = scatter(axis = scatter_36_axis_0, data = reshape_183_cast_fp16, indices = reshape_181, mode = scatter_36_mode_0, updates = reshape_182_cast_fp16, validate_indices = scatter_36_validate_indices_0)[name = tensor("scatter_36_cast_fp16")]; + tensor reshape_184_cast_fp16 = reshape(shape = shape_84, x = scatter_36_cast_fp16)[name = tensor("reshape_184_cast_fp16")]; + tensor var_8026_begin_0 = const()[name = tensor("op_8026_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor var_8026_end_0 = const()[name = tensor("op_8026_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor var_8026_end_mask_0 = const()[name = tensor("op_8026_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_8026_squeeze_mask_0 = const()[name = tensor("op_8026_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor var_8026_cast_fp16 = slice_by_index(begin = var_8026_begin_0, end = var_8026_end_0, end_mask = var_8026_end_mask_0, squeeze_mask = var_8026_squeeze_mask_0, x = reshape_184_cast_fp16)[name = tensor("op_8026_cast_fp16")]; + tensor var_8028_axis_0 = const()[name = tensor("op_8028_axis_0"), val = tensor(1)]; + tensor var_8028_mode_0 = const()[name = tensor("op_8028_mode_0"), val = tensor("update")]; + tensor var_8028_validate_indices_0 = const()[name = tensor("op_8028_validate_indices_0"), val = tensor(false)]; + tensor var_8028_cast_fp16 = scatter_along_axis(axis = var_8028_axis_0, data = var_8026_cast_fp16, indices = write_indices_37, mode = var_8028_mode_0, updates = v_37_cast_fp16, validate_indices = var_8028_validate_indices_0)[name = tensor("op_8028_cast_fp16")]; + tensor concat_130 = const()[name = tensor("concat_130"), val = tensor([1, 0, 0, 0, 0])]; + tensor concat_131 = const()[name = tensor("concat_131"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_37_internal_tensor_assign_2_stride_0 = const()[name = tensor("new_cache_37_internal_tensor_assign_2_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_37_internal_tensor_assign_2_begin_mask_0 = const()[name = tensor("new_cache_37_internal_tensor_assign_2_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_37_internal_tensor_assign_2_end_mask_0 = const()[name = tensor("new_cache_37_internal_tensor_assign_2_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_37_internal_tensor_assign_2_squeeze_mask_0 = const()[name = tensor("new_cache_37_internal_tensor_assign_2_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_85 = const()[name = tensor("shape_85"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_37 = const()[name = tensor("reduce_prod_37"), val = tensor(1048576)]; + tensor range_1d_37_start_0 = const()[name = tensor("range_1d_37_start_0"), val = tensor(0)]; + tensor range_1d_37_step_0 = const()[name = tensor("range_1d_37_step_0"), val = tensor(1)]; + tensor range_1d_37 = range_1d(end = reduce_prod_37, start = range_1d_37_start_0, step = range_1d_37_step_0)[name = tensor("range_1d_37")]; + tensor reshape_185 = reshape(shape = shape_85, x = range_1d_37)[name = tensor("reshape_185")]; + tensor slice_by_index_37 = slice_by_index(begin = concat_130, begin_mask = new_cache_37_internal_tensor_assign_2_begin_mask_0, end = concat_131, end_mask = new_cache_37_internal_tensor_assign_2_end_mask_0, squeeze_mask = new_cache_37_internal_tensor_assign_2_squeeze_mask_0, stride = new_cache_37_internal_tensor_assign_2_stride_0, x = reshape_185)[name = tensor("slice_by_index_37")]; + tensor reshape_186_shape_0 = const()[name = tensor("reshape_186_shape_0"), val = tensor([-1])]; + tensor reshape_186 = reshape(shape = reshape_186_shape_0, x = slice_by_index_37)[name = tensor("reshape_186")]; + tensor reshape_187_shape_0 = const()[name = tensor("reshape_187_shape_0"), val = tensor([-1])]; + tensor reshape_187_cast_fp16 = reshape(shape = reshape_187_shape_0, x = var_8028_cast_fp16)[name = tensor("reshape_187_cast_fp16")]; + tensor reshape_188_shape_0 = const()[name = tensor("reshape_188_shape_0"), val = tensor([-1])]; + tensor reshape_188_cast_fp16 = reshape(shape = reshape_188_shape_0, x = reshape_184_cast_fp16)[name = tensor("reshape_188_cast_fp16")]; + tensor scatter_37_mode_0 = const()[name = tensor("scatter_37_mode_0"), val = tensor("update")]; + tensor scatter_37_axis_0 = const()[name = tensor("scatter_37_axis_0"), val = tensor(0)]; + tensor scatter_37_validate_indices_0 = const()[name = tensor("scatter_37_validate_indices_0"), val = tensor(false)]; + tensor scatter_37_cast_fp16 = scatter(axis = scatter_37_axis_0, data = reshape_188_cast_fp16, indices = reshape_186, mode = scatter_37_mode_0, updates = reshape_187_cast_fp16, validate_indices = scatter_37_validate_indices_0)[name = tensor("scatter_37_cast_fp16")]; + tensor reshape_189_cast_fp16 = reshape(shape = shape_85, x = scatter_37_cast_fp16)[name = tensor("reshape_189_cast_fp16")]; + tensor reshape_189_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("reshape_189_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor keys_109_begin_0 = const()[name = tensor("keys_109_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor keys_109_end_0 = const()[name = tensor("keys_109_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor keys_109_end_mask_0 = const()[name = tensor("keys_109_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor keys_109_squeeze_mask_0 = const()[name = tensor("keys_109_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor keys_109_cast_fp16 = slice_by_index(begin = keys_109_begin_0, end = keys_109_end_0, end_mask = keys_109_end_mask_0, squeeze_mask = keys_109_squeeze_mask_0, x = reshape_189_cast_fp16)[name = tensor("keys_109_cast_fp16")]; + tensor values_109_begin_0 = const()[name = tensor("values_109_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor values_109_end_0 = const()[name = tensor("values_109_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor values_109_end_mask_0 = const()[name = tensor("values_109_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor values_109_squeeze_mask_0 = const()[name = tensor("values_109_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor values_109_cast_fp16 = slice_by_index(begin = values_109_begin_0, end = values_109_end_0, end_mask = values_109_end_mask_0, squeeze_mask = values_109_squeeze_mask_0, x = reshape_189_cast_fp16)[name = tensor("values_109_cast_fp16")]; + tensor var_8040_cast_fp16 = not_equal(x = keys_109_cast_fp16, y = keys_109_cast_fp16)[name = tensor("op_8040_cast_fp16")]; + tensor keys_111_cast_fp16 = select(a = var_504_to_fp16, b = keys_109_cast_fp16, cond = var_8040_cast_fp16)[name = tensor("keys_111_cast_fp16")]; + tensor var_8048_cast_fp16 = not_equal(x = values_109_cast_fp16, y = values_109_cast_fp16)[name = tensor("op_8048_cast_fp16")]; + tensor values_111_cast_fp16 = select(a = var_504_to_fp16, b = values_109_cast_fp16, cond = var_8048_cast_fp16)[name = tensor("values_111_cast_fp16")]; + tensor var_8072 = const()[name = tensor("op_8072"), val = tensor([0, 2, 1, 3])]; + tensor var_8085 = const()[name = tensor("op_8085"), val = tensor([1, 1, 1])]; + tensor var_8086_cast_fp16 = reshape(shape = var_8085, x = position18_to_fp16)[name = tensor("op_8086_cast_fp16")]; + tensor var_8103_to_fp16 = const()[name = tensor("op_8103_to_fp16"), val = tensor(0x1p+0)]; + tensor valid_len_37_cast_fp16 = add(x = var_8086_cast_fp16, y = var_8103_to_fp16)[name = tensor("valid_len_37_cast_fp16")]; + tensor valid_mask_37_cast_fp16 = less(x = k_positions_1_promoted_to_fp16, y = valid_len_37_cast_fp16)[name = tensor("valid_mask_37_cast_fp16")]; + tensor causal_mask_37_cast_fp16 = less_equal(x = k_positions_1_promoted_to_fp16, y = var_8086_cast_fp16)[name = tensor("causal_mask_37_cast_fp16")]; + tensor attn_mask_73 = logical_and(x = valid_mask_37_cast_fp16, y = causal_mask_37_cast_fp16)[name = tensor("attn_mask_73")]; + tensor attn_mask_75_axes_0 = const()[name = tensor("attn_mask_75_axes_0"), val = tensor([1])]; + tensor attn_mask_75 = expand_dims(axes = attn_mask_75_axes_0, x = attn_mask_73)[name = tensor("attn_mask_75")]; + tensor var_8121_transpose_x_0 = const()[name = tensor("op_8121_transpose_x_0"), val = tensor(false)]; + tensor var_8121_transpose_y_0 = const()[name = tensor("op_8121_transpose_y_0"), val = tensor(false)]; + tensor transpose_108_perm_0 = const()[name = tensor("transpose_108_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_109_perm_0 = const()[name = tensor("transpose_109_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_109 = transpose(perm = transpose_109_perm_0, x = keys_111_cast_fp16)[name = tensor("transpose_141")]; + tensor transpose_108 = transpose(perm = transpose_108_perm_0, x = q_111_cast_fp16)[name = tensor("transpose_142")]; + tensor var_8121_cast_fp16 = matmul(transpose_x = var_8121_transpose_x_0, transpose_y = var_8121_transpose_y_0, x = transpose_108, y = transpose_109)[name = tensor("op_8121_cast_fp16")]; + tensor var_8115_to_fp16 = const()[name = tensor("op_8115_to_fp16"), val = tensor([0x1p-3])]; + tensor attn_weights_109_cast_fp16 = mul(x = var_8121_cast_fp16, y = var_8115_to_fp16)[name = tensor("attn_weights_109_cast_fp16")]; + tensor var_8123 = logical_not(x = attn_mask_75)[name = tensor("op_8123")]; + tensor var_8124_to_fp16 = const()[name = tensor("op_8124_to_fp16"), val = tensor(-inf)]; + tensor attn_weights_111_cast_fp16 = select(a = var_8124_to_fp16, b = attn_weights_109_cast_fp16, cond = var_8123)[name = tensor("attn_weights_111_cast_fp16")]; + tensor var_8126 = const()[name = tensor("op_8126"), val = tensor(-1)]; + tensor attn_weights_113_cast_fp16 = softmax(axis = var_8126, x = attn_weights_111_cast_fp16)[name = tensor("attn_weights_113_cast_fp16")]; + tensor attn_output_37_transpose_x_0 = const()[name = tensor("attn_output_37_transpose_x_0"), val = tensor(false)]; + tensor attn_output_37_transpose_y_0 = const()[name = tensor("attn_output_37_transpose_y_0"), val = tensor(false)]; + tensor values_113_cast_fp16 = transpose(perm = var_8072, x = values_111_cast_fp16)[name = tensor("transpose_143")]; + tensor attn_output_37_cast_fp16 = matmul(transpose_x = attn_output_37_transpose_x_0, transpose_y = attn_output_37_transpose_y_0, x = attn_weights_113_cast_fp16, y = values_113_cast_fp16)[name = tensor("attn_output_37_cast_fp16")]; + tensor var_8134 = const()[name = tensor("op_8134"), val = tensor([0, 2, 1, 3])]; + tensor var_8137 = const()[name = tensor("op_8137"), val = tensor([1, 1, 1024])]; + tensor var_8135_cast_fp16 = transpose(perm = var_8134, x = attn_output_37_cast_fp16)[name = tensor("transpose_140")]; + tensor input_185_cast_fp16 = reshape(shape = var_8137, x = var_8135_cast_fp16)[name = tensor("input_185_cast_fp16")]; + tensor linear_74_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = attn18_out_proj_weight_cast_fp16, x = input_185_cast_fp16)[name = tensor("linear_74_cast_fp16")]; + tensor var_8143_to_fp16 = const()[name = tensor("op_8143_to_fp16"), val = tensor(0x1p+0)]; + tensor var_8144_cast_fp16 = add(x = position18_to_fp16, y = var_8143_to_fp16)[name = tensor("op_8144_cast_fp16")]; + tensor var_8144_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_8144_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor input_187_cast_fp16 = add(x = input_183_cast_fp16, y = linear_74_cast_fp16)[name = tensor("input_187_cast_fp16")]; + tensor input_189_axes_0 = const()[name = tensor("input_189_axes_0"), val = tensor([-1])]; + tensor norm18_2_weight_to_fp16 = const()[name = tensor("norm18_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(303743616)))]; + tensor norm18_2_bias_to_fp16 = const()[name = tensor("norm18_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(303745728)))]; + tensor var_8148_to_fp16 = const()[name = tensor("op_8148_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_189_cast_fp16 = layer_norm(axes = input_189_axes_0, beta = norm18_2_bias_to_fp16, epsilon = var_8148_to_fp16, gamma = norm18_2_weight_to_fp16, x = input_187_cast_fp16)[name = tensor("input_189_cast_fp16")]; + tensor linear_75_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = linear18_1_weight_cast_fp16, x = input_189_cast_fp16)[name = tensor("linear_75_cast_fp16")]; + tensor input_191_mode_0 = const()[name = tensor("input_191_mode_0"), val = tensor("EXACT")]; + tensor input_191_cast_fp16 = gelu(mode = input_191_mode_0, x = linear_75_cast_fp16)[name = tensor("input_191_cast_fp16")]; + tensor linear_76_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = linear18_2_weight_cast_fp16, x = input_191_cast_fp16)[name = tensor("linear_76_cast_fp16")]; + tensor input_193_cast_fp16 = add(x = input_187_cast_fp16, y = linear_76_cast_fp16)[name = tensor("input_193_cast_fp16")]; + tensor x_39_axes_0 = const()[name = tensor("x_39_axes_0"), val = tensor([-1])]; + tensor norm19_1_weight_to_fp16 = const()[name = tensor("norm19_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(303747840)))]; + tensor norm19_1_bias_to_fp16 = const()[name = tensor("norm19_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(303749952)))]; + tensor var_8165_to_fp16 = const()[name = tensor("op_8165_to_fp16"), val = tensor(0x1.5p-17)]; + tensor x_39_cast_fp16 = layer_norm(axes = x_39_axes_0, beta = norm19_1_bias_to_fp16, epsilon = var_8165_to_fp16, gamma = norm19_1_weight_to_fp16, x = input_193_cast_fp16)[name = tensor("x_39_cast_fp16")]; + tensor linear_77_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = attn19_in_proj_weight_cast_fp16, x = x_39_cast_fp16)[name = tensor("linear_77_cast_fp16")]; + tensor var_8201 = const()[name = tensor("op_8201"), val = tensor([1, 1, 3, 16, 64])]; + tensor qkv_39_cast_fp16 = reshape(shape = var_8201, x = linear_77_cast_fp16)[name = tensor("qkv_39_cast_fp16")]; + tensor q_115_begin_0 = const()[name = tensor("q_115_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor q_115_end_0 = const()[name = tensor("q_115_end_0"), val = tensor([1, 1, 1, 16, 64])]; + tensor q_115_end_mask_0 = const()[name = tensor("q_115_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor q_115_squeeze_mask_0 = const()[name = tensor("q_115_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor q_115_cast_fp16 = slice_by_index(begin = q_115_begin_0, end = q_115_end_0, end_mask = q_115_end_mask_0, squeeze_mask = q_115_squeeze_mask_0, x = qkv_39_cast_fp16)[name = tensor("q_115_cast_fp16")]; + tensor k_77_begin_0 = const()[name = tensor("k_77_begin_0"), val = tensor([0, 0, 1, 0, 0])]; + tensor k_77_end_0 = const()[name = tensor("k_77_end_0"), val = tensor([1, 1, 2, 16, 64])]; + tensor k_77_end_mask_0 = const()[name = tensor("k_77_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor k_77_squeeze_mask_0 = const()[name = tensor("k_77_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor k_77_cast_fp16 = slice_by_index(begin = k_77_begin_0, end = k_77_end_0, end_mask = k_77_end_mask_0, squeeze_mask = k_77_squeeze_mask_0, x = qkv_39_cast_fp16)[name = tensor("k_77_cast_fp16")]; + tensor v_39_begin_0 = const()[name = tensor("v_39_begin_0"), val = tensor([0, 0, 2, 0, 0])]; + tensor v_39_end_0 = const()[name = tensor("v_39_end_0"), val = tensor([1, 1, 3, 16, 64])]; + tensor v_39_end_mask_0 = const()[name = tensor("v_39_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor v_39_squeeze_mask_0 = const()[name = tensor("v_39_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor v_39_cast_fp16 = slice_by_index(begin = v_39_begin_0, end = v_39_end_0, end_mask = v_39_end_mask_0, squeeze_mask = v_39_squeeze_mask_0, x = qkv_39_cast_fp16)[name = tensor("v_39_cast_fp16")]; + tensor var_8305 = const()[name = tensor("op_8305"), val = tensor([1, 1, 1, 1])]; + tensor position19_to_fp16_dtype_0 = const()[name = tensor("position19_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor position19_to_fp16 = cast(dtype = position19_to_fp16_dtype_0, x = position19)[name = tensor("cast_458")]; + tensor ts_119_cast_fp16 = reshape(shape = var_8305, x = position19_to_fp16)[name = tensor("ts_119_cast_fp16")]; + tensor var_8309 = const()[name = tensor("op_8309"), val = tensor([1, 1, 16, 32, 2])]; + tensor q_complex_39_cast_fp16 = reshape(shape = var_8309, x = q_115_cast_fp16)[name = tensor("q_complex_39_cast_fp16")]; + tensor var_8313 = const()[name = tensor("op_8313"), val = tensor([1, 1, 16, 32, 2])]; + tensor k_complex_39_cast_fp16 = reshape(shape = var_8313, x = k_77_cast_fp16)[name = tensor("k_complex_39_cast_fp16")]; + tensor var_8317_begin_0 = const()[name = tensor("op_8317_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_8317_end_0 = const()[name = tensor("op_8317_end_0"), val = tensor([1, 1, 16, 32, 1])]; + tensor var_8317_end_mask_0 = const()[name = tensor("op_8317_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_8317_squeeze_mask_0 = const()[name = tensor("op_8317_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_8317_cast_fp16 = slice_by_index(begin = var_8317_begin_0, end = var_8317_end_0, end_mask = var_8317_end_mask_0, squeeze_mask = var_8317_squeeze_mask_0, x = q_complex_39_cast_fp16)[name = tensor("op_8317_cast_fp16")]; + tensor var_8325_begin_0 = const()[name = tensor("op_8325_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_8325_end_0 = const()[name = tensor("op_8325_end_0"), val = tensor([1, 1, 16, 32, 2])]; + tensor var_8325_end_mask_0 = const()[name = tensor("op_8325_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_8325_squeeze_mask_0 = const()[name = tensor("op_8325_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_8325_cast_fp16 = slice_by_index(begin = var_8325_begin_0, end = var_8325_end_0, end_mask = var_8325_end_mask_0, squeeze_mask = var_8325_squeeze_mask_0, x = q_complex_39_cast_fp16)[name = tensor("op_8325_cast_fp16")]; + tensor var_8333_begin_0 = const()[name = tensor("op_8333_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_8333_end_0 = const()[name = tensor("op_8333_end_0"), val = tensor([1, 1, 16, 32, 1])]; + tensor var_8333_end_mask_0 = const()[name = tensor("op_8333_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_8333_squeeze_mask_0 = const()[name = tensor("op_8333_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_8333_cast_fp16 = slice_by_index(begin = var_8333_begin_0, end = var_8333_end_0, end_mask = var_8333_end_mask_0, squeeze_mask = var_8333_squeeze_mask_0, x = k_complex_39_cast_fp16)[name = tensor("op_8333_cast_fp16")]; + tensor var_8341_begin_0 = const()[name = tensor("op_8341_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_8341_end_0 = const()[name = tensor("op_8341_end_0"), val = tensor([1, 1, 16, 32, 2])]; + tensor var_8341_end_mask_0 = const()[name = tensor("op_8341_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_8341_squeeze_mask_0 = const()[name = tensor("op_8341_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_8341_cast_fp16 = slice_by_index(begin = var_8341_begin_0, end = var_8341_end_0, end_mask = var_8341_end_mask_0, squeeze_mask = var_8341_squeeze_mask_0, x = k_complex_39_cast_fp16)[name = tensor("op_8341_cast_fp16")]; + tensor freqs_39_to_fp16 = const()[name = tensor("freqs_39_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(303752064)))]; + tensor var_8347_cast_fp16 = mul(x = freqs_39_to_fp16, y = ts_119_cast_fp16)[name = tensor("op_8347_cast_fp16")]; + tensor rotr_39_cast_fp16 = cos(x = var_8347_cast_fp16)[name = tensor("rotr_39_cast_fp16")]; + tensor roti_39_cast_fp16 = sin(x = var_8347_cast_fp16)[name = tensor("roti_39_cast_fp16")]; + tensor var_8351_cast_fp16 = mul(x = var_8317_cast_fp16, y = rotr_39_cast_fp16)[name = tensor("op_8351_cast_fp16")]; + tensor var_8352_cast_fp16 = mul(x = var_8325_cast_fp16, y = roti_39_cast_fp16)[name = tensor("op_8352_cast_fp16")]; + tensor qor_77_cast_fp16 = sub(x = var_8351_cast_fp16, y = var_8352_cast_fp16)[name = tensor("qor_77_cast_fp16")]; + tensor var_8355_cast_fp16 = mul(x = var_8317_cast_fp16, y = roti_39_cast_fp16)[name = tensor("op_8355_cast_fp16")]; + tensor var_8356_cast_fp16 = mul(x = var_8325_cast_fp16, y = rotr_39_cast_fp16)[name = tensor("op_8356_cast_fp16")]; + tensor qoi_77_cast_fp16 = add(x = var_8355_cast_fp16, y = var_8356_cast_fp16)[name = tensor("qoi_77_cast_fp16")]; + tensor var_8359_cast_fp16 = mul(x = var_8333_cast_fp16, y = rotr_39_cast_fp16)[name = tensor("op_8359_cast_fp16")]; + tensor var_8360_cast_fp16 = mul(x = var_8341_cast_fp16, y = roti_39_cast_fp16)[name = tensor("op_8360_cast_fp16")]; + tensor kor_77_cast_fp16 = sub(x = var_8359_cast_fp16, y = var_8360_cast_fp16)[name = tensor("kor_77_cast_fp16")]; + tensor var_8363_cast_fp16 = mul(x = var_8333_cast_fp16, y = roti_39_cast_fp16)[name = tensor("op_8363_cast_fp16")]; + tensor var_8364_cast_fp16 = mul(x = var_8341_cast_fp16, y = rotr_39_cast_fp16)[name = tensor("op_8364_cast_fp16")]; + tensor koi_77_cast_fp16 = add(x = var_8363_cast_fp16, y = var_8364_cast_fp16)[name = tensor("koi_77_cast_fp16")]; + tensor qo_39_axis_0 = const()[name = tensor("qo_39_axis_0"), val = tensor(-1)]; + tensor qo_39_cast_fp16 = stack(axis = qo_39_axis_0, values = (qor_77_cast_fp16, qoi_77_cast_fp16))[name = tensor("qo_39_cast_fp16")]; + tensor ko_39_axis_0 = const()[name = tensor("ko_39_axis_0"), val = tensor(-1)]; + tensor ko_39_cast_fp16 = stack(axis = ko_39_axis_0, values = (kor_77_cast_fp16, koi_77_cast_fp16))[name = tensor("ko_39_cast_fp16")]; + tensor var_8393 = const()[name = tensor("op_8393"), val = tensor([1, 1, 16, 64])]; + tensor q_117_cast_fp16 = reshape(shape = var_8393, x = qo_39_cast_fp16)[name = tensor("q_117_cast_fp16")]; + tensor var_8395 = const()[name = tensor("op_8395"), val = tensor([1, 1, 16, 64])]; + tensor k_79_cast_fp16 = reshape(shape = var_8395, x = ko_39_cast_fp16)[name = tensor("k_79_cast_fp16")]; + tensor _inversed_8417_y_0_to_fp16 = const()[name = tensor("_inversed_8417_y_0_to_fp16"), val = tensor(0x1p-9)]; + tensor _inversed_8417_cast_fp16 = mul(x = ts_119_cast_fp16, y = _inversed_8417_y_0_to_fp16)[name = tensor("_inversed_8417_cast_fp16")]; + tensor var_8418_cast_fp16 = floor(x = _inversed_8417_cast_fp16)[name = tensor("op_8418_cast_fp16")]; + tensor var_8419_to_fp16 = const()[name = tensor("op_8419_to_fp16"), val = tensor(0x1p+9)]; + tensor var_8420_cast_fp16 = mul(x = var_8418_cast_fp16, y = var_8419_to_fp16)[name = tensor("op_8420_cast_fp16")]; + tensor write_indices_float_79_cast_fp16 = sub(x = ts_119_cast_fp16, y = var_8420_cast_fp16)[name = tensor("write_indices_float_79_cast_fp16")]; + tensor var_8427_dtype_0 = const()[name = tensor("op_8427_dtype_0"), val = tensor("int32")]; + tensor write_indices_39_reps_0 = const()[name = tensor("write_indices_39_reps_0"), val = tensor([1, 1, 16, 64])]; + tensor write_indices_float_79_cast_fp16_to_int32 = cast(dtype = var_8427_dtype_0, x = write_indices_float_79_cast_fp16)[name = tensor("cast_457")]; + tensor write_indices_39 = tile(reps = write_indices_39_reps_0, x = write_indices_float_79_cast_fp16_to_int32)[name = tensor("write_indices_39")]; + tensor var_8435_begin_0 = const()[name = tensor("op_8435_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_8435_end_0 = const()[name = tensor("op_8435_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor var_8435_end_mask_0 = const()[name = tensor("op_8435_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_8435_squeeze_mask_0 = const()[name = tensor("op_8435_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor cache19_to_fp16_dtype_0 = const()[name = tensor("cache19_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor cache19_to_fp16 = cast(dtype = cache19_to_fp16_dtype_0, x = cache19)[name = tensor("cast_456")]; + tensor var_8435_cast_fp16 = slice_by_index(begin = var_8435_begin_0, end = var_8435_end_0, end_mask = var_8435_end_mask_0, squeeze_mask = var_8435_squeeze_mask_0, x = cache19_to_fp16)[name = tensor("op_8435_cast_fp16")]; + tensor var_8437_axis_0 = const()[name = tensor("op_8437_axis_0"), val = tensor(1)]; + tensor var_8437_mode_0 = const()[name = tensor("op_8437_mode_0"), val = tensor("update")]; + tensor var_8437_validate_indices_0 = const()[name = tensor("op_8437_validate_indices_0"), val = tensor(false)]; + tensor var_8437_cast_fp16 = scatter_along_axis(axis = var_8437_axis_0, data = var_8435_cast_fp16, indices = write_indices_39, mode = var_8437_mode_0, updates = k_79_cast_fp16, validate_indices = var_8437_validate_indices_0)[name = tensor("op_8437_cast_fp16")]; + tensor concat_135 = const()[name = tensor("concat_135"), val = tensor([0, 0, 0, 0, 0])]; + tensor concat_136 = const()[name = tensor("concat_136"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_39_internal_tensor_assign_1_stride_0 = const()[name = tensor("new_cache_39_internal_tensor_assign_1_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_39_internal_tensor_assign_1_begin_mask_0 = const()[name = tensor("new_cache_39_internal_tensor_assign_1_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_39_internal_tensor_assign_1_end_mask_0 = const()[name = tensor("new_cache_39_internal_tensor_assign_1_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_39_internal_tensor_assign_1_squeeze_mask_0 = const()[name = tensor("new_cache_39_internal_tensor_assign_1_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_86 = const()[name = tensor("shape_86"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_38 = const()[name = tensor("reduce_prod_38"), val = tensor(1048576)]; + tensor range_1d_38_start_0 = const()[name = tensor("range_1d_38_start_0"), val = tensor(0)]; + tensor range_1d_38_step_0 = const()[name = tensor("range_1d_38_step_0"), val = tensor(1)]; + tensor range_1d_38 = range_1d(end = reduce_prod_38, start = range_1d_38_start_0, step = range_1d_38_step_0)[name = tensor("range_1d_38")]; + tensor reshape_190 = reshape(shape = shape_86, x = range_1d_38)[name = tensor("reshape_190")]; + tensor slice_by_index_38 = slice_by_index(begin = concat_135, begin_mask = new_cache_39_internal_tensor_assign_1_begin_mask_0, end = concat_136, end_mask = new_cache_39_internal_tensor_assign_1_end_mask_0, squeeze_mask = new_cache_39_internal_tensor_assign_1_squeeze_mask_0, stride = new_cache_39_internal_tensor_assign_1_stride_0, x = reshape_190)[name = tensor("slice_by_index_38")]; + tensor reshape_191_shape_0 = const()[name = tensor("reshape_191_shape_0"), val = tensor([-1])]; + tensor reshape_191 = reshape(shape = reshape_191_shape_0, x = slice_by_index_38)[name = tensor("reshape_191")]; + tensor reshape_192_shape_0 = const()[name = tensor("reshape_192_shape_0"), val = tensor([-1])]; + tensor reshape_192_cast_fp16 = reshape(shape = reshape_192_shape_0, x = var_8437_cast_fp16)[name = tensor("reshape_192_cast_fp16")]; + tensor reshape_193_shape_0 = const()[name = tensor("reshape_193_shape_0"), val = tensor([-1])]; + tensor reshape_193_cast_fp16 = reshape(shape = reshape_193_shape_0, x = cache19_to_fp16)[name = tensor("reshape_193_cast_fp16")]; + tensor scatter_38_mode_0 = const()[name = tensor("scatter_38_mode_0"), val = tensor("update")]; + tensor scatter_38_axis_0 = const()[name = tensor("scatter_38_axis_0"), val = tensor(0)]; + tensor scatter_38_validate_indices_0 = const()[name = tensor("scatter_38_validate_indices_0"), val = tensor(false)]; + tensor scatter_38_cast_fp16 = scatter(axis = scatter_38_axis_0, data = reshape_193_cast_fp16, indices = reshape_191, mode = scatter_38_mode_0, updates = reshape_192_cast_fp16, validate_indices = scatter_38_validate_indices_0)[name = tensor("scatter_38_cast_fp16")]; + tensor reshape_194_cast_fp16 = reshape(shape = shape_86, x = scatter_38_cast_fp16)[name = tensor("reshape_194_cast_fp16")]; + tensor var_8445_begin_0 = const()[name = tensor("op_8445_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor var_8445_end_0 = const()[name = tensor("op_8445_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor var_8445_end_mask_0 = const()[name = tensor("op_8445_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_8445_squeeze_mask_0 = const()[name = tensor("op_8445_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor var_8445_cast_fp16 = slice_by_index(begin = var_8445_begin_0, end = var_8445_end_0, end_mask = var_8445_end_mask_0, squeeze_mask = var_8445_squeeze_mask_0, x = reshape_194_cast_fp16)[name = tensor("op_8445_cast_fp16")]; + tensor var_8447_axis_0 = const()[name = tensor("op_8447_axis_0"), val = tensor(1)]; + tensor var_8447_mode_0 = const()[name = tensor("op_8447_mode_0"), val = tensor("update")]; + tensor var_8447_validate_indices_0 = const()[name = tensor("op_8447_validate_indices_0"), val = tensor(false)]; + tensor var_8447_cast_fp16 = scatter_along_axis(axis = var_8447_axis_0, data = var_8445_cast_fp16, indices = write_indices_39, mode = var_8447_mode_0, updates = v_39_cast_fp16, validate_indices = var_8447_validate_indices_0)[name = tensor("op_8447_cast_fp16")]; + tensor concat_137 = const()[name = tensor("concat_137"), val = tensor([1, 0, 0, 0, 0])]; + tensor concat_138 = const()[name = tensor("concat_138"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_39_internal_tensor_assign_2_stride_0 = const()[name = tensor("new_cache_39_internal_tensor_assign_2_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_39_internal_tensor_assign_2_begin_mask_0 = const()[name = tensor("new_cache_39_internal_tensor_assign_2_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_39_internal_tensor_assign_2_end_mask_0 = const()[name = tensor("new_cache_39_internal_tensor_assign_2_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_39_internal_tensor_assign_2_squeeze_mask_0 = const()[name = tensor("new_cache_39_internal_tensor_assign_2_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_87 = const()[name = tensor("shape_87"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_39 = const()[name = tensor("reduce_prod_39"), val = tensor(1048576)]; + tensor range_1d_39_start_0 = const()[name = tensor("range_1d_39_start_0"), val = tensor(0)]; + tensor range_1d_39_step_0 = const()[name = tensor("range_1d_39_step_0"), val = tensor(1)]; + tensor range_1d_39 = range_1d(end = reduce_prod_39, start = range_1d_39_start_0, step = range_1d_39_step_0)[name = tensor("range_1d_39")]; + tensor reshape_195 = reshape(shape = shape_87, x = range_1d_39)[name = tensor("reshape_195")]; + tensor slice_by_index_39 = slice_by_index(begin = concat_137, begin_mask = new_cache_39_internal_tensor_assign_2_begin_mask_0, end = concat_138, end_mask = new_cache_39_internal_tensor_assign_2_end_mask_0, squeeze_mask = new_cache_39_internal_tensor_assign_2_squeeze_mask_0, stride = new_cache_39_internal_tensor_assign_2_stride_0, x = reshape_195)[name = tensor("slice_by_index_39")]; + tensor reshape_196_shape_0 = const()[name = tensor("reshape_196_shape_0"), val = tensor([-1])]; + tensor reshape_196 = reshape(shape = reshape_196_shape_0, x = slice_by_index_39)[name = tensor("reshape_196")]; + tensor reshape_197_shape_0 = const()[name = tensor("reshape_197_shape_0"), val = tensor([-1])]; + tensor reshape_197_cast_fp16 = reshape(shape = reshape_197_shape_0, x = var_8447_cast_fp16)[name = tensor("reshape_197_cast_fp16")]; + tensor reshape_198_shape_0 = const()[name = tensor("reshape_198_shape_0"), val = tensor([-1])]; + tensor reshape_198_cast_fp16 = reshape(shape = reshape_198_shape_0, x = reshape_194_cast_fp16)[name = tensor("reshape_198_cast_fp16")]; + tensor scatter_39_mode_0 = const()[name = tensor("scatter_39_mode_0"), val = tensor("update")]; + tensor scatter_39_axis_0 = const()[name = tensor("scatter_39_axis_0"), val = tensor(0)]; + tensor scatter_39_validate_indices_0 = const()[name = tensor("scatter_39_validate_indices_0"), val = tensor(false)]; + tensor scatter_39_cast_fp16 = scatter(axis = scatter_39_axis_0, data = reshape_198_cast_fp16, indices = reshape_196, mode = scatter_39_mode_0, updates = reshape_197_cast_fp16, validate_indices = scatter_39_validate_indices_0)[name = tensor("scatter_39_cast_fp16")]; + tensor reshape_199_cast_fp16 = reshape(shape = shape_87, x = scatter_39_cast_fp16)[name = tensor("reshape_199_cast_fp16")]; + tensor reshape_199_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("reshape_199_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor keys_115_begin_0 = const()[name = tensor("keys_115_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor keys_115_end_0 = const()[name = tensor("keys_115_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor keys_115_end_mask_0 = const()[name = tensor("keys_115_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor keys_115_squeeze_mask_0 = const()[name = tensor("keys_115_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor keys_115_cast_fp16 = slice_by_index(begin = keys_115_begin_0, end = keys_115_end_0, end_mask = keys_115_end_mask_0, squeeze_mask = keys_115_squeeze_mask_0, x = reshape_199_cast_fp16)[name = tensor("keys_115_cast_fp16")]; + tensor values_115_begin_0 = const()[name = tensor("values_115_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor values_115_end_0 = const()[name = tensor("values_115_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor values_115_end_mask_0 = const()[name = tensor("values_115_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor values_115_squeeze_mask_0 = const()[name = tensor("values_115_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor values_115_cast_fp16 = slice_by_index(begin = values_115_begin_0, end = values_115_end_0, end_mask = values_115_end_mask_0, squeeze_mask = values_115_squeeze_mask_0, x = reshape_199_cast_fp16)[name = tensor("values_115_cast_fp16")]; + tensor var_8459_cast_fp16 = not_equal(x = keys_115_cast_fp16, y = keys_115_cast_fp16)[name = tensor("op_8459_cast_fp16")]; + tensor keys_117_cast_fp16 = select(a = var_504_to_fp16, b = keys_115_cast_fp16, cond = var_8459_cast_fp16)[name = tensor("keys_117_cast_fp16")]; + tensor var_8467_cast_fp16 = not_equal(x = values_115_cast_fp16, y = values_115_cast_fp16)[name = tensor("op_8467_cast_fp16")]; + tensor values_117_cast_fp16 = select(a = var_504_to_fp16, b = values_115_cast_fp16, cond = var_8467_cast_fp16)[name = tensor("values_117_cast_fp16")]; + tensor var_8491 = const()[name = tensor("op_8491"), val = tensor([0, 2, 1, 3])]; + tensor var_8504 = const()[name = tensor("op_8504"), val = tensor([1, 1, 1])]; + tensor var_8505_cast_fp16 = reshape(shape = var_8504, x = position19_to_fp16)[name = tensor("op_8505_cast_fp16")]; + tensor var_8522_to_fp16 = const()[name = tensor("op_8522_to_fp16"), val = tensor(0x1p+0)]; + tensor valid_len_39_cast_fp16 = add(x = var_8505_cast_fp16, y = var_8522_to_fp16)[name = tensor("valid_len_39_cast_fp16")]; + tensor valid_mask_39_cast_fp16 = less(x = k_positions_1_promoted_to_fp16, y = valid_len_39_cast_fp16)[name = tensor("valid_mask_39_cast_fp16")]; + tensor causal_mask_39_cast_fp16 = less_equal(x = k_positions_1_promoted_to_fp16, y = var_8505_cast_fp16)[name = tensor("causal_mask_39_cast_fp16")]; + tensor attn_mask_77 = logical_and(x = valid_mask_39_cast_fp16, y = causal_mask_39_cast_fp16)[name = tensor("attn_mask_77")]; + tensor attn_mask_79_axes_0 = const()[name = tensor("attn_mask_79_axes_0"), val = tensor([1])]; + tensor attn_mask_79 = expand_dims(axes = attn_mask_79_axes_0, x = attn_mask_77)[name = tensor("attn_mask_79")]; + tensor var_8540_transpose_x_0 = const()[name = tensor("op_8540_transpose_x_0"), val = tensor(false)]; + tensor var_8540_transpose_y_0 = const()[name = tensor("op_8540_transpose_y_0"), val = tensor(false)]; + tensor transpose_110_perm_0 = const()[name = tensor("transpose_110_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_111_perm_0 = const()[name = tensor("transpose_111_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_111 = transpose(perm = transpose_111_perm_0, x = keys_117_cast_fp16)[name = tensor("transpose_137")]; + tensor transpose_110 = transpose(perm = transpose_110_perm_0, x = q_117_cast_fp16)[name = tensor("transpose_138")]; + tensor var_8540_cast_fp16 = matmul(transpose_x = var_8540_transpose_x_0, transpose_y = var_8540_transpose_y_0, x = transpose_110, y = transpose_111)[name = tensor("op_8540_cast_fp16")]; + tensor var_8534_to_fp16 = const()[name = tensor("op_8534_to_fp16"), val = tensor([0x1p-3])]; + tensor attn_weights_115_cast_fp16 = mul(x = var_8540_cast_fp16, y = var_8534_to_fp16)[name = tensor("attn_weights_115_cast_fp16")]; + tensor var_8542 = logical_not(x = attn_mask_79)[name = tensor("op_8542")]; + tensor var_8543_to_fp16 = const()[name = tensor("op_8543_to_fp16"), val = tensor(-inf)]; + tensor attn_weights_117_cast_fp16 = select(a = var_8543_to_fp16, b = attn_weights_115_cast_fp16, cond = var_8542)[name = tensor("attn_weights_117_cast_fp16")]; + tensor var_8545 = const()[name = tensor("op_8545"), val = tensor(-1)]; + tensor attn_weights_119_cast_fp16 = softmax(axis = var_8545, x = attn_weights_117_cast_fp16)[name = tensor("attn_weights_119_cast_fp16")]; + tensor attn_output_39_transpose_x_0 = const()[name = tensor("attn_output_39_transpose_x_0"), val = tensor(false)]; + tensor attn_output_39_transpose_y_0 = const()[name = tensor("attn_output_39_transpose_y_0"), val = tensor(false)]; + tensor values_119_cast_fp16 = transpose(perm = var_8491, x = values_117_cast_fp16)[name = tensor("transpose_139")]; + tensor attn_output_39_cast_fp16 = matmul(transpose_x = attn_output_39_transpose_x_0, transpose_y = attn_output_39_transpose_y_0, x = attn_weights_119_cast_fp16, y = values_119_cast_fp16)[name = tensor("attn_output_39_cast_fp16")]; + tensor var_8553 = const()[name = tensor("op_8553"), val = tensor([0, 2, 1, 3])]; + tensor var_8556 = const()[name = tensor("op_8556"), val = tensor([1, 1, 1024])]; + tensor var_8554_cast_fp16 = transpose(perm = var_8553, x = attn_output_39_cast_fp16)[name = tensor("transpose_136")]; + tensor input_195_cast_fp16 = reshape(shape = var_8556, x = var_8554_cast_fp16)[name = tensor("input_195_cast_fp16")]; + tensor linear_78_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = attn19_out_proj_weight_cast_fp16, x = input_195_cast_fp16)[name = tensor("linear_78_cast_fp16")]; + tensor var_8562_to_fp16 = const()[name = tensor("op_8562_to_fp16"), val = tensor(0x1p+0)]; + tensor var_8563_cast_fp16 = add(x = position19_to_fp16, y = var_8562_to_fp16)[name = tensor("op_8563_cast_fp16")]; + tensor var_8563_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_8563_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor input_197_cast_fp16 = add(x = input_193_cast_fp16, y = linear_78_cast_fp16)[name = tensor("input_197_cast_fp16")]; + tensor input_199_axes_0 = const()[name = tensor("input_199_axes_0"), val = tensor([-1])]; + tensor norm19_2_weight_to_fp16 = const()[name = tensor("norm19_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(303752192)))]; + tensor norm19_2_bias_to_fp16 = const()[name = tensor("norm19_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(303754304)))]; + tensor var_8567_to_fp16 = const()[name = tensor("op_8567_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_199_cast_fp16 = layer_norm(axes = input_199_axes_0, beta = norm19_2_bias_to_fp16, epsilon = var_8567_to_fp16, gamma = norm19_2_weight_to_fp16, x = input_197_cast_fp16)[name = tensor("input_199_cast_fp16")]; + tensor linear_79_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = linear19_1_weight_cast_fp16, x = input_199_cast_fp16)[name = tensor("linear_79_cast_fp16")]; + tensor input_201_mode_0 = const()[name = tensor("input_201_mode_0"), val = tensor("EXACT")]; + tensor input_201_cast_fp16 = gelu(mode = input_201_mode_0, x = linear_79_cast_fp16)[name = tensor("input_201_cast_fp16")]; + tensor linear_80_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = linear19_2_weight_cast_fp16, x = input_201_cast_fp16)[name = tensor("linear_80_cast_fp16")]; + tensor input_203_cast_fp16 = add(x = input_197_cast_fp16, y = linear_80_cast_fp16)[name = tensor("input_203_cast_fp16")]; + tensor x_41_axes_0 = const()[name = tensor("x_41_axes_0"), val = tensor([-1])]; + tensor norm20_1_weight_to_fp16 = const()[name = tensor("norm20_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(303756416)))]; + tensor norm20_1_bias_to_fp16 = const()[name = tensor("norm20_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(303758528)))]; + tensor var_8584_to_fp16 = const()[name = tensor("op_8584_to_fp16"), val = tensor(0x1.5p-17)]; + tensor x_41_cast_fp16 = layer_norm(axes = x_41_axes_0, beta = norm20_1_bias_to_fp16, epsilon = var_8584_to_fp16, gamma = norm20_1_weight_to_fp16, x = input_203_cast_fp16)[name = tensor("x_41_cast_fp16")]; + tensor linear_81_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = attn20_in_proj_weight_cast_fp16, x = x_41_cast_fp16)[name = tensor("linear_81_cast_fp16")]; + tensor var_8620 = const()[name = tensor("op_8620"), val = tensor([1, 1, 3, 16, 64])]; + tensor qkv_41_cast_fp16 = reshape(shape = var_8620, x = linear_81_cast_fp16)[name = tensor("qkv_41_cast_fp16")]; + tensor q_121_begin_0 = const()[name = tensor("q_121_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor q_121_end_0 = const()[name = tensor("q_121_end_0"), val = tensor([1, 1, 1, 16, 64])]; + tensor q_121_end_mask_0 = const()[name = tensor("q_121_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor q_121_squeeze_mask_0 = const()[name = tensor("q_121_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor q_121_cast_fp16 = slice_by_index(begin = q_121_begin_0, end = q_121_end_0, end_mask = q_121_end_mask_0, squeeze_mask = q_121_squeeze_mask_0, x = qkv_41_cast_fp16)[name = tensor("q_121_cast_fp16")]; + tensor k_81_begin_0 = const()[name = tensor("k_81_begin_0"), val = tensor([0, 0, 1, 0, 0])]; + tensor k_81_end_0 = const()[name = tensor("k_81_end_0"), val = tensor([1, 1, 2, 16, 64])]; + tensor k_81_end_mask_0 = const()[name = tensor("k_81_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor k_81_squeeze_mask_0 = const()[name = tensor("k_81_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor k_81_cast_fp16 = slice_by_index(begin = k_81_begin_0, end = k_81_end_0, end_mask = k_81_end_mask_0, squeeze_mask = k_81_squeeze_mask_0, x = qkv_41_cast_fp16)[name = tensor("k_81_cast_fp16")]; + tensor v_41_begin_0 = const()[name = tensor("v_41_begin_0"), val = tensor([0, 0, 2, 0, 0])]; + tensor v_41_end_0 = const()[name = tensor("v_41_end_0"), val = tensor([1, 1, 3, 16, 64])]; + tensor v_41_end_mask_0 = const()[name = tensor("v_41_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor v_41_squeeze_mask_0 = const()[name = tensor("v_41_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor v_41_cast_fp16 = slice_by_index(begin = v_41_begin_0, end = v_41_end_0, end_mask = v_41_end_mask_0, squeeze_mask = v_41_squeeze_mask_0, x = qkv_41_cast_fp16)[name = tensor("v_41_cast_fp16")]; + tensor var_8724 = const()[name = tensor("op_8724"), val = tensor([1, 1, 1, 1])]; + tensor position20_to_fp16_dtype_0 = const()[name = tensor("position20_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor position20_to_fp16 = cast(dtype = position20_to_fp16_dtype_0, x = position20)[name = tensor("cast_453")]; + tensor ts_125_cast_fp16 = reshape(shape = var_8724, x = position20_to_fp16)[name = tensor("ts_125_cast_fp16")]; + tensor var_8728 = const()[name = tensor("op_8728"), val = tensor([1, 1, 16, 32, 2])]; + tensor q_complex_41_cast_fp16 = reshape(shape = var_8728, x = q_121_cast_fp16)[name = tensor("q_complex_41_cast_fp16")]; + tensor var_8732 = const()[name = tensor("op_8732"), val = tensor([1, 1, 16, 32, 2])]; + tensor k_complex_41_cast_fp16 = reshape(shape = var_8732, x = k_81_cast_fp16)[name = tensor("k_complex_41_cast_fp16")]; + tensor var_8736_begin_0 = const()[name = tensor("op_8736_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_8736_end_0 = const()[name = tensor("op_8736_end_0"), val = tensor([1, 1, 16, 32, 1])]; + tensor var_8736_end_mask_0 = const()[name = tensor("op_8736_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_8736_squeeze_mask_0 = const()[name = tensor("op_8736_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_8736_cast_fp16 = slice_by_index(begin = var_8736_begin_0, end = var_8736_end_0, end_mask = var_8736_end_mask_0, squeeze_mask = var_8736_squeeze_mask_0, x = q_complex_41_cast_fp16)[name = tensor("op_8736_cast_fp16")]; + tensor var_8744_begin_0 = const()[name = tensor("op_8744_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_8744_end_0 = const()[name = tensor("op_8744_end_0"), val = tensor([1, 1, 16, 32, 2])]; + tensor var_8744_end_mask_0 = const()[name = tensor("op_8744_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_8744_squeeze_mask_0 = const()[name = tensor("op_8744_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_8744_cast_fp16 = slice_by_index(begin = var_8744_begin_0, end = var_8744_end_0, end_mask = var_8744_end_mask_0, squeeze_mask = var_8744_squeeze_mask_0, x = q_complex_41_cast_fp16)[name = tensor("op_8744_cast_fp16")]; + tensor var_8752_begin_0 = const()[name = tensor("op_8752_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_8752_end_0 = const()[name = tensor("op_8752_end_0"), val = tensor([1, 1, 16, 32, 1])]; + tensor var_8752_end_mask_0 = const()[name = tensor("op_8752_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_8752_squeeze_mask_0 = const()[name = tensor("op_8752_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_8752_cast_fp16 = slice_by_index(begin = var_8752_begin_0, end = var_8752_end_0, end_mask = var_8752_end_mask_0, squeeze_mask = var_8752_squeeze_mask_0, x = k_complex_41_cast_fp16)[name = tensor("op_8752_cast_fp16")]; + tensor var_8760_begin_0 = const()[name = tensor("op_8760_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_8760_end_0 = const()[name = tensor("op_8760_end_0"), val = tensor([1, 1, 16, 32, 2])]; + tensor var_8760_end_mask_0 = const()[name = tensor("op_8760_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_8760_squeeze_mask_0 = const()[name = tensor("op_8760_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_8760_cast_fp16 = slice_by_index(begin = var_8760_begin_0, end = var_8760_end_0, end_mask = var_8760_end_mask_0, squeeze_mask = var_8760_squeeze_mask_0, x = k_complex_41_cast_fp16)[name = tensor("op_8760_cast_fp16")]; + tensor freqs_41_to_fp16 = const()[name = tensor("freqs_41_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(303760640)))]; + tensor var_8766_cast_fp16 = mul(x = freqs_41_to_fp16, y = ts_125_cast_fp16)[name = tensor("op_8766_cast_fp16")]; + tensor rotr_41_cast_fp16 = cos(x = var_8766_cast_fp16)[name = tensor("rotr_41_cast_fp16")]; + tensor roti_41_cast_fp16 = sin(x = var_8766_cast_fp16)[name = tensor("roti_41_cast_fp16")]; + tensor var_8770_cast_fp16 = mul(x = var_8736_cast_fp16, y = rotr_41_cast_fp16)[name = tensor("op_8770_cast_fp16")]; + tensor var_8771_cast_fp16 = mul(x = var_8744_cast_fp16, y = roti_41_cast_fp16)[name = tensor("op_8771_cast_fp16")]; + tensor qor_81_cast_fp16 = sub(x = var_8770_cast_fp16, y = var_8771_cast_fp16)[name = tensor("qor_81_cast_fp16")]; + tensor var_8774_cast_fp16 = mul(x = var_8736_cast_fp16, y = roti_41_cast_fp16)[name = tensor("op_8774_cast_fp16")]; + tensor var_8775_cast_fp16 = mul(x = var_8744_cast_fp16, y = rotr_41_cast_fp16)[name = tensor("op_8775_cast_fp16")]; + tensor qoi_81_cast_fp16 = add(x = var_8774_cast_fp16, y = var_8775_cast_fp16)[name = tensor("qoi_81_cast_fp16")]; + tensor var_8778_cast_fp16 = mul(x = var_8752_cast_fp16, y = rotr_41_cast_fp16)[name = tensor("op_8778_cast_fp16")]; + tensor var_8779_cast_fp16 = mul(x = var_8760_cast_fp16, y = roti_41_cast_fp16)[name = tensor("op_8779_cast_fp16")]; + tensor kor_81_cast_fp16 = sub(x = var_8778_cast_fp16, y = var_8779_cast_fp16)[name = tensor("kor_81_cast_fp16")]; + tensor var_8782_cast_fp16 = mul(x = var_8752_cast_fp16, y = roti_41_cast_fp16)[name = tensor("op_8782_cast_fp16")]; + tensor var_8783_cast_fp16 = mul(x = var_8760_cast_fp16, y = rotr_41_cast_fp16)[name = tensor("op_8783_cast_fp16")]; + tensor koi_81_cast_fp16 = add(x = var_8782_cast_fp16, y = var_8783_cast_fp16)[name = tensor("koi_81_cast_fp16")]; + tensor qo_41_axis_0 = const()[name = tensor("qo_41_axis_0"), val = tensor(-1)]; + tensor qo_41_cast_fp16 = stack(axis = qo_41_axis_0, values = (qor_81_cast_fp16, qoi_81_cast_fp16))[name = tensor("qo_41_cast_fp16")]; + tensor ko_41_axis_0 = const()[name = tensor("ko_41_axis_0"), val = tensor(-1)]; + tensor ko_41_cast_fp16 = stack(axis = ko_41_axis_0, values = (kor_81_cast_fp16, koi_81_cast_fp16))[name = tensor("ko_41_cast_fp16")]; + tensor var_8812 = const()[name = tensor("op_8812"), val = tensor([1, 1, 16, 64])]; + tensor q_123_cast_fp16 = reshape(shape = var_8812, x = qo_41_cast_fp16)[name = tensor("q_123_cast_fp16")]; + tensor var_8814 = const()[name = tensor("op_8814"), val = tensor([1, 1, 16, 64])]; + tensor k_83_cast_fp16 = reshape(shape = var_8814, x = ko_41_cast_fp16)[name = tensor("k_83_cast_fp16")]; + tensor _inversed_8836_y_0_to_fp16 = const()[name = tensor("_inversed_8836_y_0_to_fp16"), val = tensor(0x1p-9)]; + tensor _inversed_8836_cast_fp16 = mul(x = ts_125_cast_fp16, y = _inversed_8836_y_0_to_fp16)[name = tensor("_inversed_8836_cast_fp16")]; + tensor var_8837_cast_fp16 = floor(x = _inversed_8836_cast_fp16)[name = tensor("op_8837_cast_fp16")]; + tensor var_8838_to_fp16 = const()[name = tensor("op_8838_to_fp16"), val = tensor(0x1p+9)]; + tensor var_8839_cast_fp16 = mul(x = var_8837_cast_fp16, y = var_8838_to_fp16)[name = tensor("op_8839_cast_fp16")]; + tensor write_indices_float_83_cast_fp16 = sub(x = ts_125_cast_fp16, y = var_8839_cast_fp16)[name = tensor("write_indices_float_83_cast_fp16")]; + tensor var_8846_dtype_0 = const()[name = tensor("op_8846_dtype_0"), val = tensor("int32")]; + tensor write_indices_41_reps_0 = const()[name = tensor("write_indices_41_reps_0"), val = tensor([1, 1, 16, 64])]; + tensor write_indices_float_83_cast_fp16_to_int32 = cast(dtype = var_8846_dtype_0, x = write_indices_float_83_cast_fp16)[name = tensor("cast_452")]; + tensor write_indices_41 = tile(reps = write_indices_41_reps_0, x = write_indices_float_83_cast_fp16_to_int32)[name = tensor("write_indices_41")]; + tensor var_8854_begin_0 = const()[name = tensor("op_8854_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_8854_end_0 = const()[name = tensor("op_8854_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor var_8854_end_mask_0 = const()[name = tensor("op_8854_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_8854_squeeze_mask_0 = const()[name = tensor("op_8854_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor cache20_to_fp16_dtype_0 = const()[name = tensor("cache20_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor cache20_to_fp16 = cast(dtype = cache20_to_fp16_dtype_0, x = cache20)[name = tensor("cast_451")]; + tensor var_8854_cast_fp16 = slice_by_index(begin = var_8854_begin_0, end = var_8854_end_0, end_mask = var_8854_end_mask_0, squeeze_mask = var_8854_squeeze_mask_0, x = cache20_to_fp16)[name = tensor("op_8854_cast_fp16")]; + tensor var_8856_axis_0 = const()[name = tensor("op_8856_axis_0"), val = tensor(1)]; + tensor var_8856_mode_0 = const()[name = tensor("op_8856_mode_0"), val = tensor("update")]; + tensor var_8856_validate_indices_0 = const()[name = tensor("op_8856_validate_indices_0"), val = tensor(false)]; + tensor var_8856_cast_fp16 = scatter_along_axis(axis = var_8856_axis_0, data = var_8854_cast_fp16, indices = write_indices_41, mode = var_8856_mode_0, updates = k_83_cast_fp16, validate_indices = var_8856_validate_indices_0)[name = tensor("op_8856_cast_fp16")]; + tensor concat_142 = const()[name = tensor("concat_142"), val = tensor([0, 0, 0, 0, 0])]; + tensor concat_143 = const()[name = tensor("concat_143"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_41_internal_tensor_assign_1_stride_0 = const()[name = tensor("new_cache_41_internal_tensor_assign_1_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_41_internal_tensor_assign_1_begin_mask_0 = const()[name = tensor("new_cache_41_internal_tensor_assign_1_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_41_internal_tensor_assign_1_end_mask_0 = const()[name = tensor("new_cache_41_internal_tensor_assign_1_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_41_internal_tensor_assign_1_squeeze_mask_0 = const()[name = tensor("new_cache_41_internal_tensor_assign_1_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_88 = const()[name = tensor("shape_88"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_40 = const()[name = tensor("reduce_prod_40"), val = tensor(1048576)]; + tensor range_1d_40_start_0 = const()[name = tensor("range_1d_40_start_0"), val = tensor(0)]; + tensor range_1d_40_step_0 = const()[name = tensor("range_1d_40_step_0"), val = tensor(1)]; + tensor range_1d_40 = range_1d(end = reduce_prod_40, start = range_1d_40_start_0, step = range_1d_40_step_0)[name = tensor("range_1d_40")]; + tensor reshape_200 = reshape(shape = shape_88, x = range_1d_40)[name = tensor("reshape_200")]; + tensor slice_by_index_40 = slice_by_index(begin = concat_142, begin_mask = new_cache_41_internal_tensor_assign_1_begin_mask_0, end = concat_143, end_mask = new_cache_41_internal_tensor_assign_1_end_mask_0, squeeze_mask = new_cache_41_internal_tensor_assign_1_squeeze_mask_0, stride = new_cache_41_internal_tensor_assign_1_stride_0, x = reshape_200)[name = tensor("slice_by_index_40")]; + tensor reshape_201_shape_0 = const()[name = tensor("reshape_201_shape_0"), val = tensor([-1])]; + tensor reshape_201 = reshape(shape = reshape_201_shape_0, x = slice_by_index_40)[name = tensor("reshape_201")]; + tensor reshape_202_shape_0 = const()[name = tensor("reshape_202_shape_0"), val = tensor([-1])]; + tensor reshape_202_cast_fp16 = reshape(shape = reshape_202_shape_0, x = var_8856_cast_fp16)[name = tensor("reshape_202_cast_fp16")]; + tensor reshape_203_shape_0 = const()[name = tensor("reshape_203_shape_0"), val = tensor([-1])]; + tensor reshape_203_cast_fp16 = reshape(shape = reshape_203_shape_0, x = cache20_to_fp16)[name = tensor("reshape_203_cast_fp16")]; + tensor scatter_40_mode_0 = const()[name = tensor("scatter_40_mode_0"), val = tensor("update")]; + tensor scatter_40_axis_0 = const()[name = tensor("scatter_40_axis_0"), val = tensor(0)]; + tensor scatter_40_validate_indices_0 = const()[name = tensor("scatter_40_validate_indices_0"), val = tensor(false)]; + tensor scatter_40_cast_fp16 = scatter(axis = scatter_40_axis_0, data = reshape_203_cast_fp16, indices = reshape_201, mode = scatter_40_mode_0, updates = reshape_202_cast_fp16, validate_indices = scatter_40_validate_indices_0)[name = tensor("scatter_40_cast_fp16")]; + tensor reshape_204_cast_fp16 = reshape(shape = shape_88, x = scatter_40_cast_fp16)[name = tensor("reshape_204_cast_fp16")]; + tensor var_8864_begin_0 = const()[name = tensor("op_8864_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor var_8864_end_0 = const()[name = tensor("op_8864_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor var_8864_end_mask_0 = const()[name = tensor("op_8864_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_8864_squeeze_mask_0 = const()[name = tensor("op_8864_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor var_8864_cast_fp16 = slice_by_index(begin = var_8864_begin_0, end = var_8864_end_0, end_mask = var_8864_end_mask_0, squeeze_mask = var_8864_squeeze_mask_0, x = reshape_204_cast_fp16)[name = tensor("op_8864_cast_fp16")]; + tensor var_8866_axis_0 = const()[name = tensor("op_8866_axis_0"), val = tensor(1)]; + tensor var_8866_mode_0 = const()[name = tensor("op_8866_mode_0"), val = tensor("update")]; + tensor var_8866_validate_indices_0 = const()[name = tensor("op_8866_validate_indices_0"), val = tensor(false)]; + tensor var_8866_cast_fp16 = scatter_along_axis(axis = var_8866_axis_0, data = var_8864_cast_fp16, indices = write_indices_41, mode = var_8866_mode_0, updates = v_41_cast_fp16, validate_indices = var_8866_validate_indices_0)[name = tensor("op_8866_cast_fp16")]; + tensor concat_144 = const()[name = tensor("concat_144"), val = tensor([1, 0, 0, 0, 0])]; + tensor concat_145 = const()[name = tensor("concat_145"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_41_internal_tensor_assign_2_stride_0 = const()[name = tensor("new_cache_41_internal_tensor_assign_2_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_41_internal_tensor_assign_2_begin_mask_0 = const()[name = tensor("new_cache_41_internal_tensor_assign_2_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_41_internal_tensor_assign_2_end_mask_0 = const()[name = tensor("new_cache_41_internal_tensor_assign_2_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_41_internal_tensor_assign_2_squeeze_mask_0 = const()[name = tensor("new_cache_41_internal_tensor_assign_2_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_89 = const()[name = tensor("shape_89"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_41 = const()[name = tensor("reduce_prod_41"), val = tensor(1048576)]; + tensor range_1d_41_start_0 = const()[name = tensor("range_1d_41_start_0"), val = tensor(0)]; + tensor range_1d_41_step_0 = const()[name = tensor("range_1d_41_step_0"), val = tensor(1)]; + tensor range_1d_41 = range_1d(end = reduce_prod_41, start = range_1d_41_start_0, step = range_1d_41_step_0)[name = tensor("range_1d_41")]; + tensor reshape_205 = reshape(shape = shape_89, x = range_1d_41)[name = tensor("reshape_205")]; + tensor slice_by_index_41 = slice_by_index(begin = concat_144, begin_mask = new_cache_41_internal_tensor_assign_2_begin_mask_0, end = concat_145, end_mask = new_cache_41_internal_tensor_assign_2_end_mask_0, squeeze_mask = new_cache_41_internal_tensor_assign_2_squeeze_mask_0, stride = new_cache_41_internal_tensor_assign_2_stride_0, x = reshape_205)[name = tensor("slice_by_index_41")]; + tensor reshape_206_shape_0 = const()[name = tensor("reshape_206_shape_0"), val = tensor([-1])]; + tensor reshape_206 = reshape(shape = reshape_206_shape_0, x = slice_by_index_41)[name = tensor("reshape_206")]; + tensor reshape_207_shape_0 = const()[name = tensor("reshape_207_shape_0"), val = tensor([-1])]; + tensor reshape_207_cast_fp16 = reshape(shape = reshape_207_shape_0, x = var_8866_cast_fp16)[name = tensor("reshape_207_cast_fp16")]; + tensor reshape_208_shape_0 = const()[name = tensor("reshape_208_shape_0"), val = tensor([-1])]; + tensor reshape_208_cast_fp16 = reshape(shape = reshape_208_shape_0, x = reshape_204_cast_fp16)[name = tensor("reshape_208_cast_fp16")]; + tensor scatter_41_mode_0 = const()[name = tensor("scatter_41_mode_0"), val = tensor("update")]; + tensor scatter_41_axis_0 = const()[name = tensor("scatter_41_axis_0"), val = tensor(0)]; + tensor scatter_41_validate_indices_0 = const()[name = tensor("scatter_41_validate_indices_0"), val = tensor(false)]; + tensor scatter_41_cast_fp16 = scatter(axis = scatter_41_axis_0, data = reshape_208_cast_fp16, indices = reshape_206, mode = scatter_41_mode_0, updates = reshape_207_cast_fp16, validate_indices = scatter_41_validate_indices_0)[name = tensor("scatter_41_cast_fp16")]; + tensor reshape_209_cast_fp16 = reshape(shape = shape_89, x = scatter_41_cast_fp16)[name = tensor("reshape_209_cast_fp16")]; + tensor reshape_209_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("reshape_209_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor keys_121_begin_0 = const()[name = tensor("keys_121_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor keys_121_end_0 = const()[name = tensor("keys_121_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor keys_121_end_mask_0 = const()[name = tensor("keys_121_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor keys_121_squeeze_mask_0 = const()[name = tensor("keys_121_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor keys_121_cast_fp16 = slice_by_index(begin = keys_121_begin_0, end = keys_121_end_0, end_mask = keys_121_end_mask_0, squeeze_mask = keys_121_squeeze_mask_0, x = reshape_209_cast_fp16)[name = tensor("keys_121_cast_fp16")]; + tensor values_121_begin_0 = const()[name = tensor("values_121_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor values_121_end_0 = const()[name = tensor("values_121_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor values_121_end_mask_0 = const()[name = tensor("values_121_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor values_121_squeeze_mask_0 = const()[name = tensor("values_121_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor values_121_cast_fp16 = slice_by_index(begin = values_121_begin_0, end = values_121_end_0, end_mask = values_121_end_mask_0, squeeze_mask = values_121_squeeze_mask_0, x = reshape_209_cast_fp16)[name = tensor("values_121_cast_fp16")]; + tensor var_8878_cast_fp16 = not_equal(x = keys_121_cast_fp16, y = keys_121_cast_fp16)[name = tensor("op_8878_cast_fp16")]; + tensor keys_123_cast_fp16 = select(a = var_504_to_fp16, b = keys_121_cast_fp16, cond = var_8878_cast_fp16)[name = tensor("keys_123_cast_fp16")]; + tensor var_8886_cast_fp16 = not_equal(x = values_121_cast_fp16, y = values_121_cast_fp16)[name = tensor("op_8886_cast_fp16")]; + tensor values_123_cast_fp16 = select(a = var_504_to_fp16, b = values_121_cast_fp16, cond = var_8886_cast_fp16)[name = tensor("values_123_cast_fp16")]; + tensor var_8910 = const()[name = tensor("op_8910"), val = tensor([0, 2, 1, 3])]; + tensor var_8923 = const()[name = tensor("op_8923"), val = tensor([1, 1, 1])]; + tensor var_8924_cast_fp16 = reshape(shape = var_8923, x = position20_to_fp16)[name = tensor("op_8924_cast_fp16")]; + tensor var_8941_to_fp16 = const()[name = tensor("op_8941_to_fp16"), val = tensor(0x1p+0)]; + tensor valid_len_41_cast_fp16 = add(x = var_8924_cast_fp16, y = var_8941_to_fp16)[name = tensor("valid_len_41_cast_fp16")]; + tensor valid_mask_41_cast_fp16 = less(x = k_positions_1_promoted_to_fp16, y = valid_len_41_cast_fp16)[name = tensor("valid_mask_41_cast_fp16")]; + tensor causal_mask_41_cast_fp16 = less_equal(x = k_positions_1_promoted_to_fp16, y = var_8924_cast_fp16)[name = tensor("causal_mask_41_cast_fp16")]; + tensor attn_mask_81 = logical_and(x = valid_mask_41_cast_fp16, y = causal_mask_41_cast_fp16)[name = tensor("attn_mask_81")]; + tensor attn_mask_83_axes_0 = const()[name = tensor("attn_mask_83_axes_0"), val = tensor([1])]; + tensor attn_mask_83 = expand_dims(axes = attn_mask_83_axes_0, x = attn_mask_81)[name = tensor("attn_mask_83")]; + tensor var_8959_transpose_x_0 = const()[name = tensor("op_8959_transpose_x_0"), val = tensor(false)]; + tensor var_8959_transpose_y_0 = const()[name = tensor("op_8959_transpose_y_0"), val = tensor(false)]; + tensor transpose_112_perm_0 = const()[name = tensor("transpose_112_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_113_perm_0 = const()[name = tensor("transpose_113_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_113 = transpose(perm = transpose_113_perm_0, x = keys_123_cast_fp16)[name = tensor("transpose_133")]; + tensor transpose_112 = transpose(perm = transpose_112_perm_0, x = q_123_cast_fp16)[name = tensor("transpose_134")]; + tensor var_8959_cast_fp16 = matmul(transpose_x = var_8959_transpose_x_0, transpose_y = var_8959_transpose_y_0, x = transpose_112, y = transpose_113)[name = tensor("op_8959_cast_fp16")]; + tensor var_8953_to_fp16 = const()[name = tensor("op_8953_to_fp16"), val = tensor([0x1p-3])]; + tensor attn_weights_121_cast_fp16 = mul(x = var_8959_cast_fp16, y = var_8953_to_fp16)[name = tensor("attn_weights_121_cast_fp16")]; + tensor var_8961 = logical_not(x = attn_mask_83)[name = tensor("op_8961")]; + tensor var_8962_to_fp16 = const()[name = tensor("op_8962_to_fp16"), val = tensor(-inf)]; + tensor attn_weights_123_cast_fp16 = select(a = var_8962_to_fp16, b = attn_weights_121_cast_fp16, cond = var_8961)[name = tensor("attn_weights_123_cast_fp16")]; + tensor var_8964 = const()[name = tensor("op_8964"), val = tensor(-1)]; + tensor attn_weights_125_cast_fp16 = softmax(axis = var_8964, x = attn_weights_123_cast_fp16)[name = tensor("attn_weights_125_cast_fp16")]; + tensor attn_output_41_transpose_x_0 = const()[name = tensor("attn_output_41_transpose_x_0"), val = tensor(false)]; + tensor attn_output_41_transpose_y_0 = const()[name = tensor("attn_output_41_transpose_y_0"), val = tensor(false)]; + tensor values_125_cast_fp16 = transpose(perm = var_8910, x = values_123_cast_fp16)[name = tensor("transpose_135")]; + tensor attn_output_41_cast_fp16 = matmul(transpose_x = attn_output_41_transpose_x_0, transpose_y = attn_output_41_transpose_y_0, x = attn_weights_125_cast_fp16, y = values_125_cast_fp16)[name = tensor("attn_output_41_cast_fp16")]; + tensor var_8972 = const()[name = tensor("op_8972"), val = tensor([0, 2, 1, 3])]; + tensor var_8975 = const()[name = tensor("op_8975"), val = tensor([1, 1, 1024])]; + tensor var_8973_cast_fp16 = transpose(perm = var_8972, x = attn_output_41_cast_fp16)[name = tensor("transpose_132")]; + tensor input_205_cast_fp16 = reshape(shape = var_8975, x = var_8973_cast_fp16)[name = tensor("input_205_cast_fp16")]; + tensor linear_82_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = attn20_out_proj_weight_cast_fp16, x = input_205_cast_fp16)[name = tensor("linear_82_cast_fp16")]; + tensor var_8981_to_fp16 = const()[name = tensor("op_8981_to_fp16"), val = tensor(0x1p+0)]; + tensor var_8982_cast_fp16 = add(x = position20_to_fp16, y = var_8981_to_fp16)[name = tensor("op_8982_cast_fp16")]; + tensor var_8982_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_8982_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor input_207_cast_fp16 = add(x = input_203_cast_fp16, y = linear_82_cast_fp16)[name = tensor("input_207_cast_fp16")]; + tensor input_209_axes_0 = const()[name = tensor("input_209_axes_0"), val = tensor([-1])]; + tensor norm20_2_weight_to_fp16 = const()[name = tensor("norm20_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(303760768)))]; + tensor norm20_2_bias_to_fp16 = const()[name = tensor("norm20_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(303762880)))]; + tensor var_8986_to_fp16 = const()[name = tensor("op_8986_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_209_cast_fp16 = layer_norm(axes = input_209_axes_0, beta = norm20_2_bias_to_fp16, epsilon = var_8986_to_fp16, gamma = norm20_2_weight_to_fp16, x = input_207_cast_fp16)[name = tensor("input_209_cast_fp16")]; + tensor linear_83_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = linear20_1_weight_cast_fp16, x = input_209_cast_fp16)[name = tensor("linear_83_cast_fp16")]; + tensor input_211_mode_0 = const()[name = tensor("input_211_mode_0"), val = tensor("EXACT")]; + tensor input_211_cast_fp16 = gelu(mode = input_211_mode_0, x = linear_83_cast_fp16)[name = tensor("input_211_cast_fp16")]; + tensor linear_84_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = linear20_2_weight_cast_fp16, x = input_211_cast_fp16)[name = tensor("linear_84_cast_fp16")]; + tensor input_213_cast_fp16 = add(x = input_207_cast_fp16, y = linear_84_cast_fp16)[name = tensor("input_213_cast_fp16")]; + tensor x_43_axes_0 = const()[name = tensor("x_43_axes_0"), val = tensor([-1])]; + tensor norm21_1_weight_to_fp16 = const()[name = tensor("norm21_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(303764992)))]; + tensor norm21_1_bias_to_fp16 = const()[name = tensor("norm21_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(303767104)))]; + tensor var_9003_to_fp16 = const()[name = tensor("op_9003_to_fp16"), val = tensor(0x1.5p-17)]; + tensor x_43_cast_fp16 = layer_norm(axes = x_43_axes_0, beta = norm21_1_bias_to_fp16, epsilon = var_9003_to_fp16, gamma = norm21_1_weight_to_fp16, x = input_213_cast_fp16)[name = tensor("x_43_cast_fp16")]; + tensor linear_85_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = attn21_in_proj_weight_cast_fp16, x = x_43_cast_fp16)[name = tensor("linear_85_cast_fp16")]; + tensor var_9039 = const()[name = tensor("op_9039"), val = tensor([1, 1, 3, 16, 64])]; + tensor qkv_43_cast_fp16 = reshape(shape = var_9039, x = linear_85_cast_fp16)[name = tensor("qkv_43_cast_fp16")]; + tensor q_127_begin_0 = const()[name = tensor("q_127_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor q_127_end_0 = const()[name = tensor("q_127_end_0"), val = tensor([1, 1, 1, 16, 64])]; + tensor q_127_end_mask_0 = const()[name = tensor("q_127_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor q_127_squeeze_mask_0 = const()[name = tensor("q_127_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor q_127_cast_fp16 = slice_by_index(begin = q_127_begin_0, end = q_127_end_0, end_mask = q_127_end_mask_0, squeeze_mask = q_127_squeeze_mask_0, x = qkv_43_cast_fp16)[name = tensor("q_127_cast_fp16")]; + tensor k_85_begin_0 = const()[name = tensor("k_85_begin_0"), val = tensor([0, 0, 1, 0, 0])]; + tensor k_85_end_0 = const()[name = tensor("k_85_end_0"), val = tensor([1, 1, 2, 16, 64])]; + tensor k_85_end_mask_0 = const()[name = tensor("k_85_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor k_85_squeeze_mask_0 = const()[name = tensor("k_85_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor k_85_cast_fp16 = slice_by_index(begin = k_85_begin_0, end = k_85_end_0, end_mask = k_85_end_mask_0, squeeze_mask = k_85_squeeze_mask_0, x = qkv_43_cast_fp16)[name = tensor("k_85_cast_fp16")]; + tensor v_43_begin_0 = const()[name = tensor("v_43_begin_0"), val = tensor([0, 0, 2, 0, 0])]; + tensor v_43_end_0 = const()[name = tensor("v_43_end_0"), val = tensor([1, 1, 3, 16, 64])]; + tensor v_43_end_mask_0 = const()[name = tensor("v_43_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor v_43_squeeze_mask_0 = const()[name = tensor("v_43_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor v_43_cast_fp16 = slice_by_index(begin = v_43_begin_0, end = v_43_end_0, end_mask = v_43_end_mask_0, squeeze_mask = v_43_squeeze_mask_0, x = qkv_43_cast_fp16)[name = tensor("v_43_cast_fp16")]; + tensor var_9143 = const()[name = tensor("op_9143"), val = tensor([1, 1, 1, 1])]; + tensor position21_to_fp16_dtype_0 = const()[name = tensor("position21_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor position21_to_fp16 = cast(dtype = position21_to_fp16_dtype_0, x = position21)[name = tensor("cast_448")]; + tensor ts_131_cast_fp16 = reshape(shape = var_9143, x = position21_to_fp16)[name = tensor("ts_131_cast_fp16")]; + tensor var_9147 = const()[name = tensor("op_9147"), val = tensor([1, 1, 16, 32, 2])]; + tensor q_complex_43_cast_fp16 = reshape(shape = var_9147, x = q_127_cast_fp16)[name = tensor("q_complex_43_cast_fp16")]; + tensor var_9151 = const()[name = tensor("op_9151"), val = tensor([1, 1, 16, 32, 2])]; + tensor k_complex_43_cast_fp16 = reshape(shape = var_9151, x = k_85_cast_fp16)[name = tensor("k_complex_43_cast_fp16")]; + tensor var_9155_begin_0 = const()[name = tensor("op_9155_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_9155_end_0 = const()[name = tensor("op_9155_end_0"), val = tensor([1, 1, 16, 32, 1])]; + tensor var_9155_end_mask_0 = const()[name = tensor("op_9155_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_9155_squeeze_mask_0 = const()[name = tensor("op_9155_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_9155_cast_fp16 = slice_by_index(begin = var_9155_begin_0, end = var_9155_end_0, end_mask = var_9155_end_mask_0, squeeze_mask = var_9155_squeeze_mask_0, x = q_complex_43_cast_fp16)[name = tensor("op_9155_cast_fp16")]; + tensor var_9163_begin_0 = const()[name = tensor("op_9163_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_9163_end_0 = const()[name = tensor("op_9163_end_0"), val = tensor([1, 1, 16, 32, 2])]; + tensor var_9163_end_mask_0 = const()[name = tensor("op_9163_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_9163_squeeze_mask_0 = const()[name = tensor("op_9163_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_9163_cast_fp16 = slice_by_index(begin = var_9163_begin_0, end = var_9163_end_0, end_mask = var_9163_end_mask_0, squeeze_mask = var_9163_squeeze_mask_0, x = q_complex_43_cast_fp16)[name = tensor("op_9163_cast_fp16")]; + tensor var_9171_begin_0 = const()[name = tensor("op_9171_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_9171_end_0 = const()[name = tensor("op_9171_end_0"), val = tensor([1, 1, 16, 32, 1])]; + tensor var_9171_end_mask_0 = const()[name = tensor("op_9171_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_9171_squeeze_mask_0 = const()[name = tensor("op_9171_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_9171_cast_fp16 = slice_by_index(begin = var_9171_begin_0, end = var_9171_end_0, end_mask = var_9171_end_mask_0, squeeze_mask = var_9171_squeeze_mask_0, x = k_complex_43_cast_fp16)[name = tensor("op_9171_cast_fp16")]; + tensor var_9179_begin_0 = const()[name = tensor("op_9179_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_9179_end_0 = const()[name = tensor("op_9179_end_0"), val = tensor([1, 1, 16, 32, 2])]; + tensor var_9179_end_mask_0 = const()[name = tensor("op_9179_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_9179_squeeze_mask_0 = const()[name = tensor("op_9179_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_9179_cast_fp16 = slice_by_index(begin = var_9179_begin_0, end = var_9179_end_0, end_mask = var_9179_end_mask_0, squeeze_mask = var_9179_squeeze_mask_0, x = k_complex_43_cast_fp16)[name = tensor("op_9179_cast_fp16")]; + tensor freqs_43_to_fp16 = const()[name = tensor("freqs_43_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(303769216)))]; + tensor var_9185_cast_fp16 = mul(x = freqs_43_to_fp16, y = ts_131_cast_fp16)[name = tensor("op_9185_cast_fp16")]; + tensor rotr_43_cast_fp16 = cos(x = var_9185_cast_fp16)[name = tensor("rotr_43_cast_fp16")]; + tensor roti_43_cast_fp16 = sin(x = var_9185_cast_fp16)[name = tensor("roti_43_cast_fp16")]; + tensor var_9189_cast_fp16 = mul(x = var_9155_cast_fp16, y = rotr_43_cast_fp16)[name = tensor("op_9189_cast_fp16")]; + tensor var_9190_cast_fp16 = mul(x = var_9163_cast_fp16, y = roti_43_cast_fp16)[name = tensor("op_9190_cast_fp16")]; + tensor qor_85_cast_fp16 = sub(x = var_9189_cast_fp16, y = var_9190_cast_fp16)[name = tensor("qor_85_cast_fp16")]; + tensor var_9193_cast_fp16 = mul(x = var_9155_cast_fp16, y = roti_43_cast_fp16)[name = tensor("op_9193_cast_fp16")]; + tensor var_9194_cast_fp16 = mul(x = var_9163_cast_fp16, y = rotr_43_cast_fp16)[name = tensor("op_9194_cast_fp16")]; + tensor qoi_85_cast_fp16 = add(x = var_9193_cast_fp16, y = var_9194_cast_fp16)[name = tensor("qoi_85_cast_fp16")]; + tensor var_9197_cast_fp16 = mul(x = var_9171_cast_fp16, y = rotr_43_cast_fp16)[name = tensor("op_9197_cast_fp16")]; + tensor var_9198_cast_fp16 = mul(x = var_9179_cast_fp16, y = roti_43_cast_fp16)[name = tensor("op_9198_cast_fp16")]; + tensor kor_85_cast_fp16 = sub(x = var_9197_cast_fp16, y = var_9198_cast_fp16)[name = tensor("kor_85_cast_fp16")]; + tensor var_9201_cast_fp16 = mul(x = var_9171_cast_fp16, y = roti_43_cast_fp16)[name = tensor("op_9201_cast_fp16")]; + tensor var_9202_cast_fp16 = mul(x = var_9179_cast_fp16, y = rotr_43_cast_fp16)[name = tensor("op_9202_cast_fp16")]; + tensor koi_85_cast_fp16 = add(x = var_9201_cast_fp16, y = var_9202_cast_fp16)[name = tensor("koi_85_cast_fp16")]; + tensor qo_43_axis_0 = const()[name = tensor("qo_43_axis_0"), val = tensor(-1)]; + tensor qo_43_cast_fp16 = stack(axis = qo_43_axis_0, values = (qor_85_cast_fp16, qoi_85_cast_fp16))[name = tensor("qo_43_cast_fp16")]; + tensor ko_43_axis_0 = const()[name = tensor("ko_43_axis_0"), val = tensor(-1)]; + tensor ko_43_cast_fp16 = stack(axis = ko_43_axis_0, values = (kor_85_cast_fp16, koi_85_cast_fp16))[name = tensor("ko_43_cast_fp16")]; + tensor var_9231 = const()[name = tensor("op_9231"), val = tensor([1, 1, 16, 64])]; + tensor q_129_cast_fp16 = reshape(shape = var_9231, x = qo_43_cast_fp16)[name = tensor("q_129_cast_fp16")]; + tensor var_9233 = const()[name = tensor("op_9233"), val = tensor([1, 1, 16, 64])]; + tensor k_87_cast_fp16 = reshape(shape = var_9233, x = ko_43_cast_fp16)[name = tensor("k_87_cast_fp16")]; + tensor _inversed_9255_y_0_to_fp16 = const()[name = tensor("_inversed_9255_y_0_to_fp16"), val = tensor(0x1p-9)]; + tensor _inversed_9255_cast_fp16 = mul(x = ts_131_cast_fp16, y = _inversed_9255_y_0_to_fp16)[name = tensor("_inversed_9255_cast_fp16")]; + tensor var_9256_cast_fp16 = floor(x = _inversed_9255_cast_fp16)[name = tensor("op_9256_cast_fp16")]; + tensor var_9257_to_fp16 = const()[name = tensor("op_9257_to_fp16"), val = tensor(0x1p+9)]; + tensor var_9258_cast_fp16 = mul(x = var_9256_cast_fp16, y = var_9257_to_fp16)[name = tensor("op_9258_cast_fp16")]; + tensor write_indices_float_87_cast_fp16 = sub(x = ts_131_cast_fp16, y = var_9258_cast_fp16)[name = tensor("write_indices_float_87_cast_fp16")]; + tensor var_9265_dtype_0 = const()[name = tensor("op_9265_dtype_0"), val = tensor("int32")]; + tensor write_indices_43_reps_0 = const()[name = tensor("write_indices_43_reps_0"), val = tensor([1, 1, 16, 64])]; + tensor write_indices_float_87_cast_fp16_to_int32 = cast(dtype = var_9265_dtype_0, x = write_indices_float_87_cast_fp16)[name = tensor("cast_447")]; + tensor write_indices_43 = tile(reps = write_indices_43_reps_0, x = write_indices_float_87_cast_fp16_to_int32)[name = tensor("write_indices_43")]; + tensor var_9273_begin_0 = const()[name = tensor("op_9273_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_9273_end_0 = const()[name = tensor("op_9273_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor var_9273_end_mask_0 = const()[name = tensor("op_9273_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_9273_squeeze_mask_0 = const()[name = tensor("op_9273_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor cache21_to_fp16_dtype_0 = const()[name = tensor("cache21_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor cache21_to_fp16 = cast(dtype = cache21_to_fp16_dtype_0, x = cache21)[name = tensor("cast_446")]; + tensor var_9273_cast_fp16 = slice_by_index(begin = var_9273_begin_0, end = var_9273_end_0, end_mask = var_9273_end_mask_0, squeeze_mask = var_9273_squeeze_mask_0, x = cache21_to_fp16)[name = tensor("op_9273_cast_fp16")]; + tensor var_9275_axis_0 = const()[name = tensor("op_9275_axis_0"), val = tensor(1)]; + tensor var_9275_mode_0 = const()[name = tensor("op_9275_mode_0"), val = tensor("update")]; + tensor var_9275_validate_indices_0 = const()[name = tensor("op_9275_validate_indices_0"), val = tensor(false)]; + tensor var_9275_cast_fp16 = scatter_along_axis(axis = var_9275_axis_0, data = var_9273_cast_fp16, indices = write_indices_43, mode = var_9275_mode_0, updates = k_87_cast_fp16, validate_indices = var_9275_validate_indices_0)[name = tensor("op_9275_cast_fp16")]; + tensor concat_149 = const()[name = tensor("concat_149"), val = tensor([0, 0, 0, 0, 0])]; + tensor concat_150 = const()[name = tensor("concat_150"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_43_internal_tensor_assign_1_stride_0 = const()[name = tensor("new_cache_43_internal_tensor_assign_1_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_43_internal_tensor_assign_1_begin_mask_0 = const()[name = tensor("new_cache_43_internal_tensor_assign_1_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_43_internal_tensor_assign_1_end_mask_0 = const()[name = tensor("new_cache_43_internal_tensor_assign_1_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_43_internal_tensor_assign_1_squeeze_mask_0 = const()[name = tensor("new_cache_43_internal_tensor_assign_1_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_90 = const()[name = tensor("shape_90"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_42 = const()[name = tensor("reduce_prod_42"), val = tensor(1048576)]; + tensor range_1d_42_start_0 = const()[name = tensor("range_1d_42_start_0"), val = tensor(0)]; + tensor range_1d_42_step_0 = const()[name = tensor("range_1d_42_step_0"), val = tensor(1)]; + tensor range_1d_42 = range_1d(end = reduce_prod_42, start = range_1d_42_start_0, step = range_1d_42_step_0)[name = tensor("range_1d_42")]; + tensor reshape_210 = reshape(shape = shape_90, x = range_1d_42)[name = tensor("reshape_210")]; + tensor slice_by_index_42 = slice_by_index(begin = concat_149, begin_mask = new_cache_43_internal_tensor_assign_1_begin_mask_0, end = concat_150, end_mask = new_cache_43_internal_tensor_assign_1_end_mask_0, squeeze_mask = new_cache_43_internal_tensor_assign_1_squeeze_mask_0, stride = new_cache_43_internal_tensor_assign_1_stride_0, x = reshape_210)[name = tensor("slice_by_index_42")]; + tensor reshape_211_shape_0 = const()[name = tensor("reshape_211_shape_0"), val = tensor([-1])]; + tensor reshape_211 = reshape(shape = reshape_211_shape_0, x = slice_by_index_42)[name = tensor("reshape_211")]; + tensor reshape_212_shape_0 = const()[name = tensor("reshape_212_shape_0"), val = tensor([-1])]; + tensor reshape_212_cast_fp16 = reshape(shape = reshape_212_shape_0, x = var_9275_cast_fp16)[name = tensor("reshape_212_cast_fp16")]; + tensor reshape_213_shape_0 = const()[name = tensor("reshape_213_shape_0"), val = tensor([-1])]; + tensor reshape_213_cast_fp16 = reshape(shape = reshape_213_shape_0, x = cache21_to_fp16)[name = tensor("reshape_213_cast_fp16")]; + tensor scatter_42_mode_0 = const()[name = tensor("scatter_42_mode_0"), val = tensor("update")]; + tensor scatter_42_axis_0 = const()[name = tensor("scatter_42_axis_0"), val = tensor(0)]; + tensor scatter_42_validate_indices_0 = const()[name = tensor("scatter_42_validate_indices_0"), val = tensor(false)]; + tensor scatter_42_cast_fp16 = scatter(axis = scatter_42_axis_0, data = reshape_213_cast_fp16, indices = reshape_211, mode = scatter_42_mode_0, updates = reshape_212_cast_fp16, validate_indices = scatter_42_validate_indices_0)[name = tensor("scatter_42_cast_fp16")]; + tensor reshape_214_cast_fp16 = reshape(shape = shape_90, x = scatter_42_cast_fp16)[name = tensor("reshape_214_cast_fp16")]; + tensor var_9283_begin_0 = const()[name = tensor("op_9283_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor var_9283_end_0 = const()[name = tensor("op_9283_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor var_9283_end_mask_0 = const()[name = tensor("op_9283_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_9283_squeeze_mask_0 = const()[name = tensor("op_9283_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor var_9283_cast_fp16 = slice_by_index(begin = var_9283_begin_0, end = var_9283_end_0, end_mask = var_9283_end_mask_0, squeeze_mask = var_9283_squeeze_mask_0, x = reshape_214_cast_fp16)[name = tensor("op_9283_cast_fp16")]; + tensor var_9285_axis_0 = const()[name = tensor("op_9285_axis_0"), val = tensor(1)]; + tensor var_9285_mode_0 = const()[name = tensor("op_9285_mode_0"), val = tensor("update")]; + tensor var_9285_validate_indices_0 = const()[name = tensor("op_9285_validate_indices_0"), val = tensor(false)]; + tensor var_9285_cast_fp16 = scatter_along_axis(axis = var_9285_axis_0, data = var_9283_cast_fp16, indices = write_indices_43, mode = var_9285_mode_0, updates = v_43_cast_fp16, validate_indices = var_9285_validate_indices_0)[name = tensor("op_9285_cast_fp16")]; + tensor concat_151 = const()[name = tensor("concat_151"), val = tensor([1, 0, 0, 0, 0])]; + tensor concat_152 = const()[name = tensor("concat_152"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_43_internal_tensor_assign_2_stride_0 = const()[name = tensor("new_cache_43_internal_tensor_assign_2_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_43_internal_tensor_assign_2_begin_mask_0 = const()[name = tensor("new_cache_43_internal_tensor_assign_2_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_43_internal_tensor_assign_2_end_mask_0 = const()[name = tensor("new_cache_43_internal_tensor_assign_2_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_43_internal_tensor_assign_2_squeeze_mask_0 = const()[name = tensor("new_cache_43_internal_tensor_assign_2_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_91 = const()[name = tensor("shape_91"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_43 = const()[name = tensor("reduce_prod_43"), val = tensor(1048576)]; + tensor range_1d_43_start_0 = const()[name = tensor("range_1d_43_start_0"), val = tensor(0)]; + tensor range_1d_43_step_0 = const()[name = tensor("range_1d_43_step_0"), val = tensor(1)]; + tensor range_1d_43 = range_1d(end = reduce_prod_43, start = range_1d_43_start_0, step = range_1d_43_step_0)[name = tensor("range_1d_43")]; + tensor reshape_215 = reshape(shape = shape_91, x = range_1d_43)[name = tensor("reshape_215")]; + tensor slice_by_index_43 = slice_by_index(begin = concat_151, begin_mask = new_cache_43_internal_tensor_assign_2_begin_mask_0, end = concat_152, end_mask = new_cache_43_internal_tensor_assign_2_end_mask_0, squeeze_mask = new_cache_43_internal_tensor_assign_2_squeeze_mask_0, stride = new_cache_43_internal_tensor_assign_2_stride_0, x = reshape_215)[name = tensor("slice_by_index_43")]; + tensor reshape_216_shape_0 = const()[name = tensor("reshape_216_shape_0"), val = tensor([-1])]; + tensor reshape_216 = reshape(shape = reshape_216_shape_0, x = slice_by_index_43)[name = tensor("reshape_216")]; + tensor reshape_217_shape_0 = const()[name = tensor("reshape_217_shape_0"), val = tensor([-1])]; + tensor reshape_217_cast_fp16 = reshape(shape = reshape_217_shape_0, x = var_9285_cast_fp16)[name = tensor("reshape_217_cast_fp16")]; + tensor reshape_218_shape_0 = const()[name = tensor("reshape_218_shape_0"), val = tensor([-1])]; + tensor reshape_218_cast_fp16 = reshape(shape = reshape_218_shape_0, x = reshape_214_cast_fp16)[name = tensor("reshape_218_cast_fp16")]; + tensor scatter_43_mode_0 = const()[name = tensor("scatter_43_mode_0"), val = tensor("update")]; + tensor scatter_43_axis_0 = const()[name = tensor("scatter_43_axis_0"), val = tensor(0)]; + tensor scatter_43_validate_indices_0 = const()[name = tensor("scatter_43_validate_indices_0"), val = tensor(false)]; + tensor scatter_43_cast_fp16 = scatter(axis = scatter_43_axis_0, data = reshape_218_cast_fp16, indices = reshape_216, mode = scatter_43_mode_0, updates = reshape_217_cast_fp16, validate_indices = scatter_43_validate_indices_0)[name = tensor("scatter_43_cast_fp16")]; + tensor reshape_219_cast_fp16 = reshape(shape = shape_91, x = scatter_43_cast_fp16)[name = tensor("reshape_219_cast_fp16")]; + tensor reshape_219_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("reshape_219_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor keys_127_begin_0 = const()[name = tensor("keys_127_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor keys_127_end_0 = const()[name = tensor("keys_127_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor keys_127_end_mask_0 = const()[name = tensor("keys_127_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor keys_127_squeeze_mask_0 = const()[name = tensor("keys_127_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor keys_127_cast_fp16 = slice_by_index(begin = keys_127_begin_0, end = keys_127_end_0, end_mask = keys_127_end_mask_0, squeeze_mask = keys_127_squeeze_mask_0, x = reshape_219_cast_fp16)[name = tensor("keys_127_cast_fp16")]; + tensor values_127_begin_0 = const()[name = tensor("values_127_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor values_127_end_0 = const()[name = tensor("values_127_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor values_127_end_mask_0 = const()[name = tensor("values_127_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor values_127_squeeze_mask_0 = const()[name = tensor("values_127_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor values_127_cast_fp16 = slice_by_index(begin = values_127_begin_0, end = values_127_end_0, end_mask = values_127_end_mask_0, squeeze_mask = values_127_squeeze_mask_0, x = reshape_219_cast_fp16)[name = tensor("values_127_cast_fp16")]; + tensor var_9297_cast_fp16 = not_equal(x = keys_127_cast_fp16, y = keys_127_cast_fp16)[name = tensor("op_9297_cast_fp16")]; + tensor keys_129_cast_fp16 = select(a = var_504_to_fp16, b = keys_127_cast_fp16, cond = var_9297_cast_fp16)[name = tensor("keys_129_cast_fp16")]; + tensor var_9305_cast_fp16 = not_equal(x = values_127_cast_fp16, y = values_127_cast_fp16)[name = tensor("op_9305_cast_fp16")]; + tensor values_129_cast_fp16 = select(a = var_504_to_fp16, b = values_127_cast_fp16, cond = var_9305_cast_fp16)[name = tensor("values_129_cast_fp16")]; + tensor var_9329 = const()[name = tensor("op_9329"), val = tensor([0, 2, 1, 3])]; + tensor var_9342 = const()[name = tensor("op_9342"), val = tensor([1, 1, 1])]; + tensor var_9343_cast_fp16 = reshape(shape = var_9342, x = position21_to_fp16)[name = tensor("op_9343_cast_fp16")]; + tensor var_9360_to_fp16 = const()[name = tensor("op_9360_to_fp16"), val = tensor(0x1p+0)]; + tensor valid_len_43_cast_fp16 = add(x = var_9343_cast_fp16, y = var_9360_to_fp16)[name = tensor("valid_len_43_cast_fp16")]; + tensor valid_mask_43_cast_fp16 = less(x = k_positions_1_promoted_to_fp16, y = valid_len_43_cast_fp16)[name = tensor("valid_mask_43_cast_fp16")]; + tensor causal_mask_43_cast_fp16 = less_equal(x = k_positions_1_promoted_to_fp16, y = var_9343_cast_fp16)[name = tensor("causal_mask_43_cast_fp16")]; + tensor attn_mask_85 = logical_and(x = valid_mask_43_cast_fp16, y = causal_mask_43_cast_fp16)[name = tensor("attn_mask_85")]; + tensor attn_mask_87_axes_0 = const()[name = tensor("attn_mask_87_axes_0"), val = tensor([1])]; + tensor attn_mask_87 = expand_dims(axes = attn_mask_87_axes_0, x = attn_mask_85)[name = tensor("attn_mask_87")]; + tensor var_9378_transpose_x_0 = const()[name = tensor("op_9378_transpose_x_0"), val = tensor(false)]; + tensor var_9378_transpose_y_0 = const()[name = tensor("op_9378_transpose_y_0"), val = tensor(false)]; + tensor transpose_114_perm_0 = const()[name = tensor("transpose_114_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_115_perm_0 = const()[name = tensor("transpose_115_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_115 = transpose(perm = transpose_115_perm_0, x = keys_129_cast_fp16)[name = tensor("transpose_129")]; + tensor transpose_114 = transpose(perm = transpose_114_perm_0, x = q_129_cast_fp16)[name = tensor("transpose_130")]; + tensor var_9378_cast_fp16 = matmul(transpose_x = var_9378_transpose_x_0, transpose_y = var_9378_transpose_y_0, x = transpose_114, y = transpose_115)[name = tensor("op_9378_cast_fp16")]; + tensor var_9372_to_fp16 = const()[name = tensor("op_9372_to_fp16"), val = tensor([0x1p-3])]; + tensor attn_weights_127_cast_fp16 = mul(x = var_9378_cast_fp16, y = var_9372_to_fp16)[name = tensor("attn_weights_127_cast_fp16")]; + tensor var_9380 = logical_not(x = attn_mask_87)[name = tensor("op_9380")]; + tensor var_9381_to_fp16 = const()[name = tensor("op_9381_to_fp16"), val = tensor(-inf)]; + tensor attn_weights_129_cast_fp16 = select(a = var_9381_to_fp16, b = attn_weights_127_cast_fp16, cond = var_9380)[name = tensor("attn_weights_129_cast_fp16")]; + tensor var_9383 = const()[name = tensor("op_9383"), val = tensor(-1)]; + tensor attn_weights_131_cast_fp16 = softmax(axis = var_9383, x = attn_weights_129_cast_fp16)[name = tensor("attn_weights_131_cast_fp16")]; + tensor attn_output_43_transpose_x_0 = const()[name = tensor("attn_output_43_transpose_x_0"), val = tensor(false)]; + tensor attn_output_43_transpose_y_0 = const()[name = tensor("attn_output_43_transpose_y_0"), val = tensor(false)]; + tensor values_131_cast_fp16 = transpose(perm = var_9329, x = values_129_cast_fp16)[name = tensor("transpose_131")]; + tensor attn_output_43_cast_fp16 = matmul(transpose_x = attn_output_43_transpose_x_0, transpose_y = attn_output_43_transpose_y_0, x = attn_weights_131_cast_fp16, y = values_131_cast_fp16)[name = tensor("attn_output_43_cast_fp16")]; + tensor var_9391 = const()[name = tensor("op_9391"), val = tensor([0, 2, 1, 3])]; + tensor var_9394 = const()[name = tensor("op_9394"), val = tensor([1, 1, 1024])]; + tensor var_9392_cast_fp16 = transpose(perm = var_9391, x = attn_output_43_cast_fp16)[name = tensor("transpose_128")]; + tensor input_215_cast_fp16 = reshape(shape = var_9394, x = var_9392_cast_fp16)[name = tensor("input_215_cast_fp16")]; + tensor linear_86_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = attn21_out_proj_weight_cast_fp16, x = input_215_cast_fp16)[name = tensor("linear_86_cast_fp16")]; + tensor var_9400_to_fp16 = const()[name = tensor("op_9400_to_fp16"), val = tensor(0x1p+0)]; + tensor var_9401_cast_fp16 = add(x = position21_to_fp16, y = var_9400_to_fp16)[name = tensor("op_9401_cast_fp16")]; + tensor var_9401_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_9401_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor input_217_cast_fp16 = add(x = input_213_cast_fp16, y = linear_86_cast_fp16)[name = tensor("input_217_cast_fp16")]; + tensor input_219_axes_0 = const()[name = tensor("input_219_axes_0"), val = tensor([-1])]; + tensor norm21_2_weight_to_fp16 = const()[name = tensor("norm21_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(303769344)))]; + tensor norm21_2_bias_to_fp16 = const()[name = tensor("norm21_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(303771456)))]; + tensor var_9405_to_fp16 = const()[name = tensor("op_9405_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_219_cast_fp16 = layer_norm(axes = input_219_axes_0, beta = norm21_2_bias_to_fp16, epsilon = var_9405_to_fp16, gamma = norm21_2_weight_to_fp16, x = input_217_cast_fp16)[name = tensor("input_219_cast_fp16")]; + tensor linear_87_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = linear21_1_weight_cast_fp16, x = input_219_cast_fp16)[name = tensor("linear_87_cast_fp16")]; + tensor input_221_mode_0 = const()[name = tensor("input_221_mode_0"), val = tensor("EXACT")]; + tensor input_221_cast_fp16 = gelu(mode = input_221_mode_0, x = linear_87_cast_fp16)[name = tensor("input_221_cast_fp16")]; + tensor linear_88_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = linear21_2_weight_cast_fp16, x = input_221_cast_fp16)[name = tensor("linear_88_cast_fp16")]; + tensor input_223_cast_fp16 = add(x = input_217_cast_fp16, y = linear_88_cast_fp16)[name = tensor("input_223_cast_fp16")]; + tensor x_45_axes_0 = const()[name = tensor("x_45_axes_0"), val = tensor([-1])]; + tensor norm22_1_weight_to_fp16 = const()[name = tensor("norm22_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(303773568)))]; + tensor norm22_1_bias_to_fp16 = const()[name = tensor("norm22_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(303775680)))]; + tensor var_9422_to_fp16 = const()[name = tensor("op_9422_to_fp16"), val = tensor(0x1.5p-17)]; + tensor x_45_cast_fp16 = layer_norm(axes = x_45_axes_0, beta = norm22_1_bias_to_fp16, epsilon = var_9422_to_fp16, gamma = norm22_1_weight_to_fp16, x = input_223_cast_fp16)[name = tensor("x_45_cast_fp16")]; + tensor linear_89_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = attn22_in_proj_weight_cast_fp16, x = x_45_cast_fp16)[name = tensor("linear_89_cast_fp16")]; + tensor var_9458 = const()[name = tensor("op_9458"), val = tensor([1, 1, 3, 16, 64])]; + tensor qkv_45_cast_fp16 = reshape(shape = var_9458, x = linear_89_cast_fp16)[name = tensor("qkv_45_cast_fp16")]; + tensor q_133_begin_0 = const()[name = tensor("q_133_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor q_133_end_0 = const()[name = tensor("q_133_end_0"), val = tensor([1, 1, 1, 16, 64])]; + tensor q_133_end_mask_0 = const()[name = tensor("q_133_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor q_133_squeeze_mask_0 = const()[name = tensor("q_133_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor q_133_cast_fp16 = slice_by_index(begin = q_133_begin_0, end = q_133_end_0, end_mask = q_133_end_mask_0, squeeze_mask = q_133_squeeze_mask_0, x = qkv_45_cast_fp16)[name = tensor("q_133_cast_fp16")]; + tensor k_89_begin_0 = const()[name = tensor("k_89_begin_0"), val = tensor([0, 0, 1, 0, 0])]; + tensor k_89_end_0 = const()[name = tensor("k_89_end_0"), val = tensor([1, 1, 2, 16, 64])]; + tensor k_89_end_mask_0 = const()[name = tensor("k_89_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor k_89_squeeze_mask_0 = const()[name = tensor("k_89_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor k_89_cast_fp16 = slice_by_index(begin = k_89_begin_0, end = k_89_end_0, end_mask = k_89_end_mask_0, squeeze_mask = k_89_squeeze_mask_0, x = qkv_45_cast_fp16)[name = tensor("k_89_cast_fp16")]; + tensor v_45_begin_0 = const()[name = tensor("v_45_begin_0"), val = tensor([0, 0, 2, 0, 0])]; + tensor v_45_end_0 = const()[name = tensor("v_45_end_0"), val = tensor([1, 1, 3, 16, 64])]; + tensor v_45_end_mask_0 = const()[name = tensor("v_45_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor v_45_squeeze_mask_0 = const()[name = tensor("v_45_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor v_45_cast_fp16 = slice_by_index(begin = v_45_begin_0, end = v_45_end_0, end_mask = v_45_end_mask_0, squeeze_mask = v_45_squeeze_mask_0, x = qkv_45_cast_fp16)[name = tensor("v_45_cast_fp16")]; + tensor var_9562 = const()[name = tensor("op_9562"), val = tensor([1, 1, 1, 1])]; + tensor position22_to_fp16_dtype_0 = const()[name = tensor("position22_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor position22_to_fp16 = cast(dtype = position22_to_fp16_dtype_0, x = position22)[name = tensor("cast_443")]; + tensor ts_137_cast_fp16 = reshape(shape = var_9562, x = position22_to_fp16)[name = tensor("ts_137_cast_fp16")]; + tensor var_9566 = const()[name = tensor("op_9566"), val = tensor([1, 1, 16, 32, 2])]; + tensor q_complex_45_cast_fp16 = reshape(shape = var_9566, x = q_133_cast_fp16)[name = tensor("q_complex_45_cast_fp16")]; + tensor var_9570 = const()[name = tensor("op_9570"), val = tensor([1, 1, 16, 32, 2])]; + tensor k_complex_45_cast_fp16 = reshape(shape = var_9570, x = k_89_cast_fp16)[name = tensor("k_complex_45_cast_fp16")]; + tensor var_9574_begin_0 = const()[name = tensor("op_9574_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_9574_end_0 = const()[name = tensor("op_9574_end_0"), val = tensor([1, 1, 16, 32, 1])]; + tensor var_9574_end_mask_0 = const()[name = tensor("op_9574_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_9574_squeeze_mask_0 = const()[name = tensor("op_9574_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_9574_cast_fp16 = slice_by_index(begin = var_9574_begin_0, end = var_9574_end_0, end_mask = var_9574_end_mask_0, squeeze_mask = var_9574_squeeze_mask_0, x = q_complex_45_cast_fp16)[name = tensor("op_9574_cast_fp16")]; + tensor var_9582_begin_0 = const()[name = tensor("op_9582_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_9582_end_0 = const()[name = tensor("op_9582_end_0"), val = tensor([1, 1, 16, 32, 2])]; + tensor var_9582_end_mask_0 = const()[name = tensor("op_9582_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_9582_squeeze_mask_0 = const()[name = tensor("op_9582_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_9582_cast_fp16 = slice_by_index(begin = var_9582_begin_0, end = var_9582_end_0, end_mask = var_9582_end_mask_0, squeeze_mask = var_9582_squeeze_mask_0, x = q_complex_45_cast_fp16)[name = tensor("op_9582_cast_fp16")]; + tensor var_9590_begin_0 = const()[name = tensor("op_9590_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_9590_end_0 = const()[name = tensor("op_9590_end_0"), val = tensor([1, 1, 16, 32, 1])]; + tensor var_9590_end_mask_0 = const()[name = tensor("op_9590_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_9590_squeeze_mask_0 = const()[name = tensor("op_9590_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_9590_cast_fp16 = slice_by_index(begin = var_9590_begin_0, end = var_9590_end_0, end_mask = var_9590_end_mask_0, squeeze_mask = var_9590_squeeze_mask_0, x = k_complex_45_cast_fp16)[name = tensor("op_9590_cast_fp16")]; + tensor var_9598_begin_0 = const()[name = tensor("op_9598_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_9598_end_0 = const()[name = tensor("op_9598_end_0"), val = tensor([1, 1, 16, 32, 2])]; + tensor var_9598_end_mask_0 = const()[name = tensor("op_9598_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_9598_squeeze_mask_0 = const()[name = tensor("op_9598_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_9598_cast_fp16 = slice_by_index(begin = var_9598_begin_0, end = var_9598_end_0, end_mask = var_9598_end_mask_0, squeeze_mask = var_9598_squeeze_mask_0, x = k_complex_45_cast_fp16)[name = tensor("op_9598_cast_fp16")]; + tensor freqs_45_to_fp16 = const()[name = tensor("freqs_45_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(303777792)))]; + tensor var_9604_cast_fp16 = mul(x = freqs_45_to_fp16, y = ts_137_cast_fp16)[name = tensor("op_9604_cast_fp16")]; + tensor rotr_45_cast_fp16 = cos(x = var_9604_cast_fp16)[name = tensor("rotr_45_cast_fp16")]; + tensor roti_45_cast_fp16 = sin(x = var_9604_cast_fp16)[name = tensor("roti_45_cast_fp16")]; + tensor var_9608_cast_fp16 = mul(x = var_9574_cast_fp16, y = rotr_45_cast_fp16)[name = tensor("op_9608_cast_fp16")]; + tensor var_9609_cast_fp16 = mul(x = var_9582_cast_fp16, y = roti_45_cast_fp16)[name = tensor("op_9609_cast_fp16")]; + tensor qor_89_cast_fp16 = sub(x = var_9608_cast_fp16, y = var_9609_cast_fp16)[name = tensor("qor_89_cast_fp16")]; + tensor var_9612_cast_fp16 = mul(x = var_9574_cast_fp16, y = roti_45_cast_fp16)[name = tensor("op_9612_cast_fp16")]; + tensor var_9613_cast_fp16 = mul(x = var_9582_cast_fp16, y = rotr_45_cast_fp16)[name = tensor("op_9613_cast_fp16")]; + tensor qoi_89_cast_fp16 = add(x = var_9612_cast_fp16, y = var_9613_cast_fp16)[name = tensor("qoi_89_cast_fp16")]; + tensor var_9616_cast_fp16 = mul(x = var_9590_cast_fp16, y = rotr_45_cast_fp16)[name = tensor("op_9616_cast_fp16")]; + tensor var_9617_cast_fp16 = mul(x = var_9598_cast_fp16, y = roti_45_cast_fp16)[name = tensor("op_9617_cast_fp16")]; + tensor kor_89_cast_fp16 = sub(x = var_9616_cast_fp16, y = var_9617_cast_fp16)[name = tensor("kor_89_cast_fp16")]; + tensor var_9620_cast_fp16 = mul(x = var_9590_cast_fp16, y = roti_45_cast_fp16)[name = tensor("op_9620_cast_fp16")]; + tensor var_9621_cast_fp16 = mul(x = var_9598_cast_fp16, y = rotr_45_cast_fp16)[name = tensor("op_9621_cast_fp16")]; + tensor koi_89_cast_fp16 = add(x = var_9620_cast_fp16, y = var_9621_cast_fp16)[name = tensor("koi_89_cast_fp16")]; + tensor qo_45_axis_0 = const()[name = tensor("qo_45_axis_0"), val = tensor(-1)]; + tensor qo_45_cast_fp16 = stack(axis = qo_45_axis_0, values = (qor_89_cast_fp16, qoi_89_cast_fp16))[name = tensor("qo_45_cast_fp16")]; + tensor ko_45_axis_0 = const()[name = tensor("ko_45_axis_0"), val = tensor(-1)]; + tensor ko_45_cast_fp16 = stack(axis = ko_45_axis_0, values = (kor_89_cast_fp16, koi_89_cast_fp16))[name = tensor("ko_45_cast_fp16")]; + tensor var_9650 = const()[name = tensor("op_9650"), val = tensor([1, 1, 16, 64])]; + tensor q_135_cast_fp16 = reshape(shape = var_9650, x = qo_45_cast_fp16)[name = tensor("q_135_cast_fp16")]; + tensor var_9652 = const()[name = tensor("op_9652"), val = tensor([1, 1, 16, 64])]; + tensor k_91_cast_fp16 = reshape(shape = var_9652, x = ko_45_cast_fp16)[name = tensor("k_91_cast_fp16")]; + tensor _inversed_9674_y_0_to_fp16 = const()[name = tensor("_inversed_9674_y_0_to_fp16"), val = tensor(0x1p-9)]; + tensor _inversed_9674_cast_fp16 = mul(x = ts_137_cast_fp16, y = _inversed_9674_y_0_to_fp16)[name = tensor("_inversed_9674_cast_fp16")]; + tensor var_9675_cast_fp16 = floor(x = _inversed_9674_cast_fp16)[name = tensor("op_9675_cast_fp16")]; + tensor var_9676_to_fp16 = const()[name = tensor("op_9676_to_fp16"), val = tensor(0x1p+9)]; + tensor var_9677_cast_fp16 = mul(x = var_9675_cast_fp16, y = var_9676_to_fp16)[name = tensor("op_9677_cast_fp16")]; + tensor write_indices_float_91_cast_fp16 = sub(x = ts_137_cast_fp16, y = var_9677_cast_fp16)[name = tensor("write_indices_float_91_cast_fp16")]; + tensor var_9684_dtype_0 = const()[name = tensor("op_9684_dtype_0"), val = tensor("int32")]; + tensor write_indices_45_reps_0 = const()[name = tensor("write_indices_45_reps_0"), val = tensor([1, 1, 16, 64])]; + tensor write_indices_float_91_cast_fp16_to_int32 = cast(dtype = var_9684_dtype_0, x = write_indices_float_91_cast_fp16)[name = tensor("cast_442")]; + tensor write_indices_45 = tile(reps = write_indices_45_reps_0, x = write_indices_float_91_cast_fp16_to_int32)[name = tensor("write_indices_45")]; + tensor var_9692_begin_0 = const()[name = tensor("op_9692_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_9692_end_0 = const()[name = tensor("op_9692_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor var_9692_end_mask_0 = const()[name = tensor("op_9692_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_9692_squeeze_mask_0 = const()[name = tensor("op_9692_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor cache22_to_fp16_dtype_0 = const()[name = tensor("cache22_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor cache22_to_fp16 = cast(dtype = cache22_to_fp16_dtype_0, x = cache22)[name = tensor("cast_441")]; + tensor var_9692_cast_fp16 = slice_by_index(begin = var_9692_begin_0, end = var_9692_end_0, end_mask = var_9692_end_mask_0, squeeze_mask = var_9692_squeeze_mask_0, x = cache22_to_fp16)[name = tensor("op_9692_cast_fp16")]; + tensor var_9694_axis_0 = const()[name = tensor("op_9694_axis_0"), val = tensor(1)]; + tensor var_9694_mode_0 = const()[name = tensor("op_9694_mode_0"), val = tensor("update")]; + tensor var_9694_validate_indices_0 = const()[name = tensor("op_9694_validate_indices_0"), val = tensor(false)]; + tensor var_9694_cast_fp16 = scatter_along_axis(axis = var_9694_axis_0, data = var_9692_cast_fp16, indices = write_indices_45, mode = var_9694_mode_0, updates = k_91_cast_fp16, validate_indices = var_9694_validate_indices_0)[name = tensor("op_9694_cast_fp16")]; + tensor concat_156 = const()[name = tensor("concat_156"), val = tensor([0, 0, 0, 0, 0])]; + tensor concat_157 = const()[name = tensor("concat_157"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_45_internal_tensor_assign_1_stride_0 = const()[name = tensor("new_cache_45_internal_tensor_assign_1_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_45_internal_tensor_assign_1_begin_mask_0 = const()[name = tensor("new_cache_45_internal_tensor_assign_1_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_45_internal_tensor_assign_1_end_mask_0 = const()[name = tensor("new_cache_45_internal_tensor_assign_1_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_45_internal_tensor_assign_1_squeeze_mask_0 = const()[name = tensor("new_cache_45_internal_tensor_assign_1_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_92 = const()[name = tensor("shape_92"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_44 = const()[name = tensor("reduce_prod_44"), val = tensor(1048576)]; + tensor range_1d_44_start_0 = const()[name = tensor("range_1d_44_start_0"), val = tensor(0)]; + tensor range_1d_44_step_0 = const()[name = tensor("range_1d_44_step_0"), val = tensor(1)]; + tensor range_1d_44 = range_1d(end = reduce_prod_44, start = range_1d_44_start_0, step = range_1d_44_step_0)[name = tensor("range_1d_44")]; + tensor reshape_220 = reshape(shape = shape_92, x = range_1d_44)[name = tensor("reshape_220")]; + tensor slice_by_index_44 = slice_by_index(begin = concat_156, begin_mask = new_cache_45_internal_tensor_assign_1_begin_mask_0, end = concat_157, end_mask = new_cache_45_internal_tensor_assign_1_end_mask_0, squeeze_mask = new_cache_45_internal_tensor_assign_1_squeeze_mask_0, stride = new_cache_45_internal_tensor_assign_1_stride_0, x = reshape_220)[name = tensor("slice_by_index_44")]; + tensor reshape_221_shape_0 = const()[name = tensor("reshape_221_shape_0"), val = tensor([-1])]; + tensor reshape_221 = reshape(shape = reshape_221_shape_0, x = slice_by_index_44)[name = tensor("reshape_221")]; + tensor reshape_222_shape_0 = const()[name = tensor("reshape_222_shape_0"), val = tensor([-1])]; + tensor reshape_222_cast_fp16 = reshape(shape = reshape_222_shape_0, x = var_9694_cast_fp16)[name = tensor("reshape_222_cast_fp16")]; + tensor reshape_223_shape_0 = const()[name = tensor("reshape_223_shape_0"), val = tensor([-1])]; + tensor reshape_223_cast_fp16 = reshape(shape = reshape_223_shape_0, x = cache22_to_fp16)[name = tensor("reshape_223_cast_fp16")]; + tensor scatter_44_mode_0 = const()[name = tensor("scatter_44_mode_0"), val = tensor("update")]; + tensor scatter_44_axis_0 = const()[name = tensor("scatter_44_axis_0"), val = tensor(0)]; + tensor scatter_44_validate_indices_0 = const()[name = tensor("scatter_44_validate_indices_0"), val = tensor(false)]; + tensor scatter_44_cast_fp16 = scatter(axis = scatter_44_axis_0, data = reshape_223_cast_fp16, indices = reshape_221, mode = scatter_44_mode_0, updates = reshape_222_cast_fp16, validate_indices = scatter_44_validate_indices_0)[name = tensor("scatter_44_cast_fp16")]; + tensor reshape_224_cast_fp16 = reshape(shape = shape_92, x = scatter_44_cast_fp16)[name = tensor("reshape_224_cast_fp16")]; + tensor var_9702_begin_0 = const()[name = tensor("op_9702_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor var_9702_end_0 = const()[name = tensor("op_9702_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor var_9702_end_mask_0 = const()[name = tensor("op_9702_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_9702_squeeze_mask_0 = const()[name = tensor("op_9702_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor var_9702_cast_fp16 = slice_by_index(begin = var_9702_begin_0, end = var_9702_end_0, end_mask = var_9702_end_mask_0, squeeze_mask = var_9702_squeeze_mask_0, x = reshape_224_cast_fp16)[name = tensor("op_9702_cast_fp16")]; + tensor var_9704_axis_0 = const()[name = tensor("op_9704_axis_0"), val = tensor(1)]; + tensor var_9704_mode_0 = const()[name = tensor("op_9704_mode_0"), val = tensor("update")]; + tensor var_9704_validate_indices_0 = const()[name = tensor("op_9704_validate_indices_0"), val = tensor(false)]; + tensor var_9704_cast_fp16 = scatter_along_axis(axis = var_9704_axis_0, data = var_9702_cast_fp16, indices = write_indices_45, mode = var_9704_mode_0, updates = v_45_cast_fp16, validate_indices = var_9704_validate_indices_0)[name = tensor("op_9704_cast_fp16")]; + tensor concat_158 = const()[name = tensor("concat_158"), val = tensor([1, 0, 0, 0, 0])]; + tensor concat_159 = const()[name = tensor("concat_159"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_45_internal_tensor_assign_2_stride_0 = const()[name = tensor("new_cache_45_internal_tensor_assign_2_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_45_internal_tensor_assign_2_begin_mask_0 = const()[name = tensor("new_cache_45_internal_tensor_assign_2_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_45_internal_tensor_assign_2_end_mask_0 = const()[name = tensor("new_cache_45_internal_tensor_assign_2_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_45_internal_tensor_assign_2_squeeze_mask_0 = const()[name = tensor("new_cache_45_internal_tensor_assign_2_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_93 = const()[name = tensor("shape_93"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_45 = const()[name = tensor("reduce_prod_45"), val = tensor(1048576)]; + tensor range_1d_45_start_0 = const()[name = tensor("range_1d_45_start_0"), val = tensor(0)]; + tensor range_1d_45_step_0 = const()[name = tensor("range_1d_45_step_0"), val = tensor(1)]; + tensor range_1d_45 = range_1d(end = reduce_prod_45, start = range_1d_45_start_0, step = range_1d_45_step_0)[name = tensor("range_1d_45")]; + tensor reshape_225 = reshape(shape = shape_93, x = range_1d_45)[name = tensor("reshape_225")]; + tensor slice_by_index_45 = slice_by_index(begin = concat_158, begin_mask = new_cache_45_internal_tensor_assign_2_begin_mask_0, end = concat_159, end_mask = new_cache_45_internal_tensor_assign_2_end_mask_0, squeeze_mask = new_cache_45_internal_tensor_assign_2_squeeze_mask_0, stride = new_cache_45_internal_tensor_assign_2_stride_0, x = reshape_225)[name = tensor("slice_by_index_45")]; + tensor reshape_226_shape_0 = const()[name = tensor("reshape_226_shape_0"), val = tensor([-1])]; + tensor reshape_226 = reshape(shape = reshape_226_shape_0, x = slice_by_index_45)[name = tensor("reshape_226")]; + tensor reshape_227_shape_0 = const()[name = tensor("reshape_227_shape_0"), val = tensor([-1])]; + tensor reshape_227_cast_fp16 = reshape(shape = reshape_227_shape_0, x = var_9704_cast_fp16)[name = tensor("reshape_227_cast_fp16")]; + tensor reshape_228_shape_0 = const()[name = tensor("reshape_228_shape_0"), val = tensor([-1])]; + tensor reshape_228_cast_fp16 = reshape(shape = reshape_228_shape_0, x = reshape_224_cast_fp16)[name = tensor("reshape_228_cast_fp16")]; + tensor scatter_45_mode_0 = const()[name = tensor("scatter_45_mode_0"), val = tensor("update")]; + tensor scatter_45_axis_0 = const()[name = tensor("scatter_45_axis_0"), val = tensor(0)]; + tensor scatter_45_validate_indices_0 = const()[name = tensor("scatter_45_validate_indices_0"), val = tensor(false)]; + tensor scatter_45_cast_fp16 = scatter(axis = scatter_45_axis_0, data = reshape_228_cast_fp16, indices = reshape_226, mode = scatter_45_mode_0, updates = reshape_227_cast_fp16, validate_indices = scatter_45_validate_indices_0)[name = tensor("scatter_45_cast_fp16")]; + tensor reshape_229_cast_fp16 = reshape(shape = shape_93, x = scatter_45_cast_fp16)[name = tensor("reshape_229_cast_fp16")]; + tensor reshape_229_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("reshape_229_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor keys_133_begin_0 = const()[name = tensor("keys_133_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor keys_133_end_0 = const()[name = tensor("keys_133_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor keys_133_end_mask_0 = const()[name = tensor("keys_133_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor keys_133_squeeze_mask_0 = const()[name = tensor("keys_133_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor keys_133_cast_fp16 = slice_by_index(begin = keys_133_begin_0, end = keys_133_end_0, end_mask = keys_133_end_mask_0, squeeze_mask = keys_133_squeeze_mask_0, x = reshape_229_cast_fp16)[name = tensor("keys_133_cast_fp16")]; + tensor values_133_begin_0 = const()[name = tensor("values_133_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor values_133_end_0 = const()[name = tensor("values_133_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor values_133_end_mask_0 = const()[name = tensor("values_133_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor values_133_squeeze_mask_0 = const()[name = tensor("values_133_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor values_133_cast_fp16 = slice_by_index(begin = values_133_begin_0, end = values_133_end_0, end_mask = values_133_end_mask_0, squeeze_mask = values_133_squeeze_mask_0, x = reshape_229_cast_fp16)[name = tensor("values_133_cast_fp16")]; + tensor var_9716_cast_fp16 = not_equal(x = keys_133_cast_fp16, y = keys_133_cast_fp16)[name = tensor("op_9716_cast_fp16")]; + tensor keys_135_cast_fp16 = select(a = var_504_to_fp16, b = keys_133_cast_fp16, cond = var_9716_cast_fp16)[name = tensor("keys_135_cast_fp16")]; + tensor var_9724_cast_fp16 = not_equal(x = values_133_cast_fp16, y = values_133_cast_fp16)[name = tensor("op_9724_cast_fp16")]; + tensor values_135_cast_fp16 = select(a = var_504_to_fp16, b = values_133_cast_fp16, cond = var_9724_cast_fp16)[name = tensor("values_135_cast_fp16")]; + tensor var_9748 = const()[name = tensor("op_9748"), val = tensor([0, 2, 1, 3])]; + tensor var_9761 = const()[name = tensor("op_9761"), val = tensor([1, 1, 1])]; + tensor var_9762_cast_fp16 = reshape(shape = var_9761, x = position22_to_fp16)[name = tensor("op_9762_cast_fp16")]; + tensor var_9779_to_fp16 = const()[name = tensor("op_9779_to_fp16"), val = tensor(0x1p+0)]; + tensor valid_len_45_cast_fp16 = add(x = var_9762_cast_fp16, y = var_9779_to_fp16)[name = tensor("valid_len_45_cast_fp16")]; + tensor valid_mask_45_cast_fp16 = less(x = k_positions_1_promoted_to_fp16, y = valid_len_45_cast_fp16)[name = tensor("valid_mask_45_cast_fp16")]; + tensor causal_mask_45_cast_fp16 = less_equal(x = k_positions_1_promoted_to_fp16, y = var_9762_cast_fp16)[name = tensor("causal_mask_45_cast_fp16")]; + tensor attn_mask_89 = logical_and(x = valid_mask_45_cast_fp16, y = causal_mask_45_cast_fp16)[name = tensor("attn_mask_89")]; + tensor attn_mask_91_axes_0 = const()[name = tensor("attn_mask_91_axes_0"), val = tensor([1])]; + tensor attn_mask_91 = expand_dims(axes = attn_mask_91_axes_0, x = attn_mask_89)[name = tensor("attn_mask_91")]; + tensor var_9797_transpose_x_0 = const()[name = tensor("op_9797_transpose_x_0"), val = tensor(false)]; + tensor var_9797_transpose_y_0 = const()[name = tensor("op_9797_transpose_y_0"), val = tensor(false)]; + tensor transpose_116_perm_0 = const()[name = tensor("transpose_116_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_117_perm_0 = const()[name = tensor("transpose_117_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_117 = transpose(perm = transpose_117_perm_0, x = keys_135_cast_fp16)[name = tensor("transpose_125")]; + tensor transpose_116 = transpose(perm = transpose_116_perm_0, x = q_135_cast_fp16)[name = tensor("transpose_126")]; + tensor var_9797_cast_fp16 = matmul(transpose_x = var_9797_transpose_x_0, transpose_y = var_9797_transpose_y_0, x = transpose_116, y = transpose_117)[name = tensor("op_9797_cast_fp16")]; + tensor var_9791_to_fp16 = const()[name = tensor("op_9791_to_fp16"), val = tensor([0x1p-3])]; + tensor attn_weights_133_cast_fp16 = mul(x = var_9797_cast_fp16, y = var_9791_to_fp16)[name = tensor("attn_weights_133_cast_fp16")]; + tensor var_9799 = logical_not(x = attn_mask_91)[name = tensor("op_9799")]; + tensor var_9800_to_fp16 = const()[name = tensor("op_9800_to_fp16"), val = tensor(-inf)]; + tensor attn_weights_135_cast_fp16 = select(a = var_9800_to_fp16, b = attn_weights_133_cast_fp16, cond = var_9799)[name = tensor("attn_weights_135_cast_fp16")]; + tensor var_9802 = const()[name = tensor("op_9802"), val = tensor(-1)]; + tensor attn_weights_137_cast_fp16 = softmax(axis = var_9802, x = attn_weights_135_cast_fp16)[name = tensor("attn_weights_137_cast_fp16")]; + tensor attn_output_45_transpose_x_0 = const()[name = tensor("attn_output_45_transpose_x_0"), val = tensor(false)]; + tensor attn_output_45_transpose_y_0 = const()[name = tensor("attn_output_45_transpose_y_0"), val = tensor(false)]; + tensor values_137_cast_fp16 = transpose(perm = var_9748, x = values_135_cast_fp16)[name = tensor("transpose_127")]; + tensor attn_output_45_cast_fp16 = matmul(transpose_x = attn_output_45_transpose_x_0, transpose_y = attn_output_45_transpose_y_0, x = attn_weights_137_cast_fp16, y = values_137_cast_fp16)[name = tensor("attn_output_45_cast_fp16")]; + tensor var_9810 = const()[name = tensor("op_9810"), val = tensor([0, 2, 1, 3])]; + tensor var_9813 = const()[name = tensor("op_9813"), val = tensor([1, 1, 1024])]; + tensor var_9811_cast_fp16 = transpose(perm = var_9810, x = attn_output_45_cast_fp16)[name = tensor("transpose_124")]; + tensor input_225_cast_fp16 = reshape(shape = var_9813, x = var_9811_cast_fp16)[name = tensor("input_225_cast_fp16")]; + tensor linear_90_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = attn22_out_proj_weight_cast_fp16, x = input_225_cast_fp16)[name = tensor("linear_90_cast_fp16")]; + tensor var_9819_to_fp16 = const()[name = tensor("op_9819_to_fp16"), val = tensor(0x1p+0)]; + tensor var_9820_cast_fp16 = add(x = position22_to_fp16, y = var_9819_to_fp16)[name = tensor("op_9820_cast_fp16")]; + tensor var_9820_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_9820_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor input_227_cast_fp16 = add(x = input_223_cast_fp16, y = linear_90_cast_fp16)[name = tensor("input_227_cast_fp16")]; + tensor input_229_axes_0 = const()[name = tensor("input_229_axes_0"), val = tensor([-1])]; + tensor norm22_2_weight_to_fp16 = const()[name = tensor("norm22_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(303777920)))]; + tensor norm22_2_bias_to_fp16 = const()[name = tensor("norm22_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(303780032)))]; + tensor var_9824_to_fp16 = const()[name = tensor("op_9824_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_229_cast_fp16 = layer_norm(axes = input_229_axes_0, beta = norm22_2_bias_to_fp16, epsilon = var_9824_to_fp16, gamma = norm22_2_weight_to_fp16, x = input_227_cast_fp16)[name = tensor("input_229_cast_fp16")]; + tensor linear_91_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = linear22_1_weight_cast_fp16, x = input_229_cast_fp16)[name = tensor("linear_91_cast_fp16")]; + tensor input_231_mode_0 = const()[name = tensor("input_231_mode_0"), val = tensor("EXACT")]; + tensor input_231_cast_fp16 = gelu(mode = input_231_mode_0, x = linear_91_cast_fp16)[name = tensor("input_231_cast_fp16")]; + tensor linear_92_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = linear22_2_weight_cast_fp16, x = input_231_cast_fp16)[name = tensor("linear_92_cast_fp16")]; + tensor input_233_cast_fp16 = add(x = input_227_cast_fp16, y = linear_92_cast_fp16)[name = tensor("input_233_cast_fp16")]; + tensor x_axes_0 = const()[name = tensor("x_axes_0"), val = tensor([-1])]; + tensor norm23_1_weight_to_fp16 = const()[name = tensor("norm23_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(303782144)))]; + tensor norm23_1_bias_to_fp16 = const()[name = tensor("norm23_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(303784256)))]; + tensor var_9841_to_fp16 = const()[name = tensor("op_9841_to_fp16"), val = tensor(0x1.5p-17)]; + tensor x_cast_fp16 = layer_norm(axes = x_axes_0, beta = norm23_1_bias_to_fp16, epsilon = var_9841_to_fp16, gamma = norm23_1_weight_to_fp16, x = input_233_cast_fp16)[name = tensor("x_cast_fp16")]; + tensor linear_93_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = attn23_in_proj_weight_cast_fp16, x = x_cast_fp16)[name = tensor("linear_93_cast_fp16")]; + tensor var_9877 = const()[name = tensor("op_9877"), val = tensor([1, 1, 3, 16, 64])]; + tensor qkv_cast_fp16 = reshape(shape = var_9877, x = linear_93_cast_fp16)[name = tensor("qkv_cast_fp16")]; + tensor q_139_begin_0 = const()[name = tensor("q_139_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor q_139_end_0 = const()[name = tensor("q_139_end_0"), val = tensor([1, 1, 1, 16, 64])]; + tensor q_139_end_mask_0 = const()[name = tensor("q_139_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor q_139_squeeze_mask_0 = const()[name = tensor("q_139_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor q_139_cast_fp16 = slice_by_index(begin = q_139_begin_0, end = q_139_end_0, end_mask = q_139_end_mask_0, squeeze_mask = q_139_squeeze_mask_0, x = qkv_cast_fp16)[name = tensor("q_139_cast_fp16")]; + tensor k_93_begin_0 = const()[name = tensor("k_93_begin_0"), val = tensor([0, 0, 1, 0, 0])]; + tensor k_93_end_0 = const()[name = tensor("k_93_end_0"), val = tensor([1, 1, 2, 16, 64])]; + tensor k_93_end_mask_0 = const()[name = tensor("k_93_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor k_93_squeeze_mask_0 = const()[name = tensor("k_93_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor k_93_cast_fp16 = slice_by_index(begin = k_93_begin_0, end = k_93_end_0, end_mask = k_93_end_mask_0, squeeze_mask = k_93_squeeze_mask_0, x = qkv_cast_fp16)[name = tensor("k_93_cast_fp16")]; + tensor v_begin_0 = const()[name = tensor("v_begin_0"), val = tensor([0, 0, 2, 0, 0])]; + tensor v_end_0 = const()[name = tensor("v_end_0"), val = tensor([1, 1, 3, 16, 64])]; + tensor v_end_mask_0 = const()[name = tensor("v_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor v_squeeze_mask_0 = const()[name = tensor("v_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor v_cast_fp16 = slice_by_index(begin = v_begin_0, end = v_end_0, end_mask = v_end_mask_0, squeeze_mask = v_squeeze_mask_0, x = qkv_cast_fp16)[name = tensor("v_cast_fp16")]; + tensor var_9981 = const()[name = tensor("op_9981"), val = tensor([1, 1, 1, 1])]; + tensor position23_to_fp16_dtype_0 = const()[name = tensor("position23_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor position23_to_fp16 = cast(dtype = position23_to_fp16_dtype_0, x = position23)[name = tensor("cast_438")]; + tensor ts_cast_fp16 = reshape(shape = var_9981, x = position23_to_fp16)[name = tensor("ts_cast_fp16")]; + tensor var_9985 = const()[name = tensor("op_9985"), val = tensor([1, 1, 16, 32, 2])]; + tensor q_complex_cast_fp16 = reshape(shape = var_9985, x = q_139_cast_fp16)[name = tensor("q_complex_cast_fp16")]; + tensor var_9989 = const()[name = tensor("op_9989"), val = tensor([1, 1, 16, 32, 2])]; + tensor k_complex_cast_fp16 = reshape(shape = var_9989, x = k_93_cast_fp16)[name = tensor("k_complex_cast_fp16")]; + tensor var_9993_begin_0 = const()[name = tensor("op_9993_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_9993_end_0 = const()[name = tensor("op_9993_end_0"), val = tensor([1, 1, 16, 32, 1])]; + tensor var_9993_end_mask_0 = const()[name = tensor("op_9993_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_9993_squeeze_mask_0 = const()[name = tensor("op_9993_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_9993_cast_fp16 = slice_by_index(begin = var_9993_begin_0, end = var_9993_end_0, end_mask = var_9993_end_mask_0, squeeze_mask = var_9993_squeeze_mask_0, x = q_complex_cast_fp16)[name = tensor("op_9993_cast_fp16")]; + tensor var_10001_begin_0 = const()[name = tensor("op_10001_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_10001_end_0 = const()[name = tensor("op_10001_end_0"), val = tensor([1, 1, 16, 32, 2])]; + tensor var_10001_end_mask_0 = const()[name = tensor("op_10001_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_10001_squeeze_mask_0 = const()[name = tensor("op_10001_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_10001_cast_fp16 = slice_by_index(begin = var_10001_begin_0, end = var_10001_end_0, end_mask = var_10001_end_mask_0, squeeze_mask = var_10001_squeeze_mask_0, x = q_complex_cast_fp16)[name = tensor("op_10001_cast_fp16")]; + tensor var_10009_begin_0 = const()[name = tensor("op_10009_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_10009_end_0 = const()[name = tensor("op_10009_end_0"), val = tensor([1, 1, 16, 32, 1])]; + tensor var_10009_end_mask_0 = const()[name = tensor("op_10009_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_10009_squeeze_mask_0 = const()[name = tensor("op_10009_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_10009_cast_fp16 = slice_by_index(begin = var_10009_begin_0, end = var_10009_end_0, end_mask = var_10009_end_mask_0, squeeze_mask = var_10009_squeeze_mask_0, x = k_complex_cast_fp16)[name = tensor("op_10009_cast_fp16")]; + tensor var_10017_begin_0 = const()[name = tensor("op_10017_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_10017_end_0 = const()[name = tensor("op_10017_end_0"), val = tensor([1, 1, 16, 32, 2])]; + tensor var_10017_end_mask_0 = const()[name = tensor("op_10017_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_10017_squeeze_mask_0 = const()[name = tensor("op_10017_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_10017_cast_fp16 = slice_by_index(begin = var_10017_begin_0, end = var_10017_end_0, end_mask = var_10017_end_mask_0, squeeze_mask = var_10017_squeeze_mask_0, x = k_complex_cast_fp16)[name = tensor("op_10017_cast_fp16")]; + tensor freqs_to_fp16 = const()[name = tensor("freqs_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(303786368)))]; + tensor var_10023_cast_fp16 = mul(x = freqs_to_fp16, y = ts_cast_fp16)[name = tensor("op_10023_cast_fp16")]; + tensor rotr_cast_fp16 = cos(x = var_10023_cast_fp16)[name = tensor("rotr_cast_fp16")]; + tensor roti_cast_fp16 = sin(x = var_10023_cast_fp16)[name = tensor("roti_cast_fp16")]; + tensor var_10027_cast_fp16 = mul(x = var_9993_cast_fp16, y = rotr_cast_fp16)[name = tensor("op_10027_cast_fp16")]; + tensor var_10028_cast_fp16 = mul(x = var_10001_cast_fp16, y = roti_cast_fp16)[name = tensor("op_10028_cast_fp16")]; + tensor qor_93_cast_fp16 = sub(x = var_10027_cast_fp16, y = var_10028_cast_fp16)[name = tensor("qor_93_cast_fp16")]; + tensor var_10031_cast_fp16 = mul(x = var_9993_cast_fp16, y = roti_cast_fp16)[name = tensor("op_10031_cast_fp16")]; + tensor var_10032_cast_fp16 = mul(x = var_10001_cast_fp16, y = rotr_cast_fp16)[name = tensor("op_10032_cast_fp16")]; + tensor qoi_93_cast_fp16 = add(x = var_10031_cast_fp16, y = var_10032_cast_fp16)[name = tensor("qoi_93_cast_fp16")]; + tensor var_10035_cast_fp16 = mul(x = var_10009_cast_fp16, y = rotr_cast_fp16)[name = tensor("op_10035_cast_fp16")]; + tensor var_10036_cast_fp16 = mul(x = var_10017_cast_fp16, y = roti_cast_fp16)[name = tensor("op_10036_cast_fp16")]; + tensor kor_93_cast_fp16 = sub(x = var_10035_cast_fp16, y = var_10036_cast_fp16)[name = tensor("kor_93_cast_fp16")]; + tensor var_10039_cast_fp16 = mul(x = var_10009_cast_fp16, y = roti_cast_fp16)[name = tensor("op_10039_cast_fp16")]; + tensor var_10040_cast_fp16 = mul(x = var_10017_cast_fp16, y = rotr_cast_fp16)[name = tensor("op_10040_cast_fp16")]; + tensor koi_93_cast_fp16 = add(x = var_10039_cast_fp16, y = var_10040_cast_fp16)[name = tensor("koi_93_cast_fp16")]; + tensor qo_axis_0 = const()[name = tensor("qo_axis_0"), val = tensor(-1)]; + tensor qo_cast_fp16 = stack(axis = qo_axis_0, values = (qor_93_cast_fp16, qoi_93_cast_fp16))[name = tensor("qo_cast_fp16")]; + tensor ko_axis_0 = const()[name = tensor("ko_axis_0"), val = tensor(-1)]; + tensor ko_cast_fp16 = stack(axis = ko_axis_0, values = (kor_93_cast_fp16, koi_93_cast_fp16))[name = tensor("ko_cast_fp16")]; + tensor var_10069 = const()[name = tensor("op_10069"), val = tensor([1, 1, 16, 64])]; + tensor q_141_cast_fp16 = reshape(shape = var_10069, x = qo_cast_fp16)[name = tensor("q_141_cast_fp16")]; + tensor var_10071 = const()[name = tensor("op_10071"), val = tensor([1, 1, 16, 64])]; + tensor k_cast_fp16 = reshape(shape = var_10071, x = ko_cast_fp16)[name = tensor("k_cast_fp16")]; + tensor _inversed_10093_y_0_to_fp16 = const()[name = tensor("_inversed_10093_y_0_to_fp16"), val = tensor(0x1p-9)]; + tensor _inversed_10093_cast_fp16 = mul(x = ts_cast_fp16, y = _inversed_10093_y_0_to_fp16)[name = tensor("_inversed_10093_cast_fp16")]; + tensor var_10094_cast_fp16 = floor(x = _inversed_10093_cast_fp16)[name = tensor("op_10094_cast_fp16")]; + tensor var_10095_to_fp16 = const()[name = tensor("op_10095_to_fp16"), val = tensor(0x1p+9)]; + tensor var_10096_cast_fp16 = mul(x = var_10094_cast_fp16, y = var_10095_to_fp16)[name = tensor("op_10096_cast_fp16")]; + tensor write_indices_float_cast_fp16 = sub(x = ts_cast_fp16, y = var_10096_cast_fp16)[name = tensor("write_indices_float_cast_fp16")]; + tensor var_10103_dtype_0 = const()[name = tensor("op_10103_dtype_0"), val = tensor("int32")]; + tensor write_indices_reps_0 = const()[name = tensor("write_indices_reps_0"), val = tensor([1, 1, 16, 64])]; + tensor write_indices_float_cast_fp16_to_int32 = cast(dtype = var_10103_dtype_0, x = write_indices_float_cast_fp16)[name = tensor("cast_437")]; + tensor write_indices = tile(reps = write_indices_reps_0, x = write_indices_float_cast_fp16_to_int32)[name = tensor("write_indices")]; + tensor var_10111_begin_0 = const()[name = tensor("op_10111_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_10111_end_0 = const()[name = tensor("op_10111_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor var_10111_end_mask_0 = const()[name = tensor("op_10111_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_10111_squeeze_mask_0 = const()[name = tensor("op_10111_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor cache23_to_fp16_dtype_0 = const()[name = tensor("cache23_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor cache23_to_fp16 = cast(dtype = cache23_to_fp16_dtype_0, x = cache23)[name = tensor("cast_436")]; + tensor var_10111_cast_fp16 = slice_by_index(begin = var_10111_begin_0, end = var_10111_end_0, end_mask = var_10111_end_mask_0, squeeze_mask = var_10111_squeeze_mask_0, x = cache23_to_fp16)[name = tensor("op_10111_cast_fp16")]; + tensor var_10113_axis_0 = const()[name = tensor("op_10113_axis_0"), val = tensor(1)]; + tensor var_10113_mode_0 = const()[name = tensor("op_10113_mode_0"), val = tensor("update")]; + tensor var_10113_validate_indices_0 = const()[name = tensor("op_10113_validate_indices_0"), val = tensor(false)]; + tensor var_10113_cast_fp16 = scatter_along_axis(axis = var_10113_axis_0, data = var_10111_cast_fp16, indices = write_indices, mode = var_10113_mode_0, updates = k_cast_fp16, validate_indices = var_10113_validate_indices_0)[name = tensor("op_10113_cast_fp16")]; + tensor concat_163 = const()[name = tensor("concat_163"), val = tensor([0, 0, 0, 0, 0])]; + tensor concat_164 = const()[name = tensor("concat_164"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_internal_tensor_assign_1_stride_0 = const()[name = tensor("new_cache_internal_tensor_assign_1_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_internal_tensor_assign_1_begin_mask_0 = const()[name = tensor("new_cache_internal_tensor_assign_1_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_internal_tensor_assign_1_end_mask_0 = const()[name = tensor("new_cache_internal_tensor_assign_1_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_internal_tensor_assign_1_squeeze_mask_0 = const()[name = tensor("new_cache_internal_tensor_assign_1_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_94 = const()[name = tensor("shape_94"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_46 = const()[name = tensor("reduce_prod_46"), val = tensor(1048576)]; + tensor range_1d_46_start_0 = const()[name = tensor("range_1d_46_start_0"), val = tensor(0)]; + tensor range_1d_46_step_0 = const()[name = tensor("range_1d_46_step_0"), val = tensor(1)]; + tensor range_1d_46 = range_1d(end = reduce_prod_46, start = range_1d_46_start_0, step = range_1d_46_step_0)[name = tensor("range_1d_46")]; + tensor reshape_230 = reshape(shape = shape_94, x = range_1d_46)[name = tensor("reshape_230")]; + tensor slice_by_index_46 = slice_by_index(begin = concat_163, begin_mask = new_cache_internal_tensor_assign_1_begin_mask_0, end = concat_164, end_mask = new_cache_internal_tensor_assign_1_end_mask_0, squeeze_mask = new_cache_internal_tensor_assign_1_squeeze_mask_0, stride = new_cache_internal_tensor_assign_1_stride_0, x = reshape_230)[name = tensor("slice_by_index_46")]; + tensor reshape_231_shape_0 = const()[name = tensor("reshape_231_shape_0"), val = tensor([-1])]; + tensor reshape_231 = reshape(shape = reshape_231_shape_0, x = slice_by_index_46)[name = tensor("reshape_231")]; + tensor reshape_232_shape_0 = const()[name = tensor("reshape_232_shape_0"), val = tensor([-1])]; + tensor reshape_232_cast_fp16 = reshape(shape = reshape_232_shape_0, x = var_10113_cast_fp16)[name = tensor("reshape_232_cast_fp16")]; + tensor reshape_233_shape_0 = const()[name = tensor("reshape_233_shape_0"), val = tensor([-1])]; + tensor reshape_233_cast_fp16 = reshape(shape = reshape_233_shape_0, x = cache23_to_fp16)[name = tensor("reshape_233_cast_fp16")]; + tensor scatter_46_mode_0 = const()[name = tensor("scatter_46_mode_0"), val = tensor("update")]; + tensor scatter_46_axis_0 = const()[name = tensor("scatter_46_axis_0"), val = tensor(0)]; + tensor scatter_46_validate_indices_0 = const()[name = tensor("scatter_46_validate_indices_0"), val = tensor(false)]; + tensor scatter_46_cast_fp16 = scatter(axis = scatter_46_axis_0, data = reshape_233_cast_fp16, indices = reshape_231, mode = scatter_46_mode_0, updates = reshape_232_cast_fp16, validate_indices = scatter_46_validate_indices_0)[name = tensor("scatter_46_cast_fp16")]; + tensor reshape_234_cast_fp16 = reshape(shape = shape_94, x = scatter_46_cast_fp16)[name = tensor("reshape_234_cast_fp16")]; + tensor var_10121_begin_0 = const()[name = tensor("op_10121_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor var_10121_end_0 = const()[name = tensor("op_10121_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor var_10121_end_mask_0 = const()[name = tensor("op_10121_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_10121_squeeze_mask_0 = const()[name = tensor("op_10121_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor var_10121_cast_fp16 = slice_by_index(begin = var_10121_begin_0, end = var_10121_end_0, end_mask = var_10121_end_mask_0, squeeze_mask = var_10121_squeeze_mask_0, x = reshape_234_cast_fp16)[name = tensor("op_10121_cast_fp16")]; + tensor var_10123_axis_0 = const()[name = tensor("op_10123_axis_0"), val = tensor(1)]; + tensor var_10123_mode_0 = const()[name = tensor("op_10123_mode_0"), val = tensor("update")]; + tensor var_10123_validate_indices_0 = const()[name = tensor("op_10123_validate_indices_0"), val = tensor(false)]; + tensor var_10123_cast_fp16 = scatter_along_axis(axis = var_10123_axis_0, data = var_10121_cast_fp16, indices = write_indices, mode = var_10123_mode_0, updates = v_cast_fp16, validate_indices = var_10123_validate_indices_0)[name = tensor("op_10123_cast_fp16")]; + tensor concat_165 = const()[name = tensor("concat_165"), val = tensor([1, 0, 0, 0, 0])]; + tensor concat_166 = const()[name = tensor("concat_166"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_internal_tensor_assign_2_stride_0 = const()[name = tensor("new_cache_internal_tensor_assign_2_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_internal_tensor_assign_2_begin_mask_0 = const()[name = tensor("new_cache_internal_tensor_assign_2_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_internal_tensor_assign_2_end_mask_0 = const()[name = tensor("new_cache_internal_tensor_assign_2_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_internal_tensor_assign_2_squeeze_mask_0 = const()[name = tensor("new_cache_internal_tensor_assign_2_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_95 = const()[name = tensor("shape_95"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_47 = const()[name = tensor("reduce_prod_47"), val = tensor(1048576)]; + tensor range_1d_47_start_0 = const()[name = tensor("range_1d_47_start_0"), val = tensor(0)]; + tensor range_1d_47_step_0 = const()[name = tensor("range_1d_47_step_0"), val = tensor(1)]; + tensor range_1d_47 = range_1d(end = reduce_prod_47, start = range_1d_47_start_0, step = range_1d_47_step_0)[name = tensor("range_1d_47")]; + tensor reshape_235 = reshape(shape = shape_95, x = range_1d_47)[name = tensor("reshape_235")]; + tensor slice_by_index_47 = slice_by_index(begin = concat_165, begin_mask = new_cache_internal_tensor_assign_2_begin_mask_0, end = concat_166, end_mask = new_cache_internal_tensor_assign_2_end_mask_0, squeeze_mask = new_cache_internal_tensor_assign_2_squeeze_mask_0, stride = new_cache_internal_tensor_assign_2_stride_0, x = reshape_235)[name = tensor("slice_by_index_47")]; + tensor reshape_236_shape_0 = const()[name = tensor("reshape_236_shape_0"), val = tensor([-1])]; + tensor reshape_236 = reshape(shape = reshape_236_shape_0, x = slice_by_index_47)[name = tensor("reshape_236")]; + tensor reshape_237_shape_0 = const()[name = tensor("reshape_237_shape_0"), val = tensor([-1])]; + tensor reshape_237_cast_fp16 = reshape(shape = reshape_237_shape_0, x = var_10123_cast_fp16)[name = tensor("reshape_237_cast_fp16")]; + tensor reshape_238_shape_0 = const()[name = tensor("reshape_238_shape_0"), val = tensor([-1])]; + tensor reshape_238_cast_fp16 = reshape(shape = reshape_238_shape_0, x = reshape_234_cast_fp16)[name = tensor("reshape_238_cast_fp16")]; + tensor scatter_47_mode_0 = const()[name = tensor("scatter_47_mode_0"), val = tensor("update")]; + tensor scatter_47_axis_0 = const()[name = tensor("scatter_47_axis_0"), val = tensor(0)]; + tensor scatter_47_validate_indices_0 = const()[name = tensor("scatter_47_validate_indices_0"), val = tensor(false)]; + tensor scatter_47_cast_fp16 = scatter(axis = scatter_47_axis_0, data = reshape_238_cast_fp16, indices = reshape_236, mode = scatter_47_mode_0, updates = reshape_237_cast_fp16, validate_indices = scatter_47_validate_indices_0)[name = tensor("scatter_47_cast_fp16")]; + tensor reshape_239_cast_fp16 = reshape(shape = shape_95, x = scatter_47_cast_fp16)[name = tensor("reshape_239_cast_fp16")]; + tensor reshape_239_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("reshape_239_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor keys_139_begin_0 = const()[name = tensor("keys_139_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor keys_139_end_0 = const()[name = tensor("keys_139_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor keys_139_end_mask_0 = const()[name = tensor("keys_139_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor keys_139_squeeze_mask_0 = const()[name = tensor("keys_139_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor keys_139_cast_fp16 = slice_by_index(begin = keys_139_begin_0, end = keys_139_end_0, end_mask = keys_139_end_mask_0, squeeze_mask = keys_139_squeeze_mask_0, x = reshape_239_cast_fp16)[name = tensor("keys_139_cast_fp16")]; + tensor values_139_begin_0 = const()[name = tensor("values_139_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor values_139_end_0 = const()[name = tensor("values_139_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor values_139_end_mask_0 = const()[name = tensor("values_139_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor values_139_squeeze_mask_0 = const()[name = tensor("values_139_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor values_139_cast_fp16 = slice_by_index(begin = values_139_begin_0, end = values_139_end_0, end_mask = values_139_end_mask_0, squeeze_mask = values_139_squeeze_mask_0, x = reshape_239_cast_fp16)[name = tensor("values_139_cast_fp16")]; + tensor var_10135_cast_fp16 = not_equal(x = keys_139_cast_fp16, y = keys_139_cast_fp16)[name = tensor("op_10135_cast_fp16")]; + tensor keys_141_cast_fp16 = select(a = var_504_to_fp16, b = keys_139_cast_fp16, cond = var_10135_cast_fp16)[name = tensor("keys_141_cast_fp16")]; + tensor var_10143_cast_fp16 = not_equal(x = values_139_cast_fp16, y = values_139_cast_fp16)[name = tensor("op_10143_cast_fp16")]; + tensor values_141_cast_fp16 = select(a = var_504_to_fp16, b = values_139_cast_fp16, cond = var_10143_cast_fp16)[name = tensor("values_141_cast_fp16")]; + tensor var_10167 = const()[name = tensor("op_10167"), val = tensor([0, 2, 1, 3])]; + tensor var_10180 = const()[name = tensor("op_10180"), val = tensor([1, 1, 1])]; + tensor var_10181_cast_fp16 = reshape(shape = var_10180, x = position23_to_fp16)[name = tensor("op_10181_cast_fp16")]; + tensor var_10198_to_fp16 = const()[name = tensor("op_10198_to_fp16"), val = tensor(0x1p+0)]; + tensor valid_len_cast_fp16 = add(x = var_10181_cast_fp16, y = var_10198_to_fp16)[name = tensor("valid_len_cast_fp16")]; + tensor valid_mask_cast_fp16 = less(x = k_positions_1_promoted_to_fp16, y = valid_len_cast_fp16)[name = tensor("valid_mask_cast_fp16")]; + tensor causal_mask_cast_fp16 = less_equal(x = k_positions_1_promoted_to_fp16, y = var_10181_cast_fp16)[name = tensor("causal_mask_cast_fp16")]; + tensor attn_mask_93 = logical_and(x = valid_mask_cast_fp16, y = causal_mask_cast_fp16)[name = tensor("attn_mask_93")]; + tensor attn_mask_axes_0 = const()[name = tensor("attn_mask_axes_0"), val = tensor([1])]; + tensor attn_mask = expand_dims(axes = attn_mask_axes_0, x = attn_mask_93)[name = tensor("attn_mask")]; + tensor var_10216_transpose_x_0 = const()[name = tensor("op_10216_transpose_x_0"), val = tensor(false)]; + tensor var_10216_transpose_y_0 = const()[name = tensor("op_10216_transpose_y_0"), val = tensor(false)]; + tensor transpose_118_perm_0 = const()[name = tensor("transpose_118_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_119_perm_0 = const()[name = tensor("transpose_119_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_119 = transpose(perm = transpose_119_perm_0, x = keys_141_cast_fp16)[name = tensor("transpose_121")]; + tensor transpose_118 = transpose(perm = transpose_118_perm_0, x = q_141_cast_fp16)[name = tensor("transpose_122")]; + tensor var_10216_cast_fp16 = matmul(transpose_x = var_10216_transpose_x_0, transpose_y = var_10216_transpose_y_0, x = transpose_118, y = transpose_119)[name = tensor("op_10216_cast_fp16")]; + tensor var_10210_to_fp16 = const()[name = tensor("op_10210_to_fp16"), val = tensor([0x1p-3])]; + tensor attn_weights_139_cast_fp16 = mul(x = var_10216_cast_fp16, y = var_10210_to_fp16)[name = tensor("attn_weights_139_cast_fp16")]; + tensor var_10218 = logical_not(x = attn_mask)[name = tensor("op_10218")]; + tensor var_10219_to_fp16 = const()[name = tensor("op_10219_to_fp16"), val = tensor(-inf)]; + tensor attn_weights_141_cast_fp16 = select(a = var_10219_to_fp16, b = attn_weights_139_cast_fp16, cond = var_10218)[name = tensor("attn_weights_141_cast_fp16")]; + tensor var_10221 = const()[name = tensor("op_10221"), val = tensor(-1)]; + tensor attn_weights_cast_fp16 = softmax(axis = var_10221, x = attn_weights_141_cast_fp16)[name = tensor("attn_weights_cast_fp16")]; + tensor attn_output_transpose_x_0 = const()[name = tensor("attn_output_transpose_x_0"), val = tensor(false)]; + tensor attn_output_transpose_y_0 = const()[name = tensor("attn_output_transpose_y_0"), val = tensor(false)]; + tensor values_cast_fp16 = transpose(perm = var_10167, x = values_141_cast_fp16)[name = tensor("transpose_123")]; + tensor attn_output_cast_fp16 = matmul(transpose_x = attn_output_transpose_x_0, transpose_y = attn_output_transpose_y_0, x = attn_weights_cast_fp16, y = values_cast_fp16)[name = tensor("attn_output_cast_fp16")]; + tensor var_10229 = const()[name = tensor("op_10229"), val = tensor([0, 2, 1, 3])]; + tensor var_10232 = const()[name = tensor("op_10232"), val = tensor([1, 1, 1024])]; + tensor var_10230_cast_fp16 = transpose(perm = var_10229, x = attn_output_cast_fp16)[name = tensor("transpose_120")]; + tensor input_235_cast_fp16 = reshape(shape = var_10232, x = var_10230_cast_fp16)[name = tensor("input_235_cast_fp16")]; + tensor linear_94_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = attn23_out_proj_weight_cast_fp16, x = input_235_cast_fp16)[name = tensor("linear_94_cast_fp16")]; + tensor var_10238_to_fp16 = const()[name = tensor("op_10238_to_fp16"), val = tensor(0x1p+0)]; + tensor var_10239_cast_fp16 = add(x = position23_to_fp16, y = var_10238_to_fp16)[name = tensor("op_10239_cast_fp16")]; + tensor var_10239_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_10239_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor input_237_cast_fp16 = add(x = input_233_cast_fp16, y = linear_94_cast_fp16)[name = tensor("input_237_cast_fp16")]; + tensor input_239_axes_0 = const()[name = tensor("input_239_axes_0"), val = tensor([-1])]; + tensor norm23_2_weight_to_fp16 = const()[name = tensor("norm23_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(303786496)))]; + tensor norm23_2_bias_to_fp16 = const()[name = tensor("norm23_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(303788608)))]; + tensor var_10243_to_fp16 = const()[name = tensor("op_10243_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_239_cast_fp16 = layer_norm(axes = input_239_axes_0, beta = norm23_2_bias_to_fp16, epsilon = var_10243_to_fp16, gamma = norm23_2_weight_to_fp16, x = input_237_cast_fp16)[name = tensor("input_239_cast_fp16")]; + tensor linear_95_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = linear23_1_weight_cast_fp16, x = input_239_cast_fp16)[name = tensor("linear_95_cast_fp16")]; + tensor input_241_mode_0 = const()[name = tensor("input_241_mode_0"), val = tensor("EXACT")]; + tensor input_241_cast_fp16 = gelu(mode = input_241_mode_0, x = linear_95_cast_fp16)[name = tensor("input_241_cast_fp16")]; + tensor linear_96_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = linear23_2_weight_cast_fp16, x = input_241_cast_fp16)[name = tensor("linear_96_cast_fp16")]; + tensor input_243_cast_fp16 = add(x = input_237_cast_fp16, y = linear_96_cast_fp16)[name = tensor("input_243_cast_fp16")]; + tensor input_axes_0 = const()[name = tensor("input_axes_0"), val = tensor([-1])]; + tensor out_norm_weight_to_fp16 = const()[name = tensor("out_norm_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(303790720)))]; + tensor out_norm_bias_to_fp16 = const()[name = tensor("out_norm_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(303792832)))]; + tensor var_10260_to_fp16 = const()[name = tensor("op_10260_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_cast_fp16 = layer_norm(axes = input_axes_0, beta = out_norm_bias_to_fp16, epsilon = var_10260_to_fp16, gamma = out_norm_weight_to_fp16, x = input_243_cast_fp16)[name = tensor("input_cast_fp16")]; + tensor input_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("input_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor out_eos_weight_to_fp16 = const()[name = tensor("out_eos_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(303794944)))]; + tensor out_eos_bias_to_fp16 = const()[name = tensor("out_eos_bias_to_fp16"), val = tensor([-0x1.5ap-3])]; + tensor linear_97_cast_fp16 = linear(bias = out_eos_bias_to_fp16, weight = out_eos_weight_to_fp16, x = input_cast_fp16)[name = tensor("linear_97_cast_fp16")]; + tensor linear_97_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("linear_97_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor var_10268 = cast(dtype = linear_97_cast_fp16_to_fp32_dtype_0, x = linear_97_cast_fp16)[name = tensor("cast_432")]; + tensor input = cast(dtype = input_cast_fp16_to_fp32_dtype_0, x = input_cast_fp16)[name = tensor("cast_433")]; + tensor var_10239 = cast(dtype = var_10239_cast_fp16_to_fp32_dtype_0, x = var_10239_cast_fp16)[name = tensor("cast_434")]; + tensor new_cache_internal_tensor_assign_2 = cast(dtype = reshape_239_cast_fp16_to_fp32_dtype_0, x = reshape_239_cast_fp16)[name = tensor("cast_435")]; + tensor var_9820 = cast(dtype = var_9820_cast_fp16_to_fp32_dtype_0, x = var_9820_cast_fp16)[name = tensor("cast_439")]; + tensor new_cache_45_internal_tensor_assign_2 = cast(dtype = reshape_229_cast_fp16_to_fp32_dtype_0, x = reshape_229_cast_fp16)[name = tensor("cast_440")]; + tensor var_9401 = cast(dtype = var_9401_cast_fp16_to_fp32_dtype_0, x = var_9401_cast_fp16)[name = tensor("cast_444")]; + tensor new_cache_43_internal_tensor_assign_2 = cast(dtype = reshape_219_cast_fp16_to_fp32_dtype_0, x = reshape_219_cast_fp16)[name = tensor("cast_445")]; + tensor var_8982 = cast(dtype = var_8982_cast_fp16_to_fp32_dtype_0, x = var_8982_cast_fp16)[name = tensor("cast_449")]; + tensor new_cache_41_internal_tensor_assign_2 = cast(dtype = reshape_209_cast_fp16_to_fp32_dtype_0, x = reshape_209_cast_fp16)[name = tensor("cast_450")]; + tensor var_8563 = cast(dtype = var_8563_cast_fp16_to_fp32_dtype_0, x = var_8563_cast_fp16)[name = tensor("cast_454")]; + tensor new_cache_39_internal_tensor_assign_2 = cast(dtype = reshape_199_cast_fp16_to_fp32_dtype_0, x = reshape_199_cast_fp16)[name = tensor("cast_455")]; + tensor var_8144 = cast(dtype = var_8144_cast_fp16_to_fp32_dtype_0, x = var_8144_cast_fp16)[name = tensor("cast_459")]; + tensor new_cache_37_internal_tensor_assign_2 = cast(dtype = reshape_189_cast_fp16_to_fp32_dtype_0, x = reshape_189_cast_fp16)[name = tensor("cast_460")]; + tensor var_7725 = cast(dtype = var_7725_cast_fp16_to_fp32_dtype_0, x = var_7725_cast_fp16)[name = tensor("cast_464")]; + tensor new_cache_35_internal_tensor_assign_2 = cast(dtype = reshape_179_cast_fp16_to_fp32_dtype_0, x = reshape_179_cast_fp16)[name = tensor("cast_465")]; + tensor var_7306 = cast(dtype = var_7306_cast_fp16_to_fp32_dtype_0, x = var_7306_cast_fp16)[name = tensor("cast_469")]; + tensor new_cache_33_internal_tensor_assign_2 = cast(dtype = reshape_169_cast_fp16_to_fp32_dtype_0, x = reshape_169_cast_fp16)[name = tensor("cast_470")]; + tensor var_6887 = cast(dtype = var_6887_cast_fp16_to_fp32_dtype_0, x = var_6887_cast_fp16)[name = tensor("cast_474")]; + tensor new_cache_31_internal_tensor_assign_2 = cast(dtype = reshape_159_cast_fp16_to_fp32_dtype_0, x = reshape_159_cast_fp16)[name = tensor("cast_475")]; + tensor var_6468 = cast(dtype = var_6468_cast_fp16_to_fp32_dtype_0, x = var_6468_cast_fp16)[name = tensor("cast_479")]; + tensor new_cache_29_internal_tensor_assign_2 = cast(dtype = reshape_149_cast_fp16_to_fp32_dtype_0, x = reshape_149_cast_fp16)[name = tensor("cast_480")]; + tensor var_6049 = cast(dtype = var_6049_cast_fp16_to_fp32_dtype_0, x = var_6049_cast_fp16)[name = tensor("cast_484")]; + tensor new_cache_27_internal_tensor_assign_2 = cast(dtype = reshape_139_cast_fp16_to_fp32_dtype_0, x = reshape_139_cast_fp16)[name = tensor("cast_485")]; + tensor var_5630 = cast(dtype = var_5630_cast_fp16_to_fp32_dtype_0, x = var_5630_cast_fp16)[name = tensor("cast_489")]; + tensor new_cache_25_internal_tensor_assign_2 = cast(dtype = reshape_129_cast_fp16_to_fp32_dtype_0, x = reshape_129_cast_fp16)[name = tensor("cast_490")]; + tensor var_5211 = cast(dtype = var_5211_cast_fp16_to_fp32_dtype_0, x = var_5211_cast_fp16)[name = tensor("cast_494")]; + tensor new_cache_23_internal_tensor_assign_2 = cast(dtype = reshape_119_cast_fp16_to_fp32_dtype_0, x = reshape_119_cast_fp16)[name = tensor("cast_495")]; + tensor var_4792 = cast(dtype = var_4792_cast_fp16_to_fp32_dtype_0, x = var_4792_cast_fp16)[name = tensor("cast_499")]; + tensor new_cache_21_internal_tensor_assign_2 = cast(dtype = reshape_109_cast_fp16_to_fp32_dtype_0, x = reshape_109_cast_fp16)[name = tensor("cast_500")]; + tensor var_4373 = cast(dtype = var_4373_cast_fp16_to_fp32_dtype_0, x = var_4373_cast_fp16)[name = tensor("cast_504")]; + tensor new_cache_19_internal_tensor_assign_2 = cast(dtype = reshape_99_cast_fp16_to_fp32_dtype_0, x = reshape_99_cast_fp16)[name = tensor("cast_505")]; + tensor var_3954 = cast(dtype = var_3954_cast_fp16_to_fp32_dtype_0, x = var_3954_cast_fp16)[name = tensor("cast_509")]; + tensor new_cache_17_internal_tensor_assign_2 = cast(dtype = reshape_89_cast_fp16_to_fp32_dtype_0, x = reshape_89_cast_fp16)[name = tensor("cast_510")]; + tensor var_3535 = cast(dtype = var_3535_cast_fp16_to_fp32_dtype_0, x = var_3535_cast_fp16)[name = tensor("cast_514")]; + tensor new_cache_15_internal_tensor_assign_2 = cast(dtype = reshape_79_cast_fp16_to_fp32_dtype_0, x = reshape_79_cast_fp16)[name = tensor("cast_515")]; + tensor var_3116 = cast(dtype = var_3116_cast_fp16_to_fp32_dtype_0, x = var_3116_cast_fp16)[name = tensor("cast_519")]; + tensor new_cache_13_internal_tensor_assign_2 = cast(dtype = reshape_69_cast_fp16_to_fp32_dtype_0, x = reshape_69_cast_fp16)[name = tensor("cast_520")]; + tensor var_2697 = cast(dtype = var_2697_cast_fp16_to_fp32_dtype_0, x = var_2697_cast_fp16)[name = tensor("cast_524")]; + tensor new_cache_11_internal_tensor_assign_2 = cast(dtype = reshape_59_cast_fp16_to_fp32_dtype_0, x = reshape_59_cast_fp16)[name = tensor("cast_525")]; + tensor var_2278 = cast(dtype = var_2278_cast_fp16_to_fp32_dtype_0, x = var_2278_cast_fp16)[name = tensor("cast_529")]; + tensor new_cache_9_internal_tensor_assign_2 = cast(dtype = reshape_49_cast_fp16_to_fp32_dtype_0, x = reshape_49_cast_fp16)[name = tensor("cast_530")]; + tensor var_1859 = cast(dtype = var_1859_cast_fp16_to_fp32_dtype_0, x = var_1859_cast_fp16)[name = tensor("cast_534")]; + tensor new_cache_7_internal_tensor_assign_2 = cast(dtype = reshape_39_cast_fp16_to_fp32_dtype_0, x = reshape_39_cast_fp16)[name = tensor("cast_535")]; + tensor var_1440 = cast(dtype = var_1440_cast_fp16_to_fp32_dtype_0, x = var_1440_cast_fp16)[name = tensor("cast_539")]; + tensor new_cache_5_internal_tensor_assign_2 = cast(dtype = reshape_29_cast_fp16_to_fp32_dtype_0, x = reshape_29_cast_fp16)[name = tensor("cast_540")]; + tensor var_1021 = cast(dtype = var_1021_cast_fp16_to_fp32_dtype_0, x = var_1021_cast_fp16)[name = tensor("cast_544")]; + tensor new_cache_3_internal_tensor_assign_2 = cast(dtype = reshape_19_cast_fp16_to_fp32_dtype_0, x = reshape_19_cast_fp16)[name = tensor("cast_545")]; + tensor var_602 = cast(dtype = var_602_cast_fp16_to_fp32_dtype_0, x = var_602_cast_fp16)[name = tensor("cast_549")]; + tensor new_cache_1_internal_tensor_assign_2 = cast(dtype = reshape_9_cast_fp16_to_fp32_dtype_0, x = reshape_9_cast_fp16)[name = tensor("cast_550")]; + } -> (input, var_10268, new_cache_1_internal_tensor_assign_2, var_602, new_cache_3_internal_tensor_assign_2, var_1021, new_cache_5_internal_tensor_assign_2, var_1440, new_cache_7_internal_tensor_assign_2, var_1859, new_cache_9_internal_tensor_assign_2, var_2278, new_cache_11_internal_tensor_assign_2, var_2697, new_cache_13_internal_tensor_assign_2, var_3116, new_cache_15_internal_tensor_assign_2, var_3535, new_cache_17_internal_tensor_assign_2, var_3954, new_cache_19_internal_tensor_assign_2, var_4373, new_cache_21_internal_tensor_assign_2, var_4792, new_cache_23_internal_tensor_assign_2, var_5211, new_cache_25_internal_tensor_assign_2, var_5630, new_cache_27_internal_tensor_assign_2, var_6049, new_cache_29_internal_tensor_assign_2, var_6468, new_cache_31_internal_tensor_assign_2, var_6887, new_cache_33_internal_tensor_assign_2, var_7306, new_cache_35_internal_tensor_assign_2, var_7725, new_cache_37_internal_tensor_assign_2, var_8144, new_cache_39_internal_tensor_assign_2, var_8563, new_cache_41_internal_tensor_assign_2, var_8982, new_cache_43_internal_tensor_assign_2, var_9401, new_cache_45_internal_tensor_assign_2, var_9820, new_cache_internal_tensor_assign_2, var_10239); +} \ No newline at end of file diff --git a/v2.1/portuguese_24l/flowlm_stepv2.mlmodelc/weights/weight.bin b/v2.1/portuguese_24l/flowlm_stepv2.mlmodelc/weights/weight.bin new file mode 100644 index 0000000000000000000000000000000000000000..994737a10600007719e324a659df1d8f23547759 --- /dev/null +++ b/v2.1/portuguese_24l/flowlm_stepv2.mlmodelc/weights/weight.bin @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:51ea61f00c307f2636b08f86cb7067881d45beed88e2ebf1558038fa4bf79633 +size 303797056 diff --git a/v2.1/portuguese_24l/flowlm_stepv2.mlpackage/Data/com.apple.CoreML/model.mlmodel b/v2.1/portuguese_24l/flowlm_stepv2.mlpackage/Data/com.apple.CoreML/model.mlmodel new file mode 100644 index 0000000000000000000000000000000000000000..68b78cc1784ca783f76af90d975e07db98f5ca35 --- /dev/null +++ b/v2.1/portuguese_24l/flowlm_stepv2.mlpackage/Data/com.apple.CoreML/model.mlmodel @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:257aa9c21878bc84c0464c46802c73ed36d0b5351a916458a5b9b63534a26793 +size 843748 diff --git a/v2.1/portuguese_24l/flowlm_stepv2.mlpackage/Data/com.apple.CoreML/weights/weight.bin b/v2.1/portuguese_24l/flowlm_stepv2.mlpackage/Data/com.apple.CoreML/weights/weight.bin new file mode 100644 index 0000000000000000000000000000000000000000..994737a10600007719e324a659df1d8f23547759 --- /dev/null +++ b/v2.1/portuguese_24l/flowlm_stepv2.mlpackage/Data/com.apple.CoreML/weights/weight.bin @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:51ea61f00c307f2636b08f86cb7067881d45beed88e2ebf1558038fa4bf79633 +size 303797056 diff --git a/v2.1/portuguese_24l/flowlm_stepv2.mlpackage/Manifest.json b/v2.1/portuguese_24l/flowlm_stepv2.mlpackage/Manifest.json new file mode 100644 index 0000000000000000000000000000000000000000..8f5aa954c898bb40676b1fec6e65c2922ad24369 --- /dev/null +++ b/v2.1/portuguese_24l/flowlm_stepv2.mlpackage/Manifest.json @@ -0,0 +1,18 @@ +{ + "fileFormatVersion": "1.0.0", + "itemInfoEntries": { + "52FF01D9-803D-4A0C-94E4-872815D4D545": { + "author": "com.apple.CoreML", + "description": "CoreML Model Weights", + "name": "weights", + "path": "com.apple.CoreML/weights" + }, + "7C9C4850-66E3-4DC4-AA83-5AA10140EF7C": { + "author": "com.apple.CoreML", + "description": "CoreML Model Specification", + "name": "model.mlmodel", + "path": "com.apple.CoreML/model.mlmodel" + } + }, + "rootModelIdentifier": "7C9C4850-66E3-4DC4-AA83-5AA10140EF7C" +} diff --git a/v2.1/portuguese_24l/manifest.json b/v2.1/portuguese_24l/manifest.json new file mode 100644 index 0000000000000000000000000000000000000000..fc7da98920126d1261a9c47517b1303db1599db7 --- /dev/null +++ b/v2.1/portuguese_24l/manifest.json @@ -0,0 +1,27 @@ +{ + "version": "2.1", + "base": "v2", + "kind": "optimization-reconvert", + "self_contained": true, + "note": "Same weights as v2. Re-converted for speed: fused flow decoder (100% ANE), one-shot cond prefill, fp16 flowlm. NOT a finetune.", + "new": [ + "flow_decoder_fused", + "cond_prefill", + "flowlm_step" + ], + "copied_from_v2": [ + "flowlm_stepv2.mlpackage", + "flowlm_stepv2.mlmodelc", + "mimi_decoder.mlpackage", + "mimi_decoder.mlmodelc", + "constants", + "constants_bin" + ], + "compute_units": { + "flow_decoder_fused": "all (ANE)", + "cond_prefill": "all (GPU)", + "flowlm_step": "all (GPU, fp16)", + "flowlm_stepv2": "cpuAndGpu (GPU, int8 \u2014 fastest flowlm)", + "mimi_decoder": "cpuOnly" + } +} diff --git a/v2.1/portuguese_24l/mimi_decoder.mlmodelc/analytics/coremldata.bin b/v2.1/portuguese_24l/mimi_decoder.mlmodelc/analytics/coremldata.bin new file mode 100644 index 0000000000000000000000000000000000000000..d82743c83d2238a1bc1bedd6de452a40a70662f3 --- /dev/null +++ b/v2.1/portuguese_24l/mimi_decoder.mlmodelc/analytics/coremldata.bin @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:94b7d63dfaed4d55e872da6de411c182be7be8bd554973c14e0c5d85fa8643eb +size 243 diff --git a/v2.1/portuguese_24l/mimi_decoder.mlmodelc/coremldata.bin b/v2.1/portuguese_24l/mimi_decoder.mlmodelc/coremldata.bin new file mode 100644 index 0000000000000000000000000000000000000000..c7fd734fd5accd542824b3f38a6d17b27171b809 --- /dev/null +++ b/v2.1/portuguese_24l/mimi_decoder.mlmodelc/coremldata.bin @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:f282b1bdc12dc30f18e85cad18dcb20a816810671efdefb6697dad31c5033945 +size 1777 diff --git a/v2.1/portuguese_24l/mimi_decoder.mlmodelc/metadata.json b/v2.1/portuguese_24l/mimi_decoder.mlmodelc/metadata.json new file mode 100644 index 0000000000000000000000000000000000000000..f66f9e7ff5be2ffbc44c8300f3501f67a77afaf6 --- /dev/null +++ b/v2.1/portuguese_24l/mimi_decoder.mlmodelc/metadata.json @@ -0,0 +1,573 @@ +[ + { + "metadataOutputVersion" : "3.0", + "storagePrecision" : "Float32", + "outputSchema" : [ + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1 × 1 × 1920)", + "shortDescription" : "", + "shape" : "[1, 1, 1920]", + "name" : "var_724", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1 × 512 × 16)", + "shortDescription" : "", + "shape" : "[1, 512, 16]", + "name" : "var_77", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 2 × 1 × 256 × 8 × 64)", + "shortDescription" : "", + "shape" : "[2, 1, 256, 8, 64]", + "name" : "var_210", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1)", + "shortDescription" : "", + "shape" : "[1]", + "name" : "var_741", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 2 × 1 × 256 × 8 × 64)", + "shortDescription" : "", + "shape" : "[2, 1, 256, 8, 64]", + "name" : "var_400", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1)", + "shortDescription" : "", + "shape" : "[1]", + "name" : "var_744", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1 × 512 × 6)", + "shortDescription" : "", + "shape" : "[1, 512, 6]", + "name" : "var_542", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1)", + "shortDescription" : "", + "shape" : "[1]", + "name" : "conv0_first", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1 × 256 × 6)", + "shortDescription" : "", + "shape" : "[1, 256, 6]", + "name" : "var_565", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1 × 256 × 2)", + "shortDescription" : "", + "shape" : "[1, 256, 2]", + "name" : "var_585", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1)", + "shortDescription" : "", + "shape" : "[1]", + "name" : "res0_conv0_first", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1 × 128 × 0)", + "shortDescription" : "", + "shape" : "[1, 128, 0]", + "name" : "res0_conv1_prev", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1)", + "shortDescription" : "", + "shape" : "[1]", + "name" : "res0_conv1_first", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1 × 128 × 5)", + "shortDescription" : "", + "shape" : "[1, 128, 5]", + "name" : "var_621", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1 × 128 × 2)", + "shortDescription" : "", + "shape" : "[1, 128, 2]", + "name" : "var_641", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1)", + "shortDescription" : "", + "shape" : "[1]", + "name" : "res1_conv0_first", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1 × 64 × 0)", + "shortDescription" : "", + "shape" : "[1, 64, 0]", + "name" : "res1_conv1_prev", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1)", + "shortDescription" : "", + "shape" : "[1]", + "name" : "res1_conv1_first", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1 × 64 × 4)", + "shortDescription" : "", + "shape" : "[1, 64, 4]", + "name" : "var_677", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1 × 64 × 2)", + "shortDescription" : "", + "shape" : "[1, 64, 2]", + "name" : "var_697", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1)", + "shortDescription" : "", + "shape" : "[1]", + "name" : "res2_conv0_first", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1 × 32 × 0)", + "shortDescription" : "", + "shape" : "[1, 32, 0]", + "name" : "res2_conv1_prev", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1)", + "shortDescription" : "", + "shape" : "[1]", + "name" : "res2_conv1_first", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1 × 64 × 2)", + "shortDescription" : "", + "shape" : "[1, 64, 2]", + "name" : "var_725", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1)", + "shortDescription" : "", + "shape" : "[1]", + "name" : "conv_final_first", + "type" : "MultiArray" + } + ], + "modelParameters" : [ + + ], + "specificationVersion" : 8, + "mlProgramOperationTypeHistogram" : { + "Ios17.notEqual" : 4, + "Tile" : 2, + "Ios17.lessEqual" : 2, + "Ios17.layerNorm" : 4, + "Stack" : 6, + "Ios17.greaterEqual" : 4, + "Ios17.scatterAlongAxis" : 4, + "Ios17.sub" : 17, + "Identity" : 11, + "Ios17.elu" : 10, + "Ios17.logicalNot" : 2, + "Ios17.sliceByIndex" : 35, + "Ios17.concat" : 9, + "Ios17.convTranspose" : 4, + "Ios17.floorDiv" : 4, + "Ios17.expandDims" : 7, + "Ios17.less" : 2, + "Ios17.mul" : 29, + "Ios17.transpose" : 10, + "Ios17.matmul" : 4, + "Ios16.gelu" : 2, + "Ios17.conv" : 9, + "Ios17.reshape" : 22, + "Split" : 2, + "Ios17.squeeze" : 6, + "Ios17.cast" : 6, + "Ios17.sin" : 2, + "Ios17.linear" : 8, + "Select" : 6, + "Ios16.softmax" : 2, + "Ios17.add" : 28, + "Ios17.cos" : 2, + "Ios17.logicalAnd" : 6 + }, + "computePrecision" : "Mixed (Float32, Int32)", + "isUpdatable" : "0", + "stateSchema" : [ + + ], + "availability" : { + "macOS" : "14.0", + "tvOS" : "17.0", + "visionOS" : "1.0", + "watchOS" : "10.0", + "iOS" : "17.0", + "macCatalyst" : "17.0" + }, + "modelType" : { + "name" : "MLModelType_mlProgram" + }, + "userDefinedMetadata" : { + "com.github.apple.coremltools.conversion_date" : "2026-04-24", + "com.github.apple.coremltools.source" : "torch==2.9.1", + "com.github.apple.coremltools.version" : "9.0", + "com.github.apple.coremltools.source_dialect" : "TorchScript" + }, + "inputSchema" : [ + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1 × 32)", + "shortDescription" : "", + "shape" : "[1, 32]", + "name" : "latent", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1 × 512 × 16)", + "shortDescription" : "", + "shape" : "[1, 512, 16]", + "name" : "upsample_partial", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 2 × 1 × 256 × 8 × 64)", + "shortDescription" : "", + "shape" : "[2, 1, 256, 8, 64]", + "name" : "attn0_cache", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1)", + "shortDescription" : "", + "shape" : "[1]", + "name" : "attn0_offset", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 2 × 1 × 256 × 8 × 64)", + "shortDescription" : "", + "shape" : "[2, 1, 256, 8, 64]", + "name" : "attn1_cache", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1)", + "shortDescription" : "", + "shape" : "[1]", + "name" : "attn1_offset", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1 × 512 × 6)", + "shortDescription" : "", + "shape" : "[1, 512, 6]", + "name" : "conv0_prev", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1)", + "shortDescription" : "", + "shape" : "[1]", + "name" : "conv0_first", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1 × 256 × 6)", + "shortDescription" : "", + "shape" : "[1, 256, 6]", + "name" : "convtr0_partial", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1 × 256 × 2)", + "shortDescription" : "", + "shape" : "[1, 256, 2]", + "name" : "res0_conv0_prev", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1)", + "shortDescription" : "", + "shape" : "[1]", + "name" : "res0_conv0_first", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1 × 128 × 0)", + "shortDescription" : "", + "shape" : "[1, 128, 0]", + "name" : "res0_conv1_prev", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1)", + "shortDescription" : "", + "shape" : "[1]", + "name" : "res0_conv1_first", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1 × 128 × 5)", + "shortDescription" : "", + "shape" : "[1, 128, 5]", + "name" : "convtr1_partial", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1 × 128 × 2)", + "shortDescription" : "", + "shape" : "[1, 128, 2]", + "name" : "res1_conv0_prev", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1)", + "shortDescription" : "", + "shape" : "[1]", + "name" : "res1_conv0_first", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1 × 64 × 0)", + "shortDescription" : "", + "shape" : "[1, 64, 0]", + "name" : "res1_conv1_prev", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1)", + "shortDescription" : "", + "shape" : "[1]", + "name" : "res1_conv1_first", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1 × 64 × 4)", + "shortDescription" : "", + "shape" : "[1, 64, 4]", + "name" : "convtr2_partial", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1 × 64 × 2)", + "shortDescription" : "", + "shape" : "[1, 64, 2]", + "name" : "res2_conv0_prev", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1)", + "shortDescription" : "", + "shape" : "[1]", + "name" : "res2_conv0_first", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1 × 32 × 0)", + "shortDescription" : "", + "shape" : "[1, 32, 0]", + "name" : "res2_conv1_prev", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1)", + "shortDescription" : "", + "shape" : "[1]", + "name" : "res2_conv1_first", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1 × 64 × 2)", + "shortDescription" : "", + "shape" : "[1, 64, 2]", + "name" : "conv_final_prev", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1)", + "shortDescription" : "", + "shape" : "[1]", + "name" : "conv_final_first", + "type" : "MultiArray" + } + ], + "generatedClassName" : "mimi_decoder", + "method" : "predict" + } +] \ No newline at end of file diff --git a/v2.1/portuguese_24l/mimi_decoder.mlmodelc/model.mil b/v2.1/portuguese_24l/mimi_decoder.mlmodelc/model.mil new file mode 100644 index 0000000000000000000000000000000000000000..54d22260d0813157c84e216bfdbd2ff3fd7fcaf6 --- /dev/null +++ b/v2.1/portuguese_24l/mimi_decoder.mlmodelc/model.mil @@ -0,0 +1,646 @@ +program(1.0) +[buildInfo = dict, tensor>({{"coremlc-component-MIL", "3520.4.1"}, {"coremlc-version", "3520.5.1"}, {"coremltools-component-torch", "2.9.1"}, {"coremltools-source-dialect", "TorchScript"}, {"coremltools-version", "9.0"}})] +{ + func main(tensor attn0_cache, tensor attn0_offset, tensor attn1_cache, tensor attn1_offset, tensor conv0_first, tensor conv0_prev, tensor conv_final_first, tensor conv_final_prev, tensor convtr0_partial, tensor convtr1_partial, tensor convtr2_partial, tensor latent, tensor res0_conv0_first, tensor res0_conv0_prev, tensor res0_conv1_first, tensor res0_conv1_prev, tensor res1_conv0_first, tensor res1_conv0_prev, tensor res1_conv1_first, tensor res1_conv1_prev, tensor res2_conv0_first, tensor res2_conv0_prev, tensor res2_conv1_first, tensor res2_conv1_prev, tensor upsample_partial) { + tensor emb_mean = const()[name = tensor("emb_mean"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(64)))]; + tensor emb_std = const()[name = tensor("emb_std"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(256)))]; + tensor mimi_quantizer_output_proj_weight = const()[name = tensor("mimi_quantizer_output_proj_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(448)))]; + tensor mimi_upsample_convtr_convtr_weight = const()[name = tensor("mimi_upsample_convtr_convtr_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(66048)))]; + tensor mimi_decoder_transformer_transformer_layers_0_norm1_bias = const()[name = tensor("mimi_decoder_transformer_transformer_layers_0_norm1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(131648)))]; + tensor mimi_decoder_transformer_transformer_layers_0_norm1_weight = const()[name = tensor("mimi_decoder_transformer_transformer_layers_0_norm1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(133760)))]; + tensor mimi_decoder_transformer_transformer_layers_0_self_attn_in_proj_weight = const()[name = tensor("mimi_decoder_transformer_transformer_layers_0_self_attn_in_proj_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(135872)))]; + tensor mimi_decoder_transformer_transformer_layers_0_self_attn_out_proj_weight = const()[name = tensor("mimi_decoder_transformer_transformer_layers_0_self_attn_out_proj_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(3281664)))]; + tensor mimi_decoder_transformer_transformer_layers_0_layer_scale_1_scale = const()[name = tensor("mimi_decoder_transformer_transformer_layers_0_layer_scale_1_scale"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(4330304)))]; + tensor mimi_decoder_transformer_transformer_layers_0_norm2_bias = const()[name = tensor("mimi_decoder_transformer_transformer_layers_0_norm2_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(4332416)))]; + tensor mimi_decoder_transformer_transformer_layers_0_norm2_weight = const()[name = tensor("mimi_decoder_transformer_transformer_layers_0_norm2_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(4334528)))]; + tensor mimi_decoder_transformer_transformer_layers_0_linear1_weight = const()[name = tensor("mimi_decoder_transformer_transformer_layers_0_linear1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(4336640)))]; + tensor mimi_decoder_transformer_transformer_layers_0_linear2_weight = const()[name = tensor("mimi_decoder_transformer_transformer_layers_0_linear2_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8531008)))]; + tensor mimi_decoder_transformer_transformer_layers_0_layer_scale_2_scale = const()[name = tensor("mimi_decoder_transformer_transformer_layers_0_layer_scale_2_scale"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(12725376)))]; + tensor mimi_decoder_transformer_transformer_layers_1_norm1_bias = const()[name = tensor("mimi_decoder_transformer_transformer_layers_1_norm1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(12727488)))]; + tensor mimi_decoder_transformer_transformer_layers_1_norm1_weight = const()[name = tensor("mimi_decoder_transformer_transformer_layers_1_norm1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(12729600)))]; + tensor mimi_decoder_transformer_transformer_layers_1_self_attn_in_proj_weight = const()[name = tensor("mimi_decoder_transformer_transformer_layers_1_self_attn_in_proj_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(12731712)))]; + tensor mimi_decoder_transformer_transformer_layers_1_self_attn_out_proj_weight = const()[name = tensor("mimi_decoder_transformer_transformer_layers_1_self_attn_out_proj_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15877504)))]; + tensor mimi_decoder_transformer_transformer_layers_1_layer_scale_1_scale = const()[name = tensor("mimi_decoder_transformer_transformer_layers_1_layer_scale_1_scale"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(16926144)))]; + tensor mimi_decoder_transformer_transformer_layers_1_norm2_bias = const()[name = tensor("mimi_decoder_transformer_transformer_layers_1_norm2_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(16928256)))]; + tensor mimi_decoder_transformer_transformer_layers_1_norm2_weight = const()[name = tensor("mimi_decoder_transformer_transformer_layers_1_norm2_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(16930368)))]; + tensor mimi_decoder_transformer_transformer_layers_1_linear1_weight = const()[name = tensor("mimi_decoder_transformer_transformer_layers_1_linear1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(16932480)))]; + tensor mimi_decoder_transformer_transformer_layers_1_linear2_weight = const()[name = tensor("mimi_decoder_transformer_transformer_layers_1_linear2_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21126848)))]; + tensor mimi_decoder_transformer_transformer_layers_1_layer_scale_2_scale = const()[name = tensor("mimi_decoder_transformer_transformer_layers_1_layer_scale_2_scale"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(25321216)))]; + tensor mimi_decoder_model_0_conv_bias = const()[name = tensor("mimi_decoder_model_0_conv_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(25323328)))]; + tensor mimi_decoder_model_0_conv_weight = const()[name = tensor("mimi_decoder_model_0_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(25325440)))]; + tensor mimi_decoder_model_2_convtr_bias = const()[name = tensor("mimi_decoder_model_2_convtr_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(32665536)))]; + tensor mimi_decoder_model_2_convtr_weight = const()[name = tensor("mimi_decoder_model_2_convtr_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(32666624)))]; + tensor mimi_decoder_model_3_block_1_conv_bias = const()[name = tensor("mimi_decoder_model_3_block_1_conv_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(38958144)))]; + tensor mimi_decoder_model_3_block_1_conv_weight = const()[name = tensor("mimi_decoder_model_3_block_1_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(38958720)))]; + tensor mimi_decoder_model_3_block_3_conv_bias = const()[name = tensor("mimi_decoder_model_3_block_3_conv_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(39352000)))]; + tensor mimi_decoder_model_3_block_3_conv_weight = const()[name = tensor("mimi_decoder_model_3_block_3_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(39353088)))]; + tensor mimi_decoder_model_5_convtr_bias = const()[name = tensor("mimi_decoder_model_5_convtr_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(39484224)))]; + tensor mimi_decoder_model_5_convtr_weight = const()[name = tensor("mimi_decoder_model_5_convtr_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(39484800)))]; + tensor mimi_decoder_model_6_block_1_conv_bias = const()[name = tensor("mimi_decoder_model_6_block_1_conv_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(40795584)))]; + tensor mimi_decoder_model_6_block_1_conv_weight = const()[name = tensor("mimi_decoder_model_6_block_1_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(40795904)))]; + tensor mimi_decoder_model_6_block_3_conv_bias = const()[name = tensor("mimi_decoder_model_6_block_3_conv_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(40894272)))]; + tensor mimi_decoder_model_6_block_3_conv_weight = const()[name = tensor("mimi_decoder_model_6_block_3_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(40894848)))]; + tensor mimi_decoder_model_8_convtr_bias = const()[name = tensor("mimi_decoder_model_8_convtr_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(40927680)))]; + tensor mimi_decoder_model_8_convtr_weight = const()[name = tensor("mimi_decoder_model_8_convtr_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(40928000)))]; + tensor mimi_decoder_model_9_block_1_conv_bias = const()[name = tensor("mimi_decoder_model_9_block_1_conv_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(41190208)))]; + tensor mimi_decoder_model_9_block_1_conv_weight = const()[name = tensor("mimi_decoder_model_9_block_1_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(41190400)))]; + tensor mimi_decoder_model_9_block_3_conv_bias = const()[name = tensor("mimi_decoder_model_9_block_3_conv_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(41215040)))]; + tensor mimi_decoder_model_9_block_3_conv_weight = const()[name = tensor("mimi_decoder_model_9_block_3_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(41215360)))]; + tensor mimi_decoder_model_11_conv_bias = const()[name = tensor("mimi_decoder_model_11_conv_bias"), val = tensor([0x1.28p-11])]; + tensor mimi_decoder_model_11_conv_weight = const()[name = tensor("mimi_decoder_model_11_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(41223616)))]; + tensor var_38 = mul(x = latent, y = emb_std)[name = tensor("op_38")]; + tensor denorm = add(x = var_38, y = emb_mean)[name = tensor("denorm")]; + tensor input_1_axes_0 = const()[name = tensor("input_1_axes_0"), val = tensor([-1])]; + tensor input_1 = expand_dims(axes = input_1_axes_0, x = denorm)[name = tensor("input_1")]; + tensor x_1_pad_type_0 = const()[name = tensor("x_1_pad_type_0"), val = tensor("valid")]; + tensor x_1_strides_0 = const()[name = tensor("x_1_strides_0"), val = tensor([1])]; + tensor x_1_pad_0 = const()[name = tensor("x_1_pad_0"), val = tensor([0, 0])]; + tensor x_1_dilations_0 = const()[name = tensor("x_1_dilations_0"), val = tensor([1])]; + tensor x_1_groups_0 = const()[name = tensor("x_1_groups_0"), val = tensor(1)]; + tensor x_1 = conv(dilations = x_1_dilations_0, groups = x_1_groups_0, pad = x_1_pad_0, pad_type = x_1_pad_type_0, strides = x_1_strides_0, weight = mimi_quantizer_output_proj_weight, x = input_1)[name = tensor("x_1")]; + tensor var_62 = const()[name = tensor("op_62"), val = tensor(-1)]; + tensor y_1_pad_type_0 = const()[name = tensor("y_1_pad_type_0"), val = tensor("valid")]; + tensor y_1_strides_0 = const()[name = tensor("y_1_strides_0"), val = tensor([16])]; + tensor y_1_groups_0 = const()[name = tensor("y_1_groups_0"), val = tensor(512)]; + tensor y_1_pad_0 = const()[name = tensor("y_1_pad_0"), val = tensor([0, 0])]; + tensor y_1_dilations_0 = const()[name = tensor("y_1_dilations_0"), val = tensor([1])]; + tensor y_1_has_output_shape_output_shape_0 = const()[name = tensor("y_1_has_output_shape_output_shape_0"), val = tensor([1, 512, 32])]; + tensor y_1_has_output_shape = conv_transpose(dilations = y_1_dilations_0, groups = y_1_groups_0, output_shape = y_1_has_output_shape_output_shape_0, pad = y_1_pad_0, pad_type = y_1_pad_type_0, strides = y_1_strides_0, weight = mimi_upsample_convtr_convtr_weight, x = x_1)[name = tensor("y_1_has_output_shape")]; + tensor var_72_begin_0 = const()[name = tensor("op_72_begin_0"), val = tensor([0, 0, 0])]; + tensor var_72_end_0 = const()[name = tensor("op_72_end_0"), val = tensor([1, 512, 16])]; + tensor var_72_end_mask_0 = const()[name = tensor("op_72_end_mask_0"), val = tensor([true, true, false])]; + tensor var_72 = slice_by_index(begin = var_72_begin_0, end = var_72_end_0, end_mask = var_72_end_mask_0, x = y_1_has_output_shape)[name = tensor("op_72")]; + tensor var_73 = add(x = var_72, y = upsample_partial)[name = tensor("op_73")]; + tensor var_74_begin_0 = const()[name = tensor("op_74_begin_0"), val = tensor([0, 0, 16])]; + tensor var_74_end_0 = const()[name = tensor("op_74_end_0"), val = tensor([1, 512, 32])]; + tensor var_74_end_mask_0 = const()[name = tensor("op_74_end_mask_0"), val = tensor([true, true, true])]; + tensor var_74 = slice_by_index(begin = var_74_begin_0, end = var_74_end_0, end_mask = var_74_end_mask_0, x = y_1_has_output_shape)[name = tensor("op_74")]; + tensor y_3_interleave_0 = const()[name = tensor("y_3_interleave_0"), val = tensor(false)]; + tensor y_3 = concat(axis = var_62, interleave = y_3_interleave_0, values = (var_73, var_74))[name = tensor("y_3")]; + tensor var_77_begin_0 = const()[name = tensor("op_77_begin_0"), val = tensor([0, 0, 16])]; + tensor var_77_end_0 = const()[name = tensor("op_77_end_0"), val = tensor([1, 512, 32])]; + tensor var_77_end_mask_0 = const()[name = tensor("op_77_end_mask_0"), val = tensor([true, true, true])]; + tensor var_77 = slice_by_index(begin = var_77_begin_0, end = var_77_end_0, end_mask = var_77_end_mask_0, x = y_3)[name = tensor("op_77")]; + tensor x_3_begin_0 = const()[name = tensor("x_3_begin_0"), val = tensor([0, 0, 0])]; + tensor x_3_end_0 = const()[name = tensor("x_3_end_0"), val = tensor([1, 512, 16])]; + tensor x_3_end_mask_0 = const()[name = tensor("x_3_end_mask_0"), val = tensor([true, true, false])]; + tensor x_3 = slice_by_index(begin = x_3_begin_0, end = x_3_end_0, end_mask = x_3_end_mask_0, x = y_3)[name = tensor("x_3")]; + tensor var_86 = const()[name = tensor("op_86"), val = tensor(0)]; + tensor var_91 = const()[name = tensor("op_91"), val = tensor(-1)]; + tensor var_100 = const()[name = tensor("op_100"), val = tensor(-0x1.ff933cp+127)]; + tensor var_102 = const()[name = tensor("op_102"), val = tensor(0x1.4f8b58p-17)]; + tensor input_3_perm_0 = const()[name = tensor("input_3_perm_0"), val = tensor([0, 2, 1])]; + tensor query_1_axes_0 = const()[name = tensor("query_1_axes_0"), val = tensor([-1])]; + tensor input_3 = transpose(perm = input_3_perm_0, x = x_3)[name = tensor("transpose_19")]; + tensor query_1 = layer_norm(axes = query_1_axes_0, beta = mimi_decoder_transformer_transformer_layers_0_norm1_bias, epsilon = var_102, gamma = mimi_decoder_transformer_transformer_layers_0_norm1_weight, x = input_3)[name = tensor("query_1")]; + tensor linear_0_bias_0 = const()[name = tensor("linear_0_bias_0"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(41224448)))]; + tensor projected_1 = linear(bias = linear_0_bias_0, weight = mimi_decoder_transformer_transformer_layers_0_self_attn_in_proj_weight, x = query_1)[name = tensor("linear_0")]; + tensor var_130 = const()[name = tensor("op_130"), val = tensor([1, 16, 3, 8, 64])]; + tensor packed_1 = reshape(shape = var_130, x = projected_1)[name = tensor("packed_1")]; + tensor var_132_split_sizes_0 = const()[name = tensor("op_132_split_sizes_0"), val = tensor([1, 1, 1])]; + tensor var_132_axis_0 = const()[name = tensor("op_132_axis_0"), val = tensor(2)]; + tensor var_132_0, tensor var_132_1, tensor var_132_2 = split(axis = var_132_axis_0, split_sizes = var_132_split_sizes_0, x = packed_1)[name = tensor("op_132")]; + tensor squeeze_0_axes_0 = const()[name = tensor("squeeze_0_axes_0"), val = tensor([2])]; + tensor squeeze_0 = squeeze(axes = squeeze_0_axes_0, x = var_132_0)[name = tensor("squeeze_0")]; + tensor squeeze_1_axes_0 = const()[name = tensor("squeeze_1_axes_0"), val = tensor([2])]; + tensor squeeze_1 = squeeze(axes = squeeze_1_axes_0, x = var_132_1)[name = tensor("squeeze_1")]; + tensor squeeze_2_axes_0 = const()[name = tensor("squeeze_2_axes_0"), val = tensor([2])]; + tensor squeeze_2 = squeeze(axes = squeeze_2_axes_0, x = var_132_2)[name = tensor("squeeze_2")]; + tensor offset_3_begin_0 = const()[name = tensor("offset_3_begin_0"), val = tensor([0])]; + tensor offset_3_end_0 = const()[name = tensor("offset_3_end_0"), val = tensor([1])]; + tensor offset_3_end_mask_0 = const()[name = tensor("offset_3_end_mask_0"), val = tensor([false])]; + tensor offset_3_squeeze_mask_0 = const()[name = tensor("offset_3_squeeze_mask_0"), val = tensor([true])]; + tensor offset_3 = slice_by_index(begin = offset_3_begin_0, end = offset_3_end_0, end_mask = offset_3_end_mask_0, squeeze_mask = offset_3_squeeze_mask_0, x = attn0_offset)[name = tensor("offset_3")]; + tensor freqs_1 = const()[name = tensor("freqs_1"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(41230656)))]; + tensor ts_1_promoted = const()[name = tensor("ts_1_promoted"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(41230848)))]; + tensor ts_3 = add(x = ts_1_promoted, y = offset_3)[name = tensor("ts_3")]; + tensor var_148 = const()[name = tensor("op_148"), val = tensor([-1, 1, 1])]; + tensor ts_5 = reshape(shape = var_148, x = ts_3)[name = tensor("ts_5")]; + tensor var_150 = const()[name = tensor("op_150"), val = tensor([1, 16, 8, 32, 2])]; + tensor q_3 = reshape(shape = var_150, x = squeeze_0)[name = tensor("q_3")]; + tensor var_152 = const()[name = tensor("op_152"), val = tensor([1, 16, 8, 32, 2])]; + tensor k_3 = reshape(shape = var_152, x = squeeze_1)[name = tensor("k_3")]; + tensor var_154_begin_0 = const()[name = tensor("op_154_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_154_end_0 = const()[name = tensor("op_154_end_0"), val = tensor([1, 16, 8, 32, 1])]; + tensor var_154_end_mask_0 = const()[name = tensor("op_154_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_154_squeeze_mask_0 = const()[name = tensor("op_154_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_154 = slice_by_index(begin = var_154_begin_0, end = var_154_end_0, end_mask = var_154_end_mask_0, squeeze_mask = var_154_squeeze_mask_0, x = q_3)[name = tensor("op_154")]; + tensor var_156_begin_0 = const()[name = tensor("op_156_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_156_end_0 = const()[name = tensor("op_156_end_0"), val = tensor([1, 16, 8, 32, 2])]; + tensor var_156_end_mask_0 = const()[name = tensor("op_156_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_156_squeeze_mask_0 = const()[name = tensor("op_156_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_156 = slice_by_index(begin = var_156_begin_0, end = var_156_end_0, end_mask = var_156_end_mask_0, squeeze_mask = var_156_squeeze_mask_0, x = q_3)[name = tensor("op_156")]; + tensor var_158_begin_0 = const()[name = tensor("op_158_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_158_end_0 = const()[name = tensor("op_158_end_0"), val = tensor([1, 16, 8, 32, 1])]; + tensor var_158_end_mask_0 = const()[name = tensor("op_158_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_158_squeeze_mask_0 = const()[name = tensor("op_158_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_158 = slice_by_index(begin = var_158_begin_0, end = var_158_end_0, end_mask = var_158_end_mask_0, squeeze_mask = var_158_squeeze_mask_0, x = k_3)[name = tensor("op_158")]; + tensor var_160_begin_0 = const()[name = tensor("op_160_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_160_end_0 = const()[name = tensor("op_160_end_0"), val = tensor([1, 16, 8, 32, 2])]; + tensor var_160_end_mask_0 = const()[name = tensor("op_160_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_160_squeeze_mask_0 = const()[name = tensor("op_160_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_160 = slice_by_index(begin = var_160_begin_0, end = var_160_end_0, end_mask = var_160_end_mask_0, squeeze_mask = var_160_squeeze_mask_0, x = k_3)[name = tensor("op_160")]; + tensor var_162 = mul(x = freqs_1, y = ts_5)[name = tensor("op_162")]; + tensor rotr_1 = cos(x = var_162)[name = tensor("rotr_1")]; + tensor roti_1 = sin(x = var_162)[name = tensor("roti_1")]; + tensor var_166 = mul(x = var_154, y = rotr_1)[name = tensor("op_166")]; + tensor var_167 = mul(x = var_156, y = roti_1)[name = tensor("op_167")]; + tensor qor_1 = sub(x = var_166, y = var_167)[name = tensor("qor_1")]; + tensor var_169 = mul(x = var_154, y = roti_1)[name = tensor("op_169")]; + tensor var_170 = mul(x = var_156, y = rotr_1)[name = tensor("op_170")]; + tensor qoi_1 = add(x = var_169, y = var_170)[name = tensor("qoi_1")]; + tensor var_172 = mul(x = var_158, y = rotr_1)[name = tensor("op_172")]; + tensor var_173 = mul(x = var_160, y = roti_1)[name = tensor("op_173")]; + tensor kor_1 = sub(x = var_172, y = var_173)[name = tensor("kor_1")]; + tensor var_175 = mul(x = var_158, y = roti_1)[name = tensor("op_175")]; + tensor var_176 = mul(x = var_160, y = rotr_1)[name = tensor("op_176")]; + tensor koi_1 = add(x = var_175, y = var_176)[name = tensor("koi_1")]; + tensor qo_1_axis_0 = const()[name = tensor("qo_1_axis_0"), val = tensor(-1)]; + tensor qo_1 = stack(axis = qo_1_axis_0, values = (qor_1, qoi_1))[name = tensor("qo_1")]; + tensor ko_1_axis_0 = const()[name = tensor("ko_1_axis_0"), val = tensor(-1)]; + tensor ko_1 = stack(axis = ko_1_axis_0, values = (kor_1, koi_1))[name = tensor("ko_1")]; + tensor var_186 = const()[name = tensor("op_186"), val = tensor([1, 16, 8, 64])]; + tensor q_5 = reshape(shape = var_186, x = qo_1)[name = tensor("q_5")]; + tensor var_188 = const()[name = tensor("op_188"), val = tensor([1, 16, 8, 64])]; + tensor k_5 = reshape(shape = var_188, x = ko_1)[name = tensor("k_5")]; + tensor capacity_1 = const()[name = tensor("capacity_1"), val = tensor([256])]; + tensor var_193_dtype_0 = const()[name = tensor("op_193_dtype_0"), val = tensor("int32")]; + tensor var_194 = const()[name = tensor("op_194"), val = tensor([1, 1])]; + tensor var_193 = cast(dtype = var_193_dtype_0, x = attn0_offset)[name = tensor("cast_49")]; + tensor write_base_1 = reshape(shape = var_194, x = var_193)[name = tensor("write_base_1")]; + tensor write_range_1 = const()[name = tensor("write_range_1"), val = tensor([[0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15]])]; + tensor abs_idx_1 = add(x = write_base_1, y = write_range_1)[name = tensor("abs_idx_1")]; + tensor wrapped_1_div = floor_div(x = abs_idx_1, y = capacity_1)[name = tensor("wrapped_1_div")]; + tensor wrapped_1_div_scaled = mul(x = wrapped_1_div, y = capacity_1)[name = tensor("wrapped_1_div_scaled")]; + tensor wrapped_1 = sub(x = abs_idx_1, y = wrapped_1_div_scaled)[name = tensor("wrapped_1")]; + tensor var_201 = const()[name = tensor("op_201"), val = tensor([1, 16, 1, 1])]; + tensor var_202 = reshape(shape = var_201, x = wrapped_1)[name = tensor("op_202")]; + tensor write_indexes_1_reps_0 = const()[name = tensor("write_indexes_1_reps_0"), val = tensor([1, 1, 8, 64])]; + tensor write_indexes_1 = tile(reps = write_indexes_1_reps_0, x = var_202)[name = tensor("write_indexes_1")]; + tensor var_205_begin_0 = const()[name = tensor("op_205_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_205_end_0 = const()[name = tensor("op_205_end_0"), val = tensor([1, 1, 256, 8, 64])]; + tensor var_205_end_mask_0 = const()[name = tensor("op_205_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_205_squeeze_mask_0 = const()[name = tensor("op_205_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor var_205 = slice_by_index(begin = var_205_begin_0, end = var_205_end_0, end_mask = var_205_end_mask_0, squeeze_mask = var_205_squeeze_mask_0, x = attn0_cache)[name = tensor("op_205")]; + tensor new_k_cache_1_axis_0 = const()[name = tensor("new_k_cache_1_axis_0"), val = tensor(1)]; + tensor new_k_cache_1_mode_0 = const()[name = tensor("new_k_cache_1_mode_0"), val = tensor("update")]; + tensor new_k_cache_1_validate_indices_0 = const()[name = tensor("new_k_cache_1_validate_indices_0"), val = tensor(false)]; + tensor new_k_cache_1 = scatter_along_axis(axis = new_k_cache_1_axis_0, data = var_205, indices = write_indexes_1, mode = new_k_cache_1_mode_0, updates = k_5, validate_indices = new_k_cache_1_validate_indices_0)[name = tensor("new_k_cache_1")]; + tensor var_207_begin_0 = const()[name = tensor("op_207_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor var_207_end_0 = const()[name = tensor("op_207_end_0"), val = tensor([2, 1, 256, 8, 64])]; + tensor var_207_end_mask_0 = const()[name = tensor("op_207_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_207_squeeze_mask_0 = const()[name = tensor("op_207_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor var_207 = slice_by_index(begin = var_207_begin_0, end = var_207_end_0, end_mask = var_207_end_mask_0, squeeze_mask = var_207_squeeze_mask_0, x = attn0_cache)[name = tensor("op_207")]; + tensor new_v_cache_1_axis_0 = const()[name = tensor("new_v_cache_1_axis_0"), val = tensor(1)]; + tensor new_v_cache_1_mode_0 = const()[name = tensor("new_v_cache_1_mode_0"), val = tensor("update")]; + tensor new_v_cache_1_validate_indices_0 = const()[name = tensor("new_v_cache_1_validate_indices_0"), val = tensor(false)]; + tensor new_v_cache_1 = scatter_along_axis(axis = new_v_cache_1_axis_0, data = var_207, indices = write_indexes_1, mode = new_v_cache_1_mode_0, updates = squeeze_2, validate_indices = new_v_cache_1_validate_indices_0)[name = tensor("new_v_cache_1")]; + tensor var_210_axis_0 = const()[name = tensor("op_210_axis_0"), val = tensor(0)]; + tensor var_210 = stack(axis = var_210_axis_0, values = (new_k_cache_1, new_v_cache_1))[name = tensor("op_210")]; + tensor var_211 = not_equal(x = new_k_cache_1, y = new_k_cache_1)[name = tensor("op_211")]; + tensor var_212 = const()[name = tensor("op_212"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(41230976)))]; + tensor new_k_cache_3 = select(a = var_212, b = new_k_cache_1, cond = var_211)[name = tensor("new_k_cache_3")]; + tensor var_214 = not_equal(x = new_v_cache_1, y = new_v_cache_1)[name = tensor("op_214")]; + tensor new_v_cache_3 = select(a = var_212, b = new_v_cache_1, cond = var_214)[name = tensor("new_v_cache_3")]; + tensor var_219 = const()[name = tensor("op_219"), val = tensor([0, 2, 1, 3])]; + tensor var_221 = const()[name = tensor("op_221"), val = tensor([1, 1])]; + tensor var_222 = reshape(shape = var_221, x = attn0_offset)[name = tensor("op_222")]; + tensor var_224_promoted = const()[name = tensor("op_224_promoted"), val = tensor([0x1.ep+3])]; + tensor var_225 = add(x = var_222, y = var_224_promoted)[name = tensor("op_225")]; + tensor last_pos_1_dtype_0 = const()[name = tensor("last_pos_1_dtype_0"), val = tensor("int32")]; + tensor slot_idx_1 = const()[name = tensor("slot_idx_1"), val = tensor([[0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19, 20, 21, 22, 23, 24, 25, 26, 27, 28, 29, 30, 31, 32, 33, 34, 35, 36, 37, 38, 39, 40, 41, 42, 43, 44, 45, 46, 47, 48, 49, 50, 51, 52, 53, 54, 55, 56, 57, 58, 59, 60, 61, 62, 63, 64, 65, 66, 67, 68, 69, 70, 71, 72, 73, 74, 75, 76, 77, 78, 79, 80, 81, 82, 83, 84, 85, 86, 87, 88, 89, 90, 91, 92, 93, 94, 95, 96, 97, 98, 99, 100, 101, 102, 103, 104, 105, 106, 107, 108, 109, 110, 111, 112, 113, 114, 115, 116, 117, 118, 119, 120, 121, 122, 123, 124, 125, 126, 127, 128, 129, 130, 131, 132, 133, 134, 135, 136, 137, 138, 139, 140, 141, 142, 143, 144, 145, 146, 147, 148, 149, 150, 151, 152, 153, 154, 155, 156, 157, 158, 159, 160, 161, 162, 163, 164, 165, 166, 167, 168, 169, 170, 171, 172, 173, 174, 175, 176, 177, 178, 179, 180, 181, 182, 183, 184, 185, 186, 187, 188, 189, 190, 191, 192, 193, 194, 195, 196, 197, 198, 199, 200, 201, 202, 203, 204, 205, 206, 207, 208, 209, 210, 211, 212, 213, 214, 215, 216, 217, 218, 219, 220, 221, 222, 223, 224, 225, 226, 227, 228, 229, 230, 231, 232, 233, 234, 235, 236, 237, 238, 239, 240, 241, 242, 243, 244, 245, 246, 247, 248, 249, 250, 251, 252, 253, 254, 255]])]; + tensor last_pos_1 = cast(dtype = last_pos_1_dtype_0, x = var_225)[name = tensor("cast_48")]; + tensor diff_1 = sub(x = last_pos_1, y = slot_idx_1)[name = tensor("diff_1")]; + tensor var_231_div = floor_div(x = diff_1, y = capacity_1)[name = tensor("op_231_div")]; + tensor var_231_div_scaled = mul(x = var_231_div, y = capacity_1)[name = tensor("op_231_div_scaled")]; + tensor var_231 = sub(x = diff_1, y = var_231_div_scaled)[name = tensor("op_231")]; + tensor pos_k_1 = sub(x = last_pos_1, y = var_231)[name = tensor("pos_k_1")]; + tensor var_237_promoted = const()[name = tensor("op_237_promoted"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(41755328)))]; + tensor pos_q_1 = add(x = var_222, y = var_237_promoted)[name = tensor("pos_q_1")]; + tensor var_241_axes_0 = const()[name = tensor("op_241_axes_0"), val = tensor([2])]; + tensor var_241 = expand_dims(axes = var_241_axes_0, x = pos_q_1)[name = tensor("op_241")]; + tensor var_243_axes_0 = const()[name = tensor("op_243_axes_0"), val = tensor([1])]; + tensor var_243 = expand_dims(axes = var_243_axes_0, x = pos_k_1)[name = tensor("op_243")]; + tensor var_244_promoted_dtype_0 = const()[name = tensor("op_244_promoted_dtype_0"), val = tensor("fp32")]; + tensor var_244_promoted = cast(dtype = var_244_promoted_dtype_0, x = var_243)[name = tensor("cast_47")]; + tensor delta_1 = sub(x = var_241, y = var_244_promoted)[name = tensor("delta_1")]; + tensor valid_1 = greater_equal(x = var_243, y = var_86)[name = tensor("valid_1")]; + tensor var_253 = const()[name = tensor("op_253"), val = tensor([1, 1, 1])]; + tensor var_254 = reshape(shape = var_253, x = attn0_offset)[name = tensor("op_254")]; + tensor var_256_promoted = const()[name = tensor("op_256_promoted"), val = tensor([0x1.ep+3])]; + tensor var_257 = add(x = var_254, y = var_256_promoted)[name = tensor("op_257")]; + tensor var_258 = less_equal(x = var_244_promoted, y = var_257)[name = tensor("op_258")]; + tensor valid_3 = logical_and(x = valid_1, y = var_258)[name = tensor("valid_3")]; + tensor var_86_promoted = const()[name = tensor("op_86_promoted"), val = tensor(0x0p+0)]; + tensor var_260 = greater_equal(x = delta_1, y = var_86_promoted)[name = tensor("op_260")]; + tensor attn_mask_1 = logical_and(x = valid_3, y = var_260)[name = tensor("attn_mask_1")]; + tensor var_98_promoted = const()[name = tensor("op_98_promoted"), val = tensor(0x1.f4p+7)]; + tensor var_262 = less(x = delta_1, y = var_98_promoted)[name = tensor("op_262")]; + tensor attn_mask_3 = logical_and(x = attn_mask_1, y = var_262)[name = tensor("attn_mask_3")]; + tensor attn_mask_5_axes_0 = const()[name = tensor("attn_mask_5_axes_0"), val = tensor([1])]; + tensor attn_mask_5 = expand_dims(axes = attn_mask_5_axes_0, x = attn_mask_3)[name = tensor("attn_mask_5")]; + tensor var_267_transpose_x_0 = const()[name = tensor("op_267_transpose_x_0"), val = tensor(false)]; + tensor var_267_transpose_y_0 = const()[name = tensor("op_267_transpose_y_0"), val = tensor(false)]; + tensor transpose_6_perm_0 = const()[name = tensor("transpose_6_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_7_perm_0 = const()[name = tensor("transpose_7_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_7 = transpose(perm = transpose_7_perm_0, x = new_k_cache_3)[name = tensor("transpose_16")]; + tensor transpose_6 = transpose(perm = transpose_6_perm_0, x = q_5)[name = tensor("transpose_17")]; + tensor var_267 = matmul(transpose_x = var_267_transpose_x_0, transpose_y = var_267_transpose_y_0, x = transpose_6, y = transpose_7)[name = tensor("op_267")]; + tensor var_268 = const()[name = tensor("op_268"), val = tensor(0x1p-3)]; + tensor attn_1 = mul(x = var_267, y = var_268)[name = tensor("attn_1")]; + tensor var_270 = logical_not(x = attn_mask_5)[name = tensor("op_270")]; + tensor attn_3 = select(a = var_100, b = attn_1, cond = var_270)[name = tensor("attn_3")]; + tensor attn_5 = softmax(axis = var_91, x = attn_3)[name = tensor("attn_5")]; + tensor x_5_transpose_x_0 = const()[name = tensor("x_5_transpose_x_0"), val = tensor(false)]; + tensor x_5_transpose_y_0 = const()[name = tensor("x_5_transpose_y_0"), val = tensor(false)]; + tensor v_attn_1 = transpose(perm = var_219, x = new_v_cache_3)[name = tensor("transpose_18")]; + tensor x_5 = matmul(transpose_x = x_5_transpose_x_0, transpose_y = x_5_transpose_y_0, x = attn_5, y = v_attn_1)[name = tensor("x_5")]; + tensor var_274_perm_0 = const()[name = tensor("op_274_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_275 = const()[name = tensor("op_275"), val = tensor([1, 16, 512])]; + tensor var_274 = transpose(perm = var_274_perm_0, x = x_5)[name = tensor("transpose_15")]; + tensor input_5 = reshape(shape = var_275, x = var_274)[name = tensor("input_5")]; + tensor linear_1_bias_0 = const()[name = tensor("linear_1_bias_0"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(41755456)))]; + tensor x_7 = linear(bias = linear_1_bias_0, weight = mimi_decoder_transformer_transformer_layers_0_self_attn_out_proj_weight, x = input_5)[name = tensor("linear_1")]; + tensor var_284 = mul(x = mimi_decoder_transformer_transformer_layers_0_layer_scale_1_scale, y = x_7)[name = tensor("op_284")]; + tensor input_7 = add(x = input_3, y = var_284)[name = tensor("input_7")]; + tensor input_9_axes_0 = const()[name = tensor("input_9_axes_0"), val = tensor([-1])]; + tensor input_9 = layer_norm(axes = input_9_axes_0, beta = mimi_decoder_transformer_transformer_layers_0_norm2_bias, epsilon = var_102, gamma = mimi_decoder_transformer_transformer_layers_0_norm2_weight, x = input_7)[name = tensor("input_9")]; + tensor linear_2_bias_0 = const()[name = tensor("linear_2_bias_0"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(41757568)))]; + tensor var_291 = linear(bias = linear_2_bias_0, weight = mimi_decoder_transformer_transformer_layers_0_linear1_weight, x = input_9)[name = tensor("linear_2")]; + tensor input_11_mode_0 = const()[name = tensor("input_11_mode_0"), val = tensor("EXACT")]; + tensor input_11 = gelu(mode = input_11_mode_0, x = var_291)[name = tensor("input_11")]; + tensor x_9 = linear(bias = linear_1_bias_0, weight = mimi_decoder_transformer_transformer_layers_0_linear2_weight, x = input_11)[name = tensor("linear_3")]; + tensor var_297 = mul(x = mimi_decoder_transformer_transformer_layers_0_layer_scale_2_scale, y = x_9)[name = tensor("op_297")]; + tensor input_13 = add(x = input_7, y = var_297)[name = tensor("input_13")]; + tensor query_axes_0 = const()[name = tensor("query_axes_0"), val = tensor([-1])]; + tensor query = layer_norm(axes = query_axes_0, beta = mimi_decoder_transformer_transformer_layers_1_norm1_bias, epsilon = var_102, gamma = mimi_decoder_transformer_transformer_layers_1_norm1_weight, x = input_13)[name = tensor("query")]; + tensor projected = linear(bias = linear_0_bias_0, weight = mimi_decoder_transformer_transformer_layers_1_self_attn_in_proj_weight, x = query)[name = tensor("linear_4")]; + tensor var_320 = const()[name = tensor("op_320"), val = tensor([1, 16, 3, 8, 64])]; + tensor packed = reshape(shape = var_320, x = projected)[name = tensor("packed")]; + tensor var_322_split_sizes_0 = const()[name = tensor("op_322_split_sizes_0"), val = tensor([1, 1, 1])]; + tensor var_322_axis_0 = const()[name = tensor("op_322_axis_0"), val = tensor(2)]; + tensor var_322_0, tensor var_322_1, tensor var_322_2 = split(axis = var_322_axis_0, split_sizes = var_322_split_sizes_0, x = packed)[name = tensor("op_322")]; + tensor squeeze_3_axes_0 = const()[name = tensor("squeeze_3_axes_0"), val = tensor([2])]; + tensor squeeze_3 = squeeze(axes = squeeze_3_axes_0, x = var_322_0)[name = tensor("squeeze_3")]; + tensor squeeze_4_axes_0 = const()[name = tensor("squeeze_4_axes_0"), val = tensor([2])]; + tensor squeeze_4 = squeeze(axes = squeeze_4_axes_0, x = var_322_1)[name = tensor("squeeze_4")]; + tensor squeeze_5_axes_0 = const()[name = tensor("squeeze_5_axes_0"), val = tensor([2])]; + tensor squeeze_5 = squeeze(axes = squeeze_5_axes_0, x = var_322_2)[name = tensor("squeeze_5")]; + tensor offset_begin_0 = const()[name = tensor("offset_begin_0"), val = tensor([0])]; + tensor offset_end_0 = const()[name = tensor("offset_end_0"), val = tensor([1])]; + tensor offset_end_mask_0 = const()[name = tensor("offset_end_mask_0"), val = tensor([false])]; + tensor offset_squeeze_mask_0 = const()[name = tensor("offset_squeeze_mask_0"), val = tensor([true])]; + tensor offset = slice_by_index(begin = offset_begin_0, end = offset_end_0, end_mask = offset_end_mask_0, squeeze_mask = offset_squeeze_mask_0, x = attn1_offset)[name = tensor("offset")]; + tensor freqs = const()[name = tensor("freqs"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(41765824)))]; + tensor ts_7_promoted = const()[name = tensor("ts_7_promoted"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(41766016)))]; + tensor ts_9 = add(x = ts_7_promoted, y = offset)[name = tensor("ts_9")]; + tensor var_338 = const()[name = tensor("op_338"), val = tensor([-1, 1, 1])]; + tensor ts = reshape(shape = var_338, x = ts_9)[name = tensor("ts")]; + tensor var_340 = const()[name = tensor("op_340"), val = tensor([1, 16, 8, 32, 2])]; + tensor q_9 = reshape(shape = var_340, x = squeeze_3)[name = tensor("q_9")]; + tensor var_342 = const()[name = tensor("op_342"), val = tensor([1, 16, 8, 32, 2])]; + tensor k_9 = reshape(shape = var_342, x = squeeze_4)[name = tensor("k_9")]; + tensor var_344_begin_0 = const()[name = tensor("op_344_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_344_end_0 = const()[name = tensor("op_344_end_0"), val = tensor([1, 16, 8, 32, 1])]; + tensor var_344_end_mask_0 = const()[name = tensor("op_344_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_344_squeeze_mask_0 = const()[name = tensor("op_344_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_344 = slice_by_index(begin = var_344_begin_0, end = var_344_end_0, end_mask = var_344_end_mask_0, squeeze_mask = var_344_squeeze_mask_0, x = q_9)[name = tensor("op_344")]; + tensor var_346_begin_0 = const()[name = tensor("op_346_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_346_end_0 = const()[name = tensor("op_346_end_0"), val = tensor([1, 16, 8, 32, 2])]; + tensor var_346_end_mask_0 = const()[name = tensor("op_346_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_346_squeeze_mask_0 = const()[name = tensor("op_346_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_346 = slice_by_index(begin = var_346_begin_0, end = var_346_end_0, end_mask = var_346_end_mask_0, squeeze_mask = var_346_squeeze_mask_0, x = q_9)[name = tensor("op_346")]; + tensor var_348_begin_0 = const()[name = tensor("op_348_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_348_end_0 = const()[name = tensor("op_348_end_0"), val = tensor([1, 16, 8, 32, 1])]; + tensor var_348_end_mask_0 = const()[name = tensor("op_348_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_348_squeeze_mask_0 = const()[name = tensor("op_348_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_348 = slice_by_index(begin = var_348_begin_0, end = var_348_end_0, end_mask = var_348_end_mask_0, squeeze_mask = var_348_squeeze_mask_0, x = k_9)[name = tensor("op_348")]; + tensor var_350_begin_0 = const()[name = tensor("op_350_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_350_end_0 = const()[name = tensor("op_350_end_0"), val = tensor([1, 16, 8, 32, 2])]; + tensor var_350_end_mask_0 = const()[name = tensor("op_350_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_350_squeeze_mask_0 = const()[name = tensor("op_350_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_350 = slice_by_index(begin = var_350_begin_0, end = var_350_end_0, end_mask = var_350_end_mask_0, squeeze_mask = var_350_squeeze_mask_0, x = k_9)[name = tensor("op_350")]; + tensor var_352 = mul(x = freqs, y = ts)[name = tensor("op_352")]; + tensor rotr = cos(x = var_352)[name = tensor("rotr")]; + tensor roti = sin(x = var_352)[name = tensor("roti")]; + tensor var_356 = mul(x = var_344, y = rotr)[name = tensor("op_356")]; + tensor var_357 = mul(x = var_346, y = roti)[name = tensor("op_357")]; + tensor qor_5 = sub(x = var_356, y = var_357)[name = tensor("qor_5")]; + tensor var_359 = mul(x = var_344, y = roti)[name = tensor("op_359")]; + tensor var_360 = mul(x = var_346, y = rotr)[name = tensor("op_360")]; + tensor qoi_5 = add(x = var_359, y = var_360)[name = tensor("qoi_5")]; + tensor var_362 = mul(x = var_348, y = rotr)[name = tensor("op_362")]; + tensor var_363 = mul(x = var_350, y = roti)[name = tensor("op_363")]; + tensor kor_5 = sub(x = var_362, y = var_363)[name = tensor("kor_5")]; + tensor var_365 = mul(x = var_348, y = roti)[name = tensor("op_365")]; + tensor var_366 = mul(x = var_350, y = rotr)[name = tensor("op_366")]; + tensor koi_5 = add(x = var_365, y = var_366)[name = tensor("koi_5")]; + tensor qo_axis_0 = const()[name = tensor("qo_axis_0"), val = tensor(-1)]; + tensor qo = stack(axis = qo_axis_0, values = (qor_5, qoi_5))[name = tensor("qo")]; + tensor ko_axis_0 = const()[name = tensor("ko_axis_0"), val = tensor(-1)]; + tensor ko = stack(axis = ko_axis_0, values = (kor_5, koi_5))[name = tensor("ko")]; + tensor var_376 = const()[name = tensor("op_376"), val = tensor([1, 16, 8, 64])]; + tensor q = reshape(shape = var_376, x = qo)[name = tensor("q")]; + tensor var_378 = const()[name = tensor("op_378"), val = tensor([1, 16, 8, 64])]; + tensor k = reshape(shape = var_378, x = ko)[name = tensor("k")]; + tensor capacity = const()[name = tensor("capacity"), val = tensor([256])]; + tensor var_383_dtype_0 = const()[name = tensor("op_383_dtype_0"), val = tensor("int32")]; + tensor var_384 = const()[name = tensor("op_384"), val = tensor([1, 1])]; + tensor var_383 = cast(dtype = var_383_dtype_0, x = attn1_offset)[name = tensor("cast_46")]; + tensor write_base = reshape(shape = var_384, x = var_383)[name = tensor("write_base")]; + tensor write_range = const()[name = tensor("write_range"), val = tensor([[0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15]])]; + tensor abs_idx = add(x = write_base, y = write_range)[name = tensor("abs_idx")]; + tensor wrapped_div = floor_div(x = abs_idx, y = capacity)[name = tensor("wrapped_div")]; + tensor wrapped_div_scaled = mul(x = wrapped_div, y = capacity)[name = tensor("wrapped_div_scaled")]; + tensor wrapped = sub(x = abs_idx, y = wrapped_div_scaled)[name = tensor("wrapped")]; + tensor var_391 = const()[name = tensor("op_391"), val = tensor([1, 16, 1, 1])]; + tensor var_392 = reshape(shape = var_391, x = wrapped)[name = tensor("op_392")]; + tensor write_indexes_reps_0 = const()[name = tensor("write_indexes_reps_0"), val = tensor([1, 1, 8, 64])]; + tensor write_indexes = tile(reps = write_indexes_reps_0, x = var_392)[name = tensor("write_indexes")]; + tensor var_395_begin_0 = const()[name = tensor("op_395_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_395_end_0 = const()[name = tensor("op_395_end_0"), val = tensor([1, 1, 256, 8, 64])]; + tensor var_395_end_mask_0 = const()[name = tensor("op_395_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_395_squeeze_mask_0 = const()[name = tensor("op_395_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor var_395 = slice_by_index(begin = var_395_begin_0, end = var_395_end_0, end_mask = var_395_end_mask_0, squeeze_mask = var_395_squeeze_mask_0, x = attn1_cache)[name = tensor("op_395")]; + tensor new_k_cache_5_axis_0 = const()[name = tensor("new_k_cache_5_axis_0"), val = tensor(1)]; + tensor new_k_cache_5_mode_0 = const()[name = tensor("new_k_cache_5_mode_0"), val = tensor("update")]; + tensor new_k_cache_5_validate_indices_0 = const()[name = tensor("new_k_cache_5_validate_indices_0"), val = tensor(false)]; + tensor new_k_cache_5 = scatter_along_axis(axis = new_k_cache_5_axis_0, data = var_395, indices = write_indexes, mode = new_k_cache_5_mode_0, updates = k, validate_indices = new_k_cache_5_validate_indices_0)[name = tensor("new_k_cache_5")]; + tensor var_397_begin_0 = const()[name = tensor("op_397_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor var_397_end_0 = const()[name = tensor("op_397_end_0"), val = tensor([2, 1, 256, 8, 64])]; + tensor var_397_end_mask_0 = const()[name = tensor("op_397_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_397_squeeze_mask_0 = const()[name = tensor("op_397_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor var_397 = slice_by_index(begin = var_397_begin_0, end = var_397_end_0, end_mask = var_397_end_mask_0, squeeze_mask = var_397_squeeze_mask_0, x = attn1_cache)[name = tensor("op_397")]; + tensor new_v_cache_5_axis_0 = const()[name = tensor("new_v_cache_5_axis_0"), val = tensor(1)]; + tensor new_v_cache_5_mode_0 = const()[name = tensor("new_v_cache_5_mode_0"), val = tensor("update")]; + tensor new_v_cache_5_validate_indices_0 = const()[name = tensor("new_v_cache_5_validate_indices_0"), val = tensor(false)]; + tensor new_v_cache_5 = scatter_along_axis(axis = new_v_cache_5_axis_0, data = var_397, indices = write_indexes, mode = new_v_cache_5_mode_0, updates = squeeze_5, validate_indices = new_v_cache_5_validate_indices_0)[name = tensor("new_v_cache_5")]; + tensor var_400_axis_0 = const()[name = tensor("op_400_axis_0"), val = tensor(0)]; + tensor var_400 = stack(axis = var_400_axis_0, values = (new_k_cache_5, new_v_cache_5))[name = tensor("op_400")]; + tensor var_401 = not_equal(x = new_k_cache_5, y = new_k_cache_5)[name = tensor("op_401")]; + tensor new_k_cache = select(a = var_212, b = new_k_cache_5, cond = var_401)[name = tensor("new_k_cache")]; + tensor var_404 = not_equal(x = new_v_cache_5, y = new_v_cache_5)[name = tensor("op_404")]; + tensor new_v_cache = select(a = var_212, b = new_v_cache_5, cond = var_404)[name = tensor("new_v_cache")]; + tensor var_409 = const()[name = tensor("op_409"), val = tensor([0, 2, 1, 3])]; + tensor var_411 = const()[name = tensor("op_411"), val = tensor([1, 1])]; + tensor var_412 = reshape(shape = var_411, x = attn1_offset)[name = tensor("op_412")]; + tensor var_414_promoted = const()[name = tensor("op_414_promoted"), val = tensor([0x1.ep+3])]; + tensor var_415 = add(x = var_412, y = var_414_promoted)[name = tensor("op_415")]; + tensor last_pos_dtype_0 = const()[name = tensor("last_pos_dtype_0"), val = tensor("int32")]; + tensor last_pos = cast(dtype = last_pos_dtype_0, x = var_415)[name = tensor("cast_45")]; + tensor diff = sub(x = last_pos, y = slot_idx_1)[name = tensor("diff")]; + tensor var_421_div = floor_div(x = diff, y = capacity)[name = tensor("op_421_div")]; + tensor var_421_div_scaled = mul(x = var_421_div, y = capacity)[name = tensor("op_421_div_scaled")]; + tensor var_421 = sub(x = diff, y = var_421_div_scaled)[name = tensor("op_421")]; + tensor pos_k = sub(x = last_pos, y = var_421)[name = tensor("pos_k")]; + tensor var_427_promoted = const()[name = tensor("op_427_promoted"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(41766144)))]; + tensor pos_q = add(x = var_412, y = var_427_promoted)[name = tensor("pos_q")]; + tensor var_431_axes_0 = const()[name = tensor("op_431_axes_0"), val = tensor([2])]; + tensor var_431 = expand_dims(axes = var_431_axes_0, x = pos_q)[name = tensor("op_431")]; + tensor var_433_axes_0 = const()[name = tensor("op_433_axes_0"), val = tensor([1])]; + tensor var_433 = expand_dims(axes = var_433_axes_0, x = pos_k)[name = tensor("op_433")]; + tensor var_434_promoted_dtype_0 = const()[name = tensor("op_434_promoted_dtype_0"), val = tensor("fp32")]; + tensor var_434_promoted = cast(dtype = var_434_promoted_dtype_0, x = var_433)[name = tensor("cast_44")]; + tensor delta = sub(x = var_431, y = var_434_promoted)[name = tensor("delta")]; + tensor valid_5 = greater_equal(x = var_433, y = var_86)[name = tensor("valid_5")]; + tensor var_443 = const()[name = tensor("op_443"), val = tensor([1, 1, 1])]; + tensor var_444 = reshape(shape = var_443, x = attn1_offset)[name = tensor("op_444")]; + tensor var_446_promoted = const()[name = tensor("op_446_promoted"), val = tensor([0x1.ep+3])]; + tensor var_447 = add(x = var_444, y = var_446_promoted)[name = tensor("op_447")]; + tensor var_448 = less_equal(x = var_434_promoted, y = var_447)[name = tensor("op_448")]; + tensor valid = logical_and(x = valid_5, y = var_448)[name = tensor("valid")]; + tensor var_86_promoted_1 = const()[name = tensor("op_86_promoted_1"), val = tensor(0x0p+0)]; + tensor var_450 = greater_equal(x = delta, y = var_86_promoted_1)[name = tensor("op_450")]; + tensor attn_mask_7 = logical_and(x = valid, y = var_450)[name = tensor("attn_mask_7")]; + tensor var_98_promoted_1 = const()[name = tensor("op_98_promoted_1"), val = tensor(0x1.f4p+7)]; + tensor var_452 = less(x = delta, y = var_98_promoted_1)[name = tensor("op_452")]; + tensor attn_mask_9 = logical_and(x = attn_mask_7, y = var_452)[name = tensor("attn_mask_9")]; + tensor attn_mask_axes_0 = const()[name = tensor("attn_mask_axes_0"), val = tensor([1])]; + tensor attn_mask = expand_dims(axes = attn_mask_axes_0, x = attn_mask_9)[name = tensor("attn_mask")]; + tensor var_457_transpose_x_0 = const()[name = tensor("op_457_transpose_x_0"), val = tensor(false)]; + tensor var_457_transpose_y_0 = const()[name = tensor("op_457_transpose_y_0"), val = tensor(false)]; + tensor transpose_8_perm_0 = const()[name = tensor("transpose_8_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_9_perm_0 = const()[name = tensor("transpose_9_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_9 = transpose(perm = transpose_9_perm_0, x = new_k_cache)[name = tensor("transpose_12")]; + tensor transpose_8 = transpose(perm = transpose_8_perm_0, x = q)[name = tensor("transpose_13")]; + tensor var_457 = matmul(transpose_x = var_457_transpose_x_0, transpose_y = var_457_transpose_y_0, x = transpose_8, y = transpose_9)[name = tensor("op_457")]; + tensor var_458 = const()[name = tensor("op_458"), val = tensor(0x1p-3)]; + tensor attn_7 = mul(x = var_457, y = var_458)[name = tensor("attn_7")]; + tensor var_460 = logical_not(x = attn_mask)[name = tensor("op_460")]; + tensor attn_9 = select(a = var_100, b = attn_7, cond = var_460)[name = tensor("attn_9")]; + tensor attn = softmax(axis = var_91, x = attn_9)[name = tensor("attn")]; + tensor x_11_transpose_x_0 = const()[name = tensor("x_11_transpose_x_0"), val = tensor(false)]; + tensor x_11_transpose_y_0 = const()[name = tensor("x_11_transpose_y_0"), val = tensor(false)]; + tensor v_attn = transpose(perm = var_409, x = new_v_cache)[name = tensor("transpose_14")]; + tensor x_11 = matmul(transpose_x = x_11_transpose_x_0, transpose_y = x_11_transpose_y_0, x = attn, y = v_attn)[name = tensor("x_11")]; + tensor var_464_perm_0 = const()[name = tensor("op_464_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_465 = const()[name = tensor("op_465"), val = tensor([1, 16, 512])]; + tensor var_464 = transpose(perm = var_464_perm_0, x = x_11)[name = tensor("transpose_11")]; + tensor input_15 = reshape(shape = var_465, x = var_464)[name = tensor("input_15")]; + tensor x_13 = linear(bias = linear_1_bias_0, weight = mimi_decoder_transformer_transformer_layers_1_self_attn_out_proj_weight, x = input_15)[name = tensor("linear_5")]; + tensor var_474 = mul(x = mimi_decoder_transformer_transformer_layers_1_layer_scale_1_scale, y = x_13)[name = tensor("op_474")]; + tensor input_17 = add(x = input_13, y = var_474)[name = tensor("input_17")]; + tensor input_19_axes_0 = const()[name = tensor("input_19_axes_0"), val = tensor([-1])]; + tensor input_19 = layer_norm(axes = input_19_axes_0, beta = mimi_decoder_transformer_transformer_layers_1_norm2_bias, epsilon = var_102, gamma = mimi_decoder_transformer_transformer_layers_1_norm2_weight, x = input_17)[name = tensor("input_19")]; + tensor var_481 = linear(bias = linear_2_bias_0, weight = mimi_decoder_transformer_transformer_layers_1_linear1_weight, x = input_19)[name = tensor("linear_6")]; + tensor input_21_mode_0 = const()[name = tensor("input_21_mode_0"), val = tensor("EXACT")]; + tensor input_21 = gelu(mode = input_21_mode_0, x = var_481)[name = tensor("input_21")]; + tensor x_15 = linear(bias = linear_1_bias_0, weight = mimi_decoder_transformer_transformer_layers_1_linear2_weight, x = input_21)[name = tensor("linear_7")]; + tensor var_487 = mul(x = mimi_decoder_transformer_transformer_layers_1_layer_scale_2_scale, y = x_15)[name = tensor("op_487")]; + tensor z = add(x = input_17, y = var_487)[name = tensor("z")]; + tensor x_17_perm_0 = const()[name = tensor("x_17_perm_0"), val = tensor([0, 2, 1])]; + tensor var_507 = const()[name = tensor("op_507"), val = tensor(0x1p+0)]; + tensor var_508 = const()[name = tensor("op_508"), val = tensor(-1)]; + tensor input_23_interleave_0 = const()[name = tensor("input_23_interleave_0"), val = tensor(false)]; + tensor x_17 = transpose(perm = x_17_perm_0, x = z)[name = tensor("transpose_10")]; + tensor input_23 = concat(axis = var_508, interleave = input_23_interleave_0, values = (conv0_prev, x_17))[name = tensor("input_23")]; + tensor input_25_pad_type_0 = const()[name = tensor("input_25_pad_type_0"), val = tensor("valid")]; + tensor input_25_strides_0 = const()[name = tensor("input_25_strides_0"), val = tensor([1])]; + tensor input_25_pad_0 = const()[name = tensor("input_25_pad_0"), val = tensor([0, 0])]; + tensor input_25_dilations_0 = const()[name = tensor("input_25_dilations_0"), val = tensor([1])]; + tensor input_25_groups_0 = const()[name = tensor("input_25_groups_0"), val = tensor(1)]; + tensor input_25 = conv(bias = mimi_decoder_model_0_conv_bias, dilations = input_25_dilations_0, groups = input_25_groups_0, pad = input_25_pad_0, pad_type = input_25_pad_type_0, strides = input_25_strides_0, weight = mimi_decoder_model_0_conv_weight, x = input_23)[name = tensor("input_25")]; + tensor var_542_begin_0 = const()[name = tensor("op_542_begin_0"), val = tensor([0, 0, 16])]; + tensor var_542_end_0 = const()[name = tensor("op_542_end_0"), val = tensor([1, 512, 22])]; + tensor var_542_end_mask_0 = const()[name = tensor("op_542_end_mask_0"), val = tensor([true, true, true])]; + tensor var_542 = slice_by_index(begin = var_542_begin_0, end = var_542_end_0, end_mask = var_542_end_mask_0, x = input_23)[name = tensor("op_542")]; + tensor input_27 = elu(alpha = var_507, x = input_25)[name = tensor("input_27")]; + tensor y_5_pad_type_0 = const()[name = tensor("y_5_pad_type_0"), val = tensor("valid")]; + tensor y_5_strides_0 = const()[name = tensor("y_5_strides_0"), val = tensor([6])]; + tensor y_5_pad_0 = const()[name = tensor("y_5_pad_0"), val = tensor([0, 0])]; + tensor y_5_dilations_0 = const()[name = tensor("y_5_dilations_0"), val = tensor([1])]; + tensor y_5_groups_0 = const()[name = tensor("y_5_groups_0"), val = tensor(1)]; + tensor y_5_has_output_shape_output_shape_0 = const()[name = tensor("y_5_has_output_shape_output_shape_0"), val = tensor([1, 256, 102])]; + tensor y_5_has_output_shape = conv_transpose(bias = mimi_decoder_model_2_convtr_bias, dilations = y_5_dilations_0, groups = y_5_groups_0, output_shape = y_5_has_output_shape_output_shape_0, pad = y_5_pad_0, pad_type = y_5_pad_type_0, strides = y_5_strides_0, weight = mimi_decoder_model_2_convtr_weight, x = input_27)[name = tensor("y_5_has_output_shape")]; + tensor var_557_begin_0 = const()[name = tensor("op_557_begin_0"), val = tensor([0, 0, 0])]; + tensor var_557_end_0 = const()[name = tensor("op_557_end_0"), val = tensor([1, 256, 6])]; + tensor var_557_end_mask_0 = const()[name = tensor("op_557_end_mask_0"), val = tensor([true, true, false])]; + tensor var_557 = slice_by_index(begin = var_557_begin_0, end = var_557_end_0, end_mask = var_557_end_mask_0, x = y_5_has_output_shape)[name = tensor("op_557")]; + tensor var_558 = add(x = var_557, y = convtr0_partial)[name = tensor("op_558")]; + tensor var_559_begin_0 = const()[name = tensor("op_559_begin_0"), val = tensor([0, 0, 6])]; + tensor var_559_end_0 = const()[name = tensor("op_559_end_0"), val = tensor([1, 256, 102])]; + tensor var_559_end_mask_0 = const()[name = tensor("op_559_end_mask_0"), val = tensor([true, true, true])]; + tensor var_559 = slice_by_index(begin = var_559_begin_0, end = var_559_end_0, end_mask = var_559_end_mask_0, x = y_5_has_output_shape)[name = tensor("op_559")]; + tensor y_7_interleave_0 = const()[name = tensor("y_7_interleave_0"), val = tensor(false)]; + tensor y_7 = concat(axis = var_508, interleave = y_7_interleave_0, values = (var_558, var_559))[name = tensor("y_7")]; + tensor new_partial_1_begin_0 = const()[name = tensor("new_partial_1_begin_0"), val = tensor([0, 0, 96])]; + tensor new_partial_1_end_0 = const()[name = tensor("new_partial_1_end_0"), val = tensor([1, 256, 102])]; + tensor new_partial_1_end_mask_0 = const()[name = tensor("new_partial_1_end_mask_0"), val = tensor([true, true, true])]; + tensor new_partial_1 = slice_by_index(begin = new_partial_1_begin_0, end = new_partial_1_end_0, end_mask = new_partial_1_end_mask_0, x = y_7)[name = tensor("new_partial_1")]; + tensor var_564 = const()[name = tensor("op_564"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(41766272)))]; + tensor var_565 = sub(x = new_partial_1, y = var_564)[name = tensor("op_565")]; + tensor input_29_begin_0 = const()[name = tensor("input_29_begin_0"), val = tensor([0, 0, 0])]; + tensor input_29_end_0 = const()[name = tensor("input_29_end_0"), val = tensor([1, 256, 96])]; + tensor input_29_end_mask_0 = const()[name = tensor("input_29_end_mask_0"), val = tensor([true, true, false])]; + tensor input_29 = slice_by_index(begin = input_29_begin_0, end = input_29_end_0, end_mask = input_29_end_mask_0, x = y_7)[name = tensor("input_29")]; + tensor x_19 = elu(alpha = var_507, x = input_29)[name = tensor("x_19")]; + tensor input_31_interleave_0 = const()[name = tensor("input_31_interleave_0"), val = tensor(false)]; + tensor input_31 = concat(axis = var_508, interleave = input_31_interleave_0, values = (res0_conv0_prev, x_19))[name = tensor("input_31")]; + tensor input_33_pad_type_0 = const()[name = tensor("input_33_pad_type_0"), val = tensor("valid")]; + tensor input_33_strides_0 = const()[name = tensor("input_33_strides_0"), val = tensor([1])]; + tensor input_33_pad_0 = const()[name = tensor("input_33_pad_0"), val = tensor([0, 0])]; + tensor input_33_dilations_0 = const()[name = tensor("input_33_dilations_0"), val = tensor([1])]; + tensor input_33_groups_0 = const()[name = tensor("input_33_groups_0"), val = tensor(1)]; + tensor input_33 = conv(bias = mimi_decoder_model_3_block_1_conv_bias, dilations = input_33_dilations_0, groups = input_33_groups_0, pad = input_33_pad_0, pad_type = input_33_pad_type_0, strides = input_33_strides_0, weight = mimi_decoder_model_3_block_1_conv_weight, x = input_31)[name = tensor("input_33")]; + tensor var_585_begin_0 = const()[name = tensor("op_585_begin_0"), val = tensor([0, 0, 96])]; + tensor var_585_end_0 = const()[name = tensor("op_585_end_0"), val = tensor([1, 256, 98])]; + tensor var_585_end_mask_0 = const()[name = tensor("op_585_end_mask_0"), val = tensor([true, true, true])]; + tensor var_585 = slice_by_index(begin = var_585_begin_0, end = var_585_end_0, end_mask = var_585_end_mask_0, x = input_31)[name = tensor("op_585")]; + tensor x_21 = elu(alpha = var_507, x = input_33)[name = tensor("x_21")]; + tensor v_5_pad_type_0 = const()[name = tensor("v_5_pad_type_0"), val = tensor("valid")]; + tensor v_5_strides_0 = const()[name = tensor("v_5_strides_0"), val = tensor([1])]; + tensor v_5_pad_0 = const()[name = tensor("v_5_pad_0"), val = tensor([0, 0])]; + tensor v_5_dilations_0 = const()[name = tensor("v_5_dilations_0"), val = tensor([1])]; + tensor v_5_groups_0 = const()[name = tensor("v_5_groups_0"), val = tensor(1)]; + tensor v_5 = conv(bias = mimi_decoder_model_3_block_3_conv_bias, dilations = v_5_dilations_0, groups = v_5_groups_0, pad = v_5_pad_0, pad_type = v_5_pad_type_0, strides = v_5_strides_0, weight = mimi_decoder_model_3_block_3_conv_weight, x = x_21)[name = tensor("v_5")]; + tensor input_35 = add(x = input_29, y = v_5)[name = tensor("input_35")]; + tensor input_37 = elu(alpha = var_507, x = input_35)[name = tensor("input_37")]; + tensor y_9_pad_type_0 = const()[name = tensor("y_9_pad_type_0"), val = tensor("valid")]; + tensor y_9_strides_0 = const()[name = tensor("y_9_strides_0"), val = tensor([5])]; + tensor y_9_pad_0 = const()[name = tensor("y_9_pad_0"), val = tensor([0, 0])]; + tensor y_9_dilations_0 = const()[name = tensor("y_9_dilations_0"), val = tensor([1])]; + tensor y_9_groups_0 = const()[name = tensor("y_9_groups_0"), val = tensor(1)]; + tensor y_9_has_output_shape_output_shape_0 = const()[name = tensor("y_9_has_output_shape_output_shape_0"), val = tensor([1, 128, 485])]; + tensor y_9_has_output_shape = conv_transpose(bias = mimi_decoder_model_5_convtr_bias, dilations = y_9_dilations_0, groups = y_9_groups_0, output_shape = y_9_has_output_shape_output_shape_0, pad = y_9_pad_0, pad_type = y_9_pad_type_0, strides = y_9_strides_0, weight = mimi_decoder_model_5_convtr_weight, x = input_37)[name = tensor("y_9_has_output_shape")]; + tensor var_613_begin_0 = const()[name = tensor("op_613_begin_0"), val = tensor([0, 0, 0])]; + tensor var_613_end_0 = const()[name = tensor("op_613_end_0"), val = tensor([1, 128, 5])]; + tensor var_613_end_mask_0 = const()[name = tensor("op_613_end_mask_0"), val = tensor([true, true, false])]; + tensor var_613 = slice_by_index(begin = var_613_begin_0, end = var_613_end_0, end_mask = var_613_end_mask_0, x = y_9_has_output_shape)[name = tensor("op_613")]; + tensor var_614 = add(x = var_613, y = convtr1_partial)[name = tensor("op_614")]; + tensor var_615_begin_0 = const()[name = tensor("op_615_begin_0"), val = tensor([0, 0, 5])]; + tensor var_615_end_0 = const()[name = tensor("op_615_end_0"), val = tensor([1, 128, 485])]; + tensor var_615_end_mask_0 = const()[name = tensor("op_615_end_mask_0"), val = tensor([true, true, true])]; + tensor var_615 = slice_by_index(begin = var_615_begin_0, end = var_615_end_0, end_mask = var_615_end_mask_0, x = y_9_has_output_shape)[name = tensor("op_615")]; + tensor y_11_interleave_0 = const()[name = tensor("y_11_interleave_0"), val = tensor(false)]; + tensor y_11 = concat(axis = var_508, interleave = y_11_interleave_0, values = (var_614, var_615))[name = tensor("y_11")]; + tensor new_partial_3_begin_0 = const()[name = tensor("new_partial_3_begin_0"), val = tensor([0, 0, 480])]; + tensor new_partial_3_end_0 = const()[name = tensor("new_partial_3_end_0"), val = tensor([1, 128, 485])]; + tensor new_partial_3_end_mask_0 = const()[name = tensor("new_partial_3_end_mask_0"), val = tensor([true, true, true])]; + tensor new_partial_3 = slice_by_index(begin = new_partial_3_begin_0, end = new_partial_3_end_0, end_mask = new_partial_3_end_mask_0, x = y_11)[name = tensor("new_partial_3")]; + tensor var_620 = const()[name = tensor("op_620"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(41767360)))]; + tensor var_621 = sub(x = new_partial_3, y = var_620)[name = tensor("op_621")]; + tensor input_39_begin_0 = const()[name = tensor("input_39_begin_0"), val = tensor([0, 0, 0])]; + tensor input_39_end_0 = const()[name = tensor("input_39_end_0"), val = tensor([1, 128, 480])]; + tensor input_39_end_mask_0 = const()[name = tensor("input_39_end_mask_0"), val = tensor([true, true, false])]; + tensor input_39 = slice_by_index(begin = input_39_begin_0, end = input_39_end_0, end_mask = input_39_end_mask_0, x = y_11)[name = tensor("input_39")]; + tensor x_23 = elu(alpha = var_507, x = input_39)[name = tensor("x_23")]; + tensor input_41_interleave_0 = const()[name = tensor("input_41_interleave_0"), val = tensor(false)]; + tensor input_41 = concat(axis = var_508, interleave = input_41_interleave_0, values = (res1_conv0_prev, x_23))[name = tensor("input_41")]; + tensor input_43_pad_type_0 = const()[name = tensor("input_43_pad_type_0"), val = tensor("valid")]; + tensor input_43_strides_0 = const()[name = tensor("input_43_strides_0"), val = tensor([1])]; + tensor input_43_pad_0 = const()[name = tensor("input_43_pad_0"), val = tensor([0, 0])]; + tensor input_43_dilations_0 = const()[name = tensor("input_43_dilations_0"), val = tensor([1])]; + tensor input_43_groups_0 = const()[name = tensor("input_43_groups_0"), val = tensor(1)]; + tensor input_43 = conv(bias = mimi_decoder_model_6_block_1_conv_bias, dilations = input_43_dilations_0, groups = input_43_groups_0, pad = input_43_pad_0, pad_type = input_43_pad_type_0, strides = input_43_strides_0, weight = mimi_decoder_model_6_block_1_conv_weight, x = input_41)[name = tensor("input_43")]; + tensor var_641_begin_0 = const()[name = tensor("op_641_begin_0"), val = tensor([0, 0, 480])]; + tensor var_641_end_0 = const()[name = tensor("op_641_end_0"), val = tensor([1, 128, 482])]; + tensor var_641_end_mask_0 = const()[name = tensor("op_641_end_mask_0"), val = tensor([true, true, true])]; + tensor var_641 = slice_by_index(begin = var_641_begin_0, end = var_641_end_0, end_mask = var_641_end_mask_0, x = input_41)[name = tensor("op_641")]; + tensor x_25 = elu(alpha = var_507, x = input_43)[name = tensor("x_25")]; + tensor v_7_pad_type_0 = const()[name = tensor("v_7_pad_type_0"), val = tensor("valid")]; + tensor v_7_strides_0 = const()[name = tensor("v_7_strides_0"), val = tensor([1])]; + tensor v_7_pad_0 = const()[name = tensor("v_7_pad_0"), val = tensor([0, 0])]; + tensor v_7_dilations_0 = const()[name = tensor("v_7_dilations_0"), val = tensor([1])]; + tensor v_7_groups_0 = const()[name = tensor("v_7_groups_0"), val = tensor(1)]; + tensor v_7 = conv(bias = mimi_decoder_model_6_block_3_conv_bias, dilations = v_7_dilations_0, groups = v_7_groups_0, pad = v_7_pad_0, pad_type = v_7_pad_type_0, strides = v_7_strides_0, weight = mimi_decoder_model_6_block_3_conv_weight, x = x_25)[name = tensor("v_7")]; + tensor input_45 = add(x = input_39, y = v_7)[name = tensor("input_45")]; + tensor input_47 = elu(alpha = var_507, x = input_45)[name = tensor("input_47")]; + tensor y_13_pad_type_0 = const()[name = tensor("y_13_pad_type_0"), val = tensor("valid")]; + tensor y_13_strides_0 = const()[name = tensor("y_13_strides_0"), val = tensor([4])]; + tensor y_13_pad_0 = const()[name = tensor("y_13_pad_0"), val = tensor([0, 0])]; + tensor y_13_dilations_0 = const()[name = tensor("y_13_dilations_0"), val = tensor([1])]; + tensor y_13_groups_0 = const()[name = tensor("y_13_groups_0"), val = tensor(1)]; + tensor y_13_has_output_shape_output_shape_0 = const()[name = tensor("y_13_has_output_shape_output_shape_0"), val = tensor([1, 64, 1924])]; + tensor y_13_has_output_shape = conv_transpose(bias = mimi_decoder_model_8_convtr_bias, dilations = y_13_dilations_0, groups = y_13_groups_0, output_shape = y_13_has_output_shape_output_shape_0, pad = y_13_pad_0, pad_type = y_13_pad_type_0, strides = y_13_strides_0, weight = mimi_decoder_model_8_convtr_weight, x = input_47)[name = tensor("y_13_has_output_shape")]; + tensor var_669_begin_0 = const()[name = tensor("op_669_begin_0"), val = tensor([0, 0, 0])]; + tensor var_669_end_0 = const()[name = tensor("op_669_end_0"), val = tensor([1, 64, 4])]; + tensor var_669_end_mask_0 = const()[name = tensor("op_669_end_mask_0"), val = tensor([true, true, false])]; + tensor var_669 = slice_by_index(begin = var_669_begin_0, end = var_669_end_0, end_mask = var_669_end_mask_0, x = y_13_has_output_shape)[name = tensor("op_669")]; + tensor var_670 = add(x = var_669, y = convtr2_partial)[name = tensor("op_670")]; + tensor var_671_begin_0 = const()[name = tensor("op_671_begin_0"), val = tensor([0, 0, 4])]; + tensor var_671_end_0 = const()[name = tensor("op_671_end_0"), val = tensor([1, 64, 1924])]; + tensor var_671_end_mask_0 = const()[name = tensor("op_671_end_mask_0"), val = tensor([true, true, true])]; + tensor var_671 = slice_by_index(begin = var_671_begin_0, end = var_671_end_0, end_mask = var_671_end_mask_0, x = y_13_has_output_shape)[name = tensor("op_671")]; + tensor y_interleave_0 = const()[name = tensor("y_interleave_0"), val = tensor(false)]; + tensor y = concat(axis = var_508, interleave = y_interleave_0, values = (var_670, var_671))[name = tensor("y")]; + tensor new_partial_begin_0 = const()[name = tensor("new_partial_begin_0"), val = tensor([0, 0, 1920])]; + tensor new_partial_end_0 = const()[name = tensor("new_partial_end_0"), val = tensor([1, 64, 1924])]; + tensor new_partial_end_mask_0 = const()[name = tensor("new_partial_end_mask_0"), val = tensor([true, true, true])]; + tensor new_partial = slice_by_index(begin = new_partial_begin_0, end = new_partial_end_0, end_mask = new_partial_end_mask_0, x = y)[name = tensor("new_partial")]; + tensor var_676 = const()[name = tensor("op_676"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(41767936)))]; + tensor var_677 = sub(x = new_partial, y = var_676)[name = tensor("op_677")]; + tensor input_49_begin_0 = const()[name = tensor("input_49_begin_0"), val = tensor([0, 0, 0])]; + tensor input_49_end_0 = const()[name = tensor("input_49_end_0"), val = tensor([1, 64, 1920])]; + tensor input_49_end_mask_0 = const()[name = tensor("input_49_end_mask_0"), val = tensor([true, true, false])]; + tensor input_49 = slice_by_index(begin = input_49_begin_0, end = input_49_end_0, end_mask = input_49_end_mask_0, x = y)[name = tensor("input_49")]; + tensor x_27 = elu(alpha = var_507, x = input_49)[name = tensor("x_27")]; + tensor input_51_interleave_0 = const()[name = tensor("input_51_interleave_0"), val = tensor(false)]; + tensor input_51 = concat(axis = var_508, interleave = input_51_interleave_0, values = (res2_conv0_prev, x_27))[name = tensor("input_51")]; + tensor input_53_pad_type_0 = const()[name = tensor("input_53_pad_type_0"), val = tensor("valid")]; + tensor input_53_strides_0 = const()[name = tensor("input_53_strides_0"), val = tensor([1])]; + tensor input_53_pad_0 = const()[name = tensor("input_53_pad_0"), val = tensor([0, 0])]; + tensor input_53_dilations_0 = const()[name = tensor("input_53_dilations_0"), val = tensor([1])]; + tensor input_53_groups_0 = const()[name = tensor("input_53_groups_0"), val = tensor(1)]; + tensor input_53 = conv(bias = mimi_decoder_model_9_block_1_conv_bias, dilations = input_53_dilations_0, groups = input_53_groups_0, pad = input_53_pad_0, pad_type = input_53_pad_type_0, strides = input_53_strides_0, weight = mimi_decoder_model_9_block_1_conv_weight, x = input_51)[name = tensor("input_53")]; + tensor var_697_begin_0 = const()[name = tensor("op_697_begin_0"), val = tensor([0, 0, 1920])]; + tensor var_697_end_0 = const()[name = tensor("op_697_end_0"), val = tensor([1, 64, 1922])]; + tensor var_697_end_mask_0 = const()[name = tensor("op_697_end_mask_0"), val = tensor([true, true, true])]; + tensor var_697 = slice_by_index(begin = var_697_begin_0, end = var_697_end_0, end_mask = var_697_end_mask_0, x = input_51)[name = tensor("op_697")]; + tensor x_29 = elu(alpha = var_507, x = input_53)[name = tensor("x_29")]; + tensor v_pad_type_0 = const()[name = tensor("v_pad_type_0"), val = tensor("valid")]; + tensor v_strides_0 = const()[name = tensor("v_strides_0"), val = tensor([1])]; + tensor v_pad_0 = const()[name = tensor("v_pad_0"), val = tensor([0, 0])]; + tensor v_dilations_0 = const()[name = tensor("v_dilations_0"), val = tensor([1])]; + tensor v_groups_0 = const()[name = tensor("v_groups_0"), val = tensor(1)]; + tensor v = conv(bias = mimi_decoder_model_9_block_3_conv_bias, dilations = v_dilations_0, groups = v_groups_0, pad = v_pad_0, pad_type = v_pad_type_0, strides = v_strides_0, weight = mimi_decoder_model_9_block_3_conv_weight, x = x_29)[name = tensor("v")]; + tensor input_55 = add(x = input_49, y = v)[name = tensor("input_55")]; + tensor x = elu(alpha = var_507, x = input_55)[name = tensor("x")]; + tensor input_interleave_0 = const()[name = tensor("input_interleave_0"), val = tensor(false)]; + tensor input = concat(axis = var_508, interleave = input_interleave_0, values = (conv_final_prev, x))[name = tensor("input")]; + tensor var_724_pad_type_0 = const()[name = tensor("op_724_pad_type_0"), val = tensor("valid")]; + tensor var_724_strides_0 = const()[name = tensor("op_724_strides_0"), val = tensor([1])]; + tensor var_724_pad_0 = const()[name = tensor("op_724_pad_0"), val = tensor([0, 0])]; + tensor var_724_dilations_0 = const()[name = tensor("op_724_dilations_0"), val = tensor([1])]; + tensor var_724_groups_0 = const()[name = tensor("op_724_groups_0"), val = tensor(1)]; + tensor var_724 = conv(bias = mimi_decoder_model_11_conv_bias, dilations = var_724_dilations_0, groups = var_724_groups_0, pad = var_724_pad_0, pad_type = var_724_pad_type_0, strides = var_724_strides_0, weight = mimi_decoder_model_11_conv_weight, x = input)[name = tensor("op_724")]; + tensor var_725_begin_0 = const()[name = tensor("op_725_begin_0"), val = tensor([0, 0, 1920])]; + tensor var_725_end_0 = const()[name = tensor("op_725_end_0"), val = tensor([1, 64, 1922])]; + tensor var_725_end_mask_0 = const()[name = tensor("op_725_end_mask_0"), val = tensor([true, true, true])]; + tensor var_725 = slice_by_index(begin = var_725_begin_0, end = var_725_end_0, end_mask = var_725_end_mask_0, x = input)[name = tensor("op_725")]; + tensor var_740_promoted = const()[name = tensor("op_740_promoted"), val = tensor(0x1p+4)]; + tensor var_741 = add(x = attn0_offset, y = var_740_promoted)[name = tensor("op_741")]; + tensor var_743_promoted = const()[name = tensor("op_743_promoted"), val = tensor(0x1p+4)]; + tensor var_744 = add(x = attn1_offset, y = var_743_promoted)[name = tensor("op_744")]; + tensor conv0_first_tmp = identity(x = conv0_first)[name = tensor("conv0_first_tmp")]; + tensor res0_conv0_first_tmp = identity(x = res0_conv0_first)[name = tensor("res0_conv0_first_tmp")]; + tensor res0_conv1_prev_tmp = identity(x = res0_conv1_prev)[name = tensor("res0_conv1_prev_tmp")]; + tensor res0_conv1_first_tmp = identity(x = res0_conv1_first)[name = tensor("res0_conv1_first_tmp")]; + tensor res1_conv0_first_tmp = identity(x = res1_conv0_first)[name = tensor("res1_conv0_first_tmp")]; + tensor res1_conv1_prev_tmp = identity(x = res1_conv1_prev)[name = tensor("res1_conv1_prev_tmp")]; + tensor res1_conv1_first_tmp = identity(x = res1_conv1_first)[name = tensor("res1_conv1_first_tmp")]; + tensor res2_conv0_first_tmp = identity(x = res2_conv0_first)[name = tensor("res2_conv0_first_tmp")]; + tensor res2_conv1_prev_tmp = identity(x = res2_conv1_prev)[name = tensor("res2_conv1_prev_tmp")]; + tensor res2_conv1_first_tmp = identity(x = res2_conv1_first)[name = tensor("res2_conv1_first_tmp")]; + tensor conv_final_first_tmp = identity(x = conv_final_first)[name = tensor("conv_final_first_tmp")]; + } -> (var_724, var_77, var_210, var_741, var_400, var_744, var_542, conv0_first, var_565, var_585, res0_conv0_first, res0_conv1_prev, res0_conv1_first, var_621, var_641, res1_conv0_first, res1_conv1_prev, res1_conv1_first, var_677, var_697, res2_conv0_first, res2_conv1_prev, res2_conv1_first, var_725, conv_final_first); +} \ No newline at end of file diff --git a/v2.1/portuguese_24l/mimi_decoder.mlmodelc/weights/weight.bin b/v2.1/portuguese_24l/mimi_decoder.mlmodelc/weights/weight.bin new file mode 100644 index 0000000000000000000000000000000000000000..1370ffd7f5e8989e5db9d220d47f098f2b1235d8 --- /dev/null +++ b/v2.1/portuguese_24l/mimi_decoder.mlmodelc/weights/weight.bin @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:1d5e99ea6732f93fdf15a439fb22d32172021a27cc779fd32b77f79f06a286b0 +size 41768256 diff --git a/v2.1/portuguese_24l/mimi_decoder.mlpackage/Data/com.apple.CoreML/model.mlmodel b/v2.1/portuguese_24l/mimi_decoder.mlpackage/Data/com.apple.CoreML/model.mlmodel new file mode 100644 index 0000000000000000000000000000000000000000..36408047c393b2ff6940bda3497682fd64093337 --- /dev/null +++ b/v2.1/portuguese_24l/mimi_decoder.mlpackage/Data/com.apple.CoreML/model.mlmodel @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:032d927ad265813bb42038940d6e9fa755f9a5e049946f1af6251acab49201e9 +size 93270 diff --git a/v2.1/portuguese_24l/mimi_decoder.mlpackage/Data/com.apple.CoreML/weights/weight.bin b/v2.1/portuguese_24l/mimi_decoder.mlpackage/Data/com.apple.CoreML/weights/weight.bin new file mode 100644 index 0000000000000000000000000000000000000000..1370ffd7f5e8989e5db9d220d47f098f2b1235d8 --- /dev/null +++ b/v2.1/portuguese_24l/mimi_decoder.mlpackage/Data/com.apple.CoreML/weights/weight.bin @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:1d5e99ea6732f93fdf15a439fb22d32172021a27cc779fd32b77f79f06a286b0 +size 41768256 diff --git a/v2.1/portuguese_24l/mimi_decoder.mlpackage/Manifest.json b/v2.1/portuguese_24l/mimi_decoder.mlpackage/Manifest.json new file mode 100644 index 0000000000000000000000000000000000000000..653ad47142a65155672b885daeb15652260fa045 --- /dev/null +++ b/v2.1/portuguese_24l/mimi_decoder.mlpackage/Manifest.json @@ -0,0 +1,18 @@ +{ + "fileFormatVersion": "1.0.0", + "itemInfoEntries": { + "5ACF7B47-D054-4415-8D60-B391C5CD4F03": { + "author": "com.apple.CoreML", + "description": "CoreML Model Specification", + "name": "model.mlmodel", + "path": "com.apple.CoreML/model.mlmodel" + }, + "99D8D468-7B0A-4169-94F5-441CE58FF972": { + "author": "com.apple.CoreML", + "description": "CoreML Model Weights", + "name": "weights", + "path": "com.apple.CoreML/weights" + } + }, + "rootModelIdentifier": "5ACF7B47-D054-4415-8D60-B391C5CD4F03" +} diff --git a/v2.1/spanish/cond_prefill.mlmodelc/analytics/coremldata.bin b/v2.1/spanish/cond_prefill.mlmodelc/analytics/coremldata.bin new file mode 100644 index 0000000000000000000000000000000000000000..4f59be7a3e7c35d5ad31d534613149b56e956172 --- /dev/null +++ b/v2.1/spanish/cond_prefill.mlmodelc/analytics/coremldata.bin @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:c62b80986ff07c242fa4236e3a6b98037a2947defa214c94990f17f2346c1356 +size 243 diff --git a/v2.1/spanish/cond_prefill.mlmodelc/coremldata.bin b/v2.1/spanish/cond_prefill.mlmodelc/coremldata.bin new file mode 100644 index 0000000000000000000000000000000000000000..8362e094e82b76a2165a8d2891c46e4a5da15a27 --- /dev/null +++ b/v2.1/spanish/cond_prefill.mlmodelc/coremldata.bin @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:ace0e1dbc5695b2631f2eda4d476d2744be7e2c67d8f365aa54a82d7c5093ee1 +size 1151 diff --git a/v2.1/spanish/cond_prefill.mlmodelc/model.mil b/v2.1/spanish/cond_prefill.mlmodelc/model.mil new file mode 100644 index 0000000000000000000000000000000000000000..203b091a2c5043d4b8c07f4402d5a5f0319b3c3f --- /dev/null +++ b/v2.1/spanish/cond_prefill.mlmodelc/model.mil @@ -0,0 +1,1309 @@ +program(1.0) +[buildInfo = dict, tensor>({{"coremlc-component-MIL", "3520.4.1"}, {"coremlc-version", "3520.5.1"}, {"coremltools-component-torch", "2.12.0"}, {"coremltools-source-dialect", "TorchScript"}, {"coremltools-version", "9.0"}})] +{ + func main(tensor cache0, tensor cache1, tensor cache2, tensor cache3, tensor cache4, tensor cache5, tensor conditioning, tensor position0, tensor position1, tensor position2, tensor position3, tensor position4, tensor position5, tensor valid_len) { + tensor input_3_axes_0 = const()[name = tensor("input_3_axes_0"), val = tensor([-1])]; + tensor conditioning_to_fp16_dtype_0 = const()[name = tensor("conditioning_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor norm0_1_weight_to_fp16 = const()[name = tensor("norm0_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(64)))]; + tensor norm0_1_bias_to_fp16 = const()[name = tensor("norm0_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(2176)))]; + tensor var_48_to_fp16 = const()[name = tensor("op_48_to_fp16"), val = tensor(0x1.5p-17)]; + tensor conditioning_to_fp16 = cast(dtype = conditioning_to_fp16_dtype_0, x = conditioning)[name = tensor("cast_159")]; + tensor input_3_cast_fp16 = layer_norm(axes = input_3_axes_0, beta = norm0_1_bias_to_fp16, epsilon = var_48_to_fp16, gamma = norm0_1_weight_to_fp16, x = conditioning_to_fp16)[name = tensor("input_3_cast_fp16")]; + tensor attn0_in_proj_weight_to_fp16 = const()[name = tensor("attn0_in_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(4288)))]; + tensor linear_0_bias_0_to_fp16 = const()[name = tensor("linear_0_bias_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(6295808)))]; + tensor linear_0_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = attn0_in_proj_weight_to_fp16, x = input_3_cast_fp16)[name = tensor("linear_0_cast_fp16")]; + tensor var_62 = const()[name = tensor("op_62"), val = tensor([1, 256, 3, 16, 64])]; + tensor qkv_1_cast_fp16 = reshape(shape = var_62, x = linear_0_cast_fp16)[name = tensor("qkv_1_cast_fp16")]; + tensor q_1_begin_0 = const()[name = tensor("q_1_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor q_1_end_0 = const()[name = tensor("q_1_end_0"), val = tensor([1, 256, 1, 16, 64])]; + tensor q_1_end_mask_0 = const()[name = tensor("q_1_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor q_1_squeeze_mask_0 = const()[name = tensor("q_1_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor q_1_cast_fp16 = slice_by_index(begin = q_1_begin_0, end = q_1_end_0, end_mask = q_1_end_mask_0, squeeze_mask = q_1_squeeze_mask_0, x = qkv_1_cast_fp16)[name = tensor("q_1_cast_fp16")]; + tensor k_1_begin_0 = const()[name = tensor("k_1_begin_0"), val = tensor([0, 0, 1, 0, 0])]; + tensor k_1_end_0 = const()[name = tensor("k_1_end_0"), val = tensor([1, 256, 2, 16, 64])]; + tensor k_1_end_mask_0 = const()[name = tensor("k_1_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor k_1_squeeze_mask_0 = const()[name = tensor("k_1_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor k_1_cast_fp16 = slice_by_index(begin = k_1_begin_0, end = k_1_end_0, end_mask = k_1_end_mask_0, squeeze_mask = k_1_squeeze_mask_0, x = qkv_1_cast_fp16)[name = tensor("k_1_cast_fp16")]; + tensor v_1_begin_0 = const()[name = tensor("v_1_begin_0"), val = tensor([0, 0, 2, 0, 0])]; + tensor v_1_end_0 = const()[name = tensor("v_1_end_0"), val = tensor([1, 256, 3, 16, 64])]; + tensor v_1_end_mask_0 = const()[name = tensor("v_1_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor v_1_squeeze_mask_0 = const()[name = tensor("v_1_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor v_1_cast_fp16 = slice_by_index(begin = v_1_begin_0, end = v_1_end_0, end_mask = v_1_end_mask_0, squeeze_mask = v_1_squeeze_mask_0, x = qkv_1_cast_fp16)[name = tensor("v_1_cast_fp16")]; + tensor var_120 = const()[name = tensor("op_120"), val = tensor([1, 1])]; + tensor position0_to_fp16_dtype_0 = const()[name = tensor("position0_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor position0_to_fp16 = cast(dtype = position0_to_fp16_dtype_0, x = position0)[name = tensor("cast_158")]; + tensor var_121_cast_fp16 = reshape(shape = var_120, x = position0_to_fp16)[name = tensor("op_121_cast_fp16")]; + tensor ts_1_promoted_to_fp16 = const()[name = tensor("ts_1_promoted_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(6302016)))]; + tensor ts_3_cast_fp16 = add(x = ts_1_promoted_to_fp16, y = var_121_cast_fp16)[name = tensor("ts_3_cast_fp16")]; + tensor var_128 = const()[name = tensor("op_128"), val = tensor([1, 256, 1, 1])]; + tensor ts_5_cast_fp16 = reshape(shape = var_128, x = ts_3_cast_fp16)[name = tensor("ts_5_cast_fp16")]; + tensor var_135 = const()[name = tensor("op_135"), val = tensor([1, 256, 16, 32, 2])]; + tensor q_complex_1_cast_fp16 = reshape(shape = var_135, x = q_1_cast_fp16)[name = tensor("q_complex_1_cast_fp16")]; + tensor var_142 = const()[name = tensor("op_142"), val = tensor([1, 256, 16, 32, 2])]; + tensor k_complex_1_cast_fp16 = reshape(shape = var_142, x = k_1_cast_fp16)[name = tensor("k_complex_1_cast_fp16")]; + tensor var_146_begin_0 = const()[name = tensor("op_146_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_146_end_0 = const()[name = tensor("op_146_end_0"), val = tensor([1, 256, 16, 32, 1])]; + tensor var_146_end_mask_0 = const()[name = tensor("op_146_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_146_squeeze_mask_0 = const()[name = tensor("op_146_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_146_cast_fp16 = slice_by_index(begin = var_146_begin_0, end = var_146_end_0, end_mask = var_146_end_mask_0, squeeze_mask = var_146_squeeze_mask_0, x = q_complex_1_cast_fp16)[name = tensor("op_146_cast_fp16")]; + tensor var_154_begin_0 = const()[name = tensor("op_154_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_154_end_0 = const()[name = tensor("op_154_end_0"), val = tensor([1, 256, 16, 32, 2])]; + tensor var_154_end_mask_0 = const()[name = tensor("op_154_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_154_squeeze_mask_0 = const()[name = tensor("op_154_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_154_cast_fp16 = slice_by_index(begin = var_154_begin_0, end = var_154_end_0, end_mask = var_154_end_mask_0, squeeze_mask = var_154_squeeze_mask_0, x = q_complex_1_cast_fp16)[name = tensor("op_154_cast_fp16")]; + tensor var_162_begin_0 = const()[name = tensor("op_162_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_162_end_0 = const()[name = tensor("op_162_end_0"), val = tensor([1, 256, 16, 32, 1])]; + tensor var_162_end_mask_0 = const()[name = tensor("op_162_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_162_squeeze_mask_0 = const()[name = tensor("op_162_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_162_cast_fp16 = slice_by_index(begin = var_162_begin_0, end = var_162_end_0, end_mask = var_162_end_mask_0, squeeze_mask = var_162_squeeze_mask_0, x = k_complex_1_cast_fp16)[name = tensor("op_162_cast_fp16")]; + tensor var_170_begin_0 = const()[name = tensor("op_170_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_170_end_0 = const()[name = tensor("op_170_end_0"), val = tensor([1, 256, 16, 32, 2])]; + tensor var_170_end_mask_0 = const()[name = tensor("op_170_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_170_squeeze_mask_0 = const()[name = tensor("op_170_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_170_cast_fp16 = slice_by_index(begin = var_170_begin_0, end = var_170_end_0, end_mask = var_170_end_mask_0, squeeze_mask = var_170_squeeze_mask_0, x = k_complex_1_cast_fp16)[name = tensor("op_170_cast_fp16")]; + tensor freqs_1_to_fp16 = const()[name = tensor("freqs_1_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(6302592)))]; + tensor var_176_cast_fp16 = mul(x = freqs_1_to_fp16, y = ts_5_cast_fp16)[name = tensor("op_176_cast_fp16")]; + tensor rotr_1_cast_fp16 = cos(x = var_176_cast_fp16)[name = tensor("rotr_1_cast_fp16")]; + tensor roti_1_cast_fp16 = sin(x = var_176_cast_fp16)[name = tensor("roti_1_cast_fp16")]; + tensor var_180_cast_fp16 = mul(x = var_146_cast_fp16, y = rotr_1_cast_fp16)[name = tensor("op_180_cast_fp16")]; + tensor var_181_cast_fp16 = mul(x = var_154_cast_fp16, y = roti_1_cast_fp16)[name = tensor("op_181_cast_fp16")]; + tensor qor_1_cast_fp16 = sub(x = var_180_cast_fp16, y = var_181_cast_fp16)[name = tensor("qor_1_cast_fp16")]; + tensor var_184_cast_fp16 = mul(x = var_146_cast_fp16, y = roti_1_cast_fp16)[name = tensor("op_184_cast_fp16")]; + tensor var_185_cast_fp16 = mul(x = var_154_cast_fp16, y = rotr_1_cast_fp16)[name = tensor("op_185_cast_fp16")]; + tensor qoi_1_cast_fp16 = add(x = var_184_cast_fp16, y = var_185_cast_fp16)[name = tensor("qoi_1_cast_fp16")]; + tensor var_188_cast_fp16 = mul(x = var_162_cast_fp16, y = rotr_1_cast_fp16)[name = tensor("op_188_cast_fp16")]; + tensor var_189_cast_fp16 = mul(x = var_170_cast_fp16, y = roti_1_cast_fp16)[name = tensor("op_189_cast_fp16")]; + tensor kor_1_cast_fp16 = sub(x = var_188_cast_fp16, y = var_189_cast_fp16)[name = tensor("kor_1_cast_fp16")]; + tensor var_192_cast_fp16 = mul(x = var_162_cast_fp16, y = roti_1_cast_fp16)[name = tensor("op_192_cast_fp16")]; + tensor var_193_cast_fp16 = mul(x = var_170_cast_fp16, y = rotr_1_cast_fp16)[name = tensor("op_193_cast_fp16")]; + tensor koi_1_cast_fp16 = add(x = var_192_cast_fp16, y = var_193_cast_fp16)[name = tensor("koi_1_cast_fp16")]; + tensor qo_1_axis_0 = const()[name = tensor("qo_1_axis_0"), val = tensor(-1)]; + tensor qo_1_cast_fp16 = stack(axis = qo_1_axis_0, values = (qor_1_cast_fp16, qoi_1_cast_fp16))[name = tensor("qo_1_cast_fp16")]; + tensor ko_1_axis_0 = const()[name = tensor("ko_1_axis_0"), val = tensor(-1)]; + tensor ko_1_cast_fp16 = stack(axis = ko_1_axis_0, values = (kor_1_cast_fp16, koi_1_cast_fp16))[name = tensor("ko_1_cast_fp16")]; + tensor var_226 = const()[name = tensor("op_226"), val = tensor([1, 256, 16, 64])]; + tensor q_3_cast_fp16 = reshape(shape = var_226, x = qo_1_cast_fp16)[name = tensor("q_3_cast_fp16")]; + tensor var_232 = const()[name = tensor("op_232"), val = tensor([1, 256, 16, 64])]; + tensor k_3_cast_fp16 = reshape(shape = var_232, x = ko_1_cast_fp16)[name = tensor("k_3_cast_fp16")]; + tensor var_249_promoted_to_fp16 = const()[name = tensor("op_249_promoted_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(6302720)))]; + tensor write_indices_float_1_cast_fp16 = add(x = var_121_cast_fp16, y = var_249_promoted_to_fp16)[name = tensor("write_indices_float_1_cast_fp16")]; + tensor _inversed_253_y_0_to_fp16 = const()[name = tensor("_inversed_253_y_0_to_fp16"), val = tensor(0x1p-9)]; + tensor _inversed_253_cast_fp16 = mul(x = write_indices_float_1_cast_fp16, y = _inversed_253_y_0_to_fp16)[name = tensor("_inversed_253_cast_fp16")]; + tensor var_254_cast_fp16 = floor(x = _inversed_253_cast_fp16)[name = tensor("op_254_cast_fp16")]; + tensor var_255_to_fp16 = const()[name = tensor("op_255_to_fp16"), val = tensor(0x1p+9)]; + tensor var_256_cast_fp16 = mul(x = var_254_cast_fp16, y = var_255_to_fp16)[name = tensor("op_256_cast_fp16")]; + tensor write_indices_float_3_cast_fp16 = sub(x = write_indices_float_1_cast_fp16, y = var_256_cast_fp16)[name = tensor("write_indices_float_3_cast_fp16")]; + tensor var_265 = const()[name = tensor("op_265"), val = tensor([1, 1])]; + tensor valid_len_to_fp16_dtype_0 = const()[name = tensor("valid_len_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor valid_len_to_fp16 = cast(dtype = valid_len_to_fp16_dtype_0, x = valid_len)[name = tensor("cast_157")]; + tensor var_266_cast_fp16 = reshape(shape = var_265, x = valid_len_to_fp16)[name = tensor("op_266_cast_fp16")]; + tensor token_valid_1_cast_fp16 = less(x = var_249_promoted_to_fp16, y = var_266_cast_fp16)[name = tensor("token_valid_1_cast_fp16")]; + tensor var_274_value_0_to_fp16 = const()[name = tensor("op_274_value_0_to_fp16"), val = tensor(0x1.ffp+8)]; + tensor var_274_cast_fp16 = fill_like(ref_tensor = write_indices_float_3_cast_fp16, value = var_274_value_0_to_fp16)[name = tensor("op_274_cast_fp16")]; + tensor write_indices_float_5_cast_fp16 = select(a = write_indices_float_3_cast_fp16, b = var_274_cast_fp16, cond = token_valid_1_cast_fp16)[name = tensor("write_indices_float_5_cast_fp16")]; + tensor var_280_dtype_0 = const()[name = tensor("op_280_dtype_0"), val = tensor("int32")]; + tensor var_285 = const()[name = tensor("op_285"), val = tensor([1, 256, 1, 1])]; + tensor write_indices_float_5_cast_fp16_to_int32 = cast(dtype = var_280_dtype_0, x = write_indices_float_5_cast_fp16)[name = tensor("cast_156")]; + tensor var_286 = reshape(shape = var_285, x = write_indices_float_5_cast_fp16_to_int32)[name = tensor("op_286")]; + tensor write_indices_1_reps_0 = const()[name = tensor("write_indices_1_reps_0"), val = tensor([1, 1, 16, 64])]; + tensor write_indices_1 = tile(reps = write_indices_1_reps_0, x = var_286)[name = tensor("write_indices_1")]; + tensor var_296_begin_0 = const()[name = tensor("op_296_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_296_end_0 = const()[name = tensor("op_296_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor var_296_end_mask_0 = const()[name = tensor("op_296_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_296_squeeze_mask_0 = const()[name = tensor("op_296_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor cache0_to_fp16_dtype_0 = const()[name = tensor("cache0_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor cache0_to_fp16 = cast(dtype = cache0_to_fp16_dtype_0, x = cache0)[name = tensor("cast_155")]; + tensor var_296_cast_fp16 = slice_by_index(begin = var_296_begin_0, end = var_296_end_0, end_mask = var_296_end_mask_0, squeeze_mask = var_296_squeeze_mask_0, x = cache0_to_fp16)[name = tensor("op_296_cast_fp16")]; + tensor var_298_axis_0 = const()[name = tensor("op_298_axis_0"), val = tensor(1)]; + tensor var_298_mode_0 = const()[name = tensor("op_298_mode_0"), val = tensor("update")]; + tensor var_298_validate_indices_0 = const()[name = tensor("op_298_validate_indices_0"), val = tensor(false)]; + tensor var_298_cast_fp16 = scatter_along_axis(axis = var_298_axis_0, data = var_296_cast_fp16, indices = write_indices_1, mode = var_298_mode_0, updates = k_3_cast_fp16, validate_indices = var_298_validate_indices_0)[name = tensor("op_298_cast_fp16")]; + tensor concat_2 = const()[name = tensor("concat_2"), val = tensor([0, 0, 0, 0, 0])]; + tensor concat_3 = const()[name = tensor("concat_3"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_1_internal_tensor_assign_1_stride_0 = const()[name = tensor("new_cache_1_internal_tensor_assign_1_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_1_internal_tensor_assign_1_begin_mask_0 = const()[name = tensor("new_cache_1_internal_tensor_assign_1_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_1_internal_tensor_assign_1_end_mask_0 = const()[name = tensor("new_cache_1_internal_tensor_assign_1_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_1_internal_tensor_assign_1_squeeze_mask_0 = const()[name = tensor("new_cache_1_internal_tensor_assign_1_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_10 = const()[name = tensor("shape_10"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_0 = const()[name = tensor("reduce_prod_0"), val = tensor(1048576)]; + tensor range_1d_0_start_0 = const()[name = tensor("range_1d_0_start_0"), val = tensor(0)]; + tensor range_1d_0_step_0 = const()[name = tensor("range_1d_0_step_0"), val = tensor(1)]; + tensor range_1d_0 = range_1d(end = reduce_prod_0, start = range_1d_0_start_0, step = range_1d_0_step_0)[name = tensor("range_1d_0")]; + tensor reshape_0 = reshape(shape = shape_10, x = range_1d_0)[name = tensor("reshape_0")]; + tensor slice_by_index_0 = slice_by_index(begin = concat_2, begin_mask = new_cache_1_internal_tensor_assign_1_begin_mask_0, end = concat_3, end_mask = new_cache_1_internal_tensor_assign_1_end_mask_0, squeeze_mask = new_cache_1_internal_tensor_assign_1_squeeze_mask_0, stride = new_cache_1_internal_tensor_assign_1_stride_0, x = reshape_0)[name = tensor("slice_by_index_0")]; + tensor reshape_1_shape_0 = const()[name = tensor("reshape_1_shape_0"), val = tensor([-1])]; + tensor reshape_1 = reshape(shape = reshape_1_shape_0, x = slice_by_index_0)[name = tensor("reshape_1")]; + tensor reshape_2_shape_0 = const()[name = tensor("reshape_2_shape_0"), val = tensor([-1])]; + tensor reshape_2_cast_fp16 = reshape(shape = reshape_2_shape_0, x = var_298_cast_fp16)[name = tensor("reshape_2_cast_fp16")]; + tensor reshape_3_shape_0 = const()[name = tensor("reshape_3_shape_0"), val = tensor([-1])]; + tensor reshape_3_cast_fp16 = reshape(shape = reshape_3_shape_0, x = cache0_to_fp16)[name = tensor("reshape_3_cast_fp16")]; + tensor scatter_0_mode_0 = const()[name = tensor("scatter_0_mode_0"), val = tensor("update")]; + tensor scatter_0_axis_0 = const()[name = tensor("scatter_0_axis_0"), val = tensor(0)]; + tensor scatter_0_validate_indices_0 = const()[name = tensor("scatter_0_validate_indices_0"), val = tensor(false)]; + tensor scatter_0_cast_fp16 = scatter(axis = scatter_0_axis_0, data = reshape_3_cast_fp16, indices = reshape_1, mode = scatter_0_mode_0, updates = reshape_2_cast_fp16, validate_indices = scatter_0_validate_indices_0)[name = tensor("scatter_0_cast_fp16")]; + tensor reshape_4_cast_fp16 = reshape(shape = shape_10, x = scatter_0_cast_fp16)[name = tensor("reshape_4_cast_fp16")]; + tensor var_306_begin_0 = const()[name = tensor("op_306_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor var_306_end_0 = const()[name = tensor("op_306_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor var_306_end_mask_0 = const()[name = tensor("op_306_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_306_squeeze_mask_0 = const()[name = tensor("op_306_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor var_306_cast_fp16 = slice_by_index(begin = var_306_begin_0, end = var_306_end_0, end_mask = var_306_end_mask_0, squeeze_mask = var_306_squeeze_mask_0, x = reshape_4_cast_fp16)[name = tensor("op_306_cast_fp16")]; + tensor var_308_axis_0 = const()[name = tensor("op_308_axis_0"), val = tensor(1)]; + tensor var_308_mode_0 = const()[name = tensor("op_308_mode_0"), val = tensor("update")]; + tensor var_308_validate_indices_0 = const()[name = tensor("op_308_validate_indices_0"), val = tensor(false)]; + tensor var_308_cast_fp16 = scatter_along_axis(axis = var_308_axis_0, data = var_306_cast_fp16, indices = write_indices_1, mode = var_308_mode_0, updates = v_1_cast_fp16, validate_indices = var_308_validate_indices_0)[name = tensor("op_308_cast_fp16")]; + tensor concat_4 = const()[name = tensor("concat_4"), val = tensor([1, 0, 0, 0, 0])]; + tensor concat_5 = const()[name = tensor("concat_5"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_1_internal_tensor_assign_2_stride_0 = const()[name = tensor("new_cache_1_internal_tensor_assign_2_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_1_internal_tensor_assign_2_begin_mask_0 = const()[name = tensor("new_cache_1_internal_tensor_assign_2_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_1_internal_tensor_assign_2_end_mask_0 = const()[name = tensor("new_cache_1_internal_tensor_assign_2_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_1_internal_tensor_assign_2_squeeze_mask_0 = const()[name = tensor("new_cache_1_internal_tensor_assign_2_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_11 = const()[name = tensor("shape_11"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_1 = const()[name = tensor("reduce_prod_1"), val = tensor(1048576)]; + tensor range_1d_1_start_0 = const()[name = tensor("range_1d_1_start_0"), val = tensor(0)]; + tensor range_1d_1_step_0 = const()[name = tensor("range_1d_1_step_0"), val = tensor(1)]; + tensor range_1d_1 = range_1d(end = reduce_prod_1, start = range_1d_1_start_0, step = range_1d_1_step_0)[name = tensor("range_1d_1")]; + tensor reshape_5 = reshape(shape = shape_11, x = range_1d_1)[name = tensor("reshape_5")]; + tensor slice_by_index_1 = slice_by_index(begin = concat_4, begin_mask = new_cache_1_internal_tensor_assign_2_begin_mask_0, end = concat_5, end_mask = new_cache_1_internal_tensor_assign_2_end_mask_0, squeeze_mask = new_cache_1_internal_tensor_assign_2_squeeze_mask_0, stride = new_cache_1_internal_tensor_assign_2_stride_0, x = reshape_5)[name = tensor("slice_by_index_1")]; + tensor reshape_6_shape_0 = const()[name = tensor("reshape_6_shape_0"), val = tensor([-1])]; + tensor reshape_6 = reshape(shape = reshape_6_shape_0, x = slice_by_index_1)[name = tensor("reshape_6")]; + tensor reshape_7_shape_0 = const()[name = tensor("reshape_7_shape_0"), val = tensor([-1])]; + tensor reshape_7_cast_fp16 = reshape(shape = reshape_7_shape_0, x = var_308_cast_fp16)[name = tensor("reshape_7_cast_fp16")]; + tensor reshape_8_shape_0 = const()[name = tensor("reshape_8_shape_0"), val = tensor([-1])]; + tensor reshape_8_cast_fp16 = reshape(shape = reshape_8_shape_0, x = reshape_4_cast_fp16)[name = tensor("reshape_8_cast_fp16")]; + tensor scatter_1_mode_0 = const()[name = tensor("scatter_1_mode_0"), val = tensor("update")]; + tensor scatter_1_axis_0 = const()[name = tensor("scatter_1_axis_0"), val = tensor(0)]; + tensor scatter_1_validate_indices_0 = const()[name = tensor("scatter_1_validate_indices_0"), val = tensor(false)]; + tensor scatter_1_cast_fp16 = scatter(axis = scatter_1_axis_0, data = reshape_8_cast_fp16, indices = reshape_6, mode = scatter_1_mode_0, updates = reshape_7_cast_fp16, validate_indices = scatter_1_validate_indices_0)[name = tensor("scatter_1_cast_fp16")]; + tensor reshape_9_cast_fp16 = reshape(shape = shape_11, x = scatter_1_cast_fp16)[name = tensor("reshape_9_cast_fp16")]; + tensor reshape_9_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("reshape_9_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor keys_1_begin_0 = const()[name = tensor("keys_1_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor keys_1_end_0 = const()[name = tensor("keys_1_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor keys_1_end_mask_0 = const()[name = tensor("keys_1_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor keys_1_squeeze_mask_0 = const()[name = tensor("keys_1_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor keys_1_cast_fp16 = slice_by_index(begin = keys_1_begin_0, end = keys_1_end_0, end_mask = keys_1_end_mask_0, squeeze_mask = keys_1_squeeze_mask_0, x = reshape_9_cast_fp16)[name = tensor("keys_1_cast_fp16")]; + tensor values_1_begin_0 = const()[name = tensor("values_1_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor values_1_end_0 = const()[name = tensor("values_1_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor values_1_end_mask_0 = const()[name = tensor("values_1_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor values_1_squeeze_mask_0 = const()[name = tensor("values_1_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor values_1_cast_fp16 = slice_by_index(begin = values_1_begin_0, end = values_1_end_0, end_mask = values_1_end_mask_0, squeeze_mask = values_1_squeeze_mask_0, x = reshape_9_cast_fp16)[name = tensor("values_1_cast_fp16")]; + tensor var_320_cast_fp16 = not_equal(x = keys_1_cast_fp16, y = keys_1_cast_fp16)[name = tensor("op_320_cast_fp16")]; + tensor var_326_to_fp16 = const()[name = tensor("op_326_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(6303296)))]; + tensor keys_3_cast_fp16 = select(a = var_326_to_fp16, b = keys_1_cast_fp16, cond = var_320_cast_fp16)[name = tensor("keys_3_cast_fp16")]; + tensor var_328_cast_fp16 = not_equal(x = values_1_cast_fp16, y = values_1_cast_fp16)[name = tensor("op_328_cast_fp16")]; + tensor values_3_cast_fp16 = select(a = var_326_to_fp16, b = values_1_cast_fp16, cond = var_328_cast_fp16)[name = tensor("values_3_cast_fp16")]; + tensor var_352 = const()[name = tensor("op_352"), val = tensor([0, 2, 1, 3])]; + tensor var_368 = const()[name = tensor("op_368"), val = tensor([1, 1, 1])]; + tensor var_369_cast_fp16 = reshape(shape = var_368, x = position0_to_fp16)[name = tensor("op_369_cast_fp16")]; + tensor q_offsets_1_promoted_to_fp16 = const()[name = tensor("q_offsets_1_promoted_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(7351936)))]; + tensor q_positions_1_cast_fp16 = add(x = var_369_cast_fp16, y = q_offsets_1_promoted_to_fp16)[name = tensor("q_positions_1_cast_fp16")]; + tensor var_391 = const()[name = tensor("op_391"), val = tensor([1, 1, 1])]; + tensor var_392_cast_fp16 = reshape(shape = var_391, x = valid_len_to_fp16)[name = tensor("op_392_cast_fp16")]; + tensor valid_end_1_cast_fp16 = add(x = var_369_cast_fp16, y = var_392_cast_fp16)[name = tensor("valid_end_1_cast_fp16")]; + tensor k_positions_1_promoted_to_fp16 = const()[name = tensor("k_positions_1_promoted_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(7352512)))]; + tensor valid_mask_1_cast_fp16 = less(x = k_positions_1_promoted_to_fp16, y = valid_end_1_cast_fp16)[name = tensor("valid_mask_1_cast_fp16")]; + tensor causal_mask_1_cast_fp16 = less_equal(x = k_positions_1_promoted_to_fp16, y = q_positions_1_cast_fp16)[name = tensor("causal_mask_1_cast_fp16")]; + tensor var_397 = logical_and(x = valid_mask_1_cast_fp16, y = causal_mask_1_cast_fp16)[name = tensor("op_397")]; + tensor attn_mask_1_axes_0 = const()[name = tensor("attn_mask_1_axes_0"), val = tensor([1])]; + tensor attn_mask_1 = expand_dims(axes = attn_mask_1_axes_0, x = var_397)[name = tensor("attn_mask_1")]; + tensor var_411_transpose_x_0 = const()[name = tensor("op_411_transpose_x_0"), val = tensor(false)]; + tensor var_411_transpose_y_0 = const()[name = tensor("op_411_transpose_y_0"), val = tensor(false)]; + tensor transpose_15_perm_0 = const()[name = tensor("transpose_15_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_16_perm_0 = const()[name = tensor("transpose_16_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_16 = transpose(perm = transpose_16_perm_0, x = keys_3_cast_fp16)[name = tensor("transpose_42")]; + tensor transpose_15 = transpose(perm = transpose_15_perm_0, x = q_3_cast_fp16)[name = tensor("transpose_43")]; + tensor var_411_cast_fp16 = matmul(transpose_x = var_411_transpose_x_0, transpose_y = var_411_transpose_y_0, x = transpose_15, y = transpose_16)[name = tensor("op_411_cast_fp16")]; + tensor var_405_to_fp16 = const()[name = tensor("op_405_to_fp16"), val = tensor([0x1p-3])]; + tensor attn_weights_1_cast_fp16 = mul(x = var_411_cast_fp16, y = var_405_to_fp16)[name = tensor("attn_weights_1_cast_fp16")]; + tensor var_413 = logical_not(x = attn_mask_1)[name = tensor("op_413")]; + tensor var_414_to_fp16 = const()[name = tensor("op_414_to_fp16"), val = tensor(-inf)]; + tensor attn_weights_3_cast_fp16 = select(a = var_414_to_fp16, b = attn_weights_1_cast_fp16, cond = var_413)[name = tensor("attn_weights_3_cast_fp16")]; + tensor var_416 = const()[name = tensor("op_416"), val = tensor(-1)]; + tensor attn_weights_5_cast_fp16 = softmax(axis = var_416, x = attn_weights_3_cast_fp16)[name = tensor("attn_weights_5_cast_fp16")]; + tensor attn_output_1_transpose_x_0 = const()[name = tensor("attn_output_1_transpose_x_0"), val = tensor(false)]; + tensor attn_output_1_transpose_y_0 = const()[name = tensor("attn_output_1_transpose_y_0"), val = tensor(false)]; + tensor values_5_cast_fp16 = transpose(perm = var_352, x = values_3_cast_fp16)[name = tensor("transpose_44")]; + tensor attn_output_1_cast_fp16 = matmul(transpose_x = attn_output_1_transpose_x_0, transpose_y = attn_output_1_transpose_y_0, x = attn_weights_5_cast_fp16, y = values_5_cast_fp16)[name = tensor("attn_output_1_cast_fp16")]; + tensor var_424 = const()[name = tensor("op_424"), val = tensor([0, 2, 1, 3])]; + tensor var_429 = const()[name = tensor("op_429"), val = tensor([1, 256, 1024])]; + tensor var_425_cast_fp16 = transpose(perm = var_424, x = attn_output_1_cast_fp16)[name = tensor("transpose_41")]; + tensor input_5_cast_fp16 = reshape(shape = var_429, x = var_425_cast_fp16)[name = tensor("input_5_cast_fp16")]; + tensor attn0_out_proj_weight_to_fp16 = const()[name = tensor("attn0_out_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(7353600)))]; + tensor linear_1_bias_0_to_fp16 = const()[name = tensor("linear_1_bias_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9450816)))]; + tensor linear_1_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = attn0_out_proj_weight_to_fp16, x = input_5_cast_fp16)[name = tensor("linear_1_cast_fp16")]; + tensor var_438_cast_fp16 = add(x = position0_to_fp16, y = valid_len_to_fp16)[name = tensor("op_438_cast_fp16")]; + tensor var_438_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_438_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor input_7_cast_fp16 = add(x = conditioning_to_fp16, y = linear_1_cast_fp16)[name = tensor("input_7_cast_fp16")]; + tensor input_9_axes_0 = const()[name = tensor("input_9_axes_0"), val = tensor([-1])]; + tensor norm0_2_weight_to_fp16 = const()[name = tensor("norm0_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9452928)))]; + tensor norm0_2_bias_to_fp16 = const()[name = tensor("norm0_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9455040)))]; + tensor var_442_to_fp16 = const()[name = tensor("op_442_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_9_cast_fp16 = layer_norm(axes = input_9_axes_0, beta = norm0_2_bias_to_fp16, epsilon = var_442_to_fp16, gamma = norm0_2_weight_to_fp16, x = input_7_cast_fp16)[name = tensor("input_9_cast_fp16")]; + tensor linear0_1_weight_to_fp16 = const()[name = tensor("linear0_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9457152)))]; + tensor linear_2_bias_0_to_fp16 = const()[name = tensor("linear_2_bias_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(17845824)))]; + tensor linear_2_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = linear0_1_weight_to_fp16, x = input_9_cast_fp16)[name = tensor("linear_2_cast_fp16")]; + tensor input_11_mode_0 = const()[name = tensor("input_11_mode_0"), val = tensor("EXACT")]; + tensor input_11_cast_fp16 = gelu(mode = input_11_mode_0, x = linear_2_cast_fp16)[name = tensor("input_11_cast_fp16")]; + tensor linear0_2_weight_to_fp16 = const()[name = tensor("linear0_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(17854080)))]; + tensor linear_3_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = linear0_2_weight_to_fp16, x = input_11_cast_fp16)[name = tensor("linear_3_cast_fp16")]; + tensor input_13_cast_fp16 = add(x = input_7_cast_fp16, y = linear_3_cast_fp16)[name = tensor("input_13_cast_fp16")]; + tensor input_15_axes_0 = const()[name = tensor("input_15_axes_0"), val = tensor([-1])]; + tensor norm1_1_weight_to_fp16 = const()[name = tensor("norm1_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(26242752)))]; + tensor norm1_1_bias_to_fp16 = const()[name = tensor("norm1_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(26244864)))]; + tensor var_459_to_fp16 = const()[name = tensor("op_459_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_15_cast_fp16 = layer_norm(axes = input_15_axes_0, beta = norm1_1_bias_to_fp16, epsilon = var_459_to_fp16, gamma = norm1_1_weight_to_fp16, x = input_13_cast_fp16)[name = tensor("input_15_cast_fp16")]; + tensor attn1_in_proj_weight_to_fp16 = const()[name = tensor("attn1_in_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(26246976)))]; + tensor linear_4_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = attn1_in_proj_weight_to_fp16, x = input_15_cast_fp16)[name = tensor("linear_4_cast_fp16")]; + tensor var_473 = const()[name = tensor("op_473"), val = tensor([1, 256, 3, 16, 64])]; + tensor qkv_3_cast_fp16 = reshape(shape = var_473, x = linear_4_cast_fp16)[name = tensor("qkv_3_cast_fp16")]; + tensor q_7_begin_0 = const()[name = tensor("q_7_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor q_7_end_0 = const()[name = tensor("q_7_end_0"), val = tensor([1, 256, 1, 16, 64])]; + tensor q_7_end_mask_0 = const()[name = tensor("q_7_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor q_7_squeeze_mask_0 = const()[name = tensor("q_7_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor q_7_cast_fp16 = slice_by_index(begin = q_7_begin_0, end = q_7_end_0, end_mask = q_7_end_mask_0, squeeze_mask = q_7_squeeze_mask_0, x = qkv_3_cast_fp16)[name = tensor("q_7_cast_fp16")]; + tensor k_5_begin_0 = const()[name = tensor("k_5_begin_0"), val = tensor([0, 0, 1, 0, 0])]; + tensor k_5_end_0 = const()[name = tensor("k_5_end_0"), val = tensor([1, 256, 2, 16, 64])]; + tensor k_5_end_mask_0 = const()[name = tensor("k_5_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor k_5_squeeze_mask_0 = const()[name = tensor("k_5_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor k_5_cast_fp16 = slice_by_index(begin = k_5_begin_0, end = k_5_end_0, end_mask = k_5_end_mask_0, squeeze_mask = k_5_squeeze_mask_0, x = qkv_3_cast_fp16)[name = tensor("k_5_cast_fp16")]; + tensor v_3_begin_0 = const()[name = tensor("v_3_begin_0"), val = tensor([0, 0, 2, 0, 0])]; + tensor v_3_end_0 = const()[name = tensor("v_3_end_0"), val = tensor([1, 256, 3, 16, 64])]; + tensor v_3_end_mask_0 = const()[name = tensor("v_3_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor v_3_squeeze_mask_0 = const()[name = tensor("v_3_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor v_3_cast_fp16 = slice_by_index(begin = v_3_begin_0, end = v_3_end_0, end_mask = v_3_end_mask_0, squeeze_mask = v_3_squeeze_mask_0, x = qkv_3_cast_fp16)[name = tensor("v_3_cast_fp16")]; + tensor var_531 = const()[name = tensor("op_531"), val = tensor([1, 1])]; + tensor position1_to_fp16_dtype_0 = const()[name = tensor("position1_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor position1_to_fp16 = cast(dtype = position1_to_fp16_dtype_0, x = position1)[name = tensor("cast_152")]; + tensor var_532_cast_fp16 = reshape(shape = var_531, x = position1_to_fp16)[name = tensor("op_532_cast_fp16")]; + tensor ts_9_cast_fp16 = add(x = ts_1_promoted_to_fp16, y = var_532_cast_fp16)[name = tensor("ts_9_cast_fp16")]; + tensor var_539 = const()[name = tensor("op_539"), val = tensor([1, 256, 1, 1])]; + tensor ts_11_cast_fp16 = reshape(shape = var_539, x = ts_9_cast_fp16)[name = tensor("ts_11_cast_fp16")]; + tensor var_546 = const()[name = tensor("op_546"), val = tensor([1, 256, 16, 32, 2])]; + tensor q_complex_3_cast_fp16 = reshape(shape = var_546, x = q_7_cast_fp16)[name = tensor("q_complex_3_cast_fp16")]; + tensor var_553 = const()[name = tensor("op_553"), val = tensor([1, 256, 16, 32, 2])]; + tensor k_complex_3_cast_fp16 = reshape(shape = var_553, x = k_5_cast_fp16)[name = tensor("k_complex_3_cast_fp16")]; + tensor var_557_begin_0 = const()[name = tensor("op_557_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_557_end_0 = const()[name = tensor("op_557_end_0"), val = tensor([1, 256, 16, 32, 1])]; + tensor var_557_end_mask_0 = const()[name = tensor("op_557_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_557_squeeze_mask_0 = const()[name = tensor("op_557_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_557_cast_fp16 = slice_by_index(begin = var_557_begin_0, end = var_557_end_0, end_mask = var_557_end_mask_0, squeeze_mask = var_557_squeeze_mask_0, x = q_complex_3_cast_fp16)[name = tensor("op_557_cast_fp16")]; + tensor var_565_begin_0 = const()[name = tensor("op_565_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_565_end_0 = const()[name = tensor("op_565_end_0"), val = tensor([1, 256, 16, 32, 2])]; + tensor var_565_end_mask_0 = const()[name = tensor("op_565_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_565_squeeze_mask_0 = const()[name = tensor("op_565_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_565_cast_fp16 = slice_by_index(begin = var_565_begin_0, end = var_565_end_0, end_mask = var_565_end_mask_0, squeeze_mask = var_565_squeeze_mask_0, x = q_complex_3_cast_fp16)[name = tensor("op_565_cast_fp16")]; + tensor var_573_begin_0 = const()[name = tensor("op_573_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_573_end_0 = const()[name = tensor("op_573_end_0"), val = tensor([1, 256, 16, 32, 1])]; + tensor var_573_end_mask_0 = const()[name = tensor("op_573_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_573_squeeze_mask_0 = const()[name = tensor("op_573_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_573_cast_fp16 = slice_by_index(begin = var_573_begin_0, end = var_573_end_0, end_mask = var_573_end_mask_0, squeeze_mask = var_573_squeeze_mask_0, x = k_complex_3_cast_fp16)[name = tensor("op_573_cast_fp16")]; + tensor var_581_begin_0 = const()[name = tensor("op_581_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_581_end_0 = const()[name = tensor("op_581_end_0"), val = tensor([1, 256, 16, 32, 2])]; + tensor var_581_end_mask_0 = const()[name = tensor("op_581_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_581_squeeze_mask_0 = const()[name = tensor("op_581_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_581_cast_fp16 = slice_by_index(begin = var_581_begin_0, end = var_581_end_0, end_mask = var_581_end_mask_0, squeeze_mask = var_581_squeeze_mask_0, x = k_complex_3_cast_fp16)[name = tensor("op_581_cast_fp16")]; + tensor freqs_3_to_fp16 = const()[name = tensor("freqs_3_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(32538496)))]; + tensor var_587_cast_fp16 = mul(x = freqs_3_to_fp16, y = ts_11_cast_fp16)[name = tensor("op_587_cast_fp16")]; + tensor rotr_3_cast_fp16 = cos(x = var_587_cast_fp16)[name = tensor("rotr_3_cast_fp16")]; + tensor roti_3_cast_fp16 = sin(x = var_587_cast_fp16)[name = tensor("roti_3_cast_fp16")]; + tensor var_591_cast_fp16 = mul(x = var_557_cast_fp16, y = rotr_3_cast_fp16)[name = tensor("op_591_cast_fp16")]; + tensor var_592_cast_fp16 = mul(x = var_565_cast_fp16, y = roti_3_cast_fp16)[name = tensor("op_592_cast_fp16")]; + tensor qor_5_cast_fp16 = sub(x = var_591_cast_fp16, y = var_592_cast_fp16)[name = tensor("qor_5_cast_fp16")]; + tensor var_595_cast_fp16 = mul(x = var_557_cast_fp16, y = roti_3_cast_fp16)[name = tensor("op_595_cast_fp16")]; + tensor var_596_cast_fp16 = mul(x = var_565_cast_fp16, y = rotr_3_cast_fp16)[name = tensor("op_596_cast_fp16")]; + tensor qoi_5_cast_fp16 = add(x = var_595_cast_fp16, y = var_596_cast_fp16)[name = tensor("qoi_5_cast_fp16")]; + tensor var_599_cast_fp16 = mul(x = var_573_cast_fp16, y = rotr_3_cast_fp16)[name = tensor("op_599_cast_fp16")]; + tensor var_600_cast_fp16 = mul(x = var_581_cast_fp16, y = roti_3_cast_fp16)[name = tensor("op_600_cast_fp16")]; + tensor kor_5_cast_fp16 = sub(x = var_599_cast_fp16, y = var_600_cast_fp16)[name = tensor("kor_5_cast_fp16")]; + tensor var_603_cast_fp16 = mul(x = var_573_cast_fp16, y = roti_3_cast_fp16)[name = tensor("op_603_cast_fp16")]; + tensor var_604_cast_fp16 = mul(x = var_581_cast_fp16, y = rotr_3_cast_fp16)[name = tensor("op_604_cast_fp16")]; + tensor koi_5_cast_fp16 = add(x = var_603_cast_fp16, y = var_604_cast_fp16)[name = tensor("koi_5_cast_fp16")]; + tensor qo_3_axis_0 = const()[name = tensor("qo_3_axis_0"), val = tensor(-1)]; + tensor qo_3_cast_fp16 = stack(axis = qo_3_axis_0, values = (qor_5_cast_fp16, qoi_5_cast_fp16))[name = tensor("qo_3_cast_fp16")]; + tensor ko_3_axis_0 = const()[name = tensor("ko_3_axis_0"), val = tensor(-1)]; + tensor ko_3_cast_fp16 = stack(axis = ko_3_axis_0, values = (kor_5_cast_fp16, koi_5_cast_fp16))[name = tensor("ko_3_cast_fp16")]; + tensor var_637 = const()[name = tensor("op_637"), val = tensor([1, 256, 16, 64])]; + tensor q_9_cast_fp16 = reshape(shape = var_637, x = qo_3_cast_fp16)[name = tensor("q_9_cast_fp16")]; + tensor var_643 = const()[name = tensor("op_643"), val = tensor([1, 256, 16, 64])]; + tensor k_7_cast_fp16 = reshape(shape = var_643, x = ko_3_cast_fp16)[name = tensor("k_7_cast_fp16")]; + tensor write_indices_float_7_cast_fp16 = add(x = var_532_cast_fp16, y = var_249_promoted_to_fp16)[name = tensor("write_indices_float_7_cast_fp16")]; + tensor _inversed_664_y_0_to_fp16 = const()[name = tensor("_inversed_664_y_0_to_fp16"), val = tensor(0x1p-9)]; + tensor _inversed_664_cast_fp16 = mul(x = write_indices_float_7_cast_fp16, y = _inversed_664_y_0_to_fp16)[name = tensor("_inversed_664_cast_fp16")]; + tensor var_665_cast_fp16 = floor(x = _inversed_664_cast_fp16)[name = tensor("op_665_cast_fp16")]; + tensor var_666_to_fp16 = const()[name = tensor("op_666_to_fp16"), val = tensor(0x1p+9)]; + tensor var_667_cast_fp16 = mul(x = var_665_cast_fp16, y = var_666_to_fp16)[name = tensor("op_667_cast_fp16")]; + tensor write_indices_float_9_cast_fp16 = sub(x = write_indices_float_7_cast_fp16, y = var_667_cast_fp16)[name = tensor("write_indices_float_9_cast_fp16")]; + tensor var_685_value_0_to_fp16 = const()[name = tensor("op_685_value_0_to_fp16"), val = tensor(0x1.ffp+8)]; + tensor var_685_cast_fp16 = fill_like(ref_tensor = write_indices_float_9_cast_fp16, value = var_685_value_0_to_fp16)[name = tensor("op_685_cast_fp16")]; + tensor write_indices_float_11_cast_fp16 = select(a = write_indices_float_9_cast_fp16, b = var_685_cast_fp16, cond = token_valid_1_cast_fp16)[name = tensor("write_indices_float_11_cast_fp16")]; + tensor var_691_dtype_0 = const()[name = tensor("op_691_dtype_0"), val = tensor("int32")]; + tensor var_696 = const()[name = tensor("op_696"), val = tensor([1, 256, 1, 1])]; + tensor write_indices_float_11_cast_fp16_to_int32 = cast(dtype = var_691_dtype_0, x = write_indices_float_11_cast_fp16)[name = tensor("cast_151")]; + tensor var_697 = reshape(shape = var_696, x = write_indices_float_11_cast_fp16_to_int32)[name = tensor("op_697")]; + tensor write_indices_3_reps_0 = const()[name = tensor("write_indices_3_reps_0"), val = tensor([1, 1, 16, 64])]; + tensor write_indices_3 = tile(reps = write_indices_3_reps_0, x = var_697)[name = tensor("write_indices_3")]; + tensor var_707_begin_0 = const()[name = tensor("op_707_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_707_end_0 = const()[name = tensor("op_707_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor var_707_end_mask_0 = const()[name = tensor("op_707_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_707_squeeze_mask_0 = const()[name = tensor("op_707_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor cache1_to_fp16_dtype_0 = const()[name = tensor("cache1_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor cache1_to_fp16 = cast(dtype = cache1_to_fp16_dtype_0, x = cache1)[name = tensor("cast_150")]; + tensor var_707_cast_fp16 = slice_by_index(begin = var_707_begin_0, end = var_707_end_0, end_mask = var_707_end_mask_0, squeeze_mask = var_707_squeeze_mask_0, x = cache1_to_fp16)[name = tensor("op_707_cast_fp16")]; + tensor var_709_axis_0 = const()[name = tensor("op_709_axis_0"), val = tensor(1)]; + tensor var_709_mode_0 = const()[name = tensor("op_709_mode_0"), val = tensor("update")]; + tensor var_709_validate_indices_0 = const()[name = tensor("op_709_validate_indices_0"), val = tensor(false)]; + tensor var_709_cast_fp16 = scatter_along_axis(axis = var_709_axis_0, data = var_707_cast_fp16, indices = write_indices_3, mode = var_709_mode_0, updates = k_7_cast_fp16, validate_indices = var_709_validate_indices_0)[name = tensor("op_709_cast_fp16")]; + tensor concat_10 = const()[name = tensor("concat_10"), val = tensor([0, 0, 0, 0, 0])]; + tensor concat_11 = const()[name = tensor("concat_11"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_3_internal_tensor_assign_1_stride_0 = const()[name = tensor("new_cache_3_internal_tensor_assign_1_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_3_internal_tensor_assign_1_begin_mask_0 = const()[name = tensor("new_cache_3_internal_tensor_assign_1_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_3_internal_tensor_assign_1_end_mask_0 = const()[name = tensor("new_cache_3_internal_tensor_assign_1_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_3_internal_tensor_assign_1_squeeze_mask_0 = const()[name = tensor("new_cache_3_internal_tensor_assign_1_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_12 = const()[name = tensor("shape_12"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_2 = const()[name = tensor("reduce_prod_2"), val = tensor(1048576)]; + tensor range_1d_2_start_0 = const()[name = tensor("range_1d_2_start_0"), val = tensor(0)]; + tensor range_1d_2_step_0 = const()[name = tensor("range_1d_2_step_0"), val = tensor(1)]; + tensor range_1d_2 = range_1d(end = reduce_prod_2, start = range_1d_2_start_0, step = range_1d_2_step_0)[name = tensor("range_1d_2")]; + tensor reshape_10 = reshape(shape = shape_12, x = range_1d_2)[name = tensor("reshape_10")]; + tensor slice_by_index_2 = slice_by_index(begin = concat_10, begin_mask = new_cache_3_internal_tensor_assign_1_begin_mask_0, end = concat_11, end_mask = new_cache_3_internal_tensor_assign_1_end_mask_0, squeeze_mask = new_cache_3_internal_tensor_assign_1_squeeze_mask_0, stride = new_cache_3_internal_tensor_assign_1_stride_0, x = reshape_10)[name = tensor("slice_by_index_2")]; + tensor reshape_11_shape_0 = const()[name = tensor("reshape_11_shape_0"), val = tensor([-1])]; + tensor reshape_11 = reshape(shape = reshape_11_shape_0, x = slice_by_index_2)[name = tensor("reshape_11")]; + tensor reshape_12_shape_0 = const()[name = tensor("reshape_12_shape_0"), val = tensor([-1])]; + tensor reshape_12_cast_fp16 = reshape(shape = reshape_12_shape_0, x = var_709_cast_fp16)[name = tensor("reshape_12_cast_fp16")]; + tensor reshape_13_shape_0 = const()[name = tensor("reshape_13_shape_0"), val = tensor([-1])]; + tensor reshape_13_cast_fp16 = reshape(shape = reshape_13_shape_0, x = cache1_to_fp16)[name = tensor("reshape_13_cast_fp16")]; + tensor scatter_2_mode_0 = const()[name = tensor("scatter_2_mode_0"), val = tensor("update")]; + tensor scatter_2_axis_0 = const()[name = tensor("scatter_2_axis_0"), val = tensor(0)]; + tensor scatter_2_validate_indices_0 = const()[name = tensor("scatter_2_validate_indices_0"), val = tensor(false)]; + tensor scatter_2_cast_fp16 = scatter(axis = scatter_2_axis_0, data = reshape_13_cast_fp16, indices = reshape_11, mode = scatter_2_mode_0, updates = reshape_12_cast_fp16, validate_indices = scatter_2_validate_indices_0)[name = tensor("scatter_2_cast_fp16")]; + tensor reshape_14_cast_fp16 = reshape(shape = shape_12, x = scatter_2_cast_fp16)[name = tensor("reshape_14_cast_fp16")]; + tensor var_717_begin_0 = const()[name = tensor("op_717_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor var_717_end_0 = const()[name = tensor("op_717_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor var_717_end_mask_0 = const()[name = tensor("op_717_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_717_squeeze_mask_0 = const()[name = tensor("op_717_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor var_717_cast_fp16 = slice_by_index(begin = var_717_begin_0, end = var_717_end_0, end_mask = var_717_end_mask_0, squeeze_mask = var_717_squeeze_mask_0, x = reshape_14_cast_fp16)[name = tensor("op_717_cast_fp16")]; + tensor var_719_axis_0 = const()[name = tensor("op_719_axis_0"), val = tensor(1)]; + tensor var_719_mode_0 = const()[name = tensor("op_719_mode_0"), val = tensor("update")]; + tensor var_719_validate_indices_0 = const()[name = tensor("op_719_validate_indices_0"), val = tensor(false)]; + tensor var_719_cast_fp16 = scatter_along_axis(axis = var_719_axis_0, data = var_717_cast_fp16, indices = write_indices_3, mode = var_719_mode_0, updates = v_3_cast_fp16, validate_indices = var_719_validate_indices_0)[name = tensor("op_719_cast_fp16")]; + tensor concat_12 = const()[name = tensor("concat_12"), val = tensor([1, 0, 0, 0, 0])]; + tensor concat_13 = const()[name = tensor("concat_13"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_3_internal_tensor_assign_2_stride_0 = const()[name = tensor("new_cache_3_internal_tensor_assign_2_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_3_internal_tensor_assign_2_begin_mask_0 = const()[name = tensor("new_cache_3_internal_tensor_assign_2_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_3_internal_tensor_assign_2_end_mask_0 = const()[name = tensor("new_cache_3_internal_tensor_assign_2_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_3_internal_tensor_assign_2_squeeze_mask_0 = const()[name = tensor("new_cache_3_internal_tensor_assign_2_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_13 = const()[name = tensor("shape_13"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_3 = const()[name = tensor("reduce_prod_3"), val = tensor(1048576)]; + tensor range_1d_3_start_0 = const()[name = tensor("range_1d_3_start_0"), val = tensor(0)]; + tensor range_1d_3_step_0 = const()[name = tensor("range_1d_3_step_0"), val = tensor(1)]; + tensor range_1d_3 = range_1d(end = reduce_prod_3, start = range_1d_3_start_0, step = range_1d_3_step_0)[name = tensor("range_1d_3")]; + tensor reshape_15 = reshape(shape = shape_13, x = range_1d_3)[name = tensor("reshape_15")]; + tensor slice_by_index_3 = slice_by_index(begin = concat_12, begin_mask = new_cache_3_internal_tensor_assign_2_begin_mask_0, end = concat_13, end_mask = new_cache_3_internal_tensor_assign_2_end_mask_0, squeeze_mask = new_cache_3_internal_tensor_assign_2_squeeze_mask_0, stride = new_cache_3_internal_tensor_assign_2_stride_0, x = reshape_15)[name = tensor("slice_by_index_3")]; + tensor reshape_16_shape_0 = const()[name = tensor("reshape_16_shape_0"), val = tensor([-1])]; + tensor reshape_16 = reshape(shape = reshape_16_shape_0, x = slice_by_index_3)[name = tensor("reshape_16")]; + tensor reshape_17_shape_0 = const()[name = tensor("reshape_17_shape_0"), val = tensor([-1])]; + tensor reshape_17_cast_fp16 = reshape(shape = reshape_17_shape_0, x = var_719_cast_fp16)[name = tensor("reshape_17_cast_fp16")]; + tensor reshape_18_shape_0 = const()[name = tensor("reshape_18_shape_0"), val = tensor([-1])]; + tensor reshape_18_cast_fp16 = reshape(shape = reshape_18_shape_0, x = reshape_14_cast_fp16)[name = tensor("reshape_18_cast_fp16")]; + tensor scatter_3_mode_0 = const()[name = tensor("scatter_3_mode_0"), val = tensor("update")]; + tensor scatter_3_axis_0 = const()[name = tensor("scatter_3_axis_0"), val = tensor(0)]; + tensor scatter_3_validate_indices_0 = const()[name = tensor("scatter_3_validate_indices_0"), val = tensor(false)]; + tensor scatter_3_cast_fp16 = scatter(axis = scatter_3_axis_0, data = reshape_18_cast_fp16, indices = reshape_16, mode = scatter_3_mode_0, updates = reshape_17_cast_fp16, validate_indices = scatter_3_validate_indices_0)[name = tensor("scatter_3_cast_fp16")]; + tensor reshape_19_cast_fp16 = reshape(shape = shape_13, x = scatter_3_cast_fp16)[name = tensor("reshape_19_cast_fp16")]; + tensor reshape_19_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("reshape_19_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor keys_7_begin_0 = const()[name = tensor("keys_7_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor keys_7_end_0 = const()[name = tensor("keys_7_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor keys_7_end_mask_0 = const()[name = tensor("keys_7_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor keys_7_squeeze_mask_0 = const()[name = tensor("keys_7_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor keys_7_cast_fp16 = slice_by_index(begin = keys_7_begin_0, end = keys_7_end_0, end_mask = keys_7_end_mask_0, squeeze_mask = keys_7_squeeze_mask_0, x = reshape_19_cast_fp16)[name = tensor("keys_7_cast_fp16")]; + tensor values_7_begin_0 = const()[name = tensor("values_7_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor values_7_end_0 = const()[name = tensor("values_7_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor values_7_end_mask_0 = const()[name = tensor("values_7_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor values_7_squeeze_mask_0 = const()[name = tensor("values_7_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor values_7_cast_fp16 = slice_by_index(begin = values_7_begin_0, end = values_7_end_0, end_mask = values_7_end_mask_0, squeeze_mask = values_7_squeeze_mask_0, x = reshape_19_cast_fp16)[name = tensor("values_7_cast_fp16")]; + tensor var_731_cast_fp16 = not_equal(x = keys_7_cast_fp16, y = keys_7_cast_fp16)[name = tensor("op_731_cast_fp16")]; + tensor keys_9_cast_fp16 = select(a = var_326_to_fp16, b = keys_7_cast_fp16, cond = var_731_cast_fp16)[name = tensor("keys_9_cast_fp16")]; + tensor var_739_cast_fp16 = not_equal(x = values_7_cast_fp16, y = values_7_cast_fp16)[name = tensor("op_739_cast_fp16")]; + tensor values_9_cast_fp16 = select(a = var_326_to_fp16, b = values_7_cast_fp16, cond = var_739_cast_fp16)[name = tensor("values_9_cast_fp16")]; + tensor var_763 = const()[name = tensor("op_763"), val = tensor([0, 2, 1, 3])]; + tensor var_779 = const()[name = tensor("op_779"), val = tensor([1, 1, 1])]; + tensor var_780_cast_fp16 = reshape(shape = var_779, x = position1_to_fp16)[name = tensor("op_780_cast_fp16")]; + tensor q_positions_3_cast_fp16 = add(x = var_780_cast_fp16, y = q_offsets_1_promoted_to_fp16)[name = tensor("q_positions_3_cast_fp16")]; + tensor valid_end_3_cast_fp16 = add(x = var_780_cast_fp16, y = var_392_cast_fp16)[name = tensor("valid_end_3_cast_fp16")]; + tensor valid_mask_3_cast_fp16 = less(x = k_positions_1_promoted_to_fp16, y = valid_end_3_cast_fp16)[name = tensor("valid_mask_3_cast_fp16")]; + tensor causal_mask_3_cast_fp16 = less_equal(x = k_positions_1_promoted_to_fp16, y = q_positions_3_cast_fp16)[name = tensor("causal_mask_3_cast_fp16")]; + tensor var_808 = logical_and(x = valid_mask_3_cast_fp16, y = causal_mask_3_cast_fp16)[name = tensor("op_808")]; + tensor attn_mask_3_axes_0 = const()[name = tensor("attn_mask_3_axes_0"), val = tensor([1])]; + tensor attn_mask_3 = expand_dims(axes = attn_mask_3_axes_0, x = var_808)[name = tensor("attn_mask_3")]; + tensor var_822_transpose_x_0 = const()[name = tensor("op_822_transpose_x_0"), val = tensor(false)]; + tensor var_822_transpose_y_0 = const()[name = tensor("op_822_transpose_y_0"), val = tensor(false)]; + tensor transpose_17_perm_0 = const()[name = tensor("transpose_17_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_18_perm_0 = const()[name = tensor("transpose_18_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_18 = transpose(perm = transpose_18_perm_0, x = keys_9_cast_fp16)[name = tensor("transpose_38")]; + tensor transpose_17 = transpose(perm = transpose_17_perm_0, x = q_9_cast_fp16)[name = tensor("transpose_39")]; + tensor var_822_cast_fp16 = matmul(transpose_x = var_822_transpose_x_0, transpose_y = var_822_transpose_y_0, x = transpose_17, y = transpose_18)[name = tensor("op_822_cast_fp16")]; + tensor var_816_to_fp16 = const()[name = tensor("op_816_to_fp16"), val = tensor([0x1p-3])]; + tensor attn_weights_7_cast_fp16 = mul(x = var_822_cast_fp16, y = var_816_to_fp16)[name = tensor("attn_weights_7_cast_fp16")]; + tensor var_824 = logical_not(x = attn_mask_3)[name = tensor("op_824")]; + tensor var_825_to_fp16 = const()[name = tensor("op_825_to_fp16"), val = tensor(-inf)]; + tensor attn_weights_9_cast_fp16 = select(a = var_825_to_fp16, b = attn_weights_7_cast_fp16, cond = var_824)[name = tensor("attn_weights_9_cast_fp16")]; + tensor var_827 = const()[name = tensor("op_827"), val = tensor(-1)]; + tensor attn_weights_11_cast_fp16 = softmax(axis = var_827, x = attn_weights_9_cast_fp16)[name = tensor("attn_weights_11_cast_fp16")]; + tensor attn_output_3_transpose_x_0 = const()[name = tensor("attn_output_3_transpose_x_0"), val = tensor(false)]; + tensor attn_output_3_transpose_y_0 = const()[name = tensor("attn_output_3_transpose_y_0"), val = tensor(false)]; + tensor values_11_cast_fp16 = transpose(perm = var_763, x = values_9_cast_fp16)[name = tensor("transpose_40")]; + tensor attn_output_3_cast_fp16 = matmul(transpose_x = attn_output_3_transpose_x_0, transpose_y = attn_output_3_transpose_y_0, x = attn_weights_11_cast_fp16, y = values_11_cast_fp16)[name = tensor("attn_output_3_cast_fp16")]; + tensor var_835 = const()[name = tensor("op_835"), val = tensor([0, 2, 1, 3])]; + tensor var_840 = const()[name = tensor("op_840"), val = tensor([1, 256, 1024])]; + tensor var_836_cast_fp16 = transpose(perm = var_835, x = attn_output_3_cast_fp16)[name = tensor("transpose_37")]; + tensor input_17_cast_fp16 = reshape(shape = var_840, x = var_836_cast_fp16)[name = tensor("input_17_cast_fp16")]; + tensor attn1_out_proj_weight_to_fp16 = const()[name = tensor("attn1_out_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(32538624)))]; + tensor linear_5_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = attn1_out_proj_weight_to_fp16, x = input_17_cast_fp16)[name = tensor("linear_5_cast_fp16")]; + tensor var_849_cast_fp16 = add(x = position1_to_fp16, y = valid_len_to_fp16)[name = tensor("op_849_cast_fp16")]; + tensor var_849_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_849_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor input_19_cast_fp16 = add(x = input_13_cast_fp16, y = linear_5_cast_fp16)[name = tensor("input_19_cast_fp16")]; + tensor input_21_axes_0 = const()[name = tensor("input_21_axes_0"), val = tensor([-1])]; + tensor norm1_2_weight_to_fp16 = const()[name = tensor("norm1_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(34635840)))]; + tensor norm1_2_bias_to_fp16 = const()[name = tensor("norm1_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(34637952)))]; + tensor var_853_to_fp16 = const()[name = tensor("op_853_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_21_cast_fp16 = layer_norm(axes = input_21_axes_0, beta = norm1_2_bias_to_fp16, epsilon = var_853_to_fp16, gamma = norm1_2_weight_to_fp16, x = input_19_cast_fp16)[name = tensor("input_21_cast_fp16")]; + tensor linear1_1_weight_to_fp16 = const()[name = tensor("linear1_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(34640064)))]; + tensor linear_6_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = linear1_1_weight_to_fp16, x = input_21_cast_fp16)[name = tensor("linear_6_cast_fp16")]; + tensor input_23_mode_0 = const()[name = tensor("input_23_mode_0"), val = tensor("EXACT")]; + tensor input_23_cast_fp16 = gelu(mode = input_23_mode_0, x = linear_6_cast_fp16)[name = tensor("input_23_cast_fp16")]; + tensor linear1_2_weight_to_fp16 = const()[name = tensor("linear1_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(43028736)))]; + tensor linear_7_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = linear1_2_weight_to_fp16, x = input_23_cast_fp16)[name = tensor("linear_7_cast_fp16")]; + tensor input_25_cast_fp16 = add(x = input_19_cast_fp16, y = linear_7_cast_fp16)[name = tensor("input_25_cast_fp16")]; + tensor input_27_axes_0 = const()[name = tensor("input_27_axes_0"), val = tensor([-1])]; + tensor norm2_1_weight_to_fp16 = const()[name = tensor("norm2_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(51417408)))]; + tensor norm2_1_bias_to_fp16 = const()[name = tensor("norm2_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(51419520)))]; + tensor var_870_to_fp16 = const()[name = tensor("op_870_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_27_cast_fp16 = layer_norm(axes = input_27_axes_0, beta = norm2_1_bias_to_fp16, epsilon = var_870_to_fp16, gamma = norm2_1_weight_to_fp16, x = input_25_cast_fp16)[name = tensor("input_27_cast_fp16")]; + tensor attn2_in_proj_weight_to_fp16 = const()[name = tensor("attn2_in_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(51421632)))]; + tensor linear_8_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = attn2_in_proj_weight_to_fp16, x = input_27_cast_fp16)[name = tensor("linear_8_cast_fp16")]; + tensor var_884 = const()[name = tensor("op_884"), val = tensor([1, 256, 3, 16, 64])]; + tensor qkv_5_cast_fp16 = reshape(shape = var_884, x = linear_8_cast_fp16)[name = tensor("qkv_5_cast_fp16")]; + tensor q_13_begin_0 = const()[name = tensor("q_13_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor q_13_end_0 = const()[name = tensor("q_13_end_0"), val = tensor([1, 256, 1, 16, 64])]; + tensor q_13_end_mask_0 = const()[name = tensor("q_13_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor q_13_squeeze_mask_0 = const()[name = tensor("q_13_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor q_13_cast_fp16 = slice_by_index(begin = q_13_begin_0, end = q_13_end_0, end_mask = q_13_end_mask_0, squeeze_mask = q_13_squeeze_mask_0, x = qkv_5_cast_fp16)[name = tensor("q_13_cast_fp16")]; + tensor k_9_begin_0 = const()[name = tensor("k_9_begin_0"), val = tensor([0, 0, 1, 0, 0])]; + tensor k_9_end_0 = const()[name = tensor("k_9_end_0"), val = tensor([1, 256, 2, 16, 64])]; + tensor k_9_end_mask_0 = const()[name = tensor("k_9_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor k_9_squeeze_mask_0 = const()[name = tensor("k_9_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor k_9_cast_fp16 = slice_by_index(begin = k_9_begin_0, end = k_9_end_0, end_mask = k_9_end_mask_0, squeeze_mask = k_9_squeeze_mask_0, x = qkv_5_cast_fp16)[name = tensor("k_9_cast_fp16")]; + tensor v_5_begin_0 = const()[name = tensor("v_5_begin_0"), val = tensor([0, 0, 2, 0, 0])]; + tensor v_5_end_0 = const()[name = tensor("v_5_end_0"), val = tensor([1, 256, 3, 16, 64])]; + tensor v_5_end_mask_0 = const()[name = tensor("v_5_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor v_5_squeeze_mask_0 = const()[name = tensor("v_5_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor v_5_cast_fp16 = slice_by_index(begin = v_5_begin_0, end = v_5_end_0, end_mask = v_5_end_mask_0, squeeze_mask = v_5_squeeze_mask_0, x = qkv_5_cast_fp16)[name = tensor("v_5_cast_fp16")]; + tensor var_942 = const()[name = tensor("op_942"), val = tensor([1, 1])]; + tensor position2_to_fp16_dtype_0 = const()[name = tensor("position2_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor position2_to_fp16 = cast(dtype = position2_to_fp16_dtype_0, x = position2)[name = tensor("cast_147")]; + tensor var_943_cast_fp16 = reshape(shape = var_942, x = position2_to_fp16)[name = tensor("op_943_cast_fp16")]; + tensor ts_15_cast_fp16 = add(x = ts_1_promoted_to_fp16, y = var_943_cast_fp16)[name = tensor("ts_15_cast_fp16")]; + tensor var_950 = const()[name = tensor("op_950"), val = tensor([1, 256, 1, 1])]; + tensor ts_17_cast_fp16 = reshape(shape = var_950, x = ts_15_cast_fp16)[name = tensor("ts_17_cast_fp16")]; + tensor var_957 = const()[name = tensor("op_957"), val = tensor([1, 256, 16, 32, 2])]; + tensor q_complex_5_cast_fp16 = reshape(shape = var_957, x = q_13_cast_fp16)[name = tensor("q_complex_5_cast_fp16")]; + tensor var_964 = const()[name = tensor("op_964"), val = tensor([1, 256, 16, 32, 2])]; + tensor k_complex_5_cast_fp16 = reshape(shape = var_964, x = k_9_cast_fp16)[name = tensor("k_complex_5_cast_fp16")]; + tensor var_968_begin_0 = const()[name = tensor("op_968_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_968_end_0 = const()[name = tensor("op_968_end_0"), val = tensor([1, 256, 16, 32, 1])]; + tensor var_968_end_mask_0 = const()[name = tensor("op_968_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_968_squeeze_mask_0 = const()[name = tensor("op_968_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_968_cast_fp16 = slice_by_index(begin = var_968_begin_0, end = var_968_end_0, end_mask = var_968_end_mask_0, squeeze_mask = var_968_squeeze_mask_0, x = q_complex_5_cast_fp16)[name = tensor("op_968_cast_fp16")]; + tensor var_976_begin_0 = const()[name = tensor("op_976_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_976_end_0 = const()[name = tensor("op_976_end_0"), val = tensor([1, 256, 16, 32, 2])]; + tensor var_976_end_mask_0 = const()[name = tensor("op_976_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_976_squeeze_mask_0 = const()[name = tensor("op_976_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_976_cast_fp16 = slice_by_index(begin = var_976_begin_0, end = var_976_end_0, end_mask = var_976_end_mask_0, squeeze_mask = var_976_squeeze_mask_0, x = q_complex_5_cast_fp16)[name = tensor("op_976_cast_fp16")]; + tensor var_984_begin_0 = const()[name = tensor("op_984_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_984_end_0 = const()[name = tensor("op_984_end_0"), val = tensor([1, 256, 16, 32, 1])]; + tensor var_984_end_mask_0 = const()[name = tensor("op_984_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_984_squeeze_mask_0 = const()[name = tensor("op_984_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_984_cast_fp16 = slice_by_index(begin = var_984_begin_0, end = var_984_end_0, end_mask = var_984_end_mask_0, squeeze_mask = var_984_squeeze_mask_0, x = k_complex_5_cast_fp16)[name = tensor("op_984_cast_fp16")]; + tensor var_992_begin_0 = const()[name = tensor("op_992_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_992_end_0 = const()[name = tensor("op_992_end_0"), val = tensor([1, 256, 16, 32, 2])]; + tensor var_992_end_mask_0 = const()[name = tensor("op_992_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_992_squeeze_mask_0 = const()[name = tensor("op_992_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_992_cast_fp16 = slice_by_index(begin = var_992_begin_0, end = var_992_end_0, end_mask = var_992_end_mask_0, squeeze_mask = var_992_squeeze_mask_0, x = k_complex_5_cast_fp16)[name = tensor("op_992_cast_fp16")]; + tensor freqs_5_to_fp16 = const()[name = tensor("freqs_5_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(57713152)))]; + tensor var_998_cast_fp16 = mul(x = freqs_5_to_fp16, y = ts_17_cast_fp16)[name = tensor("op_998_cast_fp16")]; + tensor rotr_5_cast_fp16 = cos(x = var_998_cast_fp16)[name = tensor("rotr_5_cast_fp16")]; + tensor roti_5_cast_fp16 = sin(x = var_998_cast_fp16)[name = tensor("roti_5_cast_fp16")]; + tensor var_1002_cast_fp16 = mul(x = var_968_cast_fp16, y = rotr_5_cast_fp16)[name = tensor("op_1002_cast_fp16")]; + tensor var_1003_cast_fp16 = mul(x = var_976_cast_fp16, y = roti_5_cast_fp16)[name = tensor("op_1003_cast_fp16")]; + tensor qor_9_cast_fp16 = sub(x = var_1002_cast_fp16, y = var_1003_cast_fp16)[name = tensor("qor_9_cast_fp16")]; + tensor var_1006_cast_fp16 = mul(x = var_968_cast_fp16, y = roti_5_cast_fp16)[name = tensor("op_1006_cast_fp16")]; + tensor var_1007_cast_fp16 = mul(x = var_976_cast_fp16, y = rotr_5_cast_fp16)[name = tensor("op_1007_cast_fp16")]; + tensor qoi_9_cast_fp16 = add(x = var_1006_cast_fp16, y = var_1007_cast_fp16)[name = tensor("qoi_9_cast_fp16")]; + tensor var_1010_cast_fp16 = mul(x = var_984_cast_fp16, y = rotr_5_cast_fp16)[name = tensor("op_1010_cast_fp16")]; + tensor var_1011_cast_fp16 = mul(x = var_992_cast_fp16, y = roti_5_cast_fp16)[name = tensor("op_1011_cast_fp16")]; + tensor kor_9_cast_fp16 = sub(x = var_1010_cast_fp16, y = var_1011_cast_fp16)[name = tensor("kor_9_cast_fp16")]; + tensor var_1014_cast_fp16 = mul(x = var_984_cast_fp16, y = roti_5_cast_fp16)[name = tensor("op_1014_cast_fp16")]; + tensor var_1015_cast_fp16 = mul(x = var_992_cast_fp16, y = rotr_5_cast_fp16)[name = tensor("op_1015_cast_fp16")]; + tensor koi_9_cast_fp16 = add(x = var_1014_cast_fp16, y = var_1015_cast_fp16)[name = tensor("koi_9_cast_fp16")]; + tensor qo_5_axis_0 = const()[name = tensor("qo_5_axis_0"), val = tensor(-1)]; + tensor qo_5_cast_fp16 = stack(axis = qo_5_axis_0, values = (qor_9_cast_fp16, qoi_9_cast_fp16))[name = tensor("qo_5_cast_fp16")]; + tensor ko_5_axis_0 = const()[name = tensor("ko_5_axis_0"), val = tensor(-1)]; + tensor ko_5_cast_fp16 = stack(axis = ko_5_axis_0, values = (kor_9_cast_fp16, koi_9_cast_fp16))[name = tensor("ko_5_cast_fp16")]; + tensor var_1048 = const()[name = tensor("op_1048"), val = tensor([1, 256, 16, 64])]; + tensor q_15_cast_fp16 = reshape(shape = var_1048, x = qo_5_cast_fp16)[name = tensor("q_15_cast_fp16")]; + tensor var_1054 = const()[name = tensor("op_1054"), val = tensor([1, 256, 16, 64])]; + tensor k_11_cast_fp16 = reshape(shape = var_1054, x = ko_5_cast_fp16)[name = tensor("k_11_cast_fp16")]; + tensor write_indices_float_13_cast_fp16 = add(x = var_943_cast_fp16, y = var_249_promoted_to_fp16)[name = tensor("write_indices_float_13_cast_fp16")]; + tensor _inversed_1075_y_0_to_fp16 = const()[name = tensor("_inversed_1075_y_0_to_fp16"), val = tensor(0x1p-9)]; + tensor _inversed_1075_cast_fp16 = mul(x = write_indices_float_13_cast_fp16, y = _inversed_1075_y_0_to_fp16)[name = tensor("_inversed_1075_cast_fp16")]; + tensor var_1076_cast_fp16 = floor(x = _inversed_1075_cast_fp16)[name = tensor("op_1076_cast_fp16")]; + tensor var_1077_to_fp16 = const()[name = tensor("op_1077_to_fp16"), val = tensor(0x1p+9)]; + tensor var_1078_cast_fp16 = mul(x = var_1076_cast_fp16, y = var_1077_to_fp16)[name = tensor("op_1078_cast_fp16")]; + tensor write_indices_float_15_cast_fp16 = sub(x = write_indices_float_13_cast_fp16, y = var_1078_cast_fp16)[name = tensor("write_indices_float_15_cast_fp16")]; + tensor var_1096_value_0_to_fp16 = const()[name = tensor("op_1096_value_0_to_fp16"), val = tensor(0x1.ffp+8)]; + tensor var_1096_cast_fp16 = fill_like(ref_tensor = write_indices_float_15_cast_fp16, value = var_1096_value_0_to_fp16)[name = tensor("op_1096_cast_fp16")]; + tensor write_indices_float_17_cast_fp16 = select(a = write_indices_float_15_cast_fp16, b = var_1096_cast_fp16, cond = token_valid_1_cast_fp16)[name = tensor("write_indices_float_17_cast_fp16")]; + tensor var_1102_dtype_0 = const()[name = tensor("op_1102_dtype_0"), val = tensor("int32")]; + tensor var_1107 = const()[name = tensor("op_1107"), val = tensor([1, 256, 1, 1])]; + tensor write_indices_float_17_cast_fp16_to_int32 = cast(dtype = var_1102_dtype_0, x = write_indices_float_17_cast_fp16)[name = tensor("cast_146")]; + tensor var_1108 = reshape(shape = var_1107, x = write_indices_float_17_cast_fp16_to_int32)[name = tensor("op_1108")]; + tensor write_indices_5_reps_0 = const()[name = tensor("write_indices_5_reps_0"), val = tensor([1, 1, 16, 64])]; + tensor write_indices_5 = tile(reps = write_indices_5_reps_0, x = var_1108)[name = tensor("write_indices_5")]; + tensor var_1118_begin_0 = const()[name = tensor("op_1118_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_1118_end_0 = const()[name = tensor("op_1118_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor var_1118_end_mask_0 = const()[name = tensor("op_1118_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_1118_squeeze_mask_0 = const()[name = tensor("op_1118_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor cache2_to_fp16_dtype_0 = const()[name = tensor("cache2_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor cache2_to_fp16 = cast(dtype = cache2_to_fp16_dtype_0, x = cache2)[name = tensor("cast_145")]; + tensor var_1118_cast_fp16 = slice_by_index(begin = var_1118_begin_0, end = var_1118_end_0, end_mask = var_1118_end_mask_0, squeeze_mask = var_1118_squeeze_mask_0, x = cache2_to_fp16)[name = tensor("op_1118_cast_fp16")]; + tensor var_1120_axis_0 = const()[name = tensor("op_1120_axis_0"), val = tensor(1)]; + tensor var_1120_mode_0 = const()[name = tensor("op_1120_mode_0"), val = tensor("update")]; + tensor var_1120_validate_indices_0 = const()[name = tensor("op_1120_validate_indices_0"), val = tensor(false)]; + tensor var_1120_cast_fp16 = scatter_along_axis(axis = var_1120_axis_0, data = var_1118_cast_fp16, indices = write_indices_5, mode = var_1120_mode_0, updates = k_11_cast_fp16, validate_indices = var_1120_validate_indices_0)[name = tensor("op_1120_cast_fp16")]; + tensor concat_18 = const()[name = tensor("concat_18"), val = tensor([0, 0, 0, 0, 0])]; + tensor concat_19 = const()[name = tensor("concat_19"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_5_internal_tensor_assign_1_stride_0 = const()[name = tensor("new_cache_5_internal_tensor_assign_1_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_5_internal_tensor_assign_1_begin_mask_0 = const()[name = tensor("new_cache_5_internal_tensor_assign_1_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_5_internal_tensor_assign_1_end_mask_0 = const()[name = tensor("new_cache_5_internal_tensor_assign_1_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_5_internal_tensor_assign_1_squeeze_mask_0 = const()[name = tensor("new_cache_5_internal_tensor_assign_1_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_14 = const()[name = tensor("shape_14"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_4 = const()[name = tensor("reduce_prod_4"), val = tensor(1048576)]; + tensor range_1d_4_start_0 = const()[name = tensor("range_1d_4_start_0"), val = tensor(0)]; + tensor range_1d_4_step_0 = const()[name = tensor("range_1d_4_step_0"), val = tensor(1)]; + tensor range_1d_4 = range_1d(end = reduce_prod_4, start = range_1d_4_start_0, step = range_1d_4_step_0)[name = tensor("range_1d_4")]; + tensor reshape_20 = reshape(shape = shape_14, x = range_1d_4)[name = tensor("reshape_20")]; + tensor slice_by_index_4 = slice_by_index(begin = concat_18, begin_mask = new_cache_5_internal_tensor_assign_1_begin_mask_0, end = concat_19, end_mask = new_cache_5_internal_tensor_assign_1_end_mask_0, squeeze_mask = new_cache_5_internal_tensor_assign_1_squeeze_mask_0, stride = new_cache_5_internal_tensor_assign_1_stride_0, x = reshape_20)[name = tensor("slice_by_index_4")]; + tensor reshape_21_shape_0 = const()[name = tensor("reshape_21_shape_0"), val = tensor([-1])]; + tensor reshape_21 = reshape(shape = reshape_21_shape_0, x = slice_by_index_4)[name = tensor("reshape_21")]; + tensor reshape_22_shape_0 = const()[name = tensor("reshape_22_shape_0"), val = tensor([-1])]; + tensor reshape_22_cast_fp16 = reshape(shape = reshape_22_shape_0, x = var_1120_cast_fp16)[name = tensor("reshape_22_cast_fp16")]; + tensor reshape_23_shape_0 = const()[name = tensor("reshape_23_shape_0"), val = tensor([-1])]; + tensor reshape_23_cast_fp16 = reshape(shape = reshape_23_shape_0, x = cache2_to_fp16)[name = tensor("reshape_23_cast_fp16")]; + tensor scatter_4_mode_0 = const()[name = tensor("scatter_4_mode_0"), val = tensor("update")]; + tensor scatter_4_axis_0 = const()[name = tensor("scatter_4_axis_0"), val = tensor(0)]; + tensor scatter_4_validate_indices_0 = const()[name = tensor("scatter_4_validate_indices_0"), val = tensor(false)]; + tensor scatter_4_cast_fp16 = scatter(axis = scatter_4_axis_0, data = reshape_23_cast_fp16, indices = reshape_21, mode = scatter_4_mode_0, updates = reshape_22_cast_fp16, validate_indices = scatter_4_validate_indices_0)[name = tensor("scatter_4_cast_fp16")]; + tensor reshape_24_cast_fp16 = reshape(shape = shape_14, x = scatter_4_cast_fp16)[name = tensor("reshape_24_cast_fp16")]; + tensor var_1128_begin_0 = const()[name = tensor("op_1128_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor var_1128_end_0 = const()[name = tensor("op_1128_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor var_1128_end_mask_0 = const()[name = tensor("op_1128_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_1128_squeeze_mask_0 = const()[name = tensor("op_1128_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor var_1128_cast_fp16 = slice_by_index(begin = var_1128_begin_0, end = var_1128_end_0, end_mask = var_1128_end_mask_0, squeeze_mask = var_1128_squeeze_mask_0, x = reshape_24_cast_fp16)[name = tensor("op_1128_cast_fp16")]; + tensor var_1130_axis_0 = const()[name = tensor("op_1130_axis_0"), val = tensor(1)]; + tensor var_1130_mode_0 = const()[name = tensor("op_1130_mode_0"), val = tensor("update")]; + tensor var_1130_validate_indices_0 = const()[name = tensor("op_1130_validate_indices_0"), val = tensor(false)]; + tensor var_1130_cast_fp16 = scatter_along_axis(axis = var_1130_axis_0, data = var_1128_cast_fp16, indices = write_indices_5, mode = var_1130_mode_0, updates = v_5_cast_fp16, validate_indices = var_1130_validate_indices_0)[name = tensor("op_1130_cast_fp16")]; + tensor concat_20 = const()[name = tensor("concat_20"), val = tensor([1, 0, 0, 0, 0])]; + tensor concat_21 = const()[name = tensor("concat_21"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_5_internal_tensor_assign_2_stride_0 = const()[name = tensor("new_cache_5_internal_tensor_assign_2_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_5_internal_tensor_assign_2_begin_mask_0 = const()[name = tensor("new_cache_5_internal_tensor_assign_2_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_5_internal_tensor_assign_2_end_mask_0 = const()[name = tensor("new_cache_5_internal_tensor_assign_2_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_5_internal_tensor_assign_2_squeeze_mask_0 = const()[name = tensor("new_cache_5_internal_tensor_assign_2_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_15 = const()[name = tensor("shape_15"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_5 = const()[name = tensor("reduce_prod_5"), val = tensor(1048576)]; + tensor range_1d_5_start_0 = const()[name = tensor("range_1d_5_start_0"), val = tensor(0)]; + tensor range_1d_5_step_0 = const()[name = tensor("range_1d_5_step_0"), val = tensor(1)]; + tensor range_1d_5 = range_1d(end = reduce_prod_5, start = range_1d_5_start_0, step = range_1d_5_step_0)[name = tensor("range_1d_5")]; + tensor reshape_25 = reshape(shape = shape_15, x = range_1d_5)[name = tensor("reshape_25")]; + tensor slice_by_index_5 = slice_by_index(begin = concat_20, begin_mask = new_cache_5_internal_tensor_assign_2_begin_mask_0, end = concat_21, end_mask = new_cache_5_internal_tensor_assign_2_end_mask_0, squeeze_mask = new_cache_5_internal_tensor_assign_2_squeeze_mask_0, stride = new_cache_5_internal_tensor_assign_2_stride_0, x = reshape_25)[name = tensor("slice_by_index_5")]; + tensor reshape_26_shape_0 = const()[name = tensor("reshape_26_shape_0"), val = tensor([-1])]; + tensor reshape_26 = reshape(shape = reshape_26_shape_0, x = slice_by_index_5)[name = tensor("reshape_26")]; + tensor reshape_27_shape_0 = const()[name = tensor("reshape_27_shape_0"), val = tensor([-1])]; + tensor reshape_27_cast_fp16 = reshape(shape = reshape_27_shape_0, x = var_1130_cast_fp16)[name = tensor("reshape_27_cast_fp16")]; + tensor reshape_28_shape_0 = const()[name = tensor("reshape_28_shape_0"), val = tensor([-1])]; + tensor reshape_28_cast_fp16 = reshape(shape = reshape_28_shape_0, x = reshape_24_cast_fp16)[name = tensor("reshape_28_cast_fp16")]; + tensor scatter_5_mode_0 = const()[name = tensor("scatter_5_mode_0"), val = tensor("update")]; + tensor scatter_5_axis_0 = const()[name = tensor("scatter_5_axis_0"), val = tensor(0)]; + tensor scatter_5_validate_indices_0 = const()[name = tensor("scatter_5_validate_indices_0"), val = tensor(false)]; + tensor scatter_5_cast_fp16 = scatter(axis = scatter_5_axis_0, data = reshape_28_cast_fp16, indices = reshape_26, mode = scatter_5_mode_0, updates = reshape_27_cast_fp16, validate_indices = scatter_5_validate_indices_0)[name = tensor("scatter_5_cast_fp16")]; + tensor reshape_29_cast_fp16 = reshape(shape = shape_15, x = scatter_5_cast_fp16)[name = tensor("reshape_29_cast_fp16")]; + tensor reshape_29_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("reshape_29_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor keys_13_begin_0 = const()[name = tensor("keys_13_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor keys_13_end_0 = const()[name = tensor("keys_13_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor keys_13_end_mask_0 = const()[name = tensor("keys_13_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor keys_13_squeeze_mask_0 = const()[name = tensor("keys_13_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor keys_13_cast_fp16 = slice_by_index(begin = keys_13_begin_0, end = keys_13_end_0, end_mask = keys_13_end_mask_0, squeeze_mask = keys_13_squeeze_mask_0, x = reshape_29_cast_fp16)[name = tensor("keys_13_cast_fp16")]; + tensor values_13_begin_0 = const()[name = tensor("values_13_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor values_13_end_0 = const()[name = tensor("values_13_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor values_13_end_mask_0 = const()[name = tensor("values_13_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor values_13_squeeze_mask_0 = const()[name = tensor("values_13_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor values_13_cast_fp16 = slice_by_index(begin = values_13_begin_0, end = values_13_end_0, end_mask = values_13_end_mask_0, squeeze_mask = values_13_squeeze_mask_0, x = reshape_29_cast_fp16)[name = tensor("values_13_cast_fp16")]; + tensor var_1142_cast_fp16 = not_equal(x = keys_13_cast_fp16, y = keys_13_cast_fp16)[name = tensor("op_1142_cast_fp16")]; + tensor keys_15_cast_fp16 = select(a = var_326_to_fp16, b = keys_13_cast_fp16, cond = var_1142_cast_fp16)[name = tensor("keys_15_cast_fp16")]; + tensor var_1150_cast_fp16 = not_equal(x = values_13_cast_fp16, y = values_13_cast_fp16)[name = tensor("op_1150_cast_fp16")]; + tensor values_15_cast_fp16 = select(a = var_326_to_fp16, b = values_13_cast_fp16, cond = var_1150_cast_fp16)[name = tensor("values_15_cast_fp16")]; + tensor var_1174 = const()[name = tensor("op_1174"), val = tensor([0, 2, 1, 3])]; + tensor var_1190 = const()[name = tensor("op_1190"), val = tensor([1, 1, 1])]; + tensor var_1191_cast_fp16 = reshape(shape = var_1190, x = position2_to_fp16)[name = tensor("op_1191_cast_fp16")]; + tensor q_positions_5_cast_fp16 = add(x = var_1191_cast_fp16, y = q_offsets_1_promoted_to_fp16)[name = tensor("q_positions_5_cast_fp16")]; + tensor valid_end_5_cast_fp16 = add(x = var_1191_cast_fp16, y = var_392_cast_fp16)[name = tensor("valid_end_5_cast_fp16")]; + tensor valid_mask_5_cast_fp16 = less(x = k_positions_1_promoted_to_fp16, y = valid_end_5_cast_fp16)[name = tensor("valid_mask_5_cast_fp16")]; + tensor causal_mask_5_cast_fp16 = less_equal(x = k_positions_1_promoted_to_fp16, y = q_positions_5_cast_fp16)[name = tensor("causal_mask_5_cast_fp16")]; + tensor var_1219 = logical_and(x = valid_mask_5_cast_fp16, y = causal_mask_5_cast_fp16)[name = tensor("op_1219")]; + tensor attn_mask_5_axes_0 = const()[name = tensor("attn_mask_5_axes_0"), val = tensor([1])]; + tensor attn_mask_5 = expand_dims(axes = attn_mask_5_axes_0, x = var_1219)[name = tensor("attn_mask_5")]; + tensor var_1233_transpose_x_0 = const()[name = tensor("op_1233_transpose_x_0"), val = tensor(false)]; + tensor var_1233_transpose_y_0 = const()[name = tensor("op_1233_transpose_y_0"), val = tensor(false)]; + tensor transpose_19_perm_0 = const()[name = tensor("transpose_19_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_20_perm_0 = const()[name = tensor("transpose_20_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_20 = transpose(perm = transpose_20_perm_0, x = keys_15_cast_fp16)[name = tensor("transpose_34")]; + tensor transpose_19 = transpose(perm = transpose_19_perm_0, x = q_15_cast_fp16)[name = tensor("transpose_35")]; + tensor var_1233_cast_fp16 = matmul(transpose_x = var_1233_transpose_x_0, transpose_y = var_1233_transpose_y_0, x = transpose_19, y = transpose_20)[name = tensor("op_1233_cast_fp16")]; + tensor var_1227_to_fp16 = const()[name = tensor("op_1227_to_fp16"), val = tensor([0x1p-3])]; + tensor attn_weights_13_cast_fp16 = mul(x = var_1233_cast_fp16, y = var_1227_to_fp16)[name = tensor("attn_weights_13_cast_fp16")]; + tensor var_1235 = logical_not(x = attn_mask_5)[name = tensor("op_1235")]; + tensor var_1236_to_fp16 = const()[name = tensor("op_1236_to_fp16"), val = tensor(-inf)]; + tensor attn_weights_15_cast_fp16 = select(a = var_1236_to_fp16, b = attn_weights_13_cast_fp16, cond = var_1235)[name = tensor("attn_weights_15_cast_fp16")]; + tensor var_1238 = const()[name = tensor("op_1238"), val = tensor(-1)]; + tensor attn_weights_17_cast_fp16 = softmax(axis = var_1238, x = attn_weights_15_cast_fp16)[name = tensor("attn_weights_17_cast_fp16")]; + tensor attn_output_5_transpose_x_0 = const()[name = tensor("attn_output_5_transpose_x_0"), val = tensor(false)]; + tensor attn_output_5_transpose_y_0 = const()[name = tensor("attn_output_5_transpose_y_0"), val = tensor(false)]; + tensor values_17_cast_fp16 = transpose(perm = var_1174, x = values_15_cast_fp16)[name = tensor("transpose_36")]; + tensor attn_output_5_cast_fp16 = matmul(transpose_x = attn_output_5_transpose_x_0, transpose_y = attn_output_5_transpose_y_0, x = attn_weights_17_cast_fp16, y = values_17_cast_fp16)[name = tensor("attn_output_5_cast_fp16")]; + tensor var_1246 = const()[name = tensor("op_1246"), val = tensor([0, 2, 1, 3])]; + tensor var_1251 = const()[name = tensor("op_1251"), val = tensor([1, 256, 1024])]; + tensor var_1247_cast_fp16 = transpose(perm = var_1246, x = attn_output_5_cast_fp16)[name = tensor("transpose_33")]; + tensor input_29_cast_fp16 = reshape(shape = var_1251, x = var_1247_cast_fp16)[name = tensor("input_29_cast_fp16")]; + tensor attn2_out_proj_weight_to_fp16 = const()[name = tensor("attn2_out_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(57713280)))]; + tensor linear_9_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = attn2_out_proj_weight_to_fp16, x = input_29_cast_fp16)[name = tensor("linear_9_cast_fp16")]; + tensor var_1260_cast_fp16 = add(x = position2_to_fp16, y = valid_len_to_fp16)[name = tensor("op_1260_cast_fp16")]; + tensor var_1260_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_1260_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor input_31_cast_fp16 = add(x = input_25_cast_fp16, y = linear_9_cast_fp16)[name = tensor("input_31_cast_fp16")]; + tensor input_33_axes_0 = const()[name = tensor("input_33_axes_0"), val = tensor([-1])]; + tensor norm2_2_weight_to_fp16 = const()[name = tensor("norm2_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(59810496)))]; + tensor norm2_2_bias_to_fp16 = const()[name = tensor("norm2_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(59812608)))]; + tensor var_1264_to_fp16 = const()[name = tensor("op_1264_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_33_cast_fp16 = layer_norm(axes = input_33_axes_0, beta = norm2_2_bias_to_fp16, epsilon = var_1264_to_fp16, gamma = norm2_2_weight_to_fp16, x = input_31_cast_fp16)[name = tensor("input_33_cast_fp16")]; + tensor linear2_1_weight_to_fp16 = const()[name = tensor("linear2_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(59814720)))]; + tensor linear_10_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = linear2_1_weight_to_fp16, x = input_33_cast_fp16)[name = tensor("linear_10_cast_fp16")]; + tensor input_35_mode_0 = const()[name = tensor("input_35_mode_0"), val = tensor("EXACT")]; + tensor input_35_cast_fp16 = gelu(mode = input_35_mode_0, x = linear_10_cast_fp16)[name = tensor("input_35_cast_fp16")]; + tensor linear2_2_weight_to_fp16 = const()[name = tensor("linear2_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(68203392)))]; + tensor linear_11_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = linear2_2_weight_to_fp16, x = input_35_cast_fp16)[name = tensor("linear_11_cast_fp16")]; + tensor input_37_cast_fp16 = add(x = input_31_cast_fp16, y = linear_11_cast_fp16)[name = tensor("input_37_cast_fp16")]; + tensor input_39_axes_0 = const()[name = tensor("input_39_axes_0"), val = tensor([-1])]; + tensor norm3_1_weight_to_fp16 = const()[name = tensor("norm3_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(76592064)))]; + tensor norm3_1_bias_to_fp16 = const()[name = tensor("norm3_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(76594176)))]; + tensor var_1281_to_fp16 = const()[name = tensor("op_1281_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_39_cast_fp16 = layer_norm(axes = input_39_axes_0, beta = norm3_1_bias_to_fp16, epsilon = var_1281_to_fp16, gamma = norm3_1_weight_to_fp16, x = input_37_cast_fp16)[name = tensor("input_39_cast_fp16")]; + tensor attn3_in_proj_weight_to_fp16 = const()[name = tensor("attn3_in_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(76596288)))]; + tensor linear_12_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = attn3_in_proj_weight_to_fp16, x = input_39_cast_fp16)[name = tensor("linear_12_cast_fp16")]; + tensor var_1295 = const()[name = tensor("op_1295"), val = tensor([1, 256, 3, 16, 64])]; + tensor qkv_7_cast_fp16 = reshape(shape = var_1295, x = linear_12_cast_fp16)[name = tensor("qkv_7_cast_fp16")]; + tensor q_19_begin_0 = const()[name = tensor("q_19_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor q_19_end_0 = const()[name = tensor("q_19_end_0"), val = tensor([1, 256, 1, 16, 64])]; + tensor q_19_end_mask_0 = const()[name = tensor("q_19_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor q_19_squeeze_mask_0 = const()[name = tensor("q_19_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor q_19_cast_fp16 = slice_by_index(begin = q_19_begin_0, end = q_19_end_0, end_mask = q_19_end_mask_0, squeeze_mask = q_19_squeeze_mask_0, x = qkv_7_cast_fp16)[name = tensor("q_19_cast_fp16")]; + tensor k_13_begin_0 = const()[name = tensor("k_13_begin_0"), val = tensor([0, 0, 1, 0, 0])]; + tensor k_13_end_0 = const()[name = tensor("k_13_end_0"), val = tensor([1, 256, 2, 16, 64])]; + tensor k_13_end_mask_0 = const()[name = tensor("k_13_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor k_13_squeeze_mask_0 = const()[name = tensor("k_13_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor k_13_cast_fp16 = slice_by_index(begin = k_13_begin_0, end = k_13_end_0, end_mask = k_13_end_mask_0, squeeze_mask = k_13_squeeze_mask_0, x = qkv_7_cast_fp16)[name = tensor("k_13_cast_fp16")]; + tensor v_7_begin_0 = const()[name = tensor("v_7_begin_0"), val = tensor([0, 0, 2, 0, 0])]; + tensor v_7_end_0 = const()[name = tensor("v_7_end_0"), val = tensor([1, 256, 3, 16, 64])]; + tensor v_7_end_mask_0 = const()[name = tensor("v_7_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor v_7_squeeze_mask_0 = const()[name = tensor("v_7_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor v_7_cast_fp16 = slice_by_index(begin = v_7_begin_0, end = v_7_end_0, end_mask = v_7_end_mask_0, squeeze_mask = v_7_squeeze_mask_0, x = qkv_7_cast_fp16)[name = tensor("v_7_cast_fp16")]; + tensor var_1353 = const()[name = tensor("op_1353"), val = tensor([1, 1])]; + tensor position3_to_fp16_dtype_0 = const()[name = tensor("position3_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor position3_to_fp16 = cast(dtype = position3_to_fp16_dtype_0, x = position3)[name = tensor("cast_142")]; + tensor var_1354_cast_fp16 = reshape(shape = var_1353, x = position3_to_fp16)[name = tensor("op_1354_cast_fp16")]; + tensor ts_21_cast_fp16 = add(x = ts_1_promoted_to_fp16, y = var_1354_cast_fp16)[name = tensor("ts_21_cast_fp16")]; + tensor var_1361 = const()[name = tensor("op_1361"), val = tensor([1, 256, 1, 1])]; + tensor ts_23_cast_fp16 = reshape(shape = var_1361, x = ts_21_cast_fp16)[name = tensor("ts_23_cast_fp16")]; + tensor var_1368 = const()[name = tensor("op_1368"), val = tensor([1, 256, 16, 32, 2])]; + tensor q_complex_7_cast_fp16 = reshape(shape = var_1368, x = q_19_cast_fp16)[name = tensor("q_complex_7_cast_fp16")]; + tensor var_1375 = const()[name = tensor("op_1375"), val = tensor([1, 256, 16, 32, 2])]; + tensor k_complex_7_cast_fp16 = reshape(shape = var_1375, x = k_13_cast_fp16)[name = tensor("k_complex_7_cast_fp16")]; + tensor var_1379_begin_0 = const()[name = tensor("op_1379_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_1379_end_0 = const()[name = tensor("op_1379_end_0"), val = tensor([1, 256, 16, 32, 1])]; + tensor var_1379_end_mask_0 = const()[name = tensor("op_1379_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_1379_squeeze_mask_0 = const()[name = tensor("op_1379_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_1379_cast_fp16 = slice_by_index(begin = var_1379_begin_0, end = var_1379_end_0, end_mask = var_1379_end_mask_0, squeeze_mask = var_1379_squeeze_mask_0, x = q_complex_7_cast_fp16)[name = tensor("op_1379_cast_fp16")]; + tensor var_1387_begin_0 = const()[name = tensor("op_1387_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_1387_end_0 = const()[name = tensor("op_1387_end_0"), val = tensor([1, 256, 16, 32, 2])]; + tensor var_1387_end_mask_0 = const()[name = tensor("op_1387_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_1387_squeeze_mask_0 = const()[name = tensor("op_1387_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_1387_cast_fp16 = slice_by_index(begin = var_1387_begin_0, end = var_1387_end_0, end_mask = var_1387_end_mask_0, squeeze_mask = var_1387_squeeze_mask_0, x = q_complex_7_cast_fp16)[name = tensor("op_1387_cast_fp16")]; + tensor var_1395_begin_0 = const()[name = tensor("op_1395_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_1395_end_0 = const()[name = tensor("op_1395_end_0"), val = tensor([1, 256, 16, 32, 1])]; + tensor var_1395_end_mask_0 = const()[name = tensor("op_1395_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_1395_squeeze_mask_0 = const()[name = tensor("op_1395_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_1395_cast_fp16 = slice_by_index(begin = var_1395_begin_0, end = var_1395_end_0, end_mask = var_1395_end_mask_0, squeeze_mask = var_1395_squeeze_mask_0, x = k_complex_7_cast_fp16)[name = tensor("op_1395_cast_fp16")]; + tensor var_1403_begin_0 = const()[name = tensor("op_1403_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_1403_end_0 = const()[name = tensor("op_1403_end_0"), val = tensor([1, 256, 16, 32, 2])]; + tensor var_1403_end_mask_0 = const()[name = tensor("op_1403_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_1403_squeeze_mask_0 = const()[name = tensor("op_1403_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_1403_cast_fp16 = slice_by_index(begin = var_1403_begin_0, end = var_1403_end_0, end_mask = var_1403_end_mask_0, squeeze_mask = var_1403_squeeze_mask_0, x = k_complex_7_cast_fp16)[name = tensor("op_1403_cast_fp16")]; + tensor freqs_7_to_fp16 = const()[name = tensor("freqs_7_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(82887808)))]; + tensor var_1409_cast_fp16 = mul(x = freqs_7_to_fp16, y = ts_23_cast_fp16)[name = tensor("op_1409_cast_fp16")]; + tensor rotr_7_cast_fp16 = cos(x = var_1409_cast_fp16)[name = tensor("rotr_7_cast_fp16")]; + tensor roti_7_cast_fp16 = sin(x = var_1409_cast_fp16)[name = tensor("roti_7_cast_fp16")]; + tensor var_1413_cast_fp16 = mul(x = var_1379_cast_fp16, y = rotr_7_cast_fp16)[name = tensor("op_1413_cast_fp16")]; + tensor var_1414_cast_fp16 = mul(x = var_1387_cast_fp16, y = roti_7_cast_fp16)[name = tensor("op_1414_cast_fp16")]; + tensor qor_13_cast_fp16 = sub(x = var_1413_cast_fp16, y = var_1414_cast_fp16)[name = tensor("qor_13_cast_fp16")]; + tensor var_1417_cast_fp16 = mul(x = var_1379_cast_fp16, y = roti_7_cast_fp16)[name = tensor("op_1417_cast_fp16")]; + tensor var_1418_cast_fp16 = mul(x = var_1387_cast_fp16, y = rotr_7_cast_fp16)[name = tensor("op_1418_cast_fp16")]; + tensor qoi_13_cast_fp16 = add(x = var_1417_cast_fp16, y = var_1418_cast_fp16)[name = tensor("qoi_13_cast_fp16")]; + tensor var_1421_cast_fp16 = mul(x = var_1395_cast_fp16, y = rotr_7_cast_fp16)[name = tensor("op_1421_cast_fp16")]; + tensor var_1422_cast_fp16 = mul(x = var_1403_cast_fp16, y = roti_7_cast_fp16)[name = tensor("op_1422_cast_fp16")]; + tensor kor_13_cast_fp16 = sub(x = var_1421_cast_fp16, y = var_1422_cast_fp16)[name = tensor("kor_13_cast_fp16")]; + tensor var_1425_cast_fp16 = mul(x = var_1395_cast_fp16, y = roti_7_cast_fp16)[name = tensor("op_1425_cast_fp16")]; + tensor var_1426_cast_fp16 = mul(x = var_1403_cast_fp16, y = rotr_7_cast_fp16)[name = tensor("op_1426_cast_fp16")]; + tensor koi_13_cast_fp16 = add(x = var_1425_cast_fp16, y = var_1426_cast_fp16)[name = tensor("koi_13_cast_fp16")]; + tensor qo_7_axis_0 = const()[name = tensor("qo_7_axis_0"), val = tensor(-1)]; + tensor qo_7_cast_fp16 = stack(axis = qo_7_axis_0, values = (qor_13_cast_fp16, qoi_13_cast_fp16))[name = tensor("qo_7_cast_fp16")]; + tensor ko_7_axis_0 = const()[name = tensor("ko_7_axis_0"), val = tensor(-1)]; + tensor ko_7_cast_fp16 = stack(axis = ko_7_axis_0, values = (kor_13_cast_fp16, koi_13_cast_fp16))[name = tensor("ko_7_cast_fp16")]; + tensor var_1459 = const()[name = tensor("op_1459"), val = tensor([1, 256, 16, 64])]; + tensor q_21_cast_fp16 = reshape(shape = var_1459, x = qo_7_cast_fp16)[name = tensor("q_21_cast_fp16")]; + tensor var_1465 = const()[name = tensor("op_1465"), val = tensor([1, 256, 16, 64])]; + tensor k_15_cast_fp16 = reshape(shape = var_1465, x = ko_7_cast_fp16)[name = tensor("k_15_cast_fp16")]; + tensor write_indices_float_19_cast_fp16 = add(x = var_1354_cast_fp16, y = var_249_promoted_to_fp16)[name = tensor("write_indices_float_19_cast_fp16")]; + tensor _inversed_1486_y_0_to_fp16 = const()[name = tensor("_inversed_1486_y_0_to_fp16"), val = tensor(0x1p-9)]; + tensor _inversed_1486_cast_fp16 = mul(x = write_indices_float_19_cast_fp16, y = _inversed_1486_y_0_to_fp16)[name = tensor("_inversed_1486_cast_fp16")]; + tensor var_1487_cast_fp16 = floor(x = _inversed_1486_cast_fp16)[name = tensor("op_1487_cast_fp16")]; + tensor var_1488_to_fp16 = const()[name = tensor("op_1488_to_fp16"), val = tensor(0x1p+9)]; + tensor var_1489_cast_fp16 = mul(x = var_1487_cast_fp16, y = var_1488_to_fp16)[name = tensor("op_1489_cast_fp16")]; + tensor write_indices_float_21_cast_fp16 = sub(x = write_indices_float_19_cast_fp16, y = var_1489_cast_fp16)[name = tensor("write_indices_float_21_cast_fp16")]; + tensor var_1507_value_0_to_fp16 = const()[name = tensor("op_1507_value_0_to_fp16"), val = tensor(0x1.ffp+8)]; + tensor var_1507_cast_fp16 = fill_like(ref_tensor = write_indices_float_21_cast_fp16, value = var_1507_value_0_to_fp16)[name = tensor("op_1507_cast_fp16")]; + tensor write_indices_float_23_cast_fp16 = select(a = write_indices_float_21_cast_fp16, b = var_1507_cast_fp16, cond = token_valid_1_cast_fp16)[name = tensor("write_indices_float_23_cast_fp16")]; + tensor var_1513_dtype_0 = const()[name = tensor("op_1513_dtype_0"), val = tensor("int32")]; + tensor var_1518 = const()[name = tensor("op_1518"), val = tensor([1, 256, 1, 1])]; + tensor write_indices_float_23_cast_fp16_to_int32 = cast(dtype = var_1513_dtype_0, x = write_indices_float_23_cast_fp16)[name = tensor("cast_141")]; + tensor var_1519 = reshape(shape = var_1518, x = write_indices_float_23_cast_fp16_to_int32)[name = tensor("op_1519")]; + tensor write_indices_7_reps_0 = const()[name = tensor("write_indices_7_reps_0"), val = tensor([1, 1, 16, 64])]; + tensor write_indices_7 = tile(reps = write_indices_7_reps_0, x = var_1519)[name = tensor("write_indices_7")]; + tensor var_1529_begin_0 = const()[name = tensor("op_1529_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_1529_end_0 = const()[name = tensor("op_1529_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor var_1529_end_mask_0 = const()[name = tensor("op_1529_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_1529_squeeze_mask_0 = const()[name = tensor("op_1529_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor cache3_to_fp16_dtype_0 = const()[name = tensor("cache3_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor cache3_to_fp16 = cast(dtype = cache3_to_fp16_dtype_0, x = cache3)[name = tensor("cast_140")]; + tensor var_1529_cast_fp16 = slice_by_index(begin = var_1529_begin_0, end = var_1529_end_0, end_mask = var_1529_end_mask_0, squeeze_mask = var_1529_squeeze_mask_0, x = cache3_to_fp16)[name = tensor("op_1529_cast_fp16")]; + tensor var_1531_axis_0 = const()[name = tensor("op_1531_axis_0"), val = tensor(1)]; + tensor var_1531_mode_0 = const()[name = tensor("op_1531_mode_0"), val = tensor("update")]; + tensor var_1531_validate_indices_0 = const()[name = tensor("op_1531_validate_indices_0"), val = tensor(false)]; + tensor var_1531_cast_fp16 = scatter_along_axis(axis = var_1531_axis_0, data = var_1529_cast_fp16, indices = write_indices_7, mode = var_1531_mode_0, updates = k_15_cast_fp16, validate_indices = var_1531_validate_indices_0)[name = tensor("op_1531_cast_fp16")]; + tensor concat_26 = const()[name = tensor("concat_26"), val = tensor([0, 0, 0, 0, 0])]; + tensor concat_27 = const()[name = tensor("concat_27"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_7_internal_tensor_assign_1_stride_0 = const()[name = tensor("new_cache_7_internal_tensor_assign_1_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_7_internal_tensor_assign_1_begin_mask_0 = const()[name = tensor("new_cache_7_internal_tensor_assign_1_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_7_internal_tensor_assign_1_end_mask_0 = const()[name = tensor("new_cache_7_internal_tensor_assign_1_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_7_internal_tensor_assign_1_squeeze_mask_0 = const()[name = tensor("new_cache_7_internal_tensor_assign_1_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_16 = const()[name = tensor("shape_16"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_6 = const()[name = tensor("reduce_prod_6"), val = tensor(1048576)]; + tensor range_1d_6_start_0 = const()[name = tensor("range_1d_6_start_0"), val = tensor(0)]; + tensor range_1d_6_step_0 = const()[name = tensor("range_1d_6_step_0"), val = tensor(1)]; + tensor range_1d_6 = range_1d(end = reduce_prod_6, start = range_1d_6_start_0, step = range_1d_6_step_0)[name = tensor("range_1d_6")]; + tensor reshape_30 = reshape(shape = shape_16, x = range_1d_6)[name = tensor("reshape_30")]; + tensor slice_by_index_6 = slice_by_index(begin = concat_26, begin_mask = new_cache_7_internal_tensor_assign_1_begin_mask_0, end = concat_27, end_mask = new_cache_7_internal_tensor_assign_1_end_mask_0, squeeze_mask = new_cache_7_internal_tensor_assign_1_squeeze_mask_0, stride = new_cache_7_internal_tensor_assign_1_stride_0, x = reshape_30)[name = tensor("slice_by_index_6")]; + tensor reshape_31_shape_0 = const()[name = tensor("reshape_31_shape_0"), val = tensor([-1])]; + tensor reshape_31 = reshape(shape = reshape_31_shape_0, x = slice_by_index_6)[name = tensor("reshape_31")]; + tensor reshape_32_shape_0 = const()[name = tensor("reshape_32_shape_0"), val = tensor([-1])]; + tensor reshape_32_cast_fp16 = reshape(shape = reshape_32_shape_0, x = var_1531_cast_fp16)[name = tensor("reshape_32_cast_fp16")]; + tensor reshape_33_shape_0 = const()[name = tensor("reshape_33_shape_0"), val = tensor([-1])]; + tensor reshape_33_cast_fp16 = reshape(shape = reshape_33_shape_0, x = cache3_to_fp16)[name = tensor("reshape_33_cast_fp16")]; + tensor scatter_6_mode_0 = const()[name = tensor("scatter_6_mode_0"), val = tensor("update")]; + tensor scatter_6_axis_0 = const()[name = tensor("scatter_6_axis_0"), val = tensor(0)]; + tensor scatter_6_validate_indices_0 = const()[name = tensor("scatter_6_validate_indices_0"), val = tensor(false)]; + tensor scatter_6_cast_fp16 = scatter(axis = scatter_6_axis_0, data = reshape_33_cast_fp16, indices = reshape_31, mode = scatter_6_mode_0, updates = reshape_32_cast_fp16, validate_indices = scatter_6_validate_indices_0)[name = tensor("scatter_6_cast_fp16")]; + tensor reshape_34_cast_fp16 = reshape(shape = shape_16, x = scatter_6_cast_fp16)[name = tensor("reshape_34_cast_fp16")]; + tensor var_1539_begin_0 = const()[name = tensor("op_1539_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor var_1539_end_0 = const()[name = tensor("op_1539_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor var_1539_end_mask_0 = const()[name = tensor("op_1539_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_1539_squeeze_mask_0 = const()[name = tensor("op_1539_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor var_1539_cast_fp16 = slice_by_index(begin = var_1539_begin_0, end = var_1539_end_0, end_mask = var_1539_end_mask_0, squeeze_mask = var_1539_squeeze_mask_0, x = reshape_34_cast_fp16)[name = tensor("op_1539_cast_fp16")]; + tensor var_1541_axis_0 = const()[name = tensor("op_1541_axis_0"), val = tensor(1)]; + tensor var_1541_mode_0 = const()[name = tensor("op_1541_mode_0"), val = tensor("update")]; + tensor var_1541_validate_indices_0 = const()[name = tensor("op_1541_validate_indices_0"), val = tensor(false)]; + tensor var_1541_cast_fp16 = scatter_along_axis(axis = var_1541_axis_0, data = var_1539_cast_fp16, indices = write_indices_7, mode = var_1541_mode_0, updates = v_7_cast_fp16, validate_indices = var_1541_validate_indices_0)[name = tensor("op_1541_cast_fp16")]; + tensor concat_28 = const()[name = tensor("concat_28"), val = tensor([1, 0, 0, 0, 0])]; + tensor concat_29 = const()[name = tensor("concat_29"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_7_internal_tensor_assign_2_stride_0 = const()[name = tensor("new_cache_7_internal_tensor_assign_2_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_7_internal_tensor_assign_2_begin_mask_0 = const()[name = tensor("new_cache_7_internal_tensor_assign_2_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_7_internal_tensor_assign_2_end_mask_0 = const()[name = tensor("new_cache_7_internal_tensor_assign_2_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_7_internal_tensor_assign_2_squeeze_mask_0 = const()[name = tensor("new_cache_7_internal_tensor_assign_2_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_17 = const()[name = tensor("shape_17"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_7 = const()[name = tensor("reduce_prod_7"), val = tensor(1048576)]; + tensor range_1d_7_start_0 = const()[name = tensor("range_1d_7_start_0"), val = tensor(0)]; + tensor range_1d_7_step_0 = const()[name = tensor("range_1d_7_step_0"), val = tensor(1)]; + tensor range_1d_7 = range_1d(end = reduce_prod_7, start = range_1d_7_start_0, step = range_1d_7_step_0)[name = tensor("range_1d_7")]; + tensor reshape_35 = reshape(shape = shape_17, x = range_1d_7)[name = tensor("reshape_35")]; + tensor slice_by_index_7 = slice_by_index(begin = concat_28, begin_mask = new_cache_7_internal_tensor_assign_2_begin_mask_0, end = concat_29, end_mask = new_cache_7_internal_tensor_assign_2_end_mask_0, squeeze_mask = new_cache_7_internal_tensor_assign_2_squeeze_mask_0, stride = new_cache_7_internal_tensor_assign_2_stride_0, x = reshape_35)[name = tensor("slice_by_index_7")]; + tensor reshape_36_shape_0 = const()[name = tensor("reshape_36_shape_0"), val = tensor([-1])]; + tensor reshape_36 = reshape(shape = reshape_36_shape_0, x = slice_by_index_7)[name = tensor("reshape_36")]; + tensor reshape_37_shape_0 = const()[name = tensor("reshape_37_shape_0"), val = tensor([-1])]; + tensor reshape_37_cast_fp16 = reshape(shape = reshape_37_shape_0, x = var_1541_cast_fp16)[name = tensor("reshape_37_cast_fp16")]; + tensor reshape_38_shape_0 = const()[name = tensor("reshape_38_shape_0"), val = tensor([-1])]; + tensor reshape_38_cast_fp16 = reshape(shape = reshape_38_shape_0, x = reshape_34_cast_fp16)[name = tensor("reshape_38_cast_fp16")]; + tensor scatter_7_mode_0 = const()[name = tensor("scatter_7_mode_0"), val = tensor("update")]; + tensor scatter_7_axis_0 = const()[name = tensor("scatter_7_axis_0"), val = tensor(0)]; + tensor scatter_7_validate_indices_0 = const()[name = tensor("scatter_7_validate_indices_0"), val = tensor(false)]; + tensor scatter_7_cast_fp16 = scatter(axis = scatter_7_axis_0, data = reshape_38_cast_fp16, indices = reshape_36, mode = scatter_7_mode_0, updates = reshape_37_cast_fp16, validate_indices = scatter_7_validate_indices_0)[name = tensor("scatter_7_cast_fp16")]; + tensor reshape_39_cast_fp16 = reshape(shape = shape_17, x = scatter_7_cast_fp16)[name = tensor("reshape_39_cast_fp16")]; + tensor reshape_39_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("reshape_39_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor keys_19_begin_0 = const()[name = tensor("keys_19_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor keys_19_end_0 = const()[name = tensor("keys_19_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor keys_19_end_mask_0 = const()[name = tensor("keys_19_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor keys_19_squeeze_mask_0 = const()[name = tensor("keys_19_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor keys_19_cast_fp16 = slice_by_index(begin = keys_19_begin_0, end = keys_19_end_0, end_mask = keys_19_end_mask_0, squeeze_mask = keys_19_squeeze_mask_0, x = reshape_39_cast_fp16)[name = tensor("keys_19_cast_fp16")]; + tensor values_19_begin_0 = const()[name = tensor("values_19_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor values_19_end_0 = const()[name = tensor("values_19_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor values_19_end_mask_0 = const()[name = tensor("values_19_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor values_19_squeeze_mask_0 = const()[name = tensor("values_19_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor values_19_cast_fp16 = slice_by_index(begin = values_19_begin_0, end = values_19_end_0, end_mask = values_19_end_mask_0, squeeze_mask = values_19_squeeze_mask_0, x = reshape_39_cast_fp16)[name = tensor("values_19_cast_fp16")]; + tensor var_1553_cast_fp16 = not_equal(x = keys_19_cast_fp16, y = keys_19_cast_fp16)[name = tensor("op_1553_cast_fp16")]; + tensor keys_21_cast_fp16 = select(a = var_326_to_fp16, b = keys_19_cast_fp16, cond = var_1553_cast_fp16)[name = tensor("keys_21_cast_fp16")]; + tensor var_1561_cast_fp16 = not_equal(x = values_19_cast_fp16, y = values_19_cast_fp16)[name = tensor("op_1561_cast_fp16")]; + tensor values_21_cast_fp16 = select(a = var_326_to_fp16, b = values_19_cast_fp16, cond = var_1561_cast_fp16)[name = tensor("values_21_cast_fp16")]; + tensor var_1585 = const()[name = tensor("op_1585"), val = tensor([0, 2, 1, 3])]; + tensor var_1601 = const()[name = tensor("op_1601"), val = tensor([1, 1, 1])]; + tensor var_1602_cast_fp16 = reshape(shape = var_1601, x = position3_to_fp16)[name = tensor("op_1602_cast_fp16")]; + tensor q_positions_7_cast_fp16 = add(x = var_1602_cast_fp16, y = q_offsets_1_promoted_to_fp16)[name = tensor("q_positions_7_cast_fp16")]; + tensor valid_end_7_cast_fp16 = add(x = var_1602_cast_fp16, y = var_392_cast_fp16)[name = tensor("valid_end_7_cast_fp16")]; + tensor valid_mask_7_cast_fp16 = less(x = k_positions_1_promoted_to_fp16, y = valid_end_7_cast_fp16)[name = tensor("valid_mask_7_cast_fp16")]; + tensor causal_mask_7_cast_fp16 = less_equal(x = k_positions_1_promoted_to_fp16, y = q_positions_7_cast_fp16)[name = tensor("causal_mask_7_cast_fp16")]; + tensor var_1630 = logical_and(x = valid_mask_7_cast_fp16, y = causal_mask_7_cast_fp16)[name = tensor("op_1630")]; + tensor attn_mask_7_axes_0 = const()[name = tensor("attn_mask_7_axes_0"), val = tensor([1])]; + tensor attn_mask_7 = expand_dims(axes = attn_mask_7_axes_0, x = var_1630)[name = tensor("attn_mask_7")]; + tensor var_1644_transpose_x_0 = const()[name = tensor("op_1644_transpose_x_0"), val = tensor(false)]; + tensor var_1644_transpose_y_0 = const()[name = tensor("op_1644_transpose_y_0"), val = tensor(false)]; + tensor transpose_21_perm_0 = const()[name = tensor("transpose_21_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_22_perm_0 = const()[name = tensor("transpose_22_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_22 = transpose(perm = transpose_22_perm_0, x = keys_21_cast_fp16)[name = tensor("transpose_30")]; + tensor transpose_21 = transpose(perm = transpose_21_perm_0, x = q_21_cast_fp16)[name = tensor("transpose_31")]; + tensor var_1644_cast_fp16 = matmul(transpose_x = var_1644_transpose_x_0, transpose_y = var_1644_transpose_y_0, x = transpose_21, y = transpose_22)[name = tensor("op_1644_cast_fp16")]; + tensor var_1638_to_fp16 = const()[name = tensor("op_1638_to_fp16"), val = tensor([0x1p-3])]; + tensor attn_weights_19_cast_fp16 = mul(x = var_1644_cast_fp16, y = var_1638_to_fp16)[name = tensor("attn_weights_19_cast_fp16")]; + tensor var_1646 = logical_not(x = attn_mask_7)[name = tensor("op_1646")]; + tensor var_1647_to_fp16 = const()[name = tensor("op_1647_to_fp16"), val = tensor(-inf)]; + tensor attn_weights_21_cast_fp16 = select(a = var_1647_to_fp16, b = attn_weights_19_cast_fp16, cond = var_1646)[name = tensor("attn_weights_21_cast_fp16")]; + tensor var_1649 = const()[name = tensor("op_1649"), val = tensor(-1)]; + tensor attn_weights_23_cast_fp16 = softmax(axis = var_1649, x = attn_weights_21_cast_fp16)[name = tensor("attn_weights_23_cast_fp16")]; + tensor attn_output_7_transpose_x_0 = const()[name = tensor("attn_output_7_transpose_x_0"), val = tensor(false)]; + tensor attn_output_7_transpose_y_0 = const()[name = tensor("attn_output_7_transpose_y_0"), val = tensor(false)]; + tensor values_23_cast_fp16 = transpose(perm = var_1585, x = values_21_cast_fp16)[name = tensor("transpose_32")]; + tensor attn_output_7_cast_fp16 = matmul(transpose_x = attn_output_7_transpose_x_0, transpose_y = attn_output_7_transpose_y_0, x = attn_weights_23_cast_fp16, y = values_23_cast_fp16)[name = tensor("attn_output_7_cast_fp16")]; + tensor var_1657 = const()[name = tensor("op_1657"), val = tensor([0, 2, 1, 3])]; + tensor var_1662 = const()[name = tensor("op_1662"), val = tensor([1, 256, 1024])]; + tensor var_1658_cast_fp16 = transpose(perm = var_1657, x = attn_output_7_cast_fp16)[name = tensor("transpose_29")]; + tensor input_41_cast_fp16 = reshape(shape = var_1662, x = var_1658_cast_fp16)[name = tensor("input_41_cast_fp16")]; + tensor attn3_out_proj_weight_to_fp16 = const()[name = tensor("attn3_out_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(82887936)))]; + tensor linear_13_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = attn3_out_proj_weight_to_fp16, x = input_41_cast_fp16)[name = tensor("linear_13_cast_fp16")]; + tensor var_1671_cast_fp16 = add(x = position3_to_fp16, y = valid_len_to_fp16)[name = tensor("op_1671_cast_fp16")]; + tensor var_1671_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_1671_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor input_43_cast_fp16 = add(x = input_37_cast_fp16, y = linear_13_cast_fp16)[name = tensor("input_43_cast_fp16")]; + tensor input_45_axes_0 = const()[name = tensor("input_45_axes_0"), val = tensor([-1])]; + tensor norm3_2_weight_to_fp16 = const()[name = tensor("norm3_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(84985152)))]; + tensor norm3_2_bias_to_fp16 = const()[name = tensor("norm3_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(84987264)))]; + tensor var_1675_to_fp16 = const()[name = tensor("op_1675_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_45_cast_fp16 = layer_norm(axes = input_45_axes_0, beta = norm3_2_bias_to_fp16, epsilon = var_1675_to_fp16, gamma = norm3_2_weight_to_fp16, x = input_43_cast_fp16)[name = tensor("input_45_cast_fp16")]; + tensor linear3_1_weight_to_fp16 = const()[name = tensor("linear3_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(84989376)))]; + tensor linear_14_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = linear3_1_weight_to_fp16, x = input_45_cast_fp16)[name = tensor("linear_14_cast_fp16")]; + tensor input_47_mode_0 = const()[name = tensor("input_47_mode_0"), val = tensor("EXACT")]; + tensor input_47_cast_fp16 = gelu(mode = input_47_mode_0, x = linear_14_cast_fp16)[name = tensor("input_47_cast_fp16")]; + tensor linear3_2_weight_to_fp16 = const()[name = tensor("linear3_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(93378048)))]; + tensor linear_15_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = linear3_2_weight_to_fp16, x = input_47_cast_fp16)[name = tensor("linear_15_cast_fp16")]; + tensor input_49_cast_fp16 = add(x = input_43_cast_fp16, y = linear_15_cast_fp16)[name = tensor("input_49_cast_fp16")]; + tensor input_51_axes_0 = const()[name = tensor("input_51_axes_0"), val = tensor([-1])]; + tensor norm4_1_weight_to_fp16 = const()[name = tensor("norm4_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(101766720)))]; + tensor norm4_1_bias_to_fp16 = const()[name = tensor("norm4_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(101768832)))]; + tensor var_1692_to_fp16 = const()[name = tensor("op_1692_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_51_cast_fp16 = layer_norm(axes = input_51_axes_0, beta = norm4_1_bias_to_fp16, epsilon = var_1692_to_fp16, gamma = norm4_1_weight_to_fp16, x = input_49_cast_fp16)[name = tensor("input_51_cast_fp16")]; + tensor attn4_in_proj_weight_to_fp16 = const()[name = tensor("attn4_in_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(101770944)))]; + tensor linear_16_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = attn4_in_proj_weight_to_fp16, x = input_51_cast_fp16)[name = tensor("linear_16_cast_fp16")]; + tensor var_1706 = const()[name = tensor("op_1706"), val = tensor([1, 256, 3, 16, 64])]; + tensor qkv_9_cast_fp16 = reshape(shape = var_1706, x = linear_16_cast_fp16)[name = tensor("qkv_9_cast_fp16")]; + tensor q_25_begin_0 = const()[name = tensor("q_25_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor q_25_end_0 = const()[name = tensor("q_25_end_0"), val = tensor([1, 256, 1, 16, 64])]; + tensor q_25_end_mask_0 = const()[name = tensor("q_25_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor q_25_squeeze_mask_0 = const()[name = tensor("q_25_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor q_25_cast_fp16 = slice_by_index(begin = q_25_begin_0, end = q_25_end_0, end_mask = q_25_end_mask_0, squeeze_mask = q_25_squeeze_mask_0, x = qkv_9_cast_fp16)[name = tensor("q_25_cast_fp16")]; + tensor k_17_begin_0 = const()[name = tensor("k_17_begin_0"), val = tensor([0, 0, 1, 0, 0])]; + tensor k_17_end_0 = const()[name = tensor("k_17_end_0"), val = tensor([1, 256, 2, 16, 64])]; + tensor k_17_end_mask_0 = const()[name = tensor("k_17_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor k_17_squeeze_mask_0 = const()[name = tensor("k_17_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor k_17_cast_fp16 = slice_by_index(begin = k_17_begin_0, end = k_17_end_0, end_mask = k_17_end_mask_0, squeeze_mask = k_17_squeeze_mask_0, x = qkv_9_cast_fp16)[name = tensor("k_17_cast_fp16")]; + tensor v_9_begin_0 = const()[name = tensor("v_9_begin_0"), val = tensor([0, 0, 2, 0, 0])]; + tensor v_9_end_0 = const()[name = tensor("v_9_end_0"), val = tensor([1, 256, 3, 16, 64])]; + tensor v_9_end_mask_0 = const()[name = tensor("v_9_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor v_9_squeeze_mask_0 = const()[name = tensor("v_9_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor v_9_cast_fp16 = slice_by_index(begin = v_9_begin_0, end = v_9_end_0, end_mask = v_9_end_mask_0, squeeze_mask = v_9_squeeze_mask_0, x = qkv_9_cast_fp16)[name = tensor("v_9_cast_fp16")]; + tensor var_1764 = const()[name = tensor("op_1764"), val = tensor([1, 1])]; + tensor position4_to_fp16_dtype_0 = const()[name = tensor("position4_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor position4_to_fp16 = cast(dtype = position4_to_fp16_dtype_0, x = position4)[name = tensor("cast_137")]; + tensor var_1765_cast_fp16 = reshape(shape = var_1764, x = position4_to_fp16)[name = tensor("op_1765_cast_fp16")]; + tensor ts_27_cast_fp16 = add(x = ts_1_promoted_to_fp16, y = var_1765_cast_fp16)[name = tensor("ts_27_cast_fp16")]; + tensor var_1772 = const()[name = tensor("op_1772"), val = tensor([1, 256, 1, 1])]; + tensor ts_29_cast_fp16 = reshape(shape = var_1772, x = ts_27_cast_fp16)[name = tensor("ts_29_cast_fp16")]; + tensor var_1779 = const()[name = tensor("op_1779"), val = tensor([1, 256, 16, 32, 2])]; + tensor q_complex_9_cast_fp16 = reshape(shape = var_1779, x = q_25_cast_fp16)[name = tensor("q_complex_9_cast_fp16")]; + tensor var_1786 = const()[name = tensor("op_1786"), val = tensor([1, 256, 16, 32, 2])]; + tensor k_complex_9_cast_fp16 = reshape(shape = var_1786, x = k_17_cast_fp16)[name = tensor("k_complex_9_cast_fp16")]; + tensor var_1790_begin_0 = const()[name = tensor("op_1790_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_1790_end_0 = const()[name = tensor("op_1790_end_0"), val = tensor([1, 256, 16, 32, 1])]; + tensor var_1790_end_mask_0 = const()[name = tensor("op_1790_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_1790_squeeze_mask_0 = const()[name = tensor("op_1790_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_1790_cast_fp16 = slice_by_index(begin = var_1790_begin_0, end = var_1790_end_0, end_mask = var_1790_end_mask_0, squeeze_mask = var_1790_squeeze_mask_0, x = q_complex_9_cast_fp16)[name = tensor("op_1790_cast_fp16")]; + tensor var_1798_begin_0 = const()[name = tensor("op_1798_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_1798_end_0 = const()[name = tensor("op_1798_end_0"), val = tensor([1, 256, 16, 32, 2])]; + tensor var_1798_end_mask_0 = const()[name = tensor("op_1798_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_1798_squeeze_mask_0 = const()[name = tensor("op_1798_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_1798_cast_fp16 = slice_by_index(begin = var_1798_begin_0, end = var_1798_end_0, end_mask = var_1798_end_mask_0, squeeze_mask = var_1798_squeeze_mask_0, x = q_complex_9_cast_fp16)[name = tensor("op_1798_cast_fp16")]; + tensor var_1806_begin_0 = const()[name = tensor("op_1806_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_1806_end_0 = const()[name = tensor("op_1806_end_0"), val = tensor([1, 256, 16, 32, 1])]; + tensor var_1806_end_mask_0 = const()[name = tensor("op_1806_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_1806_squeeze_mask_0 = const()[name = tensor("op_1806_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_1806_cast_fp16 = slice_by_index(begin = var_1806_begin_0, end = var_1806_end_0, end_mask = var_1806_end_mask_0, squeeze_mask = var_1806_squeeze_mask_0, x = k_complex_9_cast_fp16)[name = tensor("op_1806_cast_fp16")]; + tensor var_1814_begin_0 = const()[name = tensor("op_1814_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_1814_end_0 = const()[name = tensor("op_1814_end_0"), val = tensor([1, 256, 16, 32, 2])]; + tensor var_1814_end_mask_0 = const()[name = tensor("op_1814_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_1814_squeeze_mask_0 = const()[name = tensor("op_1814_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_1814_cast_fp16 = slice_by_index(begin = var_1814_begin_0, end = var_1814_end_0, end_mask = var_1814_end_mask_0, squeeze_mask = var_1814_squeeze_mask_0, x = k_complex_9_cast_fp16)[name = tensor("op_1814_cast_fp16")]; + tensor freqs_9_to_fp16 = const()[name = tensor("freqs_9_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(108062464)))]; + tensor var_1820_cast_fp16 = mul(x = freqs_9_to_fp16, y = ts_29_cast_fp16)[name = tensor("op_1820_cast_fp16")]; + tensor rotr_9_cast_fp16 = cos(x = var_1820_cast_fp16)[name = tensor("rotr_9_cast_fp16")]; + tensor roti_9_cast_fp16 = sin(x = var_1820_cast_fp16)[name = tensor("roti_9_cast_fp16")]; + tensor var_1824_cast_fp16 = mul(x = var_1790_cast_fp16, y = rotr_9_cast_fp16)[name = tensor("op_1824_cast_fp16")]; + tensor var_1825_cast_fp16 = mul(x = var_1798_cast_fp16, y = roti_9_cast_fp16)[name = tensor("op_1825_cast_fp16")]; + tensor qor_17_cast_fp16 = sub(x = var_1824_cast_fp16, y = var_1825_cast_fp16)[name = tensor("qor_17_cast_fp16")]; + tensor var_1828_cast_fp16 = mul(x = var_1790_cast_fp16, y = roti_9_cast_fp16)[name = tensor("op_1828_cast_fp16")]; + tensor var_1829_cast_fp16 = mul(x = var_1798_cast_fp16, y = rotr_9_cast_fp16)[name = tensor("op_1829_cast_fp16")]; + tensor qoi_17_cast_fp16 = add(x = var_1828_cast_fp16, y = var_1829_cast_fp16)[name = tensor("qoi_17_cast_fp16")]; + tensor var_1832_cast_fp16 = mul(x = var_1806_cast_fp16, y = rotr_9_cast_fp16)[name = tensor("op_1832_cast_fp16")]; + tensor var_1833_cast_fp16 = mul(x = var_1814_cast_fp16, y = roti_9_cast_fp16)[name = tensor("op_1833_cast_fp16")]; + tensor kor_17_cast_fp16 = sub(x = var_1832_cast_fp16, y = var_1833_cast_fp16)[name = tensor("kor_17_cast_fp16")]; + tensor var_1836_cast_fp16 = mul(x = var_1806_cast_fp16, y = roti_9_cast_fp16)[name = tensor("op_1836_cast_fp16")]; + tensor var_1837_cast_fp16 = mul(x = var_1814_cast_fp16, y = rotr_9_cast_fp16)[name = tensor("op_1837_cast_fp16")]; + tensor koi_17_cast_fp16 = add(x = var_1836_cast_fp16, y = var_1837_cast_fp16)[name = tensor("koi_17_cast_fp16")]; + tensor qo_9_axis_0 = const()[name = tensor("qo_9_axis_0"), val = tensor(-1)]; + tensor qo_9_cast_fp16 = stack(axis = qo_9_axis_0, values = (qor_17_cast_fp16, qoi_17_cast_fp16))[name = tensor("qo_9_cast_fp16")]; + tensor ko_9_axis_0 = const()[name = tensor("ko_9_axis_0"), val = tensor(-1)]; + tensor ko_9_cast_fp16 = stack(axis = ko_9_axis_0, values = (kor_17_cast_fp16, koi_17_cast_fp16))[name = tensor("ko_9_cast_fp16")]; + tensor var_1870 = const()[name = tensor("op_1870"), val = tensor([1, 256, 16, 64])]; + tensor q_27_cast_fp16 = reshape(shape = var_1870, x = qo_9_cast_fp16)[name = tensor("q_27_cast_fp16")]; + tensor var_1876 = const()[name = tensor("op_1876"), val = tensor([1, 256, 16, 64])]; + tensor k_19_cast_fp16 = reshape(shape = var_1876, x = ko_9_cast_fp16)[name = tensor("k_19_cast_fp16")]; + tensor write_indices_float_25_cast_fp16 = add(x = var_1765_cast_fp16, y = var_249_promoted_to_fp16)[name = tensor("write_indices_float_25_cast_fp16")]; + tensor _inversed_1897_y_0_to_fp16 = const()[name = tensor("_inversed_1897_y_0_to_fp16"), val = tensor(0x1p-9)]; + tensor _inversed_1897_cast_fp16 = mul(x = write_indices_float_25_cast_fp16, y = _inversed_1897_y_0_to_fp16)[name = tensor("_inversed_1897_cast_fp16")]; + tensor var_1898_cast_fp16 = floor(x = _inversed_1897_cast_fp16)[name = tensor("op_1898_cast_fp16")]; + tensor var_1899_to_fp16 = const()[name = tensor("op_1899_to_fp16"), val = tensor(0x1p+9)]; + tensor var_1900_cast_fp16 = mul(x = var_1898_cast_fp16, y = var_1899_to_fp16)[name = tensor("op_1900_cast_fp16")]; + tensor write_indices_float_27_cast_fp16 = sub(x = write_indices_float_25_cast_fp16, y = var_1900_cast_fp16)[name = tensor("write_indices_float_27_cast_fp16")]; + tensor var_1918_value_0_to_fp16 = const()[name = tensor("op_1918_value_0_to_fp16"), val = tensor(0x1.ffp+8)]; + tensor var_1918_cast_fp16 = fill_like(ref_tensor = write_indices_float_27_cast_fp16, value = var_1918_value_0_to_fp16)[name = tensor("op_1918_cast_fp16")]; + tensor write_indices_float_29_cast_fp16 = select(a = write_indices_float_27_cast_fp16, b = var_1918_cast_fp16, cond = token_valid_1_cast_fp16)[name = tensor("write_indices_float_29_cast_fp16")]; + tensor var_1924_dtype_0 = const()[name = tensor("op_1924_dtype_0"), val = tensor("int32")]; + tensor var_1929 = const()[name = tensor("op_1929"), val = tensor([1, 256, 1, 1])]; + tensor write_indices_float_29_cast_fp16_to_int32 = cast(dtype = var_1924_dtype_0, x = write_indices_float_29_cast_fp16)[name = tensor("cast_136")]; + tensor var_1930 = reshape(shape = var_1929, x = write_indices_float_29_cast_fp16_to_int32)[name = tensor("op_1930")]; + tensor write_indices_9_reps_0 = const()[name = tensor("write_indices_9_reps_0"), val = tensor([1, 1, 16, 64])]; + tensor write_indices_9 = tile(reps = write_indices_9_reps_0, x = var_1930)[name = tensor("write_indices_9")]; + tensor var_1940_begin_0 = const()[name = tensor("op_1940_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_1940_end_0 = const()[name = tensor("op_1940_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor var_1940_end_mask_0 = const()[name = tensor("op_1940_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_1940_squeeze_mask_0 = const()[name = tensor("op_1940_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor cache4_to_fp16_dtype_0 = const()[name = tensor("cache4_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor cache4_to_fp16 = cast(dtype = cache4_to_fp16_dtype_0, x = cache4)[name = tensor("cast_135")]; + tensor var_1940_cast_fp16 = slice_by_index(begin = var_1940_begin_0, end = var_1940_end_0, end_mask = var_1940_end_mask_0, squeeze_mask = var_1940_squeeze_mask_0, x = cache4_to_fp16)[name = tensor("op_1940_cast_fp16")]; + tensor var_1942_axis_0 = const()[name = tensor("op_1942_axis_0"), val = tensor(1)]; + tensor var_1942_mode_0 = const()[name = tensor("op_1942_mode_0"), val = tensor("update")]; + tensor var_1942_validate_indices_0 = const()[name = tensor("op_1942_validate_indices_0"), val = tensor(false)]; + tensor var_1942_cast_fp16 = scatter_along_axis(axis = var_1942_axis_0, data = var_1940_cast_fp16, indices = write_indices_9, mode = var_1942_mode_0, updates = k_19_cast_fp16, validate_indices = var_1942_validate_indices_0)[name = tensor("op_1942_cast_fp16")]; + tensor concat_34 = const()[name = tensor("concat_34"), val = tensor([0, 0, 0, 0, 0])]; + tensor concat_35 = const()[name = tensor("concat_35"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_9_internal_tensor_assign_1_stride_0 = const()[name = tensor("new_cache_9_internal_tensor_assign_1_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_9_internal_tensor_assign_1_begin_mask_0 = const()[name = tensor("new_cache_9_internal_tensor_assign_1_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_9_internal_tensor_assign_1_end_mask_0 = const()[name = tensor("new_cache_9_internal_tensor_assign_1_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_9_internal_tensor_assign_1_squeeze_mask_0 = const()[name = tensor("new_cache_9_internal_tensor_assign_1_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_18 = const()[name = tensor("shape_18"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_8 = const()[name = tensor("reduce_prod_8"), val = tensor(1048576)]; + tensor range_1d_8_start_0 = const()[name = tensor("range_1d_8_start_0"), val = tensor(0)]; + tensor range_1d_8_step_0 = const()[name = tensor("range_1d_8_step_0"), val = tensor(1)]; + tensor range_1d_8 = range_1d(end = reduce_prod_8, start = range_1d_8_start_0, step = range_1d_8_step_0)[name = tensor("range_1d_8")]; + tensor reshape_40 = reshape(shape = shape_18, x = range_1d_8)[name = tensor("reshape_40")]; + tensor slice_by_index_8 = slice_by_index(begin = concat_34, begin_mask = new_cache_9_internal_tensor_assign_1_begin_mask_0, end = concat_35, end_mask = new_cache_9_internal_tensor_assign_1_end_mask_0, squeeze_mask = new_cache_9_internal_tensor_assign_1_squeeze_mask_0, stride = new_cache_9_internal_tensor_assign_1_stride_0, x = reshape_40)[name = tensor("slice_by_index_8")]; + tensor reshape_41_shape_0 = const()[name = tensor("reshape_41_shape_0"), val = tensor([-1])]; + tensor reshape_41 = reshape(shape = reshape_41_shape_0, x = slice_by_index_8)[name = tensor("reshape_41")]; + tensor reshape_42_shape_0 = const()[name = tensor("reshape_42_shape_0"), val = tensor([-1])]; + tensor reshape_42_cast_fp16 = reshape(shape = reshape_42_shape_0, x = var_1942_cast_fp16)[name = tensor("reshape_42_cast_fp16")]; + tensor reshape_43_shape_0 = const()[name = tensor("reshape_43_shape_0"), val = tensor([-1])]; + tensor reshape_43_cast_fp16 = reshape(shape = reshape_43_shape_0, x = cache4_to_fp16)[name = tensor("reshape_43_cast_fp16")]; + tensor scatter_8_mode_0 = const()[name = tensor("scatter_8_mode_0"), val = tensor("update")]; + tensor scatter_8_axis_0 = const()[name = tensor("scatter_8_axis_0"), val = tensor(0)]; + tensor scatter_8_validate_indices_0 = const()[name = tensor("scatter_8_validate_indices_0"), val = tensor(false)]; + tensor scatter_8_cast_fp16 = scatter(axis = scatter_8_axis_0, data = reshape_43_cast_fp16, indices = reshape_41, mode = scatter_8_mode_0, updates = reshape_42_cast_fp16, validate_indices = scatter_8_validate_indices_0)[name = tensor("scatter_8_cast_fp16")]; + tensor reshape_44_cast_fp16 = reshape(shape = shape_18, x = scatter_8_cast_fp16)[name = tensor("reshape_44_cast_fp16")]; + tensor var_1950_begin_0 = const()[name = tensor("op_1950_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor var_1950_end_0 = const()[name = tensor("op_1950_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor var_1950_end_mask_0 = const()[name = tensor("op_1950_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_1950_squeeze_mask_0 = const()[name = tensor("op_1950_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor var_1950_cast_fp16 = slice_by_index(begin = var_1950_begin_0, end = var_1950_end_0, end_mask = var_1950_end_mask_0, squeeze_mask = var_1950_squeeze_mask_0, x = reshape_44_cast_fp16)[name = tensor("op_1950_cast_fp16")]; + tensor var_1952_axis_0 = const()[name = tensor("op_1952_axis_0"), val = tensor(1)]; + tensor var_1952_mode_0 = const()[name = tensor("op_1952_mode_0"), val = tensor("update")]; + tensor var_1952_validate_indices_0 = const()[name = tensor("op_1952_validate_indices_0"), val = tensor(false)]; + tensor var_1952_cast_fp16 = scatter_along_axis(axis = var_1952_axis_0, data = var_1950_cast_fp16, indices = write_indices_9, mode = var_1952_mode_0, updates = v_9_cast_fp16, validate_indices = var_1952_validate_indices_0)[name = tensor("op_1952_cast_fp16")]; + tensor concat_36 = const()[name = tensor("concat_36"), val = tensor([1, 0, 0, 0, 0])]; + tensor concat_37 = const()[name = tensor("concat_37"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_9_internal_tensor_assign_2_stride_0 = const()[name = tensor("new_cache_9_internal_tensor_assign_2_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_9_internal_tensor_assign_2_begin_mask_0 = const()[name = tensor("new_cache_9_internal_tensor_assign_2_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_9_internal_tensor_assign_2_end_mask_0 = const()[name = tensor("new_cache_9_internal_tensor_assign_2_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_9_internal_tensor_assign_2_squeeze_mask_0 = const()[name = tensor("new_cache_9_internal_tensor_assign_2_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_19 = const()[name = tensor("shape_19"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_9 = const()[name = tensor("reduce_prod_9"), val = tensor(1048576)]; + tensor range_1d_9_start_0 = const()[name = tensor("range_1d_9_start_0"), val = tensor(0)]; + tensor range_1d_9_step_0 = const()[name = tensor("range_1d_9_step_0"), val = tensor(1)]; + tensor range_1d_9 = range_1d(end = reduce_prod_9, start = range_1d_9_start_0, step = range_1d_9_step_0)[name = tensor("range_1d_9")]; + tensor reshape_45 = reshape(shape = shape_19, x = range_1d_9)[name = tensor("reshape_45")]; + tensor slice_by_index_9 = slice_by_index(begin = concat_36, begin_mask = new_cache_9_internal_tensor_assign_2_begin_mask_0, end = concat_37, end_mask = new_cache_9_internal_tensor_assign_2_end_mask_0, squeeze_mask = new_cache_9_internal_tensor_assign_2_squeeze_mask_0, stride = new_cache_9_internal_tensor_assign_2_stride_0, x = reshape_45)[name = tensor("slice_by_index_9")]; + tensor reshape_46_shape_0 = const()[name = tensor("reshape_46_shape_0"), val = tensor([-1])]; + tensor reshape_46 = reshape(shape = reshape_46_shape_0, x = slice_by_index_9)[name = tensor("reshape_46")]; + tensor reshape_47_shape_0 = const()[name = tensor("reshape_47_shape_0"), val = tensor([-1])]; + tensor reshape_47_cast_fp16 = reshape(shape = reshape_47_shape_0, x = var_1952_cast_fp16)[name = tensor("reshape_47_cast_fp16")]; + tensor reshape_48_shape_0 = const()[name = tensor("reshape_48_shape_0"), val = tensor([-1])]; + tensor reshape_48_cast_fp16 = reshape(shape = reshape_48_shape_0, x = reshape_44_cast_fp16)[name = tensor("reshape_48_cast_fp16")]; + tensor scatter_9_mode_0 = const()[name = tensor("scatter_9_mode_0"), val = tensor("update")]; + tensor scatter_9_axis_0 = const()[name = tensor("scatter_9_axis_0"), val = tensor(0)]; + tensor scatter_9_validate_indices_0 = const()[name = tensor("scatter_9_validate_indices_0"), val = tensor(false)]; + tensor scatter_9_cast_fp16 = scatter(axis = scatter_9_axis_0, data = reshape_48_cast_fp16, indices = reshape_46, mode = scatter_9_mode_0, updates = reshape_47_cast_fp16, validate_indices = scatter_9_validate_indices_0)[name = tensor("scatter_9_cast_fp16")]; + tensor reshape_49_cast_fp16 = reshape(shape = shape_19, x = scatter_9_cast_fp16)[name = tensor("reshape_49_cast_fp16")]; + tensor reshape_49_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("reshape_49_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor keys_25_begin_0 = const()[name = tensor("keys_25_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor keys_25_end_0 = const()[name = tensor("keys_25_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor keys_25_end_mask_0 = const()[name = tensor("keys_25_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor keys_25_squeeze_mask_0 = const()[name = tensor("keys_25_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor keys_25_cast_fp16 = slice_by_index(begin = keys_25_begin_0, end = keys_25_end_0, end_mask = keys_25_end_mask_0, squeeze_mask = keys_25_squeeze_mask_0, x = reshape_49_cast_fp16)[name = tensor("keys_25_cast_fp16")]; + tensor values_25_begin_0 = const()[name = tensor("values_25_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor values_25_end_0 = const()[name = tensor("values_25_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor values_25_end_mask_0 = const()[name = tensor("values_25_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor values_25_squeeze_mask_0 = const()[name = tensor("values_25_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor values_25_cast_fp16 = slice_by_index(begin = values_25_begin_0, end = values_25_end_0, end_mask = values_25_end_mask_0, squeeze_mask = values_25_squeeze_mask_0, x = reshape_49_cast_fp16)[name = tensor("values_25_cast_fp16")]; + tensor var_1964_cast_fp16 = not_equal(x = keys_25_cast_fp16, y = keys_25_cast_fp16)[name = tensor("op_1964_cast_fp16")]; + tensor keys_27_cast_fp16 = select(a = var_326_to_fp16, b = keys_25_cast_fp16, cond = var_1964_cast_fp16)[name = tensor("keys_27_cast_fp16")]; + tensor var_1972_cast_fp16 = not_equal(x = values_25_cast_fp16, y = values_25_cast_fp16)[name = tensor("op_1972_cast_fp16")]; + tensor values_27_cast_fp16 = select(a = var_326_to_fp16, b = values_25_cast_fp16, cond = var_1972_cast_fp16)[name = tensor("values_27_cast_fp16")]; + tensor var_1996 = const()[name = tensor("op_1996"), val = tensor([0, 2, 1, 3])]; + tensor var_2012 = const()[name = tensor("op_2012"), val = tensor([1, 1, 1])]; + tensor var_2013_cast_fp16 = reshape(shape = var_2012, x = position4_to_fp16)[name = tensor("op_2013_cast_fp16")]; + tensor q_positions_9_cast_fp16 = add(x = var_2013_cast_fp16, y = q_offsets_1_promoted_to_fp16)[name = tensor("q_positions_9_cast_fp16")]; + tensor valid_end_9_cast_fp16 = add(x = var_2013_cast_fp16, y = var_392_cast_fp16)[name = tensor("valid_end_9_cast_fp16")]; + tensor valid_mask_9_cast_fp16 = less(x = k_positions_1_promoted_to_fp16, y = valid_end_9_cast_fp16)[name = tensor("valid_mask_9_cast_fp16")]; + tensor causal_mask_9_cast_fp16 = less_equal(x = k_positions_1_promoted_to_fp16, y = q_positions_9_cast_fp16)[name = tensor("causal_mask_9_cast_fp16")]; + tensor var_2041 = logical_and(x = valid_mask_9_cast_fp16, y = causal_mask_9_cast_fp16)[name = tensor("op_2041")]; + tensor attn_mask_9_axes_0 = const()[name = tensor("attn_mask_9_axes_0"), val = tensor([1])]; + tensor attn_mask_9 = expand_dims(axes = attn_mask_9_axes_0, x = var_2041)[name = tensor("attn_mask_9")]; + tensor var_2055_transpose_x_0 = const()[name = tensor("op_2055_transpose_x_0"), val = tensor(false)]; + tensor var_2055_transpose_y_0 = const()[name = tensor("op_2055_transpose_y_0"), val = tensor(false)]; + tensor transpose_23_perm_0 = const()[name = tensor("transpose_23_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_24_perm_0 = const()[name = tensor("transpose_24_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_24 = transpose(perm = transpose_24_perm_0, x = keys_27_cast_fp16)[name = tensor("transpose_26")]; + tensor transpose_23 = transpose(perm = transpose_23_perm_0, x = q_27_cast_fp16)[name = tensor("transpose_27")]; + tensor var_2055_cast_fp16 = matmul(transpose_x = var_2055_transpose_x_0, transpose_y = var_2055_transpose_y_0, x = transpose_23, y = transpose_24)[name = tensor("op_2055_cast_fp16")]; + tensor var_2049_to_fp16 = const()[name = tensor("op_2049_to_fp16"), val = tensor([0x1p-3])]; + tensor attn_weights_25_cast_fp16 = mul(x = var_2055_cast_fp16, y = var_2049_to_fp16)[name = tensor("attn_weights_25_cast_fp16")]; + tensor var_2057 = logical_not(x = attn_mask_9)[name = tensor("op_2057")]; + tensor var_2058_to_fp16 = const()[name = tensor("op_2058_to_fp16"), val = tensor(-inf)]; + tensor attn_weights_27_cast_fp16 = select(a = var_2058_to_fp16, b = attn_weights_25_cast_fp16, cond = var_2057)[name = tensor("attn_weights_27_cast_fp16")]; + tensor var_2060 = const()[name = tensor("op_2060"), val = tensor(-1)]; + tensor attn_weights_29_cast_fp16 = softmax(axis = var_2060, x = attn_weights_27_cast_fp16)[name = tensor("attn_weights_29_cast_fp16")]; + tensor attn_output_9_transpose_x_0 = const()[name = tensor("attn_output_9_transpose_x_0"), val = tensor(false)]; + tensor attn_output_9_transpose_y_0 = const()[name = tensor("attn_output_9_transpose_y_0"), val = tensor(false)]; + tensor values_29_cast_fp16 = transpose(perm = var_1996, x = values_27_cast_fp16)[name = tensor("transpose_28")]; + tensor attn_output_9_cast_fp16 = matmul(transpose_x = attn_output_9_transpose_x_0, transpose_y = attn_output_9_transpose_y_0, x = attn_weights_29_cast_fp16, y = values_29_cast_fp16)[name = tensor("attn_output_9_cast_fp16")]; + tensor var_2068 = const()[name = tensor("op_2068"), val = tensor([0, 2, 1, 3])]; + tensor var_2073 = const()[name = tensor("op_2073"), val = tensor([1, 256, 1024])]; + tensor var_2069_cast_fp16 = transpose(perm = var_2068, x = attn_output_9_cast_fp16)[name = tensor("transpose_25")]; + tensor input_53_cast_fp16 = reshape(shape = var_2073, x = var_2069_cast_fp16)[name = tensor("input_53_cast_fp16")]; + tensor attn4_out_proj_weight_to_fp16 = const()[name = tensor("attn4_out_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(108062592)))]; + tensor linear_17_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = attn4_out_proj_weight_to_fp16, x = input_53_cast_fp16)[name = tensor("linear_17_cast_fp16")]; + tensor var_2082_cast_fp16 = add(x = position4_to_fp16, y = valid_len_to_fp16)[name = tensor("op_2082_cast_fp16")]; + tensor var_2082_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_2082_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor input_55_cast_fp16 = add(x = input_49_cast_fp16, y = linear_17_cast_fp16)[name = tensor("input_55_cast_fp16")]; + tensor input_57_axes_0 = const()[name = tensor("input_57_axes_0"), val = tensor([-1])]; + tensor norm4_2_weight_to_fp16 = const()[name = tensor("norm4_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(110159808)))]; + tensor norm4_2_bias_to_fp16 = const()[name = tensor("norm4_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(110161920)))]; + tensor var_2086_to_fp16 = const()[name = tensor("op_2086_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_57_cast_fp16 = layer_norm(axes = input_57_axes_0, beta = norm4_2_bias_to_fp16, epsilon = var_2086_to_fp16, gamma = norm4_2_weight_to_fp16, x = input_55_cast_fp16)[name = tensor("input_57_cast_fp16")]; + tensor linear4_1_weight_to_fp16 = const()[name = tensor("linear4_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(110164032)))]; + tensor linear_18_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = linear4_1_weight_to_fp16, x = input_57_cast_fp16)[name = tensor("linear_18_cast_fp16")]; + tensor input_59_mode_0 = const()[name = tensor("input_59_mode_0"), val = tensor("EXACT")]; + tensor input_59_cast_fp16 = gelu(mode = input_59_mode_0, x = linear_18_cast_fp16)[name = tensor("input_59_cast_fp16")]; + tensor linear4_2_weight_to_fp16 = const()[name = tensor("linear4_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(118552704)))]; + tensor linear_19_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = linear4_2_weight_to_fp16, x = input_59_cast_fp16)[name = tensor("linear_19_cast_fp16")]; + tensor input_61_cast_fp16 = add(x = input_55_cast_fp16, y = linear_19_cast_fp16)[name = tensor("input_61_cast_fp16")]; + tensor input_63_axes_0 = const()[name = tensor("input_63_axes_0"), val = tensor([-1])]; + tensor norm5_1_weight_to_fp16 = const()[name = tensor("norm5_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(126941376)))]; + tensor norm5_1_bias_to_fp16 = const()[name = tensor("norm5_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(126943488)))]; + tensor var_2103_to_fp16 = const()[name = tensor("op_2103_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_63_cast_fp16 = layer_norm(axes = input_63_axes_0, beta = norm5_1_bias_to_fp16, epsilon = var_2103_to_fp16, gamma = norm5_1_weight_to_fp16, x = input_61_cast_fp16)[name = tensor("input_63_cast_fp16")]; + tensor attn5_in_proj_weight_to_fp16 = const()[name = tensor("attn5_in_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(126945600)))]; + tensor linear_20_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = attn5_in_proj_weight_to_fp16, x = input_63_cast_fp16)[name = tensor("linear_20_cast_fp16")]; + tensor var_2117 = const()[name = tensor("op_2117"), val = tensor([1, 256, 3, 16, 64])]; + tensor qkv_cast_fp16 = reshape(shape = var_2117, x = linear_20_cast_fp16)[name = tensor("qkv_cast_fp16")]; + tensor k_21_begin_0 = const()[name = tensor("k_21_begin_0"), val = tensor([0, 0, 1, 0, 0])]; + tensor k_21_end_0 = const()[name = tensor("k_21_end_0"), val = tensor([1, 256, 2, 16, 64])]; + tensor k_21_end_mask_0 = const()[name = tensor("k_21_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor k_21_squeeze_mask_0 = const()[name = tensor("k_21_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor k_21_cast_fp16 = slice_by_index(begin = k_21_begin_0, end = k_21_end_0, end_mask = k_21_end_mask_0, squeeze_mask = k_21_squeeze_mask_0, x = qkv_cast_fp16)[name = tensor("k_21_cast_fp16")]; + tensor v_begin_0 = const()[name = tensor("v_begin_0"), val = tensor([0, 0, 2, 0, 0])]; + tensor v_end_0 = const()[name = tensor("v_end_0"), val = tensor([1, 256, 3, 16, 64])]; + tensor v_end_mask_0 = const()[name = tensor("v_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor v_squeeze_mask_0 = const()[name = tensor("v_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor v_cast_fp16 = slice_by_index(begin = v_begin_0, end = v_end_0, end_mask = v_end_mask_0, squeeze_mask = v_squeeze_mask_0, x = qkv_cast_fp16)[name = tensor("v_cast_fp16")]; + tensor var_2162 = const()[name = tensor("op_2162"), val = tensor([1, 1])]; + tensor position5_to_fp16_dtype_0 = const()[name = tensor("position5_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor position5_to_fp16 = cast(dtype = position5_to_fp16_dtype_0, x = position5)[name = tensor("cast_132")]; + tensor var_2163_cast_fp16 = reshape(shape = var_2162, x = position5_to_fp16)[name = tensor("op_2163_cast_fp16")]; + tensor ts_33_cast_fp16 = add(x = ts_1_promoted_to_fp16, y = var_2163_cast_fp16)[name = tensor("ts_33_cast_fp16")]; + tensor var_2170 = const()[name = tensor("op_2170"), val = tensor([1, 256, 1, 1])]; + tensor ts_cast_fp16 = reshape(shape = var_2170, x = ts_33_cast_fp16)[name = tensor("ts_cast_fp16")]; + tensor var_2177 = const()[name = tensor("op_2177"), val = tensor([1, 256, 16, 32, 2])]; + tensor k_complex_cast_fp16 = reshape(shape = var_2177, x = k_21_cast_fp16)[name = tensor("k_complex_cast_fp16")]; + tensor var_2181_begin_0 = const()[name = tensor("op_2181_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_2181_end_0 = const()[name = tensor("op_2181_end_0"), val = tensor([1, 256, 16, 32, 1])]; + tensor var_2181_end_mask_0 = const()[name = tensor("op_2181_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_2181_squeeze_mask_0 = const()[name = tensor("op_2181_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_2181_cast_fp16 = slice_by_index(begin = var_2181_begin_0, end = var_2181_end_0, end_mask = var_2181_end_mask_0, squeeze_mask = var_2181_squeeze_mask_0, x = k_complex_cast_fp16)[name = tensor("op_2181_cast_fp16")]; + tensor var_2189_begin_0 = const()[name = tensor("op_2189_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_2189_end_0 = const()[name = tensor("op_2189_end_0"), val = tensor([1, 256, 16, 32, 2])]; + tensor var_2189_end_mask_0 = const()[name = tensor("op_2189_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_2189_squeeze_mask_0 = const()[name = tensor("op_2189_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_2189_cast_fp16 = slice_by_index(begin = var_2189_begin_0, end = var_2189_end_0, end_mask = var_2189_end_mask_0, squeeze_mask = var_2189_squeeze_mask_0, x = k_complex_cast_fp16)[name = tensor("op_2189_cast_fp16")]; + tensor freqs_to_fp16 = const()[name = tensor("freqs_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(133237120)))]; + tensor var_2195_cast_fp16 = mul(x = freqs_to_fp16, y = ts_cast_fp16)[name = tensor("op_2195_cast_fp16")]; + tensor rotr_cast_fp16 = cos(x = var_2195_cast_fp16)[name = tensor("rotr_cast_fp16")]; + tensor roti_cast_fp16 = sin(x = var_2195_cast_fp16)[name = tensor("roti_cast_fp16")]; + tensor var_2199_cast_fp16 = mul(x = var_2181_cast_fp16, y = rotr_cast_fp16)[name = tensor("op_2199_cast_fp16")]; + tensor var_2200_cast_fp16 = mul(x = var_2189_cast_fp16, y = roti_cast_fp16)[name = tensor("op_2200_cast_fp16")]; + tensor kor_21_cast_fp16 = sub(x = var_2199_cast_fp16, y = var_2200_cast_fp16)[name = tensor("kor_21_cast_fp16")]; + tensor var_2203_cast_fp16 = mul(x = var_2181_cast_fp16, y = roti_cast_fp16)[name = tensor("op_2203_cast_fp16")]; + tensor var_2204_cast_fp16 = mul(x = var_2189_cast_fp16, y = rotr_cast_fp16)[name = tensor("op_2204_cast_fp16")]; + tensor koi_21_cast_fp16 = add(x = var_2203_cast_fp16, y = var_2204_cast_fp16)[name = tensor("koi_21_cast_fp16")]; + tensor ko_axis_0 = const()[name = tensor("ko_axis_0"), val = tensor(-1)]; + tensor ko_cast_fp16 = stack(axis = ko_axis_0, values = (kor_21_cast_fp16, koi_21_cast_fp16))[name = tensor("ko_cast_fp16")]; + tensor var_2224 = const()[name = tensor("op_2224"), val = tensor([1, 256, 16, 64])]; + tensor k_cast_fp16 = reshape(shape = var_2224, x = ko_cast_fp16)[name = tensor("k_cast_fp16")]; + tensor write_indices_float_31_cast_fp16 = add(x = var_2163_cast_fp16, y = var_249_promoted_to_fp16)[name = tensor("write_indices_float_31_cast_fp16")]; + tensor _inversed_2245_y_0_to_fp16 = const()[name = tensor("_inversed_2245_y_0_to_fp16"), val = tensor(0x1p-9)]; + tensor _inversed_2245_cast_fp16 = mul(x = write_indices_float_31_cast_fp16, y = _inversed_2245_y_0_to_fp16)[name = tensor("_inversed_2245_cast_fp16")]; + tensor var_2246_cast_fp16 = floor(x = _inversed_2245_cast_fp16)[name = tensor("op_2246_cast_fp16")]; + tensor var_2247_to_fp16 = const()[name = tensor("op_2247_to_fp16"), val = tensor(0x1p+9)]; + tensor var_2248_cast_fp16 = mul(x = var_2246_cast_fp16, y = var_2247_to_fp16)[name = tensor("op_2248_cast_fp16")]; + tensor write_indices_float_33_cast_fp16 = sub(x = write_indices_float_31_cast_fp16, y = var_2248_cast_fp16)[name = tensor("write_indices_float_33_cast_fp16")]; + tensor var_2266_value_0_to_fp16 = const()[name = tensor("op_2266_value_0_to_fp16"), val = tensor(0x1.ffp+8)]; + tensor var_2266_cast_fp16 = fill_like(ref_tensor = write_indices_float_33_cast_fp16, value = var_2266_value_0_to_fp16)[name = tensor("op_2266_cast_fp16")]; + tensor write_indices_float_cast_fp16 = select(a = write_indices_float_33_cast_fp16, b = var_2266_cast_fp16, cond = token_valid_1_cast_fp16)[name = tensor("write_indices_float_cast_fp16")]; + tensor var_2272_dtype_0 = const()[name = tensor("op_2272_dtype_0"), val = tensor("int32")]; + tensor var_2277 = const()[name = tensor("op_2277"), val = tensor([1, 256, 1, 1])]; + tensor write_indices_float_cast_fp16_to_int32 = cast(dtype = var_2272_dtype_0, x = write_indices_float_cast_fp16)[name = tensor("cast_131")]; + tensor var_2278 = reshape(shape = var_2277, x = write_indices_float_cast_fp16_to_int32)[name = tensor("op_2278")]; + tensor write_indices_reps_0 = const()[name = tensor("write_indices_reps_0"), val = tensor([1, 1, 16, 64])]; + tensor write_indices = tile(reps = write_indices_reps_0, x = var_2278)[name = tensor("write_indices")]; + tensor var_2288_begin_0 = const()[name = tensor("op_2288_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_2288_end_0 = const()[name = tensor("op_2288_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor var_2288_end_mask_0 = const()[name = tensor("op_2288_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_2288_squeeze_mask_0 = const()[name = tensor("op_2288_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor cache5_to_fp16_dtype_0 = const()[name = tensor("cache5_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor cache5_to_fp16 = cast(dtype = cache5_to_fp16_dtype_0, x = cache5)[name = tensor("cast_130")]; + tensor var_2288_cast_fp16 = slice_by_index(begin = var_2288_begin_0, end = var_2288_end_0, end_mask = var_2288_end_mask_0, squeeze_mask = var_2288_squeeze_mask_0, x = cache5_to_fp16)[name = tensor("op_2288_cast_fp16")]; + tensor var_2290_axis_0 = const()[name = tensor("op_2290_axis_0"), val = tensor(1)]; + tensor var_2290_mode_0 = const()[name = tensor("op_2290_mode_0"), val = tensor("update")]; + tensor var_2290_validate_indices_0 = const()[name = tensor("op_2290_validate_indices_0"), val = tensor(false)]; + tensor var_2290_cast_fp16 = scatter_along_axis(axis = var_2290_axis_0, data = var_2288_cast_fp16, indices = write_indices, mode = var_2290_mode_0, updates = k_cast_fp16, validate_indices = var_2290_validate_indices_0)[name = tensor("op_2290_cast_fp16")]; + tensor concat_42 = const()[name = tensor("concat_42"), val = tensor([0, 0, 0, 0, 0])]; + tensor concat_43 = const()[name = tensor("concat_43"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_internal_tensor_assign_1_stride_0 = const()[name = tensor("new_cache_internal_tensor_assign_1_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_internal_tensor_assign_1_begin_mask_0 = const()[name = tensor("new_cache_internal_tensor_assign_1_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_internal_tensor_assign_1_end_mask_0 = const()[name = tensor("new_cache_internal_tensor_assign_1_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_internal_tensor_assign_1_squeeze_mask_0 = const()[name = tensor("new_cache_internal_tensor_assign_1_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_20 = const()[name = tensor("shape_20"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_10 = const()[name = tensor("reduce_prod_10"), val = tensor(1048576)]; + tensor range_1d_10_start_0 = const()[name = tensor("range_1d_10_start_0"), val = tensor(0)]; + tensor range_1d_10_step_0 = const()[name = tensor("range_1d_10_step_0"), val = tensor(1)]; + tensor range_1d_10 = range_1d(end = reduce_prod_10, start = range_1d_10_start_0, step = range_1d_10_step_0)[name = tensor("range_1d_10")]; + tensor reshape_50 = reshape(shape = shape_20, x = range_1d_10)[name = tensor("reshape_50")]; + tensor slice_by_index_10 = slice_by_index(begin = concat_42, begin_mask = new_cache_internal_tensor_assign_1_begin_mask_0, end = concat_43, end_mask = new_cache_internal_tensor_assign_1_end_mask_0, squeeze_mask = new_cache_internal_tensor_assign_1_squeeze_mask_0, stride = new_cache_internal_tensor_assign_1_stride_0, x = reshape_50)[name = tensor("slice_by_index_10")]; + tensor reshape_51_shape_0 = const()[name = tensor("reshape_51_shape_0"), val = tensor([-1])]; + tensor reshape_51 = reshape(shape = reshape_51_shape_0, x = slice_by_index_10)[name = tensor("reshape_51")]; + tensor reshape_52_shape_0 = const()[name = tensor("reshape_52_shape_0"), val = tensor([-1])]; + tensor reshape_52_cast_fp16 = reshape(shape = reshape_52_shape_0, x = var_2290_cast_fp16)[name = tensor("reshape_52_cast_fp16")]; + tensor reshape_53_shape_0 = const()[name = tensor("reshape_53_shape_0"), val = tensor([-1])]; + tensor reshape_53_cast_fp16 = reshape(shape = reshape_53_shape_0, x = cache5_to_fp16)[name = tensor("reshape_53_cast_fp16")]; + tensor scatter_10_mode_0 = const()[name = tensor("scatter_10_mode_0"), val = tensor("update")]; + tensor scatter_10_axis_0 = const()[name = tensor("scatter_10_axis_0"), val = tensor(0)]; + tensor scatter_10_validate_indices_0 = const()[name = tensor("scatter_10_validate_indices_0"), val = tensor(false)]; + tensor scatter_10_cast_fp16 = scatter(axis = scatter_10_axis_0, data = reshape_53_cast_fp16, indices = reshape_51, mode = scatter_10_mode_0, updates = reshape_52_cast_fp16, validate_indices = scatter_10_validate_indices_0)[name = tensor("scatter_10_cast_fp16")]; + tensor reshape_54_cast_fp16 = reshape(shape = shape_20, x = scatter_10_cast_fp16)[name = tensor("reshape_54_cast_fp16")]; + tensor var_2298_begin_0 = const()[name = tensor("op_2298_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor var_2298_end_0 = const()[name = tensor("op_2298_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor var_2298_end_mask_0 = const()[name = tensor("op_2298_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_2298_squeeze_mask_0 = const()[name = tensor("op_2298_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor var_2298_cast_fp16 = slice_by_index(begin = var_2298_begin_0, end = var_2298_end_0, end_mask = var_2298_end_mask_0, squeeze_mask = var_2298_squeeze_mask_0, x = reshape_54_cast_fp16)[name = tensor("op_2298_cast_fp16")]; + tensor var_2300_axis_0 = const()[name = tensor("op_2300_axis_0"), val = tensor(1)]; + tensor var_2300_mode_0 = const()[name = tensor("op_2300_mode_0"), val = tensor("update")]; + tensor var_2300_validate_indices_0 = const()[name = tensor("op_2300_validate_indices_0"), val = tensor(false)]; + tensor var_2300_cast_fp16 = scatter_along_axis(axis = var_2300_axis_0, data = var_2298_cast_fp16, indices = write_indices, mode = var_2300_mode_0, updates = v_cast_fp16, validate_indices = var_2300_validate_indices_0)[name = tensor("op_2300_cast_fp16")]; + tensor concat_44 = const()[name = tensor("concat_44"), val = tensor([1, 0, 0, 0, 0])]; + tensor concat_45 = const()[name = tensor("concat_45"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_internal_tensor_assign_2_stride_0 = const()[name = tensor("new_cache_internal_tensor_assign_2_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_internal_tensor_assign_2_begin_mask_0 = const()[name = tensor("new_cache_internal_tensor_assign_2_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_internal_tensor_assign_2_end_mask_0 = const()[name = tensor("new_cache_internal_tensor_assign_2_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_internal_tensor_assign_2_squeeze_mask_0 = const()[name = tensor("new_cache_internal_tensor_assign_2_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_21 = const()[name = tensor("shape_21"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_11 = const()[name = tensor("reduce_prod_11"), val = tensor(1048576)]; + tensor range_1d_11_start_0 = const()[name = tensor("range_1d_11_start_0"), val = tensor(0)]; + tensor range_1d_11_step_0 = const()[name = tensor("range_1d_11_step_0"), val = tensor(1)]; + tensor range_1d_11 = range_1d(end = reduce_prod_11, start = range_1d_11_start_0, step = range_1d_11_step_0)[name = tensor("range_1d_11")]; + tensor reshape_55 = reshape(shape = shape_21, x = range_1d_11)[name = tensor("reshape_55")]; + tensor slice_by_index_11 = slice_by_index(begin = concat_44, begin_mask = new_cache_internal_tensor_assign_2_begin_mask_0, end = concat_45, end_mask = new_cache_internal_tensor_assign_2_end_mask_0, squeeze_mask = new_cache_internal_tensor_assign_2_squeeze_mask_0, stride = new_cache_internal_tensor_assign_2_stride_0, x = reshape_55)[name = tensor("slice_by_index_11")]; + tensor reshape_56_shape_0 = const()[name = tensor("reshape_56_shape_0"), val = tensor([-1])]; + tensor reshape_56 = reshape(shape = reshape_56_shape_0, x = slice_by_index_11)[name = tensor("reshape_56")]; + tensor reshape_57_shape_0 = const()[name = tensor("reshape_57_shape_0"), val = tensor([-1])]; + tensor reshape_57_cast_fp16 = reshape(shape = reshape_57_shape_0, x = var_2300_cast_fp16)[name = tensor("reshape_57_cast_fp16")]; + tensor reshape_58_shape_0 = const()[name = tensor("reshape_58_shape_0"), val = tensor([-1])]; + tensor reshape_58_cast_fp16 = reshape(shape = reshape_58_shape_0, x = reshape_54_cast_fp16)[name = tensor("reshape_58_cast_fp16")]; + tensor scatter_11_mode_0 = const()[name = tensor("scatter_11_mode_0"), val = tensor("update")]; + tensor scatter_11_axis_0 = const()[name = tensor("scatter_11_axis_0"), val = tensor(0)]; + tensor scatter_11_validate_indices_0 = const()[name = tensor("scatter_11_validate_indices_0"), val = tensor(false)]; + tensor scatter_11_cast_fp16 = scatter(axis = scatter_11_axis_0, data = reshape_58_cast_fp16, indices = reshape_56, mode = scatter_11_mode_0, updates = reshape_57_cast_fp16, validate_indices = scatter_11_validate_indices_0)[name = tensor("scatter_11_cast_fp16")]; + tensor reshape_59_cast_fp16 = reshape(shape = shape_21, x = scatter_11_cast_fp16)[name = tensor("reshape_59_cast_fp16")]; + tensor reshape_59_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("reshape_59_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor var_2310_cast_fp16 = add(x = position5_to_fp16, y = valid_len_to_fp16)[name = tensor("op_2310_cast_fp16")]; + tensor var_2310_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_2310_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor var_2310 = cast(dtype = var_2310_cast_fp16_to_fp32_dtype_0, x = var_2310_cast_fp16)[name = tensor("cast_128")]; + tensor new_cache_internal_tensor_assign_2 = cast(dtype = reshape_59_cast_fp16_to_fp32_dtype_0, x = reshape_59_cast_fp16)[name = tensor("cast_129")]; + tensor var_2082 = cast(dtype = var_2082_cast_fp16_to_fp32_dtype_0, x = var_2082_cast_fp16)[name = tensor("cast_133")]; + tensor new_cache_9_internal_tensor_assign_2 = cast(dtype = reshape_49_cast_fp16_to_fp32_dtype_0, x = reshape_49_cast_fp16)[name = tensor("cast_134")]; + tensor var_1671 = cast(dtype = var_1671_cast_fp16_to_fp32_dtype_0, x = var_1671_cast_fp16)[name = tensor("cast_138")]; + tensor new_cache_7_internal_tensor_assign_2 = cast(dtype = reshape_39_cast_fp16_to_fp32_dtype_0, x = reshape_39_cast_fp16)[name = tensor("cast_139")]; + tensor var_1260 = cast(dtype = var_1260_cast_fp16_to_fp32_dtype_0, x = var_1260_cast_fp16)[name = tensor("cast_143")]; + tensor new_cache_5_internal_tensor_assign_2 = cast(dtype = reshape_29_cast_fp16_to_fp32_dtype_0, x = reshape_29_cast_fp16)[name = tensor("cast_144")]; + tensor var_849 = cast(dtype = var_849_cast_fp16_to_fp32_dtype_0, x = var_849_cast_fp16)[name = tensor("cast_148")]; + tensor new_cache_3_internal_tensor_assign_2 = cast(dtype = reshape_19_cast_fp16_to_fp32_dtype_0, x = reshape_19_cast_fp16)[name = tensor("cast_149")]; + tensor var_438 = cast(dtype = var_438_cast_fp16_to_fp32_dtype_0, x = var_438_cast_fp16)[name = tensor("cast_153")]; + tensor new_cache_1_internal_tensor_assign_2 = cast(dtype = reshape_9_cast_fp16_to_fp32_dtype_0, x = reshape_9_cast_fp16)[name = tensor("cast_154")]; + } -> (new_cache_1_internal_tensor_assign_2, var_438, new_cache_3_internal_tensor_assign_2, var_849, new_cache_5_internal_tensor_assign_2, var_1260, new_cache_7_internal_tensor_assign_2, var_1671, new_cache_9_internal_tensor_assign_2, var_2082, new_cache_internal_tensor_assign_2, var_2310); +} \ No newline at end of file diff --git a/v2.1/spanish/cond_prefill.mlmodelc/weights/weight.bin b/v2.1/spanish/cond_prefill.mlmodelc/weights/weight.bin new file mode 100644 index 0000000000000000000000000000000000000000..d64a8dbf6424d8b862b22b22cbc8f435b771b971 --- /dev/null +++ b/v2.1/spanish/cond_prefill.mlmodelc/weights/weight.bin @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:f5c825c894d648bba6033981cdd1a98d24a7c89a95f58a7d3ead3c05041b9a3e +size 133237248 diff --git a/v2.1/spanish/cond_prefill.mlpackage/Data/com.apple.CoreML/model.mlmodel b/v2.1/spanish/cond_prefill.mlpackage/Data/com.apple.CoreML/model.mlmodel new file mode 100644 index 0000000000000000000000000000000000000000..32ad2352aa08613ae391cab7bfcd923d799a29f6 --- /dev/null +++ b/v2.1/spanish/cond_prefill.mlpackage/Data/com.apple.CoreML/model.mlmodel @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:7fdc4ac177e8d2fa88576982ad0f2d90f3ac7f3b2e48316cc5e58bbd40e0b8d8 +size 200480 diff --git a/v2.1/spanish/cond_prefill.mlpackage/Data/com.apple.CoreML/weights/weight.bin b/v2.1/spanish/cond_prefill.mlpackage/Data/com.apple.CoreML/weights/weight.bin new file mode 100644 index 0000000000000000000000000000000000000000..d64a8dbf6424d8b862b22b22cbc8f435b771b971 --- /dev/null +++ b/v2.1/spanish/cond_prefill.mlpackage/Data/com.apple.CoreML/weights/weight.bin @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:f5c825c894d648bba6033981cdd1a98d24a7c89a95f58a7d3ead3c05041b9a3e +size 133237248 diff --git a/v2.1/spanish/cond_prefill.mlpackage/Manifest.json b/v2.1/spanish/cond_prefill.mlpackage/Manifest.json new file mode 100644 index 0000000000000000000000000000000000000000..47bb20d7b93ce59bf9bdc3440590f9f7590ba2fb --- /dev/null +++ b/v2.1/spanish/cond_prefill.mlpackage/Manifest.json @@ -0,0 +1,18 @@ +{ + "fileFormatVersion": "1.0.0", + "itemInfoEntries": { + "17E64B67-C9C3-43FF-9D4F-B84E1C49E4E3": { + "author": "com.apple.CoreML", + "description": "CoreML Model Specification", + "name": "model.mlmodel", + "path": "com.apple.CoreML/model.mlmodel" + }, + "75CFDFFF-1B95-4C7B-A4D7-417602B23781": { + "author": "com.apple.CoreML", + "description": "CoreML Model Weights", + "name": "weights", + "path": "com.apple.CoreML/weights" + } + }, + "rootModelIdentifier": "17E64B67-C9C3-43FF-9D4F-B84E1C49E4E3" +} diff --git a/v2.1/spanish/constants/bos_emb.npy b/v2.1/spanish/constants/bos_emb.npy new file mode 100644 index 0000000000000000000000000000000000000000..08acc13b8528bed6d0cece437e8b6681a1ef1c5c --- /dev/null +++ b/v2.1/spanish/constants/bos_emb.npy @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:ee4fe6fc778497ffd84e4912da12300e4d8e94e65f53d506b30fde26b940abac +size 256 diff --git a/v2.1/spanish/constants/emb_mean.npy b/v2.1/spanish/constants/emb_mean.npy new file mode 100644 index 0000000000000000000000000000000000000000..29bee3dc5f686c58698b896bcd48f6957f79e8d7 --- /dev/null +++ b/v2.1/spanish/constants/emb_mean.npy @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:57333945fc74572a2504025b3d48ad156217828144fba9407aabb962d58028b3 +size 256 diff --git a/v2.1/spanish/constants/emb_std.npy b/v2.1/spanish/constants/emb_std.npy new file mode 100644 index 0000000000000000000000000000000000000000..674d07376296efb08e41a46352a97a9ed4349760 --- /dev/null +++ b/v2.1/spanish/constants/emb_std.npy @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:9e8ac8f4dcd6e6a3aa2d1c1fc9270b248525856da470d7f1a9b86c966dbebe69 +size 256 diff --git a/v2.1/spanish/constants/mimi_init_state.npz b/v2.1/spanish/constants/mimi_init_state.npz new file mode 100644 index 0000000000000000000000000000000000000000..3b52f383a86ef0439f9993506e4529c4c1234828 --- /dev/null +++ b/v2.1/spanish/constants/mimi_init_state.npz @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:515588e10e9c14a6b009f1cfda9fcd592542b1640ad856d5fc69e50024702a39 +size 1083108 diff --git a/v2.1/spanish/constants/quantizer_weight.npy b/v2.1/spanish/constants/quantizer_weight.npy new file mode 100644 index 0000000000000000000000000000000000000000..74fd84280b4dee8d77ab6146e9c5eec48763ba30 --- /dev/null +++ b/v2.1/spanish/constants/quantizer_weight.npy @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:b068606d26574a8a1513591f9d228b0b44b71d9c90f6c1f032f6ed6cae9f065c +size 65664 diff --git a/v2.1/spanish/constants/text_embed_table.npy b/v2.1/spanish/constants/text_embed_table.npy new file mode 100644 index 0000000000000000000000000000000000000000..c4c65eb9ed4058ff4585c59201b2ca354ed27388 --- /dev/null +++ b/v2.1/spanish/constants/text_embed_table.npy @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:98f8d62bf90a2a86c2f290f2fbdcad103594e63d329de18649dc04e98ec5affd +size 16388224 diff --git a/v2.1/spanish/constants_bin/alba.safetensors b/v2.1/spanish/constants_bin/alba.safetensors new file mode 100644 index 0000000000000000000000000000000000000000..7847720d62825a3527a0ff5a55a9e87ba3b643bf --- /dev/null +++ b/v2.1/spanish/constants_bin/alba.safetensors @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:694be334ed6dda508b125e22a6ae480036d1c99923cccd92a643a348867f3235 +size 6194424 diff --git a/v2.1/spanish/constants_bin/anna.safetensors b/v2.1/spanish/constants_bin/anna.safetensors new file mode 100644 index 0000000000000000000000000000000000000000..779799e1ac171bd9cdcbce91dd3832333e3788d1 --- /dev/null +++ b/v2.1/spanish/constants_bin/anna.safetensors @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:fe964cdad79e8c2c0d2ca30bc9dc8479c6f38b45b77512975a6e272109367337 +size 7816440 diff --git a/v2.1/spanish/constants_bin/azelma.safetensors b/v2.1/spanish/constants_bin/azelma.safetensors new file mode 100644 index 0000000000000000000000000000000000000000..edb7f410003ecfced3bc75278d17c68e3db33cce --- /dev/null +++ b/v2.1/spanish/constants_bin/azelma.safetensors @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:7fe8e3e2bf5258c73c765f5553478b95386e760ccecbb46fe24c2632b5fd7458 +size 7963896 diff --git a/v2.1/spanish/constants_bin/bill_boerst.safetensors b/v2.1/spanish/constants_bin/bill_boerst.safetensors new file mode 100644 index 0000000000000000000000000000000000000000..2ace9a4b2444d7d973f7275c40cff7adbdde18b1 --- /dev/null +++ b/v2.1/spanish/constants_bin/bill_boerst.safetensors @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:5b95f21e267bd88357fbb4fe3cba173d9e4ba39ade8f386bb004a293ab5b470f +size 6735096 diff --git a/v2.1/spanish/constants_bin/bos_before_voice.bin b/v2.1/spanish/constants_bin/bos_before_voice.bin new file mode 100644 index 0000000000000000000000000000000000000000..515a2c7f6ca6655e46172deb49bdb04f96df35c2 --- /dev/null +++ b/v2.1/spanish/constants_bin/bos_before_voice.bin @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:8e66ab2090702caa0fe1688c47e4e9f3d7d6f7e25c4b09a067f6aba82c5d1c6b +size 4096 diff --git a/v2.1/spanish/constants_bin/bos_emb.bin b/v2.1/spanish/constants_bin/bos_emb.bin new file mode 100644 index 0000000000000000000000000000000000000000..d19f64d943c43212beb80dc86eb4fba54c42e950 --- /dev/null +++ b/v2.1/spanish/constants_bin/bos_emb.bin @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:bf1e187eb178c8f2c24845149958ac50512c0f07cea217335a248f36571d6ab8 +size 128 diff --git a/v2.1/spanish/constants_bin/caro_davy.safetensors b/v2.1/spanish/constants_bin/caro_davy.safetensors new file mode 100644 index 0000000000000000000000000000000000000000..aa38f14710da803555a846930a293dacad2867e0 --- /dev/null +++ b/v2.1/spanish/constants_bin/caro_davy.safetensors @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:d07fb793439eca0e2d650e5c8b49f21b829351ca76e9525f9ab39c136ceaf079 +size 5260536 diff --git a/v2.1/spanish/constants_bin/charles.safetensors b/v2.1/spanish/constants_bin/charles.safetensors new file mode 100644 index 0000000000000000000000000000000000000000..dd89c9350ab36ded6370ee1dd25918c992d745cb --- /dev/null +++ b/v2.1/spanish/constants_bin/charles.safetensors @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:d620942926b0a30dcf002212221c89bb4f3b8b9b9bd2cc7d034c5e1c792a675f +size 6194424 diff --git a/v2.1/spanish/constants_bin/cosette.safetensors b/v2.1/spanish/constants_bin/cosette.safetensors new file mode 100644 index 0000000000000000000000000000000000000000..12cb20551b57397161fd61d7fbe476d901b58780 --- /dev/null +++ b/v2.1/spanish/constants_bin/cosette.safetensors @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:c4a00e5e1c1c953b15fb8100208948a4e38331e1b9230e2f6a0722260a301f41 +size 6194424 diff --git a/v2.1/spanish/constants_bin/eponine.safetensors b/v2.1/spanish/constants_bin/eponine.safetensors new file mode 100644 index 0000000000000000000000000000000000000000..b730b8e05e9e91b905d4e920c0effc01801af81f --- /dev/null +++ b/v2.1/spanish/constants_bin/eponine.safetensors @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:06e688d7b179125afc56009a0f9d263768daa2c9ac3a0c7d07d70e411c01d5a7 +size 6931704 diff --git a/v2.1/spanish/constants_bin/estelle.safetensors b/v2.1/spanish/constants_bin/estelle.safetensors new file mode 100644 index 0000000000000000000000000000000000000000..116292ad97a7f67932c4314d4af2d93fa8bbba07 --- /dev/null +++ b/v2.1/spanish/constants_bin/estelle.safetensors @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:2342bd79d85e80da2f39a9cb4b4a5c3db4dec79cbbecfdf01c9af7e9578ea187 +size 8258808 diff --git a/v2.1/spanish/constants_bin/eve.safetensors b/v2.1/spanish/constants_bin/eve.safetensors new file mode 100644 index 0000000000000000000000000000000000000000..e608d8eacca08a8af17608345a18978213010583 --- /dev/null +++ b/v2.1/spanish/constants_bin/eve.safetensors @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:3fdb6a2ddd06f339b015ec90aea2ea004c260fe463c9526e020249b307096a41 +size 6538488 diff --git a/v2.1/spanish/constants_bin/fantine.safetensors b/v2.1/spanish/constants_bin/fantine.safetensors new file mode 100644 index 0000000000000000000000000000000000000000..7cfbf3f797b816dfa7690c3b341c60c862c50eeb --- /dev/null +++ b/v2.1/spanish/constants_bin/fantine.safetensors @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:7941dd5a2666afc6db9dadfcd27ee0255df89b50227cbd9628a3c0f280ff6f91 +size 6538488 diff --git a/v2.1/spanish/constants_bin/george.safetensors b/v2.1/spanish/constants_bin/george.safetensors new file mode 100644 index 0000000000000000000000000000000000000000..69e45805f4e4dc2a5d286a3c5100ba657249c693 --- /dev/null +++ b/v2.1/spanish/constants_bin/george.safetensors @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:da140d63300c46c31b49ec26402a0b4efed5d773da8310cf4ccca115853cdc87 +size 6243576 diff --git a/v2.1/spanish/constants_bin/giovanni.safetensors b/v2.1/spanish/constants_bin/giovanni.safetensors new file mode 100644 index 0000000000000000000000000000000000000000..e03d45b8584a4dd3ff3235b7fe2c6b68b645d5f0 --- /dev/null +++ b/v2.1/spanish/constants_bin/giovanni.safetensors @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:aed540f9c8b2336f9ff6670fcab8d077f7bc7b076790da773a6038d7bb041cb4 +size 4621552 diff --git a/v2.1/spanish/constants_bin/jane.safetensors b/v2.1/spanish/constants_bin/jane.safetensors new file mode 100644 index 0000000000000000000000000000000000000000..30fc5a586a8cf8cab8f93bb248f7d964203b45ed --- /dev/null +++ b/v2.1/spanish/constants_bin/jane.safetensors @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:1bf9bfe3172412c44d0c7227e49eeb40b060f6833e2f6f365f0bbcde757afd09 +size 7374072 diff --git a/v2.1/spanish/constants_bin/javert.safetensors b/v2.1/spanish/constants_bin/javert.safetensors new file mode 100644 index 0000000000000000000000000000000000000000..db8e6e69684d84d431d796c38972ff6e523cc478 --- /dev/null +++ b/v2.1/spanish/constants_bin/javert.safetensors @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:8a1d914ade79e8e470de3bd0725ad8bce62081976e352929d9d958a8773eba56 +size 6194424 diff --git a/v2.1/spanish/constants_bin/jean.safetensors b/v2.1/spanish/constants_bin/jean.safetensors new file mode 100644 index 0000000000000000000000000000000000000000..adaf83b0e876afd5e9a8bdb430e31d4b53f601c5 --- /dev/null +++ b/v2.1/spanish/constants_bin/jean.safetensors @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:d3b9f7e4436a048895a6f609fc19143a381ab89b44657311644c639285bdab22 +size 6194424 diff --git a/v2.1/spanish/constants_bin/juergen.safetensors b/v2.1/spanish/constants_bin/juergen.safetensors new file mode 100644 index 0000000000000000000000000000000000000000..26164f319f279ac1eb146a743a5498616876d91b --- /dev/null +++ b/v2.1/spanish/constants_bin/juergen.safetensors @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:acb46a675179db6baeac24189eeaaa7ca581cc56c14bd60b0086014a0aae6902 +size 6243576 diff --git a/v2.1/spanish/constants_bin/lola.safetensors b/v2.1/spanish/constants_bin/lola.safetensors new file mode 100644 index 0000000000000000000000000000000000000000..eaa91785eb5e6570b0b2915a2798f06e48019dc5 --- /dev/null +++ b/v2.1/spanish/constants_bin/lola.safetensors @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:5ec51484cea179be24582c3d3326398c5175a6d4ed7e1f2eb3491f8933e5a243 +size 5948664 diff --git a/v2.1/spanish/constants_bin/marius.safetensors b/v2.1/spanish/constants_bin/marius.safetensors new file mode 100644 index 0000000000000000000000000000000000000000..05b049bd2d2b15597d375833c1314c5f4f4e2bdd --- /dev/null +++ b/v2.1/spanish/constants_bin/marius.safetensors @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:56490a863e39b33fd1ab9c9e1fdaa87127f7b73cf73f259dc308c40015abb9f9 +size 6194424 diff --git a/v2.1/spanish/constants_bin/mary.safetensors b/v2.1/spanish/constants_bin/mary.safetensors new file mode 100644 index 0000000000000000000000000000000000000000..07a64bb2df1fbff5b2a7e4047849addce612b130 --- /dev/null +++ b/v2.1/spanish/constants_bin/mary.safetensors @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:9a2ccf97785a059592b3e56c6277ec1091b997e38452c69dd3f0e10ba1096603 +size 6194424 diff --git a/v2.1/spanish/constants_bin/michael.safetensors b/v2.1/spanish/constants_bin/michael.safetensors new file mode 100644 index 0000000000000000000000000000000000000000..002e5b69c13beff6d5453d963ebfd7f820fb5cc1 --- /dev/null +++ b/v2.1/spanish/constants_bin/michael.safetensors @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:ca3dfdaff6a2bf70ef9b2ab69bc66dce6d16fb06d4d7473be92610d3e43e9e15 +size 7275768 diff --git a/v2.1/spanish/constants_bin/paul.safetensors b/v2.1/spanish/constants_bin/paul.safetensors new file mode 100644 index 0000000000000000000000000000000000000000..7dca4aeff3b23de504342248296783262b5cbf65 --- /dev/null +++ b/v2.1/spanish/constants_bin/paul.safetensors @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:fcd677f0459352f7cdf62d7a1424a663bd49fc8921fb2b8e346ef9b45c92dd04 +size 6980856 diff --git a/v2.1/spanish/constants_bin/peter_yearsley.safetensors b/v2.1/spanish/constants_bin/peter_yearsley.safetensors new file mode 100644 index 0000000000000000000000000000000000000000..f92d89f86d37313b552a5e84b6c665ca08dbc05b --- /dev/null +++ b/v2.1/spanish/constants_bin/peter_yearsley.safetensors @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:088277ee6dbb17f00ac3519dd41c0c7dd5b977b4b946f173f6122f513d1921a7 +size 3736816 diff --git a/v2.1/spanish/constants_bin/rafael.safetensors b/v2.1/spanish/constants_bin/rafael.safetensors new file mode 100644 index 0000000000000000000000000000000000000000..bb877bf186eb7411c60c31eb725eda6fd94f3706 --- /dev/null +++ b/v2.1/spanish/constants_bin/rafael.safetensors @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:9a4d50795a8a0b30d925124572b687b9a88485bb1d15a92a403b61d9d1569195 +size 6194424 diff --git a/v2.1/spanish/constants_bin/stuart_bell.safetensors b/v2.1/spanish/constants_bin/stuart_bell.safetensors new file mode 100644 index 0000000000000000000000000000000000000000..b8e6a53e06b874322eb7d4c63da364a82504b5da --- /dev/null +++ b/v2.1/spanish/constants_bin/stuart_bell.safetensors @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:3182401767b42c6c6cf42c08047c939a903c9fb8895633e0b71bca65c6b476b0 +size 5260536 diff --git a/v2.1/spanish/constants_bin/text_embed_table.bin b/v2.1/spanish/constants_bin/text_embed_table.bin new file mode 100644 index 0000000000000000000000000000000000000000..75bee386db0b25b96f57e6a9c589a481395f6107 --- /dev/null +++ b/v2.1/spanish/constants_bin/text_embed_table.bin @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:3b89ed68229c0764070670a08589d23d49219b06c47fbb6419542772eae360bf +size 16388096 diff --git a/v2.1/spanish/constants_bin/tokenizer.model b/v2.1/spanish/constants_bin/tokenizer.model new file mode 100644 index 0000000000000000000000000000000000000000..1ac69e37ac41fb9e2a4491ade3064674a8f2682c --- /dev/null +++ b/v2.1/spanish/constants_bin/tokenizer.model @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:aac2b96478e3e3068712bd5c45dc196f84b7341ea72587c33a2094957f0c1b0c +size 60895 diff --git a/v2.1/spanish/constants_bin/vera.safetensors b/v2.1/spanish/constants_bin/vera.safetensors new file mode 100644 index 0000000000000000000000000000000000000000..7d597bec9e538e1ab9a0321a2159f404ac7de036 --- /dev/null +++ b/v2.1/spanish/constants_bin/vera.safetensors @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:65787d133a1a55bc26ce50bb9cc0a992e8343b3ab006e27c2e755239512cd727 +size 6735096 diff --git a/v2.1/spanish/flow_decoder_fused.mlmodelc/analytics/coremldata.bin b/v2.1/spanish/flow_decoder_fused.mlmodelc/analytics/coremldata.bin new file mode 100644 index 0000000000000000000000000000000000000000..88cd66c9b30490c48f8539db35631c02a88c5afa --- /dev/null +++ b/v2.1/spanish/flow_decoder_fused.mlmodelc/analytics/coremldata.bin @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:c4cdea05db76a9327db7a4cc21813378b30c9d3a162d895013ad9b148751b4a7 +size 243 diff --git a/v2.1/spanish/flow_decoder_fused.mlmodelc/coremldata.bin b/v2.1/spanish/flow_decoder_fused.mlmodelc/coremldata.bin new file mode 100644 index 0000000000000000000000000000000000000000..d7587d3924ac45ea6edf4b6c9a1864ab6df43768 --- /dev/null +++ b/v2.1/spanish/flow_decoder_fused.mlmodelc/coremldata.bin @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:0c3cb82046065fad467915b0e122f9c3b011186ce45ab9f44f43c9748b5c26c2 +size 413 diff --git a/v2.1/spanish/flow_decoder_fused.mlmodelc/model.mil b/v2.1/spanish/flow_decoder_fused.mlmodelc/model.mil new file mode 100644 index 0000000000000000000000000000000000000000..c64a1f34555e7bbf9c8d4502d29f9ef52dd6b0ec --- /dev/null +++ b/v2.1/spanish/flow_decoder_fused.mlmodelc/model.mil @@ -0,0 +1,1915 @@ +program(1.0) +[buildInfo = dict, tensor>({{"coremlc-component-MIL", "3520.4.1"}, {"coremlc-version", "3520.5.1"}, {"coremltools-component-torch", "2.12.0"}, {"coremltools-source-dialect", "TorchScript"}, {"coremltools-version", "9.0"}})] +{ + func main(tensor latent_init, tensor transformer_out) { + tensor latent_init_to_fp16_dtype_0 = const()[name = tensor("latent_init_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor flow_net_input_proj_weight_to_fp16 = const()[name = tensor("flow_net_input_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(64)))]; + tensor flow_net_input_proj_bias_to_fp16 = const()[name = tensor("flow_net_input_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(32896)))]; + tensor latent_init_to_fp16 = cast(dtype = latent_init_to_fp16_dtype_0, x = latent_init)[name = tensor("cast_146")]; + tensor linear_0_cast_fp16 = linear(bias = flow_net_input_proj_bias_to_fp16, weight = flow_net_input_proj_weight_to_fp16, x = latent_init_to_fp16)[name = tensor("linear_0_cast_fp16")]; + tensor input_3_to_fp16 = const()[name = tensor("input_3_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(33984)))]; + tensor input_5_cast_fp16 = silu(x = input_3_to_fp16)[name = tensor("input_5_cast_fp16")]; + tensor flow_net_time_embed_0_mlp_2_weight_to_fp16 = const()[name = tensor("flow_net_time_embed_0_mlp_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(35072)))]; + tensor flow_net_time_embed_0_mlp_2_bias_to_fp16 = const()[name = tensor("flow_net_time_embed_0_mlp_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(559424)))]; + tensor linear_2_cast_fp16 = linear(bias = flow_net_time_embed_0_mlp_2_bias_to_fp16, weight = flow_net_time_embed_0_mlp_2_weight_to_fp16, x = input_5_cast_fp16)[name = tensor("linear_2_cast_fp16")]; + tensor reduce_mean_0_axes_0 = const()[name = tensor("reduce_mean_0_axes_0"), val = tensor([-1])]; + tensor reduce_mean_0_keep_dims_0 = const()[name = tensor("reduce_mean_0_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_0_cast_fp16 = reduce_mean(axes = reduce_mean_0_axes_0, keep_dims = reduce_mean_0_keep_dims_0, x = linear_2_cast_fp16)[name = tensor("reduce_mean_0_cast_fp16")]; + tensor sub_0_cast_fp16 = sub(x = linear_2_cast_fp16, y = reduce_mean_0_cast_fp16)[name = tensor("sub_0_cast_fp16")]; + tensor square_0_cast_fp16 = square(x = sub_0_cast_fp16)[name = tensor("square_0_cast_fp16")]; + tensor reduce_mean_1_axes_0 = const()[name = tensor("reduce_mean_1_axes_0"), val = tensor([-1])]; + tensor reduce_mean_1_keep_dims_0 = const()[name = tensor("reduce_mean_1_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_1_cast_fp16 = reduce_mean(axes = reduce_mean_1_axes_0, keep_dims = reduce_mean_1_keep_dims_0, x = square_0_cast_fp16)[name = tensor("reduce_mean_1_cast_fp16")]; + tensor real_div_0_to_fp16 = const()[name = tensor("real_div_0_to_fp16"), val = tensor(0x1.008p+0)]; + tensor mul_0_cast_fp16 = mul(x = reduce_mean_1_cast_fp16, y = real_div_0_to_fp16)[name = tensor("mul_0_cast_fp16")]; + tensor var_70_to_fp16 = const()[name = tensor("op_70_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_1_cast_fp16 = add(x = mul_0_cast_fp16, y = var_70_to_fp16)[name = tensor("var_1_cast_fp16")]; + tensor var_73_epsilon_0 = const()[name = tensor("op_73_epsilon_0"), val = tensor(0x1.197998p-40)]; + tensor var_73_cast_fp16 = rsqrt(epsilon = var_73_epsilon_0, x = var_1_cast_fp16)[name = tensor("op_73_cast_fp16")]; + tensor const_3_to_fp16 = const()[name = tensor("const_3_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(560512)))]; + tensor var_74_cast_fp16 = mul(x = const_3_to_fp16, y = var_73_cast_fp16)[name = tensor("op_74_cast_fp16")]; + tensor var_75_cast_fp16 = mul(x = linear_2_cast_fp16, y = var_74_cast_fp16)[name = tensor("op_75_cast_fp16")]; + tensor input_9_to_fp16 = const()[name = tensor("input_9_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(561600)))]; + tensor input_11_cast_fp16 = silu(x = input_9_to_fp16)[name = tensor("input_11_cast_fp16")]; + tensor flow_net_time_embed_1_mlp_2_weight_to_fp16 = const()[name = tensor("flow_net_time_embed_1_mlp_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(562688)))]; + tensor flow_net_time_embed_1_mlp_2_bias_to_fp16 = const()[name = tensor("flow_net_time_embed_1_mlp_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1087040)))]; + tensor linear_4_cast_fp16 = linear(bias = flow_net_time_embed_1_mlp_2_bias_to_fp16, weight = flow_net_time_embed_1_mlp_2_weight_to_fp16, x = input_11_cast_fp16)[name = tensor("linear_4_cast_fp16")]; + tensor reduce_mean_2_axes_0 = const()[name = tensor("reduce_mean_2_axes_0"), val = tensor([-1])]; + tensor reduce_mean_2_keep_dims_0 = const()[name = tensor("reduce_mean_2_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_2_cast_fp16 = reduce_mean(axes = reduce_mean_2_axes_0, keep_dims = reduce_mean_2_keep_dims_0, x = linear_4_cast_fp16)[name = tensor("reduce_mean_2_cast_fp16")]; + tensor sub_2_cast_fp16 = sub(x = linear_4_cast_fp16, y = reduce_mean_2_cast_fp16)[name = tensor("sub_2_cast_fp16")]; + tensor square_1_cast_fp16 = square(x = sub_2_cast_fp16)[name = tensor("square_1_cast_fp16")]; + tensor reduce_mean_3_axes_0 = const()[name = tensor("reduce_mean_3_axes_0"), val = tensor([-1])]; + tensor reduce_mean_3_keep_dims_0 = const()[name = tensor("reduce_mean_3_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_3_cast_fp16 = reduce_mean(axes = reduce_mean_3_axes_0, keep_dims = reduce_mean_3_keep_dims_0, x = square_1_cast_fp16)[name = tensor("reduce_mean_3_cast_fp16")]; + tensor real_div_1_to_fp16 = const()[name = tensor("real_div_1_to_fp16"), val = tensor(0x1.008p+0)]; + tensor mul_1_cast_fp16 = mul(x = reduce_mean_3_cast_fp16, y = real_div_1_to_fp16)[name = tensor("mul_1_cast_fp16")]; + tensor var_110_to_fp16 = const()[name = tensor("op_110_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_3_cast_fp16 = add(x = mul_1_cast_fp16, y = var_110_to_fp16)[name = tensor("var_3_cast_fp16")]; + tensor var_113_epsilon_0 = const()[name = tensor("op_113_epsilon_0"), val = tensor(0x1.197998p-40)]; + tensor var_113_cast_fp16 = rsqrt(epsilon = var_113_epsilon_0, x = var_3_cast_fp16)[name = tensor("op_113_cast_fp16")]; + tensor const_5_to_fp16 = const()[name = tensor("const_5_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1088128)))]; + tensor var_114_cast_fp16 = mul(x = const_5_to_fp16, y = var_113_cast_fp16)[name = tensor("op_114_cast_fp16")]; + tensor var_115_cast_fp16 = mul(x = linear_4_cast_fp16, y = var_114_cast_fp16)[name = tensor("op_115_cast_fp16")]; + tensor var_127_cast_fp16 = add(x = var_75_cast_fp16, y = var_115_cast_fp16)[name = tensor("op_127_cast_fp16")]; + tensor _inversed_t_combined_1_y_0_to_fp16 = const()[name = tensor("_inversed_t_combined_1_y_0_to_fp16"), val = tensor(0x1p-1)]; + tensor _inversed_t_combined_1_cast_fp16 = mul(x = var_127_cast_fp16, y = _inversed_t_combined_1_y_0_to_fp16)[name = tensor("_inversed_t_combined_1_cast_fp16")]; + tensor transformer_out_to_fp16_dtype_0 = const()[name = tensor("transformer_out_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor flow_net_cond_embed_weight_to_fp16 = const()[name = tensor("flow_net_cond_embed_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1089216)))]; + tensor flow_net_cond_embed_bias_to_fp16 = const()[name = tensor("flow_net_cond_embed_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(2137856)))]; + tensor transformer_out_to_fp16 = cast(dtype = transformer_out_to_fp16_dtype_0, x = transformer_out)[name = tensor("cast_145")]; + tensor linear_5_cast_fp16 = linear(bias = flow_net_cond_embed_bias_to_fp16, weight = flow_net_cond_embed_weight_to_fp16, x = transformer_out_to_fp16)[name = tensor("linear_5_cast_fp16")]; + tensor input_13_cast_fp16 = add(x = _inversed_t_combined_1_cast_fp16, y = linear_5_cast_fp16)[name = tensor("input_13_cast_fp16")]; + tensor input_15_cast_fp16 = silu(x = input_13_cast_fp16)[name = tensor("input_15_cast_fp16")]; + tensor flow_net_res_blocks_0_adaLN_modulation_1_weight_to_fp16 = const()[name = tensor("flow_net_res_blocks_0_adaLN_modulation_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(2138944)))]; + tensor flow_net_res_blocks_0_adaLN_modulation_1_bias_to_fp16 = const()[name = tensor("flow_net_res_blocks_0_adaLN_modulation_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(3711872)))]; + tensor linear_6_cast_fp16 = linear(bias = flow_net_res_blocks_0_adaLN_modulation_1_bias_to_fp16, weight = flow_net_res_blocks_0_adaLN_modulation_1_weight_to_fp16, x = input_15_cast_fp16)[name = tensor("linear_6_cast_fp16")]; + tensor var_142_split_sizes_0 = const()[name = tensor("op_142_split_sizes_0"), val = tensor([512, 512, 512])]; + tensor var_142_axis_0 = const()[name = tensor("op_142_axis_0"), val = tensor(-1)]; + tensor var_142_cast_fp16_0, tensor var_142_cast_fp16_1, tensor var_142_cast_fp16_2 = split(axis = var_142_axis_0, split_sizes = var_142_split_sizes_0, x = linear_6_cast_fp16)[name = tensor("op_142_cast_fp16")]; + tensor mean_1_axes_0 = const()[name = tensor("mean_1_axes_0"), val = tensor([-1])]; + tensor mean_1_keep_dims_0 = const()[name = tensor("mean_1_keep_dims_0"), val = tensor(true)]; + tensor mean_1_cast_fp16 = reduce_mean(axes = mean_1_axes_0, keep_dims = mean_1_keep_dims_0, x = linear_0_cast_fp16)[name = tensor("mean_1_cast_fp16")]; + tensor sub_4_cast_fp16 = sub(x = linear_0_cast_fp16, y = mean_1_cast_fp16)[name = tensor("sub_4_cast_fp16")]; + tensor square_2_cast_fp16 = square(x = sub_4_cast_fp16)[name = tensor("square_2_cast_fp16")]; + tensor reduce_mean_5_axes_0 = const()[name = tensor("reduce_mean_5_axes_0"), val = tensor([-1])]; + tensor reduce_mean_5_keep_dims_0 = const()[name = tensor("reduce_mean_5_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_5_cast_fp16 = reduce_mean(axes = reduce_mean_5_axes_0, keep_dims = reduce_mean_5_keep_dims_0, x = square_2_cast_fp16)[name = tensor("reduce_mean_5_cast_fp16")]; + tensor var_152_to_fp16 = const()[name = tensor("op_152_to_fp16"), val = tensor(0x1.1p-20)]; + tensor var_153_cast_fp16 = add(x = reduce_mean_5_cast_fp16, y = var_152_to_fp16)[name = tensor("op_153_cast_fp16")]; + tensor var_154_cast_fp16 = sqrt(x = var_153_cast_fp16)[name = tensor("op_154_cast_fp16")]; + tensor x_7_cast_fp16 = real_div(x = sub_4_cast_fp16, y = var_154_cast_fp16)[name = tensor("x_7_cast_fp16")]; + tensor flow_net_res_blocks_0_in_ln_weight_to_fp16 = const()[name = tensor("flow_net_res_blocks_0_in_ln_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(3715008)))]; + tensor var_156_cast_fp16 = mul(x = x_7_cast_fp16, y = flow_net_res_blocks_0_in_ln_weight_to_fp16)[name = tensor("op_156_cast_fp16")]; + tensor flow_net_res_blocks_0_in_ln_bias_to_fp16 = const()[name = tensor("flow_net_res_blocks_0_in_ln_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(3716096)))]; + tensor x_9_cast_fp16 = add(x = var_156_cast_fp16, y = flow_net_res_blocks_0_in_ln_bias_to_fp16)[name = tensor("x_9_cast_fp16")]; + tensor var_158_promoted_to_fp16 = const()[name = tensor("op_158_promoted_to_fp16"), val = tensor(0x1p+0)]; + tensor var_159_cast_fp16 = add(x = var_142_cast_fp16_1, y = var_158_promoted_to_fp16)[name = tensor("op_159_cast_fp16")]; + tensor var_160_cast_fp16 = mul(x = x_9_cast_fp16, y = var_159_cast_fp16)[name = tensor("op_160_cast_fp16")]; + tensor input_17_cast_fp16 = add(x = var_160_cast_fp16, y = var_142_cast_fp16_0)[name = tensor("input_17_cast_fp16")]; + tensor flow_net_res_blocks_0_mlp_0_weight_to_fp16 = const()[name = tensor("flow_net_res_blocks_0_mlp_0_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(3717184)))]; + tensor flow_net_res_blocks_0_mlp_0_bias_to_fp16 = const()[name = tensor("flow_net_res_blocks_0_mlp_0_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(4241536)))]; + tensor linear_7_cast_fp16 = linear(bias = flow_net_res_blocks_0_mlp_0_bias_to_fp16, weight = flow_net_res_blocks_0_mlp_0_weight_to_fp16, x = input_17_cast_fp16)[name = tensor("linear_7_cast_fp16")]; + tensor input_21_cast_fp16 = silu(x = linear_7_cast_fp16)[name = tensor("input_21_cast_fp16")]; + tensor flow_net_res_blocks_0_mlp_2_weight_to_fp16 = const()[name = tensor("flow_net_res_blocks_0_mlp_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(4242624)))]; + tensor flow_net_res_blocks_0_mlp_2_bias_to_fp16 = const()[name = tensor("flow_net_res_blocks_0_mlp_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(4766976)))]; + tensor linear_8_cast_fp16 = linear(bias = flow_net_res_blocks_0_mlp_2_bias_to_fp16, weight = flow_net_res_blocks_0_mlp_2_weight_to_fp16, x = input_21_cast_fp16)[name = tensor("linear_8_cast_fp16")]; + tensor var_171_cast_fp16 = mul(x = var_142_cast_fp16_2, y = linear_8_cast_fp16)[name = tensor("op_171_cast_fp16")]; + tensor x_11_cast_fp16 = add(x = linear_0_cast_fp16, y = var_171_cast_fp16)[name = tensor("x_11_cast_fp16")]; + tensor flow_net_res_blocks_1_adaLN_modulation_1_weight_to_fp16 = const()[name = tensor("flow_net_res_blocks_1_adaLN_modulation_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(4768064)))]; + tensor flow_net_res_blocks_1_adaLN_modulation_1_bias_to_fp16 = const()[name = tensor("flow_net_res_blocks_1_adaLN_modulation_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(6340992)))]; + tensor linear_9_cast_fp16 = linear(bias = flow_net_res_blocks_1_adaLN_modulation_1_bias_to_fp16, weight = flow_net_res_blocks_1_adaLN_modulation_1_weight_to_fp16, x = input_15_cast_fp16)[name = tensor("linear_9_cast_fp16")]; + tensor var_181_split_sizes_0 = const()[name = tensor("op_181_split_sizes_0"), val = tensor([512, 512, 512])]; + tensor var_181_axis_0 = const()[name = tensor("op_181_axis_0"), val = tensor(-1)]; + tensor var_181_cast_fp16_0, tensor var_181_cast_fp16_1, tensor var_181_cast_fp16_2 = split(axis = var_181_axis_0, split_sizes = var_181_split_sizes_0, x = linear_9_cast_fp16)[name = tensor("op_181_cast_fp16")]; + tensor mean_3_axes_0 = const()[name = tensor("mean_3_axes_0"), val = tensor([-1])]; + tensor mean_3_keep_dims_0 = const()[name = tensor("mean_3_keep_dims_0"), val = tensor(true)]; + tensor mean_3_cast_fp16 = reduce_mean(axes = mean_3_axes_0, keep_dims = mean_3_keep_dims_0, x = x_11_cast_fp16)[name = tensor("mean_3_cast_fp16")]; + tensor sub_5_cast_fp16 = sub(x = x_11_cast_fp16, y = mean_3_cast_fp16)[name = tensor("sub_5_cast_fp16")]; + tensor square_3_cast_fp16 = square(x = sub_5_cast_fp16)[name = tensor("square_3_cast_fp16")]; + tensor reduce_mean_7_axes_0 = const()[name = tensor("reduce_mean_7_axes_0"), val = tensor([-1])]; + tensor reduce_mean_7_keep_dims_0 = const()[name = tensor("reduce_mean_7_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_7_cast_fp16 = reduce_mean(axes = reduce_mean_7_axes_0, keep_dims = reduce_mean_7_keep_dims_0, x = square_3_cast_fp16)[name = tensor("reduce_mean_7_cast_fp16")]; + tensor var_191_to_fp16 = const()[name = tensor("op_191_to_fp16"), val = tensor(0x1.1p-20)]; + tensor var_192_cast_fp16 = add(x = reduce_mean_7_cast_fp16, y = var_191_to_fp16)[name = tensor("op_192_cast_fp16")]; + tensor var_193_cast_fp16 = sqrt(x = var_192_cast_fp16)[name = tensor("op_193_cast_fp16")]; + tensor x_13_cast_fp16 = real_div(x = sub_5_cast_fp16, y = var_193_cast_fp16)[name = tensor("x_13_cast_fp16")]; + tensor flow_net_res_blocks_1_in_ln_weight_to_fp16 = const()[name = tensor("flow_net_res_blocks_1_in_ln_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(6344128)))]; + tensor var_195_cast_fp16 = mul(x = x_13_cast_fp16, y = flow_net_res_blocks_1_in_ln_weight_to_fp16)[name = tensor("op_195_cast_fp16")]; + tensor flow_net_res_blocks_1_in_ln_bias_to_fp16 = const()[name = tensor("flow_net_res_blocks_1_in_ln_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(6345216)))]; + tensor x_15_cast_fp16 = add(x = var_195_cast_fp16, y = flow_net_res_blocks_1_in_ln_bias_to_fp16)[name = tensor("x_15_cast_fp16")]; + tensor var_197_promoted_to_fp16 = const()[name = tensor("op_197_promoted_to_fp16"), val = tensor(0x1p+0)]; + tensor var_198_cast_fp16 = add(x = var_181_cast_fp16_1, y = var_197_promoted_to_fp16)[name = tensor("op_198_cast_fp16")]; + tensor var_199_cast_fp16 = mul(x = x_15_cast_fp16, y = var_198_cast_fp16)[name = tensor("op_199_cast_fp16")]; + tensor input_25_cast_fp16 = add(x = var_199_cast_fp16, y = var_181_cast_fp16_0)[name = tensor("input_25_cast_fp16")]; + tensor flow_net_res_blocks_1_mlp_0_weight_to_fp16 = const()[name = tensor("flow_net_res_blocks_1_mlp_0_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(6346304)))]; + tensor flow_net_res_blocks_1_mlp_0_bias_to_fp16 = const()[name = tensor("flow_net_res_blocks_1_mlp_0_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(6870656)))]; + tensor linear_10_cast_fp16 = linear(bias = flow_net_res_blocks_1_mlp_0_bias_to_fp16, weight = flow_net_res_blocks_1_mlp_0_weight_to_fp16, x = input_25_cast_fp16)[name = tensor("linear_10_cast_fp16")]; + tensor input_29_cast_fp16 = silu(x = linear_10_cast_fp16)[name = tensor("input_29_cast_fp16")]; + tensor flow_net_res_blocks_1_mlp_2_weight_to_fp16 = const()[name = tensor("flow_net_res_blocks_1_mlp_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(6871744)))]; + tensor flow_net_res_blocks_1_mlp_2_bias_to_fp16 = const()[name = tensor("flow_net_res_blocks_1_mlp_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(7396096)))]; + tensor linear_11_cast_fp16 = linear(bias = flow_net_res_blocks_1_mlp_2_bias_to_fp16, weight = flow_net_res_blocks_1_mlp_2_weight_to_fp16, x = input_29_cast_fp16)[name = tensor("linear_11_cast_fp16")]; + tensor var_210_cast_fp16 = mul(x = var_181_cast_fp16_2, y = linear_11_cast_fp16)[name = tensor("op_210_cast_fp16")]; + tensor x_17_cast_fp16 = add(x = x_11_cast_fp16, y = var_210_cast_fp16)[name = tensor("x_17_cast_fp16")]; + tensor flow_net_res_blocks_2_adaLN_modulation_1_weight_to_fp16 = const()[name = tensor("flow_net_res_blocks_2_adaLN_modulation_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(7397184)))]; + tensor flow_net_res_blocks_2_adaLN_modulation_1_bias_to_fp16 = const()[name = tensor("flow_net_res_blocks_2_adaLN_modulation_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8970112)))]; + tensor linear_12_cast_fp16 = linear(bias = flow_net_res_blocks_2_adaLN_modulation_1_bias_to_fp16, weight = flow_net_res_blocks_2_adaLN_modulation_1_weight_to_fp16, x = input_15_cast_fp16)[name = tensor("linear_12_cast_fp16")]; + tensor var_220_split_sizes_0 = const()[name = tensor("op_220_split_sizes_0"), val = tensor([512, 512, 512])]; + tensor var_220_axis_0 = const()[name = tensor("op_220_axis_0"), val = tensor(-1)]; + tensor var_220_cast_fp16_0, tensor var_220_cast_fp16_1, tensor var_220_cast_fp16_2 = split(axis = var_220_axis_0, split_sizes = var_220_split_sizes_0, x = linear_12_cast_fp16)[name = tensor("op_220_cast_fp16")]; + tensor mean_5_axes_0 = const()[name = tensor("mean_5_axes_0"), val = tensor([-1])]; + tensor mean_5_keep_dims_0 = const()[name = tensor("mean_5_keep_dims_0"), val = tensor(true)]; + tensor mean_5_cast_fp16 = reduce_mean(axes = mean_5_axes_0, keep_dims = mean_5_keep_dims_0, x = x_17_cast_fp16)[name = tensor("mean_5_cast_fp16")]; + tensor sub_6_cast_fp16 = sub(x = x_17_cast_fp16, y = mean_5_cast_fp16)[name = tensor("sub_6_cast_fp16")]; + tensor square_4_cast_fp16 = square(x = sub_6_cast_fp16)[name = tensor("square_4_cast_fp16")]; + tensor reduce_mean_9_axes_0 = const()[name = tensor("reduce_mean_9_axes_0"), val = tensor([-1])]; + tensor reduce_mean_9_keep_dims_0 = const()[name = tensor("reduce_mean_9_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_9_cast_fp16 = reduce_mean(axes = reduce_mean_9_axes_0, keep_dims = reduce_mean_9_keep_dims_0, x = square_4_cast_fp16)[name = tensor("reduce_mean_9_cast_fp16")]; + tensor var_230_to_fp16 = const()[name = tensor("op_230_to_fp16"), val = tensor(0x1.1p-20)]; + tensor var_231_cast_fp16 = add(x = reduce_mean_9_cast_fp16, y = var_230_to_fp16)[name = tensor("op_231_cast_fp16")]; + tensor var_232_cast_fp16 = sqrt(x = var_231_cast_fp16)[name = tensor("op_232_cast_fp16")]; + tensor x_19_cast_fp16 = real_div(x = sub_6_cast_fp16, y = var_232_cast_fp16)[name = tensor("x_19_cast_fp16")]; + tensor flow_net_res_blocks_2_in_ln_weight_to_fp16 = const()[name = tensor("flow_net_res_blocks_2_in_ln_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8973248)))]; + tensor var_234_cast_fp16 = mul(x = x_19_cast_fp16, y = flow_net_res_blocks_2_in_ln_weight_to_fp16)[name = tensor("op_234_cast_fp16")]; + tensor flow_net_res_blocks_2_in_ln_bias_to_fp16 = const()[name = tensor("flow_net_res_blocks_2_in_ln_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8974336)))]; + tensor x_21_cast_fp16 = add(x = var_234_cast_fp16, y = flow_net_res_blocks_2_in_ln_bias_to_fp16)[name = tensor("x_21_cast_fp16")]; + tensor var_236_promoted_to_fp16 = const()[name = tensor("op_236_promoted_to_fp16"), val = tensor(0x1p+0)]; + tensor var_237_cast_fp16 = add(x = var_220_cast_fp16_1, y = var_236_promoted_to_fp16)[name = tensor("op_237_cast_fp16")]; + tensor var_238_cast_fp16 = mul(x = x_21_cast_fp16, y = var_237_cast_fp16)[name = tensor("op_238_cast_fp16")]; + tensor input_33_cast_fp16 = add(x = var_238_cast_fp16, y = var_220_cast_fp16_0)[name = tensor("input_33_cast_fp16")]; + tensor flow_net_res_blocks_2_mlp_0_weight_to_fp16 = const()[name = tensor("flow_net_res_blocks_2_mlp_0_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8975424)))]; + tensor flow_net_res_blocks_2_mlp_0_bias_to_fp16 = const()[name = tensor("flow_net_res_blocks_2_mlp_0_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9499776)))]; + tensor linear_13_cast_fp16 = linear(bias = flow_net_res_blocks_2_mlp_0_bias_to_fp16, weight = flow_net_res_blocks_2_mlp_0_weight_to_fp16, x = input_33_cast_fp16)[name = tensor("linear_13_cast_fp16")]; + tensor input_37_cast_fp16 = silu(x = linear_13_cast_fp16)[name = tensor("input_37_cast_fp16")]; + tensor flow_net_res_blocks_2_mlp_2_weight_to_fp16 = const()[name = tensor("flow_net_res_blocks_2_mlp_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9500864)))]; + tensor flow_net_res_blocks_2_mlp_2_bias_to_fp16 = const()[name = tensor("flow_net_res_blocks_2_mlp_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(10025216)))]; + tensor linear_14_cast_fp16 = linear(bias = flow_net_res_blocks_2_mlp_2_bias_to_fp16, weight = flow_net_res_blocks_2_mlp_2_weight_to_fp16, x = input_37_cast_fp16)[name = tensor("linear_14_cast_fp16")]; + tensor var_249_cast_fp16 = mul(x = var_220_cast_fp16_2, y = linear_14_cast_fp16)[name = tensor("op_249_cast_fp16")]; + tensor x_23_cast_fp16 = add(x = x_17_cast_fp16, y = var_249_cast_fp16)[name = tensor("x_23_cast_fp16")]; + tensor flow_net_res_blocks_3_adaLN_modulation_1_weight_to_fp16 = const()[name = tensor("flow_net_res_blocks_3_adaLN_modulation_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(10026304)))]; + tensor flow_net_res_blocks_3_adaLN_modulation_1_bias_to_fp16 = const()[name = tensor("flow_net_res_blocks_3_adaLN_modulation_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(11599232)))]; + tensor linear_15_cast_fp16 = linear(bias = flow_net_res_blocks_3_adaLN_modulation_1_bias_to_fp16, weight = flow_net_res_blocks_3_adaLN_modulation_1_weight_to_fp16, x = input_15_cast_fp16)[name = tensor("linear_15_cast_fp16")]; + tensor var_259_split_sizes_0 = const()[name = tensor("op_259_split_sizes_0"), val = tensor([512, 512, 512])]; + tensor var_259_axis_0 = const()[name = tensor("op_259_axis_0"), val = tensor(-1)]; + tensor var_259_cast_fp16_0, tensor var_259_cast_fp16_1, tensor var_259_cast_fp16_2 = split(axis = var_259_axis_0, split_sizes = var_259_split_sizes_0, x = linear_15_cast_fp16)[name = tensor("op_259_cast_fp16")]; + tensor mean_7_axes_0 = const()[name = tensor("mean_7_axes_0"), val = tensor([-1])]; + tensor mean_7_keep_dims_0 = const()[name = tensor("mean_7_keep_dims_0"), val = tensor(true)]; + tensor mean_7_cast_fp16 = reduce_mean(axes = mean_7_axes_0, keep_dims = mean_7_keep_dims_0, x = x_23_cast_fp16)[name = tensor("mean_7_cast_fp16")]; + tensor sub_7_cast_fp16 = sub(x = x_23_cast_fp16, y = mean_7_cast_fp16)[name = tensor("sub_7_cast_fp16")]; + tensor square_5_cast_fp16 = square(x = sub_7_cast_fp16)[name = tensor("square_5_cast_fp16")]; + tensor reduce_mean_11_axes_0 = const()[name = tensor("reduce_mean_11_axes_0"), val = tensor([-1])]; + tensor reduce_mean_11_keep_dims_0 = const()[name = tensor("reduce_mean_11_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_11_cast_fp16 = reduce_mean(axes = reduce_mean_11_axes_0, keep_dims = reduce_mean_11_keep_dims_0, x = square_5_cast_fp16)[name = tensor("reduce_mean_11_cast_fp16")]; + tensor var_269_to_fp16 = const()[name = tensor("op_269_to_fp16"), val = tensor(0x1.1p-20)]; + tensor var_270_cast_fp16 = add(x = reduce_mean_11_cast_fp16, y = var_269_to_fp16)[name = tensor("op_270_cast_fp16")]; + tensor var_271_cast_fp16 = sqrt(x = var_270_cast_fp16)[name = tensor("op_271_cast_fp16")]; + tensor x_25_cast_fp16 = real_div(x = sub_7_cast_fp16, y = var_271_cast_fp16)[name = tensor("x_25_cast_fp16")]; + tensor flow_net_res_blocks_3_in_ln_weight_to_fp16 = const()[name = tensor("flow_net_res_blocks_3_in_ln_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(11602368)))]; + tensor var_273_cast_fp16 = mul(x = x_25_cast_fp16, y = flow_net_res_blocks_3_in_ln_weight_to_fp16)[name = tensor("op_273_cast_fp16")]; + tensor flow_net_res_blocks_3_in_ln_bias_to_fp16 = const()[name = tensor("flow_net_res_blocks_3_in_ln_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(11603456)))]; + tensor x_27_cast_fp16 = add(x = var_273_cast_fp16, y = flow_net_res_blocks_3_in_ln_bias_to_fp16)[name = tensor("x_27_cast_fp16")]; + tensor var_275_promoted_to_fp16 = const()[name = tensor("op_275_promoted_to_fp16"), val = tensor(0x1p+0)]; + tensor var_276_cast_fp16 = add(x = var_259_cast_fp16_1, y = var_275_promoted_to_fp16)[name = tensor("op_276_cast_fp16")]; + tensor var_277_cast_fp16 = mul(x = x_27_cast_fp16, y = var_276_cast_fp16)[name = tensor("op_277_cast_fp16")]; + tensor input_41_cast_fp16 = add(x = var_277_cast_fp16, y = var_259_cast_fp16_0)[name = tensor("input_41_cast_fp16")]; + tensor flow_net_res_blocks_3_mlp_0_weight_to_fp16 = const()[name = tensor("flow_net_res_blocks_3_mlp_0_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(11604544)))]; + tensor flow_net_res_blocks_3_mlp_0_bias_to_fp16 = const()[name = tensor("flow_net_res_blocks_3_mlp_0_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(12128896)))]; + tensor linear_16_cast_fp16 = linear(bias = flow_net_res_blocks_3_mlp_0_bias_to_fp16, weight = flow_net_res_blocks_3_mlp_0_weight_to_fp16, x = input_41_cast_fp16)[name = tensor("linear_16_cast_fp16")]; + tensor input_45_cast_fp16 = silu(x = linear_16_cast_fp16)[name = tensor("input_45_cast_fp16")]; + tensor flow_net_res_blocks_3_mlp_2_weight_to_fp16 = const()[name = tensor("flow_net_res_blocks_3_mlp_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(12129984)))]; + tensor flow_net_res_blocks_3_mlp_2_bias_to_fp16 = const()[name = tensor("flow_net_res_blocks_3_mlp_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(12654336)))]; + tensor linear_17_cast_fp16 = linear(bias = flow_net_res_blocks_3_mlp_2_bias_to_fp16, weight = flow_net_res_blocks_3_mlp_2_weight_to_fp16, x = input_45_cast_fp16)[name = tensor("linear_17_cast_fp16")]; + tensor var_288_cast_fp16 = mul(x = var_259_cast_fp16_2, y = linear_17_cast_fp16)[name = tensor("op_288_cast_fp16")]; + tensor x_29_cast_fp16 = add(x = x_23_cast_fp16, y = var_288_cast_fp16)[name = tensor("x_29_cast_fp16")]; + tensor flow_net_res_blocks_4_adaLN_modulation_1_weight_to_fp16 = const()[name = tensor("flow_net_res_blocks_4_adaLN_modulation_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(12655424)))]; + tensor flow_net_res_blocks_4_adaLN_modulation_1_bias_to_fp16 = const()[name = tensor("flow_net_res_blocks_4_adaLN_modulation_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(14228352)))]; + tensor linear_18_cast_fp16 = linear(bias = flow_net_res_blocks_4_adaLN_modulation_1_bias_to_fp16, weight = flow_net_res_blocks_4_adaLN_modulation_1_weight_to_fp16, x = input_15_cast_fp16)[name = tensor("linear_18_cast_fp16")]; + tensor var_298_split_sizes_0 = const()[name = tensor("op_298_split_sizes_0"), val = tensor([512, 512, 512])]; + tensor var_298_axis_0 = const()[name = tensor("op_298_axis_0"), val = tensor(-1)]; + tensor var_298_cast_fp16_0, tensor var_298_cast_fp16_1, tensor var_298_cast_fp16_2 = split(axis = var_298_axis_0, split_sizes = var_298_split_sizes_0, x = linear_18_cast_fp16)[name = tensor("op_298_cast_fp16")]; + tensor mean_9_axes_0 = const()[name = tensor("mean_9_axes_0"), val = tensor([-1])]; + tensor mean_9_keep_dims_0 = const()[name = tensor("mean_9_keep_dims_0"), val = tensor(true)]; + tensor mean_9_cast_fp16 = reduce_mean(axes = mean_9_axes_0, keep_dims = mean_9_keep_dims_0, x = x_29_cast_fp16)[name = tensor("mean_9_cast_fp16")]; + tensor sub_8_cast_fp16 = sub(x = x_29_cast_fp16, y = mean_9_cast_fp16)[name = tensor("sub_8_cast_fp16")]; + tensor square_6_cast_fp16 = square(x = sub_8_cast_fp16)[name = tensor("square_6_cast_fp16")]; + tensor reduce_mean_13_axes_0 = const()[name = tensor("reduce_mean_13_axes_0"), val = tensor([-1])]; + tensor reduce_mean_13_keep_dims_0 = const()[name = tensor("reduce_mean_13_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_13_cast_fp16 = reduce_mean(axes = reduce_mean_13_axes_0, keep_dims = reduce_mean_13_keep_dims_0, x = square_6_cast_fp16)[name = tensor("reduce_mean_13_cast_fp16")]; + tensor var_308_to_fp16 = const()[name = tensor("op_308_to_fp16"), val = tensor(0x1.1p-20)]; + tensor var_309_cast_fp16 = add(x = reduce_mean_13_cast_fp16, y = var_308_to_fp16)[name = tensor("op_309_cast_fp16")]; + tensor var_310_cast_fp16 = sqrt(x = var_309_cast_fp16)[name = tensor("op_310_cast_fp16")]; + tensor x_31_cast_fp16 = real_div(x = sub_8_cast_fp16, y = var_310_cast_fp16)[name = tensor("x_31_cast_fp16")]; + tensor flow_net_res_blocks_4_in_ln_weight_to_fp16 = const()[name = tensor("flow_net_res_blocks_4_in_ln_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(14231488)))]; + tensor var_312_cast_fp16 = mul(x = x_31_cast_fp16, y = flow_net_res_blocks_4_in_ln_weight_to_fp16)[name = tensor("op_312_cast_fp16")]; + tensor flow_net_res_blocks_4_in_ln_bias_to_fp16 = const()[name = tensor("flow_net_res_blocks_4_in_ln_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(14232576)))]; + tensor x_33_cast_fp16 = add(x = var_312_cast_fp16, y = flow_net_res_blocks_4_in_ln_bias_to_fp16)[name = tensor("x_33_cast_fp16")]; + tensor var_314_promoted_to_fp16 = const()[name = tensor("op_314_promoted_to_fp16"), val = tensor(0x1p+0)]; + tensor var_315_cast_fp16 = add(x = var_298_cast_fp16_1, y = var_314_promoted_to_fp16)[name = tensor("op_315_cast_fp16")]; + tensor var_316_cast_fp16 = mul(x = x_33_cast_fp16, y = var_315_cast_fp16)[name = tensor("op_316_cast_fp16")]; + tensor input_49_cast_fp16 = add(x = var_316_cast_fp16, y = var_298_cast_fp16_0)[name = tensor("input_49_cast_fp16")]; + tensor flow_net_res_blocks_4_mlp_0_weight_to_fp16 = const()[name = tensor("flow_net_res_blocks_4_mlp_0_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(14233664)))]; + tensor flow_net_res_blocks_4_mlp_0_bias_to_fp16 = const()[name = tensor("flow_net_res_blocks_4_mlp_0_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(14758016)))]; + tensor linear_19_cast_fp16 = linear(bias = flow_net_res_blocks_4_mlp_0_bias_to_fp16, weight = flow_net_res_blocks_4_mlp_0_weight_to_fp16, x = input_49_cast_fp16)[name = tensor("linear_19_cast_fp16")]; + tensor input_53_cast_fp16 = silu(x = linear_19_cast_fp16)[name = tensor("input_53_cast_fp16")]; + tensor flow_net_res_blocks_4_mlp_2_weight_to_fp16 = const()[name = tensor("flow_net_res_blocks_4_mlp_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(14759104)))]; + tensor flow_net_res_blocks_4_mlp_2_bias_to_fp16 = const()[name = tensor("flow_net_res_blocks_4_mlp_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15283456)))]; + tensor linear_20_cast_fp16 = linear(bias = flow_net_res_blocks_4_mlp_2_bias_to_fp16, weight = flow_net_res_blocks_4_mlp_2_weight_to_fp16, x = input_53_cast_fp16)[name = tensor("linear_20_cast_fp16")]; + tensor var_327_cast_fp16 = mul(x = var_298_cast_fp16_2, y = linear_20_cast_fp16)[name = tensor("op_327_cast_fp16")]; + tensor x_35_cast_fp16 = add(x = x_29_cast_fp16, y = var_327_cast_fp16)[name = tensor("x_35_cast_fp16")]; + tensor flow_net_res_blocks_5_adaLN_modulation_1_weight_to_fp16 = const()[name = tensor("flow_net_res_blocks_5_adaLN_modulation_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15284544)))]; + tensor flow_net_res_blocks_5_adaLN_modulation_1_bias_to_fp16 = const()[name = tensor("flow_net_res_blocks_5_adaLN_modulation_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(16857472)))]; + tensor linear_21_cast_fp16 = linear(bias = flow_net_res_blocks_5_adaLN_modulation_1_bias_to_fp16, weight = flow_net_res_blocks_5_adaLN_modulation_1_weight_to_fp16, x = input_15_cast_fp16)[name = tensor("linear_21_cast_fp16")]; + tensor var_337_split_sizes_0 = const()[name = tensor("op_337_split_sizes_0"), val = tensor([512, 512, 512])]; + tensor var_337_axis_0 = const()[name = tensor("op_337_axis_0"), val = tensor(-1)]; + tensor var_337_cast_fp16_0, tensor var_337_cast_fp16_1, tensor var_337_cast_fp16_2 = split(axis = var_337_axis_0, split_sizes = var_337_split_sizes_0, x = linear_21_cast_fp16)[name = tensor("op_337_cast_fp16")]; + tensor mean_11_axes_0 = const()[name = tensor("mean_11_axes_0"), val = tensor([-1])]; + tensor mean_11_keep_dims_0 = const()[name = tensor("mean_11_keep_dims_0"), val = tensor(true)]; + tensor mean_11_cast_fp16 = reduce_mean(axes = mean_11_axes_0, keep_dims = mean_11_keep_dims_0, x = x_35_cast_fp16)[name = tensor("mean_11_cast_fp16")]; + tensor sub_9_cast_fp16 = sub(x = x_35_cast_fp16, y = mean_11_cast_fp16)[name = tensor("sub_9_cast_fp16")]; + tensor square_7_cast_fp16 = square(x = sub_9_cast_fp16)[name = tensor("square_7_cast_fp16")]; + tensor reduce_mean_15_axes_0 = const()[name = tensor("reduce_mean_15_axes_0"), val = tensor([-1])]; + tensor reduce_mean_15_keep_dims_0 = const()[name = tensor("reduce_mean_15_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_15_cast_fp16 = reduce_mean(axes = reduce_mean_15_axes_0, keep_dims = reduce_mean_15_keep_dims_0, x = square_7_cast_fp16)[name = tensor("reduce_mean_15_cast_fp16")]; + tensor var_347_to_fp16 = const()[name = tensor("op_347_to_fp16"), val = tensor(0x1.1p-20)]; + tensor var_348_cast_fp16 = add(x = reduce_mean_15_cast_fp16, y = var_347_to_fp16)[name = tensor("op_348_cast_fp16")]; + tensor var_349_cast_fp16 = sqrt(x = var_348_cast_fp16)[name = tensor("op_349_cast_fp16")]; + tensor x_37_cast_fp16 = real_div(x = sub_9_cast_fp16, y = var_349_cast_fp16)[name = tensor("x_37_cast_fp16")]; + tensor flow_net_res_blocks_5_in_ln_weight_to_fp16 = const()[name = tensor("flow_net_res_blocks_5_in_ln_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(16860608)))]; + tensor var_351_cast_fp16 = mul(x = x_37_cast_fp16, y = flow_net_res_blocks_5_in_ln_weight_to_fp16)[name = tensor("op_351_cast_fp16")]; + tensor flow_net_res_blocks_5_in_ln_bias_to_fp16 = const()[name = tensor("flow_net_res_blocks_5_in_ln_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(16861696)))]; + tensor x_39_cast_fp16 = add(x = var_351_cast_fp16, y = flow_net_res_blocks_5_in_ln_bias_to_fp16)[name = tensor("x_39_cast_fp16")]; + tensor var_353_promoted_to_fp16 = const()[name = tensor("op_353_promoted_to_fp16"), val = tensor(0x1p+0)]; + tensor var_354_cast_fp16 = add(x = var_337_cast_fp16_1, y = var_353_promoted_to_fp16)[name = tensor("op_354_cast_fp16")]; + tensor var_355_cast_fp16 = mul(x = x_39_cast_fp16, y = var_354_cast_fp16)[name = tensor("op_355_cast_fp16")]; + tensor input_57_cast_fp16 = add(x = var_355_cast_fp16, y = var_337_cast_fp16_0)[name = tensor("input_57_cast_fp16")]; + tensor flow_net_res_blocks_5_mlp_0_weight_to_fp16 = const()[name = tensor("flow_net_res_blocks_5_mlp_0_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(16862784)))]; + tensor flow_net_res_blocks_5_mlp_0_bias_to_fp16 = const()[name = tensor("flow_net_res_blocks_5_mlp_0_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(17387136)))]; + tensor linear_22_cast_fp16 = linear(bias = flow_net_res_blocks_5_mlp_0_bias_to_fp16, weight = flow_net_res_blocks_5_mlp_0_weight_to_fp16, x = input_57_cast_fp16)[name = tensor("linear_22_cast_fp16")]; + tensor input_61_cast_fp16 = silu(x = linear_22_cast_fp16)[name = tensor("input_61_cast_fp16")]; + tensor flow_net_res_blocks_5_mlp_2_weight_to_fp16 = const()[name = tensor("flow_net_res_blocks_5_mlp_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(17388224)))]; + tensor flow_net_res_blocks_5_mlp_2_bias_to_fp16 = const()[name = tensor("flow_net_res_blocks_5_mlp_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(17912576)))]; + tensor linear_23_cast_fp16 = linear(bias = flow_net_res_blocks_5_mlp_2_bias_to_fp16, weight = flow_net_res_blocks_5_mlp_2_weight_to_fp16, x = input_61_cast_fp16)[name = tensor("linear_23_cast_fp16")]; + tensor var_366_cast_fp16 = mul(x = var_337_cast_fp16_2, y = linear_23_cast_fp16)[name = tensor("op_366_cast_fp16")]; + tensor x_41_cast_fp16 = add(x = x_35_cast_fp16, y = var_366_cast_fp16)[name = tensor("x_41_cast_fp16")]; + tensor flow_net_final_layer_adaLN_modulation_1_weight_to_fp16 = const()[name = tensor("flow_net_final_layer_adaLN_modulation_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(17913664)))]; + tensor flow_net_final_layer_adaLN_modulation_1_bias_to_fp16 = const()[name = tensor("flow_net_final_layer_adaLN_modulation_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(18962304)))]; + tensor linear_24_cast_fp16 = linear(bias = flow_net_final_layer_adaLN_modulation_1_bias_to_fp16, weight = flow_net_final_layer_adaLN_modulation_1_weight_to_fp16, x = input_15_cast_fp16)[name = tensor("linear_24_cast_fp16")]; + tensor var_375_split_sizes_0 = const()[name = tensor("op_375_split_sizes_0"), val = tensor([512, 512])]; + tensor var_375_axis_0 = const()[name = tensor("op_375_axis_0"), val = tensor(-1)]; + tensor var_375_cast_fp16_0, tensor var_375_cast_fp16_1 = split(axis = var_375_axis_0, split_sizes = var_375_split_sizes_0, x = linear_24_cast_fp16)[name = tensor("op_375_cast_fp16")]; + tensor mean_13_axes_0 = const()[name = tensor("mean_13_axes_0"), val = tensor([-1])]; + tensor mean_13_keep_dims_0 = const()[name = tensor("mean_13_keep_dims_0"), val = tensor(true)]; + tensor mean_13_cast_fp16 = reduce_mean(axes = mean_13_axes_0, keep_dims = mean_13_keep_dims_0, x = x_41_cast_fp16)[name = tensor("mean_13_cast_fp16")]; + tensor sub_10_cast_fp16 = sub(x = x_41_cast_fp16, y = mean_13_cast_fp16)[name = tensor("sub_10_cast_fp16")]; + tensor square_8_cast_fp16 = square(x = sub_10_cast_fp16)[name = tensor("square_8_cast_fp16")]; + tensor reduce_mean_17_axes_0 = const()[name = tensor("reduce_mean_17_axes_0"), val = tensor([-1])]; + tensor reduce_mean_17_keep_dims_0 = const()[name = tensor("reduce_mean_17_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_17_cast_fp16 = reduce_mean(axes = reduce_mean_17_axes_0, keep_dims = reduce_mean_17_keep_dims_0, x = square_8_cast_fp16)[name = tensor("reduce_mean_17_cast_fp16")]; + tensor var_382_to_fp16 = const()[name = tensor("op_382_to_fp16"), val = tensor(0x1.1p-20)]; + tensor var_383_cast_fp16 = add(x = reduce_mean_17_cast_fp16, y = var_382_to_fp16)[name = tensor("op_383_cast_fp16")]; + tensor var_384_cast_fp16 = sqrt(x = var_383_cast_fp16)[name = tensor("op_384_cast_fp16")]; + tensor x_43_cast_fp16 = real_div(x = sub_10_cast_fp16, y = var_384_cast_fp16)[name = tensor("x_43_cast_fp16")]; + tensor var_386_promoted_to_fp16 = const()[name = tensor("op_386_promoted_to_fp16"), val = tensor(0x1p+0)]; + tensor var_387_cast_fp16 = add(x = var_375_cast_fp16_1, y = var_386_promoted_to_fp16)[name = tensor("op_387_cast_fp16")]; + tensor var_388_cast_fp16 = mul(x = x_43_cast_fp16, y = var_387_cast_fp16)[name = tensor("op_388_cast_fp16")]; + tensor input_65_cast_fp16 = add(x = var_388_cast_fp16, y = var_375_cast_fp16_0)[name = tensor("input_65_cast_fp16")]; + tensor flow_net_final_layer_linear_weight_to_fp16 = const()[name = tensor("flow_net_final_layer_linear_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(18964416)))]; + tensor flow_net_final_layer_linear_bias_to_fp16 = const()[name = tensor("flow_net_final_layer_linear_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(18997248)))]; + tensor linear_25_cast_fp16 = linear(bias = flow_net_final_layer_linear_bias_to_fp16, weight = flow_net_final_layer_linear_weight_to_fp16, x = input_65_cast_fp16)[name = tensor("linear_25_cast_fp16")]; + tensor var_399_to_fp16 = const()[name = tensor("op_399_to_fp16"), val = tensor(0x1p-3)]; + tensor var_400_cast_fp16 = mul(x = linear_25_cast_fp16, y = var_399_to_fp16)[name = tensor("op_400_cast_fp16")]; + tensor input_67_cast_fp16 = add(x = latent_init_to_fp16, y = var_400_cast_fp16)[name = tensor("input_67_cast_fp16")]; + tensor linear_26_cast_fp16 = linear(bias = flow_net_input_proj_bias_to_fp16, weight = flow_net_input_proj_weight_to_fp16, x = input_67_cast_fp16)[name = tensor("linear_26_cast_fp16")]; + tensor input_71_to_fp16 = const()[name = tensor("input_71_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(18997376)))]; + tensor input_73_cast_fp16 = silu(x = input_71_to_fp16)[name = tensor("input_73_cast_fp16")]; + tensor linear_28_cast_fp16 = linear(bias = flow_net_time_embed_0_mlp_2_bias_to_fp16, weight = flow_net_time_embed_0_mlp_2_weight_to_fp16, x = input_73_cast_fp16)[name = tensor("linear_28_cast_fp16")]; + tensor reduce_mean_18_axes_0 = const()[name = tensor("reduce_mean_18_axes_0"), val = tensor([-1])]; + tensor reduce_mean_18_keep_dims_0 = const()[name = tensor("reduce_mean_18_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_18_cast_fp16 = reduce_mean(axes = reduce_mean_18_axes_0, keep_dims = reduce_mean_18_keep_dims_0, x = linear_28_cast_fp16)[name = tensor("reduce_mean_18_cast_fp16")]; + tensor sub_11_cast_fp16 = sub(x = linear_28_cast_fp16, y = reduce_mean_18_cast_fp16)[name = tensor("sub_11_cast_fp16")]; + tensor square_9_cast_fp16 = square(x = sub_11_cast_fp16)[name = tensor("square_9_cast_fp16")]; + tensor reduce_mean_19_axes_0 = const()[name = tensor("reduce_mean_19_axes_0"), val = tensor([-1])]; + tensor reduce_mean_19_keep_dims_0 = const()[name = tensor("reduce_mean_19_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_19_cast_fp16 = reduce_mean(axes = reduce_mean_19_axes_0, keep_dims = reduce_mean_19_keep_dims_0, x = square_9_cast_fp16)[name = tensor("reduce_mean_19_cast_fp16")]; + tensor real_div_2_to_fp16 = const()[name = tensor("real_div_2_to_fp16"), val = tensor(0x1.008p+0)]; + tensor mul_2_cast_fp16 = mul(x = reduce_mean_19_cast_fp16, y = real_div_2_to_fp16)[name = tensor("mul_2_cast_fp16")]; + tensor var_466_to_fp16 = const()[name = tensor("op_466_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_19_cast_fp16 = add(x = mul_2_cast_fp16, y = var_466_to_fp16)[name = tensor("var_19_cast_fp16")]; + tensor var_469_epsilon_0 = const()[name = tensor("op_469_epsilon_0"), val = tensor(0x1.197998p-40)]; + tensor var_469_cast_fp16 = rsqrt(epsilon = var_469_epsilon_0, x = var_19_cast_fp16)[name = tensor("op_469_cast_fp16")]; + tensor var_470_cast_fp16 = mul(x = const_3_to_fp16, y = var_469_cast_fp16)[name = tensor("op_470_cast_fp16")]; + tensor var_471_cast_fp16 = mul(x = linear_28_cast_fp16, y = var_470_cast_fp16)[name = tensor("op_471_cast_fp16")]; + tensor input_77_to_fp16 = const()[name = tensor("input_77_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(18998464)))]; + tensor input_79_cast_fp16 = silu(x = input_77_to_fp16)[name = tensor("input_79_cast_fp16")]; + tensor linear_30_cast_fp16 = linear(bias = flow_net_time_embed_1_mlp_2_bias_to_fp16, weight = flow_net_time_embed_1_mlp_2_weight_to_fp16, x = input_79_cast_fp16)[name = tensor("linear_30_cast_fp16")]; + tensor reduce_mean_20_axes_0 = const()[name = tensor("reduce_mean_20_axes_0"), val = tensor([-1])]; + tensor reduce_mean_20_keep_dims_0 = const()[name = tensor("reduce_mean_20_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_20_cast_fp16 = reduce_mean(axes = reduce_mean_20_axes_0, keep_dims = reduce_mean_20_keep_dims_0, x = linear_30_cast_fp16)[name = tensor("reduce_mean_20_cast_fp16")]; + tensor sub_13_cast_fp16 = sub(x = linear_30_cast_fp16, y = reduce_mean_20_cast_fp16)[name = tensor("sub_13_cast_fp16")]; + tensor square_10_cast_fp16 = square(x = sub_13_cast_fp16)[name = tensor("square_10_cast_fp16")]; + tensor reduce_mean_21_axes_0 = const()[name = tensor("reduce_mean_21_axes_0"), val = tensor([-1])]; + tensor reduce_mean_21_keep_dims_0 = const()[name = tensor("reduce_mean_21_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_21_cast_fp16 = reduce_mean(axes = reduce_mean_21_axes_0, keep_dims = reduce_mean_21_keep_dims_0, x = square_10_cast_fp16)[name = tensor("reduce_mean_21_cast_fp16")]; + tensor real_div_3_to_fp16 = const()[name = tensor("real_div_3_to_fp16"), val = tensor(0x1.008p+0)]; + tensor mul_3_cast_fp16 = mul(x = reduce_mean_21_cast_fp16, y = real_div_3_to_fp16)[name = tensor("mul_3_cast_fp16")]; + tensor var_503_to_fp16 = const()[name = tensor("op_503_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_21_cast_fp16 = add(x = mul_3_cast_fp16, y = var_503_to_fp16)[name = tensor("var_21_cast_fp16")]; + tensor var_506_epsilon_0 = const()[name = tensor("op_506_epsilon_0"), val = tensor(0x1.197998p-40)]; + tensor var_506_cast_fp16 = rsqrt(epsilon = var_506_epsilon_0, x = var_21_cast_fp16)[name = tensor("op_506_cast_fp16")]; + tensor var_507_cast_fp16 = mul(x = const_5_to_fp16, y = var_506_cast_fp16)[name = tensor("op_507_cast_fp16")]; + tensor var_508_cast_fp16 = mul(x = linear_30_cast_fp16, y = var_507_cast_fp16)[name = tensor("op_508_cast_fp16")]; + tensor var_520_cast_fp16 = add(x = var_471_cast_fp16, y = var_508_cast_fp16)[name = tensor("op_520_cast_fp16")]; + tensor _inversed_t_combined_3_y_0_to_fp16 = const()[name = tensor("_inversed_t_combined_3_y_0_to_fp16"), val = tensor(0x1p-1)]; + tensor _inversed_t_combined_3_cast_fp16 = mul(x = var_520_cast_fp16, y = _inversed_t_combined_3_y_0_to_fp16)[name = tensor("_inversed_t_combined_3_cast_fp16")]; + tensor input_81_cast_fp16 = add(x = _inversed_t_combined_3_cast_fp16, y = linear_5_cast_fp16)[name = tensor("input_81_cast_fp16")]; + tensor input_83_cast_fp16 = silu(x = input_81_cast_fp16)[name = tensor("input_83_cast_fp16")]; + tensor linear_32_cast_fp16 = linear(bias = flow_net_res_blocks_0_adaLN_modulation_1_bias_to_fp16, weight = flow_net_res_blocks_0_adaLN_modulation_1_weight_to_fp16, x = input_83_cast_fp16)[name = tensor("linear_32_cast_fp16")]; + tensor var_535_split_sizes_0 = const()[name = tensor("op_535_split_sizes_0"), val = tensor([512, 512, 512])]; + tensor var_535_axis_0 = const()[name = tensor("op_535_axis_0"), val = tensor(-1)]; + tensor var_535_cast_fp16_0, tensor var_535_cast_fp16_1, tensor var_535_cast_fp16_2 = split(axis = var_535_axis_0, split_sizes = var_535_split_sizes_0, x = linear_32_cast_fp16)[name = tensor("op_535_cast_fp16")]; + tensor mean_15_axes_0 = const()[name = tensor("mean_15_axes_0"), val = tensor([-1])]; + tensor mean_15_keep_dims_0 = const()[name = tensor("mean_15_keep_dims_0"), val = tensor(true)]; + tensor mean_15_cast_fp16 = reduce_mean(axes = mean_15_axes_0, keep_dims = mean_15_keep_dims_0, x = linear_26_cast_fp16)[name = tensor("mean_15_cast_fp16")]; + tensor sub_15_cast_fp16 = sub(x = linear_26_cast_fp16, y = mean_15_cast_fp16)[name = tensor("sub_15_cast_fp16")]; + tensor square_11_cast_fp16 = square(x = sub_15_cast_fp16)[name = tensor("square_11_cast_fp16")]; + tensor reduce_mean_23_axes_0 = const()[name = tensor("reduce_mean_23_axes_0"), val = tensor([-1])]; + tensor reduce_mean_23_keep_dims_0 = const()[name = tensor("reduce_mean_23_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_23_cast_fp16 = reduce_mean(axes = reduce_mean_23_axes_0, keep_dims = reduce_mean_23_keep_dims_0, x = square_11_cast_fp16)[name = tensor("reduce_mean_23_cast_fp16")]; + tensor var_545_to_fp16 = const()[name = tensor("op_545_to_fp16"), val = tensor(0x1.1p-20)]; + tensor var_546_cast_fp16 = add(x = reduce_mean_23_cast_fp16, y = var_545_to_fp16)[name = tensor("op_546_cast_fp16")]; + tensor var_547_cast_fp16 = sqrt(x = var_546_cast_fp16)[name = tensor("op_547_cast_fp16")]; + tensor x_51_cast_fp16 = real_div(x = sub_15_cast_fp16, y = var_547_cast_fp16)[name = tensor("x_51_cast_fp16")]; + tensor var_549_cast_fp16 = mul(x = x_51_cast_fp16, y = flow_net_res_blocks_0_in_ln_weight_to_fp16)[name = tensor("op_549_cast_fp16")]; + tensor x_53_cast_fp16 = add(x = var_549_cast_fp16, y = flow_net_res_blocks_0_in_ln_bias_to_fp16)[name = tensor("x_53_cast_fp16")]; + tensor var_551_promoted_to_fp16 = const()[name = tensor("op_551_promoted_to_fp16"), val = tensor(0x1p+0)]; + tensor var_552_cast_fp16 = add(x = var_535_cast_fp16_1, y = var_551_promoted_to_fp16)[name = tensor("op_552_cast_fp16")]; + tensor var_553_cast_fp16 = mul(x = x_53_cast_fp16, y = var_552_cast_fp16)[name = tensor("op_553_cast_fp16")]; + tensor input_85_cast_fp16 = add(x = var_553_cast_fp16, y = var_535_cast_fp16_0)[name = tensor("input_85_cast_fp16")]; + tensor linear_33_cast_fp16 = linear(bias = flow_net_res_blocks_0_mlp_0_bias_to_fp16, weight = flow_net_res_blocks_0_mlp_0_weight_to_fp16, x = input_85_cast_fp16)[name = tensor("linear_33_cast_fp16")]; + tensor input_89_cast_fp16 = silu(x = linear_33_cast_fp16)[name = tensor("input_89_cast_fp16")]; + tensor linear_34_cast_fp16 = linear(bias = flow_net_res_blocks_0_mlp_2_bias_to_fp16, weight = flow_net_res_blocks_0_mlp_2_weight_to_fp16, x = input_89_cast_fp16)[name = tensor("linear_34_cast_fp16")]; + tensor var_564_cast_fp16 = mul(x = var_535_cast_fp16_2, y = linear_34_cast_fp16)[name = tensor("op_564_cast_fp16")]; + tensor x_55_cast_fp16 = add(x = linear_26_cast_fp16, y = var_564_cast_fp16)[name = tensor("x_55_cast_fp16")]; + tensor linear_35_cast_fp16 = linear(bias = flow_net_res_blocks_1_adaLN_modulation_1_bias_to_fp16, weight = flow_net_res_blocks_1_adaLN_modulation_1_weight_to_fp16, x = input_83_cast_fp16)[name = tensor("linear_35_cast_fp16")]; + tensor var_574_split_sizes_0 = const()[name = tensor("op_574_split_sizes_0"), val = tensor([512, 512, 512])]; + tensor var_574_axis_0 = const()[name = tensor("op_574_axis_0"), val = tensor(-1)]; + tensor var_574_cast_fp16_0, tensor var_574_cast_fp16_1, tensor var_574_cast_fp16_2 = split(axis = var_574_axis_0, split_sizes = var_574_split_sizes_0, x = linear_35_cast_fp16)[name = tensor("op_574_cast_fp16")]; + tensor mean_17_axes_0 = const()[name = tensor("mean_17_axes_0"), val = tensor([-1])]; + tensor mean_17_keep_dims_0 = const()[name = tensor("mean_17_keep_dims_0"), val = tensor(true)]; + tensor mean_17_cast_fp16 = reduce_mean(axes = mean_17_axes_0, keep_dims = mean_17_keep_dims_0, x = x_55_cast_fp16)[name = tensor("mean_17_cast_fp16")]; + tensor sub_16_cast_fp16 = sub(x = x_55_cast_fp16, y = mean_17_cast_fp16)[name = tensor("sub_16_cast_fp16")]; + tensor square_12_cast_fp16 = square(x = sub_16_cast_fp16)[name = tensor("square_12_cast_fp16")]; + tensor reduce_mean_25_axes_0 = const()[name = tensor("reduce_mean_25_axes_0"), val = tensor([-1])]; + tensor reduce_mean_25_keep_dims_0 = const()[name = tensor("reduce_mean_25_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_25_cast_fp16 = reduce_mean(axes = reduce_mean_25_axes_0, keep_dims = reduce_mean_25_keep_dims_0, x = square_12_cast_fp16)[name = tensor("reduce_mean_25_cast_fp16")]; + tensor var_584_to_fp16 = const()[name = tensor("op_584_to_fp16"), val = tensor(0x1.1p-20)]; + tensor var_585_cast_fp16 = add(x = reduce_mean_25_cast_fp16, y = var_584_to_fp16)[name = tensor("op_585_cast_fp16")]; + tensor var_586_cast_fp16 = sqrt(x = var_585_cast_fp16)[name = tensor("op_586_cast_fp16")]; + tensor x_57_cast_fp16 = real_div(x = sub_16_cast_fp16, y = var_586_cast_fp16)[name = tensor("x_57_cast_fp16")]; + tensor var_588_cast_fp16 = mul(x = x_57_cast_fp16, y = flow_net_res_blocks_1_in_ln_weight_to_fp16)[name = tensor("op_588_cast_fp16")]; + tensor x_59_cast_fp16 = add(x = var_588_cast_fp16, y = flow_net_res_blocks_1_in_ln_bias_to_fp16)[name = tensor("x_59_cast_fp16")]; + tensor var_590_promoted_to_fp16 = const()[name = tensor("op_590_promoted_to_fp16"), val = tensor(0x1p+0)]; + tensor var_591_cast_fp16 = add(x = var_574_cast_fp16_1, y = var_590_promoted_to_fp16)[name = tensor("op_591_cast_fp16")]; + tensor var_592_cast_fp16 = mul(x = x_59_cast_fp16, y = var_591_cast_fp16)[name = tensor("op_592_cast_fp16")]; + tensor input_93_cast_fp16 = add(x = var_592_cast_fp16, y = var_574_cast_fp16_0)[name = tensor("input_93_cast_fp16")]; + tensor linear_36_cast_fp16 = linear(bias = flow_net_res_blocks_1_mlp_0_bias_to_fp16, weight = flow_net_res_blocks_1_mlp_0_weight_to_fp16, x = input_93_cast_fp16)[name = tensor("linear_36_cast_fp16")]; + tensor input_97_cast_fp16 = silu(x = linear_36_cast_fp16)[name = tensor("input_97_cast_fp16")]; + tensor linear_37_cast_fp16 = linear(bias = flow_net_res_blocks_1_mlp_2_bias_to_fp16, weight = flow_net_res_blocks_1_mlp_2_weight_to_fp16, x = input_97_cast_fp16)[name = tensor("linear_37_cast_fp16")]; + tensor var_603_cast_fp16 = mul(x = var_574_cast_fp16_2, y = linear_37_cast_fp16)[name = tensor("op_603_cast_fp16")]; + tensor x_61_cast_fp16 = add(x = x_55_cast_fp16, y = var_603_cast_fp16)[name = tensor("x_61_cast_fp16")]; + tensor linear_38_cast_fp16 = linear(bias = flow_net_res_blocks_2_adaLN_modulation_1_bias_to_fp16, weight = flow_net_res_blocks_2_adaLN_modulation_1_weight_to_fp16, x = input_83_cast_fp16)[name = tensor("linear_38_cast_fp16")]; + tensor var_613_split_sizes_0 = const()[name = tensor("op_613_split_sizes_0"), val = tensor([512, 512, 512])]; + tensor var_613_axis_0 = const()[name = tensor("op_613_axis_0"), val = tensor(-1)]; + tensor var_613_cast_fp16_0, tensor var_613_cast_fp16_1, tensor var_613_cast_fp16_2 = split(axis = var_613_axis_0, split_sizes = var_613_split_sizes_0, x = linear_38_cast_fp16)[name = tensor("op_613_cast_fp16")]; + tensor mean_19_axes_0 = const()[name = tensor("mean_19_axes_0"), val = tensor([-1])]; + tensor mean_19_keep_dims_0 = const()[name = tensor("mean_19_keep_dims_0"), val = tensor(true)]; + tensor mean_19_cast_fp16 = reduce_mean(axes = mean_19_axes_0, keep_dims = mean_19_keep_dims_0, x = x_61_cast_fp16)[name = tensor("mean_19_cast_fp16")]; + tensor sub_17_cast_fp16 = sub(x = x_61_cast_fp16, y = mean_19_cast_fp16)[name = tensor("sub_17_cast_fp16")]; + tensor square_13_cast_fp16 = square(x = sub_17_cast_fp16)[name = tensor("square_13_cast_fp16")]; + tensor reduce_mean_27_axes_0 = const()[name = tensor("reduce_mean_27_axes_0"), val = tensor([-1])]; + tensor reduce_mean_27_keep_dims_0 = const()[name = tensor("reduce_mean_27_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_27_cast_fp16 = reduce_mean(axes = reduce_mean_27_axes_0, keep_dims = reduce_mean_27_keep_dims_0, x = square_13_cast_fp16)[name = tensor("reduce_mean_27_cast_fp16")]; + tensor var_623_to_fp16 = const()[name = tensor("op_623_to_fp16"), val = tensor(0x1.1p-20)]; + tensor var_624_cast_fp16 = add(x = reduce_mean_27_cast_fp16, y = var_623_to_fp16)[name = tensor("op_624_cast_fp16")]; + tensor var_625_cast_fp16 = sqrt(x = var_624_cast_fp16)[name = tensor("op_625_cast_fp16")]; + tensor x_63_cast_fp16 = real_div(x = sub_17_cast_fp16, y = var_625_cast_fp16)[name = tensor("x_63_cast_fp16")]; + tensor var_627_cast_fp16 = mul(x = x_63_cast_fp16, y = flow_net_res_blocks_2_in_ln_weight_to_fp16)[name = tensor("op_627_cast_fp16")]; + tensor x_65_cast_fp16 = add(x = var_627_cast_fp16, y = flow_net_res_blocks_2_in_ln_bias_to_fp16)[name = tensor("x_65_cast_fp16")]; + tensor var_629_promoted_to_fp16 = const()[name = tensor("op_629_promoted_to_fp16"), val = tensor(0x1p+0)]; + tensor var_630_cast_fp16 = add(x = var_613_cast_fp16_1, y = var_629_promoted_to_fp16)[name = tensor("op_630_cast_fp16")]; + tensor var_631_cast_fp16 = mul(x = x_65_cast_fp16, y = var_630_cast_fp16)[name = tensor("op_631_cast_fp16")]; + tensor input_101_cast_fp16 = add(x = var_631_cast_fp16, y = var_613_cast_fp16_0)[name = tensor("input_101_cast_fp16")]; + tensor linear_39_cast_fp16 = linear(bias = flow_net_res_blocks_2_mlp_0_bias_to_fp16, weight = flow_net_res_blocks_2_mlp_0_weight_to_fp16, x = input_101_cast_fp16)[name = tensor("linear_39_cast_fp16")]; + tensor input_105_cast_fp16 = silu(x = linear_39_cast_fp16)[name = tensor("input_105_cast_fp16")]; + tensor linear_40_cast_fp16 = linear(bias = flow_net_res_blocks_2_mlp_2_bias_to_fp16, weight = flow_net_res_blocks_2_mlp_2_weight_to_fp16, x = input_105_cast_fp16)[name = tensor("linear_40_cast_fp16")]; + tensor var_642_cast_fp16 = mul(x = var_613_cast_fp16_2, y = linear_40_cast_fp16)[name = tensor("op_642_cast_fp16")]; + tensor x_67_cast_fp16 = add(x = x_61_cast_fp16, y = var_642_cast_fp16)[name = tensor("x_67_cast_fp16")]; + tensor linear_41_cast_fp16 = linear(bias = flow_net_res_blocks_3_adaLN_modulation_1_bias_to_fp16, weight = flow_net_res_blocks_3_adaLN_modulation_1_weight_to_fp16, x = input_83_cast_fp16)[name = tensor("linear_41_cast_fp16")]; + tensor var_652_split_sizes_0 = const()[name = tensor("op_652_split_sizes_0"), val = tensor([512, 512, 512])]; + tensor var_652_axis_0 = const()[name = tensor("op_652_axis_0"), val = tensor(-1)]; + tensor var_652_cast_fp16_0, tensor var_652_cast_fp16_1, tensor var_652_cast_fp16_2 = split(axis = var_652_axis_0, split_sizes = var_652_split_sizes_0, x = linear_41_cast_fp16)[name = tensor("op_652_cast_fp16")]; + tensor mean_21_axes_0 = const()[name = tensor("mean_21_axes_0"), val = tensor([-1])]; + tensor mean_21_keep_dims_0 = const()[name = tensor("mean_21_keep_dims_0"), val = tensor(true)]; + tensor mean_21_cast_fp16 = reduce_mean(axes = mean_21_axes_0, keep_dims = mean_21_keep_dims_0, x = x_67_cast_fp16)[name = tensor("mean_21_cast_fp16")]; + tensor sub_18_cast_fp16 = sub(x = x_67_cast_fp16, y = mean_21_cast_fp16)[name = tensor("sub_18_cast_fp16")]; + tensor square_14_cast_fp16 = square(x = sub_18_cast_fp16)[name = tensor("square_14_cast_fp16")]; + tensor reduce_mean_29_axes_0 = const()[name = tensor("reduce_mean_29_axes_0"), val = tensor([-1])]; + tensor reduce_mean_29_keep_dims_0 = const()[name = tensor("reduce_mean_29_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_29_cast_fp16 = reduce_mean(axes = reduce_mean_29_axes_0, keep_dims = reduce_mean_29_keep_dims_0, x = square_14_cast_fp16)[name = tensor("reduce_mean_29_cast_fp16")]; + tensor var_662_to_fp16 = const()[name = tensor("op_662_to_fp16"), val = tensor(0x1.1p-20)]; + tensor var_663_cast_fp16 = add(x = reduce_mean_29_cast_fp16, y = var_662_to_fp16)[name = tensor("op_663_cast_fp16")]; + tensor var_664_cast_fp16 = sqrt(x = var_663_cast_fp16)[name = tensor("op_664_cast_fp16")]; + tensor x_69_cast_fp16 = real_div(x = sub_18_cast_fp16, y = var_664_cast_fp16)[name = tensor("x_69_cast_fp16")]; + tensor var_666_cast_fp16 = mul(x = x_69_cast_fp16, y = flow_net_res_blocks_3_in_ln_weight_to_fp16)[name = tensor("op_666_cast_fp16")]; + tensor x_71_cast_fp16 = add(x = var_666_cast_fp16, y = flow_net_res_blocks_3_in_ln_bias_to_fp16)[name = tensor("x_71_cast_fp16")]; + tensor var_668_promoted_to_fp16 = const()[name = tensor("op_668_promoted_to_fp16"), val = tensor(0x1p+0)]; + tensor var_669_cast_fp16 = add(x = var_652_cast_fp16_1, y = var_668_promoted_to_fp16)[name = tensor("op_669_cast_fp16")]; + tensor var_670_cast_fp16 = mul(x = x_71_cast_fp16, y = var_669_cast_fp16)[name = tensor("op_670_cast_fp16")]; + tensor input_109_cast_fp16 = add(x = var_670_cast_fp16, y = var_652_cast_fp16_0)[name = tensor("input_109_cast_fp16")]; + tensor linear_42_cast_fp16 = linear(bias = flow_net_res_blocks_3_mlp_0_bias_to_fp16, weight = flow_net_res_blocks_3_mlp_0_weight_to_fp16, x = input_109_cast_fp16)[name = tensor("linear_42_cast_fp16")]; + tensor input_113_cast_fp16 = silu(x = linear_42_cast_fp16)[name = tensor("input_113_cast_fp16")]; + tensor linear_43_cast_fp16 = linear(bias = flow_net_res_blocks_3_mlp_2_bias_to_fp16, weight = flow_net_res_blocks_3_mlp_2_weight_to_fp16, x = input_113_cast_fp16)[name = tensor("linear_43_cast_fp16")]; + tensor var_681_cast_fp16 = mul(x = var_652_cast_fp16_2, y = linear_43_cast_fp16)[name = tensor("op_681_cast_fp16")]; + tensor x_73_cast_fp16 = add(x = x_67_cast_fp16, y = var_681_cast_fp16)[name = tensor("x_73_cast_fp16")]; + tensor linear_44_cast_fp16 = linear(bias = flow_net_res_blocks_4_adaLN_modulation_1_bias_to_fp16, weight = flow_net_res_blocks_4_adaLN_modulation_1_weight_to_fp16, x = input_83_cast_fp16)[name = tensor("linear_44_cast_fp16")]; + tensor var_691_split_sizes_0 = const()[name = tensor("op_691_split_sizes_0"), val = tensor([512, 512, 512])]; + tensor var_691_axis_0 = const()[name = tensor("op_691_axis_0"), val = tensor(-1)]; + tensor var_691_cast_fp16_0, tensor var_691_cast_fp16_1, tensor var_691_cast_fp16_2 = split(axis = var_691_axis_0, split_sizes = var_691_split_sizes_0, x = linear_44_cast_fp16)[name = tensor("op_691_cast_fp16")]; + tensor mean_23_axes_0 = const()[name = tensor("mean_23_axes_0"), val = tensor([-1])]; + tensor mean_23_keep_dims_0 = const()[name = tensor("mean_23_keep_dims_0"), val = tensor(true)]; + tensor mean_23_cast_fp16 = reduce_mean(axes = mean_23_axes_0, keep_dims = mean_23_keep_dims_0, x = x_73_cast_fp16)[name = tensor("mean_23_cast_fp16")]; + tensor sub_19_cast_fp16 = sub(x = x_73_cast_fp16, y = mean_23_cast_fp16)[name = tensor("sub_19_cast_fp16")]; + tensor square_15_cast_fp16 = square(x = sub_19_cast_fp16)[name = tensor("square_15_cast_fp16")]; + tensor reduce_mean_31_axes_0 = const()[name = tensor("reduce_mean_31_axes_0"), val = tensor([-1])]; + tensor reduce_mean_31_keep_dims_0 = const()[name = tensor("reduce_mean_31_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_31_cast_fp16 = reduce_mean(axes = reduce_mean_31_axes_0, keep_dims = reduce_mean_31_keep_dims_0, x = square_15_cast_fp16)[name = tensor("reduce_mean_31_cast_fp16")]; + tensor var_701_to_fp16 = const()[name = tensor("op_701_to_fp16"), val = tensor(0x1.1p-20)]; + tensor var_702_cast_fp16 = add(x = reduce_mean_31_cast_fp16, y = var_701_to_fp16)[name = tensor("op_702_cast_fp16")]; + tensor var_703_cast_fp16 = sqrt(x = var_702_cast_fp16)[name = tensor("op_703_cast_fp16")]; + tensor x_75_cast_fp16 = real_div(x = sub_19_cast_fp16, y = var_703_cast_fp16)[name = tensor("x_75_cast_fp16")]; + tensor var_705_cast_fp16 = mul(x = x_75_cast_fp16, y = flow_net_res_blocks_4_in_ln_weight_to_fp16)[name = tensor("op_705_cast_fp16")]; + tensor x_77_cast_fp16 = add(x = var_705_cast_fp16, y = flow_net_res_blocks_4_in_ln_bias_to_fp16)[name = tensor("x_77_cast_fp16")]; + tensor var_707_promoted_to_fp16 = const()[name = tensor("op_707_promoted_to_fp16"), val = tensor(0x1p+0)]; + tensor var_708_cast_fp16 = add(x = var_691_cast_fp16_1, y = var_707_promoted_to_fp16)[name = tensor("op_708_cast_fp16")]; + tensor var_709_cast_fp16 = mul(x = x_77_cast_fp16, y = var_708_cast_fp16)[name = tensor("op_709_cast_fp16")]; + tensor input_117_cast_fp16 = add(x = var_709_cast_fp16, y = var_691_cast_fp16_0)[name = tensor("input_117_cast_fp16")]; + tensor linear_45_cast_fp16 = linear(bias = flow_net_res_blocks_4_mlp_0_bias_to_fp16, weight = flow_net_res_blocks_4_mlp_0_weight_to_fp16, x = input_117_cast_fp16)[name = tensor("linear_45_cast_fp16")]; + tensor input_121_cast_fp16 = silu(x = linear_45_cast_fp16)[name = tensor("input_121_cast_fp16")]; + tensor linear_46_cast_fp16 = linear(bias = flow_net_res_blocks_4_mlp_2_bias_to_fp16, weight = flow_net_res_blocks_4_mlp_2_weight_to_fp16, x = input_121_cast_fp16)[name = tensor("linear_46_cast_fp16")]; + tensor var_720_cast_fp16 = mul(x = var_691_cast_fp16_2, y = linear_46_cast_fp16)[name = tensor("op_720_cast_fp16")]; + tensor x_79_cast_fp16 = add(x = x_73_cast_fp16, y = var_720_cast_fp16)[name = tensor("x_79_cast_fp16")]; + tensor linear_47_cast_fp16 = linear(bias = flow_net_res_blocks_5_adaLN_modulation_1_bias_to_fp16, weight = flow_net_res_blocks_5_adaLN_modulation_1_weight_to_fp16, x = input_83_cast_fp16)[name = tensor("linear_47_cast_fp16")]; + tensor var_730_split_sizes_0 = const()[name = tensor("op_730_split_sizes_0"), val = tensor([512, 512, 512])]; + tensor var_730_axis_0 = const()[name = tensor("op_730_axis_0"), val = tensor(-1)]; + tensor var_730_cast_fp16_0, tensor var_730_cast_fp16_1, tensor var_730_cast_fp16_2 = split(axis = var_730_axis_0, split_sizes = var_730_split_sizes_0, x = linear_47_cast_fp16)[name = tensor("op_730_cast_fp16")]; + tensor mean_25_axes_0 = const()[name = tensor("mean_25_axes_0"), val = tensor([-1])]; + tensor mean_25_keep_dims_0 = const()[name = tensor("mean_25_keep_dims_0"), val = tensor(true)]; + tensor mean_25_cast_fp16 = reduce_mean(axes = mean_25_axes_0, keep_dims = mean_25_keep_dims_0, x = x_79_cast_fp16)[name = tensor("mean_25_cast_fp16")]; + tensor sub_20_cast_fp16 = sub(x = x_79_cast_fp16, y = mean_25_cast_fp16)[name = tensor("sub_20_cast_fp16")]; + tensor square_16_cast_fp16 = square(x = sub_20_cast_fp16)[name = tensor("square_16_cast_fp16")]; + tensor reduce_mean_33_axes_0 = const()[name = tensor("reduce_mean_33_axes_0"), val = tensor([-1])]; + tensor reduce_mean_33_keep_dims_0 = const()[name = tensor("reduce_mean_33_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_33_cast_fp16 = reduce_mean(axes = reduce_mean_33_axes_0, keep_dims = reduce_mean_33_keep_dims_0, x = square_16_cast_fp16)[name = tensor("reduce_mean_33_cast_fp16")]; + tensor var_740_to_fp16 = const()[name = tensor("op_740_to_fp16"), val = tensor(0x1.1p-20)]; + tensor var_741_cast_fp16 = add(x = reduce_mean_33_cast_fp16, y = var_740_to_fp16)[name = tensor("op_741_cast_fp16")]; + tensor var_742_cast_fp16 = sqrt(x = var_741_cast_fp16)[name = tensor("op_742_cast_fp16")]; + tensor x_81_cast_fp16 = real_div(x = sub_20_cast_fp16, y = var_742_cast_fp16)[name = tensor("x_81_cast_fp16")]; + tensor var_744_cast_fp16 = mul(x = x_81_cast_fp16, y = flow_net_res_blocks_5_in_ln_weight_to_fp16)[name = tensor("op_744_cast_fp16")]; + tensor x_83_cast_fp16 = add(x = var_744_cast_fp16, y = flow_net_res_blocks_5_in_ln_bias_to_fp16)[name = tensor("x_83_cast_fp16")]; + tensor var_746_promoted_to_fp16 = const()[name = tensor("op_746_promoted_to_fp16"), val = tensor(0x1p+0)]; + tensor var_747_cast_fp16 = add(x = var_730_cast_fp16_1, y = var_746_promoted_to_fp16)[name = tensor("op_747_cast_fp16")]; + tensor var_748_cast_fp16 = mul(x = x_83_cast_fp16, y = var_747_cast_fp16)[name = tensor("op_748_cast_fp16")]; + tensor input_125_cast_fp16 = add(x = var_748_cast_fp16, y = var_730_cast_fp16_0)[name = tensor("input_125_cast_fp16")]; + tensor linear_48_cast_fp16 = linear(bias = flow_net_res_blocks_5_mlp_0_bias_to_fp16, weight = flow_net_res_blocks_5_mlp_0_weight_to_fp16, x = input_125_cast_fp16)[name = tensor("linear_48_cast_fp16")]; + tensor input_129_cast_fp16 = silu(x = linear_48_cast_fp16)[name = tensor("input_129_cast_fp16")]; + tensor linear_49_cast_fp16 = linear(bias = flow_net_res_blocks_5_mlp_2_bias_to_fp16, weight = flow_net_res_blocks_5_mlp_2_weight_to_fp16, x = input_129_cast_fp16)[name = tensor("linear_49_cast_fp16")]; + tensor var_759_cast_fp16 = mul(x = var_730_cast_fp16_2, y = linear_49_cast_fp16)[name = tensor("op_759_cast_fp16")]; + tensor x_85_cast_fp16 = add(x = x_79_cast_fp16, y = var_759_cast_fp16)[name = tensor("x_85_cast_fp16")]; + tensor linear_50_cast_fp16 = linear(bias = flow_net_final_layer_adaLN_modulation_1_bias_to_fp16, weight = flow_net_final_layer_adaLN_modulation_1_weight_to_fp16, x = input_83_cast_fp16)[name = tensor("linear_50_cast_fp16")]; + tensor var_768_split_sizes_0 = const()[name = tensor("op_768_split_sizes_0"), val = tensor([512, 512])]; + tensor var_768_axis_0 = const()[name = tensor("op_768_axis_0"), val = tensor(-1)]; + tensor var_768_cast_fp16_0, tensor var_768_cast_fp16_1 = split(axis = var_768_axis_0, split_sizes = var_768_split_sizes_0, x = linear_50_cast_fp16)[name = tensor("op_768_cast_fp16")]; + tensor mean_27_axes_0 = const()[name = tensor("mean_27_axes_0"), val = tensor([-1])]; + tensor mean_27_keep_dims_0 = const()[name = tensor("mean_27_keep_dims_0"), val = tensor(true)]; + tensor mean_27_cast_fp16 = reduce_mean(axes = mean_27_axes_0, keep_dims = mean_27_keep_dims_0, x = x_85_cast_fp16)[name = tensor("mean_27_cast_fp16")]; + tensor sub_21_cast_fp16 = sub(x = x_85_cast_fp16, y = mean_27_cast_fp16)[name = tensor("sub_21_cast_fp16")]; + tensor square_17_cast_fp16 = square(x = sub_21_cast_fp16)[name = tensor("square_17_cast_fp16")]; + tensor reduce_mean_35_axes_0 = const()[name = tensor("reduce_mean_35_axes_0"), val = tensor([-1])]; + tensor reduce_mean_35_keep_dims_0 = const()[name = tensor("reduce_mean_35_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_35_cast_fp16 = reduce_mean(axes = reduce_mean_35_axes_0, keep_dims = reduce_mean_35_keep_dims_0, x = square_17_cast_fp16)[name = tensor("reduce_mean_35_cast_fp16")]; + tensor var_775_to_fp16 = const()[name = tensor("op_775_to_fp16"), val = tensor(0x1.1p-20)]; + tensor var_776_cast_fp16 = add(x = reduce_mean_35_cast_fp16, y = var_775_to_fp16)[name = tensor("op_776_cast_fp16")]; + tensor var_777_cast_fp16 = sqrt(x = var_776_cast_fp16)[name = tensor("op_777_cast_fp16")]; + tensor x_87_cast_fp16 = real_div(x = sub_21_cast_fp16, y = var_777_cast_fp16)[name = tensor("x_87_cast_fp16")]; + tensor var_779_promoted_to_fp16 = const()[name = tensor("op_779_promoted_to_fp16"), val = tensor(0x1p+0)]; + tensor var_780_cast_fp16 = add(x = var_768_cast_fp16_1, y = var_779_promoted_to_fp16)[name = tensor("op_780_cast_fp16")]; + tensor var_781_cast_fp16 = mul(x = x_87_cast_fp16, y = var_780_cast_fp16)[name = tensor("op_781_cast_fp16")]; + tensor input_133_cast_fp16 = add(x = var_781_cast_fp16, y = var_768_cast_fp16_0)[name = tensor("input_133_cast_fp16")]; + tensor linear_51_cast_fp16 = linear(bias = flow_net_final_layer_linear_bias_to_fp16, weight = flow_net_final_layer_linear_weight_to_fp16, x = input_133_cast_fp16)[name = tensor("linear_51_cast_fp16")]; + tensor var_792_to_fp16 = const()[name = tensor("op_792_to_fp16"), val = tensor(0x1p-3)]; + tensor var_793_cast_fp16 = mul(x = linear_51_cast_fp16, y = var_792_to_fp16)[name = tensor("op_793_cast_fp16")]; + tensor input_135_cast_fp16 = add(x = input_67_cast_fp16, y = var_793_cast_fp16)[name = tensor("input_135_cast_fp16")]; + tensor linear_52_cast_fp16 = linear(bias = flow_net_input_proj_bias_to_fp16, weight = flow_net_input_proj_weight_to_fp16, x = input_135_cast_fp16)[name = tensor("linear_52_cast_fp16")]; + tensor input_139_to_fp16 = const()[name = tensor("input_139_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(18999552)))]; + tensor input_141_cast_fp16 = silu(x = input_139_to_fp16)[name = tensor("input_141_cast_fp16")]; + tensor linear_54_cast_fp16 = linear(bias = flow_net_time_embed_0_mlp_2_bias_to_fp16, weight = flow_net_time_embed_0_mlp_2_weight_to_fp16, x = input_141_cast_fp16)[name = tensor("linear_54_cast_fp16")]; + tensor reduce_mean_36_axes_0 = const()[name = tensor("reduce_mean_36_axes_0"), val = tensor([-1])]; + tensor reduce_mean_36_keep_dims_0 = const()[name = tensor("reduce_mean_36_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_36_cast_fp16 = reduce_mean(axes = reduce_mean_36_axes_0, keep_dims = reduce_mean_36_keep_dims_0, x = linear_54_cast_fp16)[name = tensor("reduce_mean_36_cast_fp16")]; + tensor sub_22_cast_fp16 = sub(x = linear_54_cast_fp16, y = reduce_mean_36_cast_fp16)[name = tensor("sub_22_cast_fp16")]; + tensor square_18_cast_fp16 = square(x = sub_22_cast_fp16)[name = tensor("square_18_cast_fp16")]; + tensor reduce_mean_37_axes_0 = const()[name = tensor("reduce_mean_37_axes_0"), val = tensor([-1])]; + tensor reduce_mean_37_keep_dims_0 = const()[name = tensor("reduce_mean_37_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_37_cast_fp16 = reduce_mean(axes = reduce_mean_37_axes_0, keep_dims = reduce_mean_37_keep_dims_0, x = square_18_cast_fp16)[name = tensor("reduce_mean_37_cast_fp16")]; + tensor real_div_4_to_fp16 = const()[name = tensor("real_div_4_to_fp16"), val = tensor(0x1.008p+0)]; + tensor mul_4_cast_fp16 = mul(x = reduce_mean_37_cast_fp16, y = real_div_4_to_fp16)[name = tensor("mul_4_cast_fp16")]; + tensor var_859_to_fp16 = const()[name = tensor("op_859_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_37_cast_fp16 = add(x = mul_4_cast_fp16, y = var_859_to_fp16)[name = tensor("var_37_cast_fp16")]; + tensor var_862_epsilon_0 = const()[name = tensor("op_862_epsilon_0"), val = tensor(0x1.197998p-40)]; + tensor var_862_cast_fp16 = rsqrt(epsilon = var_862_epsilon_0, x = var_37_cast_fp16)[name = tensor("op_862_cast_fp16")]; + tensor var_863_cast_fp16 = mul(x = const_3_to_fp16, y = var_862_cast_fp16)[name = tensor("op_863_cast_fp16")]; + tensor var_864_cast_fp16 = mul(x = linear_54_cast_fp16, y = var_863_cast_fp16)[name = tensor("op_864_cast_fp16")]; + tensor input_145_to_fp16 = const()[name = tensor("input_145_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(19000640)))]; + tensor input_147_cast_fp16 = silu(x = input_145_to_fp16)[name = tensor("input_147_cast_fp16")]; + tensor linear_56_cast_fp16 = linear(bias = flow_net_time_embed_1_mlp_2_bias_to_fp16, weight = flow_net_time_embed_1_mlp_2_weight_to_fp16, x = input_147_cast_fp16)[name = tensor("linear_56_cast_fp16")]; + tensor reduce_mean_38_axes_0 = const()[name = tensor("reduce_mean_38_axes_0"), val = tensor([-1])]; + tensor reduce_mean_38_keep_dims_0 = const()[name = tensor("reduce_mean_38_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_38_cast_fp16 = reduce_mean(axes = reduce_mean_38_axes_0, keep_dims = reduce_mean_38_keep_dims_0, x = linear_56_cast_fp16)[name = tensor("reduce_mean_38_cast_fp16")]; + tensor sub_24_cast_fp16 = sub(x = linear_56_cast_fp16, y = reduce_mean_38_cast_fp16)[name = tensor("sub_24_cast_fp16")]; + tensor square_19_cast_fp16 = square(x = sub_24_cast_fp16)[name = tensor("square_19_cast_fp16")]; + tensor reduce_mean_39_axes_0 = const()[name = tensor("reduce_mean_39_axes_0"), val = tensor([-1])]; + tensor reduce_mean_39_keep_dims_0 = const()[name = tensor("reduce_mean_39_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_39_cast_fp16 = reduce_mean(axes = reduce_mean_39_axes_0, keep_dims = reduce_mean_39_keep_dims_0, x = square_19_cast_fp16)[name = tensor("reduce_mean_39_cast_fp16")]; + tensor real_div_5_to_fp16 = const()[name = tensor("real_div_5_to_fp16"), val = tensor(0x1.008p+0)]; + tensor mul_5_cast_fp16 = mul(x = reduce_mean_39_cast_fp16, y = real_div_5_to_fp16)[name = tensor("mul_5_cast_fp16")]; + tensor var_896_to_fp16 = const()[name = tensor("op_896_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_39_cast_fp16 = add(x = mul_5_cast_fp16, y = var_896_to_fp16)[name = tensor("var_39_cast_fp16")]; + tensor var_899_epsilon_0 = const()[name = tensor("op_899_epsilon_0"), val = tensor(0x1.197998p-40)]; + tensor var_899_cast_fp16 = rsqrt(epsilon = var_899_epsilon_0, x = var_39_cast_fp16)[name = tensor("op_899_cast_fp16")]; + tensor var_900_cast_fp16 = mul(x = const_5_to_fp16, y = var_899_cast_fp16)[name = tensor("op_900_cast_fp16")]; + tensor var_901_cast_fp16 = mul(x = linear_56_cast_fp16, y = var_900_cast_fp16)[name = tensor("op_901_cast_fp16")]; + tensor var_913_cast_fp16 = add(x = var_864_cast_fp16, y = var_901_cast_fp16)[name = tensor("op_913_cast_fp16")]; + tensor _inversed_t_combined_5_y_0_to_fp16 = const()[name = tensor("_inversed_t_combined_5_y_0_to_fp16"), val = tensor(0x1p-1)]; + tensor _inversed_t_combined_5_cast_fp16 = mul(x = var_913_cast_fp16, y = _inversed_t_combined_5_y_0_to_fp16)[name = tensor("_inversed_t_combined_5_cast_fp16")]; + tensor input_149_cast_fp16 = add(x = _inversed_t_combined_5_cast_fp16, y = linear_5_cast_fp16)[name = tensor("input_149_cast_fp16")]; + tensor input_151_cast_fp16 = silu(x = input_149_cast_fp16)[name = tensor("input_151_cast_fp16")]; + tensor linear_58_cast_fp16 = linear(bias = flow_net_res_blocks_0_adaLN_modulation_1_bias_to_fp16, weight = flow_net_res_blocks_0_adaLN_modulation_1_weight_to_fp16, x = input_151_cast_fp16)[name = tensor("linear_58_cast_fp16")]; + tensor var_928_split_sizes_0 = const()[name = tensor("op_928_split_sizes_0"), val = tensor([512, 512, 512])]; + tensor var_928_axis_0 = const()[name = tensor("op_928_axis_0"), val = tensor(-1)]; + tensor var_928_cast_fp16_0, tensor var_928_cast_fp16_1, tensor var_928_cast_fp16_2 = split(axis = var_928_axis_0, split_sizes = var_928_split_sizes_0, x = linear_58_cast_fp16)[name = tensor("op_928_cast_fp16")]; + tensor mean_29_axes_0 = const()[name = tensor("mean_29_axes_0"), val = tensor([-1])]; + tensor mean_29_keep_dims_0 = const()[name = tensor("mean_29_keep_dims_0"), val = tensor(true)]; + tensor mean_29_cast_fp16 = reduce_mean(axes = mean_29_axes_0, keep_dims = mean_29_keep_dims_0, x = linear_52_cast_fp16)[name = tensor("mean_29_cast_fp16")]; + tensor sub_26_cast_fp16 = sub(x = linear_52_cast_fp16, y = mean_29_cast_fp16)[name = tensor("sub_26_cast_fp16")]; + tensor square_20_cast_fp16 = square(x = sub_26_cast_fp16)[name = tensor("square_20_cast_fp16")]; + tensor reduce_mean_41_axes_0 = const()[name = tensor("reduce_mean_41_axes_0"), val = tensor([-1])]; + tensor reduce_mean_41_keep_dims_0 = const()[name = tensor("reduce_mean_41_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_41_cast_fp16 = reduce_mean(axes = reduce_mean_41_axes_0, keep_dims = reduce_mean_41_keep_dims_0, x = square_20_cast_fp16)[name = tensor("reduce_mean_41_cast_fp16")]; + tensor var_938_to_fp16 = const()[name = tensor("op_938_to_fp16"), val = tensor(0x1.1p-20)]; + tensor var_939_cast_fp16 = add(x = reduce_mean_41_cast_fp16, y = var_938_to_fp16)[name = tensor("op_939_cast_fp16")]; + tensor var_940_cast_fp16 = sqrt(x = var_939_cast_fp16)[name = tensor("op_940_cast_fp16")]; + tensor x_95_cast_fp16 = real_div(x = sub_26_cast_fp16, y = var_940_cast_fp16)[name = tensor("x_95_cast_fp16")]; + tensor var_942_cast_fp16 = mul(x = x_95_cast_fp16, y = flow_net_res_blocks_0_in_ln_weight_to_fp16)[name = tensor("op_942_cast_fp16")]; + tensor x_97_cast_fp16 = add(x = var_942_cast_fp16, y = flow_net_res_blocks_0_in_ln_bias_to_fp16)[name = tensor("x_97_cast_fp16")]; + tensor var_944_promoted_to_fp16 = const()[name = tensor("op_944_promoted_to_fp16"), val = tensor(0x1p+0)]; + tensor var_945_cast_fp16 = add(x = var_928_cast_fp16_1, y = var_944_promoted_to_fp16)[name = tensor("op_945_cast_fp16")]; + tensor var_946_cast_fp16 = mul(x = x_97_cast_fp16, y = var_945_cast_fp16)[name = tensor("op_946_cast_fp16")]; + tensor input_153_cast_fp16 = add(x = var_946_cast_fp16, y = var_928_cast_fp16_0)[name = tensor("input_153_cast_fp16")]; + tensor linear_59_cast_fp16 = linear(bias = flow_net_res_blocks_0_mlp_0_bias_to_fp16, weight = flow_net_res_blocks_0_mlp_0_weight_to_fp16, x = input_153_cast_fp16)[name = tensor("linear_59_cast_fp16")]; + tensor input_157_cast_fp16 = silu(x = linear_59_cast_fp16)[name = tensor("input_157_cast_fp16")]; + tensor linear_60_cast_fp16 = linear(bias = flow_net_res_blocks_0_mlp_2_bias_to_fp16, weight = flow_net_res_blocks_0_mlp_2_weight_to_fp16, x = input_157_cast_fp16)[name = tensor("linear_60_cast_fp16")]; + tensor var_957_cast_fp16 = mul(x = var_928_cast_fp16_2, y = linear_60_cast_fp16)[name = tensor("op_957_cast_fp16")]; + tensor x_99_cast_fp16 = add(x = linear_52_cast_fp16, y = var_957_cast_fp16)[name = tensor("x_99_cast_fp16")]; + tensor linear_61_cast_fp16 = linear(bias = flow_net_res_blocks_1_adaLN_modulation_1_bias_to_fp16, weight = flow_net_res_blocks_1_adaLN_modulation_1_weight_to_fp16, x = input_151_cast_fp16)[name = tensor("linear_61_cast_fp16")]; + tensor var_967_split_sizes_0 = const()[name = tensor("op_967_split_sizes_0"), val = tensor([512, 512, 512])]; + tensor var_967_axis_0 = const()[name = tensor("op_967_axis_0"), val = tensor(-1)]; + tensor var_967_cast_fp16_0, tensor var_967_cast_fp16_1, tensor var_967_cast_fp16_2 = split(axis = var_967_axis_0, split_sizes = var_967_split_sizes_0, x = linear_61_cast_fp16)[name = tensor("op_967_cast_fp16")]; + tensor mean_31_axes_0 = const()[name = tensor("mean_31_axes_0"), val = tensor([-1])]; + tensor mean_31_keep_dims_0 = const()[name = tensor("mean_31_keep_dims_0"), val = tensor(true)]; + tensor mean_31_cast_fp16 = reduce_mean(axes = mean_31_axes_0, keep_dims = mean_31_keep_dims_0, x = x_99_cast_fp16)[name = tensor("mean_31_cast_fp16")]; + tensor sub_27_cast_fp16 = sub(x = x_99_cast_fp16, y = mean_31_cast_fp16)[name = tensor("sub_27_cast_fp16")]; + tensor square_21_cast_fp16 = square(x = sub_27_cast_fp16)[name = tensor("square_21_cast_fp16")]; + tensor reduce_mean_43_axes_0 = const()[name = tensor("reduce_mean_43_axes_0"), val = tensor([-1])]; + tensor reduce_mean_43_keep_dims_0 = const()[name = tensor("reduce_mean_43_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_43_cast_fp16 = reduce_mean(axes = reduce_mean_43_axes_0, keep_dims = reduce_mean_43_keep_dims_0, x = square_21_cast_fp16)[name = tensor("reduce_mean_43_cast_fp16")]; + tensor var_977_to_fp16 = const()[name = tensor("op_977_to_fp16"), val = tensor(0x1.1p-20)]; + tensor var_978_cast_fp16 = add(x = reduce_mean_43_cast_fp16, y = var_977_to_fp16)[name = tensor("op_978_cast_fp16")]; + tensor var_979_cast_fp16 = sqrt(x = var_978_cast_fp16)[name = tensor("op_979_cast_fp16")]; + tensor x_101_cast_fp16 = real_div(x = sub_27_cast_fp16, y = var_979_cast_fp16)[name = tensor("x_101_cast_fp16")]; + tensor var_981_cast_fp16 = mul(x = x_101_cast_fp16, y = flow_net_res_blocks_1_in_ln_weight_to_fp16)[name = tensor("op_981_cast_fp16")]; + tensor x_103_cast_fp16 = add(x = var_981_cast_fp16, y = flow_net_res_blocks_1_in_ln_bias_to_fp16)[name = tensor("x_103_cast_fp16")]; + tensor var_983_promoted_to_fp16 = const()[name = tensor("op_983_promoted_to_fp16"), val = tensor(0x1p+0)]; + tensor var_984_cast_fp16 = add(x = var_967_cast_fp16_1, y = var_983_promoted_to_fp16)[name = tensor("op_984_cast_fp16")]; + tensor var_985_cast_fp16 = mul(x = x_103_cast_fp16, y = var_984_cast_fp16)[name = tensor("op_985_cast_fp16")]; + tensor input_161_cast_fp16 = add(x = var_985_cast_fp16, y = var_967_cast_fp16_0)[name = tensor("input_161_cast_fp16")]; + tensor linear_62_cast_fp16 = linear(bias = flow_net_res_blocks_1_mlp_0_bias_to_fp16, weight = flow_net_res_blocks_1_mlp_0_weight_to_fp16, x = input_161_cast_fp16)[name = tensor("linear_62_cast_fp16")]; + tensor input_165_cast_fp16 = silu(x = linear_62_cast_fp16)[name = tensor("input_165_cast_fp16")]; + tensor linear_63_cast_fp16 = linear(bias = flow_net_res_blocks_1_mlp_2_bias_to_fp16, weight = flow_net_res_blocks_1_mlp_2_weight_to_fp16, x = input_165_cast_fp16)[name = tensor("linear_63_cast_fp16")]; + tensor var_996_cast_fp16 = mul(x = var_967_cast_fp16_2, y = linear_63_cast_fp16)[name = tensor("op_996_cast_fp16")]; + tensor x_105_cast_fp16 = add(x = x_99_cast_fp16, y = var_996_cast_fp16)[name = tensor("x_105_cast_fp16")]; + tensor linear_64_cast_fp16 = linear(bias = flow_net_res_blocks_2_adaLN_modulation_1_bias_to_fp16, weight = flow_net_res_blocks_2_adaLN_modulation_1_weight_to_fp16, x = input_151_cast_fp16)[name = tensor("linear_64_cast_fp16")]; + tensor var_1006_split_sizes_0 = const()[name = tensor("op_1006_split_sizes_0"), val = tensor([512, 512, 512])]; + tensor var_1006_axis_0 = const()[name = tensor("op_1006_axis_0"), val = tensor(-1)]; + tensor var_1006_cast_fp16_0, tensor var_1006_cast_fp16_1, tensor var_1006_cast_fp16_2 = split(axis = var_1006_axis_0, split_sizes = var_1006_split_sizes_0, x = linear_64_cast_fp16)[name = tensor("op_1006_cast_fp16")]; + tensor mean_33_axes_0 = const()[name = tensor("mean_33_axes_0"), val = tensor([-1])]; + tensor mean_33_keep_dims_0 = const()[name = tensor("mean_33_keep_dims_0"), val = tensor(true)]; + tensor mean_33_cast_fp16 = reduce_mean(axes = mean_33_axes_0, keep_dims = mean_33_keep_dims_0, x = x_105_cast_fp16)[name = tensor("mean_33_cast_fp16")]; + tensor sub_28_cast_fp16 = sub(x = x_105_cast_fp16, y = mean_33_cast_fp16)[name = tensor("sub_28_cast_fp16")]; + tensor square_22_cast_fp16 = square(x = sub_28_cast_fp16)[name = tensor("square_22_cast_fp16")]; + tensor reduce_mean_45_axes_0 = const()[name = tensor("reduce_mean_45_axes_0"), val = tensor([-1])]; + tensor reduce_mean_45_keep_dims_0 = const()[name = tensor("reduce_mean_45_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_45_cast_fp16 = reduce_mean(axes = reduce_mean_45_axes_0, keep_dims = reduce_mean_45_keep_dims_0, x = square_22_cast_fp16)[name = tensor("reduce_mean_45_cast_fp16")]; + tensor var_1016_to_fp16 = const()[name = tensor("op_1016_to_fp16"), val = tensor(0x1.1p-20)]; + tensor var_1017_cast_fp16 = add(x = reduce_mean_45_cast_fp16, y = var_1016_to_fp16)[name = tensor("op_1017_cast_fp16")]; + tensor var_1018_cast_fp16 = sqrt(x = var_1017_cast_fp16)[name = tensor("op_1018_cast_fp16")]; + tensor x_107_cast_fp16 = real_div(x = sub_28_cast_fp16, y = var_1018_cast_fp16)[name = tensor("x_107_cast_fp16")]; + tensor var_1020_cast_fp16 = mul(x = x_107_cast_fp16, y = flow_net_res_blocks_2_in_ln_weight_to_fp16)[name = tensor("op_1020_cast_fp16")]; + tensor x_109_cast_fp16 = add(x = var_1020_cast_fp16, y = flow_net_res_blocks_2_in_ln_bias_to_fp16)[name = tensor("x_109_cast_fp16")]; + tensor var_1022_promoted_to_fp16 = const()[name = tensor("op_1022_promoted_to_fp16"), val = tensor(0x1p+0)]; + tensor var_1023_cast_fp16 = add(x = var_1006_cast_fp16_1, y = var_1022_promoted_to_fp16)[name = tensor("op_1023_cast_fp16")]; + tensor var_1024_cast_fp16 = mul(x = x_109_cast_fp16, y = var_1023_cast_fp16)[name = tensor("op_1024_cast_fp16")]; + tensor input_169_cast_fp16 = add(x = var_1024_cast_fp16, y = var_1006_cast_fp16_0)[name = tensor("input_169_cast_fp16")]; + tensor linear_65_cast_fp16 = linear(bias = flow_net_res_blocks_2_mlp_0_bias_to_fp16, weight = flow_net_res_blocks_2_mlp_0_weight_to_fp16, x = input_169_cast_fp16)[name = tensor("linear_65_cast_fp16")]; + tensor input_173_cast_fp16 = silu(x = linear_65_cast_fp16)[name = tensor("input_173_cast_fp16")]; + tensor linear_66_cast_fp16 = linear(bias = flow_net_res_blocks_2_mlp_2_bias_to_fp16, weight = flow_net_res_blocks_2_mlp_2_weight_to_fp16, x = input_173_cast_fp16)[name = tensor("linear_66_cast_fp16")]; + tensor var_1035_cast_fp16 = mul(x = var_1006_cast_fp16_2, y = linear_66_cast_fp16)[name = tensor("op_1035_cast_fp16")]; + tensor x_111_cast_fp16 = add(x = x_105_cast_fp16, y = var_1035_cast_fp16)[name = tensor("x_111_cast_fp16")]; + tensor linear_67_cast_fp16 = linear(bias = flow_net_res_blocks_3_adaLN_modulation_1_bias_to_fp16, weight = flow_net_res_blocks_3_adaLN_modulation_1_weight_to_fp16, x = input_151_cast_fp16)[name = tensor("linear_67_cast_fp16")]; + tensor var_1045_split_sizes_0 = const()[name = tensor("op_1045_split_sizes_0"), val = tensor([512, 512, 512])]; + tensor var_1045_axis_0 = const()[name = tensor("op_1045_axis_0"), val = tensor(-1)]; + tensor var_1045_cast_fp16_0, tensor var_1045_cast_fp16_1, tensor var_1045_cast_fp16_2 = split(axis = var_1045_axis_0, split_sizes = var_1045_split_sizes_0, x = linear_67_cast_fp16)[name = tensor("op_1045_cast_fp16")]; + tensor mean_35_axes_0 = const()[name = tensor("mean_35_axes_0"), val = tensor([-1])]; + tensor mean_35_keep_dims_0 = const()[name = tensor("mean_35_keep_dims_0"), val = tensor(true)]; + tensor mean_35_cast_fp16 = reduce_mean(axes = mean_35_axes_0, keep_dims = mean_35_keep_dims_0, x = x_111_cast_fp16)[name = tensor("mean_35_cast_fp16")]; + tensor sub_29_cast_fp16 = sub(x = x_111_cast_fp16, y = mean_35_cast_fp16)[name = tensor("sub_29_cast_fp16")]; + tensor square_23_cast_fp16 = square(x = sub_29_cast_fp16)[name = tensor("square_23_cast_fp16")]; + tensor reduce_mean_47_axes_0 = const()[name = tensor("reduce_mean_47_axes_0"), val = tensor([-1])]; + tensor reduce_mean_47_keep_dims_0 = const()[name = tensor("reduce_mean_47_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_47_cast_fp16 = reduce_mean(axes = reduce_mean_47_axes_0, keep_dims = reduce_mean_47_keep_dims_0, x = square_23_cast_fp16)[name = tensor("reduce_mean_47_cast_fp16")]; + tensor var_1055_to_fp16 = const()[name = tensor("op_1055_to_fp16"), val = tensor(0x1.1p-20)]; + tensor var_1056_cast_fp16 = add(x = reduce_mean_47_cast_fp16, y = var_1055_to_fp16)[name = tensor("op_1056_cast_fp16")]; + tensor var_1057_cast_fp16 = sqrt(x = var_1056_cast_fp16)[name = tensor("op_1057_cast_fp16")]; + tensor x_113_cast_fp16 = real_div(x = sub_29_cast_fp16, y = var_1057_cast_fp16)[name = tensor("x_113_cast_fp16")]; + tensor var_1059_cast_fp16 = mul(x = x_113_cast_fp16, y = flow_net_res_blocks_3_in_ln_weight_to_fp16)[name = tensor("op_1059_cast_fp16")]; + tensor x_115_cast_fp16 = add(x = var_1059_cast_fp16, y = flow_net_res_blocks_3_in_ln_bias_to_fp16)[name = tensor("x_115_cast_fp16")]; + tensor var_1061_promoted_to_fp16 = const()[name = tensor("op_1061_promoted_to_fp16"), val = tensor(0x1p+0)]; + tensor var_1062_cast_fp16 = add(x = var_1045_cast_fp16_1, y = var_1061_promoted_to_fp16)[name = tensor("op_1062_cast_fp16")]; + tensor var_1063_cast_fp16 = mul(x = x_115_cast_fp16, y = var_1062_cast_fp16)[name = tensor("op_1063_cast_fp16")]; + tensor input_177_cast_fp16 = add(x = var_1063_cast_fp16, y = var_1045_cast_fp16_0)[name = tensor("input_177_cast_fp16")]; + tensor linear_68_cast_fp16 = linear(bias = flow_net_res_blocks_3_mlp_0_bias_to_fp16, weight = flow_net_res_blocks_3_mlp_0_weight_to_fp16, x = input_177_cast_fp16)[name = tensor("linear_68_cast_fp16")]; + tensor input_181_cast_fp16 = silu(x = linear_68_cast_fp16)[name = tensor("input_181_cast_fp16")]; + tensor linear_69_cast_fp16 = linear(bias = flow_net_res_blocks_3_mlp_2_bias_to_fp16, weight = flow_net_res_blocks_3_mlp_2_weight_to_fp16, x = input_181_cast_fp16)[name = tensor("linear_69_cast_fp16")]; + tensor var_1074_cast_fp16 = mul(x = var_1045_cast_fp16_2, y = linear_69_cast_fp16)[name = tensor("op_1074_cast_fp16")]; + tensor x_117_cast_fp16 = add(x = x_111_cast_fp16, y = var_1074_cast_fp16)[name = tensor("x_117_cast_fp16")]; + tensor linear_70_cast_fp16 = linear(bias = flow_net_res_blocks_4_adaLN_modulation_1_bias_to_fp16, weight = flow_net_res_blocks_4_adaLN_modulation_1_weight_to_fp16, x = input_151_cast_fp16)[name = tensor("linear_70_cast_fp16")]; + tensor var_1084_split_sizes_0 = const()[name = tensor("op_1084_split_sizes_0"), val = tensor([512, 512, 512])]; + tensor var_1084_axis_0 = const()[name = tensor("op_1084_axis_0"), val = tensor(-1)]; + tensor var_1084_cast_fp16_0, tensor var_1084_cast_fp16_1, tensor var_1084_cast_fp16_2 = split(axis = var_1084_axis_0, split_sizes = var_1084_split_sizes_0, x = linear_70_cast_fp16)[name = tensor("op_1084_cast_fp16")]; + tensor mean_37_axes_0 = const()[name = tensor("mean_37_axes_0"), val = tensor([-1])]; + tensor mean_37_keep_dims_0 = const()[name = tensor("mean_37_keep_dims_0"), val = tensor(true)]; + tensor mean_37_cast_fp16 = reduce_mean(axes = mean_37_axes_0, keep_dims = mean_37_keep_dims_0, x = x_117_cast_fp16)[name = tensor("mean_37_cast_fp16")]; + tensor sub_30_cast_fp16 = sub(x = x_117_cast_fp16, y = mean_37_cast_fp16)[name = tensor("sub_30_cast_fp16")]; + tensor square_24_cast_fp16 = square(x = sub_30_cast_fp16)[name = tensor("square_24_cast_fp16")]; + tensor reduce_mean_49_axes_0 = const()[name = tensor("reduce_mean_49_axes_0"), val = tensor([-1])]; + tensor reduce_mean_49_keep_dims_0 = const()[name = tensor("reduce_mean_49_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_49_cast_fp16 = reduce_mean(axes = reduce_mean_49_axes_0, keep_dims = reduce_mean_49_keep_dims_0, x = square_24_cast_fp16)[name = tensor("reduce_mean_49_cast_fp16")]; + tensor var_1094_to_fp16 = const()[name = tensor("op_1094_to_fp16"), val = tensor(0x1.1p-20)]; + tensor var_1095_cast_fp16 = add(x = reduce_mean_49_cast_fp16, y = var_1094_to_fp16)[name = tensor("op_1095_cast_fp16")]; + tensor var_1096_cast_fp16 = sqrt(x = var_1095_cast_fp16)[name = tensor("op_1096_cast_fp16")]; + tensor x_119_cast_fp16 = real_div(x = sub_30_cast_fp16, y = var_1096_cast_fp16)[name = tensor("x_119_cast_fp16")]; + tensor var_1098_cast_fp16 = mul(x = x_119_cast_fp16, y = flow_net_res_blocks_4_in_ln_weight_to_fp16)[name = tensor("op_1098_cast_fp16")]; + tensor x_121_cast_fp16 = add(x = var_1098_cast_fp16, y = flow_net_res_blocks_4_in_ln_bias_to_fp16)[name = tensor("x_121_cast_fp16")]; + tensor var_1100_promoted_to_fp16 = const()[name = tensor("op_1100_promoted_to_fp16"), val = tensor(0x1p+0)]; + tensor var_1101_cast_fp16 = add(x = var_1084_cast_fp16_1, y = var_1100_promoted_to_fp16)[name = tensor("op_1101_cast_fp16")]; + tensor var_1102_cast_fp16 = mul(x = x_121_cast_fp16, y = var_1101_cast_fp16)[name = tensor("op_1102_cast_fp16")]; + tensor input_185_cast_fp16 = add(x = var_1102_cast_fp16, y = var_1084_cast_fp16_0)[name = tensor("input_185_cast_fp16")]; + tensor linear_71_cast_fp16 = linear(bias = flow_net_res_blocks_4_mlp_0_bias_to_fp16, weight = flow_net_res_blocks_4_mlp_0_weight_to_fp16, x = input_185_cast_fp16)[name = tensor("linear_71_cast_fp16")]; + tensor input_189_cast_fp16 = silu(x = linear_71_cast_fp16)[name = tensor("input_189_cast_fp16")]; + tensor linear_72_cast_fp16 = linear(bias = flow_net_res_blocks_4_mlp_2_bias_to_fp16, weight = flow_net_res_blocks_4_mlp_2_weight_to_fp16, x = input_189_cast_fp16)[name = tensor("linear_72_cast_fp16")]; + tensor var_1113_cast_fp16 = mul(x = var_1084_cast_fp16_2, y = linear_72_cast_fp16)[name = tensor("op_1113_cast_fp16")]; + tensor x_123_cast_fp16 = add(x = x_117_cast_fp16, y = var_1113_cast_fp16)[name = tensor("x_123_cast_fp16")]; + tensor linear_73_cast_fp16 = linear(bias = flow_net_res_blocks_5_adaLN_modulation_1_bias_to_fp16, weight = flow_net_res_blocks_5_adaLN_modulation_1_weight_to_fp16, x = input_151_cast_fp16)[name = tensor("linear_73_cast_fp16")]; + tensor var_1123_split_sizes_0 = const()[name = tensor("op_1123_split_sizes_0"), val = tensor([512, 512, 512])]; + tensor var_1123_axis_0 = const()[name = tensor("op_1123_axis_0"), val = tensor(-1)]; + tensor var_1123_cast_fp16_0, tensor var_1123_cast_fp16_1, tensor var_1123_cast_fp16_2 = split(axis = var_1123_axis_0, split_sizes = var_1123_split_sizes_0, x = linear_73_cast_fp16)[name = tensor("op_1123_cast_fp16")]; + tensor mean_39_axes_0 = const()[name = tensor("mean_39_axes_0"), val = tensor([-1])]; + tensor mean_39_keep_dims_0 = const()[name = tensor("mean_39_keep_dims_0"), val = tensor(true)]; + tensor mean_39_cast_fp16 = reduce_mean(axes = mean_39_axes_0, keep_dims = mean_39_keep_dims_0, x = x_123_cast_fp16)[name = tensor("mean_39_cast_fp16")]; + tensor sub_31_cast_fp16 = sub(x = x_123_cast_fp16, y = mean_39_cast_fp16)[name = tensor("sub_31_cast_fp16")]; + tensor square_25_cast_fp16 = square(x = sub_31_cast_fp16)[name = tensor("square_25_cast_fp16")]; + tensor reduce_mean_51_axes_0 = const()[name = tensor("reduce_mean_51_axes_0"), val = tensor([-1])]; + tensor reduce_mean_51_keep_dims_0 = const()[name = tensor("reduce_mean_51_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_51_cast_fp16 = reduce_mean(axes = reduce_mean_51_axes_0, keep_dims = reduce_mean_51_keep_dims_0, x = square_25_cast_fp16)[name = tensor("reduce_mean_51_cast_fp16")]; + tensor var_1133_to_fp16 = const()[name = tensor("op_1133_to_fp16"), val = tensor(0x1.1p-20)]; + tensor var_1134_cast_fp16 = add(x = reduce_mean_51_cast_fp16, y = var_1133_to_fp16)[name = tensor("op_1134_cast_fp16")]; + tensor var_1135_cast_fp16 = sqrt(x = var_1134_cast_fp16)[name = tensor("op_1135_cast_fp16")]; + tensor x_125_cast_fp16 = real_div(x = sub_31_cast_fp16, y = var_1135_cast_fp16)[name = tensor("x_125_cast_fp16")]; + tensor var_1137_cast_fp16 = mul(x = x_125_cast_fp16, y = flow_net_res_blocks_5_in_ln_weight_to_fp16)[name = tensor("op_1137_cast_fp16")]; + tensor x_127_cast_fp16 = add(x = var_1137_cast_fp16, y = flow_net_res_blocks_5_in_ln_bias_to_fp16)[name = tensor("x_127_cast_fp16")]; + tensor var_1139_promoted_to_fp16 = const()[name = tensor("op_1139_promoted_to_fp16"), val = tensor(0x1p+0)]; + tensor var_1140_cast_fp16 = add(x = var_1123_cast_fp16_1, y = var_1139_promoted_to_fp16)[name = tensor("op_1140_cast_fp16")]; + tensor var_1141_cast_fp16 = mul(x = x_127_cast_fp16, y = var_1140_cast_fp16)[name = tensor("op_1141_cast_fp16")]; + tensor input_193_cast_fp16 = add(x = var_1141_cast_fp16, y = var_1123_cast_fp16_0)[name = tensor("input_193_cast_fp16")]; + tensor linear_74_cast_fp16 = linear(bias = flow_net_res_blocks_5_mlp_0_bias_to_fp16, weight = flow_net_res_blocks_5_mlp_0_weight_to_fp16, x = input_193_cast_fp16)[name = tensor("linear_74_cast_fp16")]; + tensor input_197_cast_fp16 = silu(x = linear_74_cast_fp16)[name = tensor("input_197_cast_fp16")]; + tensor linear_75_cast_fp16 = linear(bias = flow_net_res_blocks_5_mlp_2_bias_to_fp16, weight = flow_net_res_blocks_5_mlp_2_weight_to_fp16, x = input_197_cast_fp16)[name = tensor("linear_75_cast_fp16")]; + tensor var_1152_cast_fp16 = mul(x = var_1123_cast_fp16_2, y = linear_75_cast_fp16)[name = tensor("op_1152_cast_fp16")]; + tensor x_129_cast_fp16 = add(x = x_123_cast_fp16, y = var_1152_cast_fp16)[name = tensor("x_129_cast_fp16")]; + tensor linear_76_cast_fp16 = linear(bias = flow_net_final_layer_adaLN_modulation_1_bias_to_fp16, weight = flow_net_final_layer_adaLN_modulation_1_weight_to_fp16, x = input_151_cast_fp16)[name = tensor("linear_76_cast_fp16")]; + tensor var_1161_split_sizes_0 = const()[name = tensor("op_1161_split_sizes_0"), val = tensor([512, 512])]; + tensor var_1161_axis_0 = const()[name = tensor("op_1161_axis_0"), val = tensor(-1)]; + tensor var_1161_cast_fp16_0, tensor var_1161_cast_fp16_1 = split(axis = var_1161_axis_0, split_sizes = var_1161_split_sizes_0, x = linear_76_cast_fp16)[name = tensor("op_1161_cast_fp16")]; + tensor mean_41_axes_0 = const()[name = tensor("mean_41_axes_0"), val = tensor([-1])]; + tensor mean_41_keep_dims_0 = const()[name = tensor("mean_41_keep_dims_0"), val = tensor(true)]; + tensor mean_41_cast_fp16 = reduce_mean(axes = mean_41_axes_0, keep_dims = mean_41_keep_dims_0, x = x_129_cast_fp16)[name = tensor("mean_41_cast_fp16")]; + tensor sub_32_cast_fp16 = sub(x = x_129_cast_fp16, y = mean_41_cast_fp16)[name = tensor("sub_32_cast_fp16")]; + tensor square_26_cast_fp16 = square(x = sub_32_cast_fp16)[name = tensor("square_26_cast_fp16")]; + tensor reduce_mean_53_axes_0 = const()[name = tensor("reduce_mean_53_axes_0"), val = tensor([-1])]; + tensor reduce_mean_53_keep_dims_0 = const()[name = tensor("reduce_mean_53_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_53_cast_fp16 = reduce_mean(axes = reduce_mean_53_axes_0, keep_dims = reduce_mean_53_keep_dims_0, x = square_26_cast_fp16)[name = tensor("reduce_mean_53_cast_fp16")]; + tensor var_1168_to_fp16 = const()[name = tensor("op_1168_to_fp16"), val = tensor(0x1.1p-20)]; + tensor var_1169_cast_fp16 = add(x = reduce_mean_53_cast_fp16, y = var_1168_to_fp16)[name = tensor("op_1169_cast_fp16")]; + tensor var_1170_cast_fp16 = sqrt(x = var_1169_cast_fp16)[name = tensor("op_1170_cast_fp16")]; + tensor x_131_cast_fp16 = real_div(x = sub_32_cast_fp16, y = var_1170_cast_fp16)[name = tensor("x_131_cast_fp16")]; + tensor var_1172_promoted_to_fp16 = const()[name = tensor("op_1172_promoted_to_fp16"), val = tensor(0x1p+0)]; + tensor var_1173_cast_fp16 = add(x = var_1161_cast_fp16_1, y = var_1172_promoted_to_fp16)[name = tensor("op_1173_cast_fp16")]; + tensor var_1174_cast_fp16 = mul(x = x_131_cast_fp16, y = var_1173_cast_fp16)[name = tensor("op_1174_cast_fp16")]; + tensor input_201_cast_fp16 = add(x = var_1174_cast_fp16, y = var_1161_cast_fp16_0)[name = tensor("input_201_cast_fp16")]; + tensor linear_77_cast_fp16 = linear(bias = flow_net_final_layer_linear_bias_to_fp16, weight = flow_net_final_layer_linear_weight_to_fp16, x = input_201_cast_fp16)[name = tensor("linear_77_cast_fp16")]; + tensor var_1185_to_fp16 = const()[name = tensor("op_1185_to_fp16"), val = tensor(0x1p-3)]; + tensor var_1186_cast_fp16 = mul(x = linear_77_cast_fp16, y = var_1185_to_fp16)[name = tensor("op_1186_cast_fp16")]; + tensor input_203_cast_fp16 = add(x = input_135_cast_fp16, y = var_1186_cast_fp16)[name = tensor("input_203_cast_fp16")]; + tensor linear_78_cast_fp16 = linear(bias = flow_net_input_proj_bias_to_fp16, weight = flow_net_input_proj_weight_to_fp16, x = input_203_cast_fp16)[name = tensor("linear_78_cast_fp16")]; + tensor input_207_to_fp16 = const()[name = tensor("input_207_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(19001728)))]; + tensor input_209_cast_fp16 = silu(x = input_207_to_fp16)[name = tensor("input_209_cast_fp16")]; + tensor linear_80_cast_fp16 = linear(bias = flow_net_time_embed_0_mlp_2_bias_to_fp16, weight = flow_net_time_embed_0_mlp_2_weight_to_fp16, x = input_209_cast_fp16)[name = tensor("linear_80_cast_fp16")]; + tensor reduce_mean_54_axes_0 = const()[name = tensor("reduce_mean_54_axes_0"), val = tensor([-1])]; + tensor reduce_mean_54_keep_dims_0 = const()[name = tensor("reduce_mean_54_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_54_cast_fp16 = reduce_mean(axes = reduce_mean_54_axes_0, keep_dims = reduce_mean_54_keep_dims_0, x = linear_80_cast_fp16)[name = tensor("reduce_mean_54_cast_fp16")]; + tensor sub_33_cast_fp16 = sub(x = linear_80_cast_fp16, y = reduce_mean_54_cast_fp16)[name = tensor("sub_33_cast_fp16")]; + tensor square_27_cast_fp16 = square(x = sub_33_cast_fp16)[name = tensor("square_27_cast_fp16")]; + tensor reduce_mean_55_axes_0 = const()[name = tensor("reduce_mean_55_axes_0"), val = tensor([-1])]; + tensor reduce_mean_55_keep_dims_0 = const()[name = tensor("reduce_mean_55_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_55_cast_fp16 = reduce_mean(axes = reduce_mean_55_axes_0, keep_dims = reduce_mean_55_keep_dims_0, x = square_27_cast_fp16)[name = tensor("reduce_mean_55_cast_fp16")]; + tensor real_div_6_to_fp16 = const()[name = tensor("real_div_6_to_fp16"), val = tensor(0x1.008p+0)]; + tensor mul_6_cast_fp16 = mul(x = reduce_mean_55_cast_fp16, y = real_div_6_to_fp16)[name = tensor("mul_6_cast_fp16")]; + tensor var_1252_to_fp16 = const()[name = tensor("op_1252_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_55_cast_fp16 = add(x = mul_6_cast_fp16, y = var_1252_to_fp16)[name = tensor("var_55_cast_fp16")]; + tensor var_1255_epsilon_0 = const()[name = tensor("op_1255_epsilon_0"), val = tensor(0x1.197998p-40)]; + tensor var_1255_cast_fp16 = rsqrt(epsilon = var_1255_epsilon_0, x = var_55_cast_fp16)[name = tensor("op_1255_cast_fp16")]; + tensor var_1256_cast_fp16 = mul(x = const_3_to_fp16, y = var_1255_cast_fp16)[name = tensor("op_1256_cast_fp16")]; + tensor var_1257_cast_fp16 = mul(x = linear_80_cast_fp16, y = var_1256_cast_fp16)[name = tensor("op_1257_cast_fp16")]; + tensor input_213_to_fp16 = const()[name = tensor("input_213_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(19002816)))]; + tensor input_215_cast_fp16 = silu(x = input_213_to_fp16)[name = tensor("input_215_cast_fp16")]; + tensor linear_82_cast_fp16 = linear(bias = flow_net_time_embed_1_mlp_2_bias_to_fp16, weight = flow_net_time_embed_1_mlp_2_weight_to_fp16, x = input_215_cast_fp16)[name = tensor("linear_82_cast_fp16")]; + tensor reduce_mean_56_axes_0 = const()[name = tensor("reduce_mean_56_axes_0"), val = tensor([-1])]; + tensor reduce_mean_56_keep_dims_0 = const()[name = tensor("reduce_mean_56_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_56_cast_fp16 = reduce_mean(axes = reduce_mean_56_axes_0, keep_dims = reduce_mean_56_keep_dims_0, x = linear_82_cast_fp16)[name = tensor("reduce_mean_56_cast_fp16")]; + tensor sub_35_cast_fp16 = sub(x = linear_82_cast_fp16, y = reduce_mean_56_cast_fp16)[name = tensor("sub_35_cast_fp16")]; + tensor square_28_cast_fp16 = square(x = sub_35_cast_fp16)[name = tensor("square_28_cast_fp16")]; + tensor reduce_mean_57_axes_0 = const()[name = tensor("reduce_mean_57_axes_0"), val = tensor([-1])]; + tensor reduce_mean_57_keep_dims_0 = const()[name = tensor("reduce_mean_57_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_57_cast_fp16 = reduce_mean(axes = reduce_mean_57_axes_0, keep_dims = reduce_mean_57_keep_dims_0, x = square_28_cast_fp16)[name = tensor("reduce_mean_57_cast_fp16")]; + tensor real_div_7_to_fp16 = const()[name = tensor("real_div_7_to_fp16"), val = tensor(0x1.008p+0)]; + tensor mul_7_cast_fp16 = mul(x = reduce_mean_57_cast_fp16, y = real_div_7_to_fp16)[name = tensor("mul_7_cast_fp16")]; + tensor var_1289_to_fp16 = const()[name = tensor("op_1289_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_57_cast_fp16 = add(x = mul_7_cast_fp16, y = var_1289_to_fp16)[name = tensor("var_57_cast_fp16")]; + tensor var_1292_epsilon_0 = const()[name = tensor("op_1292_epsilon_0"), val = tensor(0x1.197998p-40)]; + tensor var_1292_cast_fp16 = rsqrt(epsilon = var_1292_epsilon_0, x = var_57_cast_fp16)[name = tensor("op_1292_cast_fp16")]; + tensor var_1293_cast_fp16 = mul(x = const_5_to_fp16, y = var_1292_cast_fp16)[name = tensor("op_1293_cast_fp16")]; + tensor var_1294_cast_fp16 = mul(x = linear_82_cast_fp16, y = var_1293_cast_fp16)[name = tensor("op_1294_cast_fp16")]; + tensor var_1306_cast_fp16 = add(x = var_1257_cast_fp16, y = var_1294_cast_fp16)[name = tensor("op_1306_cast_fp16")]; + tensor _inversed_t_combined_7_y_0_to_fp16 = const()[name = tensor("_inversed_t_combined_7_y_0_to_fp16"), val = tensor(0x1p-1)]; + tensor _inversed_t_combined_7_cast_fp16 = mul(x = var_1306_cast_fp16, y = _inversed_t_combined_7_y_0_to_fp16)[name = tensor("_inversed_t_combined_7_cast_fp16")]; + tensor input_217_cast_fp16 = add(x = _inversed_t_combined_7_cast_fp16, y = linear_5_cast_fp16)[name = tensor("input_217_cast_fp16")]; + tensor input_219_cast_fp16 = silu(x = input_217_cast_fp16)[name = tensor("input_219_cast_fp16")]; + tensor linear_84_cast_fp16 = linear(bias = flow_net_res_blocks_0_adaLN_modulation_1_bias_to_fp16, weight = flow_net_res_blocks_0_adaLN_modulation_1_weight_to_fp16, x = input_219_cast_fp16)[name = tensor("linear_84_cast_fp16")]; + tensor var_1321_split_sizes_0 = const()[name = tensor("op_1321_split_sizes_0"), val = tensor([512, 512, 512])]; + tensor var_1321_axis_0 = const()[name = tensor("op_1321_axis_0"), val = tensor(-1)]; + tensor var_1321_cast_fp16_0, tensor var_1321_cast_fp16_1, tensor var_1321_cast_fp16_2 = split(axis = var_1321_axis_0, split_sizes = var_1321_split_sizes_0, x = linear_84_cast_fp16)[name = tensor("op_1321_cast_fp16")]; + tensor mean_43_axes_0 = const()[name = tensor("mean_43_axes_0"), val = tensor([-1])]; + tensor mean_43_keep_dims_0 = const()[name = tensor("mean_43_keep_dims_0"), val = tensor(true)]; + tensor mean_43_cast_fp16 = reduce_mean(axes = mean_43_axes_0, keep_dims = mean_43_keep_dims_0, x = linear_78_cast_fp16)[name = tensor("mean_43_cast_fp16")]; + tensor sub_37_cast_fp16 = sub(x = linear_78_cast_fp16, y = mean_43_cast_fp16)[name = tensor("sub_37_cast_fp16")]; + tensor square_29_cast_fp16 = square(x = sub_37_cast_fp16)[name = tensor("square_29_cast_fp16")]; + tensor reduce_mean_59_axes_0 = const()[name = tensor("reduce_mean_59_axes_0"), val = tensor([-1])]; + tensor reduce_mean_59_keep_dims_0 = const()[name = tensor("reduce_mean_59_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_59_cast_fp16 = reduce_mean(axes = reduce_mean_59_axes_0, keep_dims = reduce_mean_59_keep_dims_0, x = square_29_cast_fp16)[name = tensor("reduce_mean_59_cast_fp16")]; + tensor var_1331_to_fp16 = const()[name = tensor("op_1331_to_fp16"), val = tensor(0x1.1p-20)]; + tensor var_1332_cast_fp16 = add(x = reduce_mean_59_cast_fp16, y = var_1331_to_fp16)[name = tensor("op_1332_cast_fp16")]; + tensor var_1333_cast_fp16 = sqrt(x = var_1332_cast_fp16)[name = tensor("op_1333_cast_fp16")]; + tensor x_139_cast_fp16 = real_div(x = sub_37_cast_fp16, y = var_1333_cast_fp16)[name = tensor("x_139_cast_fp16")]; + tensor var_1335_cast_fp16 = mul(x = x_139_cast_fp16, y = flow_net_res_blocks_0_in_ln_weight_to_fp16)[name = tensor("op_1335_cast_fp16")]; + tensor x_141_cast_fp16 = add(x = var_1335_cast_fp16, y = flow_net_res_blocks_0_in_ln_bias_to_fp16)[name = tensor("x_141_cast_fp16")]; + tensor var_1337_promoted_to_fp16 = const()[name = tensor("op_1337_promoted_to_fp16"), val = tensor(0x1p+0)]; + tensor var_1338_cast_fp16 = add(x = var_1321_cast_fp16_1, y = var_1337_promoted_to_fp16)[name = tensor("op_1338_cast_fp16")]; + tensor var_1339_cast_fp16 = mul(x = x_141_cast_fp16, y = var_1338_cast_fp16)[name = tensor("op_1339_cast_fp16")]; + tensor input_221_cast_fp16 = add(x = var_1339_cast_fp16, y = var_1321_cast_fp16_0)[name = tensor("input_221_cast_fp16")]; + tensor linear_85_cast_fp16 = linear(bias = flow_net_res_blocks_0_mlp_0_bias_to_fp16, weight = flow_net_res_blocks_0_mlp_0_weight_to_fp16, x = input_221_cast_fp16)[name = tensor("linear_85_cast_fp16")]; + tensor input_225_cast_fp16 = silu(x = linear_85_cast_fp16)[name = tensor("input_225_cast_fp16")]; + tensor linear_86_cast_fp16 = linear(bias = flow_net_res_blocks_0_mlp_2_bias_to_fp16, weight = flow_net_res_blocks_0_mlp_2_weight_to_fp16, x = input_225_cast_fp16)[name = tensor("linear_86_cast_fp16")]; + tensor var_1350_cast_fp16 = mul(x = var_1321_cast_fp16_2, y = linear_86_cast_fp16)[name = tensor("op_1350_cast_fp16")]; + tensor x_143_cast_fp16 = add(x = linear_78_cast_fp16, y = var_1350_cast_fp16)[name = tensor("x_143_cast_fp16")]; + tensor linear_87_cast_fp16 = linear(bias = flow_net_res_blocks_1_adaLN_modulation_1_bias_to_fp16, weight = flow_net_res_blocks_1_adaLN_modulation_1_weight_to_fp16, x = input_219_cast_fp16)[name = tensor("linear_87_cast_fp16")]; + tensor var_1360_split_sizes_0 = const()[name = tensor("op_1360_split_sizes_0"), val = tensor([512, 512, 512])]; + tensor var_1360_axis_0 = const()[name = tensor("op_1360_axis_0"), val = tensor(-1)]; + tensor var_1360_cast_fp16_0, tensor var_1360_cast_fp16_1, tensor var_1360_cast_fp16_2 = split(axis = var_1360_axis_0, split_sizes = var_1360_split_sizes_0, x = linear_87_cast_fp16)[name = tensor("op_1360_cast_fp16")]; + tensor mean_45_axes_0 = const()[name = tensor("mean_45_axes_0"), val = tensor([-1])]; + tensor mean_45_keep_dims_0 = const()[name = tensor("mean_45_keep_dims_0"), val = tensor(true)]; + tensor mean_45_cast_fp16 = reduce_mean(axes = mean_45_axes_0, keep_dims = mean_45_keep_dims_0, x = x_143_cast_fp16)[name = tensor("mean_45_cast_fp16")]; + tensor sub_38_cast_fp16 = sub(x = x_143_cast_fp16, y = mean_45_cast_fp16)[name = tensor("sub_38_cast_fp16")]; + tensor square_30_cast_fp16 = square(x = sub_38_cast_fp16)[name = tensor("square_30_cast_fp16")]; + tensor reduce_mean_61_axes_0 = const()[name = tensor("reduce_mean_61_axes_0"), val = tensor([-1])]; + tensor reduce_mean_61_keep_dims_0 = const()[name = tensor("reduce_mean_61_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_61_cast_fp16 = reduce_mean(axes = reduce_mean_61_axes_0, keep_dims = reduce_mean_61_keep_dims_0, x = square_30_cast_fp16)[name = tensor("reduce_mean_61_cast_fp16")]; + tensor var_1370_to_fp16 = const()[name = tensor("op_1370_to_fp16"), val = tensor(0x1.1p-20)]; + tensor var_1371_cast_fp16 = add(x = reduce_mean_61_cast_fp16, y = var_1370_to_fp16)[name = tensor("op_1371_cast_fp16")]; + tensor var_1372_cast_fp16 = sqrt(x = var_1371_cast_fp16)[name = tensor("op_1372_cast_fp16")]; + tensor x_145_cast_fp16 = real_div(x = sub_38_cast_fp16, y = var_1372_cast_fp16)[name = tensor("x_145_cast_fp16")]; + tensor var_1374_cast_fp16 = mul(x = x_145_cast_fp16, y = flow_net_res_blocks_1_in_ln_weight_to_fp16)[name = tensor("op_1374_cast_fp16")]; + tensor x_147_cast_fp16 = add(x = var_1374_cast_fp16, y = flow_net_res_blocks_1_in_ln_bias_to_fp16)[name = tensor("x_147_cast_fp16")]; + tensor var_1376_promoted_to_fp16 = const()[name = tensor("op_1376_promoted_to_fp16"), val = tensor(0x1p+0)]; + tensor var_1377_cast_fp16 = add(x = var_1360_cast_fp16_1, y = var_1376_promoted_to_fp16)[name = tensor("op_1377_cast_fp16")]; + tensor var_1378_cast_fp16 = mul(x = x_147_cast_fp16, y = var_1377_cast_fp16)[name = tensor("op_1378_cast_fp16")]; + tensor input_229_cast_fp16 = add(x = var_1378_cast_fp16, y = var_1360_cast_fp16_0)[name = tensor("input_229_cast_fp16")]; + tensor linear_88_cast_fp16 = linear(bias = flow_net_res_blocks_1_mlp_0_bias_to_fp16, weight = flow_net_res_blocks_1_mlp_0_weight_to_fp16, x = input_229_cast_fp16)[name = tensor("linear_88_cast_fp16")]; + tensor input_233_cast_fp16 = silu(x = linear_88_cast_fp16)[name = tensor("input_233_cast_fp16")]; + tensor linear_89_cast_fp16 = linear(bias = flow_net_res_blocks_1_mlp_2_bias_to_fp16, weight = flow_net_res_blocks_1_mlp_2_weight_to_fp16, x = input_233_cast_fp16)[name = tensor("linear_89_cast_fp16")]; + tensor var_1389_cast_fp16 = mul(x = var_1360_cast_fp16_2, y = linear_89_cast_fp16)[name = tensor("op_1389_cast_fp16")]; + tensor x_149_cast_fp16 = add(x = x_143_cast_fp16, y = var_1389_cast_fp16)[name = tensor("x_149_cast_fp16")]; + tensor linear_90_cast_fp16 = linear(bias = flow_net_res_blocks_2_adaLN_modulation_1_bias_to_fp16, weight = flow_net_res_blocks_2_adaLN_modulation_1_weight_to_fp16, x = input_219_cast_fp16)[name = tensor("linear_90_cast_fp16")]; + tensor var_1399_split_sizes_0 = const()[name = tensor("op_1399_split_sizes_0"), val = tensor([512, 512, 512])]; + tensor var_1399_axis_0 = const()[name = tensor("op_1399_axis_0"), val = tensor(-1)]; + tensor var_1399_cast_fp16_0, tensor var_1399_cast_fp16_1, tensor var_1399_cast_fp16_2 = split(axis = var_1399_axis_0, split_sizes = var_1399_split_sizes_0, x = linear_90_cast_fp16)[name = tensor("op_1399_cast_fp16")]; + tensor mean_47_axes_0 = const()[name = tensor("mean_47_axes_0"), val = tensor([-1])]; + tensor mean_47_keep_dims_0 = const()[name = tensor("mean_47_keep_dims_0"), val = tensor(true)]; + tensor mean_47_cast_fp16 = reduce_mean(axes = mean_47_axes_0, keep_dims = mean_47_keep_dims_0, x = x_149_cast_fp16)[name = tensor("mean_47_cast_fp16")]; + tensor sub_39_cast_fp16 = sub(x = x_149_cast_fp16, y = mean_47_cast_fp16)[name = tensor("sub_39_cast_fp16")]; + tensor square_31_cast_fp16 = square(x = sub_39_cast_fp16)[name = tensor("square_31_cast_fp16")]; + tensor reduce_mean_63_axes_0 = const()[name = tensor("reduce_mean_63_axes_0"), val = tensor([-1])]; + tensor reduce_mean_63_keep_dims_0 = const()[name = tensor("reduce_mean_63_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_63_cast_fp16 = reduce_mean(axes = reduce_mean_63_axes_0, keep_dims = reduce_mean_63_keep_dims_0, x = square_31_cast_fp16)[name = tensor("reduce_mean_63_cast_fp16")]; + tensor var_1409_to_fp16 = const()[name = tensor("op_1409_to_fp16"), val = tensor(0x1.1p-20)]; + tensor var_1410_cast_fp16 = add(x = reduce_mean_63_cast_fp16, y = var_1409_to_fp16)[name = tensor("op_1410_cast_fp16")]; + tensor var_1411_cast_fp16 = sqrt(x = var_1410_cast_fp16)[name = tensor("op_1411_cast_fp16")]; + tensor x_151_cast_fp16 = real_div(x = sub_39_cast_fp16, y = var_1411_cast_fp16)[name = tensor("x_151_cast_fp16")]; + tensor var_1413_cast_fp16 = mul(x = x_151_cast_fp16, y = flow_net_res_blocks_2_in_ln_weight_to_fp16)[name = tensor("op_1413_cast_fp16")]; + tensor x_153_cast_fp16 = add(x = var_1413_cast_fp16, y = flow_net_res_blocks_2_in_ln_bias_to_fp16)[name = tensor("x_153_cast_fp16")]; + tensor var_1415_promoted_to_fp16 = const()[name = tensor("op_1415_promoted_to_fp16"), val = tensor(0x1p+0)]; + tensor var_1416_cast_fp16 = add(x = var_1399_cast_fp16_1, y = var_1415_promoted_to_fp16)[name = tensor("op_1416_cast_fp16")]; + tensor var_1417_cast_fp16 = mul(x = x_153_cast_fp16, y = var_1416_cast_fp16)[name = tensor("op_1417_cast_fp16")]; + tensor input_237_cast_fp16 = add(x = var_1417_cast_fp16, y = var_1399_cast_fp16_0)[name = tensor("input_237_cast_fp16")]; + tensor linear_91_cast_fp16 = linear(bias = flow_net_res_blocks_2_mlp_0_bias_to_fp16, weight = flow_net_res_blocks_2_mlp_0_weight_to_fp16, x = input_237_cast_fp16)[name = tensor("linear_91_cast_fp16")]; + tensor input_241_cast_fp16 = silu(x = linear_91_cast_fp16)[name = tensor("input_241_cast_fp16")]; + tensor linear_92_cast_fp16 = linear(bias = flow_net_res_blocks_2_mlp_2_bias_to_fp16, weight = flow_net_res_blocks_2_mlp_2_weight_to_fp16, x = input_241_cast_fp16)[name = tensor("linear_92_cast_fp16")]; + tensor var_1428_cast_fp16 = mul(x = var_1399_cast_fp16_2, y = linear_92_cast_fp16)[name = tensor("op_1428_cast_fp16")]; + tensor x_155_cast_fp16 = add(x = x_149_cast_fp16, y = var_1428_cast_fp16)[name = tensor("x_155_cast_fp16")]; + tensor linear_93_cast_fp16 = linear(bias = flow_net_res_blocks_3_adaLN_modulation_1_bias_to_fp16, weight = flow_net_res_blocks_3_adaLN_modulation_1_weight_to_fp16, x = input_219_cast_fp16)[name = tensor("linear_93_cast_fp16")]; + tensor var_1438_split_sizes_0 = const()[name = tensor("op_1438_split_sizes_0"), val = tensor([512, 512, 512])]; + tensor var_1438_axis_0 = const()[name = tensor("op_1438_axis_0"), val = tensor(-1)]; + tensor var_1438_cast_fp16_0, tensor var_1438_cast_fp16_1, tensor var_1438_cast_fp16_2 = split(axis = var_1438_axis_0, split_sizes = var_1438_split_sizes_0, x = linear_93_cast_fp16)[name = tensor("op_1438_cast_fp16")]; + tensor mean_49_axes_0 = const()[name = tensor("mean_49_axes_0"), val = tensor([-1])]; + tensor mean_49_keep_dims_0 = const()[name = tensor("mean_49_keep_dims_0"), val = tensor(true)]; + tensor mean_49_cast_fp16 = reduce_mean(axes = mean_49_axes_0, keep_dims = mean_49_keep_dims_0, x = x_155_cast_fp16)[name = tensor("mean_49_cast_fp16")]; + tensor sub_40_cast_fp16 = sub(x = x_155_cast_fp16, y = mean_49_cast_fp16)[name = tensor("sub_40_cast_fp16")]; + tensor square_32_cast_fp16 = square(x = sub_40_cast_fp16)[name = tensor("square_32_cast_fp16")]; + tensor reduce_mean_65_axes_0 = const()[name = tensor("reduce_mean_65_axes_0"), val = tensor([-1])]; + tensor reduce_mean_65_keep_dims_0 = const()[name = tensor("reduce_mean_65_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_65_cast_fp16 = reduce_mean(axes = reduce_mean_65_axes_0, keep_dims = reduce_mean_65_keep_dims_0, x = square_32_cast_fp16)[name = tensor("reduce_mean_65_cast_fp16")]; + tensor var_1448_to_fp16 = const()[name = tensor("op_1448_to_fp16"), val = tensor(0x1.1p-20)]; + tensor var_1449_cast_fp16 = add(x = reduce_mean_65_cast_fp16, y = var_1448_to_fp16)[name = tensor("op_1449_cast_fp16")]; + tensor var_1450_cast_fp16 = sqrt(x = var_1449_cast_fp16)[name = tensor("op_1450_cast_fp16")]; + tensor x_157_cast_fp16 = real_div(x = sub_40_cast_fp16, y = var_1450_cast_fp16)[name = tensor("x_157_cast_fp16")]; + tensor var_1452_cast_fp16 = mul(x = x_157_cast_fp16, y = flow_net_res_blocks_3_in_ln_weight_to_fp16)[name = tensor("op_1452_cast_fp16")]; + tensor x_159_cast_fp16 = add(x = var_1452_cast_fp16, y = flow_net_res_blocks_3_in_ln_bias_to_fp16)[name = tensor("x_159_cast_fp16")]; + tensor var_1454_promoted_to_fp16 = const()[name = tensor("op_1454_promoted_to_fp16"), val = tensor(0x1p+0)]; + tensor var_1455_cast_fp16 = add(x = var_1438_cast_fp16_1, y = var_1454_promoted_to_fp16)[name = tensor("op_1455_cast_fp16")]; + tensor var_1456_cast_fp16 = mul(x = x_159_cast_fp16, y = var_1455_cast_fp16)[name = tensor("op_1456_cast_fp16")]; + tensor input_245_cast_fp16 = add(x = var_1456_cast_fp16, y = var_1438_cast_fp16_0)[name = tensor("input_245_cast_fp16")]; + tensor linear_94_cast_fp16 = linear(bias = flow_net_res_blocks_3_mlp_0_bias_to_fp16, weight = flow_net_res_blocks_3_mlp_0_weight_to_fp16, x = input_245_cast_fp16)[name = tensor("linear_94_cast_fp16")]; + tensor input_249_cast_fp16 = silu(x = linear_94_cast_fp16)[name = tensor("input_249_cast_fp16")]; + tensor linear_95_cast_fp16 = linear(bias = flow_net_res_blocks_3_mlp_2_bias_to_fp16, weight = flow_net_res_blocks_3_mlp_2_weight_to_fp16, x = input_249_cast_fp16)[name = tensor("linear_95_cast_fp16")]; + tensor var_1467_cast_fp16 = mul(x = var_1438_cast_fp16_2, y = linear_95_cast_fp16)[name = tensor("op_1467_cast_fp16")]; + tensor x_161_cast_fp16 = add(x = x_155_cast_fp16, y = var_1467_cast_fp16)[name = tensor("x_161_cast_fp16")]; + tensor linear_96_cast_fp16 = linear(bias = flow_net_res_blocks_4_adaLN_modulation_1_bias_to_fp16, weight = flow_net_res_blocks_4_adaLN_modulation_1_weight_to_fp16, x = input_219_cast_fp16)[name = tensor("linear_96_cast_fp16")]; + tensor var_1477_split_sizes_0 = const()[name = tensor("op_1477_split_sizes_0"), val = tensor([512, 512, 512])]; + tensor var_1477_axis_0 = const()[name = tensor("op_1477_axis_0"), val = tensor(-1)]; + tensor var_1477_cast_fp16_0, tensor var_1477_cast_fp16_1, tensor var_1477_cast_fp16_2 = split(axis = var_1477_axis_0, split_sizes = var_1477_split_sizes_0, x = linear_96_cast_fp16)[name = tensor("op_1477_cast_fp16")]; + tensor mean_51_axes_0 = const()[name = tensor("mean_51_axes_0"), val = tensor([-1])]; + tensor mean_51_keep_dims_0 = const()[name = tensor("mean_51_keep_dims_0"), val = tensor(true)]; + tensor mean_51_cast_fp16 = reduce_mean(axes = mean_51_axes_0, keep_dims = mean_51_keep_dims_0, x = x_161_cast_fp16)[name = tensor("mean_51_cast_fp16")]; + tensor sub_41_cast_fp16 = sub(x = x_161_cast_fp16, y = mean_51_cast_fp16)[name = tensor("sub_41_cast_fp16")]; + tensor square_33_cast_fp16 = square(x = sub_41_cast_fp16)[name = tensor("square_33_cast_fp16")]; + tensor reduce_mean_67_axes_0 = const()[name = tensor("reduce_mean_67_axes_0"), val = tensor([-1])]; + tensor reduce_mean_67_keep_dims_0 = const()[name = tensor("reduce_mean_67_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_67_cast_fp16 = reduce_mean(axes = reduce_mean_67_axes_0, keep_dims = reduce_mean_67_keep_dims_0, x = square_33_cast_fp16)[name = tensor("reduce_mean_67_cast_fp16")]; + tensor var_1487_to_fp16 = const()[name = tensor("op_1487_to_fp16"), val = tensor(0x1.1p-20)]; + tensor var_1488_cast_fp16 = add(x = reduce_mean_67_cast_fp16, y = var_1487_to_fp16)[name = tensor("op_1488_cast_fp16")]; + tensor var_1489_cast_fp16 = sqrt(x = var_1488_cast_fp16)[name = tensor("op_1489_cast_fp16")]; + tensor x_163_cast_fp16 = real_div(x = sub_41_cast_fp16, y = var_1489_cast_fp16)[name = tensor("x_163_cast_fp16")]; + tensor var_1491_cast_fp16 = mul(x = x_163_cast_fp16, y = flow_net_res_blocks_4_in_ln_weight_to_fp16)[name = tensor("op_1491_cast_fp16")]; + tensor x_165_cast_fp16 = add(x = var_1491_cast_fp16, y = flow_net_res_blocks_4_in_ln_bias_to_fp16)[name = tensor("x_165_cast_fp16")]; + tensor var_1493_promoted_to_fp16 = const()[name = tensor("op_1493_promoted_to_fp16"), val = tensor(0x1p+0)]; + tensor var_1494_cast_fp16 = add(x = var_1477_cast_fp16_1, y = var_1493_promoted_to_fp16)[name = tensor("op_1494_cast_fp16")]; + tensor var_1495_cast_fp16 = mul(x = x_165_cast_fp16, y = var_1494_cast_fp16)[name = tensor("op_1495_cast_fp16")]; + tensor input_253_cast_fp16 = add(x = var_1495_cast_fp16, y = var_1477_cast_fp16_0)[name = tensor("input_253_cast_fp16")]; + tensor linear_97_cast_fp16 = linear(bias = flow_net_res_blocks_4_mlp_0_bias_to_fp16, weight = flow_net_res_blocks_4_mlp_0_weight_to_fp16, x = input_253_cast_fp16)[name = tensor("linear_97_cast_fp16")]; + tensor input_257_cast_fp16 = silu(x = linear_97_cast_fp16)[name = tensor("input_257_cast_fp16")]; + tensor linear_98_cast_fp16 = linear(bias = flow_net_res_blocks_4_mlp_2_bias_to_fp16, weight = flow_net_res_blocks_4_mlp_2_weight_to_fp16, x = input_257_cast_fp16)[name = tensor("linear_98_cast_fp16")]; + tensor var_1506_cast_fp16 = mul(x = var_1477_cast_fp16_2, y = linear_98_cast_fp16)[name = tensor("op_1506_cast_fp16")]; + tensor x_167_cast_fp16 = add(x = x_161_cast_fp16, y = var_1506_cast_fp16)[name = tensor("x_167_cast_fp16")]; + tensor linear_99_cast_fp16 = linear(bias = flow_net_res_blocks_5_adaLN_modulation_1_bias_to_fp16, weight = flow_net_res_blocks_5_adaLN_modulation_1_weight_to_fp16, x = input_219_cast_fp16)[name = tensor("linear_99_cast_fp16")]; + tensor var_1516_split_sizes_0 = const()[name = tensor("op_1516_split_sizes_0"), val = tensor([512, 512, 512])]; + tensor var_1516_axis_0 = const()[name = tensor("op_1516_axis_0"), val = tensor(-1)]; + tensor var_1516_cast_fp16_0, tensor var_1516_cast_fp16_1, tensor var_1516_cast_fp16_2 = split(axis = var_1516_axis_0, split_sizes = var_1516_split_sizes_0, x = linear_99_cast_fp16)[name = tensor("op_1516_cast_fp16")]; + tensor mean_53_axes_0 = const()[name = tensor("mean_53_axes_0"), val = tensor([-1])]; + tensor mean_53_keep_dims_0 = const()[name = tensor("mean_53_keep_dims_0"), val = tensor(true)]; + tensor mean_53_cast_fp16 = reduce_mean(axes = mean_53_axes_0, keep_dims = mean_53_keep_dims_0, x = x_167_cast_fp16)[name = tensor("mean_53_cast_fp16")]; + tensor sub_42_cast_fp16 = sub(x = x_167_cast_fp16, y = mean_53_cast_fp16)[name = tensor("sub_42_cast_fp16")]; + tensor square_34_cast_fp16 = square(x = sub_42_cast_fp16)[name = tensor("square_34_cast_fp16")]; + tensor reduce_mean_69_axes_0 = const()[name = tensor("reduce_mean_69_axes_0"), val = tensor([-1])]; + tensor reduce_mean_69_keep_dims_0 = const()[name = tensor("reduce_mean_69_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_69_cast_fp16 = reduce_mean(axes = reduce_mean_69_axes_0, keep_dims = reduce_mean_69_keep_dims_0, x = square_34_cast_fp16)[name = tensor("reduce_mean_69_cast_fp16")]; + tensor var_1526_to_fp16 = const()[name = tensor("op_1526_to_fp16"), val = tensor(0x1.1p-20)]; + tensor var_1527_cast_fp16 = add(x = reduce_mean_69_cast_fp16, y = var_1526_to_fp16)[name = tensor("op_1527_cast_fp16")]; + tensor var_1528_cast_fp16 = sqrt(x = var_1527_cast_fp16)[name = tensor("op_1528_cast_fp16")]; + tensor x_169_cast_fp16 = real_div(x = sub_42_cast_fp16, y = var_1528_cast_fp16)[name = tensor("x_169_cast_fp16")]; + tensor var_1530_cast_fp16 = mul(x = x_169_cast_fp16, y = flow_net_res_blocks_5_in_ln_weight_to_fp16)[name = tensor("op_1530_cast_fp16")]; + tensor x_171_cast_fp16 = add(x = var_1530_cast_fp16, y = flow_net_res_blocks_5_in_ln_bias_to_fp16)[name = tensor("x_171_cast_fp16")]; + tensor var_1532_promoted_to_fp16 = const()[name = tensor("op_1532_promoted_to_fp16"), val = tensor(0x1p+0)]; + tensor var_1533_cast_fp16 = add(x = var_1516_cast_fp16_1, y = var_1532_promoted_to_fp16)[name = tensor("op_1533_cast_fp16")]; + tensor var_1534_cast_fp16 = mul(x = x_171_cast_fp16, y = var_1533_cast_fp16)[name = tensor("op_1534_cast_fp16")]; + tensor input_261_cast_fp16 = add(x = var_1534_cast_fp16, y = var_1516_cast_fp16_0)[name = tensor("input_261_cast_fp16")]; + tensor linear_100_cast_fp16 = linear(bias = flow_net_res_blocks_5_mlp_0_bias_to_fp16, weight = flow_net_res_blocks_5_mlp_0_weight_to_fp16, x = input_261_cast_fp16)[name = tensor("linear_100_cast_fp16")]; + tensor input_265_cast_fp16 = silu(x = linear_100_cast_fp16)[name = tensor("input_265_cast_fp16")]; + tensor linear_101_cast_fp16 = linear(bias = flow_net_res_blocks_5_mlp_2_bias_to_fp16, weight = flow_net_res_blocks_5_mlp_2_weight_to_fp16, x = input_265_cast_fp16)[name = tensor("linear_101_cast_fp16")]; + tensor var_1545_cast_fp16 = mul(x = var_1516_cast_fp16_2, y = linear_101_cast_fp16)[name = tensor("op_1545_cast_fp16")]; + tensor x_173_cast_fp16 = add(x = x_167_cast_fp16, y = var_1545_cast_fp16)[name = tensor("x_173_cast_fp16")]; + tensor linear_102_cast_fp16 = linear(bias = flow_net_final_layer_adaLN_modulation_1_bias_to_fp16, weight = flow_net_final_layer_adaLN_modulation_1_weight_to_fp16, x = input_219_cast_fp16)[name = tensor("linear_102_cast_fp16")]; + tensor var_1554_split_sizes_0 = const()[name = tensor("op_1554_split_sizes_0"), val = tensor([512, 512])]; + tensor var_1554_axis_0 = const()[name = tensor("op_1554_axis_0"), val = tensor(-1)]; + tensor var_1554_cast_fp16_0, tensor var_1554_cast_fp16_1 = split(axis = var_1554_axis_0, split_sizes = var_1554_split_sizes_0, x = linear_102_cast_fp16)[name = tensor("op_1554_cast_fp16")]; + tensor mean_55_axes_0 = const()[name = tensor("mean_55_axes_0"), val = tensor([-1])]; + tensor mean_55_keep_dims_0 = const()[name = tensor("mean_55_keep_dims_0"), val = tensor(true)]; + tensor mean_55_cast_fp16 = reduce_mean(axes = mean_55_axes_0, keep_dims = mean_55_keep_dims_0, x = x_173_cast_fp16)[name = tensor("mean_55_cast_fp16")]; + tensor sub_43_cast_fp16 = sub(x = x_173_cast_fp16, y = mean_55_cast_fp16)[name = tensor("sub_43_cast_fp16")]; + tensor square_35_cast_fp16 = square(x = sub_43_cast_fp16)[name = tensor("square_35_cast_fp16")]; + tensor reduce_mean_71_axes_0 = const()[name = tensor("reduce_mean_71_axes_0"), val = tensor([-1])]; + tensor reduce_mean_71_keep_dims_0 = const()[name = tensor("reduce_mean_71_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_71_cast_fp16 = reduce_mean(axes = reduce_mean_71_axes_0, keep_dims = reduce_mean_71_keep_dims_0, x = square_35_cast_fp16)[name = tensor("reduce_mean_71_cast_fp16")]; + tensor var_1561_to_fp16 = const()[name = tensor("op_1561_to_fp16"), val = tensor(0x1.1p-20)]; + tensor var_1562_cast_fp16 = add(x = reduce_mean_71_cast_fp16, y = var_1561_to_fp16)[name = tensor("op_1562_cast_fp16")]; + tensor var_1563_cast_fp16 = sqrt(x = var_1562_cast_fp16)[name = tensor("op_1563_cast_fp16")]; + tensor x_175_cast_fp16 = real_div(x = sub_43_cast_fp16, y = var_1563_cast_fp16)[name = tensor("x_175_cast_fp16")]; + tensor var_1565_promoted_to_fp16 = const()[name = tensor("op_1565_promoted_to_fp16"), val = tensor(0x1p+0)]; + tensor var_1566_cast_fp16 = add(x = var_1554_cast_fp16_1, y = var_1565_promoted_to_fp16)[name = tensor("op_1566_cast_fp16")]; + tensor var_1567_cast_fp16 = mul(x = x_175_cast_fp16, y = var_1566_cast_fp16)[name = tensor("op_1567_cast_fp16")]; + tensor input_269_cast_fp16 = add(x = var_1567_cast_fp16, y = var_1554_cast_fp16_0)[name = tensor("input_269_cast_fp16")]; + tensor linear_103_cast_fp16 = linear(bias = flow_net_final_layer_linear_bias_to_fp16, weight = flow_net_final_layer_linear_weight_to_fp16, x = input_269_cast_fp16)[name = tensor("linear_103_cast_fp16")]; + tensor var_1578_to_fp16 = const()[name = tensor("op_1578_to_fp16"), val = tensor(0x1p-3)]; + tensor var_1579_cast_fp16 = mul(x = linear_103_cast_fp16, y = var_1578_to_fp16)[name = tensor("op_1579_cast_fp16")]; + tensor input_271_cast_fp16 = add(x = input_203_cast_fp16, y = var_1579_cast_fp16)[name = tensor("input_271_cast_fp16")]; + tensor linear_104_cast_fp16 = linear(bias = flow_net_input_proj_bias_to_fp16, weight = flow_net_input_proj_weight_to_fp16, x = input_271_cast_fp16)[name = tensor("linear_104_cast_fp16")]; + tensor input_275_to_fp16 = const()[name = tensor("input_275_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(19003904)))]; + tensor input_277_cast_fp16 = silu(x = input_275_to_fp16)[name = tensor("input_277_cast_fp16")]; + tensor linear_106_cast_fp16 = linear(bias = flow_net_time_embed_0_mlp_2_bias_to_fp16, weight = flow_net_time_embed_0_mlp_2_weight_to_fp16, x = input_277_cast_fp16)[name = tensor("linear_106_cast_fp16")]; + tensor reduce_mean_72_axes_0 = const()[name = tensor("reduce_mean_72_axes_0"), val = tensor([-1])]; + tensor reduce_mean_72_keep_dims_0 = const()[name = tensor("reduce_mean_72_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_72_cast_fp16 = reduce_mean(axes = reduce_mean_72_axes_0, keep_dims = reduce_mean_72_keep_dims_0, x = linear_106_cast_fp16)[name = tensor("reduce_mean_72_cast_fp16")]; + tensor sub_44_cast_fp16 = sub(x = linear_106_cast_fp16, y = reduce_mean_72_cast_fp16)[name = tensor("sub_44_cast_fp16")]; + tensor square_36_cast_fp16 = square(x = sub_44_cast_fp16)[name = tensor("square_36_cast_fp16")]; + tensor reduce_mean_73_axes_0 = const()[name = tensor("reduce_mean_73_axes_0"), val = tensor([-1])]; + tensor reduce_mean_73_keep_dims_0 = const()[name = tensor("reduce_mean_73_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_73_cast_fp16 = reduce_mean(axes = reduce_mean_73_axes_0, keep_dims = reduce_mean_73_keep_dims_0, x = square_36_cast_fp16)[name = tensor("reduce_mean_73_cast_fp16")]; + tensor real_div_8_to_fp16 = const()[name = tensor("real_div_8_to_fp16"), val = tensor(0x1.008p+0)]; + tensor mul_8_cast_fp16 = mul(x = reduce_mean_73_cast_fp16, y = real_div_8_to_fp16)[name = tensor("mul_8_cast_fp16")]; + tensor var_1645_to_fp16 = const()[name = tensor("op_1645_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_73_cast_fp16_0 = add(x = mul_8_cast_fp16, y = var_1645_to_fp16)[name = tensor("var_73_cast_fp16")]; + tensor var_1648_epsilon_0 = const()[name = tensor("op_1648_epsilon_0"), val = tensor(0x1.197998p-40)]; + tensor var_1648_cast_fp16 = rsqrt(epsilon = var_1648_epsilon_0, x = var_73_cast_fp16_0)[name = tensor("op_1648_cast_fp16")]; + tensor var_1649_cast_fp16 = mul(x = const_3_to_fp16, y = var_1648_cast_fp16)[name = tensor("op_1649_cast_fp16")]; + tensor var_1650_cast_fp16 = mul(x = linear_106_cast_fp16, y = var_1649_cast_fp16)[name = tensor("op_1650_cast_fp16")]; + tensor input_281_to_fp16 = const()[name = tensor("input_281_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(19004992)))]; + tensor input_283_cast_fp16 = silu(x = input_281_to_fp16)[name = tensor("input_283_cast_fp16")]; + tensor linear_108_cast_fp16 = linear(bias = flow_net_time_embed_1_mlp_2_bias_to_fp16, weight = flow_net_time_embed_1_mlp_2_weight_to_fp16, x = input_283_cast_fp16)[name = tensor("linear_108_cast_fp16")]; + tensor reduce_mean_74_axes_0 = const()[name = tensor("reduce_mean_74_axes_0"), val = tensor([-1])]; + tensor reduce_mean_74_keep_dims_0 = const()[name = tensor("reduce_mean_74_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_74_cast_fp16 = reduce_mean(axes = reduce_mean_74_axes_0, keep_dims = reduce_mean_74_keep_dims_0, x = linear_108_cast_fp16)[name = tensor("reduce_mean_74_cast_fp16")]; + tensor sub_46_cast_fp16 = sub(x = linear_108_cast_fp16, y = reduce_mean_74_cast_fp16)[name = tensor("sub_46_cast_fp16")]; + tensor square_37_cast_fp16 = square(x = sub_46_cast_fp16)[name = tensor("square_37_cast_fp16")]; + tensor reduce_mean_75_axes_0 = const()[name = tensor("reduce_mean_75_axes_0"), val = tensor([-1])]; + tensor reduce_mean_75_keep_dims_0 = const()[name = tensor("reduce_mean_75_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_75_cast_fp16 = reduce_mean(axes = reduce_mean_75_axes_0, keep_dims = reduce_mean_75_keep_dims_0, x = square_37_cast_fp16)[name = tensor("reduce_mean_75_cast_fp16")]; + tensor real_div_9_to_fp16 = const()[name = tensor("real_div_9_to_fp16"), val = tensor(0x1.008p+0)]; + tensor mul_9_cast_fp16 = mul(x = reduce_mean_75_cast_fp16, y = real_div_9_to_fp16)[name = tensor("mul_9_cast_fp16")]; + tensor var_1682_to_fp16 = const()[name = tensor("op_1682_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_75_cast_fp16_0 = add(x = mul_9_cast_fp16, y = var_1682_to_fp16)[name = tensor("var_75_cast_fp16")]; + tensor var_1685_epsilon_0 = const()[name = tensor("op_1685_epsilon_0"), val = tensor(0x1.197998p-40)]; + tensor var_1685_cast_fp16 = rsqrt(epsilon = var_1685_epsilon_0, x = var_75_cast_fp16_0)[name = tensor("op_1685_cast_fp16")]; + tensor var_1686_cast_fp16 = mul(x = const_5_to_fp16, y = var_1685_cast_fp16)[name = tensor("op_1686_cast_fp16")]; + tensor var_1687_cast_fp16 = mul(x = linear_108_cast_fp16, y = var_1686_cast_fp16)[name = tensor("op_1687_cast_fp16")]; + tensor var_1699_cast_fp16 = add(x = var_1650_cast_fp16, y = var_1687_cast_fp16)[name = tensor("op_1699_cast_fp16")]; + tensor _inversed_t_combined_9_y_0_to_fp16 = const()[name = tensor("_inversed_t_combined_9_y_0_to_fp16"), val = tensor(0x1p-1)]; + tensor _inversed_t_combined_9_cast_fp16 = mul(x = var_1699_cast_fp16, y = _inversed_t_combined_9_y_0_to_fp16)[name = tensor("_inversed_t_combined_9_cast_fp16")]; + tensor input_285_cast_fp16 = add(x = _inversed_t_combined_9_cast_fp16, y = linear_5_cast_fp16)[name = tensor("input_285_cast_fp16")]; + tensor input_287_cast_fp16 = silu(x = input_285_cast_fp16)[name = tensor("input_287_cast_fp16")]; + tensor linear_110_cast_fp16 = linear(bias = flow_net_res_blocks_0_adaLN_modulation_1_bias_to_fp16, weight = flow_net_res_blocks_0_adaLN_modulation_1_weight_to_fp16, x = input_287_cast_fp16)[name = tensor("linear_110_cast_fp16")]; + tensor var_1714_split_sizes_0 = const()[name = tensor("op_1714_split_sizes_0"), val = tensor([512, 512, 512])]; + tensor var_1714_axis_0 = const()[name = tensor("op_1714_axis_0"), val = tensor(-1)]; + tensor var_1714_cast_fp16_0, tensor var_1714_cast_fp16_1, tensor var_1714_cast_fp16_2 = split(axis = var_1714_axis_0, split_sizes = var_1714_split_sizes_0, x = linear_110_cast_fp16)[name = tensor("op_1714_cast_fp16")]; + tensor mean_57_axes_0 = const()[name = tensor("mean_57_axes_0"), val = tensor([-1])]; + tensor mean_57_keep_dims_0 = const()[name = tensor("mean_57_keep_dims_0"), val = tensor(true)]; + tensor mean_57_cast_fp16 = reduce_mean(axes = mean_57_axes_0, keep_dims = mean_57_keep_dims_0, x = linear_104_cast_fp16)[name = tensor("mean_57_cast_fp16")]; + tensor sub_48_cast_fp16 = sub(x = linear_104_cast_fp16, y = mean_57_cast_fp16)[name = tensor("sub_48_cast_fp16")]; + tensor square_38_cast_fp16 = square(x = sub_48_cast_fp16)[name = tensor("square_38_cast_fp16")]; + tensor reduce_mean_77_axes_0 = const()[name = tensor("reduce_mean_77_axes_0"), val = tensor([-1])]; + tensor reduce_mean_77_keep_dims_0 = const()[name = tensor("reduce_mean_77_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_77_cast_fp16 = reduce_mean(axes = reduce_mean_77_axes_0, keep_dims = reduce_mean_77_keep_dims_0, x = square_38_cast_fp16)[name = tensor("reduce_mean_77_cast_fp16")]; + tensor var_1724_to_fp16 = const()[name = tensor("op_1724_to_fp16"), val = tensor(0x1.1p-20)]; + tensor var_1725_cast_fp16 = add(x = reduce_mean_77_cast_fp16, y = var_1724_to_fp16)[name = tensor("op_1725_cast_fp16")]; + tensor var_1726_cast_fp16 = sqrt(x = var_1725_cast_fp16)[name = tensor("op_1726_cast_fp16")]; + tensor x_183_cast_fp16 = real_div(x = sub_48_cast_fp16, y = var_1726_cast_fp16)[name = tensor("x_183_cast_fp16")]; + tensor var_1728_cast_fp16 = mul(x = x_183_cast_fp16, y = flow_net_res_blocks_0_in_ln_weight_to_fp16)[name = tensor("op_1728_cast_fp16")]; + tensor x_185_cast_fp16 = add(x = var_1728_cast_fp16, y = flow_net_res_blocks_0_in_ln_bias_to_fp16)[name = tensor("x_185_cast_fp16")]; + tensor var_1730_promoted_to_fp16 = const()[name = tensor("op_1730_promoted_to_fp16"), val = tensor(0x1p+0)]; + tensor var_1731_cast_fp16 = add(x = var_1714_cast_fp16_1, y = var_1730_promoted_to_fp16)[name = tensor("op_1731_cast_fp16")]; + tensor var_1732_cast_fp16 = mul(x = x_185_cast_fp16, y = var_1731_cast_fp16)[name = tensor("op_1732_cast_fp16")]; + tensor input_289_cast_fp16 = add(x = var_1732_cast_fp16, y = var_1714_cast_fp16_0)[name = tensor("input_289_cast_fp16")]; + tensor linear_111_cast_fp16 = linear(bias = flow_net_res_blocks_0_mlp_0_bias_to_fp16, weight = flow_net_res_blocks_0_mlp_0_weight_to_fp16, x = input_289_cast_fp16)[name = tensor("linear_111_cast_fp16")]; + tensor input_293_cast_fp16 = silu(x = linear_111_cast_fp16)[name = tensor("input_293_cast_fp16")]; + tensor linear_112_cast_fp16 = linear(bias = flow_net_res_blocks_0_mlp_2_bias_to_fp16, weight = flow_net_res_blocks_0_mlp_2_weight_to_fp16, x = input_293_cast_fp16)[name = tensor("linear_112_cast_fp16")]; + tensor var_1743_cast_fp16 = mul(x = var_1714_cast_fp16_2, y = linear_112_cast_fp16)[name = tensor("op_1743_cast_fp16")]; + tensor x_187_cast_fp16 = add(x = linear_104_cast_fp16, y = var_1743_cast_fp16)[name = tensor("x_187_cast_fp16")]; + tensor linear_113_cast_fp16 = linear(bias = flow_net_res_blocks_1_adaLN_modulation_1_bias_to_fp16, weight = flow_net_res_blocks_1_adaLN_modulation_1_weight_to_fp16, x = input_287_cast_fp16)[name = tensor("linear_113_cast_fp16")]; + tensor var_1753_split_sizes_0 = const()[name = tensor("op_1753_split_sizes_0"), val = tensor([512, 512, 512])]; + tensor var_1753_axis_0 = const()[name = tensor("op_1753_axis_0"), val = tensor(-1)]; + tensor var_1753_cast_fp16_0, tensor var_1753_cast_fp16_1, tensor var_1753_cast_fp16_2 = split(axis = var_1753_axis_0, split_sizes = var_1753_split_sizes_0, x = linear_113_cast_fp16)[name = tensor("op_1753_cast_fp16")]; + tensor mean_59_axes_0 = const()[name = tensor("mean_59_axes_0"), val = tensor([-1])]; + tensor mean_59_keep_dims_0 = const()[name = tensor("mean_59_keep_dims_0"), val = tensor(true)]; + tensor mean_59_cast_fp16 = reduce_mean(axes = mean_59_axes_0, keep_dims = mean_59_keep_dims_0, x = x_187_cast_fp16)[name = tensor("mean_59_cast_fp16")]; + tensor sub_49_cast_fp16 = sub(x = x_187_cast_fp16, y = mean_59_cast_fp16)[name = tensor("sub_49_cast_fp16")]; + tensor square_39_cast_fp16 = square(x = sub_49_cast_fp16)[name = tensor("square_39_cast_fp16")]; + tensor reduce_mean_79_axes_0 = const()[name = tensor("reduce_mean_79_axes_0"), val = tensor([-1])]; + tensor reduce_mean_79_keep_dims_0 = const()[name = tensor("reduce_mean_79_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_79_cast_fp16 = reduce_mean(axes = reduce_mean_79_axes_0, keep_dims = reduce_mean_79_keep_dims_0, x = square_39_cast_fp16)[name = tensor("reduce_mean_79_cast_fp16")]; + tensor var_1763_to_fp16 = const()[name = tensor("op_1763_to_fp16"), val = tensor(0x1.1p-20)]; + tensor var_1764_cast_fp16 = add(x = reduce_mean_79_cast_fp16, y = var_1763_to_fp16)[name = tensor("op_1764_cast_fp16")]; + tensor var_1765_cast_fp16 = sqrt(x = var_1764_cast_fp16)[name = tensor("op_1765_cast_fp16")]; + tensor x_189_cast_fp16 = real_div(x = sub_49_cast_fp16, y = var_1765_cast_fp16)[name = tensor("x_189_cast_fp16")]; + tensor var_1767_cast_fp16 = mul(x = x_189_cast_fp16, y = flow_net_res_blocks_1_in_ln_weight_to_fp16)[name = tensor("op_1767_cast_fp16")]; + tensor x_191_cast_fp16 = add(x = var_1767_cast_fp16, y = flow_net_res_blocks_1_in_ln_bias_to_fp16)[name = tensor("x_191_cast_fp16")]; + tensor var_1769_promoted_to_fp16 = const()[name = tensor("op_1769_promoted_to_fp16"), val = tensor(0x1p+0)]; + tensor var_1770_cast_fp16 = add(x = var_1753_cast_fp16_1, y = var_1769_promoted_to_fp16)[name = tensor("op_1770_cast_fp16")]; + tensor var_1771_cast_fp16 = mul(x = x_191_cast_fp16, y = var_1770_cast_fp16)[name = tensor("op_1771_cast_fp16")]; + tensor input_297_cast_fp16 = add(x = var_1771_cast_fp16, y = var_1753_cast_fp16_0)[name = tensor("input_297_cast_fp16")]; + tensor linear_114_cast_fp16 = linear(bias = flow_net_res_blocks_1_mlp_0_bias_to_fp16, weight = flow_net_res_blocks_1_mlp_0_weight_to_fp16, x = input_297_cast_fp16)[name = tensor("linear_114_cast_fp16")]; + tensor input_301_cast_fp16 = silu(x = linear_114_cast_fp16)[name = tensor("input_301_cast_fp16")]; + tensor linear_115_cast_fp16 = linear(bias = flow_net_res_blocks_1_mlp_2_bias_to_fp16, weight = flow_net_res_blocks_1_mlp_2_weight_to_fp16, x = input_301_cast_fp16)[name = tensor("linear_115_cast_fp16")]; + tensor var_1782_cast_fp16 = mul(x = var_1753_cast_fp16_2, y = linear_115_cast_fp16)[name = tensor("op_1782_cast_fp16")]; + tensor x_193_cast_fp16 = add(x = x_187_cast_fp16, y = var_1782_cast_fp16)[name = tensor("x_193_cast_fp16")]; + tensor linear_116_cast_fp16 = linear(bias = flow_net_res_blocks_2_adaLN_modulation_1_bias_to_fp16, weight = flow_net_res_blocks_2_adaLN_modulation_1_weight_to_fp16, x = input_287_cast_fp16)[name = tensor("linear_116_cast_fp16")]; + tensor var_1792_split_sizes_0 = const()[name = tensor("op_1792_split_sizes_0"), val = tensor([512, 512, 512])]; + tensor var_1792_axis_0 = const()[name = tensor("op_1792_axis_0"), val = tensor(-1)]; + tensor var_1792_cast_fp16_0, tensor var_1792_cast_fp16_1, tensor var_1792_cast_fp16_2 = split(axis = var_1792_axis_0, split_sizes = var_1792_split_sizes_0, x = linear_116_cast_fp16)[name = tensor("op_1792_cast_fp16")]; + tensor mean_61_axes_0 = const()[name = tensor("mean_61_axes_0"), val = tensor([-1])]; + tensor mean_61_keep_dims_0 = const()[name = tensor("mean_61_keep_dims_0"), val = tensor(true)]; + tensor mean_61_cast_fp16 = reduce_mean(axes = mean_61_axes_0, keep_dims = mean_61_keep_dims_0, x = x_193_cast_fp16)[name = tensor("mean_61_cast_fp16")]; + tensor sub_50_cast_fp16 = sub(x = x_193_cast_fp16, y = mean_61_cast_fp16)[name = tensor("sub_50_cast_fp16")]; + tensor square_40_cast_fp16 = square(x = sub_50_cast_fp16)[name = tensor("square_40_cast_fp16")]; + tensor reduce_mean_81_axes_0 = const()[name = tensor("reduce_mean_81_axes_0"), val = tensor([-1])]; + tensor reduce_mean_81_keep_dims_0 = const()[name = tensor("reduce_mean_81_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_81_cast_fp16 = reduce_mean(axes = reduce_mean_81_axes_0, keep_dims = reduce_mean_81_keep_dims_0, x = square_40_cast_fp16)[name = tensor("reduce_mean_81_cast_fp16")]; + tensor var_1802_to_fp16 = const()[name = tensor("op_1802_to_fp16"), val = tensor(0x1.1p-20)]; + tensor var_1803_cast_fp16 = add(x = reduce_mean_81_cast_fp16, y = var_1802_to_fp16)[name = tensor("op_1803_cast_fp16")]; + tensor var_1804_cast_fp16 = sqrt(x = var_1803_cast_fp16)[name = tensor("op_1804_cast_fp16")]; + tensor x_195_cast_fp16 = real_div(x = sub_50_cast_fp16, y = var_1804_cast_fp16)[name = tensor("x_195_cast_fp16")]; + tensor var_1806_cast_fp16 = mul(x = x_195_cast_fp16, y = flow_net_res_blocks_2_in_ln_weight_to_fp16)[name = tensor("op_1806_cast_fp16")]; + tensor x_197_cast_fp16 = add(x = var_1806_cast_fp16, y = flow_net_res_blocks_2_in_ln_bias_to_fp16)[name = tensor("x_197_cast_fp16")]; + tensor var_1808_promoted_to_fp16 = const()[name = tensor("op_1808_promoted_to_fp16"), val = tensor(0x1p+0)]; + tensor var_1809_cast_fp16 = add(x = var_1792_cast_fp16_1, y = var_1808_promoted_to_fp16)[name = tensor("op_1809_cast_fp16")]; + tensor var_1810_cast_fp16 = mul(x = x_197_cast_fp16, y = var_1809_cast_fp16)[name = tensor("op_1810_cast_fp16")]; + tensor input_305_cast_fp16 = add(x = var_1810_cast_fp16, y = var_1792_cast_fp16_0)[name = tensor("input_305_cast_fp16")]; + tensor linear_117_cast_fp16 = linear(bias = flow_net_res_blocks_2_mlp_0_bias_to_fp16, weight = flow_net_res_blocks_2_mlp_0_weight_to_fp16, x = input_305_cast_fp16)[name = tensor("linear_117_cast_fp16")]; + tensor input_309_cast_fp16 = silu(x = linear_117_cast_fp16)[name = tensor("input_309_cast_fp16")]; + tensor linear_118_cast_fp16 = linear(bias = flow_net_res_blocks_2_mlp_2_bias_to_fp16, weight = flow_net_res_blocks_2_mlp_2_weight_to_fp16, x = input_309_cast_fp16)[name = tensor("linear_118_cast_fp16")]; + tensor var_1821_cast_fp16 = mul(x = var_1792_cast_fp16_2, y = linear_118_cast_fp16)[name = tensor("op_1821_cast_fp16")]; + tensor x_199_cast_fp16 = add(x = x_193_cast_fp16, y = var_1821_cast_fp16)[name = tensor("x_199_cast_fp16")]; + tensor linear_119_cast_fp16 = linear(bias = flow_net_res_blocks_3_adaLN_modulation_1_bias_to_fp16, weight = flow_net_res_blocks_3_adaLN_modulation_1_weight_to_fp16, x = input_287_cast_fp16)[name = tensor("linear_119_cast_fp16")]; + tensor var_1831_split_sizes_0 = const()[name = tensor("op_1831_split_sizes_0"), val = tensor([512, 512, 512])]; + tensor var_1831_axis_0 = const()[name = tensor("op_1831_axis_0"), val = tensor(-1)]; + tensor var_1831_cast_fp16_0, tensor var_1831_cast_fp16_1, tensor var_1831_cast_fp16_2 = split(axis = var_1831_axis_0, split_sizes = var_1831_split_sizes_0, x = linear_119_cast_fp16)[name = tensor("op_1831_cast_fp16")]; + tensor mean_63_axes_0 = const()[name = tensor("mean_63_axes_0"), val = tensor([-1])]; + tensor mean_63_keep_dims_0 = const()[name = tensor("mean_63_keep_dims_0"), val = tensor(true)]; + tensor mean_63_cast_fp16 = reduce_mean(axes = mean_63_axes_0, keep_dims = mean_63_keep_dims_0, x = x_199_cast_fp16)[name = tensor("mean_63_cast_fp16")]; + tensor sub_51_cast_fp16 = sub(x = x_199_cast_fp16, y = mean_63_cast_fp16)[name = tensor("sub_51_cast_fp16")]; + tensor square_41_cast_fp16 = square(x = sub_51_cast_fp16)[name = tensor("square_41_cast_fp16")]; + tensor reduce_mean_83_axes_0 = const()[name = tensor("reduce_mean_83_axes_0"), val = tensor([-1])]; + tensor reduce_mean_83_keep_dims_0 = const()[name = tensor("reduce_mean_83_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_83_cast_fp16 = reduce_mean(axes = reduce_mean_83_axes_0, keep_dims = reduce_mean_83_keep_dims_0, x = square_41_cast_fp16)[name = tensor("reduce_mean_83_cast_fp16")]; + tensor var_1841_to_fp16 = const()[name = tensor("op_1841_to_fp16"), val = tensor(0x1.1p-20)]; + tensor var_1842_cast_fp16 = add(x = reduce_mean_83_cast_fp16, y = var_1841_to_fp16)[name = tensor("op_1842_cast_fp16")]; + tensor var_1843_cast_fp16 = sqrt(x = var_1842_cast_fp16)[name = tensor("op_1843_cast_fp16")]; + tensor x_201_cast_fp16 = real_div(x = sub_51_cast_fp16, y = var_1843_cast_fp16)[name = tensor("x_201_cast_fp16")]; + tensor var_1845_cast_fp16 = mul(x = x_201_cast_fp16, y = flow_net_res_blocks_3_in_ln_weight_to_fp16)[name = tensor("op_1845_cast_fp16")]; + tensor x_203_cast_fp16 = add(x = var_1845_cast_fp16, y = flow_net_res_blocks_3_in_ln_bias_to_fp16)[name = tensor("x_203_cast_fp16")]; + tensor var_1847_promoted_to_fp16 = const()[name = tensor("op_1847_promoted_to_fp16"), val = tensor(0x1p+0)]; + tensor var_1848_cast_fp16 = add(x = var_1831_cast_fp16_1, y = var_1847_promoted_to_fp16)[name = tensor("op_1848_cast_fp16")]; + tensor var_1849_cast_fp16 = mul(x = x_203_cast_fp16, y = var_1848_cast_fp16)[name = tensor("op_1849_cast_fp16")]; + tensor input_313_cast_fp16 = add(x = var_1849_cast_fp16, y = var_1831_cast_fp16_0)[name = tensor("input_313_cast_fp16")]; + tensor linear_120_cast_fp16 = linear(bias = flow_net_res_blocks_3_mlp_0_bias_to_fp16, weight = flow_net_res_blocks_3_mlp_0_weight_to_fp16, x = input_313_cast_fp16)[name = tensor("linear_120_cast_fp16")]; + tensor input_317_cast_fp16 = silu(x = linear_120_cast_fp16)[name = tensor("input_317_cast_fp16")]; + tensor linear_121_cast_fp16 = linear(bias = flow_net_res_blocks_3_mlp_2_bias_to_fp16, weight = flow_net_res_blocks_3_mlp_2_weight_to_fp16, x = input_317_cast_fp16)[name = tensor("linear_121_cast_fp16")]; + tensor var_1860_cast_fp16 = mul(x = var_1831_cast_fp16_2, y = linear_121_cast_fp16)[name = tensor("op_1860_cast_fp16")]; + tensor x_205_cast_fp16 = add(x = x_199_cast_fp16, y = var_1860_cast_fp16)[name = tensor("x_205_cast_fp16")]; + tensor linear_122_cast_fp16 = linear(bias = flow_net_res_blocks_4_adaLN_modulation_1_bias_to_fp16, weight = flow_net_res_blocks_4_adaLN_modulation_1_weight_to_fp16, x = input_287_cast_fp16)[name = tensor("linear_122_cast_fp16")]; + tensor var_1870_split_sizes_0 = const()[name = tensor("op_1870_split_sizes_0"), val = tensor([512, 512, 512])]; + tensor var_1870_axis_0 = const()[name = tensor("op_1870_axis_0"), val = tensor(-1)]; + tensor var_1870_cast_fp16_0, tensor var_1870_cast_fp16_1, tensor var_1870_cast_fp16_2 = split(axis = var_1870_axis_0, split_sizes = var_1870_split_sizes_0, x = linear_122_cast_fp16)[name = tensor("op_1870_cast_fp16")]; + tensor mean_65_axes_0 = const()[name = tensor("mean_65_axes_0"), val = tensor([-1])]; + tensor mean_65_keep_dims_0 = const()[name = tensor("mean_65_keep_dims_0"), val = tensor(true)]; + tensor mean_65_cast_fp16 = reduce_mean(axes = mean_65_axes_0, keep_dims = mean_65_keep_dims_0, x = x_205_cast_fp16)[name = tensor("mean_65_cast_fp16")]; + tensor sub_52_cast_fp16 = sub(x = x_205_cast_fp16, y = mean_65_cast_fp16)[name = tensor("sub_52_cast_fp16")]; + tensor square_42_cast_fp16 = square(x = sub_52_cast_fp16)[name = tensor("square_42_cast_fp16")]; + tensor reduce_mean_85_axes_0 = const()[name = tensor("reduce_mean_85_axes_0"), val = tensor([-1])]; + tensor reduce_mean_85_keep_dims_0 = const()[name = tensor("reduce_mean_85_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_85_cast_fp16 = reduce_mean(axes = reduce_mean_85_axes_0, keep_dims = reduce_mean_85_keep_dims_0, x = square_42_cast_fp16)[name = tensor("reduce_mean_85_cast_fp16")]; + tensor var_1880_to_fp16 = const()[name = tensor("op_1880_to_fp16"), val = tensor(0x1.1p-20)]; + tensor var_1881_cast_fp16 = add(x = reduce_mean_85_cast_fp16, y = var_1880_to_fp16)[name = tensor("op_1881_cast_fp16")]; + tensor var_1882_cast_fp16 = sqrt(x = var_1881_cast_fp16)[name = tensor("op_1882_cast_fp16")]; + tensor x_207_cast_fp16 = real_div(x = sub_52_cast_fp16, y = var_1882_cast_fp16)[name = tensor("x_207_cast_fp16")]; + tensor var_1884_cast_fp16 = mul(x = x_207_cast_fp16, y = flow_net_res_blocks_4_in_ln_weight_to_fp16)[name = tensor("op_1884_cast_fp16")]; + tensor x_209_cast_fp16 = add(x = var_1884_cast_fp16, y = flow_net_res_blocks_4_in_ln_bias_to_fp16)[name = tensor("x_209_cast_fp16")]; + tensor var_1886_promoted_to_fp16 = const()[name = tensor("op_1886_promoted_to_fp16"), val = tensor(0x1p+0)]; + tensor var_1887_cast_fp16 = add(x = var_1870_cast_fp16_1, y = var_1886_promoted_to_fp16)[name = tensor("op_1887_cast_fp16")]; + tensor var_1888_cast_fp16 = mul(x = x_209_cast_fp16, y = var_1887_cast_fp16)[name = tensor("op_1888_cast_fp16")]; + tensor input_321_cast_fp16 = add(x = var_1888_cast_fp16, y = var_1870_cast_fp16_0)[name = tensor("input_321_cast_fp16")]; + tensor linear_123_cast_fp16 = linear(bias = flow_net_res_blocks_4_mlp_0_bias_to_fp16, weight = flow_net_res_blocks_4_mlp_0_weight_to_fp16, x = input_321_cast_fp16)[name = tensor("linear_123_cast_fp16")]; + tensor input_325_cast_fp16 = silu(x = linear_123_cast_fp16)[name = tensor("input_325_cast_fp16")]; + tensor linear_124_cast_fp16 = linear(bias = flow_net_res_blocks_4_mlp_2_bias_to_fp16, weight = flow_net_res_blocks_4_mlp_2_weight_to_fp16, x = input_325_cast_fp16)[name = tensor("linear_124_cast_fp16")]; + tensor var_1899_cast_fp16 = mul(x = var_1870_cast_fp16_2, y = linear_124_cast_fp16)[name = tensor("op_1899_cast_fp16")]; + tensor x_211_cast_fp16 = add(x = x_205_cast_fp16, y = var_1899_cast_fp16)[name = tensor("x_211_cast_fp16")]; + tensor linear_125_cast_fp16 = linear(bias = flow_net_res_blocks_5_adaLN_modulation_1_bias_to_fp16, weight = flow_net_res_blocks_5_adaLN_modulation_1_weight_to_fp16, x = input_287_cast_fp16)[name = tensor("linear_125_cast_fp16")]; + tensor var_1909_split_sizes_0 = const()[name = tensor("op_1909_split_sizes_0"), val = tensor([512, 512, 512])]; + tensor var_1909_axis_0 = const()[name = tensor("op_1909_axis_0"), val = tensor(-1)]; + tensor var_1909_cast_fp16_0, tensor var_1909_cast_fp16_1, tensor var_1909_cast_fp16_2 = split(axis = var_1909_axis_0, split_sizes = var_1909_split_sizes_0, x = linear_125_cast_fp16)[name = tensor("op_1909_cast_fp16")]; + tensor mean_67_axes_0 = const()[name = tensor("mean_67_axes_0"), val = tensor([-1])]; + tensor mean_67_keep_dims_0 = const()[name = tensor("mean_67_keep_dims_0"), val = tensor(true)]; + tensor mean_67_cast_fp16 = reduce_mean(axes = mean_67_axes_0, keep_dims = mean_67_keep_dims_0, x = x_211_cast_fp16)[name = tensor("mean_67_cast_fp16")]; + tensor sub_53_cast_fp16 = sub(x = x_211_cast_fp16, y = mean_67_cast_fp16)[name = tensor("sub_53_cast_fp16")]; + tensor square_43_cast_fp16 = square(x = sub_53_cast_fp16)[name = tensor("square_43_cast_fp16")]; + tensor reduce_mean_87_axes_0 = const()[name = tensor("reduce_mean_87_axes_0"), val = tensor([-1])]; + tensor reduce_mean_87_keep_dims_0 = const()[name = tensor("reduce_mean_87_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_87_cast_fp16 = reduce_mean(axes = reduce_mean_87_axes_0, keep_dims = reduce_mean_87_keep_dims_0, x = square_43_cast_fp16)[name = tensor("reduce_mean_87_cast_fp16")]; + tensor var_1919_to_fp16 = const()[name = tensor("op_1919_to_fp16"), val = tensor(0x1.1p-20)]; + tensor var_1920_cast_fp16 = add(x = reduce_mean_87_cast_fp16, y = var_1919_to_fp16)[name = tensor("op_1920_cast_fp16")]; + tensor var_1921_cast_fp16 = sqrt(x = var_1920_cast_fp16)[name = tensor("op_1921_cast_fp16")]; + tensor x_213_cast_fp16 = real_div(x = sub_53_cast_fp16, y = var_1921_cast_fp16)[name = tensor("x_213_cast_fp16")]; + tensor var_1923_cast_fp16 = mul(x = x_213_cast_fp16, y = flow_net_res_blocks_5_in_ln_weight_to_fp16)[name = tensor("op_1923_cast_fp16")]; + tensor x_215_cast_fp16 = add(x = var_1923_cast_fp16, y = flow_net_res_blocks_5_in_ln_bias_to_fp16)[name = tensor("x_215_cast_fp16")]; + tensor var_1925_promoted_to_fp16 = const()[name = tensor("op_1925_promoted_to_fp16"), val = tensor(0x1p+0)]; + tensor var_1926_cast_fp16 = add(x = var_1909_cast_fp16_1, y = var_1925_promoted_to_fp16)[name = tensor("op_1926_cast_fp16")]; + tensor var_1927_cast_fp16 = mul(x = x_215_cast_fp16, y = var_1926_cast_fp16)[name = tensor("op_1927_cast_fp16")]; + tensor input_329_cast_fp16 = add(x = var_1927_cast_fp16, y = var_1909_cast_fp16_0)[name = tensor("input_329_cast_fp16")]; + tensor linear_126_cast_fp16 = linear(bias = flow_net_res_blocks_5_mlp_0_bias_to_fp16, weight = flow_net_res_blocks_5_mlp_0_weight_to_fp16, x = input_329_cast_fp16)[name = tensor("linear_126_cast_fp16")]; + tensor input_333_cast_fp16 = silu(x = linear_126_cast_fp16)[name = tensor("input_333_cast_fp16")]; + tensor linear_127_cast_fp16 = linear(bias = flow_net_res_blocks_5_mlp_2_bias_to_fp16, weight = flow_net_res_blocks_5_mlp_2_weight_to_fp16, x = input_333_cast_fp16)[name = tensor("linear_127_cast_fp16")]; + tensor var_1938_cast_fp16 = mul(x = var_1909_cast_fp16_2, y = linear_127_cast_fp16)[name = tensor("op_1938_cast_fp16")]; + tensor x_217_cast_fp16 = add(x = x_211_cast_fp16, y = var_1938_cast_fp16)[name = tensor("x_217_cast_fp16")]; + tensor linear_128_cast_fp16 = linear(bias = flow_net_final_layer_adaLN_modulation_1_bias_to_fp16, weight = flow_net_final_layer_adaLN_modulation_1_weight_to_fp16, x = input_287_cast_fp16)[name = tensor("linear_128_cast_fp16")]; + tensor var_1947_split_sizes_0 = const()[name = tensor("op_1947_split_sizes_0"), val = tensor([512, 512])]; + tensor var_1947_axis_0 = const()[name = tensor("op_1947_axis_0"), val = tensor(-1)]; + tensor var_1947_cast_fp16_0, tensor var_1947_cast_fp16_1 = split(axis = var_1947_axis_0, split_sizes = var_1947_split_sizes_0, x = linear_128_cast_fp16)[name = tensor("op_1947_cast_fp16")]; + tensor mean_69_axes_0 = const()[name = tensor("mean_69_axes_0"), val = tensor([-1])]; + tensor mean_69_keep_dims_0 = const()[name = tensor("mean_69_keep_dims_0"), val = tensor(true)]; + tensor mean_69_cast_fp16 = reduce_mean(axes = mean_69_axes_0, keep_dims = mean_69_keep_dims_0, x = x_217_cast_fp16)[name = tensor("mean_69_cast_fp16")]; + tensor sub_54_cast_fp16 = sub(x = x_217_cast_fp16, y = mean_69_cast_fp16)[name = tensor("sub_54_cast_fp16")]; + tensor square_44_cast_fp16 = square(x = sub_54_cast_fp16)[name = tensor("square_44_cast_fp16")]; + tensor reduce_mean_89_axes_0 = const()[name = tensor("reduce_mean_89_axes_0"), val = tensor([-1])]; + tensor reduce_mean_89_keep_dims_0 = const()[name = tensor("reduce_mean_89_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_89_cast_fp16 = reduce_mean(axes = reduce_mean_89_axes_0, keep_dims = reduce_mean_89_keep_dims_0, x = square_44_cast_fp16)[name = tensor("reduce_mean_89_cast_fp16")]; + tensor var_1954_to_fp16 = const()[name = tensor("op_1954_to_fp16"), val = tensor(0x1.1p-20)]; + tensor var_1955_cast_fp16 = add(x = reduce_mean_89_cast_fp16, y = var_1954_to_fp16)[name = tensor("op_1955_cast_fp16")]; + tensor var_1956_cast_fp16 = sqrt(x = var_1955_cast_fp16)[name = tensor("op_1956_cast_fp16")]; + tensor x_219_cast_fp16 = real_div(x = sub_54_cast_fp16, y = var_1956_cast_fp16)[name = tensor("x_219_cast_fp16")]; + tensor var_1958_promoted_to_fp16 = const()[name = tensor("op_1958_promoted_to_fp16"), val = tensor(0x1p+0)]; + tensor var_1959_cast_fp16 = add(x = var_1947_cast_fp16_1, y = var_1958_promoted_to_fp16)[name = tensor("op_1959_cast_fp16")]; + tensor var_1960_cast_fp16 = mul(x = x_219_cast_fp16, y = var_1959_cast_fp16)[name = tensor("op_1960_cast_fp16")]; + tensor input_337_cast_fp16 = add(x = var_1960_cast_fp16, y = var_1947_cast_fp16_0)[name = tensor("input_337_cast_fp16")]; + tensor linear_129_cast_fp16 = linear(bias = flow_net_final_layer_linear_bias_to_fp16, weight = flow_net_final_layer_linear_weight_to_fp16, x = input_337_cast_fp16)[name = tensor("linear_129_cast_fp16")]; + tensor var_1971_to_fp16 = const()[name = tensor("op_1971_to_fp16"), val = tensor(0x1p-3)]; + tensor var_1972_cast_fp16 = mul(x = linear_129_cast_fp16, y = var_1971_to_fp16)[name = tensor("op_1972_cast_fp16")]; + tensor input_339_cast_fp16 = add(x = input_271_cast_fp16, y = var_1972_cast_fp16)[name = tensor("input_339_cast_fp16")]; + tensor linear_130_cast_fp16 = linear(bias = flow_net_input_proj_bias_to_fp16, weight = flow_net_input_proj_weight_to_fp16, x = input_339_cast_fp16)[name = tensor("linear_130_cast_fp16")]; + tensor input_343_to_fp16 = const()[name = tensor("input_343_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(19006080)))]; + tensor input_345_cast_fp16 = silu(x = input_343_to_fp16)[name = tensor("input_345_cast_fp16")]; + tensor linear_132_cast_fp16 = linear(bias = flow_net_time_embed_0_mlp_2_bias_to_fp16, weight = flow_net_time_embed_0_mlp_2_weight_to_fp16, x = input_345_cast_fp16)[name = tensor("linear_132_cast_fp16")]; + tensor reduce_mean_90_axes_0 = const()[name = tensor("reduce_mean_90_axes_0"), val = tensor([-1])]; + tensor reduce_mean_90_keep_dims_0 = const()[name = tensor("reduce_mean_90_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_90_cast_fp16 = reduce_mean(axes = reduce_mean_90_axes_0, keep_dims = reduce_mean_90_keep_dims_0, x = linear_132_cast_fp16)[name = tensor("reduce_mean_90_cast_fp16")]; + tensor sub_55_cast_fp16 = sub(x = linear_132_cast_fp16, y = reduce_mean_90_cast_fp16)[name = tensor("sub_55_cast_fp16")]; + tensor square_45_cast_fp16 = square(x = sub_55_cast_fp16)[name = tensor("square_45_cast_fp16")]; + tensor reduce_mean_91_axes_0 = const()[name = tensor("reduce_mean_91_axes_0"), val = tensor([-1])]; + tensor reduce_mean_91_keep_dims_0 = const()[name = tensor("reduce_mean_91_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_91_cast_fp16 = reduce_mean(axes = reduce_mean_91_axes_0, keep_dims = reduce_mean_91_keep_dims_0, x = square_45_cast_fp16)[name = tensor("reduce_mean_91_cast_fp16")]; + tensor real_div_10_to_fp16 = const()[name = tensor("real_div_10_to_fp16"), val = tensor(0x1.008p+0)]; + tensor mul_10_cast_fp16 = mul(x = reduce_mean_91_cast_fp16, y = real_div_10_to_fp16)[name = tensor("mul_10_cast_fp16")]; + tensor var_2038_to_fp16 = const()[name = tensor("op_2038_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_91_cast_fp16 = add(x = mul_10_cast_fp16, y = var_2038_to_fp16)[name = tensor("var_91_cast_fp16")]; + tensor var_2041_epsilon_0 = const()[name = tensor("op_2041_epsilon_0"), val = tensor(0x1.197998p-40)]; + tensor var_2041_cast_fp16 = rsqrt(epsilon = var_2041_epsilon_0, x = var_91_cast_fp16)[name = tensor("op_2041_cast_fp16")]; + tensor var_2042_cast_fp16 = mul(x = const_3_to_fp16, y = var_2041_cast_fp16)[name = tensor("op_2042_cast_fp16")]; + tensor var_2043_cast_fp16 = mul(x = linear_132_cast_fp16, y = var_2042_cast_fp16)[name = tensor("op_2043_cast_fp16")]; + tensor input_349_to_fp16 = const()[name = tensor("input_349_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(19007168)))]; + tensor input_351_cast_fp16 = silu(x = input_349_to_fp16)[name = tensor("input_351_cast_fp16")]; + tensor linear_134_cast_fp16 = linear(bias = flow_net_time_embed_1_mlp_2_bias_to_fp16, weight = flow_net_time_embed_1_mlp_2_weight_to_fp16, x = input_351_cast_fp16)[name = tensor("linear_134_cast_fp16")]; + tensor reduce_mean_92_axes_0 = const()[name = tensor("reduce_mean_92_axes_0"), val = tensor([-1])]; + tensor reduce_mean_92_keep_dims_0 = const()[name = tensor("reduce_mean_92_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_92_cast_fp16 = reduce_mean(axes = reduce_mean_92_axes_0, keep_dims = reduce_mean_92_keep_dims_0, x = linear_134_cast_fp16)[name = tensor("reduce_mean_92_cast_fp16")]; + tensor sub_57_cast_fp16 = sub(x = linear_134_cast_fp16, y = reduce_mean_92_cast_fp16)[name = tensor("sub_57_cast_fp16")]; + tensor square_46_cast_fp16 = square(x = sub_57_cast_fp16)[name = tensor("square_46_cast_fp16")]; + tensor reduce_mean_93_axes_0 = const()[name = tensor("reduce_mean_93_axes_0"), val = tensor([-1])]; + tensor reduce_mean_93_keep_dims_0 = const()[name = tensor("reduce_mean_93_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_93_cast_fp16 = reduce_mean(axes = reduce_mean_93_axes_0, keep_dims = reduce_mean_93_keep_dims_0, x = square_46_cast_fp16)[name = tensor("reduce_mean_93_cast_fp16")]; + tensor real_div_11_to_fp16 = const()[name = tensor("real_div_11_to_fp16"), val = tensor(0x1.008p+0)]; + tensor mul_11_cast_fp16 = mul(x = reduce_mean_93_cast_fp16, y = real_div_11_to_fp16)[name = tensor("mul_11_cast_fp16")]; + tensor var_2075_to_fp16 = const()[name = tensor("op_2075_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_93_cast_fp16 = add(x = mul_11_cast_fp16, y = var_2075_to_fp16)[name = tensor("var_93_cast_fp16")]; + tensor var_2078_epsilon_0 = const()[name = tensor("op_2078_epsilon_0"), val = tensor(0x1.197998p-40)]; + tensor var_2078_cast_fp16 = rsqrt(epsilon = var_2078_epsilon_0, x = var_93_cast_fp16)[name = tensor("op_2078_cast_fp16")]; + tensor var_2079_cast_fp16 = mul(x = const_5_to_fp16, y = var_2078_cast_fp16)[name = tensor("op_2079_cast_fp16")]; + tensor var_2080_cast_fp16 = mul(x = linear_134_cast_fp16, y = var_2079_cast_fp16)[name = tensor("op_2080_cast_fp16")]; + tensor var_2092_cast_fp16 = add(x = var_2043_cast_fp16, y = var_2080_cast_fp16)[name = tensor("op_2092_cast_fp16")]; + tensor _inversed_t_combined_11_y_0_to_fp16 = const()[name = tensor("_inversed_t_combined_11_y_0_to_fp16"), val = tensor(0x1p-1)]; + tensor _inversed_t_combined_11_cast_fp16 = mul(x = var_2092_cast_fp16, y = _inversed_t_combined_11_y_0_to_fp16)[name = tensor("_inversed_t_combined_11_cast_fp16")]; + tensor input_353_cast_fp16 = add(x = _inversed_t_combined_11_cast_fp16, y = linear_5_cast_fp16)[name = tensor("input_353_cast_fp16")]; + tensor input_355_cast_fp16 = silu(x = input_353_cast_fp16)[name = tensor("input_355_cast_fp16")]; + tensor linear_136_cast_fp16 = linear(bias = flow_net_res_blocks_0_adaLN_modulation_1_bias_to_fp16, weight = flow_net_res_blocks_0_adaLN_modulation_1_weight_to_fp16, x = input_355_cast_fp16)[name = tensor("linear_136_cast_fp16")]; + tensor var_2107_split_sizes_0 = const()[name = tensor("op_2107_split_sizes_0"), val = tensor([512, 512, 512])]; + tensor var_2107_axis_0 = const()[name = tensor("op_2107_axis_0"), val = tensor(-1)]; + tensor var_2107_cast_fp16_0, tensor var_2107_cast_fp16_1, tensor var_2107_cast_fp16_2 = split(axis = var_2107_axis_0, split_sizes = var_2107_split_sizes_0, x = linear_136_cast_fp16)[name = tensor("op_2107_cast_fp16")]; + tensor mean_71_axes_0 = const()[name = tensor("mean_71_axes_0"), val = tensor([-1])]; + tensor mean_71_keep_dims_0 = const()[name = tensor("mean_71_keep_dims_0"), val = tensor(true)]; + tensor mean_71_cast_fp16 = reduce_mean(axes = mean_71_axes_0, keep_dims = mean_71_keep_dims_0, x = linear_130_cast_fp16)[name = tensor("mean_71_cast_fp16")]; + tensor sub_59_cast_fp16 = sub(x = linear_130_cast_fp16, y = mean_71_cast_fp16)[name = tensor("sub_59_cast_fp16")]; + tensor square_47_cast_fp16 = square(x = sub_59_cast_fp16)[name = tensor("square_47_cast_fp16")]; + tensor reduce_mean_95_axes_0 = const()[name = tensor("reduce_mean_95_axes_0"), val = tensor([-1])]; + tensor reduce_mean_95_keep_dims_0 = const()[name = tensor("reduce_mean_95_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_95_cast_fp16 = reduce_mean(axes = reduce_mean_95_axes_0, keep_dims = reduce_mean_95_keep_dims_0, x = square_47_cast_fp16)[name = tensor("reduce_mean_95_cast_fp16")]; + tensor var_2117_to_fp16 = const()[name = tensor("op_2117_to_fp16"), val = tensor(0x1.1p-20)]; + tensor var_2118_cast_fp16 = add(x = reduce_mean_95_cast_fp16, y = var_2117_to_fp16)[name = tensor("op_2118_cast_fp16")]; + tensor var_2119_cast_fp16 = sqrt(x = var_2118_cast_fp16)[name = tensor("op_2119_cast_fp16")]; + tensor x_227_cast_fp16 = real_div(x = sub_59_cast_fp16, y = var_2119_cast_fp16)[name = tensor("x_227_cast_fp16")]; + tensor var_2121_cast_fp16 = mul(x = x_227_cast_fp16, y = flow_net_res_blocks_0_in_ln_weight_to_fp16)[name = tensor("op_2121_cast_fp16")]; + tensor x_229_cast_fp16 = add(x = var_2121_cast_fp16, y = flow_net_res_blocks_0_in_ln_bias_to_fp16)[name = tensor("x_229_cast_fp16")]; + tensor var_2123_promoted_to_fp16 = const()[name = tensor("op_2123_promoted_to_fp16"), val = tensor(0x1p+0)]; + tensor var_2124_cast_fp16 = add(x = var_2107_cast_fp16_1, y = var_2123_promoted_to_fp16)[name = tensor("op_2124_cast_fp16")]; + tensor var_2125_cast_fp16 = mul(x = x_229_cast_fp16, y = var_2124_cast_fp16)[name = tensor("op_2125_cast_fp16")]; + tensor input_357_cast_fp16 = add(x = var_2125_cast_fp16, y = var_2107_cast_fp16_0)[name = tensor("input_357_cast_fp16")]; + tensor linear_137_cast_fp16 = linear(bias = flow_net_res_blocks_0_mlp_0_bias_to_fp16, weight = flow_net_res_blocks_0_mlp_0_weight_to_fp16, x = input_357_cast_fp16)[name = tensor("linear_137_cast_fp16")]; + tensor input_361_cast_fp16 = silu(x = linear_137_cast_fp16)[name = tensor("input_361_cast_fp16")]; + tensor linear_138_cast_fp16 = linear(bias = flow_net_res_blocks_0_mlp_2_bias_to_fp16, weight = flow_net_res_blocks_0_mlp_2_weight_to_fp16, x = input_361_cast_fp16)[name = tensor("linear_138_cast_fp16")]; + tensor var_2136_cast_fp16 = mul(x = var_2107_cast_fp16_2, y = linear_138_cast_fp16)[name = tensor("op_2136_cast_fp16")]; + tensor x_231_cast_fp16 = add(x = linear_130_cast_fp16, y = var_2136_cast_fp16)[name = tensor("x_231_cast_fp16")]; + tensor linear_139_cast_fp16 = linear(bias = flow_net_res_blocks_1_adaLN_modulation_1_bias_to_fp16, weight = flow_net_res_blocks_1_adaLN_modulation_1_weight_to_fp16, x = input_355_cast_fp16)[name = tensor("linear_139_cast_fp16")]; + tensor var_2146_split_sizes_0 = const()[name = tensor("op_2146_split_sizes_0"), val = tensor([512, 512, 512])]; + tensor var_2146_axis_0 = const()[name = tensor("op_2146_axis_0"), val = tensor(-1)]; + tensor var_2146_cast_fp16_0, tensor var_2146_cast_fp16_1, tensor var_2146_cast_fp16_2 = split(axis = var_2146_axis_0, split_sizes = var_2146_split_sizes_0, x = linear_139_cast_fp16)[name = tensor("op_2146_cast_fp16")]; + tensor mean_73_axes_0 = const()[name = tensor("mean_73_axes_0"), val = tensor([-1])]; + tensor mean_73_keep_dims_0 = const()[name = tensor("mean_73_keep_dims_0"), val = tensor(true)]; + tensor mean_73_cast_fp16 = reduce_mean(axes = mean_73_axes_0, keep_dims = mean_73_keep_dims_0, x = x_231_cast_fp16)[name = tensor("mean_73_cast_fp16")]; + tensor sub_60_cast_fp16 = sub(x = x_231_cast_fp16, y = mean_73_cast_fp16)[name = tensor("sub_60_cast_fp16")]; + tensor square_48_cast_fp16 = square(x = sub_60_cast_fp16)[name = tensor("square_48_cast_fp16")]; + tensor reduce_mean_97_axes_0 = const()[name = tensor("reduce_mean_97_axes_0"), val = tensor([-1])]; + tensor reduce_mean_97_keep_dims_0 = const()[name = tensor("reduce_mean_97_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_97_cast_fp16 = reduce_mean(axes = reduce_mean_97_axes_0, keep_dims = reduce_mean_97_keep_dims_0, x = square_48_cast_fp16)[name = tensor("reduce_mean_97_cast_fp16")]; + tensor var_2156_to_fp16 = const()[name = tensor("op_2156_to_fp16"), val = tensor(0x1.1p-20)]; + tensor var_2157_cast_fp16 = add(x = reduce_mean_97_cast_fp16, y = var_2156_to_fp16)[name = tensor("op_2157_cast_fp16")]; + tensor var_2158_cast_fp16 = sqrt(x = var_2157_cast_fp16)[name = tensor("op_2158_cast_fp16")]; + tensor x_233_cast_fp16 = real_div(x = sub_60_cast_fp16, y = var_2158_cast_fp16)[name = tensor("x_233_cast_fp16")]; + tensor var_2160_cast_fp16 = mul(x = x_233_cast_fp16, y = flow_net_res_blocks_1_in_ln_weight_to_fp16)[name = tensor("op_2160_cast_fp16")]; + tensor x_235_cast_fp16 = add(x = var_2160_cast_fp16, y = flow_net_res_blocks_1_in_ln_bias_to_fp16)[name = tensor("x_235_cast_fp16")]; + tensor var_2162_promoted_to_fp16 = const()[name = tensor("op_2162_promoted_to_fp16"), val = tensor(0x1p+0)]; + tensor var_2163_cast_fp16 = add(x = var_2146_cast_fp16_1, y = var_2162_promoted_to_fp16)[name = tensor("op_2163_cast_fp16")]; + tensor var_2164_cast_fp16 = mul(x = x_235_cast_fp16, y = var_2163_cast_fp16)[name = tensor("op_2164_cast_fp16")]; + tensor input_365_cast_fp16 = add(x = var_2164_cast_fp16, y = var_2146_cast_fp16_0)[name = tensor("input_365_cast_fp16")]; + tensor linear_140_cast_fp16 = linear(bias = flow_net_res_blocks_1_mlp_0_bias_to_fp16, weight = flow_net_res_blocks_1_mlp_0_weight_to_fp16, x = input_365_cast_fp16)[name = tensor("linear_140_cast_fp16")]; + tensor input_369_cast_fp16 = silu(x = linear_140_cast_fp16)[name = tensor("input_369_cast_fp16")]; + tensor linear_141_cast_fp16 = linear(bias = flow_net_res_blocks_1_mlp_2_bias_to_fp16, weight = flow_net_res_blocks_1_mlp_2_weight_to_fp16, x = input_369_cast_fp16)[name = tensor("linear_141_cast_fp16")]; + tensor var_2175_cast_fp16 = mul(x = var_2146_cast_fp16_2, y = linear_141_cast_fp16)[name = tensor("op_2175_cast_fp16")]; + tensor x_237_cast_fp16 = add(x = x_231_cast_fp16, y = var_2175_cast_fp16)[name = tensor("x_237_cast_fp16")]; + tensor linear_142_cast_fp16 = linear(bias = flow_net_res_blocks_2_adaLN_modulation_1_bias_to_fp16, weight = flow_net_res_blocks_2_adaLN_modulation_1_weight_to_fp16, x = input_355_cast_fp16)[name = tensor("linear_142_cast_fp16")]; + tensor var_2185_split_sizes_0 = const()[name = tensor("op_2185_split_sizes_0"), val = tensor([512, 512, 512])]; + tensor var_2185_axis_0 = const()[name = tensor("op_2185_axis_0"), val = tensor(-1)]; + tensor var_2185_cast_fp16_0, tensor var_2185_cast_fp16_1, tensor var_2185_cast_fp16_2 = split(axis = var_2185_axis_0, split_sizes = var_2185_split_sizes_0, x = linear_142_cast_fp16)[name = tensor("op_2185_cast_fp16")]; + tensor mean_75_axes_0 = const()[name = tensor("mean_75_axes_0"), val = tensor([-1])]; + tensor mean_75_keep_dims_0 = const()[name = tensor("mean_75_keep_dims_0"), val = tensor(true)]; + tensor mean_75_cast_fp16 = reduce_mean(axes = mean_75_axes_0, keep_dims = mean_75_keep_dims_0, x = x_237_cast_fp16)[name = tensor("mean_75_cast_fp16")]; + tensor sub_61_cast_fp16 = sub(x = x_237_cast_fp16, y = mean_75_cast_fp16)[name = tensor("sub_61_cast_fp16")]; + tensor square_49_cast_fp16 = square(x = sub_61_cast_fp16)[name = tensor("square_49_cast_fp16")]; + tensor reduce_mean_99_axes_0 = const()[name = tensor("reduce_mean_99_axes_0"), val = tensor([-1])]; + tensor reduce_mean_99_keep_dims_0 = const()[name = tensor("reduce_mean_99_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_99_cast_fp16 = reduce_mean(axes = reduce_mean_99_axes_0, keep_dims = reduce_mean_99_keep_dims_0, x = square_49_cast_fp16)[name = tensor("reduce_mean_99_cast_fp16")]; + tensor var_2195_to_fp16 = const()[name = tensor("op_2195_to_fp16"), val = tensor(0x1.1p-20)]; + tensor var_2196_cast_fp16 = add(x = reduce_mean_99_cast_fp16, y = var_2195_to_fp16)[name = tensor("op_2196_cast_fp16")]; + tensor var_2197_cast_fp16 = sqrt(x = var_2196_cast_fp16)[name = tensor("op_2197_cast_fp16")]; + tensor x_239_cast_fp16 = real_div(x = sub_61_cast_fp16, y = var_2197_cast_fp16)[name = tensor("x_239_cast_fp16")]; + tensor var_2199_cast_fp16 = mul(x = x_239_cast_fp16, y = flow_net_res_blocks_2_in_ln_weight_to_fp16)[name = tensor("op_2199_cast_fp16")]; + tensor x_241_cast_fp16 = add(x = var_2199_cast_fp16, y = flow_net_res_blocks_2_in_ln_bias_to_fp16)[name = tensor("x_241_cast_fp16")]; + tensor var_2201_promoted_to_fp16 = const()[name = tensor("op_2201_promoted_to_fp16"), val = tensor(0x1p+0)]; + tensor var_2202_cast_fp16 = add(x = var_2185_cast_fp16_1, y = var_2201_promoted_to_fp16)[name = tensor("op_2202_cast_fp16")]; + tensor var_2203_cast_fp16 = mul(x = x_241_cast_fp16, y = var_2202_cast_fp16)[name = tensor("op_2203_cast_fp16")]; + tensor input_373_cast_fp16 = add(x = var_2203_cast_fp16, y = var_2185_cast_fp16_0)[name = tensor("input_373_cast_fp16")]; + tensor linear_143_cast_fp16 = linear(bias = flow_net_res_blocks_2_mlp_0_bias_to_fp16, weight = flow_net_res_blocks_2_mlp_0_weight_to_fp16, x = input_373_cast_fp16)[name = tensor("linear_143_cast_fp16")]; + tensor input_377_cast_fp16 = silu(x = linear_143_cast_fp16)[name = tensor("input_377_cast_fp16")]; + tensor linear_144_cast_fp16 = linear(bias = flow_net_res_blocks_2_mlp_2_bias_to_fp16, weight = flow_net_res_blocks_2_mlp_2_weight_to_fp16, x = input_377_cast_fp16)[name = tensor("linear_144_cast_fp16")]; + tensor var_2214_cast_fp16 = mul(x = var_2185_cast_fp16_2, y = linear_144_cast_fp16)[name = tensor("op_2214_cast_fp16")]; + tensor x_243_cast_fp16 = add(x = x_237_cast_fp16, y = var_2214_cast_fp16)[name = tensor("x_243_cast_fp16")]; + tensor linear_145_cast_fp16 = linear(bias = flow_net_res_blocks_3_adaLN_modulation_1_bias_to_fp16, weight = flow_net_res_blocks_3_adaLN_modulation_1_weight_to_fp16, x = input_355_cast_fp16)[name = tensor("linear_145_cast_fp16")]; + tensor var_2224_split_sizes_0 = const()[name = tensor("op_2224_split_sizes_0"), val = tensor([512, 512, 512])]; + tensor var_2224_axis_0 = const()[name = tensor("op_2224_axis_0"), val = tensor(-1)]; + tensor var_2224_cast_fp16_0, tensor var_2224_cast_fp16_1, tensor var_2224_cast_fp16_2 = split(axis = var_2224_axis_0, split_sizes = var_2224_split_sizes_0, x = linear_145_cast_fp16)[name = tensor("op_2224_cast_fp16")]; + tensor mean_77_axes_0 = const()[name = tensor("mean_77_axes_0"), val = tensor([-1])]; + tensor mean_77_keep_dims_0 = const()[name = tensor("mean_77_keep_dims_0"), val = tensor(true)]; + tensor mean_77_cast_fp16 = reduce_mean(axes = mean_77_axes_0, keep_dims = mean_77_keep_dims_0, x = x_243_cast_fp16)[name = tensor("mean_77_cast_fp16")]; + tensor sub_62_cast_fp16 = sub(x = x_243_cast_fp16, y = mean_77_cast_fp16)[name = tensor("sub_62_cast_fp16")]; + tensor square_50_cast_fp16 = square(x = sub_62_cast_fp16)[name = tensor("square_50_cast_fp16")]; + tensor reduce_mean_101_axes_0 = const()[name = tensor("reduce_mean_101_axes_0"), val = tensor([-1])]; + tensor reduce_mean_101_keep_dims_0 = const()[name = tensor("reduce_mean_101_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_101_cast_fp16 = reduce_mean(axes = reduce_mean_101_axes_0, keep_dims = reduce_mean_101_keep_dims_0, x = square_50_cast_fp16)[name = tensor("reduce_mean_101_cast_fp16")]; + tensor var_2234_to_fp16 = const()[name = tensor("op_2234_to_fp16"), val = tensor(0x1.1p-20)]; + tensor var_2235_cast_fp16 = add(x = reduce_mean_101_cast_fp16, y = var_2234_to_fp16)[name = tensor("op_2235_cast_fp16")]; + tensor var_2236_cast_fp16 = sqrt(x = var_2235_cast_fp16)[name = tensor("op_2236_cast_fp16")]; + tensor x_245_cast_fp16 = real_div(x = sub_62_cast_fp16, y = var_2236_cast_fp16)[name = tensor("x_245_cast_fp16")]; + tensor var_2238_cast_fp16 = mul(x = x_245_cast_fp16, y = flow_net_res_blocks_3_in_ln_weight_to_fp16)[name = tensor("op_2238_cast_fp16")]; + tensor x_247_cast_fp16 = add(x = var_2238_cast_fp16, y = flow_net_res_blocks_3_in_ln_bias_to_fp16)[name = tensor("x_247_cast_fp16")]; + tensor var_2240_promoted_to_fp16 = const()[name = tensor("op_2240_promoted_to_fp16"), val = tensor(0x1p+0)]; + tensor var_2241_cast_fp16 = add(x = var_2224_cast_fp16_1, y = var_2240_promoted_to_fp16)[name = tensor("op_2241_cast_fp16")]; + tensor var_2242_cast_fp16 = mul(x = x_247_cast_fp16, y = var_2241_cast_fp16)[name = tensor("op_2242_cast_fp16")]; + tensor input_381_cast_fp16 = add(x = var_2242_cast_fp16, y = var_2224_cast_fp16_0)[name = tensor("input_381_cast_fp16")]; + tensor linear_146_cast_fp16 = linear(bias = flow_net_res_blocks_3_mlp_0_bias_to_fp16, weight = flow_net_res_blocks_3_mlp_0_weight_to_fp16, x = input_381_cast_fp16)[name = tensor("linear_146_cast_fp16")]; + tensor input_385_cast_fp16 = silu(x = linear_146_cast_fp16)[name = tensor("input_385_cast_fp16")]; + tensor linear_147_cast_fp16 = linear(bias = flow_net_res_blocks_3_mlp_2_bias_to_fp16, weight = flow_net_res_blocks_3_mlp_2_weight_to_fp16, x = input_385_cast_fp16)[name = tensor("linear_147_cast_fp16")]; + tensor var_2253_cast_fp16 = mul(x = var_2224_cast_fp16_2, y = linear_147_cast_fp16)[name = tensor("op_2253_cast_fp16")]; + tensor x_249_cast_fp16 = add(x = x_243_cast_fp16, y = var_2253_cast_fp16)[name = tensor("x_249_cast_fp16")]; + tensor linear_148_cast_fp16 = linear(bias = flow_net_res_blocks_4_adaLN_modulation_1_bias_to_fp16, weight = flow_net_res_blocks_4_adaLN_modulation_1_weight_to_fp16, x = input_355_cast_fp16)[name = tensor("linear_148_cast_fp16")]; + tensor var_2263_split_sizes_0 = const()[name = tensor("op_2263_split_sizes_0"), val = tensor([512, 512, 512])]; + tensor var_2263_axis_0 = const()[name = tensor("op_2263_axis_0"), val = tensor(-1)]; + tensor var_2263_cast_fp16_0, tensor var_2263_cast_fp16_1, tensor var_2263_cast_fp16_2 = split(axis = var_2263_axis_0, split_sizes = var_2263_split_sizes_0, x = linear_148_cast_fp16)[name = tensor("op_2263_cast_fp16")]; + tensor mean_79_axes_0 = const()[name = tensor("mean_79_axes_0"), val = tensor([-1])]; + tensor mean_79_keep_dims_0 = const()[name = tensor("mean_79_keep_dims_0"), val = tensor(true)]; + tensor mean_79_cast_fp16 = reduce_mean(axes = mean_79_axes_0, keep_dims = mean_79_keep_dims_0, x = x_249_cast_fp16)[name = tensor("mean_79_cast_fp16")]; + tensor sub_63_cast_fp16 = sub(x = x_249_cast_fp16, y = mean_79_cast_fp16)[name = tensor("sub_63_cast_fp16")]; + tensor square_51_cast_fp16 = square(x = sub_63_cast_fp16)[name = tensor("square_51_cast_fp16")]; + tensor reduce_mean_103_axes_0 = const()[name = tensor("reduce_mean_103_axes_0"), val = tensor([-1])]; + tensor reduce_mean_103_keep_dims_0 = const()[name = tensor("reduce_mean_103_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_103_cast_fp16 = reduce_mean(axes = reduce_mean_103_axes_0, keep_dims = reduce_mean_103_keep_dims_0, x = square_51_cast_fp16)[name = tensor("reduce_mean_103_cast_fp16")]; + tensor var_2273_to_fp16 = const()[name = tensor("op_2273_to_fp16"), val = tensor(0x1.1p-20)]; + tensor var_2274_cast_fp16 = add(x = reduce_mean_103_cast_fp16, y = var_2273_to_fp16)[name = tensor("op_2274_cast_fp16")]; + tensor var_2275_cast_fp16 = sqrt(x = var_2274_cast_fp16)[name = tensor("op_2275_cast_fp16")]; + tensor x_251_cast_fp16 = real_div(x = sub_63_cast_fp16, y = var_2275_cast_fp16)[name = tensor("x_251_cast_fp16")]; + tensor var_2277_cast_fp16 = mul(x = x_251_cast_fp16, y = flow_net_res_blocks_4_in_ln_weight_to_fp16)[name = tensor("op_2277_cast_fp16")]; + tensor x_253_cast_fp16 = add(x = var_2277_cast_fp16, y = flow_net_res_blocks_4_in_ln_bias_to_fp16)[name = tensor("x_253_cast_fp16")]; + tensor var_2279_promoted_to_fp16 = const()[name = tensor("op_2279_promoted_to_fp16"), val = tensor(0x1p+0)]; + tensor var_2280_cast_fp16 = add(x = var_2263_cast_fp16_1, y = var_2279_promoted_to_fp16)[name = tensor("op_2280_cast_fp16")]; + tensor var_2281_cast_fp16 = mul(x = x_253_cast_fp16, y = var_2280_cast_fp16)[name = tensor("op_2281_cast_fp16")]; + tensor input_389_cast_fp16 = add(x = var_2281_cast_fp16, y = var_2263_cast_fp16_0)[name = tensor("input_389_cast_fp16")]; + tensor linear_149_cast_fp16 = linear(bias = flow_net_res_blocks_4_mlp_0_bias_to_fp16, weight = flow_net_res_blocks_4_mlp_0_weight_to_fp16, x = input_389_cast_fp16)[name = tensor("linear_149_cast_fp16")]; + tensor input_393_cast_fp16 = silu(x = linear_149_cast_fp16)[name = tensor("input_393_cast_fp16")]; + tensor linear_150_cast_fp16 = linear(bias = flow_net_res_blocks_4_mlp_2_bias_to_fp16, weight = flow_net_res_blocks_4_mlp_2_weight_to_fp16, x = input_393_cast_fp16)[name = tensor("linear_150_cast_fp16")]; + tensor var_2292_cast_fp16 = mul(x = var_2263_cast_fp16_2, y = linear_150_cast_fp16)[name = tensor("op_2292_cast_fp16")]; + tensor x_255_cast_fp16 = add(x = x_249_cast_fp16, y = var_2292_cast_fp16)[name = tensor("x_255_cast_fp16")]; + tensor linear_151_cast_fp16 = linear(bias = flow_net_res_blocks_5_adaLN_modulation_1_bias_to_fp16, weight = flow_net_res_blocks_5_adaLN_modulation_1_weight_to_fp16, x = input_355_cast_fp16)[name = tensor("linear_151_cast_fp16")]; + tensor var_2302_split_sizes_0 = const()[name = tensor("op_2302_split_sizes_0"), val = tensor([512, 512, 512])]; + tensor var_2302_axis_0 = const()[name = tensor("op_2302_axis_0"), val = tensor(-1)]; + tensor var_2302_cast_fp16_0, tensor var_2302_cast_fp16_1, tensor var_2302_cast_fp16_2 = split(axis = var_2302_axis_0, split_sizes = var_2302_split_sizes_0, x = linear_151_cast_fp16)[name = tensor("op_2302_cast_fp16")]; + tensor mean_81_axes_0 = const()[name = tensor("mean_81_axes_0"), val = tensor([-1])]; + tensor mean_81_keep_dims_0 = const()[name = tensor("mean_81_keep_dims_0"), val = tensor(true)]; + tensor mean_81_cast_fp16 = reduce_mean(axes = mean_81_axes_0, keep_dims = mean_81_keep_dims_0, x = x_255_cast_fp16)[name = tensor("mean_81_cast_fp16")]; + tensor sub_64_cast_fp16 = sub(x = x_255_cast_fp16, y = mean_81_cast_fp16)[name = tensor("sub_64_cast_fp16")]; + tensor square_52_cast_fp16 = square(x = sub_64_cast_fp16)[name = tensor("square_52_cast_fp16")]; + tensor reduce_mean_105_axes_0 = const()[name = tensor("reduce_mean_105_axes_0"), val = tensor([-1])]; + tensor reduce_mean_105_keep_dims_0 = const()[name = tensor("reduce_mean_105_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_105_cast_fp16 = reduce_mean(axes = reduce_mean_105_axes_0, keep_dims = reduce_mean_105_keep_dims_0, x = square_52_cast_fp16)[name = tensor("reduce_mean_105_cast_fp16")]; + tensor var_2312_to_fp16 = const()[name = tensor("op_2312_to_fp16"), val = tensor(0x1.1p-20)]; + tensor var_2313_cast_fp16 = add(x = reduce_mean_105_cast_fp16, y = var_2312_to_fp16)[name = tensor("op_2313_cast_fp16")]; + tensor var_2314_cast_fp16 = sqrt(x = var_2313_cast_fp16)[name = tensor("op_2314_cast_fp16")]; + tensor x_257_cast_fp16 = real_div(x = sub_64_cast_fp16, y = var_2314_cast_fp16)[name = tensor("x_257_cast_fp16")]; + tensor var_2316_cast_fp16 = mul(x = x_257_cast_fp16, y = flow_net_res_blocks_5_in_ln_weight_to_fp16)[name = tensor("op_2316_cast_fp16")]; + tensor x_259_cast_fp16 = add(x = var_2316_cast_fp16, y = flow_net_res_blocks_5_in_ln_bias_to_fp16)[name = tensor("x_259_cast_fp16")]; + tensor var_2318_promoted_to_fp16 = const()[name = tensor("op_2318_promoted_to_fp16"), val = tensor(0x1p+0)]; + tensor var_2319_cast_fp16 = add(x = var_2302_cast_fp16_1, y = var_2318_promoted_to_fp16)[name = tensor("op_2319_cast_fp16")]; + tensor var_2320_cast_fp16 = mul(x = x_259_cast_fp16, y = var_2319_cast_fp16)[name = tensor("op_2320_cast_fp16")]; + tensor input_397_cast_fp16 = add(x = var_2320_cast_fp16, y = var_2302_cast_fp16_0)[name = tensor("input_397_cast_fp16")]; + tensor linear_152_cast_fp16 = linear(bias = flow_net_res_blocks_5_mlp_0_bias_to_fp16, weight = flow_net_res_blocks_5_mlp_0_weight_to_fp16, x = input_397_cast_fp16)[name = tensor("linear_152_cast_fp16")]; + tensor input_401_cast_fp16 = silu(x = linear_152_cast_fp16)[name = tensor("input_401_cast_fp16")]; + tensor linear_153_cast_fp16 = linear(bias = flow_net_res_blocks_5_mlp_2_bias_to_fp16, weight = flow_net_res_blocks_5_mlp_2_weight_to_fp16, x = input_401_cast_fp16)[name = tensor("linear_153_cast_fp16")]; + tensor var_2331_cast_fp16 = mul(x = var_2302_cast_fp16_2, y = linear_153_cast_fp16)[name = tensor("op_2331_cast_fp16")]; + tensor x_261_cast_fp16 = add(x = x_255_cast_fp16, y = var_2331_cast_fp16)[name = tensor("x_261_cast_fp16")]; + tensor linear_154_cast_fp16 = linear(bias = flow_net_final_layer_adaLN_modulation_1_bias_to_fp16, weight = flow_net_final_layer_adaLN_modulation_1_weight_to_fp16, x = input_355_cast_fp16)[name = tensor("linear_154_cast_fp16")]; + tensor var_2340_split_sizes_0 = const()[name = tensor("op_2340_split_sizes_0"), val = tensor([512, 512])]; + tensor var_2340_axis_0 = const()[name = tensor("op_2340_axis_0"), val = tensor(-1)]; + tensor var_2340_cast_fp16_0, tensor var_2340_cast_fp16_1 = split(axis = var_2340_axis_0, split_sizes = var_2340_split_sizes_0, x = linear_154_cast_fp16)[name = tensor("op_2340_cast_fp16")]; + tensor mean_83_axes_0 = const()[name = tensor("mean_83_axes_0"), val = tensor([-1])]; + tensor mean_83_keep_dims_0 = const()[name = tensor("mean_83_keep_dims_0"), val = tensor(true)]; + tensor mean_83_cast_fp16 = reduce_mean(axes = mean_83_axes_0, keep_dims = mean_83_keep_dims_0, x = x_261_cast_fp16)[name = tensor("mean_83_cast_fp16")]; + tensor sub_65_cast_fp16 = sub(x = x_261_cast_fp16, y = mean_83_cast_fp16)[name = tensor("sub_65_cast_fp16")]; + tensor square_53_cast_fp16 = square(x = sub_65_cast_fp16)[name = tensor("square_53_cast_fp16")]; + tensor reduce_mean_107_axes_0 = const()[name = tensor("reduce_mean_107_axes_0"), val = tensor([-1])]; + tensor reduce_mean_107_keep_dims_0 = const()[name = tensor("reduce_mean_107_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_107_cast_fp16 = reduce_mean(axes = reduce_mean_107_axes_0, keep_dims = reduce_mean_107_keep_dims_0, x = square_53_cast_fp16)[name = tensor("reduce_mean_107_cast_fp16")]; + tensor var_2347_to_fp16 = const()[name = tensor("op_2347_to_fp16"), val = tensor(0x1.1p-20)]; + tensor var_2348_cast_fp16 = add(x = reduce_mean_107_cast_fp16, y = var_2347_to_fp16)[name = tensor("op_2348_cast_fp16")]; + tensor var_2349_cast_fp16 = sqrt(x = var_2348_cast_fp16)[name = tensor("op_2349_cast_fp16")]; + tensor x_263_cast_fp16 = real_div(x = sub_65_cast_fp16, y = var_2349_cast_fp16)[name = tensor("x_263_cast_fp16")]; + tensor var_2351_promoted_to_fp16 = const()[name = tensor("op_2351_promoted_to_fp16"), val = tensor(0x1p+0)]; + tensor var_2352_cast_fp16 = add(x = var_2340_cast_fp16_1, y = var_2351_promoted_to_fp16)[name = tensor("op_2352_cast_fp16")]; + tensor var_2353_cast_fp16 = mul(x = x_263_cast_fp16, y = var_2352_cast_fp16)[name = tensor("op_2353_cast_fp16")]; + tensor input_405_cast_fp16 = add(x = var_2353_cast_fp16, y = var_2340_cast_fp16_0)[name = tensor("input_405_cast_fp16")]; + tensor linear_155_cast_fp16 = linear(bias = flow_net_final_layer_linear_bias_to_fp16, weight = flow_net_final_layer_linear_weight_to_fp16, x = input_405_cast_fp16)[name = tensor("linear_155_cast_fp16")]; + tensor var_2364_to_fp16 = const()[name = tensor("op_2364_to_fp16"), val = tensor(0x1p-3)]; + tensor var_2365_cast_fp16 = mul(x = linear_155_cast_fp16, y = var_2364_to_fp16)[name = tensor("op_2365_cast_fp16")]; + tensor input_407_cast_fp16 = add(x = input_339_cast_fp16, y = var_2365_cast_fp16)[name = tensor("input_407_cast_fp16")]; + tensor linear_156_cast_fp16 = linear(bias = flow_net_input_proj_bias_to_fp16, weight = flow_net_input_proj_weight_to_fp16, x = input_407_cast_fp16)[name = tensor("linear_156_cast_fp16")]; + tensor input_411_to_fp16 = const()[name = tensor("input_411_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(19008256)))]; + tensor input_413_cast_fp16 = silu(x = input_411_to_fp16)[name = tensor("input_413_cast_fp16")]; + tensor linear_158_cast_fp16 = linear(bias = flow_net_time_embed_0_mlp_2_bias_to_fp16, weight = flow_net_time_embed_0_mlp_2_weight_to_fp16, x = input_413_cast_fp16)[name = tensor("linear_158_cast_fp16")]; + tensor reduce_mean_108_axes_0 = const()[name = tensor("reduce_mean_108_axes_0"), val = tensor([-1])]; + tensor reduce_mean_108_keep_dims_0 = const()[name = tensor("reduce_mean_108_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_108_cast_fp16 = reduce_mean(axes = reduce_mean_108_axes_0, keep_dims = reduce_mean_108_keep_dims_0, x = linear_158_cast_fp16)[name = tensor("reduce_mean_108_cast_fp16")]; + tensor sub_66_cast_fp16 = sub(x = linear_158_cast_fp16, y = reduce_mean_108_cast_fp16)[name = tensor("sub_66_cast_fp16")]; + tensor square_54_cast_fp16 = square(x = sub_66_cast_fp16)[name = tensor("square_54_cast_fp16")]; + tensor reduce_mean_109_axes_0 = const()[name = tensor("reduce_mean_109_axes_0"), val = tensor([-1])]; + tensor reduce_mean_109_keep_dims_0 = const()[name = tensor("reduce_mean_109_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_109_cast_fp16 = reduce_mean(axes = reduce_mean_109_axes_0, keep_dims = reduce_mean_109_keep_dims_0, x = square_54_cast_fp16)[name = tensor("reduce_mean_109_cast_fp16")]; + tensor real_div_12_to_fp16 = const()[name = tensor("real_div_12_to_fp16"), val = tensor(0x1.008p+0)]; + tensor mul_12_cast_fp16 = mul(x = reduce_mean_109_cast_fp16, y = real_div_12_to_fp16)[name = tensor("mul_12_cast_fp16")]; + tensor var_2431_to_fp16 = const()[name = tensor("op_2431_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_109_cast_fp16 = add(x = mul_12_cast_fp16, y = var_2431_to_fp16)[name = tensor("var_109_cast_fp16")]; + tensor var_2434_epsilon_0 = const()[name = tensor("op_2434_epsilon_0"), val = tensor(0x1.197998p-40)]; + tensor var_2434_cast_fp16 = rsqrt(epsilon = var_2434_epsilon_0, x = var_109_cast_fp16)[name = tensor("op_2434_cast_fp16")]; + tensor var_2435_cast_fp16 = mul(x = const_3_to_fp16, y = var_2434_cast_fp16)[name = tensor("op_2435_cast_fp16")]; + tensor var_2436_cast_fp16 = mul(x = linear_158_cast_fp16, y = var_2435_cast_fp16)[name = tensor("op_2436_cast_fp16")]; + tensor input_417_to_fp16 = const()[name = tensor("input_417_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(19009344)))]; + tensor input_419_cast_fp16 = silu(x = input_417_to_fp16)[name = tensor("input_419_cast_fp16")]; + tensor linear_160_cast_fp16 = linear(bias = flow_net_time_embed_1_mlp_2_bias_to_fp16, weight = flow_net_time_embed_1_mlp_2_weight_to_fp16, x = input_419_cast_fp16)[name = tensor("linear_160_cast_fp16")]; + tensor reduce_mean_110_axes_0 = const()[name = tensor("reduce_mean_110_axes_0"), val = tensor([-1])]; + tensor reduce_mean_110_keep_dims_0 = const()[name = tensor("reduce_mean_110_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_110_cast_fp16 = reduce_mean(axes = reduce_mean_110_axes_0, keep_dims = reduce_mean_110_keep_dims_0, x = linear_160_cast_fp16)[name = tensor("reduce_mean_110_cast_fp16")]; + tensor sub_68_cast_fp16 = sub(x = linear_160_cast_fp16, y = reduce_mean_110_cast_fp16)[name = tensor("sub_68_cast_fp16")]; + tensor square_55_cast_fp16 = square(x = sub_68_cast_fp16)[name = tensor("square_55_cast_fp16")]; + tensor reduce_mean_111_axes_0 = const()[name = tensor("reduce_mean_111_axes_0"), val = tensor([-1])]; + tensor reduce_mean_111_keep_dims_0 = const()[name = tensor("reduce_mean_111_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_111_cast_fp16 = reduce_mean(axes = reduce_mean_111_axes_0, keep_dims = reduce_mean_111_keep_dims_0, x = square_55_cast_fp16)[name = tensor("reduce_mean_111_cast_fp16")]; + tensor real_div_13_to_fp16 = const()[name = tensor("real_div_13_to_fp16"), val = tensor(0x1.008p+0)]; + tensor mul_13_cast_fp16 = mul(x = reduce_mean_111_cast_fp16, y = real_div_13_to_fp16)[name = tensor("mul_13_cast_fp16")]; + tensor var_2468_to_fp16 = const()[name = tensor("op_2468_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_111_cast_fp16 = add(x = mul_13_cast_fp16, y = var_2468_to_fp16)[name = tensor("var_111_cast_fp16")]; + tensor var_2471_epsilon_0 = const()[name = tensor("op_2471_epsilon_0"), val = tensor(0x1.197998p-40)]; + tensor var_2471_cast_fp16 = rsqrt(epsilon = var_2471_epsilon_0, x = var_111_cast_fp16)[name = tensor("op_2471_cast_fp16")]; + tensor var_2472_cast_fp16 = mul(x = const_5_to_fp16, y = var_2471_cast_fp16)[name = tensor("op_2472_cast_fp16")]; + tensor var_2473_cast_fp16 = mul(x = linear_160_cast_fp16, y = var_2472_cast_fp16)[name = tensor("op_2473_cast_fp16")]; + tensor var_2485_cast_fp16 = add(x = var_2436_cast_fp16, y = var_2473_cast_fp16)[name = tensor("op_2485_cast_fp16")]; + tensor _inversed_t_combined_13_y_0_to_fp16 = const()[name = tensor("_inversed_t_combined_13_y_0_to_fp16"), val = tensor(0x1p-1)]; + tensor _inversed_t_combined_13_cast_fp16 = mul(x = var_2485_cast_fp16, y = _inversed_t_combined_13_y_0_to_fp16)[name = tensor("_inversed_t_combined_13_cast_fp16")]; + tensor input_421_cast_fp16 = add(x = _inversed_t_combined_13_cast_fp16, y = linear_5_cast_fp16)[name = tensor("input_421_cast_fp16")]; + tensor input_423_cast_fp16 = silu(x = input_421_cast_fp16)[name = tensor("input_423_cast_fp16")]; + tensor linear_162_cast_fp16 = linear(bias = flow_net_res_blocks_0_adaLN_modulation_1_bias_to_fp16, weight = flow_net_res_blocks_0_adaLN_modulation_1_weight_to_fp16, x = input_423_cast_fp16)[name = tensor("linear_162_cast_fp16")]; + tensor var_2500_split_sizes_0 = const()[name = tensor("op_2500_split_sizes_0"), val = tensor([512, 512, 512])]; + tensor var_2500_axis_0 = const()[name = tensor("op_2500_axis_0"), val = tensor(-1)]; + tensor var_2500_cast_fp16_0, tensor var_2500_cast_fp16_1, tensor var_2500_cast_fp16_2 = split(axis = var_2500_axis_0, split_sizes = var_2500_split_sizes_0, x = linear_162_cast_fp16)[name = tensor("op_2500_cast_fp16")]; + tensor mean_85_axes_0 = const()[name = tensor("mean_85_axes_0"), val = tensor([-1])]; + tensor mean_85_keep_dims_0 = const()[name = tensor("mean_85_keep_dims_0"), val = tensor(true)]; + tensor mean_85_cast_fp16 = reduce_mean(axes = mean_85_axes_0, keep_dims = mean_85_keep_dims_0, x = linear_156_cast_fp16)[name = tensor("mean_85_cast_fp16")]; + tensor sub_70_cast_fp16 = sub(x = linear_156_cast_fp16, y = mean_85_cast_fp16)[name = tensor("sub_70_cast_fp16")]; + tensor square_56_cast_fp16 = square(x = sub_70_cast_fp16)[name = tensor("square_56_cast_fp16")]; + tensor reduce_mean_113_axes_0 = const()[name = tensor("reduce_mean_113_axes_0"), val = tensor([-1])]; + tensor reduce_mean_113_keep_dims_0 = const()[name = tensor("reduce_mean_113_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_113_cast_fp16 = reduce_mean(axes = reduce_mean_113_axes_0, keep_dims = reduce_mean_113_keep_dims_0, x = square_56_cast_fp16)[name = tensor("reduce_mean_113_cast_fp16")]; + tensor var_2510_to_fp16 = const()[name = tensor("op_2510_to_fp16"), val = tensor(0x1.1p-20)]; + tensor var_2511_cast_fp16 = add(x = reduce_mean_113_cast_fp16, y = var_2510_to_fp16)[name = tensor("op_2511_cast_fp16")]; + tensor var_2512_cast_fp16 = sqrt(x = var_2511_cast_fp16)[name = tensor("op_2512_cast_fp16")]; + tensor x_271_cast_fp16 = real_div(x = sub_70_cast_fp16, y = var_2512_cast_fp16)[name = tensor("x_271_cast_fp16")]; + tensor var_2514_cast_fp16 = mul(x = x_271_cast_fp16, y = flow_net_res_blocks_0_in_ln_weight_to_fp16)[name = tensor("op_2514_cast_fp16")]; + tensor x_273_cast_fp16 = add(x = var_2514_cast_fp16, y = flow_net_res_blocks_0_in_ln_bias_to_fp16)[name = tensor("x_273_cast_fp16")]; + tensor var_2516_promoted_to_fp16 = const()[name = tensor("op_2516_promoted_to_fp16"), val = tensor(0x1p+0)]; + tensor var_2517_cast_fp16 = add(x = var_2500_cast_fp16_1, y = var_2516_promoted_to_fp16)[name = tensor("op_2517_cast_fp16")]; + tensor var_2518_cast_fp16 = mul(x = x_273_cast_fp16, y = var_2517_cast_fp16)[name = tensor("op_2518_cast_fp16")]; + tensor input_425_cast_fp16 = add(x = var_2518_cast_fp16, y = var_2500_cast_fp16_0)[name = tensor("input_425_cast_fp16")]; + tensor linear_163_cast_fp16 = linear(bias = flow_net_res_blocks_0_mlp_0_bias_to_fp16, weight = flow_net_res_blocks_0_mlp_0_weight_to_fp16, x = input_425_cast_fp16)[name = tensor("linear_163_cast_fp16")]; + tensor input_429_cast_fp16 = silu(x = linear_163_cast_fp16)[name = tensor("input_429_cast_fp16")]; + tensor linear_164_cast_fp16 = linear(bias = flow_net_res_blocks_0_mlp_2_bias_to_fp16, weight = flow_net_res_blocks_0_mlp_2_weight_to_fp16, x = input_429_cast_fp16)[name = tensor("linear_164_cast_fp16")]; + tensor var_2529_cast_fp16 = mul(x = var_2500_cast_fp16_2, y = linear_164_cast_fp16)[name = tensor("op_2529_cast_fp16")]; + tensor x_275_cast_fp16 = add(x = linear_156_cast_fp16, y = var_2529_cast_fp16)[name = tensor("x_275_cast_fp16")]; + tensor linear_165_cast_fp16 = linear(bias = flow_net_res_blocks_1_adaLN_modulation_1_bias_to_fp16, weight = flow_net_res_blocks_1_adaLN_modulation_1_weight_to_fp16, x = input_423_cast_fp16)[name = tensor("linear_165_cast_fp16")]; + tensor var_2539_split_sizes_0 = const()[name = tensor("op_2539_split_sizes_0"), val = tensor([512, 512, 512])]; + tensor var_2539_axis_0 = const()[name = tensor("op_2539_axis_0"), val = tensor(-1)]; + tensor var_2539_cast_fp16_0, tensor var_2539_cast_fp16_1, tensor var_2539_cast_fp16_2 = split(axis = var_2539_axis_0, split_sizes = var_2539_split_sizes_0, x = linear_165_cast_fp16)[name = tensor("op_2539_cast_fp16")]; + tensor mean_87_axes_0 = const()[name = tensor("mean_87_axes_0"), val = tensor([-1])]; + tensor mean_87_keep_dims_0 = const()[name = tensor("mean_87_keep_dims_0"), val = tensor(true)]; + tensor mean_87_cast_fp16 = reduce_mean(axes = mean_87_axes_0, keep_dims = mean_87_keep_dims_0, x = x_275_cast_fp16)[name = tensor("mean_87_cast_fp16")]; + tensor sub_71_cast_fp16 = sub(x = x_275_cast_fp16, y = mean_87_cast_fp16)[name = tensor("sub_71_cast_fp16")]; + tensor square_57_cast_fp16 = square(x = sub_71_cast_fp16)[name = tensor("square_57_cast_fp16")]; + tensor reduce_mean_115_axes_0 = const()[name = tensor("reduce_mean_115_axes_0"), val = tensor([-1])]; + tensor reduce_mean_115_keep_dims_0 = const()[name = tensor("reduce_mean_115_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_115_cast_fp16 = reduce_mean(axes = reduce_mean_115_axes_0, keep_dims = reduce_mean_115_keep_dims_0, x = square_57_cast_fp16)[name = tensor("reduce_mean_115_cast_fp16")]; + tensor var_2549_to_fp16 = const()[name = tensor("op_2549_to_fp16"), val = tensor(0x1.1p-20)]; + tensor var_2550_cast_fp16 = add(x = reduce_mean_115_cast_fp16, y = var_2549_to_fp16)[name = tensor("op_2550_cast_fp16")]; + tensor var_2551_cast_fp16 = sqrt(x = var_2550_cast_fp16)[name = tensor("op_2551_cast_fp16")]; + tensor x_277_cast_fp16 = real_div(x = sub_71_cast_fp16, y = var_2551_cast_fp16)[name = tensor("x_277_cast_fp16")]; + tensor var_2553_cast_fp16 = mul(x = x_277_cast_fp16, y = flow_net_res_blocks_1_in_ln_weight_to_fp16)[name = tensor("op_2553_cast_fp16")]; + tensor x_279_cast_fp16 = add(x = var_2553_cast_fp16, y = flow_net_res_blocks_1_in_ln_bias_to_fp16)[name = tensor("x_279_cast_fp16")]; + tensor var_2555_promoted_to_fp16 = const()[name = tensor("op_2555_promoted_to_fp16"), val = tensor(0x1p+0)]; + tensor var_2556_cast_fp16 = add(x = var_2539_cast_fp16_1, y = var_2555_promoted_to_fp16)[name = tensor("op_2556_cast_fp16")]; + tensor var_2557_cast_fp16 = mul(x = x_279_cast_fp16, y = var_2556_cast_fp16)[name = tensor("op_2557_cast_fp16")]; + tensor input_433_cast_fp16 = add(x = var_2557_cast_fp16, y = var_2539_cast_fp16_0)[name = tensor("input_433_cast_fp16")]; + tensor linear_166_cast_fp16 = linear(bias = flow_net_res_blocks_1_mlp_0_bias_to_fp16, weight = flow_net_res_blocks_1_mlp_0_weight_to_fp16, x = input_433_cast_fp16)[name = tensor("linear_166_cast_fp16")]; + tensor input_437_cast_fp16 = silu(x = linear_166_cast_fp16)[name = tensor("input_437_cast_fp16")]; + tensor linear_167_cast_fp16 = linear(bias = flow_net_res_blocks_1_mlp_2_bias_to_fp16, weight = flow_net_res_blocks_1_mlp_2_weight_to_fp16, x = input_437_cast_fp16)[name = tensor("linear_167_cast_fp16")]; + tensor var_2568_cast_fp16 = mul(x = var_2539_cast_fp16_2, y = linear_167_cast_fp16)[name = tensor("op_2568_cast_fp16")]; + tensor x_281_cast_fp16 = add(x = x_275_cast_fp16, y = var_2568_cast_fp16)[name = tensor("x_281_cast_fp16")]; + tensor linear_168_cast_fp16 = linear(bias = flow_net_res_blocks_2_adaLN_modulation_1_bias_to_fp16, weight = flow_net_res_blocks_2_adaLN_modulation_1_weight_to_fp16, x = input_423_cast_fp16)[name = tensor("linear_168_cast_fp16")]; + tensor var_2578_split_sizes_0 = const()[name = tensor("op_2578_split_sizes_0"), val = tensor([512, 512, 512])]; + tensor var_2578_axis_0 = const()[name = tensor("op_2578_axis_0"), val = tensor(-1)]; + tensor var_2578_cast_fp16_0, tensor var_2578_cast_fp16_1, tensor var_2578_cast_fp16_2 = split(axis = var_2578_axis_0, split_sizes = var_2578_split_sizes_0, x = linear_168_cast_fp16)[name = tensor("op_2578_cast_fp16")]; + tensor mean_89_axes_0 = const()[name = tensor("mean_89_axes_0"), val = tensor([-1])]; + tensor mean_89_keep_dims_0 = const()[name = tensor("mean_89_keep_dims_0"), val = tensor(true)]; + tensor mean_89_cast_fp16 = reduce_mean(axes = mean_89_axes_0, keep_dims = mean_89_keep_dims_0, x = x_281_cast_fp16)[name = tensor("mean_89_cast_fp16")]; + tensor sub_72_cast_fp16 = sub(x = x_281_cast_fp16, y = mean_89_cast_fp16)[name = tensor("sub_72_cast_fp16")]; + tensor square_58_cast_fp16 = square(x = sub_72_cast_fp16)[name = tensor("square_58_cast_fp16")]; + tensor reduce_mean_117_axes_0 = const()[name = tensor("reduce_mean_117_axes_0"), val = tensor([-1])]; + tensor reduce_mean_117_keep_dims_0 = const()[name = tensor("reduce_mean_117_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_117_cast_fp16 = reduce_mean(axes = reduce_mean_117_axes_0, keep_dims = reduce_mean_117_keep_dims_0, x = square_58_cast_fp16)[name = tensor("reduce_mean_117_cast_fp16")]; + tensor var_2588_to_fp16 = const()[name = tensor("op_2588_to_fp16"), val = tensor(0x1.1p-20)]; + tensor var_2589_cast_fp16 = add(x = reduce_mean_117_cast_fp16, y = var_2588_to_fp16)[name = tensor("op_2589_cast_fp16")]; + tensor var_2590_cast_fp16 = sqrt(x = var_2589_cast_fp16)[name = tensor("op_2590_cast_fp16")]; + tensor x_283_cast_fp16 = real_div(x = sub_72_cast_fp16, y = var_2590_cast_fp16)[name = tensor("x_283_cast_fp16")]; + tensor var_2592_cast_fp16 = mul(x = x_283_cast_fp16, y = flow_net_res_blocks_2_in_ln_weight_to_fp16)[name = tensor("op_2592_cast_fp16")]; + tensor x_285_cast_fp16 = add(x = var_2592_cast_fp16, y = flow_net_res_blocks_2_in_ln_bias_to_fp16)[name = tensor("x_285_cast_fp16")]; + tensor var_2594_promoted_to_fp16 = const()[name = tensor("op_2594_promoted_to_fp16"), val = tensor(0x1p+0)]; + tensor var_2595_cast_fp16 = add(x = var_2578_cast_fp16_1, y = var_2594_promoted_to_fp16)[name = tensor("op_2595_cast_fp16")]; + tensor var_2596_cast_fp16 = mul(x = x_285_cast_fp16, y = var_2595_cast_fp16)[name = tensor("op_2596_cast_fp16")]; + tensor input_441_cast_fp16 = add(x = var_2596_cast_fp16, y = var_2578_cast_fp16_0)[name = tensor("input_441_cast_fp16")]; + tensor linear_169_cast_fp16 = linear(bias = flow_net_res_blocks_2_mlp_0_bias_to_fp16, weight = flow_net_res_blocks_2_mlp_0_weight_to_fp16, x = input_441_cast_fp16)[name = tensor("linear_169_cast_fp16")]; + tensor input_445_cast_fp16 = silu(x = linear_169_cast_fp16)[name = tensor("input_445_cast_fp16")]; + tensor linear_170_cast_fp16 = linear(bias = flow_net_res_blocks_2_mlp_2_bias_to_fp16, weight = flow_net_res_blocks_2_mlp_2_weight_to_fp16, x = input_445_cast_fp16)[name = tensor("linear_170_cast_fp16")]; + tensor var_2607_cast_fp16 = mul(x = var_2578_cast_fp16_2, y = linear_170_cast_fp16)[name = tensor("op_2607_cast_fp16")]; + tensor x_287_cast_fp16 = add(x = x_281_cast_fp16, y = var_2607_cast_fp16)[name = tensor("x_287_cast_fp16")]; + tensor linear_171_cast_fp16 = linear(bias = flow_net_res_blocks_3_adaLN_modulation_1_bias_to_fp16, weight = flow_net_res_blocks_3_adaLN_modulation_1_weight_to_fp16, x = input_423_cast_fp16)[name = tensor("linear_171_cast_fp16")]; + tensor var_2617_split_sizes_0 = const()[name = tensor("op_2617_split_sizes_0"), val = tensor([512, 512, 512])]; + tensor var_2617_axis_0 = const()[name = tensor("op_2617_axis_0"), val = tensor(-1)]; + tensor var_2617_cast_fp16_0, tensor var_2617_cast_fp16_1, tensor var_2617_cast_fp16_2 = split(axis = var_2617_axis_0, split_sizes = var_2617_split_sizes_0, x = linear_171_cast_fp16)[name = tensor("op_2617_cast_fp16")]; + tensor mean_91_axes_0 = const()[name = tensor("mean_91_axes_0"), val = tensor([-1])]; + tensor mean_91_keep_dims_0 = const()[name = tensor("mean_91_keep_dims_0"), val = tensor(true)]; + tensor mean_91_cast_fp16 = reduce_mean(axes = mean_91_axes_0, keep_dims = mean_91_keep_dims_0, x = x_287_cast_fp16)[name = tensor("mean_91_cast_fp16")]; + tensor sub_73_cast_fp16 = sub(x = x_287_cast_fp16, y = mean_91_cast_fp16)[name = tensor("sub_73_cast_fp16")]; + tensor square_59_cast_fp16 = square(x = sub_73_cast_fp16)[name = tensor("square_59_cast_fp16")]; + tensor reduce_mean_119_axes_0 = const()[name = tensor("reduce_mean_119_axes_0"), val = tensor([-1])]; + tensor reduce_mean_119_keep_dims_0 = const()[name = tensor("reduce_mean_119_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_119_cast_fp16 = reduce_mean(axes = reduce_mean_119_axes_0, keep_dims = reduce_mean_119_keep_dims_0, x = square_59_cast_fp16)[name = tensor("reduce_mean_119_cast_fp16")]; + tensor var_2627_to_fp16 = const()[name = tensor("op_2627_to_fp16"), val = tensor(0x1.1p-20)]; + tensor var_2628_cast_fp16 = add(x = reduce_mean_119_cast_fp16, y = var_2627_to_fp16)[name = tensor("op_2628_cast_fp16")]; + tensor var_2629_cast_fp16 = sqrt(x = var_2628_cast_fp16)[name = tensor("op_2629_cast_fp16")]; + tensor x_289_cast_fp16 = real_div(x = sub_73_cast_fp16, y = var_2629_cast_fp16)[name = tensor("x_289_cast_fp16")]; + tensor var_2631_cast_fp16 = mul(x = x_289_cast_fp16, y = flow_net_res_blocks_3_in_ln_weight_to_fp16)[name = tensor("op_2631_cast_fp16")]; + tensor x_291_cast_fp16 = add(x = var_2631_cast_fp16, y = flow_net_res_blocks_3_in_ln_bias_to_fp16)[name = tensor("x_291_cast_fp16")]; + tensor var_2633_promoted_to_fp16 = const()[name = tensor("op_2633_promoted_to_fp16"), val = tensor(0x1p+0)]; + tensor var_2634_cast_fp16 = add(x = var_2617_cast_fp16_1, y = var_2633_promoted_to_fp16)[name = tensor("op_2634_cast_fp16")]; + tensor var_2635_cast_fp16 = mul(x = x_291_cast_fp16, y = var_2634_cast_fp16)[name = tensor("op_2635_cast_fp16")]; + tensor input_449_cast_fp16 = add(x = var_2635_cast_fp16, y = var_2617_cast_fp16_0)[name = tensor("input_449_cast_fp16")]; + tensor linear_172_cast_fp16 = linear(bias = flow_net_res_blocks_3_mlp_0_bias_to_fp16, weight = flow_net_res_blocks_3_mlp_0_weight_to_fp16, x = input_449_cast_fp16)[name = tensor("linear_172_cast_fp16")]; + tensor input_453_cast_fp16 = silu(x = linear_172_cast_fp16)[name = tensor("input_453_cast_fp16")]; + tensor linear_173_cast_fp16 = linear(bias = flow_net_res_blocks_3_mlp_2_bias_to_fp16, weight = flow_net_res_blocks_3_mlp_2_weight_to_fp16, x = input_453_cast_fp16)[name = tensor("linear_173_cast_fp16")]; + tensor var_2646_cast_fp16 = mul(x = var_2617_cast_fp16_2, y = linear_173_cast_fp16)[name = tensor("op_2646_cast_fp16")]; + tensor x_293_cast_fp16 = add(x = x_287_cast_fp16, y = var_2646_cast_fp16)[name = tensor("x_293_cast_fp16")]; + tensor linear_174_cast_fp16 = linear(bias = flow_net_res_blocks_4_adaLN_modulation_1_bias_to_fp16, weight = flow_net_res_blocks_4_adaLN_modulation_1_weight_to_fp16, x = input_423_cast_fp16)[name = tensor("linear_174_cast_fp16")]; + tensor var_2656_split_sizes_0 = const()[name = tensor("op_2656_split_sizes_0"), val = tensor([512, 512, 512])]; + tensor var_2656_axis_0 = const()[name = tensor("op_2656_axis_0"), val = tensor(-1)]; + tensor var_2656_cast_fp16_0, tensor var_2656_cast_fp16_1, tensor var_2656_cast_fp16_2 = split(axis = var_2656_axis_0, split_sizes = var_2656_split_sizes_0, x = linear_174_cast_fp16)[name = tensor("op_2656_cast_fp16")]; + tensor mean_93_axes_0 = const()[name = tensor("mean_93_axes_0"), val = tensor([-1])]; + tensor mean_93_keep_dims_0 = const()[name = tensor("mean_93_keep_dims_0"), val = tensor(true)]; + tensor mean_93_cast_fp16 = reduce_mean(axes = mean_93_axes_0, keep_dims = mean_93_keep_dims_0, x = x_293_cast_fp16)[name = tensor("mean_93_cast_fp16")]; + tensor sub_74_cast_fp16 = sub(x = x_293_cast_fp16, y = mean_93_cast_fp16)[name = tensor("sub_74_cast_fp16")]; + tensor square_60_cast_fp16 = square(x = sub_74_cast_fp16)[name = tensor("square_60_cast_fp16")]; + tensor reduce_mean_121_axes_0 = const()[name = tensor("reduce_mean_121_axes_0"), val = tensor([-1])]; + tensor reduce_mean_121_keep_dims_0 = const()[name = tensor("reduce_mean_121_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_121_cast_fp16 = reduce_mean(axes = reduce_mean_121_axes_0, keep_dims = reduce_mean_121_keep_dims_0, x = square_60_cast_fp16)[name = tensor("reduce_mean_121_cast_fp16")]; + tensor var_2666_to_fp16 = const()[name = tensor("op_2666_to_fp16"), val = tensor(0x1.1p-20)]; + tensor var_2667_cast_fp16 = add(x = reduce_mean_121_cast_fp16, y = var_2666_to_fp16)[name = tensor("op_2667_cast_fp16")]; + tensor var_2668_cast_fp16 = sqrt(x = var_2667_cast_fp16)[name = tensor("op_2668_cast_fp16")]; + tensor x_295_cast_fp16 = real_div(x = sub_74_cast_fp16, y = var_2668_cast_fp16)[name = tensor("x_295_cast_fp16")]; + tensor var_2670_cast_fp16 = mul(x = x_295_cast_fp16, y = flow_net_res_blocks_4_in_ln_weight_to_fp16)[name = tensor("op_2670_cast_fp16")]; + tensor x_297_cast_fp16 = add(x = var_2670_cast_fp16, y = flow_net_res_blocks_4_in_ln_bias_to_fp16)[name = tensor("x_297_cast_fp16")]; + tensor var_2672_promoted_to_fp16 = const()[name = tensor("op_2672_promoted_to_fp16"), val = tensor(0x1p+0)]; + tensor var_2673_cast_fp16 = add(x = var_2656_cast_fp16_1, y = var_2672_promoted_to_fp16)[name = tensor("op_2673_cast_fp16")]; + tensor var_2674_cast_fp16 = mul(x = x_297_cast_fp16, y = var_2673_cast_fp16)[name = tensor("op_2674_cast_fp16")]; + tensor input_457_cast_fp16 = add(x = var_2674_cast_fp16, y = var_2656_cast_fp16_0)[name = tensor("input_457_cast_fp16")]; + tensor linear_175_cast_fp16 = linear(bias = flow_net_res_blocks_4_mlp_0_bias_to_fp16, weight = flow_net_res_blocks_4_mlp_0_weight_to_fp16, x = input_457_cast_fp16)[name = tensor("linear_175_cast_fp16")]; + tensor input_461_cast_fp16 = silu(x = linear_175_cast_fp16)[name = tensor("input_461_cast_fp16")]; + tensor linear_176_cast_fp16 = linear(bias = flow_net_res_blocks_4_mlp_2_bias_to_fp16, weight = flow_net_res_blocks_4_mlp_2_weight_to_fp16, x = input_461_cast_fp16)[name = tensor("linear_176_cast_fp16")]; + tensor var_2685_cast_fp16 = mul(x = var_2656_cast_fp16_2, y = linear_176_cast_fp16)[name = tensor("op_2685_cast_fp16")]; + tensor x_299_cast_fp16 = add(x = x_293_cast_fp16, y = var_2685_cast_fp16)[name = tensor("x_299_cast_fp16")]; + tensor linear_177_cast_fp16 = linear(bias = flow_net_res_blocks_5_adaLN_modulation_1_bias_to_fp16, weight = flow_net_res_blocks_5_adaLN_modulation_1_weight_to_fp16, x = input_423_cast_fp16)[name = tensor("linear_177_cast_fp16")]; + tensor var_2695_split_sizes_0 = const()[name = tensor("op_2695_split_sizes_0"), val = tensor([512, 512, 512])]; + tensor var_2695_axis_0 = const()[name = tensor("op_2695_axis_0"), val = tensor(-1)]; + tensor var_2695_cast_fp16_0, tensor var_2695_cast_fp16_1, tensor var_2695_cast_fp16_2 = split(axis = var_2695_axis_0, split_sizes = var_2695_split_sizes_0, x = linear_177_cast_fp16)[name = tensor("op_2695_cast_fp16")]; + tensor mean_95_axes_0 = const()[name = tensor("mean_95_axes_0"), val = tensor([-1])]; + tensor mean_95_keep_dims_0 = const()[name = tensor("mean_95_keep_dims_0"), val = tensor(true)]; + tensor mean_95_cast_fp16 = reduce_mean(axes = mean_95_axes_0, keep_dims = mean_95_keep_dims_0, x = x_299_cast_fp16)[name = tensor("mean_95_cast_fp16")]; + tensor sub_75_cast_fp16 = sub(x = x_299_cast_fp16, y = mean_95_cast_fp16)[name = tensor("sub_75_cast_fp16")]; + tensor square_61_cast_fp16 = square(x = sub_75_cast_fp16)[name = tensor("square_61_cast_fp16")]; + tensor reduce_mean_123_axes_0 = const()[name = tensor("reduce_mean_123_axes_0"), val = tensor([-1])]; + tensor reduce_mean_123_keep_dims_0 = const()[name = tensor("reduce_mean_123_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_123_cast_fp16 = reduce_mean(axes = reduce_mean_123_axes_0, keep_dims = reduce_mean_123_keep_dims_0, x = square_61_cast_fp16)[name = tensor("reduce_mean_123_cast_fp16")]; + tensor var_2705_to_fp16 = const()[name = tensor("op_2705_to_fp16"), val = tensor(0x1.1p-20)]; + tensor var_2706_cast_fp16 = add(x = reduce_mean_123_cast_fp16, y = var_2705_to_fp16)[name = tensor("op_2706_cast_fp16")]; + tensor var_2707_cast_fp16 = sqrt(x = var_2706_cast_fp16)[name = tensor("op_2707_cast_fp16")]; + tensor x_301_cast_fp16 = real_div(x = sub_75_cast_fp16, y = var_2707_cast_fp16)[name = tensor("x_301_cast_fp16")]; + tensor var_2709_cast_fp16 = mul(x = x_301_cast_fp16, y = flow_net_res_blocks_5_in_ln_weight_to_fp16)[name = tensor("op_2709_cast_fp16")]; + tensor x_303_cast_fp16 = add(x = var_2709_cast_fp16, y = flow_net_res_blocks_5_in_ln_bias_to_fp16)[name = tensor("x_303_cast_fp16")]; + tensor var_2711_promoted_to_fp16 = const()[name = tensor("op_2711_promoted_to_fp16"), val = tensor(0x1p+0)]; + tensor var_2712_cast_fp16 = add(x = var_2695_cast_fp16_1, y = var_2711_promoted_to_fp16)[name = tensor("op_2712_cast_fp16")]; + tensor var_2713_cast_fp16 = mul(x = x_303_cast_fp16, y = var_2712_cast_fp16)[name = tensor("op_2713_cast_fp16")]; + tensor input_465_cast_fp16 = add(x = var_2713_cast_fp16, y = var_2695_cast_fp16_0)[name = tensor("input_465_cast_fp16")]; + tensor linear_178_cast_fp16 = linear(bias = flow_net_res_blocks_5_mlp_0_bias_to_fp16, weight = flow_net_res_blocks_5_mlp_0_weight_to_fp16, x = input_465_cast_fp16)[name = tensor("linear_178_cast_fp16")]; + tensor input_469_cast_fp16 = silu(x = linear_178_cast_fp16)[name = tensor("input_469_cast_fp16")]; + tensor linear_179_cast_fp16 = linear(bias = flow_net_res_blocks_5_mlp_2_bias_to_fp16, weight = flow_net_res_blocks_5_mlp_2_weight_to_fp16, x = input_469_cast_fp16)[name = tensor("linear_179_cast_fp16")]; + tensor var_2724_cast_fp16 = mul(x = var_2695_cast_fp16_2, y = linear_179_cast_fp16)[name = tensor("op_2724_cast_fp16")]; + tensor x_305_cast_fp16 = add(x = x_299_cast_fp16, y = var_2724_cast_fp16)[name = tensor("x_305_cast_fp16")]; + tensor linear_180_cast_fp16 = linear(bias = flow_net_final_layer_adaLN_modulation_1_bias_to_fp16, weight = flow_net_final_layer_adaLN_modulation_1_weight_to_fp16, x = input_423_cast_fp16)[name = tensor("linear_180_cast_fp16")]; + tensor var_2733_split_sizes_0 = const()[name = tensor("op_2733_split_sizes_0"), val = tensor([512, 512])]; + tensor var_2733_axis_0 = const()[name = tensor("op_2733_axis_0"), val = tensor(-1)]; + tensor var_2733_cast_fp16_0, tensor var_2733_cast_fp16_1 = split(axis = var_2733_axis_0, split_sizes = var_2733_split_sizes_0, x = linear_180_cast_fp16)[name = tensor("op_2733_cast_fp16")]; + tensor mean_97_axes_0 = const()[name = tensor("mean_97_axes_0"), val = tensor([-1])]; + tensor mean_97_keep_dims_0 = const()[name = tensor("mean_97_keep_dims_0"), val = tensor(true)]; + tensor mean_97_cast_fp16 = reduce_mean(axes = mean_97_axes_0, keep_dims = mean_97_keep_dims_0, x = x_305_cast_fp16)[name = tensor("mean_97_cast_fp16")]; + tensor sub_76_cast_fp16 = sub(x = x_305_cast_fp16, y = mean_97_cast_fp16)[name = tensor("sub_76_cast_fp16")]; + tensor square_62_cast_fp16 = square(x = sub_76_cast_fp16)[name = tensor("square_62_cast_fp16")]; + tensor reduce_mean_125_axes_0 = const()[name = tensor("reduce_mean_125_axes_0"), val = tensor([-1])]; + tensor reduce_mean_125_keep_dims_0 = const()[name = tensor("reduce_mean_125_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_125_cast_fp16 = reduce_mean(axes = reduce_mean_125_axes_0, keep_dims = reduce_mean_125_keep_dims_0, x = square_62_cast_fp16)[name = tensor("reduce_mean_125_cast_fp16")]; + tensor var_2740_to_fp16 = const()[name = tensor("op_2740_to_fp16"), val = tensor(0x1.1p-20)]; + tensor var_2741_cast_fp16 = add(x = reduce_mean_125_cast_fp16, y = var_2740_to_fp16)[name = tensor("op_2741_cast_fp16")]; + tensor var_2742_cast_fp16 = sqrt(x = var_2741_cast_fp16)[name = tensor("op_2742_cast_fp16")]; + tensor x_307_cast_fp16 = real_div(x = sub_76_cast_fp16, y = var_2742_cast_fp16)[name = tensor("x_307_cast_fp16")]; + tensor var_2744_promoted_to_fp16 = const()[name = tensor("op_2744_promoted_to_fp16"), val = tensor(0x1p+0)]; + tensor var_2745_cast_fp16 = add(x = var_2733_cast_fp16_1, y = var_2744_promoted_to_fp16)[name = tensor("op_2745_cast_fp16")]; + tensor var_2746_cast_fp16 = mul(x = x_307_cast_fp16, y = var_2745_cast_fp16)[name = tensor("op_2746_cast_fp16")]; + tensor input_473_cast_fp16 = add(x = var_2746_cast_fp16, y = var_2733_cast_fp16_0)[name = tensor("input_473_cast_fp16")]; + tensor linear_181_cast_fp16 = linear(bias = flow_net_final_layer_linear_bias_to_fp16, weight = flow_net_final_layer_linear_weight_to_fp16, x = input_473_cast_fp16)[name = tensor("linear_181_cast_fp16")]; + tensor var_2757_to_fp16 = const()[name = tensor("op_2757_to_fp16"), val = tensor(0x1p-3)]; + tensor var_2758_cast_fp16 = mul(x = linear_181_cast_fp16, y = var_2757_to_fp16)[name = tensor("op_2758_cast_fp16")]; + tensor input_475_cast_fp16 = add(x = input_407_cast_fp16, y = var_2758_cast_fp16)[name = tensor("input_475_cast_fp16")]; + tensor linear_182_cast_fp16 = linear(bias = flow_net_input_proj_bias_to_fp16, weight = flow_net_input_proj_weight_to_fp16, x = input_475_cast_fp16)[name = tensor("linear_182_cast_fp16")]; + tensor input_479_to_fp16 = const()[name = tensor("input_479_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(19010432)))]; + tensor input_481_cast_fp16 = silu(x = input_479_to_fp16)[name = tensor("input_481_cast_fp16")]; + tensor linear_184_cast_fp16 = linear(bias = flow_net_time_embed_0_mlp_2_bias_to_fp16, weight = flow_net_time_embed_0_mlp_2_weight_to_fp16, x = input_481_cast_fp16)[name = tensor("linear_184_cast_fp16")]; + tensor reduce_mean_126_axes_0 = const()[name = tensor("reduce_mean_126_axes_0"), val = tensor([-1])]; + tensor reduce_mean_126_keep_dims_0 = const()[name = tensor("reduce_mean_126_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_126_cast_fp16 = reduce_mean(axes = reduce_mean_126_axes_0, keep_dims = reduce_mean_126_keep_dims_0, x = linear_184_cast_fp16)[name = tensor("reduce_mean_126_cast_fp16")]; + tensor sub_77_cast_fp16 = sub(x = linear_184_cast_fp16, y = reduce_mean_126_cast_fp16)[name = tensor("sub_77_cast_fp16")]; + tensor square_63_cast_fp16 = square(x = sub_77_cast_fp16)[name = tensor("square_63_cast_fp16")]; + tensor reduce_mean_127_axes_0 = const()[name = tensor("reduce_mean_127_axes_0"), val = tensor([-1])]; + tensor reduce_mean_127_keep_dims_0 = const()[name = tensor("reduce_mean_127_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_127_cast_fp16 = reduce_mean(axes = reduce_mean_127_axes_0, keep_dims = reduce_mean_127_keep_dims_0, x = square_63_cast_fp16)[name = tensor("reduce_mean_127_cast_fp16")]; + tensor real_div_14_to_fp16 = const()[name = tensor("real_div_14_to_fp16"), val = tensor(0x1.008p+0)]; + tensor mul_14_cast_fp16 = mul(x = reduce_mean_127_cast_fp16, y = real_div_14_to_fp16)[name = tensor("mul_14_cast_fp16")]; + tensor var_2824_to_fp16 = const()[name = tensor("op_2824_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_127_cast_fp16_0 = add(x = mul_14_cast_fp16, y = var_2824_to_fp16)[name = tensor("var_127_cast_fp16")]; + tensor var_2827_epsilon_0 = const()[name = tensor("op_2827_epsilon_0"), val = tensor(0x1.197998p-40)]; + tensor var_2827_cast_fp16 = rsqrt(epsilon = var_2827_epsilon_0, x = var_127_cast_fp16_0)[name = tensor("op_2827_cast_fp16")]; + tensor var_2828_cast_fp16 = mul(x = const_3_to_fp16, y = var_2827_cast_fp16)[name = tensor("op_2828_cast_fp16")]; + tensor var_2829_cast_fp16 = mul(x = linear_184_cast_fp16, y = var_2828_cast_fp16)[name = tensor("op_2829_cast_fp16")]; + tensor input_485_to_fp16 = const()[name = tensor("input_485_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(19011520)))]; + tensor input_487_cast_fp16 = silu(x = input_485_to_fp16)[name = tensor("input_487_cast_fp16")]; + tensor linear_186_cast_fp16 = linear(bias = flow_net_time_embed_1_mlp_2_bias_to_fp16, weight = flow_net_time_embed_1_mlp_2_weight_to_fp16, x = input_487_cast_fp16)[name = tensor("linear_186_cast_fp16")]; + tensor reduce_mean_128_axes_0 = const()[name = tensor("reduce_mean_128_axes_0"), val = tensor([-1])]; + tensor reduce_mean_128_keep_dims_0 = const()[name = tensor("reduce_mean_128_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_128_cast_fp16 = reduce_mean(axes = reduce_mean_128_axes_0, keep_dims = reduce_mean_128_keep_dims_0, x = linear_186_cast_fp16)[name = tensor("reduce_mean_128_cast_fp16")]; + tensor sub_79_cast_fp16 = sub(x = linear_186_cast_fp16, y = reduce_mean_128_cast_fp16)[name = tensor("sub_79_cast_fp16")]; + tensor square_64_cast_fp16 = square(x = sub_79_cast_fp16)[name = tensor("square_64_cast_fp16")]; + tensor reduce_mean_129_axes_0 = const()[name = tensor("reduce_mean_129_axes_0"), val = tensor([-1])]; + tensor reduce_mean_129_keep_dims_0 = const()[name = tensor("reduce_mean_129_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_129_cast_fp16 = reduce_mean(axes = reduce_mean_129_axes_0, keep_dims = reduce_mean_129_keep_dims_0, x = square_64_cast_fp16)[name = tensor("reduce_mean_129_cast_fp16")]; + tensor real_div_15_to_fp16 = const()[name = tensor("real_div_15_to_fp16"), val = tensor(0x1.008p+0)]; + tensor mul_15_cast_fp16 = mul(x = reduce_mean_129_cast_fp16, y = real_div_15_to_fp16)[name = tensor("mul_15_cast_fp16")]; + tensor var_2851_to_fp16 = const()[name = tensor("op_2851_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_129_cast_fp16 = add(x = mul_15_cast_fp16, y = var_2851_to_fp16)[name = tensor("var_129_cast_fp16")]; + tensor var_2854_epsilon_0 = const()[name = tensor("op_2854_epsilon_0"), val = tensor(0x1.197998p-40)]; + tensor var_2854_cast_fp16 = rsqrt(epsilon = var_2854_epsilon_0, x = var_129_cast_fp16)[name = tensor("op_2854_cast_fp16")]; + tensor var_2855_cast_fp16 = mul(x = const_5_to_fp16, y = var_2854_cast_fp16)[name = tensor("op_2855_cast_fp16")]; + tensor var_2856_cast_fp16 = mul(x = linear_186_cast_fp16, y = var_2855_cast_fp16)[name = tensor("op_2856_cast_fp16")]; + tensor var_2858_cast_fp16 = add(x = var_2829_cast_fp16, y = var_2856_cast_fp16)[name = tensor("op_2858_cast_fp16")]; + tensor _inversed_t_combined_y_0_to_fp16 = const()[name = tensor("_inversed_t_combined_y_0_to_fp16"), val = tensor(0x1p-1)]; + tensor _inversed_t_combined_cast_fp16 = mul(x = var_2858_cast_fp16, y = _inversed_t_combined_y_0_to_fp16)[name = tensor("_inversed_t_combined_cast_fp16")]; + tensor input_489_cast_fp16 = add(x = _inversed_t_combined_cast_fp16, y = linear_5_cast_fp16)[name = tensor("input_489_cast_fp16")]; + tensor input_491_cast_fp16 = silu(x = input_489_cast_fp16)[name = tensor("input_491_cast_fp16")]; + tensor linear_188_cast_fp16 = linear(bias = flow_net_res_blocks_0_adaLN_modulation_1_bias_to_fp16, weight = flow_net_res_blocks_0_adaLN_modulation_1_weight_to_fp16, x = input_491_cast_fp16)[name = tensor("linear_188_cast_fp16")]; + tensor var_2873_split_sizes_0 = const()[name = tensor("op_2873_split_sizes_0"), val = tensor([512, 512, 512])]; + tensor var_2873_axis_0 = const()[name = tensor("op_2873_axis_0"), val = tensor(-1)]; + tensor var_2873_cast_fp16_0, tensor var_2873_cast_fp16_1, tensor var_2873_cast_fp16_2 = split(axis = var_2873_axis_0, split_sizes = var_2873_split_sizes_0, x = linear_188_cast_fp16)[name = tensor("op_2873_cast_fp16")]; + tensor mean_99_axes_0 = const()[name = tensor("mean_99_axes_0"), val = tensor([-1])]; + tensor mean_99_keep_dims_0 = const()[name = tensor("mean_99_keep_dims_0"), val = tensor(true)]; + tensor mean_99_cast_fp16 = reduce_mean(axes = mean_99_axes_0, keep_dims = mean_99_keep_dims_0, x = linear_182_cast_fp16)[name = tensor("mean_99_cast_fp16")]; + tensor sub_81_cast_fp16 = sub(x = linear_182_cast_fp16, y = mean_99_cast_fp16)[name = tensor("sub_81_cast_fp16")]; + tensor square_65_cast_fp16 = square(x = sub_81_cast_fp16)[name = tensor("square_65_cast_fp16")]; + tensor reduce_mean_131_axes_0 = const()[name = tensor("reduce_mean_131_axes_0"), val = tensor([-1])]; + tensor reduce_mean_131_keep_dims_0 = const()[name = tensor("reduce_mean_131_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_131_cast_fp16 = reduce_mean(axes = reduce_mean_131_axes_0, keep_dims = reduce_mean_131_keep_dims_0, x = square_65_cast_fp16)[name = tensor("reduce_mean_131_cast_fp16")]; + tensor var_2883_to_fp16 = const()[name = tensor("op_2883_to_fp16"), val = tensor(0x1.1p-20)]; + tensor var_2884_cast_fp16 = add(x = reduce_mean_131_cast_fp16, y = var_2883_to_fp16)[name = tensor("op_2884_cast_fp16")]; + tensor var_2885_cast_fp16 = sqrt(x = var_2884_cast_fp16)[name = tensor("op_2885_cast_fp16")]; + tensor x_315_cast_fp16 = real_div(x = sub_81_cast_fp16, y = var_2885_cast_fp16)[name = tensor("x_315_cast_fp16")]; + tensor var_2887_cast_fp16 = mul(x = x_315_cast_fp16, y = flow_net_res_blocks_0_in_ln_weight_to_fp16)[name = tensor("op_2887_cast_fp16")]; + tensor x_317_cast_fp16 = add(x = var_2887_cast_fp16, y = flow_net_res_blocks_0_in_ln_bias_to_fp16)[name = tensor("x_317_cast_fp16")]; + tensor var_2889_promoted_to_fp16 = const()[name = tensor("op_2889_promoted_to_fp16"), val = tensor(0x1p+0)]; + tensor var_2890_cast_fp16 = add(x = var_2873_cast_fp16_1, y = var_2889_promoted_to_fp16)[name = tensor("op_2890_cast_fp16")]; + tensor var_2891_cast_fp16 = mul(x = x_317_cast_fp16, y = var_2890_cast_fp16)[name = tensor("op_2891_cast_fp16")]; + tensor input_493_cast_fp16 = add(x = var_2891_cast_fp16, y = var_2873_cast_fp16_0)[name = tensor("input_493_cast_fp16")]; + tensor linear_189_cast_fp16 = linear(bias = flow_net_res_blocks_0_mlp_0_bias_to_fp16, weight = flow_net_res_blocks_0_mlp_0_weight_to_fp16, x = input_493_cast_fp16)[name = tensor("linear_189_cast_fp16")]; + tensor input_497_cast_fp16 = silu(x = linear_189_cast_fp16)[name = tensor("input_497_cast_fp16")]; + tensor linear_190_cast_fp16 = linear(bias = flow_net_res_blocks_0_mlp_2_bias_to_fp16, weight = flow_net_res_blocks_0_mlp_2_weight_to_fp16, x = input_497_cast_fp16)[name = tensor("linear_190_cast_fp16")]; + tensor var_2902_cast_fp16 = mul(x = var_2873_cast_fp16_2, y = linear_190_cast_fp16)[name = tensor("op_2902_cast_fp16")]; + tensor x_319_cast_fp16 = add(x = linear_182_cast_fp16, y = var_2902_cast_fp16)[name = tensor("x_319_cast_fp16")]; + tensor linear_191_cast_fp16 = linear(bias = flow_net_res_blocks_1_adaLN_modulation_1_bias_to_fp16, weight = flow_net_res_blocks_1_adaLN_modulation_1_weight_to_fp16, x = input_491_cast_fp16)[name = tensor("linear_191_cast_fp16")]; + tensor var_2912_split_sizes_0 = const()[name = tensor("op_2912_split_sizes_0"), val = tensor([512, 512, 512])]; + tensor var_2912_axis_0 = const()[name = tensor("op_2912_axis_0"), val = tensor(-1)]; + tensor var_2912_cast_fp16_0, tensor var_2912_cast_fp16_1, tensor var_2912_cast_fp16_2 = split(axis = var_2912_axis_0, split_sizes = var_2912_split_sizes_0, x = linear_191_cast_fp16)[name = tensor("op_2912_cast_fp16")]; + tensor mean_101_axes_0 = const()[name = tensor("mean_101_axes_0"), val = tensor([-1])]; + tensor mean_101_keep_dims_0 = const()[name = tensor("mean_101_keep_dims_0"), val = tensor(true)]; + tensor mean_101_cast_fp16 = reduce_mean(axes = mean_101_axes_0, keep_dims = mean_101_keep_dims_0, x = x_319_cast_fp16)[name = tensor("mean_101_cast_fp16")]; + tensor sub_82_cast_fp16 = sub(x = x_319_cast_fp16, y = mean_101_cast_fp16)[name = tensor("sub_82_cast_fp16")]; + tensor square_66_cast_fp16 = square(x = sub_82_cast_fp16)[name = tensor("square_66_cast_fp16")]; + tensor reduce_mean_133_axes_0 = const()[name = tensor("reduce_mean_133_axes_0"), val = tensor([-1])]; + tensor reduce_mean_133_keep_dims_0 = const()[name = tensor("reduce_mean_133_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_133_cast_fp16 = reduce_mean(axes = reduce_mean_133_axes_0, keep_dims = reduce_mean_133_keep_dims_0, x = square_66_cast_fp16)[name = tensor("reduce_mean_133_cast_fp16")]; + tensor var_2922_to_fp16 = const()[name = tensor("op_2922_to_fp16"), val = tensor(0x1.1p-20)]; + tensor var_2923_cast_fp16 = add(x = reduce_mean_133_cast_fp16, y = var_2922_to_fp16)[name = tensor("op_2923_cast_fp16")]; + tensor var_2924_cast_fp16 = sqrt(x = var_2923_cast_fp16)[name = tensor("op_2924_cast_fp16")]; + tensor x_321_cast_fp16 = real_div(x = sub_82_cast_fp16, y = var_2924_cast_fp16)[name = tensor("x_321_cast_fp16")]; + tensor var_2926_cast_fp16 = mul(x = x_321_cast_fp16, y = flow_net_res_blocks_1_in_ln_weight_to_fp16)[name = tensor("op_2926_cast_fp16")]; + tensor x_323_cast_fp16 = add(x = var_2926_cast_fp16, y = flow_net_res_blocks_1_in_ln_bias_to_fp16)[name = tensor("x_323_cast_fp16")]; + tensor var_2928_promoted_to_fp16 = const()[name = tensor("op_2928_promoted_to_fp16"), val = tensor(0x1p+0)]; + tensor var_2929_cast_fp16 = add(x = var_2912_cast_fp16_1, y = var_2928_promoted_to_fp16)[name = tensor("op_2929_cast_fp16")]; + tensor var_2930_cast_fp16 = mul(x = x_323_cast_fp16, y = var_2929_cast_fp16)[name = tensor("op_2930_cast_fp16")]; + tensor input_501_cast_fp16 = add(x = var_2930_cast_fp16, y = var_2912_cast_fp16_0)[name = tensor("input_501_cast_fp16")]; + tensor linear_192_cast_fp16 = linear(bias = flow_net_res_blocks_1_mlp_0_bias_to_fp16, weight = flow_net_res_blocks_1_mlp_0_weight_to_fp16, x = input_501_cast_fp16)[name = tensor("linear_192_cast_fp16")]; + tensor input_505_cast_fp16 = silu(x = linear_192_cast_fp16)[name = tensor("input_505_cast_fp16")]; + tensor linear_193_cast_fp16 = linear(bias = flow_net_res_blocks_1_mlp_2_bias_to_fp16, weight = flow_net_res_blocks_1_mlp_2_weight_to_fp16, x = input_505_cast_fp16)[name = tensor("linear_193_cast_fp16")]; + tensor var_2941_cast_fp16 = mul(x = var_2912_cast_fp16_2, y = linear_193_cast_fp16)[name = tensor("op_2941_cast_fp16")]; + tensor x_325_cast_fp16 = add(x = x_319_cast_fp16, y = var_2941_cast_fp16)[name = tensor("x_325_cast_fp16")]; + tensor linear_194_cast_fp16 = linear(bias = flow_net_res_blocks_2_adaLN_modulation_1_bias_to_fp16, weight = flow_net_res_blocks_2_adaLN_modulation_1_weight_to_fp16, x = input_491_cast_fp16)[name = tensor("linear_194_cast_fp16")]; + tensor var_2951_split_sizes_0 = const()[name = tensor("op_2951_split_sizes_0"), val = tensor([512, 512, 512])]; + tensor var_2951_axis_0 = const()[name = tensor("op_2951_axis_0"), val = tensor(-1)]; + tensor var_2951_cast_fp16_0, tensor var_2951_cast_fp16_1, tensor var_2951_cast_fp16_2 = split(axis = var_2951_axis_0, split_sizes = var_2951_split_sizes_0, x = linear_194_cast_fp16)[name = tensor("op_2951_cast_fp16")]; + tensor mean_103_axes_0 = const()[name = tensor("mean_103_axes_0"), val = tensor([-1])]; + tensor mean_103_keep_dims_0 = const()[name = tensor("mean_103_keep_dims_0"), val = tensor(true)]; + tensor mean_103_cast_fp16 = reduce_mean(axes = mean_103_axes_0, keep_dims = mean_103_keep_dims_0, x = x_325_cast_fp16)[name = tensor("mean_103_cast_fp16")]; + tensor sub_83_cast_fp16 = sub(x = x_325_cast_fp16, y = mean_103_cast_fp16)[name = tensor("sub_83_cast_fp16")]; + tensor square_67_cast_fp16 = square(x = sub_83_cast_fp16)[name = tensor("square_67_cast_fp16")]; + tensor reduce_mean_135_axes_0 = const()[name = tensor("reduce_mean_135_axes_0"), val = tensor([-1])]; + tensor reduce_mean_135_keep_dims_0 = const()[name = tensor("reduce_mean_135_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_135_cast_fp16 = reduce_mean(axes = reduce_mean_135_axes_0, keep_dims = reduce_mean_135_keep_dims_0, x = square_67_cast_fp16)[name = tensor("reduce_mean_135_cast_fp16")]; + tensor var_2961_to_fp16 = const()[name = tensor("op_2961_to_fp16"), val = tensor(0x1.1p-20)]; + tensor var_2962_cast_fp16 = add(x = reduce_mean_135_cast_fp16, y = var_2961_to_fp16)[name = tensor("op_2962_cast_fp16")]; + tensor var_2963_cast_fp16 = sqrt(x = var_2962_cast_fp16)[name = tensor("op_2963_cast_fp16")]; + tensor x_327_cast_fp16 = real_div(x = sub_83_cast_fp16, y = var_2963_cast_fp16)[name = tensor("x_327_cast_fp16")]; + tensor var_2965_cast_fp16 = mul(x = x_327_cast_fp16, y = flow_net_res_blocks_2_in_ln_weight_to_fp16)[name = tensor("op_2965_cast_fp16")]; + tensor x_329_cast_fp16 = add(x = var_2965_cast_fp16, y = flow_net_res_blocks_2_in_ln_bias_to_fp16)[name = tensor("x_329_cast_fp16")]; + tensor var_2967_promoted_to_fp16 = const()[name = tensor("op_2967_promoted_to_fp16"), val = tensor(0x1p+0)]; + tensor var_2968_cast_fp16 = add(x = var_2951_cast_fp16_1, y = var_2967_promoted_to_fp16)[name = tensor("op_2968_cast_fp16")]; + tensor var_2969_cast_fp16 = mul(x = x_329_cast_fp16, y = var_2968_cast_fp16)[name = tensor("op_2969_cast_fp16")]; + tensor input_509_cast_fp16 = add(x = var_2969_cast_fp16, y = var_2951_cast_fp16_0)[name = tensor("input_509_cast_fp16")]; + tensor linear_195_cast_fp16 = linear(bias = flow_net_res_blocks_2_mlp_0_bias_to_fp16, weight = flow_net_res_blocks_2_mlp_0_weight_to_fp16, x = input_509_cast_fp16)[name = tensor("linear_195_cast_fp16")]; + tensor input_513_cast_fp16 = silu(x = linear_195_cast_fp16)[name = tensor("input_513_cast_fp16")]; + tensor linear_196_cast_fp16 = linear(bias = flow_net_res_blocks_2_mlp_2_bias_to_fp16, weight = flow_net_res_blocks_2_mlp_2_weight_to_fp16, x = input_513_cast_fp16)[name = tensor("linear_196_cast_fp16")]; + tensor var_2980_cast_fp16 = mul(x = var_2951_cast_fp16_2, y = linear_196_cast_fp16)[name = tensor("op_2980_cast_fp16")]; + tensor x_331_cast_fp16 = add(x = x_325_cast_fp16, y = var_2980_cast_fp16)[name = tensor("x_331_cast_fp16")]; + tensor linear_197_cast_fp16 = linear(bias = flow_net_res_blocks_3_adaLN_modulation_1_bias_to_fp16, weight = flow_net_res_blocks_3_adaLN_modulation_1_weight_to_fp16, x = input_491_cast_fp16)[name = tensor("linear_197_cast_fp16")]; + tensor var_2990_split_sizes_0 = const()[name = tensor("op_2990_split_sizes_0"), val = tensor([512, 512, 512])]; + tensor var_2990_axis_0 = const()[name = tensor("op_2990_axis_0"), val = tensor(-1)]; + tensor var_2990_cast_fp16_0, tensor var_2990_cast_fp16_1, tensor var_2990_cast_fp16_2 = split(axis = var_2990_axis_0, split_sizes = var_2990_split_sizes_0, x = linear_197_cast_fp16)[name = tensor("op_2990_cast_fp16")]; + tensor mean_105_axes_0 = const()[name = tensor("mean_105_axes_0"), val = tensor([-1])]; + tensor mean_105_keep_dims_0 = const()[name = tensor("mean_105_keep_dims_0"), val = tensor(true)]; + tensor mean_105_cast_fp16 = reduce_mean(axes = mean_105_axes_0, keep_dims = mean_105_keep_dims_0, x = x_331_cast_fp16)[name = tensor("mean_105_cast_fp16")]; + tensor sub_84_cast_fp16 = sub(x = x_331_cast_fp16, y = mean_105_cast_fp16)[name = tensor("sub_84_cast_fp16")]; + tensor square_68_cast_fp16 = square(x = sub_84_cast_fp16)[name = tensor("square_68_cast_fp16")]; + tensor reduce_mean_137_axes_0 = const()[name = tensor("reduce_mean_137_axes_0"), val = tensor([-1])]; + tensor reduce_mean_137_keep_dims_0 = const()[name = tensor("reduce_mean_137_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_137_cast_fp16 = reduce_mean(axes = reduce_mean_137_axes_0, keep_dims = reduce_mean_137_keep_dims_0, x = square_68_cast_fp16)[name = tensor("reduce_mean_137_cast_fp16")]; + tensor var_3000_to_fp16 = const()[name = tensor("op_3000_to_fp16"), val = tensor(0x1.1p-20)]; + tensor var_3001_cast_fp16 = add(x = reduce_mean_137_cast_fp16, y = var_3000_to_fp16)[name = tensor("op_3001_cast_fp16")]; + tensor var_3002_cast_fp16 = sqrt(x = var_3001_cast_fp16)[name = tensor("op_3002_cast_fp16")]; + tensor x_333_cast_fp16 = real_div(x = sub_84_cast_fp16, y = var_3002_cast_fp16)[name = tensor("x_333_cast_fp16")]; + tensor var_3004_cast_fp16 = mul(x = x_333_cast_fp16, y = flow_net_res_blocks_3_in_ln_weight_to_fp16)[name = tensor("op_3004_cast_fp16")]; + tensor x_335_cast_fp16 = add(x = var_3004_cast_fp16, y = flow_net_res_blocks_3_in_ln_bias_to_fp16)[name = tensor("x_335_cast_fp16")]; + tensor var_3006_promoted_to_fp16 = const()[name = tensor("op_3006_promoted_to_fp16"), val = tensor(0x1p+0)]; + tensor var_3007_cast_fp16 = add(x = var_2990_cast_fp16_1, y = var_3006_promoted_to_fp16)[name = tensor("op_3007_cast_fp16")]; + tensor var_3008_cast_fp16 = mul(x = x_335_cast_fp16, y = var_3007_cast_fp16)[name = tensor("op_3008_cast_fp16")]; + tensor input_517_cast_fp16 = add(x = var_3008_cast_fp16, y = var_2990_cast_fp16_0)[name = tensor("input_517_cast_fp16")]; + tensor linear_198_cast_fp16 = linear(bias = flow_net_res_blocks_3_mlp_0_bias_to_fp16, weight = flow_net_res_blocks_3_mlp_0_weight_to_fp16, x = input_517_cast_fp16)[name = tensor("linear_198_cast_fp16")]; + tensor input_521_cast_fp16 = silu(x = linear_198_cast_fp16)[name = tensor("input_521_cast_fp16")]; + tensor linear_199_cast_fp16 = linear(bias = flow_net_res_blocks_3_mlp_2_bias_to_fp16, weight = flow_net_res_blocks_3_mlp_2_weight_to_fp16, x = input_521_cast_fp16)[name = tensor("linear_199_cast_fp16")]; + tensor var_3019_cast_fp16 = mul(x = var_2990_cast_fp16_2, y = linear_199_cast_fp16)[name = tensor("op_3019_cast_fp16")]; + tensor x_337_cast_fp16 = add(x = x_331_cast_fp16, y = var_3019_cast_fp16)[name = tensor("x_337_cast_fp16")]; + tensor linear_200_cast_fp16 = linear(bias = flow_net_res_blocks_4_adaLN_modulation_1_bias_to_fp16, weight = flow_net_res_blocks_4_adaLN_modulation_1_weight_to_fp16, x = input_491_cast_fp16)[name = tensor("linear_200_cast_fp16")]; + tensor var_3029_split_sizes_0 = const()[name = tensor("op_3029_split_sizes_0"), val = tensor([512, 512, 512])]; + tensor var_3029_axis_0 = const()[name = tensor("op_3029_axis_0"), val = tensor(-1)]; + tensor var_3029_cast_fp16_0, tensor var_3029_cast_fp16_1, tensor var_3029_cast_fp16_2 = split(axis = var_3029_axis_0, split_sizes = var_3029_split_sizes_0, x = linear_200_cast_fp16)[name = tensor("op_3029_cast_fp16")]; + tensor mean_107_axes_0 = const()[name = tensor("mean_107_axes_0"), val = tensor([-1])]; + tensor mean_107_keep_dims_0 = const()[name = tensor("mean_107_keep_dims_0"), val = tensor(true)]; + tensor mean_107_cast_fp16 = reduce_mean(axes = mean_107_axes_0, keep_dims = mean_107_keep_dims_0, x = x_337_cast_fp16)[name = tensor("mean_107_cast_fp16")]; + tensor sub_85_cast_fp16 = sub(x = x_337_cast_fp16, y = mean_107_cast_fp16)[name = tensor("sub_85_cast_fp16")]; + tensor square_69_cast_fp16 = square(x = sub_85_cast_fp16)[name = tensor("square_69_cast_fp16")]; + tensor reduce_mean_139_axes_0 = const()[name = tensor("reduce_mean_139_axes_0"), val = tensor([-1])]; + tensor reduce_mean_139_keep_dims_0 = const()[name = tensor("reduce_mean_139_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_139_cast_fp16 = reduce_mean(axes = reduce_mean_139_axes_0, keep_dims = reduce_mean_139_keep_dims_0, x = square_69_cast_fp16)[name = tensor("reduce_mean_139_cast_fp16")]; + tensor var_3039_to_fp16 = const()[name = tensor("op_3039_to_fp16"), val = tensor(0x1.1p-20)]; + tensor var_3040_cast_fp16 = add(x = reduce_mean_139_cast_fp16, y = var_3039_to_fp16)[name = tensor("op_3040_cast_fp16")]; + tensor var_3041_cast_fp16 = sqrt(x = var_3040_cast_fp16)[name = tensor("op_3041_cast_fp16")]; + tensor x_339_cast_fp16 = real_div(x = sub_85_cast_fp16, y = var_3041_cast_fp16)[name = tensor("x_339_cast_fp16")]; + tensor var_3043_cast_fp16 = mul(x = x_339_cast_fp16, y = flow_net_res_blocks_4_in_ln_weight_to_fp16)[name = tensor("op_3043_cast_fp16")]; + tensor x_341_cast_fp16 = add(x = var_3043_cast_fp16, y = flow_net_res_blocks_4_in_ln_bias_to_fp16)[name = tensor("x_341_cast_fp16")]; + tensor var_3045_promoted_to_fp16 = const()[name = tensor("op_3045_promoted_to_fp16"), val = tensor(0x1p+0)]; + tensor var_3046_cast_fp16 = add(x = var_3029_cast_fp16_1, y = var_3045_promoted_to_fp16)[name = tensor("op_3046_cast_fp16")]; + tensor var_3047_cast_fp16 = mul(x = x_341_cast_fp16, y = var_3046_cast_fp16)[name = tensor("op_3047_cast_fp16")]; + tensor input_525_cast_fp16 = add(x = var_3047_cast_fp16, y = var_3029_cast_fp16_0)[name = tensor("input_525_cast_fp16")]; + tensor linear_201_cast_fp16 = linear(bias = flow_net_res_blocks_4_mlp_0_bias_to_fp16, weight = flow_net_res_blocks_4_mlp_0_weight_to_fp16, x = input_525_cast_fp16)[name = tensor("linear_201_cast_fp16")]; + tensor input_529_cast_fp16 = silu(x = linear_201_cast_fp16)[name = tensor("input_529_cast_fp16")]; + tensor linear_202_cast_fp16 = linear(bias = flow_net_res_blocks_4_mlp_2_bias_to_fp16, weight = flow_net_res_blocks_4_mlp_2_weight_to_fp16, x = input_529_cast_fp16)[name = tensor("linear_202_cast_fp16")]; + tensor var_3058_cast_fp16 = mul(x = var_3029_cast_fp16_2, y = linear_202_cast_fp16)[name = tensor("op_3058_cast_fp16")]; + tensor x_343_cast_fp16 = add(x = x_337_cast_fp16, y = var_3058_cast_fp16)[name = tensor("x_343_cast_fp16")]; + tensor linear_203_cast_fp16 = linear(bias = flow_net_res_blocks_5_adaLN_modulation_1_bias_to_fp16, weight = flow_net_res_blocks_5_adaLN_modulation_1_weight_to_fp16, x = input_491_cast_fp16)[name = tensor("linear_203_cast_fp16")]; + tensor var_3068_split_sizes_0 = const()[name = tensor("op_3068_split_sizes_0"), val = tensor([512, 512, 512])]; + tensor var_3068_axis_0 = const()[name = tensor("op_3068_axis_0"), val = tensor(-1)]; + tensor var_3068_cast_fp16_0, tensor var_3068_cast_fp16_1, tensor var_3068_cast_fp16_2 = split(axis = var_3068_axis_0, split_sizes = var_3068_split_sizes_0, x = linear_203_cast_fp16)[name = tensor("op_3068_cast_fp16")]; + tensor mean_109_axes_0 = const()[name = tensor("mean_109_axes_0"), val = tensor([-1])]; + tensor mean_109_keep_dims_0 = const()[name = tensor("mean_109_keep_dims_0"), val = tensor(true)]; + tensor mean_109_cast_fp16 = reduce_mean(axes = mean_109_axes_0, keep_dims = mean_109_keep_dims_0, x = x_343_cast_fp16)[name = tensor("mean_109_cast_fp16")]; + tensor sub_86_cast_fp16 = sub(x = x_343_cast_fp16, y = mean_109_cast_fp16)[name = tensor("sub_86_cast_fp16")]; + tensor square_70_cast_fp16 = square(x = sub_86_cast_fp16)[name = tensor("square_70_cast_fp16")]; + tensor reduce_mean_141_axes_0 = const()[name = tensor("reduce_mean_141_axes_0"), val = tensor([-1])]; + tensor reduce_mean_141_keep_dims_0 = const()[name = tensor("reduce_mean_141_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_141_cast_fp16 = reduce_mean(axes = reduce_mean_141_axes_0, keep_dims = reduce_mean_141_keep_dims_0, x = square_70_cast_fp16)[name = tensor("reduce_mean_141_cast_fp16")]; + tensor var_3078_to_fp16 = const()[name = tensor("op_3078_to_fp16"), val = tensor(0x1.1p-20)]; + tensor var_3079_cast_fp16 = add(x = reduce_mean_141_cast_fp16, y = var_3078_to_fp16)[name = tensor("op_3079_cast_fp16")]; + tensor var_3080_cast_fp16 = sqrt(x = var_3079_cast_fp16)[name = tensor("op_3080_cast_fp16")]; + tensor x_345_cast_fp16 = real_div(x = sub_86_cast_fp16, y = var_3080_cast_fp16)[name = tensor("x_345_cast_fp16")]; + tensor var_3082_cast_fp16 = mul(x = x_345_cast_fp16, y = flow_net_res_blocks_5_in_ln_weight_to_fp16)[name = tensor("op_3082_cast_fp16")]; + tensor x_347_cast_fp16 = add(x = var_3082_cast_fp16, y = flow_net_res_blocks_5_in_ln_bias_to_fp16)[name = tensor("x_347_cast_fp16")]; + tensor var_3084_promoted_to_fp16 = const()[name = tensor("op_3084_promoted_to_fp16"), val = tensor(0x1p+0)]; + tensor var_3085_cast_fp16 = add(x = var_3068_cast_fp16_1, y = var_3084_promoted_to_fp16)[name = tensor("op_3085_cast_fp16")]; + tensor var_3086_cast_fp16 = mul(x = x_347_cast_fp16, y = var_3085_cast_fp16)[name = tensor("op_3086_cast_fp16")]; + tensor input_533_cast_fp16 = add(x = var_3086_cast_fp16, y = var_3068_cast_fp16_0)[name = tensor("input_533_cast_fp16")]; + tensor linear_204_cast_fp16 = linear(bias = flow_net_res_blocks_5_mlp_0_bias_to_fp16, weight = flow_net_res_blocks_5_mlp_0_weight_to_fp16, x = input_533_cast_fp16)[name = tensor("linear_204_cast_fp16")]; + tensor input_537_cast_fp16 = silu(x = linear_204_cast_fp16)[name = tensor("input_537_cast_fp16")]; + tensor linear_205_cast_fp16 = linear(bias = flow_net_res_blocks_5_mlp_2_bias_to_fp16, weight = flow_net_res_blocks_5_mlp_2_weight_to_fp16, x = input_537_cast_fp16)[name = tensor("linear_205_cast_fp16")]; + tensor var_3097_cast_fp16 = mul(x = var_3068_cast_fp16_2, y = linear_205_cast_fp16)[name = tensor("op_3097_cast_fp16")]; + tensor x_349_cast_fp16 = add(x = x_343_cast_fp16, y = var_3097_cast_fp16)[name = tensor("x_349_cast_fp16")]; + tensor linear_206_cast_fp16 = linear(bias = flow_net_final_layer_adaLN_modulation_1_bias_to_fp16, weight = flow_net_final_layer_adaLN_modulation_1_weight_to_fp16, x = input_491_cast_fp16)[name = tensor("linear_206_cast_fp16")]; + tensor var_3106_split_sizes_0 = const()[name = tensor("op_3106_split_sizes_0"), val = tensor([512, 512])]; + tensor var_3106_axis_0 = const()[name = tensor("op_3106_axis_0"), val = tensor(-1)]; + tensor var_3106_cast_fp16_0, tensor var_3106_cast_fp16_1 = split(axis = var_3106_axis_0, split_sizes = var_3106_split_sizes_0, x = linear_206_cast_fp16)[name = tensor("op_3106_cast_fp16")]; + tensor mean_axes_0 = const()[name = tensor("mean_axes_0"), val = tensor([-1])]; + tensor mean_keep_dims_0 = const()[name = tensor("mean_keep_dims_0"), val = tensor(true)]; + tensor mean_cast_fp16 = reduce_mean(axes = mean_axes_0, keep_dims = mean_keep_dims_0, x = x_349_cast_fp16)[name = tensor("mean_cast_fp16")]; + tensor sub_87_cast_fp16 = sub(x = x_349_cast_fp16, y = mean_cast_fp16)[name = tensor("sub_87_cast_fp16")]; + tensor square_71_cast_fp16 = square(x = sub_87_cast_fp16)[name = tensor("square_71_cast_fp16")]; + tensor reduce_mean_143_axes_0 = const()[name = tensor("reduce_mean_143_axes_0"), val = tensor([-1])]; + tensor reduce_mean_143_keep_dims_0 = const()[name = tensor("reduce_mean_143_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_143_cast_fp16 = reduce_mean(axes = reduce_mean_143_axes_0, keep_dims = reduce_mean_143_keep_dims_0, x = square_71_cast_fp16)[name = tensor("reduce_mean_143_cast_fp16")]; + tensor var_3113_to_fp16 = const()[name = tensor("op_3113_to_fp16"), val = tensor(0x1.1p-20)]; + tensor var_3114_cast_fp16 = add(x = reduce_mean_143_cast_fp16, y = var_3113_to_fp16)[name = tensor("op_3114_cast_fp16")]; + tensor var_3115_cast_fp16 = sqrt(x = var_3114_cast_fp16)[name = tensor("op_3115_cast_fp16")]; + tensor x_cast_fp16 = real_div(x = sub_87_cast_fp16, y = var_3115_cast_fp16)[name = tensor("x_cast_fp16")]; + tensor var_3117_promoted_to_fp16 = const()[name = tensor("op_3117_promoted_to_fp16"), val = tensor(0x1p+0)]; + tensor var_3118_cast_fp16 = add(x = var_3106_cast_fp16_1, y = var_3117_promoted_to_fp16)[name = tensor("op_3118_cast_fp16")]; + tensor var_3119_cast_fp16 = mul(x = x_cast_fp16, y = var_3118_cast_fp16)[name = tensor("op_3119_cast_fp16")]; + tensor input_cast_fp16 = add(x = var_3119_cast_fp16, y = var_3106_cast_fp16_0)[name = tensor("input_cast_fp16")]; + tensor linear_207_cast_fp16 = linear(bias = flow_net_final_layer_linear_bias_to_fp16, weight = flow_net_final_layer_linear_weight_to_fp16, x = input_cast_fp16)[name = tensor("linear_207_cast_fp16")]; + tensor var_3124_to_fp16 = const()[name = tensor("op_3124_to_fp16"), val = tensor(0x1p-3)]; + tensor var_3125_cast_fp16 = mul(x = linear_207_cast_fp16, y = var_3124_to_fp16)[name = tensor("op_3125_cast_fp16")]; + tensor var_3127_cast_fp16 = add(x = input_475_cast_fp16, y = var_3125_cast_fp16)[name = tensor("op_3127_cast_fp16")]; + tensor var_3127_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_3127_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor latent_final = cast(dtype = var_3127_cast_fp16_to_fp32_dtype_0, x = var_3127_cast_fp16)[name = tensor("cast_144")]; + } -> (latent_final); +} \ No newline at end of file diff --git a/v2.1/spanish/flow_decoder_fused.mlmodelc/weights/weight.bin b/v2.1/spanish/flow_decoder_fused.mlmodelc/weights/weight.bin new file mode 100644 index 0000000000000000000000000000000000000000..dfd39700353f5054362fffed76322ddd3ef07910 --- /dev/null +++ b/v2.1/spanish/flow_decoder_fused.mlmodelc/weights/weight.bin @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:83ee3374f6946210512b80b904e534ae6224e3cb72402ab3b80924cbff7d35eb +size 19012608 diff --git a/v2.1/spanish/flow_decoder_fused.mlpackage/Data/com.apple.CoreML/model.mlmodel b/v2.1/spanish/flow_decoder_fused.mlpackage/Data/com.apple.CoreML/model.mlmodel new file mode 100644 index 0000000000000000000000000000000000000000..934783d202f2fc1e63654e2883348619329325a2 --- /dev/null +++ b/v2.1/spanish/flow_decoder_fused.mlpackage/Data/com.apple.CoreML/model.mlmodel @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:84a40edef63d4e1cbfcf4301102401cc2909e0ef38a7cce48adf9e0b55df2138 +size 301733 diff --git a/v2.1/spanish/flow_decoder_fused.mlpackage/Data/com.apple.CoreML/weights/weight.bin b/v2.1/spanish/flow_decoder_fused.mlpackage/Data/com.apple.CoreML/weights/weight.bin new file mode 100644 index 0000000000000000000000000000000000000000..dfd39700353f5054362fffed76322ddd3ef07910 --- /dev/null +++ b/v2.1/spanish/flow_decoder_fused.mlpackage/Data/com.apple.CoreML/weights/weight.bin @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:83ee3374f6946210512b80b904e534ae6224e3cb72402ab3b80924cbff7d35eb +size 19012608 diff --git a/v2.1/spanish/flow_decoder_fused.mlpackage/Manifest.json b/v2.1/spanish/flow_decoder_fused.mlpackage/Manifest.json new file mode 100644 index 0000000000000000000000000000000000000000..92760e43760dd4c0fe44a03502b3c46870407311 --- /dev/null +++ b/v2.1/spanish/flow_decoder_fused.mlpackage/Manifest.json @@ -0,0 +1,18 @@ +{ + "fileFormatVersion": "1.0.0", + "itemInfoEntries": { + "01631C84-DEAF-4EF1-81C7-72597ED978EA": { + "author": "com.apple.CoreML", + "description": "CoreML Model Weights", + "name": "weights", + "path": "com.apple.CoreML/weights" + }, + "7B14ADCB-D7B5-4AD0-A8A8-4A065273C78F": { + "author": "com.apple.CoreML", + "description": "CoreML Model Specification", + "name": "model.mlmodel", + "path": "com.apple.CoreML/model.mlmodel" + } + }, + "rootModelIdentifier": "7B14ADCB-D7B5-4AD0-A8A8-4A065273C78F" +} diff --git a/v2.1/spanish/flowlm_step.mlmodelc/analytics/coremldata.bin b/v2.1/spanish/flowlm_step.mlmodelc/analytics/coremldata.bin new file mode 100644 index 0000000000000000000000000000000000000000..aec8eb672b12ca32599a1a93bf7e4c1a6961b83b --- /dev/null +++ b/v2.1/spanish/flowlm_step.mlmodelc/analytics/coremldata.bin @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:77bbd30262295e6ec1b2c44c04486a4909199478991d93c352516bc225dddecc +size 243 diff --git a/v2.1/spanish/flowlm_step.mlmodelc/coremldata.bin b/v2.1/spanish/flowlm_step.mlmodelc/coremldata.bin new file mode 100644 index 0000000000000000000000000000000000000000..bf246db8c4c7c4c371402356537b028ee18b571d --- /dev/null +++ b/v2.1/spanish/flowlm_step.mlmodelc/coremldata.bin @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:2eb1dc5a354006ca4a8fc9bbbd0638ee2b4db19b2500c1ddd16ccf4274442d4f +size 1191 diff --git a/v2.1/spanish/flowlm_step.mlmodelc/model.mil b/v2.1/spanish/flowlm_step.mlmodelc/model.mil new file mode 100644 index 0000000000000000000000000000000000000000..80e5ccd89c07bb9a586e16c697d270e91a6a397c --- /dev/null +++ b/v2.1/spanish/flowlm_step.mlmodelc/model.mil @@ -0,0 +1,1359 @@ +program(1.0) +[buildInfo = dict, tensor>({{"coremlc-component-MIL", "3520.4.1"}, {"coremlc-version", "3520.5.1"}, {"coremltools-component-torch", "2.12.0"}, {"coremltools-source-dialect", "TorchScript"}, {"coremltools-version", "9.0"}})] +{ + func main(tensor bos_emb, tensor cache0, tensor cache1, tensor cache2, tensor cache3, tensor cache4, tensor cache5, tensor position0, tensor position1, tensor position2, tensor position3, tensor position4, tensor position5, tensor sequence) { + tensor sequence_to_fp16_dtype_0 = const()[name = tensor("sequence_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor sequence_to_fp16 = cast(dtype = sequence_to_fp16_dtype_0, x = sequence)[name = tensor("cast_141")]; + tensor var_54_cast_fp16 = not_equal(x = sequence_to_fp16, y = sequence_to_fp16)[name = tensor("op_54_cast_fp16")]; + tensor expand_dims_0_axes_0 = const()[name = tensor("expand_dims_0_axes_0"), val = tensor([0, 1])]; + tensor bos_emb_to_fp16_dtype_0 = const()[name = tensor("bos_emb_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor bos_emb_to_fp16 = cast(dtype = bos_emb_to_fp16_dtype_0, x = bos_emb)[name = tensor("cast_140")]; + tensor expand_dims_0_cast_fp16 = expand_dims(axes = expand_dims_0_axes_0, x = bos_emb_to_fp16)[name = tensor("expand_dims_0_cast_fp16")]; + tensor input_1_cast_fp16 = select(a = expand_dims_0_cast_fp16, b = sequence_to_fp16, cond = var_54_cast_fp16)[name = tensor("input_1_cast_fp16")]; + tensor input_linear_weight_to_fp16 = const()[name = tensor("input_linear_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(64)))]; + tensor linear_0_bias_0_to_fp16 = const()[name = tensor("linear_0_bias_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(65664)))]; + tensor linear_0_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = input_linear_weight_to_fp16, x = input_1_cast_fp16)[name = tensor("linear_0_cast_fp16")]; + tensor input_5_axes_0 = const()[name = tensor("input_5_axes_0"), val = tensor([-1])]; + tensor norm0_1_weight_to_fp16 = const()[name = tensor("norm0_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(67776)))]; + tensor norm0_1_bias_to_fp16 = const()[name = tensor("norm0_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(69888)))]; + tensor var_60_to_fp16 = const()[name = tensor("op_60_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_5_cast_fp16 = layer_norm(axes = input_5_axes_0, beta = norm0_1_bias_to_fp16, epsilon = var_60_to_fp16, gamma = norm0_1_weight_to_fp16, x = linear_0_cast_fp16)[name = tensor("input_5_cast_fp16")]; + tensor attn0_in_proj_weight_to_fp16 = const()[name = tensor("attn0_in_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(72000)))]; + tensor linear_1_bias_0_to_fp16 = const()[name = tensor("linear_1_bias_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(6363520)))]; + tensor linear_1_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = attn0_in_proj_weight_to_fp16, x = input_5_cast_fp16)[name = tensor("linear_1_cast_fp16")]; + tensor var_74 = const()[name = tensor("op_74"), val = tensor([1, 1, 3, 16, 64])]; + tensor qkv_1_cast_fp16 = reshape(shape = var_74, x = linear_1_cast_fp16)[name = tensor("qkv_1_cast_fp16")]; + tensor q_1_begin_0 = const()[name = tensor("q_1_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor q_1_end_0 = const()[name = tensor("q_1_end_0"), val = tensor([1, 1, 1, 16, 64])]; + tensor q_1_end_mask_0 = const()[name = tensor("q_1_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor q_1_squeeze_mask_0 = const()[name = tensor("q_1_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor q_1_cast_fp16 = slice_by_index(begin = q_1_begin_0, end = q_1_end_0, end_mask = q_1_end_mask_0, squeeze_mask = q_1_squeeze_mask_0, x = qkv_1_cast_fp16)[name = tensor("q_1_cast_fp16")]; + tensor k_1_begin_0 = const()[name = tensor("k_1_begin_0"), val = tensor([0, 0, 1, 0, 0])]; + tensor k_1_end_0 = const()[name = tensor("k_1_end_0"), val = tensor([1, 1, 2, 16, 64])]; + tensor k_1_end_mask_0 = const()[name = tensor("k_1_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor k_1_squeeze_mask_0 = const()[name = tensor("k_1_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor k_1_cast_fp16 = slice_by_index(begin = k_1_begin_0, end = k_1_end_0, end_mask = k_1_end_mask_0, squeeze_mask = k_1_squeeze_mask_0, x = qkv_1_cast_fp16)[name = tensor("k_1_cast_fp16")]; + tensor v_1_begin_0 = const()[name = tensor("v_1_begin_0"), val = tensor([0, 0, 2, 0, 0])]; + tensor v_1_end_0 = const()[name = tensor("v_1_end_0"), val = tensor([1, 1, 3, 16, 64])]; + tensor v_1_end_mask_0 = const()[name = tensor("v_1_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor v_1_squeeze_mask_0 = const()[name = tensor("v_1_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor v_1_cast_fp16 = slice_by_index(begin = v_1_begin_0, end = v_1_end_0, end_mask = v_1_end_mask_0, squeeze_mask = v_1_squeeze_mask_0, x = qkv_1_cast_fp16)[name = tensor("v_1_cast_fp16")]; + tensor var_140 = const()[name = tensor("op_140"), val = tensor([1, 1, 1, 1])]; + tensor position0_to_fp16_dtype_0 = const()[name = tensor("position0_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor position0_to_fp16 = cast(dtype = position0_to_fp16_dtype_0, x = position0)[name = tensor("cast_139")]; + tensor ts_5_cast_fp16 = reshape(shape = var_140, x = position0_to_fp16)[name = tensor("ts_5_cast_fp16")]; + tensor var_147 = const()[name = tensor("op_147"), val = tensor([1, 1, 16, 32, 2])]; + tensor q_complex_1_cast_fp16 = reshape(shape = var_147, x = q_1_cast_fp16)[name = tensor("q_complex_1_cast_fp16")]; + tensor var_154 = const()[name = tensor("op_154"), val = tensor([1, 1, 16, 32, 2])]; + tensor k_complex_1_cast_fp16 = reshape(shape = var_154, x = k_1_cast_fp16)[name = tensor("k_complex_1_cast_fp16")]; + tensor var_158_begin_0 = const()[name = tensor("op_158_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_158_end_0 = const()[name = tensor("op_158_end_0"), val = tensor([1, 1, 16, 32, 1])]; + tensor var_158_end_mask_0 = const()[name = tensor("op_158_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_158_squeeze_mask_0 = const()[name = tensor("op_158_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_158_cast_fp16 = slice_by_index(begin = var_158_begin_0, end = var_158_end_0, end_mask = var_158_end_mask_0, squeeze_mask = var_158_squeeze_mask_0, x = q_complex_1_cast_fp16)[name = tensor("op_158_cast_fp16")]; + tensor var_166_begin_0 = const()[name = tensor("op_166_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_166_end_0 = const()[name = tensor("op_166_end_0"), val = tensor([1, 1, 16, 32, 2])]; + tensor var_166_end_mask_0 = const()[name = tensor("op_166_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_166_squeeze_mask_0 = const()[name = tensor("op_166_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_166_cast_fp16 = slice_by_index(begin = var_166_begin_0, end = var_166_end_0, end_mask = var_166_end_mask_0, squeeze_mask = var_166_squeeze_mask_0, x = q_complex_1_cast_fp16)[name = tensor("op_166_cast_fp16")]; + tensor var_174_begin_0 = const()[name = tensor("op_174_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_174_end_0 = const()[name = tensor("op_174_end_0"), val = tensor([1, 1, 16, 32, 1])]; + tensor var_174_end_mask_0 = const()[name = tensor("op_174_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_174_squeeze_mask_0 = const()[name = tensor("op_174_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_174_cast_fp16 = slice_by_index(begin = var_174_begin_0, end = var_174_end_0, end_mask = var_174_end_mask_0, squeeze_mask = var_174_squeeze_mask_0, x = k_complex_1_cast_fp16)[name = tensor("op_174_cast_fp16")]; + tensor var_182_begin_0 = const()[name = tensor("op_182_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_182_end_0 = const()[name = tensor("op_182_end_0"), val = tensor([1, 1, 16, 32, 2])]; + tensor var_182_end_mask_0 = const()[name = tensor("op_182_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_182_squeeze_mask_0 = const()[name = tensor("op_182_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_182_cast_fp16 = slice_by_index(begin = var_182_begin_0, end = var_182_end_0, end_mask = var_182_end_mask_0, squeeze_mask = var_182_squeeze_mask_0, x = k_complex_1_cast_fp16)[name = tensor("op_182_cast_fp16")]; + tensor freqs_1_to_fp16 = const()[name = tensor("freqs_1_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(6369728)))]; + tensor var_188_cast_fp16 = mul(x = freqs_1_to_fp16, y = ts_5_cast_fp16)[name = tensor("op_188_cast_fp16")]; + tensor rotr_1_cast_fp16 = cos(x = var_188_cast_fp16)[name = tensor("rotr_1_cast_fp16")]; + tensor roti_1_cast_fp16 = sin(x = var_188_cast_fp16)[name = tensor("roti_1_cast_fp16")]; + tensor var_192_cast_fp16 = mul(x = var_158_cast_fp16, y = rotr_1_cast_fp16)[name = tensor("op_192_cast_fp16")]; + tensor var_193_cast_fp16 = mul(x = var_166_cast_fp16, y = roti_1_cast_fp16)[name = tensor("op_193_cast_fp16")]; + tensor qor_1_cast_fp16 = sub(x = var_192_cast_fp16, y = var_193_cast_fp16)[name = tensor("qor_1_cast_fp16")]; + tensor var_196_cast_fp16 = mul(x = var_158_cast_fp16, y = roti_1_cast_fp16)[name = tensor("op_196_cast_fp16")]; + tensor var_197_cast_fp16 = mul(x = var_166_cast_fp16, y = rotr_1_cast_fp16)[name = tensor("op_197_cast_fp16")]; + tensor qoi_1_cast_fp16 = add(x = var_196_cast_fp16, y = var_197_cast_fp16)[name = tensor("qoi_1_cast_fp16")]; + tensor var_200_cast_fp16 = mul(x = var_174_cast_fp16, y = rotr_1_cast_fp16)[name = tensor("op_200_cast_fp16")]; + tensor var_201_cast_fp16 = mul(x = var_182_cast_fp16, y = roti_1_cast_fp16)[name = tensor("op_201_cast_fp16")]; + tensor kor_1_cast_fp16 = sub(x = var_200_cast_fp16, y = var_201_cast_fp16)[name = tensor("kor_1_cast_fp16")]; + tensor var_204_cast_fp16 = mul(x = var_174_cast_fp16, y = roti_1_cast_fp16)[name = tensor("op_204_cast_fp16")]; + tensor var_205_cast_fp16 = mul(x = var_182_cast_fp16, y = rotr_1_cast_fp16)[name = tensor("op_205_cast_fp16")]; + tensor koi_1_cast_fp16 = add(x = var_204_cast_fp16, y = var_205_cast_fp16)[name = tensor("koi_1_cast_fp16")]; + tensor qo_1_axis_0 = const()[name = tensor("qo_1_axis_0"), val = tensor(-1)]; + tensor qo_1_cast_fp16 = stack(axis = qo_1_axis_0, values = (qor_1_cast_fp16, qoi_1_cast_fp16))[name = tensor("qo_1_cast_fp16")]; + tensor ko_1_axis_0 = const()[name = tensor("ko_1_axis_0"), val = tensor(-1)]; + tensor ko_1_cast_fp16 = stack(axis = ko_1_axis_0, values = (kor_1_cast_fp16, koi_1_cast_fp16))[name = tensor("ko_1_cast_fp16")]; + tensor var_238 = const()[name = tensor("op_238"), val = tensor([1, 1, 16, 64])]; + tensor q_3_cast_fp16 = reshape(shape = var_238, x = qo_1_cast_fp16)[name = tensor("q_3_cast_fp16")]; + tensor var_244 = const()[name = tensor("op_244"), val = tensor([1, 1, 16, 64])]; + tensor k_3_cast_fp16 = reshape(shape = var_244, x = ko_1_cast_fp16)[name = tensor("k_3_cast_fp16")]; + tensor _inversed_269_y_0_to_fp16 = const()[name = tensor("_inversed_269_y_0_to_fp16"), val = tensor(0x1p-9)]; + tensor _inversed_269_cast_fp16 = mul(x = ts_5_cast_fp16, y = _inversed_269_y_0_to_fp16)[name = tensor("_inversed_269_cast_fp16")]; + tensor var_270_cast_fp16 = floor(x = _inversed_269_cast_fp16)[name = tensor("op_270_cast_fp16")]; + tensor var_271_to_fp16 = const()[name = tensor("op_271_to_fp16"), val = tensor(0x1p+9)]; + tensor var_272_cast_fp16 = mul(x = var_270_cast_fp16, y = var_271_to_fp16)[name = tensor("op_272_cast_fp16")]; + tensor write_indices_float_3_cast_fp16 = sub(x = ts_5_cast_fp16, y = var_272_cast_fp16)[name = tensor("write_indices_float_3_cast_fp16")]; + tensor var_279_dtype_0 = const()[name = tensor("op_279_dtype_0"), val = tensor("int32")]; + tensor write_indices_1_reps_0 = const()[name = tensor("write_indices_1_reps_0"), val = tensor([1, 1, 16, 64])]; + tensor write_indices_float_3_cast_fp16_to_int32 = cast(dtype = var_279_dtype_0, x = write_indices_float_3_cast_fp16)[name = tensor("cast_138")]; + tensor write_indices_1 = tile(reps = write_indices_1_reps_0, x = write_indices_float_3_cast_fp16_to_int32)[name = tensor("write_indices_1")]; + tensor var_289_begin_0 = const()[name = tensor("op_289_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_289_end_0 = const()[name = tensor("op_289_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor var_289_end_mask_0 = const()[name = tensor("op_289_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_289_squeeze_mask_0 = const()[name = tensor("op_289_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor cache0_to_fp16_dtype_0 = const()[name = tensor("cache0_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor cache0_to_fp16 = cast(dtype = cache0_to_fp16_dtype_0, x = cache0)[name = tensor("cast_137")]; + tensor var_289_cast_fp16 = slice_by_index(begin = var_289_begin_0, end = var_289_end_0, end_mask = var_289_end_mask_0, squeeze_mask = var_289_squeeze_mask_0, x = cache0_to_fp16)[name = tensor("op_289_cast_fp16")]; + tensor var_291_axis_0 = const()[name = tensor("op_291_axis_0"), val = tensor(1)]; + tensor var_291_mode_0 = const()[name = tensor("op_291_mode_0"), val = tensor("update")]; + tensor var_291_validate_indices_0 = const()[name = tensor("op_291_validate_indices_0"), val = tensor(false)]; + tensor var_291_cast_fp16 = scatter_along_axis(axis = var_291_axis_0, data = var_289_cast_fp16, indices = write_indices_1, mode = var_291_mode_0, updates = k_3_cast_fp16, validate_indices = var_291_validate_indices_0)[name = tensor("op_291_cast_fp16")]; + tensor concat_2 = const()[name = tensor("concat_2"), val = tensor([0, 0, 0, 0, 0])]; + tensor concat_3 = const()[name = tensor("concat_3"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_1_internal_tensor_assign_1_stride_0 = const()[name = tensor("new_cache_1_internal_tensor_assign_1_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_1_internal_tensor_assign_1_begin_mask_0 = const()[name = tensor("new_cache_1_internal_tensor_assign_1_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_1_internal_tensor_assign_1_end_mask_0 = const()[name = tensor("new_cache_1_internal_tensor_assign_1_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_1_internal_tensor_assign_1_squeeze_mask_0 = const()[name = tensor("new_cache_1_internal_tensor_assign_1_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_12 = const()[name = tensor("shape_12"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_0 = const()[name = tensor("reduce_prod_0"), val = tensor(1048576)]; + tensor range_1d_0_start_0 = const()[name = tensor("range_1d_0_start_0"), val = tensor(0)]; + tensor range_1d_0_step_0 = const()[name = tensor("range_1d_0_step_0"), val = tensor(1)]; + tensor range_1d_0 = range_1d(end = reduce_prod_0, start = range_1d_0_start_0, step = range_1d_0_step_0)[name = tensor("range_1d_0")]; + tensor reshape_0 = reshape(shape = shape_12, x = range_1d_0)[name = tensor("reshape_0")]; + tensor slice_by_index_0 = slice_by_index(begin = concat_2, begin_mask = new_cache_1_internal_tensor_assign_1_begin_mask_0, end = concat_3, end_mask = new_cache_1_internal_tensor_assign_1_end_mask_0, squeeze_mask = new_cache_1_internal_tensor_assign_1_squeeze_mask_0, stride = new_cache_1_internal_tensor_assign_1_stride_0, x = reshape_0)[name = tensor("slice_by_index_0")]; + tensor reshape_1_shape_0 = const()[name = tensor("reshape_1_shape_0"), val = tensor([-1])]; + tensor reshape_1 = reshape(shape = reshape_1_shape_0, x = slice_by_index_0)[name = tensor("reshape_1")]; + tensor reshape_2_shape_0 = const()[name = tensor("reshape_2_shape_0"), val = tensor([-1])]; + tensor reshape_2_cast_fp16 = reshape(shape = reshape_2_shape_0, x = var_291_cast_fp16)[name = tensor("reshape_2_cast_fp16")]; + tensor reshape_3_shape_0 = const()[name = tensor("reshape_3_shape_0"), val = tensor([-1])]; + tensor reshape_3_cast_fp16 = reshape(shape = reshape_3_shape_0, x = cache0_to_fp16)[name = tensor("reshape_3_cast_fp16")]; + tensor scatter_0_mode_0 = const()[name = tensor("scatter_0_mode_0"), val = tensor("update")]; + tensor scatter_0_axis_0 = const()[name = tensor("scatter_0_axis_0"), val = tensor(0)]; + tensor scatter_0_validate_indices_0 = const()[name = tensor("scatter_0_validate_indices_0"), val = tensor(false)]; + tensor scatter_0_cast_fp16 = scatter(axis = scatter_0_axis_0, data = reshape_3_cast_fp16, indices = reshape_1, mode = scatter_0_mode_0, updates = reshape_2_cast_fp16, validate_indices = scatter_0_validate_indices_0)[name = tensor("scatter_0_cast_fp16")]; + tensor reshape_4_cast_fp16 = reshape(shape = shape_12, x = scatter_0_cast_fp16)[name = tensor("reshape_4_cast_fp16")]; + tensor var_299_begin_0 = const()[name = tensor("op_299_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor var_299_end_0 = const()[name = tensor("op_299_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor var_299_end_mask_0 = const()[name = tensor("op_299_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_299_squeeze_mask_0 = const()[name = tensor("op_299_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor var_299_cast_fp16 = slice_by_index(begin = var_299_begin_0, end = var_299_end_0, end_mask = var_299_end_mask_0, squeeze_mask = var_299_squeeze_mask_0, x = reshape_4_cast_fp16)[name = tensor("op_299_cast_fp16")]; + tensor var_301_axis_0 = const()[name = tensor("op_301_axis_0"), val = tensor(1)]; + tensor var_301_mode_0 = const()[name = tensor("op_301_mode_0"), val = tensor("update")]; + tensor var_301_validate_indices_0 = const()[name = tensor("op_301_validate_indices_0"), val = tensor(false)]; + tensor var_301_cast_fp16 = scatter_along_axis(axis = var_301_axis_0, data = var_299_cast_fp16, indices = write_indices_1, mode = var_301_mode_0, updates = v_1_cast_fp16, validate_indices = var_301_validate_indices_0)[name = tensor("op_301_cast_fp16")]; + tensor concat_4 = const()[name = tensor("concat_4"), val = tensor([1, 0, 0, 0, 0])]; + tensor concat_5 = const()[name = tensor("concat_5"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_1_internal_tensor_assign_2_stride_0 = const()[name = tensor("new_cache_1_internal_tensor_assign_2_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_1_internal_tensor_assign_2_begin_mask_0 = const()[name = tensor("new_cache_1_internal_tensor_assign_2_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_1_internal_tensor_assign_2_end_mask_0 = const()[name = tensor("new_cache_1_internal_tensor_assign_2_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_1_internal_tensor_assign_2_squeeze_mask_0 = const()[name = tensor("new_cache_1_internal_tensor_assign_2_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_13 = const()[name = tensor("shape_13"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_1 = const()[name = tensor("reduce_prod_1"), val = tensor(1048576)]; + tensor range_1d_1_start_0 = const()[name = tensor("range_1d_1_start_0"), val = tensor(0)]; + tensor range_1d_1_step_0 = const()[name = tensor("range_1d_1_step_0"), val = tensor(1)]; + tensor range_1d_1 = range_1d(end = reduce_prod_1, start = range_1d_1_start_0, step = range_1d_1_step_0)[name = tensor("range_1d_1")]; + tensor reshape_5 = reshape(shape = shape_13, x = range_1d_1)[name = tensor("reshape_5")]; + tensor slice_by_index_1 = slice_by_index(begin = concat_4, begin_mask = new_cache_1_internal_tensor_assign_2_begin_mask_0, end = concat_5, end_mask = new_cache_1_internal_tensor_assign_2_end_mask_0, squeeze_mask = new_cache_1_internal_tensor_assign_2_squeeze_mask_0, stride = new_cache_1_internal_tensor_assign_2_stride_0, x = reshape_5)[name = tensor("slice_by_index_1")]; + tensor reshape_6_shape_0 = const()[name = tensor("reshape_6_shape_0"), val = tensor([-1])]; + tensor reshape_6 = reshape(shape = reshape_6_shape_0, x = slice_by_index_1)[name = tensor("reshape_6")]; + tensor reshape_7_shape_0 = const()[name = tensor("reshape_7_shape_0"), val = tensor([-1])]; + tensor reshape_7_cast_fp16 = reshape(shape = reshape_7_shape_0, x = var_301_cast_fp16)[name = tensor("reshape_7_cast_fp16")]; + tensor reshape_8_shape_0 = const()[name = tensor("reshape_8_shape_0"), val = tensor([-1])]; + tensor reshape_8_cast_fp16 = reshape(shape = reshape_8_shape_0, x = reshape_4_cast_fp16)[name = tensor("reshape_8_cast_fp16")]; + tensor scatter_1_mode_0 = const()[name = tensor("scatter_1_mode_0"), val = tensor("update")]; + tensor scatter_1_axis_0 = const()[name = tensor("scatter_1_axis_0"), val = tensor(0)]; + tensor scatter_1_validate_indices_0 = const()[name = tensor("scatter_1_validate_indices_0"), val = tensor(false)]; + tensor scatter_1_cast_fp16 = scatter(axis = scatter_1_axis_0, data = reshape_8_cast_fp16, indices = reshape_6, mode = scatter_1_mode_0, updates = reshape_7_cast_fp16, validate_indices = scatter_1_validate_indices_0)[name = tensor("scatter_1_cast_fp16")]; + tensor reshape_9_cast_fp16 = reshape(shape = shape_13, x = scatter_1_cast_fp16)[name = tensor("reshape_9_cast_fp16")]; + tensor reshape_9_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("reshape_9_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor keys_1_begin_0 = const()[name = tensor("keys_1_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor keys_1_end_0 = const()[name = tensor("keys_1_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor keys_1_end_mask_0 = const()[name = tensor("keys_1_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor keys_1_squeeze_mask_0 = const()[name = tensor("keys_1_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor keys_1_cast_fp16 = slice_by_index(begin = keys_1_begin_0, end = keys_1_end_0, end_mask = keys_1_end_mask_0, squeeze_mask = keys_1_squeeze_mask_0, x = reshape_9_cast_fp16)[name = tensor("keys_1_cast_fp16")]; + tensor values_1_begin_0 = const()[name = tensor("values_1_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor values_1_end_0 = const()[name = tensor("values_1_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor values_1_end_mask_0 = const()[name = tensor("values_1_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor values_1_squeeze_mask_0 = const()[name = tensor("values_1_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor values_1_cast_fp16 = slice_by_index(begin = values_1_begin_0, end = values_1_end_0, end_mask = values_1_end_mask_0, squeeze_mask = values_1_squeeze_mask_0, x = reshape_9_cast_fp16)[name = tensor("values_1_cast_fp16")]; + tensor var_313_cast_fp16 = not_equal(x = keys_1_cast_fp16, y = keys_1_cast_fp16)[name = tensor("op_313_cast_fp16")]; + tensor var_319_to_fp16 = const()[name = tensor("op_319_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(6369856)))]; + tensor keys_3_cast_fp16 = select(a = var_319_to_fp16, b = keys_1_cast_fp16, cond = var_313_cast_fp16)[name = tensor("keys_3_cast_fp16")]; + tensor var_321_cast_fp16 = not_equal(x = values_1_cast_fp16, y = values_1_cast_fp16)[name = tensor("op_321_cast_fp16")]; + tensor values_3_cast_fp16 = select(a = var_319_to_fp16, b = values_1_cast_fp16, cond = var_321_cast_fp16)[name = tensor("values_3_cast_fp16")]; + tensor var_345 = const()[name = tensor("op_345"), val = tensor([0, 2, 1, 3])]; + tensor var_361 = const()[name = tensor("op_361"), val = tensor([1, 1, 1])]; + tensor var_362_cast_fp16 = reshape(shape = var_361, x = position0_to_fp16)[name = tensor("op_362_cast_fp16")]; + tensor var_382_to_fp16 = const()[name = tensor("op_382_to_fp16"), val = tensor(0x1p+0)]; + tensor valid_len_1_cast_fp16 = add(x = var_362_cast_fp16, y = var_382_to_fp16)[name = tensor("valid_len_1_cast_fp16")]; + tensor k_positions_1_promoted_to_fp16 = const()[name = tensor("k_positions_1_promoted_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(7418496)))]; + tensor valid_mask_1_cast_fp16 = less(x = k_positions_1_promoted_to_fp16, y = valid_len_1_cast_fp16)[name = tensor("valid_mask_1_cast_fp16")]; + tensor causal_mask_1_cast_fp16 = less_equal(x = k_positions_1_promoted_to_fp16, y = var_362_cast_fp16)[name = tensor("causal_mask_1_cast_fp16")]; + tensor attn_mask_1 = logical_and(x = valid_mask_1_cast_fp16, y = causal_mask_1_cast_fp16)[name = tensor("attn_mask_1")]; + tensor attn_mask_3_axes_0 = const()[name = tensor("attn_mask_3_axes_0"), val = tensor([1])]; + tensor attn_mask_3 = expand_dims(axes = attn_mask_3_axes_0, x = attn_mask_1)[name = tensor("attn_mask_3")]; + tensor var_400_transpose_x_0 = const()[name = tensor("op_400_transpose_x_0"), val = tensor(false)]; + tensor var_400_transpose_y_0 = const()[name = tensor("op_400_transpose_y_0"), val = tensor(false)]; + tensor transpose_18_perm_0 = const()[name = tensor("transpose_18_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_19_perm_0 = const()[name = tensor("transpose_19_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_19 = transpose(perm = transpose_19_perm_0, x = keys_3_cast_fp16)[name = tensor("transpose_51")]; + tensor transpose_18 = transpose(perm = transpose_18_perm_0, x = q_3_cast_fp16)[name = tensor("transpose_52")]; + tensor var_400_cast_fp16 = matmul(transpose_x = var_400_transpose_x_0, transpose_y = var_400_transpose_y_0, x = transpose_18, y = transpose_19)[name = tensor("op_400_cast_fp16")]; + tensor var_394_to_fp16 = const()[name = tensor("op_394_to_fp16"), val = tensor([0x1p-3])]; + tensor attn_weights_1_cast_fp16 = mul(x = var_400_cast_fp16, y = var_394_to_fp16)[name = tensor("attn_weights_1_cast_fp16")]; + tensor var_402 = logical_not(x = attn_mask_3)[name = tensor("op_402")]; + tensor var_403_to_fp16 = const()[name = tensor("op_403_to_fp16"), val = tensor(-inf)]; + tensor attn_weights_3_cast_fp16 = select(a = var_403_to_fp16, b = attn_weights_1_cast_fp16, cond = var_402)[name = tensor("attn_weights_3_cast_fp16")]; + tensor var_405 = const()[name = tensor("op_405"), val = tensor(-1)]; + tensor attn_weights_5_cast_fp16 = softmax(axis = var_405, x = attn_weights_3_cast_fp16)[name = tensor("attn_weights_5_cast_fp16")]; + tensor attn_output_1_transpose_x_0 = const()[name = tensor("attn_output_1_transpose_x_0"), val = tensor(false)]; + tensor attn_output_1_transpose_y_0 = const()[name = tensor("attn_output_1_transpose_y_0"), val = tensor(false)]; + tensor values_5_cast_fp16 = transpose(perm = var_345, x = values_3_cast_fp16)[name = tensor("transpose_53")]; + tensor attn_output_1_cast_fp16 = matmul(transpose_x = attn_output_1_transpose_x_0, transpose_y = attn_output_1_transpose_y_0, x = attn_weights_5_cast_fp16, y = values_5_cast_fp16)[name = tensor("attn_output_1_cast_fp16")]; + tensor var_413 = const()[name = tensor("op_413"), val = tensor([0, 2, 1, 3])]; + tensor var_418 = const()[name = tensor("op_418"), val = tensor([1, 1, 1024])]; + tensor var_414_cast_fp16 = transpose(perm = var_413, x = attn_output_1_cast_fp16)[name = tensor("transpose_50")]; + tensor input_7_cast_fp16 = reshape(shape = var_418, x = var_414_cast_fp16)[name = tensor("input_7_cast_fp16")]; + tensor attn0_out_proj_weight_to_fp16 = const()[name = tensor("attn0_out_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(7419584)))]; + tensor linear_2_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = attn0_out_proj_weight_to_fp16, x = input_7_cast_fp16)[name = tensor("linear_2_cast_fp16")]; + tensor var_424_to_fp16 = const()[name = tensor("op_424_to_fp16"), val = tensor(0x1p+0)]; + tensor var_425_cast_fp16 = add(x = position0_to_fp16, y = var_424_to_fp16)[name = tensor("op_425_cast_fp16")]; + tensor var_425_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_425_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor input_9_cast_fp16 = add(x = linear_0_cast_fp16, y = linear_2_cast_fp16)[name = tensor("input_9_cast_fp16")]; + tensor input_11_axes_0 = const()[name = tensor("input_11_axes_0"), val = tensor([-1])]; + tensor norm0_2_weight_to_fp16 = const()[name = tensor("norm0_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9516800)))]; + tensor norm0_2_bias_to_fp16 = const()[name = tensor("norm0_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9518912)))]; + tensor var_429_to_fp16 = const()[name = tensor("op_429_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_11_cast_fp16 = layer_norm(axes = input_11_axes_0, beta = norm0_2_bias_to_fp16, epsilon = var_429_to_fp16, gamma = norm0_2_weight_to_fp16, x = input_9_cast_fp16)[name = tensor("input_11_cast_fp16")]; + tensor linear0_1_weight_to_fp16 = const()[name = tensor("linear0_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9521024)))]; + tensor linear_3_bias_0_to_fp16 = const()[name = tensor("linear_3_bias_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(17909696)))]; + tensor linear_3_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = linear0_1_weight_to_fp16, x = input_11_cast_fp16)[name = tensor("linear_3_cast_fp16")]; + tensor input_13_mode_0 = const()[name = tensor("input_13_mode_0"), val = tensor("EXACT")]; + tensor input_13_cast_fp16 = gelu(mode = input_13_mode_0, x = linear_3_cast_fp16)[name = tensor("input_13_cast_fp16")]; + tensor linear0_2_weight_to_fp16 = const()[name = tensor("linear0_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(17917952)))]; + tensor linear_4_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = linear0_2_weight_to_fp16, x = input_13_cast_fp16)[name = tensor("linear_4_cast_fp16")]; + tensor input_15_cast_fp16 = add(x = input_9_cast_fp16, y = linear_4_cast_fp16)[name = tensor("input_15_cast_fp16")]; + tensor input_17_axes_0 = const()[name = tensor("input_17_axes_0"), val = tensor([-1])]; + tensor norm1_1_weight_to_fp16 = const()[name = tensor("norm1_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(26306624)))]; + tensor norm1_1_bias_to_fp16 = const()[name = tensor("norm1_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(26308736)))]; + tensor var_446_to_fp16 = const()[name = tensor("op_446_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_17_cast_fp16 = layer_norm(axes = input_17_axes_0, beta = norm1_1_bias_to_fp16, epsilon = var_446_to_fp16, gamma = norm1_1_weight_to_fp16, x = input_15_cast_fp16)[name = tensor("input_17_cast_fp16")]; + tensor attn1_in_proj_weight_to_fp16 = const()[name = tensor("attn1_in_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(26310848)))]; + tensor linear_5_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = attn1_in_proj_weight_to_fp16, x = input_17_cast_fp16)[name = tensor("linear_5_cast_fp16")]; + tensor var_460 = const()[name = tensor("op_460"), val = tensor([1, 1, 3, 16, 64])]; + tensor qkv_3_cast_fp16 = reshape(shape = var_460, x = linear_5_cast_fp16)[name = tensor("qkv_3_cast_fp16")]; + tensor q_7_begin_0 = const()[name = tensor("q_7_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor q_7_end_0 = const()[name = tensor("q_7_end_0"), val = tensor([1, 1, 1, 16, 64])]; + tensor q_7_end_mask_0 = const()[name = tensor("q_7_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor q_7_squeeze_mask_0 = const()[name = tensor("q_7_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor q_7_cast_fp16 = slice_by_index(begin = q_7_begin_0, end = q_7_end_0, end_mask = q_7_end_mask_0, squeeze_mask = q_7_squeeze_mask_0, x = qkv_3_cast_fp16)[name = tensor("q_7_cast_fp16")]; + tensor k_5_begin_0 = const()[name = tensor("k_5_begin_0"), val = tensor([0, 0, 1, 0, 0])]; + tensor k_5_end_0 = const()[name = tensor("k_5_end_0"), val = tensor([1, 1, 2, 16, 64])]; + tensor k_5_end_mask_0 = const()[name = tensor("k_5_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor k_5_squeeze_mask_0 = const()[name = tensor("k_5_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor k_5_cast_fp16 = slice_by_index(begin = k_5_begin_0, end = k_5_end_0, end_mask = k_5_end_mask_0, squeeze_mask = k_5_squeeze_mask_0, x = qkv_3_cast_fp16)[name = tensor("k_5_cast_fp16")]; + tensor v_3_begin_0 = const()[name = tensor("v_3_begin_0"), val = tensor([0, 0, 2, 0, 0])]; + tensor v_3_end_0 = const()[name = tensor("v_3_end_0"), val = tensor([1, 1, 3, 16, 64])]; + tensor v_3_end_mask_0 = const()[name = tensor("v_3_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor v_3_squeeze_mask_0 = const()[name = tensor("v_3_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor v_3_cast_fp16 = slice_by_index(begin = v_3_begin_0, end = v_3_end_0, end_mask = v_3_end_mask_0, squeeze_mask = v_3_squeeze_mask_0, x = qkv_3_cast_fp16)[name = tensor("v_3_cast_fp16")]; + tensor var_526 = const()[name = tensor("op_526"), val = tensor([1, 1, 1, 1])]; + tensor position1_to_fp16_dtype_0 = const()[name = tensor("position1_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor position1_to_fp16 = cast(dtype = position1_to_fp16_dtype_0, x = position1)[name = tensor("cast_134")]; + tensor ts_11_cast_fp16 = reshape(shape = var_526, x = position1_to_fp16)[name = tensor("ts_11_cast_fp16")]; + tensor var_533 = const()[name = tensor("op_533"), val = tensor([1, 1, 16, 32, 2])]; + tensor q_complex_3_cast_fp16 = reshape(shape = var_533, x = q_7_cast_fp16)[name = tensor("q_complex_3_cast_fp16")]; + tensor var_540 = const()[name = tensor("op_540"), val = tensor([1, 1, 16, 32, 2])]; + tensor k_complex_3_cast_fp16 = reshape(shape = var_540, x = k_5_cast_fp16)[name = tensor("k_complex_3_cast_fp16")]; + tensor var_544_begin_0 = const()[name = tensor("op_544_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_544_end_0 = const()[name = tensor("op_544_end_0"), val = tensor([1, 1, 16, 32, 1])]; + tensor var_544_end_mask_0 = const()[name = tensor("op_544_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_544_squeeze_mask_0 = const()[name = tensor("op_544_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_544_cast_fp16 = slice_by_index(begin = var_544_begin_0, end = var_544_end_0, end_mask = var_544_end_mask_0, squeeze_mask = var_544_squeeze_mask_0, x = q_complex_3_cast_fp16)[name = tensor("op_544_cast_fp16")]; + tensor var_552_begin_0 = const()[name = tensor("op_552_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_552_end_0 = const()[name = tensor("op_552_end_0"), val = tensor([1, 1, 16, 32, 2])]; + tensor var_552_end_mask_0 = const()[name = tensor("op_552_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_552_squeeze_mask_0 = const()[name = tensor("op_552_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_552_cast_fp16 = slice_by_index(begin = var_552_begin_0, end = var_552_end_0, end_mask = var_552_end_mask_0, squeeze_mask = var_552_squeeze_mask_0, x = q_complex_3_cast_fp16)[name = tensor("op_552_cast_fp16")]; + tensor var_560_begin_0 = const()[name = tensor("op_560_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_560_end_0 = const()[name = tensor("op_560_end_0"), val = tensor([1, 1, 16, 32, 1])]; + tensor var_560_end_mask_0 = const()[name = tensor("op_560_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_560_squeeze_mask_0 = const()[name = tensor("op_560_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_560_cast_fp16 = slice_by_index(begin = var_560_begin_0, end = var_560_end_0, end_mask = var_560_end_mask_0, squeeze_mask = var_560_squeeze_mask_0, x = k_complex_3_cast_fp16)[name = tensor("op_560_cast_fp16")]; + tensor var_568_begin_0 = const()[name = tensor("op_568_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_568_end_0 = const()[name = tensor("op_568_end_0"), val = tensor([1, 1, 16, 32, 2])]; + tensor var_568_end_mask_0 = const()[name = tensor("op_568_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_568_squeeze_mask_0 = const()[name = tensor("op_568_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_568_cast_fp16 = slice_by_index(begin = var_568_begin_0, end = var_568_end_0, end_mask = var_568_end_mask_0, squeeze_mask = var_568_squeeze_mask_0, x = k_complex_3_cast_fp16)[name = tensor("op_568_cast_fp16")]; + tensor freqs_3_to_fp16 = const()[name = tensor("freqs_3_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(32602368)))]; + tensor var_574_cast_fp16 = mul(x = freqs_3_to_fp16, y = ts_11_cast_fp16)[name = tensor("op_574_cast_fp16")]; + tensor rotr_3_cast_fp16 = cos(x = var_574_cast_fp16)[name = tensor("rotr_3_cast_fp16")]; + tensor roti_3_cast_fp16 = sin(x = var_574_cast_fp16)[name = tensor("roti_3_cast_fp16")]; + tensor var_578_cast_fp16 = mul(x = var_544_cast_fp16, y = rotr_3_cast_fp16)[name = tensor("op_578_cast_fp16")]; + tensor var_579_cast_fp16 = mul(x = var_552_cast_fp16, y = roti_3_cast_fp16)[name = tensor("op_579_cast_fp16")]; + tensor qor_5_cast_fp16 = sub(x = var_578_cast_fp16, y = var_579_cast_fp16)[name = tensor("qor_5_cast_fp16")]; + tensor var_582_cast_fp16 = mul(x = var_544_cast_fp16, y = roti_3_cast_fp16)[name = tensor("op_582_cast_fp16")]; + tensor var_583_cast_fp16 = mul(x = var_552_cast_fp16, y = rotr_3_cast_fp16)[name = tensor("op_583_cast_fp16")]; + tensor qoi_5_cast_fp16 = add(x = var_582_cast_fp16, y = var_583_cast_fp16)[name = tensor("qoi_5_cast_fp16")]; + tensor var_586_cast_fp16 = mul(x = var_560_cast_fp16, y = rotr_3_cast_fp16)[name = tensor("op_586_cast_fp16")]; + tensor var_587_cast_fp16 = mul(x = var_568_cast_fp16, y = roti_3_cast_fp16)[name = tensor("op_587_cast_fp16")]; + tensor kor_5_cast_fp16 = sub(x = var_586_cast_fp16, y = var_587_cast_fp16)[name = tensor("kor_5_cast_fp16")]; + tensor var_590_cast_fp16 = mul(x = var_560_cast_fp16, y = roti_3_cast_fp16)[name = tensor("op_590_cast_fp16")]; + tensor var_591_cast_fp16 = mul(x = var_568_cast_fp16, y = rotr_3_cast_fp16)[name = tensor("op_591_cast_fp16")]; + tensor koi_5_cast_fp16 = add(x = var_590_cast_fp16, y = var_591_cast_fp16)[name = tensor("koi_5_cast_fp16")]; + tensor qo_3_axis_0 = const()[name = tensor("qo_3_axis_0"), val = tensor(-1)]; + tensor qo_3_cast_fp16 = stack(axis = qo_3_axis_0, values = (qor_5_cast_fp16, qoi_5_cast_fp16))[name = tensor("qo_3_cast_fp16")]; + tensor ko_3_axis_0 = const()[name = tensor("ko_3_axis_0"), val = tensor(-1)]; + tensor ko_3_cast_fp16 = stack(axis = ko_3_axis_0, values = (kor_5_cast_fp16, koi_5_cast_fp16))[name = tensor("ko_3_cast_fp16")]; + tensor var_624 = const()[name = tensor("op_624"), val = tensor([1, 1, 16, 64])]; + tensor q_9_cast_fp16 = reshape(shape = var_624, x = qo_3_cast_fp16)[name = tensor("q_9_cast_fp16")]; + tensor var_630 = const()[name = tensor("op_630"), val = tensor([1, 1, 16, 64])]; + tensor k_7_cast_fp16 = reshape(shape = var_630, x = ko_3_cast_fp16)[name = tensor("k_7_cast_fp16")]; + tensor _inversed_655_y_0_to_fp16 = const()[name = tensor("_inversed_655_y_0_to_fp16"), val = tensor(0x1p-9)]; + tensor _inversed_655_cast_fp16 = mul(x = ts_11_cast_fp16, y = _inversed_655_y_0_to_fp16)[name = tensor("_inversed_655_cast_fp16")]; + tensor var_656_cast_fp16 = floor(x = _inversed_655_cast_fp16)[name = tensor("op_656_cast_fp16")]; + tensor var_657_to_fp16 = const()[name = tensor("op_657_to_fp16"), val = tensor(0x1p+9)]; + tensor var_658_cast_fp16 = mul(x = var_656_cast_fp16, y = var_657_to_fp16)[name = tensor("op_658_cast_fp16")]; + tensor write_indices_float_7_cast_fp16 = sub(x = ts_11_cast_fp16, y = var_658_cast_fp16)[name = tensor("write_indices_float_7_cast_fp16")]; + tensor var_665_dtype_0 = const()[name = tensor("op_665_dtype_0"), val = tensor("int32")]; + tensor write_indices_3_reps_0 = const()[name = tensor("write_indices_3_reps_0"), val = tensor([1, 1, 16, 64])]; + tensor write_indices_float_7_cast_fp16_to_int32 = cast(dtype = var_665_dtype_0, x = write_indices_float_7_cast_fp16)[name = tensor("cast_133")]; + tensor write_indices_3 = tile(reps = write_indices_3_reps_0, x = write_indices_float_7_cast_fp16_to_int32)[name = tensor("write_indices_3")]; + tensor var_675_begin_0 = const()[name = tensor("op_675_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_675_end_0 = const()[name = tensor("op_675_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor var_675_end_mask_0 = const()[name = tensor("op_675_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_675_squeeze_mask_0 = const()[name = tensor("op_675_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor cache1_to_fp16_dtype_0 = const()[name = tensor("cache1_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor cache1_to_fp16 = cast(dtype = cache1_to_fp16_dtype_0, x = cache1)[name = tensor("cast_132")]; + tensor var_675_cast_fp16 = slice_by_index(begin = var_675_begin_0, end = var_675_end_0, end_mask = var_675_end_mask_0, squeeze_mask = var_675_squeeze_mask_0, x = cache1_to_fp16)[name = tensor("op_675_cast_fp16")]; + tensor var_677_axis_0 = const()[name = tensor("op_677_axis_0"), val = tensor(1)]; + tensor var_677_mode_0 = const()[name = tensor("op_677_mode_0"), val = tensor("update")]; + tensor var_677_validate_indices_0 = const()[name = tensor("op_677_validate_indices_0"), val = tensor(false)]; + tensor var_677_cast_fp16 = scatter_along_axis(axis = var_677_axis_0, data = var_675_cast_fp16, indices = write_indices_3, mode = var_677_mode_0, updates = k_7_cast_fp16, validate_indices = var_677_validate_indices_0)[name = tensor("op_677_cast_fp16")]; + tensor concat_9 = const()[name = tensor("concat_9"), val = tensor([0, 0, 0, 0, 0])]; + tensor concat_10 = const()[name = tensor("concat_10"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_3_internal_tensor_assign_1_stride_0 = const()[name = tensor("new_cache_3_internal_tensor_assign_1_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_3_internal_tensor_assign_1_begin_mask_0 = const()[name = tensor("new_cache_3_internal_tensor_assign_1_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_3_internal_tensor_assign_1_end_mask_0 = const()[name = tensor("new_cache_3_internal_tensor_assign_1_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_3_internal_tensor_assign_1_squeeze_mask_0 = const()[name = tensor("new_cache_3_internal_tensor_assign_1_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_14 = const()[name = tensor("shape_14"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_2 = const()[name = tensor("reduce_prod_2"), val = tensor(1048576)]; + tensor range_1d_2_start_0 = const()[name = tensor("range_1d_2_start_0"), val = tensor(0)]; + tensor range_1d_2_step_0 = const()[name = tensor("range_1d_2_step_0"), val = tensor(1)]; + tensor range_1d_2 = range_1d(end = reduce_prod_2, start = range_1d_2_start_0, step = range_1d_2_step_0)[name = tensor("range_1d_2")]; + tensor reshape_10 = reshape(shape = shape_14, x = range_1d_2)[name = tensor("reshape_10")]; + tensor slice_by_index_2 = slice_by_index(begin = concat_9, begin_mask = new_cache_3_internal_tensor_assign_1_begin_mask_0, end = concat_10, end_mask = new_cache_3_internal_tensor_assign_1_end_mask_0, squeeze_mask = new_cache_3_internal_tensor_assign_1_squeeze_mask_0, stride = new_cache_3_internal_tensor_assign_1_stride_0, x = reshape_10)[name = tensor("slice_by_index_2")]; + tensor reshape_11_shape_0 = const()[name = tensor("reshape_11_shape_0"), val = tensor([-1])]; + tensor reshape_11 = reshape(shape = reshape_11_shape_0, x = slice_by_index_2)[name = tensor("reshape_11")]; + tensor reshape_12_shape_0 = const()[name = tensor("reshape_12_shape_0"), val = tensor([-1])]; + tensor reshape_12_cast_fp16 = reshape(shape = reshape_12_shape_0, x = var_677_cast_fp16)[name = tensor("reshape_12_cast_fp16")]; + tensor reshape_13_shape_0 = const()[name = tensor("reshape_13_shape_0"), val = tensor([-1])]; + tensor reshape_13_cast_fp16 = reshape(shape = reshape_13_shape_0, x = cache1_to_fp16)[name = tensor("reshape_13_cast_fp16")]; + tensor scatter_2_mode_0 = const()[name = tensor("scatter_2_mode_0"), val = tensor("update")]; + tensor scatter_2_axis_0 = const()[name = tensor("scatter_2_axis_0"), val = tensor(0)]; + tensor scatter_2_validate_indices_0 = const()[name = tensor("scatter_2_validate_indices_0"), val = tensor(false)]; + tensor scatter_2_cast_fp16 = scatter(axis = scatter_2_axis_0, data = reshape_13_cast_fp16, indices = reshape_11, mode = scatter_2_mode_0, updates = reshape_12_cast_fp16, validate_indices = scatter_2_validate_indices_0)[name = tensor("scatter_2_cast_fp16")]; + tensor reshape_14_cast_fp16 = reshape(shape = shape_14, x = scatter_2_cast_fp16)[name = tensor("reshape_14_cast_fp16")]; + tensor var_685_begin_0 = const()[name = tensor("op_685_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor var_685_end_0 = const()[name = tensor("op_685_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor var_685_end_mask_0 = const()[name = tensor("op_685_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_685_squeeze_mask_0 = const()[name = tensor("op_685_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor var_685_cast_fp16 = slice_by_index(begin = var_685_begin_0, end = var_685_end_0, end_mask = var_685_end_mask_0, squeeze_mask = var_685_squeeze_mask_0, x = reshape_14_cast_fp16)[name = tensor("op_685_cast_fp16")]; + tensor var_687_axis_0 = const()[name = tensor("op_687_axis_0"), val = tensor(1)]; + tensor var_687_mode_0 = const()[name = tensor("op_687_mode_0"), val = tensor("update")]; + tensor var_687_validate_indices_0 = const()[name = tensor("op_687_validate_indices_0"), val = tensor(false)]; + tensor var_687_cast_fp16 = scatter_along_axis(axis = var_687_axis_0, data = var_685_cast_fp16, indices = write_indices_3, mode = var_687_mode_0, updates = v_3_cast_fp16, validate_indices = var_687_validate_indices_0)[name = tensor("op_687_cast_fp16")]; + tensor concat_11 = const()[name = tensor("concat_11"), val = tensor([1, 0, 0, 0, 0])]; + tensor concat_12 = const()[name = tensor("concat_12"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_3_internal_tensor_assign_2_stride_0 = const()[name = tensor("new_cache_3_internal_tensor_assign_2_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_3_internal_tensor_assign_2_begin_mask_0 = const()[name = tensor("new_cache_3_internal_tensor_assign_2_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_3_internal_tensor_assign_2_end_mask_0 = const()[name = tensor("new_cache_3_internal_tensor_assign_2_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_3_internal_tensor_assign_2_squeeze_mask_0 = const()[name = tensor("new_cache_3_internal_tensor_assign_2_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_15 = const()[name = tensor("shape_15"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_3 = const()[name = tensor("reduce_prod_3"), val = tensor(1048576)]; + tensor range_1d_3_start_0 = const()[name = tensor("range_1d_3_start_0"), val = tensor(0)]; + tensor range_1d_3_step_0 = const()[name = tensor("range_1d_3_step_0"), val = tensor(1)]; + tensor range_1d_3 = range_1d(end = reduce_prod_3, start = range_1d_3_start_0, step = range_1d_3_step_0)[name = tensor("range_1d_3")]; + tensor reshape_15 = reshape(shape = shape_15, x = range_1d_3)[name = tensor("reshape_15")]; + tensor slice_by_index_3 = slice_by_index(begin = concat_11, begin_mask = new_cache_3_internal_tensor_assign_2_begin_mask_0, end = concat_12, end_mask = new_cache_3_internal_tensor_assign_2_end_mask_0, squeeze_mask = new_cache_3_internal_tensor_assign_2_squeeze_mask_0, stride = new_cache_3_internal_tensor_assign_2_stride_0, x = reshape_15)[name = tensor("slice_by_index_3")]; + tensor reshape_16_shape_0 = const()[name = tensor("reshape_16_shape_0"), val = tensor([-1])]; + tensor reshape_16 = reshape(shape = reshape_16_shape_0, x = slice_by_index_3)[name = tensor("reshape_16")]; + tensor reshape_17_shape_0 = const()[name = tensor("reshape_17_shape_0"), val = tensor([-1])]; + tensor reshape_17_cast_fp16 = reshape(shape = reshape_17_shape_0, x = var_687_cast_fp16)[name = tensor("reshape_17_cast_fp16")]; + tensor reshape_18_shape_0 = const()[name = tensor("reshape_18_shape_0"), val = tensor([-1])]; + tensor reshape_18_cast_fp16 = reshape(shape = reshape_18_shape_0, x = reshape_14_cast_fp16)[name = tensor("reshape_18_cast_fp16")]; + tensor scatter_3_mode_0 = const()[name = tensor("scatter_3_mode_0"), val = tensor("update")]; + tensor scatter_3_axis_0 = const()[name = tensor("scatter_3_axis_0"), val = tensor(0)]; + tensor scatter_3_validate_indices_0 = const()[name = tensor("scatter_3_validate_indices_0"), val = tensor(false)]; + tensor scatter_3_cast_fp16 = scatter(axis = scatter_3_axis_0, data = reshape_18_cast_fp16, indices = reshape_16, mode = scatter_3_mode_0, updates = reshape_17_cast_fp16, validate_indices = scatter_3_validate_indices_0)[name = tensor("scatter_3_cast_fp16")]; + tensor reshape_19_cast_fp16 = reshape(shape = shape_15, x = scatter_3_cast_fp16)[name = tensor("reshape_19_cast_fp16")]; + tensor reshape_19_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("reshape_19_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor keys_7_begin_0 = const()[name = tensor("keys_7_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor keys_7_end_0 = const()[name = tensor("keys_7_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor keys_7_end_mask_0 = const()[name = tensor("keys_7_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor keys_7_squeeze_mask_0 = const()[name = tensor("keys_7_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor keys_7_cast_fp16 = slice_by_index(begin = keys_7_begin_0, end = keys_7_end_0, end_mask = keys_7_end_mask_0, squeeze_mask = keys_7_squeeze_mask_0, x = reshape_19_cast_fp16)[name = tensor("keys_7_cast_fp16")]; + tensor values_7_begin_0 = const()[name = tensor("values_7_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor values_7_end_0 = const()[name = tensor("values_7_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor values_7_end_mask_0 = const()[name = tensor("values_7_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor values_7_squeeze_mask_0 = const()[name = tensor("values_7_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor values_7_cast_fp16 = slice_by_index(begin = values_7_begin_0, end = values_7_end_0, end_mask = values_7_end_mask_0, squeeze_mask = values_7_squeeze_mask_0, x = reshape_19_cast_fp16)[name = tensor("values_7_cast_fp16")]; + tensor var_699_cast_fp16 = not_equal(x = keys_7_cast_fp16, y = keys_7_cast_fp16)[name = tensor("op_699_cast_fp16")]; + tensor keys_9_cast_fp16 = select(a = var_319_to_fp16, b = keys_7_cast_fp16, cond = var_699_cast_fp16)[name = tensor("keys_9_cast_fp16")]; + tensor var_707_cast_fp16 = not_equal(x = values_7_cast_fp16, y = values_7_cast_fp16)[name = tensor("op_707_cast_fp16")]; + tensor values_9_cast_fp16 = select(a = var_319_to_fp16, b = values_7_cast_fp16, cond = var_707_cast_fp16)[name = tensor("values_9_cast_fp16")]; + tensor var_731 = const()[name = tensor("op_731"), val = tensor([0, 2, 1, 3])]; + tensor var_747 = const()[name = tensor("op_747"), val = tensor([1, 1, 1])]; + tensor var_748_cast_fp16 = reshape(shape = var_747, x = position1_to_fp16)[name = tensor("op_748_cast_fp16")]; + tensor var_768_to_fp16 = const()[name = tensor("op_768_to_fp16"), val = tensor(0x1p+0)]; + tensor valid_len_3_cast_fp16 = add(x = var_748_cast_fp16, y = var_768_to_fp16)[name = tensor("valid_len_3_cast_fp16")]; + tensor valid_mask_3_cast_fp16 = less(x = k_positions_1_promoted_to_fp16, y = valid_len_3_cast_fp16)[name = tensor("valid_mask_3_cast_fp16")]; + tensor causal_mask_3_cast_fp16 = less_equal(x = k_positions_1_promoted_to_fp16, y = var_748_cast_fp16)[name = tensor("causal_mask_3_cast_fp16")]; + tensor attn_mask_5 = logical_and(x = valid_mask_3_cast_fp16, y = causal_mask_3_cast_fp16)[name = tensor("attn_mask_5")]; + tensor attn_mask_7_axes_0 = const()[name = tensor("attn_mask_7_axes_0"), val = tensor([1])]; + tensor attn_mask_7 = expand_dims(axes = attn_mask_7_axes_0, x = attn_mask_5)[name = tensor("attn_mask_7")]; + tensor var_786_transpose_x_0 = const()[name = tensor("op_786_transpose_x_0"), val = tensor(false)]; + tensor var_786_transpose_y_0 = const()[name = tensor("op_786_transpose_y_0"), val = tensor(false)]; + tensor transpose_20_perm_0 = const()[name = tensor("transpose_20_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_21_perm_0 = const()[name = tensor("transpose_21_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_21 = transpose(perm = transpose_21_perm_0, x = keys_9_cast_fp16)[name = tensor("transpose_47")]; + tensor transpose_20 = transpose(perm = transpose_20_perm_0, x = q_9_cast_fp16)[name = tensor("transpose_48")]; + tensor var_786_cast_fp16 = matmul(transpose_x = var_786_transpose_x_0, transpose_y = var_786_transpose_y_0, x = transpose_20, y = transpose_21)[name = tensor("op_786_cast_fp16")]; + tensor var_780_to_fp16 = const()[name = tensor("op_780_to_fp16"), val = tensor([0x1p-3])]; + tensor attn_weights_7_cast_fp16 = mul(x = var_786_cast_fp16, y = var_780_to_fp16)[name = tensor("attn_weights_7_cast_fp16")]; + tensor var_788 = logical_not(x = attn_mask_7)[name = tensor("op_788")]; + tensor var_789_to_fp16 = const()[name = tensor("op_789_to_fp16"), val = tensor(-inf)]; + tensor attn_weights_9_cast_fp16 = select(a = var_789_to_fp16, b = attn_weights_7_cast_fp16, cond = var_788)[name = tensor("attn_weights_9_cast_fp16")]; + tensor var_791 = const()[name = tensor("op_791"), val = tensor(-1)]; + tensor attn_weights_11_cast_fp16 = softmax(axis = var_791, x = attn_weights_9_cast_fp16)[name = tensor("attn_weights_11_cast_fp16")]; + tensor attn_output_3_transpose_x_0 = const()[name = tensor("attn_output_3_transpose_x_0"), val = tensor(false)]; + tensor attn_output_3_transpose_y_0 = const()[name = tensor("attn_output_3_transpose_y_0"), val = tensor(false)]; + tensor values_11_cast_fp16 = transpose(perm = var_731, x = values_9_cast_fp16)[name = tensor("transpose_49")]; + tensor attn_output_3_cast_fp16 = matmul(transpose_x = attn_output_3_transpose_x_0, transpose_y = attn_output_3_transpose_y_0, x = attn_weights_11_cast_fp16, y = values_11_cast_fp16)[name = tensor("attn_output_3_cast_fp16")]; + tensor var_799 = const()[name = tensor("op_799"), val = tensor([0, 2, 1, 3])]; + tensor var_804 = const()[name = tensor("op_804"), val = tensor([1, 1, 1024])]; + tensor var_800_cast_fp16 = transpose(perm = var_799, x = attn_output_3_cast_fp16)[name = tensor("transpose_46")]; + tensor input_19_cast_fp16 = reshape(shape = var_804, x = var_800_cast_fp16)[name = tensor("input_19_cast_fp16")]; + tensor attn1_out_proj_weight_to_fp16 = const()[name = tensor("attn1_out_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(32602496)))]; + tensor linear_6_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = attn1_out_proj_weight_to_fp16, x = input_19_cast_fp16)[name = tensor("linear_6_cast_fp16")]; + tensor var_810_to_fp16 = const()[name = tensor("op_810_to_fp16"), val = tensor(0x1p+0)]; + tensor var_811_cast_fp16 = add(x = position1_to_fp16, y = var_810_to_fp16)[name = tensor("op_811_cast_fp16")]; + tensor var_811_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_811_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor input_21_cast_fp16 = add(x = input_15_cast_fp16, y = linear_6_cast_fp16)[name = tensor("input_21_cast_fp16")]; + tensor input_23_axes_0 = const()[name = tensor("input_23_axes_0"), val = tensor([-1])]; + tensor norm1_2_weight_to_fp16 = const()[name = tensor("norm1_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(34699712)))]; + tensor norm1_2_bias_to_fp16 = const()[name = tensor("norm1_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(34701824)))]; + tensor var_815_to_fp16 = const()[name = tensor("op_815_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_23_cast_fp16 = layer_norm(axes = input_23_axes_0, beta = norm1_2_bias_to_fp16, epsilon = var_815_to_fp16, gamma = norm1_2_weight_to_fp16, x = input_21_cast_fp16)[name = tensor("input_23_cast_fp16")]; + tensor linear1_1_weight_to_fp16 = const()[name = tensor("linear1_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(34703936)))]; + tensor linear_7_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = linear1_1_weight_to_fp16, x = input_23_cast_fp16)[name = tensor("linear_7_cast_fp16")]; + tensor input_25_mode_0 = const()[name = tensor("input_25_mode_0"), val = tensor("EXACT")]; + tensor input_25_cast_fp16 = gelu(mode = input_25_mode_0, x = linear_7_cast_fp16)[name = tensor("input_25_cast_fp16")]; + tensor linear1_2_weight_to_fp16 = const()[name = tensor("linear1_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(43092608)))]; + tensor linear_8_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = linear1_2_weight_to_fp16, x = input_25_cast_fp16)[name = tensor("linear_8_cast_fp16")]; + tensor input_27_cast_fp16 = add(x = input_21_cast_fp16, y = linear_8_cast_fp16)[name = tensor("input_27_cast_fp16")]; + tensor input_29_axes_0 = const()[name = tensor("input_29_axes_0"), val = tensor([-1])]; + tensor norm2_1_weight_to_fp16 = const()[name = tensor("norm2_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(51481280)))]; + tensor norm2_1_bias_to_fp16 = const()[name = tensor("norm2_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(51483392)))]; + tensor var_832_to_fp16 = const()[name = tensor("op_832_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_29_cast_fp16 = layer_norm(axes = input_29_axes_0, beta = norm2_1_bias_to_fp16, epsilon = var_832_to_fp16, gamma = norm2_1_weight_to_fp16, x = input_27_cast_fp16)[name = tensor("input_29_cast_fp16")]; + tensor attn2_in_proj_weight_to_fp16 = const()[name = tensor("attn2_in_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(51485504)))]; + tensor linear_9_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = attn2_in_proj_weight_to_fp16, x = input_29_cast_fp16)[name = tensor("linear_9_cast_fp16")]; + tensor var_846 = const()[name = tensor("op_846"), val = tensor([1, 1, 3, 16, 64])]; + tensor qkv_5_cast_fp16 = reshape(shape = var_846, x = linear_9_cast_fp16)[name = tensor("qkv_5_cast_fp16")]; + tensor q_13_begin_0 = const()[name = tensor("q_13_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor q_13_end_0 = const()[name = tensor("q_13_end_0"), val = tensor([1, 1, 1, 16, 64])]; + tensor q_13_end_mask_0 = const()[name = tensor("q_13_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor q_13_squeeze_mask_0 = const()[name = tensor("q_13_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor q_13_cast_fp16 = slice_by_index(begin = q_13_begin_0, end = q_13_end_0, end_mask = q_13_end_mask_0, squeeze_mask = q_13_squeeze_mask_0, x = qkv_5_cast_fp16)[name = tensor("q_13_cast_fp16")]; + tensor k_9_begin_0 = const()[name = tensor("k_9_begin_0"), val = tensor([0, 0, 1, 0, 0])]; + tensor k_9_end_0 = const()[name = tensor("k_9_end_0"), val = tensor([1, 1, 2, 16, 64])]; + tensor k_9_end_mask_0 = const()[name = tensor("k_9_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor k_9_squeeze_mask_0 = const()[name = tensor("k_9_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor k_9_cast_fp16 = slice_by_index(begin = k_9_begin_0, end = k_9_end_0, end_mask = k_9_end_mask_0, squeeze_mask = k_9_squeeze_mask_0, x = qkv_5_cast_fp16)[name = tensor("k_9_cast_fp16")]; + tensor v_5_begin_0 = const()[name = tensor("v_5_begin_0"), val = tensor([0, 0, 2, 0, 0])]; + tensor v_5_end_0 = const()[name = tensor("v_5_end_0"), val = tensor([1, 1, 3, 16, 64])]; + tensor v_5_end_mask_0 = const()[name = tensor("v_5_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor v_5_squeeze_mask_0 = const()[name = tensor("v_5_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor v_5_cast_fp16 = slice_by_index(begin = v_5_begin_0, end = v_5_end_0, end_mask = v_5_end_mask_0, squeeze_mask = v_5_squeeze_mask_0, x = qkv_5_cast_fp16)[name = tensor("v_5_cast_fp16")]; + tensor var_912 = const()[name = tensor("op_912"), val = tensor([1, 1, 1, 1])]; + tensor position2_to_fp16_dtype_0 = const()[name = tensor("position2_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor position2_to_fp16 = cast(dtype = position2_to_fp16_dtype_0, x = position2)[name = tensor("cast_129")]; + tensor ts_17_cast_fp16 = reshape(shape = var_912, x = position2_to_fp16)[name = tensor("ts_17_cast_fp16")]; + tensor var_919 = const()[name = tensor("op_919"), val = tensor([1, 1, 16, 32, 2])]; + tensor q_complex_5_cast_fp16 = reshape(shape = var_919, x = q_13_cast_fp16)[name = tensor("q_complex_5_cast_fp16")]; + tensor var_926 = const()[name = tensor("op_926"), val = tensor([1, 1, 16, 32, 2])]; + tensor k_complex_5_cast_fp16 = reshape(shape = var_926, x = k_9_cast_fp16)[name = tensor("k_complex_5_cast_fp16")]; + tensor var_930_begin_0 = const()[name = tensor("op_930_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_930_end_0 = const()[name = tensor("op_930_end_0"), val = tensor([1, 1, 16, 32, 1])]; + tensor var_930_end_mask_0 = const()[name = tensor("op_930_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_930_squeeze_mask_0 = const()[name = tensor("op_930_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_930_cast_fp16 = slice_by_index(begin = var_930_begin_0, end = var_930_end_0, end_mask = var_930_end_mask_0, squeeze_mask = var_930_squeeze_mask_0, x = q_complex_5_cast_fp16)[name = tensor("op_930_cast_fp16")]; + tensor var_938_begin_0 = const()[name = tensor("op_938_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_938_end_0 = const()[name = tensor("op_938_end_0"), val = tensor([1, 1, 16, 32, 2])]; + tensor var_938_end_mask_0 = const()[name = tensor("op_938_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_938_squeeze_mask_0 = const()[name = tensor("op_938_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_938_cast_fp16 = slice_by_index(begin = var_938_begin_0, end = var_938_end_0, end_mask = var_938_end_mask_0, squeeze_mask = var_938_squeeze_mask_0, x = q_complex_5_cast_fp16)[name = tensor("op_938_cast_fp16")]; + tensor var_946_begin_0 = const()[name = tensor("op_946_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_946_end_0 = const()[name = tensor("op_946_end_0"), val = tensor([1, 1, 16, 32, 1])]; + tensor var_946_end_mask_0 = const()[name = tensor("op_946_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_946_squeeze_mask_0 = const()[name = tensor("op_946_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_946_cast_fp16 = slice_by_index(begin = var_946_begin_0, end = var_946_end_0, end_mask = var_946_end_mask_0, squeeze_mask = var_946_squeeze_mask_0, x = k_complex_5_cast_fp16)[name = tensor("op_946_cast_fp16")]; + tensor var_954_begin_0 = const()[name = tensor("op_954_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_954_end_0 = const()[name = tensor("op_954_end_0"), val = tensor([1, 1, 16, 32, 2])]; + tensor var_954_end_mask_0 = const()[name = tensor("op_954_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_954_squeeze_mask_0 = const()[name = tensor("op_954_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_954_cast_fp16 = slice_by_index(begin = var_954_begin_0, end = var_954_end_0, end_mask = var_954_end_mask_0, squeeze_mask = var_954_squeeze_mask_0, x = k_complex_5_cast_fp16)[name = tensor("op_954_cast_fp16")]; + tensor freqs_5_to_fp16 = const()[name = tensor("freqs_5_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(57777024)))]; + tensor var_960_cast_fp16 = mul(x = freqs_5_to_fp16, y = ts_17_cast_fp16)[name = tensor("op_960_cast_fp16")]; + tensor rotr_5_cast_fp16 = cos(x = var_960_cast_fp16)[name = tensor("rotr_5_cast_fp16")]; + tensor roti_5_cast_fp16 = sin(x = var_960_cast_fp16)[name = tensor("roti_5_cast_fp16")]; + tensor var_964_cast_fp16 = mul(x = var_930_cast_fp16, y = rotr_5_cast_fp16)[name = tensor("op_964_cast_fp16")]; + tensor var_965_cast_fp16 = mul(x = var_938_cast_fp16, y = roti_5_cast_fp16)[name = tensor("op_965_cast_fp16")]; + tensor qor_9_cast_fp16 = sub(x = var_964_cast_fp16, y = var_965_cast_fp16)[name = tensor("qor_9_cast_fp16")]; + tensor var_968_cast_fp16 = mul(x = var_930_cast_fp16, y = roti_5_cast_fp16)[name = tensor("op_968_cast_fp16")]; + tensor var_969_cast_fp16 = mul(x = var_938_cast_fp16, y = rotr_5_cast_fp16)[name = tensor("op_969_cast_fp16")]; + tensor qoi_9_cast_fp16 = add(x = var_968_cast_fp16, y = var_969_cast_fp16)[name = tensor("qoi_9_cast_fp16")]; + tensor var_972_cast_fp16 = mul(x = var_946_cast_fp16, y = rotr_5_cast_fp16)[name = tensor("op_972_cast_fp16")]; + tensor var_973_cast_fp16 = mul(x = var_954_cast_fp16, y = roti_5_cast_fp16)[name = tensor("op_973_cast_fp16")]; + tensor kor_9_cast_fp16 = sub(x = var_972_cast_fp16, y = var_973_cast_fp16)[name = tensor("kor_9_cast_fp16")]; + tensor var_976_cast_fp16 = mul(x = var_946_cast_fp16, y = roti_5_cast_fp16)[name = tensor("op_976_cast_fp16")]; + tensor var_977_cast_fp16 = mul(x = var_954_cast_fp16, y = rotr_5_cast_fp16)[name = tensor("op_977_cast_fp16")]; + tensor koi_9_cast_fp16 = add(x = var_976_cast_fp16, y = var_977_cast_fp16)[name = tensor("koi_9_cast_fp16")]; + tensor qo_5_axis_0 = const()[name = tensor("qo_5_axis_0"), val = tensor(-1)]; + tensor qo_5_cast_fp16 = stack(axis = qo_5_axis_0, values = (qor_9_cast_fp16, qoi_9_cast_fp16))[name = tensor("qo_5_cast_fp16")]; + tensor ko_5_axis_0 = const()[name = tensor("ko_5_axis_0"), val = tensor(-1)]; + tensor ko_5_cast_fp16 = stack(axis = ko_5_axis_0, values = (kor_9_cast_fp16, koi_9_cast_fp16))[name = tensor("ko_5_cast_fp16")]; + tensor var_1010 = const()[name = tensor("op_1010"), val = tensor([1, 1, 16, 64])]; + tensor q_15_cast_fp16 = reshape(shape = var_1010, x = qo_5_cast_fp16)[name = tensor("q_15_cast_fp16")]; + tensor var_1016 = const()[name = tensor("op_1016"), val = tensor([1, 1, 16, 64])]; + tensor k_11_cast_fp16 = reshape(shape = var_1016, x = ko_5_cast_fp16)[name = tensor("k_11_cast_fp16")]; + tensor _inversed_1041_y_0_to_fp16 = const()[name = tensor("_inversed_1041_y_0_to_fp16"), val = tensor(0x1p-9)]; + tensor _inversed_1041_cast_fp16 = mul(x = ts_17_cast_fp16, y = _inversed_1041_y_0_to_fp16)[name = tensor("_inversed_1041_cast_fp16")]; + tensor var_1042_cast_fp16 = floor(x = _inversed_1041_cast_fp16)[name = tensor("op_1042_cast_fp16")]; + tensor var_1043_to_fp16 = const()[name = tensor("op_1043_to_fp16"), val = tensor(0x1p+9)]; + tensor var_1044_cast_fp16 = mul(x = var_1042_cast_fp16, y = var_1043_to_fp16)[name = tensor("op_1044_cast_fp16")]; + tensor write_indices_float_11_cast_fp16 = sub(x = ts_17_cast_fp16, y = var_1044_cast_fp16)[name = tensor("write_indices_float_11_cast_fp16")]; + tensor var_1051_dtype_0 = const()[name = tensor("op_1051_dtype_0"), val = tensor("int32")]; + tensor write_indices_5_reps_0 = const()[name = tensor("write_indices_5_reps_0"), val = tensor([1, 1, 16, 64])]; + tensor write_indices_float_11_cast_fp16_to_int32 = cast(dtype = var_1051_dtype_0, x = write_indices_float_11_cast_fp16)[name = tensor("cast_128")]; + tensor write_indices_5 = tile(reps = write_indices_5_reps_0, x = write_indices_float_11_cast_fp16_to_int32)[name = tensor("write_indices_5")]; + tensor var_1061_begin_0 = const()[name = tensor("op_1061_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_1061_end_0 = const()[name = tensor("op_1061_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor var_1061_end_mask_0 = const()[name = tensor("op_1061_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_1061_squeeze_mask_0 = const()[name = tensor("op_1061_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor cache2_to_fp16_dtype_0 = const()[name = tensor("cache2_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor cache2_to_fp16 = cast(dtype = cache2_to_fp16_dtype_0, x = cache2)[name = tensor("cast_127")]; + tensor var_1061_cast_fp16 = slice_by_index(begin = var_1061_begin_0, end = var_1061_end_0, end_mask = var_1061_end_mask_0, squeeze_mask = var_1061_squeeze_mask_0, x = cache2_to_fp16)[name = tensor("op_1061_cast_fp16")]; + tensor var_1063_axis_0 = const()[name = tensor("op_1063_axis_0"), val = tensor(1)]; + tensor var_1063_mode_0 = const()[name = tensor("op_1063_mode_0"), val = tensor("update")]; + tensor var_1063_validate_indices_0 = const()[name = tensor("op_1063_validate_indices_0"), val = tensor(false)]; + tensor var_1063_cast_fp16 = scatter_along_axis(axis = var_1063_axis_0, data = var_1061_cast_fp16, indices = write_indices_5, mode = var_1063_mode_0, updates = k_11_cast_fp16, validate_indices = var_1063_validate_indices_0)[name = tensor("op_1063_cast_fp16")]; + tensor concat_16 = const()[name = tensor("concat_16"), val = tensor([0, 0, 0, 0, 0])]; + tensor concat_17 = const()[name = tensor("concat_17"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_5_internal_tensor_assign_1_stride_0 = const()[name = tensor("new_cache_5_internal_tensor_assign_1_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_5_internal_tensor_assign_1_begin_mask_0 = const()[name = tensor("new_cache_5_internal_tensor_assign_1_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_5_internal_tensor_assign_1_end_mask_0 = const()[name = tensor("new_cache_5_internal_tensor_assign_1_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_5_internal_tensor_assign_1_squeeze_mask_0 = const()[name = tensor("new_cache_5_internal_tensor_assign_1_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_16 = const()[name = tensor("shape_16"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_4 = const()[name = tensor("reduce_prod_4"), val = tensor(1048576)]; + tensor range_1d_4_start_0 = const()[name = tensor("range_1d_4_start_0"), val = tensor(0)]; + tensor range_1d_4_step_0 = const()[name = tensor("range_1d_4_step_0"), val = tensor(1)]; + tensor range_1d_4 = range_1d(end = reduce_prod_4, start = range_1d_4_start_0, step = range_1d_4_step_0)[name = tensor("range_1d_4")]; + tensor reshape_20 = reshape(shape = shape_16, x = range_1d_4)[name = tensor("reshape_20")]; + tensor slice_by_index_4 = slice_by_index(begin = concat_16, begin_mask = new_cache_5_internal_tensor_assign_1_begin_mask_0, end = concat_17, end_mask = new_cache_5_internal_tensor_assign_1_end_mask_0, squeeze_mask = new_cache_5_internal_tensor_assign_1_squeeze_mask_0, stride = new_cache_5_internal_tensor_assign_1_stride_0, x = reshape_20)[name = tensor("slice_by_index_4")]; + tensor reshape_21_shape_0 = const()[name = tensor("reshape_21_shape_0"), val = tensor([-1])]; + tensor reshape_21 = reshape(shape = reshape_21_shape_0, x = slice_by_index_4)[name = tensor("reshape_21")]; + tensor reshape_22_shape_0 = const()[name = tensor("reshape_22_shape_0"), val = tensor([-1])]; + tensor reshape_22_cast_fp16 = reshape(shape = reshape_22_shape_0, x = var_1063_cast_fp16)[name = tensor("reshape_22_cast_fp16")]; + tensor reshape_23_shape_0 = const()[name = tensor("reshape_23_shape_0"), val = tensor([-1])]; + tensor reshape_23_cast_fp16 = reshape(shape = reshape_23_shape_0, x = cache2_to_fp16)[name = tensor("reshape_23_cast_fp16")]; + tensor scatter_4_mode_0 = const()[name = tensor("scatter_4_mode_0"), val = tensor("update")]; + tensor scatter_4_axis_0 = const()[name = tensor("scatter_4_axis_0"), val = tensor(0)]; + tensor scatter_4_validate_indices_0 = const()[name = tensor("scatter_4_validate_indices_0"), val = tensor(false)]; + tensor scatter_4_cast_fp16 = scatter(axis = scatter_4_axis_0, data = reshape_23_cast_fp16, indices = reshape_21, mode = scatter_4_mode_0, updates = reshape_22_cast_fp16, validate_indices = scatter_4_validate_indices_0)[name = tensor("scatter_4_cast_fp16")]; + tensor reshape_24_cast_fp16 = reshape(shape = shape_16, x = scatter_4_cast_fp16)[name = tensor("reshape_24_cast_fp16")]; + tensor var_1071_begin_0 = const()[name = tensor("op_1071_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor var_1071_end_0 = const()[name = tensor("op_1071_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor var_1071_end_mask_0 = const()[name = tensor("op_1071_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_1071_squeeze_mask_0 = const()[name = tensor("op_1071_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor var_1071_cast_fp16 = slice_by_index(begin = var_1071_begin_0, end = var_1071_end_0, end_mask = var_1071_end_mask_0, squeeze_mask = var_1071_squeeze_mask_0, x = reshape_24_cast_fp16)[name = tensor("op_1071_cast_fp16")]; + tensor var_1073_axis_0 = const()[name = tensor("op_1073_axis_0"), val = tensor(1)]; + tensor var_1073_mode_0 = const()[name = tensor("op_1073_mode_0"), val = tensor("update")]; + tensor var_1073_validate_indices_0 = const()[name = tensor("op_1073_validate_indices_0"), val = tensor(false)]; + tensor var_1073_cast_fp16 = scatter_along_axis(axis = var_1073_axis_0, data = var_1071_cast_fp16, indices = write_indices_5, mode = var_1073_mode_0, updates = v_5_cast_fp16, validate_indices = var_1073_validate_indices_0)[name = tensor("op_1073_cast_fp16")]; + tensor concat_18 = const()[name = tensor("concat_18"), val = tensor([1, 0, 0, 0, 0])]; + tensor concat_19 = const()[name = tensor("concat_19"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_5_internal_tensor_assign_2_stride_0 = const()[name = tensor("new_cache_5_internal_tensor_assign_2_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_5_internal_tensor_assign_2_begin_mask_0 = const()[name = tensor("new_cache_5_internal_tensor_assign_2_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_5_internal_tensor_assign_2_end_mask_0 = const()[name = tensor("new_cache_5_internal_tensor_assign_2_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_5_internal_tensor_assign_2_squeeze_mask_0 = const()[name = tensor("new_cache_5_internal_tensor_assign_2_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_17 = const()[name = tensor("shape_17"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_5 = const()[name = tensor("reduce_prod_5"), val = tensor(1048576)]; + tensor range_1d_5_start_0 = const()[name = tensor("range_1d_5_start_0"), val = tensor(0)]; + tensor range_1d_5_step_0 = const()[name = tensor("range_1d_5_step_0"), val = tensor(1)]; + tensor range_1d_5 = range_1d(end = reduce_prod_5, start = range_1d_5_start_0, step = range_1d_5_step_0)[name = tensor("range_1d_5")]; + tensor reshape_25 = reshape(shape = shape_17, x = range_1d_5)[name = tensor("reshape_25")]; + tensor slice_by_index_5 = slice_by_index(begin = concat_18, begin_mask = new_cache_5_internal_tensor_assign_2_begin_mask_0, end = concat_19, end_mask = new_cache_5_internal_tensor_assign_2_end_mask_0, squeeze_mask = new_cache_5_internal_tensor_assign_2_squeeze_mask_0, stride = new_cache_5_internal_tensor_assign_2_stride_0, x = reshape_25)[name = tensor("slice_by_index_5")]; + tensor reshape_26_shape_0 = const()[name = tensor("reshape_26_shape_0"), val = tensor([-1])]; + tensor reshape_26 = reshape(shape = reshape_26_shape_0, x = slice_by_index_5)[name = tensor("reshape_26")]; + tensor reshape_27_shape_0 = const()[name = tensor("reshape_27_shape_0"), val = tensor([-1])]; + tensor reshape_27_cast_fp16 = reshape(shape = reshape_27_shape_0, x = var_1073_cast_fp16)[name = tensor("reshape_27_cast_fp16")]; + tensor reshape_28_shape_0 = const()[name = tensor("reshape_28_shape_0"), val = tensor([-1])]; + tensor reshape_28_cast_fp16 = reshape(shape = reshape_28_shape_0, x = reshape_24_cast_fp16)[name = tensor("reshape_28_cast_fp16")]; + tensor scatter_5_mode_0 = const()[name = tensor("scatter_5_mode_0"), val = tensor("update")]; + tensor scatter_5_axis_0 = const()[name = tensor("scatter_5_axis_0"), val = tensor(0)]; + tensor scatter_5_validate_indices_0 = const()[name = tensor("scatter_5_validate_indices_0"), val = tensor(false)]; + tensor scatter_5_cast_fp16 = scatter(axis = scatter_5_axis_0, data = reshape_28_cast_fp16, indices = reshape_26, mode = scatter_5_mode_0, updates = reshape_27_cast_fp16, validate_indices = scatter_5_validate_indices_0)[name = tensor("scatter_5_cast_fp16")]; + tensor reshape_29_cast_fp16 = reshape(shape = shape_17, x = scatter_5_cast_fp16)[name = tensor("reshape_29_cast_fp16")]; + tensor reshape_29_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("reshape_29_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor keys_13_begin_0 = const()[name = tensor("keys_13_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor keys_13_end_0 = const()[name = tensor("keys_13_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor keys_13_end_mask_0 = const()[name = tensor("keys_13_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor keys_13_squeeze_mask_0 = const()[name = tensor("keys_13_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor keys_13_cast_fp16 = slice_by_index(begin = keys_13_begin_0, end = keys_13_end_0, end_mask = keys_13_end_mask_0, squeeze_mask = keys_13_squeeze_mask_0, x = reshape_29_cast_fp16)[name = tensor("keys_13_cast_fp16")]; + tensor values_13_begin_0 = const()[name = tensor("values_13_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor values_13_end_0 = const()[name = tensor("values_13_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor values_13_end_mask_0 = const()[name = tensor("values_13_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor values_13_squeeze_mask_0 = const()[name = tensor("values_13_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor values_13_cast_fp16 = slice_by_index(begin = values_13_begin_0, end = values_13_end_0, end_mask = values_13_end_mask_0, squeeze_mask = values_13_squeeze_mask_0, x = reshape_29_cast_fp16)[name = tensor("values_13_cast_fp16")]; + tensor var_1085_cast_fp16 = not_equal(x = keys_13_cast_fp16, y = keys_13_cast_fp16)[name = tensor("op_1085_cast_fp16")]; + tensor keys_15_cast_fp16 = select(a = var_319_to_fp16, b = keys_13_cast_fp16, cond = var_1085_cast_fp16)[name = tensor("keys_15_cast_fp16")]; + tensor var_1093_cast_fp16 = not_equal(x = values_13_cast_fp16, y = values_13_cast_fp16)[name = tensor("op_1093_cast_fp16")]; + tensor values_15_cast_fp16 = select(a = var_319_to_fp16, b = values_13_cast_fp16, cond = var_1093_cast_fp16)[name = tensor("values_15_cast_fp16")]; + tensor var_1117 = const()[name = tensor("op_1117"), val = tensor([0, 2, 1, 3])]; + tensor var_1133 = const()[name = tensor("op_1133"), val = tensor([1, 1, 1])]; + tensor var_1134_cast_fp16 = reshape(shape = var_1133, x = position2_to_fp16)[name = tensor("op_1134_cast_fp16")]; + tensor var_1154_to_fp16 = const()[name = tensor("op_1154_to_fp16"), val = tensor(0x1p+0)]; + tensor valid_len_5_cast_fp16 = add(x = var_1134_cast_fp16, y = var_1154_to_fp16)[name = tensor("valid_len_5_cast_fp16")]; + tensor valid_mask_5_cast_fp16 = less(x = k_positions_1_promoted_to_fp16, y = valid_len_5_cast_fp16)[name = tensor("valid_mask_5_cast_fp16")]; + tensor causal_mask_5_cast_fp16 = less_equal(x = k_positions_1_promoted_to_fp16, y = var_1134_cast_fp16)[name = tensor("causal_mask_5_cast_fp16")]; + tensor attn_mask_9 = logical_and(x = valid_mask_5_cast_fp16, y = causal_mask_5_cast_fp16)[name = tensor("attn_mask_9")]; + tensor attn_mask_11_axes_0 = const()[name = tensor("attn_mask_11_axes_0"), val = tensor([1])]; + tensor attn_mask_11 = expand_dims(axes = attn_mask_11_axes_0, x = attn_mask_9)[name = tensor("attn_mask_11")]; + tensor var_1172_transpose_x_0 = const()[name = tensor("op_1172_transpose_x_0"), val = tensor(false)]; + tensor var_1172_transpose_y_0 = const()[name = tensor("op_1172_transpose_y_0"), val = tensor(false)]; + tensor transpose_22_perm_0 = const()[name = tensor("transpose_22_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_23_perm_0 = const()[name = tensor("transpose_23_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_23 = transpose(perm = transpose_23_perm_0, x = keys_15_cast_fp16)[name = tensor("transpose_43")]; + tensor transpose_22 = transpose(perm = transpose_22_perm_0, x = q_15_cast_fp16)[name = tensor("transpose_44")]; + tensor var_1172_cast_fp16 = matmul(transpose_x = var_1172_transpose_x_0, transpose_y = var_1172_transpose_y_0, x = transpose_22, y = transpose_23)[name = tensor("op_1172_cast_fp16")]; + tensor var_1166_to_fp16 = const()[name = tensor("op_1166_to_fp16"), val = tensor([0x1p-3])]; + tensor attn_weights_13_cast_fp16 = mul(x = var_1172_cast_fp16, y = var_1166_to_fp16)[name = tensor("attn_weights_13_cast_fp16")]; + tensor var_1174 = logical_not(x = attn_mask_11)[name = tensor("op_1174")]; + tensor var_1175_to_fp16 = const()[name = tensor("op_1175_to_fp16"), val = tensor(-inf)]; + tensor attn_weights_15_cast_fp16 = select(a = var_1175_to_fp16, b = attn_weights_13_cast_fp16, cond = var_1174)[name = tensor("attn_weights_15_cast_fp16")]; + tensor var_1177 = const()[name = tensor("op_1177"), val = tensor(-1)]; + tensor attn_weights_17_cast_fp16 = softmax(axis = var_1177, x = attn_weights_15_cast_fp16)[name = tensor("attn_weights_17_cast_fp16")]; + tensor attn_output_5_transpose_x_0 = const()[name = tensor("attn_output_5_transpose_x_0"), val = tensor(false)]; + tensor attn_output_5_transpose_y_0 = const()[name = tensor("attn_output_5_transpose_y_0"), val = tensor(false)]; + tensor values_17_cast_fp16 = transpose(perm = var_1117, x = values_15_cast_fp16)[name = tensor("transpose_45")]; + tensor attn_output_5_cast_fp16 = matmul(transpose_x = attn_output_5_transpose_x_0, transpose_y = attn_output_5_transpose_y_0, x = attn_weights_17_cast_fp16, y = values_17_cast_fp16)[name = tensor("attn_output_5_cast_fp16")]; + tensor var_1185 = const()[name = tensor("op_1185"), val = tensor([0, 2, 1, 3])]; + tensor var_1190 = const()[name = tensor("op_1190"), val = tensor([1, 1, 1024])]; + tensor var_1186_cast_fp16 = transpose(perm = var_1185, x = attn_output_5_cast_fp16)[name = tensor("transpose_42")]; + tensor input_31_cast_fp16 = reshape(shape = var_1190, x = var_1186_cast_fp16)[name = tensor("input_31_cast_fp16")]; + tensor attn2_out_proj_weight_to_fp16 = const()[name = tensor("attn2_out_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(57777152)))]; + tensor linear_10_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = attn2_out_proj_weight_to_fp16, x = input_31_cast_fp16)[name = tensor("linear_10_cast_fp16")]; + tensor var_1196_to_fp16 = const()[name = tensor("op_1196_to_fp16"), val = tensor(0x1p+0)]; + tensor var_1197_cast_fp16 = add(x = position2_to_fp16, y = var_1196_to_fp16)[name = tensor("op_1197_cast_fp16")]; + tensor var_1197_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_1197_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor input_33_cast_fp16 = add(x = input_27_cast_fp16, y = linear_10_cast_fp16)[name = tensor("input_33_cast_fp16")]; + tensor input_35_axes_0 = const()[name = tensor("input_35_axes_0"), val = tensor([-1])]; + tensor norm2_2_weight_to_fp16 = const()[name = tensor("norm2_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(59874368)))]; + tensor norm2_2_bias_to_fp16 = const()[name = tensor("norm2_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(59876480)))]; + tensor var_1201_to_fp16 = const()[name = tensor("op_1201_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_35_cast_fp16 = layer_norm(axes = input_35_axes_0, beta = norm2_2_bias_to_fp16, epsilon = var_1201_to_fp16, gamma = norm2_2_weight_to_fp16, x = input_33_cast_fp16)[name = tensor("input_35_cast_fp16")]; + tensor linear2_1_weight_to_fp16 = const()[name = tensor("linear2_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(59878592)))]; + tensor linear_11_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = linear2_1_weight_to_fp16, x = input_35_cast_fp16)[name = tensor("linear_11_cast_fp16")]; + tensor input_37_mode_0 = const()[name = tensor("input_37_mode_0"), val = tensor("EXACT")]; + tensor input_37_cast_fp16 = gelu(mode = input_37_mode_0, x = linear_11_cast_fp16)[name = tensor("input_37_cast_fp16")]; + tensor linear2_2_weight_to_fp16 = const()[name = tensor("linear2_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(68267264)))]; + tensor linear_12_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = linear2_2_weight_to_fp16, x = input_37_cast_fp16)[name = tensor("linear_12_cast_fp16")]; + tensor input_39_cast_fp16 = add(x = input_33_cast_fp16, y = linear_12_cast_fp16)[name = tensor("input_39_cast_fp16")]; + tensor input_41_axes_0 = const()[name = tensor("input_41_axes_0"), val = tensor([-1])]; + tensor norm3_1_weight_to_fp16 = const()[name = tensor("norm3_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(76655936)))]; + tensor norm3_1_bias_to_fp16 = const()[name = tensor("norm3_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(76658048)))]; + tensor var_1218_to_fp16 = const()[name = tensor("op_1218_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_41_cast_fp16 = layer_norm(axes = input_41_axes_0, beta = norm3_1_bias_to_fp16, epsilon = var_1218_to_fp16, gamma = norm3_1_weight_to_fp16, x = input_39_cast_fp16)[name = tensor("input_41_cast_fp16")]; + tensor attn3_in_proj_weight_to_fp16 = const()[name = tensor("attn3_in_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(76660160)))]; + tensor linear_13_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = attn3_in_proj_weight_to_fp16, x = input_41_cast_fp16)[name = tensor("linear_13_cast_fp16")]; + tensor var_1232 = const()[name = tensor("op_1232"), val = tensor([1, 1, 3, 16, 64])]; + tensor qkv_7_cast_fp16 = reshape(shape = var_1232, x = linear_13_cast_fp16)[name = tensor("qkv_7_cast_fp16")]; + tensor q_19_begin_0 = const()[name = tensor("q_19_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor q_19_end_0 = const()[name = tensor("q_19_end_0"), val = tensor([1, 1, 1, 16, 64])]; + tensor q_19_end_mask_0 = const()[name = tensor("q_19_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor q_19_squeeze_mask_0 = const()[name = tensor("q_19_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor q_19_cast_fp16 = slice_by_index(begin = q_19_begin_0, end = q_19_end_0, end_mask = q_19_end_mask_0, squeeze_mask = q_19_squeeze_mask_0, x = qkv_7_cast_fp16)[name = tensor("q_19_cast_fp16")]; + tensor k_13_begin_0 = const()[name = tensor("k_13_begin_0"), val = tensor([0, 0, 1, 0, 0])]; + tensor k_13_end_0 = const()[name = tensor("k_13_end_0"), val = tensor([1, 1, 2, 16, 64])]; + tensor k_13_end_mask_0 = const()[name = tensor("k_13_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor k_13_squeeze_mask_0 = const()[name = tensor("k_13_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor k_13_cast_fp16 = slice_by_index(begin = k_13_begin_0, end = k_13_end_0, end_mask = k_13_end_mask_0, squeeze_mask = k_13_squeeze_mask_0, x = qkv_7_cast_fp16)[name = tensor("k_13_cast_fp16")]; + tensor v_7_begin_0 = const()[name = tensor("v_7_begin_0"), val = tensor([0, 0, 2, 0, 0])]; + tensor v_7_end_0 = const()[name = tensor("v_7_end_0"), val = tensor([1, 1, 3, 16, 64])]; + tensor v_7_end_mask_0 = const()[name = tensor("v_7_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor v_7_squeeze_mask_0 = const()[name = tensor("v_7_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor v_7_cast_fp16 = slice_by_index(begin = v_7_begin_0, end = v_7_end_0, end_mask = v_7_end_mask_0, squeeze_mask = v_7_squeeze_mask_0, x = qkv_7_cast_fp16)[name = tensor("v_7_cast_fp16")]; + tensor var_1298 = const()[name = tensor("op_1298"), val = tensor([1, 1, 1, 1])]; + tensor position3_to_fp16_dtype_0 = const()[name = tensor("position3_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor position3_to_fp16 = cast(dtype = position3_to_fp16_dtype_0, x = position3)[name = tensor("cast_124")]; + tensor ts_23_cast_fp16 = reshape(shape = var_1298, x = position3_to_fp16)[name = tensor("ts_23_cast_fp16")]; + tensor var_1305 = const()[name = tensor("op_1305"), val = tensor([1, 1, 16, 32, 2])]; + tensor q_complex_7_cast_fp16 = reshape(shape = var_1305, x = q_19_cast_fp16)[name = tensor("q_complex_7_cast_fp16")]; + tensor var_1312 = const()[name = tensor("op_1312"), val = tensor([1, 1, 16, 32, 2])]; + tensor k_complex_7_cast_fp16 = reshape(shape = var_1312, x = k_13_cast_fp16)[name = tensor("k_complex_7_cast_fp16")]; + tensor var_1316_begin_0 = const()[name = tensor("op_1316_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_1316_end_0 = const()[name = tensor("op_1316_end_0"), val = tensor([1, 1, 16, 32, 1])]; + tensor var_1316_end_mask_0 = const()[name = tensor("op_1316_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_1316_squeeze_mask_0 = const()[name = tensor("op_1316_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_1316_cast_fp16 = slice_by_index(begin = var_1316_begin_0, end = var_1316_end_0, end_mask = var_1316_end_mask_0, squeeze_mask = var_1316_squeeze_mask_0, x = q_complex_7_cast_fp16)[name = tensor("op_1316_cast_fp16")]; + tensor var_1324_begin_0 = const()[name = tensor("op_1324_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_1324_end_0 = const()[name = tensor("op_1324_end_0"), val = tensor([1, 1, 16, 32, 2])]; + tensor var_1324_end_mask_0 = const()[name = tensor("op_1324_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_1324_squeeze_mask_0 = const()[name = tensor("op_1324_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_1324_cast_fp16 = slice_by_index(begin = var_1324_begin_0, end = var_1324_end_0, end_mask = var_1324_end_mask_0, squeeze_mask = var_1324_squeeze_mask_0, x = q_complex_7_cast_fp16)[name = tensor("op_1324_cast_fp16")]; + tensor var_1332_begin_0 = const()[name = tensor("op_1332_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_1332_end_0 = const()[name = tensor("op_1332_end_0"), val = tensor([1, 1, 16, 32, 1])]; + tensor var_1332_end_mask_0 = const()[name = tensor("op_1332_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_1332_squeeze_mask_0 = const()[name = tensor("op_1332_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_1332_cast_fp16 = slice_by_index(begin = var_1332_begin_0, end = var_1332_end_0, end_mask = var_1332_end_mask_0, squeeze_mask = var_1332_squeeze_mask_0, x = k_complex_7_cast_fp16)[name = tensor("op_1332_cast_fp16")]; + tensor var_1340_begin_0 = const()[name = tensor("op_1340_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_1340_end_0 = const()[name = tensor("op_1340_end_0"), val = tensor([1, 1, 16, 32, 2])]; + tensor var_1340_end_mask_0 = const()[name = tensor("op_1340_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_1340_squeeze_mask_0 = const()[name = tensor("op_1340_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_1340_cast_fp16 = slice_by_index(begin = var_1340_begin_0, end = var_1340_end_0, end_mask = var_1340_end_mask_0, squeeze_mask = var_1340_squeeze_mask_0, x = k_complex_7_cast_fp16)[name = tensor("op_1340_cast_fp16")]; + tensor freqs_7_to_fp16 = const()[name = tensor("freqs_7_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(82951680)))]; + tensor var_1346_cast_fp16 = mul(x = freqs_7_to_fp16, y = ts_23_cast_fp16)[name = tensor("op_1346_cast_fp16")]; + tensor rotr_7_cast_fp16 = cos(x = var_1346_cast_fp16)[name = tensor("rotr_7_cast_fp16")]; + tensor roti_7_cast_fp16 = sin(x = var_1346_cast_fp16)[name = tensor("roti_7_cast_fp16")]; + tensor var_1350_cast_fp16 = mul(x = var_1316_cast_fp16, y = rotr_7_cast_fp16)[name = tensor("op_1350_cast_fp16")]; + tensor var_1351_cast_fp16 = mul(x = var_1324_cast_fp16, y = roti_7_cast_fp16)[name = tensor("op_1351_cast_fp16")]; + tensor qor_13_cast_fp16 = sub(x = var_1350_cast_fp16, y = var_1351_cast_fp16)[name = tensor("qor_13_cast_fp16")]; + tensor var_1354_cast_fp16 = mul(x = var_1316_cast_fp16, y = roti_7_cast_fp16)[name = tensor("op_1354_cast_fp16")]; + tensor var_1355_cast_fp16 = mul(x = var_1324_cast_fp16, y = rotr_7_cast_fp16)[name = tensor("op_1355_cast_fp16")]; + tensor qoi_13_cast_fp16 = add(x = var_1354_cast_fp16, y = var_1355_cast_fp16)[name = tensor("qoi_13_cast_fp16")]; + tensor var_1358_cast_fp16 = mul(x = var_1332_cast_fp16, y = rotr_7_cast_fp16)[name = tensor("op_1358_cast_fp16")]; + tensor var_1359_cast_fp16 = mul(x = var_1340_cast_fp16, y = roti_7_cast_fp16)[name = tensor("op_1359_cast_fp16")]; + tensor kor_13_cast_fp16 = sub(x = var_1358_cast_fp16, y = var_1359_cast_fp16)[name = tensor("kor_13_cast_fp16")]; + tensor var_1362_cast_fp16 = mul(x = var_1332_cast_fp16, y = roti_7_cast_fp16)[name = tensor("op_1362_cast_fp16")]; + tensor var_1363_cast_fp16 = mul(x = var_1340_cast_fp16, y = rotr_7_cast_fp16)[name = tensor("op_1363_cast_fp16")]; + tensor koi_13_cast_fp16 = add(x = var_1362_cast_fp16, y = var_1363_cast_fp16)[name = tensor("koi_13_cast_fp16")]; + tensor qo_7_axis_0 = const()[name = tensor("qo_7_axis_0"), val = tensor(-1)]; + tensor qo_7_cast_fp16 = stack(axis = qo_7_axis_0, values = (qor_13_cast_fp16, qoi_13_cast_fp16))[name = tensor("qo_7_cast_fp16")]; + tensor ko_7_axis_0 = const()[name = tensor("ko_7_axis_0"), val = tensor(-1)]; + tensor ko_7_cast_fp16 = stack(axis = ko_7_axis_0, values = (kor_13_cast_fp16, koi_13_cast_fp16))[name = tensor("ko_7_cast_fp16")]; + tensor var_1396 = const()[name = tensor("op_1396"), val = tensor([1, 1, 16, 64])]; + tensor q_21_cast_fp16 = reshape(shape = var_1396, x = qo_7_cast_fp16)[name = tensor("q_21_cast_fp16")]; + tensor var_1402 = const()[name = tensor("op_1402"), val = tensor([1, 1, 16, 64])]; + tensor k_15_cast_fp16 = reshape(shape = var_1402, x = ko_7_cast_fp16)[name = tensor("k_15_cast_fp16")]; + tensor _inversed_1427_y_0_to_fp16 = const()[name = tensor("_inversed_1427_y_0_to_fp16"), val = tensor(0x1p-9)]; + tensor _inversed_1427_cast_fp16 = mul(x = ts_23_cast_fp16, y = _inversed_1427_y_0_to_fp16)[name = tensor("_inversed_1427_cast_fp16")]; + tensor var_1428_cast_fp16 = floor(x = _inversed_1427_cast_fp16)[name = tensor("op_1428_cast_fp16")]; + tensor var_1429_to_fp16 = const()[name = tensor("op_1429_to_fp16"), val = tensor(0x1p+9)]; + tensor var_1430_cast_fp16 = mul(x = var_1428_cast_fp16, y = var_1429_to_fp16)[name = tensor("op_1430_cast_fp16")]; + tensor write_indices_float_15_cast_fp16 = sub(x = ts_23_cast_fp16, y = var_1430_cast_fp16)[name = tensor("write_indices_float_15_cast_fp16")]; + tensor var_1437_dtype_0 = const()[name = tensor("op_1437_dtype_0"), val = tensor("int32")]; + tensor write_indices_7_reps_0 = const()[name = tensor("write_indices_7_reps_0"), val = tensor([1, 1, 16, 64])]; + tensor write_indices_float_15_cast_fp16_to_int32 = cast(dtype = var_1437_dtype_0, x = write_indices_float_15_cast_fp16)[name = tensor("cast_123")]; + tensor write_indices_7 = tile(reps = write_indices_7_reps_0, x = write_indices_float_15_cast_fp16_to_int32)[name = tensor("write_indices_7")]; + tensor var_1447_begin_0 = const()[name = tensor("op_1447_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_1447_end_0 = const()[name = tensor("op_1447_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor var_1447_end_mask_0 = const()[name = tensor("op_1447_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_1447_squeeze_mask_0 = const()[name = tensor("op_1447_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor cache3_to_fp16_dtype_0 = const()[name = tensor("cache3_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor cache3_to_fp16 = cast(dtype = cache3_to_fp16_dtype_0, x = cache3)[name = tensor("cast_122")]; + tensor var_1447_cast_fp16 = slice_by_index(begin = var_1447_begin_0, end = var_1447_end_0, end_mask = var_1447_end_mask_0, squeeze_mask = var_1447_squeeze_mask_0, x = cache3_to_fp16)[name = tensor("op_1447_cast_fp16")]; + tensor var_1449_axis_0 = const()[name = tensor("op_1449_axis_0"), val = tensor(1)]; + tensor var_1449_mode_0 = const()[name = tensor("op_1449_mode_0"), val = tensor("update")]; + tensor var_1449_validate_indices_0 = const()[name = tensor("op_1449_validate_indices_0"), val = tensor(false)]; + tensor var_1449_cast_fp16 = scatter_along_axis(axis = var_1449_axis_0, data = var_1447_cast_fp16, indices = write_indices_7, mode = var_1449_mode_0, updates = k_15_cast_fp16, validate_indices = var_1449_validate_indices_0)[name = tensor("op_1449_cast_fp16")]; + tensor concat_23 = const()[name = tensor("concat_23"), val = tensor([0, 0, 0, 0, 0])]; + tensor concat_24 = const()[name = tensor("concat_24"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_7_internal_tensor_assign_1_stride_0 = const()[name = tensor("new_cache_7_internal_tensor_assign_1_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_7_internal_tensor_assign_1_begin_mask_0 = const()[name = tensor("new_cache_7_internal_tensor_assign_1_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_7_internal_tensor_assign_1_end_mask_0 = const()[name = tensor("new_cache_7_internal_tensor_assign_1_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_7_internal_tensor_assign_1_squeeze_mask_0 = const()[name = tensor("new_cache_7_internal_tensor_assign_1_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_18 = const()[name = tensor("shape_18"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_6 = const()[name = tensor("reduce_prod_6"), val = tensor(1048576)]; + tensor range_1d_6_start_0 = const()[name = tensor("range_1d_6_start_0"), val = tensor(0)]; + tensor range_1d_6_step_0 = const()[name = tensor("range_1d_6_step_0"), val = tensor(1)]; + tensor range_1d_6 = range_1d(end = reduce_prod_6, start = range_1d_6_start_0, step = range_1d_6_step_0)[name = tensor("range_1d_6")]; + tensor reshape_30 = reshape(shape = shape_18, x = range_1d_6)[name = tensor("reshape_30")]; + tensor slice_by_index_6 = slice_by_index(begin = concat_23, begin_mask = new_cache_7_internal_tensor_assign_1_begin_mask_0, end = concat_24, end_mask = new_cache_7_internal_tensor_assign_1_end_mask_0, squeeze_mask = new_cache_7_internal_tensor_assign_1_squeeze_mask_0, stride = new_cache_7_internal_tensor_assign_1_stride_0, x = reshape_30)[name = tensor("slice_by_index_6")]; + tensor reshape_31_shape_0 = const()[name = tensor("reshape_31_shape_0"), val = tensor([-1])]; + tensor reshape_31 = reshape(shape = reshape_31_shape_0, x = slice_by_index_6)[name = tensor("reshape_31")]; + tensor reshape_32_shape_0 = const()[name = tensor("reshape_32_shape_0"), val = tensor([-1])]; + tensor reshape_32_cast_fp16 = reshape(shape = reshape_32_shape_0, x = var_1449_cast_fp16)[name = tensor("reshape_32_cast_fp16")]; + tensor reshape_33_shape_0 = const()[name = tensor("reshape_33_shape_0"), val = tensor([-1])]; + tensor reshape_33_cast_fp16 = reshape(shape = reshape_33_shape_0, x = cache3_to_fp16)[name = tensor("reshape_33_cast_fp16")]; + tensor scatter_6_mode_0 = const()[name = tensor("scatter_6_mode_0"), val = tensor("update")]; + tensor scatter_6_axis_0 = const()[name = tensor("scatter_6_axis_0"), val = tensor(0)]; + tensor scatter_6_validate_indices_0 = const()[name = tensor("scatter_6_validate_indices_0"), val = tensor(false)]; + tensor scatter_6_cast_fp16 = scatter(axis = scatter_6_axis_0, data = reshape_33_cast_fp16, indices = reshape_31, mode = scatter_6_mode_0, updates = reshape_32_cast_fp16, validate_indices = scatter_6_validate_indices_0)[name = tensor("scatter_6_cast_fp16")]; + tensor reshape_34_cast_fp16 = reshape(shape = shape_18, x = scatter_6_cast_fp16)[name = tensor("reshape_34_cast_fp16")]; + tensor var_1457_begin_0 = const()[name = tensor("op_1457_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor var_1457_end_0 = const()[name = tensor("op_1457_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor var_1457_end_mask_0 = const()[name = tensor("op_1457_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_1457_squeeze_mask_0 = const()[name = tensor("op_1457_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor var_1457_cast_fp16 = slice_by_index(begin = var_1457_begin_0, end = var_1457_end_0, end_mask = var_1457_end_mask_0, squeeze_mask = var_1457_squeeze_mask_0, x = reshape_34_cast_fp16)[name = tensor("op_1457_cast_fp16")]; + tensor var_1459_axis_0 = const()[name = tensor("op_1459_axis_0"), val = tensor(1)]; + tensor var_1459_mode_0 = const()[name = tensor("op_1459_mode_0"), val = tensor("update")]; + tensor var_1459_validate_indices_0 = const()[name = tensor("op_1459_validate_indices_0"), val = tensor(false)]; + tensor var_1459_cast_fp16 = scatter_along_axis(axis = var_1459_axis_0, data = var_1457_cast_fp16, indices = write_indices_7, mode = var_1459_mode_0, updates = v_7_cast_fp16, validate_indices = var_1459_validate_indices_0)[name = tensor("op_1459_cast_fp16")]; + tensor concat_25 = const()[name = tensor("concat_25"), val = tensor([1, 0, 0, 0, 0])]; + tensor concat_26 = const()[name = tensor("concat_26"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_7_internal_tensor_assign_2_stride_0 = const()[name = tensor("new_cache_7_internal_tensor_assign_2_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_7_internal_tensor_assign_2_begin_mask_0 = const()[name = tensor("new_cache_7_internal_tensor_assign_2_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_7_internal_tensor_assign_2_end_mask_0 = const()[name = tensor("new_cache_7_internal_tensor_assign_2_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_7_internal_tensor_assign_2_squeeze_mask_0 = const()[name = tensor("new_cache_7_internal_tensor_assign_2_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_19 = const()[name = tensor("shape_19"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_7 = const()[name = tensor("reduce_prod_7"), val = tensor(1048576)]; + tensor range_1d_7_start_0 = const()[name = tensor("range_1d_7_start_0"), val = tensor(0)]; + tensor range_1d_7_step_0 = const()[name = tensor("range_1d_7_step_0"), val = tensor(1)]; + tensor range_1d_7 = range_1d(end = reduce_prod_7, start = range_1d_7_start_0, step = range_1d_7_step_0)[name = tensor("range_1d_7")]; + tensor reshape_35 = reshape(shape = shape_19, x = range_1d_7)[name = tensor("reshape_35")]; + tensor slice_by_index_7 = slice_by_index(begin = concat_25, begin_mask = new_cache_7_internal_tensor_assign_2_begin_mask_0, end = concat_26, end_mask = new_cache_7_internal_tensor_assign_2_end_mask_0, squeeze_mask = new_cache_7_internal_tensor_assign_2_squeeze_mask_0, stride = new_cache_7_internal_tensor_assign_2_stride_0, x = reshape_35)[name = tensor("slice_by_index_7")]; + tensor reshape_36_shape_0 = const()[name = tensor("reshape_36_shape_0"), val = tensor([-1])]; + tensor reshape_36 = reshape(shape = reshape_36_shape_0, x = slice_by_index_7)[name = tensor("reshape_36")]; + tensor reshape_37_shape_0 = const()[name = tensor("reshape_37_shape_0"), val = tensor([-1])]; + tensor reshape_37_cast_fp16 = reshape(shape = reshape_37_shape_0, x = var_1459_cast_fp16)[name = tensor("reshape_37_cast_fp16")]; + tensor reshape_38_shape_0 = const()[name = tensor("reshape_38_shape_0"), val = tensor([-1])]; + tensor reshape_38_cast_fp16 = reshape(shape = reshape_38_shape_0, x = reshape_34_cast_fp16)[name = tensor("reshape_38_cast_fp16")]; + tensor scatter_7_mode_0 = const()[name = tensor("scatter_7_mode_0"), val = tensor("update")]; + tensor scatter_7_axis_0 = const()[name = tensor("scatter_7_axis_0"), val = tensor(0)]; + tensor scatter_7_validate_indices_0 = const()[name = tensor("scatter_7_validate_indices_0"), val = tensor(false)]; + tensor scatter_7_cast_fp16 = scatter(axis = scatter_7_axis_0, data = reshape_38_cast_fp16, indices = reshape_36, mode = scatter_7_mode_0, updates = reshape_37_cast_fp16, validate_indices = scatter_7_validate_indices_0)[name = tensor("scatter_7_cast_fp16")]; + tensor reshape_39_cast_fp16 = reshape(shape = shape_19, x = scatter_7_cast_fp16)[name = tensor("reshape_39_cast_fp16")]; + tensor reshape_39_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("reshape_39_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor keys_19_begin_0 = const()[name = tensor("keys_19_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor keys_19_end_0 = const()[name = tensor("keys_19_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor keys_19_end_mask_0 = const()[name = tensor("keys_19_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor keys_19_squeeze_mask_0 = const()[name = tensor("keys_19_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor keys_19_cast_fp16 = slice_by_index(begin = keys_19_begin_0, end = keys_19_end_0, end_mask = keys_19_end_mask_0, squeeze_mask = keys_19_squeeze_mask_0, x = reshape_39_cast_fp16)[name = tensor("keys_19_cast_fp16")]; + tensor values_19_begin_0 = const()[name = tensor("values_19_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor values_19_end_0 = const()[name = tensor("values_19_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor values_19_end_mask_0 = const()[name = tensor("values_19_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor values_19_squeeze_mask_0 = const()[name = tensor("values_19_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor values_19_cast_fp16 = slice_by_index(begin = values_19_begin_0, end = values_19_end_0, end_mask = values_19_end_mask_0, squeeze_mask = values_19_squeeze_mask_0, x = reshape_39_cast_fp16)[name = tensor("values_19_cast_fp16")]; + tensor var_1471_cast_fp16 = not_equal(x = keys_19_cast_fp16, y = keys_19_cast_fp16)[name = tensor("op_1471_cast_fp16")]; + tensor keys_21_cast_fp16 = select(a = var_319_to_fp16, b = keys_19_cast_fp16, cond = var_1471_cast_fp16)[name = tensor("keys_21_cast_fp16")]; + tensor var_1479_cast_fp16 = not_equal(x = values_19_cast_fp16, y = values_19_cast_fp16)[name = tensor("op_1479_cast_fp16")]; + tensor values_21_cast_fp16 = select(a = var_319_to_fp16, b = values_19_cast_fp16, cond = var_1479_cast_fp16)[name = tensor("values_21_cast_fp16")]; + tensor var_1503 = const()[name = tensor("op_1503"), val = tensor([0, 2, 1, 3])]; + tensor var_1519 = const()[name = tensor("op_1519"), val = tensor([1, 1, 1])]; + tensor var_1520_cast_fp16 = reshape(shape = var_1519, x = position3_to_fp16)[name = tensor("op_1520_cast_fp16")]; + tensor var_1540_to_fp16 = const()[name = tensor("op_1540_to_fp16"), val = tensor(0x1p+0)]; + tensor valid_len_7_cast_fp16 = add(x = var_1520_cast_fp16, y = var_1540_to_fp16)[name = tensor("valid_len_7_cast_fp16")]; + tensor valid_mask_7_cast_fp16 = less(x = k_positions_1_promoted_to_fp16, y = valid_len_7_cast_fp16)[name = tensor("valid_mask_7_cast_fp16")]; + tensor causal_mask_7_cast_fp16 = less_equal(x = k_positions_1_promoted_to_fp16, y = var_1520_cast_fp16)[name = tensor("causal_mask_7_cast_fp16")]; + tensor attn_mask_13 = logical_and(x = valid_mask_7_cast_fp16, y = causal_mask_7_cast_fp16)[name = tensor("attn_mask_13")]; + tensor attn_mask_15_axes_0 = const()[name = tensor("attn_mask_15_axes_0"), val = tensor([1])]; + tensor attn_mask_15 = expand_dims(axes = attn_mask_15_axes_0, x = attn_mask_13)[name = tensor("attn_mask_15")]; + tensor var_1558_transpose_x_0 = const()[name = tensor("op_1558_transpose_x_0"), val = tensor(false)]; + tensor var_1558_transpose_y_0 = const()[name = tensor("op_1558_transpose_y_0"), val = tensor(false)]; + tensor transpose_24_perm_0 = const()[name = tensor("transpose_24_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_25_perm_0 = const()[name = tensor("transpose_25_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_25 = transpose(perm = transpose_25_perm_0, x = keys_21_cast_fp16)[name = tensor("transpose_39")]; + tensor transpose_24 = transpose(perm = transpose_24_perm_0, x = q_21_cast_fp16)[name = tensor("transpose_40")]; + tensor var_1558_cast_fp16 = matmul(transpose_x = var_1558_transpose_x_0, transpose_y = var_1558_transpose_y_0, x = transpose_24, y = transpose_25)[name = tensor("op_1558_cast_fp16")]; + tensor var_1552_to_fp16 = const()[name = tensor("op_1552_to_fp16"), val = tensor([0x1p-3])]; + tensor attn_weights_19_cast_fp16 = mul(x = var_1558_cast_fp16, y = var_1552_to_fp16)[name = tensor("attn_weights_19_cast_fp16")]; + tensor var_1560 = logical_not(x = attn_mask_15)[name = tensor("op_1560")]; + tensor var_1561_to_fp16 = const()[name = tensor("op_1561_to_fp16"), val = tensor(-inf)]; + tensor attn_weights_21_cast_fp16 = select(a = var_1561_to_fp16, b = attn_weights_19_cast_fp16, cond = var_1560)[name = tensor("attn_weights_21_cast_fp16")]; + tensor var_1563 = const()[name = tensor("op_1563"), val = tensor(-1)]; + tensor attn_weights_23_cast_fp16 = softmax(axis = var_1563, x = attn_weights_21_cast_fp16)[name = tensor("attn_weights_23_cast_fp16")]; + tensor attn_output_7_transpose_x_0 = const()[name = tensor("attn_output_7_transpose_x_0"), val = tensor(false)]; + tensor attn_output_7_transpose_y_0 = const()[name = tensor("attn_output_7_transpose_y_0"), val = tensor(false)]; + tensor values_23_cast_fp16 = transpose(perm = var_1503, x = values_21_cast_fp16)[name = tensor("transpose_41")]; + tensor attn_output_7_cast_fp16 = matmul(transpose_x = attn_output_7_transpose_x_0, transpose_y = attn_output_7_transpose_y_0, x = attn_weights_23_cast_fp16, y = values_23_cast_fp16)[name = tensor("attn_output_7_cast_fp16")]; + tensor var_1571 = const()[name = tensor("op_1571"), val = tensor([0, 2, 1, 3])]; + tensor var_1576 = const()[name = tensor("op_1576"), val = tensor([1, 1, 1024])]; + tensor var_1572_cast_fp16 = transpose(perm = var_1571, x = attn_output_7_cast_fp16)[name = tensor("transpose_38")]; + tensor input_43_cast_fp16 = reshape(shape = var_1576, x = var_1572_cast_fp16)[name = tensor("input_43_cast_fp16")]; + tensor attn3_out_proj_weight_to_fp16 = const()[name = tensor("attn3_out_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(82951808)))]; + tensor linear_14_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = attn3_out_proj_weight_to_fp16, x = input_43_cast_fp16)[name = tensor("linear_14_cast_fp16")]; + tensor var_1582_to_fp16 = const()[name = tensor("op_1582_to_fp16"), val = tensor(0x1p+0)]; + tensor var_1583_cast_fp16 = add(x = position3_to_fp16, y = var_1582_to_fp16)[name = tensor("op_1583_cast_fp16")]; + tensor var_1583_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_1583_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor input_45_cast_fp16 = add(x = input_39_cast_fp16, y = linear_14_cast_fp16)[name = tensor("input_45_cast_fp16")]; + tensor input_47_axes_0 = const()[name = tensor("input_47_axes_0"), val = tensor([-1])]; + tensor norm3_2_weight_to_fp16 = const()[name = tensor("norm3_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(85049024)))]; + tensor norm3_2_bias_to_fp16 = const()[name = tensor("norm3_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(85051136)))]; + tensor var_1587_to_fp16 = const()[name = tensor("op_1587_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_47_cast_fp16 = layer_norm(axes = input_47_axes_0, beta = norm3_2_bias_to_fp16, epsilon = var_1587_to_fp16, gamma = norm3_2_weight_to_fp16, x = input_45_cast_fp16)[name = tensor("input_47_cast_fp16")]; + tensor linear3_1_weight_to_fp16 = const()[name = tensor("linear3_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(85053248)))]; + tensor linear_15_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = linear3_1_weight_to_fp16, x = input_47_cast_fp16)[name = tensor("linear_15_cast_fp16")]; + tensor input_49_mode_0 = const()[name = tensor("input_49_mode_0"), val = tensor("EXACT")]; + tensor input_49_cast_fp16 = gelu(mode = input_49_mode_0, x = linear_15_cast_fp16)[name = tensor("input_49_cast_fp16")]; + tensor linear3_2_weight_to_fp16 = const()[name = tensor("linear3_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(93441920)))]; + tensor linear_16_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = linear3_2_weight_to_fp16, x = input_49_cast_fp16)[name = tensor("linear_16_cast_fp16")]; + tensor input_51_cast_fp16 = add(x = input_45_cast_fp16, y = linear_16_cast_fp16)[name = tensor("input_51_cast_fp16")]; + tensor input_53_axes_0 = const()[name = tensor("input_53_axes_0"), val = tensor([-1])]; + tensor norm4_1_weight_to_fp16 = const()[name = tensor("norm4_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(101830592)))]; + tensor norm4_1_bias_to_fp16 = const()[name = tensor("norm4_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(101832704)))]; + tensor var_1604_to_fp16 = const()[name = tensor("op_1604_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_53_cast_fp16 = layer_norm(axes = input_53_axes_0, beta = norm4_1_bias_to_fp16, epsilon = var_1604_to_fp16, gamma = norm4_1_weight_to_fp16, x = input_51_cast_fp16)[name = tensor("input_53_cast_fp16")]; + tensor attn4_in_proj_weight_to_fp16 = const()[name = tensor("attn4_in_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(101834816)))]; + tensor linear_17_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = attn4_in_proj_weight_to_fp16, x = input_53_cast_fp16)[name = tensor("linear_17_cast_fp16")]; + tensor var_1618 = const()[name = tensor("op_1618"), val = tensor([1, 1, 3, 16, 64])]; + tensor qkv_9_cast_fp16 = reshape(shape = var_1618, x = linear_17_cast_fp16)[name = tensor("qkv_9_cast_fp16")]; + tensor q_25_begin_0 = const()[name = tensor("q_25_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor q_25_end_0 = const()[name = tensor("q_25_end_0"), val = tensor([1, 1, 1, 16, 64])]; + tensor q_25_end_mask_0 = const()[name = tensor("q_25_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor q_25_squeeze_mask_0 = const()[name = tensor("q_25_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor q_25_cast_fp16 = slice_by_index(begin = q_25_begin_0, end = q_25_end_0, end_mask = q_25_end_mask_0, squeeze_mask = q_25_squeeze_mask_0, x = qkv_9_cast_fp16)[name = tensor("q_25_cast_fp16")]; + tensor k_17_begin_0 = const()[name = tensor("k_17_begin_0"), val = tensor([0, 0, 1, 0, 0])]; + tensor k_17_end_0 = const()[name = tensor("k_17_end_0"), val = tensor([1, 1, 2, 16, 64])]; + tensor k_17_end_mask_0 = const()[name = tensor("k_17_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor k_17_squeeze_mask_0 = const()[name = tensor("k_17_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor k_17_cast_fp16 = slice_by_index(begin = k_17_begin_0, end = k_17_end_0, end_mask = k_17_end_mask_0, squeeze_mask = k_17_squeeze_mask_0, x = qkv_9_cast_fp16)[name = tensor("k_17_cast_fp16")]; + tensor v_9_begin_0 = const()[name = tensor("v_9_begin_0"), val = tensor([0, 0, 2, 0, 0])]; + tensor v_9_end_0 = const()[name = tensor("v_9_end_0"), val = tensor([1, 1, 3, 16, 64])]; + tensor v_9_end_mask_0 = const()[name = tensor("v_9_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor v_9_squeeze_mask_0 = const()[name = tensor("v_9_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor v_9_cast_fp16 = slice_by_index(begin = v_9_begin_0, end = v_9_end_0, end_mask = v_9_end_mask_0, squeeze_mask = v_9_squeeze_mask_0, x = qkv_9_cast_fp16)[name = tensor("v_9_cast_fp16")]; + tensor var_1684 = const()[name = tensor("op_1684"), val = tensor([1, 1, 1, 1])]; + tensor position4_to_fp16_dtype_0 = const()[name = tensor("position4_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor position4_to_fp16 = cast(dtype = position4_to_fp16_dtype_0, x = position4)[name = tensor("cast_119")]; + tensor ts_29_cast_fp16 = reshape(shape = var_1684, x = position4_to_fp16)[name = tensor("ts_29_cast_fp16")]; + tensor var_1691 = const()[name = tensor("op_1691"), val = tensor([1, 1, 16, 32, 2])]; + tensor q_complex_9_cast_fp16 = reshape(shape = var_1691, x = q_25_cast_fp16)[name = tensor("q_complex_9_cast_fp16")]; + tensor var_1698 = const()[name = tensor("op_1698"), val = tensor([1, 1, 16, 32, 2])]; + tensor k_complex_9_cast_fp16 = reshape(shape = var_1698, x = k_17_cast_fp16)[name = tensor("k_complex_9_cast_fp16")]; + tensor var_1702_begin_0 = const()[name = tensor("op_1702_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_1702_end_0 = const()[name = tensor("op_1702_end_0"), val = tensor([1, 1, 16, 32, 1])]; + tensor var_1702_end_mask_0 = const()[name = tensor("op_1702_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_1702_squeeze_mask_0 = const()[name = tensor("op_1702_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_1702_cast_fp16 = slice_by_index(begin = var_1702_begin_0, end = var_1702_end_0, end_mask = var_1702_end_mask_0, squeeze_mask = var_1702_squeeze_mask_0, x = q_complex_9_cast_fp16)[name = tensor("op_1702_cast_fp16")]; + tensor var_1710_begin_0 = const()[name = tensor("op_1710_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_1710_end_0 = const()[name = tensor("op_1710_end_0"), val = tensor([1, 1, 16, 32, 2])]; + tensor var_1710_end_mask_0 = const()[name = tensor("op_1710_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_1710_squeeze_mask_0 = const()[name = tensor("op_1710_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_1710_cast_fp16 = slice_by_index(begin = var_1710_begin_0, end = var_1710_end_0, end_mask = var_1710_end_mask_0, squeeze_mask = var_1710_squeeze_mask_0, x = q_complex_9_cast_fp16)[name = tensor("op_1710_cast_fp16")]; + tensor var_1718_begin_0 = const()[name = tensor("op_1718_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_1718_end_0 = const()[name = tensor("op_1718_end_0"), val = tensor([1, 1, 16, 32, 1])]; + tensor var_1718_end_mask_0 = const()[name = tensor("op_1718_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_1718_squeeze_mask_0 = const()[name = tensor("op_1718_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_1718_cast_fp16 = slice_by_index(begin = var_1718_begin_0, end = var_1718_end_0, end_mask = var_1718_end_mask_0, squeeze_mask = var_1718_squeeze_mask_0, x = k_complex_9_cast_fp16)[name = tensor("op_1718_cast_fp16")]; + tensor var_1726_begin_0 = const()[name = tensor("op_1726_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_1726_end_0 = const()[name = tensor("op_1726_end_0"), val = tensor([1, 1, 16, 32, 2])]; + tensor var_1726_end_mask_0 = const()[name = tensor("op_1726_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_1726_squeeze_mask_0 = const()[name = tensor("op_1726_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_1726_cast_fp16 = slice_by_index(begin = var_1726_begin_0, end = var_1726_end_0, end_mask = var_1726_end_mask_0, squeeze_mask = var_1726_squeeze_mask_0, x = k_complex_9_cast_fp16)[name = tensor("op_1726_cast_fp16")]; + tensor freqs_9_to_fp16 = const()[name = tensor("freqs_9_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(108126336)))]; + tensor var_1732_cast_fp16 = mul(x = freqs_9_to_fp16, y = ts_29_cast_fp16)[name = tensor("op_1732_cast_fp16")]; + tensor rotr_9_cast_fp16 = cos(x = var_1732_cast_fp16)[name = tensor("rotr_9_cast_fp16")]; + tensor roti_9_cast_fp16 = sin(x = var_1732_cast_fp16)[name = tensor("roti_9_cast_fp16")]; + tensor var_1736_cast_fp16 = mul(x = var_1702_cast_fp16, y = rotr_9_cast_fp16)[name = tensor("op_1736_cast_fp16")]; + tensor var_1737_cast_fp16 = mul(x = var_1710_cast_fp16, y = roti_9_cast_fp16)[name = tensor("op_1737_cast_fp16")]; + tensor qor_17_cast_fp16 = sub(x = var_1736_cast_fp16, y = var_1737_cast_fp16)[name = tensor("qor_17_cast_fp16")]; + tensor var_1740_cast_fp16 = mul(x = var_1702_cast_fp16, y = roti_9_cast_fp16)[name = tensor("op_1740_cast_fp16")]; + tensor var_1741_cast_fp16 = mul(x = var_1710_cast_fp16, y = rotr_9_cast_fp16)[name = tensor("op_1741_cast_fp16")]; + tensor qoi_17_cast_fp16 = add(x = var_1740_cast_fp16, y = var_1741_cast_fp16)[name = tensor("qoi_17_cast_fp16")]; + tensor var_1744_cast_fp16 = mul(x = var_1718_cast_fp16, y = rotr_9_cast_fp16)[name = tensor("op_1744_cast_fp16")]; + tensor var_1745_cast_fp16 = mul(x = var_1726_cast_fp16, y = roti_9_cast_fp16)[name = tensor("op_1745_cast_fp16")]; + tensor kor_17_cast_fp16 = sub(x = var_1744_cast_fp16, y = var_1745_cast_fp16)[name = tensor("kor_17_cast_fp16")]; + tensor var_1748_cast_fp16 = mul(x = var_1718_cast_fp16, y = roti_9_cast_fp16)[name = tensor("op_1748_cast_fp16")]; + tensor var_1749_cast_fp16 = mul(x = var_1726_cast_fp16, y = rotr_9_cast_fp16)[name = tensor("op_1749_cast_fp16")]; + tensor koi_17_cast_fp16 = add(x = var_1748_cast_fp16, y = var_1749_cast_fp16)[name = tensor("koi_17_cast_fp16")]; + tensor qo_9_axis_0 = const()[name = tensor("qo_9_axis_0"), val = tensor(-1)]; + tensor qo_9_cast_fp16 = stack(axis = qo_9_axis_0, values = (qor_17_cast_fp16, qoi_17_cast_fp16))[name = tensor("qo_9_cast_fp16")]; + tensor ko_9_axis_0 = const()[name = tensor("ko_9_axis_0"), val = tensor(-1)]; + tensor ko_9_cast_fp16 = stack(axis = ko_9_axis_0, values = (kor_17_cast_fp16, koi_17_cast_fp16))[name = tensor("ko_9_cast_fp16")]; + tensor var_1782 = const()[name = tensor("op_1782"), val = tensor([1, 1, 16, 64])]; + tensor q_27_cast_fp16 = reshape(shape = var_1782, x = qo_9_cast_fp16)[name = tensor("q_27_cast_fp16")]; + tensor var_1788 = const()[name = tensor("op_1788"), val = tensor([1, 1, 16, 64])]; + tensor k_19_cast_fp16 = reshape(shape = var_1788, x = ko_9_cast_fp16)[name = tensor("k_19_cast_fp16")]; + tensor _inversed_1813_y_0_to_fp16 = const()[name = tensor("_inversed_1813_y_0_to_fp16"), val = tensor(0x1p-9)]; + tensor _inversed_1813_cast_fp16 = mul(x = ts_29_cast_fp16, y = _inversed_1813_y_0_to_fp16)[name = tensor("_inversed_1813_cast_fp16")]; + tensor var_1814_cast_fp16 = floor(x = _inversed_1813_cast_fp16)[name = tensor("op_1814_cast_fp16")]; + tensor var_1815_to_fp16 = const()[name = tensor("op_1815_to_fp16"), val = tensor(0x1p+9)]; + tensor var_1816_cast_fp16 = mul(x = var_1814_cast_fp16, y = var_1815_to_fp16)[name = tensor("op_1816_cast_fp16")]; + tensor write_indices_float_19_cast_fp16 = sub(x = ts_29_cast_fp16, y = var_1816_cast_fp16)[name = tensor("write_indices_float_19_cast_fp16")]; + tensor var_1823_dtype_0 = const()[name = tensor("op_1823_dtype_0"), val = tensor("int32")]; + tensor write_indices_9_reps_0 = const()[name = tensor("write_indices_9_reps_0"), val = tensor([1, 1, 16, 64])]; + tensor write_indices_float_19_cast_fp16_to_int32 = cast(dtype = var_1823_dtype_0, x = write_indices_float_19_cast_fp16)[name = tensor("cast_118")]; + tensor write_indices_9 = tile(reps = write_indices_9_reps_0, x = write_indices_float_19_cast_fp16_to_int32)[name = tensor("write_indices_9")]; + tensor var_1833_begin_0 = const()[name = tensor("op_1833_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_1833_end_0 = const()[name = tensor("op_1833_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor var_1833_end_mask_0 = const()[name = tensor("op_1833_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_1833_squeeze_mask_0 = const()[name = tensor("op_1833_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor cache4_to_fp16_dtype_0 = const()[name = tensor("cache4_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor cache4_to_fp16 = cast(dtype = cache4_to_fp16_dtype_0, x = cache4)[name = tensor("cast_117")]; + tensor var_1833_cast_fp16 = slice_by_index(begin = var_1833_begin_0, end = var_1833_end_0, end_mask = var_1833_end_mask_0, squeeze_mask = var_1833_squeeze_mask_0, x = cache4_to_fp16)[name = tensor("op_1833_cast_fp16")]; + tensor var_1835_axis_0 = const()[name = tensor("op_1835_axis_0"), val = tensor(1)]; + tensor var_1835_mode_0 = const()[name = tensor("op_1835_mode_0"), val = tensor("update")]; + tensor var_1835_validate_indices_0 = const()[name = tensor("op_1835_validate_indices_0"), val = tensor(false)]; + tensor var_1835_cast_fp16 = scatter_along_axis(axis = var_1835_axis_0, data = var_1833_cast_fp16, indices = write_indices_9, mode = var_1835_mode_0, updates = k_19_cast_fp16, validate_indices = var_1835_validate_indices_0)[name = tensor("op_1835_cast_fp16")]; + tensor concat_30 = const()[name = tensor("concat_30"), val = tensor([0, 0, 0, 0, 0])]; + tensor concat_31 = const()[name = tensor("concat_31"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_9_internal_tensor_assign_1_stride_0 = const()[name = tensor("new_cache_9_internal_tensor_assign_1_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_9_internal_tensor_assign_1_begin_mask_0 = const()[name = tensor("new_cache_9_internal_tensor_assign_1_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_9_internal_tensor_assign_1_end_mask_0 = const()[name = tensor("new_cache_9_internal_tensor_assign_1_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_9_internal_tensor_assign_1_squeeze_mask_0 = const()[name = tensor("new_cache_9_internal_tensor_assign_1_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_20 = const()[name = tensor("shape_20"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_8 = const()[name = tensor("reduce_prod_8"), val = tensor(1048576)]; + tensor range_1d_8_start_0 = const()[name = tensor("range_1d_8_start_0"), val = tensor(0)]; + tensor range_1d_8_step_0 = const()[name = tensor("range_1d_8_step_0"), val = tensor(1)]; + tensor range_1d_8 = range_1d(end = reduce_prod_8, start = range_1d_8_start_0, step = range_1d_8_step_0)[name = tensor("range_1d_8")]; + tensor reshape_40 = reshape(shape = shape_20, x = range_1d_8)[name = tensor("reshape_40")]; + tensor slice_by_index_8 = slice_by_index(begin = concat_30, begin_mask = new_cache_9_internal_tensor_assign_1_begin_mask_0, end = concat_31, end_mask = new_cache_9_internal_tensor_assign_1_end_mask_0, squeeze_mask = new_cache_9_internal_tensor_assign_1_squeeze_mask_0, stride = new_cache_9_internal_tensor_assign_1_stride_0, x = reshape_40)[name = tensor("slice_by_index_8")]; + tensor reshape_41_shape_0 = const()[name = tensor("reshape_41_shape_0"), val = tensor([-1])]; + tensor reshape_41 = reshape(shape = reshape_41_shape_0, x = slice_by_index_8)[name = tensor("reshape_41")]; + tensor reshape_42_shape_0 = const()[name = tensor("reshape_42_shape_0"), val = tensor([-1])]; + tensor reshape_42_cast_fp16 = reshape(shape = reshape_42_shape_0, x = var_1835_cast_fp16)[name = tensor("reshape_42_cast_fp16")]; + tensor reshape_43_shape_0 = const()[name = tensor("reshape_43_shape_0"), val = tensor([-1])]; + tensor reshape_43_cast_fp16 = reshape(shape = reshape_43_shape_0, x = cache4_to_fp16)[name = tensor("reshape_43_cast_fp16")]; + tensor scatter_8_mode_0 = const()[name = tensor("scatter_8_mode_0"), val = tensor("update")]; + tensor scatter_8_axis_0 = const()[name = tensor("scatter_8_axis_0"), val = tensor(0)]; + tensor scatter_8_validate_indices_0 = const()[name = tensor("scatter_8_validate_indices_0"), val = tensor(false)]; + tensor scatter_8_cast_fp16 = scatter(axis = scatter_8_axis_0, data = reshape_43_cast_fp16, indices = reshape_41, mode = scatter_8_mode_0, updates = reshape_42_cast_fp16, validate_indices = scatter_8_validate_indices_0)[name = tensor("scatter_8_cast_fp16")]; + tensor reshape_44_cast_fp16 = reshape(shape = shape_20, x = scatter_8_cast_fp16)[name = tensor("reshape_44_cast_fp16")]; + tensor var_1843_begin_0 = const()[name = tensor("op_1843_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor var_1843_end_0 = const()[name = tensor("op_1843_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor var_1843_end_mask_0 = const()[name = tensor("op_1843_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_1843_squeeze_mask_0 = const()[name = tensor("op_1843_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor var_1843_cast_fp16 = slice_by_index(begin = var_1843_begin_0, end = var_1843_end_0, end_mask = var_1843_end_mask_0, squeeze_mask = var_1843_squeeze_mask_0, x = reshape_44_cast_fp16)[name = tensor("op_1843_cast_fp16")]; + tensor var_1845_axis_0 = const()[name = tensor("op_1845_axis_0"), val = tensor(1)]; + tensor var_1845_mode_0 = const()[name = tensor("op_1845_mode_0"), val = tensor("update")]; + tensor var_1845_validate_indices_0 = const()[name = tensor("op_1845_validate_indices_0"), val = tensor(false)]; + tensor var_1845_cast_fp16 = scatter_along_axis(axis = var_1845_axis_0, data = var_1843_cast_fp16, indices = write_indices_9, mode = var_1845_mode_0, updates = v_9_cast_fp16, validate_indices = var_1845_validate_indices_0)[name = tensor("op_1845_cast_fp16")]; + tensor concat_32 = const()[name = tensor("concat_32"), val = tensor([1, 0, 0, 0, 0])]; + tensor concat_33 = const()[name = tensor("concat_33"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_9_internal_tensor_assign_2_stride_0 = const()[name = tensor("new_cache_9_internal_tensor_assign_2_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_9_internal_tensor_assign_2_begin_mask_0 = const()[name = tensor("new_cache_9_internal_tensor_assign_2_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_9_internal_tensor_assign_2_end_mask_0 = const()[name = tensor("new_cache_9_internal_tensor_assign_2_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_9_internal_tensor_assign_2_squeeze_mask_0 = const()[name = tensor("new_cache_9_internal_tensor_assign_2_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_21 = const()[name = tensor("shape_21"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_9 = const()[name = tensor("reduce_prod_9"), val = tensor(1048576)]; + tensor range_1d_9_start_0 = const()[name = tensor("range_1d_9_start_0"), val = tensor(0)]; + tensor range_1d_9_step_0 = const()[name = tensor("range_1d_9_step_0"), val = tensor(1)]; + tensor range_1d_9 = range_1d(end = reduce_prod_9, start = range_1d_9_start_0, step = range_1d_9_step_0)[name = tensor("range_1d_9")]; + tensor reshape_45 = reshape(shape = shape_21, x = range_1d_9)[name = tensor("reshape_45")]; + tensor slice_by_index_9 = slice_by_index(begin = concat_32, begin_mask = new_cache_9_internal_tensor_assign_2_begin_mask_0, end = concat_33, end_mask = new_cache_9_internal_tensor_assign_2_end_mask_0, squeeze_mask = new_cache_9_internal_tensor_assign_2_squeeze_mask_0, stride = new_cache_9_internal_tensor_assign_2_stride_0, x = reshape_45)[name = tensor("slice_by_index_9")]; + tensor reshape_46_shape_0 = const()[name = tensor("reshape_46_shape_0"), val = tensor([-1])]; + tensor reshape_46 = reshape(shape = reshape_46_shape_0, x = slice_by_index_9)[name = tensor("reshape_46")]; + tensor reshape_47_shape_0 = const()[name = tensor("reshape_47_shape_0"), val = tensor([-1])]; + tensor reshape_47_cast_fp16 = reshape(shape = reshape_47_shape_0, x = var_1845_cast_fp16)[name = tensor("reshape_47_cast_fp16")]; + tensor reshape_48_shape_0 = const()[name = tensor("reshape_48_shape_0"), val = tensor([-1])]; + tensor reshape_48_cast_fp16 = reshape(shape = reshape_48_shape_0, x = reshape_44_cast_fp16)[name = tensor("reshape_48_cast_fp16")]; + tensor scatter_9_mode_0 = const()[name = tensor("scatter_9_mode_0"), val = tensor("update")]; + tensor scatter_9_axis_0 = const()[name = tensor("scatter_9_axis_0"), val = tensor(0)]; + tensor scatter_9_validate_indices_0 = const()[name = tensor("scatter_9_validate_indices_0"), val = tensor(false)]; + tensor scatter_9_cast_fp16 = scatter(axis = scatter_9_axis_0, data = reshape_48_cast_fp16, indices = reshape_46, mode = scatter_9_mode_0, updates = reshape_47_cast_fp16, validate_indices = scatter_9_validate_indices_0)[name = tensor("scatter_9_cast_fp16")]; + tensor reshape_49_cast_fp16 = reshape(shape = shape_21, x = scatter_9_cast_fp16)[name = tensor("reshape_49_cast_fp16")]; + tensor reshape_49_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("reshape_49_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor keys_25_begin_0 = const()[name = tensor("keys_25_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor keys_25_end_0 = const()[name = tensor("keys_25_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor keys_25_end_mask_0 = const()[name = tensor("keys_25_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor keys_25_squeeze_mask_0 = const()[name = tensor("keys_25_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor keys_25_cast_fp16 = slice_by_index(begin = keys_25_begin_0, end = keys_25_end_0, end_mask = keys_25_end_mask_0, squeeze_mask = keys_25_squeeze_mask_0, x = reshape_49_cast_fp16)[name = tensor("keys_25_cast_fp16")]; + tensor values_25_begin_0 = const()[name = tensor("values_25_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor values_25_end_0 = const()[name = tensor("values_25_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor values_25_end_mask_0 = const()[name = tensor("values_25_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor values_25_squeeze_mask_0 = const()[name = tensor("values_25_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor values_25_cast_fp16 = slice_by_index(begin = values_25_begin_0, end = values_25_end_0, end_mask = values_25_end_mask_0, squeeze_mask = values_25_squeeze_mask_0, x = reshape_49_cast_fp16)[name = tensor("values_25_cast_fp16")]; + tensor var_1857_cast_fp16 = not_equal(x = keys_25_cast_fp16, y = keys_25_cast_fp16)[name = tensor("op_1857_cast_fp16")]; + tensor keys_27_cast_fp16 = select(a = var_319_to_fp16, b = keys_25_cast_fp16, cond = var_1857_cast_fp16)[name = tensor("keys_27_cast_fp16")]; + tensor var_1865_cast_fp16 = not_equal(x = values_25_cast_fp16, y = values_25_cast_fp16)[name = tensor("op_1865_cast_fp16")]; + tensor values_27_cast_fp16 = select(a = var_319_to_fp16, b = values_25_cast_fp16, cond = var_1865_cast_fp16)[name = tensor("values_27_cast_fp16")]; + tensor var_1889 = const()[name = tensor("op_1889"), val = tensor([0, 2, 1, 3])]; + tensor var_1905 = const()[name = tensor("op_1905"), val = tensor([1, 1, 1])]; + tensor var_1906_cast_fp16 = reshape(shape = var_1905, x = position4_to_fp16)[name = tensor("op_1906_cast_fp16")]; + tensor var_1926_to_fp16 = const()[name = tensor("op_1926_to_fp16"), val = tensor(0x1p+0)]; + tensor valid_len_9_cast_fp16 = add(x = var_1906_cast_fp16, y = var_1926_to_fp16)[name = tensor("valid_len_9_cast_fp16")]; + tensor valid_mask_9_cast_fp16 = less(x = k_positions_1_promoted_to_fp16, y = valid_len_9_cast_fp16)[name = tensor("valid_mask_9_cast_fp16")]; + tensor causal_mask_9_cast_fp16 = less_equal(x = k_positions_1_promoted_to_fp16, y = var_1906_cast_fp16)[name = tensor("causal_mask_9_cast_fp16")]; + tensor attn_mask_17 = logical_and(x = valid_mask_9_cast_fp16, y = causal_mask_9_cast_fp16)[name = tensor("attn_mask_17")]; + tensor attn_mask_19_axes_0 = const()[name = tensor("attn_mask_19_axes_0"), val = tensor([1])]; + tensor attn_mask_19 = expand_dims(axes = attn_mask_19_axes_0, x = attn_mask_17)[name = tensor("attn_mask_19")]; + tensor var_1944_transpose_x_0 = const()[name = tensor("op_1944_transpose_x_0"), val = tensor(false)]; + tensor var_1944_transpose_y_0 = const()[name = tensor("op_1944_transpose_y_0"), val = tensor(false)]; + tensor transpose_26_perm_0 = const()[name = tensor("transpose_26_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_27_perm_0 = const()[name = tensor("transpose_27_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_27 = transpose(perm = transpose_27_perm_0, x = keys_27_cast_fp16)[name = tensor("transpose_35")]; + tensor transpose_26 = transpose(perm = transpose_26_perm_0, x = q_27_cast_fp16)[name = tensor("transpose_36")]; + tensor var_1944_cast_fp16 = matmul(transpose_x = var_1944_transpose_x_0, transpose_y = var_1944_transpose_y_0, x = transpose_26, y = transpose_27)[name = tensor("op_1944_cast_fp16")]; + tensor var_1938_to_fp16 = const()[name = tensor("op_1938_to_fp16"), val = tensor([0x1p-3])]; + tensor attn_weights_25_cast_fp16 = mul(x = var_1944_cast_fp16, y = var_1938_to_fp16)[name = tensor("attn_weights_25_cast_fp16")]; + tensor var_1946 = logical_not(x = attn_mask_19)[name = tensor("op_1946")]; + tensor var_1947_to_fp16 = const()[name = tensor("op_1947_to_fp16"), val = tensor(-inf)]; + tensor attn_weights_27_cast_fp16 = select(a = var_1947_to_fp16, b = attn_weights_25_cast_fp16, cond = var_1946)[name = tensor("attn_weights_27_cast_fp16")]; + tensor var_1949 = const()[name = tensor("op_1949"), val = tensor(-1)]; + tensor attn_weights_29_cast_fp16 = softmax(axis = var_1949, x = attn_weights_27_cast_fp16)[name = tensor("attn_weights_29_cast_fp16")]; + tensor attn_output_9_transpose_x_0 = const()[name = tensor("attn_output_9_transpose_x_0"), val = tensor(false)]; + tensor attn_output_9_transpose_y_0 = const()[name = tensor("attn_output_9_transpose_y_0"), val = tensor(false)]; + tensor values_29_cast_fp16 = transpose(perm = var_1889, x = values_27_cast_fp16)[name = tensor("transpose_37")]; + tensor attn_output_9_cast_fp16 = matmul(transpose_x = attn_output_9_transpose_x_0, transpose_y = attn_output_9_transpose_y_0, x = attn_weights_29_cast_fp16, y = values_29_cast_fp16)[name = tensor("attn_output_9_cast_fp16")]; + tensor var_1957 = const()[name = tensor("op_1957"), val = tensor([0, 2, 1, 3])]; + tensor var_1962 = const()[name = tensor("op_1962"), val = tensor([1, 1, 1024])]; + tensor var_1958_cast_fp16 = transpose(perm = var_1957, x = attn_output_9_cast_fp16)[name = tensor("transpose_34")]; + tensor input_55_cast_fp16 = reshape(shape = var_1962, x = var_1958_cast_fp16)[name = tensor("input_55_cast_fp16")]; + tensor attn4_out_proj_weight_to_fp16 = const()[name = tensor("attn4_out_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(108126464)))]; + tensor linear_18_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = attn4_out_proj_weight_to_fp16, x = input_55_cast_fp16)[name = tensor("linear_18_cast_fp16")]; + tensor var_1968_to_fp16 = const()[name = tensor("op_1968_to_fp16"), val = tensor(0x1p+0)]; + tensor var_1969_cast_fp16 = add(x = position4_to_fp16, y = var_1968_to_fp16)[name = tensor("op_1969_cast_fp16")]; + tensor var_1969_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_1969_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor input_57_cast_fp16 = add(x = input_51_cast_fp16, y = linear_18_cast_fp16)[name = tensor("input_57_cast_fp16")]; + tensor input_59_axes_0 = const()[name = tensor("input_59_axes_0"), val = tensor([-1])]; + tensor norm4_2_weight_to_fp16 = const()[name = tensor("norm4_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(110223680)))]; + tensor norm4_2_bias_to_fp16 = const()[name = tensor("norm4_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(110225792)))]; + tensor var_1973_to_fp16 = const()[name = tensor("op_1973_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_59_cast_fp16 = layer_norm(axes = input_59_axes_0, beta = norm4_2_bias_to_fp16, epsilon = var_1973_to_fp16, gamma = norm4_2_weight_to_fp16, x = input_57_cast_fp16)[name = tensor("input_59_cast_fp16")]; + tensor linear4_1_weight_to_fp16 = const()[name = tensor("linear4_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(110227904)))]; + tensor linear_19_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = linear4_1_weight_to_fp16, x = input_59_cast_fp16)[name = tensor("linear_19_cast_fp16")]; + tensor input_61_mode_0 = const()[name = tensor("input_61_mode_0"), val = tensor("EXACT")]; + tensor input_61_cast_fp16 = gelu(mode = input_61_mode_0, x = linear_19_cast_fp16)[name = tensor("input_61_cast_fp16")]; + tensor linear4_2_weight_to_fp16 = const()[name = tensor("linear4_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(118616576)))]; + tensor linear_20_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = linear4_2_weight_to_fp16, x = input_61_cast_fp16)[name = tensor("linear_20_cast_fp16")]; + tensor input_63_cast_fp16 = add(x = input_57_cast_fp16, y = linear_20_cast_fp16)[name = tensor("input_63_cast_fp16")]; + tensor input_65_axes_0 = const()[name = tensor("input_65_axes_0"), val = tensor([-1])]; + tensor norm5_1_weight_to_fp16 = const()[name = tensor("norm5_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(127005248)))]; + tensor norm5_1_bias_to_fp16 = const()[name = tensor("norm5_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(127007360)))]; + tensor var_1990_to_fp16 = const()[name = tensor("op_1990_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_65_cast_fp16 = layer_norm(axes = input_65_axes_0, beta = norm5_1_bias_to_fp16, epsilon = var_1990_to_fp16, gamma = norm5_1_weight_to_fp16, x = input_63_cast_fp16)[name = tensor("input_65_cast_fp16")]; + tensor attn5_in_proj_weight_to_fp16 = const()[name = tensor("attn5_in_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(127009472)))]; + tensor linear_21_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = attn5_in_proj_weight_to_fp16, x = input_65_cast_fp16)[name = tensor("linear_21_cast_fp16")]; + tensor var_2004 = const()[name = tensor("op_2004"), val = tensor([1, 1, 3, 16, 64])]; + tensor qkv_cast_fp16 = reshape(shape = var_2004, x = linear_21_cast_fp16)[name = tensor("qkv_cast_fp16")]; + tensor q_31_begin_0 = const()[name = tensor("q_31_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor q_31_end_0 = const()[name = tensor("q_31_end_0"), val = tensor([1, 1, 1, 16, 64])]; + tensor q_31_end_mask_0 = const()[name = tensor("q_31_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor q_31_squeeze_mask_0 = const()[name = tensor("q_31_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor q_31_cast_fp16 = slice_by_index(begin = q_31_begin_0, end = q_31_end_0, end_mask = q_31_end_mask_0, squeeze_mask = q_31_squeeze_mask_0, x = qkv_cast_fp16)[name = tensor("q_31_cast_fp16")]; + tensor k_21_begin_0 = const()[name = tensor("k_21_begin_0"), val = tensor([0, 0, 1, 0, 0])]; + tensor k_21_end_0 = const()[name = tensor("k_21_end_0"), val = tensor([1, 1, 2, 16, 64])]; + tensor k_21_end_mask_0 = const()[name = tensor("k_21_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor k_21_squeeze_mask_0 = const()[name = tensor("k_21_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor k_21_cast_fp16 = slice_by_index(begin = k_21_begin_0, end = k_21_end_0, end_mask = k_21_end_mask_0, squeeze_mask = k_21_squeeze_mask_0, x = qkv_cast_fp16)[name = tensor("k_21_cast_fp16")]; + tensor v_begin_0 = const()[name = tensor("v_begin_0"), val = tensor([0, 0, 2, 0, 0])]; + tensor v_end_0 = const()[name = tensor("v_end_0"), val = tensor([1, 1, 3, 16, 64])]; + tensor v_end_mask_0 = const()[name = tensor("v_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor v_squeeze_mask_0 = const()[name = tensor("v_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor v_cast_fp16 = slice_by_index(begin = v_begin_0, end = v_end_0, end_mask = v_end_mask_0, squeeze_mask = v_squeeze_mask_0, x = qkv_cast_fp16)[name = tensor("v_cast_fp16")]; + tensor var_2070 = const()[name = tensor("op_2070"), val = tensor([1, 1, 1, 1])]; + tensor position5_to_fp16_dtype_0 = const()[name = tensor("position5_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor position5_to_fp16 = cast(dtype = position5_to_fp16_dtype_0, x = position5)[name = tensor("cast_114")]; + tensor ts_cast_fp16 = reshape(shape = var_2070, x = position5_to_fp16)[name = tensor("ts_cast_fp16")]; + tensor var_2077 = const()[name = tensor("op_2077"), val = tensor([1, 1, 16, 32, 2])]; + tensor q_complex_cast_fp16 = reshape(shape = var_2077, x = q_31_cast_fp16)[name = tensor("q_complex_cast_fp16")]; + tensor var_2084 = const()[name = tensor("op_2084"), val = tensor([1, 1, 16, 32, 2])]; + tensor k_complex_cast_fp16 = reshape(shape = var_2084, x = k_21_cast_fp16)[name = tensor("k_complex_cast_fp16")]; + tensor var_2088_begin_0 = const()[name = tensor("op_2088_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_2088_end_0 = const()[name = tensor("op_2088_end_0"), val = tensor([1, 1, 16, 32, 1])]; + tensor var_2088_end_mask_0 = const()[name = tensor("op_2088_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_2088_squeeze_mask_0 = const()[name = tensor("op_2088_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_2088_cast_fp16 = slice_by_index(begin = var_2088_begin_0, end = var_2088_end_0, end_mask = var_2088_end_mask_0, squeeze_mask = var_2088_squeeze_mask_0, x = q_complex_cast_fp16)[name = tensor("op_2088_cast_fp16")]; + tensor var_2096_begin_0 = const()[name = tensor("op_2096_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_2096_end_0 = const()[name = tensor("op_2096_end_0"), val = tensor([1, 1, 16, 32, 2])]; + tensor var_2096_end_mask_0 = const()[name = tensor("op_2096_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_2096_squeeze_mask_0 = const()[name = tensor("op_2096_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_2096_cast_fp16 = slice_by_index(begin = var_2096_begin_0, end = var_2096_end_0, end_mask = var_2096_end_mask_0, squeeze_mask = var_2096_squeeze_mask_0, x = q_complex_cast_fp16)[name = tensor("op_2096_cast_fp16")]; + tensor var_2104_begin_0 = const()[name = tensor("op_2104_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_2104_end_0 = const()[name = tensor("op_2104_end_0"), val = tensor([1, 1, 16, 32, 1])]; + tensor var_2104_end_mask_0 = const()[name = tensor("op_2104_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_2104_squeeze_mask_0 = const()[name = tensor("op_2104_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_2104_cast_fp16 = slice_by_index(begin = var_2104_begin_0, end = var_2104_end_0, end_mask = var_2104_end_mask_0, squeeze_mask = var_2104_squeeze_mask_0, x = k_complex_cast_fp16)[name = tensor("op_2104_cast_fp16")]; + tensor var_2112_begin_0 = const()[name = tensor("op_2112_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_2112_end_0 = const()[name = tensor("op_2112_end_0"), val = tensor([1, 1, 16, 32, 2])]; + tensor var_2112_end_mask_0 = const()[name = tensor("op_2112_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_2112_squeeze_mask_0 = const()[name = tensor("op_2112_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_2112_cast_fp16 = slice_by_index(begin = var_2112_begin_0, end = var_2112_end_0, end_mask = var_2112_end_mask_0, squeeze_mask = var_2112_squeeze_mask_0, x = k_complex_cast_fp16)[name = tensor("op_2112_cast_fp16")]; + tensor freqs_to_fp16 = const()[name = tensor("freqs_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(133300992)))]; + tensor var_2118_cast_fp16 = mul(x = freqs_to_fp16, y = ts_cast_fp16)[name = tensor("op_2118_cast_fp16")]; + tensor rotr_cast_fp16 = cos(x = var_2118_cast_fp16)[name = tensor("rotr_cast_fp16")]; + tensor roti_cast_fp16 = sin(x = var_2118_cast_fp16)[name = tensor("roti_cast_fp16")]; + tensor var_2122_cast_fp16 = mul(x = var_2088_cast_fp16, y = rotr_cast_fp16)[name = tensor("op_2122_cast_fp16")]; + tensor var_2123_cast_fp16 = mul(x = var_2096_cast_fp16, y = roti_cast_fp16)[name = tensor("op_2123_cast_fp16")]; + tensor qor_21_cast_fp16 = sub(x = var_2122_cast_fp16, y = var_2123_cast_fp16)[name = tensor("qor_21_cast_fp16")]; + tensor var_2126_cast_fp16 = mul(x = var_2088_cast_fp16, y = roti_cast_fp16)[name = tensor("op_2126_cast_fp16")]; + tensor var_2127_cast_fp16 = mul(x = var_2096_cast_fp16, y = rotr_cast_fp16)[name = tensor("op_2127_cast_fp16")]; + tensor qoi_21_cast_fp16 = add(x = var_2126_cast_fp16, y = var_2127_cast_fp16)[name = tensor("qoi_21_cast_fp16")]; + tensor var_2130_cast_fp16 = mul(x = var_2104_cast_fp16, y = rotr_cast_fp16)[name = tensor("op_2130_cast_fp16")]; + tensor var_2131_cast_fp16 = mul(x = var_2112_cast_fp16, y = roti_cast_fp16)[name = tensor("op_2131_cast_fp16")]; + tensor kor_21_cast_fp16 = sub(x = var_2130_cast_fp16, y = var_2131_cast_fp16)[name = tensor("kor_21_cast_fp16")]; + tensor var_2134_cast_fp16 = mul(x = var_2104_cast_fp16, y = roti_cast_fp16)[name = tensor("op_2134_cast_fp16")]; + tensor var_2135_cast_fp16 = mul(x = var_2112_cast_fp16, y = rotr_cast_fp16)[name = tensor("op_2135_cast_fp16")]; + tensor koi_21_cast_fp16 = add(x = var_2134_cast_fp16, y = var_2135_cast_fp16)[name = tensor("koi_21_cast_fp16")]; + tensor qo_axis_0 = const()[name = tensor("qo_axis_0"), val = tensor(-1)]; + tensor qo_cast_fp16 = stack(axis = qo_axis_0, values = (qor_21_cast_fp16, qoi_21_cast_fp16))[name = tensor("qo_cast_fp16")]; + tensor ko_axis_0 = const()[name = tensor("ko_axis_0"), val = tensor(-1)]; + tensor ko_cast_fp16 = stack(axis = ko_axis_0, values = (kor_21_cast_fp16, koi_21_cast_fp16))[name = tensor("ko_cast_fp16")]; + tensor var_2168 = const()[name = tensor("op_2168"), val = tensor([1, 1, 16, 64])]; + tensor q_33_cast_fp16 = reshape(shape = var_2168, x = qo_cast_fp16)[name = tensor("q_33_cast_fp16")]; + tensor var_2174 = const()[name = tensor("op_2174"), val = tensor([1, 1, 16, 64])]; + tensor k_cast_fp16 = reshape(shape = var_2174, x = ko_cast_fp16)[name = tensor("k_cast_fp16")]; + tensor _inversed_2199_y_0_to_fp16 = const()[name = tensor("_inversed_2199_y_0_to_fp16"), val = tensor(0x1p-9)]; + tensor _inversed_2199_cast_fp16 = mul(x = ts_cast_fp16, y = _inversed_2199_y_0_to_fp16)[name = tensor("_inversed_2199_cast_fp16")]; + tensor var_2200_cast_fp16 = floor(x = _inversed_2199_cast_fp16)[name = tensor("op_2200_cast_fp16")]; + tensor var_2201_to_fp16 = const()[name = tensor("op_2201_to_fp16"), val = tensor(0x1p+9)]; + tensor var_2202_cast_fp16 = mul(x = var_2200_cast_fp16, y = var_2201_to_fp16)[name = tensor("op_2202_cast_fp16")]; + tensor write_indices_float_cast_fp16 = sub(x = ts_cast_fp16, y = var_2202_cast_fp16)[name = tensor("write_indices_float_cast_fp16")]; + tensor var_2209_dtype_0 = const()[name = tensor("op_2209_dtype_0"), val = tensor("int32")]; + tensor write_indices_reps_0 = const()[name = tensor("write_indices_reps_0"), val = tensor([1, 1, 16, 64])]; + tensor write_indices_float_cast_fp16_to_int32 = cast(dtype = var_2209_dtype_0, x = write_indices_float_cast_fp16)[name = tensor("cast_113")]; + tensor write_indices = tile(reps = write_indices_reps_0, x = write_indices_float_cast_fp16_to_int32)[name = tensor("write_indices")]; + tensor var_2219_begin_0 = const()[name = tensor("op_2219_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_2219_end_0 = const()[name = tensor("op_2219_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor var_2219_end_mask_0 = const()[name = tensor("op_2219_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_2219_squeeze_mask_0 = const()[name = tensor("op_2219_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor cache5_to_fp16_dtype_0 = const()[name = tensor("cache5_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor cache5_to_fp16 = cast(dtype = cache5_to_fp16_dtype_0, x = cache5)[name = tensor("cast_112")]; + tensor var_2219_cast_fp16 = slice_by_index(begin = var_2219_begin_0, end = var_2219_end_0, end_mask = var_2219_end_mask_0, squeeze_mask = var_2219_squeeze_mask_0, x = cache5_to_fp16)[name = tensor("op_2219_cast_fp16")]; + tensor var_2221_axis_0 = const()[name = tensor("op_2221_axis_0"), val = tensor(1)]; + tensor var_2221_mode_0 = const()[name = tensor("op_2221_mode_0"), val = tensor("update")]; + tensor var_2221_validate_indices_0 = const()[name = tensor("op_2221_validate_indices_0"), val = tensor(false)]; + tensor var_2221_cast_fp16 = scatter_along_axis(axis = var_2221_axis_0, data = var_2219_cast_fp16, indices = write_indices, mode = var_2221_mode_0, updates = k_cast_fp16, validate_indices = var_2221_validate_indices_0)[name = tensor("op_2221_cast_fp16")]; + tensor concat_37 = const()[name = tensor("concat_37"), val = tensor([0, 0, 0, 0, 0])]; + tensor concat_38 = const()[name = tensor("concat_38"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_internal_tensor_assign_1_stride_0 = const()[name = tensor("new_cache_internal_tensor_assign_1_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_internal_tensor_assign_1_begin_mask_0 = const()[name = tensor("new_cache_internal_tensor_assign_1_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_internal_tensor_assign_1_end_mask_0 = const()[name = tensor("new_cache_internal_tensor_assign_1_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_internal_tensor_assign_1_squeeze_mask_0 = const()[name = tensor("new_cache_internal_tensor_assign_1_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_22 = const()[name = tensor("shape_22"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_10 = const()[name = tensor("reduce_prod_10"), val = tensor(1048576)]; + tensor range_1d_10_start_0 = const()[name = tensor("range_1d_10_start_0"), val = tensor(0)]; + tensor range_1d_10_step_0 = const()[name = tensor("range_1d_10_step_0"), val = tensor(1)]; + tensor range_1d_10 = range_1d(end = reduce_prod_10, start = range_1d_10_start_0, step = range_1d_10_step_0)[name = tensor("range_1d_10")]; + tensor reshape_50 = reshape(shape = shape_22, x = range_1d_10)[name = tensor("reshape_50")]; + tensor slice_by_index_10 = slice_by_index(begin = concat_37, begin_mask = new_cache_internal_tensor_assign_1_begin_mask_0, end = concat_38, end_mask = new_cache_internal_tensor_assign_1_end_mask_0, squeeze_mask = new_cache_internal_tensor_assign_1_squeeze_mask_0, stride = new_cache_internal_tensor_assign_1_stride_0, x = reshape_50)[name = tensor("slice_by_index_10")]; + tensor reshape_51_shape_0 = const()[name = tensor("reshape_51_shape_0"), val = tensor([-1])]; + tensor reshape_51 = reshape(shape = reshape_51_shape_0, x = slice_by_index_10)[name = tensor("reshape_51")]; + tensor reshape_52_shape_0 = const()[name = tensor("reshape_52_shape_0"), val = tensor([-1])]; + tensor reshape_52_cast_fp16 = reshape(shape = reshape_52_shape_0, x = var_2221_cast_fp16)[name = tensor("reshape_52_cast_fp16")]; + tensor reshape_53_shape_0 = const()[name = tensor("reshape_53_shape_0"), val = tensor([-1])]; + tensor reshape_53_cast_fp16 = reshape(shape = reshape_53_shape_0, x = cache5_to_fp16)[name = tensor("reshape_53_cast_fp16")]; + tensor scatter_10_mode_0 = const()[name = tensor("scatter_10_mode_0"), val = tensor("update")]; + tensor scatter_10_axis_0 = const()[name = tensor("scatter_10_axis_0"), val = tensor(0)]; + tensor scatter_10_validate_indices_0 = const()[name = tensor("scatter_10_validate_indices_0"), val = tensor(false)]; + tensor scatter_10_cast_fp16 = scatter(axis = scatter_10_axis_0, data = reshape_53_cast_fp16, indices = reshape_51, mode = scatter_10_mode_0, updates = reshape_52_cast_fp16, validate_indices = scatter_10_validate_indices_0)[name = tensor("scatter_10_cast_fp16")]; + tensor reshape_54_cast_fp16 = reshape(shape = shape_22, x = scatter_10_cast_fp16)[name = tensor("reshape_54_cast_fp16")]; + tensor var_2229_begin_0 = const()[name = tensor("op_2229_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor var_2229_end_0 = const()[name = tensor("op_2229_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor var_2229_end_mask_0 = const()[name = tensor("op_2229_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_2229_squeeze_mask_0 = const()[name = tensor("op_2229_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor var_2229_cast_fp16 = slice_by_index(begin = var_2229_begin_0, end = var_2229_end_0, end_mask = var_2229_end_mask_0, squeeze_mask = var_2229_squeeze_mask_0, x = reshape_54_cast_fp16)[name = tensor("op_2229_cast_fp16")]; + tensor var_2231_axis_0 = const()[name = tensor("op_2231_axis_0"), val = tensor(1)]; + tensor var_2231_mode_0 = const()[name = tensor("op_2231_mode_0"), val = tensor("update")]; + tensor var_2231_validate_indices_0 = const()[name = tensor("op_2231_validate_indices_0"), val = tensor(false)]; + tensor var_2231_cast_fp16 = scatter_along_axis(axis = var_2231_axis_0, data = var_2229_cast_fp16, indices = write_indices, mode = var_2231_mode_0, updates = v_cast_fp16, validate_indices = var_2231_validate_indices_0)[name = tensor("op_2231_cast_fp16")]; + tensor concat_39 = const()[name = tensor("concat_39"), val = tensor([1, 0, 0, 0, 0])]; + tensor concat_40 = const()[name = tensor("concat_40"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_internal_tensor_assign_2_stride_0 = const()[name = tensor("new_cache_internal_tensor_assign_2_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_internal_tensor_assign_2_begin_mask_0 = const()[name = tensor("new_cache_internal_tensor_assign_2_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_internal_tensor_assign_2_end_mask_0 = const()[name = tensor("new_cache_internal_tensor_assign_2_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_internal_tensor_assign_2_squeeze_mask_0 = const()[name = tensor("new_cache_internal_tensor_assign_2_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_23 = const()[name = tensor("shape_23"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_11 = const()[name = tensor("reduce_prod_11"), val = tensor(1048576)]; + tensor range_1d_11_start_0 = const()[name = tensor("range_1d_11_start_0"), val = tensor(0)]; + tensor range_1d_11_step_0 = const()[name = tensor("range_1d_11_step_0"), val = tensor(1)]; + tensor range_1d_11 = range_1d(end = reduce_prod_11, start = range_1d_11_start_0, step = range_1d_11_step_0)[name = tensor("range_1d_11")]; + tensor reshape_55 = reshape(shape = shape_23, x = range_1d_11)[name = tensor("reshape_55")]; + tensor slice_by_index_11 = slice_by_index(begin = concat_39, begin_mask = new_cache_internal_tensor_assign_2_begin_mask_0, end = concat_40, end_mask = new_cache_internal_tensor_assign_2_end_mask_0, squeeze_mask = new_cache_internal_tensor_assign_2_squeeze_mask_0, stride = new_cache_internal_tensor_assign_2_stride_0, x = reshape_55)[name = tensor("slice_by_index_11")]; + tensor reshape_56_shape_0 = const()[name = tensor("reshape_56_shape_0"), val = tensor([-1])]; + tensor reshape_56 = reshape(shape = reshape_56_shape_0, x = slice_by_index_11)[name = tensor("reshape_56")]; + tensor reshape_57_shape_0 = const()[name = tensor("reshape_57_shape_0"), val = tensor([-1])]; + tensor reshape_57_cast_fp16 = reshape(shape = reshape_57_shape_0, x = var_2231_cast_fp16)[name = tensor("reshape_57_cast_fp16")]; + tensor reshape_58_shape_0 = const()[name = tensor("reshape_58_shape_0"), val = tensor([-1])]; + tensor reshape_58_cast_fp16 = reshape(shape = reshape_58_shape_0, x = reshape_54_cast_fp16)[name = tensor("reshape_58_cast_fp16")]; + tensor scatter_11_mode_0 = const()[name = tensor("scatter_11_mode_0"), val = tensor("update")]; + tensor scatter_11_axis_0 = const()[name = tensor("scatter_11_axis_0"), val = tensor(0)]; + tensor scatter_11_validate_indices_0 = const()[name = tensor("scatter_11_validate_indices_0"), val = tensor(false)]; + tensor scatter_11_cast_fp16 = scatter(axis = scatter_11_axis_0, data = reshape_58_cast_fp16, indices = reshape_56, mode = scatter_11_mode_0, updates = reshape_57_cast_fp16, validate_indices = scatter_11_validate_indices_0)[name = tensor("scatter_11_cast_fp16")]; + tensor reshape_59_cast_fp16 = reshape(shape = shape_23, x = scatter_11_cast_fp16)[name = tensor("reshape_59_cast_fp16")]; + tensor reshape_59_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("reshape_59_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor keys_31_begin_0 = const()[name = tensor("keys_31_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor keys_31_end_0 = const()[name = tensor("keys_31_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor keys_31_end_mask_0 = const()[name = tensor("keys_31_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor keys_31_squeeze_mask_0 = const()[name = tensor("keys_31_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor keys_31_cast_fp16 = slice_by_index(begin = keys_31_begin_0, end = keys_31_end_0, end_mask = keys_31_end_mask_0, squeeze_mask = keys_31_squeeze_mask_0, x = reshape_59_cast_fp16)[name = tensor("keys_31_cast_fp16")]; + tensor values_31_begin_0 = const()[name = tensor("values_31_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor values_31_end_0 = const()[name = tensor("values_31_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor values_31_end_mask_0 = const()[name = tensor("values_31_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor values_31_squeeze_mask_0 = const()[name = tensor("values_31_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor values_31_cast_fp16 = slice_by_index(begin = values_31_begin_0, end = values_31_end_0, end_mask = values_31_end_mask_0, squeeze_mask = values_31_squeeze_mask_0, x = reshape_59_cast_fp16)[name = tensor("values_31_cast_fp16")]; + tensor var_2243_cast_fp16 = not_equal(x = keys_31_cast_fp16, y = keys_31_cast_fp16)[name = tensor("op_2243_cast_fp16")]; + tensor keys_33_cast_fp16 = select(a = var_319_to_fp16, b = keys_31_cast_fp16, cond = var_2243_cast_fp16)[name = tensor("keys_33_cast_fp16")]; + tensor var_2251_cast_fp16 = not_equal(x = values_31_cast_fp16, y = values_31_cast_fp16)[name = tensor("op_2251_cast_fp16")]; + tensor values_33_cast_fp16 = select(a = var_319_to_fp16, b = values_31_cast_fp16, cond = var_2251_cast_fp16)[name = tensor("values_33_cast_fp16")]; + tensor var_2275 = const()[name = tensor("op_2275"), val = tensor([0, 2, 1, 3])]; + tensor var_2291 = const()[name = tensor("op_2291"), val = tensor([1, 1, 1])]; + tensor var_2292_cast_fp16 = reshape(shape = var_2291, x = position5_to_fp16)[name = tensor("op_2292_cast_fp16")]; + tensor var_2312_to_fp16 = const()[name = tensor("op_2312_to_fp16"), val = tensor(0x1p+0)]; + tensor valid_len_cast_fp16 = add(x = var_2292_cast_fp16, y = var_2312_to_fp16)[name = tensor("valid_len_cast_fp16")]; + tensor valid_mask_cast_fp16 = less(x = k_positions_1_promoted_to_fp16, y = valid_len_cast_fp16)[name = tensor("valid_mask_cast_fp16")]; + tensor causal_mask_cast_fp16 = less_equal(x = k_positions_1_promoted_to_fp16, y = var_2292_cast_fp16)[name = tensor("causal_mask_cast_fp16")]; + tensor attn_mask_21 = logical_and(x = valid_mask_cast_fp16, y = causal_mask_cast_fp16)[name = tensor("attn_mask_21")]; + tensor attn_mask_axes_0 = const()[name = tensor("attn_mask_axes_0"), val = tensor([1])]; + tensor attn_mask = expand_dims(axes = attn_mask_axes_0, x = attn_mask_21)[name = tensor("attn_mask")]; + tensor var_2330_transpose_x_0 = const()[name = tensor("op_2330_transpose_x_0"), val = tensor(false)]; + tensor var_2330_transpose_y_0 = const()[name = tensor("op_2330_transpose_y_0"), val = tensor(false)]; + tensor transpose_28_perm_0 = const()[name = tensor("transpose_28_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_29_perm_0 = const()[name = tensor("transpose_29_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_29 = transpose(perm = transpose_29_perm_0, x = keys_33_cast_fp16)[name = tensor("transpose_31")]; + tensor transpose_28 = transpose(perm = transpose_28_perm_0, x = q_33_cast_fp16)[name = tensor("transpose_32")]; + tensor var_2330_cast_fp16 = matmul(transpose_x = var_2330_transpose_x_0, transpose_y = var_2330_transpose_y_0, x = transpose_28, y = transpose_29)[name = tensor("op_2330_cast_fp16")]; + tensor var_2324_to_fp16 = const()[name = tensor("op_2324_to_fp16"), val = tensor([0x1p-3])]; + tensor attn_weights_31_cast_fp16 = mul(x = var_2330_cast_fp16, y = var_2324_to_fp16)[name = tensor("attn_weights_31_cast_fp16")]; + tensor var_2332 = logical_not(x = attn_mask)[name = tensor("op_2332")]; + tensor var_2333_to_fp16 = const()[name = tensor("op_2333_to_fp16"), val = tensor(-inf)]; + tensor attn_weights_33_cast_fp16 = select(a = var_2333_to_fp16, b = attn_weights_31_cast_fp16, cond = var_2332)[name = tensor("attn_weights_33_cast_fp16")]; + tensor var_2335 = const()[name = tensor("op_2335"), val = tensor(-1)]; + tensor attn_weights_cast_fp16 = softmax(axis = var_2335, x = attn_weights_33_cast_fp16)[name = tensor("attn_weights_cast_fp16")]; + tensor attn_output_transpose_x_0 = const()[name = tensor("attn_output_transpose_x_0"), val = tensor(false)]; + tensor attn_output_transpose_y_0 = const()[name = tensor("attn_output_transpose_y_0"), val = tensor(false)]; + tensor values_cast_fp16 = transpose(perm = var_2275, x = values_33_cast_fp16)[name = tensor("transpose_33")]; + tensor attn_output_cast_fp16 = matmul(transpose_x = attn_output_transpose_x_0, transpose_y = attn_output_transpose_y_0, x = attn_weights_cast_fp16, y = values_cast_fp16)[name = tensor("attn_output_cast_fp16")]; + tensor var_2343 = const()[name = tensor("op_2343"), val = tensor([0, 2, 1, 3])]; + tensor var_2348 = const()[name = tensor("op_2348"), val = tensor([1, 1, 1024])]; + tensor var_2344_cast_fp16 = transpose(perm = var_2343, x = attn_output_cast_fp16)[name = tensor("transpose_30")]; + tensor input_67_cast_fp16 = reshape(shape = var_2348, x = var_2344_cast_fp16)[name = tensor("input_67_cast_fp16")]; + tensor attn5_out_proj_weight_to_fp16 = const()[name = tensor("attn5_out_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(133301120)))]; + tensor linear_22_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = attn5_out_proj_weight_to_fp16, x = input_67_cast_fp16)[name = tensor("linear_22_cast_fp16")]; + tensor var_2354_to_fp16 = const()[name = tensor("op_2354_to_fp16"), val = tensor(0x1p+0)]; + tensor var_2355_cast_fp16 = add(x = position5_to_fp16, y = var_2354_to_fp16)[name = tensor("op_2355_cast_fp16")]; + tensor var_2355_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_2355_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor input_69_cast_fp16 = add(x = input_63_cast_fp16, y = linear_22_cast_fp16)[name = tensor("input_69_cast_fp16")]; + tensor input_71_axes_0 = const()[name = tensor("input_71_axes_0"), val = tensor([-1])]; + tensor norm5_2_weight_to_fp16 = const()[name = tensor("norm5_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(135398336)))]; + tensor norm5_2_bias_to_fp16 = const()[name = tensor("norm5_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(135400448)))]; + tensor var_2359_to_fp16 = const()[name = tensor("op_2359_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_71_cast_fp16 = layer_norm(axes = input_71_axes_0, beta = norm5_2_bias_to_fp16, epsilon = var_2359_to_fp16, gamma = norm5_2_weight_to_fp16, x = input_69_cast_fp16)[name = tensor("input_71_cast_fp16")]; + tensor linear5_1_weight_to_fp16 = const()[name = tensor("linear5_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(135402560)))]; + tensor linear_23_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = linear5_1_weight_to_fp16, x = input_71_cast_fp16)[name = tensor("linear_23_cast_fp16")]; + tensor input_73_mode_0 = const()[name = tensor("input_73_mode_0"), val = tensor("EXACT")]; + tensor input_73_cast_fp16 = gelu(mode = input_73_mode_0, x = linear_23_cast_fp16)[name = tensor("input_73_cast_fp16")]; + tensor linear5_2_weight_to_fp16 = const()[name = tensor("linear5_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(143791232)))]; + tensor linear_24_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = linear5_2_weight_to_fp16, x = input_73_cast_fp16)[name = tensor("linear_24_cast_fp16")]; + tensor input_75_cast_fp16 = add(x = input_69_cast_fp16, y = linear_24_cast_fp16)[name = tensor("input_75_cast_fp16")]; + tensor input_axes_0 = const()[name = tensor("input_axes_0"), val = tensor([-1])]; + tensor out_norm_weight_to_fp16 = const()[name = tensor("out_norm_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(152179904)))]; + tensor out_norm_bias_to_fp16 = const()[name = tensor("out_norm_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(152182016)))]; + tensor var_2376_to_fp16 = const()[name = tensor("op_2376_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_cast_fp16 = layer_norm(axes = input_axes_0, beta = out_norm_bias_to_fp16, epsilon = var_2376_to_fp16, gamma = out_norm_weight_to_fp16, x = input_75_cast_fp16)[name = tensor("input_cast_fp16")]; + tensor input_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("input_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor out_eos_weight_to_fp16 = const()[name = tensor("out_eos_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(152184128)))]; + tensor out_eos_bias_to_fp16 = const()[name = tensor("out_eos_bias_to_fp16"), val = tensor([-0x1.5ep-3])]; + tensor linear_25_cast_fp16 = linear(bias = out_eos_bias_to_fp16, weight = out_eos_weight_to_fp16, x = input_cast_fp16)[name = tensor("linear_25_cast_fp16")]; + tensor linear_25_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("linear_25_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor var_2384 = cast(dtype = linear_25_cast_fp16_to_fp32_dtype_0, x = linear_25_cast_fp16)[name = tensor("cast_108")]; + tensor input = cast(dtype = input_cast_fp16_to_fp32_dtype_0, x = input_cast_fp16)[name = tensor("cast_109")]; + tensor var_2355 = cast(dtype = var_2355_cast_fp16_to_fp32_dtype_0, x = var_2355_cast_fp16)[name = tensor("cast_110")]; + tensor new_cache_internal_tensor_assign_2 = cast(dtype = reshape_59_cast_fp16_to_fp32_dtype_0, x = reshape_59_cast_fp16)[name = tensor("cast_111")]; + tensor var_1969 = cast(dtype = var_1969_cast_fp16_to_fp32_dtype_0, x = var_1969_cast_fp16)[name = tensor("cast_115")]; + tensor new_cache_9_internal_tensor_assign_2 = cast(dtype = reshape_49_cast_fp16_to_fp32_dtype_0, x = reshape_49_cast_fp16)[name = tensor("cast_116")]; + tensor var_1583 = cast(dtype = var_1583_cast_fp16_to_fp32_dtype_0, x = var_1583_cast_fp16)[name = tensor("cast_120")]; + tensor new_cache_7_internal_tensor_assign_2 = cast(dtype = reshape_39_cast_fp16_to_fp32_dtype_0, x = reshape_39_cast_fp16)[name = tensor("cast_121")]; + tensor var_1197 = cast(dtype = var_1197_cast_fp16_to_fp32_dtype_0, x = var_1197_cast_fp16)[name = tensor("cast_125")]; + tensor new_cache_5_internal_tensor_assign_2 = cast(dtype = reshape_29_cast_fp16_to_fp32_dtype_0, x = reshape_29_cast_fp16)[name = tensor("cast_126")]; + tensor var_811 = cast(dtype = var_811_cast_fp16_to_fp32_dtype_0, x = var_811_cast_fp16)[name = tensor("cast_130")]; + tensor new_cache_3_internal_tensor_assign_2 = cast(dtype = reshape_19_cast_fp16_to_fp32_dtype_0, x = reshape_19_cast_fp16)[name = tensor("cast_131")]; + tensor var_425 = cast(dtype = var_425_cast_fp16_to_fp32_dtype_0, x = var_425_cast_fp16)[name = tensor("cast_135")]; + tensor new_cache_1_internal_tensor_assign_2 = cast(dtype = reshape_9_cast_fp16_to_fp32_dtype_0, x = reshape_9_cast_fp16)[name = tensor("cast_136")]; + } -> (input, var_2384, new_cache_1_internal_tensor_assign_2, var_425, new_cache_3_internal_tensor_assign_2, var_811, new_cache_5_internal_tensor_assign_2, var_1197, new_cache_7_internal_tensor_assign_2, var_1583, new_cache_9_internal_tensor_assign_2, var_1969, new_cache_internal_tensor_assign_2, var_2355); +} \ No newline at end of file diff --git a/v2.1/spanish/flowlm_step.mlmodelc/weights/weight.bin b/v2.1/spanish/flowlm_step.mlmodelc/weights/weight.bin new file mode 100644 index 0000000000000000000000000000000000000000..80b5486ff7d3e5a00f09d1bdd1fd68064920b670 --- /dev/null +++ b/v2.1/spanish/flowlm_step.mlmodelc/weights/weight.bin @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:4db061b3a9791e49fbf991441ea9571b876dd3bb22ddcb896723b1ebe78257a7 +size 152186240 diff --git a/v2.1/spanish/flowlm_step.mlpackage/Data/com.apple.CoreML/model.mlmodel b/v2.1/spanish/flowlm_step.mlpackage/Data/com.apple.CoreML/model.mlmodel new file mode 100644 index 0000000000000000000000000000000000000000..613622a63125cd836c2f102786bae46dc63a1a7c --- /dev/null +++ b/v2.1/spanish/flowlm_step.mlpackage/Data/com.apple.CoreML/model.mlmodel @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:2ee44328cb68ff655e8c30557ba798c3af2d63b9f8bb77c36060768f9bde9963 +size 207618 diff --git a/v2.1/spanish/flowlm_step.mlpackage/Data/com.apple.CoreML/weights/weight.bin b/v2.1/spanish/flowlm_step.mlpackage/Data/com.apple.CoreML/weights/weight.bin new file mode 100644 index 0000000000000000000000000000000000000000..80b5486ff7d3e5a00f09d1bdd1fd68064920b670 --- /dev/null +++ b/v2.1/spanish/flowlm_step.mlpackage/Data/com.apple.CoreML/weights/weight.bin @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:4db061b3a9791e49fbf991441ea9571b876dd3bb22ddcb896723b1ebe78257a7 +size 152186240 diff --git a/v2.1/spanish/flowlm_step.mlpackage/Manifest.json b/v2.1/spanish/flowlm_step.mlpackage/Manifest.json new file mode 100644 index 0000000000000000000000000000000000000000..d7047cacfc695b17007eb5d3545bd944c197161b --- /dev/null +++ b/v2.1/spanish/flowlm_step.mlpackage/Manifest.json @@ -0,0 +1,18 @@ +{ + "fileFormatVersion": "1.0.0", + "itemInfoEntries": { + "9F0521CD-8B2F-4C2F-A48A-5C3BE7A59725": { + "author": "com.apple.CoreML", + "description": "CoreML Model Weights", + "name": "weights", + "path": "com.apple.CoreML/weights" + }, + "B5972182-D583-43B8-A65A-96607BFE86C1": { + "author": "com.apple.CoreML", + "description": "CoreML Model Specification", + "name": "model.mlmodel", + "path": "com.apple.CoreML/model.mlmodel" + } + }, + "rootModelIdentifier": "B5972182-D583-43B8-A65A-96607BFE86C1" +} diff --git a/v2.1/spanish/flowlm_stepv2.mlmodelc/analytics/coremldata.bin b/v2.1/spanish/flowlm_stepv2.mlmodelc/analytics/coremldata.bin new file mode 100644 index 0000000000000000000000000000000000000000..60caa1f5a0dff80bf8d04c9228249864ccaa8537 --- /dev/null +++ b/v2.1/spanish/flowlm_stepv2.mlmodelc/analytics/coremldata.bin @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:ce87ab1ebddab10c69587ff105c0c93aa21cc674bf01869fd892e1ebf3a2c936 +size 243 diff --git a/v2.1/spanish/flowlm_stepv2.mlmodelc/coremldata.bin b/v2.1/spanish/flowlm_stepv2.mlmodelc/coremldata.bin new file mode 100644 index 0000000000000000000000000000000000000000..f85f00238ad4d275b2d42d263256131ca81a58e1 --- /dev/null +++ b/v2.1/spanish/flowlm_stepv2.mlmodelc/coremldata.bin @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:a28ab65e7aab97a93da35e319d1a05633a939ae92c84b77a5a6727a2f84bab43 +size 1190 diff --git a/v2.1/spanish/flowlm_stepv2.mlmodelc/metadata.json b/v2.1/spanish/flowlm_stepv2.mlmodelc/metadata.json new file mode 100644 index 0000000000000000000000000000000000000000..f2969c83e1cf7a6025c56eb3502b7ec456a89630 --- /dev/null +++ b/v2.1/spanish/flowlm_stepv2.mlmodelc/metadata.json @@ -0,0 +1,348 @@ +[ + { + "metadataOutputVersion" : "3.0", + "storagePrecision" : "Mixed (Float16, Int8)", + "outputSchema" : [ + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1 × 1 × 1024)", + "shortDescription" : "", + "shape" : "[1, 1, 1024]", + "name" : "input", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1 × 1 × 1)", + "shortDescription" : "", + "shape" : "[1, 1, 1]", + "name" : "var_2582", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 2 × 1 × 512 × 16 × 64)", + "shortDescription" : "", + "shape" : "[2, 1, 512, 16, 64]", + "name" : "new_cache_1_internal_tensor_assign_2", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1)", + "shortDescription" : "", + "shape" : "[1]", + "name" : "var_458", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 2 × 1 × 512 × 16 × 64)", + "shortDescription" : "", + "shape" : "[2, 1, 512, 16, 64]", + "name" : "new_cache_3_internal_tensor_assign_2", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1)", + "shortDescription" : "", + "shape" : "[1]", + "name" : "var_877", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 2 × 1 × 512 × 16 × 64)", + "shortDescription" : "", + "shape" : "[2, 1, 512, 16, 64]", + "name" : "new_cache_5_internal_tensor_assign_2", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1)", + "shortDescription" : "", + "shape" : "[1]", + "name" : "var_1296", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 2 × 1 × 512 × 16 × 64)", + "shortDescription" : "", + "shape" : "[2, 1, 512, 16, 64]", + "name" : "new_cache_7_internal_tensor_assign_2", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1)", + "shortDescription" : "", + "shape" : "[1]", + "name" : "var_1715", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 2 × 1 × 512 × 16 × 64)", + "shortDescription" : "", + "shape" : "[2, 1, 512, 16, 64]", + "name" : "new_cache_9_internal_tensor_assign_2", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1)", + "shortDescription" : "", + "shape" : "[1]", + "name" : "var_2134", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 2 × 1 × 512 × 16 × 64)", + "shortDescription" : "", + "shape" : "[2, 1, 512, 16, 64]", + "name" : "new_cache_internal_tensor_assign_2", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1)", + "shortDescription" : "", + "shape" : "[1]", + "name" : "var_2553", + "type" : "MultiArray" + } + ], + "modelParameters" : [ + + ], + "specificationVersion" : 8, + "mlProgramOperationTypeHistogram" : { + "Ios17.cos" : 6, + "Range1d" : 12, + "Ios17.notEqual" : 13, + "Ios17.floor" : 6, + "Ios17.reshape" : 108, + "Ios17.logicalAnd" : 6, + "Ios16.softmax" : 6, + "Ios17.scatter" : 12, + "Ios17.matmul" : 12, + "Ios17.sin" : 6, + "Ios17.transpose" : 24, + "Select" : 19, + "Ios17.expandDims" : 7, + "Ios17.add" : 36, + "Tile" : 6, + "Ios17.sliceByIndex" : 78, + "Ios17.lessEqual" : 6, + "Ios17.scatterAlongAxis" : 12, + "Ios17.logicalNot" : 6, + "Ios17.layerNorm" : 13, + "Ios16.constexprAffineDequantize" : 24, + "Ios17.cast" : 34, + "Ios17.sub" : 18, + "Ios17.less" : 6, + "Ios16.gelu" : 6, + "Ios17.linear" : 26, + "Stack" : 12, + "Ios17.mul" : 72 + }, + "computePrecision" : "Mixed (Float16, Float32, Int32)", + "isUpdatable" : "0", + "stateSchema" : [ + + ], + "availability" : { + "macOS" : "14.0", + "tvOS" : "17.0", + "visionOS" : "1.0", + "watchOS" : "10.0", + "iOS" : "17.0", + "macCatalyst" : "17.0" + }, + "modelType" : { + "name" : "MLModelType_mlProgram" + }, + "userDefinedMetadata" : { + "com.github.apple.coremltools.conversion_date" : "2026-04-27", + "com.github.apple.coremltools.source" : "torch==2.9.1", + "com.github.apple.coremltools.version" : "9.0", + "com.github.apple.coremltools.source_dialect" : "TorchScript" + }, + "inputSchema" : [ + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1 × 1 × 32)", + "shortDescription" : "", + "shape" : "[1, 1, 32]", + "name" : "sequence", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 32)", + "shortDescription" : "", + "shape" : "[32]", + "name" : "bos_emb", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 2 × 1 × 512 × 16 × 64)", + "shortDescription" : "", + "shape" : "[2, 1, 512, 16, 64]", + "name" : "cache0", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1)", + "shortDescription" : "", + "shape" : "[1]", + "name" : "position0", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 2 × 1 × 512 × 16 × 64)", + "shortDescription" : "", + "shape" : "[2, 1, 512, 16, 64]", + "name" : "cache1", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1)", + "shortDescription" : "", + "shape" : "[1]", + "name" : "position1", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 2 × 1 × 512 × 16 × 64)", + "shortDescription" : "", + "shape" : "[2, 1, 512, 16, 64]", + "name" : "cache2", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1)", + "shortDescription" : "", + "shape" : "[1]", + "name" : "position2", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 2 × 1 × 512 × 16 × 64)", + "shortDescription" : "", + "shape" : "[2, 1, 512, 16, 64]", + "name" : "cache3", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1)", + "shortDescription" : "", + "shape" : "[1]", + "name" : "position3", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 2 × 1 × 512 × 16 × 64)", + "shortDescription" : "", + "shape" : "[2, 1, 512, 16, 64]", + "name" : "cache4", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1)", + "shortDescription" : "", + "shape" : "[1]", + "name" : "position4", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 2 × 1 × 512 × 16 × 64)", + "shortDescription" : "", + "shape" : "[2, 1, 512, 16, 64]", + "name" : "cache5", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1)", + "shortDescription" : "", + "shape" : "[1]", + "name" : "position5", + "type" : "MultiArray" + } + ], + "generatedClassName" : "flowlm_stepv2", + "method" : "predict" + } +] \ No newline at end of file diff --git a/v2.1/spanish/flowlm_stepv2.mlmodelc/model.mil b/v2.1/spanish/flowlm_stepv2.mlmodelc/model.mil new file mode 100644 index 0000000000000000000000000000000000000000..5dfb3652b7cbf13e3c282080dc36d539596dee0e --- /dev/null +++ b/v2.1/spanish/flowlm_stepv2.mlmodelc/model.mil @@ -0,0 +1,1359 @@ +program(1.0) +[buildInfo = dict, tensor>({{"coremlc-component-MIL", "3520.4.1"}, {"coremlc-version", "3520.5.1"}, {"coremltools-component-torch", "2.9.1"}, {"coremltools-source-dialect", "TorchScript"}, {"coremltools-version", "9.0"}})] +{ + func main(tensor bos_emb, tensor cache0, tensor cache1, tensor cache2, tensor cache3, tensor cache4, tensor cache5, tensor position0, tensor position1, tensor position2, tensor position3, tensor position4, tensor position5, tensor sequence) { + tensor attn0_in_proj_weight_cast_fp16 = constexpr_affine_dequantize()[axis = tensor(0), name = tensor("attn0_in_proj_weight_cast_fp16"), quantized_data = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(64))), scale = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(3148992))), zero_point = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(3145856)))]; + tensor attn0_out_proj_weight_cast_fp16 = constexpr_affine_dequantize()[axis = tensor(0), name = tensor("attn0_out_proj_weight_cast_fp16"), quantized_data = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(3155200))), scale = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(4204928))), zero_point = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(4203840)))]; + tensor linear0_1_weight_cast_fp16 = constexpr_affine_dequantize()[axis = tensor(0), name = tensor("linear0_1_weight_cast_fp16"), quantized_data = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(4207040))), scale = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8405568))), zero_point = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8401408)))]; + tensor linear0_2_weight_cast_fp16 = constexpr_affine_dequantize()[axis = tensor(0), name = tensor("linear0_2_weight_cast_fp16"), quantized_data = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8413824))), scale = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(12608192))), zero_point = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(4203840)))]; + tensor attn1_in_proj_weight_cast_fp16 = constexpr_affine_dequantize()[axis = tensor(0), name = tensor("attn1_in_proj_weight_cast_fp16"), quantized_data = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(12610304))), scale = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15756096))), zero_point = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(3145856)))]; + tensor attn1_out_proj_weight_cast_fp16 = constexpr_affine_dequantize()[axis = tensor(0), name = tensor("attn1_out_proj_weight_cast_fp16"), quantized_data = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15762304))), scale = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(16810944))), zero_point = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(4203840)))]; + tensor linear1_1_weight_cast_fp16 = constexpr_affine_dequantize()[axis = tensor(0), name = tensor("linear1_1_weight_cast_fp16"), quantized_data = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(16813056))), scale = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21007424))), zero_point = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8401408)))]; + tensor linear1_2_weight_cast_fp16 = constexpr_affine_dequantize()[axis = tensor(0), name = tensor("linear1_2_weight_cast_fp16"), quantized_data = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21015680))), scale = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(25210048))), zero_point = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(4203840)))]; + tensor attn2_in_proj_weight_cast_fp16 = constexpr_affine_dequantize()[axis = tensor(0), name = tensor("attn2_in_proj_weight_cast_fp16"), quantized_data = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(25212160))), scale = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28357952))), zero_point = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(3145856)))]; + tensor attn2_out_proj_weight_cast_fp16 = constexpr_affine_dequantize()[axis = tensor(0), name = tensor("attn2_out_proj_weight_cast_fp16"), quantized_data = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28364160))), scale = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(29412800))), zero_point = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(4203840)))]; + tensor linear2_1_weight_cast_fp16 = constexpr_affine_dequantize()[axis = tensor(0), name = tensor("linear2_1_weight_cast_fp16"), quantized_data = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(29414912))), scale = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(33609280))), zero_point = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8401408)))]; + tensor linear2_2_weight_cast_fp16 = constexpr_affine_dequantize()[axis = tensor(0), name = tensor("linear2_2_weight_cast_fp16"), quantized_data = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(33617536))), scale = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(37811904))), zero_point = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(4203840)))]; + tensor attn3_in_proj_weight_cast_fp16 = constexpr_affine_dequantize()[axis = tensor(0), name = tensor("attn3_in_proj_weight_cast_fp16"), quantized_data = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(37814016))), scale = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(40959808))), zero_point = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(3145856)))]; + tensor attn3_out_proj_weight_cast_fp16 = constexpr_affine_dequantize()[axis = tensor(0), name = tensor("attn3_out_proj_weight_cast_fp16"), quantized_data = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(40966016))), scale = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(42014656))), zero_point = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(4203840)))]; + tensor linear3_1_weight_cast_fp16 = constexpr_affine_dequantize()[axis = tensor(0), name = tensor("linear3_1_weight_cast_fp16"), quantized_data = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(42016768))), scale = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(46211136))), zero_point = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8401408)))]; + tensor linear3_2_weight_cast_fp16 = constexpr_affine_dequantize()[axis = tensor(0), name = tensor("linear3_2_weight_cast_fp16"), quantized_data = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(46219392))), scale = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(50413760))), zero_point = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(4203840)))]; + tensor attn4_in_proj_weight_cast_fp16 = constexpr_affine_dequantize()[axis = tensor(0), name = tensor("attn4_in_proj_weight_cast_fp16"), quantized_data = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(50415872))), scale = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(53561664))), zero_point = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(3145856)))]; + tensor attn4_out_proj_weight_cast_fp16 = constexpr_affine_dequantize()[axis = tensor(0), name = tensor("attn4_out_proj_weight_cast_fp16"), quantized_data = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(53567872))), scale = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(54616512))), zero_point = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(4203840)))]; + tensor linear4_1_weight_cast_fp16 = constexpr_affine_dequantize()[axis = tensor(0), name = tensor("linear4_1_weight_cast_fp16"), quantized_data = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(54618624))), scale = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(58812992))), zero_point = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8401408)))]; + tensor linear4_2_weight_cast_fp16 = constexpr_affine_dequantize()[axis = tensor(0), name = tensor("linear4_2_weight_cast_fp16"), quantized_data = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(58821248))), scale = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(63015616))), zero_point = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(4203840)))]; + tensor attn5_in_proj_weight_cast_fp16 = constexpr_affine_dequantize()[axis = tensor(0), name = tensor("attn5_in_proj_weight_cast_fp16"), quantized_data = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(63017728))), scale = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(66163520))), zero_point = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(3145856)))]; + tensor attn5_out_proj_weight_cast_fp16 = constexpr_affine_dequantize()[axis = tensor(0), name = tensor("attn5_out_proj_weight_cast_fp16"), quantized_data = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(66169728))), scale = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(67218368))), zero_point = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(4203840)))]; + tensor linear5_1_weight_cast_fp16 = constexpr_affine_dequantize()[axis = tensor(0), name = tensor("linear5_1_weight_cast_fp16"), quantized_data = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(67220480))), scale = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(71414848))), zero_point = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8401408)))]; + tensor linear5_2_weight_cast_fp16 = constexpr_affine_dequantize()[axis = tensor(0), name = tensor("linear5_2_weight_cast_fp16"), quantized_data = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(71423104))), scale = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(75617472))), zero_point = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(4203840)))]; + tensor sequence_to_fp16_dtype_0 = const()[name = tensor("sequence_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor sequence_to_fp16 = cast(dtype = sequence_to_fp16_dtype_0, x = sequence)[name = tensor("cast_141")]; + tensor var_54_cast_fp16 = not_equal(x = sequence_to_fp16, y = sequence_to_fp16)[name = tensor("op_54_cast_fp16")]; + tensor expand_dims_0_axes_0 = const()[name = tensor("expand_dims_0_axes_0"), val = tensor([0, 1])]; + tensor bos_emb_to_fp16_dtype_0 = const()[name = tensor("bos_emb_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor bos_emb_to_fp16 = cast(dtype = bos_emb_to_fp16_dtype_0, x = bos_emb)[name = tensor("cast_140")]; + tensor expand_dims_0_cast_fp16 = expand_dims(axes = expand_dims_0_axes_0, x = bos_emb_to_fp16)[name = tensor("expand_dims_0_cast_fp16")]; + tensor input_1_cast_fp16 = select(a = expand_dims_0_cast_fp16, b = sequence_to_fp16, cond = var_54_cast_fp16)[name = tensor("input_1_cast_fp16")]; + tensor input_linear_weight_to_fp16 = const()[name = tensor("input_linear_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(75619584)))]; + tensor linear_0_bias_0_to_fp16 = const()[name = tensor("linear_0_bias_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(75685184)))]; + tensor linear_0_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = input_linear_weight_to_fp16, x = input_1_cast_fp16)[name = tensor("linear_0_cast_fp16")]; + tensor x_1_axes_0 = const()[name = tensor("x_1_axes_0"), val = tensor([-1])]; + tensor norm0_1_weight_to_fp16 = const()[name = tensor("norm0_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(75687296)))]; + tensor norm0_1_bias_to_fp16 = const()[name = tensor("norm0_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(75689408)))]; + tensor var_60_to_fp16 = const()[name = tensor("op_60_to_fp16"), val = tensor(0x1.5p-17)]; + tensor x_1_cast_fp16 = layer_norm(axes = x_1_axes_0, beta = norm0_1_bias_to_fp16, epsilon = var_60_to_fp16, gamma = norm0_1_weight_to_fp16, x = linear_0_cast_fp16)[name = tensor("x_1_cast_fp16")]; + tensor linear_1_bias_0_to_fp16 = const()[name = tensor("linear_1_bias_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(75691520)))]; + tensor linear_1_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = attn0_in_proj_weight_cast_fp16, x = x_1_cast_fp16)[name = tensor("linear_1_cast_fp16")]; + tensor var_96 = const()[name = tensor("op_96"), val = tensor([1, 1, 3, 16, 64])]; + tensor qkv_1_cast_fp16 = reshape(shape = var_96, x = linear_1_cast_fp16)[name = tensor("qkv_1_cast_fp16")]; + tensor q_1_begin_0 = const()[name = tensor("q_1_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor q_1_end_0 = const()[name = tensor("q_1_end_0"), val = tensor([1, 1, 1, 16, 64])]; + tensor q_1_end_mask_0 = const()[name = tensor("q_1_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor q_1_squeeze_mask_0 = const()[name = tensor("q_1_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor q_1_cast_fp16 = slice_by_index(begin = q_1_begin_0, end = q_1_end_0, end_mask = q_1_end_mask_0, squeeze_mask = q_1_squeeze_mask_0, x = qkv_1_cast_fp16)[name = tensor("q_1_cast_fp16")]; + tensor k_1_begin_0 = const()[name = tensor("k_1_begin_0"), val = tensor([0, 0, 1, 0, 0])]; + tensor k_1_end_0 = const()[name = tensor("k_1_end_0"), val = tensor([1, 1, 2, 16, 64])]; + tensor k_1_end_mask_0 = const()[name = tensor("k_1_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor k_1_squeeze_mask_0 = const()[name = tensor("k_1_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor k_1_cast_fp16 = slice_by_index(begin = k_1_begin_0, end = k_1_end_0, end_mask = k_1_end_mask_0, squeeze_mask = k_1_squeeze_mask_0, x = qkv_1_cast_fp16)[name = tensor("k_1_cast_fp16")]; + tensor v_1_begin_0 = const()[name = tensor("v_1_begin_0"), val = tensor([0, 0, 2, 0, 0])]; + tensor v_1_end_0 = const()[name = tensor("v_1_end_0"), val = tensor([1, 1, 3, 16, 64])]; + tensor v_1_end_mask_0 = const()[name = tensor("v_1_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor v_1_squeeze_mask_0 = const()[name = tensor("v_1_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor v_1_cast_fp16 = slice_by_index(begin = v_1_begin_0, end = v_1_end_0, end_mask = v_1_end_mask_0, squeeze_mask = v_1_squeeze_mask_0, x = qkv_1_cast_fp16)[name = tensor("v_1_cast_fp16")]; + tensor var_200 = const()[name = tensor("op_200"), val = tensor([1, 1, 1, 1])]; + tensor position0_to_fp16_dtype_0 = const()[name = tensor("position0_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor position0_to_fp16 = cast(dtype = position0_to_fp16_dtype_0, x = position0)[name = tensor("cast_139")]; + tensor ts_5_cast_fp16 = reshape(shape = var_200, x = position0_to_fp16)[name = tensor("ts_5_cast_fp16")]; + tensor var_204 = const()[name = tensor("op_204"), val = tensor([1, 1, 16, 32, 2])]; + tensor q_complex_1_cast_fp16 = reshape(shape = var_204, x = q_1_cast_fp16)[name = tensor("q_complex_1_cast_fp16")]; + tensor var_208 = const()[name = tensor("op_208"), val = tensor([1, 1, 16, 32, 2])]; + tensor k_complex_1_cast_fp16 = reshape(shape = var_208, x = k_1_cast_fp16)[name = tensor("k_complex_1_cast_fp16")]; + tensor var_212_begin_0 = const()[name = tensor("op_212_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_212_end_0 = const()[name = tensor("op_212_end_0"), val = tensor([1, 1, 16, 32, 1])]; + tensor var_212_end_mask_0 = const()[name = tensor("op_212_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_212_squeeze_mask_0 = const()[name = tensor("op_212_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_212_cast_fp16 = slice_by_index(begin = var_212_begin_0, end = var_212_end_0, end_mask = var_212_end_mask_0, squeeze_mask = var_212_squeeze_mask_0, x = q_complex_1_cast_fp16)[name = tensor("op_212_cast_fp16")]; + tensor var_220_begin_0 = const()[name = tensor("op_220_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_220_end_0 = const()[name = tensor("op_220_end_0"), val = tensor([1, 1, 16, 32, 2])]; + tensor var_220_end_mask_0 = const()[name = tensor("op_220_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_220_squeeze_mask_0 = const()[name = tensor("op_220_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_220_cast_fp16 = slice_by_index(begin = var_220_begin_0, end = var_220_end_0, end_mask = var_220_end_mask_0, squeeze_mask = var_220_squeeze_mask_0, x = q_complex_1_cast_fp16)[name = tensor("op_220_cast_fp16")]; + tensor var_228_begin_0 = const()[name = tensor("op_228_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_228_end_0 = const()[name = tensor("op_228_end_0"), val = tensor([1, 1, 16, 32, 1])]; + tensor var_228_end_mask_0 = const()[name = tensor("op_228_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_228_squeeze_mask_0 = const()[name = tensor("op_228_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_228_cast_fp16 = slice_by_index(begin = var_228_begin_0, end = var_228_end_0, end_mask = var_228_end_mask_0, squeeze_mask = var_228_squeeze_mask_0, x = k_complex_1_cast_fp16)[name = tensor("op_228_cast_fp16")]; + tensor var_236_begin_0 = const()[name = tensor("op_236_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_236_end_0 = const()[name = tensor("op_236_end_0"), val = tensor([1, 1, 16, 32, 2])]; + tensor var_236_end_mask_0 = const()[name = tensor("op_236_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_236_squeeze_mask_0 = const()[name = tensor("op_236_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_236_cast_fp16 = slice_by_index(begin = var_236_begin_0, end = var_236_end_0, end_mask = var_236_end_mask_0, squeeze_mask = var_236_squeeze_mask_0, x = k_complex_1_cast_fp16)[name = tensor("op_236_cast_fp16")]; + tensor freqs_1_to_fp16 = const()[name = tensor("freqs_1_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(75697728)))]; + tensor var_242_cast_fp16 = mul(x = freqs_1_to_fp16, y = ts_5_cast_fp16)[name = tensor("op_242_cast_fp16")]; + tensor rotr_1_cast_fp16 = cos(x = var_242_cast_fp16)[name = tensor("rotr_1_cast_fp16")]; + tensor roti_1_cast_fp16 = sin(x = var_242_cast_fp16)[name = tensor("roti_1_cast_fp16")]; + tensor var_246_cast_fp16 = mul(x = var_212_cast_fp16, y = rotr_1_cast_fp16)[name = tensor("op_246_cast_fp16")]; + tensor var_247_cast_fp16 = mul(x = var_220_cast_fp16, y = roti_1_cast_fp16)[name = tensor("op_247_cast_fp16")]; + tensor qor_1_cast_fp16 = sub(x = var_246_cast_fp16, y = var_247_cast_fp16)[name = tensor("qor_1_cast_fp16")]; + tensor var_250_cast_fp16 = mul(x = var_212_cast_fp16, y = roti_1_cast_fp16)[name = tensor("op_250_cast_fp16")]; + tensor var_251_cast_fp16 = mul(x = var_220_cast_fp16, y = rotr_1_cast_fp16)[name = tensor("op_251_cast_fp16")]; + tensor qoi_1_cast_fp16 = add(x = var_250_cast_fp16, y = var_251_cast_fp16)[name = tensor("qoi_1_cast_fp16")]; + tensor var_254_cast_fp16 = mul(x = var_228_cast_fp16, y = rotr_1_cast_fp16)[name = tensor("op_254_cast_fp16")]; + tensor var_255_cast_fp16 = mul(x = var_236_cast_fp16, y = roti_1_cast_fp16)[name = tensor("op_255_cast_fp16")]; + tensor kor_1_cast_fp16 = sub(x = var_254_cast_fp16, y = var_255_cast_fp16)[name = tensor("kor_1_cast_fp16")]; + tensor var_258_cast_fp16 = mul(x = var_228_cast_fp16, y = roti_1_cast_fp16)[name = tensor("op_258_cast_fp16")]; + tensor var_259_cast_fp16 = mul(x = var_236_cast_fp16, y = rotr_1_cast_fp16)[name = tensor("op_259_cast_fp16")]; + tensor koi_1_cast_fp16 = add(x = var_258_cast_fp16, y = var_259_cast_fp16)[name = tensor("koi_1_cast_fp16")]; + tensor qo_1_axis_0 = const()[name = tensor("qo_1_axis_0"), val = tensor(-1)]; + tensor qo_1_cast_fp16 = stack(axis = qo_1_axis_0, values = (qor_1_cast_fp16, qoi_1_cast_fp16))[name = tensor("qo_1_cast_fp16")]; + tensor ko_1_axis_0 = const()[name = tensor("ko_1_axis_0"), val = tensor(-1)]; + tensor ko_1_cast_fp16 = stack(axis = ko_1_axis_0, values = (kor_1_cast_fp16, koi_1_cast_fp16))[name = tensor("ko_1_cast_fp16")]; + tensor var_288 = const()[name = tensor("op_288"), val = tensor([1, 1, 16, 64])]; + tensor q_3_cast_fp16 = reshape(shape = var_288, x = qo_1_cast_fp16)[name = tensor("q_3_cast_fp16")]; + tensor var_290 = const()[name = tensor("op_290"), val = tensor([1, 1, 16, 64])]; + tensor k_3_cast_fp16 = reshape(shape = var_290, x = ko_1_cast_fp16)[name = tensor("k_3_cast_fp16")]; + tensor _inversed_312_y_0_to_fp16 = const()[name = tensor("_inversed_312_y_0_to_fp16"), val = tensor(0x1p-9)]; + tensor _inversed_312_cast_fp16 = mul(x = ts_5_cast_fp16, y = _inversed_312_y_0_to_fp16)[name = tensor("_inversed_312_cast_fp16")]; + tensor var_313_cast_fp16 = floor(x = _inversed_312_cast_fp16)[name = tensor("op_313_cast_fp16")]; + tensor var_314_to_fp16 = const()[name = tensor("op_314_to_fp16"), val = tensor(0x1p+9)]; + tensor var_315_cast_fp16 = mul(x = var_313_cast_fp16, y = var_314_to_fp16)[name = tensor("op_315_cast_fp16")]; + tensor write_indices_float_3_cast_fp16 = sub(x = ts_5_cast_fp16, y = var_315_cast_fp16)[name = tensor("write_indices_float_3_cast_fp16")]; + tensor var_322_dtype_0 = const()[name = tensor("op_322_dtype_0"), val = tensor("int32")]; + tensor write_indices_1_reps_0 = const()[name = tensor("write_indices_1_reps_0"), val = tensor([1, 1, 16, 64])]; + tensor write_indices_float_3_cast_fp16_to_int32 = cast(dtype = var_322_dtype_0, x = write_indices_float_3_cast_fp16)[name = tensor("cast_138")]; + tensor write_indices_1 = tile(reps = write_indices_1_reps_0, x = write_indices_float_3_cast_fp16_to_int32)[name = tensor("write_indices_1")]; + tensor var_330_begin_0 = const()[name = tensor("op_330_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_330_end_0 = const()[name = tensor("op_330_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor var_330_end_mask_0 = const()[name = tensor("op_330_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_330_squeeze_mask_0 = const()[name = tensor("op_330_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor cache0_to_fp16_dtype_0 = const()[name = tensor("cache0_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor cache0_to_fp16 = cast(dtype = cache0_to_fp16_dtype_0, x = cache0)[name = tensor("cast_137")]; + tensor var_330_cast_fp16 = slice_by_index(begin = var_330_begin_0, end = var_330_end_0, end_mask = var_330_end_mask_0, squeeze_mask = var_330_squeeze_mask_0, x = cache0_to_fp16)[name = tensor("op_330_cast_fp16")]; + tensor var_332_axis_0 = const()[name = tensor("op_332_axis_0"), val = tensor(1)]; + tensor var_332_mode_0 = const()[name = tensor("op_332_mode_0"), val = tensor("update")]; + tensor var_332_validate_indices_0 = const()[name = tensor("op_332_validate_indices_0"), val = tensor(false)]; + tensor var_332_cast_fp16 = scatter_along_axis(axis = var_332_axis_0, data = var_330_cast_fp16, indices = write_indices_1, mode = var_332_mode_0, updates = k_3_cast_fp16, validate_indices = var_332_validate_indices_0)[name = tensor("op_332_cast_fp16")]; + tensor concat_2 = const()[name = tensor("concat_2"), val = tensor([0, 0, 0, 0, 0])]; + tensor concat_3 = const()[name = tensor("concat_3"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_1_internal_tensor_assign_1_stride_0 = const()[name = tensor("new_cache_1_internal_tensor_assign_1_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_1_internal_tensor_assign_1_begin_mask_0 = const()[name = tensor("new_cache_1_internal_tensor_assign_1_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_1_internal_tensor_assign_1_end_mask_0 = const()[name = tensor("new_cache_1_internal_tensor_assign_1_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_1_internal_tensor_assign_1_squeeze_mask_0 = const()[name = tensor("new_cache_1_internal_tensor_assign_1_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_12 = const()[name = tensor("shape_12"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_0 = const()[name = tensor("reduce_prod_0"), val = tensor(1048576)]; + tensor range_1d_0_start_0 = const()[name = tensor("range_1d_0_start_0"), val = tensor(0)]; + tensor range_1d_0_step_0 = const()[name = tensor("range_1d_0_step_0"), val = tensor(1)]; + tensor range_1d_0 = range_1d(end = reduce_prod_0, start = range_1d_0_start_0, step = range_1d_0_step_0)[name = tensor("range_1d_0")]; + tensor reshape_0 = reshape(shape = shape_12, x = range_1d_0)[name = tensor("reshape_0")]; + tensor slice_by_index_0 = slice_by_index(begin = concat_2, begin_mask = new_cache_1_internal_tensor_assign_1_begin_mask_0, end = concat_3, end_mask = new_cache_1_internal_tensor_assign_1_end_mask_0, squeeze_mask = new_cache_1_internal_tensor_assign_1_squeeze_mask_0, stride = new_cache_1_internal_tensor_assign_1_stride_0, x = reshape_0)[name = tensor("slice_by_index_0")]; + tensor reshape_1_shape_0 = const()[name = tensor("reshape_1_shape_0"), val = tensor([-1])]; + tensor reshape_1 = reshape(shape = reshape_1_shape_0, x = slice_by_index_0)[name = tensor("reshape_1")]; + tensor reshape_2_shape_0 = const()[name = tensor("reshape_2_shape_0"), val = tensor([-1])]; + tensor reshape_2_cast_fp16 = reshape(shape = reshape_2_shape_0, x = var_332_cast_fp16)[name = tensor("reshape_2_cast_fp16")]; + tensor reshape_3_shape_0 = const()[name = tensor("reshape_3_shape_0"), val = tensor([-1])]; + tensor reshape_3_cast_fp16 = reshape(shape = reshape_3_shape_0, x = cache0_to_fp16)[name = tensor("reshape_3_cast_fp16")]; + tensor scatter_0_mode_0 = const()[name = tensor("scatter_0_mode_0"), val = tensor("update")]; + tensor scatter_0_axis_0 = const()[name = tensor("scatter_0_axis_0"), val = tensor(0)]; + tensor scatter_0_validate_indices_0 = const()[name = tensor("scatter_0_validate_indices_0"), val = tensor(false)]; + tensor scatter_0_cast_fp16 = scatter(axis = scatter_0_axis_0, data = reshape_3_cast_fp16, indices = reshape_1, mode = scatter_0_mode_0, updates = reshape_2_cast_fp16, validate_indices = scatter_0_validate_indices_0)[name = tensor("scatter_0_cast_fp16")]; + tensor reshape_4_cast_fp16 = reshape(shape = shape_12, x = scatter_0_cast_fp16)[name = tensor("reshape_4_cast_fp16")]; + tensor var_340_begin_0 = const()[name = tensor("op_340_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor var_340_end_0 = const()[name = tensor("op_340_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor var_340_end_mask_0 = const()[name = tensor("op_340_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_340_squeeze_mask_0 = const()[name = tensor("op_340_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor var_340_cast_fp16 = slice_by_index(begin = var_340_begin_0, end = var_340_end_0, end_mask = var_340_end_mask_0, squeeze_mask = var_340_squeeze_mask_0, x = reshape_4_cast_fp16)[name = tensor("op_340_cast_fp16")]; + tensor var_342_axis_0 = const()[name = tensor("op_342_axis_0"), val = tensor(1)]; + tensor var_342_mode_0 = const()[name = tensor("op_342_mode_0"), val = tensor("update")]; + tensor var_342_validate_indices_0 = const()[name = tensor("op_342_validate_indices_0"), val = tensor(false)]; + tensor var_342_cast_fp16 = scatter_along_axis(axis = var_342_axis_0, data = var_340_cast_fp16, indices = write_indices_1, mode = var_342_mode_0, updates = v_1_cast_fp16, validate_indices = var_342_validate_indices_0)[name = tensor("op_342_cast_fp16")]; + tensor concat_4 = const()[name = tensor("concat_4"), val = tensor([1, 0, 0, 0, 0])]; + tensor concat_5 = const()[name = tensor("concat_5"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_1_internal_tensor_assign_2_stride_0 = const()[name = tensor("new_cache_1_internal_tensor_assign_2_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_1_internal_tensor_assign_2_begin_mask_0 = const()[name = tensor("new_cache_1_internal_tensor_assign_2_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_1_internal_tensor_assign_2_end_mask_0 = const()[name = tensor("new_cache_1_internal_tensor_assign_2_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_1_internal_tensor_assign_2_squeeze_mask_0 = const()[name = tensor("new_cache_1_internal_tensor_assign_2_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_13 = const()[name = tensor("shape_13"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_1 = const()[name = tensor("reduce_prod_1"), val = tensor(1048576)]; + tensor range_1d_1_start_0 = const()[name = tensor("range_1d_1_start_0"), val = tensor(0)]; + tensor range_1d_1_step_0 = const()[name = tensor("range_1d_1_step_0"), val = tensor(1)]; + tensor range_1d_1 = range_1d(end = reduce_prod_1, start = range_1d_1_start_0, step = range_1d_1_step_0)[name = tensor("range_1d_1")]; + tensor reshape_5 = reshape(shape = shape_13, x = range_1d_1)[name = tensor("reshape_5")]; + tensor slice_by_index_1 = slice_by_index(begin = concat_4, begin_mask = new_cache_1_internal_tensor_assign_2_begin_mask_0, end = concat_5, end_mask = new_cache_1_internal_tensor_assign_2_end_mask_0, squeeze_mask = new_cache_1_internal_tensor_assign_2_squeeze_mask_0, stride = new_cache_1_internal_tensor_assign_2_stride_0, x = reshape_5)[name = tensor("slice_by_index_1")]; + tensor reshape_6_shape_0 = const()[name = tensor("reshape_6_shape_0"), val = tensor([-1])]; + tensor reshape_6 = reshape(shape = reshape_6_shape_0, x = slice_by_index_1)[name = tensor("reshape_6")]; + tensor reshape_7_shape_0 = const()[name = tensor("reshape_7_shape_0"), val = tensor([-1])]; + tensor reshape_7_cast_fp16 = reshape(shape = reshape_7_shape_0, x = var_342_cast_fp16)[name = tensor("reshape_7_cast_fp16")]; + tensor reshape_8_shape_0 = const()[name = tensor("reshape_8_shape_0"), val = tensor([-1])]; + tensor reshape_8_cast_fp16 = reshape(shape = reshape_8_shape_0, x = reshape_4_cast_fp16)[name = tensor("reshape_8_cast_fp16")]; + tensor scatter_1_mode_0 = const()[name = tensor("scatter_1_mode_0"), val = tensor("update")]; + tensor scatter_1_axis_0 = const()[name = tensor("scatter_1_axis_0"), val = tensor(0)]; + tensor scatter_1_validate_indices_0 = const()[name = tensor("scatter_1_validate_indices_0"), val = tensor(false)]; + tensor scatter_1_cast_fp16 = scatter(axis = scatter_1_axis_0, data = reshape_8_cast_fp16, indices = reshape_6, mode = scatter_1_mode_0, updates = reshape_7_cast_fp16, validate_indices = scatter_1_validate_indices_0)[name = tensor("scatter_1_cast_fp16")]; + tensor reshape_9_cast_fp16 = reshape(shape = shape_13, x = scatter_1_cast_fp16)[name = tensor("reshape_9_cast_fp16")]; + tensor reshape_9_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("reshape_9_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor keys_1_begin_0 = const()[name = tensor("keys_1_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor keys_1_end_0 = const()[name = tensor("keys_1_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor keys_1_end_mask_0 = const()[name = tensor("keys_1_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor keys_1_squeeze_mask_0 = const()[name = tensor("keys_1_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor keys_1_cast_fp16 = slice_by_index(begin = keys_1_begin_0, end = keys_1_end_0, end_mask = keys_1_end_mask_0, squeeze_mask = keys_1_squeeze_mask_0, x = reshape_9_cast_fp16)[name = tensor("keys_1_cast_fp16")]; + tensor values_1_begin_0 = const()[name = tensor("values_1_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor values_1_end_0 = const()[name = tensor("values_1_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor values_1_end_mask_0 = const()[name = tensor("values_1_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor values_1_squeeze_mask_0 = const()[name = tensor("values_1_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor values_1_cast_fp16 = slice_by_index(begin = values_1_begin_0, end = values_1_end_0, end_mask = values_1_end_mask_0, squeeze_mask = values_1_squeeze_mask_0, x = reshape_9_cast_fp16)[name = tensor("values_1_cast_fp16")]; + tensor var_354_cast_fp16 = not_equal(x = keys_1_cast_fp16, y = keys_1_cast_fp16)[name = tensor("op_354_cast_fp16")]; + tensor var_360_to_fp16 = const()[name = tensor("op_360_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(75697856)))]; + tensor keys_3_cast_fp16 = select(a = var_360_to_fp16, b = keys_1_cast_fp16, cond = var_354_cast_fp16)[name = tensor("keys_3_cast_fp16")]; + tensor var_362_cast_fp16 = not_equal(x = values_1_cast_fp16, y = values_1_cast_fp16)[name = tensor("op_362_cast_fp16")]; + tensor values_3_cast_fp16 = select(a = var_360_to_fp16, b = values_1_cast_fp16, cond = var_362_cast_fp16)[name = tensor("values_3_cast_fp16")]; + tensor var_386 = const()[name = tensor("op_386"), val = tensor([0, 2, 1, 3])]; + tensor var_399 = const()[name = tensor("op_399"), val = tensor([1, 1, 1])]; + tensor var_400_cast_fp16 = reshape(shape = var_399, x = position0_to_fp16)[name = tensor("op_400_cast_fp16")]; + tensor var_417_to_fp16 = const()[name = tensor("op_417_to_fp16"), val = tensor(0x1p+0)]; + tensor valid_len_1_cast_fp16 = add(x = var_400_cast_fp16, y = var_417_to_fp16)[name = tensor("valid_len_1_cast_fp16")]; + tensor k_positions_1_promoted_to_fp16 = const()[name = tensor("k_positions_1_promoted_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(76746496)))]; + tensor valid_mask_1_cast_fp16 = less(x = k_positions_1_promoted_to_fp16, y = valid_len_1_cast_fp16)[name = tensor("valid_mask_1_cast_fp16")]; + tensor causal_mask_1_cast_fp16 = less_equal(x = k_positions_1_promoted_to_fp16, y = var_400_cast_fp16)[name = tensor("causal_mask_1_cast_fp16")]; + tensor attn_mask_1 = logical_and(x = valid_mask_1_cast_fp16, y = causal_mask_1_cast_fp16)[name = tensor("attn_mask_1")]; + tensor attn_mask_3_axes_0 = const()[name = tensor("attn_mask_3_axes_0"), val = tensor([1])]; + tensor attn_mask_3 = expand_dims(axes = attn_mask_3_axes_0, x = attn_mask_1)[name = tensor("attn_mask_3")]; + tensor var_435_transpose_x_0 = const()[name = tensor("op_435_transpose_x_0"), val = tensor(false)]; + tensor var_435_transpose_y_0 = const()[name = tensor("op_435_transpose_y_0"), val = tensor(false)]; + tensor transpose_18_perm_0 = const()[name = tensor("transpose_18_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_19_perm_0 = const()[name = tensor("transpose_19_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_19 = transpose(perm = transpose_19_perm_0, x = keys_3_cast_fp16)[name = tensor("transpose_51")]; + tensor transpose_18 = transpose(perm = transpose_18_perm_0, x = q_3_cast_fp16)[name = tensor("transpose_52")]; + tensor var_435_cast_fp16 = matmul(transpose_x = var_435_transpose_x_0, transpose_y = var_435_transpose_y_0, x = transpose_18, y = transpose_19)[name = tensor("op_435_cast_fp16")]; + tensor var_429_to_fp16 = const()[name = tensor("op_429_to_fp16"), val = tensor([0x1p-3])]; + tensor attn_weights_1_cast_fp16 = mul(x = var_435_cast_fp16, y = var_429_to_fp16)[name = tensor("attn_weights_1_cast_fp16")]; + tensor var_437 = logical_not(x = attn_mask_3)[name = tensor("op_437")]; + tensor var_438_to_fp16 = const()[name = tensor("op_438_to_fp16"), val = tensor(-inf)]; + tensor attn_weights_3_cast_fp16 = select(a = var_438_to_fp16, b = attn_weights_1_cast_fp16, cond = var_437)[name = tensor("attn_weights_3_cast_fp16")]; + tensor var_440 = const()[name = tensor("op_440"), val = tensor(-1)]; + tensor attn_weights_5_cast_fp16 = softmax(axis = var_440, x = attn_weights_3_cast_fp16)[name = tensor("attn_weights_5_cast_fp16")]; + tensor attn_output_1_transpose_x_0 = const()[name = tensor("attn_output_1_transpose_x_0"), val = tensor(false)]; + tensor attn_output_1_transpose_y_0 = const()[name = tensor("attn_output_1_transpose_y_0"), val = tensor(false)]; + tensor values_5_cast_fp16 = transpose(perm = var_386, x = values_3_cast_fp16)[name = tensor("transpose_53")]; + tensor attn_output_1_cast_fp16 = matmul(transpose_x = attn_output_1_transpose_x_0, transpose_y = attn_output_1_transpose_y_0, x = attn_weights_5_cast_fp16, y = values_5_cast_fp16)[name = tensor("attn_output_1_cast_fp16")]; + tensor var_448 = const()[name = tensor("op_448"), val = tensor([0, 2, 1, 3])]; + tensor var_451 = const()[name = tensor("op_451"), val = tensor([1, 1, 1024])]; + tensor var_449_cast_fp16 = transpose(perm = var_448, x = attn_output_1_cast_fp16)[name = tensor("transpose_50")]; + tensor input_5_cast_fp16 = reshape(shape = var_451, x = var_449_cast_fp16)[name = tensor("input_5_cast_fp16")]; + tensor linear_2_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = attn0_out_proj_weight_cast_fp16, x = input_5_cast_fp16)[name = tensor("linear_2_cast_fp16")]; + tensor var_457_to_fp16 = const()[name = tensor("op_457_to_fp16"), val = tensor(0x1p+0)]; + tensor var_458_cast_fp16 = add(x = position0_to_fp16, y = var_457_to_fp16)[name = tensor("op_458_cast_fp16")]; + tensor var_458_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_458_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor input_7_cast_fp16 = add(x = linear_0_cast_fp16, y = linear_2_cast_fp16)[name = tensor("input_7_cast_fp16")]; + tensor input_9_axes_0 = const()[name = tensor("input_9_axes_0"), val = tensor([-1])]; + tensor norm0_2_weight_to_fp16 = const()[name = tensor("norm0_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(76747584)))]; + tensor norm0_2_bias_to_fp16 = const()[name = tensor("norm0_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(76749696)))]; + tensor var_462_to_fp16 = const()[name = tensor("op_462_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_9_cast_fp16 = layer_norm(axes = input_9_axes_0, beta = norm0_2_bias_to_fp16, epsilon = var_462_to_fp16, gamma = norm0_2_weight_to_fp16, x = input_7_cast_fp16)[name = tensor("input_9_cast_fp16")]; + tensor linear_3_bias_0_to_fp16 = const()[name = tensor("linear_3_bias_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(76751808)))]; + tensor linear_3_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = linear0_1_weight_cast_fp16, x = input_9_cast_fp16)[name = tensor("linear_3_cast_fp16")]; + tensor input_11_mode_0 = const()[name = tensor("input_11_mode_0"), val = tensor("EXACT")]; + tensor input_11_cast_fp16 = gelu(mode = input_11_mode_0, x = linear_3_cast_fp16)[name = tensor("input_11_cast_fp16")]; + tensor linear_4_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = linear0_2_weight_cast_fp16, x = input_11_cast_fp16)[name = tensor("linear_4_cast_fp16")]; + tensor input_13_cast_fp16 = add(x = input_7_cast_fp16, y = linear_4_cast_fp16)[name = tensor("input_13_cast_fp16")]; + tensor x_3_axes_0 = const()[name = tensor("x_3_axes_0"), val = tensor([-1])]; + tensor norm1_1_weight_to_fp16 = const()[name = tensor("norm1_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(76760064)))]; + tensor norm1_1_bias_to_fp16 = const()[name = tensor("norm1_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(76762176)))]; + tensor var_479_to_fp16 = const()[name = tensor("op_479_to_fp16"), val = tensor(0x1.5p-17)]; + tensor x_3_cast_fp16 = layer_norm(axes = x_3_axes_0, beta = norm1_1_bias_to_fp16, epsilon = var_479_to_fp16, gamma = norm1_1_weight_to_fp16, x = input_13_cast_fp16)[name = tensor("x_3_cast_fp16")]; + tensor linear_5_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = attn1_in_proj_weight_cast_fp16, x = x_3_cast_fp16)[name = tensor("linear_5_cast_fp16")]; + tensor var_515 = const()[name = tensor("op_515"), val = tensor([1, 1, 3, 16, 64])]; + tensor qkv_3_cast_fp16 = reshape(shape = var_515, x = linear_5_cast_fp16)[name = tensor("qkv_3_cast_fp16")]; + tensor q_7_begin_0 = const()[name = tensor("q_7_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor q_7_end_0 = const()[name = tensor("q_7_end_0"), val = tensor([1, 1, 1, 16, 64])]; + tensor q_7_end_mask_0 = const()[name = tensor("q_7_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor q_7_squeeze_mask_0 = const()[name = tensor("q_7_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor q_7_cast_fp16 = slice_by_index(begin = q_7_begin_0, end = q_7_end_0, end_mask = q_7_end_mask_0, squeeze_mask = q_7_squeeze_mask_0, x = qkv_3_cast_fp16)[name = tensor("q_7_cast_fp16")]; + tensor k_5_begin_0 = const()[name = tensor("k_5_begin_0"), val = tensor([0, 0, 1, 0, 0])]; + tensor k_5_end_0 = const()[name = tensor("k_5_end_0"), val = tensor([1, 1, 2, 16, 64])]; + tensor k_5_end_mask_0 = const()[name = tensor("k_5_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor k_5_squeeze_mask_0 = const()[name = tensor("k_5_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor k_5_cast_fp16 = slice_by_index(begin = k_5_begin_0, end = k_5_end_0, end_mask = k_5_end_mask_0, squeeze_mask = k_5_squeeze_mask_0, x = qkv_3_cast_fp16)[name = tensor("k_5_cast_fp16")]; + tensor v_3_begin_0 = const()[name = tensor("v_3_begin_0"), val = tensor([0, 0, 2, 0, 0])]; + tensor v_3_end_0 = const()[name = tensor("v_3_end_0"), val = tensor([1, 1, 3, 16, 64])]; + tensor v_3_end_mask_0 = const()[name = tensor("v_3_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor v_3_squeeze_mask_0 = const()[name = tensor("v_3_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor v_3_cast_fp16 = slice_by_index(begin = v_3_begin_0, end = v_3_end_0, end_mask = v_3_end_mask_0, squeeze_mask = v_3_squeeze_mask_0, x = qkv_3_cast_fp16)[name = tensor("v_3_cast_fp16")]; + tensor var_619 = const()[name = tensor("op_619"), val = tensor([1, 1, 1, 1])]; + tensor position1_to_fp16_dtype_0 = const()[name = tensor("position1_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor position1_to_fp16 = cast(dtype = position1_to_fp16_dtype_0, x = position1)[name = tensor("cast_134")]; + tensor ts_11_cast_fp16 = reshape(shape = var_619, x = position1_to_fp16)[name = tensor("ts_11_cast_fp16")]; + tensor var_623 = const()[name = tensor("op_623"), val = tensor([1, 1, 16, 32, 2])]; + tensor q_complex_3_cast_fp16 = reshape(shape = var_623, x = q_7_cast_fp16)[name = tensor("q_complex_3_cast_fp16")]; + tensor var_627 = const()[name = tensor("op_627"), val = tensor([1, 1, 16, 32, 2])]; + tensor k_complex_3_cast_fp16 = reshape(shape = var_627, x = k_5_cast_fp16)[name = tensor("k_complex_3_cast_fp16")]; + tensor var_631_begin_0 = const()[name = tensor("op_631_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_631_end_0 = const()[name = tensor("op_631_end_0"), val = tensor([1, 1, 16, 32, 1])]; + tensor var_631_end_mask_0 = const()[name = tensor("op_631_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_631_squeeze_mask_0 = const()[name = tensor("op_631_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_631_cast_fp16 = slice_by_index(begin = var_631_begin_0, end = var_631_end_0, end_mask = var_631_end_mask_0, squeeze_mask = var_631_squeeze_mask_0, x = q_complex_3_cast_fp16)[name = tensor("op_631_cast_fp16")]; + tensor var_639_begin_0 = const()[name = tensor("op_639_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_639_end_0 = const()[name = tensor("op_639_end_0"), val = tensor([1, 1, 16, 32, 2])]; + tensor var_639_end_mask_0 = const()[name = tensor("op_639_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_639_squeeze_mask_0 = const()[name = tensor("op_639_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_639_cast_fp16 = slice_by_index(begin = var_639_begin_0, end = var_639_end_0, end_mask = var_639_end_mask_0, squeeze_mask = var_639_squeeze_mask_0, x = q_complex_3_cast_fp16)[name = tensor("op_639_cast_fp16")]; + tensor var_647_begin_0 = const()[name = tensor("op_647_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_647_end_0 = const()[name = tensor("op_647_end_0"), val = tensor([1, 1, 16, 32, 1])]; + tensor var_647_end_mask_0 = const()[name = tensor("op_647_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_647_squeeze_mask_0 = const()[name = tensor("op_647_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_647_cast_fp16 = slice_by_index(begin = var_647_begin_0, end = var_647_end_0, end_mask = var_647_end_mask_0, squeeze_mask = var_647_squeeze_mask_0, x = k_complex_3_cast_fp16)[name = tensor("op_647_cast_fp16")]; + tensor var_655_begin_0 = const()[name = tensor("op_655_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_655_end_0 = const()[name = tensor("op_655_end_0"), val = tensor([1, 1, 16, 32, 2])]; + tensor var_655_end_mask_0 = const()[name = tensor("op_655_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_655_squeeze_mask_0 = const()[name = tensor("op_655_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_655_cast_fp16 = slice_by_index(begin = var_655_begin_0, end = var_655_end_0, end_mask = var_655_end_mask_0, squeeze_mask = var_655_squeeze_mask_0, x = k_complex_3_cast_fp16)[name = tensor("op_655_cast_fp16")]; + tensor freqs_3_to_fp16 = const()[name = tensor("freqs_3_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(76764288)))]; + tensor var_661_cast_fp16 = mul(x = freqs_3_to_fp16, y = ts_11_cast_fp16)[name = tensor("op_661_cast_fp16")]; + tensor rotr_3_cast_fp16 = cos(x = var_661_cast_fp16)[name = tensor("rotr_3_cast_fp16")]; + tensor roti_3_cast_fp16 = sin(x = var_661_cast_fp16)[name = tensor("roti_3_cast_fp16")]; + tensor var_665_cast_fp16 = mul(x = var_631_cast_fp16, y = rotr_3_cast_fp16)[name = tensor("op_665_cast_fp16")]; + tensor var_666_cast_fp16 = mul(x = var_639_cast_fp16, y = roti_3_cast_fp16)[name = tensor("op_666_cast_fp16")]; + tensor qor_5_cast_fp16 = sub(x = var_665_cast_fp16, y = var_666_cast_fp16)[name = tensor("qor_5_cast_fp16")]; + tensor var_669_cast_fp16 = mul(x = var_631_cast_fp16, y = roti_3_cast_fp16)[name = tensor("op_669_cast_fp16")]; + tensor var_670_cast_fp16 = mul(x = var_639_cast_fp16, y = rotr_3_cast_fp16)[name = tensor("op_670_cast_fp16")]; + tensor qoi_5_cast_fp16 = add(x = var_669_cast_fp16, y = var_670_cast_fp16)[name = tensor("qoi_5_cast_fp16")]; + tensor var_673_cast_fp16 = mul(x = var_647_cast_fp16, y = rotr_3_cast_fp16)[name = tensor("op_673_cast_fp16")]; + tensor var_674_cast_fp16 = mul(x = var_655_cast_fp16, y = roti_3_cast_fp16)[name = tensor("op_674_cast_fp16")]; + tensor kor_5_cast_fp16 = sub(x = var_673_cast_fp16, y = var_674_cast_fp16)[name = tensor("kor_5_cast_fp16")]; + tensor var_677_cast_fp16 = mul(x = var_647_cast_fp16, y = roti_3_cast_fp16)[name = tensor("op_677_cast_fp16")]; + tensor var_678_cast_fp16 = mul(x = var_655_cast_fp16, y = rotr_3_cast_fp16)[name = tensor("op_678_cast_fp16")]; + tensor koi_5_cast_fp16 = add(x = var_677_cast_fp16, y = var_678_cast_fp16)[name = tensor("koi_5_cast_fp16")]; + tensor qo_3_axis_0 = const()[name = tensor("qo_3_axis_0"), val = tensor(-1)]; + tensor qo_3_cast_fp16 = stack(axis = qo_3_axis_0, values = (qor_5_cast_fp16, qoi_5_cast_fp16))[name = tensor("qo_3_cast_fp16")]; + tensor ko_3_axis_0 = const()[name = tensor("ko_3_axis_0"), val = tensor(-1)]; + tensor ko_3_cast_fp16 = stack(axis = ko_3_axis_0, values = (kor_5_cast_fp16, koi_5_cast_fp16))[name = tensor("ko_3_cast_fp16")]; + tensor var_707 = const()[name = tensor("op_707"), val = tensor([1, 1, 16, 64])]; + tensor q_9_cast_fp16 = reshape(shape = var_707, x = qo_3_cast_fp16)[name = tensor("q_9_cast_fp16")]; + tensor var_709 = const()[name = tensor("op_709"), val = tensor([1, 1, 16, 64])]; + tensor k_7_cast_fp16 = reshape(shape = var_709, x = ko_3_cast_fp16)[name = tensor("k_7_cast_fp16")]; + tensor _inversed_731_y_0_to_fp16 = const()[name = tensor("_inversed_731_y_0_to_fp16"), val = tensor(0x1p-9)]; + tensor _inversed_731_cast_fp16 = mul(x = ts_11_cast_fp16, y = _inversed_731_y_0_to_fp16)[name = tensor("_inversed_731_cast_fp16")]; + tensor var_732_cast_fp16 = floor(x = _inversed_731_cast_fp16)[name = tensor("op_732_cast_fp16")]; + tensor var_733_to_fp16 = const()[name = tensor("op_733_to_fp16"), val = tensor(0x1p+9)]; + tensor var_734_cast_fp16 = mul(x = var_732_cast_fp16, y = var_733_to_fp16)[name = tensor("op_734_cast_fp16")]; + tensor write_indices_float_7_cast_fp16 = sub(x = ts_11_cast_fp16, y = var_734_cast_fp16)[name = tensor("write_indices_float_7_cast_fp16")]; + tensor var_741_dtype_0 = const()[name = tensor("op_741_dtype_0"), val = tensor("int32")]; + tensor write_indices_3_reps_0 = const()[name = tensor("write_indices_3_reps_0"), val = tensor([1, 1, 16, 64])]; + tensor write_indices_float_7_cast_fp16_to_int32 = cast(dtype = var_741_dtype_0, x = write_indices_float_7_cast_fp16)[name = tensor("cast_133")]; + tensor write_indices_3 = tile(reps = write_indices_3_reps_0, x = write_indices_float_7_cast_fp16_to_int32)[name = tensor("write_indices_3")]; + tensor var_749_begin_0 = const()[name = tensor("op_749_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_749_end_0 = const()[name = tensor("op_749_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor var_749_end_mask_0 = const()[name = tensor("op_749_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_749_squeeze_mask_0 = const()[name = tensor("op_749_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor cache1_to_fp16_dtype_0 = const()[name = tensor("cache1_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor cache1_to_fp16 = cast(dtype = cache1_to_fp16_dtype_0, x = cache1)[name = tensor("cast_132")]; + tensor var_749_cast_fp16 = slice_by_index(begin = var_749_begin_0, end = var_749_end_0, end_mask = var_749_end_mask_0, squeeze_mask = var_749_squeeze_mask_0, x = cache1_to_fp16)[name = tensor("op_749_cast_fp16")]; + tensor var_751_axis_0 = const()[name = tensor("op_751_axis_0"), val = tensor(1)]; + tensor var_751_mode_0 = const()[name = tensor("op_751_mode_0"), val = tensor("update")]; + tensor var_751_validate_indices_0 = const()[name = tensor("op_751_validate_indices_0"), val = tensor(false)]; + tensor var_751_cast_fp16 = scatter_along_axis(axis = var_751_axis_0, data = var_749_cast_fp16, indices = write_indices_3, mode = var_751_mode_0, updates = k_7_cast_fp16, validate_indices = var_751_validate_indices_0)[name = tensor("op_751_cast_fp16")]; + tensor concat_9 = const()[name = tensor("concat_9"), val = tensor([0, 0, 0, 0, 0])]; + tensor concat_10 = const()[name = tensor("concat_10"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_3_internal_tensor_assign_1_stride_0 = const()[name = tensor("new_cache_3_internal_tensor_assign_1_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_3_internal_tensor_assign_1_begin_mask_0 = const()[name = tensor("new_cache_3_internal_tensor_assign_1_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_3_internal_tensor_assign_1_end_mask_0 = const()[name = tensor("new_cache_3_internal_tensor_assign_1_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_3_internal_tensor_assign_1_squeeze_mask_0 = const()[name = tensor("new_cache_3_internal_tensor_assign_1_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_14 = const()[name = tensor("shape_14"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_2 = const()[name = tensor("reduce_prod_2"), val = tensor(1048576)]; + tensor range_1d_2_start_0 = const()[name = tensor("range_1d_2_start_0"), val = tensor(0)]; + tensor range_1d_2_step_0 = const()[name = tensor("range_1d_2_step_0"), val = tensor(1)]; + tensor range_1d_2 = range_1d(end = reduce_prod_2, start = range_1d_2_start_0, step = range_1d_2_step_0)[name = tensor("range_1d_2")]; + tensor reshape_10 = reshape(shape = shape_14, x = range_1d_2)[name = tensor("reshape_10")]; + tensor slice_by_index_2 = slice_by_index(begin = concat_9, begin_mask = new_cache_3_internal_tensor_assign_1_begin_mask_0, end = concat_10, end_mask = new_cache_3_internal_tensor_assign_1_end_mask_0, squeeze_mask = new_cache_3_internal_tensor_assign_1_squeeze_mask_0, stride = new_cache_3_internal_tensor_assign_1_stride_0, x = reshape_10)[name = tensor("slice_by_index_2")]; + tensor reshape_11_shape_0 = const()[name = tensor("reshape_11_shape_0"), val = tensor([-1])]; + tensor reshape_11 = reshape(shape = reshape_11_shape_0, x = slice_by_index_2)[name = tensor("reshape_11")]; + tensor reshape_12_shape_0 = const()[name = tensor("reshape_12_shape_0"), val = tensor([-1])]; + tensor reshape_12_cast_fp16 = reshape(shape = reshape_12_shape_0, x = var_751_cast_fp16)[name = tensor("reshape_12_cast_fp16")]; + tensor reshape_13_shape_0 = const()[name = tensor("reshape_13_shape_0"), val = tensor([-1])]; + tensor reshape_13_cast_fp16 = reshape(shape = reshape_13_shape_0, x = cache1_to_fp16)[name = tensor("reshape_13_cast_fp16")]; + tensor scatter_2_mode_0 = const()[name = tensor("scatter_2_mode_0"), val = tensor("update")]; + tensor scatter_2_axis_0 = const()[name = tensor("scatter_2_axis_0"), val = tensor(0)]; + tensor scatter_2_validate_indices_0 = const()[name = tensor("scatter_2_validate_indices_0"), val = tensor(false)]; + tensor scatter_2_cast_fp16 = scatter(axis = scatter_2_axis_0, data = reshape_13_cast_fp16, indices = reshape_11, mode = scatter_2_mode_0, updates = reshape_12_cast_fp16, validate_indices = scatter_2_validate_indices_0)[name = tensor("scatter_2_cast_fp16")]; + tensor reshape_14_cast_fp16 = reshape(shape = shape_14, x = scatter_2_cast_fp16)[name = tensor("reshape_14_cast_fp16")]; + tensor var_759_begin_0 = const()[name = tensor("op_759_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor var_759_end_0 = const()[name = tensor("op_759_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor var_759_end_mask_0 = const()[name = tensor("op_759_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_759_squeeze_mask_0 = const()[name = tensor("op_759_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor var_759_cast_fp16 = slice_by_index(begin = var_759_begin_0, end = var_759_end_0, end_mask = var_759_end_mask_0, squeeze_mask = var_759_squeeze_mask_0, x = reshape_14_cast_fp16)[name = tensor("op_759_cast_fp16")]; + tensor var_761_axis_0 = const()[name = tensor("op_761_axis_0"), val = tensor(1)]; + tensor var_761_mode_0 = const()[name = tensor("op_761_mode_0"), val = tensor("update")]; + tensor var_761_validate_indices_0 = const()[name = tensor("op_761_validate_indices_0"), val = tensor(false)]; + tensor var_761_cast_fp16 = scatter_along_axis(axis = var_761_axis_0, data = var_759_cast_fp16, indices = write_indices_3, mode = var_761_mode_0, updates = v_3_cast_fp16, validate_indices = var_761_validate_indices_0)[name = tensor("op_761_cast_fp16")]; + tensor concat_11 = const()[name = tensor("concat_11"), val = tensor([1, 0, 0, 0, 0])]; + tensor concat_12 = const()[name = tensor("concat_12"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_3_internal_tensor_assign_2_stride_0 = const()[name = tensor("new_cache_3_internal_tensor_assign_2_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_3_internal_tensor_assign_2_begin_mask_0 = const()[name = tensor("new_cache_3_internal_tensor_assign_2_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_3_internal_tensor_assign_2_end_mask_0 = const()[name = tensor("new_cache_3_internal_tensor_assign_2_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_3_internal_tensor_assign_2_squeeze_mask_0 = const()[name = tensor("new_cache_3_internal_tensor_assign_2_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_15 = const()[name = tensor("shape_15"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_3 = const()[name = tensor("reduce_prod_3"), val = tensor(1048576)]; + tensor range_1d_3_start_0 = const()[name = tensor("range_1d_3_start_0"), val = tensor(0)]; + tensor range_1d_3_step_0 = const()[name = tensor("range_1d_3_step_0"), val = tensor(1)]; + tensor range_1d_3 = range_1d(end = reduce_prod_3, start = range_1d_3_start_0, step = range_1d_3_step_0)[name = tensor("range_1d_3")]; + tensor reshape_15 = reshape(shape = shape_15, x = range_1d_3)[name = tensor("reshape_15")]; + tensor slice_by_index_3 = slice_by_index(begin = concat_11, begin_mask = new_cache_3_internal_tensor_assign_2_begin_mask_0, end = concat_12, end_mask = new_cache_3_internal_tensor_assign_2_end_mask_0, squeeze_mask = new_cache_3_internal_tensor_assign_2_squeeze_mask_0, stride = new_cache_3_internal_tensor_assign_2_stride_0, x = reshape_15)[name = tensor("slice_by_index_3")]; + tensor reshape_16_shape_0 = const()[name = tensor("reshape_16_shape_0"), val = tensor([-1])]; + tensor reshape_16 = reshape(shape = reshape_16_shape_0, x = slice_by_index_3)[name = tensor("reshape_16")]; + tensor reshape_17_shape_0 = const()[name = tensor("reshape_17_shape_0"), val = tensor([-1])]; + tensor reshape_17_cast_fp16 = reshape(shape = reshape_17_shape_0, x = var_761_cast_fp16)[name = tensor("reshape_17_cast_fp16")]; + tensor reshape_18_shape_0 = const()[name = tensor("reshape_18_shape_0"), val = tensor([-1])]; + tensor reshape_18_cast_fp16 = reshape(shape = reshape_18_shape_0, x = reshape_14_cast_fp16)[name = tensor("reshape_18_cast_fp16")]; + tensor scatter_3_mode_0 = const()[name = tensor("scatter_3_mode_0"), val = tensor("update")]; + tensor scatter_3_axis_0 = const()[name = tensor("scatter_3_axis_0"), val = tensor(0)]; + tensor scatter_3_validate_indices_0 = const()[name = tensor("scatter_3_validate_indices_0"), val = tensor(false)]; + tensor scatter_3_cast_fp16 = scatter(axis = scatter_3_axis_0, data = reshape_18_cast_fp16, indices = reshape_16, mode = scatter_3_mode_0, updates = reshape_17_cast_fp16, validate_indices = scatter_3_validate_indices_0)[name = tensor("scatter_3_cast_fp16")]; + tensor reshape_19_cast_fp16 = reshape(shape = shape_15, x = scatter_3_cast_fp16)[name = tensor("reshape_19_cast_fp16")]; + tensor reshape_19_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("reshape_19_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor keys_7_begin_0 = const()[name = tensor("keys_7_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor keys_7_end_0 = const()[name = tensor("keys_7_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor keys_7_end_mask_0 = const()[name = tensor("keys_7_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor keys_7_squeeze_mask_0 = const()[name = tensor("keys_7_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor keys_7_cast_fp16 = slice_by_index(begin = keys_7_begin_0, end = keys_7_end_0, end_mask = keys_7_end_mask_0, squeeze_mask = keys_7_squeeze_mask_0, x = reshape_19_cast_fp16)[name = tensor("keys_7_cast_fp16")]; + tensor values_7_begin_0 = const()[name = tensor("values_7_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor values_7_end_0 = const()[name = tensor("values_7_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor values_7_end_mask_0 = const()[name = tensor("values_7_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor values_7_squeeze_mask_0 = const()[name = tensor("values_7_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor values_7_cast_fp16 = slice_by_index(begin = values_7_begin_0, end = values_7_end_0, end_mask = values_7_end_mask_0, squeeze_mask = values_7_squeeze_mask_0, x = reshape_19_cast_fp16)[name = tensor("values_7_cast_fp16")]; + tensor var_773_cast_fp16 = not_equal(x = keys_7_cast_fp16, y = keys_7_cast_fp16)[name = tensor("op_773_cast_fp16")]; + tensor keys_9_cast_fp16 = select(a = var_360_to_fp16, b = keys_7_cast_fp16, cond = var_773_cast_fp16)[name = tensor("keys_9_cast_fp16")]; + tensor var_781_cast_fp16 = not_equal(x = values_7_cast_fp16, y = values_7_cast_fp16)[name = tensor("op_781_cast_fp16")]; + tensor values_9_cast_fp16 = select(a = var_360_to_fp16, b = values_7_cast_fp16, cond = var_781_cast_fp16)[name = tensor("values_9_cast_fp16")]; + tensor var_805 = const()[name = tensor("op_805"), val = tensor([0, 2, 1, 3])]; + tensor var_818 = const()[name = tensor("op_818"), val = tensor([1, 1, 1])]; + tensor var_819_cast_fp16 = reshape(shape = var_818, x = position1_to_fp16)[name = tensor("op_819_cast_fp16")]; + tensor var_836_to_fp16 = const()[name = tensor("op_836_to_fp16"), val = tensor(0x1p+0)]; + tensor valid_len_3_cast_fp16 = add(x = var_819_cast_fp16, y = var_836_to_fp16)[name = tensor("valid_len_3_cast_fp16")]; + tensor valid_mask_3_cast_fp16 = less(x = k_positions_1_promoted_to_fp16, y = valid_len_3_cast_fp16)[name = tensor("valid_mask_3_cast_fp16")]; + tensor causal_mask_3_cast_fp16 = less_equal(x = k_positions_1_promoted_to_fp16, y = var_819_cast_fp16)[name = tensor("causal_mask_3_cast_fp16")]; + tensor attn_mask_5 = logical_and(x = valid_mask_3_cast_fp16, y = causal_mask_3_cast_fp16)[name = tensor("attn_mask_5")]; + tensor attn_mask_7_axes_0 = const()[name = tensor("attn_mask_7_axes_0"), val = tensor([1])]; + tensor attn_mask_7 = expand_dims(axes = attn_mask_7_axes_0, x = attn_mask_5)[name = tensor("attn_mask_7")]; + tensor var_854_transpose_x_0 = const()[name = tensor("op_854_transpose_x_0"), val = tensor(false)]; + tensor var_854_transpose_y_0 = const()[name = tensor("op_854_transpose_y_0"), val = tensor(false)]; + tensor transpose_20_perm_0 = const()[name = tensor("transpose_20_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_21_perm_0 = const()[name = tensor("transpose_21_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_21 = transpose(perm = transpose_21_perm_0, x = keys_9_cast_fp16)[name = tensor("transpose_47")]; + tensor transpose_20 = transpose(perm = transpose_20_perm_0, x = q_9_cast_fp16)[name = tensor("transpose_48")]; + tensor var_854_cast_fp16 = matmul(transpose_x = var_854_transpose_x_0, transpose_y = var_854_transpose_y_0, x = transpose_20, y = transpose_21)[name = tensor("op_854_cast_fp16")]; + tensor var_848_to_fp16 = const()[name = tensor("op_848_to_fp16"), val = tensor([0x1p-3])]; + tensor attn_weights_7_cast_fp16 = mul(x = var_854_cast_fp16, y = var_848_to_fp16)[name = tensor("attn_weights_7_cast_fp16")]; + tensor var_856 = logical_not(x = attn_mask_7)[name = tensor("op_856")]; + tensor var_857_to_fp16 = const()[name = tensor("op_857_to_fp16"), val = tensor(-inf)]; + tensor attn_weights_9_cast_fp16 = select(a = var_857_to_fp16, b = attn_weights_7_cast_fp16, cond = var_856)[name = tensor("attn_weights_9_cast_fp16")]; + tensor var_859 = const()[name = tensor("op_859"), val = tensor(-1)]; + tensor attn_weights_11_cast_fp16 = softmax(axis = var_859, x = attn_weights_9_cast_fp16)[name = tensor("attn_weights_11_cast_fp16")]; + tensor attn_output_3_transpose_x_0 = const()[name = tensor("attn_output_3_transpose_x_0"), val = tensor(false)]; + tensor attn_output_3_transpose_y_0 = const()[name = tensor("attn_output_3_transpose_y_0"), val = tensor(false)]; + tensor values_11_cast_fp16 = transpose(perm = var_805, x = values_9_cast_fp16)[name = tensor("transpose_49")]; + tensor attn_output_3_cast_fp16 = matmul(transpose_x = attn_output_3_transpose_x_0, transpose_y = attn_output_3_transpose_y_0, x = attn_weights_11_cast_fp16, y = values_11_cast_fp16)[name = tensor("attn_output_3_cast_fp16")]; + tensor var_867 = const()[name = tensor("op_867"), val = tensor([0, 2, 1, 3])]; + tensor var_870 = const()[name = tensor("op_870"), val = tensor([1, 1, 1024])]; + tensor var_868_cast_fp16 = transpose(perm = var_867, x = attn_output_3_cast_fp16)[name = tensor("transpose_46")]; + tensor input_15_cast_fp16 = reshape(shape = var_870, x = var_868_cast_fp16)[name = tensor("input_15_cast_fp16")]; + tensor linear_6_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = attn1_out_proj_weight_cast_fp16, x = input_15_cast_fp16)[name = tensor("linear_6_cast_fp16")]; + tensor var_876_to_fp16 = const()[name = tensor("op_876_to_fp16"), val = tensor(0x1p+0)]; + tensor var_877_cast_fp16 = add(x = position1_to_fp16, y = var_876_to_fp16)[name = tensor("op_877_cast_fp16")]; + tensor var_877_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_877_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor input_17_cast_fp16 = add(x = input_13_cast_fp16, y = linear_6_cast_fp16)[name = tensor("input_17_cast_fp16")]; + tensor input_19_axes_0 = const()[name = tensor("input_19_axes_0"), val = tensor([-1])]; + tensor norm1_2_weight_to_fp16 = const()[name = tensor("norm1_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(76764416)))]; + tensor norm1_2_bias_to_fp16 = const()[name = tensor("norm1_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(76766528)))]; + tensor var_881_to_fp16 = const()[name = tensor("op_881_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_19_cast_fp16 = layer_norm(axes = input_19_axes_0, beta = norm1_2_bias_to_fp16, epsilon = var_881_to_fp16, gamma = norm1_2_weight_to_fp16, x = input_17_cast_fp16)[name = tensor("input_19_cast_fp16")]; + tensor linear_7_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = linear1_1_weight_cast_fp16, x = input_19_cast_fp16)[name = tensor("linear_7_cast_fp16")]; + tensor input_21_mode_0 = const()[name = tensor("input_21_mode_0"), val = tensor("EXACT")]; + tensor input_21_cast_fp16 = gelu(mode = input_21_mode_0, x = linear_7_cast_fp16)[name = tensor("input_21_cast_fp16")]; + tensor linear_8_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = linear1_2_weight_cast_fp16, x = input_21_cast_fp16)[name = tensor("linear_8_cast_fp16")]; + tensor input_23_cast_fp16 = add(x = input_17_cast_fp16, y = linear_8_cast_fp16)[name = tensor("input_23_cast_fp16")]; + tensor x_5_axes_0 = const()[name = tensor("x_5_axes_0"), val = tensor([-1])]; + tensor norm2_1_weight_to_fp16 = const()[name = tensor("norm2_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(76768640)))]; + tensor norm2_1_bias_to_fp16 = const()[name = tensor("norm2_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(76770752)))]; + tensor var_898_to_fp16 = const()[name = tensor("op_898_to_fp16"), val = tensor(0x1.5p-17)]; + tensor x_5_cast_fp16 = layer_norm(axes = x_5_axes_0, beta = norm2_1_bias_to_fp16, epsilon = var_898_to_fp16, gamma = norm2_1_weight_to_fp16, x = input_23_cast_fp16)[name = tensor("x_5_cast_fp16")]; + tensor linear_9_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = attn2_in_proj_weight_cast_fp16, x = x_5_cast_fp16)[name = tensor("linear_9_cast_fp16")]; + tensor var_934 = const()[name = tensor("op_934"), val = tensor([1, 1, 3, 16, 64])]; + tensor qkv_5_cast_fp16 = reshape(shape = var_934, x = linear_9_cast_fp16)[name = tensor("qkv_5_cast_fp16")]; + tensor q_13_begin_0 = const()[name = tensor("q_13_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor q_13_end_0 = const()[name = tensor("q_13_end_0"), val = tensor([1, 1, 1, 16, 64])]; + tensor q_13_end_mask_0 = const()[name = tensor("q_13_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor q_13_squeeze_mask_0 = const()[name = tensor("q_13_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor q_13_cast_fp16 = slice_by_index(begin = q_13_begin_0, end = q_13_end_0, end_mask = q_13_end_mask_0, squeeze_mask = q_13_squeeze_mask_0, x = qkv_5_cast_fp16)[name = tensor("q_13_cast_fp16")]; + tensor k_9_begin_0 = const()[name = tensor("k_9_begin_0"), val = tensor([0, 0, 1, 0, 0])]; + tensor k_9_end_0 = const()[name = tensor("k_9_end_0"), val = tensor([1, 1, 2, 16, 64])]; + tensor k_9_end_mask_0 = const()[name = tensor("k_9_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor k_9_squeeze_mask_0 = const()[name = tensor("k_9_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor k_9_cast_fp16 = slice_by_index(begin = k_9_begin_0, end = k_9_end_0, end_mask = k_9_end_mask_0, squeeze_mask = k_9_squeeze_mask_0, x = qkv_5_cast_fp16)[name = tensor("k_9_cast_fp16")]; + tensor v_5_begin_0 = const()[name = tensor("v_5_begin_0"), val = tensor([0, 0, 2, 0, 0])]; + tensor v_5_end_0 = const()[name = tensor("v_5_end_0"), val = tensor([1, 1, 3, 16, 64])]; + tensor v_5_end_mask_0 = const()[name = tensor("v_5_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor v_5_squeeze_mask_0 = const()[name = tensor("v_5_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor v_5_cast_fp16 = slice_by_index(begin = v_5_begin_0, end = v_5_end_0, end_mask = v_5_end_mask_0, squeeze_mask = v_5_squeeze_mask_0, x = qkv_5_cast_fp16)[name = tensor("v_5_cast_fp16")]; + tensor var_1038 = const()[name = tensor("op_1038"), val = tensor([1, 1, 1, 1])]; + tensor position2_to_fp16_dtype_0 = const()[name = tensor("position2_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor position2_to_fp16 = cast(dtype = position2_to_fp16_dtype_0, x = position2)[name = tensor("cast_129")]; + tensor ts_17_cast_fp16 = reshape(shape = var_1038, x = position2_to_fp16)[name = tensor("ts_17_cast_fp16")]; + tensor var_1042 = const()[name = tensor("op_1042"), val = tensor([1, 1, 16, 32, 2])]; + tensor q_complex_5_cast_fp16 = reshape(shape = var_1042, x = q_13_cast_fp16)[name = tensor("q_complex_5_cast_fp16")]; + tensor var_1046 = const()[name = tensor("op_1046"), val = tensor([1, 1, 16, 32, 2])]; + tensor k_complex_5_cast_fp16 = reshape(shape = var_1046, x = k_9_cast_fp16)[name = tensor("k_complex_5_cast_fp16")]; + tensor var_1050_begin_0 = const()[name = tensor("op_1050_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_1050_end_0 = const()[name = tensor("op_1050_end_0"), val = tensor([1, 1, 16, 32, 1])]; + tensor var_1050_end_mask_0 = const()[name = tensor("op_1050_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_1050_squeeze_mask_0 = const()[name = tensor("op_1050_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_1050_cast_fp16 = slice_by_index(begin = var_1050_begin_0, end = var_1050_end_0, end_mask = var_1050_end_mask_0, squeeze_mask = var_1050_squeeze_mask_0, x = q_complex_5_cast_fp16)[name = tensor("op_1050_cast_fp16")]; + tensor var_1058_begin_0 = const()[name = tensor("op_1058_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_1058_end_0 = const()[name = tensor("op_1058_end_0"), val = tensor([1, 1, 16, 32, 2])]; + tensor var_1058_end_mask_0 = const()[name = tensor("op_1058_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_1058_squeeze_mask_0 = const()[name = tensor("op_1058_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_1058_cast_fp16 = slice_by_index(begin = var_1058_begin_0, end = var_1058_end_0, end_mask = var_1058_end_mask_0, squeeze_mask = var_1058_squeeze_mask_0, x = q_complex_5_cast_fp16)[name = tensor("op_1058_cast_fp16")]; + tensor var_1066_begin_0 = const()[name = tensor("op_1066_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_1066_end_0 = const()[name = tensor("op_1066_end_0"), val = tensor([1, 1, 16, 32, 1])]; + tensor var_1066_end_mask_0 = const()[name = tensor("op_1066_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_1066_squeeze_mask_0 = const()[name = tensor("op_1066_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_1066_cast_fp16 = slice_by_index(begin = var_1066_begin_0, end = var_1066_end_0, end_mask = var_1066_end_mask_0, squeeze_mask = var_1066_squeeze_mask_0, x = k_complex_5_cast_fp16)[name = tensor("op_1066_cast_fp16")]; + tensor var_1074_begin_0 = const()[name = tensor("op_1074_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_1074_end_0 = const()[name = tensor("op_1074_end_0"), val = tensor([1, 1, 16, 32, 2])]; + tensor var_1074_end_mask_0 = const()[name = tensor("op_1074_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_1074_squeeze_mask_0 = const()[name = tensor("op_1074_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_1074_cast_fp16 = slice_by_index(begin = var_1074_begin_0, end = var_1074_end_0, end_mask = var_1074_end_mask_0, squeeze_mask = var_1074_squeeze_mask_0, x = k_complex_5_cast_fp16)[name = tensor("op_1074_cast_fp16")]; + tensor freqs_5_to_fp16 = const()[name = tensor("freqs_5_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(76772864)))]; + tensor var_1080_cast_fp16 = mul(x = freqs_5_to_fp16, y = ts_17_cast_fp16)[name = tensor("op_1080_cast_fp16")]; + tensor rotr_5_cast_fp16 = cos(x = var_1080_cast_fp16)[name = tensor("rotr_5_cast_fp16")]; + tensor roti_5_cast_fp16 = sin(x = var_1080_cast_fp16)[name = tensor("roti_5_cast_fp16")]; + tensor var_1084_cast_fp16 = mul(x = var_1050_cast_fp16, y = rotr_5_cast_fp16)[name = tensor("op_1084_cast_fp16")]; + tensor var_1085_cast_fp16 = mul(x = var_1058_cast_fp16, y = roti_5_cast_fp16)[name = tensor("op_1085_cast_fp16")]; + tensor qor_9_cast_fp16 = sub(x = var_1084_cast_fp16, y = var_1085_cast_fp16)[name = tensor("qor_9_cast_fp16")]; + tensor var_1088_cast_fp16 = mul(x = var_1050_cast_fp16, y = roti_5_cast_fp16)[name = tensor("op_1088_cast_fp16")]; + tensor var_1089_cast_fp16 = mul(x = var_1058_cast_fp16, y = rotr_5_cast_fp16)[name = tensor("op_1089_cast_fp16")]; + tensor qoi_9_cast_fp16 = add(x = var_1088_cast_fp16, y = var_1089_cast_fp16)[name = tensor("qoi_9_cast_fp16")]; + tensor var_1092_cast_fp16 = mul(x = var_1066_cast_fp16, y = rotr_5_cast_fp16)[name = tensor("op_1092_cast_fp16")]; + tensor var_1093_cast_fp16 = mul(x = var_1074_cast_fp16, y = roti_5_cast_fp16)[name = tensor("op_1093_cast_fp16")]; + tensor kor_9_cast_fp16 = sub(x = var_1092_cast_fp16, y = var_1093_cast_fp16)[name = tensor("kor_9_cast_fp16")]; + tensor var_1096_cast_fp16 = mul(x = var_1066_cast_fp16, y = roti_5_cast_fp16)[name = tensor("op_1096_cast_fp16")]; + tensor var_1097_cast_fp16 = mul(x = var_1074_cast_fp16, y = rotr_5_cast_fp16)[name = tensor("op_1097_cast_fp16")]; + tensor koi_9_cast_fp16 = add(x = var_1096_cast_fp16, y = var_1097_cast_fp16)[name = tensor("koi_9_cast_fp16")]; + tensor qo_5_axis_0 = const()[name = tensor("qo_5_axis_0"), val = tensor(-1)]; + tensor qo_5_cast_fp16 = stack(axis = qo_5_axis_0, values = (qor_9_cast_fp16, qoi_9_cast_fp16))[name = tensor("qo_5_cast_fp16")]; + tensor ko_5_axis_0 = const()[name = tensor("ko_5_axis_0"), val = tensor(-1)]; + tensor ko_5_cast_fp16 = stack(axis = ko_5_axis_0, values = (kor_9_cast_fp16, koi_9_cast_fp16))[name = tensor("ko_5_cast_fp16")]; + tensor var_1126 = const()[name = tensor("op_1126"), val = tensor([1, 1, 16, 64])]; + tensor q_15_cast_fp16 = reshape(shape = var_1126, x = qo_5_cast_fp16)[name = tensor("q_15_cast_fp16")]; + tensor var_1128 = const()[name = tensor("op_1128"), val = tensor([1, 1, 16, 64])]; + tensor k_11_cast_fp16 = reshape(shape = var_1128, x = ko_5_cast_fp16)[name = tensor("k_11_cast_fp16")]; + tensor _inversed_1150_y_0_to_fp16 = const()[name = tensor("_inversed_1150_y_0_to_fp16"), val = tensor(0x1p-9)]; + tensor _inversed_1150_cast_fp16 = mul(x = ts_17_cast_fp16, y = _inversed_1150_y_0_to_fp16)[name = tensor("_inversed_1150_cast_fp16")]; + tensor var_1151_cast_fp16 = floor(x = _inversed_1150_cast_fp16)[name = tensor("op_1151_cast_fp16")]; + tensor var_1152_to_fp16 = const()[name = tensor("op_1152_to_fp16"), val = tensor(0x1p+9)]; + tensor var_1153_cast_fp16 = mul(x = var_1151_cast_fp16, y = var_1152_to_fp16)[name = tensor("op_1153_cast_fp16")]; + tensor write_indices_float_11_cast_fp16 = sub(x = ts_17_cast_fp16, y = var_1153_cast_fp16)[name = tensor("write_indices_float_11_cast_fp16")]; + tensor var_1160_dtype_0 = const()[name = tensor("op_1160_dtype_0"), val = tensor("int32")]; + tensor write_indices_5_reps_0 = const()[name = tensor("write_indices_5_reps_0"), val = tensor([1, 1, 16, 64])]; + tensor write_indices_float_11_cast_fp16_to_int32 = cast(dtype = var_1160_dtype_0, x = write_indices_float_11_cast_fp16)[name = tensor("cast_128")]; + tensor write_indices_5 = tile(reps = write_indices_5_reps_0, x = write_indices_float_11_cast_fp16_to_int32)[name = tensor("write_indices_5")]; + tensor var_1168_begin_0 = const()[name = tensor("op_1168_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_1168_end_0 = const()[name = tensor("op_1168_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor var_1168_end_mask_0 = const()[name = tensor("op_1168_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_1168_squeeze_mask_0 = const()[name = tensor("op_1168_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor cache2_to_fp16_dtype_0 = const()[name = tensor("cache2_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor cache2_to_fp16 = cast(dtype = cache2_to_fp16_dtype_0, x = cache2)[name = tensor("cast_127")]; + tensor var_1168_cast_fp16 = slice_by_index(begin = var_1168_begin_0, end = var_1168_end_0, end_mask = var_1168_end_mask_0, squeeze_mask = var_1168_squeeze_mask_0, x = cache2_to_fp16)[name = tensor("op_1168_cast_fp16")]; + tensor var_1170_axis_0 = const()[name = tensor("op_1170_axis_0"), val = tensor(1)]; + tensor var_1170_mode_0 = const()[name = tensor("op_1170_mode_0"), val = tensor("update")]; + tensor var_1170_validate_indices_0 = const()[name = tensor("op_1170_validate_indices_0"), val = tensor(false)]; + tensor var_1170_cast_fp16 = scatter_along_axis(axis = var_1170_axis_0, data = var_1168_cast_fp16, indices = write_indices_5, mode = var_1170_mode_0, updates = k_11_cast_fp16, validate_indices = var_1170_validate_indices_0)[name = tensor("op_1170_cast_fp16")]; + tensor concat_16 = const()[name = tensor("concat_16"), val = tensor([0, 0, 0, 0, 0])]; + tensor concat_17 = const()[name = tensor("concat_17"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_5_internal_tensor_assign_1_stride_0 = const()[name = tensor("new_cache_5_internal_tensor_assign_1_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_5_internal_tensor_assign_1_begin_mask_0 = const()[name = tensor("new_cache_5_internal_tensor_assign_1_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_5_internal_tensor_assign_1_end_mask_0 = const()[name = tensor("new_cache_5_internal_tensor_assign_1_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_5_internal_tensor_assign_1_squeeze_mask_0 = const()[name = tensor("new_cache_5_internal_tensor_assign_1_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_16 = const()[name = tensor("shape_16"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_4 = const()[name = tensor("reduce_prod_4"), val = tensor(1048576)]; + tensor range_1d_4_start_0 = const()[name = tensor("range_1d_4_start_0"), val = tensor(0)]; + tensor range_1d_4_step_0 = const()[name = tensor("range_1d_4_step_0"), val = tensor(1)]; + tensor range_1d_4 = range_1d(end = reduce_prod_4, start = range_1d_4_start_0, step = range_1d_4_step_0)[name = tensor("range_1d_4")]; + tensor reshape_20 = reshape(shape = shape_16, x = range_1d_4)[name = tensor("reshape_20")]; + tensor slice_by_index_4 = slice_by_index(begin = concat_16, begin_mask = new_cache_5_internal_tensor_assign_1_begin_mask_0, end = concat_17, end_mask = new_cache_5_internal_tensor_assign_1_end_mask_0, squeeze_mask = new_cache_5_internal_tensor_assign_1_squeeze_mask_0, stride = new_cache_5_internal_tensor_assign_1_stride_0, x = reshape_20)[name = tensor("slice_by_index_4")]; + tensor reshape_21_shape_0 = const()[name = tensor("reshape_21_shape_0"), val = tensor([-1])]; + tensor reshape_21 = reshape(shape = reshape_21_shape_0, x = slice_by_index_4)[name = tensor("reshape_21")]; + tensor reshape_22_shape_0 = const()[name = tensor("reshape_22_shape_0"), val = tensor([-1])]; + tensor reshape_22_cast_fp16 = reshape(shape = reshape_22_shape_0, x = var_1170_cast_fp16)[name = tensor("reshape_22_cast_fp16")]; + tensor reshape_23_shape_0 = const()[name = tensor("reshape_23_shape_0"), val = tensor([-1])]; + tensor reshape_23_cast_fp16 = reshape(shape = reshape_23_shape_0, x = cache2_to_fp16)[name = tensor("reshape_23_cast_fp16")]; + tensor scatter_4_mode_0 = const()[name = tensor("scatter_4_mode_0"), val = tensor("update")]; + tensor scatter_4_axis_0 = const()[name = tensor("scatter_4_axis_0"), val = tensor(0)]; + tensor scatter_4_validate_indices_0 = const()[name = tensor("scatter_4_validate_indices_0"), val = tensor(false)]; + tensor scatter_4_cast_fp16 = scatter(axis = scatter_4_axis_0, data = reshape_23_cast_fp16, indices = reshape_21, mode = scatter_4_mode_0, updates = reshape_22_cast_fp16, validate_indices = scatter_4_validate_indices_0)[name = tensor("scatter_4_cast_fp16")]; + tensor reshape_24_cast_fp16 = reshape(shape = shape_16, x = scatter_4_cast_fp16)[name = tensor("reshape_24_cast_fp16")]; + tensor var_1178_begin_0 = const()[name = tensor("op_1178_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor var_1178_end_0 = const()[name = tensor("op_1178_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor var_1178_end_mask_0 = const()[name = tensor("op_1178_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_1178_squeeze_mask_0 = const()[name = tensor("op_1178_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor var_1178_cast_fp16 = slice_by_index(begin = var_1178_begin_0, end = var_1178_end_0, end_mask = var_1178_end_mask_0, squeeze_mask = var_1178_squeeze_mask_0, x = reshape_24_cast_fp16)[name = tensor("op_1178_cast_fp16")]; + tensor var_1180_axis_0 = const()[name = tensor("op_1180_axis_0"), val = tensor(1)]; + tensor var_1180_mode_0 = const()[name = tensor("op_1180_mode_0"), val = tensor("update")]; + tensor var_1180_validate_indices_0 = const()[name = tensor("op_1180_validate_indices_0"), val = tensor(false)]; + tensor var_1180_cast_fp16 = scatter_along_axis(axis = var_1180_axis_0, data = var_1178_cast_fp16, indices = write_indices_5, mode = var_1180_mode_0, updates = v_5_cast_fp16, validate_indices = var_1180_validate_indices_0)[name = tensor("op_1180_cast_fp16")]; + tensor concat_18 = const()[name = tensor("concat_18"), val = tensor([1, 0, 0, 0, 0])]; + tensor concat_19 = const()[name = tensor("concat_19"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_5_internal_tensor_assign_2_stride_0 = const()[name = tensor("new_cache_5_internal_tensor_assign_2_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_5_internal_tensor_assign_2_begin_mask_0 = const()[name = tensor("new_cache_5_internal_tensor_assign_2_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_5_internal_tensor_assign_2_end_mask_0 = const()[name = tensor("new_cache_5_internal_tensor_assign_2_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_5_internal_tensor_assign_2_squeeze_mask_0 = const()[name = tensor("new_cache_5_internal_tensor_assign_2_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_17 = const()[name = tensor("shape_17"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_5 = const()[name = tensor("reduce_prod_5"), val = tensor(1048576)]; + tensor range_1d_5_start_0 = const()[name = tensor("range_1d_5_start_0"), val = tensor(0)]; + tensor range_1d_5_step_0 = const()[name = tensor("range_1d_5_step_0"), val = tensor(1)]; + tensor range_1d_5 = range_1d(end = reduce_prod_5, start = range_1d_5_start_0, step = range_1d_5_step_0)[name = tensor("range_1d_5")]; + tensor reshape_25 = reshape(shape = shape_17, x = range_1d_5)[name = tensor("reshape_25")]; + tensor slice_by_index_5 = slice_by_index(begin = concat_18, begin_mask = new_cache_5_internal_tensor_assign_2_begin_mask_0, end = concat_19, end_mask = new_cache_5_internal_tensor_assign_2_end_mask_0, squeeze_mask = new_cache_5_internal_tensor_assign_2_squeeze_mask_0, stride = new_cache_5_internal_tensor_assign_2_stride_0, x = reshape_25)[name = tensor("slice_by_index_5")]; + tensor reshape_26_shape_0 = const()[name = tensor("reshape_26_shape_0"), val = tensor([-1])]; + tensor reshape_26 = reshape(shape = reshape_26_shape_0, x = slice_by_index_5)[name = tensor("reshape_26")]; + tensor reshape_27_shape_0 = const()[name = tensor("reshape_27_shape_0"), val = tensor([-1])]; + tensor reshape_27_cast_fp16 = reshape(shape = reshape_27_shape_0, x = var_1180_cast_fp16)[name = tensor("reshape_27_cast_fp16")]; + tensor reshape_28_shape_0 = const()[name = tensor("reshape_28_shape_0"), val = tensor([-1])]; + tensor reshape_28_cast_fp16 = reshape(shape = reshape_28_shape_0, x = reshape_24_cast_fp16)[name = tensor("reshape_28_cast_fp16")]; + tensor scatter_5_mode_0 = const()[name = tensor("scatter_5_mode_0"), val = tensor("update")]; + tensor scatter_5_axis_0 = const()[name = tensor("scatter_5_axis_0"), val = tensor(0)]; + tensor scatter_5_validate_indices_0 = const()[name = tensor("scatter_5_validate_indices_0"), val = tensor(false)]; + tensor scatter_5_cast_fp16 = scatter(axis = scatter_5_axis_0, data = reshape_28_cast_fp16, indices = reshape_26, mode = scatter_5_mode_0, updates = reshape_27_cast_fp16, validate_indices = scatter_5_validate_indices_0)[name = tensor("scatter_5_cast_fp16")]; + tensor reshape_29_cast_fp16 = reshape(shape = shape_17, x = scatter_5_cast_fp16)[name = tensor("reshape_29_cast_fp16")]; + tensor reshape_29_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("reshape_29_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor keys_13_begin_0 = const()[name = tensor("keys_13_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor keys_13_end_0 = const()[name = tensor("keys_13_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor keys_13_end_mask_0 = const()[name = tensor("keys_13_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor keys_13_squeeze_mask_0 = const()[name = tensor("keys_13_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor keys_13_cast_fp16 = slice_by_index(begin = keys_13_begin_0, end = keys_13_end_0, end_mask = keys_13_end_mask_0, squeeze_mask = keys_13_squeeze_mask_0, x = reshape_29_cast_fp16)[name = tensor("keys_13_cast_fp16")]; + tensor values_13_begin_0 = const()[name = tensor("values_13_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor values_13_end_0 = const()[name = tensor("values_13_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor values_13_end_mask_0 = const()[name = tensor("values_13_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor values_13_squeeze_mask_0 = const()[name = tensor("values_13_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor values_13_cast_fp16 = slice_by_index(begin = values_13_begin_0, end = values_13_end_0, end_mask = values_13_end_mask_0, squeeze_mask = values_13_squeeze_mask_0, x = reshape_29_cast_fp16)[name = tensor("values_13_cast_fp16")]; + tensor var_1192_cast_fp16 = not_equal(x = keys_13_cast_fp16, y = keys_13_cast_fp16)[name = tensor("op_1192_cast_fp16")]; + tensor keys_15_cast_fp16 = select(a = var_360_to_fp16, b = keys_13_cast_fp16, cond = var_1192_cast_fp16)[name = tensor("keys_15_cast_fp16")]; + tensor var_1200_cast_fp16 = not_equal(x = values_13_cast_fp16, y = values_13_cast_fp16)[name = tensor("op_1200_cast_fp16")]; + tensor values_15_cast_fp16 = select(a = var_360_to_fp16, b = values_13_cast_fp16, cond = var_1200_cast_fp16)[name = tensor("values_15_cast_fp16")]; + tensor var_1224 = const()[name = tensor("op_1224"), val = tensor([0, 2, 1, 3])]; + tensor var_1237 = const()[name = tensor("op_1237"), val = tensor([1, 1, 1])]; + tensor var_1238_cast_fp16 = reshape(shape = var_1237, x = position2_to_fp16)[name = tensor("op_1238_cast_fp16")]; + tensor var_1255_to_fp16 = const()[name = tensor("op_1255_to_fp16"), val = tensor(0x1p+0)]; + tensor valid_len_5_cast_fp16 = add(x = var_1238_cast_fp16, y = var_1255_to_fp16)[name = tensor("valid_len_5_cast_fp16")]; + tensor valid_mask_5_cast_fp16 = less(x = k_positions_1_promoted_to_fp16, y = valid_len_5_cast_fp16)[name = tensor("valid_mask_5_cast_fp16")]; + tensor causal_mask_5_cast_fp16 = less_equal(x = k_positions_1_promoted_to_fp16, y = var_1238_cast_fp16)[name = tensor("causal_mask_5_cast_fp16")]; + tensor attn_mask_9 = logical_and(x = valid_mask_5_cast_fp16, y = causal_mask_5_cast_fp16)[name = tensor("attn_mask_9")]; + tensor attn_mask_11_axes_0 = const()[name = tensor("attn_mask_11_axes_0"), val = tensor([1])]; + tensor attn_mask_11 = expand_dims(axes = attn_mask_11_axes_0, x = attn_mask_9)[name = tensor("attn_mask_11")]; + tensor var_1273_transpose_x_0 = const()[name = tensor("op_1273_transpose_x_0"), val = tensor(false)]; + tensor var_1273_transpose_y_0 = const()[name = tensor("op_1273_transpose_y_0"), val = tensor(false)]; + tensor transpose_22_perm_0 = const()[name = tensor("transpose_22_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_23_perm_0 = const()[name = tensor("transpose_23_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_23 = transpose(perm = transpose_23_perm_0, x = keys_15_cast_fp16)[name = tensor("transpose_43")]; + tensor transpose_22 = transpose(perm = transpose_22_perm_0, x = q_15_cast_fp16)[name = tensor("transpose_44")]; + tensor var_1273_cast_fp16 = matmul(transpose_x = var_1273_transpose_x_0, transpose_y = var_1273_transpose_y_0, x = transpose_22, y = transpose_23)[name = tensor("op_1273_cast_fp16")]; + tensor var_1267_to_fp16 = const()[name = tensor("op_1267_to_fp16"), val = tensor([0x1p-3])]; + tensor attn_weights_13_cast_fp16 = mul(x = var_1273_cast_fp16, y = var_1267_to_fp16)[name = tensor("attn_weights_13_cast_fp16")]; + tensor var_1275 = logical_not(x = attn_mask_11)[name = tensor("op_1275")]; + tensor var_1276_to_fp16 = const()[name = tensor("op_1276_to_fp16"), val = tensor(-inf)]; + tensor attn_weights_15_cast_fp16 = select(a = var_1276_to_fp16, b = attn_weights_13_cast_fp16, cond = var_1275)[name = tensor("attn_weights_15_cast_fp16")]; + tensor var_1278 = const()[name = tensor("op_1278"), val = tensor(-1)]; + tensor attn_weights_17_cast_fp16 = softmax(axis = var_1278, x = attn_weights_15_cast_fp16)[name = tensor("attn_weights_17_cast_fp16")]; + tensor attn_output_5_transpose_x_0 = const()[name = tensor("attn_output_5_transpose_x_0"), val = tensor(false)]; + tensor attn_output_5_transpose_y_0 = const()[name = tensor("attn_output_5_transpose_y_0"), val = tensor(false)]; + tensor values_17_cast_fp16 = transpose(perm = var_1224, x = values_15_cast_fp16)[name = tensor("transpose_45")]; + tensor attn_output_5_cast_fp16 = matmul(transpose_x = attn_output_5_transpose_x_0, transpose_y = attn_output_5_transpose_y_0, x = attn_weights_17_cast_fp16, y = values_17_cast_fp16)[name = tensor("attn_output_5_cast_fp16")]; + tensor var_1286 = const()[name = tensor("op_1286"), val = tensor([0, 2, 1, 3])]; + tensor var_1289 = const()[name = tensor("op_1289"), val = tensor([1, 1, 1024])]; + tensor var_1287_cast_fp16 = transpose(perm = var_1286, x = attn_output_5_cast_fp16)[name = tensor("transpose_42")]; + tensor input_25_cast_fp16 = reshape(shape = var_1289, x = var_1287_cast_fp16)[name = tensor("input_25_cast_fp16")]; + tensor linear_10_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = attn2_out_proj_weight_cast_fp16, x = input_25_cast_fp16)[name = tensor("linear_10_cast_fp16")]; + tensor var_1295_to_fp16 = const()[name = tensor("op_1295_to_fp16"), val = tensor(0x1p+0)]; + tensor var_1296_cast_fp16 = add(x = position2_to_fp16, y = var_1295_to_fp16)[name = tensor("op_1296_cast_fp16")]; + tensor var_1296_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_1296_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor input_27_cast_fp16 = add(x = input_23_cast_fp16, y = linear_10_cast_fp16)[name = tensor("input_27_cast_fp16")]; + tensor input_29_axes_0 = const()[name = tensor("input_29_axes_0"), val = tensor([-1])]; + tensor norm2_2_weight_to_fp16 = const()[name = tensor("norm2_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(76772992)))]; + tensor norm2_2_bias_to_fp16 = const()[name = tensor("norm2_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(76775104)))]; + tensor var_1300_to_fp16 = const()[name = tensor("op_1300_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_29_cast_fp16 = layer_norm(axes = input_29_axes_0, beta = norm2_2_bias_to_fp16, epsilon = var_1300_to_fp16, gamma = norm2_2_weight_to_fp16, x = input_27_cast_fp16)[name = tensor("input_29_cast_fp16")]; + tensor linear_11_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = linear2_1_weight_cast_fp16, x = input_29_cast_fp16)[name = tensor("linear_11_cast_fp16")]; + tensor input_31_mode_0 = const()[name = tensor("input_31_mode_0"), val = tensor("EXACT")]; + tensor input_31_cast_fp16 = gelu(mode = input_31_mode_0, x = linear_11_cast_fp16)[name = tensor("input_31_cast_fp16")]; + tensor linear_12_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = linear2_2_weight_cast_fp16, x = input_31_cast_fp16)[name = tensor("linear_12_cast_fp16")]; + tensor input_33_cast_fp16 = add(x = input_27_cast_fp16, y = linear_12_cast_fp16)[name = tensor("input_33_cast_fp16")]; + tensor x_7_axes_0 = const()[name = tensor("x_7_axes_0"), val = tensor([-1])]; + tensor norm3_1_weight_to_fp16 = const()[name = tensor("norm3_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(76777216)))]; + tensor norm3_1_bias_to_fp16 = const()[name = tensor("norm3_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(76779328)))]; + tensor var_1317_to_fp16 = const()[name = tensor("op_1317_to_fp16"), val = tensor(0x1.5p-17)]; + tensor x_7_cast_fp16 = layer_norm(axes = x_7_axes_0, beta = norm3_1_bias_to_fp16, epsilon = var_1317_to_fp16, gamma = norm3_1_weight_to_fp16, x = input_33_cast_fp16)[name = tensor("x_7_cast_fp16")]; + tensor linear_13_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = attn3_in_proj_weight_cast_fp16, x = x_7_cast_fp16)[name = tensor("linear_13_cast_fp16")]; + tensor var_1353 = const()[name = tensor("op_1353"), val = tensor([1, 1, 3, 16, 64])]; + tensor qkv_7_cast_fp16 = reshape(shape = var_1353, x = linear_13_cast_fp16)[name = tensor("qkv_7_cast_fp16")]; + tensor q_19_begin_0 = const()[name = tensor("q_19_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor q_19_end_0 = const()[name = tensor("q_19_end_0"), val = tensor([1, 1, 1, 16, 64])]; + tensor q_19_end_mask_0 = const()[name = tensor("q_19_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor q_19_squeeze_mask_0 = const()[name = tensor("q_19_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor q_19_cast_fp16 = slice_by_index(begin = q_19_begin_0, end = q_19_end_0, end_mask = q_19_end_mask_0, squeeze_mask = q_19_squeeze_mask_0, x = qkv_7_cast_fp16)[name = tensor("q_19_cast_fp16")]; + tensor k_13_begin_0 = const()[name = tensor("k_13_begin_0"), val = tensor([0, 0, 1, 0, 0])]; + tensor k_13_end_0 = const()[name = tensor("k_13_end_0"), val = tensor([1, 1, 2, 16, 64])]; + tensor k_13_end_mask_0 = const()[name = tensor("k_13_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor k_13_squeeze_mask_0 = const()[name = tensor("k_13_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor k_13_cast_fp16 = slice_by_index(begin = k_13_begin_0, end = k_13_end_0, end_mask = k_13_end_mask_0, squeeze_mask = k_13_squeeze_mask_0, x = qkv_7_cast_fp16)[name = tensor("k_13_cast_fp16")]; + tensor v_7_begin_0 = const()[name = tensor("v_7_begin_0"), val = tensor([0, 0, 2, 0, 0])]; + tensor v_7_end_0 = const()[name = tensor("v_7_end_0"), val = tensor([1, 1, 3, 16, 64])]; + tensor v_7_end_mask_0 = const()[name = tensor("v_7_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor v_7_squeeze_mask_0 = const()[name = tensor("v_7_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor v_7_cast_fp16 = slice_by_index(begin = v_7_begin_0, end = v_7_end_0, end_mask = v_7_end_mask_0, squeeze_mask = v_7_squeeze_mask_0, x = qkv_7_cast_fp16)[name = tensor("v_7_cast_fp16")]; + tensor var_1457 = const()[name = tensor("op_1457"), val = tensor([1, 1, 1, 1])]; + tensor position3_to_fp16_dtype_0 = const()[name = tensor("position3_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor position3_to_fp16 = cast(dtype = position3_to_fp16_dtype_0, x = position3)[name = tensor("cast_124")]; + tensor ts_23_cast_fp16 = reshape(shape = var_1457, x = position3_to_fp16)[name = tensor("ts_23_cast_fp16")]; + tensor var_1461 = const()[name = tensor("op_1461"), val = tensor([1, 1, 16, 32, 2])]; + tensor q_complex_7_cast_fp16 = reshape(shape = var_1461, x = q_19_cast_fp16)[name = tensor("q_complex_7_cast_fp16")]; + tensor var_1465 = const()[name = tensor("op_1465"), val = tensor([1, 1, 16, 32, 2])]; + tensor k_complex_7_cast_fp16 = reshape(shape = var_1465, x = k_13_cast_fp16)[name = tensor("k_complex_7_cast_fp16")]; + tensor var_1469_begin_0 = const()[name = tensor("op_1469_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_1469_end_0 = const()[name = tensor("op_1469_end_0"), val = tensor([1, 1, 16, 32, 1])]; + tensor var_1469_end_mask_0 = const()[name = tensor("op_1469_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_1469_squeeze_mask_0 = const()[name = tensor("op_1469_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_1469_cast_fp16 = slice_by_index(begin = var_1469_begin_0, end = var_1469_end_0, end_mask = var_1469_end_mask_0, squeeze_mask = var_1469_squeeze_mask_0, x = q_complex_7_cast_fp16)[name = tensor("op_1469_cast_fp16")]; + tensor var_1477_begin_0 = const()[name = tensor("op_1477_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_1477_end_0 = const()[name = tensor("op_1477_end_0"), val = tensor([1, 1, 16, 32, 2])]; + tensor var_1477_end_mask_0 = const()[name = tensor("op_1477_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_1477_squeeze_mask_0 = const()[name = tensor("op_1477_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_1477_cast_fp16 = slice_by_index(begin = var_1477_begin_0, end = var_1477_end_0, end_mask = var_1477_end_mask_0, squeeze_mask = var_1477_squeeze_mask_0, x = q_complex_7_cast_fp16)[name = tensor("op_1477_cast_fp16")]; + tensor var_1485_begin_0 = const()[name = tensor("op_1485_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_1485_end_0 = const()[name = tensor("op_1485_end_0"), val = tensor([1, 1, 16, 32, 1])]; + tensor var_1485_end_mask_0 = const()[name = tensor("op_1485_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_1485_squeeze_mask_0 = const()[name = tensor("op_1485_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_1485_cast_fp16 = slice_by_index(begin = var_1485_begin_0, end = var_1485_end_0, end_mask = var_1485_end_mask_0, squeeze_mask = var_1485_squeeze_mask_0, x = k_complex_7_cast_fp16)[name = tensor("op_1485_cast_fp16")]; + tensor var_1493_begin_0 = const()[name = tensor("op_1493_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_1493_end_0 = const()[name = tensor("op_1493_end_0"), val = tensor([1, 1, 16, 32, 2])]; + tensor var_1493_end_mask_0 = const()[name = tensor("op_1493_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_1493_squeeze_mask_0 = const()[name = tensor("op_1493_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_1493_cast_fp16 = slice_by_index(begin = var_1493_begin_0, end = var_1493_end_0, end_mask = var_1493_end_mask_0, squeeze_mask = var_1493_squeeze_mask_0, x = k_complex_7_cast_fp16)[name = tensor("op_1493_cast_fp16")]; + tensor freqs_7_to_fp16 = const()[name = tensor("freqs_7_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(76781440)))]; + tensor var_1499_cast_fp16 = mul(x = freqs_7_to_fp16, y = ts_23_cast_fp16)[name = tensor("op_1499_cast_fp16")]; + tensor rotr_7_cast_fp16 = cos(x = var_1499_cast_fp16)[name = tensor("rotr_7_cast_fp16")]; + tensor roti_7_cast_fp16 = sin(x = var_1499_cast_fp16)[name = tensor("roti_7_cast_fp16")]; + tensor var_1503_cast_fp16 = mul(x = var_1469_cast_fp16, y = rotr_7_cast_fp16)[name = tensor("op_1503_cast_fp16")]; + tensor var_1504_cast_fp16 = mul(x = var_1477_cast_fp16, y = roti_7_cast_fp16)[name = tensor("op_1504_cast_fp16")]; + tensor qor_13_cast_fp16 = sub(x = var_1503_cast_fp16, y = var_1504_cast_fp16)[name = tensor("qor_13_cast_fp16")]; + tensor var_1507_cast_fp16 = mul(x = var_1469_cast_fp16, y = roti_7_cast_fp16)[name = tensor("op_1507_cast_fp16")]; + tensor var_1508_cast_fp16 = mul(x = var_1477_cast_fp16, y = rotr_7_cast_fp16)[name = tensor("op_1508_cast_fp16")]; + tensor qoi_13_cast_fp16 = add(x = var_1507_cast_fp16, y = var_1508_cast_fp16)[name = tensor("qoi_13_cast_fp16")]; + tensor var_1511_cast_fp16 = mul(x = var_1485_cast_fp16, y = rotr_7_cast_fp16)[name = tensor("op_1511_cast_fp16")]; + tensor var_1512_cast_fp16 = mul(x = var_1493_cast_fp16, y = roti_7_cast_fp16)[name = tensor("op_1512_cast_fp16")]; + tensor kor_13_cast_fp16 = sub(x = var_1511_cast_fp16, y = var_1512_cast_fp16)[name = tensor("kor_13_cast_fp16")]; + tensor var_1515_cast_fp16 = mul(x = var_1485_cast_fp16, y = roti_7_cast_fp16)[name = tensor("op_1515_cast_fp16")]; + tensor var_1516_cast_fp16 = mul(x = var_1493_cast_fp16, y = rotr_7_cast_fp16)[name = tensor("op_1516_cast_fp16")]; + tensor koi_13_cast_fp16 = add(x = var_1515_cast_fp16, y = var_1516_cast_fp16)[name = tensor("koi_13_cast_fp16")]; + tensor qo_7_axis_0 = const()[name = tensor("qo_7_axis_0"), val = tensor(-1)]; + tensor qo_7_cast_fp16 = stack(axis = qo_7_axis_0, values = (qor_13_cast_fp16, qoi_13_cast_fp16))[name = tensor("qo_7_cast_fp16")]; + tensor ko_7_axis_0 = const()[name = tensor("ko_7_axis_0"), val = tensor(-1)]; + tensor ko_7_cast_fp16 = stack(axis = ko_7_axis_0, values = (kor_13_cast_fp16, koi_13_cast_fp16))[name = tensor("ko_7_cast_fp16")]; + tensor var_1545 = const()[name = tensor("op_1545"), val = tensor([1, 1, 16, 64])]; + tensor q_21_cast_fp16 = reshape(shape = var_1545, x = qo_7_cast_fp16)[name = tensor("q_21_cast_fp16")]; + tensor var_1547 = const()[name = tensor("op_1547"), val = tensor([1, 1, 16, 64])]; + tensor k_15_cast_fp16 = reshape(shape = var_1547, x = ko_7_cast_fp16)[name = tensor("k_15_cast_fp16")]; + tensor _inversed_1569_y_0_to_fp16 = const()[name = tensor("_inversed_1569_y_0_to_fp16"), val = tensor(0x1p-9)]; + tensor _inversed_1569_cast_fp16 = mul(x = ts_23_cast_fp16, y = _inversed_1569_y_0_to_fp16)[name = tensor("_inversed_1569_cast_fp16")]; + tensor var_1570_cast_fp16 = floor(x = _inversed_1569_cast_fp16)[name = tensor("op_1570_cast_fp16")]; + tensor var_1571_to_fp16 = const()[name = tensor("op_1571_to_fp16"), val = tensor(0x1p+9)]; + tensor var_1572_cast_fp16 = mul(x = var_1570_cast_fp16, y = var_1571_to_fp16)[name = tensor("op_1572_cast_fp16")]; + tensor write_indices_float_15_cast_fp16 = sub(x = ts_23_cast_fp16, y = var_1572_cast_fp16)[name = tensor("write_indices_float_15_cast_fp16")]; + tensor var_1579_dtype_0 = const()[name = tensor("op_1579_dtype_0"), val = tensor("int32")]; + tensor write_indices_7_reps_0 = const()[name = tensor("write_indices_7_reps_0"), val = tensor([1, 1, 16, 64])]; + tensor write_indices_float_15_cast_fp16_to_int32 = cast(dtype = var_1579_dtype_0, x = write_indices_float_15_cast_fp16)[name = tensor("cast_123")]; + tensor write_indices_7 = tile(reps = write_indices_7_reps_0, x = write_indices_float_15_cast_fp16_to_int32)[name = tensor("write_indices_7")]; + tensor var_1587_begin_0 = const()[name = tensor("op_1587_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_1587_end_0 = const()[name = tensor("op_1587_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor var_1587_end_mask_0 = const()[name = tensor("op_1587_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_1587_squeeze_mask_0 = const()[name = tensor("op_1587_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor cache3_to_fp16_dtype_0 = const()[name = tensor("cache3_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor cache3_to_fp16 = cast(dtype = cache3_to_fp16_dtype_0, x = cache3)[name = tensor("cast_122")]; + tensor var_1587_cast_fp16 = slice_by_index(begin = var_1587_begin_0, end = var_1587_end_0, end_mask = var_1587_end_mask_0, squeeze_mask = var_1587_squeeze_mask_0, x = cache3_to_fp16)[name = tensor("op_1587_cast_fp16")]; + tensor var_1589_axis_0 = const()[name = tensor("op_1589_axis_0"), val = tensor(1)]; + tensor var_1589_mode_0 = const()[name = tensor("op_1589_mode_0"), val = tensor("update")]; + tensor var_1589_validate_indices_0 = const()[name = tensor("op_1589_validate_indices_0"), val = tensor(false)]; + tensor var_1589_cast_fp16 = scatter_along_axis(axis = var_1589_axis_0, data = var_1587_cast_fp16, indices = write_indices_7, mode = var_1589_mode_0, updates = k_15_cast_fp16, validate_indices = var_1589_validate_indices_0)[name = tensor("op_1589_cast_fp16")]; + tensor concat_23 = const()[name = tensor("concat_23"), val = tensor([0, 0, 0, 0, 0])]; + tensor concat_24 = const()[name = tensor("concat_24"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_7_internal_tensor_assign_1_stride_0 = const()[name = tensor("new_cache_7_internal_tensor_assign_1_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_7_internal_tensor_assign_1_begin_mask_0 = const()[name = tensor("new_cache_7_internal_tensor_assign_1_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_7_internal_tensor_assign_1_end_mask_0 = const()[name = tensor("new_cache_7_internal_tensor_assign_1_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_7_internal_tensor_assign_1_squeeze_mask_0 = const()[name = tensor("new_cache_7_internal_tensor_assign_1_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_18 = const()[name = tensor("shape_18"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_6 = const()[name = tensor("reduce_prod_6"), val = tensor(1048576)]; + tensor range_1d_6_start_0 = const()[name = tensor("range_1d_6_start_0"), val = tensor(0)]; + tensor range_1d_6_step_0 = const()[name = tensor("range_1d_6_step_0"), val = tensor(1)]; + tensor range_1d_6 = range_1d(end = reduce_prod_6, start = range_1d_6_start_0, step = range_1d_6_step_0)[name = tensor("range_1d_6")]; + tensor reshape_30 = reshape(shape = shape_18, x = range_1d_6)[name = tensor("reshape_30")]; + tensor slice_by_index_6 = slice_by_index(begin = concat_23, begin_mask = new_cache_7_internal_tensor_assign_1_begin_mask_0, end = concat_24, end_mask = new_cache_7_internal_tensor_assign_1_end_mask_0, squeeze_mask = new_cache_7_internal_tensor_assign_1_squeeze_mask_0, stride = new_cache_7_internal_tensor_assign_1_stride_0, x = reshape_30)[name = tensor("slice_by_index_6")]; + tensor reshape_31_shape_0 = const()[name = tensor("reshape_31_shape_0"), val = tensor([-1])]; + tensor reshape_31 = reshape(shape = reshape_31_shape_0, x = slice_by_index_6)[name = tensor("reshape_31")]; + tensor reshape_32_shape_0 = const()[name = tensor("reshape_32_shape_0"), val = tensor([-1])]; + tensor reshape_32_cast_fp16 = reshape(shape = reshape_32_shape_0, x = var_1589_cast_fp16)[name = tensor("reshape_32_cast_fp16")]; + tensor reshape_33_shape_0 = const()[name = tensor("reshape_33_shape_0"), val = tensor([-1])]; + tensor reshape_33_cast_fp16 = reshape(shape = reshape_33_shape_0, x = cache3_to_fp16)[name = tensor("reshape_33_cast_fp16")]; + tensor scatter_6_mode_0 = const()[name = tensor("scatter_6_mode_0"), val = tensor("update")]; + tensor scatter_6_axis_0 = const()[name = tensor("scatter_6_axis_0"), val = tensor(0)]; + tensor scatter_6_validate_indices_0 = const()[name = tensor("scatter_6_validate_indices_0"), val = tensor(false)]; + tensor scatter_6_cast_fp16 = scatter(axis = scatter_6_axis_0, data = reshape_33_cast_fp16, indices = reshape_31, mode = scatter_6_mode_0, updates = reshape_32_cast_fp16, validate_indices = scatter_6_validate_indices_0)[name = tensor("scatter_6_cast_fp16")]; + tensor reshape_34_cast_fp16 = reshape(shape = shape_18, x = scatter_6_cast_fp16)[name = tensor("reshape_34_cast_fp16")]; + tensor var_1597_begin_0 = const()[name = tensor("op_1597_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor var_1597_end_0 = const()[name = tensor("op_1597_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor var_1597_end_mask_0 = const()[name = tensor("op_1597_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_1597_squeeze_mask_0 = const()[name = tensor("op_1597_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor var_1597_cast_fp16 = slice_by_index(begin = var_1597_begin_0, end = var_1597_end_0, end_mask = var_1597_end_mask_0, squeeze_mask = var_1597_squeeze_mask_0, x = reshape_34_cast_fp16)[name = tensor("op_1597_cast_fp16")]; + tensor var_1599_axis_0 = const()[name = tensor("op_1599_axis_0"), val = tensor(1)]; + tensor var_1599_mode_0 = const()[name = tensor("op_1599_mode_0"), val = tensor("update")]; + tensor var_1599_validate_indices_0 = const()[name = tensor("op_1599_validate_indices_0"), val = tensor(false)]; + tensor var_1599_cast_fp16 = scatter_along_axis(axis = var_1599_axis_0, data = var_1597_cast_fp16, indices = write_indices_7, mode = var_1599_mode_0, updates = v_7_cast_fp16, validate_indices = var_1599_validate_indices_0)[name = tensor("op_1599_cast_fp16")]; + tensor concat_25 = const()[name = tensor("concat_25"), val = tensor([1, 0, 0, 0, 0])]; + tensor concat_26 = const()[name = tensor("concat_26"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_7_internal_tensor_assign_2_stride_0 = const()[name = tensor("new_cache_7_internal_tensor_assign_2_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_7_internal_tensor_assign_2_begin_mask_0 = const()[name = tensor("new_cache_7_internal_tensor_assign_2_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_7_internal_tensor_assign_2_end_mask_0 = const()[name = tensor("new_cache_7_internal_tensor_assign_2_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_7_internal_tensor_assign_2_squeeze_mask_0 = const()[name = tensor("new_cache_7_internal_tensor_assign_2_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_19 = const()[name = tensor("shape_19"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_7 = const()[name = tensor("reduce_prod_7"), val = tensor(1048576)]; + tensor range_1d_7_start_0 = const()[name = tensor("range_1d_7_start_0"), val = tensor(0)]; + tensor range_1d_7_step_0 = const()[name = tensor("range_1d_7_step_0"), val = tensor(1)]; + tensor range_1d_7 = range_1d(end = reduce_prod_7, start = range_1d_7_start_0, step = range_1d_7_step_0)[name = tensor("range_1d_7")]; + tensor reshape_35 = reshape(shape = shape_19, x = range_1d_7)[name = tensor("reshape_35")]; + tensor slice_by_index_7 = slice_by_index(begin = concat_25, begin_mask = new_cache_7_internal_tensor_assign_2_begin_mask_0, end = concat_26, end_mask = new_cache_7_internal_tensor_assign_2_end_mask_0, squeeze_mask = new_cache_7_internal_tensor_assign_2_squeeze_mask_0, stride = new_cache_7_internal_tensor_assign_2_stride_0, x = reshape_35)[name = tensor("slice_by_index_7")]; + tensor reshape_36_shape_0 = const()[name = tensor("reshape_36_shape_0"), val = tensor([-1])]; + tensor reshape_36 = reshape(shape = reshape_36_shape_0, x = slice_by_index_7)[name = tensor("reshape_36")]; + tensor reshape_37_shape_0 = const()[name = tensor("reshape_37_shape_0"), val = tensor([-1])]; + tensor reshape_37_cast_fp16 = reshape(shape = reshape_37_shape_0, x = var_1599_cast_fp16)[name = tensor("reshape_37_cast_fp16")]; + tensor reshape_38_shape_0 = const()[name = tensor("reshape_38_shape_0"), val = tensor([-1])]; + tensor reshape_38_cast_fp16 = reshape(shape = reshape_38_shape_0, x = reshape_34_cast_fp16)[name = tensor("reshape_38_cast_fp16")]; + tensor scatter_7_mode_0 = const()[name = tensor("scatter_7_mode_0"), val = tensor("update")]; + tensor scatter_7_axis_0 = const()[name = tensor("scatter_7_axis_0"), val = tensor(0)]; + tensor scatter_7_validate_indices_0 = const()[name = tensor("scatter_7_validate_indices_0"), val = tensor(false)]; + tensor scatter_7_cast_fp16 = scatter(axis = scatter_7_axis_0, data = reshape_38_cast_fp16, indices = reshape_36, mode = scatter_7_mode_0, updates = reshape_37_cast_fp16, validate_indices = scatter_7_validate_indices_0)[name = tensor("scatter_7_cast_fp16")]; + tensor reshape_39_cast_fp16 = reshape(shape = shape_19, x = scatter_7_cast_fp16)[name = tensor("reshape_39_cast_fp16")]; + tensor reshape_39_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("reshape_39_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor keys_19_begin_0 = const()[name = tensor("keys_19_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor keys_19_end_0 = const()[name = tensor("keys_19_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor keys_19_end_mask_0 = const()[name = tensor("keys_19_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor keys_19_squeeze_mask_0 = const()[name = tensor("keys_19_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor keys_19_cast_fp16 = slice_by_index(begin = keys_19_begin_0, end = keys_19_end_0, end_mask = keys_19_end_mask_0, squeeze_mask = keys_19_squeeze_mask_0, x = reshape_39_cast_fp16)[name = tensor("keys_19_cast_fp16")]; + tensor values_19_begin_0 = const()[name = tensor("values_19_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor values_19_end_0 = const()[name = tensor("values_19_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor values_19_end_mask_0 = const()[name = tensor("values_19_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor values_19_squeeze_mask_0 = const()[name = tensor("values_19_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor values_19_cast_fp16 = slice_by_index(begin = values_19_begin_0, end = values_19_end_0, end_mask = values_19_end_mask_0, squeeze_mask = values_19_squeeze_mask_0, x = reshape_39_cast_fp16)[name = tensor("values_19_cast_fp16")]; + tensor var_1611_cast_fp16 = not_equal(x = keys_19_cast_fp16, y = keys_19_cast_fp16)[name = tensor("op_1611_cast_fp16")]; + tensor keys_21_cast_fp16 = select(a = var_360_to_fp16, b = keys_19_cast_fp16, cond = var_1611_cast_fp16)[name = tensor("keys_21_cast_fp16")]; + tensor var_1619_cast_fp16 = not_equal(x = values_19_cast_fp16, y = values_19_cast_fp16)[name = tensor("op_1619_cast_fp16")]; + tensor values_21_cast_fp16 = select(a = var_360_to_fp16, b = values_19_cast_fp16, cond = var_1619_cast_fp16)[name = tensor("values_21_cast_fp16")]; + tensor var_1643 = const()[name = tensor("op_1643"), val = tensor([0, 2, 1, 3])]; + tensor var_1656 = const()[name = tensor("op_1656"), val = tensor([1, 1, 1])]; + tensor var_1657_cast_fp16 = reshape(shape = var_1656, x = position3_to_fp16)[name = tensor("op_1657_cast_fp16")]; + tensor var_1674_to_fp16 = const()[name = tensor("op_1674_to_fp16"), val = tensor(0x1p+0)]; + tensor valid_len_7_cast_fp16 = add(x = var_1657_cast_fp16, y = var_1674_to_fp16)[name = tensor("valid_len_7_cast_fp16")]; + tensor valid_mask_7_cast_fp16 = less(x = k_positions_1_promoted_to_fp16, y = valid_len_7_cast_fp16)[name = tensor("valid_mask_7_cast_fp16")]; + tensor causal_mask_7_cast_fp16 = less_equal(x = k_positions_1_promoted_to_fp16, y = var_1657_cast_fp16)[name = tensor("causal_mask_7_cast_fp16")]; + tensor attn_mask_13 = logical_and(x = valid_mask_7_cast_fp16, y = causal_mask_7_cast_fp16)[name = tensor("attn_mask_13")]; + tensor attn_mask_15_axes_0 = const()[name = tensor("attn_mask_15_axes_0"), val = tensor([1])]; + tensor attn_mask_15 = expand_dims(axes = attn_mask_15_axes_0, x = attn_mask_13)[name = tensor("attn_mask_15")]; + tensor var_1692_transpose_x_0 = const()[name = tensor("op_1692_transpose_x_0"), val = tensor(false)]; + tensor var_1692_transpose_y_0 = const()[name = tensor("op_1692_transpose_y_0"), val = tensor(false)]; + tensor transpose_24_perm_0 = const()[name = tensor("transpose_24_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_25_perm_0 = const()[name = tensor("transpose_25_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_25 = transpose(perm = transpose_25_perm_0, x = keys_21_cast_fp16)[name = tensor("transpose_39")]; + tensor transpose_24 = transpose(perm = transpose_24_perm_0, x = q_21_cast_fp16)[name = tensor("transpose_40")]; + tensor var_1692_cast_fp16 = matmul(transpose_x = var_1692_transpose_x_0, transpose_y = var_1692_transpose_y_0, x = transpose_24, y = transpose_25)[name = tensor("op_1692_cast_fp16")]; + tensor var_1686_to_fp16 = const()[name = tensor("op_1686_to_fp16"), val = tensor([0x1p-3])]; + tensor attn_weights_19_cast_fp16 = mul(x = var_1692_cast_fp16, y = var_1686_to_fp16)[name = tensor("attn_weights_19_cast_fp16")]; + tensor var_1694 = logical_not(x = attn_mask_15)[name = tensor("op_1694")]; + tensor var_1695_to_fp16 = const()[name = tensor("op_1695_to_fp16"), val = tensor(-inf)]; + tensor attn_weights_21_cast_fp16 = select(a = var_1695_to_fp16, b = attn_weights_19_cast_fp16, cond = var_1694)[name = tensor("attn_weights_21_cast_fp16")]; + tensor var_1697 = const()[name = tensor("op_1697"), val = tensor(-1)]; + tensor attn_weights_23_cast_fp16 = softmax(axis = var_1697, x = attn_weights_21_cast_fp16)[name = tensor("attn_weights_23_cast_fp16")]; + tensor attn_output_7_transpose_x_0 = const()[name = tensor("attn_output_7_transpose_x_0"), val = tensor(false)]; + tensor attn_output_7_transpose_y_0 = const()[name = tensor("attn_output_7_transpose_y_0"), val = tensor(false)]; + tensor values_23_cast_fp16 = transpose(perm = var_1643, x = values_21_cast_fp16)[name = tensor("transpose_41")]; + tensor attn_output_7_cast_fp16 = matmul(transpose_x = attn_output_7_transpose_x_0, transpose_y = attn_output_7_transpose_y_0, x = attn_weights_23_cast_fp16, y = values_23_cast_fp16)[name = tensor("attn_output_7_cast_fp16")]; + tensor var_1705 = const()[name = tensor("op_1705"), val = tensor([0, 2, 1, 3])]; + tensor var_1708 = const()[name = tensor("op_1708"), val = tensor([1, 1, 1024])]; + tensor var_1706_cast_fp16 = transpose(perm = var_1705, x = attn_output_7_cast_fp16)[name = tensor("transpose_38")]; + tensor input_35_cast_fp16 = reshape(shape = var_1708, x = var_1706_cast_fp16)[name = tensor("input_35_cast_fp16")]; + tensor linear_14_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = attn3_out_proj_weight_cast_fp16, x = input_35_cast_fp16)[name = tensor("linear_14_cast_fp16")]; + tensor var_1714_to_fp16 = const()[name = tensor("op_1714_to_fp16"), val = tensor(0x1p+0)]; + tensor var_1715_cast_fp16 = add(x = position3_to_fp16, y = var_1714_to_fp16)[name = tensor("op_1715_cast_fp16")]; + tensor var_1715_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_1715_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor input_37_cast_fp16 = add(x = input_33_cast_fp16, y = linear_14_cast_fp16)[name = tensor("input_37_cast_fp16")]; + tensor input_39_axes_0 = const()[name = tensor("input_39_axes_0"), val = tensor([-1])]; + tensor norm3_2_weight_to_fp16 = const()[name = tensor("norm3_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(76781568)))]; + tensor norm3_2_bias_to_fp16 = const()[name = tensor("norm3_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(76783680)))]; + tensor var_1719_to_fp16 = const()[name = tensor("op_1719_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_39_cast_fp16 = layer_norm(axes = input_39_axes_0, beta = norm3_2_bias_to_fp16, epsilon = var_1719_to_fp16, gamma = norm3_2_weight_to_fp16, x = input_37_cast_fp16)[name = tensor("input_39_cast_fp16")]; + tensor linear_15_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = linear3_1_weight_cast_fp16, x = input_39_cast_fp16)[name = tensor("linear_15_cast_fp16")]; + tensor input_41_mode_0 = const()[name = tensor("input_41_mode_0"), val = tensor("EXACT")]; + tensor input_41_cast_fp16 = gelu(mode = input_41_mode_0, x = linear_15_cast_fp16)[name = tensor("input_41_cast_fp16")]; + tensor linear_16_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = linear3_2_weight_cast_fp16, x = input_41_cast_fp16)[name = tensor("linear_16_cast_fp16")]; + tensor input_43_cast_fp16 = add(x = input_37_cast_fp16, y = linear_16_cast_fp16)[name = tensor("input_43_cast_fp16")]; + tensor x_9_axes_0 = const()[name = tensor("x_9_axes_0"), val = tensor([-1])]; + tensor norm4_1_weight_to_fp16 = const()[name = tensor("norm4_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(76785792)))]; + tensor norm4_1_bias_to_fp16 = const()[name = tensor("norm4_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(76787904)))]; + tensor var_1736_to_fp16 = const()[name = tensor("op_1736_to_fp16"), val = tensor(0x1.5p-17)]; + tensor x_9_cast_fp16 = layer_norm(axes = x_9_axes_0, beta = norm4_1_bias_to_fp16, epsilon = var_1736_to_fp16, gamma = norm4_1_weight_to_fp16, x = input_43_cast_fp16)[name = tensor("x_9_cast_fp16")]; + tensor linear_17_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = attn4_in_proj_weight_cast_fp16, x = x_9_cast_fp16)[name = tensor("linear_17_cast_fp16")]; + tensor var_1772 = const()[name = tensor("op_1772"), val = tensor([1, 1, 3, 16, 64])]; + tensor qkv_9_cast_fp16 = reshape(shape = var_1772, x = linear_17_cast_fp16)[name = tensor("qkv_9_cast_fp16")]; + tensor q_25_begin_0 = const()[name = tensor("q_25_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor q_25_end_0 = const()[name = tensor("q_25_end_0"), val = tensor([1, 1, 1, 16, 64])]; + tensor q_25_end_mask_0 = const()[name = tensor("q_25_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor q_25_squeeze_mask_0 = const()[name = tensor("q_25_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor q_25_cast_fp16 = slice_by_index(begin = q_25_begin_0, end = q_25_end_0, end_mask = q_25_end_mask_0, squeeze_mask = q_25_squeeze_mask_0, x = qkv_9_cast_fp16)[name = tensor("q_25_cast_fp16")]; + tensor k_17_begin_0 = const()[name = tensor("k_17_begin_0"), val = tensor([0, 0, 1, 0, 0])]; + tensor k_17_end_0 = const()[name = tensor("k_17_end_0"), val = tensor([1, 1, 2, 16, 64])]; + tensor k_17_end_mask_0 = const()[name = tensor("k_17_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor k_17_squeeze_mask_0 = const()[name = tensor("k_17_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor k_17_cast_fp16 = slice_by_index(begin = k_17_begin_0, end = k_17_end_0, end_mask = k_17_end_mask_0, squeeze_mask = k_17_squeeze_mask_0, x = qkv_9_cast_fp16)[name = tensor("k_17_cast_fp16")]; + tensor v_9_begin_0 = const()[name = tensor("v_9_begin_0"), val = tensor([0, 0, 2, 0, 0])]; + tensor v_9_end_0 = const()[name = tensor("v_9_end_0"), val = tensor([1, 1, 3, 16, 64])]; + tensor v_9_end_mask_0 = const()[name = tensor("v_9_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor v_9_squeeze_mask_0 = const()[name = tensor("v_9_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor v_9_cast_fp16 = slice_by_index(begin = v_9_begin_0, end = v_9_end_0, end_mask = v_9_end_mask_0, squeeze_mask = v_9_squeeze_mask_0, x = qkv_9_cast_fp16)[name = tensor("v_9_cast_fp16")]; + tensor var_1876 = const()[name = tensor("op_1876"), val = tensor([1, 1, 1, 1])]; + tensor position4_to_fp16_dtype_0 = const()[name = tensor("position4_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor position4_to_fp16 = cast(dtype = position4_to_fp16_dtype_0, x = position4)[name = tensor("cast_119")]; + tensor ts_29_cast_fp16 = reshape(shape = var_1876, x = position4_to_fp16)[name = tensor("ts_29_cast_fp16")]; + tensor var_1880 = const()[name = tensor("op_1880"), val = tensor([1, 1, 16, 32, 2])]; + tensor q_complex_9_cast_fp16 = reshape(shape = var_1880, x = q_25_cast_fp16)[name = tensor("q_complex_9_cast_fp16")]; + tensor var_1884 = const()[name = tensor("op_1884"), val = tensor([1, 1, 16, 32, 2])]; + tensor k_complex_9_cast_fp16 = reshape(shape = var_1884, x = k_17_cast_fp16)[name = tensor("k_complex_9_cast_fp16")]; + tensor var_1888_begin_0 = const()[name = tensor("op_1888_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_1888_end_0 = const()[name = tensor("op_1888_end_0"), val = tensor([1, 1, 16, 32, 1])]; + tensor var_1888_end_mask_0 = const()[name = tensor("op_1888_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_1888_squeeze_mask_0 = const()[name = tensor("op_1888_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_1888_cast_fp16 = slice_by_index(begin = var_1888_begin_0, end = var_1888_end_0, end_mask = var_1888_end_mask_0, squeeze_mask = var_1888_squeeze_mask_0, x = q_complex_9_cast_fp16)[name = tensor("op_1888_cast_fp16")]; + tensor var_1896_begin_0 = const()[name = tensor("op_1896_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_1896_end_0 = const()[name = tensor("op_1896_end_0"), val = tensor([1, 1, 16, 32, 2])]; + tensor var_1896_end_mask_0 = const()[name = tensor("op_1896_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_1896_squeeze_mask_0 = const()[name = tensor("op_1896_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_1896_cast_fp16 = slice_by_index(begin = var_1896_begin_0, end = var_1896_end_0, end_mask = var_1896_end_mask_0, squeeze_mask = var_1896_squeeze_mask_0, x = q_complex_9_cast_fp16)[name = tensor("op_1896_cast_fp16")]; + tensor var_1904_begin_0 = const()[name = tensor("op_1904_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_1904_end_0 = const()[name = tensor("op_1904_end_0"), val = tensor([1, 1, 16, 32, 1])]; + tensor var_1904_end_mask_0 = const()[name = tensor("op_1904_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_1904_squeeze_mask_0 = const()[name = tensor("op_1904_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_1904_cast_fp16 = slice_by_index(begin = var_1904_begin_0, end = var_1904_end_0, end_mask = var_1904_end_mask_0, squeeze_mask = var_1904_squeeze_mask_0, x = k_complex_9_cast_fp16)[name = tensor("op_1904_cast_fp16")]; + tensor var_1912_begin_0 = const()[name = tensor("op_1912_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_1912_end_0 = const()[name = tensor("op_1912_end_0"), val = tensor([1, 1, 16, 32, 2])]; + tensor var_1912_end_mask_0 = const()[name = tensor("op_1912_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_1912_squeeze_mask_0 = const()[name = tensor("op_1912_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_1912_cast_fp16 = slice_by_index(begin = var_1912_begin_0, end = var_1912_end_0, end_mask = var_1912_end_mask_0, squeeze_mask = var_1912_squeeze_mask_0, x = k_complex_9_cast_fp16)[name = tensor("op_1912_cast_fp16")]; + tensor freqs_9_to_fp16 = const()[name = tensor("freqs_9_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(76790016)))]; + tensor var_1918_cast_fp16 = mul(x = freqs_9_to_fp16, y = ts_29_cast_fp16)[name = tensor("op_1918_cast_fp16")]; + tensor rotr_9_cast_fp16 = cos(x = var_1918_cast_fp16)[name = tensor("rotr_9_cast_fp16")]; + tensor roti_9_cast_fp16 = sin(x = var_1918_cast_fp16)[name = tensor("roti_9_cast_fp16")]; + tensor var_1922_cast_fp16 = mul(x = var_1888_cast_fp16, y = rotr_9_cast_fp16)[name = tensor("op_1922_cast_fp16")]; + tensor var_1923_cast_fp16 = mul(x = var_1896_cast_fp16, y = roti_9_cast_fp16)[name = tensor("op_1923_cast_fp16")]; + tensor qor_17_cast_fp16 = sub(x = var_1922_cast_fp16, y = var_1923_cast_fp16)[name = tensor("qor_17_cast_fp16")]; + tensor var_1926_cast_fp16 = mul(x = var_1888_cast_fp16, y = roti_9_cast_fp16)[name = tensor("op_1926_cast_fp16")]; + tensor var_1927_cast_fp16 = mul(x = var_1896_cast_fp16, y = rotr_9_cast_fp16)[name = tensor("op_1927_cast_fp16")]; + tensor qoi_17_cast_fp16 = add(x = var_1926_cast_fp16, y = var_1927_cast_fp16)[name = tensor("qoi_17_cast_fp16")]; + tensor var_1930_cast_fp16 = mul(x = var_1904_cast_fp16, y = rotr_9_cast_fp16)[name = tensor("op_1930_cast_fp16")]; + tensor var_1931_cast_fp16 = mul(x = var_1912_cast_fp16, y = roti_9_cast_fp16)[name = tensor("op_1931_cast_fp16")]; + tensor kor_17_cast_fp16 = sub(x = var_1930_cast_fp16, y = var_1931_cast_fp16)[name = tensor("kor_17_cast_fp16")]; + tensor var_1934_cast_fp16 = mul(x = var_1904_cast_fp16, y = roti_9_cast_fp16)[name = tensor("op_1934_cast_fp16")]; + tensor var_1935_cast_fp16 = mul(x = var_1912_cast_fp16, y = rotr_9_cast_fp16)[name = tensor("op_1935_cast_fp16")]; + tensor koi_17_cast_fp16 = add(x = var_1934_cast_fp16, y = var_1935_cast_fp16)[name = tensor("koi_17_cast_fp16")]; + tensor qo_9_axis_0 = const()[name = tensor("qo_9_axis_0"), val = tensor(-1)]; + tensor qo_9_cast_fp16 = stack(axis = qo_9_axis_0, values = (qor_17_cast_fp16, qoi_17_cast_fp16))[name = tensor("qo_9_cast_fp16")]; + tensor ko_9_axis_0 = const()[name = tensor("ko_9_axis_0"), val = tensor(-1)]; + tensor ko_9_cast_fp16 = stack(axis = ko_9_axis_0, values = (kor_17_cast_fp16, koi_17_cast_fp16))[name = tensor("ko_9_cast_fp16")]; + tensor var_1964 = const()[name = tensor("op_1964"), val = tensor([1, 1, 16, 64])]; + tensor q_27_cast_fp16 = reshape(shape = var_1964, x = qo_9_cast_fp16)[name = tensor("q_27_cast_fp16")]; + tensor var_1966 = const()[name = tensor("op_1966"), val = tensor([1, 1, 16, 64])]; + tensor k_19_cast_fp16 = reshape(shape = var_1966, x = ko_9_cast_fp16)[name = tensor("k_19_cast_fp16")]; + tensor _inversed_1988_y_0_to_fp16 = const()[name = tensor("_inversed_1988_y_0_to_fp16"), val = tensor(0x1p-9)]; + tensor _inversed_1988_cast_fp16 = mul(x = ts_29_cast_fp16, y = _inversed_1988_y_0_to_fp16)[name = tensor("_inversed_1988_cast_fp16")]; + tensor var_1989_cast_fp16 = floor(x = _inversed_1988_cast_fp16)[name = tensor("op_1989_cast_fp16")]; + tensor var_1990_to_fp16 = const()[name = tensor("op_1990_to_fp16"), val = tensor(0x1p+9)]; + tensor var_1991_cast_fp16 = mul(x = var_1989_cast_fp16, y = var_1990_to_fp16)[name = tensor("op_1991_cast_fp16")]; + tensor write_indices_float_19_cast_fp16 = sub(x = ts_29_cast_fp16, y = var_1991_cast_fp16)[name = tensor("write_indices_float_19_cast_fp16")]; + tensor var_1998_dtype_0 = const()[name = tensor("op_1998_dtype_0"), val = tensor("int32")]; + tensor write_indices_9_reps_0 = const()[name = tensor("write_indices_9_reps_0"), val = tensor([1, 1, 16, 64])]; + tensor write_indices_float_19_cast_fp16_to_int32 = cast(dtype = var_1998_dtype_0, x = write_indices_float_19_cast_fp16)[name = tensor("cast_118")]; + tensor write_indices_9 = tile(reps = write_indices_9_reps_0, x = write_indices_float_19_cast_fp16_to_int32)[name = tensor("write_indices_9")]; + tensor var_2006_begin_0 = const()[name = tensor("op_2006_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_2006_end_0 = const()[name = tensor("op_2006_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor var_2006_end_mask_0 = const()[name = tensor("op_2006_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_2006_squeeze_mask_0 = const()[name = tensor("op_2006_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor cache4_to_fp16_dtype_0 = const()[name = tensor("cache4_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor cache4_to_fp16 = cast(dtype = cache4_to_fp16_dtype_0, x = cache4)[name = tensor("cast_117")]; + tensor var_2006_cast_fp16 = slice_by_index(begin = var_2006_begin_0, end = var_2006_end_0, end_mask = var_2006_end_mask_0, squeeze_mask = var_2006_squeeze_mask_0, x = cache4_to_fp16)[name = tensor("op_2006_cast_fp16")]; + tensor var_2008_axis_0 = const()[name = tensor("op_2008_axis_0"), val = tensor(1)]; + tensor var_2008_mode_0 = const()[name = tensor("op_2008_mode_0"), val = tensor("update")]; + tensor var_2008_validate_indices_0 = const()[name = tensor("op_2008_validate_indices_0"), val = tensor(false)]; + tensor var_2008_cast_fp16 = scatter_along_axis(axis = var_2008_axis_0, data = var_2006_cast_fp16, indices = write_indices_9, mode = var_2008_mode_0, updates = k_19_cast_fp16, validate_indices = var_2008_validate_indices_0)[name = tensor("op_2008_cast_fp16")]; + tensor concat_30 = const()[name = tensor("concat_30"), val = tensor([0, 0, 0, 0, 0])]; + tensor concat_31 = const()[name = tensor("concat_31"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_9_internal_tensor_assign_1_stride_0 = const()[name = tensor("new_cache_9_internal_tensor_assign_1_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_9_internal_tensor_assign_1_begin_mask_0 = const()[name = tensor("new_cache_9_internal_tensor_assign_1_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_9_internal_tensor_assign_1_end_mask_0 = const()[name = tensor("new_cache_9_internal_tensor_assign_1_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_9_internal_tensor_assign_1_squeeze_mask_0 = const()[name = tensor("new_cache_9_internal_tensor_assign_1_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_20 = const()[name = tensor("shape_20"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_8 = const()[name = tensor("reduce_prod_8"), val = tensor(1048576)]; + tensor range_1d_8_start_0 = const()[name = tensor("range_1d_8_start_0"), val = tensor(0)]; + tensor range_1d_8_step_0 = const()[name = tensor("range_1d_8_step_0"), val = tensor(1)]; + tensor range_1d_8 = range_1d(end = reduce_prod_8, start = range_1d_8_start_0, step = range_1d_8_step_0)[name = tensor("range_1d_8")]; + tensor reshape_40 = reshape(shape = shape_20, x = range_1d_8)[name = tensor("reshape_40")]; + tensor slice_by_index_8 = slice_by_index(begin = concat_30, begin_mask = new_cache_9_internal_tensor_assign_1_begin_mask_0, end = concat_31, end_mask = new_cache_9_internal_tensor_assign_1_end_mask_0, squeeze_mask = new_cache_9_internal_tensor_assign_1_squeeze_mask_0, stride = new_cache_9_internal_tensor_assign_1_stride_0, x = reshape_40)[name = tensor("slice_by_index_8")]; + tensor reshape_41_shape_0 = const()[name = tensor("reshape_41_shape_0"), val = tensor([-1])]; + tensor reshape_41 = reshape(shape = reshape_41_shape_0, x = slice_by_index_8)[name = tensor("reshape_41")]; + tensor reshape_42_shape_0 = const()[name = tensor("reshape_42_shape_0"), val = tensor([-1])]; + tensor reshape_42_cast_fp16 = reshape(shape = reshape_42_shape_0, x = var_2008_cast_fp16)[name = tensor("reshape_42_cast_fp16")]; + tensor reshape_43_shape_0 = const()[name = tensor("reshape_43_shape_0"), val = tensor([-1])]; + tensor reshape_43_cast_fp16 = reshape(shape = reshape_43_shape_0, x = cache4_to_fp16)[name = tensor("reshape_43_cast_fp16")]; + tensor scatter_8_mode_0 = const()[name = tensor("scatter_8_mode_0"), val = tensor("update")]; + tensor scatter_8_axis_0 = const()[name = tensor("scatter_8_axis_0"), val = tensor(0)]; + tensor scatter_8_validate_indices_0 = const()[name = tensor("scatter_8_validate_indices_0"), val = tensor(false)]; + tensor scatter_8_cast_fp16 = scatter(axis = scatter_8_axis_0, data = reshape_43_cast_fp16, indices = reshape_41, mode = scatter_8_mode_0, updates = reshape_42_cast_fp16, validate_indices = scatter_8_validate_indices_0)[name = tensor("scatter_8_cast_fp16")]; + tensor reshape_44_cast_fp16 = reshape(shape = shape_20, x = scatter_8_cast_fp16)[name = tensor("reshape_44_cast_fp16")]; + tensor var_2016_begin_0 = const()[name = tensor("op_2016_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor var_2016_end_0 = const()[name = tensor("op_2016_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor var_2016_end_mask_0 = const()[name = tensor("op_2016_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_2016_squeeze_mask_0 = const()[name = tensor("op_2016_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor var_2016_cast_fp16 = slice_by_index(begin = var_2016_begin_0, end = var_2016_end_0, end_mask = var_2016_end_mask_0, squeeze_mask = var_2016_squeeze_mask_0, x = reshape_44_cast_fp16)[name = tensor("op_2016_cast_fp16")]; + tensor var_2018_axis_0 = const()[name = tensor("op_2018_axis_0"), val = tensor(1)]; + tensor var_2018_mode_0 = const()[name = tensor("op_2018_mode_0"), val = tensor("update")]; + tensor var_2018_validate_indices_0 = const()[name = tensor("op_2018_validate_indices_0"), val = tensor(false)]; + tensor var_2018_cast_fp16 = scatter_along_axis(axis = var_2018_axis_0, data = var_2016_cast_fp16, indices = write_indices_9, mode = var_2018_mode_0, updates = v_9_cast_fp16, validate_indices = var_2018_validate_indices_0)[name = tensor("op_2018_cast_fp16")]; + tensor concat_32 = const()[name = tensor("concat_32"), val = tensor([1, 0, 0, 0, 0])]; + tensor concat_33 = const()[name = tensor("concat_33"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_9_internal_tensor_assign_2_stride_0 = const()[name = tensor("new_cache_9_internal_tensor_assign_2_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_9_internal_tensor_assign_2_begin_mask_0 = const()[name = tensor("new_cache_9_internal_tensor_assign_2_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_9_internal_tensor_assign_2_end_mask_0 = const()[name = tensor("new_cache_9_internal_tensor_assign_2_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_9_internal_tensor_assign_2_squeeze_mask_0 = const()[name = tensor("new_cache_9_internal_tensor_assign_2_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_21 = const()[name = tensor("shape_21"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_9 = const()[name = tensor("reduce_prod_9"), val = tensor(1048576)]; + tensor range_1d_9_start_0 = const()[name = tensor("range_1d_9_start_0"), val = tensor(0)]; + tensor range_1d_9_step_0 = const()[name = tensor("range_1d_9_step_0"), val = tensor(1)]; + tensor range_1d_9 = range_1d(end = reduce_prod_9, start = range_1d_9_start_0, step = range_1d_9_step_0)[name = tensor("range_1d_9")]; + tensor reshape_45 = reshape(shape = shape_21, x = range_1d_9)[name = tensor("reshape_45")]; + tensor slice_by_index_9 = slice_by_index(begin = concat_32, begin_mask = new_cache_9_internal_tensor_assign_2_begin_mask_0, end = concat_33, end_mask = new_cache_9_internal_tensor_assign_2_end_mask_0, squeeze_mask = new_cache_9_internal_tensor_assign_2_squeeze_mask_0, stride = new_cache_9_internal_tensor_assign_2_stride_0, x = reshape_45)[name = tensor("slice_by_index_9")]; + tensor reshape_46_shape_0 = const()[name = tensor("reshape_46_shape_0"), val = tensor([-1])]; + tensor reshape_46 = reshape(shape = reshape_46_shape_0, x = slice_by_index_9)[name = tensor("reshape_46")]; + tensor reshape_47_shape_0 = const()[name = tensor("reshape_47_shape_0"), val = tensor([-1])]; + tensor reshape_47_cast_fp16 = reshape(shape = reshape_47_shape_0, x = var_2018_cast_fp16)[name = tensor("reshape_47_cast_fp16")]; + tensor reshape_48_shape_0 = const()[name = tensor("reshape_48_shape_0"), val = tensor([-1])]; + tensor reshape_48_cast_fp16 = reshape(shape = reshape_48_shape_0, x = reshape_44_cast_fp16)[name = tensor("reshape_48_cast_fp16")]; + tensor scatter_9_mode_0 = const()[name = tensor("scatter_9_mode_0"), val = tensor("update")]; + tensor scatter_9_axis_0 = const()[name = tensor("scatter_9_axis_0"), val = tensor(0)]; + tensor scatter_9_validate_indices_0 = const()[name = tensor("scatter_9_validate_indices_0"), val = tensor(false)]; + tensor scatter_9_cast_fp16 = scatter(axis = scatter_9_axis_0, data = reshape_48_cast_fp16, indices = reshape_46, mode = scatter_9_mode_0, updates = reshape_47_cast_fp16, validate_indices = scatter_9_validate_indices_0)[name = tensor("scatter_9_cast_fp16")]; + tensor reshape_49_cast_fp16 = reshape(shape = shape_21, x = scatter_9_cast_fp16)[name = tensor("reshape_49_cast_fp16")]; + tensor reshape_49_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("reshape_49_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor keys_25_begin_0 = const()[name = tensor("keys_25_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor keys_25_end_0 = const()[name = tensor("keys_25_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor keys_25_end_mask_0 = const()[name = tensor("keys_25_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor keys_25_squeeze_mask_0 = const()[name = tensor("keys_25_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor keys_25_cast_fp16 = slice_by_index(begin = keys_25_begin_0, end = keys_25_end_0, end_mask = keys_25_end_mask_0, squeeze_mask = keys_25_squeeze_mask_0, x = reshape_49_cast_fp16)[name = tensor("keys_25_cast_fp16")]; + tensor values_25_begin_0 = const()[name = tensor("values_25_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor values_25_end_0 = const()[name = tensor("values_25_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor values_25_end_mask_0 = const()[name = tensor("values_25_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor values_25_squeeze_mask_0 = const()[name = tensor("values_25_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor values_25_cast_fp16 = slice_by_index(begin = values_25_begin_0, end = values_25_end_0, end_mask = values_25_end_mask_0, squeeze_mask = values_25_squeeze_mask_0, x = reshape_49_cast_fp16)[name = tensor("values_25_cast_fp16")]; + tensor var_2030_cast_fp16 = not_equal(x = keys_25_cast_fp16, y = keys_25_cast_fp16)[name = tensor("op_2030_cast_fp16")]; + tensor keys_27_cast_fp16 = select(a = var_360_to_fp16, b = keys_25_cast_fp16, cond = var_2030_cast_fp16)[name = tensor("keys_27_cast_fp16")]; + tensor var_2038_cast_fp16 = not_equal(x = values_25_cast_fp16, y = values_25_cast_fp16)[name = tensor("op_2038_cast_fp16")]; + tensor values_27_cast_fp16 = select(a = var_360_to_fp16, b = values_25_cast_fp16, cond = var_2038_cast_fp16)[name = tensor("values_27_cast_fp16")]; + tensor var_2062 = const()[name = tensor("op_2062"), val = tensor([0, 2, 1, 3])]; + tensor var_2075 = const()[name = tensor("op_2075"), val = tensor([1, 1, 1])]; + tensor var_2076_cast_fp16 = reshape(shape = var_2075, x = position4_to_fp16)[name = tensor("op_2076_cast_fp16")]; + tensor var_2093_to_fp16 = const()[name = tensor("op_2093_to_fp16"), val = tensor(0x1p+0)]; + tensor valid_len_9_cast_fp16 = add(x = var_2076_cast_fp16, y = var_2093_to_fp16)[name = tensor("valid_len_9_cast_fp16")]; + tensor valid_mask_9_cast_fp16 = less(x = k_positions_1_promoted_to_fp16, y = valid_len_9_cast_fp16)[name = tensor("valid_mask_9_cast_fp16")]; + tensor causal_mask_9_cast_fp16 = less_equal(x = k_positions_1_promoted_to_fp16, y = var_2076_cast_fp16)[name = tensor("causal_mask_9_cast_fp16")]; + tensor attn_mask_17 = logical_and(x = valid_mask_9_cast_fp16, y = causal_mask_9_cast_fp16)[name = tensor("attn_mask_17")]; + tensor attn_mask_19_axes_0 = const()[name = tensor("attn_mask_19_axes_0"), val = tensor([1])]; + tensor attn_mask_19 = expand_dims(axes = attn_mask_19_axes_0, x = attn_mask_17)[name = tensor("attn_mask_19")]; + tensor var_2111_transpose_x_0 = const()[name = tensor("op_2111_transpose_x_0"), val = tensor(false)]; + tensor var_2111_transpose_y_0 = const()[name = tensor("op_2111_transpose_y_0"), val = tensor(false)]; + tensor transpose_26_perm_0 = const()[name = tensor("transpose_26_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_27_perm_0 = const()[name = tensor("transpose_27_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_27 = transpose(perm = transpose_27_perm_0, x = keys_27_cast_fp16)[name = tensor("transpose_35")]; + tensor transpose_26 = transpose(perm = transpose_26_perm_0, x = q_27_cast_fp16)[name = tensor("transpose_36")]; + tensor var_2111_cast_fp16 = matmul(transpose_x = var_2111_transpose_x_0, transpose_y = var_2111_transpose_y_0, x = transpose_26, y = transpose_27)[name = tensor("op_2111_cast_fp16")]; + tensor var_2105_to_fp16 = const()[name = tensor("op_2105_to_fp16"), val = tensor([0x1p-3])]; + tensor attn_weights_25_cast_fp16 = mul(x = var_2111_cast_fp16, y = var_2105_to_fp16)[name = tensor("attn_weights_25_cast_fp16")]; + tensor var_2113 = logical_not(x = attn_mask_19)[name = tensor("op_2113")]; + tensor var_2114_to_fp16 = const()[name = tensor("op_2114_to_fp16"), val = tensor(-inf)]; + tensor attn_weights_27_cast_fp16 = select(a = var_2114_to_fp16, b = attn_weights_25_cast_fp16, cond = var_2113)[name = tensor("attn_weights_27_cast_fp16")]; + tensor var_2116 = const()[name = tensor("op_2116"), val = tensor(-1)]; + tensor attn_weights_29_cast_fp16 = softmax(axis = var_2116, x = attn_weights_27_cast_fp16)[name = tensor("attn_weights_29_cast_fp16")]; + tensor attn_output_9_transpose_x_0 = const()[name = tensor("attn_output_9_transpose_x_0"), val = tensor(false)]; + tensor attn_output_9_transpose_y_0 = const()[name = tensor("attn_output_9_transpose_y_0"), val = tensor(false)]; + tensor values_29_cast_fp16 = transpose(perm = var_2062, x = values_27_cast_fp16)[name = tensor("transpose_37")]; + tensor attn_output_9_cast_fp16 = matmul(transpose_x = attn_output_9_transpose_x_0, transpose_y = attn_output_9_transpose_y_0, x = attn_weights_29_cast_fp16, y = values_29_cast_fp16)[name = tensor("attn_output_9_cast_fp16")]; + tensor var_2124 = const()[name = tensor("op_2124"), val = tensor([0, 2, 1, 3])]; + tensor var_2127 = const()[name = tensor("op_2127"), val = tensor([1, 1, 1024])]; + tensor var_2125_cast_fp16 = transpose(perm = var_2124, x = attn_output_9_cast_fp16)[name = tensor("transpose_34")]; + tensor input_45_cast_fp16 = reshape(shape = var_2127, x = var_2125_cast_fp16)[name = tensor("input_45_cast_fp16")]; + tensor linear_18_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = attn4_out_proj_weight_cast_fp16, x = input_45_cast_fp16)[name = tensor("linear_18_cast_fp16")]; + tensor var_2133_to_fp16 = const()[name = tensor("op_2133_to_fp16"), val = tensor(0x1p+0)]; + tensor var_2134_cast_fp16 = add(x = position4_to_fp16, y = var_2133_to_fp16)[name = tensor("op_2134_cast_fp16")]; + tensor var_2134_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_2134_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor input_47_cast_fp16 = add(x = input_43_cast_fp16, y = linear_18_cast_fp16)[name = tensor("input_47_cast_fp16")]; + tensor input_49_axes_0 = const()[name = tensor("input_49_axes_0"), val = tensor([-1])]; + tensor norm4_2_weight_to_fp16 = const()[name = tensor("norm4_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(76790144)))]; + tensor norm4_2_bias_to_fp16 = const()[name = tensor("norm4_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(76792256)))]; + tensor var_2138_to_fp16 = const()[name = tensor("op_2138_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_49_cast_fp16 = layer_norm(axes = input_49_axes_0, beta = norm4_2_bias_to_fp16, epsilon = var_2138_to_fp16, gamma = norm4_2_weight_to_fp16, x = input_47_cast_fp16)[name = tensor("input_49_cast_fp16")]; + tensor linear_19_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = linear4_1_weight_cast_fp16, x = input_49_cast_fp16)[name = tensor("linear_19_cast_fp16")]; + tensor input_51_mode_0 = const()[name = tensor("input_51_mode_0"), val = tensor("EXACT")]; + tensor input_51_cast_fp16 = gelu(mode = input_51_mode_0, x = linear_19_cast_fp16)[name = tensor("input_51_cast_fp16")]; + tensor linear_20_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = linear4_2_weight_cast_fp16, x = input_51_cast_fp16)[name = tensor("linear_20_cast_fp16")]; + tensor input_53_cast_fp16 = add(x = input_47_cast_fp16, y = linear_20_cast_fp16)[name = tensor("input_53_cast_fp16")]; + tensor x_axes_0 = const()[name = tensor("x_axes_0"), val = tensor([-1])]; + tensor norm5_1_weight_to_fp16 = const()[name = tensor("norm5_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(76794368)))]; + tensor norm5_1_bias_to_fp16 = const()[name = tensor("norm5_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(76796480)))]; + tensor var_2155_to_fp16 = const()[name = tensor("op_2155_to_fp16"), val = tensor(0x1.5p-17)]; + tensor x_cast_fp16 = layer_norm(axes = x_axes_0, beta = norm5_1_bias_to_fp16, epsilon = var_2155_to_fp16, gamma = norm5_1_weight_to_fp16, x = input_53_cast_fp16)[name = tensor("x_cast_fp16")]; + tensor linear_21_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = attn5_in_proj_weight_cast_fp16, x = x_cast_fp16)[name = tensor("linear_21_cast_fp16")]; + tensor var_2191 = const()[name = tensor("op_2191"), val = tensor([1, 1, 3, 16, 64])]; + tensor qkv_cast_fp16 = reshape(shape = var_2191, x = linear_21_cast_fp16)[name = tensor("qkv_cast_fp16")]; + tensor q_31_begin_0 = const()[name = tensor("q_31_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor q_31_end_0 = const()[name = tensor("q_31_end_0"), val = tensor([1, 1, 1, 16, 64])]; + tensor q_31_end_mask_0 = const()[name = tensor("q_31_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor q_31_squeeze_mask_0 = const()[name = tensor("q_31_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor q_31_cast_fp16 = slice_by_index(begin = q_31_begin_0, end = q_31_end_0, end_mask = q_31_end_mask_0, squeeze_mask = q_31_squeeze_mask_0, x = qkv_cast_fp16)[name = tensor("q_31_cast_fp16")]; + tensor k_21_begin_0 = const()[name = tensor("k_21_begin_0"), val = tensor([0, 0, 1, 0, 0])]; + tensor k_21_end_0 = const()[name = tensor("k_21_end_0"), val = tensor([1, 1, 2, 16, 64])]; + tensor k_21_end_mask_0 = const()[name = tensor("k_21_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor k_21_squeeze_mask_0 = const()[name = tensor("k_21_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor k_21_cast_fp16 = slice_by_index(begin = k_21_begin_0, end = k_21_end_0, end_mask = k_21_end_mask_0, squeeze_mask = k_21_squeeze_mask_0, x = qkv_cast_fp16)[name = tensor("k_21_cast_fp16")]; + tensor v_begin_0 = const()[name = tensor("v_begin_0"), val = tensor([0, 0, 2, 0, 0])]; + tensor v_end_0 = const()[name = tensor("v_end_0"), val = tensor([1, 1, 3, 16, 64])]; + tensor v_end_mask_0 = const()[name = tensor("v_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor v_squeeze_mask_0 = const()[name = tensor("v_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor v_cast_fp16 = slice_by_index(begin = v_begin_0, end = v_end_0, end_mask = v_end_mask_0, squeeze_mask = v_squeeze_mask_0, x = qkv_cast_fp16)[name = tensor("v_cast_fp16")]; + tensor var_2295 = const()[name = tensor("op_2295"), val = tensor([1, 1, 1, 1])]; + tensor position5_to_fp16_dtype_0 = const()[name = tensor("position5_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor position5_to_fp16 = cast(dtype = position5_to_fp16_dtype_0, x = position5)[name = tensor("cast_114")]; + tensor ts_cast_fp16 = reshape(shape = var_2295, x = position5_to_fp16)[name = tensor("ts_cast_fp16")]; + tensor var_2299 = const()[name = tensor("op_2299"), val = tensor([1, 1, 16, 32, 2])]; + tensor q_complex_cast_fp16 = reshape(shape = var_2299, x = q_31_cast_fp16)[name = tensor("q_complex_cast_fp16")]; + tensor var_2303 = const()[name = tensor("op_2303"), val = tensor([1, 1, 16, 32, 2])]; + tensor k_complex_cast_fp16 = reshape(shape = var_2303, x = k_21_cast_fp16)[name = tensor("k_complex_cast_fp16")]; + tensor var_2307_begin_0 = const()[name = tensor("op_2307_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_2307_end_0 = const()[name = tensor("op_2307_end_0"), val = tensor([1, 1, 16, 32, 1])]; + tensor var_2307_end_mask_0 = const()[name = tensor("op_2307_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_2307_squeeze_mask_0 = const()[name = tensor("op_2307_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_2307_cast_fp16 = slice_by_index(begin = var_2307_begin_0, end = var_2307_end_0, end_mask = var_2307_end_mask_0, squeeze_mask = var_2307_squeeze_mask_0, x = q_complex_cast_fp16)[name = tensor("op_2307_cast_fp16")]; + tensor var_2315_begin_0 = const()[name = tensor("op_2315_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_2315_end_0 = const()[name = tensor("op_2315_end_0"), val = tensor([1, 1, 16, 32, 2])]; + tensor var_2315_end_mask_0 = const()[name = tensor("op_2315_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_2315_squeeze_mask_0 = const()[name = tensor("op_2315_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_2315_cast_fp16 = slice_by_index(begin = var_2315_begin_0, end = var_2315_end_0, end_mask = var_2315_end_mask_0, squeeze_mask = var_2315_squeeze_mask_0, x = q_complex_cast_fp16)[name = tensor("op_2315_cast_fp16")]; + tensor var_2323_begin_0 = const()[name = tensor("op_2323_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_2323_end_0 = const()[name = tensor("op_2323_end_0"), val = tensor([1, 1, 16, 32, 1])]; + tensor var_2323_end_mask_0 = const()[name = tensor("op_2323_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_2323_squeeze_mask_0 = const()[name = tensor("op_2323_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_2323_cast_fp16 = slice_by_index(begin = var_2323_begin_0, end = var_2323_end_0, end_mask = var_2323_end_mask_0, squeeze_mask = var_2323_squeeze_mask_0, x = k_complex_cast_fp16)[name = tensor("op_2323_cast_fp16")]; + tensor var_2331_begin_0 = const()[name = tensor("op_2331_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_2331_end_0 = const()[name = tensor("op_2331_end_0"), val = tensor([1, 1, 16, 32, 2])]; + tensor var_2331_end_mask_0 = const()[name = tensor("op_2331_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_2331_squeeze_mask_0 = const()[name = tensor("op_2331_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_2331_cast_fp16 = slice_by_index(begin = var_2331_begin_0, end = var_2331_end_0, end_mask = var_2331_end_mask_0, squeeze_mask = var_2331_squeeze_mask_0, x = k_complex_cast_fp16)[name = tensor("op_2331_cast_fp16")]; + tensor freqs_to_fp16 = const()[name = tensor("freqs_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(76798592)))]; + tensor var_2337_cast_fp16 = mul(x = freqs_to_fp16, y = ts_cast_fp16)[name = tensor("op_2337_cast_fp16")]; + tensor rotr_cast_fp16 = cos(x = var_2337_cast_fp16)[name = tensor("rotr_cast_fp16")]; + tensor roti_cast_fp16 = sin(x = var_2337_cast_fp16)[name = tensor("roti_cast_fp16")]; + tensor var_2341_cast_fp16 = mul(x = var_2307_cast_fp16, y = rotr_cast_fp16)[name = tensor("op_2341_cast_fp16")]; + tensor var_2342_cast_fp16 = mul(x = var_2315_cast_fp16, y = roti_cast_fp16)[name = tensor("op_2342_cast_fp16")]; + tensor qor_21_cast_fp16 = sub(x = var_2341_cast_fp16, y = var_2342_cast_fp16)[name = tensor("qor_21_cast_fp16")]; + tensor var_2345_cast_fp16 = mul(x = var_2307_cast_fp16, y = roti_cast_fp16)[name = tensor("op_2345_cast_fp16")]; + tensor var_2346_cast_fp16 = mul(x = var_2315_cast_fp16, y = rotr_cast_fp16)[name = tensor("op_2346_cast_fp16")]; + tensor qoi_21_cast_fp16 = add(x = var_2345_cast_fp16, y = var_2346_cast_fp16)[name = tensor("qoi_21_cast_fp16")]; + tensor var_2349_cast_fp16 = mul(x = var_2323_cast_fp16, y = rotr_cast_fp16)[name = tensor("op_2349_cast_fp16")]; + tensor var_2350_cast_fp16 = mul(x = var_2331_cast_fp16, y = roti_cast_fp16)[name = tensor("op_2350_cast_fp16")]; + tensor kor_21_cast_fp16 = sub(x = var_2349_cast_fp16, y = var_2350_cast_fp16)[name = tensor("kor_21_cast_fp16")]; + tensor var_2353_cast_fp16 = mul(x = var_2323_cast_fp16, y = roti_cast_fp16)[name = tensor("op_2353_cast_fp16")]; + tensor var_2354_cast_fp16 = mul(x = var_2331_cast_fp16, y = rotr_cast_fp16)[name = tensor("op_2354_cast_fp16")]; + tensor koi_21_cast_fp16 = add(x = var_2353_cast_fp16, y = var_2354_cast_fp16)[name = tensor("koi_21_cast_fp16")]; + tensor qo_axis_0 = const()[name = tensor("qo_axis_0"), val = tensor(-1)]; + tensor qo_cast_fp16 = stack(axis = qo_axis_0, values = (qor_21_cast_fp16, qoi_21_cast_fp16))[name = tensor("qo_cast_fp16")]; + tensor ko_axis_0 = const()[name = tensor("ko_axis_0"), val = tensor(-1)]; + tensor ko_cast_fp16 = stack(axis = ko_axis_0, values = (kor_21_cast_fp16, koi_21_cast_fp16))[name = tensor("ko_cast_fp16")]; + tensor var_2383 = const()[name = tensor("op_2383"), val = tensor([1, 1, 16, 64])]; + tensor q_33_cast_fp16 = reshape(shape = var_2383, x = qo_cast_fp16)[name = tensor("q_33_cast_fp16")]; + tensor var_2385 = const()[name = tensor("op_2385"), val = tensor([1, 1, 16, 64])]; + tensor k_cast_fp16 = reshape(shape = var_2385, x = ko_cast_fp16)[name = tensor("k_cast_fp16")]; + tensor _inversed_2407_y_0_to_fp16 = const()[name = tensor("_inversed_2407_y_0_to_fp16"), val = tensor(0x1p-9)]; + tensor _inversed_2407_cast_fp16 = mul(x = ts_cast_fp16, y = _inversed_2407_y_0_to_fp16)[name = tensor("_inversed_2407_cast_fp16")]; + tensor var_2408_cast_fp16 = floor(x = _inversed_2407_cast_fp16)[name = tensor("op_2408_cast_fp16")]; + tensor var_2409_to_fp16 = const()[name = tensor("op_2409_to_fp16"), val = tensor(0x1p+9)]; + tensor var_2410_cast_fp16 = mul(x = var_2408_cast_fp16, y = var_2409_to_fp16)[name = tensor("op_2410_cast_fp16")]; + tensor write_indices_float_cast_fp16 = sub(x = ts_cast_fp16, y = var_2410_cast_fp16)[name = tensor("write_indices_float_cast_fp16")]; + tensor var_2417_dtype_0 = const()[name = tensor("op_2417_dtype_0"), val = tensor("int32")]; + tensor write_indices_reps_0 = const()[name = tensor("write_indices_reps_0"), val = tensor([1, 1, 16, 64])]; + tensor write_indices_float_cast_fp16_to_int32 = cast(dtype = var_2417_dtype_0, x = write_indices_float_cast_fp16)[name = tensor("cast_113")]; + tensor write_indices = tile(reps = write_indices_reps_0, x = write_indices_float_cast_fp16_to_int32)[name = tensor("write_indices")]; + tensor var_2425_begin_0 = const()[name = tensor("op_2425_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_2425_end_0 = const()[name = tensor("op_2425_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor var_2425_end_mask_0 = const()[name = tensor("op_2425_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_2425_squeeze_mask_0 = const()[name = tensor("op_2425_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor cache5_to_fp16_dtype_0 = const()[name = tensor("cache5_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor cache5_to_fp16 = cast(dtype = cache5_to_fp16_dtype_0, x = cache5)[name = tensor("cast_112")]; + tensor var_2425_cast_fp16 = slice_by_index(begin = var_2425_begin_0, end = var_2425_end_0, end_mask = var_2425_end_mask_0, squeeze_mask = var_2425_squeeze_mask_0, x = cache5_to_fp16)[name = tensor("op_2425_cast_fp16")]; + tensor var_2427_axis_0 = const()[name = tensor("op_2427_axis_0"), val = tensor(1)]; + tensor var_2427_mode_0 = const()[name = tensor("op_2427_mode_0"), val = tensor("update")]; + tensor var_2427_validate_indices_0 = const()[name = tensor("op_2427_validate_indices_0"), val = tensor(false)]; + tensor var_2427_cast_fp16 = scatter_along_axis(axis = var_2427_axis_0, data = var_2425_cast_fp16, indices = write_indices, mode = var_2427_mode_0, updates = k_cast_fp16, validate_indices = var_2427_validate_indices_0)[name = tensor("op_2427_cast_fp16")]; + tensor concat_37 = const()[name = tensor("concat_37"), val = tensor([0, 0, 0, 0, 0])]; + tensor concat_38 = const()[name = tensor("concat_38"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_internal_tensor_assign_1_stride_0 = const()[name = tensor("new_cache_internal_tensor_assign_1_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_internal_tensor_assign_1_begin_mask_0 = const()[name = tensor("new_cache_internal_tensor_assign_1_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_internal_tensor_assign_1_end_mask_0 = const()[name = tensor("new_cache_internal_tensor_assign_1_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_internal_tensor_assign_1_squeeze_mask_0 = const()[name = tensor("new_cache_internal_tensor_assign_1_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_22 = const()[name = tensor("shape_22"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_10 = const()[name = tensor("reduce_prod_10"), val = tensor(1048576)]; + tensor range_1d_10_start_0 = const()[name = tensor("range_1d_10_start_0"), val = tensor(0)]; + tensor range_1d_10_step_0 = const()[name = tensor("range_1d_10_step_0"), val = tensor(1)]; + tensor range_1d_10 = range_1d(end = reduce_prod_10, start = range_1d_10_start_0, step = range_1d_10_step_0)[name = tensor("range_1d_10")]; + tensor reshape_50 = reshape(shape = shape_22, x = range_1d_10)[name = tensor("reshape_50")]; + tensor slice_by_index_10 = slice_by_index(begin = concat_37, begin_mask = new_cache_internal_tensor_assign_1_begin_mask_0, end = concat_38, end_mask = new_cache_internal_tensor_assign_1_end_mask_0, squeeze_mask = new_cache_internal_tensor_assign_1_squeeze_mask_0, stride = new_cache_internal_tensor_assign_1_stride_0, x = reshape_50)[name = tensor("slice_by_index_10")]; + tensor reshape_51_shape_0 = const()[name = tensor("reshape_51_shape_0"), val = tensor([-1])]; + tensor reshape_51 = reshape(shape = reshape_51_shape_0, x = slice_by_index_10)[name = tensor("reshape_51")]; + tensor reshape_52_shape_0 = const()[name = tensor("reshape_52_shape_0"), val = tensor([-1])]; + tensor reshape_52_cast_fp16 = reshape(shape = reshape_52_shape_0, x = var_2427_cast_fp16)[name = tensor("reshape_52_cast_fp16")]; + tensor reshape_53_shape_0 = const()[name = tensor("reshape_53_shape_0"), val = tensor([-1])]; + tensor reshape_53_cast_fp16 = reshape(shape = reshape_53_shape_0, x = cache5_to_fp16)[name = tensor("reshape_53_cast_fp16")]; + tensor scatter_10_mode_0 = const()[name = tensor("scatter_10_mode_0"), val = tensor("update")]; + tensor scatter_10_axis_0 = const()[name = tensor("scatter_10_axis_0"), val = tensor(0)]; + tensor scatter_10_validate_indices_0 = const()[name = tensor("scatter_10_validate_indices_0"), val = tensor(false)]; + tensor scatter_10_cast_fp16 = scatter(axis = scatter_10_axis_0, data = reshape_53_cast_fp16, indices = reshape_51, mode = scatter_10_mode_0, updates = reshape_52_cast_fp16, validate_indices = scatter_10_validate_indices_0)[name = tensor("scatter_10_cast_fp16")]; + tensor reshape_54_cast_fp16 = reshape(shape = shape_22, x = scatter_10_cast_fp16)[name = tensor("reshape_54_cast_fp16")]; + tensor var_2435_begin_0 = const()[name = tensor("op_2435_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor var_2435_end_0 = const()[name = tensor("op_2435_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor var_2435_end_mask_0 = const()[name = tensor("op_2435_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_2435_squeeze_mask_0 = const()[name = tensor("op_2435_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor var_2435_cast_fp16 = slice_by_index(begin = var_2435_begin_0, end = var_2435_end_0, end_mask = var_2435_end_mask_0, squeeze_mask = var_2435_squeeze_mask_0, x = reshape_54_cast_fp16)[name = tensor("op_2435_cast_fp16")]; + tensor var_2437_axis_0 = const()[name = tensor("op_2437_axis_0"), val = tensor(1)]; + tensor var_2437_mode_0 = const()[name = tensor("op_2437_mode_0"), val = tensor("update")]; + tensor var_2437_validate_indices_0 = const()[name = tensor("op_2437_validate_indices_0"), val = tensor(false)]; + tensor var_2437_cast_fp16 = scatter_along_axis(axis = var_2437_axis_0, data = var_2435_cast_fp16, indices = write_indices, mode = var_2437_mode_0, updates = v_cast_fp16, validate_indices = var_2437_validate_indices_0)[name = tensor("op_2437_cast_fp16")]; + tensor concat_39 = const()[name = tensor("concat_39"), val = tensor([1, 0, 0, 0, 0])]; + tensor concat_40 = const()[name = tensor("concat_40"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_internal_tensor_assign_2_stride_0 = const()[name = tensor("new_cache_internal_tensor_assign_2_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_internal_tensor_assign_2_begin_mask_0 = const()[name = tensor("new_cache_internal_tensor_assign_2_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_internal_tensor_assign_2_end_mask_0 = const()[name = tensor("new_cache_internal_tensor_assign_2_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_internal_tensor_assign_2_squeeze_mask_0 = const()[name = tensor("new_cache_internal_tensor_assign_2_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_23 = const()[name = tensor("shape_23"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_11 = const()[name = tensor("reduce_prod_11"), val = tensor(1048576)]; + tensor range_1d_11_start_0 = const()[name = tensor("range_1d_11_start_0"), val = tensor(0)]; + tensor range_1d_11_step_0 = const()[name = tensor("range_1d_11_step_0"), val = tensor(1)]; + tensor range_1d_11 = range_1d(end = reduce_prod_11, start = range_1d_11_start_0, step = range_1d_11_step_0)[name = tensor("range_1d_11")]; + tensor reshape_55 = reshape(shape = shape_23, x = range_1d_11)[name = tensor("reshape_55")]; + tensor slice_by_index_11 = slice_by_index(begin = concat_39, begin_mask = new_cache_internal_tensor_assign_2_begin_mask_0, end = concat_40, end_mask = new_cache_internal_tensor_assign_2_end_mask_0, squeeze_mask = new_cache_internal_tensor_assign_2_squeeze_mask_0, stride = new_cache_internal_tensor_assign_2_stride_0, x = reshape_55)[name = tensor("slice_by_index_11")]; + tensor reshape_56_shape_0 = const()[name = tensor("reshape_56_shape_0"), val = tensor([-1])]; + tensor reshape_56 = reshape(shape = reshape_56_shape_0, x = slice_by_index_11)[name = tensor("reshape_56")]; + tensor reshape_57_shape_0 = const()[name = tensor("reshape_57_shape_0"), val = tensor([-1])]; + tensor reshape_57_cast_fp16 = reshape(shape = reshape_57_shape_0, x = var_2437_cast_fp16)[name = tensor("reshape_57_cast_fp16")]; + tensor reshape_58_shape_0 = const()[name = tensor("reshape_58_shape_0"), val = tensor([-1])]; + tensor reshape_58_cast_fp16 = reshape(shape = reshape_58_shape_0, x = reshape_54_cast_fp16)[name = tensor("reshape_58_cast_fp16")]; + tensor scatter_11_mode_0 = const()[name = tensor("scatter_11_mode_0"), val = tensor("update")]; + tensor scatter_11_axis_0 = const()[name = tensor("scatter_11_axis_0"), val = tensor(0)]; + tensor scatter_11_validate_indices_0 = const()[name = tensor("scatter_11_validate_indices_0"), val = tensor(false)]; + tensor scatter_11_cast_fp16 = scatter(axis = scatter_11_axis_0, data = reshape_58_cast_fp16, indices = reshape_56, mode = scatter_11_mode_0, updates = reshape_57_cast_fp16, validate_indices = scatter_11_validate_indices_0)[name = tensor("scatter_11_cast_fp16")]; + tensor reshape_59_cast_fp16 = reshape(shape = shape_23, x = scatter_11_cast_fp16)[name = tensor("reshape_59_cast_fp16")]; + tensor reshape_59_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("reshape_59_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor keys_31_begin_0 = const()[name = tensor("keys_31_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor keys_31_end_0 = const()[name = tensor("keys_31_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor keys_31_end_mask_0 = const()[name = tensor("keys_31_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor keys_31_squeeze_mask_0 = const()[name = tensor("keys_31_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor keys_31_cast_fp16 = slice_by_index(begin = keys_31_begin_0, end = keys_31_end_0, end_mask = keys_31_end_mask_0, squeeze_mask = keys_31_squeeze_mask_0, x = reshape_59_cast_fp16)[name = tensor("keys_31_cast_fp16")]; + tensor values_31_begin_0 = const()[name = tensor("values_31_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor values_31_end_0 = const()[name = tensor("values_31_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor values_31_end_mask_0 = const()[name = tensor("values_31_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor values_31_squeeze_mask_0 = const()[name = tensor("values_31_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor values_31_cast_fp16 = slice_by_index(begin = values_31_begin_0, end = values_31_end_0, end_mask = values_31_end_mask_0, squeeze_mask = values_31_squeeze_mask_0, x = reshape_59_cast_fp16)[name = tensor("values_31_cast_fp16")]; + tensor var_2449_cast_fp16 = not_equal(x = keys_31_cast_fp16, y = keys_31_cast_fp16)[name = tensor("op_2449_cast_fp16")]; + tensor keys_33_cast_fp16 = select(a = var_360_to_fp16, b = keys_31_cast_fp16, cond = var_2449_cast_fp16)[name = tensor("keys_33_cast_fp16")]; + tensor var_2457_cast_fp16 = not_equal(x = values_31_cast_fp16, y = values_31_cast_fp16)[name = tensor("op_2457_cast_fp16")]; + tensor values_33_cast_fp16 = select(a = var_360_to_fp16, b = values_31_cast_fp16, cond = var_2457_cast_fp16)[name = tensor("values_33_cast_fp16")]; + tensor var_2481 = const()[name = tensor("op_2481"), val = tensor([0, 2, 1, 3])]; + tensor var_2494 = const()[name = tensor("op_2494"), val = tensor([1, 1, 1])]; + tensor var_2495_cast_fp16 = reshape(shape = var_2494, x = position5_to_fp16)[name = tensor("op_2495_cast_fp16")]; + tensor var_2512_to_fp16 = const()[name = tensor("op_2512_to_fp16"), val = tensor(0x1p+0)]; + tensor valid_len_cast_fp16 = add(x = var_2495_cast_fp16, y = var_2512_to_fp16)[name = tensor("valid_len_cast_fp16")]; + tensor valid_mask_cast_fp16 = less(x = k_positions_1_promoted_to_fp16, y = valid_len_cast_fp16)[name = tensor("valid_mask_cast_fp16")]; + tensor causal_mask_cast_fp16 = less_equal(x = k_positions_1_promoted_to_fp16, y = var_2495_cast_fp16)[name = tensor("causal_mask_cast_fp16")]; + tensor attn_mask_21 = logical_and(x = valid_mask_cast_fp16, y = causal_mask_cast_fp16)[name = tensor("attn_mask_21")]; + tensor attn_mask_axes_0 = const()[name = tensor("attn_mask_axes_0"), val = tensor([1])]; + tensor attn_mask = expand_dims(axes = attn_mask_axes_0, x = attn_mask_21)[name = tensor("attn_mask")]; + tensor var_2530_transpose_x_0 = const()[name = tensor("op_2530_transpose_x_0"), val = tensor(false)]; + tensor var_2530_transpose_y_0 = const()[name = tensor("op_2530_transpose_y_0"), val = tensor(false)]; + tensor transpose_28_perm_0 = const()[name = tensor("transpose_28_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_29_perm_0 = const()[name = tensor("transpose_29_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_29 = transpose(perm = transpose_29_perm_0, x = keys_33_cast_fp16)[name = tensor("transpose_31")]; + tensor transpose_28 = transpose(perm = transpose_28_perm_0, x = q_33_cast_fp16)[name = tensor("transpose_32")]; + tensor var_2530_cast_fp16 = matmul(transpose_x = var_2530_transpose_x_0, transpose_y = var_2530_transpose_y_0, x = transpose_28, y = transpose_29)[name = tensor("op_2530_cast_fp16")]; + tensor var_2524_to_fp16 = const()[name = tensor("op_2524_to_fp16"), val = tensor([0x1p-3])]; + tensor attn_weights_31_cast_fp16 = mul(x = var_2530_cast_fp16, y = var_2524_to_fp16)[name = tensor("attn_weights_31_cast_fp16")]; + tensor var_2532 = logical_not(x = attn_mask)[name = tensor("op_2532")]; + tensor var_2533_to_fp16 = const()[name = tensor("op_2533_to_fp16"), val = tensor(-inf)]; + tensor attn_weights_33_cast_fp16 = select(a = var_2533_to_fp16, b = attn_weights_31_cast_fp16, cond = var_2532)[name = tensor("attn_weights_33_cast_fp16")]; + tensor var_2535 = const()[name = tensor("op_2535"), val = tensor(-1)]; + tensor attn_weights_cast_fp16 = softmax(axis = var_2535, x = attn_weights_33_cast_fp16)[name = tensor("attn_weights_cast_fp16")]; + tensor attn_output_transpose_x_0 = const()[name = tensor("attn_output_transpose_x_0"), val = tensor(false)]; + tensor attn_output_transpose_y_0 = const()[name = tensor("attn_output_transpose_y_0"), val = tensor(false)]; + tensor values_cast_fp16 = transpose(perm = var_2481, x = values_33_cast_fp16)[name = tensor("transpose_33")]; + tensor attn_output_cast_fp16 = matmul(transpose_x = attn_output_transpose_x_0, transpose_y = attn_output_transpose_y_0, x = attn_weights_cast_fp16, y = values_cast_fp16)[name = tensor("attn_output_cast_fp16")]; + tensor var_2543 = const()[name = tensor("op_2543"), val = tensor([0, 2, 1, 3])]; + tensor var_2546 = const()[name = tensor("op_2546"), val = tensor([1, 1, 1024])]; + tensor var_2544_cast_fp16 = transpose(perm = var_2543, x = attn_output_cast_fp16)[name = tensor("transpose_30")]; + tensor input_55_cast_fp16 = reshape(shape = var_2546, x = var_2544_cast_fp16)[name = tensor("input_55_cast_fp16")]; + tensor linear_22_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = attn5_out_proj_weight_cast_fp16, x = input_55_cast_fp16)[name = tensor("linear_22_cast_fp16")]; + tensor var_2552_to_fp16 = const()[name = tensor("op_2552_to_fp16"), val = tensor(0x1p+0)]; + tensor var_2553_cast_fp16 = add(x = position5_to_fp16, y = var_2552_to_fp16)[name = tensor("op_2553_cast_fp16")]; + tensor var_2553_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_2553_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor input_57_cast_fp16 = add(x = input_53_cast_fp16, y = linear_22_cast_fp16)[name = tensor("input_57_cast_fp16")]; + tensor input_59_axes_0 = const()[name = tensor("input_59_axes_0"), val = tensor([-1])]; + tensor norm5_2_weight_to_fp16 = const()[name = tensor("norm5_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(76798720)))]; + tensor norm5_2_bias_to_fp16 = const()[name = tensor("norm5_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(76800832)))]; + tensor var_2557_to_fp16 = const()[name = tensor("op_2557_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_59_cast_fp16 = layer_norm(axes = input_59_axes_0, beta = norm5_2_bias_to_fp16, epsilon = var_2557_to_fp16, gamma = norm5_2_weight_to_fp16, x = input_57_cast_fp16)[name = tensor("input_59_cast_fp16")]; + tensor linear_23_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = linear5_1_weight_cast_fp16, x = input_59_cast_fp16)[name = tensor("linear_23_cast_fp16")]; + tensor input_61_mode_0 = const()[name = tensor("input_61_mode_0"), val = tensor("EXACT")]; + tensor input_61_cast_fp16 = gelu(mode = input_61_mode_0, x = linear_23_cast_fp16)[name = tensor("input_61_cast_fp16")]; + tensor linear_24_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = linear5_2_weight_cast_fp16, x = input_61_cast_fp16)[name = tensor("linear_24_cast_fp16")]; + tensor input_63_cast_fp16 = add(x = input_57_cast_fp16, y = linear_24_cast_fp16)[name = tensor("input_63_cast_fp16")]; + tensor input_axes_0 = const()[name = tensor("input_axes_0"), val = tensor([-1])]; + tensor out_norm_weight_to_fp16 = const()[name = tensor("out_norm_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(76802944)))]; + tensor out_norm_bias_to_fp16 = const()[name = tensor("out_norm_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(76805056)))]; + tensor var_2574_to_fp16 = const()[name = tensor("op_2574_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_cast_fp16 = layer_norm(axes = input_axes_0, beta = out_norm_bias_to_fp16, epsilon = var_2574_to_fp16, gamma = out_norm_weight_to_fp16, x = input_63_cast_fp16)[name = tensor("input_cast_fp16")]; + tensor input_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("input_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor out_eos_weight_to_fp16 = const()[name = tensor("out_eos_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(76807168)))]; + tensor out_eos_bias_to_fp16 = const()[name = tensor("out_eos_bias_to_fp16"), val = tensor([-0x1.5ep-3])]; + tensor linear_25_cast_fp16 = linear(bias = out_eos_bias_to_fp16, weight = out_eos_weight_to_fp16, x = input_cast_fp16)[name = tensor("linear_25_cast_fp16")]; + tensor linear_25_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("linear_25_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor var_2582 = cast(dtype = linear_25_cast_fp16_to_fp32_dtype_0, x = linear_25_cast_fp16)[name = tensor("cast_108")]; + tensor input = cast(dtype = input_cast_fp16_to_fp32_dtype_0, x = input_cast_fp16)[name = tensor("cast_109")]; + tensor var_2553 = cast(dtype = var_2553_cast_fp16_to_fp32_dtype_0, x = var_2553_cast_fp16)[name = tensor("cast_110")]; + tensor new_cache_internal_tensor_assign_2 = cast(dtype = reshape_59_cast_fp16_to_fp32_dtype_0, x = reshape_59_cast_fp16)[name = tensor("cast_111")]; + tensor var_2134 = cast(dtype = var_2134_cast_fp16_to_fp32_dtype_0, x = var_2134_cast_fp16)[name = tensor("cast_115")]; + tensor new_cache_9_internal_tensor_assign_2 = cast(dtype = reshape_49_cast_fp16_to_fp32_dtype_0, x = reshape_49_cast_fp16)[name = tensor("cast_116")]; + tensor var_1715 = cast(dtype = var_1715_cast_fp16_to_fp32_dtype_0, x = var_1715_cast_fp16)[name = tensor("cast_120")]; + tensor new_cache_7_internal_tensor_assign_2 = cast(dtype = reshape_39_cast_fp16_to_fp32_dtype_0, x = reshape_39_cast_fp16)[name = tensor("cast_121")]; + tensor var_1296 = cast(dtype = var_1296_cast_fp16_to_fp32_dtype_0, x = var_1296_cast_fp16)[name = tensor("cast_125")]; + tensor new_cache_5_internal_tensor_assign_2 = cast(dtype = reshape_29_cast_fp16_to_fp32_dtype_0, x = reshape_29_cast_fp16)[name = tensor("cast_126")]; + tensor var_877 = cast(dtype = var_877_cast_fp16_to_fp32_dtype_0, x = var_877_cast_fp16)[name = tensor("cast_130")]; + tensor new_cache_3_internal_tensor_assign_2 = cast(dtype = reshape_19_cast_fp16_to_fp32_dtype_0, x = reshape_19_cast_fp16)[name = tensor("cast_131")]; + tensor var_458 = cast(dtype = var_458_cast_fp16_to_fp32_dtype_0, x = var_458_cast_fp16)[name = tensor("cast_135")]; + tensor new_cache_1_internal_tensor_assign_2 = cast(dtype = reshape_9_cast_fp16_to_fp32_dtype_0, x = reshape_9_cast_fp16)[name = tensor("cast_136")]; + } -> (input, var_2582, new_cache_1_internal_tensor_assign_2, var_458, new_cache_3_internal_tensor_assign_2, var_877, new_cache_5_internal_tensor_assign_2, var_1296, new_cache_7_internal_tensor_assign_2, var_1715, new_cache_9_internal_tensor_assign_2, var_2134, new_cache_internal_tensor_assign_2, var_2553); +} \ No newline at end of file diff --git a/v2.1/spanish/flowlm_stepv2.mlmodelc/weights/weight.bin b/v2.1/spanish/flowlm_stepv2.mlmodelc/weights/weight.bin new file mode 100644 index 0000000000000000000000000000000000000000..472d377d65e3cace81ea37119c8780ebb1693514 --- /dev/null +++ b/v2.1/spanish/flowlm_stepv2.mlmodelc/weights/weight.bin @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:791404730749303b2bf25b66efc4d7c62d8072149ff6930b987f4708b6e8fab6 +size 76809280 diff --git a/v2.1/spanish/flowlm_stepv2.mlpackage/Data/com.apple.CoreML/model.mlmodel b/v2.1/spanish/flowlm_stepv2.mlpackage/Data/com.apple.CoreML/model.mlmodel new file mode 100644 index 0000000000000000000000000000000000000000..d9e00aa86fa9faa4914607234b1b24d369da8065 --- /dev/null +++ b/v2.1/spanish/flowlm_stepv2.mlpackage/Data/com.apple.CoreML/model.mlmodel @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:d1d50934ff1cf719e053e26d9fcb512a7f817072231f80c0ecd20c65bc97eece +size 212305 diff --git a/v2.1/spanish/flowlm_stepv2.mlpackage/Data/com.apple.CoreML/weights/weight.bin b/v2.1/spanish/flowlm_stepv2.mlpackage/Data/com.apple.CoreML/weights/weight.bin new file mode 100644 index 0000000000000000000000000000000000000000..472d377d65e3cace81ea37119c8780ebb1693514 --- /dev/null +++ b/v2.1/spanish/flowlm_stepv2.mlpackage/Data/com.apple.CoreML/weights/weight.bin @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:791404730749303b2bf25b66efc4d7c62d8072149ff6930b987f4708b6e8fab6 +size 76809280 diff --git a/v2.1/spanish/flowlm_stepv2.mlpackage/Manifest.json b/v2.1/spanish/flowlm_stepv2.mlpackage/Manifest.json new file mode 100644 index 0000000000000000000000000000000000000000..74285b9e4f6c5ba228d95750111bc40d91f522a2 --- /dev/null +++ b/v2.1/spanish/flowlm_stepv2.mlpackage/Manifest.json @@ -0,0 +1,18 @@ +{ + "fileFormatVersion": "1.0.0", + "itemInfoEntries": { + "12AA4F89-C398-4D78-9636-F21E0C7D6389": { + "author": "com.apple.CoreML", + "description": "CoreML Model Weights", + "name": "weights", + "path": "com.apple.CoreML/weights" + }, + "4A2D2A7A-15E5-435D-884B-B301529104DB": { + "author": "com.apple.CoreML", + "description": "CoreML Model Specification", + "name": "model.mlmodel", + "path": "com.apple.CoreML/model.mlmodel" + } + }, + "rootModelIdentifier": "4A2D2A7A-15E5-435D-884B-B301529104DB" +} diff --git a/v2.1/spanish/manifest.json b/v2.1/spanish/manifest.json new file mode 100644 index 0000000000000000000000000000000000000000..fc7da98920126d1261a9c47517b1303db1599db7 --- /dev/null +++ b/v2.1/spanish/manifest.json @@ -0,0 +1,27 @@ +{ + "version": "2.1", + "base": "v2", + "kind": "optimization-reconvert", + "self_contained": true, + "note": "Same weights as v2. Re-converted for speed: fused flow decoder (100% ANE), one-shot cond prefill, fp16 flowlm. NOT a finetune.", + "new": [ + "flow_decoder_fused", + "cond_prefill", + "flowlm_step" + ], + "copied_from_v2": [ + "flowlm_stepv2.mlpackage", + "flowlm_stepv2.mlmodelc", + "mimi_decoder.mlpackage", + "mimi_decoder.mlmodelc", + "constants", + "constants_bin" + ], + "compute_units": { + "flow_decoder_fused": "all (ANE)", + "cond_prefill": "all (GPU)", + "flowlm_step": "all (GPU, fp16)", + "flowlm_stepv2": "cpuAndGpu (GPU, int8 \u2014 fastest flowlm)", + "mimi_decoder": "cpuOnly" + } +} diff --git a/v2.1/spanish/mimi_decoder.mlmodelc/analytics/coremldata.bin b/v2.1/spanish/mimi_decoder.mlmodelc/analytics/coremldata.bin new file mode 100644 index 0000000000000000000000000000000000000000..33fc253f346b098f6bafa6bd13d660c647e3f9d3 --- /dev/null +++ b/v2.1/spanish/mimi_decoder.mlmodelc/analytics/coremldata.bin @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:9579ddbd51fa2b122af38330c7d61edb0007fbe0f006df032f2dda9724f55298 +size 243 diff --git a/v2.1/spanish/mimi_decoder.mlmodelc/coremldata.bin b/v2.1/spanish/mimi_decoder.mlmodelc/coremldata.bin new file mode 100644 index 0000000000000000000000000000000000000000..678e6f7df0b2bdc9ca6ac48e8b058a71fbb1ca59 --- /dev/null +++ b/v2.1/spanish/mimi_decoder.mlmodelc/coremldata.bin @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:512b82e9346f0e77811c84a3fc4ccfb5751721c3c87162a135bbe3c2bf99771c +size 1777 diff --git a/v2.1/spanish/mimi_decoder.mlmodelc/metadata.json b/v2.1/spanish/mimi_decoder.mlmodelc/metadata.json new file mode 100644 index 0000000000000000000000000000000000000000..f66f9e7ff5be2ffbc44c8300f3501f67a77afaf6 --- /dev/null +++ b/v2.1/spanish/mimi_decoder.mlmodelc/metadata.json @@ -0,0 +1,573 @@ +[ + { + "metadataOutputVersion" : "3.0", + "storagePrecision" : "Float32", + "outputSchema" : [ + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1 × 1 × 1920)", + "shortDescription" : "", + "shape" : "[1, 1, 1920]", + "name" : "var_724", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1 × 512 × 16)", + "shortDescription" : "", + "shape" : "[1, 512, 16]", + "name" : "var_77", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 2 × 1 × 256 × 8 × 64)", + "shortDescription" : "", + "shape" : "[2, 1, 256, 8, 64]", + "name" : "var_210", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1)", + "shortDescription" : "", + "shape" : "[1]", + "name" : "var_741", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 2 × 1 × 256 × 8 × 64)", + "shortDescription" : "", + "shape" : "[2, 1, 256, 8, 64]", + "name" : "var_400", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1)", + "shortDescription" : "", + "shape" : "[1]", + "name" : "var_744", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1 × 512 × 6)", + "shortDescription" : "", + "shape" : "[1, 512, 6]", + "name" : "var_542", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1)", + "shortDescription" : "", + "shape" : "[1]", + "name" : "conv0_first", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1 × 256 × 6)", + "shortDescription" : "", + "shape" : "[1, 256, 6]", + "name" : "var_565", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1 × 256 × 2)", + "shortDescription" : "", + "shape" : "[1, 256, 2]", + "name" : "var_585", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1)", + "shortDescription" : "", + "shape" : "[1]", + "name" : "res0_conv0_first", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1 × 128 × 0)", + "shortDescription" : "", + "shape" : "[1, 128, 0]", + "name" : "res0_conv1_prev", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1)", + "shortDescription" : "", + "shape" : "[1]", + "name" : "res0_conv1_first", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1 × 128 × 5)", + "shortDescription" : "", + "shape" : "[1, 128, 5]", + "name" : "var_621", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1 × 128 × 2)", + "shortDescription" : "", + "shape" : "[1, 128, 2]", + "name" : "var_641", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1)", + "shortDescription" : "", + "shape" : "[1]", + "name" : "res1_conv0_first", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1 × 64 × 0)", + "shortDescription" : "", + "shape" : "[1, 64, 0]", + "name" : "res1_conv1_prev", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1)", + "shortDescription" : "", + "shape" : "[1]", + "name" : "res1_conv1_first", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1 × 64 × 4)", + "shortDescription" : "", + "shape" : "[1, 64, 4]", + "name" : "var_677", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1 × 64 × 2)", + "shortDescription" : "", + "shape" : "[1, 64, 2]", + "name" : "var_697", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1)", + "shortDescription" : "", + "shape" : "[1]", + "name" : "res2_conv0_first", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1 × 32 × 0)", + "shortDescription" : "", + "shape" : "[1, 32, 0]", + "name" : "res2_conv1_prev", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1)", + "shortDescription" : "", + "shape" : "[1]", + "name" : "res2_conv1_first", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1 × 64 × 2)", + "shortDescription" : "", + "shape" : "[1, 64, 2]", + "name" : "var_725", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1)", + "shortDescription" : "", + "shape" : "[1]", + "name" : "conv_final_first", + "type" : "MultiArray" + } + ], + "modelParameters" : [ + + ], + "specificationVersion" : 8, + "mlProgramOperationTypeHistogram" : { + "Ios17.notEqual" : 4, + "Tile" : 2, + "Ios17.lessEqual" : 2, + "Ios17.layerNorm" : 4, + "Stack" : 6, + "Ios17.greaterEqual" : 4, + "Ios17.scatterAlongAxis" : 4, + "Ios17.sub" : 17, + "Identity" : 11, + "Ios17.elu" : 10, + "Ios17.logicalNot" : 2, + "Ios17.sliceByIndex" : 35, + "Ios17.concat" : 9, + "Ios17.convTranspose" : 4, + "Ios17.floorDiv" : 4, + "Ios17.expandDims" : 7, + "Ios17.less" : 2, + "Ios17.mul" : 29, + "Ios17.transpose" : 10, + "Ios17.matmul" : 4, + "Ios16.gelu" : 2, + "Ios17.conv" : 9, + "Ios17.reshape" : 22, + "Split" : 2, + "Ios17.squeeze" : 6, + "Ios17.cast" : 6, + "Ios17.sin" : 2, + "Ios17.linear" : 8, + "Select" : 6, + "Ios16.softmax" : 2, + "Ios17.add" : 28, + "Ios17.cos" : 2, + "Ios17.logicalAnd" : 6 + }, + "computePrecision" : "Mixed (Float32, Int32)", + "isUpdatable" : "0", + "stateSchema" : [ + + ], + "availability" : { + "macOS" : "14.0", + "tvOS" : "17.0", + "visionOS" : "1.0", + "watchOS" : "10.0", + "iOS" : "17.0", + "macCatalyst" : "17.0" + }, + "modelType" : { + "name" : "MLModelType_mlProgram" + }, + "userDefinedMetadata" : { + "com.github.apple.coremltools.conversion_date" : "2026-04-24", + "com.github.apple.coremltools.source" : "torch==2.9.1", + "com.github.apple.coremltools.version" : "9.0", + "com.github.apple.coremltools.source_dialect" : "TorchScript" + }, + "inputSchema" : [ + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1 × 32)", + "shortDescription" : "", + "shape" : "[1, 32]", + "name" : "latent", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1 × 512 × 16)", + "shortDescription" : "", + "shape" : "[1, 512, 16]", + "name" : "upsample_partial", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 2 × 1 × 256 × 8 × 64)", + "shortDescription" : "", + "shape" : "[2, 1, 256, 8, 64]", + "name" : "attn0_cache", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1)", + "shortDescription" : "", + "shape" : "[1]", + "name" : "attn0_offset", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 2 × 1 × 256 × 8 × 64)", + "shortDescription" : "", + "shape" : "[2, 1, 256, 8, 64]", + "name" : "attn1_cache", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1)", + "shortDescription" : "", + "shape" : "[1]", + "name" : "attn1_offset", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1 × 512 × 6)", + "shortDescription" : "", + "shape" : "[1, 512, 6]", + "name" : "conv0_prev", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1)", + "shortDescription" : "", + "shape" : "[1]", + "name" : "conv0_first", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1 × 256 × 6)", + "shortDescription" : "", + "shape" : "[1, 256, 6]", + "name" : "convtr0_partial", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1 × 256 × 2)", + "shortDescription" : "", + "shape" : "[1, 256, 2]", + "name" : "res0_conv0_prev", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1)", + "shortDescription" : "", + "shape" : "[1]", + "name" : "res0_conv0_first", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1 × 128 × 0)", + "shortDescription" : "", + "shape" : "[1, 128, 0]", + "name" : "res0_conv1_prev", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1)", + "shortDescription" : "", + "shape" : "[1]", + "name" : "res0_conv1_first", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1 × 128 × 5)", + "shortDescription" : "", + "shape" : "[1, 128, 5]", + "name" : "convtr1_partial", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1 × 128 × 2)", + "shortDescription" : "", + "shape" : "[1, 128, 2]", + "name" : "res1_conv0_prev", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1)", + "shortDescription" : "", + "shape" : "[1]", + "name" : "res1_conv0_first", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1 × 64 × 0)", + "shortDescription" : "", + "shape" : "[1, 64, 0]", + "name" : "res1_conv1_prev", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1)", + "shortDescription" : "", + "shape" : "[1]", + "name" : "res1_conv1_first", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1 × 64 × 4)", + "shortDescription" : "", + "shape" : "[1, 64, 4]", + "name" : "convtr2_partial", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1 × 64 × 2)", + "shortDescription" : "", + "shape" : "[1, 64, 2]", + "name" : "res2_conv0_prev", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1)", + "shortDescription" : "", + "shape" : "[1]", + "name" : "res2_conv0_first", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1 × 32 × 0)", + "shortDescription" : "", + "shape" : "[1, 32, 0]", + "name" : "res2_conv1_prev", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1)", + "shortDescription" : "", + "shape" : "[1]", + "name" : "res2_conv1_first", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1 × 64 × 2)", + "shortDescription" : "", + "shape" : "[1, 64, 2]", + "name" : "conv_final_prev", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1)", + "shortDescription" : "", + "shape" : "[1]", + "name" : "conv_final_first", + "type" : "MultiArray" + } + ], + "generatedClassName" : "mimi_decoder", + "method" : "predict" + } +] \ No newline at end of file diff --git a/v2.1/spanish/mimi_decoder.mlmodelc/model.mil b/v2.1/spanish/mimi_decoder.mlmodelc/model.mil new file mode 100644 index 0000000000000000000000000000000000000000..312fdf981100e6ea1ead8963b7a3aa07c21d32b8 --- /dev/null +++ b/v2.1/spanish/mimi_decoder.mlmodelc/model.mil @@ -0,0 +1,646 @@ +program(1.0) +[buildInfo = dict, tensor>({{"coremlc-component-MIL", "3520.4.1"}, {"coremlc-version", "3520.5.1"}, {"coremltools-component-torch", "2.9.1"}, {"coremltools-source-dialect", "TorchScript"}, {"coremltools-version", "9.0"}})] +{ + func main(tensor attn0_cache, tensor attn0_offset, tensor attn1_cache, tensor attn1_offset, tensor conv0_first, tensor conv0_prev, tensor conv_final_first, tensor conv_final_prev, tensor convtr0_partial, tensor convtr1_partial, tensor convtr2_partial, tensor latent, tensor res0_conv0_first, tensor res0_conv0_prev, tensor res0_conv1_first, tensor res0_conv1_prev, tensor res1_conv0_first, tensor res1_conv0_prev, tensor res1_conv1_first, tensor res1_conv1_prev, tensor res2_conv0_first, tensor res2_conv0_prev, tensor res2_conv1_first, tensor res2_conv1_prev, tensor upsample_partial) { + tensor emb_mean = const()[name = tensor("emb_mean"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(64)))]; + tensor emb_std = const()[name = tensor("emb_std"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(256)))]; + tensor mimi_quantizer_output_proj_weight = const()[name = tensor("mimi_quantizer_output_proj_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(448)))]; + tensor mimi_upsample_convtr_convtr_weight = const()[name = tensor("mimi_upsample_convtr_convtr_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(66048)))]; + tensor mimi_decoder_transformer_transformer_layers_0_norm1_bias = const()[name = tensor("mimi_decoder_transformer_transformer_layers_0_norm1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(131648)))]; + tensor mimi_decoder_transformer_transformer_layers_0_norm1_weight = const()[name = tensor("mimi_decoder_transformer_transformer_layers_0_norm1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(133760)))]; + tensor mimi_decoder_transformer_transformer_layers_0_self_attn_in_proj_weight = const()[name = tensor("mimi_decoder_transformer_transformer_layers_0_self_attn_in_proj_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(135872)))]; + tensor mimi_decoder_transformer_transformer_layers_0_self_attn_out_proj_weight = const()[name = tensor("mimi_decoder_transformer_transformer_layers_0_self_attn_out_proj_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(3281664)))]; + tensor mimi_decoder_transformer_transformer_layers_0_layer_scale_1_scale = const()[name = tensor("mimi_decoder_transformer_transformer_layers_0_layer_scale_1_scale"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(4330304)))]; + tensor mimi_decoder_transformer_transformer_layers_0_norm2_bias = const()[name = tensor("mimi_decoder_transformer_transformer_layers_0_norm2_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(4332416)))]; + tensor mimi_decoder_transformer_transformer_layers_0_norm2_weight = const()[name = tensor("mimi_decoder_transformer_transformer_layers_0_norm2_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(4334528)))]; + tensor mimi_decoder_transformer_transformer_layers_0_linear1_weight = const()[name = tensor("mimi_decoder_transformer_transformer_layers_0_linear1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(4336640)))]; + tensor mimi_decoder_transformer_transformer_layers_0_linear2_weight = const()[name = tensor("mimi_decoder_transformer_transformer_layers_0_linear2_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8531008)))]; + tensor mimi_decoder_transformer_transformer_layers_0_layer_scale_2_scale = const()[name = tensor("mimi_decoder_transformer_transformer_layers_0_layer_scale_2_scale"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(12725376)))]; + tensor mimi_decoder_transformer_transformer_layers_1_norm1_bias = const()[name = tensor("mimi_decoder_transformer_transformer_layers_1_norm1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(12727488)))]; + tensor mimi_decoder_transformer_transformer_layers_1_norm1_weight = const()[name = tensor("mimi_decoder_transformer_transformer_layers_1_norm1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(12729600)))]; + tensor mimi_decoder_transformer_transformer_layers_1_self_attn_in_proj_weight = const()[name = tensor("mimi_decoder_transformer_transformer_layers_1_self_attn_in_proj_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(12731712)))]; + tensor mimi_decoder_transformer_transformer_layers_1_self_attn_out_proj_weight = const()[name = tensor("mimi_decoder_transformer_transformer_layers_1_self_attn_out_proj_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15877504)))]; + tensor mimi_decoder_transformer_transformer_layers_1_layer_scale_1_scale = const()[name = tensor("mimi_decoder_transformer_transformer_layers_1_layer_scale_1_scale"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(16926144)))]; + tensor mimi_decoder_transformer_transformer_layers_1_norm2_bias = const()[name = tensor("mimi_decoder_transformer_transformer_layers_1_norm2_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(16928256)))]; + tensor mimi_decoder_transformer_transformer_layers_1_norm2_weight = const()[name = tensor("mimi_decoder_transformer_transformer_layers_1_norm2_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(16930368)))]; + tensor mimi_decoder_transformer_transformer_layers_1_linear1_weight = const()[name = tensor("mimi_decoder_transformer_transformer_layers_1_linear1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(16932480)))]; + tensor mimi_decoder_transformer_transformer_layers_1_linear2_weight = const()[name = tensor("mimi_decoder_transformer_transformer_layers_1_linear2_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21126848)))]; + tensor mimi_decoder_transformer_transformer_layers_1_layer_scale_2_scale = const()[name = tensor("mimi_decoder_transformer_transformer_layers_1_layer_scale_2_scale"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(25321216)))]; + tensor mimi_decoder_model_0_conv_bias = const()[name = tensor("mimi_decoder_model_0_conv_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(25323328)))]; + tensor mimi_decoder_model_0_conv_weight = const()[name = tensor("mimi_decoder_model_0_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(25325440)))]; + tensor mimi_decoder_model_2_convtr_bias = const()[name = tensor("mimi_decoder_model_2_convtr_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(32665536)))]; + tensor mimi_decoder_model_2_convtr_weight = const()[name = tensor("mimi_decoder_model_2_convtr_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(32666624)))]; + tensor mimi_decoder_model_3_block_1_conv_bias = const()[name = tensor("mimi_decoder_model_3_block_1_conv_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(38958144)))]; + tensor mimi_decoder_model_3_block_1_conv_weight = const()[name = tensor("mimi_decoder_model_3_block_1_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(38958720)))]; + tensor mimi_decoder_model_3_block_3_conv_bias = const()[name = tensor("mimi_decoder_model_3_block_3_conv_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(39352000)))]; + tensor mimi_decoder_model_3_block_3_conv_weight = const()[name = tensor("mimi_decoder_model_3_block_3_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(39353088)))]; + tensor mimi_decoder_model_5_convtr_bias = const()[name = tensor("mimi_decoder_model_5_convtr_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(39484224)))]; + tensor mimi_decoder_model_5_convtr_weight = const()[name = tensor("mimi_decoder_model_5_convtr_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(39484800)))]; + tensor mimi_decoder_model_6_block_1_conv_bias = const()[name = tensor("mimi_decoder_model_6_block_1_conv_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(40795584)))]; + tensor mimi_decoder_model_6_block_1_conv_weight = const()[name = tensor("mimi_decoder_model_6_block_1_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(40795904)))]; + tensor mimi_decoder_model_6_block_3_conv_bias = const()[name = tensor("mimi_decoder_model_6_block_3_conv_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(40894272)))]; + tensor mimi_decoder_model_6_block_3_conv_weight = const()[name = tensor("mimi_decoder_model_6_block_3_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(40894848)))]; + tensor mimi_decoder_model_8_convtr_bias = const()[name = tensor("mimi_decoder_model_8_convtr_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(40927680)))]; + tensor mimi_decoder_model_8_convtr_weight = const()[name = tensor("mimi_decoder_model_8_convtr_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(40928000)))]; + tensor mimi_decoder_model_9_block_1_conv_bias = const()[name = tensor("mimi_decoder_model_9_block_1_conv_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(41190208)))]; + tensor mimi_decoder_model_9_block_1_conv_weight = const()[name = tensor("mimi_decoder_model_9_block_1_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(41190400)))]; + tensor mimi_decoder_model_9_block_3_conv_bias = const()[name = tensor("mimi_decoder_model_9_block_3_conv_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(41215040)))]; + tensor mimi_decoder_model_9_block_3_conv_weight = const()[name = tensor("mimi_decoder_model_9_block_3_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(41215360)))]; + tensor mimi_decoder_model_11_conv_bias = const()[name = tensor("mimi_decoder_model_11_conv_bias"), val = tensor([-0x1.96p-16])]; + tensor mimi_decoder_model_11_conv_weight = const()[name = tensor("mimi_decoder_model_11_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(41223616)))]; + tensor var_38 = mul(x = latent, y = emb_std)[name = tensor("op_38")]; + tensor denorm = add(x = var_38, y = emb_mean)[name = tensor("denorm")]; + tensor input_1_axes_0 = const()[name = tensor("input_1_axes_0"), val = tensor([-1])]; + tensor input_1 = expand_dims(axes = input_1_axes_0, x = denorm)[name = tensor("input_1")]; + tensor x_1_pad_type_0 = const()[name = tensor("x_1_pad_type_0"), val = tensor("valid")]; + tensor x_1_strides_0 = const()[name = tensor("x_1_strides_0"), val = tensor([1])]; + tensor x_1_pad_0 = const()[name = tensor("x_1_pad_0"), val = tensor([0, 0])]; + tensor x_1_dilations_0 = const()[name = tensor("x_1_dilations_0"), val = tensor([1])]; + tensor x_1_groups_0 = const()[name = tensor("x_1_groups_0"), val = tensor(1)]; + tensor x_1 = conv(dilations = x_1_dilations_0, groups = x_1_groups_0, pad = x_1_pad_0, pad_type = x_1_pad_type_0, strides = x_1_strides_0, weight = mimi_quantizer_output_proj_weight, x = input_1)[name = tensor("x_1")]; + tensor var_62 = const()[name = tensor("op_62"), val = tensor(-1)]; + tensor y_1_pad_type_0 = const()[name = tensor("y_1_pad_type_0"), val = tensor("valid")]; + tensor y_1_strides_0 = const()[name = tensor("y_1_strides_0"), val = tensor([16])]; + tensor y_1_groups_0 = const()[name = tensor("y_1_groups_0"), val = tensor(512)]; + tensor y_1_pad_0 = const()[name = tensor("y_1_pad_0"), val = tensor([0, 0])]; + tensor y_1_dilations_0 = const()[name = tensor("y_1_dilations_0"), val = tensor([1])]; + tensor y_1_has_output_shape_output_shape_0 = const()[name = tensor("y_1_has_output_shape_output_shape_0"), val = tensor([1, 512, 32])]; + tensor y_1_has_output_shape = conv_transpose(dilations = y_1_dilations_0, groups = y_1_groups_0, output_shape = y_1_has_output_shape_output_shape_0, pad = y_1_pad_0, pad_type = y_1_pad_type_0, strides = y_1_strides_0, weight = mimi_upsample_convtr_convtr_weight, x = x_1)[name = tensor("y_1_has_output_shape")]; + tensor var_72_begin_0 = const()[name = tensor("op_72_begin_0"), val = tensor([0, 0, 0])]; + tensor var_72_end_0 = const()[name = tensor("op_72_end_0"), val = tensor([1, 512, 16])]; + tensor var_72_end_mask_0 = const()[name = tensor("op_72_end_mask_0"), val = tensor([true, true, false])]; + tensor var_72 = slice_by_index(begin = var_72_begin_0, end = var_72_end_0, end_mask = var_72_end_mask_0, x = y_1_has_output_shape)[name = tensor("op_72")]; + tensor var_73 = add(x = var_72, y = upsample_partial)[name = tensor("op_73")]; + tensor var_74_begin_0 = const()[name = tensor("op_74_begin_0"), val = tensor([0, 0, 16])]; + tensor var_74_end_0 = const()[name = tensor("op_74_end_0"), val = tensor([1, 512, 32])]; + tensor var_74_end_mask_0 = const()[name = tensor("op_74_end_mask_0"), val = tensor([true, true, true])]; + tensor var_74 = slice_by_index(begin = var_74_begin_0, end = var_74_end_0, end_mask = var_74_end_mask_0, x = y_1_has_output_shape)[name = tensor("op_74")]; + tensor y_3_interleave_0 = const()[name = tensor("y_3_interleave_0"), val = tensor(false)]; + tensor y_3 = concat(axis = var_62, interleave = y_3_interleave_0, values = (var_73, var_74))[name = tensor("y_3")]; + tensor var_77_begin_0 = const()[name = tensor("op_77_begin_0"), val = tensor([0, 0, 16])]; + tensor var_77_end_0 = const()[name = tensor("op_77_end_0"), val = tensor([1, 512, 32])]; + tensor var_77_end_mask_0 = const()[name = tensor("op_77_end_mask_0"), val = tensor([true, true, true])]; + tensor var_77 = slice_by_index(begin = var_77_begin_0, end = var_77_end_0, end_mask = var_77_end_mask_0, x = y_3)[name = tensor("op_77")]; + tensor x_3_begin_0 = const()[name = tensor("x_3_begin_0"), val = tensor([0, 0, 0])]; + tensor x_3_end_0 = const()[name = tensor("x_3_end_0"), val = tensor([1, 512, 16])]; + tensor x_3_end_mask_0 = const()[name = tensor("x_3_end_mask_0"), val = tensor([true, true, false])]; + tensor x_3 = slice_by_index(begin = x_3_begin_0, end = x_3_end_0, end_mask = x_3_end_mask_0, x = y_3)[name = tensor("x_3")]; + tensor var_86 = const()[name = tensor("op_86"), val = tensor(0)]; + tensor var_91 = const()[name = tensor("op_91"), val = tensor(-1)]; + tensor var_100 = const()[name = tensor("op_100"), val = tensor(-0x1.ff933cp+127)]; + tensor var_102 = const()[name = tensor("op_102"), val = tensor(0x1.4f8b58p-17)]; + tensor input_3_perm_0 = const()[name = tensor("input_3_perm_0"), val = tensor([0, 2, 1])]; + tensor query_1_axes_0 = const()[name = tensor("query_1_axes_0"), val = tensor([-1])]; + tensor input_3 = transpose(perm = input_3_perm_0, x = x_3)[name = tensor("transpose_19")]; + tensor query_1 = layer_norm(axes = query_1_axes_0, beta = mimi_decoder_transformer_transformer_layers_0_norm1_bias, epsilon = var_102, gamma = mimi_decoder_transformer_transformer_layers_0_norm1_weight, x = input_3)[name = tensor("query_1")]; + tensor linear_0_bias_0 = const()[name = tensor("linear_0_bias_0"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(41224448)))]; + tensor projected_1 = linear(bias = linear_0_bias_0, weight = mimi_decoder_transformer_transformer_layers_0_self_attn_in_proj_weight, x = query_1)[name = tensor("linear_0")]; + tensor var_130 = const()[name = tensor("op_130"), val = tensor([1, 16, 3, 8, 64])]; + tensor packed_1 = reshape(shape = var_130, x = projected_1)[name = tensor("packed_1")]; + tensor var_132_split_sizes_0 = const()[name = tensor("op_132_split_sizes_0"), val = tensor([1, 1, 1])]; + tensor var_132_axis_0 = const()[name = tensor("op_132_axis_0"), val = tensor(2)]; + tensor var_132_0, tensor var_132_1, tensor var_132_2 = split(axis = var_132_axis_0, split_sizes = var_132_split_sizes_0, x = packed_1)[name = tensor("op_132")]; + tensor squeeze_0_axes_0 = const()[name = tensor("squeeze_0_axes_0"), val = tensor([2])]; + tensor squeeze_0 = squeeze(axes = squeeze_0_axes_0, x = var_132_0)[name = tensor("squeeze_0")]; + tensor squeeze_1_axes_0 = const()[name = tensor("squeeze_1_axes_0"), val = tensor([2])]; + tensor squeeze_1 = squeeze(axes = squeeze_1_axes_0, x = var_132_1)[name = tensor("squeeze_1")]; + tensor squeeze_2_axes_0 = const()[name = tensor("squeeze_2_axes_0"), val = tensor([2])]; + tensor squeeze_2 = squeeze(axes = squeeze_2_axes_0, x = var_132_2)[name = tensor("squeeze_2")]; + tensor offset_3_begin_0 = const()[name = tensor("offset_3_begin_0"), val = tensor([0])]; + tensor offset_3_end_0 = const()[name = tensor("offset_3_end_0"), val = tensor([1])]; + tensor offset_3_end_mask_0 = const()[name = tensor("offset_3_end_mask_0"), val = tensor([false])]; + tensor offset_3_squeeze_mask_0 = const()[name = tensor("offset_3_squeeze_mask_0"), val = tensor([true])]; + tensor offset_3 = slice_by_index(begin = offset_3_begin_0, end = offset_3_end_0, end_mask = offset_3_end_mask_0, squeeze_mask = offset_3_squeeze_mask_0, x = attn0_offset)[name = tensor("offset_3")]; + tensor freqs_1 = const()[name = tensor("freqs_1"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(41230656)))]; + tensor ts_1_promoted = const()[name = tensor("ts_1_promoted"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(41230848)))]; + tensor ts_3 = add(x = ts_1_promoted, y = offset_3)[name = tensor("ts_3")]; + tensor var_148 = const()[name = tensor("op_148"), val = tensor([-1, 1, 1])]; + tensor ts_5 = reshape(shape = var_148, x = ts_3)[name = tensor("ts_5")]; + tensor var_150 = const()[name = tensor("op_150"), val = tensor([1, 16, 8, 32, 2])]; + tensor q_3 = reshape(shape = var_150, x = squeeze_0)[name = tensor("q_3")]; + tensor var_152 = const()[name = tensor("op_152"), val = tensor([1, 16, 8, 32, 2])]; + tensor k_3 = reshape(shape = var_152, x = squeeze_1)[name = tensor("k_3")]; + tensor var_154_begin_0 = const()[name = tensor("op_154_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_154_end_0 = const()[name = tensor("op_154_end_0"), val = tensor([1, 16, 8, 32, 1])]; + tensor var_154_end_mask_0 = const()[name = tensor("op_154_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_154_squeeze_mask_0 = const()[name = tensor("op_154_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_154 = slice_by_index(begin = var_154_begin_0, end = var_154_end_0, end_mask = var_154_end_mask_0, squeeze_mask = var_154_squeeze_mask_0, x = q_3)[name = tensor("op_154")]; + tensor var_156_begin_0 = const()[name = tensor("op_156_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_156_end_0 = const()[name = tensor("op_156_end_0"), val = tensor([1, 16, 8, 32, 2])]; + tensor var_156_end_mask_0 = const()[name = tensor("op_156_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_156_squeeze_mask_0 = const()[name = tensor("op_156_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_156 = slice_by_index(begin = var_156_begin_0, end = var_156_end_0, end_mask = var_156_end_mask_0, squeeze_mask = var_156_squeeze_mask_0, x = q_3)[name = tensor("op_156")]; + tensor var_158_begin_0 = const()[name = tensor("op_158_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_158_end_0 = const()[name = tensor("op_158_end_0"), val = tensor([1, 16, 8, 32, 1])]; + tensor var_158_end_mask_0 = const()[name = tensor("op_158_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_158_squeeze_mask_0 = const()[name = tensor("op_158_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_158 = slice_by_index(begin = var_158_begin_0, end = var_158_end_0, end_mask = var_158_end_mask_0, squeeze_mask = var_158_squeeze_mask_0, x = k_3)[name = tensor("op_158")]; + tensor var_160_begin_0 = const()[name = tensor("op_160_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_160_end_0 = const()[name = tensor("op_160_end_0"), val = tensor([1, 16, 8, 32, 2])]; + tensor var_160_end_mask_0 = const()[name = tensor("op_160_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_160_squeeze_mask_0 = const()[name = tensor("op_160_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_160 = slice_by_index(begin = var_160_begin_0, end = var_160_end_0, end_mask = var_160_end_mask_0, squeeze_mask = var_160_squeeze_mask_0, x = k_3)[name = tensor("op_160")]; + tensor var_162 = mul(x = freqs_1, y = ts_5)[name = tensor("op_162")]; + tensor rotr_1 = cos(x = var_162)[name = tensor("rotr_1")]; + tensor roti_1 = sin(x = var_162)[name = tensor("roti_1")]; + tensor var_166 = mul(x = var_154, y = rotr_1)[name = tensor("op_166")]; + tensor var_167 = mul(x = var_156, y = roti_1)[name = tensor("op_167")]; + tensor qor_1 = sub(x = var_166, y = var_167)[name = tensor("qor_1")]; + tensor var_169 = mul(x = var_154, y = roti_1)[name = tensor("op_169")]; + tensor var_170 = mul(x = var_156, y = rotr_1)[name = tensor("op_170")]; + tensor qoi_1 = add(x = var_169, y = var_170)[name = tensor("qoi_1")]; + tensor var_172 = mul(x = var_158, y = rotr_1)[name = tensor("op_172")]; + tensor var_173 = mul(x = var_160, y = roti_1)[name = tensor("op_173")]; + tensor kor_1 = sub(x = var_172, y = var_173)[name = tensor("kor_1")]; + tensor var_175 = mul(x = var_158, y = roti_1)[name = tensor("op_175")]; + tensor var_176 = mul(x = var_160, y = rotr_1)[name = tensor("op_176")]; + tensor koi_1 = add(x = var_175, y = var_176)[name = tensor("koi_1")]; + tensor qo_1_axis_0 = const()[name = tensor("qo_1_axis_0"), val = tensor(-1)]; + tensor qo_1 = stack(axis = qo_1_axis_0, values = (qor_1, qoi_1))[name = tensor("qo_1")]; + tensor ko_1_axis_0 = const()[name = tensor("ko_1_axis_0"), val = tensor(-1)]; + tensor ko_1 = stack(axis = ko_1_axis_0, values = (kor_1, koi_1))[name = tensor("ko_1")]; + tensor var_186 = const()[name = tensor("op_186"), val = tensor([1, 16, 8, 64])]; + tensor q_5 = reshape(shape = var_186, x = qo_1)[name = tensor("q_5")]; + tensor var_188 = const()[name = tensor("op_188"), val = tensor([1, 16, 8, 64])]; + tensor k_5 = reshape(shape = var_188, x = ko_1)[name = tensor("k_5")]; + tensor capacity_1 = const()[name = tensor("capacity_1"), val = tensor([256])]; + tensor var_193_dtype_0 = const()[name = tensor("op_193_dtype_0"), val = tensor("int32")]; + tensor var_194 = const()[name = tensor("op_194"), val = tensor([1, 1])]; + tensor var_193 = cast(dtype = var_193_dtype_0, x = attn0_offset)[name = tensor("cast_49")]; + tensor write_base_1 = reshape(shape = var_194, x = var_193)[name = tensor("write_base_1")]; + tensor write_range_1 = const()[name = tensor("write_range_1"), val = tensor([[0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15]])]; + tensor abs_idx_1 = add(x = write_base_1, y = write_range_1)[name = tensor("abs_idx_1")]; + tensor wrapped_1_div = floor_div(x = abs_idx_1, y = capacity_1)[name = tensor("wrapped_1_div")]; + tensor wrapped_1_div_scaled = mul(x = wrapped_1_div, y = capacity_1)[name = tensor("wrapped_1_div_scaled")]; + tensor wrapped_1 = sub(x = abs_idx_1, y = wrapped_1_div_scaled)[name = tensor("wrapped_1")]; + tensor var_201 = const()[name = tensor("op_201"), val = tensor([1, 16, 1, 1])]; + tensor var_202 = reshape(shape = var_201, x = wrapped_1)[name = tensor("op_202")]; + tensor write_indexes_1_reps_0 = const()[name = tensor("write_indexes_1_reps_0"), val = tensor([1, 1, 8, 64])]; + tensor write_indexes_1 = tile(reps = write_indexes_1_reps_0, x = var_202)[name = tensor("write_indexes_1")]; + tensor var_205_begin_0 = const()[name = tensor("op_205_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_205_end_0 = const()[name = tensor("op_205_end_0"), val = tensor([1, 1, 256, 8, 64])]; + tensor var_205_end_mask_0 = const()[name = tensor("op_205_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_205_squeeze_mask_0 = const()[name = tensor("op_205_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor var_205 = slice_by_index(begin = var_205_begin_0, end = var_205_end_0, end_mask = var_205_end_mask_0, squeeze_mask = var_205_squeeze_mask_0, x = attn0_cache)[name = tensor("op_205")]; + tensor new_k_cache_1_axis_0 = const()[name = tensor("new_k_cache_1_axis_0"), val = tensor(1)]; + tensor new_k_cache_1_mode_0 = const()[name = tensor("new_k_cache_1_mode_0"), val = tensor("update")]; + tensor new_k_cache_1_validate_indices_0 = const()[name = tensor("new_k_cache_1_validate_indices_0"), val = tensor(false)]; + tensor new_k_cache_1 = scatter_along_axis(axis = new_k_cache_1_axis_0, data = var_205, indices = write_indexes_1, mode = new_k_cache_1_mode_0, updates = k_5, validate_indices = new_k_cache_1_validate_indices_0)[name = tensor("new_k_cache_1")]; + tensor var_207_begin_0 = const()[name = tensor("op_207_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor var_207_end_0 = const()[name = tensor("op_207_end_0"), val = tensor([2, 1, 256, 8, 64])]; + tensor var_207_end_mask_0 = const()[name = tensor("op_207_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_207_squeeze_mask_0 = const()[name = tensor("op_207_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor var_207 = slice_by_index(begin = var_207_begin_0, end = var_207_end_0, end_mask = var_207_end_mask_0, squeeze_mask = var_207_squeeze_mask_0, x = attn0_cache)[name = tensor("op_207")]; + tensor new_v_cache_1_axis_0 = const()[name = tensor("new_v_cache_1_axis_0"), val = tensor(1)]; + tensor new_v_cache_1_mode_0 = const()[name = tensor("new_v_cache_1_mode_0"), val = tensor("update")]; + tensor new_v_cache_1_validate_indices_0 = const()[name = tensor("new_v_cache_1_validate_indices_0"), val = tensor(false)]; + tensor new_v_cache_1 = scatter_along_axis(axis = new_v_cache_1_axis_0, data = var_207, indices = write_indexes_1, mode = new_v_cache_1_mode_0, updates = squeeze_2, validate_indices = new_v_cache_1_validate_indices_0)[name = tensor("new_v_cache_1")]; + tensor var_210_axis_0 = const()[name = tensor("op_210_axis_0"), val = tensor(0)]; + tensor var_210 = stack(axis = var_210_axis_0, values = (new_k_cache_1, new_v_cache_1))[name = tensor("op_210")]; + tensor var_211 = not_equal(x = new_k_cache_1, y = new_k_cache_1)[name = tensor("op_211")]; + tensor var_212 = const()[name = tensor("op_212"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(41230976)))]; + tensor new_k_cache_3 = select(a = var_212, b = new_k_cache_1, cond = var_211)[name = tensor("new_k_cache_3")]; + tensor var_214 = not_equal(x = new_v_cache_1, y = new_v_cache_1)[name = tensor("op_214")]; + tensor new_v_cache_3 = select(a = var_212, b = new_v_cache_1, cond = var_214)[name = tensor("new_v_cache_3")]; + tensor var_219 = const()[name = tensor("op_219"), val = tensor([0, 2, 1, 3])]; + tensor var_221 = const()[name = tensor("op_221"), val = tensor([1, 1])]; + tensor var_222 = reshape(shape = var_221, x = attn0_offset)[name = tensor("op_222")]; + tensor var_224_promoted = const()[name = tensor("op_224_promoted"), val = tensor([0x1.ep+3])]; + tensor var_225 = add(x = var_222, y = var_224_promoted)[name = tensor("op_225")]; + tensor last_pos_1_dtype_0 = const()[name = tensor("last_pos_1_dtype_0"), val = tensor("int32")]; + tensor slot_idx_1 = const()[name = tensor("slot_idx_1"), val = tensor([[0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19, 20, 21, 22, 23, 24, 25, 26, 27, 28, 29, 30, 31, 32, 33, 34, 35, 36, 37, 38, 39, 40, 41, 42, 43, 44, 45, 46, 47, 48, 49, 50, 51, 52, 53, 54, 55, 56, 57, 58, 59, 60, 61, 62, 63, 64, 65, 66, 67, 68, 69, 70, 71, 72, 73, 74, 75, 76, 77, 78, 79, 80, 81, 82, 83, 84, 85, 86, 87, 88, 89, 90, 91, 92, 93, 94, 95, 96, 97, 98, 99, 100, 101, 102, 103, 104, 105, 106, 107, 108, 109, 110, 111, 112, 113, 114, 115, 116, 117, 118, 119, 120, 121, 122, 123, 124, 125, 126, 127, 128, 129, 130, 131, 132, 133, 134, 135, 136, 137, 138, 139, 140, 141, 142, 143, 144, 145, 146, 147, 148, 149, 150, 151, 152, 153, 154, 155, 156, 157, 158, 159, 160, 161, 162, 163, 164, 165, 166, 167, 168, 169, 170, 171, 172, 173, 174, 175, 176, 177, 178, 179, 180, 181, 182, 183, 184, 185, 186, 187, 188, 189, 190, 191, 192, 193, 194, 195, 196, 197, 198, 199, 200, 201, 202, 203, 204, 205, 206, 207, 208, 209, 210, 211, 212, 213, 214, 215, 216, 217, 218, 219, 220, 221, 222, 223, 224, 225, 226, 227, 228, 229, 230, 231, 232, 233, 234, 235, 236, 237, 238, 239, 240, 241, 242, 243, 244, 245, 246, 247, 248, 249, 250, 251, 252, 253, 254, 255]])]; + tensor last_pos_1 = cast(dtype = last_pos_1_dtype_0, x = var_225)[name = tensor("cast_48")]; + tensor diff_1 = sub(x = last_pos_1, y = slot_idx_1)[name = tensor("diff_1")]; + tensor var_231_div = floor_div(x = diff_1, y = capacity_1)[name = tensor("op_231_div")]; + tensor var_231_div_scaled = mul(x = var_231_div, y = capacity_1)[name = tensor("op_231_div_scaled")]; + tensor var_231 = sub(x = diff_1, y = var_231_div_scaled)[name = tensor("op_231")]; + tensor pos_k_1 = sub(x = last_pos_1, y = var_231)[name = tensor("pos_k_1")]; + tensor var_237_promoted = const()[name = tensor("op_237_promoted"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(41755328)))]; + tensor pos_q_1 = add(x = var_222, y = var_237_promoted)[name = tensor("pos_q_1")]; + tensor var_241_axes_0 = const()[name = tensor("op_241_axes_0"), val = tensor([2])]; + tensor var_241 = expand_dims(axes = var_241_axes_0, x = pos_q_1)[name = tensor("op_241")]; + tensor var_243_axes_0 = const()[name = tensor("op_243_axes_0"), val = tensor([1])]; + tensor var_243 = expand_dims(axes = var_243_axes_0, x = pos_k_1)[name = tensor("op_243")]; + tensor var_244_promoted_dtype_0 = const()[name = tensor("op_244_promoted_dtype_0"), val = tensor("fp32")]; + tensor var_244_promoted = cast(dtype = var_244_promoted_dtype_0, x = var_243)[name = tensor("cast_47")]; + tensor delta_1 = sub(x = var_241, y = var_244_promoted)[name = tensor("delta_1")]; + tensor valid_1 = greater_equal(x = var_243, y = var_86)[name = tensor("valid_1")]; + tensor var_253 = const()[name = tensor("op_253"), val = tensor([1, 1, 1])]; + tensor var_254 = reshape(shape = var_253, x = attn0_offset)[name = tensor("op_254")]; + tensor var_256_promoted = const()[name = tensor("op_256_promoted"), val = tensor([0x1.ep+3])]; + tensor var_257 = add(x = var_254, y = var_256_promoted)[name = tensor("op_257")]; + tensor var_258 = less_equal(x = var_244_promoted, y = var_257)[name = tensor("op_258")]; + tensor valid_3 = logical_and(x = valid_1, y = var_258)[name = tensor("valid_3")]; + tensor var_86_promoted = const()[name = tensor("op_86_promoted"), val = tensor(0x0p+0)]; + tensor var_260 = greater_equal(x = delta_1, y = var_86_promoted)[name = tensor("op_260")]; + tensor attn_mask_1 = logical_and(x = valid_3, y = var_260)[name = tensor("attn_mask_1")]; + tensor var_98_promoted = const()[name = tensor("op_98_promoted"), val = tensor(0x1.f4p+7)]; + tensor var_262 = less(x = delta_1, y = var_98_promoted)[name = tensor("op_262")]; + tensor attn_mask_3 = logical_and(x = attn_mask_1, y = var_262)[name = tensor("attn_mask_3")]; + tensor attn_mask_5_axes_0 = const()[name = tensor("attn_mask_5_axes_0"), val = tensor([1])]; + tensor attn_mask_5 = expand_dims(axes = attn_mask_5_axes_0, x = attn_mask_3)[name = tensor("attn_mask_5")]; + tensor var_267_transpose_x_0 = const()[name = tensor("op_267_transpose_x_0"), val = tensor(false)]; + tensor var_267_transpose_y_0 = const()[name = tensor("op_267_transpose_y_0"), val = tensor(false)]; + tensor transpose_6_perm_0 = const()[name = tensor("transpose_6_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_7_perm_0 = const()[name = tensor("transpose_7_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_7 = transpose(perm = transpose_7_perm_0, x = new_k_cache_3)[name = tensor("transpose_16")]; + tensor transpose_6 = transpose(perm = transpose_6_perm_0, x = q_5)[name = tensor("transpose_17")]; + tensor var_267 = matmul(transpose_x = var_267_transpose_x_0, transpose_y = var_267_transpose_y_0, x = transpose_6, y = transpose_7)[name = tensor("op_267")]; + tensor var_268 = const()[name = tensor("op_268"), val = tensor(0x1p-3)]; + tensor attn_1 = mul(x = var_267, y = var_268)[name = tensor("attn_1")]; + tensor var_270 = logical_not(x = attn_mask_5)[name = tensor("op_270")]; + tensor attn_3 = select(a = var_100, b = attn_1, cond = var_270)[name = tensor("attn_3")]; + tensor attn_5 = softmax(axis = var_91, x = attn_3)[name = tensor("attn_5")]; + tensor x_5_transpose_x_0 = const()[name = tensor("x_5_transpose_x_0"), val = tensor(false)]; + tensor x_5_transpose_y_0 = const()[name = tensor("x_5_transpose_y_0"), val = tensor(false)]; + tensor v_attn_1 = transpose(perm = var_219, x = new_v_cache_3)[name = tensor("transpose_18")]; + tensor x_5 = matmul(transpose_x = x_5_transpose_x_0, transpose_y = x_5_transpose_y_0, x = attn_5, y = v_attn_1)[name = tensor("x_5")]; + tensor var_274_perm_0 = const()[name = tensor("op_274_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_275 = const()[name = tensor("op_275"), val = tensor([1, 16, 512])]; + tensor var_274 = transpose(perm = var_274_perm_0, x = x_5)[name = tensor("transpose_15")]; + tensor input_5 = reshape(shape = var_275, x = var_274)[name = tensor("input_5")]; + tensor linear_1_bias_0 = const()[name = tensor("linear_1_bias_0"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(41755456)))]; + tensor x_7 = linear(bias = linear_1_bias_0, weight = mimi_decoder_transformer_transformer_layers_0_self_attn_out_proj_weight, x = input_5)[name = tensor("linear_1")]; + tensor var_284 = mul(x = mimi_decoder_transformer_transformer_layers_0_layer_scale_1_scale, y = x_7)[name = tensor("op_284")]; + tensor input_7 = add(x = input_3, y = var_284)[name = tensor("input_7")]; + tensor input_9_axes_0 = const()[name = tensor("input_9_axes_0"), val = tensor([-1])]; + tensor input_9 = layer_norm(axes = input_9_axes_0, beta = mimi_decoder_transformer_transformer_layers_0_norm2_bias, epsilon = var_102, gamma = mimi_decoder_transformer_transformer_layers_0_norm2_weight, x = input_7)[name = tensor("input_9")]; + tensor linear_2_bias_0 = const()[name = tensor("linear_2_bias_0"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(41757568)))]; + tensor var_291 = linear(bias = linear_2_bias_0, weight = mimi_decoder_transformer_transformer_layers_0_linear1_weight, x = input_9)[name = tensor("linear_2")]; + tensor input_11_mode_0 = const()[name = tensor("input_11_mode_0"), val = tensor("EXACT")]; + tensor input_11 = gelu(mode = input_11_mode_0, x = var_291)[name = tensor("input_11")]; + tensor x_9 = linear(bias = linear_1_bias_0, weight = mimi_decoder_transformer_transformer_layers_0_linear2_weight, x = input_11)[name = tensor("linear_3")]; + tensor var_297 = mul(x = mimi_decoder_transformer_transformer_layers_0_layer_scale_2_scale, y = x_9)[name = tensor("op_297")]; + tensor input_13 = add(x = input_7, y = var_297)[name = tensor("input_13")]; + tensor query_axes_0 = const()[name = tensor("query_axes_0"), val = tensor([-1])]; + tensor query = layer_norm(axes = query_axes_0, beta = mimi_decoder_transformer_transformer_layers_1_norm1_bias, epsilon = var_102, gamma = mimi_decoder_transformer_transformer_layers_1_norm1_weight, x = input_13)[name = tensor("query")]; + tensor projected = linear(bias = linear_0_bias_0, weight = mimi_decoder_transformer_transformer_layers_1_self_attn_in_proj_weight, x = query)[name = tensor("linear_4")]; + tensor var_320 = const()[name = tensor("op_320"), val = tensor([1, 16, 3, 8, 64])]; + tensor packed = reshape(shape = var_320, x = projected)[name = tensor("packed")]; + tensor var_322_split_sizes_0 = const()[name = tensor("op_322_split_sizes_0"), val = tensor([1, 1, 1])]; + tensor var_322_axis_0 = const()[name = tensor("op_322_axis_0"), val = tensor(2)]; + tensor var_322_0, tensor var_322_1, tensor var_322_2 = split(axis = var_322_axis_0, split_sizes = var_322_split_sizes_0, x = packed)[name = tensor("op_322")]; + tensor squeeze_3_axes_0 = const()[name = tensor("squeeze_3_axes_0"), val = tensor([2])]; + tensor squeeze_3 = squeeze(axes = squeeze_3_axes_0, x = var_322_0)[name = tensor("squeeze_3")]; + tensor squeeze_4_axes_0 = const()[name = tensor("squeeze_4_axes_0"), val = tensor([2])]; + tensor squeeze_4 = squeeze(axes = squeeze_4_axes_0, x = var_322_1)[name = tensor("squeeze_4")]; + tensor squeeze_5_axes_0 = const()[name = tensor("squeeze_5_axes_0"), val = tensor([2])]; + tensor squeeze_5 = squeeze(axes = squeeze_5_axes_0, x = var_322_2)[name = tensor("squeeze_5")]; + tensor offset_begin_0 = const()[name = tensor("offset_begin_0"), val = tensor([0])]; + tensor offset_end_0 = const()[name = tensor("offset_end_0"), val = tensor([1])]; + tensor offset_end_mask_0 = const()[name = tensor("offset_end_mask_0"), val = tensor([false])]; + tensor offset_squeeze_mask_0 = const()[name = tensor("offset_squeeze_mask_0"), val = tensor([true])]; + tensor offset = slice_by_index(begin = offset_begin_0, end = offset_end_0, end_mask = offset_end_mask_0, squeeze_mask = offset_squeeze_mask_0, x = attn1_offset)[name = tensor("offset")]; + tensor freqs = const()[name = tensor("freqs"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(41765824)))]; + tensor ts_7_promoted = const()[name = tensor("ts_7_promoted"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(41766016)))]; + tensor ts_9 = add(x = ts_7_promoted, y = offset)[name = tensor("ts_9")]; + tensor var_338 = const()[name = tensor("op_338"), val = tensor([-1, 1, 1])]; + tensor ts = reshape(shape = var_338, x = ts_9)[name = tensor("ts")]; + tensor var_340 = const()[name = tensor("op_340"), val = tensor([1, 16, 8, 32, 2])]; + tensor q_9 = reshape(shape = var_340, x = squeeze_3)[name = tensor("q_9")]; + tensor var_342 = const()[name = tensor("op_342"), val = tensor([1, 16, 8, 32, 2])]; + tensor k_9 = reshape(shape = var_342, x = squeeze_4)[name = tensor("k_9")]; + tensor var_344_begin_0 = const()[name = tensor("op_344_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_344_end_0 = const()[name = tensor("op_344_end_0"), val = tensor([1, 16, 8, 32, 1])]; + tensor var_344_end_mask_0 = const()[name = tensor("op_344_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_344_squeeze_mask_0 = const()[name = tensor("op_344_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_344 = slice_by_index(begin = var_344_begin_0, end = var_344_end_0, end_mask = var_344_end_mask_0, squeeze_mask = var_344_squeeze_mask_0, x = q_9)[name = tensor("op_344")]; + tensor var_346_begin_0 = const()[name = tensor("op_346_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_346_end_0 = const()[name = tensor("op_346_end_0"), val = tensor([1, 16, 8, 32, 2])]; + tensor var_346_end_mask_0 = const()[name = tensor("op_346_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_346_squeeze_mask_0 = const()[name = tensor("op_346_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_346 = slice_by_index(begin = var_346_begin_0, end = var_346_end_0, end_mask = var_346_end_mask_0, squeeze_mask = var_346_squeeze_mask_0, x = q_9)[name = tensor("op_346")]; + tensor var_348_begin_0 = const()[name = tensor("op_348_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_348_end_0 = const()[name = tensor("op_348_end_0"), val = tensor([1, 16, 8, 32, 1])]; + tensor var_348_end_mask_0 = const()[name = tensor("op_348_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_348_squeeze_mask_0 = const()[name = tensor("op_348_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_348 = slice_by_index(begin = var_348_begin_0, end = var_348_end_0, end_mask = var_348_end_mask_0, squeeze_mask = var_348_squeeze_mask_0, x = k_9)[name = tensor("op_348")]; + tensor var_350_begin_0 = const()[name = tensor("op_350_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_350_end_0 = const()[name = tensor("op_350_end_0"), val = tensor([1, 16, 8, 32, 2])]; + tensor var_350_end_mask_0 = const()[name = tensor("op_350_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_350_squeeze_mask_0 = const()[name = tensor("op_350_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_350 = slice_by_index(begin = var_350_begin_0, end = var_350_end_0, end_mask = var_350_end_mask_0, squeeze_mask = var_350_squeeze_mask_0, x = k_9)[name = tensor("op_350")]; + tensor var_352 = mul(x = freqs, y = ts)[name = tensor("op_352")]; + tensor rotr = cos(x = var_352)[name = tensor("rotr")]; + tensor roti = sin(x = var_352)[name = tensor("roti")]; + tensor var_356 = mul(x = var_344, y = rotr)[name = tensor("op_356")]; + tensor var_357 = mul(x = var_346, y = roti)[name = tensor("op_357")]; + tensor qor_5 = sub(x = var_356, y = var_357)[name = tensor("qor_5")]; + tensor var_359 = mul(x = var_344, y = roti)[name = tensor("op_359")]; + tensor var_360 = mul(x = var_346, y = rotr)[name = tensor("op_360")]; + tensor qoi_5 = add(x = var_359, y = var_360)[name = tensor("qoi_5")]; + tensor var_362 = mul(x = var_348, y = rotr)[name = tensor("op_362")]; + tensor var_363 = mul(x = var_350, y = roti)[name = tensor("op_363")]; + tensor kor_5 = sub(x = var_362, y = var_363)[name = tensor("kor_5")]; + tensor var_365 = mul(x = var_348, y = roti)[name = tensor("op_365")]; + tensor var_366 = mul(x = var_350, y = rotr)[name = tensor("op_366")]; + tensor koi_5 = add(x = var_365, y = var_366)[name = tensor("koi_5")]; + tensor qo_axis_0 = const()[name = tensor("qo_axis_0"), val = tensor(-1)]; + tensor qo = stack(axis = qo_axis_0, values = (qor_5, qoi_5))[name = tensor("qo")]; + tensor ko_axis_0 = const()[name = tensor("ko_axis_0"), val = tensor(-1)]; + tensor ko = stack(axis = ko_axis_0, values = (kor_5, koi_5))[name = tensor("ko")]; + tensor var_376 = const()[name = tensor("op_376"), val = tensor([1, 16, 8, 64])]; + tensor q = reshape(shape = var_376, x = qo)[name = tensor("q")]; + tensor var_378 = const()[name = tensor("op_378"), val = tensor([1, 16, 8, 64])]; + tensor k = reshape(shape = var_378, x = ko)[name = tensor("k")]; + tensor capacity = const()[name = tensor("capacity"), val = tensor([256])]; + tensor var_383_dtype_0 = const()[name = tensor("op_383_dtype_0"), val = tensor("int32")]; + tensor var_384 = const()[name = tensor("op_384"), val = tensor([1, 1])]; + tensor var_383 = cast(dtype = var_383_dtype_0, x = attn1_offset)[name = tensor("cast_46")]; + tensor write_base = reshape(shape = var_384, x = var_383)[name = tensor("write_base")]; + tensor write_range = const()[name = tensor("write_range"), val = tensor([[0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15]])]; + tensor abs_idx = add(x = write_base, y = write_range)[name = tensor("abs_idx")]; + tensor wrapped_div = floor_div(x = abs_idx, y = capacity)[name = tensor("wrapped_div")]; + tensor wrapped_div_scaled = mul(x = wrapped_div, y = capacity)[name = tensor("wrapped_div_scaled")]; + tensor wrapped = sub(x = abs_idx, y = wrapped_div_scaled)[name = tensor("wrapped")]; + tensor var_391 = const()[name = tensor("op_391"), val = tensor([1, 16, 1, 1])]; + tensor var_392 = reshape(shape = var_391, x = wrapped)[name = tensor("op_392")]; + tensor write_indexes_reps_0 = const()[name = tensor("write_indexes_reps_0"), val = tensor([1, 1, 8, 64])]; + tensor write_indexes = tile(reps = write_indexes_reps_0, x = var_392)[name = tensor("write_indexes")]; + tensor var_395_begin_0 = const()[name = tensor("op_395_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_395_end_0 = const()[name = tensor("op_395_end_0"), val = tensor([1, 1, 256, 8, 64])]; + tensor var_395_end_mask_0 = const()[name = tensor("op_395_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_395_squeeze_mask_0 = const()[name = tensor("op_395_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor var_395 = slice_by_index(begin = var_395_begin_0, end = var_395_end_0, end_mask = var_395_end_mask_0, squeeze_mask = var_395_squeeze_mask_0, x = attn1_cache)[name = tensor("op_395")]; + tensor new_k_cache_5_axis_0 = const()[name = tensor("new_k_cache_5_axis_0"), val = tensor(1)]; + tensor new_k_cache_5_mode_0 = const()[name = tensor("new_k_cache_5_mode_0"), val = tensor("update")]; + tensor new_k_cache_5_validate_indices_0 = const()[name = tensor("new_k_cache_5_validate_indices_0"), val = tensor(false)]; + tensor new_k_cache_5 = scatter_along_axis(axis = new_k_cache_5_axis_0, data = var_395, indices = write_indexes, mode = new_k_cache_5_mode_0, updates = k, validate_indices = new_k_cache_5_validate_indices_0)[name = tensor("new_k_cache_5")]; + tensor var_397_begin_0 = const()[name = tensor("op_397_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor var_397_end_0 = const()[name = tensor("op_397_end_0"), val = tensor([2, 1, 256, 8, 64])]; + tensor var_397_end_mask_0 = const()[name = tensor("op_397_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_397_squeeze_mask_0 = const()[name = tensor("op_397_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor var_397 = slice_by_index(begin = var_397_begin_0, end = var_397_end_0, end_mask = var_397_end_mask_0, squeeze_mask = var_397_squeeze_mask_0, x = attn1_cache)[name = tensor("op_397")]; + tensor new_v_cache_5_axis_0 = const()[name = tensor("new_v_cache_5_axis_0"), val = tensor(1)]; + tensor new_v_cache_5_mode_0 = const()[name = tensor("new_v_cache_5_mode_0"), val = tensor("update")]; + tensor new_v_cache_5_validate_indices_0 = const()[name = tensor("new_v_cache_5_validate_indices_0"), val = tensor(false)]; + tensor new_v_cache_5 = scatter_along_axis(axis = new_v_cache_5_axis_0, data = var_397, indices = write_indexes, mode = new_v_cache_5_mode_0, updates = squeeze_5, validate_indices = new_v_cache_5_validate_indices_0)[name = tensor("new_v_cache_5")]; + tensor var_400_axis_0 = const()[name = tensor("op_400_axis_0"), val = tensor(0)]; + tensor var_400 = stack(axis = var_400_axis_0, values = (new_k_cache_5, new_v_cache_5))[name = tensor("op_400")]; + tensor var_401 = not_equal(x = new_k_cache_5, y = new_k_cache_5)[name = tensor("op_401")]; + tensor new_k_cache = select(a = var_212, b = new_k_cache_5, cond = var_401)[name = tensor("new_k_cache")]; + tensor var_404 = not_equal(x = new_v_cache_5, y = new_v_cache_5)[name = tensor("op_404")]; + tensor new_v_cache = select(a = var_212, b = new_v_cache_5, cond = var_404)[name = tensor("new_v_cache")]; + tensor var_409 = const()[name = tensor("op_409"), val = tensor([0, 2, 1, 3])]; + tensor var_411 = const()[name = tensor("op_411"), val = tensor([1, 1])]; + tensor var_412 = reshape(shape = var_411, x = attn1_offset)[name = tensor("op_412")]; + tensor var_414_promoted = const()[name = tensor("op_414_promoted"), val = tensor([0x1.ep+3])]; + tensor var_415 = add(x = var_412, y = var_414_promoted)[name = tensor("op_415")]; + tensor last_pos_dtype_0 = const()[name = tensor("last_pos_dtype_0"), val = tensor("int32")]; + tensor last_pos = cast(dtype = last_pos_dtype_0, x = var_415)[name = tensor("cast_45")]; + tensor diff = sub(x = last_pos, y = slot_idx_1)[name = tensor("diff")]; + tensor var_421_div = floor_div(x = diff, y = capacity)[name = tensor("op_421_div")]; + tensor var_421_div_scaled = mul(x = var_421_div, y = capacity)[name = tensor("op_421_div_scaled")]; + tensor var_421 = sub(x = diff, y = var_421_div_scaled)[name = tensor("op_421")]; + tensor pos_k = sub(x = last_pos, y = var_421)[name = tensor("pos_k")]; + tensor var_427_promoted = const()[name = tensor("op_427_promoted"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(41766144)))]; + tensor pos_q = add(x = var_412, y = var_427_promoted)[name = tensor("pos_q")]; + tensor var_431_axes_0 = const()[name = tensor("op_431_axes_0"), val = tensor([2])]; + tensor var_431 = expand_dims(axes = var_431_axes_0, x = pos_q)[name = tensor("op_431")]; + tensor var_433_axes_0 = const()[name = tensor("op_433_axes_0"), val = tensor([1])]; + tensor var_433 = expand_dims(axes = var_433_axes_0, x = pos_k)[name = tensor("op_433")]; + tensor var_434_promoted_dtype_0 = const()[name = tensor("op_434_promoted_dtype_0"), val = tensor("fp32")]; + tensor var_434_promoted = cast(dtype = var_434_promoted_dtype_0, x = var_433)[name = tensor("cast_44")]; + tensor delta = sub(x = var_431, y = var_434_promoted)[name = tensor("delta")]; + tensor valid_5 = greater_equal(x = var_433, y = var_86)[name = tensor("valid_5")]; + tensor var_443 = const()[name = tensor("op_443"), val = tensor([1, 1, 1])]; + tensor var_444 = reshape(shape = var_443, x = attn1_offset)[name = tensor("op_444")]; + tensor var_446_promoted = const()[name = tensor("op_446_promoted"), val = tensor([0x1.ep+3])]; + tensor var_447 = add(x = var_444, y = var_446_promoted)[name = tensor("op_447")]; + tensor var_448 = less_equal(x = var_434_promoted, y = var_447)[name = tensor("op_448")]; + tensor valid = logical_and(x = valid_5, y = var_448)[name = tensor("valid")]; + tensor var_86_promoted_1 = const()[name = tensor("op_86_promoted_1"), val = tensor(0x0p+0)]; + tensor var_450 = greater_equal(x = delta, y = var_86_promoted_1)[name = tensor("op_450")]; + tensor attn_mask_7 = logical_and(x = valid, y = var_450)[name = tensor("attn_mask_7")]; + tensor var_98_promoted_1 = const()[name = tensor("op_98_promoted_1"), val = tensor(0x1.f4p+7)]; + tensor var_452 = less(x = delta, y = var_98_promoted_1)[name = tensor("op_452")]; + tensor attn_mask_9 = logical_and(x = attn_mask_7, y = var_452)[name = tensor("attn_mask_9")]; + tensor attn_mask_axes_0 = const()[name = tensor("attn_mask_axes_0"), val = tensor([1])]; + tensor attn_mask = expand_dims(axes = attn_mask_axes_0, x = attn_mask_9)[name = tensor("attn_mask")]; + tensor var_457_transpose_x_0 = const()[name = tensor("op_457_transpose_x_0"), val = tensor(false)]; + tensor var_457_transpose_y_0 = const()[name = tensor("op_457_transpose_y_0"), val = tensor(false)]; + tensor transpose_8_perm_0 = const()[name = tensor("transpose_8_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_9_perm_0 = const()[name = tensor("transpose_9_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_9 = transpose(perm = transpose_9_perm_0, x = new_k_cache)[name = tensor("transpose_12")]; + tensor transpose_8 = transpose(perm = transpose_8_perm_0, x = q)[name = tensor("transpose_13")]; + tensor var_457 = matmul(transpose_x = var_457_transpose_x_0, transpose_y = var_457_transpose_y_0, x = transpose_8, y = transpose_9)[name = tensor("op_457")]; + tensor var_458 = const()[name = tensor("op_458"), val = tensor(0x1p-3)]; + tensor attn_7 = mul(x = var_457, y = var_458)[name = tensor("attn_7")]; + tensor var_460 = logical_not(x = attn_mask)[name = tensor("op_460")]; + tensor attn_9 = select(a = var_100, b = attn_7, cond = var_460)[name = tensor("attn_9")]; + tensor attn = softmax(axis = var_91, x = attn_9)[name = tensor("attn")]; + tensor x_11_transpose_x_0 = const()[name = tensor("x_11_transpose_x_0"), val = tensor(false)]; + tensor x_11_transpose_y_0 = const()[name = tensor("x_11_transpose_y_0"), val = tensor(false)]; + tensor v_attn = transpose(perm = var_409, x = new_v_cache)[name = tensor("transpose_14")]; + tensor x_11 = matmul(transpose_x = x_11_transpose_x_0, transpose_y = x_11_transpose_y_0, x = attn, y = v_attn)[name = tensor("x_11")]; + tensor var_464_perm_0 = const()[name = tensor("op_464_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_465 = const()[name = tensor("op_465"), val = tensor([1, 16, 512])]; + tensor var_464 = transpose(perm = var_464_perm_0, x = x_11)[name = tensor("transpose_11")]; + tensor input_15 = reshape(shape = var_465, x = var_464)[name = tensor("input_15")]; + tensor x_13 = linear(bias = linear_1_bias_0, weight = mimi_decoder_transformer_transformer_layers_1_self_attn_out_proj_weight, x = input_15)[name = tensor("linear_5")]; + tensor var_474 = mul(x = mimi_decoder_transformer_transformer_layers_1_layer_scale_1_scale, y = x_13)[name = tensor("op_474")]; + tensor input_17 = add(x = input_13, y = var_474)[name = tensor("input_17")]; + tensor input_19_axes_0 = const()[name = tensor("input_19_axes_0"), val = tensor([-1])]; + tensor input_19 = layer_norm(axes = input_19_axes_0, beta = mimi_decoder_transformer_transformer_layers_1_norm2_bias, epsilon = var_102, gamma = mimi_decoder_transformer_transformer_layers_1_norm2_weight, x = input_17)[name = tensor("input_19")]; + tensor var_481 = linear(bias = linear_2_bias_0, weight = mimi_decoder_transformer_transformer_layers_1_linear1_weight, x = input_19)[name = tensor("linear_6")]; + tensor input_21_mode_0 = const()[name = tensor("input_21_mode_0"), val = tensor("EXACT")]; + tensor input_21 = gelu(mode = input_21_mode_0, x = var_481)[name = tensor("input_21")]; + tensor x_15 = linear(bias = linear_1_bias_0, weight = mimi_decoder_transformer_transformer_layers_1_linear2_weight, x = input_21)[name = tensor("linear_7")]; + tensor var_487 = mul(x = mimi_decoder_transformer_transformer_layers_1_layer_scale_2_scale, y = x_15)[name = tensor("op_487")]; + tensor z = add(x = input_17, y = var_487)[name = tensor("z")]; + tensor x_17_perm_0 = const()[name = tensor("x_17_perm_0"), val = tensor([0, 2, 1])]; + tensor var_507 = const()[name = tensor("op_507"), val = tensor(0x1p+0)]; + tensor var_508 = const()[name = tensor("op_508"), val = tensor(-1)]; + tensor input_23_interleave_0 = const()[name = tensor("input_23_interleave_0"), val = tensor(false)]; + tensor x_17 = transpose(perm = x_17_perm_0, x = z)[name = tensor("transpose_10")]; + tensor input_23 = concat(axis = var_508, interleave = input_23_interleave_0, values = (conv0_prev, x_17))[name = tensor("input_23")]; + tensor input_25_pad_type_0 = const()[name = tensor("input_25_pad_type_0"), val = tensor("valid")]; + tensor input_25_strides_0 = const()[name = tensor("input_25_strides_0"), val = tensor([1])]; + tensor input_25_pad_0 = const()[name = tensor("input_25_pad_0"), val = tensor([0, 0])]; + tensor input_25_dilations_0 = const()[name = tensor("input_25_dilations_0"), val = tensor([1])]; + tensor input_25_groups_0 = const()[name = tensor("input_25_groups_0"), val = tensor(1)]; + tensor input_25 = conv(bias = mimi_decoder_model_0_conv_bias, dilations = input_25_dilations_0, groups = input_25_groups_0, pad = input_25_pad_0, pad_type = input_25_pad_type_0, strides = input_25_strides_0, weight = mimi_decoder_model_0_conv_weight, x = input_23)[name = tensor("input_25")]; + tensor var_542_begin_0 = const()[name = tensor("op_542_begin_0"), val = tensor([0, 0, 16])]; + tensor var_542_end_0 = const()[name = tensor("op_542_end_0"), val = tensor([1, 512, 22])]; + tensor var_542_end_mask_0 = const()[name = tensor("op_542_end_mask_0"), val = tensor([true, true, true])]; + tensor var_542 = slice_by_index(begin = var_542_begin_0, end = var_542_end_0, end_mask = var_542_end_mask_0, x = input_23)[name = tensor("op_542")]; + tensor input_27 = elu(alpha = var_507, x = input_25)[name = tensor("input_27")]; + tensor y_5_pad_type_0 = const()[name = tensor("y_5_pad_type_0"), val = tensor("valid")]; + tensor y_5_strides_0 = const()[name = tensor("y_5_strides_0"), val = tensor([6])]; + tensor y_5_pad_0 = const()[name = tensor("y_5_pad_0"), val = tensor([0, 0])]; + tensor y_5_dilations_0 = const()[name = tensor("y_5_dilations_0"), val = tensor([1])]; + tensor y_5_groups_0 = const()[name = tensor("y_5_groups_0"), val = tensor(1)]; + tensor y_5_has_output_shape_output_shape_0 = const()[name = tensor("y_5_has_output_shape_output_shape_0"), val = tensor([1, 256, 102])]; + tensor y_5_has_output_shape = conv_transpose(bias = mimi_decoder_model_2_convtr_bias, dilations = y_5_dilations_0, groups = y_5_groups_0, output_shape = y_5_has_output_shape_output_shape_0, pad = y_5_pad_0, pad_type = y_5_pad_type_0, strides = y_5_strides_0, weight = mimi_decoder_model_2_convtr_weight, x = input_27)[name = tensor("y_5_has_output_shape")]; + tensor var_557_begin_0 = const()[name = tensor("op_557_begin_0"), val = tensor([0, 0, 0])]; + tensor var_557_end_0 = const()[name = tensor("op_557_end_0"), val = tensor([1, 256, 6])]; + tensor var_557_end_mask_0 = const()[name = tensor("op_557_end_mask_0"), val = tensor([true, true, false])]; + tensor var_557 = slice_by_index(begin = var_557_begin_0, end = var_557_end_0, end_mask = var_557_end_mask_0, x = y_5_has_output_shape)[name = tensor("op_557")]; + tensor var_558 = add(x = var_557, y = convtr0_partial)[name = tensor("op_558")]; + tensor var_559_begin_0 = const()[name = tensor("op_559_begin_0"), val = tensor([0, 0, 6])]; + tensor var_559_end_0 = const()[name = tensor("op_559_end_0"), val = tensor([1, 256, 102])]; + tensor var_559_end_mask_0 = const()[name = tensor("op_559_end_mask_0"), val = tensor([true, true, true])]; + tensor var_559 = slice_by_index(begin = var_559_begin_0, end = var_559_end_0, end_mask = var_559_end_mask_0, x = y_5_has_output_shape)[name = tensor("op_559")]; + tensor y_7_interleave_0 = const()[name = tensor("y_7_interleave_0"), val = tensor(false)]; + tensor y_7 = concat(axis = var_508, interleave = y_7_interleave_0, values = (var_558, var_559))[name = tensor("y_7")]; + tensor new_partial_1_begin_0 = const()[name = tensor("new_partial_1_begin_0"), val = tensor([0, 0, 96])]; + tensor new_partial_1_end_0 = const()[name = tensor("new_partial_1_end_0"), val = tensor([1, 256, 102])]; + tensor new_partial_1_end_mask_0 = const()[name = tensor("new_partial_1_end_mask_0"), val = tensor([true, true, true])]; + tensor new_partial_1 = slice_by_index(begin = new_partial_1_begin_0, end = new_partial_1_end_0, end_mask = new_partial_1_end_mask_0, x = y_7)[name = tensor("new_partial_1")]; + tensor var_564 = const()[name = tensor("op_564"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(41766272)))]; + tensor var_565 = sub(x = new_partial_1, y = var_564)[name = tensor("op_565")]; + tensor input_29_begin_0 = const()[name = tensor("input_29_begin_0"), val = tensor([0, 0, 0])]; + tensor input_29_end_0 = const()[name = tensor("input_29_end_0"), val = tensor([1, 256, 96])]; + tensor input_29_end_mask_0 = const()[name = tensor("input_29_end_mask_0"), val = tensor([true, true, false])]; + tensor input_29 = slice_by_index(begin = input_29_begin_0, end = input_29_end_0, end_mask = input_29_end_mask_0, x = y_7)[name = tensor("input_29")]; + tensor x_19 = elu(alpha = var_507, x = input_29)[name = tensor("x_19")]; + tensor input_31_interleave_0 = const()[name = tensor("input_31_interleave_0"), val = tensor(false)]; + tensor input_31 = concat(axis = var_508, interleave = input_31_interleave_0, values = (res0_conv0_prev, x_19))[name = tensor("input_31")]; + tensor input_33_pad_type_0 = const()[name = tensor("input_33_pad_type_0"), val = tensor("valid")]; + tensor input_33_strides_0 = const()[name = tensor("input_33_strides_0"), val = tensor([1])]; + tensor input_33_pad_0 = const()[name = tensor("input_33_pad_0"), val = tensor([0, 0])]; + tensor input_33_dilations_0 = const()[name = tensor("input_33_dilations_0"), val = tensor([1])]; + tensor input_33_groups_0 = const()[name = tensor("input_33_groups_0"), val = tensor(1)]; + tensor input_33 = conv(bias = mimi_decoder_model_3_block_1_conv_bias, dilations = input_33_dilations_0, groups = input_33_groups_0, pad = input_33_pad_0, pad_type = input_33_pad_type_0, strides = input_33_strides_0, weight = mimi_decoder_model_3_block_1_conv_weight, x = input_31)[name = tensor("input_33")]; + tensor var_585_begin_0 = const()[name = tensor("op_585_begin_0"), val = tensor([0, 0, 96])]; + tensor var_585_end_0 = const()[name = tensor("op_585_end_0"), val = tensor([1, 256, 98])]; + tensor var_585_end_mask_0 = const()[name = tensor("op_585_end_mask_0"), val = tensor([true, true, true])]; + tensor var_585 = slice_by_index(begin = var_585_begin_0, end = var_585_end_0, end_mask = var_585_end_mask_0, x = input_31)[name = tensor("op_585")]; + tensor x_21 = elu(alpha = var_507, x = input_33)[name = tensor("x_21")]; + tensor v_5_pad_type_0 = const()[name = tensor("v_5_pad_type_0"), val = tensor("valid")]; + tensor v_5_strides_0 = const()[name = tensor("v_5_strides_0"), val = tensor([1])]; + tensor v_5_pad_0 = const()[name = tensor("v_5_pad_0"), val = tensor([0, 0])]; + tensor v_5_dilations_0 = const()[name = tensor("v_5_dilations_0"), val = tensor([1])]; + tensor v_5_groups_0 = const()[name = tensor("v_5_groups_0"), val = tensor(1)]; + tensor v_5 = conv(bias = mimi_decoder_model_3_block_3_conv_bias, dilations = v_5_dilations_0, groups = v_5_groups_0, pad = v_5_pad_0, pad_type = v_5_pad_type_0, strides = v_5_strides_0, weight = mimi_decoder_model_3_block_3_conv_weight, x = x_21)[name = tensor("v_5")]; + tensor input_35 = add(x = input_29, y = v_5)[name = tensor("input_35")]; + tensor input_37 = elu(alpha = var_507, x = input_35)[name = tensor("input_37")]; + tensor y_9_pad_type_0 = const()[name = tensor("y_9_pad_type_0"), val = tensor("valid")]; + tensor y_9_strides_0 = const()[name = tensor("y_9_strides_0"), val = tensor([5])]; + tensor y_9_pad_0 = const()[name = tensor("y_9_pad_0"), val = tensor([0, 0])]; + tensor y_9_dilations_0 = const()[name = tensor("y_9_dilations_0"), val = tensor([1])]; + tensor y_9_groups_0 = const()[name = tensor("y_9_groups_0"), val = tensor(1)]; + tensor y_9_has_output_shape_output_shape_0 = const()[name = tensor("y_9_has_output_shape_output_shape_0"), val = tensor([1, 128, 485])]; + tensor y_9_has_output_shape = conv_transpose(bias = mimi_decoder_model_5_convtr_bias, dilations = y_9_dilations_0, groups = y_9_groups_0, output_shape = y_9_has_output_shape_output_shape_0, pad = y_9_pad_0, pad_type = y_9_pad_type_0, strides = y_9_strides_0, weight = mimi_decoder_model_5_convtr_weight, x = input_37)[name = tensor("y_9_has_output_shape")]; + tensor var_613_begin_0 = const()[name = tensor("op_613_begin_0"), val = tensor([0, 0, 0])]; + tensor var_613_end_0 = const()[name = tensor("op_613_end_0"), val = tensor([1, 128, 5])]; + tensor var_613_end_mask_0 = const()[name = tensor("op_613_end_mask_0"), val = tensor([true, true, false])]; + tensor var_613 = slice_by_index(begin = var_613_begin_0, end = var_613_end_0, end_mask = var_613_end_mask_0, x = y_9_has_output_shape)[name = tensor("op_613")]; + tensor var_614 = add(x = var_613, y = convtr1_partial)[name = tensor("op_614")]; + tensor var_615_begin_0 = const()[name = tensor("op_615_begin_0"), val = tensor([0, 0, 5])]; + tensor var_615_end_0 = const()[name = tensor("op_615_end_0"), val = tensor([1, 128, 485])]; + tensor var_615_end_mask_0 = const()[name = tensor("op_615_end_mask_0"), val = tensor([true, true, true])]; + tensor var_615 = slice_by_index(begin = var_615_begin_0, end = var_615_end_0, end_mask = var_615_end_mask_0, x = y_9_has_output_shape)[name = tensor("op_615")]; + tensor y_11_interleave_0 = const()[name = tensor("y_11_interleave_0"), val = tensor(false)]; + tensor y_11 = concat(axis = var_508, interleave = y_11_interleave_0, values = (var_614, var_615))[name = tensor("y_11")]; + tensor new_partial_3_begin_0 = const()[name = tensor("new_partial_3_begin_0"), val = tensor([0, 0, 480])]; + tensor new_partial_3_end_0 = const()[name = tensor("new_partial_3_end_0"), val = tensor([1, 128, 485])]; + tensor new_partial_3_end_mask_0 = const()[name = tensor("new_partial_3_end_mask_0"), val = tensor([true, true, true])]; + tensor new_partial_3 = slice_by_index(begin = new_partial_3_begin_0, end = new_partial_3_end_0, end_mask = new_partial_3_end_mask_0, x = y_11)[name = tensor("new_partial_3")]; + tensor var_620 = const()[name = tensor("op_620"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(41767360)))]; + tensor var_621 = sub(x = new_partial_3, y = var_620)[name = tensor("op_621")]; + tensor input_39_begin_0 = const()[name = tensor("input_39_begin_0"), val = tensor([0, 0, 0])]; + tensor input_39_end_0 = const()[name = tensor("input_39_end_0"), val = tensor([1, 128, 480])]; + tensor input_39_end_mask_0 = const()[name = tensor("input_39_end_mask_0"), val = tensor([true, true, false])]; + tensor input_39 = slice_by_index(begin = input_39_begin_0, end = input_39_end_0, end_mask = input_39_end_mask_0, x = y_11)[name = tensor("input_39")]; + tensor x_23 = elu(alpha = var_507, x = input_39)[name = tensor("x_23")]; + tensor input_41_interleave_0 = const()[name = tensor("input_41_interleave_0"), val = tensor(false)]; + tensor input_41 = concat(axis = var_508, interleave = input_41_interleave_0, values = (res1_conv0_prev, x_23))[name = tensor("input_41")]; + tensor input_43_pad_type_0 = const()[name = tensor("input_43_pad_type_0"), val = tensor("valid")]; + tensor input_43_strides_0 = const()[name = tensor("input_43_strides_0"), val = tensor([1])]; + tensor input_43_pad_0 = const()[name = tensor("input_43_pad_0"), val = tensor([0, 0])]; + tensor input_43_dilations_0 = const()[name = tensor("input_43_dilations_0"), val = tensor([1])]; + tensor input_43_groups_0 = const()[name = tensor("input_43_groups_0"), val = tensor(1)]; + tensor input_43 = conv(bias = mimi_decoder_model_6_block_1_conv_bias, dilations = input_43_dilations_0, groups = input_43_groups_0, pad = input_43_pad_0, pad_type = input_43_pad_type_0, strides = input_43_strides_0, weight = mimi_decoder_model_6_block_1_conv_weight, x = input_41)[name = tensor("input_43")]; + tensor var_641_begin_0 = const()[name = tensor("op_641_begin_0"), val = tensor([0, 0, 480])]; + tensor var_641_end_0 = const()[name = tensor("op_641_end_0"), val = tensor([1, 128, 482])]; + tensor var_641_end_mask_0 = const()[name = tensor("op_641_end_mask_0"), val = tensor([true, true, true])]; + tensor var_641 = slice_by_index(begin = var_641_begin_0, end = var_641_end_0, end_mask = var_641_end_mask_0, x = input_41)[name = tensor("op_641")]; + tensor x_25 = elu(alpha = var_507, x = input_43)[name = tensor("x_25")]; + tensor v_7_pad_type_0 = const()[name = tensor("v_7_pad_type_0"), val = tensor("valid")]; + tensor v_7_strides_0 = const()[name = tensor("v_7_strides_0"), val = tensor([1])]; + tensor v_7_pad_0 = const()[name = tensor("v_7_pad_0"), val = tensor([0, 0])]; + tensor v_7_dilations_0 = const()[name = tensor("v_7_dilations_0"), val = tensor([1])]; + tensor v_7_groups_0 = const()[name = tensor("v_7_groups_0"), val = tensor(1)]; + tensor v_7 = conv(bias = mimi_decoder_model_6_block_3_conv_bias, dilations = v_7_dilations_0, groups = v_7_groups_0, pad = v_7_pad_0, pad_type = v_7_pad_type_0, strides = v_7_strides_0, weight = mimi_decoder_model_6_block_3_conv_weight, x = x_25)[name = tensor("v_7")]; + tensor input_45 = add(x = input_39, y = v_7)[name = tensor("input_45")]; + tensor input_47 = elu(alpha = var_507, x = input_45)[name = tensor("input_47")]; + tensor y_13_pad_type_0 = const()[name = tensor("y_13_pad_type_0"), val = tensor("valid")]; + tensor y_13_strides_0 = const()[name = tensor("y_13_strides_0"), val = tensor([4])]; + tensor y_13_pad_0 = const()[name = tensor("y_13_pad_0"), val = tensor([0, 0])]; + tensor y_13_dilations_0 = const()[name = tensor("y_13_dilations_0"), val = tensor([1])]; + tensor y_13_groups_0 = const()[name = tensor("y_13_groups_0"), val = tensor(1)]; + tensor y_13_has_output_shape_output_shape_0 = const()[name = tensor("y_13_has_output_shape_output_shape_0"), val = tensor([1, 64, 1924])]; + tensor y_13_has_output_shape = conv_transpose(bias = mimi_decoder_model_8_convtr_bias, dilations = y_13_dilations_0, groups = y_13_groups_0, output_shape = y_13_has_output_shape_output_shape_0, pad = y_13_pad_0, pad_type = y_13_pad_type_0, strides = y_13_strides_0, weight = mimi_decoder_model_8_convtr_weight, x = input_47)[name = tensor("y_13_has_output_shape")]; + tensor var_669_begin_0 = const()[name = tensor("op_669_begin_0"), val = tensor([0, 0, 0])]; + tensor var_669_end_0 = const()[name = tensor("op_669_end_0"), val = tensor([1, 64, 4])]; + tensor var_669_end_mask_0 = const()[name = tensor("op_669_end_mask_0"), val = tensor([true, true, false])]; + tensor var_669 = slice_by_index(begin = var_669_begin_0, end = var_669_end_0, end_mask = var_669_end_mask_0, x = y_13_has_output_shape)[name = tensor("op_669")]; + tensor var_670 = add(x = var_669, y = convtr2_partial)[name = tensor("op_670")]; + tensor var_671_begin_0 = const()[name = tensor("op_671_begin_0"), val = tensor([0, 0, 4])]; + tensor var_671_end_0 = const()[name = tensor("op_671_end_0"), val = tensor([1, 64, 1924])]; + tensor var_671_end_mask_0 = const()[name = tensor("op_671_end_mask_0"), val = tensor([true, true, true])]; + tensor var_671 = slice_by_index(begin = var_671_begin_0, end = var_671_end_0, end_mask = var_671_end_mask_0, x = y_13_has_output_shape)[name = tensor("op_671")]; + tensor y_interleave_0 = const()[name = tensor("y_interleave_0"), val = tensor(false)]; + tensor y = concat(axis = var_508, interleave = y_interleave_0, values = (var_670, var_671))[name = tensor("y")]; + tensor new_partial_begin_0 = const()[name = tensor("new_partial_begin_0"), val = tensor([0, 0, 1920])]; + tensor new_partial_end_0 = const()[name = tensor("new_partial_end_0"), val = tensor([1, 64, 1924])]; + tensor new_partial_end_mask_0 = const()[name = tensor("new_partial_end_mask_0"), val = tensor([true, true, true])]; + tensor new_partial = slice_by_index(begin = new_partial_begin_0, end = new_partial_end_0, end_mask = new_partial_end_mask_0, x = y)[name = tensor("new_partial")]; + tensor var_676 = const()[name = tensor("op_676"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(41767936)))]; + tensor var_677 = sub(x = new_partial, y = var_676)[name = tensor("op_677")]; + tensor input_49_begin_0 = const()[name = tensor("input_49_begin_0"), val = tensor([0, 0, 0])]; + tensor input_49_end_0 = const()[name = tensor("input_49_end_0"), val = tensor([1, 64, 1920])]; + tensor input_49_end_mask_0 = const()[name = tensor("input_49_end_mask_0"), val = tensor([true, true, false])]; + tensor input_49 = slice_by_index(begin = input_49_begin_0, end = input_49_end_0, end_mask = input_49_end_mask_0, x = y)[name = tensor("input_49")]; + tensor x_27 = elu(alpha = var_507, x = input_49)[name = tensor("x_27")]; + tensor input_51_interleave_0 = const()[name = tensor("input_51_interleave_0"), val = tensor(false)]; + tensor input_51 = concat(axis = var_508, interleave = input_51_interleave_0, values = (res2_conv0_prev, x_27))[name = tensor("input_51")]; + tensor input_53_pad_type_0 = const()[name = tensor("input_53_pad_type_0"), val = tensor("valid")]; + tensor input_53_strides_0 = const()[name = tensor("input_53_strides_0"), val = tensor([1])]; + tensor input_53_pad_0 = const()[name = tensor("input_53_pad_0"), val = tensor([0, 0])]; + tensor input_53_dilations_0 = const()[name = tensor("input_53_dilations_0"), val = tensor([1])]; + tensor input_53_groups_0 = const()[name = tensor("input_53_groups_0"), val = tensor(1)]; + tensor input_53 = conv(bias = mimi_decoder_model_9_block_1_conv_bias, dilations = input_53_dilations_0, groups = input_53_groups_0, pad = input_53_pad_0, pad_type = input_53_pad_type_0, strides = input_53_strides_0, weight = mimi_decoder_model_9_block_1_conv_weight, x = input_51)[name = tensor("input_53")]; + tensor var_697_begin_0 = const()[name = tensor("op_697_begin_0"), val = tensor([0, 0, 1920])]; + tensor var_697_end_0 = const()[name = tensor("op_697_end_0"), val = tensor([1, 64, 1922])]; + tensor var_697_end_mask_0 = const()[name = tensor("op_697_end_mask_0"), val = tensor([true, true, true])]; + tensor var_697 = slice_by_index(begin = var_697_begin_0, end = var_697_end_0, end_mask = var_697_end_mask_0, x = input_51)[name = tensor("op_697")]; + tensor x_29 = elu(alpha = var_507, x = input_53)[name = tensor("x_29")]; + tensor v_pad_type_0 = const()[name = tensor("v_pad_type_0"), val = tensor("valid")]; + tensor v_strides_0 = const()[name = tensor("v_strides_0"), val = tensor([1])]; + tensor v_pad_0 = const()[name = tensor("v_pad_0"), val = tensor([0, 0])]; + tensor v_dilations_0 = const()[name = tensor("v_dilations_0"), val = tensor([1])]; + tensor v_groups_0 = const()[name = tensor("v_groups_0"), val = tensor(1)]; + tensor v = conv(bias = mimi_decoder_model_9_block_3_conv_bias, dilations = v_dilations_0, groups = v_groups_0, pad = v_pad_0, pad_type = v_pad_type_0, strides = v_strides_0, weight = mimi_decoder_model_9_block_3_conv_weight, x = x_29)[name = tensor("v")]; + tensor input_55 = add(x = input_49, y = v)[name = tensor("input_55")]; + tensor x = elu(alpha = var_507, x = input_55)[name = tensor("x")]; + tensor input_interleave_0 = const()[name = tensor("input_interleave_0"), val = tensor(false)]; + tensor input = concat(axis = var_508, interleave = input_interleave_0, values = (conv_final_prev, x))[name = tensor("input")]; + tensor var_724_pad_type_0 = const()[name = tensor("op_724_pad_type_0"), val = tensor("valid")]; + tensor var_724_strides_0 = const()[name = tensor("op_724_strides_0"), val = tensor([1])]; + tensor var_724_pad_0 = const()[name = tensor("op_724_pad_0"), val = tensor([0, 0])]; + tensor var_724_dilations_0 = const()[name = tensor("op_724_dilations_0"), val = tensor([1])]; + tensor var_724_groups_0 = const()[name = tensor("op_724_groups_0"), val = tensor(1)]; + tensor var_724 = conv(bias = mimi_decoder_model_11_conv_bias, dilations = var_724_dilations_0, groups = var_724_groups_0, pad = var_724_pad_0, pad_type = var_724_pad_type_0, strides = var_724_strides_0, weight = mimi_decoder_model_11_conv_weight, x = input)[name = tensor("op_724")]; + tensor var_725_begin_0 = const()[name = tensor("op_725_begin_0"), val = tensor([0, 0, 1920])]; + tensor var_725_end_0 = const()[name = tensor("op_725_end_0"), val = tensor([1, 64, 1922])]; + tensor var_725_end_mask_0 = const()[name = tensor("op_725_end_mask_0"), val = tensor([true, true, true])]; + tensor var_725 = slice_by_index(begin = var_725_begin_0, end = var_725_end_0, end_mask = var_725_end_mask_0, x = input)[name = tensor("op_725")]; + tensor var_740_promoted = const()[name = tensor("op_740_promoted"), val = tensor(0x1p+4)]; + tensor var_741 = add(x = attn0_offset, y = var_740_promoted)[name = tensor("op_741")]; + tensor var_743_promoted = const()[name = tensor("op_743_promoted"), val = tensor(0x1p+4)]; + tensor var_744 = add(x = attn1_offset, y = var_743_promoted)[name = tensor("op_744")]; + tensor conv0_first_tmp = identity(x = conv0_first)[name = tensor("conv0_first_tmp")]; + tensor res0_conv0_first_tmp = identity(x = res0_conv0_first)[name = tensor("res0_conv0_first_tmp")]; + tensor res0_conv1_prev_tmp = identity(x = res0_conv1_prev)[name = tensor("res0_conv1_prev_tmp")]; + tensor res0_conv1_first_tmp = identity(x = res0_conv1_first)[name = tensor("res0_conv1_first_tmp")]; + tensor res1_conv0_first_tmp = identity(x = res1_conv0_first)[name = tensor("res1_conv0_first_tmp")]; + tensor res1_conv1_prev_tmp = identity(x = res1_conv1_prev)[name = tensor("res1_conv1_prev_tmp")]; + tensor res1_conv1_first_tmp = identity(x = res1_conv1_first)[name = tensor("res1_conv1_first_tmp")]; + tensor res2_conv0_first_tmp = identity(x = res2_conv0_first)[name = tensor("res2_conv0_first_tmp")]; + tensor res2_conv1_prev_tmp = identity(x = res2_conv1_prev)[name = tensor("res2_conv1_prev_tmp")]; + tensor res2_conv1_first_tmp = identity(x = res2_conv1_first)[name = tensor("res2_conv1_first_tmp")]; + tensor conv_final_first_tmp = identity(x = conv_final_first)[name = tensor("conv_final_first_tmp")]; + } -> (var_724, var_77, var_210, var_741, var_400, var_744, var_542, conv0_first, var_565, var_585, res0_conv0_first, res0_conv1_prev, res0_conv1_first, var_621, var_641, res1_conv0_first, res1_conv1_prev, res1_conv1_first, var_677, var_697, res2_conv0_first, res2_conv1_prev, res2_conv1_first, var_725, conv_final_first); +} \ No newline at end of file diff --git a/v2.1/spanish/mimi_decoder.mlmodelc/weights/weight.bin b/v2.1/spanish/mimi_decoder.mlmodelc/weights/weight.bin new file mode 100644 index 0000000000000000000000000000000000000000..bd5e7b9637062a46e05dbd4dc1bda16bdc3045d8 --- /dev/null +++ b/v2.1/spanish/mimi_decoder.mlmodelc/weights/weight.bin @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:bb7420c1509b2cffe63c553fa375975d9600df265bbb60aa8e200626f849b8e3 +size 41768256 diff --git a/v2.1/spanish/mimi_decoder.mlpackage/Data/com.apple.CoreML/model.mlmodel b/v2.1/spanish/mimi_decoder.mlpackage/Data/com.apple.CoreML/model.mlmodel new file mode 100644 index 0000000000000000000000000000000000000000..bc94dc792723464cb9949cff6c8c954f7d3d8e80 --- /dev/null +++ b/v2.1/spanish/mimi_decoder.mlpackage/Data/com.apple.CoreML/model.mlmodel @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:33cb046e2869fcbe544fb43acf7e7cbe2273d6f3e3eb70fc9f1cd2acdbbd7d61 +size 93270 diff --git a/v2.1/spanish/mimi_decoder.mlpackage/Data/com.apple.CoreML/weights/weight.bin b/v2.1/spanish/mimi_decoder.mlpackage/Data/com.apple.CoreML/weights/weight.bin new file mode 100644 index 0000000000000000000000000000000000000000..bd5e7b9637062a46e05dbd4dc1bda16bdc3045d8 --- /dev/null +++ b/v2.1/spanish/mimi_decoder.mlpackage/Data/com.apple.CoreML/weights/weight.bin @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:bb7420c1509b2cffe63c553fa375975d9600df265bbb60aa8e200626f849b8e3 +size 41768256 diff --git a/v2.1/spanish/mimi_decoder.mlpackage/Manifest.json b/v2.1/spanish/mimi_decoder.mlpackage/Manifest.json new file mode 100644 index 0000000000000000000000000000000000000000..d716e89eb11724543854229d4ce47a0bdca7415a --- /dev/null +++ b/v2.1/spanish/mimi_decoder.mlpackage/Manifest.json @@ -0,0 +1,18 @@ +{ + "fileFormatVersion": "1.0.0", + "itemInfoEntries": { + "63F87D76-6332-482C-8294-AAA26AE17F2F": { + "author": "com.apple.CoreML", + "description": "CoreML Model Specification", + "name": "model.mlmodel", + "path": "com.apple.CoreML/model.mlmodel" + }, + "7D0E8E2C-33BD-4E94-AE53-CE9AE9D3B011": { + "author": "com.apple.CoreML", + "description": "CoreML Model Weights", + "name": "weights", + "path": "com.apple.CoreML/weights" + } + }, + "rootModelIdentifier": "63F87D76-6332-482C-8294-AAA26AE17F2F" +} diff --git a/v2.1/spanish_24l/cond_prefill.mlmodelc/analytics/coremldata.bin b/v2.1/spanish_24l/cond_prefill.mlmodelc/analytics/coremldata.bin new file mode 100644 index 0000000000000000000000000000000000000000..565bed81d3733fcd79a34a402f0f9e782154963e --- /dev/null +++ b/v2.1/spanish_24l/cond_prefill.mlmodelc/analytics/coremldata.bin @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:70e757cd26676c1408ace61805e2202482d9b399c0f8e0bdff5cf1c3773ead2e +size 243 diff --git a/v2.1/spanish_24l/cond_prefill.mlmodelc/coremldata.bin b/v2.1/spanish_24l/cond_prefill.mlmodelc/coremldata.bin new file mode 100644 index 0000000000000000000000000000000000000000..1785fb31531e3bce1dc283e9568e181375d86086 --- /dev/null +++ b/v2.1/spanish_24l/cond_prefill.mlmodelc/coremldata.bin @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:6c02efb89c164fdef27a53f29dea316d25d1ae4ae798741528257942b0e7ad35 +size 3519 diff --git a/v2.1/spanish_24l/cond_prefill.mlmodelc/model.mil b/v2.1/spanish_24l/cond_prefill.mlmodelc/model.mil new file mode 100644 index 0000000000000000000000000000000000000000..7028cd94630c33bb5143ac20b6b86ca016eb1d25 --- /dev/null +++ b/v2.1/spanish_24l/cond_prefill.mlmodelc/model.mil @@ -0,0 +1,5431 @@ +program(1.0) +[buildInfo = dict, tensor>({{"coremlc-component-MIL", "3520.4.1"}, {"coremlc-version", "3520.5.1"}, {"coremltools-component-torch", "2.12.0"}, {"coremltools-source-dialect", "TorchScript"}, {"coremltools-version", "9.0"}})] +{ + func main(tensor cache0, tensor cache1, tensor cache10, tensor cache11, tensor cache12, tensor cache13, tensor cache14, tensor cache15, tensor cache16, tensor cache17, tensor cache18, tensor cache19, tensor cache2, tensor cache20, tensor cache21, tensor cache22, tensor cache23, tensor cache3, tensor cache4, tensor cache5, tensor cache6, tensor cache7, tensor cache8, tensor cache9, tensor conditioning, tensor position0, tensor position1, tensor position10, tensor position11, tensor position12, tensor position13, tensor position14, tensor position15, tensor position16, tensor position17, tensor position18, tensor position19, tensor position2, tensor position20, tensor position21, tensor position22, tensor position23, tensor position3, tensor position4, tensor position5, tensor position6, tensor position7, tensor position8, tensor position9, tensor valid_len) { + tensor input_3_axes_0 = const()[name = tensor("input_3_axes_0"), val = tensor([-1])]; + tensor conditioning_to_fp16_dtype_0 = const()[name = tensor("conditioning_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor norm0_1_weight_to_fp16 = const()[name = tensor("norm0_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(64)))]; + tensor norm0_1_bias_to_fp16 = const()[name = tensor("norm0_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(2176)))]; + tensor var_192_to_fp16 = const()[name = tensor("op_192_to_fp16"), val = tensor(0x1.5p-17)]; + tensor conditioning_to_fp16 = cast(dtype = conditioning_to_fp16_dtype_0, x = conditioning)[name = tensor("cast_663")]; + tensor input_3_cast_fp16 = layer_norm(axes = input_3_axes_0, beta = norm0_1_bias_to_fp16, epsilon = var_192_to_fp16, gamma = norm0_1_weight_to_fp16, x = conditioning_to_fp16)[name = tensor("input_3_cast_fp16")]; + tensor attn0_in_proj_weight_to_fp16 = const()[name = tensor("attn0_in_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(4288)))]; + tensor linear_0_bias_0_to_fp16 = const()[name = tensor("linear_0_bias_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(6295808)))]; + tensor linear_0_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = attn0_in_proj_weight_to_fp16, x = input_3_cast_fp16)[name = tensor("linear_0_cast_fp16")]; + tensor var_206 = const()[name = tensor("op_206"), val = tensor([1, 256, 3, 16, 64])]; + tensor qkv_1_cast_fp16 = reshape(shape = var_206, x = linear_0_cast_fp16)[name = tensor("qkv_1_cast_fp16")]; + tensor q_1_begin_0 = const()[name = tensor("q_1_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor q_1_end_0 = const()[name = tensor("q_1_end_0"), val = tensor([1, 256, 1, 16, 64])]; + tensor q_1_end_mask_0 = const()[name = tensor("q_1_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor q_1_squeeze_mask_0 = const()[name = tensor("q_1_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor q_1_cast_fp16 = slice_by_index(begin = q_1_begin_0, end = q_1_end_0, end_mask = q_1_end_mask_0, squeeze_mask = q_1_squeeze_mask_0, x = qkv_1_cast_fp16)[name = tensor("q_1_cast_fp16")]; + tensor k_1_begin_0 = const()[name = tensor("k_1_begin_0"), val = tensor([0, 0, 1, 0, 0])]; + tensor k_1_end_0 = const()[name = tensor("k_1_end_0"), val = tensor([1, 256, 2, 16, 64])]; + tensor k_1_end_mask_0 = const()[name = tensor("k_1_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor k_1_squeeze_mask_0 = const()[name = tensor("k_1_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor k_1_cast_fp16 = slice_by_index(begin = k_1_begin_0, end = k_1_end_0, end_mask = k_1_end_mask_0, squeeze_mask = k_1_squeeze_mask_0, x = qkv_1_cast_fp16)[name = tensor("k_1_cast_fp16")]; + tensor v_1_begin_0 = const()[name = tensor("v_1_begin_0"), val = tensor([0, 0, 2, 0, 0])]; + tensor v_1_end_0 = const()[name = tensor("v_1_end_0"), val = tensor([1, 256, 3, 16, 64])]; + tensor v_1_end_mask_0 = const()[name = tensor("v_1_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor v_1_squeeze_mask_0 = const()[name = tensor("v_1_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor v_1_cast_fp16 = slice_by_index(begin = v_1_begin_0, end = v_1_end_0, end_mask = v_1_end_mask_0, squeeze_mask = v_1_squeeze_mask_0, x = qkv_1_cast_fp16)[name = tensor("v_1_cast_fp16")]; + tensor var_264 = const()[name = tensor("op_264"), val = tensor([1, 1])]; + tensor position0_to_fp16_dtype_0 = const()[name = tensor("position0_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor position0_to_fp16 = cast(dtype = position0_to_fp16_dtype_0, x = position0)[name = tensor("cast_662")]; + tensor var_265_cast_fp16 = reshape(shape = var_264, x = position0_to_fp16)[name = tensor("op_265_cast_fp16")]; + tensor ts_1_promoted_to_fp16 = const()[name = tensor("ts_1_promoted_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(6302016)))]; + tensor ts_3_cast_fp16 = add(x = ts_1_promoted_to_fp16, y = var_265_cast_fp16)[name = tensor("ts_3_cast_fp16")]; + tensor var_272 = const()[name = tensor("op_272"), val = tensor([1, 256, 1, 1])]; + tensor ts_5_cast_fp16 = reshape(shape = var_272, x = ts_3_cast_fp16)[name = tensor("ts_5_cast_fp16")]; + tensor var_279 = const()[name = tensor("op_279"), val = tensor([1, 256, 16, 32, 2])]; + tensor q_complex_1_cast_fp16 = reshape(shape = var_279, x = q_1_cast_fp16)[name = tensor("q_complex_1_cast_fp16")]; + tensor var_286 = const()[name = tensor("op_286"), val = tensor([1, 256, 16, 32, 2])]; + tensor k_complex_1_cast_fp16 = reshape(shape = var_286, x = k_1_cast_fp16)[name = tensor("k_complex_1_cast_fp16")]; + tensor var_290_begin_0 = const()[name = tensor("op_290_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_290_end_0 = const()[name = tensor("op_290_end_0"), val = tensor([1, 256, 16, 32, 1])]; + tensor var_290_end_mask_0 = const()[name = tensor("op_290_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_290_squeeze_mask_0 = const()[name = tensor("op_290_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_290_cast_fp16 = slice_by_index(begin = var_290_begin_0, end = var_290_end_0, end_mask = var_290_end_mask_0, squeeze_mask = var_290_squeeze_mask_0, x = q_complex_1_cast_fp16)[name = tensor("op_290_cast_fp16")]; + tensor var_298_begin_0 = const()[name = tensor("op_298_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_298_end_0 = const()[name = tensor("op_298_end_0"), val = tensor([1, 256, 16, 32, 2])]; + tensor var_298_end_mask_0 = const()[name = tensor("op_298_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_298_squeeze_mask_0 = const()[name = tensor("op_298_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_298_cast_fp16 = slice_by_index(begin = var_298_begin_0, end = var_298_end_0, end_mask = var_298_end_mask_0, squeeze_mask = var_298_squeeze_mask_0, x = q_complex_1_cast_fp16)[name = tensor("op_298_cast_fp16")]; + tensor var_306_begin_0 = const()[name = tensor("op_306_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_306_end_0 = const()[name = tensor("op_306_end_0"), val = tensor([1, 256, 16, 32, 1])]; + tensor var_306_end_mask_0 = const()[name = tensor("op_306_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_306_squeeze_mask_0 = const()[name = tensor("op_306_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_306_cast_fp16 = slice_by_index(begin = var_306_begin_0, end = var_306_end_0, end_mask = var_306_end_mask_0, squeeze_mask = var_306_squeeze_mask_0, x = k_complex_1_cast_fp16)[name = tensor("op_306_cast_fp16")]; + tensor var_314_begin_0 = const()[name = tensor("op_314_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_314_end_0 = const()[name = tensor("op_314_end_0"), val = tensor([1, 256, 16, 32, 2])]; + tensor var_314_end_mask_0 = const()[name = tensor("op_314_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_314_squeeze_mask_0 = const()[name = tensor("op_314_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_314_cast_fp16 = slice_by_index(begin = var_314_begin_0, end = var_314_end_0, end_mask = var_314_end_mask_0, squeeze_mask = var_314_squeeze_mask_0, x = k_complex_1_cast_fp16)[name = tensor("op_314_cast_fp16")]; + tensor freqs_1_to_fp16 = const()[name = tensor("freqs_1_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(6302592)))]; + tensor var_320_cast_fp16 = mul(x = freqs_1_to_fp16, y = ts_5_cast_fp16)[name = tensor("op_320_cast_fp16")]; + tensor rotr_1_cast_fp16 = cos(x = var_320_cast_fp16)[name = tensor("rotr_1_cast_fp16")]; + tensor roti_1_cast_fp16 = sin(x = var_320_cast_fp16)[name = tensor("roti_1_cast_fp16")]; + tensor var_324_cast_fp16 = mul(x = var_290_cast_fp16, y = rotr_1_cast_fp16)[name = tensor("op_324_cast_fp16")]; + tensor var_325_cast_fp16 = mul(x = var_298_cast_fp16, y = roti_1_cast_fp16)[name = tensor("op_325_cast_fp16")]; + tensor qor_1_cast_fp16 = sub(x = var_324_cast_fp16, y = var_325_cast_fp16)[name = tensor("qor_1_cast_fp16")]; + tensor var_328_cast_fp16 = mul(x = var_290_cast_fp16, y = roti_1_cast_fp16)[name = tensor("op_328_cast_fp16")]; + tensor var_329_cast_fp16 = mul(x = var_298_cast_fp16, y = rotr_1_cast_fp16)[name = tensor("op_329_cast_fp16")]; + tensor qoi_1_cast_fp16 = add(x = var_328_cast_fp16, y = var_329_cast_fp16)[name = tensor("qoi_1_cast_fp16")]; + tensor var_332_cast_fp16 = mul(x = var_306_cast_fp16, y = rotr_1_cast_fp16)[name = tensor("op_332_cast_fp16")]; + tensor var_333_cast_fp16 = mul(x = var_314_cast_fp16, y = roti_1_cast_fp16)[name = tensor("op_333_cast_fp16")]; + tensor kor_1_cast_fp16 = sub(x = var_332_cast_fp16, y = var_333_cast_fp16)[name = tensor("kor_1_cast_fp16")]; + tensor var_336_cast_fp16 = mul(x = var_306_cast_fp16, y = roti_1_cast_fp16)[name = tensor("op_336_cast_fp16")]; + tensor var_337_cast_fp16 = mul(x = var_314_cast_fp16, y = rotr_1_cast_fp16)[name = tensor("op_337_cast_fp16")]; + tensor koi_1_cast_fp16 = add(x = var_336_cast_fp16, y = var_337_cast_fp16)[name = tensor("koi_1_cast_fp16")]; + tensor qo_1_axis_0 = const()[name = tensor("qo_1_axis_0"), val = tensor(-1)]; + tensor qo_1_cast_fp16 = stack(axis = qo_1_axis_0, values = (qor_1_cast_fp16, qoi_1_cast_fp16))[name = tensor("qo_1_cast_fp16")]; + tensor ko_1_axis_0 = const()[name = tensor("ko_1_axis_0"), val = tensor(-1)]; + tensor ko_1_cast_fp16 = stack(axis = ko_1_axis_0, values = (kor_1_cast_fp16, koi_1_cast_fp16))[name = tensor("ko_1_cast_fp16")]; + tensor var_370 = const()[name = tensor("op_370"), val = tensor([1, 256, 16, 64])]; + tensor q_3_cast_fp16 = reshape(shape = var_370, x = qo_1_cast_fp16)[name = tensor("q_3_cast_fp16")]; + tensor var_376 = const()[name = tensor("op_376"), val = tensor([1, 256, 16, 64])]; + tensor k_3_cast_fp16 = reshape(shape = var_376, x = ko_1_cast_fp16)[name = tensor("k_3_cast_fp16")]; + tensor var_393_promoted_to_fp16 = const()[name = tensor("op_393_promoted_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(6302720)))]; + tensor write_indices_float_1_cast_fp16 = add(x = var_265_cast_fp16, y = var_393_promoted_to_fp16)[name = tensor("write_indices_float_1_cast_fp16")]; + tensor _inversed_397_y_0_to_fp16 = const()[name = tensor("_inversed_397_y_0_to_fp16"), val = tensor(0x1p-9)]; + tensor _inversed_397_cast_fp16 = mul(x = write_indices_float_1_cast_fp16, y = _inversed_397_y_0_to_fp16)[name = tensor("_inversed_397_cast_fp16")]; + tensor var_398_cast_fp16 = floor(x = _inversed_397_cast_fp16)[name = tensor("op_398_cast_fp16")]; + tensor var_399_to_fp16 = const()[name = tensor("op_399_to_fp16"), val = tensor(0x1p+9)]; + tensor var_400_cast_fp16 = mul(x = var_398_cast_fp16, y = var_399_to_fp16)[name = tensor("op_400_cast_fp16")]; + tensor write_indices_float_3_cast_fp16 = sub(x = write_indices_float_1_cast_fp16, y = var_400_cast_fp16)[name = tensor("write_indices_float_3_cast_fp16")]; + tensor var_409 = const()[name = tensor("op_409"), val = tensor([1, 1])]; + tensor valid_len_to_fp16_dtype_0 = const()[name = tensor("valid_len_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor valid_len_to_fp16 = cast(dtype = valid_len_to_fp16_dtype_0, x = valid_len)[name = tensor("cast_661")]; + tensor var_410_cast_fp16 = reshape(shape = var_409, x = valid_len_to_fp16)[name = tensor("op_410_cast_fp16")]; + tensor token_valid_1_cast_fp16 = less(x = var_393_promoted_to_fp16, y = var_410_cast_fp16)[name = tensor("token_valid_1_cast_fp16")]; + tensor var_418_value_0_to_fp16 = const()[name = tensor("op_418_value_0_to_fp16"), val = tensor(0x1.ffp+8)]; + tensor var_418_cast_fp16 = fill_like(ref_tensor = write_indices_float_3_cast_fp16, value = var_418_value_0_to_fp16)[name = tensor("op_418_cast_fp16")]; + tensor write_indices_float_5_cast_fp16 = select(a = write_indices_float_3_cast_fp16, b = var_418_cast_fp16, cond = token_valid_1_cast_fp16)[name = tensor("write_indices_float_5_cast_fp16")]; + tensor var_424_dtype_0 = const()[name = tensor("op_424_dtype_0"), val = tensor("int32")]; + tensor var_429 = const()[name = tensor("op_429"), val = tensor([1, 256, 1, 1])]; + tensor write_indices_float_5_cast_fp16_to_int32 = cast(dtype = var_424_dtype_0, x = write_indices_float_5_cast_fp16)[name = tensor("cast_660")]; + tensor var_430 = reshape(shape = var_429, x = write_indices_float_5_cast_fp16_to_int32)[name = tensor("op_430")]; + tensor write_indices_1_reps_0 = const()[name = tensor("write_indices_1_reps_0"), val = tensor([1, 1, 16, 64])]; + tensor write_indices_1 = tile(reps = write_indices_1_reps_0, x = var_430)[name = tensor("write_indices_1")]; + tensor var_440_begin_0 = const()[name = tensor("op_440_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_440_end_0 = const()[name = tensor("op_440_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor var_440_end_mask_0 = const()[name = tensor("op_440_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_440_squeeze_mask_0 = const()[name = tensor("op_440_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor cache0_to_fp16_dtype_0 = const()[name = tensor("cache0_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor cache0_to_fp16 = cast(dtype = cache0_to_fp16_dtype_0, x = cache0)[name = tensor("cast_659")]; + tensor var_440_cast_fp16 = slice_by_index(begin = var_440_begin_0, end = var_440_end_0, end_mask = var_440_end_mask_0, squeeze_mask = var_440_squeeze_mask_0, x = cache0_to_fp16)[name = tensor("op_440_cast_fp16")]; + tensor var_442_axis_0 = const()[name = tensor("op_442_axis_0"), val = tensor(1)]; + tensor var_442_mode_0 = const()[name = tensor("op_442_mode_0"), val = tensor("update")]; + tensor var_442_validate_indices_0 = const()[name = tensor("op_442_validate_indices_0"), val = tensor(false)]; + tensor var_442_cast_fp16 = scatter_along_axis(axis = var_442_axis_0, data = var_440_cast_fp16, indices = write_indices_1, mode = var_442_mode_0, updates = k_3_cast_fp16, validate_indices = var_442_validate_indices_0)[name = tensor("op_442_cast_fp16")]; + tensor concat_2 = const()[name = tensor("concat_2"), val = tensor([0, 0, 0, 0, 0])]; + tensor concat_3 = const()[name = tensor("concat_3"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_1_internal_tensor_assign_1_stride_0 = const()[name = tensor("new_cache_1_internal_tensor_assign_1_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_1_internal_tensor_assign_1_begin_mask_0 = const()[name = tensor("new_cache_1_internal_tensor_assign_1_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_1_internal_tensor_assign_1_end_mask_0 = const()[name = tensor("new_cache_1_internal_tensor_assign_1_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_1_internal_tensor_assign_1_squeeze_mask_0 = const()[name = tensor("new_cache_1_internal_tensor_assign_1_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_46 = const()[name = tensor("shape_46"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_0 = const()[name = tensor("reduce_prod_0"), val = tensor(1048576)]; + tensor range_1d_0_start_0 = const()[name = tensor("range_1d_0_start_0"), val = tensor(0)]; + tensor range_1d_0_step_0 = const()[name = tensor("range_1d_0_step_0"), val = tensor(1)]; + tensor range_1d_0 = range_1d(end = reduce_prod_0, start = range_1d_0_start_0, step = range_1d_0_step_0)[name = tensor("range_1d_0")]; + tensor reshape_0 = reshape(shape = shape_46, x = range_1d_0)[name = tensor("reshape_0")]; + tensor slice_by_index_0 = slice_by_index(begin = concat_2, begin_mask = new_cache_1_internal_tensor_assign_1_begin_mask_0, end = concat_3, end_mask = new_cache_1_internal_tensor_assign_1_end_mask_0, squeeze_mask = new_cache_1_internal_tensor_assign_1_squeeze_mask_0, stride = new_cache_1_internal_tensor_assign_1_stride_0, x = reshape_0)[name = tensor("slice_by_index_0")]; + tensor reshape_1_shape_0 = const()[name = tensor("reshape_1_shape_0"), val = tensor([-1])]; + tensor reshape_1 = reshape(shape = reshape_1_shape_0, x = slice_by_index_0)[name = tensor("reshape_1")]; + tensor reshape_2_shape_0 = const()[name = tensor("reshape_2_shape_0"), val = tensor([-1])]; + tensor reshape_2_cast_fp16 = reshape(shape = reshape_2_shape_0, x = var_442_cast_fp16)[name = tensor("reshape_2_cast_fp16")]; + tensor reshape_3_shape_0 = const()[name = tensor("reshape_3_shape_0"), val = tensor([-1])]; + tensor reshape_3_cast_fp16 = reshape(shape = reshape_3_shape_0, x = cache0_to_fp16)[name = tensor("reshape_3_cast_fp16")]; + tensor scatter_0_mode_0 = const()[name = tensor("scatter_0_mode_0"), val = tensor("update")]; + tensor scatter_0_axis_0 = const()[name = tensor("scatter_0_axis_0"), val = tensor(0)]; + tensor scatter_0_validate_indices_0 = const()[name = tensor("scatter_0_validate_indices_0"), val = tensor(false)]; + tensor scatter_0_cast_fp16 = scatter(axis = scatter_0_axis_0, data = reshape_3_cast_fp16, indices = reshape_1, mode = scatter_0_mode_0, updates = reshape_2_cast_fp16, validate_indices = scatter_0_validate_indices_0)[name = tensor("scatter_0_cast_fp16")]; + tensor reshape_4_cast_fp16 = reshape(shape = shape_46, x = scatter_0_cast_fp16)[name = tensor("reshape_4_cast_fp16")]; + tensor var_450_begin_0 = const()[name = tensor("op_450_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor var_450_end_0 = const()[name = tensor("op_450_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor var_450_end_mask_0 = const()[name = tensor("op_450_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_450_squeeze_mask_0 = const()[name = tensor("op_450_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor var_450_cast_fp16 = slice_by_index(begin = var_450_begin_0, end = var_450_end_0, end_mask = var_450_end_mask_0, squeeze_mask = var_450_squeeze_mask_0, x = reshape_4_cast_fp16)[name = tensor("op_450_cast_fp16")]; + tensor var_452_axis_0 = const()[name = tensor("op_452_axis_0"), val = tensor(1)]; + tensor var_452_mode_0 = const()[name = tensor("op_452_mode_0"), val = tensor("update")]; + tensor var_452_validate_indices_0 = const()[name = tensor("op_452_validate_indices_0"), val = tensor(false)]; + tensor var_452_cast_fp16 = scatter_along_axis(axis = var_452_axis_0, data = var_450_cast_fp16, indices = write_indices_1, mode = var_452_mode_0, updates = v_1_cast_fp16, validate_indices = var_452_validate_indices_0)[name = tensor("op_452_cast_fp16")]; + tensor concat_4 = const()[name = tensor("concat_4"), val = tensor([1, 0, 0, 0, 0])]; + tensor concat_5 = const()[name = tensor("concat_5"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_1_internal_tensor_assign_2_stride_0 = const()[name = tensor("new_cache_1_internal_tensor_assign_2_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_1_internal_tensor_assign_2_begin_mask_0 = const()[name = tensor("new_cache_1_internal_tensor_assign_2_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_1_internal_tensor_assign_2_end_mask_0 = const()[name = tensor("new_cache_1_internal_tensor_assign_2_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_1_internal_tensor_assign_2_squeeze_mask_0 = const()[name = tensor("new_cache_1_internal_tensor_assign_2_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_47 = const()[name = tensor("shape_47"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_1 = const()[name = tensor("reduce_prod_1"), val = tensor(1048576)]; + tensor range_1d_1_start_0 = const()[name = tensor("range_1d_1_start_0"), val = tensor(0)]; + tensor range_1d_1_step_0 = const()[name = tensor("range_1d_1_step_0"), val = tensor(1)]; + tensor range_1d_1 = range_1d(end = reduce_prod_1, start = range_1d_1_start_0, step = range_1d_1_step_0)[name = tensor("range_1d_1")]; + tensor reshape_5 = reshape(shape = shape_47, x = range_1d_1)[name = tensor("reshape_5")]; + tensor slice_by_index_1 = slice_by_index(begin = concat_4, begin_mask = new_cache_1_internal_tensor_assign_2_begin_mask_0, end = concat_5, end_mask = new_cache_1_internal_tensor_assign_2_end_mask_0, squeeze_mask = new_cache_1_internal_tensor_assign_2_squeeze_mask_0, stride = new_cache_1_internal_tensor_assign_2_stride_0, x = reshape_5)[name = tensor("slice_by_index_1")]; + tensor reshape_6_shape_0 = const()[name = tensor("reshape_6_shape_0"), val = tensor([-1])]; + tensor reshape_6 = reshape(shape = reshape_6_shape_0, x = slice_by_index_1)[name = tensor("reshape_6")]; + tensor reshape_7_shape_0 = const()[name = tensor("reshape_7_shape_0"), val = tensor([-1])]; + tensor reshape_7_cast_fp16 = reshape(shape = reshape_7_shape_0, x = var_452_cast_fp16)[name = tensor("reshape_7_cast_fp16")]; + tensor reshape_8_shape_0 = const()[name = tensor("reshape_8_shape_0"), val = tensor([-1])]; + tensor reshape_8_cast_fp16 = reshape(shape = reshape_8_shape_0, x = reshape_4_cast_fp16)[name = tensor("reshape_8_cast_fp16")]; + tensor scatter_1_mode_0 = const()[name = tensor("scatter_1_mode_0"), val = tensor("update")]; + tensor scatter_1_axis_0 = const()[name = tensor("scatter_1_axis_0"), val = tensor(0)]; + tensor scatter_1_validate_indices_0 = const()[name = tensor("scatter_1_validate_indices_0"), val = tensor(false)]; + tensor scatter_1_cast_fp16 = scatter(axis = scatter_1_axis_0, data = reshape_8_cast_fp16, indices = reshape_6, mode = scatter_1_mode_0, updates = reshape_7_cast_fp16, validate_indices = scatter_1_validate_indices_0)[name = tensor("scatter_1_cast_fp16")]; + tensor reshape_9_cast_fp16 = reshape(shape = shape_47, x = scatter_1_cast_fp16)[name = tensor("reshape_9_cast_fp16")]; + tensor reshape_9_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("reshape_9_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor keys_1_begin_0 = const()[name = tensor("keys_1_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor keys_1_end_0 = const()[name = tensor("keys_1_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor keys_1_end_mask_0 = const()[name = tensor("keys_1_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor keys_1_squeeze_mask_0 = const()[name = tensor("keys_1_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor keys_1_cast_fp16 = slice_by_index(begin = keys_1_begin_0, end = keys_1_end_0, end_mask = keys_1_end_mask_0, squeeze_mask = keys_1_squeeze_mask_0, x = reshape_9_cast_fp16)[name = tensor("keys_1_cast_fp16")]; + tensor values_1_begin_0 = const()[name = tensor("values_1_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor values_1_end_0 = const()[name = tensor("values_1_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor values_1_end_mask_0 = const()[name = tensor("values_1_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor values_1_squeeze_mask_0 = const()[name = tensor("values_1_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor values_1_cast_fp16 = slice_by_index(begin = values_1_begin_0, end = values_1_end_0, end_mask = values_1_end_mask_0, squeeze_mask = values_1_squeeze_mask_0, x = reshape_9_cast_fp16)[name = tensor("values_1_cast_fp16")]; + tensor var_464_cast_fp16 = not_equal(x = keys_1_cast_fp16, y = keys_1_cast_fp16)[name = tensor("op_464_cast_fp16")]; + tensor var_470_to_fp16 = const()[name = tensor("op_470_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(6303296)))]; + tensor keys_3_cast_fp16 = select(a = var_470_to_fp16, b = keys_1_cast_fp16, cond = var_464_cast_fp16)[name = tensor("keys_3_cast_fp16")]; + tensor var_472_cast_fp16 = not_equal(x = values_1_cast_fp16, y = values_1_cast_fp16)[name = tensor("op_472_cast_fp16")]; + tensor values_3_cast_fp16 = select(a = var_470_to_fp16, b = values_1_cast_fp16, cond = var_472_cast_fp16)[name = tensor("values_3_cast_fp16")]; + tensor var_496 = const()[name = tensor("op_496"), val = tensor([0, 2, 1, 3])]; + tensor var_512 = const()[name = tensor("op_512"), val = tensor([1, 1, 1])]; + tensor var_513_cast_fp16 = reshape(shape = var_512, x = position0_to_fp16)[name = tensor("op_513_cast_fp16")]; + tensor q_offsets_1_promoted_to_fp16 = const()[name = tensor("q_offsets_1_promoted_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(7351936)))]; + tensor q_positions_1_cast_fp16 = add(x = var_513_cast_fp16, y = q_offsets_1_promoted_to_fp16)[name = tensor("q_positions_1_cast_fp16")]; + tensor var_535 = const()[name = tensor("op_535"), val = tensor([1, 1, 1])]; + tensor var_536_cast_fp16 = reshape(shape = var_535, x = valid_len_to_fp16)[name = tensor("op_536_cast_fp16")]; + tensor valid_end_1_cast_fp16 = add(x = var_513_cast_fp16, y = var_536_cast_fp16)[name = tensor("valid_end_1_cast_fp16")]; + tensor k_positions_1_promoted_to_fp16 = const()[name = tensor("k_positions_1_promoted_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(7352512)))]; + tensor valid_mask_1_cast_fp16 = less(x = k_positions_1_promoted_to_fp16, y = valid_end_1_cast_fp16)[name = tensor("valid_mask_1_cast_fp16")]; + tensor causal_mask_1_cast_fp16 = less_equal(x = k_positions_1_promoted_to_fp16, y = q_positions_1_cast_fp16)[name = tensor("causal_mask_1_cast_fp16")]; + tensor var_541 = logical_and(x = valid_mask_1_cast_fp16, y = causal_mask_1_cast_fp16)[name = tensor("op_541")]; + tensor attn_mask_1_axes_0 = const()[name = tensor("attn_mask_1_axes_0"), val = tensor([1])]; + tensor attn_mask_1 = expand_dims(axes = attn_mask_1_axes_0, x = var_541)[name = tensor("attn_mask_1")]; + tensor var_555_transpose_x_0 = const()[name = tensor("op_555_transpose_x_0"), val = tensor(false)]; + tensor var_555_transpose_y_0 = const()[name = tensor("op_555_transpose_y_0"), val = tensor(false)]; + tensor transpose_69_perm_0 = const()[name = tensor("transpose_69_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_70_perm_0 = const()[name = tensor("transpose_70_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_70 = transpose(perm = transpose_70_perm_0, x = keys_3_cast_fp16)[name = tensor("transpose_204")]; + tensor transpose_69 = transpose(perm = transpose_69_perm_0, x = q_3_cast_fp16)[name = tensor("transpose_205")]; + tensor var_555_cast_fp16 = matmul(transpose_x = var_555_transpose_x_0, transpose_y = var_555_transpose_y_0, x = transpose_69, y = transpose_70)[name = tensor("op_555_cast_fp16")]; + tensor var_549_to_fp16 = const()[name = tensor("op_549_to_fp16"), val = tensor([0x1p-3])]; + tensor attn_weights_1_cast_fp16 = mul(x = var_555_cast_fp16, y = var_549_to_fp16)[name = tensor("attn_weights_1_cast_fp16")]; + tensor var_557 = logical_not(x = attn_mask_1)[name = tensor("op_557")]; + tensor var_558_to_fp16 = const()[name = tensor("op_558_to_fp16"), val = tensor(-inf)]; + tensor attn_weights_3_cast_fp16 = select(a = var_558_to_fp16, b = attn_weights_1_cast_fp16, cond = var_557)[name = tensor("attn_weights_3_cast_fp16")]; + tensor var_560 = const()[name = tensor("op_560"), val = tensor(-1)]; + tensor attn_weights_5_cast_fp16 = softmax(axis = var_560, x = attn_weights_3_cast_fp16)[name = tensor("attn_weights_5_cast_fp16")]; + tensor attn_output_1_transpose_x_0 = const()[name = tensor("attn_output_1_transpose_x_0"), val = tensor(false)]; + tensor attn_output_1_transpose_y_0 = const()[name = tensor("attn_output_1_transpose_y_0"), val = tensor(false)]; + tensor values_5_cast_fp16 = transpose(perm = var_496, x = values_3_cast_fp16)[name = tensor("transpose_206")]; + tensor attn_output_1_cast_fp16 = matmul(transpose_x = attn_output_1_transpose_x_0, transpose_y = attn_output_1_transpose_y_0, x = attn_weights_5_cast_fp16, y = values_5_cast_fp16)[name = tensor("attn_output_1_cast_fp16")]; + tensor var_568 = const()[name = tensor("op_568"), val = tensor([0, 2, 1, 3])]; + tensor var_573 = const()[name = tensor("op_573"), val = tensor([1, 256, 1024])]; + tensor var_569_cast_fp16 = transpose(perm = var_568, x = attn_output_1_cast_fp16)[name = tensor("transpose_203")]; + tensor input_5_cast_fp16 = reshape(shape = var_573, x = var_569_cast_fp16)[name = tensor("input_5_cast_fp16")]; + tensor attn0_out_proj_weight_to_fp16 = const()[name = tensor("attn0_out_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(7353600)))]; + tensor linear_1_bias_0_to_fp16 = const()[name = tensor("linear_1_bias_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9450816)))]; + tensor linear_1_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = attn0_out_proj_weight_to_fp16, x = input_5_cast_fp16)[name = tensor("linear_1_cast_fp16")]; + tensor var_582_cast_fp16 = add(x = position0_to_fp16, y = valid_len_to_fp16)[name = tensor("op_582_cast_fp16")]; + tensor var_582_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_582_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor input_7_cast_fp16 = add(x = conditioning_to_fp16, y = linear_1_cast_fp16)[name = tensor("input_7_cast_fp16")]; + tensor input_9_axes_0 = const()[name = tensor("input_9_axes_0"), val = tensor([-1])]; + tensor norm0_2_weight_to_fp16 = const()[name = tensor("norm0_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9452928)))]; + tensor norm0_2_bias_to_fp16 = const()[name = tensor("norm0_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9455040)))]; + tensor var_586_to_fp16 = const()[name = tensor("op_586_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_9_cast_fp16 = layer_norm(axes = input_9_axes_0, beta = norm0_2_bias_to_fp16, epsilon = var_586_to_fp16, gamma = norm0_2_weight_to_fp16, x = input_7_cast_fp16)[name = tensor("input_9_cast_fp16")]; + tensor linear0_1_weight_to_fp16 = const()[name = tensor("linear0_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9457152)))]; + tensor linear_2_bias_0_to_fp16 = const()[name = tensor("linear_2_bias_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(17845824)))]; + tensor linear_2_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = linear0_1_weight_to_fp16, x = input_9_cast_fp16)[name = tensor("linear_2_cast_fp16")]; + tensor input_11_mode_0 = const()[name = tensor("input_11_mode_0"), val = tensor("EXACT")]; + tensor input_11_cast_fp16 = gelu(mode = input_11_mode_0, x = linear_2_cast_fp16)[name = tensor("input_11_cast_fp16")]; + tensor linear0_2_weight_to_fp16 = const()[name = tensor("linear0_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(17854080)))]; + tensor linear_3_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = linear0_2_weight_to_fp16, x = input_11_cast_fp16)[name = tensor("linear_3_cast_fp16")]; + tensor input_13_cast_fp16 = add(x = input_7_cast_fp16, y = linear_3_cast_fp16)[name = tensor("input_13_cast_fp16")]; + tensor input_15_axes_0 = const()[name = tensor("input_15_axes_0"), val = tensor([-1])]; + tensor norm1_1_weight_to_fp16 = const()[name = tensor("norm1_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(26242752)))]; + tensor norm1_1_bias_to_fp16 = const()[name = tensor("norm1_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(26244864)))]; + tensor var_603_to_fp16 = const()[name = tensor("op_603_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_15_cast_fp16 = layer_norm(axes = input_15_axes_0, beta = norm1_1_bias_to_fp16, epsilon = var_603_to_fp16, gamma = norm1_1_weight_to_fp16, x = input_13_cast_fp16)[name = tensor("input_15_cast_fp16")]; + tensor attn1_in_proj_weight_to_fp16 = const()[name = tensor("attn1_in_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(26246976)))]; + tensor linear_4_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = attn1_in_proj_weight_to_fp16, x = input_15_cast_fp16)[name = tensor("linear_4_cast_fp16")]; + tensor var_617 = const()[name = tensor("op_617"), val = tensor([1, 256, 3, 16, 64])]; + tensor qkv_3_cast_fp16 = reshape(shape = var_617, x = linear_4_cast_fp16)[name = tensor("qkv_3_cast_fp16")]; + tensor q_7_begin_0 = const()[name = tensor("q_7_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor q_7_end_0 = const()[name = tensor("q_7_end_0"), val = tensor([1, 256, 1, 16, 64])]; + tensor q_7_end_mask_0 = const()[name = tensor("q_7_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor q_7_squeeze_mask_0 = const()[name = tensor("q_7_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor q_7_cast_fp16 = slice_by_index(begin = q_7_begin_0, end = q_7_end_0, end_mask = q_7_end_mask_0, squeeze_mask = q_7_squeeze_mask_0, x = qkv_3_cast_fp16)[name = tensor("q_7_cast_fp16")]; + tensor k_5_begin_0 = const()[name = tensor("k_5_begin_0"), val = tensor([0, 0, 1, 0, 0])]; + tensor k_5_end_0 = const()[name = tensor("k_5_end_0"), val = tensor([1, 256, 2, 16, 64])]; + tensor k_5_end_mask_0 = const()[name = tensor("k_5_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor k_5_squeeze_mask_0 = const()[name = tensor("k_5_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor k_5_cast_fp16 = slice_by_index(begin = k_5_begin_0, end = k_5_end_0, end_mask = k_5_end_mask_0, squeeze_mask = k_5_squeeze_mask_0, x = qkv_3_cast_fp16)[name = tensor("k_5_cast_fp16")]; + tensor v_3_begin_0 = const()[name = tensor("v_3_begin_0"), val = tensor([0, 0, 2, 0, 0])]; + tensor v_3_end_0 = const()[name = tensor("v_3_end_0"), val = tensor([1, 256, 3, 16, 64])]; + tensor v_3_end_mask_0 = const()[name = tensor("v_3_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor v_3_squeeze_mask_0 = const()[name = tensor("v_3_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor v_3_cast_fp16 = slice_by_index(begin = v_3_begin_0, end = v_3_end_0, end_mask = v_3_end_mask_0, squeeze_mask = v_3_squeeze_mask_0, x = qkv_3_cast_fp16)[name = tensor("v_3_cast_fp16")]; + tensor var_675 = const()[name = tensor("op_675"), val = tensor([1, 1])]; + tensor position1_to_fp16_dtype_0 = const()[name = tensor("position1_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor position1_to_fp16 = cast(dtype = position1_to_fp16_dtype_0, x = position1)[name = tensor("cast_656")]; + tensor var_676_cast_fp16 = reshape(shape = var_675, x = position1_to_fp16)[name = tensor("op_676_cast_fp16")]; + tensor ts_9_cast_fp16 = add(x = ts_1_promoted_to_fp16, y = var_676_cast_fp16)[name = tensor("ts_9_cast_fp16")]; + tensor var_683 = const()[name = tensor("op_683"), val = tensor([1, 256, 1, 1])]; + tensor ts_11_cast_fp16 = reshape(shape = var_683, x = ts_9_cast_fp16)[name = tensor("ts_11_cast_fp16")]; + tensor var_690 = const()[name = tensor("op_690"), val = tensor([1, 256, 16, 32, 2])]; + tensor q_complex_3_cast_fp16 = reshape(shape = var_690, x = q_7_cast_fp16)[name = tensor("q_complex_3_cast_fp16")]; + tensor var_697 = const()[name = tensor("op_697"), val = tensor([1, 256, 16, 32, 2])]; + tensor k_complex_3_cast_fp16 = reshape(shape = var_697, x = k_5_cast_fp16)[name = tensor("k_complex_3_cast_fp16")]; + tensor var_701_begin_0 = const()[name = tensor("op_701_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_701_end_0 = const()[name = tensor("op_701_end_0"), val = tensor([1, 256, 16, 32, 1])]; + tensor var_701_end_mask_0 = const()[name = tensor("op_701_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_701_squeeze_mask_0 = const()[name = tensor("op_701_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_701_cast_fp16 = slice_by_index(begin = var_701_begin_0, end = var_701_end_0, end_mask = var_701_end_mask_0, squeeze_mask = var_701_squeeze_mask_0, x = q_complex_3_cast_fp16)[name = tensor("op_701_cast_fp16")]; + tensor var_709_begin_0 = const()[name = tensor("op_709_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_709_end_0 = const()[name = tensor("op_709_end_0"), val = tensor([1, 256, 16, 32, 2])]; + tensor var_709_end_mask_0 = const()[name = tensor("op_709_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_709_squeeze_mask_0 = const()[name = tensor("op_709_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_709_cast_fp16 = slice_by_index(begin = var_709_begin_0, end = var_709_end_0, end_mask = var_709_end_mask_0, squeeze_mask = var_709_squeeze_mask_0, x = q_complex_3_cast_fp16)[name = tensor("op_709_cast_fp16")]; + tensor var_717_begin_0 = const()[name = tensor("op_717_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_717_end_0 = const()[name = tensor("op_717_end_0"), val = tensor([1, 256, 16, 32, 1])]; + tensor var_717_end_mask_0 = const()[name = tensor("op_717_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_717_squeeze_mask_0 = const()[name = tensor("op_717_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_717_cast_fp16 = slice_by_index(begin = var_717_begin_0, end = var_717_end_0, end_mask = var_717_end_mask_0, squeeze_mask = var_717_squeeze_mask_0, x = k_complex_3_cast_fp16)[name = tensor("op_717_cast_fp16")]; + tensor var_725_begin_0 = const()[name = tensor("op_725_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_725_end_0 = const()[name = tensor("op_725_end_0"), val = tensor([1, 256, 16, 32, 2])]; + tensor var_725_end_mask_0 = const()[name = tensor("op_725_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_725_squeeze_mask_0 = const()[name = tensor("op_725_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_725_cast_fp16 = slice_by_index(begin = var_725_begin_0, end = var_725_end_0, end_mask = var_725_end_mask_0, squeeze_mask = var_725_squeeze_mask_0, x = k_complex_3_cast_fp16)[name = tensor("op_725_cast_fp16")]; + tensor freqs_3_to_fp16 = const()[name = tensor("freqs_3_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(32538496)))]; + tensor var_731_cast_fp16 = mul(x = freqs_3_to_fp16, y = ts_11_cast_fp16)[name = tensor("op_731_cast_fp16")]; + tensor rotr_3_cast_fp16 = cos(x = var_731_cast_fp16)[name = tensor("rotr_3_cast_fp16")]; + tensor roti_3_cast_fp16 = sin(x = var_731_cast_fp16)[name = tensor("roti_3_cast_fp16")]; + tensor var_735_cast_fp16 = mul(x = var_701_cast_fp16, y = rotr_3_cast_fp16)[name = tensor("op_735_cast_fp16")]; + tensor var_736_cast_fp16 = mul(x = var_709_cast_fp16, y = roti_3_cast_fp16)[name = tensor("op_736_cast_fp16")]; + tensor qor_5_cast_fp16 = sub(x = var_735_cast_fp16, y = var_736_cast_fp16)[name = tensor("qor_5_cast_fp16")]; + tensor var_739_cast_fp16 = mul(x = var_701_cast_fp16, y = roti_3_cast_fp16)[name = tensor("op_739_cast_fp16")]; + tensor var_740_cast_fp16 = mul(x = var_709_cast_fp16, y = rotr_3_cast_fp16)[name = tensor("op_740_cast_fp16")]; + tensor qoi_5_cast_fp16 = add(x = var_739_cast_fp16, y = var_740_cast_fp16)[name = tensor("qoi_5_cast_fp16")]; + tensor var_743_cast_fp16 = mul(x = var_717_cast_fp16, y = rotr_3_cast_fp16)[name = tensor("op_743_cast_fp16")]; + tensor var_744_cast_fp16 = mul(x = var_725_cast_fp16, y = roti_3_cast_fp16)[name = tensor("op_744_cast_fp16")]; + tensor kor_5_cast_fp16 = sub(x = var_743_cast_fp16, y = var_744_cast_fp16)[name = tensor("kor_5_cast_fp16")]; + tensor var_747_cast_fp16 = mul(x = var_717_cast_fp16, y = roti_3_cast_fp16)[name = tensor("op_747_cast_fp16")]; + tensor var_748_cast_fp16 = mul(x = var_725_cast_fp16, y = rotr_3_cast_fp16)[name = tensor("op_748_cast_fp16")]; + tensor koi_5_cast_fp16 = add(x = var_747_cast_fp16, y = var_748_cast_fp16)[name = tensor("koi_5_cast_fp16")]; + tensor qo_3_axis_0 = const()[name = tensor("qo_3_axis_0"), val = tensor(-1)]; + tensor qo_3_cast_fp16 = stack(axis = qo_3_axis_0, values = (qor_5_cast_fp16, qoi_5_cast_fp16))[name = tensor("qo_3_cast_fp16")]; + tensor ko_3_axis_0 = const()[name = tensor("ko_3_axis_0"), val = tensor(-1)]; + tensor ko_3_cast_fp16 = stack(axis = ko_3_axis_0, values = (kor_5_cast_fp16, koi_5_cast_fp16))[name = tensor("ko_3_cast_fp16")]; + tensor var_781 = const()[name = tensor("op_781"), val = tensor([1, 256, 16, 64])]; + tensor q_9_cast_fp16 = reshape(shape = var_781, x = qo_3_cast_fp16)[name = tensor("q_9_cast_fp16")]; + tensor var_787 = const()[name = tensor("op_787"), val = tensor([1, 256, 16, 64])]; + tensor k_7_cast_fp16 = reshape(shape = var_787, x = ko_3_cast_fp16)[name = tensor("k_7_cast_fp16")]; + tensor write_indices_float_7_cast_fp16 = add(x = var_676_cast_fp16, y = var_393_promoted_to_fp16)[name = tensor("write_indices_float_7_cast_fp16")]; + tensor _inversed_808_y_0_to_fp16 = const()[name = tensor("_inversed_808_y_0_to_fp16"), val = tensor(0x1p-9)]; + tensor _inversed_808_cast_fp16 = mul(x = write_indices_float_7_cast_fp16, y = _inversed_808_y_0_to_fp16)[name = tensor("_inversed_808_cast_fp16")]; + tensor var_809_cast_fp16 = floor(x = _inversed_808_cast_fp16)[name = tensor("op_809_cast_fp16")]; + tensor var_810_to_fp16 = const()[name = tensor("op_810_to_fp16"), val = tensor(0x1p+9)]; + tensor var_811_cast_fp16 = mul(x = var_809_cast_fp16, y = var_810_to_fp16)[name = tensor("op_811_cast_fp16")]; + tensor write_indices_float_9_cast_fp16 = sub(x = write_indices_float_7_cast_fp16, y = var_811_cast_fp16)[name = tensor("write_indices_float_9_cast_fp16")]; + tensor var_829_value_0_to_fp16 = const()[name = tensor("op_829_value_0_to_fp16"), val = tensor(0x1.ffp+8)]; + tensor var_829_cast_fp16 = fill_like(ref_tensor = write_indices_float_9_cast_fp16, value = var_829_value_0_to_fp16)[name = tensor("op_829_cast_fp16")]; + tensor write_indices_float_11_cast_fp16 = select(a = write_indices_float_9_cast_fp16, b = var_829_cast_fp16, cond = token_valid_1_cast_fp16)[name = tensor("write_indices_float_11_cast_fp16")]; + tensor var_835_dtype_0 = const()[name = tensor("op_835_dtype_0"), val = tensor("int32")]; + tensor var_840 = const()[name = tensor("op_840"), val = tensor([1, 256, 1, 1])]; + tensor write_indices_float_11_cast_fp16_to_int32 = cast(dtype = var_835_dtype_0, x = write_indices_float_11_cast_fp16)[name = tensor("cast_655")]; + tensor var_841 = reshape(shape = var_840, x = write_indices_float_11_cast_fp16_to_int32)[name = tensor("op_841")]; + tensor write_indices_3_reps_0 = const()[name = tensor("write_indices_3_reps_0"), val = tensor([1, 1, 16, 64])]; + tensor write_indices_3 = tile(reps = write_indices_3_reps_0, x = var_841)[name = tensor("write_indices_3")]; + tensor var_851_begin_0 = const()[name = tensor("op_851_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_851_end_0 = const()[name = tensor("op_851_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor var_851_end_mask_0 = const()[name = tensor("op_851_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_851_squeeze_mask_0 = const()[name = tensor("op_851_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor cache1_to_fp16_dtype_0 = const()[name = tensor("cache1_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor cache1_to_fp16 = cast(dtype = cache1_to_fp16_dtype_0, x = cache1)[name = tensor("cast_654")]; + tensor var_851_cast_fp16 = slice_by_index(begin = var_851_begin_0, end = var_851_end_0, end_mask = var_851_end_mask_0, squeeze_mask = var_851_squeeze_mask_0, x = cache1_to_fp16)[name = tensor("op_851_cast_fp16")]; + tensor var_853_axis_0 = const()[name = tensor("op_853_axis_0"), val = tensor(1)]; + tensor var_853_mode_0 = const()[name = tensor("op_853_mode_0"), val = tensor("update")]; + tensor var_853_validate_indices_0 = const()[name = tensor("op_853_validate_indices_0"), val = tensor(false)]; + tensor var_853_cast_fp16 = scatter_along_axis(axis = var_853_axis_0, data = var_851_cast_fp16, indices = write_indices_3, mode = var_853_mode_0, updates = k_7_cast_fp16, validate_indices = var_853_validate_indices_0)[name = tensor("op_853_cast_fp16")]; + tensor concat_10 = const()[name = tensor("concat_10"), val = tensor([0, 0, 0, 0, 0])]; + tensor concat_11 = const()[name = tensor("concat_11"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_3_internal_tensor_assign_1_stride_0 = const()[name = tensor("new_cache_3_internal_tensor_assign_1_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_3_internal_tensor_assign_1_begin_mask_0 = const()[name = tensor("new_cache_3_internal_tensor_assign_1_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_3_internal_tensor_assign_1_end_mask_0 = const()[name = tensor("new_cache_3_internal_tensor_assign_1_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_3_internal_tensor_assign_1_squeeze_mask_0 = const()[name = tensor("new_cache_3_internal_tensor_assign_1_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_48 = const()[name = tensor("shape_48"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_2 = const()[name = tensor("reduce_prod_2"), val = tensor(1048576)]; + tensor range_1d_2_start_0 = const()[name = tensor("range_1d_2_start_0"), val = tensor(0)]; + tensor range_1d_2_step_0 = const()[name = tensor("range_1d_2_step_0"), val = tensor(1)]; + tensor range_1d_2 = range_1d(end = reduce_prod_2, start = range_1d_2_start_0, step = range_1d_2_step_0)[name = tensor("range_1d_2")]; + tensor reshape_10 = reshape(shape = shape_48, x = range_1d_2)[name = tensor("reshape_10")]; + tensor slice_by_index_2 = slice_by_index(begin = concat_10, begin_mask = new_cache_3_internal_tensor_assign_1_begin_mask_0, end = concat_11, end_mask = new_cache_3_internal_tensor_assign_1_end_mask_0, squeeze_mask = new_cache_3_internal_tensor_assign_1_squeeze_mask_0, stride = new_cache_3_internal_tensor_assign_1_stride_0, x = reshape_10)[name = tensor("slice_by_index_2")]; + tensor reshape_11_shape_0 = const()[name = tensor("reshape_11_shape_0"), val = tensor([-1])]; + tensor reshape_11 = reshape(shape = reshape_11_shape_0, x = slice_by_index_2)[name = tensor("reshape_11")]; + tensor reshape_12_shape_0 = const()[name = tensor("reshape_12_shape_0"), val = tensor([-1])]; + tensor reshape_12_cast_fp16 = reshape(shape = reshape_12_shape_0, x = var_853_cast_fp16)[name = tensor("reshape_12_cast_fp16")]; + tensor reshape_13_shape_0 = const()[name = tensor("reshape_13_shape_0"), val = tensor([-1])]; + tensor reshape_13_cast_fp16 = reshape(shape = reshape_13_shape_0, x = cache1_to_fp16)[name = tensor("reshape_13_cast_fp16")]; + tensor scatter_2_mode_0 = const()[name = tensor("scatter_2_mode_0"), val = tensor("update")]; + tensor scatter_2_axis_0 = const()[name = tensor("scatter_2_axis_0"), val = tensor(0)]; + tensor scatter_2_validate_indices_0 = const()[name = tensor("scatter_2_validate_indices_0"), val = tensor(false)]; + tensor scatter_2_cast_fp16 = scatter(axis = scatter_2_axis_0, data = reshape_13_cast_fp16, indices = reshape_11, mode = scatter_2_mode_0, updates = reshape_12_cast_fp16, validate_indices = scatter_2_validate_indices_0)[name = tensor("scatter_2_cast_fp16")]; + tensor reshape_14_cast_fp16 = reshape(shape = shape_48, x = scatter_2_cast_fp16)[name = tensor("reshape_14_cast_fp16")]; + tensor var_861_begin_0 = const()[name = tensor("op_861_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor var_861_end_0 = const()[name = tensor("op_861_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor var_861_end_mask_0 = const()[name = tensor("op_861_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_861_squeeze_mask_0 = const()[name = tensor("op_861_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor var_861_cast_fp16 = slice_by_index(begin = var_861_begin_0, end = var_861_end_0, end_mask = var_861_end_mask_0, squeeze_mask = var_861_squeeze_mask_0, x = reshape_14_cast_fp16)[name = tensor("op_861_cast_fp16")]; + tensor var_863_axis_0 = const()[name = tensor("op_863_axis_0"), val = tensor(1)]; + tensor var_863_mode_0 = const()[name = tensor("op_863_mode_0"), val = tensor("update")]; + tensor var_863_validate_indices_0 = const()[name = tensor("op_863_validate_indices_0"), val = tensor(false)]; + tensor var_863_cast_fp16 = scatter_along_axis(axis = var_863_axis_0, data = var_861_cast_fp16, indices = write_indices_3, mode = var_863_mode_0, updates = v_3_cast_fp16, validate_indices = var_863_validate_indices_0)[name = tensor("op_863_cast_fp16")]; + tensor concat_12 = const()[name = tensor("concat_12"), val = tensor([1, 0, 0, 0, 0])]; + tensor concat_13 = const()[name = tensor("concat_13"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_3_internal_tensor_assign_2_stride_0 = const()[name = tensor("new_cache_3_internal_tensor_assign_2_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_3_internal_tensor_assign_2_begin_mask_0 = const()[name = tensor("new_cache_3_internal_tensor_assign_2_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_3_internal_tensor_assign_2_end_mask_0 = const()[name = tensor("new_cache_3_internal_tensor_assign_2_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_3_internal_tensor_assign_2_squeeze_mask_0 = const()[name = tensor("new_cache_3_internal_tensor_assign_2_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_49 = const()[name = tensor("shape_49"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_3 = const()[name = tensor("reduce_prod_3"), val = tensor(1048576)]; + tensor range_1d_3_start_0 = const()[name = tensor("range_1d_3_start_0"), val = tensor(0)]; + tensor range_1d_3_step_0 = const()[name = tensor("range_1d_3_step_0"), val = tensor(1)]; + tensor range_1d_3 = range_1d(end = reduce_prod_3, start = range_1d_3_start_0, step = range_1d_3_step_0)[name = tensor("range_1d_3")]; + tensor reshape_15 = reshape(shape = shape_49, x = range_1d_3)[name = tensor("reshape_15")]; + tensor slice_by_index_3 = slice_by_index(begin = concat_12, begin_mask = new_cache_3_internal_tensor_assign_2_begin_mask_0, end = concat_13, end_mask = new_cache_3_internal_tensor_assign_2_end_mask_0, squeeze_mask = new_cache_3_internal_tensor_assign_2_squeeze_mask_0, stride = new_cache_3_internal_tensor_assign_2_stride_0, x = reshape_15)[name = tensor("slice_by_index_3")]; + tensor reshape_16_shape_0 = const()[name = tensor("reshape_16_shape_0"), val = tensor([-1])]; + tensor reshape_16 = reshape(shape = reshape_16_shape_0, x = slice_by_index_3)[name = tensor("reshape_16")]; + tensor reshape_17_shape_0 = const()[name = tensor("reshape_17_shape_0"), val = tensor([-1])]; + tensor reshape_17_cast_fp16 = reshape(shape = reshape_17_shape_0, x = var_863_cast_fp16)[name = tensor("reshape_17_cast_fp16")]; + tensor reshape_18_shape_0 = const()[name = tensor("reshape_18_shape_0"), val = tensor([-1])]; + tensor reshape_18_cast_fp16 = reshape(shape = reshape_18_shape_0, x = reshape_14_cast_fp16)[name = tensor("reshape_18_cast_fp16")]; + tensor scatter_3_mode_0 = const()[name = tensor("scatter_3_mode_0"), val = tensor("update")]; + tensor scatter_3_axis_0 = const()[name = tensor("scatter_3_axis_0"), val = tensor(0)]; + tensor scatter_3_validate_indices_0 = const()[name = tensor("scatter_3_validate_indices_0"), val = tensor(false)]; + tensor scatter_3_cast_fp16 = scatter(axis = scatter_3_axis_0, data = reshape_18_cast_fp16, indices = reshape_16, mode = scatter_3_mode_0, updates = reshape_17_cast_fp16, validate_indices = scatter_3_validate_indices_0)[name = tensor("scatter_3_cast_fp16")]; + tensor reshape_19_cast_fp16 = reshape(shape = shape_49, x = scatter_3_cast_fp16)[name = tensor("reshape_19_cast_fp16")]; + tensor reshape_19_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("reshape_19_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor keys_7_begin_0 = const()[name = tensor("keys_7_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor keys_7_end_0 = const()[name = tensor("keys_7_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor keys_7_end_mask_0 = const()[name = tensor("keys_7_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor keys_7_squeeze_mask_0 = const()[name = tensor("keys_7_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor keys_7_cast_fp16 = slice_by_index(begin = keys_7_begin_0, end = keys_7_end_0, end_mask = keys_7_end_mask_0, squeeze_mask = keys_7_squeeze_mask_0, x = reshape_19_cast_fp16)[name = tensor("keys_7_cast_fp16")]; + tensor values_7_begin_0 = const()[name = tensor("values_7_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor values_7_end_0 = const()[name = tensor("values_7_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor values_7_end_mask_0 = const()[name = tensor("values_7_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor values_7_squeeze_mask_0 = const()[name = tensor("values_7_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor values_7_cast_fp16 = slice_by_index(begin = values_7_begin_0, end = values_7_end_0, end_mask = values_7_end_mask_0, squeeze_mask = values_7_squeeze_mask_0, x = reshape_19_cast_fp16)[name = tensor("values_7_cast_fp16")]; + tensor var_875_cast_fp16 = not_equal(x = keys_7_cast_fp16, y = keys_7_cast_fp16)[name = tensor("op_875_cast_fp16")]; + tensor keys_9_cast_fp16 = select(a = var_470_to_fp16, b = keys_7_cast_fp16, cond = var_875_cast_fp16)[name = tensor("keys_9_cast_fp16")]; + tensor var_883_cast_fp16 = not_equal(x = values_7_cast_fp16, y = values_7_cast_fp16)[name = tensor("op_883_cast_fp16")]; + tensor values_9_cast_fp16 = select(a = var_470_to_fp16, b = values_7_cast_fp16, cond = var_883_cast_fp16)[name = tensor("values_9_cast_fp16")]; + tensor var_907 = const()[name = tensor("op_907"), val = tensor([0, 2, 1, 3])]; + tensor var_923 = const()[name = tensor("op_923"), val = tensor([1, 1, 1])]; + tensor var_924_cast_fp16 = reshape(shape = var_923, x = position1_to_fp16)[name = tensor("op_924_cast_fp16")]; + tensor q_positions_3_cast_fp16 = add(x = var_924_cast_fp16, y = q_offsets_1_promoted_to_fp16)[name = tensor("q_positions_3_cast_fp16")]; + tensor valid_end_3_cast_fp16 = add(x = var_924_cast_fp16, y = var_536_cast_fp16)[name = tensor("valid_end_3_cast_fp16")]; + tensor valid_mask_3_cast_fp16 = less(x = k_positions_1_promoted_to_fp16, y = valid_end_3_cast_fp16)[name = tensor("valid_mask_3_cast_fp16")]; + tensor causal_mask_3_cast_fp16 = less_equal(x = k_positions_1_promoted_to_fp16, y = q_positions_3_cast_fp16)[name = tensor("causal_mask_3_cast_fp16")]; + tensor var_952 = logical_and(x = valid_mask_3_cast_fp16, y = causal_mask_3_cast_fp16)[name = tensor("op_952")]; + tensor attn_mask_3_axes_0 = const()[name = tensor("attn_mask_3_axes_0"), val = tensor([1])]; + tensor attn_mask_3 = expand_dims(axes = attn_mask_3_axes_0, x = var_952)[name = tensor("attn_mask_3")]; + tensor var_966_transpose_x_0 = const()[name = tensor("op_966_transpose_x_0"), val = tensor(false)]; + tensor var_966_transpose_y_0 = const()[name = tensor("op_966_transpose_y_0"), val = tensor(false)]; + tensor transpose_71_perm_0 = const()[name = tensor("transpose_71_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_72_perm_0 = const()[name = tensor("transpose_72_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_72 = transpose(perm = transpose_72_perm_0, x = keys_9_cast_fp16)[name = tensor("transpose_200")]; + tensor transpose_71 = transpose(perm = transpose_71_perm_0, x = q_9_cast_fp16)[name = tensor("transpose_201")]; + tensor var_966_cast_fp16 = matmul(transpose_x = var_966_transpose_x_0, transpose_y = var_966_transpose_y_0, x = transpose_71, y = transpose_72)[name = tensor("op_966_cast_fp16")]; + tensor var_960_to_fp16 = const()[name = tensor("op_960_to_fp16"), val = tensor([0x1p-3])]; + tensor attn_weights_7_cast_fp16 = mul(x = var_966_cast_fp16, y = var_960_to_fp16)[name = tensor("attn_weights_7_cast_fp16")]; + tensor var_968 = logical_not(x = attn_mask_3)[name = tensor("op_968")]; + tensor var_969_to_fp16 = const()[name = tensor("op_969_to_fp16"), val = tensor(-inf)]; + tensor attn_weights_9_cast_fp16 = select(a = var_969_to_fp16, b = attn_weights_7_cast_fp16, cond = var_968)[name = tensor("attn_weights_9_cast_fp16")]; + tensor var_971 = const()[name = tensor("op_971"), val = tensor(-1)]; + tensor attn_weights_11_cast_fp16 = softmax(axis = var_971, x = attn_weights_9_cast_fp16)[name = tensor("attn_weights_11_cast_fp16")]; + tensor attn_output_3_transpose_x_0 = const()[name = tensor("attn_output_3_transpose_x_0"), val = tensor(false)]; + tensor attn_output_3_transpose_y_0 = const()[name = tensor("attn_output_3_transpose_y_0"), val = tensor(false)]; + tensor values_11_cast_fp16 = transpose(perm = var_907, x = values_9_cast_fp16)[name = tensor("transpose_202")]; + tensor attn_output_3_cast_fp16 = matmul(transpose_x = attn_output_3_transpose_x_0, transpose_y = attn_output_3_transpose_y_0, x = attn_weights_11_cast_fp16, y = values_11_cast_fp16)[name = tensor("attn_output_3_cast_fp16")]; + tensor var_979 = const()[name = tensor("op_979"), val = tensor([0, 2, 1, 3])]; + tensor var_984 = const()[name = tensor("op_984"), val = tensor([1, 256, 1024])]; + tensor var_980_cast_fp16 = transpose(perm = var_979, x = attn_output_3_cast_fp16)[name = tensor("transpose_199")]; + tensor input_17_cast_fp16 = reshape(shape = var_984, x = var_980_cast_fp16)[name = tensor("input_17_cast_fp16")]; + tensor attn1_out_proj_weight_to_fp16 = const()[name = tensor("attn1_out_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(32538624)))]; + tensor linear_5_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = attn1_out_proj_weight_to_fp16, x = input_17_cast_fp16)[name = tensor("linear_5_cast_fp16")]; + tensor var_993_cast_fp16 = add(x = position1_to_fp16, y = valid_len_to_fp16)[name = tensor("op_993_cast_fp16")]; + tensor var_993_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_993_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor input_19_cast_fp16 = add(x = input_13_cast_fp16, y = linear_5_cast_fp16)[name = tensor("input_19_cast_fp16")]; + tensor input_21_axes_0 = const()[name = tensor("input_21_axes_0"), val = tensor([-1])]; + tensor norm1_2_weight_to_fp16 = const()[name = tensor("norm1_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(34635840)))]; + tensor norm1_2_bias_to_fp16 = const()[name = tensor("norm1_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(34637952)))]; + tensor var_997_to_fp16 = const()[name = tensor("op_997_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_21_cast_fp16 = layer_norm(axes = input_21_axes_0, beta = norm1_2_bias_to_fp16, epsilon = var_997_to_fp16, gamma = norm1_2_weight_to_fp16, x = input_19_cast_fp16)[name = tensor("input_21_cast_fp16")]; + tensor linear1_1_weight_to_fp16 = const()[name = tensor("linear1_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(34640064)))]; + tensor linear_6_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = linear1_1_weight_to_fp16, x = input_21_cast_fp16)[name = tensor("linear_6_cast_fp16")]; + tensor input_23_mode_0 = const()[name = tensor("input_23_mode_0"), val = tensor("EXACT")]; + tensor input_23_cast_fp16 = gelu(mode = input_23_mode_0, x = linear_6_cast_fp16)[name = tensor("input_23_cast_fp16")]; + tensor linear1_2_weight_to_fp16 = const()[name = tensor("linear1_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(43028736)))]; + tensor linear_7_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = linear1_2_weight_to_fp16, x = input_23_cast_fp16)[name = tensor("linear_7_cast_fp16")]; + tensor input_25_cast_fp16 = add(x = input_19_cast_fp16, y = linear_7_cast_fp16)[name = tensor("input_25_cast_fp16")]; + tensor input_27_axes_0 = const()[name = tensor("input_27_axes_0"), val = tensor([-1])]; + tensor norm2_1_weight_to_fp16 = const()[name = tensor("norm2_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(51417408)))]; + tensor norm2_1_bias_to_fp16 = const()[name = tensor("norm2_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(51419520)))]; + tensor var_1014_to_fp16 = const()[name = tensor("op_1014_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_27_cast_fp16 = layer_norm(axes = input_27_axes_0, beta = norm2_1_bias_to_fp16, epsilon = var_1014_to_fp16, gamma = norm2_1_weight_to_fp16, x = input_25_cast_fp16)[name = tensor("input_27_cast_fp16")]; + tensor attn2_in_proj_weight_to_fp16 = const()[name = tensor("attn2_in_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(51421632)))]; + tensor linear_8_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = attn2_in_proj_weight_to_fp16, x = input_27_cast_fp16)[name = tensor("linear_8_cast_fp16")]; + tensor var_1028 = const()[name = tensor("op_1028"), val = tensor([1, 256, 3, 16, 64])]; + tensor qkv_5_cast_fp16 = reshape(shape = var_1028, x = linear_8_cast_fp16)[name = tensor("qkv_5_cast_fp16")]; + tensor q_13_begin_0 = const()[name = tensor("q_13_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor q_13_end_0 = const()[name = tensor("q_13_end_0"), val = tensor([1, 256, 1, 16, 64])]; + tensor q_13_end_mask_0 = const()[name = tensor("q_13_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor q_13_squeeze_mask_0 = const()[name = tensor("q_13_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor q_13_cast_fp16 = slice_by_index(begin = q_13_begin_0, end = q_13_end_0, end_mask = q_13_end_mask_0, squeeze_mask = q_13_squeeze_mask_0, x = qkv_5_cast_fp16)[name = tensor("q_13_cast_fp16")]; + tensor k_9_begin_0 = const()[name = tensor("k_9_begin_0"), val = tensor([0, 0, 1, 0, 0])]; + tensor k_9_end_0 = const()[name = tensor("k_9_end_0"), val = tensor([1, 256, 2, 16, 64])]; + tensor k_9_end_mask_0 = const()[name = tensor("k_9_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor k_9_squeeze_mask_0 = const()[name = tensor("k_9_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor k_9_cast_fp16 = slice_by_index(begin = k_9_begin_0, end = k_9_end_0, end_mask = k_9_end_mask_0, squeeze_mask = k_9_squeeze_mask_0, x = qkv_5_cast_fp16)[name = tensor("k_9_cast_fp16")]; + tensor v_5_begin_0 = const()[name = tensor("v_5_begin_0"), val = tensor([0, 0, 2, 0, 0])]; + tensor v_5_end_0 = const()[name = tensor("v_5_end_0"), val = tensor([1, 256, 3, 16, 64])]; + tensor v_5_end_mask_0 = const()[name = tensor("v_5_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor v_5_squeeze_mask_0 = const()[name = tensor("v_5_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor v_5_cast_fp16 = slice_by_index(begin = v_5_begin_0, end = v_5_end_0, end_mask = v_5_end_mask_0, squeeze_mask = v_5_squeeze_mask_0, x = qkv_5_cast_fp16)[name = tensor("v_5_cast_fp16")]; + tensor var_1086 = const()[name = tensor("op_1086"), val = tensor([1, 1])]; + tensor position2_to_fp16_dtype_0 = const()[name = tensor("position2_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor position2_to_fp16 = cast(dtype = position2_to_fp16_dtype_0, x = position2)[name = tensor("cast_651")]; + tensor var_1087_cast_fp16 = reshape(shape = var_1086, x = position2_to_fp16)[name = tensor("op_1087_cast_fp16")]; + tensor ts_15_cast_fp16 = add(x = ts_1_promoted_to_fp16, y = var_1087_cast_fp16)[name = tensor("ts_15_cast_fp16")]; + tensor var_1094 = const()[name = tensor("op_1094"), val = tensor([1, 256, 1, 1])]; + tensor ts_17_cast_fp16 = reshape(shape = var_1094, x = ts_15_cast_fp16)[name = tensor("ts_17_cast_fp16")]; + tensor var_1101 = const()[name = tensor("op_1101"), val = tensor([1, 256, 16, 32, 2])]; + tensor q_complex_5_cast_fp16 = reshape(shape = var_1101, x = q_13_cast_fp16)[name = tensor("q_complex_5_cast_fp16")]; + tensor var_1108 = const()[name = tensor("op_1108"), val = tensor([1, 256, 16, 32, 2])]; + tensor k_complex_5_cast_fp16 = reshape(shape = var_1108, x = k_9_cast_fp16)[name = tensor("k_complex_5_cast_fp16")]; + tensor var_1112_begin_0 = const()[name = tensor("op_1112_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_1112_end_0 = const()[name = tensor("op_1112_end_0"), val = tensor([1, 256, 16, 32, 1])]; + tensor var_1112_end_mask_0 = const()[name = tensor("op_1112_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_1112_squeeze_mask_0 = const()[name = tensor("op_1112_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_1112_cast_fp16 = slice_by_index(begin = var_1112_begin_0, end = var_1112_end_0, end_mask = var_1112_end_mask_0, squeeze_mask = var_1112_squeeze_mask_0, x = q_complex_5_cast_fp16)[name = tensor("op_1112_cast_fp16")]; + tensor var_1120_begin_0 = const()[name = tensor("op_1120_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_1120_end_0 = const()[name = tensor("op_1120_end_0"), val = tensor([1, 256, 16, 32, 2])]; + tensor var_1120_end_mask_0 = const()[name = tensor("op_1120_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_1120_squeeze_mask_0 = const()[name = tensor("op_1120_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_1120_cast_fp16 = slice_by_index(begin = var_1120_begin_0, end = var_1120_end_0, end_mask = var_1120_end_mask_0, squeeze_mask = var_1120_squeeze_mask_0, x = q_complex_5_cast_fp16)[name = tensor("op_1120_cast_fp16")]; + tensor var_1128_begin_0 = const()[name = tensor("op_1128_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_1128_end_0 = const()[name = tensor("op_1128_end_0"), val = tensor([1, 256, 16, 32, 1])]; + tensor var_1128_end_mask_0 = const()[name = tensor("op_1128_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_1128_squeeze_mask_0 = const()[name = tensor("op_1128_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_1128_cast_fp16 = slice_by_index(begin = var_1128_begin_0, end = var_1128_end_0, end_mask = var_1128_end_mask_0, squeeze_mask = var_1128_squeeze_mask_0, x = k_complex_5_cast_fp16)[name = tensor("op_1128_cast_fp16")]; + tensor var_1136_begin_0 = const()[name = tensor("op_1136_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_1136_end_0 = const()[name = tensor("op_1136_end_0"), val = tensor([1, 256, 16, 32, 2])]; + tensor var_1136_end_mask_0 = const()[name = tensor("op_1136_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_1136_squeeze_mask_0 = const()[name = tensor("op_1136_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_1136_cast_fp16 = slice_by_index(begin = var_1136_begin_0, end = var_1136_end_0, end_mask = var_1136_end_mask_0, squeeze_mask = var_1136_squeeze_mask_0, x = k_complex_5_cast_fp16)[name = tensor("op_1136_cast_fp16")]; + tensor freqs_5_to_fp16 = const()[name = tensor("freqs_5_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(57713152)))]; + tensor var_1142_cast_fp16 = mul(x = freqs_5_to_fp16, y = ts_17_cast_fp16)[name = tensor("op_1142_cast_fp16")]; + tensor rotr_5_cast_fp16 = cos(x = var_1142_cast_fp16)[name = tensor("rotr_5_cast_fp16")]; + tensor roti_5_cast_fp16 = sin(x = var_1142_cast_fp16)[name = tensor("roti_5_cast_fp16")]; + tensor var_1146_cast_fp16 = mul(x = var_1112_cast_fp16, y = rotr_5_cast_fp16)[name = tensor("op_1146_cast_fp16")]; + tensor var_1147_cast_fp16 = mul(x = var_1120_cast_fp16, y = roti_5_cast_fp16)[name = tensor("op_1147_cast_fp16")]; + tensor qor_9_cast_fp16 = sub(x = var_1146_cast_fp16, y = var_1147_cast_fp16)[name = tensor("qor_9_cast_fp16")]; + tensor var_1150_cast_fp16 = mul(x = var_1112_cast_fp16, y = roti_5_cast_fp16)[name = tensor("op_1150_cast_fp16")]; + tensor var_1151_cast_fp16 = mul(x = var_1120_cast_fp16, y = rotr_5_cast_fp16)[name = tensor("op_1151_cast_fp16")]; + tensor qoi_9_cast_fp16 = add(x = var_1150_cast_fp16, y = var_1151_cast_fp16)[name = tensor("qoi_9_cast_fp16")]; + tensor var_1154_cast_fp16 = mul(x = var_1128_cast_fp16, y = rotr_5_cast_fp16)[name = tensor("op_1154_cast_fp16")]; + tensor var_1155_cast_fp16 = mul(x = var_1136_cast_fp16, y = roti_5_cast_fp16)[name = tensor("op_1155_cast_fp16")]; + tensor kor_9_cast_fp16 = sub(x = var_1154_cast_fp16, y = var_1155_cast_fp16)[name = tensor("kor_9_cast_fp16")]; + tensor var_1158_cast_fp16 = mul(x = var_1128_cast_fp16, y = roti_5_cast_fp16)[name = tensor("op_1158_cast_fp16")]; + tensor var_1159_cast_fp16 = mul(x = var_1136_cast_fp16, y = rotr_5_cast_fp16)[name = tensor("op_1159_cast_fp16")]; + tensor koi_9_cast_fp16 = add(x = var_1158_cast_fp16, y = var_1159_cast_fp16)[name = tensor("koi_9_cast_fp16")]; + tensor qo_5_axis_0 = const()[name = tensor("qo_5_axis_0"), val = tensor(-1)]; + tensor qo_5_cast_fp16 = stack(axis = qo_5_axis_0, values = (qor_9_cast_fp16, qoi_9_cast_fp16))[name = tensor("qo_5_cast_fp16")]; + tensor ko_5_axis_0 = const()[name = tensor("ko_5_axis_0"), val = tensor(-1)]; + tensor ko_5_cast_fp16 = stack(axis = ko_5_axis_0, values = (kor_9_cast_fp16, koi_9_cast_fp16))[name = tensor("ko_5_cast_fp16")]; + tensor var_1192 = const()[name = tensor("op_1192"), val = tensor([1, 256, 16, 64])]; + tensor q_15_cast_fp16 = reshape(shape = var_1192, x = qo_5_cast_fp16)[name = tensor("q_15_cast_fp16")]; + tensor var_1198 = const()[name = tensor("op_1198"), val = tensor([1, 256, 16, 64])]; + tensor k_11_cast_fp16 = reshape(shape = var_1198, x = ko_5_cast_fp16)[name = tensor("k_11_cast_fp16")]; + tensor write_indices_float_13_cast_fp16 = add(x = var_1087_cast_fp16, y = var_393_promoted_to_fp16)[name = tensor("write_indices_float_13_cast_fp16")]; + tensor _inversed_1219_y_0_to_fp16 = const()[name = tensor("_inversed_1219_y_0_to_fp16"), val = tensor(0x1p-9)]; + tensor _inversed_1219_cast_fp16 = mul(x = write_indices_float_13_cast_fp16, y = _inversed_1219_y_0_to_fp16)[name = tensor("_inversed_1219_cast_fp16")]; + tensor var_1220_cast_fp16 = floor(x = _inversed_1219_cast_fp16)[name = tensor("op_1220_cast_fp16")]; + tensor var_1221_to_fp16 = const()[name = tensor("op_1221_to_fp16"), val = tensor(0x1p+9)]; + tensor var_1222_cast_fp16 = mul(x = var_1220_cast_fp16, y = var_1221_to_fp16)[name = tensor("op_1222_cast_fp16")]; + tensor write_indices_float_15_cast_fp16 = sub(x = write_indices_float_13_cast_fp16, y = var_1222_cast_fp16)[name = tensor("write_indices_float_15_cast_fp16")]; + tensor var_1240_value_0_to_fp16 = const()[name = tensor("op_1240_value_0_to_fp16"), val = tensor(0x1.ffp+8)]; + tensor var_1240_cast_fp16 = fill_like(ref_tensor = write_indices_float_15_cast_fp16, value = var_1240_value_0_to_fp16)[name = tensor("op_1240_cast_fp16")]; + tensor write_indices_float_17_cast_fp16 = select(a = write_indices_float_15_cast_fp16, b = var_1240_cast_fp16, cond = token_valid_1_cast_fp16)[name = tensor("write_indices_float_17_cast_fp16")]; + tensor var_1246_dtype_0 = const()[name = tensor("op_1246_dtype_0"), val = tensor("int32")]; + tensor var_1251 = const()[name = tensor("op_1251"), val = tensor([1, 256, 1, 1])]; + tensor write_indices_float_17_cast_fp16_to_int32 = cast(dtype = var_1246_dtype_0, x = write_indices_float_17_cast_fp16)[name = tensor("cast_650")]; + tensor var_1252 = reshape(shape = var_1251, x = write_indices_float_17_cast_fp16_to_int32)[name = tensor("op_1252")]; + tensor write_indices_5_reps_0 = const()[name = tensor("write_indices_5_reps_0"), val = tensor([1, 1, 16, 64])]; + tensor write_indices_5 = tile(reps = write_indices_5_reps_0, x = var_1252)[name = tensor("write_indices_5")]; + tensor var_1262_begin_0 = const()[name = tensor("op_1262_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_1262_end_0 = const()[name = tensor("op_1262_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor var_1262_end_mask_0 = const()[name = tensor("op_1262_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_1262_squeeze_mask_0 = const()[name = tensor("op_1262_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor cache2_to_fp16_dtype_0 = const()[name = tensor("cache2_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor cache2_to_fp16 = cast(dtype = cache2_to_fp16_dtype_0, x = cache2)[name = tensor("cast_649")]; + tensor var_1262_cast_fp16 = slice_by_index(begin = var_1262_begin_0, end = var_1262_end_0, end_mask = var_1262_end_mask_0, squeeze_mask = var_1262_squeeze_mask_0, x = cache2_to_fp16)[name = tensor("op_1262_cast_fp16")]; + tensor var_1264_axis_0 = const()[name = tensor("op_1264_axis_0"), val = tensor(1)]; + tensor var_1264_mode_0 = const()[name = tensor("op_1264_mode_0"), val = tensor("update")]; + tensor var_1264_validate_indices_0 = const()[name = tensor("op_1264_validate_indices_0"), val = tensor(false)]; + tensor var_1264_cast_fp16 = scatter_along_axis(axis = var_1264_axis_0, data = var_1262_cast_fp16, indices = write_indices_5, mode = var_1264_mode_0, updates = k_11_cast_fp16, validate_indices = var_1264_validate_indices_0)[name = tensor("op_1264_cast_fp16")]; + tensor concat_18 = const()[name = tensor("concat_18"), val = tensor([0, 0, 0, 0, 0])]; + tensor concat_19 = const()[name = tensor("concat_19"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_5_internal_tensor_assign_1_stride_0 = const()[name = tensor("new_cache_5_internal_tensor_assign_1_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_5_internal_tensor_assign_1_begin_mask_0 = const()[name = tensor("new_cache_5_internal_tensor_assign_1_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_5_internal_tensor_assign_1_end_mask_0 = const()[name = tensor("new_cache_5_internal_tensor_assign_1_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_5_internal_tensor_assign_1_squeeze_mask_0 = const()[name = tensor("new_cache_5_internal_tensor_assign_1_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_50 = const()[name = tensor("shape_50"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_4 = const()[name = tensor("reduce_prod_4"), val = tensor(1048576)]; + tensor range_1d_4_start_0 = const()[name = tensor("range_1d_4_start_0"), val = tensor(0)]; + tensor range_1d_4_step_0 = const()[name = tensor("range_1d_4_step_0"), val = tensor(1)]; + tensor range_1d_4 = range_1d(end = reduce_prod_4, start = range_1d_4_start_0, step = range_1d_4_step_0)[name = tensor("range_1d_4")]; + tensor reshape_20 = reshape(shape = shape_50, x = range_1d_4)[name = tensor("reshape_20")]; + tensor slice_by_index_4 = slice_by_index(begin = concat_18, begin_mask = new_cache_5_internal_tensor_assign_1_begin_mask_0, end = concat_19, end_mask = new_cache_5_internal_tensor_assign_1_end_mask_0, squeeze_mask = new_cache_5_internal_tensor_assign_1_squeeze_mask_0, stride = new_cache_5_internal_tensor_assign_1_stride_0, x = reshape_20)[name = tensor("slice_by_index_4")]; + tensor reshape_21_shape_0 = const()[name = tensor("reshape_21_shape_0"), val = tensor([-1])]; + tensor reshape_21 = reshape(shape = reshape_21_shape_0, x = slice_by_index_4)[name = tensor("reshape_21")]; + tensor reshape_22_shape_0 = const()[name = tensor("reshape_22_shape_0"), val = tensor([-1])]; + tensor reshape_22_cast_fp16 = reshape(shape = reshape_22_shape_0, x = var_1264_cast_fp16)[name = tensor("reshape_22_cast_fp16")]; + tensor reshape_23_shape_0 = const()[name = tensor("reshape_23_shape_0"), val = tensor([-1])]; + tensor reshape_23_cast_fp16 = reshape(shape = reshape_23_shape_0, x = cache2_to_fp16)[name = tensor("reshape_23_cast_fp16")]; + tensor scatter_4_mode_0 = const()[name = tensor("scatter_4_mode_0"), val = tensor("update")]; + tensor scatter_4_axis_0 = const()[name = tensor("scatter_4_axis_0"), val = tensor(0)]; + tensor scatter_4_validate_indices_0 = const()[name = tensor("scatter_4_validate_indices_0"), val = tensor(false)]; + tensor scatter_4_cast_fp16 = scatter(axis = scatter_4_axis_0, data = reshape_23_cast_fp16, indices = reshape_21, mode = scatter_4_mode_0, updates = reshape_22_cast_fp16, validate_indices = scatter_4_validate_indices_0)[name = tensor("scatter_4_cast_fp16")]; + tensor reshape_24_cast_fp16 = reshape(shape = shape_50, x = scatter_4_cast_fp16)[name = tensor("reshape_24_cast_fp16")]; + tensor var_1272_begin_0 = const()[name = tensor("op_1272_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor var_1272_end_0 = const()[name = tensor("op_1272_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor var_1272_end_mask_0 = const()[name = tensor("op_1272_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_1272_squeeze_mask_0 = const()[name = tensor("op_1272_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor var_1272_cast_fp16 = slice_by_index(begin = var_1272_begin_0, end = var_1272_end_0, end_mask = var_1272_end_mask_0, squeeze_mask = var_1272_squeeze_mask_0, x = reshape_24_cast_fp16)[name = tensor("op_1272_cast_fp16")]; + tensor var_1274_axis_0 = const()[name = tensor("op_1274_axis_0"), val = tensor(1)]; + tensor var_1274_mode_0 = const()[name = tensor("op_1274_mode_0"), val = tensor("update")]; + tensor var_1274_validate_indices_0 = const()[name = tensor("op_1274_validate_indices_0"), val = tensor(false)]; + tensor var_1274_cast_fp16 = scatter_along_axis(axis = var_1274_axis_0, data = var_1272_cast_fp16, indices = write_indices_5, mode = var_1274_mode_0, updates = v_5_cast_fp16, validate_indices = var_1274_validate_indices_0)[name = tensor("op_1274_cast_fp16")]; + tensor concat_20 = const()[name = tensor("concat_20"), val = tensor([1, 0, 0, 0, 0])]; + tensor concat_21 = const()[name = tensor("concat_21"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_5_internal_tensor_assign_2_stride_0 = const()[name = tensor("new_cache_5_internal_tensor_assign_2_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_5_internal_tensor_assign_2_begin_mask_0 = const()[name = tensor("new_cache_5_internal_tensor_assign_2_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_5_internal_tensor_assign_2_end_mask_0 = const()[name = tensor("new_cache_5_internal_tensor_assign_2_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_5_internal_tensor_assign_2_squeeze_mask_0 = const()[name = tensor("new_cache_5_internal_tensor_assign_2_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_51 = const()[name = tensor("shape_51"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_5 = const()[name = tensor("reduce_prod_5"), val = tensor(1048576)]; + tensor range_1d_5_start_0 = const()[name = tensor("range_1d_5_start_0"), val = tensor(0)]; + tensor range_1d_5_step_0 = const()[name = tensor("range_1d_5_step_0"), val = tensor(1)]; + tensor range_1d_5 = range_1d(end = reduce_prod_5, start = range_1d_5_start_0, step = range_1d_5_step_0)[name = tensor("range_1d_5")]; + tensor reshape_25 = reshape(shape = shape_51, x = range_1d_5)[name = tensor("reshape_25")]; + tensor slice_by_index_5 = slice_by_index(begin = concat_20, begin_mask = new_cache_5_internal_tensor_assign_2_begin_mask_0, end = concat_21, end_mask = new_cache_5_internal_tensor_assign_2_end_mask_0, squeeze_mask = new_cache_5_internal_tensor_assign_2_squeeze_mask_0, stride = new_cache_5_internal_tensor_assign_2_stride_0, x = reshape_25)[name = tensor("slice_by_index_5")]; + tensor reshape_26_shape_0 = const()[name = tensor("reshape_26_shape_0"), val = tensor([-1])]; + tensor reshape_26 = reshape(shape = reshape_26_shape_0, x = slice_by_index_5)[name = tensor("reshape_26")]; + tensor reshape_27_shape_0 = const()[name = tensor("reshape_27_shape_0"), val = tensor([-1])]; + tensor reshape_27_cast_fp16 = reshape(shape = reshape_27_shape_0, x = var_1274_cast_fp16)[name = tensor("reshape_27_cast_fp16")]; + tensor reshape_28_shape_0 = const()[name = tensor("reshape_28_shape_0"), val = tensor([-1])]; + tensor reshape_28_cast_fp16 = reshape(shape = reshape_28_shape_0, x = reshape_24_cast_fp16)[name = tensor("reshape_28_cast_fp16")]; + tensor scatter_5_mode_0 = const()[name = tensor("scatter_5_mode_0"), val = tensor("update")]; + tensor scatter_5_axis_0 = const()[name = tensor("scatter_5_axis_0"), val = tensor(0)]; + tensor scatter_5_validate_indices_0 = const()[name = tensor("scatter_5_validate_indices_0"), val = tensor(false)]; + tensor scatter_5_cast_fp16 = scatter(axis = scatter_5_axis_0, data = reshape_28_cast_fp16, indices = reshape_26, mode = scatter_5_mode_0, updates = reshape_27_cast_fp16, validate_indices = scatter_5_validate_indices_0)[name = tensor("scatter_5_cast_fp16")]; + tensor reshape_29_cast_fp16 = reshape(shape = shape_51, x = scatter_5_cast_fp16)[name = tensor("reshape_29_cast_fp16")]; + tensor reshape_29_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("reshape_29_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor keys_13_begin_0 = const()[name = tensor("keys_13_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor keys_13_end_0 = const()[name = tensor("keys_13_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor keys_13_end_mask_0 = const()[name = tensor("keys_13_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor keys_13_squeeze_mask_0 = const()[name = tensor("keys_13_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor keys_13_cast_fp16 = slice_by_index(begin = keys_13_begin_0, end = keys_13_end_0, end_mask = keys_13_end_mask_0, squeeze_mask = keys_13_squeeze_mask_0, x = reshape_29_cast_fp16)[name = tensor("keys_13_cast_fp16")]; + tensor values_13_begin_0 = const()[name = tensor("values_13_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor values_13_end_0 = const()[name = tensor("values_13_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor values_13_end_mask_0 = const()[name = tensor("values_13_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor values_13_squeeze_mask_0 = const()[name = tensor("values_13_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor values_13_cast_fp16 = slice_by_index(begin = values_13_begin_0, end = values_13_end_0, end_mask = values_13_end_mask_0, squeeze_mask = values_13_squeeze_mask_0, x = reshape_29_cast_fp16)[name = tensor("values_13_cast_fp16")]; + tensor var_1286_cast_fp16 = not_equal(x = keys_13_cast_fp16, y = keys_13_cast_fp16)[name = tensor("op_1286_cast_fp16")]; + tensor keys_15_cast_fp16 = select(a = var_470_to_fp16, b = keys_13_cast_fp16, cond = var_1286_cast_fp16)[name = tensor("keys_15_cast_fp16")]; + tensor var_1294_cast_fp16 = not_equal(x = values_13_cast_fp16, y = values_13_cast_fp16)[name = tensor("op_1294_cast_fp16")]; + tensor values_15_cast_fp16 = select(a = var_470_to_fp16, b = values_13_cast_fp16, cond = var_1294_cast_fp16)[name = tensor("values_15_cast_fp16")]; + tensor var_1318 = const()[name = tensor("op_1318"), val = tensor([0, 2, 1, 3])]; + tensor var_1334 = const()[name = tensor("op_1334"), val = tensor([1, 1, 1])]; + tensor var_1335_cast_fp16 = reshape(shape = var_1334, x = position2_to_fp16)[name = tensor("op_1335_cast_fp16")]; + tensor q_positions_5_cast_fp16 = add(x = var_1335_cast_fp16, y = q_offsets_1_promoted_to_fp16)[name = tensor("q_positions_5_cast_fp16")]; + tensor valid_end_5_cast_fp16 = add(x = var_1335_cast_fp16, y = var_536_cast_fp16)[name = tensor("valid_end_5_cast_fp16")]; + tensor valid_mask_5_cast_fp16 = less(x = k_positions_1_promoted_to_fp16, y = valid_end_5_cast_fp16)[name = tensor("valid_mask_5_cast_fp16")]; + tensor causal_mask_5_cast_fp16 = less_equal(x = k_positions_1_promoted_to_fp16, y = q_positions_5_cast_fp16)[name = tensor("causal_mask_5_cast_fp16")]; + tensor var_1363 = logical_and(x = valid_mask_5_cast_fp16, y = causal_mask_5_cast_fp16)[name = tensor("op_1363")]; + tensor attn_mask_5_axes_0 = const()[name = tensor("attn_mask_5_axes_0"), val = tensor([1])]; + tensor attn_mask_5 = expand_dims(axes = attn_mask_5_axes_0, x = var_1363)[name = tensor("attn_mask_5")]; + tensor var_1377_transpose_x_0 = const()[name = tensor("op_1377_transpose_x_0"), val = tensor(false)]; + tensor var_1377_transpose_y_0 = const()[name = tensor("op_1377_transpose_y_0"), val = tensor(false)]; + tensor transpose_73_perm_0 = const()[name = tensor("transpose_73_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_74_perm_0 = const()[name = tensor("transpose_74_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_74 = transpose(perm = transpose_74_perm_0, x = keys_15_cast_fp16)[name = tensor("transpose_196")]; + tensor transpose_73 = transpose(perm = transpose_73_perm_0, x = q_15_cast_fp16)[name = tensor("transpose_197")]; + tensor var_1377_cast_fp16 = matmul(transpose_x = var_1377_transpose_x_0, transpose_y = var_1377_transpose_y_0, x = transpose_73, y = transpose_74)[name = tensor("op_1377_cast_fp16")]; + tensor var_1371_to_fp16 = const()[name = tensor("op_1371_to_fp16"), val = tensor([0x1p-3])]; + tensor attn_weights_13_cast_fp16 = mul(x = var_1377_cast_fp16, y = var_1371_to_fp16)[name = tensor("attn_weights_13_cast_fp16")]; + tensor var_1379 = logical_not(x = attn_mask_5)[name = tensor("op_1379")]; + tensor var_1380_to_fp16 = const()[name = tensor("op_1380_to_fp16"), val = tensor(-inf)]; + tensor attn_weights_15_cast_fp16 = select(a = var_1380_to_fp16, b = attn_weights_13_cast_fp16, cond = var_1379)[name = tensor("attn_weights_15_cast_fp16")]; + tensor var_1382 = const()[name = tensor("op_1382"), val = tensor(-1)]; + tensor attn_weights_17_cast_fp16 = softmax(axis = var_1382, x = attn_weights_15_cast_fp16)[name = tensor("attn_weights_17_cast_fp16")]; + tensor attn_output_5_transpose_x_0 = const()[name = tensor("attn_output_5_transpose_x_0"), val = tensor(false)]; + tensor attn_output_5_transpose_y_0 = const()[name = tensor("attn_output_5_transpose_y_0"), val = tensor(false)]; + tensor values_17_cast_fp16 = transpose(perm = var_1318, x = values_15_cast_fp16)[name = tensor("transpose_198")]; + tensor attn_output_5_cast_fp16 = matmul(transpose_x = attn_output_5_transpose_x_0, transpose_y = attn_output_5_transpose_y_0, x = attn_weights_17_cast_fp16, y = values_17_cast_fp16)[name = tensor("attn_output_5_cast_fp16")]; + tensor var_1390 = const()[name = tensor("op_1390"), val = tensor([0, 2, 1, 3])]; + tensor var_1395 = const()[name = tensor("op_1395"), val = tensor([1, 256, 1024])]; + tensor var_1391_cast_fp16 = transpose(perm = var_1390, x = attn_output_5_cast_fp16)[name = tensor("transpose_195")]; + tensor input_29_cast_fp16 = reshape(shape = var_1395, x = var_1391_cast_fp16)[name = tensor("input_29_cast_fp16")]; + tensor attn2_out_proj_weight_to_fp16 = const()[name = tensor("attn2_out_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(57713280)))]; + tensor linear_9_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = attn2_out_proj_weight_to_fp16, x = input_29_cast_fp16)[name = tensor("linear_9_cast_fp16")]; + tensor var_1404_cast_fp16 = add(x = position2_to_fp16, y = valid_len_to_fp16)[name = tensor("op_1404_cast_fp16")]; + tensor var_1404_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_1404_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor input_31_cast_fp16 = add(x = input_25_cast_fp16, y = linear_9_cast_fp16)[name = tensor("input_31_cast_fp16")]; + tensor input_33_axes_0 = const()[name = tensor("input_33_axes_0"), val = tensor([-1])]; + tensor norm2_2_weight_to_fp16 = const()[name = tensor("norm2_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(59810496)))]; + tensor norm2_2_bias_to_fp16 = const()[name = tensor("norm2_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(59812608)))]; + tensor var_1408_to_fp16 = const()[name = tensor("op_1408_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_33_cast_fp16 = layer_norm(axes = input_33_axes_0, beta = norm2_2_bias_to_fp16, epsilon = var_1408_to_fp16, gamma = norm2_2_weight_to_fp16, x = input_31_cast_fp16)[name = tensor("input_33_cast_fp16")]; + tensor linear2_1_weight_to_fp16 = const()[name = tensor("linear2_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(59814720)))]; + tensor linear_10_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = linear2_1_weight_to_fp16, x = input_33_cast_fp16)[name = tensor("linear_10_cast_fp16")]; + tensor input_35_mode_0 = const()[name = tensor("input_35_mode_0"), val = tensor("EXACT")]; + tensor input_35_cast_fp16 = gelu(mode = input_35_mode_0, x = linear_10_cast_fp16)[name = tensor("input_35_cast_fp16")]; + tensor linear2_2_weight_to_fp16 = const()[name = tensor("linear2_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(68203392)))]; + tensor linear_11_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = linear2_2_weight_to_fp16, x = input_35_cast_fp16)[name = tensor("linear_11_cast_fp16")]; + tensor input_37_cast_fp16 = add(x = input_31_cast_fp16, y = linear_11_cast_fp16)[name = tensor("input_37_cast_fp16")]; + tensor input_39_axes_0 = const()[name = tensor("input_39_axes_0"), val = tensor([-1])]; + tensor norm3_1_weight_to_fp16 = const()[name = tensor("norm3_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(76592064)))]; + tensor norm3_1_bias_to_fp16 = const()[name = tensor("norm3_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(76594176)))]; + tensor var_1425_to_fp16 = const()[name = tensor("op_1425_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_39_cast_fp16 = layer_norm(axes = input_39_axes_0, beta = norm3_1_bias_to_fp16, epsilon = var_1425_to_fp16, gamma = norm3_1_weight_to_fp16, x = input_37_cast_fp16)[name = tensor("input_39_cast_fp16")]; + tensor attn3_in_proj_weight_to_fp16 = const()[name = tensor("attn3_in_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(76596288)))]; + tensor linear_12_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = attn3_in_proj_weight_to_fp16, x = input_39_cast_fp16)[name = tensor("linear_12_cast_fp16")]; + tensor var_1439 = const()[name = tensor("op_1439"), val = tensor([1, 256, 3, 16, 64])]; + tensor qkv_7_cast_fp16 = reshape(shape = var_1439, x = linear_12_cast_fp16)[name = tensor("qkv_7_cast_fp16")]; + tensor q_19_begin_0 = const()[name = tensor("q_19_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor q_19_end_0 = const()[name = tensor("q_19_end_0"), val = tensor([1, 256, 1, 16, 64])]; + tensor q_19_end_mask_0 = const()[name = tensor("q_19_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor q_19_squeeze_mask_0 = const()[name = tensor("q_19_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor q_19_cast_fp16 = slice_by_index(begin = q_19_begin_0, end = q_19_end_0, end_mask = q_19_end_mask_0, squeeze_mask = q_19_squeeze_mask_0, x = qkv_7_cast_fp16)[name = tensor("q_19_cast_fp16")]; + tensor k_13_begin_0 = const()[name = tensor("k_13_begin_0"), val = tensor([0, 0, 1, 0, 0])]; + tensor k_13_end_0 = const()[name = tensor("k_13_end_0"), val = tensor([1, 256, 2, 16, 64])]; + tensor k_13_end_mask_0 = const()[name = tensor("k_13_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor k_13_squeeze_mask_0 = const()[name = tensor("k_13_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor k_13_cast_fp16 = slice_by_index(begin = k_13_begin_0, end = k_13_end_0, end_mask = k_13_end_mask_0, squeeze_mask = k_13_squeeze_mask_0, x = qkv_7_cast_fp16)[name = tensor("k_13_cast_fp16")]; + tensor v_7_begin_0 = const()[name = tensor("v_7_begin_0"), val = tensor([0, 0, 2, 0, 0])]; + tensor v_7_end_0 = const()[name = tensor("v_7_end_0"), val = tensor([1, 256, 3, 16, 64])]; + tensor v_7_end_mask_0 = const()[name = tensor("v_7_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor v_7_squeeze_mask_0 = const()[name = tensor("v_7_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor v_7_cast_fp16 = slice_by_index(begin = v_7_begin_0, end = v_7_end_0, end_mask = v_7_end_mask_0, squeeze_mask = v_7_squeeze_mask_0, x = qkv_7_cast_fp16)[name = tensor("v_7_cast_fp16")]; + tensor var_1497 = const()[name = tensor("op_1497"), val = tensor([1, 1])]; + tensor position3_to_fp16_dtype_0 = const()[name = tensor("position3_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor position3_to_fp16 = cast(dtype = position3_to_fp16_dtype_0, x = position3)[name = tensor("cast_646")]; + tensor var_1498_cast_fp16 = reshape(shape = var_1497, x = position3_to_fp16)[name = tensor("op_1498_cast_fp16")]; + tensor ts_21_cast_fp16 = add(x = ts_1_promoted_to_fp16, y = var_1498_cast_fp16)[name = tensor("ts_21_cast_fp16")]; + tensor var_1505 = const()[name = tensor("op_1505"), val = tensor([1, 256, 1, 1])]; + tensor ts_23_cast_fp16 = reshape(shape = var_1505, x = ts_21_cast_fp16)[name = tensor("ts_23_cast_fp16")]; + tensor var_1512 = const()[name = tensor("op_1512"), val = tensor([1, 256, 16, 32, 2])]; + tensor q_complex_7_cast_fp16 = reshape(shape = var_1512, x = q_19_cast_fp16)[name = tensor("q_complex_7_cast_fp16")]; + tensor var_1519 = const()[name = tensor("op_1519"), val = tensor([1, 256, 16, 32, 2])]; + tensor k_complex_7_cast_fp16 = reshape(shape = var_1519, x = k_13_cast_fp16)[name = tensor("k_complex_7_cast_fp16")]; + tensor var_1523_begin_0 = const()[name = tensor("op_1523_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_1523_end_0 = const()[name = tensor("op_1523_end_0"), val = tensor([1, 256, 16, 32, 1])]; + tensor var_1523_end_mask_0 = const()[name = tensor("op_1523_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_1523_squeeze_mask_0 = const()[name = tensor("op_1523_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_1523_cast_fp16 = slice_by_index(begin = var_1523_begin_0, end = var_1523_end_0, end_mask = var_1523_end_mask_0, squeeze_mask = var_1523_squeeze_mask_0, x = q_complex_7_cast_fp16)[name = tensor("op_1523_cast_fp16")]; + tensor var_1531_begin_0 = const()[name = tensor("op_1531_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_1531_end_0 = const()[name = tensor("op_1531_end_0"), val = tensor([1, 256, 16, 32, 2])]; + tensor var_1531_end_mask_0 = const()[name = tensor("op_1531_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_1531_squeeze_mask_0 = const()[name = tensor("op_1531_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_1531_cast_fp16 = slice_by_index(begin = var_1531_begin_0, end = var_1531_end_0, end_mask = var_1531_end_mask_0, squeeze_mask = var_1531_squeeze_mask_0, x = q_complex_7_cast_fp16)[name = tensor("op_1531_cast_fp16")]; + tensor var_1539_begin_0 = const()[name = tensor("op_1539_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_1539_end_0 = const()[name = tensor("op_1539_end_0"), val = tensor([1, 256, 16, 32, 1])]; + tensor var_1539_end_mask_0 = const()[name = tensor("op_1539_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_1539_squeeze_mask_0 = const()[name = tensor("op_1539_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_1539_cast_fp16 = slice_by_index(begin = var_1539_begin_0, end = var_1539_end_0, end_mask = var_1539_end_mask_0, squeeze_mask = var_1539_squeeze_mask_0, x = k_complex_7_cast_fp16)[name = tensor("op_1539_cast_fp16")]; + tensor var_1547_begin_0 = const()[name = tensor("op_1547_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_1547_end_0 = const()[name = tensor("op_1547_end_0"), val = tensor([1, 256, 16, 32, 2])]; + tensor var_1547_end_mask_0 = const()[name = tensor("op_1547_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_1547_squeeze_mask_0 = const()[name = tensor("op_1547_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_1547_cast_fp16 = slice_by_index(begin = var_1547_begin_0, end = var_1547_end_0, end_mask = var_1547_end_mask_0, squeeze_mask = var_1547_squeeze_mask_0, x = k_complex_7_cast_fp16)[name = tensor("op_1547_cast_fp16")]; + tensor freqs_7_to_fp16 = const()[name = tensor("freqs_7_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(82887808)))]; + tensor var_1553_cast_fp16 = mul(x = freqs_7_to_fp16, y = ts_23_cast_fp16)[name = tensor("op_1553_cast_fp16")]; + tensor rotr_7_cast_fp16 = cos(x = var_1553_cast_fp16)[name = tensor("rotr_7_cast_fp16")]; + tensor roti_7_cast_fp16 = sin(x = var_1553_cast_fp16)[name = tensor("roti_7_cast_fp16")]; + tensor var_1557_cast_fp16 = mul(x = var_1523_cast_fp16, y = rotr_7_cast_fp16)[name = tensor("op_1557_cast_fp16")]; + tensor var_1558_cast_fp16 = mul(x = var_1531_cast_fp16, y = roti_7_cast_fp16)[name = tensor("op_1558_cast_fp16")]; + tensor qor_13_cast_fp16 = sub(x = var_1557_cast_fp16, y = var_1558_cast_fp16)[name = tensor("qor_13_cast_fp16")]; + tensor var_1561_cast_fp16 = mul(x = var_1523_cast_fp16, y = roti_7_cast_fp16)[name = tensor("op_1561_cast_fp16")]; + tensor var_1562_cast_fp16 = mul(x = var_1531_cast_fp16, y = rotr_7_cast_fp16)[name = tensor("op_1562_cast_fp16")]; + tensor qoi_13_cast_fp16 = add(x = var_1561_cast_fp16, y = var_1562_cast_fp16)[name = tensor("qoi_13_cast_fp16")]; + tensor var_1565_cast_fp16 = mul(x = var_1539_cast_fp16, y = rotr_7_cast_fp16)[name = tensor("op_1565_cast_fp16")]; + tensor var_1566_cast_fp16 = mul(x = var_1547_cast_fp16, y = roti_7_cast_fp16)[name = tensor("op_1566_cast_fp16")]; + tensor kor_13_cast_fp16 = sub(x = var_1565_cast_fp16, y = var_1566_cast_fp16)[name = tensor("kor_13_cast_fp16")]; + tensor var_1569_cast_fp16 = mul(x = var_1539_cast_fp16, y = roti_7_cast_fp16)[name = tensor("op_1569_cast_fp16")]; + tensor var_1570_cast_fp16 = mul(x = var_1547_cast_fp16, y = rotr_7_cast_fp16)[name = tensor("op_1570_cast_fp16")]; + tensor koi_13_cast_fp16 = add(x = var_1569_cast_fp16, y = var_1570_cast_fp16)[name = tensor("koi_13_cast_fp16")]; + tensor qo_7_axis_0 = const()[name = tensor("qo_7_axis_0"), val = tensor(-1)]; + tensor qo_7_cast_fp16 = stack(axis = qo_7_axis_0, values = (qor_13_cast_fp16, qoi_13_cast_fp16))[name = tensor("qo_7_cast_fp16")]; + tensor ko_7_axis_0 = const()[name = tensor("ko_7_axis_0"), val = tensor(-1)]; + tensor ko_7_cast_fp16 = stack(axis = ko_7_axis_0, values = (kor_13_cast_fp16, koi_13_cast_fp16))[name = tensor("ko_7_cast_fp16")]; + tensor var_1603 = const()[name = tensor("op_1603"), val = tensor([1, 256, 16, 64])]; + tensor q_21_cast_fp16 = reshape(shape = var_1603, x = qo_7_cast_fp16)[name = tensor("q_21_cast_fp16")]; + tensor var_1609 = const()[name = tensor("op_1609"), val = tensor([1, 256, 16, 64])]; + tensor k_15_cast_fp16 = reshape(shape = var_1609, x = ko_7_cast_fp16)[name = tensor("k_15_cast_fp16")]; + tensor write_indices_float_19_cast_fp16 = add(x = var_1498_cast_fp16, y = var_393_promoted_to_fp16)[name = tensor("write_indices_float_19_cast_fp16")]; + tensor _inversed_1630_y_0_to_fp16 = const()[name = tensor("_inversed_1630_y_0_to_fp16"), val = tensor(0x1p-9)]; + tensor _inversed_1630_cast_fp16 = mul(x = write_indices_float_19_cast_fp16, y = _inversed_1630_y_0_to_fp16)[name = tensor("_inversed_1630_cast_fp16")]; + tensor var_1631_cast_fp16 = floor(x = _inversed_1630_cast_fp16)[name = tensor("op_1631_cast_fp16")]; + tensor var_1632_to_fp16 = const()[name = tensor("op_1632_to_fp16"), val = tensor(0x1p+9)]; + tensor var_1633_cast_fp16 = mul(x = var_1631_cast_fp16, y = var_1632_to_fp16)[name = tensor("op_1633_cast_fp16")]; + tensor write_indices_float_21_cast_fp16 = sub(x = write_indices_float_19_cast_fp16, y = var_1633_cast_fp16)[name = tensor("write_indices_float_21_cast_fp16")]; + tensor var_1651_value_0_to_fp16 = const()[name = tensor("op_1651_value_0_to_fp16"), val = tensor(0x1.ffp+8)]; + tensor var_1651_cast_fp16 = fill_like(ref_tensor = write_indices_float_21_cast_fp16, value = var_1651_value_0_to_fp16)[name = tensor("op_1651_cast_fp16")]; + tensor write_indices_float_23_cast_fp16 = select(a = write_indices_float_21_cast_fp16, b = var_1651_cast_fp16, cond = token_valid_1_cast_fp16)[name = tensor("write_indices_float_23_cast_fp16")]; + tensor var_1657_dtype_0 = const()[name = tensor("op_1657_dtype_0"), val = tensor("int32")]; + tensor var_1662 = const()[name = tensor("op_1662"), val = tensor([1, 256, 1, 1])]; + tensor write_indices_float_23_cast_fp16_to_int32 = cast(dtype = var_1657_dtype_0, x = write_indices_float_23_cast_fp16)[name = tensor("cast_645")]; + tensor var_1663 = reshape(shape = var_1662, x = write_indices_float_23_cast_fp16_to_int32)[name = tensor("op_1663")]; + tensor write_indices_7_reps_0 = const()[name = tensor("write_indices_7_reps_0"), val = tensor([1, 1, 16, 64])]; + tensor write_indices_7 = tile(reps = write_indices_7_reps_0, x = var_1663)[name = tensor("write_indices_7")]; + tensor var_1673_begin_0 = const()[name = tensor("op_1673_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_1673_end_0 = const()[name = tensor("op_1673_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor var_1673_end_mask_0 = const()[name = tensor("op_1673_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_1673_squeeze_mask_0 = const()[name = tensor("op_1673_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor cache3_to_fp16_dtype_0 = const()[name = tensor("cache3_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor cache3_to_fp16 = cast(dtype = cache3_to_fp16_dtype_0, x = cache3)[name = tensor("cast_644")]; + tensor var_1673_cast_fp16 = slice_by_index(begin = var_1673_begin_0, end = var_1673_end_0, end_mask = var_1673_end_mask_0, squeeze_mask = var_1673_squeeze_mask_0, x = cache3_to_fp16)[name = tensor("op_1673_cast_fp16")]; + tensor var_1675_axis_0 = const()[name = tensor("op_1675_axis_0"), val = tensor(1)]; + tensor var_1675_mode_0 = const()[name = tensor("op_1675_mode_0"), val = tensor("update")]; + tensor var_1675_validate_indices_0 = const()[name = tensor("op_1675_validate_indices_0"), val = tensor(false)]; + tensor var_1675_cast_fp16 = scatter_along_axis(axis = var_1675_axis_0, data = var_1673_cast_fp16, indices = write_indices_7, mode = var_1675_mode_0, updates = k_15_cast_fp16, validate_indices = var_1675_validate_indices_0)[name = tensor("op_1675_cast_fp16")]; + tensor concat_26 = const()[name = tensor("concat_26"), val = tensor([0, 0, 0, 0, 0])]; + tensor concat_27 = const()[name = tensor("concat_27"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_7_internal_tensor_assign_1_stride_0 = const()[name = tensor("new_cache_7_internal_tensor_assign_1_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_7_internal_tensor_assign_1_begin_mask_0 = const()[name = tensor("new_cache_7_internal_tensor_assign_1_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_7_internal_tensor_assign_1_end_mask_0 = const()[name = tensor("new_cache_7_internal_tensor_assign_1_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_7_internal_tensor_assign_1_squeeze_mask_0 = const()[name = tensor("new_cache_7_internal_tensor_assign_1_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_52 = const()[name = tensor("shape_52"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_6 = const()[name = tensor("reduce_prod_6"), val = tensor(1048576)]; + tensor range_1d_6_start_0 = const()[name = tensor("range_1d_6_start_0"), val = tensor(0)]; + tensor range_1d_6_step_0 = const()[name = tensor("range_1d_6_step_0"), val = tensor(1)]; + tensor range_1d_6 = range_1d(end = reduce_prod_6, start = range_1d_6_start_0, step = range_1d_6_step_0)[name = tensor("range_1d_6")]; + tensor reshape_30 = reshape(shape = shape_52, x = range_1d_6)[name = tensor("reshape_30")]; + tensor slice_by_index_6 = slice_by_index(begin = concat_26, begin_mask = new_cache_7_internal_tensor_assign_1_begin_mask_0, end = concat_27, end_mask = new_cache_7_internal_tensor_assign_1_end_mask_0, squeeze_mask = new_cache_7_internal_tensor_assign_1_squeeze_mask_0, stride = new_cache_7_internal_tensor_assign_1_stride_0, x = reshape_30)[name = tensor("slice_by_index_6")]; + tensor reshape_31_shape_0 = const()[name = tensor("reshape_31_shape_0"), val = tensor([-1])]; + tensor reshape_31 = reshape(shape = reshape_31_shape_0, x = slice_by_index_6)[name = tensor("reshape_31")]; + tensor reshape_32_shape_0 = const()[name = tensor("reshape_32_shape_0"), val = tensor([-1])]; + tensor reshape_32_cast_fp16 = reshape(shape = reshape_32_shape_0, x = var_1675_cast_fp16)[name = tensor("reshape_32_cast_fp16")]; + tensor reshape_33_shape_0 = const()[name = tensor("reshape_33_shape_0"), val = tensor([-1])]; + tensor reshape_33_cast_fp16 = reshape(shape = reshape_33_shape_0, x = cache3_to_fp16)[name = tensor("reshape_33_cast_fp16")]; + tensor scatter_6_mode_0 = const()[name = tensor("scatter_6_mode_0"), val = tensor("update")]; + tensor scatter_6_axis_0 = const()[name = tensor("scatter_6_axis_0"), val = tensor(0)]; + tensor scatter_6_validate_indices_0 = const()[name = tensor("scatter_6_validate_indices_0"), val = tensor(false)]; + tensor scatter_6_cast_fp16 = scatter(axis = scatter_6_axis_0, data = reshape_33_cast_fp16, indices = reshape_31, mode = scatter_6_mode_0, updates = reshape_32_cast_fp16, validate_indices = scatter_6_validate_indices_0)[name = tensor("scatter_6_cast_fp16")]; + tensor reshape_34_cast_fp16 = reshape(shape = shape_52, x = scatter_6_cast_fp16)[name = tensor("reshape_34_cast_fp16")]; + tensor var_1683_begin_0 = const()[name = tensor("op_1683_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor var_1683_end_0 = const()[name = tensor("op_1683_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor var_1683_end_mask_0 = const()[name = tensor("op_1683_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_1683_squeeze_mask_0 = const()[name = tensor("op_1683_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor var_1683_cast_fp16 = slice_by_index(begin = var_1683_begin_0, end = var_1683_end_0, end_mask = var_1683_end_mask_0, squeeze_mask = var_1683_squeeze_mask_0, x = reshape_34_cast_fp16)[name = tensor("op_1683_cast_fp16")]; + tensor var_1685_axis_0 = const()[name = tensor("op_1685_axis_0"), val = tensor(1)]; + tensor var_1685_mode_0 = const()[name = tensor("op_1685_mode_0"), val = tensor("update")]; + tensor var_1685_validate_indices_0 = const()[name = tensor("op_1685_validate_indices_0"), val = tensor(false)]; + tensor var_1685_cast_fp16 = scatter_along_axis(axis = var_1685_axis_0, data = var_1683_cast_fp16, indices = write_indices_7, mode = var_1685_mode_0, updates = v_7_cast_fp16, validate_indices = var_1685_validate_indices_0)[name = tensor("op_1685_cast_fp16")]; + tensor concat_28 = const()[name = tensor("concat_28"), val = tensor([1, 0, 0, 0, 0])]; + tensor concat_29 = const()[name = tensor("concat_29"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_7_internal_tensor_assign_2_stride_0 = const()[name = tensor("new_cache_7_internal_tensor_assign_2_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_7_internal_tensor_assign_2_begin_mask_0 = const()[name = tensor("new_cache_7_internal_tensor_assign_2_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_7_internal_tensor_assign_2_end_mask_0 = const()[name = tensor("new_cache_7_internal_tensor_assign_2_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_7_internal_tensor_assign_2_squeeze_mask_0 = const()[name = tensor("new_cache_7_internal_tensor_assign_2_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_53 = const()[name = tensor("shape_53"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_7 = const()[name = tensor("reduce_prod_7"), val = tensor(1048576)]; + tensor range_1d_7_start_0 = const()[name = tensor("range_1d_7_start_0"), val = tensor(0)]; + tensor range_1d_7_step_0 = const()[name = tensor("range_1d_7_step_0"), val = tensor(1)]; + tensor range_1d_7 = range_1d(end = reduce_prod_7, start = range_1d_7_start_0, step = range_1d_7_step_0)[name = tensor("range_1d_7")]; + tensor reshape_35 = reshape(shape = shape_53, x = range_1d_7)[name = tensor("reshape_35")]; + tensor slice_by_index_7 = slice_by_index(begin = concat_28, begin_mask = new_cache_7_internal_tensor_assign_2_begin_mask_0, end = concat_29, end_mask = new_cache_7_internal_tensor_assign_2_end_mask_0, squeeze_mask = new_cache_7_internal_tensor_assign_2_squeeze_mask_0, stride = new_cache_7_internal_tensor_assign_2_stride_0, x = reshape_35)[name = tensor("slice_by_index_7")]; + tensor reshape_36_shape_0 = const()[name = tensor("reshape_36_shape_0"), val = tensor([-1])]; + tensor reshape_36 = reshape(shape = reshape_36_shape_0, x = slice_by_index_7)[name = tensor("reshape_36")]; + tensor reshape_37_shape_0 = const()[name = tensor("reshape_37_shape_0"), val = tensor([-1])]; + tensor reshape_37_cast_fp16 = reshape(shape = reshape_37_shape_0, x = var_1685_cast_fp16)[name = tensor("reshape_37_cast_fp16")]; + tensor reshape_38_shape_0 = const()[name = tensor("reshape_38_shape_0"), val = tensor([-1])]; + tensor reshape_38_cast_fp16 = reshape(shape = reshape_38_shape_0, x = reshape_34_cast_fp16)[name = tensor("reshape_38_cast_fp16")]; + tensor scatter_7_mode_0 = const()[name = tensor("scatter_7_mode_0"), val = tensor("update")]; + tensor scatter_7_axis_0 = const()[name = tensor("scatter_7_axis_0"), val = tensor(0)]; + tensor scatter_7_validate_indices_0 = const()[name = tensor("scatter_7_validate_indices_0"), val = tensor(false)]; + tensor scatter_7_cast_fp16 = scatter(axis = scatter_7_axis_0, data = reshape_38_cast_fp16, indices = reshape_36, mode = scatter_7_mode_0, updates = reshape_37_cast_fp16, validate_indices = scatter_7_validate_indices_0)[name = tensor("scatter_7_cast_fp16")]; + tensor reshape_39_cast_fp16 = reshape(shape = shape_53, x = scatter_7_cast_fp16)[name = tensor("reshape_39_cast_fp16")]; + tensor reshape_39_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("reshape_39_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor keys_19_begin_0 = const()[name = tensor("keys_19_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor keys_19_end_0 = const()[name = tensor("keys_19_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor keys_19_end_mask_0 = const()[name = tensor("keys_19_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor keys_19_squeeze_mask_0 = const()[name = tensor("keys_19_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor keys_19_cast_fp16 = slice_by_index(begin = keys_19_begin_0, end = keys_19_end_0, end_mask = keys_19_end_mask_0, squeeze_mask = keys_19_squeeze_mask_0, x = reshape_39_cast_fp16)[name = tensor("keys_19_cast_fp16")]; + tensor values_19_begin_0 = const()[name = tensor("values_19_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor values_19_end_0 = const()[name = tensor("values_19_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor values_19_end_mask_0 = const()[name = tensor("values_19_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor values_19_squeeze_mask_0 = const()[name = tensor("values_19_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor values_19_cast_fp16 = slice_by_index(begin = values_19_begin_0, end = values_19_end_0, end_mask = values_19_end_mask_0, squeeze_mask = values_19_squeeze_mask_0, x = reshape_39_cast_fp16)[name = tensor("values_19_cast_fp16")]; + tensor var_1697_cast_fp16 = not_equal(x = keys_19_cast_fp16, y = keys_19_cast_fp16)[name = tensor("op_1697_cast_fp16")]; + tensor keys_21_cast_fp16 = select(a = var_470_to_fp16, b = keys_19_cast_fp16, cond = var_1697_cast_fp16)[name = tensor("keys_21_cast_fp16")]; + tensor var_1705_cast_fp16 = not_equal(x = values_19_cast_fp16, y = values_19_cast_fp16)[name = tensor("op_1705_cast_fp16")]; + tensor values_21_cast_fp16 = select(a = var_470_to_fp16, b = values_19_cast_fp16, cond = var_1705_cast_fp16)[name = tensor("values_21_cast_fp16")]; + tensor var_1729 = const()[name = tensor("op_1729"), val = tensor([0, 2, 1, 3])]; + tensor var_1745 = const()[name = tensor("op_1745"), val = tensor([1, 1, 1])]; + tensor var_1746_cast_fp16 = reshape(shape = var_1745, x = position3_to_fp16)[name = tensor("op_1746_cast_fp16")]; + tensor q_positions_7_cast_fp16 = add(x = var_1746_cast_fp16, y = q_offsets_1_promoted_to_fp16)[name = tensor("q_positions_7_cast_fp16")]; + tensor valid_end_7_cast_fp16 = add(x = var_1746_cast_fp16, y = var_536_cast_fp16)[name = tensor("valid_end_7_cast_fp16")]; + tensor valid_mask_7_cast_fp16 = less(x = k_positions_1_promoted_to_fp16, y = valid_end_7_cast_fp16)[name = tensor("valid_mask_7_cast_fp16")]; + tensor causal_mask_7_cast_fp16 = less_equal(x = k_positions_1_promoted_to_fp16, y = q_positions_7_cast_fp16)[name = tensor("causal_mask_7_cast_fp16")]; + tensor var_1774 = logical_and(x = valid_mask_7_cast_fp16, y = causal_mask_7_cast_fp16)[name = tensor("op_1774")]; + tensor attn_mask_7_axes_0 = const()[name = tensor("attn_mask_7_axes_0"), val = tensor([1])]; + tensor attn_mask_7 = expand_dims(axes = attn_mask_7_axes_0, x = var_1774)[name = tensor("attn_mask_7")]; + tensor var_1788_transpose_x_0 = const()[name = tensor("op_1788_transpose_x_0"), val = tensor(false)]; + tensor var_1788_transpose_y_0 = const()[name = tensor("op_1788_transpose_y_0"), val = tensor(false)]; + tensor transpose_75_perm_0 = const()[name = tensor("transpose_75_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_76_perm_0 = const()[name = tensor("transpose_76_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_76 = transpose(perm = transpose_76_perm_0, x = keys_21_cast_fp16)[name = tensor("transpose_192")]; + tensor transpose_75 = transpose(perm = transpose_75_perm_0, x = q_21_cast_fp16)[name = tensor("transpose_193")]; + tensor var_1788_cast_fp16 = matmul(transpose_x = var_1788_transpose_x_0, transpose_y = var_1788_transpose_y_0, x = transpose_75, y = transpose_76)[name = tensor("op_1788_cast_fp16")]; + tensor var_1782_to_fp16 = const()[name = tensor("op_1782_to_fp16"), val = tensor([0x1p-3])]; + tensor attn_weights_19_cast_fp16 = mul(x = var_1788_cast_fp16, y = var_1782_to_fp16)[name = tensor("attn_weights_19_cast_fp16")]; + tensor var_1790 = logical_not(x = attn_mask_7)[name = tensor("op_1790")]; + tensor var_1791_to_fp16 = const()[name = tensor("op_1791_to_fp16"), val = tensor(-inf)]; + tensor attn_weights_21_cast_fp16 = select(a = var_1791_to_fp16, b = attn_weights_19_cast_fp16, cond = var_1790)[name = tensor("attn_weights_21_cast_fp16")]; + tensor var_1793 = const()[name = tensor("op_1793"), val = tensor(-1)]; + tensor attn_weights_23_cast_fp16 = softmax(axis = var_1793, x = attn_weights_21_cast_fp16)[name = tensor("attn_weights_23_cast_fp16")]; + tensor attn_output_7_transpose_x_0 = const()[name = tensor("attn_output_7_transpose_x_0"), val = tensor(false)]; + tensor attn_output_7_transpose_y_0 = const()[name = tensor("attn_output_7_transpose_y_0"), val = tensor(false)]; + tensor values_23_cast_fp16 = transpose(perm = var_1729, x = values_21_cast_fp16)[name = tensor("transpose_194")]; + tensor attn_output_7_cast_fp16 = matmul(transpose_x = attn_output_7_transpose_x_0, transpose_y = attn_output_7_transpose_y_0, x = attn_weights_23_cast_fp16, y = values_23_cast_fp16)[name = tensor("attn_output_7_cast_fp16")]; + tensor var_1801 = const()[name = tensor("op_1801"), val = tensor([0, 2, 1, 3])]; + tensor var_1806 = const()[name = tensor("op_1806"), val = tensor([1, 256, 1024])]; + tensor var_1802_cast_fp16 = transpose(perm = var_1801, x = attn_output_7_cast_fp16)[name = tensor("transpose_191")]; + tensor input_41_cast_fp16 = reshape(shape = var_1806, x = var_1802_cast_fp16)[name = tensor("input_41_cast_fp16")]; + tensor attn3_out_proj_weight_to_fp16 = const()[name = tensor("attn3_out_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(82887936)))]; + tensor linear_13_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = attn3_out_proj_weight_to_fp16, x = input_41_cast_fp16)[name = tensor("linear_13_cast_fp16")]; + tensor var_1815_cast_fp16 = add(x = position3_to_fp16, y = valid_len_to_fp16)[name = tensor("op_1815_cast_fp16")]; + tensor var_1815_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_1815_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor input_43_cast_fp16 = add(x = input_37_cast_fp16, y = linear_13_cast_fp16)[name = tensor("input_43_cast_fp16")]; + tensor input_45_axes_0 = const()[name = tensor("input_45_axes_0"), val = tensor([-1])]; + tensor norm3_2_weight_to_fp16 = const()[name = tensor("norm3_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(84985152)))]; + tensor norm3_2_bias_to_fp16 = const()[name = tensor("norm3_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(84987264)))]; + tensor var_1819_to_fp16 = const()[name = tensor("op_1819_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_45_cast_fp16 = layer_norm(axes = input_45_axes_0, beta = norm3_2_bias_to_fp16, epsilon = var_1819_to_fp16, gamma = norm3_2_weight_to_fp16, x = input_43_cast_fp16)[name = tensor("input_45_cast_fp16")]; + tensor linear3_1_weight_to_fp16 = const()[name = tensor("linear3_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(84989376)))]; + tensor linear_14_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = linear3_1_weight_to_fp16, x = input_45_cast_fp16)[name = tensor("linear_14_cast_fp16")]; + tensor input_47_mode_0 = const()[name = tensor("input_47_mode_0"), val = tensor("EXACT")]; + tensor input_47_cast_fp16 = gelu(mode = input_47_mode_0, x = linear_14_cast_fp16)[name = tensor("input_47_cast_fp16")]; + tensor linear3_2_weight_to_fp16 = const()[name = tensor("linear3_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(93378048)))]; + tensor linear_15_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = linear3_2_weight_to_fp16, x = input_47_cast_fp16)[name = tensor("linear_15_cast_fp16")]; + tensor input_49_cast_fp16 = add(x = input_43_cast_fp16, y = linear_15_cast_fp16)[name = tensor("input_49_cast_fp16")]; + tensor input_51_axes_0 = const()[name = tensor("input_51_axes_0"), val = tensor([-1])]; + tensor norm4_1_weight_to_fp16 = const()[name = tensor("norm4_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(101766720)))]; + tensor norm4_1_bias_to_fp16 = const()[name = tensor("norm4_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(101768832)))]; + tensor var_1836_to_fp16 = const()[name = tensor("op_1836_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_51_cast_fp16 = layer_norm(axes = input_51_axes_0, beta = norm4_1_bias_to_fp16, epsilon = var_1836_to_fp16, gamma = norm4_1_weight_to_fp16, x = input_49_cast_fp16)[name = tensor("input_51_cast_fp16")]; + tensor attn4_in_proj_weight_to_fp16 = const()[name = tensor("attn4_in_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(101770944)))]; + tensor linear_16_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = attn4_in_proj_weight_to_fp16, x = input_51_cast_fp16)[name = tensor("linear_16_cast_fp16")]; + tensor var_1850 = const()[name = tensor("op_1850"), val = tensor([1, 256, 3, 16, 64])]; + tensor qkv_9_cast_fp16 = reshape(shape = var_1850, x = linear_16_cast_fp16)[name = tensor("qkv_9_cast_fp16")]; + tensor q_25_begin_0 = const()[name = tensor("q_25_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor q_25_end_0 = const()[name = tensor("q_25_end_0"), val = tensor([1, 256, 1, 16, 64])]; + tensor q_25_end_mask_0 = const()[name = tensor("q_25_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor q_25_squeeze_mask_0 = const()[name = tensor("q_25_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor q_25_cast_fp16 = slice_by_index(begin = q_25_begin_0, end = q_25_end_0, end_mask = q_25_end_mask_0, squeeze_mask = q_25_squeeze_mask_0, x = qkv_9_cast_fp16)[name = tensor("q_25_cast_fp16")]; + tensor k_17_begin_0 = const()[name = tensor("k_17_begin_0"), val = tensor([0, 0, 1, 0, 0])]; + tensor k_17_end_0 = const()[name = tensor("k_17_end_0"), val = tensor([1, 256, 2, 16, 64])]; + tensor k_17_end_mask_0 = const()[name = tensor("k_17_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor k_17_squeeze_mask_0 = const()[name = tensor("k_17_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor k_17_cast_fp16 = slice_by_index(begin = k_17_begin_0, end = k_17_end_0, end_mask = k_17_end_mask_0, squeeze_mask = k_17_squeeze_mask_0, x = qkv_9_cast_fp16)[name = tensor("k_17_cast_fp16")]; + tensor v_9_begin_0 = const()[name = tensor("v_9_begin_0"), val = tensor([0, 0, 2, 0, 0])]; + tensor v_9_end_0 = const()[name = tensor("v_9_end_0"), val = tensor([1, 256, 3, 16, 64])]; + tensor v_9_end_mask_0 = const()[name = tensor("v_9_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor v_9_squeeze_mask_0 = const()[name = tensor("v_9_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor v_9_cast_fp16 = slice_by_index(begin = v_9_begin_0, end = v_9_end_0, end_mask = v_9_end_mask_0, squeeze_mask = v_9_squeeze_mask_0, x = qkv_9_cast_fp16)[name = tensor("v_9_cast_fp16")]; + tensor var_1908 = const()[name = tensor("op_1908"), val = tensor([1, 1])]; + tensor position4_to_fp16_dtype_0 = const()[name = tensor("position4_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor position4_to_fp16 = cast(dtype = position4_to_fp16_dtype_0, x = position4)[name = tensor("cast_641")]; + tensor var_1909_cast_fp16 = reshape(shape = var_1908, x = position4_to_fp16)[name = tensor("op_1909_cast_fp16")]; + tensor ts_27_cast_fp16 = add(x = ts_1_promoted_to_fp16, y = var_1909_cast_fp16)[name = tensor("ts_27_cast_fp16")]; + tensor var_1916 = const()[name = tensor("op_1916"), val = tensor([1, 256, 1, 1])]; + tensor ts_29_cast_fp16 = reshape(shape = var_1916, x = ts_27_cast_fp16)[name = tensor("ts_29_cast_fp16")]; + tensor var_1923 = const()[name = tensor("op_1923"), val = tensor([1, 256, 16, 32, 2])]; + tensor q_complex_9_cast_fp16 = reshape(shape = var_1923, x = q_25_cast_fp16)[name = tensor("q_complex_9_cast_fp16")]; + tensor var_1930 = const()[name = tensor("op_1930"), val = tensor([1, 256, 16, 32, 2])]; + tensor k_complex_9_cast_fp16 = reshape(shape = var_1930, x = k_17_cast_fp16)[name = tensor("k_complex_9_cast_fp16")]; + tensor var_1934_begin_0 = const()[name = tensor("op_1934_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_1934_end_0 = const()[name = tensor("op_1934_end_0"), val = tensor([1, 256, 16, 32, 1])]; + tensor var_1934_end_mask_0 = const()[name = tensor("op_1934_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_1934_squeeze_mask_0 = const()[name = tensor("op_1934_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_1934_cast_fp16 = slice_by_index(begin = var_1934_begin_0, end = var_1934_end_0, end_mask = var_1934_end_mask_0, squeeze_mask = var_1934_squeeze_mask_0, x = q_complex_9_cast_fp16)[name = tensor("op_1934_cast_fp16")]; + tensor var_1942_begin_0 = const()[name = tensor("op_1942_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_1942_end_0 = const()[name = tensor("op_1942_end_0"), val = tensor([1, 256, 16, 32, 2])]; + tensor var_1942_end_mask_0 = const()[name = tensor("op_1942_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_1942_squeeze_mask_0 = const()[name = tensor("op_1942_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_1942_cast_fp16 = slice_by_index(begin = var_1942_begin_0, end = var_1942_end_0, end_mask = var_1942_end_mask_0, squeeze_mask = var_1942_squeeze_mask_0, x = q_complex_9_cast_fp16)[name = tensor("op_1942_cast_fp16")]; + tensor var_1950_begin_0 = const()[name = tensor("op_1950_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_1950_end_0 = const()[name = tensor("op_1950_end_0"), val = tensor([1, 256, 16, 32, 1])]; + tensor var_1950_end_mask_0 = const()[name = tensor("op_1950_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_1950_squeeze_mask_0 = const()[name = tensor("op_1950_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_1950_cast_fp16 = slice_by_index(begin = var_1950_begin_0, end = var_1950_end_0, end_mask = var_1950_end_mask_0, squeeze_mask = var_1950_squeeze_mask_0, x = k_complex_9_cast_fp16)[name = tensor("op_1950_cast_fp16")]; + tensor var_1958_begin_0 = const()[name = tensor("op_1958_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_1958_end_0 = const()[name = tensor("op_1958_end_0"), val = tensor([1, 256, 16, 32, 2])]; + tensor var_1958_end_mask_0 = const()[name = tensor("op_1958_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_1958_squeeze_mask_0 = const()[name = tensor("op_1958_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_1958_cast_fp16 = slice_by_index(begin = var_1958_begin_0, end = var_1958_end_0, end_mask = var_1958_end_mask_0, squeeze_mask = var_1958_squeeze_mask_0, x = k_complex_9_cast_fp16)[name = tensor("op_1958_cast_fp16")]; + tensor freqs_9_to_fp16 = const()[name = tensor("freqs_9_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(108062464)))]; + tensor var_1964_cast_fp16 = mul(x = freqs_9_to_fp16, y = ts_29_cast_fp16)[name = tensor("op_1964_cast_fp16")]; + tensor rotr_9_cast_fp16 = cos(x = var_1964_cast_fp16)[name = tensor("rotr_9_cast_fp16")]; + tensor roti_9_cast_fp16 = sin(x = var_1964_cast_fp16)[name = tensor("roti_9_cast_fp16")]; + tensor var_1968_cast_fp16 = mul(x = var_1934_cast_fp16, y = rotr_9_cast_fp16)[name = tensor("op_1968_cast_fp16")]; + tensor var_1969_cast_fp16 = mul(x = var_1942_cast_fp16, y = roti_9_cast_fp16)[name = tensor("op_1969_cast_fp16")]; + tensor qor_17_cast_fp16 = sub(x = var_1968_cast_fp16, y = var_1969_cast_fp16)[name = tensor("qor_17_cast_fp16")]; + tensor var_1972_cast_fp16 = mul(x = var_1934_cast_fp16, y = roti_9_cast_fp16)[name = tensor("op_1972_cast_fp16")]; + tensor var_1973_cast_fp16 = mul(x = var_1942_cast_fp16, y = rotr_9_cast_fp16)[name = tensor("op_1973_cast_fp16")]; + tensor qoi_17_cast_fp16 = add(x = var_1972_cast_fp16, y = var_1973_cast_fp16)[name = tensor("qoi_17_cast_fp16")]; + tensor var_1976_cast_fp16 = mul(x = var_1950_cast_fp16, y = rotr_9_cast_fp16)[name = tensor("op_1976_cast_fp16")]; + tensor var_1977_cast_fp16 = mul(x = var_1958_cast_fp16, y = roti_9_cast_fp16)[name = tensor("op_1977_cast_fp16")]; + tensor kor_17_cast_fp16 = sub(x = var_1976_cast_fp16, y = var_1977_cast_fp16)[name = tensor("kor_17_cast_fp16")]; + tensor var_1980_cast_fp16 = mul(x = var_1950_cast_fp16, y = roti_9_cast_fp16)[name = tensor("op_1980_cast_fp16")]; + tensor var_1981_cast_fp16 = mul(x = var_1958_cast_fp16, y = rotr_9_cast_fp16)[name = tensor("op_1981_cast_fp16")]; + tensor koi_17_cast_fp16 = add(x = var_1980_cast_fp16, y = var_1981_cast_fp16)[name = tensor("koi_17_cast_fp16")]; + tensor qo_9_axis_0 = const()[name = tensor("qo_9_axis_0"), val = tensor(-1)]; + tensor qo_9_cast_fp16 = stack(axis = qo_9_axis_0, values = (qor_17_cast_fp16, qoi_17_cast_fp16))[name = tensor("qo_9_cast_fp16")]; + tensor ko_9_axis_0 = const()[name = tensor("ko_9_axis_0"), val = tensor(-1)]; + tensor ko_9_cast_fp16 = stack(axis = ko_9_axis_0, values = (kor_17_cast_fp16, koi_17_cast_fp16))[name = tensor("ko_9_cast_fp16")]; + tensor var_2014 = const()[name = tensor("op_2014"), val = tensor([1, 256, 16, 64])]; + tensor q_27_cast_fp16 = reshape(shape = var_2014, x = qo_9_cast_fp16)[name = tensor("q_27_cast_fp16")]; + tensor var_2020 = const()[name = tensor("op_2020"), val = tensor([1, 256, 16, 64])]; + tensor k_19_cast_fp16 = reshape(shape = var_2020, x = ko_9_cast_fp16)[name = tensor("k_19_cast_fp16")]; + tensor write_indices_float_25_cast_fp16 = add(x = var_1909_cast_fp16, y = var_393_promoted_to_fp16)[name = tensor("write_indices_float_25_cast_fp16")]; + tensor _inversed_2041_y_0_to_fp16 = const()[name = tensor("_inversed_2041_y_0_to_fp16"), val = tensor(0x1p-9)]; + tensor _inversed_2041_cast_fp16 = mul(x = write_indices_float_25_cast_fp16, y = _inversed_2041_y_0_to_fp16)[name = tensor("_inversed_2041_cast_fp16")]; + tensor var_2042_cast_fp16 = floor(x = _inversed_2041_cast_fp16)[name = tensor("op_2042_cast_fp16")]; + tensor var_2043_to_fp16 = const()[name = tensor("op_2043_to_fp16"), val = tensor(0x1p+9)]; + tensor var_2044_cast_fp16 = mul(x = var_2042_cast_fp16, y = var_2043_to_fp16)[name = tensor("op_2044_cast_fp16")]; + tensor write_indices_float_27_cast_fp16 = sub(x = write_indices_float_25_cast_fp16, y = var_2044_cast_fp16)[name = tensor("write_indices_float_27_cast_fp16")]; + tensor var_2062_value_0_to_fp16 = const()[name = tensor("op_2062_value_0_to_fp16"), val = tensor(0x1.ffp+8)]; + tensor var_2062_cast_fp16 = fill_like(ref_tensor = write_indices_float_27_cast_fp16, value = var_2062_value_0_to_fp16)[name = tensor("op_2062_cast_fp16")]; + tensor write_indices_float_29_cast_fp16 = select(a = write_indices_float_27_cast_fp16, b = var_2062_cast_fp16, cond = token_valid_1_cast_fp16)[name = tensor("write_indices_float_29_cast_fp16")]; + tensor var_2068_dtype_0 = const()[name = tensor("op_2068_dtype_0"), val = tensor("int32")]; + tensor var_2073 = const()[name = tensor("op_2073"), val = tensor([1, 256, 1, 1])]; + tensor write_indices_float_29_cast_fp16_to_int32 = cast(dtype = var_2068_dtype_0, x = write_indices_float_29_cast_fp16)[name = tensor("cast_640")]; + tensor var_2074 = reshape(shape = var_2073, x = write_indices_float_29_cast_fp16_to_int32)[name = tensor("op_2074")]; + tensor write_indices_9_reps_0 = const()[name = tensor("write_indices_9_reps_0"), val = tensor([1, 1, 16, 64])]; + tensor write_indices_9 = tile(reps = write_indices_9_reps_0, x = var_2074)[name = tensor("write_indices_9")]; + tensor var_2084_begin_0 = const()[name = tensor("op_2084_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_2084_end_0 = const()[name = tensor("op_2084_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor var_2084_end_mask_0 = const()[name = tensor("op_2084_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_2084_squeeze_mask_0 = const()[name = tensor("op_2084_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor cache4_to_fp16_dtype_0 = const()[name = tensor("cache4_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor cache4_to_fp16 = cast(dtype = cache4_to_fp16_dtype_0, x = cache4)[name = tensor("cast_639")]; + tensor var_2084_cast_fp16 = slice_by_index(begin = var_2084_begin_0, end = var_2084_end_0, end_mask = var_2084_end_mask_0, squeeze_mask = var_2084_squeeze_mask_0, x = cache4_to_fp16)[name = tensor("op_2084_cast_fp16")]; + tensor var_2086_axis_0 = const()[name = tensor("op_2086_axis_0"), val = tensor(1)]; + tensor var_2086_mode_0 = const()[name = tensor("op_2086_mode_0"), val = tensor("update")]; + tensor var_2086_validate_indices_0 = const()[name = tensor("op_2086_validate_indices_0"), val = tensor(false)]; + tensor var_2086_cast_fp16 = scatter_along_axis(axis = var_2086_axis_0, data = var_2084_cast_fp16, indices = write_indices_9, mode = var_2086_mode_0, updates = k_19_cast_fp16, validate_indices = var_2086_validate_indices_0)[name = tensor("op_2086_cast_fp16")]; + tensor concat_34 = const()[name = tensor("concat_34"), val = tensor([0, 0, 0, 0, 0])]; + tensor concat_35 = const()[name = tensor("concat_35"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_9_internal_tensor_assign_1_stride_0 = const()[name = tensor("new_cache_9_internal_tensor_assign_1_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_9_internal_tensor_assign_1_begin_mask_0 = const()[name = tensor("new_cache_9_internal_tensor_assign_1_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_9_internal_tensor_assign_1_end_mask_0 = const()[name = tensor("new_cache_9_internal_tensor_assign_1_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_9_internal_tensor_assign_1_squeeze_mask_0 = const()[name = tensor("new_cache_9_internal_tensor_assign_1_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_54 = const()[name = tensor("shape_54"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_8 = const()[name = tensor("reduce_prod_8"), val = tensor(1048576)]; + tensor range_1d_8_start_0 = const()[name = tensor("range_1d_8_start_0"), val = tensor(0)]; + tensor range_1d_8_step_0 = const()[name = tensor("range_1d_8_step_0"), val = tensor(1)]; + tensor range_1d_8 = range_1d(end = reduce_prod_8, start = range_1d_8_start_0, step = range_1d_8_step_0)[name = tensor("range_1d_8")]; + tensor reshape_40 = reshape(shape = shape_54, x = range_1d_8)[name = tensor("reshape_40")]; + tensor slice_by_index_8 = slice_by_index(begin = concat_34, begin_mask = new_cache_9_internal_tensor_assign_1_begin_mask_0, end = concat_35, end_mask = new_cache_9_internal_tensor_assign_1_end_mask_0, squeeze_mask = new_cache_9_internal_tensor_assign_1_squeeze_mask_0, stride = new_cache_9_internal_tensor_assign_1_stride_0, x = reshape_40)[name = tensor("slice_by_index_8")]; + tensor reshape_41_shape_0 = const()[name = tensor("reshape_41_shape_0"), val = tensor([-1])]; + tensor reshape_41 = reshape(shape = reshape_41_shape_0, x = slice_by_index_8)[name = tensor("reshape_41")]; + tensor reshape_42_shape_0 = const()[name = tensor("reshape_42_shape_0"), val = tensor([-1])]; + tensor reshape_42_cast_fp16 = reshape(shape = reshape_42_shape_0, x = var_2086_cast_fp16)[name = tensor("reshape_42_cast_fp16")]; + tensor reshape_43_shape_0 = const()[name = tensor("reshape_43_shape_0"), val = tensor([-1])]; + tensor reshape_43_cast_fp16 = reshape(shape = reshape_43_shape_0, x = cache4_to_fp16)[name = tensor("reshape_43_cast_fp16")]; + tensor scatter_8_mode_0 = const()[name = tensor("scatter_8_mode_0"), val = tensor("update")]; + tensor scatter_8_axis_0 = const()[name = tensor("scatter_8_axis_0"), val = tensor(0)]; + tensor scatter_8_validate_indices_0 = const()[name = tensor("scatter_8_validate_indices_0"), val = tensor(false)]; + tensor scatter_8_cast_fp16 = scatter(axis = scatter_8_axis_0, data = reshape_43_cast_fp16, indices = reshape_41, mode = scatter_8_mode_0, updates = reshape_42_cast_fp16, validate_indices = scatter_8_validate_indices_0)[name = tensor("scatter_8_cast_fp16")]; + tensor reshape_44_cast_fp16 = reshape(shape = shape_54, x = scatter_8_cast_fp16)[name = tensor("reshape_44_cast_fp16")]; + tensor var_2094_begin_0 = const()[name = tensor("op_2094_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor var_2094_end_0 = const()[name = tensor("op_2094_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor var_2094_end_mask_0 = const()[name = tensor("op_2094_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_2094_squeeze_mask_0 = const()[name = tensor("op_2094_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor var_2094_cast_fp16 = slice_by_index(begin = var_2094_begin_0, end = var_2094_end_0, end_mask = var_2094_end_mask_0, squeeze_mask = var_2094_squeeze_mask_0, x = reshape_44_cast_fp16)[name = tensor("op_2094_cast_fp16")]; + tensor var_2096_axis_0 = const()[name = tensor("op_2096_axis_0"), val = tensor(1)]; + tensor var_2096_mode_0 = const()[name = tensor("op_2096_mode_0"), val = tensor("update")]; + tensor var_2096_validate_indices_0 = const()[name = tensor("op_2096_validate_indices_0"), val = tensor(false)]; + tensor var_2096_cast_fp16 = scatter_along_axis(axis = var_2096_axis_0, data = var_2094_cast_fp16, indices = write_indices_9, mode = var_2096_mode_0, updates = v_9_cast_fp16, validate_indices = var_2096_validate_indices_0)[name = tensor("op_2096_cast_fp16")]; + tensor concat_36 = const()[name = tensor("concat_36"), val = tensor([1, 0, 0, 0, 0])]; + tensor concat_37 = const()[name = tensor("concat_37"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_9_internal_tensor_assign_2_stride_0 = const()[name = tensor("new_cache_9_internal_tensor_assign_2_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_9_internal_tensor_assign_2_begin_mask_0 = const()[name = tensor("new_cache_9_internal_tensor_assign_2_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_9_internal_tensor_assign_2_end_mask_0 = const()[name = tensor("new_cache_9_internal_tensor_assign_2_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_9_internal_tensor_assign_2_squeeze_mask_0 = const()[name = tensor("new_cache_9_internal_tensor_assign_2_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_55 = const()[name = tensor("shape_55"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_9 = const()[name = tensor("reduce_prod_9"), val = tensor(1048576)]; + tensor range_1d_9_start_0 = const()[name = tensor("range_1d_9_start_0"), val = tensor(0)]; + tensor range_1d_9_step_0 = const()[name = tensor("range_1d_9_step_0"), val = tensor(1)]; + tensor range_1d_9 = range_1d(end = reduce_prod_9, start = range_1d_9_start_0, step = range_1d_9_step_0)[name = tensor("range_1d_9")]; + tensor reshape_45 = reshape(shape = shape_55, x = range_1d_9)[name = tensor("reshape_45")]; + tensor slice_by_index_9 = slice_by_index(begin = concat_36, begin_mask = new_cache_9_internal_tensor_assign_2_begin_mask_0, end = concat_37, end_mask = new_cache_9_internal_tensor_assign_2_end_mask_0, squeeze_mask = new_cache_9_internal_tensor_assign_2_squeeze_mask_0, stride = new_cache_9_internal_tensor_assign_2_stride_0, x = reshape_45)[name = tensor("slice_by_index_9")]; + tensor reshape_46_shape_0 = const()[name = tensor("reshape_46_shape_0"), val = tensor([-1])]; + tensor reshape_46 = reshape(shape = reshape_46_shape_0, x = slice_by_index_9)[name = tensor("reshape_46")]; + tensor reshape_47_shape_0 = const()[name = tensor("reshape_47_shape_0"), val = tensor([-1])]; + tensor reshape_47_cast_fp16 = reshape(shape = reshape_47_shape_0, x = var_2096_cast_fp16)[name = tensor("reshape_47_cast_fp16")]; + tensor reshape_48_shape_0 = const()[name = tensor("reshape_48_shape_0"), val = tensor([-1])]; + tensor reshape_48_cast_fp16 = reshape(shape = reshape_48_shape_0, x = reshape_44_cast_fp16)[name = tensor("reshape_48_cast_fp16")]; + tensor scatter_9_mode_0 = const()[name = tensor("scatter_9_mode_0"), val = tensor("update")]; + tensor scatter_9_axis_0 = const()[name = tensor("scatter_9_axis_0"), val = tensor(0)]; + tensor scatter_9_validate_indices_0 = const()[name = tensor("scatter_9_validate_indices_0"), val = tensor(false)]; + tensor scatter_9_cast_fp16 = scatter(axis = scatter_9_axis_0, data = reshape_48_cast_fp16, indices = reshape_46, mode = scatter_9_mode_0, updates = reshape_47_cast_fp16, validate_indices = scatter_9_validate_indices_0)[name = tensor("scatter_9_cast_fp16")]; + tensor reshape_49_cast_fp16 = reshape(shape = shape_55, x = scatter_9_cast_fp16)[name = tensor("reshape_49_cast_fp16")]; + tensor reshape_49_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("reshape_49_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor keys_25_begin_0 = const()[name = tensor("keys_25_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor keys_25_end_0 = const()[name = tensor("keys_25_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor keys_25_end_mask_0 = const()[name = tensor("keys_25_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor keys_25_squeeze_mask_0 = const()[name = tensor("keys_25_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor keys_25_cast_fp16 = slice_by_index(begin = keys_25_begin_0, end = keys_25_end_0, end_mask = keys_25_end_mask_0, squeeze_mask = keys_25_squeeze_mask_0, x = reshape_49_cast_fp16)[name = tensor("keys_25_cast_fp16")]; + tensor values_25_begin_0 = const()[name = tensor("values_25_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor values_25_end_0 = const()[name = tensor("values_25_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor values_25_end_mask_0 = const()[name = tensor("values_25_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor values_25_squeeze_mask_0 = const()[name = tensor("values_25_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor values_25_cast_fp16 = slice_by_index(begin = values_25_begin_0, end = values_25_end_0, end_mask = values_25_end_mask_0, squeeze_mask = values_25_squeeze_mask_0, x = reshape_49_cast_fp16)[name = tensor("values_25_cast_fp16")]; + tensor var_2108_cast_fp16 = not_equal(x = keys_25_cast_fp16, y = keys_25_cast_fp16)[name = tensor("op_2108_cast_fp16")]; + tensor keys_27_cast_fp16 = select(a = var_470_to_fp16, b = keys_25_cast_fp16, cond = var_2108_cast_fp16)[name = tensor("keys_27_cast_fp16")]; + tensor var_2116_cast_fp16 = not_equal(x = values_25_cast_fp16, y = values_25_cast_fp16)[name = tensor("op_2116_cast_fp16")]; + tensor values_27_cast_fp16 = select(a = var_470_to_fp16, b = values_25_cast_fp16, cond = var_2116_cast_fp16)[name = tensor("values_27_cast_fp16")]; + tensor var_2140 = const()[name = tensor("op_2140"), val = tensor([0, 2, 1, 3])]; + tensor var_2156 = const()[name = tensor("op_2156"), val = tensor([1, 1, 1])]; + tensor var_2157_cast_fp16 = reshape(shape = var_2156, x = position4_to_fp16)[name = tensor("op_2157_cast_fp16")]; + tensor q_positions_9_cast_fp16 = add(x = var_2157_cast_fp16, y = q_offsets_1_promoted_to_fp16)[name = tensor("q_positions_9_cast_fp16")]; + tensor valid_end_9_cast_fp16 = add(x = var_2157_cast_fp16, y = var_536_cast_fp16)[name = tensor("valid_end_9_cast_fp16")]; + tensor valid_mask_9_cast_fp16 = less(x = k_positions_1_promoted_to_fp16, y = valid_end_9_cast_fp16)[name = tensor("valid_mask_9_cast_fp16")]; + tensor causal_mask_9_cast_fp16 = less_equal(x = k_positions_1_promoted_to_fp16, y = q_positions_9_cast_fp16)[name = tensor("causal_mask_9_cast_fp16")]; + tensor var_2185 = logical_and(x = valid_mask_9_cast_fp16, y = causal_mask_9_cast_fp16)[name = tensor("op_2185")]; + tensor attn_mask_9_axes_0 = const()[name = tensor("attn_mask_9_axes_0"), val = tensor([1])]; + tensor attn_mask_9 = expand_dims(axes = attn_mask_9_axes_0, x = var_2185)[name = tensor("attn_mask_9")]; + tensor var_2199_transpose_x_0 = const()[name = tensor("op_2199_transpose_x_0"), val = tensor(false)]; + tensor var_2199_transpose_y_0 = const()[name = tensor("op_2199_transpose_y_0"), val = tensor(false)]; + tensor transpose_77_perm_0 = const()[name = tensor("transpose_77_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_78_perm_0 = const()[name = tensor("transpose_78_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_78 = transpose(perm = transpose_78_perm_0, x = keys_27_cast_fp16)[name = tensor("transpose_188")]; + tensor transpose_77 = transpose(perm = transpose_77_perm_0, x = q_27_cast_fp16)[name = tensor("transpose_189")]; + tensor var_2199_cast_fp16 = matmul(transpose_x = var_2199_transpose_x_0, transpose_y = var_2199_transpose_y_0, x = transpose_77, y = transpose_78)[name = tensor("op_2199_cast_fp16")]; + tensor var_2193_to_fp16 = const()[name = tensor("op_2193_to_fp16"), val = tensor([0x1p-3])]; + tensor attn_weights_25_cast_fp16 = mul(x = var_2199_cast_fp16, y = var_2193_to_fp16)[name = tensor("attn_weights_25_cast_fp16")]; + tensor var_2201 = logical_not(x = attn_mask_9)[name = tensor("op_2201")]; + tensor var_2202_to_fp16 = const()[name = tensor("op_2202_to_fp16"), val = tensor(-inf)]; + tensor attn_weights_27_cast_fp16 = select(a = var_2202_to_fp16, b = attn_weights_25_cast_fp16, cond = var_2201)[name = tensor("attn_weights_27_cast_fp16")]; + tensor var_2204 = const()[name = tensor("op_2204"), val = tensor(-1)]; + tensor attn_weights_29_cast_fp16 = softmax(axis = var_2204, x = attn_weights_27_cast_fp16)[name = tensor("attn_weights_29_cast_fp16")]; + tensor attn_output_9_transpose_x_0 = const()[name = tensor("attn_output_9_transpose_x_0"), val = tensor(false)]; + tensor attn_output_9_transpose_y_0 = const()[name = tensor("attn_output_9_transpose_y_0"), val = tensor(false)]; + tensor values_29_cast_fp16 = transpose(perm = var_2140, x = values_27_cast_fp16)[name = tensor("transpose_190")]; + tensor attn_output_9_cast_fp16 = matmul(transpose_x = attn_output_9_transpose_x_0, transpose_y = attn_output_9_transpose_y_0, x = attn_weights_29_cast_fp16, y = values_29_cast_fp16)[name = tensor("attn_output_9_cast_fp16")]; + tensor var_2212 = const()[name = tensor("op_2212"), val = tensor([0, 2, 1, 3])]; + tensor var_2217 = const()[name = tensor("op_2217"), val = tensor([1, 256, 1024])]; + tensor var_2213_cast_fp16 = transpose(perm = var_2212, x = attn_output_9_cast_fp16)[name = tensor("transpose_187")]; + tensor input_53_cast_fp16 = reshape(shape = var_2217, x = var_2213_cast_fp16)[name = tensor("input_53_cast_fp16")]; + tensor attn4_out_proj_weight_to_fp16 = const()[name = tensor("attn4_out_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(108062592)))]; + tensor linear_17_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = attn4_out_proj_weight_to_fp16, x = input_53_cast_fp16)[name = tensor("linear_17_cast_fp16")]; + tensor var_2226_cast_fp16 = add(x = position4_to_fp16, y = valid_len_to_fp16)[name = tensor("op_2226_cast_fp16")]; + tensor var_2226_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_2226_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor input_55_cast_fp16 = add(x = input_49_cast_fp16, y = linear_17_cast_fp16)[name = tensor("input_55_cast_fp16")]; + tensor input_57_axes_0 = const()[name = tensor("input_57_axes_0"), val = tensor([-1])]; + tensor norm4_2_weight_to_fp16 = const()[name = tensor("norm4_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(110159808)))]; + tensor norm4_2_bias_to_fp16 = const()[name = tensor("norm4_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(110161920)))]; + tensor var_2230_to_fp16 = const()[name = tensor("op_2230_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_57_cast_fp16 = layer_norm(axes = input_57_axes_0, beta = norm4_2_bias_to_fp16, epsilon = var_2230_to_fp16, gamma = norm4_2_weight_to_fp16, x = input_55_cast_fp16)[name = tensor("input_57_cast_fp16")]; + tensor linear4_1_weight_to_fp16 = const()[name = tensor("linear4_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(110164032)))]; + tensor linear_18_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = linear4_1_weight_to_fp16, x = input_57_cast_fp16)[name = tensor("linear_18_cast_fp16")]; + tensor input_59_mode_0 = const()[name = tensor("input_59_mode_0"), val = tensor("EXACT")]; + tensor input_59_cast_fp16 = gelu(mode = input_59_mode_0, x = linear_18_cast_fp16)[name = tensor("input_59_cast_fp16")]; + tensor linear4_2_weight_to_fp16 = const()[name = tensor("linear4_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(118552704)))]; + tensor linear_19_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = linear4_2_weight_to_fp16, x = input_59_cast_fp16)[name = tensor("linear_19_cast_fp16")]; + tensor input_61_cast_fp16 = add(x = input_55_cast_fp16, y = linear_19_cast_fp16)[name = tensor("input_61_cast_fp16")]; + tensor input_63_axes_0 = const()[name = tensor("input_63_axes_0"), val = tensor([-1])]; + tensor norm5_1_weight_to_fp16 = const()[name = tensor("norm5_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(126941376)))]; + tensor norm5_1_bias_to_fp16 = const()[name = tensor("norm5_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(126943488)))]; + tensor var_2247_to_fp16 = const()[name = tensor("op_2247_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_63_cast_fp16 = layer_norm(axes = input_63_axes_0, beta = norm5_1_bias_to_fp16, epsilon = var_2247_to_fp16, gamma = norm5_1_weight_to_fp16, x = input_61_cast_fp16)[name = tensor("input_63_cast_fp16")]; + tensor attn5_in_proj_weight_to_fp16 = const()[name = tensor("attn5_in_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(126945600)))]; + tensor linear_20_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = attn5_in_proj_weight_to_fp16, x = input_63_cast_fp16)[name = tensor("linear_20_cast_fp16")]; + tensor var_2261 = const()[name = tensor("op_2261"), val = tensor([1, 256, 3, 16, 64])]; + tensor qkv_11_cast_fp16 = reshape(shape = var_2261, x = linear_20_cast_fp16)[name = tensor("qkv_11_cast_fp16")]; + tensor q_31_begin_0 = const()[name = tensor("q_31_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor q_31_end_0 = const()[name = tensor("q_31_end_0"), val = tensor([1, 256, 1, 16, 64])]; + tensor q_31_end_mask_0 = const()[name = tensor("q_31_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor q_31_squeeze_mask_0 = const()[name = tensor("q_31_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor q_31_cast_fp16 = slice_by_index(begin = q_31_begin_0, end = q_31_end_0, end_mask = q_31_end_mask_0, squeeze_mask = q_31_squeeze_mask_0, x = qkv_11_cast_fp16)[name = tensor("q_31_cast_fp16")]; + tensor k_21_begin_0 = const()[name = tensor("k_21_begin_0"), val = tensor([0, 0, 1, 0, 0])]; + tensor k_21_end_0 = const()[name = tensor("k_21_end_0"), val = tensor([1, 256, 2, 16, 64])]; + tensor k_21_end_mask_0 = const()[name = tensor("k_21_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor k_21_squeeze_mask_0 = const()[name = tensor("k_21_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor k_21_cast_fp16 = slice_by_index(begin = k_21_begin_0, end = k_21_end_0, end_mask = k_21_end_mask_0, squeeze_mask = k_21_squeeze_mask_0, x = qkv_11_cast_fp16)[name = tensor("k_21_cast_fp16")]; + tensor v_11_begin_0 = const()[name = tensor("v_11_begin_0"), val = tensor([0, 0, 2, 0, 0])]; + tensor v_11_end_0 = const()[name = tensor("v_11_end_0"), val = tensor([1, 256, 3, 16, 64])]; + tensor v_11_end_mask_0 = const()[name = tensor("v_11_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor v_11_squeeze_mask_0 = const()[name = tensor("v_11_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor v_11_cast_fp16 = slice_by_index(begin = v_11_begin_0, end = v_11_end_0, end_mask = v_11_end_mask_0, squeeze_mask = v_11_squeeze_mask_0, x = qkv_11_cast_fp16)[name = tensor("v_11_cast_fp16")]; + tensor var_2319 = const()[name = tensor("op_2319"), val = tensor([1, 1])]; + tensor position5_to_fp16_dtype_0 = const()[name = tensor("position5_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor position5_to_fp16 = cast(dtype = position5_to_fp16_dtype_0, x = position5)[name = tensor("cast_636")]; + tensor var_2320_cast_fp16 = reshape(shape = var_2319, x = position5_to_fp16)[name = tensor("op_2320_cast_fp16")]; + tensor ts_33_cast_fp16 = add(x = ts_1_promoted_to_fp16, y = var_2320_cast_fp16)[name = tensor("ts_33_cast_fp16")]; + tensor var_2327 = const()[name = tensor("op_2327"), val = tensor([1, 256, 1, 1])]; + tensor ts_35_cast_fp16 = reshape(shape = var_2327, x = ts_33_cast_fp16)[name = tensor("ts_35_cast_fp16")]; + tensor var_2334 = const()[name = tensor("op_2334"), val = tensor([1, 256, 16, 32, 2])]; + tensor q_complex_11_cast_fp16 = reshape(shape = var_2334, x = q_31_cast_fp16)[name = tensor("q_complex_11_cast_fp16")]; + tensor var_2341 = const()[name = tensor("op_2341"), val = tensor([1, 256, 16, 32, 2])]; + tensor k_complex_11_cast_fp16 = reshape(shape = var_2341, x = k_21_cast_fp16)[name = tensor("k_complex_11_cast_fp16")]; + tensor var_2345_begin_0 = const()[name = tensor("op_2345_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_2345_end_0 = const()[name = tensor("op_2345_end_0"), val = tensor([1, 256, 16, 32, 1])]; + tensor var_2345_end_mask_0 = const()[name = tensor("op_2345_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_2345_squeeze_mask_0 = const()[name = tensor("op_2345_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_2345_cast_fp16 = slice_by_index(begin = var_2345_begin_0, end = var_2345_end_0, end_mask = var_2345_end_mask_0, squeeze_mask = var_2345_squeeze_mask_0, x = q_complex_11_cast_fp16)[name = tensor("op_2345_cast_fp16")]; + tensor var_2353_begin_0 = const()[name = tensor("op_2353_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_2353_end_0 = const()[name = tensor("op_2353_end_0"), val = tensor([1, 256, 16, 32, 2])]; + tensor var_2353_end_mask_0 = const()[name = tensor("op_2353_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_2353_squeeze_mask_0 = const()[name = tensor("op_2353_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_2353_cast_fp16 = slice_by_index(begin = var_2353_begin_0, end = var_2353_end_0, end_mask = var_2353_end_mask_0, squeeze_mask = var_2353_squeeze_mask_0, x = q_complex_11_cast_fp16)[name = tensor("op_2353_cast_fp16")]; + tensor var_2361_begin_0 = const()[name = tensor("op_2361_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_2361_end_0 = const()[name = tensor("op_2361_end_0"), val = tensor([1, 256, 16, 32, 1])]; + tensor var_2361_end_mask_0 = const()[name = tensor("op_2361_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_2361_squeeze_mask_0 = const()[name = tensor("op_2361_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_2361_cast_fp16 = slice_by_index(begin = var_2361_begin_0, end = var_2361_end_0, end_mask = var_2361_end_mask_0, squeeze_mask = var_2361_squeeze_mask_0, x = k_complex_11_cast_fp16)[name = tensor("op_2361_cast_fp16")]; + tensor var_2369_begin_0 = const()[name = tensor("op_2369_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_2369_end_0 = const()[name = tensor("op_2369_end_0"), val = tensor([1, 256, 16, 32, 2])]; + tensor var_2369_end_mask_0 = const()[name = tensor("op_2369_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_2369_squeeze_mask_0 = const()[name = tensor("op_2369_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_2369_cast_fp16 = slice_by_index(begin = var_2369_begin_0, end = var_2369_end_0, end_mask = var_2369_end_mask_0, squeeze_mask = var_2369_squeeze_mask_0, x = k_complex_11_cast_fp16)[name = tensor("op_2369_cast_fp16")]; + tensor freqs_11_to_fp16 = const()[name = tensor("freqs_11_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(133237120)))]; + tensor var_2375_cast_fp16 = mul(x = freqs_11_to_fp16, y = ts_35_cast_fp16)[name = tensor("op_2375_cast_fp16")]; + tensor rotr_11_cast_fp16 = cos(x = var_2375_cast_fp16)[name = tensor("rotr_11_cast_fp16")]; + tensor roti_11_cast_fp16 = sin(x = var_2375_cast_fp16)[name = tensor("roti_11_cast_fp16")]; + tensor var_2379_cast_fp16 = mul(x = var_2345_cast_fp16, y = rotr_11_cast_fp16)[name = tensor("op_2379_cast_fp16")]; + tensor var_2380_cast_fp16 = mul(x = var_2353_cast_fp16, y = roti_11_cast_fp16)[name = tensor("op_2380_cast_fp16")]; + tensor qor_21_cast_fp16 = sub(x = var_2379_cast_fp16, y = var_2380_cast_fp16)[name = tensor("qor_21_cast_fp16")]; + tensor var_2383_cast_fp16 = mul(x = var_2345_cast_fp16, y = roti_11_cast_fp16)[name = tensor("op_2383_cast_fp16")]; + tensor var_2384_cast_fp16 = mul(x = var_2353_cast_fp16, y = rotr_11_cast_fp16)[name = tensor("op_2384_cast_fp16")]; + tensor qoi_21_cast_fp16 = add(x = var_2383_cast_fp16, y = var_2384_cast_fp16)[name = tensor("qoi_21_cast_fp16")]; + tensor var_2387_cast_fp16 = mul(x = var_2361_cast_fp16, y = rotr_11_cast_fp16)[name = tensor("op_2387_cast_fp16")]; + tensor var_2388_cast_fp16 = mul(x = var_2369_cast_fp16, y = roti_11_cast_fp16)[name = tensor("op_2388_cast_fp16")]; + tensor kor_21_cast_fp16 = sub(x = var_2387_cast_fp16, y = var_2388_cast_fp16)[name = tensor("kor_21_cast_fp16")]; + tensor var_2391_cast_fp16 = mul(x = var_2361_cast_fp16, y = roti_11_cast_fp16)[name = tensor("op_2391_cast_fp16")]; + tensor var_2392_cast_fp16 = mul(x = var_2369_cast_fp16, y = rotr_11_cast_fp16)[name = tensor("op_2392_cast_fp16")]; + tensor koi_21_cast_fp16 = add(x = var_2391_cast_fp16, y = var_2392_cast_fp16)[name = tensor("koi_21_cast_fp16")]; + tensor qo_11_axis_0 = const()[name = tensor("qo_11_axis_0"), val = tensor(-1)]; + tensor qo_11_cast_fp16 = stack(axis = qo_11_axis_0, values = (qor_21_cast_fp16, qoi_21_cast_fp16))[name = tensor("qo_11_cast_fp16")]; + tensor ko_11_axis_0 = const()[name = tensor("ko_11_axis_0"), val = tensor(-1)]; + tensor ko_11_cast_fp16 = stack(axis = ko_11_axis_0, values = (kor_21_cast_fp16, koi_21_cast_fp16))[name = tensor("ko_11_cast_fp16")]; + tensor var_2425 = const()[name = tensor("op_2425"), val = tensor([1, 256, 16, 64])]; + tensor q_33_cast_fp16 = reshape(shape = var_2425, x = qo_11_cast_fp16)[name = tensor("q_33_cast_fp16")]; + tensor var_2431 = const()[name = tensor("op_2431"), val = tensor([1, 256, 16, 64])]; + tensor k_23_cast_fp16 = reshape(shape = var_2431, x = ko_11_cast_fp16)[name = tensor("k_23_cast_fp16")]; + tensor write_indices_float_31_cast_fp16 = add(x = var_2320_cast_fp16, y = var_393_promoted_to_fp16)[name = tensor("write_indices_float_31_cast_fp16")]; + tensor _inversed_2452_y_0_to_fp16 = const()[name = tensor("_inversed_2452_y_0_to_fp16"), val = tensor(0x1p-9)]; + tensor _inversed_2452_cast_fp16 = mul(x = write_indices_float_31_cast_fp16, y = _inversed_2452_y_0_to_fp16)[name = tensor("_inversed_2452_cast_fp16")]; + tensor var_2453_cast_fp16 = floor(x = _inversed_2452_cast_fp16)[name = tensor("op_2453_cast_fp16")]; + tensor var_2454_to_fp16 = const()[name = tensor("op_2454_to_fp16"), val = tensor(0x1p+9)]; + tensor var_2455_cast_fp16 = mul(x = var_2453_cast_fp16, y = var_2454_to_fp16)[name = tensor("op_2455_cast_fp16")]; + tensor write_indices_float_33_cast_fp16 = sub(x = write_indices_float_31_cast_fp16, y = var_2455_cast_fp16)[name = tensor("write_indices_float_33_cast_fp16")]; + tensor var_2473_value_0_to_fp16 = const()[name = tensor("op_2473_value_0_to_fp16"), val = tensor(0x1.ffp+8)]; + tensor var_2473_cast_fp16 = fill_like(ref_tensor = write_indices_float_33_cast_fp16, value = var_2473_value_0_to_fp16)[name = tensor("op_2473_cast_fp16")]; + tensor write_indices_float_35_cast_fp16 = select(a = write_indices_float_33_cast_fp16, b = var_2473_cast_fp16, cond = token_valid_1_cast_fp16)[name = tensor("write_indices_float_35_cast_fp16")]; + tensor var_2479_dtype_0 = const()[name = tensor("op_2479_dtype_0"), val = tensor("int32")]; + tensor var_2484 = const()[name = tensor("op_2484"), val = tensor([1, 256, 1, 1])]; + tensor write_indices_float_35_cast_fp16_to_int32 = cast(dtype = var_2479_dtype_0, x = write_indices_float_35_cast_fp16)[name = tensor("cast_635")]; + tensor var_2485 = reshape(shape = var_2484, x = write_indices_float_35_cast_fp16_to_int32)[name = tensor("op_2485")]; + tensor write_indices_11_reps_0 = const()[name = tensor("write_indices_11_reps_0"), val = tensor([1, 1, 16, 64])]; + tensor write_indices_11 = tile(reps = write_indices_11_reps_0, x = var_2485)[name = tensor("write_indices_11")]; + tensor var_2495_begin_0 = const()[name = tensor("op_2495_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_2495_end_0 = const()[name = tensor("op_2495_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor var_2495_end_mask_0 = const()[name = tensor("op_2495_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_2495_squeeze_mask_0 = const()[name = tensor("op_2495_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor cache5_to_fp16_dtype_0 = const()[name = tensor("cache5_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor cache5_to_fp16 = cast(dtype = cache5_to_fp16_dtype_0, x = cache5)[name = tensor("cast_634")]; + tensor var_2495_cast_fp16 = slice_by_index(begin = var_2495_begin_0, end = var_2495_end_0, end_mask = var_2495_end_mask_0, squeeze_mask = var_2495_squeeze_mask_0, x = cache5_to_fp16)[name = tensor("op_2495_cast_fp16")]; + tensor var_2497_axis_0 = const()[name = tensor("op_2497_axis_0"), val = tensor(1)]; + tensor var_2497_mode_0 = const()[name = tensor("op_2497_mode_0"), val = tensor("update")]; + tensor var_2497_validate_indices_0 = const()[name = tensor("op_2497_validate_indices_0"), val = tensor(false)]; + tensor var_2497_cast_fp16 = scatter_along_axis(axis = var_2497_axis_0, data = var_2495_cast_fp16, indices = write_indices_11, mode = var_2497_mode_0, updates = k_23_cast_fp16, validate_indices = var_2497_validate_indices_0)[name = tensor("op_2497_cast_fp16")]; + tensor concat_42 = const()[name = tensor("concat_42"), val = tensor([0, 0, 0, 0, 0])]; + tensor concat_43 = const()[name = tensor("concat_43"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_11_internal_tensor_assign_1_stride_0 = const()[name = tensor("new_cache_11_internal_tensor_assign_1_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_11_internal_tensor_assign_1_begin_mask_0 = const()[name = tensor("new_cache_11_internal_tensor_assign_1_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_11_internal_tensor_assign_1_end_mask_0 = const()[name = tensor("new_cache_11_internal_tensor_assign_1_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_11_internal_tensor_assign_1_squeeze_mask_0 = const()[name = tensor("new_cache_11_internal_tensor_assign_1_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_56 = const()[name = tensor("shape_56"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_10 = const()[name = tensor("reduce_prod_10"), val = tensor(1048576)]; + tensor range_1d_10_start_0 = const()[name = tensor("range_1d_10_start_0"), val = tensor(0)]; + tensor range_1d_10_step_0 = const()[name = tensor("range_1d_10_step_0"), val = tensor(1)]; + tensor range_1d_10 = range_1d(end = reduce_prod_10, start = range_1d_10_start_0, step = range_1d_10_step_0)[name = tensor("range_1d_10")]; + tensor reshape_50 = reshape(shape = shape_56, x = range_1d_10)[name = tensor("reshape_50")]; + tensor slice_by_index_10 = slice_by_index(begin = concat_42, begin_mask = new_cache_11_internal_tensor_assign_1_begin_mask_0, end = concat_43, end_mask = new_cache_11_internal_tensor_assign_1_end_mask_0, squeeze_mask = new_cache_11_internal_tensor_assign_1_squeeze_mask_0, stride = new_cache_11_internal_tensor_assign_1_stride_0, x = reshape_50)[name = tensor("slice_by_index_10")]; + tensor reshape_51_shape_0 = const()[name = tensor("reshape_51_shape_0"), val = tensor([-1])]; + tensor reshape_51 = reshape(shape = reshape_51_shape_0, x = slice_by_index_10)[name = tensor("reshape_51")]; + tensor reshape_52_shape_0 = const()[name = tensor("reshape_52_shape_0"), val = tensor([-1])]; + tensor reshape_52_cast_fp16 = reshape(shape = reshape_52_shape_0, x = var_2497_cast_fp16)[name = tensor("reshape_52_cast_fp16")]; + tensor reshape_53_shape_0 = const()[name = tensor("reshape_53_shape_0"), val = tensor([-1])]; + tensor reshape_53_cast_fp16 = reshape(shape = reshape_53_shape_0, x = cache5_to_fp16)[name = tensor("reshape_53_cast_fp16")]; + tensor scatter_10_mode_0 = const()[name = tensor("scatter_10_mode_0"), val = tensor("update")]; + tensor scatter_10_axis_0 = const()[name = tensor("scatter_10_axis_0"), val = tensor(0)]; + tensor scatter_10_validate_indices_0 = const()[name = tensor("scatter_10_validate_indices_0"), val = tensor(false)]; + tensor scatter_10_cast_fp16 = scatter(axis = scatter_10_axis_0, data = reshape_53_cast_fp16, indices = reshape_51, mode = scatter_10_mode_0, updates = reshape_52_cast_fp16, validate_indices = scatter_10_validate_indices_0)[name = tensor("scatter_10_cast_fp16")]; + tensor reshape_54_cast_fp16 = reshape(shape = shape_56, x = scatter_10_cast_fp16)[name = tensor("reshape_54_cast_fp16")]; + tensor var_2505_begin_0 = const()[name = tensor("op_2505_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor var_2505_end_0 = const()[name = tensor("op_2505_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor var_2505_end_mask_0 = const()[name = tensor("op_2505_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_2505_squeeze_mask_0 = const()[name = tensor("op_2505_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor var_2505_cast_fp16 = slice_by_index(begin = var_2505_begin_0, end = var_2505_end_0, end_mask = var_2505_end_mask_0, squeeze_mask = var_2505_squeeze_mask_0, x = reshape_54_cast_fp16)[name = tensor("op_2505_cast_fp16")]; + tensor var_2507_axis_0 = const()[name = tensor("op_2507_axis_0"), val = tensor(1)]; + tensor var_2507_mode_0 = const()[name = tensor("op_2507_mode_0"), val = tensor("update")]; + tensor var_2507_validate_indices_0 = const()[name = tensor("op_2507_validate_indices_0"), val = tensor(false)]; + tensor var_2507_cast_fp16 = scatter_along_axis(axis = var_2507_axis_0, data = var_2505_cast_fp16, indices = write_indices_11, mode = var_2507_mode_0, updates = v_11_cast_fp16, validate_indices = var_2507_validate_indices_0)[name = tensor("op_2507_cast_fp16")]; + tensor concat_44 = const()[name = tensor("concat_44"), val = tensor([1, 0, 0, 0, 0])]; + tensor concat_45 = const()[name = tensor("concat_45"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_11_internal_tensor_assign_2_stride_0 = const()[name = tensor("new_cache_11_internal_tensor_assign_2_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_11_internal_tensor_assign_2_begin_mask_0 = const()[name = tensor("new_cache_11_internal_tensor_assign_2_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_11_internal_tensor_assign_2_end_mask_0 = const()[name = tensor("new_cache_11_internal_tensor_assign_2_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_11_internal_tensor_assign_2_squeeze_mask_0 = const()[name = tensor("new_cache_11_internal_tensor_assign_2_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_57 = const()[name = tensor("shape_57"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_11 = const()[name = tensor("reduce_prod_11"), val = tensor(1048576)]; + tensor range_1d_11_start_0 = const()[name = tensor("range_1d_11_start_0"), val = tensor(0)]; + tensor range_1d_11_step_0 = const()[name = tensor("range_1d_11_step_0"), val = tensor(1)]; + tensor range_1d_11 = range_1d(end = reduce_prod_11, start = range_1d_11_start_0, step = range_1d_11_step_0)[name = tensor("range_1d_11")]; + tensor reshape_55 = reshape(shape = shape_57, x = range_1d_11)[name = tensor("reshape_55")]; + tensor slice_by_index_11 = slice_by_index(begin = concat_44, begin_mask = new_cache_11_internal_tensor_assign_2_begin_mask_0, end = concat_45, end_mask = new_cache_11_internal_tensor_assign_2_end_mask_0, squeeze_mask = new_cache_11_internal_tensor_assign_2_squeeze_mask_0, stride = new_cache_11_internal_tensor_assign_2_stride_0, x = reshape_55)[name = tensor("slice_by_index_11")]; + tensor reshape_56_shape_0 = const()[name = tensor("reshape_56_shape_0"), val = tensor([-1])]; + tensor reshape_56 = reshape(shape = reshape_56_shape_0, x = slice_by_index_11)[name = tensor("reshape_56")]; + tensor reshape_57_shape_0 = const()[name = tensor("reshape_57_shape_0"), val = tensor([-1])]; + tensor reshape_57_cast_fp16 = reshape(shape = reshape_57_shape_0, x = var_2507_cast_fp16)[name = tensor("reshape_57_cast_fp16")]; + tensor reshape_58_shape_0 = const()[name = tensor("reshape_58_shape_0"), val = tensor([-1])]; + tensor reshape_58_cast_fp16 = reshape(shape = reshape_58_shape_0, x = reshape_54_cast_fp16)[name = tensor("reshape_58_cast_fp16")]; + tensor scatter_11_mode_0 = const()[name = tensor("scatter_11_mode_0"), val = tensor("update")]; + tensor scatter_11_axis_0 = const()[name = tensor("scatter_11_axis_0"), val = tensor(0)]; + tensor scatter_11_validate_indices_0 = const()[name = tensor("scatter_11_validate_indices_0"), val = tensor(false)]; + tensor scatter_11_cast_fp16 = scatter(axis = scatter_11_axis_0, data = reshape_58_cast_fp16, indices = reshape_56, mode = scatter_11_mode_0, updates = reshape_57_cast_fp16, validate_indices = scatter_11_validate_indices_0)[name = tensor("scatter_11_cast_fp16")]; + tensor reshape_59_cast_fp16 = reshape(shape = shape_57, x = scatter_11_cast_fp16)[name = tensor("reshape_59_cast_fp16")]; + tensor reshape_59_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("reshape_59_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor keys_31_begin_0 = const()[name = tensor("keys_31_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor keys_31_end_0 = const()[name = tensor("keys_31_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor keys_31_end_mask_0 = const()[name = tensor("keys_31_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor keys_31_squeeze_mask_0 = const()[name = tensor("keys_31_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor keys_31_cast_fp16 = slice_by_index(begin = keys_31_begin_0, end = keys_31_end_0, end_mask = keys_31_end_mask_0, squeeze_mask = keys_31_squeeze_mask_0, x = reshape_59_cast_fp16)[name = tensor("keys_31_cast_fp16")]; + tensor values_31_begin_0 = const()[name = tensor("values_31_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor values_31_end_0 = const()[name = tensor("values_31_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor values_31_end_mask_0 = const()[name = tensor("values_31_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor values_31_squeeze_mask_0 = const()[name = tensor("values_31_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor values_31_cast_fp16 = slice_by_index(begin = values_31_begin_0, end = values_31_end_0, end_mask = values_31_end_mask_0, squeeze_mask = values_31_squeeze_mask_0, x = reshape_59_cast_fp16)[name = tensor("values_31_cast_fp16")]; + tensor var_2519_cast_fp16 = not_equal(x = keys_31_cast_fp16, y = keys_31_cast_fp16)[name = tensor("op_2519_cast_fp16")]; + tensor keys_33_cast_fp16 = select(a = var_470_to_fp16, b = keys_31_cast_fp16, cond = var_2519_cast_fp16)[name = tensor("keys_33_cast_fp16")]; + tensor var_2527_cast_fp16 = not_equal(x = values_31_cast_fp16, y = values_31_cast_fp16)[name = tensor("op_2527_cast_fp16")]; + tensor values_33_cast_fp16 = select(a = var_470_to_fp16, b = values_31_cast_fp16, cond = var_2527_cast_fp16)[name = tensor("values_33_cast_fp16")]; + tensor var_2551 = const()[name = tensor("op_2551"), val = tensor([0, 2, 1, 3])]; + tensor var_2567 = const()[name = tensor("op_2567"), val = tensor([1, 1, 1])]; + tensor var_2568_cast_fp16 = reshape(shape = var_2567, x = position5_to_fp16)[name = tensor("op_2568_cast_fp16")]; + tensor q_positions_11_cast_fp16 = add(x = var_2568_cast_fp16, y = q_offsets_1_promoted_to_fp16)[name = tensor("q_positions_11_cast_fp16")]; + tensor valid_end_11_cast_fp16 = add(x = var_2568_cast_fp16, y = var_536_cast_fp16)[name = tensor("valid_end_11_cast_fp16")]; + tensor valid_mask_11_cast_fp16 = less(x = k_positions_1_promoted_to_fp16, y = valid_end_11_cast_fp16)[name = tensor("valid_mask_11_cast_fp16")]; + tensor causal_mask_11_cast_fp16 = less_equal(x = k_positions_1_promoted_to_fp16, y = q_positions_11_cast_fp16)[name = tensor("causal_mask_11_cast_fp16")]; + tensor var_2596 = logical_and(x = valid_mask_11_cast_fp16, y = causal_mask_11_cast_fp16)[name = tensor("op_2596")]; + tensor attn_mask_11_axes_0 = const()[name = tensor("attn_mask_11_axes_0"), val = tensor([1])]; + tensor attn_mask_11 = expand_dims(axes = attn_mask_11_axes_0, x = var_2596)[name = tensor("attn_mask_11")]; + tensor var_2610_transpose_x_0 = const()[name = tensor("op_2610_transpose_x_0"), val = tensor(false)]; + tensor var_2610_transpose_y_0 = const()[name = tensor("op_2610_transpose_y_0"), val = tensor(false)]; + tensor transpose_79_perm_0 = const()[name = tensor("transpose_79_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_80_perm_0 = const()[name = tensor("transpose_80_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_80 = transpose(perm = transpose_80_perm_0, x = keys_33_cast_fp16)[name = tensor("transpose_184")]; + tensor transpose_79 = transpose(perm = transpose_79_perm_0, x = q_33_cast_fp16)[name = tensor("transpose_185")]; + tensor var_2610_cast_fp16 = matmul(transpose_x = var_2610_transpose_x_0, transpose_y = var_2610_transpose_y_0, x = transpose_79, y = transpose_80)[name = tensor("op_2610_cast_fp16")]; + tensor var_2604_to_fp16 = const()[name = tensor("op_2604_to_fp16"), val = tensor([0x1p-3])]; + tensor attn_weights_31_cast_fp16 = mul(x = var_2610_cast_fp16, y = var_2604_to_fp16)[name = tensor("attn_weights_31_cast_fp16")]; + tensor var_2612 = logical_not(x = attn_mask_11)[name = tensor("op_2612")]; + tensor var_2613_to_fp16 = const()[name = tensor("op_2613_to_fp16"), val = tensor(-inf)]; + tensor attn_weights_33_cast_fp16 = select(a = var_2613_to_fp16, b = attn_weights_31_cast_fp16, cond = var_2612)[name = tensor("attn_weights_33_cast_fp16")]; + tensor var_2615 = const()[name = tensor("op_2615"), val = tensor(-1)]; + tensor attn_weights_35_cast_fp16 = softmax(axis = var_2615, x = attn_weights_33_cast_fp16)[name = tensor("attn_weights_35_cast_fp16")]; + tensor attn_output_11_transpose_x_0 = const()[name = tensor("attn_output_11_transpose_x_0"), val = tensor(false)]; + tensor attn_output_11_transpose_y_0 = const()[name = tensor("attn_output_11_transpose_y_0"), val = tensor(false)]; + tensor values_35_cast_fp16 = transpose(perm = var_2551, x = values_33_cast_fp16)[name = tensor("transpose_186")]; + tensor attn_output_11_cast_fp16 = matmul(transpose_x = attn_output_11_transpose_x_0, transpose_y = attn_output_11_transpose_y_0, x = attn_weights_35_cast_fp16, y = values_35_cast_fp16)[name = tensor("attn_output_11_cast_fp16")]; + tensor var_2623 = const()[name = tensor("op_2623"), val = tensor([0, 2, 1, 3])]; + tensor var_2628 = const()[name = tensor("op_2628"), val = tensor([1, 256, 1024])]; + tensor var_2624_cast_fp16 = transpose(perm = var_2623, x = attn_output_11_cast_fp16)[name = tensor("transpose_183")]; + tensor input_65_cast_fp16 = reshape(shape = var_2628, x = var_2624_cast_fp16)[name = tensor("input_65_cast_fp16")]; + tensor attn5_out_proj_weight_to_fp16 = const()[name = tensor("attn5_out_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(133237248)))]; + tensor linear_21_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = attn5_out_proj_weight_to_fp16, x = input_65_cast_fp16)[name = tensor("linear_21_cast_fp16")]; + tensor var_2637_cast_fp16 = add(x = position5_to_fp16, y = valid_len_to_fp16)[name = tensor("op_2637_cast_fp16")]; + tensor var_2637_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_2637_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor input_67_cast_fp16 = add(x = input_61_cast_fp16, y = linear_21_cast_fp16)[name = tensor("input_67_cast_fp16")]; + tensor input_69_axes_0 = const()[name = tensor("input_69_axes_0"), val = tensor([-1])]; + tensor norm5_2_weight_to_fp16 = const()[name = tensor("norm5_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(135334464)))]; + tensor norm5_2_bias_to_fp16 = const()[name = tensor("norm5_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(135336576)))]; + tensor var_2641_to_fp16 = const()[name = tensor("op_2641_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_69_cast_fp16 = layer_norm(axes = input_69_axes_0, beta = norm5_2_bias_to_fp16, epsilon = var_2641_to_fp16, gamma = norm5_2_weight_to_fp16, x = input_67_cast_fp16)[name = tensor("input_69_cast_fp16")]; + tensor linear5_1_weight_to_fp16 = const()[name = tensor("linear5_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(135338688)))]; + tensor linear_22_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = linear5_1_weight_to_fp16, x = input_69_cast_fp16)[name = tensor("linear_22_cast_fp16")]; + tensor input_71_mode_0 = const()[name = tensor("input_71_mode_0"), val = tensor("EXACT")]; + tensor input_71_cast_fp16 = gelu(mode = input_71_mode_0, x = linear_22_cast_fp16)[name = tensor("input_71_cast_fp16")]; + tensor linear5_2_weight_to_fp16 = const()[name = tensor("linear5_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(143727360)))]; + tensor linear_23_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = linear5_2_weight_to_fp16, x = input_71_cast_fp16)[name = tensor("linear_23_cast_fp16")]; + tensor input_73_cast_fp16 = add(x = input_67_cast_fp16, y = linear_23_cast_fp16)[name = tensor("input_73_cast_fp16")]; + tensor input_75_axes_0 = const()[name = tensor("input_75_axes_0"), val = tensor([-1])]; + tensor norm6_1_weight_to_fp16 = const()[name = tensor("norm6_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(152116032)))]; + tensor norm6_1_bias_to_fp16 = const()[name = tensor("norm6_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(152118144)))]; + tensor var_2658_to_fp16 = const()[name = tensor("op_2658_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_75_cast_fp16 = layer_norm(axes = input_75_axes_0, beta = norm6_1_bias_to_fp16, epsilon = var_2658_to_fp16, gamma = norm6_1_weight_to_fp16, x = input_73_cast_fp16)[name = tensor("input_75_cast_fp16")]; + tensor attn6_in_proj_weight_to_fp16 = const()[name = tensor("attn6_in_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(152120256)))]; + tensor linear_24_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = attn6_in_proj_weight_to_fp16, x = input_75_cast_fp16)[name = tensor("linear_24_cast_fp16")]; + tensor var_2672 = const()[name = tensor("op_2672"), val = tensor([1, 256, 3, 16, 64])]; + tensor qkv_13_cast_fp16 = reshape(shape = var_2672, x = linear_24_cast_fp16)[name = tensor("qkv_13_cast_fp16")]; + tensor q_37_begin_0 = const()[name = tensor("q_37_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor q_37_end_0 = const()[name = tensor("q_37_end_0"), val = tensor([1, 256, 1, 16, 64])]; + tensor q_37_end_mask_0 = const()[name = tensor("q_37_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor q_37_squeeze_mask_0 = const()[name = tensor("q_37_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor q_37_cast_fp16 = slice_by_index(begin = q_37_begin_0, end = q_37_end_0, end_mask = q_37_end_mask_0, squeeze_mask = q_37_squeeze_mask_0, x = qkv_13_cast_fp16)[name = tensor("q_37_cast_fp16")]; + tensor k_25_begin_0 = const()[name = tensor("k_25_begin_0"), val = tensor([0, 0, 1, 0, 0])]; + tensor k_25_end_0 = const()[name = tensor("k_25_end_0"), val = tensor([1, 256, 2, 16, 64])]; + tensor k_25_end_mask_0 = const()[name = tensor("k_25_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor k_25_squeeze_mask_0 = const()[name = tensor("k_25_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor k_25_cast_fp16 = slice_by_index(begin = k_25_begin_0, end = k_25_end_0, end_mask = k_25_end_mask_0, squeeze_mask = k_25_squeeze_mask_0, x = qkv_13_cast_fp16)[name = tensor("k_25_cast_fp16")]; + tensor v_13_begin_0 = const()[name = tensor("v_13_begin_0"), val = tensor([0, 0, 2, 0, 0])]; + tensor v_13_end_0 = const()[name = tensor("v_13_end_0"), val = tensor([1, 256, 3, 16, 64])]; + tensor v_13_end_mask_0 = const()[name = tensor("v_13_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor v_13_squeeze_mask_0 = const()[name = tensor("v_13_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor v_13_cast_fp16 = slice_by_index(begin = v_13_begin_0, end = v_13_end_0, end_mask = v_13_end_mask_0, squeeze_mask = v_13_squeeze_mask_0, x = qkv_13_cast_fp16)[name = tensor("v_13_cast_fp16")]; + tensor var_2730 = const()[name = tensor("op_2730"), val = tensor([1, 1])]; + tensor position6_to_fp16_dtype_0 = const()[name = tensor("position6_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor position6_to_fp16 = cast(dtype = position6_to_fp16_dtype_0, x = position6)[name = tensor("cast_631")]; + tensor var_2731_cast_fp16 = reshape(shape = var_2730, x = position6_to_fp16)[name = tensor("op_2731_cast_fp16")]; + tensor ts_39_cast_fp16 = add(x = ts_1_promoted_to_fp16, y = var_2731_cast_fp16)[name = tensor("ts_39_cast_fp16")]; + tensor var_2738 = const()[name = tensor("op_2738"), val = tensor([1, 256, 1, 1])]; + tensor ts_41_cast_fp16 = reshape(shape = var_2738, x = ts_39_cast_fp16)[name = tensor("ts_41_cast_fp16")]; + tensor var_2745 = const()[name = tensor("op_2745"), val = tensor([1, 256, 16, 32, 2])]; + tensor q_complex_13_cast_fp16 = reshape(shape = var_2745, x = q_37_cast_fp16)[name = tensor("q_complex_13_cast_fp16")]; + tensor var_2752 = const()[name = tensor("op_2752"), val = tensor([1, 256, 16, 32, 2])]; + tensor k_complex_13_cast_fp16 = reshape(shape = var_2752, x = k_25_cast_fp16)[name = tensor("k_complex_13_cast_fp16")]; + tensor var_2756_begin_0 = const()[name = tensor("op_2756_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_2756_end_0 = const()[name = tensor("op_2756_end_0"), val = tensor([1, 256, 16, 32, 1])]; + tensor var_2756_end_mask_0 = const()[name = tensor("op_2756_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_2756_squeeze_mask_0 = const()[name = tensor("op_2756_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_2756_cast_fp16 = slice_by_index(begin = var_2756_begin_0, end = var_2756_end_0, end_mask = var_2756_end_mask_0, squeeze_mask = var_2756_squeeze_mask_0, x = q_complex_13_cast_fp16)[name = tensor("op_2756_cast_fp16")]; + tensor var_2764_begin_0 = const()[name = tensor("op_2764_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_2764_end_0 = const()[name = tensor("op_2764_end_0"), val = tensor([1, 256, 16, 32, 2])]; + tensor var_2764_end_mask_0 = const()[name = tensor("op_2764_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_2764_squeeze_mask_0 = const()[name = tensor("op_2764_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_2764_cast_fp16 = slice_by_index(begin = var_2764_begin_0, end = var_2764_end_0, end_mask = var_2764_end_mask_0, squeeze_mask = var_2764_squeeze_mask_0, x = q_complex_13_cast_fp16)[name = tensor("op_2764_cast_fp16")]; + tensor var_2772_begin_0 = const()[name = tensor("op_2772_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_2772_end_0 = const()[name = tensor("op_2772_end_0"), val = tensor([1, 256, 16, 32, 1])]; + tensor var_2772_end_mask_0 = const()[name = tensor("op_2772_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_2772_squeeze_mask_0 = const()[name = tensor("op_2772_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_2772_cast_fp16 = slice_by_index(begin = var_2772_begin_0, end = var_2772_end_0, end_mask = var_2772_end_mask_0, squeeze_mask = var_2772_squeeze_mask_0, x = k_complex_13_cast_fp16)[name = tensor("op_2772_cast_fp16")]; + tensor var_2780_begin_0 = const()[name = tensor("op_2780_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_2780_end_0 = const()[name = tensor("op_2780_end_0"), val = tensor([1, 256, 16, 32, 2])]; + tensor var_2780_end_mask_0 = const()[name = tensor("op_2780_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_2780_squeeze_mask_0 = const()[name = tensor("op_2780_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_2780_cast_fp16 = slice_by_index(begin = var_2780_begin_0, end = var_2780_end_0, end_mask = var_2780_end_mask_0, squeeze_mask = var_2780_squeeze_mask_0, x = k_complex_13_cast_fp16)[name = tensor("op_2780_cast_fp16")]; + tensor freqs_13_to_fp16 = const()[name = tensor("freqs_13_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(158411776)))]; + tensor var_2786_cast_fp16 = mul(x = freqs_13_to_fp16, y = ts_41_cast_fp16)[name = tensor("op_2786_cast_fp16")]; + tensor rotr_13_cast_fp16 = cos(x = var_2786_cast_fp16)[name = tensor("rotr_13_cast_fp16")]; + tensor roti_13_cast_fp16 = sin(x = var_2786_cast_fp16)[name = tensor("roti_13_cast_fp16")]; + tensor var_2790_cast_fp16 = mul(x = var_2756_cast_fp16, y = rotr_13_cast_fp16)[name = tensor("op_2790_cast_fp16")]; + tensor var_2791_cast_fp16 = mul(x = var_2764_cast_fp16, y = roti_13_cast_fp16)[name = tensor("op_2791_cast_fp16")]; + tensor qor_25_cast_fp16 = sub(x = var_2790_cast_fp16, y = var_2791_cast_fp16)[name = tensor("qor_25_cast_fp16")]; + tensor var_2794_cast_fp16 = mul(x = var_2756_cast_fp16, y = roti_13_cast_fp16)[name = tensor("op_2794_cast_fp16")]; + tensor var_2795_cast_fp16 = mul(x = var_2764_cast_fp16, y = rotr_13_cast_fp16)[name = tensor("op_2795_cast_fp16")]; + tensor qoi_25_cast_fp16 = add(x = var_2794_cast_fp16, y = var_2795_cast_fp16)[name = tensor("qoi_25_cast_fp16")]; + tensor var_2798_cast_fp16 = mul(x = var_2772_cast_fp16, y = rotr_13_cast_fp16)[name = tensor("op_2798_cast_fp16")]; + tensor var_2799_cast_fp16 = mul(x = var_2780_cast_fp16, y = roti_13_cast_fp16)[name = tensor("op_2799_cast_fp16")]; + tensor kor_25_cast_fp16 = sub(x = var_2798_cast_fp16, y = var_2799_cast_fp16)[name = tensor("kor_25_cast_fp16")]; + tensor var_2802_cast_fp16 = mul(x = var_2772_cast_fp16, y = roti_13_cast_fp16)[name = tensor("op_2802_cast_fp16")]; + tensor var_2803_cast_fp16 = mul(x = var_2780_cast_fp16, y = rotr_13_cast_fp16)[name = tensor("op_2803_cast_fp16")]; + tensor koi_25_cast_fp16 = add(x = var_2802_cast_fp16, y = var_2803_cast_fp16)[name = tensor("koi_25_cast_fp16")]; + tensor qo_13_axis_0 = const()[name = tensor("qo_13_axis_0"), val = tensor(-1)]; + tensor qo_13_cast_fp16 = stack(axis = qo_13_axis_0, values = (qor_25_cast_fp16, qoi_25_cast_fp16))[name = tensor("qo_13_cast_fp16")]; + tensor ko_13_axis_0 = const()[name = tensor("ko_13_axis_0"), val = tensor(-1)]; + tensor ko_13_cast_fp16 = stack(axis = ko_13_axis_0, values = (kor_25_cast_fp16, koi_25_cast_fp16))[name = tensor("ko_13_cast_fp16")]; + tensor var_2836 = const()[name = tensor("op_2836"), val = tensor([1, 256, 16, 64])]; + tensor q_39_cast_fp16 = reshape(shape = var_2836, x = qo_13_cast_fp16)[name = tensor("q_39_cast_fp16")]; + tensor var_2842 = const()[name = tensor("op_2842"), val = tensor([1, 256, 16, 64])]; + tensor k_27_cast_fp16 = reshape(shape = var_2842, x = ko_13_cast_fp16)[name = tensor("k_27_cast_fp16")]; + tensor write_indices_float_37_cast_fp16 = add(x = var_2731_cast_fp16, y = var_393_promoted_to_fp16)[name = tensor("write_indices_float_37_cast_fp16")]; + tensor _inversed_2863_y_0_to_fp16 = const()[name = tensor("_inversed_2863_y_0_to_fp16"), val = tensor(0x1p-9)]; + tensor _inversed_2863_cast_fp16 = mul(x = write_indices_float_37_cast_fp16, y = _inversed_2863_y_0_to_fp16)[name = tensor("_inversed_2863_cast_fp16")]; + tensor var_2864_cast_fp16 = floor(x = _inversed_2863_cast_fp16)[name = tensor("op_2864_cast_fp16")]; + tensor var_2865_to_fp16 = const()[name = tensor("op_2865_to_fp16"), val = tensor(0x1p+9)]; + tensor var_2866_cast_fp16 = mul(x = var_2864_cast_fp16, y = var_2865_to_fp16)[name = tensor("op_2866_cast_fp16")]; + tensor write_indices_float_39_cast_fp16 = sub(x = write_indices_float_37_cast_fp16, y = var_2866_cast_fp16)[name = tensor("write_indices_float_39_cast_fp16")]; + tensor var_2884_value_0_to_fp16 = const()[name = tensor("op_2884_value_0_to_fp16"), val = tensor(0x1.ffp+8)]; + tensor var_2884_cast_fp16 = fill_like(ref_tensor = write_indices_float_39_cast_fp16, value = var_2884_value_0_to_fp16)[name = tensor("op_2884_cast_fp16")]; + tensor write_indices_float_41_cast_fp16 = select(a = write_indices_float_39_cast_fp16, b = var_2884_cast_fp16, cond = token_valid_1_cast_fp16)[name = tensor("write_indices_float_41_cast_fp16")]; + tensor var_2890_dtype_0 = const()[name = tensor("op_2890_dtype_0"), val = tensor("int32")]; + tensor var_2895 = const()[name = tensor("op_2895"), val = tensor([1, 256, 1, 1])]; + tensor write_indices_float_41_cast_fp16_to_int32 = cast(dtype = var_2890_dtype_0, x = write_indices_float_41_cast_fp16)[name = tensor("cast_630")]; + tensor var_2896 = reshape(shape = var_2895, x = write_indices_float_41_cast_fp16_to_int32)[name = tensor("op_2896")]; + tensor write_indices_13_reps_0 = const()[name = tensor("write_indices_13_reps_0"), val = tensor([1, 1, 16, 64])]; + tensor write_indices_13 = tile(reps = write_indices_13_reps_0, x = var_2896)[name = tensor("write_indices_13")]; + tensor var_2906_begin_0 = const()[name = tensor("op_2906_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_2906_end_0 = const()[name = tensor("op_2906_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor var_2906_end_mask_0 = const()[name = tensor("op_2906_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_2906_squeeze_mask_0 = const()[name = tensor("op_2906_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor cache6_to_fp16_dtype_0 = const()[name = tensor("cache6_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor cache6_to_fp16 = cast(dtype = cache6_to_fp16_dtype_0, x = cache6)[name = tensor("cast_629")]; + tensor var_2906_cast_fp16 = slice_by_index(begin = var_2906_begin_0, end = var_2906_end_0, end_mask = var_2906_end_mask_0, squeeze_mask = var_2906_squeeze_mask_0, x = cache6_to_fp16)[name = tensor("op_2906_cast_fp16")]; + tensor var_2908_axis_0 = const()[name = tensor("op_2908_axis_0"), val = tensor(1)]; + tensor var_2908_mode_0 = const()[name = tensor("op_2908_mode_0"), val = tensor("update")]; + tensor var_2908_validate_indices_0 = const()[name = tensor("op_2908_validate_indices_0"), val = tensor(false)]; + tensor var_2908_cast_fp16 = scatter_along_axis(axis = var_2908_axis_0, data = var_2906_cast_fp16, indices = write_indices_13, mode = var_2908_mode_0, updates = k_27_cast_fp16, validate_indices = var_2908_validate_indices_0)[name = tensor("op_2908_cast_fp16")]; + tensor concat_50 = const()[name = tensor("concat_50"), val = tensor([0, 0, 0, 0, 0])]; + tensor concat_51 = const()[name = tensor("concat_51"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_13_internal_tensor_assign_1_stride_0 = const()[name = tensor("new_cache_13_internal_tensor_assign_1_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_13_internal_tensor_assign_1_begin_mask_0 = const()[name = tensor("new_cache_13_internal_tensor_assign_1_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_13_internal_tensor_assign_1_end_mask_0 = const()[name = tensor("new_cache_13_internal_tensor_assign_1_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_13_internal_tensor_assign_1_squeeze_mask_0 = const()[name = tensor("new_cache_13_internal_tensor_assign_1_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_58 = const()[name = tensor("shape_58"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_12 = const()[name = tensor("reduce_prod_12"), val = tensor(1048576)]; + tensor range_1d_12_start_0 = const()[name = tensor("range_1d_12_start_0"), val = tensor(0)]; + tensor range_1d_12_step_0 = const()[name = tensor("range_1d_12_step_0"), val = tensor(1)]; + tensor range_1d_12 = range_1d(end = reduce_prod_12, start = range_1d_12_start_0, step = range_1d_12_step_0)[name = tensor("range_1d_12")]; + tensor reshape_60 = reshape(shape = shape_58, x = range_1d_12)[name = tensor("reshape_60")]; + tensor slice_by_index_12 = slice_by_index(begin = concat_50, begin_mask = new_cache_13_internal_tensor_assign_1_begin_mask_0, end = concat_51, end_mask = new_cache_13_internal_tensor_assign_1_end_mask_0, squeeze_mask = new_cache_13_internal_tensor_assign_1_squeeze_mask_0, stride = new_cache_13_internal_tensor_assign_1_stride_0, x = reshape_60)[name = tensor("slice_by_index_12")]; + tensor reshape_61_shape_0 = const()[name = tensor("reshape_61_shape_0"), val = tensor([-1])]; + tensor reshape_61 = reshape(shape = reshape_61_shape_0, x = slice_by_index_12)[name = tensor("reshape_61")]; + tensor reshape_62_shape_0 = const()[name = tensor("reshape_62_shape_0"), val = tensor([-1])]; + tensor reshape_62_cast_fp16 = reshape(shape = reshape_62_shape_0, x = var_2908_cast_fp16)[name = tensor("reshape_62_cast_fp16")]; + tensor reshape_63_shape_0 = const()[name = tensor("reshape_63_shape_0"), val = tensor([-1])]; + tensor reshape_63_cast_fp16 = reshape(shape = reshape_63_shape_0, x = cache6_to_fp16)[name = tensor("reshape_63_cast_fp16")]; + tensor scatter_12_mode_0 = const()[name = tensor("scatter_12_mode_0"), val = tensor("update")]; + tensor scatter_12_axis_0 = const()[name = tensor("scatter_12_axis_0"), val = tensor(0)]; + tensor scatter_12_validate_indices_0 = const()[name = tensor("scatter_12_validate_indices_0"), val = tensor(false)]; + tensor scatter_12_cast_fp16 = scatter(axis = scatter_12_axis_0, data = reshape_63_cast_fp16, indices = reshape_61, mode = scatter_12_mode_0, updates = reshape_62_cast_fp16, validate_indices = scatter_12_validate_indices_0)[name = tensor("scatter_12_cast_fp16")]; + tensor reshape_64_cast_fp16 = reshape(shape = shape_58, x = scatter_12_cast_fp16)[name = tensor("reshape_64_cast_fp16")]; + tensor var_2916_begin_0 = const()[name = tensor("op_2916_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor var_2916_end_0 = const()[name = tensor("op_2916_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor var_2916_end_mask_0 = const()[name = tensor("op_2916_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_2916_squeeze_mask_0 = const()[name = tensor("op_2916_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor var_2916_cast_fp16 = slice_by_index(begin = var_2916_begin_0, end = var_2916_end_0, end_mask = var_2916_end_mask_0, squeeze_mask = var_2916_squeeze_mask_0, x = reshape_64_cast_fp16)[name = tensor("op_2916_cast_fp16")]; + tensor var_2918_axis_0 = const()[name = tensor("op_2918_axis_0"), val = tensor(1)]; + tensor var_2918_mode_0 = const()[name = tensor("op_2918_mode_0"), val = tensor("update")]; + tensor var_2918_validate_indices_0 = const()[name = tensor("op_2918_validate_indices_0"), val = tensor(false)]; + tensor var_2918_cast_fp16 = scatter_along_axis(axis = var_2918_axis_0, data = var_2916_cast_fp16, indices = write_indices_13, mode = var_2918_mode_0, updates = v_13_cast_fp16, validate_indices = var_2918_validate_indices_0)[name = tensor("op_2918_cast_fp16")]; + tensor concat_52 = const()[name = tensor("concat_52"), val = tensor([1, 0, 0, 0, 0])]; + tensor concat_53 = const()[name = tensor("concat_53"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_13_internal_tensor_assign_2_stride_0 = const()[name = tensor("new_cache_13_internal_tensor_assign_2_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_13_internal_tensor_assign_2_begin_mask_0 = const()[name = tensor("new_cache_13_internal_tensor_assign_2_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_13_internal_tensor_assign_2_end_mask_0 = const()[name = tensor("new_cache_13_internal_tensor_assign_2_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_13_internal_tensor_assign_2_squeeze_mask_0 = const()[name = tensor("new_cache_13_internal_tensor_assign_2_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_59 = const()[name = tensor("shape_59"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_13 = const()[name = tensor("reduce_prod_13"), val = tensor(1048576)]; + tensor range_1d_13_start_0 = const()[name = tensor("range_1d_13_start_0"), val = tensor(0)]; + tensor range_1d_13_step_0 = const()[name = tensor("range_1d_13_step_0"), val = tensor(1)]; + tensor range_1d_13 = range_1d(end = reduce_prod_13, start = range_1d_13_start_0, step = range_1d_13_step_0)[name = tensor("range_1d_13")]; + tensor reshape_65 = reshape(shape = shape_59, x = range_1d_13)[name = tensor("reshape_65")]; + tensor slice_by_index_13 = slice_by_index(begin = concat_52, begin_mask = new_cache_13_internal_tensor_assign_2_begin_mask_0, end = concat_53, end_mask = new_cache_13_internal_tensor_assign_2_end_mask_0, squeeze_mask = new_cache_13_internal_tensor_assign_2_squeeze_mask_0, stride = new_cache_13_internal_tensor_assign_2_stride_0, x = reshape_65)[name = tensor("slice_by_index_13")]; + tensor reshape_66_shape_0 = const()[name = tensor("reshape_66_shape_0"), val = tensor([-1])]; + tensor reshape_66 = reshape(shape = reshape_66_shape_0, x = slice_by_index_13)[name = tensor("reshape_66")]; + tensor reshape_67_shape_0 = const()[name = tensor("reshape_67_shape_0"), val = tensor([-1])]; + tensor reshape_67_cast_fp16 = reshape(shape = reshape_67_shape_0, x = var_2918_cast_fp16)[name = tensor("reshape_67_cast_fp16")]; + tensor reshape_68_shape_0 = const()[name = tensor("reshape_68_shape_0"), val = tensor([-1])]; + tensor reshape_68_cast_fp16 = reshape(shape = reshape_68_shape_0, x = reshape_64_cast_fp16)[name = tensor("reshape_68_cast_fp16")]; + tensor scatter_13_mode_0 = const()[name = tensor("scatter_13_mode_0"), val = tensor("update")]; + tensor scatter_13_axis_0 = const()[name = tensor("scatter_13_axis_0"), val = tensor(0)]; + tensor scatter_13_validate_indices_0 = const()[name = tensor("scatter_13_validate_indices_0"), val = tensor(false)]; + tensor scatter_13_cast_fp16 = scatter(axis = scatter_13_axis_0, data = reshape_68_cast_fp16, indices = reshape_66, mode = scatter_13_mode_0, updates = reshape_67_cast_fp16, validate_indices = scatter_13_validate_indices_0)[name = tensor("scatter_13_cast_fp16")]; + tensor reshape_69_cast_fp16 = reshape(shape = shape_59, x = scatter_13_cast_fp16)[name = tensor("reshape_69_cast_fp16")]; + tensor reshape_69_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("reshape_69_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor keys_37_begin_0 = const()[name = tensor("keys_37_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor keys_37_end_0 = const()[name = tensor("keys_37_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor keys_37_end_mask_0 = const()[name = tensor("keys_37_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor keys_37_squeeze_mask_0 = const()[name = tensor("keys_37_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor keys_37_cast_fp16 = slice_by_index(begin = keys_37_begin_0, end = keys_37_end_0, end_mask = keys_37_end_mask_0, squeeze_mask = keys_37_squeeze_mask_0, x = reshape_69_cast_fp16)[name = tensor("keys_37_cast_fp16")]; + tensor values_37_begin_0 = const()[name = tensor("values_37_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor values_37_end_0 = const()[name = tensor("values_37_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor values_37_end_mask_0 = const()[name = tensor("values_37_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor values_37_squeeze_mask_0 = const()[name = tensor("values_37_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor values_37_cast_fp16 = slice_by_index(begin = values_37_begin_0, end = values_37_end_0, end_mask = values_37_end_mask_0, squeeze_mask = values_37_squeeze_mask_0, x = reshape_69_cast_fp16)[name = tensor("values_37_cast_fp16")]; + tensor var_2930_cast_fp16 = not_equal(x = keys_37_cast_fp16, y = keys_37_cast_fp16)[name = tensor("op_2930_cast_fp16")]; + tensor keys_39_cast_fp16 = select(a = var_470_to_fp16, b = keys_37_cast_fp16, cond = var_2930_cast_fp16)[name = tensor("keys_39_cast_fp16")]; + tensor var_2938_cast_fp16 = not_equal(x = values_37_cast_fp16, y = values_37_cast_fp16)[name = tensor("op_2938_cast_fp16")]; + tensor values_39_cast_fp16 = select(a = var_470_to_fp16, b = values_37_cast_fp16, cond = var_2938_cast_fp16)[name = tensor("values_39_cast_fp16")]; + tensor var_2962 = const()[name = tensor("op_2962"), val = tensor([0, 2, 1, 3])]; + tensor var_2978 = const()[name = tensor("op_2978"), val = tensor([1, 1, 1])]; + tensor var_2979_cast_fp16 = reshape(shape = var_2978, x = position6_to_fp16)[name = tensor("op_2979_cast_fp16")]; + tensor q_positions_13_cast_fp16 = add(x = var_2979_cast_fp16, y = q_offsets_1_promoted_to_fp16)[name = tensor("q_positions_13_cast_fp16")]; + tensor valid_end_13_cast_fp16 = add(x = var_2979_cast_fp16, y = var_536_cast_fp16)[name = tensor("valid_end_13_cast_fp16")]; + tensor valid_mask_13_cast_fp16 = less(x = k_positions_1_promoted_to_fp16, y = valid_end_13_cast_fp16)[name = tensor("valid_mask_13_cast_fp16")]; + tensor causal_mask_13_cast_fp16 = less_equal(x = k_positions_1_promoted_to_fp16, y = q_positions_13_cast_fp16)[name = tensor("causal_mask_13_cast_fp16")]; + tensor var_3007 = logical_and(x = valid_mask_13_cast_fp16, y = causal_mask_13_cast_fp16)[name = tensor("op_3007")]; + tensor attn_mask_13_axes_0 = const()[name = tensor("attn_mask_13_axes_0"), val = tensor([1])]; + tensor attn_mask_13 = expand_dims(axes = attn_mask_13_axes_0, x = var_3007)[name = tensor("attn_mask_13")]; + tensor var_3021_transpose_x_0 = const()[name = tensor("op_3021_transpose_x_0"), val = tensor(false)]; + tensor var_3021_transpose_y_0 = const()[name = tensor("op_3021_transpose_y_0"), val = tensor(false)]; + tensor transpose_81_perm_0 = const()[name = tensor("transpose_81_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_82_perm_0 = const()[name = tensor("transpose_82_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_82 = transpose(perm = transpose_82_perm_0, x = keys_39_cast_fp16)[name = tensor("transpose_180")]; + tensor transpose_81 = transpose(perm = transpose_81_perm_0, x = q_39_cast_fp16)[name = tensor("transpose_181")]; + tensor var_3021_cast_fp16 = matmul(transpose_x = var_3021_transpose_x_0, transpose_y = var_3021_transpose_y_0, x = transpose_81, y = transpose_82)[name = tensor("op_3021_cast_fp16")]; + tensor var_3015_to_fp16 = const()[name = tensor("op_3015_to_fp16"), val = tensor([0x1p-3])]; + tensor attn_weights_37_cast_fp16 = mul(x = var_3021_cast_fp16, y = var_3015_to_fp16)[name = tensor("attn_weights_37_cast_fp16")]; + tensor var_3023 = logical_not(x = attn_mask_13)[name = tensor("op_3023")]; + tensor var_3024_to_fp16 = const()[name = tensor("op_3024_to_fp16"), val = tensor(-inf)]; + tensor attn_weights_39_cast_fp16 = select(a = var_3024_to_fp16, b = attn_weights_37_cast_fp16, cond = var_3023)[name = tensor("attn_weights_39_cast_fp16")]; + tensor var_3026 = const()[name = tensor("op_3026"), val = tensor(-1)]; + tensor attn_weights_41_cast_fp16 = softmax(axis = var_3026, x = attn_weights_39_cast_fp16)[name = tensor("attn_weights_41_cast_fp16")]; + tensor attn_output_13_transpose_x_0 = const()[name = tensor("attn_output_13_transpose_x_0"), val = tensor(false)]; + tensor attn_output_13_transpose_y_0 = const()[name = tensor("attn_output_13_transpose_y_0"), val = tensor(false)]; + tensor values_41_cast_fp16 = transpose(perm = var_2962, x = values_39_cast_fp16)[name = tensor("transpose_182")]; + tensor attn_output_13_cast_fp16 = matmul(transpose_x = attn_output_13_transpose_x_0, transpose_y = attn_output_13_transpose_y_0, x = attn_weights_41_cast_fp16, y = values_41_cast_fp16)[name = tensor("attn_output_13_cast_fp16")]; + tensor var_3034 = const()[name = tensor("op_3034"), val = tensor([0, 2, 1, 3])]; + tensor var_3039 = const()[name = tensor("op_3039"), val = tensor([1, 256, 1024])]; + tensor var_3035_cast_fp16 = transpose(perm = var_3034, x = attn_output_13_cast_fp16)[name = tensor("transpose_179")]; + tensor input_77_cast_fp16 = reshape(shape = var_3039, x = var_3035_cast_fp16)[name = tensor("input_77_cast_fp16")]; + tensor attn6_out_proj_weight_to_fp16 = const()[name = tensor("attn6_out_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(158411904)))]; + tensor linear_25_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = attn6_out_proj_weight_to_fp16, x = input_77_cast_fp16)[name = tensor("linear_25_cast_fp16")]; + tensor var_3048_cast_fp16 = add(x = position6_to_fp16, y = valid_len_to_fp16)[name = tensor("op_3048_cast_fp16")]; + tensor var_3048_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_3048_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor input_79_cast_fp16 = add(x = input_73_cast_fp16, y = linear_25_cast_fp16)[name = tensor("input_79_cast_fp16")]; + tensor input_81_axes_0 = const()[name = tensor("input_81_axes_0"), val = tensor([-1])]; + tensor norm6_2_weight_to_fp16 = const()[name = tensor("norm6_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(160509120)))]; + tensor norm6_2_bias_to_fp16 = const()[name = tensor("norm6_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(160511232)))]; + tensor var_3052_to_fp16 = const()[name = tensor("op_3052_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_81_cast_fp16 = layer_norm(axes = input_81_axes_0, beta = norm6_2_bias_to_fp16, epsilon = var_3052_to_fp16, gamma = norm6_2_weight_to_fp16, x = input_79_cast_fp16)[name = tensor("input_81_cast_fp16")]; + tensor linear6_1_weight_to_fp16 = const()[name = tensor("linear6_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(160513344)))]; + tensor linear_26_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = linear6_1_weight_to_fp16, x = input_81_cast_fp16)[name = tensor("linear_26_cast_fp16")]; + tensor input_83_mode_0 = const()[name = tensor("input_83_mode_0"), val = tensor("EXACT")]; + tensor input_83_cast_fp16 = gelu(mode = input_83_mode_0, x = linear_26_cast_fp16)[name = tensor("input_83_cast_fp16")]; + tensor linear6_2_weight_to_fp16 = const()[name = tensor("linear6_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(168902016)))]; + tensor linear_27_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = linear6_2_weight_to_fp16, x = input_83_cast_fp16)[name = tensor("linear_27_cast_fp16")]; + tensor input_85_cast_fp16 = add(x = input_79_cast_fp16, y = linear_27_cast_fp16)[name = tensor("input_85_cast_fp16")]; + tensor input_87_axes_0 = const()[name = tensor("input_87_axes_0"), val = tensor([-1])]; + tensor norm7_1_weight_to_fp16 = const()[name = tensor("norm7_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(177290688)))]; + tensor norm7_1_bias_to_fp16 = const()[name = tensor("norm7_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(177292800)))]; + tensor var_3069_to_fp16 = const()[name = tensor("op_3069_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_87_cast_fp16 = layer_norm(axes = input_87_axes_0, beta = norm7_1_bias_to_fp16, epsilon = var_3069_to_fp16, gamma = norm7_1_weight_to_fp16, x = input_85_cast_fp16)[name = tensor("input_87_cast_fp16")]; + tensor attn7_in_proj_weight_to_fp16 = const()[name = tensor("attn7_in_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(177294912)))]; + tensor linear_28_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = attn7_in_proj_weight_to_fp16, x = input_87_cast_fp16)[name = tensor("linear_28_cast_fp16")]; + tensor var_3083 = const()[name = tensor("op_3083"), val = tensor([1, 256, 3, 16, 64])]; + tensor qkv_15_cast_fp16 = reshape(shape = var_3083, x = linear_28_cast_fp16)[name = tensor("qkv_15_cast_fp16")]; + tensor q_43_begin_0 = const()[name = tensor("q_43_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor q_43_end_0 = const()[name = tensor("q_43_end_0"), val = tensor([1, 256, 1, 16, 64])]; + tensor q_43_end_mask_0 = const()[name = tensor("q_43_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor q_43_squeeze_mask_0 = const()[name = tensor("q_43_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor q_43_cast_fp16 = slice_by_index(begin = q_43_begin_0, end = q_43_end_0, end_mask = q_43_end_mask_0, squeeze_mask = q_43_squeeze_mask_0, x = qkv_15_cast_fp16)[name = tensor("q_43_cast_fp16")]; + tensor k_29_begin_0 = const()[name = tensor("k_29_begin_0"), val = tensor([0, 0, 1, 0, 0])]; + tensor k_29_end_0 = const()[name = tensor("k_29_end_0"), val = tensor([1, 256, 2, 16, 64])]; + tensor k_29_end_mask_0 = const()[name = tensor("k_29_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor k_29_squeeze_mask_0 = const()[name = tensor("k_29_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor k_29_cast_fp16 = slice_by_index(begin = k_29_begin_0, end = k_29_end_0, end_mask = k_29_end_mask_0, squeeze_mask = k_29_squeeze_mask_0, x = qkv_15_cast_fp16)[name = tensor("k_29_cast_fp16")]; + tensor v_15_begin_0 = const()[name = tensor("v_15_begin_0"), val = tensor([0, 0, 2, 0, 0])]; + tensor v_15_end_0 = const()[name = tensor("v_15_end_0"), val = tensor([1, 256, 3, 16, 64])]; + tensor v_15_end_mask_0 = const()[name = tensor("v_15_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor v_15_squeeze_mask_0 = const()[name = tensor("v_15_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor v_15_cast_fp16 = slice_by_index(begin = v_15_begin_0, end = v_15_end_0, end_mask = v_15_end_mask_0, squeeze_mask = v_15_squeeze_mask_0, x = qkv_15_cast_fp16)[name = tensor("v_15_cast_fp16")]; + tensor var_3141 = const()[name = tensor("op_3141"), val = tensor([1, 1])]; + tensor position7_to_fp16_dtype_0 = const()[name = tensor("position7_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor position7_to_fp16 = cast(dtype = position7_to_fp16_dtype_0, x = position7)[name = tensor("cast_626")]; + tensor var_3142_cast_fp16 = reshape(shape = var_3141, x = position7_to_fp16)[name = tensor("op_3142_cast_fp16")]; + tensor ts_45_cast_fp16 = add(x = ts_1_promoted_to_fp16, y = var_3142_cast_fp16)[name = tensor("ts_45_cast_fp16")]; + tensor var_3149 = const()[name = tensor("op_3149"), val = tensor([1, 256, 1, 1])]; + tensor ts_47_cast_fp16 = reshape(shape = var_3149, x = ts_45_cast_fp16)[name = tensor("ts_47_cast_fp16")]; + tensor var_3156 = const()[name = tensor("op_3156"), val = tensor([1, 256, 16, 32, 2])]; + tensor q_complex_15_cast_fp16 = reshape(shape = var_3156, x = q_43_cast_fp16)[name = tensor("q_complex_15_cast_fp16")]; + tensor var_3163 = const()[name = tensor("op_3163"), val = tensor([1, 256, 16, 32, 2])]; + tensor k_complex_15_cast_fp16 = reshape(shape = var_3163, x = k_29_cast_fp16)[name = tensor("k_complex_15_cast_fp16")]; + tensor var_3167_begin_0 = const()[name = tensor("op_3167_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_3167_end_0 = const()[name = tensor("op_3167_end_0"), val = tensor([1, 256, 16, 32, 1])]; + tensor var_3167_end_mask_0 = const()[name = tensor("op_3167_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_3167_squeeze_mask_0 = const()[name = tensor("op_3167_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_3167_cast_fp16 = slice_by_index(begin = var_3167_begin_0, end = var_3167_end_0, end_mask = var_3167_end_mask_0, squeeze_mask = var_3167_squeeze_mask_0, x = q_complex_15_cast_fp16)[name = tensor("op_3167_cast_fp16")]; + tensor var_3175_begin_0 = const()[name = tensor("op_3175_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_3175_end_0 = const()[name = tensor("op_3175_end_0"), val = tensor([1, 256, 16, 32, 2])]; + tensor var_3175_end_mask_0 = const()[name = tensor("op_3175_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_3175_squeeze_mask_0 = const()[name = tensor("op_3175_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_3175_cast_fp16 = slice_by_index(begin = var_3175_begin_0, end = var_3175_end_0, end_mask = var_3175_end_mask_0, squeeze_mask = var_3175_squeeze_mask_0, x = q_complex_15_cast_fp16)[name = tensor("op_3175_cast_fp16")]; + tensor var_3183_begin_0 = const()[name = tensor("op_3183_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_3183_end_0 = const()[name = tensor("op_3183_end_0"), val = tensor([1, 256, 16, 32, 1])]; + tensor var_3183_end_mask_0 = const()[name = tensor("op_3183_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_3183_squeeze_mask_0 = const()[name = tensor("op_3183_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_3183_cast_fp16 = slice_by_index(begin = var_3183_begin_0, end = var_3183_end_0, end_mask = var_3183_end_mask_0, squeeze_mask = var_3183_squeeze_mask_0, x = k_complex_15_cast_fp16)[name = tensor("op_3183_cast_fp16")]; + tensor var_3191_begin_0 = const()[name = tensor("op_3191_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_3191_end_0 = const()[name = tensor("op_3191_end_0"), val = tensor([1, 256, 16, 32, 2])]; + tensor var_3191_end_mask_0 = const()[name = tensor("op_3191_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_3191_squeeze_mask_0 = const()[name = tensor("op_3191_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_3191_cast_fp16 = slice_by_index(begin = var_3191_begin_0, end = var_3191_end_0, end_mask = var_3191_end_mask_0, squeeze_mask = var_3191_squeeze_mask_0, x = k_complex_15_cast_fp16)[name = tensor("op_3191_cast_fp16")]; + tensor freqs_15_to_fp16 = const()[name = tensor("freqs_15_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(183586432)))]; + tensor var_3197_cast_fp16 = mul(x = freqs_15_to_fp16, y = ts_47_cast_fp16)[name = tensor("op_3197_cast_fp16")]; + tensor rotr_15_cast_fp16 = cos(x = var_3197_cast_fp16)[name = tensor("rotr_15_cast_fp16")]; + tensor roti_15_cast_fp16 = sin(x = var_3197_cast_fp16)[name = tensor("roti_15_cast_fp16")]; + tensor var_3201_cast_fp16 = mul(x = var_3167_cast_fp16, y = rotr_15_cast_fp16)[name = tensor("op_3201_cast_fp16")]; + tensor var_3202_cast_fp16 = mul(x = var_3175_cast_fp16, y = roti_15_cast_fp16)[name = tensor("op_3202_cast_fp16")]; + tensor qor_29_cast_fp16 = sub(x = var_3201_cast_fp16, y = var_3202_cast_fp16)[name = tensor("qor_29_cast_fp16")]; + tensor var_3205_cast_fp16 = mul(x = var_3167_cast_fp16, y = roti_15_cast_fp16)[name = tensor("op_3205_cast_fp16")]; + tensor var_3206_cast_fp16 = mul(x = var_3175_cast_fp16, y = rotr_15_cast_fp16)[name = tensor("op_3206_cast_fp16")]; + tensor qoi_29_cast_fp16 = add(x = var_3205_cast_fp16, y = var_3206_cast_fp16)[name = tensor("qoi_29_cast_fp16")]; + tensor var_3209_cast_fp16 = mul(x = var_3183_cast_fp16, y = rotr_15_cast_fp16)[name = tensor("op_3209_cast_fp16")]; + tensor var_3210_cast_fp16 = mul(x = var_3191_cast_fp16, y = roti_15_cast_fp16)[name = tensor("op_3210_cast_fp16")]; + tensor kor_29_cast_fp16 = sub(x = var_3209_cast_fp16, y = var_3210_cast_fp16)[name = tensor("kor_29_cast_fp16")]; + tensor var_3213_cast_fp16 = mul(x = var_3183_cast_fp16, y = roti_15_cast_fp16)[name = tensor("op_3213_cast_fp16")]; + tensor var_3214_cast_fp16 = mul(x = var_3191_cast_fp16, y = rotr_15_cast_fp16)[name = tensor("op_3214_cast_fp16")]; + tensor koi_29_cast_fp16 = add(x = var_3213_cast_fp16, y = var_3214_cast_fp16)[name = tensor("koi_29_cast_fp16")]; + tensor qo_15_axis_0 = const()[name = tensor("qo_15_axis_0"), val = tensor(-1)]; + tensor qo_15_cast_fp16 = stack(axis = qo_15_axis_0, values = (qor_29_cast_fp16, qoi_29_cast_fp16))[name = tensor("qo_15_cast_fp16")]; + tensor ko_15_axis_0 = const()[name = tensor("ko_15_axis_0"), val = tensor(-1)]; + tensor ko_15_cast_fp16 = stack(axis = ko_15_axis_0, values = (kor_29_cast_fp16, koi_29_cast_fp16))[name = tensor("ko_15_cast_fp16")]; + tensor var_3247 = const()[name = tensor("op_3247"), val = tensor([1, 256, 16, 64])]; + tensor q_45_cast_fp16 = reshape(shape = var_3247, x = qo_15_cast_fp16)[name = tensor("q_45_cast_fp16")]; + tensor var_3253 = const()[name = tensor("op_3253"), val = tensor([1, 256, 16, 64])]; + tensor k_31_cast_fp16 = reshape(shape = var_3253, x = ko_15_cast_fp16)[name = tensor("k_31_cast_fp16")]; + tensor write_indices_float_43_cast_fp16 = add(x = var_3142_cast_fp16, y = var_393_promoted_to_fp16)[name = tensor("write_indices_float_43_cast_fp16")]; + tensor _inversed_3274_y_0_to_fp16 = const()[name = tensor("_inversed_3274_y_0_to_fp16"), val = tensor(0x1p-9)]; + tensor _inversed_3274_cast_fp16 = mul(x = write_indices_float_43_cast_fp16, y = _inversed_3274_y_0_to_fp16)[name = tensor("_inversed_3274_cast_fp16")]; + tensor var_3275_cast_fp16 = floor(x = _inversed_3274_cast_fp16)[name = tensor("op_3275_cast_fp16")]; + tensor var_3276_to_fp16 = const()[name = tensor("op_3276_to_fp16"), val = tensor(0x1p+9)]; + tensor var_3277_cast_fp16 = mul(x = var_3275_cast_fp16, y = var_3276_to_fp16)[name = tensor("op_3277_cast_fp16")]; + tensor write_indices_float_45_cast_fp16 = sub(x = write_indices_float_43_cast_fp16, y = var_3277_cast_fp16)[name = tensor("write_indices_float_45_cast_fp16")]; + tensor var_3295_value_0_to_fp16 = const()[name = tensor("op_3295_value_0_to_fp16"), val = tensor(0x1.ffp+8)]; + tensor var_3295_cast_fp16 = fill_like(ref_tensor = write_indices_float_45_cast_fp16, value = var_3295_value_0_to_fp16)[name = tensor("op_3295_cast_fp16")]; + tensor write_indices_float_47_cast_fp16 = select(a = write_indices_float_45_cast_fp16, b = var_3295_cast_fp16, cond = token_valid_1_cast_fp16)[name = tensor("write_indices_float_47_cast_fp16")]; + tensor var_3301_dtype_0 = const()[name = tensor("op_3301_dtype_0"), val = tensor("int32")]; + tensor var_3306 = const()[name = tensor("op_3306"), val = tensor([1, 256, 1, 1])]; + tensor write_indices_float_47_cast_fp16_to_int32 = cast(dtype = var_3301_dtype_0, x = write_indices_float_47_cast_fp16)[name = tensor("cast_625")]; + tensor var_3307 = reshape(shape = var_3306, x = write_indices_float_47_cast_fp16_to_int32)[name = tensor("op_3307")]; + tensor write_indices_15_reps_0 = const()[name = tensor("write_indices_15_reps_0"), val = tensor([1, 1, 16, 64])]; + tensor write_indices_15 = tile(reps = write_indices_15_reps_0, x = var_3307)[name = tensor("write_indices_15")]; + tensor var_3317_begin_0 = const()[name = tensor("op_3317_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_3317_end_0 = const()[name = tensor("op_3317_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor var_3317_end_mask_0 = const()[name = tensor("op_3317_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_3317_squeeze_mask_0 = const()[name = tensor("op_3317_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor cache7_to_fp16_dtype_0 = const()[name = tensor("cache7_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor cache7_to_fp16 = cast(dtype = cache7_to_fp16_dtype_0, x = cache7)[name = tensor("cast_624")]; + tensor var_3317_cast_fp16 = slice_by_index(begin = var_3317_begin_0, end = var_3317_end_0, end_mask = var_3317_end_mask_0, squeeze_mask = var_3317_squeeze_mask_0, x = cache7_to_fp16)[name = tensor("op_3317_cast_fp16")]; + tensor var_3319_axis_0 = const()[name = tensor("op_3319_axis_0"), val = tensor(1)]; + tensor var_3319_mode_0 = const()[name = tensor("op_3319_mode_0"), val = tensor("update")]; + tensor var_3319_validate_indices_0 = const()[name = tensor("op_3319_validate_indices_0"), val = tensor(false)]; + tensor var_3319_cast_fp16 = scatter_along_axis(axis = var_3319_axis_0, data = var_3317_cast_fp16, indices = write_indices_15, mode = var_3319_mode_0, updates = k_31_cast_fp16, validate_indices = var_3319_validate_indices_0)[name = tensor("op_3319_cast_fp16")]; + tensor concat_58 = const()[name = tensor("concat_58"), val = tensor([0, 0, 0, 0, 0])]; + tensor concat_59 = const()[name = tensor("concat_59"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_15_internal_tensor_assign_1_stride_0 = const()[name = tensor("new_cache_15_internal_tensor_assign_1_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_15_internal_tensor_assign_1_begin_mask_0 = const()[name = tensor("new_cache_15_internal_tensor_assign_1_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_15_internal_tensor_assign_1_end_mask_0 = const()[name = tensor("new_cache_15_internal_tensor_assign_1_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_15_internal_tensor_assign_1_squeeze_mask_0 = const()[name = tensor("new_cache_15_internal_tensor_assign_1_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_60 = const()[name = tensor("shape_60"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_14 = const()[name = tensor("reduce_prod_14"), val = tensor(1048576)]; + tensor range_1d_14_start_0 = const()[name = tensor("range_1d_14_start_0"), val = tensor(0)]; + tensor range_1d_14_step_0 = const()[name = tensor("range_1d_14_step_0"), val = tensor(1)]; + tensor range_1d_14 = range_1d(end = reduce_prod_14, start = range_1d_14_start_0, step = range_1d_14_step_0)[name = tensor("range_1d_14")]; + tensor reshape_70 = reshape(shape = shape_60, x = range_1d_14)[name = tensor("reshape_70")]; + tensor slice_by_index_14 = slice_by_index(begin = concat_58, begin_mask = new_cache_15_internal_tensor_assign_1_begin_mask_0, end = concat_59, end_mask = new_cache_15_internal_tensor_assign_1_end_mask_0, squeeze_mask = new_cache_15_internal_tensor_assign_1_squeeze_mask_0, stride = new_cache_15_internal_tensor_assign_1_stride_0, x = reshape_70)[name = tensor("slice_by_index_14")]; + tensor reshape_71_shape_0 = const()[name = tensor("reshape_71_shape_0"), val = tensor([-1])]; + tensor reshape_71 = reshape(shape = reshape_71_shape_0, x = slice_by_index_14)[name = tensor("reshape_71")]; + tensor reshape_72_shape_0 = const()[name = tensor("reshape_72_shape_0"), val = tensor([-1])]; + tensor reshape_72_cast_fp16 = reshape(shape = reshape_72_shape_0, x = var_3319_cast_fp16)[name = tensor("reshape_72_cast_fp16")]; + tensor reshape_73_shape_0 = const()[name = tensor("reshape_73_shape_0"), val = tensor([-1])]; + tensor reshape_73_cast_fp16 = reshape(shape = reshape_73_shape_0, x = cache7_to_fp16)[name = tensor("reshape_73_cast_fp16")]; + tensor scatter_14_mode_0 = const()[name = tensor("scatter_14_mode_0"), val = tensor("update")]; + tensor scatter_14_axis_0 = const()[name = tensor("scatter_14_axis_0"), val = tensor(0)]; + tensor scatter_14_validate_indices_0 = const()[name = tensor("scatter_14_validate_indices_0"), val = tensor(false)]; + tensor scatter_14_cast_fp16 = scatter(axis = scatter_14_axis_0, data = reshape_73_cast_fp16, indices = reshape_71, mode = scatter_14_mode_0, updates = reshape_72_cast_fp16, validate_indices = scatter_14_validate_indices_0)[name = tensor("scatter_14_cast_fp16")]; + tensor reshape_74_cast_fp16 = reshape(shape = shape_60, x = scatter_14_cast_fp16)[name = tensor("reshape_74_cast_fp16")]; + tensor var_3327_begin_0 = const()[name = tensor("op_3327_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor var_3327_end_0 = const()[name = tensor("op_3327_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor var_3327_end_mask_0 = const()[name = tensor("op_3327_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_3327_squeeze_mask_0 = const()[name = tensor("op_3327_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor var_3327_cast_fp16 = slice_by_index(begin = var_3327_begin_0, end = var_3327_end_0, end_mask = var_3327_end_mask_0, squeeze_mask = var_3327_squeeze_mask_0, x = reshape_74_cast_fp16)[name = tensor("op_3327_cast_fp16")]; + tensor var_3329_axis_0 = const()[name = tensor("op_3329_axis_0"), val = tensor(1)]; + tensor var_3329_mode_0 = const()[name = tensor("op_3329_mode_0"), val = tensor("update")]; + tensor var_3329_validate_indices_0 = const()[name = tensor("op_3329_validate_indices_0"), val = tensor(false)]; + tensor var_3329_cast_fp16 = scatter_along_axis(axis = var_3329_axis_0, data = var_3327_cast_fp16, indices = write_indices_15, mode = var_3329_mode_0, updates = v_15_cast_fp16, validate_indices = var_3329_validate_indices_0)[name = tensor("op_3329_cast_fp16")]; + tensor concat_60 = const()[name = tensor("concat_60"), val = tensor([1, 0, 0, 0, 0])]; + tensor concat_61 = const()[name = tensor("concat_61"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_15_internal_tensor_assign_2_stride_0 = const()[name = tensor("new_cache_15_internal_tensor_assign_2_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_15_internal_tensor_assign_2_begin_mask_0 = const()[name = tensor("new_cache_15_internal_tensor_assign_2_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_15_internal_tensor_assign_2_end_mask_0 = const()[name = tensor("new_cache_15_internal_tensor_assign_2_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_15_internal_tensor_assign_2_squeeze_mask_0 = const()[name = tensor("new_cache_15_internal_tensor_assign_2_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_61 = const()[name = tensor("shape_61"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_15 = const()[name = tensor("reduce_prod_15"), val = tensor(1048576)]; + tensor range_1d_15_start_0 = const()[name = tensor("range_1d_15_start_0"), val = tensor(0)]; + tensor range_1d_15_step_0 = const()[name = tensor("range_1d_15_step_0"), val = tensor(1)]; + tensor range_1d_15 = range_1d(end = reduce_prod_15, start = range_1d_15_start_0, step = range_1d_15_step_0)[name = tensor("range_1d_15")]; + tensor reshape_75 = reshape(shape = shape_61, x = range_1d_15)[name = tensor("reshape_75")]; + tensor slice_by_index_15 = slice_by_index(begin = concat_60, begin_mask = new_cache_15_internal_tensor_assign_2_begin_mask_0, end = concat_61, end_mask = new_cache_15_internal_tensor_assign_2_end_mask_0, squeeze_mask = new_cache_15_internal_tensor_assign_2_squeeze_mask_0, stride = new_cache_15_internal_tensor_assign_2_stride_0, x = reshape_75)[name = tensor("slice_by_index_15")]; + tensor reshape_76_shape_0 = const()[name = tensor("reshape_76_shape_0"), val = tensor([-1])]; + tensor reshape_76 = reshape(shape = reshape_76_shape_0, x = slice_by_index_15)[name = tensor("reshape_76")]; + tensor reshape_77_shape_0 = const()[name = tensor("reshape_77_shape_0"), val = tensor([-1])]; + tensor reshape_77_cast_fp16 = reshape(shape = reshape_77_shape_0, x = var_3329_cast_fp16)[name = tensor("reshape_77_cast_fp16")]; + tensor reshape_78_shape_0 = const()[name = tensor("reshape_78_shape_0"), val = tensor([-1])]; + tensor reshape_78_cast_fp16 = reshape(shape = reshape_78_shape_0, x = reshape_74_cast_fp16)[name = tensor("reshape_78_cast_fp16")]; + tensor scatter_15_mode_0 = const()[name = tensor("scatter_15_mode_0"), val = tensor("update")]; + tensor scatter_15_axis_0 = const()[name = tensor("scatter_15_axis_0"), val = tensor(0)]; + tensor scatter_15_validate_indices_0 = const()[name = tensor("scatter_15_validate_indices_0"), val = tensor(false)]; + tensor scatter_15_cast_fp16 = scatter(axis = scatter_15_axis_0, data = reshape_78_cast_fp16, indices = reshape_76, mode = scatter_15_mode_0, updates = reshape_77_cast_fp16, validate_indices = scatter_15_validate_indices_0)[name = tensor("scatter_15_cast_fp16")]; + tensor reshape_79_cast_fp16 = reshape(shape = shape_61, x = scatter_15_cast_fp16)[name = tensor("reshape_79_cast_fp16")]; + tensor reshape_79_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("reshape_79_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor keys_43_begin_0 = const()[name = tensor("keys_43_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor keys_43_end_0 = const()[name = tensor("keys_43_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor keys_43_end_mask_0 = const()[name = tensor("keys_43_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor keys_43_squeeze_mask_0 = const()[name = tensor("keys_43_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor keys_43_cast_fp16 = slice_by_index(begin = keys_43_begin_0, end = keys_43_end_0, end_mask = keys_43_end_mask_0, squeeze_mask = keys_43_squeeze_mask_0, x = reshape_79_cast_fp16)[name = tensor("keys_43_cast_fp16")]; + tensor values_43_begin_0 = const()[name = tensor("values_43_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor values_43_end_0 = const()[name = tensor("values_43_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor values_43_end_mask_0 = const()[name = tensor("values_43_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor values_43_squeeze_mask_0 = const()[name = tensor("values_43_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor values_43_cast_fp16 = slice_by_index(begin = values_43_begin_0, end = values_43_end_0, end_mask = values_43_end_mask_0, squeeze_mask = values_43_squeeze_mask_0, x = reshape_79_cast_fp16)[name = tensor("values_43_cast_fp16")]; + tensor var_3341_cast_fp16 = not_equal(x = keys_43_cast_fp16, y = keys_43_cast_fp16)[name = tensor("op_3341_cast_fp16")]; + tensor keys_45_cast_fp16 = select(a = var_470_to_fp16, b = keys_43_cast_fp16, cond = var_3341_cast_fp16)[name = tensor("keys_45_cast_fp16")]; + tensor var_3349_cast_fp16 = not_equal(x = values_43_cast_fp16, y = values_43_cast_fp16)[name = tensor("op_3349_cast_fp16")]; + tensor values_45_cast_fp16 = select(a = var_470_to_fp16, b = values_43_cast_fp16, cond = var_3349_cast_fp16)[name = tensor("values_45_cast_fp16")]; + tensor var_3373 = const()[name = tensor("op_3373"), val = tensor([0, 2, 1, 3])]; + tensor var_3389 = const()[name = tensor("op_3389"), val = tensor([1, 1, 1])]; + tensor var_3390_cast_fp16 = reshape(shape = var_3389, x = position7_to_fp16)[name = tensor("op_3390_cast_fp16")]; + tensor q_positions_15_cast_fp16 = add(x = var_3390_cast_fp16, y = q_offsets_1_promoted_to_fp16)[name = tensor("q_positions_15_cast_fp16")]; + tensor valid_end_15_cast_fp16 = add(x = var_3390_cast_fp16, y = var_536_cast_fp16)[name = tensor("valid_end_15_cast_fp16")]; + tensor valid_mask_15_cast_fp16 = less(x = k_positions_1_promoted_to_fp16, y = valid_end_15_cast_fp16)[name = tensor("valid_mask_15_cast_fp16")]; + tensor causal_mask_15_cast_fp16 = less_equal(x = k_positions_1_promoted_to_fp16, y = q_positions_15_cast_fp16)[name = tensor("causal_mask_15_cast_fp16")]; + tensor var_3418 = logical_and(x = valid_mask_15_cast_fp16, y = causal_mask_15_cast_fp16)[name = tensor("op_3418")]; + tensor attn_mask_15_axes_0 = const()[name = tensor("attn_mask_15_axes_0"), val = tensor([1])]; + tensor attn_mask_15 = expand_dims(axes = attn_mask_15_axes_0, x = var_3418)[name = tensor("attn_mask_15")]; + tensor var_3432_transpose_x_0 = const()[name = tensor("op_3432_transpose_x_0"), val = tensor(false)]; + tensor var_3432_transpose_y_0 = const()[name = tensor("op_3432_transpose_y_0"), val = tensor(false)]; + tensor transpose_83_perm_0 = const()[name = tensor("transpose_83_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_84_perm_0 = const()[name = tensor("transpose_84_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_84 = transpose(perm = transpose_84_perm_0, x = keys_45_cast_fp16)[name = tensor("transpose_176")]; + tensor transpose_83 = transpose(perm = transpose_83_perm_0, x = q_45_cast_fp16)[name = tensor("transpose_177")]; + tensor var_3432_cast_fp16 = matmul(transpose_x = var_3432_transpose_x_0, transpose_y = var_3432_transpose_y_0, x = transpose_83, y = transpose_84)[name = tensor("op_3432_cast_fp16")]; + tensor var_3426_to_fp16 = const()[name = tensor("op_3426_to_fp16"), val = tensor([0x1p-3])]; + tensor attn_weights_43_cast_fp16 = mul(x = var_3432_cast_fp16, y = var_3426_to_fp16)[name = tensor("attn_weights_43_cast_fp16")]; + tensor var_3434 = logical_not(x = attn_mask_15)[name = tensor("op_3434")]; + tensor var_3435_to_fp16 = const()[name = tensor("op_3435_to_fp16"), val = tensor(-inf)]; + tensor attn_weights_45_cast_fp16 = select(a = var_3435_to_fp16, b = attn_weights_43_cast_fp16, cond = var_3434)[name = tensor("attn_weights_45_cast_fp16")]; + tensor var_3437 = const()[name = tensor("op_3437"), val = tensor(-1)]; + tensor attn_weights_47_cast_fp16 = softmax(axis = var_3437, x = attn_weights_45_cast_fp16)[name = tensor("attn_weights_47_cast_fp16")]; + tensor attn_output_15_transpose_x_0 = const()[name = tensor("attn_output_15_transpose_x_0"), val = tensor(false)]; + tensor attn_output_15_transpose_y_0 = const()[name = tensor("attn_output_15_transpose_y_0"), val = tensor(false)]; + tensor values_47_cast_fp16 = transpose(perm = var_3373, x = values_45_cast_fp16)[name = tensor("transpose_178")]; + tensor attn_output_15_cast_fp16 = matmul(transpose_x = attn_output_15_transpose_x_0, transpose_y = attn_output_15_transpose_y_0, x = attn_weights_47_cast_fp16, y = values_47_cast_fp16)[name = tensor("attn_output_15_cast_fp16")]; + tensor var_3445 = const()[name = tensor("op_3445"), val = tensor([0, 2, 1, 3])]; + tensor var_3450 = const()[name = tensor("op_3450"), val = tensor([1, 256, 1024])]; + tensor var_3446_cast_fp16 = transpose(perm = var_3445, x = attn_output_15_cast_fp16)[name = tensor("transpose_175")]; + tensor input_89_cast_fp16 = reshape(shape = var_3450, x = var_3446_cast_fp16)[name = tensor("input_89_cast_fp16")]; + tensor attn7_out_proj_weight_to_fp16 = const()[name = tensor("attn7_out_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(183586560)))]; + tensor linear_29_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = attn7_out_proj_weight_to_fp16, x = input_89_cast_fp16)[name = tensor("linear_29_cast_fp16")]; + tensor var_3459_cast_fp16 = add(x = position7_to_fp16, y = valid_len_to_fp16)[name = tensor("op_3459_cast_fp16")]; + tensor var_3459_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_3459_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor input_91_cast_fp16 = add(x = input_85_cast_fp16, y = linear_29_cast_fp16)[name = tensor("input_91_cast_fp16")]; + tensor input_93_axes_0 = const()[name = tensor("input_93_axes_0"), val = tensor([-1])]; + tensor norm7_2_weight_to_fp16 = const()[name = tensor("norm7_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(185683776)))]; + tensor norm7_2_bias_to_fp16 = const()[name = tensor("norm7_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(185685888)))]; + tensor var_3463_to_fp16 = const()[name = tensor("op_3463_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_93_cast_fp16 = layer_norm(axes = input_93_axes_0, beta = norm7_2_bias_to_fp16, epsilon = var_3463_to_fp16, gamma = norm7_2_weight_to_fp16, x = input_91_cast_fp16)[name = tensor("input_93_cast_fp16")]; + tensor linear7_1_weight_to_fp16 = const()[name = tensor("linear7_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(185688000)))]; + tensor linear_30_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = linear7_1_weight_to_fp16, x = input_93_cast_fp16)[name = tensor("linear_30_cast_fp16")]; + tensor input_95_mode_0 = const()[name = tensor("input_95_mode_0"), val = tensor("EXACT")]; + tensor input_95_cast_fp16 = gelu(mode = input_95_mode_0, x = linear_30_cast_fp16)[name = tensor("input_95_cast_fp16")]; + tensor linear7_2_weight_to_fp16 = const()[name = tensor("linear7_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(194076672)))]; + tensor linear_31_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = linear7_2_weight_to_fp16, x = input_95_cast_fp16)[name = tensor("linear_31_cast_fp16")]; + tensor input_97_cast_fp16 = add(x = input_91_cast_fp16, y = linear_31_cast_fp16)[name = tensor("input_97_cast_fp16")]; + tensor input_99_axes_0 = const()[name = tensor("input_99_axes_0"), val = tensor([-1])]; + tensor norm8_1_weight_to_fp16 = const()[name = tensor("norm8_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(202465344)))]; + tensor norm8_1_bias_to_fp16 = const()[name = tensor("norm8_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(202467456)))]; + tensor var_3480_to_fp16 = const()[name = tensor("op_3480_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_99_cast_fp16 = layer_norm(axes = input_99_axes_0, beta = norm8_1_bias_to_fp16, epsilon = var_3480_to_fp16, gamma = norm8_1_weight_to_fp16, x = input_97_cast_fp16)[name = tensor("input_99_cast_fp16")]; + tensor attn8_in_proj_weight_to_fp16 = const()[name = tensor("attn8_in_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(202469568)))]; + tensor linear_32_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = attn8_in_proj_weight_to_fp16, x = input_99_cast_fp16)[name = tensor("linear_32_cast_fp16")]; + tensor var_3494 = const()[name = tensor("op_3494"), val = tensor([1, 256, 3, 16, 64])]; + tensor qkv_17_cast_fp16 = reshape(shape = var_3494, x = linear_32_cast_fp16)[name = tensor("qkv_17_cast_fp16")]; + tensor q_49_begin_0 = const()[name = tensor("q_49_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor q_49_end_0 = const()[name = tensor("q_49_end_0"), val = tensor([1, 256, 1, 16, 64])]; + tensor q_49_end_mask_0 = const()[name = tensor("q_49_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor q_49_squeeze_mask_0 = const()[name = tensor("q_49_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor q_49_cast_fp16 = slice_by_index(begin = q_49_begin_0, end = q_49_end_0, end_mask = q_49_end_mask_0, squeeze_mask = q_49_squeeze_mask_0, x = qkv_17_cast_fp16)[name = tensor("q_49_cast_fp16")]; + tensor k_33_begin_0 = const()[name = tensor("k_33_begin_0"), val = tensor([0, 0, 1, 0, 0])]; + tensor k_33_end_0 = const()[name = tensor("k_33_end_0"), val = tensor([1, 256, 2, 16, 64])]; + tensor k_33_end_mask_0 = const()[name = tensor("k_33_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor k_33_squeeze_mask_0 = const()[name = tensor("k_33_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor k_33_cast_fp16 = slice_by_index(begin = k_33_begin_0, end = k_33_end_0, end_mask = k_33_end_mask_0, squeeze_mask = k_33_squeeze_mask_0, x = qkv_17_cast_fp16)[name = tensor("k_33_cast_fp16")]; + tensor v_17_begin_0 = const()[name = tensor("v_17_begin_0"), val = tensor([0, 0, 2, 0, 0])]; + tensor v_17_end_0 = const()[name = tensor("v_17_end_0"), val = tensor([1, 256, 3, 16, 64])]; + tensor v_17_end_mask_0 = const()[name = tensor("v_17_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor v_17_squeeze_mask_0 = const()[name = tensor("v_17_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor v_17_cast_fp16 = slice_by_index(begin = v_17_begin_0, end = v_17_end_0, end_mask = v_17_end_mask_0, squeeze_mask = v_17_squeeze_mask_0, x = qkv_17_cast_fp16)[name = tensor("v_17_cast_fp16")]; + tensor var_3552 = const()[name = tensor("op_3552"), val = tensor([1, 1])]; + tensor position8_to_fp16_dtype_0 = const()[name = tensor("position8_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor position8_to_fp16 = cast(dtype = position8_to_fp16_dtype_0, x = position8)[name = tensor("cast_621")]; + tensor var_3553_cast_fp16 = reshape(shape = var_3552, x = position8_to_fp16)[name = tensor("op_3553_cast_fp16")]; + tensor ts_51_cast_fp16 = add(x = ts_1_promoted_to_fp16, y = var_3553_cast_fp16)[name = tensor("ts_51_cast_fp16")]; + tensor var_3560 = const()[name = tensor("op_3560"), val = tensor([1, 256, 1, 1])]; + tensor ts_53_cast_fp16 = reshape(shape = var_3560, x = ts_51_cast_fp16)[name = tensor("ts_53_cast_fp16")]; + tensor var_3567 = const()[name = tensor("op_3567"), val = tensor([1, 256, 16, 32, 2])]; + tensor q_complex_17_cast_fp16 = reshape(shape = var_3567, x = q_49_cast_fp16)[name = tensor("q_complex_17_cast_fp16")]; + tensor var_3574 = const()[name = tensor("op_3574"), val = tensor([1, 256, 16, 32, 2])]; + tensor k_complex_17_cast_fp16 = reshape(shape = var_3574, x = k_33_cast_fp16)[name = tensor("k_complex_17_cast_fp16")]; + tensor var_3578_begin_0 = const()[name = tensor("op_3578_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_3578_end_0 = const()[name = tensor("op_3578_end_0"), val = tensor([1, 256, 16, 32, 1])]; + tensor var_3578_end_mask_0 = const()[name = tensor("op_3578_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_3578_squeeze_mask_0 = const()[name = tensor("op_3578_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_3578_cast_fp16 = slice_by_index(begin = var_3578_begin_0, end = var_3578_end_0, end_mask = var_3578_end_mask_0, squeeze_mask = var_3578_squeeze_mask_0, x = q_complex_17_cast_fp16)[name = tensor("op_3578_cast_fp16")]; + tensor var_3586_begin_0 = const()[name = tensor("op_3586_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_3586_end_0 = const()[name = tensor("op_3586_end_0"), val = tensor([1, 256, 16, 32, 2])]; + tensor var_3586_end_mask_0 = const()[name = tensor("op_3586_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_3586_squeeze_mask_0 = const()[name = tensor("op_3586_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_3586_cast_fp16 = slice_by_index(begin = var_3586_begin_0, end = var_3586_end_0, end_mask = var_3586_end_mask_0, squeeze_mask = var_3586_squeeze_mask_0, x = q_complex_17_cast_fp16)[name = tensor("op_3586_cast_fp16")]; + tensor var_3594_begin_0 = const()[name = tensor("op_3594_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_3594_end_0 = const()[name = tensor("op_3594_end_0"), val = tensor([1, 256, 16, 32, 1])]; + tensor var_3594_end_mask_0 = const()[name = tensor("op_3594_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_3594_squeeze_mask_0 = const()[name = tensor("op_3594_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_3594_cast_fp16 = slice_by_index(begin = var_3594_begin_0, end = var_3594_end_0, end_mask = var_3594_end_mask_0, squeeze_mask = var_3594_squeeze_mask_0, x = k_complex_17_cast_fp16)[name = tensor("op_3594_cast_fp16")]; + tensor var_3602_begin_0 = const()[name = tensor("op_3602_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_3602_end_0 = const()[name = tensor("op_3602_end_0"), val = tensor([1, 256, 16, 32, 2])]; + tensor var_3602_end_mask_0 = const()[name = tensor("op_3602_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_3602_squeeze_mask_0 = const()[name = tensor("op_3602_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_3602_cast_fp16 = slice_by_index(begin = var_3602_begin_0, end = var_3602_end_0, end_mask = var_3602_end_mask_0, squeeze_mask = var_3602_squeeze_mask_0, x = k_complex_17_cast_fp16)[name = tensor("op_3602_cast_fp16")]; + tensor freqs_17_to_fp16 = const()[name = tensor("freqs_17_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(208761088)))]; + tensor var_3608_cast_fp16 = mul(x = freqs_17_to_fp16, y = ts_53_cast_fp16)[name = tensor("op_3608_cast_fp16")]; + tensor rotr_17_cast_fp16 = cos(x = var_3608_cast_fp16)[name = tensor("rotr_17_cast_fp16")]; + tensor roti_17_cast_fp16 = sin(x = var_3608_cast_fp16)[name = tensor("roti_17_cast_fp16")]; + tensor var_3612_cast_fp16 = mul(x = var_3578_cast_fp16, y = rotr_17_cast_fp16)[name = tensor("op_3612_cast_fp16")]; + tensor var_3613_cast_fp16 = mul(x = var_3586_cast_fp16, y = roti_17_cast_fp16)[name = tensor("op_3613_cast_fp16")]; + tensor qor_33_cast_fp16 = sub(x = var_3612_cast_fp16, y = var_3613_cast_fp16)[name = tensor("qor_33_cast_fp16")]; + tensor var_3616_cast_fp16 = mul(x = var_3578_cast_fp16, y = roti_17_cast_fp16)[name = tensor("op_3616_cast_fp16")]; + tensor var_3617_cast_fp16 = mul(x = var_3586_cast_fp16, y = rotr_17_cast_fp16)[name = tensor("op_3617_cast_fp16")]; + tensor qoi_33_cast_fp16 = add(x = var_3616_cast_fp16, y = var_3617_cast_fp16)[name = tensor("qoi_33_cast_fp16")]; + tensor var_3620_cast_fp16 = mul(x = var_3594_cast_fp16, y = rotr_17_cast_fp16)[name = tensor("op_3620_cast_fp16")]; + tensor var_3621_cast_fp16 = mul(x = var_3602_cast_fp16, y = roti_17_cast_fp16)[name = tensor("op_3621_cast_fp16")]; + tensor kor_33_cast_fp16 = sub(x = var_3620_cast_fp16, y = var_3621_cast_fp16)[name = tensor("kor_33_cast_fp16")]; + tensor var_3624_cast_fp16 = mul(x = var_3594_cast_fp16, y = roti_17_cast_fp16)[name = tensor("op_3624_cast_fp16")]; + tensor var_3625_cast_fp16 = mul(x = var_3602_cast_fp16, y = rotr_17_cast_fp16)[name = tensor("op_3625_cast_fp16")]; + tensor koi_33_cast_fp16 = add(x = var_3624_cast_fp16, y = var_3625_cast_fp16)[name = tensor("koi_33_cast_fp16")]; + tensor qo_17_axis_0 = const()[name = tensor("qo_17_axis_0"), val = tensor(-1)]; + tensor qo_17_cast_fp16 = stack(axis = qo_17_axis_0, values = (qor_33_cast_fp16, qoi_33_cast_fp16))[name = tensor("qo_17_cast_fp16")]; + tensor ko_17_axis_0 = const()[name = tensor("ko_17_axis_0"), val = tensor(-1)]; + tensor ko_17_cast_fp16 = stack(axis = ko_17_axis_0, values = (kor_33_cast_fp16, koi_33_cast_fp16))[name = tensor("ko_17_cast_fp16")]; + tensor var_3658 = const()[name = tensor("op_3658"), val = tensor([1, 256, 16, 64])]; + tensor q_51_cast_fp16 = reshape(shape = var_3658, x = qo_17_cast_fp16)[name = tensor("q_51_cast_fp16")]; + tensor var_3664 = const()[name = tensor("op_3664"), val = tensor([1, 256, 16, 64])]; + tensor k_35_cast_fp16 = reshape(shape = var_3664, x = ko_17_cast_fp16)[name = tensor("k_35_cast_fp16")]; + tensor write_indices_float_49_cast_fp16 = add(x = var_3553_cast_fp16, y = var_393_promoted_to_fp16)[name = tensor("write_indices_float_49_cast_fp16")]; + tensor _inversed_3685_y_0_to_fp16 = const()[name = tensor("_inversed_3685_y_0_to_fp16"), val = tensor(0x1p-9)]; + tensor _inversed_3685_cast_fp16 = mul(x = write_indices_float_49_cast_fp16, y = _inversed_3685_y_0_to_fp16)[name = tensor("_inversed_3685_cast_fp16")]; + tensor var_3686_cast_fp16 = floor(x = _inversed_3685_cast_fp16)[name = tensor("op_3686_cast_fp16")]; + tensor var_3687_to_fp16 = const()[name = tensor("op_3687_to_fp16"), val = tensor(0x1p+9)]; + tensor var_3688_cast_fp16 = mul(x = var_3686_cast_fp16, y = var_3687_to_fp16)[name = tensor("op_3688_cast_fp16")]; + tensor write_indices_float_51_cast_fp16 = sub(x = write_indices_float_49_cast_fp16, y = var_3688_cast_fp16)[name = tensor("write_indices_float_51_cast_fp16")]; + tensor var_3706_value_0_to_fp16 = const()[name = tensor("op_3706_value_0_to_fp16"), val = tensor(0x1.ffp+8)]; + tensor var_3706_cast_fp16 = fill_like(ref_tensor = write_indices_float_51_cast_fp16, value = var_3706_value_0_to_fp16)[name = tensor("op_3706_cast_fp16")]; + tensor write_indices_float_53_cast_fp16 = select(a = write_indices_float_51_cast_fp16, b = var_3706_cast_fp16, cond = token_valid_1_cast_fp16)[name = tensor("write_indices_float_53_cast_fp16")]; + tensor var_3712_dtype_0 = const()[name = tensor("op_3712_dtype_0"), val = tensor("int32")]; + tensor var_3717 = const()[name = tensor("op_3717"), val = tensor([1, 256, 1, 1])]; + tensor write_indices_float_53_cast_fp16_to_int32 = cast(dtype = var_3712_dtype_0, x = write_indices_float_53_cast_fp16)[name = tensor("cast_620")]; + tensor var_3718 = reshape(shape = var_3717, x = write_indices_float_53_cast_fp16_to_int32)[name = tensor("op_3718")]; + tensor write_indices_17_reps_0 = const()[name = tensor("write_indices_17_reps_0"), val = tensor([1, 1, 16, 64])]; + tensor write_indices_17 = tile(reps = write_indices_17_reps_0, x = var_3718)[name = tensor("write_indices_17")]; + tensor var_3728_begin_0 = const()[name = tensor("op_3728_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_3728_end_0 = const()[name = tensor("op_3728_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor var_3728_end_mask_0 = const()[name = tensor("op_3728_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_3728_squeeze_mask_0 = const()[name = tensor("op_3728_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor cache8_to_fp16_dtype_0 = const()[name = tensor("cache8_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor cache8_to_fp16 = cast(dtype = cache8_to_fp16_dtype_0, x = cache8)[name = tensor("cast_619")]; + tensor var_3728_cast_fp16 = slice_by_index(begin = var_3728_begin_0, end = var_3728_end_0, end_mask = var_3728_end_mask_0, squeeze_mask = var_3728_squeeze_mask_0, x = cache8_to_fp16)[name = tensor("op_3728_cast_fp16")]; + tensor var_3730_axis_0 = const()[name = tensor("op_3730_axis_0"), val = tensor(1)]; + tensor var_3730_mode_0 = const()[name = tensor("op_3730_mode_0"), val = tensor("update")]; + tensor var_3730_validate_indices_0 = const()[name = tensor("op_3730_validate_indices_0"), val = tensor(false)]; + tensor var_3730_cast_fp16 = scatter_along_axis(axis = var_3730_axis_0, data = var_3728_cast_fp16, indices = write_indices_17, mode = var_3730_mode_0, updates = k_35_cast_fp16, validate_indices = var_3730_validate_indices_0)[name = tensor("op_3730_cast_fp16")]; + tensor concat_66 = const()[name = tensor("concat_66"), val = tensor([0, 0, 0, 0, 0])]; + tensor concat_67 = const()[name = tensor("concat_67"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_17_internal_tensor_assign_1_stride_0 = const()[name = tensor("new_cache_17_internal_tensor_assign_1_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_17_internal_tensor_assign_1_begin_mask_0 = const()[name = tensor("new_cache_17_internal_tensor_assign_1_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_17_internal_tensor_assign_1_end_mask_0 = const()[name = tensor("new_cache_17_internal_tensor_assign_1_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_17_internal_tensor_assign_1_squeeze_mask_0 = const()[name = tensor("new_cache_17_internal_tensor_assign_1_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_62 = const()[name = tensor("shape_62"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_16 = const()[name = tensor("reduce_prod_16"), val = tensor(1048576)]; + tensor range_1d_16_start_0 = const()[name = tensor("range_1d_16_start_0"), val = tensor(0)]; + tensor range_1d_16_step_0 = const()[name = tensor("range_1d_16_step_0"), val = tensor(1)]; + tensor range_1d_16 = range_1d(end = reduce_prod_16, start = range_1d_16_start_0, step = range_1d_16_step_0)[name = tensor("range_1d_16")]; + tensor reshape_80 = reshape(shape = shape_62, x = range_1d_16)[name = tensor("reshape_80")]; + tensor slice_by_index_16 = slice_by_index(begin = concat_66, begin_mask = new_cache_17_internal_tensor_assign_1_begin_mask_0, end = concat_67, end_mask = new_cache_17_internal_tensor_assign_1_end_mask_0, squeeze_mask = new_cache_17_internal_tensor_assign_1_squeeze_mask_0, stride = new_cache_17_internal_tensor_assign_1_stride_0, x = reshape_80)[name = tensor("slice_by_index_16")]; + tensor reshape_81_shape_0 = const()[name = tensor("reshape_81_shape_0"), val = tensor([-1])]; + tensor reshape_81 = reshape(shape = reshape_81_shape_0, x = slice_by_index_16)[name = tensor("reshape_81")]; + tensor reshape_82_shape_0 = const()[name = tensor("reshape_82_shape_0"), val = tensor([-1])]; + tensor reshape_82_cast_fp16 = reshape(shape = reshape_82_shape_0, x = var_3730_cast_fp16)[name = tensor("reshape_82_cast_fp16")]; + tensor reshape_83_shape_0 = const()[name = tensor("reshape_83_shape_0"), val = tensor([-1])]; + tensor reshape_83_cast_fp16 = reshape(shape = reshape_83_shape_0, x = cache8_to_fp16)[name = tensor("reshape_83_cast_fp16")]; + tensor scatter_16_mode_0 = const()[name = tensor("scatter_16_mode_0"), val = tensor("update")]; + tensor scatter_16_axis_0 = const()[name = tensor("scatter_16_axis_0"), val = tensor(0)]; + tensor scatter_16_validate_indices_0 = const()[name = tensor("scatter_16_validate_indices_0"), val = tensor(false)]; + tensor scatter_16_cast_fp16 = scatter(axis = scatter_16_axis_0, data = reshape_83_cast_fp16, indices = reshape_81, mode = scatter_16_mode_0, updates = reshape_82_cast_fp16, validate_indices = scatter_16_validate_indices_0)[name = tensor("scatter_16_cast_fp16")]; + tensor reshape_84_cast_fp16 = reshape(shape = shape_62, x = scatter_16_cast_fp16)[name = tensor("reshape_84_cast_fp16")]; + tensor var_3738_begin_0 = const()[name = tensor("op_3738_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor var_3738_end_0 = const()[name = tensor("op_3738_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor var_3738_end_mask_0 = const()[name = tensor("op_3738_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_3738_squeeze_mask_0 = const()[name = tensor("op_3738_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor var_3738_cast_fp16 = slice_by_index(begin = var_3738_begin_0, end = var_3738_end_0, end_mask = var_3738_end_mask_0, squeeze_mask = var_3738_squeeze_mask_0, x = reshape_84_cast_fp16)[name = tensor("op_3738_cast_fp16")]; + tensor var_3740_axis_0 = const()[name = tensor("op_3740_axis_0"), val = tensor(1)]; + tensor var_3740_mode_0 = const()[name = tensor("op_3740_mode_0"), val = tensor("update")]; + tensor var_3740_validate_indices_0 = const()[name = tensor("op_3740_validate_indices_0"), val = tensor(false)]; + tensor var_3740_cast_fp16 = scatter_along_axis(axis = var_3740_axis_0, data = var_3738_cast_fp16, indices = write_indices_17, mode = var_3740_mode_0, updates = v_17_cast_fp16, validate_indices = var_3740_validate_indices_0)[name = tensor("op_3740_cast_fp16")]; + tensor concat_68 = const()[name = tensor("concat_68"), val = tensor([1, 0, 0, 0, 0])]; + tensor concat_69 = const()[name = tensor("concat_69"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_17_internal_tensor_assign_2_stride_0 = const()[name = tensor("new_cache_17_internal_tensor_assign_2_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_17_internal_tensor_assign_2_begin_mask_0 = const()[name = tensor("new_cache_17_internal_tensor_assign_2_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_17_internal_tensor_assign_2_end_mask_0 = const()[name = tensor("new_cache_17_internal_tensor_assign_2_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_17_internal_tensor_assign_2_squeeze_mask_0 = const()[name = tensor("new_cache_17_internal_tensor_assign_2_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_63 = const()[name = tensor("shape_63"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_17 = const()[name = tensor("reduce_prod_17"), val = tensor(1048576)]; + tensor range_1d_17_start_0 = const()[name = tensor("range_1d_17_start_0"), val = tensor(0)]; + tensor range_1d_17_step_0 = const()[name = tensor("range_1d_17_step_0"), val = tensor(1)]; + tensor range_1d_17 = range_1d(end = reduce_prod_17, start = range_1d_17_start_0, step = range_1d_17_step_0)[name = tensor("range_1d_17")]; + tensor reshape_85 = reshape(shape = shape_63, x = range_1d_17)[name = tensor("reshape_85")]; + tensor slice_by_index_17 = slice_by_index(begin = concat_68, begin_mask = new_cache_17_internal_tensor_assign_2_begin_mask_0, end = concat_69, end_mask = new_cache_17_internal_tensor_assign_2_end_mask_0, squeeze_mask = new_cache_17_internal_tensor_assign_2_squeeze_mask_0, stride = new_cache_17_internal_tensor_assign_2_stride_0, x = reshape_85)[name = tensor("slice_by_index_17")]; + tensor reshape_86_shape_0 = const()[name = tensor("reshape_86_shape_0"), val = tensor([-1])]; + tensor reshape_86 = reshape(shape = reshape_86_shape_0, x = slice_by_index_17)[name = tensor("reshape_86")]; + tensor reshape_87_shape_0 = const()[name = tensor("reshape_87_shape_0"), val = tensor([-1])]; + tensor reshape_87_cast_fp16 = reshape(shape = reshape_87_shape_0, x = var_3740_cast_fp16)[name = tensor("reshape_87_cast_fp16")]; + tensor reshape_88_shape_0 = const()[name = tensor("reshape_88_shape_0"), val = tensor([-1])]; + tensor reshape_88_cast_fp16 = reshape(shape = reshape_88_shape_0, x = reshape_84_cast_fp16)[name = tensor("reshape_88_cast_fp16")]; + tensor scatter_17_mode_0 = const()[name = tensor("scatter_17_mode_0"), val = tensor("update")]; + tensor scatter_17_axis_0 = const()[name = tensor("scatter_17_axis_0"), val = tensor(0)]; + tensor scatter_17_validate_indices_0 = const()[name = tensor("scatter_17_validate_indices_0"), val = tensor(false)]; + tensor scatter_17_cast_fp16 = scatter(axis = scatter_17_axis_0, data = reshape_88_cast_fp16, indices = reshape_86, mode = scatter_17_mode_0, updates = reshape_87_cast_fp16, validate_indices = scatter_17_validate_indices_0)[name = tensor("scatter_17_cast_fp16")]; + tensor reshape_89_cast_fp16 = reshape(shape = shape_63, x = scatter_17_cast_fp16)[name = tensor("reshape_89_cast_fp16")]; + tensor reshape_89_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("reshape_89_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor keys_49_begin_0 = const()[name = tensor("keys_49_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor keys_49_end_0 = const()[name = tensor("keys_49_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor keys_49_end_mask_0 = const()[name = tensor("keys_49_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor keys_49_squeeze_mask_0 = const()[name = tensor("keys_49_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor keys_49_cast_fp16 = slice_by_index(begin = keys_49_begin_0, end = keys_49_end_0, end_mask = keys_49_end_mask_0, squeeze_mask = keys_49_squeeze_mask_0, x = reshape_89_cast_fp16)[name = tensor("keys_49_cast_fp16")]; + tensor values_49_begin_0 = const()[name = tensor("values_49_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor values_49_end_0 = const()[name = tensor("values_49_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor values_49_end_mask_0 = const()[name = tensor("values_49_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor values_49_squeeze_mask_0 = const()[name = tensor("values_49_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor values_49_cast_fp16 = slice_by_index(begin = values_49_begin_0, end = values_49_end_0, end_mask = values_49_end_mask_0, squeeze_mask = values_49_squeeze_mask_0, x = reshape_89_cast_fp16)[name = tensor("values_49_cast_fp16")]; + tensor var_3752_cast_fp16 = not_equal(x = keys_49_cast_fp16, y = keys_49_cast_fp16)[name = tensor("op_3752_cast_fp16")]; + tensor keys_51_cast_fp16 = select(a = var_470_to_fp16, b = keys_49_cast_fp16, cond = var_3752_cast_fp16)[name = tensor("keys_51_cast_fp16")]; + tensor var_3760_cast_fp16 = not_equal(x = values_49_cast_fp16, y = values_49_cast_fp16)[name = tensor("op_3760_cast_fp16")]; + tensor values_51_cast_fp16 = select(a = var_470_to_fp16, b = values_49_cast_fp16, cond = var_3760_cast_fp16)[name = tensor("values_51_cast_fp16")]; + tensor var_3784 = const()[name = tensor("op_3784"), val = tensor([0, 2, 1, 3])]; + tensor var_3800 = const()[name = tensor("op_3800"), val = tensor([1, 1, 1])]; + tensor var_3801_cast_fp16 = reshape(shape = var_3800, x = position8_to_fp16)[name = tensor("op_3801_cast_fp16")]; + tensor q_positions_17_cast_fp16 = add(x = var_3801_cast_fp16, y = q_offsets_1_promoted_to_fp16)[name = tensor("q_positions_17_cast_fp16")]; + tensor valid_end_17_cast_fp16 = add(x = var_3801_cast_fp16, y = var_536_cast_fp16)[name = tensor("valid_end_17_cast_fp16")]; + tensor valid_mask_17_cast_fp16 = less(x = k_positions_1_promoted_to_fp16, y = valid_end_17_cast_fp16)[name = tensor("valid_mask_17_cast_fp16")]; + tensor causal_mask_17_cast_fp16 = less_equal(x = k_positions_1_promoted_to_fp16, y = q_positions_17_cast_fp16)[name = tensor("causal_mask_17_cast_fp16")]; + tensor var_3829 = logical_and(x = valid_mask_17_cast_fp16, y = causal_mask_17_cast_fp16)[name = tensor("op_3829")]; + tensor attn_mask_17_axes_0 = const()[name = tensor("attn_mask_17_axes_0"), val = tensor([1])]; + tensor attn_mask_17 = expand_dims(axes = attn_mask_17_axes_0, x = var_3829)[name = tensor("attn_mask_17")]; + tensor var_3843_transpose_x_0 = const()[name = tensor("op_3843_transpose_x_0"), val = tensor(false)]; + tensor var_3843_transpose_y_0 = const()[name = tensor("op_3843_transpose_y_0"), val = tensor(false)]; + tensor transpose_85_perm_0 = const()[name = tensor("transpose_85_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_86_perm_0 = const()[name = tensor("transpose_86_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_86 = transpose(perm = transpose_86_perm_0, x = keys_51_cast_fp16)[name = tensor("transpose_172")]; + tensor transpose_85 = transpose(perm = transpose_85_perm_0, x = q_51_cast_fp16)[name = tensor("transpose_173")]; + tensor var_3843_cast_fp16 = matmul(transpose_x = var_3843_transpose_x_0, transpose_y = var_3843_transpose_y_0, x = transpose_85, y = transpose_86)[name = tensor("op_3843_cast_fp16")]; + tensor var_3837_to_fp16 = const()[name = tensor("op_3837_to_fp16"), val = tensor([0x1p-3])]; + tensor attn_weights_49_cast_fp16 = mul(x = var_3843_cast_fp16, y = var_3837_to_fp16)[name = tensor("attn_weights_49_cast_fp16")]; + tensor var_3845 = logical_not(x = attn_mask_17)[name = tensor("op_3845")]; + tensor var_3846_to_fp16 = const()[name = tensor("op_3846_to_fp16"), val = tensor(-inf)]; + tensor attn_weights_51_cast_fp16 = select(a = var_3846_to_fp16, b = attn_weights_49_cast_fp16, cond = var_3845)[name = tensor("attn_weights_51_cast_fp16")]; + tensor var_3848 = const()[name = tensor("op_3848"), val = tensor(-1)]; + tensor attn_weights_53_cast_fp16 = softmax(axis = var_3848, x = attn_weights_51_cast_fp16)[name = tensor("attn_weights_53_cast_fp16")]; + tensor attn_output_17_transpose_x_0 = const()[name = tensor("attn_output_17_transpose_x_0"), val = tensor(false)]; + tensor attn_output_17_transpose_y_0 = const()[name = tensor("attn_output_17_transpose_y_0"), val = tensor(false)]; + tensor values_53_cast_fp16 = transpose(perm = var_3784, x = values_51_cast_fp16)[name = tensor("transpose_174")]; + tensor attn_output_17_cast_fp16 = matmul(transpose_x = attn_output_17_transpose_x_0, transpose_y = attn_output_17_transpose_y_0, x = attn_weights_53_cast_fp16, y = values_53_cast_fp16)[name = tensor("attn_output_17_cast_fp16")]; + tensor var_3856 = const()[name = tensor("op_3856"), val = tensor([0, 2, 1, 3])]; + tensor var_3861 = const()[name = tensor("op_3861"), val = tensor([1, 256, 1024])]; + tensor var_3857_cast_fp16 = transpose(perm = var_3856, x = attn_output_17_cast_fp16)[name = tensor("transpose_171")]; + tensor input_101_cast_fp16 = reshape(shape = var_3861, x = var_3857_cast_fp16)[name = tensor("input_101_cast_fp16")]; + tensor attn8_out_proj_weight_to_fp16 = const()[name = tensor("attn8_out_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(208761216)))]; + tensor linear_33_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = attn8_out_proj_weight_to_fp16, x = input_101_cast_fp16)[name = tensor("linear_33_cast_fp16")]; + tensor var_3870_cast_fp16 = add(x = position8_to_fp16, y = valid_len_to_fp16)[name = tensor("op_3870_cast_fp16")]; + tensor var_3870_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_3870_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor input_103_cast_fp16 = add(x = input_97_cast_fp16, y = linear_33_cast_fp16)[name = tensor("input_103_cast_fp16")]; + tensor input_105_axes_0 = const()[name = tensor("input_105_axes_0"), val = tensor([-1])]; + tensor norm8_2_weight_to_fp16 = const()[name = tensor("norm8_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(210858432)))]; + tensor norm8_2_bias_to_fp16 = const()[name = tensor("norm8_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(210860544)))]; + tensor var_3874_to_fp16 = const()[name = tensor("op_3874_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_105_cast_fp16 = layer_norm(axes = input_105_axes_0, beta = norm8_2_bias_to_fp16, epsilon = var_3874_to_fp16, gamma = norm8_2_weight_to_fp16, x = input_103_cast_fp16)[name = tensor("input_105_cast_fp16")]; + tensor linear8_1_weight_to_fp16 = const()[name = tensor("linear8_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(210862656)))]; + tensor linear_34_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = linear8_1_weight_to_fp16, x = input_105_cast_fp16)[name = tensor("linear_34_cast_fp16")]; + tensor input_107_mode_0 = const()[name = tensor("input_107_mode_0"), val = tensor("EXACT")]; + tensor input_107_cast_fp16 = gelu(mode = input_107_mode_0, x = linear_34_cast_fp16)[name = tensor("input_107_cast_fp16")]; + tensor linear8_2_weight_to_fp16 = const()[name = tensor("linear8_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(219251328)))]; + tensor linear_35_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = linear8_2_weight_to_fp16, x = input_107_cast_fp16)[name = tensor("linear_35_cast_fp16")]; + tensor input_109_cast_fp16 = add(x = input_103_cast_fp16, y = linear_35_cast_fp16)[name = tensor("input_109_cast_fp16")]; + tensor input_111_axes_0 = const()[name = tensor("input_111_axes_0"), val = tensor([-1])]; + tensor norm9_1_weight_to_fp16 = const()[name = tensor("norm9_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(227640000)))]; + tensor norm9_1_bias_to_fp16 = const()[name = tensor("norm9_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(227642112)))]; + tensor var_3891_to_fp16 = const()[name = tensor("op_3891_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_111_cast_fp16 = layer_norm(axes = input_111_axes_0, beta = norm9_1_bias_to_fp16, epsilon = var_3891_to_fp16, gamma = norm9_1_weight_to_fp16, x = input_109_cast_fp16)[name = tensor("input_111_cast_fp16")]; + tensor attn9_in_proj_weight_to_fp16 = const()[name = tensor("attn9_in_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(227644224)))]; + tensor linear_36_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = attn9_in_proj_weight_to_fp16, x = input_111_cast_fp16)[name = tensor("linear_36_cast_fp16")]; + tensor var_3905 = const()[name = tensor("op_3905"), val = tensor([1, 256, 3, 16, 64])]; + tensor qkv_19_cast_fp16 = reshape(shape = var_3905, x = linear_36_cast_fp16)[name = tensor("qkv_19_cast_fp16")]; + tensor q_55_begin_0 = const()[name = tensor("q_55_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor q_55_end_0 = const()[name = tensor("q_55_end_0"), val = tensor([1, 256, 1, 16, 64])]; + tensor q_55_end_mask_0 = const()[name = tensor("q_55_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor q_55_squeeze_mask_0 = const()[name = tensor("q_55_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor q_55_cast_fp16 = slice_by_index(begin = q_55_begin_0, end = q_55_end_0, end_mask = q_55_end_mask_0, squeeze_mask = q_55_squeeze_mask_0, x = qkv_19_cast_fp16)[name = tensor("q_55_cast_fp16")]; + tensor k_37_begin_0 = const()[name = tensor("k_37_begin_0"), val = tensor([0, 0, 1, 0, 0])]; + tensor k_37_end_0 = const()[name = tensor("k_37_end_0"), val = tensor([1, 256, 2, 16, 64])]; + tensor k_37_end_mask_0 = const()[name = tensor("k_37_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor k_37_squeeze_mask_0 = const()[name = tensor("k_37_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor k_37_cast_fp16 = slice_by_index(begin = k_37_begin_0, end = k_37_end_0, end_mask = k_37_end_mask_0, squeeze_mask = k_37_squeeze_mask_0, x = qkv_19_cast_fp16)[name = tensor("k_37_cast_fp16")]; + tensor v_19_begin_0 = const()[name = tensor("v_19_begin_0"), val = tensor([0, 0, 2, 0, 0])]; + tensor v_19_end_0 = const()[name = tensor("v_19_end_0"), val = tensor([1, 256, 3, 16, 64])]; + tensor v_19_end_mask_0 = const()[name = tensor("v_19_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor v_19_squeeze_mask_0 = const()[name = tensor("v_19_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor v_19_cast_fp16 = slice_by_index(begin = v_19_begin_0, end = v_19_end_0, end_mask = v_19_end_mask_0, squeeze_mask = v_19_squeeze_mask_0, x = qkv_19_cast_fp16)[name = tensor("v_19_cast_fp16")]; + tensor var_3963 = const()[name = tensor("op_3963"), val = tensor([1, 1])]; + tensor position9_to_fp16_dtype_0 = const()[name = tensor("position9_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor position9_to_fp16 = cast(dtype = position9_to_fp16_dtype_0, x = position9)[name = tensor("cast_616")]; + tensor var_3964_cast_fp16 = reshape(shape = var_3963, x = position9_to_fp16)[name = tensor("op_3964_cast_fp16")]; + tensor ts_57_cast_fp16 = add(x = ts_1_promoted_to_fp16, y = var_3964_cast_fp16)[name = tensor("ts_57_cast_fp16")]; + tensor var_3971 = const()[name = tensor("op_3971"), val = tensor([1, 256, 1, 1])]; + tensor ts_59_cast_fp16 = reshape(shape = var_3971, x = ts_57_cast_fp16)[name = tensor("ts_59_cast_fp16")]; + tensor var_3978 = const()[name = tensor("op_3978"), val = tensor([1, 256, 16, 32, 2])]; + tensor q_complex_19_cast_fp16 = reshape(shape = var_3978, x = q_55_cast_fp16)[name = tensor("q_complex_19_cast_fp16")]; + tensor var_3985 = const()[name = tensor("op_3985"), val = tensor([1, 256, 16, 32, 2])]; + tensor k_complex_19_cast_fp16 = reshape(shape = var_3985, x = k_37_cast_fp16)[name = tensor("k_complex_19_cast_fp16")]; + tensor var_3989_begin_0 = const()[name = tensor("op_3989_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_3989_end_0 = const()[name = tensor("op_3989_end_0"), val = tensor([1, 256, 16, 32, 1])]; + tensor var_3989_end_mask_0 = const()[name = tensor("op_3989_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_3989_squeeze_mask_0 = const()[name = tensor("op_3989_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_3989_cast_fp16 = slice_by_index(begin = var_3989_begin_0, end = var_3989_end_0, end_mask = var_3989_end_mask_0, squeeze_mask = var_3989_squeeze_mask_0, x = q_complex_19_cast_fp16)[name = tensor("op_3989_cast_fp16")]; + tensor var_3997_begin_0 = const()[name = tensor("op_3997_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_3997_end_0 = const()[name = tensor("op_3997_end_0"), val = tensor([1, 256, 16, 32, 2])]; + tensor var_3997_end_mask_0 = const()[name = tensor("op_3997_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_3997_squeeze_mask_0 = const()[name = tensor("op_3997_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_3997_cast_fp16 = slice_by_index(begin = var_3997_begin_0, end = var_3997_end_0, end_mask = var_3997_end_mask_0, squeeze_mask = var_3997_squeeze_mask_0, x = q_complex_19_cast_fp16)[name = tensor("op_3997_cast_fp16")]; + tensor var_4005_begin_0 = const()[name = tensor("op_4005_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_4005_end_0 = const()[name = tensor("op_4005_end_0"), val = tensor([1, 256, 16, 32, 1])]; + tensor var_4005_end_mask_0 = const()[name = tensor("op_4005_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_4005_squeeze_mask_0 = const()[name = tensor("op_4005_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_4005_cast_fp16 = slice_by_index(begin = var_4005_begin_0, end = var_4005_end_0, end_mask = var_4005_end_mask_0, squeeze_mask = var_4005_squeeze_mask_0, x = k_complex_19_cast_fp16)[name = tensor("op_4005_cast_fp16")]; + tensor var_4013_begin_0 = const()[name = tensor("op_4013_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_4013_end_0 = const()[name = tensor("op_4013_end_0"), val = tensor([1, 256, 16, 32, 2])]; + tensor var_4013_end_mask_0 = const()[name = tensor("op_4013_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_4013_squeeze_mask_0 = const()[name = tensor("op_4013_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_4013_cast_fp16 = slice_by_index(begin = var_4013_begin_0, end = var_4013_end_0, end_mask = var_4013_end_mask_0, squeeze_mask = var_4013_squeeze_mask_0, x = k_complex_19_cast_fp16)[name = tensor("op_4013_cast_fp16")]; + tensor freqs_19_to_fp16 = const()[name = tensor("freqs_19_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(233935744)))]; + tensor var_4019_cast_fp16 = mul(x = freqs_19_to_fp16, y = ts_59_cast_fp16)[name = tensor("op_4019_cast_fp16")]; + tensor rotr_19_cast_fp16 = cos(x = var_4019_cast_fp16)[name = tensor("rotr_19_cast_fp16")]; + tensor roti_19_cast_fp16 = sin(x = var_4019_cast_fp16)[name = tensor("roti_19_cast_fp16")]; + tensor var_4023_cast_fp16 = mul(x = var_3989_cast_fp16, y = rotr_19_cast_fp16)[name = tensor("op_4023_cast_fp16")]; + tensor var_4024_cast_fp16 = mul(x = var_3997_cast_fp16, y = roti_19_cast_fp16)[name = tensor("op_4024_cast_fp16")]; + tensor qor_37_cast_fp16 = sub(x = var_4023_cast_fp16, y = var_4024_cast_fp16)[name = tensor("qor_37_cast_fp16")]; + tensor var_4027_cast_fp16 = mul(x = var_3989_cast_fp16, y = roti_19_cast_fp16)[name = tensor("op_4027_cast_fp16")]; + tensor var_4028_cast_fp16 = mul(x = var_3997_cast_fp16, y = rotr_19_cast_fp16)[name = tensor("op_4028_cast_fp16")]; + tensor qoi_37_cast_fp16 = add(x = var_4027_cast_fp16, y = var_4028_cast_fp16)[name = tensor("qoi_37_cast_fp16")]; + tensor var_4031_cast_fp16 = mul(x = var_4005_cast_fp16, y = rotr_19_cast_fp16)[name = tensor("op_4031_cast_fp16")]; + tensor var_4032_cast_fp16 = mul(x = var_4013_cast_fp16, y = roti_19_cast_fp16)[name = tensor("op_4032_cast_fp16")]; + tensor kor_37_cast_fp16 = sub(x = var_4031_cast_fp16, y = var_4032_cast_fp16)[name = tensor("kor_37_cast_fp16")]; + tensor var_4035_cast_fp16 = mul(x = var_4005_cast_fp16, y = roti_19_cast_fp16)[name = tensor("op_4035_cast_fp16")]; + tensor var_4036_cast_fp16 = mul(x = var_4013_cast_fp16, y = rotr_19_cast_fp16)[name = tensor("op_4036_cast_fp16")]; + tensor koi_37_cast_fp16 = add(x = var_4035_cast_fp16, y = var_4036_cast_fp16)[name = tensor("koi_37_cast_fp16")]; + tensor qo_19_axis_0 = const()[name = tensor("qo_19_axis_0"), val = tensor(-1)]; + tensor qo_19_cast_fp16 = stack(axis = qo_19_axis_0, values = (qor_37_cast_fp16, qoi_37_cast_fp16))[name = tensor("qo_19_cast_fp16")]; + tensor ko_19_axis_0 = const()[name = tensor("ko_19_axis_0"), val = tensor(-1)]; + tensor ko_19_cast_fp16 = stack(axis = ko_19_axis_0, values = (kor_37_cast_fp16, koi_37_cast_fp16))[name = tensor("ko_19_cast_fp16")]; + tensor var_4069 = const()[name = tensor("op_4069"), val = tensor([1, 256, 16, 64])]; + tensor q_57_cast_fp16 = reshape(shape = var_4069, x = qo_19_cast_fp16)[name = tensor("q_57_cast_fp16")]; + tensor var_4075 = const()[name = tensor("op_4075"), val = tensor([1, 256, 16, 64])]; + tensor k_39_cast_fp16 = reshape(shape = var_4075, x = ko_19_cast_fp16)[name = tensor("k_39_cast_fp16")]; + tensor write_indices_float_55_cast_fp16 = add(x = var_3964_cast_fp16, y = var_393_promoted_to_fp16)[name = tensor("write_indices_float_55_cast_fp16")]; + tensor _inversed_4096_y_0_to_fp16 = const()[name = tensor("_inversed_4096_y_0_to_fp16"), val = tensor(0x1p-9)]; + tensor _inversed_4096_cast_fp16 = mul(x = write_indices_float_55_cast_fp16, y = _inversed_4096_y_0_to_fp16)[name = tensor("_inversed_4096_cast_fp16")]; + tensor var_4097_cast_fp16 = floor(x = _inversed_4096_cast_fp16)[name = tensor("op_4097_cast_fp16")]; + tensor var_4098_to_fp16 = const()[name = tensor("op_4098_to_fp16"), val = tensor(0x1p+9)]; + tensor var_4099_cast_fp16 = mul(x = var_4097_cast_fp16, y = var_4098_to_fp16)[name = tensor("op_4099_cast_fp16")]; + tensor write_indices_float_57_cast_fp16 = sub(x = write_indices_float_55_cast_fp16, y = var_4099_cast_fp16)[name = tensor("write_indices_float_57_cast_fp16")]; + tensor var_4117_value_0_to_fp16 = const()[name = tensor("op_4117_value_0_to_fp16"), val = tensor(0x1.ffp+8)]; + tensor var_4117_cast_fp16 = fill_like(ref_tensor = write_indices_float_57_cast_fp16, value = var_4117_value_0_to_fp16)[name = tensor("op_4117_cast_fp16")]; + tensor write_indices_float_59_cast_fp16 = select(a = write_indices_float_57_cast_fp16, b = var_4117_cast_fp16, cond = token_valid_1_cast_fp16)[name = tensor("write_indices_float_59_cast_fp16")]; + tensor var_4123_dtype_0 = const()[name = tensor("op_4123_dtype_0"), val = tensor("int32")]; + tensor var_4128 = const()[name = tensor("op_4128"), val = tensor([1, 256, 1, 1])]; + tensor write_indices_float_59_cast_fp16_to_int32 = cast(dtype = var_4123_dtype_0, x = write_indices_float_59_cast_fp16)[name = tensor("cast_615")]; + tensor var_4129 = reshape(shape = var_4128, x = write_indices_float_59_cast_fp16_to_int32)[name = tensor("op_4129")]; + tensor write_indices_19_reps_0 = const()[name = tensor("write_indices_19_reps_0"), val = tensor([1, 1, 16, 64])]; + tensor write_indices_19 = tile(reps = write_indices_19_reps_0, x = var_4129)[name = tensor("write_indices_19")]; + tensor var_4139_begin_0 = const()[name = tensor("op_4139_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_4139_end_0 = const()[name = tensor("op_4139_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor var_4139_end_mask_0 = const()[name = tensor("op_4139_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_4139_squeeze_mask_0 = const()[name = tensor("op_4139_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor cache9_to_fp16_dtype_0 = const()[name = tensor("cache9_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor cache9_to_fp16 = cast(dtype = cache9_to_fp16_dtype_0, x = cache9)[name = tensor("cast_614")]; + tensor var_4139_cast_fp16 = slice_by_index(begin = var_4139_begin_0, end = var_4139_end_0, end_mask = var_4139_end_mask_0, squeeze_mask = var_4139_squeeze_mask_0, x = cache9_to_fp16)[name = tensor("op_4139_cast_fp16")]; + tensor var_4141_axis_0 = const()[name = tensor("op_4141_axis_0"), val = tensor(1)]; + tensor var_4141_mode_0 = const()[name = tensor("op_4141_mode_0"), val = tensor("update")]; + tensor var_4141_validate_indices_0 = const()[name = tensor("op_4141_validate_indices_0"), val = tensor(false)]; + tensor var_4141_cast_fp16 = scatter_along_axis(axis = var_4141_axis_0, data = var_4139_cast_fp16, indices = write_indices_19, mode = var_4141_mode_0, updates = k_39_cast_fp16, validate_indices = var_4141_validate_indices_0)[name = tensor("op_4141_cast_fp16")]; + tensor concat_74 = const()[name = tensor("concat_74"), val = tensor([0, 0, 0, 0, 0])]; + tensor concat_75 = const()[name = tensor("concat_75"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_19_internal_tensor_assign_1_stride_0 = const()[name = tensor("new_cache_19_internal_tensor_assign_1_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_19_internal_tensor_assign_1_begin_mask_0 = const()[name = tensor("new_cache_19_internal_tensor_assign_1_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_19_internal_tensor_assign_1_end_mask_0 = const()[name = tensor("new_cache_19_internal_tensor_assign_1_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_19_internal_tensor_assign_1_squeeze_mask_0 = const()[name = tensor("new_cache_19_internal_tensor_assign_1_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_64 = const()[name = tensor("shape_64"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_18 = const()[name = tensor("reduce_prod_18"), val = tensor(1048576)]; + tensor range_1d_18_start_0 = const()[name = tensor("range_1d_18_start_0"), val = tensor(0)]; + tensor range_1d_18_step_0 = const()[name = tensor("range_1d_18_step_0"), val = tensor(1)]; + tensor range_1d_18 = range_1d(end = reduce_prod_18, start = range_1d_18_start_0, step = range_1d_18_step_0)[name = tensor("range_1d_18")]; + tensor reshape_90 = reshape(shape = shape_64, x = range_1d_18)[name = tensor("reshape_90")]; + tensor slice_by_index_18 = slice_by_index(begin = concat_74, begin_mask = new_cache_19_internal_tensor_assign_1_begin_mask_0, end = concat_75, end_mask = new_cache_19_internal_tensor_assign_1_end_mask_0, squeeze_mask = new_cache_19_internal_tensor_assign_1_squeeze_mask_0, stride = new_cache_19_internal_tensor_assign_1_stride_0, x = reshape_90)[name = tensor("slice_by_index_18")]; + tensor reshape_91_shape_0 = const()[name = tensor("reshape_91_shape_0"), val = tensor([-1])]; + tensor reshape_91 = reshape(shape = reshape_91_shape_0, x = slice_by_index_18)[name = tensor("reshape_91")]; + tensor reshape_92_shape_0 = const()[name = tensor("reshape_92_shape_0"), val = tensor([-1])]; + tensor reshape_92_cast_fp16 = reshape(shape = reshape_92_shape_0, x = var_4141_cast_fp16)[name = tensor("reshape_92_cast_fp16")]; + tensor reshape_93_shape_0 = const()[name = tensor("reshape_93_shape_0"), val = tensor([-1])]; + tensor reshape_93_cast_fp16 = reshape(shape = reshape_93_shape_0, x = cache9_to_fp16)[name = tensor("reshape_93_cast_fp16")]; + tensor scatter_18_mode_0 = const()[name = tensor("scatter_18_mode_0"), val = tensor("update")]; + tensor scatter_18_axis_0 = const()[name = tensor("scatter_18_axis_0"), val = tensor(0)]; + tensor scatter_18_validate_indices_0 = const()[name = tensor("scatter_18_validate_indices_0"), val = tensor(false)]; + tensor scatter_18_cast_fp16 = scatter(axis = scatter_18_axis_0, data = reshape_93_cast_fp16, indices = reshape_91, mode = scatter_18_mode_0, updates = reshape_92_cast_fp16, validate_indices = scatter_18_validate_indices_0)[name = tensor("scatter_18_cast_fp16")]; + tensor reshape_94_cast_fp16 = reshape(shape = shape_64, x = scatter_18_cast_fp16)[name = tensor("reshape_94_cast_fp16")]; + tensor var_4149_begin_0 = const()[name = tensor("op_4149_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor var_4149_end_0 = const()[name = tensor("op_4149_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor var_4149_end_mask_0 = const()[name = tensor("op_4149_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_4149_squeeze_mask_0 = const()[name = tensor("op_4149_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor var_4149_cast_fp16 = slice_by_index(begin = var_4149_begin_0, end = var_4149_end_0, end_mask = var_4149_end_mask_0, squeeze_mask = var_4149_squeeze_mask_0, x = reshape_94_cast_fp16)[name = tensor("op_4149_cast_fp16")]; + tensor var_4151_axis_0 = const()[name = tensor("op_4151_axis_0"), val = tensor(1)]; + tensor var_4151_mode_0 = const()[name = tensor("op_4151_mode_0"), val = tensor("update")]; + tensor var_4151_validate_indices_0 = const()[name = tensor("op_4151_validate_indices_0"), val = tensor(false)]; + tensor var_4151_cast_fp16 = scatter_along_axis(axis = var_4151_axis_0, data = var_4149_cast_fp16, indices = write_indices_19, mode = var_4151_mode_0, updates = v_19_cast_fp16, validate_indices = var_4151_validate_indices_0)[name = tensor("op_4151_cast_fp16")]; + tensor concat_76 = const()[name = tensor("concat_76"), val = tensor([1, 0, 0, 0, 0])]; + tensor concat_77 = const()[name = tensor("concat_77"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_19_internal_tensor_assign_2_stride_0 = const()[name = tensor("new_cache_19_internal_tensor_assign_2_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_19_internal_tensor_assign_2_begin_mask_0 = const()[name = tensor("new_cache_19_internal_tensor_assign_2_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_19_internal_tensor_assign_2_end_mask_0 = const()[name = tensor("new_cache_19_internal_tensor_assign_2_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_19_internal_tensor_assign_2_squeeze_mask_0 = const()[name = tensor("new_cache_19_internal_tensor_assign_2_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_65 = const()[name = tensor("shape_65"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_19 = const()[name = tensor("reduce_prod_19"), val = tensor(1048576)]; + tensor range_1d_19_start_0 = const()[name = tensor("range_1d_19_start_0"), val = tensor(0)]; + tensor range_1d_19_step_0 = const()[name = tensor("range_1d_19_step_0"), val = tensor(1)]; + tensor range_1d_19 = range_1d(end = reduce_prod_19, start = range_1d_19_start_0, step = range_1d_19_step_0)[name = tensor("range_1d_19")]; + tensor reshape_95 = reshape(shape = shape_65, x = range_1d_19)[name = tensor("reshape_95")]; + tensor slice_by_index_19 = slice_by_index(begin = concat_76, begin_mask = new_cache_19_internal_tensor_assign_2_begin_mask_0, end = concat_77, end_mask = new_cache_19_internal_tensor_assign_2_end_mask_0, squeeze_mask = new_cache_19_internal_tensor_assign_2_squeeze_mask_0, stride = new_cache_19_internal_tensor_assign_2_stride_0, x = reshape_95)[name = tensor("slice_by_index_19")]; + tensor reshape_96_shape_0 = const()[name = tensor("reshape_96_shape_0"), val = tensor([-1])]; + tensor reshape_96 = reshape(shape = reshape_96_shape_0, x = slice_by_index_19)[name = tensor("reshape_96")]; + tensor reshape_97_shape_0 = const()[name = tensor("reshape_97_shape_0"), val = tensor([-1])]; + tensor reshape_97_cast_fp16 = reshape(shape = reshape_97_shape_0, x = var_4151_cast_fp16)[name = tensor("reshape_97_cast_fp16")]; + tensor reshape_98_shape_0 = const()[name = tensor("reshape_98_shape_0"), val = tensor([-1])]; + tensor reshape_98_cast_fp16 = reshape(shape = reshape_98_shape_0, x = reshape_94_cast_fp16)[name = tensor("reshape_98_cast_fp16")]; + tensor scatter_19_mode_0 = const()[name = tensor("scatter_19_mode_0"), val = tensor("update")]; + tensor scatter_19_axis_0 = const()[name = tensor("scatter_19_axis_0"), val = tensor(0)]; + tensor scatter_19_validate_indices_0 = const()[name = tensor("scatter_19_validate_indices_0"), val = tensor(false)]; + tensor scatter_19_cast_fp16 = scatter(axis = scatter_19_axis_0, data = reshape_98_cast_fp16, indices = reshape_96, mode = scatter_19_mode_0, updates = reshape_97_cast_fp16, validate_indices = scatter_19_validate_indices_0)[name = tensor("scatter_19_cast_fp16")]; + tensor reshape_99_cast_fp16 = reshape(shape = shape_65, x = scatter_19_cast_fp16)[name = tensor("reshape_99_cast_fp16")]; + tensor reshape_99_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("reshape_99_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor keys_55_begin_0 = const()[name = tensor("keys_55_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor keys_55_end_0 = const()[name = tensor("keys_55_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor keys_55_end_mask_0 = const()[name = tensor("keys_55_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor keys_55_squeeze_mask_0 = const()[name = tensor("keys_55_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor keys_55_cast_fp16 = slice_by_index(begin = keys_55_begin_0, end = keys_55_end_0, end_mask = keys_55_end_mask_0, squeeze_mask = keys_55_squeeze_mask_0, x = reshape_99_cast_fp16)[name = tensor("keys_55_cast_fp16")]; + tensor values_55_begin_0 = const()[name = tensor("values_55_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor values_55_end_0 = const()[name = tensor("values_55_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor values_55_end_mask_0 = const()[name = tensor("values_55_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor values_55_squeeze_mask_0 = const()[name = tensor("values_55_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor values_55_cast_fp16 = slice_by_index(begin = values_55_begin_0, end = values_55_end_0, end_mask = values_55_end_mask_0, squeeze_mask = values_55_squeeze_mask_0, x = reshape_99_cast_fp16)[name = tensor("values_55_cast_fp16")]; + tensor var_4163_cast_fp16 = not_equal(x = keys_55_cast_fp16, y = keys_55_cast_fp16)[name = tensor("op_4163_cast_fp16")]; + tensor keys_57_cast_fp16 = select(a = var_470_to_fp16, b = keys_55_cast_fp16, cond = var_4163_cast_fp16)[name = tensor("keys_57_cast_fp16")]; + tensor var_4171_cast_fp16 = not_equal(x = values_55_cast_fp16, y = values_55_cast_fp16)[name = tensor("op_4171_cast_fp16")]; + tensor values_57_cast_fp16 = select(a = var_470_to_fp16, b = values_55_cast_fp16, cond = var_4171_cast_fp16)[name = tensor("values_57_cast_fp16")]; + tensor var_4195 = const()[name = tensor("op_4195"), val = tensor([0, 2, 1, 3])]; + tensor var_4211 = const()[name = tensor("op_4211"), val = tensor([1, 1, 1])]; + tensor var_4212_cast_fp16 = reshape(shape = var_4211, x = position9_to_fp16)[name = tensor("op_4212_cast_fp16")]; + tensor q_positions_19_cast_fp16 = add(x = var_4212_cast_fp16, y = q_offsets_1_promoted_to_fp16)[name = tensor("q_positions_19_cast_fp16")]; + tensor valid_end_19_cast_fp16 = add(x = var_4212_cast_fp16, y = var_536_cast_fp16)[name = tensor("valid_end_19_cast_fp16")]; + tensor valid_mask_19_cast_fp16 = less(x = k_positions_1_promoted_to_fp16, y = valid_end_19_cast_fp16)[name = tensor("valid_mask_19_cast_fp16")]; + tensor causal_mask_19_cast_fp16 = less_equal(x = k_positions_1_promoted_to_fp16, y = q_positions_19_cast_fp16)[name = tensor("causal_mask_19_cast_fp16")]; + tensor var_4240 = logical_and(x = valid_mask_19_cast_fp16, y = causal_mask_19_cast_fp16)[name = tensor("op_4240")]; + tensor attn_mask_19_axes_0 = const()[name = tensor("attn_mask_19_axes_0"), val = tensor([1])]; + tensor attn_mask_19 = expand_dims(axes = attn_mask_19_axes_0, x = var_4240)[name = tensor("attn_mask_19")]; + tensor var_4254_transpose_x_0 = const()[name = tensor("op_4254_transpose_x_0"), val = tensor(false)]; + tensor var_4254_transpose_y_0 = const()[name = tensor("op_4254_transpose_y_0"), val = tensor(false)]; + tensor transpose_87_perm_0 = const()[name = tensor("transpose_87_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_88_perm_0 = const()[name = tensor("transpose_88_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_88 = transpose(perm = transpose_88_perm_0, x = keys_57_cast_fp16)[name = tensor("transpose_168")]; + tensor transpose_87 = transpose(perm = transpose_87_perm_0, x = q_57_cast_fp16)[name = tensor("transpose_169")]; + tensor var_4254_cast_fp16 = matmul(transpose_x = var_4254_transpose_x_0, transpose_y = var_4254_transpose_y_0, x = transpose_87, y = transpose_88)[name = tensor("op_4254_cast_fp16")]; + tensor var_4248_to_fp16 = const()[name = tensor("op_4248_to_fp16"), val = tensor([0x1p-3])]; + tensor attn_weights_55_cast_fp16 = mul(x = var_4254_cast_fp16, y = var_4248_to_fp16)[name = tensor("attn_weights_55_cast_fp16")]; + tensor var_4256 = logical_not(x = attn_mask_19)[name = tensor("op_4256")]; + tensor var_4257_to_fp16 = const()[name = tensor("op_4257_to_fp16"), val = tensor(-inf)]; + tensor attn_weights_57_cast_fp16 = select(a = var_4257_to_fp16, b = attn_weights_55_cast_fp16, cond = var_4256)[name = tensor("attn_weights_57_cast_fp16")]; + tensor var_4259 = const()[name = tensor("op_4259"), val = tensor(-1)]; + tensor attn_weights_59_cast_fp16 = softmax(axis = var_4259, x = attn_weights_57_cast_fp16)[name = tensor("attn_weights_59_cast_fp16")]; + tensor attn_output_19_transpose_x_0 = const()[name = tensor("attn_output_19_transpose_x_0"), val = tensor(false)]; + tensor attn_output_19_transpose_y_0 = const()[name = tensor("attn_output_19_transpose_y_0"), val = tensor(false)]; + tensor values_59_cast_fp16 = transpose(perm = var_4195, x = values_57_cast_fp16)[name = tensor("transpose_170")]; + tensor attn_output_19_cast_fp16 = matmul(transpose_x = attn_output_19_transpose_x_0, transpose_y = attn_output_19_transpose_y_0, x = attn_weights_59_cast_fp16, y = values_59_cast_fp16)[name = tensor("attn_output_19_cast_fp16")]; + tensor var_4267 = const()[name = tensor("op_4267"), val = tensor([0, 2, 1, 3])]; + tensor var_4272 = const()[name = tensor("op_4272"), val = tensor([1, 256, 1024])]; + tensor var_4268_cast_fp16 = transpose(perm = var_4267, x = attn_output_19_cast_fp16)[name = tensor("transpose_167")]; + tensor input_113_cast_fp16 = reshape(shape = var_4272, x = var_4268_cast_fp16)[name = tensor("input_113_cast_fp16")]; + tensor attn9_out_proj_weight_to_fp16 = const()[name = tensor("attn9_out_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(233935872)))]; + tensor linear_37_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = attn9_out_proj_weight_to_fp16, x = input_113_cast_fp16)[name = tensor("linear_37_cast_fp16")]; + tensor var_4281_cast_fp16 = add(x = position9_to_fp16, y = valid_len_to_fp16)[name = tensor("op_4281_cast_fp16")]; + tensor var_4281_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_4281_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor input_115_cast_fp16 = add(x = input_109_cast_fp16, y = linear_37_cast_fp16)[name = tensor("input_115_cast_fp16")]; + tensor input_117_axes_0 = const()[name = tensor("input_117_axes_0"), val = tensor([-1])]; + tensor norm9_2_weight_to_fp16 = const()[name = tensor("norm9_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(236033088)))]; + tensor norm9_2_bias_to_fp16 = const()[name = tensor("norm9_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(236035200)))]; + tensor var_4285_to_fp16 = const()[name = tensor("op_4285_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_117_cast_fp16 = layer_norm(axes = input_117_axes_0, beta = norm9_2_bias_to_fp16, epsilon = var_4285_to_fp16, gamma = norm9_2_weight_to_fp16, x = input_115_cast_fp16)[name = tensor("input_117_cast_fp16")]; + tensor linear9_1_weight_to_fp16 = const()[name = tensor("linear9_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(236037312)))]; + tensor linear_38_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = linear9_1_weight_to_fp16, x = input_117_cast_fp16)[name = tensor("linear_38_cast_fp16")]; + tensor input_119_mode_0 = const()[name = tensor("input_119_mode_0"), val = tensor("EXACT")]; + tensor input_119_cast_fp16 = gelu(mode = input_119_mode_0, x = linear_38_cast_fp16)[name = tensor("input_119_cast_fp16")]; + tensor linear9_2_weight_to_fp16 = const()[name = tensor("linear9_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(244425984)))]; + tensor linear_39_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = linear9_2_weight_to_fp16, x = input_119_cast_fp16)[name = tensor("linear_39_cast_fp16")]; + tensor input_121_cast_fp16 = add(x = input_115_cast_fp16, y = linear_39_cast_fp16)[name = tensor("input_121_cast_fp16")]; + tensor input_123_axes_0 = const()[name = tensor("input_123_axes_0"), val = tensor([-1])]; + tensor norm10_1_weight_to_fp16 = const()[name = tensor("norm10_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(252814656)))]; + tensor norm10_1_bias_to_fp16 = const()[name = tensor("norm10_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(252816768)))]; + tensor var_4302_to_fp16 = const()[name = tensor("op_4302_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_123_cast_fp16 = layer_norm(axes = input_123_axes_0, beta = norm10_1_bias_to_fp16, epsilon = var_4302_to_fp16, gamma = norm10_1_weight_to_fp16, x = input_121_cast_fp16)[name = tensor("input_123_cast_fp16")]; + tensor attn10_in_proj_weight_to_fp16 = const()[name = tensor("attn10_in_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(252818880)))]; + tensor linear_40_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = attn10_in_proj_weight_to_fp16, x = input_123_cast_fp16)[name = tensor("linear_40_cast_fp16")]; + tensor var_4316 = const()[name = tensor("op_4316"), val = tensor([1, 256, 3, 16, 64])]; + tensor qkv_21_cast_fp16 = reshape(shape = var_4316, x = linear_40_cast_fp16)[name = tensor("qkv_21_cast_fp16")]; + tensor q_61_begin_0 = const()[name = tensor("q_61_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor q_61_end_0 = const()[name = tensor("q_61_end_0"), val = tensor([1, 256, 1, 16, 64])]; + tensor q_61_end_mask_0 = const()[name = tensor("q_61_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor q_61_squeeze_mask_0 = const()[name = tensor("q_61_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor q_61_cast_fp16 = slice_by_index(begin = q_61_begin_0, end = q_61_end_0, end_mask = q_61_end_mask_0, squeeze_mask = q_61_squeeze_mask_0, x = qkv_21_cast_fp16)[name = tensor("q_61_cast_fp16")]; + tensor k_41_begin_0 = const()[name = tensor("k_41_begin_0"), val = tensor([0, 0, 1, 0, 0])]; + tensor k_41_end_0 = const()[name = tensor("k_41_end_0"), val = tensor([1, 256, 2, 16, 64])]; + tensor k_41_end_mask_0 = const()[name = tensor("k_41_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor k_41_squeeze_mask_0 = const()[name = tensor("k_41_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor k_41_cast_fp16 = slice_by_index(begin = k_41_begin_0, end = k_41_end_0, end_mask = k_41_end_mask_0, squeeze_mask = k_41_squeeze_mask_0, x = qkv_21_cast_fp16)[name = tensor("k_41_cast_fp16")]; + tensor v_21_begin_0 = const()[name = tensor("v_21_begin_0"), val = tensor([0, 0, 2, 0, 0])]; + tensor v_21_end_0 = const()[name = tensor("v_21_end_0"), val = tensor([1, 256, 3, 16, 64])]; + tensor v_21_end_mask_0 = const()[name = tensor("v_21_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor v_21_squeeze_mask_0 = const()[name = tensor("v_21_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor v_21_cast_fp16 = slice_by_index(begin = v_21_begin_0, end = v_21_end_0, end_mask = v_21_end_mask_0, squeeze_mask = v_21_squeeze_mask_0, x = qkv_21_cast_fp16)[name = tensor("v_21_cast_fp16")]; + tensor var_4374 = const()[name = tensor("op_4374"), val = tensor([1, 1])]; + tensor position10_to_fp16_dtype_0 = const()[name = tensor("position10_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor position10_to_fp16 = cast(dtype = position10_to_fp16_dtype_0, x = position10)[name = tensor("cast_611")]; + tensor var_4375_cast_fp16 = reshape(shape = var_4374, x = position10_to_fp16)[name = tensor("op_4375_cast_fp16")]; + tensor ts_63_cast_fp16 = add(x = ts_1_promoted_to_fp16, y = var_4375_cast_fp16)[name = tensor("ts_63_cast_fp16")]; + tensor var_4382 = const()[name = tensor("op_4382"), val = tensor([1, 256, 1, 1])]; + tensor ts_65_cast_fp16 = reshape(shape = var_4382, x = ts_63_cast_fp16)[name = tensor("ts_65_cast_fp16")]; + tensor var_4389 = const()[name = tensor("op_4389"), val = tensor([1, 256, 16, 32, 2])]; + tensor q_complex_21_cast_fp16 = reshape(shape = var_4389, x = q_61_cast_fp16)[name = tensor("q_complex_21_cast_fp16")]; + tensor var_4396 = const()[name = tensor("op_4396"), val = tensor([1, 256, 16, 32, 2])]; + tensor k_complex_21_cast_fp16 = reshape(shape = var_4396, x = k_41_cast_fp16)[name = tensor("k_complex_21_cast_fp16")]; + tensor var_4400_begin_0 = const()[name = tensor("op_4400_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_4400_end_0 = const()[name = tensor("op_4400_end_0"), val = tensor([1, 256, 16, 32, 1])]; + tensor var_4400_end_mask_0 = const()[name = tensor("op_4400_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_4400_squeeze_mask_0 = const()[name = tensor("op_4400_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_4400_cast_fp16 = slice_by_index(begin = var_4400_begin_0, end = var_4400_end_0, end_mask = var_4400_end_mask_0, squeeze_mask = var_4400_squeeze_mask_0, x = q_complex_21_cast_fp16)[name = tensor("op_4400_cast_fp16")]; + tensor var_4408_begin_0 = const()[name = tensor("op_4408_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_4408_end_0 = const()[name = tensor("op_4408_end_0"), val = tensor([1, 256, 16, 32, 2])]; + tensor var_4408_end_mask_0 = const()[name = tensor("op_4408_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_4408_squeeze_mask_0 = const()[name = tensor("op_4408_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_4408_cast_fp16 = slice_by_index(begin = var_4408_begin_0, end = var_4408_end_0, end_mask = var_4408_end_mask_0, squeeze_mask = var_4408_squeeze_mask_0, x = q_complex_21_cast_fp16)[name = tensor("op_4408_cast_fp16")]; + tensor var_4416_begin_0 = const()[name = tensor("op_4416_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_4416_end_0 = const()[name = tensor("op_4416_end_0"), val = tensor([1, 256, 16, 32, 1])]; + tensor var_4416_end_mask_0 = const()[name = tensor("op_4416_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_4416_squeeze_mask_0 = const()[name = tensor("op_4416_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_4416_cast_fp16 = slice_by_index(begin = var_4416_begin_0, end = var_4416_end_0, end_mask = var_4416_end_mask_0, squeeze_mask = var_4416_squeeze_mask_0, x = k_complex_21_cast_fp16)[name = tensor("op_4416_cast_fp16")]; + tensor var_4424_begin_0 = const()[name = tensor("op_4424_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_4424_end_0 = const()[name = tensor("op_4424_end_0"), val = tensor([1, 256, 16, 32, 2])]; + tensor var_4424_end_mask_0 = const()[name = tensor("op_4424_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_4424_squeeze_mask_0 = const()[name = tensor("op_4424_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_4424_cast_fp16 = slice_by_index(begin = var_4424_begin_0, end = var_4424_end_0, end_mask = var_4424_end_mask_0, squeeze_mask = var_4424_squeeze_mask_0, x = k_complex_21_cast_fp16)[name = tensor("op_4424_cast_fp16")]; + tensor freqs_21_to_fp16 = const()[name = tensor("freqs_21_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(259110400)))]; + tensor var_4430_cast_fp16 = mul(x = freqs_21_to_fp16, y = ts_65_cast_fp16)[name = tensor("op_4430_cast_fp16")]; + tensor rotr_21_cast_fp16 = cos(x = var_4430_cast_fp16)[name = tensor("rotr_21_cast_fp16")]; + tensor roti_21_cast_fp16 = sin(x = var_4430_cast_fp16)[name = tensor("roti_21_cast_fp16")]; + tensor var_4434_cast_fp16 = mul(x = var_4400_cast_fp16, y = rotr_21_cast_fp16)[name = tensor("op_4434_cast_fp16")]; + tensor var_4435_cast_fp16 = mul(x = var_4408_cast_fp16, y = roti_21_cast_fp16)[name = tensor("op_4435_cast_fp16")]; + tensor qor_41_cast_fp16 = sub(x = var_4434_cast_fp16, y = var_4435_cast_fp16)[name = tensor("qor_41_cast_fp16")]; + tensor var_4438_cast_fp16 = mul(x = var_4400_cast_fp16, y = roti_21_cast_fp16)[name = tensor("op_4438_cast_fp16")]; + tensor var_4439_cast_fp16 = mul(x = var_4408_cast_fp16, y = rotr_21_cast_fp16)[name = tensor("op_4439_cast_fp16")]; + tensor qoi_41_cast_fp16 = add(x = var_4438_cast_fp16, y = var_4439_cast_fp16)[name = tensor("qoi_41_cast_fp16")]; + tensor var_4442_cast_fp16 = mul(x = var_4416_cast_fp16, y = rotr_21_cast_fp16)[name = tensor("op_4442_cast_fp16")]; + tensor var_4443_cast_fp16 = mul(x = var_4424_cast_fp16, y = roti_21_cast_fp16)[name = tensor("op_4443_cast_fp16")]; + tensor kor_41_cast_fp16 = sub(x = var_4442_cast_fp16, y = var_4443_cast_fp16)[name = tensor("kor_41_cast_fp16")]; + tensor var_4446_cast_fp16 = mul(x = var_4416_cast_fp16, y = roti_21_cast_fp16)[name = tensor("op_4446_cast_fp16")]; + tensor var_4447_cast_fp16 = mul(x = var_4424_cast_fp16, y = rotr_21_cast_fp16)[name = tensor("op_4447_cast_fp16")]; + tensor koi_41_cast_fp16 = add(x = var_4446_cast_fp16, y = var_4447_cast_fp16)[name = tensor("koi_41_cast_fp16")]; + tensor qo_21_axis_0 = const()[name = tensor("qo_21_axis_0"), val = tensor(-1)]; + tensor qo_21_cast_fp16 = stack(axis = qo_21_axis_0, values = (qor_41_cast_fp16, qoi_41_cast_fp16))[name = tensor("qo_21_cast_fp16")]; + tensor ko_21_axis_0 = const()[name = tensor("ko_21_axis_0"), val = tensor(-1)]; + tensor ko_21_cast_fp16 = stack(axis = ko_21_axis_0, values = (kor_41_cast_fp16, koi_41_cast_fp16))[name = tensor("ko_21_cast_fp16")]; + tensor var_4480 = const()[name = tensor("op_4480"), val = tensor([1, 256, 16, 64])]; + tensor q_63_cast_fp16 = reshape(shape = var_4480, x = qo_21_cast_fp16)[name = tensor("q_63_cast_fp16")]; + tensor var_4486 = const()[name = tensor("op_4486"), val = tensor([1, 256, 16, 64])]; + tensor k_43_cast_fp16 = reshape(shape = var_4486, x = ko_21_cast_fp16)[name = tensor("k_43_cast_fp16")]; + tensor write_indices_float_61_cast_fp16 = add(x = var_4375_cast_fp16, y = var_393_promoted_to_fp16)[name = tensor("write_indices_float_61_cast_fp16")]; + tensor _inversed_4507_y_0_to_fp16 = const()[name = tensor("_inversed_4507_y_0_to_fp16"), val = tensor(0x1p-9)]; + tensor _inversed_4507_cast_fp16 = mul(x = write_indices_float_61_cast_fp16, y = _inversed_4507_y_0_to_fp16)[name = tensor("_inversed_4507_cast_fp16")]; + tensor var_4508_cast_fp16 = floor(x = _inversed_4507_cast_fp16)[name = tensor("op_4508_cast_fp16")]; + tensor var_4509_to_fp16 = const()[name = tensor("op_4509_to_fp16"), val = tensor(0x1p+9)]; + tensor var_4510_cast_fp16 = mul(x = var_4508_cast_fp16, y = var_4509_to_fp16)[name = tensor("op_4510_cast_fp16")]; + tensor write_indices_float_63_cast_fp16 = sub(x = write_indices_float_61_cast_fp16, y = var_4510_cast_fp16)[name = tensor("write_indices_float_63_cast_fp16")]; + tensor var_4528_value_0_to_fp16 = const()[name = tensor("op_4528_value_0_to_fp16"), val = tensor(0x1.ffp+8)]; + tensor var_4528_cast_fp16 = fill_like(ref_tensor = write_indices_float_63_cast_fp16, value = var_4528_value_0_to_fp16)[name = tensor("op_4528_cast_fp16")]; + tensor write_indices_float_65_cast_fp16 = select(a = write_indices_float_63_cast_fp16, b = var_4528_cast_fp16, cond = token_valid_1_cast_fp16)[name = tensor("write_indices_float_65_cast_fp16")]; + tensor var_4534_dtype_0 = const()[name = tensor("op_4534_dtype_0"), val = tensor("int32")]; + tensor var_4539 = const()[name = tensor("op_4539"), val = tensor([1, 256, 1, 1])]; + tensor write_indices_float_65_cast_fp16_to_int32 = cast(dtype = var_4534_dtype_0, x = write_indices_float_65_cast_fp16)[name = tensor("cast_610")]; + tensor var_4540 = reshape(shape = var_4539, x = write_indices_float_65_cast_fp16_to_int32)[name = tensor("op_4540")]; + tensor write_indices_21_reps_0 = const()[name = tensor("write_indices_21_reps_0"), val = tensor([1, 1, 16, 64])]; + tensor write_indices_21 = tile(reps = write_indices_21_reps_0, x = var_4540)[name = tensor("write_indices_21")]; + tensor var_4550_begin_0 = const()[name = tensor("op_4550_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_4550_end_0 = const()[name = tensor("op_4550_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor var_4550_end_mask_0 = const()[name = tensor("op_4550_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_4550_squeeze_mask_0 = const()[name = tensor("op_4550_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor cache10_to_fp16_dtype_0 = const()[name = tensor("cache10_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor cache10_to_fp16 = cast(dtype = cache10_to_fp16_dtype_0, x = cache10)[name = tensor("cast_609")]; + tensor var_4550_cast_fp16 = slice_by_index(begin = var_4550_begin_0, end = var_4550_end_0, end_mask = var_4550_end_mask_0, squeeze_mask = var_4550_squeeze_mask_0, x = cache10_to_fp16)[name = tensor("op_4550_cast_fp16")]; + tensor var_4552_axis_0 = const()[name = tensor("op_4552_axis_0"), val = tensor(1)]; + tensor var_4552_mode_0 = const()[name = tensor("op_4552_mode_0"), val = tensor("update")]; + tensor var_4552_validate_indices_0 = const()[name = tensor("op_4552_validate_indices_0"), val = tensor(false)]; + tensor var_4552_cast_fp16 = scatter_along_axis(axis = var_4552_axis_0, data = var_4550_cast_fp16, indices = write_indices_21, mode = var_4552_mode_0, updates = k_43_cast_fp16, validate_indices = var_4552_validate_indices_0)[name = tensor("op_4552_cast_fp16")]; + tensor concat_82 = const()[name = tensor("concat_82"), val = tensor([0, 0, 0, 0, 0])]; + tensor concat_83 = const()[name = tensor("concat_83"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_21_internal_tensor_assign_1_stride_0 = const()[name = tensor("new_cache_21_internal_tensor_assign_1_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_21_internal_tensor_assign_1_begin_mask_0 = const()[name = tensor("new_cache_21_internal_tensor_assign_1_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_21_internal_tensor_assign_1_end_mask_0 = const()[name = tensor("new_cache_21_internal_tensor_assign_1_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_21_internal_tensor_assign_1_squeeze_mask_0 = const()[name = tensor("new_cache_21_internal_tensor_assign_1_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_66 = const()[name = tensor("shape_66"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_20 = const()[name = tensor("reduce_prod_20"), val = tensor(1048576)]; + tensor range_1d_20_start_0 = const()[name = tensor("range_1d_20_start_0"), val = tensor(0)]; + tensor range_1d_20_step_0 = const()[name = tensor("range_1d_20_step_0"), val = tensor(1)]; + tensor range_1d_20 = range_1d(end = reduce_prod_20, start = range_1d_20_start_0, step = range_1d_20_step_0)[name = tensor("range_1d_20")]; + tensor reshape_100 = reshape(shape = shape_66, x = range_1d_20)[name = tensor("reshape_100")]; + tensor slice_by_index_20 = slice_by_index(begin = concat_82, begin_mask = new_cache_21_internal_tensor_assign_1_begin_mask_0, end = concat_83, end_mask = new_cache_21_internal_tensor_assign_1_end_mask_0, squeeze_mask = new_cache_21_internal_tensor_assign_1_squeeze_mask_0, stride = new_cache_21_internal_tensor_assign_1_stride_0, x = reshape_100)[name = tensor("slice_by_index_20")]; + tensor reshape_101_shape_0 = const()[name = tensor("reshape_101_shape_0"), val = tensor([-1])]; + tensor reshape_101 = reshape(shape = reshape_101_shape_0, x = slice_by_index_20)[name = tensor("reshape_101")]; + tensor reshape_102_shape_0 = const()[name = tensor("reshape_102_shape_0"), val = tensor([-1])]; + tensor reshape_102_cast_fp16 = reshape(shape = reshape_102_shape_0, x = var_4552_cast_fp16)[name = tensor("reshape_102_cast_fp16")]; + tensor reshape_103_shape_0 = const()[name = tensor("reshape_103_shape_0"), val = tensor([-1])]; + tensor reshape_103_cast_fp16 = reshape(shape = reshape_103_shape_0, x = cache10_to_fp16)[name = tensor("reshape_103_cast_fp16")]; + tensor scatter_20_mode_0 = const()[name = tensor("scatter_20_mode_0"), val = tensor("update")]; + tensor scatter_20_axis_0 = const()[name = tensor("scatter_20_axis_0"), val = tensor(0)]; + tensor scatter_20_validate_indices_0 = const()[name = tensor("scatter_20_validate_indices_0"), val = tensor(false)]; + tensor scatter_20_cast_fp16 = scatter(axis = scatter_20_axis_0, data = reshape_103_cast_fp16, indices = reshape_101, mode = scatter_20_mode_0, updates = reshape_102_cast_fp16, validate_indices = scatter_20_validate_indices_0)[name = tensor("scatter_20_cast_fp16")]; + tensor reshape_104_cast_fp16 = reshape(shape = shape_66, x = scatter_20_cast_fp16)[name = tensor("reshape_104_cast_fp16")]; + tensor var_4560_begin_0 = const()[name = tensor("op_4560_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor var_4560_end_0 = const()[name = tensor("op_4560_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor var_4560_end_mask_0 = const()[name = tensor("op_4560_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_4560_squeeze_mask_0 = const()[name = tensor("op_4560_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor var_4560_cast_fp16 = slice_by_index(begin = var_4560_begin_0, end = var_4560_end_0, end_mask = var_4560_end_mask_0, squeeze_mask = var_4560_squeeze_mask_0, x = reshape_104_cast_fp16)[name = tensor("op_4560_cast_fp16")]; + tensor var_4562_axis_0 = const()[name = tensor("op_4562_axis_0"), val = tensor(1)]; + tensor var_4562_mode_0 = const()[name = tensor("op_4562_mode_0"), val = tensor("update")]; + tensor var_4562_validate_indices_0 = const()[name = tensor("op_4562_validate_indices_0"), val = tensor(false)]; + tensor var_4562_cast_fp16 = scatter_along_axis(axis = var_4562_axis_0, data = var_4560_cast_fp16, indices = write_indices_21, mode = var_4562_mode_0, updates = v_21_cast_fp16, validate_indices = var_4562_validate_indices_0)[name = tensor("op_4562_cast_fp16")]; + tensor concat_84 = const()[name = tensor("concat_84"), val = tensor([1, 0, 0, 0, 0])]; + tensor concat_85 = const()[name = tensor("concat_85"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_21_internal_tensor_assign_2_stride_0 = const()[name = tensor("new_cache_21_internal_tensor_assign_2_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_21_internal_tensor_assign_2_begin_mask_0 = const()[name = tensor("new_cache_21_internal_tensor_assign_2_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_21_internal_tensor_assign_2_end_mask_0 = const()[name = tensor("new_cache_21_internal_tensor_assign_2_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_21_internal_tensor_assign_2_squeeze_mask_0 = const()[name = tensor("new_cache_21_internal_tensor_assign_2_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_67 = const()[name = tensor("shape_67"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_21 = const()[name = tensor("reduce_prod_21"), val = tensor(1048576)]; + tensor range_1d_21_start_0 = const()[name = tensor("range_1d_21_start_0"), val = tensor(0)]; + tensor range_1d_21_step_0 = const()[name = tensor("range_1d_21_step_0"), val = tensor(1)]; + tensor range_1d_21 = range_1d(end = reduce_prod_21, start = range_1d_21_start_0, step = range_1d_21_step_0)[name = tensor("range_1d_21")]; + tensor reshape_105 = reshape(shape = shape_67, x = range_1d_21)[name = tensor("reshape_105")]; + tensor slice_by_index_21 = slice_by_index(begin = concat_84, begin_mask = new_cache_21_internal_tensor_assign_2_begin_mask_0, end = concat_85, end_mask = new_cache_21_internal_tensor_assign_2_end_mask_0, squeeze_mask = new_cache_21_internal_tensor_assign_2_squeeze_mask_0, stride = new_cache_21_internal_tensor_assign_2_stride_0, x = reshape_105)[name = tensor("slice_by_index_21")]; + tensor reshape_106_shape_0 = const()[name = tensor("reshape_106_shape_0"), val = tensor([-1])]; + tensor reshape_106 = reshape(shape = reshape_106_shape_0, x = slice_by_index_21)[name = tensor("reshape_106")]; + tensor reshape_107_shape_0 = const()[name = tensor("reshape_107_shape_0"), val = tensor([-1])]; + tensor reshape_107_cast_fp16 = reshape(shape = reshape_107_shape_0, x = var_4562_cast_fp16)[name = tensor("reshape_107_cast_fp16")]; + tensor reshape_108_shape_0 = const()[name = tensor("reshape_108_shape_0"), val = tensor([-1])]; + tensor reshape_108_cast_fp16 = reshape(shape = reshape_108_shape_0, x = reshape_104_cast_fp16)[name = tensor("reshape_108_cast_fp16")]; + tensor scatter_21_mode_0 = const()[name = tensor("scatter_21_mode_0"), val = tensor("update")]; + tensor scatter_21_axis_0 = const()[name = tensor("scatter_21_axis_0"), val = tensor(0)]; + tensor scatter_21_validate_indices_0 = const()[name = tensor("scatter_21_validate_indices_0"), val = tensor(false)]; + tensor scatter_21_cast_fp16 = scatter(axis = scatter_21_axis_0, data = reshape_108_cast_fp16, indices = reshape_106, mode = scatter_21_mode_0, updates = reshape_107_cast_fp16, validate_indices = scatter_21_validate_indices_0)[name = tensor("scatter_21_cast_fp16")]; + tensor reshape_109_cast_fp16 = reshape(shape = shape_67, x = scatter_21_cast_fp16)[name = tensor("reshape_109_cast_fp16")]; + tensor reshape_109_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("reshape_109_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor keys_61_begin_0 = const()[name = tensor("keys_61_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor keys_61_end_0 = const()[name = tensor("keys_61_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor keys_61_end_mask_0 = const()[name = tensor("keys_61_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor keys_61_squeeze_mask_0 = const()[name = tensor("keys_61_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor keys_61_cast_fp16 = slice_by_index(begin = keys_61_begin_0, end = keys_61_end_0, end_mask = keys_61_end_mask_0, squeeze_mask = keys_61_squeeze_mask_0, x = reshape_109_cast_fp16)[name = tensor("keys_61_cast_fp16")]; + tensor values_61_begin_0 = const()[name = tensor("values_61_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor values_61_end_0 = const()[name = tensor("values_61_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor values_61_end_mask_0 = const()[name = tensor("values_61_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor values_61_squeeze_mask_0 = const()[name = tensor("values_61_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor values_61_cast_fp16 = slice_by_index(begin = values_61_begin_0, end = values_61_end_0, end_mask = values_61_end_mask_0, squeeze_mask = values_61_squeeze_mask_0, x = reshape_109_cast_fp16)[name = tensor("values_61_cast_fp16")]; + tensor var_4574_cast_fp16 = not_equal(x = keys_61_cast_fp16, y = keys_61_cast_fp16)[name = tensor("op_4574_cast_fp16")]; + tensor keys_63_cast_fp16 = select(a = var_470_to_fp16, b = keys_61_cast_fp16, cond = var_4574_cast_fp16)[name = tensor("keys_63_cast_fp16")]; + tensor var_4582_cast_fp16 = not_equal(x = values_61_cast_fp16, y = values_61_cast_fp16)[name = tensor("op_4582_cast_fp16")]; + tensor values_63_cast_fp16 = select(a = var_470_to_fp16, b = values_61_cast_fp16, cond = var_4582_cast_fp16)[name = tensor("values_63_cast_fp16")]; + tensor var_4606 = const()[name = tensor("op_4606"), val = tensor([0, 2, 1, 3])]; + tensor var_4622 = const()[name = tensor("op_4622"), val = tensor([1, 1, 1])]; + tensor var_4623_cast_fp16 = reshape(shape = var_4622, x = position10_to_fp16)[name = tensor("op_4623_cast_fp16")]; + tensor q_positions_21_cast_fp16 = add(x = var_4623_cast_fp16, y = q_offsets_1_promoted_to_fp16)[name = tensor("q_positions_21_cast_fp16")]; + tensor valid_end_21_cast_fp16 = add(x = var_4623_cast_fp16, y = var_536_cast_fp16)[name = tensor("valid_end_21_cast_fp16")]; + tensor valid_mask_21_cast_fp16 = less(x = k_positions_1_promoted_to_fp16, y = valid_end_21_cast_fp16)[name = tensor("valid_mask_21_cast_fp16")]; + tensor causal_mask_21_cast_fp16 = less_equal(x = k_positions_1_promoted_to_fp16, y = q_positions_21_cast_fp16)[name = tensor("causal_mask_21_cast_fp16")]; + tensor var_4651 = logical_and(x = valid_mask_21_cast_fp16, y = causal_mask_21_cast_fp16)[name = tensor("op_4651")]; + tensor attn_mask_21_axes_0 = const()[name = tensor("attn_mask_21_axes_0"), val = tensor([1])]; + tensor attn_mask_21 = expand_dims(axes = attn_mask_21_axes_0, x = var_4651)[name = tensor("attn_mask_21")]; + tensor var_4665_transpose_x_0 = const()[name = tensor("op_4665_transpose_x_0"), val = tensor(false)]; + tensor var_4665_transpose_y_0 = const()[name = tensor("op_4665_transpose_y_0"), val = tensor(false)]; + tensor transpose_89_perm_0 = const()[name = tensor("transpose_89_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_90_perm_0 = const()[name = tensor("transpose_90_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_90 = transpose(perm = transpose_90_perm_0, x = keys_63_cast_fp16)[name = tensor("transpose_164")]; + tensor transpose_89 = transpose(perm = transpose_89_perm_0, x = q_63_cast_fp16)[name = tensor("transpose_165")]; + tensor var_4665_cast_fp16 = matmul(transpose_x = var_4665_transpose_x_0, transpose_y = var_4665_transpose_y_0, x = transpose_89, y = transpose_90)[name = tensor("op_4665_cast_fp16")]; + tensor var_4659_to_fp16 = const()[name = tensor("op_4659_to_fp16"), val = tensor([0x1p-3])]; + tensor attn_weights_61_cast_fp16 = mul(x = var_4665_cast_fp16, y = var_4659_to_fp16)[name = tensor("attn_weights_61_cast_fp16")]; + tensor var_4667 = logical_not(x = attn_mask_21)[name = tensor("op_4667")]; + tensor var_4668_to_fp16 = const()[name = tensor("op_4668_to_fp16"), val = tensor(-inf)]; + tensor attn_weights_63_cast_fp16 = select(a = var_4668_to_fp16, b = attn_weights_61_cast_fp16, cond = var_4667)[name = tensor("attn_weights_63_cast_fp16")]; + tensor var_4670 = const()[name = tensor("op_4670"), val = tensor(-1)]; + tensor attn_weights_65_cast_fp16 = softmax(axis = var_4670, x = attn_weights_63_cast_fp16)[name = tensor("attn_weights_65_cast_fp16")]; + tensor attn_output_21_transpose_x_0 = const()[name = tensor("attn_output_21_transpose_x_0"), val = tensor(false)]; + tensor attn_output_21_transpose_y_0 = const()[name = tensor("attn_output_21_transpose_y_0"), val = tensor(false)]; + tensor values_65_cast_fp16 = transpose(perm = var_4606, x = values_63_cast_fp16)[name = tensor("transpose_166")]; + tensor attn_output_21_cast_fp16 = matmul(transpose_x = attn_output_21_transpose_x_0, transpose_y = attn_output_21_transpose_y_0, x = attn_weights_65_cast_fp16, y = values_65_cast_fp16)[name = tensor("attn_output_21_cast_fp16")]; + tensor var_4678 = const()[name = tensor("op_4678"), val = tensor([0, 2, 1, 3])]; + tensor var_4683 = const()[name = tensor("op_4683"), val = tensor([1, 256, 1024])]; + tensor var_4679_cast_fp16 = transpose(perm = var_4678, x = attn_output_21_cast_fp16)[name = tensor("transpose_163")]; + tensor input_125_cast_fp16 = reshape(shape = var_4683, x = var_4679_cast_fp16)[name = tensor("input_125_cast_fp16")]; + tensor attn10_out_proj_weight_to_fp16 = const()[name = tensor("attn10_out_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(259110528)))]; + tensor linear_41_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = attn10_out_proj_weight_to_fp16, x = input_125_cast_fp16)[name = tensor("linear_41_cast_fp16")]; + tensor var_4692_cast_fp16 = add(x = position10_to_fp16, y = valid_len_to_fp16)[name = tensor("op_4692_cast_fp16")]; + tensor var_4692_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_4692_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor input_127_cast_fp16 = add(x = input_121_cast_fp16, y = linear_41_cast_fp16)[name = tensor("input_127_cast_fp16")]; + tensor input_129_axes_0 = const()[name = tensor("input_129_axes_0"), val = tensor([-1])]; + tensor norm10_2_weight_to_fp16 = const()[name = tensor("norm10_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(261207744)))]; + tensor norm10_2_bias_to_fp16 = const()[name = tensor("norm10_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(261209856)))]; + tensor var_4696_to_fp16 = const()[name = tensor("op_4696_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_129_cast_fp16 = layer_norm(axes = input_129_axes_0, beta = norm10_2_bias_to_fp16, epsilon = var_4696_to_fp16, gamma = norm10_2_weight_to_fp16, x = input_127_cast_fp16)[name = tensor("input_129_cast_fp16")]; + tensor linear10_1_weight_to_fp16 = const()[name = tensor("linear10_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(261211968)))]; + tensor linear_42_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = linear10_1_weight_to_fp16, x = input_129_cast_fp16)[name = tensor("linear_42_cast_fp16")]; + tensor input_131_mode_0 = const()[name = tensor("input_131_mode_0"), val = tensor("EXACT")]; + tensor input_131_cast_fp16 = gelu(mode = input_131_mode_0, x = linear_42_cast_fp16)[name = tensor("input_131_cast_fp16")]; + tensor linear10_2_weight_to_fp16 = const()[name = tensor("linear10_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(269600640)))]; + tensor linear_43_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = linear10_2_weight_to_fp16, x = input_131_cast_fp16)[name = tensor("linear_43_cast_fp16")]; + tensor input_133_cast_fp16 = add(x = input_127_cast_fp16, y = linear_43_cast_fp16)[name = tensor("input_133_cast_fp16")]; + tensor input_135_axes_0 = const()[name = tensor("input_135_axes_0"), val = tensor([-1])]; + tensor norm11_1_weight_to_fp16 = const()[name = tensor("norm11_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(277989312)))]; + tensor norm11_1_bias_to_fp16 = const()[name = tensor("norm11_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(277991424)))]; + tensor var_4713_to_fp16 = const()[name = tensor("op_4713_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_135_cast_fp16 = layer_norm(axes = input_135_axes_0, beta = norm11_1_bias_to_fp16, epsilon = var_4713_to_fp16, gamma = norm11_1_weight_to_fp16, x = input_133_cast_fp16)[name = tensor("input_135_cast_fp16")]; + tensor attn11_in_proj_weight_to_fp16 = const()[name = tensor("attn11_in_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(277993536)))]; + tensor linear_44_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = attn11_in_proj_weight_to_fp16, x = input_135_cast_fp16)[name = tensor("linear_44_cast_fp16")]; + tensor var_4727 = const()[name = tensor("op_4727"), val = tensor([1, 256, 3, 16, 64])]; + tensor qkv_23_cast_fp16 = reshape(shape = var_4727, x = linear_44_cast_fp16)[name = tensor("qkv_23_cast_fp16")]; + tensor q_67_begin_0 = const()[name = tensor("q_67_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor q_67_end_0 = const()[name = tensor("q_67_end_0"), val = tensor([1, 256, 1, 16, 64])]; + tensor q_67_end_mask_0 = const()[name = tensor("q_67_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor q_67_squeeze_mask_0 = const()[name = tensor("q_67_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor q_67_cast_fp16 = slice_by_index(begin = q_67_begin_0, end = q_67_end_0, end_mask = q_67_end_mask_0, squeeze_mask = q_67_squeeze_mask_0, x = qkv_23_cast_fp16)[name = tensor("q_67_cast_fp16")]; + tensor k_45_begin_0 = const()[name = tensor("k_45_begin_0"), val = tensor([0, 0, 1, 0, 0])]; + tensor k_45_end_0 = const()[name = tensor("k_45_end_0"), val = tensor([1, 256, 2, 16, 64])]; + tensor k_45_end_mask_0 = const()[name = tensor("k_45_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor k_45_squeeze_mask_0 = const()[name = tensor("k_45_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor k_45_cast_fp16 = slice_by_index(begin = k_45_begin_0, end = k_45_end_0, end_mask = k_45_end_mask_0, squeeze_mask = k_45_squeeze_mask_0, x = qkv_23_cast_fp16)[name = tensor("k_45_cast_fp16")]; + tensor v_23_begin_0 = const()[name = tensor("v_23_begin_0"), val = tensor([0, 0, 2, 0, 0])]; + tensor v_23_end_0 = const()[name = tensor("v_23_end_0"), val = tensor([1, 256, 3, 16, 64])]; + tensor v_23_end_mask_0 = const()[name = tensor("v_23_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor v_23_squeeze_mask_0 = const()[name = tensor("v_23_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor v_23_cast_fp16 = slice_by_index(begin = v_23_begin_0, end = v_23_end_0, end_mask = v_23_end_mask_0, squeeze_mask = v_23_squeeze_mask_0, x = qkv_23_cast_fp16)[name = tensor("v_23_cast_fp16")]; + tensor var_4785 = const()[name = tensor("op_4785"), val = tensor([1, 1])]; + tensor position11_to_fp16_dtype_0 = const()[name = tensor("position11_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor position11_to_fp16 = cast(dtype = position11_to_fp16_dtype_0, x = position11)[name = tensor("cast_606")]; + tensor var_4786_cast_fp16 = reshape(shape = var_4785, x = position11_to_fp16)[name = tensor("op_4786_cast_fp16")]; + tensor ts_69_cast_fp16 = add(x = ts_1_promoted_to_fp16, y = var_4786_cast_fp16)[name = tensor("ts_69_cast_fp16")]; + tensor var_4793 = const()[name = tensor("op_4793"), val = tensor([1, 256, 1, 1])]; + tensor ts_71_cast_fp16 = reshape(shape = var_4793, x = ts_69_cast_fp16)[name = tensor("ts_71_cast_fp16")]; + tensor var_4800 = const()[name = tensor("op_4800"), val = tensor([1, 256, 16, 32, 2])]; + tensor q_complex_23_cast_fp16 = reshape(shape = var_4800, x = q_67_cast_fp16)[name = tensor("q_complex_23_cast_fp16")]; + tensor var_4807 = const()[name = tensor("op_4807"), val = tensor([1, 256, 16, 32, 2])]; + tensor k_complex_23_cast_fp16 = reshape(shape = var_4807, x = k_45_cast_fp16)[name = tensor("k_complex_23_cast_fp16")]; + tensor var_4811_begin_0 = const()[name = tensor("op_4811_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_4811_end_0 = const()[name = tensor("op_4811_end_0"), val = tensor([1, 256, 16, 32, 1])]; + tensor var_4811_end_mask_0 = const()[name = tensor("op_4811_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_4811_squeeze_mask_0 = const()[name = tensor("op_4811_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_4811_cast_fp16 = slice_by_index(begin = var_4811_begin_0, end = var_4811_end_0, end_mask = var_4811_end_mask_0, squeeze_mask = var_4811_squeeze_mask_0, x = q_complex_23_cast_fp16)[name = tensor("op_4811_cast_fp16")]; + tensor var_4819_begin_0 = const()[name = tensor("op_4819_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_4819_end_0 = const()[name = tensor("op_4819_end_0"), val = tensor([1, 256, 16, 32, 2])]; + tensor var_4819_end_mask_0 = const()[name = tensor("op_4819_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_4819_squeeze_mask_0 = const()[name = tensor("op_4819_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_4819_cast_fp16 = slice_by_index(begin = var_4819_begin_0, end = var_4819_end_0, end_mask = var_4819_end_mask_0, squeeze_mask = var_4819_squeeze_mask_0, x = q_complex_23_cast_fp16)[name = tensor("op_4819_cast_fp16")]; + tensor var_4827_begin_0 = const()[name = tensor("op_4827_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_4827_end_0 = const()[name = tensor("op_4827_end_0"), val = tensor([1, 256, 16, 32, 1])]; + tensor var_4827_end_mask_0 = const()[name = tensor("op_4827_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_4827_squeeze_mask_0 = const()[name = tensor("op_4827_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_4827_cast_fp16 = slice_by_index(begin = var_4827_begin_0, end = var_4827_end_0, end_mask = var_4827_end_mask_0, squeeze_mask = var_4827_squeeze_mask_0, x = k_complex_23_cast_fp16)[name = tensor("op_4827_cast_fp16")]; + tensor var_4835_begin_0 = const()[name = tensor("op_4835_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_4835_end_0 = const()[name = tensor("op_4835_end_0"), val = tensor([1, 256, 16, 32, 2])]; + tensor var_4835_end_mask_0 = const()[name = tensor("op_4835_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_4835_squeeze_mask_0 = const()[name = tensor("op_4835_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_4835_cast_fp16 = slice_by_index(begin = var_4835_begin_0, end = var_4835_end_0, end_mask = var_4835_end_mask_0, squeeze_mask = var_4835_squeeze_mask_0, x = k_complex_23_cast_fp16)[name = tensor("op_4835_cast_fp16")]; + tensor freqs_23_to_fp16 = const()[name = tensor("freqs_23_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(284285056)))]; + tensor var_4841_cast_fp16 = mul(x = freqs_23_to_fp16, y = ts_71_cast_fp16)[name = tensor("op_4841_cast_fp16")]; + tensor rotr_23_cast_fp16 = cos(x = var_4841_cast_fp16)[name = tensor("rotr_23_cast_fp16")]; + tensor roti_23_cast_fp16 = sin(x = var_4841_cast_fp16)[name = tensor("roti_23_cast_fp16")]; + tensor var_4845_cast_fp16 = mul(x = var_4811_cast_fp16, y = rotr_23_cast_fp16)[name = tensor("op_4845_cast_fp16")]; + tensor var_4846_cast_fp16 = mul(x = var_4819_cast_fp16, y = roti_23_cast_fp16)[name = tensor("op_4846_cast_fp16")]; + tensor qor_45_cast_fp16 = sub(x = var_4845_cast_fp16, y = var_4846_cast_fp16)[name = tensor("qor_45_cast_fp16")]; + tensor var_4849_cast_fp16 = mul(x = var_4811_cast_fp16, y = roti_23_cast_fp16)[name = tensor("op_4849_cast_fp16")]; + tensor var_4850_cast_fp16 = mul(x = var_4819_cast_fp16, y = rotr_23_cast_fp16)[name = tensor("op_4850_cast_fp16")]; + tensor qoi_45_cast_fp16 = add(x = var_4849_cast_fp16, y = var_4850_cast_fp16)[name = tensor("qoi_45_cast_fp16")]; + tensor var_4853_cast_fp16 = mul(x = var_4827_cast_fp16, y = rotr_23_cast_fp16)[name = tensor("op_4853_cast_fp16")]; + tensor var_4854_cast_fp16 = mul(x = var_4835_cast_fp16, y = roti_23_cast_fp16)[name = tensor("op_4854_cast_fp16")]; + tensor kor_45_cast_fp16 = sub(x = var_4853_cast_fp16, y = var_4854_cast_fp16)[name = tensor("kor_45_cast_fp16")]; + tensor var_4857_cast_fp16 = mul(x = var_4827_cast_fp16, y = roti_23_cast_fp16)[name = tensor("op_4857_cast_fp16")]; + tensor var_4858_cast_fp16 = mul(x = var_4835_cast_fp16, y = rotr_23_cast_fp16)[name = tensor("op_4858_cast_fp16")]; + tensor koi_45_cast_fp16 = add(x = var_4857_cast_fp16, y = var_4858_cast_fp16)[name = tensor("koi_45_cast_fp16")]; + tensor qo_23_axis_0 = const()[name = tensor("qo_23_axis_0"), val = tensor(-1)]; + tensor qo_23_cast_fp16 = stack(axis = qo_23_axis_0, values = (qor_45_cast_fp16, qoi_45_cast_fp16))[name = tensor("qo_23_cast_fp16")]; + tensor ko_23_axis_0 = const()[name = tensor("ko_23_axis_0"), val = tensor(-1)]; + tensor ko_23_cast_fp16 = stack(axis = ko_23_axis_0, values = (kor_45_cast_fp16, koi_45_cast_fp16))[name = tensor("ko_23_cast_fp16")]; + tensor var_4891 = const()[name = tensor("op_4891"), val = tensor([1, 256, 16, 64])]; + tensor q_69_cast_fp16 = reshape(shape = var_4891, x = qo_23_cast_fp16)[name = tensor("q_69_cast_fp16")]; + tensor var_4897 = const()[name = tensor("op_4897"), val = tensor([1, 256, 16, 64])]; + tensor k_47_cast_fp16 = reshape(shape = var_4897, x = ko_23_cast_fp16)[name = tensor("k_47_cast_fp16")]; + tensor write_indices_float_67_cast_fp16 = add(x = var_4786_cast_fp16, y = var_393_promoted_to_fp16)[name = tensor("write_indices_float_67_cast_fp16")]; + tensor _inversed_4918_y_0_to_fp16 = const()[name = tensor("_inversed_4918_y_0_to_fp16"), val = tensor(0x1p-9)]; + tensor _inversed_4918_cast_fp16 = mul(x = write_indices_float_67_cast_fp16, y = _inversed_4918_y_0_to_fp16)[name = tensor("_inversed_4918_cast_fp16")]; + tensor var_4919_cast_fp16 = floor(x = _inversed_4918_cast_fp16)[name = tensor("op_4919_cast_fp16")]; + tensor var_4920_to_fp16 = const()[name = tensor("op_4920_to_fp16"), val = tensor(0x1p+9)]; + tensor var_4921_cast_fp16 = mul(x = var_4919_cast_fp16, y = var_4920_to_fp16)[name = tensor("op_4921_cast_fp16")]; + tensor write_indices_float_69_cast_fp16 = sub(x = write_indices_float_67_cast_fp16, y = var_4921_cast_fp16)[name = tensor("write_indices_float_69_cast_fp16")]; + tensor var_4939_value_0_to_fp16 = const()[name = tensor("op_4939_value_0_to_fp16"), val = tensor(0x1.ffp+8)]; + tensor var_4939_cast_fp16 = fill_like(ref_tensor = write_indices_float_69_cast_fp16, value = var_4939_value_0_to_fp16)[name = tensor("op_4939_cast_fp16")]; + tensor write_indices_float_71_cast_fp16 = select(a = write_indices_float_69_cast_fp16, b = var_4939_cast_fp16, cond = token_valid_1_cast_fp16)[name = tensor("write_indices_float_71_cast_fp16")]; + tensor var_4945_dtype_0 = const()[name = tensor("op_4945_dtype_0"), val = tensor("int32")]; + tensor var_4950 = const()[name = tensor("op_4950"), val = tensor([1, 256, 1, 1])]; + tensor write_indices_float_71_cast_fp16_to_int32 = cast(dtype = var_4945_dtype_0, x = write_indices_float_71_cast_fp16)[name = tensor("cast_605")]; + tensor var_4951 = reshape(shape = var_4950, x = write_indices_float_71_cast_fp16_to_int32)[name = tensor("op_4951")]; + tensor write_indices_23_reps_0 = const()[name = tensor("write_indices_23_reps_0"), val = tensor([1, 1, 16, 64])]; + tensor write_indices_23 = tile(reps = write_indices_23_reps_0, x = var_4951)[name = tensor("write_indices_23")]; + tensor var_4961_begin_0 = const()[name = tensor("op_4961_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_4961_end_0 = const()[name = tensor("op_4961_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor var_4961_end_mask_0 = const()[name = tensor("op_4961_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_4961_squeeze_mask_0 = const()[name = tensor("op_4961_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor cache11_to_fp16_dtype_0 = const()[name = tensor("cache11_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor cache11_to_fp16 = cast(dtype = cache11_to_fp16_dtype_0, x = cache11)[name = tensor("cast_604")]; + tensor var_4961_cast_fp16 = slice_by_index(begin = var_4961_begin_0, end = var_4961_end_0, end_mask = var_4961_end_mask_0, squeeze_mask = var_4961_squeeze_mask_0, x = cache11_to_fp16)[name = tensor("op_4961_cast_fp16")]; + tensor var_4963_axis_0 = const()[name = tensor("op_4963_axis_0"), val = tensor(1)]; + tensor var_4963_mode_0 = const()[name = tensor("op_4963_mode_0"), val = tensor("update")]; + tensor var_4963_validate_indices_0 = const()[name = tensor("op_4963_validate_indices_0"), val = tensor(false)]; + tensor var_4963_cast_fp16 = scatter_along_axis(axis = var_4963_axis_0, data = var_4961_cast_fp16, indices = write_indices_23, mode = var_4963_mode_0, updates = k_47_cast_fp16, validate_indices = var_4963_validate_indices_0)[name = tensor("op_4963_cast_fp16")]; + tensor concat_90 = const()[name = tensor("concat_90"), val = tensor([0, 0, 0, 0, 0])]; + tensor concat_91 = const()[name = tensor("concat_91"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_23_internal_tensor_assign_1_stride_0 = const()[name = tensor("new_cache_23_internal_tensor_assign_1_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_23_internal_tensor_assign_1_begin_mask_0 = const()[name = tensor("new_cache_23_internal_tensor_assign_1_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_23_internal_tensor_assign_1_end_mask_0 = const()[name = tensor("new_cache_23_internal_tensor_assign_1_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_23_internal_tensor_assign_1_squeeze_mask_0 = const()[name = tensor("new_cache_23_internal_tensor_assign_1_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_68 = const()[name = tensor("shape_68"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_22 = const()[name = tensor("reduce_prod_22"), val = tensor(1048576)]; + tensor range_1d_22_start_0 = const()[name = tensor("range_1d_22_start_0"), val = tensor(0)]; + tensor range_1d_22_step_0 = const()[name = tensor("range_1d_22_step_0"), val = tensor(1)]; + tensor range_1d_22 = range_1d(end = reduce_prod_22, start = range_1d_22_start_0, step = range_1d_22_step_0)[name = tensor("range_1d_22")]; + tensor reshape_110 = reshape(shape = shape_68, x = range_1d_22)[name = tensor("reshape_110")]; + tensor slice_by_index_22 = slice_by_index(begin = concat_90, begin_mask = new_cache_23_internal_tensor_assign_1_begin_mask_0, end = concat_91, end_mask = new_cache_23_internal_tensor_assign_1_end_mask_0, squeeze_mask = new_cache_23_internal_tensor_assign_1_squeeze_mask_0, stride = new_cache_23_internal_tensor_assign_1_stride_0, x = reshape_110)[name = tensor("slice_by_index_22")]; + tensor reshape_111_shape_0 = const()[name = tensor("reshape_111_shape_0"), val = tensor([-1])]; + tensor reshape_111 = reshape(shape = reshape_111_shape_0, x = slice_by_index_22)[name = tensor("reshape_111")]; + tensor reshape_112_shape_0 = const()[name = tensor("reshape_112_shape_0"), val = tensor([-1])]; + tensor reshape_112_cast_fp16 = reshape(shape = reshape_112_shape_0, x = var_4963_cast_fp16)[name = tensor("reshape_112_cast_fp16")]; + tensor reshape_113_shape_0 = const()[name = tensor("reshape_113_shape_0"), val = tensor([-1])]; + tensor reshape_113_cast_fp16 = reshape(shape = reshape_113_shape_0, x = cache11_to_fp16)[name = tensor("reshape_113_cast_fp16")]; + tensor scatter_22_mode_0 = const()[name = tensor("scatter_22_mode_0"), val = tensor("update")]; + tensor scatter_22_axis_0 = const()[name = tensor("scatter_22_axis_0"), val = tensor(0)]; + tensor scatter_22_validate_indices_0 = const()[name = tensor("scatter_22_validate_indices_0"), val = tensor(false)]; + tensor scatter_22_cast_fp16 = scatter(axis = scatter_22_axis_0, data = reshape_113_cast_fp16, indices = reshape_111, mode = scatter_22_mode_0, updates = reshape_112_cast_fp16, validate_indices = scatter_22_validate_indices_0)[name = tensor("scatter_22_cast_fp16")]; + tensor reshape_114_cast_fp16 = reshape(shape = shape_68, x = scatter_22_cast_fp16)[name = tensor("reshape_114_cast_fp16")]; + tensor var_4971_begin_0 = const()[name = tensor("op_4971_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor var_4971_end_0 = const()[name = tensor("op_4971_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor var_4971_end_mask_0 = const()[name = tensor("op_4971_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_4971_squeeze_mask_0 = const()[name = tensor("op_4971_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor var_4971_cast_fp16 = slice_by_index(begin = var_4971_begin_0, end = var_4971_end_0, end_mask = var_4971_end_mask_0, squeeze_mask = var_4971_squeeze_mask_0, x = reshape_114_cast_fp16)[name = tensor("op_4971_cast_fp16")]; + tensor var_4973_axis_0 = const()[name = tensor("op_4973_axis_0"), val = tensor(1)]; + tensor var_4973_mode_0 = const()[name = tensor("op_4973_mode_0"), val = tensor("update")]; + tensor var_4973_validate_indices_0 = const()[name = tensor("op_4973_validate_indices_0"), val = tensor(false)]; + tensor var_4973_cast_fp16 = scatter_along_axis(axis = var_4973_axis_0, data = var_4971_cast_fp16, indices = write_indices_23, mode = var_4973_mode_0, updates = v_23_cast_fp16, validate_indices = var_4973_validate_indices_0)[name = tensor("op_4973_cast_fp16")]; + tensor concat_92 = const()[name = tensor("concat_92"), val = tensor([1, 0, 0, 0, 0])]; + tensor concat_93 = const()[name = tensor("concat_93"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_23_internal_tensor_assign_2_stride_0 = const()[name = tensor("new_cache_23_internal_tensor_assign_2_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_23_internal_tensor_assign_2_begin_mask_0 = const()[name = tensor("new_cache_23_internal_tensor_assign_2_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_23_internal_tensor_assign_2_end_mask_0 = const()[name = tensor("new_cache_23_internal_tensor_assign_2_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_23_internal_tensor_assign_2_squeeze_mask_0 = const()[name = tensor("new_cache_23_internal_tensor_assign_2_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_69 = const()[name = tensor("shape_69"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_23 = const()[name = tensor("reduce_prod_23"), val = tensor(1048576)]; + tensor range_1d_23_start_0 = const()[name = tensor("range_1d_23_start_0"), val = tensor(0)]; + tensor range_1d_23_step_0 = const()[name = tensor("range_1d_23_step_0"), val = tensor(1)]; + tensor range_1d_23 = range_1d(end = reduce_prod_23, start = range_1d_23_start_0, step = range_1d_23_step_0)[name = tensor("range_1d_23")]; + tensor reshape_115 = reshape(shape = shape_69, x = range_1d_23)[name = tensor("reshape_115")]; + tensor slice_by_index_23 = slice_by_index(begin = concat_92, begin_mask = new_cache_23_internal_tensor_assign_2_begin_mask_0, end = concat_93, end_mask = new_cache_23_internal_tensor_assign_2_end_mask_0, squeeze_mask = new_cache_23_internal_tensor_assign_2_squeeze_mask_0, stride = new_cache_23_internal_tensor_assign_2_stride_0, x = reshape_115)[name = tensor("slice_by_index_23")]; + tensor reshape_116_shape_0 = const()[name = tensor("reshape_116_shape_0"), val = tensor([-1])]; + tensor reshape_116 = reshape(shape = reshape_116_shape_0, x = slice_by_index_23)[name = tensor("reshape_116")]; + tensor reshape_117_shape_0 = const()[name = tensor("reshape_117_shape_0"), val = tensor([-1])]; + tensor reshape_117_cast_fp16 = reshape(shape = reshape_117_shape_0, x = var_4973_cast_fp16)[name = tensor("reshape_117_cast_fp16")]; + tensor reshape_118_shape_0 = const()[name = tensor("reshape_118_shape_0"), val = tensor([-1])]; + tensor reshape_118_cast_fp16 = reshape(shape = reshape_118_shape_0, x = reshape_114_cast_fp16)[name = tensor("reshape_118_cast_fp16")]; + tensor scatter_23_mode_0 = const()[name = tensor("scatter_23_mode_0"), val = tensor("update")]; + tensor scatter_23_axis_0 = const()[name = tensor("scatter_23_axis_0"), val = tensor(0)]; + tensor scatter_23_validate_indices_0 = const()[name = tensor("scatter_23_validate_indices_0"), val = tensor(false)]; + tensor scatter_23_cast_fp16 = scatter(axis = scatter_23_axis_0, data = reshape_118_cast_fp16, indices = reshape_116, mode = scatter_23_mode_0, updates = reshape_117_cast_fp16, validate_indices = scatter_23_validate_indices_0)[name = tensor("scatter_23_cast_fp16")]; + tensor reshape_119_cast_fp16 = reshape(shape = shape_69, x = scatter_23_cast_fp16)[name = tensor("reshape_119_cast_fp16")]; + tensor reshape_119_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("reshape_119_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor keys_67_begin_0 = const()[name = tensor("keys_67_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor keys_67_end_0 = const()[name = tensor("keys_67_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor keys_67_end_mask_0 = const()[name = tensor("keys_67_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor keys_67_squeeze_mask_0 = const()[name = tensor("keys_67_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor keys_67_cast_fp16 = slice_by_index(begin = keys_67_begin_0, end = keys_67_end_0, end_mask = keys_67_end_mask_0, squeeze_mask = keys_67_squeeze_mask_0, x = reshape_119_cast_fp16)[name = tensor("keys_67_cast_fp16")]; + tensor values_67_begin_0 = const()[name = tensor("values_67_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor values_67_end_0 = const()[name = tensor("values_67_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor values_67_end_mask_0 = const()[name = tensor("values_67_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor values_67_squeeze_mask_0 = const()[name = tensor("values_67_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor values_67_cast_fp16 = slice_by_index(begin = values_67_begin_0, end = values_67_end_0, end_mask = values_67_end_mask_0, squeeze_mask = values_67_squeeze_mask_0, x = reshape_119_cast_fp16)[name = tensor("values_67_cast_fp16")]; + tensor var_4985_cast_fp16 = not_equal(x = keys_67_cast_fp16, y = keys_67_cast_fp16)[name = tensor("op_4985_cast_fp16")]; + tensor keys_69_cast_fp16 = select(a = var_470_to_fp16, b = keys_67_cast_fp16, cond = var_4985_cast_fp16)[name = tensor("keys_69_cast_fp16")]; + tensor var_4993_cast_fp16 = not_equal(x = values_67_cast_fp16, y = values_67_cast_fp16)[name = tensor("op_4993_cast_fp16")]; + tensor values_69_cast_fp16 = select(a = var_470_to_fp16, b = values_67_cast_fp16, cond = var_4993_cast_fp16)[name = tensor("values_69_cast_fp16")]; + tensor var_5017 = const()[name = tensor("op_5017"), val = tensor([0, 2, 1, 3])]; + tensor var_5033 = const()[name = tensor("op_5033"), val = tensor([1, 1, 1])]; + tensor var_5034_cast_fp16 = reshape(shape = var_5033, x = position11_to_fp16)[name = tensor("op_5034_cast_fp16")]; + tensor q_positions_23_cast_fp16 = add(x = var_5034_cast_fp16, y = q_offsets_1_promoted_to_fp16)[name = tensor("q_positions_23_cast_fp16")]; + tensor valid_end_23_cast_fp16 = add(x = var_5034_cast_fp16, y = var_536_cast_fp16)[name = tensor("valid_end_23_cast_fp16")]; + tensor valid_mask_23_cast_fp16 = less(x = k_positions_1_promoted_to_fp16, y = valid_end_23_cast_fp16)[name = tensor("valid_mask_23_cast_fp16")]; + tensor causal_mask_23_cast_fp16 = less_equal(x = k_positions_1_promoted_to_fp16, y = q_positions_23_cast_fp16)[name = tensor("causal_mask_23_cast_fp16")]; + tensor var_5062 = logical_and(x = valid_mask_23_cast_fp16, y = causal_mask_23_cast_fp16)[name = tensor("op_5062")]; + tensor attn_mask_23_axes_0 = const()[name = tensor("attn_mask_23_axes_0"), val = tensor([1])]; + tensor attn_mask_23 = expand_dims(axes = attn_mask_23_axes_0, x = var_5062)[name = tensor("attn_mask_23")]; + tensor var_5076_transpose_x_0 = const()[name = tensor("op_5076_transpose_x_0"), val = tensor(false)]; + tensor var_5076_transpose_y_0 = const()[name = tensor("op_5076_transpose_y_0"), val = tensor(false)]; + tensor transpose_91_perm_0 = const()[name = tensor("transpose_91_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_92_perm_0 = const()[name = tensor("transpose_92_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_92 = transpose(perm = transpose_92_perm_0, x = keys_69_cast_fp16)[name = tensor("transpose_160")]; + tensor transpose_91 = transpose(perm = transpose_91_perm_0, x = q_69_cast_fp16)[name = tensor("transpose_161")]; + tensor var_5076_cast_fp16 = matmul(transpose_x = var_5076_transpose_x_0, transpose_y = var_5076_transpose_y_0, x = transpose_91, y = transpose_92)[name = tensor("op_5076_cast_fp16")]; + tensor var_5070_to_fp16 = const()[name = tensor("op_5070_to_fp16"), val = tensor([0x1p-3])]; + tensor attn_weights_67_cast_fp16 = mul(x = var_5076_cast_fp16, y = var_5070_to_fp16)[name = tensor("attn_weights_67_cast_fp16")]; + tensor var_5078 = logical_not(x = attn_mask_23)[name = tensor("op_5078")]; + tensor var_5079_to_fp16 = const()[name = tensor("op_5079_to_fp16"), val = tensor(-inf)]; + tensor attn_weights_69_cast_fp16 = select(a = var_5079_to_fp16, b = attn_weights_67_cast_fp16, cond = var_5078)[name = tensor("attn_weights_69_cast_fp16")]; + tensor var_5081 = const()[name = tensor("op_5081"), val = tensor(-1)]; + tensor attn_weights_71_cast_fp16 = softmax(axis = var_5081, x = attn_weights_69_cast_fp16)[name = tensor("attn_weights_71_cast_fp16")]; + tensor attn_output_23_transpose_x_0 = const()[name = tensor("attn_output_23_transpose_x_0"), val = tensor(false)]; + tensor attn_output_23_transpose_y_0 = const()[name = tensor("attn_output_23_transpose_y_0"), val = tensor(false)]; + tensor values_71_cast_fp16 = transpose(perm = var_5017, x = values_69_cast_fp16)[name = tensor("transpose_162")]; + tensor attn_output_23_cast_fp16 = matmul(transpose_x = attn_output_23_transpose_x_0, transpose_y = attn_output_23_transpose_y_0, x = attn_weights_71_cast_fp16, y = values_71_cast_fp16)[name = tensor("attn_output_23_cast_fp16")]; + tensor var_5089 = const()[name = tensor("op_5089"), val = tensor([0, 2, 1, 3])]; + tensor var_5094 = const()[name = tensor("op_5094"), val = tensor([1, 256, 1024])]; + tensor var_5090_cast_fp16 = transpose(perm = var_5089, x = attn_output_23_cast_fp16)[name = tensor("transpose_159")]; + tensor input_137_cast_fp16 = reshape(shape = var_5094, x = var_5090_cast_fp16)[name = tensor("input_137_cast_fp16")]; + tensor attn11_out_proj_weight_to_fp16 = const()[name = tensor("attn11_out_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(284285184)))]; + tensor linear_45_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = attn11_out_proj_weight_to_fp16, x = input_137_cast_fp16)[name = tensor("linear_45_cast_fp16")]; + tensor var_5103_cast_fp16 = add(x = position11_to_fp16, y = valid_len_to_fp16)[name = tensor("op_5103_cast_fp16")]; + tensor var_5103_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_5103_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor input_139_cast_fp16 = add(x = input_133_cast_fp16, y = linear_45_cast_fp16)[name = tensor("input_139_cast_fp16")]; + tensor input_141_axes_0 = const()[name = tensor("input_141_axes_0"), val = tensor([-1])]; + tensor norm11_2_weight_to_fp16 = const()[name = tensor("norm11_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(286382400)))]; + tensor norm11_2_bias_to_fp16 = const()[name = tensor("norm11_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(286384512)))]; + tensor var_5107_to_fp16 = const()[name = tensor("op_5107_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_141_cast_fp16 = layer_norm(axes = input_141_axes_0, beta = norm11_2_bias_to_fp16, epsilon = var_5107_to_fp16, gamma = norm11_2_weight_to_fp16, x = input_139_cast_fp16)[name = tensor("input_141_cast_fp16")]; + tensor linear11_1_weight_to_fp16 = const()[name = tensor("linear11_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(286386624)))]; + tensor linear_46_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = linear11_1_weight_to_fp16, x = input_141_cast_fp16)[name = tensor("linear_46_cast_fp16")]; + tensor input_143_mode_0 = const()[name = tensor("input_143_mode_0"), val = tensor("EXACT")]; + tensor input_143_cast_fp16 = gelu(mode = input_143_mode_0, x = linear_46_cast_fp16)[name = tensor("input_143_cast_fp16")]; + tensor linear11_2_weight_to_fp16 = const()[name = tensor("linear11_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(294775296)))]; + tensor linear_47_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = linear11_2_weight_to_fp16, x = input_143_cast_fp16)[name = tensor("linear_47_cast_fp16")]; + tensor input_145_cast_fp16 = add(x = input_139_cast_fp16, y = linear_47_cast_fp16)[name = tensor("input_145_cast_fp16")]; + tensor input_147_axes_0 = const()[name = tensor("input_147_axes_0"), val = tensor([-1])]; + tensor norm12_1_weight_to_fp16 = const()[name = tensor("norm12_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(303163968)))]; + tensor norm12_1_bias_to_fp16 = const()[name = tensor("norm12_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(303166080)))]; + tensor var_5124_to_fp16 = const()[name = tensor("op_5124_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_147_cast_fp16 = layer_norm(axes = input_147_axes_0, beta = norm12_1_bias_to_fp16, epsilon = var_5124_to_fp16, gamma = norm12_1_weight_to_fp16, x = input_145_cast_fp16)[name = tensor("input_147_cast_fp16")]; + tensor attn12_in_proj_weight_to_fp16 = const()[name = tensor("attn12_in_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(303168192)))]; + tensor linear_48_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = attn12_in_proj_weight_to_fp16, x = input_147_cast_fp16)[name = tensor("linear_48_cast_fp16")]; + tensor var_5138 = const()[name = tensor("op_5138"), val = tensor([1, 256, 3, 16, 64])]; + tensor qkv_25_cast_fp16 = reshape(shape = var_5138, x = linear_48_cast_fp16)[name = tensor("qkv_25_cast_fp16")]; + tensor q_73_begin_0 = const()[name = tensor("q_73_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor q_73_end_0 = const()[name = tensor("q_73_end_0"), val = tensor([1, 256, 1, 16, 64])]; + tensor q_73_end_mask_0 = const()[name = tensor("q_73_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor q_73_squeeze_mask_0 = const()[name = tensor("q_73_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor q_73_cast_fp16 = slice_by_index(begin = q_73_begin_0, end = q_73_end_0, end_mask = q_73_end_mask_0, squeeze_mask = q_73_squeeze_mask_0, x = qkv_25_cast_fp16)[name = tensor("q_73_cast_fp16")]; + tensor k_49_begin_0 = const()[name = tensor("k_49_begin_0"), val = tensor([0, 0, 1, 0, 0])]; + tensor k_49_end_0 = const()[name = tensor("k_49_end_0"), val = tensor([1, 256, 2, 16, 64])]; + tensor k_49_end_mask_0 = const()[name = tensor("k_49_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor k_49_squeeze_mask_0 = const()[name = tensor("k_49_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor k_49_cast_fp16 = slice_by_index(begin = k_49_begin_0, end = k_49_end_0, end_mask = k_49_end_mask_0, squeeze_mask = k_49_squeeze_mask_0, x = qkv_25_cast_fp16)[name = tensor("k_49_cast_fp16")]; + tensor v_25_begin_0 = const()[name = tensor("v_25_begin_0"), val = tensor([0, 0, 2, 0, 0])]; + tensor v_25_end_0 = const()[name = tensor("v_25_end_0"), val = tensor([1, 256, 3, 16, 64])]; + tensor v_25_end_mask_0 = const()[name = tensor("v_25_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor v_25_squeeze_mask_0 = const()[name = tensor("v_25_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor v_25_cast_fp16 = slice_by_index(begin = v_25_begin_0, end = v_25_end_0, end_mask = v_25_end_mask_0, squeeze_mask = v_25_squeeze_mask_0, x = qkv_25_cast_fp16)[name = tensor("v_25_cast_fp16")]; + tensor var_5196 = const()[name = tensor("op_5196"), val = tensor([1, 1])]; + tensor position12_to_fp16_dtype_0 = const()[name = tensor("position12_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor position12_to_fp16 = cast(dtype = position12_to_fp16_dtype_0, x = position12)[name = tensor("cast_601")]; + tensor var_5197_cast_fp16 = reshape(shape = var_5196, x = position12_to_fp16)[name = tensor("op_5197_cast_fp16")]; + tensor ts_75_cast_fp16 = add(x = ts_1_promoted_to_fp16, y = var_5197_cast_fp16)[name = tensor("ts_75_cast_fp16")]; + tensor var_5204 = const()[name = tensor("op_5204"), val = tensor([1, 256, 1, 1])]; + tensor ts_77_cast_fp16 = reshape(shape = var_5204, x = ts_75_cast_fp16)[name = tensor("ts_77_cast_fp16")]; + tensor var_5211 = const()[name = tensor("op_5211"), val = tensor([1, 256, 16, 32, 2])]; + tensor q_complex_25_cast_fp16 = reshape(shape = var_5211, x = q_73_cast_fp16)[name = tensor("q_complex_25_cast_fp16")]; + tensor var_5218 = const()[name = tensor("op_5218"), val = tensor([1, 256, 16, 32, 2])]; + tensor k_complex_25_cast_fp16 = reshape(shape = var_5218, x = k_49_cast_fp16)[name = tensor("k_complex_25_cast_fp16")]; + tensor var_5222_begin_0 = const()[name = tensor("op_5222_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_5222_end_0 = const()[name = tensor("op_5222_end_0"), val = tensor([1, 256, 16, 32, 1])]; + tensor var_5222_end_mask_0 = const()[name = tensor("op_5222_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_5222_squeeze_mask_0 = const()[name = tensor("op_5222_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_5222_cast_fp16 = slice_by_index(begin = var_5222_begin_0, end = var_5222_end_0, end_mask = var_5222_end_mask_0, squeeze_mask = var_5222_squeeze_mask_0, x = q_complex_25_cast_fp16)[name = tensor("op_5222_cast_fp16")]; + tensor var_5230_begin_0 = const()[name = tensor("op_5230_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_5230_end_0 = const()[name = tensor("op_5230_end_0"), val = tensor([1, 256, 16, 32, 2])]; + tensor var_5230_end_mask_0 = const()[name = tensor("op_5230_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_5230_squeeze_mask_0 = const()[name = tensor("op_5230_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_5230_cast_fp16 = slice_by_index(begin = var_5230_begin_0, end = var_5230_end_0, end_mask = var_5230_end_mask_0, squeeze_mask = var_5230_squeeze_mask_0, x = q_complex_25_cast_fp16)[name = tensor("op_5230_cast_fp16")]; + tensor var_5238_begin_0 = const()[name = tensor("op_5238_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_5238_end_0 = const()[name = tensor("op_5238_end_0"), val = tensor([1, 256, 16, 32, 1])]; + tensor var_5238_end_mask_0 = const()[name = tensor("op_5238_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_5238_squeeze_mask_0 = const()[name = tensor("op_5238_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_5238_cast_fp16 = slice_by_index(begin = var_5238_begin_0, end = var_5238_end_0, end_mask = var_5238_end_mask_0, squeeze_mask = var_5238_squeeze_mask_0, x = k_complex_25_cast_fp16)[name = tensor("op_5238_cast_fp16")]; + tensor var_5246_begin_0 = const()[name = tensor("op_5246_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_5246_end_0 = const()[name = tensor("op_5246_end_0"), val = tensor([1, 256, 16, 32, 2])]; + tensor var_5246_end_mask_0 = const()[name = tensor("op_5246_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_5246_squeeze_mask_0 = const()[name = tensor("op_5246_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_5246_cast_fp16 = slice_by_index(begin = var_5246_begin_0, end = var_5246_end_0, end_mask = var_5246_end_mask_0, squeeze_mask = var_5246_squeeze_mask_0, x = k_complex_25_cast_fp16)[name = tensor("op_5246_cast_fp16")]; + tensor freqs_25_to_fp16 = const()[name = tensor("freqs_25_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(309459712)))]; + tensor var_5252_cast_fp16 = mul(x = freqs_25_to_fp16, y = ts_77_cast_fp16)[name = tensor("op_5252_cast_fp16")]; + tensor rotr_25_cast_fp16 = cos(x = var_5252_cast_fp16)[name = tensor("rotr_25_cast_fp16")]; + tensor roti_25_cast_fp16 = sin(x = var_5252_cast_fp16)[name = tensor("roti_25_cast_fp16")]; + tensor var_5256_cast_fp16 = mul(x = var_5222_cast_fp16, y = rotr_25_cast_fp16)[name = tensor("op_5256_cast_fp16")]; + tensor var_5257_cast_fp16 = mul(x = var_5230_cast_fp16, y = roti_25_cast_fp16)[name = tensor("op_5257_cast_fp16")]; + tensor qor_49_cast_fp16 = sub(x = var_5256_cast_fp16, y = var_5257_cast_fp16)[name = tensor("qor_49_cast_fp16")]; + tensor var_5260_cast_fp16 = mul(x = var_5222_cast_fp16, y = roti_25_cast_fp16)[name = tensor("op_5260_cast_fp16")]; + tensor var_5261_cast_fp16 = mul(x = var_5230_cast_fp16, y = rotr_25_cast_fp16)[name = tensor("op_5261_cast_fp16")]; + tensor qoi_49_cast_fp16 = add(x = var_5260_cast_fp16, y = var_5261_cast_fp16)[name = tensor("qoi_49_cast_fp16")]; + tensor var_5264_cast_fp16 = mul(x = var_5238_cast_fp16, y = rotr_25_cast_fp16)[name = tensor("op_5264_cast_fp16")]; + tensor var_5265_cast_fp16 = mul(x = var_5246_cast_fp16, y = roti_25_cast_fp16)[name = tensor("op_5265_cast_fp16")]; + tensor kor_49_cast_fp16 = sub(x = var_5264_cast_fp16, y = var_5265_cast_fp16)[name = tensor("kor_49_cast_fp16")]; + tensor var_5268_cast_fp16 = mul(x = var_5238_cast_fp16, y = roti_25_cast_fp16)[name = tensor("op_5268_cast_fp16")]; + tensor var_5269_cast_fp16 = mul(x = var_5246_cast_fp16, y = rotr_25_cast_fp16)[name = tensor("op_5269_cast_fp16")]; + tensor koi_49_cast_fp16 = add(x = var_5268_cast_fp16, y = var_5269_cast_fp16)[name = tensor("koi_49_cast_fp16")]; + tensor qo_25_axis_0 = const()[name = tensor("qo_25_axis_0"), val = tensor(-1)]; + tensor qo_25_cast_fp16 = stack(axis = qo_25_axis_0, values = (qor_49_cast_fp16, qoi_49_cast_fp16))[name = tensor("qo_25_cast_fp16")]; + tensor ko_25_axis_0 = const()[name = tensor("ko_25_axis_0"), val = tensor(-1)]; + tensor ko_25_cast_fp16 = stack(axis = ko_25_axis_0, values = (kor_49_cast_fp16, koi_49_cast_fp16))[name = tensor("ko_25_cast_fp16")]; + tensor var_5302 = const()[name = tensor("op_5302"), val = tensor([1, 256, 16, 64])]; + tensor q_75_cast_fp16 = reshape(shape = var_5302, x = qo_25_cast_fp16)[name = tensor("q_75_cast_fp16")]; + tensor var_5308 = const()[name = tensor("op_5308"), val = tensor([1, 256, 16, 64])]; + tensor k_51_cast_fp16 = reshape(shape = var_5308, x = ko_25_cast_fp16)[name = tensor("k_51_cast_fp16")]; + tensor write_indices_float_73_cast_fp16 = add(x = var_5197_cast_fp16, y = var_393_promoted_to_fp16)[name = tensor("write_indices_float_73_cast_fp16")]; + tensor _inversed_5329_y_0_to_fp16 = const()[name = tensor("_inversed_5329_y_0_to_fp16"), val = tensor(0x1p-9)]; + tensor _inversed_5329_cast_fp16 = mul(x = write_indices_float_73_cast_fp16, y = _inversed_5329_y_0_to_fp16)[name = tensor("_inversed_5329_cast_fp16")]; + tensor var_5330_cast_fp16 = floor(x = _inversed_5329_cast_fp16)[name = tensor("op_5330_cast_fp16")]; + tensor var_5331_to_fp16 = const()[name = tensor("op_5331_to_fp16"), val = tensor(0x1p+9)]; + tensor var_5332_cast_fp16 = mul(x = var_5330_cast_fp16, y = var_5331_to_fp16)[name = tensor("op_5332_cast_fp16")]; + tensor write_indices_float_75_cast_fp16 = sub(x = write_indices_float_73_cast_fp16, y = var_5332_cast_fp16)[name = tensor("write_indices_float_75_cast_fp16")]; + tensor var_5350_value_0_to_fp16 = const()[name = tensor("op_5350_value_0_to_fp16"), val = tensor(0x1.ffp+8)]; + tensor var_5350_cast_fp16 = fill_like(ref_tensor = write_indices_float_75_cast_fp16, value = var_5350_value_0_to_fp16)[name = tensor("op_5350_cast_fp16")]; + tensor write_indices_float_77_cast_fp16 = select(a = write_indices_float_75_cast_fp16, b = var_5350_cast_fp16, cond = token_valid_1_cast_fp16)[name = tensor("write_indices_float_77_cast_fp16")]; + tensor var_5356_dtype_0 = const()[name = tensor("op_5356_dtype_0"), val = tensor("int32")]; + tensor var_5361 = const()[name = tensor("op_5361"), val = tensor([1, 256, 1, 1])]; + tensor write_indices_float_77_cast_fp16_to_int32 = cast(dtype = var_5356_dtype_0, x = write_indices_float_77_cast_fp16)[name = tensor("cast_600")]; + tensor var_5362 = reshape(shape = var_5361, x = write_indices_float_77_cast_fp16_to_int32)[name = tensor("op_5362")]; + tensor write_indices_25_reps_0 = const()[name = tensor("write_indices_25_reps_0"), val = tensor([1, 1, 16, 64])]; + tensor write_indices_25 = tile(reps = write_indices_25_reps_0, x = var_5362)[name = tensor("write_indices_25")]; + tensor var_5372_begin_0 = const()[name = tensor("op_5372_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_5372_end_0 = const()[name = tensor("op_5372_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor var_5372_end_mask_0 = const()[name = tensor("op_5372_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_5372_squeeze_mask_0 = const()[name = tensor("op_5372_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor cache12_to_fp16_dtype_0 = const()[name = tensor("cache12_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor cache12_to_fp16 = cast(dtype = cache12_to_fp16_dtype_0, x = cache12)[name = tensor("cast_599")]; + tensor var_5372_cast_fp16 = slice_by_index(begin = var_5372_begin_0, end = var_5372_end_0, end_mask = var_5372_end_mask_0, squeeze_mask = var_5372_squeeze_mask_0, x = cache12_to_fp16)[name = tensor("op_5372_cast_fp16")]; + tensor var_5374_axis_0 = const()[name = tensor("op_5374_axis_0"), val = tensor(1)]; + tensor var_5374_mode_0 = const()[name = tensor("op_5374_mode_0"), val = tensor("update")]; + tensor var_5374_validate_indices_0 = const()[name = tensor("op_5374_validate_indices_0"), val = tensor(false)]; + tensor var_5374_cast_fp16 = scatter_along_axis(axis = var_5374_axis_0, data = var_5372_cast_fp16, indices = write_indices_25, mode = var_5374_mode_0, updates = k_51_cast_fp16, validate_indices = var_5374_validate_indices_0)[name = tensor("op_5374_cast_fp16")]; + tensor concat_98 = const()[name = tensor("concat_98"), val = tensor([0, 0, 0, 0, 0])]; + tensor concat_99 = const()[name = tensor("concat_99"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_25_internal_tensor_assign_1_stride_0 = const()[name = tensor("new_cache_25_internal_tensor_assign_1_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_25_internal_tensor_assign_1_begin_mask_0 = const()[name = tensor("new_cache_25_internal_tensor_assign_1_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_25_internal_tensor_assign_1_end_mask_0 = const()[name = tensor("new_cache_25_internal_tensor_assign_1_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_25_internal_tensor_assign_1_squeeze_mask_0 = const()[name = tensor("new_cache_25_internal_tensor_assign_1_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_70 = const()[name = tensor("shape_70"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_24 = const()[name = tensor("reduce_prod_24"), val = tensor(1048576)]; + tensor range_1d_24_start_0 = const()[name = tensor("range_1d_24_start_0"), val = tensor(0)]; + tensor range_1d_24_step_0 = const()[name = tensor("range_1d_24_step_0"), val = tensor(1)]; + tensor range_1d_24 = range_1d(end = reduce_prod_24, start = range_1d_24_start_0, step = range_1d_24_step_0)[name = tensor("range_1d_24")]; + tensor reshape_120 = reshape(shape = shape_70, x = range_1d_24)[name = tensor("reshape_120")]; + tensor slice_by_index_24 = slice_by_index(begin = concat_98, begin_mask = new_cache_25_internal_tensor_assign_1_begin_mask_0, end = concat_99, end_mask = new_cache_25_internal_tensor_assign_1_end_mask_0, squeeze_mask = new_cache_25_internal_tensor_assign_1_squeeze_mask_0, stride = new_cache_25_internal_tensor_assign_1_stride_0, x = reshape_120)[name = tensor("slice_by_index_24")]; + tensor reshape_121_shape_0 = const()[name = tensor("reshape_121_shape_0"), val = tensor([-1])]; + tensor reshape_121 = reshape(shape = reshape_121_shape_0, x = slice_by_index_24)[name = tensor("reshape_121")]; + tensor reshape_122_shape_0 = const()[name = tensor("reshape_122_shape_0"), val = tensor([-1])]; + tensor reshape_122_cast_fp16 = reshape(shape = reshape_122_shape_0, x = var_5374_cast_fp16)[name = tensor("reshape_122_cast_fp16")]; + tensor reshape_123_shape_0 = const()[name = tensor("reshape_123_shape_0"), val = tensor([-1])]; + tensor reshape_123_cast_fp16 = reshape(shape = reshape_123_shape_0, x = cache12_to_fp16)[name = tensor("reshape_123_cast_fp16")]; + tensor scatter_24_mode_0 = const()[name = tensor("scatter_24_mode_0"), val = tensor("update")]; + tensor scatter_24_axis_0 = const()[name = tensor("scatter_24_axis_0"), val = tensor(0)]; + tensor scatter_24_validate_indices_0 = const()[name = tensor("scatter_24_validate_indices_0"), val = tensor(false)]; + tensor scatter_24_cast_fp16 = scatter(axis = scatter_24_axis_0, data = reshape_123_cast_fp16, indices = reshape_121, mode = scatter_24_mode_0, updates = reshape_122_cast_fp16, validate_indices = scatter_24_validate_indices_0)[name = tensor("scatter_24_cast_fp16")]; + tensor reshape_124_cast_fp16 = reshape(shape = shape_70, x = scatter_24_cast_fp16)[name = tensor("reshape_124_cast_fp16")]; + tensor var_5382_begin_0 = const()[name = tensor("op_5382_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor var_5382_end_0 = const()[name = tensor("op_5382_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor var_5382_end_mask_0 = const()[name = tensor("op_5382_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_5382_squeeze_mask_0 = const()[name = tensor("op_5382_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor var_5382_cast_fp16 = slice_by_index(begin = var_5382_begin_0, end = var_5382_end_0, end_mask = var_5382_end_mask_0, squeeze_mask = var_5382_squeeze_mask_0, x = reshape_124_cast_fp16)[name = tensor("op_5382_cast_fp16")]; + tensor var_5384_axis_0 = const()[name = tensor("op_5384_axis_0"), val = tensor(1)]; + tensor var_5384_mode_0 = const()[name = tensor("op_5384_mode_0"), val = tensor("update")]; + tensor var_5384_validate_indices_0 = const()[name = tensor("op_5384_validate_indices_0"), val = tensor(false)]; + tensor var_5384_cast_fp16 = scatter_along_axis(axis = var_5384_axis_0, data = var_5382_cast_fp16, indices = write_indices_25, mode = var_5384_mode_0, updates = v_25_cast_fp16, validate_indices = var_5384_validate_indices_0)[name = tensor("op_5384_cast_fp16")]; + tensor concat_100 = const()[name = tensor("concat_100"), val = tensor([1, 0, 0, 0, 0])]; + tensor concat_101 = const()[name = tensor("concat_101"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_25_internal_tensor_assign_2_stride_0 = const()[name = tensor("new_cache_25_internal_tensor_assign_2_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_25_internal_tensor_assign_2_begin_mask_0 = const()[name = tensor("new_cache_25_internal_tensor_assign_2_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_25_internal_tensor_assign_2_end_mask_0 = const()[name = tensor("new_cache_25_internal_tensor_assign_2_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_25_internal_tensor_assign_2_squeeze_mask_0 = const()[name = tensor("new_cache_25_internal_tensor_assign_2_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_71 = const()[name = tensor("shape_71"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_25 = const()[name = tensor("reduce_prod_25"), val = tensor(1048576)]; + tensor range_1d_25_start_0 = const()[name = tensor("range_1d_25_start_0"), val = tensor(0)]; + tensor range_1d_25_step_0 = const()[name = tensor("range_1d_25_step_0"), val = tensor(1)]; + tensor range_1d_25 = range_1d(end = reduce_prod_25, start = range_1d_25_start_0, step = range_1d_25_step_0)[name = tensor("range_1d_25")]; + tensor reshape_125 = reshape(shape = shape_71, x = range_1d_25)[name = tensor("reshape_125")]; + tensor slice_by_index_25 = slice_by_index(begin = concat_100, begin_mask = new_cache_25_internal_tensor_assign_2_begin_mask_0, end = concat_101, end_mask = new_cache_25_internal_tensor_assign_2_end_mask_0, squeeze_mask = new_cache_25_internal_tensor_assign_2_squeeze_mask_0, stride = new_cache_25_internal_tensor_assign_2_stride_0, x = reshape_125)[name = tensor("slice_by_index_25")]; + tensor reshape_126_shape_0 = const()[name = tensor("reshape_126_shape_0"), val = tensor([-1])]; + tensor reshape_126 = reshape(shape = reshape_126_shape_0, x = slice_by_index_25)[name = tensor("reshape_126")]; + tensor reshape_127_shape_0 = const()[name = tensor("reshape_127_shape_0"), val = tensor([-1])]; + tensor reshape_127_cast_fp16 = reshape(shape = reshape_127_shape_0, x = var_5384_cast_fp16)[name = tensor("reshape_127_cast_fp16")]; + tensor reshape_128_shape_0 = const()[name = tensor("reshape_128_shape_0"), val = tensor([-1])]; + tensor reshape_128_cast_fp16 = reshape(shape = reshape_128_shape_0, x = reshape_124_cast_fp16)[name = tensor("reshape_128_cast_fp16")]; + tensor scatter_25_mode_0 = const()[name = tensor("scatter_25_mode_0"), val = tensor("update")]; + tensor scatter_25_axis_0 = const()[name = tensor("scatter_25_axis_0"), val = tensor(0)]; + tensor scatter_25_validate_indices_0 = const()[name = tensor("scatter_25_validate_indices_0"), val = tensor(false)]; + tensor scatter_25_cast_fp16 = scatter(axis = scatter_25_axis_0, data = reshape_128_cast_fp16, indices = reshape_126, mode = scatter_25_mode_0, updates = reshape_127_cast_fp16, validate_indices = scatter_25_validate_indices_0)[name = tensor("scatter_25_cast_fp16")]; + tensor reshape_129_cast_fp16 = reshape(shape = shape_71, x = scatter_25_cast_fp16)[name = tensor("reshape_129_cast_fp16")]; + tensor reshape_129_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("reshape_129_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor keys_73_begin_0 = const()[name = tensor("keys_73_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor keys_73_end_0 = const()[name = tensor("keys_73_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor keys_73_end_mask_0 = const()[name = tensor("keys_73_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor keys_73_squeeze_mask_0 = const()[name = tensor("keys_73_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor keys_73_cast_fp16 = slice_by_index(begin = keys_73_begin_0, end = keys_73_end_0, end_mask = keys_73_end_mask_0, squeeze_mask = keys_73_squeeze_mask_0, x = reshape_129_cast_fp16)[name = tensor("keys_73_cast_fp16")]; + tensor values_73_begin_0 = const()[name = tensor("values_73_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor values_73_end_0 = const()[name = tensor("values_73_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor values_73_end_mask_0 = const()[name = tensor("values_73_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor values_73_squeeze_mask_0 = const()[name = tensor("values_73_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor values_73_cast_fp16 = slice_by_index(begin = values_73_begin_0, end = values_73_end_0, end_mask = values_73_end_mask_0, squeeze_mask = values_73_squeeze_mask_0, x = reshape_129_cast_fp16)[name = tensor("values_73_cast_fp16")]; + tensor var_5396_cast_fp16 = not_equal(x = keys_73_cast_fp16, y = keys_73_cast_fp16)[name = tensor("op_5396_cast_fp16")]; + tensor keys_75_cast_fp16 = select(a = var_470_to_fp16, b = keys_73_cast_fp16, cond = var_5396_cast_fp16)[name = tensor("keys_75_cast_fp16")]; + tensor var_5404_cast_fp16 = not_equal(x = values_73_cast_fp16, y = values_73_cast_fp16)[name = tensor("op_5404_cast_fp16")]; + tensor values_75_cast_fp16 = select(a = var_470_to_fp16, b = values_73_cast_fp16, cond = var_5404_cast_fp16)[name = tensor("values_75_cast_fp16")]; + tensor var_5428 = const()[name = tensor("op_5428"), val = tensor([0, 2, 1, 3])]; + tensor var_5444 = const()[name = tensor("op_5444"), val = tensor([1, 1, 1])]; + tensor var_5445_cast_fp16 = reshape(shape = var_5444, x = position12_to_fp16)[name = tensor("op_5445_cast_fp16")]; + tensor q_positions_25_cast_fp16 = add(x = var_5445_cast_fp16, y = q_offsets_1_promoted_to_fp16)[name = tensor("q_positions_25_cast_fp16")]; + tensor valid_end_25_cast_fp16 = add(x = var_5445_cast_fp16, y = var_536_cast_fp16)[name = tensor("valid_end_25_cast_fp16")]; + tensor valid_mask_25_cast_fp16 = less(x = k_positions_1_promoted_to_fp16, y = valid_end_25_cast_fp16)[name = tensor("valid_mask_25_cast_fp16")]; + tensor causal_mask_25_cast_fp16 = less_equal(x = k_positions_1_promoted_to_fp16, y = q_positions_25_cast_fp16)[name = tensor("causal_mask_25_cast_fp16")]; + tensor var_5473 = logical_and(x = valid_mask_25_cast_fp16, y = causal_mask_25_cast_fp16)[name = tensor("op_5473")]; + tensor attn_mask_25_axes_0 = const()[name = tensor("attn_mask_25_axes_0"), val = tensor([1])]; + tensor attn_mask_25 = expand_dims(axes = attn_mask_25_axes_0, x = var_5473)[name = tensor("attn_mask_25")]; + tensor var_5487_transpose_x_0 = const()[name = tensor("op_5487_transpose_x_0"), val = tensor(false)]; + tensor var_5487_transpose_y_0 = const()[name = tensor("op_5487_transpose_y_0"), val = tensor(false)]; + tensor transpose_93_perm_0 = const()[name = tensor("transpose_93_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_94_perm_0 = const()[name = tensor("transpose_94_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_94 = transpose(perm = transpose_94_perm_0, x = keys_75_cast_fp16)[name = tensor("transpose_156")]; + tensor transpose_93 = transpose(perm = transpose_93_perm_0, x = q_75_cast_fp16)[name = tensor("transpose_157")]; + tensor var_5487_cast_fp16 = matmul(transpose_x = var_5487_transpose_x_0, transpose_y = var_5487_transpose_y_0, x = transpose_93, y = transpose_94)[name = tensor("op_5487_cast_fp16")]; + tensor var_5481_to_fp16 = const()[name = tensor("op_5481_to_fp16"), val = tensor([0x1p-3])]; + tensor attn_weights_73_cast_fp16 = mul(x = var_5487_cast_fp16, y = var_5481_to_fp16)[name = tensor("attn_weights_73_cast_fp16")]; + tensor var_5489 = logical_not(x = attn_mask_25)[name = tensor("op_5489")]; + tensor var_5490_to_fp16 = const()[name = tensor("op_5490_to_fp16"), val = tensor(-inf)]; + tensor attn_weights_75_cast_fp16 = select(a = var_5490_to_fp16, b = attn_weights_73_cast_fp16, cond = var_5489)[name = tensor("attn_weights_75_cast_fp16")]; + tensor var_5492 = const()[name = tensor("op_5492"), val = tensor(-1)]; + tensor attn_weights_77_cast_fp16 = softmax(axis = var_5492, x = attn_weights_75_cast_fp16)[name = tensor("attn_weights_77_cast_fp16")]; + tensor attn_output_25_transpose_x_0 = const()[name = tensor("attn_output_25_transpose_x_0"), val = tensor(false)]; + tensor attn_output_25_transpose_y_0 = const()[name = tensor("attn_output_25_transpose_y_0"), val = tensor(false)]; + tensor values_77_cast_fp16 = transpose(perm = var_5428, x = values_75_cast_fp16)[name = tensor("transpose_158")]; + tensor attn_output_25_cast_fp16 = matmul(transpose_x = attn_output_25_transpose_x_0, transpose_y = attn_output_25_transpose_y_0, x = attn_weights_77_cast_fp16, y = values_77_cast_fp16)[name = tensor("attn_output_25_cast_fp16")]; + tensor var_5500 = const()[name = tensor("op_5500"), val = tensor([0, 2, 1, 3])]; + tensor var_5505 = const()[name = tensor("op_5505"), val = tensor([1, 256, 1024])]; + tensor var_5501_cast_fp16 = transpose(perm = var_5500, x = attn_output_25_cast_fp16)[name = tensor("transpose_155")]; + tensor input_149_cast_fp16 = reshape(shape = var_5505, x = var_5501_cast_fp16)[name = tensor("input_149_cast_fp16")]; + tensor attn12_out_proj_weight_to_fp16 = const()[name = tensor("attn12_out_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(309459840)))]; + tensor linear_49_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = attn12_out_proj_weight_to_fp16, x = input_149_cast_fp16)[name = tensor("linear_49_cast_fp16")]; + tensor var_5514_cast_fp16 = add(x = position12_to_fp16, y = valid_len_to_fp16)[name = tensor("op_5514_cast_fp16")]; + tensor var_5514_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_5514_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor input_151_cast_fp16 = add(x = input_145_cast_fp16, y = linear_49_cast_fp16)[name = tensor("input_151_cast_fp16")]; + tensor input_153_axes_0 = const()[name = tensor("input_153_axes_0"), val = tensor([-1])]; + tensor norm12_2_weight_to_fp16 = const()[name = tensor("norm12_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(311557056)))]; + tensor norm12_2_bias_to_fp16 = const()[name = tensor("norm12_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(311559168)))]; + tensor var_5518_to_fp16 = const()[name = tensor("op_5518_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_153_cast_fp16 = layer_norm(axes = input_153_axes_0, beta = norm12_2_bias_to_fp16, epsilon = var_5518_to_fp16, gamma = norm12_2_weight_to_fp16, x = input_151_cast_fp16)[name = tensor("input_153_cast_fp16")]; + tensor linear12_1_weight_to_fp16 = const()[name = tensor("linear12_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(311561280)))]; + tensor linear_50_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = linear12_1_weight_to_fp16, x = input_153_cast_fp16)[name = tensor("linear_50_cast_fp16")]; + tensor input_155_mode_0 = const()[name = tensor("input_155_mode_0"), val = tensor("EXACT")]; + tensor input_155_cast_fp16 = gelu(mode = input_155_mode_0, x = linear_50_cast_fp16)[name = tensor("input_155_cast_fp16")]; + tensor linear12_2_weight_to_fp16 = const()[name = tensor("linear12_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(319949952)))]; + tensor linear_51_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = linear12_2_weight_to_fp16, x = input_155_cast_fp16)[name = tensor("linear_51_cast_fp16")]; + tensor input_157_cast_fp16 = add(x = input_151_cast_fp16, y = linear_51_cast_fp16)[name = tensor("input_157_cast_fp16")]; + tensor input_159_axes_0 = const()[name = tensor("input_159_axes_0"), val = tensor([-1])]; + tensor norm13_1_weight_to_fp16 = const()[name = tensor("norm13_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(328338624)))]; + tensor norm13_1_bias_to_fp16 = const()[name = tensor("norm13_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(328340736)))]; + tensor var_5535_to_fp16 = const()[name = tensor("op_5535_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_159_cast_fp16 = layer_norm(axes = input_159_axes_0, beta = norm13_1_bias_to_fp16, epsilon = var_5535_to_fp16, gamma = norm13_1_weight_to_fp16, x = input_157_cast_fp16)[name = tensor("input_159_cast_fp16")]; + tensor attn13_in_proj_weight_to_fp16 = const()[name = tensor("attn13_in_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(328342848)))]; + tensor linear_52_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = attn13_in_proj_weight_to_fp16, x = input_159_cast_fp16)[name = tensor("linear_52_cast_fp16")]; + tensor var_5549 = const()[name = tensor("op_5549"), val = tensor([1, 256, 3, 16, 64])]; + tensor qkv_27_cast_fp16 = reshape(shape = var_5549, x = linear_52_cast_fp16)[name = tensor("qkv_27_cast_fp16")]; + tensor q_79_begin_0 = const()[name = tensor("q_79_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor q_79_end_0 = const()[name = tensor("q_79_end_0"), val = tensor([1, 256, 1, 16, 64])]; + tensor q_79_end_mask_0 = const()[name = tensor("q_79_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor q_79_squeeze_mask_0 = const()[name = tensor("q_79_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor q_79_cast_fp16 = slice_by_index(begin = q_79_begin_0, end = q_79_end_0, end_mask = q_79_end_mask_0, squeeze_mask = q_79_squeeze_mask_0, x = qkv_27_cast_fp16)[name = tensor("q_79_cast_fp16")]; + tensor k_53_begin_0 = const()[name = tensor("k_53_begin_0"), val = tensor([0, 0, 1, 0, 0])]; + tensor k_53_end_0 = const()[name = tensor("k_53_end_0"), val = tensor([1, 256, 2, 16, 64])]; + tensor k_53_end_mask_0 = const()[name = tensor("k_53_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor k_53_squeeze_mask_0 = const()[name = tensor("k_53_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor k_53_cast_fp16 = slice_by_index(begin = k_53_begin_0, end = k_53_end_0, end_mask = k_53_end_mask_0, squeeze_mask = k_53_squeeze_mask_0, x = qkv_27_cast_fp16)[name = tensor("k_53_cast_fp16")]; + tensor v_27_begin_0 = const()[name = tensor("v_27_begin_0"), val = tensor([0, 0, 2, 0, 0])]; + tensor v_27_end_0 = const()[name = tensor("v_27_end_0"), val = tensor([1, 256, 3, 16, 64])]; + tensor v_27_end_mask_0 = const()[name = tensor("v_27_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor v_27_squeeze_mask_0 = const()[name = tensor("v_27_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor v_27_cast_fp16 = slice_by_index(begin = v_27_begin_0, end = v_27_end_0, end_mask = v_27_end_mask_0, squeeze_mask = v_27_squeeze_mask_0, x = qkv_27_cast_fp16)[name = tensor("v_27_cast_fp16")]; + tensor var_5607 = const()[name = tensor("op_5607"), val = tensor([1, 1])]; + tensor position13_to_fp16_dtype_0 = const()[name = tensor("position13_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor position13_to_fp16 = cast(dtype = position13_to_fp16_dtype_0, x = position13)[name = tensor("cast_596")]; + tensor var_5608_cast_fp16 = reshape(shape = var_5607, x = position13_to_fp16)[name = tensor("op_5608_cast_fp16")]; + tensor ts_81_cast_fp16 = add(x = ts_1_promoted_to_fp16, y = var_5608_cast_fp16)[name = tensor("ts_81_cast_fp16")]; + tensor var_5615 = const()[name = tensor("op_5615"), val = tensor([1, 256, 1, 1])]; + tensor ts_83_cast_fp16 = reshape(shape = var_5615, x = ts_81_cast_fp16)[name = tensor("ts_83_cast_fp16")]; + tensor var_5622 = const()[name = tensor("op_5622"), val = tensor([1, 256, 16, 32, 2])]; + tensor q_complex_27_cast_fp16 = reshape(shape = var_5622, x = q_79_cast_fp16)[name = tensor("q_complex_27_cast_fp16")]; + tensor var_5629 = const()[name = tensor("op_5629"), val = tensor([1, 256, 16, 32, 2])]; + tensor k_complex_27_cast_fp16 = reshape(shape = var_5629, x = k_53_cast_fp16)[name = tensor("k_complex_27_cast_fp16")]; + tensor var_5633_begin_0 = const()[name = tensor("op_5633_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_5633_end_0 = const()[name = tensor("op_5633_end_0"), val = tensor([1, 256, 16, 32, 1])]; + tensor var_5633_end_mask_0 = const()[name = tensor("op_5633_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_5633_squeeze_mask_0 = const()[name = tensor("op_5633_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_5633_cast_fp16 = slice_by_index(begin = var_5633_begin_0, end = var_5633_end_0, end_mask = var_5633_end_mask_0, squeeze_mask = var_5633_squeeze_mask_0, x = q_complex_27_cast_fp16)[name = tensor("op_5633_cast_fp16")]; + tensor var_5641_begin_0 = const()[name = tensor("op_5641_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_5641_end_0 = const()[name = tensor("op_5641_end_0"), val = tensor([1, 256, 16, 32, 2])]; + tensor var_5641_end_mask_0 = const()[name = tensor("op_5641_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_5641_squeeze_mask_0 = const()[name = tensor("op_5641_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_5641_cast_fp16 = slice_by_index(begin = var_5641_begin_0, end = var_5641_end_0, end_mask = var_5641_end_mask_0, squeeze_mask = var_5641_squeeze_mask_0, x = q_complex_27_cast_fp16)[name = tensor("op_5641_cast_fp16")]; + tensor var_5649_begin_0 = const()[name = tensor("op_5649_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_5649_end_0 = const()[name = tensor("op_5649_end_0"), val = tensor([1, 256, 16, 32, 1])]; + tensor var_5649_end_mask_0 = const()[name = tensor("op_5649_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_5649_squeeze_mask_0 = const()[name = tensor("op_5649_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_5649_cast_fp16 = slice_by_index(begin = var_5649_begin_0, end = var_5649_end_0, end_mask = var_5649_end_mask_0, squeeze_mask = var_5649_squeeze_mask_0, x = k_complex_27_cast_fp16)[name = tensor("op_5649_cast_fp16")]; + tensor var_5657_begin_0 = const()[name = tensor("op_5657_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_5657_end_0 = const()[name = tensor("op_5657_end_0"), val = tensor([1, 256, 16, 32, 2])]; + tensor var_5657_end_mask_0 = const()[name = tensor("op_5657_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_5657_squeeze_mask_0 = const()[name = tensor("op_5657_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_5657_cast_fp16 = slice_by_index(begin = var_5657_begin_0, end = var_5657_end_0, end_mask = var_5657_end_mask_0, squeeze_mask = var_5657_squeeze_mask_0, x = k_complex_27_cast_fp16)[name = tensor("op_5657_cast_fp16")]; + tensor freqs_27_to_fp16 = const()[name = tensor("freqs_27_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(334634368)))]; + tensor var_5663_cast_fp16 = mul(x = freqs_27_to_fp16, y = ts_83_cast_fp16)[name = tensor("op_5663_cast_fp16")]; + tensor rotr_27_cast_fp16 = cos(x = var_5663_cast_fp16)[name = tensor("rotr_27_cast_fp16")]; + tensor roti_27_cast_fp16 = sin(x = var_5663_cast_fp16)[name = tensor("roti_27_cast_fp16")]; + tensor var_5667_cast_fp16 = mul(x = var_5633_cast_fp16, y = rotr_27_cast_fp16)[name = tensor("op_5667_cast_fp16")]; + tensor var_5668_cast_fp16 = mul(x = var_5641_cast_fp16, y = roti_27_cast_fp16)[name = tensor("op_5668_cast_fp16")]; + tensor qor_53_cast_fp16 = sub(x = var_5667_cast_fp16, y = var_5668_cast_fp16)[name = tensor("qor_53_cast_fp16")]; + tensor var_5671_cast_fp16 = mul(x = var_5633_cast_fp16, y = roti_27_cast_fp16)[name = tensor("op_5671_cast_fp16")]; + tensor var_5672_cast_fp16 = mul(x = var_5641_cast_fp16, y = rotr_27_cast_fp16)[name = tensor("op_5672_cast_fp16")]; + tensor qoi_53_cast_fp16 = add(x = var_5671_cast_fp16, y = var_5672_cast_fp16)[name = tensor("qoi_53_cast_fp16")]; + tensor var_5675_cast_fp16 = mul(x = var_5649_cast_fp16, y = rotr_27_cast_fp16)[name = tensor("op_5675_cast_fp16")]; + tensor var_5676_cast_fp16 = mul(x = var_5657_cast_fp16, y = roti_27_cast_fp16)[name = tensor("op_5676_cast_fp16")]; + tensor kor_53_cast_fp16 = sub(x = var_5675_cast_fp16, y = var_5676_cast_fp16)[name = tensor("kor_53_cast_fp16")]; + tensor var_5679_cast_fp16 = mul(x = var_5649_cast_fp16, y = roti_27_cast_fp16)[name = tensor("op_5679_cast_fp16")]; + tensor var_5680_cast_fp16 = mul(x = var_5657_cast_fp16, y = rotr_27_cast_fp16)[name = tensor("op_5680_cast_fp16")]; + tensor koi_53_cast_fp16 = add(x = var_5679_cast_fp16, y = var_5680_cast_fp16)[name = tensor("koi_53_cast_fp16")]; + tensor qo_27_axis_0 = const()[name = tensor("qo_27_axis_0"), val = tensor(-1)]; + tensor qo_27_cast_fp16 = stack(axis = qo_27_axis_0, values = (qor_53_cast_fp16, qoi_53_cast_fp16))[name = tensor("qo_27_cast_fp16")]; + tensor ko_27_axis_0 = const()[name = tensor("ko_27_axis_0"), val = tensor(-1)]; + tensor ko_27_cast_fp16 = stack(axis = ko_27_axis_0, values = (kor_53_cast_fp16, koi_53_cast_fp16))[name = tensor("ko_27_cast_fp16")]; + tensor var_5713 = const()[name = tensor("op_5713"), val = tensor([1, 256, 16, 64])]; + tensor q_81_cast_fp16 = reshape(shape = var_5713, x = qo_27_cast_fp16)[name = tensor("q_81_cast_fp16")]; + tensor var_5719 = const()[name = tensor("op_5719"), val = tensor([1, 256, 16, 64])]; + tensor k_55_cast_fp16 = reshape(shape = var_5719, x = ko_27_cast_fp16)[name = tensor("k_55_cast_fp16")]; + tensor write_indices_float_79_cast_fp16 = add(x = var_5608_cast_fp16, y = var_393_promoted_to_fp16)[name = tensor("write_indices_float_79_cast_fp16")]; + tensor _inversed_5740_y_0_to_fp16 = const()[name = tensor("_inversed_5740_y_0_to_fp16"), val = tensor(0x1p-9)]; + tensor _inversed_5740_cast_fp16 = mul(x = write_indices_float_79_cast_fp16, y = _inversed_5740_y_0_to_fp16)[name = tensor("_inversed_5740_cast_fp16")]; + tensor var_5741_cast_fp16 = floor(x = _inversed_5740_cast_fp16)[name = tensor("op_5741_cast_fp16")]; + tensor var_5742_to_fp16 = const()[name = tensor("op_5742_to_fp16"), val = tensor(0x1p+9)]; + tensor var_5743_cast_fp16 = mul(x = var_5741_cast_fp16, y = var_5742_to_fp16)[name = tensor("op_5743_cast_fp16")]; + tensor write_indices_float_81_cast_fp16 = sub(x = write_indices_float_79_cast_fp16, y = var_5743_cast_fp16)[name = tensor("write_indices_float_81_cast_fp16")]; + tensor var_5761_value_0_to_fp16 = const()[name = tensor("op_5761_value_0_to_fp16"), val = tensor(0x1.ffp+8)]; + tensor var_5761_cast_fp16 = fill_like(ref_tensor = write_indices_float_81_cast_fp16, value = var_5761_value_0_to_fp16)[name = tensor("op_5761_cast_fp16")]; + tensor write_indices_float_83_cast_fp16 = select(a = write_indices_float_81_cast_fp16, b = var_5761_cast_fp16, cond = token_valid_1_cast_fp16)[name = tensor("write_indices_float_83_cast_fp16")]; + tensor var_5767_dtype_0 = const()[name = tensor("op_5767_dtype_0"), val = tensor("int32")]; + tensor var_5772 = const()[name = tensor("op_5772"), val = tensor([1, 256, 1, 1])]; + tensor write_indices_float_83_cast_fp16_to_int32 = cast(dtype = var_5767_dtype_0, x = write_indices_float_83_cast_fp16)[name = tensor("cast_595")]; + tensor var_5773 = reshape(shape = var_5772, x = write_indices_float_83_cast_fp16_to_int32)[name = tensor("op_5773")]; + tensor write_indices_27_reps_0 = const()[name = tensor("write_indices_27_reps_0"), val = tensor([1, 1, 16, 64])]; + tensor write_indices_27 = tile(reps = write_indices_27_reps_0, x = var_5773)[name = tensor("write_indices_27")]; + tensor var_5783_begin_0 = const()[name = tensor("op_5783_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_5783_end_0 = const()[name = tensor("op_5783_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor var_5783_end_mask_0 = const()[name = tensor("op_5783_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_5783_squeeze_mask_0 = const()[name = tensor("op_5783_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor cache13_to_fp16_dtype_0 = const()[name = tensor("cache13_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor cache13_to_fp16 = cast(dtype = cache13_to_fp16_dtype_0, x = cache13)[name = tensor("cast_594")]; + tensor var_5783_cast_fp16 = slice_by_index(begin = var_5783_begin_0, end = var_5783_end_0, end_mask = var_5783_end_mask_0, squeeze_mask = var_5783_squeeze_mask_0, x = cache13_to_fp16)[name = tensor("op_5783_cast_fp16")]; + tensor var_5785_axis_0 = const()[name = tensor("op_5785_axis_0"), val = tensor(1)]; + tensor var_5785_mode_0 = const()[name = tensor("op_5785_mode_0"), val = tensor("update")]; + tensor var_5785_validate_indices_0 = const()[name = tensor("op_5785_validate_indices_0"), val = tensor(false)]; + tensor var_5785_cast_fp16 = scatter_along_axis(axis = var_5785_axis_0, data = var_5783_cast_fp16, indices = write_indices_27, mode = var_5785_mode_0, updates = k_55_cast_fp16, validate_indices = var_5785_validate_indices_0)[name = tensor("op_5785_cast_fp16")]; + tensor concat_106 = const()[name = tensor("concat_106"), val = tensor([0, 0, 0, 0, 0])]; + tensor concat_107 = const()[name = tensor("concat_107"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_27_internal_tensor_assign_1_stride_0 = const()[name = tensor("new_cache_27_internal_tensor_assign_1_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_27_internal_tensor_assign_1_begin_mask_0 = const()[name = tensor("new_cache_27_internal_tensor_assign_1_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_27_internal_tensor_assign_1_end_mask_0 = const()[name = tensor("new_cache_27_internal_tensor_assign_1_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_27_internal_tensor_assign_1_squeeze_mask_0 = const()[name = tensor("new_cache_27_internal_tensor_assign_1_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_72 = const()[name = tensor("shape_72"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_26 = const()[name = tensor("reduce_prod_26"), val = tensor(1048576)]; + tensor range_1d_26_start_0 = const()[name = tensor("range_1d_26_start_0"), val = tensor(0)]; + tensor range_1d_26_step_0 = const()[name = tensor("range_1d_26_step_0"), val = tensor(1)]; + tensor range_1d_26 = range_1d(end = reduce_prod_26, start = range_1d_26_start_0, step = range_1d_26_step_0)[name = tensor("range_1d_26")]; + tensor reshape_130 = reshape(shape = shape_72, x = range_1d_26)[name = tensor("reshape_130")]; + tensor slice_by_index_26 = slice_by_index(begin = concat_106, begin_mask = new_cache_27_internal_tensor_assign_1_begin_mask_0, end = concat_107, end_mask = new_cache_27_internal_tensor_assign_1_end_mask_0, squeeze_mask = new_cache_27_internal_tensor_assign_1_squeeze_mask_0, stride = new_cache_27_internal_tensor_assign_1_stride_0, x = reshape_130)[name = tensor("slice_by_index_26")]; + tensor reshape_131_shape_0 = const()[name = tensor("reshape_131_shape_0"), val = tensor([-1])]; + tensor reshape_131 = reshape(shape = reshape_131_shape_0, x = slice_by_index_26)[name = tensor("reshape_131")]; + tensor reshape_132_shape_0 = const()[name = tensor("reshape_132_shape_0"), val = tensor([-1])]; + tensor reshape_132_cast_fp16 = reshape(shape = reshape_132_shape_0, x = var_5785_cast_fp16)[name = tensor("reshape_132_cast_fp16")]; + tensor reshape_133_shape_0 = const()[name = tensor("reshape_133_shape_0"), val = tensor([-1])]; + tensor reshape_133_cast_fp16 = reshape(shape = reshape_133_shape_0, x = cache13_to_fp16)[name = tensor("reshape_133_cast_fp16")]; + tensor scatter_26_mode_0 = const()[name = tensor("scatter_26_mode_0"), val = tensor("update")]; + tensor scatter_26_axis_0 = const()[name = tensor("scatter_26_axis_0"), val = tensor(0)]; + tensor scatter_26_validate_indices_0 = const()[name = tensor("scatter_26_validate_indices_0"), val = tensor(false)]; + tensor scatter_26_cast_fp16 = scatter(axis = scatter_26_axis_0, data = reshape_133_cast_fp16, indices = reshape_131, mode = scatter_26_mode_0, updates = reshape_132_cast_fp16, validate_indices = scatter_26_validate_indices_0)[name = tensor("scatter_26_cast_fp16")]; + tensor reshape_134_cast_fp16 = reshape(shape = shape_72, x = scatter_26_cast_fp16)[name = tensor("reshape_134_cast_fp16")]; + tensor var_5793_begin_0 = const()[name = tensor("op_5793_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor var_5793_end_0 = const()[name = tensor("op_5793_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor var_5793_end_mask_0 = const()[name = tensor("op_5793_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_5793_squeeze_mask_0 = const()[name = tensor("op_5793_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor var_5793_cast_fp16 = slice_by_index(begin = var_5793_begin_0, end = var_5793_end_0, end_mask = var_5793_end_mask_0, squeeze_mask = var_5793_squeeze_mask_0, x = reshape_134_cast_fp16)[name = tensor("op_5793_cast_fp16")]; + tensor var_5795_axis_0 = const()[name = tensor("op_5795_axis_0"), val = tensor(1)]; + tensor var_5795_mode_0 = const()[name = tensor("op_5795_mode_0"), val = tensor("update")]; + tensor var_5795_validate_indices_0 = const()[name = tensor("op_5795_validate_indices_0"), val = tensor(false)]; + tensor var_5795_cast_fp16 = scatter_along_axis(axis = var_5795_axis_0, data = var_5793_cast_fp16, indices = write_indices_27, mode = var_5795_mode_0, updates = v_27_cast_fp16, validate_indices = var_5795_validate_indices_0)[name = tensor("op_5795_cast_fp16")]; + tensor concat_108 = const()[name = tensor("concat_108"), val = tensor([1, 0, 0, 0, 0])]; + tensor concat_109 = const()[name = tensor("concat_109"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_27_internal_tensor_assign_2_stride_0 = const()[name = tensor("new_cache_27_internal_tensor_assign_2_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_27_internal_tensor_assign_2_begin_mask_0 = const()[name = tensor("new_cache_27_internal_tensor_assign_2_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_27_internal_tensor_assign_2_end_mask_0 = const()[name = tensor("new_cache_27_internal_tensor_assign_2_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_27_internal_tensor_assign_2_squeeze_mask_0 = const()[name = tensor("new_cache_27_internal_tensor_assign_2_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_73 = const()[name = tensor("shape_73"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_27 = const()[name = tensor("reduce_prod_27"), val = tensor(1048576)]; + tensor range_1d_27_start_0 = const()[name = tensor("range_1d_27_start_0"), val = tensor(0)]; + tensor range_1d_27_step_0 = const()[name = tensor("range_1d_27_step_0"), val = tensor(1)]; + tensor range_1d_27 = range_1d(end = reduce_prod_27, start = range_1d_27_start_0, step = range_1d_27_step_0)[name = tensor("range_1d_27")]; + tensor reshape_135 = reshape(shape = shape_73, x = range_1d_27)[name = tensor("reshape_135")]; + tensor slice_by_index_27 = slice_by_index(begin = concat_108, begin_mask = new_cache_27_internal_tensor_assign_2_begin_mask_0, end = concat_109, end_mask = new_cache_27_internal_tensor_assign_2_end_mask_0, squeeze_mask = new_cache_27_internal_tensor_assign_2_squeeze_mask_0, stride = new_cache_27_internal_tensor_assign_2_stride_0, x = reshape_135)[name = tensor("slice_by_index_27")]; + tensor reshape_136_shape_0 = const()[name = tensor("reshape_136_shape_0"), val = tensor([-1])]; + tensor reshape_136 = reshape(shape = reshape_136_shape_0, x = slice_by_index_27)[name = tensor("reshape_136")]; + tensor reshape_137_shape_0 = const()[name = tensor("reshape_137_shape_0"), val = tensor([-1])]; + tensor reshape_137_cast_fp16 = reshape(shape = reshape_137_shape_0, x = var_5795_cast_fp16)[name = tensor("reshape_137_cast_fp16")]; + tensor reshape_138_shape_0 = const()[name = tensor("reshape_138_shape_0"), val = tensor([-1])]; + tensor reshape_138_cast_fp16 = reshape(shape = reshape_138_shape_0, x = reshape_134_cast_fp16)[name = tensor("reshape_138_cast_fp16")]; + tensor scatter_27_mode_0 = const()[name = tensor("scatter_27_mode_0"), val = tensor("update")]; + tensor scatter_27_axis_0 = const()[name = tensor("scatter_27_axis_0"), val = tensor(0)]; + tensor scatter_27_validate_indices_0 = const()[name = tensor("scatter_27_validate_indices_0"), val = tensor(false)]; + tensor scatter_27_cast_fp16 = scatter(axis = scatter_27_axis_0, data = reshape_138_cast_fp16, indices = reshape_136, mode = scatter_27_mode_0, updates = reshape_137_cast_fp16, validate_indices = scatter_27_validate_indices_0)[name = tensor("scatter_27_cast_fp16")]; + tensor reshape_139_cast_fp16 = reshape(shape = shape_73, x = scatter_27_cast_fp16)[name = tensor("reshape_139_cast_fp16")]; + tensor reshape_139_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("reshape_139_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor keys_79_begin_0 = const()[name = tensor("keys_79_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor keys_79_end_0 = const()[name = tensor("keys_79_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor keys_79_end_mask_0 = const()[name = tensor("keys_79_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor keys_79_squeeze_mask_0 = const()[name = tensor("keys_79_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor keys_79_cast_fp16 = slice_by_index(begin = keys_79_begin_0, end = keys_79_end_0, end_mask = keys_79_end_mask_0, squeeze_mask = keys_79_squeeze_mask_0, x = reshape_139_cast_fp16)[name = tensor("keys_79_cast_fp16")]; + tensor values_79_begin_0 = const()[name = tensor("values_79_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor values_79_end_0 = const()[name = tensor("values_79_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor values_79_end_mask_0 = const()[name = tensor("values_79_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor values_79_squeeze_mask_0 = const()[name = tensor("values_79_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor values_79_cast_fp16 = slice_by_index(begin = values_79_begin_0, end = values_79_end_0, end_mask = values_79_end_mask_0, squeeze_mask = values_79_squeeze_mask_0, x = reshape_139_cast_fp16)[name = tensor("values_79_cast_fp16")]; + tensor var_5807_cast_fp16 = not_equal(x = keys_79_cast_fp16, y = keys_79_cast_fp16)[name = tensor("op_5807_cast_fp16")]; + tensor keys_81_cast_fp16 = select(a = var_470_to_fp16, b = keys_79_cast_fp16, cond = var_5807_cast_fp16)[name = tensor("keys_81_cast_fp16")]; + tensor var_5815_cast_fp16 = not_equal(x = values_79_cast_fp16, y = values_79_cast_fp16)[name = tensor("op_5815_cast_fp16")]; + tensor values_81_cast_fp16 = select(a = var_470_to_fp16, b = values_79_cast_fp16, cond = var_5815_cast_fp16)[name = tensor("values_81_cast_fp16")]; + tensor var_5839 = const()[name = tensor("op_5839"), val = tensor([0, 2, 1, 3])]; + tensor var_5855 = const()[name = tensor("op_5855"), val = tensor([1, 1, 1])]; + tensor var_5856_cast_fp16 = reshape(shape = var_5855, x = position13_to_fp16)[name = tensor("op_5856_cast_fp16")]; + tensor q_positions_27_cast_fp16 = add(x = var_5856_cast_fp16, y = q_offsets_1_promoted_to_fp16)[name = tensor("q_positions_27_cast_fp16")]; + tensor valid_end_27_cast_fp16 = add(x = var_5856_cast_fp16, y = var_536_cast_fp16)[name = tensor("valid_end_27_cast_fp16")]; + tensor valid_mask_27_cast_fp16 = less(x = k_positions_1_promoted_to_fp16, y = valid_end_27_cast_fp16)[name = tensor("valid_mask_27_cast_fp16")]; + tensor causal_mask_27_cast_fp16 = less_equal(x = k_positions_1_promoted_to_fp16, y = q_positions_27_cast_fp16)[name = tensor("causal_mask_27_cast_fp16")]; + tensor var_5884 = logical_and(x = valid_mask_27_cast_fp16, y = causal_mask_27_cast_fp16)[name = tensor("op_5884")]; + tensor attn_mask_27_axes_0 = const()[name = tensor("attn_mask_27_axes_0"), val = tensor([1])]; + tensor attn_mask_27 = expand_dims(axes = attn_mask_27_axes_0, x = var_5884)[name = tensor("attn_mask_27")]; + tensor var_5898_transpose_x_0 = const()[name = tensor("op_5898_transpose_x_0"), val = tensor(false)]; + tensor var_5898_transpose_y_0 = const()[name = tensor("op_5898_transpose_y_0"), val = tensor(false)]; + tensor transpose_95_perm_0 = const()[name = tensor("transpose_95_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_96_perm_0 = const()[name = tensor("transpose_96_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_96 = transpose(perm = transpose_96_perm_0, x = keys_81_cast_fp16)[name = tensor("transpose_152")]; + tensor transpose_95 = transpose(perm = transpose_95_perm_0, x = q_81_cast_fp16)[name = tensor("transpose_153")]; + tensor var_5898_cast_fp16 = matmul(transpose_x = var_5898_transpose_x_0, transpose_y = var_5898_transpose_y_0, x = transpose_95, y = transpose_96)[name = tensor("op_5898_cast_fp16")]; + tensor var_5892_to_fp16 = const()[name = tensor("op_5892_to_fp16"), val = tensor([0x1p-3])]; + tensor attn_weights_79_cast_fp16 = mul(x = var_5898_cast_fp16, y = var_5892_to_fp16)[name = tensor("attn_weights_79_cast_fp16")]; + tensor var_5900 = logical_not(x = attn_mask_27)[name = tensor("op_5900")]; + tensor var_5901_to_fp16 = const()[name = tensor("op_5901_to_fp16"), val = tensor(-inf)]; + tensor attn_weights_81_cast_fp16 = select(a = var_5901_to_fp16, b = attn_weights_79_cast_fp16, cond = var_5900)[name = tensor("attn_weights_81_cast_fp16")]; + tensor var_5903 = const()[name = tensor("op_5903"), val = tensor(-1)]; + tensor attn_weights_83_cast_fp16 = softmax(axis = var_5903, x = attn_weights_81_cast_fp16)[name = tensor("attn_weights_83_cast_fp16")]; + tensor attn_output_27_transpose_x_0 = const()[name = tensor("attn_output_27_transpose_x_0"), val = tensor(false)]; + tensor attn_output_27_transpose_y_0 = const()[name = tensor("attn_output_27_transpose_y_0"), val = tensor(false)]; + tensor values_83_cast_fp16 = transpose(perm = var_5839, x = values_81_cast_fp16)[name = tensor("transpose_154")]; + tensor attn_output_27_cast_fp16 = matmul(transpose_x = attn_output_27_transpose_x_0, transpose_y = attn_output_27_transpose_y_0, x = attn_weights_83_cast_fp16, y = values_83_cast_fp16)[name = tensor("attn_output_27_cast_fp16")]; + tensor var_5911 = const()[name = tensor("op_5911"), val = tensor([0, 2, 1, 3])]; + tensor var_5916 = const()[name = tensor("op_5916"), val = tensor([1, 256, 1024])]; + tensor var_5912_cast_fp16 = transpose(perm = var_5911, x = attn_output_27_cast_fp16)[name = tensor("transpose_151")]; + tensor input_161_cast_fp16 = reshape(shape = var_5916, x = var_5912_cast_fp16)[name = tensor("input_161_cast_fp16")]; + tensor attn13_out_proj_weight_to_fp16 = const()[name = tensor("attn13_out_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(334634496)))]; + tensor linear_53_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = attn13_out_proj_weight_to_fp16, x = input_161_cast_fp16)[name = tensor("linear_53_cast_fp16")]; + tensor var_5925_cast_fp16 = add(x = position13_to_fp16, y = valid_len_to_fp16)[name = tensor("op_5925_cast_fp16")]; + tensor var_5925_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_5925_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor input_163_cast_fp16 = add(x = input_157_cast_fp16, y = linear_53_cast_fp16)[name = tensor("input_163_cast_fp16")]; + tensor input_165_axes_0 = const()[name = tensor("input_165_axes_0"), val = tensor([-1])]; + tensor norm13_2_weight_to_fp16 = const()[name = tensor("norm13_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(336731712)))]; + tensor norm13_2_bias_to_fp16 = const()[name = tensor("norm13_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(336733824)))]; + tensor var_5929_to_fp16 = const()[name = tensor("op_5929_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_165_cast_fp16 = layer_norm(axes = input_165_axes_0, beta = norm13_2_bias_to_fp16, epsilon = var_5929_to_fp16, gamma = norm13_2_weight_to_fp16, x = input_163_cast_fp16)[name = tensor("input_165_cast_fp16")]; + tensor linear13_1_weight_to_fp16 = const()[name = tensor("linear13_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(336735936)))]; + tensor linear_54_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = linear13_1_weight_to_fp16, x = input_165_cast_fp16)[name = tensor("linear_54_cast_fp16")]; + tensor input_167_mode_0 = const()[name = tensor("input_167_mode_0"), val = tensor("EXACT")]; + tensor input_167_cast_fp16 = gelu(mode = input_167_mode_0, x = linear_54_cast_fp16)[name = tensor("input_167_cast_fp16")]; + tensor linear13_2_weight_to_fp16 = const()[name = tensor("linear13_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(345124608)))]; + tensor linear_55_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = linear13_2_weight_to_fp16, x = input_167_cast_fp16)[name = tensor("linear_55_cast_fp16")]; + tensor input_169_cast_fp16 = add(x = input_163_cast_fp16, y = linear_55_cast_fp16)[name = tensor("input_169_cast_fp16")]; + tensor input_171_axes_0 = const()[name = tensor("input_171_axes_0"), val = tensor([-1])]; + tensor norm14_1_weight_to_fp16 = const()[name = tensor("norm14_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(353513280)))]; + tensor norm14_1_bias_to_fp16 = const()[name = tensor("norm14_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(353515392)))]; + tensor var_5946_to_fp16 = const()[name = tensor("op_5946_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_171_cast_fp16 = layer_norm(axes = input_171_axes_0, beta = norm14_1_bias_to_fp16, epsilon = var_5946_to_fp16, gamma = norm14_1_weight_to_fp16, x = input_169_cast_fp16)[name = tensor("input_171_cast_fp16")]; + tensor attn14_in_proj_weight_to_fp16 = const()[name = tensor("attn14_in_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(353517504)))]; + tensor linear_56_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = attn14_in_proj_weight_to_fp16, x = input_171_cast_fp16)[name = tensor("linear_56_cast_fp16")]; + tensor var_5960 = const()[name = tensor("op_5960"), val = tensor([1, 256, 3, 16, 64])]; + tensor qkv_29_cast_fp16 = reshape(shape = var_5960, x = linear_56_cast_fp16)[name = tensor("qkv_29_cast_fp16")]; + tensor q_85_begin_0 = const()[name = tensor("q_85_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor q_85_end_0 = const()[name = tensor("q_85_end_0"), val = tensor([1, 256, 1, 16, 64])]; + tensor q_85_end_mask_0 = const()[name = tensor("q_85_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor q_85_squeeze_mask_0 = const()[name = tensor("q_85_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor q_85_cast_fp16 = slice_by_index(begin = q_85_begin_0, end = q_85_end_0, end_mask = q_85_end_mask_0, squeeze_mask = q_85_squeeze_mask_0, x = qkv_29_cast_fp16)[name = tensor("q_85_cast_fp16")]; + tensor k_57_begin_0 = const()[name = tensor("k_57_begin_0"), val = tensor([0, 0, 1, 0, 0])]; + tensor k_57_end_0 = const()[name = tensor("k_57_end_0"), val = tensor([1, 256, 2, 16, 64])]; + tensor k_57_end_mask_0 = const()[name = tensor("k_57_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor k_57_squeeze_mask_0 = const()[name = tensor("k_57_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor k_57_cast_fp16 = slice_by_index(begin = k_57_begin_0, end = k_57_end_0, end_mask = k_57_end_mask_0, squeeze_mask = k_57_squeeze_mask_0, x = qkv_29_cast_fp16)[name = tensor("k_57_cast_fp16")]; + tensor v_29_begin_0 = const()[name = tensor("v_29_begin_0"), val = tensor([0, 0, 2, 0, 0])]; + tensor v_29_end_0 = const()[name = tensor("v_29_end_0"), val = tensor([1, 256, 3, 16, 64])]; + tensor v_29_end_mask_0 = const()[name = tensor("v_29_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor v_29_squeeze_mask_0 = const()[name = tensor("v_29_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor v_29_cast_fp16 = slice_by_index(begin = v_29_begin_0, end = v_29_end_0, end_mask = v_29_end_mask_0, squeeze_mask = v_29_squeeze_mask_0, x = qkv_29_cast_fp16)[name = tensor("v_29_cast_fp16")]; + tensor var_6018 = const()[name = tensor("op_6018"), val = tensor([1, 1])]; + tensor position14_to_fp16_dtype_0 = const()[name = tensor("position14_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor position14_to_fp16 = cast(dtype = position14_to_fp16_dtype_0, x = position14)[name = tensor("cast_591")]; + tensor var_6019_cast_fp16 = reshape(shape = var_6018, x = position14_to_fp16)[name = tensor("op_6019_cast_fp16")]; + tensor ts_87_cast_fp16 = add(x = ts_1_promoted_to_fp16, y = var_6019_cast_fp16)[name = tensor("ts_87_cast_fp16")]; + tensor var_6026 = const()[name = tensor("op_6026"), val = tensor([1, 256, 1, 1])]; + tensor ts_89_cast_fp16 = reshape(shape = var_6026, x = ts_87_cast_fp16)[name = tensor("ts_89_cast_fp16")]; + tensor var_6033 = const()[name = tensor("op_6033"), val = tensor([1, 256, 16, 32, 2])]; + tensor q_complex_29_cast_fp16 = reshape(shape = var_6033, x = q_85_cast_fp16)[name = tensor("q_complex_29_cast_fp16")]; + tensor var_6040 = const()[name = tensor("op_6040"), val = tensor([1, 256, 16, 32, 2])]; + tensor k_complex_29_cast_fp16 = reshape(shape = var_6040, x = k_57_cast_fp16)[name = tensor("k_complex_29_cast_fp16")]; + tensor var_6044_begin_0 = const()[name = tensor("op_6044_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_6044_end_0 = const()[name = tensor("op_6044_end_0"), val = tensor([1, 256, 16, 32, 1])]; + tensor var_6044_end_mask_0 = const()[name = tensor("op_6044_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_6044_squeeze_mask_0 = const()[name = tensor("op_6044_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_6044_cast_fp16 = slice_by_index(begin = var_6044_begin_0, end = var_6044_end_0, end_mask = var_6044_end_mask_0, squeeze_mask = var_6044_squeeze_mask_0, x = q_complex_29_cast_fp16)[name = tensor("op_6044_cast_fp16")]; + tensor var_6052_begin_0 = const()[name = tensor("op_6052_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_6052_end_0 = const()[name = tensor("op_6052_end_0"), val = tensor([1, 256, 16, 32, 2])]; + tensor var_6052_end_mask_0 = const()[name = tensor("op_6052_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_6052_squeeze_mask_0 = const()[name = tensor("op_6052_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_6052_cast_fp16 = slice_by_index(begin = var_6052_begin_0, end = var_6052_end_0, end_mask = var_6052_end_mask_0, squeeze_mask = var_6052_squeeze_mask_0, x = q_complex_29_cast_fp16)[name = tensor("op_6052_cast_fp16")]; + tensor var_6060_begin_0 = const()[name = tensor("op_6060_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_6060_end_0 = const()[name = tensor("op_6060_end_0"), val = tensor([1, 256, 16, 32, 1])]; + tensor var_6060_end_mask_0 = const()[name = tensor("op_6060_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_6060_squeeze_mask_0 = const()[name = tensor("op_6060_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_6060_cast_fp16 = slice_by_index(begin = var_6060_begin_0, end = var_6060_end_0, end_mask = var_6060_end_mask_0, squeeze_mask = var_6060_squeeze_mask_0, x = k_complex_29_cast_fp16)[name = tensor("op_6060_cast_fp16")]; + tensor var_6068_begin_0 = const()[name = tensor("op_6068_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_6068_end_0 = const()[name = tensor("op_6068_end_0"), val = tensor([1, 256, 16, 32, 2])]; + tensor var_6068_end_mask_0 = const()[name = tensor("op_6068_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_6068_squeeze_mask_0 = const()[name = tensor("op_6068_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_6068_cast_fp16 = slice_by_index(begin = var_6068_begin_0, end = var_6068_end_0, end_mask = var_6068_end_mask_0, squeeze_mask = var_6068_squeeze_mask_0, x = k_complex_29_cast_fp16)[name = tensor("op_6068_cast_fp16")]; + tensor freqs_29_to_fp16 = const()[name = tensor("freqs_29_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(359809024)))]; + tensor var_6074_cast_fp16 = mul(x = freqs_29_to_fp16, y = ts_89_cast_fp16)[name = tensor("op_6074_cast_fp16")]; + tensor rotr_29_cast_fp16 = cos(x = var_6074_cast_fp16)[name = tensor("rotr_29_cast_fp16")]; + tensor roti_29_cast_fp16 = sin(x = var_6074_cast_fp16)[name = tensor("roti_29_cast_fp16")]; + tensor var_6078_cast_fp16 = mul(x = var_6044_cast_fp16, y = rotr_29_cast_fp16)[name = tensor("op_6078_cast_fp16")]; + tensor var_6079_cast_fp16 = mul(x = var_6052_cast_fp16, y = roti_29_cast_fp16)[name = tensor("op_6079_cast_fp16")]; + tensor qor_57_cast_fp16 = sub(x = var_6078_cast_fp16, y = var_6079_cast_fp16)[name = tensor("qor_57_cast_fp16")]; + tensor var_6082_cast_fp16 = mul(x = var_6044_cast_fp16, y = roti_29_cast_fp16)[name = tensor("op_6082_cast_fp16")]; + tensor var_6083_cast_fp16 = mul(x = var_6052_cast_fp16, y = rotr_29_cast_fp16)[name = tensor("op_6083_cast_fp16")]; + tensor qoi_57_cast_fp16 = add(x = var_6082_cast_fp16, y = var_6083_cast_fp16)[name = tensor("qoi_57_cast_fp16")]; + tensor var_6086_cast_fp16 = mul(x = var_6060_cast_fp16, y = rotr_29_cast_fp16)[name = tensor("op_6086_cast_fp16")]; + tensor var_6087_cast_fp16 = mul(x = var_6068_cast_fp16, y = roti_29_cast_fp16)[name = tensor("op_6087_cast_fp16")]; + tensor kor_57_cast_fp16 = sub(x = var_6086_cast_fp16, y = var_6087_cast_fp16)[name = tensor("kor_57_cast_fp16")]; + tensor var_6090_cast_fp16 = mul(x = var_6060_cast_fp16, y = roti_29_cast_fp16)[name = tensor("op_6090_cast_fp16")]; + tensor var_6091_cast_fp16 = mul(x = var_6068_cast_fp16, y = rotr_29_cast_fp16)[name = tensor("op_6091_cast_fp16")]; + tensor koi_57_cast_fp16 = add(x = var_6090_cast_fp16, y = var_6091_cast_fp16)[name = tensor("koi_57_cast_fp16")]; + tensor qo_29_axis_0 = const()[name = tensor("qo_29_axis_0"), val = tensor(-1)]; + tensor qo_29_cast_fp16 = stack(axis = qo_29_axis_0, values = (qor_57_cast_fp16, qoi_57_cast_fp16))[name = tensor("qo_29_cast_fp16")]; + tensor ko_29_axis_0 = const()[name = tensor("ko_29_axis_0"), val = tensor(-1)]; + tensor ko_29_cast_fp16 = stack(axis = ko_29_axis_0, values = (kor_57_cast_fp16, koi_57_cast_fp16))[name = tensor("ko_29_cast_fp16")]; + tensor var_6124 = const()[name = tensor("op_6124"), val = tensor([1, 256, 16, 64])]; + tensor q_87_cast_fp16 = reshape(shape = var_6124, x = qo_29_cast_fp16)[name = tensor("q_87_cast_fp16")]; + tensor var_6130 = const()[name = tensor("op_6130"), val = tensor([1, 256, 16, 64])]; + tensor k_59_cast_fp16 = reshape(shape = var_6130, x = ko_29_cast_fp16)[name = tensor("k_59_cast_fp16")]; + tensor write_indices_float_85_cast_fp16 = add(x = var_6019_cast_fp16, y = var_393_promoted_to_fp16)[name = tensor("write_indices_float_85_cast_fp16")]; + tensor _inversed_6151_y_0_to_fp16 = const()[name = tensor("_inversed_6151_y_0_to_fp16"), val = tensor(0x1p-9)]; + tensor _inversed_6151_cast_fp16 = mul(x = write_indices_float_85_cast_fp16, y = _inversed_6151_y_0_to_fp16)[name = tensor("_inversed_6151_cast_fp16")]; + tensor var_6152_cast_fp16 = floor(x = _inversed_6151_cast_fp16)[name = tensor("op_6152_cast_fp16")]; + tensor var_6153_to_fp16 = const()[name = tensor("op_6153_to_fp16"), val = tensor(0x1p+9)]; + tensor var_6154_cast_fp16 = mul(x = var_6152_cast_fp16, y = var_6153_to_fp16)[name = tensor("op_6154_cast_fp16")]; + tensor write_indices_float_87_cast_fp16 = sub(x = write_indices_float_85_cast_fp16, y = var_6154_cast_fp16)[name = tensor("write_indices_float_87_cast_fp16")]; + tensor var_6172_value_0_to_fp16 = const()[name = tensor("op_6172_value_0_to_fp16"), val = tensor(0x1.ffp+8)]; + tensor var_6172_cast_fp16 = fill_like(ref_tensor = write_indices_float_87_cast_fp16, value = var_6172_value_0_to_fp16)[name = tensor("op_6172_cast_fp16")]; + tensor write_indices_float_89_cast_fp16 = select(a = write_indices_float_87_cast_fp16, b = var_6172_cast_fp16, cond = token_valid_1_cast_fp16)[name = tensor("write_indices_float_89_cast_fp16")]; + tensor var_6178_dtype_0 = const()[name = tensor("op_6178_dtype_0"), val = tensor("int32")]; + tensor var_6183 = const()[name = tensor("op_6183"), val = tensor([1, 256, 1, 1])]; + tensor write_indices_float_89_cast_fp16_to_int32 = cast(dtype = var_6178_dtype_0, x = write_indices_float_89_cast_fp16)[name = tensor("cast_590")]; + tensor var_6184 = reshape(shape = var_6183, x = write_indices_float_89_cast_fp16_to_int32)[name = tensor("op_6184")]; + tensor write_indices_29_reps_0 = const()[name = tensor("write_indices_29_reps_0"), val = tensor([1, 1, 16, 64])]; + tensor write_indices_29 = tile(reps = write_indices_29_reps_0, x = var_6184)[name = tensor("write_indices_29")]; + tensor var_6194_begin_0 = const()[name = tensor("op_6194_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_6194_end_0 = const()[name = tensor("op_6194_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor var_6194_end_mask_0 = const()[name = tensor("op_6194_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_6194_squeeze_mask_0 = const()[name = tensor("op_6194_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor cache14_to_fp16_dtype_0 = const()[name = tensor("cache14_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor cache14_to_fp16 = cast(dtype = cache14_to_fp16_dtype_0, x = cache14)[name = tensor("cast_589")]; + tensor var_6194_cast_fp16 = slice_by_index(begin = var_6194_begin_0, end = var_6194_end_0, end_mask = var_6194_end_mask_0, squeeze_mask = var_6194_squeeze_mask_0, x = cache14_to_fp16)[name = tensor("op_6194_cast_fp16")]; + tensor var_6196_axis_0 = const()[name = tensor("op_6196_axis_0"), val = tensor(1)]; + tensor var_6196_mode_0 = const()[name = tensor("op_6196_mode_0"), val = tensor("update")]; + tensor var_6196_validate_indices_0 = const()[name = tensor("op_6196_validate_indices_0"), val = tensor(false)]; + tensor var_6196_cast_fp16 = scatter_along_axis(axis = var_6196_axis_0, data = var_6194_cast_fp16, indices = write_indices_29, mode = var_6196_mode_0, updates = k_59_cast_fp16, validate_indices = var_6196_validate_indices_0)[name = tensor("op_6196_cast_fp16")]; + tensor concat_114 = const()[name = tensor("concat_114"), val = tensor([0, 0, 0, 0, 0])]; + tensor concat_115 = const()[name = tensor("concat_115"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_29_internal_tensor_assign_1_stride_0 = const()[name = tensor("new_cache_29_internal_tensor_assign_1_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_29_internal_tensor_assign_1_begin_mask_0 = const()[name = tensor("new_cache_29_internal_tensor_assign_1_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_29_internal_tensor_assign_1_end_mask_0 = const()[name = tensor("new_cache_29_internal_tensor_assign_1_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_29_internal_tensor_assign_1_squeeze_mask_0 = const()[name = tensor("new_cache_29_internal_tensor_assign_1_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_74 = const()[name = tensor("shape_74"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_28 = const()[name = tensor("reduce_prod_28"), val = tensor(1048576)]; + tensor range_1d_28_start_0 = const()[name = tensor("range_1d_28_start_0"), val = tensor(0)]; + tensor range_1d_28_step_0 = const()[name = tensor("range_1d_28_step_0"), val = tensor(1)]; + tensor range_1d_28 = range_1d(end = reduce_prod_28, start = range_1d_28_start_0, step = range_1d_28_step_0)[name = tensor("range_1d_28")]; + tensor reshape_140 = reshape(shape = shape_74, x = range_1d_28)[name = tensor("reshape_140")]; + tensor slice_by_index_28 = slice_by_index(begin = concat_114, begin_mask = new_cache_29_internal_tensor_assign_1_begin_mask_0, end = concat_115, end_mask = new_cache_29_internal_tensor_assign_1_end_mask_0, squeeze_mask = new_cache_29_internal_tensor_assign_1_squeeze_mask_0, stride = new_cache_29_internal_tensor_assign_1_stride_0, x = reshape_140)[name = tensor("slice_by_index_28")]; + tensor reshape_141_shape_0 = const()[name = tensor("reshape_141_shape_0"), val = tensor([-1])]; + tensor reshape_141 = reshape(shape = reshape_141_shape_0, x = slice_by_index_28)[name = tensor("reshape_141")]; + tensor reshape_142_shape_0 = const()[name = tensor("reshape_142_shape_0"), val = tensor([-1])]; + tensor reshape_142_cast_fp16 = reshape(shape = reshape_142_shape_0, x = var_6196_cast_fp16)[name = tensor("reshape_142_cast_fp16")]; + tensor reshape_143_shape_0 = const()[name = tensor("reshape_143_shape_0"), val = tensor([-1])]; + tensor reshape_143_cast_fp16 = reshape(shape = reshape_143_shape_0, x = cache14_to_fp16)[name = tensor("reshape_143_cast_fp16")]; + tensor scatter_28_mode_0 = const()[name = tensor("scatter_28_mode_0"), val = tensor("update")]; + tensor scatter_28_axis_0 = const()[name = tensor("scatter_28_axis_0"), val = tensor(0)]; + tensor scatter_28_validate_indices_0 = const()[name = tensor("scatter_28_validate_indices_0"), val = tensor(false)]; + tensor scatter_28_cast_fp16 = scatter(axis = scatter_28_axis_0, data = reshape_143_cast_fp16, indices = reshape_141, mode = scatter_28_mode_0, updates = reshape_142_cast_fp16, validate_indices = scatter_28_validate_indices_0)[name = tensor("scatter_28_cast_fp16")]; + tensor reshape_144_cast_fp16 = reshape(shape = shape_74, x = scatter_28_cast_fp16)[name = tensor("reshape_144_cast_fp16")]; + tensor var_6204_begin_0 = const()[name = tensor("op_6204_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor var_6204_end_0 = const()[name = tensor("op_6204_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor var_6204_end_mask_0 = const()[name = tensor("op_6204_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_6204_squeeze_mask_0 = const()[name = tensor("op_6204_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor var_6204_cast_fp16 = slice_by_index(begin = var_6204_begin_0, end = var_6204_end_0, end_mask = var_6204_end_mask_0, squeeze_mask = var_6204_squeeze_mask_0, x = reshape_144_cast_fp16)[name = tensor("op_6204_cast_fp16")]; + tensor var_6206_axis_0 = const()[name = tensor("op_6206_axis_0"), val = tensor(1)]; + tensor var_6206_mode_0 = const()[name = tensor("op_6206_mode_0"), val = tensor("update")]; + tensor var_6206_validate_indices_0 = const()[name = tensor("op_6206_validate_indices_0"), val = tensor(false)]; + tensor var_6206_cast_fp16 = scatter_along_axis(axis = var_6206_axis_0, data = var_6204_cast_fp16, indices = write_indices_29, mode = var_6206_mode_0, updates = v_29_cast_fp16, validate_indices = var_6206_validate_indices_0)[name = tensor("op_6206_cast_fp16")]; + tensor concat_116 = const()[name = tensor("concat_116"), val = tensor([1, 0, 0, 0, 0])]; + tensor concat_117 = const()[name = tensor("concat_117"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_29_internal_tensor_assign_2_stride_0 = const()[name = tensor("new_cache_29_internal_tensor_assign_2_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_29_internal_tensor_assign_2_begin_mask_0 = const()[name = tensor("new_cache_29_internal_tensor_assign_2_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_29_internal_tensor_assign_2_end_mask_0 = const()[name = tensor("new_cache_29_internal_tensor_assign_2_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_29_internal_tensor_assign_2_squeeze_mask_0 = const()[name = tensor("new_cache_29_internal_tensor_assign_2_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_75 = const()[name = tensor("shape_75"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_29 = const()[name = tensor("reduce_prod_29"), val = tensor(1048576)]; + tensor range_1d_29_start_0 = const()[name = tensor("range_1d_29_start_0"), val = tensor(0)]; + tensor range_1d_29_step_0 = const()[name = tensor("range_1d_29_step_0"), val = tensor(1)]; + tensor range_1d_29 = range_1d(end = reduce_prod_29, start = range_1d_29_start_0, step = range_1d_29_step_0)[name = tensor("range_1d_29")]; + tensor reshape_145 = reshape(shape = shape_75, x = range_1d_29)[name = tensor("reshape_145")]; + tensor slice_by_index_29 = slice_by_index(begin = concat_116, begin_mask = new_cache_29_internal_tensor_assign_2_begin_mask_0, end = concat_117, end_mask = new_cache_29_internal_tensor_assign_2_end_mask_0, squeeze_mask = new_cache_29_internal_tensor_assign_2_squeeze_mask_0, stride = new_cache_29_internal_tensor_assign_2_stride_0, x = reshape_145)[name = tensor("slice_by_index_29")]; + tensor reshape_146_shape_0 = const()[name = tensor("reshape_146_shape_0"), val = tensor([-1])]; + tensor reshape_146 = reshape(shape = reshape_146_shape_0, x = slice_by_index_29)[name = tensor("reshape_146")]; + tensor reshape_147_shape_0 = const()[name = tensor("reshape_147_shape_0"), val = tensor([-1])]; + tensor reshape_147_cast_fp16 = reshape(shape = reshape_147_shape_0, x = var_6206_cast_fp16)[name = tensor("reshape_147_cast_fp16")]; + tensor reshape_148_shape_0 = const()[name = tensor("reshape_148_shape_0"), val = tensor([-1])]; + tensor reshape_148_cast_fp16 = reshape(shape = reshape_148_shape_0, x = reshape_144_cast_fp16)[name = tensor("reshape_148_cast_fp16")]; + tensor scatter_29_mode_0 = const()[name = tensor("scatter_29_mode_0"), val = tensor("update")]; + tensor scatter_29_axis_0 = const()[name = tensor("scatter_29_axis_0"), val = tensor(0)]; + tensor scatter_29_validate_indices_0 = const()[name = tensor("scatter_29_validate_indices_0"), val = tensor(false)]; + tensor scatter_29_cast_fp16 = scatter(axis = scatter_29_axis_0, data = reshape_148_cast_fp16, indices = reshape_146, mode = scatter_29_mode_0, updates = reshape_147_cast_fp16, validate_indices = scatter_29_validate_indices_0)[name = tensor("scatter_29_cast_fp16")]; + tensor reshape_149_cast_fp16 = reshape(shape = shape_75, x = scatter_29_cast_fp16)[name = tensor("reshape_149_cast_fp16")]; + tensor reshape_149_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("reshape_149_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor keys_85_begin_0 = const()[name = tensor("keys_85_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor keys_85_end_0 = const()[name = tensor("keys_85_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor keys_85_end_mask_0 = const()[name = tensor("keys_85_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor keys_85_squeeze_mask_0 = const()[name = tensor("keys_85_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor keys_85_cast_fp16 = slice_by_index(begin = keys_85_begin_0, end = keys_85_end_0, end_mask = keys_85_end_mask_0, squeeze_mask = keys_85_squeeze_mask_0, x = reshape_149_cast_fp16)[name = tensor("keys_85_cast_fp16")]; + tensor values_85_begin_0 = const()[name = tensor("values_85_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor values_85_end_0 = const()[name = tensor("values_85_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor values_85_end_mask_0 = const()[name = tensor("values_85_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor values_85_squeeze_mask_0 = const()[name = tensor("values_85_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor values_85_cast_fp16 = slice_by_index(begin = values_85_begin_0, end = values_85_end_0, end_mask = values_85_end_mask_0, squeeze_mask = values_85_squeeze_mask_0, x = reshape_149_cast_fp16)[name = tensor("values_85_cast_fp16")]; + tensor var_6218_cast_fp16 = not_equal(x = keys_85_cast_fp16, y = keys_85_cast_fp16)[name = tensor("op_6218_cast_fp16")]; + tensor keys_87_cast_fp16 = select(a = var_470_to_fp16, b = keys_85_cast_fp16, cond = var_6218_cast_fp16)[name = tensor("keys_87_cast_fp16")]; + tensor var_6226_cast_fp16 = not_equal(x = values_85_cast_fp16, y = values_85_cast_fp16)[name = tensor("op_6226_cast_fp16")]; + tensor values_87_cast_fp16 = select(a = var_470_to_fp16, b = values_85_cast_fp16, cond = var_6226_cast_fp16)[name = tensor("values_87_cast_fp16")]; + tensor var_6250 = const()[name = tensor("op_6250"), val = tensor([0, 2, 1, 3])]; + tensor var_6266 = const()[name = tensor("op_6266"), val = tensor([1, 1, 1])]; + tensor var_6267_cast_fp16 = reshape(shape = var_6266, x = position14_to_fp16)[name = tensor("op_6267_cast_fp16")]; + tensor q_positions_29_cast_fp16 = add(x = var_6267_cast_fp16, y = q_offsets_1_promoted_to_fp16)[name = tensor("q_positions_29_cast_fp16")]; + tensor valid_end_29_cast_fp16 = add(x = var_6267_cast_fp16, y = var_536_cast_fp16)[name = tensor("valid_end_29_cast_fp16")]; + tensor valid_mask_29_cast_fp16 = less(x = k_positions_1_promoted_to_fp16, y = valid_end_29_cast_fp16)[name = tensor("valid_mask_29_cast_fp16")]; + tensor causal_mask_29_cast_fp16 = less_equal(x = k_positions_1_promoted_to_fp16, y = q_positions_29_cast_fp16)[name = tensor("causal_mask_29_cast_fp16")]; + tensor var_6295 = logical_and(x = valid_mask_29_cast_fp16, y = causal_mask_29_cast_fp16)[name = tensor("op_6295")]; + tensor attn_mask_29_axes_0 = const()[name = tensor("attn_mask_29_axes_0"), val = tensor([1])]; + tensor attn_mask_29 = expand_dims(axes = attn_mask_29_axes_0, x = var_6295)[name = tensor("attn_mask_29")]; + tensor var_6309_transpose_x_0 = const()[name = tensor("op_6309_transpose_x_0"), val = tensor(false)]; + tensor var_6309_transpose_y_0 = const()[name = tensor("op_6309_transpose_y_0"), val = tensor(false)]; + tensor transpose_97_perm_0 = const()[name = tensor("transpose_97_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_98_perm_0 = const()[name = tensor("transpose_98_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_98 = transpose(perm = transpose_98_perm_0, x = keys_87_cast_fp16)[name = tensor("transpose_148")]; + tensor transpose_97 = transpose(perm = transpose_97_perm_0, x = q_87_cast_fp16)[name = tensor("transpose_149")]; + tensor var_6309_cast_fp16 = matmul(transpose_x = var_6309_transpose_x_0, transpose_y = var_6309_transpose_y_0, x = transpose_97, y = transpose_98)[name = tensor("op_6309_cast_fp16")]; + tensor var_6303_to_fp16 = const()[name = tensor("op_6303_to_fp16"), val = tensor([0x1p-3])]; + tensor attn_weights_85_cast_fp16 = mul(x = var_6309_cast_fp16, y = var_6303_to_fp16)[name = tensor("attn_weights_85_cast_fp16")]; + tensor var_6311 = logical_not(x = attn_mask_29)[name = tensor("op_6311")]; + tensor var_6312_to_fp16 = const()[name = tensor("op_6312_to_fp16"), val = tensor(-inf)]; + tensor attn_weights_87_cast_fp16 = select(a = var_6312_to_fp16, b = attn_weights_85_cast_fp16, cond = var_6311)[name = tensor("attn_weights_87_cast_fp16")]; + tensor var_6314 = const()[name = tensor("op_6314"), val = tensor(-1)]; + tensor attn_weights_89_cast_fp16 = softmax(axis = var_6314, x = attn_weights_87_cast_fp16)[name = tensor("attn_weights_89_cast_fp16")]; + tensor attn_output_29_transpose_x_0 = const()[name = tensor("attn_output_29_transpose_x_0"), val = tensor(false)]; + tensor attn_output_29_transpose_y_0 = const()[name = tensor("attn_output_29_transpose_y_0"), val = tensor(false)]; + tensor values_89_cast_fp16 = transpose(perm = var_6250, x = values_87_cast_fp16)[name = tensor("transpose_150")]; + tensor attn_output_29_cast_fp16 = matmul(transpose_x = attn_output_29_transpose_x_0, transpose_y = attn_output_29_transpose_y_0, x = attn_weights_89_cast_fp16, y = values_89_cast_fp16)[name = tensor("attn_output_29_cast_fp16")]; + tensor var_6322 = const()[name = tensor("op_6322"), val = tensor([0, 2, 1, 3])]; + tensor var_6327 = const()[name = tensor("op_6327"), val = tensor([1, 256, 1024])]; + tensor var_6323_cast_fp16 = transpose(perm = var_6322, x = attn_output_29_cast_fp16)[name = tensor("transpose_147")]; + tensor input_173_cast_fp16 = reshape(shape = var_6327, x = var_6323_cast_fp16)[name = tensor("input_173_cast_fp16")]; + tensor attn14_out_proj_weight_to_fp16 = const()[name = tensor("attn14_out_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(359809152)))]; + tensor linear_57_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = attn14_out_proj_weight_to_fp16, x = input_173_cast_fp16)[name = tensor("linear_57_cast_fp16")]; + tensor var_6336_cast_fp16 = add(x = position14_to_fp16, y = valid_len_to_fp16)[name = tensor("op_6336_cast_fp16")]; + tensor var_6336_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_6336_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor input_175_cast_fp16 = add(x = input_169_cast_fp16, y = linear_57_cast_fp16)[name = tensor("input_175_cast_fp16")]; + tensor input_177_axes_0 = const()[name = tensor("input_177_axes_0"), val = tensor([-1])]; + tensor norm14_2_weight_to_fp16 = const()[name = tensor("norm14_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(361906368)))]; + tensor norm14_2_bias_to_fp16 = const()[name = tensor("norm14_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(361908480)))]; + tensor var_6340_to_fp16 = const()[name = tensor("op_6340_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_177_cast_fp16 = layer_norm(axes = input_177_axes_0, beta = norm14_2_bias_to_fp16, epsilon = var_6340_to_fp16, gamma = norm14_2_weight_to_fp16, x = input_175_cast_fp16)[name = tensor("input_177_cast_fp16")]; + tensor linear14_1_weight_to_fp16 = const()[name = tensor("linear14_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(361910592)))]; + tensor linear_58_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = linear14_1_weight_to_fp16, x = input_177_cast_fp16)[name = tensor("linear_58_cast_fp16")]; + tensor input_179_mode_0 = const()[name = tensor("input_179_mode_0"), val = tensor("EXACT")]; + tensor input_179_cast_fp16 = gelu(mode = input_179_mode_0, x = linear_58_cast_fp16)[name = tensor("input_179_cast_fp16")]; + tensor linear14_2_weight_to_fp16 = const()[name = tensor("linear14_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(370299264)))]; + tensor linear_59_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = linear14_2_weight_to_fp16, x = input_179_cast_fp16)[name = tensor("linear_59_cast_fp16")]; + tensor input_181_cast_fp16 = add(x = input_175_cast_fp16, y = linear_59_cast_fp16)[name = tensor("input_181_cast_fp16")]; + tensor input_183_axes_0 = const()[name = tensor("input_183_axes_0"), val = tensor([-1])]; + tensor norm15_1_weight_to_fp16 = const()[name = tensor("norm15_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(378687936)))]; + tensor norm15_1_bias_to_fp16 = const()[name = tensor("norm15_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(378690048)))]; + tensor var_6357_to_fp16 = const()[name = tensor("op_6357_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_183_cast_fp16 = layer_norm(axes = input_183_axes_0, beta = norm15_1_bias_to_fp16, epsilon = var_6357_to_fp16, gamma = norm15_1_weight_to_fp16, x = input_181_cast_fp16)[name = tensor("input_183_cast_fp16")]; + tensor attn15_in_proj_weight_to_fp16 = const()[name = tensor("attn15_in_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(378692160)))]; + tensor linear_60_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = attn15_in_proj_weight_to_fp16, x = input_183_cast_fp16)[name = tensor("linear_60_cast_fp16")]; + tensor var_6371 = const()[name = tensor("op_6371"), val = tensor([1, 256, 3, 16, 64])]; + tensor qkv_31_cast_fp16 = reshape(shape = var_6371, x = linear_60_cast_fp16)[name = tensor("qkv_31_cast_fp16")]; + tensor q_91_begin_0 = const()[name = tensor("q_91_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor q_91_end_0 = const()[name = tensor("q_91_end_0"), val = tensor([1, 256, 1, 16, 64])]; + tensor q_91_end_mask_0 = const()[name = tensor("q_91_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor q_91_squeeze_mask_0 = const()[name = tensor("q_91_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor q_91_cast_fp16 = slice_by_index(begin = q_91_begin_0, end = q_91_end_0, end_mask = q_91_end_mask_0, squeeze_mask = q_91_squeeze_mask_0, x = qkv_31_cast_fp16)[name = tensor("q_91_cast_fp16")]; + tensor k_61_begin_0 = const()[name = tensor("k_61_begin_0"), val = tensor([0, 0, 1, 0, 0])]; + tensor k_61_end_0 = const()[name = tensor("k_61_end_0"), val = tensor([1, 256, 2, 16, 64])]; + tensor k_61_end_mask_0 = const()[name = tensor("k_61_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor k_61_squeeze_mask_0 = const()[name = tensor("k_61_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor k_61_cast_fp16 = slice_by_index(begin = k_61_begin_0, end = k_61_end_0, end_mask = k_61_end_mask_0, squeeze_mask = k_61_squeeze_mask_0, x = qkv_31_cast_fp16)[name = tensor("k_61_cast_fp16")]; + tensor v_31_begin_0 = const()[name = tensor("v_31_begin_0"), val = tensor([0, 0, 2, 0, 0])]; + tensor v_31_end_0 = const()[name = tensor("v_31_end_0"), val = tensor([1, 256, 3, 16, 64])]; + tensor v_31_end_mask_0 = const()[name = tensor("v_31_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor v_31_squeeze_mask_0 = const()[name = tensor("v_31_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor v_31_cast_fp16 = slice_by_index(begin = v_31_begin_0, end = v_31_end_0, end_mask = v_31_end_mask_0, squeeze_mask = v_31_squeeze_mask_0, x = qkv_31_cast_fp16)[name = tensor("v_31_cast_fp16")]; + tensor var_6429 = const()[name = tensor("op_6429"), val = tensor([1, 1])]; + tensor position15_to_fp16_dtype_0 = const()[name = tensor("position15_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor position15_to_fp16 = cast(dtype = position15_to_fp16_dtype_0, x = position15)[name = tensor("cast_586")]; + tensor var_6430_cast_fp16 = reshape(shape = var_6429, x = position15_to_fp16)[name = tensor("op_6430_cast_fp16")]; + tensor ts_93_cast_fp16 = add(x = ts_1_promoted_to_fp16, y = var_6430_cast_fp16)[name = tensor("ts_93_cast_fp16")]; + tensor var_6437 = const()[name = tensor("op_6437"), val = tensor([1, 256, 1, 1])]; + tensor ts_95_cast_fp16 = reshape(shape = var_6437, x = ts_93_cast_fp16)[name = tensor("ts_95_cast_fp16")]; + tensor var_6444 = const()[name = tensor("op_6444"), val = tensor([1, 256, 16, 32, 2])]; + tensor q_complex_31_cast_fp16 = reshape(shape = var_6444, x = q_91_cast_fp16)[name = tensor("q_complex_31_cast_fp16")]; + tensor var_6451 = const()[name = tensor("op_6451"), val = tensor([1, 256, 16, 32, 2])]; + tensor k_complex_31_cast_fp16 = reshape(shape = var_6451, x = k_61_cast_fp16)[name = tensor("k_complex_31_cast_fp16")]; + tensor var_6455_begin_0 = const()[name = tensor("op_6455_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_6455_end_0 = const()[name = tensor("op_6455_end_0"), val = tensor([1, 256, 16, 32, 1])]; + tensor var_6455_end_mask_0 = const()[name = tensor("op_6455_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_6455_squeeze_mask_0 = const()[name = tensor("op_6455_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_6455_cast_fp16 = slice_by_index(begin = var_6455_begin_0, end = var_6455_end_0, end_mask = var_6455_end_mask_0, squeeze_mask = var_6455_squeeze_mask_0, x = q_complex_31_cast_fp16)[name = tensor("op_6455_cast_fp16")]; + tensor var_6463_begin_0 = const()[name = tensor("op_6463_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_6463_end_0 = const()[name = tensor("op_6463_end_0"), val = tensor([1, 256, 16, 32, 2])]; + tensor var_6463_end_mask_0 = const()[name = tensor("op_6463_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_6463_squeeze_mask_0 = const()[name = tensor("op_6463_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_6463_cast_fp16 = slice_by_index(begin = var_6463_begin_0, end = var_6463_end_0, end_mask = var_6463_end_mask_0, squeeze_mask = var_6463_squeeze_mask_0, x = q_complex_31_cast_fp16)[name = tensor("op_6463_cast_fp16")]; + tensor var_6471_begin_0 = const()[name = tensor("op_6471_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_6471_end_0 = const()[name = tensor("op_6471_end_0"), val = tensor([1, 256, 16, 32, 1])]; + tensor var_6471_end_mask_0 = const()[name = tensor("op_6471_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_6471_squeeze_mask_0 = const()[name = tensor("op_6471_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_6471_cast_fp16 = slice_by_index(begin = var_6471_begin_0, end = var_6471_end_0, end_mask = var_6471_end_mask_0, squeeze_mask = var_6471_squeeze_mask_0, x = k_complex_31_cast_fp16)[name = tensor("op_6471_cast_fp16")]; + tensor var_6479_begin_0 = const()[name = tensor("op_6479_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_6479_end_0 = const()[name = tensor("op_6479_end_0"), val = tensor([1, 256, 16, 32, 2])]; + tensor var_6479_end_mask_0 = const()[name = tensor("op_6479_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_6479_squeeze_mask_0 = const()[name = tensor("op_6479_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_6479_cast_fp16 = slice_by_index(begin = var_6479_begin_0, end = var_6479_end_0, end_mask = var_6479_end_mask_0, squeeze_mask = var_6479_squeeze_mask_0, x = k_complex_31_cast_fp16)[name = tensor("op_6479_cast_fp16")]; + tensor freqs_31_to_fp16 = const()[name = tensor("freqs_31_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(384983680)))]; + tensor var_6485_cast_fp16 = mul(x = freqs_31_to_fp16, y = ts_95_cast_fp16)[name = tensor("op_6485_cast_fp16")]; + tensor rotr_31_cast_fp16 = cos(x = var_6485_cast_fp16)[name = tensor("rotr_31_cast_fp16")]; + tensor roti_31_cast_fp16 = sin(x = var_6485_cast_fp16)[name = tensor("roti_31_cast_fp16")]; + tensor var_6489_cast_fp16 = mul(x = var_6455_cast_fp16, y = rotr_31_cast_fp16)[name = tensor("op_6489_cast_fp16")]; + tensor var_6490_cast_fp16 = mul(x = var_6463_cast_fp16, y = roti_31_cast_fp16)[name = tensor("op_6490_cast_fp16")]; + tensor qor_61_cast_fp16 = sub(x = var_6489_cast_fp16, y = var_6490_cast_fp16)[name = tensor("qor_61_cast_fp16")]; + tensor var_6493_cast_fp16 = mul(x = var_6455_cast_fp16, y = roti_31_cast_fp16)[name = tensor("op_6493_cast_fp16")]; + tensor var_6494_cast_fp16 = mul(x = var_6463_cast_fp16, y = rotr_31_cast_fp16)[name = tensor("op_6494_cast_fp16")]; + tensor qoi_61_cast_fp16 = add(x = var_6493_cast_fp16, y = var_6494_cast_fp16)[name = tensor("qoi_61_cast_fp16")]; + tensor var_6497_cast_fp16 = mul(x = var_6471_cast_fp16, y = rotr_31_cast_fp16)[name = tensor("op_6497_cast_fp16")]; + tensor var_6498_cast_fp16 = mul(x = var_6479_cast_fp16, y = roti_31_cast_fp16)[name = tensor("op_6498_cast_fp16")]; + tensor kor_61_cast_fp16 = sub(x = var_6497_cast_fp16, y = var_6498_cast_fp16)[name = tensor("kor_61_cast_fp16")]; + tensor var_6501_cast_fp16 = mul(x = var_6471_cast_fp16, y = roti_31_cast_fp16)[name = tensor("op_6501_cast_fp16")]; + tensor var_6502_cast_fp16 = mul(x = var_6479_cast_fp16, y = rotr_31_cast_fp16)[name = tensor("op_6502_cast_fp16")]; + tensor koi_61_cast_fp16 = add(x = var_6501_cast_fp16, y = var_6502_cast_fp16)[name = tensor("koi_61_cast_fp16")]; + tensor qo_31_axis_0 = const()[name = tensor("qo_31_axis_0"), val = tensor(-1)]; + tensor qo_31_cast_fp16 = stack(axis = qo_31_axis_0, values = (qor_61_cast_fp16, qoi_61_cast_fp16))[name = tensor("qo_31_cast_fp16")]; + tensor ko_31_axis_0 = const()[name = tensor("ko_31_axis_0"), val = tensor(-1)]; + tensor ko_31_cast_fp16 = stack(axis = ko_31_axis_0, values = (kor_61_cast_fp16, koi_61_cast_fp16))[name = tensor("ko_31_cast_fp16")]; + tensor var_6535 = const()[name = tensor("op_6535"), val = tensor([1, 256, 16, 64])]; + tensor q_93_cast_fp16 = reshape(shape = var_6535, x = qo_31_cast_fp16)[name = tensor("q_93_cast_fp16")]; + tensor var_6541 = const()[name = tensor("op_6541"), val = tensor([1, 256, 16, 64])]; + tensor k_63_cast_fp16 = reshape(shape = var_6541, x = ko_31_cast_fp16)[name = tensor("k_63_cast_fp16")]; + tensor write_indices_float_91_cast_fp16 = add(x = var_6430_cast_fp16, y = var_393_promoted_to_fp16)[name = tensor("write_indices_float_91_cast_fp16")]; + tensor _inversed_6562_y_0_to_fp16 = const()[name = tensor("_inversed_6562_y_0_to_fp16"), val = tensor(0x1p-9)]; + tensor _inversed_6562_cast_fp16 = mul(x = write_indices_float_91_cast_fp16, y = _inversed_6562_y_0_to_fp16)[name = tensor("_inversed_6562_cast_fp16")]; + tensor var_6563_cast_fp16 = floor(x = _inversed_6562_cast_fp16)[name = tensor("op_6563_cast_fp16")]; + tensor var_6564_to_fp16 = const()[name = tensor("op_6564_to_fp16"), val = tensor(0x1p+9)]; + tensor var_6565_cast_fp16 = mul(x = var_6563_cast_fp16, y = var_6564_to_fp16)[name = tensor("op_6565_cast_fp16")]; + tensor write_indices_float_93_cast_fp16 = sub(x = write_indices_float_91_cast_fp16, y = var_6565_cast_fp16)[name = tensor("write_indices_float_93_cast_fp16")]; + tensor var_6583_value_0_to_fp16 = const()[name = tensor("op_6583_value_0_to_fp16"), val = tensor(0x1.ffp+8)]; + tensor var_6583_cast_fp16 = fill_like(ref_tensor = write_indices_float_93_cast_fp16, value = var_6583_value_0_to_fp16)[name = tensor("op_6583_cast_fp16")]; + tensor write_indices_float_95_cast_fp16 = select(a = write_indices_float_93_cast_fp16, b = var_6583_cast_fp16, cond = token_valid_1_cast_fp16)[name = tensor("write_indices_float_95_cast_fp16")]; + tensor var_6589_dtype_0 = const()[name = tensor("op_6589_dtype_0"), val = tensor("int32")]; + tensor var_6594 = const()[name = tensor("op_6594"), val = tensor([1, 256, 1, 1])]; + tensor write_indices_float_95_cast_fp16_to_int32 = cast(dtype = var_6589_dtype_0, x = write_indices_float_95_cast_fp16)[name = tensor("cast_585")]; + tensor var_6595 = reshape(shape = var_6594, x = write_indices_float_95_cast_fp16_to_int32)[name = tensor("op_6595")]; + tensor write_indices_31_reps_0 = const()[name = tensor("write_indices_31_reps_0"), val = tensor([1, 1, 16, 64])]; + tensor write_indices_31 = tile(reps = write_indices_31_reps_0, x = var_6595)[name = tensor("write_indices_31")]; + tensor var_6605_begin_0 = const()[name = tensor("op_6605_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_6605_end_0 = const()[name = tensor("op_6605_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor var_6605_end_mask_0 = const()[name = tensor("op_6605_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_6605_squeeze_mask_0 = const()[name = tensor("op_6605_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor cache15_to_fp16_dtype_0 = const()[name = tensor("cache15_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor cache15_to_fp16 = cast(dtype = cache15_to_fp16_dtype_0, x = cache15)[name = tensor("cast_584")]; + tensor var_6605_cast_fp16 = slice_by_index(begin = var_6605_begin_0, end = var_6605_end_0, end_mask = var_6605_end_mask_0, squeeze_mask = var_6605_squeeze_mask_0, x = cache15_to_fp16)[name = tensor("op_6605_cast_fp16")]; + tensor var_6607_axis_0 = const()[name = tensor("op_6607_axis_0"), val = tensor(1)]; + tensor var_6607_mode_0 = const()[name = tensor("op_6607_mode_0"), val = tensor("update")]; + tensor var_6607_validate_indices_0 = const()[name = tensor("op_6607_validate_indices_0"), val = tensor(false)]; + tensor var_6607_cast_fp16 = scatter_along_axis(axis = var_6607_axis_0, data = var_6605_cast_fp16, indices = write_indices_31, mode = var_6607_mode_0, updates = k_63_cast_fp16, validate_indices = var_6607_validate_indices_0)[name = tensor("op_6607_cast_fp16")]; + tensor concat_122 = const()[name = tensor("concat_122"), val = tensor([0, 0, 0, 0, 0])]; + tensor concat_123 = const()[name = tensor("concat_123"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_31_internal_tensor_assign_1_stride_0 = const()[name = tensor("new_cache_31_internal_tensor_assign_1_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_31_internal_tensor_assign_1_begin_mask_0 = const()[name = tensor("new_cache_31_internal_tensor_assign_1_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_31_internal_tensor_assign_1_end_mask_0 = const()[name = tensor("new_cache_31_internal_tensor_assign_1_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_31_internal_tensor_assign_1_squeeze_mask_0 = const()[name = tensor("new_cache_31_internal_tensor_assign_1_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_76 = const()[name = tensor("shape_76"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_30 = const()[name = tensor("reduce_prod_30"), val = tensor(1048576)]; + tensor range_1d_30_start_0 = const()[name = tensor("range_1d_30_start_0"), val = tensor(0)]; + tensor range_1d_30_step_0 = const()[name = tensor("range_1d_30_step_0"), val = tensor(1)]; + tensor range_1d_30 = range_1d(end = reduce_prod_30, start = range_1d_30_start_0, step = range_1d_30_step_0)[name = tensor("range_1d_30")]; + tensor reshape_150 = reshape(shape = shape_76, x = range_1d_30)[name = tensor("reshape_150")]; + tensor slice_by_index_30 = slice_by_index(begin = concat_122, begin_mask = new_cache_31_internal_tensor_assign_1_begin_mask_0, end = concat_123, end_mask = new_cache_31_internal_tensor_assign_1_end_mask_0, squeeze_mask = new_cache_31_internal_tensor_assign_1_squeeze_mask_0, stride = new_cache_31_internal_tensor_assign_1_stride_0, x = reshape_150)[name = tensor("slice_by_index_30")]; + tensor reshape_151_shape_0 = const()[name = tensor("reshape_151_shape_0"), val = tensor([-1])]; + tensor reshape_151 = reshape(shape = reshape_151_shape_0, x = slice_by_index_30)[name = tensor("reshape_151")]; + tensor reshape_152_shape_0 = const()[name = tensor("reshape_152_shape_0"), val = tensor([-1])]; + tensor reshape_152_cast_fp16 = reshape(shape = reshape_152_shape_0, x = var_6607_cast_fp16)[name = tensor("reshape_152_cast_fp16")]; + tensor reshape_153_shape_0 = const()[name = tensor("reshape_153_shape_0"), val = tensor([-1])]; + tensor reshape_153_cast_fp16 = reshape(shape = reshape_153_shape_0, x = cache15_to_fp16)[name = tensor("reshape_153_cast_fp16")]; + tensor scatter_30_mode_0 = const()[name = tensor("scatter_30_mode_0"), val = tensor("update")]; + tensor scatter_30_axis_0 = const()[name = tensor("scatter_30_axis_0"), val = tensor(0)]; + tensor scatter_30_validate_indices_0 = const()[name = tensor("scatter_30_validate_indices_0"), val = tensor(false)]; + tensor scatter_30_cast_fp16 = scatter(axis = scatter_30_axis_0, data = reshape_153_cast_fp16, indices = reshape_151, mode = scatter_30_mode_0, updates = reshape_152_cast_fp16, validate_indices = scatter_30_validate_indices_0)[name = tensor("scatter_30_cast_fp16")]; + tensor reshape_154_cast_fp16 = reshape(shape = shape_76, x = scatter_30_cast_fp16)[name = tensor("reshape_154_cast_fp16")]; + tensor var_6615_begin_0 = const()[name = tensor("op_6615_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor var_6615_end_0 = const()[name = tensor("op_6615_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor var_6615_end_mask_0 = const()[name = tensor("op_6615_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_6615_squeeze_mask_0 = const()[name = tensor("op_6615_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor var_6615_cast_fp16 = slice_by_index(begin = var_6615_begin_0, end = var_6615_end_0, end_mask = var_6615_end_mask_0, squeeze_mask = var_6615_squeeze_mask_0, x = reshape_154_cast_fp16)[name = tensor("op_6615_cast_fp16")]; + tensor var_6617_axis_0 = const()[name = tensor("op_6617_axis_0"), val = tensor(1)]; + tensor var_6617_mode_0 = const()[name = tensor("op_6617_mode_0"), val = tensor("update")]; + tensor var_6617_validate_indices_0 = const()[name = tensor("op_6617_validate_indices_0"), val = tensor(false)]; + tensor var_6617_cast_fp16 = scatter_along_axis(axis = var_6617_axis_0, data = var_6615_cast_fp16, indices = write_indices_31, mode = var_6617_mode_0, updates = v_31_cast_fp16, validate_indices = var_6617_validate_indices_0)[name = tensor("op_6617_cast_fp16")]; + tensor concat_124 = const()[name = tensor("concat_124"), val = tensor([1, 0, 0, 0, 0])]; + tensor concat_125 = const()[name = tensor("concat_125"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_31_internal_tensor_assign_2_stride_0 = const()[name = tensor("new_cache_31_internal_tensor_assign_2_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_31_internal_tensor_assign_2_begin_mask_0 = const()[name = tensor("new_cache_31_internal_tensor_assign_2_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_31_internal_tensor_assign_2_end_mask_0 = const()[name = tensor("new_cache_31_internal_tensor_assign_2_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_31_internal_tensor_assign_2_squeeze_mask_0 = const()[name = tensor("new_cache_31_internal_tensor_assign_2_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_77 = const()[name = tensor("shape_77"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_31 = const()[name = tensor("reduce_prod_31"), val = tensor(1048576)]; + tensor range_1d_31_start_0 = const()[name = tensor("range_1d_31_start_0"), val = tensor(0)]; + tensor range_1d_31_step_0 = const()[name = tensor("range_1d_31_step_0"), val = tensor(1)]; + tensor range_1d_31 = range_1d(end = reduce_prod_31, start = range_1d_31_start_0, step = range_1d_31_step_0)[name = tensor("range_1d_31")]; + tensor reshape_155 = reshape(shape = shape_77, x = range_1d_31)[name = tensor("reshape_155")]; + tensor slice_by_index_31 = slice_by_index(begin = concat_124, begin_mask = new_cache_31_internal_tensor_assign_2_begin_mask_0, end = concat_125, end_mask = new_cache_31_internal_tensor_assign_2_end_mask_0, squeeze_mask = new_cache_31_internal_tensor_assign_2_squeeze_mask_0, stride = new_cache_31_internal_tensor_assign_2_stride_0, x = reshape_155)[name = tensor("slice_by_index_31")]; + tensor reshape_156_shape_0 = const()[name = tensor("reshape_156_shape_0"), val = tensor([-1])]; + tensor reshape_156 = reshape(shape = reshape_156_shape_0, x = slice_by_index_31)[name = tensor("reshape_156")]; + tensor reshape_157_shape_0 = const()[name = tensor("reshape_157_shape_0"), val = tensor([-1])]; + tensor reshape_157_cast_fp16 = reshape(shape = reshape_157_shape_0, x = var_6617_cast_fp16)[name = tensor("reshape_157_cast_fp16")]; + tensor reshape_158_shape_0 = const()[name = tensor("reshape_158_shape_0"), val = tensor([-1])]; + tensor reshape_158_cast_fp16 = reshape(shape = reshape_158_shape_0, x = reshape_154_cast_fp16)[name = tensor("reshape_158_cast_fp16")]; + tensor scatter_31_mode_0 = const()[name = tensor("scatter_31_mode_0"), val = tensor("update")]; + tensor scatter_31_axis_0 = const()[name = tensor("scatter_31_axis_0"), val = tensor(0)]; + tensor scatter_31_validate_indices_0 = const()[name = tensor("scatter_31_validate_indices_0"), val = tensor(false)]; + tensor scatter_31_cast_fp16 = scatter(axis = scatter_31_axis_0, data = reshape_158_cast_fp16, indices = reshape_156, mode = scatter_31_mode_0, updates = reshape_157_cast_fp16, validate_indices = scatter_31_validate_indices_0)[name = tensor("scatter_31_cast_fp16")]; + tensor reshape_159_cast_fp16 = reshape(shape = shape_77, x = scatter_31_cast_fp16)[name = tensor("reshape_159_cast_fp16")]; + tensor reshape_159_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("reshape_159_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor keys_91_begin_0 = const()[name = tensor("keys_91_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor keys_91_end_0 = const()[name = tensor("keys_91_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor keys_91_end_mask_0 = const()[name = tensor("keys_91_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor keys_91_squeeze_mask_0 = const()[name = tensor("keys_91_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor keys_91_cast_fp16 = slice_by_index(begin = keys_91_begin_0, end = keys_91_end_0, end_mask = keys_91_end_mask_0, squeeze_mask = keys_91_squeeze_mask_0, x = reshape_159_cast_fp16)[name = tensor("keys_91_cast_fp16")]; + tensor values_91_begin_0 = const()[name = tensor("values_91_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor values_91_end_0 = const()[name = tensor("values_91_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor values_91_end_mask_0 = const()[name = tensor("values_91_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor values_91_squeeze_mask_0 = const()[name = tensor("values_91_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor values_91_cast_fp16 = slice_by_index(begin = values_91_begin_0, end = values_91_end_0, end_mask = values_91_end_mask_0, squeeze_mask = values_91_squeeze_mask_0, x = reshape_159_cast_fp16)[name = tensor("values_91_cast_fp16")]; + tensor var_6629_cast_fp16 = not_equal(x = keys_91_cast_fp16, y = keys_91_cast_fp16)[name = tensor("op_6629_cast_fp16")]; + tensor keys_93_cast_fp16 = select(a = var_470_to_fp16, b = keys_91_cast_fp16, cond = var_6629_cast_fp16)[name = tensor("keys_93_cast_fp16")]; + tensor var_6637_cast_fp16 = not_equal(x = values_91_cast_fp16, y = values_91_cast_fp16)[name = tensor("op_6637_cast_fp16")]; + tensor values_93_cast_fp16 = select(a = var_470_to_fp16, b = values_91_cast_fp16, cond = var_6637_cast_fp16)[name = tensor("values_93_cast_fp16")]; + tensor var_6661 = const()[name = tensor("op_6661"), val = tensor([0, 2, 1, 3])]; + tensor var_6677 = const()[name = tensor("op_6677"), val = tensor([1, 1, 1])]; + tensor var_6678_cast_fp16 = reshape(shape = var_6677, x = position15_to_fp16)[name = tensor("op_6678_cast_fp16")]; + tensor q_positions_31_cast_fp16 = add(x = var_6678_cast_fp16, y = q_offsets_1_promoted_to_fp16)[name = tensor("q_positions_31_cast_fp16")]; + tensor valid_end_31_cast_fp16 = add(x = var_6678_cast_fp16, y = var_536_cast_fp16)[name = tensor("valid_end_31_cast_fp16")]; + tensor valid_mask_31_cast_fp16 = less(x = k_positions_1_promoted_to_fp16, y = valid_end_31_cast_fp16)[name = tensor("valid_mask_31_cast_fp16")]; + tensor causal_mask_31_cast_fp16 = less_equal(x = k_positions_1_promoted_to_fp16, y = q_positions_31_cast_fp16)[name = tensor("causal_mask_31_cast_fp16")]; + tensor var_6706 = logical_and(x = valid_mask_31_cast_fp16, y = causal_mask_31_cast_fp16)[name = tensor("op_6706")]; + tensor attn_mask_31_axes_0 = const()[name = tensor("attn_mask_31_axes_0"), val = tensor([1])]; + tensor attn_mask_31 = expand_dims(axes = attn_mask_31_axes_0, x = var_6706)[name = tensor("attn_mask_31")]; + tensor var_6720_transpose_x_0 = const()[name = tensor("op_6720_transpose_x_0"), val = tensor(false)]; + tensor var_6720_transpose_y_0 = const()[name = tensor("op_6720_transpose_y_0"), val = tensor(false)]; + tensor transpose_99_perm_0 = const()[name = tensor("transpose_99_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_100_perm_0 = const()[name = tensor("transpose_100_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_100 = transpose(perm = transpose_100_perm_0, x = keys_93_cast_fp16)[name = tensor("transpose_144")]; + tensor transpose_99 = transpose(perm = transpose_99_perm_0, x = q_93_cast_fp16)[name = tensor("transpose_145")]; + tensor var_6720_cast_fp16 = matmul(transpose_x = var_6720_transpose_x_0, transpose_y = var_6720_transpose_y_0, x = transpose_99, y = transpose_100)[name = tensor("op_6720_cast_fp16")]; + tensor var_6714_to_fp16 = const()[name = tensor("op_6714_to_fp16"), val = tensor([0x1p-3])]; + tensor attn_weights_91_cast_fp16 = mul(x = var_6720_cast_fp16, y = var_6714_to_fp16)[name = tensor("attn_weights_91_cast_fp16")]; + tensor var_6722 = logical_not(x = attn_mask_31)[name = tensor("op_6722")]; + tensor var_6723_to_fp16 = const()[name = tensor("op_6723_to_fp16"), val = tensor(-inf)]; + tensor attn_weights_93_cast_fp16 = select(a = var_6723_to_fp16, b = attn_weights_91_cast_fp16, cond = var_6722)[name = tensor("attn_weights_93_cast_fp16")]; + tensor var_6725 = const()[name = tensor("op_6725"), val = tensor(-1)]; + tensor attn_weights_95_cast_fp16 = softmax(axis = var_6725, x = attn_weights_93_cast_fp16)[name = tensor("attn_weights_95_cast_fp16")]; + tensor attn_output_31_transpose_x_0 = const()[name = tensor("attn_output_31_transpose_x_0"), val = tensor(false)]; + tensor attn_output_31_transpose_y_0 = const()[name = tensor("attn_output_31_transpose_y_0"), val = tensor(false)]; + tensor values_95_cast_fp16 = transpose(perm = var_6661, x = values_93_cast_fp16)[name = tensor("transpose_146")]; + tensor attn_output_31_cast_fp16 = matmul(transpose_x = attn_output_31_transpose_x_0, transpose_y = attn_output_31_transpose_y_0, x = attn_weights_95_cast_fp16, y = values_95_cast_fp16)[name = tensor("attn_output_31_cast_fp16")]; + tensor var_6733 = const()[name = tensor("op_6733"), val = tensor([0, 2, 1, 3])]; + tensor var_6738 = const()[name = tensor("op_6738"), val = tensor([1, 256, 1024])]; + tensor var_6734_cast_fp16 = transpose(perm = var_6733, x = attn_output_31_cast_fp16)[name = tensor("transpose_143")]; + tensor input_185_cast_fp16 = reshape(shape = var_6738, x = var_6734_cast_fp16)[name = tensor("input_185_cast_fp16")]; + tensor attn15_out_proj_weight_to_fp16 = const()[name = tensor("attn15_out_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(384983808)))]; + tensor linear_61_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = attn15_out_proj_weight_to_fp16, x = input_185_cast_fp16)[name = tensor("linear_61_cast_fp16")]; + tensor var_6747_cast_fp16 = add(x = position15_to_fp16, y = valid_len_to_fp16)[name = tensor("op_6747_cast_fp16")]; + tensor var_6747_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_6747_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor input_187_cast_fp16 = add(x = input_181_cast_fp16, y = linear_61_cast_fp16)[name = tensor("input_187_cast_fp16")]; + tensor input_189_axes_0 = const()[name = tensor("input_189_axes_0"), val = tensor([-1])]; + tensor norm15_2_weight_to_fp16 = const()[name = tensor("norm15_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(387081024)))]; + tensor norm15_2_bias_to_fp16 = const()[name = tensor("norm15_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(387083136)))]; + tensor var_6751_to_fp16 = const()[name = tensor("op_6751_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_189_cast_fp16 = layer_norm(axes = input_189_axes_0, beta = norm15_2_bias_to_fp16, epsilon = var_6751_to_fp16, gamma = norm15_2_weight_to_fp16, x = input_187_cast_fp16)[name = tensor("input_189_cast_fp16")]; + tensor linear15_1_weight_to_fp16 = const()[name = tensor("linear15_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(387085248)))]; + tensor linear_62_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = linear15_1_weight_to_fp16, x = input_189_cast_fp16)[name = tensor("linear_62_cast_fp16")]; + tensor input_191_mode_0 = const()[name = tensor("input_191_mode_0"), val = tensor("EXACT")]; + tensor input_191_cast_fp16 = gelu(mode = input_191_mode_0, x = linear_62_cast_fp16)[name = tensor("input_191_cast_fp16")]; + tensor linear15_2_weight_to_fp16 = const()[name = tensor("linear15_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(395473920)))]; + tensor linear_63_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = linear15_2_weight_to_fp16, x = input_191_cast_fp16)[name = tensor("linear_63_cast_fp16")]; + tensor input_193_cast_fp16 = add(x = input_187_cast_fp16, y = linear_63_cast_fp16)[name = tensor("input_193_cast_fp16")]; + tensor input_195_axes_0 = const()[name = tensor("input_195_axes_0"), val = tensor([-1])]; + tensor norm16_1_weight_to_fp16 = const()[name = tensor("norm16_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(403862592)))]; + tensor norm16_1_bias_to_fp16 = const()[name = tensor("norm16_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(403864704)))]; + tensor var_6768_to_fp16 = const()[name = tensor("op_6768_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_195_cast_fp16 = layer_norm(axes = input_195_axes_0, beta = norm16_1_bias_to_fp16, epsilon = var_6768_to_fp16, gamma = norm16_1_weight_to_fp16, x = input_193_cast_fp16)[name = tensor("input_195_cast_fp16")]; + tensor attn16_in_proj_weight_to_fp16 = const()[name = tensor("attn16_in_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(403866816)))]; + tensor linear_64_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = attn16_in_proj_weight_to_fp16, x = input_195_cast_fp16)[name = tensor("linear_64_cast_fp16")]; + tensor var_6782 = const()[name = tensor("op_6782"), val = tensor([1, 256, 3, 16, 64])]; + tensor qkv_33_cast_fp16 = reshape(shape = var_6782, x = linear_64_cast_fp16)[name = tensor("qkv_33_cast_fp16")]; + tensor q_97_begin_0 = const()[name = tensor("q_97_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor q_97_end_0 = const()[name = tensor("q_97_end_0"), val = tensor([1, 256, 1, 16, 64])]; + tensor q_97_end_mask_0 = const()[name = tensor("q_97_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor q_97_squeeze_mask_0 = const()[name = tensor("q_97_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor q_97_cast_fp16 = slice_by_index(begin = q_97_begin_0, end = q_97_end_0, end_mask = q_97_end_mask_0, squeeze_mask = q_97_squeeze_mask_0, x = qkv_33_cast_fp16)[name = tensor("q_97_cast_fp16")]; + tensor k_65_begin_0 = const()[name = tensor("k_65_begin_0"), val = tensor([0, 0, 1, 0, 0])]; + tensor k_65_end_0 = const()[name = tensor("k_65_end_0"), val = tensor([1, 256, 2, 16, 64])]; + tensor k_65_end_mask_0 = const()[name = tensor("k_65_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor k_65_squeeze_mask_0 = const()[name = tensor("k_65_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor k_65_cast_fp16 = slice_by_index(begin = k_65_begin_0, end = k_65_end_0, end_mask = k_65_end_mask_0, squeeze_mask = k_65_squeeze_mask_0, x = qkv_33_cast_fp16)[name = tensor("k_65_cast_fp16")]; + tensor v_33_begin_0 = const()[name = tensor("v_33_begin_0"), val = tensor([0, 0, 2, 0, 0])]; + tensor v_33_end_0 = const()[name = tensor("v_33_end_0"), val = tensor([1, 256, 3, 16, 64])]; + tensor v_33_end_mask_0 = const()[name = tensor("v_33_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor v_33_squeeze_mask_0 = const()[name = tensor("v_33_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor v_33_cast_fp16 = slice_by_index(begin = v_33_begin_0, end = v_33_end_0, end_mask = v_33_end_mask_0, squeeze_mask = v_33_squeeze_mask_0, x = qkv_33_cast_fp16)[name = tensor("v_33_cast_fp16")]; + tensor var_6840 = const()[name = tensor("op_6840"), val = tensor([1, 1])]; + tensor position16_to_fp16_dtype_0 = const()[name = tensor("position16_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor position16_to_fp16 = cast(dtype = position16_to_fp16_dtype_0, x = position16)[name = tensor("cast_581")]; + tensor var_6841_cast_fp16 = reshape(shape = var_6840, x = position16_to_fp16)[name = tensor("op_6841_cast_fp16")]; + tensor ts_99_cast_fp16 = add(x = ts_1_promoted_to_fp16, y = var_6841_cast_fp16)[name = tensor("ts_99_cast_fp16")]; + tensor var_6848 = const()[name = tensor("op_6848"), val = tensor([1, 256, 1, 1])]; + tensor ts_101_cast_fp16 = reshape(shape = var_6848, x = ts_99_cast_fp16)[name = tensor("ts_101_cast_fp16")]; + tensor var_6855 = const()[name = tensor("op_6855"), val = tensor([1, 256, 16, 32, 2])]; + tensor q_complex_33_cast_fp16 = reshape(shape = var_6855, x = q_97_cast_fp16)[name = tensor("q_complex_33_cast_fp16")]; + tensor var_6862 = const()[name = tensor("op_6862"), val = tensor([1, 256, 16, 32, 2])]; + tensor k_complex_33_cast_fp16 = reshape(shape = var_6862, x = k_65_cast_fp16)[name = tensor("k_complex_33_cast_fp16")]; + tensor var_6866_begin_0 = const()[name = tensor("op_6866_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_6866_end_0 = const()[name = tensor("op_6866_end_0"), val = tensor([1, 256, 16, 32, 1])]; + tensor var_6866_end_mask_0 = const()[name = tensor("op_6866_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_6866_squeeze_mask_0 = const()[name = tensor("op_6866_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_6866_cast_fp16 = slice_by_index(begin = var_6866_begin_0, end = var_6866_end_0, end_mask = var_6866_end_mask_0, squeeze_mask = var_6866_squeeze_mask_0, x = q_complex_33_cast_fp16)[name = tensor("op_6866_cast_fp16")]; + tensor var_6874_begin_0 = const()[name = tensor("op_6874_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_6874_end_0 = const()[name = tensor("op_6874_end_0"), val = tensor([1, 256, 16, 32, 2])]; + tensor var_6874_end_mask_0 = const()[name = tensor("op_6874_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_6874_squeeze_mask_0 = const()[name = tensor("op_6874_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_6874_cast_fp16 = slice_by_index(begin = var_6874_begin_0, end = var_6874_end_0, end_mask = var_6874_end_mask_0, squeeze_mask = var_6874_squeeze_mask_0, x = q_complex_33_cast_fp16)[name = tensor("op_6874_cast_fp16")]; + tensor var_6882_begin_0 = const()[name = tensor("op_6882_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_6882_end_0 = const()[name = tensor("op_6882_end_0"), val = tensor([1, 256, 16, 32, 1])]; + tensor var_6882_end_mask_0 = const()[name = tensor("op_6882_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_6882_squeeze_mask_0 = const()[name = tensor("op_6882_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_6882_cast_fp16 = slice_by_index(begin = var_6882_begin_0, end = var_6882_end_0, end_mask = var_6882_end_mask_0, squeeze_mask = var_6882_squeeze_mask_0, x = k_complex_33_cast_fp16)[name = tensor("op_6882_cast_fp16")]; + tensor var_6890_begin_0 = const()[name = tensor("op_6890_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_6890_end_0 = const()[name = tensor("op_6890_end_0"), val = tensor([1, 256, 16, 32, 2])]; + tensor var_6890_end_mask_0 = const()[name = tensor("op_6890_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_6890_squeeze_mask_0 = const()[name = tensor("op_6890_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_6890_cast_fp16 = slice_by_index(begin = var_6890_begin_0, end = var_6890_end_0, end_mask = var_6890_end_mask_0, squeeze_mask = var_6890_squeeze_mask_0, x = k_complex_33_cast_fp16)[name = tensor("op_6890_cast_fp16")]; + tensor freqs_33_to_fp16 = const()[name = tensor("freqs_33_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(410158336)))]; + tensor var_6896_cast_fp16 = mul(x = freqs_33_to_fp16, y = ts_101_cast_fp16)[name = tensor("op_6896_cast_fp16")]; + tensor rotr_33_cast_fp16 = cos(x = var_6896_cast_fp16)[name = tensor("rotr_33_cast_fp16")]; + tensor roti_33_cast_fp16 = sin(x = var_6896_cast_fp16)[name = tensor("roti_33_cast_fp16")]; + tensor var_6900_cast_fp16 = mul(x = var_6866_cast_fp16, y = rotr_33_cast_fp16)[name = tensor("op_6900_cast_fp16")]; + tensor var_6901_cast_fp16 = mul(x = var_6874_cast_fp16, y = roti_33_cast_fp16)[name = tensor("op_6901_cast_fp16")]; + tensor qor_65_cast_fp16 = sub(x = var_6900_cast_fp16, y = var_6901_cast_fp16)[name = tensor("qor_65_cast_fp16")]; + tensor var_6904_cast_fp16 = mul(x = var_6866_cast_fp16, y = roti_33_cast_fp16)[name = tensor("op_6904_cast_fp16")]; + tensor var_6905_cast_fp16 = mul(x = var_6874_cast_fp16, y = rotr_33_cast_fp16)[name = tensor("op_6905_cast_fp16")]; + tensor qoi_65_cast_fp16 = add(x = var_6904_cast_fp16, y = var_6905_cast_fp16)[name = tensor("qoi_65_cast_fp16")]; + tensor var_6908_cast_fp16 = mul(x = var_6882_cast_fp16, y = rotr_33_cast_fp16)[name = tensor("op_6908_cast_fp16")]; + tensor var_6909_cast_fp16 = mul(x = var_6890_cast_fp16, y = roti_33_cast_fp16)[name = tensor("op_6909_cast_fp16")]; + tensor kor_65_cast_fp16 = sub(x = var_6908_cast_fp16, y = var_6909_cast_fp16)[name = tensor("kor_65_cast_fp16")]; + tensor var_6912_cast_fp16 = mul(x = var_6882_cast_fp16, y = roti_33_cast_fp16)[name = tensor("op_6912_cast_fp16")]; + tensor var_6913_cast_fp16 = mul(x = var_6890_cast_fp16, y = rotr_33_cast_fp16)[name = tensor("op_6913_cast_fp16")]; + tensor koi_65_cast_fp16 = add(x = var_6912_cast_fp16, y = var_6913_cast_fp16)[name = tensor("koi_65_cast_fp16")]; + tensor qo_33_axis_0 = const()[name = tensor("qo_33_axis_0"), val = tensor(-1)]; + tensor qo_33_cast_fp16 = stack(axis = qo_33_axis_0, values = (qor_65_cast_fp16, qoi_65_cast_fp16))[name = tensor("qo_33_cast_fp16")]; + tensor ko_33_axis_0 = const()[name = tensor("ko_33_axis_0"), val = tensor(-1)]; + tensor ko_33_cast_fp16 = stack(axis = ko_33_axis_0, values = (kor_65_cast_fp16, koi_65_cast_fp16))[name = tensor("ko_33_cast_fp16")]; + tensor var_6946 = const()[name = tensor("op_6946"), val = tensor([1, 256, 16, 64])]; + tensor q_99_cast_fp16 = reshape(shape = var_6946, x = qo_33_cast_fp16)[name = tensor("q_99_cast_fp16")]; + tensor var_6952 = const()[name = tensor("op_6952"), val = tensor([1, 256, 16, 64])]; + tensor k_67_cast_fp16 = reshape(shape = var_6952, x = ko_33_cast_fp16)[name = tensor("k_67_cast_fp16")]; + tensor write_indices_float_97_cast_fp16 = add(x = var_6841_cast_fp16, y = var_393_promoted_to_fp16)[name = tensor("write_indices_float_97_cast_fp16")]; + tensor _inversed_6973_y_0_to_fp16 = const()[name = tensor("_inversed_6973_y_0_to_fp16"), val = tensor(0x1p-9)]; + tensor _inversed_6973_cast_fp16 = mul(x = write_indices_float_97_cast_fp16, y = _inversed_6973_y_0_to_fp16)[name = tensor("_inversed_6973_cast_fp16")]; + tensor var_6974_cast_fp16 = floor(x = _inversed_6973_cast_fp16)[name = tensor("op_6974_cast_fp16")]; + tensor var_6975_to_fp16 = const()[name = tensor("op_6975_to_fp16"), val = tensor(0x1p+9)]; + tensor var_6976_cast_fp16 = mul(x = var_6974_cast_fp16, y = var_6975_to_fp16)[name = tensor("op_6976_cast_fp16")]; + tensor write_indices_float_99_cast_fp16 = sub(x = write_indices_float_97_cast_fp16, y = var_6976_cast_fp16)[name = tensor("write_indices_float_99_cast_fp16")]; + tensor var_6994_value_0_to_fp16 = const()[name = tensor("op_6994_value_0_to_fp16"), val = tensor(0x1.ffp+8)]; + tensor var_6994_cast_fp16 = fill_like(ref_tensor = write_indices_float_99_cast_fp16, value = var_6994_value_0_to_fp16)[name = tensor("op_6994_cast_fp16")]; + tensor write_indices_float_101_cast_fp16 = select(a = write_indices_float_99_cast_fp16, b = var_6994_cast_fp16, cond = token_valid_1_cast_fp16)[name = tensor("write_indices_float_101_cast_fp16")]; + tensor var_7000_dtype_0 = const()[name = tensor("op_7000_dtype_0"), val = tensor("int32")]; + tensor var_7005 = const()[name = tensor("op_7005"), val = tensor([1, 256, 1, 1])]; + tensor write_indices_float_101_cast_fp16_to_int32 = cast(dtype = var_7000_dtype_0, x = write_indices_float_101_cast_fp16)[name = tensor("cast_580")]; + tensor var_7006 = reshape(shape = var_7005, x = write_indices_float_101_cast_fp16_to_int32)[name = tensor("op_7006")]; + tensor write_indices_33_reps_0 = const()[name = tensor("write_indices_33_reps_0"), val = tensor([1, 1, 16, 64])]; + tensor write_indices_33 = tile(reps = write_indices_33_reps_0, x = var_7006)[name = tensor("write_indices_33")]; + tensor var_7016_begin_0 = const()[name = tensor("op_7016_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_7016_end_0 = const()[name = tensor("op_7016_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor var_7016_end_mask_0 = const()[name = tensor("op_7016_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_7016_squeeze_mask_0 = const()[name = tensor("op_7016_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor cache16_to_fp16_dtype_0 = const()[name = tensor("cache16_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor cache16_to_fp16 = cast(dtype = cache16_to_fp16_dtype_0, x = cache16)[name = tensor("cast_579")]; + tensor var_7016_cast_fp16 = slice_by_index(begin = var_7016_begin_0, end = var_7016_end_0, end_mask = var_7016_end_mask_0, squeeze_mask = var_7016_squeeze_mask_0, x = cache16_to_fp16)[name = tensor("op_7016_cast_fp16")]; + tensor var_7018_axis_0 = const()[name = tensor("op_7018_axis_0"), val = tensor(1)]; + tensor var_7018_mode_0 = const()[name = tensor("op_7018_mode_0"), val = tensor("update")]; + tensor var_7018_validate_indices_0 = const()[name = tensor("op_7018_validate_indices_0"), val = tensor(false)]; + tensor var_7018_cast_fp16 = scatter_along_axis(axis = var_7018_axis_0, data = var_7016_cast_fp16, indices = write_indices_33, mode = var_7018_mode_0, updates = k_67_cast_fp16, validate_indices = var_7018_validate_indices_0)[name = tensor("op_7018_cast_fp16")]; + tensor concat_130 = const()[name = tensor("concat_130"), val = tensor([0, 0, 0, 0, 0])]; + tensor concat_131 = const()[name = tensor("concat_131"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_33_internal_tensor_assign_1_stride_0 = const()[name = tensor("new_cache_33_internal_tensor_assign_1_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_33_internal_tensor_assign_1_begin_mask_0 = const()[name = tensor("new_cache_33_internal_tensor_assign_1_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_33_internal_tensor_assign_1_end_mask_0 = const()[name = tensor("new_cache_33_internal_tensor_assign_1_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_33_internal_tensor_assign_1_squeeze_mask_0 = const()[name = tensor("new_cache_33_internal_tensor_assign_1_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_78 = const()[name = tensor("shape_78"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_32 = const()[name = tensor("reduce_prod_32"), val = tensor(1048576)]; + tensor range_1d_32_start_0 = const()[name = tensor("range_1d_32_start_0"), val = tensor(0)]; + tensor range_1d_32_step_0 = const()[name = tensor("range_1d_32_step_0"), val = tensor(1)]; + tensor range_1d_32 = range_1d(end = reduce_prod_32, start = range_1d_32_start_0, step = range_1d_32_step_0)[name = tensor("range_1d_32")]; + tensor reshape_160 = reshape(shape = shape_78, x = range_1d_32)[name = tensor("reshape_160")]; + tensor slice_by_index_32 = slice_by_index(begin = concat_130, begin_mask = new_cache_33_internal_tensor_assign_1_begin_mask_0, end = concat_131, end_mask = new_cache_33_internal_tensor_assign_1_end_mask_0, squeeze_mask = new_cache_33_internal_tensor_assign_1_squeeze_mask_0, stride = new_cache_33_internal_tensor_assign_1_stride_0, x = reshape_160)[name = tensor("slice_by_index_32")]; + tensor reshape_161_shape_0 = const()[name = tensor("reshape_161_shape_0"), val = tensor([-1])]; + tensor reshape_161 = reshape(shape = reshape_161_shape_0, x = slice_by_index_32)[name = tensor("reshape_161")]; + tensor reshape_162_shape_0 = const()[name = tensor("reshape_162_shape_0"), val = tensor([-1])]; + tensor reshape_162_cast_fp16 = reshape(shape = reshape_162_shape_0, x = var_7018_cast_fp16)[name = tensor("reshape_162_cast_fp16")]; + tensor reshape_163_shape_0 = const()[name = tensor("reshape_163_shape_0"), val = tensor([-1])]; + tensor reshape_163_cast_fp16 = reshape(shape = reshape_163_shape_0, x = cache16_to_fp16)[name = tensor("reshape_163_cast_fp16")]; + tensor scatter_32_mode_0 = const()[name = tensor("scatter_32_mode_0"), val = tensor("update")]; + tensor scatter_32_axis_0 = const()[name = tensor("scatter_32_axis_0"), val = tensor(0)]; + tensor scatter_32_validate_indices_0 = const()[name = tensor("scatter_32_validate_indices_0"), val = tensor(false)]; + tensor scatter_32_cast_fp16 = scatter(axis = scatter_32_axis_0, data = reshape_163_cast_fp16, indices = reshape_161, mode = scatter_32_mode_0, updates = reshape_162_cast_fp16, validate_indices = scatter_32_validate_indices_0)[name = tensor("scatter_32_cast_fp16")]; + tensor reshape_164_cast_fp16 = reshape(shape = shape_78, x = scatter_32_cast_fp16)[name = tensor("reshape_164_cast_fp16")]; + tensor var_7026_begin_0 = const()[name = tensor("op_7026_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor var_7026_end_0 = const()[name = tensor("op_7026_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor var_7026_end_mask_0 = const()[name = tensor("op_7026_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_7026_squeeze_mask_0 = const()[name = tensor("op_7026_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor var_7026_cast_fp16 = slice_by_index(begin = var_7026_begin_0, end = var_7026_end_0, end_mask = var_7026_end_mask_0, squeeze_mask = var_7026_squeeze_mask_0, x = reshape_164_cast_fp16)[name = tensor("op_7026_cast_fp16")]; + tensor var_7028_axis_0 = const()[name = tensor("op_7028_axis_0"), val = tensor(1)]; + tensor var_7028_mode_0 = const()[name = tensor("op_7028_mode_0"), val = tensor("update")]; + tensor var_7028_validate_indices_0 = const()[name = tensor("op_7028_validate_indices_0"), val = tensor(false)]; + tensor var_7028_cast_fp16 = scatter_along_axis(axis = var_7028_axis_0, data = var_7026_cast_fp16, indices = write_indices_33, mode = var_7028_mode_0, updates = v_33_cast_fp16, validate_indices = var_7028_validate_indices_0)[name = tensor("op_7028_cast_fp16")]; + tensor concat_132 = const()[name = tensor("concat_132"), val = tensor([1, 0, 0, 0, 0])]; + tensor concat_133 = const()[name = tensor("concat_133"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_33_internal_tensor_assign_2_stride_0 = const()[name = tensor("new_cache_33_internal_tensor_assign_2_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_33_internal_tensor_assign_2_begin_mask_0 = const()[name = tensor("new_cache_33_internal_tensor_assign_2_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_33_internal_tensor_assign_2_end_mask_0 = const()[name = tensor("new_cache_33_internal_tensor_assign_2_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_33_internal_tensor_assign_2_squeeze_mask_0 = const()[name = tensor("new_cache_33_internal_tensor_assign_2_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_79 = const()[name = tensor("shape_79"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_33 = const()[name = tensor("reduce_prod_33"), val = tensor(1048576)]; + tensor range_1d_33_start_0 = const()[name = tensor("range_1d_33_start_0"), val = tensor(0)]; + tensor range_1d_33_step_0 = const()[name = tensor("range_1d_33_step_0"), val = tensor(1)]; + tensor range_1d_33 = range_1d(end = reduce_prod_33, start = range_1d_33_start_0, step = range_1d_33_step_0)[name = tensor("range_1d_33")]; + tensor reshape_165 = reshape(shape = shape_79, x = range_1d_33)[name = tensor("reshape_165")]; + tensor slice_by_index_33 = slice_by_index(begin = concat_132, begin_mask = new_cache_33_internal_tensor_assign_2_begin_mask_0, end = concat_133, end_mask = new_cache_33_internal_tensor_assign_2_end_mask_0, squeeze_mask = new_cache_33_internal_tensor_assign_2_squeeze_mask_0, stride = new_cache_33_internal_tensor_assign_2_stride_0, x = reshape_165)[name = tensor("slice_by_index_33")]; + tensor reshape_166_shape_0 = const()[name = tensor("reshape_166_shape_0"), val = tensor([-1])]; + tensor reshape_166 = reshape(shape = reshape_166_shape_0, x = slice_by_index_33)[name = tensor("reshape_166")]; + tensor reshape_167_shape_0 = const()[name = tensor("reshape_167_shape_0"), val = tensor([-1])]; + tensor reshape_167_cast_fp16 = reshape(shape = reshape_167_shape_0, x = var_7028_cast_fp16)[name = tensor("reshape_167_cast_fp16")]; + tensor reshape_168_shape_0 = const()[name = tensor("reshape_168_shape_0"), val = tensor([-1])]; + tensor reshape_168_cast_fp16 = reshape(shape = reshape_168_shape_0, x = reshape_164_cast_fp16)[name = tensor("reshape_168_cast_fp16")]; + tensor scatter_33_mode_0 = const()[name = tensor("scatter_33_mode_0"), val = tensor("update")]; + tensor scatter_33_axis_0 = const()[name = tensor("scatter_33_axis_0"), val = tensor(0)]; + tensor scatter_33_validate_indices_0 = const()[name = tensor("scatter_33_validate_indices_0"), val = tensor(false)]; + tensor scatter_33_cast_fp16 = scatter(axis = scatter_33_axis_0, data = reshape_168_cast_fp16, indices = reshape_166, mode = scatter_33_mode_0, updates = reshape_167_cast_fp16, validate_indices = scatter_33_validate_indices_0)[name = tensor("scatter_33_cast_fp16")]; + tensor reshape_169_cast_fp16 = reshape(shape = shape_79, x = scatter_33_cast_fp16)[name = tensor("reshape_169_cast_fp16")]; + tensor reshape_169_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("reshape_169_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor keys_97_begin_0 = const()[name = tensor("keys_97_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor keys_97_end_0 = const()[name = tensor("keys_97_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor keys_97_end_mask_0 = const()[name = tensor("keys_97_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor keys_97_squeeze_mask_0 = const()[name = tensor("keys_97_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor keys_97_cast_fp16 = slice_by_index(begin = keys_97_begin_0, end = keys_97_end_0, end_mask = keys_97_end_mask_0, squeeze_mask = keys_97_squeeze_mask_0, x = reshape_169_cast_fp16)[name = tensor("keys_97_cast_fp16")]; + tensor values_97_begin_0 = const()[name = tensor("values_97_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor values_97_end_0 = const()[name = tensor("values_97_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor values_97_end_mask_0 = const()[name = tensor("values_97_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor values_97_squeeze_mask_0 = const()[name = tensor("values_97_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor values_97_cast_fp16 = slice_by_index(begin = values_97_begin_0, end = values_97_end_0, end_mask = values_97_end_mask_0, squeeze_mask = values_97_squeeze_mask_0, x = reshape_169_cast_fp16)[name = tensor("values_97_cast_fp16")]; + tensor var_7040_cast_fp16 = not_equal(x = keys_97_cast_fp16, y = keys_97_cast_fp16)[name = tensor("op_7040_cast_fp16")]; + tensor keys_99_cast_fp16 = select(a = var_470_to_fp16, b = keys_97_cast_fp16, cond = var_7040_cast_fp16)[name = tensor("keys_99_cast_fp16")]; + tensor var_7048_cast_fp16 = not_equal(x = values_97_cast_fp16, y = values_97_cast_fp16)[name = tensor("op_7048_cast_fp16")]; + tensor values_99_cast_fp16 = select(a = var_470_to_fp16, b = values_97_cast_fp16, cond = var_7048_cast_fp16)[name = tensor("values_99_cast_fp16")]; + tensor var_7072 = const()[name = tensor("op_7072"), val = tensor([0, 2, 1, 3])]; + tensor var_7088 = const()[name = tensor("op_7088"), val = tensor([1, 1, 1])]; + tensor var_7089_cast_fp16 = reshape(shape = var_7088, x = position16_to_fp16)[name = tensor("op_7089_cast_fp16")]; + tensor q_positions_33_cast_fp16 = add(x = var_7089_cast_fp16, y = q_offsets_1_promoted_to_fp16)[name = tensor("q_positions_33_cast_fp16")]; + tensor valid_end_33_cast_fp16 = add(x = var_7089_cast_fp16, y = var_536_cast_fp16)[name = tensor("valid_end_33_cast_fp16")]; + tensor valid_mask_33_cast_fp16 = less(x = k_positions_1_promoted_to_fp16, y = valid_end_33_cast_fp16)[name = tensor("valid_mask_33_cast_fp16")]; + tensor causal_mask_33_cast_fp16 = less_equal(x = k_positions_1_promoted_to_fp16, y = q_positions_33_cast_fp16)[name = tensor("causal_mask_33_cast_fp16")]; + tensor var_7117 = logical_and(x = valid_mask_33_cast_fp16, y = causal_mask_33_cast_fp16)[name = tensor("op_7117")]; + tensor attn_mask_33_axes_0 = const()[name = tensor("attn_mask_33_axes_0"), val = tensor([1])]; + tensor attn_mask_33 = expand_dims(axes = attn_mask_33_axes_0, x = var_7117)[name = tensor("attn_mask_33")]; + tensor var_7131_transpose_x_0 = const()[name = tensor("op_7131_transpose_x_0"), val = tensor(false)]; + tensor var_7131_transpose_y_0 = const()[name = tensor("op_7131_transpose_y_0"), val = tensor(false)]; + tensor transpose_101_perm_0 = const()[name = tensor("transpose_101_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_102_perm_0 = const()[name = tensor("transpose_102_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_102 = transpose(perm = transpose_102_perm_0, x = keys_99_cast_fp16)[name = tensor("transpose_140")]; + tensor transpose_101 = transpose(perm = transpose_101_perm_0, x = q_99_cast_fp16)[name = tensor("transpose_141")]; + tensor var_7131_cast_fp16 = matmul(transpose_x = var_7131_transpose_x_0, transpose_y = var_7131_transpose_y_0, x = transpose_101, y = transpose_102)[name = tensor("op_7131_cast_fp16")]; + tensor var_7125_to_fp16 = const()[name = tensor("op_7125_to_fp16"), val = tensor([0x1p-3])]; + tensor attn_weights_97_cast_fp16 = mul(x = var_7131_cast_fp16, y = var_7125_to_fp16)[name = tensor("attn_weights_97_cast_fp16")]; + tensor var_7133 = logical_not(x = attn_mask_33)[name = tensor("op_7133")]; + tensor var_7134_to_fp16 = const()[name = tensor("op_7134_to_fp16"), val = tensor(-inf)]; + tensor attn_weights_99_cast_fp16 = select(a = var_7134_to_fp16, b = attn_weights_97_cast_fp16, cond = var_7133)[name = tensor("attn_weights_99_cast_fp16")]; + tensor var_7136 = const()[name = tensor("op_7136"), val = tensor(-1)]; + tensor attn_weights_101_cast_fp16 = softmax(axis = var_7136, x = attn_weights_99_cast_fp16)[name = tensor("attn_weights_101_cast_fp16")]; + tensor attn_output_33_transpose_x_0 = const()[name = tensor("attn_output_33_transpose_x_0"), val = tensor(false)]; + tensor attn_output_33_transpose_y_0 = const()[name = tensor("attn_output_33_transpose_y_0"), val = tensor(false)]; + tensor values_101_cast_fp16 = transpose(perm = var_7072, x = values_99_cast_fp16)[name = tensor("transpose_142")]; + tensor attn_output_33_cast_fp16 = matmul(transpose_x = attn_output_33_transpose_x_0, transpose_y = attn_output_33_transpose_y_0, x = attn_weights_101_cast_fp16, y = values_101_cast_fp16)[name = tensor("attn_output_33_cast_fp16")]; + tensor var_7144 = const()[name = tensor("op_7144"), val = tensor([0, 2, 1, 3])]; + tensor var_7149 = const()[name = tensor("op_7149"), val = tensor([1, 256, 1024])]; + tensor var_7145_cast_fp16 = transpose(perm = var_7144, x = attn_output_33_cast_fp16)[name = tensor("transpose_139")]; + tensor input_197_cast_fp16 = reshape(shape = var_7149, x = var_7145_cast_fp16)[name = tensor("input_197_cast_fp16")]; + tensor attn16_out_proj_weight_to_fp16 = const()[name = tensor("attn16_out_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(410158464)))]; + tensor linear_65_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = attn16_out_proj_weight_to_fp16, x = input_197_cast_fp16)[name = tensor("linear_65_cast_fp16")]; + tensor var_7158_cast_fp16 = add(x = position16_to_fp16, y = valid_len_to_fp16)[name = tensor("op_7158_cast_fp16")]; + tensor var_7158_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_7158_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor input_199_cast_fp16 = add(x = input_193_cast_fp16, y = linear_65_cast_fp16)[name = tensor("input_199_cast_fp16")]; + tensor input_201_axes_0 = const()[name = tensor("input_201_axes_0"), val = tensor([-1])]; + tensor norm16_2_weight_to_fp16 = const()[name = tensor("norm16_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(412255680)))]; + tensor norm16_2_bias_to_fp16 = const()[name = tensor("norm16_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(412257792)))]; + tensor var_7162_to_fp16 = const()[name = tensor("op_7162_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_201_cast_fp16 = layer_norm(axes = input_201_axes_0, beta = norm16_2_bias_to_fp16, epsilon = var_7162_to_fp16, gamma = norm16_2_weight_to_fp16, x = input_199_cast_fp16)[name = tensor("input_201_cast_fp16")]; + tensor linear16_1_weight_to_fp16 = const()[name = tensor("linear16_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(412259904)))]; + tensor linear_66_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = linear16_1_weight_to_fp16, x = input_201_cast_fp16)[name = tensor("linear_66_cast_fp16")]; + tensor input_203_mode_0 = const()[name = tensor("input_203_mode_0"), val = tensor("EXACT")]; + tensor input_203_cast_fp16 = gelu(mode = input_203_mode_0, x = linear_66_cast_fp16)[name = tensor("input_203_cast_fp16")]; + tensor linear16_2_weight_to_fp16 = const()[name = tensor("linear16_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(420648576)))]; + tensor linear_67_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = linear16_2_weight_to_fp16, x = input_203_cast_fp16)[name = tensor("linear_67_cast_fp16")]; + tensor input_205_cast_fp16 = add(x = input_199_cast_fp16, y = linear_67_cast_fp16)[name = tensor("input_205_cast_fp16")]; + tensor input_207_axes_0 = const()[name = tensor("input_207_axes_0"), val = tensor([-1])]; + tensor norm17_1_weight_to_fp16 = const()[name = tensor("norm17_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(429037248)))]; + tensor norm17_1_bias_to_fp16 = const()[name = tensor("norm17_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(429039360)))]; + tensor var_7179_to_fp16 = const()[name = tensor("op_7179_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_207_cast_fp16 = layer_norm(axes = input_207_axes_0, beta = norm17_1_bias_to_fp16, epsilon = var_7179_to_fp16, gamma = norm17_1_weight_to_fp16, x = input_205_cast_fp16)[name = tensor("input_207_cast_fp16")]; + tensor attn17_in_proj_weight_to_fp16 = const()[name = tensor("attn17_in_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(429041472)))]; + tensor linear_68_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = attn17_in_proj_weight_to_fp16, x = input_207_cast_fp16)[name = tensor("linear_68_cast_fp16")]; + tensor var_7193 = const()[name = tensor("op_7193"), val = tensor([1, 256, 3, 16, 64])]; + tensor qkv_35_cast_fp16 = reshape(shape = var_7193, x = linear_68_cast_fp16)[name = tensor("qkv_35_cast_fp16")]; + tensor q_103_begin_0 = const()[name = tensor("q_103_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor q_103_end_0 = const()[name = tensor("q_103_end_0"), val = tensor([1, 256, 1, 16, 64])]; + tensor q_103_end_mask_0 = const()[name = tensor("q_103_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor q_103_squeeze_mask_0 = const()[name = tensor("q_103_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor q_103_cast_fp16 = slice_by_index(begin = q_103_begin_0, end = q_103_end_0, end_mask = q_103_end_mask_0, squeeze_mask = q_103_squeeze_mask_0, x = qkv_35_cast_fp16)[name = tensor("q_103_cast_fp16")]; + tensor k_69_begin_0 = const()[name = tensor("k_69_begin_0"), val = tensor([0, 0, 1, 0, 0])]; + tensor k_69_end_0 = const()[name = tensor("k_69_end_0"), val = tensor([1, 256, 2, 16, 64])]; + tensor k_69_end_mask_0 = const()[name = tensor("k_69_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor k_69_squeeze_mask_0 = const()[name = tensor("k_69_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor k_69_cast_fp16 = slice_by_index(begin = k_69_begin_0, end = k_69_end_0, end_mask = k_69_end_mask_0, squeeze_mask = k_69_squeeze_mask_0, x = qkv_35_cast_fp16)[name = tensor("k_69_cast_fp16")]; + tensor v_35_begin_0 = const()[name = tensor("v_35_begin_0"), val = tensor([0, 0, 2, 0, 0])]; + tensor v_35_end_0 = const()[name = tensor("v_35_end_0"), val = tensor([1, 256, 3, 16, 64])]; + tensor v_35_end_mask_0 = const()[name = tensor("v_35_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor v_35_squeeze_mask_0 = const()[name = tensor("v_35_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor v_35_cast_fp16 = slice_by_index(begin = v_35_begin_0, end = v_35_end_0, end_mask = v_35_end_mask_0, squeeze_mask = v_35_squeeze_mask_0, x = qkv_35_cast_fp16)[name = tensor("v_35_cast_fp16")]; + tensor var_7251 = const()[name = tensor("op_7251"), val = tensor([1, 1])]; + tensor position17_to_fp16_dtype_0 = const()[name = tensor("position17_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor position17_to_fp16 = cast(dtype = position17_to_fp16_dtype_0, x = position17)[name = tensor("cast_576")]; + tensor var_7252_cast_fp16 = reshape(shape = var_7251, x = position17_to_fp16)[name = tensor("op_7252_cast_fp16")]; + tensor ts_105_cast_fp16 = add(x = ts_1_promoted_to_fp16, y = var_7252_cast_fp16)[name = tensor("ts_105_cast_fp16")]; + tensor var_7259 = const()[name = tensor("op_7259"), val = tensor([1, 256, 1, 1])]; + tensor ts_107_cast_fp16 = reshape(shape = var_7259, x = ts_105_cast_fp16)[name = tensor("ts_107_cast_fp16")]; + tensor var_7266 = const()[name = tensor("op_7266"), val = tensor([1, 256, 16, 32, 2])]; + tensor q_complex_35_cast_fp16 = reshape(shape = var_7266, x = q_103_cast_fp16)[name = tensor("q_complex_35_cast_fp16")]; + tensor var_7273 = const()[name = tensor("op_7273"), val = tensor([1, 256, 16, 32, 2])]; + tensor k_complex_35_cast_fp16 = reshape(shape = var_7273, x = k_69_cast_fp16)[name = tensor("k_complex_35_cast_fp16")]; + tensor var_7277_begin_0 = const()[name = tensor("op_7277_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_7277_end_0 = const()[name = tensor("op_7277_end_0"), val = tensor([1, 256, 16, 32, 1])]; + tensor var_7277_end_mask_0 = const()[name = tensor("op_7277_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_7277_squeeze_mask_0 = const()[name = tensor("op_7277_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_7277_cast_fp16 = slice_by_index(begin = var_7277_begin_0, end = var_7277_end_0, end_mask = var_7277_end_mask_0, squeeze_mask = var_7277_squeeze_mask_0, x = q_complex_35_cast_fp16)[name = tensor("op_7277_cast_fp16")]; + tensor var_7285_begin_0 = const()[name = tensor("op_7285_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_7285_end_0 = const()[name = tensor("op_7285_end_0"), val = tensor([1, 256, 16, 32, 2])]; + tensor var_7285_end_mask_0 = const()[name = tensor("op_7285_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_7285_squeeze_mask_0 = const()[name = tensor("op_7285_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_7285_cast_fp16 = slice_by_index(begin = var_7285_begin_0, end = var_7285_end_0, end_mask = var_7285_end_mask_0, squeeze_mask = var_7285_squeeze_mask_0, x = q_complex_35_cast_fp16)[name = tensor("op_7285_cast_fp16")]; + tensor var_7293_begin_0 = const()[name = tensor("op_7293_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_7293_end_0 = const()[name = tensor("op_7293_end_0"), val = tensor([1, 256, 16, 32, 1])]; + tensor var_7293_end_mask_0 = const()[name = tensor("op_7293_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_7293_squeeze_mask_0 = const()[name = tensor("op_7293_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_7293_cast_fp16 = slice_by_index(begin = var_7293_begin_0, end = var_7293_end_0, end_mask = var_7293_end_mask_0, squeeze_mask = var_7293_squeeze_mask_0, x = k_complex_35_cast_fp16)[name = tensor("op_7293_cast_fp16")]; + tensor var_7301_begin_0 = const()[name = tensor("op_7301_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_7301_end_0 = const()[name = tensor("op_7301_end_0"), val = tensor([1, 256, 16, 32, 2])]; + tensor var_7301_end_mask_0 = const()[name = tensor("op_7301_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_7301_squeeze_mask_0 = const()[name = tensor("op_7301_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_7301_cast_fp16 = slice_by_index(begin = var_7301_begin_0, end = var_7301_end_0, end_mask = var_7301_end_mask_0, squeeze_mask = var_7301_squeeze_mask_0, x = k_complex_35_cast_fp16)[name = tensor("op_7301_cast_fp16")]; + tensor freqs_35_to_fp16 = const()[name = tensor("freqs_35_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(435332992)))]; + tensor var_7307_cast_fp16 = mul(x = freqs_35_to_fp16, y = ts_107_cast_fp16)[name = tensor("op_7307_cast_fp16")]; + tensor rotr_35_cast_fp16 = cos(x = var_7307_cast_fp16)[name = tensor("rotr_35_cast_fp16")]; + tensor roti_35_cast_fp16 = sin(x = var_7307_cast_fp16)[name = tensor("roti_35_cast_fp16")]; + tensor var_7311_cast_fp16 = mul(x = var_7277_cast_fp16, y = rotr_35_cast_fp16)[name = tensor("op_7311_cast_fp16")]; + tensor var_7312_cast_fp16 = mul(x = var_7285_cast_fp16, y = roti_35_cast_fp16)[name = tensor("op_7312_cast_fp16")]; + tensor qor_69_cast_fp16 = sub(x = var_7311_cast_fp16, y = var_7312_cast_fp16)[name = tensor("qor_69_cast_fp16")]; + tensor var_7315_cast_fp16 = mul(x = var_7277_cast_fp16, y = roti_35_cast_fp16)[name = tensor("op_7315_cast_fp16")]; + tensor var_7316_cast_fp16 = mul(x = var_7285_cast_fp16, y = rotr_35_cast_fp16)[name = tensor("op_7316_cast_fp16")]; + tensor qoi_69_cast_fp16 = add(x = var_7315_cast_fp16, y = var_7316_cast_fp16)[name = tensor("qoi_69_cast_fp16")]; + tensor var_7319_cast_fp16 = mul(x = var_7293_cast_fp16, y = rotr_35_cast_fp16)[name = tensor("op_7319_cast_fp16")]; + tensor var_7320_cast_fp16 = mul(x = var_7301_cast_fp16, y = roti_35_cast_fp16)[name = tensor("op_7320_cast_fp16")]; + tensor kor_69_cast_fp16 = sub(x = var_7319_cast_fp16, y = var_7320_cast_fp16)[name = tensor("kor_69_cast_fp16")]; + tensor var_7323_cast_fp16 = mul(x = var_7293_cast_fp16, y = roti_35_cast_fp16)[name = tensor("op_7323_cast_fp16")]; + tensor var_7324_cast_fp16 = mul(x = var_7301_cast_fp16, y = rotr_35_cast_fp16)[name = tensor("op_7324_cast_fp16")]; + tensor koi_69_cast_fp16 = add(x = var_7323_cast_fp16, y = var_7324_cast_fp16)[name = tensor("koi_69_cast_fp16")]; + tensor qo_35_axis_0 = const()[name = tensor("qo_35_axis_0"), val = tensor(-1)]; + tensor qo_35_cast_fp16 = stack(axis = qo_35_axis_0, values = (qor_69_cast_fp16, qoi_69_cast_fp16))[name = tensor("qo_35_cast_fp16")]; + tensor ko_35_axis_0 = const()[name = tensor("ko_35_axis_0"), val = tensor(-1)]; + tensor ko_35_cast_fp16 = stack(axis = ko_35_axis_0, values = (kor_69_cast_fp16, koi_69_cast_fp16))[name = tensor("ko_35_cast_fp16")]; + tensor var_7357 = const()[name = tensor("op_7357"), val = tensor([1, 256, 16, 64])]; + tensor q_105_cast_fp16 = reshape(shape = var_7357, x = qo_35_cast_fp16)[name = tensor("q_105_cast_fp16")]; + tensor var_7363 = const()[name = tensor("op_7363"), val = tensor([1, 256, 16, 64])]; + tensor k_71_cast_fp16 = reshape(shape = var_7363, x = ko_35_cast_fp16)[name = tensor("k_71_cast_fp16")]; + tensor write_indices_float_103_cast_fp16 = add(x = var_7252_cast_fp16, y = var_393_promoted_to_fp16)[name = tensor("write_indices_float_103_cast_fp16")]; + tensor _inversed_7384_y_0_to_fp16 = const()[name = tensor("_inversed_7384_y_0_to_fp16"), val = tensor(0x1p-9)]; + tensor _inversed_7384_cast_fp16 = mul(x = write_indices_float_103_cast_fp16, y = _inversed_7384_y_0_to_fp16)[name = tensor("_inversed_7384_cast_fp16")]; + tensor var_7385_cast_fp16 = floor(x = _inversed_7384_cast_fp16)[name = tensor("op_7385_cast_fp16")]; + tensor var_7386_to_fp16 = const()[name = tensor("op_7386_to_fp16"), val = tensor(0x1p+9)]; + tensor var_7387_cast_fp16 = mul(x = var_7385_cast_fp16, y = var_7386_to_fp16)[name = tensor("op_7387_cast_fp16")]; + tensor write_indices_float_105_cast_fp16 = sub(x = write_indices_float_103_cast_fp16, y = var_7387_cast_fp16)[name = tensor("write_indices_float_105_cast_fp16")]; + tensor var_7405_value_0_to_fp16 = const()[name = tensor("op_7405_value_0_to_fp16"), val = tensor(0x1.ffp+8)]; + tensor var_7405_cast_fp16 = fill_like(ref_tensor = write_indices_float_105_cast_fp16, value = var_7405_value_0_to_fp16)[name = tensor("op_7405_cast_fp16")]; + tensor write_indices_float_107_cast_fp16 = select(a = write_indices_float_105_cast_fp16, b = var_7405_cast_fp16, cond = token_valid_1_cast_fp16)[name = tensor("write_indices_float_107_cast_fp16")]; + tensor var_7411_dtype_0 = const()[name = tensor("op_7411_dtype_0"), val = tensor("int32")]; + tensor var_7416 = const()[name = tensor("op_7416"), val = tensor([1, 256, 1, 1])]; + tensor write_indices_float_107_cast_fp16_to_int32 = cast(dtype = var_7411_dtype_0, x = write_indices_float_107_cast_fp16)[name = tensor("cast_575")]; + tensor var_7417 = reshape(shape = var_7416, x = write_indices_float_107_cast_fp16_to_int32)[name = tensor("op_7417")]; + tensor write_indices_35_reps_0 = const()[name = tensor("write_indices_35_reps_0"), val = tensor([1, 1, 16, 64])]; + tensor write_indices_35 = tile(reps = write_indices_35_reps_0, x = var_7417)[name = tensor("write_indices_35")]; + tensor var_7427_begin_0 = const()[name = tensor("op_7427_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_7427_end_0 = const()[name = tensor("op_7427_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor var_7427_end_mask_0 = const()[name = tensor("op_7427_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_7427_squeeze_mask_0 = const()[name = tensor("op_7427_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor cache17_to_fp16_dtype_0 = const()[name = tensor("cache17_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor cache17_to_fp16 = cast(dtype = cache17_to_fp16_dtype_0, x = cache17)[name = tensor("cast_574")]; + tensor var_7427_cast_fp16 = slice_by_index(begin = var_7427_begin_0, end = var_7427_end_0, end_mask = var_7427_end_mask_0, squeeze_mask = var_7427_squeeze_mask_0, x = cache17_to_fp16)[name = tensor("op_7427_cast_fp16")]; + tensor var_7429_axis_0 = const()[name = tensor("op_7429_axis_0"), val = tensor(1)]; + tensor var_7429_mode_0 = const()[name = tensor("op_7429_mode_0"), val = tensor("update")]; + tensor var_7429_validate_indices_0 = const()[name = tensor("op_7429_validate_indices_0"), val = tensor(false)]; + tensor var_7429_cast_fp16 = scatter_along_axis(axis = var_7429_axis_0, data = var_7427_cast_fp16, indices = write_indices_35, mode = var_7429_mode_0, updates = k_71_cast_fp16, validate_indices = var_7429_validate_indices_0)[name = tensor("op_7429_cast_fp16")]; + tensor concat_138 = const()[name = tensor("concat_138"), val = tensor([0, 0, 0, 0, 0])]; + tensor concat_139 = const()[name = tensor("concat_139"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_35_internal_tensor_assign_1_stride_0 = const()[name = tensor("new_cache_35_internal_tensor_assign_1_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_35_internal_tensor_assign_1_begin_mask_0 = const()[name = tensor("new_cache_35_internal_tensor_assign_1_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_35_internal_tensor_assign_1_end_mask_0 = const()[name = tensor("new_cache_35_internal_tensor_assign_1_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_35_internal_tensor_assign_1_squeeze_mask_0 = const()[name = tensor("new_cache_35_internal_tensor_assign_1_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_80 = const()[name = tensor("shape_80"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_34 = const()[name = tensor("reduce_prod_34"), val = tensor(1048576)]; + tensor range_1d_34_start_0 = const()[name = tensor("range_1d_34_start_0"), val = tensor(0)]; + tensor range_1d_34_step_0 = const()[name = tensor("range_1d_34_step_0"), val = tensor(1)]; + tensor range_1d_34 = range_1d(end = reduce_prod_34, start = range_1d_34_start_0, step = range_1d_34_step_0)[name = tensor("range_1d_34")]; + tensor reshape_170 = reshape(shape = shape_80, x = range_1d_34)[name = tensor("reshape_170")]; + tensor slice_by_index_34 = slice_by_index(begin = concat_138, begin_mask = new_cache_35_internal_tensor_assign_1_begin_mask_0, end = concat_139, end_mask = new_cache_35_internal_tensor_assign_1_end_mask_0, squeeze_mask = new_cache_35_internal_tensor_assign_1_squeeze_mask_0, stride = new_cache_35_internal_tensor_assign_1_stride_0, x = reshape_170)[name = tensor("slice_by_index_34")]; + tensor reshape_171_shape_0 = const()[name = tensor("reshape_171_shape_0"), val = tensor([-1])]; + tensor reshape_171 = reshape(shape = reshape_171_shape_0, x = slice_by_index_34)[name = tensor("reshape_171")]; + tensor reshape_172_shape_0 = const()[name = tensor("reshape_172_shape_0"), val = tensor([-1])]; + tensor reshape_172_cast_fp16 = reshape(shape = reshape_172_shape_0, x = var_7429_cast_fp16)[name = tensor("reshape_172_cast_fp16")]; + tensor reshape_173_shape_0 = const()[name = tensor("reshape_173_shape_0"), val = tensor([-1])]; + tensor reshape_173_cast_fp16 = reshape(shape = reshape_173_shape_0, x = cache17_to_fp16)[name = tensor("reshape_173_cast_fp16")]; + tensor scatter_34_mode_0 = const()[name = tensor("scatter_34_mode_0"), val = tensor("update")]; + tensor scatter_34_axis_0 = const()[name = tensor("scatter_34_axis_0"), val = tensor(0)]; + tensor scatter_34_validate_indices_0 = const()[name = tensor("scatter_34_validate_indices_0"), val = tensor(false)]; + tensor scatter_34_cast_fp16 = scatter(axis = scatter_34_axis_0, data = reshape_173_cast_fp16, indices = reshape_171, mode = scatter_34_mode_0, updates = reshape_172_cast_fp16, validate_indices = scatter_34_validate_indices_0)[name = tensor("scatter_34_cast_fp16")]; + tensor reshape_174_cast_fp16 = reshape(shape = shape_80, x = scatter_34_cast_fp16)[name = tensor("reshape_174_cast_fp16")]; + tensor var_7437_begin_0 = const()[name = tensor("op_7437_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor var_7437_end_0 = const()[name = tensor("op_7437_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor var_7437_end_mask_0 = const()[name = tensor("op_7437_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_7437_squeeze_mask_0 = const()[name = tensor("op_7437_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor var_7437_cast_fp16 = slice_by_index(begin = var_7437_begin_0, end = var_7437_end_0, end_mask = var_7437_end_mask_0, squeeze_mask = var_7437_squeeze_mask_0, x = reshape_174_cast_fp16)[name = tensor("op_7437_cast_fp16")]; + tensor var_7439_axis_0 = const()[name = tensor("op_7439_axis_0"), val = tensor(1)]; + tensor var_7439_mode_0 = const()[name = tensor("op_7439_mode_0"), val = tensor("update")]; + tensor var_7439_validate_indices_0 = const()[name = tensor("op_7439_validate_indices_0"), val = tensor(false)]; + tensor var_7439_cast_fp16 = scatter_along_axis(axis = var_7439_axis_0, data = var_7437_cast_fp16, indices = write_indices_35, mode = var_7439_mode_0, updates = v_35_cast_fp16, validate_indices = var_7439_validate_indices_0)[name = tensor("op_7439_cast_fp16")]; + tensor concat_140 = const()[name = tensor("concat_140"), val = tensor([1, 0, 0, 0, 0])]; + tensor concat_141 = const()[name = tensor("concat_141"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_35_internal_tensor_assign_2_stride_0 = const()[name = tensor("new_cache_35_internal_tensor_assign_2_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_35_internal_tensor_assign_2_begin_mask_0 = const()[name = tensor("new_cache_35_internal_tensor_assign_2_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_35_internal_tensor_assign_2_end_mask_0 = const()[name = tensor("new_cache_35_internal_tensor_assign_2_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_35_internal_tensor_assign_2_squeeze_mask_0 = const()[name = tensor("new_cache_35_internal_tensor_assign_2_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_81 = const()[name = tensor("shape_81"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_35 = const()[name = tensor("reduce_prod_35"), val = tensor(1048576)]; + tensor range_1d_35_start_0 = const()[name = tensor("range_1d_35_start_0"), val = tensor(0)]; + tensor range_1d_35_step_0 = const()[name = tensor("range_1d_35_step_0"), val = tensor(1)]; + tensor range_1d_35 = range_1d(end = reduce_prod_35, start = range_1d_35_start_0, step = range_1d_35_step_0)[name = tensor("range_1d_35")]; + tensor reshape_175 = reshape(shape = shape_81, x = range_1d_35)[name = tensor("reshape_175")]; + tensor slice_by_index_35 = slice_by_index(begin = concat_140, begin_mask = new_cache_35_internal_tensor_assign_2_begin_mask_0, end = concat_141, end_mask = new_cache_35_internal_tensor_assign_2_end_mask_0, squeeze_mask = new_cache_35_internal_tensor_assign_2_squeeze_mask_0, stride = new_cache_35_internal_tensor_assign_2_stride_0, x = reshape_175)[name = tensor("slice_by_index_35")]; + tensor reshape_176_shape_0 = const()[name = tensor("reshape_176_shape_0"), val = tensor([-1])]; + tensor reshape_176 = reshape(shape = reshape_176_shape_0, x = slice_by_index_35)[name = tensor("reshape_176")]; + tensor reshape_177_shape_0 = const()[name = tensor("reshape_177_shape_0"), val = tensor([-1])]; + tensor reshape_177_cast_fp16 = reshape(shape = reshape_177_shape_0, x = var_7439_cast_fp16)[name = tensor("reshape_177_cast_fp16")]; + tensor reshape_178_shape_0 = const()[name = tensor("reshape_178_shape_0"), val = tensor([-1])]; + tensor reshape_178_cast_fp16 = reshape(shape = reshape_178_shape_0, x = reshape_174_cast_fp16)[name = tensor("reshape_178_cast_fp16")]; + tensor scatter_35_mode_0 = const()[name = tensor("scatter_35_mode_0"), val = tensor("update")]; + tensor scatter_35_axis_0 = const()[name = tensor("scatter_35_axis_0"), val = tensor(0)]; + tensor scatter_35_validate_indices_0 = const()[name = tensor("scatter_35_validate_indices_0"), val = tensor(false)]; + tensor scatter_35_cast_fp16 = scatter(axis = scatter_35_axis_0, data = reshape_178_cast_fp16, indices = reshape_176, mode = scatter_35_mode_0, updates = reshape_177_cast_fp16, validate_indices = scatter_35_validate_indices_0)[name = tensor("scatter_35_cast_fp16")]; + tensor reshape_179_cast_fp16 = reshape(shape = shape_81, x = scatter_35_cast_fp16)[name = tensor("reshape_179_cast_fp16")]; + tensor reshape_179_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("reshape_179_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor keys_103_begin_0 = const()[name = tensor("keys_103_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor keys_103_end_0 = const()[name = tensor("keys_103_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor keys_103_end_mask_0 = const()[name = tensor("keys_103_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor keys_103_squeeze_mask_0 = const()[name = tensor("keys_103_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor keys_103_cast_fp16 = slice_by_index(begin = keys_103_begin_0, end = keys_103_end_0, end_mask = keys_103_end_mask_0, squeeze_mask = keys_103_squeeze_mask_0, x = reshape_179_cast_fp16)[name = tensor("keys_103_cast_fp16")]; + tensor values_103_begin_0 = const()[name = tensor("values_103_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor values_103_end_0 = const()[name = tensor("values_103_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor values_103_end_mask_0 = const()[name = tensor("values_103_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor values_103_squeeze_mask_0 = const()[name = tensor("values_103_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor values_103_cast_fp16 = slice_by_index(begin = values_103_begin_0, end = values_103_end_0, end_mask = values_103_end_mask_0, squeeze_mask = values_103_squeeze_mask_0, x = reshape_179_cast_fp16)[name = tensor("values_103_cast_fp16")]; + tensor var_7451_cast_fp16 = not_equal(x = keys_103_cast_fp16, y = keys_103_cast_fp16)[name = tensor("op_7451_cast_fp16")]; + tensor keys_105_cast_fp16 = select(a = var_470_to_fp16, b = keys_103_cast_fp16, cond = var_7451_cast_fp16)[name = tensor("keys_105_cast_fp16")]; + tensor var_7459_cast_fp16 = not_equal(x = values_103_cast_fp16, y = values_103_cast_fp16)[name = tensor("op_7459_cast_fp16")]; + tensor values_105_cast_fp16 = select(a = var_470_to_fp16, b = values_103_cast_fp16, cond = var_7459_cast_fp16)[name = tensor("values_105_cast_fp16")]; + tensor var_7483 = const()[name = tensor("op_7483"), val = tensor([0, 2, 1, 3])]; + tensor var_7499 = const()[name = tensor("op_7499"), val = tensor([1, 1, 1])]; + tensor var_7500_cast_fp16 = reshape(shape = var_7499, x = position17_to_fp16)[name = tensor("op_7500_cast_fp16")]; + tensor q_positions_35_cast_fp16 = add(x = var_7500_cast_fp16, y = q_offsets_1_promoted_to_fp16)[name = tensor("q_positions_35_cast_fp16")]; + tensor valid_end_35_cast_fp16 = add(x = var_7500_cast_fp16, y = var_536_cast_fp16)[name = tensor("valid_end_35_cast_fp16")]; + tensor valid_mask_35_cast_fp16 = less(x = k_positions_1_promoted_to_fp16, y = valid_end_35_cast_fp16)[name = tensor("valid_mask_35_cast_fp16")]; + tensor causal_mask_35_cast_fp16 = less_equal(x = k_positions_1_promoted_to_fp16, y = q_positions_35_cast_fp16)[name = tensor("causal_mask_35_cast_fp16")]; + tensor var_7528 = logical_and(x = valid_mask_35_cast_fp16, y = causal_mask_35_cast_fp16)[name = tensor("op_7528")]; + tensor attn_mask_35_axes_0 = const()[name = tensor("attn_mask_35_axes_0"), val = tensor([1])]; + tensor attn_mask_35 = expand_dims(axes = attn_mask_35_axes_0, x = var_7528)[name = tensor("attn_mask_35")]; + tensor var_7542_transpose_x_0 = const()[name = tensor("op_7542_transpose_x_0"), val = tensor(false)]; + tensor var_7542_transpose_y_0 = const()[name = tensor("op_7542_transpose_y_0"), val = tensor(false)]; + tensor transpose_103_perm_0 = const()[name = tensor("transpose_103_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_104_perm_0 = const()[name = tensor("transpose_104_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_104 = transpose(perm = transpose_104_perm_0, x = keys_105_cast_fp16)[name = tensor("transpose_136")]; + tensor transpose_103 = transpose(perm = transpose_103_perm_0, x = q_105_cast_fp16)[name = tensor("transpose_137")]; + tensor var_7542_cast_fp16 = matmul(transpose_x = var_7542_transpose_x_0, transpose_y = var_7542_transpose_y_0, x = transpose_103, y = transpose_104)[name = tensor("op_7542_cast_fp16")]; + tensor var_7536_to_fp16 = const()[name = tensor("op_7536_to_fp16"), val = tensor([0x1p-3])]; + tensor attn_weights_103_cast_fp16 = mul(x = var_7542_cast_fp16, y = var_7536_to_fp16)[name = tensor("attn_weights_103_cast_fp16")]; + tensor var_7544 = logical_not(x = attn_mask_35)[name = tensor("op_7544")]; + tensor var_7545_to_fp16 = const()[name = tensor("op_7545_to_fp16"), val = tensor(-inf)]; + tensor attn_weights_105_cast_fp16 = select(a = var_7545_to_fp16, b = attn_weights_103_cast_fp16, cond = var_7544)[name = tensor("attn_weights_105_cast_fp16")]; + tensor var_7547 = const()[name = tensor("op_7547"), val = tensor(-1)]; + tensor attn_weights_107_cast_fp16 = softmax(axis = var_7547, x = attn_weights_105_cast_fp16)[name = tensor("attn_weights_107_cast_fp16")]; + tensor attn_output_35_transpose_x_0 = const()[name = tensor("attn_output_35_transpose_x_0"), val = tensor(false)]; + tensor attn_output_35_transpose_y_0 = const()[name = tensor("attn_output_35_transpose_y_0"), val = tensor(false)]; + tensor values_107_cast_fp16 = transpose(perm = var_7483, x = values_105_cast_fp16)[name = tensor("transpose_138")]; + tensor attn_output_35_cast_fp16 = matmul(transpose_x = attn_output_35_transpose_x_0, transpose_y = attn_output_35_transpose_y_0, x = attn_weights_107_cast_fp16, y = values_107_cast_fp16)[name = tensor("attn_output_35_cast_fp16")]; + tensor var_7555 = const()[name = tensor("op_7555"), val = tensor([0, 2, 1, 3])]; + tensor var_7560 = const()[name = tensor("op_7560"), val = tensor([1, 256, 1024])]; + tensor var_7556_cast_fp16 = transpose(perm = var_7555, x = attn_output_35_cast_fp16)[name = tensor("transpose_135")]; + tensor input_209_cast_fp16 = reshape(shape = var_7560, x = var_7556_cast_fp16)[name = tensor("input_209_cast_fp16")]; + tensor attn17_out_proj_weight_to_fp16 = const()[name = tensor("attn17_out_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(435333120)))]; + tensor linear_69_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = attn17_out_proj_weight_to_fp16, x = input_209_cast_fp16)[name = tensor("linear_69_cast_fp16")]; + tensor var_7569_cast_fp16 = add(x = position17_to_fp16, y = valid_len_to_fp16)[name = tensor("op_7569_cast_fp16")]; + tensor var_7569_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_7569_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor input_211_cast_fp16 = add(x = input_205_cast_fp16, y = linear_69_cast_fp16)[name = tensor("input_211_cast_fp16")]; + tensor input_213_axes_0 = const()[name = tensor("input_213_axes_0"), val = tensor([-1])]; + tensor norm17_2_weight_to_fp16 = const()[name = tensor("norm17_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(437430336)))]; + tensor norm17_2_bias_to_fp16 = const()[name = tensor("norm17_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(437432448)))]; + tensor var_7573_to_fp16 = const()[name = tensor("op_7573_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_213_cast_fp16 = layer_norm(axes = input_213_axes_0, beta = norm17_2_bias_to_fp16, epsilon = var_7573_to_fp16, gamma = norm17_2_weight_to_fp16, x = input_211_cast_fp16)[name = tensor("input_213_cast_fp16")]; + tensor linear17_1_weight_to_fp16 = const()[name = tensor("linear17_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(437434560)))]; + tensor linear_70_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = linear17_1_weight_to_fp16, x = input_213_cast_fp16)[name = tensor("linear_70_cast_fp16")]; + tensor input_215_mode_0 = const()[name = tensor("input_215_mode_0"), val = tensor("EXACT")]; + tensor input_215_cast_fp16 = gelu(mode = input_215_mode_0, x = linear_70_cast_fp16)[name = tensor("input_215_cast_fp16")]; + tensor linear17_2_weight_to_fp16 = const()[name = tensor("linear17_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(445823232)))]; + tensor linear_71_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = linear17_2_weight_to_fp16, x = input_215_cast_fp16)[name = tensor("linear_71_cast_fp16")]; + tensor input_217_cast_fp16 = add(x = input_211_cast_fp16, y = linear_71_cast_fp16)[name = tensor("input_217_cast_fp16")]; + tensor input_219_axes_0 = const()[name = tensor("input_219_axes_0"), val = tensor([-1])]; + tensor norm18_1_weight_to_fp16 = const()[name = tensor("norm18_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(454211904)))]; + tensor norm18_1_bias_to_fp16 = const()[name = tensor("norm18_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(454214016)))]; + tensor var_7590_to_fp16 = const()[name = tensor("op_7590_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_219_cast_fp16 = layer_norm(axes = input_219_axes_0, beta = norm18_1_bias_to_fp16, epsilon = var_7590_to_fp16, gamma = norm18_1_weight_to_fp16, x = input_217_cast_fp16)[name = tensor("input_219_cast_fp16")]; + tensor attn18_in_proj_weight_to_fp16 = const()[name = tensor("attn18_in_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(454216128)))]; + tensor linear_72_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = attn18_in_proj_weight_to_fp16, x = input_219_cast_fp16)[name = tensor("linear_72_cast_fp16")]; + tensor var_7604 = const()[name = tensor("op_7604"), val = tensor([1, 256, 3, 16, 64])]; + tensor qkv_37_cast_fp16 = reshape(shape = var_7604, x = linear_72_cast_fp16)[name = tensor("qkv_37_cast_fp16")]; + tensor q_109_begin_0 = const()[name = tensor("q_109_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor q_109_end_0 = const()[name = tensor("q_109_end_0"), val = tensor([1, 256, 1, 16, 64])]; + tensor q_109_end_mask_0 = const()[name = tensor("q_109_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor q_109_squeeze_mask_0 = const()[name = tensor("q_109_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor q_109_cast_fp16 = slice_by_index(begin = q_109_begin_0, end = q_109_end_0, end_mask = q_109_end_mask_0, squeeze_mask = q_109_squeeze_mask_0, x = qkv_37_cast_fp16)[name = tensor("q_109_cast_fp16")]; + tensor k_73_begin_0 = const()[name = tensor("k_73_begin_0"), val = tensor([0, 0, 1, 0, 0])]; + tensor k_73_end_0 = const()[name = tensor("k_73_end_0"), val = tensor([1, 256, 2, 16, 64])]; + tensor k_73_end_mask_0 = const()[name = tensor("k_73_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor k_73_squeeze_mask_0 = const()[name = tensor("k_73_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor k_73_cast_fp16 = slice_by_index(begin = k_73_begin_0, end = k_73_end_0, end_mask = k_73_end_mask_0, squeeze_mask = k_73_squeeze_mask_0, x = qkv_37_cast_fp16)[name = tensor("k_73_cast_fp16")]; + tensor v_37_begin_0 = const()[name = tensor("v_37_begin_0"), val = tensor([0, 0, 2, 0, 0])]; + tensor v_37_end_0 = const()[name = tensor("v_37_end_0"), val = tensor([1, 256, 3, 16, 64])]; + tensor v_37_end_mask_0 = const()[name = tensor("v_37_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor v_37_squeeze_mask_0 = const()[name = tensor("v_37_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor v_37_cast_fp16 = slice_by_index(begin = v_37_begin_0, end = v_37_end_0, end_mask = v_37_end_mask_0, squeeze_mask = v_37_squeeze_mask_0, x = qkv_37_cast_fp16)[name = tensor("v_37_cast_fp16")]; + tensor var_7662 = const()[name = tensor("op_7662"), val = tensor([1, 1])]; + tensor position18_to_fp16_dtype_0 = const()[name = tensor("position18_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor position18_to_fp16 = cast(dtype = position18_to_fp16_dtype_0, x = position18)[name = tensor("cast_571")]; + tensor var_7663_cast_fp16 = reshape(shape = var_7662, x = position18_to_fp16)[name = tensor("op_7663_cast_fp16")]; + tensor ts_111_cast_fp16 = add(x = ts_1_promoted_to_fp16, y = var_7663_cast_fp16)[name = tensor("ts_111_cast_fp16")]; + tensor var_7670 = const()[name = tensor("op_7670"), val = tensor([1, 256, 1, 1])]; + tensor ts_113_cast_fp16 = reshape(shape = var_7670, x = ts_111_cast_fp16)[name = tensor("ts_113_cast_fp16")]; + tensor var_7677 = const()[name = tensor("op_7677"), val = tensor([1, 256, 16, 32, 2])]; + tensor q_complex_37_cast_fp16 = reshape(shape = var_7677, x = q_109_cast_fp16)[name = tensor("q_complex_37_cast_fp16")]; + tensor var_7684 = const()[name = tensor("op_7684"), val = tensor([1, 256, 16, 32, 2])]; + tensor k_complex_37_cast_fp16 = reshape(shape = var_7684, x = k_73_cast_fp16)[name = tensor("k_complex_37_cast_fp16")]; + tensor var_7688_begin_0 = const()[name = tensor("op_7688_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_7688_end_0 = const()[name = tensor("op_7688_end_0"), val = tensor([1, 256, 16, 32, 1])]; + tensor var_7688_end_mask_0 = const()[name = tensor("op_7688_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_7688_squeeze_mask_0 = const()[name = tensor("op_7688_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_7688_cast_fp16 = slice_by_index(begin = var_7688_begin_0, end = var_7688_end_0, end_mask = var_7688_end_mask_0, squeeze_mask = var_7688_squeeze_mask_0, x = q_complex_37_cast_fp16)[name = tensor("op_7688_cast_fp16")]; + tensor var_7696_begin_0 = const()[name = tensor("op_7696_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_7696_end_0 = const()[name = tensor("op_7696_end_0"), val = tensor([1, 256, 16, 32, 2])]; + tensor var_7696_end_mask_0 = const()[name = tensor("op_7696_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_7696_squeeze_mask_0 = const()[name = tensor("op_7696_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_7696_cast_fp16 = slice_by_index(begin = var_7696_begin_0, end = var_7696_end_0, end_mask = var_7696_end_mask_0, squeeze_mask = var_7696_squeeze_mask_0, x = q_complex_37_cast_fp16)[name = tensor("op_7696_cast_fp16")]; + tensor var_7704_begin_0 = const()[name = tensor("op_7704_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_7704_end_0 = const()[name = tensor("op_7704_end_0"), val = tensor([1, 256, 16, 32, 1])]; + tensor var_7704_end_mask_0 = const()[name = tensor("op_7704_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_7704_squeeze_mask_0 = const()[name = tensor("op_7704_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_7704_cast_fp16 = slice_by_index(begin = var_7704_begin_0, end = var_7704_end_0, end_mask = var_7704_end_mask_0, squeeze_mask = var_7704_squeeze_mask_0, x = k_complex_37_cast_fp16)[name = tensor("op_7704_cast_fp16")]; + tensor var_7712_begin_0 = const()[name = tensor("op_7712_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_7712_end_0 = const()[name = tensor("op_7712_end_0"), val = tensor([1, 256, 16, 32, 2])]; + tensor var_7712_end_mask_0 = const()[name = tensor("op_7712_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_7712_squeeze_mask_0 = const()[name = tensor("op_7712_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_7712_cast_fp16 = slice_by_index(begin = var_7712_begin_0, end = var_7712_end_0, end_mask = var_7712_end_mask_0, squeeze_mask = var_7712_squeeze_mask_0, x = k_complex_37_cast_fp16)[name = tensor("op_7712_cast_fp16")]; + tensor freqs_37_to_fp16 = const()[name = tensor("freqs_37_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(460507648)))]; + tensor var_7718_cast_fp16 = mul(x = freqs_37_to_fp16, y = ts_113_cast_fp16)[name = tensor("op_7718_cast_fp16")]; + tensor rotr_37_cast_fp16 = cos(x = var_7718_cast_fp16)[name = tensor("rotr_37_cast_fp16")]; + tensor roti_37_cast_fp16 = sin(x = var_7718_cast_fp16)[name = tensor("roti_37_cast_fp16")]; + tensor var_7722_cast_fp16 = mul(x = var_7688_cast_fp16, y = rotr_37_cast_fp16)[name = tensor("op_7722_cast_fp16")]; + tensor var_7723_cast_fp16 = mul(x = var_7696_cast_fp16, y = roti_37_cast_fp16)[name = tensor("op_7723_cast_fp16")]; + tensor qor_73_cast_fp16 = sub(x = var_7722_cast_fp16, y = var_7723_cast_fp16)[name = tensor("qor_73_cast_fp16")]; + tensor var_7726_cast_fp16 = mul(x = var_7688_cast_fp16, y = roti_37_cast_fp16)[name = tensor("op_7726_cast_fp16")]; + tensor var_7727_cast_fp16 = mul(x = var_7696_cast_fp16, y = rotr_37_cast_fp16)[name = tensor("op_7727_cast_fp16")]; + tensor qoi_73_cast_fp16 = add(x = var_7726_cast_fp16, y = var_7727_cast_fp16)[name = tensor("qoi_73_cast_fp16")]; + tensor var_7730_cast_fp16 = mul(x = var_7704_cast_fp16, y = rotr_37_cast_fp16)[name = tensor("op_7730_cast_fp16")]; + tensor var_7731_cast_fp16 = mul(x = var_7712_cast_fp16, y = roti_37_cast_fp16)[name = tensor("op_7731_cast_fp16")]; + tensor kor_73_cast_fp16 = sub(x = var_7730_cast_fp16, y = var_7731_cast_fp16)[name = tensor("kor_73_cast_fp16")]; + tensor var_7734_cast_fp16 = mul(x = var_7704_cast_fp16, y = roti_37_cast_fp16)[name = tensor("op_7734_cast_fp16")]; + tensor var_7735_cast_fp16 = mul(x = var_7712_cast_fp16, y = rotr_37_cast_fp16)[name = tensor("op_7735_cast_fp16")]; + tensor koi_73_cast_fp16 = add(x = var_7734_cast_fp16, y = var_7735_cast_fp16)[name = tensor("koi_73_cast_fp16")]; + tensor qo_37_axis_0 = const()[name = tensor("qo_37_axis_0"), val = tensor(-1)]; + tensor qo_37_cast_fp16 = stack(axis = qo_37_axis_0, values = (qor_73_cast_fp16, qoi_73_cast_fp16))[name = tensor("qo_37_cast_fp16")]; + tensor ko_37_axis_0 = const()[name = tensor("ko_37_axis_0"), val = tensor(-1)]; + tensor ko_37_cast_fp16 = stack(axis = ko_37_axis_0, values = (kor_73_cast_fp16, koi_73_cast_fp16))[name = tensor("ko_37_cast_fp16")]; + tensor var_7768 = const()[name = tensor("op_7768"), val = tensor([1, 256, 16, 64])]; + tensor q_111_cast_fp16 = reshape(shape = var_7768, x = qo_37_cast_fp16)[name = tensor("q_111_cast_fp16")]; + tensor var_7774 = const()[name = tensor("op_7774"), val = tensor([1, 256, 16, 64])]; + tensor k_75_cast_fp16 = reshape(shape = var_7774, x = ko_37_cast_fp16)[name = tensor("k_75_cast_fp16")]; + tensor write_indices_float_109_cast_fp16 = add(x = var_7663_cast_fp16, y = var_393_promoted_to_fp16)[name = tensor("write_indices_float_109_cast_fp16")]; + tensor _inversed_7795_y_0_to_fp16 = const()[name = tensor("_inversed_7795_y_0_to_fp16"), val = tensor(0x1p-9)]; + tensor _inversed_7795_cast_fp16 = mul(x = write_indices_float_109_cast_fp16, y = _inversed_7795_y_0_to_fp16)[name = tensor("_inversed_7795_cast_fp16")]; + tensor var_7796_cast_fp16 = floor(x = _inversed_7795_cast_fp16)[name = tensor("op_7796_cast_fp16")]; + tensor var_7797_to_fp16 = const()[name = tensor("op_7797_to_fp16"), val = tensor(0x1p+9)]; + tensor var_7798_cast_fp16 = mul(x = var_7796_cast_fp16, y = var_7797_to_fp16)[name = tensor("op_7798_cast_fp16")]; + tensor write_indices_float_111_cast_fp16 = sub(x = write_indices_float_109_cast_fp16, y = var_7798_cast_fp16)[name = tensor("write_indices_float_111_cast_fp16")]; + tensor var_7816_value_0_to_fp16 = const()[name = tensor("op_7816_value_0_to_fp16"), val = tensor(0x1.ffp+8)]; + tensor var_7816_cast_fp16 = fill_like(ref_tensor = write_indices_float_111_cast_fp16, value = var_7816_value_0_to_fp16)[name = tensor("op_7816_cast_fp16")]; + tensor write_indices_float_113_cast_fp16 = select(a = write_indices_float_111_cast_fp16, b = var_7816_cast_fp16, cond = token_valid_1_cast_fp16)[name = tensor("write_indices_float_113_cast_fp16")]; + tensor var_7822_dtype_0 = const()[name = tensor("op_7822_dtype_0"), val = tensor("int32")]; + tensor var_7827 = const()[name = tensor("op_7827"), val = tensor([1, 256, 1, 1])]; + tensor write_indices_float_113_cast_fp16_to_int32 = cast(dtype = var_7822_dtype_0, x = write_indices_float_113_cast_fp16)[name = tensor("cast_570")]; + tensor var_7828 = reshape(shape = var_7827, x = write_indices_float_113_cast_fp16_to_int32)[name = tensor("op_7828")]; + tensor write_indices_37_reps_0 = const()[name = tensor("write_indices_37_reps_0"), val = tensor([1, 1, 16, 64])]; + tensor write_indices_37 = tile(reps = write_indices_37_reps_0, x = var_7828)[name = tensor("write_indices_37")]; + tensor var_7838_begin_0 = const()[name = tensor("op_7838_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_7838_end_0 = const()[name = tensor("op_7838_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor var_7838_end_mask_0 = const()[name = tensor("op_7838_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_7838_squeeze_mask_0 = const()[name = tensor("op_7838_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor cache18_to_fp16_dtype_0 = const()[name = tensor("cache18_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor cache18_to_fp16 = cast(dtype = cache18_to_fp16_dtype_0, x = cache18)[name = tensor("cast_569")]; + tensor var_7838_cast_fp16 = slice_by_index(begin = var_7838_begin_0, end = var_7838_end_0, end_mask = var_7838_end_mask_0, squeeze_mask = var_7838_squeeze_mask_0, x = cache18_to_fp16)[name = tensor("op_7838_cast_fp16")]; + tensor var_7840_axis_0 = const()[name = tensor("op_7840_axis_0"), val = tensor(1)]; + tensor var_7840_mode_0 = const()[name = tensor("op_7840_mode_0"), val = tensor("update")]; + tensor var_7840_validate_indices_0 = const()[name = tensor("op_7840_validate_indices_0"), val = tensor(false)]; + tensor var_7840_cast_fp16 = scatter_along_axis(axis = var_7840_axis_0, data = var_7838_cast_fp16, indices = write_indices_37, mode = var_7840_mode_0, updates = k_75_cast_fp16, validate_indices = var_7840_validate_indices_0)[name = tensor("op_7840_cast_fp16")]; + tensor concat_146 = const()[name = tensor("concat_146"), val = tensor([0, 0, 0, 0, 0])]; + tensor concat_147 = const()[name = tensor("concat_147"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_37_internal_tensor_assign_1_stride_0 = const()[name = tensor("new_cache_37_internal_tensor_assign_1_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_37_internal_tensor_assign_1_begin_mask_0 = const()[name = tensor("new_cache_37_internal_tensor_assign_1_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_37_internal_tensor_assign_1_end_mask_0 = const()[name = tensor("new_cache_37_internal_tensor_assign_1_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_37_internal_tensor_assign_1_squeeze_mask_0 = const()[name = tensor("new_cache_37_internal_tensor_assign_1_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_82 = const()[name = tensor("shape_82"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_36 = const()[name = tensor("reduce_prod_36"), val = tensor(1048576)]; + tensor range_1d_36_start_0 = const()[name = tensor("range_1d_36_start_0"), val = tensor(0)]; + tensor range_1d_36_step_0 = const()[name = tensor("range_1d_36_step_0"), val = tensor(1)]; + tensor range_1d_36 = range_1d(end = reduce_prod_36, start = range_1d_36_start_0, step = range_1d_36_step_0)[name = tensor("range_1d_36")]; + tensor reshape_180 = reshape(shape = shape_82, x = range_1d_36)[name = tensor("reshape_180")]; + tensor slice_by_index_36 = slice_by_index(begin = concat_146, begin_mask = new_cache_37_internal_tensor_assign_1_begin_mask_0, end = concat_147, end_mask = new_cache_37_internal_tensor_assign_1_end_mask_0, squeeze_mask = new_cache_37_internal_tensor_assign_1_squeeze_mask_0, stride = new_cache_37_internal_tensor_assign_1_stride_0, x = reshape_180)[name = tensor("slice_by_index_36")]; + tensor reshape_181_shape_0 = const()[name = tensor("reshape_181_shape_0"), val = tensor([-1])]; + tensor reshape_181 = reshape(shape = reshape_181_shape_0, x = slice_by_index_36)[name = tensor("reshape_181")]; + tensor reshape_182_shape_0 = const()[name = tensor("reshape_182_shape_0"), val = tensor([-1])]; + tensor reshape_182_cast_fp16 = reshape(shape = reshape_182_shape_0, x = var_7840_cast_fp16)[name = tensor("reshape_182_cast_fp16")]; + tensor reshape_183_shape_0 = const()[name = tensor("reshape_183_shape_0"), val = tensor([-1])]; + tensor reshape_183_cast_fp16 = reshape(shape = reshape_183_shape_0, x = cache18_to_fp16)[name = tensor("reshape_183_cast_fp16")]; + tensor scatter_36_mode_0 = const()[name = tensor("scatter_36_mode_0"), val = tensor("update")]; + tensor scatter_36_axis_0 = const()[name = tensor("scatter_36_axis_0"), val = tensor(0)]; + tensor scatter_36_validate_indices_0 = const()[name = tensor("scatter_36_validate_indices_0"), val = tensor(false)]; + tensor scatter_36_cast_fp16 = scatter(axis = scatter_36_axis_0, data = reshape_183_cast_fp16, indices = reshape_181, mode = scatter_36_mode_0, updates = reshape_182_cast_fp16, validate_indices = scatter_36_validate_indices_0)[name = tensor("scatter_36_cast_fp16")]; + tensor reshape_184_cast_fp16 = reshape(shape = shape_82, x = scatter_36_cast_fp16)[name = tensor("reshape_184_cast_fp16")]; + tensor var_7848_begin_0 = const()[name = tensor("op_7848_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor var_7848_end_0 = const()[name = tensor("op_7848_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor var_7848_end_mask_0 = const()[name = tensor("op_7848_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_7848_squeeze_mask_0 = const()[name = tensor("op_7848_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor var_7848_cast_fp16 = slice_by_index(begin = var_7848_begin_0, end = var_7848_end_0, end_mask = var_7848_end_mask_0, squeeze_mask = var_7848_squeeze_mask_0, x = reshape_184_cast_fp16)[name = tensor("op_7848_cast_fp16")]; + tensor var_7850_axis_0 = const()[name = tensor("op_7850_axis_0"), val = tensor(1)]; + tensor var_7850_mode_0 = const()[name = tensor("op_7850_mode_0"), val = tensor("update")]; + tensor var_7850_validate_indices_0 = const()[name = tensor("op_7850_validate_indices_0"), val = tensor(false)]; + tensor var_7850_cast_fp16 = scatter_along_axis(axis = var_7850_axis_0, data = var_7848_cast_fp16, indices = write_indices_37, mode = var_7850_mode_0, updates = v_37_cast_fp16, validate_indices = var_7850_validate_indices_0)[name = tensor("op_7850_cast_fp16")]; + tensor concat_148 = const()[name = tensor("concat_148"), val = tensor([1, 0, 0, 0, 0])]; + tensor concat_149 = const()[name = tensor("concat_149"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_37_internal_tensor_assign_2_stride_0 = const()[name = tensor("new_cache_37_internal_tensor_assign_2_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_37_internal_tensor_assign_2_begin_mask_0 = const()[name = tensor("new_cache_37_internal_tensor_assign_2_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_37_internal_tensor_assign_2_end_mask_0 = const()[name = tensor("new_cache_37_internal_tensor_assign_2_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_37_internal_tensor_assign_2_squeeze_mask_0 = const()[name = tensor("new_cache_37_internal_tensor_assign_2_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_83 = const()[name = tensor("shape_83"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_37 = const()[name = tensor("reduce_prod_37"), val = tensor(1048576)]; + tensor range_1d_37_start_0 = const()[name = tensor("range_1d_37_start_0"), val = tensor(0)]; + tensor range_1d_37_step_0 = const()[name = tensor("range_1d_37_step_0"), val = tensor(1)]; + tensor range_1d_37 = range_1d(end = reduce_prod_37, start = range_1d_37_start_0, step = range_1d_37_step_0)[name = tensor("range_1d_37")]; + tensor reshape_185 = reshape(shape = shape_83, x = range_1d_37)[name = tensor("reshape_185")]; + tensor slice_by_index_37 = slice_by_index(begin = concat_148, begin_mask = new_cache_37_internal_tensor_assign_2_begin_mask_0, end = concat_149, end_mask = new_cache_37_internal_tensor_assign_2_end_mask_0, squeeze_mask = new_cache_37_internal_tensor_assign_2_squeeze_mask_0, stride = new_cache_37_internal_tensor_assign_2_stride_0, x = reshape_185)[name = tensor("slice_by_index_37")]; + tensor reshape_186_shape_0 = const()[name = tensor("reshape_186_shape_0"), val = tensor([-1])]; + tensor reshape_186 = reshape(shape = reshape_186_shape_0, x = slice_by_index_37)[name = tensor("reshape_186")]; + tensor reshape_187_shape_0 = const()[name = tensor("reshape_187_shape_0"), val = tensor([-1])]; + tensor reshape_187_cast_fp16 = reshape(shape = reshape_187_shape_0, x = var_7850_cast_fp16)[name = tensor("reshape_187_cast_fp16")]; + tensor reshape_188_shape_0 = const()[name = tensor("reshape_188_shape_0"), val = tensor([-1])]; + tensor reshape_188_cast_fp16 = reshape(shape = reshape_188_shape_0, x = reshape_184_cast_fp16)[name = tensor("reshape_188_cast_fp16")]; + tensor scatter_37_mode_0 = const()[name = tensor("scatter_37_mode_0"), val = tensor("update")]; + tensor scatter_37_axis_0 = const()[name = tensor("scatter_37_axis_0"), val = tensor(0)]; + tensor scatter_37_validate_indices_0 = const()[name = tensor("scatter_37_validate_indices_0"), val = tensor(false)]; + tensor scatter_37_cast_fp16 = scatter(axis = scatter_37_axis_0, data = reshape_188_cast_fp16, indices = reshape_186, mode = scatter_37_mode_0, updates = reshape_187_cast_fp16, validate_indices = scatter_37_validate_indices_0)[name = tensor("scatter_37_cast_fp16")]; + tensor reshape_189_cast_fp16 = reshape(shape = shape_83, x = scatter_37_cast_fp16)[name = tensor("reshape_189_cast_fp16")]; + tensor reshape_189_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("reshape_189_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor keys_109_begin_0 = const()[name = tensor("keys_109_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor keys_109_end_0 = const()[name = tensor("keys_109_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor keys_109_end_mask_0 = const()[name = tensor("keys_109_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor keys_109_squeeze_mask_0 = const()[name = tensor("keys_109_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor keys_109_cast_fp16 = slice_by_index(begin = keys_109_begin_0, end = keys_109_end_0, end_mask = keys_109_end_mask_0, squeeze_mask = keys_109_squeeze_mask_0, x = reshape_189_cast_fp16)[name = tensor("keys_109_cast_fp16")]; + tensor values_109_begin_0 = const()[name = tensor("values_109_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor values_109_end_0 = const()[name = tensor("values_109_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor values_109_end_mask_0 = const()[name = tensor("values_109_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor values_109_squeeze_mask_0 = const()[name = tensor("values_109_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor values_109_cast_fp16 = slice_by_index(begin = values_109_begin_0, end = values_109_end_0, end_mask = values_109_end_mask_0, squeeze_mask = values_109_squeeze_mask_0, x = reshape_189_cast_fp16)[name = tensor("values_109_cast_fp16")]; + tensor var_7862_cast_fp16 = not_equal(x = keys_109_cast_fp16, y = keys_109_cast_fp16)[name = tensor("op_7862_cast_fp16")]; + tensor keys_111_cast_fp16 = select(a = var_470_to_fp16, b = keys_109_cast_fp16, cond = var_7862_cast_fp16)[name = tensor("keys_111_cast_fp16")]; + tensor var_7870_cast_fp16 = not_equal(x = values_109_cast_fp16, y = values_109_cast_fp16)[name = tensor("op_7870_cast_fp16")]; + tensor values_111_cast_fp16 = select(a = var_470_to_fp16, b = values_109_cast_fp16, cond = var_7870_cast_fp16)[name = tensor("values_111_cast_fp16")]; + tensor var_7894 = const()[name = tensor("op_7894"), val = tensor([0, 2, 1, 3])]; + tensor var_7910 = const()[name = tensor("op_7910"), val = tensor([1, 1, 1])]; + tensor var_7911_cast_fp16 = reshape(shape = var_7910, x = position18_to_fp16)[name = tensor("op_7911_cast_fp16")]; + tensor q_positions_37_cast_fp16 = add(x = var_7911_cast_fp16, y = q_offsets_1_promoted_to_fp16)[name = tensor("q_positions_37_cast_fp16")]; + tensor valid_end_37_cast_fp16 = add(x = var_7911_cast_fp16, y = var_536_cast_fp16)[name = tensor("valid_end_37_cast_fp16")]; + tensor valid_mask_37_cast_fp16 = less(x = k_positions_1_promoted_to_fp16, y = valid_end_37_cast_fp16)[name = tensor("valid_mask_37_cast_fp16")]; + tensor causal_mask_37_cast_fp16 = less_equal(x = k_positions_1_promoted_to_fp16, y = q_positions_37_cast_fp16)[name = tensor("causal_mask_37_cast_fp16")]; + tensor var_7939 = logical_and(x = valid_mask_37_cast_fp16, y = causal_mask_37_cast_fp16)[name = tensor("op_7939")]; + tensor attn_mask_37_axes_0 = const()[name = tensor("attn_mask_37_axes_0"), val = tensor([1])]; + tensor attn_mask_37 = expand_dims(axes = attn_mask_37_axes_0, x = var_7939)[name = tensor("attn_mask_37")]; + tensor var_7953_transpose_x_0 = const()[name = tensor("op_7953_transpose_x_0"), val = tensor(false)]; + tensor var_7953_transpose_y_0 = const()[name = tensor("op_7953_transpose_y_0"), val = tensor(false)]; + tensor transpose_105_perm_0 = const()[name = tensor("transpose_105_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_106_perm_0 = const()[name = tensor("transpose_106_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_106 = transpose(perm = transpose_106_perm_0, x = keys_111_cast_fp16)[name = tensor("transpose_132")]; + tensor transpose_105 = transpose(perm = transpose_105_perm_0, x = q_111_cast_fp16)[name = tensor("transpose_133")]; + tensor var_7953_cast_fp16 = matmul(transpose_x = var_7953_transpose_x_0, transpose_y = var_7953_transpose_y_0, x = transpose_105, y = transpose_106)[name = tensor("op_7953_cast_fp16")]; + tensor var_7947_to_fp16 = const()[name = tensor("op_7947_to_fp16"), val = tensor([0x1p-3])]; + tensor attn_weights_109_cast_fp16 = mul(x = var_7953_cast_fp16, y = var_7947_to_fp16)[name = tensor("attn_weights_109_cast_fp16")]; + tensor var_7955 = logical_not(x = attn_mask_37)[name = tensor("op_7955")]; + tensor var_7956_to_fp16 = const()[name = tensor("op_7956_to_fp16"), val = tensor(-inf)]; + tensor attn_weights_111_cast_fp16 = select(a = var_7956_to_fp16, b = attn_weights_109_cast_fp16, cond = var_7955)[name = tensor("attn_weights_111_cast_fp16")]; + tensor var_7958 = const()[name = tensor("op_7958"), val = tensor(-1)]; + tensor attn_weights_113_cast_fp16 = softmax(axis = var_7958, x = attn_weights_111_cast_fp16)[name = tensor("attn_weights_113_cast_fp16")]; + tensor attn_output_37_transpose_x_0 = const()[name = tensor("attn_output_37_transpose_x_0"), val = tensor(false)]; + tensor attn_output_37_transpose_y_0 = const()[name = tensor("attn_output_37_transpose_y_0"), val = tensor(false)]; + tensor values_113_cast_fp16 = transpose(perm = var_7894, x = values_111_cast_fp16)[name = tensor("transpose_134")]; + tensor attn_output_37_cast_fp16 = matmul(transpose_x = attn_output_37_transpose_x_0, transpose_y = attn_output_37_transpose_y_0, x = attn_weights_113_cast_fp16, y = values_113_cast_fp16)[name = tensor("attn_output_37_cast_fp16")]; + tensor var_7966 = const()[name = tensor("op_7966"), val = tensor([0, 2, 1, 3])]; + tensor var_7971 = const()[name = tensor("op_7971"), val = tensor([1, 256, 1024])]; + tensor var_7967_cast_fp16 = transpose(perm = var_7966, x = attn_output_37_cast_fp16)[name = tensor("transpose_131")]; + tensor input_221_cast_fp16 = reshape(shape = var_7971, x = var_7967_cast_fp16)[name = tensor("input_221_cast_fp16")]; + tensor attn18_out_proj_weight_to_fp16 = const()[name = tensor("attn18_out_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(460507776)))]; + tensor linear_73_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = attn18_out_proj_weight_to_fp16, x = input_221_cast_fp16)[name = tensor("linear_73_cast_fp16")]; + tensor var_7980_cast_fp16 = add(x = position18_to_fp16, y = valid_len_to_fp16)[name = tensor("op_7980_cast_fp16")]; + tensor var_7980_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_7980_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor input_223_cast_fp16 = add(x = input_217_cast_fp16, y = linear_73_cast_fp16)[name = tensor("input_223_cast_fp16")]; + tensor input_225_axes_0 = const()[name = tensor("input_225_axes_0"), val = tensor([-1])]; + tensor norm18_2_weight_to_fp16 = const()[name = tensor("norm18_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(462604992)))]; + tensor norm18_2_bias_to_fp16 = const()[name = tensor("norm18_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(462607104)))]; + tensor var_7984_to_fp16 = const()[name = tensor("op_7984_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_225_cast_fp16 = layer_norm(axes = input_225_axes_0, beta = norm18_2_bias_to_fp16, epsilon = var_7984_to_fp16, gamma = norm18_2_weight_to_fp16, x = input_223_cast_fp16)[name = tensor("input_225_cast_fp16")]; + tensor linear18_1_weight_to_fp16 = const()[name = tensor("linear18_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(462609216)))]; + tensor linear_74_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = linear18_1_weight_to_fp16, x = input_225_cast_fp16)[name = tensor("linear_74_cast_fp16")]; + tensor input_227_mode_0 = const()[name = tensor("input_227_mode_0"), val = tensor("EXACT")]; + tensor input_227_cast_fp16 = gelu(mode = input_227_mode_0, x = linear_74_cast_fp16)[name = tensor("input_227_cast_fp16")]; + tensor linear18_2_weight_to_fp16 = const()[name = tensor("linear18_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(470997888)))]; + tensor linear_75_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = linear18_2_weight_to_fp16, x = input_227_cast_fp16)[name = tensor("linear_75_cast_fp16")]; + tensor input_229_cast_fp16 = add(x = input_223_cast_fp16, y = linear_75_cast_fp16)[name = tensor("input_229_cast_fp16")]; + tensor input_231_axes_0 = const()[name = tensor("input_231_axes_0"), val = tensor([-1])]; + tensor norm19_1_weight_to_fp16 = const()[name = tensor("norm19_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(479386560)))]; + tensor norm19_1_bias_to_fp16 = const()[name = tensor("norm19_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(479388672)))]; + tensor var_8001_to_fp16 = const()[name = tensor("op_8001_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_231_cast_fp16 = layer_norm(axes = input_231_axes_0, beta = norm19_1_bias_to_fp16, epsilon = var_8001_to_fp16, gamma = norm19_1_weight_to_fp16, x = input_229_cast_fp16)[name = tensor("input_231_cast_fp16")]; + tensor attn19_in_proj_weight_to_fp16 = const()[name = tensor("attn19_in_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(479390784)))]; + tensor linear_76_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = attn19_in_proj_weight_to_fp16, x = input_231_cast_fp16)[name = tensor("linear_76_cast_fp16")]; + tensor var_8015 = const()[name = tensor("op_8015"), val = tensor([1, 256, 3, 16, 64])]; + tensor qkv_39_cast_fp16 = reshape(shape = var_8015, x = linear_76_cast_fp16)[name = tensor("qkv_39_cast_fp16")]; + tensor q_115_begin_0 = const()[name = tensor("q_115_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor q_115_end_0 = const()[name = tensor("q_115_end_0"), val = tensor([1, 256, 1, 16, 64])]; + tensor q_115_end_mask_0 = const()[name = tensor("q_115_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor q_115_squeeze_mask_0 = const()[name = tensor("q_115_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor q_115_cast_fp16 = slice_by_index(begin = q_115_begin_0, end = q_115_end_0, end_mask = q_115_end_mask_0, squeeze_mask = q_115_squeeze_mask_0, x = qkv_39_cast_fp16)[name = tensor("q_115_cast_fp16")]; + tensor k_77_begin_0 = const()[name = tensor("k_77_begin_0"), val = tensor([0, 0, 1, 0, 0])]; + tensor k_77_end_0 = const()[name = tensor("k_77_end_0"), val = tensor([1, 256, 2, 16, 64])]; + tensor k_77_end_mask_0 = const()[name = tensor("k_77_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor k_77_squeeze_mask_0 = const()[name = tensor("k_77_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor k_77_cast_fp16 = slice_by_index(begin = k_77_begin_0, end = k_77_end_0, end_mask = k_77_end_mask_0, squeeze_mask = k_77_squeeze_mask_0, x = qkv_39_cast_fp16)[name = tensor("k_77_cast_fp16")]; + tensor v_39_begin_0 = const()[name = tensor("v_39_begin_0"), val = tensor([0, 0, 2, 0, 0])]; + tensor v_39_end_0 = const()[name = tensor("v_39_end_0"), val = tensor([1, 256, 3, 16, 64])]; + tensor v_39_end_mask_0 = const()[name = tensor("v_39_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor v_39_squeeze_mask_0 = const()[name = tensor("v_39_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor v_39_cast_fp16 = slice_by_index(begin = v_39_begin_0, end = v_39_end_0, end_mask = v_39_end_mask_0, squeeze_mask = v_39_squeeze_mask_0, x = qkv_39_cast_fp16)[name = tensor("v_39_cast_fp16")]; + tensor var_8073 = const()[name = tensor("op_8073"), val = tensor([1, 1])]; + tensor position19_to_fp16_dtype_0 = const()[name = tensor("position19_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor position19_to_fp16 = cast(dtype = position19_to_fp16_dtype_0, x = position19)[name = tensor("cast_566")]; + tensor var_8074_cast_fp16 = reshape(shape = var_8073, x = position19_to_fp16)[name = tensor("op_8074_cast_fp16")]; + tensor ts_117_cast_fp16 = add(x = ts_1_promoted_to_fp16, y = var_8074_cast_fp16)[name = tensor("ts_117_cast_fp16")]; + tensor var_8081 = const()[name = tensor("op_8081"), val = tensor([1, 256, 1, 1])]; + tensor ts_119_cast_fp16 = reshape(shape = var_8081, x = ts_117_cast_fp16)[name = tensor("ts_119_cast_fp16")]; + tensor var_8088 = const()[name = tensor("op_8088"), val = tensor([1, 256, 16, 32, 2])]; + tensor q_complex_39_cast_fp16 = reshape(shape = var_8088, x = q_115_cast_fp16)[name = tensor("q_complex_39_cast_fp16")]; + tensor var_8095 = const()[name = tensor("op_8095"), val = tensor([1, 256, 16, 32, 2])]; + tensor k_complex_39_cast_fp16 = reshape(shape = var_8095, x = k_77_cast_fp16)[name = tensor("k_complex_39_cast_fp16")]; + tensor var_8099_begin_0 = const()[name = tensor("op_8099_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_8099_end_0 = const()[name = tensor("op_8099_end_0"), val = tensor([1, 256, 16, 32, 1])]; + tensor var_8099_end_mask_0 = const()[name = tensor("op_8099_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_8099_squeeze_mask_0 = const()[name = tensor("op_8099_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_8099_cast_fp16 = slice_by_index(begin = var_8099_begin_0, end = var_8099_end_0, end_mask = var_8099_end_mask_0, squeeze_mask = var_8099_squeeze_mask_0, x = q_complex_39_cast_fp16)[name = tensor("op_8099_cast_fp16")]; + tensor var_8107_begin_0 = const()[name = tensor("op_8107_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_8107_end_0 = const()[name = tensor("op_8107_end_0"), val = tensor([1, 256, 16, 32, 2])]; + tensor var_8107_end_mask_0 = const()[name = tensor("op_8107_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_8107_squeeze_mask_0 = const()[name = tensor("op_8107_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_8107_cast_fp16 = slice_by_index(begin = var_8107_begin_0, end = var_8107_end_0, end_mask = var_8107_end_mask_0, squeeze_mask = var_8107_squeeze_mask_0, x = q_complex_39_cast_fp16)[name = tensor("op_8107_cast_fp16")]; + tensor var_8115_begin_0 = const()[name = tensor("op_8115_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_8115_end_0 = const()[name = tensor("op_8115_end_0"), val = tensor([1, 256, 16, 32, 1])]; + tensor var_8115_end_mask_0 = const()[name = tensor("op_8115_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_8115_squeeze_mask_0 = const()[name = tensor("op_8115_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_8115_cast_fp16 = slice_by_index(begin = var_8115_begin_0, end = var_8115_end_0, end_mask = var_8115_end_mask_0, squeeze_mask = var_8115_squeeze_mask_0, x = k_complex_39_cast_fp16)[name = tensor("op_8115_cast_fp16")]; + tensor var_8123_begin_0 = const()[name = tensor("op_8123_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_8123_end_0 = const()[name = tensor("op_8123_end_0"), val = tensor([1, 256, 16, 32, 2])]; + tensor var_8123_end_mask_0 = const()[name = tensor("op_8123_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_8123_squeeze_mask_0 = const()[name = tensor("op_8123_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_8123_cast_fp16 = slice_by_index(begin = var_8123_begin_0, end = var_8123_end_0, end_mask = var_8123_end_mask_0, squeeze_mask = var_8123_squeeze_mask_0, x = k_complex_39_cast_fp16)[name = tensor("op_8123_cast_fp16")]; + tensor freqs_39_to_fp16 = const()[name = tensor("freqs_39_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(485682304)))]; + tensor var_8129_cast_fp16 = mul(x = freqs_39_to_fp16, y = ts_119_cast_fp16)[name = tensor("op_8129_cast_fp16")]; + tensor rotr_39_cast_fp16 = cos(x = var_8129_cast_fp16)[name = tensor("rotr_39_cast_fp16")]; + tensor roti_39_cast_fp16 = sin(x = var_8129_cast_fp16)[name = tensor("roti_39_cast_fp16")]; + tensor var_8133_cast_fp16 = mul(x = var_8099_cast_fp16, y = rotr_39_cast_fp16)[name = tensor("op_8133_cast_fp16")]; + tensor var_8134_cast_fp16 = mul(x = var_8107_cast_fp16, y = roti_39_cast_fp16)[name = tensor("op_8134_cast_fp16")]; + tensor qor_77_cast_fp16 = sub(x = var_8133_cast_fp16, y = var_8134_cast_fp16)[name = tensor("qor_77_cast_fp16")]; + tensor var_8137_cast_fp16 = mul(x = var_8099_cast_fp16, y = roti_39_cast_fp16)[name = tensor("op_8137_cast_fp16")]; + tensor var_8138_cast_fp16 = mul(x = var_8107_cast_fp16, y = rotr_39_cast_fp16)[name = tensor("op_8138_cast_fp16")]; + tensor qoi_77_cast_fp16 = add(x = var_8137_cast_fp16, y = var_8138_cast_fp16)[name = tensor("qoi_77_cast_fp16")]; + tensor var_8141_cast_fp16 = mul(x = var_8115_cast_fp16, y = rotr_39_cast_fp16)[name = tensor("op_8141_cast_fp16")]; + tensor var_8142_cast_fp16 = mul(x = var_8123_cast_fp16, y = roti_39_cast_fp16)[name = tensor("op_8142_cast_fp16")]; + tensor kor_77_cast_fp16 = sub(x = var_8141_cast_fp16, y = var_8142_cast_fp16)[name = tensor("kor_77_cast_fp16")]; + tensor var_8145_cast_fp16 = mul(x = var_8115_cast_fp16, y = roti_39_cast_fp16)[name = tensor("op_8145_cast_fp16")]; + tensor var_8146_cast_fp16 = mul(x = var_8123_cast_fp16, y = rotr_39_cast_fp16)[name = tensor("op_8146_cast_fp16")]; + tensor koi_77_cast_fp16 = add(x = var_8145_cast_fp16, y = var_8146_cast_fp16)[name = tensor("koi_77_cast_fp16")]; + tensor qo_39_axis_0 = const()[name = tensor("qo_39_axis_0"), val = tensor(-1)]; + tensor qo_39_cast_fp16 = stack(axis = qo_39_axis_0, values = (qor_77_cast_fp16, qoi_77_cast_fp16))[name = tensor("qo_39_cast_fp16")]; + tensor ko_39_axis_0 = const()[name = tensor("ko_39_axis_0"), val = tensor(-1)]; + tensor ko_39_cast_fp16 = stack(axis = ko_39_axis_0, values = (kor_77_cast_fp16, koi_77_cast_fp16))[name = tensor("ko_39_cast_fp16")]; + tensor var_8179 = const()[name = tensor("op_8179"), val = tensor([1, 256, 16, 64])]; + tensor q_117_cast_fp16 = reshape(shape = var_8179, x = qo_39_cast_fp16)[name = tensor("q_117_cast_fp16")]; + tensor var_8185 = const()[name = tensor("op_8185"), val = tensor([1, 256, 16, 64])]; + tensor k_79_cast_fp16 = reshape(shape = var_8185, x = ko_39_cast_fp16)[name = tensor("k_79_cast_fp16")]; + tensor write_indices_float_115_cast_fp16 = add(x = var_8074_cast_fp16, y = var_393_promoted_to_fp16)[name = tensor("write_indices_float_115_cast_fp16")]; + tensor _inversed_8206_y_0_to_fp16 = const()[name = tensor("_inversed_8206_y_0_to_fp16"), val = tensor(0x1p-9)]; + tensor _inversed_8206_cast_fp16 = mul(x = write_indices_float_115_cast_fp16, y = _inversed_8206_y_0_to_fp16)[name = tensor("_inversed_8206_cast_fp16")]; + tensor var_8207_cast_fp16 = floor(x = _inversed_8206_cast_fp16)[name = tensor("op_8207_cast_fp16")]; + tensor var_8208_to_fp16 = const()[name = tensor("op_8208_to_fp16"), val = tensor(0x1p+9)]; + tensor var_8209_cast_fp16 = mul(x = var_8207_cast_fp16, y = var_8208_to_fp16)[name = tensor("op_8209_cast_fp16")]; + tensor write_indices_float_117_cast_fp16 = sub(x = write_indices_float_115_cast_fp16, y = var_8209_cast_fp16)[name = tensor("write_indices_float_117_cast_fp16")]; + tensor var_8227_value_0_to_fp16 = const()[name = tensor("op_8227_value_0_to_fp16"), val = tensor(0x1.ffp+8)]; + tensor var_8227_cast_fp16 = fill_like(ref_tensor = write_indices_float_117_cast_fp16, value = var_8227_value_0_to_fp16)[name = tensor("op_8227_cast_fp16")]; + tensor write_indices_float_119_cast_fp16 = select(a = write_indices_float_117_cast_fp16, b = var_8227_cast_fp16, cond = token_valid_1_cast_fp16)[name = tensor("write_indices_float_119_cast_fp16")]; + tensor var_8233_dtype_0 = const()[name = tensor("op_8233_dtype_0"), val = tensor("int32")]; + tensor var_8238 = const()[name = tensor("op_8238"), val = tensor([1, 256, 1, 1])]; + tensor write_indices_float_119_cast_fp16_to_int32 = cast(dtype = var_8233_dtype_0, x = write_indices_float_119_cast_fp16)[name = tensor("cast_565")]; + tensor var_8239 = reshape(shape = var_8238, x = write_indices_float_119_cast_fp16_to_int32)[name = tensor("op_8239")]; + tensor write_indices_39_reps_0 = const()[name = tensor("write_indices_39_reps_0"), val = tensor([1, 1, 16, 64])]; + tensor write_indices_39 = tile(reps = write_indices_39_reps_0, x = var_8239)[name = tensor("write_indices_39")]; + tensor var_8249_begin_0 = const()[name = tensor("op_8249_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_8249_end_0 = const()[name = tensor("op_8249_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor var_8249_end_mask_0 = const()[name = tensor("op_8249_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_8249_squeeze_mask_0 = const()[name = tensor("op_8249_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor cache19_to_fp16_dtype_0 = const()[name = tensor("cache19_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor cache19_to_fp16 = cast(dtype = cache19_to_fp16_dtype_0, x = cache19)[name = tensor("cast_564")]; + tensor var_8249_cast_fp16 = slice_by_index(begin = var_8249_begin_0, end = var_8249_end_0, end_mask = var_8249_end_mask_0, squeeze_mask = var_8249_squeeze_mask_0, x = cache19_to_fp16)[name = tensor("op_8249_cast_fp16")]; + tensor var_8251_axis_0 = const()[name = tensor("op_8251_axis_0"), val = tensor(1)]; + tensor var_8251_mode_0 = const()[name = tensor("op_8251_mode_0"), val = tensor("update")]; + tensor var_8251_validate_indices_0 = const()[name = tensor("op_8251_validate_indices_0"), val = tensor(false)]; + tensor var_8251_cast_fp16 = scatter_along_axis(axis = var_8251_axis_0, data = var_8249_cast_fp16, indices = write_indices_39, mode = var_8251_mode_0, updates = k_79_cast_fp16, validate_indices = var_8251_validate_indices_0)[name = tensor("op_8251_cast_fp16")]; + tensor concat_154 = const()[name = tensor("concat_154"), val = tensor([0, 0, 0, 0, 0])]; + tensor concat_155 = const()[name = tensor("concat_155"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_39_internal_tensor_assign_1_stride_0 = const()[name = tensor("new_cache_39_internal_tensor_assign_1_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_39_internal_tensor_assign_1_begin_mask_0 = const()[name = tensor("new_cache_39_internal_tensor_assign_1_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_39_internal_tensor_assign_1_end_mask_0 = const()[name = tensor("new_cache_39_internal_tensor_assign_1_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_39_internal_tensor_assign_1_squeeze_mask_0 = const()[name = tensor("new_cache_39_internal_tensor_assign_1_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_84 = const()[name = tensor("shape_84"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_38 = const()[name = tensor("reduce_prod_38"), val = tensor(1048576)]; + tensor range_1d_38_start_0 = const()[name = tensor("range_1d_38_start_0"), val = tensor(0)]; + tensor range_1d_38_step_0 = const()[name = tensor("range_1d_38_step_0"), val = tensor(1)]; + tensor range_1d_38 = range_1d(end = reduce_prod_38, start = range_1d_38_start_0, step = range_1d_38_step_0)[name = tensor("range_1d_38")]; + tensor reshape_190 = reshape(shape = shape_84, x = range_1d_38)[name = tensor("reshape_190")]; + tensor slice_by_index_38 = slice_by_index(begin = concat_154, begin_mask = new_cache_39_internal_tensor_assign_1_begin_mask_0, end = concat_155, end_mask = new_cache_39_internal_tensor_assign_1_end_mask_0, squeeze_mask = new_cache_39_internal_tensor_assign_1_squeeze_mask_0, stride = new_cache_39_internal_tensor_assign_1_stride_0, x = reshape_190)[name = tensor("slice_by_index_38")]; + tensor reshape_191_shape_0 = const()[name = tensor("reshape_191_shape_0"), val = tensor([-1])]; + tensor reshape_191 = reshape(shape = reshape_191_shape_0, x = slice_by_index_38)[name = tensor("reshape_191")]; + tensor reshape_192_shape_0 = const()[name = tensor("reshape_192_shape_0"), val = tensor([-1])]; + tensor reshape_192_cast_fp16 = reshape(shape = reshape_192_shape_0, x = var_8251_cast_fp16)[name = tensor("reshape_192_cast_fp16")]; + tensor reshape_193_shape_0 = const()[name = tensor("reshape_193_shape_0"), val = tensor([-1])]; + tensor reshape_193_cast_fp16 = reshape(shape = reshape_193_shape_0, x = cache19_to_fp16)[name = tensor("reshape_193_cast_fp16")]; + tensor scatter_38_mode_0 = const()[name = tensor("scatter_38_mode_0"), val = tensor("update")]; + tensor scatter_38_axis_0 = const()[name = tensor("scatter_38_axis_0"), val = tensor(0)]; + tensor scatter_38_validate_indices_0 = const()[name = tensor("scatter_38_validate_indices_0"), val = tensor(false)]; + tensor scatter_38_cast_fp16 = scatter(axis = scatter_38_axis_0, data = reshape_193_cast_fp16, indices = reshape_191, mode = scatter_38_mode_0, updates = reshape_192_cast_fp16, validate_indices = scatter_38_validate_indices_0)[name = tensor("scatter_38_cast_fp16")]; + tensor reshape_194_cast_fp16 = reshape(shape = shape_84, x = scatter_38_cast_fp16)[name = tensor("reshape_194_cast_fp16")]; + tensor var_8259_begin_0 = const()[name = tensor("op_8259_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor var_8259_end_0 = const()[name = tensor("op_8259_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor var_8259_end_mask_0 = const()[name = tensor("op_8259_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_8259_squeeze_mask_0 = const()[name = tensor("op_8259_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor var_8259_cast_fp16 = slice_by_index(begin = var_8259_begin_0, end = var_8259_end_0, end_mask = var_8259_end_mask_0, squeeze_mask = var_8259_squeeze_mask_0, x = reshape_194_cast_fp16)[name = tensor("op_8259_cast_fp16")]; + tensor var_8261_axis_0 = const()[name = tensor("op_8261_axis_0"), val = tensor(1)]; + tensor var_8261_mode_0 = const()[name = tensor("op_8261_mode_0"), val = tensor("update")]; + tensor var_8261_validate_indices_0 = const()[name = tensor("op_8261_validate_indices_0"), val = tensor(false)]; + tensor var_8261_cast_fp16 = scatter_along_axis(axis = var_8261_axis_0, data = var_8259_cast_fp16, indices = write_indices_39, mode = var_8261_mode_0, updates = v_39_cast_fp16, validate_indices = var_8261_validate_indices_0)[name = tensor("op_8261_cast_fp16")]; + tensor concat_156 = const()[name = tensor("concat_156"), val = tensor([1, 0, 0, 0, 0])]; + tensor concat_157 = const()[name = tensor("concat_157"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_39_internal_tensor_assign_2_stride_0 = const()[name = tensor("new_cache_39_internal_tensor_assign_2_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_39_internal_tensor_assign_2_begin_mask_0 = const()[name = tensor("new_cache_39_internal_tensor_assign_2_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_39_internal_tensor_assign_2_end_mask_0 = const()[name = tensor("new_cache_39_internal_tensor_assign_2_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_39_internal_tensor_assign_2_squeeze_mask_0 = const()[name = tensor("new_cache_39_internal_tensor_assign_2_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_85 = const()[name = tensor("shape_85"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_39 = const()[name = tensor("reduce_prod_39"), val = tensor(1048576)]; + tensor range_1d_39_start_0 = const()[name = tensor("range_1d_39_start_0"), val = tensor(0)]; + tensor range_1d_39_step_0 = const()[name = tensor("range_1d_39_step_0"), val = tensor(1)]; + tensor range_1d_39 = range_1d(end = reduce_prod_39, start = range_1d_39_start_0, step = range_1d_39_step_0)[name = tensor("range_1d_39")]; + tensor reshape_195 = reshape(shape = shape_85, x = range_1d_39)[name = tensor("reshape_195")]; + tensor slice_by_index_39 = slice_by_index(begin = concat_156, begin_mask = new_cache_39_internal_tensor_assign_2_begin_mask_0, end = concat_157, end_mask = new_cache_39_internal_tensor_assign_2_end_mask_0, squeeze_mask = new_cache_39_internal_tensor_assign_2_squeeze_mask_0, stride = new_cache_39_internal_tensor_assign_2_stride_0, x = reshape_195)[name = tensor("slice_by_index_39")]; + tensor reshape_196_shape_0 = const()[name = tensor("reshape_196_shape_0"), val = tensor([-1])]; + tensor reshape_196 = reshape(shape = reshape_196_shape_0, x = slice_by_index_39)[name = tensor("reshape_196")]; + tensor reshape_197_shape_0 = const()[name = tensor("reshape_197_shape_0"), val = tensor([-1])]; + tensor reshape_197_cast_fp16 = reshape(shape = reshape_197_shape_0, x = var_8261_cast_fp16)[name = tensor("reshape_197_cast_fp16")]; + tensor reshape_198_shape_0 = const()[name = tensor("reshape_198_shape_0"), val = tensor([-1])]; + tensor reshape_198_cast_fp16 = reshape(shape = reshape_198_shape_0, x = reshape_194_cast_fp16)[name = tensor("reshape_198_cast_fp16")]; + tensor scatter_39_mode_0 = const()[name = tensor("scatter_39_mode_0"), val = tensor("update")]; + tensor scatter_39_axis_0 = const()[name = tensor("scatter_39_axis_0"), val = tensor(0)]; + tensor scatter_39_validate_indices_0 = const()[name = tensor("scatter_39_validate_indices_0"), val = tensor(false)]; + tensor scatter_39_cast_fp16 = scatter(axis = scatter_39_axis_0, data = reshape_198_cast_fp16, indices = reshape_196, mode = scatter_39_mode_0, updates = reshape_197_cast_fp16, validate_indices = scatter_39_validate_indices_0)[name = tensor("scatter_39_cast_fp16")]; + tensor reshape_199_cast_fp16 = reshape(shape = shape_85, x = scatter_39_cast_fp16)[name = tensor("reshape_199_cast_fp16")]; + tensor reshape_199_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("reshape_199_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor keys_115_begin_0 = const()[name = tensor("keys_115_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor keys_115_end_0 = const()[name = tensor("keys_115_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor keys_115_end_mask_0 = const()[name = tensor("keys_115_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor keys_115_squeeze_mask_0 = const()[name = tensor("keys_115_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor keys_115_cast_fp16 = slice_by_index(begin = keys_115_begin_0, end = keys_115_end_0, end_mask = keys_115_end_mask_0, squeeze_mask = keys_115_squeeze_mask_0, x = reshape_199_cast_fp16)[name = tensor("keys_115_cast_fp16")]; + tensor values_115_begin_0 = const()[name = tensor("values_115_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor values_115_end_0 = const()[name = tensor("values_115_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor values_115_end_mask_0 = const()[name = tensor("values_115_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor values_115_squeeze_mask_0 = const()[name = tensor("values_115_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor values_115_cast_fp16 = slice_by_index(begin = values_115_begin_0, end = values_115_end_0, end_mask = values_115_end_mask_0, squeeze_mask = values_115_squeeze_mask_0, x = reshape_199_cast_fp16)[name = tensor("values_115_cast_fp16")]; + tensor var_8273_cast_fp16 = not_equal(x = keys_115_cast_fp16, y = keys_115_cast_fp16)[name = tensor("op_8273_cast_fp16")]; + tensor keys_117_cast_fp16 = select(a = var_470_to_fp16, b = keys_115_cast_fp16, cond = var_8273_cast_fp16)[name = tensor("keys_117_cast_fp16")]; + tensor var_8281_cast_fp16 = not_equal(x = values_115_cast_fp16, y = values_115_cast_fp16)[name = tensor("op_8281_cast_fp16")]; + tensor values_117_cast_fp16 = select(a = var_470_to_fp16, b = values_115_cast_fp16, cond = var_8281_cast_fp16)[name = tensor("values_117_cast_fp16")]; + tensor var_8305 = const()[name = tensor("op_8305"), val = tensor([0, 2, 1, 3])]; + tensor var_8321 = const()[name = tensor("op_8321"), val = tensor([1, 1, 1])]; + tensor var_8322_cast_fp16 = reshape(shape = var_8321, x = position19_to_fp16)[name = tensor("op_8322_cast_fp16")]; + tensor q_positions_39_cast_fp16 = add(x = var_8322_cast_fp16, y = q_offsets_1_promoted_to_fp16)[name = tensor("q_positions_39_cast_fp16")]; + tensor valid_end_39_cast_fp16 = add(x = var_8322_cast_fp16, y = var_536_cast_fp16)[name = tensor("valid_end_39_cast_fp16")]; + tensor valid_mask_39_cast_fp16 = less(x = k_positions_1_promoted_to_fp16, y = valid_end_39_cast_fp16)[name = tensor("valid_mask_39_cast_fp16")]; + tensor causal_mask_39_cast_fp16 = less_equal(x = k_positions_1_promoted_to_fp16, y = q_positions_39_cast_fp16)[name = tensor("causal_mask_39_cast_fp16")]; + tensor var_8350 = logical_and(x = valid_mask_39_cast_fp16, y = causal_mask_39_cast_fp16)[name = tensor("op_8350")]; + tensor attn_mask_39_axes_0 = const()[name = tensor("attn_mask_39_axes_0"), val = tensor([1])]; + tensor attn_mask_39 = expand_dims(axes = attn_mask_39_axes_0, x = var_8350)[name = tensor("attn_mask_39")]; + tensor var_8364_transpose_x_0 = const()[name = tensor("op_8364_transpose_x_0"), val = tensor(false)]; + tensor var_8364_transpose_y_0 = const()[name = tensor("op_8364_transpose_y_0"), val = tensor(false)]; + tensor transpose_107_perm_0 = const()[name = tensor("transpose_107_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_108_perm_0 = const()[name = tensor("transpose_108_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_108 = transpose(perm = transpose_108_perm_0, x = keys_117_cast_fp16)[name = tensor("transpose_128")]; + tensor transpose_107 = transpose(perm = transpose_107_perm_0, x = q_117_cast_fp16)[name = tensor("transpose_129")]; + tensor var_8364_cast_fp16 = matmul(transpose_x = var_8364_transpose_x_0, transpose_y = var_8364_transpose_y_0, x = transpose_107, y = transpose_108)[name = tensor("op_8364_cast_fp16")]; + tensor var_8358_to_fp16 = const()[name = tensor("op_8358_to_fp16"), val = tensor([0x1p-3])]; + tensor attn_weights_115_cast_fp16 = mul(x = var_8364_cast_fp16, y = var_8358_to_fp16)[name = tensor("attn_weights_115_cast_fp16")]; + tensor var_8366 = logical_not(x = attn_mask_39)[name = tensor("op_8366")]; + tensor var_8367_to_fp16 = const()[name = tensor("op_8367_to_fp16"), val = tensor(-inf)]; + tensor attn_weights_117_cast_fp16 = select(a = var_8367_to_fp16, b = attn_weights_115_cast_fp16, cond = var_8366)[name = tensor("attn_weights_117_cast_fp16")]; + tensor var_8369 = const()[name = tensor("op_8369"), val = tensor(-1)]; + tensor attn_weights_119_cast_fp16 = softmax(axis = var_8369, x = attn_weights_117_cast_fp16)[name = tensor("attn_weights_119_cast_fp16")]; + tensor attn_output_39_transpose_x_0 = const()[name = tensor("attn_output_39_transpose_x_0"), val = tensor(false)]; + tensor attn_output_39_transpose_y_0 = const()[name = tensor("attn_output_39_transpose_y_0"), val = tensor(false)]; + tensor values_119_cast_fp16 = transpose(perm = var_8305, x = values_117_cast_fp16)[name = tensor("transpose_130")]; + tensor attn_output_39_cast_fp16 = matmul(transpose_x = attn_output_39_transpose_x_0, transpose_y = attn_output_39_transpose_y_0, x = attn_weights_119_cast_fp16, y = values_119_cast_fp16)[name = tensor("attn_output_39_cast_fp16")]; + tensor var_8377 = const()[name = tensor("op_8377"), val = tensor([0, 2, 1, 3])]; + tensor var_8382 = const()[name = tensor("op_8382"), val = tensor([1, 256, 1024])]; + tensor var_8378_cast_fp16 = transpose(perm = var_8377, x = attn_output_39_cast_fp16)[name = tensor("transpose_127")]; + tensor input_233_cast_fp16 = reshape(shape = var_8382, x = var_8378_cast_fp16)[name = tensor("input_233_cast_fp16")]; + tensor attn19_out_proj_weight_to_fp16 = const()[name = tensor("attn19_out_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(485682432)))]; + tensor linear_77_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = attn19_out_proj_weight_to_fp16, x = input_233_cast_fp16)[name = tensor("linear_77_cast_fp16")]; + tensor var_8391_cast_fp16 = add(x = position19_to_fp16, y = valid_len_to_fp16)[name = tensor("op_8391_cast_fp16")]; + tensor var_8391_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_8391_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor input_235_cast_fp16 = add(x = input_229_cast_fp16, y = linear_77_cast_fp16)[name = tensor("input_235_cast_fp16")]; + tensor input_237_axes_0 = const()[name = tensor("input_237_axes_0"), val = tensor([-1])]; + tensor norm19_2_weight_to_fp16 = const()[name = tensor("norm19_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(487779648)))]; + tensor norm19_2_bias_to_fp16 = const()[name = tensor("norm19_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(487781760)))]; + tensor var_8395_to_fp16 = const()[name = tensor("op_8395_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_237_cast_fp16 = layer_norm(axes = input_237_axes_0, beta = norm19_2_bias_to_fp16, epsilon = var_8395_to_fp16, gamma = norm19_2_weight_to_fp16, x = input_235_cast_fp16)[name = tensor("input_237_cast_fp16")]; + tensor linear19_1_weight_to_fp16 = const()[name = tensor("linear19_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(487783872)))]; + tensor linear_78_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = linear19_1_weight_to_fp16, x = input_237_cast_fp16)[name = tensor("linear_78_cast_fp16")]; + tensor input_239_mode_0 = const()[name = tensor("input_239_mode_0"), val = tensor("EXACT")]; + tensor input_239_cast_fp16 = gelu(mode = input_239_mode_0, x = linear_78_cast_fp16)[name = tensor("input_239_cast_fp16")]; + tensor linear19_2_weight_to_fp16 = const()[name = tensor("linear19_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(496172544)))]; + tensor linear_79_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = linear19_2_weight_to_fp16, x = input_239_cast_fp16)[name = tensor("linear_79_cast_fp16")]; + tensor input_241_cast_fp16 = add(x = input_235_cast_fp16, y = linear_79_cast_fp16)[name = tensor("input_241_cast_fp16")]; + tensor input_243_axes_0 = const()[name = tensor("input_243_axes_0"), val = tensor([-1])]; + tensor norm20_1_weight_to_fp16 = const()[name = tensor("norm20_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(504561216)))]; + tensor norm20_1_bias_to_fp16 = const()[name = tensor("norm20_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(504563328)))]; + tensor var_8412_to_fp16 = const()[name = tensor("op_8412_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_243_cast_fp16 = layer_norm(axes = input_243_axes_0, beta = norm20_1_bias_to_fp16, epsilon = var_8412_to_fp16, gamma = norm20_1_weight_to_fp16, x = input_241_cast_fp16)[name = tensor("input_243_cast_fp16")]; + tensor attn20_in_proj_weight_to_fp16 = const()[name = tensor("attn20_in_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(504565440)))]; + tensor linear_80_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = attn20_in_proj_weight_to_fp16, x = input_243_cast_fp16)[name = tensor("linear_80_cast_fp16")]; + tensor var_8426 = const()[name = tensor("op_8426"), val = tensor([1, 256, 3, 16, 64])]; + tensor qkv_41_cast_fp16 = reshape(shape = var_8426, x = linear_80_cast_fp16)[name = tensor("qkv_41_cast_fp16")]; + tensor q_121_begin_0 = const()[name = tensor("q_121_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor q_121_end_0 = const()[name = tensor("q_121_end_0"), val = tensor([1, 256, 1, 16, 64])]; + tensor q_121_end_mask_0 = const()[name = tensor("q_121_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor q_121_squeeze_mask_0 = const()[name = tensor("q_121_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor q_121_cast_fp16 = slice_by_index(begin = q_121_begin_0, end = q_121_end_0, end_mask = q_121_end_mask_0, squeeze_mask = q_121_squeeze_mask_0, x = qkv_41_cast_fp16)[name = tensor("q_121_cast_fp16")]; + tensor k_81_begin_0 = const()[name = tensor("k_81_begin_0"), val = tensor([0, 0, 1, 0, 0])]; + tensor k_81_end_0 = const()[name = tensor("k_81_end_0"), val = tensor([1, 256, 2, 16, 64])]; + tensor k_81_end_mask_0 = const()[name = tensor("k_81_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor k_81_squeeze_mask_0 = const()[name = tensor("k_81_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor k_81_cast_fp16 = slice_by_index(begin = k_81_begin_0, end = k_81_end_0, end_mask = k_81_end_mask_0, squeeze_mask = k_81_squeeze_mask_0, x = qkv_41_cast_fp16)[name = tensor("k_81_cast_fp16")]; + tensor v_41_begin_0 = const()[name = tensor("v_41_begin_0"), val = tensor([0, 0, 2, 0, 0])]; + tensor v_41_end_0 = const()[name = tensor("v_41_end_0"), val = tensor([1, 256, 3, 16, 64])]; + tensor v_41_end_mask_0 = const()[name = tensor("v_41_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor v_41_squeeze_mask_0 = const()[name = tensor("v_41_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor v_41_cast_fp16 = slice_by_index(begin = v_41_begin_0, end = v_41_end_0, end_mask = v_41_end_mask_0, squeeze_mask = v_41_squeeze_mask_0, x = qkv_41_cast_fp16)[name = tensor("v_41_cast_fp16")]; + tensor var_8484 = const()[name = tensor("op_8484"), val = tensor([1, 1])]; + tensor position20_to_fp16_dtype_0 = const()[name = tensor("position20_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor position20_to_fp16 = cast(dtype = position20_to_fp16_dtype_0, x = position20)[name = tensor("cast_561")]; + tensor var_8485_cast_fp16 = reshape(shape = var_8484, x = position20_to_fp16)[name = tensor("op_8485_cast_fp16")]; + tensor ts_123_cast_fp16 = add(x = ts_1_promoted_to_fp16, y = var_8485_cast_fp16)[name = tensor("ts_123_cast_fp16")]; + tensor var_8492 = const()[name = tensor("op_8492"), val = tensor([1, 256, 1, 1])]; + tensor ts_125_cast_fp16 = reshape(shape = var_8492, x = ts_123_cast_fp16)[name = tensor("ts_125_cast_fp16")]; + tensor var_8499 = const()[name = tensor("op_8499"), val = tensor([1, 256, 16, 32, 2])]; + tensor q_complex_41_cast_fp16 = reshape(shape = var_8499, x = q_121_cast_fp16)[name = tensor("q_complex_41_cast_fp16")]; + tensor var_8506 = const()[name = tensor("op_8506"), val = tensor([1, 256, 16, 32, 2])]; + tensor k_complex_41_cast_fp16 = reshape(shape = var_8506, x = k_81_cast_fp16)[name = tensor("k_complex_41_cast_fp16")]; + tensor var_8510_begin_0 = const()[name = tensor("op_8510_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_8510_end_0 = const()[name = tensor("op_8510_end_0"), val = tensor([1, 256, 16, 32, 1])]; + tensor var_8510_end_mask_0 = const()[name = tensor("op_8510_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_8510_squeeze_mask_0 = const()[name = tensor("op_8510_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_8510_cast_fp16 = slice_by_index(begin = var_8510_begin_0, end = var_8510_end_0, end_mask = var_8510_end_mask_0, squeeze_mask = var_8510_squeeze_mask_0, x = q_complex_41_cast_fp16)[name = tensor("op_8510_cast_fp16")]; + tensor var_8518_begin_0 = const()[name = tensor("op_8518_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_8518_end_0 = const()[name = tensor("op_8518_end_0"), val = tensor([1, 256, 16, 32, 2])]; + tensor var_8518_end_mask_0 = const()[name = tensor("op_8518_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_8518_squeeze_mask_0 = const()[name = tensor("op_8518_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_8518_cast_fp16 = slice_by_index(begin = var_8518_begin_0, end = var_8518_end_0, end_mask = var_8518_end_mask_0, squeeze_mask = var_8518_squeeze_mask_0, x = q_complex_41_cast_fp16)[name = tensor("op_8518_cast_fp16")]; + tensor var_8526_begin_0 = const()[name = tensor("op_8526_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_8526_end_0 = const()[name = tensor("op_8526_end_0"), val = tensor([1, 256, 16, 32, 1])]; + tensor var_8526_end_mask_0 = const()[name = tensor("op_8526_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_8526_squeeze_mask_0 = const()[name = tensor("op_8526_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_8526_cast_fp16 = slice_by_index(begin = var_8526_begin_0, end = var_8526_end_0, end_mask = var_8526_end_mask_0, squeeze_mask = var_8526_squeeze_mask_0, x = k_complex_41_cast_fp16)[name = tensor("op_8526_cast_fp16")]; + tensor var_8534_begin_0 = const()[name = tensor("op_8534_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_8534_end_0 = const()[name = tensor("op_8534_end_0"), val = tensor([1, 256, 16, 32, 2])]; + tensor var_8534_end_mask_0 = const()[name = tensor("op_8534_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_8534_squeeze_mask_0 = const()[name = tensor("op_8534_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_8534_cast_fp16 = slice_by_index(begin = var_8534_begin_0, end = var_8534_end_0, end_mask = var_8534_end_mask_0, squeeze_mask = var_8534_squeeze_mask_0, x = k_complex_41_cast_fp16)[name = tensor("op_8534_cast_fp16")]; + tensor freqs_41_to_fp16 = const()[name = tensor("freqs_41_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(510856960)))]; + tensor var_8540_cast_fp16 = mul(x = freqs_41_to_fp16, y = ts_125_cast_fp16)[name = tensor("op_8540_cast_fp16")]; + tensor rotr_41_cast_fp16 = cos(x = var_8540_cast_fp16)[name = tensor("rotr_41_cast_fp16")]; + tensor roti_41_cast_fp16 = sin(x = var_8540_cast_fp16)[name = tensor("roti_41_cast_fp16")]; + tensor var_8544_cast_fp16 = mul(x = var_8510_cast_fp16, y = rotr_41_cast_fp16)[name = tensor("op_8544_cast_fp16")]; + tensor var_8545_cast_fp16 = mul(x = var_8518_cast_fp16, y = roti_41_cast_fp16)[name = tensor("op_8545_cast_fp16")]; + tensor qor_81_cast_fp16 = sub(x = var_8544_cast_fp16, y = var_8545_cast_fp16)[name = tensor("qor_81_cast_fp16")]; + tensor var_8548_cast_fp16 = mul(x = var_8510_cast_fp16, y = roti_41_cast_fp16)[name = tensor("op_8548_cast_fp16")]; + tensor var_8549_cast_fp16 = mul(x = var_8518_cast_fp16, y = rotr_41_cast_fp16)[name = tensor("op_8549_cast_fp16")]; + tensor qoi_81_cast_fp16 = add(x = var_8548_cast_fp16, y = var_8549_cast_fp16)[name = tensor("qoi_81_cast_fp16")]; + tensor var_8552_cast_fp16 = mul(x = var_8526_cast_fp16, y = rotr_41_cast_fp16)[name = tensor("op_8552_cast_fp16")]; + tensor var_8553_cast_fp16 = mul(x = var_8534_cast_fp16, y = roti_41_cast_fp16)[name = tensor("op_8553_cast_fp16")]; + tensor kor_81_cast_fp16 = sub(x = var_8552_cast_fp16, y = var_8553_cast_fp16)[name = tensor("kor_81_cast_fp16")]; + tensor var_8556_cast_fp16 = mul(x = var_8526_cast_fp16, y = roti_41_cast_fp16)[name = tensor("op_8556_cast_fp16")]; + tensor var_8557_cast_fp16 = mul(x = var_8534_cast_fp16, y = rotr_41_cast_fp16)[name = tensor("op_8557_cast_fp16")]; + tensor koi_81_cast_fp16 = add(x = var_8556_cast_fp16, y = var_8557_cast_fp16)[name = tensor("koi_81_cast_fp16")]; + tensor qo_41_axis_0 = const()[name = tensor("qo_41_axis_0"), val = tensor(-1)]; + tensor qo_41_cast_fp16 = stack(axis = qo_41_axis_0, values = (qor_81_cast_fp16, qoi_81_cast_fp16))[name = tensor("qo_41_cast_fp16")]; + tensor ko_41_axis_0 = const()[name = tensor("ko_41_axis_0"), val = tensor(-1)]; + tensor ko_41_cast_fp16 = stack(axis = ko_41_axis_0, values = (kor_81_cast_fp16, koi_81_cast_fp16))[name = tensor("ko_41_cast_fp16")]; + tensor var_8590 = const()[name = tensor("op_8590"), val = tensor([1, 256, 16, 64])]; + tensor q_123_cast_fp16 = reshape(shape = var_8590, x = qo_41_cast_fp16)[name = tensor("q_123_cast_fp16")]; + tensor var_8596 = const()[name = tensor("op_8596"), val = tensor([1, 256, 16, 64])]; + tensor k_83_cast_fp16 = reshape(shape = var_8596, x = ko_41_cast_fp16)[name = tensor("k_83_cast_fp16")]; + tensor write_indices_float_121_cast_fp16 = add(x = var_8485_cast_fp16, y = var_393_promoted_to_fp16)[name = tensor("write_indices_float_121_cast_fp16")]; + tensor _inversed_8617_y_0_to_fp16 = const()[name = tensor("_inversed_8617_y_0_to_fp16"), val = tensor(0x1p-9)]; + tensor _inversed_8617_cast_fp16 = mul(x = write_indices_float_121_cast_fp16, y = _inversed_8617_y_0_to_fp16)[name = tensor("_inversed_8617_cast_fp16")]; + tensor var_8618_cast_fp16 = floor(x = _inversed_8617_cast_fp16)[name = tensor("op_8618_cast_fp16")]; + tensor var_8619_to_fp16 = const()[name = tensor("op_8619_to_fp16"), val = tensor(0x1p+9)]; + tensor var_8620_cast_fp16 = mul(x = var_8618_cast_fp16, y = var_8619_to_fp16)[name = tensor("op_8620_cast_fp16")]; + tensor write_indices_float_123_cast_fp16 = sub(x = write_indices_float_121_cast_fp16, y = var_8620_cast_fp16)[name = tensor("write_indices_float_123_cast_fp16")]; + tensor var_8638_value_0_to_fp16 = const()[name = tensor("op_8638_value_0_to_fp16"), val = tensor(0x1.ffp+8)]; + tensor var_8638_cast_fp16 = fill_like(ref_tensor = write_indices_float_123_cast_fp16, value = var_8638_value_0_to_fp16)[name = tensor("op_8638_cast_fp16")]; + tensor write_indices_float_125_cast_fp16 = select(a = write_indices_float_123_cast_fp16, b = var_8638_cast_fp16, cond = token_valid_1_cast_fp16)[name = tensor("write_indices_float_125_cast_fp16")]; + tensor var_8644_dtype_0 = const()[name = tensor("op_8644_dtype_0"), val = tensor("int32")]; + tensor var_8649 = const()[name = tensor("op_8649"), val = tensor([1, 256, 1, 1])]; + tensor write_indices_float_125_cast_fp16_to_int32 = cast(dtype = var_8644_dtype_0, x = write_indices_float_125_cast_fp16)[name = tensor("cast_560")]; + tensor var_8650 = reshape(shape = var_8649, x = write_indices_float_125_cast_fp16_to_int32)[name = tensor("op_8650")]; + tensor write_indices_41_reps_0 = const()[name = tensor("write_indices_41_reps_0"), val = tensor([1, 1, 16, 64])]; + tensor write_indices_41 = tile(reps = write_indices_41_reps_0, x = var_8650)[name = tensor("write_indices_41")]; + tensor var_8660_begin_0 = const()[name = tensor("op_8660_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_8660_end_0 = const()[name = tensor("op_8660_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor var_8660_end_mask_0 = const()[name = tensor("op_8660_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_8660_squeeze_mask_0 = const()[name = tensor("op_8660_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor cache20_to_fp16_dtype_0 = const()[name = tensor("cache20_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor cache20_to_fp16 = cast(dtype = cache20_to_fp16_dtype_0, x = cache20)[name = tensor("cast_559")]; + tensor var_8660_cast_fp16 = slice_by_index(begin = var_8660_begin_0, end = var_8660_end_0, end_mask = var_8660_end_mask_0, squeeze_mask = var_8660_squeeze_mask_0, x = cache20_to_fp16)[name = tensor("op_8660_cast_fp16")]; + tensor var_8662_axis_0 = const()[name = tensor("op_8662_axis_0"), val = tensor(1)]; + tensor var_8662_mode_0 = const()[name = tensor("op_8662_mode_0"), val = tensor("update")]; + tensor var_8662_validate_indices_0 = const()[name = tensor("op_8662_validate_indices_0"), val = tensor(false)]; + tensor var_8662_cast_fp16 = scatter_along_axis(axis = var_8662_axis_0, data = var_8660_cast_fp16, indices = write_indices_41, mode = var_8662_mode_0, updates = k_83_cast_fp16, validate_indices = var_8662_validate_indices_0)[name = tensor("op_8662_cast_fp16")]; + tensor concat_162 = const()[name = tensor("concat_162"), val = tensor([0, 0, 0, 0, 0])]; + tensor concat_163 = const()[name = tensor("concat_163"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_41_internal_tensor_assign_1_stride_0 = const()[name = tensor("new_cache_41_internal_tensor_assign_1_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_41_internal_tensor_assign_1_begin_mask_0 = const()[name = tensor("new_cache_41_internal_tensor_assign_1_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_41_internal_tensor_assign_1_end_mask_0 = const()[name = tensor("new_cache_41_internal_tensor_assign_1_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_41_internal_tensor_assign_1_squeeze_mask_0 = const()[name = tensor("new_cache_41_internal_tensor_assign_1_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_86 = const()[name = tensor("shape_86"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_40 = const()[name = tensor("reduce_prod_40"), val = tensor(1048576)]; + tensor range_1d_40_start_0 = const()[name = tensor("range_1d_40_start_0"), val = tensor(0)]; + tensor range_1d_40_step_0 = const()[name = tensor("range_1d_40_step_0"), val = tensor(1)]; + tensor range_1d_40 = range_1d(end = reduce_prod_40, start = range_1d_40_start_0, step = range_1d_40_step_0)[name = tensor("range_1d_40")]; + tensor reshape_200 = reshape(shape = shape_86, x = range_1d_40)[name = tensor("reshape_200")]; + tensor slice_by_index_40 = slice_by_index(begin = concat_162, begin_mask = new_cache_41_internal_tensor_assign_1_begin_mask_0, end = concat_163, end_mask = new_cache_41_internal_tensor_assign_1_end_mask_0, squeeze_mask = new_cache_41_internal_tensor_assign_1_squeeze_mask_0, stride = new_cache_41_internal_tensor_assign_1_stride_0, x = reshape_200)[name = tensor("slice_by_index_40")]; + tensor reshape_201_shape_0 = const()[name = tensor("reshape_201_shape_0"), val = tensor([-1])]; + tensor reshape_201 = reshape(shape = reshape_201_shape_0, x = slice_by_index_40)[name = tensor("reshape_201")]; + tensor reshape_202_shape_0 = const()[name = tensor("reshape_202_shape_0"), val = tensor([-1])]; + tensor reshape_202_cast_fp16 = reshape(shape = reshape_202_shape_0, x = var_8662_cast_fp16)[name = tensor("reshape_202_cast_fp16")]; + tensor reshape_203_shape_0 = const()[name = tensor("reshape_203_shape_0"), val = tensor([-1])]; + tensor reshape_203_cast_fp16 = reshape(shape = reshape_203_shape_0, x = cache20_to_fp16)[name = tensor("reshape_203_cast_fp16")]; + tensor scatter_40_mode_0 = const()[name = tensor("scatter_40_mode_0"), val = tensor("update")]; + tensor scatter_40_axis_0 = const()[name = tensor("scatter_40_axis_0"), val = tensor(0)]; + tensor scatter_40_validate_indices_0 = const()[name = tensor("scatter_40_validate_indices_0"), val = tensor(false)]; + tensor scatter_40_cast_fp16 = scatter(axis = scatter_40_axis_0, data = reshape_203_cast_fp16, indices = reshape_201, mode = scatter_40_mode_0, updates = reshape_202_cast_fp16, validate_indices = scatter_40_validate_indices_0)[name = tensor("scatter_40_cast_fp16")]; + tensor reshape_204_cast_fp16 = reshape(shape = shape_86, x = scatter_40_cast_fp16)[name = tensor("reshape_204_cast_fp16")]; + tensor var_8670_begin_0 = const()[name = tensor("op_8670_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor var_8670_end_0 = const()[name = tensor("op_8670_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor var_8670_end_mask_0 = const()[name = tensor("op_8670_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_8670_squeeze_mask_0 = const()[name = tensor("op_8670_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor var_8670_cast_fp16 = slice_by_index(begin = var_8670_begin_0, end = var_8670_end_0, end_mask = var_8670_end_mask_0, squeeze_mask = var_8670_squeeze_mask_0, x = reshape_204_cast_fp16)[name = tensor("op_8670_cast_fp16")]; + tensor var_8672_axis_0 = const()[name = tensor("op_8672_axis_0"), val = tensor(1)]; + tensor var_8672_mode_0 = const()[name = tensor("op_8672_mode_0"), val = tensor("update")]; + tensor var_8672_validate_indices_0 = const()[name = tensor("op_8672_validate_indices_0"), val = tensor(false)]; + tensor var_8672_cast_fp16 = scatter_along_axis(axis = var_8672_axis_0, data = var_8670_cast_fp16, indices = write_indices_41, mode = var_8672_mode_0, updates = v_41_cast_fp16, validate_indices = var_8672_validate_indices_0)[name = tensor("op_8672_cast_fp16")]; + tensor concat_164 = const()[name = tensor("concat_164"), val = tensor([1, 0, 0, 0, 0])]; + tensor concat_165 = const()[name = tensor("concat_165"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_41_internal_tensor_assign_2_stride_0 = const()[name = tensor("new_cache_41_internal_tensor_assign_2_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_41_internal_tensor_assign_2_begin_mask_0 = const()[name = tensor("new_cache_41_internal_tensor_assign_2_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_41_internal_tensor_assign_2_end_mask_0 = const()[name = tensor("new_cache_41_internal_tensor_assign_2_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_41_internal_tensor_assign_2_squeeze_mask_0 = const()[name = tensor("new_cache_41_internal_tensor_assign_2_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_87 = const()[name = tensor("shape_87"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_41 = const()[name = tensor("reduce_prod_41"), val = tensor(1048576)]; + tensor range_1d_41_start_0 = const()[name = tensor("range_1d_41_start_0"), val = tensor(0)]; + tensor range_1d_41_step_0 = const()[name = tensor("range_1d_41_step_0"), val = tensor(1)]; + tensor range_1d_41 = range_1d(end = reduce_prod_41, start = range_1d_41_start_0, step = range_1d_41_step_0)[name = tensor("range_1d_41")]; + tensor reshape_205 = reshape(shape = shape_87, x = range_1d_41)[name = tensor("reshape_205")]; + tensor slice_by_index_41 = slice_by_index(begin = concat_164, begin_mask = new_cache_41_internal_tensor_assign_2_begin_mask_0, end = concat_165, end_mask = new_cache_41_internal_tensor_assign_2_end_mask_0, squeeze_mask = new_cache_41_internal_tensor_assign_2_squeeze_mask_0, stride = new_cache_41_internal_tensor_assign_2_stride_0, x = reshape_205)[name = tensor("slice_by_index_41")]; + tensor reshape_206_shape_0 = const()[name = tensor("reshape_206_shape_0"), val = tensor([-1])]; + tensor reshape_206 = reshape(shape = reshape_206_shape_0, x = slice_by_index_41)[name = tensor("reshape_206")]; + tensor reshape_207_shape_0 = const()[name = tensor("reshape_207_shape_0"), val = tensor([-1])]; + tensor reshape_207_cast_fp16 = reshape(shape = reshape_207_shape_0, x = var_8672_cast_fp16)[name = tensor("reshape_207_cast_fp16")]; + tensor reshape_208_shape_0 = const()[name = tensor("reshape_208_shape_0"), val = tensor([-1])]; + tensor reshape_208_cast_fp16 = reshape(shape = reshape_208_shape_0, x = reshape_204_cast_fp16)[name = tensor("reshape_208_cast_fp16")]; + tensor scatter_41_mode_0 = const()[name = tensor("scatter_41_mode_0"), val = tensor("update")]; + tensor scatter_41_axis_0 = const()[name = tensor("scatter_41_axis_0"), val = tensor(0)]; + tensor scatter_41_validate_indices_0 = const()[name = tensor("scatter_41_validate_indices_0"), val = tensor(false)]; + tensor scatter_41_cast_fp16 = scatter(axis = scatter_41_axis_0, data = reshape_208_cast_fp16, indices = reshape_206, mode = scatter_41_mode_0, updates = reshape_207_cast_fp16, validate_indices = scatter_41_validate_indices_0)[name = tensor("scatter_41_cast_fp16")]; + tensor reshape_209_cast_fp16 = reshape(shape = shape_87, x = scatter_41_cast_fp16)[name = tensor("reshape_209_cast_fp16")]; + tensor reshape_209_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("reshape_209_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor keys_121_begin_0 = const()[name = tensor("keys_121_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor keys_121_end_0 = const()[name = tensor("keys_121_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor keys_121_end_mask_0 = const()[name = tensor("keys_121_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor keys_121_squeeze_mask_0 = const()[name = tensor("keys_121_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor keys_121_cast_fp16 = slice_by_index(begin = keys_121_begin_0, end = keys_121_end_0, end_mask = keys_121_end_mask_0, squeeze_mask = keys_121_squeeze_mask_0, x = reshape_209_cast_fp16)[name = tensor("keys_121_cast_fp16")]; + tensor values_121_begin_0 = const()[name = tensor("values_121_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor values_121_end_0 = const()[name = tensor("values_121_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor values_121_end_mask_0 = const()[name = tensor("values_121_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor values_121_squeeze_mask_0 = const()[name = tensor("values_121_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor values_121_cast_fp16 = slice_by_index(begin = values_121_begin_0, end = values_121_end_0, end_mask = values_121_end_mask_0, squeeze_mask = values_121_squeeze_mask_0, x = reshape_209_cast_fp16)[name = tensor("values_121_cast_fp16")]; + tensor var_8684_cast_fp16 = not_equal(x = keys_121_cast_fp16, y = keys_121_cast_fp16)[name = tensor("op_8684_cast_fp16")]; + tensor keys_123_cast_fp16 = select(a = var_470_to_fp16, b = keys_121_cast_fp16, cond = var_8684_cast_fp16)[name = tensor("keys_123_cast_fp16")]; + tensor var_8692_cast_fp16 = not_equal(x = values_121_cast_fp16, y = values_121_cast_fp16)[name = tensor("op_8692_cast_fp16")]; + tensor values_123_cast_fp16 = select(a = var_470_to_fp16, b = values_121_cast_fp16, cond = var_8692_cast_fp16)[name = tensor("values_123_cast_fp16")]; + tensor var_8716 = const()[name = tensor("op_8716"), val = tensor([0, 2, 1, 3])]; + tensor var_8732 = const()[name = tensor("op_8732"), val = tensor([1, 1, 1])]; + tensor var_8733_cast_fp16 = reshape(shape = var_8732, x = position20_to_fp16)[name = tensor("op_8733_cast_fp16")]; + tensor q_positions_41_cast_fp16 = add(x = var_8733_cast_fp16, y = q_offsets_1_promoted_to_fp16)[name = tensor("q_positions_41_cast_fp16")]; + tensor valid_end_41_cast_fp16 = add(x = var_8733_cast_fp16, y = var_536_cast_fp16)[name = tensor("valid_end_41_cast_fp16")]; + tensor valid_mask_41_cast_fp16 = less(x = k_positions_1_promoted_to_fp16, y = valid_end_41_cast_fp16)[name = tensor("valid_mask_41_cast_fp16")]; + tensor causal_mask_41_cast_fp16 = less_equal(x = k_positions_1_promoted_to_fp16, y = q_positions_41_cast_fp16)[name = tensor("causal_mask_41_cast_fp16")]; + tensor var_8761 = logical_and(x = valid_mask_41_cast_fp16, y = causal_mask_41_cast_fp16)[name = tensor("op_8761")]; + tensor attn_mask_41_axes_0 = const()[name = tensor("attn_mask_41_axes_0"), val = tensor([1])]; + tensor attn_mask_41 = expand_dims(axes = attn_mask_41_axes_0, x = var_8761)[name = tensor("attn_mask_41")]; + tensor var_8775_transpose_x_0 = const()[name = tensor("op_8775_transpose_x_0"), val = tensor(false)]; + tensor var_8775_transpose_y_0 = const()[name = tensor("op_8775_transpose_y_0"), val = tensor(false)]; + tensor transpose_109_perm_0 = const()[name = tensor("transpose_109_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_110_perm_0 = const()[name = tensor("transpose_110_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_110 = transpose(perm = transpose_110_perm_0, x = keys_123_cast_fp16)[name = tensor("transpose_124")]; + tensor transpose_109 = transpose(perm = transpose_109_perm_0, x = q_123_cast_fp16)[name = tensor("transpose_125")]; + tensor var_8775_cast_fp16 = matmul(transpose_x = var_8775_transpose_x_0, transpose_y = var_8775_transpose_y_0, x = transpose_109, y = transpose_110)[name = tensor("op_8775_cast_fp16")]; + tensor var_8769_to_fp16 = const()[name = tensor("op_8769_to_fp16"), val = tensor([0x1p-3])]; + tensor attn_weights_121_cast_fp16 = mul(x = var_8775_cast_fp16, y = var_8769_to_fp16)[name = tensor("attn_weights_121_cast_fp16")]; + tensor var_8777 = logical_not(x = attn_mask_41)[name = tensor("op_8777")]; + tensor var_8778_to_fp16 = const()[name = tensor("op_8778_to_fp16"), val = tensor(-inf)]; + tensor attn_weights_123_cast_fp16 = select(a = var_8778_to_fp16, b = attn_weights_121_cast_fp16, cond = var_8777)[name = tensor("attn_weights_123_cast_fp16")]; + tensor var_8780 = const()[name = tensor("op_8780"), val = tensor(-1)]; + tensor attn_weights_125_cast_fp16 = softmax(axis = var_8780, x = attn_weights_123_cast_fp16)[name = tensor("attn_weights_125_cast_fp16")]; + tensor attn_output_41_transpose_x_0 = const()[name = tensor("attn_output_41_transpose_x_0"), val = tensor(false)]; + tensor attn_output_41_transpose_y_0 = const()[name = tensor("attn_output_41_transpose_y_0"), val = tensor(false)]; + tensor values_125_cast_fp16 = transpose(perm = var_8716, x = values_123_cast_fp16)[name = tensor("transpose_126")]; + tensor attn_output_41_cast_fp16 = matmul(transpose_x = attn_output_41_transpose_x_0, transpose_y = attn_output_41_transpose_y_0, x = attn_weights_125_cast_fp16, y = values_125_cast_fp16)[name = tensor("attn_output_41_cast_fp16")]; + tensor var_8788 = const()[name = tensor("op_8788"), val = tensor([0, 2, 1, 3])]; + tensor var_8793 = const()[name = tensor("op_8793"), val = tensor([1, 256, 1024])]; + tensor var_8789_cast_fp16 = transpose(perm = var_8788, x = attn_output_41_cast_fp16)[name = tensor("transpose_123")]; + tensor input_245_cast_fp16 = reshape(shape = var_8793, x = var_8789_cast_fp16)[name = tensor("input_245_cast_fp16")]; + tensor attn20_out_proj_weight_to_fp16 = const()[name = tensor("attn20_out_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(510857088)))]; + tensor linear_81_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = attn20_out_proj_weight_to_fp16, x = input_245_cast_fp16)[name = tensor("linear_81_cast_fp16")]; + tensor var_8802_cast_fp16 = add(x = position20_to_fp16, y = valid_len_to_fp16)[name = tensor("op_8802_cast_fp16")]; + tensor var_8802_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_8802_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor input_247_cast_fp16 = add(x = input_241_cast_fp16, y = linear_81_cast_fp16)[name = tensor("input_247_cast_fp16")]; + tensor input_249_axes_0 = const()[name = tensor("input_249_axes_0"), val = tensor([-1])]; + tensor norm20_2_weight_to_fp16 = const()[name = tensor("norm20_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(512954304)))]; + tensor norm20_2_bias_to_fp16 = const()[name = tensor("norm20_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(512956416)))]; + tensor var_8806_to_fp16 = const()[name = tensor("op_8806_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_249_cast_fp16 = layer_norm(axes = input_249_axes_0, beta = norm20_2_bias_to_fp16, epsilon = var_8806_to_fp16, gamma = norm20_2_weight_to_fp16, x = input_247_cast_fp16)[name = tensor("input_249_cast_fp16")]; + tensor linear20_1_weight_to_fp16 = const()[name = tensor("linear20_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(512958528)))]; + tensor linear_82_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = linear20_1_weight_to_fp16, x = input_249_cast_fp16)[name = tensor("linear_82_cast_fp16")]; + tensor input_251_mode_0 = const()[name = tensor("input_251_mode_0"), val = tensor("EXACT")]; + tensor input_251_cast_fp16 = gelu(mode = input_251_mode_0, x = linear_82_cast_fp16)[name = tensor("input_251_cast_fp16")]; + tensor linear20_2_weight_to_fp16 = const()[name = tensor("linear20_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(521347200)))]; + tensor linear_83_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = linear20_2_weight_to_fp16, x = input_251_cast_fp16)[name = tensor("linear_83_cast_fp16")]; + tensor input_253_cast_fp16 = add(x = input_247_cast_fp16, y = linear_83_cast_fp16)[name = tensor("input_253_cast_fp16")]; + tensor input_255_axes_0 = const()[name = tensor("input_255_axes_0"), val = tensor([-1])]; + tensor norm21_1_weight_to_fp16 = const()[name = tensor("norm21_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(529735872)))]; + tensor norm21_1_bias_to_fp16 = const()[name = tensor("norm21_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(529737984)))]; + tensor var_8823_to_fp16 = const()[name = tensor("op_8823_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_255_cast_fp16 = layer_norm(axes = input_255_axes_0, beta = norm21_1_bias_to_fp16, epsilon = var_8823_to_fp16, gamma = norm21_1_weight_to_fp16, x = input_253_cast_fp16)[name = tensor("input_255_cast_fp16")]; + tensor attn21_in_proj_weight_to_fp16 = const()[name = tensor("attn21_in_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(529740096)))]; + tensor linear_84_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = attn21_in_proj_weight_to_fp16, x = input_255_cast_fp16)[name = tensor("linear_84_cast_fp16")]; + tensor var_8837 = const()[name = tensor("op_8837"), val = tensor([1, 256, 3, 16, 64])]; + tensor qkv_43_cast_fp16 = reshape(shape = var_8837, x = linear_84_cast_fp16)[name = tensor("qkv_43_cast_fp16")]; + tensor q_127_begin_0 = const()[name = tensor("q_127_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor q_127_end_0 = const()[name = tensor("q_127_end_0"), val = tensor([1, 256, 1, 16, 64])]; + tensor q_127_end_mask_0 = const()[name = tensor("q_127_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor q_127_squeeze_mask_0 = const()[name = tensor("q_127_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor q_127_cast_fp16 = slice_by_index(begin = q_127_begin_0, end = q_127_end_0, end_mask = q_127_end_mask_0, squeeze_mask = q_127_squeeze_mask_0, x = qkv_43_cast_fp16)[name = tensor("q_127_cast_fp16")]; + tensor k_85_begin_0 = const()[name = tensor("k_85_begin_0"), val = tensor([0, 0, 1, 0, 0])]; + tensor k_85_end_0 = const()[name = tensor("k_85_end_0"), val = tensor([1, 256, 2, 16, 64])]; + tensor k_85_end_mask_0 = const()[name = tensor("k_85_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor k_85_squeeze_mask_0 = const()[name = tensor("k_85_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor k_85_cast_fp16 = slice_by_index(begin = k_85_begin_0, end = k_85_end_0, end_mask = k_85_end_mask_0, squeeze_mask = k_85_squeeze_mask_0, x = qkv_43_cast_fp16)[name = tensor("k_85_cast_fp16")]; + tensor v_43_begin_0 = const()[name = tensor("v_43_begin_0"), val = tensor([0, 0, 2, 0, 0])]; + tensor v_43_end_0 = const()[name = tensor("v_43_end_0"), val = tensor([1, 256, 3, 16, 64])]; + tensor v_43_end_mask_0 = const()[name = tensor("v_43_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor v_43_squeeze_mask_0 = const()[name = tensor("v_43_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor v_43_cast_fp16 = slice_by_index(begin = v_43_begin_0, end = v_43_end_0, end_mask = v_43_end_mask_0, squeeze_mask = v_43_squeeze_mask_0, x = qkv_43_cast_fp16)[name = tensor("v_43_cast_fp16")]; + tensor var_8895 = const()[name = tensor("op_8895"), val = tensor([1, 1])]; + tensor position21_to_fp16_dtype_0 = const()[name = tensor("position21_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor position21_to_fp16 = cast(dtype = position21_to_fp16_dtype_0, x = position21)[name = tensor("cast_556")]; + tensor var_8896_cast_fp16 = reshape(shape = var_8895, x = position21_to_fp16)[name = tensor("op_8896_cast_fp16")]; + tensor ts_129_cast_fp16 = add(x = ts_1_promoted_to_fp16, y = var_8896_cast_fp16)[name = tensor("ts_129_cast_fp16")]; + tensor var_8903 = const()[name = tensor("op_8903"), val = tensor([1, 256, 1, 1])]; + tensor ts_131_cast_fp16 = reshape(shape = var_8903, x = ts_129_cast_fp16)[name = tensor("ts_131_cast_fp16")]; + tensor var_8910 = const()[name = tensor("op_8910"), val = tensor([1, 256, 16, 32, 2])]; + tensor q_complex_43_cast_fp16 = reshape(shape = var_8910, x = q_127_cast_fp16)[name = tensor("q_complex_43_cast_fp16")]; + tensor var_8917 = const()[name = tensor("op_8917"), val = tensor([1, 256, 16, 32, 2])]; + tensor k_complex_43_cast_fp16 = reshape(shape = var_8917, x = k_85_cast_fp16)[name = tensor("k_complex_43_cast_fp16")]; + tensor var_8921_begin_0 = const()[name = tensor("op_8921_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_8921_end_0 = const()[name = tensor("op_8921_end_0"), val = tensor([1, 256, 16, 32, 1])]; + tensor var_8921_end_mask_0 = const()[name = tensor("op_8921_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_8921_squeeze_mask_0 = const()[name = tensor("op_8921_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_8921_cast_fp16 = slice_by_index(begin = var_8921_begin_0, end = var_8921_end_0, end_mask = var_8921_end_mask_0, squeeze_mask = var_8921_squeeze_mask_0, x = q_complex_43_cast_fp16)[name = tensor("op_8921_cast_fp16")]; + tensor var_8929_begin_0 = const()[name = tensor("op_8929_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_8929_end_0 = const()[name = tensor("op_8929_end_0"), val = tensor([1, 256, 16, 32, 2])]; + tensor var_8929_end_mask_0 = const()[name = tensor("op_8929_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_8929_squeeze_mask_0 = const()[name = tensor("op_8929_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_8929_cast_fp16 = slice_by_index(begin = var_8929_begin_0, end = var_8929_end_0, end_mask = var_8929_end_mask_0, squeeze_mask = var_8929_squeeze_mask_0, x = q_complex_43_cast_fp16)[name = tensor("op_8929_cast_fp16")]; + tensor var_8937_begin_0 = const()[name = tensor("op_8937_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_8937_end_0 = const()[name = tensor("op_8937_end_0"), val = tensor([1, 256, 16, 32, 1])]; + tensor var_8937_end_mask_0 = const()[name = tensor("op_8937_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_8937_squeeze_mask_0 = const()[name = tensor("op_8937_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_8937_cast_fp16 = slice_by_index(begin = var_8937_begin_0, end = var_8937_end_0, end_mask = var_8937_end_mask_0, squeeze_mask = var_8937_squeeze_mask_0, x = k_complex_43_cast_fp16)[name = tensor("op_8937_cast_fp16")]; + tensor var_8945_begin_0 = const()[name = tensor("op_8945_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_8945_end_0 = const()[name = tensor("op_8945_end_0"), val = tensor([1, 256, 16, 32, 2])]; + tensor var_8945_end_mask_0 = const()[name = tensor("op_8945_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_8945_squeeze_mask_0 = const()[name = tensor("op_8945_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_8945_cast_fp16 = slice_by_index(begin = var_8945_begin_0, end = var_8945_end_0, end_mask = var_8945_end_mask_0, squeeze_mask = var_8945_squeeze_mask_0, x = k_complex_43_cast_fp16)[name = tensor("op_8945_cast_fp16")]; + tensor freqs_43_to_fp16 = const()[name = tensor("freqs_43_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(536031616)))]; + tensor var_8951_cast_fp16 = mul(x = freqs_43_to_fp16, y = ts_131_cast_fp16)[name = tensor("op_8951_cast_fp16")]; + tensor rotr_43_cast_fp16 = cos(x = var_8951_cast_fp16)[name = tensor("rotr_43_cast_fp16")]; + tensor roti_43_cast_fp16 = sin(x = var_8951_cast_fp16)[name = tensor("roti_43_cast_fp16")]; + tensor var_8955_cast_fp16 = mul(x = var_8921_cast_fp16, y = rotr_43_cast_fp16)[name = tensor("op_8955_cast_fp16")]; + tensor var_8956_cast_fp16 = mul(x = var_8929_cast_fp16, y = roti_43_cast_fp16)[name = tensor("op_8956_cast_fp16")]; + tensor qor_85_cast_fp16 = sub(x = var_8955_cast_fp16, y = var_8956_cast_fp16)[name = tensor("qor_85_cast_fp16")]; + tensor var_8959_cast_fp16 = mul(x = var_8921_cast_fp16, y = roti_43_cast_fp16)[name = tensor("op_8959_cast_fp16")]; + tensor var_8960_cast_fp16 = mul(x = var_8929_cast_fp16, y = rotr_43_cast_fp16)[name = tensor("op_8960_cast_fp16")]; + tensor qoi_85_cast_fp16 = add(x = var_8959_cast_fp16, y = var_8960_cast_fp16)[name = tensor("qoi_85_cast_fp16")]; + tensor var_8963_cast_fp16 = mul(x = var_8937_cast_fp16, y = rotr_43_cast_fp16)[name = tensor("op_8963_cast_fp16")]; + tensor var_8964_cast_fp16 = mul(x = var_8945_cast_fp16, y = roti_43_cast_fp16)[name = tensor("op_8964_cast_fp16")]; + tensor kor_85_cast_fp16 = sub(x = var_8963_cast_fp16, y = var_8964_cast_fp16)[name = tensor("kor_85_cast_fp16")]; + tensor var_8967_cast_fp16 = mul(x = var_8937_cast_fp16, y = roti_43_cast_fp16)[name = tensor("op_8967_cast_fp16")]; + tensor var_8968_cast_fp16 = mul(x = var_8945_cast_fp16, y = rotr_43_cast_fp16)[name = tensor("op_8968_cast_fp16")]; + tensor koi_85_cast_fp16 = add(x = var_8967_cast_fp16, y = var_8968_cast_fp16)[name = tensor("koi_85_cast_fp16")]; + tensor qo_43_axis_0 = const()[name = tensor("qo_43_axis_0"), val = tensor(-1)]; + tensor qo_43_cast_fp16 = stack(axis = qo_43_axis_0, values = (qor_85_cast_fp16, qoi_85_cast_fp16))[name = tensor("qo_43_cast_fp16")]; + tensor ko_43_axis_0 = const()[name = tensor("ko_43_axis_0"), val = tensor(-1)]; + tensor ko_43_cast_fp16 = stack(axis = ko_43_axis_0, values = (kor_85_cast_fp16, koi_85_cast_fp16))[name = tensor("ko_43_cast_fp16")]; + tensor var_9001 = const()[name = tensor("op_9001"), val = tensor([1, 256, 16, 64])]; + tensor q_129_cast_fp16 = reshape(shape = var_9001, x = qo_43_cast_fp16)[name = tensor("q_129_cast_fp16")]; + tensor var_9007 = const()[name = tensor("op_9007"), val = tensor([1, 256, 16, 64])]; + tensor k_87_cast_fp16 = reshape(shape = var_9007, x = ko_43_cast_fp16)[name = tensor("k_87_cast_fp16")]; + tensor write_indices_float_127_cast_fp16 = add(x = var_8896_cast_fp16, y = var_393_promoted_to_fp16)[name = tensor("write_indices_float_127_cast_fp16")]; + tensor _inversed_9028_y_0_to_fp16 = const()[name = tensor("_inversed_9028_y_0_to_fp16"), val = tensor(0x1p-9)]; + tensor _inversed_9028_cast_fp16 = mul(x = write_indices_float_127_cast_fp16, y = _inversed_9028_y_0_to_fp16)[name = tensor("_inversed_9028_cast_fp16")]; + tensor var_9029_cast_fp16 = floor(x = _inversed_9028_cast_fp16)[name = tensor("op_9029_cast_fp16")]; + tensor var_9030_to_fp16 = const()[name = tensor("op_9030_to_fp16"), val = tensor(0x1p+9)]; + tensor var_9031_cast_fp16 = mul(x = var_9029_cast_fp16, y = var_9030_to_fp16)[name = tensor("op_9031_cast_fp16")]; + tensor write_indices_float_129_cast_fp16 = sub(x = write_indices_float_127_cast_fp16, y = var_9031_cast_fp16)[name = tensor("write_indices_float_129_cast_fp16")]; + tensor var_9049_value_0_to_fp16 = const()[name = tensor("op_9049_value_0_to_fp16"), val = tensor(0x1.ffp+8)]; + tensor var_9049_cast_fp16 = fill_like(ref_tensor = write_indices_float_129_cast_fp16, value = var_9049_value_0_to_fp16)[name = tensor("op_9049_cast_fp16")]; + tensor write_indices_float_131_cast_fp16 = select(a = write_indices_float_129_cast_fp16, b = var_9049_cast_fp16, cond = token_valid_1_cast_fp16)[name = tensor("write_indices_float_131_cast_fp16")]; + tensor var_9055_dtype_0 = const()[name = tensor("op_9055_dtype_0"), val = tensor("int32")]; + tensor var_9060 = const()[name = tensor("op_9060"), val = tensor([1, 256, 1, 1])]; + tensor write_indices_float_131_cast_fp16_to_int32 = cast(dtype = var_9055_dtype_0, x = write_indices_float_131_cast_fp16)[name = tensor("cast_555")]; + tensor var_9061 = reshape(shape = var_9060, x = write_indices_float_131_cast_fp16_to_int32)[name = tensor("op_9061")]; + tensor write_indices_43_reps_0 = const()[name = tensor("write_indices_43_reps_0"), val = tensor([1, 1, 16, 64])]; + tensor write_indices_43 = tile(reps = write_indices_43_reps_0, x = var_9061)[name = tensor("write_indices_43")]; + tensor var_9071_begin_0 = const()[name = tensor("op_9071_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_9071_end_0 = const()[name = tensor("op_9071_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor var_9071_end_mask_0 = const()[name = tensor("op_9071_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_9071_squeeze_mask_0 = const()[name = tensor("op_9071_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor cache21_to_fp16_dtype_0 = const()[name = tensor("cache21_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor cache21_to_fp16 = cast(dtype = cache21_to_fp16_dtype_0, x = cache21)[name = tensor("cast_554")]; + tensor var_9071_cast_fp16 = slice_by_index(begin = var_9071_begin_0, end = var_9071_end_0, end_mask = var_9071_end_mask_0, squeeze_mask = var_9071_squeeze_mask_0, x = cache21_to_fp16)[name = tensor("op_9071_cast_fp16")]; + tensor var_9073_axis_0 = const()[name = tensor("op_9073_axis_0"), val = tensor(1)]; + tensor var_9073_mode_0 = const()[name = tensor("op_9073_mode_0"), val = tensor("update")]; + tensor var_9073_validate_indices_0 = const()[name = tensor("op_9073_validate_indices_0"), val = tensor(false)]; + tensor var_9073_cast_fp16 = scatter_along_axis(axis = var_9073_axis_0, data = var_9071_cast_fp16, indices = write_indices_43, mode = var_9073_mode_0, updates = k_87_cast_fp16, validate_indices = var_9073_validate_indices_0)[name = tensor("op_9073_cast_fp16")]; + tensor concat_170 = const()[name = tensor("concat_170"), val = tensor([0, 0, 0, 0, 0])]; + tensor concat_171 = const()[name = tensor("concat_171"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_43_internal_tensor_assign_1_stride_0 = const()[name = tensor("new_cache_43_internal_tensor_assign_1_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_43_internal_tensor_assign_1_begin_mask_0 = const()[name = tensor("new_cache_43_internal_tensor_assign_1_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_43_internal_tensor_assign_1_end_mask_0 = const()[name = tensor("new_cache_43_internal_tensor_assign_1_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_43_internal_tensor_assign_1_squeeze_mask_0 = const()[name = tensor("new_cache_43_internal_tensor_assign_1_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_88 = const()[name = tensor("shape_88"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_42 = const()[name = tensor("reduce_prod_42"), val = tensor(1048576)]; + tensor range_1d_42_start_0 = const()[name = tensor("range_1d_42_start_0"), val = tensor(0)]; + tensor range_1d_42_step_0 = const()[name = tensor("range_1d_42_step_0"), val = tensor(1)]; + tensor range_1d_42 = range_1d(end = reduce_prod_42, start = range_1d_42_start_0, step = range_1d_42_step_0)[name = tensor("range_1d_42")]; + tensor reshape_210 = reshape(shape = shape_88, x = range_1d_42)[name = tensor("reshape_210")]; + tensor slice_by_index_42 = slice_by_index(begin = concat_170, begin_mask = new_cache_43_internal_tensor_assign_1_begin_mask_0, end = concat_171, end_mask = new_cache_43_internal_tensor_assign_1_end_mask_0, squeeze_mask = new_cache_43_internal_tensor_assign_1_squeeze_mask_0, stride = new_cache_43_internal_tensor_assign_1_stride_0, x = reshape_210)[name = tensor("slice_by_index_42")]; + tensor reshape_211_shape_0 = const()[name = tensor("reshape_211_shape_0"), val = tensor([-1])]; + tensor reshape_211 = reshape(shape = reshape_211_shape_0, x = slice_by_index_42)[name = tensor("reshape_211")]; + tensor reshape_212_shape_0 = const()[name = tensor("reshape_212_shape_0"), val = tensor([-1])]; + tensor reshape_212_cast_fp16 = reshape(shape = reshape_212_shape_0, x = var_9073_cast_fp16)[name = tensor("reshape_212_cast_fp16")]; + tensor reshape_213_shape_0 = const()[name = tensor("reshape_213_shape_0"), val = tensor([-1])]; + tensor reshape_213_cast_fp16 = reshape(shape = reshape_213_shape_0, x = cache21_to_fp16)[name = tensor("reshape_213_cast_fp16")]; + tensor scatter_42_mode_0 = const()[name = tensor("scatter_42_mode_0"), val = tensor("update")]; + tensor scatter_42_axis_0 = const()[name = tensor("scatter_42_axis_0"), val = tensor(0)]; + tensor scatter_42_validate_indices_0 = const()[name = tensor("scatter_42_validate_indices_0"), val = tensor(false)]; + tensor scatter_42_cast_fp16 = scatter(axis = scatter_42_axis_0, data = reshape_213_cast_fp16, indices = reshape_211, mode = scatter_42_mode_0, updates = reshape_212_cast_fp16, validate_indices = scatter_42_validate_indices_0)[name = tensor("scatter_42_cast_fp16")]; + tensor reshape_214_cast_fp16 = reshape(shape = shape_88, x = scatter_42_cast_fp16)[name = tensor("reshape_214_cast_fp16")]; + tensor var_9081_begin_0 = const()[name = tensor("op_9081_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor var_9081_end_0 = const()[name = tensor("op_9081_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor var_9081_end_mask_0 = const()[name = tensor("op_9081_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_9081_squeeze_mask_0 = const()[name = tensor("op_9081_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor var_9081_cast_fp16 = slice_by_index(begin = var_9081_begin_0, end = var_9081_end_0, end_mask = var_9081_end_mask_0, squeeze_mask = var_9081_squeeze_mask_0, x = reshape_214_cast_fp16)[name = tensor("op_9081_cast_fp16")]; + tensor var_9083_axis_0 = const()[name = tensor("op_9083_axis_0"), val = tensor(1)]; + tensor var_9083_mode_0 = const()[name = tensor("op_9083_mode_0"), val = tensor("update")]; + tensor var_9083_validate_indices_0 = const()[name = tensor("op_9083_validate_indices_0"), val = tensor(false)]; + tensor var_9083_cast_fp16 = scatter_along_axis(axis = var_9083_axis_0, data = var_9081_cast_fp16, indices = write_indices_43, mode = var_9083_mode_0, updates = v_43_cast_fp16, validate_indices = var_9083_validate_indices_0)[name = tensor("op_9083_cast_fp16")]; + tensor concat_172 = const()[name = tensor("concat_172"), val = tensor([1, 0, 0, 0, 0])]; + tensor concat_173 = const()[name = tensor("concat_173"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_43_internal_tensor_assign_2_stride_0 = const()[name = tensor("new_cache_43_internal_tensor_assign_2_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_43_internal_tensor_assign_2_begin_mask_0 = const()[name = tensor("new_cache_43_internal_tensor_assign_2_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_43_internal_tensor_assign_2_end_mask_0 = const()[name = tensor("new_cache_43_internal_tensor_assign_2_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_43_internal_tensor_assign_2_squeeze_mask_0 = const()[name = tensor("new_cache_43_internal_tensor_assign_2_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_89 = const()[name = tensor("shape_89"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_43 = const()[name = tensor("reduce_prod_43"), val = tensor(1048576)]; + tensor range_1d_43_start_0 = const()[name = tensor("range_1d_43_start_0"), val = tensor(0)]; + tensor range_1d_43_step_0 = const()[name = tensor("range_1d_43_step_0"), val = tensor(1)]; + tensor range_1d_43 = range_1d(end = reduce_prod_43, start = range_1d_43_start_0, step = range_1d_43_step_0)[name = tensor("range_1d_43")]; + tensor reshape_215 = reshape(shape = shape_89, x = range_1d_43)[name = tensor("reshape_215")]; + tensor slice_by_index_43 = slice_by_index(begin = concat_172, begin_mask = new_cache_43_internal_tensor_assign_2_begin_mask_0, end = concat_173, end_mask = new_cache_43_internal_tensor_assign_2_end_mask_0, squeeze_mask = new_cache_43_internal_tensor_assign_2_squeeze_mask_0, stride = new_cache_43_internal_tensor_assign_2_stride_0, x = reshape_215)[name = tensor("slice_by_index_43")]; + tensor reshape_216_shape_0 = const()[name = tensor("reshape_216_shape_0"), val = tensor([-1])]; + tensor reshape_216 = reshape(shape = reshape_216_shape_0, x = slice_by_index_43)[name = tensor("reshape_216")]; + tensor reshape_217_shape_0 = const()[name = tensor("reshape_217_shape_0"), val = tensor([-1])]; + tensor reshape_217_cast_fp16 = reshape(shape = reshape_217_shape_0, x = var_9083_cast_fp16)[name = tensor("reshape_217_cast_fp16")]; + tensor reshape_218_shape_0 = const()[name = tensor("reshape_218_shape_0"), val = tensor([-1])]; + tensor reshape_218_cast_fp16 = reshape(shape = reshape_218_shape_0, x = reshape_214_cast_fp16)[name = tensor("reshape_218_cast_fp16")]; + tensor scatter_43_mode_0 = const()[name = tensor("scatter_43_mode_0"), val = tensor("update")]; + tensor scatter_43_axis_0 = const()[name = tensor("scatter_43_axis_0"), val = tensor(0)]; + tensor scatter_43_validate_indices_0 = const()[name = tensor("scatter_43_validate_indices_0"), val = tensor(false)]; + tensor scatter_43_cast_fp16 = scatter(axis = scatter_43_axis_0, data = reshape_218_cast_fp16, indices = reshape_216, mode = scatter_43_mode_0, updates = reshape_217_cast_fp16, validate_indices = scatter_43_validate_indices_0)[name = tensor("scatter_43_cast_fp16")]; + tensor reshape_219_cast_fp16 = reshape(shape = shape_89, x = scatter_43_cast_fp16)[name = tensor("reshape_219_cast_fp16")]; + tensor reshape_219_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("reshape_219_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor keys_127_begin_0 = const()[name = tensor("keys_127_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor keys_127_end_0 = const()[name = tensor("keys_127_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor keys_127_end_mask_0 = const()[name = tensor("keys_127_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor keys_127_squeeze_mask_0 = const()[name = tensor("keys_127_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor keys_127_cast_fp16 = slice_by_index(begin = keys_127_begin_0, end = keys_127_end_0, end_mask = keys_127_end_mask_0, squeeze_mask = keys_127_squeeze_mask_0, x = reshape_219_cast_fp16)[name = tensor("keys_127_cast_fp16")]; + tensor values_127_begin_0 = const()[name = tensor("values_127_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor values_127_end_0 = const()[name = tensor("values_127_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor values_127_end_mask_0 = const()[name = tensor("values_127_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor values_127_squeeze_mask_0 = const()[name = tensor("values_127_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor values_127_cast_fp16 = slice_by_index(begin = values_127_begin_0, end = values_127_end_0, end_mask = values_127_end_mask_0, squeeze_mask = values_127_squeeze_mask_0, x = reshape_219_cast_fp16)[name = tensor("values_127_cast_fp16")]; + tensor var_9095_cast_fp16 = not_equal(x = keys_127_cast_fp16, y = keys_127_cast_fp16)[name = tensor("op_9095_cast_fp16")]; + tensor keys_129_cast_fp16 = select(a = var_470_to_fp16, b = keys_127_cast_fp16, cond = var_9095_cast_fp16)[name = tensor("keys_129_cast_fp16")]; + tensor var_9103_cast_fp16 = not_equal(x = values_127_cast_fp16, y = values_127_cast_fp16)[name = tensor("op_9103_cast_fp16")]; + tensor values_129_cast_fp16 = select(a = var_470_to_fp16, b = values_127_cast_fp16, cond = var_9103_cast_fp16)[name = tensor("values_129_cast_fp16")]; + tensor var_9127 = const()[name = tensor("op_9127"), val = tensor([0, 2, 1, 3])]; + tensor var_9143 = const()[name = tensor("op_9143"), val = tensor([1, 1, 1])]; + tensor var_9144_cast_fp16 = reshape(shape = var_9143, x = position21_to_fp16)[name = tensor("op_9144_cast_fp16")]; + tensor q_positions_43_cast_fp16 = add(x = var_9144_cast_fp16, y = q_offsets_1_promoted_to_fp16)[name = tensor("q_positions_43_cast_fp16")]; + tensor valid_end_43_cast_fp16 = add(x = var_9144_cast_fp16, y = var_536_cast_fp16)[name = tensor("valid_end_43_cast_fp16")]; + tensor valid_mask_43_cast_fp16 = less(x = k_positions_1_promoted_to_fp16, y = valid_end_43_cast_fp16)[name = tensor("valid_mask_43_cast_fp16")]; + tensor causal_mask_43_cast_fp16 = less_equal(x = k_positions_1_promoted_to_fp16, y = q_positions_43_cast_fp16)[name = tensor("causal_mask_43_cast_fp16")]; + tensor var_9172 = logical_and(x = valid_mask_43_cast_fp16, y = causal_mask_43_cast_fp16)[name = tensor("op_9172")]; + tensor attn_mask_43_axes_0 = const()[name = tensor("attn_mask_43_axes_0"), val = tensor([1])]; + tensor attn_mask_43 = expand_dims(axes = attn_mask_43_axes_0, x = var_9172)[name = tensor("attn_mask_43")]; + tensor var_9186_transpose_x_0 = const()[name = tensor("op_9186_transpose_x_0"), val = tensor(false)]; + tensor var_9186_transpose_y_0 = const()[name = tensor("op_9186_transpose_y_0"), val = tensor(false)]; + tensor transpose_111_perm_0 = const()[name = tensor("transpose_111_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_112_perm_0 = const()[name = tensor("transpose_112_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_112 = transpose(perm = transpose_112_perm_0, x = keys_129_cast_fp16)[name = tensor("transpose_120")]; + tensor transpose_111 = transpose(perm = transpose_111_perm_0, x = q_129_cast_fp16)[name = tensor("transpose_121")]; + tensor var_9186_cast_fp16 = matmul(transpose_x = var_9186_transpose_x_0, transpose_y = var_9186_transpose_y_0, x = transpose_111, y = transpose_112)[name = tensor("op_9186_cast_fp16")]; + tensor var_9180_to_fp16 = const()[name = tensor("op_9180_to_fp16"), val = tensor([0x1p-3])]; + tensor attn_weights_127_cast_fp16 = mul(x = var_9186_cast_fp16, y = var_9180_to_fp16)[name = tensor("attn_weights_127_cast_fp16")]; + tensor var_9188 = logical_not(x = attn_mask_43)[name = tensor("op_9188")]; + tensor var_9189_to_fp16 = const()[name = tensor("op_9189_to_fp16"), val = tensor(-inf)]; + tensor attn_weights_129_cast_fp16 = select(a = var_9189_to_fp16, b = attn_weights_127_cast_fp16, cond = var_9188)[name = tensor("attn_weights_129_cast_fp16")]; + tensor var_9191 = const()[name = tensor("op_9191"), val = tensor(-1)]; + tensor attn_weights_131_cast_fp16 = softmax(axis = var_9191, x = attn_weights_129_cast_fp16)[name = tensor("attn_weights_131_cast_fp16")]; + tensor attn_output_43_transpose_x_0 = const()[name = tensor("attn_output_43_transpose_x_0"), val = tensor(false)]; + tensor attn_output_43_transpose_y_0 = const()[name = tensor("attn_output_43_transpose_y_0"), val = tensor(false)]; + tensor values_131_cast_fp16 = transpose(perm = var_9127, x = values_129_cast_fp16)[name = tensor("transpose_122")]; + tensor attn_output_43_cast_fp16 = matmul(transpose_x = attn_output_43_transpose_x_0, transpose_y = attn_output_43_transpose_y_0, x = attn_weights_131_cast_fp16, y = values_131_cast_fp16)[name = tensor("attn_output_43_cast_fp16")]; + tensor var_9199 = const()[name = tensor("op_9199"), val = tensor([0, 2, 1, 3])]; + tensor var_9204 = const()[name = tensor("op_9204"), val = tensor([1, 256, 1024])]; + tensor var_9200_cast_fp16 = transpose(perm = var_9199, x = attn_output_43_cast_fp16)[name = tensor("transpose_119")]; + tensor input_257_cast_fp16 = reshape(shape = var_9204, x = var_9200_cast_fp16)[name = tensor("input_257_cast_fp16")]; + tensor attn21_out_proj_weight_to_fp16 = const()[name = tensor("attn21_out_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(536031744)))]; + tensor linear_85_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = attn21_out_proj_weight_to_fp16, x = input_257_cast_fp16)[name = tensor("linear_85_cast_fp16")]; + tensor var_9213_cast_fp16 = add(x = position21_to_fp16, y = valid_len_to_fp16)[name = tensor("op_9213_cast_fp16")]; + tensor var_9213_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_9213_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor input_259_cast_fp16 = add(x = input_253_cast_fp16, y = linear_85_cast_fp16)[name = tensor("input_259_cast_fp16")]; + tensor input_261_axes_0 = const()[name = tensor("input_261_axes_0"), val = tensor([-1])]; + tensor norm21_2_weight_to_fp16 = const()[name = tensor("norm21_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(538128960)))]; + tensor norm21_2_bias_to_fp16 = const()[name = tensor("norm21_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(538131072)))]; + tensor var_9217_to_fp16 = const()[name = tensor("op_9217_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_261_cast_fp16 = layer_norm(axes = input_261_axes_0, beta = norm21_2_bias_to_fp16, epsilon = var_9217_to_fp16, gamma = norm21_2_weight_to_fp16, x = input_259_cast_fp16)[name = tensor("input_261_cast_fp16")]; + tensor linear21_1_weight_to_fp16 = const()[name = tensor("linear21_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(538133184)))]; + tensor linear_86_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = linear21_1_weight_to_fp16, x = input_261_cast_fp16)[name = tensor("linear_86_cast_fp16")]; + tensor input_263_mode_0 = const()[name = tensor("input_263_mode_0"), val = tensor("EXACT")]; + tensor input_263_cast_fp16 = gelu(mode = input_263_mode_0, x = linear_86_cast_fp16)[name = tensor("input_263_cast_fp16")]; + tensor linear21_2_weight_to_fp16 = const()[name = tensor("linear21_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(546521856)))]; + tensor linear_87_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = linear21_2_weight_to_fp16, x = input_263_cast_fp16)[name = tensor("linear_87_cast_fp16")]; + tensor input_265_cast_fp16 = add(x = input_259_cast_fp16, y = linear_87_cast_fp16)[name = tensor("input_265_cast_fp16")]; + tensor input_267_axes_0 = const()[name = tensor("input_267_axes_0"), val = tensor([-1])]; + tensor norm22_1_weight_to_fp16 = const()[name = tensor("norm22_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(554910528)))]; + tensor norm22_1_bias_to_fp16 = const()[name = tensor("norm22_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(554912640)))]; + tensor var_9234_to_fp16 = const()[name = tensor("op_9234_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_267_cast_fp16 = layer_norm(axes = input_267_axes_0, beta = norm22_1_bias_to_fp16, epsilon = var_9234_to_fp16, gamma = norm22_1_weight_to_fp16, x = input_265_cast_fp16)[name = tensor("input_267_cast_fp16")]; + tensor attn22_in_proj_weight_to_fp16 = const()[name = tensor("attn22_in_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(554914752)))]; + tensor linear_88_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = attn22_in_proj_weight_to_fp16, x = input_267_cast_fp16)[name = tensor("linear_88_cast_fp16")]; + tensor var_9248 = const()[name = tensor("op_9248"), val = tensor([1, 256, 3, 16, 64])]; + tensor qkv_45_cast_fp16 = reshape(shape = var_9248, x = linear_88_cast_fp16)[name = tensor("qkv_45_cast_fp16")]; + tensor q_133_begin_0 = const()[name = tensor("q_133_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor q_133_end_0 = const()[name = tensor("q_133_end_0"), val = tensor([1, 256, 1, 16, 64])]; + tensor q_133_end_mask_0 = const()[name = tensor("q_133_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor q_133_squeeze_mask_0 = const()[name = tensor("q_133_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor q_133_cast_fp16 = slice_by_index(begin = q_133_begin_0, end = q_133_end_0, end_mask = q_133_end_mask_0, squeeze_mask = q_133_squeeze_mask_0, x = qkv_45_cast_fp16)[name = tensor("q_133_cast_fp16")]; + tensor k_89_begin_0 = const()[name = tensor("k_89_begin_0"), val = tensor([0, 0, 1, 0, 0])]; + tensor k_89_end_0 = const()[name = tensor("k_89_end_0"), val = tensor([1, 256, 2, 16, 64])]; + tensor k_89_end_mask_0 = const()[name = tensor("k_89_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor k_89_squeeze_mask_0 = const()[name = tensor("k_89_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor k_89_cast_fp16 = slice_by_index(begin = k_89_begin_0, end = k_89_end_0, end_mask = k_89_end_mask_0, squeeze_mask = k_89_squeeze_mask_0, x = qkv_45_cast_fp16)[name = tensor("k_89_cast_fp16")]; + tensor v_45_begin_0 = const()[name = tensor("v_45_begin_0"), val = tensor([0, 0, 2, 0, 0])]; + tensor v_45_end_0 = const()[name = tensor("v_45_end_0"), val = tensor([1, 256, 3, 16, 64])]; + tensor v_45_end_mask_0 = const()[name = tensor("v_45_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor v_45_squeeze_mask_0 = const()[name = tensor("v_45_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor v_45_cast_fp16 = slice_by_index(begin = v_45_begin_0, end = v_45_end_0, end_mask = v_45_end_mask_0, squeeze_mask = v_45_squeeze_mask_0, x = qkv_45_cast_fp16)[name = tensor("v_45_cast_fp16")]; + tensor var_9306 = const()[name = tensor("op_9306"), val = tensor([1, 1])]; + tensor position22_to_fp16_dtype_0 = const()[name = tensor("position22_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor position22_to_fp16 = cast(dtype = position22_to_fp16_dtype_0, x = position22)[name = tensor("cast_551")]; + tensor var_9307_cast_fp16 = reshape(shape = var_9306, x = position22_to_fp16)[name = tensor("op_9307_cast_fp16")]; + tensor ts_135_cast_fp16 = add(x = ts_1_promoted_to_fp16, y = var_9307_cast_fp16)[name = tensor("ts_135_cast_fp16")]; + tensor var_9314 = const()[name = tensor("op_9314"), val = tensor([1, 256, 1, 1])]; + tensor ts_137_cast_fp16 = reshape(shape = var_9314, x = ts_135_cast_fp16)[name = tensor("ts_137_cast_fp16")]; + tensor var_9321 = const()[name = tensor("op_9321"), val = tensor([1, 256, 16, 32, 2])]; + tensor q_complex_45_cast_fp16 = reshape(shape = var_9321, x = q_133_cast_fp16)[name = tensor("q_complex_45_cast_fp16")]; + tensor var_9328 = const()[name = tensor("op_9328"), val = tensor([1, 256, 16, 32, 2])]; + tensor k_complex_45_cast_fp16 = reshape(shape = var_9328, x = k_89_cast_fp16)[name = tensor("k_complex_45_cast_fp16")]; + tensor var_9332_begin_0 = const()[name = tensor("op_9332_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_9332_end_0 = const()[name = tensor("op_9332_end_0"), val = tensor([1, 256, 16, 32, 1])]; + tensor var_9332_end_mask_0 = const()[name = tensor("op_9332_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_9332_squeeze_mask_0 = const()[name = tensor("op_9332_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_9332_cast_fp16 = slice_by_index(begin = var_9332_begin_0, end = var_9332_end_0, end_mask = var_9332_end_mask_0, squeeze_mask = var_9332_squeeze_mask_0, x = q_complex_45_cast_fp16)[name = tensor("op_9332_cast_fp16")]; + tensor var_9340_begin_0 = const()[name = tensor("op_9340_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_9340_end_0 = const()[name = tensor("op_9340_end_0"), val = tensor([1, 256, 16, 32, 2])]; + tensor var_9340_end_mask_0 = const()[name = tensor("op_9340_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_9340_squeeze_mask_0 = const()[name = tensor("op_9340_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_9340_cast_fp16 = slice_by_index(begin = var_9340_begin_0, end = var_9340_end_0, end_mask = var_9340_end_mask_0, squeeze_mask = var_9340_squeeze_mask_0, x = q_complex_45_cast_fp16)[name = tensor("op_9340_cast_fp16")]; + tensor var_9348_begin_0 = const()[name = tensor("op_9348_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_9348_end_0 = const()[name = tensor("op_9348_end_0"), val = tensor([1, 256, 16, 32, 1])]; + tensor var_9348_end_mask_0 = const()[name = tensor("op_9348_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_9348_squeeze_mask_0 = const()[name = tensor("op_9348_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_9348_cast_fp16 = slice_by_index(begin = var_9348_begin_0, end = var_9348_end_0, end_mask = var_9348_end_mask_0, squeeze_mask = var_9348_squeeze_mask_0, x = k_complex_45_cast_fp16)[name = tensor("op_9348_cast_fp16")]; + tensor var_9356_begin_0 = const()[name = tensor("op_9356_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_9356_end_0 = const()[name = tensor("op_9356_end_0"), val = tensor([1, 256, 16, 32, 2])]; + tensor var_9356_end_mask_0 = const()[name = tensor("op_9356_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_9356_squeeze_mask_0 = const()[name = tensor("op_9356_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_9356_cast_fp16 = slice_by_index(begin = var_9356_begin_0, end = var_9356_end_0, end_mask = var_9356_end_mask_0, squeeze_mask = var_9356_squeeze_mask_0, x = k_complex_45_cast_fp16)[name = tensor("op_9356_cast_fp16")]; + tensor freqs_45_to_fp16 = const()[name = tensor("freqs_45_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(561206272)))]; + tensor var_9362_cast_fp16 = mul(x = freqs_45_to_fp16, y = ts_137_cast_fp16)[name = tensor("op_9362_cast_fp16")]; + tensor rotr_45_cast_fp16 = cos(x = var_9362_cast_fp16)[name = tensor("rotr_45_cast_fp16")]; + tensor roti_45_cast_fp16 = sin(x = var_9362_cast_fp16)[name = tensor("roti_45_cast_fp16")]; + tensor var_9366_cast_fp16 = mul(x = var_9332_cast_fp16, y = rotr_45_cast_fp16)[name = tensor("op_9366_cast_fp16")]; + tensor var_9367_cast_fp16 = mul(x = var_9340_cast_fp16, y = roti_45_cast_fp16)[name = tensor("op_9367_cast_fp16")]; + tensor qor_89_cast_fp16 = sub(x = var_9366_cast_fp16, y = var_9367_cast_fp16)[name = tensor("qor_89_cast_fp16")]; + tensor var_9370_cast_fp16 = mul(x = var_9332_cast_fp16, y = roti_45_cast_fp16)[name = tensor("op_9370_cast_fp16")]; + tensor var_9371_cast_fp16 = mul(x = var_9340_cast_fp16, y = rotr_45_cast_fp16)[name = tensor("op_9371_cast_fp16")]; + tensor qoi_89_cast_fp16 = add(x = var_9370_cast_fp16, y = var_9371_cast_fp16)[name = tensor("qoi_89_cast_fp16")]; + tensor var_9374_cast_fp16 = mul(x = var_9348_cast_fp16, y = rotr_45_cast_fp16)[name = tensor("op_9374_cast_fp16")]; + tensor var_9375_cast_fp16 = mul(x = var_9356_cast_fp16, y = roti_45_cast_fp16)[name = tensor("op_9375_cast_fp16")]; + tensor kor_89_cast_fp16 = sub(x = var_9374_cast_fp16, y = var_9375_cast_fp16)[name = tensor("kor_89_cast_fp16")]; + tensor var_9378_cast_fp16 = mul(x = var_9348_cast_fp16, y = roti_45_cast_fp16)[name = tensor("op_9378_cast_fp16")]; + tensor var_9379_cast_fp16 = mul(x = var_9356_cast_fp16, y = rotr_45_cast_fp16)[name = tensor("op_9379_cast_fp16")]; + tensor koi_89_cast_fp16 = add(x = var_9378_cast_fp16, y = var_9379_cast_fp16)[name = tensor("koi_89_cast_fp16")]; + tensor qo_45_axis_0 = const()[name = tensor("qo_45_axis_0"), val = tensor(-1)]; + tensor qo_45_cast_fp16 = stack(axis = qo_45_axis_0, values = (qor_89_cast_fp16, qoi_89_cast_fp16))[name = tensor("qo_45_cast_fp16")]; + tensor ko_45_axis_0 = const()[name = tensor("ko_45_axis_0"), val = tensor(-1)]; + tensor ko_45_cast_fp16 = stack(axis = ko_45_axis_0, values = (kor_89_cast_fp16, koi_89_cast_fp16))[name = tensor("ko_45_cast_fp16")]; + tensor var_9412 = const()[name = tensor("op_9412"), val = tensor([1, 256, 16, 64])]; + tensor q_135_cast_fp16 = reshape(shape = var_9412, x = qo_45_cast_fp16)[name = tensor("q_135_cast_fp16")]; + tensor var_9418 = const()[name = tensor("op_9418"), val = tensor([1, 256, 16, 64])]; + tensor k_91_cast_fp16 = reshape(shape = var_9418, x = ko_45_cast_fp16)[name = tensor("k_91_cast_fp16")]; + tensor write_indices_float_133_cast_fp16 = add(x = var_9307_cast_fp16, y = var_393_promoted_to_fp16)[name = tensor("write_indices_float_133_cast_fp16")]; + tensor _inversed_9439_y_0_to_fp16 = const()[name = tensor("_inversed_9439_y_0_to_fp16"), val = tensor(0x1p-9)]; + tensor _inversed_9439_cast_fp16 = mul(x = write_indices_float_133_cast_fp16, y = _inversed_9439_y_0_to_fp16)[name = tensor("_inversed_9439_cast_fp16")]; + tensor var_9440_cast_fp16 = floor(x = _inversed_9439_cast_fp16)[name = tensor("op_9440_cast_fp16")]; + tensor var_9441_to_fp16 = const()[name = tensor("op_9441_to_fp16"), val = tensor(0x1p+9)]; + tensor var_9442_cast_fp16 = mul(x = var_9440_cast_fp16, y = var_9441_to_fp16)[name = tensor("op_9442_cast_fp16")]; + tensor write_indices_float_135_cast_fp16 = sub(x = write_indices_float_133_cast_fp16, y = var_9442_cast_fp16)[name = tensor("write_indices_float_135_cast_fp16")]; + tensor var_9460_value_0_to_fp16 = const()[name = tensor("op_9460_value_0_to_fp16"), val = tensor(0x1.ffp+8)]; + tensor var_9460_cast_fp16 = fill_like(ref_tensor = write_indices_float_135_cast_fp16, value = var_9460_value_0_to_fp16)[name = tensor("op_9460_cast_fp16")]; + tensor write_indices_float_137_cast_fp16 = select(a = write_indices_float_135_cast_fp16, b = var_9460_cast_fp16, cond = token_valid_1_cast_fp16)[name = tensor("write_indices_float_137_cast_fp16")]; + tensor var_9466_dtype_0 = const()[name = tensor("op_9466_dtype_0"), val = tensor("int32")]; + tensor var_9471 = const()[name = tensor("op_9471"), val = tensor([1, 256, 1, 1])]; + tensor write_indices_float_137_cast_fp16_to_int32 = cast(dtype = var_9466_dtype_0, x = write_indices_float_137_cast_fp16)[name = tensor("cast_550")]; + tensor var_9472 = reshape(shape = var_9471, x = write_indices_float_137_cast_fp16_to_int32)[name = tensor("op_9472")]; + tensor write_indices_45_reps_0 = const()[name = tensor("write_indices_45_reps_0"), val = tensor([1, 1, 16, 64])]; + tensor write_indices_45 = tile(reps = write_indices_45_reps_0, x = var_9472)[name = tensor("write_indices_45")]; + tensor var_9482_begin_0 = const()[name = tensor("op_9482_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_9482_end_0 = const()[name = tensor("op_9482_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor var_9482_end_mask_0 = const()[name = tensor("op_9482_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_9482_squeeze_mask_0 = const()[name = tensor("op_9482_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor cache22_to_fp16_dtype_0 = const()[name = tensor("cache22_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor cache22_to_fp16 = cast(dtype = cache22_to_fp16_dtype_0, x = cache22)[name = tensor("cast_549")]; + tensor var_9482_cast_fp16 = slice_by_index(begin = var_9482_begin_0, end = var_9482_end_0, end_mask = var_9482_end_mask_0, squeeze_mask = var_9482_squeeze_mask_0, x = cache22_to_fp16)[name = tensor("op_9482_cast_fp16")]; + tensor var_9484_axis_0 = const()[name = tensor("op_9484_axis_0"), val = tensor(1)]; + tensor var_9484_mode_0 = const()[name = tensor("op_9484_mode_0"), val = tensor("update")]; + tensor var_9484_validate_indices_0 = const()[name = tensor("op_9484_validate_indices_0"), val = tensor(false)]; + tensor var_9484_cast_fp16 = scatter_along_axis(axis = var_9484_axis_0, data = var_9482_cast_fp16, indices = write_indices_45, mode = var_9484_mode_0, updates = k_91_cast_fp16, validate_indices = var_9484_validate_indices_0)[name = tensor("op_9484_cast_fp16")]; + tensor concat_178 = const()[name = tensor("concat_178"), val = tensor([0, 0, 0, 0, 0])]; + tensor concat_179 = const()[name = tensor("concat_179"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_45_internal_tensor_assign_1_stride_0 = const()[name = tensor("new_cache_45_internal_tensor_assign_1_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_45_internal_tensor_assign_1_begin_mask_0 = const()[name = tensor("new_cache_45_internal_tensor_assign_1_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_45_internal_tensor_assign_1_end_mask_0 = const()[name = tensor("new_cache_45_internal_tensor_assign_1_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_45_internal_tensor_assign_1_squeeze_mask_0 = const()[name = tensor("new_cache_45_internal_tensor_assign_1_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_90 = const()[name = tensor("shape_90"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_44 = const()[name = tensor("reduce_prod_44"), val = tensor(1048576)]; + tensor range_1d_44_start_0 = const()[name = tensor("range_1d_44_start_0"), val = tensor(0)]; + tensor range_1d_44_step_0 = const()[name = tensor("range_1d_44_step_0"), val = tensor(1)]; + tensor range_1d_44 = range_1d(end = reduce_prod_44, start = range_1d_44_start_0, step = range_1d_44_step_0)[name = tensor("range_1d_44")]; + tensor reshape_220 = reshape(shape = shape_90, x = range_1d_44)[name = tensor("reshape_220")]; + tensor slice_by_index_44 = slice_by_index(begin = concat_178, begin_mask = new_cache_45_internal_tensor_assign_1_begin_mask_0, end = concat_179, end_mask = new_cache_45_internal_tensor_assign_1_end_mask_0, squeeze_mask = new_cache_45_internal_tensor_assign_1_squeeze_mask_0, stride = new_cache_45_internal_tensor_assign_1_stride_0, x = reshape_220)[name = tensor("slice_by_index_44")]; + tensor reshape_221_shape_0 = const()[name = tensor("reshape_221_shape_0"), val = tensor([-1])]; + tensor reshape_221 = reshape(shape = reshape_221_shape_0, x = slice_by_index_44)[name = tensor("reshape_221")]; + tensor reshape_222_shape_0 = const()[name = tensor("reshape_222_shape_0"), val = tensor([-1])]; + tensor reshape_222_cast_fp16 = reshape(shape = reshape_222_shape_0, x = var_9484_cast_fp16)[name = tensor("reshape_222_cast_fp16")]; + tensor reshape_223_shape_0 = const()[name = tensor("reshape_223_shape_0"), val = tensor([-1])]; + tensor reshape_223_cast_fp16 = reshape(shape = reshape_223_shape_0, x = cache22_to_fp16)[name = tensor("reshape_223_cast_fp16")]; + tensor scatter_44_mode_0 = const()[name = tensor("scatter_44_mode_0"), val = tensor("update")]; + tensor scatter_44_axis_0 = const()[name = tensor("scatter_44_axis_0"), val = tensor(0)]; + tensor scatter_44_validate_indices_0 = const()[name = tensor("scatter_44_validate_indices_0"), val = tensor(false)]; + tensor scatter_44_cast_fp16 = scatter(axis = scatter_44_axis_0, data = reshape_223_cast_fp16, indices = reshape_221, mode = scatter_44_mode_0, updates = reshape_222_cast_fp16, validate_indices = scatter_44_validate_indices_0)[name = tensor("scatter_44_cast_fp16")]; + tensor reshape_224_cast_fp16 = reshape(shape = shape_90, x = scatter_44_cast_fp16)[name = tensor("reshape_224_cast_fp16")]; + tensor var_9492_begin_0 = const()[name = tensor("op_9492_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor var_9492_end_0 = const()[name = tensor("op_9492_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor var_9492_end_mask_0 = const()[name = tensor("op_9492_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_9492_squeeze_mask_0 = const()[name = tensor("op_9492_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor var_9492_cast_fp16 = slice_by_index(begin = var_9492_begin_0, end = var_9492_end_0, end_mask = var_9492_end_mask_0, squeeze_mask = var_9492_squeeze_mask_0, x = reshape_224_cast_fp16)[name = tensor("op_9492_cast_fp16")]; + tensor var_9494_axis_0 = const()[name = tensor("op_9494_axis_0"), val = tensor(1)]; + tensor var_9494_mode_0 = const()[name = tensor("op_9494_mode_0"), val = tensor("update")]; + tensor var_9494_validate_indices_0 = const()[name = tensor("op_9494_validate_indices_0"), val = tensor(false)]; + tensor var_9494_cast_fp16 = scatter_along_axis(axis = var_9494_axis_0, data = var_9492_cast_fp16, indices = write_indices_45, mode = var_9494_mode_0, updates = v_45_cast_fp16, validate_indices = var_9494_validate_indices_0)[name = tensor("op_9494_cast_fp16")]; + tensor concat_180 = const()[name = tensor("concat_180"), val = tensor([1, 0, 0, 0, 0])]; + tensor concat_181 = const()[name = tensor("concat_181"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_45_internal_tensor_assign_2_stride_0 = const()[name = tensor("new_cache_45_internal_tensor_assign_2_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_45_internal_tensor_assign_2_begin_mask_0 = const()[name = tensor("new_cache_45_internal_tensor_assign_2_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_45_internal_tensor_assign_2_end_mask_0 = const()[name = tensor("new_cache_45_internal_tensor_assign_2_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_45_internal_tensor_assign_2_squeeze_mask_0 = const()[name = tensor("new_cache_45_internal_tensor_assign_2_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_91 = const()[name = tensor("shape_91"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_45 = const()[name = tensor("reduce_prod_45"), val = tensor(1048576)]; + tensor range_1d_45_start_0 = const()[name = tensor("range_1d_45_start_0"), val = tensor(0)]; + tensor range_1d_45_step_0 = const()[name = tensor("range_1d_45_step_0"), val = tensor(1)]; + tensor range_1d_45 = range_1d(end = reduce_prod_45, start = range_1d_45_start_0, step = range_1d_45_step_0)[name = tensor("range_1d_45")]; + tensor reshape_225 = reshape(shape = shape_91, x = range_1d_45)[name = tensor("reshape_225")]; + tensor slice_by_index_45 = slice_by_index(begin = concat_180, begin_mask = new_cache_45_internal_tensor_assign_2_begin_mask_0, end = concat_181, end_mask = new_cache_45_internal_tensor_assign_2_end_mask_0, squeeze_mask = new_cache_45_internal_tensor_assign_2_squeeze_mask_0, stride = new_cache_45_internal_tensor_assign_2_stride_0, x = reshape_225)[name = tensor("slice_by_index_45")]; + tensor reshape_226_shape_0 = const()[name = tensor("reshape_226_shape_0"), val = tensor([-1])]; + tensor reshape_226 = reshape(shape = reshape_226_shape_0, x = slice_by_index_45)[name = tensor("reshape_226")]; + tensor reshape_227_shape_0 = const()[name = tensor("reshape_227_shape_0"), val = tensor([-1])]; + tensor reshape_227_cast_fp16 = reshape(shape = reshape_227_shape_0, x = var_9494_cast_fp16)[name = tensor("reshape_227_cast_fp16")]; + tensor reshape_228_shape_0 = const()[name = tensor("reshape_228_shape_0"), val = tensor([-1])]; + tensor reshape_228_cast_fp16 = reshape(shape = reshape_228_shape_0, x = reshape_224_cast_fp16)[name = tensor("reshape_228_cast_fp16")]; + tensor scatter_45_mode_0 = const()[name = tensor("scatter_45_mode_0"), val = tensor("update")]; + tensor scatter_45_axis_0 = const()[name = tensor("scatter_45_axis_0"), val = tensor(0)]; + tensor scatter_45_validate_indices_0 = const()[name = tensor("scatter_45_validate_indices_0"), val = tensor(false)]; + tensor scatter_45_cast_fp16 = scatter(axis = scatter_45_axis_0, data = reshape_228_cast_fp16, indices = reshape_226, mode = scatter_45_mode_0, updates = reshape_227_cast_fp16, validate_indices = scatter_45_validate_indices_0)[name = tensor("scatter_45_cast_fp16")]; + tensor reshape_229_cast_fp16 = reshape(shape = shape_91, x = scatter_45_cast_fp16)[name = tensor("reshape_229_cast_fp16")]; + tensor reshape_229_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("reshape_229_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor keys_133_begin_0 = const()[name = tensor("keys_133_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor keys_133_end_0 = const()[name = tensor("keys_133_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor keys_133_end_mask_0 = const()[name = tensor("keys_133_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor keys_133_squeeze_mask_0 = const()[name = tensor("keys_133_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor keys_133_cast_fp16 = slice_by_index(begin = keys_133_begin_0, end = keys_133_end_0, end_mask = keys_133_end_mask_0, squeeze_mask = keys_133_squeeze_mask_0, x = reshape_229_cast_fp16)[name = tensor("keys_133_cast_fp16")]; + tensor values_133_begin_0 = const()[name = tensor("values_133_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor values_133_end_0 = const()[name = tensor("values_133_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor values_133_end_mask_0 = const()[name = tensor("values_133_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor values_133_squeeze_mask_0 = const()[name = tensor("values_133_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor values_133_cast_fp16 = slice_by_index(begin = values_133_begin_0, end = values_133_end_0, end_mask = values_133_end_mask_0, squeeze_mask = values_133_squeeze_mask_0, x = reshape_229_cast_fp16)[name = tensor("values_133_cast_fp16")]; + tensor var_9506_cast_fp16 = not_equal(x = keys_133_cast_fp16, y = keys_133_cast_fp16)[name = tensor("op_9506_cast_fp16")]; + tensor keys_135_cast_fp16 = select(a = var_470_to_fp16, b = keys_133_cast_fp16, cond = var_9506_cast_fp16)[name = tensor("keys_135_cast_fp16")]; + tensor var_9514_cast_fp16 = not_equal(x = values_133_cast_fp16, y = values_133_cast_fp16)[name = tensor("op_9514_cast_fp16")]; + tensor values_135_cast_fp16 = select(a = var_470_to_fp16, b = values_133_cast_fp16, cond = var_9514_cast_fp16)[name = tensor("values_135_cast_fp16")]; + tensor var_9538 = const()[name = tensor("op_9538"), val = tensor([0, 2, 1, 3])]; + tensor var_9554 = const()[name = tensor("op_9554"), val = tensor([1, 1, 1])]; + tensor var_9555_cast_fp16 = reshape(shape = var_9554, x = position22_to_fp16)[name = tensor("op_9555_cast_fp16")]; + tensor q_positions_45_cast_fp16 = add(x = var_9555_cast_fp16, y = q_offsets_1_promoted_to_fp16)[name = tensor("q_positions_45_cast_fp16")]; + tensor valid_end_45_cast_fp16 = add(x = var_9555_cast_fp16, y = var_536_cast_fp16)[name = tensor("valid_end_45_cast_fp16")]; + tensor valid_mask_45_cast_fp16 = less(x = k_positions_1_promoted_to_fp16, y = valid_end_45_cast_fp16)[name = tensor("valid_mask_45_cast_fp16")]; + tensor causal_mask_45_cast_fp16 = less_equal(x = k_positions_1_promoted_to_fp16, y = q_positions_45_cast_fp16)[name = tensor("causal_mask_45_cast_fp16")]; + tensor var_9583 = logical_and(x = valid_mask_45_cast_fp16, y = causal_mask_45_cast_fp16)[name = tensor("op_9583")]; + tensor attn_mask_45_axes_0 = const()[name = tensor("attn_mask_45_axes_0"), val = tensor([1])]; + tensor attn_mask_45 = expand_dims(axes = attn_mask_45_axes_0, x = var_9583)[name = tensor("attn_mask_45")]; + tensor var_9597_transpose_x_0 = const()[name = tensor("op_9597_transpose_x_0"), val = tensor(false)]; + tensor var_9597_transpose_y_0 = const()[name = tensor("op_9597_transpose_y_0"), val = tensor(false)]; + tensor transpose_113_perm_0 = const()[name = tensor("transpose_113_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_114_perm_0 = const()[name = tensor("transpose_114_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_114 = transpose(perm = transpose_114_perm_0, x = keys_135_cast_fp16)[name = tensor("transpose_116")]; + tensor transpose_113 = transpose(perm = transpose_113_perm_0, x = q_135_cast_fp16)[name = tensor("transpose_117")]; + tensor var_9597_cast_fp16 = matmul(transpose_x = var_9597_transpose_x_0, transpose_y = var_9597_transpose_y_0, x = transpose_113, y = transpose_114)[name = tensor("op_9597_cast_fp16")]; + tensor var_9591_to_fp16 = const()[name = tensor("op_9591_to_fp16"), val = tensor([0x1p-3])]; + tensor attn_weights_133_cast_fp16 = mul(x = var_9597_cast_fp16, y = var_9591_to_fp16)[name = tensor("attn_weights_133_cast_fp16")]; + tensor var_9599 = logical_not(x = attn_mask_45)[name = tensor("op_9599")]; + tensor var_9600_to_fp16 = const()[name = tensor("op_9600_to_fp16"), val = tensor(-inf)]; + tensor attn_weights_135_cast_fp16 = select(a = var_9600_to_fp16, b = attn_weights_133_cast_fp16, cond = var_9599)[name = tensor("attn_weights_135_cast_fp16")]; + tensor var_9602 = const()[name = tensor("op_9602"), val = tensor(-1)]; + tensor attn_weights_137_cast_fp16 = softmax(axis = var_9602, x = attn_weights_135_cast_fp16)[name = tensor("attn_weights_137_cast_fp16")]; + tensor attn_output_45_transpose_x_0 = const()[name = tensor("attn_output_45_transpose_x_0"), val = tensor(false)]; + tensor attn_output_45_transpose_y_0 = const()[name = tensor("attn_output_45_transpose_y_0"), val = tensor(false)]; + tensor values_137_cast_fp16 = transpose(perm = var_9538, x = values_135_cast_fp16)[name = tensor("transpose_118")]; + tensor attn_output_45_cast_fp16 = matmul(transpose_x = attn_output_45_transpose_x_0, transpose_y = attn_output_45_transpose_y_0, x = attn_weights_137_cast_fp16, y = values_137_cast_fp16)[name = tensor("attn_output_45_cast_fp16")]; + tensor var_9610 = const()[name = tensor("op_9610"), val = tensor([0, 2, 1, 3])]; + tensor var_9615 = const()[name = tensor("op_9615"), val = tensor([1, 256, 1024])]; + tensor var_9611_cast_fp16 = transpose(perm = var_9610, x = attn_output_45_cast_fp16)[name = tensor("transpose_115")]; + tensor input_269_cast_fp16 = reshape(shape = var_9615, x = var_9611_cast_fp16)[name = tensor("input_269_cast_fp16")]; + tensor attn22_out_proj_weight_to_fp16 = const()[name = tensor("attn22_out_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(561206400)))]; + tensor linear_89_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = attn22_out_proj_weight_to_fp16, x = input_269_cast_fp16)[name = tensor("linear_89_cast_fp16")]; + tensor var_9624_cast_fp16 = add(x = position22_to_fp16, y = valid_len_to_fp16)[name = tensor("op_9624_cast_fp16")]; + tensor var_9624_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_9624_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor input_271_cast_fp16 = add(x = input_265_cast_fp16, y = linear_89_cast_fp16)[name = tensor("input_271_cast_fp16")]; + tensor input_273_axes_0 = const()[name = tensor("input_273_axes_0"), val = tensor([-1])]; + tensor norm22_2_weight_to_fp16 = const()[name = tensor("norm22_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(563303616)))]; + tensor norm22_2_bias_to_fp16 = const()[name = tensor("norm22_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(563305728)))]; + tensor var_9628_to_fp16 = const()[name = tensor("op_9628_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_273_cast_fp16 = layer_norm(axes = input_273_axes_0, beta = norm22_2_bias_to_fp16, epsilon = var_9628_to_fp16, gamma = norm22_2_weight_to_fp16, x = input_271_cast_fp16)[name = tensor("input_273_cast_fp16")]; + tensor linear22_1_weight_to_fp16 = const()[name = tensor("linear22_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(563307840)))]; + tensor linear_90_cast_fp16 = linear(bias = linear_2_bias_0_to_fp16, weight = linear22_1_weight_to_fp16, x = input_273_cast_fp16)[name = tensor("linear_90_cast_fp16")]; + tensor input_275_mode_0 = const()[name = tensor("input_275_mode_0"), val = tensor("EXACT")]; + tensor input_275_cast_fp16 = gelu(mode = input_275_mode_0, x = linear_90_cast_fp16)[name = tensor("input_275_cast_fp16")]; + tensor linear22_2_weight_to_fp16 = const()[name = tensor("linear22_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(571696512)))]; + tensor linear_91_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = linear22_2_weight_to_fp16, x = input_275_cast_fp16)[name = tensor("linear_91_cast_fp16")]; + tensor input_277_cast_fp16 = add(x = input_271_cast_fp16, y = linear_91_cast_fp16)[name = tensor("input_277_cast_fp16")]; + tensor input_279_axes_0 = const()[name = tensor("input_279_axes_0"), val = tensor([-1])]; + tensor norm23_1_weight_to_fp16 = const()[name = tensor("norm23_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(580085184)))]; + tensor norm23_1_bias_to_fp16 = const()[name = tensor("norm23_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(580087296)))]; + tensor var_9645_to_fp16 = const()[name = tensor("op_9645_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_279_cast_fp16 = layer_norm(axes = input_279_axes_0, beta = norm23_1_bias_to_fp16, epsilon = var_9645_to_fp16, gamma = norm23_1_weight_to_fp16, x = input_277_cast_fp16)[name = tensor("input_279_cast_fp16")]; + tensor attn23_in_proj_weight_to_fp16 = const()[name = tensor("attn23_in_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(580089408)))]; + tensor linear_92_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = attn23_in_proj_weight_to_fp16, x = input_279_cast_fp16)[name = tensor("linear_92_cast_fp16")]; + tensor var_9659 = const()[name = tensor("op_9659"), val = tensor([1, 256, 3, 16, 64])]; + tensor qkv_cast_fp16 = reshape(shape = var_9659, x = linear_92_cast_fp16)[name = tensor("qkv_cast_fp16")]; + tensor k_93_begin_0 = const()[name = tensor("k_93_begin_0"), val = tensor([0, 0, 1, 0, 0])]; + tensor k_93_end_0 = const()[name = tensor("k_93_end_0"), val = tensor([1, 256, 2, 16, 64])]; + tensor k_93_end_mask_0 = const()[name = tensor("k_93_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor k_93_squeeze_mask_0 = const()[name = tensor("k_93_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor k_93_cast_fp16 = slice_by_index(begin = k_93_begin_0, end = k_93_end_0, end_mask = k_93_end_mask_0, squeeze_mask = k_93_squeeze_mask_0, x = qkv_cast_fp16)[name = tensor("k_93_cast_fp16")]; + tensor v_begin_0 = const()[name = tensor("v_begin_0"), val = tensor([0, 0, 2, 0, 0])]; + tensor v_end_0 = const()[name = tensor("v_end_0"), val = tensor([1, 256, 3, 16, 64])]; + tensor v_end_mask_0 = const()[name = tensor("v_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor v_squeeze_mask_0 = const()[name = tensor("v_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor v_cast_fp16 = slice_by_index(begin = v_begin_0, end = v_end_0, end_mask = v_end_mask_0, squeeze_mask = v_squeeze_mask_0, x = qkv_cast_fp16)[name = tensor("v_cast_fp16")]; + tensor var_9704 = const()[name = tensor("op_9704"), val = tensor([1, 1])]; + tensor position23_to_fp16_dtype_0 = const()[name = tensor("position23_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor position23_to_fp16 = cast(dtype = position23_to_fp16_dtype_0, x = position23)[name = tensor("cast_546")]; + tensor var_9705_cast_fp16 = reshape(shape = var_9704, x = position23_to_fp16)[name = tensor("op_9705_cast_fp16")]; + tensor ts_141_cast_fp16 = add(x = ts_1_promoted_to_fp16, y = var_9705_cast_fp16)[name = tensor("ts_141_cast_fp16")]; + tensor var_9712 = const()[name = tensor("op_9712"), val = tensor([1, 256, 1, 1])]; + tensor ts_cast_fp16 = reshape(shape = var_9712, x = ts_141_cast_fp16)[name = tensor("ts_cast_fp16")]; + tensor var_9719 = const()[name = tensor("op_9719"), val = tensor([1, 256, 16, 32, 2])]; + tensor k_complex_cast_fp16 = reshape(shape = var_9719, x = k_93_cast_fp16)[name = tensor("k_complex_cast_fp16")]; + tensor var_9723_begin_0 = const()[name = tensor("op_9723_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_9723_end_0 = const()[name = tensor("op_9723_end_0"), val = tensor([1, 256, 16, 32, 1])]; + tensor var_9723_end_mask_0 = const()[name = tensor("op_9723_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_9723_squeeze_mask_0 = const()[name = tensor("op_9723_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_9723_cast_fp16 = slice_by_index(begin = var_9723_begin_0, end = var_9723_end_0, end_mask = var_9723_end_mask_0, squeeze_mask = var_9723_squeeze_mask_0, x = k_complex_cast_fp16)[name = tensor("op_9723_cast_fp16")]; + tensor var_9731_begin_0 = const()[name = tensor("op_9731_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_9731_end_0 = const()[name = tensor("op_9731_end_0"), val = tensor([1, 256, 16, 32, 2])]; + tensor var_9731_end_mask_0 = const()[name = tensor("op_9731_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_9731_squeeze_mask_0 = const()[name = tensor("op_9731_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_9731_cast_fp16 = slice_by_index(begin = var_9731_begin_0, end = var_9731_end_0, end_mask = var_9731_end_mask_0, squeeze_mask = var_9731_squeeze_mask_0, x = k_complex_cast_fp16)[name = tensor("op_9731_cast_fp16")]; + tensor freqs_to_fp16 = const()[name = tensor("freqs_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(586380928)))]; + tensor var_9737_cast_fp16 = mul(x = freqs_to_fp16, y = ts_cast_fp16)[name = tensor("op_9737_cast_fp16")]; + tensor rotr_cast_fp16 = cos(x = var_9737_cast_fp16)[name = tensor("rotr_cast_fp16")]; + tensor roti_cast_fp16 = sin(x = var_9737_cast_fp16)[name = tensor("roti_cast_fp16")]; + tensor var_9741_cast_fp16 = mul(x = var_9723_cast_fp16, y = rotr_cast_fp16)[name = tensor("op_9741_cast_fp16")]; + tensor var_9742_cast_fp16 = mul(x = var_9731_cast_fp16, y = roti_cast_fp16)[name = tensor("op_9742_cast_fp16")]; + tensor kor_93_cast_fp16 = sub(x = var_9741_cast_fp16, y = var_9742_cast_fp16)[name = tensor("kor_93_cast_fp16")]; + tensor var_9745_cast_fp16 = mul(x = var_9723_cast_fp16, y = roti_cast_fp16)[name = tensor("op_9745_cast_fp16")]; + tensor var_9746_cast_fp16 = mul(x = var_9731_cast_fp16, y = rotr_cast_fp16)[name = tensor("op_9746_cast_fp16")]; + tensor koi_93_cast_fp16 = add(x = var_9745_cast_fp16, y = var_9746_cast_fp16)[name = tensor("koi_93_cast_fp16")]; + tensor ko_axis_0 = const()[name = tensor("ko_axis_0"), val = tensor(-1)]; + tensor ko_cast_fp16 = stack(axis = ko_axis_0, values = (kor_93_cast_fp16, koi_93_cast_fp16))[name = tensor("ko_cast_fp16")]; + tensor var_9766 = const()[name = tensor("op_9766"), val = tensor([1, 256, 16, 64])]; + tensor k_cast_fp16 = reshape(shape = var_9766, x = ko_cast_fp16)[name = tensor("k_cast_fp16")]; + tensor write_indices_float_139_cast_fp16 = add(x = var_9705_cast_fp16, y = var_393_promoted_to_fp16)[name = tensor("write_indices_float_139_cast_fp16")]; + tensor _inversed_9787_y_0_to_fp16 = const()[name = tensor("_inversed_9787_y_0_to_fp16"), val = tensor(0x1p-9)]; + tensor _inversed_9787_cast_fp16 = mul(x = write_indices_float_139_cast_fp16, y = _inversed_9787_y_0_to_fp16)[name = tensor("_inversed_9787_cast_fp16")]; + tensor var_9788_cast_fp16 = floor(x = _inversed_9787_cast_fp16)[name = tensor("op_9788_cast_fp16")]; + tensor var_9789_to_fp16 = const()[name = tensor("op_9789_to_fp16"), val = tensor(0x1p+9)]; + tensor var_9790_cast_fp16 = mul(x = var_9788_cast_fp16, y = var_9789_to_fp16)[name = tensor("op_9790_cast_fp16")]; + tensor write_indices_float_141_cast_fp16 = sub(x = write_indices_float_139_cast_fp16, y = var_9790_cast_fp16)[name = tensor("write_indices_float_141_cast_fp16")]; + tensor var_9808_value_0_to_fp16 = const()[name = tensor("op_9808_value_0_to_fp16"), val = tensor(0x1.ffp+8)]; + tensor var_9808_cast_fp16 = fill_like(ref_tensor = write_indices_float_141_cast_fp16, value = var_9808_value_0_to_fp16)[name = tensor("op_9808_cast_fp16")]; + tensor write_indices_float_cast_fp16 = select(a = write_indices_float_141_cast_fp16, b = var_9808_cast_fp16, cond = token_valid_1_cast_fp16)[name = tensor("write_indices_float_cast_fp16")]; + tensor var_9814_dtype_0 = const()[name = tensor("op_9814_dtype_0"), val = tensor("int32")]; + tensor var_9819 = const()[name = tensor("op_9819"), val = tensor([1, 256, 1, 1])]; + tensor write_indices_float_cast_fp16_to_int32 = cast(dtype = var_9814_dtype_0, x = write_indices_float_cast_fp16)[name = tensor("cast_545")]; + tensor var_9820 = reshape(shape = var_9819, x = write_indices_float_cast_fp16_to_int32)[name = tensor("op_9820")]; + tensor write_indices_reps_0 = const()[name = tensor("write_indices_reps_0"), val = tensor([1, 1, 16, 64])]; + tensor write_indices = tile(reps = write_indices_reps_0, x = var_9820)[name = tensor("write_indices")]; + tensor var_9830_begin_0 = const()[name = tensor("op_9830_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_9830_end_0 = const()[name = tensor("op_9830_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor var_9830_end_mask_0 = const()[name = tensor("op_9830_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_9830_squeeze_mask_0 = const()[name = tensor("op_9830_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor cache23_to_fp16_dtype_0 = const()[name = tensor("cache23_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor cache23_to_fp16 = cast(dtype = cache23_to_fp16_dtype_0, x = cache23)[name = tensor("cast_544")]; + tensor var_9830_cast_fp16 = slice_by_index(begin = var_9830_begin_0, end = var_9830_end_0, end_mask = var_9830_end_mask_0, squeeze_mask = var_9830_squeeze_mask_0, x = cache23_to_fp16)[name = tensor("op_9830_cast_fp16")]; + tensor var_9832_axis_0 = const()[name = tensor("op_9832_axis_0"), val = tensor(1)]; + tensor var_9832_mode_0 = const()[name = tensor("op_9832_mode_0"), val = tensor("update")]; + tensor var_9832_validate_indices_0 = const()[name = tensor("op_9832_validate_indices_0"), val = tensor(false)]; + tensor var_9832_cast_fp16 = scatter_along_axis(axis = var_9832_axis_0, data = var_9830_cast_fp16, indices = write_indices, mode = var_9832_mode_0, updates = k_cast_fp16, validate_indices = var_9832_validate_indices_0)[name = tensor("op_9832_cast_fp16")]; + tensor concat_186 = const()[name = tensor("concat_186"), val = tensor([0, 0, 0, 0, 0])]; + tensor concat_187 = const()[name = tensor("concat_187"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_internal_tensor_assign_1_stride_0 = const()[name = tensor("new_cache_internal_tensor_assign_1_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_internal_tensor_assign_1_begin_mask_0 = const()[name = tensor("new_cache_internal_tensor_assign_1_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_internal_tensor_assign_1_end_mask_0 = const()[name = tensor("new_cache_internal_tensor_assign_1_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_internal_tensor_assign_1_squeeze_mask_0 = const()[name = tensor("new_cache_internal_tensor_assign_1_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_92 = const()[name = tensor("shape_92"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_46 = const()[name = tensor("reduce_prod_46"), val = tensor(1048576)]; + tensor range_1d_46_start_0 = const()[name = tensor("range_1d_46_start_0"), val = tensor(0)]; + tensor range_1d_46_step_0 = const()[name = tensor("range_1d_46_step_0"), val = tensor(1)]; + tensor range_1d_46 = range_1d(end = reduce_prod_46, start = range_1d_46_start_0, step = range_1d_46_step_0)[name = tensor("range_1d_46")]; + tensor reshape_230 = reshape(shape = shape_92, x = range_1d_46)[name = tensor("reshape_230")]; + tensor slice_by_index_46 = slice_by_index(begin = concat_186, begin_mask = new_cache_internal_tensor_assign_1_begin_mask_0, end = concat_187, end_mask = new_cache_internal_tensor_assign_1_end_mask_0, squeeze_mask = new_cache_internal_tensor_assign_1_squeeze_mask_0, stride = new_cache_internal_tensor_assign_1_stride_0, x = reshape_230)[name = tensor("slice_by_index_46")]; + tensor reshape_231_shape_0 = const()[name = tensor("reshape_231_shape_0"), val = tensor([-1])]; + tensor reshape_231 = reshape(shape = reshape_231_shape_0, x = slice_by_index_46)[name = tensor("reshape_231")]; + tensor reshape_232_shape_0 = const()[name = tensor("reshape_232_shape_0"), val = tensor([-1])]; + tensor reshape_232_cast_fp16 = reshape(shape = reshape_232_shape_0, x = var_9832_cast_fp16)[name = tensor("reshape_232_cast_fp16")]; + tensor reshape_233_shape_0 = const()[name = tensor("reshape_233_shape_0"), val = tensor([-1])]; + tensor reshape_233_cast_fp16 = reshape(shape = reshape_233_shape_0, x = cache23_to_fp16)[name = tensor("reshape_233_cast_fp16")]; + tensor scatter_46_mode_0 = const()[name = tensor("scatter_46_mode_0"), val = tensor("update")]; + tensor scatter_46_axis_0 = const()[name = tensor("scatter_46_axis_0"), val = tensor(0)]; + tensor scatter_46_validate_indices_0 = const()[name = tensor("scatter_46_validate_indices_0"), val = tensor(false)]; + tensor scatter_46_cast_fp16 = scatter(axis = scatter_46_axis_0, data = reshape_233_cast_fp16, indices = reshape_231, mode = scatter_46_mode_0, updates = reshape_232_cast_fp16, validate_indices = scatter_46_validate_indices_0)[name = tensor("scatter_46_cast_fp16")]; + tensor reshape_234_cast_fp16 = reshape(shape = shape_92, x = scatter_46_cast_fp16)[name = tensor("reshape_234_cast_fp16")]; + tensor var_9840_begin_0 = const()[name = tensor("op_9840_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor var_9840_end_0 = const()[name = tensor("op_9840_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor var_9840_end_mask_0 = const()[name = tensor("op_9840_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_9840_squeeze_mask_0 = const()[name = tensor("op_9840_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor var_9840_cast_fp16 = slice_by_index(begin = var_9840_begin_0, end = var_9840_end_0, end_mask = var_9840_end_mask_0, squeeze_mask = var_9840_squeeze_mask_0, x = reshape_234_cast_fp16)[name = tensor("op_9840_cast_fp16")]; + tensor var_9842_axis_0 = const()[name = tensor("op_9842_axis_0"), val = tensor(1)]; + tensor var_9842_mode_0 = const()[name = tensor("op_9842_mode_0"), val = tensor("update")]; + tensor var_9842_validate_indices_0 = const()[name = tensor("op_9842_validate_indices_0"), val = tensor(false)]; + tensor var_9842_cast_fp16 = scatter_along_axis(axis = var_9842_axis_0, data = var_9840_cast_fp16, indices = write_indices, mode = var_9842_mode_0, updates = v_cast_fp16, validate_indices = var_9842_validate_indices_0)[name = tensor("op_9842_cast_fp16")]; + tensor concat_188 = const()[name = tensor("concat_188"), val = tensor([1, 0, 0, 0, 0])]; + tensor concat_189 = const()[name = tensor("concat_189"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_internal_tensor_assign_2_stride_0 = const()[name = tensor("new_cache_internal_tensor_assign_2_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_internal_tensor_assign_2_begin_mask_0 = const()[name = tensor("new_cache_internal_tensor_assign_2_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_internal_tensor_assign_2_end_mask_0 = const()[name = tensor("new_cache_internal_tensor_assign_2_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_internal_tensor_assign_2_squeeze_mask_0 = const()[name = tensor("new_cache_internal_tensor_assign_2_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_93 = const()[name = tensor("shape_93"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_47 = const()[name = tensor("reduce_prod_47"), val = tensor(1048576)]; + tensor range_1d_47_start_0 = const()[name = tensor("range_1d_47_start_0"), val = tensor(0)]; + tensor range_1d_47_step_0 = const()[name = tensor("range_1d_47_step_0"), val = tensor(1)]; + tensor range_1d_47 = range_1d(end = reduce_prod_47, start = range_1d_47_start_0, step = range_1d_47_step_0)[name = tensor("range_1d_47")]; + tensor reshape_235 = reshape(shape = shape_93, x = range_1d_47)[name = tensor("reshape_235")]; + tensor slice_by_index_47 = slice_by_index(begin = concat_188, begin_mask = new_cache_internal_tensor_assign_2_begin_mask_0, end = concat_189, end_mask = new_cache_internal_tensor_assign_2_end_mask_0, squeeze_mask = new_cache_internal_tensor_assign_2_squeeze_mask_0, stride = new_cache_internal_tensor_assign_2_stride_0, x = reshape_235)[name = tensor("slice_by_index_47")]; + tensor reshape_236_shape_0 = const()[name = tensor("reshape_236_shape_0"), val = tensor([-1])]; + tensor reshape_236 = reshape(shape = reshape_236_shape_0, x = slice_by_index_47)[name = tensor("reshape_236")]; + tensor reshape_237_shape_0 = const()[name = tensor("reshape_237_shape_0"), val = tensor([-1])]; + tensor reshape_237_cast_fp16 = reshape(shape = reshape_237_shape_0, x = var_9842_cast_fp16)[name = tensor("reshape_237_cast_fp16")]; + tensor reshape_238_shape_0 = const()[name = tensor("reshape_238_shape_0"), val = tensor([-1])]; + tensor reshape_238_cast_fp16 = reshape(shape = reshape_238_shape_0, x = reshape_234_cast_fp16)[name = tensor("reshape_238_cast_fp16")]; + tensor scatter_47_mode_0 = const()[name = tensor("scatter_47_mode_0"), val = tensor("update")]; + tensor scatter_47_axis_0 = const()[name = tensor("scatter_47_axis_0"), val = tensor(0)]; + tensor scatter_47_validate_indices_0 = const()[name = tensor("scatter_47_validate_indices_0"), val = tensor(false)]; + tensor scatter_47_cast_fp16 = scatter(axis = scatter_47_axis_0, data = reshape_238_cast_fp16, indices = reshape_236, mode = scatter_47_mode_0, updates = reshape_237_cast_fp16, validate_indices = scatter_47_validate_indices_0)[name = tensor("scatter_47_cast_fp16")]; + tensor reshape_239_cast_fp16 = reshape(shape = shape_93, x = scatter_47_cast_fp16)[name = tensor("reshape_239_cast_fp16")]; + tensor reshape_239_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("reshape_239_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor var_9852_cast_fp16 = add(x = position23_to_fp16, y = valid_len_to_fp16)[name = tensor("op_9852_cast_fp16")]; + tensor var_9852_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_9852_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor var_9852 = cast(dtype = var_9852_cast_fp16_to_fp32_dtype_0, x = var_9852_cast_fp16)[name = tensor("cast_542")]; + tensor new_cache_internal_tensor_assign_2 = cast(dtype = reshape_239_cast_fp16_to_fp32_dtype_0, x = reshape_239_cast_fp16)[name = tensor("cast_543")]; + tensor var_9624 = cast(dtype = var_9624_cast_fp16_to_fp32_dtype_0, x = var_9624_cast_fp16)[name = tensor("cast_547")]; + tensor new_cache_45_internal_tensor_assign_2 = cast(dtype = reshape_229_cast_fp16_to_fp32_dtype_0, x = reshape_229_cast_fp16)[name = tensor("cast_548")]; + tensor var_9213 = cast(dtype = var_9213_cast_fp16_to_fp32_dtype_0, x = var_9213_cast_fp16)[name = tensor("cast_552")]; + tensor new_cache_43_internal_tensor_assign_2 = cast(dtype = reshape_219_cast_fp16_to_fp32_dtype_0, x = reshape_219_cast_fp16)[name = tensor("cast_553")]; + tensor var_8802 = cast(dtype = var_8802_cast_fp16_to_fp32_dtype_0, x = var_8802_cast_fp16)[name = tensor("cast_557")]; + tensor new_cache_41_internal_tensor_assign_2 = cast(dtype = reshape_209_cast_fp16_to_fp32_dtype_0, x = reshape_209_cast_fp16)[name = tensor("cast_558")]; + tensor var_8391 = cast(dtype = var_8391_cast_fp16_to_fp32_dtype_0, x = var_8391_cast_fp16)[name = tensor("cast_562")]; + tensor new_cache_39_internal_tensor_assign_2 = cast(dtype = reshape_199_cast_fp16_to_fp32_dtype_0, x = reshape_199_cast_fp16)[name = tensor("cast_563")]; + tensor var_7980 = cast(dtype = var_7980_cast_fp16_to_fp32_dtype_0, x = var_7980_cast_fp16)[name = tensor("cast_567")]; + tensor new_cache_37_internal_tensor_assign_2 = cast(dtype = reshape_189_cast_fp16_to_fp32_dtype_0, x = reshape_189_cast_fp16)[name = tensor("cast_568")]; + tensor var_7569 = cast(dtype = var_7569_cast_fp16_to_fp32_dtype_0, x = var_7569_cast_fp16)[name = tensor("cast_572")]; + tensor new_cache_35_internal_tensor_assign_2 = cast(dtype = reshape_179_cast_fp16_to_fp32_dtype_0, x = reshape_179_cast_fp16)[name = tensor("cast_573")]; + tensor var_7158 = cast(dtype = var_7158_cast_fp16_to_fp32_dtype_0, x = var_7158_cast_fp16)[name = tensor("cast_577")]; + tensor new_cache_33_internal_tensor_assign_2 = cast(dtype = reshape_169_cast_fp16_to_fp32_dtype_0, x = reshape_169_cast_fp16)[name = tensor("cast_578")]; + tensor var_6747 = cast(dtype = var_6747_cast_fp16_to_fp32_dtype_0, x = var_6747_cast_fp16)[name = tensor("cast_582")]; + tensor new_cache_31_internal_tensor_assign_2 = cast(dtype = reshape_159_cast_fp16_to_fp32_dtype_0, x = reshape_159_cast_fp16)[name = tensor("cast_583")]; + tensor var_6336 = cast(dtype = var_6336_cast_fp16_to_fp32_dtype_0, x = var_6336_cast_fp16)[name = tensor("cast_587")]; + tensor new_cache_29_internal_tensor_assign_2 = cast(dtype = reshape_149_cast_fp16_to_fp32_dtype_0, x = reshape_149_cast_fp16)[name = tensor("cast_588")]; + tensor var_5925 = cast(dtype = var_5925_cast_fp16_to_fp32_dtype_0, x = var_5925_cast_fp16)[name = tensor("cast_592")]; + tensor new_cache_27_internal_tensor_assign_2 = cast(dtype = reshape_139_cast_fp16_to_fp32_dtype_0, x = reshape_139_cast_fp16)[name = tensor("cast_593")]; + tensor var_5514 = cast(dtype = var_5514_cast_fp16_to_fp32_dtype_0, x = var_5514_cast_fp16)[name = tensor("cast_597")]; + tensor new_cache_25_internal_tensor_assign_2 = cast(dtype = reshape_129_cast_fp16_to_fp32_dtype_0, x = reshape_129_cast_fp16)[name = tensor("cast_598")]; + tensor var_5103 = cast(dtype = var_5103_cast_fp16_to_fp32_dtype_0, x = var_5103_cast_fp16)[name = tensor("cast_602")]; + tensor new_cache_23_internal_tensor_assign_2 = cast(dtype = reshape_119_cast_fp16_to_fp32_dtype_0, x = reshape_119_cast_fp16)[name = tensor("cast_603")]; + tensor var_4692 = cast(dtype = var_4692_cast_fp16_to_fp32_dtype_0, x = var_4692_cast_fp16)[name = tensor("cast_607")]; + tensor new_cache_21_internal_tensor_assign_2 = cast(dtype = reshape_109_cast_fp16_to_fp32_dtype_0, x = reshape_109_cast_fp16)[name = tensor("cast_608")]; + tensor var_4281 = cast(dtype = var_4281_cast_fp16_to_fp32_dtype_0, x = var_4281_cast_fp16)[name = tensor("cast_612")]; + tensor new_cache_19_internal_tensor_assign_2 = cast(dtype = reshape_99_cast_fp16_to_fp32_dtype_0, x = reshape_99_cast_fp16)[name = tensor("cast_613")]; + tensor var_3870 = cast(dtype = var_3870_cast_fp16_to_fp32_dtype_0, x = var_3870_cast_fp16)[name = tensor("cast_617")]; + tensor new_cache_17_internal_tensor_assign_2 = cast(dtype = reshape_89_cast_fp16_to_fp32_dtype_0, x = reshape_89_cast_fp16)[name = tensor("cast_618")]; + tensor var_3459 = cast(dtype = var_3459_cast_fp16_to_fp32_dtype_0, x = var_3459_cast_fp16)[name = tensor("cast_622")]; + tensor new_cache_15_internal_tensor_assign_2 = cast(dtype = reshape_79_cast_fp16_to_fp32_dtype_0, x = reshape_79_cast_fp16)[name = tensor("cast_623")]; + tensor var_3048 = cast(dtype = var_3048_cast_fp16_to_fp32_dtype_0, x = var_3048_cast_fp16)[name = tensor("cast_627")]; + tensor new_cache_13_internal_tensor_assign_2 = cast(dtype = reshape_69_cast_fp16_to_fp32_dtype_0, x = reshape_69_cast_fp16)[name = tensor("cast_628")]; + tensor var_2637 = cast(dtype = var_2637_cast_fp16_to_fp32_dtype_0, x = var_2637_cast_fp16)[name = tensor("cast_632")]; + tensor new_cache_11_internal_tensor_assign_2 = cast(dtype = reshape_59_cast_fp16_to_fp32_dtype_0, x = reshape_59_cast_fp16)[name = tensor("cast_633")]; + tensor var_2226 = cast(dtype = var_2226_cast_fp16_to_fp32_dtype_0, x = var_2226_cast_fp16)[name = tensor("cast_637")]; + tensor new_cache_9_internal_tensor_assign_2 = cast(dtype = reshape_49_cast_fp16_to_fp32_dtype_0, x = reshape_49_cast_fp16)[name = tensor("cast_638")]; + tensor var_1815 = cast(dtype = var_1815_cast_fp16_to_fp32_dtype_0, x = var_1815_cast_fp16)[name = tensor("cast_642")]; + tensor new_cache_7_internal_tensor_assign_2 = cast(dtype = reshape_39_cast_fp16_to_fp32_dtype_0, x = reshape_39_cast_fp16)[name = tensor("cast_643")]; + tensor var_1404 = cast(dtype = var_1404_cast_fp16_to_fp32_dtype_0, x = var_1404_cast_fp16)[name = tensor("cast_647")]; + tensor new_cache_5_internal_tensor_assign_2 = cast(dtype = reshape_29_cast_fp16_to_fp32_dtype_0, x = reshape_29_cast_fp16)[name = tensor("cast_648")]; + tensor var_993 = cast(dtype = var_993_cast_fp16_to_fp32_dtype_0, x = var_993_cast_fp16)[name = tensor("cast_652")]; + tensor new_cache_3_internal_tensor_assign_2 = cast(dtype = reshape_19_cast_fp16_to_fp32_dtype_0, x = reshape_19_cast_fp16)[name = tensor("cast_653")]; + tensor var_582 = cast(dtype = var_582_cast_fp16_to_fp32_dtype_0, x = var_582_cast_fp16)[name = tensor("cast_657")]; + tensor new_cache_1_internal_tensor_assign_2 = cast(dtype = reshape_9_cast_fp16_to_fp32_dtype_0, x = reshape_9_cast_fp16)[name = tensor("cast_658")]; + } -> (new_cache_1_internal_tensor_assign_2, var_582, new_cache_3_internal_tensor_assign_2, var_993, new_cache_5_internal_tensor_assign_2, var_1404, new_cache_7_internal_tensor_assign_2, var_1815, new_cache_9_internal_tensor_assign_2, var_2226, new_cache_11_internal_tensor_assign_2, var_2637, new_cache_13_internal_tensor_assign_2, var_3048, new_cache_15_internal_tensor_assign_2, var_3459, new_cache_17_internal_tensor_assign_2, var_3870, new_cache_19_internal_tensor_assign_2, var_4281, new_cache_21_internal_tensor_assign_2, var_4692, new_cache_23_internal_tensor_assign_2, var_5103, new_cache_25_internal_tensor_assign_2, var_5514, new_cache_27_internal_tensor_assign_2, var_5925, new_cache_29_internal_tensor_assign_2, var_6336, new_cache_31_internal_tensor_assign_2, var_6747, new_cache_33_internal_tensor_assign_2, var_7158, new_cache_35_internal_tensor_assign_2, var_7569, new_cache_37_internal_tensor_assign_2, var_7980, new_cache_39_internal_tensor_assign_2, var_8391, new_cache_41_internal_tensor_assign_2, var_8802, new_cache_43_internal_tensor_assign_2, var_9213, new_cache_45_internal_tensor_assign_2, var_9624, new_cache_internal_tensor_assign_2, var_9852); +} \ No newline at end of file diff --git a/v2.1/spanish_24l/cond_prefill.mlmodelc/weights/weight.bin b/v2.1/spanish_24l/cond_prefill.mlmodelc/weights/weight.bin new file mode 100644 index 0000000000000000000000000000000000000000..a16790546e89c6208605e52db1a1bc4819c7fec7 --- /dev/null +++ b/v2.1/spanish_24l/cond_prefill.mlmodelc/weights/weight.bin @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:cfb4f6ca73314577454e22fc73bafda7e2375bf7755992d8c577462a16a05999 +size 586381056 diff --git a/v2.1/spanish_24l/cond_prefill.mlpackage/Data/com.apple.CoreML/model.mlmodel b/v2.1/spanish_24l/cond_prefill.mlpackage/Data/com.apple.CoreML/model.mlmodel new file mode 100644 index 0000000000000000000000000000000000000000..62038c517ca82921495ec65dee28f141769c3661 --- /dev/null +++ b/v2.1/spanish_24l/cond_prefill.mlpackage/Data/com.apple.CoreML/model.mlmodel @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:28af45e579af38a8b0108f6f7c14e2ddbf5683ca1b8928c2c667285747e1bda4 +size 842494 diff --git a/v2.1/spanish_24l/cond_prefill.mlpackage/Data/com.apple.CoreML/weights/weight.bin b/v2.1/spanish_24l/cond_prefill.mlpackage/Data/com.apple.CoreML/weights/weight.bin new file mode 100644 index 0000000000000000000000000000000000000000..a16790546e89c6208605e52db1a1bc4819c7fec7 --- /dev/null +++ b/v2.1/spanish_24l/cond_prefill.mlpackage/Data/com.apple.CoreML/weights/weight.bin @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:cfb4f6ca73314577454e22fc73bafda7e2375bf7755992d8c577462a16a05999 +size 586381056 diff --git a/v2.1/spanish_24l/cond_prefill.mlpackage/Manifest.json b/v2.1/spanish_24l/cond_prefill.mlpackage/Manifest.json new file mode 100644 index 0000000000000000000000000000000000000000..e0d62d446d7746c2a5404c782352a7dce814dbac --- /dev/null +++ b/v2.1/spanish_24l/cond_prefill.mlpackage/Manifest.json @@ -0,0 +1,18 @@ +{ + "fileFormatVersion": "1.0.0", + "itemInfoEntries": { + "5D2D6281-CB59-4544-8971-AC50562F7C22": { + "author": "com.apple.CoreML", + "description": "CoreML Model Specification", + "name": "model.mlmodel", + "path": "com.apple.CoreML/model.mlmodel" + }, + "80FCA390-E211-4F16-B673-E7474368CCC1": { + "author": "com.apple.CoreML", + "description": "CoreML Model Weights", + "name": "weights", + "path": "com.apple.CoreML/weights" + } + }, + "rootModelIdentifier": "5D2D6281-CB59-4544-8971-AC50562F7C22" +} diff --git a/v2.1/spanish_24l/constants/bos_emb.npy b/v2.1/spanish_24l/constants/bos_emb.npy new file mode 100644 index 0000000000000000000000000000000000000000..651ab5522e06a3457d600894f5d88db2f849207f --- /dev/null +++ b/v2.1/spanish_24l/constants/bos_emb.npy @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:8532d97622913da0c1a76292a24365b6fce6762986d27ee3c310a91d437cd5c6 +size 256 diff --git a/v2.1/spanish_24l/constants/emb_mean.npy b/v2.1/spanish_24l/constants/emb_mean.npy new file mode 100644 index 0000000000000000000000000000000000000000..29bee3dc5f686c58698b896bcd48f6957f79e8d7 --- /dev/null +++ b/v2.1/spanish_24l/constants/emb_mean.npy @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:57333945fc74572a2504025b3d48ad156217828144fba9407aabb962d58028b3 +size 256 diff --git a/v2.1/spanish_24l/constants/emb_std.npy b/v2.1/spanish_24l/constants/emb_std.npy new file mode 100644 index 0000000000000000000000000000000000000000..674d07376296efb08e41a46352a97a9ed4349760 --- /dev/null +++ b/v2.1/spanish_24l/constants/emb_std.npy @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:9e8ac8f4dcd6e6a3aa2d1c1fc9270b248525856da470d7f1a9b86c966dbebe69 +size 256 diff --git a/v2.1/spanish_24l/constants/mimi_init_state.npz b/v2.1/spanish_24l/constants/mimi_init_state.npz new file mode 100644 index 0000000000000000000000000000000000000000..3b52f383a86ef0439f9993506e4529c4c1234828 --- /dev/null +++ b/v2.1/spanish_24l/constants/mimi_init_state.npz @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:515588e10e9c14a6b009f1cfda9fcd592542b1640ad856d5fc69e50024702a39 +size 1083108 diff --git a/v2.1/spanish_24l/constants/quantizer_weight.npy b/v2.1/spanish_24l/constants/quantizer_weight.npy new file mode 100644 index 0000000000000000000000000000000000000000..74fd84280b4dee8d77ab6146e9c5eec48763ba30 --- /dev/null +++ b/v2.1/spanish_24l/constants/quantizer_weight.npy @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:b068606d26574a8a1513591f9d228b0b44b71d9c90f6c1f032f6ed6cae9f065c +size 65664 diff --git a/v2.1/spanish_24l/constants/text_embed_table.npy b/v2.1/spanish_24l/constants/text_embed_table.npy new file mode 100644 index 0000000000000000000000000000000000000000..ffca44d45743c3679467bb3d72e8feb423826265 --- /dev/null +++ b/v2.1/spanish_24l/constants/text_embed_table.npy @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:889cb9feac3f8f74718b6f37b58015afa57ebcb84d221a683fed265d10da4ae5 +size 16388224 diff --git a/v2.1/spanish_24l/constants_bin/alba.safetensors b/v2.1/spanish_24l/constants_bin/alba.safetensors new file mode 100644 index 0000000000000000000000000000000000000000..1fc8875a679415ae86216768dc224a3093da7598 --- /dev/null +++ b/v2.1/spanish_24l/constants_bin/alba.safetensors @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:4fb1780fda68177656f7da0c781db388a5b9dd4fc35a47b27f24600b16854984 +size 24777760 diff --git a/v2.1/spanish_24l/constants_bin/anna.safetensors b/v2.1/spanish_24l/constants_bin/anna.safetensors new file mode 100644 index 0000000000000000000000000000000000000000..ab1f3a5143e9f2a9fc221f875a7427a89ba0c027 --- /dev/null +++ b/v2.1/spanish_24l/constants_bin/anna.safetensors @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:5190915b6ced491154d619f32dfc7f7829ce1d633f0615223dae4c55b57535e0 +size 31265824 diff --git a/v2.1/spanish_24l/constants_bin/azelma.safetensors b/v2.1/spanish_24l/constants_bin/azelma.safetensors new file mode 100644 index 0000000000000000000000000000000000000000..4a3495f3ca6ad30bf4fc46f4ddae70b1c933d287 --- /dev/null +++ b/v2.1/spanish_24l/constants_bin/azelma.safetensors @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:25cc32b8bbfee482a1062f6e81414991253b47f3c7b16b2f3288c93ec1e3b94d +size 31855648 diff --git a/v2.1/spanish_24l/constants_bin/bill_boerst.safetensors b/v2.1/spanish_24l/constants_bin/bill_boerst.safetensors new file mode 100644 index 0000000000000000000000000000000000000000..06bf269a35c5d5e8d34a16ac9b6e9c2d9640f50d --- /dev/null +++ b/v2.1/spanish_24l/constants_bin/bill_boerst.safetensors @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:4c4e8ad3d0e8c87657f2a3b32b3b3d70943935cdbd6917d90668bfbe69cc92be +size 26940448 diff --git a/v2.1/spanish_24l/constants_bin/bos_before_voice.bin b/v2.1/spanish_24l/constants_bin/bos_before_voice.bin new file mode 100644 index 0000000000000000000000000000000000000000..0c46d3e4ddc39d34ef8f5cf03dc4244e61fa3279 --- /dev/null +++ b/v2.1/spanish_24l/constants_bin/bos_before_voice.bin @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:3b7d90522ddc0d1daad2518f625f0cb33bf403b151b20817763dca57e23a50c8 +size 4096 diff --git a/v2.1/spanish_24l/constants_bin/bos_emb.bin b/v2.1/spanish_24l/constants_bin/bos_emb.bin new file mode 100644 index 0000000000000000000000000000000000000000..a5ce34f707f20cdde60d286b2b1f468d6ea2d54c --- /dev/null +++ b/v2.1/spanish_24l/constants_bin/bos_emb.bin @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:fd0a8c7f577da4168f83914060af72bb93c832f2294ac41aad9933d727079e40 +size 128 diff --git a/v2.1/spanish_24l/constants_bin/caro_davy.safetensors b/v2.1/spanish_24l/constants_bin/caro_davy.safetensors new file mode 100644 index 0000000000000000000000000000000000000000..b8e7fe379a56521fd3e187259c2067f5ae7bcd2c --- /dev/null +++ b/v2.1/spanish_24l/constants_bin/caro_davy.safetensors @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:1e421c80de42dad99fd36a1d26307e52fb9eb9feeeacbcd1d359971e90929f2d +size 21042200 diff --git a/v2.1/spanish_24l/constants_bin/charles.safetensors b/v2.1/spanish_24l/constants_bin/charles.safetensors new file mode 100644 index 0000000000000000000000000000000000000000..63e941eb1a667aed77973fed71da1f6496820a2a --- /dev/null +++ b/v2.1/spanish_24l/constants_bin/charles.safetensors @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:17ea33581e5d14ec4d1f2e87621695ae36d9b5a5558d21d6830f0da88ff9c704 +size 24777760 diff --git a/v2.1/spanish_24l/constants_bin/cosette.safetensors b/v2.1/spanish_24l/constants_bin/cosette.safetensors new file mode 100644 index 0000000000000000000000000000000000000000..aef4b807bc2a566055246ffd56f89ef9e0c1bc4f --- /dev/null +++ b/v2.1/spanish_24l/constants_bin/cosette.safetensors @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:7eddd3552a292fbb8a90d882eed81bdd56f41b23f587d2493928b3e08d9a287e +size 24777760 diff --git a/v2.1/spanish_24l/constants_bin/eponine.safetensors b/v2.1/spanish_24l/constants_bin/eponine.safetensors new file mode 100644 index 0000000000000000000000000000000000000000..919990cb7959d39997a0b6b2a42964c357179956 --- /dev/null +++ b/v2.1/spanish_24l/constants_bin/eponine.safetensors @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:b652112b2de9155bcc3ba0cd4c7279c1632d0cc39cd527c1085468e926d2640c +size 27726880 diff --git a/v2.1/spanish_24l/constants_bin/estelle.safetensors b/v2.1/spanish_24l/constants_bin/estelle.safetensors new file mode 100644 index 0000000000000000000000000000000000000000..64a5c852d61dd10d49ea91937d629347c683a500 --- /dev/null +++ b/v2.1/spanish_24l/constants_bin/estelle.safetensors @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:b16a48de7b9f2f2afe4350dd32fd5a01e64b873e258a990464dc8f2e1582b344 +size 33035296 diff --git a/v2.1/spanish_24l/constants_bin/eve.safetensors b/v2.1/spanish_24l/constants_bin/eve.safetensors new file mode 100644 index 0000000000000000000000000000000000000000..84b7c6cf241fc0d09b93d54f852bd770dbc12cb0 --- /dev/null +++ b/v2.1/spanish_24l/constants_bin/eve.safetensors @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:373fa084cf1d5773445b6dd559614ffd1cf954c32517ccb9862bf5b735d961e4 +size 26154016 diff --git a/v2.1/spanish_24l/constants_bin/fantine.safetensors b/v2.1/spanish_24l/constants_bin/fantine.safetensors new file mode 100644 index 0000000000000000000000000000000000000000..51e9f8f54520c185fadb5e88bfc0d301a38591ab --- /dev/null +++ b/v2.1/spanish_24l/constants_bin/fantine.safetensors @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:c10a16bac8ec3d98e23927baad1471d8959884cb99622b41734491b517627e60 +size 26154016 diff --git a/v2.1/spanish_24l/constants_bin/george.safetensors b/v2.1/spanish_24l/constants_bin/george.safetensors new file mode 100644 index 0000000000000000000000000000000000000000..a7d36fa4bd1ac995195ac1a01d521822441fbcaa --- /dev/null +++ b/v2.1/spanish_24l/constants_bin/george.safetensors @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:bfc8a6e4e01060d3843e0ab19912739697824d5a84de42618986a4abd65154bb +size 24974368 diff --git a/v2.1/spanish_24l/constants_bin/giovanni.safetensors b/v2.1/spanish_24l/constants_bin/giovanni.safetensors new file mode 100644 index 0000000000000000000000000000000000000000..81574818a2fc10eda23f15a0a7e56dd2b0e3c7fb --- /dev/null +++ b/v2.1/spanish_24l/constants_bin/giovanni.safetensors @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:e8386d054b1f2d9af2bcc48c20d5273716a96ef04adece489af287c5dac20a68 +size 18486272 diff --git a/v2.1/spanish_24l/constants_bin/jane.safetensors b/v2.1/spanish_24l/constants_bin/jane.safetensors new file mode 100644 index 0000000000000000000000000000000000000000..556dea997d06deb2687e806c5743243679f14f4e --- /dev/null +++ b/v2.1/spanish_24l/constants_bin/jane.safetensors @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:2993ab3450e140cf508ef11549e3421636d552c7ff7dd94370d348e0a27407dd +size 29496352 diff --git a/v2.1/spanish_24l/constants_bin/javert.safetensors b/v2.1/spanish_24l/constants_bin/javert.safetensors new file mode 100644 index 0000000000000000000000000000000000000000..999f1481e92bf88fa28286995a0f8e169aa7c646 --- /dev/null +++ b/v2.1/spanish_24l/constants_bin/javert.safetensors @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:02568ad8bc470966b7ad38ed2952c3e68465a9183d2a7e4d910d0424d016216d +size 24777760 diff --git a/v2.1/spanish_24l/constants_bin/jean.safetensors b/v2.1/spanish_24l/constants_bin/jean.safetensors new file mode 100644 index 0000000000000000000000000000000000000000..4eacb79cd9f52e2b6d930c3beda54f7e4b3b7b8d --- /dev/null +++ b/v2.1/spanish_24l/constants_bin/jean.safetensors @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:876c50c9c200f0f32c9b4f9e591dd11b72dc8419b99773b6e4c82d8404fb6c20 +size 24777760 diff --git a/v2.1/spanish_24l/constants_bin/juergen.safetensors b/v2.1/spanish_24l/constants_bin/juergen.safetensors new file mode 100644 index 0000000000000000000000000000000000000000..aa760323726c0e7f3760888d8434bfd702a8592d --- /dev/null +++ b/v2.1/spanish_24l/constants_bin/juergen.safetensors @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:eef78e93daac817e08825a0f2acfe3968c6480f73eb05fa8c2034a2b9ec17a19 +size 24974368 diff --git a/v2.1/spanish_24l/constants_bin/lola.safetensors b/v2.1/spanish_24l/constants_bin/lola.safetensors new file mode 100644 index 0000000000000000000000000000000000000000..f8caa446799f295f5b464e2ad483ef7888d76f28 --- /dev/null +++ b/v2.1/spanish_24l/constants_bin/lola.safetensors @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:07b36d8bfd72fd0e9aa091b5f56241a3b6c5b251f7ff466be43b7000e0efc661 +size 23794712 diff --git a/v2.1/spanish_24l/constants_bin/marius.safetensors b/v2.1/spanish_24l/constants_bin/marius.safetensors new file mode 100644 index 0000000000000000000000000000000000000000..e183b5f19c9a65fb51ae0562dd2b4d22b5391ac6 --- /dev/null +++ b/v2.1/spanish_24l/constants_bin/marius.safetensors @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:35b9499e5fbbca09d12887916dec00e5a3affc81c67d1535f5fb696802bd19a7 +size 24777760 diff --git a/v2.1/spanish_24l/constants_bin/mary.safetensors b/v2.1/spanish_24l/constants_bin/mary.safetensors new file mode 100644 index 0000000000000000000000000000000000000000..9ea182a607152df8e2cc74a65415fd6bbc20d7ac --- /dev/null +++ b/v2.1/spanish_24l/constants_bin/mary.safetensors @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:01bf1479ac0f914adec0c7ca325d49d239bfec03ddb9e30a28e231253be43457 +size 24777760 diff --git a/v2.1/spanish_24l/constants_bin/michael.safetensors b/v2.1/spanish_24l/constants_bin/michael.safetensors new file mode 100644 index 0000000000000000000000000000000000000000..f21eef98e4eb2ab60d9273b7e7c24d9627bd224b --- /dev/null +++ b/v2.1/spanish_24l/constants_bin/michael.safetensors @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:6d3227ee02a7fee6d0c581289a6fb23a0502f00777a9ab6ea780cb9b03bf4c84 +size 29103136 diff --git a/v2.1/spanish_24l/constants_bin/paul.safetensors b/v2.1/spanish_24l/constants_bin/paul.safetensors new file mode 100644 index 0000000000000000000000000000000000000000..59ebb209c0f1dcf2bc583c0116800cebc15652ec --- /dev/null +++ b/v2.1/spanish_24l/constants_bin/paul.safetensors @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:a9ecf89514ea906d77e5aed6485566ce3714cbc5effb9652ac53d94db3251d3a +size 27923488 diff --git a/v2.1/spanish_24l/constants_bin/peter_yearsley.safetensors b/v2.1/spanish_24l/constants_bin/peter_yearsley.safetensors new file mode 100644 index 0000000000000000000000000000000000000000..a3a4e4265cb86a6700e8df7a5edf06fce99b78d6 --- /dev/null +++ b/v2.1/spanish_24l/constants_bin/peter_yearsley.safetensors @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:719f6cbcb7ec48751229de0354c4d1cf3e75bc75063ef9455896efd8da88ff9b +size 14947320 diff --git a/v2.1/spanish_24l/constants_bin/rafael.safetensors b/v2.1/spanish_24l/constants_bin/rafael.safetensors new file mode 100644 index 0000000000000000000000000000000000000000..fec5a95d217dd5761b168f6bf7c79274dcae25aa --- /dev/null +++ b/v2.1/spanish_24l/constants_bin/rafael.safetensors @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:bae1a031a903e83fbff721655b2ebe4286559873d07de464cc2748ce5c9b9103 +size 24777760 diff --git a/v2.1/spanish_24l/constants_bin/stuart_bell.safetensors b/v2.1/spanish_24l/constants_bin/stuart_bell.safetensors new file mode 100644 index 0000000000000000000000000000000000000000..74dba3d2a8af9cef369c35447720d3c16d7496de --- /dev/null +++ b/v2.1/spanish_24l/constants_bin/stuart_bell.safetensors @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:76ec9ffc71b1c59ff8ae4be33ae4eb3761d7f4987c9c11405f9cc63db889bf8f +size 21042200 diff --git a/v2.1/spanish_24l/constants_bin/text_embed_table.bin b/v2.1/spanish_24l/constants_bin/text_embed_table.bin new file mode 100644 index 0000000000000000000000000000000000000000..ad349298138a112e04674cb3cdfdc3e129a11e1f --- /dev/null +++ b/v2.1/spanish_24l/constants_bin/text_embed_table.bin @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:56fd41a3073bc03365d09ae1c0f814515216ec8a123f13f02fad47cba2dafcfe +size 16388096 diff --git a/v2.1/spanish_24l/constants_bin/tokenizer.model b/v2.1/spanish_24l/constants_bin/tokenizer.model new file mode 100644 index 0000000000000000000000000000000000000000..1ac69e37ac41fb9e2a4491ade3064674a8f2682c --- /dev/null +++ b/v2.1/spanish_24l/constants_bin/tokenizer.model @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:aac2b96478e3e3068712bd5c45dc196f84b7341ea72587c33a2094957f0c1b0c +size 60895 diff --git a/v2.1/spanish_24l/constants_bin/vera.safetensors b/v2.1/spanish_24l/constants_bin/vera.safetensors new file mode 100644 index 0000000000000000000000000000000000000000..3f0a5b33943f205c8d68302c93a7ea5215046ebf --- /dev/null +++ b/v2.1/spanish_24l/constants_bin/vera.safetensors @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:0b5364212a6b0994475cc3a98919ac6439a435d92731ae8518b180607e7c690e +size 26940448 diff --git a/v2.1/spanish_24l/flow_decoder_fused.mlmodelc/analytics/coremldata.bin b/v2.1/spanish_24l/flow_decoder_fused.mlmodelc/analytics/coremldata.bin new file mode 100644 index 0000000000000000000000000000000000000000..605d9a74c499ea542c6afa2793265d37d1b9d20a --- /dev/null +++ b/v2.1/spanish_24l/flow_decoder_fused.mlmodelc/analytics/coremldata.bin @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:a9967185ef10f7dbd3a4ad1b513b365dfc7276fa56b609b18fce157c992ee752 +size 243 diff --git a/v2.1/spanish_24l/flow_decoder_fused.mlmodelc/coremldata.bin b/v2.1/spanish_24l/flow_decoder_fused.mlmodelc/coremldata.bin new file mode 100644 index 0000000000000000000000000000000000000000..498f5971adf5e33b7dc53bba02ecba9e4116601a --- /dev/null +++ b/v2.1/spanish_24l/flow_decoder_fused.mlmodelc/coremldata.bin @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:e5ed2f6b3a95e3025402f669d98ad09634bd2783c48e3dada3394d9dd4fbe81c +size 413 diff --git a/v2.1/spanish_24l/flow_decoder_fused.mlmodelc/model.mil b/v2.1/spanish_24l/flow_decoder_fused.mlmodelc/model.mil new file mode 100644 index 0000000000000000000000000000000000000000..c64a1f34555e7bbf9c8d4502d29f9ef52dd6b0ec --- /dev/null +++ b/v2.1/spanish_24l/flow_decoder_fused.mlmodelc/model.mil @@ -0,0 +1,1915 @@ +program(1.0) +[buildInfo = dict, tensor>({{"coremlc-component-MIL", "3520.4.1"}, {"coremlc-version", "3520.5.1"}, {"coremltools-component-torch", "2.12.0"}, {"coremltools-source-dialect", "TorchScript"}, {"coremltools-version", "9.0"}})] +{ + func main(tensor latent_init, tensor transformer_out) { + tensor latent_init_to_fp16_dtype_0 = const()[name = tensor("latent_init_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor flow_net_input_proj_weight_to_fp16 = const()[name = tensor("flow_net_input_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(64)))]; + tensor flow_net_input_proj_bias_to_fp16 = const()[name = tensor("flow_net_input_proj_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(32896)))]; + tensor latent_init_to_fp16 = cast(dtype = latent_init_to_fp16_dtype_0, x = latent_init)[name = tensor("cast_146")]; + tensor linear_0_cast_fp16 = linear(bias = flow_net_input_proj_bias_to_fp16, weight = flow_net_input_proj_weight_to_fp16, x = latent_init_to_fp16)[name = tensor("linear_0_cast_fp16")]; + tensor input_3_to_fp16 = const()[name = tensor("input_3_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(33984)))]; + tensor input_5_cast_fp16 = silu(x = input_3_to_fp16)[name = tensor("input_5_cast_fp16")]; + tensor flow_net_time_embed_0_mlp_2_weight_to_fp16 = const()[name = tensor("flow_net_time_embed_0_mlp_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(35072)))]; + tensor flow_net_time_embed_0_mlp_2_bias_to_fp16 = const()[name = tensor("flow_net_time_embed_0_mlp_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(559424)))]; + tensor linear_2_cast_fp16 = linear(bias = flow_net_time_embed_0_mlp_2_bias_to_fp16, weight = flow_net_time_embed_0_mlp_2_weight_to_fp16, x = input_5_cast_fp16)[name = tensor("linear_2_cast_fp16")]; + tensor reduce_mean_0_axes_0 = const()[name = tensor("reduce_mean_0_axes_0"), val = tensor([-1])]; + tensor reduce_mean_0_keep_dims_0 = const()[name = tensor("reduce_mean_0_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_0_cast_fp16 = reduce_mean(axes = reduce_mean_0_axes_0, keep_dims = reduce_mean_0_keep_dims_0, x = linear_2_cast_fp16)[name = tensor("reduce_mean_0_cast_fp16")]; + tensor sub_0_cast_fp16 = sub(x = linear_2_cast_fp16, y = reduce_mean_0_cast_fp16)[name = tensor("sub_0_cast_fp16")]; + tensor square_0_cast_fp16 = square(x = sub_0_cast_fp16)[name = tensor("square_0_cast_fp16")]; + tensor reduce_mean_1_axes_0 = const()[name = tensor("reduce_mean_1_axes_0"), val = tensor([-1])]; + tensor reduce_mean_1_keep_dims_0 = const()[name = tensor("reduce_mean_1_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_1_cast_fp16 = reduce_mean(axes = reduce_mean_1_axes_0, keep_dims = reduce_mean_1_keep_dims_0, x = square_0_cast_fp16)[name = tensor("reduce_mean_1_cast_fp16")]; + tensor real_div_0_to_fp16 = const()[name = tensor("real_div_0_to_fp16"), val = tensor(0x1.008p+0)]; + tensor mul_0_cast_fp16 = mul(x = reduce_mean_1_cast_fp16, y = real_div_0_to_fp16)[name = tensor("mul_0_cast_fp16")]; + tensor var_70_to_fp16 = const()[name = tensor("op_70_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_1_cast_fp16 = add(x = mul_0_cast_fp16, y = var_70_to_fp16)[name = tensor("var_1_cast_fp16")]; + tensor var_73_epsilon_0 = const()[name = tensor("op_73_epsilon_0"), val = tensor(0x1.197998p-40)]; + tensor var_73_cast_fp16 = rsqrt(epsilon = var_73_epsilon_0, x = var_1_cast_fp16)[name = tensor("op_73_cast_fp16")]; + tensor const_3_to_fp16 = const()[name = tensor("const_3_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(560512)))]; + tensor var_74_cast_fp16 = mul(x = const_3_to_fp16, y = var_73_cast_fp16)[name = tensor("op_74_cast_fp16")]; + tensor var_75_cast_fp16 = mul(x = linear_2_cast_fp16, y = var_74_cast_fp16)[name = tensor("op_75_cast_fp16")]; + tensor input_9_to_fp16 = const()[name = tensor("input_9_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(561600)))]; + tensor input_11_cast_fp16 = silu(x = input_9_to_fp16)[name = tensor("input_11_cast_fp16")]; + tensor flow_net_time_embed_1_mlp_2_weight_to_fp16 = const()[name = tensor("flow_net_time_embed_1_mlp_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(562688)))]; + tensor flow_net_time_embed_1_mlp_2_bias_to_fp16 = const()[name = tensor("flow_net_time_embed_1_mlp_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1087040)))]; + tensor linear_4_cast_fp16 = linear(bias = flow_net_time_embed_1_mlp_2_bias_to_fp16, weight = flow_net_time_embed_1_mlp_2_weight_to_fp16, x = input_11_cast_fp16)[name = tensor("linear_4_cast_fp16")]; + tensor reduce_mean_2_axes_0 = const()[name = tensor("reduce_mean_2_axes_0"), val = tensor([-1])]; + tensor reduce_mean_2_keep_dims_0 = const()[name = tensor("reduce_mean_2_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_2_cast_fp16 = reduce_mean(axes = reduce_mean_2_axes_0, keep_dims = reduce_mean_2_keep_dims_0, x = linear_4_cast_fp16)[name = tensor("reduce_mean_2_cast_fp16")]; + tensor sub_2_cast_fp16 = sub(x = linear_4_cast_fp16, y = reduce_mean_2_cast_fp16)[name = tensor("sub_2_cast_fp16")]; + tensor square_1_cast_fp16 = square(x = sub_2_cast_fp16)[name = tensor("square_1_cast_fp16")]; + tensor reduce_mean_3_axes_0 = const()[name = tensor("reduce_mean_3_axes_0"), val = tensor([-1])]; + tensor reduce_mean_3_keep_dims_0 = const()[name = tensor("reduce_mean_3_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_3_cast_fp16 = reduce_mean(axes = reduce_mean_3_axes_0, keep_dims = reduce_mean_3_keep_dims_0, x = square_1_cast_fp16)[name = tensor("reduce_mean_3_cast_fp16")]; + tensor real_div_1_to_fp16 = const()[name = tensor("real_div_1_to_fp16"), val = tensor(0x1.008p+0)]; + tensor mul_1_cast_fp16 = mul(x = reduce_mean_3_cast_fp16, y = real_div_1_to_fp16)[name = tensor("mul_1_cast_fp16")]; + tensor var_110_to_fp16 = const()[name = tensor("op_110_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_3_cast_fp16 = add(x = mul_1_cast_fp16, y = var_110_to_fp16)[name = tensor("var_3_cast_fp16")]; + tensor var_113_epsilon_0 = const()[name = tensor("op_113_epsilon_0"), val = tensor(0x1.197998p-40)]; + tensor var_113_cast_fp16 = rsqrt(epsilon = var_113_epsilon_0, x = var_3_cast_fp16)[name = tensor("op_113_cast_fp16")]; + tensor const_5_to_fp16 = const()[name = tensor("const_5_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1088128)))]; + tensor var_114_cast_fp16 = mul(x = const_5_to_fp16, y = var_113_cast_fp16)[name = tensor("op_114_cast_fp16")]; + tensor var_115_cast_fp16 = mul(x = linear_4_cast_fp16, y = var_114_cast_fp16)[name = tensor("op_115_cast_fp16")]; + tensor var_127_cast_fp16 = add(x = var_75_cast_fp16, y = var_115_cast_fp16)[name = tensor("op_127_cast_fp16")]; + tensor _inversed_t_combined_1_y_0_to_fp16 = const()[name = tensor("_inversed_t_combined_1_y_0_to_fp16"), val = tensor(0x1p-1)]; + tensor _inversed_t_combined_1_cast_fp16 = mul(x = var_127_cast_fp16, y = _inversed_t_combined_1_y_0_to_fp16)[name = tensor("_inversed_t_combined_1_cast_fp16")]; + tensor transformer_out_to_fp16_dtype_0 = const()[name = tensor("transformer_out_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor flow_net_cond_embed_weight_to_fp16 = const()[name = tensor("flow_net_cond_embed_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(1089216)))]; + tensor flow_net_cond_embed_bias_to_fp16 = const()[name = tensor("flow_net_cond_embed_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(2137856)))]; + tensor transformer_out_to_fp16 = cast(dtype = transformer_out_to_fp16_dtype_0, x = transformer_out)[name = tensor("cast_145")]; + tensor linear_5_cast_fp16 = linear(bias = flow_net_cond_embed_bias_to_fp16, weight = flow_net_cond_embed_weight_to_fp16, x = transformer_out_to_fp16)[name = tensor("linear_5_cast_fp16")]; + tensor input_13_cast_fp16 = add(x = _inversed_t_combined_1_cast_fp16, y = linear_5_cast_fp16)[name = tensor("input_13_cast_fp16")]; + tensor input_15_cast_fp16 = silu(x = input_13_cast_fp16)[name = tensor("input_15_cast_fp16")]; + tensor flow_net_res_blocks_0_adaLN_modulation_1_weight_to_fp16 = const()[name = tensor("flow_net_res_blocks_0_adaLN_modulation_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(2138944)))]; + tensor flow_net_res_blocks_0_adaLN_modulation_1_bias_to_fp16 = const()[name = tensor("flow_net_res_blocks_0_adaLN_modulation_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(3711872)))]; + tensor linear_6_cast_fp16 = linear(bias = flow_net_res_blocks_0_adaLN_modulation_1_bias_to_fp16, weight = flow_net_res_blocks_0_adaLN_modulation_1_weight_to_fp16, x = input_15_cast_fp16)[name = tensor("linear_6_cast_fp16")]; + tensor var_142_split_sizes_0 = const()[name = tensor("op_142_split_sizes_0"), val = tensor([512, 512, 512])]; + tensor var_142_axis_0 = const()[name = tensor("op_142_axis_0"), val = tensor(-1)]; + tensor var_142_cast_fp16_0, tensor var_142_cast_fp16_1, tensor var_142_cast_fp16_2 = split(axis = var_142_axis_0, split_sizes = var_142_split_sizes_0, x = linear_6_cast_fp16)[name = tensor("op_142_cast_fp16")]; + tensor mean_1_axes_0 = const()[name = tensor("mean_1_axes_0"), val = tensor([-1])]; + tensor mean_1_keep_dims_0 = const()[name = tensor("mean_1_keep_dims_0"), val = tensor(true)]; + tensor mean_1_cast_fp16 = reduce_mean(axes = mean_1_axes_0, keep_dims = mean_1_keep_dims_0, x = linear_0_cast_fp16)[name = tensor("mean_1_cast_fp16")]; + tensor sub_4_cast_fp16 = sub(x = linear_0_cast_fp16, y = mean_1_cast_fp16)[name = tensor("sub_4_cast_fp16")]; + tensor square_2_cast_fp16 = square(x = sub_4_cast_fp16)[name = tensor("square_2_cast_fp16")]; + tensor reduce_mean_5_axes_0 = const()[name = tensor("reduce_mean_5_axes_0"), val = tensor([-1])]; + tensor reduce_mean_5_keep_dims_0 = const()[name = tensor("reduce_mean_5_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_5_cast_fp16 = reduce_mean(axes = reduce_mean_5_axes_0, keep_dims = reduce_mean_5_keep_dims_0, x = square_2_cast_fp16)[name = tensor("reduce_mean_5_cast_fp16")]; + tensor var_152_to_fp16 = const()[name = tensor("op_152_to_fp16"), val = tensor(0x1.1p-20)]; + tensor var_153_cast_fp16 = add(x = reduce_mean_5_cast_fp16, y = var_152_to_fp16)[name = tensor("op_153_cast_fp16")]; + tensor var_154_cast_fp16 = sqrt(x = var_153_cast_fp16)[name = tensor("op_154_cast_fp16")]; + tensor x_7_cast_fp16 = real_div(x = sub_4_cast_fp16, y = var_154_cast_fp16)[name = tensor("x_7_cast_fp16")]; + tensor flow_net_res_blocks_0_in_ln_weight_to_fp16 = const()[name = tensor("flow_net_res_blocks_0_in_ln_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(3715008)))]; + tensor var_156_cast_fp16 = mul(x = x_7_cast_fp16, y = flow_net_res_blocks_0_in_ln_weight_to_fp16)[name = tensor("op_156_cast_fp16")]; + tensor flow_net_res_blocks_0_in_ln_bias_to_fp16 = const()[name = tensor("flow_net_res_blocks_0_in_ln_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(3716096)))]; + tensor x_9_cast_fp16 = add(x = var_156_cast_fp16, y = flow_net_res_blocks_0_in_ln_bias_to_fp16)[name = tensor("x_9_cast_fp16")]; + tensor var_158_promoted_to_fp16 = const()[name = tensor("op_158_promoted_to_fp16"), val = tensor(0x1p+0)]; + tensor var_159_cast_fp16 = add(x = var_142_cast_fp16_1, y = var_158_promoted_to_fp16)[name = tensor("op_159_cast_fp16")]; + tensor var_160_cast_fp16 = mul(x = x_9_cast_fp16, y = var_159_cast_fp16)[name = tensor("op_160_cast_fp16")]; + tensor input_17_cast_fp16 = add(x = var_160_cast_fp16, y = var_142_cast_fp16_0)[name = tensor("input_17_cast_fp16")]; + tensor flow_net_res_blocks_0_mlp_0_weight_to_fp16 = const()[name = tensor("flow_net_res_blocks_0_mlp_0_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(3717184)))]; + tensor flow_net_res_blocks_0_mlp_0_bias_to_fp16 = const()[name = tensor("flow_net_res_blocks_0_mlp_0_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(4241536)))]; + tensor linear_7_cast_fp16 = linear(bias = flow_net_res_blocks_0_mlp_0_bias_to_fp16, weight = flow_net_res_blocks_0_mlp_0_weight_to_fp16, x = input_17_cast_fp16)[name = tensor("linear_7_cast_fp16")]; + tensor input_21_cast_fp16 = silu(x = linear_7_cast_fp16)[name = tensor("input_21_cast_fp16")]; + tensor flow_net_res_blocks_0_mlp_2_weight_to_fp16 = const()[name = tensor("flow_net_res_blocks_0_mlp_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(4242624)))]; + tensor flow_net_res_blocks_0_mlp_2_bias_to_fp16 = const()[name = tensor("flow_net_res_blocks_0_mlp_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(4766976)))]; + tensor linear_8_cast_fp16 = linear(bias = flow_net_res_blocks_0_mlp_2_bias_to_fp16, weight = flow_net_res_blocks_0_mlp_2_weight_to_fp16, x = input_21_cast_fp16)[name = tensor("linear_8_cast_fp16")]; + tensor var_171_cast_fp16 = mul(x = var_142_cast_fp16_2, y = linear_8_cast_fp16)[name = tensor("op_171_cast_fp16")]; + tensor x_11_cast_fp16 = add(x = linear_0_cast_fp16, y = var_171_cast_fp16)[name = tensor("x_11_cast_fp16")]; + tensor flow_net_res_blocks_1_adaLN_modulation_1_weight_to_fp16 = const()[name = tensor("flow_net_res_blocks_1_adaLN_modulation_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(4768064)))]; + tensor flow_net_res_blocks_1_adaLN_modulation_1_bias_to_fp16 = const()[name = tensor("flow_net_res_blocks_1_adaLN_modulation_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(6340992)))]; + tensor linear_9_cast_fp16 = linear(bias = flow_net_res_blocks_1_adaLN_modulation_1_bias_to_fp16, weight = flow_net_res_blocks_1_adaLN_modulation_1_weight_to_fp16, x = input_15_cast_fp16)[name = tensor("linear_9_cast_fp16")]; + tensor var_181_split_sizes_0 = const()[name = tensor("op_181_split_sizes_0"), val = tensor([512, 512, 512])]; + tensor var_181_axis_0 = const()[name = tensor("op_181_axis_0"), val = tensor(-1)]; + tensor var_181_cast_fp16_0, tensor var_181_cast_fp16_1, tensor var_181_cast_fp16_2 = split(axis = var_181_axis_0, split_sizes = var_181_split_sizes_0, x = linear_9_cast_fp16)[name = tensor("op_181_cast_fp16")]; + tensor mean_3_axes_0 = const()[name = tensor("mean_3_axes_0"), val = tensor([-1])]; + tensor mean_3_keep_dims_0 = const()[name = tensor("mean_3_keep_dims_0"), val = tensor(true)]; + tensor mean_3_cast_fp16 = reduce_mean(axes = mean_3_axes_0, keep_dims = mean_3_keep_dims_0, x = x_11_cast_fp16)[name = tensor("mean_3_cast_fp16")]; + tensor sub_5_cast_fp16 = sub(x = x_11_cast_fp16, y = mean_3_cast_fp16)[name = tensor("sub_5_cast_fp16")]; + tensor square_3_cast_fp16 = square(x = sub_5_cast_fp16)[name = tensor("square_3_cast_fp16")]; + tensor reduce_mean_7_axes_0 = const()[name = tensor("reduce_mean_7_axes_0"), val = tensor([-1])]; + tensor reduce_mean_7_keep_dims_0 = const()[name = tensor("reduce_mean_7_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_7_cast_fp16 = reduce_mean(axes = reduce_mean_7_axes_0, keep_dims = reduce_mean_7_keep_dims_0, x = square_3_cast_fp16)[name = tensor("reduce_mean_7_cast_fp16")]; + tensor var_191_to_fp16 = const()[name = tensor("op_191_to_fp16"), val = tensor(0x1.1p-20)]; + tensor var_192_cast_fp16 = add(x = reduce_mean_7_cast_fp16, y = var_191_to_fp16)[name = tensor("op_192_cast_fp16")]; + tensor var_193_cast_fp16 = sqrt(x = var_192_cast_fp16)[name = tensor("op_193_cast_fp16")]; + tensor x_13_cast_fp16 = real_div(x = sub_5_cast_fp16, y = var_193_cast_fp16)[name = tensor("x_13_cast_fp16")]; + tensor flow_net_res_blocks_1_in_ln_weight_to_fp16 = const()[name = tensor("flow_net_res_blocks_1_in_ln_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(6344128)))]; + tensor var_195_cast_fp16 = mul(x = x_13_cast_fp16, y = flow_net_res_blocks_1_in_ln_weight_to_fp16)[name = tensor("op_195_cast_fp16")]; + tensor flow_net_res_blocks_1_in_ln_bias_to_fp16 = const()[name = tensor("flow_net_res_blocks_1_in_ln_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(6345216)))]; + tensor x_15_cast_fp16 = add(x = var_195_cast_fp16, y = flow_net_res_blocks_1_in_ln_bias_to_fp16)[name = tensor("x_15_cast_fp16")]; + tensor var_197_promoted_to_fp16 = const()[name = tensor("op_197_promoted_to_fp16"), val = tensor(0x1p+0)]; + tensor var_198_cast_fp16 = add(x = var_181_cast_fp16_1, y = var_197_promoted_to_fp16)[name = tensor("op_198_cast_fp16")]; + tensor var_199_cast_fp16 = mul(x = x_15_cast_fp16, y = var_198_cast_fp16)[name = tensor("op_199_cast_fp16")]; + tensor input_25_cast_fp16 = add(x = var_199_cast_fp16, y = var_181_cast_fp16_0)[name = tensor("input_25_cast_fp16")]; + tensor flow_net_res_blocks_1_mlp_0_weight_to_fp16 = const()[name = tensor("flow_net_res_blocks_1_mlp_0_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(6346304)))]; + tensor flow_net_res_blocks_1_mlp_0_bias_to_fp16 = const()[name = tensor("flow_net_res_blocks_1_mlp_0_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(6870656)))]; + tensor linear_10_cast_fp16 = linear(bias = flow_net_res_blocks_1_mlp_0_bias_to_fp16, weight = flow_net_res_blocks_1_mlp_0_weight_to_fp16, x = input_25_cast_fp16)[name = tensor("linear_10_cast_fp16")]; + tensor input_29_cast_fp16 = silu(x = linear_10_cast_fp16)[name = tensor("input_29_cast_fp16")]; + tensor flow_net_res_blocks_1_mlp_2_weight_to_fp16 = const()[name = tensor("flow_net_res_blocks_1_mlp_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(6871744)))]; + tensor flow_net_res_blocks_1_mlp_2_bias_to_fp16 = const()[name = tensor("flow_net_res_blocks_1_mlp_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(7396096)))]; + tensor linear_11_cast_fp16 = linear(bias = flow_net_res_blocks_1_mlp_2_bias_to_fp16, weight = flow_net_res_blocks_1_mlp_2_weight_to_fp16, x = input_29_cast_fp16)[name = tensor("linear_11_cast_fp16")]; + tensor var_210_cast_fp16 = mul(x = var_181_cast_fp16_2, y = linear_11_cast_fp16)[name = tensor("op_210_cast_fp16")]; + tensor x_17_cast_fp16 = add(x = x_11_cast_fp16, y = var_210_cast_fp16)[name = tensor("x_17_cast_fp16")]; + tensor flow_net_res_blocks_2_adaLN_modulation_1_weight_to_fp16 = const()[name = tensor("flow_net_res_blocks_2_adaLN_modulation_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(7397184)))]; + tensor flow_net_res_blocks_2_adaLN_modulation_1_bias_to_fp16 = const()[name = tensor("flow_net_res_blocks_2_adaLN_modulation_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8970112)))]; + tensor linear_12_cast_fp16 = linear(bias = flow_net_res_blocks_2_adaLN_modulation_1_bias_to_fp16, weight = flow_net_res_blocks_2_adaLN_modulation_1_weight_to_fp16, x = input_15_cast_fp16)[name = tensor("linear_12_cast_fp16")]; + tensor var_220_split_sizes_0 = const()[name = tensor("op_220_split_sizes_0"), val = tensor([512, 512, 512])]; + tensor var_220_axis_0 = const()[name = tensor("op_220_axis_0"), val = tensor(-1)]; + tensor var_220_cast_fp16_0, tensor var_220_cast_fp16_1, tensor var_220_cast_fp16_2 = split(axis = var_220_axis_0, split_sizes = var_220_split_sizes_0, x = linear_12_cast_fp16)[name = tensor("op_220_cast_fp16")]; + tensor mean_5_axes_0 = const()[name = tensor("mean_5_axes_0"), val = tensor([-1])]; + tensor mean_5_keep_dims_0 = const()[name = tensor("mean_5_keep_dims_0"), val = tensor(true)]; + tensor mean_5_cast_fp16 = reduce_mean(axes = mean_5_axes_0, keep_dims = mean_5_keep_dims_0, x = x_17_cast_fp16)[name = tensor("mean_5_cast_fp16")]; + tensor sub_6_cast_fp16 = sub(x = x_17_cast_fp16, y = mean_5_cast_fp16)[name = tensor("sub_6_cast_fp16")]; + tensor square_4_cast_fp16 = square(x = sub_6_cast_fp16)[name = tensor("square_4_cast_fp16")]; + tensor reduce_mean_9_axes_0 = const()[name = tensor("reduce_mean_9_axes_0"), val = tensor([-1])]; + tensor reduce_mean_9_keep_dims_0 = const()[name = tensor("reduce_mean_9_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_9_cast_fp16 = reduce_mean(axes = reduce_mean_9_axes_0, keep_dims = reduce_mean_9_keep_dims_0, x = square_4_cast_fp16)[name = tensor("reduce_mean_9_cast_fp16")]; + tensor var_230_to_fp16 = const()[name = tensor("op_230_to_fp16"), val = tensor(0x1.1p-20)]; + tensor var_231_cast_fp16 = add(x = reduce_mean_9_cast_fp16, y = var_230_to_fp16)[name = tensor("op_231_cast_fp16")]; + tensor var_232_cast_fp16 = sqrt(x = var_231_cast_fp16)[name = tensor("op_232_cast_fp16")]; + tensor x_19_cast_fp16 = real_div(x = sub_6_cast_fp16, y = var_232_cast_fp16)[name = tensor("x_19_cast_fp16")]; + tensor flow_net_res_blocks_2_in_ln_weight_to_fp16 = const()[name = tensor("flow_net_res_blocks_2_in_ln_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8973248)))]; + tensor var_234_cast_fp16 = mul(x = x_19_cast_fp16, y = flow_net_res_blocks_2_in_ln_weight_to_fp16)[name = tensor("op_234_cast_fp16")]; + tensor flow_net_res_blocks_2_in_ln_bias_to_fp16 = const()[name = tensor("flow_net_res_blocks_2_in_ln_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8974336)))]; + tensor x_21_cast_fp16 = add(x = var_234_cast_fp16, y = flow_net_res_blocks_2_in_ln_bias_to_fp16)[name = tensor("x_21_cast_fp16")]; + tensor var_236_promoted_to_fp16 = const()[name = tensor("op_236_promoted_to_fp16"), val = tensor(0x1p+0)]; + tensor var_237_cast_fp16 = add(x = var_220_cast_fp16_1, y = var_236_promoted_to_fp16)[name = tensor("op_237_cast_fp16")]; + tensor var_238_cast_fp16 = mul(x = x_21_cast_fp16, y = var_237_cast_fp16)[name = tensor("op_238_cast_fp16")]; + tensor input_33_cast_fp16 = add(x = var_238_cast_fp16, y = var_220_cast_fp16_0)[name = tensor("input_33_cast_fp16")]; + tensor flow_net_res_blocks_2_mlp_0_weight_to_fp16 = const()[name = tensor("flow_net_res_blocks_2_mlp_0_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8975424)))]; + tensor flow_net_res_blocks_2_mlp_0_bias_to_fp16 = const()[name = tensor("flow_net_res_blocks_2_mlp_0_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9499776)))]; + tensor linear_13_cast_fp16 = linear(bias = flow_net_res_blocks_2_mlp_0_bias_to_fp16, weight = flow_net_res_blocks_2_mlp_0_weight_to_fp16, x = input_33_cast_fp16)[name = tensor("linear_13_cast_fp16")]; + tensor input_37_cast_fp16 = silu(x = linear_13_cast_fp16)[name = tensor("input_37_cast_fp16")]; + tensor flow_net_res_blocks_2_mlp_2_weight_to_fp16 = const()[name = tensor("flow_net_res_blocks_2_mlp_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9500864)))]; + tensor flow_net_res_blocks_2_mlp_2_bias_to_fp16 = const()[name = tensor("flow_net_res_blocks_2_mlp_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(10025216)))]; + tensor linear_14_cast_fp16 = linear(bias = flow_net_res_blocks_2_mlp_2_bias_to_fp16, weight = flow_net_res_blocks_2_mlp_2_weight_to_fp16, x = input_37_cast_fp16)[name = tensor("linear_14_cast_fp16")]; + tensor var_249_cast_fp16 = mul(x = var_220_cast_fp16_2, y = linear_14_cast_fp16)[name = tensor("op_249_cast_fp16")]; + tensor x_23_cast_fp16 = add(x = x_17_cast_fp16, y = var_249_cast_fp16)[name = tensor("x_23_cast_fp16")]; + tensor flow_net_res_blocks_3_adaLN_modulation_1_weight_to_fp16 = const()[name = tensor("flow_net_res_blocks_3_adaLN_modulation_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(10026304)))]; + tensor flow_net_res_blocks_3_adaLN_modulation_1_bias_to_fp16 = const()[name = tensor("flow_net_res_blocks_3_adaLN_modulation_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(11599232)))]; + tensor linear_15_cast_fp16 = linear(bias = flow_net_res_blocks_3_adaLN_modulation_1_bias_to_fp16, weight = flow_net_res_blocks_3_adaLN_modulation_1_weight_to_fp16, x = input_15_cast_fp16)[name = tensor("linear_15_cast_fp16")]; + tensor var_259_split_sizes_0 = const()[name = tensor("op_259_split_sizes_0"), val = tensor([512, 512, 512])]; + tensor var_259_axis_0 = const()[name = tensor("op_259_axis_0"), val = tensor(-1)]; + tensor var_259_cast_fp16_0, tensor var_259_cast_fp16_1, tensor var_259_cast_fp16_2 = split(axis = var_259_axis_0, split_sizes = var_259_split_sizes_0, x = linear_15_cast_fp16)[name = tensor("op_259_cast_fp16")]; + tensor mean_7_axes_0 = const()[name = tensor("mean_7_axes_0"), val = tensor([-1])]; + tensor mean_7_keep_dims_0 = const()[name = tensor("mean_7_keep_dims_0"), val = tensor(true)]; + tensor mean_7_cast_fp16 = reduce_mean(axes = mean_7_axes_0, keep_dims = mean_7_keep_dims_0, x = x_23_cast_fp16)[name = tensor("mean_7_cast_fp16")]; + tensor sub_7_cast_fp16 = sub(x = x_23_cast_fp16, y = mean_7_cast_fp16)[name = tensor("sub_7_cast_fp16")]; + tensor square_5_cast_fp16 = square(x = sub_7_cast_fp16)[name = tensor("square_5_cast_fp16")]; + tensor reduce_mean_11_axes_0 = const()[name = tensor("reduce_mean_11_axes_0"), val = tensor([-1])]; + tensor reduce_mean_11_keep_dims_0 = const()[name = tensor("reduce_mean_11_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_11_cast_fp16 = reduce_mean(axes = reduce_mean_11_axes_0, keep_dims = reduce_mean_11_keep_dims_0, x = square_5_cast_fp16)[name = tensor("reduce_mean_11_cast_fp16")]; + tensor var_269_to_fp16 = const()[name = tensor("op_269_to_fp16"), val = tensor(0x1.1p-20)]; + tensor var_270_cast_fp16 = add(x = reduce_mean_11_cast_fp16, y = var_269_to_fp16)[name = tensor("op_270_cast_fp16")]; + tensor var_271_cast_fp16 = sqrt(x = var_270_cast_fp16)[name = tensor("op_271_cast_fp16")]; + tensor x_25_cast_fp16 = real_div(x = sub_7_cast_fp16, y = var_271_cast_fp16)[name = tensor("x_25_cast_fp16")]; + tensor flow_net_res_blocks_3_in_ln_weight_to_fp16 = const()[name = tensor("flow_net_res_blocks_3_in_ln_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(11602368)))]; + tensor var_273_cast_fp16 = mul(x = x_25_cast_fp16, y = flow_net_res_blocks_3_in_ln_weight_to_fp16)[name = tensor("op_273_cast_fp16")]; + tensor flow_net_res_blocks_3_in_ln_bias_to_fp16 = const()[name = tensor("flow_net_res_blocks_3_in_ln_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(11603456)))]; + tensor x_27_cast_fp16 = add(x = var_273_cast_fp16, y = flow_net_res_blocks_3_in_ln_bias_to_fp16)[name = tensor("x_27_cast_fp16")]; + tensor var_275_promoted_to_fp16 = const()[name = tensor("op_275_promoted_to_fp16"), val = tensor(0x1p+0)]; + tensor var_276_cast_fp16 = add(x = var_259_cast_fp16_1, y = var_275_promoted_to_fp16)[name = tensor("op_276_cast_fp16")]; + tensor var_277_cast_fp16 = mul(x = x_27_cast_fp16, y = var_276_cast_fp16)[name = tensor("op_277_cast_fp16")]; + tensor input_41_cast_fp16 = add(x = var_277_cast_fp16, y = var_259_cast_fp16_0)[name = tensor("input_41_cast_fp16")]; + tensor flow_net_res_blocks_3_mlp_0_weight_to_fp16 = const()[name = tensor("flow_net_res_blocks_3_mlp_0_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(11604544)))]; + tensor flow_net_res_blocks_3_mlp_0_bias_to_fp16 = const()[name = tensor("flow_net_res_blocks_3_mlp_0_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(12128896)))]; + tensor linear_16_cast_fp16 = linear(bias = flow_net_res_blocks_3_mlp_0_bias_to_fp16, weight = flow_net_res_blocks_3_mlp_0_weight_to_fp16, x = input_41_cast_fp16)[name = tensor("linear_16_cast_fp16")]; + tensor input_45_cast_fp16 = silu(x = linear_16_cast_fp16)[name = tensor("input_45_cast_fp16")]; + tensor flow_net_res_blocks_3_mlp_2_weight_to_fp16 = const()[name = tensor("flow_net_res_blocks_3_mlp_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(12129984)))]; + tensor flow_net_res_blocks_3_mlp_2_bias_to_fp16 = const()[name = tensor("flow_net_res_blocks_3_mlp_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(12654336)))]; + tensor linear_17_cast_fp16 = linear(bias = flow_net_res_blocks_3_mlp_2_bias_to_fp16, weight = flow_net_res_blocks_3_mlp_2_weight_to_fp16, x = input_45_cast_fp16)[name = tensor("linear_17_cast_fp16")]; + tensor var_288_cast_fp16 = mul(x = var_259_cast_fp16_2, y = linear_17_cast_fp16)[name = tensor("op_288_cast_fp16")]; + tensor x_29_cast_fp16 = add(x = x_23_cast_fp16, y = var_288_cast_fp16)[name = tensor("x_29_cast_fp16")]; + tensor flow_net_res_blocks_4_adaLN_modulation_1_weight_to_fp16 = const()[name = tensor("flow_net_res_blocks_4_adaLN_modulation_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(12655424)))]; + tensor flow_net_res_blocks_4_adaLN_modulation_1_bias_to_fp16 = const()[name = tensor("flow_net_res_blocks_4_adaLN_modulation_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(14228352)))]; + tensor linear_18_cast_fp16 = linear(bias = flow_net_res_blocks_4_adaLN_modulation_1_bias_to_fp16, weight = flow_net_res_blocks_4_adaLN_modulation_1_weight_to_fp16, x = input_15_cast_fp16)[name = tensor("linear_18_cast_fp16")]; + tensor var_298_split_sizes_0 = const()[name = tensor("op_298_split_sizes_0"), val = tensor([512, 512, 512])]; + tensor var_298_axis_0 = const()[name = tensor("op_298_axis_0"), val = tensor(-1)]; + tensor var_298_cast_fp16_0, tensor var_298_cast_fp16_1, tensor var_298_cast_fp16_2 = split(axis = var_298_axis_0, split_sizes = var_298_split_sizes_0, x = linear_18_cast_fp16)[name = tensor("op_298_cast_fp16")]; + tensor mean_9_axes_0 = const()[name = tensor("mean_9_axes_0"), val = tensor([-1])]; + tensor mean_9_keep_dims_0 = const()[name = tensor("mean_9_keep_dims_0"), val = tensor(true)]; + tensor mean_9_cast_fp16 = reduce_mean(axes = mean_9_axes_0, keep_dims = mean_9_keep_dims_0, x = x_29_cast_fp16)[name = tensor("mean_9_cast_fp16")]; + tensor sub_8_cast_fp16 = sub(x = x_29_cast_fp16, y = mean_9_cast_fp16)[name = tensor("sub_8_cast_fp16")]; + tensor square_6_cast_fp16 = square(x = sub_8_cast_fp16)[name = tensor("square_6_cast_fp16")]; + tensor reduce_mean_13_axes_0 = const()[name = tensor("reduce_mean_13_axes_0"), val = tensor([-1])]; + tensor reduce_mean_13_keep_dims_0 = const()[name = tensor("reduce_mean_13_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_13_cast_fp16 = reduce_mean(axes = reduce_mean_13_axes_0, keep_dims = reduce_mean_13_keep_dims_0, x = square_6_cast_fp16)[name = tensor("reduce_mean_13_cast_fp16")]; + tensor var_308_to_fp16 = const()[name = tensor("op_308_to_fp16"), val = tensor(0x1.1p-20)]; + tensor var_309_cast_fp16 = add(x = reduce_mean_13_cast_fp16, y = var_308_to_fp16)[name = tensor("op_309_cast_fp16")]; + tensor var_310_cast_fp16 = sqrt(x = var_309_cast_fp16)[name = tensor("op_310_cast_fp16")]; + tensor x_31_cast_fp16 = real_div(x = sub_8_cast_fp16, y = var_310_cast_fp16)[name = tensor("x_31_cast_fp16")]; + tensor flow_net_res_blocks_4_in_ln_weight_to_fp16 = const()[name = tensor("flow_net_res_blocks_4_in_ln_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(14231488)))]; + tensor var_312_cast_fp16 = mul(x = x_31_cast_fp16, y = flow_net_res_blocks_4_in_ln_weight_to_fp16)[name = tensor("op_312_cast_fp16")]; + tensor flow_net_res_blocks_4_in_ln_bias_to_fp16 = const()[name = tensor("flow_net_res_blocks_4_in_ln_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(14232576)))]; + tensor x_33_cast_fp16 = add(x = var_312_cast_fp16, y = flow_net_res_blocks_4_in_ln_bias_to_fp16)[name = tensor("x_33_cast_fp16")]; + tensor var_314_promoted_to_fp16 = const()[name = tensor("op_314_promoted_to_fp16"), val = tensor(0x1p+0)]; + tensor var_315_cast_fp16 = add(x = var_298_cast_fp16_1, y = var_314_promoted_to_fp16)[name = tensor("op_315_cast_fp16")]; + tensor var_316_cast_fp16 = mul(x = x_33_cast_fp16, y = var_315_cast_fp16)[name = tensor("op_316_cast_fp16")]; + tensor input_49_cast_fp16 = add(x = var_316_cast_fp16, y = var_298_cast_fp16_0)[name = tensor("input_49_cast_fp16")]; + tensor flow_net_res_blocks_4_mlp_0_weight_to_fp16 = const()[name = tensor("flow_net_res_blocks_4_mlp_0_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(14233664)))]; + tensor flow_net_res_blocks_4_mlp_0_bias_to_fp16 = const()[name = tensor("flow_net_res_blocks_4_mlp_0_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(14758016)))]; + tensor linear_19_cast_fp16 = linear(bias = flow_net_res_blocks_4_mlp_0_bias_to_fp16, weight = flow_net_res_blocks_4_mlp_0_weight_to_fp16, x = input_49_cast_fp16)[name = tensor("linear_19_cast_fp16")]; + tensor input_53_cast_fp16 = silu(x = linear_19_cast_fp16)[name = tensor("input_53_cast_fp16")]; + tensor flow_net_res_blocks_4_mlp_2_weight_to_fp16 = const()[name = tensor("flow_net_res_blocks_4_mlp_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(14759104)))]; + tensor flow_net_res_blocks_4_mlp_2_bias_to_fp16 = const()[name = tensor("flow_net_res_blocks_4_mlp_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15283456)))]; + tensor linear_20_cast_fp16 = linear(bias = flow_net_res_blocks_4_mlp_2_bias_to_fp16, weight = flow_net_res_blocks_4_mlp_2_weight_to_fp16, x = input_53_cast_fp16)[name = tensor("linear_20_cast_fp16")]; + tensor var_327_cast_fp16 = mul(x = var_298_cast_fp16_2, y = linear_20_cast_fp16)[name = tensor("op_327_cast_fp16")]; + tensor x_35_cast_fp16 = add(x = x_29_cast_fp16, y = var_327_cast_fp16)[name = tensor("x_35_cast_fp16")]; + tensor flow_net_res_blocks_5_adaLN_modulation_1_weight_to_fp16 = const()[name = tensor("flow_net_res_blocks_5_adaLN_modulation_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15284544)))]; + tensor flow_net_res_blocks_5_adaLN_modulation_1_bias_to_fp16 = const()[name = tensor("flow_net_res_blocks_5_adaLN_modulation_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(16857472)))]; + tensor linear_21_cast_fp16 = linear(bias = flow_net_res_blocks_5_adaLN_modulation_1_bias_to_fp16, weight = flow_net_res_blocks_5_adaLN_modulation_1_weight_to_fp16, x = input_15_cast_fp16)[name = tensor("linear_21_cast_fp16")]; + tensor var_337_split_sizes_0 = const()[name = tensor("op_337_split_sizes_0"), val = tensor([512, 512, 512])]; + tensor var_337_axis_0 = const()[name = tensor("op_337_axis_0"), val = tensor(-1)]; + tensor var_337_cast_fp16_0, tensor var_337_cast_fp16_1, tensor var_337_cast_fp16_2 = split(axis = var_337_axis_0, split_sizes = var_337_split_sizes_0, x = linear_21_cast_fp16)[name = tensor("op_337_cast_fp16")]; + tensor mean_11_axes_0 = const()[name = tensor("mean_11_axes_0"), val = tensor([-1])]; + tensor mean_11_keep_dims_0 = const()[name = tensor("mean_11_keep_dims_0"), val = tensor(true)]; + tensor mean_11_cast_fp16 = reduce_mean(axes = mean_11_axes_0, keep_dims = mean_11_keep_dims_0, x = x_35_cast_fp16)[name = tensor("mean_11_cast_fp16")]; + tensor sub_9_cast_fp16 = sub(x = x_35_cast_fp16, y = mean_11_cast_fp16)[name = tensor("sub_9_cast_fp16")]; + tensor square_7_cast_fp16 = square(x = sub_9_cast_fp16)[name = tensor("square_7_cast_fp16")]; + tensor reduce_mean_15_axes_0 = const()[name = tensor("reduce_mean_15_axes_0"), val = tensor([-1])]; + tensor reduce_mean_15_keep_dims_0 = const()[name = tensor("reduce_mean_15_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_15_cast_fp16 = reduce_mean(axes = reduce_mean_15_axes_0, keep_dims = reduce_mean_15_keep_dims_0, x = square_7_cast_fp16)[name = tensor("reduce_mean_15_cast_fp16")]; + tensor var_347_to_fp16 = const()[name = tensor("op_347_to_fp16"), val = tensor(0x1.1p-20)]; + tensor var_348_cast_fp16 = add(x = reduce_mean_15_cast_fp16, y = var_347_to_fp16)[name = tensor("op_348_cast_fp16")]; + tensor var_349_cast_fp16 = sqrt(x = var_348_cast_fp16)[name = tensor("op_349_cast_fp16")]; + tensor x_37_cast_fp16 = real_div(x = sub_9_cast_fp16, y = var_349_cast_fp16)[name = tensor("x_37_cast_fp16")]; + tensor flow_net_res_blocks_5_in_ln_weight_to_fp16 = const()[name = tensor("flow_net_res_blocks_5_in_ln_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(16860608)))]; + tensor var_351_cast_fp16 = mul(x = x_37_cast_fp16, y = flow_net_res_blocks_5_in_ln_weight_to_fp16)[name = tensor("op_351_cast_fp16")]; + tensor flow_net_res_blocks_5_in_ln_bias_to_fp16 = const()[name = tensor("flow_net_res_blocks_5_in_ln_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(16861696)))]; + tensor x_39_cast_fp16 = add(x = var_351_cast_fp16, y = flow_net_res_blocks_5_in_ln_bias_to_fp16)[name = tensor("x_39_cast_fp16")]; + tensor var_353_promoted_to_fp16 = const()[name = tensor("op_353_promoted_to_fp16"), val = tensor(0x1p+0)]; + tensor var_354_cast_fp16 = add(x = var_337_cast_fp16_1, y = var_353_promoted_to_fp16)[name = tensor("op_354_cast_fp16")]; + tensor var_355_cast_fp16 = mul(x = x_39_cast_fp16, y = var_354_cast_fp16)[name = tensor("op_355_cast_fp16")]; + tensor input_57_cast_fp16 = add(x = var_355_cast_fp16, y = var_337_cast_fp16_0)[name = tensor("input_57_cast_fp16")]; + tensor flow_net_res_blocks_5_mlp_0_weight_to_fp16 = const()[name = tensor("flow_net_res_blocks_5_mlp_0_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(16862784)))]; + tensor flow_net_res_blocks_5_mlp_0_bias_to_fp16 = const()[name = tensor("flow_net_res_blocks_5_mlp_0_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(17387136)))]; + tensor linear_22_cast_fp16 = linear(bias = flow_net_res_blocks_5_mlp_0_bias_to_fp16, weight = flow_net_res_blocks_5_mlp_0_weight_to_fp16, x = input_57_cast_fp16)[name = tensor("linear_22_cast_fp16")]; + tensor input_61_cast_fp16 = silu(x = linear_22_cast_fp16)[name = tensor("input_61_cast_fp16")]; + tensor flow_net_res_blocks_5_mlp_2_weight_to_fp16 = const()[name = tensor("flow_net_res_blocks_5_mlp_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(17388224)))]; + tensor flow_net_res_blocks_5_mlp_2_bias_to_fp16 = const()[name = tensor("flow_net_res_blocks_5_mlp_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(17912576)))]; + tensor linear_23_cast_fp16 = linear(bias = flow_net_res_blocks_5_mlp_2_bias_to_fp16, weight = flow_net_res_blocks_5_mlp_2_weight_to_fp16, x = input_61_cast_fp16)[name = tensor("linear_23_cast_fp16")]; + tensor var_366_cast_fp16 = mul(x = var_337_cast_fp16_2, y = linear_23_cast_fp16)[name = tensor("op_366_cast_fp16")]; + tensor x_41_cast_fp16 = add(x = x_35_cast_fp16, y = var_366_cast_fp16)[name = tensor("x_41_cast_fp16")]; + tensor flow_net_final_layer_adaLN_modulation_1_weight_to_fp16 = const()[name = tensor("flow_net_final_layer_adaLN_modulation_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(17913664)))]; + tensor flow_net_final_layer_adaLN_modulation_1_bias_to_fp16 = const()[name = tensor("flow_net_final_layer_adaLN_modulation_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(18962304)))]; + tensor linear_24_cast_fp16 = linear(bias = flow_net_final_layer_adaLN_modulation_1_bias_to_fp16, weight = flow_net_final_layer_adaLN_modulation_1_weight_to_fp16, x = input_15_cast_fp16)[name = tensor("linear_24_cast_fp16")]; + tensor var_375_split_sizes_0 = const()[name = tensor("op_375_split_sizes_0"), val = tensor([512, 512])]; + tensor var_375_axis_0 = const()[name = tensor("op_375_axis_0"), val = tensor(-1)]; + tensor var_375_cast_fp16_0, tensor var_375_cast_fp16_1 = split(axis = var_375_axis_0, split_sizes = var_375_split_sizes_0, x = linear_24_cast_fp16)[name = tensor("op_375_cast_fp16")]; + tensor mean_13_axes_0 = const()[name = tensor("mean_13_axes_0"), val = tensor([-1])]; + tensor mean_13_keep_dims_0 = const()[name = tensor("mean_13_keep_dims_0"), val = tensor(true)]; + tensor mean_13_cast_fp16 = reduce_mean(axes = mean_13_axes_0, keep_dims = mean_13_keep_dims_0, x = x_41_cast_fp16)[name = tensor("mean_13_cast_fp16")]; + tensor sub_10_cast_fp16 = sub(x = x_41_cast_fp16, y = mean_13_cast_fp16)[name = tensor("sub_10_cast_fp16")]; + tensor square_8_cast_fp16 = square(x = sub_10_cast_fp16)[name = tensor("square_8_cast_fp16")]; + tensor reduce_mean_17_axes_0 = const()[name = tensor("reduce_mean_17_axes_0"), val = tensor([-1])]; + tensor reduce_mean_17_keep_dims_0 = const()[name = tensor("reduce_mean_17_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_17_cast_fp16 = reduce_mean(axes = reduce_mean_17_axes_0, keep_dims = reduce_mean_17_keep_dims_0, x = square_8_cast_fp16)[name = tensor("reduce_mean_17_cast_fp16")]; + tensor var_382_to_fp16 = const()[name = tensor("op_382_to_fp16"), val = tensor(0x1.1p-20)]; + tensor var_383_cast_fp16 = add(x = reduce_mean_17_cast_fp16, y = var_382_to_fp16)[name = tensor("op_383_cast_fp16")]; + tensor var_384_cast_fp16 = sqrt(x = var_383_cast_fp16)[name = tensor("op_384_cast_fp16")]; + tensor x_43_cast_fp16 = real_div(x = sub_10_cast_fp16, y = var_384_cast_fp16)[name = tensor("x_43_cast_fp16")]; + tensor var_386_promoted_to_fp16 = const()[name = tensor("op_386_promoted_to_fp16"), val = tensor(0x1p+0)]; + tensor var_387_cast_fp16 = add(x = var_375_cast_fp16_1, y = var_386_promoted_to_fp16)[name = tensor("op_387_cast_fp16")]; + tensor var_388_cast_fp16 = mul(x = x_43_cast_fp16, y = var_387_cast_fp16)[name = tensor("op_388_cast_fp16")]; + tensor input_65_cast_fp16 = add(x = var_388_cast_fp16, y = var_375_cast_fp16_0)[name = tensor("input_65_cast_fp16")]; + tensor flow_net_final_layer_linear_weight_to_fp16 = const()[name = tensor("flow_net_final_layer_linear_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(18964416)))]; + tensor flow_net_final_layer_linear_bias_to_fp16 = const()[name = tensor("flow_net_final_layer_linear_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(18997248)))]; + tensor linear_25_cast_fp16 = linear(bias = flow_net_final_layer_linear_bias_to_fp16, weight = flow_net_final_layer_linear_weight_to_fp16, x = input_65_cast_fp16)[name = tensor("linear_25_cast_fp16")]; + tensor var_399_to_fp16 = const()[name = tensor("op_399_to_fp16"), val = tensor(0x1p-3)]; + tensor var_400_cast_fp16 = mul(x = linear_25_cast_fp16, y = var_399_to_fp16)[name = tensor("op_400_cast_fp16")]; + tensor input_67_cast_fp16 = add(x = latent_init_to_fp16, y = var_400_cast_fp16)[name = tensor("input_67_cast_fp16")]; + tensor linear_26_cast_fp16 = linear(bias = flow_net_input_proj_bias_to_fp16, weight = flow_net_input_proj_weight_to_fp16, x = input_67_cast_fp16)[name = tensor("linear_26_cast_fp16")]; + tensor input_71_to_fp16 = const()[name = tensor("input_71_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(18997376)))]; + tensor input_73_cast_fp16 = silu(x = input_71_to_fp16)[name = tensor("input_73_cast_fp16")]; + tensor linear_28_cast_fp16 = linear(bias = flow_net_time_embed_0_mlp_2_bias_to_fp16, weight = flow_net_time_embed_0_mlp_2_weight_to_fp16, x = input_73_cast_fp16)[name = tensor("linear_28_cast_fp16")]; + tensor reduce_mean_18_axes_0 = const()[name = tensor("reduce_mean_18_axes_0"), val = tensor([-1])]; + tensor reduce_mean_18_keep_dims_0 = const()[name = tensor("reduce_mean_18_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_18_cast_fp16 = reduce_mean(axes = reduce_mean_18_axes_0, keep_dims = reduce_mean_18_keep_dims_0, x = linear_28_cast_fp16)[name = tensor("reduce_mean_18_cast_fp16")]; + tensor sub_11_cast_fp16 = sub(x = linear_28_cast_fp16, y = reduce_mean_18_cast_fp16)[name = tensor("sub_11_cast_fp16")]; + tensor square_9_cast_fp16 = square(x = sub_11_cast_fp16)[name = tensor("square_9_cast_fp16")]; + tensor reduce_mean_19_axes_0 = const()[name = tensor("reduce_mean_19_axes_0"), val = tensor([-1])]; + tensor reduce_mean_19_keep_dims_0 = const()[name = tensor("reduce_mean_19_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_19_cast_fp16 = reduce_mean(axes = reduce_mean_19_axes_0, keep_dims = reduce_mean_19_keep_dims_0, x = square_9_cast_fp16)[name = tensor("reduce_mean_19_cast_fp16")]; + tensor real_div_2_to_fp16 = const()[name = tensor("real_div_2_to_fp16"), val = tensor(0x1.008p+0)]; + tensor mul_2_cast_fp16 = mul(x = reduce_mean_19_cast_fp16, y = real_div_2_to_fp16)[name = tensor("mul_2_cast_fp16")]; + tensor var_466_to_fp16 = const()[name = tensor("op_466_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_19_cast_fp16 = add(x = mul_2_cast_fp16, y = var_466_to_fp16)[name = tensor("var_19_cast_fp16")]; + tensor var_469_epsilon_0 = const()[name = tensor("op_469_epsilon_0"), val = tensor(0x1.197998p-40)]; + tensor var_469_cast_fp16 = rsqrt(epsilon = var_469_epsilon_0, x = var_19_cast_fp16)[name = tensor("op_469_cast_fp16")]; + tensor var_470_cast_fp16 = mul(x = const_3_to_fp16, y = var_469_cast_fp16)[name = tensor("op_470_cast_fp16")]; + tensor var_471_cast_fp16 = mul(x = linear_28_cast_fp16, y = var_470_cast_fp16)[name = tensor("op_471_cast_fp16")]; + tensor input_77_to_fp16 = const()[name = tensor("input_77_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(18998464)))]; + tensor input_79_cast_fp16 = silu(x = input_77_to_fp16)[name = tensor("input_79_cast_fp16")]; + tensor linear_30_cast_fp16 = linear(bias = flow_net_time_embed_1_mlp_2_bias_to_fp16, weight = flow_net_time_embed_1_mlp_2_weight_to_fp16, x = input_79_cast_fp16)[name = tensor("linear_30_cast_fp16")]; + tensor reduce_mean_20_axes_0 = const()[name = tensor("reduce_mean_20_axes_0"), val = tensor([-1])]; + tensor reduce_mean_20_keep_dims_0 = const()[name = tensor("reduce_mean_20_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_20_cast_fp16 = reduce_mean(axes = reduce_mean_20_axes_0, keep_dims = reduce_mean_20_keep_dims_0, x = linear_30_cast_fp16)[name = tensor("reduce_mean_20_cast_fp16")]; + tensor sub_13_cast_fp16 = sub(x = linear_30_cast_fp16, y = reduce_mean_20_cast_fp16)[name = tensor("sub_13_cast_fp16")]; + tensor square_10_cast_fp16 = square(x = sub_13_cast_fp16)[name = tensor("square_10_cast_fp16")]; + tensor reduce_mean_21_axes_0 = const()[name = tensor("reduce_mean_21_axes_0"), val = tensor([-1])]; + tensor reduce_mean_21_keep_dims_0 = const()[name = tensor("reduce_mean_21_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_21_cast_fp16 = reduce_mean(axes = reduce_mean_21_axes_0, keep_dims = reduce_mean_21_keep_dims_0, x = square_10_cast_fp16)[name = tensor("reduce_mean_21_cast_fp16")]; + tensor real_div_3_to_fp16 = const()[name = tensor("real_div_3_to_fp16"), val = tensor(0x1.008p+0)]; + tensor mul_3_cast_fp16 = mul(x = reduce_mean_21_cast_fp16, y = real_div_3_to_fp16)[name = tensor("mul_3_cast_fp16")]; + tensor var_503_to_fp16 = const()[name = tensor("op_503_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_21_cast_fp16 = add(x = mul_3_cast_fp16, y = var_503_to_fp16)[name = tensor("var_21_cast_fp16")]; + tensor var_506_epsilon_0 = const()[name = tensor("op_506_epsilon_0"), val = tensor(0x1.197998p-40)]; + tensor var_506_cast_fp16 = rsqrt(epsilon = var_506_epsilon_0, x = var_21_cast_fp16)[name = tensor("op_506_cast_fp16")]; + tensor var_507_cast_fp16 = mul(x = const_5_to_fp16, y = var_506_cast_fp16)[name = tensor("op_507_cast_fp16")]; + tensor var_508_cast_fp16 = mul(x = linear_30_cast_fp16, y = var_507_cast_fp16)[name = tensor("op_508_cast_fp16")]; + tensor var_520_cast_fp16 = add(x = var_471_cast_fp16, y = var_508_cast_fp16)[name = tensor("op_520_cast_fp16")]; + tensor _inversed_t_combined_3_y_0_to_fp16 = const()[name = tensor("_inversed_t_combined_3_y_0_to_fp16"), val = tensor(0x1p-1)]; + tensor _inversed_t_combined_3_cast_fp16 = mul(x = var_520_cast_fp16, y = _inversed_t_combined_3_y_0_to_fp16)[name = tensor("_inversed_t_combined_3_cast_fp16")]; + tensor input_81_cast_fp16 = add(x = _inversed_t_combined_3_cast_fp16, y = linear_5_cast_fp16)[name = tensor("input_81_cast_fp16")]; + tensor input_83_cast_fp16 = silu(x = input_81_cast_fp16)[name = tensor("input_83_cast_fp16")]; + tensor linear_32_cast_fp16 = linear(bias = flow_net_res_blocks_0_adaLN_modulation_1_bias_to_fp16, weight = flow_net_res_blocks_0_adaLN_modulation_1_weight_to_fp16, x = input_83_cast_fp16)[name = tensor("linear_32_cast_fp16")]; + tensor var_535_split_sizes_0 = const()[name = tensor("op_535_split_sizes_0"), val = tensor([512, 512, 512])]; + tensor var_535_axis_0 = const()[name = tensor("op_535_axis_0"), val = tensor(-1)]; + tensor var_535_cast_fp16_0, tensor var_535_cast_fp16_1, tensor var_535_cast_fp16_2 = split(axis = var_535_axis_0, split_sizes = var_535_split_sizes_0, x = linear_32_cast_fp16)[name = tensor("op_535_cast_fp16")]; + tensor mean_15_axes_0 = const()[name = tensor("mean_15_axes_0"), val = tensor([-1])]; + tensor mean_15_keep_dims_0 = const()[name = tensor("mean_15_keep_dims_0"), val = tensor(true)]; + tensor mean_15_cast_fp16 = reduce_mean(axes = mean_15_axes_0, keep_dims = mean_15_keep_dims_0, x = linear_26_cast_fp16)[name = tensor("mean_15_cast_fp16")]; + tensor sub_15_cast_fp16 = sub(x = linear_26_cast_fp16, y = mean_15_cast_fp16)[name = tensor("sub_15_cast_fp16")]; + tensor square_11_cast_fp16 = square(x = sub_15_cast_fp16)[name = tensor("square_11_cast_fp16")]; + tensor reduce_mean_23_axes_0 = const()[name = tensor("reduce_mean_23_axes_0"), val = tensor([-1])]; + tensor reduce_mean_23_keep_dims_0 = const()[name = tensor("reduce_mean_23_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_23_cast_fp16 = reduce_mean(axes = reduce_mean_23_axes_0, keep_dims = reduce_mean_23_keep_dims_0, x = square_11_cast_fp16)[name = tensor("reduce_mean_23_cast_fp16")]; + tensor var_545_to_fp16 = const()[name = tensor("op_545_to_fp16"), val = tensor(0x1.1p-20)]; + tensor var_546_cast_fp16 = add(x = reduce_mean_23_cast_fp16, y = var_545_to_fp16)[name = tensor("op_546_cast_fp16")]; + tensor var_547_cast_fp16 = sqrt(x = var_546_cast_fp16)[name = tensor("op_547_cast_fp16")]; + tensor x_51_cast_fp16 = real_div(x = sub_15_cast_fp16, y = var_547_cast_fp16)[name = tensor("x_51_cast_fp16")]; + tensor var_549_cast_fp16 = mul(x = x_51_cast_fp16, y = flow_net_res_blocks_0_in_ln_weight_to_fp16)[name = tensor("op_549_cast_fp16")]; + tensor x_53_cast_fp16 = add(x = var_549_cast_fp16, y = flow_net_res_blocks_0_in_ln_bias_to_fp16)[name = tensor("x_53_cast_fp16")]; + tensor var_551_promoted_to_fp16 = const()[name = tensor("op_551_promoted_to_fp16"), val = tensor(0x1p+0)]; + tensor var_552_cast_fp16 = add(x = var_535_cast_fp16_1, y = var_551_promoted_to_fp16)[name = tensor("op_552_cast_fp16")]; + tensor var_553_cast_fp16 = mul(x = x_53_cast_fp16, y = var_552_cast_fp16)[name = tensor("op_553_cast_fp16")]; + tensor input_85_cast_fp16 = add(x = var_553_cast_fp16, y = var_535_cast_fp16_0)[name = tensor("input_85_cast_fp16")]; + tensor linear_33_cast_fp16 = linear(bias = flow_net_res_blocks_0_mlp_0_bias_to_fp16, weight = flow_net_res_blocks_0_mlp_0_weight_to_fp16, x = input_85_cast_fp16)[name = tensor("linear_33_cast_fp16")]; + tensor input_89_cast_fp16 = silu(x = linear_33_cast_fp16)[name = tensor("input_89_cast_fp16")]; + tensor linear_34_cast_fp16 = linear(bias = flow_net_res_blocks_0_mlp_2_bias_to_fp16, weight = flow_net_res_blocks_0_mlp_2_weight_to_fp16, x = input_89_cast_fp16)[name = tensor("linear_34_cast_fp16")]; + tensor var_564_cast_fp16 = mul(x = var_535_cast_fp16_2, y = linear_34_cast_fp16)[name = tensor("op_564_cast_fp16")]; + tensor x_55_cast_fp16 = add(x = linear_26_cast_fp16, y = var_564_cast_fp16)[name = tensor("x_55_cast_fp16")]; + tensor linear_35_cast_fp16 = linear(bias = flow_net_res_blocks_1_adaLN_modulation_1_bias_to_fp16, weight = flow_net_res_blocks_1_adaLN_modulation_1_weight_to_fp16, x = input_83_cast_fp16)[name = tensor("linear_35_cast_fp16")]; + tensor var_574_split_sizes_0 = const()[name = tensor("op_574_split_sizes_0"), val = tensor([512, 512, 512])]; + tensor var_574_axis_0 = const()[name = tensor("op_574_axis_0"), val = tensor(-1)]; + tensor var_574_cast_fp16_0, tensor var_574_cast_fp16_1, tensor var_574_cast_fp16_2 = split(axis = var_574_axis_0, split_sizes = var_574_split_sizes_0, x = linear_35_cast_fp16)[name = tensor("op_574_cast_fp16")]; + tensor mean_17_axes_0 = const()[name = tensor("mean_17_axes_0"), val = tensor([-1])]; + tensor mean_17_keep_dims_0 = const()[name = tensor("mean_17_keep_dims_0"), val = tensor(true)]; + tensor mean_17_cast_fp16 = reduce_mean(axes = mean_17_axes_0, keep_dims = mean_17_keep_dims_0, x = x_55_cast_fp16)[name = tensor("mean_17_cast_fp16")]; + tensor sub_16_cast_fp16 = sub(x = x_55_cast_fp16, y = mean_17_cast_fp16)[name = tensor("sub_16_cast_fp16")]; + tensor square_12_cast_fp16 = square(x = sub_16_cast_fp16)[name = tensor("square_12_cast_fp16")]; + tensor reduce_mean_25_axes_0 = const()[name = tensor("reduce_mean_25_axes_0"), val = tensor([-1])]; + tensor reduce_mean_25_keep_dims_0 = const()[name = tensor("reduce_mean_25_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_25_cast_fp16 = reduce_mean(axes = reduce_mean_25_axes_0, keep_dims = reduce_mean_25_keep_dims_0, x = square_12_cast_fp16)[name = tensor("reduce_mean_25_cast_fp16")]; + tensor var_584_to_fp16 = const()[name = tensor("op_584_to_fp16"), val = tensor(0x1.1p-20)]; + tensor var_585_cast_fp16 = add(x = reduce_mean_25_cast_fp16, y = var_584_to_fp16)[name = tensor("op_585_cast_fp16")]; + tensor var_586_cast_fp16 = sqrt(x = var_585_cast_fp16)[name = tensor("op_586_cast_fp16")]; + tensor x_57_cast_fp16 = real_div(x = sub_16_cast_fp16, y = var_586_cast_fp16)[name = tensor("x_57_cast_fp16")]; + tensor var_588_cast_fp16 = mul(x = x_57_cast_fp16, y = flow_net_res_blocks_1_in_ln_weight_to_fp16)[name = tensor("op_588_cast_fp16")]; + tensor x_59_cast_fp16 = add(x = var_588_cast_fp16, y = flow_net_res_blocks_1_in_ln_bias_to_fp16)[name = tensor("x_59_cast_fp16")]; + tensor var_590_promoted_to_fp16 = const()[name = tensor("op_590_promoted_to_fp16"), val = tensor(0x1p+0)]; + tensor var_591_cast_fp16 = add(x = var_574_cast_fp16_1, y = var_590_promoted_to_fp16)[name = tensor("op_591_cast_fp16")]; + tensor var_592_cast_fp16 = mul(x = x_59_cast_fp16, y = var_591_cast_fp16)[name = tensor("op_592_cast_fp16")]; + tensor input_93_cast_fp16 = add(x = var_592_cast_fp16, y = var_574_cast_fp16_0)[name = tensor("input_93_cast_fp16")]; + tensor linear_36_cast_fp16 = linear(bias = flow_net_res_blocks_1_mlp_0_bias_to_fp16, weight = flow_net_res_blocks_1_mlp_0_weight_to_fp16, x = input_93_cast_fp16)[name = tensor("linear_36_cast_fp16")]; + tensor input_97_cast_fp16 = silu(x = linear_36_cast_fp16)[name = tensor("input_97_cast_fp16")]; + tensor linear_37_cast_fp16 = linear(bias = flow_net_res_blocks_1_mlp_2_bias_to_fp16, weight = flow_net_res_blocks_1_mlp_2_weight_to_fp16, x = input_97_cast_fp16)[name = tensor("linear_37_cast_fp16")]; + tensor var_603_cast_fp16 = mul(x = var_574_cast_fp16_2, y = linear_37_cast_fp16)[name = tensor("op_603_cast_fp16")]; + tensor x_61_cast_fp16 = add(x = x_55_cast_fp16, y = var_603_cast_fp16)[name = tensor("x_61_cast_fp16")]; + tensor linear_38_cast_fp16 = linear(bias = flow_net_res_blocks_2_adaLN_modulation_1_bias_to_fp16, weight = flow_net_res_blocks_2_adaLN_modulation_1_weight_to_fp16, x = input_83_cast_fp16)[name = tensor("linear_38_cast_fp16")]; + tensor var_613_split_sizes_0 = const()[name = tensor("op_613_split_sizes_0"), val = tensor([512, 512, 512])]; + tensor var_613_axis_0 = const()[name = tensor("op_613_axis_0"), val = tensor(-1)]; + tensor var_613_cast_fp16_0, tensor var_613_cast_fp16_1, tensor var_613_cast_fp16_2 = split(axis = var_613_axis_0, split_sizes = var_613_split_sizes_0, x = linear_38_cast_fp16)[name = tensor("op_613_cast_fp16")]; + tensor mean_19_axes_0 = const()[name = tensor("mean_19_axes_0"), val = tensor([-1])]; + tensor mean_19_keep_dims_0 = const()[name = tensor("mean_19_keep_dims_0"), val = tensor(true)]; + tensor mean_19_cast_fp16 = reduce_mean(axes = mean_19_axes_0, keep_dims = mean_19_keep_dims_0, x = x_61_cast_fp16)[name = tensor("mean_19_cast_fp16")]; + tensor sub_17_cast_fp16 = sub(x = x_61_cast_fp16, y = mean_19_cast_fp16)[name = tensor("sub_17_cast_fp16")]; + tensor square_13_cast_fp16 = square(x = sub_17_cast_fp16)[name = tensor("square_13_cast_fp16")]; + tensor reduce_mean_27_axes_0 = const()[name = tensor("reduce_mean_27_axes_0"), val = tensor([-1])]; + tensor reduce_mean_27_keep_dims_0 = const()[name = tensor("reduce_mean_27_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_27_cast_fp16 = reduce_mean(axes = reduce_mean_27_axes_0, keep_dims = reduce_mean_27_keep_dims_0, x = square_13_cast_fp16)[name = tensor("reduce_mean_27_cast_fp16")]; + tensor var_623_to_fp16 = const()[name = tensor("op_623_to_fp16"), val = tensor(0x1.1p-20)]; + tensor var_624_cast_fp16 = add(x = reduce_mean_27_cast_fp16, y = var_623_to_fp16)[name = tensor("op_624_cast_fp16")]; + tensor var_625_cast_fp16 = sqrt(x = var_624_cast_fp16)[name = tensor("op_625_cast_fp16")]; + tensor x_63_cast_fp16 = real_div(x = sub_17_cast_fp16, y = var_625_cast_fp16)[name = tensor("x_63_cast_fp16")]; + tensor var_627_cast_fp16 = mul(x = x_63_cast_fp16, y = flow_net_res_blocks_2_in_ln_weight_to_fp16)[name = tensor("op_627_cast_fp16")]; + tensor x_65_cast_fp16 = add(x = var_627_cast_fp16, y = flow_net_res_blocks_2_in_ln_bias_to_fp16)[name = tensor("x_65_cast_fp16")]; + tensor var_629_promoted_to_fp16 = const()[name = tensor("op_629_promoted_to_fp16"), val = tensor(0x1p+0)]; + tensor var_630_cast_fp16 = add(x = var_613_cast_fp16_1, y = var_629_promoted_to_fp16)[name = tensor("op_630_cast_fp16")]; + tensor var_631_cast_fp16 = mul(x = x_65_cast_fp16, y = var_630_cast_fp16)[name = tensor("op_631_cast_fp16")]; + tensor input_101_cast_fp16 = add(x = var_631_cast_fp16, y = var_613_cast_fp16_0)[name = tensor("input_101_cast_fp16")]; + tensor linear_39_cast_fp16 = linear(bias = flow_net_res_blocks_2_mlp_0_bias_to_fp16, weight = flow_net_res_blocks_2_mlp_0_weight_to_fp16, x = input_101_cast_fp16)[name = tensor("linear_39_cast_fp16")]; + tensor input_105_cast_fp16 = silu(x = linear_39_cast_fp16)[name = tensor("input_105_cast_fp16")]; + tensor linear_40_cast_fp16 = linear(bias = flow_net_res_blocks_2_mlp_2_bias_to_fp16, weight = flow_net_res_blocks_2_mlp_2_weight_to_fp16, x = input_105_cast_fp16)[name = tensor("linear_40_cast_fp16")]; + tensor var_642_cast_fp16 = mul(x = var_613_cast_fp16_2, y = linear_40_cast_fp16)[name = tensor("op_642_cast_fp16")]; + tensor x_67_cast_fp16 = add(x = x_61_cast_fp16, y = var_642_cast_fp16)[name = tensor("x_67_cast_fp16")]; + tensor linear_41_cast_fp16 = linear(bias = flow_net_res_blocks_3_adaLN_modulation_1_bias_to_fp16, weight = flow_net_res_blocks_3_adaLN_modulation_1_weight_to_fp16, x = input_83_cast_fp16)[name = tensor("linear_41_cast_fp16")]; + tensor var_652_split_sizes_0 = const()[name = tensor("op_652_split_sizes_0"), val = tensor([512, 512, 512])]; + tensor var_652_axis_0 = const()[name = tensor("op_652_axis_0"), val = tensor(-1)]; + tensor var_652_cast_fp16_0, tensor var_652_cast_fp16_1, tensor var_652_cast_fp16_2 = split(axis = var_652_axis_0, split_sizes = var_652_split_sizes_0, x = linear_41_cast_fp16)[name = tensor("op_652_cast_fp16")]; + tensor mean_21_axes_0 = const()[name = tensor("mean_21_axes_0"), val = tensor([-1])]; + tensor mean_21_keep_dims_0 = const()[name = tensor("mean_21_keep_dims_0"), val = tensor(true)]; + tensor mean_21_cast_fp16 = reduce_mean(axes = mean_21_axes_0, keep_dims = mean_21_keep_dims_0, x = x_67_cast_fp16)[name = tensor("mean_21_cast_fp16")]; + tensor sub_18_cast_fp16 = sub(x = x_67_cast_fp16, y = mean_21_cast_fp16)[name = tensor("sub_18_cast_fp16")]; + tensor square_14_cast_fp16 = square(x = sub_18_cast_fp16)[name = tensor("square_14_cast_fp16")]; + tensor reduce_mean_29_axes_0 = const()[name = tensor("reduce_mean_29_axes_0"), val = tensor([-1])]; + tensor reduce_mean_29_keep_dims_0 = const()[name = tensor("reduce_mean_29_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_29_cast_fp16 = reduce_mean(axes = reduce_mean_29_axes_0, keep_dims = reduce_mean_29_keep_dims_0, x = square_14_cast_fp16)[name = tensor("reduce_mean_29_cast_fp16")]; + tensor var_662_to_fp16 = const()[name = tensor("op_662_to_fp16"), val = tensor(0x1.1p-20)]; + tensor var_663_cast_fp16 = add(x = reduce_mean_29_cast_fp16, y = var_662_to_fp16)[name = tensor("op_663_cast_fp16")]; + tensor var_664_cast_fp16 = sqrt(x = var_663_cast_fp16)[name = tensor("op_664_cast_fp16")]; + tensor x_69_cast_fp16 = real_div(x = sub_18_cast_fp16, y = var_664_cast_fp16)[name = tensor("x_69_cast_fp16")]; + tensor var_666_cast_fp16 = mul(x = x_69_cast_fp16, y = flow_net_res_blocks_3_in_ln_weight_to_fp16)[name = tensor("op_666_cast_fp16")]; + tensor x_71_cast_fp16 = add(x = var_666_cast_fp16, y = flow_net_res_blocks_3_in_ln_bias_to_fp16)[name = tensor("x_71_cast_fp16")]; + tensor var_668_promoted_to_fp16 = const()[name = tensor("op_668_promoted_to_fp16"), val = tensor(0x1p+0)]; + tensor var_669_cast_fp16 = add(x = var_652_cast_fp16_1, y = var_668_promoted_to_fp16)[name = tensor("op_669_cast_fp16")]; + tensor var_670_cast_fp16 = mul(x = x_71_cast_fp16, y = var_669_cast_fp16)[name = tensor("op_670_cast_fp16")]; + tensor input_109_cast_fp16 = add(x = var_670_cast_fp16, y = var_652_cast_fp16_0)[name = tensor("input_109_cast_fp16")]; + tensor linear_42_cast_fp16 = linear(bias = flow_net_res_blocks_3_mlp_0_bias_to_fp16, weight = flow_net_res_blocks_3_mlp_0_weight_to_fp16, x = input_109_cast_fp16)[name = tensor("linear_42_cast_fp16")]; + tensor input_113_cast_fp16 = silu(x = linear_42_cast_fp16)[name = tensor("input_113_cast_fp16")]; + tensor linear_43_cast_fp16 = linear(bias = flow_net_res_blocks_3_mlp_2_bias_to_fp16, weight = flow_net_res_blocks_3_mlp_2_weight_to_fp16, x = input_113_cast_fp16)[name = tensor("linear_43_cast_fp16")]; + tensor var_681_cast_fp16 = mul(x = var_652_cast_fp16_2, y = linear_43_cast_fp16)[name = tensor("op_681_cast_fp16")]; + tensor x_73_cast_fp16 = add(x = x_67_cast_fp16, y = var_681_cast_fp16)[name = tensor("x_73_cast_fp16")]; + tensor linear_44_cast_fp16 = linear(bias = flow_net_res_blocks_4_adaLN_modulation_1_bias_to_fp16, weight = flow_net_res_blocks_4_adaLN_modulation_1_weight_to_fp16, x = input_83_cast_fp16)[name = tensor("linear_44_cast_fp16")]; + tensor var_691_split_sizes_0 = const()[name = tensor("op_691_split_sizes_0"), val = tensor([512, 512, 512])]; + tensor var_691_axis_0 = const()[name = tensor("op_691_axis_0"), val = tensor(-1)]; + tensor var_691_cast_fp16_0, tensor var_691_cast_fp16_1, tensor var_691_cast_fp16_2 = split(axis = var_691_axis_0, split_sizes = var_691_split_sizes_0, x = linear_44_cast_fp16)[name = tensor("op_691_cast_fp16")]; + tensor mean_23_axes_0 = const()[name = tensor("mean_23_axes_0"), val = tensor([-1])]; + tensor mean_23_keep_dims_0 = const()[name = tensor("mean_23_keep_dims_0"), val = tensor(true)]; + tensor mean_23_cast_fp16 = reduce_mean(axes = mean_23_axes_0, keep_dims = mean_23_keep_dims_0, x = x_73_cast_fp16)[name = tensor("mean_23_cast_fp16")]; + tensor sub_19_cast_fp16 = sub(x = x_73_cast_fp16, y = mean_23_cast_fp16)[name = tensor("sub_19_cast_fp16")]; + tensor square_15_cast_fp16 = square(x = sub_19_cast_fp16)[name = tensor("square_15_cast_fp16")]; + tensor reduce_mean_31_axes_0 = const()[name = tensor("reduce_mean_31_axes_0"), val = tensor([-1])]; + tensor reduce_mean_31_keep_dims_0 = const()[name = tensor("reduce_mean_31_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_31_cast_fp16 = reduce_mean(axes = reduce_mean_31_axes_0, keep_dims = reduce_mean_31_keep_dims_0, x = square_15_cast_fp16)[name = tensor("reduce_mean_31_cast_fp16")]; + tensor var_701_to_fp16 = const()[name = tensor("op_701_to_fp16"), val = tensor(0x1.1p-20)]; + tensor var_702_cast_fp16 = add(x = reduce_mean_31_cast_fp16, y = var_701_to_fp16)[name = tensor("op_702_cast_fp16")]; + tensor var_703_cast_fp16 = sqrt(x = var_702_cast_fp16)[name = tensor("op_703_cast_fp16")]; + tensor x_75_cast_fp16 = real_div(x = sub_19_cast_fp16, y = var_703_cast_fp16)[name = tensor("x_75_cast_fp16")]; + tensor var_705_cast_fp16 = mul(x = x_75_cast_fp16, y = flow_net_res_blocks_4_in_ln_weight_to_fp16)[name = tensor("op_705_cast_fp16")]; + tensor x_77_cast_fp16 = add(x = var_705_cast_fp16, y = flow_net_res_blocks_4_in_ln_bias_to_fp16)[name = tensor("x_77_cast_fp16")]; + tensor var_707_promoted_to_fp16 = const()[name = tensor("op_707_promoted_to_fp16"), val = tensor(0x1p+0)]; + tensor var_708_cast_fp16 = add(x = var_691_cast_fp16_1, y = var_707_promoted_to_fp16)[name = tensor("op_708_cast_fp16")]; + tensor var_709_cast_fp16 = mul(x = x_77_cast_fp16, y = var_708_cast_fp16)[name = tensor("op_709_cast_fp16")]; + tensor input_117_cast_fp16 = add(x = var_709_cast_fp16, y = var_691_cast_fp16_0)[name = tensor("input_117_cast_fp16")]; + tensor linear_45_cast_fp16 = linear(bias = flow_net_res_blocks_4_mlp_0_bias_to_fp16, weight = flow_net_res_blocks_4_mlp_0_weight_to_fp16, x = input_117_cast_fp16)[name = tensor("linear_45_cast_fp16")]; + tensor input_121_cast_fp16 = silu(x = linear_45_cast_fp16)[name = tensor("input_121_cast_fp16")]; + tensor linear_46_cast_fp16 = linear(bias = flow_net_res_blocks_4_mlp_2_bias_to_fp16, weight = flow_net_res_blocks_4_mlp_2_weight_to_fp16, x = input_121_cast_fp16)[name = tensor("linear_46_cast_fp16")]; + tensor var_720_cast_fp16 = mul(x = var_691_cast_fp16_2, y = linear_46_cast_fp16)[name = tensor("op_720_cast_fp16")]; + tensor x_79_cast_fp16 = add(x = x_73_cast_fp16, y = var_720_cast_fp16)[name = tensor("x_79_cast_fp16")]; + tensor linear_47_cast_fp16 = linear(bias = flow_net_res_blocks_5_adaLN_modulation_1_bias_to_fp16, weight = flow_net_res_blocks_5_adaLN_modulation_1_weight_to_fp16, x = input_83_cast_fp16)[name = tensor("linear_47_cast_fp16")]; + tensor var_730_split_sizes_0 = const()[name = tensor("op_730_split_sizes_0"), val = tensor([512, 512, 512])]; + tensor var_730_axis_0 = const()[name = tensor("op_730_axis_0"), val = tensor(-1)]; + tensor var_730_cast_fp16_0, tensor var_730_cast_fp16_1, tensor var_730_cast_fp16_2 = split(axis = var_730_axis_0, split_sizes = var_730_split_sizes_0, x = linear_47_cast_fp16)[name = tensor("op_730_cast_fp16")]; + tensor mean_25_axes_0 = const()[name = tensor("mean_25_axes_0"), val = tensor([-1])]; + tensor mean_25_keep_dims_0 = const()[name = tensor("mean_25_keep_dims_0"), val = tensor(true)]; + tensor mean_25_cast_fp16 = reduce_mean(axes = mean_25_axes_0, keep_dims = mean_25_keep_dims_0, x = x_79_cast_fp16)[name = tensor("mean_25_cast_fp16")]; + tensor sub_20_cast_fp16 = sub(x = x_79_cast_fp16, y = mean_25_cast_fp16)[name = tensor("sub_20_cast_fp16")]; + tensor square_16_cast_fp16 = square(x = sub_20_cast_fp16)[name = tensor("square_16_cast_fp16")]; + tensor reduce_mean_33_axes_0 = const()[name = tensor("reduce_mean_33_axes_0"), val = tensor([-1])]; + tensor reduce_mean_33_keep_dims_0 = const()[name = tensor("reduce_mean_33_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_33_cast_fp16 = reduce_mean(axes = reduce_mean_33_axes_0, keep_dims = reduce_mean_33_keep_dims_0, x = square_16_cast_fp16)[name = tensor("reduce_mean_33_cast_fp16")]; + tensor var_740_to_fp16 = const()[name = tensor("op_740_to_fp16"), val = tensor(0x1.1p-20)]; + tensor var_741_cast_fp16 = add(x = reduce_mean_33_cast_fp16, y = var_740_to_fp16)[name = tensor("op_741_cast_fp16")]; + tensor var_742_cast_fp16 = sqrt(x = var_741_cast_fp16)[name = tensor("op_742_cast_fp16")]; + tensor x_81_cast_fp16 = real_div(x = sub_20_cast_fp16, y = var_742_cast_fp16)[name = tensor("x_81_cast_fp16")]; + tensor var_744_cast_fp16 = mul(x = x_81_cast_fp16, y = flow_net_res_blocks_5_in_ln_weight_to_fp16)[name = tensor("op_744_cast_fp16")]; + tensor x_83_cast_fp16 = add(x = var_744_cast_fp16, y = flow_net_res_blocks_5_in_ln_bias_to_fp16)[name = tensor("x_83_cast_fp16")]; + tensor var_746_promoted_to_fp16 = const()[name = tensor("op_746_promoted_to_fp16"), val = tensor(0x1p+0)]; + tensor var_747_cast_fp16 = add(x = var_730_cast_fp16_1, y = var_746_promoted_to_fp16)[name = tensor("op_747_cast_fp16")]; + tensor var_748_cast_fp16 = mul(x = x_83_cast_fp16, y = var_747_cast_fp16)[name = tensor("op_748_cast_fp16")]; + tensor input_125_cast_fp16 = add(x = var_748_cast_fp16, y = var_730_cast_fp16_0)[name = tensor("input_125_cast_fp16")]; + tensor linear_48_cast_fp16 = linear(bias = flow_net_res_blocks_5_mlp_0_bias_to_fp16, weight = flow_net_res_blocks_5_mlp_0_weight_to_fp16, x = input_125_cast_fp16)[name = tensor("linear_48_cast_fp16")]; + tensor input_129_cast_fp16 = silu(x = linear_48_cast_fp16)[name = tensor("input_129_cast_fp16")]; + tensor linear_49_cast_fp16 = linear(bias = flow_net_res_blocks_5_mlp_2_bias_to_fp16, weight = flow_net_res_blocks_5_mlp_2_weight_to_fp16, x = input_129_cast_fp16)[name = tensor("linear_49_cast_fp16")]; + tensor var_759_cast_fp16 = mul(x = var_730_cast_fp16_2, y = linear_49_cast_fp16)[name = tensor("op_759_cast_fp16")]; + tensor x_85_cast_fp16 = add(x = x_79_cast_fp16, y = var_759_cast_fp16)[name = tensor("x_85_cast_fp16")]; + tensor linear_50_cast_fp16 = linear(bias = flow_net_final_layer_adaLN_modulation_1_bias_to_fp16, weight = flow_net_final_layer_adaLN_modulation_1_weight_to_fp16, x = input_83_cast_fp16)[name = tensor("linear_50_cast_fp16")]; + tensor var_768_split_sizes_0 = const()[name = tensor("op_768_split_sizes_0"), val = tensor([512, 512])]; + tensor var_768_axis_0 = const()[name = tensor("op_768_axis_0"), val = tensor(-1)]; + tensor var_768_cast_fp16_0, tensor var_768_cast_fp16_1 = split(axis = var_768_axis_0, split_sizes = var_768_split_sizes_0, x = linear_50_cast_fp16)[name = tensor("op_768_cast_fp16")]; + tensor mean_27_axes_0 = const()[name = tensor("mean_27_axes_0"), val = tensor([-1])]; + tensor mean_27_keep_dims_0 = const()[name = tensor("mean_27_keep_dims_0"), val = tensor(true)]; + tensor mean_27_cast_fp16 = reduce_mean(axes = mean_27_axes_0, keep_dims = mean_27_keep_dims_0, x = x_85_cast_fp16)[name = tensor("mean_27_cast_fp16")]; + tensor sub_21_cast_fp16 = sub(x = x_85_cast_fp16, y = mean_27_cast_fp16)[name = tensor("sub_21_cast_fp16")]; + tensor square_17_cast_fp16 = square(x = sub_21_cast_fp16)[name = tensor("square_17_cast_fp16")]; + tensor reduce_mean_35_axes_0 = const()[name = tensor("reduce_mean_35_axes_0"), val = tensor([-1])]; + tensor reduce_mean_35_keep_dims_0 = const()[name = tensor("reduce_mean_35_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_35_cast_fp16 = reduce_mean(axes = reduce_mean_35_axes_0, keep_dims = reduce_mean_35_keep_dims_0, x = square_17_cast_fp16)[name = tensor("reduce_mean_35_cast_fp16")]; + tensor var_775_to_fp16 = const()[name = tensor("op_775_to_fp16"), val = tensor(0x1.1p-20)]; + tensor var_776_cast_fp16 = add(x = reduce_mean_35_cast_fp16, y = var_775_to_fp16)[name = tensor("op_776_cast_fp16")]; + tensor var_777_cast_fp16 = sqrt(x = var_776_cast_fp16)[name = tensor("op_777_cast_fp16")]; + tensor x_87_cast_fp16 = real_div(x = sub_21_cast_fp16, y = var_777_cast_fp16)[name = tensor("x_87_cast_fp16")]; + tensor var_779_promoted_to_fp16 = const()[name = tensor("op_779_promoted_to_fp16"), val = tensor(0x1p+0)]; + tensor var_780_cast_fp16 = add(x = var_768_cast_fp16_1, y = var_779_promoted_to_fp16)[name = tensor("op_780_cast_fp16")]; + tensor var_781_cast_fp16 = mul(x = x_87_cast_fp16, y = var_780_cast_fp16)[name = tensor("op_781_cast_fp16")]; + tensor input_133_cast_fp16 = add(x = var_781_cast_fp16, y = var_768_cast_fp16_0)[name = tensor("input_133_cast_fp16")]; + tensor linear_51_cast_fp16 = linear(bias = flow_net_final_layer_linear_bias_to_fp16, weight = flow_net_final_layer_linear_weight_to_fp16, x = input_133_cast_fp16)[name = tensor("linear_51_cast_fp16")]; + tensor var_792_to_fp16 = const()[name = tensor("op_792_to_fp16"), val = tensor(0x1p-3)]; + tensor var_793_cast_fp16 = mul(x = linear_51_cast_fp16, y = var_792_to_fp16)[name = tensor("op_793_cast_fp16")]; + tensor input_135_cast_fp16 = add(x = input_67_cast_fp16, y = var_793_cast_fp16)[name = tensor("input_135_cast_fp16")]; + tensor linear_52_cast_fp16 = linear(bias = flow_net_input_proj_bias_to_fp16, weight = flow_net_input_proj_weight_to_fp16, x = input_135_cast_fp16)[name = tensor("linear_52_cast_fp16")]; + tensor input_139_to_fp16 = const()[name = tensor("input_139_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(18999552)))]; + tensor input_141_cast_fp16 = silu(x = input_139_to_fp16)[name = tensor("input_141_cast_fp16")]; + tensor linear_54_cast_fp16 = linear(bias = flow_net_time_embed_0_mlp_2_bias_to_fp16, weight = flow_net_time_embed_0_mlp_2_weight_to_fp16, x = input_141_cast_fp16)[name = tensor("linear_54_cast_fp16")]; + tensor reduce_mean_36_axes_0 = const()[name = tensor("reduce_mean_36_axes_0"), val = tensor([-1])]; + tensor reduce_mean_36_keep_dims_0 = const()[name = tensor("reduce_mean_36_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_36_cast_fp16 = reduce_mean(axes = reduce_mean_36_axes_0, keep_dims = reduce_mean_36_keep_dims_0, x = linear_54_cast_fp16)[name = tensor("reduce_mean_36_cast_fp16")]; + tensor sub_22_cast_fp16 = sub(x = linear_54_cast_fp16, y = reduce_mean_36_cast_fp16)[name = tensor("sub_22_cast_fp16")]; + tensor square_18_cast_fp16 = square(x = sub_22_cast_fp16)[name = tensor("square_18_cast_fp16")]; + tensor reduce_mean_37_axes_0 = const()[name = tensor("reduce_mean_37_axes_0"), val = tensor([-1])]; + tensor reduce_mean_37_keep_dims_0 = const()[name = tensor("reduce_mean_37_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_37_cast_fp16 = reduce_mean(axes = reduce_mean_37_axes_0, keep_dims = reduce_mean_37_keep_dims_0, x = square_18_cast_fp16)[name = tensor("reduce_mean_37_cast_fp16")]; + tensor real_div_4_to_fp16 = const()[name = tensor("real_div_4_to_fp16"), val = tensor(0x1.008p+0)]; + tensor mul_4_cast_fp16 = mul(x = reduce_mean_37_cast_fp16, y = real_div_4_to_fp16)[name = tensor("mul_4_cast_fp16")]; + tensor var_859_to_fp16 = const()[name = tensor("op_859_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_37_cast_fp16 = add(x = mul_4_cast_fp16, y = var_859_to_fp16)[name = tensor("var_37_cast_fp16")]; + tensor var_862_epsilon_0 = const()[name = tensor("op_862_epsilon_0"), val = tensor(0x1.197998p-40)]; + tensor var_862_cast_fp16 = rsqrt(epsilon = var_862_epsilon_0, x = var_37_cast_fp16)[name = tensor("op_862_cast_fp16")]; + tensor var_863_cast_fp16 = mul(x = const_3_to_fp16, y = var_862_cast_fp16)[name = tensor("op_863_cast_fp16")]; + tensor var_864_cast_fp16 = mul(x = linear_54_cast_fp16, y = var_863_cast_fp16)[name = tensor("op_864_cast_fp16")]; + tensor input_145_to_fp16 = const()[name = tensor("input_145_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(19000640)))]; + tensor input_147_cast_fp16 = silu(x = input_145_to_fp16)[name = tensor("input_147_cast_fp16")]; + tensor linear_56_cast_fp16 = linear(bias = flow_net_time_embed_1_mlp_2_bias_to_fp16, weight = flow_net_time_embed_1_mlp_2_weight_to_fp16, x = input_147_cast_fp16)[name = tensor("linear_56_cast_fp16")]; + tensor reduce_mean_38_axes_0 = const()[name = tensor("reduce_mean_38_axes_0"), val = tensor([-1])]; + tensor reduce_mean_38_keep_dims_0 = const()[name = tensor("reduce_mean_38_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_38_cast_fp16 = reduce_mean(axes = reduce_mean_38_axes_0, keep_dims = reduce_mean_38_keep_dims_0, x = linear_56_cast_fp16)[name = tensor("reduce_mean_38_cast_fp16")]; + tensor sub_24_cast_fp16 = sub(x = linear_56_cast_fp16, y = reduce_mean_38_cast_fp16)[name = tensor("sub_24_cast_fp16")]; + tensor square_19_cast_fp16 = square(x = sub_24_cast_fp16)[name = tensor("square_19_cast_fp16")]; + tensor reduce_mean_39_axes_0 = const()[name = tensor("reduce_mean_39_axes_0"), val = tensor([-1])]; + tensor reduce_mean_39_keep_dims_0 = const()[name = tensor("reduce_mean_39_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_39_cast_fp16 = reduce_mean(axes = reduce_mean_39_axes_0, keep_dims = reduce_mean_39_keep_dims_0, x = square_19_cast_fp16)[name = tensor("reduce_mean_39_cast_fp16")]; + tensor real_div_5_to_fp16 = const()[name = tensor("real_div_5_to_fp16"), val = tensor(0x1.008p+0)]; + tensor mul_5_cast_fp16 = mul(x = reduce_mean_39_cast_fp16, y = real_div_5_to_fp16)[name = tensor("mul_5_cast_fp16")]; + tensor var_896_to_fp16 = const()[name = tensor("op_896_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_39_cast_fp16 = add(x = mul_5_cast_fp16, y = var_896_to_fp16)[name = tensor("var_39_cast_fp16")]; + tensor var_899_epsilon_0 = const()[name = tensor("op_899_epsilon_0"), val = tensor(0x1.197998p-40)]; + tensor var_899_cast_fp16 = rsqrt(epsilon = var_899_epsilon_0, x = var_39_cast_fp16)[name = tensor("op_899_cast_fp16")]; + tensor var_900_cast_fp16 = mul(x = const_5_to_fp16, y = var_899_cast_fp16)[name = tensor("op_900_cast_fp16")]; + tensor var_901_cast_fp16 = mul(x = linear_56_cast_fp16, y = var_900_cast_fp16)[name = tensor("op_901_cast_fp16")]; + tensor var_913_cast_fp16 = add(x = var_864_cast_fp16, y = var_901_cast_fp16)[name = tensor("op_913_cast_fp16")]; + tensor _inversed_t_combined_5_y_0_to_fp16 = const()[name = tensor("_inversed_t_combined_5_y_0_to_fp16"), val = tensor(0x1p-1)]; + tensor _inversed_t_combined_5_cast_fp16 = mul(x = var_913_cast_fp16, y = _inversed_t_combined_5_y_0_to_fp16)[name = tensor("_inversed_t_combined_5_cast_fp16")]; + tensor input_149_cast_fp16 = add(x = _inversed_t_combined_5_cast_fp16, y = linear_5_cast_fp16)[name = tensor("input_149_cast_fp16")]; + tensor input_151_cast_fp16 = silu(x = input_149_cast_fp16)[name = tensor("input_151_cast_fp16")]; + tensor linear_58_cast_fp16 = linear(bias = flow_net_res_blocks_0_adaLN_modulation_1_bias_to_fp16, weight = flow_net_res_blocks_0_adaLN_modulation_1_weight_to_fp16, x = input_151_cast_fp16)[name = tensor("linear_58_cast_fp16")]; + tensor var_928_split_sizes_0 = const()[name = tensor("op_928_split_sizes_0"), val = tensor([512, 512, 512])]; + tensor var_928_axis_0 = const()[name = tensor("op_928_axis_0"), val = tensor(-1)]; + tensor var_928_cast_fp16_0, tensor var_928_cast_fp16_1, tensor var_928_cast_fp16_2 = split(axis = var_928_axis_0, split_sizes = var_928_split_sizes_0, x = linear_58_cast_fp16)[name = tensor("op_928_cast_fp16")]; + tensor mean_29_axes_0 = const()[name = tensor("mean_29_axes_0"), val = tensor([-1])]; + tensor mean_29_keep_dims_0 = const()[name = tensor("mean_29_keep_dims_0"), val = tensor(true)]; + tensor mean_29_cast_fp16 = reduce_mean(axes = mean_29_axes_0, keep_dims = mean_29_keep_dims_0, x = linear_52_cast_fp16)[name = tensor("mean_29_cast_fp16")]; + tensor sub_26_cast_fp16 = sub(x = linear_52_cast_fp16, y = mean_29_cast_fp16)[name = tensor("sub_26_cast_fp16")]; + tensor square_20_cast_fp16 = square(x = sub_26_cast_fp16)[name = tensor("square_20_cast_fp16")]; + tensor reduce_mean_41_axes_0 = const()[name = tensor("reduce_mean_41_axes_0"), val = tensor([-1])]; + tensor reduce_mean_41_keep_dims_0 = const()[name = tensor("reduce_mean_41_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_41_cast_fp16 = reduce_mean(axes = reduce_mean_41_axes_0, keep_dims = reduce_mean_41_keep_dims_0, x = square_20_cast_fp16)[name = tensor("reduce_mean_41_cast_fp16")]; + tensor var_938_to_fp16 = const()[name = tensor("op_938_to_fp16"), val = tensor(0x1.1p-20)]; + tensor var_939_cast_fp16 = add(x = reduce_mean_41_cast_fp16, y = var_938_to_fp16)[name = tensor("op_939_cast_fp16")]; + tensor var_940_cast_fp16 = sqrt(x = var_939_cast_fp16)[name = tensor("op_940_cast_fp16")]; + tensor x_95_cast_fp16 = real_div(x = sub_26_cast_fp16, y = var_940_cast_fp16)[name = tensor("x_95_cast_fp16")]; + tensor var_942_cast_fp16 = mul(x = x_95_cast_fp16, y = flow_net_res_blocks_0_in_ln_weight_to_fp16)[name = tensor("op_942_cast_fp16")]; + tensor x_97_cast_fp16 = add(x = var_942_cast_fp16, y = flow_net_res_blocks_0_in_ln_bias_to_fp16)[name = tensor("x_97_cast_fp16")]; + tensor var_944_promoted_to_fp16 = const()[name = tensor("op_944_promoted_to_fp16"), val = tensor(0x1p+0)]; + tensor var_945_cast_fp16 = add(x = var_928_cast_fp16_1, y = var_944_promoted_to_fp16)[name = tensor("op_945_cast_fp16")]; + tensor var_946_cast_fp16 = mul(x = x_97_cast_fp16, y = var_945_cast_fp16)[name = tensor("op_946_cast_fp16")]; + tensor input_153_cast_fp16 = add(x = var_946_cast_fp16, y = var_928_cast_fp16_0)[name = tensor("input_153_cast_fp16")]; + tensor linear_59_cast_fp16 = linear(bias = flow_net_res_blocks_0_mlp_0_bias_to_fp16, weight = flow_net_res_blocks_0_mlp_0_weight_to_fp16, x = input_153_cast_fp16)[name = tensor("linear_59_cast_fp16")]; + tensor input_157_cast_fp16 = silu(x = linear_59_cast_fp16)[name = tensor("input_157_cast_fp16")]; + tensor linear_60_cast_fp16 = linear(bias = flow_net_res_blocks_0_mlp_2_bias_to_fp16, weight = flow_net_res_blocks_0_mlp_2_weight_to_fp16, x = input_157_cast_fp16)[name = tensor("linear_60_cast_fp16")]; + tensor var_957_cast_fp16 = mul(x = var_928_cast_fp16_2, y = linear_60_cast_fp16)[name = tensor("op_957_cast_fp16")]; + tensor x_99_cast_fp16 = add(x = linear_52_cast_fp16, y = var_957_cast_fp16)[name = tensor("x_99_cast_fp16")]; + tensor linear_61_cast_fp16 = linear(bias = flow_net_res_blocks_1_adaLN_modulation_1_bias_to_fp16, weight = flow_net_res_blocks_1_adaLN_modulation_1_weight_to_fp16, x = input_151_cast_fp16)[name = tensor("linear_61_cast_fp16")]; + tensor var_967_split_sizes_0 = const()[name = tensor("op_967_split_sizes_0"), val = tensor([512, 512, 512])]; + tensor var_967_axis_0 = const()[name = tensor("op_967_axis_0"), val = tensor(-1)]; + tensor var_967_cast_fp16_0, tensor var_967_cast_fp16_1, tensor var_967_cast_fp16_2 = split(axis = var_967_axis_0, split_sizes = var_967_split_sizes_0, x = linear_61_cast_fp16)[name = tensor("op_967_cast_fp16")]; + tensor mean_31_axes_0 = const()[name = tensor("mean_31_axes_0"), val = tensor([-1])]; + tensor mean_31_keep_dims_0 = const()[name = tensor("mean_31_keep_dims_0"), val = tensor(true)]; + tensor mean_31_cast_fp16 = reduce_mean(axes = mean_31_axes_0, keep_dims = mean_31_keep_dims_0, x = x_99_cast_fp16)[name = tensor("mean_31_cast_fp16")]; + tensor sub_27_cast_fp16 = sub(x = x_99_cast_fp16, y = mean_31_cast_fp16)[name = tensor("sub_27_cast_fp16")]; + tensor square_21_cast_fp16 = square(x = sub_27_cast_fp16)[name = tensor("square_21_cast_fp16")]; + tensor reduce_mean_43_axes_0 = const()[name = tensor("reduce_mean_43_axes_0"), val = tensor([-1])]; + tensor reduce_mean_43_keep_dims_0 = const()[name = tensor("reduce_mean_43_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_43_cast_fp16 = reduce_mean(axes = reduce_mean_43_axes_0, keep_dims = reduce_mean_43_keep_dims_0, x = square_21_cast_fp16)[name = tensor("reduce_mean_43_cast_fp16")]; + tensor var_977_to_fp16 = const()[name = tensor("op_977_to_fp16"), val = tensor(0x1.1p-20)]; + tensor var_978_cast_fp16 = add(x = reduce_mean_43_cast_fp16, y = var_977_to_fp16)[name = tensor("op_978_cast_fp16")]; + tensor var_979_cast_fp16 = sqrt(x = var_978_cast_fp16)[name = tensor("op_979_cast_fp16")]; + tensor x_101_cast_fp16 = real_div(x = sub_27_cast_fp16, y = var_979_cast_fp16)[name = tensor("x_101_cast_fp16")]; + tensor var_981_cast_fp16 = mul(x = x_101_cast_fp16, y = flow_net_res_blocks_1_in_ln_weight_to_fp16)[name = tensor("op_981_cast_fp16")]; + tensor x_103_cast_fp16 = add(x = var_981_cast_fp16, y = flow_net_res_blocks_1_in_ln_bias_to_fp16)[name = tensor("x_103_cast_fp16")]; + tensor var_983_promoted_to_fp16 = const()[name = tensor("op_983_promoted_to_fp16"), val = tensor(0x1p+0)]; + tensor var_984_cast_fp16 = add(x = var_967_cast_fp16_1, y = var_983_promoted_to_fp16)[name = tensor("op_984_cast_fp16")]; + tensor var_985_cast_fp16 = mul(x = x_103_cast_fp16, y = var_984_cast_fp16)[name = tensor("op_985_cast_fp16")]; + tensor input_161_cast_fp16 = add(x = var_985_cast_fp16, y = var_967_cast_fp16_0)[name = tensor("input_161_cast_fp16")]; + tensor linear_62_cast_fp16 = linear(bias = flow_net_res_blocks_1_mlp_0_bias_to_fp16, weight = flow_net_res_blocks_1_mlp_0_weight_to_fp16, x = input_161_cast_fp16)[name = tensor("linear_62_cast_fp16")]; + tensor input_165_cast_fp16 = silu(x = linear_62_cast_fp16)[name = tensor("input_165_cast_fp16")]; + tensor linear_63_cast_fp16 = linear(bias = flow_net_res_blocks_1_mlp_2_bias_to_fp16, weight = flow_net_res_blocks_1_mlp_2_weight_to_fp16, x = input_165_cast_fp16)[name = tensor("linear_63_cast_fp16")]; + tensor var_996_cast_fp16 = mul(x = var_967_cast_fp16_2, y = linear_63_cast_fp16)[name = tensor("op_996_cast_fp16")]; + tensor x_105_cast_fp16 = add(x = x_99_cast_fp16, y = var_996_cast_fp16)[name = tensor("x_105_cast_fp16")]; + tensor linear_64_cast_fp16 = linear(bias = flow_net_res_blocks_2_adaLN_modulation_1_bias_to_fp16, weight = flow_net_res_blocks_2_adaLN_modulation_1_weight_to_fp16, x = input_151_cast_fp16)[name = tensor("linear_64_cast_fp16")]; + tensor var_1006_split_sizes_0 = const()[name = tensor("op_1006_split_sizes_0"), val = tensor([512, 512, 512])]; + tensor var_1006_axis_0 = const()[name = tensor("op_1006_axis_0"), val = tensor(-1)]; + tensor var_1006_cast_fp16_0, tensor var_1006_cast_fp16_1, tensor var_1006_cast_fp16_2 = split(axis = var_1006_axis_0, split_sizes = var_1006_split_sizes_0, x = linear_64_cast_fp16)[name = tensor("op_1006_cast_fp16")]; + tensor mean_33_axes_0 = const()[name = tensor("mean_33_axes_0"), val = tensor([-1])]; + tensor mean_33_keep_dims_0 = const()[name = tensor("mean_33_keep_dims_0"), val = tensor(true)]; + tensor mean_33_cast_fp16 = reduce_mean(axes = mean_33_axes_0, keep_dims = mean_33_keep_dims_0, x = x_105_cast_fp16)[name = tensor("mean_33_cast_fp16")]; + tensor sub_28_cast_fp16 = sub(x = x_105_cast_fp16, y = mean_33_cast_fp16)[name = tensor("sub_28_cast_fp16")]; + tensor square_22_cast_fp16 = square(x = sub_28_cast_fp16)[name = tensor("square_22_cast_fp16")]; + tensor reduce_mean_45_axes_0 = const()[name = tensor("reduce_mean_45_axes_0"), val = tensor([-1])]; + tensor reduce_mean_45_keep_dims_0 = const()[name = tensor("reduce_mean_45_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_45_cast_fp16 = reduce_mean(axes = reduce_mean_45_axes_0, keep_dims = reduce_mean_45_keep_dims_0, x = square_22_cast_fp16)[name = tensor("reduce_mean_45_cast_fp16")]; + tensor var_1016_to_fp16 = const()[name = tensor("op_1016_to_fp16"), val = tensor(0x1.1p-20)]; + tensor var_1017_cast_fp16 = add(x = reduce_mean_45_cast_fp16, y = var_1016_to_fp16)[name = tensor("op_1017_cast_fp16")]; + tensor var_1018_cast_fp16 = sqrt(x = var_1017_cast_fp16)[name = tensor("op_1018_cast_fp16")]; + tensor x_107_cast_fp16 = real_div(x = sub_28_cast_fp16, y = var_1018_cast_fp16)[name = tensor("x_107_cast_fp16")]; + tensor var_1020_cast_fp16 = mul(x = x_107_cast_fp16, y = flow_net_res_blocks_2_in_ln_weight_to_fp16)[name = tensor("op_1020_cast_fp16")]; + tensor x_109_cast_fp16 = add(x = var_1020_cast_fp16, y = flow_net_res_blocks_2_in_ln_bias_to_fp16)[name = tensor("x_109_cast_fp16")]; + tensor var_1022_promoted_to_fp16 = const()[name = tensor("op_1022_promoted_to_fp16"), val = tensor(0x1p+0)]; + tensor var_1023_cast_fp16 = add(x = var_1006_cast_fp16_1, y = var_1022_promoted_to_fp16)[name = tensor("op_1023_cast_fp16")]; + tensor var_1024_cast_fp16 = mul(x = x_109_cast_fp16, y = var_1023_cast_fp16)[name = tensor("op_1024_cast_fp16")]; + tensor input_169_cast_fp16 = add(x = var_1024_cast_fp16, y = var_1006_cast_fp16_0)[name = tensor("input_169_cast_fp16")]; + tensor linear_65_cast_fp16 = linear(bias = flow_net_res_blocks_2_mlp_0_bias_to_fp16, weight = flow_net_res_blocks_2_mlp_0_weight_to_fp16, x = input_169_cast_fp16)[name = tensor("linear_65_cast_fp16")]; + tensor input_173_cast_fp16 = silu(x = linear_65_cast_fp16)[name = tensor("input_173_cast_fp16")]; + tensor linear_66_cast_fp16 = linear(bias = flow_net_res_blocks_2_mlp_2_bias_to_fp16, weight = flow_net_res_blocks_2_mlp_2_weight_to_fp16, x = input_173_cast_fp16)[name = tensor("linear_66_cast_fp16")]; + tensor var_1035_cast_fp16 = mul(x = var_1006_cast_fp16_2, y = linear_66_cast_fp16)[name = tensor("op_1035_cast_fp16")]; + tensor x_111_cast_fp16 = add(x = x_105_cast_fp16, y = var_1035_cast_fp16)[name = tensor("x_111_cast_fp16")]; + tensor linear_67_cast_fp16 = linear(bias = flow_net_res_blocks_3_adaLN_modulation_1_bias_to_fp16, weight = flow_net_res_blocks_3_adaLN_modulation_1_weight_to_fp16, x = input_151_cast_fp16)[name = tensor("linear_67_cast_fp16")]; + tensor var_1045_split_sizes_0 = const()[name = tensor("op_1045_split_sizes_0"), val = tensor([512, 512, 512])]; + tensor var_1045_axis_0 = const()[name = tensor("op_1045_axis_0"), val = tensor(-1)]; + tensor var_1045_cast_fp16_0, tensor var_1045_cast_fp16_1, tensor var_1045_cast_fp16_2 = split(axis = var_1045_axis_0, split_sizes = var_1045_split_sizes_0, x = linear_67_cast_fp16)[name = tensor("op_1045_cast_fp16")]; + tensor mean_35_axes_0 = const()[name = tensor("mean_35_axes_0"), val = tensor([-1])]; + tensor mean_35_keep_dims_0 = const()[name = tensor("mean_35_keep_dims_0"), val = tensor(true)]; + tensor mean_35_cast_fp16 = reduce_mean(axes = mean_35_axes_0, keep_dims = mean_35_keep_dims_0, x = x_111_cast_fp16)[name = tensor("mean_35_cast_fp16")]; + tensor sub_29_cast_fp16 = sub(x = x_111_cast_fp16, y = mean_35_cast_fp16)[name = tensor("sub_29_cast_fp16")]; + tensor square_23_cast_fp16 = square(x = sub_29_cast_fp16)[name = tensor("square_23_cast_fp16")]; + tensor reduce_mean_47_axes_0 = const()[name = tensor("reduce_mean_47_axes_0"), val = tensor([-1])]; + tensor reduce_mean_47_keep_dims_0 = const()[name = tensor("reduce_mean_47_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_47_cast_fp16 = reduce_mean(axes = reduce_mean_47_axes_0, keep_dims = reduce_mean_47_keep_dims_0, x = square_23_cast_fp16)[name = tensor("reduce_mean_47_cast_fp16")]; + tensor var_1055_to_fp16 = const()[name = tensor("op_1055_to_fp16"), val = tensor(0x1.1p-20)]; + tensor var_1056_cast_fp16 = add(x = reduce_mean_47_cast_fp16, y = var_1055_to_fp16)[name = tensor("op_1056_cast_fp16")]; + tensor var_1057_cast_fp16 = sqrt(x = var_1056_cast_fp16)[name = tensor("op_1057_cast_fp16")]; + tensor x_113_cast_fp16 = real_div(x = sub_29_cast_fp16, y = var_1057_cast_fp16)[name = tensor("x_113_cast_fp16")]; + tensor var_1059_cast_fp16 = mul(x = x_113_cast_fp16, y = flow_net_res_blocks_3_in_ln_weight_to_fp16)[name = tensor("op_1059_cast_fp16")]; + tensor x_115_cast_fp16 = add(x = var_1059_cast_fp16, y = flow_net_res_blocks_3_in_ln_bias_to_fp16)[name = tensor("x_115_cast_fp16")]; + tensor var_1061_promoted_to_fp16 = const()[name = tensor("op_1061_promoted_to_fp16"), val = tensor(0x1p+0)]; + tensor var_1062_cast_fp16 = add(x = var_1045_cast_fp16_1, y = var_1061_promoted_to_fp16)[name = tensor("op_1062_cast_fp16")]; + tensor var_1063_cast_fp16 = mul(x = x_115_cast_fp16, y = var_1062_cast_fp16)[name = tensor("op_1063_cast_fp16")]; + tensor input_177_cast_fp16 = add(x = var_1063_cast_fp16, y = var_1045_cast_fp16_0)[name = tensor("input_177_cast_fp16")]; + tensor linear_68_cast_fp16 = linear(bias = flow_net_res_blocks_3_mlp_0_bias_to_fp16, weight = flow_net_res_blocks_3_mlp_0_weight_to_fp16, x = input_177_cast_fp16)[name = tensor("linear_68_cast_fp16")]; + tensor input_181_cast_fp16 = silu(x = linear_68_cast_fp16)[name = tensor("input_181_cast_fp16")]; + tensor linear_69_cast_fp16 = linear(bias = flow_net_res_blocks_3_mlp_2_bias_to_fp16, weight = flow_net_res_blocks_3_mlp_2_weight_to_fp16, x = input_181_cast_fp16)[name = tensor("linear_69_cast_fp16")]; + tensor var_1074_cast_fp16 = mul(x = var_1045_cast_fp16_2, y = linear_69_cast_fp16)[name = tensor("op_1074_cast_fp16")]; + tensor x_117_cast_fp16 = add(x = x_111_cast_fp16, y = var_1074_cast_fp16)[name = tensor("x_117_cast_fp16")]; + tensor linear_70_cast_fp16 = linear(bias = flow_net_res_blocks_4_adaLN_modulation_1_bias_to_fp16, weight = flow_net_res_blocks_4_adaLN_modulation_1_weight_to_fp16, x = input_151_cast_fp16)[name = tensor("linear_70_cast_fp16")]; + tensor var_1084_split_sizes_0 = const()[name = tensor("op_1084_split_sizes_0"), val = tensor([512, 512, 512])]; + tensor var_1084_axis_0 = const()[name = tensor("op_1084_axis_0"), val = tensor(-1)]; + tensor var_1084_cast_fp16_0, tensor var_1084_cast_fp16_1, tensor var_1084_cast_fp16_2 = split(axis = var_1084_axis_0, split_sizes = var_1084_split_sizes_0, x = linear_70_cast_fp16)[name = tensor("op_1084_cast_fp16")]; + tensor mean_37_axes_0 = const()[name = tensor("mean_37_axes_0"), val = tensor([-1])]; + tensor mean_37_keep_dims_0 = const()[name = tensor("mean_37_keep_dims_0"), val = tensor(true)]; + tensor mean_37_cast_fp16 = reduce_mean(axes = mean_37_axes_0, keep_dims = mean_37_keep_dims_0, x = x_117_cast_fp16)[name = tensor("mean_37_cast_fp16")]; + tensor sub_30_cast_fp16 = sub(x = x_117_cast_fp16, y = mean_37_cast_fp16)[name = tensor("sub_30_cast_fp16")]; + tensor square_24_cast_fp16 = square(x = sub_30_cast_fp16)[name = tensor("square_24_cast_fp16")]; + tensor reduce_mean_49_axes_0 = const()[name = tensor("reduce_mean_49_axes_0"), val = tensor([-1])]; + tensor reduce_mean_49_keep_dims_0 = const()[name = tensor("reduce_mean_49_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_49_cast_fp16 = reduce_mean(axes = reduce_mean_49_axes_0, keep_dims = reduce_mean_49_keep_dims_0, x = square_24_cast_fp16)[name = tensor("reduce_mean_49_cast_fp16")]; + tensor var_1094_to_fp16 = const()[name = tensor("op_1094_to_fp16"), val = tensor(0x1.1p-20)]; + tensor var_1095_cast_fp16 = add(x = reduce_mean_49_cast_fp16, y = var_1094_to_fp16)[name = tensor("op_1095_cast_fp16")]; + tensor var_1096_cast_fp16 = sqrt(x = var_1095_cast_fp16)[name = tensor("op_1096_cast_fp16")]; + tensor x_119_cast_fp16 = real_div(x = sub_30_cast_fp16, y = var_1096_cast_fp16)[name = tensor("x_119_cast_fp16")]; + tensor var_1098_cast_fp16 = mul(x = x_119_cast_fp16, y = flow_net_res_blocks_4_in_ln_weight_to_fp16)[name = tensor("op_1098_cast_fp16")]; + tensor x_121_cast_fp16 = add(x = var_1098_cast_fp16, y = flow_net_res_blocks_4_in_ln_bias_to_fp16)[name = tensor("x_121_cast_fp16")]; + tensor var_1100_promoted_to_fp16 = const()[name = tensor("op_1100_promoted_to_fp16"), val = tensor(0x1p+0)]; + tensor var_1101_cast_fp16 = add(x = var_1084_cast_fp16_1, y = var_1100_promoted_to_fp16)[name = tensor("op_1101_cast_fp16")]; + tensor var_1102_cast_fp16 = mul(x = x_121_cast_fp16, y = var_1101_cast_fp16)[name = tensor("op_1102_cast_fp16")]; + tensor input_185_cast_fp16 = add(x = var_1102_cast_fp16, y = var_1084_cast_fp16_0)[name = tensor("input_185_cast_fp16")]; + tensor linear_71_cast_fp16 = linear(bias = flow_net_res_blocks_4_mlp_0_bias_to_fp16, weight = flow_net_res_blocks_4_mlp_0_weight_to_fp16, x = input_185_cast_fp16)[name = tensor("linear_71_cast_fp16")]; + tensor input_189_cast_fp16 = silu(x = linear_71_cast_fp16)[name = tensor("input_189_cast_fp16")]; + tensor linear_72_cast_fp16 = linear(bias = flow_net_res_blocks_4_mlp_2_bias_to_fp16, weight = flow_net_res_blocks_4_mlp_2_weight_to_fp16, x = input_189_cast_fp16)[name = tensor("linear_72_cast_fp16")]; + tensor var_1113_cast_fp16 = mul(x = var_1084_cast_fp16_2, y = linear_72_cast_fp16)[name = tensor("op_1113_cast_fp16")]; + tensor x_123_cast_fp16 = add(x = x_117_cast_fp16, y = var_1113_cast_fp16)[name = tensor("x_123_cast_fp16")]; + tensor linear_73_cast_fp16 = linear(bias = flow_net_res_blocks_5_adaLN_modulation_1_bias_to_fp16, weight = flow_net_res_blocks_5_adaLN_modulation_1_weight_to_fp16, x = input_151_cast_fp16)[name = tensor("linear_73_cast_fp16")]; + tensor var_1123_split_sizes_0 = const()[name = tensor("op_1123_split_sizes_0"), val = tensor([512, 512, 512])]; + tensor var_1123_axis_0 = const()[name = tensor("op_1123_axis_0"), val = tensor(-1)]; + tensor var_1123_cast_fp16_0, tensor var_1123_cast_fp16_1, tensor var_1123_cast_fp16_2 = split(axis = var_1123_axis_0, split_sizes = var_1123_split_sizes_0, x = linear_73_cast_fp16)[name = tensor("op_1123_cast_fp16")]; + tensor mean_39_axes_0 = const()[name = tensor("mean_39_axes_0"), val = tensor([-1])]; + tensor mean_39_keep_dims_0 = const()[name = tensor("mean_39_keep_dims_0"), val = tensor(true)]; + tensor mean_39_cast_fp16 = reduce_mean(axes = mean_39_axes_0, keep_dims = mean_39_keep_dims_0, x = x_123_cast_fp16)[name = tensor("mean_39_cast_fp16")]; + tensor sub_31_cast_fp16 = sub(x = x_123_cast_fp16, y = mean_39_cast_fp16)[name = tensor("sub_31_cast_fp16")]; + tensor square_25_cast_fp16 = square(x = sub_31_cast_fp16)[name = tensor("square_25_cast_fp16")]; + tensor reduce_mean_51_axes_0 = const()[name = tensor("reduce_mean_51_axes_0"), val = tensor([-1])]; + tensor reduce_mean_51_keep_dims_0 = const()[name = tensor("reduce_mean_51_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_51_cast_fp16 = reduce_mean(axes = reduce_mean_51_axes_0, keep_dims = reduce_mean_51_keep_dims_0, x = square_25_cast_fp16)[name = tensor("reduce_mean_51_cast_fp16")]; + tensor var_1133_to_fp16 = const()[name = tensor("op_1133_to_fp16"), val = tensor(0x1.1p-20)]; + tensor var_1134_cast_fp16 = add(x = reduce_mean_51_cast_fp16, y = var_1133_to_fp16)[name = tensor("op_1134_cast_fp16")]; + tensor var_1135_cast_fp16 = sqrt(x = var_1134_cast_fp16)[name = tensor("op_1135_cast_fp16")]; + tensor x_125_cast_fp16 = real_div(x = sub_31_cast_fp16, y = var_1135_cast_fp16)[name = tensor("x_125_cast_fp16")]; + tensor var_1137_cast_fp16 = mul(x = x_125_cast_fp16, y = flow_net_res_blocks_5_in_ln_weight_to_fp16)[name = tensor("op_1137_cast_fp16")]; + tensor x_127_cast_fp16 = add(x = var_1137_cast_fp16, y = flow_net_res_blocks_5_in_ln_bias_to_fp16)[name = tensor("x_127_cast_fp16")]; + tensor var_1139_promoted_to_fp16 = const()[name = tensor("op_1139_promoted_to_fp16"), val = tensor(0x1p+0)]; + tensor var_1140_cast_fp16 = add(x = var_1123_cast_fp16_1, y = var_1139_promoted_to_fp16)[name = tensor("op_1140_cast_fp16")]; + tensor var_1141_cast_fp16 = mul(x = x_127_cast_fp16, y = var_1140_cast_fp16)[name = tensor("op_1141_cast_fp16")]; + tensor input_193_cast_fp16 = add(x = var_1141_cast_fp16, y = var_1123_cast_fp16_0)[name = tensor("input_193_cast_fp16")]; + tensor linear_74_cast_fp16 = linear(bias = flow_net_res_blocks_5_mlp_0_bias_to_fp16, weight = flow_net_res_blocks_5_mlp_0_weight_to_fp16, x = input_193_cast_fp16)[name = tensor("linear_74_cast_fp16")]; + tensor input_197_cast_fp16 = silu(x = linear_74_cast_fp16)[name = tensor("input_197_cast_fp16")]; + tensor linear_75_cast_fp16 = linear(bias = flow_net_res_blocks_5_mlp_2_bias_to_fp16, weight = flow_net_res_blocks_5_mlp_2_weight_to_fp16, x = input_197_cast_fp16)[name = tensor("linear_75_cast_fp16")]; + tensor var_1152_cast_fp16 = mul(x = var_1123_cast_fp16_2, y = linear_75_cast_fp16)[name = tensor("op_1152_cast_fp16")]; + tensor x_129_cast_fp16 = add(x = x_123_cast_fp16, y = var_1152_cast_fp16)[name = tensor("x_129_cast_fp16")]; + tensor linear_76_cast_fp16 = linear(bias = flow_net_final_layer_adaLN_modulation_1_bias_to_fp16, weight = flow_net_final_layer_adaLN_modulation_1_weight_to_fp16, x = input_151_cast_fp16)[name = tensor("linear_76_cast_fp16")]; + tensor var_1161_split_sizes_0 = const()[name = tensor("op_1161_split_sizes_0"), val = tensor([512, 512])]; + tensor var_1161_axis_0 = const()[name = tensor("op_1161_axis_0"), val = tensor(-1)]; + tensor var_1161_cast_fp16_0, tensor var_1161_cast_fp16_1 = split(axis = var_1161_axis_0, split_sizes = var_1161_split_sizes_0, x = linear_76_cast_fp16)[name = tensor("op_1161_cast_fp16")]; + tensor mean_41_axes_0 = const()[name = tensor("mean_41_axes_0"), val = tensor([-1])]; + tensor mean_41_keep_dims_0 = const()[name = tensor("mean_41_keep_dims_0"), val = tensor(true)]; + tensor mean_41_cast_fp16 = reduce_mean(axes = mean_41_axes_0, keep_dims = mean_41_keep_dims_0, x = x_129_cast_fp16)[name = tensor("mean_41_cast_fp16")]; + tensor sub_32_cast_fp16 = sub(x = x_129_cast_fp16, y = mean_41_cast_fp16)[name = tensor("sub_32_cast_fp16")]; + tensor square_26_cast_fp16 = square(x = sub_32_cast_fp16)[name = tensor("square_26_cast_fp16")]; + tensor reduce_mean_53_axes_0 = const()[name = tensor("reduce_mean_53_axes_0"), val = tensor([-1])]; + tensor reduce_mean_53_keep_dims_0 = const()[name = tensor("reduce_mean_53_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_53_cast_fp16 = reduce_mean(axes = reduce_mean_53_axes_0, keep_dims = reduce_mean_53_keep_dims_0, x = square_26_cast_fp16)[name = tensor("reduce_mean_53_cast_fp16")]; + tensor var_1168_to_fp16 = const()[name = tensor("op_1168_to_fp16"), val = tensor(0x1.1p-20)]; + tensor var_1169_cast_fp16 = add(x = reduce_mean_53_cast_fp16, y = var_1168_to_fp16)[name = tensor("op_1169_cast_fp16")]; + tensor var_1170_cast_fp16 = sqrt(x = var_1169_cast_fp16)[name = tensor("op_1170_cast_fp16")]; + tensor x_131_cast_fp16 = real_div(x = sub_32_cast_fp16, y = var_1170_cast_fp16)[name = tensor("x_131_cast_fp16")]; + tensor var_1172_promoted_to_fp16 = const()[name = tensor("op_1172_promoted_to_fp16"), val = tensor(0x1p+0)]; + tensor var_1173_cast_fp16 = add(x = var_1161_cast_fp16_1, y = var_1172_promoted_to_fp16)[name = tensor("op_1173_cast_fp16")]; + tensor var_1174_cast_fp16 = mul(x = x_131_cast_fp16, y = var_1173_cast_fp16)[name = tensor("op_1174_cast_fp16")]; + tensor input_201_cast_fp16 = add(x = var_1174_cast_fp16, y = var_1161_cast_fp16_0)[name = tensor("input_201_cast_fp16")]; + tensor linear_77_cast_fp16 = linear(bias = flow_net_final_layer_linear_bias_to_fp16, weight = flow_net_final_layer_linear_weight_to_fp16, x = input_201_cast_fp16)[name = tensor("linear_77_cast_fp16")]; + tensor var_1185_to_fp16 = const()[name = tensor("op_1185_to_fp16"), val = tensor(0x1p-3)]; + tensor var_1186_cast_fp16 = mul(x = linear_77_cast_fp16, y = var_1185_to_fp16)[name = tensor("op_1186_cast_fp16")]; + tensor input_203_cast_fp16 = add(x = input_135_cast_fp16, y = var_1186_cast_fp16)[name = tensor("input_203_cast_fp16")]; + tensor linear_78_cast_fp16 = linear(bias = flow_net_input_proj_bias_to_fp16, weight = flow_net_input_proj_weight_to_fp16, x = input_203_cast_fp16)[name = tensor("linear_78_cast_fp16")]; + tensor input_207_to_fp16 = const()[name = tensor("input_207_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(19001728)))]; + tensor input_209_cast_fp16 = silu(x = input_207_to_fp16)[name = tensor("input_209_cast_fp16")]; + tensor linear_80_cast_fp16 = linear(bias = flow_net_time_embed_0_mlp_2_bias_to_fp16, weight = flow_net_time_embed_0_mlp_2_weight_to_fp16, x = input_209_cast_fp16)[name = tensor("linear_80_cast_fp16")]; + tensor reduce_mean_54_axes_0 = const()[name = tensor("reduce_mean_54_axes_0"), val = tensor([-1])]; + tensor reduce_mean_54_keep_dims_0 = const()[name = tensor("reduce_mean_54_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_54_cast_fp16 = reduce_mean(axes = reduce_mean_54_axes_0, keep_dims = reduce_mean_54_keep_dims_0, x = linear_80_cast_fp16)[name = tensor("reduce_mean_54_cast_fp16")]; + tensor sub_33_cast_fp16 = sub(x = linear_80_cast_fp16, y = reduce_mean_54_cast_fp16)[name = tensor("sub_33_cast_fp16")]; + tensor square_27_cast_fp16 = square(x = sub_33_cast_fp16)[name = tensor("square_27_cast_fp16")]; + tensor reduce_mean_55_axes_0 = const()[name = tensor("reduce_mean_55_axes_0"), val = tensor([-1])]; + tensor reduce_mean_55_keep_dims_0 = const()[name = tensor("reduce_mean_55_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_55_cast_fp16 = reduce_mean(axes = reduce_mean_55_axes_0, keep_dims = reduce_mean_55_keep_dims_0, x = square_27_cast_fp16)[name = tensor("reduce_mean_55_cast_fp16")]; + tensor real_div_6_to_fp16 = const()[name = tensor("real_div_6_to_fp16"), val = tensor(0x1.008p+0)]; + tensor mul_6_cast_fp16 = mul(x = reduce_mean_55_cast_fp16, y = real_div_6_to_fp16)[name = tensor("mul_6_cast_fp16")]; + tensor var_1252_to_fp16 = const()[name = tensor("op_1252_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_55_cast_fp16 = add(x = mul_6_cast_fp16, y = var_1252_to_fp16)[name = tensor("var_55_cast_fp16")]; + tensor var_1255_epsilon_0 = const()[name = tensor("op_1255_epsilon_0"), val = tensor(0x1.197998p-40)]; + tensor var_1255_cast_fp16 = rsqrt(epsilon = var_1255_epsilon_0, x = var_55_cast_fp16)[name = tensor("op_1255_cast_fp16")]; + tensor var_1256_cast_fp16 = mul(x = const_3_to_fp16, y = var_1255_cast_fp16)[name = tensor("op_1256_cast_fp16")]; + tensor var_1257_cast_fp16 = mul(x = linear_80_cast_fp16, y = var_1256_cast_fp16)[name = tensor("op_1257_cast_fp16")]; + tensor input_213_to_fp16 = const()[name = tensor("input_213_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(19002816)))]; + tensor input_215_cast_fp16 = silu(x = input_213_to_fp16)[name = tensor("input_215_cast_fp16")]; + tensor linear_82_cast_fp16 = linear(bias = flow_net_time_embed_1_mlp_2_bias_to_fp16, weight = flow_net_time_embed_1_mlp_2_weight_to_fp16, x = input_215_cast_fp16)[name = tensor("linear_82_cast_fp16")]; + tensor reduce_mean_56_axes_0 = const()[name = tensor("reduce_mean_56_axes_0"), val = tensor([-1])]; + tensor reduce_mean_56_keep_dims_0 = const()[name = tensor("reduce_mean_56_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_56_cast_fp16 = reduce_mean(axes = reduce_mean_56_axes_0, keep_dims = reduce_mean_56_keep_dims_0, x = linear_82_cast_fp16)[name = tensor("reduce_mean_56_cast_fp16")]; + tensor sub_35_cast_fp16 = sub(x = linear_82_cast_fp16, y = reduce_mean_56_cast_fp16)[name = tensor("sub_35_cast_fp16")]; + tensor square_28_cast_fp16 = square(x = sub_35_cast_fp16)[name = tensor("square_28_cast_fp16")]; + tensor reduce_mean_57_axes_0 = const()[name = tensor("reduce_mean_57_axes_0"), val = tensor([-1])]; + tensor reduce_mean_57_keep_dims_0 = const()[name = tensor("reduce_mean_57_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_57_cast_fp16 = reduce_mean(axes = reduce_mean_57_axes_0, keep_dims = reduce_mean_57_keep_dims_0, x = square_28_cast_fp16)[name = tensor("reduce_mean_57_cast_fp16")]; + tensor real_div_7_to_fp16 = const()[name = tensor("real_div_7_to_fp16"), val = tensor(0x1.008p+0)]; + tensor mul_7_cast_fp16 = mul(x = reduce_mean_57_cast_fp16, y = real_div_7_to_fp16)[name = tensor("mul_7_cast_fp16")]; + tensor var_1289_to_fp16 = const()[name = tensor("op_1289_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_57_cast_fp16 = add(x = mul_7_cast_fp16, y = var_1289_to_fp16)[name = tensor("var_57_cast_fp16")]; + tensor var_1292_epsilon_0 = const()[name = tensor("op_1292_epsilon_0"), val = tensor(0x1.197998p-40)]; + tensor var_1292_cast_fp16 = rsqrt(epsilon = var_1292_epsilon_0, x = var_57_cast_fp16)[name = tensor("op_1292_cast_fp16")]; + tensor var_1293_cast_fp16 = mul(x = const_5_to_fp16, y = var_1292_cast_fp16)[name = tensor("op_1293_cast_fp16")]; + tensor var_1294_cast_fp16 = mul(x = linear_82_cast_fp16, y = var_1293_cast_fp16)[name = tensor("op_1294_cast_fp16")]; + tensor var_1306_cast_fp16 = add(x = var_1257_cast_fp16, y = var_1294_cast_fp16)[name = tensor("op_1306_cast_fp16")]; + tensor _inversed_t_combined_7_y_0_to_fp16 = const()[name = tensor("_inversed_t_combined_7_y_0_to_fp16"), val = tensor(0x1p-1)]; + tensor _inversed_t_combined_7_cast_fp16 = mul(x = var_1306_cast_fp16, y = _inversed_t_combined_7_y_0_to_fp16)[name = tensor("_inversed_t_combined_7_cast_fp16")]; + tensor input_217_cast_fp16 = add(x = _inversed_t_combined_7_cast_fp16, y = linear_5_cast_fp16)[name = tensor("input_217_cast_fp16")]; + tensor input_219_cast_fp16 = silu(x = input_217_cast_fp16)[name = tensor("input_219_cast_fp16")]; + tensor linear_84_cast_fp16 = linear(bias = flow_net_res_blocks_0_adaLN_modulation_1_bias_to_fp16, weight = flow_net_res_blocks_0_adaLN_modulation_1_weight_to_fp16, x = input_219_cast_fp16)[name = tensor("linear_84_cast_fp16")]; + tensor var_1321_split_sizes_0 = const()[name = tensor("op_1321_split_sizes_0"), val = tensor([512, 512, 512])]; + tensor var_1321_axis_0 = const()[name = tensor("op_1321_axis_0"), val = tensor(-1)]; + tensor var_1321_cast_fp16_0, tensor var_1321_cast_fp16_1, tensor var_1321_cast_fp16_2 = split(axis = var_1321_axis_0, split_sizes = var_1321_split_sizes_0, x = linear_84_cast_fp16)[name = tensor("op_1321_cast_fp16")]; + tensor mean_43_axes_0 = const()[name = tensor("mean_43_axes_0"), val = tensor([-1])]; + tensor mean_43_keep_dims_0 = const()[name = tensor("mean_43_keep_dims_0"), val = tensor(true)]; + tensor mean_43_cast_fp16 = reduce_mean(axes = mean_43_axes_0, keep_dims = mean_43_keep_dims_0, x = linear_78_cast_fp16)[name = tensor("mean_43_cast_fp16")]; + tensor sub_37_cast_fp16 = sub(x = linear_78_cast_fp16, y = mean_43_cast_fp16)[name = tensor("sub_37_cast_fp16")]; + tensor square_29_cast_fp16 = square(x = sub_37_cast_fp16)[name = tensor("square_29_cast_fp16")]; + tensor reduce_mean_59_axes_0 = const()[name = tensor("reduce_mean_59_axes_0"), val = tensor([-1])]; + tensor reduce_mean_59_keep_dims_0 = const()[name = tensor("reduce_mean_59_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_59_cast_fp16 = reduce_mean(axes = reduce_mean_59_axes_0, keep_dims = reduce_mean_59_keep_dims_0, x = square_29_cast_fp16)[name = tensor("reduce_mean_59_cast_fp16")]; + tensor var_1331_to_fp16 = const()[name = tensor("op_1331_to_fp16"), val = tensor(0x1.1p-20)]; + tensor var_1332_cast_fp16 = add(x = reduce_mean_59_cast_fp16, y = var_1331_to_fp16)[name = tensor("op_1332_cast_fp16")]; + tensor var_1333_cast_fp16 = sqrt(x = var_1332_cast_fp16)[name = tensor("op_1333_cast_fp16")]; + tensor x_139_cast_fp16 = real_div(x = sub_37_cast_fp16, y = var_1333_cast_fp16)[name = tensor("x_139_cast_fp16")]; + tensor var_1335_cast_fp16 = mul(x = x_139_cast_fp16, y = flow_net_res_blocks_0_in_ln_weight_to_fp16)[name = tensor("op_1335_cast_fp16")]; + tensor x_141_cast_fp16 = add(x = var_1335_cast_fp16, y = flow_net_res_blocks_0_in_ln_bias_to_fp16)[name = tensor("x_141_cast_fp16")]; + tensor var_1337_promoted_to_fp16 = const()[name = tensor("op_1337_promoted_to_fp16"), val = tensor(0x1p+0)]; + tensor var_1338_cast_fp16 = add(x = var_1321_cast_fp16_1, y = var_1337_promoted_to_fp16)[name = tensor("op_1338_cast_fp16")]; + tensor var_1339_cast_fp16 = mul(x = x_141_cast_fp16, y = var_1338_cast_fp16)[name = tensor("op_1339_cast_fp16")]; + tensor input_221_cast_fp16 = add(x = var_1339_cast_fp16, y = var_1321_cast_fp16_0)[name = tensor("input_221_cast_fp16")]; + tensor linear_85_cast_fp16 = linear(bias = flow_net_res_blocks_0_mlp_0_bias_to_fp16, weight = flow_net_res_blocks_0_mlp_0_weight_to_fp16, x = input_221_cast_fp16)[name = tensor("linear_85_cast_fp16")]; + tensor input_225_cast_fp16 = silu(x = linear_85_cast_fp16)[name = tensor("input_225_cast_fp16")]; + tensor linear_86_cast_fp16 = linear(bias = flow_net_res_blocks_0_mlp_2_bias_to_fp16, weight = flow_net_res_blocks_0_mlp_2_weight_to_fp16, x = input_225_cast_fp16)[name = tensor("linear_86_cast_fp16")]; + tensor var_1350_cast_fp16 = mul(x = var_1321_cast_fp16_2, y = linear_86_cast_fp16)[name = tensor("op_1350_cast_fp16")]; + tensor x_143_cast_fp16 = add(x = linear_78_cast_fp16, y = var_1350_cast_fp16)[name = tensor("x_143_cast_fp16")]; + tensor linear_87_cast_fp16 = linear(bias = flow_net_res_blocks_1_adaLN_modulation_1_bias_to_fp16, weight = flow_net_res_blocks_1_adaLN_modulation_1_weight_to_fp16, x = input_219_cast_fp16)[name = tensor("linear_87_cast_fp16")]; + tensor var_1360_split_sizes_0 = const()[name = tensor("op_1360_split_sizes_0"), val = tensor([512, 512, 512])]; + tensor var_1360_axis_0 = const()[name = tensor("op_1360_axis_0"), val = tensor(-1)]; + tensor var_1360_cast_fp16_0, tensor var_1360_cast_fp16_1, tensor var_1360_cast_fp16_2 = split(axis = var_1360_axis_0, split_sizes = var_1360_split_sizes_0, x = linear_87_cast_fp16)[name = tensor("op_1360_cast_fp16")]; + tensor mean_45_axes_0 = const()[name = tensor("mean_45_axes_0"), val = tensor([-1])]; + tensor mean_45_keep_dims_0 = const()[name = tensor("mean_45_keep_dims_0"), val = tensor(true)]; + tensor mean_45_cast_fp16 = reduce_mean(axes = mean_45_axes_0, keep_dims = mean_45_keep_dims_0, x = x_143_cast_fp16)[name = tensor("mean_45_cast_fp16")]; + tensor sub_38_cast_fp16 = sub(x = x_143_cast_fp16, y = mean_45_cast_fp16)[name = tensor("sub_38_cast_fp16")]; + tensor square_30_cast_fp16 = square(x = sub_38_cast_fp16)[name = tensor("square_30_cast_fp16")]; + tensor reduce_mean_61_axes_0 = const()[name = tensor("reduce_mean_61_axes_0"), val = tensor([-1])]; + tensor reduce_mean_61_keep_dims_0 = const()[name = tensor("reduce_mean_61_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_61_cast_fp16 = reduce_mean(axes = reduce_mean_61_axes_0, keep_dims = reduce_mean_61_keep_dims_0, x = square_30_cast_fp16)[name = tensor("reduce_mean_61_cast_fp16")]; + tensor var_1370_to_fp16 = const()[name = tensor("op_1370_to_fp16"), val = tensor(0x1.1p-20)]; + tensor var_1371_cast_fp16 = add(x = reduce_mean_61_cast_fp16, y = var_1370_to_fp16)[name = tensor("op_1371_cast_fp16")]; + tensor var_1372_cast_fp16 = sqrt(x = var_1371_cast_fp16)[name = tensor("op_1372_cast_fp16")]; + tensor x_145_cast_fp16 = real_div(x = sub_38_cast_fp16, y = var_1372_cast_fp16)[name = tensor("x_145_cast_fp16")]; + tensor var_1374_cast_fp16 = mul(x = x_145_cast_fp16, y = flow_net_res_blocks_1_in_ln_weight_to_fp16)[name = tensor("op_1374_cast_fp16")]; + tensor x_147_cast_fp16 = add(x = var_1374_cast_fp16, y = flow_net_res_blocks_1_in_ln_bias_to_fp16)[name = tensor("x_147_cast_fp16")]; + tensor var_1376_promoted_to_fp16 = const()[name = tensor("op_1376_promoted_to_fp16"), val = tensor(0x1p+0)]; + tensor var_1377_cast_fp16 = add(x = var_1360_cast_fp16_1, y = var_1376_promoted_to_fp16)[name = tensor("op_1377_cast_fp16")]; + tensor var_1378_cast_fp16 = mul(x = x_147_cast_fp16, y = var_1377_cast_fp16)[name = tensor("op_1378_cast_fp16")]; + tensor input_229_cast_fp16 = add(x = var_1378_cast_fp16, y = var_1360_cast_fp16_0)[name = tensor("input_229_cast_fp16")]; + tensor linear_88_cast_fp16 = linear(bias = flow_net_res_blocks_1_mlp_0_bias_to_fp16, weight = flow_net_res_blocks_1_mlp_0_weight_to_fp16, x = input_229_cast_fp16)[name = tensor("linear_88_cast_fp16")]; + tensor input_233_cast_fp16 = silu(x = linear_88_cast_fp16)[name = tensor("input_233_cast_fp16")]; + tensor linear_89_cast_fp16 = linear(bias = flow_net_res_blocks_1_mlp_2_bias_to_fp16, weight = flow_net_res_blocks_1_mlp_2_weight_to_fp16, x = input_233_cast_fp16)[name = tensor("linear_89_cast_fp16")]; + tensor var_1389_cast_fp16 = mul(x = var_1360_cast_fp16_2, y = linear_89_cast_fp16)[name = tensor("op_1389_cast_fp16")]; + tensor x_149_cast_fp16 = add(x = x_143_cast_fp16, y = var_1389_cast_fp16)[name = tensor("x_149_cast_fp16")]; + tensor linear_90_cast_fp16 = linear(bias = flow_net_res_blocks_2_adaLN_modulation_1_bias_to_fp16, weight = flow_net_res_blocks_2_adaLN_modulation_1_weight_to_fp16, x = input_219_cast_fp16)[name = tensor("linear_90_cast_fp16")]; + tensor var_1399_split_sizes_0 = const()[name = tensor("op_1399_split_sizes_0"), val = tensor([512, 512, 512])]; + tensor var_1399_axis_0 = const()[name = tensor("op_1399_axis_0"), val = tensor(-1)]; + tensor var_1399_cast_fp16_0, tensor var_1399_cast_fp16_1, tensor var_1399_cast_fp16_2 = split(axis = var_1399_axis_0, split_sizes = var_1399_split_sizes_0, x = linear_90_cast_fp16)[name = tensor("op_1399_cast_fp16")]; + tensor mean_47_axes_0 = const()[name = tensor("mean_47_axes_0"), val = tensor([-1])]; + tensor mean_47_keep_dims_0 = const()[name = tensor("mean_47_keep_dims_0"), val = tensor(true)]; + tensor mean_47_cast_fp16 = reduce_mean(axes = mean_47_axes_0, keep_dims = mean_47_keep_dims_0, x = x_149_cast_fp16)[name = tensor("mean_47_cast_fp16")]; + tensor sub_39_cast_fp16 = sub(x = x_149_cast_fp16, y = mean_47_cast_fp16)[name = tensor("sub_39_cast_fp16")]; + tensor square_31_cast_fp16 = square(x = sub_39_cast_fp16)[name = tensor("square_31_cast_fp16")]; + tensor reduce_mean_63_axes_0 = const()[name = tensor("reduce_mean_63_axes_0"), val = tensor([-1])]; + tensor reduce_mean_63_keep_dims_0 = const()[name = tensor("reduce_mean_63_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_63_cast_fp16 = reduce_mean(axes = reduce_mean_63_axes_0, keep_dims = reduce_mean_63_keep_dims_0, x = square_31_cast_fp16)[name = tensor("reduce_mean_63_cast_fp16")]; + tensor var_1409_to_fp16 = const()[name = tensor("op_1409_to_fp16"), val = tensor(0x1.1p-20)]; + tensor var_1410_cast_fp16 = add(x = reduce_mean_63_cast_fp16, y = var_1409_to_fp16)[name = tensor("op_1410_cast_fp16")]; + tensor var_1411_cast_fp16 = sqrt(x = var_1410_cast_fp16)[name = tensor("op_1411_cast_fp16")]; + tensor x_151_cast_fp16 = real_div(x = sub_39_cast_fp16, y = var_1411_cast_fp16)[name = tensor("x_151_cast_fp16")]; + tensor var_1413_cast_fp16 = mul(x = x_151_cast_fp16, y = flow_net_res_blocks_2_in_ln_weight_to_fp16)[name = tensor("op_1413_cast_fp16")]; + tensor x_153_cast_fp16 = add(x = var_1413_cast_fp16, y = flow_net_res_blocks_2_in_ln_bias_to_fp16)[name = tensor("x_153_cast_fp16")]; + tensor var_1415_promoted_to_fp16 = const()[name = tensor("op_1415_promoted_to_fp16"), val = tensor(0x1p+0)]; + tensor var_1416_cast_fp16 = add(x = var_1399_cast_fp16_1, y = var_1415_promoted_to_fp16)[name = tensor("op_1416_cast_fp16")]; + tensor var_1417_cast_fp16 = mul(x = x_153_cast_fp16, y = var_1416_cast_fp16)[name = tensor("op_1417_cast_fp16")]; + tensor input_237_cast_fp16 = add(x = var_1417_cast_fp16, y = var_1399_cast_fp16_0)[name = tensor("input_237_cast_fp16")]; + tensor linear_91_cast_fp16 = linear(bias = flow_net_res_blocks_2_mlp_0_bias_to_fp16, weight = flow_net_res_blocks_2_mlp_0_weight_to_fp16, x = input_237_cast_fp16)[name = tensor("linear_91_cast_fp16")]; + tensor input_241_cast_fp16 = silu(x = linear_91_cast_fp16)[name = tensor("input_241_cast_fp16")]; + tensor linear_92_cast_fp16 = linear(bias = flow_net_res_blocks_2_mlp_2_bias_to_fp16, weight = flow_net_res_blocks_2_mlp_2_weight_to_fp16, x = input_241_cast_fp16)[name = tensor("linear_92_cast_fp16")]; + tensor var_1428_cast_fp16 = mul(x = var_1399_cast_fp16_2, y = linear_92_cast_fp16)[name = tensor("op_1428_cast_fp16")]; + tensor x_155_cast_fp16 = add(x = x_149_cast_fp16, y = var_1428_cast_fp16)[name = tensor("x_155_cast_fp16")]; + tensor linear_93_cast_fp16 = linear(bias = flow_net_res_blocks_3_adaLN_modulation_1_bias_to_fp16, weight = flow_net_res_blocks_3_adaLN_modulation_1_weight_to_fp16, x = input_219_cast_fp16)[name = tensor("linear_93_cast_fp16")]; + tensor var_1438_split_sizes_0 = const()[name = tensor("op_1438_split_sizes_0"), val = tensor([512, 512, 512])]; + tensor var_1438_axis_0 = const()[name = tensor("op_1438_axis_0"), val = tensor(-1)]; + tensor var_1438_cast_fp16_0, tensor var_1438_cast_fp16_1, tensor var_1438_cast_fp16_2 = split(axis = var_1438_axis_0, split_sizes = var_1438_split_sizes_0, x = linear_93_cast_fp16)[name = tensor("op_1438_cast_fp16")]; + tensor mean_49_axes_0 = const()[name = tensor("mean_49_axes_0"), val = tensor([-1])]; + tensor mean_49_keep_dims_0 = const()[name = tensor("mean_49_keep_dims_0"), val = tensor(true)]; + tensor mean_49_cast_fp16 = reduce_mean(axes = mean_49_axes_0, keep_dims = mean_49_keep_dims_0, x = x_155_cast_fp16)[name = tensor("mean_49_cast_fp16")]; + tensor sub_40_cast_fp16 = sub(x = x_155_cast_fp16, y = mean_49_cast_fp16)[name = tensor("sub_40_cast_fp16")]; + tensor square_32_cast_fp16 = square(x = sub_40_cast_fp16)[name = tensor("square_32_cast_fp16")]; + tensor reduce_mean_65_axes_0 = const()[name = tensor("reduce_mean_65_axes_0"), val = tensor([-1])]; + tensor reduce_mean_65_keep_dims_0 = const()[name = tensor("reduce_mean_65_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_65_cast_fp16 = reduce_mean(axes = reduce_mean_65_axes_0, keep_dims = reduce_mean_65_keep_dims_0, x = square_32_cast_fp16)[name = tensor("reduce_mean_65_cast_fp16")]; + tensor var_1448_to_fp16 = const()[name = tensor("op_1448_to_fp16"), val = tensor(0x1.1p-20)]; + tensor var_1449_cast_fp16 = add(x = reduce_mean_65_cast_fp16, y = var_1448_to_fp16)[name = tensor("op_1449_cast_fp16")]; + tensor var_1450_cast_fp16 = sqrt(x = var_1449_cast_fp16)[name = tensor("op_1450_cast_fp16")]; + tensor x_157_cast_fp16 = real_div(x = sub_40_cast_fp16, y = var_1450_cast_fp16)[name = tensor("x_157_cast_fp16")]; + tensor var_1452_cast_fp16 = mul(x = x_157_cast_fp16, y = flow_net_res_blocks_3_in_ln_weight_to_fp16)[name = tensor("op_1452_cast_fp16")]; + tensor x_159_cast_fp16 = add(x = var_1452_cast_fp16, y = flow_net_res_blocks_3_in_ln_bias_to_fp16)[name = tensor("x_159_cast_fp16")]; + tensor var_1454_promoted_to_fp16 = const()[name = tensor("op_1454_promoted_to_fp16"), val = tensor(0x1p+0)]; + tensor var_1455_cast_fp16 = add(x = var_1438_cast_fp16_1, y = var_1454_promoted_to_fp16)[name = tensor("op_1455_cast_fp16")]; + tensor var_1456_cast_fp16 = mul(x = x_159_cast_fp16, y = var_1455_cast_fp16)[name = tensor("op_1456_cast_fp16")]; + tensor input_245_cast_fp16 = add(x = var_1456_cast_fp16, y = var_1438_cast_fp16_0)[name = tensor("input_245_cast_fp16")]; + tensor linear_94_cast_fp16 = linear(bias = flow_net_res_blocks_3_mlp_0_bias_to_fp16, weight = flow_net_res_blocks_3_mlp_0_weight_to_fp16, x = input_245_cast_fp16)[name = tensor("linear_94_cast_fp16")]; + tensor input_249_cast_fp16 = silu(x = linear_94_cast_fp16)[name = tensor("input_249_cast_fp16")]; + tensor linear_95_cast_fp16 = linear(bias = flow_net_res_blocks_3_mlp_2_bias_to_fp16, weight = flow_net_res_blocks_3_mlp_2_weight_to_fp16, x = input_249_cast_fp16)[name = tensor("linear_95_cast_fp16")]; + tensor var_1467_cast_fp16 = mul(x = var_1438_cast_fp16_2, y = linear_95_cast_fp16)[name = tensor("op_1467_cast_fp16")]; + tensor x_161_cast_fp16 = add(x = x_155_cast_fp16, y = var_1467_cast_fp16)[name = tensor("x_161_cast_fp16")]; + tensor linear_96_cast_fp16 = linear(bias = flow_net_res_blocks_4_adaLN_modulation_1_bias_to_fp16, weight = flow_net_res_blocks_4_adaLN_modulation_1_weight_to_fp16, x = input_219_cast_fp16)[name = tensor("linear_96_cast_fp16")]; + tensor var_1477_split_sizes_0 = const()[name = tensor("op_1477_split_sizes_0"), val = tensor([512, 512, 512])]; + tensor var_1477_axis_0 = const()[name = tensor("op_1477_axis_0"), val = tensor(-1)]; + tensor var_1477_cast_fp16_0, tensor var_1477_cast_fp16_1, tensor var_1477_cast_fp16_2 = split(axis = var_1477_axis_0, split_sizes = var_1477_split_sizes_0, x = linear_96_cast_fp16)[name = tensor("op_1477_cast_fp16")]; + tensor mean_51_axes_0 = const()[name = tensor("mean_51_axes_0"), val = tensor([-1])]; + tensor mean_51_keep_dims_0 = const()[name = tensor("mean_51_keep_dims_0"), val = tensor(true)]; + tensor mean_51_cast_fp16 = reduce_mean(axes = mean_51_axes_0, keep_dims = mean_51_keep_dims_0, x = x_161_cast_fp16)[name = tensor("mean_51_cast_fp16")]; + tensor sub_41_cast_fp16 = sub(x = x_161_cast_fp16, y = mean_51_cast_fp16)[name = tensor("sub_41_cast_fp16")]; + tensor square_33_cast_fp16 = square(x = sub_41_cast_fp16)[name = tensor("square_33_cast_fp16")]; + tensor reduce_mean_67_axes_0 = const()[name = tensor("reduce_mean_67_axes_0"), val = tensor([-1])]; + tensor reduce_mean_67_keep_dims_0 = const()[name = tensor("reduce_mean_67_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_67_cast_fp16 = reduce_mean(axes = reduce_mean_67_axes_0, keep_dims = reduce_mean_67_keep_dims_0, x = square_33_cast_fp16)[name = tensor("reduce_mean_67_cast_fp16")]; + tensor var_1487_to_fp16 = const()[name = tensor("op_1487_to_fp16"), val = tensor(0x1.1p-20)]; + tensor var_1488_cast_fp16 = add(x = reduce_mean_67_cast_fp16, y = var_1487_to_fp16)[name = tensor("op_1488_cast_fp16")]; + tensor var_1489_cast_fp16 = sqrt(x = var_1488_cast_fp16)[name = tensor("op_1489_cast_fp16")]; + tensor x_163_cast_fp16 = real_div(x = sub_41_cast_fp16, y = var_1489_cast_fp16)[name = tensor("x_163_cast_fp16")]; + tensor var_1491_cast_fp16 = mul(x = x_163_cast_fp16, y = flow_net_res_blocks_4_in_ln_weight_to_fp16)[name = tensor("op_1491_cast_fp16")]; + tensor x_165_cast_fp16 = add(x = var_1491_cast_fp16, y = flow_net_res_blocks_4_in_ln_bias_to_fp16)[name = tensor("x_165_cast_fp16")]; + tensor var_1493_promoted_to_fp16 = const()[name = tensor("op_1493_promoted_to_fp16"), val = tensor(0x1p+0)]; + tensor var_1494_cast_fp16 = add(x = var_1477_cast_fp16_1, y = var_1493_promoted_to_fp16)[name = tensor("op_1494_cast_fp16")]; + tensor var_1495_cast_fp16 = mul(x = x_165_cast_fp16, y = var_1494_cast_fp16)[name = tensor("op_1495_cast_fp16")]; + tensor input_253_cast_fp16 = add(x = var_1495_cast_fp16, y = var_1477_cast_fp16_0)[name = tensor("input_253_cast_fp16")]; + tensor linear_97_cast_fp16 = linear(bias = flow_net_res_blocks_4_mlp_0_bias_to_fp16, weight = flow_net_res_blocks_4_mlp_0_weight_to_fp16, x = input_253_cast_fp16)[name = tensor("linear_97_cast_fp16")]; + tensor input_257_cast_fp16 = silu(x = linear_97_cast_fp16)[name = tensor("input_257_cast_fp16")]; + tensor linear_98_cast_fp16 = linear(bias = flow_net_res_blocks_4_mlp_2_bias_to_fp16, weight = flow_net_res_blocks_4_mlp_2_weight_to_fp16, x = input_257_cast_fp16)[name = tensor("linear_98_cast_fp16")]; + tensor var_1506_cast_fp16 = mul(x = var_1477_cast_fp16_2, y = linear_98_cast_fp16)[name = tensor("op_1506_cast_fp16")]; + tensor x_167_cast_fp16 = add(x = x_161_cast_fp16, y = var_1506_cast_fp16)[name = tensor("x_167_cast_fp16")]; + tensor linear_99_cast_fp16 = linear(bias = flow_net_res_blocks_5_adaLN_modulation_1_bias_to_fp16, weight = flow_net_res_blocks_5_adaLN_modulation_1_weight_to_fp16, x = input_219_cast_fp16)[name = tensor("linear_99_cast_fp16")]; + tensor var_1516_split_sizes_0 = const()[name = tensor("op_1516_split_sizes_0"), val = tensor([512, 512, 512])]; + tensor var_1516_axis_0 = const()[name = tensor("op_1516_axis_0"), val = tensor(-1)]; + tensor var_1516_cast_fp16_0, tensor var_1516_cast_fp16_1, tensor var_1516_cast_fp16_2 = split(axis = var_1516_axis_0, split_sizes = var_1516_split_sizes_0, x = linear_99_cast_fp16)[name = tensor("op_1516_cast_fp16")]; + tensor mean_53_axes_0 = const()[name = tensor("mean_53_axes_0"), val = tensor([-1])]; + tensor mean_53_keep_dims_0 = const()[name = tensor("mean_53_keep_dims_0"), val = tensor(true)]; + tensor mean_53_cast_fp16 = reduce_mean(axes = mean_53_axes_0, keep_dims = mean_53_keep_dims_0, x = x_167_cast_fp16)[name = tensor("mean_53_cast_fp16")]; + tensor sub_42_cast_fp16 = sub(x = x_167_cast_fp16, y = mean_53_cast_fp16)[name = tensor("sub_42_cast_fp16")]; + tensor square_34_cast_fp16 = square(x = sub_42_cast_fp16)[name = tensor("square_34_cast_fp16")]; + tensor reduce_mean_69_axes_0 = const()[name = tensor("reduce_mean_69_axes_0"), val = tensor([-1])]; + tensor reduce_mean_69_keep_dims_0 = const()[name = tensor("reduce_mean_69_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_69_cast_fp16 = reduce_mean(axes = reduce_mean_69_axes_0, keep_dims = reduce_mean_69_keep_dims_0, x = square_34_cast_fp16)[name = tensor("reduce_mean_69_cast_fp16")]; + tensor var_1526_to_fp16 = const()[name = tensor("op_1526_to_fp16"), val = tensor(0x1.1p-20)]; + tensor var_1527_cast_fp16 = add(x = reduce_mean_69_cast_fp16, y = var_1526_to_fp16)[name = tensor("op_1527_cast_fp16")]; + tensor var_1528_cast_fp16 = sqrt(x = var_1527_cast_fp16)[name = tensor("op_1528_cast_fp16")]; + tensor x_169_cast_fp16 = real_div(x = sub_42_cast_fp16, y = var_1528_cast_fp16)[name = tensor("x_169_cast_fp16")]; + tensor var_1530_cast_fp16 = mul(x = x_169_cast_fp16, y = flow_net_res_blocks_5_in_ln_weight_to_fp16)[name = tensor("op_1530_cast_fp16")]; + tensor x_171_cast_fp16 = add(x = var_1530_cast_fp16, y = flow_net_res_blocks_5_in_ln_bias_to_fp16)[name = tensor("x_171_cast_fp16")]; + tensor var_1532_promoted_to_fp16 = const()[name = tensor("op_1532_promoted_to_fp16"), val = tensor(0x1p+0)]; + tensor var_1533_cast_fp16 = add(x = var_1516_cast_fp16_1, y = var_1532_promoted_to_fp16)[name = tensor("op_1533_cast_fp16")]; + tensor var_1534_cast_fp16 = mul(x = x_171_cast_fp16, y = var_1533_cast_fp16)[name = tensor("op_1534_cast_fp16")]; + tensor input_261_cast_fp16 = add(x = var_1534_cast_fp16, y = var_1516_cast_fp16_0)[name = tensor("input_261_cast_fp16")]; + tensor linear_100_cast_fp16 = linear(bias = flow_net_res_blocks_5_mlp_0_bias_to_fp16, weight = flow_net_res_blocks_5_mlp_0_weight_to_fp16, x = input_261_cast_fp16)[name = tensor("linear_100_cast_fp16")]; + tensor input_265_cast_fp16 = silu(x = linear_100_cast_fp16)[name = tensor("input_265_cast_fp16")]; + tensor linear_101_cast_fp16 = linear(bias = flow_net_res_blocks_5_mlp_2_bias_to_fp16, weight = flow_net_res_blocks_5_mlp_2_weight_to_fp16, x = input_265_cast_fp16)[name = tensor("linear_101_cast_fp16")]; + tensor var_1545_cast_fp16 = mul(x = var_1516_cast_fp16_2, y = linear_101_cast_fp16)[name = tensor("op_1545_cast_fp16")]; + tensor x_173_cast_fp16 = add(x = x_167_cast_fp16, y = var_1545_cast_fp16)[name = tensor("x_173_cast_fp16")]; + tensor linear_102_cast_fp16 = linear(bias = flow_net_final_layer_adaLN_modulation_1_bias_to_fp16, weight = flow_net_final_layer_adaLN_modulation_1_weight_to_fp16, x = input_219_cast_fp16)[name = tensor("linear_102_cast_fp16")]; + tensor var_1554_split_sizes_0 = const()[name = tensor("op_1554_split_sizes_0"), val = tensor([512, 512])]; + tensor var_1554_axis_0 = const()[name = tensor("op_1554_axis_0"), val = tensor(-1)]; + tensor var_1554_cast_fp16_0, tensor var_1554_cast_fp16_1 = split(axis = var_1554_axis_0, split_sizes = var_1554_split_sizes_0, x = linear_102_cast_fp16)[name = tensor("op_1554_cast_fp16")]; + tensor mean_55_axes_0 = const()[name = tensor("mean_55_axes_0"), val = tensor([-1])]; + tensor mean_55_keep_dims_0 = const()[name = tensor("mean_55_keep_dims_0"), val = tensor(true)]; + tensor mean_55_cast_fp16 = reduce_mean(axes = mean_55_axes_0, keep_dims = mean_55_keep_dims_0, x = x_173_cast_fp16)[name = tensor("mean_55_cast_fp16")]; + tensor sub_43_cast_fp16 = sub(x = x_173_cast_fp16, y = mean_55_cast_fp16)[name = tensor("sub_43_cast_fp16")]; + tensor square_35_cast_fp16 = square(x = sub_43_cast_fp16)[name = tensor("square_35_cast_fp16")]; + tensor reduce_mean_71_axes_0 = const()[name = tensor("reduce_mean_71_axes_0"), val = tensor([-1])]; + tensor reduce_mean_71_keep_dims_0 = const()[name = tensor("reduce_mean_71_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_71_cast_fp16 = reduce_mean(axes = reduce_mean_71_axes_0, keep_dims = reduce_mean_71_keep_dims_0, x = square_35_cast_fp16)[name = tensor("reduce_mean_71_cast_fp16")]; + tensor var_1561_to_fp16 = const()[name = tensor("op_1561_to_fp16"), val = tensor(0x1.1p-20)]; + tensor var_1562_cast_fp16 = add(x = reduce_mean_71_cast_fp16, y = var_1561_to_fp16)[name = tensor("op_1562_cast_fp16")]; + tensor var_1563_cast_fp16 = sqrt(x = var_1562_cast_fp16)[name = tensor("op_1563_cast_fp16")]; + tensor x_175_cast_fp16 = real_div(x = sub_43_cast_fp16, y = var_1563_cast_fp16)[name = tensor("x_175_cast_fp16")]; + tensor var_1565_promoted_to_fp16 = const()[name = tensor("op_1565_promoted_to_fp16"), val = tensor(0x1p+0)]; + tensor var_1566_cast_fp16 = add(x = var_1554_cast_fp16_1, y = var_1565_promoted_to_fp16)[name = tensor("op_1566_cast_fp16")]; + tensor var_1567_cast_fp16 = mul(x = x_175_cast_fp16, y = var_1566_cast_fp16)[name = tensor("op_1567_cast_fp16")]; + tensor input_269_cast_fp16 = add(x = var_1567_cast_fp16, y = var_1554_cast_fp16_0)[name = tensor("input_269_cast_fp16")]; + tensor linear_103_cast_fp16 = linear(bias = flow_net_final_layer_linear_bias_to_fp16, weight = flow_net_final_layer_linear_weight_to_fp16, x = input_269_cast_fp16)[name = tensor("linear_103_cast_fp16")]; + tensor var_1578_to_fp16 = const()[name = tensor("op_1578_to_fp16"), val = tensor(0x1p-3)]; + tensor var_1579_cast_fp16 = mul(x = linear_103_cast_fp16, y = var_1578_to_fp16)[name = tensor("op_1579_cast_fp16")]; + tensor input_271_cast_fp16 = add(x = input_203_cast_fp16, y = var_1579_cast_fp16)[name = tensor("input_271_cast_fp16")]; + tensor linear_104_cast_fp16 = linear(bias = flow_net_input_proj_bias_to_fp16, weight = flow_net_input_proj_weight_to_fp16, x = input_271_cast_fp16)[name = tensor("linear_104_cast_fp16")]; + tensor input_275_to_fp16 = const()[name = tensor("input_275_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(19003904)))]; + tensor input_277_cast_fp16 = silu(x = input_275_to_fp16)[name = tensor("input_277_cast_fp16")]; + tensor linear_106_cast_fp16 = linear(bias = flow_net_time_embed_0_mlp_2_bias_to_fp16, weight = flow_net_time_embed_0_mlp_2_weight_to_fp16, x = input_277_cast_fp16)[name = tensor("linear_106_cast_fp16")]; + tensor reduce_mean_72_axes_0 = const()[name = tensor("reduce_mean_72_axes_0"), val = tensor([-1])]; + tensor reduce_mean_72_keep_dims_0 = const()[name = tensor("reduce_mean_72_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_72_cast_fp16 = reduce_mean(axes = reduce_mean_72_axes_0, keep_dims = reduce_mean_72_keep_dims_0, x = linear_106_cast_fp16)[name = tensor("reduce_mean_72_cast_fp16")]; + tensor sub_44_cast_fp16 = sub(x = linear_106_cast_fp16, y = reduce_mean_72_cast_fp16)[name = tensor("sub_44_cast_fp16")]; + tensor square_36_cast_fp16 = square(x = sub_44_cast_fp16)[name = tensor("square_36_cast_fp16")]; + tensor reduce_mean_73_axes_0 = const()[name = tensor("reduce_mean_73_axes_0"), val = tensor([-1])]; + tensor reduce_mean_73_keep_dims_0 = const()[name = tensor("reduce_mean_73_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_73_cast_fp16 = reduce_mean(axes = reduce_mean_73_axes_0, keep_dims = reduce_mean_73_keep_dims_0, x = square_36_cast_fp16)[name = tensor("reduce_mean_73_cast_fp16")]; + tensor real_div_8_to_fp16 = const()[name = tensor("real_div_8_to_fp16"), val = tensor(0x1.008p+0)]; + tensor mul_8_cast_fp16 = mul(x = reduce_mean_73_cast_fp16, y = real_div_8_to_fp16)[name = tensor("mul_8_cast_fp16")]; + tensor var_1645_to_fp16 = const()[name = tensor("op_1645_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_73_cast_fp16_0 = add(x = mul_8_cast_fp16, y = var_1645_to_fp16)[name = tensor("var_73_cast_fp16")]; + tensor var_1648_epsilon_0 = const()[name = tensor("op_1648_epsilon_0"), val = tensor(0x1.197998p-40)]; + tensor var_1648_cast_fp16 = rsqrt(epsilon = var_1648_epsilon_0, x = var_73_cast_fp16_0)[name = tensor("op_1648_cast_fp16")]; + tensor var_1649_cast_fp16 = mul(x = const_3_to_fp16, y = var_1648_cast_fp16)[name = tensor("op_1649_cast_fp16")]; + tensor var_1650_cast_fp16 = mul(x = linear_106_cast_fp16, y = var_1649_cast_fp16)[name = tensor("op_1650_cast_fp16")]; + tensor input_281_to_fp16 = const()[name = tensor("input_281_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(19004992)))]; + tensor input_283_cast_fp16 = silu(x = input_281_to_fp16)[name = tensor("input_283_cast_fp16")]; + tensor linear_108_cast_fp16 = linear(bias = flow_net_time_embed_1_mlp_2_bias_to_fp16, weight = flow_net_time_embed_1_mlp_2_weight_to_fp16, x = input_283_cast_fp16)[name = tensor("linear_108_cast_fp16")]; + tensor reduce_mean_74_axes_0 = const()[name = tensor("reduce_mean_74_axes_0"), val = tensor([-1])]; + tensor reduce_mean_74_keep_dims_0 = const()[name = tensor("reduce_mean_74_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_74_cast_fp16 = reduce_mean(axes = reduce_mean_74_axes_0, keep_dims = reduce_mean_74_keep_dims_0, x = linear_108_cast_fp16)[name = tensor("reduce_mean_74_cast_fp16")]; + tensor sub_46_cast_fp16 = sub(x = linear_108_cast_fp16, y = reduce_mean_74_cast_fp16)[name = tensor("sub_46_cast_fp16")]; + tensor square_37_cast_fp16 = square(x = sub_46_cast_fp16)[name = tensor("square_37_cast_fp16")]; + tensor reduce_mean_75_axes_0 = const()[name = tensor("reduce_mean_75_axes_0"), val = tensor([-1])]; + tensor reduce_mean_75_keep_dims_0 = const()[name = tensor("reduce_mean_75_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_75_cast_fp16 = reduce_mean(axes = reduce_mean_75_axes_0, keep_dims = reduce_mean_75_keep_dims_0, x = square_37_cast_fp16)[name = tensor("reduce_mean_75_cast_fp16")]; + tensor real_div_9_to_fp16 = const()[name = tensor("real_div_9_to_fp16"), val = tensor(0x1.008p+0)]; + tensor mul_9_cast_fp16 = mul(x = reduce_mean_75_cast_fp16, y = real_div_9_to_fp16)[name = tensor("mul_9_cast_fp16")]; + tensor var_1682_to_fp16 = const()[name = tensor("op_1682_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_75_cast_fp16_0 = add(x = mul_9_cast_fp16, y = var_1682_to_fp16)[name = tensor("var_75_cast_fp16")]; + tensor var_1685_epsilon_0 = const()[name = tensor("op_1685_epsilon_0"), val = tensor(0x1.197998p-40)]; + tensor var_1685_cast_fp16 = rsqrt(epsilon = var_1685_epsilon_0, x = var_75_cast_fp16_0)[name = tensor("op_1685_cast_fp16")]; + tensor var_1686_cast_fp16 = mul(x = const_5_to_fp16, y = var_1685_cast_fp16)[name = tensor("op_1686_cast_fp16")]; + tensor var_1687_cast_fp16 = mul(x = linear_108_cast_fp16, y = var_1686_cast_fp16)[name = tensor("op_1687_cast_fp16")]; + tensor var_1699_cast_fp16 = add(x = var_1650_cast_fp16, y = var_1687_cast_fp16)[name = tensor("op_1699_cast_fp16")]; + tensor _inversed_t_combined_9_y_0_to_fp16 = const()[name = tensor("_inversed_t_combined_9_y_0_to_fp16"), val = tensor(0x1p-1)]; + tensor _inversed_t_combined_9_cast_fp16 = mul(x = var_1699_cast_fp16, y = _inversed_t_combined_9_y_0_to_fp16)[name = tensor("_inversed_t_combined_9_cast_fp16")]; + tensor input_285_cast_fp16 = add(x = _inversed_t_combined_9_cast_fp16, y = linear_5_cast_fp16)[name = tensor("input_285_cast_fp16")]; + tensor input_287_cast_fp16 = silu(x = input_285_cast_fp16)[name = tensor("input_287_cast_fp16")]; + tensor linear_110_cast_fp16 = linear(bias = flow_net_res_blocks_0_adaLN_modulation_1_bias_to_fp16, weight = flow_net_res_blocks_0_adaLN_modulation_1_weight_to_fp16, x = input_287_cast_fp16)[name = tensor("linear_110_cast_fp16")]; + tensor var_1714_split_sizes_0 = const()[name = tensor("op_1714_split_sizes_0"), val = tensor([512, 512, 512])]; + tensor var_1714_axis_0 = const()[name = tensor("op_1714_axis_0"), val = tensor(-1)]; + tensor var_1714_cast_fp16_0, tensor var_1714_cast_fp16_1, tensor var_1714_cast_fp16_2 = split(axis = var_1714_axis_0, split_sizes = var_1714_split_sizes_0, x = linear_110_cast_fp16)[name = tensor("op_1714_cast_fp16")]; + tensor mean_57_axes_0 = const()[name = tensor("mean_57_axes_0"), val = tensor([-1])]; + tensor mean_57_keep_dims_0 = const()[name = tensor("mean_57_keep_dims_0"), val = tensor(true)]; + tensor mean_57_cast_fp16 = reduce_mean(axes = mean_57_axes_0, keep_dims = mean_57_keep_dims_0, x = linear_104_cast_fp16)[name = tensor("mean_57_cast_fp16")]; + tensor sub_48_cast_fp16 = sub(x = linear_104_cast_fp16, y = mean_57_cast_fp16)[name = tensor("sub_48_cast_fp16")]; + tensor square_38_cast_fp16 = square(x = sub_48_cast_fp16)[name = tensor("square_38_cast_fp16")]; + tensor reduce_mean_77_axes_0 = const()[name = tensor("reduce_mean_77_axes_0"), val = tensor([-1])]; + tensor reduce_mean_77_keep_dims_0 = const()[name = tensor("reduce_mean_77_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_77_cast_fp16 = reduce_mean(axes = reduce_mean_77_axes_0, keep_dims = reduce_mean_77_keep_dims_0, x = square_38_cast_fp16)[name = tensor("reduce_mean_77_cast_fp16")]; + tensor var_1724_to_fp16 = const()[name = tensor("op_1724_to_fp16"), val = tensor(0x1.1p-20)]; + tensor var_1725_cast_fp16 = add(x = reduce_mean_77_cast_fp16, y = var_1724_to_fp16)[name = tensor("op_1725_cast_fp16")]; + tensor var_1726_cast_fp16 = sqrt(x = var_1725_cast_fp16)[name = tensor("op_1726_cast_fp16")]; + tensor x_183_cast_fp16 = real_div(x = sub_48_cast_fp16, y = var_1726_cast_fp16)[name = tensor("x_183_cast_fp16")]; + tensor var_1728_cast_fp16 = mul(x = x_183_cast_fp16, y = flow_net_res_blocks_0_in_ln_weight_to_fp16)[name = tensor("op_1728_cast_fp16")]; + tensor x_185_cast_fp16 = add(x = var_1728_cast_fp16, y = flow_net_res_blocks_0_in_ln_bias_to_fp16)[name = tensor("x_185_cast_fp16")]; + tensor var_1730_promoted_to_fp16 = const()[name = tensor("op_1730_promoted_to_fp16"), val = tensor(0x1p+0)]; + tensor var_1731_cast_fp16 = add(x = var_1714_cast_fp16_1, y = var_1730_promoted_to_fp16)[name = tensor("op_1731_cast_fp16")]; + tensor var_1732_cast_fp16 = mul(x = x_185_cast_fp16, y = var_1731_cast_fp16)[name = tensor("op_1732_cast_fp16")]; + tensor input_289_cast_fp16 = add(x = var_1732_cast_fp16, y = var_1714_cast_fp16_0)[name = tensor("input_289_cast_fp16")]; + tensor linear_111_cast_fp16 = linear(bias = flow_net_res_blocks_0_mlp_0_bias_to_fp16, weight = flow_net_res_blocks_0_mlp_0_weight_to_fp16, x = input_289_cast_fp16)[name = tensor("linear_111_cast_fp16")]; + tensor input_293_cast_fp16 = silu(x = linear_111_cast_fp16)[name = tensor("input_293_cast_fp16")]; + tensor linear_112_cast_fp16 = linear(bias = flow_net_res_blocks_0_mlp_2_bias_to_fp16, weight = flow_net_res_blocks_0_mlp_2_weight_to_fp16, x = input_293_cast_fp16)[name = tensor("linear_112_cast_fp16")]; + tensor var_1743_cast_fp16 = mul(x = var_1714_cast_fp16_2, y = linear_112_cast_fp16)[name = tensor("op_1743_cast_fp16")]; + tensor x_187_cast_fp16 = add(x = linear_104_cast_fp16, y = var_1743_cast_fp16)[name = tensor("x_187_cast_fp16")]; + tensor linear_113_cast_fp16 = linear(bias = flow_net_res_blocks_1_adaLN_modulation_1_bias_to_fp16, weight = flow_net_res_blocks_1_adaLN_modulation_1_weight_to_fp16, x = input_287_cast_fp16)[name = tensor("linear_113_cast_fp16")]; + tensor var_1753_split_sizes_0 = const()[name = tensor("op_1753_split_sizes_0"), val = tensor([512, 512, 512])]; + tensor var_1753_axis_0 = const()[name = tensor("op_1753_axis_0"), val = tensor(-1)]; + tensor var_1753_cast_fp16_0, tensor var_1753_cast_fp16_1, tensor var_1753_cast_fp16_2 = split(axis = var_1753_axis_0, split_sizes = var_1753_split_sizes_0, x = linear_113_cast_fp16)[name = tensor("op_1753_cast_fp16")]; + tensor mean_59_axes_0 = const()[name = tensor("mean_59_axes_0"), val = tensor([-1])]; + tensor mean_59_keep_dims_0 = const()[name = tensor("mean_59_keep_dims_0"), val = tensor(true)]; + tensor mean_59_cast_fp16 = reduce_mean(axes = mean_59_axes_0, keep_dims = mean_59_keep_dims_0, x = x_187_cast_fp16)[name = tensor("mean_59_cast_fp16")]; + tensor sub_49_cast_fp16 = sub(x = x_187_cast_fp16, y = mean_59_cast_fp16)[name = tensor("sub_49_cast_fp16")]; + tensor square_39_cast_fp16 = square(x = sub_49_cast_fp16)[name = tensor("square_39_cast_fp16")]; + tensor reduce_mean_79_axes_0 = const()[name = tensor("reduce_mean_79_axes_0"), val = tensor([-1])]; + tensor reduce_mean_79_keep_dims_0 = const()[name = tensor("reduce_mean_79_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_79_cast_fp16 = reduce_mean(axes = reduce_mean_79_axes_0, keep_dims = reduce_mean_79_keep_dims_0, x = square_39_cast_fp16)[name = tensor("reduce_mean_79_cast_fp16")]; + tensor var_1763_to_fp16 = const()[name = tensor("op_1763_to_fp16"), val = tensor(0x1.1p-20)]; + tensor var_1764_cast_fp16 = add(x = reduce_mean_79_cast_fp16, y = var_1763_to_fp16)[name = tensor("op_1764_cast_fp16")]; + tensor var_1765_cast_fp16 = sqrt(x = var_1764_cast_fp16)[name = tensor("op_1765_cast_fp16")]; + tensor x_189_cast_fp16 = real_div(x = sub_49_cast_fp16, y = var_1765_cast_fp16)[name = tensor("x_189_cast_fp16")]; + tensor var_1767_cast_fp16 = mul(x = x_189_cast_fp16, y = flow_net_res_blocks_1_in_ln_weight_to_fp16)[name = tensor("op_1767_cast_fp16")]; + tensor x_191_cast_fp16 = add(x = var_1767_cast_fp16, y = flow_net_res_blocks_1_in_ln_bias_to_fp16)[name = tensor("x_191_cast_fp16")]; + tensor var_1769_promoted_to_fp16 = const()[name = tensor("op_1769_promoted_to_fp16"), val = tensor(0x1p+0)]; + tensor var_1770_cast_fp16 = add(x = var_1753_cast_fp16_1, y = var_1769_promoted_to_fp16)[name = tensor("op_1770_cast_fp16")]; + tensor var_1771_cast_fp16 = mul(x = x_191_cast_fp16, y = var_1770_cast_fp16)[name = tensor("op_1771_cast_fp16")]; + tensor input_297_cast_fp16 = add(x = var_1771_cast_fp16, y = var_1753_cast_fp16_0)[name = tensor("input_297_cast_fp16")]; + tensor linear_114_cast_fp16 = linear(bias = flow_net_res_blocks_1_mlp_0_bias_to_fp16, weight = flow_net_res_blocks_1_mlp_0_weight_to_fp16, x = input_297_cast_fp16)[name = tensor("linear_114_cast_fp16")]; + tensor input_301_cast_fp16 = silu(x = linear_114_cast_fp16)[name = tensor("input_301_cast_fp16")]; + tensor linear_115_cast_fp16 = linear(bias = flow_net_res_blocks_1_mlp_2_bias_to_fp16, weight = flow_net_res_blocks_1_mlp_2_weight_to_fp16, x = input_301_cast_fp16)[name = tensor("linear_115_cast_fp16")]; + tensor var_1782_cast_fp16 = mul(x = var_1753_cast_fp16_2, y = linear_115_cast_fp16)[name = tensor("op_1782_cast_fp16")]; + tensor x_193_cast_fp16 = add(x = x_187_cast_fp16, y = var_1782_cast_fp16)[name = tensor("x_193_cast_fp16")]; + tensor linear_116_cast_fp16 = linear(bias = flow_net_res_blocks_2_adaLN_modulation_1_bias_to_fp16, weight = flow_net_res_blocks_2_adaLN_modulation_1_weight_to_fp16, x = input_287_cast_fp16)[name = tensor("linear_116_cast_fp16")]; + tensor var_1792_split_sizes_0 = const()[name = tensor("op_1792_split_sizes_0"), val = tensor([512, 512, 512])]; + tensor var_1792_axis_0 = const()[name = tensor("op_1792_axis_0"), val = tensor(-1)]; + tensor var_1792_cast_fp16_0, tensor var_1792_cast_fp16_1, tensor var_1792_cast_fp16_2 = split(axis = var_1792_axis_0, split_sizes = var_1792_split_sizes_0, x = linear_116_cast_fp16)[name = tensor("op_1792_cast_fp16")]; + tensor mean_61_axes_0 = const()[name = tensor("mean_61_axes_0"), val = tensor([-1])]; + tensor mean_61_keep_dims_0 = const()[name = tensor("mean_61_keep_dims_0"), val = tensor(true)]; + tensor mean_61_cast_fp16 = reduce_mean(axes = mean_61_axes_0, keep_dims = mean_61_keep_dims_0, x = x_193_cast_fp16)[name = tensor("mean_61_cast_fp16")]; + tensor sub_50_cast_fp16 = sub(x = x_193_cast_fp16, y = mean_61_cast_fp16)[name = tensor("sub_50_cast_fp16")]; + tensor square_40_cast_fp16 = square(x = sub_50_cast_fp16)[name = tensor("square_40_cast_fp16")]; + tensor reduce_mean_81_axes_0 = const()[name = tensor("reduce_mean_81_axes_0"), val = tensor([-1])]; + tensor reduce_mean_81_keep_dims_0 = const()[name = tensor("reduce_mean_81_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_81_cast_fp16 = reduce_mean(axes = reduce_mean_81_axes_0, keep_dims = reduce_mean_81_keep_dims_0, x = square_40_cast_fp16)[name = tensor("reduce_mean_81_cast_fp16")]; + tensor var_1802_to_fp16 = const()[name = tensor("op_1802_to_fp16"), val = tensor(0x1.1p-20)]; + tensor var_1803_cast_fp16 = add(x = reduce_mean_81_cast_fp16, y = var_1802_to_fp16)[name = tensor("op_1803_cast_fp16")]; + tensor var_1804_cast_fp16 = sqrt(x = var_1803_cast_fp16)[name = tensor("op_1804_cast_fp16")]; + tensor x_195_cast_fp16 = real_div(x = sub_50_cast_fp16, y = var_1804_cast_fp16)[name = tensor("x_195_cast_fp16")]; + tensor var_1806_cast_fp16 = mul(x = x_195_cast_fp16, y = flow_net_res_blocks_2_in_ln_weight_to_fp16)[name = tensor("op_1806_cast_fp16")]; + tensor x_197_cast_fp16 = add(x = var_1806_cast_fp16, y = flow_net_res_blocks_2_in_ln_bias_to_fp16)[name = tensor("x_197_cast_fp16")]; + tensor var_1808_promoted_to_fp16 = const()[name = tensor("op_1808_promoted_to_fp16"), val = tensor(0x1p+0)]; + tensor var_1809_cast_fp16 = add(x = var_1792_cast_fp16_1, y = var_1808_promoted_to_fp16)[name = tensor("op_1809_cast_fp16")]; + tensor var_1810_cast_fp16 = mul(x = x_197_cast_fp16, y = var_1809_cast_fp16)[name = tensor("op_1810_cast_fp16")]; + tensor input_305_cast_fp16 = add(x = var_1810_cast_fp16, y = var_1792_cast_fp16_0)[name = tensor("input_305_cast_fp16")]; + tensor linear_117_cast_fp16 = linear(bias = flow_net_res_blocks_2_mlp_0_bias_to_fp16, weight = flow_net_res_blocks_2_mlp_0_weight_to_fp16, x = input_305_cast_fp16)[name = tensor("linear_117_cast_fp16")]; + tensor input_309_cast_fp16 = silu(x = linear_117_cast_fp16)[name = tensor("input_309_cast_fp16")]; + tensor linear_118_cast_fp16 = linear(bias = flow_net_res_blocks_2_mlp_2_bias_to_fp16, weight = flow_net_res_blocks_2_mlp_2_weight_to_fp16, x = input_309_cast_fp16)[name = tensor("linear_118_cast_fp16")]; + tensor var_1821_cast_fp16 = mul(x = var_1792_cast_fp16_2, y = linear_118_cast_fp16)[name = tensor("op_1821_cast_fp16")]; + tensor x_199_cast_fp16 = add(x = x_193_cast_fp16, y = var_1821_cast_fp16)[name = tensor("x_199_cast_fp16")]; + tensor linear_119_cast_fp16 = linear(bias = flow_net_res_blocks_3_adaLN_modulation_1_bias_to_fp16, weight = flow_net_res_blocks_3_adaLN_modulation_1_weight_to_fp16, x = input_287_cast_fp16)[name = tensor("linear_119_cast_fp16")]; + tensor var_1831_split_sizes_0 = const()[name = tensor("op_1831_split_sizes_0"), val = tensor([512, 512, 512])]; + tensor var_1831_axis_0 = const()[name = tensor("op_1831_axis_0"), val = tensor(-1)]; + tensor var_1831_cast_fp16_0, tensor var_1831_cast_fp16_1, tensor var_1831_cast_fp16_2 = split(axis = var_1831_axis_0, split_sizes = var_1831_split_sizes_0, x = linear_119_cast_fp16)[name = tensor("op_1831_cast_fp16")]; + tensor mean_63_axes_0 = const()[name = tensor("mean_63_axes_0"), val = tensor([-1])]; + tensor mean_63_keep_dims_0 = const()[name = tensor("mean_63_keep_dims_0"), val = tensor(true)]; + tensor mean_63_cast_fp16 = reduce_mean(axes = mean_63_axes_0, keep_dims = mean_63_keep_dims_0, x = x_199_cast_fp16)[name = tensor("mean_63_cast_fp16")]; + tensor sub_51_cast_fp16 = sub(x = x_199_cast_fp16, y = mean_63_cast_fp16)[name = tensor("sub_51_cast_fp16")]; + tensor square_41_cast_fp16 = square(x = sub_51_cast_fp16)[name = tensor("square_41_cast_fp16")]; + tensor reduce_mean_83_axes_0 = const()[name = tensor("reduce_mean_83_axes_0"), val = tensor([-1])]; + tensor reduce_mean_83_keep_dims_0 = const()[name = tensor("reduce_mean_83_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_83_cast_fp16 = reduce_mean(axes = reduce_mean_83_axes_0, keep_dims = reduce_mean_83_keep_dims_0, x = square_41_cast_fp16)[name = tensor("reduce_mean_83_cast_fp16")]; + tensor var_1841_to_fp16 = const()[name = tensor("op_1841_to_fp16"), val = tensor(0x1.1p-20)]; + tensor var_1842_cast_fp16 = add(x = reduce_mean_83_cast_fp16, y = var_1841_to_fp16)[name = tensor("op_1842_cast_fp16")]; + tensor var_1843_cast_fp16 = sqrt(x = var_1842_cast_fp16)[name = tensor("op_1843_cast_fp16")]; + tensor x_201_cast_fp16 = real_div(x = sub_51_cast_fp16, y = var_1843_cast_fp16)[name = tensor("x_201_cast_fp16")]; + tensor var_1845_cast_fp16 = mul(x = x_201_cast_fp16, y = flow_net_res_blocks_3_in_ln_weight_to_fp16)[name = tensor("op_1845_cast_fp16")]; + tensor x_203_cast_fp16 = add(x = var_1845_cast_fp16, y = flow_net_res_blocks_3_in_ln_bias_to_fp16)[name = tensor("x_203_cast_fp16")]; + tensor var_1847_promoted_to_fp16 = const()[name = tensor("op_1847_promoted_to_fp16"), val = tensor(0x1p+0)]; + tensor var_1848_cast_fp16 = add(x = var_1831_cast_fp16_1, y = var_1847_promoted_to_fp16)[name = tensor("op_1848_cast_fp16")]; + tensor var_1849_cast_fp16 = mul(x = x_203_cast_fp16, y = var_1848_cast_fp16)[name = tensor("op_1849_cast_fp16")]; + tensor input_313_cast_fp16 = add(x = var_1849_cast_fp16, y = var_1831_cast_fp16_0)[name = tensor("input_313_cast_fp16")]; + tensor linear_120_cast_fp16 = linear(bias = flow_net_res_blocks_3_mlp_0_bias_to_fp16, weight = flow_net_res_blocks_3_mlp_0_weight_to_fp16, x = input_313_cast_fp16)[name = tensor("linear_120_cast_fp16")]; + tensor input_317_cast_fp16 = silu(x = linear_120_cast_fp16)[name = tensor("input_317_cast_fp16")]; + tensor linear_121_cast_fp16 = linear(bias = flow_net_res_blocks_3_mlp_2_bias_to_fp16, weight = flow_net_res_blocks_3_mlp_2_weight_to_fp16, x = input_317_cast_fp16)[name = tensor("linear_121_cast_fp16")]; + tensor var_1860_cast_fp16 = mul(x = var_1831_cast_fp16_2, y = linear_121_cast_fp16)[name = tensor("op_1860_cast_fp16")]; + tensor x_205_cast_fp16 = add(x = x_199_cast_fp16, y = var_1860_cast_fp16)[name = tensor("x_205_cast_fp16")]; + tensor linear_122_cast_fp16 = linear(bias = flow_net_res_blocks_4_adaLN_modulation_1_bias_to_fp16, weight = flow_net_res_blocks_4_adaLN_modulation_1_weight_to_fp16, x = input_287_cast_fp16)[name = tensor("linear_122_cast_fp16")]; + tensor var_1870_split_sizes_0 = const()[name = tensor("op_1870_split_sizes_0"), val = tensor([512, 512, 512])]; + tensor var_1870_axis_0 = const()[name = tensor("op_1870_axis_0"), val = tensor(-1)]; + tensor var_1870_cast_fp16_0, tensor var_1870_cast_fp16_1, tensor var_1870_cast_fp16_2 = split(axis = var_1870_axis_0, split_sizes = var_1870_split_sizes_0, x = linear_122_cast_fp16)[name = tensor("op_1870_cast_fp16")]; + tensor mean_65_axes_0 = const()[name = tensor("mean_65_axes_0"), val = tensor([-1])]; + tensor mean_65_keep_dims_0 = const()[name = tensor("mean_65_keep_dims_0"), val = tensor(true)]; + tensor mean_65_cast_fp16 = reduce_mean(axes = mean_65_axes_0, keep_dims = mean_65_keep_dims_0, x = x_205_cast_fp16)[name = tensor("mean_65_cast_fp16")]; + tensor sub_52_cast_fp16 = sub(x = x_205_cast_fp16, y = mean_65_cast_fp16)[name = tensor("sub_52_cast_fp16")]; + tensor square_42_cast_fp16 = square(x = sub_52_cast_fp16)[name = tensor("square_42_cast_fp16")]; + tensor reduce_mean_85_axes_0 = const()[name = tensor("reduce_mean_85_axes_0"), val = tensor([-1])]; + tensor reduce_mean_85_keep_dims_0 = const()[name = tensor("reduce_mean_85_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_85_cast_fp16 = reduce_mean(axes = reduce_mean_85_axes_0, keep_dims = reduce_mean_85_keep_dims_0, x = square_42_cast_fp16)[name = tensor("reduce_mean_85_cast_fp16")]; + tensor var_1880_to_fp16 = const()[name = tensor("op_1880_to_fp16"), val = tensor(0x1.1p-20)]; + tensor var_1881_cast_fp16 = add(x = reduce_mean_85_cast_fp16, y = var_1880_to_fp16)[name = tensor("op_1881_cast_fp16")]; + tensor var_1882_cast_fp16 = sqrt(x = var_1881_cast_fp16)[name = tensor("op_1882_cast_fp16")]; + tensor x_207_cast_fp16 = real_div(x = sub_52_cast_fp16, y = var_1882_cast_fp16)[name = tensor("x_207_cast_fp16")]; + tensor var_1884_cast_fp16 = mul(x = x_207_cast_fp16, y = flow_net_res_blocks_4_in_ln_weight_to_fp16)[name = tensor("op_1884_cast_fp16")]; + tensor x_209_cast_fp16 = add(x = var_1884_cast_fp16, y = flow_net_res_blocks_4_in_ln_bias_to_fp16)[name = tensor("x_209_cast_fp16")]; + tensor var_1886_promoted_to_fp16 = const()[name = tensor("op_1886_promoted_to_fp16"), val = tensor(0x1p+0)]; + tensor var_1887_cast_fp16 = add(x = var_1870_cast_fp16_1, y = var_1886_promoted_to_fp16)[name = tensor("op_1887_cast_fp16")]; + tensor var_1888_cast_fp16 = mul(x = x_209_cast_fp16, y = var_1887_cast_fp16)[name = tensor("op_1888_cast_fp16")]; + tensor input_321_cast_fp16 = add(x = var_1888_cast_fp16, y = var_1870_cast_fp16_0)[name = tensor("input_321_cast_fp16")]; + tensor linear_123_cast_fp16 = linear(bias = flow_net_res_blocks_4_mlp_0_bias_to_fp16, weight = flow_net_res_blocks_4_mlp_0_weight_to_fp16, x = input_321_cast_fp16)[name = tensor("linear_123_cast_fp16")]; + tensor input_325_cast_fp16 = silu(x = linear_123_cast_fp16)[name = tensor("input_325_cast_fp16")]; + tensor linear_124_cast_fp16 = linear(bias = flow_net_res_blocks_4_mlp_2_bias_to_fp16, weight = flow_net_res_blocks_4_mlp_2_weight_to_fp16, x = input_325_cast_fp16)[name = tensor("linear_124_cast_fp16")]; + tensor var_1899_cast_fp16 = mul(x = var_1870_cast_fp16_2, y = linear_124_cast_fp16)[name = tensor("op_1899_cast_fp16")]; + tensor x_211_cast_fp16 = add(x = x_205_cast_fp16, y = var_1899_cast_fp16)[name = tensor("x_211_cast_fp16")]; + tensor linear_125_cast_fp16 = linear(bias = flow_net_res_blocks_5_adaLN_modulation_1_bias_to_fp16, weight = flow_net_res_blocks_5_adaLN_modulation_1_weight_to_fp16, x = input_287_cast_fp16)[name = tensor("linear_125_cast_fp16")]; + tensor var_1909_split_sizes_0 = const()[name = tensor("op_1909_split_sizes_0"), val = tensor([512, 512, 512])]; + tensor var_1909_axis_0 = const()[name = tensor("op_1909_axis_0"), val = tensor(-1)]; + tensor var_1909_cast_fp16_0, tensor var_1909_cast_fp16_1, tensor var_1909_cast_fp16_2 = split(axis = var_1909_axis_0, split_sizes = var_1909_split_sizes_0, x = linear_125_cast_fp16)[name = tensor("op_1909_cast_fp16")]; + tensor mean_67_axes_0 = const()[name = tensor("mean_67_axes_0"), val = tensor([-1])]; + tensor mean_67_keep_dims_0 = const()[name = tensor("mean_67_keep_dims_0"), val = tensor(true)]; + tensor mean_67_cast_fp16 = reduce_mean(axes = mean_67_axes_0, keep_dims = mean_67_keep_dims_0, x = x_211_cast_fp16)[name = tensor("mean_67_cast_fp16")]; + tensor sub_53_cast_fp16 = sub(x = x_211_cast_fp16, y = mean_67_cast_fp16)[name = tensor("sub_53_cast_fp16")]; + tensor square_43_cast_fp16 = square(x = sub_53_cast_fp16)[name = tensor("square_43_cast_fp16")]; + tensor reduce_mean_87_axes_0 = const()[name = tensor("reduce_mean_87_axes_0"), val = tensor([-1])]; + tensor reduce_mean_87_keep_dims_0 = const()[name = tensor("reduce_mean_87_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_87_cast_fp16 = reduce_mean(axes = reduce_mean_87_axes_0, keep_dims = reduce_mean_87_keep_dims_0, x = square_43_cast_fp16)[name = tensor("reduce_mean_87_cast_fp16")]; + tensor var_1919_to_fp16 = const()[name = tensor("op_1919_to_fp16"), val = tensor(0x1.1p-20)]; + tensor var_1920_cast_fp16 = add(x = reduce_mean_87_cast_fp16, y = var_1919_to_fp16)[name = tensor("op_1920_cast_fp16")]; + tensor var_1921_cast_fp16 = sqrt(x = var_1920_cast_fp16)[name = tensor("op_1921_cast_fp16")]; + tensor x_213_cast_fp16 = real_div(x = sub_53_cast_fp16, y = var_1921_cast_fp16)[name = tensor("x_213_cast_fp16")]; + tensor var_1923_cast_fp16 = mul(x = x_213_cast_fp16, y = flow_net_res_blocks_5_in_ln_weight_to_fp16)[name = tensor("op_1923_cast_fp16")]; + tensor x_215_cast_fp16 = add(x = var_1923_cast_fp16, y = flow_net_res_blocks_5_in_ln_bias_to_fp16)[name = tensor("x_215_cast_fp16")]; + tensor var_1925_promoted_to_fp16 = const()[name = tensor("op_1925_promoted_to_fp16"), val = tensor(0x1p+0)]; + tensor var_1926_cast_fp16 = add(x = var_1909_cast_fp16_1, y = var_1925_promoted_to_fp16)[name = tensor("op_1926_cast_fp16")]; + tensor var_1927_cast_fp16 = mul(x = x_215_cast_fp16, y = var_1926_cast_fp16)[name = tensor("op_1927_cast_fp16")]; + tensor input_329_cast_fp16 = add(x = var_1927_cast_fp16, y = var_1909_cast_fp16_0)[name = tensor("input_329_cast_fp16")]; + tensor linear_126_cast_fp16 = linear(bias = flow_net_res_blocks_5_mlp_0_bias_to_fp16, weight = flow_net_res_blocks_5_mlp_0_weight_to_fp16, x = input_329_cast_fp16)[name = tensor("linear_126_cast_fp16")]; + tensor input_333_cast_fp16 = silu(x = linear_126_cast_fp16)[name = tensor("input_333_cast_fp16")]; + tensor linear_127_cast_fp16 = linear(bias = flow_net_res_blocks_5_mlp_2_bias_to_fp16, weight = flow_net_res_blocks_5_mlp_2_weight_to_fp16, x = input_333_cast_fp16)[name = tensor("linear_127_cast_fp16")]; + tensor var_1938_cast_fp16 = mul(x = var_1909_cast_fp16_2, y = linear_127_cast_fp16)[name = tensor("op_1938_cast_fp16")]; + tensor x_217_cast_fp16 = add(x = x_211_cast_fp16, y = var_1938_cast_fp16)[name = tensor("x_217_cast_fp16")]; + tensor linear_128_cast_fp16 = linear(bias = flow_net_final_layer_adaLN_modulation_1_bias_to_fp16, weight = flow_net_final_layer_adaLN_modulation_1_weight_to_fp16, x = input_287_cast_fp16)[name = tensor("linear_128_cast_fp16")]; + tensor var_1947_split_sizes_0 = const()[name = tensor("op_1947_split_sizes_0"), val = tensor([512, 512])]; + tensor var_1947_axis_0 = const()[name = tensor("op_1947_axis_0"), val = tensor(-1)]; + tensor var_1947_cast_fp16_0, tensor var_1947_cast_fp16_1 = split(axis = var_1947_axis_0, split_sizes = var_1947_split_sizes_0, x = linear_128_cast_fp16)[name = tensor("op_1947_cast_fp16")]; + tensor mean_69_axes_0 = const()[name = tensor("mean_69_axes_0"), val = tensor([-1])]; + tensor mean_69_keep_dims_0 = const()[name = tensor("mean_69_keep_dims_0"), val = tensor(true)]; + tensor mean_69_cast_fp16 = reduce_mean(axes = mean_69_axes_0, keep_dims = mean_69_keep_dims_0, x = x_217_cast_fp16)[name = tensor("mean_69_cast_fp16")]; + tensor sub_54_cast_fp16 = sub(x = x_217_cast_fp16, y = mean_69_cast_fp16)[name = tensor("sub_54_cast_fp16")]; + tensor square_44_cast_fp16 = square(x = sub_54_cast_fp16)[name = tensor("square_44_cast_fp16")]; + tensor reduce_mean_89_axes_0 = const()[name = tensor("reduce_mean_89_axes_0"), val = tensor([-1])]; + tensor reduce_mean_89_keep_dims_0 = const()[name = tensor("reduce_mean_89_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_89_cast_fp16 = reduce_mean(axes = reduce_mean_89_axes_0, keep_dims = reduce_mean_89_keep_dims_0, x = square_44_cast_fp16)[name = tensor("reduce_mean_89_cast_fp16")]; + tensor var_1954_to_fp16 = const()[name = tensor("op_1954_to_fp16"), val = tensor(0x1.1p-20)]; + tensor var_1955_cast_fp16 = add(x = reduce_mean_89_cast_fp16, y = var_1954_to_fp16)[name = tensor("op_1955_cast_fp16")]; + tensor var_1956_cast_fp16 = sqrt(x = var_1955_cast_fp16)[name = tensor("op_1956_cast_fp16")]; + tensor x_219_cast_fp16 = real_div(x = sub_54_cast_fp16, y = var_1956_cast_fp16)[name = tensor("x_219_cast_fp16")]; + tensor var_1958_promoted_to_fp16 = const()[name = tensor("op_1958_promoted_to_fp16"), val = tensor(0x1p+0)]; + tensor var_1959_cast_fp16 = add(x = var_1947_cast_fp16_1, y = var_1958_promoted_to_fp16)[name = tensor("op_1959_cast_fp16")]; + tensor var_1960_cast_fp16 = mul(x = x_219_cast_fp16, y = var_1959_cast_fp16)[name = tensor("op_1960_cast_fp16")]; + tensor input_337_cast_fp16 = add(x = var_1960_cast_fp16, y = var_1947_cast_fp16_0)[name = tensor("input_337_cast_fp16")]; + tensor linear_129_cast_fp16 = linear(bias = flow_net_final_layer_linear_bias_to_fp16, weight = flow_net_final_layer_linear_weight_to_fp16, x = input_337_cast_fp16)[name = tensor("linear_129_cast_fp16")]; + tensor var_1971_to_fp16 = const()[name = tensor("op_1971_to_fp16"), val = tensor(0x1p-3)]; + tensor var_1972_cast_fp16 = mul(x = linear_129_cast_fp16, y = var_1971_to_fp16)[name = tensor("op_1972_cast_fp16")]; + tensor input_339_cast_fp16 = add(x = input_271_cast_fp16, y = var_1972_cast_fp16)[name = tensor("input_339_cast_fp16")]; + tensor linear_130_cast_fp16 = linear(bias = flow_net_input_proj_bias_to_fp16, weight = flow_net_input_proj_weight_to_fp16, x = input_339_cast_fp16)[name = tensor("linear_130_cast_fp16")]; + tensor input_343_to_fp16 = const()[name = tensor("input_343_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(19006080)))]; + tensor input_345_cast_fp16 = silu(x = input_343_to_fp16)[name = tensor("input_345_cast_fp16")]; + tensor linear_132_cast_fp16 = linear(bias = flow_net_time_embed_0_mlp_2_bias_to_fp16, weight = flow_net_time_embed_0_mlp_2_weight_to_fp16, x = input_345_cast_fp16)[name = tensor("linear_132_cast_fp16")]; + tensor reduce_mean_90_axes_0 = const()[name = tensor("reduce_mean_90_axes_0"), val = tensor([-1])]; + tensor reduce_mean_90_keep_dims_0 = const()[name = tensor("reduce_mean_90_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_90_cast_fp16 = reduce_mean(axes = reduce_mean_90_axes_0, keep_dims = reduce_mean_90_keep_dims_0, x = linear_132_cast_fp16)[name = tensor("reduce_mean_90_cast_fp16")]; + tensor sub_55_cast_fp16 = sub(x = linear_132_cast_fp16, y = reduce_mean_90_cast_fp16)[name = tensor("sub_55_cast_fp16")]; + tensor square_45_cast_fp16 = square(x = sub_55_cast_fp16)[name = tensor("square_45_cast_fp16")]; + tensor reduce_mean_91_axes_0 = const()[name = tensor("reduce_mean_91_axes_0"), val = tensor([-1])]; + tensor reduce_mean_91_keep_dims_0 = const()[name = tensor("reduce_mean_91_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_91_cast_fp16 = reduce_mean(axes = reduce_mean_91_axes_0, keep_dims = reduce_mean_91_keep_dims_0, x = square_45_cast_fp16)[name = tensor("reduce_mean_91_cast_fp16")]; + tensor real_div_10_to_fp16 = const()[name = tensor("real_div_10_to_fp16"), val = tensor(0x1.008p+0)]; + tensor mul_10_cast_fp16 = mul(x = reduce_mean_91_cast_fp16, y = real_div_10_to_fp16)[name = tensor("mul_10_cast_fp16")]; + tensor var_2038_to_fp16 = const()[name = tensor("op_2038_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_91_cast_fp16 = add(x = mul_10_cast_fp16, y = var_2038_to_fp16)[name = tensor("var_91_cast_fp16")]; + tensor var_2041_epsilon_0 = const()[name = tensor("op_2041_epsilon_0"), val = tensor(0x1.197998p-40)]; + tensor var_2041_cast_fp16 = rsqrt(epsilon = var_2041_epsilon_0, x = var_91_cast_fp16)[name = tensor("op_2041_cast_fp16")]; + tensor var_2042_cast_fp16 = mul(x = const_3_to_fp16, y = var_2041_cast_fp16)[name = tensor("op_2042_cast_fp16")]; + tensor var_2043_cast_fp16 = mul(x = linear_132_cast_fp16, y = var_2042_cast_fp16)[name = tensor("op_2043_cast_fp16")]; + tensor input_349_to_fp16 = const()[name = tensor("input_349_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(19007168)))]; + tensor input_351_cast_fp16 = silu(x = input_349_to_fp16)[name = tensor("input_351_cast_fp16")]; + tensor linear_134_cast_fp16 = linear(bias = flow_net_time_embed_1_mlp_2_bias_to_fp16, weight = flow_net_time_embed_1_mlp_2_weight_to_fp16, x = input_351_cast_fp16)[name = tensor("linear_134_cast_fp16")]; + tensor reduce_mean_92_axes_0 = const()[name = tensor("reduce_mean_92_axes_0"), val = tensor([-1])]; + tensor reduce_mean_92_keep_dims_0 = const()[name = tensor("reduce_mean_92_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_92_cast_fp16 = reduce_mean(axes = reduce_mean_92_axes_0, keep_dims = reduce_mean_92_keep_dims_0, x = linear_134_cast_fp16)[name = tensor("reduce_mean_92_cast_fp16")]; + tensor sub_57_cast_fp16 = sub(x = linear_134_cast_fp16, y = reduce_mean_92_cast_fp16)[name = tensor("sub_57_cast_fp16")]; + tensor square_46_cast_fp16 = square(x = sub_57_cast_fp16)[name = tensor("square_46_cast_fp16")]; + tensor reduce_mean_93_axes_0 = const()[name = tensor("reduce_mean_93_axes_0"), val = tensor([-1])]; + tensor reduce_mean_93_keep_dims_0 = const()[name = tensor("reduce_mean_93_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_93_cast_fp16 = reduce_mean(axes = reduce_mean_93_axes_0, keep_dims = reduce_mean_93_keep_dims_0, x = square_46_cast_fp16)[name = tensor("reduce_mean_93_cast_fp16")]; + tensor real_div_11_to_fp16 = const()[name = tensor("real_div_11_to_fp16"), val = tensor(0x1.008p+0)]; + tensor mul_11_cast_fp16 = mul(x = reduce_mean_93_cast_fp16, y = real_div_11_to_fp16)[name = tensor("mul_11_cast_fp16")]; + tensor var_2075_to_fp16 = const()[name = tensor("op_2075_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_93_cast_fp16 = add(x = mul_11_cast_fp16, y = var_2075_to_fp16)[name = tensor("var_93_cast_fp16")]; + tensor var_2078_epsilon_0 = const()[name = tensor("op_2078_epsilon_0"), val = tensor(0x1.197998p-40)]; + tensor var_2078_cast_fp16 = rsqrt(epsilon = var_2078_epsilon_0, x = var_93_cast_fp16)[name = tensor("op_2078_cast_fp16")]; + tensor var_2079_cast_fp16 = mul(x = const_5_to_fp16, y = var_2078_cast_fp16)[name = tensor("op_2079_cast_fp16")]; + tensor var_2080_cast_fp16 = mul(x = linear_134_cast_fp16, y = var_2079_cast_fp16)[name = tensor("op_2080_cast_fp16")]; + tensor var_2092_cast_fp16 = add(x = var_2043_cast_fp16, y = var_2080_cast_fp16)[name = tensor("op_2092_cast_fp16")]; + tensor _inversed_t_combined_11_y_0_to_fp16 = const()[name = tensor("_inversed_t_combined_11_y_0_to_fp16"), val = tensor(0x1p-1)]; + tensor _inversed_t_combined_11_cast_fp16 = mul(x = var_2092_cast_fp16, y = _inversed_t_combined_11_y_0_to_fp16)[name = tensor("_inversed_t_combined_11_cast_fp16")]; + tensor input_353_cast_fp16 = add(x = _inversed_t_combined_11_cast_fp16, y = linear_5_cast_fp16)[name = tensor("input_353_cast_fp16")]; + tensor input_355_cast_fp16 = silu(x = input_353_cast_fp16)[name = tensor("input_355_cast_fp16")]; + tensor linear_136_cast_fp16 = linear(bias = flow_net_res_blocks_0_adaLN_modulation_1_bias_to_fp16, weight = flow_net_res_blocks_0_adaLN_modulation_1_weight_to_fp16, x = input_355_cast_fp16)[name = tensor("linear_136_cast_fp16")]; + tensor var_2107_split_sizes_0 = const()[name = tensor("op_2107_split_sizes_0"), val = tensor([512, 512, 512])]; + tensor var_2107_axis_0 = const()[name = tensor("op_2107_axis_0"), val = tensor(-1)]; + tensor var_2107_cast_fp16_0, tensor var_2107_cast_fp16_1, tensor var_2107_cast_fp16_2 = split(axis = var_2107_axis_0, split_sizes = var_2107_split_sizes_0, x = linear_136_cast_fp16)[name = tensor("op_2107_cast_fp16")]; + tensor mean_71_axes_0 = const()[name = tensor("mean_71_axes_0"), val = tensor([-1])]; + tensor mean_71_keep_dims_0 = const()[name = tensor("mean_71_keep_dims_0"), val = tensor(true)]; + tensor mean_71_cast_fp16 = reduce_mean(axes = mean_71_axes_0, keep_dims = mean_71_keep_dims_0, x = linear_130_cast_fp16)[name = tensor("mean_71_cast_fp16")]; + tensor sub_59_cast_fp16 = sub(x = linear_130_cast_fp16, y = mean_71_cast_fp16)[name = tensor("sub_59_cast_fp16")]; + tensor square_47_cast_fp16 = square(x = sub_59_cast_fp16)[name = tensor("square_47_cast_fp16")]; + tensor reduce_mean_95_axes_0 = const()[name = tensor("reduce_mean_95_axes_0"), val = tensor([-1])]; + tensor reduce_mean_95_keep_dims_0 = const()[name = tensor("reduce_mean_95_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_95_cast_fp16 = reduce_mean(axes = reduce_mean_95_axes_0, keep_dims = reduce_mean_95_keep_dims_0, x = square_47_cast_fp16)[name = tensor("reduce_mean_95_cast_fp16")]; + tensor var_2117_to_fp16 = const()[name = tensor("op_2117_to_fp16"), val = tensor(0x1.1p-20)]; + tensor var_2118_cast_fp16 = add(x = reduce_mean_95_cast_fp16, y = var_2117_to_fp16)[name = tensor("op_2118_cast_fp16")]; + tensor var_2119_cast_fp16 = sqrt(x = var_2118_cast_fp16)[name = tensor("op_2119_cast_fp16")]; + tensor x_227_cast_fp16 = real_div(x = sub_59_cast_fp16, y = var_2119_cast_fp16)[name = tensor("x_227_cast_fp16")]; + tensor var_2121_cast_fp16 = mul(x = x_227_cast_fp16, y = flow_net_res_blocks_0_in_ln_weight_to_fp16)[name = tensor("op_2121_cast_fp16")]; + tensor x_229_cast_fp16 = add(x = var_2121_cast_fp16, y = flow_net_res_blocks_0_in_ln_bias_to_fp16)[name = tensor("x_229_cast_fp16")]; + tensor var_2123_promoted_to_fp16 = const()[name = tensor("op_2123_promoted_to_fp16"), val = tensor(0x1p+0)]; + tensor var_2124_cast_fp16 = add(x = var_2107_cast_fp16_1, y = var_2123_promoted_to_fp16)[name = tensor("op_2124_cast_fp16")]; + tensor var_2125_cast_fp16 = mul(x = x_229_cast_fp16, y = var_2124_cast_fp16)[name = tensor("op_2125_cast_fp16")]; + tensor input_357_cast_fp16 = add(x = var_2125_cast_fp16, y = var_2107_cast_fp16_0)[name = tensor("input_357_cast_fp16")]; + tensor linear_137_cast_fp16 = linear(bias = flow_net_res_blocks_0_mlp_0_bias_to_fp16, weight = flow_net_res_blocks_0_mlp_0_weight_to_fp16, x = input_357_cast_fp16)[name = tensor("linear_137_cast_fp16")]; + tensor input_361_cast_fp16 = silu(x = linear_137_cast_fp16)[name = tensor("input_361_cast_fp16")]; + tensor linear_138_cast_fp16 = linear(bias = flow_net_res_blocks_0_mlp_2_bias_to_fp16, weight = flow_net_res_blocks_0_mlp_2_weight_to_fp16, x = input_361_cast_fp16)[name = tensor("linear_138_cast_fp16")]; + tensor var_2136_cast_fp16 = mul(x = var_2107_cast_fp16_2, y = linear_138_cast_fp16)[name = tensor("op_2136_cast_fp16")]; + tensor x_231_cast_fp16 = add(x = linear_130_cast_fp16, y = var_2136_cast_fp16)[name = tensor("x_231_cast_fp16")]; + tensor linear_139_cast_fp16 = linear(bias = flow_net_res_blocks_1_adaLN_modulation_1_bias_to_fp16, weight = flow_net_res_blocks_1_adaLN_modulation_1_weight_to_fp16, x = input_355_cast_fp16)[name = tensor("linear_139_cast_fp16")]; + tensor var_2146_split_sizes_0 = const()[name = tensor("op_2146_split_sizes_0"), val = tensor([512, 512, 512])]; + tensor var_2146_axis_0 = const()[name = tensor("op_2146_axis_0"), val = tensor(-1)]; + tensor var_2146_cast_fp16_0, tensor var_2146_cast_fp16_1, tensor var_2146_cast_fp16_2 = split(axis = var_2146_axis_0, split_sizes = var_2146_split_sizes_0, x = linear_139_cast_fp16)[name = tensor("op_2146_cast_fp16")]; + tensor mean_73_axes_0 = const()[name = tensor("mean_73_axes_0"), val = tensor([-1])]; + tensor mean_73_keep_dims_0 = const()[name = tensor("mean_73_keep_dims_0"), val = tensor(true)]; + tensor mean_73_cast_fp16 = reduce_mean(axes = mean_73_axes_0, keep_dims = mean_73_keep_dims_0, x = x_231_cast_fp16)[name = tensor("mean_73_cast_fp16")]; + tensor sub_60_cast_fp16 = sub(x = x_231_cast_fp16, y = mean_73_cast_fp16)[name = tensor("sub_60_cast_fp16")]; + tensor square_48_cast_fp16 = square(x = sub_60_cast_fp16)[name = tensor("square_48_cast_fp16")]; + tensor reduce_mean_97_axes_0 = const()[name = tensor("reduce_mean_97_axes_0"), val = tensor([-1])]; + tensor reduce_mean_97_keep_dims_0 = const()[name = tensor("reduce_mean_97_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_97_cast_fp16 = reduce_mean(axes = reduce_mean_97_axes_0, keep_dims = reduce_mean_97_keep_dims_0, x = square_48_cast_fp16)[name = tensor("reduce_mean_97_cast_fp16")]; + tensor var_2156_to_fp16 = const()[name = tensor("op_2156_to_fp16"), val = tensor(0x1.1p-20)]; + tensor var_2157_cast_fp16 = add(x = reduce_mean_97_cast_fp16, y = var_2156_to_fp16)[name = tensor("op_2157_cast_fp16")]; + tensor var_2158_cast_fp16 = sqrt(x = var_2157_cast_fp16)[name = tensor("op_2158_cast_fp16")]; + tensor x_233_cast_fp16 = real_div(x = sub_60_cast_fp16, y = var_2158_cast_fp16)[name = tensor("x_233_cast_fp16")]; + tensor var_2160_cast_fp16 = mul(x = x_233_cast_fp16, y = flow_net_res_blocks_1_in_ln_weight_to_fp16)[name = tensor("op_2160_cast_fp16")]; + tensor x_235_cast_fp16 = add(x = var_2160_cast_fp16, y = flow_net_res_blocks_1_in_ln_bias_to_fp16)[name = tensor("x_235_cast_fp16")]; + tensor var_2162_promoted_to_fp16 = const()[name = tensor("op_2162_promoted_to_fp16"), val = tensor(0x1p+0)]; + tensor var_2163_cast_fp16 = add(x = var_2146_cast_fp16_1, y = var_2162_promoted_to_fp16)[name = tensor("op_2163_cast_fp16")]; + tensor var_2164_cast_fp16 = mul(x = x_235_cast_fp16, y = var_2163_cast_fp16)[name = tensor("op_2164_cast_fp16")]; + tensor input_365_cast_fp16 = add(x = var_2164_cast_fp16, y = var_2146_cast_fp16_0)[name = tensor("input_365_cast_fp16")]; + tensor linear_140_cast_fp16 = linear(bias = flow_net_res_blocks_1_mlp_0_bias_to_fp16, weight = flow_net_res_blocks_1_mlp_0_weight_to_fp16, x = input_365_cast_fp16)[name = tensor("linear_140_cast_fp16")]; + tensor input_369_cast_fp16 = silu(x = linear_140_cast_fp16)[name = tensor("input_369_cast_fp16")]; + tensor linear_141_cast_fp16 = linear(bias = flow_net_res_blocks_1_mlp_2_bias_to_fp16, weight = flow_net_res_blocks_1_mlp_2_weight_to_fp16, x = input_369_cast_fp16)[name = tensor("linear_141_cast_fp16")]; + tensor var_2175_cast_fp16 = mul(x = var_2146_cast_fp16_2, y = linear_141_cast_fp16)[name = tensor("op_2175_cast_fp16")]; + tensor x_237_cast_fp16 = add(x = x_231_cast_fp16, y = var_2175_cast_fp16)[name = tensor("x_237_cast_fp16")]; + tensor linear_142_cast_fp16 = linear(bias = flow_net_res_blocks_2_adaLN_modulation_1_bias_to_fp16, weight = flow_net_res_blocks_2_adaLN_modulation_1_weight_to_fp16, x = input_355_cast_fp16)[name = tensor("linear_142_cast_fp16")]; + tensor var_2185_split_sizes_0 = const()[name = tensor("op_2185_split_sizes_0"), val = tensor([512, 512, 512])]; + tensor var_2185_axis_0 = const()[name = tensor("op_2185_axis_0"), val = tensor(-1)]; + tensor var_2185_cast_fp16_0, tensor var_2185_cast_fp16_1, tensor var_2185_cast_fp16_2 = split(axis = var_2185_axis_0, split_sizes = var_2185_split_sizes_0, x = linear_142_cast_fp16)[name = tensor("op_2185_cast_fp16")]; + tensor mean_75_axes_0 = const()[name = tensor("mean_75_axes_0"), val = tensor([-1])]; + tensor mean_75_keep_dims_0 = const()[name = tensor("mean_75_keep_dims_0"), val = tensor(true)]; + tensor mean_75_cast_fp16 = reduce_mean(axes = mean_75_axes_0, keep_dims = mean_75_keep_dims_0, x = x_237_cast_fp16)[name = tensor("mean_75_cast_fp16")]; + tensor sub_61_cast_fp16 = sub(x = x_237_cast_fp16, y = mean_75_cast_fp16)[name = tensor("sub_61_cast_fp16")]; + tensor square_49_cast_fp16 = square(x = sub_61_cast_fp16)[name = tensor("square_49_cast_fp16")]; + tensor reduce_mean_99_axes_0 = const()[name = tensor("reduce_mean_99_axes_0"), val = tensor([-1])]; + tensor reduce_mean_99_keep_dims_0 = const()[name = tensor("reduce_mean_99_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_99_cast_fp16 = reduce_mean(axes = reduce_mean_99_axes_0, keep_dims = reduce_mean_99_keep_dims_0, x = square_49_cast_fp16)[name = tensor("reduce_mean_99_cast_fp16")]; + tensor var_2195_to_fp16 = const()[name = tensor("op_2195_to_fp16"), val = tensor(0x1.1p-20)]; + tensor var_2196_cast_fp16 = add(x = reduce_mean_99_cast_fp16, y = var_2195_to_fp16)[name = tensor("op_2196_cast_fp16")]; + tensor var_2197_cast_fp16 = sqrt(x = var_2196_cast_fp16)[name = tensor("op_2197_cast_fp16")]; + tensor x_239_cast_fp16 = real_div(x = sub_61_cast_fp16, y = var_2197_cast_fp16)[name = tensor("x_239_cast_fp16")]; + tensor var_2199_cast_fp16 = mul(x = x_239_cast_fp16, y = flow_net_res_blocks_2_in_ln_weight_to_fp16)[name = tensor("op_2199_cast_fp16")]; + tensor x_241_cast_fp16 = add(x = var_2199_cast_fp16, y = flow_net_res_blocks_2_in_ln_bias_to_fp16)[name = tensor("x_241_cast_fp16")]; + tensor var_2201_promoted_to_fp16 = const()[name = tensor("op_2201_promoted_to_fp16"), val = tensor(0x1p+0)]; + tensor var_2202_cast_fp16 = add(x = var_2185_cast_fp16_1, y = var_2201_promoted_to_fp16)[name = tensor("op_2202_cast_fp16")]; + tensor var_2203_cast_fp16 = mul(x = x_241_cast_fp16, y = var_2202_cast_fp16)[name = tensor("op_2203_cast_fp16")]; + tensor input_373_cast_fp16 = add(x = var_2203_cast_fp16, y = var_2185_cast_fp16_0)[name = tensor("input_373_cast_fp16")]; + tensor linear_143_cast_fp16 = linear(bias = flow_net_res_blocks_2_mlp_0_bias_to_fp16, weight = flow_net_res_blocks_2_mlp_0_weight_to_fp16, x = input_373_cast_fp16)[name = tensor("linear_143_cast_fp16")]; + tensor input_377_cast_fp16 = silu(x = linear_143_cast_fp16)[name = tensor("input_377_cast_fp16")]; + tensor linear_144_cast_fp16 = linear(bias = flow_net_res_blocks_2_mlp_2_bias_to_fp16, weight = flow_net_res_blocks_2_mlp_2_weight_to_fp16, x = input_377_cast_fp16)[name = tensor("linear_144_cast_fp16")]; + tensor var_2214_cast_fp16 = mul(x = var_2185_cast_fp16_2, y = linear_144_cast_fp16)[name = tensor("op_2214_cast_fp16")]; + tensor x_243_cast_fp16 = add(x = x_237_cast_fp16, y = var_2214_cast_fp16)[name = tensor("x_243_cast_fp16")]; + tensor linear_145_cast_fp16 = linear(bias = flow_net_res_blocks_3_adaLN_modulation_1_bias_to_fp16, weight = flow_net_res_blocks_3_adaLN_modulation_1_weight_to_fp16, x = input_355_cast_fp16)[name = tensor("linear_145_cast_fp16")]; + tensor var_2224_split_sizes_0 = const()[name = tensor("op_2224_split_sizes_0"), val = tensor([512, 512, 512])]; + tensor var_2224_axis_0 = const()[name = tensor("op_2224_axis_0"), val = tensor(-1)]; + tensor var_2224_cast_fp16_0, tensor var_2224_cast_fp16_1, tensor var_2224_cast_fp16_2 = split(axis = var_2224_axis_0, split_sizes = var_2224_split_sizes_0, x = linear_145_cast_fp16)[name = tensor("op_2224_cast_fp16")]; + tensor mean_77_axes_0 = const()[name = tensor("mean_77_axes_0"), val = tensor([-1])]; + tensor mean_77_keep_dims_0 = const()[name = tensor("mean_77_keep_dims_0"), val = tensor(true)]; + tensor mean_77_cast_fp16 = reduce_mean(axes = mean_77_axes_0, keep_dims = mean_77_keep_dims_0, x = x_243_cast_fp16)[name = tensor("mean_77_cast_fp16")]; + tensor sub_62_cast_fp16 = sub(x = x_243_cast_fp16, y = mean_77_cast_fp16)[name = tensor("sub_62_cast_fp16")]; + tensor square_50_cast_fp16 = square(x = sub_62_cast_fp16)[name = tensor("square_50_cast_fp16")]; + tensor reduce_mean_101_axes_0 = const()[name = tensor("reduce_mean_101_axes_0"), val = tensor([-1])]; + tensor reduce_mean_101_keep_dims_0 = const()[name = tensor("reduce_mean_101_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_101_cast_fp16 = reduce_mean(axes = reduce_mean_101_axes_0, keep_dims = reduce_mean_101_keep_dims_0, x = square_50_cast_fp16)[name = tensor("reduce_mean_101_cast_fp16")]; + tensor var_2234_to_fp16 = const()[name = tensor("op_2234_to_fp16"), val = tensor(0x1.1p-20)]; + tensor var_2235_cast_fp16 = add(x = reduce_mean_101_cast_fp16, y = var_2234_to_fp16)[name = tensor("op_2235_cast_fp16")]; + tensor var_2236_cast_fp16 = sqrt(x = var_2235_cast_fp16)[name = tensor("op_2236_cast_fp16")]; + tensor x_245_cast_fp16 = real_div(x = sub_62_cast_fp16, y = var_2236_cast_fp16)[name = tensor("x_245_cast_fp16")]; + tensor var_2238_cast_fp16 = mul(x = x_245_cast_fp16, y = flow_net_res_blocks_3_in_ln_weight_to_fp16)[name = tensor("op_2238_cast_fp16")]; + tensor x_247_cast_fp16 = add(x = var_2238_cast_fp16, y = flow_net_res_blocks_3_in_ln_bias_to_fp16)[name = tensor("x_247_cast_fp16")]; + tensor var_2240_promoted_to_fp16 = const()[name = tensor("op_2240_promoted_to_fp16"), val = tensor(0x1p+0)]; + tensor var_2241_cast_fp16 = add(x = var_2224_cast_fp16_1, y = var_2240_promoted_to_fp16)[name = tensor("op_2241_cast_fp16")]; + tensor var_2242_cast_fp16 = mul(x = x_247_cast_fp16, y = var_2241_cast_fp16)[name = tensor("op_2242_cast_fp16")]; + tensor input_381_cast_fp16 = add(x = var_2242_cast_fp16, y = var_2224_cast_fp16_0)[name = tensor("input_381_cast_fp16")]; + tensor linear_146_cast_fp16 = linear(bias = flow_net_res_blocks_3_mlp_0_bias_to_fp16, weight = flow_net_res_blocks_3_mlp_0_weight_to_fp16, x = input_381_cast_fp16)[name = tensor("linear_146_cast_fp16")]; + tensor input_385_cast_fp16 = silu(x = linear_146_cast_fp16)[name = tensor("input_385_cast_fp16")]; + tensor linear_147_cast_fp16 = linear(bias = flow_net_res_blocks_3_mlp_2_bias_to_fp16, weight = flow_net_res_blocks_3_mlp_2_weight_to_fp16, x = input_385_cast_fp16)[name = tensor("linear_147_cast_fp16")]; + tensor var_2253_cast_fp16 = mul(x = var_2224_cast_fp16_2, y = linear_147_cast_fp16)[name = tensor("op_2253_cast_fp16")]; + tensor x_249_cast_fp16 = add(x = x_243_cast_fp16, y = var_2253_cast_fp16)[name = tensor("x_249_cast_fp16")]; + tensor linear_148_cast_fp16 = linear(bias = flow_net_res_blocks_4_adaLN_modulation_1_bias_to_fp16, weight = flow_net_res_blocks_4_adaLN_modulation_1_weight_to_fp16, x = input_355_cast_fp16)[name = tensor("linear_148_cast_fp16")]; + tensor var_2263_split_sizes_0 = const()[name = tensor("op_2263_split_sizes_0"), val = tensor([512, 512, 512])]; + tensor var_2263_axis_0 = const()[name = tensor("op_2263_axis_0"), val = tensor(-1)]; + tensor var_2263_cast_fp16_0, tensor var_2263_cast_fp16_1, tensor var_2263_cast_fp16_2 = split(axis = var_2263_axis_0, split_sizes = var_2263_split_sizes_0, x = linear_148_cast_fp16)[name = tensor("op_2263_cast_fp16")]; + tensor mean_79_axes_0 = const()[name = tensor("mean_79_axes_0"), val = tensor([-1])]; + tensor mean_79_keep_dims_0 = const()[name = tensor("mean_79_keep_dims_0"), val = tensor(true)]; + tensor mean_79_cast_fp16 = reduce_mean(axes = mean_79_axes_0, keep_dims = mean_79_keep_dims_0, x = x_249_cast_fp16)[name = tensor("mean_79_cast_fp16")]; + tensor sub_63_cast_fp16 = sub(x = x_249_cast_fp16, y = mean_79_cast_fp16)[name = tensor("sub_63_cast_fp16")]; + tensor square_51_cast_fp16 = square(x = sub_63_cast_fp16)[name = tensor("square_51_cast_fp16")]; + tensor reduce_mean_103_axes_0 = const()[name = tensor("reduce_mean_103_axes_0"), val = tensor([-1])]; + tensor reduce_mean_103_keep_dims_0 = const()[name = tensor("reduce_mean_103_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_103_cast_fp16 = reduce_mean(axes = reduce_mean_103_axes_0, keep_dims = reduce_mean_103_keep_dims_0, x = square_51_cast_fp16)[name = tensor("reduce_mean_103_cast_fp16")]; + tensor var_2273_to_fp16 = const()[name = tensor("op_2273_to_fp16"), val = tensor(0x1.1p-20)]; + tensor var_2274_cast_fp16 = add(x = reduce_mean_103_cast_fp16, y = var_2273_to_fp16)[name = tensor("op_2274_cast_fp16")]; + tensor var_2275_cast_fp16 = sqrt(x = var_2274_cast_fp16)[name = tensor("op_2275_cast_fp16")]; + tensor x_251_cast_fp16 = real_div(x = sub_63_cast_fp16, y = var_2275_cast_fp16)[name = tensor("x_251_cast_fp16")]; + tensor var_2277_cast_fp16 = mul(x = x_251_cast_fp16, y = flow_net_res_blocks_4_in_ln_weight_to_fp16)[name = tensor("op_2277_cast_fp16")]; + tensor x_253_cast_fp16 = add(x = var_2277_cast_fp16, y = flow_net_res_blocks_4_in_ln_bias_to_fp16)[name = tensor("x_253_cast_fp16")]; + tensor var_2279_promoted_to_fp16 = const()[name = tensor("op_2279_promoted_to_fp16"), val = tensor(0x1p+0)]; + tensor var_2280_cast_fp16 = add(x = var_2263_cast_fp16_1, y = var_2279_promoted_to_fp16)[name = tensor("op_2280_cast_fp16")]; + tensor var_2281_cast_fp16 = mul(x = x_253_cast_fp16, y = var_2280_cast_fp16)[name = tensor("op_2281_cast_fp16")]; + tensor input_389_cast_fp16 = add(x = var_2281_cast_fp16, y = var_2263_cast_fp16_0)[name = tensor("input_389_cast_fp16")]; + tensor linear_149_cast_fp16 = linear(bias = flow_net_res_blocks_4_mlp_0_bias_to_fp16, weight = flow_net_res_blocks_4_mlp_0_weight_to_fp16, x = input_389_cast_fp16)[name = tensor("linear_149_cast_fp16")]; + tensor input_393_cast_fp16 = silu(x = linear_149_cast_fp16)[name = tensor("input_393_cast_fp16")]; + tensor linear_150_cast_fp16 = linear(bias = flow_net_res_blocks_4_mlp_2_bias_to_fp16, weight = flow_net_res_blocks_4_mlp_2_weight_to_fp16, x = input_393_cast_fp16)[name = tensor("linear_150_cast_fp16")]; + tensor var_2292_cast_fp16 = mul(x = var_2263_cast_fp16_2, y = linear_150_cast_fp16)[name = tensor("op_2292_cast_fp16")]; + tensor x_255_cast_fp16 = add(x = x_249_cast_fp16, y = var_2292_cast_fp16)[name = tensor("x_255_cast_fp16")]; + tensor linear_151_cast_fp16 = linear(bias = flow_net_res_blocks_5_adaLN_modulation_1_bias_to_fp16, weight = flow_net_res_blocks_5_adaLN_modulation_1_weight_to_fp16, x = input_355_cast_fp16)[name = tensor("linear_151_cast_fp16")]; + tensor var_2302_split_sizes_0 = const()[name = tensor("op_2302_split_sizes_0"), val = tensor([512, 512, 512])]; + tensor var_2302_axis_0 = const()[name = tensor("op_2302_axis_0"), val = tensor(-1)]; + tensor var_2302_cast_fp16_0, tensor var_2302_cast_fp16_1, tensor var_2302_cast_fp16_2 = split(axis = var_2302_axis_0, split_sizes = var_2302_split_sizes_0, x = linear_151_cast_fp16)[name = tensor("op_2302_cast_fp16")]; + tensor mean_81_axes_0 = const()[name = tensor("mean_81_axes_0"), val = tensor([-1])]; + tensor mean_81_keep_dims_0 = const()[name = tensor("mean_81_keep_dims_0"), val = tensor(true)]; + tensor mean_81_cast_fp16 = reduce_mean(axes = mean_81_axes_0, keep_dims = mean_81_keep_dims_0, x = x_255_cast_fp16)[name = tensor("mean_81_cast_fp16")]; + tensor sub_64_cast_fp16 = sub(x = x_255_cast_fp16, y = mean_81_cast_fp16)[name = tensor("sub_64_cast_fp16")]; + tensor square_52_cast_fp16 = square(x = sub_64_cast_fp16)[name = tensor("square_52_cast_fp16")]; + tensor reduce_mean_105_axes_0 = const()[name = tensor("reduce_mean_105_axes_0"), val = tensor([-1])]; + tensor reduce_mean_105_keep_dims_0 = const()[name = tensor("reduce_mean_105_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_105_cast_fp16 = reduce_mean(axes = reduce_mean_105_axes_0, keep_dims = reduce_mean_105_keep_dims_0, x = square_52_cast_fp16)[name = tensor("reduce_mean_105_cast_fp16")]; + tensor var_2312_to_fp16 = const()[name = tensor("op_2312_to_fp16"), val = tensor(0x1.1p-20)]; + tensor var_2313_cast_fp16 = add(x = reduce_mean_105_cast_fp16, y = var_2312_to_fp16)[name = tensor("op_2313_cast_fp16")]; + tensor var_2314_cast_fp16 = sqrt(x = var_2313_cast_fp16)[name = tensor("op_2314_cast_fp16")]; + tensor x_257_cast_fp16 = real_div(x = sub_64_cast_fp16, y = var_2314_cast_fp16)[name = tensor("x_257_cast_fp16")]; + tensor var_2316_cast_fp16 = mul(x = x_257_cast_fp16, y = flow_net_res_blocks_5_in_ln_weight_to_fp16)[name = tensor("op_2316_cast_fp16")]; + tensor x_259_cast_fp16 = add(x = var_2316_cast_fp16, y = flow_net_res_blocks_5_in_ln_bias_to_fp16)[name = tensor("x_259_cast_fp16")]; + tensor var_2318_promoted_to_fp16 = const()[name = tensor("op_2318_promoted_to_fp16"), val = tensor(0x1p+0)]; + tensor var_2319_cast_fp16 = add(x = var_2302_cast_fp16_1, y = var_2318_promoted_to_fp16)[name = tensor("op_2319_cast_fp16")]; + tensor var_2320_cast_fp16 = mul(x = x_259_cast_fp16, y = var_2319_cast_fp16)[name = tensor("op_2320_cast_fp16")]; + tensor input_397_cast_fp16 = add(x = var_2320_cast_fp16, y = var_2302_cast_fp16_0)[name = tensor("input_397_cast_fp16")]; + tensor linear_152_cast_fp16 = linear(bias = flow_net_res_blocks_5_mlp_0_bias_to_fp16, weight = flow_net_res_blocks_5_mlp_0_weight_to_fp16, x = input_397_cast_fp16)[name = tensor("linear_152_cast_fp16")]; + tensor input_401_cast_fp16 = silu(x = linear_152_cast_fp16)[name = tensor("input_401_cast_fp16")]; + tensor linear_153_cast_fp16 = linear(bias = flow_net_res_blocks_5_mlp_2_bias_to_fp16, weight = flow_net_res_blocks_5_mlp_2_weight_to_fp16, x = input_401_cast_fp16)[name = tensor("linear_153_cast_fp16")]; + tensor var_2331_cast_fp16 = mul(x = var_2302_cast_fp16_2, y = linear_153_cast_fp16)[name = tensor("op_2331_cast_fp16")]; + tensor x_261_cast_fp16 = add(x = x_255_cast_fp16, y = var_2331_cast_fp16)[name = tensor("x_261_cast_fp16")]; + tensor linear_154_cast_fp16 = linear(bias = flow_net_final_layer_adaLN_modulation_1_bias_to_fp16, weight = flow_net_final_layer_adaLN_modulation_1_weight_to_fp16, x = input_355_cast_fp16)[name = tensor("linear_154_cast_fp16")]; + tensor var_2340_split_sizes_0 = const()[name = tensor("op_2340_split_sizes_0"), val = tensor([512, 512])]; + tensor var_2340_axis_0 = const()[name = tensor("op_2340_axis_0"), val = tensor(-1)]; + tensor var_2340_cast_fp16_0, tensor var_2340_cast_fp16_1 = split(axis = var_2340_axis_0, split_sizes = var_2340_split_sizes_0, x = linear_154_cast_fp16)[name = tensor("op_2340_cast_fp16")]; + tensor mean_83_axes_0 = const()[name = tensor("mean_83_axes_0"), val = tensor([-1])]; + tensor mean_83_keep_dims_0 = const()[name = tensor("mean_83_keep_dims_0"), val = tensor(true)]; + tensor mean_83_cast_fp16 = reduce_mean(axes = mean_83_axes_0, keep_dims = mean_83_keep_dims_0, x = x_261_cast_fp16)[name = tensor("mean_83_cast_fp16")]; + tensor sub_65_cast_fp16 = sub(x = x_261_cast_fp16, y = mean_83_cast_fp16)[name = tensor("sub_65_cast_fp16")]; + tensor square_53_cast_fp16 = square(x = sub_65_cast_fp16)[name = tensor("square_53_cast_fp16")]; + tensor reduce_mean_107_axes_0 = const()[name = tensor("reduce_mean_107_axes_0"), val = tensor([-1])]; + tensor reduce_mean_107_keep_dims_0 = const()[name = tensor("reduce_mean_107_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_107_cast_fp16 = reduce_mean(axes = reduce_mean_107_axes_0, keep_dims = reduce_mean_107_keep_dims_0, x = square_53_cast_fp16)[name = tensor("reduce_mean_107_cast_fp16")]; + tensor var_2347_to_fp16 = const()[name = tensor("op_2347_to_fp16"), val = tensor(0x1.1p-20)]; + tensor var_2348_cast_fp16 = add(x = reduce_mean_107_cast_fp16, y = var_2347_to_fp16)[name = tensor("op_2348_cast_fp16")]; + tensor var_2349_cast_fp16 = sqrt(x = var_2348_cast_fp16)[name = tensor("op_2349_cast_fp16")]; + tensor x_263_cast_fp16 = real_div(x = sub_65_cast_fp16, y = var_2349_cast_fp16)[name = tensor("x_263_cast_fp16")]; + tensor var_2351_promoted_to_fp16 = const()[name = tensor("op_2351_promoted_to_fp16"), val = tensor(0x1p+0)]; + tensor var_2352_cast_fp16 = add(x = var_2340_cast_fp16_1, y = var_2351_promoted_to_fp16)[name = tensor("op_2352_cast_fp16")]; + tensor var_2353_cast_fp16 = mul(x = x_263_cast_fp16, y = var_2352_cast_fp16)[name = tensor("op_2353_cast_fp16")]; + tensor input_405_cast_fp16 = add(x = var_2353_cast_fp16, y = var_2340_cast_fp16_0)[name = tensor("input_405_cast_fp16")]; + tensor linear_155_cast_fp16 = linear(bias = flow_net_final_layer_linear_bias_to_fp16, weight = flow_net_final_layer_linear_weight_to_fp16, x = input_405_cast_fp16)[name = tensor("linear_155_cast_fp16")]; + tensor var_2364_to_fp16 = const()[name = tensor("op_2364_to_fp16"), val = tensor(0x1p-3)]; + tensor var_2365_cast_fp16 = mul(x = linear_155_cast_fp16, y = var_2364_to_fp16)[name = tensor("op_2365_cast_fp16")]; + tensor input_407_cast_fp16 = add(x = input_339_cast_fp16, y = var_2365_cast_fp16)[name = tensor("input_407_cast_fp16")]; + tensor linear_156_cast_fp16 = linear(bias = flow_net_input_proj_bias_to_fp16, weight = flow_net_input_proj_weight_to_fp16, x = input_407_cast_fp16)[name = tensor("linear_156_cast_fp16")]; + tensor input_411_to_fp16 = const()[name = tensor("input_411_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(19008256)))]; + tensor input_413_cast_fp16 = silu(x = input_411_to_fp16)[name = tensor("input_413_cast_fp16")]; + tensor linear_158_cast_fp16 = linear(bias = flow_net_time_embed_0_mlp_2_bias_to_fp16, weight = flow_net_time_embed_0_mlp_2_weight_to_fp16, x = input_413_cast_fp16)[name = tensor("linear_158_cast_fp16")]; + tensor reduce_mean_108_axes_0 = const()[name = tensor("reduce_mean_108_axes_0"), val = tensor([-1])]; + tensor reduce_mean_108_keep_dims_0 = const()[name = tensor("reduce_mean_108_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_108_cast_fp16 = reduce_mean(axes = reduce_mean_108_axes_0, keep_dims = reduce_mean_108_keep_dims_0, x = linear_158_cast_fp16)[name = tensor("reduce_mean_108_cast_fp16")]; + tensor sub_66_cast_fp16 = sub(x = linear_158_cast_fp16, y = reduce_mean_108_cast_fp16)[name = tensor("sub_66_cast_fp16")]; + tensor square_54_cast_fp16 = square(x = sub_66_cast_fp16)[name = tensor("square_54_cast_fp16")]; + tensor reduce_mean_109_axes_0 = const()[name = tensor("reduce_mean_109_axes_0"), val = tensor([-1])]; + tensor reduce_mean_109_keep_dims_0 = const()[name = tensor("reduce_mean_109_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_109_cast_fp16 = reduce_mean(axes = reduce_mean_109_axes_0, keep_dims = reduce_mean_109_keep_dims_0, x = square_54_cast_fp16)[name = tensor("reduce_mean_109_cast_fp16")]; + tensor real_div_12_to_fp16 = const()[name = tensor("real_div_12_to_fp16"), val = tensor(0x1.008p+0)]; + tensor mul_12_cast_fp16 = mul(x = reduce_mean_109_cast_fp16, y = real_div_12_to_fp16)[name = tensor("mul_12_cast_fp16")]; + tensor var_2431_to_fp16 = const()[name = tensor("op_2431_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_109_cast_fp16 = add(x = mul_12_cast_fp16, y = var_2431_to_fp16)[name = tensor("var_109_cast_fp16")]; + tensor var_2434_epsilon_0 = const()[name = tensor("op_2434_epsilon_0"), val = tensor(0x1.197998p-40)]; + tensor var_2434_cast_fp16 = rsqrt(epsilon = var_2434_epsilon_0, x = var_109_cast_fp16)[name = tensor("op_2434_cast_fp16")]; + tensor var_2435_cast_fp16 = mul(x = const_3_to_fp16, y = var_2434_cast_fp16)[name = tensor("op_2435_cast_fp16")]; + tensor var_2436_cast_fp16 = mul(x = linear_158_cast_fp16, y = var_2435_cast_fp16)[name = tensor("op_2436_cast_fp16")]; + tensor input_417_to_fp16 = const()[name = tensor("input_417_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(19009344)))]; + tensor input_419_cast_fp16 = silu(x = input_417_to_fp16)[name = tensor("input_419_cast_fp16")]; + tensor linear_160_cast_fp16 = linear(bias = flow_net_time_embed_1_mlp_2_bias_to_fp16, weight = flow_net_time_embed_1_mlp_2_weight_to_fp16, x = input_419_cast_fp16)[name = tensor("linear_160_cast_fp16")]; + tensor reduce_mean_110_axes_0 = const()[name = tensor("reduce_mean_110_axes_0"), val = tensor([-1])]; + tensor reduce_mean_110_keep_dims_0 = const()[name = tensor("reduce_mean_110_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_110_cast_fp16 = reduce_mean(axes = reduce_mean_110_axes_0, keep_dims = reduce_mean_110_keep_dims_0, x = linear_160_cast_fp16)[name = tensor("reduce_mean_110_cast_fp16")]; + tensor sub_68_cast_fp16 = sub(x = linear_160_cast_fp16, y = reduce_mean_110_cast_fp16)[name = tensor("sub_68_cast_fp16")]; + tensor square_55_cast_fp16 = square(x = sub_68_cast_fp16)[name = tensor("square_55_cast_fp16")]; + tensor reduce_mean_111_axes_0 = const()[name = tensor("reduce_mean_111_axes_0"), val = tensor([-1])]; + tensor reduce_mean_111_keep_dims_0 = const()[name = tensor("reduce_mean_111_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_111_cast_fp16 = reduce_mean(axes = reduce_mean_111_axes_0, keep_dims = reduce_mean_111_keep_dims_0, x = square_55_cast_fp16)[name = tensor("reduce_mean_111_cast_fp16")]; + tensor real_div_13_to_fp16 = const()[name = tensor("real_div_13_to_fp16"), val = tensor(0x1.008p+0)]; + tensor mul_13_cast_fp16 = mul(x = reduce_mean_111_cast_fp16, y = real_div_13_to_fp16)[name = tensor("mul_13_cast_fp16")]; + tensor var_2468_to_fp16 = const()[name = tensor("op_2468_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_111_cast_fp16 = add(x = mul_13_cast_fp16, y = var_2468_to_fp16)[name = tensor("var_111_cast_fp16")]; + tensor var_2471_epsilon_0 = const()[name = tensor("op_2471_epsilon_0"), val = tensor(0x1.197998p-40)]; + tensor var_2471_cast_fp16 = rsqrt(epsilon = var_2471_epsilon_0, x = var_111_cast_fp16)[name = tensor("op_2471_cast_fp16")]; + tensor var_2472_cast_fp16 = mul(x = const_5_to_fp16, y = var_2471_cast_fp16)[name = tensor("op_2472_cast_fp16")]; + tensor var_2473_cast_fp16 = mul(x = linear_160_cast_fp16, y = var_2472_cast_fp16)[name = tensor("op_2473_cast_fp16")]; + tensor var_2485_cast_fp16 = add(x = var_2436_cast_fp16, y = var_2473_cast_fp16)[name = tensor("op_2485_cast_fp16")]; + tensor _inversed_t_combined_13_y_0_to_fp16 = const()[name = tensor("_inversed_t_combined_13_y_0_to_fp16"), val = tensor(0x1p-1)]; + tensor _inversed_t_combined_13_cast_fp16 = mul(x = var_2485_cast_fp16, y = _inversed_t_combined_13_y_0_to_fp16)[name = tensor("_inversed_t_combined_13_cast_fp16")]; + tensor input_421_cast_fp16 = add(x = _inversed_t_combined_13_cast_fp16, y = linear_5_cast_fp16)[name = tensor("input_421_cast_fp16")]; + tensor input_423_cast_fp16 = silu(x = input_421_cast_fp16)[name = tensor("input_423_cast_fp16")]; + tensor linear_162_cast_fp16 = linear(bias = flow_net_res_blocks_0_adaLN_modulation_1_bias_to_fp16, weight = flow_net_res_blocks_0_adaLN_modulation_1_weight_to_fp16, x = input_423_cast_fp16)[name = tensor("linear_162_cast_fp16")]; + tensor var_2500_split_sizes_0 = const()[name = tensor("op_2500_split_sizes_0"), val = tensor([512, 512, 512])]; + tensor var_2500_axis_0 = const()[name = tensor("op_2500_axis_0"), val = tensor(-1)]; + tensor var_2500_cast_fp16_0, tensor var_2500_cast_fp16_1, tensor var_2500_cast_fp16_2 = split(axis = var_2500_axis_0, split_sizes = var_2500_split_sizes_0, x = linear_162_cast_fp16)[name = tensor("op_2500_cast_fp16")]; + tensor mean_85_axes_0 = const()[name = tensor("mean_85_axes_0"), val = tensor([-1])]; + tensor mean_85_keep_dims_0 = const()[name = tensor("mean_85_keep_dims_0"), val = tensor(true)]; + tensor mean_85_cast_fp16 = reduce_mean(axes = mean_85_axes_0, keep_dims = mean_85_keep_dims_0, x = linear_156_cast_fp16)[name = tensor("mean_85_cast_fp16")]; + tensor sub_70_cast_fp16 = sub(x = linear_156_cast_fp16, y = mean_85_cast_fp16)[name = tensor("sub_70_cast_fp16")]; + tensor square_56_cast_fp16 = square(x = sub_70_cast_fp16)[name = tensor("square_56_cast_fp16")]; + tensor reduce_mean_113_axes_0 = const()[name = tensor("reduce_mean_113_axes_0"), val = tensor([-1])]; + tensor reduce_mean_113_keep_dims_0 = const()[name = tensor("reduce_mean_113_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_113_cast_fp16 = reduce_mean(axes = reduce_mean_113_axes_0, keep_dims = reduce_mean_113_keep_dims_0, x = square_56_cast_fp16)[name = tensor("reduce_mean_113_cast_fp16")]; + tensor var_2510_to_fp16 = const()[name = tensor("op_2510_to_fp16"), val = tensor(0x1.1p-20)]; + tensor var_2511_cast_fp16 = add(x = reduce_mean_113_cast_fp16, y = var_2510_to_fp16)[name = tensor("op_2511_cast_fp16")]; + tensor var_2512_cast_fp16 = sqrt(x = var_2511_cast_fp16)[name = tensor("op_2512_cast_fp16")]; + tensor x_271_cast_fp16 = real_div(x = sub_70_cast_fp16, y = var_2512_cast_fp16)[name = tensor("x_271_cast_fp16")]; + tensor var_2514_cast_fp16 = mul(x = x_271_cast_fp16, y = flow_net_res_blocks_0_in_ln_weight_to_fp16)[name = tensor("op_2514_cast_fp16")]; + tensor x_273_cast_fp16 = add(x = var_2514_cast_fp16, y = flow_net_res_blocks_0_in_ln_bias_to_fp16)[name = tensor("x_273_cast_fp16")]; + tensor var_2516_promoted_to_fp16 = const()[name = tensor("op_2516_promoted_to_fp16"), val = tensor(0x1p+0)]; + tensor var_2517_cast_fp16 = add(x = var_2500_cast_fp16_1, y = var_2516_promoted_to_fp16)[name = tensor("op_2517_cast_fp16")]; + tensor var_2518_cast_fp16 = mul(x = x_273_cast_fp16, y = var_2517_cast_fp16)[name = tensor("op_2518_cast_fp16")]; + tensor input_425_cast_fp16 = add(x = var_2518_cast_fp16, y = var_2500_cast_fp16_0)[name = tensor("input_425_cast_fp16")]; + tensor linear_163_cast_fp16 = linear(bias = flow_net_res_blocks_0_mlp_0_bias_to_fp16, weight = flow_net_res_blocks_0_mlp_0_weight_to_fp16, x = input_425_cast_fp16)[name = tensor("linear_163_cast_fp16")]; + tensor input_429_cast_fp16 = silu(x = linear_163_cast_fp16)[name = tensor("input_429_cast_fp16")]; + tensor linear_164_cast_fp16 = linear(bias = flow_net_res_blocks_0_mlp_2_bias_to_fp16, weight = flow_net_res_blocks_0_mlp_2_weight_to_fp16, x = input_429_cast_fp16)[name = tensor("linear_164_cast_fp16")]; + tensor var_2529_cast_fp16 = mul(x = var_2500_cast_fp16_2, y = linear_164_cast_fp16)[name = tensor("op_2529_cast_fp16")]; + tensor x_275_cast_fp16 = add(x = linear_156_cast_fp16, y = var_2529_cast_fp16)[name = tensor("x_275_cast_fp16")]; + tensor linear_165_cast_fp16 = linear(bias = flow_net_res_blocks_1_adaLN_modulation_1_bias_to_fp16, weight = flow_net_res_blocks_1_adaLN_modulation_1_weight_to_fp16, x = input_423_cast_fp16)[name = tensor("linear_165_cast_fp16")]; + tensor var_2539_split_sizes_0 = const()[name = tensor("op_2539_split_sizes_0"), val = tensor([512, 512, 512])]; + tensor var_2539_axis_0 = const()[name = tensor("op_2539_axis_0"), val = tensor(-1)]; + tensor var_2539_cast_fp16_0, tensor var_2539_cast_fp16_1, tensor var_2539_cast_fp16_2 = split(axis = var_2539_axis_0, split_sizes = var_2539_split_sizes_0, x = linear_165_cast_fp16)[name = tensor("op_2539_cast_fp16")]; + tensor mean_87_axes_0 = const()[name = tensor("mean_87_axes_0"), val = tensor([-1])]; + tensor mean_87_keep_dims_0 = const()[name = tensor("mean_87_keep_dims_0"), val = tensor(true)]; + tensor mean_87_cast_fp16 = reduce_mean(axes = mean_87_axes_0, keep_dims = mean_87_keep_dims_0, x = x_275_cast_fp16)[name = tensor("mean_87_cast_fp16")]; + tensor sub_71_cast_fp16 = sub(x = x_275_cast_fp16, y = mean_87_cast_fp16)[name = tensor("sub_71_cast_fp16")]; + tensor square_57_cast_fp16 = square(x = sub_71_cast_fp16)[name = tensor("square_57_cast_fp16")]; + tensor reduce_mean_115_axes_0 = const()[name = tensor("reduce_mean_115_axes_0"), val = tensor([-1])]; + tensor reduce_mean_115_keep_dims_0 = const()[name = tensor("reduce_mean_115_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_115_cast_fp16 = reduce_mean(axes = reduce_mean_115_axes_0, keep_dims = reduce_mean_115_keep_dims_0, x = square_57_cast_fp16)[name = tensor("reduce_mean_115_cast_fp16")]; + tensor var_2549_to_fp16 = const()[name = tensor("op_2549_to_fp16"), val = tensor(0x1.1p-20)]; + tensor var_2550_cast_fp16 = add(x = reduce_mean_115_cast_fp16, y = var_2549_to_fp16)[name = tensor("op_2550_cast_fp16")]; + tensor var_2551_cast_fp16 = sqrt(x = var_2550_cast_fp16)[name = tensor("op_2551_cast_fp16")]; + tensor x_277_cast_fp16 = real_div(x = sub_71_cast_fp16, y = var_2551_cast_fp16)[name = tensor("x_277_cast_fp16")]; + tensor var_2553_cast_fp16 = mul(x = x_277_cast_fp16, y = flow_net_res_blocks_1_in_ln_weight_to_fp16)[name = tensor("op_2553_cast_fp16")]; + tensor x_279_cast_fp16 = add(x = var_2553_cast_fp16, y = flow_net_res_blocks_1_in_ln_bias_to_fp16)[name = tensor("x_279_cast_fp16")]; + tensor var_2555_promoted_to_fp16 = const()[name = tensor("op_2555_promoted_to_fp16"), val = tensor(0x1p+0)]; + tensor var_2556_cast_fp16 = add(x = var_2539_cast_fp16_1, y = var_2555_promoted_to_fp16)[name = tensor("op_2556_cast_fp16")]; + tensor var_2557_cast_fp16 = mul(x = x_279_cast_fp16, y = var_2556_cast_fp16)[name = tensor("op_2557_cast_fp16")]; + tensor input_433_cast_fp16 = add(x = var_2557_cast_fp16, y = var_2539_cast_fp16_0)[name = tensor("input_433_cast_fp16")]; + tensor linear_166_cast_fp16 = linear(bias = flow_net_res_blocks_1_mlp_0_bias_to_fp16, weight = flow_net_res_blocks_1_mlp_0_weight_to_fp16, x = input_433_cast_fp16)[name = tensor("linear_166_cast_fp16")]; + tensor input_437_cast_fp16 = silu(x = linear_166_cast_fp16)[name = tensor("input_437_cast_fp16")]; + tensor linear_167_cast_fp16 = linear(bias = flow_net_res_blocks_1_mlp_2_bias_to_fp16, weight = flow_net_res_blocks_1_mlp_2_weight_to_fp16, x = input_437_cast_fp16)[name = tensor("linear_167_cast_fp16")]; + tensor var_2568_cast_fp16 = mul(x = var_2539_cast_fp16_2, y = linear_167_cast_fp16)[name = tensor("op_2568_cast_fp16")]; + tensor x_281_cast_fp16 = add(x = x_275_cast_fp16, y = var_2568_cast_fp16)[name = tensor("x_281_cast_fp16")]; + tensor linear_168_cast_fp16 = linear(bias = flow_net_res_blocks_2_adaLN_modulation_1_bias_to_fp16, weight = flow_net_res_blocks_2_adaLN_modulation_1_weight_to_fp16, x = input_423_cast_fp16)[name = tensor("linear_168_cast_fp16")]; + tensor var_2578_split_sizes_0 = const()[name = tensor("op_2578_split_sizes_0"), val = tensor([512, 512, 512])]; + tensor var_2578_axis_0 = const()[name = tensor("op_2578_axis_0"), val = tensor(-1)]; + tensor var_2578_cast_fp16_0, tensor var_2578_cast_fp16_1, tensor var_2578_cast_fp16_2 = split(axis = var_2578_axis_0, split_sizes = var_2578_split_sizes_0, x = linear_168_cast_fp16)[name = tensor("op_2578_cast_fp16")]; + tensor mean_89_axes_0 = const()[name = tensor("mean_89_axes_0"), val = tensor([-1])]; + tensor mean_89_keep_dims_0 = const()[name = tensor("mean_89_keep_dims_0"), val = tensor(true)]; + tensor mean_89_cast_fp16 = reduce_mean(axes = mean_89_axes_0, keep_dims = mean_89_keep_dims_0, x = x_281_cast_fp16)[name = tensor("mean_89_cast_fp16")]; + tensor sub_72_cast_fp16 = sub(x = x_281_cast_fp16, y = mean_89_cast_fp16)[name = tensor("sub_72_cast_fp16")]; + tensor square_58_cast_fp16 = square(x = sub_72_cast_fp16)[name = tensor("square_58_cast_fp16")]; + tensor reduce_mean_117_axes_0 = const()[name = tensor("reduce_mean_117_axes_0"), val = tensor([-1])]; + tensor reduce_mean_117_keep_dims_0 = const()[name = tensor("reduce_mean_117_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_117_cast_fp16 = reduce_mean(axes = reduce_mean_117_axes_0, keep_dims = reduce_mean_117_keep_dims_0, x = square_58_cast_fp16)[name = tensor("reduce_mean_117_cast_fp16")]; + tensor var_2588_to_fp16 = const()[name = tensor("op_2588_to_fp16"), val = tensor(0x1.1p-20)]; + tensor var_2589_cast_fp16 = add(x = reduce_mean_117_cast_fp16, y = var_2588_to_fp16)[name = tensor("op_2589_cast_fp16")]; + tensor var_2590_cast_fp16 = sqrt(x = var_2589_cast_fp16)[name = tensor("op_2590_cast_fp16")]; + tensor x_283_cast_fp16 = real_div(x = sub_72_cast_fp16, y = var_2590_cast_fp16)[name = tensor("x_283_cast_fp16")]; + tensor var_2592_cast_fp16 = mul(x = x_283_cast_fp16, y = flow_net_res_blocks_2_in_ln_weight_to_fp16)[name = tensor("op_2592_cast_fp16")]; + tensor x_285_cast_fp16 = add(x = var_2592_cast_fp16, y = flow_net_res_blocks_2_in_ln_bias_to_fp16)[name = tensor("x_285_cast_fp16")]; + tensor var_2594_promoted_to_fp16 = const()[name = tensor("op_2594_promoted_to_fp16"), val = tensor(0x1p+0)]; + tensor var_2595_cast_fp16 = add(x = var_2578_cast_fp16_1, y = var_2594_promoted_to_fp16)[name = tensor("op_2595_cast_fp16")]; + tensor var_2596_cast_fp16 = mul(x = x_285_cast_fp16, y = var_2595_cast_fp16)[name = tensor("op_2596_cast_fp16")]; + tensor input_441_cast_fp16 = add(x = var_2596_cast_fp16, y = var_2578_cast_fp16_0)[name = tensor("input_441_cast_fp16")]; + tensor linear_169_cast_fp16 = linear(bias = flow_net_res_blocks_2_mlp_0_bias_to_fp16, weight = flow_net_res_blocks_2_mlp_0_weight_to_fp16, x = input_441_cast_fp16)[name = tensor("linear_169_cast_fp16")]; + tensor input_445_cast_fp16 = silu(x = linear_169_cast_fp16)[name = tensor("input_445_cast_fp16")]; + tensor linear_170_cast_fp16 = linear(bias = flow_net_res_blocks_2_mlp_2_bias_to_fp16, weight = flow_net_res_blocks_2_mlp_2_weight_to_fp16, x = input_445_cast_fp16)[name = tensor("linear_170_cast_fp16")]; + tensor var_2607_cast_fp16 = mul(x = var_2578_cast_fp16_2, y = linear_170_cast_fp16)[name = tensor("op_2607_cast_fp16")]; + tensor x_287_cast_fp16 = add(x = x_281_cast_fp16, y = var_2607_cast_fp16)[name = tensor("x_287_cast_fp16")]; + tensor linear_171_cast_fp16 = linear(bias = flow_net_res_blocks_3_adaLN_modulation_1_bias_to_fp16, weight = flow_net_res_blocks_3_adaLN_modulation_1_weight_to_fp16, x = input_423_cast_fp16)[name = tensor("linear_171_cast_fp16")]; + tensor var_2617_split_sizes_0 = const()[name = tensor("op_2617_split_sizes_0"), val = tensor([512, 512, 512])]; + tensor var_2617_axis_0 = const()[name = tensor("op_2617_axis_0"), val = tensor(-1)]; + tensor var_2617_cast_fp16_0, tensor var_2617_cast_fp16_1, tensor var_2617_cast_fp16_2 = split(axis = var_2617_axis_0, split_sizes = var_2617_split_sizes_0, x = linear_171_cast_fp16)[name = tensor("op_2617_cast_fp16")]; + tensor mean_91_axes_0 = const()[name = tensor("mean_91_axes_0"), val = tensor([-1])]; + tensor mean_91_keep_dims_0 = const()[name = tensor("mean_91_keep_dims_0"), val = tensor(true)]; + tensor mean_91_cast_fp16 = reduce_mean(axes = mean_91_axes_0, keep_dims = mean_91_keep_dims_0, x = x_287_cast_fp16)[name = tensor("mean_91_cast_fp16")]; + tensor sub_73_cast_fp16 = sub(x = x_287_cast_fp16, y = mean_91_cast_fp16)[name = tensor("sub_73_cast_fp16")]; + tensor square_59_cast_fp16 = square(x = sub_73_cast_fp16)[name = tensor("square_59_cast_fp16")]; + tensor reduce_mean_119_axes_0 = const()[name = tensor("reduce_mean_119_axes_0"), val = tensor([-1])]; + tensor reduce_mean_119_keep_dims_0 = const()[name = tensor("reduce_mean_119_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_119_cast_fp16 = reduce_mean(axes = reduce_mean_119_axes_0, keep_dims = reduce_mean_119_keep_dims_0, x = square_59_cast_fp16)[name = tensor("reduce_mean_119_cast_fp16")]; + tensor var_2627_to_fp16 = const()[name = tensor("op_2627_to_fp16"), val = tensor(0x1.1p-20)]; + tensor var_2628_cast_fp16 = add(x = reduce_mean_119_cast_fp16, y = var_2627_to_fp16)[name = tensor("op_2628_cast_fp16")]; + tensor var_2629_cast_fp16 = sqrt(x = var_2628_cast_fp16)[name = tensor("op_2629_cast_fp16")]; + tensor x_289_cast_fp16 = real_div(x = sub_73_cast_fp16, y = var_2629_cast_fp16)[name = tensor("x_289_cast_fp16")]; + tensor var_2631_cast_fp16 = mul(x = x_289_cast_fp16, y = flow_net_res_blocks_3_in_ln_weight_to_fp16)[name = tensor("op_2631_cast_fp16")]; + tensor x_291_cast_fp16 = add(x = var_2631_cast_fp16, y = flow_net_res_blocks_3_in_ln_bias_to_fp16)[name = tensor("x_291_cast_fp16")]; + tensor var_2633_promoted_to_fp16 = const()[name = tensor("op_2633_promoted_to_fp16"), val = tensor(0x1p+0)]; + tensor var_2634_cast_fp16 = add(x = var_2617_cast_fp16_1, y = var_2633_promoted_to_fp16)[name = tensor("op_2634_cast_fp16")]; + tensor var_2635_cast_fp16 = mul(x = x_291_cast_fp16, y = var_2634_cast_fp16)[name = tensor("op_2635_cast_fp16")]; + tensor input_449_cast_fp16 = add(x = var_2635_cast_fp16, y = var_2617_cast_fp16_0)[name = tensor("input_449_cast_fp16")]; + tensor linear_172_cast_fp16 = linear(bias = flow_net_res_blocks_3_mlp_0_bias_to_fp16, weight = flow_net_res_blocks_3_mlp_0_weight_to_fp16, x = input_449_cast_fp16)[name = tensor("linear_172_cast_fp16")]; + tensor input_453_cast_fp16 = silu(x = linear_172_cast_fp16)[name = tensor("input_453_cast_fp16")]; + tensor linear_173_cast_fp16 = linear(bias = flow_net_res_blocks_3_mlp_2_bias_to_fp16, weight = flow_net_res_blocks_3_mlp_2_weight_to_fp16, x = input_453_cast_fp16)[name = tensor("linear_173_cast_fp16")]; + tensor var_2646_cast_fp16 = mul(x = var_2617_cast_fp16_2, y = linear_173_cast_fp16)[name = tensor("op_2646_cast_fp16")]; + tensor x_293_cast_fp16 = add(x = x_287_cast_fp16, y = var_2646_cast_fp16)[name = tensor("x_293_cast_fp16")]; + tensor linear_174_cast_fp16 = linear(bias = flow_net_res_blocks_4_adaLN_modulation_1_bias_to_fp16, weight = flow_net_res_blocks_4_adaLN_modulation_1_weight_to_fp16, x = input_423_cast_fp16)[name = tensor("linear_174_cast_fp16")]; + tensor var_2656_split_sizes_0 = const()[name = tensor("op_2656_split_sizes_0"), val = tensor([512, 512, 512])]; + tensor var_2656_axis_0 = const()[name = tensor("op_2656_axis_0"), val = tensor(-1)]; + tensor var_2656_cast_fp16_0, tensor var_2656_cast_fp16_1, tensor var_2656_cast_fp16_2 = split(axis = var_2656_axis_0, split_sizes = var_2656_split_sizes_0, x = linear_174_cast_fp16)[name = tensor("op_2656_cast_fp16")]; + tensor mean_93_axes_0 = const()[name = tensor("mean_93_axes_0"), val = tensor([-1])]; + tensor mean_93_keep_dims_0 = const()[name = tensor("mean_93_keep_dims_0"), val = tensor(true)]; + tensor mean_93_cast_fp16 = reduce_mean(axes = mean_93_axes_0, keep_dims = mean_93_keep_dims_0, x = x_293_cast_fp16)[name = tensor("mean_93_cast_fp16")]; + tensor sub_74_cast_fp16 = sub(x = x_293_cast_fp16, y = mean_93_cast_fp16)[name = tensor("sub_74_cast_fp16")]; + tensor square_60_cast_fp16 = square(x = sub_74_cast_fp16)[name = tensor("square_60_cast_fp16")]; + tensor reduce_mean_121_axes_0 = const()[name = tensor("reduce_mean_121_axes_0"), val = tensor([-1])]; + tensor reduce_mean_121_keep_dims_0 = const()[name = tensor("reduce_mean_121_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_121_cast_fp16 = reduce_mean(axes = reduce_mean_121_axes_0, keep_dims = reduce_mean_121_keep_dims_0, x = square_60_cast_fp16)[name = tensor("reduce_mean_121_cast_fp16")]; + tensor var_2666_to_fp16 = const()[name = tensor("op_2666_to_fp16"), val = tensor(0x1.1p-20)]; + tensor var_2667_cast_fp16 = add(x = reduce_mean_121_cast_fp16, y = var_2666_to_fp16)[name = tensor("op_2667_cast_fp16")]; + tensor var_2668_cast_fp16 = sqrt(x = var_2667_cast_fp16)[name = tensor("op_2668_cast_fp16")]; + tensor x_295_cast_fp16 = real_div(x = sub_74_cast_fp16, y = var_2668_cast_fp16)[name = tensor("x_295_cast_fp16")]; + tensor var_2670_cast_fp16 = mul(x = x_295_cast_fp16, y = flow_net_res_blocks_4_in_ln_weight_to_fp16)[name = tensor("op_2670_cast_fp16")]; + tensor x_297_cast_fp16 = add(x = var_2670_cast_fp16, y = flow_net_res_blocks_4_in_ln_bias_to_fp16)[name = tensor("x_297_cast_fp16")]; + tensor var_2672_promoted_to_fp16 = const()[name = tensor("op_2672_promoted_to_fp16"), val = tensor(0x1p+0)]; + tensor var_2673_cast_fp16 = add(x = var_2656_cast_fp16_1, y = var_2672_promoted_to_fp16)[name = tensor("op_2673_cast_fp16")]; + tensor var_2674_cast_fp16 = mul(x = x_297_cast_fp16, y = var_2673_cast_fp16)[name = tensor("op_2674_cast_fp16")]; + tensor input_457_cast_fp16 = add(x = var_2674_cast_fp16, y = var_2656_cast_fp16_0)[name = tensor("input_457_cast_fp16")]; + tensor linear_175_cast_fp16 = linear(bias = flow_net_res_blocks_4_mlp_0_bias_to_fp16, weight = flow_net_res_blocks_4_mlp_0_weight_to_fp16, x = input_457_cast_fp16)[name = tensor("linear_175_cast_fp16")]; + tensor input_461_cast_fp16 = silu(x = linear_175_cast_fp16)[name = tensor("input_461_cast_fp16")]; + tensor linear_176_cast_fp16 = linear(bias = flow_net_res_blocks_4_mlp_2_bias_to_fp16, weight = flow_net_res_blocks_4_mlp_2_weight_to_fp16, x = input_461_cast_fp16)[name = tensor("linear_176_cast_fp16")]; + tensor var_2685_cast_fp16 = mul(x = var_2656_cast_fp16_2, y = linear_176_cast_fp16)[name = tensor("op_2685_cast_fp16")]; + tensor x_299_cast_fp16 = add(x = x_293_cast_fp16, y = var_2685_cast_fp16)[name = tensor("x_299_cast_fp16")]; + tensor linear_177_cast_fp16 = linear(bias = flow_net_res_blocks_5_adaLN_modulation_1_bias_to_fp16, weight = flow_net_res_blocks_5_adaLN_modulation_1_weight_to_fp16, x = input_423_cast_fp16)[name = tensor("linear_177_cast_fp16")]; + tensor var_2695_split_sizes_0 = const()[name = tensor("op_2695_split_sizes_0"), val = tensor([512, 512, 512])]; + tensor var_2695_axis_0 = const()[name = tensor("op_2695_axis_0"), val = tensor(-1)]; + tensor var_2695_cast_fp16_0, tensor var_2695_cast_fp16_1, tensor var_2695_cast_fp16_2 = split(axis = var_2695_axis_0, split_sizes = var_2695_split_sizes_0, x = linear_177_cast_fp16)[name = tensor("op_2695_cast_fp16")]; + tensor mean_95_axes_0 = const()[name = tensor("mean_95_axes_0"), val = tensor([-1])]; + tensor mean_95_keep_dims_0 = const()[name = tensor("mean_95_keep_dims_0"), val = tensor(true)]; + tensor mean_95_cast_fp16 = reduce_mean(axes = mean_95_axes_0, keep_dims = mean_95_keep_dims_0, x = x_299_cast_fp16)[name = tensor("mean_95_cast_fp16")]; + tensor sub_75_cast_fp16 = sub(x = x_299_cast_fp16, y = mean_95_cast_fp16)[name = tensor("sub_75_cast_fp16")]; + tensor square_61_cast_fp16 = square(x = sub_75_cast_fp16)[name = tensor("square_61_cast_fp16")]; + tensor reduce_mean_123_axes_0 = const()[name = tensor("reduce_mean_123_axes_0"), val = tensor([-1])]; + tensor reduce_mean_123_keep_dims_0 = const()[name = tensor("reduce_mean_123_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_123_cast_fp16 = reduce_mean(axes = reduce_mean_123_axes_0, keep_dims = reduce_mean_123_keep_dims_0, x = square_61_cast_fp16)[name = tensor("reduce_mean_123_cast_fp16")]; + tensor var_2705_to_fp16 = const()[name = tensor("op_2705_to_fp16"), val = tensor(0x1.1p-20)]; + tensor var_2706_cast_fp16 = add(x = reduce_mean_123_cast_fp16, y = var_2705_to_fp16)[name = tensor("op_2706_cast_fp16")]; + tensor var_2707_cast_fp16 = sqrt(x = var_2706_cast_fp16)[name = tensor("op_2707_cast_fp16")]; + tensor x_301_cast_fp16 = real_div(x = sub_75_cast_fp16, y = var_2707_cast_fp16)[name = tensor("x_301_cast_fp16")]; + tensor var_2709_cast_fp16 = mul(x = x_301_cast_fp16, y = flow_net_res_blocks_5_in_ln_weight_to_fp16)[name = tensor("op_2709_cast_fp16")]; + tensor x_303_cast_fp16 = add(x = var_2709_cast_fp16, y = flow_net_res_blocks_5_in_ln_bias_to_fp16)[name = tensor("x_303_cast_fp16")]; + tensor var_2711_promoted_to_fp16 = const()[name = tensor("op_2711_promoted_to_fp16"), val = tensor(0x1p+0)]; + tensor var_2712_cast_fp16 = add(x = var_2695_cast_fp16_1, y = var_2711_promoted_to_fp16)[name = tensor("op_2712_cast_fp16")]; + tensor var_2713_cast_fp16 = mul(x = x_303_cast_fp16, y = var_2712_cast_fp16)[name = tensor("op_2713_cast_fp16")]; + tensor input_465_cast_fp16 = add(x = var_2713_cast_fp16, y = var_2695_cast_fp16_0)[name = tensor("input_465_cast_fp16")]; + tensor linear_178_cast_fp16 = linear(bias = flow_net_res_blocks_5_mlp_0_bias_to_fp16, weight = flow_net_res_blocks_5_mlp_0_weight_to_fp16, x = input_465_cast_fp16)[name = tensor("linear_178_cast_fp16")]; + tensor input_469_cast_fp16 = silu(x = linear_178_cast_fp16)[name = tensor("input_469_cast_fp16")]; + tensor linear_179_cast_fp16 = linear(bias = flow_net_res_blocks_5_mlp_2_bias_to_fp16, weight = flow_net_res_blocks_5_mlp_2_weight_to_fp16, x = input_469_cast_fp16)[name = tensor("linear_179_cast_fp16")]; + tensor var_2724_cast_fp16 = mul(x = var_2695_cast_fp16_2, y = linear_179_cast_fp16)[name = tensor("op_2724_cast_fp16")]; + tensor x_305_cast_fp16 = add(x = x_299_cast_fp16, y = var_2724_cast_fp16)[name = tensor("x_305_cast_fp16")]; + tensor linear_180_cast_fp16 = linear(bias = flow_net_final_layer_adaLN_modulation_1_bias_to_fp16, weight = flow_net_final_layer_adaLN_modulation_1_weight_to_fp16, x = input_423_cast_fp16)[name = tensor("linear_180_cast_fp16")]; + tensor var_2733_split_sizes_0 = const()[name = tensor("op_2733_split_sizes_0"), val = tensor([512, 512])]; + tensor var_2733_axis_0 = const()[name = tensor("op_2733_axis_0"), val = tensor(-1)]; + tensor var_2733_cast_fp16_0, tensor var_2733_cast_fp16_1 = split(axis = var_2733_axis_0, split_sizes = var_2733_split_sizes_0, x = linear_180_cast_fp16)[name = tensor("op_2733_cast_fp16")]; + tensor mean_97_axes_0 = const()[name = tensor("mean_97_axes_0"), val = tensor([-1])]; + tensor mean_97_keep_dims_0 = const()[name = tensor("mean_97_keep_dims_0"), val = tensor(true)]; + tensor mean_97_cast_fp16 = reduce_mean(axes = mean_97_axes_0, keep_dims = mean_97_keep_dims_0, x = x_305_cast_fp16)[name = tensor("mean_97_cast_fp16")]; + tensor sub_76_cast_fp16 = sub(x = x_305_cast_fp16, y = mean_97_cast_fp16)[name = tensor("sub_76_cast_fp16")]; + tensor square_62_cast_fp16 = square(x = sub_76_cast_fp16)[name = tensor("square_62_cast_fp16")]; + tensor reduce_mean_125_axes_0 = const()[name = tensor("reduce_mean_125_axes_0"), val = tensor([-1])]; + tensor reduce_mean_125_keep_dims_0 = const()[name = tensor("reduce_mean_125_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_125_cast_fp16 = reduce_mean(axes = reduce_mean_125_axes_0, keep_dims = reduce_mean_125_keep_dims_0, x = square_62_cast_fp16)[name = tensor("reduce_mean_125_cast_fp16")]; + tensor var_2740_to_fp16 = const()[name = tensor("op_2740_to_fp16"), val = tensor(0x1.1p-20)]; + tensor var_2741_cast_fp16 = add(x = reduce_mean_125_cast_fp16, y = var_2740_to_fp16)[name = tensor("op_2741_cast_fp16")]; + tensor var_2742_cast_fp16 = sqrt(x = var_2741_cast_fp16)[name = tensor("op_2742_cast_fp16")]; + tensor x_307_cast_fp16 = real_div(x = sub_76_cast_fp16, y = var_2742_cast_fp16)[name = tensor("x_307_cast_fp16")]; + tensor var_2744_promoted_to_fp16 = const()[name = tensor("op_2744_promoted_to_fp16"), val = tensor(0x1p+0)]; + tensor var_2745_cast_fp16 = add(x = var_2733_cast_fp16_1, y = var_2744_promoted_to_fp16)[name = tensor("op_2745_cast_fp16")]; + tensor var_2746_cast_fp16 = mul(x = x_307_cast_fp16, y = var_2745_cast_fp16)[name = tensor("op_2746_cast_fp16")]; + tensor input_473_cast_fp16 = add(x = var_2746_cast_fp16, y = var_2733_cast_fp16_0)[name = tensor("input_473_cast_fp16")]; + tensor linear_181_cast_fp16 = linear(bias = flow_net_final_layer_linear_bias_to_fp16, weight = flow_net_final_layer_linear_weight_to_fp16, x = input_473_cast_fp16)[name = tensor("linear_181_cast_fp16")]; + tensor var_2757_to_fp16 = const()[name = tensor("op_2757_to_fp16"), val = tensor(0x1p-3)]; + tensor var_2758_cast_fp16 = mul(x = linear_181_cast_fp16, y = var_2757_to_fp16)[name = tensor("op_2758_cast_fp16")]; + tensor input_475_cast_fp16 = add(x = input_407_cast_fp16, y = var_2758_cast_fp16)[name = tensor("input_475_cast_fp16")]; + tensor linear_182_cast_fp16 = linear(bias = flow_net_input_proj_bias_to_fp16, weight = flow_net_input_proj_weight_to_fp16, x = input_475_cast_fp16)[name = tensor("linear_182_cast_fp16")]; + tensor input_479_to_fp16 = const()[name = tensor("input_479_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(19010432)))]; + tensor input_481_cast_fp16 = silu(x = input_479_to_fp16)[name = tensor("input_481_cast_fp16")]; + tensor linear_184_cast_fp16 = linear(bias = flow_net_time_embed_0_mlp_2_bias_to_fp16, weight = flow_net_time_embed_0_mlp_2_weight_to_fp16, x = input_481_cast_fp16)[name = tensor("linear_184_cast_fp16")]; + tensor reduce_mean_126_axes_0 = const()[name = tensor("reduce_mean_126_axes_0"), val = tensor([-1])]; + tensor reduce_mean_126_keep_dims_0 = const()[name = tensor("reduce_mean_126_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_126_cast_fp16 = reduce_mean(axes = reduce_mean_126_axes_0, keep_dims = reduce_mean_126_keep_dims_0, x = linear_184_cast_fp16)[name = tensor("reduce_mean_126_cast_fp16")]; + tensor sub_77_cast_fp16 = sub(x = linear_184_cast_fp16, y = reduce_mean_126_cast_fp16)[name = tensor("sub_77_cast_fp16")]; + tensor square_63_cast_fp16 = square(x = sub_77_cast_fp16)[name = tensor("square_63_cast_fp16")]; + tensor reduce_mean_127_axes_0 = const()[name = tensor("reduce_mean_127_axes_0"), val = tensor([-1])]; + tensor reduce_mean_127_keep_dims_0 = const()[name = tensor("reduce_mean_127_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_127_cast_fp16 = reduce_mean(axes = reduce_mean_127_axes_0, keep_dims = reduce_mean_127_keep_dims_0, x = square_63_cast_fp16)[name = tensor("reduce_mean_127_cast_fp16")]; + tensor real_div_14_to_fp16 = const()[name = tensor("real_div_14_to_fp16"), val = tensor(0x1.008p+0)]; + tensor mul_14_cast_fp16 = mul(x = reduce_mean_127_cast_fp16, y = real_div_14_to_fp16)[name = tensor("mul_14_cast_fp16")]; + tensor var_2824_to_fp16 = const()[name = tensor("op_2824_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_127_cast_fp16_0 = add(x = mul_14_cast_fp16, y = var_2824_to_fp16)[name = tensor("var_127_cast_fp16")]; + tensor var_2827_epsilon_0 = const()[name = tensor("op_2827_epsilon_0"), val = tensor(0x1.197998p-40)]; + tensor var_2827_cast_fp16 = rsqrt(epsilon = var_2827_epsilon_0, x = var_127_cast_fp16_0)[name = tensor("op_2827_cast_fp16")]; + tensor var_2828_cast_fp16 = mul(x = const_3_to_fp16, y = var_2827_cast_fp16)[name = tensor("op_2828_cast_fp16")]; + tensor var_2829_cast_fp16 = mul(x = linear_184_cast_fp16, y = var_2828_cast_fp16)[name = tensor("op_2829_cast_fp16")]; + tensor input_485_to_fp16 = const()[name = tensor("input_485_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(19011520)))]; + tensor input_487_cast_fp16 = silu(x = input_485_to_fp16)[name = tensor("input_487_cast_fp16")]; + tensor linear_186_cast_fp16 = linear(bias = flow_net_time_embed_1_mlp_2_bias_to_fp16, weight = flow_net_time_embed_1_mlp_2_weight_to_fp16, x = input_487_cast_fp16)[name = tensor("linear_186_cast_fp16")]; + tensor reduce_mean_128_axes_0 = const()[name = tensor("reduce_mean_128_axes_0"), val = tensor([-1])]; + tensor reduce_mean_128_keep_dims_0 = const()[name = tensor("reduce_mean_128_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_128_cast_fp16 = reduce_mean(axes = reduce_mean_128_axes_0, keep_dims = reduce_mean_128_keep_dims_0, x = linear_186_cast_fp16)[name = tensor("reduce_mean_128_cast_fp16")]; + tensor sub_79_cast_fp16 = sub(x = linear_186_cast_fp16, y = reduce_mean_128_cast_fp16)[name = tensor("sub_79_cast_fp16")]; + tensor square_64_cast_fp16 = square(x = sub_79_cast_fp16)[name = tensor("square_64_cast_fp16")]; + tensor reduce_mean_129_axes_0 = const()[name = tensor("reduce_mean_129_axes_0"), val = tensor([-1])]; + tensor reduce_mean_129_keep_dims_0 = const()[name = tensor("reduce_mean_129_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_129_cast_fp16 = reduce_mean(axes = reduce_mean_129_axes_0, keep_dims = reduce_mean_129_keep_dims_0, x = square_64_cast_fp16)[name = tensor("reduce_mean_129_cast_fp16")]; + tensor real_div_15_to_fp16 = const()[name = tensor("real_div_15_to_fp16"), val = tensor(0x1.008p+0)]; + tensor mul_15_cast_fp16 = mul(x = reduce_mean_129_cast_fp16, y = real_div_15_to_fp16)[name = tensor("mul_15_cast_fp16")]; + tensor var_2851_to_fp16 = const()[name = tensor("op_2851_to_fp16"), val = tensor(0x1.5p-17)]; + tensor var_129_cast_fp16 = add(x = mul_15_cast_fp16, y = var_2851_to_fp16)[name = tensor("var_129_cast_fp16")]; + tensor var_2854_epsilon_0 = const()[name = tensor("op_2854_epsilon_0"), val = tensor(0x1.197998p-40)]; + tensor var_2854_cast_fp16 = rsqrt(epsilon = var_2854_epsilon_0, x = var_129_cast_fp16)[name = tensor("op_2854_cast_fp16")]; + tensor var_2855_cast_fp16 = mul(x = const_5_to_fp16, y = var_2854_cast_fp16)[name = tensor("op_2855_cast_fp16")]; + tensor var_2856_cast_fp16 = mul(x = linear_186_cast_fp16, y = var_2855_cast_fp16)[name = tensor("op_2856_cast_fp16")]; + tensor var_2858_cast_fp16 = add(x = var_2829_cast_fp16, y = var_2856_cast_fp16)[name = tensor("op_2858_cast_fp16")]; + tensor _inversed_t_combined_y_0_to_fp16 = const()[name = tensor("_inversed_t_combined_y_0_to_fp16"), val = tensor(0x1p-1)]; + tensor _inversed_t_combined_cast_fp16 = mul(x = var_2858_cast_fp16, y = _inversed_t_combined_y_0_to_fp16)[name = tensor("_inversed_t_combined_cast_fp16")]; + tensor input_489_cast_fp16 = add(x = _inversed_t_combined_cast_fp16, y = linear_5_cast_fp16)[name = tensor("input_489_cast_fp16")]; + tensor input_491_cast_fp16 = silu(x = input_489_cast_fp16)[name = tensor("input_491_cast_fp16")]; + tensor linear_188_cast_fp16 = linear(bias = flow_net_res_blocks_0_adaLN_modulation_1_bias_to_fp16, weight = flow_net_res_blocks_0_adaLN_modulation_1_weight_to_fp16, x = input_491_cast_fp16)[name = tensor("linear_188_cast_fp16")]; + tensor var_2873_split_sizes_0 = const()[name = tensor("op_2873_split_sizes_0"), val = tensor([512, 512, 512])]; + tensor var_2873_axis_0 = const()[name = tensor("op_2873_axis_0"), val = tensor(-1)]; + tensor var_2873_cast_fp16_0, tensor var_2873_cast_fp16_1, tensor var_2873_cast_fp16_2 = split(axis = var_2873_axis_0, split_sizes = var_2873_split_sizes_0, x = linear_188_cast_fp16)[name = tensor("op_2873_cast_fp16")]; + tensor mean_99_axes_0 = const()[name = tensor("mean_99_axes_0"), val = tensor([-1])]; + tensor mean_99_keep_dims_0 = const()[name = tensor("mean_99_keep_dims_0"), val = tensor(true)]; + tensor mean_99_cast_fp16 = reduce_mean(axes = mean_99_axes_0, keep_dims = mean_99_keep_dims_0, x = linear_182_cast_fp16)[name = tensor("mean_99_cast_fp16")]; + tensor sub_81_cast_fp16 = sub(x = linear_182_cast_fp16, y = mean_99_cast_fp16)[name = tensor("sub_81_cast_fp16")]; + tensor square_65_cast_fp16 = square(x = sub_81_cast_fp16)[name = tensor("square_65_cast_fp16")]; + tensor reduce_mean_131_axes_0 = const()[name = tensor("reduce_mean_131_axes_0"), val = tensor([-1])]; + tensor reduce_mean_131_keep_dims_0 = const()[name = tensor("reduce_mean_131_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_131_cast_fp16 = reduce_mean(axes = reduce_mean_131_axes_0, keep_dims = reduce_mean_131_keep_dims_0, x = square_65_cast_fp16)[name = tensor("reduce_mean_131_cast_fp16")]; + tensor var_2883_to_fp16 = const()[name = tensor("op_2883_to_fp16"), val = tensor(0x1.1p-20)]; + tensor var_2884_cast_fp16 = add(x = reduce_mean_131_cast_fp16, y = var_2883_to_fp16)[name = tensor("op_2884_cast_fp16")]; + tensor var_2885_cast_fp16 = sqrt(x = var_2884_cast_fp16)[name = tensor("op_2885_cast_fp16")]; + tensor x_315_cast_fp16 = real_div(x = sub_81_cast_fp16, y = var_2885_cast_fp16)[name = tensor("x_315_cast_fp16")]; + tensor var_2887_cast_fp16 = mul(x = x_315_cast_fp16, y = flow_net_res_blocks_0_in_ln_weight_to_fp16)[name = tensor("op_2887_cast_fp16")]; + tensor x_317_cast_fp16 = add(x = var_2887_cast_fp16, y = flow_net_res_blocks_0_in_ln_bias_to_fp16)[name = tensor("x_317_cast_fp16")]; + tensor var_2889_promoted_to_fp16 = const()[name = tensor("op_2889_promoted_to_fp16"), val = tensor(0x1p+0)]; + tensor var_2890_cast_fp16 = add(x = var_2873_cast_fp16_1, y = var_2889_promoted_to_fp16)[name = tensor("op_2890_cast_fp16")]; + tensor var_2891_cast_fp16 = mul(x = x_317_cast_fp16, y = var_2890_cast_fp16)[name = tensor("op_2891_cast_fp16")]; + tensor input_493_cast_fp16 = add(x = var_2891_cast_fp16, y = var_2873_cast_fp16_0)[name = tensor("input_493_cast_fp16")]; + tensor linear_189_cast_fp16 = linear(bias = flow_net_res_blocks_0_mlp_0_bias_to_fp16, weight = flow_net_res_blocks_0_mlp_0_weight_to_fp16, x = input_493_cast_fp16)[name = tensor("linear_189_cast_fp16")]; + tensor input_497_cast_fp16 = silu(x = linear_189_cast_fp16)[name = tensor("input_497_cast_fp16")]; + tensor linear_190_cast_fp16 = linear(bias = flow_net_res_blocks_0_mlp_2_bias_to_fp16, weight = flow_net_res_blocks_0_mlp_2_weight_to_fp16, x = input_497_cast_fp16)[name = tensor("linear_190_cast_fp16")]; + tensor var_2902_cast_fp16 = mul(x = var_2873_cast_fp16_2, y = linear_190_cast_fp16)[name = tensor("op_2902_cast_fp16")]; + tensor x_319_cast_fp16 = add(x = linear_182_cast_fp16, y = var_2902_cast_fp16)[name = tensor("x_319_cast_fp16")]; + tensor linear_191_cast_fp16 = linear(bias = flow_net_res_blocks_1_adaLN_modulation_1_bias_to_fp16, weight = flow_net_res_blocks_1_adaLN_modulation_1_weight_to_fp16, x = input_491_cast_fp16)[name = tensor("linear_191_cast_fp16")]; + tensor var_2912_split_sizes_0 = const()[name = tensor("op_2912_split_sizes_0"), val = tensor([512, 512, 512])]; + tensor var_2912_axis_0 = const()[name = tensor("op_2912_axis_0"), val = tensor(-1)]; + tensor var_2912_cast_fp16_0, tensor var_2912_cast_fp16_1, tensor var_2912_cast_fp16_2 = split(axis = var_2912_axis_0, split_sizes = var_2912_split_sizes_0, x = linear_191_cast_fp16)[name = tensor("op_2912_cast_fp16")]; + tensor mean_101_axes_0 = const()[name = tensor("mean_101_axes_0"), val = tensor([-1])]; + tensor mean_101_keep_dims_0 = const()[name = tensor("mean_101_keep_dims_0"), val = tensor(true)]; + tensor mean_101_cast_fp16 = reduce_mean(axes = mean_101_axes_0, keep_dims = mean_101_keep_dims_0, x = x_319_cast_fp16)[name = tensor("mean_101_cast_fp16")]; + tensor sub_82_cast_fp16 = sub(x = x_319_cast_fp16, y = mean_101_cast_fp16)[name = tensor("sub_82_cast_fp16")]; + tensor square_66_cast_fp16 = square(x = sub_82_cast_fp16)[name = tensor("square_66_cast_fp16")]; + tensor reduce_mean_133_axes_0 = const()[name = tensor("reduce_mean_133_axes_0"), val = tensor([-1])]; + tensor reduce_mean_133_keep_dims_0 = const()[name = tensor("reduce_mean_133_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_133_cast_fp16 = reduce_mean(axes = reduce_mean_133_axes_0, keep_dims = reduce_mean_133_keep_dims_0, x = square_66_cast_fp16)[name = tensor("reduce_mean_133_cast_fp16")]; + tensor var_2922_to_fp16 = const()[name = tensor("op_2922_to_fp16"), val = tensor(0x1.1p-20)]; + tensor var_2923_cast_fp16 = add(x = reduce_mean_133_cast_fp16, y = var_2922_to_fp16)[name = tensor("op_2923_cast_fp16")]; + tensor var_2924_cast_fp16 = sqrt(x = var_2923_cast_fp16)[name = tensor("op_2924_cast_fp16")]; + tensor x_321_cast_fp16 = real_div(x = sub_82_cast_fp16, y = var_2924_cast_fp16)[name = tensor("x_321_cast_fp16")]; + tensor var_2926_cast_fp16 = mul(x = x_321_cast_fp16, y = flow_net_res_blocks_1_in_ln_weight_to_fp16)[name = tensor("op_2926_cast_fp16")]; + tensor x_323_cast_fp16 = add(x = var_2926_cast_fp16, y = flow_net_res_blocks_1_in_ln_bias_to_fp16)[name = tensor("x_323_cast_fp16")]; + tensor var_2928_promoted_to_fp16 = const()[name = tensor("op_2928_promoted_to_fp16"), val = tensor(0x1p+0)]; + tensor var_2929_cast_fp16 = add(x = var_2912_cast_fp16_1, y = var_2928_promoted_to_fp16)[name = tensor("op_2929_cast_fp16")]; + tensor var_2930_cast_fp16 = mul(x = x_323_cast_fp16, y = var_2929_cast_fp16)[name = tensor("op_2930_cast_fp16")]; + tensor input_501_cast_fp16 = add(x = var_2930_cast_fp16, y = var_2912_cast_fp16_0)[name = tensor("input_501_cast_fp16")]; + tensor linear_192_cast_fp16 = linear(bias = flow_net_res_blocks_1_mlp_0_bias_to_fp16, weight = flow_net_res_blocks_1_mlp_0_weight_to_fp16, x = input_501_cast_fp16)[name = tensor("linear_192_cast_fp16")]; + tensor input_505_cast_fp16 = silu(x = linear_192_cast_fp16)[name = tensor("input_505_cast_fp16")]; + tensor linear_193_cast_fp16 = linear(bias = flow_net_res_blocks_1_mlp_2_bias_to_fp16, weight = flow_net_res_blocks_1_mlp_2_weight_to_fp16, x = input_505_cast_fp16)[name = tensor("linear_193_cast_fp16")]; + tensor var_2941_cast_fp16 = mul(x = var_2912_cast_fp16_2, y = linear_193_cast_fp16)[name = tensor("op_2941_cast_fp16")]; + tensor x_325_cast_fp16 = add(x = x_319_cast_fp16, y = var_2941_cast_fp16)[name = tensor("x_325_cast_fp16")]; + tensor linear_194_cast_fp16 = linear(bias = flow_net_res_blocks_2_adaLN_modulation_1_bias_to_fp16, weight = flow_net_res_blocks_2_adaLN_modulation_1_weight_to_fp16, x = input_491_cast_fp16)[name = tensor("linear_194_cast_fp16")]; + tensor var_2951_split_sizes_0 = const()[name = tensor("op_2951_split_sizes_0"), val = tensor([512, 512, 512])]; + tensor var_2951_axis_0 = const()[name = tensor("op_2951_axis_0"), val = tensor(-1)]; + tensor var_2951_cast_fp16_0, tensor var_2951_cast_fp16_1, tensor var_2951_cast_fp16_2 = split(axis = var_2951_axis_0, split_sizes = var_2951_split_sizes_0, x = linear_194_cast_fp16)[name = tensor("op_2951_cast_fp16")]; + tensor mean_103_axes_0 = const()[name = tensor("mean_103_axes_0"), val = tensor([-1])]; + tensor mean_103_keep_dims_0 = const()[name = tensor("mean_103_keep_dims_0"), val = tensor(true)]; + tensor mean_103_cast_fp16 = reduce_mean(axes = mean_103_axes_0, keep_dims = mean_103_keep_dims_0, x = x_325_cast_fp16)[name = tensor("mean_103_cast_fp16")]; + tensor sub_83_cast_fp16 = sub(x = x_325_cast_fp16, y = mean_103_cast_fp16)[name = tensor("sub_83_cast_fp16")]; + tensor square_67_cast_fp16 = square(x = sub_83_cast_fp16)[name = tensor("square_67_cast_fp16")]; + tensor reduce_mean_135_axes_0 = const()[name = tensor("reduce_mean_135_axes_0"), val = tensor([-1])]; + tensor reduce_mean_135_keep_dims_0 = const()[name = tensor("reduce_mean_135_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_135_cast_fp16 = reduce_mean(axes = reduce_mean_135_axes_0, keep_dims = reduce_mean_135_keep_dims_0, x = square_67_cast_fp16)[name = tensor("reduce_mean_135_cast_fp16")]; + tensor var_2961_to_fp16 = const()[name = tensor("op_2961_to_fp16"), val = tensor(0x1.1p-20)]; + tensor var_2962_cast_fp16 = add(x = reduce_mean_135_cast_fp16, y = var_2961_to_fp16)[name = tensor("op_2962_cast_fp16")]; + tensor var_2963_cast_fp16 = sqrt(x = var_2962_cast_fp16)[name = tensor("op_2963_cast_fp16")]; + tensor x_327_cast_fp16 = real_div(x = sub_83_cast_fp16, y = var_2963_cast_fp16)[name = tensor("x_327_cast_fp16")]; + tensor var_2965_cast_fp16 = mul(x = x_327_cast_fp16, y = flow_net_res_blocks_2_in_ln_weight_to_fp16)[name = tensor("op_2965_cast_fp16")]; + tensor x_329_cast_fp16 = add(x = var_2965_cast_fp16, y = flow_net_res_blocks_2_in_ln_bias_to_fp16)[name = tensor("x_329_cast_fp16")]; + tensor var_2967_promoted_to_fp16 = const()[name = tensor("op_2967_promoted_to_fp16"), val = tensor(0x1p+0)]; + tensor var_2968_cast_fp16 = add(x = var_2951_cast_fp16_1, y = var_2967_promoted_to_fp16)[name = tensor("op_2968_cast_fp16")]; + tensor var_2969_cast_fp16 = mul(x = x_329_cast_fp16, y = var_2968_cast_fp16)[name = tensor("op_2969_cast_fp16")]; + tensor input_509_cast_fp16 = add(x = var_2969_cast_fp16, y = var_2951_cast_fp16_0)[name = tensor("input_509_cast_fp16")]; + tensor linear_195_cast_fp16 = linear(bias = flow_net_res_blocks_2_mlp_0_bias_to_fp16, weight = flow_net_res_blocks_2_mlp_0_weight_to_fp16, x = input_509_cast_fp16)[name = tensor("linear_195_cast_fp16")]; + tensor input_513_cast_fp16 = silu(x = linear_195_cast_fp16)[name = tensor("input_513_cast_fp16")]; + tensor linear_196_cast_fp16 = linear(bias = flow_net_res_blocks_2_mlp_2_bias_to_fp16, weight = flow_net_res_blocks_2_mlp_2_weight_to_fp16, x = input_513_cast_fp16)[name = tensor("linear_196_cast_fp16")]; + tensor var_2980_cast_fp16 = mul(x = var_2951_cast_fp16_2, y = linear_196_cast_fp16)[name = tensor("op_2980_cast_fp16")]; + tensor x_331_cast_fp16 = add(x = x_325_cast_fp16, y = var_2980_cast_fp16)[name = tensor("x_331_cast_fp16")]; + tensor linear_197_cast_fp16 = linear(bias = flow_net_res_blocks_3_adaLN_modulation_1_bias_to_fp16, weight = flow_net_res_blocks_3_adaLN_modulation_1_weight_to_fp16, x = input_491_cast_fp16)[name = tensor("linear_197_cast_fp16")]; + tensor var_2990_split_sizes_0 = const()[name = tensor("op_2990_split_sizes_0"), val = tensor([512, 512, 512])]; + tensor var_2990_axis_0 = const()[name = tensor("op_2990_axis_0"), val = tensor(-1)]; + tensor var_2990_cast_fp16_0, tensor var_2990_cast_fp16_1, tensor var_2990_cast_fp16_2 = split(axis = var_2990_axis_0, split_sizes = var_2990_split_sizes_0, x = linear_197_cast_fp16)[name = tensor("op_2990_cast_fp16")]; + tensor mean_105_axes_0 = const()[name = tensor("mean_105_axes_0"), val = tensor([-1])]; + tensor mean_105_keep_dims_0 = const()[name = tensor("mean_105_keep_dims_0"), val = tensor(true)]; + tensor mean_105_cast_fp16 = reduce_mean(axes = mean_105_axes_0, keep_dims = mean_105_keep_dims_0, x = x_331_cast_fp16)[name = tensor("mean_105_cast_fp16")]; + tensor sub_84_cast_fp16 = sub(x = x_331_cast_fp16, y = mean_105_cast_fp16)[name = tensor("sub_84_cast_fp16")]; + tensor square_68_cast_fp16 = square(x = sub_84_cast_fp16)[name = tensor("square_68_cast_fp16")]; + tensor reduce_mean_137_axes_0 = const()[name = tensor("reduce_mean_137_axes_0"), val = tensor([-1])]; + tensor reduce_mean_137_keep_dims_0 = const()[name = tensor("reduce_mean_137_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_137_cast_fp16 = reduce_mean(axes = reduce_mean_137_axes_0, keep_dims = reduce_mean_137_keep_dims_0, x = square_68_cast_fp16)[name = tensor("reduce_mean_137_cast_fp16")]; + tensor var_3000_to_fp16 = const()[name = tensor("op_3000_to_fp16"), val = tensor(0x1.1p-20)]; + tensor var_3001_cast_fp16 = add(x = reduce_mean_137_cast_fp16, y = var_3000_to_fp16)[name = tensor("op_3001_cast_fp16")]; + tensor var_3002_cast_fp16 = sqrt(x = var_3001_cast_fp16)[name = tensor("op_3002_cast_fp16")]; + tensor x_333_cast_fp16 = real_div(x = sub_84_cast_fp16, y = var_3002_cast_fp16)[name = tensor("x_333_cast_fp16")]; + tensor var_3004_cast_fp16 = mul(x = x_333_cast_fp16, y = flow_net_res_blocks_3_in_ln_weight_to_fp16)[name = tensor("op_3004_cast_fp16")]; + tensor x_335_cast_fp16 = add(x = var_3004_cast_fp16, y = flow_net_res_blocks_3_in_ln_bias_to_fp16)[name = tensor("x_335_cast_fp16")]; + tensor var_3006_promoted_to_fp16 = const()[name = tensor("op_3006_promoted_to_fp16"), val = tensor(0x1p+0)]; + tensor var_3007_cast_fp16 = add(x = var_2990_cast_fp16_1, y = var_3006_promoted_to_fp16)[name = tensor("op_3007_cast_fp16")]; + tensor var_3008_cast_fp16 = mul(x = x_335_cast_fp16, y = var_3007_cast_fp16)[name = tensor("op_3008_cast_fp16")]; + tensor input_517_cast_fp16 = add(x = var_3008_cast_fp16, y = var_2990_cast_fp16_0)[name = tensor("input_517_cast_fp16")]; + tensor linear_198_cast_fp16 = linear(bias = flow_net_res_blocks_3_mlp_0_bias_to_fp16, weight = flow_net_res_blocks_3_mlp_0_weight_to_fp16, x = input_517_cast_fp16)[name = tensor("linear_198_cast_fp16")]; + tensor input_521_cast_fp16 = silu(x = linear_198_cast_fp16)[name = tensor("input_521_cast_fp16")]; + tensor linear_199_cast_fp16 = linear(bias = flow_net_res_blocks_3_mlp_2_bias_to_fp16, weight = flow_net_res_blocks_3_mlp_2_weight_to_fp16, x = input_521_cast_fp16)[name = tensor("linear_199_cast_fp16")]; + tensor var_3019_cast_fp16 = mul(x = var_2990_cast_fp16_2, y = linear_199_cast_fp16)[name = tensor("op_3019_cast_fp16")]; + tensor x_337_cast_fp16 = add(x = x_331_cast_fp16, y = var_3019_cast_fp16)[name = tensor("x_337_cast_fp16")]; + tensor linear_200_cast_fp16 = linear(bias = flow_net_res_blocks_4_adaLN_modulation_1_bias_to_fp16, weight = flow_net_res_blocks_4_adaLN_modulation_1_weight_to_fp16, x = input_491_cast_fp16)[name = tensor("linear_200_cast_fp16")]; + tensor var_3029_split_sizes_0 = const()[name = tensor("op_3029_split_sizes_0"), val = tensor([512, 512, 512])]; + tensor var_3029_axis_0 = const()[name = tensor("op_3029_axis_0"), val = tensor(-1)]; + tensor var_3029_cast_fp16_0, tensor var_3029_cast_fp16_1, tensor var_3029_cast_fp16_2 = split(axis = var_3029_axis_0, split_sizes = var_3029_split_sizes_0, x = linear_200_cast_fp16)[name = tensor("op_3029_cast_fp16")]; + tensor mean_107_axes_0 = const()[name = tensor("mean_107_axes_0"), val = tensor([-1])]; + tensor mean_107_keep_dims_0 = const()[name = tensor("mean_107_keep_dims_0"), val = tensor(true)]; + tensor mean_107_cast_fp16 = reduce_mean(axes = mean_107_axes_0, keep_dims = mean_107_keep_dims_0, x = x_337_cast_fp16)[name = tensor("mean_107_cast_fp16")]; + tensor sub_85_cast_fp16 = sub(x = x_337_cast_fp16, y = mean_107_cast_fp16)[name = tensor("sub_85_cast_fp16")]; + tensor square_69_cast_fp16 = square(x = sub_85_cast_fp16)[name = tensor("square_69_cast_fp16")]; + tensor reduce_mean_139_axes_0 = const()[name = tensor("reduce_mean_139_axes_0"), val = tensor([-1])]; + tensor reduce_mean_139_keep_dims_0 = const()[name = tensor("reduce_mean_139_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_139_cast_fp16 = reduce_mean(axes = reduce_mean_139_axes_0, keep_dims = reduce_mean_139_keep_dims_0, x = square_69_cast_fp16)[name = tensor("reduce_mean_139_cast_fp16")]; + tensor var_3039_to_fp16 = const()[name = tensor("op_3039_to_fp16"), val = tensor(0x1.1p-20)]; + tensor var_3040_cast_fp16 = add(x = reduce_mean_139_cast_fp16, y = var_3039_to_fp16)[name = tensor("op_3040_cast_fp16")]; + tensor var_3041_cast_fp16 = sqrt(x = var_3040_cast_fp16)[name = tensor("op_3041_cast_fp16")]; + tensor x_339_cast_fp16 = real_div(x = sub_85_cast_fp16, y = var_3041_cast_fp16)[name = tensor("x_339_cast_fp16")]; + tensor var_3043_cast_fp16 = mul(x = x_339_cast_fp16, y = flow_net_res_blocks_4_in_ln_weight_to_fp16)[name = tensor("op_3043_cast_fp16")]; + tensor x_341_cast_fp16 = add(x = var_3043_cast_fp16, y = flow_net_res_blocks_4_in_ln_bias_to_fp16)[name = tensor("x_341_cast_fp16")]; + tensor var_3045_promoted_to_fp16 = const()[name = tensor("op_3045_promoted_to_fp16"), val = tensor(0x1p+0)]; + tensor var_3046_cast_fp16 = add(x = var_3029_cast_fp16_1, y = var_3045_promoted_to_fp16)[name = tensor("op_3046_cast_fp16")]; + tensor var_3047_cast_fp16 = mul(x = x_341_cast_fp16, y = var_3046_cast_fp16)[name = tensor("op_3047_cast_fp16")]; + tensor input_525_cast_fp16 = add(x = var_3047_cast_fp16, y = var_3029_cast_fp16_0)[name = tensor("input_525_cast_fp16")]; + tensor linear_201_cast_fp16 = linear(bias = flow_net_res_blocks_4_mlp_0_bias_to_fp16, weight = flow_net_res_blocks_4_mlp_0_weight_to_fp16, x = input_525_cast_fp16)[name = tensor("linear_201_cast_fp16")]; + tensor input_529_cast_fp16 = silu(x = linear_201_cast_fp16)[name = tensor("input_529_cast_fp16")]; + tensor linear_202_cast_fp16 = linear(bias = flow_net_res_blocks_4_mlp_2_bias_to_fp16, weight = flow_net_res_blocks_4_mlp_2_weight_to_fp16, x = input_529_cast_fp16)[name = tensor("linear_202_cast_fp16")]; + tensor var_3058_cast_fp16 = mul(x = var_3029_cast_fp16_2, y = linear_202_cast_fp16)[name = tensor("op_3058_cast_fp16")]; + tensor x_343_cast_fp16 = add(x = x_337_cast_fp16, y = var_3058_cast_fp16)[name = tensor("x_343_cast_fp16")]; + tensor linear_203_cast_fp16 = linear(bias = flow_net_res_blocks_5_adaLN_modulation_1_bias_to_fp16, weight = flow_net_res_blocks_5_adaLN_modulation_1_weight_to_fp16, x = input_491_cast_fp16)[name = tensor("linear_203_cast_fp16")]; + tensor var_3068_split_sizes_0 = const()[name = tensor("op_3068_split_sizes_0"), val = tensor([512, 512, 512])]; + tensor var_3068_axis_0 = const()[name = tensor("op_3068_axis_0"), val = tensor(-1)]; + tensor var_3068_cast_fp16_0, tensor var_3068_cast_fp16_1, tensor var_3068_cast_fp16_2 = split(axis = var_3068_axis_0, split_sizes = var_3068_split_sizes_0, x = linear_203_cast_fp16)[name = tensor("op_3068_cast_fp16")]; + tensor mean_109_axes_0 = const()[name = tensor("mean_109_axes_0"), val = tensor([-1])]; + tensor mean_109_keep_dims_0 = const()[name = tensor("mean_109_keep_dims_0"), val = tensor(true)]; + tensor mean_109_cast_fp16 = reduce_mean(axes = mean_109_axes_0, keep_dims = mean_109_keep_dims_0, x = x_343_cast_fp16)[name = tensor("mean_109_cast_fp16")]; + tensor sub_86_cast_fp16 = sub(x = x_343_cast_fp16, y = mean_109_cast_fp16)[name = tensor("sub_86_cast_fp16")]; + tensor square_70_cast_fp16 = square(x = sub_86_cast_fp16)[name = tensor("square_70_cast_fp16")]; + tensor reduce_mean_141_axes_0 = const()[name = tensor("reduce_mean_141_axes_0"), val = tensor([-1])]; + tensor reduce_mean_141_keep_dims_0 = const()[name = tensor("reduce_mean_141_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_141_cast_fp16 = reduce_mean(axes = reduce_mean_141_axes_0, keep_dims = reduce_mean_141_keep_dims_0, x = square_70_cast_fp16)[name = tensor("reduce_mean_141_cast_fp16")]; + tensor var_3078_to_fp16 = const()[name = tensor("op_3078_to_fp16"), val = tensor(0x1.1p-20)]; + tensor var_3079_cast_fp16 = add(x = reduce_mean_141_cast_fp16, y = var_3078_to_fp16)[name = tensor("op_3079_cast_fp16")]; + tensor var_3080_cast_fp16 = sqrt(x = var_3079_cast_fp16)[name = tensor("op_3080_cast_fp16")]; + tensor x_345_cast_fp16 = real_div(x = sub_86_cast_fp16, y = var_3080_cast_fp16)[name = tensor("x_345_cast_fp16")]; + tensor var_3082_cast_fp16 = mul(x = x_345_cast_fp16, y = flow_net_res_blocks_5_in_ln_weight_to_fp16)[name = tensor("op_3082_cast_fp16")]; + tensor x_347_cast_fp16 = add(x = var_3082_cast_fp16, y = flow_net_res_blocks_5_in_ln_bias_to_fp16)[name = tensor("x_347_cast_fp16")]; + tensor var_3084_promoted_to_fp16 = const()[name = tensor("op_3084_promoted_to_fp16"), val = tensor(0x1p+0)]; + tensor var_3085_cast_fp16 = add(x = var_3068_cast_fp16_1, y = var_3084_promoted_to_fp16)[name = tensor("op_3085_cast_fp16")]; + tensor var_3086_cast_fp16 = mul(x = x_347_cast_fp16, y = var_3085_cast_fp16)[name = tensor("op_3086_cast_fp16")]; + tensor input_533_cast_fp16 = add(x = var_3086_cast_fp16, y = var_3068_cast_fp16_0)[name = tensor("input_533_cast_fp16")]; + tensor linear_204_cast_fp16 = linear(bias = flow_net_res_blocks_5_mlp_0_bias_to_fp16, weight = flow_net_res_blocks_5_mlp_0_weight_to_fp16, x = input_533_cast_fp16)[name = tensor("linear_204_cast_fp16")]; + tensor input_537_cast_fp16 = silu(x = linear_204_cast_fp16)[name = tensor("input_537_cast_fp16")]; + tensor linear_205_cast_fp16 = linear(bias = flow_net_res_blocks_5_mlp_2_bias_to_fp16, weight = flow_net_res_blocks_5_mlp_2_weight_to_fp16, x = input_537_cast_fp16)[name = tensor("linear_205_cast_fp16")]; + tensor var_3097_cast_fp16 = mul(x = var_3068_cast_fp16_2, y = linear_205_cast_fp16)[name = tensor("op_3097_cast_fp16")]; + tensor x_349_cast_fp16 = add(x = x_343_cast_fp16, y = var_3097_cast_fp16)[name = tensor("x_349_cast_fp16")]; + tensor linear_206_cast_fp16 = linear(bias = flow_net_final_layer_adaLN_modulation_1_bias_to_fp16, weight = flow_net_final_layer_adaLN_modulation_1_weight_to_fp16, x = input_491_cast_fp16)[name = tensor("linear_206_cast_fp16")]; + tensor var_3106_split_sizes_0 = const()[name = tensor("op_3106_split_sizes_0"), val = tensor([512, 512])]; + tensor var_3106_axis_0 = const()[name = tensor("op_3106_axis_0"), val = tensor(-1)]; + tensor var_3106_cast_fp16_0, tensor var_3106_cast_fp16_1 = split(axis = var_3106_axis_0, split_sizes = var_3106_split_sizes_0, x = linear_206_cast_fp16)[name = tensor("op_3106_cast_fp16")]; + tensor mean_axes_0 = const()[name = tensor("mean_axes_0"), val = tensor([-1])]; + tensor mean_keep_dims_0 = const()[name = tensor("mean_keep_dims_0"), val = tensor(true)]; + tensor mean_cast_fp16 = reduce_mean(axes = mean_axes_0, keep_dims = mean_keep_dims_0, x = x_349_cast_fp16)[name = tensor("mean_cast_fp16")]; + tensor sub_87_cast_fp16 = sub(x = x_349_cast_fp16, y = mean_cast_fp16)[name = tensor("sub_87_cast_fp16")]; + tensor square_71_cast_fp16 = square(x = sub_87_cast_fp16)[name = tensor("square_71_cast_fp16")]; + tensor reduce_mean_143_axes_0 = const()[name = tensor("reduce_mean_143_axes_0"), val = tensor([-1])]; + tensor reduce_mean_143_keep_dims_0 = const()[name = tensor("reduce_mean_143_keep_dims_0"), val = tensor(true)]; + tensor reduce_mean_143_cast_fp16 = reduce_mean(axes = reduce_mean_143_axes_0, keep_dims = reduce_mean_143_keep_dims_0, x = square_71_cast_fp16)[name = tensor("reduce_mean_143_cast_fp16")]; + tensor var_3113_to_fp16 = const()[name = tensor("op_3113_to_fp16"), val = tensor(0x1.1p-20)]; + tensor var_3114_cast_fp16 = add(x = reduce_mean_143_cast_fp16, y = var_3113_to_fp16)[name = tensor("op_3114_cast_fp16")]; + tensor var_3115_cast_fp16 = sqrt(x = var_3114_cast_fp16)[name = tensor("op_3115_cast_fp16")]; + tensor x_cast_fp16 = real_div(x = sub_87_cast_fp16, y = var_3115_cast_fp16)[name = tensor("x_cast_fp16")]; + tensor var_3117_promoted_to_fp16 = const()[name = tensor("op_3117_promoted_to_fp16"), val = tensor(0x1p+0)]; + tensor var_3118_cast_fp16 = add(x = var_3106_cast_fp16_1, y = var_3117_promoted_to_fp16)[name = tensor("op_3118_cast_fp16")]; + tensor var_3119_cast_fp16 = mul(x = x_cast_fp16, y = var_3118_cast_fp16)[name = tensor("op_3119_cast_fp16")]; + tensor input_cast_fp16 = add(x = var_3119_cast_fp16, y = var_3106_cast_fp16_0)[name = tensor("input_cast_fp16")]; + tensor linear_207_cast_fp16 = linear(bias = flow_net_final_layer_linear_bias_to_fp16, weight = flow_net_final_layer_linear_weight_to_fp16, x = input_cast_fp16)[name = tensor("linear_207_cast_fp16")]; + tensor var_3124_to_fp16 = const()[name = tensor("op_3124_to_fp16"), val = tensor(0x1p-3)]; + tensor var_3125_cast_fp16 = mul(x = linear_207_cast_fp16, y = var_3124_to_fp16)[name = tensor("op_3125_cast_fp16")]; + tensor var_3127_cast_fp16 = add(x = input_475_cast_fp16, y = var_3125_cast_fp16)[name = tensor("op_3127_cast_fp16")]; + tensor var_3127_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_3127_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor latent_final = cast(dtype = var_3127_cast_fp16_to_fp32_dtype_0, x = var_3127_cast_fp16)[name = tensor("cast_144")]; + } -> (latent_final); +} \ No newline at end of file diff --git a/v2.1/spanish_24l/flow_decoder_fused.mlmodelc/weights/weight.bin b/v2.1/spanish_24l/flow_decoder_fused.mlmodelc/weights/weight.bin new file mode 100644 index 0000000000000000000000000000000000000000..dfd39700353f5054362fffed76322ddd3ef07910 --- /dev/null +++ b/v2.1/spanish_24l/flow_decoder_fused.mlmodelc/weights/weight.bin @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:83ee3374f6946210512b80b904e534ae6224e3cb72402ab3b80924cbff7d35eb +size 19012608 diff --git a/v2.1/spanish_24l/flow_decoder_fused.mlpackage/Data/com.apple.CoreML/model.mlmodel b/v2.1/spanish_24l/flow_decoder_fused.mlpackage/Data/com.apple.CoreML/model.mlmodel new file mode 100644 index 0000000000000000000000000000000000000000..462e4fbb84349b657d20340fb372e30d0b2ffef3 --- /dev/null +++ b/v2.1/spanish_24l/flow_decoder_fused.mlpackage/Data/com.apple.CoreML/model.mlmodel @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:44dd6c7f169674d27cd30f4ebd847864d5a0a440c4ac3de39e637f8634baf7d8 +size 301733 diff --git a/v2.1/spanish_24l/flow_decoder_fused.mlpackage/Data/com.apple.CoreML/weights/weight.bin b/v2.1/spanish_24l/flow_decoder_fused.mlpackage/Data/com.apple.CoreML/weights/weight.bin new file mode 100644 index 0000000000000000000000000000000000000000..dfd39700353f5054362fffed76322ddd3ef07910 --- /dev/null +++ b/v2.1/spanish_24l/flow_decoder_fused.mlpackage/Data/com.apple.CoreML/weights/weight.bin @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:83ee3374f6946210512b80b904e534ae6224e3cb72402ab3b80924cbff7d35eb +size 19012608 diff --git a/v2.1/spanish_24l/flow_decoder_fused.mlpackage/Manifest.json b/v2.1/spanish_24l/flow_decoder_fused.mlpackage/Manifest.json new file mode 100644 index 0000000000000000000000000000000000000000..45c61a3a9d42a5e0979ab1ce5663409def5dd893 --- /dev/null +++ b/v2.1/spanish_24l/flow_decoder_fused.mlpackage/Manifest.json @@ -0,0 +1,18 @@ +{ + "fileFormatVersion": "1.0.0", + "itemInfoEntries": { + "32D3ED37-63C5-42C3-9567-E0EDCA83861D": { + "author": "com.apple.CoreML", + "description": "CoreML Model Specification", + "name": "model.mlmodel", + "path": "com.apple.CoreML/model.mlmodel" + }, + "6C829322-C177-4EB9-AB7D-266D6E58E4F8": { + "author": "com.apple.CoreML", + "description": "CoreML Model Weights", + "name": "weights", + "path": "com.apple.CoreML/weights" + } + }, + "rootModelIdentifier": "32D3ED37-63C5-42C3-9567-E0EDCA83861D" +} diff --git a/v2.1/spanish_24l/flowlm_step.mlmodelc/analytics/coremldata.bin b/v2.1/spanish_24l/flowlm_step.mlmodelc/analytics/coremldata.bin new file mode 100644 index 0000000000000000000000000000000000000000..30df004d6f3ba5bc8535b209c1a702c9eb697b0b --- /dev/null +++ b/v2.1/spanish_24l/flowlm_step.mlmodelc/analytics/coremldata.bin @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:c01dde6de57db10638fe0c1ab8aa93478422e3dc7282245bbde91c21664aebc2 +size 243 diff --git a/v2.1/spanish_24l/flowlm_step.mlmodelc/coremldata.bin b/v2.1/spanish_24l/flowlm_step.mlmodelc/coremldata.bin new file mode 100644 index 0000000000000000000000000000000000000000..7e0817ecc0ee1c91e82478bde02cf0c0b6681e36 --- /dev/null +++ b/v2.1/spanish_24l/flowlm_step.mlmodelc/coremldata.bin @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:01d0728c1cc614210371e48471f0b2054a7306c64f9b6bd60b23fabe1a982f1a +size 3559 diff --git a/v2.1/spanish_24l/flowlm_step.mlmodelc/model.mil b/v2.1/spanish_24l/flowlm_step.mlmodelc/model.mil new file mode 100644 index 0000000000000000000000000000000000000000..653a884fdb32dd43dfe7d26c9cad9bb7dea939c8 --- /dev/null +++ b/v2.1/spanish_24l/flowlm_step.mlmodelc/model.mil @@ -0,0 +1,5337 @@ +program(1.0) +[buildInfo = dict, tensor>({{"coremlc-component-MIL", "3520.4.1"}, {"coremlc-version", "3520.5.1"}, {"coremltools-component-torch", "2.12.0"}, {"coremltools-source-dialect", "TorchScript"}, {"coremltools-version", "9.0"}})] +{ + func main(tensor bos_emb, tensor cache0, tensor cache1, tensor cache10, tensor cache11, tensor cache12, tensor cache13, tensor cache14, tensor cache15, tensor cache16, tensor cache17, tensor cache18, tensor cache19, tensor cache2, tensor cache20, tensor cache21, tensor cache22, tensor cache23, tensor cache3, tensor cache4, tensor cache5, tensor cache6, tensor cache7, tensor cache8, tensor cache9, tensor position0, tensor position1, tensor position10, tensor position11, tensor position12, tensor position13, tensor position14, tensor position15, tensor position16, tensor position17, tensor position18, tensor position19, tensor position2, tensor position20, tensor position21, tensor position22, tensor position23, tensor position3, tensor position4, tensor position5, tensor position6, tensor position7, tensor position8, tensor position9, tensor sequence) { + tensor sequence_to_fp16_dtype_0 = const()[name = tensor("sequence_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor sequence_to_fp16 = cast(dtype = sequence_to_fp16_dtype_0, x = sequence)[name = tensor("cast_555")]; + tensor var_198_cast_fp16 = not_equal(x = sequence_to_fp16, y = sequence_to_fp16)[name = tensor("op_198_cast_fp16")]; + tensor expand_dims_0_axes_0 = const()[name = tensor("expand_dims_0_axes_0"), val = tensor([0, 1])]; + tensor bos_emb_to_fp16_dtype_0 = const()[name = tensor("bos_emb_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor bos_emb_to_fp16 = cast(dtype = bos_emb_to_fp16_dtype_0, x = bos_emb)[name = tensor("cast_554")]; + tensor expand_dims_0_cast_fp16 = expand_dims(axes = expand_dims_0_axes_0, x = bos_emb_to_fp16)[name = tensor("expand_dims_0_cast_fp16")]; + tensor input_1_cast_fp16 = select(a = expand_dims_0_cast_fp16, b = sequence_to_fp16, cond = var_198_cast_fp16)[name = tensor("input_1_cast_fp16")]; + tensor input_linear_weight_to_fp16 = const()[name = tensor("input_linear_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(64)))]; + tensor linear_0_bias_0_to_fp16 = const()[name = tensor("linear_0_bias_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(65664)))]; + tensor linear_0_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = input_linear_weight_to_fp16, x = input_1_cast_fp16)[name = tensor("linear_0_cast_fp16")]; + tensor input_5_axes_0 = const()[name = tensor("input_5_axes_0"), val = tensor([-1])]; + tensor norm0_1_weight_to_fp16 = const()[name = tensor("norm0_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(67776)))]; + tensor norm0_1_bias_to_fp16 = const()[name = tensor("norm0_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(69888)))]; + tensor var_204_to_fp16 = const()[name = tensor("op_204_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_5_cast_fp16 = layer_norm(axes = input_5_axes_0, beta = norm0_1_bias_to_fp16, epsilon = var_204_to_fp16, gamma = norm0_1_weight_to_fp16, x = linear_0_cast_fp16)[name = tensor("input_5_cast_fp16")]; + tensor attn0_in_proj_weight_to_fp16 = const()[name = tensor("attn0_in_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(72000)))]; + tensor linear_1_bias_0_to_fp16 = const()[name = tensor("linear_1_bias_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(6363520)))]; + tensor linear_1_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = attn0_in_proj_weight_to_fp16, x = input_5_cast_fp16)[name = tensor("linear_1_cast_fp16")]; + tensor var_218 = const()[name = tensor("op_218"), val = tensor([1, 1, 3, 16, 64])]; + tensor qkv_1_cast_fp16 = reshape(shape = var_218, x = linear_1_cast_fp16)[name = tensor("qkv_1_cast_fp16")]; + tensor q_1_begin_0 = const()[name = tensor("q_1_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor q_1_end_0 = const()[name = tensor("q_1_end_0"), val = tensor([1, 1, 1, 16, 64])]; + tensor q_1_end_mask_0 = const()[name = tensor("q_1_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor q_1_squeeze_mask_0 = const()[name = tensor("q_1_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor q_1_cast_fp16 = slice_by_index(begin = q_1_begin_0, end = q_1_end_0, end_mask = q_1_end_mask_0, squeeze_mask = q_1_squeeze_mask_0, x = qkv_1_cast_fp16)[name = tensor("q_1_cast_fp16")]; + tensor k_1_begin_0 = const()[name = tensor("k_1_begin_0"), val = tensor([0, 0, 1, 0, 0])]; + tensor k_1_end_0 = const()[name = tensor("k_1_end_0"), val = tensor([1, 1, 2, 16, 64])]; + tensor k_1_end_mask_0 = const()[name = tensor("k_1_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor k_1_squeeze_mask_0 = const()[name = tensor("k_1_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor k_1_cast_fp16 = slice_by_index(begin = k_1_begin_0, end = k_1_end_0, end_mask = k_1_end_mask_0, squeeze_mask = k_1_squeeze_mask_0, x = qkv_1_cast_fp16)[name = tensor("k_1_cast_fp16")]; + tensor v_1_begin_0 = const()[name = tensor("v_1_begin_0"), val = tensor([0, 0, 2, 0, 0])]; + tensor v_1_end_0 = const()[name = tensor("v_1_end_0"), val = tensor([1, 1, 3, 16, 64])]; + tensor v_1_end_mask_0 = const()[name = tensor("v_1_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor v_1_squeeze_mask_0 = const()[name = tensor("v_1_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor v_1_cast_fp16 = slice_by_index(begin = v_1_begin_0, end = v_1_end_0, end_mask = v_1_end_mask_0, squeeze_mask = v_1_squeeze_mask_0, x = qkv_1_cast_fp16)[name = tensor("v_1_cast_fp16")]; + tensor var_284 = const()[name = tensor("op_284"), val = tensor([1, 1, 1, 1])]; + tensor position0_to_fp16_dtype_0 = const()[name = tensor("position0_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor position0_to_fp16 = cast(dtype = position0_to_fp16_dtype_0, x = position0)[name = tensor("cast_553")]; + tensor ts_5_cast_fp16 = reshape(shape = var_284, x = position0_to_fp16)[name = tensor("ts_5_cast_fp16")]; + tensor var_291 = const()[name = tensor("op_291"), val = tensor([1, 1, 16, 32, 2])]; + tensor q_complex_1_cast_fp16 = reshape(shape = var_291, x = q_1_cast_fp16)[name = tensor("q_complex_1_cast_fp16")]; + tensor var_298 = const()[name = tensor("op_298"), val = tensor([1, 1, 16, 32, 2])]; + tensor k_complex_1_cast_fp16 = reshape(shape = var_298, x = k_1_cast_fp16)[name = tensor("k_complex_1_cast_fp16")]; + tensor var_302_begin_0 = const()[name = tensor("op_302_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_302_end_0 = const()[name = tensor("op_302_end_0"), val = tensor([1, 1, 16, 32, 1])]; + tensor var_302_end_mask_0 = const()[name = tensor("op_302_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_302_squeeze_mask_0 = const()[name = tensor("op_302_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_302_cast_fp16 = slice_by_index(begin = var_302_begin_0, end = var_302_end_0, end_mask = var_302_end_mask_0, squeeze_mask = var_302_squeeze_mask_0, x = q_complex_1_cast_fp16)[name = tensor("op_302_cast_fp16")]; + tensor var_310_begin_0 = const()[name = tensor("op_310_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_310_end_0 = const()[name = tensor("op_310_end_0"), val = tensor([1, 1, 16, 32, 2])]; + tensor var_310_end_mask_0 = const()[name = tensor("op_310_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_310_squeeze_mask_0 = const()[name = tensor("op_310_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_310_cast_fp16 = slice_by_index(begin = var_310_begin_0, end = var_310_end_0, end_mask = var_310_end_mask_0, squeeze_mask = var_310_squeeze_mask_0, x = q_complex_1_cast_fp16)[name = tensor("op_310_cast_fp16")]; + tensor var_318_begin_0 = const()[name = tensor("op_318_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_318_end_0 = const()[name = tensor("op_318_end_0"), val = tensor([1, 1, 16, 32, 1])]; + tensor var_318_end_mask_0 = const()[name = tensor("op_318_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_318_squeeze_mask_0 = const()[name = tensor("op_318_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_318_cast_fp16 = slice_by_index(begin = var_318_begin_0, end = var_318_end_0, end_mask = var_318_end_mask_0, squeeze_mask = var_318_squeeze_mask_0, x = k_complex_1_cast_fp16)[name = tensor("op_318_cast_fp16")]; + tensor var_326_begin_0 = const()[name = tensor("op_326_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_326_end_0 = const()[name = tensor("op_326_end_0"), val = tensor([1, 1, 16, 32, 2])]; + tensor var_326_end_mask_0 = const()[name = tensor("op_326_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_326_squeeze_mask_0 = const()[name = tensor("op_326_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_326_cast_fp16 = slice_by_index(begin = var_326_begin_0, end = var_326_end_0, end_mask = var_326_end_mask_0, squeeze_mask = var_326_squeeze_mask_0, x = k_complex_1_cast_fp16)[name = tensor("op_326_cast_fp16")]; + tensor freqs_1_to_fp16 = const()[name = tensor("freqs_1_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(6369728)))]; + tensor var_332_cast_fp16 = mul(x = freqs_1_to_fp16, y = ts_5_cast_fp16)[name = tensor("op_332_cast_fp16")]; + tensor rotr_1_cast_fp16 = cos(x = var_332_cast_fp16)[name = tensor("rotr_1_cast_fp16")]; + tensor roti_1_cast_fp16 = sin(x = var_332_cast_fp16)[name = tensor("roti_1_cast_fp16")]; + tensor var_336_cast_fp16 = mul(x = var_302_cast_fp16, y = rotr_1_cast_fp16)[name = tensor("op_336_cast_fp16")]; + tensor var_337_cast_fp16 = mul(x = var_310_cast_fp16, y = roti_1_cast_fp16)[name = tensor("op_337_cast_fp16")]; + tensor qor_1_cast_fp16 = sub(x = var_336_cast_fp16, y = var_337_cast_fp16)[name = tensor("qor_1_cast_fp16")]; + tensor var_340_cast_fp16 = mul(x = var_302_cast_fp16, y = roti_1_cast_fp16)[name = tensor("op_340_cast_fp16")]; + tensor var_341_cast_fp16 = mul(x = var_310_cast_fp16, y = rotr_1_cast_fp16)[name = tensor("op_341_cast_fp16")]; + tensor qoi_1_cast_fp16 = add(x = var_340_cast_fp16, y = var_341_cast_fp16)[name = tensor("qoi_1_cast_fp16")]; + tensor var_344_cast_fp16 = mul(x = var_318_cast_fp16, y = rotr_1_cast_fp16)[name = tensor("op_344_cast_fp16")]; + tensor var_345_cast_fp16 = mul(x = var_326_cast_fp16, y = roti_1_cast_fp16)[name = tensor("op_345_cast_fp16")]; + tensor kor_1_cast_fp16 = sub(x = var_344_cast_fp16, y = var_345_cast_fp16)[name = tensor("kor_1_cast_fp16")]; + tensor var_348_cast_fp16 = mul(x = var_318_cast_fp16, y = roti_1_cast_fp16)[name = tensor("op_348_cast_fp16")]; + tensor var_349_cast_fp16 = mul(x = var_326_cast_fp16, y = rotr_1_cast_fp16)[name = tensor("op_349_cast_fp16")]; + tensor koi_1_cast_fp16 = add(x = var_348_cast_fp16, y = var_349_cast_fp16)[name = tensor("koi_1_cast_fp16")]; + tensor qo_1_axis_0 = const()[name = tensor("qo_1_axis_0"), val = tensor(-1)]; + tensor qo_1_cast_fp16 = stack(axis = qo_1_axis_0, values = (qor_1_cast_fp16, qoi_1_cast_fp16))[name = tensor("qo_1_cast_fp16")]; + tensor ko_1_axis_0 = const()[name = tensor("ko_1_axis_0"), val = tensor(-1)]; + tensor ko_1_cast_fp16 = stack(axis = ko_1_axis_0, values = (kor_1_cast_fp16, koi_1_cast_fp16))[name = tensor("ko_1_cast_fp16")]; + tensor var_382 = const()[name = tensor("op_382"), val = tensor([1, 1, 16, 64])]; + tensor q_3_cast_fp16 = reshape(shape = var_382, x = qo_1_cast_fp16)[name = tensor("q_3_cast_fp16")]; + tensor var_388 = const()[name = tensor("op_388"), val = tensor([1, 1, 16, 64])]; + tensor k_3_cast_fp16 = reshape(shape = var_388, x = ko_1_cast_fp16)[name = tensor("k_3_cast_fp16")]; + tensor _inversed_413_y_0_to_fp16 = const()[name = tensor("_inversed_413_y_0_to_fp16"), val = tensor(0x1p-9)]; + tensor _inversed_413_cast_fp16 = mul(x = ts_5_cast_fp16, y = _inversed_413_y_0_to_fp16)[name = tensor("_inversed_413_cast_fp16")]; + tensor var_414_cast_fp16 = floor(x = _inversed_413_cast_fp16)[name = tensor("op_414_cast_fp16")]; + tensor var_415_to_fp16 = const()[name = tensor("op_415_to_fp16"), val = tensor(0x1p+9)]; + tensor var_416_cast_fp16 = mul(x = var_414_cast_fp16, y = var_415_to_fp16)[name = tensor("op_416_cast_fp16")]; + tensor write_indices_float_3_cast_fp16 = sub(x = ts_5_cast_fp16, y = var_416_cast_fp16)[name = tensor("write_indices_float_3_cast_fp16")]; + tensor var_423_dtype_0 = const()[name = tensor("op_423_dtype_0"), val = tensor("int32")]; + tensor write_indices_1_reps_0 = const()[name = tensor("write_indices_1_reps_0"), val = tensor([1, 1, 16, 64])]; + tensor write_indices_float_3_cast_fp16_to_int32 = cast(dtype = var_423_dtype_0, x = write_indices_float_3_cast_fp16)[name = tensor("cast_552")]; + tensor write_indices_1 = tile(reps = write_indices_1_reps_0, x = write_indices_float_3_cast_fp16_to_int32)[name = tensor("write_indices_1")]; + tensor var_433_begin_0 = const()[name = tensor("op_433_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_433_end_0 = const()[name = tensor("op_433_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor var_433_end_mask_0 = const()[name = tensor("op_433_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_433_squeeze_mask_0 = const()[name = tensor("op_433_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor cache0_to_fp16_dtype_0 = const()[name = tensor("cache0_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor cache0_to_fp16 = cast(dtype = cache0_to_fp16_dtype_0, x = cache0)[name = tensor("cast_551")]; + tensor var_433_cast_fp16 = slice_by_index(begin = var_433_begin_0, end = var_433_end_0, end_mask = var_433_end_mask_0, squeeze_mask = var_433_squeeze_mask_0, x = cache0_to_fp16)[name = tensor("op_433_cast_fp16")]; + tensor var_435_axis_0 = const()[name = tensor("op_435_axis_0"), val = tensor(1)]; + tensor var_435_mode_0 = const()[name = tensor("op_435_mode_0"), val = tensor("update")]; + tensor var_435_validate_indices_0 = const()[name = tensor("op_435_validate_indices_0"), val = tensor(false)]; + tensor var_435_cast_fp16 = scatter_along_axis(axis = var_435_axis_0, data = var_433_cast_fp16, indices = write_indices_1, mode = var_435_mode_0, updates = k_3_cast_fp16, validate_indices = var_435_validate_indices_0)[name = tensor("op_435_cast_fp16")]; + tensor concat_2 = const()[name = tensor("concat_2"), val = tensor([0, 0, 0, 0, 0])]; + tensor concat_3 = const()[name = tensor("concat_3"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_1_internal_tensor_assign_1_stride_0 = const()[name = tensor("new_cache_1_internal_tensor_assign_1_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_1_internal_tensor_assign_1_begin_mask_0 = const()[name = tensor("new_cache_1_internal_tensor_assign_1_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_1_internal_tensor_assign_1_end_mask_0 = const()[name = tensor("new_cache_1_internal_tensor_assign_1_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_1_internal_tensor_assign_1_squeeze_mask_0 = const()[name = tensor("new_cache_1_internal_tensor_assign_1_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_48 = const()[name = tensor("shape_48"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_0 = const()[name = tensor("reduce_prod_0"), val = tensor(1048576)]; + tensor range_1d_0_start_0 = const()[name = tensor("range_1d_0_start_0"), val = tensor(0)]; + tensor range_1d_0_step_0 = const()[name = tensor("range_1d_0_step_0"), val = tensor(1)]; + tensor range_1d_0 = range_1d(end = reduce_prod_0, start = range_1d_0_start_0, step = range_1d_0_step_0)[name = tensor("range_1d_0")]; + tensor reshape_0 = reshape(shape = shape_48, x = range_1d_0)[name = tensor("reshape_0")]; + tensor slice_by_index_0 = slice_by_index(begin = concat_2, begin_mask = new_cache_1_internal_tensor_assign_1_begin_mask_0, end = concat_3, end_mask = new_cache_1_internal_tensor_assign_1_end_mask_0, squeeze_mask = new_cache_1_internal_tensor_assign_1_squeeze_mask_0, stride = new_cache_1_internal_tensor_assign_1_stride_0, x = reshape_0)[name = tensor("slice_by_index_0")]; + tensor reshape_1_shape_0 = const()[name = tensor("reshape_1_shape_0"), val = tensor([-1])]; + tensor reshape_1 = reshape(shape = reshape_1_shape_0, x = slice_by_index_0)[name = tensor("reshape_1")]; + tensor reshape_2_shape_0 = const()[name = tensor("reshape_2_shape_0"), val = tensor([-1])]; + tensor reshape_2_cast_fp16 = reshape(shape = reshape_2_shape_0, x = var_435_cast_fp16)[name = tensor("reshape_2_cast_fp16")]; + tensor reshape_3_shape_0 = const()[name = tensor("reshape_3_shape_0"), val = tensor([-1])]; + tensor reshape_3_cast_fp16 = reshape(shape = reshape_3_shape_0, x = cache0_to_fp16)[name = tensor("reshape_3_cast_fp16")]; + tensor scatter_0_mode_0 = const()[name = tensor("scatter_0_mode_0"), val = tensor("update")]; + tensor scatter_0_axis_0 = const()[name = tensor("scatter_0_axis_0"), val = tensor(0)]; + tensor scatter_0_validate_indices_0 = const()[name = tensor("scatter_0_validate_indices_0"), val = tensor(false)]; + tensor scatter_0_cast_fp16 = scatter(axis = scatter_0_axis_0, data = reshape_3_cast_fp16, indices = reshape_1, mode = scatter_0_mode_0, updates = reshape_2_cast_fp16, validate_indices = scatter_0_validate_indices_0)[name = tensor("scatter_0_cast_fp16")]; + tensor reshape_4_cast_fp16 = reshape(shape = shape_48, x = scatter_0_cast_fp16)[name = tensor("reshape_4_cast_fp16")]; + tensor var_443_begin_0 = const()[name = tensor("op_443_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor var_443_end_0 = const()[name = tensor("op_443_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor var_443_end_mask_0 = const()[name = tensor("op_443_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_443_squeeze_mask_0 = const()[name = tensor("op_443_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor var_443_cast_fp16 = slice_by_index(begin = var_443_begin_0, end = var_443_end_0, end_mask = var_443_end_mask_0, squeeze_mask = var_443_squeeze_mask_0, x = reshape_4_cast_fp16)[name = tensor("op_443_cast_fp16")]; + tensor var_445_axis_0 = const()[name = tensor("op_445_axis_0"), val = tensor(1)]; + tensor var_445_mode_0 = const()[name = tensor("op_445_mode_0"), val = tensor("update")]; + tensor var_445_validate_indices_0 = const()[name = tensor("op_445_validate_indices_0"), val = tensor(false)]; + tensor var_445_cast_fp16 = scatter_along_axis(axis = var_445_axis_0, data = var_443_cast_fp16, indices = write_indices_1, mode = var_445_mode_0, updates = v_1_cast_fp16, validate_indices = var_445_validate_indices_0)[name = tensor("op_445_cast_fp16")]; + tensor concat_4 = const()[name = tensor("concat_4"), val = tensor([1, 0, 0, 0, 0])]; + tensor concat_5 = const()[name = tensor("concat_5"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_1_internal_tensor_assign_2_stride_0 = const()[name = tensor("new_cache_1_internal_tensor_assign_2_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_1_internal_tensor_assign_2_begin_mask_0 = const()[name = tensor("new_cache_1_internal_tensor_assign_2_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_1_internal_tensor_assign_2_end_mask_0 = const()[name = tensor("new_cache_1_internal_tensor_assign_2_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_1_internal_tensor_assign_2_squeeze_mask_0 = const()[name = tensor("new_cache_1_internal_tensor_assign_2_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_49 = const()[name = tensor("shape_49"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_1 = const()[name = tensor("reduce_prod_1"), val = tensor(1048576)]; + tensor range_1d_1_start_0 = const()[name = tensor("range_1d_1_start_0"), val = tensor(0)]; + tensor range_1d_1_step_0 = const()[name = tensor("range_1d_1_step_0"), val = tensor(1)]; + tensor range_1d_1 = range_1d(end = reduce_prod_1, start = range_1d_1_start_0, step = range_1d_1_step_0)[name = tensor("range_1d_1")]; + tensor reshape_5 = reshape(shape = shape_49, x = range_1d_1)[name = tensor("reshape_5")]; + tensor slice_by_index_1 = slice_by_index(begin = concat_4, begin_mask = new_cache_1_internal_tensor_assign_2_begin_mask_0, end = concat_5, end_mask = new_cache_1_internal_tensor_assign_2_end_mask_0, squeeze_mask = new_cache_1_internal_tensor_assign_2_squeeze_mask_0, stride = new_cache_1_internal_tensor_assign_2_stride_0, x = reshape_5)[name = tensor("slice_by_index_1")]; + tensor reshape_6_shape_0 = const()[name = tensor("reshape_6_shape_0"), val = tensor([-1])]; + tensor reshape_6 = reshape(shape = reshape_6_shape_0, x = slice_by_index_1)[name = tensor("reshape_6")]; + tensor reshape_7_shape_0 = const()[name = tensor("reshape_7_shape_0"), val = tensor([-1])]; + tensor reshape_7_cast_fp16 = reshape(shape = reshape_7_shape_0, x = var_445_cast_fp16)[name = tensor("reshape_7_cast_fp16")]; + tensor reshape_8_shape_0 = const()[name = tensor("reshape_8_shape_0"), val = tensor([-1])]; + tensor reshape_8_cast_fp16 = reshape(shape = reshape_8_shape_0, x = reshape_4_cast_fp16)[name = tensor("reshape_8_cast_fp16")]; + tensor scatter_1_mode_0 = const()[name = tensor("scatter_1_mode_0"), val = tensor("update")]; + tensor scatter_1_axis_0 = const()[name = tensor("scatter_1_axis_0"), val = tensor(0)]; + tensor scatter_1_validate_indices_0 = const()[name = tensor("scatter_1_validate_indices_0"), val = tensor(false)]; + tensor scatter_1_cast_fp16 = scatter(axis = scatter_1_axis_0, data = reshape_8_cast_fp16, indices = reshape_6, mode = scatter_1_mode_0, updates = reshape_7_cast_fp16, validate_indices = scatter_1_validate_indices_0)[name = tensor("scatter_1_cast_fp16")]; + tensor reshape_9_cast_fp16 = reshape(shape = shape_49, x = scatter_1_cast_fp16)[name = tensor("reshape_9_cast_fp16")]; + tensor reshape_9_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("reshape_9_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor keys_1_begin_0 = const()[name = tensor("keys_1_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor keys_1_end_0 = const()[name = tensor("keys_1_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor keys_1_end_mask_0 = const()[name = tensor("keys_1_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor keys_1_squeeze_mask_0 = const()[name = tensor("keys_1_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor keys_1_cast_fp16 = slice_by_index(begin = keys_1_begin_0, end = keys_1_end_0, end_mask = keys_1_end_mask_0, squeeze_mask = keys_1_squeeze_mask_0, x = reshape_9_cast_fp16)[name = tensor("keys_1_cast_fp16")]; + tensor values_1_begin_0 = const()[name = tensor("values_1_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor values_1_end_0 = const()[name = tensor("values_1_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor values_1_end_mask_0 = const()[name = tensor("values_1_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor values_1_squeeze_mask_0 = const()[name = tensor("values_1_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor values_1_cast_fp16 = slice_by_index(begin = values_1_begin_0, end = values_1_end_0, end_mask = values_1_end_mask_0, squeeze_mask = values_1_squeeze_mask_0, x = reshape_9_cast_fp16)[name = tensor("values_1_cast_fp16")]; + tensor var_457_cast_fp16 = not_equal(x = keys_1_cast_fp16, y = keys_1_cast_fp16)[name = tensor("op_457_cast_fp16")]; + tensor var_463_to_fp16 = const()[name = tensor("op_463_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(6369856)))]; + tensor keys_3_cast_fp16 = select(a = var_463_to_fp16, b = keys_1_cast_fp16, cond = var_457_cast_fp16)[name = tensor("keys_3_cast_fp16")]; + tensor var_465_cast_fp16 = not_equal(x = values_1_cast_fp16, y = values_1_cast_fp16)[name = tensor("op_465_cast_fp16")]; + tensor values_3_cast_fp16 = select(a = var_463_to_fp16, b = values_1_cast_fp16, cond = var_465_cast_fp16)[name = tensor("values_3_cast_fp16")]; + tensor var_489 = const()[name = tensor("op_489"), val = tensor([0, 2, 1, 3])]; + tensor var_505 = const()[name = tensor("op_505"), val = tensor([1, 1, 1])]; + tensor var_506_cast_fp16 = reshape(shape = var_505, x = position0_to_fp16)[name = tensor("op_506_cast_fp16")]; + tensor var_526_to_fp16 = const()[name = tensor("op_526_to_fp16"), val = tensor(0x1p+0)]; + tensor valid_len_1_cast_fp16 = add(x = var_506_cast_fp16, y = var_526_to_fp16)[name = tensor("valid_len_1_cast_fp16")]; + tensor k_positions_1_promoted_to_fp16 = const()[name = tensor("k_positions_1_promoted_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(7418496)))]; + tensor valid_mask_1_cast_fp16 = less(x = k_positions_1_promoted_to_fp16, y = valid_len_1_cast_fp16)[name = tensor("valid_mask_1_cast_fp16")]; + tensor causal_mask_1_cast_fp16 = less_equal(x = k_positions_1_promoted_to_fp16, y = var_506_cast_fp16)[name = tensor("causal_mask_1_cast_fp16")]; + tensor attn_mask_1 = logical_and(x = valid_mask_1_cast_fp16, y = causal_mask_1_cast_fp16)[name = tensor("attn_mask_1")]; + tensor attn_mask_3_axes_0 = const()[name = tensor("attn_mask_3_axes_0"), val = tensor([1])]; + tensor attn_mask_3 = expand_dims(axes = attn_mask_3_axes_0, x = attn_mask_1)[name = tensor("attn_mask_3")]; + tensor var_544_transpose_x_0 = const()[name = tensor("op_544_transpose_x_0"), val = tensor(false)]; + tensor var_544_transpose_y_0 = const()[name = tensor("op_544_transpose_y_0"), val = tensor(false)]; + tensor transpose_72_perm_0 = const()[name = tensor("transpose_72_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_73_perm_0 = const()[name = tensor("transpose_73_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_73 = transpose(perm = transpose_73_perm_0, x = keys_3_cast_fp16)[name = tensor("transpose_213")]; + tensor transpose_72 = transpose(perm = transpose_72_perm_0, x = q_3_cast_fp16)[name = tensor("transpose_214")]; + tensor var_544_cast_fp16 = matmul(transpose_x = var_544_transpose_x_0, transpose_y = var_544_transpose_y_0, x = transpose_72, y = transpose_73)[name = tensor("op_544_cast_fp16")]; + tensor var_538_to_fp16 = const()[name = tensor("op_538_to_fp16"), val = tensor([0x1p-3])]; + tensor attn_weights_1_cast_fp16 = mul(x = var_544_cast_fp16, y = var_538_to_fp16)[name = tensor("attn_weights_1_cast_fp16")]; + tensor var_546 = logical_not(x = attn_mask_3)[name = tensor("op_546")]; + tensor var_547_to_fp16 = const()[name = tensor("op_547_to_fp16"), val = tensor(-inf)]; + tensor attn_weights_3_cast_fp16 = select(a = var_547_to_fp16, b = attn_weights_1_cast_fp16, cond = var_546)[name = tensor("attn_weights_3_cast_fp16")]; + tensor var_549 = const()[name = tensor("op_549"), val = tensor(-1)]; + tensor attn_weights_5_cast_fp16 = softmax(axis = var_549, x = attn_weights_3_cast_fp16)[name = tensor("attn_weights_5_cast_fp16")]; + tensor attn_output_1_transpose_x_0 = const()[name = tensor("attn_output_1_transpose_x_0"), val = tensor(false)]; + tensor attn_output_1_transpose_y_0 = const()[name = tensor("attn_output_1_transpose_y_0"), val = tensor(false)]; + tensor values_5_cast_fp16 = transpose(perm = var_489, x = values_3_cast_fp16)[name = tensor("transpose_215")]; + tensor attn_output_1_cast_fp16 = matmul(transpose_x = attn_output_1_transpose_x_0, transpose_y = attn_output_1_transpose_y_0, x = attn_weights_5_cast_fp16, y = values_5_cast_fp16)[name = tensor("attn_output_1_cast_fp16")]; + tensor var_557 = const()[name = tensor("op_557"), val = tensor([0, 2, 1, 3])]; + tensor var_562 = const()[name = tensor("op_562"), val = tensor([1, 1, 1024])]; + tensor var_558_cast_fp16 = transpose(perm = var_557, x = attn_output_1_cast_fp16)[name = tensor("transpose_212")]; + tensor input_7_cast_fp16 = reshape(shape = var_562, x = var_558_cast_fp16)[name = tensor("input_7_cast_fp16")]; + tensor attn0_out_proj_weight_to_fp16 = const()[name = tensor("attn0_out_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(7419584)))]; + tensor linear_2_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = attn0_out_proj_weight_to_fp16, x = input_7_cast_fp16)[name = tensor("linear_2_cast_fp16")]; + tensor var_568_to_fp16 = const()[name = tensor("op_568_to_fp16"), val = tensor(0x1p+0)]; + tensor var_569_cast_fp16 = add(x = position0_to_fp16, y = var_568_to_fp16)[name = tensor("op_569_cast_fp16")]; + tensor var_569_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_569_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor input_9_cast_fp16 = add(x = linear_0_cast_fp16, y = linear_2_cast_fp16)[name = tensor("input_9_cast_fp16")]; + tensor input_11_axes_0 = const()[name = tensor("input_11_axes_0"), val = tensor([-1])]; + tensor norm0_2_weight_to_fp16 = const()[name = tensor("norm0_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9516800)))]; + tensor norm0_2_bias_to_fp16 = const()[name = tensor("norm0_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9518912)))]; + tensor var_573_to_fp16 = const()[name = tensor("op_573_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_11_cast_fp16 = layer_norm(axes = input_11_axes_0, beta = norm0_2_bias_to_fp16, epsilon = var_573_to_fp16, gamma = norm0_2_weight_to_fp16, x = input_9_cast_fp16)[name = tensor("input_11_cast_fp16")]; + tensor linear0_1_weight_to_fp16 = const()[name = tensor("linear0_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(9521024)))]; + tensor linear_3_bias_0_to_fp16 = const()[name = tensor("linear_3_bias_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(17909696)))]; + tensor linear_3_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = linear0_1_weight_to_fp16, x = input_11_cast_fp16)[name = tensor("linear_3_cast_fp16")]; + tensor input_13_mode_0 = const()[name = tensor("input_13_mode_0"), val = tensor("EXACT")]; + tensor input_13_cast_fp16 = gelu(mode = input_13_mode_0, x = linear_3_cast_fp16)[name = tensor("input_13_cast_fp16")]; + tensor linear0_2_weight_to_fp16 = const()[name = tensor("linear0_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(17917952)))]; + tensor linear_4_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = linear0_2_weight_to_fp16, x = input_13_cast_fp16)[name = tensor("linear_4_cast_fp16")]; + tensor input_15_cast_fp16 = add(x = input_9_cast_fp16, y = linear_4_cast_fp16)[name = tensor("input_15_cast_fp16")]; + tensor input_17_axes_0 = const()[name = tensor("input_17_axes_0"), val = tensor([-1])]; + tensor norm1_1_weight_to_fp16 = const()[name = tensor("norm1_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(26306624)))]; + tensor norm1_1_bias_to_fp16 = const()[name = tensor("norm1_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(26308736)))]; + tensor var_590_to_fp16 = const()[name = tensor("op_590_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_17_cast_fp16 = layer_norm(axes = input_17_axes_0, beta = norm1_1_bias_to_fp16, epsilon = var_590_to_fp16, gamma = norm1_1_weight_to_fp16, x = input_15_cast_fp16)[name = tensor("input_17_cast_fp16")]; + tensor attn1_in_proj_weight_to_fp16 = const()[name = tensor("attn1_in_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(26310848)))]; + tensor linear_5_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = attn1_in_proj_weight_to_fp16, x = input_17_cast_fp16)[name = tensor("linear_5_cast_fp16")]; + tensor var_604 = const()[name = tensor("op_604"), val = tensor([1, 1, 3, 16, 64])]; + tensor qkv_3_cast_fp16 = reshape(shape = var_604, x = linear_5_cast_fp16)[name = tensor("qkv_3_cast_fp16")]; + tensor q_7_begin_0 = const()[name = tensor("q_7_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor q_7_end_0 = const()[name = tensor("q_7_end_0"), val = tensor([1, 1, 1, 16, 64])]; + tensor q_7_end_mask_0 = const()[name = tensor("q_7_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor q_7_squeeze_mask_0 = const()[name = tensor("q_7_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor q_7_cast_fp16 = slice_by_index(begin = q_7_begin_0, end = q_7_end_0, end_mask = q_7_end_mask_0, squeeze_mask = q_7_squeeze_mask_0, x = qkv_3_cast_fp16)[name = tensor("q_7_cast_fp16")]; + tensor k_5_begin_0 = const()[name = tensor("k_5_begin_0"), val = tensor([0, 0, 1, 0, 0])]; + tensor k_5_end_0 = const()[name = tensor("k_5_end_0"), val = tensor([1, 1, 2, 16, 64])]; + tensor k_5_end_mask_0 = const()[name = tensor("k_5_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor k_5_squeeze_mask_0 = const()[name = tensor("k_5_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor k_5_cast_fp16 = slice_by_index(begin = k_5_begin_0, end = k_5_end_0, end_mask = k_5_end_mask_0, squeeze_mask = k_5_squeeze_mask_0, x = qkv_3_cast_fp16)[name = tensor("k_5_cast_fp16")]; + tensor v_3_begin_0 = const()[name = tensor("v_3_begin_0"), val = tensor([0, 0, 2, 0, 0])]; + tensor v_3_end_0 = const()[name = tensor("v_3_end_0"), val = tensor([1, 1, 3, 16, 64])]; + tensor v_3_end_mask_0 = const()[name = tensor("v_3_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor v_3_squeeze_mask_0 = const()[name = tensor("v_3_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor v_3_cast_fp16 = slice_by_index(begin = v_3_begin_0, end = v_3_end_0, end_mask = v_3_end_mask_0, squeeze_mask = v_3_squeeze_mask_0, x = qkv_3_cast_fp16)[name = tensor("v_3_cast_fp16")]; + tensor var_670 = const()[name = tensor("op_670"), val = tensor([1, 1, 1, 1])]; + tensor position1_to_fp16_dtype_0 = const()[name = tensor("position1_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor position1_to_fp16 = cast(dtype = position1_to_fp16_dtype_0, x = position1)[name = tensor("cast_548")]; + tensor ts_11_cast_fp16 = reshape(shape = var_670, x = position1_to_fp16)[name = tensor("ts_11_cast_fp16")]; + tensor var_677 = const()[name = tensor("op_677"), val = tensor([1, 1, 16, 32, 2])]; + tensor q_complex_3_cast_fp16 = reshape(shape = var_677, x = q_7_cast_fp16)[name = tensor("q_complex_3_cast_fp16")]; + tensor var_684 = const()[name = tensor("op_684"), val = tensor([1, 1, 16, 32, 2])]; + tensor k_complex_3_cast_fp16 = reshape(shape = var_684, x = k_5_cast_fp16)[name = tensor("k_complex_3_cast_fp16")]; + tensor var_688_begin_0 = const()[name = tensor("op_688_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_688_end_0 = const()[name = tensor("op_688_end_0"), val = tensor([1, 1, 16, 32, 1])]; + tensor var_688_end_mask_0 = const()[name = tensor("op_688_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_688_squeeze_mask_0 = const()[name = tensor("op_688_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_688_cast_fp16 = slice_by_index(begin = var_688_begin_0, end = var_688_end_0, end_mask = var_688_end_mask_0, squeeze_mask = var_688_squeeze_mask_0, x = q_complex_3_cast_fp16)[name = tensor("op_688_cast_fp16")]; + tensor var_696_begin_0 = const()[name = tensor("op_696_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_696_end_0 = const()[name = tensor("op_696_end_0"), val = tensor([1, 1, 16, 32, 2])]; + tensor var_696_end_mask_0 = const()[name = tensor("op_696_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_696_squeeze_mask_0 = const()[name = tensor("op_696_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_696_cast_fp16 = slice_by_index(begin = var_696_begin_0, end = var_696_end_0, end_mask = var_696_end_mask_0, squeeze_mask = var_696_squeeze_mask_0, x = q_complex_3_cast_fp16)[name = tensor("op_696_cast_fp16")]; + tensor var_704_begin_0 = const()[name = tensor("op_704_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_704_end_0 = const()[name = tensor("op_704_end_0"), val = tensor([1, 1, 16, 32, 1])]; + tensor var_704_end_mask_0 = const()[name = tensor("op_704_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_704_squeeze_mask_0 = const()[name = tensor("op_704_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_704_cast_fp16 = slice_by_index(begin = var_704_begin_0, end = var_704_end_0, end_mask = var_704_end_mask_0, squeeze_mask = var_704_squeeze_mask_0, x = k_complex_3_cast_fp16)[name = tensor("op_704_cast_fp16")]; + tensor var_712_begin_0 = const()[name = tensor("op_712_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_712_end_0 = const()[name = tensor("op_712_end_0"), val = tensor([1, 1, 16, 32, 2])]; + tensor var_712_end_mask_0 = const()[name = tensor("op_712_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_712_squeeze_mask_0 = const()[name = tensor("op_712_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_712_cast_fp16 = slice_by_index(begin = var_712_begin_0, end = var_712_end_0, end_mask = var_712_end_mask_0, squeeze_mask = var_712_squeeze_mask_0, x = k_complex_3_cast_fp16)[name = tensor("op_712_cast_fp16")]; + tensor freqs_3_to_fp16 = const()[name = tensor("freqs_3_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(32602368)))]; + tensor var_718_cast_fp16 = mul(x = freqs_3_to_fp16, y = ts_11_cast_fp16)[name = tensor("op_718_cast_fp16")]; + tensor rotr_3_cast_fp16 = cos(x = var_718_cast_fp16)[name = tensor("rotr_3_cast_fp16")]; + tensor roti_3_cast_fp16 = sin(x = var_718_cast_fp16)[name = tensor("roti_3_cast_fp16")]; + tensor var_722_cast_fp16 = mul(x = var_688_cast_fp16, y = rotr_3_cast_fp16)[name = tensor("op_722_cast_fp16")]; + tensor var_723_cast_fp16 = mul(x = var_696_cast_fp16, y = roti_3_cast_fp16)[name = tensor("op_723_cast_fp16")]; + tensor qor_5_cast_fp16 = sub(x = var_722_cast_fp16, y = var_723_cast_fp16)[name = tensor("qor_5_cast_fp16")]; + tensor var_726_cast_fp16 = mul(x = var_688_cast_fp16, y = roti_3_cast_fp16)[name = tensor("op_726_cast_fp16")]; + tensor var_727_cast_fp16 = mul(x = var_696_cast_fp16, y = rotr_3_cast_fp16)[name = tensor("op_727_cast_fp16")]; + tensor qoi_5_cast_fp16 = add(x = var_726_cast_fp16, y = var_727_cast_fp16)[name = tensor("qoi_5_cast_fp16")]; + tensor var_730_cast_fp16 = mul(x = var_704_cast_fp16, y = rotr_3_cast_fp16)[name = tensor("op_730_cast_fp16")]; + tensor var_731_cast_fp16 = mul(x = var_712_cast_fp16, y = roti_3_cast_fp16)[name = tensor("op_731_cast_fp16")]; + tensor kor_5_cast_fp16 = sub(x = var_730_cast_fp16, y = var_731_cast_fp16)[name = tensor("kor_5_cast_fp16")]; + tensor var_734_cast_fp16 = mul(x = var_704_cast_fp16, y = roti_3_cast_fp16)[name = tensor("op_734_cast_fp16")]; + tensor var_735_cast_fp16 = mul(x = var_712_cast_fp16, y = rotr_3_cast_fp16)[name = tensor("op_735_cast_fp16")]; + tensor koi_5_cast_fp16 = add(x = var_734_cast_fp16, y = var_735_cast_fp16)[name = tensor("koi_5_cast_fp16")]; + tensor qo_3_axis_0 = const()[name = tensor("qo_3_axis_0"), val = tensor(-1)]; + tensor qo_3_cast_fp16 = stack(axis = qo_3_axis_0, values = (qor_5_cast_fp16, qoi_5_cast_fp16))[name = tensor("qo_3_cast_fp16")]; + tensor ko_3_axis_0 = const()[name = tensor("ko_3_axis_0"), val = tensor(-1)]; + tensor ko_3_cast_fp16 = stack(axis = ko_3_axis_0, values = (kor_5_cast_fp16, koi_5_cast_fp16))[name = tensor("ko_3_cast_fp16")]; + tensor var_768 = const()[name = tensor("op_768"), val = tensor([1, 1, 16, 64])]; + tensor q_9_cast_fp16 = reshape(shape = var_768, x = qo_3_cast_fp16)[name = tensor("q_9_cast_fp16")]; + tensor var_774 = const()[name = tensor("op_774"), val = tensor([1, 1, 16, 64])]; + tensor k_7_cast_fp16 = reshape(shape = var_774, x = ko_3_cast_fp16)[name = tensor("k_7_cast_fp16")]; + tensor _inversed_799_y_0_to_fp16 = const()[name = tensor("_inversed_799_y_0_to_fp16"), val = tensor(0x1p-9)]; + tensor _inversed_799_cast_fp16 = mul(x = ts_11_cast_fp16, y = _inversed_799_y_0_to_fp16)[name = tensor("_inversed_799_cast_fp16")]; + tensor var_800_cast_fp16 = floor(x = _inversed_799_cast_fp16)[name = tensor("op_800_cast_fp16")]; + tensor var_801_to_fp16 = const()[name = tensor("op_801_to_fp16"), val = tensor(0x1p+9)]; + tensor var_802_cast_fp16 = mul(x = var_800_cast_fp16, y = var_801_to_fp16)[name = tensor("op_802_cast_fp16")]; + tensor write_indices_float_7_cast_fp16 = sub(x = ts_11_cast_fp16, y = var_802_cast_fp16)[name = tensor("write_indices_float_7_cast_fp16")]; + tensor var_809_dtype_0 = const()[name = tensor("op_809_dtype_0"), val = tensor("int32")]; + tensor write_indices_3_reps_0 = const()[name = tensor("write_indices_3_reps_0"), val = tensor([1, 1, 16, 64])]; + tensor write_indices_float_7_cast_fp16_to_int32 = cast(dtype = var_809_dtype_0, x = write_indices_float_7_cast_fp16)[name = tensor("cast_547")]; + tensor write_indices_3 = tile(reps = write_indices_3_reps_0, x = write_indices_float_7_cast_fp16_to_int32)[name = tensor("write_indices_3")]; + tensor var_819_begin_0 = const()[name = tensor("op_819_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_819_end_0 = const()[name = tensor("op_819_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor var_819_end_mask_0 = const()[name = tensor("op_819_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_819_squeeze_mask_0 = const()[name = tensor("op_819_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor cache1_to_fp16_dtype_0 = const()[name = tensor("cache1_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor cache1_to_fp16 = cast(dtype = cache1_to_fp16_dtype_0, x = cache1)[name = tensor("cast_546")]; + tensor var_819_cast_fp16 = slice_by_index(begin = var_819_begin_0, end = var_819_end_0, end_mask = var_819_end_mask_0, squeeze_mask = var_819_squeeze_mask_0, x = cache1_to_fp16)[name = tensor("op_819_cast_fp16")]; + tensor var_821_axis_0 = const()[name = tensor("op_821_axis_0"), val = tensor(1)]; + tensor var_821_mode_0 = const()[name = tensor("op_821_mode_0"), val = tensor("update")]; + tensor var_821_validate_indices_0 = const()[name = tensor("op_821_validate_indices_0"), val = tensor(false)]; + tensor var_821_cast_fp16 = scatter_along_axis(axis = var_821_axis_0, data = var_819_cast_fp16, indices = write_indices_3, mode = var_821_mode_0, updates = k_7_cast_fp16, validate_indices = var_821_validate_indices_0)[name = tensor("op_821_cast_fp16")]; + tensor concat_9 = const()[name = tensor("concat_9"), val = tensor([0, 0, 0, 0, 0])]; + tensor concat_10 = const()[name = tensor("concat_10"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_3_internal_tensor_assign_1_stride_0 = const()[name = tensor("new_cache_3_internal_tensor_assign_1_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_3_internal_tensor_assign_1_begin_mask_0 = const()[name = tensor("new_cache_3_internal_tensor_assign_1_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_3_internal_tensor_assign_1_end_mask_0 = const()[name = tensor("new_cache_3_internal_tensor_assign_1_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_3_internal_tensor_assign_1_squeeze_mask_0 = const()[name = tensor("new_cache_3_internal_tensor_assign_1_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_50 = const()[name = tensor("shape_50"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_2 = const()[name = tensor("reduce_prod_2"), val = tensor(1048576)]; + tensor range_1d_2_start_0 = const()[name = tensor("range_1d_2_start_0"), val = tensor(0)]; + tensor range_1d_2_step_0 = const()[name = tensor("range_1d_2_step_0"), val = tensor(1)]; + tensor range_1d_2 = range_1d(end = reduce_prod_2, start = range_1d_2_start_0, step = range_1d_2_step_0)[name = tensor("range_1d_2")]; + tensor reshape_10 = reshape(shape = shape_50, x = range_1d_2)[name = tensor("reshape_10")]; + tensor slice_by_index_2 = slice_by_index(begin = concat_9, begin_mask = new_cache_3_internal_tensor_assign_1_begin_mask_0, end = concat_10, end_mask = new_cache_3_internal_tensor_assign_1_end_mask_0, squeeze_mask = new_cache_3_internal_tensor_assign_1_squeeze_mask_0, stride = new_cache_3_internal_tensor_assign_1_stride_0, x = reshape_10)[name = tensor("slice_by_index_2")]; + tensor reshape_11_shape_0 = const()[name = tensor("reshape_11_shape_0"), val = tensor([-1])]; + tensor reshape_11 = reshape(shape = reshape_11_shape_0, x = slice_by_index_2)[name = tensor("reshape_11")]; + tensor reshape_12_shape_0 = const()[name = tensor("reshape_12_shape_0"), val = tensor([-1])]; + tensor reshape_12_cast_fp16 = reshape(shape = reshape_12_shape_0, x = var_821_cast_fp16)[name = tensor("reshape_12_cast_fp16")]; + tensor reshape_13_shape_0 = const()[name = tensor("reshape_13_shape_0"), val = tensor([-1])]; + tensor reshape_13_cast_fp16 = reshape(shape = reshape_13_shape_0, x = cache1_to_fp16)[name = tensor("reshape_13_cast_fp16")]; + tensor scatter_2_mode_0 = const()[name = tensor("scatter_2_mode_0"), val = tensor("update")]; + tensor scatter_2_axis_0 = const()[name = tensor("scatter_2_axis_0"), val = tensor(0)]; + tensor scatter_2_validate_indices_0 = const()[name = tensor("scatter_2_validate_indices_0"), val = tensor(false)]; + tensor scatter_2_cast_fp16 = scatter(axis = scatter_2_axis_0, data = reshape_13_cast_fp16, indices = reshape_11, mode = scatter_2_mode_0, updates = reshape_12_cast_fp16, validate_indices = scatter_2_validate_indices_0)[name = tensor("scatter_2_cast_fp16")]; + tensor reshape_14_cast_fp16 = reshape(shape = shape_50, x = scatter_2_cast_fp16)[name = tensor("reshape_14_cast_fp16")]; + tensor var_829_begin_0 = const()[name = tensor("op_829_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor var_829_end_0 = const()[name = tensor("op_829_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor var_829_end_mask_0 = const()[name = tensor("op_829_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_829_squeeze_mask_0 = const()[name = tensor("op_829_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor var_829_cast_fp16 = slice_by_index(begin = var_829_begin_0, end = var_829_end_0, end_mask = var_829_end_mask_0, squeeze_mask = var_829_squeeze_mask_0, x = reshape_14_cast_fp16)[name = tensor("op_829_cast_fp16")]; + tensor var_831_axis_0 = const()[name = tensor("op_831_axis_0"), val = tensor(1)]; + tensor var_831_mode_0 = const()[name = tensor("op_831_mode_0"), val = tensor("update")]; + tensor var_831_validate_indices_0 = const()[name = tensor("op_831_validate_indices_0"), val = tensor(false)]; + tensor var_831_cast_fp16 = scatter_along_axis(axis = var_831_axis_0, data = var_829_cast_fp16, indices = write_indices_3, mode = var_831_mode_0, updates = v_3_cast_fp16, validate_indices = var_831_validate_indices_0)[name = tensor("op_831_cast_fp16")]; + tensor concat_11 = const()[name = tensor("concat_11"), val = tensor([1, 0, 0, 0, 0])]; + tensor concat_12 = const()[name = tensor("concat_12"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_3_internal_tensor_assign_2_stride_0 = const()[name = tensor("new_cache_3_internal_tensor_assign_2_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_3_internal_tensor_assign_2_begin_mask_0 = const()[name = tensor("new_cache_3_internal_tensor_assign_2_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_3_internal_tensor_assign_2_end_mask_0 = const()[name = tensor("new_cache_3_internal_tensor_assign_2_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_3_internal_tensor_assign_2_squeeze_mask_0 = const()[name = tensor("new_cache_3_internal_tensor_assign_2_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_51 = const()[name = tensor("shape_51"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_3 = const()[name = tensor("reduce_prod_3"), val = tensor(1048576)]; + tensor range_1d_3_start_0 = const()[name = tensor("range_1d_3_start_0"), val = tensor(0)]; + tensor range_1d_3_step_0 = const()[name = tensor("range_1d_3_step_0"), val = tensor(1)]; + tensor range_1d_3 = range_1d(end = reduce_prod_3, start = range_1d_3_start_0, step = range_1d_3_step_0)[name = tensor("range_1d_3")]; + tensor reshape_15 = reshape(shape = shape_51, x = range_1d_3)[name = tensor("reshape_15")]; + tensor slice_by_index_3 = slice_by_index(begin = concat_11, begin_mask = new_cache_3_internal_tensor_assign_2_begin_mask_0, end = concat_12, end_mask = new_cache_3_internal_tensor_assign_2_end_mask_0, squeeze_mask = new_cache_3_internal_tensor_assign_2_squeeze_mask_0, stride = new_cache_3_internal_tensor_assign_2_stride_0, x = reshape_15)[name = tensor("slice_by_index_3")]; + tensor reshape_16_shape_0 = const()[name = tensor("reshape_16_shape_0"), val = tensor([-1])]; + tensor reshape_16 = reshape(shape = reshape_16_shape_0, x = slice_by_index_3)[name = tensor("reshape_16")]; + tensor reshape_17_shape_0 = const()[name = tensor("reshape_17_shape_0"), val = tensor([-1])]; + tensor reshape_17_cast_fp16 = reshape(shape = reshape_17_shape_0, x = var_831_cast_fp16)[name = tensor("reshape_17_cast_fp16")]; + tensor reshape_18_shape_0 = const()[name = tensor("reshape_18_shape_0"), val = tensor([-1])]; + tensor reshape_18_cast_fp16 = reshape(shape = reshape_18_shape_0, x = reshape_14_cast_fp16)[name = tensor("reshape_18_cast_fp16")]; + tensor scatter_3_mode_0 = const()[name = tensor("scatter_3_mode_0"), val = tensor("update")]; + tensor scatter_3_axis_0 = const()[name = tensor("scatter_3_axis_0"), val = tensor(0)]; + tensor scatter_3_validate_indices_0 = const()[name = tensor("scatter_3_validate_indices_0"), val = tensor(false)]; + tensor scatter_3_cast_fp16 = scatter(axis = scatter_3_axis_0, data = reshape_18_cast_fp16, indices = reshape_16, mode = scatter_3_mode_0, updates = reshape_17_cast_fp16, validate_indices = scatter_3_validate_indices_0)[name = tensor("scatter_3_cast_fp16")]; + tensor reshape_19_cast_fp16 = reshape(shape = shape_51, x = scatter_3_cast_fp16)[name = tensor("reshape_19_cast_fp16")]; + tensor reshape_19_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("reshape_19_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor keys_7_begin_0 = const()[name = tensor("keys_7_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor keys_7_end_0 = const()[name = tensor("keys_7_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor keys_7_end_mask_0 = const()[name = tensor("keys_7_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor keys_7_squeeze_mask_0 = const()[name = tensor("keys_7_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor keys_7_cast_fp16 = slice_by_index(begin = keys_7_begin_0, end = keys_7_end_0, end_mask = keys_7_end_mask_0, squeeze_mask = keys_7_squeeze_mask_0, x = reshape_19_cast_fp16)[name = tensor("keys_7_cast_fp16")]; + tensor values_7_begin_0 = const()[name = tensor("values_7_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor values_7_end_0 = const()[name = tensor("values_7_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor values_7_end_mask_0 = const()[name = tensor("values_7_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor values_7_squeeze_mask_0 = const()[name = tensor("values_7_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor values_7_cast_fp16 = slice_by_index(begin = values_7_begin_0, end = values_7_end_0, end_mask = values_7_end_mask_0, squeeze_mask = values_7_squeeze_mask_0, x = reshape_19_cast_fp16)[name = tensor("values_7_cast_fp16")]; + tensor var_843_cast_fp16 = not_equal(x = keys_7_cast_fp16, y = keys_7_cast_fp16)[name = tensor("op_843_cast_fp16")]; + tensor keys_9_cast_fp16 = select(a = var_463_to_fp16, b = keys_7_cast_fp16, cond = var_843_cast_fp16)[name = tensor("keys_9_cast_fp16")]; + tensor var_851_cast_fp16 = not_equal(x = values_7_cast_fp16, y = values_7_cast_fp16)[name = tensor("op_851_cast_fp16")]; + tensor values_9_cast_fp16 = select(a = var_463_to_fp16, b = values_7_cast_fp16, cond = var_851_cast_fp16)[name = tensor("values_9_cast_fp16")]; + tensor var_875 = const()[name = tensor("op_875"), val = tensor([0, 2, 1, 3])]; + tensor var_891 = const()[name = tensor("op_891"), val = tensor([1, 1, 1])]; + tensor var_892_cast_fp16 = reshape(shape = var_891, x = position1_to_fp16)[name = tensor("op_892_cast_fp16")]; + tensor var_912_to_fp16 = const()[name = tensor("op_912_to_fp16"), val = tensor(0x1p+0)]; + tensor valid_len_3_cast_fp16 = add(x = var_892_cast_fp16, y = var_912_to_fp16)[name = tensor("valid_len_3_cast_fp16")]; + tensor valid_mask_3_cast_fp16 = less(x = k_positions_1_promoted_to_fp16, y = valid_len_3_cast_fp16)[name = tensor("valid_mask_3_cast_fp16")]; + tensor causal_mask_3_cast_fp16 = less_equal(x = k_positions_1_promoted_to_fp16, y = var_892_cast_fp16)[name = tensor("causal_mask_3_cast_fp16")]; + tensor attn_mask_5 = logical_and(x = valid_mask_3_cast_fp16, y = causal_mask_3_cast_fp16)[name = tensor("attn_mask_5")]; + tensor attn_mask_7_axes_0 = const()[name = tensor("attn_mask_7_axes_0"), val = tensor([1])]; + tensor attn_mask_7 = expand_dims(axes = attn_mask_7_axes_0, x = attn_mask_5)[name = tensor("attn_mask_7")]; + tensor var_930_transpose_x_0 = const()[name = tensor("op_930_transpose_x_0"), val = tensor(false)]; + tensor var_930_transpose_y_0 = const()[name = tensor("op_930_transpose_y_0"), val = tensor(false)]; + tensor transpose_74_perm_0 = const()[name = tensor("transpose_74_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_75_perm_0 = const()[name = tensor("transpose_75_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_75 = transpose(perm = transpose_75_perm_0, x = keys_9_cast_fp16)[name = tensor("transpose_209")]; + tensor transpose_74 = transpose(perm = transpose_74_perm_0, x = q_9_cast_fp16)[name = tensor("transpose_210")]; + tensor var_930_cast_fp16 = matmul(transpose_x = var_930_transpose_x_0, transpose_y = var_930_transpose_y_0, x = transpose_74, y = transpose_75)[name = tensor("op_930_cast_fp16")]; + tensor var_924_to_fp16 = const()[name = tensor("op_924_to_fp16"), val = tensor([0x1p-3])]; + tensor attn_weights_7_cast_fp16 = mul(x = var_930_cast_fp16, y = var_924_to_fp16)[name = tensor("attn_weights_7_cast_fp16")]; + tensor var_932 = logical_not(x = attn_mask_7)[name = tensor("op_932")]; + tensor var_933_to_fp16 = const()[name = tensor("op_933_to_fp16"), val = tensor(-inf)]; + tensor attn_weights_9_cast_fp16 = select(a = var_933_to_fp16, b = attn_weights_7_cast_fp16, cond = var_932)[name = tensor("attn_weights_9_cast_fp16")]; + tensor var_935 = const()[name = tensor("op_935"), val = tensor(-1)]; + tensor attn_weights_11_cast_fp16 = softmax(axis = var_935, x = attn_weights_9_cast_fp16)[name = tensor("attn_weights_11_cast_fp16")]; + tensor attn_output_3_transpose_x_0 = const()[name = tensor("attn_output_3_transpose_x_0"), val = tensor(false)]; + tensor attn_output_3_transpose_y_0 = const()[name = tensor("attn_output_3_transpose_y_0"), val = tensor(false)]; + tensor values_11_cast_fp16 = transpose(perm = var_875, x = values_9_cast_fp16)[name = tensor("transpose_211")]; + tensor attn_output_3_cast_fp16 = matmul(transpose_x = attn_output_3_transpose_x_0, transpose_y = attn_output_3_transpose_y_0, x = attn_weights_11_cast_fp16, y = values_11_cast_fp16)[name = tensor("attn_output_3_cast_fp16")]; + tensor var_943 = const()[name = tensor("op_943"), val = tensor([0, 2, 1, 3])]; + tensor var_948 = const()[name = tensor("op_948"), val = tensor([1, 1, 1024])]; + tensor var_944_cast_fp16 = transpose(perm = var_943, x = attn_output_3_cast_fp16)[name = tensor("transpose_208")]; + tensor input_19_cast_fp16 = reshape(shape = var_948, x = var_944_cast_fp16)[name = tensor("input_19_cast_fp16")]; + tensor attn1_out_proj_weight_to_fp16 = const()[name = tensor("attn1_out_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(32602496)))]; + tensor linear_6_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = attn1_out_proj_weight_to_fp16, x = input_19_cast_fp16)[name = tensor("linear_6_cast_fp16")]; + tensor var_954_to_fp16 = const()[name = tensor("op_954_to_fp16"), val = tensor(0x1p+0)]; + tensor var_955_cast_fp16 = add(x = position1_to_fp16, y = var_954_to_fp16)[name = tensor("op_955_cast_fp16")]; + tensor var_955_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_955_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor input_21_cast_fp16 = add(x = input_15_cast_fp16, y = linear_6_cast_fp16)[name = tensor("input_21_cast_fp16")]; + tensor input_23_axes_0 = const()[name = tensor("input_23_axes_0"), val = tensor([-1])]; + tensor norm1_2_weight_to_fp16 = const()[name = tensor("norm1_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(34699712)))]; + tensor norm1_2_bias_to_fp16 = const()[name = tensor("norm1_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(34701824)))]; + tensor var_959_to_fp16 = const()[name = tensor("op_959_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_23_cast_fp16 = layer_norm(axes = input_23_axes_0, beta = norm1_2_bias_to_fp16, epsilon = var_959_to_fp16, gamma = norm1_2_weight_to_fp16, x = input_21_cast_fp16)[name = tensor("input_23_cast_fp16")]; + tensor linear1_1_weight_to_fp16 = const()[name = tensor("linear1_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(34703936)))]; + tensor linear_7_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = linear1_1_weight_to_fp16, x = input_23_cast_fp16)[name = tensor("linear_7_cast_fp16")]; + tensor input_25_mode_0 = const()[name = tensor("input_25_mode_0"), val = tensor("EXACT")]; + tensor input_25_cast_fp16 = gelu(mode = input_25_mode_0, x = linear_7_cast_fp16)[name = tensor("input_25_cast_fp16")]; + tensor linear1_2_weight_to_fp16 = const()[name = tensor("linear1_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(43092608)))]; + tensor linear_8_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = linear1_2_weight_to_fp16, x = input_25_cast_fp16)[name = tensor("linear_8_cast_fp16")]; + tensor input_27_cast_fp16 = add(x = input_21_cast_fp16, y = linear_8_cast_fp16)[name = tensor("input_27_cast_fp16")]; + tensor input_29_axes_0 = const()[name = tensor("input_29_axes_0"), val = tensor([-1])]; + tensor norm2_1_weight_to_fp16 = const()[name = tensor("norm2_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(51481280)))]; + tensor norm2_1_bias_to_fp16 = const()[name = tensor("norm2_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(51483392)))]; + tensor var_976_to_fp16 = const()[name = tensor("op_976_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_29_cast_fp16 = layer_norm(axes = input_29_axes_0, beta = norm2_1_bias_to_fp16, epsilon = var_976_to_fp16, gamma = norm2_1_weight_to_fp16, x = input_27_cast_fp16)[name = tensor("input_29_cast_fp16")]; + tensor attn2_in_proj_weight_to_fp16 = const()[name = tensor("attn2_in_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(51485504)))]; + tensor linear_9_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = attn2_in_proj_weight_to_fp16, x = input_29_cast_fp16)[name = tensor("linear_9_cast_fp16")]; + tensor var_990 = const()[name = tensor("op_990"), val = tensor([1, 1, 3, 16, 64])]; + tensor qkv_5_cast_fp16 = reshape(shape = var_990, x = linear_9_cast_fp16)[name = tensor("qkv_5_cast_fp16")]; + tensor q_13_begin_0 = const()[name = tensor("q_13_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor q_13_end_0 = const()[name = tensor("q_13_end_0"), val = tensor([1, 1, 1, 16, 64])]; + tensor q_13_end_mask_0 = const()[name = tensor("q_13_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor q_13_squeeze_mask_0 = const()[name = tensor("q_13_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor q_13_cast_fp16 = slice_by_index(begin = q_13_begin_0, end = q_13_end_0, end_mask = q_13_end_mask_0, squeeze_mask = q_13_squeeze_mask_0, x = qkv_5_cast_fp16)[name = tensor("q_13_cast_fp16")]; + tensor k_9_begin_0 = const()[name = tensor("k_9_begin_0"), val = tensor([0, 0, 1, 0, 0])]; + tensor k_9_end_0 = const()[name = tensor("k_9_end_0"), val = tensor([1, 1, 2, 16, 64])]; + tensor k_9_end_mask_0 = const()[name = tensor("k_9_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor k_9_squeeze_mask_0 = const()[name = tensor("k_9_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor k_9_cast_fp16 = slice_by_index(begin = k_9_begin_0, end = k_9_end_0, end_mask = k_9_end_mask_0, squeeze_mask = k_9_squeeze_mask_0, x = qkv_5_cast_fp16)[name = tensor("k_9_cast_fp16")]; + tensor v_5_begin_0 = const()[name = tensor("v_5_begin_0"), val = tensor([0, 0, 2, 0, 0])]; + tensor v_5_end_0 = const()[name = tensor("v_5_end_0"), val = tensor([1, 1, 3, 16, 64])]; + tensor v_5_end_mask_0 = const()[name = tensor("v_5_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor v_5_squeeze_mask_0 = const()[name = tensor("v_5_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor v_5_cast_fp16 = slice_by_index(begin = v_5_begin_0, end = v_5_end_0, end_mask = v_5_end_mask_0, squeeze_mask = v_5_squeeze_mask_0, x = qkv_5_cast_fp16)[name = tensor("v_5_cast_fp16")]; + tensor var_1056 = const()[name = tensor("op_1056"), val = tensor([1, 1, 1, 1])]; + tensor position2_to_fp16_dtype_0 = const()[name = tensor("position2_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor position2_to_fp16 = cast(dtype = position2_to_fp16_dtype_0, x = position2)[name = tensor("cast_543")]; + tensor ts_17_cast_fp16 = reshape(shape = var_1056, x = position2_to_fp16)[name = tensor("ts_17_cast_fp16")]; + tensor var_1063 = const()[name = tensor("op_1063"), val = tensor([1, 1, 16, 32, 2])]; + tensor q_complex_5_cast_fp16 = reshape(shape = var_1063, x = q_13_cast_fp16)[name = tensor("q_complex_5_cast_fp16")]; + tensor var_1070 = const()[name = tensor("op_1070"), val = tensor([1, 1, 16, 32, 2])]; + tensor k_complex_5_cast_fp16 = reshape(shape = var_1070, x = k_9_cast_fp16)[name = tensor("k_complex_5_cast_fp16")]; + tensor var_1074_begin_0 = const()[name = tensor("op_1074_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_1074_end_0 = const()[name = tensor("op_1074_end_0"), val = tensor([1, 1, 16, 32, 1])]; + tensor var_1074_end_mask_0 = const()[name = tensor("op_1074_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_1074_squeeze_mask_0 = const()[name = tensor("op_1074_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_1074_cast_fp16 = slice_by_index(begin = var_1074_begin_0, end = var_1074_end_0, end_mask = var_1074_end_mask_0, squeeze_mask = var_1074_squeeze_mask_0, x = q_complex_5_cast_fp16)[name = tensor("op_1074_cast_fp16")]; + tensor var_1082_begin_0 = const()[name = tensor("op_1082_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_1082_end_0 = const()[name = tensor("op_1082_end_0"), val = tensor([1, 1, 16, 32, 2])]; + tensor var_1082_end_mask_0 = const()[name = tensor("op_1082_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_1082_squeeze_mask_0 = const()[name = tensor("op_1082_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_1082_cast_fp16 = slice_by_index(begin = var_1082_begin_0, end = var_1082_end_0, end_mask = var_1082_end_mask_0, squeeze_mask = var_1082_squeeze_mask_0, x = q_complex_5_cast_fp16)[name = tensor("op_1082_cast_fp16")]; + tensor var_1090_begin_0 = const()[name = tensor("op_1090_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_1090_end_0 = const()[name = tensor("op_1090_end_0"), val = tensor([1, 1, 16, 32, 1])]; + tensor var_1090_end_mask_0 = const()[name = tensor("op_1090_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_1090_squeeze_mask_0 = const()[name = tensor("op_1090_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_1090_cast_fp16 = slice_by_index(begin = var_1090_begin_0, end = var_1090_end_0, end_mask = var_1090_end_mask_0, squeeze_mask = var_1090_squeeze_mask_0, x = k_complex_5_cast_fp16)[name = tensor("op_1090_cast_fp16")]; + tensor var_1098_begin_0 = const()[name = tensor("op_1098_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_1098_end_0 = const()[name = tensor("op_1098_end_0"), val = tensor([1, 1, 16, 32, 2])]; + tensor var_1098_end_mask_0 = const()[name = tensor("op_1098_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_1098_squeeze_mask_0 = const()[name = tensor("op_1098_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_1098_cast_fp16 = slice_by_index(begin = var_1098_begin_0, end = var_1098_end_0, end_mask = var_1098_end_mask_0, squeeze_mask = var_1098_squeeze_mask_0, x = k_complex_5_cast_fp16)[name = tensor("op_1098_cast_fp16")]; + tensor freqs_5_to_fp16 = const()[name = tensor("freqs_5_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(57777024)))]; + tensor var_1104_cast_fp16 = mul(x = freqs_5_to_fp16, y = ts_17_cast_fp16)[name = tensor("op_1104_cast_fp16")]; + tensor rotr_5_cast_fp16 = cos(x = var_1104_cast_fp16)[name = tensor("rotr_5_cast_fp16")]; + tensor roti_5_cast_fp16 = sin(x = var_1104_cast_fp16)[name = tensor("roti_5_cast_fp16")]; + tensor var_1108_cast_fp16 = mul(x = var_1074_cast_fp16, y = rotr_5_cast_fp16)[name = tensor("op_1108_cast_fp16")]; + tensor var_1109_cast_fp16 = mul(x = var_1082_cast_fp16, y = roti_5_cast_fp16)[name = tensor("op_1109_cast_fp16")]; + tensor qor_9_cast_fp16 = sub(x = var_1108_cast_fp16, y = var_1109_cast_fp16)[name = tensor("qor_9_cast_fp16")]; + tensor var_1112_cast_fp16 = mul(x = var_1074_cast_fp16, y = roti_5_cast_fp16)[name = tensor("op_1112_cast_fp16")]; + tensor var_1113_cast_fp16 = mul(x = var_1082_cast_fp16, y = rotr_5_cast_fp16)[name = tensor("op_1113_cast_fp16")]; + tensor qoi_9_cast_fp16 = add(x = var_1112_cast_fp16, y = var_1113_cast_fp16)[name = tensor("qoi_9_cast_fp16")]; + tensor var_1116_cast_fp16 = mul(x = var_1090_cast_fp16, y = rotr_5_cast_fp16)[name = tensor("op_1116_cast_fp16")]; + tensor var_1117_cast_fp16 = mul(x = var_1098_cast_fp16, y = roti_5_cast_fp16)[name = tensor("op_1117_cast_fp16")]; + tensor kor_9_cast_fp16 = sub(x = var_1116_cast_fp16, y = var_1117_cast_fp16)[name = tensor("kor_9_cast_fp16")]; + tensor var_1120_cast_fp16 = mul(x = var_1090_cast_fp16, y = roti_5_cast_fp16)[name = tensor("op_1120_cast_fp16")]; + tensor var_1121_cast_fp16 = mul(x = var_1098_cast_fp16, y = rotr_5_cast_fp16)[name = tensor("op_1121_cast_fp16")]; + tensor koi_9_cast_fp16 = add(x = var_1120_cast_fp16, y = var_1121_cast_fp16)[name = tensor("koi_9_cast_fp16")]; + tensor qo_5_axis_0 = const()[name = tensor("qo_5_axis_0"), val = tensor(-1)]; + tensor qo_5_cast_fp16 = stack(axis = qo_5_axis_0, values = (qor_9_cast_fp16, qoi_9_cast_fp16))[name = tensor("qo_5_cast_fp16")]; + tensor ko_5_axis_0 = const()[name = tensor("ko_5_axis_0"), val = tensor(-1)]; + tensor ko_5_cast_fp16 = stack(axis = ko_5_axis_0, values = (kor_9_cast_fp16, koi_9_cast_fp16))[name = tensor("ko_5_cast_fp16")]; + tensor var_1154 = const()[name = tensor("op_1154"), val = tensor([1, 1, 16, 64])]; + tensor q_15_cast_fp16 = reshape(shape = var_1154, x = qo_5_cast_fp16)[name = tensor("q_15_cast_fp16")]; + tensor var_1160 = const()[name = tensor("op_1160"), val = tensor([1, 1, 16, 64])]; + tensor k_11_cast_fp16 = reshape(shape = var_1160, x = ko_5_cast_fp16)[name = tensor("k_11_cast_fp16")]; + tensor _inversed_1185_y_0_to_fp16 = const()[name = tensor("_inversed_1185_y_0_to_fp16"), val = tensor(0x1p-9)]; + tensor _inversed_1185_cast_fp16 = mul(x = ts_17_cast_fp16, y = _inversed_1185_y_0_to_fp16)[name = tensor("_inversed_1185_cast_fp16")]; + tensor var_1186_cast_fp16 = floor(x = _inversed_1185_cast_fp16)[name = tensor("op_1186_cast_fp16")]; + tensor var_1187_to_fp16 = const()[name = tensor("op_1187_to_fp16"), val = tensor(0x1p+9)]; + tensor var_1188_cast_fp16 = mul(x = var_1186_cast_fp16, y = var_1187_to_fp16)[name = tensor("op_1188_cast_fp16")]; + tensor write_indices_float_11_cast_fp16 = sub(x = ts_17_cast_fp16, y = var_1188_cast_fp16)[name = tensor("write_indices_float_11_cast_fp16")]; + tensor var_1195_dtype_0 = const()[name = tensor("op_1195_dtype_0"), val = tensor("int32")]; + tensor write_indices_5_reps_0 = const()[name = tensor("write_indices_5_reps_0"), val = tensor([1, 1, 16, 64])]; + tensor write_indices_float_11_cast_fp16_to_int32 = cast(dtype = var_1195_dtype_0, x = write_indices_float_11_cast_fp16)[name = tensor("cast_542")]; + tensor write_indices_5 = tile(reps = write_indices_5_reps_0, x = write_indices_float_11_cast_fp16_to_int32)[name = tensor("write_indices_5")]; + tensor var_1205_begin_0 = const()[name = tensor("op_1205_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_1205_end_0 = const()[name = tensor("op_1205_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor var_1205_end_mask_0 = const()[name = tensor("op_1205_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_1205_squeeze_mask_0 = const()[name = tensor("op_1205_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor cache2_to_fp16_dtype_0 = const()[name = tensor("cache2_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor cache2_to_fp16 = cast(dtype = cache2_to_fp16_dtype_0, x = cache2)[name = tensor("cast_541")]; + tensor var_1205_cast_fp16 = slice_by_index(begin = var_1205_begin_0, end = var_1205_end_0, end_mask = var_1205_end_mask_0, squeeze_mask = var_1205_squeeze_mask_0, x = cache2_to_fp16)[name = tensor("op_1205_cast_fp16")]; + tensor var_1207_axis_0 = const()[name = tensor("op_1207_axis_0"), val = tensor(1)]; + tensor var_1207_mode_0 = const()[name = tensor("op_1207_mode_0"), val = tensor("update")]; + tensor var_1207_validate_indices_0 = const()[name = tensor("op_1207_validate_indices_0"), val = tensor(false)]; + tensor var_1207_cast_fp16 = scatter_along_axis(axis = var_1207_axis_0, data = var_1205_cast_fp16, indices = write_indices_5, mode = var_1207_mode_0, updates = k_11_cast_fp16, validate_indices = var_1207_validate_indices_0)[name = tensor("op_1207_cast_fp16")]; + tensor concat_16 = const()[name = tensor("concat_16"), val = tensor([0, 0, 0, 0, 0])]; + tensor concat_17 = const()[name = tensor("concat_17"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_5_internal_tensor_assign_1_stride_0 = const()[name = tensor("new_cache_5_internal_tensor_assign_1_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_5_internal_tensor_assign_1_begin_mask_0 = const()[name = tensor("new_cache_5_internal_tensor_assign_1_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_5_internal_tensor_assign_1_end_mask_0 = const()[name = tensor("new_cache_5_internal_tensor_assign_1_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_5_internal_tensor_assign_1_squeeze_mask_0 = const()[name = tensor("new_cache_5_internal_tensor_assign_1_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_52 = const()[name = tensor("shape_52"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_4 = const()[name = tensor("reduce_prod_4"), val = tensor(1048576)]; + tensor range_1d_4_start_0 = const()[name = tensor("range_1d_4_start_0"), val = tensor(0)]; + tensor range_1d_4_step_0 = const()[name = tensor("range_1d_4_step_0"), val = tensor(1)]; + tensor range_1d_4 = range_1d(end = reduce_prod_4, start = range_1d_4_start_0, step = range_1d_4_step_0)[name = tensor("range_1d_4")]; + tensor reshape_20 = reshape(shape = shape_52, x = range_1d_4)[name = tensor("reshape_20")]; + tensor slice_by_index_4 = slice_by_index(begin = concat_16, begin_mask = new_cache_5_internal_tensor_assign_1_begin_mask_0, end = concat_17, end_mask = new_cache_5_internal_tensor_assign_1_end_mask_0, squeeze_mask = new_cache_5_internal_tensor_assign_1_squeeze_mask_0, stride = new_cache_5_internal_tensor_assign_1_stride_0, x = reshape_20)[name = tensor("slice_by_index_4")]; + tensor reshape_21_shape_0 = const()[name = tensor("reshape_21_shape_0"), val = tensor([-1])]; + tensor reshape_21 = reshape(shape = reshape_21_shape_0, x = slice_by_index_4)[name = tensor("reshape_21")]; + tensor reshape_22_shape_0 = const()[name = tensor("reshape_22_shape_0"), val = tensor([-1])]; + tensor reshape_22_cast_fp16 = reshape(shape = reshape_22_shape_0, x = var_1207_cast_fp16)[name = tensor("reshape_22_cast_fp16")]; + tensor reshape_23_shape_0 = const()[name = tensor("reshape_23_shape_0"), val = tensor([-1])]; + tensor reshape_23_cast_fp16 = reshape(shape = reshape_23_shape_0, x = cache2_to_fp16)[name = tensor("reshape_23_cast_fp16")]; + tensor scatter_4_mode_0 = const()[name = tensor("scatter_4_mode_0"), val = tensor("update")]; + tensor scatter_4_axis_0 = const()[name = tensor("scatter_4_axis_0"), val = tensor(0)]; + tensor scatter_4_validate_indices_0 = const()[name = tensor("scatter_4_validate_indices_0"), val = tensor(false)]; + tensor scatter_4_cast_fp16 = scatter(axis = scatter_4_axis_0, data = reshape_23_cast_fp16, indices = reshape_21, mode = scatter_4_mode_0, updates = reshape_22_cast_fp16, validate_indices = scatter_4_validate_indices_0)[name = tensor("scatter_4_cast_fp16")]; + tensor reshape_24_cast_fp16 = reshape(shape = shape_52, x = scatter_4_cast_fp16)[name = tensor("reshape_24_cast_fp16")]; + tensor var_1215_begin_0 = const()[name = tensor("op_1215_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor var_1215_end_0 = const()[name = tensor("op_1215_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor var_1215_end_mask_0 = const()[name = tensor("op_1215_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_1215_squeeze_mask_0 = const()[name = tensor("op_1215_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor var_1215_cast_fp16 = slice_by_index(begin = var_1215_begin_0, end = var_1215_end_0, end_mask = var_1215_end_mask_0, squeeze_mask = var_1215_squeeze_mask_0, x = reshape_24_cast_fp16)[name = tensor("op_1215_cast_fp16")]; + tensor var_1217_axis_0 = const()[name = tensor("op_1217_axis_0"), val = tensor(1)]; + tensor var_1217_mode_0 = const()[name = tensor("op_1217_mode_0"), val = tensor("update")]; + tensor var_1217_validate_indices_0 = const()[name = tensor("op_1217_validate_indices_0"), val = tensor(false)]; + tensor var_1217_cast_fp16 = scatter_along_axis(axis = var_1217_axis_0, data = var_1215_cast_fp16, indices = write_indices_5, mode = var_1217_mode_0, updates = v_5_cast_fp16, validate_indices = var_1217_validate_indices_0)[name = tensor("op_1217_cast_fp16")]; + tensor concat_18 = const()[name = tensor("concat_18"), val = tensor([1, 0, 0, 0, 0])]; + tensor concat_19 = const()[name = tensor("concat_19"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_5_internal_tensor_assign_2_stride_0 = const()[name = tensor("new_cache_5_internal_tensor_assign_2_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_5_internal_tensor_assign_2_begin_mask_0 = const()[name = tensor("new_cache_5_internal_tensor_assign_2_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_5_internal_tensor_assign_2_end_mask_0 = const()[name = tensor("new_cache_5_internal_tensor_assign_2_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_5_internal_tensor_assign_2_squeeze_mask_0 = const()[name = tensor("new_cache_5_internal_tensor_assign_2_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_53 = const()[name = tensor("shape_53"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_5 = const()[name = tensor("reduce_prod_5"), val = tensor(1048576)]; + tensor range_1d_5_start_0 = const()[name = tensor("range_1d_5_start_0"), val = tensor(0)]; + tensor range_1d_5_step_0 = const()[name = tensor("range_1d_5_step_0"), val = tensor(1)]; + tensor range_1d_5 = range_1d(end = reduce_prod_5, start = range_1d_5_start_0, step = range_1d_5_step_0)[name = tensor("range_1d_5")]; + tensor reshape_25 = reshape(shape = shape_53, x = range_1d_5)[name = tensor("reshape_25")]; + tensor slice_by_index_5 = slice_by_index(begin = concat_18, begin_mask = new_cache_5_internal_tensor_assign_2_begin_mask_0, end = concat_19, end_mask = new_cache_5_internal_tensor_assign_2_end_mask_0, squeeze_mask = new_cache_5_internal_tensor_assign_2_squeeze_mask_0, stride = new_cache_5_internal_tensor_assign_2_stride_0, x = reshape_25)[name = tensor("slice_by_index_5")]; + tensor reshape_26_shape_0 = const()[name = tensor("reshape_26_shape_0"), val = tensor([-1])]; + tensor reshape_26 = reshape(shape = reshape_26_shape_0, x = slice_by_index_5)[name = tensor("reshape_26")]; + tensor reshape_27_shape_0 = const()[name = tensor("reshape_27_shape_0"), val = tensor([-1])]; + tensor reshape_27_cast_fp16 = reshape(shape = reshape_27_shape_0, x = var_1217_cast_fp16)[name = tensor("reshape_27_cast_fp16")]; + tensor reshape_28_shape_0 = const()[name = tensor("reshape_28_shape_0"), val = tensor([-1])]; + tensor reshape_28_cast_fp16 = reshape(shape = reshape_28_shape_0, x = reshape_24_cast_fp16)[name = tensor("reshape_28_cast_fp16")]; + tensor scatter_5_mode_0 = const()[name = tensor("scatter_5_mode_0"), val = tensor("update")]; + tensor scatter_5_axis_0 = const()[name = tensor("scatter_5_axis_0"), val = tensor(0)]; + tensor scatter_5_validate_indices_0 = const()[name = tensor("scatter_5_validate_indices_0"), val = tensor(false)]; + tensor scatter_5_cast_fp16 = scatter(axis = scatter_5_axis_0, data = reshape_28_cast_fp16, indices = reshape_26, mode = scatter_5_mode_0, updates = reshape_27_cast_fp16, validate_indices = scatter_5_validate_indices_0)[name = tensor("scatter_5_cast_fp16")]; + tensor reshape_29_cast_fp16 = reshape(shape = shape_53, x = scatter_5_cast_fp16)[name = tensor("reshape_29_cast_fp16")]; + tensor reshape_29_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("reshape_29_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor keys_13_begin_0 = const()[name = tensor("keys_13_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor keys_13_end_0 = const()[name = tensor("keys_13_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor keys_13_end_mask_0 = const()[name = tensor("keys_13_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor keys_13_squeeze_mask_0 = const()[name = tensor("keys_13_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor keys_13_cast_fp16 = slice_by_index(begin = keys_13_begin_0, end = keys_13_end_0, end_mask = keys_13_end_mask_0, squeeze_mask = keys_13_squeeze_mask_0, x = reshape_29_cast_fp16)[name = tensor("keys_13_cast_fp16")]; + tensor values_13_begin_0 = const()[name = tensor("values_13_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor values_13_end_0 = const()[name = tensor("values_13_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor values_13_end_mask_0 = const()[name = tensor("values_13_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor values_13_squeeze_mask_0 = const()[name = tensor("values_13_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor values_13_cast_fp16 = slice_by_index(begin = values_13_begin_0, end = values_13_end_0, end_mask = values_13_end_mask_0, squeeze_mask = values_13_squeeze_mask_0, x = reshape_29_cast_fp16)[name = tensor("values_13_cast_fp16")]; + tensor var_1229_cast_fp16 = not_equal(x = keys_13_cast_fp16, y = keys_13_cast_fp16)[name = tensor("op_1229_cast_fp16")]; + tensor keys_15_cast_fp16 = select(a = var_463_to_fp16, b = keys_13_cast_fp16, cond = var_1229_cast_fp16)[name = tensor("keys_15_cast_fp16")]; + tensor var_1237_cast_fp16 = not_equal(x = values_13_cast_fp16, y = values_13_cast_fp16)[name = tensor("op_1237_cast_fp16")]; + tensor values_15_cast_fp16 = select(a = var_463_to_fp16, b = values_13_cast_fp16, cond = var_1237_cast_fp16)[name = tensor("values_15_cast_fp16")]; + tensor var_1261 = const()[name = tensor("op_1261"), val = tensor([0, 2, 1, 3])]; + tensor var_1277 = const()[name = tensor("op_1277"), val = tensor([1, 1, 1])]; + tensor var_1278_cast_fp16 = reshape(shape = var_1277, x = position2_to_fp16)[name = tensor("op_1278_cast_fp16")]; + tensor var_1298_to_fp16 = const()[name = tensor("op_1298_to_fp16"), val = tensor(0x1p+0)]; + tensor valid_len_5_cast_fp16 = add(x = var_1278_cast_fp16, y = var_1298_to_fp16)[name = tensor("valid_len_5_cast_fp16")]; + tensor valid_mask_5_cast_fp16 = less(x = k_positions_1_promoted_to_fp16, y = valid_len_5_cast_fp16)[name = tensor("valid_mask_5_cast_fp16")]; + tensor causal_mask_5_cast_fp16 = less_equal(x = k_positions_1_promoted_to_fp16, y = var_1278_cast_fp16)[name = tensor("causal_mask_5_cast_fp16")]; + tensor attn_mask_9 = logical_and(x = valid_mask_5_cast_fp16, y = causal_mask_5_cast_fp16)[name = tensor("attn_mask_9")]; + tensor attn_mask_11_axes_0 = const()[name = tensor("attn_mask_11_axes_0"), val = tensor([1])]; + tensor attn_mask_11 = expand_dims(axes = attn_mask_11_axes_0, x = attn_mask_9)[name = tensor("attn_mask_11")]; + tensor var_1316_transpose_x_0 = const()[name = tensor("op_1316_transpose_x_0"), val = tensor(false)]; + tensor var_1316_transpose_y_0 = const()[name = tensor("op_1316_transpose_y_0"), val = tensor(false)]; + tensor transpose_76_perm_0 = const()[name = tensor("transpose_76_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_77_perm_0 = const()[name = tensor("transpose_77_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_77 = transpose(perm = transpose_77_perm_0, x = keys_15_cast_fp16)[name = tensor("transpose_205")]; + tensor transpose_76 = transpose(perm = transpose_76_perm_0, x = q_15_cast_fp16)[name = tensor("transpose_206")]; + tensor var_1316_cast_fp16 = matmul(transpose_x = var_1316_transpose_x_0, transpose_y = var_1316_transpose_y_0, x = transpose_76, y = transpose_77)[name = tensor("op_1316_cast_fp16")]; + tensor var_1310_to_fp16 = const()[name = tensor("op_1310_to_fp16"), val = tensor([0x1p-3])]; + tensor attn_weights_13_cast_fp16 = mul(x = var_1316_cast_fp16, y = var_1310_to_fp16)[name = tensor("attn_weights_13_cast_fp16")]; + tensor var_1318 = logical_not(x = attn_mask_11)[name = tensor("op_1318")]; + tensor var_1319_to_fp16 = const()[name = tensor("op_1319_to_fp16"), val = tensor(-inf)]; + tensor attn_weights_15_cast_fp16 = select(a = var_1319_to_fp16, b = attn_weights_13_cast_fp16, cond = var_1318)[name = tensor("attn_weights_15_cast_fp16")]; + tensor var_1321 = const()[name = tensor("op_1321"), val = tensor(-1)]; + tensor attn_weights_17_cast_fp16 = softmax(axis = var_1321, x = attn_weights_15_cast_fp16)[name = tensor("attn_weights_17_cast_fp16")]; + tensor attn_output_5_transpose_x_0 = const()[name = tensor("attn_output_5_transpose_x_0"), val = tensor(false)]; + tensor attn_output_5_transpose_y_0 = const()[name = tensor("attn_output_5_transpose_y_0"), val = tensor(false)]; + tensor values_17_cast_fp16 = transpose(perm = var_1261, x = values_15_cast_fp16)[name = tensor("transpose_207")]; + tensor attn_output_5_cast_fp16 = matmul(transpose_x = attn_output_5_transpose_x_0, transpose_y = attn_output_5_transpose_y_0, x = attn_weights_17_cast_fp16, y = values_17_cast_fp16)[name = tensor("attn_output_5_cast_fp16")]; + tensor var_1329 = const()[name = tensor("op_1329"), val = tensor([0, 2, 1, 3])]; + tensor var_1334 = const()[name = tensor("op_1334"), val = tensor([1, 1, 1024])]; + tensor var_1330_cast_fp16 = transpose(perm = var_1329, x = attn_output_5_cast_fp16)[name = tensor("transpose_204")]; + tensor input_31_cast_fp16 = reshape(shape = var_1334, x = var_1330_cast_fp16)[name = tensor("input_31_cast_fp16")]; + tensor attn2_out_proj_weight_to_fp16 = const()[name = tensor("attn2_out_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(57777152)))]; + tensor linear_10_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = attn2_out_proj_weight_to_fp16, x = input_31_cast_fp16)[name = tensor("linear_10_cast_fp16")]; + tensor var_1340_to_fp16 = const()[name = tensor("op_1340_to_fp16"), val = tensor(0x1p+0)]; + tensor var_1341_cast_fp16 = add(x = position2_to_fp16, y = var_1340_to_fp16)[name = tensor("op_1341_cast_fp16")]; + tensor var_1341_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_1341_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor input_33_cast_fp16 = add(x = input_27_cast_fp16, y = linear_10_cast_fp16)[name = tensor("input_33_cast_fp16")]; + tensor input_35_axes_0 = const()[name = tensor("input_35_axes_0"), val = tensor([-1])]; + tensor norm2_2_weight_to_fp16 = const()[name = tensor("norm2_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(59874368)))]; + tensor norm2_2_bias_to_fp16 = const()[name = tensor("norm2_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(59876480)))]; + tensor var_1345_to_fp16 = const()[name = tensor("op_1345_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_35_cast_fp16 = layer_norm(axes = input_35_axes_0, beta = norm2_2_bias_to_fp16, epsilon = var_1345_to_fp16, gamma = norm2_2_weight_to_fp16, x = input_33_cast_fp16)[name = tensor("input_35_cast_fp16")]; + tensor linear2_1_weight_to_fp16 = const()[name = tensor("linear2_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(59878592)))]; + tensor linear_11_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = linear2_1_weight_to_fp16, x = input_35_cast_fp16)[name = tensor("linear_11_cast_fp16")]; + tensor input_37_mode_0 = const()[name = tensor("input_37_mode_0"), val = tensor("EXACT")]; + tensor input_37_cast_fp16 = gelu(mode = input_37_mode_0, x = linear_11_cast_fp16)[name = tensor("input_37_cast_fp16")]; + tensor linear2_2_weight_to_fp16 = const()[name = tensor("linear2_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(68267264)))]; + tensor linear_12_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = linear2_2_weight_to_fp16, x = input_37_cast_fp16)[name = tensor("linear_12_cast_fp16")]; + tensor input_39_cast_fp16 = add(x = input_33_cast_fp16, y = linear_12_cast_fp16)[name = tensor("input_39_cast_fp16")]; + tensor input_41_axes_0 = const()[name = tensor("input_41_axes_0"), val = tensor([-1])]; + tensor norm3_1_weight_to_fp16 = const()[name = tensor("norm3_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(76655936)))]; + tensor norm3_1_bias_to_fp16 = const()[name = tensor("norm3_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(76658048)))]; + tensor var_1362_to_fp16 = const()[name = tensor("op_1362_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_41_cast_fp16 = layer_norm(axes = input_41_axes_0, beta = norm3_1_bias_to_fp16, epsilon = var_1362_to_fp16, gamma = norm3_1_weight_to_fp16, x = input_39_cast_fp16)[name = tensor("input_41_cast_fp16")]; + tensor attn3_in_proj_weight_to_fp16 = const()[name = tensor("attn3_in_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(76660160)))]; + tensor linear_13_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = attn3_in_proj_weight_to_fp16, x = input_41_cast_fp16)[name = tensor("linear_13_cast_fp16")]; + tensor var_1376 = const()[name = tensor("op_1376"), val = tensor([1, 1, 3, 16, 64])]; + tensor qkv_7_cast_fp16 = reshape(shape = var_1376, x = linear_13_cast_fp16)[name = tensor("qkv_7_cast_fp16")]; + tensor q_19_begin_0 = const()[name = tensor("q_19_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor q_19_end_0 = const()[name = tensor("q_19_end_0"), val = tensor([1, 1, 1, 16, 64])]; + tensor q_19_end_mask_0 = const()[name = tensor("q_19_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor q_19_squeeze_mask_0 = const()[name = tensor("q_19_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor q_19_cast_fp16 = slice_by_index(begin = q_19_begin_0, end = q_19_end_0, end_mask = q_19_end_mask_0, squeeze_mask = q_19_squeeze_mask_0, x = qkv_7_cast_fp16)[name = tensor("q_19_cast_fp16")]; + tensor k_13_begin_0 = const()[name = tensor("k_13_begin_0"), val = tensor([0, 0, 1, 0, 0])]; + tensor k_13_end_0 = const()[name = tensor("k_13_end_0"), val = tensor([1, 1, 2, 16, 64])]; + tensor k_13_end_mask_0 = const()[name = tensor("k_13_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor k_13_squeeze_mask_0 = const()[name = tensor("k_13_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor k_13_cast_fp16 = slice_by_index(begin = k_13_begin_0, end = k_13_end_0, end_mask = k_13_end_mask_0, squeeze_mask = k_13_squeeze_mask_0, x = qkv_7_cast_fp16)[name = tensor("k_13_cast_fp16")]; + tensor v_7_begin_0 = const()[name = tensor("v_7_begin_0"), val = tensor([0, 0, 2, 0, 0])]; + tensor v_7_end_0 = const()[name = tensor("v_7_end_0"), val = tensor([1, 1, 3, 16, 64])]; + tensor v_7_end_mask_0 = const()[name = tensor("v_7_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor v_7_squeeze_mask_0 = const()[name = tensor("v_7_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor v_7_cast_fp16 = slice_by_index(begin = v_7_begin_0, end = v_7_end_0, end_mask = v_7_end_mask_0, squeeze_mask = v_7_squeeze_mask_0, x = qkv_7_cast_fp16)[name = tensor("v_7_cast_fp16")]; + tensor var_1442 = const()[name = tensor("op_1442"), val = tensor([1, 1, 1, 1])]; + tensor position3_to_fp16_dtype_0 = const()[name = tensor("position3_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor position3_to_fp16 = cast(dtype = position3_to_fp16_dtype_0, x = position3)[name = tensor("cast_538")]; + tensor ts_23_cast_fp16 = reshape(shape = var_1442, x = position3_to_fp16)[name = tensor("ts_23_cast_fp16")]; + tensor var_1449 = const()[name = tensor("op_1449"), val = tensor([1, 1, 16, 32, 2])]; + tensor q_complex_7_cast_fp16 = reshape(shape = var_1449, x = q_19_cast_fp16)[name = tensor("q_complex_7_cast_fp16")]; + tensor var_1456 = const()[name = tensor("op_1456"), val = tensor([1, 1, 16, 32, 2])]; + tensor k_complex_7_cast_fp16 = reshape(shape = var_1456, x = k_13_cast_fp16)[name = tensor("k_complex_7_cast_fp16")]; + tensor var_1460_begin_0 = const()[name = tensor("op_1460_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_1460_end_0 = const()[name = tensor("op_1460_end_0"), val = tensor([1, 1, 16, 32, 1])]; + tensor var_1460_end_mask_0 = const()[name = tensor("op_1460_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_1460_squeeze_mask_0 = const()[name = tensor("op_1460_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_1460_cast_fp16 = slice_by_index(begin = var_1460_begin_0, end = var_1460_end_0, end_mask = var_1460_end_mask_0, squeeze_mask = var_1460_squeeze_mask_0, x = q_complex_7_cast_fp16)[name = tensor("op_1460_cast_fp16")]; + tensor var_1468_begin_0 = const()[name = tensor("op_1468_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_1468_end_0 = const()[name = tensor("op_1468_end_0"), val = tensor([1, 1, 16, 32, 2])]; + tensor var_1468_end_mask_0 = const()[name = tensor("op_1468_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_1468_squeeze_mask_0 = const()[name = tensor("op_1468_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_1468_cast_fp16 = slice_by_index(begin = var_1468_begin_0, end = var_1468_end_0, end_mask = var_1468_end_mask_0, squeeze_mask = var_1468_squeeze_mask_0, x = q_complex_7_cast_fp16)[name = tensor("op_1468_cast_fp16")]; + tensor var_1476_begin_0 = const()[name = tensor("op_1476_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_1476_end_0 = const()[name = tensor("op_1476_end_0"), val = tensor([1, 1, 16, 32, 1])]; + tensor var_1476_end_mask_0 = const()[name = tensor("op_1476_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_1476_squeeze_mask_0 = const()[name = tensor("op_1476_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_1476_cast_fp16 = slice_by_index(begin = var_1476_begin_0, end = var_1476_end_0, end_mask = var_1476_end_mask_0, squeeze_mask = var_1476_squeeze_mask_0, x = k_complex_7_cast_fp16)[name = tensor("op_1476_cast_fp16")]; + tensor var_1484_begin_0 = const()[name = tensor("op_1484_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_1484_end_0 = const()[name = tensor("op_1484_end_0"), val = tensor([1, 1, 16, 32, 2])]; + tensor var_1484_end_mask_0 = const()[name = tensor("op_1484_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_1484_squeeze_mask_0 = const()[name = tensor("op_1484_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_1484_cast_fp16 = slice_by_index(begin = var_1484_begin_0, end = var_1484_end_0, end_mask = var_1484_end_mask_0, squeeze_mask = var_1484_squeeze_mask_0, x = k_complex_7_cast_fp16)[name = tensor("op_1484_cast_fp16")]; + tensor freqs_7_to_fp16 = const()[name = tensor("freqs_7_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(82951680)))]; + tensor var_1490_cast_fp16 = mul(x = freqs_7_to_fp16, y = ts_23_cast_fp16)[name = tensor("op_1490_cast_fp16")]; + tensor rotr_7_cast_fp16 = cos(x = var_1490_cast_fp16)[name = tensor("rotr_7_cast_fp16")]; + tensor roti_7_cast_fp16 = sin(x = var_1490_cast_fp16)[name = tensor("roti_7_cast_fp16")]; + tensor var_1494_cast_fp16 = mul(x = var_1460_cast_fp16, y = rotr_7_cast_fp16)[name = tensor("op_1494_cast_fp16")]; + tensor var_1495_cast_fp16 = mul(x = var_1468_cast_fp16, y = roti_7_cast_fp16)[name = tensor("op_1495_cast_fp16")]; + tensor qor_13_cast_fp16 = sub(x = var_1494_cast_fp16, y = var_1495_cast_fp16)[name = tensor("qor_13_cast_fp16")]; + tensor var_1498_cast_fp16 = mul(x = var_1460_cast_fp16, y = roti_7_cast_fp16)[name = tensor("op_1498_cast_fp16")]; + tensor var_1499_cast_fp16 = mul(x = var_1468_cast_fp16, y = rotr_7_cast_fp16)[name = tensor("op_1499_cast_fp16")]; + tensor qoi_13_cast_fp16 = add(x = var_1498_cast_fp16, y = var_1499_cast_fp16)[name = tensor("qoi_13_cast_fp16")]; + tensor var_1502_cast_fp16 = mul(x = var_1476_cast_fp16, y = rotr_7_cast_fp16)[name = tensor("op_1502_cast_fp16")]; + tensor var_1503_cast_fp16 = mul(x = var_1484_cast_fp16, y = roti_7_cast_fp16)[name = tensor("op_1503_cast_fp16")]; + tensor kor_13_cast_fp16 = sub(x = var_1502_cast_fp16, y = var_1503_cast_fp16)[name = tensor("kor_13_cast_fp16")]; + tensor var_1506_cast_fp16 = mul(x = var_1476_cast_fp16, y = roti_7_cast_fp16)[name = tensor("op_1506_cast_fp16")]; + tensor var_1507_cast_fp16 = mul(x = var_1484_cast_fp16, y = rotr_7_cast_fp16)[name = tensor("op_1507_cast_fp16")]; + tensor koi_13_cast_fp16 = add(x = var_1506_cast_fp16, y = var_1507_cast_fp16)[name = tensor("koi_13_cast_fp16")]; + tensor qo_7_axis_0 = const()[name = tensor("qo_7_axis_0"), val = tensor(-1)]; + tensor qo_7_cast_fp16 = stack(axis = qo_7_axis_0, values = (qor_13_cast_fp16, qoi_13_cast_fp16))[name = tensor("qo_7_cast_fp16")]; + tensor ko_7_axis_0 = const()[name = tensor("ko_7_axis_0"), val = tensor(-1)]; + tensor ko_7_cast_fp16 = stack(axis = ko_7_axis_0, values = (kor_13_cast_fp16, koi_13_cast_fp16))[name = tensor("ko_7_cast_fp16")]; + tensor var_1540 = const()[name = tensor("op_1540"), val = tensor([1, 1, 16, 64])]; + tensor q_21_cast_fp16 = reshape(shape = var_1540, x = qo_7_cast_fp16)[name = tensor("q_21_cast_fp16")]; + tensor var_1546 = const()[name = tensor("op_1546"), val = tensor([1, 1, 16, 64])]; + tensor k_15_cast_fp16 = reshape(shape = var_1546, x = ko_7_cast_fp16)[name = tensor("k_15_cast_fp16")]; + tensor _inversed_1571_y_0_to_fp16 = const()[name = tensor("_inversed_1571_y_0_to_fp16"), val = tensor(0x1p-9)]; + tensor _inversed_1571_cast_fp16 = mul(x = ts_23_cast_fp16, y = _inversed_1571_y_0_to_fp16)[name = tensor("_inversed_1571_cast_fp16")]; + tensor var_1572_cast_fp16 = floor(x = _inversed_1571_cast_fp16)[name = tensor("op_1572_cast_fp16")]; + tensor var_1573_to_fp16 = const()[name = tensor("op_1573_to_fp16"), val = tensor(0x1p+9)]; + tensor var_1574_cast_fp16 = mul(x = var_1572_cast_fp16, y = var_1573_to_fp16)[name = tensor("op_1574_cast_fp16")]; + tensor write_indices_float_15_cast_fp16 = sub(x = ts_23_cast_fp16, y = var_1574_cast_fp16)[name = tensor("write_indices_float_15_cast_fp16")]; + tensor var_1581_dtype_0 = const()[name = tensor("op_1581_dtype_0"), val = tensor("int32")]; + tensor write_indices_7_reps_0 = const()[name = tensor("write_indices_7_reps_0"), val = tensor([1, 1, 16, 64])]; + tensor write_indices_float_15_cast_fp16_to_int32 = cast(dtype = var_1581_dtype_0, x = write_indices_float_15_cast_fp16)[name = tensor("cast_537")]; + tensor write_indices_7 = tile(reps = write_indices_7_reps_0, x = write_indices_float_15_cast_fp16_to_int32)[name = tensor("write_indices_7")]; + tensor var_1591_begin_0 = const()[name = tensor("op_1591_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_1591_end_0 = const()[name = tensor("op_1591_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor var_1591_end_mask_0 = const()[name = tensor("op_1591_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_1591_squeeze_mask_0 = const()[name = tensor("op_1591_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor cache3_to_fp16_dtype_0 = const()[name = tensor("cache3_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor cache3_to_fp16 = cast(dtype = cache3_to_fp16_dtype_0, x = cache3)[name = tensor("cast_536")]; + tensor var_1591_cast_fp16 = slice_by_index(begin = var_1591_begin_0, end = var_1591_end_0, end_mask = var_1591_end_mask_0, squeeze_mask = var_1591_squeeze_mask_0, x = cache3_to_fp16)[name = tensor("op_1591_cast_fp16")]; + tensor var_1593_axis_0 = const()[name = tensor("op_1593_axis_0"), val = tensor(1)]; + tensor var_1593_mode_0 = const()[name = tensor("op_1593_mode_0"), val = tensor("update")]; + tensor var_1593_validate_indices_0 = const()[name = tensor("op_1593_validate_indices_0"), val = tensor(false)]; + tensor var_1593_cast_fp16 = scatter_along_axis(axis = var_1593_axis_0, data = var_1591_cast_fp16, indices = write_indices_7, mode = var_1593_mode_0, updates = k_15_cast_fp16, validate_indices = var_1593_validate_indices_0)[name = tensor("op_1593_cast_fp16")]; + tensor concat_23 = const()[name = tensor("concat_23"), val = tensor([0, 0, 0, 0, 0])]; + tensor concat_24 = const()[name = tensor("concat_24"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_7_internal_tensor_assign_1_stride_0 = const()[name = tensor("new_cache_7_internal_tensor_assign_1_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_7_internal_tensor_assign_1_begin_mask_0 = const()[name = tensor("new_cache_7_internal_tensor_assign_1_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_7_internal_tensor_assign_1_end_mask_0 = const()[name = tensor("new_cache_7_internal_tensor_assign_1_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_7_internal_tensor_assign_1_squeeze_mask_0 = const()[name = tensor("new_cache_7_internal_tensor_assign_1_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_54 = const()[name = tensor("shape_54"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_6 = const()[name = tensor("reduce_prod_6"), val = tensor(1048576)]; + tensor range_1d_6_start_0 = const()[name = tensor("range_1d_6_start_0"), val = tensor(0)]; + tensor range_1d_6_step_0 = const()[name = tensor("range_1d_6_step_0"), val = tensor(1)]; + tensor range_1d_6 = range_1d(end = reduce_prod_6, start = range_1d_6_start_0, step = range_1d_6_step_0)[name = tensor("range_1d_6")]; + tensor reshape_30 = reshape(shape = shape_54, x = range_1d_6)[name = tensor("reshape_30")]; + tensor slice_by_index_6 = slice_by_index(begin = concat_23, begin_mask = new_cache_7_internal_tensor_assign_1_begin_mask_0, end = concat_24, end_mask = new_cache_7_internal_tensor_assign_1_end_mask_0, squeeze_mask = new_cache_7_internal_tensor_assign_1_squeeze_mask_0, stride = new_cache_7_internal_tensor_assign_1_stride_0, x = reshape_30)[name = tensor("slice_by_index_6")]; + tensor reshape_31_shape_0 = const()[name = tensor("reshape_31_shape_0"), val = tensor([-1])]; + tensor reshape_31 = reshape(shape = reshape_31_shape_0, x = slice_by_index_6)[name = tensor("reshape_31")]; + tensor reshape_32_shape_0 = const()[name = tensor("reshape_32_shape_0"), val = tensor([-1])]; + tensor reshape_32_cast_fp16 = reshape(shape = reshape_32_shape_0, x = var_1593_cast_fp16)[name = tensor("reshape_32_cast_fp16")]; + tensor reshape_33_shape_0 = const()[name = tensor("reshape_33_shape_0"), val = tensor([-1])]; + tensor reshape_33_cast_fp16 = reshape(shape = reshape_33_shape_0, x = cache3_to_fp16)[name = tensor("reshape_33_cast_fp16")]; + tensor scatter_6_mode_0 = const()[name = tensor("scatter_6_mode_0"), val = tensor("update")]; + tensor scatter_6_axis_0 = const()[name = tensor("scatter_6_axis_0"), val = tensor(0)]; + tensor scatter_6_validate_indices_0 = const()[name = tensor("scatter_6_validate_indices_0"), val = tensor(false)]; + tensor scatter_6_cast_fp16 = scatter(axis = scatter_6_axis_0, data = reshape_33_cast_fp16, indices = reshape_31, mode = scatter_6_mode_0, updates = reshape_32_cast_fp16, validate_indices = scatter_6_validate_indices_0)[name = tensor("scatter_6_cast_fp16")]; + tensor reshape_34_cast_fp16 = reshape(shape = shape_54, x = scatter_6_cast_fp16)[name = tensor("reshape_34_cast_fp16")]; + tensor var_1601_begin_0 = const()[name = tensor("op_1601_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor var_1601_end_0 = const()[name = tensor("op_1601_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor var_1601_end_mask_0 = const()[name = tensor("op_1601_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_1601_squeeze_mask_0 = const()[name = tensor("op_1601_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor var_1601_cast_fp16 = slice_by_index(begin = var_1601_begin_0, end = var_1601_end_0, end_mask = var_1601_end_mask_0, squeeze_mask = var_1601_squeeze_mask_0, x = reshape_34_cast_fp16)[name = tensor("op_1601_cast_fp16")]; + tensor var_1603_axis_0 = const()[name = tensor("op_1603_axis_0"), val = tensor(1)]; + tensor var_1603_mode_0 = const()[name = tensor("op_1603_mode_0"), val = tensor("update")]; + tensor var_1603_validate_indices_0 = const()[name = tensor("op_1603_validate_indices_0"), val = tensor(false)]; + tensor var_1603_cast_fp16 = scatter_along_axis(axis = var_1603_axis_0, data = var_1601_cast_fp16, indices = write_indices_7, mode = var_1603_mode_0, updates = v_7_cast_fp16, validate_indices = var_1603_validate_indices_0)[name = tensor("op_1603_cast_fp16")]; + tensor concat_25 = const()[name = tensor("concat_25"), val = tensor([1, 0, 0, 0, 0])]; + tensor concat_26 = const()[name = tensor("concat_26"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_7_internal_tensor_assign_2_stride_0 = const()[name = tensor("new_cache_7_internal_tensor_assign_2_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_7_internal_tensor_assign_2_begin_mask_0 = const()[name = tensor("new_cache_7_internal_tensor_assign_2_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_7_internal_tensor_assign_2_end_mask_0 = const()[name = tensor("new_cache_7_internal_tensor_assign_2_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_7_internal_tensor_assign_2_squeeze_mask_0 = const()[name = tensor("new_cache_7_internal_tensor_assign_2_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_55 = const()[name = tensor("shape_55"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_7 = const()[name = tensor("reduce_prod_7"), val = tensor(1048576)]; + tensor range_1d_7_start_0 = const()[name = tensor("range_1d_7_start_0"), val = tensor(0)]; + tensor range_1d_7_step_0 = const()[name = tensor("range_1d_7_step_0"), val = tensor(1)]; + tensor range_1d_7 = range_1d(end = reduce_prod_7, start = range_1d_7_start_0, step = range_1d_7_step_0)[name = tensor("range_1d_7")]; + tensor reshape_35 = reshape(shape = shape_55, x = range_1d_7)[name = tensor("reshape_35")]; + tensor slice_by_index_7 = slice_by_index(begin = concat_25, begin_mask = new_cache_7_internal_tensor_assign_2_begin_mask_0, end = concat_26, end_mask = new_cache_7_internal_tensor_assign_2_end_mask_0, squeeze_mask = new_cache_7_internal_tensor_assign_2_squeeze_mask_0, stride = new_cache_7_internal_tensor_assign_2_stride_0, x = reshape_35)[name = tensor("slice_by_index_7")]; + tensor reshape_36_shape_0 = const()[name = tensor("reshape_36_shape_0"), val = tensor([-1])]; + tensor reshape_36 = reshape(shape = reshape_36_shape_0, x = slice_by_index_7)[name = tensor("reshape_36")]; + tensor reshape_37_shape_0 = const()[name = tensor("reshape_37_shape_0"), val = tensor([-1])]; + tensor reshape_37_cast_fp16 = reshape(shape = reshape_37_shape_0, x = var_1603_cast_fp16)[name = tensor("reshape_37_cast_fp16")]; + tensor reshape_38_shape_0 = const()[name = tensor("reshape_38_shape_0"), val = tensor([-1])]; + tensor reshape_38_cast_fp16 = reshape(shape = reshape_38_shape_0, x = reshape_34_cast_fp16)[name = tensor("reshape_38_cast_fp16")]; + tensor scatter_7_mode_0 = const()[name = tensor("scatter_7_mode_0"), val = tensor("update")]; + tensor scatter_7_axis_0 = const()[name = tensor("scatter_7_axis_0"), val = tensor(0)]; + tensor scatter_7_validate_indices_0 = const()[name = tensor("scatter_7_validate_indices_0"), val = tensor(false)]; + tensor scatter_7_cast_fp16 = scatter(axis = scatter_7_axis_0, data = reshape_38_cast_fp16, indices = reshape_36, mode = scatter_7_mode_0, updates = reshape_37_cast_fp16, validate_indices = scatter_7_validate_indices_0)[name = tensor("scatter_7_cast_fp16")]; + tensor reshape_39_cast_fp16 = reshape(shape = shape_55, x = scatter_7_cast_fp16)[name = tensor("reshape_39_cast_fp16")]; + tensor reshape_39_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("reshape_39_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor keys_19_begin_0 = const()[name = tensor("keys_19_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor keys_19_end_0 = const()[name = tensor("keys_19_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor keys_19_end_mask_0 = const()[name = tensor("keys_19_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor keys_19_squeeze_mask_0 = const()[name = tensor("keys_19_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor keys_19_cast_fp16 = slice_by_index(begin = keys_19_begin_0, end = keys_19_end_0, end_mask = keys_19_end_mask_0, squeeze_mask = keys_19_squeeze_mask_0, x = reshape_39_cast_fp16)[name = tensor("keys_19_cast_fp16")]; + tensor values_19_begin_0 = const()[name = tensor("values_19_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor values_19_end_0 = const()[name = tensor("values_19_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor values_19_end_mask_0 = const()[name = tensor("values_19_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor values_19_squeeze_mask_0 = const()[name = tensor("values_19_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor values_19_cast_fp16 = slice_by_index(begin = values_19_begin_0, end = values_19_end_0, end_mask = values_19_end_mask_0, squeeze_mask = values_19_squeeze_mask_0, x = reshape_39_cast_fp16)[name = tensor("values_19_cast_fp16")]; + tensor var_1615_cast_fp16 = not_equal(x = keys_19_cast_fp16, y = keys_19_cast_fp16)[name = tensor("op_1615_cast_fp16")]; + tensor keys_21_cast_fp16 = select(a = var_463_to_fp16, b = keys_19_cast_fp16, cond = var_1615_cast_fp16)[name = tensor("keys_21_cast_fp16")]; + tensor var_1623_cast_fp16 = not_equal(x = values_19_cast_fp16, y = values_19_cast_fp16)[name = tensor("op_1623_cast_fp16")]; + tensor values_21_cast_fp16 = select(a = var_463_to_fp16, b = values_19_cast_fp16, cond = var_1623_cast_fp16)[name = tensor("values_21_cast_fp16")]; + tensor var_1647 = const()[name = tensor("op_1647"), val = tensor([0, 2, 1, 3])]; + tensor var_1663 = const()[name = tensor("op_1663"), val = tensor([1, 1, 1])]; + tensor var_1664_cast_fp16 = reshape(shape = var_1663, x = position3_to_fp16)[name = tensor("op_1664_cast_fp16")]; + tensor var_1684_to_fp16 = const()[name = tensor("op_1684_to_fp16"), val = tensor(0x1p+0)]; + tensor valid_len_7_cast_fp16 = add(x = var_1664_cast_fp16, y = var_1684_to_fp16)[name = tensor("valid_len_7_cast_fp16")]; + tensor valid_mask_7_cast_fp16 = less(x = k_positions_1_promoted_to_fp16, y = valid_len_7_cast_fp16)[name = tensor("valid_mask_7_cast_fp16")]; + tensor causal_mask_7_cast_fp16 = less_equal(x = k_positions_1_promoted_to_fp16, y = var_1664_cast_fp16)[name = tensor("causal_mask_7_cast_fp16")]; + tensor attn_mask_13 = logical_and(x = valid_mask_7_cast_fp16, y = causal_mask_7_cast_fp16)[name = tensor("attn_mask_13")]; + tensor attn_mask_15_axes_0 = const()[name = tensor("attn_mask_15_axes_0"), val = tensor([1])]; + tensor attn_mask_15 = expand_dims(axes = attn_mask_15_axes_0, x = attn_mask_13)[name = tensor("attn_mask_15")]; + tensor var_1702_transpose_x_0 = const()[name = tensor("op_1702_transpose_x_0"), val = tensor(false)]; + tensor var_1702_transpose_y_0 = const()[name = tensor("op_1702_transpose_y_0"), val = tensor(false)]; + tensor transpose_78_perm_0 = const()[name = tensor("transpose_78_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_79_perm_0 = const()[name = tensor("transpose_79_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_79 = transpose(perm = transpose_79_perm_0, x = keys_21_cast_fp16)[name = tensor("transpose_201")]; + tensor transpose_78 = transpose(perm = transpose_78_perm_0, x = q_21_cast_fp16)[name = tensor("transpose_202")]; + tensor var_1702_cast_fp16 = matmul(transpose_x = var_1702_transpose_x_0, transpose_y = var_1702_transpose_y_0, x = transpose_78, y = transpose_79)[name = tensor("op_1702_cast_fp16")]; + tensor var_1696_to_fp16 = const()[name = tensor("op_1696_to_fp16"), val = tensor([0x1p-3])]; + tensor attn_weights_19_cast_fp16 = mul(x = var_1702_cast_fp16, y = var_1696_to_fp16)[name = tensor("attn_weights_19_cast_fp16")]; + tensor var_1704 = logical_not(x = attn_mask_15)[name = tensor("op_1704")]; + tensor var_1705_to_fp16 = const()[name = tensor("op_1705_to_fp16"), val = tensor(-inf)]; + tensor attn_weights_21_cast_fp16 = select(a = var_1705_to_fp16, b = attn_weights_19_cast_fp16, cond = var_1704)[name = tensor("attn_weights_21_cast_fp16")]; + tensor var_1707 = const()[name = tensor("op_1707"), val = tensor(-1)]; + tensor attn_weights_23_cast_fp16 = softmax(axis = var_1707, x = attn_weights_21_cast_fp16)[name = tensor("attn_weights_23_cast_fp16")]; + tensor attn_output_7_transpose_x_0 = const()[name = tensor("attn_output_7_transpose_x_0"), val = tensor(false)]; + tensor attn_output_7_transpose_y_0 = const()[name = tensor("attn_output_7_transpose_y_0"), val = tensor(false)]; + tensor values_23_cast_fp16 = transpose(perm = var_1647, x = values_21_cast_fp16)[name = tensor("transpose_203")]; + tensor attn_output_7_cast_fp16 = matmul(transpose_x = attn_output_7_transpose_x_0, transpose_y = attn_output_7_transpose_y_0, x = attn_weights_23_cast_fp16, y = values_23_cast_fp16)[name = tensor("attn_output_7_cast_fp16")]; + tensor var_1715 = const()[name = tensor("op_1715"), val = tensor([0, 2, 1, 3])]; + tensor var_1720 = const()[name = tensor("op_1720"), val = tensor([1, 1, 1024])]; + tensor var_1716_cast_fp16 = transpose(perm = var_1715, x = attn_output_7_cast_fp16)[name = tensor("transpose_200")]; + tensor input_43_cast_fp16 = reshape(shape = var_1720, x = var_1716_cast_fp16)[name = tensor("input_43_cast_fp16")]; + tensor attn3_out_proj_weight_to_fp16 = const()[name = tensor("attn3_out_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(82951808)))]; + tensor linear_14_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = attn3_out_proj_weight_to_fp16, x = input_43_cast_fp16)[name = tensor("linear_14_cast_fp16")]; + tensor var_1726_to_fp16 = const()[name = tensor("op_1726_to_fp16"), val = tensor(0x1p+0)]; + tensor var_1727_cast_fp16 = add(x = position3_to_fp16, y = var_1726_to_fp16)[name = tensor("op_1727_cast_fp16")]; + tensor var_1727_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_1727_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor input_45_cast_fp16 = add(x = input_39_cast_fp16, y = linear_14_cast_fp16)[name = tensor("input_45_cast_fp16")]; + tensor input_47_axes_0 = const()[name = tensor("input_47_axes_0"), val = tensor([-1])]; + tensor norm3_2_weight_to_fp16 = const()[name = tensor("norm3_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(85049024)))]; + tensor norm3_2_bias_to_fp16 = const()[name = tensor("norm3_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(85051136)))]; + tensor var_1731_to_fp16 = const()[name = tensor("op_1731_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_47_cast_fp16 = layer_norm(axes = input_47_axes_0, beta = norm3_2_bias_to_fp16, epsilon = var_1731_to_fp16, gamma = norm3_2_weight_to_fp16, x = input_45_cast_fp16)[name = tensor("input_47_cast_fp16")]; + tensor linear3_1_weight_to_fp16 = const()[name = tensor("linear3_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(85053248)))]; + tensor linear_15_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = linear3_1_weight_to_fp16, x = input_47_cast_fp16)[name = tensor("linear_15_cast_fp16")]; + tensor input_49_mode_0 = const()[name = tensor("input_49_mode_0"), val = tensor("EXACT")]; + tensor input_49_cast_fp16 = gelu(mode = input_49_mode_0, x = linear_15_cast_fp16)[name = tensor("input_49_cast_fp16")]; + tensor linear3_2_weight_to_fp16 = const()[name = tensor("linear3_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(93441920)))]; + tensor linear_16_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = linear3_2_weight_to_fp16, x = input_49_cast_fp16)[name = tensor("linear_16_cast_fp16")]; + tensor input_51_cast_fp16 = add(x = input_45_cast_fp16, y = linear_16_cast_fp16)[name = tensor("input_51_cast_fp16")]; + tensor input_53_axes_0 = const()[name = tensor("input_53_axes_0"), val = tensor([-1])]; + tensor norm4_1_weight_to_fp16 = const()[name = tensor("norm4_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(101830592)))]; + tensor norm4_1_bias_to_fp16 = const()[name = tensor("norm4_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(101832704)))]; + tensor var_1748_to_fp16 = const()[name = tensor("op_1748_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_53_cast_fp16 = layer_norm(axes = input_53_axes_0, beta = norm4_1_bias_to_fp16, epsilon = var_1748_to_fp16, gamma = norm4_1_weight_to_fp16, x = input_51_cast_fp16)[name = tensor("input_53_cast_fp16")]; + tensor attn4_in_proj_weight_to_fp16 = const()[name = tensor("attn4_in_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(101834816)))]; + tensor linear_17_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = attn4_in_proj_weight_to_fp16, x = input_53_cast_fp16)[name = tensor("linear_17_cast_fp16")]; + tensor var_1762 = const()[name = tensor("op_1762"), val = tensor([1, 1, 3, 16, 64])]; + tensor qkv_9_cast_fp16 = reshape(shape = var_1762, x = linear_17_cast_fp16)[name = tensor("qkv_9_cast_fp16")]; + tensor q_25_begin_0 = const()[name = tensor("q_25_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor q_25_end_0 = const()[name = tensor("q_25_end_0"), val = tensor([1, 1, 1, 16, 64])]; + tensor q_25_end_mask_0 = const()[name = tensor("q_25_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor q_25_squeeze_mask_0 = const()[name = tensor("q_25_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor q_25_cast_fp16 = slice_by_index(begin = q_25_begin_0, end = q_25_end_0, end_mask = q_25_end_mask_0, squeeze_mask = q_25_squeeze_mask_0, x = qkv_9_cast_fp16)[name = tensor("q_25_cast_fp16")]; + tensor k_17_begin_0 = const()[name = tensor("k_17_begin_0"), val = tensor([0, 0, 1, 0, 0])]; + tensor k_17_end_0 = const()[name = tensor("k_17_end_0"), val = tensor([1, 1, 2, 16, 64])]; + tensor k_17_end_mask_0 = const()[name = tensor("k_17_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor k_17_squeeze_mask_0 = const()[name = tensor("k_17_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor k_17_cast_fp16 = slice_by_index(begin = k_17_begin_0, end = k_17_end_0, end_mask = k_17_end_mask_0, squeeze_mask = k_17_squeeze_mask_0, x = qkv_9_cast_fp16)[name = tensor("k_17_cast_fp16")]; + tensor v_9_begin_0 = const()[name = tensor("v_9_begin_0"), val = tensor([0, 0, 2, 0, 0])]; + tensor v_9_end_0 = const()[name = tensor("v_9_end_0"), val = tensor([1, 1, 3, 16, 64])]; + tensor v_9_end_mask_0 = const()[name = tensor("v_9_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor v_9_squeeze_mask_0 = const()[name = tensor("v_9_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor v_9_cast_fp16 = slice_by_index(begin = v_9_begin_0, end = v_9_end_0, end_mask = v_9_end_mask_0, squeeze_mask = v_9_squeeze_mask_0, x = qkv_9_cast_fp16)[name = tensor("v_9_cast_fp16")]; + tensor var_1828 = const()[name = tensor("op_1828"), val = tensor([1, 1, 1, 1])]; + tensor position4_to_fp16_dtype_0 = const()[name = tensor("position4_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor position4_to_fp16 = cast(dtype = position4_to_fp16_dtype_0, x = position4)[name = tensor("cast_533")]; + tensor ts_29_cast_fp16 = reshape(shape = var_1828, x = position4_to_fp16)[name = tensor("ts_29_cast_fp16")]; + tensor var_1835 = const()[name = tensor("op_1835"), val = tensor([1, 1, 16, 32, 2])]; + tensor q_complex_9_cast_fp16 = reshape(shape = var_1835, x = q_25_cast_fp16)[name = tensor("q_complex_9_cast_fp16")]; + tensor var_1842 = const()[name = tensor("op_1842"), val = tensor([1, 1, 16, 32, 2])]; + tensor k_complex_9_cast_fp16 = reshape(shape = var_1842, x = k_17_cast_fp16)[name = tensor("k_complex_9_cast_fp16")]; + tensor var_1846_begin_0 = const()[name = tensor("op_1846_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_1846_end_0 = const()[name = tensor("op_1846_end_0"), val = tensor([1, 1, 16, 32, 1])]; + tensor var_1846_end_mask_0 = const()[name = tensor("op_1846_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_1846_squeeze_mask_0 = const()[name = tensor("op_1846_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_1846_cast_fp16 = slice_by_index(begin = var_1846_begin_0, end = var_1846_end_0, end_mask = var_1846_end_mask_0, squeeze_mask = var_1846_squeeze_mask_0, x = q_complex_9_cast_fp16)[name = tensor("op_1846_cast_fp16")]; + tensor var_1854_begin_0 = const()[name = tensor("op_1854_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_1854_end_0 = const()[name = tensor("op_1854_end_0"), val = tensor([1, 1, 16, 32, 2])]; + tensor var_1854_end_mask_0 = const()[name = tensor("op_1854_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_1854_squeeze_mask_0 = const()[name = tensor("op_1854_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_1854_cast_fp16 = slice_by_index(begin = var_1854_begin_0, end = var_1854_end_0, end_mask = var_1854_end_mask_0, squeeze_mask = var_1854_squeeze_mask_0, x = q_complex_9_cast_fp16)[name = tensor("op_1854_cast_fp16")]; + tensor var_1862_begin_0 = const()[name = tensor("op_1862_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_1862_end_0 = const()[name = tensor("op_1862_end_0"), val = tensor([1, 1, 16, 32, 1])]; + tensor var_1862_end_mask_0 = const()[name = tensor("op_1862_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_1862_squeeze_mask_0 = const()[name = tensor("op_1862_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_1862_cast_fp16 = slice_by_index(begin = var_1862_begin_0, end = var_1862_end_0, end_mask = var_1862_end_mask_0, squeeze_mask = var_1862_squeeze_mask_0, x = k_complex_9_cast_fp16)[name = tensor("op_1862_cast_fp16")]; + tensor var_1870_begin_0 = const()[name = tensor("op_1870_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_1870_end_0 = const()[name = tensor("op_1870_end_0"), val = tensor([1, 1, 16, 32, 2])]; + tensor var_1870_end_mask_0 = const()[name = tensor("op_1870_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_1870_squeeze_mask_0 = const()[name = tensor("op_1870_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_1870_cast_fp16 = slice_by_index(begin = var_1870_begin_0, end = var_1870_end_0, end_mask = var_1870_end_mask_0, squeeze_mask = var_1870_squeeze_mask_0, x = k_complex_9_cast_fp16)[name = tensor("op_1870_cast_fp16")]; + tensor freqs_9_to_fp16 = const()[name = tensor("freqs_9_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(108126336)))]; + tensor var_1876_cast_fp16 = mul(x = freqs_9_to_fp16, y = ts_29_cast_fp16)[name = tensor("op_1876_cast_fp16")]; + tensor rotr_9_cast_fp16 = cos(x = var_1876_cast_fp16)[name = tensor("rotr_9_cast_fp16")]; + tensor roti_9_cast_fp16 = sin(x = var_1876_cast_fp16)[name = tensor("roti_9_cast_fp16")]; + tensor var_1880_cast_fp16 = mul(x = var_1846_cast_fp16, y = rotr_9_cast_fp16)[name = tensor("op_1880_cast_fp16")]; + tensor var_1881_cast_fp16 = mul(x = var_1854_cast_fp16, y = roti_9_cast_fp16)[name = tensor("op_1881_cast_fp16")]; + tensor qor_17_cast_fp16 = sub(x = var_1880_cast_fp16, y = var_1881_cast_fp16)[name = tensor("qor_17_cast_fp16")]; + tensor var_1884_cast_fp16 = mul(x = var_1846_cast_fp16, y = roti_9_cast_fp16)[name = tensor("op_1884_cast_fp16")]; + tensor var_1885_cast_fp16 = mul(x = var_1854_cast_fp16, y = rotr_9_cast_fp16)[name = tensor("op_1885_cast_fp16")]; + tensor qoi_17_cast_fp16 = add(x = var_1884_cast_fp16, y = var_1885_cast_fp16)[name = tensor("qoi_17_cast_fp16")]; + tensor var_1888_cast_fp16 = mul(x = var_1862_cast_fp16, y = rotr_9_cast_fp16)[name = tensor("op_1888_cast_fp16")]; + tensor var_1889_cast_fp16 = mul(x = var_1870_cast_fp16, y = roti_9_cast_fp16)[name = tensor("op_1889_cast_fp16")]; + tensor kor_17_cast_fp16 = sub(x = var_1888_cast_fp16, y = var_1889_cast_fp16)[name = tensor("kor_17_cast_fp16")]; + tensor var_1892_cast_fp16 = mul(x = var_1862_cast_fp16, y = roti_9_cast_fp16)[name = tensor("op_1892_cast_fp16")]; + tensor var_1893_cast_fp16 = mul(x = var_1870_cast_fp16, y = rotr_9_cast_fp16)[name = tensor("op_1893_cast_fp16")]; + tensor koi_17_cast_fp16 = add(x = var_1892_cast_fp16, y = var_1893_cast_fp16)[name = tensor("koi_17_cast_fp16")]; + tensor qo_9_axis_0 = const()[name = tensor("qo_9_axis_0"), val = tensor(-1)]; + tensor qo_9_cast_fp16 = stack(axis = qo_9_axis_0, values = (qor_17_cast_fp16, qoi_17_cast_fp16))[name = tensor("qo_9_cast_fp16")]; + tensor ko_9_axis_0 = const()[name = tensor("ko_9_axis_0"), val = tensor(-1)]; + tensor ko_9_cast_fp16 = stack(axis = ko_9_axis_0, values = (kor_17_cast_fp16, koi_17_cast_fp16))[name = tensor("ko_9_cast_fp16")]; + tensor var_1926 = const()[name = tensor("op_1926"), val = tensor([1, 1, 16, 64])]; + tensor q_27_cast_fp16 = reshape(shape = var_1926, x = qo_9_cast_fp16)[name = tensor("q_27_cast_fp16")]; + tensor var_1932 = const()[name = tensor("op_1932"), val = tensor([1, 1, 16, 64])]; + tensor k_19_cast_fp16 = reshape(shape = var_1932, x = ko_9_cast_fp16)[name = tensor("k_19_cast_fp16")]; + tensor _inversed_1957_y_0_to_fp16 = const()[name = tensor("_inversed_1957_y_0_to_fp16"), val = tensor(0x1p-9)]; + tensor _inversed_1957_cast_fp16 = mul(x = ts_29_cast_fp16, y = _inversed_1957_y_0_to_fp16)[name = tensor("_inversed_1957_cast_fp16")]; + tensor var_1958_cast_fp16 = floor(x = _inversed_1957_cast_fp16)[name = tensor("op_1958_cast_fp16")]; + tensor var_1959_to_fp16 = const()[name = tensor("op_1959_to_fp16"), val = tensor(0x1p+9)]; + tensor var_1960_cast_fp16 = mul(x = var_1958_cast_fp16, y = var_1959_to_fp16)[name = tensor("op_1960_cast_fp16")]; + tensor write_indices_float_19_cast_fp16 = sub(x = ts_29_cast_fp16, y = var_1960_cast_fp16)[name = tensor("write_indices_float_19_cast_fp16")]; + tensor var_1967_dtype_0 = const()[name = tensor("op_1967_dtype_0"), val = tensor("int32")]; + tensor write_indices_9_reps_0 = const()[name = tensor("write_indices_9_reps_0"), val = tensor([1, 1, 16, 64])]; + tensor write_indices_float_19_cast_fp16_to_int32 = cast(dtype = var_1967_dtype_0, x = write_indices_float_19_cast_fp16)[name = tensor("cast_532")]; + tensor write_indices_9 = tile(reps = write_indices_9_reps_0, x = write_indices_float_19_cast_fp16_to_int32)[name = tensor("write_indices_9")]; + tensor var_1977_begin_0 = const()[name = tensor("op_1977_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_1977_end_0 = const()[name = tensor("op_1977_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor var_1977_end_mask_0 = const()[name = tensor("op_1977_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_1977_squeeze_mask_0 = const()[name = tensor("op_1977_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor cache4_to_fp16_dtype_0 = const()[name = tensor("cache4_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor cache4_to_fp16 = cast(dtype = cache4_to_fp16_dtype_0, x = cache4)[name = tensor("cast_531")]; + tensor var_1977_cast_fp16 = slice_by_index(begin = var_1977_begin_0, end = var_1977_end_0, end_mask = var_1977_end_mask_0, squeeze_mask = var_1977_squeeze_mask_0, x = cache4_to_fp16)[name = tensor("op_1977_cast_fp16")]; + tensor var_1979_axis_0 = const()[name = tensor("op_1979_axis_0"), val = tensor(1)]; + tensor var_1979_mode_0 = const()[name = tensor("op_1979_mode_0"), val = tensor("update")]; + tensor var_1979_validate_indices_0 = const()[name = tensor("op_1979_validate_indices_0"), val = tensor(false)]; + tensor var_1979_cast_fp16 = scatter_along_axis(axis = var_1979_axis_0, data = var_1977_cast_fp16, indices = write_indices_9, mode = var_1979_mode_0, updates = k_19_cast_fp16, validate_indices = var_1979_validate_indices_0)[name = tensor("op_1979_cast_fp16")]; + tensor concat_30 = const()[name = tensor("concat_30"), val = tensor([0, 0, 0, 0, 0])]; + tensor concat_31 = const()[name = tensor("concat_31"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_9_internal_tensor_assign_1_stride_0 = const()[name = tensor("new_cache_9_internal_tensor_assign_1_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_9_internal_tensor_assign_1_begin_mask_0 = const()[name = tensor("new_cache_9_internal_tensor_assign_1_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_9_internal_tensor_assign_1_end_mask_0 = const()[name = tensor("new_cache_9_internal_tensor_assign_1_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_9_internal_tensor_assign_1_squeeze_mask_0 = const()[name = tensor("new_cache_9_internal_tensor_assign_1_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_56 = const()[name = tensor("shape_56"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_8 = const()[name = tensor("reduce_prod_8"), val = tensor(1048576)]; + tensor range_1d_8_start_0 = const()[name = tensor("range_1d_8_start_0"), val = tensor(0)]; + tensor range_1d_8_step_0 = const()[name = tensor("range_1d_8_step_0"), val = tensor(1)]; + tensor range_1d_8 = range_1d(end = reduce_prod_8, start = range_1d_8_start_0, step = range_1d_8_step_0)[name = tensor("range_1d_8")]; + tensor reshape_40 = reshape(shape = shape_56, x = range_1d_8)[name = tensor("reshape_40")]; + tensor slice_by_index_8 = slice_by_index(begin = concat_30, begin_mask = new_cache_9_internal_tensor_assign_1_begin_mask_0, end = concat_31, end_mask = new_cache_9_internal_tensor_assign_1_end_mask_0, squeeze_mask = new_cache_9_internal_tensor_assign_1_squeeze_mask_0, stride = new_cache_9_internal_tensor_assign_1_stride_0, x = reshape_40)[name = tensor("slice_by_index_8")]; + tensor reshape_41_shape_0 = const()[name = tensor("reshape_41_shape_0"), val = tensor([-1])]; + tensor reshape_41 = reshape(shape = reshape_41_shape_0, x = slice_by_index_8)[name = tensor("reshape_41")]; + tensor reshape_42_shape_0 = const()[name = tensor("reshape_42_shape_0"), val = tensor([-1])]; + tensor reshape_42_cast_fp16 = reshape(shape = reshape_42_shape_0, x = var_1979_cast_fp16)[name = tensor("reshape_42_cast_fp16")]; + tensor reshape_43_shape_0 = const()[name = tensor("reshape_43_shape_0"), val = tensor([-1])]; + tensor reshape_43_cast_fp16 = reshape(shape = reshape_43_shape_0, x = cache4_to_fp16)[name = tensor("reshape_43_cast_fp16")]; + tensor scatter_8_mode_0 = const()[name = tensor("scatter_8_mode_0"), val = tensor("update")]; + tensor scatter_8_axis_0 = const()[name = tensor("scatter_8_axis_0"), val = tensor(0)]; + tensor scatter_8_validate_indices_0 = const()[name = tensor("scatter_8_validate_indices_0"), val = tensor(false)]; + tensor scatter_8_cast_fp16 = scatter(axis = scatter_8_axis_0, data = reshape_43_cast_fp16, indices = reshape_41, mode = scatter_8_mode_0, updates = reshape_42_cast_fp16, validate_indices = scatter_8_validate_indices_0)[name = tensor("scatter_8_cast_fp16")]; + tensor reshape_44_cast_fp16 = reshape(shape = shape_56, x = scatter_8_cast_fp16)[name = tensor("reshape_44_cast_fp16")]; + tensor var_1987_begin_0 = const()[name = tensor("op_1987_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor var_1987_end_0 = const()[name = tensor("op_1987_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor var_1987_end_mask_0 = const()[name = tensor("op_1987_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_1987_squeeze_mask_0 = const()[name = tensor("op_1987_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor var_1987_cast_fp16 = slice_by_index(begin = var_1987_begin_0, end = var_1987_end_0, end_mask = var_1987_end_mask_0, squeeze_mask = var_1987_squeeze_mask_0, x = reshape_44_cast_fp16)[name = tensor("op_1987_cast_fp16")]; + tensor var_1989_axis_0 = const()[name = tensor("op_1989_axis_0"), val = tensor(1)]; + tensor var_1989_mode_0 = const()[name = tensor("op_1989_mode_0"), val = tensor("update")]; + tensor var_1989_validate_indices_0 = const()[name = tensor("op_1989_validate_indices_0"), val = tensor(false)]; + tensor var_1989_cast_fp16 = scatter_along_axis(axis = var_1989_axis_0, data = var_1987_cast_fp16, indices = write_indices_9, mode = var_1989_mode_0, updates = v_9_cast_fp16, validate_indices = var_1989_validate_indices_0)[name = tensor("op_1989_cast_fp16")]; + tensor concat_32 = const()[name = tensor("concat_32"), val = tensor([1, 0, 0, 0, 0])]; + tensor concat_33 = const()[name = tensor("concat_33"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_9_internal_tensor_assign_2_stride_0 = const()[name = tensor("new_cache_9_internal_tensor_assign_2_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_9_internal_tensor_assign_2_begin_mask_0 = const()[name = tensor("new_cache_9_internal_tensor_assign_2_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_9_internal_tensor_assign_2_end_mask_0 = const()[name = tensor("new_cache_9_internal_tensor_assign_2_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_9_internal_tensor_assign_2_squeeze_mask_0 = const()[name = tensor("new_cache_9_internal_tensor_assign_2_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_57 = const()[name = tensor("shape_57"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_9 = const()[name = tensor("reduce_prod_9"), val = tensor(1048576)]; + tensor range_1d_9_start_0 = const()[name = tensor("range_1d_9_start_0"), val = tensor(0)]; + tensor range_1d_9_step_0 = const()[name = tensor("range_1d_9_step_0"), val = tensor(1)]; + tensor range_1d_9 = range_1d(end = reduce_prod_9, start = range_1d_9_start_0, step = range_1d_9_step_0)[name = tensor("range_1d_9")]; + tensor reshape_45 = reshape(shape = shape_57, x = range_1d_9)[name = tensor("reshape_45")]; + tensor slice_by_index_9 = slice_by_index(begin = concat_32, begin_mask = new_cache_9_internal_tensor_assign_2_begin_mask_0, end = concat_33, end_mask = new_cache_9_internal_tensor_assign_2_end_mask_0, squeeze_mask = new_cache_9_internal_tensor_assign_2_squeeze_mask_0, stride = new_cache_9_internal_tensor_assign_2_stride_0, x = reshape_45)[name = tensor("slice_by_index_9")]; + tensor reshape_46_shape_0 = const()[name = tensor("reshape_46_shape_0"), val = tensor([-1])]; + tensor reshape_46 = reshape(shape = reshape_46_shape_0, x = slice_by_index_9)[name = tensor("reshape_46")]; + tensor reshape_47_shape_0 = const()[name = tensor("reshape_47_shape_0"), val = tensor([-1])]; + tensor reshape_47_cast_fp16 = reshape(shape = reshape_47_shape_0, x = var_1989_cast_fp16)[name = tensor("reshape_47_cast_fp16")]; + tensor reshape_48_shape_0 = const()[name = tensor("reshape_48_shape_0"), val = tensor([-1])]; + tensor reshape_48_cast_fp16 = reshape(shape = reshape_48_shape_0, x = reshape_44_cast_fp16)[name = tensor("reshape_48_cast_fp16")]; + tensor scatter_9_mode_0 = const()[name = tensor("scatter_9_mode_0"), val = tensor("update")]; + tensor scatter_9_axis_0 = const()[name = tensor("scatter_9_axis_0"), val = tensor(0)]; + tensor scatter_9_validate_indices_0 = const()[name = tensor("scatter_9_validate_indices_0"), val = tensor(false)]; + tensor scatter_9_cast_fp16 = scatter(axis = scatter_9_axis_0, data = reshape_48_cast_fp16, indices = reshape_46, mode = scatter_9_mode_0, updates = reshape_47_cast_fp16, validate_indices = scatter_9_validate_indices_0)[name = tensor("scatter_9_cast_fp16")]; + tensor reshape_49_cast_fp16 = reshape(shape = shape_57, x = scatter_9_cast_fp16)[name = tensor("reshape_49_cast_fp16")]; + tensor reshape_49_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("reshape_49_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor keys_25_begin_0 = const()[name = tensor("keys_25_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor keys_25_end_0 = const()[name = tensor("keys_25_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor keys_25_end_mask_0 = const()[name = tensor("keys_25_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor keys_25_squeeze_mask_0 = const()[name = tensor("keys_25_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor keys_25_cast_fp16 = slice_by_index(begin = keys_25_begin_0, end = keys_25_end_0, end_mask = keys_25_end_mask_0, squeeze_mask = keys_25_squeeze_mask_0, x = reshape_49_cast_fp16)[name = tensor("keys_25_cast_fp16")]; + tensor values_25_begin_0 = const()[name = tensor("values_25_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor values_25_end_0 = const()[name = tensor("values_25_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor values_25_end_mask_0 = const()[name = tensor("values_25_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor values_25_squeeze_mask_0 = const()[name = tensor("values_25_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor values_25_cast_fp16 = slice_by_index(begin = values_25_begin_0, end = values_25_end_0, end_mask = values_25_end_mask_0, squeeze_mask = values_25_squeeze_mask_0, x = reshape_49_cast_fp16)[name = tensor("values_25_cast_fp16")]; + tensor var_2001_cast_fp16 = not_equal(x = keys_25_cast_fp16, y = keys_25_cast_fp16)[name = tensor("op_2001_cast_fp16")]; + tensor keys_27_cast_fp16 = select(a = var_463_to_fp16, b = keys_25_cast_fp16, cond = var_2001_cast_fp16)[name = tensor("keys_27_cast_fp16")]; + tensor var_2009_cast_fp16 = not_equal(x = values_25_cast_fp16, y = values_25_cast_fp16)[name = tensor("op_2009_cast_fp16")]; + tensor values_27_cast_fp16 = select(a = var_463_to_fp16, b = values_25_cast_fp16, cond = var_2009_cast_fp16)[name = tensor("values_27_cast_fp16")]; + tensor var_2033 = const()[name = tensor("op_2033"), val = tensor([0, 2, 1, 3])]; + tensor var_2049 = const()[name = tensor("op_2049"), val = tensor([1, 1, 1])]; + tensor var_2050_cast_fp16 = reshape(shape = var_2049, x = position4_to_fp16)[name = tensor("op_2050_cast_fp16")]; + tensor var_2070_to_fp16 = const()[name = tensor("op_2070_to_fp16"), val = tensor(0x1p+0)]; + tensor valid_len_9_cast_fp16 = add(x = var_2050_cast_fp16, y = var_2070_to_fp16)[name = tensor("valid_len_9_cast_fp16")]; + tensor valid_mask_9_cast_fp16 = less(x = k_positions_1_promoted_to_fp16, y = valid_len_9_cast_fp16)[name = tensor("valid_mask_9_cast_fp16")]; + tensor causal_mask_9_cast_fp16 = less_equal(x = k_positions_1_promoted_to_fp16, y = var_2050_cast_fp16)[name = tensor("causal_mask_9_cast_fp16")]; + tensor attn_mask_17 = logical_and(x = valid_mask_9_cast_fp16, y = causal_mask_9_cast_fp16)[name = tensor("attn_mask_17")]; + tensor attn_mask_19_axes_0 = const()[name = tensor("attn_mask_19_axes_0"), val = tensor([1])]; + tensor attn_mask_19 = expand_dims(axes = attn_mask_19_axes_0, x = attn_mask_17)[name = tensor("attn_mask_19")]; + tensor var_2088_transpose_x_0 = const()[name = tensor("op_2088_transpose_x_0"), val = tensor(false)]; + tensor var_2088_transpose_y_0 = const()[name = tensor("op_2088_transpose_y_0"), val = tensor(false)]; + tensor transpose_80_perm_0 = const()[name = tensor("transpose_80_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_81_perm_0 = const()[name = tensor("transpose_81_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_81 = transpose(perm = transpose_81_perm_0, x = keys_27_cast_fp16)[name = tensor("transpose_197")]; + tensor transpose_80 = transpose(perm = transpose_80_perm_0, x = q_27_cast_fp16)[name = tensor("transpose_198")]; + tensor var_2088_cast_fp16 = matmul(transpose_x = var_2088_transpose_x_0, transpose_y = var_2088_transpose_y_0, x = transpose_80, y = transpose_81)[name = tensor("op_2088_cast_fp16")]; + tensor var_2082_to_fp16 = const()[name = tensor("op_2082_to_fp16"), val = tensor([0x1p-3])]; + tensor attn_weights_25_cast_fp16 = mul(x = var_2088_cast_fp16, y = var_2082_to_fp16)[name = tensor("attn_weights_25_cast_fp16")]; + tensor var_2090 = logical_not(x = attn_mask_19)[name = tensor("op_2090")]; + tensor var_2091_to_fp16 = const()[name = tensor("op_2091_to_fp16"), val = tensor(-inf)]; + tensor attn_weights_27_cast_fp16 = select(a = var_2091_to_fp16, b = attn_weights_25_cast_fp16, cond = var_2090)[name = tensor("attn_weights_27_cast_fp16")]; + tensor var_2093 = const()[name = tensor("op_2093"), val = tensor(-1)]; + tensor attn_weights_29_cast_fp16 = softmax(axis = var_2093, x = attn_weights_27_cast_fp16)[name = tensor("attn_weights_29_cast_fp16")]; + tensor attn_output_9_transpose_x_0 = const()[name = tensor("attn_output_9_transpose_x_0"), val = tensor(false)]; + tensor attn_output_9_transpose_y_0 = const()[name = tensor("attn_output_9_transpose_y_0"), val = tensor(false)]; + tensor values_29_cast_fp16 = transpose(perm = var_2033, x = values_27_cast_fp16)[name = tensor("transpose_199")]; + tensor attn_output_9_cast_fp16 = matmul(transpose_x = attn_output_9_transpose_x_0, transpose_y = attn_output_9_transpose_y_0, x = attn_weights_29_cast_fp16, y = values_29_cast_fp16)[name = tensor("attn_output_9_cast_fp16")]; + tensor var_2101 = const()[name = tensor("op_2101"), val = tensor([0, 2, 1, 3])]; + tensor var_2106 = const()[name = tensor("op_2106"), val = tensor([1, 1, 1024])]; + tensor var_2102_cast_fp16 = transpose(perm = var_2101, x = attn_output_9_cast_fp16)[name = tensor("transpose_196")]; + tensor input_55_cast_fp16 = reshape(shape = var_2106, x = var_2102_cast_fp16)[name = tensor("input_55_cast_fp16")]; + tensor attn4_out_proj_weight_to_fp16 = const()[name = tensor("attn4_out_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(108126464)))]; + tensor linear_18_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = attn4_out_proj_weight_to_fp16, x = input_55_cast_fp16)[name = tensor("linear_18_cast_fp16")]; + tensor var_2112_to_fp16 = const()[name = tensor("op_2112_to_fp16"), val = tensor(0x1p+0)]; + tensor var_2113_cast_fp16 = add(x = position4_to_fp16, y = var_2112_to_fp16)[name = tensor("op_2113_cast_fp16")]; + tensor var_2113_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_2113_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor input_57_cast_fp16 = add(x = input_51_cast_fp16, y = linear_18_cast_fp16)[name = tensor("input_57_cast_fp16")]; + tensor input_59_axes_0 = const()[name = tensor("input_59_axes_0"), val = tensor([-1])]; + tensor norm4_2_weight_to_fp16 = const()[name = tensor("norm4_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(110223680)))]; + tensor norm4_2_bias_to_fp16 = const()[name = tensor("norm4_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(110225792)))]; + tensor var_2117_to_fp16 = const()[name = tensor("op_2117_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_59_cast_fp16 = layer_norm(axes = input_59_axes_0, beta = norm4_2_bias_to_fp16, epsilon = var_2117_to_fp16, gamma = norm4_2_weight_to_fp16, x = input_57_cast_fp16)[name = tensor("input_59_cast_fp16")]; + tensor linear4_1_weight_to_fp16 = const()[name = tensor("linear4_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(110227904)))]; + tensor linear_19_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = linear4_1_weight_to_fp16, x = input_59_cast_fp16)[name = tensor("linear_19_cast_fp16")]; + tensor input_61_mode_0 = const()[name = tensor("input_61_mode_0"), val = tensor("EXACT")]; + tensor input_61_cast_fp16 = gelu(mode = input_61_mode_0, x = linear_19_cast_fp16)[name = tensor("input_61_cast_fp16")]; + tensor linear4_2_weight_to_fp16 = const()[name = tensor("linear4_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(118616576)))]; + tensor linear_20_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = linear4_2_weight_to_fp16, x = input_61_cast_fp16)[name = tensor("linear_20_cast_fp16")]; + tensor input_63_cast_fp16 = add(x = input_57_cast_fp16, y = linear_20_cast_fp16)[name = tensor("input_63_cast_fp16")]; + tensor input_65_axes_0 = const()[name = tensor("input_65_axes_0"), val = tensor([-1])]; + tensor norm5_1_weight_to_fp16 = const()[name = tensor("norm5_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(127005248)))]; + tensor norm5_1_bias_to_fp16 = const()[name = tensor("norm5_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(127007360)))]; + tensor var_2134_to_fp16 = const()[name = tensor("op_2134_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_65_cast_fp16 = layer_norm(axes = input_65_axes_0, beta = norm5_1_bias_to_fp16, epsilon = var_2134_to_fp16, gamma = norm5_1_weight_to_fp16, x = input_63_cast_fp16)[name = tensor("input_65_cast_fp16")]; + tensor attn5_in_proj_weight_to_fp16 = const()[name = tensor("attn5_in_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(127009472)))]; + tensor linear_21_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = attn5_in_proj_weight_to_fp16, x = input_65_cast_fp16)[name = tensor("linear_21_cast_fp16")]; + tensor var_2148 = const()[name = tensor("op_2148"), val = tensor([1, 1, 3, 16, 64])]; + tensor qkv_11_cast_fp16 = reshape(shape = var_2148, x = linear_21_cast_fp16)[name = tensor("qkv_11_cast_fp16")]; + tensor q_31_begin_0 = const()[name = tensor("q_31_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor q_31_end_0 = const()[name = tensor("q_31_end_0"), val = tensor([1, 1, 1, 16, 64])]; + tensor q_31_end_mask_0 = const()[name = tensor("q_31_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor q_31_squeeze_mask_0 = const()[name = tensor("q_31_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor q_31_cast_fp16 = slice_by_index(begin = q_31_begin_0, end = q_31_end_0, end_mask = q_31_end_mask_0, squeeze_mask = q_31_squeeze_mask_0, x = qkv_11_cast_fp16)[name = tensor("q_31_cast_fp16")]; + tensor k_21_begin_0 = const()[name = tensor("k_21_begin_0"), val = tensor([0, 0, 1, 0, 0])]; + tensor k_21_end_0 = const()[name = tensor("k_21_end_0"), val = tensor([1, 1, 2, 16, 64])]; + tensor k_21_end_mask_0 = const()[name = tensor("k_21_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor k_21_squeeze_mask_0 = const()[name = tensor("k_21_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor k_21_cast_fp16 = slice_by_index(begin = k_21_begin_0, end = k_21_end_0, end_mask = k_21_end_mask_0, squeeze_mask = k_21_squeeze_mask_0, x = qkv_11_cast_fp16)[name = tensor("k_21_cast_fp16")]; + tensor v_11_begin_0 = const()[name = tensor("v_11_begin_0"), val = tensor([0, 0, 2, 0, 0])]; + tensor v_11_end_0 = const()[name = tensor("v_11_end_0"), val = tensor([1, 1, 3, 16, 64])]; + tensor v_11_end_mask_0 = const()[name = tensor("v_11_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor v_11_squeeze_mask_0 = const()[name = tensor("v_11_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor v_11_cast_fp16 = slice_by_index(begin = v_11_begin_0, end = v_11_end_0, end_mask = v_11_end_mask_0, squeeze_mask = v_11_squeeze_mask_0, x = qkv_11_cast_fp16)[name = tensor("v_11_cast_fp16")]; + tensor var_2214 = const()[name = tensor("op_2214"), val = tensor([1, 1, 1, 1])]; + tensor position5_to_fp16_dtype_0 = const()[name = tensor("position5_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor position5_to_fp16 = cast(dtype = position5_to_fp16_dtype_0, x = position5)[name = tensor("cast_528")]; + tensor ts_35_cast_fp16 = reshape(shape = var_2214, x = position5_to_fp16)[name = tensor("ts_35_cast_fp16")]; + tensor var_2221 = const()[name = tensor("op_2221"), val = tensor([1, 1, 16, 32, 2])]; + tensor q_complex_11_cast_fp16 = reshape(shape = var_2221, x = q_31_cast_fp16)[name = tensor("q_complex_11_cast_fp16")]; + tensor var_2228 = const()[name = tensor("op_2228"), val = tensor([1, 1, 16, 32, 2])]; + tensor k_complex_11_cast_fp16 = reshape(shape = var_2228, x = k_21_cast_fp16)[name = tensor("k_complex_11_cast_fp16")]; + tensor var_2232_begin_0 = const()[name = tensor("op_2232_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_2232_end_0 = const()[name = tensor("op_2232_end_0"), val = tensor([1, 1, 16, 32, 1])]; + tensor var_2232_end_mask_0 = const()[name = tensor("op_2232_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_2232_squeeze_mask_0 = const()[name = tensor("op_2232_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_2232_cast_fp16 = slice_by_index(begin = var_2232_begin_0, end = var_2232_end_0, end_mask = var_2232_end_mask_0, squeeze_mask = var_2232_squeeze_mask_0, x = q_complex_11_cast_fp16)[name = tensor("op_2232_cast_fp16")]; + tensor var_2240_begin_0 = const()[name = tensor("op_2240_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_2240_end_0 = const()[name = tensor("op_2240_end_0"), val = tensor([1, 1, 16, 32, 2])]; + tensor var_2240_end_mask_0 = const()[name = tensor("op_2240_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_2240_squeeze_mask_0 = const()[name = tensor("op_2240_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_2240_cast_fp16 = slice_by_index(begin = var_2240_begin_0, end = var_2240_end_0, end_mask = var_2240_end_mask_0, squeeze_mask = var_2240_squeeze_mask_0, x = q_complex_11_cast_fp16)[name = tensor("op_2240_cast_fp16")]; + tensor var_2248_begin_0 = const()[name = tensor("op_2248_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_2248_end_0 = const()[name = tensor("op_2248_end_0"), val = tensor([1, 1, 16, 32, 1])]; + tensor var_2248_end_mask_0 = const()[name = tensor("op_2248_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_2248_squeeze_mask_0 = const()[name = tensor("op_2248_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_2248_cast_fp16 = slice_by_index(begin = var_2248_begin_0, end = var_2248_end_0, end_mask = var_2248_end_mask_0, squeeze_mask = var_2248_squeeze_mask_0, x = k_complex_11_cast_fp16)[name = tensor("op_2248_cast_fp16")]; + tensor var_2256_begin_0 = const()[name = tensor("op_2256_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_2256_end_0 = const()[name = tensor("op_2256_end_0"), val = tensor([1, 1, 16, 32, 2])]; + tensor var_2256_end_mask_0 = const()[name = tensor("op_2256_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_2256_squeeze_mask_0 = const()[name = tensor("op_2256_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_2256_cast_fp16 = slice_by_index(begin = var_2256_begin_0, end = var_2256_end_0, end_mask = var_2256_end_mask_0, squeeze_mask = var_2256_squeeze_mask_0, x = k_complex_11_cast_fp16)[name = tensor("op_2256_cast_fp16")]; + tensor freqs_11_to_fp16 = const()[name = tensor("freqs_11_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(133300992)))]; + tensor var_2262_cast_fp16 = mul(x = freqs_11_to_fp16, y = ts_35_cast_fp16)[name = tensor("op_2262_cast_fp16")]; + tensor rotr_11_cast_fp16 = cos(x = var_2262_cast_fp16)[name = tensor("rotr_11_cast_fp16")]; + tensor roti_11_cast_fp16 = sin(x = var_2262_cast_fp16)[name = tensor("roti_11_cast_fp16")]; + tensor var_2266_cast_fp16 = mul(x = var_2232_cast_fp16, y = rotr_11_cast_fp16)[name = tensor("op_2266_cast_fp16")]; + tensor var_2267_cast_fp16 = mul(x = var_2240_cast_fp16, y = roti_11_cast_fp16)[name = tensor("op_2267_cast_fp16")]; + tensor qor_21_cast_fp16 = sub(x = var_2266_cast_fp16, y = var_2267_cast_fp16)[name = tensor("qor_21_cast_fp16")]; + tensor var_2270_cast_fp16 = mul(x = var_2232_cast_fp16, y = roti_11_cast_fp16)[name = tensor("op_2270_cast_fp16")]; + tensor var_2271_cast_fp16 = mul(x = var_2240_cast_fp16, y = rotr_11_cast_fp16)[name = tensor("op_2271_cast_fp16")]; + tensor qoi_21_cast_fp16 = add(x = var_2270_cast_fp16, y = var_2271_cast_fp16)[name = tensor("qoi_21_cast_fp16")]; + tensor var_2274_cast_fp16 = mul(x = var_2248_cast_fp16, y = rotr_11_cast_fp16)[name = tensor("op_2274_cast_fp16")]; + tensor var_2275_cast_fp16 = mul(x = var_2256_cast_fp16, y = roti_11_cast_fp16)[name = tensor("op_2275_cast_fp16")]; + tensor kor_21_cast_fp16 = sub(x = var_2274_cast_fp16, y = var_2275_cast_fp16)[name = tensor("kor_21_cast_fp16")]; + tensor var_2278_cast_fp16 = mul(x = var_2248_cast_fp16, y = roti_11_cast_fp16)[name = tensor("op_2278_cast_fp16")]; + tensor var_2279_cast_fp16 = mul(x = var_2256_cast_fp16, y = rotr_11_cast_fp16)[name = tensor("op_2279_cast_fp16")]; + tensor koi_21_cast_fp16 = add(x = var_2278_cast_fp16, y = var_2279_cast_fp16)[name = tensor("koi_21_cast_fp16")]; + tensor qo_11_axis_0 = const()[name = tensor("qo_11_axis_0"), val = tensor(-1)]; + tensor qo_11_cast_fp16 = stack(axis = qo_11_axis_0, values = (qor_21_cast_fp16, qoi_21_cast_fp16))[name = tensor("qo_11_cast_fp16")]; + tensor ko_11_axis_0 = const()[name = tensor("ko_11_axis_0"), val = tensor(-1)]; + tensor ko_11_cast_fp16 = stack(axis = ko_11_axis_0, values = (kor_21_cast_fp16, koi_21_cast_fp16))[name = tensor("ko_11_cast_fp16")]; + tensor var_2312 = const()[name = tensor("op_2312"), val = tensor([1, 1, 16, 64])]; + tensor q_33_cast_fp16 = reshape(shape = var_2312, x = qo_11_cast_fp16)[name = tensor("q_33_cast_fp16")]; + tensor var_2318 = const()[name = tensor("op_2318"), val = tensor([1, 1, 16, 64])]; + tensor k_23_cast_fp16 = reshape(shape = var_2318, x = ko_11_cast_fp16)[name = tensor("k_23_cast_fp16")]; + tensor _inversed_2343_y_0_to_fp16 = const()[name = tensor("_inversed_2343_y_0_to_fp16"), val = tensor(0x1p-9)]; + tensor _inversed_2343_cast_fp16 = mul(x = ts_35_cast_fp16, y = _inversed_2343_y_0_to_fp16)[name = tensor("_inversed_2343_cast_fp16")]; + tensor var_2344_cast_fp16 = floor(x = _inversed_2343_cast_fp16)[name = tensor("op_2344_cast_fp16")]; + tensor var_2345_to_fp16 = const()[name = tensor("op_2345_to_fp16"), val = tensor(0x1p+9)]; + tensor var_2346_cast_fp16 = mul(x = var_2344_cast_fp16, y = var_2345_to_fp16)[name = tensor("op_2346_cast_fp16")]; + tensor write_indices_float_23_cast_fp16 = sub(x = ts_35_cast_fp16, y = var_2346_cast_fp16)[name = tensor("write_indices_float_23_cast_fp16")]; + tensor var_2353_dtype_0 = const()[name = tensor("op_2353_dtype_0"), val = tensor("int32")]; + tensor write_indices_11_reps_0 = const()[name = tensor("write_indices_11_reps_0"), val = tensor([1, 1, 16, 64])]; + tensor write_indices_float_23_cast_fp16_to_int32 = cast(dtype = var_2353_dtype_0, x = write_indices_float_23_cast_fp16)[name = tensor("cast_527")]; + tensor write_indices_11 = tile(reps = write_indices_11_reps_0, x = write_indices_float_23_cast_fp16_to_int32)[name = tensor("write_indices_11")]; + tensor var_2363_begin_0 = const()[name = tensor("op_2363_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_2363_end_0 = const()[name = tensor("op_2363_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor var_2363_end_mask_0 = const()[name = tensor("op_2363_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_2363_squeeze_mask_0 = const()[name = tensor("op_2363_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor cache5_to_fp16_dtype_0 = const()[name = tensor("cache5_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor cache5_to_fp16 = cast(dtype = cache5_to_fp16_dtype_0, x = cache5)[name = tensor("cast_526")]; + tensor var_2363_cast_fp16 = slice_by_index(begin = var_2363_begin_0, end = var_2363_end_0, end_mask = var_2363_end_mask_0, squeeze_mask = var_2363_squeeze_mask_0, x = cache5_to_fp16)[name = tensor("op_2363_cast_fp16")]; + tensor var_2365_axis_0 = const()[name = tensor("op_2365_axis_0"), val = tensor(1)]; + tensor var_2365_mode_0 = const()[name = tensor("op_2365_mode_0"), val = tensor("update")]; + tensor var_2365_validate_indices_0 = const()[name = tensor("op_2365_validate_indices_0"), val = tensor(false)]; + tensor var_2365_cast_fp16 = scatter_along_axis(axis = var_2365_axis_0, data = var_2363_cast_fp16, indices = write_indices_11, mode = var_2365_mode_0, updates = k_23_cast_fp16, validate_indices = var_2365_validate_indices_0)[name = tensor("op_2365_cast_fp16")]; + tensor concat_37 = const()[name = tensor("concat_37"), val = tensor([0, 0, 0, 0, 0])]; + tensor concat_38 = const()[name = tensor("concat_38"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_11_internal_tensor_assign_1_stride_0 = const()[name = tensor("new_cache_11_internal_tensor_assign_1_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_11_internal_tensor_assign_1_begin_mask_0 = const()[name = tensor("new_cache_11_internal_tensor_assign_1_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_11_internal_tensor_assign_1_end_mask_0 = const()[name = tensor("new_cache_11_internal_tensor_assign_1_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_11_internal_tensor_assign_1_squeeze_mask_0 = const()[name = tensor("new_cache_11_internal_tensor_assign_1_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_58 = const()[name = tensor("shape_58"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_10 = const()[name = tensor("reduce_prod_10"), val = tensor(1048576)]; + tensor range_1d_10_start_0 = const()[name = tensor("range_1d_10_start_0"), val = tensor(0)]; + tensor range_1d_10_step_0 = const()[name = tensor("range_1d_10_step_0"), val = tensor(1)]; + tensor range_1d_10 = range_1d(end = reduce_prod_10, start = range_1d_10_start_0, step = range_1d_10_step_0)[name = tensor("range_1d_10")]; + tensor reshape_50 = reshape(shape = shape_58, x = range_1d_10)[name = tensor("reshape_50")]; + tensor slice_by_index_10 = slice_by_index(begin = concat_37, begin_mask = new_cache_11_internal_tensor_assign_1_begin_mask_0, end = concat_38, end_mask = new_cache_11_internal_tensor_assign_1_end_mask_0, squeeze_mask = new_cache_11_internal_tensor_assign_1_squeeze_mask_0, stride = new_cache_11_internal_tensor_assign_1_stride_0, x = reshape_50)[name = tensor("slice_by_index_10")]; + tensor reshape_51_shape_0 = const()[name = tensor("reshape_51_shape_0"), val = tensor([-1])]; + tensor reshape_51 = reshape(shape = reshape_51_shape_0, x = slice_by_index_10)[name = tensor("reshape_51")]; + tensor reshape_52_shape_0 = const()[name = tensor("reshape_52_shape_0"), val = tensor([-1])]; + tensor reshape_52_cast_fp16 = reshape(shape = reshape_52_shape_0, x = var_2365_cast_fp16)[name = tensor("reshape_52_cast_fp16")]; + tensor reshape_53_shape_0 = const()[name = tensor("reshape_53_shape_0"), val = tensor([-1])]; + tensor reshape_53_cast_fp16 = reshape(shape = reshape_53_shape_0, x = cache5_to_fp16)[name = tensor("reshape_53_cast_fp16")]; + tensor scatter_10_mode_0 = const()[name = tensor("scatter_10_mode_0"), val = tensor("update")]; + tensor scatter_10_axis_0 = const()[name = tensor("scatter_10_axis_0"), val = tensor(0)]; + tensor scatter_10_validate_indices_0 = const()[name = tensor("scatter_10_validate_indices_0"), val = tensor(false)]; + tensor scatter_10_cast_fp16 = scatter(axis = scatter_10_axis_0, data = reshape_53_cast_fp16, indices = reshape_51, mode = scatter_10_mode_0, updates = reshape_52_cast_fp16, validate_indices = scatter_10_validate_indices_0)[name = tensor("scatter_10_cast_fp16")]; + tensor reshape_54_cast_fp16 = reshape(shape = shape_58, x = scatter_10_cast_fp16)[name = tensor("reshape_54_cast_fp16")]; + tensor var_2373_begin_0 = const()[name = tensor("op_2373_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor var_2373_end_0 = const()[name = tensor("op_2373_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor var_2373_end_mask_0 = const()[name = tensor("op_2373_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_2373_squeeze_mask_0 = const()[name = tensor("op_2373_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor var_2373_cast_fp16 = slice_by_index(begin = var_2373_begin_0, end = var_2373_end_0, end_mask = var_2373_end_mask_0, squeeze_mask = var_2373_squeeze_mask_0, x = reshape_54_cast_fp16)[name = tensor("op_2373_cast_fp16")]; + tensor var_2375_axis_0 = const()[name = tensor("op_2375_axis_0"), val = tensor(1)]; + tensor var_2375_mode_0 = const()[name = tensor("op_2375_mode_0"), val = tensor("update")]; + tensor var_2375_validate_indices_0 = const()[name = tensor("op_2375_validate_indices_0"), val = tensor(false)]; + tensor var_2375_cast_fp16 = scatter_along_axis(axis = var_2375_axis_0, data = var_2373_cast_fp16, indices = write_indices_11, mode = var_2375_mode_0, updates = v_11_cast_fp16, validate_indices = var_2375_validate_indices_0)[name = tensor("op_2375_cast_fp16")]; + tensor concat_39 = const()[name = tensor("concat_39"), val = tensor([1, 0, 0, 0, 0])]; + tensor concat_40 = const()[name = tensor("concat_40"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_11_internal_tensor_assign_2_stride_0 = const()[name = tensor("new_cache_11_internal_tensor_assign_2_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_11_internal_tensor_assign_2_begin_mask_0 = const()[name = tensor("new_cache_11_internal_tensor_assign_2_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_11_internal_tensor_assign_2_end_mask_0 = const()[name = tensor("new_cache_11_internal_tensor_assign_2_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_11_internal_tensor_assign_2_squeeze_mask_0 = const()[name = tensor("new_cache_11_internal_tensor_assign_2_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_59 = const()[name = tensor("shape_59"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_11 = const()[name = tensor("reduce_prod_11"), val = tensor(1048576)]; + tensor range_1d_11_start_0 = const()[name = tensor("range_1d_11_start_0"), val = tensor(0)]; + tensor range_1d_11_step_0 = const()[name = tensor("range_1d_11_step_0"), val = tensor(1)]; + tensor range_1d_11 = range_1d(end = reduce_prod_11, start = range_1d_11_start_0, step = range_1d_11_step_0)[name = tensor("range_1d_11")]; + tensor reshape_55 = reshape(shape = shape_59, x = range_1d_11)[name = tensor("reshape_55")]; + tensor slice_by_index_11 = slice_by_index(begin = concat_39, begin_mask = new_cache_11_internal_tensor_assign_2_begin_mask_0, end = concat_40, end_mask = new_cache_11_internal_tensor_assign_2_end_mask_0, squeeze_mask = new_cache_11_internal_tensor_assign_2_squeeze_mask_0, stride = new_cache_11_internal_tensor_assign_2_stride_0, x = reshape_55)[name = tensor("slice_by_index_11")]; + tensor reshape_56_shape_0 = const()[name = tensor("reshape_56_shape_0"), val = tensor([-1])]; + tensor reshape_56 = reshape(shape = reshape_56_shape_0, x = slice_by_index_11)[name = tensor("reshape_56")]; + tensor reshape_57_shape_0 = const()[name = tensor("reshape_57_shape_0"), val = tensor([-1])]; + tensor reshape_57_cast_fp16 = reshape(shape = reshape_57_shape_0, x = var_2375_cast_fp16)[name = tensor("reshape_57_cast_fp16")]; + tensor reshape_58_shape_0 = const()[name = tensor("reshape_58_shape_0"), val = tensor([-1])]; + tensor reshape_58_cast_fp16 = reshape(shape = reshape_58_shape_0, x = reshape_54_cast_fp16)[name = tensor("reshape_58_cast_fp16")]; + tensor scatter_11_mode_0 = const()[name = tensor("scatter_11_mode_0"), val = tensor("update")]; + tensor scatter_11_axis_0 = const()[name = tensor("scatter_11_axis_0"), val = tensor(0)]; + tensor scatter_11_validate_indices_0 = const()[name = tensor("scatter_11_validate_indices_0"), val = tensor(false)]; + tensor scatter_11_cast_fp16 = scatter(axis = scatter_11_axis_0, data = reshape_58_cast_fp16, indices = reshape_56, mode = scatter_11_mode_0, updates = reshape_57_cast_fp16, validate_indices = scatter_11_validate_indices_0)[name = tensor("scatter_11_cast_fp16")]; + tensor reshape_59_cast_fp16 = reshape(shape = shape_59, x = scatter_11_cast_fp16)[name = tensor("reshape_59_cast_fp16")]; + tensor reshape_59_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("reshape_59_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor keys_31_begin_0 = const()[name = tensor("keys_31_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor keys_31_end_0 = const()[name = tensor("keys_31_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor keys_31_end_mask_0 = const()[name = tensor("keys_31_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor keys_31_squeeze_mask_0 = const()[name = tensor("keys_31_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor keys_31_cast_fp16 = slice_by_index(begin = keys_31_begin_0, end = keys_31_end_0, end_mask = keys_31_end_mask_0, squeeze_mask = keys_31_squeeze_mask_0, x = reshape_59_cast_fp16)[name = tensor("keys_31_cast_fp16")]; + tensor values_31_begin_0 = const()[name = tensor("values_31_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor values_31_end_0 = const()[name = tensor("values_31_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor values_31_end_mask_0 = const()[name = tensor("values_31_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor values_31_squeeze_mask_0 = const()[name = tensor("values_31_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor values_31_cast_fp16 = slice_by_index(begin = values_31_begin_0, end = values_31_end_0, end_mask = values_31_end_mask_0, squeeze_mask = values_31_squeeze_mask_0, x = reshape_59_cast_fp16)[name = tensor("values_31_cast_fp16")]; + tensor var_2387_cast_fp16 = not_equal(x = keys_31_cast_fp16, y = keys_31_cast_fp16)[name = tensor("op_2387_cast_fp16")]; + tensor keys_33_cast_fp16 = select(a = var_463_to_fp16, b = keys_31_cast_fp16, cond = var_2387_cast_fp16)[name = tensor("keys_33_cast_fp16")]; + tensor var_2395_cast_fp16 = not_equal(x = values_31_cast_fp16, y = values_31_cast_fp16)[name = tensor("op_2395_cast_fp16")]; + tensor values_33_cast_fp16 = select(a = var_463_to_fp16, b = values_31_cast_fp16, cond = var_2395_cast_fp16)[name = tensor("values_33_cast_fp16")]; + tensor var_2419 = const()[name = tensor("op_2419"), val = tensor([0, 2, 1, 3])]; + tensor var_2435 = const()[name = tensor("op_2435"), val = tensor([1, 1, 1])]; + tensor var_2436_cast_fp16 = reshape(shape = var_2435, x = position5_to_fp16)[name = tensor("op_2436_cast_fp16")]; + tensor var_2456_to_fp16 = const()[name = tensor("op_2456_to_fp16"), val = tensor(0x1p+0)]; + tensor valid_len_11_cast_fp16 = add(x = var_2436_cast_fp16, y = var_2456_to_fp16)[name = tensor("valid_len_11_cast_fp16")]; + tensor valid_mask_11_cast_fp16 = less(x = k_positions_1_promoted_to_fp16, y = valid_len_11_cast_fp16)[name = tensor("valid_mask_11_cast_fp16")]; + tensor causal_mask_11_cast_fp16 = less_equal(x = k_positions_1_promoted_to_fp16, y = var_2436_cast_fp16)[name = tensor("causal_mask_11_cast_fp16")]; + tensor attn_mask_21 = logical_and(x = valid_mask_11_cast_fp16, y = causal_mask_11_cast_fp16)[name = tensor("attn_mask_21")]; + tensor attn_mask_23_axes_0 = const()[name = tensor("attn_mask_23_axes_0"), val = tensor([1])]; + tensor attn_mask_23 = expand_dims(axes = attn_mask_23_axes_0, x = attn_mask_21)[name = tensor("attn_mask_23")]; + tensor var_2474_transpose_x_0 = const()[name = tensor("op_2474_transpose_x_0"), val = tensor(false)]; + tensor var_2474_transpose_y_0 = const()[name = tensor("op_2474_transpose_y_0"), val = tensor(false)]; + tensor transpose_82_perm_0 = const()[name = tensor("transpose_82_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_83_perm_0 = const()[name = tensor("transpose_83_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_83 = transpose(perm = transpose_83_perm_0, x = keys_33_cast_fp16)[name = tensor("transpose_193")]; + tensor transpose_82 = transpose(perm = transpose_82_perm_0, x = q_33_cast_fp16)[name = tensor("transpose_194")]; + tensor var_2474_cast_fp16 = matmul(transpose_x = var_2474_transpose_x_0, transpose_y = var_2474_transpose_y_0, x = transpose_82, y = transpose_83)[name = tensor("op_2474_cast_fp16")]; + tensor var_2468_to_fp16 = const()[name = tensor("op_2468_to_fp16"), val = tensor([0x1p-3])]; + tensor attn_weights_31_cast_fp16 = mul(x = var_2474_cast_fp16, y = var_2468_to_fp16)[name = tensor("attn_weights_31_cast_fp16")]; + tensor var_2476 = logical_not(x = attn_mask_23)[name = tensor("op_2476")]; + tensor var_2477_to_fp16 = const()[name = tensor("op_2477_to_fp16"), val = tensor(-inf)]; + tensor attn_weights_33_cast_fp16 = select(a = var_2477_to_fp16, b = attn_weights_31_cast_fp16, cond = var_2476)[name = tensor("attn_weights_33_cast_fp16")]; + tensor var_2479 = const()[name = tensor("op_2479"), val = tensor(-1)]; + tensor attn_weights_35_cast_fp16 = softmax(axis = var_2479, x = attn_weights_33_cast_fp16)[name = tensor("attn_weights_35_cast_fp16")]; + tensor attn_output_11_transpose_x_0 = const()[name = tensor("attn_output_11_transpose_x_0"), val = tensor(false)]; + tensor attn_output_11_transpose_y_0 = const()[name = tensor("attn_output_11_transpose_y_0"), val = tensor(false)]; + tensor values_35_cast_fp16 = transpose(perm = var_2419, x = values_33_cast_fp16)[name = tensor("transpose_195")]; + tensor attn_output_11_cast_fp16 = matmul(transpose_x = attn_output_11_transpose_x_0, transpose_y = attn_output_11_transpose_y_0, x = attn_weights_35_cast_fp16, y = values_35_cast_fp16)[name = tensor("attn_output_11_cast_fp16")]; + tensor var_2487 = const()[name = tensor("op_2487"), val = tensor([0, 2, 1, 3])]; + tensor var_2492 = const()[name = tensor("op_2492"), val = tensor([1, 1, 1024])]; + tensor var_2488_cast_fp16 = transpose(perm = var_2487, x = attn_output_11_cast_fp16)[name = tensor("transpose_192")]; + tensor input_67_cast_fp16 = reshape(shape = var_2492, x = var_2488_cast_fp16)[name = tensor("input_67_cast_fp16")]; + tensor attn5_out_proj_weight_to_fp16 = const()[name = tensor("attn5_out_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(133301120)))]; + tensor linear_22_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = attn5_out_proj_weight_to_fp16, x = input_67_cast_fp16)[name = tensor("linear_22_cast_fp16")]; + tensor var_2498_to_fp16 = const()[name = tensor("op_2498_to_fp16"), val = tensor(0x1p+0)]; + tensor var_2499_cast_fp16 = add(x = position5_to_fp16, y = var_2498_to_fp16)[name = tensor("op_2499_cast_fp16")]; + tensor var_2499_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_2499_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor input_69_cast_fp16 = add(x = input_63_cast_fp16, y = linear_22_cast_fp16)[name = tensor("input_69_cast_fp16")]; + tensor input_71_axes_0 = const()[name = tensor("input_71_axes_0"), val = tensor([-1])]; + tensor norm5_2_weight_to_fp16 = const()[name = tensor("norm5_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(135398336)))]; + tensor norm5_2_bias_to_fp16 = const()[name = tensor("norm5_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(135400448)))]; + tensor var_2503_to_fp16 = const()[name = tensor("op_2503_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_71_cast_fp16 = layer_norm(axes = input_71_axes_0, beta = norm5_2_bias_to_fp16, epsilon = var_2503_to_fp16, gamma = norm5_2_weight_to_fp16, x = input_69_cast_fp16)[name = tensor("input_71_cast_fp16")]; + tensor linear5_1_weight_to_fp16 = const()[name = tensor("linear5_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(135402560)))]; + tensor linear_23_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = linear5_1_weight_to_fp16, x = input_71_cast_fp16)[name = tensor("linear_23_cast_fp16")]; + tensor input_73_mode_0 = const()[name = tensor("input_73_mode_0"), val = tensor("EXACT")]; + tensor input_73_cast_fp16 = gelu(mode = input_73_mode_0, x = linear_23_cast_fp16)[name = tensor("input_73_cast_fp16")]; + tensor linear5_2_weight_to_fp16 = const()[name = tensor("linear5_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(143791232)))]; + tensor linear_24_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = linear5_2_weight_to_fp16, x = input_73_cast_fp16)[name = tensor("linear_24_cast_fp16")]; + tensor input_75_cast_fp16 = add(x = input_69_cast_fp16, y = linear_24_cast_fp16)[name = tensor("input_75_cast_fp16")]; + tensor input_77_axes_0 = const()[name = tensor("input_77_axes_0"), val = tensor([-1])]; + tensor norm6_1_weight_to_fp16 = const()[name = tensor("norm6_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(152179904)))]; + tensor norm6_1_bias_to_fp16 = const()[name = tensor("norm6_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(152182016)))]; + tensor var_2520_to_fp16 = const()[name = tensor("op_2520_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_77_cast_fp16 = layer_norm(axes = input_77_axes_0, beta = norm6_1_bias_to_fp16, epsilon = var_2520_to_fp16, gamma = norm6_1_weight_to_fp16, x = input_75_cast_fp16)[name = tensor("input_77_cast_fp16")]; + tensor attn6_in_proj_weight_to_fp16 = const()[name = tensor("attn6_in_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(152184128)))]; + tensor linear_25_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = attn6_in_proj_weight_to_fp16, x = input_77_cast_fp16)[name = tensor("linear_25_cast_fp16")]; + tensor var_2534 = const()[name = tensor("op_2534"), val = tensor([1, 1, 3, 16, 64])]; + tensor qkv_13_cast_fp16 = reshape(shape = var_2534, x = linear_25_cast_fp16)[name = tensor("qkv_13_cast_fp16")]; + tensor q_37_begin_0 = const()[name = tensor("q_37_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor q_37_end_0 = const()[name = tensor("q_37_end_0"), val = tensor([1, 1, 1, 16, 64])]; + tensor q_37_end_mask_0 = const()[name = tensor("q_37_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor q_37_squeeze_mask_0 = const()[name = tensor("q_37_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor q_37_cast_fp16 = slice_by_index(begin = q_37_begin_0, end = q_37_end_0, end_mask = q_37_end_mask_0, squeeze_mask = q_37_squeeze_mask_0, x = qkv_13_cast_fp16)[name = tensor("q_37_cast_fp16")]; + tensor k_25_begin_0 = const()[name = tensor("k_25_begin_0"), val = tensor([0, 0, 1, 0, 0])]; + tensor k_25_end_0 = const()[name = tensor("k_25_end_0"), val = tensor([1, 1, 2, 16, 64])]; + tensor k_25_end_mask_0 = const()[name = tensor("k_25_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor k_25_squeeze_mask_0 = const()[name = tensor("k_25_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor k_25_cast_fp16 = slice_by_index(begin = k_25_begin_0, end = k_25_end_0, end_mask = k_25_end_mask_0, squeeze_mask = k_25_squeeze_mask_0, x = qkv_13_cast_fp16)[name = tensor("k_25_cast_fp16")]; + tensor v_13_begin_0 = const()[name = tensor("v_13_begin_0"), val = tensor([0, 0, 2, 0, 0])]; + tensor v_13_end_0 = const()[name = tensor("v_13_end_0"), val = tensor([1, 1, 3, 16, 64])]; + tensor v_13_end_mask_0 = const()[name = tensor("v_13_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor v_13_squeeze_mask_0 = const()[name = tensor("v_13_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor v_13_cast_fp16 = slice_by_index(begin = v_13_begin_0, end = v_13_end_0, end_mask = v_13_end_mask_0, squeeze_mask = v_13_squeeze_mask_0, x = qkv_13_cast_fp16)[name = tensor("v_13_cast_fp16")]; + tensor var_2600 = const()[name = tensor("op_2600"), val = tensor([1, 1, 1, 1])]; + tensor position6_to_fp16_dtype_0 = const()[name = tensor("position6_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor position6_to_fp16 = cast(dtype = position6_to_fp16_dtype_0, x = position6)[name = tensor("cast_523")]; + tensor ts_41_cast_fp16 = reshape(shape = var_2600, x = position6_to_fp16)[name = tensor("ts_41_cast_fp16")]; + tensor var_2607 = const()[name = tensor("op_2607"), val = tensor([1, 1, 16, 32, 2])]; + tensor q_complex_13_cast_fp16 = reshape(shape = var_2607, x = q_37_cast_fp16)[name = tensor("q_complex_13_cast_fp16")]; + tensor var_2614 = const()[name = tensor("op_2614"), val = tensor([1, 1, 16, 32, 2])]; + tensor k_complex_13_cast_fp16 = reshape(shape = var_2614, x = k_25_cast_fp16)[name = tensor("k_complex_13_cast_fp16")]; + tensor var_2618_begin_0 = const()[name = tensor("op_2618_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_2618_end_0 = const()[name = tensor("op_2618_end_0"), val = tensor([1, 1, 16, 32, 1])]; + tensor var_2618_end_mask_0 = const()[name = tensor("op_2618_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_2618_squeeze_mask_0 = const()[name = tensor("op_2618_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_2618_cast_fp16 = slice_by_index(begin = var_2618_begin_0, end = var_2618_end_0, end_mask = var_2618_end_mask_0, squeeze_mask = var_2618_squeeze_mask_0, x = q_complex_13_cast_fp16)[name = tensor("op_2618_cast_fp16")]; + tensor var_2626_begin_0 = const()[name = tensor("op_2626_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_2626_end_0 = const()[name = tensor("op_2626_end_0"), val = tensor([1, 1, 16, 32, 2])]; + tensor var_2626_end_mask_0 = const()[name = tensor("op_2626_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_2626_squeeze_mask_0 = const()[name = tensor("op_2626_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_2626_cast_fp16 = slice_by_index(begin = var_2626_begin_0, end = var_2626_end_0, end_mask = var_2626_end_mask_0, squeeze_mask = var_2626_squeeze_mask_0, x = q_complex_13_cast_fp16)[name = tensor("op_2626_cast_fp16")]; + tensor var_2634_begin_0 = const()[name = tensor("op_2634_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_2634_end_0 = const()[name = tensor("op_2634_end_0"), val = tensor([1, 1, 16, 32, 1])]; + tensor var_2634_end_mask_0 = const()[name = tensor("op_2634_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_2634_squeeze_mask_0 = const()[name = tensor("op_2634_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_2634_cast_fp16 = slice_by_index(begin = var_2634_begin_0, end = var_2634_end_0, end_mask = var_2634_end_mask_0, squeeze_mask = var_2634_squeeze_mask_0, x = k_complex_13_cast_fp16)[name = tensor("op_2634_cast_fp16")]; + tensor var_2642_begin_0 = const()[name = tensor("op_2642_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_2642_end_0 = const()[name = tensor("op_2642_end_0"), val = tensor([1, 1, 16, 32, 2])]; + tensor var_2642_end_mask_0 = const()[name = tensor("op_2642_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_2642_squeeze_mask_0 = const()[name = tensor("op_2642_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_2642_cast_fp16 = slice_by_index(begin = var_2642_begin_0, end = var_2642_end_0, end_mask = var_2642_end_mask_0, squeeze_mask = var_2642_squeeze_mask_0, x = k_complex_13_cast_fp16)[name = tensor("op_2642_cast_fp16")]; + tensor freqs_13_to_fp16 = const()[name = tensor("freqs_13_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(158475648)))]; + tensor var_2648_cast_fp16 = mul(x = freqs_13_to_fp16, y = ts_41_cast_fp16)[name = tensor("op_2648_cast_fp16")]; + tensor rotr_13_cast_fp16 = cos(x = var_2648_cast_fp16)[name = tensor("rotr_13_cast_fp16")]; + tensor roti_13_cast_fp16 = sin(x = var_2648_cast_fp16)[name = tensor("roti_13_cast_fp16")]; + tensor var_2652_cast_fp16 = mul(x = var_2618_cast_fp16, y = rotr_13_cast_fp16)[name = tensor("op_2652_cast_fp16")]; + tensor var_2653_cast_fp16 = mul(x = var_2626_cast_fp16, y = roti_13_cast_fp16)[name = tensor("op_2653_cast_fp16")]; + tensor qor_25_cast_fp16 = sub(x = var_2652_cast_fp16, y = var_2653_cast_fp16)[name = tensor("qor_25_cast_fp16")]; + tensor var_2656_cast_fp16 = mul(x = var_2618_cast_fp16, y = roti_13_cast_fp16)[name = tensor("op_2656_cast_fp16")]; + tensor var_2657_cast_fp16 = mul(x = var_2626_cast_fp16, y = rotr_13_cast_fp16)[name = tensor("op_2657_cast_fp16")]; + tensor qoi_25_cast_fp16 = add(x = var_2656_cast_fp16, y = var_2657_cast_fp16)[name = tensor("qoi_25_cast_fp16")]; + tensor var_2660_cast_fp16 = mul(x = var_2634_cast_fp16, y = rotr_13_cast_fp16)[name = tensor("op_2660_cast_fp16")]; + tensor var_2661_cast_fp16 = mul(x = var_2642_cast_fp16, y = roti_13_cast_fp16)[name = tensor("op_2661_cast_fp16")]; + tensor kor_25_cast_fp16 = sub(x = var_2660_cast_fp16, y = var_2661_cast_fp16)[name = tensor("kor_25_cast_fp16")]; + tensor var_2664_cast_fp16 = mul(x = var_2634_cast_fp16, y = roti_13_cast_fp16)[name = tensor("op_2664_cast_fp16")]; + tensor var_2665_cast_fp16 = mul(x = var_2642_cast_fp16, y = rotr_13_cast_fp16)[name = tensor("op_2665_cast_fp16")]; + tensor koi_25_cast_fp16 = add(x = var_2664_cast_fp16, y = var_2665_cast_fp16)[name = tensor("koi_25_cast_fp16")]; + tensor qo_13_axis_0 = const()[name = tensor("qo_13_axis_0"), val = tensor(-1)]; + tensor qo_13_cast_fp16 = stack(axis = qo_13_axis_0, values = (qor_25_cast_fp16, qoi_25_cast_fp16))[name = tensor("qo_13_cast_fp16")]; + tensor ko_13_axis_0 = const()[name = tensor("ko_13_axis_0"), val = tensor(-1)]; + tensor ko_13_cast_fp16 = stack(axis = ko_13_axis_0, values = (kor_25_cast_fp16, koi_25_cast_fp16))[name = tensor("ko_13_cast_fp16")]; + tensor var_2698 = const()[name = tensor("op_2698"), val = tensor([1, 1, 16, 64])]; + tensor q_39_cast_fp16 = reshape(shape = var_2698, x = qo_13_cast_fp16)[name = tensor("q_39_cast_fp16")]; + tensor var_2704 = const()[name = tensor("op_2704"), val = tensor([1, 1, 16, 64])]; + tensor k_27_cast_fp16 = reshape(shape = var_2704, x = ko_13_cast_fp16)[name = tensor("k_27_cast_fp16")]; + tensor _inversed_2729_y_0_to_fp16 = const()[name = tensor("_inversed_2729_y_0_to_fp16"), val = tensor(0x1p-9)]; + tensor _inversed_2729_cast_fp16 = mul(x = ts_41_cast_fp16, y = _inversed_2729_y_0_to_fp16)[name = tensor("_inversed_2729_cast_fp16")]; + tensor var_2730_cast_fp16 = floor(x = _inversed_2729_cast_fp16)[name = tensor("op_2730_cast_fp16")]; + tensor var_2731_to_fp16 = const()[name = tensor("op_2731_to_fp16"), val = tensor(0x1p+9)]; + tensor var_2732_cast_fp16 = mul(x = var_2730_cast_fp16, y = var_2731_to_fp16)[name = tensor("op_2732_cast_fp16")]; + tensor write_indices_float_27_cast_fp16 = sub(x = ts_41_cast_fp16, y = var_2732_cast_fp16)[name = tensor("write_indices_float_27_cast_fp16")]; + tensor var_2739_dtype_0 = const()[name = tensor("op_2739_dtype_0"), val = tensor("int32")]; + tensor write_indices_13_reps_0 = const()[name = tensor("write_indices_13_reps_0"), val = tensor([1, 1, 16, 64])]; + tensor write_indices_float_27_cast_fp16_to_int32 = cast(dtype = var_2739_dtype_0, x = write_indices_float_27_cast_fp16)[name = tensor("cast_522")]; + tensor write_indices_13 = tile(reps = write_indices_13_reps_0, x = write_indices_float_27_cast_fp16_to_int32)[name = tensor("write_indices_13")]; + tensor var_2749_begin_0 = const()[name = tensor("op_2749_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_2749_end_0 = const()[name = tensor("op_2749_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor var_2749_end_mask_0 = const()[name = tensor("op_2749_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_2749_squeeze_mask_0 = const()[name = tensor("op_2749_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor cache6_to_fp16_dtype_0 = const()[name = tensor("cache6_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor cache6_to_fp16 = cast(dtype = cache6_to_fp16_dtype_0, x = cache6)[name = tensor("cast_521")]; + tensor var_2749_cast_fp16 = slice_by_index(begin = var_2749_begin_0, end = var_2749_end_0, end_mask = var_2749_end_mask_0, squeeze_mask = var_2749_squeeze_mask_0, x = cache6_to_fp16)[name = tensor("op_2749_cast_fp16")]; + tensor var_2751_axis_0 = const()[name = tensor("op_2751_axis_0"), val = tensor(1)]; + tensor var_2751_mode_0 = const()[name = tensor("op_2751_mode_0"), val = tensor("update")]; + tensor var_2751_validate_indices_0 = const()[name = tensor("op_2751_validate_indices_0"), val = tensor(false)]; + tensor var_2751_cast_fp16 = scatter_along_axis(axis = var_2751_axis_0, data = var_2749_cast_fp16, indices = write_indices_13, mode = var_2751_mode_0, updates = k_27_cast_fp16, validate_indices = var_2751_validate_indices_0)[name = tensor("op_2751_cast_fp16")]; + tensor concat_44 = const()[name = tensor("concat_44"), val = tensor([0, 0, 0, 0, 0])]; + tensor concat_45 = const()[name = tensor("concat_45"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_13_internal_tensor_assign_1_stride_0 = const()[name = tensor("new_cache_13_internal_tensor_assign_1_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_13_internal_tensor_assign_1_begin_mask_0 = const()[name = tensor("new_cache_13_internal_tensor_assign_1_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_13_internal_tensor_assign_1_end_mask_0 = const()[name = tensor("new_cache_13_internal_tensor_assign_1_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_13_internal_tensor_assign_1_squeeze_mask_0 = const()[name = tensor("new_cache_13_internal_tensor_assign_1_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_60 = const()[name = tensor("shape_60"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_12 = const()[name = tensor("reduce_prod_12"), val = tensor(1048576)]; + tensor range_1d_12_start_0 = const()[name = tensor("range_1d_12_start_0"), val = tensor(0)]; + tensor range_1d_12_step_0 = const()[name = tensor("range_1d_12_step_0"), val = tensor(1)]; + tensor range_1d_12 = range_1d(end = reduce_prod_12, start = range_1d_12_start_0, step = range_1d_12_step_0)[name = tensor("range_1d_12")]; + tensor reshape_60 = reshape(shape = shape_60, x = range_1d_12)[name = tensor("reshape_60")]; + tensor slice_by_index_12 = slice_by_index(begin = concat_44, begin_mask = new_cache_13_internal_tensor_assign_1_begin_mask_0, end = concat_45, end_mask = new_cache_13_internal_tensor_assign_1_end_mask_0, squeeze_mask = new_cache_13_internal_tensor_assign_1_squeeze_mask_0, stride = new_cache_13_internal_tensor_assign_1_stride_0, x = reshape_60)[name = tensor("slice_by_index_12")]; + tensor reshape_61_shape_0 = const()[name = tensor("reshape_61_shape_0"), val = tensor([-1])]; + tensor reshape_61 = reshape(shape = reshape_61_shape_0, x = slice_by_index_12)[name = tensor("reshape_61")]; + tensor reshape_62_shape_0 = const()[name = tensor("reshape_62_shape_0"), val = tensor([-1])]; + tensor reshape_62_cast_fp16 = reshape(shape = reshape_62_shape_0, x = var_2751_cast_fp16)[name = tensor("reshape_62_cast_fp16")]; + tensor reshape_63_shape_0 = const()[name = tensor("reshape_63_shape_0"), val = tensor([-1])]; + tensor reshape_63_cast_fp16 = reshape(shape = reshape_63_shape_0, x = cache6_to_fp16)[name = tensor("reshape_63_cast_fp16")]; + tensor scatter_12_mode_0 = const()[name = tensor("scatter_12_mode_0"), val = tensor("update")]; + tensor scatter_12_axis_0 = const()[name = tensor("scatter_12_axis_0"), val = tensor(0)]; + tensor scatter_12_validate_indices_0 = const()[name = tensor("scatter_12_validate_indices_0"), val = tensor(false)]; + tensor scatter_12_cast_fp16 = scatter(axis = scatter_12_axis_0, data = reshape_63_cast_fp16, indices = reshape_61, mode = scatter_12_mode_0, updates = reshape_62_cast_fp16, validate_indices = scatter_12_validate_indices_0)[name = tensor("scatter_12_cast_fp16")]; + tensor reshape_64_cast_fp16 = reshape(shape = shape_60, x = scatter_12_cast_fp16)[name = tensor("reshape_64_cast_fp16")]; + tensor var_2759_begin_0 = const()[name = tensor("op_2759_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor var_2759_end_0 = const()[name = tensor("op_2759_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor var_2759_end_mask_0 = const()[name = tensor("op_2759_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_2759_squeeze_mask_0 = const()[name = tensor("op_2759_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor var_2759_cast_fp16 = slice_by_index(begin = var_2759_begin_0, end = var_2759_end_0, end_mask = var_2759_end_mask_0, squeeze_mask = var_2759_squeeze_mask_0, x = reshape_64_cast_fp16)[name = tensor("op_2759_cast_fp16")]; + tensor var_2761_axis_0 = const()[name = tensor("op_2761_axis_0"), val = tensor(1)]; + tensor var_2761_mode_0 = const()[name = tensor("op_2761_mode_0"), val = tensor("update")]; + tensor var_2761_validate_indices_0 = const()[name = tensor("op_2761_validate_indices_0"), val = tensor(false)]; + tensor var_2761_cast_fp16 = scatter_along_axis(axis = var_2761_axis_0, data = var_2759_cast_fp16, indices = write_indices_13, mode = var_2761_mode_0, updates = v_13_cast_fp16, validate_indices = var_2761_validate_indices_0)[name = tensor("op_2761_cast_fp16")]; + tensor concat_46 = const()[name = tensor("concat_46"), val = tensor([1, 0, 0, 0, 0])]; + tensor concat_47 = const()[name = tensor("concat_47"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_13_internal_tensor_assign_2_stride_0 = const()[name = tensor("new_cache_13_internal_tensor_assign_2_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_13_internal_tensor_assign_2_begin_mask_0 = const()[name = tensor("new_cache_13_internal_tensor_assign_2_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_13_internal_tensor_assign_2_end_mask_0 = const()[name = tensor("new_cache_13_internal_tensor_assign_2_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_13_internal_tensor_assign_2_squeeze_mask_0 = const()[name = tensor("new_cache_13_internal_tensor_assign_2_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_61 = const()[name = tensor("shape_61"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_13 = const()[name = tensor("reduce_prod_13"), val = tensor(1048576)]; + tensor range_1d_13_start_0 = const()[name = tensor("range_1d_13_start_0"), val = tensor(0)]; + tensor range_1d_13_step_0 = const()[name = tensor("range_1d_13_step_0"), val = tensor(1)]; + tensor range_1d_13 = range_1d(end = reduce_prod_13, start = range_1d_13_start_0, step = range_1d_13_step_0)[name = tensor("range_1d_13")]; + tensor reshape_65 = reshape(shape = shape_61, x = range_1d_13)[name = tensor("reshape_65")]; + tensor slice_by_index_13 = slice_by_index(begin = concat_46, begin_mask = new_cache_13_internal_tensor_assign_2_begin_mask_0, end = concat_47, end_mask = new_cache_13_internal_tensor_assign_2_end_mask_0, squeeze_mask = new_cache_13_internal_tensor_assign_2_squeeze_mask_0, stride = new_cache_13_internal_tensor_assign_2_stride_0, x = reshape_65)[name = tensor("slice_by_index_13")]; + tensor reshape_66_shape_0 = const()[name = tensor("reshape_66_shape_0"), val = tensor([-1])]; + tensor reshape_66 = reshape(shape = reshape_66_shape_0, x = slice_by_index_13)[name = tensor("reshape_66")]; + tensor reshape_67_shape_0 = const()[name = tensor("reshape_67_shape_0"), val = tensor([-1])]; + tensor reshape_67_cast_fp16 = reshape(shape = reshape_67_shape_0, x = var_2761_cast_fp16)[name = tensor("reshape_67_cast_fp16")]; + tensor reshape_68_shape_0 = const()[name = tensor("reshape_68_shape_0"), val = tensor([-1])]; + tensor reshape_68_cast_fp16 = reshape(shape = reshape_68_shape_0, x = reshape_64_cast_fp16)[name = tensor("reshape_68_cast_fp16")]; + tensor scatter_13_mode_0 = const()[name = tensor("scatter_13_mode_0"), val = tensor("update")]; + tensor scatter_13_axis_0 = const()[name = tensor("scatter_13_axis_0"), val = tensor(0)]; + tensor scatter_13_validate_indices_0 = const()[name = tensor("scatter_13_validate_indices_0"), val = tensor(false)]; + tensor scatter_13_cast_fp16 = scatter(axis = scatter_13_axis_0, data = reshape_68_cast_fp16, indices = reshape_66, mode = scatter_13_mode_0, updates = reshape_67_cast_fp16, validate_indices = scatter_13_validate_indices_0)[name = tensor("scatter_13_cast_fp16")]; + tensor reshape_69_cast_fp16 = reshape(shape = shape_61, x = scatter_13_cast_fp16)[name = tensor("reshape_69_cast_fp16")]; + tensor reshape_69_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("reshape_69_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor keys_37_begin_0 = const()[name = tensor("keys_37_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor keys_37_end_0 = const()[name = tensor("keys_37_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor keys_37_end_mask_0 = const()[name = tensor("keys_37_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor keys_37_squeeze_mask_0 = const()[name = tensor("keys_37_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor keys_37_cast_fp16 = slice_by_index(begin = keys_37_begin_0, end = keys_37_end_0, end_mask = keys_37_end_mask_0, squeeze_mask = keys_37_squeeze_mask_0, x = reshape_69_cast_fp16)[name = tensor("keys_37_cast_fp16")]; + tensor values_37_begin_0 = const()[name = tensor("values_37_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor values_37_end_0 = const()[name = tensor("values_37_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor values_37_end_mask_0 = const()[name = tensor("values_37_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor values_37_squeeze_mask_0 = const()[name = tensor("values_37_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor values_37_cast_fp16 = slice_by_index(begin = values_37_begin_0, end = values_37_end_0, end_mask = values_37_end_mask_0, squeeze_mask = values_37_squeeze_mask_0, x = reshape_69_cast_fp16)[name = tensor("values_37_cast_fp16")]; + tensor var_2773_cast_fp16 = not_equal(x = keys_37_cast_fp16, y = keys_37_cast_fp16)[name = tensor("op_2773_cast_fp16")]; + tensor keys_39_cast_fp16 = select(a = var_463_to_fp16, b = keys_37_cast_fp16, cond = var_2773_cast_fp16)[name = tensor("keys_39_cast_fp16")]; + tensor var_2781_cast_fp16 = not_equal(x = values_37_cast_fp16, y = values_37_cast_fp16)[name = tensor("op_2781_cast_fp16")]; + tensor values_39_cast_fp16 = select(a = var_463_to_fp16, b = values_37_cast_fp16, cond = var_2781_cast_fp16)[name = tensor("values_39_cast_fp16")]; + tensor var_2805 = const()[name = tensor("op_2805"), val = tensor([0, 2, 1, 3])]; + tensor var_2821 = const()[name = tensor("op_2821"), val = tensor([1, 1, 1])]; + tensor var_2822_cast_fp16 = reshape(shape = var_2821, x = position6_to_fp16)[name = tensor("op_2822_cast_fp16")]; + tensor var_2842_to_fp16 = const()[name = tensor("op_2842_to_fp16"), val = tensor(0x1p+0)]; + tensor valid_len_13_cast_fp16 = add(x = var_2822_cast_fp16, y = var_2842_to_fp16)[name = tensor("valid_len_13_cast_fp16")]; + tensor valid_mask_13_cast_fp16 = less(x = k_positions_1_promoted_to_fp16, y = valid_len_13_cast_fp16)[name = tensor("valid_mask_13_cast_fp16")]; + tensor causal_mask_13_cast_fp16 = less_equal(x = k_positions_1_promoted_to_fp16, y = var_2822_cast_fp16)[name = tensor("causal_mask_13_cast_fp16")]; + tensor attn_mask_25 = logical_and(x = valid_mask_13_cast_fp16, y = causal_mask_13_cast_fp16)[name = tensor("attn_mask_25")]; + tensor attn_mask_27_axes_0 = const()[name = tensor("attn_mask_27_axes_0"), val = tensor([1])]; + tensor attn_mask_27 = expand_dims(axes = attn_mask_27_axes_0, x = attn_mask_25)[name = tensor("attn_mask_27")]; + tensor var_2860_transpose_x_0 = const()[name = tensor("op_2860_transpose_x_0"), val = tensor(false)]; + tensor var_2860_transpose_y_0 = const()[name = tensor("op_2860_transpose_y_0"), val = tensor(false)]; + tensor transpose_84_perm_0 = const()[name = tensor("transpose_84_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_85_perm_0 = const()[name = tensor("transpose_85_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_85 = transpose(perm = transpose_85_perm_0, x = keys_39_cast_fp16)[name = tensor("transpose_189")]; + tensor transpose_84 = transpose(perm = transpose_84_perm_0, x = q_39_cast_fp16)[name = tensor("transpose_190")]; + tensor var_2860_cast_fp16 = matmul(transpose_x = var_2860_transpose_x_0, transpose_y = var_2860_transpose_y_0, x = transpose_84, y = transpose_85)[name = tensor("op_2860_cast_fp16")]; + tensor var_2854_to_fp16 = const()[name = tensor("op_2854_to_fp16"), val = tensor([0x1p-3])]; + tensor attn_weights_37_cast_fp16 = mul(x = var_2860_cast_fp16, y = var_2854_to_fp16)[name = tensor("attn_weights_37_cast_fp16")]; + tensor var_2862 = logical_not(x = attn_mask_27)[name = tensor("op_2862")]; + tensor var_2863_to_fp16 = const()[name = tensor("op_2863_to_fp16"), val = tensor(-inf)]; + tensor attn_weights_39_cast_fp16 = select(a = var_2863_to_fp16, b = attn_weights_37_cast_fp16, cond = var_2862)[name = tensor("attn_weights_39_cast_fp16")]; + tensor var_2865 = const()[name = tensor("op_2865"), val = tensor(-1)]; + tensor attn_weights_41_cast_fp16 = softmax(axis = var_2865, x = attn_weights_39_cast_fp16)[name = tensor("attn_weights_41_cast_fp16")]; + tensor attn_output_13_transpose_x_0 = const()[name = tensor("attn_output_13_transpose_x_0"), val = tensor(false)]; + tensor attn_output_13_transpose_y_0 = const()[name = tensor("attn_output_13_transpose_y_0"), val = tensor(false)]; + tensor values_41_cast_fp16 = transpose(perm = var_2805, x = values_39_cast_fp16)[name = tensor("transpose_191")]; + tensor attn_output_13_cast_fp16 = matmul(transpose_x = attn_output_13_transpose_x_0, transpose_y = attn_output_13_transpose_y_0, x = attn_weights_41_cast_fp16, y = values_41_cast_fp16)[name = tensor("attn_output_13_cast_fp16")]; + tensor var_2873 = const()[name = tensor("op_2873"), val = tensor([0, 2, 1, 3])]; + tensor var_2878 = const()[name = tensor("op_2878"), val = tensor([1, 1, 1024])]; + tensor var_2874_cast_fp16 = transpose(perm = var_2873, x = attn_output_13_cast_fp16)[name = tensor("transpose_188")]; + tensor input_79_cast_fp16 = reshape(shape = var_2878, x = var_2874_cast_fp16)[name = tensor("input_79_cast_fp16")]; + tensor attn6_out_proj_weight_to_fp16 = const()[name = tensor("attn6_out_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(158475776)))]; + tensor linear_26_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = attn6_out_proj_weight_to_fp16, x = input_79_cast_fp16)[name = tensor("linear_26_cast_fp16")]; + tensor var_2884_to_fp16 = const()[name = tensor("op_2884_to_fp16"), val = tensor(0x1p+0)]; + tensor var_2885_cast_fp16 = add(x = position6_to_fp16, y = var_2884_to_fp16)[name = tensor("op_2885_cast_fp16")]; + tensor var_2885_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_2885_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor input_81_cast_fp16 = add(x = input_75_cast_fp16, y = linear_26_cast_fp16)[name = tensor("input_81_cast_fp16")]; + tensor input_83_axes_0 = const()[name = tensor("input_83_axes_0"), val = tensor([-1])]; + tensor norm6_2_weight_to_fp16 = const()[name = tensor("norm6_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(160572992)))]; + tensor norm6_2_bias_to_fp16 = const()[name = tensor("norm6_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(160575104)))]; + tensor var_2889_to_fp16 = const()[name = tensor("op_2889_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_83_cast_fp16 = layer_norm(axes = input_83_axes_0, beta = norm6_2_bias_to_fp16, epsilon = var_2889_to_fp16, gamma = norm6_2_weight_to_fp16, x = input_81_cast_fp16)[name = tensor("input_83_cast_fp16")]; + tensor linear6_1_weight_to_fp16 = const()[name = tensor("linear6_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(160577216)))]; + tensor linear_27_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = linear6_1_weight_to_fp16, x = input_83_cast_fp16)[name = tensor("linear_27_cast_fp16")]; + tensor input_85_mode_0 = const()[name = tensor("input_85_mode_0"), val = tensor("EXACT")]; + tensor input_85_cast_fp16 = gelu(mode = input_85_mode_0, x = linear_27_cast_fp16)[name = tensor("input_85_cast_fp16")]; + tensor linear6_2_weight_to_fp16 = const()[name = tensor("linear6_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(168965888)))]; + tensor linear_28_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = linear6_2_weight_to_fp16, x = input_85_cast_fp16)[name = tensor("linear_28_cast_fp16")]; + tensor input_87_cast_fp16 = add(x = input_81_cast_fp16, y = linear_28_cast_fp16)[name = tensor("input_87_cast_fp16")]; + tensor input_89_axes_0 = const()[name = tensor("input_89_axes_0"), val = tensor([-1])]; + tensor norm7_1_weight_to_fp16 = const()[name = tensor("norm7_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(177354560)))]; + tensor norm7_1_bias_to_fp16 = const()[name = tensor("norm7_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(177356672)))]; + tensor var_2906_to_fp16 = const()[name = tensor("op_2906_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_89_cast_fp16 = layer_norm(axes = input_89_axes_0, beta = norm7_1_bias_to_fp16, epsilon = var_2906_to_fp16, gamma = norm7_1_weight_to_fp16, x = input_87_cast_fp16)[name = tensor("input_89_cast_fp16")]; + tensor attn7_in_proj_weight_to_fp16 = const()[name = tensor("attn7_in_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(177358784)))]; + tensor linear_29_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = attn7_in_proj_weight_to_fp16, x = input_89_cast_fp16)[name = tensor("linear_29_cast_fp16")]; + tensor var_2920 = const()[name = tensor("op_2920"), val = tensor([1, 1, 3, 16, 64])]; + tensor qkv_15_cast_fp16 = reshape(shape = var_2920, x = linear_29_cast_fp16)[name = tensor("qkv_15_cast_fp16")]; + tensor q_43_begin_0 = const()[name = tensor("q_43_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor q_43_end_0 = const()[name = tensor("q_43_end_0"), val = tensor([1, 1, 1, 16, 64])]; + tensor q_43_end_mask_0 = const()[name = tensor("q_43_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor q_43_squeeze_mask_0 = const()[name = tensor("q_43_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor q_43_cast_fp16 = slice_by_index(begin = q_43_begin_0, end = q_43_end_0, end_mask = q_43_end_mask_0, squeeze_mask = q_43_squeeze_mask_0, x = qkv_15_cast_fp16)[name = tensor("q_43_cast_fp16")]; + tensor k_29_begin_0 = const()[name = tensor("k_29_begin_0"), val = tensor([0, 0, 1, 0, 0])]; + tensor k_29_end_0 = const()[name = tensor("k_29_end_0"), val = tensor([1, 1, 2, 16, 64])]; + tensor k_29_end_mask_0 = const()[name = tensor("k_29_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor k_29_squeeze_mask_0 = const()[name = tensor("k_29_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor k_29_cast_fp16 = slice_by_index(begin = k_29_begin_0, end = k_29_end_0, end_mask = k_29_end_mask_0, squeeze_mask = k_29_squeeze_mask_0, x = qkv_15_cast_fp16)[name = tensor("k_29_cast_fp16")]; + tensor v_15_begin_0 = const()[name = tensor("v_15_begin_0"), val = tensor([0, 0, 2, 0, 0])]; + tensor v_15_end_0 = const()[name = tensor("v_15_end_0"), val = tensor([1, 1, 3, 16, 64])]; + tensor v_15_end_mask_0 = const()[name = tensor("v_15_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor v_15_squeeze_mask_0 = const()[name = tensor("v_15_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor v_15_cast_fp16 = slice_by_index(begin = v_15_begin_0, end = v_15_end_0, end_mask = v_15_end_mask_0, squeeze_mask = v_15_squeeze_mask_0, x = qkv_15_cast_fp16)[name = tensor("v_15_cast_fp16")]; + tensor var_2986 = const()[name = tensor("op_2986"), val = tensor([1, 1, 1, 1])]; + tensor position7_to_fp16_dtype_0 = const()[name = tensor("position7_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor position7_to_fp16 = cast(dtype = position7_to_fp16_dtype_0, x = position7)[name = tensor("cast_518")]; + tensor ts_47_cast_fp16 = reshape(shape = var_2986, x = position7_to_fp16)[name = tensor("ts_47_cast_fp16")]; + tensor var_2993 = const()[name = tensor("op_2993"), val = tensor([1, 1, 16, 32, 2])]; + tensor q_complex_15_cast_fp16 = reshape(shape = var_2993, x = q_43_cast_fp16)[name = tensor("q_complex_15_cast_fp16")]; + tensor var_3000 = const()[name = tensor("op_3000"), val = tensor([1, 1, 16, 32, 2])]; + tensor k_complex_15_cast_fp16 = reshape(shape = var_3000, x = k_29_cast_fp16)[name = tensor("k_complex_15_cast_fp16")]; + tensor var_3004_begin_0 = const()[name = tensor("op_3004_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_3004_end_0 = const()[name = tensor("op_3004_end_0"), val = tensor([1, 1, 16, 32, 1])]; + tensor var_3004_end_mask_0 = const()[name = tensor("op_3004_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_3004_squeeze_mask_0 = const()[name = tensor("op_3004_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_3004_cast_fp16 = slice_by_index(begin = var_3004_begin_0, end = var_3004_end_0, end_mask = var_3004_end_mask_0, squeeze_mask = var_3004_squeeze_mask_0, x = q_complex_15_cast_fp16)[name = tensor("op_3004_cast_fp16")]; + tensor var_3012_begin_0 = const()[name = tensor("op_3012_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_3012_end_0 = const()[name = tensor("op_3012_end_0"), val = tensor([1, 1, 16, 32, 2])]; + tensor var_3012_end_mask_0 = const()[name = tensor("op_3012_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_3012_squeeze_mask_0 = const()[name = tensor("op_3012_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_3012_cast_fp16 = slice_by_index(begin = var_3012_begin_0, end = var_3012_end_0, end_mask = var_3012_end_mask_0, squeeze_mask = var_3012_squeeze_mask_0, x = q_complex_15_cast_fp16)[name = tensor("op_3012_cast_fp16")]; + tensor var_3020_begin_0 = const()[name = tensor("op_3020_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_3020_end_0 = const()[name = tensor("op_3020_end_0"), val = tensor([1, 1, 16, 32, 1])]; + tensor var_3020_end_mask_0 = const()[name = tensor("op_3020_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_3020_squeeze_mask_0 = const()[name = tensor("op_3020_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_3020_cast_fp16 = slice_by_index(begin = var_3020_begin_0, end = var_3020_end_0, end_mask = var_3020_end_mask_0, squeeze_mask = var_3020_squeeze_mask_0, x = k_complex_15_cast_fp16)[name = tensor("op_3020_cast_fp16")]; + tensor var_3028_begin_0 = const()[name = tensor("op_3028_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_3028_end_0 = const()[name = tensor("op_3028_end_0"), val = tensor([1, 1, 16, 32, 2])]; + tensor var_3028_end_mask_0 = const()[name = tensor("op_3028_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_3028_squeeze_mask_0 = const()[name = tensor("op_3028_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_3028_cast_fp16 = slice_by_index(begin = var_3028_begin_0, end = var_3028_end_0, end_mask = var_3028_end_mask_0, squeeze_mask = var_3028_squeeze_mask_0, x = k_complex_15_cast_fp16)[name = tensor("op_3028_cast_fp16")]; + tensor freqs_15_to_fp16 = const()[name = tensor("freqs_15_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(183650304)))]; + tensor var_3034_cast_fp16 = mul(x = freqs_15_to_fp16, y = ts_47_cast_fp16)[name = tensor("op_3034_cast_fp16")]; + tensor rotr_15_cast_fp16 = cos(x = var_3034_cast_fp16)[name = tensor("rotr_15_cast_fp16")]; + tensor roti_15_cast_fp16 = sin(x = var_3034_cast_fp16)[name = tensor("roti_15_cast_fp16")]; + tensor var_3038_cast_fp16 = mul(x = var_3004_cast_fp16, y = rotr_15_cast_fp16)[name = tensor("op_3038_cast_fp16")]; + tensor var_3039_cast_fp16 = mul(x = var_3012_cast_fp16, y = roti_15_cast_fp16)[name = tensor("op_3039_cast_fp16")]; + tensor qor_29_cast_fp16 = sub(x = var_3038_cast_fp16, y = var_3039_cast_fp16)[name = tensor("qor_29_cast_fp16")]; + tensor var_3042_cast_fp16 = mul(x = var_3004_cast_fp16, y = roti_15_cast_fp16)[name = tensor("op_3042_cast_fp16")]; + tensor var_3043_cast_fp16 = mul(x = var_3012_cast_fp16, y = rotr_15_cast_fp16)[name = tensor("op_3043_cast_fp16")]; + tensor qoi_29_cast_fp16 = add(x = var_3042_cast_fp16, y = var_3043_cast_fp16)[name = tensor("qoi_29_cast_fp16")]; + tensor var_3046_cast_fp16 = mul(x = var_3020_cast_fp16, y = rotr_15_cast_fp16)[name = tensor("op_3046_cast_fp16")]; + tensor var_3047_cast_fp16 = mul(x = var_3028_cast_fp16, y = roti_15_cast_fp16)[name = tensor("op_3047_cast_fp16")]; + tensor kor_29_cast_fp16 = sub(x = var_3046_cast_fp16, y = var_3047_cast_fp16)[name = tensor("kor_29_cast_fp16")]; + tensor var_3050_cast_fp16 = mul(x = var_3020_cast_fp16, y = roti_15_cast_fp16)[name = tensor("op_3050_cast_fp16")]; + tensor var_3051_cast_fp16 = mul(x = var_3028_cast_fp16, y = rotr_15_cast_fp16)[name = tensor("op_3051_cast_fp16")]; + tensor koi_29_cast_fp16 = add(x = var_3050_cast_fp16, y = var_3051_cast_fp16)[name = tensor("koi_29_cast_fp16")]; + tensor qo_15_axis_0 = const()[name = tensor("qo_15_axis_0"), val = tensor(-1)]; + tensor qo_15_cast_fp16 = stack(axis = qo_15_axis_0, values = (qor_29_cast_fp16, qoi_29_cast_fp16))[name = tensor("qo_15_cast_fp16")]; + tensor ko_15_axis_0 = const()[name = tensor("ko_15_axis_0"), val = tensor(-1)]; + tensor ko_15_cast_fp16 = stack(axis = ko_15_axis_0, values = (kor_29_cast_fp16, koi_29_cast_fp16))[name = tensor("ko_15_cast_fp16")]; + tensor var_3084 = const()[name = tensor("op_3084"), val = tensor([1, 1, 16, 64])]; + tensor q_45_cast_fp16 = reshape(shape = var_3084, x = qo_15_cast_fp16)[name = tensor("q_45_cast_fp16")]; + tensor var_3090 = const()[name = tensor("op_3090"), val = tensor([1, 1, 16, 64])]; + tensor k_31_cast_fp16 = reshape(shape = var_3090, x = ko_15_cast_fp16)[name = tensor("k_31_cast_fp16")]; + tensor _inversed_3115_y_0_to_fp16 = const()[name = tensor("_inversed_3115_y_0_to_fp16"), val = tensor(0x1p-9)]; + tensor _inversed_3115_cast_fp16 = mul(x = ts_47_cast_fp16, y = _inversed_3115_y_0_to_fp16)[name = tensor("_inversed_3115_cast_fp16")]; + tensor var_3116_cast_fp16 = floor(x = _inversed_3115_cast_fp16)[name = tensor("op_3116_cast_fp16")]; + tensor var_3117_to_fp16 = const()[name = tensor("op_3117_to_fp16"), val = tensor(0x1p+9)]; + tensor var_3118_cast_fp16 = mul(x = var_3116_cast_fp16, y = var_3117_to_fp16)[name = tensor("op_3118_cast_fp16")]; + tensor write_indices_float_31_cast_fp16 = sub(x = ts_47_cast_fp16, y = var_3118_cast_fp16)[name = tensor("write_indices_float_31_cast_fp16")]; + tensor var_3125_dtype_0 = const()[name = tensor("op_3125_dtype_0"), val = tensor("int32")]; + tensor write_indices_15_reps_0 = const()[name = tensor("write_indices_15_reps_0"), val = tensor([1, 1, 16, 64])]; + tensor write_indices_float_31_cast_fp16_to_int32 = cast(dtype = var_3125_dtype_0, x = write_indices_float_31_cast_fp16)[name = tensor("cast_517")]; + tensor write_indices_15 = tile(reps = write_indices_15_reps_0, x = write_indices_float_31_cast_fp16_to_int32)[name = tensor("write_indices_15")]; + tensor var_3135_begin_0 = const()[name = tensor("op_3135_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_3135_end_0 = const()[name = tensor("op_3135_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor var_3135_end_mask_0 = const()[name = tensor("op_3135_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_3135_squeeze_mask_0 = const()[name = tensor("op_3135_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor cache7_to_fp16_dtype_0 = const()[name = tensor("cache7_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor cache7_to_fp16 = cast(dtype = cache7_to_fp16_dtype_0, x = cache7)[name = tensor("cast_516")]; + tensor var_3135_cast_fp16 = slice_by_index(begin = var_3135_begin_0, end = var_3135_end_0, end_mask = var_3135_end_mask_0, squeeze_mask = var_3135_squeeze_mask_0, x = cache7_to_fp16)[name = tensor("op_3135_cast_fp16")]; + tensor var_3137_axis_0 = const()[name = tensor("op_3137_axis_0"), val = tensor(1)]; + tensor var_3137_mode_0 = const()[name = tensor("op_3137_mode_0"), val = tensor("update")]; + tensor var_3137_validate_indices_0 = const()[name = tensor("op_3137_validate_indices_0"), val = tensor(false)]; + tensor var_3137_cast_fp16 = scatter_along_axis(axis = var_3137_axis_0, data = var_3135_cast_fp16, indices = write_indices_15, mode = var_3137_mode_0, updates = k_31_cast_fp16, validate_indices = var_3137_validate_indices_0)[name = tensor("op_3137_cast_fp16")]; + tensor concat_51 = const()[name = tensor("concat_51"), val = tensor([0, 0, 0, 0, 0])]; + tensor concat_52 = const()[name = tensor("concat_52"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_15_internal_tensor_assign_1_stride_0 = const()[name = tensor("new_cache_15_internal_tensor_assign_1_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_15_internal_tensor_assign_1_begin_mask_0 = const()[name = tensor("new_cache_15_internal_tensor_assign_1_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_15_internal_tensor_assign_1_end_mask_0 = const()[name = tensor("new_cache_15_internal_tensor_assign_1_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_15_internal_tensor_assign_1_squeeze_mask_0 = const()[name = tensor("new_cache_15_internal_tensor_assign_1_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_62 = const()[name = tensor("shape_62"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_14 = const()[name = tensor("reduce_prod_14"), val = tensor(1048576)]; + tensor range_1d_14_start_0 = const()[name = tensor("range_1d_14_start_0"), val = tensor(0)]; + tensor range_1d_14_step_0 = const()[name = tensor("range_1d_14_step_0"), val = tensor(1)]; + tensor range_1d_14 = range_1d(end = reduce_prod_14, start = range_1d_14_start_0, step = range_1d_14_step_0)[name = tensor("range_1d_14")]; + tensor reshape_70 = reshape(shape = shape_62, x = range_1d_14)[name = tensor("reshape_70")]; + tensor slice_by_index_14 = slice_by_index(begin = concat_51, begin_mask = new_cache_15_internal_tensor_assign_1_begin_mask_0, end = concat_52, end_mask = new_cache_15_internal_tensor_assign_1_end_mask_0, squeeze_mask = new_cache_15_internal_tensor_assign_1_squeeze_mask_0, stride = new_cache_15_internal_tensor_assign_1_stride_0, x = reshape_70)[name = tensor("slice_by_index_14")]; + tensor reshape_71_shape_0 = const()[name = tensor("reshape_71_shape_0"), val = tensor([-1])]; + tensor reshape_71 = reshape(shape = reshape_71_shape_0, x = slice_by_index_14)[name = tensor("reshape_71")]; + tensor reshape_72_shape_0 = const()[name = tensor("reshape_72_shape_0"), val = tensor([-1])]; + tensor reshape_72_cast_fp16 = reshape(shape = reshape_72_shape_0, x = var_3137_cast_fp16)[name = tensor("reshape_72_cast_fp16")]; + tensor reshape_73_shape_0 = const()[name = tensor("reshape_73_shape_0"), val = tensor([-1])]; + tensor reshape_73_cast_fp16 = reshape(shape = reshape_73_shape_0, x = cache7_to_fp16)[name = tensor("reshape_73_cast_fp16")]; + tensor scatter_14_mode_0 = const()[name = tensor("scatter_14_mode_0"), val = tensor("update")]; + tensor scatter_14_axis_0 = const()[name = tensor("scatter_14_axis_0"), val = tensor(0)]; + tensor scatter_14_validate_indices_0 = const()[name = tensor("scatter_14_validate_indices_0"), val = tensor(false)]; + tensor scatter_14_cast_fp16 = scatter(axis = scatter_14_axis_0, data = reshape_73_cast_fp16, indices = reshape_71, mode = scatter_14_mode_0, updates = reshape_72_cast_fp16, validate_indices = scatter_14_validate_indices_0)[name = tensor("scatter_14_cast_fp16")]; + tensor reshape_74_cast_fp16 = reshape(shape = shape_62, x = scatter_14_cast_fp16)[name = tensor("reshape_74_cast_fp16")]; + tensor var_3145_begin_0 = const()[name = tensor("op_3145_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor var_3145_end_0 = const()[name = tensor("op_3145_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor var_3145_end_mask_0 = const()[name = tensor("op_3145_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_3145_squeeze_mask_0 = const()[name = tensor("op_3145_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor var_3145_cast_fp16 = slice_by_index(begin = var_3145_begin_0, end = var_3145_end_0, end_mask = var_3145_end_mask_0, squeeze_mask = var_3145_squeeze_mask_0, x = reshape_74_cast_fp16)[name = tensor("op_3145_cast_fp16")]; + tensor var_3147_axis_0 = const()[name = tensor("op_3147_axis_0"), val = tensor(1)]; + tensor var_3147_mode_0 = const()[name = tensor("op_3147_mode_0"), val = tensor("update")]; + tensor var_3147_validate_indices_0 = const()[name = tensor("op_3147_validate_indices_0"), val = tensor(false)]; + tensor var_3147_cast_fp16 = scatter_along_axis(axis = var_3147_axis_0, data = var_3145_cast_fp16, indices = write_indices_15, mode = var_3147_mode_0, updates = v_15_cast_fp16, validate_indices = var_3147_validate_indices_0)[name = tensor("op_3147_cast_fp16")]; + tensor concat_53 = const()[name = tensor("concat_53"), val = tensor([1, 0, 0, 0, 0])]; + tensor concat_54 = const()[name = tensor("concat_54"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_15_internal_tensor_assign_2_stride_0 = const()[name = tensor("new_cache_15_internal_tensor_assign_2_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_15_internal_tensor_assign_2_begin_mask_0 = const()[name = tensor("new_cache_15_internal_tensor_assign_2_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_15_internal_tensor_assign_2_end_mask_0 = const()[name = tensor("new_cache_15_internal_tensor_assign_2_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_15_internal_tensor_assign_2_squeeze_mask_0 = const()[name = tensor("new_cache_15_internal_tensor_assign_2_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_63 = const()[name = tensor("shape_63"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_15 = const()[name = tensor("reduce_prod_15"), val = tensor(1048576)]; + tensor range_1d_15_start_0 = const()[name = tensor("range_1d_15_start_0"), val = tensor(0)]; + tensor range_1d_15_step_0 = const()[name = tensor("range_1d_15_step_0"), val = tensor(1)]; + tensor range_1d_15 = range_1d(end = reduce_prod_15, start = range_1d_15_start_0, step = range_1d_15_step_0)[name = tensor("range_1d_15")]; + tensor reshape_75 = reshape(shape = shape_63, x = range_1d_15)[name = tensor("reshape_75")]; + tensor slice_by_index_15 = slice_by_index(begin = concat_53, begin_mask = new_cache_15_internal_tensor_assign_2_begin_mask_0, end = concat_54, end_mask = new_cache_15_internal_tensor_assign_2_end_mask_0, squeeze_mask = new_cache_15_internal_tensor_assign_2_squeeze_mask_0, stride = new_cache_15_internal_tensor_assign_2_stride_0, x = reshape_75)[name = tensor("slice_by_index_15")]; + tensor reshape_76_shape_0 = const()[name = tensor("reshape_76_shape_0"), val = tensor([-1])]; + tensor reshape_76 = reshape(shape = reshape_76_shape_0, x = slice_by_index_15)[name = tensor("reshape_76")]; + tensor reshape_77_shape_0 = const()[name = tensor("reshape_77_shape_0"), val = tensor([-1])]; + tensor reshape_77_cast_fp16 = reshape(shape = reshape_77_shape_0, x = var_3147_cast_fp16)[name = tensor("reshape_77_cast_fp16")]; + tensor reshape_78_shape_0 = const()[name = tensor("reshape_78_shape_0"), val = tensor([-1])]; + tensor reshape_78_cast_fp16 = reshape(shape = reshape_78_shape_0, x = reshape_74_cast_fp16)[name = tensor("reshape_78_cast_fp16")]; + tensor scatter_15_mode_0 = const()[name = tensor("scatter_15_mode_0"), val = tensor("update")]; + tensor scatter_15_axis_0 = const()[name = tensor("scatter_15_axis_0"), val = tensor(0)]; + tensor scatter_15_validate_indices_0 = const()[name = tensor("scatter_15_validate_indices_0"), val = tensor(false)]; + tensor scatter_15_cast_fp16 = scatter(axis = scatter_15_axis_0, data = reshape_78_cast_fp16, indices = reshape_76, mode = scatter_15_mode_0, updates = reshape_77_cast_fp16, validate_indices = scatter_15_validate_indices_0)[name = tensor("scatter_15_cast_fp16")]; + tensor reshape_79_cast_fp16 = reshape(shape = shape_63, x = scatter_15_cast_fp16)[name = tensor("reshape_79_cast_fp16")]; + tensor reshape_79_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("reshape_79_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor keys_43_begin_0 = const()[name = tensor("keys_43_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor keys_43_end_0 = const()[name = tensor("keys_43_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor keys_43_end_mask_0 = const()[name = tensor("keys_43_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor keys_43_squeeze_mask_0 = const()[name = tensor("keys_43_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor keys_43_cast_fp16 = slice_by_index(begin = keys_43_begin_0, end = keys_43_end_0, end_mask = keys_43_end_mask_0, squeeze_mask = keys_43_squeeze_mask_0, x = reshape_79_cast_fp16)[name = tensor("keys_43_cast_fp16")]; + tensor values_43_begin_0 = const()[name = tensor("values_43_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor values_43_end_0 = const()[name = tensor("values_43_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor values_43_end_mask_0 = const()[name = tensor("values_43_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor values_43_squeeze_mask_0 = const()[name = tensor("values_43_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor values_43_cast_fp16 = slice_by_index(begin = values_43_begin_0, end = values_43_end_0, end_mask = values_43_end_mask_0, squeeze_mask = values_43_squeeze_mask_0, x = reshape_79_cast_fp16)[name = tensor("values_43_cast_fp16")]; + tensor var_3159_cast_fp16 = not_equal(x = keys_43_cast_fp16, y = keys_43_cast_fp16)[name = tensor("op_3159_cast_fp16")]; + tensor keys_45_cast_fp16 = select(a = var_463_to_fp16, b = keys_43_cast_fp16, cond = var_3159_cast_fp16)[name = tensor("keys_45_cast_fp16")]; + tensor var_3167_cast_fp16 = not_equal(x = values_43_cast_fp16, y = values_43_cast_fp16)[name = tensor("op_3167_cast_fp16")]; + tensor values_45_cast_fp16 = select(a = var_463_to_fp16, b = values_43_cast_fp16, cond = var_3167_cast_fp16)[name = tensor("values_45_cast_fp16")]; + tensor var_3191 = const()[name = tensor("op_3191"), val = tensor([0, 2, 1, 3])]; + tensor var_3207 = const()[name = tensor("op_3207"), val = tensor([1, 1, 1])]; + tensor var_3208_cast_fp16 = reshape(shape = var_3207, x = position7_to_fp16)[name = tensor("op_3208_cast_fp16")]; + tensor var_3228_to_fp16 = const()[name = tensor("op_3228_to_fp16"), val = tensor(0x1p+0)]; + tensor valid_len_15_cast_fp16 = add(x = var_3208_cast_fp16, y = var_3228_to_fp16)[name = tensor("valid_len_15_cast_fp16")]; + tensor valid_mask_15_cast_fp16 = less(x = k_positions_1_promoted_to_fp16, y = valid_len_15_cast_fp16)[name = tensor("valid_mask_15_cast_fp16")]; + tensor causal_mask_15_cast_fp16 = less_equal(x = k_positions_1_promoted_to_fp16, y = var_3208_cast_fp16)[name = tensor("causal_mask_15_cast_fp16")]; + tensor attn_mask_29 = logical_and(x = valid_mask_15_cast_fp16, y = causal_mask_15_cast_fp16)[name = tensor("attn_mask_29")]; + tensor attn_mask_31_axes_0 = const()[name = tensor("attn_mask_31_axes_0"), val = tensor([1])]; + tensor attn_mask_31 = expand_dims(axes = attn_mask_31_axes_0, x = attn_mask_29)[name = tensor("attn_mask_31")]; + tensor var_3246_transpose_x_0 = const()[name = tensor("op_3246_transpose_x_0"), val = tensor(false)]; + tensor var_3246_transpose_y_0 = const()[name = tensor("op_3246_transpose_y_0"), val = tensor(false)]; + tensor transpose_86_perm_0 = const()[name = tensor("transpose_86_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_87_perm_0 = const()[name = tensor("transpose_87_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_87 = transpose(perm = transpose_87_perm_0, x = keys_45_cast_fp16)[name = tensor("transpose_185")]; + tensor transpose_86 = transpose(perm = transpose_86_perm_0, x = q_45_cast_fp16)[name = tensor("transpose_186")]; + tensor var_3246_cast_fp16 = matmul(transpose_x = var_3246_transpose_x_0, transpose_y = var_3246_transpose_y_0, x = transpose_86, y = transpose_87)[name = tensor("op_3246_cast_fp16")]; + tensor var_3240_to_fp16 = const()[name = tensor("op_3240_to_fp16"), val = tensor([0x1p-3])]; + tensor attn_weights_43_cast_fp16 = mul(x = var_3246_cast_fp16, y = var_3240_to_fp16)[name = tensor("attn_weights_43_cast_fp16")]; + tensor var_3248 = logical_not(x = attn_mask_31)[name = tensor("op_3248")]; + tensor var_3249_to_fp16 = const()[name = tensor("op_3249_to_fp16"), val = tensor(-inf)]; + tensor attn_weights_45_cast_fp16 = select(a = var_3249_to_fp16, b = attn_weights_43_cast_fp16, cond = var_3248)[name = tensor("attn_weights_45_cast_fp16")]; + tensor var_3251 = const()[name = tensor("op_3251"), val = tensor(-1)]; + tensor attn_weights_47_cast_fp16 = softmax(axis = var_3251, x = attn_weights_45_cast_fp16)[name = tensor("attn_weights_47_cast_fp16")]; + tensor attn_output_15_transpose_x_0 = const()[name = tensor("attn_output_15_transpose_x_0"), val = tensor(false)]; + tensor attn_output_15_transpose_y_0 = const()[name = tensor("attn_output_15_transpose_y_0"), val = tensor(false)]; + tensor values_47_cast_fp16 = transpose(perm = var_3191, x = values_45_cast_fp16)[name = tensor("transpose_187")]; + tensor attn_output_15_cast_fp16 = matmul(transpose_x = attn_output_15_transpose_x_0, transpose_y = attn_output_15_transpose_y_0, x = attn_weights_47_cast_fp16, y = values_47_cast_fp16)[name = tensor("attn_output_15_cast_fp16")]; + tensor var_3259 = const()[name = tensor("op_3259"), val = tensor([0, 2, 1, 3])]; + tensor var_3264 = const()[name = tensor("op_3264"), val = tensor([1, 1, 1024])]; + tensor var_3260_cast_fp16 = transpose(perm = var_3259, x = attn_output_15_cast_fp16)[name = tensor("transpose_184")]; + tensor input_91_cast_fp16 = reshape(shape = var_3264, x = var_3260_cast_fp16)[name = tensor("input_91_cast_fp16")]; + tensor attn7_out_proj_weight_to_fp16 = const()[name = tensor("attn7_out_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(183650432)))]; + tensor linear_30_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = attn7_out_proj_weight_to_fp16, x = input_91_cast_fp16)[name = tensor("linear_30_cast_fp16")]; + tensor var_3270_to_fp16 = const()[name = tensor("op_3270_to_fp16"), val = tensor(0x1p+0)]; + tensor var_3271_cast_fp16 = add(x = position7_to_fp16, y = var_3270_to_fp16)[name = tensor("op_3271_cast_fp16")]; + tensor var_3271_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_3271_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor input_93_cast_fp16 = add(x = input_87_cast_fp16, y = linear_30_cast_fp16)[name = tensor("input_93_cast_fp16")]; + tensor input_95_axes_0 = const()[name = tensor("input_95_axes_0"), val = tensor([-1])]; + tensor norm7_2_weight_to_fp16 = const()[name = tensor("norm7_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(185747648)))]; + tensor norm7_2_bias_to_fp16 = const()[name = tensor("norm7_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(185749760)))]; + tensor var_3275_to_fp16 = const()[name = tensor("op_3275_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_95_cast_fp16 = layer_norm(axes = input_95_axes_0, beta = norm7_2_bias_to_fp16, epsilon = var_3275_to_fp16, gamma = norm7_2_weight_to_fp16, x = input_93_cast_fp16)[name = tensor("input_95_cast_fp16")]; + tensor linear7_1_weight_to_fp16 = const()[name = tensor("linear7_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(185751872)))]; + tensor linear_31_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = linear7_1_weight_to_fp16, x = input_95_cast_fp16)[name = tensor("linear_31_cast_fp16")]; + tensor input_97_mode_0 = const()[name = tensor("input_97_mode_0"), val = tensor("EXACT")]; + tensor input_97_cast_fp16 = gelu(mode = input_97_mode_0, x = linear_31_cast_fp16)[name = tensor("input_97_cast_fp16")]; + tensor linear7_2_weight_to_fp16 = const()[name = tensor("linear7_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(194140544)))]; + tensor linear_32_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = linear7_2_weight_to_fp16, x = input_97_cast_fp16)[name = tensor("linear_32_cast_fp16")]; + tensor input_99_cast_fp16 = add(x = input_93_cast_fp16, y = linear_32_cast_fp16)[name = tensor("input_99_cast_fp16")]; + tensor input_101_axes_0 = const()[name = tensor("input_101_axes_0"), val = tensor([-1])]; + tensor norm8_1_weight_to_fp16 = const()[name = tensor("norm8_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(202529216)))]; + tensor norm8_1_bias_to_fp16 = const()[name = tensor("norm8_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(202531328)))]; + tensor var_3292_to_fp16 = const()[name = tensor("op_3292_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_101_cast_fp16 = layer_norm(axes = input_101_axes_0, beta = norm8_1_bias_to_fp16, epsilon = var_3292_to_fp16, gamma = norm8_1_weight_to_fp16, x = input_99_cast_fp16)[name = tensor("input_101_cast_fp16")]; + tensor attn8_in_proj_weight_to_fp16 = const()[name = tensor("attn8_in_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(202533440)))]; + tensor linear_33_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = attn8_in_proj_weight_to_fp16, x = input_101_cast_fp16)[name = tensor("linear_33_cast_fp16")]; + tensor var_3306 = const()[name = tensor("op_3306"), val = tensor([1, 1, 3, 16, 64])]; + tensor qkv_17_cast_fp16 = reshape(shape = var_3306, x = linear_33_cast_fp16)[name = tensor("qkv_17_cast_fp16")]; + tensor q_49_begin_0 = const()[name = tensor("q_49_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor q_49_end_0 = const()[name = tensor("q_49_end_0"), val = tensor([1, 1, 1, 16, 64])]; + tensor q_49_end_mask_0 = const()[name = tensor("q_49_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor q_49_squeeze_mask_0 = const()[name = tensor("q_49_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor q_49_cast_fp16 = slice_by_index(begin = q_49_begin_0, end = q_49_end_0, end_mask = q_49_end_mask_0, squeeze_mask = q_49_squeeze_mask_0, x = qkv_17_cast_fp16)[name = tensor("q_49_cast_fp16")]; + tensor k_33_begin_0 = const()[name = tensor("k_33_begin_0"), val = tensor([0, 0, 1, 0, 0])]; + tensor k_33_end_0 = const()[name = tensor("k_33_end_0"), val = tensor([1, 1, 2, 16, 64])]; + tensor k_33_end_mask_0 = const()[name = tensor("k_33_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor k_33_squeeze_mask_0 = const()[name = tensor("k_33_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor k_33_cast_fp16 = slice_by_index(begin = k_33_begin_0, end = k_33_end_0, end_mask = k_33_end_mask_0, squeeze_mask = k_33_squeeze_mask_0, x = qkv_17_cast_fp16)[name = tensor("k_33_cast_fp16")]; + tensor v_17_begin_0 = const()[name = tensor("v_17_begin_0"), val = tensor([0, 0, 2, 0, 0])]; + tensor v_17_end_0 = const()[name = tensor("v_17_end_0"), val = tensor([1, 1, 3, 16, 64])]; + tensor v_17_end_mask_0 = const()[name = tensor("v_17_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor v_17_squeeze_mask_0 = const()[name = tensor("v_17_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor v_17_cast_fp16 = slice_by_index(begin = v_17_begin_0, end = v_17_end_0, end_mask = v_17_end_mask_0, squeeze_mask = v_17_squeeze_mask_0, x = qkv_17_cast_fp16)[name = tensor("v_17_cast_fp16")]; + tensor var_3372 = const()[name = tensor("op_3372"), val = tensor([1, 1, 1, 1])]; + tensor position8_to_fp16_dtype_0 = const()[name = tensor("position8_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor position8_to_fp16 = cast(dtype = position8_to_fp16_dtype_0, x = position8)[name = tensor("cast_513")]; + tensor ts_53_cast_fp16 = reshape(shape = var_3372, x = position8_to_fp16)[name = tensor("ts_53_cast_fp16")]; + tensor var_3379 = const()[name = tensor("op_3379"), val = tensor([1, 1, 16, 32, 2])]; + tensor q_complex_17_cast_fp16 = reshape(shape = var_3379, x = q_49_cast_fp16)[name = tensor("q_complex_17_cast_fp16")]; + tensor var_3386 = const()[name = tensor("op_3386"), val = tensor([1, 1, 16, 32, 2])]; + tensor k_complex_17_cast_fp16 = reshape(shape = var_3386, x = k_33_cast_fp16)[name = tensor("k_complex_17_cast_fp16")]; + tensor var_3390_begin_0 = const()[name = tensor("op_3390_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_3390_end_0 = const()[name = tensor("op_3390_end_0"), val = tensor([1, 1, 16, 32, 1])]; + tensor var_3390_end_mask_0 = const()[name = tensor("op_3390_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_3390_squeeze_mask_0 = const()[name = tensor("op_3390_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_3390_cast_fp16 = slice_by_index(begin = var_3390_begin_0, end = var_3390_end_0, end_mask = var_3390_end_mask_0, squeeze_mask = var_3390_squeeze_mask_0, x = q_complex_17_cast_fp16)[name = tensor("op_3390_cast_fp16")]; + tensor var_3398_begin_0 = const()[name = tensor("op_3398_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_3398_end_0 = const()[name = tensor("op_3398_end_0"), val = tensor([1, 1, 16, 32, 2])]; + tensor var_3398_end_mask_0 = const()[name = tensor("op_3398_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_3398_squeeze_mask_0 = const()[name = tensor("op_3398_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_3398_cast_fp16 = slice_by_index(begin = var_3398_begin_0, end = var_3398_end_0, end_mask = var_3398_end_mask_0, squeeze_mask = var_3398_squeeze_mask_0, x = q_complex_17_cast_fp16)[name = tensor("op_3398_cast_fp16")]; + tensor var_3406_begin_0 = const()[name = tensor("op_3406_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_3406_end_0 = const()[name = tensor("op_3406_end_0"), val = tensor([1, 1, 16, 32, 1])]; + tensor var_3406_end_mask_0 = const()[name = tensor("op_3406_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_3406_squeeze_mask_0 = const()[name = tensor("op_3406_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_3406_cast_fp16 = slice_by_index(begin = var_3406_begin_0, end = var_3406_end_0, end_mask = var_3406_end_mask_0, squeeze_mask = var_3406_squeeze_mask_0, x = k_complex_17_cast_fp16)[name = tensor("op_3406_cast_fp16")]; + tensor var_3414_begin_0 = const()[name = tensor("op_3414_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_3414_end_0 = const()[name = tensor("op_3414_end_0"), val = tensor([1, 1, 16, 32, 2])]; + tensor var_3414_end_mask_0 = const()[name = tensor("op_3414_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_3414_squeeze_mask_0 = const()[name = tensor("op_3414_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_3414_cast_fp16 = slice_by_index(begin = var_3414_begin_0, end = var_3414_end_0, end_mask = var_3414_end_mask_0, squeeze_mask = var_3414_squeeze_mask_0, x = k_complex_17_cast_fp16)[name = tensor("op_3414_cast_fp16")]; + tensor freqs_17_to_fp16 = const()[name = tensor("freqs_17_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(208824960)))]; + tensor var_3420_cast_fp16 = mul(x = freqs_17_to_fp16, y = ts_53_cast_fp16)[name = tensor("op_3420_cast_fp16")]; + tensor rotr_17_cast_fp16 = cos(x = var_3420_cast_fp16)[name = tensor("rotr_17_cast_fp16")]; + tensor roti_17_cast_fp16 = sin(x = var_3420_cast_fp16)[name = tensor("roti_17_cast_fp16")]; + tensor var_3424_cast_fp16 = mul(x = var_3390_cast_fp16, y = rotr_17_cast_fp16)[name = tensor("op_3424_cast_fp16")]; + tensor var_3425_cast_fp16 = mul(x = var_3398_cast_fp16, y = roti_17_cast_fp16)[name = tensor("op_3425_cast_fp16")]; + tensor qor_33_cast_fp16 = sub(x = var_3424_cast_fp16, y = var_3425_cast_fp16)[name = tensor("qor_33_cast_fp16")]; + tensor var_3428_cast_fp16 = mul(x = var_3390_cast_fp16, y = roti_17_cast_fp16)[name = tensor("op_3428_cast_fp16")]; + tensor var_3429_cast_fp16 = mul(x = var_3398_cast_fp16, y = rotr_17_cast_fp16)[name = tensor("op_3429_cast_fp16")]; + tensor qoi_33_cast_fp16 = add(x = var_3428_cast_fp16, y = var_3429_cast_fp16)[name = tensor("qoi_33_cast_fp16")]; + tensor var_3432_cast_fp16 = mul(x = var_3406_cast_fp16, y = rotr_17_cast_fp16)[name = tensor("op_3432_cast_fp16")]; + tensor var_3433_cast_fp16 = mul(x = var_3414_cast_fp16, y = roti_17_cast_fp16)[name = tensor("op_3433_cast_fp16")]; + tensor kor_33_cast_fp16 = sub(x = var_3432_cast_fp16, y = var_3433_cast_fp16)[name = tensor("kor_33_cast_fp16")]; + tensor var_3436_cast_fp16 = mul(x = var_3406_cast_fp16, y = roti_17_cast_fp16)[name = tensor("op_3436_cast_fp16")]; + tensor var_3437_cast_fp16 = mul(x = var_3414_cast_fp16, y = rotr_17_cast_fp16)[name = tensor("op_3437_cast_fp16")]; + tensor koi_33_cast_fp16 = add(x = var_3436_cast_fp16, y = var_3437_cast_fp16)[name = tensor("koi_33_cast_fp16")]; + tensor qo_17_axis_0 = const()[name = tensor("qo_17_axis_0"), val = tensor(-1)]; + tensor qo_17_cast_fp16 = stack(axis = qo_17_axis_0, values = (qor_33_cast_fp16, qoi_33_cast_fp16))[name = tensor("qo_17_cast_fp16")]; + tensor ko_17_axis_0 = const()[name = tensor("ko_17_axis_0"), val = tensor(-1)]; + tensor ko_17_cast_fp16 = stack(axis = ko_17_axis_0, values = (kor_33_cast_fp16, koi_33_cast_fp16))[name = tensor("ko_17_cast_fp16")]; + tensor var_3470 = const()[name = tensor("op_3470"), val = tensor([1, 1, 16, 64])]; + tensor q_51_cast_fp16 = reshape(shape = var_3470, x = qo_17_cast_fp16)[name = tensor("q_51_cast_fp16")]; + tensor var_3476 = const()[name = tensor("op_3476"), val = tensor([1, 1, 16, 64])]; + tensor k_35_cast_fp16 = reshape(shape = var_3476, x = ko_17_cast_fp16)[name = tensor("k_35_cast_fp16")]; + tensor _inversed_3501_y_0_to_fp16 = const()[name = tensor("_inversed_3501_y_0_to_fp16"), val = tensor(0x1p-9)]; + tensor _inversed_3501_cast_fp16 = mul(x = ts_53_cast_fp16, y = _inversed_3501_y_0_to_fp16)[name = tensor("_inversed_3501_cast_fp16")]; + tensor var_3502_cast_fp16 = floor(x = _inversed_3501_cast_fp16)[name = tensor("op_3502_cast_fp16")]; + tensor var_3503_to_fp16 = const()[name = tensor("op_3503_to_fp16"), val = tensor(0x1p+9)]; + tensor var_3504_cast_fp16 = mul(x = var_3502_cast_fp16, y = var_3503_to_fp16)[name = tensor("op_3504_cast_fp16")]; + tensor write_indices_float_35_cast_fp16 = sub(x = ts_53_cast_fp16, y = var_3504_cast_fp16)[name = tensor("write_indices_float_35_cast_fp16")]; + tensor var_3511_dtype_0 = const()[name = tensor("op_3511_dtype_0"), val = tensor("int32")]; + tensor write_indices_17_reps_0 = const()[name = tensor("write_indices_17_reps_0"), val = tensor([1, 1, 16, 64])]; + tensor write_indices_float_35_cast_fp16_to_int32 = cast(dtype = var_3511_dtype_0, x = write_indices_float_35_cast_fp16)[name = tensor("cast_512")]; + tensor write_indices_17 = tile(reps = write_indices_17_reps_0, x = write_indices_float_35_cast_fp16_to_int32)[name = tensor("write_indices_17")]; + tensor var_3521_begin_0 = const()[name = tensor("op_3521_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_3521_end_0 = const()[name = tensor("op_3521_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor var_3521_end_mask_0 = const()[name = tensor("op_3521_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_3521_squeeze_mask_0 = const()[name = tensor("op_3521_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor cache8_to_fp16_dtype_0 = const()[name = tensor("cache8_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor cache8_to_fp16 = cast(dtype = cache8_to_fp16_dtype_0, x = cache8)[name = tensor("cast_511")]; + tensor var_3521_cast_fp16 = slice_by_index(begin = var_3521_begin_0, end = var_3521_end_0, end_mask = var_3521_end_mask_0, squeeze_mask = var_3521_squeeze_mask_0, x = cache8_to_fp16)[name = tensor("op_3521_cast_fp16")]; + tensor var_3523_axis_0 = const()[name = tensor("op_3523_axis_0"), val = tensor(1)]; + tensor var_3523_mode_0 = const()[name = tensor("op_3523_mode_0"), val = tensor("update")]; + tensor var_3523_validate_indices_0 = const()[name = tensor("op_3523_validate_indices_0"), val = tensor(false)]; + tensor var_3523_cast_fp16 = scatter_along_axis(axis = var_3523_axis_0, data = var_3521_cast_fp16, indices = write_indices_17, mode = var_3523_mode_0, updates = k_35_cast_fp16, validate_indices = var_3523_validate_indices_0)[name = tensor("op_3523_cast_fp16")]; + tensor concat_58 = const()[name = tensor("concat_58"), val = tensor([0, 0, 0, 0, 0])]; + tensor concat_59 = const()[name = tensor("concat_59"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_17_internal_tensor_assign_1_stride_0 = const()[name = tensor("new_cache_17_internal_tensor_assign_1_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_17_internal_tensor_assign_1_begin_mask_0 = const()[name = tensor("new_cache_17_internal_tensor_assign_1_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_17_internal_tensor_assign_1_end_mask_0 = const()[name = tensor("new_cache_17_internal_tensor_assign_1_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_17_internal_tensor_assign_1_squeeze_mask_0 = const()[name = tensor("new_cache_17_internal_tensor_assign_1_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_64 = const()[name = tensor("shape_64"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_16 = const()[name = tensor("reduce_prod_16"), val = tensor(1048576)]; + tensor range_1d_16_start_0 = const()[name = tensor("range_1d_16_start_0"), val = tensor(0)]; + tensor range_1d_16_step_0 = const()[name = tensor("range_1d_16_step_0"), val = tensor(1)]; + tensor range_1d_16 = range_1d(end = reduce_prod_16, start = range_1d_16_start_0, step = range_1d_16_step_0)[name = tensor("range_1d_16")]; + tensor reshape_80 = reshape(shape = shape_64, x = range_1d_16)[name = tensor("reshape_80")]; + tensor slice_by_index_16 = slice_by_index(begin = concat_58, begin_mask = new_cache_17_internal_tensor_assign_1_begin_mask_0, end = concat_59, end_mask = new_cache_17_internal_tensor_assign_1_end_mask_0, squeeze_mask = new_cache_17_internal_tensor_assign_1_squeeze_mask_0, stride = new_cache_17_internal_tensor_assign_1_stride_0, x = reshape_80)[name = tensor("slice_by_index_16")]; + tensor reshape_81_shape_0 = const()[name = tensor("reshape_81_shape_0"), val = tensor([-1])]; + tensor reshape_81 = reshape(shape = reshape_81_shape_0, x = slice_by_index_16)[name = tensor("reshape_81")]; + tensor reshape_82_shape_0 = const()[name = tensor("reshape_82_shape_0"), val = tensor([-1])]; + tensor reshape_82_cast_fp16 = reshape(shape = reshape_82_shape_0, x = var_3523_cast_fp16)[name = tensor("reshape_82_cast_fp16")]; + tensor reshape_83_shape_0 = const()[name = tensor("reshape_83_shape_0"), val = tensor([-1])]; + tensor reshape_83_cast_fp16 = reshape(shape = reshape_83_shape_0, x = cache8_to_fp16)[name = tensor("reshape_83_cast_fp16")]; + tensor scatter_16_mode_0 = const()[name = tensor("scatter_16_mode_0"), val = tensor("update")]; + tensor scatter_16_axis_0 = const()[name = tensor("scatter_16_axis_0"), val = tensor(0)]; + tensor scatter_16_validate_indices_0 = const()[name = tensor("scatter_16_validate_indices_0"), val = tensor(false)]; + tensor scatter_16_cast_fp16 = scatter(axis = scatter_16_axis_0, data = reshape_83_cast_fp16, indices = reshape_81, mode = scatter_16_mode_0, updates = reshape_82_cast_fp16, validate_indices = scatter_16_validate_indices_0)[name = tensor("scatter_16_cast_fp16")]; + tensor reshape_84_cast_fp16 = reshape(shape = shape_64, x = scatter_16_cast_fp16)[name = tensor("reshape_84_cast_fp16")]; + tensor var_3531_begin_0 = const()[name = tensor("op_3531_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor var_3531_end_0 = const()[name = tensor("op_3531_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor var_3531_end_mask_0 = const()[name = tensor("op_3531_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_3531_squeeze_mask_0 = const()[name = tensor("op_3531_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor var_3531_cast_fp16 = slice_by_index(begin = var_3531_begin_0, end = var_3531_end_0, end_mask = var_3531_end_mask_0, squeeze_mask = var_3531_squeeze_mask_0, x = reshape_84_cast_fp16)[name = tensor("op_3531_cast_fp16")]; + tensor var_3533_axis_0 = const()[name = tensor("op_3533_axis_0"), val = tensor(1)]; + tensor var_3533_mode_0 = const()[name = tensor("op_3533_mode_0"), val = tensor("update")]; + tensor var_3533_validate_indices_0 = const()[name = tensor("op_3533_validate_indices_0"), val = tensor(false)]; + tensor var_3533_cast_fp16 = scatter_along_axis(axis = var_3533_axis_0, data = var_3531_cast_fp16, indices = write_indices_17, mode = var_3533_mode_0, updates = v_17_cast_fp16, validate_indices = var_3533_validate_indices_0)[name = tensor("op_3533_cast_fp16")]; + tensor concat_60 = const()[name = tensor("concat_60"), val = tensor([1, 0, 0, 0, 0])]; + tensor concat_61 = const()[name = tensor("concat_61"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_17_internal_tensor_assign_2_stride_0 = const()[name = tensor("new_cache_17_internal_tensor_assign_2_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_17_internal_tensor_assign_2_begin_mask_0 = const()[name = tensor("new_cache_17_internal_tensor_assign_2_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_17_internal_tensor_assign_2_end_mask_0 = const()[name = tensor("new_cache_17_internal_tensor_assign_2_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_17_internal_tensor_assign_2_squeeze_mask_0 = const()[name = tensor("new_cache_17_internal_tensor_assign_2_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_65 = const()[name = tensor("shape_65"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_17 = const()[name = tensor("reduce_prod_17"), val = tensor(1048576)]; + tensor range_1d_17_start_0 = const()[name = tensor("range_1d_17_start_0"), val = tensor(0)]; + tensor range_1d_17_step_0 = const()[name = tensor("range_1d_17_step_0"), val = tensor(1)]; + tensor range_1d_17 = range_1d(end = reduce_prod_17, start = range_1d_17_start_0, step = range_1d_17_step_0)[name = tensor("range_1d_17")]; + tensor reshape_85 = reshape(shape = shape_65, x = range_1d_17)[name = tensor("reshape_85")]; + tensor slice_by_index_17 = slice_by_index(begin = concat_60, begin_mask = new_cache_17_internal_tensor_assign_2_begin_mask_0, end = concat_61, end_mask = new_cache_17_internal_tensor_assign_2_end_mask_0, squeeze_mask = new_cache_17_internal_tensor_assign_2_squeeze_mask_0, stride = new_cache_17_internal_tensor_assign_2_stride_0, x = reshape_85)[name = tensor("slice_by_index_17")]; + tensor reshape_86_shape_0 = const()[name = tensor("reshape_86_shape_0"), val = tensor([-1])]; + tensor reshape_86 = reshape(shape = reshape_86_shape_0, x = slice_by_index_17)[name = tensor("reshape_86")]; + tensor reshape_87_shape_0 = const()[name = tensor("reshape_87_shape_0"), val = tensor([-1])]; + tensor reshape_87_cast_fp16 = reshape(shape = reshape_87_shape_0, x = var_3533_cast_fp16)[name = tensor("reshape_87_cast_fp16")]; + tensor reshape_88_shape_0 = const()[name = tensor("reshape_88_shape_0"), val = tensor([-1])]; + tensor reshape_88_cast_fp16 = reshape(shape = reshape_88_shape_0, x = reshape_84_cast_fp16)[name = tensor("reshape_88_cast_fp16")]; + tensor scatter_17_mode_0 = const()[name = tensor("scatter_17_mode_0"), val = tensor("update")]; + tensor scatter_17_axis_0 = const()[name = tensor("scatter_17_axis_0"), val = tensor(0)]; + tensor scatter_17_validate_indices_0 = const()[name = tensor("scatter_17_validate_indices_0"), val = tensor(false)]; + tensor scatter_17_cast_fp16 = scatter(axis = scatter_17_axis_0, data = reshape_88_cast_fp16, indices = reshape_86, mode = scatter_17_mode_0, updates = reshape_87_cast_fp16, validate_indices = scatter_17_validate_indices_0)[name = tensor("scatter_17_cast_fp16")]; + tensor reshape_89_cast_fp16 = reshape(shape = shape_65, x = scatter_17_cast_fp16)[name = tensor("reshape_89_cast_fp16")]; + tensor reshape_89_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("reshape_89_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor keys_49_begin_0 = const()[name = tensor("keys_49_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor keys_49_end_0 = const()[name = tensor("keys_49_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor keys_49_end_mask_0 = const()[name = tensor("keys_49_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor keys_49_squeeze_mask_0 = const()[name = tensor("keys_49_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor keys_49_cast_fp16 = slice_by_index(begin = keys_49_begin_0, end = keys_49_end_0, end_mask = keys_49_end_mask_0, squeeze_mask = keys_49_squeeze_mask_0, x = reshape_89_cast_fp16)[name = tensor("keys_49_cast_fp16")]; + tensor values_49_begin_0 = const()[name = tensor("values_49_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor values_49_end_0 = const()[name = tensor("values_49_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor values_49_end_mask_0 = const()[name = tensor("values_49_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor values_49_squeeze_mask_0 = const()[name = tensor("values_49_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor values_49_cast_fp16 = slice_by_index(begin = values_49_begin_0, end = values_49_end_0, end_mask = values_49_end_mask_0, squeeze_mask = values_49_squeeze_mask_0, x = reshape_89_cast_fp16)[name = tensor("values_49_cast_fp16")]; + tensor var_3545_cast_fp16 = not_equal(x = keys_49_cast_fp16, y = keys_49_cast_fp16)[name = tensor("op_3545_cast_fp16")]; + tensor keys_51_cast_fp16 = select(a = var_463_to_fp16, b = keys_49_cast_fp16, cond = var_3545_cast_fp16)[name = tensor("keys_51_cast_fp16")]; + tensor var_3553_cast_fp16 = not_equal(x = values_49_cast_fp16, y = values_49_cast_fp16)[name = tensor("op_3553_cast_fp16")]; + tensor values_51_cast_fp16 = select(a = var_463_to_fp16, b = values_49_cast_fp16, cond = var_3553_cast_fp16)[name = tensor("values_51_cast_fp16")]; + tensor var_3577 = const()[name = tensor("op_3577"), val = tensor([0, 2, 1, 3])]; + tensor var_3593 = const()[name = tensor("op_3593"), val = tensor([1, 1, 1])]; + tensor var_3594_cast_fp16 = reshape(shape = var_3593, x = position8_to_fp16)[name = tensor("op_3594_cast_fp16")]; + tensor var_3614_to_fp16 = const()[name = tensor("op_3614_to_fp16"), val = tensor(0x1p+0)]; + tensor valid_len_17_cast_fp16 = add(x = var_3594_cast_fp16, y = var_3614_to_fp16)[name = tensor("valid_len_17_cast_fp16")]; + tensor valid_mask_17_cast_fp16 = less(x = k_positions_1_promoted_to_fp16, y = valid_len_17_cast_fp16)[name = tensor("valid_mask_17_cast_fp16")]; + tensor causal_mask_17_cast_fp16 = less_equal(x = k_positions_1_promoted_to_fp16, y = var_3594_cast_fp16)[name = tensor("causal_mask_17_cast_fp16")]; + tensor attn_mask_33 = logical_and(x = valid_mask_17_cast_fp16, y = causal_mask_17_cast_fp16)[name = tensor("attn_mask_33")]; + tensor attn_mask_35_axes_0 = const()[name = tensor("attn_mask_35_axes_0"), val = tensor([1])]; + tensor attn_mask_35 = expand_dims(axes = attn_mask_35_axes_0, x = attn_mask_33)[name = tensor("attn_mask_35")]; + tensor var_3632_transpose_x_0 = const()[name = tensor("op_3632_transpose_x_0"), val = tensor(false)]; + tensor var_3632_transpose_y_0 = const()[name = tensor("op_3632_transpose_y_0"), val = tensor(false)]; + tensor transpose_88_perm_0 = const()[name = tensor("transpose_88_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_89_perm_0 = const()[name = tensor("transpose_89_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_89 = transpose(perm = transpose_89_perm_0, x = keys_51_cast_fp16)[name = tensor("transpose_181")]; + tensor transpose_88 = transpose(perm = transpose_88_perm_0, x = q_51_cast_fp16)[name = tensor("transpose_182")]; + tensor var_3632_cast_fp16 = matmul(transpose_x = var_3632_transpose_x_0, transpose_y = var_3632_transpose_y_0, x = transpose_88, y = transpose_89)[name = tensor("op_3632_cast_fp16")]; + tensor var_3626_to_fp16 = const()[name = tensor("op_3626_to_fp16"), val = tensor([0x1p-3])]; + tensor attn_weights_49_cast_fp16 = mul(x = var_3632_cast_fp16, y = var_3626_to_fp16)[name = tensor("attn_weights_49_cast_fp16")]; + tensor var_3634 = logical_not(x = attn_mask_35)[name = tensor("op_3634")]; + tensor var_3635_to_fp16 = const()[name = tensor("op_3635_to_fp16"), val = tensor(-inf)]; + tensor attn_weights_51_cast_fp16 = select(a = var_3635_to_fp16, b = attn_weights_49_cast_fp16, cond = var_3634)[name = tensor("attn_weights_51_cast_fp16")]; + tensor var_3637 = const()[name = tensor("op_3637"), val = tensor(-1)]; + tensor attn_weights_53_cast_fp16 = softmax(axis = var_3637, x = attn_weights_51_cast_fp16)[name = tensor("attn_weights_53_cast_fp16")]; + tensor attn_output_17_transpose_x_0 = const()[name = tensor("attn_output_17_transpose_x_0"), val = tensor(false)]; + tensor attn_output_17_transpose_y_0 = const()[name = tensor("attn_output_17_transpose_y_0"), val = tensor(false)]; + tensor values_53_cast_fp16 = transpose(perm = var_3577, x = values_51_cast_fp16)[name = tensor("transpose_183")]; + tensor attn_output_17_cast_fp16 = matmul(transpose_x = attn_output_17_transpose_x_0, transpose_y = attn_output_17_transpose_y_0, x = attn_weights_53_cast_fp16, y = values_53_cast_fp16)[name = tensor("attn_output_17_cast_fp16")]; + tensor var_3645 = const()[name = tensor("op_3645"), val = tensor([0, 2, 1, 3])]; + tensor var_3650 = const()[name = tensor("op_3650"), val = tensor([1, 1, 1024])]; + tensor var_3646_cast_fp16 = transpose(perm = var_3645, x = attn_output_17_cast_fp16)[name = tensor("transpose_180")]; + tensor input_103_cast_fp16 = reshape(shape = var_3650, x = var_3646_cast_fp16)[name = tensor("input_103_cast_fp16")]; + tensor attn8_out_proj_weight_to_fp16 = const()[name = tensor("attn8_out_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(208825088)))]; + tensor linear_34_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = attn8_out_proj_weight_to_fp16, x = input_103_cast_fp16)[name = tensor("linear_34_cast_fp16")]; + tensor var_3656_to_fp16 = const()[name = tensor("op_3656_to_fp16"), val = tensor(0x1p+0)]; + tensor var_3657_cast_fp16 = add(x = position8_to_fp16, y = var_3656_to_fp16)[name = tensor("op_3657_cast_fp16")]; + tensor var_3657_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_3657_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor input_105_cast_fp16 = add(x = input_99_cast_fp16, y = linear_34_cast_fp16)[name = tensor("input_105_cast_fp16")]; + tensor input_107_axes_0 = const()[name = tensor("input_107_axes_0"), val = tensor([-1])]; + tensor norm8_2_weight_to_fp16 = const()[name = tensor("norm8_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(210922304)))]; + tensor norm8_2_bias_to_fp16 = const()[name = tensor("norm8_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(210924416)))]; + tensor var_3661_to_fp16 = const()[name = tensor("op_3661_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_107_cast_fp16 = layer_norm(axes = input_107_axes_0, beta = norm8_2_bias_to_fp16, epsilon = var_3661_to_fp16, gamma = norm8_2_weight_to_fp16, x = input_105_cast_fp16)[name = tensor("input_107_cast_fp16")]; + tensor linear8_1_weight_to_fp16 = const()[name = tensor("linear8_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(210926528)))]; + tensor linear_35_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = linear8_1_weight_to_fp16, x = input_107_cast_fp16)[name = tensor("linear_35_cast_fp16")]; + tensor input_109_mode_0 = const()[name = tensor("input_109_mode_0"), val = tensor("EXACT")]; + tensor input_109_cast_fp16 = gelu(mode = input_109_mode_0, x = linear_35_cast_fp16)[name = tensor("input_109_cast_fp16")]; + tensor linear8_2_weight_to_fp16 = const()[name = tensor("linear8_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(219315200)))]; + tensor linear_36_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = linear8_2_weight_to_fp16, x = input_109_cast_fp16)[name = tensor("linear_36_cast_fp16")]; + tensor input_111_cast_fp16 = add(x = input_105_cast_fp16, y = linear_36_cast_fp16)[name = tensor("input_111_cast_fp16")]; + tensor input_113_axes_0 = const()[name = tensor("input_113_axes_0"), val = tensor([-1])]; + tensor norm9_1_weight_to_fp16 = const()[name = tensor("norm9_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(227703872)))]; + tensor norm9_1_bias_to_fp16 = const()[name = tensor("norm9_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(227705984)))]; + tensor var_3678_to_fp16 = const()[name = tensor("op_3678_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_113_cast_fp16 = layer_norm(axes = input_113_axes_0, beta = norm9_1_bias_to_fp16, epsilon = var_3678_to_fp16, gamma = norm9_1_weight_to_fp16, x = input_111_cast_fp16)[name = tensor("input_113_cast_fp16")]; + tensor attn9_in_proj_weight_to_fp16 = const()[name = tensor("attn9_in_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(227708096)))]; + tensor linear_37_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = attn9_in_proj_weight_to_fp16, x = input_113_cast_fp16)[name = tensor("linear_37_cast_fp16")]; + tensor var_3692 = const()[name = tensor("op_3692"), val = tensor([1, 1, 3, 16, 64])]; + tensor qkv_19_cast_fp16 = reshape(shape = var_3692, x = linear_37_cast_fp16)[name = tensor("qkv_19_cast_fp16")]; + tensor q_55_begin_0 = const()[name = tensor("q_55_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor q_55_end_0 = const()[name = tensor("q_55_end_0"), val = tensor([1, 1, 1, 16, 64])]; + tensor q_55_end_mask_0 = const()[name = tensor("q_55_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor q_55_squeeze_mask_0 = const()[name = tensor("q_55_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor q_55_cast_fp16 = slice_by_index(begin = q_55_begin_0, end = q_55_end_0, end_mask = q_55_end_mask_0, squeeze_mask = q_55_squeeze_mask_0, x = qkv_19_cast_fp16)[name = tensor("q_55_cast_fp16")]; + tensor k_37_begin_0 = const()[name = tensor("k_37_begin_0"), val = tensor([0, 0, 1, 0, 0])]; + tensor k_37_end_0 = const()[name = tensor("k_37_end_0"), val = tensor([1, 1, 2, 16, 64])]; + tensor k_37_end_mask_0 = const()[name = tensor("k_37_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor k_37_squeeze_mask_0 = const()[name = tensor("k_37_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor k_37_cast_fp16 = slice_by_index(begin = k_37_begin_0, end = k_37_end_0, end_mask = k_37_end_mask_0, squeeze_mask = k_37_squeeze_mask_0, x = qkv_19_cast_fp16)[name = tensor("k_37_cast_fp16")]; + tensor v_19_begin_0 = const()[name = tensor("v_19_begin_0"), val = tensor([0, 0, 2, 0, 0])]; + tensor v_19_end_0 = const()[name = tensor("v_19_end_0"), val = tensor([1, 1, 3, 16, 64])]; + tensor v_19_end_mask_0 = const()[name = tensor("v_19_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor v_19_squeeze_mask_0 = const()[name = tensor("v_19_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor v_19_cast_fp16 = slice_by_index(begin = v_19_begin_0, end = v_19_end_0, end_mask = v_19_end_mask_0, squeeze_mask = v_19_squeeze_mask_0, x = qkv_19_cast_fp16)[name = tensor("v_19_cast_fp16")]; + tensor var_3758 = const()[name = tensor("op_3758"), val = tensor([1, 1, 1, 1])]; + tensor position9_to_fp16_dtype_0 = const()[name = tensor("position9_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor position9_to_fp16 = cast(dtype = position9_to_fp16_dtype_0, x = position9)[name = tensor("cast_508")]; + tensor ts_59_cast_fp16 = reshape(shape = var_3758, x = position9_to_fp16)[name = tensor("ts_59_cast_fp16")]; + tensor var_3765 = const()[name = tensor("op_3765"), val = tensor([1, 1, 16, 32, 2])]; + tensor q_complex_19_cast_fp16 = reshape(shape = var_3765, x = q_55_cast_fp16)[name = tensor("q_complex_19_cast_fp16")]; + tensor var_3772 = const()[name = tensor("op_3772"), val = tensor([1, 1, 16, 32, 2])]; + tensor k_complex_19_cast_fp16 = reshape(shape = var_3772, x = k_37_cast_fp16)[name = tensor("k_complex_19_cast_fp16")]; + tensor var_3776_begin_0 = const()[name = tensor("op_3776_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_3776_end_0 = const()[name = tensor("op_3776_end_0"), val = tensor([1, 1, 16, 32, 1])]; + tensor var_3776_end_mask_0 = const()[name = tensor("op_3776_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_3776_squeeze_mask_0 = const()[name = tensor("op_3776_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_3776_cast_fp16 = slice_by_index(begin = var_3776_begin_0, end = var_3776_end_0, end_mask = var_3776_end_mask_0, squeeze_mask = var_3776_squeeze_mask_0, x = q_complex_19_cast_fp16)[name = tensor("op_3776_cast_fp16")]; + tensor var_3784_begin_0 = const()[name = tensor("op_3784_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_3784_end_0 = const()[name = tensor("op_3784_end_0"), val = tensor([1, 1, 16, 32, 2])]; + tensor var_3784_end_mask_0 = const()[name = tensor("op_3784_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_3784_squeeze_mask_0 = const()[name = tensor("op_3784_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_3784_cast_fp16 = slice_by_index(begin = var_3784_begin_0, end = var_3784_end_0, end_mask = var_3784_end_mask_0, squeeze_mask = var_3784_squeeze_mask_0, x = q_complex_19_cast_fp16)[name = tensor("op_3784_cast_fp16")]; + tensor var_3792_begin_0 = const()[name = tensor("op_3792_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_3792_end_0 = const()[name = tensor("op_3792_end_0"), val = tensor([1, 1, 16, 32, 1])]; + tensor var_3792_end_mask_0 = const()[name = tensor("op_3792_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_3792_squeeze_mask_0 = const()[name = tensor("op_3792_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_3792_cast_fp16 = slice_by_index(begin = var_3792_begin_0, end = var_3792_end_0, end_mask = var_3792_end_mask_0, squeeze_mask = var_3792_squeeze_mask_0, x = k_complex_19_cast_fp16)[name = tensor("op_3792_cast_fp16")]; + tensor var_3800_begin_0 = const()[name = tensor("op_3800_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_3800_end_0 = const()[name = tensor("op_3800_end_0"), val = tensor([1, 1, 16, 32, 2])]; + tensor var_3800_end_mask_0 = const()[name = tensor("op_3800_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_3800_squeeze_mask_0 = const()[name = tensor("op_3800_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_3800_cast_fp16 = slice_by_index(begin = var_3800_begin_0, end = var_3800_end_0, end_mask = var_3800_end_mask_0, squeeze_mask = var_3800_squeeze_mask_0, x = k_complex_19_cast_fp16)[name = tensor("op_3800_cast_fp16")]; + tensor freqs_19_to_fp16 = const()[name = tensor("freqs_19_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(233999616)))]; + tensor var_3806_cast_fp16 = mul(x = freqs_19_to_fp16, y = ts_59_cast_fp16)[name = tensor("op_3806_cast_fp16")]; + tensor rotr_19_cast_fp16 = cos(x = var_3806_cast_fp16)[name = tensor("rotr_19_cast_fp16")]; + tensor roti_19_cast_fp16 = sin(x = var_3806_cast_fp16)[name = tensor("roti_19_cast_fp16")]; + tensor var_3810_cast_fp16 = mul(x = var_3776_cast_fp16, y = rotr_19_cast_fp16)[name = tensor("op_3810_cast_fp16")]; + tensor var_3811_cast_fp16 = mul(x = var_3784_cast_fp16, y = roti_19_cast_fp16)[name = tensor("op_3811_cast_fp16")]; + tensor qor_37_cast_fp16 = sub(x = var_3810_cast_fp16, y = var_3811_cast_fp16)[name = tensor("qor_37_cast_fp16")]; + tensor var_3814_cast_fp16 = mul(x = var_3776_cast_fp16, y = roti_19_cast_fp16)[name = tensor("op_3814_cast_fp16")]; + tensor var_3815_cast_fp16 = mul(x = var_3784_cast_fp16, y = rotr_19_cast_fp16)[name = tensor("op_3815_cast_fp16")]; + tensor qoi_37_cast_fp16 = add(x = var_3814_cast_fp16, y = var_3815_cast_fp16)[name = tensor("qoi_37_cast_fp16")]; + tensor var_3818_cast_fp16 = mul(x = var_3792_cast_fp16, y = rotr_19_cast_fp16)[name = tensor("op_3818_cast_fp16")]; + tensor var_3819_cast_fp16 = mul(x = var_3800_cast_fp16, y = roti_19_cast_fp16)[name = tensor("op_3819_cast_fp16")]; + tensor kor_37_cast_fp16 = sub(x = var_3818_cast_fp16, y = var_3819_cast_fp16)[name = tensor("kor_37_cast_fp16")]; + tensor var_3822_cast_fp16 = mul(x = var_3792_cast_fp16, y = roti_19_cast_fp16)[name = tensor("op_3822_cast_fp16")]; + tensor var_3823_cast_fp16 = mul(x = var_3800_cast_fp16, y = rotr_19_cast_fp16)[name = tensor("op_3823_cast_fp16")]; + tensor koi_37_cast_fp16 = add(x = var_3822_cast_fp16, y = var_3823_cast_fp16)[name = tensor("koi_37_cast_fp16")]; + tensor qo_19_axis_0 = const()[name = tensor("qo_19_axis_0"), val = tensor(-1)]; + tensor qo_19_cast_fp16 = stack(axis = qo_19_axis_0, values = (qor_37_cast_fp16, qoi_37_cast_fp16))[name = tensor("qo_19_cast_fp16")]; + tensor ko_19_axis_0 = const()[name = tensor("ko_19_axis_0"), val = tensor(-1)]; + tensor ko_19_cast_fp16 = stack(axis = ko_19_axis_0, values = (kor_37_cast_fp16, koi_37_cast_fp16))[name = tensor("ko_19_cast_fp16")]; + tensor var_3856 = const()[name = tensor("op_3856"), val = tensor([1, 1, 16, 64])]; + tensor q_57_cast_fp16 = reshape(shape = var_3856, x = qo_19_cast_fp16)[name = tensor("q_57_cast_fp16")]; + tensor var_3862 = const()[name = tensor("op_3862"), val = tensor([1, 1, 16, 64])]; + tensor k_39_cast_fp16 = reshape(shape = var_3862, x = ko_19_cast_fp16)[name = tensor("k_39_cast_fp16")]; + tensor _inversed_3887_y_0_to_fp16 = const()[name = tensor("_inversed_3887_y_0_to_fp16"), val = tensor(0x1p-9)]; + tensor _inversed_3887_cast_fp16 = mul(x = ts_59_cast_fp16, y = _inversed_3887_y_0_to_fp16)[name = tensor("_inversed_3887_cast_fp16")]; + tensor var_3888_cast_fp16 = floor(x = _inversed_3887_cast_fp16)[name = tensor("op_3888_cast_fp16")]; + tensor var_3889_to_fp16 = const()[name = tensor("op_3889_to_fp16"), val = tensor(0x1p+9)]; + tensor var_3890_cast_fp16 = mul(x = var_3888_cast_fp16, y = var_3889_to_fp16)[name = tensor("op_3890_cast_fp16")]; + tensor write_indices_float_39_cast_fp16 = sub(x = ts_59_cast_fp16, y = var_3890_cast_fp16)[name = tensor("write_indices_float_39_cast_fp16")]; + tensor var_3897_dtype_0 = const()[name = tensor("op_3897_dtype_0"), val = tensor("int32")]; + tensor write_indices_19_reps_0 = const()[name = tensor("write_indices_19_reps_0"), val = tensor([1, 1, 16, 64])]; + tensor write_indices_float_39_cast_fp16_to_int32 = cast(dtype = var_3897_dtype_0, x = write_indices_float_39_cast_fp16)[name = tensor("cast_507")]; + tensor write_indices_19 = tile(reps = write_indices_19_reps_0, x = write_indices_float_39_cast_fp16_to_int32)[name = tensor("write_indices_19")]; + tensor var_3907_begin_0 = const()[name = tensor("op_3907_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_3907_end_0 = const()[name = tensor("op_3907_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor var_3907_end_mask_0 = const()[name = tensor("op_3907_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_3907_squeeze_mask_0 = const()[name = tensor("op_3907_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor cache9_to_fp16_dtype_0 = const()[name = tensor("cache9_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor cache9_to_fp16 = cast(dtype = cache9_to_fp16_dtype_0, x = cache9)[name = tensor("cast_506")]; + tensor var_3907_cast_fp16 = slice_by_index(begin = var_3907_begin_0, end = var_3907_end_0, end_mask = var_3907_end_mask_0, squeeze_mask = var_3907_squeeze_mask_0, x = cache9_to_fp16)[name = tensor("op_3907_cast_fp16")]; + tensor var_3909_axis_0 = const()[name = tensor("op_3909_axis_0"), val = tensor(1)]; + tensor var_3909_mode_0 = const()[name = tensor("op_3909_mode_0"), val = tensor("update")]; + tensor var_3909_validate_indices_0 = const()[name = tensor("op_3909_validate_indices_0"), val = tensor(false)]; + tensor var_3909_cast_fp16 = scatter_along_axis(axis = var_3909_axis_0, data = var_3907_cast_fp16, indices = write_indices_19, mode = var_3909_mode_0, updates = k_39_cast_fp16, validate_indices = var_3909_validate_indices_0)[name = tensor("op_3909_cast_fp16")]; + tensor concat_65 = const()[name = tensor("concat_65"), val = tensor([0, 0, 0, 0, 0])]; + tensor concat_66 = const()[name = tensor("concat_66"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_19_internal_tensor_assign_1_stride_0 = const()[name = tensor("new_cache_19_internal_tensor_assign_1_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_19_internal_tensor_assign_1_begin_mask_0 = const()[name = tensor("new_cache_19_internal_tensor_assign_1_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_19_internal_tensor_assign_1_end_mask_0 = const()[name = tensor("new_cache_19_internal_tensor_assign_1_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_19_internal_tensor_assign_1_squeeze_mask_0 = const()[name = tensor("new_cache_19_internal_tensor_assign_1_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_66 = const()[name = tensor("shape_66"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_18 = const()[name = tensor("reduce_prod_18"), val = tensor(1048576)]; + tensor range_1d_18_start_0 = const()[name = tensor("range_1d_18_start_0"), val = tensor(0)]; + tensor range_1d_18_step_0 = const()[name = tensor("range_1d_18_step_0"), val = tensor(1)]; + tensor range_1d_18 = range_1d(end = reduce_prod_18, start = range_1d_18_start_0, step = range_1d_18_step_0)[name = tensor("range_1d_18")]; + tensor reshape_90 = reshape(shape = shape_66, x = range_1d_18)[name = tensor("reshape_90")]; + tensor slice_by_index_18 = slice_by_index(begin = concat_65, begin_mask = new_cache_19_internal_tensor_assign_1_begin_mask_0, end = concat_66, end_mask = new_cache_19_internal_tensor_assign_1_end_mask_0, squeeze_mask = new_cache_19_internal_tensor_assign_1_squeeze_mask_0, stride = new_cache_19_internal_tensor_assign_1_stride_0, x = reshape_90)[name = tensor("slice_by_index_18")]; + tensor reshape_91_shape_0 = const()[name = tensor("reshape_91_shape_0"), val = tensor([-1])]; + tensor reshape_91 = reshape(shape = reshape_91_shape_0, x = slice_by_index_18)[name = tensor("reshape_91")]; + tensor reshape_92_shape_0 = const()[name = tensor("reshape_92_shape_0"), val = tensor([-1])]; + tensor reshape_92_cast_fp16 = reshape(shape = reshape_92_shape_0, x = var_3909_cast_fp16)[name = tensor("reshape_92_cast_fp16")]; + tensor reshape_93_shape_0 = const()[name = tensor("reshape_93_shape_0"), val = tensor([-1])]; + tensor reshape_93_cast_fp16 = reshape(shape = reshape_93_shape_0, x = cache9_to_fp16)[name = tensor("reshape_93_cast_fp16")]; + tensor scatter_18_mode_0 = const()[name = tensor("scatter_18_mode_0"), val = tensor("update")]; + tensor scatter_18_axis_0 = const()[name = tensor("scatter_18_axis_0"), val = tensor(0)]; + tensor scatter_18_validate_indices_0 = const()[name = tensor("scatter_18_validate_indices_0"), val = tensor(false)]; + tensor scatter_18_cast_fp16 = scatter(axis = scatter_18_axis_0, data = reshape_93_cast_fp16, indices = reshape_91, mode = scatter_18_mode_0, updates = reshape_92_cast_fp16, validate_indices = scatter_18_validate_indices_0)[name = tensor("scatter_18_cast_fp16")]; + tensor reshape_94_cast_fp16 = reshape(shape = shape_66, x = scatter_18_cast_fp16)[name = tensor("reshape_94_cast_fp16")]; + tensor var_3917_begin_0 = const()[name = tensor("op_3917_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor var_3917_end_0 = const()[name = tensor("op_3917_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor var_3917_end_mask_0 = const()[name = tensor("op_3917_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_3917_squeeze_mask_0 = const()[name = tensor("op_3917_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor var_3917_cast_fp16 = slice_by_index(begin = var_3917_begin_0, end = var_3917_end_0, end_mask = var_3917_end_mask_0, squeeze_mask = var_3917_squeeze_mask_0, x = reshape_94_cast_fp16)[name = tensor("op_3917_cast_fp16")]; + tensor var_3919_axis_0 = const()[name = tensor("op_3919_axis_0"), val = tensor(1)]; + tensor var_3919_mode_0 = const()[name = tensor("op_3919_mode_0"), val = tensor("update")]; + tensor var_3919_validate_indices_0 = const()[name = tensor("op_3919_validate_indices_0"), val = tensor(false)]; + tensor var_3919_cast_fp16 = scatter_along_axis(axis = var_3919_axis_0, data = var_3917_cast_fp16, indices = write_indices_19, mode = var_3919_mode_0, updates = v_19_cast_fp16, validate_indices = var_3919_validate_indices_0)[name = tensor("op_3919_cast_fp16")]; + tensor concat_67 = const()[name = tensor("concat_67"), val = tensor([1, 0, 0, 0, 0])]; + tensor concat_68 = const()[name = tensor("concat_68"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_19_internal_tensor_assign_2_stride_0 = const()[name = tensor("new_cache_19_internal_tensor_assign_2_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_19_internal_tensor_assign_2_begin_mask_0 = const()[name = tensor("new_cache_19_internal_tensor_assign_2_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_19_internal_tensor_assign_2_end_mask_0 = const()[name = tensor("new_cache_19_internal_tensor_assign_2_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_19_internal_tensor_assign_2_squeeze_mask_0 = const()[name = tensor("new_cache_19_internal_tensor_assign_2_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_67 = const()[name = tensor("shape_67"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_19 = const()[name = tensor("reduce_prod_19"), val = tensor(1048576)]; + tensor range_1d_19_start_0 = const()[name = tensor("range_1d_19_start_0"), val = tensor(0)]; + tensor range_1d_19_step_0 = const()[name = tensor("range_1d_19_step_0"), val = tensor(1)]; + tensor range_1d_19 = range_1d(end = reduce_prod_19, start = range_1d_19_start_0, step = range_1d_19_step_0)[name = tensor("range_1d_19")]; + tensor reshape_95 = reshape(shape = shape_67, x = range_1d_19)[name = tensor("reshape_95")]; + tensor slice_by_index_19 = slice_by_index(begin = concat_67, begin_mask = new_cache_19_internal_tensor_assign_2_begin_mask_0, end = concat_68, end_mask = new_cache_19_internal_tensor_assign_2_end_mask_0, squeeze_mask = new_cache_19_internal_tensor_assign_2_squeeze_mask_0, stride = new_cache_19_internal_tensor_assign_2_stride_0, x = reshape_95)[name = tensor("slice_by_index_19")]; + tensor reshape_96_shape_0 = const()[name = tensor("reshape_96_shape_0"), val = tensor([-1])]; + tensor reshape_96 = reshape(shape = reshape_96_shape_0, x = slice_by_index_19)[name = tensor("reshape_96")]; + tensor reshape_97_shape_0 = const()[name = tensor("reshape_97_shape_0"), val = tensor([-1])]; + tensor reshape_97_cast_fp16 = reshape(shape = reshape_97_shape_0, x = var_3919_cast_fp16)[name = tensor("reshape_97_cast_fp16")]; + tensor reshape_98_shape_0 = const()[name = tensor("reshape_98_shape_0"), val = tensor([-1])]; + tensor reshape_98_cast_fp16 = reshape(shape = reshape_98_shape_0, x = reshape_94_cast_fp16)[name = tensor("reshape_98_cast_fp16")]; + tensor scatter_19_mode_0 = const()[name = tensor("scatter_19_mode_0"), val = tensor("update")]; + tensor scatter_19_axis_0 = const()[name = tensor("scatter_19_axis_0"), val = tensor(0)]; + tensor scatter_19_validate_indices_0 = const()[name = tensor("scatter_19_validate_indices_0"), val = tensor(false)]; + tensor scatter_19_cast_fp16 = scatter(axis = scatter_19_axis_0, data = reshape_98_cast_fp16, indices = reshape_96, mode = scatter_19_mode_0, updates = reshape_97_cast_fp16, validate_indices = scatter_19_validate_indices_0)[name = tensor("scatter_19_cast_fp16")]; + tensor reshape_99_cast_fp16 = reshape(shape = shape_67, x = scatter_19_cast_fp16)[name = tensor("reshape_99_cast_fp16")]; + tensor reshape_99_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("reshape_99_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor keys_55_begin_0 = const()[name = tensor("keys_55_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor keys_55_end_0 = const()[name = tensor("keys_55_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor keys_55_end_mask_0 = const()[name = tensor("keys_55_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor keys_55_squeeze_mask_0 = const()[name = tensor("keys_55_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor keys_55_cast_fp16 = slice_by_index(begin = keys_55_begin_0, end = keys_55_end_0, end_mask = keys_55_end_mask_0, squeeze_mask = keys_55_squeeze_mask_0, x = reshape_99_cast_fp16)[name = tensor("keys_55_cast_fp16")]; + tensor values_55_begin_0 = const()[name = tensor("values_55_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor values_55_end_0 = const()[name = tensor("values_55_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor values_55_end_mask_0 = const()[name = tensor("values_55_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor values_55_squeeze_mask_0 = const()[name = tensor("values_55_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor values_55_cast_fp16 = slice_by_index(begin = values_55_begin_0, end = values_55_end_0, end_mask = values_55_end_mask_0, squeeze_mask = values_55_squeeze_mask_0, x = reshape_99_cast_fp16)[name = tensor("values_55_cast_fp16")]; + tensor var_3931_cast_fp16 = not_equal(x = keys_55_cast_fp16, y = keys_55_cast_fp16)[name = tensor("op_3931_cast_fp16")]; + tensor keys_57_cast_fp16 = select(a = var_463_to_fp16, b = keys_55_cast_fp16, cond = var_3931_cast_fp16)[name = tensor("keys_57_cast_fp16")]; + tensor var_3939_cast_fp16 = not_equal(x = values_55_cast_fp16, y = values_55_cast_fp16)[name = tensor("op_3939_cast_fp16")]; + tensor values_57_cast_fp16 = select(a = var_463_to_fp16, b = values_55_cast_fp16, cond = var_3939_cast_fp16)[name = tensor("values_57_cast_fp16")]; + tensor var_3963 = const()[name = tensor("op_3963"), val = tensor([0, 2, 1, 3])]; + tensor var_3979 = const()[name = tensor("op_3979"), val = tensor([1, 1, 1])]; + tensor var_3980_cast_fp16 = reshape(shape = var_3979, x = position9_to_fp16)[name = tensor("op_3980_cast_fp16")]; + tensor var_4000_to_fp16 = const()[name = tensor("op_4000_to_fp16"), val = tensor(0x1p+0)]; + tensor valid_len_19_cast_fp16 = add(x = var_3980_cast_fp16, y = var_4000_to_fp16)[name = tensor("valid_len_19_cast_fp16")]; + tensor valid_mask_19_cast_fp16 = less(x = k_positions_1_promoted_to_fp16, y = valid_len_19_cast_fp16)[name = tensor("valid_mask_19_cast_fp16")]; + tensor causal_mask_19_cast_fp16 = less_equal(x = k_positions_1_promoted_to_fp16, y = var_3980_cast_fp16)[name = tensor("causal_mask_19_cast_fp16")]; + tensor attn_mask_37 = logical_and(x = valid_mask_19_cast_fp16, y = causal_mask_19_cast_fp16)[name = tensor("attn_mask_37")]; + tensor attn_mask_39_axes_0 = const()[name = tensor("attn_mask_39_axes_0"), val = tensor([1])]; + tensor attn_mask_39 = expand_dims(axes = attn_mask_39_axes_0, x = attn_mask_37)[name = tensor("attn_mask_39")]; + tensor var_4018_transpose_x_0 = const()[name = tensor("op_4018_transpose_x_0"), val = tensor(false)]; + tensor var_4018_transpose_y_0 = const()[name = tensor("op_4018_transpose_y_0"), val = tensor(false)]; + tensor transpose_90_perm_0 = const()[name = tensor("transpose_90_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_91_perm_0 = const()[name = tensor("transpose_91_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_91 = transpose(perm = transpose_91_perm_0, x = keys_57_cast_fp16)[name = tensor("transpose_177")]; + tensor transpose_90 = transpose(perm = transpose_90_perm_0, x = q_57_cast_fp16)[name = tensor("transpose_178")]; + tensor var_4018_cast_fp16 = matmul(transpose_x = var_4018_transpose_x_0, transpose_y = var_4018_transpose_y_0, x = transpose_90, y = transpose_91)[name = tensor("op_4018_cast_fp16")]; + tensor var_4012_to_fp16 = const()[name = tensor("op_4012_to_fp16"), val = tensor([0x1p-3])]; + tensor attn_weights_55_cast_fp16 = mul(x = var_4018_cast_fp16, y = var_4012_to_fp16)[name = tensor("attn_weights_55_cast_fp16")]; + tensor var_4020 = logical_not(x = attn_mask_39)[name = tensor("op_4020")]; + tensor var_4021_to_fp16 = const()[name = tensor("op_4021_to_fp16"), val = tensor(-inf)]; + tensor attn_weights_57_cast_fp16 = select(a = var_4021_to_fp16, b = attn_weights_55_cast_fp16, cond = var_4020)[name = tensor("attn_weights_57_cast_fp16")]; + tensor var_4023 = const()[name = tensor("op_4023"), val = tensor(-1)]; + tensor attn_weights_59_cast_fp16 = softmax(axis = var_4023, x = attn_weights_57_cast_fp16)[name = tensor("attn_weights_59_cast_fp16")]; + tensor attn_output_19_transpose_x_0 = const()[name = tensor("attn_output_19_transpose_x_0"), val = tensor(false)]; + tensor attn_output_19_transpose_y_0 = const()[name = tensor("attn_output_19_transpose_y_0"), val = tensor(false)]; + tensor values_59_cast_fp16 = transpose(perm = var_3963, x = values_57_cast_fp16)[name = tensor("transpose_179")]; + tensor attn_output_19_cast_fp16 = matmul(transpose_x = attn_output_19_transpose_x_0, transpose_y = attn_output_19_transpose_y_0, x = attn_weights_59_cast_fp16, y = values_59_cast_fp16)[name = tensor("attn_output_19_cast_fp16")]; + tensor var_4031 = const()[name = tensor("op_4031"), val = tensor([0, 2, 1, 3])]; + tensor var_4036 = const()[name = tensor("op_4036"), val = tensor([1, 1, 1024])]; + tensor var_4032_cast_fp16 = transpose(perm = var_4031, x = attn_output_19_cast_fp16)[name = tensor("transpose_176")]; + tensor input_115_cast_fp16 = reshape(shape = var_4036, x = var_4032_cast_fp16)[name = tensor("input_115_cast_fp16")]; + tensor attn9_out_proj_weight_to_fp16 = const()[name = tensor("attn9_out_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(233999744)))]; + tensor linear_38_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = attn9_out_proj_weight_to_fp16, x = input_115_cast_fp16)[name = tensor("linear_38_cast_fp16")]; + tensor var_4042_to_fp16 = const()[name = tensor("op_4042_to_fp16"), val = tensor(0x1p+0)]; + tensor var_4043_cast_fp16 = add(x = position9_to_fp16, y = var_4042_to_fp16)[name = tensor("op_4043_cast_fp16")]; + tensor var_4043_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_4043_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor input_117_cast_fp16 = add(x = input_111_cast_fp16, y = linear_38_cast_fp16)[name = tensor("input_117_cast_fp16")]; + tensor input_119_axes_0 = const()[name = tensor("input_119_axes_0"), val = tensor([-1])]; + tensor norm9_2_weight_to_fp16 = const()[name = tensor("norm9_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(236096960)))]; + tensor norm9_2_bias_to_fp16 = const()[name = tensor("norm9_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(236099072)))]; + tensor var_4047_to_fp16 = const()[name = tensor("op_4047_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_119_cast_fp16 = layer_norm(axes = input_119_axes_0, beta = norm9_2_bias_to_fp16, epsilon = var_4047_to_fp16, gamma = norm9_2_weight_to_fp16, x = input_117_cast_fp16)[name = tensor("input_119_cast_fp16")]; + tensor linear9_1_weight_to_fp16 = const()[name = tensor("linear9_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(236101184)))]; + tensor linear_39_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = linear9_1_weight_to_fp16, x = input_119_cast_fp16)[name = tensor("linear_39_cast_fp16")]; + tensor input_121_mode_0 = const()[name = tensor("input_121_mode_0"), val = tensor("EXACT")]; + tensor input_121_cast_fp16 = gelu(mode = input_121_mode_0, x = linear_39_cast_fp16)[name = tensor("input_121_cast_fp16")]; + tensor linear9_2_weight_to_fp16 = const()[name = tensor("linear9_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(244489856)))]; + tensor linear_40_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = linear9_2_weight_to_fp16, x = input_121_cast_fp16)[name = tensor("linear_40_cast_fp16")]; + tensor input_123_cast_fp16 = add(x = input_117_cast_fp16, y = linear_40_cast_fp16)[name = tensor("input_123_cast_fp16")]; + tensor input_125_axes_0 = const()[name = tensor("input_125_axes_0"), val = tensor([-1])]; + tensor norm10_1_weight_to_fp16 = const()[name = tensor("norm10_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(252878528)))]; + tensor norm10_1_bias_to_fp16 = const()[name = tensor("norm10_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(252880640)))]; + tensor var_4064_to_fp16 = const()[name = tensor("op_4064_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_125_cast_fp16 = layer_norm(axes = input_125_axes_0, beta = norm10_1_bias_to_fp16, epsilon = var_4064_to_fp16, gamma = norm10_1_weight_to_fp16, x = input_123_cast_fp16)[name = tensor("input_125_cast_fp16")]; + tensor attn10_in_proj_weight_to_fp16 = const()[name = tensor("attn10_in_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(252882752)))]; + tensor linear_41_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = attn10_in_proj_weight_to_fp16, x = input_125_cast_fp16)[name = tensor("linear_41_cast_fp16")]; + tensor var_4078 = const()[name = tensor("op_4078"), val = tensor([1, 1, 3, 16, 64])]; + tensor qkv_21_cast_fp16 = reshape(shape = var_4078, x = linear_41_cast_fp16)[name = tensor("qkv_21_cast_fp16")]; + tensor q_61_begin_0 = const()[name = tensor("q_61_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor q_61_end_0 = const()[name = tensor("q_61_end_0"), val = tensor([1, 1, 1, 16, 64])]; + tensor q_61_end_mask_0 = const()[name = tensor("q_61_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor q_61_squeeze_mask_0 = const()[name = tensor("q_61_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor q_61_cast_fp16 = slice_by_index(begin = q_61_begin_0, end = q_61_end_0, end_mask = q_61_end_mask_0, squeeze_mask = q_61_squeeze_mask_0, x = qkv_21_cast_fp16)[name = tensor("q_61_cast_fp16")]; + tensor k_41_begin_0 = const()[name = tensor("k_41_begin_0"), val = tensor([0, 0, 1, 0, 0])]; + tensor k_41_end_0 = const()[name = tensor("k_41_end_0"), val = tensor([1, 1, 2, 16, 64])]; + tensor k_41_end_mask_0 = const()[name = tensor("k_41_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor k_41_squeeze_mask_0 = const()[name = tensor("k_41_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor k_41_cast_fp16 = slice_by_index(begin = k_41_begin_0, end = k_41_end_0, end_mask = k_41_end_mask_0, squeeze_mask = k_41_squeeze_mask_0, x = qkv_21_cast_fp16)[name = tensor("k_41_cast_fp16")]; + tensor v_21_begin_0 = const()[name = tensor("v_21_begin_0"), val = tensor([0, 0, 2, 0, 0])]; + tensor v_21_end_0 = const()[name = tensor("v_21_end_0"), val = tensor([1, 1, 3, 16, 64])]; + tensor v_21_end_mask_0 = const()[name = tensor("v_21_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor v_21_squeeze_mask_0 = const()[name = tensor("v_21_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor v_21_cast_fp16 = slice_by_index(begin = v_21_begin_0, end = v_21_end_0, end_mask = v_21_end_mask_0, squeeze_mask = v_21_squeeze_mask_0, x = qkv_21_cast_fp16)[name = tensor("v_21_cast_fp16")]; + tensor var_4144 = const()[name = tensor("op_4144"), val = tensor([1, 1, 1, 1])]; + tensor position10_to_fp16_dtype_0 = const()[name = tensor("position10_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor position10_to_fp16 = cast(dtype = position10_to_fp16_dtype_0, x = position10)[name = tensor("cast_503")]; + tensor ts_65_cast_fp16 = reshape(shape = var_4144, x = position10_to_fp16)[name = tensor("ts_65_cast_fp16")]; + tensor var_4151 = const()[name = tensor("op_4151"), val = tensor([1, 1, 16, 32, 2])]; + tensor q_complex_21_cast_fp16 = reshape(shape = var_4151, x = q_61_cast_fp16)[name = tensor("q_complex_21_cast_fp16")]; + tensor var_4158 = const()[name = tensor("op_4158"), val = tensor([1, 1, 16, 32, 2])]; + tensor k_complex_21_cast_fp16 = reshape(shape = var_4158, x = k_41_cast_fp16)[name = tensor("k_complex_21_cast_fp16")]; + tensor var_4162_begin_0 = const()[name = tensor("op_4162_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_4162_end_0 = const()[name = tensor("op_4162_end_0"), val = tensor([1, 1, 16, 32, 1])]; + tensor var_4162_end_mask_0 = const()[name = tensor("op_4162_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_4162_squeeze_mask_0 = const()[name = tensor("op_4162_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_4162_cast_fp16 = slice_by_index(begin = var_4162_begin_0, end = var_4162_end_0, end_mask = var_4162_end_mask_0, squeeze_mask = var_4162_squeeze_mask_0, x = q_complex_21_cast_fp16)[name = tensor("op_4162_cast_fp16")]; + tensor var_4170_begin_0 = const()[name = tensor("op_4170_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_4170_end_0 = const()[name = tensor("op_4170_end_0"), val = tensor([1, 1, 16, 32, 2])]; + tensor var_4170_end_mask_0 = const()[name = tensor("op_4170_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_4170_squeeze_mask_0 = const()[name = tensor("op_4170_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_4170_cast_fp16 = slice_by_index(begin = var_4170_begin_0, end = var_4170_end_0, end_mask = var_4170_end_mask_0, squeeze_mask = var_4170_squeeze_mask_0, x = q_complex_21_cast_fp16)[name = tensor("op_4170_cast_fp16")]; + tensor var_4178_begin_0 = const()[name = tensor("op_4178_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_4178_end_0 = const()[name = tensor("op_4178_end_0"), val = tensor([1, 1, 16, 32, 1])]; + tensor var_4178_end_mask_0 = const()[name = tensor("op_4178_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_4178_squeeze_mask_0 = const()[name = tensor("op_4178_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_4178_cast_fp16 = slice_by_index(begin = var_4178_begin_0, end = var_4178_end_0, end_mask = var_4178_end_mask_0, squeeze_mask = var_4178_squeeze_mask_0, x = k_complex_21_cast_fp16)[name = tensor("op_4178_cast_fp16")]; + tensor var_4186_begin_0 = const()[name = tensor("op_4186_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_4186_end_0 = const()[name = tensor("op_4186_end_0"), val = tensor([1, 1, 16, 32, 2])]; + tensor var_4186_end_mask_0 = const()[name = tensor("op_4186_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_4186_squeeze_mask_0 = const()[name = tensor("op_4186_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_4186_cast_fp16 = slice_by_index(begin = var_4186_begin_0, end = var_4186_end_0, end_mask = var_4186_end_mask_0, squeeze_mask = var_4186_squeeze_mask_0, x = k_complex_21_cast_fp16)[name = tensor("op_4186_cast_fp16")]; + tensor freqs_21_to_fp16 = const()[name = tensor("freqs_21_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(259174272)))]; + tensor var_4192_cast_fp16 = mul(x = freqs_21_to_fp16, y = ts_65_cast_fp16)[name = tensor("op_4192_cast_fp16")]; + tensor rotr_21_cast_fp16 = cos(x = var_4192_cast_fp16)[name = tensor("rotr_21_cast_fp16")]; + tensor roti_21_cast_fp16 = sin(x = var_4192_cast_fp16)[name = tensor("roti_21_cast_fp16")]; + tensor var_4196_cast_fp16 = mul(x = var_4162_cast_fp16, y = rotr_21_cast_fp16)[name = tensor("op_4196_cast_fp16")]; + tensor var_4197_cast_fp16 = mul(x = var_4170_cast_fp16, y = roti_21_cast_fp16)[name = tensor("op_4197_cast_fp16")]; + tensor qor_41_cast_fp16 = sub(x = var_4196_cast_fp16, y = var_4197_cast_fp16)[name = tensor("qor_41_cast_fp16")]; + tensor var_4200_cast_fp16 = mul(x = var_4162_cast_fp16, y = roti_21_cast_fp16)[name = tensor("op_4200_cast_fp16")]; + tensor var_4201_cast_fp16 = mul(x = var_4170_cast_fp16, y = rotr_21_cast_fp16)[name = tensor("op_4201_cast_fp16")]; + tensor qoi_41_cast_fp16 = add(x = var_4200_cast_fp16, y = var_4201_cast_fp16)[name = tensor("qoi_41_cast_fp16")]; + tensor var_4204_cast_fp16 = mul(x = var_4178_cast_fp16, y = rotr_21_cast_fp16)[name = tensor("op_4204_cast_fp16")]; + tensor var_4205_cast_fp16 = mul(x = var_4186_cast_fp16, y = roti_21_cast_fp16)[name = tensor("op_4205_cast_fp16")]; + tensor kor_41_cast_fp16 = sub(x = var_4204_cast_fp16, y = var_4205_cast_fp16)[name = tensor("kor_41_cast_fp16")]; + tensor var_4208_cast_fp16 = mul(x = var_4178_cast_fp16, y = roti_21_cast_fp16)[name = tensor("op_4208_cast_fp16")]; + tensor var_4209_cast_fp16 = mul(x = var_4186_cast_fp16, y = rotr_21_cast_fp16)[name = tensor("op_4209_cast_fp16")]; + tensor koi_41_cast_fp16 = add(x = var_4208_cast_fp16, y = var_4209_cast_fp16)[name = tensor("koi_41_cast_fp16")]; + tensor qo_21_axis_0 = const()[name = tensor("qo_21_axis_0"), val = tensor(-1)]; + tensor qo_21_cast_fp16 = stack(axis = qo_21_axis_0, values = (qor_41_cast_fp16, qoi_41_cast_fp16))[name = tensor("qo_21_cast_fp16")]; + tensor ko_21_axis_0 = const()[name = tensor("ko_21_axis_0"), val = tensor(-1)]; + tensor ko_21_cast_fp16 = stack(axis = ko_21_axis_0, values = (kor_41_cast_fp16, koi_41_cast_fp16))[name = tensor("ko_21_cast_fp16")]; + tensor var_4242 = const()[name = tensor("op_4242"), val = tensor([1, 1, 16, 64])]; + tensor q_63_cast_fp16 = reshape(shape = var_4242, x = qo_21_cast_fp16)[name = tensor("q_63_cast_fp16")]; + tensor var_4248 = const()[name = tensor("op_4248"), val = tensor([1, 1, 16, 64])]; + tensor k_43_cast_fp16 = reshape(shape = var_4248, x = ko_21_cast_fp16)[name = tensor("k_43_cast_fp16")]; + tensor _inversed_4273_y_0_to_fp16 = const()[name = tensor("_inversed_4273_y_0_to_fp16"), val = tensor(0x1p-9)]; + tensor _inversed_4273_cast_fp16 = mul(x = ts_65_cast_fp16, y = _inversed_4273_y_0_to_fp16)[name = tensor("_inversed_4273_cast_fp16")]; + tensor var_4274_cast_fp16 = floor(x = _inversed_4273_cast_fp16)[name = tensor("op_4274_cast_fp16")]; + tensor var_4275_to_fp16 = const()[name = tensor("op_4275_to_fp16"), val = tensor(0x1p+9)]; + tensor var_4276_cast_fp16 = mul(x = var_4274_cast_fp16, y = var_4275_to_fp16)[name = tensor("op_4276_cast_fp16")]; + tensor write_indices_float_43_cast_fp16 = sub(x = ts_65_cast_fp16, y = var_4276_cast_fp16)[name = tensor("write_indices_float_43_cast_fp16")]; + tensor var_4283_dtype_0 = const()[name = tensor("op_4283_dtype_0"), val = tensor("int32")]; + tensor write_indices_21_reps_0 = const()[name = tensor("write_indices_21_reps_0"), val = tensor([1, 1, 16, 64])]; + tensor write_indices_float_43_cast_fp16_to_int32 = cast(dtype = var_4283_dtype_0, x = write_indices_float_43_cast_fp16)[name = tensor("cast_502")]; + tensor write_indices_21 = tile(reps = write_indices_21_reps_0, x = write_indices_float_43_cast_fp16_to_int32)[name = tensor("write_indices_21")]; + tensor var_4293_begin_0 = const()[name = tensor("op_4293_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_4293_end_0 = const()[name = tensor("op_4293_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor var_4293_end_mask_0 = const()[name = tensor("op_4293_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_4293_squeeze_mask_0 = const()[name = tensor("op_4293_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor cache10_to_fp16_dtype_0 = const()[name = tensor("cache10_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor cache10_to_fp16 = cast(dtype = cache10_to_fp16_dtype_0, x = cache10)[name = tensor("cast_501")]; + tensor var_4293_cast_fp16 = slice_by_index(begin = var_4293_begin_0, end = var_4293_end_0, end_mask = var_4293_end_mask_0, squeeze_mask = var_4293_squeeze_mask_0, x = cache10_to_fp16)[name = tensor("op_4293_cast_fp16")]; + tensor var_4295_axis_0 = const()[name = tensor("op_4295_axis_0"), val = tensor(1)]; + tensor var_4295_mode_0 = const()[name = tensor("op_4295_mode_0"), val = tensor("update")]; + tensor var_4295_validate_indices_0 = const()[name = tensor("op_4295_validate_indices_0"), val = tensor(false)]; + tensor var_4295_cast_fp16 = scatter_along_axis(axis = var_4295_axis_0, data = var_4293_cast_fp16, indices = write_indices_21, mode = var_4295_mode_0, updates = k_43_cast_fp16, validate_indices = var_4295_validate_indices_0)[name = tensor("op_4295_cast_fp16")]; + tensor concat_72 = const()[name = tensor("concat_72"), val = tensor([0, 0, 0, 0, 0])]; + tensor concat_73 = const()[name = tensor("concat_73"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_21_internal_tensor_assign_1_stride_0 = const()[name = tensor("new_cache_21_internal_tensor_assign_1_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_21_internal_tensor_assign_1_begin_mask_0 = const()[name = tensor("new_cache_21_internal_tensor_assign_1_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_21_internal_tensor_assign_1_end_mask_0 = const()[name = tensor("new_cache_21_internal_tensor_assign_1_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_21_internal_tensor_assign_1_squeeze_mask_0 = const()[name = tensor("new_cache_21_internal_tensor_assign_1_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_68 = const()[name = tensor("shape_68"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_20 = const()[name = tensor("reduce_prod_20"), val = tensor(1048576)]; + tensor range_1d_20_start_0 = const()[name = tensor("range_1d_20_start_0"), val = tensor(0)]; + tensor range_1d_20_step_0 = const()[name = tensor("range_1d_20_step_0"), val = tensor(1)]; + tensor range_1d_20 = range_1d(end = reduce_prod_20, start = range_1d_20_start_0, step = range_1d_20_step_0)[name = tensor("range_1d_20")]; + tensor reshape_100 = reshape(shape = shape_68, x = range_1d_20)[name = tensor("reshape_100")]; + tensor slice_by_index_20 = slice_by_index(begin = concat_72, begin_mask = new_cache_21_internal_tensor_assign_1_begin_mask_0, end = concat_73, end_mask = new_cache_21_internal_tensor_assign_1_end_mask_0, squeeze_mask = new_cache_21_internal_tensor_assign_1_squeeze_mask_0, stride = new_cache_21_internal_tensor_assign_1_stride_0, x = reshape_100)[name = tensor("slice_by_index_20")]; + tensor reshape_101_shape_0 = const()[name = tensor("reshape_101_shape_0"), val = tensor([-1])]; + tensor reshape_101 = reshape(shape = reshape_101_shape_0, x = slice_by_index_20)[name = tensor("reshape_101")]; + tensor reshape_102_shape_0 = const()[name = tensor("reshape_102_shape_0"), val = tensor([-1])]; + tensor reshape_102_cast_fp16 = reshape(shape = reshape_102_shape_0, x = var_4295_cast_fp16)[name = tensor("reshape_102_cast_fp16")]; + tensor reshape_103_shape_0 = const()[name = tensor("reshape_103_shape_0"), val = tensor([-1])]; + tensor reshape_103_cast_fp16 = reshape(shape = reshape_103_shape_0, x = cache10_to_fp16)[name = tensor("reshape_103_cast_fp16")]; + tensor scatter_20_mode_0 = const()[name = tensor("scatter_20_mode_0"), val = tensor("update")]; + tensor scatter_20_axis_0 = const()[name = tensor("scatter_20_axis_0"), val = tensor(0)]; + tensor scatter_20_validate_indices_0 = const()[name = tensor("scatter_20_validate_indices_0"), val = tensor(false)]; + tensor scatter_20_cast_fp16 = scatter(axis = scatter_20_axis_0, data = reshape_103_cast_fp16, indices = reshape_101, mode = scatter_20_mode_0, updates = reshape_102_cast_fp16, validate_indices = scatter_20_validate_indices_0)[name = tensor("scatter_20_cast_fp16")]; + tensor reshape_104_cast_fp16 = reshape(shape = shape_68, x = scatter_20_cast_fp16)[name = tensor("reshape_104_cast_fp16")]; + tensor var_4303_begin_0 = const()[name = tensor("op_4303_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor var_4303_end_0 = const()[name = tensor("op_4303_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor var_4303_end_mask_0 = const()[name = tensor("op_4303_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_4303_squeeze_mask_0 = const()[name = tensor("op_4303_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor var_4303_cast_fp16 = slice_by_index(begin = var_4303_begin_0, end = var_4303_end_0, end_mask = var_4303_end_mask_0, squeeze_mask = var_4303_squeeze_mask_0, x = reshape_104_cast_fp16)[name = tensor("op_4303_cast_fp16")]; + tensor var_4305_axis_0 = const()[name = tensor("op_4305_axis_0"), val = tensor(1)]; + tensor var_4305_mode_0 = const()[name = tensor("op_4305_mode_0"), val = tensor("update")]; + tensor var_4305_validate_indices_0 = const()[name = tensor("op_4305_validate_indices_0"), val = tensor(false)]; + tensor var_4305_cast_fp16 = scatter_along_axis(axis = var_4305_axis_0, data = var_4303_cast_fp16, indices = write_indices_21, mode = var_4305_mode_0, updates = v_21_cast_fp16, validate_indices = var_4305_validate_indices_0)[name = tensor("op_4305_cast_fp16")]; + tensor concat_74 = const()[name = tensor("concat_74"), val = tensor([1, 0, 0, 0, 0])]; + tensor concat_75 = const()[name = tensor("concat_75"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_21_internal_tensor_assign_2_stride_0 = const()[name = tensor("new_cache_21_internal_tensor_assign_2_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_21_internal_tensor_assign_2_begin_mask_0 = const()[name = tensor("new_cache_21_internal_tensor_assign_2_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_21_internal_tensor_assign_2_end_mask_0 = const()[name = tensor("new_cache_21_internal_tensor_assign_2_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_21_internal_tensor_assign_2_squeeze_mask_0 = const()[name = tensor("new_cache_21_internal_tensor_assign_2_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_69 = const()[name = tensor("shape_69"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_21 = const()[name = tensor("reduce_prod_21"), val = tensor(1048576)]; + tensor range_1d_21_start_0 = const()[name = tensor("range_1d_21_start_0"), val = tensor(0)]; + tensor range_1d_21_step_0 = const()[name = tensor("range_1d_21_step_0"), val = tensor(1)]; + tensor range_1d_21 = range_1d(end = reduce_prod_21, start = range_1d_21_start_0, step = range_1d_21_step_0)[name = tensor("range_1d_21")]; + tensor reshape_105 = reshape(shape = shape_69, x = range_1d_21)[name = tensor("reshape_105")]; + tensor slice_by_index_21 = slice_by_index(begin = concat_74, begin_mask = new_cache_21_internal_tensor_assign_2_begin_mask_0, end = concat_75, end_mask = new_cache_21_internal_tensor_assign_2_end_mask_0, squeeze_mask = new_cache_21_internal_tensor_assign_2_squeeze_mask_0, stride = new_cache_21_internal_tensor_assign_2_stride_0, x = reshape_105)[name = tensor("slice_by_index_21")]; + tensor reshape_106_shape_0 = const()[name = tensor("reshape_106_shape_0"), val = tensor([-1])]; + tensor reshape_106 = reshape(shape = reshape_106_shape_0, x = slice_by_index_21)[name = tensor("reshape_106")]; + tensor reshape_107_shape_0 = const()[name = tensor("reshape_107_shape_0"), val = tensor([-1])]; + tensor reshape_107_cast_fp16 = reshape(shape = reshape_107_shape_0, x = var_4305_cast_fp16)[name = tensor("reshape_107_cast_fp16")]; + tensor reshape_108_shape_0 = const()[name = tensor("reshape_108_shape_0"), val = tensor([-1])]; + tensor reshape_108_cast_fp16 = reshape(shape = reshape_108_shape_0, x = reshape_104_cast_fp16)[name = tensor("reshape_108_cast_fp16")]; + tensor scatter_21_mode_0 = const()[name = tensor("scatter_21_mode_0"), val = tensor("update")]; + tensor scatter_21_axis_0 = const()[name = tensor("scatter_21_axis_0"), val = tensor(0)]; + tensor scatter_21_validate_indices_0 = const()[name = tensor("scatter_21_validate_indices_0"), val = tensor(false)]; + tensor scatter_21_cast_fp16 = scatter(axis = scatter_21_axis_0, data = reshape_108_cast_fp16, indices = reshape_106, mode = scatter_21_mode_0, updates = reshape_107_cast_fp16, validate_indices = scatter_21_validate_indices_0)[name = tensor("scatter_21_cast_fp16")]; + tensor reshape_109_cast_fp16 = reshape(shape = shape_69, x = scatter_21_cast_fp16)[name = tensor("reshape_109_cast_fp16")]; + tensor reshape_109_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("reshape_109_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor keys_61_begin_0 = const()[name = tensor("keys_61_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor keys_61_end_0 = const()[name = tensor("keys_61_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor keys_61_end_mask_0 = const()[name = tensor("keys_61_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor keys_61_squeeze_mask_0 = const()[name = tensor("keys_61_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor keys_61_cast_fp16 = slice_by_index(begin = keys_61_begin_0, end = keys_61_end_0, end_mask = keys_61_end_mask_0, squeeze_mask = keys_61_squeeze_mask_0, x = reshape_109_cast_fp16)[name = tensor("keys_61_cast_fp16")]; + tensor values_61_begin_0 = const()[name = tensor("values_61_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor values_61_end_0 = const()[name = tensor("values_61_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor values_61_end_mask_0 = const()[name = tensor("values_61_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor values_61_squeeze_mask_0 = const()[name = tensor("values_61_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor values_61_cast_fp16 = slice_by_index(begin = values_61_begin_0, end = values_61_end_0, end_mask = values_61_end_mask_0, squeeze_mask = values_61_squeeze_mask_0, x = reshape_109_cast_fp16)[name = tensor("values_61_cast_fp16")]; + tensor var_4317_cast_fp16 = not_equal(x = keys_61_cast_fp16, y = keys_61_cast_fp16)[name = tensor("op_4317_cast_fp16")]; + tensor keys_63_cast_fp16 = select(a = var_463_to_fp16, b = keys_61_cast_fp16, cond = var_4317_cast_fp16)[name = tensor("keys_63_cast_fp16")]; + tensor var_4325_cast_fp16 = not_equal(x = values_61_cast_fp16, y = values_61_cast_fp16)[name = tensor("op_4325_cast_fp16")]; + tensor values_63_cast_fp16 = select(a = var_463_to_fp16, b = values_61_cast_fp16, cond = var_4325_cast_fp16)[name = tensor("values_63_cast_fp16")]; + tensor var_4349 = const()[name = tensor("op_4349"), val = tensor([0, 2, 1, 3])]; + tensor var_4365 = const()[name = tensor("op_4365"), val = tensor([1, 1, 1])]; + tensor var_4366_cast_fp16 = reshape(shape = var_4365, x = position10_to_fp16)[name = tensor("op_4366_cast_fp16")]; + tensor var_4386_to_fp16 = const()[name = tensor("op_4386_to_fp16"), val = tensor(0x1p+0)]; + tensor valid_len_21_cast_fp16 = add(x = var_4366_cast_fp16, y = var_4386_to_fp16)[name = tensor("valid_len_21_cast_fp16")]; + tensor valid_mask_21_cast_fp16 = less(x = k_positions_1_promoted_to_fp16, y = valid_len_21_cast_fp16)[name = tensor("valid_mask_21_cast_fp16")]; + tensor causal_mask_21_cast_fp16 = less_equal(x = k_positions_1_promoted_to_fp16, y = var_4366_cast_fp16)[name = tensor("causal_mask_21_cast_fp16")]; + tensor attn_mask_41 = logical_and(x = valid_mask_21_cast_fp16, y = causal_mask_21_cast_fp16)[name = tensor("attn_mask_41")]; + tensor attn_mask_43_axes_0 = const()[name = tensor("attn_mask_43_axes_0"), val = tensor([1])]; + tensor attn_mask_43 = expand_dims(axes = attn_mask_43_axes_0, x = attn_mask_41)[name = tensor("attn_mask_43")]; + tensor var_4404_transpose_x_0 = const()[name = tensor("op_4404_transpose_x_0"), val = tensor(false)]; + tensor var_4404_transpose_y_0 = const()[name = tensor("op_4404_transpose_y_0"), val = tensor(false)]; + tensor transpose_92_perm_0 = const()[name = tensor("transpose_92_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_93_perm_0 = const()[name = tensor("transpose_93_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_93 = transpose(perm = transpose_93_perm_0, x = keys_63_cast_fp16)[name = tensor("transpose_173")]; + tensor transpose_92 = transpose(perm = transpose_92_perm_0, x = q_63_cast_fp16)[name = tensor("transpose_174")]; + tensor var_4404_cast_fp16 = matmul(transpose_x = var_4404_transpose_x_0, transpose_y = var_4404_transpose_y_0, x = transpose_92, y = transpose_93)[name = tensor("op_4404_cast_fp16")]; + tensor var_4398_to_fp16 = const()[name = tensor("op_4398_to_fp16"), val = tensor([0x1p-3])]; + tensor attn_weights_61_cast_fp16 = mul(x = var_4404_cast_fp16, y = var_4398_to_fp16)[name = tensor("attn_weights_61_cast_fp16")]; + tensor var_4406 = logical_not(x = attn_mask_43)[name = tensor("op_4406")]; + tensor var_4407_to_fp16 = const()[name = tensor("op_4407_to_fp16"), val = tensor(-inf)]; + tensor attn_weights_63_cast_fp16 = select(a = var_4407_to_fp16, b = attn_weights_61_cast_fp16, cond = var_4406)[name = tensor("attn_weights_63_cast_fp16")]; + tensor var_4409 = const()[name = tensor("op_4409"), val = tensor(-1)]; + tensor attn_weights_65_cast_fp16 = softmax(axis = var_4409, x = attn_weights_63_cast_fp16)[name = tensor("attn_weights_65_cast_fp16")]; + tensor attn_output_21_transpose_x_0 = const()[name = tensor("attn_output_21_transpose_x_0"), val = tensor(false)]; + tensor attn_output_21_transpose_y_0 = const()[name = tensor("attn_output_21_transpose_y_0"), val = tensor(false)]; + tensor values_65_cast_fp16 = transpose(perm = var_4349, x = values_63_cast_fp16)[name = tensor("transpose_175")]; + tensor attn_output_21_cast_fp16 = matmul(transpose_x = attn_output_21_transpose_x_0, transpose_y = attn_output_21_transpose_y_0, x = attn_weights_65_cast_fp16, y = values_65_cast_fp16)[name = tensor("attn_output_21_cast_fp16")]; + tensor var_4417 = const()[name = tensor("op_4417"), val = tensor([0, 2, 1, 3])]; + tensor var_4422 = const()[name = tensor("op_4422"), val = tensor([1, 1, 1024])]; + tensor var_4418_cast_fp16 = transpose(perm = var_4417, x = attn_output_21_cast_fp16)[name = tensor("transpose_172")]; + tensor input_127_cast_fp16 = reshape(shape = var_4422, x = var_4418_cast_fp16)[name = tensor("input_127_cast_fp16")]; + tensor attn10_out_proj_weight_to_fp16 = const()[name = tensor("attn10_out_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(259174400)))]; + tensor linear_42_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = attn10_out_proj_weight_to_fp16, x = input_127_cast_fp16)[name = tensor("linear_42_cast_fp16")]; + tensor var_4428_to_fp16 = const()[name = tensor("op_4428_to_fp16"), val = tensor(0x1p+0)]; + tensor var_4429_cast_fp16 = add(x = position10_to_fp16, y = var_4428_to_fp16)[name = tensor("op_4429_cast_fp16")]; + tensor var_4429_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_4429_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor input_129_cast_fp16 = add(x = input_123_cast_fp16, y = linear_42_cast_fp16)[name = tensor("input_129_cast_fp16")]; + tensor input_131_axes_0 = const()[name = tensor("input_131_axes_0"), val = tensor([-1])]; + tensor norm10_2_weight_to_fp16 = const()[name = tensor("norm10_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(261271616)))]; + tensor norm10_2_bias_to_fp16 = const()[name = tensor("norm10_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(261273728)))]; + tensor var_4433_to_fp16 = const()[name = tensor("op_4433_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_131_cast_fp16 = layer_norm(axes = input_131_axes_0, beta = norm10_2_bias_to_fp16, epsilon = var_4433_to_fp16, gamma = norm10_2_weight_to_fp16, x = input_129_cast_fp16)[name = tensor("input_131_cast_fp16")]; + tensor linear10_1_weight_to_fp16 = const()[name = tensor("linear10_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(261275840)))]; + tensor linear_43_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = linear10_1_weight_to_fp16, x = input_131_cast_fp16)[name = tensor("linear_43_cast_fp16")]; + tensor input_133_mode_0 = const()[name = tensor("input_133_mode_0"), val = tensor("EXACT")]; + tensor input_133_cast_fp16 = gelu(mode = input_133_mode_0, x = linear_43_cast_fp16)[name = tensor("input_133_cast_fp16")]; + tensor linear10_2_weight_to_fp16 = const()[name = tensor("linear10_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(269664512)))]; + tensor linear_44_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = linear10_2_weight_to_fp16, x = input_133_cast_fp16)[name = tensor("linear_44_cast_fp16")]; + tensor input_135_cast_fp16 = add(x = input_129_cast_fp16, y = linear_44_cast_fp16)[name = tensor("input_135_cast_fp16")]; + tensor input_137_axes_0 = const()[name = tensor("input_137_axes_0"), val = tensor([-1])]; + tensor norm11_1_weight_to_fp16 = const()[name = tensor("norm11_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(278053184)))]; + tensor norm11_1_bias_to_fp16 = const()[name = tensor("norm11_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(278055296)))]; + tensor var_4450_to_fp16 = const()[name = tensor("op_4450_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_137_cast_fp16 = layer_norm(axes = input_137_axes_0, beta = norm11_1_bias_to_fp16, epsilon = var_4450_to_fp16, gamma = norm11_1_weight_to_fp16, x = input_135_cast_fp16)[name = tensor("input_137_cast_fp16")]; + tensor attn11_in_proj_weight_to_fp16 = const()[name = tensor("attn11_in_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(278057408)))]; + tensor linear_45_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = attn11_in_proj_weight_to_fp16, x = input_137_cast_fp16)[name = tensor("linear_45_cast_fp16")]; + tensor var_4464 = const()[name = tensor("op_4464"), val = tensor([1, 1, 3, 16, 64])]; + tensor qkv_23_cast_fp16 = reshape(shape = var_4464, x = linear_45_cast_fp16)[name = tensor("qkv_23_cast_fp16")]; + tensor q_67_begin_0 = const()[name = tensor("q_67_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor q_67_end_0 = const()[name = tensor("q_67_end_0"), val = tensor([1, 1, 1, 16, 64])]; + tensor q_67_end_mask_0 = const()[name = tensor("q_67_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor q_67_squeeze_mask_0 = const()[name = tensor("q_67_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor q_67_cast_fp16 = slice_by_index(begin = q_67_begin_0, end = q_67_end_0, end_mask = q_67_end_mask_0, squeeze_mask = q_67_squeeze_mask_0, x = qkv_23_cast_fp16)[name = tensor("q_67_cast_fp16")]; + tensor k_45_begin_0 = const()[name = tensor("k_45_begin_0"), val = tensor([0, 0, 1, 0, 0])]; + tensor k_45_end_0 = const()[name = tensor("k_45_end_0"), val = tensor([1, 1, 2, 16, 64])]; + tensor k_45_end_mask_0 = const()[name = tensor("k_45_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor k_45_squeeze_mask_0 = const()[name = tensor("k_45_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor k_45_cast_fp16 = slice_by_index(begin = k_45_begin_0, end = k_45_end_0, end_mask = k_45_end_mask_0, squeeze_mask = k_45_squeeze_mask_0, x = qkv_23_cast_fp16)[name = tensor("k_45_cast_fp16")]; + tensor v_23_begin_0 = const()[name = tensor("v_23_begin_0"), val = tensor([0, 0, 2, 0, 0])]; + tensor v_23_end_0 = const()[name = tensor("v_23_end_0"), val = tensor([1, 1, 3, 16, 64])]; + tensor v_23_end_mask_0 = const()[name = tensor("v_23_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor v_23_squeeze_mask_0 = const()[name = tensor("v_23_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor v_23_cast_fp16 = slice_by_index(begin = v_23_begin_0, end = v_23_end_0, end_mask = v_23_end_mask_0, squeeze_mask = v_23_squeeze_mask_0, x = qkv_23_cast_fp16)[name = tensor("v_23_cast_fp16")]; + tensor var_4530 = const()[name = tensor("op_4530"), val = tensor([1, 1, 1, 1])]; + tensor position11_to_fp16_dtype_0 = const()[name = tensor("position11_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor position11_to_fp16 = cast(dtype = position11_to_fp16_dtype_0, x = position11)[name = tensor("cast_498")]; + tensor ts_71_cast_fp16 = reshape(shape = var_4530, x = position11_to_fp16)[name = tensor("ts_71_cast_fp16")]; + tensor var_4537 = const()[name = tensor("op_4537"), val = tensor([1, 1, 16, 32, 2])]; + tensor q_complex_23_cast_fp16 = reshape(shape = var_4537, x = q_67_cast_fp16)[name = tensor("q_complex_23_cast_fp16")]; + tensor var_4544 = const()[name = tensor("op_4544"), val = tensor([1, 1, 16, 32, 2])]; + tensor k_complex_23_cast_fp16 = reshape(shape = var_4544, x = k_45_cast_fp16)[name = tensor("k_complex_23_cast_fp16")]; + tensor var_4548_begin_0 = const()[name = tensor("op_4548_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_4548_end_0 = const()[name = tensor("op_4548_end_0"), val = tensor([1, 1, 16, 32, 1])]; + tensor var_4548_end_mask_0 = const()[name = tensor("op_4548_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_4548_squeeze_mask_0 = const()[name = tensor("op_4548_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_4548_cast_fp16 = slice_by_index(begin = var_4548_begin_0, end = var_4548_end_0, end_mask = var_4548_end_mask_0, squeeze_mask = var_4548_squeeze_mask_0, x = q_complex_23_cast_fp16)[name = tensor("op_4548_cast_fp16")]; + tensor var_4556_begin_0 = const()[name = tensor("op_4556_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_4556_end_0 = const()[name = tensor("op_4556_end_0"), val = tensor([1, 1, 16, 32, 2])]; + tensor var_4556_end_mask_0 = const()[name = tensor("op_4556_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_4556_squeeze_mask_0 = const()[name = tensor("op_4556_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_4556_cast_fp16 = slice_by_index(begin = var_4556_begin_0, end = var_4556_end_0, end_mask = var_4556_end_mask_0, squeeze_mask = var_4556_squeeze_mask_0, x = q_complex_23_cast_fp16)[name = tensor("op_4556_cast_fp16")]; + tensor var_4564_begin_0 = const()[name = tensor("op_4564_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_4564_end_0 = const()[name = tensor("op_4564_end_0"), val = tensor([1, 1, 16, 32, 1])]; + tensor var_4564_end_mask_0 = const()[name = tensor("op_4564_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_4564_squeeze_mask_0 = const()[name = tensor("op_4564_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_4564_cast_fp16 = slice_by_index(begin = var_4564_begin_0, end = var_4564_end_0, end_mask = var_4564_end_mask_0, squeeze_mask = var_4564_squeeze_mask_0, x = k_complex_23_cast_fp16)[name = tensor("op_4564_cast_fp16")]; + tensor var_4572_begin_0 = const()[name = tensor("op_4572_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_4572_end_0 = const()[name = tensor("op_4572_end_0"), val = tensor([1, 1, 16, 32, 2])]; + tensor var_4572_end_mask_0 = const()[name = tensor("op_4572_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_4572_squeeze_mask_0 = const()[name = tensor("op_4572_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_4572_cast_fp16 = slice_by_index(begin = var_4572_begin_0, end = var_4572_end_0, end_mask = var_4572_end_mask_0, squeeze_mask = var_4572_squeeze_mask_0, x = k_complex_23_cast_fp16)[name = tensor("op_4572_cast_fp16")]; + tensor freqs_23_to_fp16 = const()[name = tensor("freqs_23_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(284348928)))]; + tensor var_4578_cast_fp16 = mul(x = freqs_23_to_fp16, y = ts_71_cast_fp16)[name = tensor("op_4578_cast_fp16")]; + tensor rotr_23_cast_fp16 = cos(x = var_4578_cast_fp16)[name = tensor("rotr_23_cast_fp16")]; + tensor roti_23_cast_fp16 = sin(x = var_4578_cast_fp16)[name = tensor("roti_23_cast_fp16")]; + tensor var_4582_cast_fp16 = mul(x = var_4548_cast_fp16, y = rotr_23_cast_fp16)[name = tensor("op_4582_cast_fp16")]; + tensor var_4583_cast_fp16 = mul(x = var_4556_cast_fp16, y = roti_23_cast_fp16)[name = tensor("op_4583_cast_fp16")]; + tensor qor_45_cast_fp16 = sub(x = var_4582_cast_fp16, y = var_4583_cast_fp16)[name = tensor("qor_45_cast_fp16")]; + tensor var_4586_cast_fp16 = mul(x = var_4548_cast_fp16, y = roti_23_cast_fp16)[name = tensor("op_4586_cast_fp16")]; + tensor var_4587_cast_fp16 = mul(x = var_4556_cast_fp16, y = rotr_23_cast_fp16)[name = tensor("op_4587_cast_fp16")]; + tensor qoi_45_cast_fp16 = add(x = var_4586_cast_fp16, y = var_4587_cast_fp16)[name = tensor("qoi_45_cast_fp16")]; + tensor var_4590_cast_fp16 = mul(x = var_4564_cast_fp16, y = rotr_23_cast_fp16)[name = tensor("op_4590_cast_fp16")]; + tensor var_4591_cast_fp16 = mul(x = var_4572_cast_fp16, y = roti_23_cast_fp16)[name = tensor("op_4591_cast_fp16")]; + tensor kor_45_cast_fp16 = sub(x = var_4590_cast_fp16, y = var_4591_cast_fp16)[name = tensor("kor_45_cast_fp16")]; + tensor var_4594_cast_fp16 = mul(x = var_4564_cast_fp16, y = roti_23_cast_fp16)[name = tensor("op_4594_cast_fp16")]; + tensor var_4595_cast_fp16 = mul(x = var_4572_cast_fp16, y = rotr_23_cast_fp16)[name = tensor("op_4595_cast_fp16")]; + tensor koi_45_cast_fp16 = add(x = var_4594_cast_fp16, y = var_4595_cast_fp16)[name = tensor("koi_45_cast_fp16")]; + tensor qo_23_axis_0 = const()[name = tensor("qo_23_axis_0"), val = tensor(-1)]; + tensor qo_23_cast_fp16 = stack(axis = qo_23_axis_0, values = (qor_45_cast_fp16, qoi_45_cast_fp16))[name = tensor("qo_23_cast_fp16")]; + tensor ko_23_axis_0 = const()[name = tensor("ko_23_axis_0"), val = tensor(-1)]; + tensor ko_23_cast_fp16 = stack(axis = ko_23_axis_0, values = (kor_45_cast_fp16, koi_45_cast_fp16))[name = tensor("ko_23_cast_fp16")]; + tensor var_4628 = const()[name = tensor("op_4628"), val = tensor([1, 1, 16, 64])]; + tensor q_69_cast_fp16 = reshape(shape = var_4628, x = qo_23_cast_fp16)[name = tensor("q_69_cast_fp16")]; + tensor var_4634 = const()[name = tensor("op_4634"), val = tensor([1, 1, 16, 64])]; + tensor k_47_cast_fp16 = reshape(shape = var_4634, x = ko_23_cast_fp16)[name = tensor("k_47_cast_fp16")]; + tensor _inversed_4659_y_0_to_fp16 = const()[name = tensor("_inversed_4659_y_0_to_fp16"), val = tensor(0x1p-9)]; + tensor _inversed_4659_cast_fp16 = mul(x = ts_71_cast_fp16, y = _inversed_4659_y_0_to_fp16)[name = tensor("_inversed_4659_cast_fp16")]; + tensor var_4660_cast_fp16 = floor(x = _inversed_4659_cast_fp16)[name = tensor("op_4660_cast_fp16")]; + tensor var_4661_to_fp16 = const()[name = tensor("op_4661_to_fp16"), val = tensor(0x1p+9)]; + tensor var_4662_cast_fp16 = mul(x = var_4660_cast_fp16, y = var_4661_to_fp16)[name = tensor("op_4662_cast_fp16")]; + tensor write_indices_float_47_cast_fp16 = sub(x = ts_71_cast_fp16, y = var_4662_cast_fp16)[name = tensor("write_indices_float_47_cast_fp16")]; + tensor var_4669_dtype_0 = const()[name = tensor("op_4669_dtype_0"), val = tensor("int32")]; + tensor write_indices_23_reps_0 = const()[name = tensor("write_indices_23_reps_0"), val = tensor([1, 1, 16, 64])]; + tensor write_indices_float_47_cast_fp16_to_int32 = cast(dtype = var_4669_dtype_0, x = write_indices_float_47_cast_fp16)[name = tensor("cast_497")]; + tensor write_indices_23 = tile(reps = write_indices_23_reps_0, x = write_indices_float_47_cast_fp16_to_int32)[name = tensor("write_indices_23")]; + tensor var_4679_begin_0 = const()[name = tensor("op_4679_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_4679_end_0 = const()[name = tensor("op_4679_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor var_4679_end_mask_0 = const()[name = tensor("op_4679_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_4679_squeeze_mask_0 = const()[name = tensor("op_4679_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor cache11_to_fp16_dtype_0 = const()[name = tensor("cache11_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor cache11_to_fp16 = cast(dtype = cache11_to_fp16_dtype_0, x = cache11)[name = tensor("cast_496")]; + tensor var_4679_cast_fp16 = slice_by_index(begin = var_4679_begin_0, end = var_4679_end_0, end_mask = var_4679_end_mask_0, squeeze_mask = var_4679_squeeze_mask_0, x = cache11_to_fp16)[name = tensor("op_4679_cast_fp16")]; + tensor var_4681_axis_0 = const()[name = tensor("op_4681_axis_0"), val = tensor(1)]; + tensor var_4681_mode_0 = const()[name = tensor("op_4681_mode_0"), val = tensor("update")]; + tensor var_4681_validate_indices_0 = const()[name = tensor("op_4681_validate_indices_0"), val = tensor(false)]; + tensor var_4681_cast_fp16 = scatter_along_axis(axis = var_4681_axis_0, data = var_4679_cast_fp16, indices = write_indices_23, mode = var_4681_mode_0, updates = k_47_cast_fp16, validate_indices = var_4681_validate_indices_0)[name = tensor("op_4681_cast_fp16")]; + tensor concat_79 = const()[name = tensor("concat_79"), val = tensor([0, 0, 0, 0, 0])]; + tensor concat_80 = const()[name = tensor("concat_80"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_23_internal_tensor_assign_1_stride_0 = const()[name = tensor("new_cache_23_internal_tensor_assign_1_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_23_internal_tensor_assign_1_begin_mask_0 = const()[name = tensor("new_cache_23_internal_tensor_assign_1_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_23_internal_tensor_assign_1_end_mask_0 = const()[name = tensor("new_cache_23_internal_tensor_assign_1_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_23_internal_tensor_assign_1_squeeze_mask_0 = const()[name = tensor("new_cache_23_internal_tensor_assign_1_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_70 = const()[name = tensor("shape_70"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_22 = const()[name = tensor("reduce_prod_22"), val = tensor(1048576)]; + tensor range_1d_22_start_0 = const()[name = tensor("range_1d_22_start_0"), val = tensor(0)]; + tensor range_1d_22_step_0 = const()[name = tensor("range_1d_22_step_0"), val = tensor(1)]; + tensor range_1d_22 = range_1d(end = reduce_prod_22, start = range_1d_22_start_0, step = range_1d_22_step_0)[name = tensor("range_1d_22")]; + tensor reshape_110 = reshape(shape = shape_70, x = range_1d_22)[name = tensor("reshape_110")]; + tensor slice_by_index_22 = slice_by_index(begin = concat_79, begin_mask = new_cache_23_internal_tensor_assign_1_begin_mask_0, end = concat_80, end_mask = new_cache_23_internal_tensor_assign_1_end_mask_0, squeeze_mask = new_cache_23_internal_tensor_assign_1_squeeze_mask_0, stride = new_cache_23_internal_tensor_assign_1_stride_0, x = reshape_110)[name = tensor("slice_by_index_22")]; + tensor reshape_111_shape_0 = const()[name = tensor("reshape_111_shape_0"), val = tensor([-1])]; + tensor reshape_111 = reshape(shape = reshape_111_shape_0, x = slice_by_index_22)[name = tensor("reshape_111")]; + tensor reshape_112_shape_0 = const()[name = tensor("reshape_112_shape_0"), val = tensor([-1])]; + tensor reshape_112_cast_fp16 = reshape(shape = reshape_112_shape_0, x = var_4681_cast_fp16)[name = tensor("reshape_112_cast_fp16")]; + tensor reshape_113_shape_0 = const()[name = tensor("reshape_113_shape_0"), val = tensor([-1])]; + tensor reshape_113_cast_fp16 = reshape(shape = reshape_113_shape_0, x = cache11_to_fp16)[name = tensor("reshape_113_cast_fp16")]; + tensor scatter_22_mode_0 = const()[name = tensor("scatter_22_mode_0"), val = tensor("update")]; + tensor scatter_22_axis_0 = const()[name = tensor("scatter_22_axis_0"), val = tensor(0)]; + tensor scatter_22_validate_indices_0 = const()[name = tensor("scatter_22_validate_indices_0"), val = tensor(false)]; + tensor scatter_22_cast_fp16 = scatter(axis = scatter_22_axis_0, data = reshape_113_cast_fp16, indices = reshape_111, mode = scatter_22_mode_0, updates = reshape_112_cast_fp16, validate_indices = scatter_22_validate_indices_0)[name = tensor("scatter_22_cast_fp16")]; + tensor reshape_114_cast_fp16 = reshape(shape = shape_70, x = scatter_22_cast_fp16)[name = tensor("reshape_114_cast_fp16")]; + tensor var_4689_begin_0 = const()[name = tensor("op_4689_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor var_4689_end_0 = const()[name = tensor("op_4689_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor var_4689_end_mask_0 = const()[name = tensor("op_4689_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_4689_squeeze_mask_0 = const()[name = tensor("op_4689_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor var_4689_cast_fp16 = slice_by_index(begin = var_4689_begin_0, end = var_4689_end_0, end_mask = var_4689_end_mask_0, squeeze_mask = var_4689_squeeze_mask_0, x = reshape_114_cast_fp16)[name = tensor("op_4689_cast_fp16")]; + tensor var_4691_axis_0 = const()[name = tensor("op_4691_axis_0"), val = tensor(1)]; + tensor var_4691_mode_0 = const()[name = tensor("op_4691_mode_0"), val = tensor("update")]; + tensor var_4691_validate_indices_0 = const()[name = tensor("op_4691_validate_indices_0"), val = tensor(false)]; + tensor var_4691_cast_fp16 = scatter_along_axis(axis = var_4691_axis_0, data = var_4689_cast_fp16, indices = write_indices_23, mode = var_4691_mode_0, updates = v_23_cast_fp16, validate_indices = var_4691_validate_indices_0)[name = tensor("op_4691_cast_fp16")]; + tensor concat_81 = const()[name = tensor("concat_81"), val = tensor([1, 0, 0, 0, 0])]; + tensor concat_82 = const()[name = tensor("concat_82"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_23_internal_tensor_assign_2_stride_0 = const()[name = tensor("new_cache_23_internal_tensor_assign_2_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_23_internal_tensor_assign_2_begin_mask_0 = const()[name = tensor("new_cache_23_internal_tensor_assign_2_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_23_internal_tensor_assign_2_end_mask_0 = const()[name = tensor("new_cache_23_internal_tensor_assign_2_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_23_internal_tensor_assign_2_squeeze_mask_0 = const()[name = tensor("new_cache_23_internal_tensor_assign_2_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_71 = const()[name = tensor("shape_71"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_23 = const()[name = tensor("reduce_prod_23"), val = tensor(1048576)]; + tensor range_1d_23_start_0 = const()[name = tensor("range_1d_23_start_0"), val = tensor(0)]; + tensor range_1d_23_step_0 = const()[name = tensor("range_1d_23_step_0"), val = tensor(1)]; + tensor range_1d_23 = range_1d(end = reduce_prod_23, start = range_1d_23_start_0, step = range_1d_23_step_0)[name = tensor("range_1d_23")]; + tensor reshape_115 = reshape(shape = shape_71, x = range_1d_23)[name = tensor("reshape_115")]; + tensor slice_by_index_23 = slice_by_index(begin = concat_81, begin_mask = new_cache_23_internal_tensor_assign_2_begin_mask_0, end = concat_82, end_mask = new_cache_23_internal_tensor_assign_2_end_mask_0, squeeze_mask = new_cache_23_internal_tensor_assign_2_squeeze_mask_0, stride = new_cache_23_internal_tensor_assign_2_stride_0, x = reshape_115)[name = tensor("slice_by_index_23")]; + tensor reshape_116_shape_0 = const()[name = tensor("reshape_116_shape_0"), val = tensor([-1])]; + tensor reshape_116 = reshape(shape = reshape_116_shape_0, x = slice_by_index_23)[name = tensor("reshape_116")]; + tensor reshape_117_shape_0 = const()[name = tensor("reshape_117_shape_0"), val = tensor([-1])]; + tensor reshape_117_cast_fp16 = reshape(shape = reshape_117_shape_0, x = var_4691_cast_fp16)[name = tensor("reshape_117_cast_fp16")]; + tensor reshape_118_shape_0 = const()[name = tensor("reshape_118_shape_0"), val = tensor([-1])]; + tensor reshape_118_cast_fp16 = reshape(shape = reshape_118_shape_0, x = reshape_114_cast_fp16)[name = tensor("reshape_118_cast_fp16")]; + tensor scatter_23_mode_0 = const()[name = tensor("scatter_23_mode_0"), val = tensor("update")]; + tensor scatter_23_axis_0 = const()[name = tensor("scatter_23_axis_0"), val = tensor(0)]; + tensor scatter_23_validate_indices_0 = const()[name = tensor("scatter_23_validate_indices_0"), val = tensor(false)]; + tensor scatter_23_cast_fp16 = scatter(axis = scatter_23_axis_0, data = reshape_118_cast_fp16, indices = reshape_116, mode = scatter_23_mode_0, updates = reshape_117_cast_fp16, validate_indices = scatter_23_validate_indices_0)[name = tensor("scatter_23_cast_fp16")]; + tensor reshape_119_cast_fp16 = reshape(shape = shape_71, x = scatter_23_cast_fp16)[name = tensor("reshape_119_cast_fp16")]; + tensor reshape_119_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("reshape_119_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor keys_67_begin_0 = const()[name = tensor("keys_67_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor keys_67_end_0 = const()[name = tensor("keys_67_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor keys_67_end_mask_0 = const()[name = tensor("keys_67_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor keys_67_squeeze_mask_0 = const()[name = tensor("keys_67_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor keys_67_cast_fp16 = slice_by_index(begin = keys_67_begin_0, end = keys_67_end_0, end_mask = keys_67_end_mask_0, squeeze_mask = keys_67_squeeze_mask_0, x = reshape_119_cast_fp16)[name = tensor("keys_67_cast_fp16")]; + tensor values_67_begin_0 = const()[name = tensor("values_67_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor values_67_end_0 = const()[name = tensor("values_67_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor values_67_end_mask_0 = const()[name = tensor("values_67_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor values_67_squeeze_mask_0 = const()[name = tensor("values_67_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor values_67_cast_fp16 = slice_by_index(begin = values_67_begin_0, end = values_67_end_0, end_mask = values_67_end_mask_0, squeeze_mask = values_67_squeeze_mask_0, x = reshape_119_cast_fp16)[name = tensor("values_67_cast_fp16")]; + tensor var_4703_cast_fp16 = not_equal(x = keys_67_cast_fp16, y = keys_67_cast_fp16)[name = tensor("op_4703_cast_fp16")]; + tensor keys_69_cast_fp16 = select(a = var_463_to_fp16, b = keys_67_cast_fp16, cond = var_4703_cast_fp16)[name = tensor("keys_69_cast_fp16")]; + tensor var_4711_cast_fp16 = not_equal(x = values_67_cast_fp16, y = values_67_cast_fp16)[name = tensor("op_4711_cast_fp16")]; + tensor values_69_cast_fp16 = select(a = var_463_to_fp16, b = values_67_cast_fp16, cond = var_4711_cast_fp16)[name = tensor("values_69_cast_fp16")]; + tensor var_4735 = const()[name = tensor("op_4735"), val = tensor([0, 2, 1, 3])]; + tensor var_4751 = const()[name = tensor("op_4751"), val = tensor([1, 1, 1])]; + tensor var_4752_cast_fp16 = reshape(shape = var_4751, x = position11_to_fp16)[name = tensor("op_4752_cast_fp16")]; + tensor var_4772_to_fp16 = const()[name = tensor("op_4772_to_fp16"), val = tensor(0x1p+0)]; + tensor valid_len_23_cast_fp16 = add(x = var_4752_cast_fp16, y = var_4772_to_fp16)[name = tensor("valid_len_23_cast_fp16")]; + tensor valid_mask_23_cast_fp16 = less(x = k_positions_1_promoted_to_fp16, y = valid_len_23_cast_fp16)[name = tensor("valid_mask_23_cast_fp16")]; + tensor causal_mask_23_cast_fp16 = less_equal(x = k_positions_1_promoted_to_fp16, y = var_4752_cast_fp16)[name = tensor("causal_mask_23_cast_fp16")]; + tensor attn_mask_45 = logical_and(x = valid_mask_23_cast_fp16, y = causal_mask_23_cast_fp16)[name = tensor("attn_mask_45")]; + tensor attn_mask_47_axes_0 = const()[name = tensor("attn_mask_47_axes_0"), val = tensor([1])]; + tensor attn_mask_47 = expand_dims(axes = attn_mask_47_axes_0, x = attn_mask_45)[name = tensor("attn_mask_47")]; + tensor var_4790_transpose_x_0 = const()[name = tensor("op_4790_transpose_x_0"), val = tensor(false)]; + tensor var_4790_transpose_y_0 = const()[name = tensor("op_4790_transpose_y_0"), val = tensor(false)]; + tensor transpose_94_perm_0 = const()[name = tensor("transpose_94_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_95_perm_0 = const()[name = tensor("transpose_95_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_95 = transpose(perm = transpose_95_perm_0, x = keys_69_cast_fp16)[name = tensor("transpose_169")]; + tensor transpose_94 = transpose(perm = transpose_94_perm_0, x = q_69_cast_fp16)[name = tensor("transpose_170")]; + tensor var_4790_cast_fp16 = matmul(transpose_x = var_4790_transpose_x_0, transpose_y = var_4790_transpose_y_0, x = transpose_94, y = transpose_95)[name = tensor("op_4790_cast_fp16")]; + tensor var_4784_to_fp16 = const()[name = tensor("op_4784_to_fp16"), val = tensor([0x1p-3])]; + tensor attn_weights_67_cast_fp16 = mul(x = var_4790_cast_fp16, y = var_4784_to_fp16)[name = tensor("attn_weights_67_cast_fp16")]; + tensor var_4792 = logical_not(x = attn_mask_47)[name = tensor("op_4792")]; + tensor var_4793_to_fp16 = const()[name = tensor("op_4793_to_fp16"), val = tensor(-inf)]; + tensor attn_weights_69_cast_fp16 = select(a = var_4793_to_fp16, b = attn_weights_67_cast_fp16, cond = var_4792)[name = tensor("attn_weights_69_cast_fp16")]; + tensor var_4795 = const()[name = tensor("op_4795"), val = tensor(-1)]; + tensor attn_weights_71_cast_fp16 = softmax(axis = var_4795, x = attn_weights_69_cast_fp16)[name = tensor("attn_weights_71_cast_fp16")]; + tensor attn_output_23_transpose_x_0 = const()[name = tensor("attn_output_23_transpose_x_0"), val = tensor(false)]; + tensor attn_output_23_transpose_y_0 = const()[name = tensor("attn_output_23_transpose_y_0"), val = tensor(false)]; + tensor values_71_cast_fp16 = transpose(perm = var_4735, x = values_69_cast_fp16)[name = tensor("transpose_171")]; + tensor attn_output_23_cast_fp16 = matmul(transpose_x = attn_output_23_transpose_x_0, transpose_y = attn_output_23_transpose_y_0, x = attn_weights_71_cast_fp16, y = values_71_cast_fp16)[name = tensor("attn_output_23_cast_fp16")]; + tensor var_4803 = const()[name = tensor("op_4803"), val = tensor([0, 2, 1, 3])]; + tensor var_4808 = const()[name = tensor("op_4808"), val = tensor([1, 1, 1024])]; + tensor var_4804_cast_fp16 = transpose(perm = var_4803, x = attn_output_23_cast_fp16)[name = tensor("transpose_168")]; + tensor input_139_cast_fp16 = reshape(shape = var_4808, x = var_4804_cast_fp16)[name = tensor("input_139_cast_fp16")]; + tensor attn11_out_proj_weight_to_fp16 = const()[name = tensor("attn11_out_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(284349056)))]; + tensor linear_46_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = attn11_out_proj_weight_to_fp16, x = input_139_cast_fp16)[name = tensor("linear_46_cast_fp16")]; + tensor var_4814_to_fp16 = const()[name = tensor("op_4814_to_fp16"), val = tensor(0x1p+0)]; + tensor var_4815_cast_fp16 = add(x = position11_to_fp16, y = var_4814_to_fp16)[name = tensor("op_4815_cast_fp16")]; + tensor var_4815_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_4815_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor input_141_cast_fp16 = add(x = input_135_cast_fp16, y = linear_46_cast_fp16)[name = tensor("input_141_cast_fp16")]; + tensor input_143_axes_0 = const()[name = tensor("input_143_axes_0"), val = tensor([-1])]; + tensor norm11_2_weight_to_fp16 = const()[name = tensor("norm11_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(286446272)))]; + tensor norm11_2_bias_to_fp16 = const()[name = tensor("norm11_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(286448384)))]; + tensor var_4819_to_fp16 = const()[name = tensor("op_4819_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_143_cast_fp16 = layer_norm(axes = input_143_axes_0, beta = norm11_2_bias_to_fp16, epsilon = var_4819_to_fp16, gamma = norm11_2_weight_to_fp16, x = input_141_cast_fp16)[name = tensor("input_143_cast_fp16")]; + tensor linear11_1_weight_to_fp16 = const()[name = tensor("linear11_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(286450496)))]; + tensor linear_47_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = linear11_1_weight_to_fp16, x = input_143_cast_fp16)[name = tensor("linear_47_cast_fp16")]; + tensor input_145_mode_0 = const()[name = tensor("input_145_mode_0"), val = tensor("EXACT")]; + tensor input_145_cast_fp16 = gelu(mode = input_145_mode_0, x = linear_47_cast_fp16)[name = tensor("input_145_cast_fp16")]; + tensor linear11_2_weight_to_fp16 = const()[name = tensor("linear11_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(294839168)))]; + tensor linear_48_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = linear11_2_weight_to_fp16, x = input_145_cast_fp16)[name = tensor("linear_48_cast_fp16")]; + tensor input_147_cast_fp16 = add(x = input_141_cast_fp16, y = linear_48_cast_fp16)[name = tensor("input_147_cast_fp16")]; + tensor input_149_axes_0 = const()[name = tensor("input_149_axes_0"), val = tensor([-1])]; + tensor norm12_1_weight_to_fp16 = const()[name = tensor("norm12_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(303227840)))]; + tensor norm12_1_bias_to_fp16 = const()[name = tensor("norm12_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(303229952)))]; + tensor var_4836_to_fp16 = const()[name = tensor("op_4836_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_149_cast_fp16 = layer_norm(axes = input_149_axes_0, beta = norm12_1_bias_to_fp16, epsilon = var_4836_to_fp16, gamma = norm12_1_weight_to_fp16, x = input_147_cast_fp16)[name = tensor("input_149_cast_fp16")]; + tensor attn12_in_proj_weight_to_fp16 = const()[name = tensor("attn12_in_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(303232064)))]; + tensor linear_49_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = attn12_in_proj_weight_to_fp16, x = input_149_cast_fp16)[name = tensor("linear_49_cast_fp16")]; + tensor var_4850 = const()[name = tensor("op_4850"), val = tensor([1, 1, 3, 16, 64])]; + tensor qkv_25_cast_fp16 = reshape(shape = var_4850, x = linear_49_cast_fp16)[name = tensor("qkv_25_cast_fp16")]; + tensor q_73_begin_0 = const()[name = tensor("q_73_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor q_73_end_0 = const()[name = tensor("q_73_end_0"), val = tensor([1, 1, 1, 16, 64])]; + tensor q_73_end_mask_0 = const()[name = tensor("q_73_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor q_73_squeeze_mask_0 = const()[name = tensor("q_73_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor q_73_cast_fp16 = slice_by_index(begin = q_73_begin_0, end = q_73_end_0, end_mask = q_73_end_mask_0, squeeze_mask = q_73_squeeze_mask_0, x = qkv_25_cast_fp16)[name = tensor("q_73_cast_fp16")]; + tensor k_49_begin_0 = const()[name = tensor("k_49_begin_0"), val = tensor([0, 0, 1, 0, 0])]; + tensor k_49_end_0 = const()[name = tensor("k_49_end_0"), val = tensor([1, 1, 2, 16, 64])]; + tensor k_49_end_mask_0 = const()[name = tensor("k_49_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor k_49_squeeze_mask_0 = const()[name = tensor("k_49_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor k_49_cast_fp16 = slice_by_index(begin = k_49_begin_0, end = k_49_end_0, end_mask = k_49_end_mask_0, squeeze_mask = k_49_squeeze_mask_0, x = qkv_25_cast_fp16)[name = tensor("k_49_cast_fp16")]; + tensor v_25_begin_0 = const()[name = tensor("v_25_begin_0"), val = tensor([0, 0, 2, 0, 0])]; + tensor v_25_end_0 = const()[name = tensor("v_25_end_0"), val = tensor([1, 1, 3, 16, 64])]; + tensor v_25_end_mask_0 = const()[name = tensor("v_25_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor v_25_squeeze_mask_0 = const()[name = tensor("v_25_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor v_25_cast_fp16 = slice_by_index(begin = v_25_begin_0, end = v_25_end_0, end_mask = v_25_end_mask_0, squeeze_mask = v_25_squeeze_mask_0, x = qkv_25_cast_fp16)[name = tensor("v_25_cast_fp16")]; + tensor var_4916 = const()[name = tensor("op_4916"), val = tensor([1, 1, 1, 1])]; + tensor position12_to_fp16_dtype_0 = const()[name = tensor("position12_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor position12_to_fp16 = cast(dtype = position12_to_fp16_dtype_0, x = position12)[name = tensor("cast_493")]; + tensor ts_77_cast_fp16 = reshape(shape = var_4916, x = position12_to_fp16)[name = tensor("ts_77_cast_fp16")]; + tensor var_4923 = const()[name = tensor("op_4923"), val = tensor([1, 1, 16, 32, 2])]; + tensor q_complex_25_cast_fp16 = reshape(shape = var_4923, x = q_73_cast_fp16)[name = tensor("q_complex_25_cast_fp16")]; + tensor var_4930 = const()[name = tensor("op_4930"), val = tensor([1, 1, 16, 32, 2])]; + tensor k_complex_25_cast_fp16 = reshape(shape = var_4930, x = k_49_cast_fp16)[name = tensor("k_complex_25_cast_fp16")]; + tensor var_4934_begin_0 = const()[name = tensor("op_4934_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_4934_end_0 = const()[name = tensor("op_4934_end_0"), val = tensor([1, 1, 16, 32, 1])]; + tensor var_4934_end_mask_0 = const()[name = tensor("op_4934_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_4934_squeeze_mask_0 = const()[name = tensor("op_4934_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_4934_cast_fp16 = slice_by_index(begin = var_4934_begin_0, end = var_4934_end_0, end_mask = var_4934_end_mask_0, squeeze_mask = var_4934_squeeze_mask_0, x = q_complex_25_cast_fp16)[name = tensor("op_4934_cast_fp16")]; + tensor var_4942_begin_0 = const()[name = tensor("op_4942_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_4942_end_0 = const()[name = tensor("op_4942_end_0"), val = tensor([1, 1, 16, 32, 2])]; + tensor var_4942_end_mask_0 = const()[name = tensor("op_4942_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_4942_squeeze_mask_0 = const()[name = tensor("op_4942_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_4942_cast_fp16 = slice_by_index(begin = var_4942_begin_0, end = var_4942_end_0, end_mask = var_4942_end_mask_0, squeeze_mask = var_4942_squeeze_mask_0, x = q_complex_25_cast_fp16)[name = tensor("op_4942_cast_fp16")]; + tensor var_4950_begin_0 = const()[name = tensor("op_4950_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_4950_end_0 = const()[name = tensor("op_4950_end_0"), val = tensor([1, 1, 16, 32, 1])]; + tensor var_4950_end_mask_0 = const()[name = tensor("op_4950_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_4950_squeeze_mask_0 = const()[name = tensor("op_4950_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_4950_cast_fp16 = slice_by_index(begin = var_4950_begin_0, end = var_4950_end_0, end_mask = var_4950_end_mask_0, squeeze_mask = var_4950_squeeze_mask_0, x = k_complex_25_cast_fp16)[name = tensor("op_4950_cast_fp16")]; + tensor var_4958_begin_0 = const()[name = tensor("op_4958_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_4958_end_0 = const()[name = tensor("op_4958_end_0"), val = tensor([1, 1, 16, 32, 2])]; + tensor var_4958_end_mask_0 = const()[name = tensor("op_4958_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_4958_squeeze_mask_0 = const()[name = tensor("op_4958_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_4958_cast_fp16 = slice_by_index(begin = var_4958_begin_0, end = var_4958_end_0, end_mask = var_4958_end_mask_0, squeeze_mask = var_4958_squeeze_mask_0, x = k_complex_25_cast_fp16)[name = tensor("op_4958_cast_fp16")]; + tensor freqs_25_to_fp16 = const()[name = tensor("freqs_25_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(309523584)))]; + tensor var_4964_cast_fp16 = mul(x = freqs_25_to_fp16, y = ts_77_cast_fp16)[name = tensor("op_4964_cast_fp16")]; + tensor rotr_25_cast_fp16 = cos(x = var_4964_cast_fp16)[name = tensor("rotr_25_cast_fp16")]; + tensor roti_25_cast_fp16 = sin(x = var_4964_cast_fp16)[name = tensor("roti_25_cast_fp16")]; + tensor var_4968_cast_fp16 = mul(x = var_4934_cast_fp16, y = rotr_25_cast_fp16)[name = tensor("op_4968_cast_fp16")]; + tensor var_4969_cast_fp16 = mul(x = var_4942_cast_fp16, y = roti_25_cast_fp16)[name = tensor("op_4969_cast_fp16")]; + tensor qor_49_cast_fp16 = sub(x = var_4968_cast_fp16, y = var_4969_cast_fp16)[name = tensor("qor_49_cast_fp16")]; + tensor var_4972_cast_fp16 = mul(x = var_4934_cast_fp16, y = roti_25_cast_fp16)[name = tensor("op_4972_cast_fp16")]; + tensor var_4973_cast_fp16 = mul(x = var_4942_cast_fp16, y = rotr_25_cast_fp16)[name = tensor("op_4973_cast_fp16")]; + tensor qoi_49_cast_fp16 = add(x = var_4972_cast_fp16, y = var_4973_cast_fp16)[name = tensor("qoi_49_cast_fp16")]; + tensor var_4976_cast_fp16 = mul(x = var_4950_cast_fp16, y = rotr_25_cast_fp16)[name = tensor("op_4976_cast_fp16")]; + tensor var_4977_cast_fp16 = mul(x = var_4958_cast_fp16, y = roti_25_cast_fp16)[name = tensor("op_4977_cast_fp16")]; + tensor kor_49_cast_fp16 = sub(x = var_4976_cast_fp16, y = var_4977_cast_fp16)[name = tensor("kor_49_cast_fp16")]; + tensor var_4980_cast_fp16 = mul(x = var_4950_cast_fp16, y = roti_25_cast_fp16)[name = tensor("op_4980_cast_fp16")]; + tensor var_4981_cast_fp16 = mul(x = var_4958_cast_fp16, y = rotr_25_cast_fp16)[name = tensor("op_4981_cast_fp16")]; + tensor koi_49_cast_fp16 = add(x = var_4980_cast_fp16, y = var_4981_cast_fp16)[name = tensor("koi_49_cast_fp16")]; + tensor qo_25_axis_0 = const()[name = tensor("qo_25_axis_0"), val = tensor(-1)]; + tensor qo_25_cast_fp16 = stack(axis = qo_25_axis_0, values = (qor_49_cast_fp16, qoi_49_cast_fp16))[name = tensor("qo_25_cast_fp16")]; + tensor ko_25_axis_0 = const()[name = tensor("ko_25_axis_0"), val = tensor(-1)]; + tensor ko_25_cast_fp16 = stack(axis = ko_25_axis_0, values = (kor_49_cast_fp16, koi_49_cast_fp16))[name = tensor("ko_25_cast_fp16")]; + tensor var_5014 = const()[name = tensor("op_5014"), val = tensor([1, 1, 16, 64])]; + tensor q_75_cast_fp16 = reshape(shape = var_5014, x = qo_25_cast_fp16)[name = tensor("q_75_cast_fp16")]; + tensor var_5020 = const()[name = tensor("op_5020"), val = tensor([1, 1, 16, 64])]; + tensor k_51_cast_fp16 = reshape(shape = var_5020, x = ko_25_cast_fp16)[name = tensor("k_51_cast_fp16")]; + tensor _inversed_5045_y_0_to_fp16 = const()[name = tensor("_inversed_5045_y_0_to_fp16"), val = tensor(0x1p-9)]; + tensor _inversed_5045_cast_fp16 = mul(x = ts_77_cast_fp16, y = _inversed_5045_y_0_to_fp16)[name = tensor("_inversed_5045_cast_fp16")]; + tensor var_5046_cast_fp16 = floor(x = _inversed_5045_cast_fp16)[name = tensor("op_5046_cast_fp16")]; + tensor var_5047_to_fp16 = const()[name = tensor("op_5047_to_fp16"), val = tensor(0x1p+9)]; + tensor var_5048_cast_fp16 = mul(x = var_5046_cast_fp16, y = var_5047_to_fp16)[name = tensor("op_5048_cast_fp16")]; + tensor write_indices_float_51_cast_fp16 = sub(x = ts_77_cast_fp16, y = var_5048_cast_fp16)[name = tensor("write_indices_float_51_cast_fp16")]; + tensor var_5055_dtype_0 = const()[name = tensor("op_5055_dtype_0"), val = tensor("int32")]; + tensor write_indices_25_reps_0 = const()[name = tensor("write_indices_25_reps_0"), val = tensor([1, 1, 16, 64])]; + tensor write_indices_float_51_cast_fp16_to_int32 = cast(dtype = var_5055_dtype_0, x = write_indices_float_51_cast_fp16)[name = tensor("cast_492")]; + tensor write_indices_25 = tile(reps = write_indices_25_reps_0, x = write_indices_float_51_cast_fp16_to_int32)[name = tensor("write_indices_25")]; + tensor var_5065_begin_0 = const()[name = tensor("op_5065_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_5065_end_0 = const()[name = tensor("op_5065_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor var_5065_end_mask_0 = const()[name = tensor("op_5065_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_5065_squeeze_mask_0 = const()[name = tensor("op_5065_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor cache12_to_fp16_dtype_0 = const()[name = tensor("cache12_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor cache12_to_fp16 = cast(dtype = cache12_to_fp16_dtype_0, x = cache12)[name = tensor("cast_491")]; + tensor var_5065_cast_fp16 = slice_by_index(begin = var_5065_begin_0, end = var_5065_end_0, end_mask = var_5065_end_mask_0, squeeze_mask = var_5065_squeeze_mask_0, x = cache12_to_fp16)[name = tensor("op_5065_cast_fp16")]; + tensor var_5067_axis_0 = const()[name = tensor("op_5067_axis_0"), val = tensor(1)]; + tensor var_5067_mode_0 = const()[name = tensor("op_5067_mode_0"), val = tensor("update")]; + tensor var_5067_validate_indices_0 = const()[name = tensor("op_5067_validate_indices_0"), val = tensor(false)]; + tensor var_5067_cast_fp16 = scatter_along_axis(axis = var_5067_axis_0, data = var_5065_cast_fp16, indices = write_indices_25, mode = var_5067_mode_0, updates = k_51_cast_fp16, validate_indices = var_5067_validate_indices_0)[name = tensor("op_5067_cast_fp16")]; + tensor concat_86 = const()[name = tensor("concat_86"), val = tensor([0, 0, 0, 0, 0])]; + tensor concat_87 = const()[name = tensor("concat_87"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_25_internal_tensor_assign_1_stride_0 = const()[name = tensor("new_cache_25_internal_tensor_assign_1_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_25_internal_tensor_assign_1_begin_mask_0 = const()[name = tensor("new_cache_25_internal_tensor_assign_1_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_25_internal_tensor_assign_1_end_mask_0 = const()[name = tensor("new_cache_25_internal_tensor_assign_1_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_25_internal_tensor_assign_1_squeeze_mask_0 = const()[name = tensor("new_cache_25_internal_tensor_assign_1_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_72 = const()[name = tensor("shape_72"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_24 = const()[name = tensor("reduce_prod_24"), val = tensor(1048576)]; + tensor range_1d_24_start_0 = const()[name = tensor("range_1d_24_start_0"), val = tensor(0)]; + tensor range_1d_24_step_0 = const()[name = tensor("range_1d_24_step_0"), val = tensor(1)]; + tensor range_1d_24 = range_1d(end = reduce_prod_24, start = range_1d_24_start_0, step = range_1d_24_step_0)[name = tensor("range_1d_24")]; + tensor reshape_120 = reshape(shape = shape_72, x = range_1d_24)[name = tensor("reshape_120")]; + tensor slice_by_index_24 = slice_by_index(begin = concat_86, begin_mask = new_cache_25_internal_tensor_assign_1_begin_mask_0, end = concat_87, end_mask = new_cache_25_internal_tensor_assign_1_end_mask_0, squeeze_mask = new_cache_25_internal_tensor_assign_1_squeeze_mask_0, stride = new_cache_25_internal_tensor_assign_1_stride_0, x = reshape_120)[name = tensor("slice_by_index_24")]; + tensor reshape_121_shape_0 = const()[name = tensor("reshape_121_shape_0"), val = tensor([-1])]; + tensor reshape_121 = reshape(shape = reshape_121_shape_0, x = slice_by_index_24)[name = tensor("reshape_121")]; + tensor reshape_122_shape_0 = const()[name = tensor("reshape_122_shape_0"), val = tensor([-1])]; + tensor reshape_122_cast_fp16 = reshape(shape = reshape_122_shape_0, x = var_5067_cast_fp16)[name = tensor("reshape_122_cast_fp16")]; + tensor reshape_123_shape_0 = const()[name = tensor("reshape_123_shape_0"), val = tensor([-1])]; + tensor reshape_123_cast_fp16 = reshape(shape = reshape_123_shape_0, x = cache12_to_fp16)[name = tensor("reshape_123_cast_fp16")]; + tensor scatter_24_mode_0 = const()[name = tensor("scatter_24_mode_0"), val = tensor("update")]; + tensor scatter_24_axis_0 = const()[name = tensor("scatter_24_axis_0"), val = tensor(0)]; + tensor scatter_24_validate_indices_0 = const()[name = tensor("scatter_24_validate_indices_0"), val = tensor(false)]; + tensor scatter_24_cast_fp16 = scatter(axis = scatter_24_axis_0, data = reshape_123_cast_fp16, indices = reshape_121, mode = scatter_24_mode_0, updates = reshape_122_cast_fp16, validate_indices = scatter_24_validate_indices_0)[name = tensor("scatter_24_cast_fp16")]; + tensor reshape_124_cast_fp16 = reshape(shape = shape_72, x = scatter_24_cast_fp16)[name = tensor("reshape_124_cast_fp16")]; + tensor var_5075_begin_0 = const()[name = tensor("op_5075_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor var_5075_end_0 = const()[name = tensor("op_5075_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor var_5075_end_mask_0 = const()[name = tensor("op_5075_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_5075_squeeze_mask_0 = const()[name = tensor("op_5075_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor var_5075_cast_fp16 = slice_by_index(begin = var_5075_begin_0, end = var_5075_end_0, end_mask = var_5075_end_mask_0, squeeze_mask = var_5075_squeeze_mask_0, x = reshape_124_cast_fp16)[name = tensor("op_5075_cast_fp16")]; + tensor var_5077_axis_0 = const()[name = tensor("op_5077_axis_0"), val = tensor(1)]; + tensor var_5077_mode_0 = const()[name = tensor("op_5077_mode_0"), val = tensor("update")]; + tensor var_5077_validate_indices_0 = const()[name = tensor("op_5077_validate_indices_0"), val = tensor(false)]; + tensor var_5077_cast_fp16 = scatter_along_axis(axis = var_5077_axis_0, data = var_5075_cast_fp16, indices = write_indices_25, mode = var_5077_mode_0, updates = v_25_cast_fp16, validate_indices = var_5077_validate_indices_0)[name = tensor("op_5077_cast_fp16")]; + tensor concat_88 = const()[name = tensor("concat_88"), val = tensor([1, 0, 0, 0, 0])]; + tensor concat_89 = const()[name = tensor("concat_89"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_25_internal_tensor_assign_2_stride_0 = const()[name = tensor("new_cache_25_internal_tensor_assign_2_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_25_internal_tensor_assign_2_begin_mask_0 = const()[name = tensor("new_cache_25_internal_tensor_assign_2_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_25_internal_tensor_assign_2_end_mask_0 = const()[name = tensor("new_cache_25_internal_tensor_assign_2_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_25_internal_tensor_assign_2_squeeze_mask_0 = const()[name = tensor("new_cache_25_internal_tensor_assign_2_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_73 = const()[name = tensor("shape_73"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_25 = const()[name = tensor("reduce_prod_25"), val = tensor(1048576)]; + tensor range_1d_25_start_0 = const()[name = tensor("range_1d_25_start_0"), val = tensor(0)]; + tensor range_1d_25_step_0 = const()[name = tensor("range_1d_25_step_0"), val = tensor(1)]; + tensor range_1d_25 = range_1d(end = reduce_prod_25, start = range_1d_25_start_0, step = range_1d_25_step_0)[name = tensor("range_1d_25")]; + tensor reshape_125 = reshape(shape = shape_73, x = range_1d_25)[name = tensor("reshape_125")]; + tensor slice_by_index_25 = slice_by_index(begin = concat_88, begin_mask = new_cache_25_internal_tensor_assign_2_begin_mask_0, end = concat_89, end_mask = new_cache_25_internal_tensor_assign_2_end_mask_0, squeeze_mask = new_cache_25_internal_tensor_assign_2_squeeze_mask_0, stride = new_cache_25_internal_tensor_assign_2_stride_0, x = reshape_125)[name = tensor("slice_by_index_25")]; + tensor reshape_126_shape_0 = const()[name = tensor("reshape_126_shape_0"), val = tensor([-1])]; + tensor reshape_126 = reshape(shape = reshape_126_shape_0, x = slice_by_index_25)[name = tensor("reshape_126")]; + tensor reshape_127_shape_0 = const()[name = tensor("reshape_127_shape_0"), val = tensor([-1])]; + tensor reshape_127_cast_fp16 = reshape(shape = reshape_127_shape_0, x = var_5077_cast_fp16)[name = tensor("reshape_127_cast_fp16")]; + tensor reshape_128_shape_0 = const()[name = tensor("reshape_128_shape_0"), val = tensor([-1])]; + tensor reshape_128_cast_fp16 = reshape(shape = reshape_128_shape_0, x = reshape_124_cast_fp16)[name = tensor("reshape_128_cast_fp16")]; + tensor scatter_25_mode_0 = const()[name = tensor("scatter_25_mode_0"), val = tensor("update")]; + tensor scatter_25_axis_0 = const()[name = tensor("scatter_25_axis_0"), val = tensor(0)]; + tensor scatter_25_validate_indices_0 = const()[name = tensor("scatter_25_validate_indices_0"), val = tensor(false)]; + tensor scatter_25_cast_fp16 = scatter(axis = scatter_25_axis_0, data = reshape_128_cast_fp16, indices = reshape_126, mode = scatter_25_mode_0, updates = reshape_127_cast_fp16, validate_indices = scatter_25_validate_indices_0)[name = tensor("scatter_25_cast_fp16")]; + tensor reshape_129_cast_fp16 = reshape(shape = shape_73, x = scatter_25_cast_fp16)[name = tensor("reshape_129_cast_fp16")]; + tensor reshape_129_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("reshape_129_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor keys_73_begin_0 = const()[name = tensor("keys_73_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor keys_73_end_0 = const()[name = tensor("keys_73_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor keys_73_end_mask_0 = const()[name = tensor("keys_73_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor keys_73_squeeze_mask_0 = const()[name = tensor("keys_73_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor keys_73_cast_fp16 = slice_by_index(begin = keys_73_begin_0, end = keys_73_end_0, end_mask = keys_73_end_mask_0, squeeze_mask = keys_73_squeeze_mask_0, x = reshape_129_cast_fp16)[name = tensor("keys_73_cast_fp16")]; + tensor values_73_begin_0 = const()[name = tensor("values_73_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor values_73_end_0 = const()[name = tensor("values_73_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor values_73_end_mask_0 = const()[name = tensor("values_73_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor values_73_squeeze_mask_0 = const()[name = tensor("values_73_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor values_73_cast_fp16 = slice_by_index(begin = values_73_begin_0, end = values_73_end_0, end_mask = values_73_end_mask_0, squeeze_mask = values_73_squeeze_mask_0, x = reshape_129_cast_fp16)[name = tensor("values_73_cast_fp16")]; + tensor var_5089_cast_fp16 = not_equal(x = keys_73_cast_fp16, y = keys_73_cast_fp16)[name = tensor("op_5089_cast_fp16")]; + tensor keys_75_cast_fp16 = select(a = var_463_to_fp16, b = keys_73_cast_fp16, cond = var_5089_cast_fp16)[name = tensor("keys_75_cast_fp16")]; + tensor var_5097_cast_fp16 = not_equal(x = values_73_cast_fp16, y = values_73_cast_fp16)[name = tensor("op_5097_cast_fp16")]; + tensor values_75_cast_fp16 = select(a = var_463_to_fp16, b = values_73_cast_fp16, cond = var_5097_cast_fp16)[name = tensor("values_75_cast_fp16")]; + tensor var_5121 = const()[name = tensor("op_5121"), val = tensor([0, 2, 1, 3])]; + tensor var_5137 = const()[name = tensor("op_5137"), val = tensor([1, 1, 1])]; + tensor var_5138_cast_fp16 = reshape(shape = var_5137, x = position12_to_fp16)[name = tensor("op_5138_cast_fp16")]; + tensor var_5158_to_fp16 = const()[name = tensor("op_5158_to_fp16"), val = tensor(0x1p+0)]; + tensor valid_len_25_cast_fp16 = add(x = var_5138_cast_fp16, y = var_5158_to_fp16)[name = tensor("valid_len_25_cast_fp16")]; + tensor valid_mask_25_cast_fp16 = less(x = k_positions_1_promoted_to_fp16, y = valid_len_25_cast_fp16)[name = tensor("valid_mask_25_cast_fp16")]; + tensor causal_mask_25_cast_fp16 = less_equal(x = k_positions_1_promoted_to_fp16, y = var_5138_cast_fp16)[name = tensor("causal_mask_25_cast_fp16")]; + tensor attn_mask_49 = logical_and(x = valid_mask_25_cast_fp16, y = causal_mask_25_cast_fp16)[name = tensor("attn_mask_49")]; + tensor attn_mask_51_axes_0 = const()[name = tensor("attn_mask_51_axes_0"), val = tensor([1])]; + tensor attn_mask_51 = expand_dims(axes = attn_mask_51_axes_0, x = attn_mask_49)[name = tensor("attn_mask_51")]; + tensor var_5176_transpose_x_0 = const()[name = tensor("op_5176_transpose_x_0"), val = tensor(false)]; + tensor var_5176_transpose_y_0 = const()[name = tensor("op_5176_transpose_y_0"), val = tensor(false)]; + tensor transpose_96_perm_0 = const()[name = tensor("transpose_96_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_97_perm_0 = const()[name = tensor("transpose_97_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_97 = transpose(perm = transpose_97_perm_0, x = keys_75_cast_fp16)[name = tensor("transpose_165")]; + tensor transpose_96 = transpose(perm = transpose_96_perm_0, x = q_75_cast_fp16)[name = tensor("transpose_166")]; + tensor var_5176_cast_fp16 = matmul(transpose_x = var_5176_transpose_x_0, transpose_y = var_5176_transpose_y_0, x = transpose_96, y = transpose_97)[name = tensor("op_5176_cast_fp16")]; + tensor var_5170_to_fp16 = const()[name = tensor("op_5170_to_fp16"), val = tensor([0x1p-3])]; + tensor attn_weights_73_cast_fp16 = mul(x = var_5176_cast_fp16, y = var_5170_to_fp16)[name = tensor("attn_weights_73_cast_fp16")]; + tensor var_5178 = logical_not(x = attn_mask_51)[name = tensor("op_5178")]; + tensor var_5179_to_fp16 = const()[name = tensor("op_5179_to_fp16"), val = tensor(-inf)]; + tensor attn_weights_75_cast_fp16 = select(a = var_5179_to_fp16, b = attn_weights_73_cast_fp16, cond = var_5178)[name = tensor("attn_weights_75_cast_fp16")]; + tensor var_5181 = const()[name = tensor("op_5181"), val = tensor(-1)]; + tensor attn_weights_77_cast_fp16 = softmax(axis = var_5181, x = attn_weights_75_cast_fp16)[name = tensor("attn_weights_77_cast_fp16")]; + tensor attn_output_25_transpose_x_0 = const()[name = tensor("attn_output_25_transpose_x_0"), val = tensor(false)]; + tensor attn_output_25_transpose_y_0 = const()[name = tensor("attn_output_25_transpose_y_0"), val = tensor(false)]; + tensor values_77_cast_fp16 = transpose(perm = var_5121, x = values_75_cast_fp16)[name = tensor("transpose_167")]; + tensor attn_output_25_cast_fp16 = matmul(transpose_x = attn_output_25_transpose_x_0, transpose_y = attn_output_25_transpose_y_0, x = attn_weights_77_cast_fp16, y = values_77_cast_fp16)[name = tensor("attn_output_25_cast_fp16")]; + tensor var_5189 = const()[name = tensor("op_5189"), val = tensor([0, 2, 1, 3])]; + tensor var_5194 = const()[name = tensor("op_5194"), val = tensor([1, 1, 1024])]; + tensor var_5190_cast_fp16 = transpose(perm = var_5189, x = attn_output_25_cast_fp16)[name = tensor("transpose_164")]; + tensor input_151_cast_fp16 = reshape(shape = var_5194, x = var_5190_cast_fp16)[name = tensor("input_151_cast_fp16")]; + tensor attn12_out_proj_weight_to_fp16 = const()[name = tensor("attn12_out_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(309523712)))]; + tensor linear_50_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = attn12_out_proj_weight_to_fp16, x = input_151_cast_fp16)[name = tensor("linear_50_cast_fp16")]; + tensor var_5200_to_fp16 = const()[name = tensor("op_5200_to_fp16"), val = tensor(0x1p+0)]; + tensor var_5201_cast_fp16 = add(x = position12_to_fp16, y = var_5200_to_fp16)[name = tensor("op_5201_cast_fp16")]; + tensor var_5201_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_5201_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor input_153_cast_fp16 = add(x = input_147_cast_fp16, y = linear_50_cast_fp16)[name = tensor("input_153_cast_fp16")]; + tensor input_155_axes_0 = const()[name = tensor("input_155_axes_0"), val = tensor([-1])]; + tensor norm12_2_weight_to_fp16 = const()[name = tensor("norm12_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(311620928)))]; + tensor norm12_2_bias_to_fp16 = const()[name = tensor("norm12_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(311623040)))]; + tensor var_5205_to_fp16 = const()[name = tensor("op_5205_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_155_cast_fp16 = layer_norm(axes = input_155_axes_0, beta = norm12_2_bias_to_fp16, epsilon = var_5205_to_fp16, gamma = norm12_2_weight_to_fp16, x = input_153_cast_fp16)[name = tensor("input_155_cast_fp16")]; + tensor linear12_1_weight_to_fp16 = const()[name = tensor("linear12_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(311625152)))]; + tensor linear_51_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = linear12_1_weight_to_fp16, x = input_155_cast_fp16)[name = tensor("linear_51_cast_fp16")]; + tensor input_157_mode_0 = const()[name = tensor("input_157_mode_0"), val = tensor("EXACT")]; + tensor input_157_cast_fp16 = gelu(mode = input_157_mode_0, x = linear_51_cast_fp16)[name = tensor("input_157_cast_fp16")]; + tensor linear12_2_weight_to_fp16 = const()[name = tensor("linear12_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(320013824)))]; + tensor linear_52_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = linear12_2_weight_to_fp16, x = input_157_cast_fp16)[name = tensor("linear_52_cast_fp16")]; + tensor input_159_cast_fp16 = add(x = input_153_cast_fp16, y = linear_52_cast_fp16)[name = tensor("input_159_cast_fp16")]; + tensor input_161_axes_0 = const()[name = tensor("input_161_axes_0"), val = tensor([-1])]; + tensor norm13_1_weight_to_fp16 = const()[name = tensor("norm13_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(328402496)))]; + tensor norm13_1_bias_to_fp16 = const()[name = tensor("norm13_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(328404608)))]; + tensor var_5222_to_fp16 = const()[name = tensor("op_5222_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_161_cast_fp16 = layer_norm(axes = input_161_axes_0, beta = norm13_1_bias_to_fp16, epsilon = var_5222_to_fp16, gamma = norm13_1_weight_to_fp16, x = input_159_cast_fp16)[name = tensor("input_161_cast_fp16")]; + tensor attn13_in_proj_weight_to_fp16 = const()[name = tensor("attn13_in_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(328406720)))]; + tensor linear_53_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = attn13_in_proj_weight_to_fp16, x = input_161_cast_fp16)[name = tensor("linear_53_cast_fp16")]; + tensor var_5236 = const()[name = tensor("op_5236"), val = tensor([1, 1, 3, 16, 64])]; + tensor qkv_27_cast_fp16 = reshape(shape = var_5236, x = linear_53_cast_fp16)[name = tensor("qkv_27_cast_fp16")]; + tensor q_79_begin_0 = const()[name = tensor("q_79_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor q_79_end_0 = const()[name = tensor("q_79_end_0"), val = tensor([1, 1, 1, 16, 64])]; + tensor q_79_end_mask_0 = const()[name = tensor("q_79_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor q_79_squeeze_mask_0 = const()[name = tensor("q_79_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor q_79_cast_fp16 = slice_by_index(begin = q_79_begin_0, end = q_79_end_0, end_mask = q_79_end_mask_0, squeeze_mask = q_79_squeeze_mask_0, x = qkv_27_cast_fp16)[name = tensor("q_79_cast_fp16")]; + tensor k_53_begin_0 = const()[name = tensor("k_53_begin_0"), val = tensor([0, 0, 1, 0, 0])]; + tensor k_53_end_0 = const()[name = tensor("k_53_end_0"), val = tensor([1, 1, 2, 16, 64])]; + tensor k_53_end_mask_0 = const()[name = tensor("k_53_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor k_53_squeeze_mask_0 = const()[name = tensor("k_53_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor k_53_cast_fp16 = slice_by_index(begin = k_53_begin_0, end = k_53_end_0, end_mask = k_53_end_mask_0, squeeze_mask = k_53_squeeze_mask_0, x = qkv_27_cast_fp16)[name = tensor("k_53_cast_fp16")]; + tensor v_27_begin_0 = const()[name = tensor("v_27_begin_0"), val = tensor([0, 0, 2, 0, 0])]; + tensor v_27_end_0 = const()[name = tensor("v_27_end_0"), val = tensor([1, 1, 3, 16, 64])]; + tensor v_27_end_mask_0 = const()[name = tensor("v_27_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor v_27_squeeze_mask_0 = const()[name = tensor("v_27_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor v_27_cast_fp16 = slice_by_index(begin = v_27_begin_0, end = v_27_end_0, end_mask = v_27_end_mask_0, squeeze_mask = v_27_squeeze_mask_0, x = qkv_27_cast_fp16)[name = tensor("v_27_cast_fp16")]; + tensor var_5302 = const()[name = tensor("op_5302"), val = tensor([1, 1, 1, 1])]; + tensor position13_to_fp16_dtype_0 = const()[name = tensor("position13_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor position13_to_fp16 = cast(dtype = position13_to_fp16_dtype_0, x = position13)[name = tensor("cast_488")]; + tensor ts_83_cast_fp16 = reshape(shape = var_5302, x = position13_to_fp16)[name = tensor("ts_83_cast_fp16")]; + tensor var_5309 = const()[name = tensor("op_5309"), val = tensor([1, 1, 16, 32, 2])]; + tensor q_complex_27_cast_fp16 = reshape(shape = var_5309, x = q_79_cast_fp16)[name = tensor("q_complex_27_cast_fp16")]; + tensor var_5316 = const()[name = tensor("op_5316"), val = tensor([1, 1, 16, 32, 2])]; + tensor k_complex_27_cast_fp16 = reshape(shape = var_5316, x = k_53_cast_fp16)[name = tensor("k_complex_27_cast_fp16")]; + tensor var_5320_begin_0 = const()[name = tensor("op_5320_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_5320_end_0 = const()[name = tensor("op_5320_end_0"), val = tensor([1, 1, 16, 32, 1])]; + tensor var_5320_end_mask_0 = const()[name = tensor("op_5320_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_5320_squeeze_mask_0 = const()[name = tensor("op_5320_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_5320_cast_fp16 = slice_by_index(begin = var_5320_begin_0, end = var_5320_end_0, end_mask = var_5320_end_mask_0, squeeze_mask = var_5320_squeeze_mask_0, x = q_complex_27_cast_fp16)[name = tensor("op_5320_cast_fp16")]; + tensor var_5328_begin_0 = const()[name = tensor("op_5328_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_5328_end_0 = const()[name = tensor("op_5328_end_0"), val = tensor([1, 1, 16, 32, 2])]; + tensor var_5328_end_mask_0 = const()[name = tensor("op_5328_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_5328_squeeze_mask_0 = const()[name = tensor("op_5328_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_5328_cast_fp16 = slice_by_index(begin = var_5328_begin_0, end = var_5328_end_0, end_mask = var_5328_end_mask_0, squeeze_mask = var_5328_squeeze_mask_0, x = q_complex_27_cast_fp16)[name = tensor("op_5328_cast_fp16")]; + tensor var_5336_begin_0 = const()[name = tensor("op_5336_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_5336_end_0 = const()[name = tensor("op_5336_end_0"), val = tensor([1, 1, 16, 32, 1])]; + tensor var_5336_end_mask_0 = const()[name = tensor("op_5336_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_5336_squeeze_mask_0 = const()[name = tensor("op_5336_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_5336_cast_fp16 = slice_by_index(begin = var_5336_begin_0, end = var_5336_end_0, end_mask = var_5336_end_mask_0, squeeze_mask = var_5336_squeeze_mask_0, x = k_complex_27_cast_fp16)[name = tensor("op_5336_cast_fp16")]; + tensor var_5344_begin_0 = const()[name = tensor("op_5344_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_5344_end_0 = const()[name = tensor("op_5344_end_0"), val = tensor([1, 1, 16, 32, 2])]; + tensor var_5344_end_mask_0 = const()[name = tensor("op_5344_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_5344_squeeze_mask_0 = const()[name = tensor("op_5344_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_5344_cast_fp16 = slice_by_index(begin = var_5344_begin_0, end = var_5344_end_0, end_mask = var_5344_end_mask_0, squeeze_mask = var_5344_squeeze_mask_0, x = k_complex_27_cast_fp16)[name = tensor("op_5344_cast_fp16")]; + tensor freqs_27_to_fp16 = const()[name = tensor("freqs_27_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(334698240)))]; + tensor var_5350_cast_fp16 = mul(x = freqs_27_to_fp16, y = ts_83_cast_fp16)[name = tensor("op_5350_cast_fp16")]; + tensor rotr_27_cast_fp16 = cos(x = var_5350_cast_fp16)[name = tensor("rotr_27_cast_fp16")]; + tensor roti_27_cast_fp16 = sin(x = var_5350_cast_fp16)[name = tensor("roti_27_cast_fp16")]; + tensor var_5354_cast_fp16 = mul(x = var_5320_cast_fp16, y = rotr_27_cast_fp16)[name = tensor("op_5354_cast_fp16")]; + tensor var_5355_cast_fp16 = mul(x = var_5328_cast_fp16, y = roti_27_cast_fp16)[name = tensor("op_5355_cast_fp16")]; + tensor qor_53_cast_fp16 = sub(x = var_5354_cast_fp16, y = var_5355_cast_fp16)[name = tensor("qor_53_cast_fp16")]; + tensor var_5358_cast_fp16 = mul(x = var_5320_cast_fp16, y = roti_27_cast_fp16)[name = tensor("op_5358_cast_fp16")]; + tensor var_5359_cast_fp16 = mul(x = var_5328_cast_fp16, y = rotr_27_cast_fp16)[name = tensor("op_5359_cast_fp16")]; + tensor qoi_53_cast_fp16 = add(x = var_5358_cast_fp16, y = var_5359_cast_fp16)[name = tensor("qoi_53_cast_fp16")]; + tensor var_5362_cast_fp16 = mul(x = var_5336_cast_fp16, y = rotr_27_cast_fp16)[name = tensor("op_5362_cast_fp16")]; + tensor var_5363_cast_fp16 = mul(x = var_5344_cast_fp16, y = roti_27_cast_fp16)[name = tensor("op_5363_cast_fp16")]; + tensor kor_53_cast_fp16 = sub(x = var_5362_cast_fp16, y = var_5363_cast_fp16)[name = tensor("kor_53_cast_fp16")]; + tensor var_5366_cast_fp16 = mul(x = var_5336_cast_fp16, y = roti_27_cast_fp16)[name = tensor("op_5366_cast_fp16")]; + tensor var_5367_cast_fp16 = mul(x = var_5344_cast_fp16, y = rotr_27_cast_fp16)[name = tensor("op_5367_cast_fp16")]; + tensor koi_53_cast_fp16 = add(x = var_5366_cast_fp16, y = var_5367_cast_fp16)[name = tensor("koi_53_cast_fp16")]; + tensor qo_27_axis_0 = const()[name = tensor("qo_27_axis_0"), val = tensor(-1)]; + tensor qo_27_cast_fp16 = stack(axis = qo_27_axis_0, values = (qor_53_cast_fp16, qoi_53_cast_fp16))[name = tensor("qo_27_cast_fp16")]; + tensor ko_27_axis_0 = const()[name = tensor("ko_27_axis_0"), val = tensor(-1)]; + tensor ko_27_cast_fp16 = stack(axis = ko_27_axis_0, values = (kor_53_cast_fp16, koi_53_cast_fp16))[name = tensor("ko_27_cast_fp16")]; + tensor var_5400 = const()[name = tensor("op_5400"), val = tensor([1, 1, 16, 64])]; + tensor q_81_cast_fp16 = reshape(shape = var_5400, x = qo_27_cast_fp16)[name = tensor("q_81_cast_fp16")]; + tensor var_5406 = const()[name = tensor("op_5406"), val = tensor([1, 1, 16, 64])]; + tensor k_55_cast_fp16 = reshape(shape = var_5406, x = ko_27_cast_fp16)[name = tensor("k_55_cast_fp16")]; + tensor _inversed_5431_y_0_to_fp16 = const()[name = tensor("_inversed_5431_y_0_to_fp16"), val = tensor(0x1p-9)]; + tensor _inversed_5431_cast_fp16 = mul(x = ts_83_cast_fp16, y = _inversed_5431_y_0_to_fp16)[name = tensor("_inversed_5431_cast_fp16")]; + tensor var_5432_cast_fp16 = floor(x = _inversed_5431_cast_fp16)[name = tensor("op_5432_cast_fp16")]; + tensor var_5433_to_fp16 = const()[name = tensor("op_5433_to_fp16"), val = tensor(0x1p+9)]; + tensor var_5434_cast_fp16 = mul(x = var_5432_cast_fp16, y = var_5433_to_fp16)[name = tensor("op_5434_cast_fp16")]; + tensor write_indices_float_55_cast_fp16 = sub(x = ts_83_cast_fp16, y = var_5434_cast_fp16)[name = tensor("write_indices_float_55_cast_fp16")]; + tensor var_5441_dtype_0 = const()[name = tensor("op_5441_dtype_0"), val = tensor("int32")]; + tensor write_indices_27_reps_0 = const()[name = tensor("write_indices_27_reps_0"), val = tensor([1, 1, 16, 64])]; + tensor write_indices_float_55_cast_fp16_to_int32 = cast(dtype = var_5441_dtype_0, x = write_indices_float_55_cast_fp16)[name = tensor("cast_487")]; + tensor write_indices_27 = tile(reps = write_indices_27_reps_0, x = write_indices_float_55_cast_fp16_to_int32)[name = tensor("write_indices_27")]; + tensor var_5451_begin_0 = const()[name = tensor("op_5451_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_5451_end_0 = const()[name = tensor("op_5451_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor var_5451_end_mask_0 = const()[name = tensor("op_5451_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_5451_squeeze_mask_0 = const()[name = tensor("op_5451_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor cache13_to_fp16_dtype_0 = const()[name = tensor("cache13_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor cache13_to_fp16 = cast(dtype = cache13_to_fp16_dtype_0, x = cache13)[name = tensor("cast_486")]; + tensor var_5451_cast_fp16 = slice_by_index(begin = var_5451_begin_0, end = var_5451_end_0, end_mask = var_5451_end_mask_0, squeeze_mask = var_5451_squeeze_mask_0, x = cache13_to_fp16)[name = tensor("op_5451_cast_fp16")]; + tensor var_5453_axis_0 = const()[name = tensor("op_5453_axis_0"), val = tensor(1)]; + tensor var_5453_mode_0 = const()[name = tensor("op_5453_mode_0"), val = tensor("update")]; + tensor var_5453_validate_indices_0 = const()[name = tensor("op_5453_validate_indices_0"), val = tensor(false)]; + tensor var_5453_cast_fp16 = scatter_along_axis(axis = var_5453_axis_0, data = var_5451_cast_fp16, indices = write_indices_27, mode = var_5453_mode_0, updates = k_55_cast_fp16, validate_indices = var_5453_validate_indices_0)[name = tensor("op_5453_cast_fp16")]; + tensor concat_93 = const()[name = tensor("concat_93"), val = tensor([0, 0, 0, 0, 0])]; + tensor concat_94 = const()[name = tensor("concat_94"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_27_internal_tensor_assign_1_stride_0 = const()[name = tensor("new_cache_27_internal_tensor_assign_1_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_27_internal_tensor_assign_1_begin_mask_0 = const()[name = tensor("new_cache_27_internal_tensor_assign_1_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_27_internal_tensor_assign_1_end_mask_0 = const()[name = tensor("new_cache_27_internal_tensor_assign_1_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_27_internal_tensor_assign_1_squeeze_mask_0 = const()[name = tensor("new_cache_27_internal_tensor_assign_1_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_74 = const()[name = tensor("shape_74"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_26 = const()[name = tensor("reduce_prod_26"), val = tensor(1048576)]; + tensor range_1d_26_start_0 = const()[name = tensor("range_1d_26_start_0"), val = tensor(0)]; + tensor range_1d_26_step_0 = const()[name = tensor("range_1d_26_step_0"), val = tensor(1)]; + tensor range_1d_26 = range_1d(end = reduce_prod_26, start = range_1d_26_start_0, step = range_1d_26_step_0)[name = tensor("range_1d_26")]; + tensor reshape_130 = reshape(shape = shape_74, x = range_1d_26)[name = tensor("reshape_130")]; + tensor slice_by_index_26 = slice_by_index(begin = concat_93, begin_mask = new_cache_27_internal_tensor_assign_1_begin_mask_0, end = concat_94, end_mask = new_cache_27_internal_tensor_assign_1_end_mask_0, squeeze_mask = new_cache_27_internal_tensor_assign_1_squeeze_mask_0, stride = new_cache_27_internal_tensor_assign_1_stride_0, x = reshape_130)[name = tensor("slice_by_index_26")]; + tensor reshape_131_shape_0 = const()[name = tensor("reshape_131_shape_0"), val = tensor([-1])]; + tensor reshape_131 = reshape(shape = reshape_131_shape_0, x = slice_by_index_26)[name = tensor("reshape_131")]; + tensor reshape_132_shape_0 = const()[name = tensor("reshape_132_shape_0"), val = tensor([-1])]; + tensor reshape_132_cast_fp16 = reshape(shape = reshape_132_shape_0, x = var_5453_cast_fp16)[name = tensor("reshape_132_cast_fp16")]; + tensor reshape_133_shape_0 = const()[name = tensor("reshape_133_shape_0"), val = tensor([-1])]; + tensor reshape_133_cast_fp16 = reshape(shape = reshape_133_shape_0, x = cache13_to_fp16)[name = tensor("reshape_133_cast_fp16")]; + tensor scatter_26_mode_0 = const()[name = tensor("scatter_26_mode_0"), val = tensor("update")]; + tensor scatter_26_axis_0 = const()[name = tensor("scatter_26_axis_0"), val = tensor(0)]; + tensor scatter_26_validate_indices_0 = const()[name = tensor("scatter_26_validate_indices_0"), val = tensor(false)]; + tensor scatter_26_cast_fp16 = scatter(axis = scatter_26_axis_0, data = reshape_133_cast_fp16, indices = reshape_131, mode = scatter_26_mode_0, updates = reshape_132_cast_fp16, validate_indices = scatter_26_validate_indices_0)[name = tensor("scatter_26_cast_fp16")]; + tensor reshape_134_cast_fp16 = reshape(shape = shape_74, x = scatter_26_cast_fp16)[name = tensor("reshape_134_cast_fp16")]; + tensor var_5461_begin_0 = const()[name = tensor("op_5461_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor var_5461_end_0 = const()[name = tensor("op_5461_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor var_5461_end_mask_0 = const()[name = tensor("op_5461_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_5461_squeeze_mask_0 = const()[name = tensor("op_5461_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor var_5461_cast_fp16 = slice_by_index(begin = var_5461_begin_0, end = var_5461_end_0, end_mask = var_5461_end_mask_0, squeeze_mask = var_5461_squeeze_mask_0, x = reshape_134_cast_fp16)[name = tensor("op_5461_cast_fp16")]; + tensor var_5463_axis_0 = const()[name = tensor("op_5463_axis_0"), val = tensor(1)]; + tensor var_5463_mode_0 = const()[name = tensor("op_5463_mode_0"), val = tensor("update")]; + tensor var_5463_validate_indices_0 = const()[name = tensor("op_5463_validate_indices_0"), val = tensor(false)]; + tensor var_5463_cast_fp16 = scatter_along_axis(axis = var_5463_axis_0, data = var_5461_cast_fp16, indices = write_indices_27, mode = var_5463_mode_0, updates = v_27_cast_fp16, validate_indices = var_5463_validate_indices_0)[name = tensor("op_5463_cast_fp16")]; + tensor concat_95 = const()[name = tensor("concat_95"), val = tensor([1, 0, 0, 0, 0])]; + tensor concat_96 = const()[name = tensor("concat_96"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_27_internal_tensor_assign_2_stride_0 = const()[name = tensor("new_cache_27_internal_tensor_assign_2_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_27_internal_tensor_assign_2_begin_mask_0 = const()[name = tensor("new_cache_27_internal_tensor_assign_2_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_27_internal_tensor_assign_2_end_mask_0 = const()[name = tensor("new_cache_27_internal_tensor_assign_2_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_27_internal_tensor_assign_2_squeeze_mask_0 = const()[name = tensor("new_cache_27_internal_tensor_assign_2_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_75 = const()[name = tensor("shape_75"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_27 = const()[name = tensor("reduce_prod_27"), val = tensor(1048576)]; + tensor range_1d_27_start_0 = const()[name = tensor("range_1d_27_start_0"), val = tensor(0)]; + tensor range_1d_27_step_0 = const()[name = tensor("range_1d_27_step_0"), val = tensor(1)]; + tensor range_1d_27 = range_1d(end = reduce_prod_27, start = range_1d_27_start_0, step = range_1d_27_step_0)[name = tensor("range_1d_27")]; + tensor reshape_135 = reshape(shape = shape_75, x = range_1d_27)[name = tensor("reshape_135")]; + tensor slice_by_index_27 = slice_by_index(begin = concat_95, begin_mask = new_cache_27_internal_tensor_assign_2_begin_mask_0, end = concat_96, end_mask = new_cache_27_internal_tensor_assign_2_end_mask_0, squeeze_mask = new_cache_27_internal_tensor_assign_2_squeeze_mask_0, stride = new_cache_27_internal_tensor_assign_2_stride_0, x = reshape_135)[name = tensor("slice_by_index_27")]; + tensor reshape_136_shape_0 = const()[name = tensor("reshape_136_shape_0"), val = tensor([-1])]; + tensor reshape_136 = reshape(shape = reshape_136_shape_0, x = slice_by_index_27)[name = tensor("reshape_136")]; + tensor reshape_137_shape_0 = const()[name = tensor("reshape_137_shape_0"), val = tensor([-1])]; + tensor reshape_137_cast_fp16 = reshape(shape = reshape_137_shape_0, x = var_5463_cast_fp16)[name = tensor("reshape_137_cast_fp16")]; + tensor reshape_138_shape_0 = const()[name = tensor("reshape_138_shape_0"), val = tensor([-1])]; + tensor reshape_138_cast_fp16 = reshape(shape = reshape_138_shape_0, x = reshape_134_cast_fp16)[name = tensor("reshape_138_cast_fp16")]; + tensor scatter_27_mode_0 = const()[name = tensor("scatter_27_mode_0"), val = tensor("update")]; + tensor scatter_27_axis_0 = const()[name = tensor("scatter_27_axis_0"), val = tensor(0)]; + tensor scatter_27_validate_indices_0 = const()[name = tensor("scatter_27_validate_indices_0"), val = tensor(false)]; + tensor scatter_27_cast_fp16 = scatter(axis = scatter_27_axis_0, data = reshape_138_cast_fp16, indices = reshape_136, mode = scatter_27_mode_0, updates = reshape_137_cast_fp16, validate_indices = scatter_27_validate_indices_0)[name = tensor("scatter_27_cast_fp16")]; + tensor reshape_139_cast_fp16 = reshape(shape = shape_75, x = scatter_27_cast_fp16)[name = tensor("reshape_139_cast_fp16")]; + tensor reshape_139_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("reshape_139_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor keys_79_begin_0 = const()[name = tensor("keys_79_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor keys_79_end_0 = const()[name = tensor("keys_79_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor keys_79_end_mask_0 = const()[name = tensor("keys_79_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor keys_79_squeeze_mask_0 = const()[name = tensor("keys_79_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor keys_79_cast_fp16 = slice_by_index(begin = keys_79_begin_0, end = keys_79_end_0, end_mask = keys_79_end_mask_0, squeeze_mask = keys_79_squeeze_mask_0, x = reshape_139_cast_fp16)[name = tensor("keys_79_cast_fp16")]; + tensor values_79_begin_0 = const()[name = tensor("values_79_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor values_79_end_0 = const()[name = tensor("values_79_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor values_79_end_mask_0 = const()[name = tensor("values_79_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor values_79_squeeze_mask_0 = const()[name = tensor("values_79_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor values_79_cast_fp16 = slice_by_index(begin = values_79_begin_0, end = values_79_end_0, end_mask = values_79_end_mask_0, squeeze_mask = values_79_squeeze_mask_0, x = reshape_139_cast_fp16)[name = tensor("values_79_cast_fp16")]; + tensor var_5475_cast_fp16 = not_equal(x = keys_79_cast_fp16, y = keys_79_cast_fp16)[name = tensor("op_5475_cast_fp16")]; + tensor keys_81_cast_fp16 = select(a = var_463_to_fp16, b = keys_79_cast_fp16, cond = var_5475_cast_fp16)[name = tensor("keys_81_cast_fp16")]; + tensor var_5483_cast_fp16 = not_equal(x = values_79_cast_fp16, y = values_79_cast_fp16)[name = tensor("op_5483_cast_fp16")]; + tensor values_81_cast_fp16 = select(a = var_463_to_fp16, b = values_79_cast_fp16, cond = var_5483_cast_fp16)[name = tensor("values_81_cast_fp16")]; + tensor var_5507 = const()[name = tensor("op_5507"), val = tensor([0, 2, 1, 3])]; + tensor var_5523 = const()[name = tensor("op_5523"), val = tensor([1, 1, 1])]; + tensor var_5524_cast_fp16 = reshape(shape = var_5523, x = position13_to_fp16)[name = tensor("op_5524_cast_fp16")]; + tensor var_5544_to_fp16 = const()[name = tensor("op_5544_to_fp16"), val = tensor(0x1p+0)]; + tensor valid_len_27_cast_fp16 = add(x = var_5524_cast_fp16, y = var_5544_to_fp16)[name = tensor("valid_len_27_cast_fp16")]; + tensor valid_mask_27_cast_fp16 = less(x = k_positions_1_promoted_to_fp16, y = valid_len_27_cast_fp16)[name = tensor("valid_mask_27_cast_fp16")]; + tensor causal_mask_27_cast_fp16 = less_equal(x = k_positions_1_promoted_to_fp16, y = var_5524_cast_fp16)[name = tensor("causal_mask_27_cast_fp16")]; + tensor attn_mask_53 = logical_and(x = valid_mask_27_cast_fp16, y = causal_mask_27_cast_fp16)[name = tensor("attn_mask_53")]; + tensor attn_mask_55_axes_0 = const()[name = tensor("attn_mask_55_axes_0"), val = tensor([1])]; + tensor attn_mask_55 = expand_dims(axes = attn_mask_55_axes_0, x = attn_mask_53)[name = tensor("attn_mask_55")]; + tensor var_5562_transpose_x_0 = const()[name = tensor("op_5562_transpose_x_0"), val = tensor(false)]; + tensor var_5562_transpose_y_0 = const()[name = tensor("op_5562_transpose_y_0"), val = tensor(false)]; + tensor transpose_98_perm_0 = const()[name = tensor("transpose_98_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_99_perm_0 = const()[name = tensor("transpose_99_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_99 = transpose(perm = transpose_99_perm_0, x = keys_81_cast_fp16)[name = tensor("transpose_161")]; + tensor transpose_98 = transpose(perm = transpose_98_perm_0, x = q_81_cast_fp16)[name = tensor("transpose_162")]; + tensor var_5562_cast_fp16 = matmul(transpose_x = var_5562_transpose_x_0, transpose_y = var_5562_transpose_y_0, x = transpose_98, y = transpose_99)[name = tensor("op_5562_cast_fp16")]; + tensor var_5556_to_fp16 = const()[name = tensor("op_5556_to_fp16"), val = tensor([0x1p-3])]; + tensor attn_weights_79_cast_fp16 = mul(x = var_5562_cast_fp16, y = var_5556_to_fp16)[name = tensor("attn_weights_79_cast_fp16")]; + tensor var_5564 = logical_not(x = attn_mask_55)[name = tensor("op_5564")]; + tensor var_5565_to_fp16 = const()[name = tensor("op_5565_to_fp16"), val = tensor(-inf)]; + tensor attn_weights_81_cast_fp16 = select(a = var_5565_to_fp16, b = attn_weights_79_cast_fp16, cond = var_5564)[name = tensor("attn_weights_81_cast_fp16")]; + tensor var_5567 = const()[name = tensor("op_5567"), val = tensor(-1)]; + tensor attn_weights_83_cast_fp16 = softmax(axis = var_5567, x = attn_weights_81_cast_fp16)[name = tensor("attn_weights_83_cast_fp16")]; + tensor attn_output_27_transpose_x_0 = const()[name = tensor("attn_output_27_transpose_x_0"), val = tensor(false)]; + tensor attn_output_27_transpose_y_0 = const()[name = tensor("attn_output_27_transpose_y_0"), val = tensor(false)]; + tensor values_83_cast_fp16 = transpose(perm = var_5507, x = values_81_cast_fp16)[name = tensor("transpose_163")]; + tensor attn_output_27_cast_fp16 = matmul(transpose_x = attn_output_27_transpose_x_0, transpose_y = attn_output_27_transpose_y_0, x = attn_weights_83_cast_fp16, y = values_83_cast_fp16)[name = tensor("attn_output_27_cast_fp16")]; + tensor var_5575 = const()[name = tensor("op_5575"), val = tensor([0, 2, 1, 3])]; + tensor var_5580 = const()[name = tensor("op_5580"), val = tensor([1, 1, 1024])]; + tensor var_5576_cast_fp16 = transpose(perm = var_5575, x = attn_output_27_cast_fp16)[name = tensor("transpose_160")]; + tensor input_163_cast_fp16 = reshape(shape = var_5580, x = var_5576_cast_fp16)[name = tensor("input_163_cast_fp16")]; + tensor attn13_out_proj_weight_to_fp16 = const()[name = tensor("attn13_out_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(334698368)))]; + tensor linear_54_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = attn13_out_proj_weight_to_fp16, x = input_163_cast_fp16)[name = tensor("linear_54_cast_fp16")]; + tensor var_5586_to_fp16 = const()[name = tensor("op_5586_to_fp16"), val = tensor(0x1p+0)]; + tensor var_5587_cast_fp16 = add(x = position13_to_fp16, y = var_5586_to_fp16)[name = tensor("op_5587_cast_fp16")]; + tensor var_5587_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_5587_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor input_165_cast_fp16 = add(x = input_159_cast_fp16, y = linear_54_cast_fp16)[name = tensor("input_165_cast_fp16")]; + tensor input_167_axes_0 = const()[name = tensor("input_167_axes_0"), val = tensor([-1])]; + tensor norm13_2_weight_to_fp16 = const()[name = tensor("norm13_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(336795584)))]; + tensor norm13_2_bias_to_fp16 = const()[name = tensor("norm13_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(336797696)))]; + tensor var_5591_to_fp16 = const()[name = tensor("op_5591_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_167_cast_fp16 = layer_norm(axes = input_167_axes_0, beta = norm13_2_bias_to_fp16, epsilon = var_5591_to_fp16, gamma = norm13_2_weight_to_fp16, x = input_165_cast_fp16)[name = tensor("input_167_cast_fp16")]; + tensor linear13_1_weight_to_fp16 = const()[name = tensor("linear13_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(336799808)))]; + tensor linear_55_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = linear13_1_weight_to_fp16, x = input_167_cast_fp16)[name = tensor("linear_55_cast_fp16")]; + tensor input_169_mode_0 = const()[name = tensor("input_169_mode_0"), val = tensor("EXACT")]; + tensor input_169_cast_fp16 = gelu(mode = input_169_mode_0, x = linear_55_cast_fp16)[name = tensor("input_169_cast_fp16")]; + tensor linear13_2_weight_to_fp16 = const()[name = tensor("linear13_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(345188480)))]; + tensor linear_56_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = linear13_2_weight_to_fp16, x = input_169_cast_fp16)[name = tensor("linear_56_cast_fp16")]; + tensor input_171_cast_fp16 = add(x = input_165_cast_fp16, y = linear_56_cast_fp16)[name = tensor("input_171_cast_fp16")]; + tensor input_173_axes_0 = const()[name = tensor("input_173_axes_0"), val = tensor([-1])]; + tensor norm14_1_weight_to_fp16 = const()[name = tensor("norm14_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(353577152)))]; + tensor norm14_1_bias_to_fp16 = const()[name = tensor("norm14_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(353579264)))]; + tensor var_5608_to_fp16 = const()[name = tensor("op_5608_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_173_cast_fp16 = layer_norm(axes = input_173_axes_0, beta = norm14_1_bias_to_fp16, epsilon = var_5608_to_fp16, gamma = norm14_1_weight_to_fp16, x = input_171_cast_fp16)[name = tensor("input_173_cast_fp16")]; + tensor attn14_in_proj_weight_to_fp16 = const()[name = tensor("attn14_in_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(353581376)))]; + tensor linear_57_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = attn14_in_proj_weight_to_fp16, x = input_173_cast_fp16)[name = tensor("linear_57_cast_fp16")]; + tensor var_5622 = const()[name = tensor("op_5622"), val = tensor([1, 1, 3, 16, 64])]; + tensor qkv_29_cast_fp16 = reshape(shape = var_5622, x = linear_57_cast_fp16)[name = tensor("qkv_29_cast_fp16")]; + tensor q_85_begin_0 = const()[name = tensor("q_85_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor q_85_end_0 = const()[name = tensor("q_85_end_0"), val = tensor([1, 1, 1, 16, 64])]; + tensor q_85_end_mask_0 = const()[name = tensor("q_85_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor q_85_squeeze_mask_0 = const()[name = tensor("q_85_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor q_85_cast_fp16 = slice_by_index(begin = q_85_begin_0, end = q_85_end_0, end_mask = q_85_end_mask_0, squeeze_mask = q_85_squeeze_mask_0, x = qkv_29_cast_fp16)[name = tensor("q_85_cast_fp16")]; + tensor k_57_begin_0 = const()[name = tensor("k_57_begin_0"), val = tensor([0, 0, 1, 0, 0])]; + tensor k_57_end_0 = const()[name = tensor("k_57_end_0"), val = tensor([1, 1, 2, 16, 64])]; + tensor k_57_end_mask_0 = const()[name = tensor("k_57_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor k_57_squeeze_mask_0 = const()[name = tensor("k_57_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor k_57_cast_fp16 = slice_by_index(begin = k_57_begin_0, end = k_57_end_0, end_mask = k_57_end_mask_0, squeeze_mask = k_57_squeeze_mask_0, x = qkv_29_cast_fp16)[name = tensor("k_57_cast_fp16")]; + tensor v_29_begin_0 = const()[name = tensor("v_29_begin_0"), val = tensor([0, 0, 2, 0, 0])]; + tensor v_29_end_0 = const()[name = tensor("v_29_end_0"), val = tensor([1, 1, 3, 16, 64])]; + tensor v_29_end_mask_0 = const()[name = tensor("v_29_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor v_29_squeeze_mask_0 = const()[name = tensor("v_29_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor v_29_cast_fp16 = slice_by_index(begin = v_29_begin_0, end = v_29_end_0, end_mask = v_29_end_mask_0, squeeze_mask = v_29_squeeze_mask_0, x = qkv_29_cast_fp16)[name = tensor("v_29_cast_fp16")]; + tensor var_5688 = const()[name = tensor("op_5688"), val = tensor([1, 1, 1, 1])]; + tensor position14_to_fp16_dtype_0 = const()[name = tensor("position14_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor position14_to_fp16 = cast(dtype = position14_to_fp16_dtype_0, x = position14)[name = tensor("cast_483")]; + tensor ts_89_cast_fp16 = reshape(shape = var_5688, x = position14_to_fp16)[name = tensor("ts_89_cast_fp16")]; + tensor var_5695 = const()[name = tensor("op_5695"), val = tensor([1, 1, 16, 32, 2])]; + tensor q_complex_29_cast_fp16 = reshape(shape = var_5695, x = q_85_cast_fp16)[name = tensor("q_complex_29_cast_fp16")]; + tensor var_5702 = const()[name = tensor("op_5702"), val = tensor([1, 1, 16, 32, 2])]; + tensor k_complex_29_cast_fp16 = reshape(shape = var_5702, x = k_57_cast_fp16)[name = tensor("k_complex_29_cast_fp16")]; + tensor var_5706_begin_0 = const()[name = tensor("op_5706_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_5706_end_0 = const()[name = tensor("op_5706_end_0"), val = tensor([1, 1, 16, 32, 1])]; + tensor var_5706_end_mask_0 = const()[name = tensor("op_5706_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_5706_squeeze_mask_0 = const()[name = tensor("op_5706_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_5706_cast_fp16 = slice_by_index(begin = var_5706_begin_0, end = var_5706_end_0, end_mask = var_5706_end_mask_0, squeeze_mask = var_5706_squeeze_mask_0, x = q_complex_29_cast_fp16)[name = tensor("op_5706_cast_fp16")]; + tensor var_5714_begin_0 = const()[name = tensor("op_5714_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_5714_end_0 = const()[name = tensor("op_5714_end_0"), val = tensor([1, 1, 16, 32, 2])]; + tensor var_5714_end_mask_0 = const()[name = tensor("op_5714_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_5714_squeeze_mask_0 = const()[name = tensor("op_5714_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_5714_cast_fp16 = slice_by_index(begin = var_5714_begin_0, end = var_5714_end_0, end_mask = var_5714_end_mask_0, squeeze_mask = var_5714_squeeze_mask_0, x = q_complex_29_cast_fp16)[name = tensor("op_5714_cast_fp16")]; + tensor var_5722_begin_0 = const()[name = tensor("op_5722_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_5722_end_0 = const()[name = tensor("op_5722_end_0"), val = tensor([1, 1, 16, 32, 1])]; + tensor var_5722_end_mask_0 = const()[name = tensor("op_5722_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_5722_squeeze_mask_0 = const()[name = tensor("op_5722_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_5722_cast_fp16 = slice_by_index(begin = var_5722_begin_0, end = var_5722_end_0, end_mask = var_5722_end_mask_0, squeeze_mask = var_5722_squeeze_mask_0, x = k_complex_29_cast_fp16)[name = tensor("op_5722_cast_fp16")]; + tensor var_5730_begin_0 = const()[name = tensor("op_5730_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_5730_end_0 = const()[name = tensor("op_5730_end_0"), val = tensor([1, 1, 16, 32, 2])]; + tensor var_5730_end_mask_0 = const()[name = tensor("op_5730_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_5730_squeeze_mask_0 = const()[name = tensor("op_5730_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_5730_cast_fp16 = slice_by_index(begin = var_5730_begin_0, end = var_5730_end_0, end_mask = var_5730_end_mask_0, squeeze_mask = var_5730_squeeze_mask_0, x = k_complex_29_cast_fp16)[name = tensor("op_5730_cast_fp16")]; + tensor freqs_29_to_fp16 = const()[name = tensor("freqs_29_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(359872896)))]; + tensor var_5736_cast_fp16 = mul(x = freqs_29_to_fp16, y = ts_89_cast_fp16)[name = tensor("op_5736_cast_fp16")]; + tensor rotr_29_cast_fp16 = cos(x = var_5736_cast_fp16)[name = tensor("rotr_29_cast_fp16")]; + tensor roti_29_cast_fp16 = sin(x = var_5736_cast_fp16)[name = tensor("roti_29_cast_fp16")]; + tensor var_5740_cast_fp16 = mul(x = var_5706_cast_fp16, y = rotr_29_cast_fp16)[name = tensor("op_5740_cast_fp16")]; + tensor var_5741_cast_fp16 = mul(x = var_5714_cast_fp16, y = roti_29_cast_fp16)[name = tensor("op_5741_cast_fp16")]; + tensor qor_57_cast_fp16 = sub(x = var_5740_cast_fp16, y = var_5741_cast_fp16)[name = tensor("qor_57_cast_fp16")]; + tensor var_5744_cast_fp16 = mul(x = var_5706_cast_fp16, y = roti_29_cast_fp16)[name = tensor("op_5744_cast_fp16")]; + tensor var_5745_cast_fp16 = mul(x = var_5714_cast_fp16, y = rotr_29_cast_fp16)[name = tensor("op_5745_cast_fp16")]; + tensor qoi_57_cast_fp16 = add(x = var_5744_cast_fp16, y = var_5745_cast_fp16)[name = tensor("qoi_57_cast_fp16")]; + tensor var_5748_cast_fp16 = mul(x = var_5722_cast_fp16, y = rotr_29_cast_fp16)[name = tensor("op_5748_cast_fp16")]; + tensor var_5749_cast_fp16 = mul(x = var_5730_cast_fp16, y = roti_29_cast_fp16)[name = tensor("op_5749_cast_fp16")]; + tensor kor_57_cast_fp16 = sub(x = var_5748_cast_fp16, y = var_5749_cast_fp16)[name = tensor("kor_57_cast_fp16")]; + tensor var_5752_cast_fp16 = mul(x = var_5722_cast_fp16, y = roti_29_cast_fp16)[name = tensor("op_5752_cast_fp16")]; + tensor var_5753_cast_fp16 = mul(x = var_5730_cast_fp16, y = rotr_29_cast_fp16)[name = tensor("op_5753_cast_fp16")]; + tensor koi_57_cast_fp16 = add(x = var_5752_cast_fp16, y = var_5753_cast_fp16)[name = tensor("koi_57_cast_fp16")]; + tensor qo_29_axis_0 = const()[name = tensor("qo_29_axis_0"), val = tensor(-1)]; + tensor qo_29_cast_fp16 = stack(axis = qo_29_axis_0, values = (qor_57_cast_fp16, qoi_57_cast_fp16))[name = tensor("qo_29_cast_fp16")]; + tensor ko_29_axis_0 = const()[name = tensor("ko_29_axis_0"), val = tensor(-1)]; + tensor ko_29_cast_fp16 = stack(axis = ko_29_axis_0, values = (kor_57_cast_fp16, koi_57_cast_fp16))[name = tensor("ko_29_cast_fp16")]; + tensor var_5786 = const()[name = tensor("op_5786"), val = tensor([1, 1, 16, 64])]; + tensor q_87_cast_fp16 = reshape(shape = var_5786, x = qo_29_cast_fp16)[name = tensor("q_87_cast_fp16")]; + tensor var_5792 = const()[name = tensor("op_5792"), val = tensor([1, 1, 16, 64])]; + tensor k_59_cast_fp16 = reshape(shape = var_5792, x = ko_29_cast_fp16)[name = tensor("k_59_cast_fp16")]; + tensor _inversed_5817_y_0_to_fp16 = const()[name = tensor("_inversed_5817_y_0_to_fp16"), val = tensor(0x1p-9)]; + tensor _inversed_5817_cast_fp16 = mul(x = ts_89_cast_fp16, y = _inversed_5817_y_0_to_fp16)[name = tensor("_inversed_5817_cast_fp16")]; + tensor var_5818_cast_fp16 = floor(x = _inversed_5817_cast_fp16)[name = tensor("op_5818_cast_fp16")]; + tensor var_5819_to_fp16 = const()[name = tensor("op_5819_to_fp16"), val = tensor(0x1p+9)]; + tensor var_5820_cast_fp16 = mul(x = var_5818_cast_fp16, y = var_5819_to_fp16)[name = tensor("op_5820_cast_fp16")]; + tensor write_indices_float_59_cast_fp16 = sub(x = ts_89_cast_fp16, y = var_5820_cast_fp16)[name = tensor("write_indices_float_59_cast_fp16")]; + tensor var_5827_dtype_0 = const()[name = tensor("op_5827_dtype_0"), val = tensor("int32")]; + tensor write_indices_29_reps_0 = const()[name = tensor("write_indices_29_reps_0"), val = tensor([1, 1, 16, 64])]; + tensor write_indices_float_59_cast_fp16_to_int32 = cast(dtype = var_5827_dtype_0, x = write_indices_float_59_cast_fp16)[name = tensor("cast_482")]; + tensor write_indices_29 = tile(reps = write_indices_29_reps_0, x = write_indices_float_59_cast_fp16_to_int32)[name = tensor("write_indices_29")]; + tensor var_5837_begin_0 = const()[name = tensor("op_5837_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_5837_end_0 = const()[name = tensor("op_5837_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor var_5837_end_mask_0 = const()[name = tensor("op_5837_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_5837_squeeze_mask_0 = const()[name = tensor("op_5837_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor cache14_to_fp16_dtype_0 = const()[name = tensor("cache14_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor cache14_to_fp16 = cast(dtype = cache14_to_fp16_dtype_0, x = cache14)[name = tensor("cast_481")]; + tensor var_5837_cast_fp16 = slice_by_index(begin = var_5837_begin_0, end = var_5837_end_0, end_mask = var_5837_end_mask_0, squeeze_mask = var_5837_squeeze_mask_0, x = cache14_to_fp16)[name = tensor("op_5837_cast_fp16")]; + tensor var_5839_axis_0 = const()[name = tensor("op_5839_axis_0"), val = tensor(1)]; + tensor var_5839_mode_0 = const()[name = tensor("op_5839_mode_0"), val = tensor("update")]; + tensor var_5839_validate_indices_0 = const()[name = tensor("op_5839_validate_indices_0"), val = tensor(false)]; + tensor var_5839_cast_fp16 = scatter_along_axis(axis = var_5839_axis_0, data = var_5837_cast_fp16, indices = write_indices_29, mode = var_5839_mode_0, updates = k_59_cast_fp16, validate_indices = var_5839_validate_indices_0)[name = tensor("op_5839_cast_fp16")]; + tensor concat_100 = const()[name = tensor("concat_100"), val = tensor([0, 0, 0, 0, 0])]; + tensor concat_101 = const()[name = tensor("concat_101"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_29_internal_tensor_assign_1_stride_0 = const()[name = tensor("new_cache_29_internal_tensor_assign_1_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_29_internal_tensor_assign_1_begin_mask_0 = const()[name = tensor("new_cache_29_internal_tensor_assign_1_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_29_internal_tensor_assign_1_end_mask_0 = const()[name = tensor("new_cache_29_internal_tensor_assign_1_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_29_internal_tensor_assign_1_squeeze_mask_0 = const()[name = tensor("new_cache_29_internal_tensor_assign_1_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_76 = const()[name = tensor("shape_76"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_28 = const()[name = tensor("reduce_prod_28"), val = tensor(1048576)]; + tensor range_1d_28_start_0 = const()[name = tensor("range_1d_28_start_0"), val = tensor(0)]; + tensor range_1d_28_step_0 = const()[name = tensor("range_1d_28_step_0"), val = tensor(1)]; + tensor range_1d_28 = range_1d(end = reduce_prod_28, start = range_1d_28_start_0, step = range_1d_28_step_0)[name = tensor("range_1d_28")]; + tensor reshape_140 = reshape(shape = shape_76, x = range_1d_28)[name = tensor("reshape_140")]; + tensor slice_by_index_28 = slice_by_index(begin = concat_100, begin_mask = new_cache_29_internal_tensor_assign_1_begin_mask_0, end = concat_101, end_mask = new_cache_29_internal_tensor_assign_1_end_mask_0, squeeze_mask = new_cache_29_internal_tensor_assign_1_squeeze_mask_0, stride = new_cache_29_internal_tensor_assign_1_stride_0, x = reshape_140)[name = tensor("slice_by_index_28")]; + tensor reshape_141_shape_0 = const()[name = tensor("reshape_141_shape_0"), val = tensor([-1])]; + tensor reshape_141 = reshape(shape = reshape_141_shape_0, x = slice_by_index_28)[name = tensor("reshape_141")]; + tensor reshape_142_shape_0 = const()[name = tensor("reshape_142_shape_0"), val = tensor([-1])]; + tensor reshape_142_cast_fp16 = reshape(shape = reshape_142_shape_0, x = var_5839_cast_fp16)[name = tensor("reshape_142_cast_fp16")]; + tensor reshape_143_shape_0 = const()[name = tensor("reshape_143_shape_0"), val = tensor([-1])]; + tensor reshape_143_cast_fp16 = reshape(shape = reshape_143_shape_0, x = cache14_to_fp16)[name = tensor("reshape_143_cast_fp16")]; + tensor scatter_28_mode_0 = const()[name = tensor("scatter_28_mode_0"), val = tensor("update")]; + tensor scatter_28_axis_0 = const()[name = tensor("scatter_28_axis_0"), val = tensor(0)]; + tensor scatter_28_validate_indices_0 = const()[name = tensor("scatter_28_validate_indices_0"), val = tensor(false)]; + tensor scatter_28_cast_fp16 = scatter(axis = scatter_28_axis_0, data = reshape_143_cast_fp16, indices = reshape_141, mode = scatter_28_mode_0, updates = reshape_142_cast_fp16, validate_indices = scatter_28_validate_indices_0)[name = tensor("scatter_28_cast_fp16")]; + tensor reshape_144_cast_fp16 = reshape(shape = shape_76, x = scatter_28_cast_fp16)[name = tensor("reshape_144_cast_fp16")]; + tensor var_5847_begin_0 = const()[name = tensor("op_5847_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor var_5847_end_0 = const()[name = tensor("op_5847_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor var_5847_end_mask_0 = const()[name = tensor("op_5847_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_5847_squeeze_mask_0 = const()[name = tensor("op_5847_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor var_5847_cast_fp16 = slice_by_index(begin = var_5847_begin_0, end = var_5847_end_0, end_mask = var_5847_end_mask_0, squeeze_mask = var_5847_squeeze_mask_0, x = reshape_144_cast_fp16)[name = tensor("op_5847_cast_fp16")]; + tensor var_5849_axis_0 = const()[name = tensor("op_5849_axis_0"), val = tensor(1)]; + tensor var_5849_mode_0 = const()[name = tensor("op_5849_mode_0"), val = tensor("update")]; + tensor var_5849_validate_indices_0 = const()[name = tensor("op_5849_validate_indices_0"), val = tensor(false)]; + tensor var_5849_cast_fp16 = scatter_along_axis(axis = var_5849_axis_0, data = var_5847_cast_fp16, indices = write_indices_29, mode = var_5849_mode_0, updates = v_29_cast_fp16, validate_indices = var_5849_validate_indices_0)[name = tensor("op_5849_cast_fp16")]; + tensor concat_102 = const()[name = tensor("concat_102"), val = tensor([1, 0, 0, 0, 0])]; + tensor concat_103 = const()[name = tensor("concat_103"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_29_internal_tensor_assign_2_stride_0 = const()[name = tensor("new_cache_29_internal_tensor_assign_2_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_29_internal_tensor_assign_2_begin_mask_0 = const()[name = tensor("new_cache_29_internal_tensor_assign_2_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_29_internal_tensor_assign_2_end_mask_0 = const()[name = tensor("new_cache_29_internal_tensor_assign_2_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_29_internal_tensor_assign_2_squeeze_mask_0 = const()[name = tensor("new_cache_29_internal_tensor_assign_2_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_77 = const()[name = tensor("shape_77"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_29 = const()[name = tensor("reduce_prod_29"), val = tensor(1048576)]; + tensor range_1d_29_start_0 = const()[name = tensor("range_1d_29_start_0"), val = tensor(0)]; + tensor range_1d_29_step_0 = const()[name = tensor("range_1d_29_step_0"), val = tensor(1)]; + tensor range_1d_29 = range_1d(end = reduce_prod_29, start = range_1d_29_start_0, step = range_1d_29_step_0)[name = tensor("range_1d_29")]; + tensor reshape_145 = reshape(shape = shape_77, x = range_1d_29)[name = tensor("reshape_145")]; + tensor slice_by_index_29 = slice_by_index(begin = concat_102, begin_mask = new_cache_29_internal_tensor_assign_2_begin_mask_0, end = concat_103, end_mask = new_cache_29_internal_tensor_assign_2_end_mask_0, squeeze_mask = new_cache_29_internal_tensor_assign_2_squeeze_mask_0, stride = new_cache_29_internal_tensor_assign_2_stride_0, x = reshape_145)[name = tensor("slice_by_index_29")]; + tensor reshape_146_shape_0 = const()[name = tensor("reshape_146_shape_0"), val = tensor([-1])]; + tensor reshape_146 = reshape(shape = reshape_146_shape_0, x = slice_by_index_29)[name = tensor("reshape_146")]; + tensor reshape_147_shape_0 = const()[name = tensor("reshape_147_shape_0"), val = tensor([-1])]; + tensor reshape_147_cast_fp16 = reshape(shape = reshape_147_shape_0, x = var_5849_cast_fp16)[name = tensor("reshape_147_cast_fp16")]; + tensor reshape_148_shape_0 = const()[name = tensor("reshape_148_shape_0"), val = tensor([-1])]; + tensor reshape_148_cast_fp16 = reshape(shape = reshape_148_shape_0, x = reshape_144_cast_fp16)[name = tensor("reshape_148_cast_fp16")]; + tensor scatter_29_mode_0 = const()[name = tensor("scatter_29_mode_0"), val = tensor("update")]; + tensor scatter_29_axis_0 = const()[name = tensor("scatter_29_axis_0"), val = tensor(0)]; + tensor scatter_29_validate_indices_0 = const()[name = tensor("scatter_29_validate_indices_0"), val = tensor(false)]; + tensor scatter_29_cast_fp16 = scatter(axis = scatter_29_axis_0, data = reshape_148_cast_fp16, indices = reshape_146, mode = scatter_29_mode_0, updates = reshape_147_cast_fp16, validate_indices = scatter_29_validate_indices_0)[name = tensor("scatter_29_cast_fp16")]; + tensor reshape_149_cast_fp16 = reshape(shape = shape_77, x = scatter_29_cast_fp16)[name = tensor("reshape_149_cast_fp16")]; + tensor reshape_149_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("reshape_149_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor keys_85_begin_0 = const()[name = tensor("keys_85_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor keys_85_end_0 = const()[name = tensor("keys_85_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor keys_85_end_mask_0 = const()[name = tensor("keys_85_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor keys_85_squeeze_mask_0 = const()[name = tensor("keys_85_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor keys_85_cast_fp16 = slice_by_index(begin = keys_85_begin_0, end = keys_85_end_0, end_mask = keys_85_end_mask_0, squeeze_mask = keys_85_squeeze_mask_0, x = reshape_149_cast_fp16)[name = tensor("keys_85_cast_fp16")]; + tensor values_85_begin_0 = const()[name = tensor("values_85_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor values_85_end_0 = const()[name = tensor("values_85_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor values_85_end_mask_0 = const()[name = tensor("values_85_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor values_85_squeeze_mask_0 = const()[name = tensor("values_85_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor values_85_cast_fp16 = slice_by_index(begin = values_85_begin_0, end = values_85_end_0, end_mask = values_85_end_mask_0, squeeze_mask = values_85_squeeze_mask_0, x = reshape_149_cast_fp16)[name = tensor("values_85_cast_fp16")]; + tensor var_5861_cast_fp16 = not_equal(x = keys_85_cast_fp16, y = keys_85_cast_fp16)[name = tensor("op_5861_cast_fp16")]; + tensor keys_87_cast_fp16 = select(a = var_463_to_fp16, b = keys_85_cast_fp16, cond = var_5861_cast_fp16)[name = tensor("keys_87_cast_fp16")]; + tensor var_5869_cast_fp16 = not_equal(x = values_85_cast_fp16, y = values_85_cast_fp16)[name = tensor("op_5869_cast_fp16")]; + tensor values_87_cast_fp16 = select(a = var_463_to_fp16, b = values_85_cast_fp16, cond = var_5869_cast_fp16)[name = tensor("values_87_cast_fp16")]; + tensor var_5893 = const()[name = tensor("op_5893"), val = tensor([0, 2, 1, 3])]; + tensor var_5909 = const()[name = tensor("op_5909"), val = tensor([1, 1, 1])]; + tensor var_5910_cast_fp16 = reshape(shape = var_5909, x = position14_to_fp16)[name = tensor("op_5910_cast_fp16")]; + tensor var_5930_to_fp16 = const()[name = tensor("op_5930_to_fp16"), val = tensor(0x1p+0)]; + tensor valid_len_29_cast_fp16 = add(x = var_5910_cast_fp16, y = var_5930_to_fp16)[name = tensor("valid_len_29_cast_fp16")]; + tensor valid_mask_29_cast_fp16 = less(x = k_positions_1_promoted_to_fp16, y = valid_len_29_cast_fp16)[name = tensor("valid_mask_29_cast_fp16")]; + tensor causal_mask_29_cast_fp16 = less_equal(x = k_positions_1_promoted_to_fp16, y = var_5910_cast_fp16)[name = tensor("causal_mask_29_cast_fp16")]; + tensor attn_mask_57 = logical_and(x = valid_mask_29_cast_fp16, y = causal_mask_29_cast_fp16)[name = tensor("attn_mask_57")]; + tensor attn_mask_59_axes_0 = const()[name = tensor("attn_mask_59_axes_0"), val = tensor([1])]; + tensor attn_mask_59 = expand_dims(axes = attn_mask_59_axes_0, x = attn_mask_57)[name = tensor("attn_mask_59")]; + tensor var_5948_transpose_x_0 = const()[name = tensor("op_5948_transpose_x_0"), val = tensor(false)]; + tensor var_5948_transpose_y_0 = const()[name = tensor("op_5948_transpose_y_0"), val = tensor(false)]; + tensor transpose_100_perm_0 = const()[name = tensor("transpose_100_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_101_perm_0 = const()[name = tensor("transpose_101_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_101 = transpose(perm = transpose_101_perm_0, x = keys_87_cast_fp16)[name = tensor("transpose_157")]; + tensor transpose_100 = transpose(perm = transpose_100_perm_0, x = q_87_cast_fp16)[name = tensor("transpose_158")]; + tensor var_5948_cast_fp16 = matmul(transpose_x = var_5948_transpose_x_0, transpose_y = var_5948_transpose_y_0, x = transpose_100, y = transpose_101)[name = tensor("op_5948_cast_fp16")]; + tensor var_5942_to_fp16 = const()[name = tensor("op_5942_to_fp16"), val = tensor([0x1p-3])]; + tensor attn_weights_85_cast_fp16 = mul(x = var_5948_cast_fp16, y = var_5942_to_fp16)[name = tensor("attn_weights_85_cast_fp16")]; + tensor var_5950 = logical_not(x = attn_mask_59)[name = tensor("op_5950")]; + tensor var_5951_to_fp16 = const()[name = tensor("op_5951_to_fp16"), val = tensor(-inf)]; + tensor attn_weights_87_cast_fp16 = select(a = var_5951_to_fp16, b = attn_weights_85_cast_fp16, cond = var_5950)[name = tensor("attn_weights_87_cast_fp16")]; + tensor var_5953 = const()[name = tensor("op_5953"), val = tensor(-1)]; + tensor attn_weights_89_cast_fp16 = softmax(axis = var_5953, x = attn_weights_87_cast_fp16)[name = tensor("attn_weights_89_cast_fp16")]; + tensor attn_output_29_transpose_x_0 = const()[name = tensor("attn_output_29_transpose_x_0"), val = tensor(false)]; + tensor attn_output_29_transpose_y_0 = const()[name = tensor("attn_output_29_transpose_y_0"), val = tensor(false)]; + tensor values_89_cast_fp16 = transpose(perm = var_5893, x = values_87_cast_fp16)[name = tensor("transpose_159")]; + tensor attn_output_29_cast_fp16 = matmul(transpose_x = attn_output_29_transpose_x_0, transpose_y = attn_output_29_transpose_y_0, x = attn_weights_89_cast_fp16, y = values_89_cast_fp16)[name = tensor("attn_output_29_cast_fp16")]; + tensor var_5961 = const()[name = tensor("op_5961"), val = tensor([0, 2, 1, 3])]; + tensor var_5966 = const()[name = tensor("op_5966"), val = tensor([1, 1, 1024])]; + tensor var_5962_cast_fp16 = transpose(perm = var_5961, x = attn_output_29_cast_fp16)[name = tensor("transpose_156")]; + tensor input_175_cast_fp16 = reshape(shape = var_5966, x = var_5962_cast_fp16)[name = tensor("input_175_cast_fp16")]; + tensor attn14_out_proj_weight_to_fp16 = const()[name = tensor("attn14_out_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(359873024)))]; + tensor linear_58_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = attn14_out_proj_weight_to_fp16, x = input_175_cast_fp16)[name = tensor("linear_58_cast_fp16")]; + tensor var_5972_to_fp16 = const()[name = tensor("op_5972_to_fp16"), val = tensor(0x1p+0)]; + tensor var_5973_cast_fp16 = add(x = position14_to_fp16, y = var_5972_to_fp16)[name = tensor("op_5973_cast_fp16")]; + tensor var_5973_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_5973_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor input_177_cast_fp16 = add(x = input_171_cast_fp16, y = linear_58_cast_fp16)[name = tensor("input_177_cast_fp16")]; + tensor input_179_axes_0 = const()[name = tensor("input_179_axes_0"), val = tensor([-1])]; + tensor norm14_2_weight_to_fp16 = const()[name = tensor("norm14_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(361970240)))]; + tensor norm14_2_bias_to_fp16 = const()[name = tensor("norm14_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(361972352)))]; + tensor var_5977_to_fp16 = const()[name = tensor("op_5977_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_179_cast_fp16 = layer_norm(axes = input_179_axes_0, beta = norm14_2_bias_to_fp16, epsilon = var_5977_to_fp16, gamma = norm14_2_weight_to_fp16, x = input_177_cast_fp16)[name = tensor("input_179_cast_fp16")]; + tensor linear14_1_weight_to_fp16 = const()[name = tensor("linear14_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(361974464)))]; + tensor linear_59_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = linear14_1_weight_to_fp16, x = input_179_cast_fp16)[name = tensor("linear_59_cast_fp16")]; + tensor input_181_mode_0 = const()[name = tensor("input_181_mode_0"), val = tensor("EXACT")]; + tensor input_181_cast_fp16 = gelu(mode = input_181_mode_0, x = linear_59_cast_fp16)[name = tensor("input_181_cast_fp16")]; + tensor linear14_2_weight_to_fp16 = const()[name = tensor("linear14_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(370363136)))]; + tensor linear_60_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = linear14_2_weight_to_fp16, x = input_181_cast_fp16)[name = tensor("linear_60_cast_fp16")]; + tensor input_183_cast_fp16 = add(x = input_177_cast_fp16, y = linear_60_cast_fp16)[name = tensor("input_183_cast_fp16")]; + tensor input_185_axes_0 = const()[name = tensor("input_185_axes_0"), val = tensor([-1])]; + tensor norm15_1_weight_to_fp16 = const()[name = tensor("norm15_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(378751808)))]; + tensor norm15_1_bias_to_fp16 = const()[name = tensor("norm15_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(378753920)))]; + tensor var_5994_to_fp16 = const()[name = tensor("op_5994_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_185_cast_fp16 = layer_norm(axes = input_185_axes_0, beta = norm15_1_bias_to_fp16, epsilon = var_5994_to_fp16, gamma = norm15_1_weight_to_fp16, x = input_183_cast_fp16)[name = tensor("input_185_cast_fp16")]; + tensor attn15_in_proj_weight_to_fp16 = const()[name = tensor("attn15_in_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(378756032)))]; + tensor linear_61_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = attn15_in_proj_weight_to_fp16, x = input_185_cast_fp16)[name = tensor("linear_61_cast_fp16")]; + tensor var_6008 = const()[name = tensor("op_6008"), val = tensor([1, 1, 3, 16, 64])]; + tensor qkv_31_cast_fp16 = reshape(shape = var_6008, x = linear_61_cast_fp16)[name = tensor("qkv_31_cast_fp16")]; + tensor q_91_begin_0 = const()[name = tensor("q_91_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor q_91_end_0 = const()[name = tensor("q_91_end_0"), val = tensor([1, 1, 1, 16, 64])]; + tensor q_91_end_mask_0 = const()[name = tensor("q_91_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor q_91_squeeze_mask_0 = const()[name = tensor("q_91_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor q_91_cast_fp16 = slice_by_index(begin = q_91_begin_0, end = q_91_end_0, end_mask = q_91_end_mask_0, squeeze_mask = q_91_squeeze_mask_0, x = qkv_31_cast_fp16)[name = tensor("q_91_cast_fp16")]; + tensor k_61_begin_0 = const()[name = tensor("k_61_begin_0"), val = tensor([0, 0, 1, 0, 0])]; + tensor k_61_end_0 = const()[name = tensor("k_61_end_0"), val = tensor([1, 1, 2, 16, 64])]; + tensor k_61_end_mask_0 = const()[name = tensor("k_61_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor k_61_squeeze_mask_0 = const()[name = tensor("k_61_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor k_61_cast_fp16 = slice_by_index(begin = k_61_begin_0, end = k_61_end_0, end_mask = k_61_end_mask_0, squeeze_mask = k_61_squeeze_mask_0, x = qkv_31_cast_fp16)[name = tensor("k_61_cast_fp16")]; + tensor v_31_begin_0 = const()[name = tensor("v_31_begin_0"), val = tensor([0, 0, 2, 0, 0])]; + tensor v_31_end_0 = const()[name = tensor("v_31_end_0"), val = tensor([1, 1, 3, 16, 64])]; + tensor v_31_end_mask_0 = const()[name = tensor("v_31_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor v_31_squeeze_mask_0 = const()[name = tensor("v_31_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor v_31_cast_fp16 = slice_by_index(begin = v_31_begin_0, end = v_31_end_0, end_mask = v_31_end_mask_0, squeeze_mask = v_31_squeeze_mask_0, x = qkv_31_cast_fp16)[name = tensor("v_31_cast_fp16")]; + tensor var_6074 = const()[name = tensor("op_6074"), val = tensor([1, 1, 1, 1])]; + tensor position15_to_fp16_dtype_0 = const()[name = tensor("position15_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor position15_to_fp16 = cast(dtype = position15_to_fp16_dtype_0, x = position15)[name = tensor("cast_478")]; + tensor ts_95_cast_fp16 = reshape(shape = var_6074, x = position15_to_fp16)[name = tensor("ts_95_cast_fp16")]; + tensor var_6081 = const()[name = tensor("op_6081"), val = tensor([1, 1, 16, 32, 2])]; + tensor q_complex_31_cast_fp16 = reshape(shape = var_6081, x = q_91_cast_fp16)[name = tensor("q_complex_31_cast_fp16")]; + tensor var_6088 = const()[name = tensor("op_6088"), val = tensor([1, 1, 16, 32, 2])]; + tensor k_complex_31_cast_fp16 = reshape(shape = var_6088, x = k_61_cast_fp16)[name = tensor("k_complex_31_cast_fp16")]; + tensor var_6092_begin_0 = const()[name = tensor("op_6092_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_6092_end_0 = const()[name = tensor("op_6092_end_0"), val = tensor([1, 1, 16, 32, 1])]; + tensor var_6092_end_mask_0 = const()[name = tensor("op_6092_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_6092_squeeze_mask_0 = const()[name = tensor("op_6092_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_6092_cast_fp16 = slice_by_index(begin = var_6092_begin_0, end = var_6092_end_0, end_mask = var_6092_end_mask_0, squeeze_mask = var_6092_squeeze_mask_0, x = q_complex_31_cast_fp16)[name = tensor("op_6092_cast_fp16")]; + tensor var_6100_begin_0 = const()[name = tensor("op_6100_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_6100_end_0 = const()[name = tensor("op_6100_end_0"), val = tensor([1, 1, 16, 32, 2])]; + tensor var_6100_end_mask_0 = const()[name = tensor("op_6100_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_6100_squeeze_mask_0 = const()[name = tensor("op_6100_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_6100_cast_fp16 = slice_by_index(begin = var_6100_begin_0, end = var_6100_end_0, end_mask = var_6100_end_mask_0, squeeze_mask = var_6100_squeeze_mask_0, x = q_complex_31_cast_fp16)[name = tensor("op_6100_cast_fp16")]; + tensor var_6108_begin_0 = const()[name = tensor("op_6108_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_6108_end_0 = const()[name = tensor("op_6108_end_0"), val = tensor([1, 1, 16, 32, 1])]; + tensor var_6108_end_mask_0 = const()[name = tensor("op_6108_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_6108_squeeze_mask_0 = const()[name = tensor("op_6108_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_6108_cast_fp16 = slice_by_index(begin = var_6108_begin_0, end = var_6108_end_0, end_mask = var_6108_end_mask_0, squeeze_mask = var_6108_squeeze_mask_0, x = k_complex_31_cast_fp16)[name = tensor("op_6108_cast_fp16")]; + tensor var_6116_begin_0 = const()[name = tensor("op_6116_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_6116_end_0 = const()[name = tensor("op_6116_end_0"), val = tensor([1, 1, 16, 32, 2])]; + tensor var_6116_end_mask_0 = const()[name = tensor("op_6116_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_6116_squeeze_mask_0 = const()[name = tensor("op_6116_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_6116_cast_fp16 = slice_by_index(begin = var_6116_begin_0, end = var_6116_end_0, end_mask = var_6116_end_mask_0, squeeze_mask = var_6116_squeeze_mask_0, x = k_complex_31_cast_fp16)[name = tensor("op_6116_cast_fp16")]; + tensor freqs_31_to_fp16 = const()[name = tensor("freqs_31_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(385047552)))]; + tensor var_6122_cast_fp16 = mul(x = freqs_31_to_fp16, y = ts_95_cast_fp16)[name = tensor("op_6122_cast_fp16")]; + tensor rotr_31_cast_fp16 = cos(x = var_6122_cast_fp16)[name = tensor("rotr_31_cast_fp16")]; + tensor roti_31_cast_fp16 = sin(x = var_6122_cast_fp16)[name = tensor("roti_31_cast_fp16")]; + tensor var_6126_cast_fp16 = mul(x = var_6092_cast_fp16, y = rotr_31_cast_fp16)[name = tensor("op_6126_cast_fp16")]; + tensor var_6127_cast_fp16 = mul(x = var_6100_cast_fp16, y = roti_31_cast_fp16)[name = tensor("op_6127_cast_fp16")]; + tensor qor_61_cast_fp16 = sub(x = var_6126_cast_fp16, y = var_6127_cast_fp16)[name = tensor("qor_61_cast_fp16")]; + tensor var_6130_cast_fp16 = mul(x = var_6092_cast_fp16, y = roti_31_cast_fp16)[name = tensor("op_6130_cast_fp16")]; + tensor var_6131_cast_fp16 = mul(x = var_6100_cast_fp16, y = rotr_31_cast_fp16)[name = tensor("op_6131_cast_fp16")]; + tensor qoi_61_cast_fp16 = add(x = var_6130_cast_fp16, y = var_6131_cast_fp16)[name = tensor("qoi_61_cast_fp16")]; + tensor var_6134_cast_fp16 = mul(x = var_6108_cast_fp16, y = rotr_31_cast_fp16)[name = tensor("op_6134_cast_fp16")]; + tensor var_6135_cast_fp16 = mul(x = var_6116_cast_fp16, y = roti_31_cast_fp16)[name = tensor("op_6135_cast_fp16")]; + tensor kor_61_cast_fp16 = sub(x = var_6134_cast_fp16, y = var_6135_cast_fp16)[name = tensor("kor_61_cast_fp16")]; + tensor var_6138_cast_fp16 = mul(x = var_6108_cast_fp16, y = roti_31_cast_fp16)[name = tensor("op_6138_cast_fp16")]; + tensor var_6139_cast_fp16 = mul(x = var_6116_cast_fp16, y = rotr_31_cast_fp16)[name = tensor("op_6139_cast_fp16")]; + tensor koi_61_cast_fp16 = add(x = var_6138_cast_fp16, y = var_6139_cast_fp16)[name = tensor("koi_61_cast_fp16")]; + tensor qo_31_axis_0 = const()[name = tensor("qo_31_axis_0"), val = tensor(-1)]; + tensor qo_31_cast_fp16 = stack(axis = qo_31_axis_0, values = (qor_61_cast_fp16, qoi_61_cast_fp16))[name = tensor("qo_31_cast_fp16")]; + tensor ko_31_axis_0 = const()[name = tensor("ko_31_axis_0"), val = tensor(-1)]; + tensor ko_31_cast_fp16 = stack(axis = ko_31_axis_0, values = (kor_61_cast_fp16, koi_61_cast_fp16))[name = tensor("ko_31_cast_fp16")]; + tensor var_6172 = const()[name = tensor("op_6172"), val = tensor([1, 1, 16, 64])]; + tensor q_93_cast_fp16 = reshape(shape = var_6172, x = qo_31_cast_fp16)[name = tensor("q_93_cast_fp16")]; + tensor var_6178 = const()[name = tensor("op_6178"), val = tensor([1, 1, 16, 64])]; + tensor k_63_cast_fp16 = reshape(shape = var_6178, x = ko_31_cast_fp16)[name = tensor("k_63_cast_fp16")]; + tensor _inversed_6203_y_0_to_fp16 = const()[name = tensor("_inversed_6203_y_0_to_fp16"), val = tensor(0x1p-9)]; + tensor _inversed_6203_cast_fp16 = mul(x = ts_95_cast_fp16, y = _inversed_6203_y_0_to_fp16)[name = tensor("_inversed_6203_cast_fp16")]; + tensor var_6204_cast_fp16 = floor(x = _inversed_6203_cast_fp16)[name = tensor("op_6204_cast_fp16")]; + tensor var_6205_to_fp16 = const()[name = tensor("op_6205_to_fp16"), val = tensor(0x1p+9)]; + tensor var_6206_cast_fp16 = mul(x = var_6204_cast_fp16, y = var_6205_to_fp16)[name = tensor("op_6206_cast_fp16")]; + tensor write_indices_float_63_cast_fp16 = sub(x = ts_95_cast_fp16, y = var_6206_cast_fp16)[name = tensor("write_indices_float_63_cast_fp16")]; + tensor var_6213_dtype_0 = const()[name = tensor("op_6213_dtype_0"), val = tensor("int32")]; + tensor write_indices_31_reps_0 = const()[name = tensor("write_indices_31_reps_0"), val = tensor([1, 1, 16, 64])]; + tensor write_indices_float_63_cast_fp16_to_int32 = cast(dtype = var_6213_dtype_0, x = write_indices_float_63_cast_fp16)[name = tensor("cast_477")]; + tensor write_indices_31 = tile(reps = write_indices_31_reps_0, x = write_indices_float_63_cast_fp16_to_int32)[name = tensor("write_indices_31")]; + tensor var_6223_begin_0 = const()[name = tensor("op_6223_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_6223_end_0 = const()[name = tensor("op_6223_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor var_6223_end_mask_0 = const()[name = tensor("op_6223_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_6223_squeeze_mask_0 = const()[name = tensor("op_6223_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor cache15_to_fp16_dtype_0 = const()[name = tensor("cache15_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor cache15_to_fp16 = cast(dtype = cache15_to_fp16_dtype_0, x = cache15)[name = tensor("cast_476")]; + tensor var_6223_cast_fp16 = slice_by_index(begin = var_6223_begin_0, end = var_6223_end_0, end_mask = var_6223_end_mask_0, squeeze_mask = var_6223_squeeze_mask_0, x = cache15_to_fp16)[name = tensor("op_6223_cast_fp16")]; + tensor var_6225_axis_0 = const()[name = tensor("op_6225_axis_0"), val = tensor(1)]; + tensor var_6225_mode_0 = const()[name = tensor("op_6225_mode_0"), val = tensor("update")]; + tensor var_6225_validate_indices_0 = const()[name = tensor("op_6225_validate_indices_0"), val = tensor(false)]; + tensor var_6225_cast_fp16 = scatter_along_axis(axis = var_6225_axis_0, data = var_6223_cast_fp16, indices = write_indices_31, mode = var_6225_mode_0, updates = k_63_cast_fp16, validate_indices = var_6225_validate_indices_0)[name = tensor("op_6225_cast_fp16")]; + tensor concat_107 = const()[name = tensor("concat_107"), val = tensor([0, 0, 0, 0, 0])]; + tensor concat_108 = const()[name = tensor("concat_108"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_31_internal_tensor_assign_1_stride_0 = const()[name = tensor("new_cache_31_internal_tensor_assign_1_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_31_internal_tensor_assign_1_begin_mask_0 = const()[name = tensor("new_cache_31_internal_tensor_assign_1_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_31_internal_tensor_assign_1_end_mask_0 = const()[name = tensor("new_cache_31_internal_tensor_assign_1_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_31_internal_tensor_assign_1_squeeze_mask_0 = const()[name = tensor("new_cache_31_internal_tensor_assign_1_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_78 = const()[name = tensor("shape_78"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_30 = const()[name = tensor("reduce_prod_30"), val = tensor(1048576)]; + tensor range_1d_30_start_0 = const()[name = tensor("range_1d_30_start_0"), val = tensor(0)]; + tensor range_1d_30_step_0 = const()[name = tensor("range_1d_30_step_0"), val = tensor(1)]; + tensor range_1d_30 = range_1d(end = reduce_prod_30, start = range_1d_30_start_0, step = range_1d_30_step_0)[name = tensor("range_1d_30")]; + tensor reshape_150 = reshape(shape = shape_78, x = range_1d_30)[name = tensor("reshape_150")]; + tensor slice_by_index_30 = slice_by_index(begin = concat_107, begin_mask = new_cache_31_internal_tensor_assign_1_begin_mask_0, end = concat_108, end_mask = new_cache_31_internal_tensor_assign_1_end_mask_0, squeeze_mask = new_cache_31_internal_tensor_assign_1_squeeze_mask_0, stride = new_cache_31_internal_tensor_assign_1_stride_0, x = reshape_150)[name = tensor("slice_by_index_30")]; + tensor reshape_151_shape_0 = const()[name = tensor("reshape_151_shape_0"), val = tensor([-1])]; + tensor reshape_151 = reshape(shape = reshape_151_shape_0, x = slice_by_index_30)[name = tensor("reshape_151")]; + tensor reshape_152_shape_0 = const()[name = tensor("reshape_152_shape_0"), val = tensor([-1])]; + tensor reshape_152_cast_fp16 = reshape(shape = reshape_152_shape_0, x = var_6225_cast_fp16)[name = tensor("reshape_152_cast_fp16")]; + tensor reshape_153_shape_0 = const()[name = tensor("reshape_153_shape_0"), val = tensor([-1])]; + tensor reshape_153_cast_fp16 = reshape(shape = reshape_153_shape_0, x = cache15_to_fp16)[name = tensor("reshape_153_cast_fp16")]; + tensor scatter_30_mode_0 = const()[name = tensor("scatter_30_mode_0"), val = tensor("update")]; + tensor scatter_30_axis_0 = const()[name = tensor("scatter_30_axis_0"), val = tensor(0)]; + tensor scatter_30_validate_indices_0 = const()[name = tensor("scatter_30_validate_indices_0"), val = tensor(false)]; + tensor scatter_30_cast_fp16 = scatter(axis = scatter_30_axis_0, data = reshape_153_cast_fp16, indices = reshape_151, mode = scatter_30_mode_0, updates = reshape_152_cast_fp16, validate_indices = scatter_30_validate_indices_0)[name = tensor("scatter_30_cast_fp16")]; + tensor reshape_154_cast_fp16 = reshape(shape = shape_78, x = scatter_30_cast_fp16)[name = tensor("reshape_154_cast_fp16")]; + tensor var_6233_begin_0 = const()[name = tensor("op_6233_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor var_6233_end_0 = const()[name = tensor("op_6233_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor var_6233_end_mask_0 = const()[name = tensor("op_6233_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_6233_squeeze_mask_0 = const()[name = tensor("op_6233_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor var_6233_cast_fp16 = slice_by_index(begin = var_6233_begin_0, end = var_6233_end_0, end_mask = var_6233_end_mask_0, squeeze_mask = var_6233_squeeze_mask_0, x = reshape_154_cast_fp16)[name = tensor("op_6233_cast_fp16")]; + tensor var_6235_axis_0 = const()[name = tensor("op_6235_axis_0"), val = tensor(1)]; + tensor var_6235_mode_0 = const()[name = tensor("op_6235_mode_0"), val = tensor("update")]; + tensor var_6235_validate_indices_0 = const()[name = tensor("op_6235_validate_indices_0"), val = tensor(false)]; + tensor var_6235_cast_fp16 = scatter_along_axis(axis = var_6235_axis_0, data = var_6233_cast_fp16, indices = write_indices_31, mode = var_6235_mode_0, updates = v_31_cast_fp16, validate_indices = var_6235_validate_indices_0)[name = tensor("op_6235_cast_fp16")]; + tensor concat_109 = const()[name = tensor("concat_109"), val = tensor([1, 0, 0, 0, 0])]; + tensor concat_110 = const()[name = tensor("concat_110"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_31_internal_tensor_assign_2_stride_0 = const()[name = tensor("new_cache_31_internal_tensor_assign_2_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_31_internal_tensor_assign_2_begin_mask_0 = const()[name = tensor("new_cache_31_internal_tensor_assign_2_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_31_internal_tensor_assign_2_end_mask_0 = const()[name = tensor("new_cache_31_internal_tensor_assign_2_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_31_internal_tensor_assign_2_squeeze_mask_0 = const()[name = tensor("new_cache_31_internal_tensor_assign_2_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_79 = const()[name = tensor("shape_79"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_31 = const()[name = tensor("reduce_prod_31"), val = tensor(1048576)]; + tensor range_1d_31_start_0 = const()[name = tensor("range_1d_31_start_0"), val = tensor(0)]; + tensor range_1d_31_step_0 = const()[name = tensor("range_1d_31_step_0"), val = tensor(1)]; + tensor range_1d_31 = range_1d(end = reduce_prod_31, start = range_1d_31_start_0, step = range_1d_31_step_0)[name = tensor("range_1d_31")]; + tensor reshape_155 = reshape(shape = shape_79, x = range_1d_31)[name = tensor("reshape_155")]; + tensor slice_by_index_31 = slice_by_index(begin = concat_109, begin_mask = new_cache_31_internal_tensor_assign_2_begin_mask_0, end = concat_110, end_mask = new_cache_31_internal_tensor_assign_2_end_mask_0, squeeze_mask = new_cache_31_internal_tensor_assign_2_squeeze_mask_0, stride = new_cache_31_internal_tensor_assign_2_stride_0, x = reshape_155)[name = tensor("slice_by_index_31")]; + tensor reshape_156_shape_0 = const()[name = tensor("reshape_156_shape_0"), val = tensor([-1])]; + tensor reshape_156 = reshape(shape = reshape_156_shape_0, x = slice_by_index_31)[name = tensor("reshape_156")]; + tensor reshape_157_shape_0 = const()[name = tensor("reshape_157_shape_0"), val = tensor([-1])]; + tensor reshape_157_cast_fp16 = reshape(shape = reshape_157_shape_0, x = var_6235_cast_fp16)[name = tensor("reshape_157_cast_fp16")]; + tensor reshape_158_shape_0 = const()[name = tensor("reshape_158_shape_0"), val = tensor([-1])]; + tensor reshape_158_cast_fp16 = reshape(shape = reshape_158_shape_0, x = reshape_154_cast_fp16)[name = tensor("reshape_158_cast_fp16")]; + tensor scatter_31_mode_0 = const()[name = tensor("scatter_31_mode_0"), val = tensor("update")]; + tensor scatter_31_axis_0 = const()[name = tensor("scatter_31_axis_0"), val = tensor(0)]; + tensor scatter_31_validate_indices_0 = const()[name = tensor("scatter_31_validate_indices_0"), val = tensor(false)]; + tensor scatter_31_cast_fp16 = scatter(axis = scatter_31_axis_0, data = reshape_158_cast_fp16, indices = reshape_156, mode = scatter_31_mode_0, updates = reshape_157_cast_fp16, validate_indices = scatter_31_validate_indices_0)[name = tensor("scatter_31_cast_fp16")]; + tensor reshape_159_cast_fp16 = reshape(shape = shape_79, x = scatter_31_cast_fp16)[name = tensor("reshape_159_cast_fp16")]; + tensor reshape_159_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("reshape_159_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor keys_91_begin_0 = const()[name = tensor("keys_91_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor keys_91_end_0 = const()[name = tensor("keys_91_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor keys_91_end_mask_0 = const()[name = tensor("keys_91_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor keys_91_squeeze_mask_0 = const()[name = tensor("keys_91_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor keys_91_cast_fp16 = slice_by_index(begin = keys_91_begin_0, end = keys_91_end_0, end_mask = keys_91_end_mask_0, squeeze_mask = keys_91_squeeze_mask_0, x = reshape_159_cast_fp16)[name = tensor("keys_91_cast_fp16")]; + tensor values_91_begin_0 = const()[name = tensor("values_91_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor values_91_end_0 = const()[name = tensor("values_91_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor values_91_end_mask_0 = const()[name = tensor("values_91_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor values_91_squeeze_mask_0 = const()[name = tensor("values_91_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor values_91_cast_fp16 = slice_by_index(begin = values_91_begin_0, end = values_91_end_0, end_mask = values_91_end_mask_0, squeeze_mask = values_91_squeeze_mask_0, x = reshape_159_cast_fp16)[name = tensor("values_91_cast_fp16")]; + tensor var_6247_cast_fp16 = not_equal(x = keys_91_cast_fp16, y = keys_91_cast_fp16)[name = tensor("op_6247_cast_fp16")]; + tensor keys_93_cast_fp16 = select(a = var_463_to_fp16, b = keys_91_cast_fp16, cond = var_6247_cast_fp16)[name = tensor("keys_93_cast_fp16")]; + tensor var_6255_cast_fp16 = not_equal(x = values_91_cast_fp16, y = values_91_cast_fp16)[name = tensor("op_6255_cast_fp16")]; + tensor values_93_cast_fp16 = select(a = var_463_to_fp16, b = values_91_cast_fp16, cond = var_6255_cast_fp16)[name = tensor("values_93_cast_fp16")]; + tensor var_6279 = const()[name = tensor("op_6279"), val = tensor([0, 2, 1, 3])]; + tensor var_6295 = const()[name = tensor("op_6295"), val = tensor([1, 1, 1])]; + tensor var_6296_cast_fp16 = reshape(shape = var_6295, x = position15_to_fp16)[name = tensor("op_6296_cast_fp16")]; + tensor var_6316_to_fp16 = const()[name = tensor("op_6316_to_fp16"), val = tensor(0x1p+0)]; + tensor valid_len_31_cast_fp16 = add(x = var_6296_cast_fp16, y = var_6316_to_fp16)[name = tensor("valid_len_31_cast_fp16")]; + tensor valid_mask_31_cast_fp16 = less(x = k_positions_1_promoted_to_fp16, y = valid_len_31_cast_fp16)[name = tensor("valid_mask_31_cast_fp16")]; + tensor causal_mask_31_cast_fp16 = less_equal(x = k_positions_1_promoted_to_fp16, y = var_6296_cast_fp16)[name = tensor("causal_mask_31_cast_fp16")]; + tensor attn_mask_61 = logical_and(x = valid_mask_31_cast_fp16, y = causal_mask_31_cast_fp16)[name = tensor("attn_mask_61")]; + tensor attn_mask_63_axes_0 = const()[name = tensor("attn_mask_63_axes_0"), val = tensor([1])]; + tensor attn_mask_63 = expand_dims(axes = attn_mask_63_axes_0, x = attn_mask_61)[name = tensor("attn_mask_63")]; + tensor var_6334_transpose_x_0 = const()[name = tensor("op_6334_transpose_x_0"), val = tensor(false)]; + tensor var_6334_transpose_y_0 = const()[name = tensor("op_6334_transpose_y_0"), val = tensor(false)]; + tensor transpose_102_perm_0 = const()[name = tensor("transpose_102_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_103_perm_0 = const()[name = tensor("transpose_103_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_103 = transpose(perm = transpose_103_perm_0, x = keys_93_cast_fp16)[name = tensor("transpose_153")]; + tensor transpose_102 = transpose(perm = transpose_102_perm_0, x = q_93_cast_fp16)[name = tensor("transpose_154")]; + tensor var_6334_cast_fp16 = matmul(transpose_x = var_6334_transpose_x_0, transpose_y = var_6334_transpose_y_0, x = transpose_102, y = transpose_103)[name = tensor("op_6334_cast_fp16")]; + tensor var_6328_to_fp16 = const()[name = tensor("op_6328_to_fp16"), val = tensor([0x1p-3])]; + tensor attn_weights_91_cast_fp16 = mul(x = var_6334_cast_fp16, y = var_6328_to_fp16)[name = tensor("attn_weights_91_cast_fp16")]; + tensor var_6336 = logical_not(x = attn_mask_63)[name = tensor("op_6336")]; + tensor var_6337_to_fp16 = const()[name = tensor("op_6337_to_fp16"), val = tensor(-inf)]; + tensor attn_weights_93_cast_fp16 = select(a = var_6337_to_fp16, b = attn_weights_91_cast_fp16, cond = var_6336)[name = tensor("attn_weights_93_cast_fp16")]; + tensor var_6339 = const()[name = tensor("op_6339"), val = tensor(-1)]; + tensor attn_weights_95_cast_fp16 = softmax(axis = var_6339, x = attn_weights_93_cast_fp16)[name = tensor("attn_weights_95_cast_fp16")]; + tensor attn_output_31_transpose_x_0 = const()[name = tensor("attn_output_31_transpose_x_0"), val = tensor(false)]; + tensor attn_output_31_transpose_y_0 = const()[name = tensor("attn_output_31_transpose_y_0"), val = tensor(false)]; + tensor values_95_cast_fp16 = transpose(perm = var_6279, x = values_93_cast_fp16)[name = tensor("transpose_155")]; + tensor attn_output_31_cast_fp16 = matmul(transpose_x = attn_output_31_transpose_x_0, transpose_y = attn_output_31_transpose_y_0, x = attn_weights_95_cast_fp16, y = values_95_cast_fp16)[name = tensor("attn_output_31_cast_fp16")]; + tensor var_6347 = const()[name = tensor("op_6347"), val = tensor([0, 2, 1, 3])]; + tensor var_6352 = const()[name = tensor("op_6352"), val = tensor([1, 1, 1024])]; + tensor var_6348_cast_fp16 = transpose(perm = var_6347, x = attn_output_31_cast_fp16)[name = tensor("transpose_152")]; + tensor input_187_cast_fp16 = reshape(shape = var_6352, x = var_6348_cast_fp16)[name = tensor("input_187_cast_fp16")]; + tensor attn15_out_proj_weight_to_fp16 = const()[name = tensor("attn15_out_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(385047680)))]; + tensor linear_62_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = attn15_out_proj_weight_to_fp16, x = input_187_cast_fp16)[name = tensor("linear_62_cast_fp16")]; + tensor var_6358_to_fp16 = const()[name = tensor("op_6358_to_fp16"), val = tensor(0x1p+0)]; + tensor var_6359_cast_fp16 = add(x = position15_to_fp16, y = var_6358_to_fp16)[name = tensor("op_6359_cast_fp16")]; + tensor var_6359_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_6359_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor input_189_cast_fp16 = add(x = input_183_cast_fp16, y = linear_62_cast_fp16)[name = tensor("input_189_cast_fp16")]; + tensor input_191_axes_0 = const()[name = tensor("input_191_axes_0"), val = tensor([-1])]; + tensor norm15_2_weight_to_fp16 = const()[name = tensor("norm15_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(387144896)))]; + tensor norm15_2_bias_to_fp16 = const()[name = tensor("norm15_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(387147008)))]; + tensor var_6363_to_fp16 = const()[name = tensor("op_6363_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_191_cast_fp16 = layer_norm(axes = input_191_axes_0, beta = norm15_2_bias_to_fp16, epsilon = var_6363_to_fp16, gamma = norm15_2_weight_to_fp16, x = input_189_cast_fp16)[name = tensor("input_191_cast_fp16")]; + tensor linear15_1_weight_to_fp16 = const()[name = tensor("linear15_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(387149120)))]; + tensor linear_63_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = linear15_1_weight_to_fp16, x = input_191_cast_fp16)[name = tensor("linear_63_cast_fp16")]; + tensor input_193_mode_0 = const()[name = tensor("input_193_mode_0"), val = tensor("EXACT")]; + tensor input_193_cast_fp16 = gelu(mode = input_193_mode_0, x = linear_63_cast_fp16)[name = tensor("input_193_cast_fp16")]; + tensor linear15_2_weight_to_fp16 = const()[name = tensor("linear15_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(395537792)))]; + tensor linear_64_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = linear15_2_weight_to_fp16, x = input_193_cast_fp16)[name = tensor("linear_64_cast_fp16")]; + tensor input_195_cast_fp16 = add(x = input_189_cast_fp16, y = linear_64_cast_fp16)[name = tensor("input_195_cast_fp16")]; + tensor input_197_axes_0 = const()[name = tensor("input_197_axes_0"), val = tensor([-1])]; + tensor norm16_1_weight_to_fp16 = const()[name = tensor("norm16_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(403926464)))]; + tensor norm16_1_bias_to_fp16 = const()[name = tensor("norm16_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(403928576)))]; + tensor var_6380_to_fp16 = const()[name = tensor("op_6380_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_197_cast_fp16 = layer_norm(axes = input_197_axes_0, beta = norm16_1_bias_to_fp16, epsilon = var_6380_to_fp16, gamma = norm16_1_weight_to_fp16, x = input_195_cast_fp16)[name = tensor("input_197_cast_fp16")]; + tensor attn16_in_proj_weight_to_fp16 = const()[name = tensor("attn16_in_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(403930688)))]; + tensor linear_65_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = attn16_in_proj_weight_to_fp16, x = input_197_cast_fp16)[name = tensor("linear_65_cast_fp16")]; + tensor var_6394 = const()[name = tensor("op_6394"), val = tensor([1, 1, 3, 16, 64])]; + tensor qkv_33_cast_fp16 = reshape(shape = var_6394, x = linear_65_cast_fp16)[name = tensor("qkv_33_cast_fp16")]; + tensor q_97_begin_0 = const()[name = tensor("q_97_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor q_97_end_0 = const()[name = tensor("q_97_end_0"), val = tensor([1, 1, 1, 16, 64])]; + tensor q_97_end_mask_0 = const()[name = tensor("q_97_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor q_97_squeeze_mask_0 = const()[name = tensor("q_97_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor q_97_cast_fp16 = slice_by_index(begin = q_97_begin_0, end = q_97_end_0, end_mask = q_97_end_mask_0, squeeze_mask = q_97_squeeze_mask_0, x = qkv_33_cast_fp16)[name = tensor("q_97_cast_fp16")]; + tensor k_65_begin_0 = const()[name = tensor("k_65_begin_0"), val = tensor([0, 0, 1, 0, 0])]; + tensor k_65_end_0 = const()[name = tensor("k_65_end_0"), val = tensor([1, 1, 2, 16, 64])]; + tensor k_65_end_mask_0 = const()[name = tensor("k_65_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor k_65_squeeze_mask_0 = const()[name = tensor("k_65_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor k_65_cast_fp16 = slice_by_index(begin = k_65_begin_0, end = k_65_end_0, end_mask = k_65_end_mask_0, squeeze_mask = k_65_squeeze_mask_0, x = qkv_33_cast_fp16)[name = tensor("k_65_cast_fp16")]; + tensor v_33_begin_0 = const()[name = tensor("v_33_begin_0"), val = tensor([0, 0, 2, 0, 0])]; + tensor v_33_end_0 = const()[name = tensor("v_33_end_0"), val = tensor([1, 1, 3, 16, 64])]; + tensor v_33_end_mask_0 = const()[name = tensor("v_33_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor v_33_squeeze_mask_0 = const()[name = tensor("v_33_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor v_33_cast_fp16 = slice_by_index(begin = v_33_begin_0, end = v_33_end_0, end_mask = v_33_end_mask_0, squeeze_mask = v_33_squeeze_mask_0, x = qkv_33_cast_fp16)[name = tensor("v_33_cast_fp16")]; + tensor var_6460 = const()[name = tensor("op_6460"), val = tensor([1, 1, 1, 1])]; + tensor position16_to_fp16_dtype_0 = const()[name = tensor("position16_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor position16_to_fp16 = cast(dtype = position16_to_fp16_dtype_0, x = position16)[name = tensor("cast_473")]; + tensor ts_101_cast_fp16 = reshape(shape = var_6460, x = position16_to_fp16)[name = tensor("ts_101_cast_fp16")]; + tensor var_6467 = const()[name = tensor("op_6467"), val = tensor([1, 1, 16, 32, 2])]; + tensor q_complex_33_cast_fp16 = reshape(shape = var_6467, x = q_97_cast_fp16)[name = tensor("q_complex_33_cast_fp16")]; + tensor var_6474 = const()[name = tensor("op_6474"), val = tensor([1, 1, 16, 32, 2])]; + tensor k_complex_33_cast_fp16 = reshape(shape = var_6474, x = k_65_cast_fp16)[name = tensor("k_complex_33_cast_fp16")]; + tensor var_6478_begin_0 = const()[name = tensor("op_6478_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_6478_end_0 = const()[name = tensor("op_6478_end_0"), val = tensor([1, 1, 16, 32, 1])]; + tensor var_6478_end_mask_0 = const()[name = tensor("op_6478_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_6478_squeeze_mask_0 = const()[name = tensor("op_6478_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_6478_cast_fp16 = slice_by_index(begin = var_6478_begin_0, end = var_6478_end_0, end_mask = var_6478_end_mask_0, squeeze_mask = var_6478_squeeze_mask_0, x = q_complex_33_cast_fp16)[name = tensor("op_6478_cast_fp16")]; + tensor var_6486_begin_0 = const()[name = tensor("op_6486_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_6486_end_0 = const()[name = tensor("op_6486_end_0"), val = tensor([1, 1, 16, 32, 2])]; + tensor var_6486_end_mask_0 = const()[name = tensor("op_6486_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_6486_squeeze_mask_0 = const()[name = tensor("op_6486_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_6486_cast_fp16 = slice_by_index(begin = var_6486_begin_0, end = var_6486_end_0, end_mask = var_6486_end_mask_0, squeeze_mask = var_6486_squeeze_mask_0, x = q_complex_33_cast_fp16)[name = tensor("op_6486_cast_fp16")]; + tensor var_6494_begin_0 = const()[name = tensor("op_6494_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_6494_end_0 = const()[name = tensor("op_6494_end_0"), val = tensor([1, 1, 16, 32, 1])]; + tensor var_6494_end_mask_0 = const()[name = tensor("op_6494_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_6494_squeeze_mask_0 = const()[name = tensor("op_6494_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_6494_cast_fp16 = slice_by_index(begin = var_6494_begin_0, end = var_6494_end_0, end_mask = var_6494_end_mask_0, squeeze_mask = var_6494_squeeze_mask_0, x = k_complex_33_cast_fp16)[name = tensor("op_6494_cast_fp16")]; + tensor var_6502_begin_0 = const()[name = tensor("op_6502_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_6502_end_0 = const()[name = tensor("op_6502_end_0"), val = tensor([1, 1, 16, 32, 2])]; + tensor var_6502_end_mask_0 = const()[name = tensor("op_6502_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_6502_squeeze_mask_0 = const()[name = tensor("op_6502_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_6502_cast_fp16 = slice_by_index(begin = var_6502_begin_0, end = var_6502_end_0, end_mask = var_6502_end_mask_0, squeeze_mask = var_6502_squeeze_mask_0, x = k_complex_33_cast_fp16)[name = tensor("op_6502_cast_fp16")]; + tensor freqs_33_to_fp16 = const()[name = tensor("freqs_33_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(410222208)))]; + tensor var_6508_cast_fp16 = mul(x = freqs_33_to_fp16, y = ts_101_cast_fp16)[name = tensor("op_6508_cast_fp16")]; + tensor rotr_33_cast_fp16 = cos(x = var_6508_cast_fp16)[name = tensor("rotr_33_cast_fp16")]; + tensor roti_33_cast_fp16 = sin(x = var_6508_cast_fp16)[name = tensor("roti_33_cast_fp16")]; + tensor var_6512_cast_fp16 = mul(x = var_6478_cast_fp16, y = rotr_33_cast_fp16)[name = tensor("op_6512_cast_fp16")]; + tensor var_6513_cast_fp16 = mul(x = var_6486_cast_fp16, y = roti_33_cast_fp16)[name = tensor("op_6513_cast_fp16")]; + tensor qor_65_cast_fp16 = sub(x = var_6512_cast_fp16, y = var_6513_cast_fp16)[name = tensor("qor_65_cast_fp16")]; + tensor var_6516_cast_fp16 = mul(x = var_6478_cast_fp16, y = roti_33_cast_fp16)[name = tensor("op_6516_cast_fp16")]; + tensor var_6517_cast_fp16 = mul(x = var_6486_cast_fp16, y = rotr_33_cast_fp16)[name = tensor("op_6517_cast_fp16")]; + tensor qoi_65_cast_fp16 = add(x = var_6516_cast_fp16, y = var_6517_cast_fp16)[name = tensor("qoi_65_cast_fp16")]; + tensor var_6520_cast_fp16 = mul(x = var_6494_cast_fp16, y = rotr_33_cast_fp16)[name = tensor("op_6520_cast_fp16")]; + tensor var_6521_cast_fp16 = mul(x = var_6502_cast_fp16, y = roti_33_cast_fp16)[name = tensor("op_6521_cast_fp16")]; + tensor kor_65_cast_fp16 = sub(x = var_6520_cast_fp16, y = var_6521_cast_fp16)[name = tensor("kor_65_cast_fp16")]; + tensor var_6524_cast_fp16 = mul(x = var_6494_cast_fp16, y = roti_33_cast_fp16)[name = tensor("op_6524_cast_fp16")]; + tensor var_6525_cast_fp16 = mul(x = var_6502_cast_fp16, y = rotr_33_cast_fp16)[name = tensor("op_6525_cast_fp16")]; + tensor koi_65_cast_fp16 = add(x = var_6524_cast_fp16, y = var_6525_cast_fp16)[name = tensor("koi_65_cast_fp16")]; + tensor qo_33_axis_0 = const()[name = tensor("qo_33_axis_0"), val = tensor(-1)]; + tensor qo_33_cast_fp16 = stack(axis = qo_33_axis_0, values = (qor_65_cast_fp16, qoi_65_cast_fp16))[name = tensor("qo_33_cast_fp16")]; + tensor ko_33_axis_0 = const()[name = tensor("ko_33_axis_0"), val = tensor(-1)]; + tensor ko_33_cast_fp16 = stack(axis = ko_33_axis_0, values = (kor_65_cast_fp16, koi_65_cast_fp16))[name = tensor("ko_33_cast_fp16")]; + tensor var_6558 = const()[name = tensor("op_6558"), val = tensor([1, 1, 16, 64])]; + tensor q_99_cast_fp16 = reshape(shape = var_6558, x = qo_33_cast_fp16)[name = tensor("q_99_cast_fp16")]; + tensor var_6564 = const()[name = tensor("op_6564"), val = tensor([1, 1, 16, 64])]; + tensor k_67_cast_fp16 = reshape(shape = var_6564, x = ko_33_cast_fp16)[name = tensor("k_67_cast_fp16")]; + tensor _inversed_6589_y_0_to_fp16 = const()[name = tensor("_inversed_6589_y_0_to_fp16"), val = tensor(0x1p-9)]; + tensor _inversed_6589_cast_fp16 = mul(x = ts_101_cast_fp16, y = _inversed_6589_y_0_to_fp16)[name = tensor("_inversed_6589_cast_fp16")]; + tensor var_6590_cast_fp16 = floor(x = _inversed_6589_cast_fp16)[name = tensor("op_6590_cast_fp16")]; + tensor var_6591_to_fp16 = const()[name = tensor("op_6591_to_fp16"), val = tensor(0x1p+9)]; + tensor var_6592_cast_fp16 = mul(x = var_6590_cast_fp16, y = var_6591_to_fp16)[name = tensor("op_6592_cast_fp16")]; + tensor write_indices_float_67_cast_fp16 = sub(x = ts_101_cast_fp16, y = var_6592_cast_fp16)[name = tensor("write_indices_float_67_cast_fp16")]; + tensor var_6599_dtype_0 = const()[name = tensor("op_6599_dtype_0"), val = tensor("int32")]; + tensor write_indices_33_reps_0 = const()[name = tensor("write_indices_33_reps_0"), val = tensor([1, 1, 16, 64])]; + tensor write_indices_float_67_cast_fp16_to_int32 = cast(dtype = var_6599_dtype_0, x = write_indices_float_67_cast_fp16)[name = tensor("cast_472")]; + tensor write_indices_33 = tile(reps = write_indices_33_reps_0, x = write_indices_float_67_cast_fp16_to_int32)[name = tensor("write_indices_33")]; + tensor var_6609_begin_0 = const()[name = tensor("op_6609_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_6609_end_0 = const()[name = tensor("op_6609_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor var_6609_end_mask_0 = const()[name = tensor("op_6609_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_6609_squeeze_mask_0 = const()[name = tensor("op_6609_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor cache16_to_fp16_dtype_0 = const()[name = tensor("cache16_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor cache16_to_fp16 = cast(dtype = cache16_to_fp16_dtype_0, x = cache16)[name = tensor("cast_471")]; + tensor var_6609_cast_fp16 = slice_by_index(begin = var_6609_begin_0, end = var_6609_end_0, end_mask = var_6609_end_mask_0, squeeze_mask = var_6609_squeeze_mask_0, x = cache16_to_fp16)[name = tensor("op_6609_cast_fp16")]; + tensor var_6611_axis_0 = const()[name = tensor("op_6611_axis_0"), val = tensor(1)]; + tensor var_6611_mode_0 = const()[name = tensor("op_6611_mode_0"), val = tensor("update")]; + tensor var_6611_validate_indices_0 = const()[name = tensor("op_6611_validate_indices_0"), val = tensor(false)]; + tensor var_6611_cast_fp16 = scatter_along_axis(axis = var_6611_axis_0, data = var_6609_cast_fp16, indices = write_indices_33, mode = var_6611_mode_0, updates = k_67_cast_fp16, validate_indices = var_6611_validate_indices_0)[name = tensor("op_6611_cast_fp16")]; + tensor concat_114 = const()[name = tensor("concat_114"), val = tensor([0, 0, 0, 0, 0])]; + tensor concat_115 = const()[name = tensor("concat_115"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_33_internal_tensor_assign_1_stride_0 = const()[name = tensor("new_cache_33_internal_tensor_assign_1_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_33_internal_tensor_assign_1_begin_mask_0 = const()[name = tensor("new_cache_33_internal_tensor_assign_1_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_33_internal_tensor_assign_1_end_mask_0 = const()[name = tensor("new_cache_33_internal_tensor_assign_1_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_33_internal_tensor_assign_1_squeeze_mask_0 = const()[name = tensor("new_cache_33_internal_tensor_assign_1_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_80 = const()[name = tensor("shape_80"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_32 = const()[name = tensor("reduce_prod_32"), val = tensor(1048576)]; + tensor range_1d_32_start_0 = const()[name = tensor("range_1d_32_start_0"), val = tensor(0)]; + tensor range_1d_32_step_0 = const()[name = tensor("range_1d_32_step_0"), val = tensor(1)]; + tensor range_1d_32 = range_1d(end = reduce_prod_32, start = range_1d_32_start_0, step = range_1d_32_step_0)[name = tensor("range_1d_32")]; + tensor reshape_160 = reshape(shape = shape_80, x = range_1d_32)[name = tensor("reshape_160")]; + tensor slice_by_index_32 = slice_by_index(begin = concat_114, begin_mask = new_cache_33_internal_tensor_assign_1_begin_mask_0, end = concat_115, end_mask = new_cache_33_internal_tensor_assign_1_end_mask_0, squeeze_mask = new_cache_33_internal_tensor_assign_1_squeeze_mask_0, stride = new_cache_33_internal_tensor_assign_1_stride_0, x = reshape_160)[name = tensor("slice_by_index_32")]; + tensor reshape_161_shape_0 = const()[name = tensor("reshape_161_shape_0"), val = tensor([-1])]; + tensor reshape_161 = reshape(shape = reshape_161_shape_0, x = slice_by_index_32)[name = tensor("reshape_161")]; + tensor reshape_162_shape_0 = const()[name = tensor("reshape_162_shape_0"), val = tensor([-1])]; + tensor reshape_162_cast_fp16 = reshape(shape = reshape_162_shape_0, x = var_6611_cast_fp16)[name = tensor("reshape_162_cast_fp16")]; + tensor reshape_163_shape_0 = const()[name = tensor("reshape_163_shape_0"), val = tensor([-1])]; + tensor reshape_163_cast_fp16 = reshape(shape = reshape_163_shape_0, x = cache16_to_fp16)[name = tensor("reshape_163_cast_fp16")]; + tensor scatter_32_mode_0 = const()[name = tensor("scatter_32_mode_0"), val = tensor("update")]; + tensor scatter_32_axis_0 = const()[name = tensor("scatter_32_axis_0"), val = tensor(0)]; + tensor scatter_32_validate_indices_0 = const()[name = tensor("scatter_32_validate_indices_0"), val = tensor(false)]; + tensor scatter_32_cast_fp16 = scatter(axis = scatter_32_axis_0, data = reshape_163_cast_fp16, indices = reshape_161, mode = scatter_32_mode_0, updates = reshape_162_cast_fp16, validate_indices = scatter_32_validate_indices_0)[name = tensor("scatter_32_cast_fp16")]; + tensor reshape_164_cast_fp16 = reshape(shape = shape_80, x = scatter_32_cast_fp16)[name = tensor("reshape_164_cast_fp16")]; + tensor var_6619_begin_0 = const()[name = tensor("op_6619_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor var_6619_end_0 = const()[name = tensor("op_6619_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor var_6619_end_mask_0 = const()[name = tensor("op_6619_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_6619_squeeze_mask_0 = const()[name = tensor("op_6619_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor var_6619_cast_fp16 = slice_by_index(begin = var_6619_begin_0, end = var_6619_end_0, end_mask = var_6619_end_mask_0, squeeze_mask = var_6619_squeeze_mask_0, x = reshape_164_cast_fp16)[name = tensor("op_6619_cast_fp16")]; + tensor var_6621_axis_0 = const()[name = tensor("op_6621_axis_0"), val = tensor(1)]; + tensor var_6621_mode_0 = const()[name = tensor("op_6621_mode_0"), val = tensor("update")]; + tensor var_6621_validate_indices_0 = const()[name = tensor("op_6621_validate_indices_0"), val = tensor(false)]; + tensor var_6621_cast_fp16 = scatter_along_axis(axis = var_6621_axis_0, data = var_6619_cast_fp16, indices = write_indices_33, mode = var_6621_mode_0, updates = v_33_cast_fp16, validate_indices = var_6621_validate_indices_0)[name = tensor("op_6621_cast_fp16")]; + tensor concat_116 = const()[name = tensor("concat_116"), val = tensor([1, 0, 0, 0, 0])]; + tensor concat_117 = const()[name = tensor("concat_117"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_33_internal_tensor_assign_2_stride_0 = const()[name = tensor("new_cache_33_internal_tensor_assign_2_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_33_internal_tensor_assign_2_begin_mask_0 = const()[name = tensor("new_cache_33_internal_tensor_assign_2_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_33_internal_tensor_assign_2_end_mask_0 = const()[name = tensor("new_cache_33_internal_tensor_assign_2_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_33_internal_tensor_assign_2_squeeze_mask_0 = const()[name = tensor("new_cache_33_internal_tensor_assign_2_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_81 = const()[name = tensor("shape_81"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_33 = const()[name = tensor("reduce_prod_33"), val = tensor(1048576)]; + tensor range_1d_33_start_0 = const()[name = tensor("range_1d_33_start_0"), val = tensor(0)]; + tensor range_1d_33_step_0 = const()[name = tensor("range_1d_33_step_0"), val = tensor(1)]; + tensor range_1d_33 = range_1d(end = reduce_prod_33, start = range_1d_33_start_0, step = range_1d_33_step_0)[name = tensor("range_1d_33")]; + tensor reshape_165 = reshape(shape = shape_81, x = range_1d_33)[name = tensor("reshape_165")]; + tensor slice_by_index_33 = slice_by_index(begin = concat_116, begin_mask = new_cache_33_internal_tensor_assign_2_begin_mask_0, end = concat_117, end_mask = new_cache_33_internal_tensor_assign_2_end_mask_0, squeeze_mask = new_cache_33_internal_tensor_assign_2_squeeze_mask_0, stride = new_cache_33_internal_tensor_assign_2_stride_0, x = reshape_165)[name = tensor("slice_by_index_33")]; + tensor reshape_166_shape_0 = const()[name = tensor("reshape_166_shape_0"), val = tensor([-1])]; + tensor reshape_166 = reshape(shape = reshape_166_shape_0, x = slice_by_index_33)[name = tensor("reshape_166")]; + tensor reshape_167_shape_0 = const()[name = tensor("reshape_167_shape_0"), val = tensor([-1])]; + tensor reshape_167_cast_fp16 = reshape(shape = reshape_167_shape_0, x = var_6621_cast_fp16)[name = tensor("reshape_167_cast_fp16")]; + tensor reshape_168_shape_0 = const()[name = tensor("reshape_168_shape_0"), val = tensor([-1])]; + tensor reshape_168_cast_fp16 = reshape(shape = reshape_168_shape_0, x = reshape_164_cast_fp16)[name = tensor("reshape_168_cast_fp16")]; + tensor scatter_33_mode_0 = const()[name = tensor("scatter_33_mode_0"), val = tensor("update")]; + tensor scatter_33_axis_0 = const()[name = tensor("scatter_33_axis_0"), val = tensor(0)]; + tensor scatter_33_validate_indices_0 = const()[name = tensor("scatter_33_validate_indices_0"), val = tensor(false)]; + tensor scatter_33_cast_fp16 = scatter(axis = scatter_33_axis_0, data = reshape_168_cast_fp16, indices = reshape_166, mode = scatter_33_mode_0, updates = reshape_167_cast_fp16, validate_indices = scatter_33_validate_indices_0)[name = tensor("scatter_33_cast_fp16")]; + tensor reshape_169_cast_fp16 = reshape(shape = shape_81, x = scatter_33_cast_fp16)[name = tensor("reshape_169_cast_fp16")]; + tensor reshape_169_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("reshape_169_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor keys_97_begin_0 = const()[name = tensor("keys_97_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor keys_97_end_0 = const()[name = tensor("keys_97_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor keys_97_end_mask_0 = const()[name = tensor("keys_97_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor keys_97_squeeze_mask_0 = const()[name = tensor("keys_97_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor keys_97_cast_fp16 = slice_by_index(begin = keys_97_begin_0, end = keys_97_end_0, end_mask = keys_97_end_mask_0, squeeze_mask = keys_97_squeeze_mask_0, x = reshape_169_cast_fp16)[name = tensor("keys_97_cast_fp16")]; + tensor values_97_begin_0 = const()[name = tensor("values_97_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor values_97_end_0 = const()[name = tensor("values_97_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor values_97_end_mask_0 = const()[name = tensor("values_97_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor values_97_squeeze_mask_0 = const()[name = tensor("values_97_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor values_97_cast_fp16 = slice_by_index(begin = values_97_begin_0, end = values_97_end_0, end_mask = values_97_end_mask_0, squeeze_mask = values_97_squeeze_mask_0, x = reshape_169_cast_fp16)[name = tensor("values_97_cast_fp16")]; + tensor var_6633_cast_fp16 = not_equal(x = keys_97_cast_fp16, y = keys_97_cast_fp16)[name = tensor("op_6633_cast_fp16")]; + tensor keys_99_cast_fp16 = select(a = var_463_to_fp16, b = keys_97_cast_fp16, cond = var_6633_cast_fp16)[name = tensor("keys_99_cast_fp16")]; + tensor var_6641_cast_fp16 = not_equal(x = values_97_cast_fp16, y = values_97_cast_fp16)[name = tensor("op_6641_cast_fp16")]; + tensor values_99_cast_fp16 = select(a = var_463_to_fp16, b = values_97_cast_fp16, cond = var_6641_cast_fp16)[name = tensor("values_99_cast_fp16")]; + tensor var_6665 = const()[name = tensor("op_6665"), val = tensor([0, 2, 1, 3])]; + tensor var_6681 = const()[name = tensor("op_6681"), val = tensor([1, 1, 1])]; + tensor var_6682_cast_fp16 = reshape(shape = var_6681, x = position16_to_fp16)[name = tensor("op_6682_cast_fp16")]; + tensor var_6702_to_fp16 = const()[name = tensor("op_6702_to_fp16"), val = tensor(0x1p+0)]; + tensor valid_len_33_cast_fp16 = add(x = var_6682_cast_fp16, y = var_6702_to_fp16)[name = tensor("valid_len_33_cast_fp16")]; + tensor valid_mask_33_cast_fp16 = less(x = k_positions_1_promoted_to_fp16, y = valid_len_33_cast_fp16)[name = tensor("valid_mask_33_cast_fp16")]; + tensor causal_mask_33_cast_fp16 = less_equal(x = k_positions_1_promoted_to_fp16, y = var_6682_cast_fp16)[name = tensor("causal_mask_33_cast_fp16")]; + tensor attn_mask_65 = logical_and(x = valid_mask_33_cast_fp16, y = causal_mask_33_cast_fp16)[name = tensor("attn_mask_65")]; + tensor attn_mask_67_axes_0 = const()[name = tensor("attn_mask_67_axes_0"), val = tensor([1])]; + tensor attn_mask_67 = expand_dims(axes = attn_mask_67_axes_0, x = attn_mask_65)[name = tensor("attn_mask_67")]; + tensor var_6720_transpose_x_0 = const()[name = tensor("op_6720_transpose_x_0"), val = tensor(false)]; + tensor var_6720_transpose_y_0 = const()[name = tensor("op_6720_transpose_y_0"), val = tensor(false)]; + tensor transpose_104_perm_0 = const()[name = tensor("transpose_104_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_105_perm_0 = const()[name = tensor("transpose_105_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_105 = transpose(perm = transpose_105_perm_0, x = keys_99_cast_fp16)[name = tensor("transpose_149")]; + tensor transpose_104 = transpose(perm = transpose_104_perm_0, x = q_99_cast_fp16)[name = tensor("transpose_150")]; + tensor var_6720_cast_fp16 = matmul(transpose_x = var_6720_transpose_x_0, transpose_y = var_6720_transpose_y_0, x = transpose_104, y = transpose_105)[name = tensor("op_6720_cast_fp16")]; + tensor var_6714_to_fp16 = const()[name = tensor("op_6714_to_fp16"), val = tensor([0x1p-3])]; + tensor attn_weights_97_cast_fp16 = mul(x = var_6720_cast_fp16, y = var_6714_to_fp16)[name = tensor("attn_weights_97_cast_fp16")]; + tensor var_6722 = logical_not(x = attn_mask_67)[name = tensor("op_6722")]; + tensor var_6723_to_fp16 = const()[name = tensor("op_6723_to_fp16"), val = tensor(-inf)]; + tensor attn_weights_99_cast_fp16 = select(a = var_6723_to_fp16, b = attn_weights_97_cast_fp16, cond = var_6722)[name = tensor("attn_weights_99_cast_fp16")]; + tensor var_6725 = const()[name = tensor("op_6725"), val = tensor(-1)]; + tensor attn_weights_101_cast_fp16 = softmax(axis = var_6725, x = attn_weights_99_cast_fp16)[name = tensor("attn_weights_101_cast_fp16")]; + tensor attn_output_33_transpose_x_0 = const()[name = tensor("attn_output_33_transpose_x_0"), val = tensor(false)]; + tensor attn_output_33_transpose_y_0 = const()[name = tensor("attn_output_33_transpose_y_0"), val = tensor(false)]; + tensor values_101_cast_fp16 = transpose(perm = var_6665, x = values_99_cast_fp16)[name = tensor("transpose_151")]; + tensor attn_output_33_cast_fp16 = matmul(transpose_x = attn_output_33_transpose_x_0, transpose_y = attn_output_33_transpose_y_0, x = attn_weights_101_cast_fp16, y = values_101_cast_fp16)[name = tensor("attn_output_33_cast_fp16")]; + tensor var_6733 = const()[name = tensor("op_6733"), val = tensor([0, 2, 1, 3])]; + tensor var_6738 = const()[name = tensor("op_6738"), val = tensor([1, 1, 1024])]; + tensor var_6734_cast_fp16 = transpose(perm = var_6733, x = attn_output_33_cast_fp16)[name = tensor("transpose_148")]; + tensor input_199_cast_fp16 = reshape(shape = var_6738, x = var_6734_cast_fp16)[name = tensor("input_199_cast_fp16")]; + tensor attn16_out_proj_weight_to_fp16 = const()[name = tensor("attn16_out_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(410222336)))]; + tensor linear_66_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = attn16_out_proj_weight_to_fp16, x = input_199_cast_fp16)[name = tensor("linear_66_cast_fp16")]; + tensor var_6744_to_fp16 = const()[name = tensor("op_6744_to_fp16"), val = tensor(0x1p+0)]; + tensor var_6745_cast_fp16 = add(x = position16_to_fp16, y = var_6744_to_fp16)[name = tensor("op_6745_cast_fp16")]; + tensor var_6745_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_6745_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor input_201_cast_fp16 = add(x = input_195_cast_fp16, y = linear_66_cast_fp16)[name = tensor("input_201_cast_fp16")]; + tensor input_203_axes_0 = const()[name = tensor("input_203_axes_0"), val = tensor([-1])]; + tensor norm16_2_weight_to_fp16 = const()[name = tensor("norm16_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(412319552)))]; + tensor norm16_2_bias_to_fp16 = const()[name = tensor("norm16_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(412321664)))]; + tensor var_6749_to_fp16 = const()[name = tensor("op_6749_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_203_cast_fp16 = layer_norm(axes = input_203_axes_0, beta = norm16_2_bias_to_fp16, epsilon = var_6749_to_fp16, gamma = norm16_2_weight_to_fp16, x = input_201_cast_fp16)[name = tensor("input_203_cast_fp16")]; + tensor linear16_1_weight_to_fp16 = const()[name = tensor("linear16_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(412323776)))]; + tensor linear_67_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = linear16_1_weight_to_fp16, x = input_203_cast_fp16)[name = tensor("linear_67_cast_fp16")]; + tensor input_205_mode_0 = const()[name = tensor("input_205_mode_0"), val = tensor("EXACT")]; + tensor input_205_cast_fp16 = gelu(mode = input_205_mode_0, x = linear_67_cast_fp16)[name = tensor("input_205_cast_fp16")]; + tensor linear16_2_weight_to_fp16 = const()[name = tensor("linear16_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(420712448)))]; + tensor linear_68_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = linear16_2_weight_to_fp16, x = input_205_cast_fp16)[name = tensor("linear_68_cast_fp16")]; + tensor input_207_cast_fp16 = add(x = input_201_cast_fp16, y = linear_68_cast_fp16)[name = tensor("input_207_cast_fp16")]; + tensor input_209_axes_0 = const()[name = tensor("input_209_axes_0"), val = tensor([-1])]; + tensor norm17_1_weight_to_fp16 = const()[name = tensor("norm17_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(429101120)))]; + tensor norm17_1_bias_to_fp16 = const()[name = tensor("norm17_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(429103232)))]; + tensor var_6766_to_fp16 = const()[name = tensor("op_6766_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_209_cast_fp16 = layer_norm(axes = input_209_axes_0, beta = norm17_1_bias_to_fp16, epsilon = var_6766_to_fp16, gamma = norm17_1_weight_to_fp16, x = input_207_cast_fp16)[name = tensor("input_209_cast_fp16")]; + tensor attn17_in_proj_weight_to_fp16 = const()[name = tensor("attn17_in_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(429105344)))]; + tensor linear_69_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = attn17_in_proj_weight_to_fp16, x = input_209_cast_fp16)[name = tensor("linear_69_cast_fp16")]; + tensor var_6780 = const()[name = tensor("op_6780"), val = tensor([1, 1, 3, 16, 64])]; + tensor qkv_35_cast_fp16 = reshape(shape = var_6780, x = linear_69_cast_fp16)[name = tensor("qkv_35_cast_fp16")]; + tensor q_103_begin_0 = const()[name = tensor("q_103_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor q_103_end_0 = const()[name = tensor("q_103_end_0"), val = tensor([1, 1, 1, 16, 64])]; + tensor q_103_end_mask_0 = const()[name = tensor("q_103_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor q_103_squeeze_mask_0 = const()[name = tensor("q_103_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor q_103_cast_fp16 = slice_by_index(begin = q_103_begin_0, end = q_103_end_0, end_mask = q_103_end_mask_0, squeeze_mask = q_103_squeeze_mask_0, x = qkv_35_cast_fp16)[name = tensor("q_103_cast_fp16")]; + tensor k_69_begin_0 = const()[name = tensor("k_69_begin_0"), val = tensor([0, 0, 1, 0, 0])]; + tensor k_69_end_0 = const()[name = tensor("k_69_end_0"), val = tensor([1, 1, 2, 16, 64])]; + tensor k_69_end_mask_0 = const()[name = tensor("k_69_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor k_69_squeeze_mask_0 = const()[name = tensor("k_69_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor k_69_cast_fp16 = slice_by_index(begin = k_69_begin_0, end = k_69_end_0, end_mask = k_69_end_mask_0, squeeze_mask = k_69_squeeze_mask_0, x = qkv_35_cast_fp16)[name = tensor("k_69_cast_fp16")]; + tensor v_35_begin_0 = const()[name = tensor("v_35_begin_0"), val = tensor([0, 0, 2, 0, 0])]; + tensor v_35_end_0 = const()[name = tensor("v_35_end_0"), val = tensor([1, 1, 3, 16, 64])]; + tensor v_35_end_mask_0 = const()[name = tensor("v_35_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor v_35_squeeze_mask_0 = const()[name = tensor("v_35_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor v_35_cast_fp16 = slice_by_index(begin = v_35_begin_0, end = v_35_end_0, end_mask = v_35_end_mask_0, squeeze_mask = v_35_squeeze_mask_0, x = qkv_35_cast_fp16)[name = tensor("v_35_cast_fp16")]; + tensor var_6846 = const()[name = tensor("op_6846"), val = tensor([1, 1, 1, 1])]; + tensor position17_to_fp16_dtype_0 = const()[name = tensor("position17_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor position17_to_fp16 = cast(dtype = position17_to_fp16_dtype_0, x = position17)[name = tensor("cast_468")]; + tensor ts_107_cast_fp16 = reshape(shape = var_6846, x = position17_to_fp16)[name = tensor("ts_107_cast_fp16")]; + tensor var_6853 = const()[name = tensor("op_6853"), val = tensor([1, 1, 16, 32, 2])]; + tensor q_complex_35_cast_fp16 = reshape(shape = var_6853, x = q_103_cast_fp16)[name = tensor("q_complex_35_cast_fp16")]; + tensor var_6860 = const()[name = tensor("op_6860"), val = tensor([1, 1, 16, 32, 2])]; + tensor k_complex_35_cast_fp16 = reshape(shape = var_6860, x = k_69_cast_fp16)[name = tensor("k_complex_35_cast_fp16")]; + tensor var_6864_begin_0 = const()[name = tensor("op_6864_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_6864_end_0 = const()[name = tensor("op_6864_end_0"), val = tensor([1, 1, 16, 32, 1])]; + tensor var_6864_end_mask_0 = const()[name = tensor("op_6864_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_6864_squeeze_mask_0 = const()[name = tensor("op_6864_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_6864_cast_fp16 = slice_by_index(begin = var_6864_begin_0, end = var_6864_end_0, end_mask = var_6864_end_mask_0, squeeze_mask = var_6864_squeeze_mask_0, x = q_complex_35_cast_fp16)[name = tensor("op_6864_cast_fp16")]; + tensor var_6872_begin_0 = const()[name = tensor("op_6872_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_6872_end_0 = const()[name = tensor("op_6872_end_0"), val = tensor([1, 1, 16, 32, 2])]; + tensor var_6872_end_mask_0 = const()[name = tensor("op_6872_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_6872_squeeze_mask_0 = const()[name = tensor("op_6872_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_6872_cast_fp16 = slice_by_index(begin = var_6872_begin_0, end = var_6872_end_0, end_mask = var_6872_end_mask_0, squeeze_mask = var_6872_squeeze_mask_0, x = q_complex_35_cast_fp16)[name = tensor("op_6872_cast_fp16")]; + tensor var_6880_begin_0 = const()[name = tensor("op_6880_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_6880_end_0 = const()[name = tensor("op_6880_end_0"), val = tensor([1, 1, 16, 32, 1])]; + tensor var_6880_end_mask_0 = const()[name = tensor("op_6880_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_6880_squeeze_mask_0 = const()[name = tensor("op_6880_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_6880_cast_fp16 = slice_by_index(begin = var_6880_begin_0, end = var_6880_end_0, end_mask = var_6880_end_mask_0, squeeze_mask = var_6880_squeeze_mask_0, x = k_complex_35_cast_fp16)[name = tensor("op_6880_cast_fp16")]; + tensor var_6888_begin_0 = const()[name = tensor("op_6888_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_6888_end_0 = const()[name = tensor("op_6888_end_0"), val = tensor([1, 1, 16, 32, 2])]; + tensor var_6888_end_mask_0 = const()[name = tensor("op_6888_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_6888_squeeze_mask_0 = const()[name = tensor("op_6888_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_6888_cast_fp16 = slice_by_index(begin = var_6888_begin_0, end = var_6888_end_0, end_mask = var_6888_end_mask_0, squeeze_mask = var_6888_squeeze_mask_0, x = k_complex_35_cast_fp16)[name = tensor("op_6888_cast_fp16")]; + tensor freqs_35_to_fp16 = const()[name = tensor("freqs_35_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(435396864)))]; + tensor var_6894_cast_fp16 = mul(x = freqs_35_to_fp16, y = ts_107_cast_fp16)[name = tensor("op_6894_cast_fp16")]; + tensor rotr_35_cast_fp16 = cos(x = var_6894_cast_fp16)[name = tensor("rotr_35_cast_fp16")]; + tensor roti_35_cast_fp16 = sin(x = var_6894_cast_fp16)[name = tensor("roti_35_cast_fp16")]; + tensor var_6898_cast_fp16 = mul(x = var_6864_cast_fp16, y = rotr_35_cast_fp16)[name = tensor("op_6898_cast_fp16")]; + tensor var_6899_cast_fp16 = mul(x = var_6872_cast_fp16, y = roti_35_cast_fp16)[name = tensor("op_6899_cast_fp16")]; + tensor qor_69_cast_fp16 = sub(x = var_6898_cast_fp16, y = var_6899_cast_fp16)[name = tensor("qor_69_cast_fp16")]; + tensor var_6902_cast_fp16 = mul(x = var_6864_cast_fp16, y = roti_35_cast_fp16)[name = tensor("op_6902_cast_fp16")]; + tensor var_6903_cast_fp16 = mul(x = var_6872_cast_fp16, y = rotr_35_cast_fp16)[name = tensor("op_6903_cast_fp16")]; + tensor qoi_69_cast_fp16 = add(x = var_6902_cast_fp16, y = var_6903_cast_fp16)[name = tensor("qoi_69_cast_fp16")]; + tensor var_6906_cast_fp16 = mul(x = var_6880_cast_fp16, y = rotr_35_cast_fp16)[name = tensor("op_6906_cast_fp16")]; + tensor var_6907_cast_fp16 = mul(x = var_6888_cast_fp16, y = roti_35_cast_fp16)[name = tensor("op_6907_cast_fp16")]; + tensor kor_69_cast_fp16 = sub(x = var_6906_cast_fp16, y = var_6907_cast_fp16)[name = tensor("kor_69_cast_fp16")]; + tensor var_6910_cast_fp16 = mul(x = var_6880_cast_fp16, y = roti_35_cast_fp16)[name = tensor("op_6910_cast_fp16")]; + tensor var_6911_cast_fp16 = mul(x = var_6888_cast_fp16, y = rotr_35_cast_fp16)[name = tensor("op_6911_cast_fp16")]; + tensor koi_69_cast_fp16 = add(x = var_6910_cast_fp16, y = var_6911_cast_fp16)[name = tensor("koi_69_cast_fp16")]; + tensor qo_35_axis_0 = const()[name = tensor("qo_35_axis_0"), val = tensor(-1)]; + tensor qo_35_cast_fp16 = stack(axis = qo_35_axis_0, values = (qor_69_cast_fp16, qoi_69_cast_fp16))[name = tensor("qo_35_cast_fp16")]; + tensor ko_35_axis_0 = const()[name = tensor("ko_35_axis_0"), val = tensor(-1)]; + tensor ko_35_cast_fp16 = stack(axis = ko_35_axis_0, values = (kor_69_cast_fp16, koi_69_cast_fp16))[name = tensor("ko_35_cast_fp16")]; + tensor var_6944 = const()[name = tensor("op_6944"), val = tensor([1, 1, 16, 64])]; + tensor q_105_cast_fp16 = reshape(shape = var_6944, x = qo_35_cast_fp16)[name = tensor("q_105_cast_fp16")]; + tensor var_6950 = const()[name = tensor("op_6950"), val = tensor([1, 1, 16, 64])]; + tensor k_71_cast_fp16 = reshape(shape = var_6950, x = ko_35_cast_fp16)[name = tensor("k_71_cast_fp16")]; + tensor _inversed_6975_y_0_to_fp16 = const()[name = tensor("_inversed_6975_y_0_to_fp16"), val = tensor(0x1p-9)]; + tensor _inversed_6975_cast_fp16 = mul(x = ts_107_cast_fp16, y = _inversed_6975_y_0_to_fp16)[name = tensor("_inversed_6975_cast_fp16")]; + tensor var_6976_cast_fp16 = floor(x = _inversed_6975_cast_fp16)[name = tensor("op_6976_cast_fp16")]; + tensor var_6977_to_fp16 = const()[name = tensor("op_6977_to_fp16"), val = tensor(0x1p+9)]; + tensor var_6978_cast_fp16 = mul(x = var_6976_cast_fp16, y = var_6977_to_fp16)[name = tensor("op_6978_cast_fp16")]; + tensor write_indices_float_71_cast_fp16 = sub(x = ts_107_cast_fp16, y = var_6978_cast_fp16)[name = tensor("write_indices_float_71_cast_fp16")]; + tensor var_6985_dtype_0 = const()[name = tensor("op_6985_dtype_0"), val = tensor("int32")]; + tensor write_indices_35_reps_0 = const()[name = tensor("write_indices_35_reps_0"), val = tensor([1, 1, 16, 64])]; + tensor write_indices_float_71_cast_fp16_to_int32 = cast(dtype = var_6985_dtype_0, x = write_indices_float_71_cast_fp16)[name = tensor("cast_467")]; + tensor write_indices_35 = tile(reps = write_indices_35_reps_0, x = write_indices_float_71_cast_fp16_to_int32)[name = tensor("write_indices_35")]; + tensor var_6995_begin_0 = const()[name = tensor("op_6995_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_6995_end_0 = const()[name = tensor("op_6995_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor var_6995_end_mask_0 = const()[name = tensor("op_6995_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_6995_squeeze_mask_0 = const()[name = tensor("op_6995_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor cache17_to_fp16_dtype_0 = const()[name = tensor("cache17_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor cache17_to_fp16 = cast(dtype = cache17_to_fp16_dtype_0, x = cache17)[name = tensor("cast_466")]; + tensor var_6995_cast_fp16 = slice_by_index(begin = var_6995_begin_0, end = var_6995_end_0, end_mask = var_6995_end_mask_0, squeeze_mask = var_6995_squeeze_mask_0, x = cache17_to_fp16)[name = tensor("op_6995_cast_fp16")]; + tensor var_6997_axis_0 = const()[name = tensor("op_6997_axis_0"), val = tensor(1)]; + tensor var_6997_mode_0 = const()[name = tensor("op_6997_mode_0"), val = tensor("update")]; + tensor var_6997_validate_indices_0 = const()[name = tensor("op_6997_validate_indices_0"), val = tensor(false)]; + tensor var_6997_cast_fp16 = scatter_along_axis(axis = var_6997_axis_0, data = var_6995_cast_fp16, indices = write_indices_35, mode = var_6997_mode_0, updates = k_71_cast_fp16, validate_indices = var_6997_validate_indices_0)[name = tensor("op_6997_cast_fp16")]; + tensor concat_121 = const()[name = tensor("concat_121"), val = tensor([0, 0, 0, 0, 0])]; + tensor concat_122 = const()[name = tensor("concat_122"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_35_internal_tensor_assign_1_stride_0 = const()[name = tensor("new_cache_35_internal_tensor_assign_1_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_35_internal_tensor_assign_1_begin_mask_0 = const()[name = tensor("new_cache_35_internal_tensor_assign_1_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_35_internal_tensor_assign_1_end_mask_0 = const()[name = tensor("new_cache_35_internal_tensor_assign_1_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_35_internal_tensor_assign_1_squeeze_mask_0 = const()[name = tensor("new_cache_35_internal_tensor_assign_1_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_82 = const()[name = tensor("shape_82"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_34 = const()[name = tensor("reduce_prod_34"), val = tensor(1048576)]; + tensor range_1d_34_start_0 = const()[name = tensor("range_1d_34_start_0"), val = tensor(0)]; + tensor range_1d_34_step_0 = const()[name = tensor("range_1d_34_step_0"), val = tensor(1)]; + tensor range_1d_34 = range_1d(end = reduce_prod_34, start = range_1d_34_start_0, step = range_1d_34_step_0)[name = tensor("range_1d_34")]; + tensor reshape_170 = reshape(shape = shape_82, x = range_1d_34)[name = tensor("reshape_170")]; + tensor slice_by_index_34 = slice_by_index(begin = concat_121, begin_mask = new_cache_35_internal_tensor_assign_1_begin_mask_0, end = concat_122, end_mask = new_cache_35_internal_tensor_assign_1_end_mask_0, squeeze_mask = new_cache_35_internal_tensor_assign_1_squeeze_mask_0, stride = new_cache_35_internal_tensor_assign_1_stride_0, x = reshape_170)[name = tensor("slice_by_index_34")]; + tensor reshape_171_shape_0 = const()[name = tensor("reshape_171_shape_0"), val = tensor([-1])]; + tensor reshape_171 = reshape(shape = reshape_171_shape_0, x = slice_by_index_34)[name = tensor("reshape_171")]; + tensor reshape_172_shape_0 = const()[name = tensor("reshape_172_shape_0"), val = tensor([-1])]; + tensor reshape_172_cast_fp16 = reshape(shape = reshape_172_shape_0, x = var_6997_cast_fp16)[name = tensor("reshape_172_cast_fp16")]; + tensor reshape_173_shape_0 = const()[name = tensor("reshape_173_shape_0"), val = tensor([-1])]; + tensor reshape_173_cast_fp16 = reshape(shape = reshape_173_shape_0, x = cache17_to_fp16)[name = tensor("reshape_173_cast_fp16")]; + tensor scatter_34_mode_0 = const()[name = tensor("scatter_34_mode_0"), val = tensor("update")]; + tensor scatter_34_axis_0 = const()[name = tensor("scatter_34_axis_0"), val = tensor(0)]; + tensor scatter_34_validate_indices_0 = const()[name = tensor("scatter_34_validate_indices_0"), val = tensor(false)]; + tensor scatter_34_cast_fp16 = scatter(axis = scatter_34_axis_0, data = reshape_173_cast_fp16, indices = reshape_171, mode = scatter_34_mode_0, updates = reshape_172_cast_fp16, validate_indices = scatter_34_validate_indices_0)[name = tensor("scatter_34_cast_fp16")]; + tensor reshape_174_cast_fp16 = reshape(shape = shape_82, x = scatter_34_cast_fp16)[name = tensor("reshape_174_cast_fp16")]; + tensor var_7005_begin_0 = const()[name = tensor("op_7005_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor var_7005_end_0 = const()[name = tensor("op_7005_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor var_7005_end_mask_0 = const()[name = tensor("op_7005_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_7005_squeeze_mask_0 = const()[name = tensor("op_7005_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor var_7005_cast_fp16 = slice_by_index(begin = var_7005_begin_0, end = var_7005_end_0, end_mask = var_7005_end_mask_0, squeeze_mask = var_7005_squeeze_mask_0, x = reshape_174_cast_fp16)[name = tensor("op_7005_cast_fp16")]; + tensor var_7007_axis_0 = const()[name = tensor("op_7007_axis_0"), val = tensor(1)]; + tensor var_7007_mode_0 = const()[name = tensor("op_7007_mode_0"), val = tensor("update")]; + tensor var_7007_validate_indices_0 = const()[name = tensor("op_7007_validate_indices_0"), val = tensor(false)]; + tensor var_7007_cast_fp16 = scatter_along_axis(axis = var_7007_axis_0, data = var_7005_cast_fp16, indices = write_indices_35, mode = var_7007_mode_0, updates = v_35_cast_fp16, validate_indices = var_7007_validate_indices_0)[name = tensor("op_7007_cast_fp16")]; + tensor concat_123 = const()[name = tensor("concat_123"), val = tensor([1, 0, 0, 0, 0])]; + tensor concat_124 = const()[name = tensor("concat_124"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_35_internal_tensor_assign_2_stride_0 = const()[name = tensor("new_cache_35_internal_tensor_assign_2_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_35_internal_tensor_assign_2_begin_mask_0 = const()[name = tensor("new_cache_35_internal_tensor_assign_2_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_35_internal_tensor_assign_2_end_mask_0 = const()[name = tensor("new_cache_35_internal_tensor_assign_2_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_35_internal_tensor_assign_2_squeeze_mask_0 = const()[name = tensor("new_cache_35_internal_tensor_assign_2_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_83 = const()[name = tensor("shape_83"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_35 = const()[name = tensor("reduce_prod_35"), val = tensor(1048576)]; + tensor range_1d_35_start_0 = const()[name = tensor("range_1d_35_start_0"), val = tensor(0)]; + tensor range_1d_35_step_0 = const()[name = tensor("range_1d_35_step_0"), val = tensor(1)]; + tensor range_1d_35 = range_1d(end = reduce_prod_35, start = range_1d_35_start_0, step = range_1d_35_step_0)[name = tensor("range_1d_35")]; + tensor reshape_175 = reshape(shape = shape_83, x = range_1d_35)[name = tensor("reshape_175")]; + tensor slice_by_index_35 = slice_by_index(begin = concat_123, begin_mask = new_cache_35_internal_tensor_assign_2_begin_mask_0, end = concat_124, end_mask = new_cache_35_internal_tensor_assign_2_end_mask_0, squeeze_mask = new_cache_35_internal_tensor_assign_2_squeeze_mask_0, stride = new_cache_35_internal_tensor_assign_2_stride_0, x = reshape_175)[name = tensor("slice_by_index_35")]; + tensor reshape_176_shape_0 = const()[name = tensor("reshape_176_shape_0"), val = tensor([-1])]; + tensor reshape_176 = reshape(shape = reshape_176_shape_0, x = slice_by_index_35)[name = tensor("reshape_176")]; + tensor reshape_177_shape_0 = const()[name = tensor("reshape_177_shape_0"), val = tensor([-1])]; + tensor reshape_177_cast_fp16 = reshape(shape = reshape_177_shape_0, x = var_7007_cast_fp16)[name = tensor("reshape_177_cast_fp16")]; + tensor reshape_178_shape_0 = const()[name = tensor("reshape_178_shape_0"), val = tensor([-1])]; + tensor reshape_178_cast_fp16 = reshape(shape = reshape_178_shape_0, x = reshape_174_cast_fp16)[name = tensor("reshape_178_cast_fp16")]; + tensor scatter_35_mode_0 = const()[name = tensor("scatter_35_mode_0"), val = tensor("update")]; + tensor scatter_35_axis_0 = const()[name = tensor("scatter_35_axis_0"), val = tensor(0)]; + tensor scatter_35_validate_indices_0 = const()[name = tensor("scatter_35_validate_indices_0"), val = tensor(false)]; + tensor scatter_35_cast_fp16 = scatter(axis = scatter_35_axis_0, data = reshape_178_cast_fp16, indices = reshape_176, mode = scatter_35_mode_0, updates = reshape_177_cast_fp16, validate_indices = scatter_35_validate_indices_0)[name = tensor("scatter_35_cast_fp16")]; + tensor reshape_179_cast_fp16 = reshape(shape = shape_83, x = scatter_35_cast_fp16)[name = tensor("reshape_179_cast_fp16")]; + tensor reshape_179_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("reshape_179_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor keys_103_begin_0 = const()[name = tensor("keys_103_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor keys_103_end_0 = const()[name = tensor("keys_103_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor keys_103_end_mask_0 = const()[name = tensor("keys_103_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor keys_103_squeeze_mask_0 = const()[name = tensor("keys_103_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor keys_103_cast_fp16 = slice_by_index(begin = keys_103_begin_0, end = keys_103_end_0, end_mask = keys_103_end_mask_0, squeeze_mask = keys_103_squeeze_mask_0, x = reshape_179_cast_fp16)[name = tensor("keys_103_cast_fp16")]; + tensor values_103_begin_0 = const()[name = tensor("values_103_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor values_103_end_0 = const()[name = tensor("values_103_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor values_103_end_mask_0 = const()[name = tensor("values_103_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor values_103_squeeze_mask_0 = const()[name = tensor("values_103_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor values_103_cast_fp16 = slice_by_index(begin = values_103_begin_0, end = values_103_end_0, end_mask = values_103_end_mask_0, squeeze_mask = values_103_squeeze_mask_0, x = reshape_179_cast_fp16)[name = tensor("values_103_cast_fp16")]; + tensor var_7019_cast_fp16 = not_equal(x = keys_103_cast_fp16, y = keys_103_cast_fp16)[name = tensor("op_7019_cast_fp16")]; + tensor keys_105_cast_fp16 = select(a = var_463_to_fp16, b = keys_103_cast_fp16, cond = var_7019_cast_fp16)[name = tensor("keys_105_cast_fp16")]; + tensor var_7027_cast_fp16 = not_equal(x = values_103_cast_fp16, y = values_103_cast_fp16)[name = tensor("op_7027_cast_fp16")]; + tensor values_105_cast_fp16 = select(a = var_463_to_fp16, b = values_103_cast_fp16, cond = var_7027_cast_fp16)[name = tensor("values_105_cast_fp16")]; + tensor var_7051 = const()[name = tensor("op_7051"), val = tensor([0, 2, 1, 3])]; + tensor var_7067 = const()[name = tensor("op_7067"), val = tensor([1, 1, 1])]; + tensor var_7068_cast_fp16 = reshape(shape = var_7067, x = position17_to_fp16)[name = tensor("op_7068_cast_fp16")]; + tensor var_7088_to_fp16 = const()[name = tensor("op_7088_to_fp16"), val = tensor(0x1p+0)]; + tensor valid_len_35_cast_fp16 = add(x = var_7068_cast_fp16, y = var_7088_to_fp16)[name = tensor("valid_len_35_cast_fp16")]; + tensor valid_mask_35_cast_fp16 = less(x = k_positions_1_promoted_to_fp16, y = valid_len_35_cast_fp16)[name = tensor("valid_mask_35_cast_fp16")]; + tensor causal_mask_35_cast_fp16 = less_equal(x = k_positions_1_promoted_to_fp16, y = var_7068_cast_fp16)[name = tensor("causal_mask_35_cast_fp16")]; + tensor attn_mask_69 = logical_and(x = valid_mask_35_cast_fp16, y = causal_mask_35_cast_fp16)[name = tensor("attn_mask_69")]; + tensor attn_mask_71_axes_0 = const()[name = tensor("attn_mask_71_axes_0"), val = tensor([1])]; + tensor attn_mask_71 = expand_dims(axes = attn_mask_71_axes_0, x = attn_mask_69)[name = tensor("attn_mask_71")]; + tensor var_7106_transpose_x_0 = const()[name = tensor("op_7106_transpose_x_0"), val = tensor(false)]; + tensor var_7106_transpose_y_0 = const()[name = tensor("op_7106_transpose_y_0"), val = tensor(false)]; + tensor transpose_106_perm_0 = const()[name = tensor("transpose_106_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_107_perm_0 = const()[name = tensor("transpose_107_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_107 = transpose(perm = transpose_107_perm_0, x = keys_105_cast_fp16)[name = tensor("transpose_145")]; + tensor transpose_106 = transpose(perm = transpose_106_perm_0, x = q_105_cast_fp16)[name = tensor("transpose_146")]; + tensor var_7106_cast_fp16 = matmul(transpose_x = var_7106_transpose_x_0, transpose_y = var_7106_transpose_y_0, x = transpose_106, y = transpose_107)[name = tensor("op_7106_cast_fp16")]; + tensor var_7100_to_fp16 = const()[name = tensor("op_7100_to_fp16"), val = tensor([0x1p-3])]; + tensor attn_weights_103_cast_fp16 = mul(x = var_7106_cast_fp16, y = var_7100_to_fp16)[name = tensor("attn_weights_103_cast_fp16")]; + tensor var_7108 = logical_not(x = attn_mask_71)[name = tensor("op_7108")]; + tensor var_7109_to_fp16 = const()[name = tensor("op_7109_to_fp16"), val = tensor(-inf)]; + tensor attn_weights_105_cast_fp16 = select(a = var_7109_to_fp16, b = attn_weights_103_cast_fp16, cond = var_7108)[name = tensor("attn_weights_105_cast_fp16")]; + tensor var_7111 = const()[name = tensor("op_7111"), val = tensor(-1)]; + tensor attn_weights_107_cast_fp16 = softmax(axis = var_7111, x = attn_weights_105_cast_fp16)[name = tensor("attn_weights_107_cast_fp16")]; + tensor attn_output_35_transpose_x_0 = const()[name = tensor("attn_output_35_transpose_x_0"), val = tensor(false)]; + tensor attn_output_35_transpose_y_0 = const()[name = tensor("attn_output_35_transpose_y_0"), val = tensor(false)]; + tensor values_107_cast_fp16 = transpose(perm = var_7051, x = values_105_cast_fp16)[name = tensor("transpose_147")]; + tensor attn_output_35_cast_fp16 = matmul(transpose_x = attn_output_35_transpose_x_0, transpose_y = attn_output_35_transpose_y_0, x = attn_weights_107_cast_fp16, y = values_107_cast_fp16)[name = tensor("attn_output_35_cast_fp16")]; + tensor var_7119 = const()[name = tensor("op_7119"), val = tensor([0, 2, 1, 3])]; + tensor var_7124 = const()[name = tensor("op_7124"), val = tensor([1, 1, 1024])]; + tensor var_7120_cast_fp16 = transpose(perm = var_7119, x = attn_output_35_cast_fp16)[name = tensor("transpose_144")]; + tensor input_211_cast_fp16 = reshape(shape = var_7124, x = var_7120_cast_fp16)[name = tensor("input_211_cast_fp16")]; + tensor attn17_out_proj_weight_to_fp16 = const()[name = tensor("attn17_out_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(435396992)))]; + tensor linear_70_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = attn17_out_proj_weight_to_fp16, x = input_211_cast_fp16)[name = tensor("linear_70_cast_fp16")]; + tensor var_7130_to_fp16 = const()[name = tensor("op_7130_to_fp16"), val = tensor(0x1p+0)]; + tensor var_7131_cast_fp16 = add(x = position17_to_fp16, y = var_7130_to_fp16)[name = tensor("op_7131_cast_fp16")]; + tensor var_7131_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_7131_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor input_213_cast_fp16 = add(x = input_207_cast_fp16, y = linear_70_cast_fp16)[name = tensor("input_213_cast_fp16")]; + tensor input_215_axes_0 = const()[name = tensor("input_215_axes_0"), val = tensor([-1])]; + tensor norm17_2_weight_to_fp16 = const()[name = tensor("norm17_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(437494208)))]; + tensor norm17_2_bias_to_fp16 = const()[name = tensor("norm17_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(437496320)))]; + tensor var_7135_to_fp16 = const()[name = tensor("op_7135_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_215_cast_fp16 = layer_norm(axes = input_215_axes_0, beta = norm17_2_bias_to_fp16, epsilon = var_7135_to_fp16, gamma = norm17_2_weight_to_fp16, x = input_213_cast_fp16)[name = tensor("input_215_cast_fp16")]; + tensor linear17_1_weight_to_fp16 = const()[name = tensor("linear17_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(437498432)))]; + tensor linear_71_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = linear17_1_weight_to_fp16, x = input_215_cast_fp16)[name = tensor("linear_71_cast_fp16")]; + tensor input_217_mode_0 = const()[name = tensor("input_217_mode_0"), val = tensor("EXACT")]; + tensor input_217_cast_fp16 = gelu(mode = input_217_mode_0, x = linear_71_cast_fp16)[name = tensor("input_217_cast_fp16")]; + tensor linear17_2_weight_to_fp16 = const()[name = tensor("linear17_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(445887104)))]; + tensor linear_72_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = linear17_2_weight_to_fp16, x = input_217_cast_fp16)[name = tensor("linear_72_cast_fp16")]; + tensor input_219_cast_fp16 = add(x = input_213_cast_fp16, y = linear_72_cast_fp16)[name = tensor("input_219_cast_fp16")]; + tensor input_221_axes_0 = const()[name = tensor("input_221_axes_0"), val = tensor([-1])]; + tensor norm18_1_weight_to_fp16 = const()[name = tensor("norm18_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(454275776)))]; + tensor norm18_1_bias_to_fp16 = const()[name = tensor("norm18_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(454277888)))]; + tensor var_7152_to_fp16 = const()[name = tensor("op_7152_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_221_cast_fp16 = layer_norm(axes = input_221_axes_0, beta = norm18_1_bias_to_fp16, epsilon = var_7152_to_fp16, gamma = norm18_1_weight_to_fp16, x = input_219_cast_fp16)[name = tensor("input_221_cast_fp16")]; + tensor attn18_in_proj_weight_to_fp16 = const()[name = tensor("attn18_in_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(454280000)))]; + tensor linear_73_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = attn18_in_proj_weight_to_fp16, x = input_221_cast_fp16)[name = tensor("linear_73_cast_fp16")]; + tensor var_7166 = const()[name = tensor("op_7166"), val = tensor([1, 1, 3, 16, 64])]; + tensor qkv_37_cast_fp16 = reshape(shape = var_7166, x = linear_73_cast_fp16)[name = tensor("qkv_37_cast_fp16")]; + tensor q_109_begin_0 = const()[name = tensor("q_109_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor q_109_end_0 = const()[name = tensor("q_109_end_0"), val = tensor([1, 1, 1, 16, 64])]; + tensor q_109_end_mask_0 = const()[name = tensor("q_109_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor q_109_squeeze_mask_0 = const()[name = tensor("q_109_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor q_109_cast_fp16 = slice_by_index(begin = q_109_begin_0, end = q_109_end_0, end_mask = q_109_end_mask_0, squeeze_mask = q_109_squeeze_mask_0, x = qkv_37_cast_fp16)[name = tensor("q_109_cast_fp16")]; + tensor k_73_begin_0 = const()[name = tensor("k_73_begin_0"), val = tensor([0, 0, 1, 0, 0])]; + tensor k_73_end_0 = const()[name = tensor("k_73_end_0"), val = tensor([1, 1, 2, 16, 64])]; + tensor k_73_end_mask_0 = const()[name = tensor("k_73_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor k_73_squeeze_mask_0 = const()[name = tensor("k_73_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor k_73_cast_fp16 = slice_by_index(begin = k_73_begin_0, end = k_73_end_0, end_mask = k_73_end_mask_0, squeeze_mask = k_73_squeeze_mask_0, x = qkv_37_cast_fp16)[name = tensor("k_73_cast_fp16")]; + tensor v_37_begin_0 = const()[name = tensor("v_37_begin_0"), val = tensor([0, 0, 2, 0, 0])]; + tensor v_37_end_0 = const()[name = tensor("v_37_end_0"), val = tensor([1, 1, 3, 16, 64])]; + tensor v_37_end_mask_0 = const()[name = tensor("v_37_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor v_37_squeeze_mask_0 = const()[name = tensor("v_37_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor v_37_cast_fp16 = slice_by_index(begin = v_37_begin_0, end = v_37_end_0, end_mask = v_37_end_mask_0, squeeze_mask = v_37_squeeze_mask_0, x = qkv_37_cast_fp16)[name = tensor("v_37_cast_fp16")]; + tensor var_7232 = const()[name = tensor("op_7232"), val = tensor([1, 1, 1, 1])]; + tensor position18_to_fp16_dtype_0 = const()[name = tensor("position18_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor position18_to_fp16 = cast(dtype = position18_to_fp16_dtype_0, x = position18)[name = tensor("cast_463")]; + tensor ts_113_cast_fp16 = reshape(shape = var_7232, x = position18_to_fp16)[name = tensor("ts_113_cast_fp16")]; + tensor var_7239 = const()[name = tensor("op_7239"), val = tensor([1, 1, 16, 32, 2])]; + tensor q_complex_37_cast_fp16 = reshape(shape = var_7239, x = q_109_cast_fp16)[name = tensor("q_complex_37_cast_fp16")]; + tensor var_7246 = const()[name = tensor("op_7246"), val = tensor([1, 1, 16, 32, 2])]; + tensor k_complex_37_cast_fp16 = reshape(shape = var_7246, x = k_73_cast_fp16)[name = tensor("k_complex_37_cast_fp16")]; + tensor var_7250_begin_0 = const()[name = tensor("op_7250_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_7250_end_0 = const()[name = tensor("op_7250_end_0"), val = tensor([1, 1, 16, 32, 1])]; + tensor var_7250_end_mask_0 = const()[name = tensor("op_7250_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_7250_squeeze_mask_0 = const()[name = tensor("op_7250_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_7250_cast_fp16 = slice_by_index(begin = var_7250_begin_0, end = var_7250_end_0, end_mask = var_7250_end_mask_0, squeeze_mask = var_7250_squeeze_mask_0, x = q_complex_37_cast_fp16)[name = tensor("op_7250_cast_fp16")]; + tensor var_7258_begin_0 = const()[name = tensor("op_7258_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_7258_end_0 = const()[name = tensor("op_7258_end_0"), val = tensor([1, 1, 16, 32, 2])]; + tensor var_7258_end_mask_0 = const()[name = tensor("op_7258_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_7258_squeeze_mask_0 = const()[name = tensor("op_7258_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_7258_cast_fp16 = slice_by_index(begin = var_7258_begin_0, end = var_7258_end_0, end_mask = var_7258_end_mask_0, squeeze_mask = var_7258_squeeze_mask_0, x = q_complex_37_cast_fp16)[name = tensor("op_7258_cast_fp16")]; + tensor var_7266_begin_0 = const()[name = tensor("op_7266_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_7266_end_0 = const()[name = tensor("op_7266_end_0"), val = tensor([1, 1, 16, 32, 1])]; + tensor var_7266_end_mask_0 = const()[name = tensor("op_7266_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_7266_squeeze_mask_0 = const()[name = tensor("op_7266_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_7266_cast_fp16 = slice_by_index(begin = var_7266_begin_0, end = var_7266_end_0, end_mask = var_7266_end_mask_0, squeeze_mask = var_7266_squeeze_mask_0, x = k_complex_37_cast_fp16)[name = tensor("op_7266_cast_fp16")]; + tensor var_7274_begin_0 = const()[name = tensor("op_7274_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_7274_end_0 = const()[name = tensor("op_7274_end_0"), val = tensor([1, 1, 16, 32, 2])]; + tensor var_7274_end_mask_0 = const()[name = tensor("op_7274_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_7274_squeeze_mask_0 = const()[name = tensor("op_7274_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_7274_cast_fp16 = slice_by_index(begin = var_7274_begin_0, end = var_7274_end_0, end_mask = var_7274_end_mask_0, squeeze_mask = var_7274_squeeze_mask_0, x = k_complex_37_cast_fp16)[name = tensor("op_7274_cast_fp16")]; + tensor freqs_37_to_fp16 = const()[name = tensor("freqs_37_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(460571520)))]; + tensor var_7280_cast_fp16 = mul(x = freqs_37_to_fp16, y = ts_113_cast_fp16)[name = tensor("op_7280_cast_fp16")]; + tensor rotr_37_cast_fp16 = cos(x = var_7280_cast_fp16)[name = tensor("rotr_37_cast_fp16")]; + tensor roti_37_cast_fp16 = sin(x = var_7280_cast_fp16)[name = tensor("roti_37_cast_fp16")]; + tensor var_7284_cast_fp16 = mul(x = var_7250_cast_fp16, y = rotr_37_cast_fp16)[name = tensor("op_7284_cast_fp16")]; + tensor var_7285_cast_fp16 = mul(x = var_7258_cast_fp16, y = roti_37_cast_fp16)[name = tensor("op_7285_cast_fp16")]; + tensor qor_73_cast_fp16 = sub(x = var_7284_cast_fp16, y = var_7285_cast_fp16)[name = tensor("qor_73_cast_fp16")]; + tensor var_7288_cast_fp16 = mul(x = var_7250_cast_fp16, y = roti_37_cast_fp16)[name = tensor("op_7288_cast_fp16")]; + tensor var_7289_cast_fp16 = mul(x = var_7258_cast_fp16, y = rotr_37_cast_fp16)[name = tensor("op_7289_cast_fp16")]; + tensor qoi_73_cast_fp16 = add(x = var_7288_cast_fp16, y = var_7289_cast_fp16)[name = tensor("qoi_73_cast_fp16")]; + tensor var_7292_cast_fp16 = mul(x = var_7266_cast_fp16, y = rotr_37_cast_fp16)[name = tensor("op_7292_cast_fp16")]; + tensor var_7293_cast_fp16 = mul(x = var_7274_cast_fp16, y = roti_37_cast_fp16)[name = tensor("op_7293_cast_fp16")]; + tensor kor_73_cast_fp16 = sub(x = var_7292_cast_fp16, y = var_7293_cast_fp16)[name = tensor("kor_73_cast_fp16")]; + tensor var_7296_cast_fp16 = mul(x = var_7266_cast_fp16, y = roti_37_cast_fp16)[name = tensor("op_7296_cast_fp16")]; + tensor var_7297_cast_fp16 = mul(x = var_7274_cast_fp16, y = rotr_37_cast_fp16)[name = tensor("op_7297_cast_fp16")]; + tensor koi_73_cast_fp16 = add(x = var_7296_cast_fp16, y = var_7297_cast_fp16)[name = tensor("koi_73_cast_fp16")]; + tensor qo_37_axis_0 = const()[name = tensor("qo_37_axis_0"), val = tensor(-1)]; + tensor qo_37_cast_fp16 = stack(axis = qo_37_axis_0, values = (qor_73_cast_fp16, qoi_73_cast_fp16))[name = tensor("qo_37_cast_fp16")]; + tensor ko_37_axis_0 = const()[name = tensor("ko_37_axis_0"), val = tensor(-1)]; + tensor ko_37_cast_fp16 = stack(axis = ko_37_axis_0, values = (kor_73_cast_fp16, koi_73_cast_fp16))[name = tensor("ko_37_cast_fp16")]; + tensor var_7330 = const()[name = tensor("op_7330"), val = tensor([1, 1, 16, 64])]; + tensor q_111_cast_fp16 = reshape(shape = var_7330, x = qo_37_cast_fp16)[name = tensor("q_111_cast_fp16")]; + tensor var_7336 = const()[name = tensor("op_7336"), val = tensor([1, 1, 16, 64])]; + tensor k_75_cast_fp16 = reshape(shape = var_7336, x = ko_37_cast_fp16)[name = tensor("k_75_cast_fp16")]; + tensor _inversed_7361_y_0_to_fp16 = const()[name = tensor("_inversed_7361_y_0_to_fp16"), val = tensor(0x1p-9)]; + tensor _inversed_7361_cast_fp16 = mul(x = ts_113_cast_fp16, y = _inversed_7361_y_0_to_fp16)[name = tensor("_inversed_7361_cast_fp16")]; + tensor var_7362_cast_fp16 = floor(x = _inversed_7361_cast_fp16)[name = tensor("op_7362_cast_fp16")]; + tensor var_7363_to_fp16 = const()[name = tensor("op_7363_to_fp16"), val = tensor(0x1p+9)]; + tensor var_7364_cast_fp16 = mul(x = var_7362_cast_fp16, y = var_7363_to_fp16)[name = tensor("op_7364_cast_fp16")]; + tensor write_indices_float_75_cast_fp16 = sub(x = ts_113_cast_fp16, y = var_7364_cast_fp16)[name = tensor("write_indices_float_75_cast_fp16")]; + tensor var_7371_dtype_0 = const()[name = tensor("op_7371_dtype_0"), val = tensor("int32")]; + tensor write_indices_37_reps_0 = const()[name = tensor("write_indices_37_reps_0"), val = tensor([1, 1, 16, 64])]; + tensor write_indices_float_75_cast_fp16_to_int32 = cast(dtype = var_7371_dtype_0, x = write_indices_float_75_cast_fp16)[name = tensor("cast_462")]; + tensor write_indices_37 = tile(reps = write_indices_37_reps_0, x = write_indices_float_75_cast_fp16_to_int32)[name = tensor("write_indices_37")]; + tensor var_7381_begin_0 = const()[name = tensor("op_7381_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_7381_end_0 = const()[name = tensor("op_7381_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor var_7381_end_mask_0 = const()[name = tensor("op_7381_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_7381_squeeze_mask_0 = const()[name = tensor("op_7381_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor cache18_to_fp16_dtype_0 = const()[name = tensor("cache18_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor cache18_to_fp16 = cast(dtype = cache18_to_fp16_dtype_0, x = cache18)[name = tensor("cast_461")]; + tensor var_7381_cast_fp16 = slice_by_index(begin = var_7381_begin_0, end = var_7381_end_0, end_mask = var_7381_end_mask_0, squeeze_mask = var_7381_squeeze_mask_0, x = cache18_to_fp16)[name = tensor("op_7381_cast_fp16")]; + tensor var_7383_axis_0 = const()[name = tensor("op_7383_axis_0"), val = tensor(1)]; + tensor var_7383_mode_0 = const()[name = tensor("op_7383_mode_0"), val = tensor("update")]; + tensor var_7383_validate_indices_0 = const()[name = tensor("op_7383_validate_indices_0"), val = tensor(false)]; + tensor var_7383_cast_fp16 = scatter_along_axis(axis = var_7383_axis_0, data = var_7381_cast_fp16, indices = write_indices_37, mode = var_7383_mode_0, updates = k_75_cast_fp16, validate_indices = var_7383_validate_indices_0)[name = tensor("op_7383_cast_fp16")]; + tensor concat_128 = const()[name = tensor("concat_128"), val = tensor([0, 0, 0, 0, 0])]; + tensor concat_129 = const()[name = tensor("concat_129"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_37_internal_tensor_assign_1_stride_0 = const()[name = tensor("new_cache_37_internal_tensor_assign_1_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_37_internal_tensor_assign_1_begin_mask_0 = const()[name = tensor("new_cache_37_internal_tensor_assign_1_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_37_internal_tensor_assign_1_end_mask_0 = const()[name = tensor("new_cache_37_internal_tensor_assign_1_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_37_internal_tensor_assign_1_squeeze_mask_0 = const()[name = tensor("new_cache_37_internal_tensor_assign_1_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_84 = const()[name = tensor("shape_84"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_36 = const()[name = tensor("reduce_prod_36"), val = tensor(1048576)]; + tensor range_1d_36_start_0 = const()[name = tensor("range_1d_36_start_0"), val = tensor(0)]; + tensor range_1d_36_step_0 = const()[name = tensor("range_1d_36_step_0"), val = tensor(1)]; + tensor range_1d_36 = range_1d(end = reduce_prod_36, start = range_1d_36_start_0, step = range_1d_36_step_0)[name = tensor("range_1d_36")]; + tensor reshape_180 = reshape(shape = shape_84, x = range_1d_36)[name = tensor("reshape_180")]; + tensor slice_by_index_36 = slice_by_index(begin = concat_128, begin_mask = new_cache_37_internal_tensor_assign_1_begin_mask_0, end = concat_129, end_mask = new_cache_37_internal_tensor_assign_1_end_mask_0, squeeze_mask = new_cache_37_internal_tensor_assign_1_squeeze_mask_0, stride = new_cache_37_internal_tensor_assign_1_stride_0, x = reshape_180)[name = tensor("slice_by_index_36")]; + tensor reshape_181_shape_0 = const()[name = tensor("reshape_181_shape_0"), val = tensor([-1])]; + tensor reshape_181 = reshape(shape = reshape_181_shape_0, x = slice_by_index_36)[name = tensor("reshape_181")]; + tensor reshape_182_shape_0 = const()[name = tensor("reshape_182_shape_0"), val = tensor([-1])]; + tensor reshape_182_cast_fp16 = reshape(shape = reshape_182_shape_0, x = var_7383_cast_fp16)[name = tensor("reshape_182_cast_fp16")]; + tensor reshape_183_shape_0 = const()[name = tensor("reshape_183_shape_0"), val = tensor([-1])]; + tensor reshape_183_cast_fp16 = reshape(shape = reshape_183_shape_0, x = cache18_to_fp16)[name = tensor("reshape_183_cast_fp16")]; + tensor scatter_36_mode_0 = const()[name = tensor("scatter_36_mode_0"), val = tensor("update")]; + tensor scatter_36_axis_0 = const()[name = tensor("scatter_36_axis_0"), val = tensor(0)]; + tensor scatter_36_validate_indices_0 = const()[name = tensor("scatter_36_validate_indices_0"), val = tensor(false)]; + tensor scatter_36_cast_fp16 = scatter(axis = scatter_36_axis_0, data = reshape_183_cast_fp16, indices = reshape_181, mode = scatter_36_mode_0, updates = reshape_182_cast_fp16, validate_indices = scatter_36_validate_indices_0)[name = tensor("scatter_36_cast_fp16")]; + tensor reshape_184_cast_fp16 = reshape(shape = shape_84, x = scatter_36_cast_fp16)[name = tensor("reshape_184_cast_fp16")]; + tensor var_7391_begin_0 = const()[name = tensor("op_7391_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor var_7391_end_0 = const()[name = tensor("op_7391_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor var_7391_end_mask_0 = const()[name = tensor("op_7391_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_7391_squeeze_mask_0 = const()[name = tensor("op_7391_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor var_7391_cast_fp16 = slice_by_index(begin = var_7391_begin_0, end = var_7391_end_0, end_mask = var_7391_end_mask_0, squeeze_mask = var_7391_squeeze_mask_0, x = reshape_184_cast_fp16)[name = tensor("op_7391_cast_fp16")]; + tensor var_7393_axis_0 = const()[name = tensor("op_7393_axis_0"), val = tensor(1)]; + tensor var_7393_mode_0 = const()[name = tensor("op_7393_mode_0"), val = tensor("update")]; + tensor var_7393_validate_indices_0 = const()[name = tensor("op_7393_validate_indices_0"), val = tensor(false)]; + tensor var_7393_cast_fp16 = scatter_along_axis(axis = var_7393_axis_0, data = var_7391_cast_fp16, indices = write_indices_37, mode = var_7393_mode_0, updates = v_37_cast_fp16, validate_indices = var_7393_validate_indices_0)[name = tensor("op_7393_cast_fp16")]; + tensor concat_130 = const()[name = tensor("concat_130"), val = tensor([1, 0, 0, 0, 0])]; + tensor concat_131 = const()[name = tensor("concat_131"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_37_internal_tensor_assign_2_stride_0 = const()[name = tensor("new_cache_37_internal_tensor_assign_2_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_37_internal_tensor_assign_2_begin_mask_0 = const()[name = tensor("new_cache_37_internal_tensor_assign_2_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_37_internal_tensor_assign_2_end_mask_0 = const()[name = tensor("new_cache_37_internal_tensor_assign_2_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_37_internal_tensor_assign_2_squeeze_mask_0 = const()[name = tensor("new_cache_37_internal_tensor_assign_2_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_85 = const()[name = tensor("shape_85"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_37 = const()[name = tensor("reduce_prod_37"), val = tensor(1048576)]; + tensor range_1d_37_start_0 = const()[name = tensor("range_1d_37_start_0"), val = tensor(0)]; + tensor range_1d_37_step_0 = const()[name = tensor("range_1d_37_step_0"), val = tensor(1)]; + tensor range_1d_37 = range_1d(end = reduce_prod_37, start = range_1d_37_start_0, step = range_1d_37_step_0)[name = tensor("range_1d_37")]; + tensor reshape_185 = reshape(shape = shape_85, x = range_1d_37)[name = tensor("reshape_185")]; + tensor slice_by_index_37 = slice_by_index(begin = concat_130, begin_mask = new_cache_37_internal_tensor_assign_2_begin_mask_0, end = concat_131, end_mask = new_cache_37_internal_tensor_assign_2_end_mask_0, squeeze_mask = new_cache_37_internal_tensor_assign_2_squeeze_mask_0, stride = new_cache_37_internal_tensor_assign_2_stride_0, x = reshape_185)[name = tensor("slice_by_index_37")]; + tensor reshape_186_shape_0 = const()[name = tensor("reshape_186_shape_0"), val = tensor([-1])]; + tensor reshape_186 = reshape(shape = reshape_186_shape_0, x = slice_by_index_37)[name = tensor("reshape_186")]; + tensor reshape_187_shape_0 = const()[name = tensor("reshape_187_shape_0"), val = tensor([-1])]; + tensor reshape_187_cast_fp16 = reshape(shape = reshape_187_shape_0, x = var_7393_cast_fp16)[name = tensor("reshape_187_cast_fp16")]; + tensor reshape_188_shape_0 = const()[name = tensor("reshape_188_shape_0"), val = tensor([-1])]; + tensor reshape_188_cast_fp16 = reshape(shape = reshape_188_shape_0, x = reshape_184_cast_fp16)[name = tensor("reshape_188_cast_fp16")]; + tensor scatter_37_mode_0 = const()[name = tensor("scatter_37_mode_0"), val = tensor("update")]; + tensor scatter_37_axis_0 = const()[name = tensor("scatter_37_axis_0"), val = tensor(0)]; + tensor scatter_37_validate_indices_0 = const()[name = tensor("scatter_37_validate_indices_0"), val = tensor(false)]; + tensor scatter_37_cast_fp16 = scatter(axis = scatter_37_axis_0, data = reshape_188_cast_fp16, indices = reshape_186, mode = scatter_37_mode_0, updates = reshape_187_cast_fp16, validate_indices = scatter_37_validate_indices_0)[name = tensor("scatter_37_cast_fp16")]; + tensor reshape_189_cast_fp16 = reshape(shape = shape_85, x = scatter_37_cast_fp16)[name = tensor("reshape_189_cast_fp16")]; + tensor reshape_189_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("reshape_189_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor keys_109_begin_0 = const()[name = tensor("keys_109_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor keys_109_end_0 = const()[name = tensor("keys_109_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor keys_109_end_mask_0 = const()[name = tensor("keys_109_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor keys_109_squeeze_mask_0 = const()[name = tensor("keys_109_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor keys_109_cast_fp16 = slice_by_index(begin = keys_109_begin_0, end = keys_109_end_0, end_mask = keys_109_end_mask_0, squeeze_mask = keys_109_squeeze_mask_0, x = reshape_189_cast_fp16)[name = tensor("keys_109_cast_fp16")]; + tensor values_109_begin_0 = const()[name = tensor("values_109_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor values_109_end_0 = const()[name = tensor("values_109_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor values_109_end_mask_0 = const()[name = tensor("values_109_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor values_109_squeeze_mask_0 = const()[name = tensor("values_109_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor values_109_cast_fp16 = slice_by_index(begin = values_109_begin_0, end = values_109_end_0, end_mask = values_109_end_mask_0, squeeze_mask = values_109_squeeze_mask_0, x = reshape_189_cast_fp16)[name = tensor("values_109_cast_fp16")]; + tensor var_7405_cast_fp16 = not_equal(x = keys_109_cast_fp16, y = keys_109_cast_fp16)[name = tensor("op_7405_cast_fp16")]; + tensor keys_111_cast_fp16 = select(a = var_463_to_fp16, b = keys_109_cast_fp16, cond = var_7405_cast_fp16)[name = tensor("keys_111_cast_fp16")]; + tensor var_7413_cast_fp16 = not_equal(x = values_109_cast_fp16, y = values_109_cast_fp16)[name = tensor("op_7413_cast_fp16")]; + tensor values_111_cast_fp16 = select(a = var_463_to_fp16, b = values_109_cast_fp16, cond = var_7413_cast_fp16)[name = tensor("values_111_cast_fp16")]; + tensor var_7437 = const()[name = tensor("op_7437"), val = tensor([0, 2, 1, 3])]; + tensor var_7453 = const()[name = tensor("op_7453"), val = tensor([1, 1, 1])]; + tensor var_7454_cast_fp16 = reshape(shape = var_7453, x = position18_to_fp16)[name = tensor("op_7454_cast_fp16")]; + tensor var_7474_to_fp16 = const()[name = tensor("op_7474_to_fp16"), val = tensor(0x1p+0)]; + tensor valid_len_37_cast_fp16 = add(x = var_7454_cast_fp16, y = var_7474_to_fp16)[name = tensor("valid_len_37_cast_fp16")]; + tensor valid_mask_37_cast_fp16 = less(x = k_positions_1_promoted_to_fp16, y = valid_len_37_cast_fp16)[name = tensor("valid_mask_37_cast_fp16")]; + tensor causal_mask_37_cast_fp16 = less_equal(x = k_positions_1_promoted_to_fp16, y = var_7454_cast_fp16)[name = tensor("causal_mask_37_cast_fp16")]; + tensor attn_mask_73 = logical_and(x = valid_mask_37_cast_fp16, y = causal_mask_37_cast_fp16)[name = tensor("attn_mask_73")]; + tensor attn_mask_75_axes_0 = const()[name = tensor("attn_mask_75_axes_0"), val = tensor([1])]; + tensor attn_mask_75 = expand_dims(axes = attn_mask_75_axes_0, x = attn_mask_73)[name = tensor("attn_mask_75")]; + tensor var_7492_transpose_x_0 = const()[name = tensor("op_7492_transpose_x_0"), val = tensor(false)]; + tensor var_7492_transpose_y_0 = const()[name = tensor("op_7492_transpose_y_0"), val = tensor(false)]; + tensor transpose_108_perm_0 = const()[name = tensor("transpose_108_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_109_perm_0 = const()[name = tensor("transpose_109_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_109 = transpose(perm = transpose_109_perm_0, x = keys_111_cast_fp16)[name = tensor("transpose_141")]; + tensor transpose_108 = transpose(perm = transpose_108_perm_0, x = q_111_cast_fp16)[name = tensor("transpose_142")]; + tensor var_7492_cast_fp16 = matmul(transpose_x = var_7492_transpose_x_0, transpose_y = var_7492_transpose_y_0, x = transpose_108, y = transpose_109)[name = tensor("op_7492_cast_fp16")]; + tensor var_7486_to_fp16 = const()[name = tensor("op_7486_to_fp16"), val = tensor([0x1p-3])]; + tensor attn_weights_109_cast_fp16 = mul(x = var_7492_cast_fp16, y = var_7486_to_fp16)[name = tensor("attn_weights_109_cast_fp16")]; + tensor var_7494 = logical_not(x = attn_mask_75)[name = tensor("op_7494")]; + tensor var_7495_to_fp16 = const()[name = tensor("op_7495_to_fp16"), val = tensor(-inf)]; + tensor attn_weights_111_cast_fp16 = select(a = var_7495_to_fp16, b = attn_weights_109_cast_fp16, cond = var_7494)[name = tensor("attn_weights_111_cast_fp16")]; + tensor var_7497 = const()[name = tensor("op_7497"), val = tensor(-1)]; + tensor attn_weights_113_cast_fp16 = softmax(axis = var_7497, x = attn_weights_111_cast_fp16)[name = tensor("attn_weights_113_cast_fp16")]; + tensor attn_output_37_transpose_x_0 = const()[name = tensor("attn_output_37_transpose_x_0"), val = tensor(false)]; + tensor attn_output_37_transpose_y_0 = const()[name = tensor("attn_output_37_transpose_y_0"), val = tensor(false)]; + tensor values_113_cast_fp16 = transpose(perm = var_7437, x = values_111_cast_fp16)[name = tensor("transpose_143")]; + tensor attn_output_37_cast_fp16 = matmul(transpose_x = attn_output_37_transpose_x_0, transpose_y = attn_output_37_transpose_y_0, x = attn_weights_113_cast_fp16, y = values_113_cast_fp16)[name = tensor("attn_output_37_cast_fp16")]; + tensor var_7505 = const()[name = tensor("op_7505"), val = tensor([0, 2, 1, 3])]; + tensor var_7510 = const()[name = tensor("op_7510"), val = tensor([1, 1, 1024])]; + tensor var_7506_cast_fp16 = transpose(perm = var_7505, x = attn_output_37_cast_fp16)[name = tensor("transpose_140")]; + tensor input_223_cast_fp16 = reshape(shape = var_7510, x = var_7506_cast_fp16)[name = tensor("input_223_cast_fp16")]; + tensor attn18_out_proj_weight_to_fp16 = const()[name = tensor("attn18_out_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(460571648)))]; + tensor linear_74_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = attn18_out_proj_weight_to_fp16, x = input_223_cast_fp16)[name = tensor("linear_74_cast_fp16")]; + tensor var_7516_to_fp16 = const()[name = tensor("op_7516_to_fp16"), val = tensor(0x1p+0)]; + tensor var_7517_cast_fp16 = add(x = position18_to_fp16, y = var_7516_to_fp16)[name = tensor("op_7517_cast_fp16")]; + tensor var_7517_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_7517_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor input_225_cast_fp16 = add(x = input_219_cast_fp16, y = linear_74_cast_fp16)[name = tensor("input_225_cast_fp16")]; + tensor input_227_axes_0 = const()[name = tensor("input_227_axes_0"), val = tensor([-1])]; + tensor norm18_2_weight_to_fp16 = const()[name = tensor("norm18_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(462668864)))]; + tensor norm18_2_bias_to_fp16 = const()[name = tensor("norm18_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(462670976)))]; + tensor var_7521_to_fp16 = const()[name = tensor("op_7521_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_227_cast_fp16 = layer_norm(axes = input_227_axes_0, beta = norm18_2_bias_to_fp16, epsilon = var_7521_to_fp16, gamma = norm18_2_weight_to_fp16, x = input_225_cast_fp16)[name = tensor("input_227_cast_fp16")]; + tensor linear18_1_weight_to_fp16 = const()[name = tensor("linear18_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(462673088)))]; + tensor linear_75_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = linear18_1_weight_to_fp16, x = input_227_cast_fp16)[name = tensor("linear_75_cast_fp16")]; + tensor input_229_mode_0 = const()[name = tensor("input_229_mode_0"), val = tensor("EXACT")]; + tensor input_229_cast_fp16 = gelu(mode = input_229_mode_0, x = linear_75_cast_fp16)[name = tensor("input_229_cast_fp16")]; + tensor linear18_2_weight_to_fp16 = const()[name = tensor("linear18_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(471061760)))]; + tensor linear_76_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = linear18_2_weight_to_fp16, x = input_229_cast_fp16)[name = tensor("linear_76_cast_fp16")]; + tensor input_231_cast_fp16 = add(x = input_225_cast_fp16, y = linear_76_cast_fp16)[name = tensor("input_231_cast_fp16")]; + tensor input_233_axes_0 = const()[name = tensor("input_233_axes_0"), val = tensor([-1])]; + tensor norm19_1_weight_to_fp16 = const()[name = tensor("norm19_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(479450432)))]; + tensor norm19_1_bias_to_fp16 = const()[name = tensor("norm19_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(479452544)))]; + tensor var_7538_to_fp16 = const()[name = tensor("op_7538_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_233_cast_fp16 = layer_norm(axes = input_233_axes_0, beta = norm19_1_bias_to_fp16, epsilon = var_7538_to_fp16, gamma = norm19_1_weight_to_fp16, x = input_231_cast_fp16)[name = tensor("input_233_cast_fp16")]; + tensor attn19_in_proj_weight_to_fp16 = const()[name = tensor("attn19_in_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(479454656)))]; + tensor linear_77_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = attn19_in_proj_weight_to_fp16, x = input_233_cast_fp16)[name = tensor("linear_77_cast_fp16")]; + tensor var_7552 = const()[name = tensor("op_7552"), val = tensor([1, 1, 3, 16, 64])]; + tensor qkv_39_cast_fp16 = reshape(shape = var_7552, x = linear_77_cast_fp16)[name = tensor("qkv_39_cast_fp16")]; + tensor q_115_begin_0 = const()[name = tensor("q_115_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor q_115_end_0 = const()[name = tensor("q_115_end_0"), val = tensor([1, 1, 1, 16, 64])]; + tensor q_115_end_mask_0 = const()[name = tensor("q_115_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor q_115_squeeze_mask_0 = const()[name = tensor("q_115_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor q_115_cast_fp16 = slice_by_index(begin = q_115_begin_0, end = q_115_end_0, end_mask = q_115_end_mask_0, squeeze_mask = q_115_squeeze_mask_0, x = qkv_39_cast_fp16)[name = tensor("q_115_cast_fp16")]; + tensor k_77_begin_0 = const()[name = tensor("k_77_begin_0"), val = tensor([0, 0, 1, 0, 0])]; + tensor k_77_end_0 = const()[name = tensor("k_77_end_0"), val = tensor([1, 1, 2, 16, 64])]; + tensor k_77_end_mask_0 = const()[name = tensor("k_77_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor k_77_squeeze_mask_0 = const()[name = tensor("k_77_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor k_77_cast_fp16 = slice_by_index(begin = k_77_begin_0, end = k_77_end_0, end_mask = k_77_end_mask_0, squeeze_mask = k_77_squeeze_mask_0, x = qkv_39_cast_fp16)[name = tensor("k_77_cast_fp16")]; + tensor v_39_begin_0 = const()[name = tensor("v_39_begin_0"), val = tensor([0, 0, 2, 0, 0])]; + tensor v_39_end_0 = const()[name = tensor("v_39_end_0"), val = tensor([1, 1, 3, 16, 64])]; + tensor v_39_end_mask_0 = const()[name = tensor("v_39_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor v_39_squeeze_mask_0 = const()[name = tensor("v_39_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor v_39_cast_fp16 = slice_by_index(begin = v_39_begin_0, end = v_39_end_0, end_mask = v_39_end_mask_0, squeeze_mask = v_39_squeeze_mask_0, x = qkv_39_cast_fp16)[name = tensor("v_39_cast_fp16")]; + tensor var_7618 = const()[name = tensor("op_7618"), val = tensor([1, 1, 1, 1])]; + tensor position19_to_fp16_dtype_0 = const()[name = tensor("position19_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor position19_to_fp16 = cast(dtype = position19_to_fp16_dtype_0, x = position19)[name = tensor("cast_458")]; + tensor ts_119_cast_fp16 = reshape(shape = var_7618, x = position19_to_fp16)[name = tensor("ts_119_cast_fp16")]; + tensor var_7625 = const()[name = tensor("op_7625"), val = tensor([1, 1, 16, 32, 2])]; + tensor q_complex_39_cast_fp16 = reshape(shape = var_7625, x = q_115_cast_fp16)[name = tensor("q_complex_39_cast_fp16")]; + tensor var_7632 = const()[name = tensor("op_7632"), val = tensor([1, 1, 16, 32, 2])]; + tensor k_complex_39_cast_fp16 = reshape(shape = var_7632, x = k_77_cast_fp16)[name = tensor("k_complex_39_cast_fp16")]; + tensor var_7636_begin_0 = const()[name = tensor("op_7636_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_7636_end_0 = const()[name = tensor("op_7636_end_0"), val = tensor([1, 1, 16, 32, 1])]; + tensor var_7636_end_mask_0 = const()[name = tensor("op_7636_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_7636_squeeze_mask_0 = const()[name = tensor("op_7636_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_7636_cast_fp16 = slice_by_index(begin = var_7636_begin_0, end = var_7636_end_0, end_mask = var_7636_end_mask_0, squeeze_mask = var_7636_squeeze_mask_0, x = q_complex_39_cast_fp16)[name = tensor("op_7636_cast_fp16")]; + tensor var_7644_begin_0 = const()[name = tensor("op_7644_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_7644_end_0 = const()[name = tensor("op_7644_end_0"), val = tensor([1, 1, 16, 32, 2])]; + tensor var_7644_end_mask_0 = const()[name = tensor("op_7644_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_7644_squeeze_mask_0 = const()[name = tensor("op_7644_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_7644_cast_fp16 = slice_by_index(begin = var_7644_begin_0, end = var_7644_end_0, end_mask = var_7644_end_mask_0, squeeze_mask = var_7644_squeeze_mask_0, x = q_complex_39_cast_fp16)[name = tensor("op_7644_cast_fp16")]; + tensor var_7652_begin_0 = const()[name = tensor("op_7652_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_7652_end_0 = const()[name = tensor("op_7652_end_0"), val = tensor([1, 1, 16, 32, 1])]; + tensor var_7652_end_mask_0 = const()[name = tensor("op_7652_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_7652_squeeze_mask_0 = const()[name = tensor("op_7652_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_7652_cast_fp16 = slice_by_index(begin = var_7652_begin_0, end = var_7652_end_0, end_mask = var_7652_end_mask_0, squeeze_mask = var_7652_squeeze_mask_0, x = k_complex_39_cast_fp16)[name = tensor("op_7652_cast_fp16")]; + tensor var_7660_begin_0 = const()[name = tensor("op_7660_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_7660_end_0 = const()[name = tensor("op_7660_end_0"), val = tensor([1, 1, 16, 32, 2])]; + tensor var_7660_end_mask_0 = const()[name = tensor("op_7660_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_7660_squeeze_mask_0 = const()[name = tensor("op_7660_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_7660_cast_fp16 = slice_by_index(begin = var_7660_begin_0, end = var_7660_end_0, end_mask = var_7660_end_mask_0, squeeze_mask = var_7660_squeeze_mask_0, x = k_complex_39_cast_fp16)[name = tensor("op_7660_cast_fp16")]; + tensor freqs_39_to_fp16 = const()[name = tensor("freqs_39_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(485746176)))]; + tensor var_7666_cast_fp16 = mul(x = freqs_39_to_fp16, y = ts_119_cast_fp16)[name = tensor("op_7666_cast_fp16")]; + tensor rotr_39_cast_fp16 = cos(x = var_7666_cast_fp16)[name = tensor("rotr_39_cast_fp16")]; + tensor roti_39_cast_fp16 = sin(x = var_7666_cast_fp16)[name = tensor("roti_39_cast_fp16")]; + tensor var_7670_cast_fp16 = mul(x = var_7636_cast_fp16, y = rotr_39_cast_fp16)[name = tensor("op_7670_cast_fp16")]; + tensor var_7671_cast_fp16 = mul(x = var_7644_cast_fp16, y = roti_39_cast_fp16)[name = tensor("op_7671_cast_fp16")]; + tensor qor_77_cast_fp16 = sub(x = var_7670_cast_fp16, y = var_7671_cast_fp16)[name = tensor("qor_77_cast_fp16")]; + tensor var_7674_cast_fp16 = mul(x = var_7636_cast_fp16, y = roti_39_cast_fp16)[name = tensor("op_7674_cast_fp16")]; + tensor var_7675_cast_fp16 = mul(x = var_7644_cast_fp16, y = rotr_39_cast_fp16)[name = tensor("op_7675_cast_fp16")]; + tensor qoi_77_cast_fp16 = add(x = var_7674_cast_fp16, y = var_7675_cast_fp16)[name = tensor("qoi_77_cast_fp16")]; + tensor var_7678_cast_fp16 = mul(x = var_7652_cast_fp16, y = rotr_39_cast_fp16)[name = tensor("op_7678_cast_fp16")]; + tensor var_7679_cast_fp16 = mul(x = var_7660_cast_fp16, y = roti_39_cast_fp16)[name = tensor("op_7679_cast_fp16")]; + tensor kor_77_cast_fp16 = sub(x = var_7678_cast_fp16, y = var_7679_cast_fp16)[name = tensor("kor_77_cast_fp16")]; + tensor var_7682_cast_fp16 = mul(x = var_7652_cast_fp16, y = roti_39_cast_fp16)[name = tensor("op_7682_cast_fp16")]; + tensor var_7683_cast_fp16 = mul(x = var_7660_cast_fp16, y = rotr_39_cast_fp16)[name = tensor("op_7683_cast_fp16")]; + tensor koi_77_cast_fp16 = add(x = var_7682_cast_fp16, y = var_7683_cast_fp16)[name = tensor("koi_77_cast_fp16")]; + tensor qo_39_axis_0 = const()[name = tensor("qo_39_axis_0"), val = tensor(-1)]; + tensor qo_39_cast_fp16 = stack(axis = qo_39_axis_0, values = (qor_77_cast_fp16, qoi_77_cast_fp16))[name = tensor("qo_39_cast_fp16")]; + tensor ko_39_axis_0 = const()[name = tensor("ko_39_axis_0"), val = tensor(-1)]; + tensor ko_39_cast_fp16 = stack(axis = ko_39_axis_0, values = (kor_77_cast_fp16, koi_77_cast_fp16))[name = tensor("ko_39_cast_fp16")]; + tensor var_7716 = const()[name = tensor("op_7716"), val = tensor([1, 1, 16, 64])]; + tensor q_117_cast_fp16 = reshape(shape = var_7716, x = qo_39_cast_fp16)[name = tensor("q_117_cast_fp16")]; + tensor var_7722 = const()[name = tensor("op_7722"), val = tensor([1, 1, 16, 64])]; + tensor k_79_cast_fp16 = reshape(shape = var_7722, x = ko_39_cast_fp16)[name = tensor("k_79_cast_fp16")]; + tensor _inversed_7747_y_0_to_fp16 = const()[name = tensor("_inversed_7747_y_0_to_fp16"), val = tensor(0x1p-9)]; + tensor _inversed_7747_cast_fp16 = mul(x = ts_119_cast_fp16, y = _inversed_7747_y_0_to_fp16)[name = tensor("_inversed_7747_cast_fp16")]; + tensor var_7748_cast_fp16 = floor(x = _inversed_7747_cast_fp16)[name = tensor("op_7748_cast_fp16")]; + tensor var_7749_to_fp16 = const()[name = tensor("op_7749_to_fp16"), val = tensor(0x1p+9)]; + tensor var_7750_cast_fp16 = mul(x = var_7748_cast_fp16, y = var_7749_to_fp16)[name = tensor("op_7750_cast_fp16")]; + tensor write_indices_float_79_cast_fp16 = sub(x = ts_119_cast_fp16, y = var_7750_cast_fp16)[name = tensor("write_indices_float_79_cast_fp16")]; + tensor var_7757_dtype_0 = const()[name = tensor("op_7757_dtype_0"), val = tensor("int32")]; + tensor write_indices_39_reps_0 = const()[name = tensor("write_indices_39_reps_0"), val = tensor([1, 1, 16, 64])]; + tensor write_indices_float_79_cast_fp16_to_int32 = cast(dtype = var_7757_dtype_0, x = write_indices_float_79_cast_fp16)[name = tensor("cast_457")]; + tensor write_indices_39 = tile(reps = write_indices_39_reps_0, x = write_indices_float_79_cast_fp16_to_int32)[name = tensor("write_indices_39")]; + tensor var_7767_begin_0 = const()[name = tensor("op_7767_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_7767_end_0 = const()[name = tensor("op_7767_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor var_7767_end_mask_0 = const()[name = tensor("op_7767_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_7767_squeeze_mask_0 = const()[name = tensor("op_7767_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor cache19_to_fp16_dtype_0 = const()[name = tensor("cache19_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor cache19_to_fp16 = cast(dtype = cache19_to_fp16_dtype_0, x = cache19)[name = tensor("cast_456")]; + tensor var_7767_cast_fp16 = slice_by_index(begin = var_7767_begin_0, end = var_7767_end_0, end_mask = var_7767_end_mask_0, squeeze_mask = var_7767_squeeze_mask_0, x = cache19_to_fp16)[name = tensor("op_7767_cast_fp16")]; + tensor var_7769_axis_0 = const()[name = tensor("op_7769_axis_0"), val = tensor(1)]; + tensor var_7769_mode_0 = const()[name = tensor("op_7769_mode_0"), val = tensor("update")]; + tensor var_7769_validate_indices_0 = const()[name = tensor("op_7769_validate_indices_0"), val = tensor(false)]; + tensor var_7769_cast_fp16 = scatter_along_axis(axis = var_7769_axis_0, data = var_7767_cast_fp16, indices = write_indices_39, mode = var_7769_mode_0, updates = k_79_cast_fp16, validate_indices = var_7769_validate_indices_0)[name = tensor("op_7769_cast_fp16")]; + tensor concat_135 = const()[name = tensor("concat_135"), val = tensor([0, 0, 0, 0, 0])]; + tensor concat_136 = const()[name = tensor("concat_136"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_39_internal_tensor_assign_1_stride_0 = const()[name = tensor("new_cache_39_internal_tensor_assign_1_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_39_internal_tensor_assign_1_begin_mask_0 = const()[name = tensor("new_cache_39_internal_tensor_assign_1_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_39_internal_tensor_assign_1_end_mask_0 = const()[name = tensor("new_cache_39_internal_tensor_assign_1_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_39_internal_tensor_assign_1_squeeze_mask_0 = const()[name = tensor("new_cache_39_internal_tensor_assign_1_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_86 = const()[name = tensor("shape_86"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_38 = const()[name = tensor("reduce_prod_38"), val = tensor(1048576)]; + tensor range_1d_38_start_0 = const()[name = tensor("range_1d_38_start_0"), val = tensor(0)]; + tensor range_1d_38_step_0 = const()[name = tensor("range_1d_38_step_0"), val = tensor(1)]; + tensor range_1d_38 = range_1d(end = reduce_prod_38, start = range_1d_38_start_0, step = range_1d_38_step_0)[name = tensor("range_1d_38")]; + tensor reshape_190 = reshape(shape = shape_86, x = range_1d_38)[name = tensor("reshape_190")]; + tensor slice_by_index_38 = slice_by_index(begin = concat_135, begin_mask = new_cache_39_internal_tensor_assign_1_begin_mask_0, end = concat_136, end_mask = new_cache_39_internal_tensor_assign_1_end_mask_0, squeeze_mask = new_cache_39_internal_tensor_assign_1_squeeze_mask_0, stride = new_cache_39_internal_tensor_assign_1_stride_0, x = reshape_190)[name = tensor("slice_by_index_38")]; + tensor reshape_191_shape_0 = const()[name = tensor("reshape_191_shape_0"), val = tensor([-1])]; + tensor reshape_191 = reshape(shape = reshape_191_shape_0, x = slice_by_index_38)[name = tensor("reshape_191")]; + tensor reshape_192_shape_0 = const()[name = tensor("reshape_192_shape_0"), val = tensor([-1])]; + tensor reshape_192_cast_fp16 = reshape(shape = reshape_192_shape_0, x = var_7769_cast_fp16)[name = tensor("reshape_192_cast_fp16")]; + tensor reshape_193_shape_0 = const()[name = tensor("reshape_193_shape_0"), val = tensor([-1])]; + tensor reshape_193_cast_fp16 = reshape(shape = reshape_193_shape_0, x = cache19_to_fp16)[name = tensor("reshape_193_cast_fp16")]; + tensor scatter_38_mode_0 = const()[name = tensor("scatter_38_mode_0"), val = tensor("update")]; + tensor scatter_38_axis_0 = const()[name = tensor("scatter_38_axis_0"), val = tensor(0)]; + tensor scatter_38_validate_indices_0 = const()[name = tensor("scatter_38_validate_indices_0"), val = tensor(false)]; + tensor scatter_38_cast_fp16 = scatter(axis = scatter_38_axis_0, data = reshape_193_cast_fp16, indices = reshape_191, mode = scatter_38_mode_0, updates = reshape_192_cast_fp16, validate_indices = scatter_38_validate_indices_0)[name = tensor("scatter_38_cast_fp16")]; + tensor reshape_194_cast_fp16 = reshape(shape = shape_86, x = scatter_38_cast_fp16)[name = tensor("reshape_194_cast_fp16")]; + tensor var_7777_begin_0 = const()[name = tensor("op_7777_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor var_7777_end_0 = const()[name = tensor("op_7777_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor var_7777_end_mask_0 = const()[name = tensor("op_7777_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_7777_squeeze_mask_0 = const()[name = tensor("op_7777_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor var_7777_cast_fp16 = slice_by_index(begin = var_7777_begin_0, end = var_7777_end_0, end_mask = var_7777_end_mask_0, squeeze_mask = var_7777_squeeze_mask_0, x = reshape_194_cast_fp16)[name = tensor("op_7777_cast_fp16")]; + tensor var_7779_axis_0 = const()[name = tensor("op_7779_axis_0"), val = tensor(1)]; + tensor var_7779_mode_0 = const()[name = tensor("op_7779_mode_0"), val = tensor("update")]; + tensor var_7779_validate_indices_0 = const()[name = tensor("op_7779_validate_indices_0"), val = tensor(false)]; + tensor var_7779_cast_fp16 = scatter_along_axis(axis = var_7779_axis_0, data = var_7777_cast_fp16, indices = write_indices_39, mode = var_7779_mode_0, updates = v_39_cast_fp16, validate_indices = var_7779_validate_indices_0)[name = tensor("op_7779_cast_fp16")]; + tensor concat_137 = const()[name = tensor("concat_137"), val = tensor([1, 0, 0, 0, 0])]; + tensor concat_138 = const()[name = tensor("concat_138"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_39_internal_tensor_assign_2_stride_0 = const()[name = tensor("new_cache_39_internal_tensor_assign_2_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_39_internal_tensor_assign_2_begin_mask_0 = const()[name = tensor("new_cache_39_internal_tensor_assign_2_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_39_internal_tensor_assign_2_end_mask_0 = const()[name = tensor("new_cache_39_internal_tensor_assign_2_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_39_internal_tensor_assign_2_squeeze_mask_0 = const()[name = tensor("new_cache_39_internal_tensor_assign_2_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_87 = const()[name = tensor("shape_87"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_39 = const()[name = tensor("reduce_prod_39"), val = tensor(1048576)]; + tensor range_1d_39_start_0 = const()[name = tensor("range_1d_39_start_0"), val = tensor(0)]; + tensor range_1d_39_step_0 = const()[name = tensor("range_1d_39_step_0"), val = tensor(1)]; + tensor range_1d_39 = range_1d(end = reduce_prod_39, start = range_1d_39_start_0, step = range_1d_39_step_0)[name = tensor("range_1d_39")]; + tensor reshape_195 = reshape(shape = shape_87, x = range_1d_39)[name = tensor("reshape_195")]; + tensor slice_by_index_39 = slice_by_index(begin = concat_137, begin_mask = new_cache_39_internal_tensor_assign_2_begin_mask_0, end = concat_138, end_mask = new_cache_39_internal_tensor_assign_2_end_mask_0, squeeze_mask = new_cache_39_internal_tensor_assign_2_squeeze_mask_0, stride = new_cache_39_internal_tensor_assign_2_stride_0, x = reshape_195)[name = tensor("slice_by_index_39")]; + tensor reshape_196_shape_0 = const()[name = tensor("reshape_196_shape_0"), val = tensor([-1])]; + tensor reshape_196 = reshape(shape = reshape_196_shape_0, x = slice_by_index_39)[name = tensor("reshape_196")]; + tensor reshape_197_shape_0 = const()[name = tensor("reshape_197_shape_0"), val = tensor([-1])]; + tensor reshape_197_cast_fp16 = reshape(shape = reshape_197_shape_0, x = var_7779_cast_fp16)[name = tensor("reshape_197_cast_fp16")]; + tensor reshape_198_shape_0 = const()[name = tensor("reshape_198_shape_0"), val = tensor([-1])]; + tensor reshape_198_cast_fp16 = reshape(shape = reshape_198_shape_0, x = reshape_194_cast_fp16)[name = tensor("reshape_198_cast_fp16")]; + tensor scatter_39_mode_0 = const()[name = tensor("scatter_39_mode_0"), val = tensor("update")]; + tensor scatter_39_axis_0 = const()[name = tensor("scatter_39_axis_0"), val = tensor(0)]; + tensor scatter_39_validate_indices_0 = const()[name = tensor("scatter_39_validate_indices_0"), val = tensor(false)]; + tensor scatter_39_cast_fp16 = scatter(axis = scatter_39_axis_0, data = reshape_198_cast_fp16, indices = reshape_196, mode = scatter_39_mode_0, updates = reshape_197_cast_fp16, validate_indices = scatter_39_validate_indices_0)[name = tensor("scatter_39_cast_fp16")]; + tensor reshape_199_cast_fp16 = reshape(shape = shape_87, x = scatter_39_cast_fp16)[name = tensor("reshape_199_cast_fp16")]; + tensor reshape_199_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("reshape_199_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor keys_115_begin_0 = const()[name = tensor("keys_115_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor keys_115_end_0 = const()[name = tensor("keys_115_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor keys_115_end_mask_0 = const()[name = tensor("keys_115_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor keys_115_squeeze_mask_0 = const()[name = tensor("keys_115_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor keys_115_cast_fp16 = slice_by_index(begin = keys_115_begin_0, end = keys_115_end_0, end_mask = keys_115_end_mask_0, squeeze_mask = keys_115_squeeze_mask_0, x = reshape_199_cast_fp16)[name = tensor("keys_115_cast_fp16")]; + tensor values_115_begin_0 = const()[name = tensor("values_115_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor values_115_end_0 = const()[name = tensor("values_115_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor values_115_end_mask_0 = const()[name = tensor("values_115_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor values_115_squeeze_mask_0 = const()[name = tensor("values_115_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor values_115_cast_fp16 = slice_by_index(begin = values_115_begin_0, end = values_115_end_0, end_mask = values_115_end_mask_0, squeeze_mask = values_115_squeeze_mask_0, x = reshape_199_cast_fp16)[name = tensor("values_115_cast_fp16")]; + tensor var_7791_cast_fp16 = not_equal(x = keys_115_cast_fp16, y = keys_115_cast_fp16)[name = tensor("op_7791_cast_fp16")]; + tensor keys_117_cast_fp16 = select(a = var_463_to_fp16, b = keys_115_cast_fp16, cond = var_7791_cast_fp16)[name = tensor("keys_117_cast_fp16")]; + tensor var_7799_cast_fp16 = not_equal(x = values_115_cast_fp16, y = values_115_cast_fp16)[name = tensor("op_7799_cast_fp16")]; + tensor values_117_cast_fp16 = select(a = var_463_to_fp16, b = values_115_cast_fp16, cond = var_7799_cast_fp16)[name = tensor("values_117_cast_fp16")]; + tensor var_7823 = const()[name = tensor("op_7823"), val = tensor([0, 2, 1, 3])]; + tensor var_7839 = const()[name = tensor("op_7839"), val = tensor([1, 1, 1])]; + tensor var_7840_cast_fp16 = reshape(shape = var_7839, x = position19_to_fp16)[name = tensor("op_7840_cast_fp16")]; + tensor var_7860_to_fp16 = const()[name = tensor("op_7860_to_fp16"), val = tensor(0x1p+0)]; + tensor valid_len_39_cast_fp16 = add(x = var_7840_cast_fp16, y = var_7860_to_fp16)[name = tensor("valid_len_39_cast_fp16")]; + tensor valid_mask_39_cast_fp16 = less(x = k_positions_1_promoted_to_fp16, y = valid_len_39_cast_fp16)[name = tensor("valid_mask_39_cast_fp16")]; + tensor causal_mask_39_cast_fp16 = less_equal(x = k_positions_1_promoted_to_fp16, y = var_7840_cast_fp16)[name = tensor("causal_mask_39_cast_fp16")]; + tensor attn_mask_77 = logical_and(x = valid_mask_39_cast_fp16, y = causal_mask_39_cast_fp16)[name = tensor("attn_mask_77")]; + tensor attn_mask_79_axes_0 = const()[name = tensor("attn_mask_79_axes_0"), val = tensor([1])]; + tensor attn_mask_79 = expand_dims(axes = attn_mask_79_axes_0, x = attn_mask_77)[name = tensor("attn_mask_79")]; + tensor var_7878_transpose_x_0 = const()[name = tensor("op_7878_transpose_x_0"), val = tensor(false)]; + tensor var_7878_transpose_y_0 = const()[name = tensor("op_7878_transpose_y_0"), val = tensor(false)]; + tensor transpose_110_perm_0 = const()[name = tensor("transpose_110_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_111_perm_0 = const()[name = tensor("transpose_111_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_111 = transpose(perm = transpose_111_perm_0, x = keys_117_cast_fp16)[name = tensor("transpose_137")]; + tensor transpose_110 = transpose(perm = transpose_110_perm_0, x = q_117_cast_fp16)[name = tensor("transpose_138")]; + tensor var_7878_cast_fp16 = matmul(transpose_x = var_7878_transpose_x_0, transpose_y = var_7878_transpose_y_0, x = transpose_110, y = transpose_111)[name = tensor("op_7878_cast_fp16")]; + tensor var_7872_to_fp16 = const()[name = tensor("op_7872_to_fp16"), val = tensor([0x1p-3])]; + tensor attn_weights_115_cast_fp16 = mul(x = var_7878_cast_fp16, y = var_7872_to_fp16)[name = tensor("attn_weights_115_cast_fp16")]; + tensor var_7880 = logical_not(x = attn_mask_79)[name = tensor("op_7880")]; + tensor var_7881_to_fp16 = const()[name = tensor("op_7881_to_fp16"), val = tensor(-inf)]; + tensor attn_weights_117_cast_fp16 = select(a = var_7881_to_fp16, b = attn_weights_115_cast_fp16, cond = var_7880)[name = tensor("attn_weights_117_cast_fp16")]; + tensor var_7883 = const()[name = tensor("op_7883"), val = tensor(-1)]; + tensor attn_weights_119_cast_fp16 = softmax(axis = var_7883, x = attn_weights_117_cast_fp16)[name = tensor("attn_weights_119_cast_fp16")]; + tensor attn_output_39_transpose_x_0 = const()[name = tensor("attn_output_39_transpose_x_0"), val = tensor(false)]; + tensor attn_output_39_transpose_y_0 = const()[name = tensor("attn_output_39_transpose_y_0"), val = tensor(false)]; + tensor values_119_cast_fp16 = transpose(perm = var_7823, x = values_117_cast_fp16)[name = tensor("transpose_139")]; + tensor attn_output_39_cast_fp16 = matmul(transpose_x = attn_output_39_transpose_x_0, transpose_y = attn_output_39_transpose_y_0, x = attn_weights_119_cast_fp16, y = values_119_cast_fp16)[name = tensor("attn_output_39_cast_fp16")]; + tensor var_7891 = const()[name = tensor("op_7891"), val = tensor([0, 2, 1, 3])]; + tensor var_7896 = const()[name = tensor("op_7896"), val = tensor([1, 1, 1024])]; + tensor var_7892_cast_fp16 = transpose(perm = var_7891, x = attn_output_39_cast_fp16)[name = tensor("transpose_136")]; + tensor input_235_cast_fp16 = reshape(shape = var_7896, x = var_7892_cast_fp16)[name = tensor("input_235_cast_fp16")]; + tensor attn19_out_proj_weight_to_fp16 = const()[name = tensor("attn19_out_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(485746304)))]; + tensor linear_78_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = attn19_out_proj_weight_to_fp16, x = input_235_cast_fp16)[name = tensor("linear_78_cast_fp16")]; + tensor var_7902_to_fp16 = const()[name = tensor("op_7902_to_fp16"), val = tensor(0x1p+0)]; + tensor var_7903_cast_fp16 = add(x = position19_to_fp16, y = var_7902_to_fp16)[name = tensor("op_7903_cast_fp16")]; + tensor var_7903_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_7903_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor input_237_cast_fp16 = add(x = input_231_cast_fp16, y = linear_78_cast_fp16)[name = tensor("input_237_cast_fp16")]; + tensor input_239_axes_0 = const()[name = tensor("input_239_axes_0"), val = tensor([-1])]; + tensor norm19_2_weight_to_fp16 = const()[name = tensor("norm19_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(487843520)))]; + tensor norm19_2_bias_to_fp16 = const()[name = tensor("norm19_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(487845632)))]; + tensor var_7907_to_fp16 = const()[name = tensor("op_7907_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_239_cast_fp16 = layer_norm(axes = input_239_axes_0, beta = norm19_2_bias_to_fp16, epsilon = var_7907_to_fp16, gamma = norm19_2_weight_to_fp16, x = input_237_cast_fp16)[name = tensor("input_239_cast_fp16")]; + tensor linear19_1_weight_to_fp16 = const()[name = tensor("linear19_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(487847744)))]; + tensor linear_79_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = linear19_1_weight_to_fp16, x = input_239_cast_fp16)[name = tensor("linear_79_cast_fp16")]; + tensor input_241_mode_0 = const()[name = tensor("input_241_mode_0"), val = tensor("EXACT")]; + tensor input_241_cast_fp16 = gelu(mode = input_241_mode_0, x = linear_79_cast_fp16)[name = tensor("input_241_cast_fp16")]; + tensor linear19_2_weight_to_fp16 = const()[name = tensor("linear19_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(496236416)))]; + tensor linear_80_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = linear19_2_weight_to_fp16, x = input_241_cast_fp16)[name = tensor("linear_80_cast_fp16")]; + tensor input_243_cast_fp16 = add(x = input_237_cast_fp16, y = linear_80_cast_fp16)[name = tensor("input_243_cast_fp16")]; + tensor input_245_axes_0 = const()[name = tensor("input_245_axes_0"), val = tensor([-1])]; + tensor norm20_1_weight_to_fp16 = const()[name = tensor("norm20_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(504625088)))]; + tensor norm20_1_bias_to_fp16 = const()[name = tensor("norm20_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(504627200)))]; + tensor var_7924_to_fp16 = const()[name = tensor("op_7924_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_245_cast_fp16 = layer_norm(axes = input_245_axes_0, beta = norm20_1_bias_to_fp16, epsilon = var_7924_to_fp16, gamma = norm20_1_weight_to_fp16, x = input_243_cast_fp16)[name = tensor("input_245_cast_fp16")]; + tensor attn20_in_proj_weight_to_fp16 = const()[name = tensor("attn20_in_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(504629312)))]; + tensor linear_81_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = attn20_in_proj_weight_to_fp16, x = input_245_cast_fp16)[name = tensor("linear_81_cast_fp16")]; + tensor var_7938 = const()[name = tensor("op_7938"), val = tensor([1, 1, 3, 16, 64])]; + tensor qkv_41_cast_fp16 = reshape(shape = var_7938, x = linear_81_cast_fp16)[name = tensor("qkv_41_cast_fp16")]; + tensor q_121_begin_0 = const()[name = tensor("q_121_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor q_121_end_0 = const()[name = tensor("q_121_end_0"), val = tensor([1, 1, 1, 16, 64])]; + tensor q_121_end_mask_0 = const()[name = tensor("q_121_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor q_121_squeeze_mask_0 = const()[name = tensor("q_121_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor q_121_cast_fp16 = slice_by_index(begin = q_121_begin_0, end = q_121_end_0, end_mask = q_121_end_mask_0, squeeze_mask = q_121_squeeze_mask_0, x = qkv_41_cast_fp16)[name = tensor("q_121_cast_fp16")]; + tensor k_81_begin_0 = const()[name = tensor("k_81_begin_0"), val = tensor([0, 0, 1, 0, 0])]; + tensor k_81_end_0 = const()[name = tensor("k_81_end_0"), val = tensor([1, 1, 2, 16, 64])]; + tensor k_81_end_mask_0 = const()[name = tensor("k_81_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor k_81_squeeze_mask_0 = const()[name = tensor("k_81_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor k_81_cast_fp16 = slice_by_index(begin = k_81_begin_0, end = k_81_end_0, end_mask = k_81_end_mask_0, squeeze_mask = k_81_squeeze_mask_0, x = qkv_41_cast_fp16)[name = tensor("k_81_cast_fp16")]; + tensor v_41_begin_0 = const()[name = tensor("v_41_begin_0"), val = tensor([0, 0, 2, 0, 0])]; + tensor v_41_end_0 = const()[name = tensor("v_41_end_0"), val = tensor([1, 1, 3, 16, 64])]; + tensor v_41_end_mask_0 = const()[name = tensor("v_41_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor v_41_squeeze_mask_0 = const()[name = tensor("v_41_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor v_41_cast_fp16 = slice_by_index(begin = v_41_begin_0, end = v_41_end_0, end_mask = v_41_end_mask_0, squeeze_mask = v_41_squeeze_mask_0, x = qkv_41_cast_fp16)[name = tensor("v_41_cast_fp16")]; + tensor var_8004 = const()[name = tensor("op_8004"), val = tensor([1, 1, 1, 1])]; + tensor position20_to_fp16_dtype_0 = const()[name = tensor("position20_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor position20_to_fp16 = cast(dtype = position20_to_fp16_dtype_0, x = position20)[name = tensor("cast_453")]; + tensor ts_125_cast_fp16 = reshape(shape = var_8004, x = position20_to_fp16)[name = tensor("ts_125_cast_fp16")]; + tensor var_8011 = const()[name = tensor("op_8011"), val = tensor([1, 1, 16, 32, 2])]; + tensor q_complex_41_cast_fp16 = reshape(shape = var_8011, x = q_121_cast_fp16)[name = tensor("q_complex_41_cast_fp16")]; + tensor var_8018 = const()[name = tensor("op_8018"), val = tensor([1, 1, 16, 32, 2])]; + tensor k_complex_41_cast_fp16 = reshape(shape = var_8018, x = k_81_cast_fp16)[name = tensor("k_complex_41_cast_fp16")]; + tensor var_8022_begin_0 = const()[name = tensor("op_8022_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_8022_end_0 = const()[name = tensor("op_8022_end_0"), val = tensor([1, 1, 16, 32, 1])]; + tensor var_8022_end_mask_0 = const()[name = tensor("op_8022_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_8022_squeeze_mask_0 = const()[name = tensor("op_8022_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_8022_cast_fp16 = slice_by_index(begin = var_8022_begin_0, end = var_8022_end_0, end_mask = var_8022_end_mask_0, squeeze_mask = var_8022_squeeze_mask_0, x = q_complex_41_cast_fp16)[name = tensor("op_8022_cast_fp16")]; + tensor var_8030_begin_0 = const()[name = tensor("op_8030_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_8030_end_0 = const()[name = tensor("op_8030_end_0"), val = tensor([1, 1, 16, 32, 2])]; + tensor var_8030_end_mask_0 = const()[name = tensor("op_8030_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_8030_squeeze_mask_0 = const()[name = tensor("op_8030_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_8030_cast_fp16 = slice_by_index(begin = var_8030_begin_0, end = var_8030_end_0, end_mask = var_8030_end_mask_0, squeeze_mask = var_8030_squeeze_mask_0, x = q_complex_41_cast_fp16)[name = tensor("op_8030_cast_fp16")]; + tensor var_8038_begin_0 = const()[name = tensor("op_8038_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_8038_end_0 = const()[name = tensor("op_8038_end_0"), val = tensor([1, 1, 16, 32, 1])]; + tensor var_8038_end_mask_0 = const()[name = tensor("op_8038_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_8038_squeeze_mask_0 = const()[name = tensor("op_8038_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_8038_cast_fp16 = slice_by_index(begin = var_8038_begin_0, end = var_8038_end_0, end_mask = var_8038_end_mask_0, squeeze_mask = var_8038_squeeze_mask_0, x = k_complex_41_cast_fp16)[name = tensor("op_8038_cast_fp16")]; + tensor var_8046_begin_0 = const()[name = tensor("op_8046_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_8046_end_0 = const()[name = tensor("op_8046_end_0"), val = tensor([1, 1, 16, 32, 2])]; + tensor var_8046_end_mask_0 = const()[name = tensor("op_8046_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_8046_squeeze_mask_0 = const()[name = tensor("op_8046_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_8046_cast_fp16 = slice_by_index(begin = var_8046_begin_0, end = var_8046_end_0, end_mask = var_8046_end_mask_0, squeeze_mask = var_8046_squeeze_mask_0, x = k_complex_41_cast_fp16)[name = tensor("op_8046_cast_fp16")]; + tensor freqs_41_to_fp16 = const()[name = tensor("freqs_41_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(510920832)))]; + tensor var_8052_cast_fp16 = mul(x = freqs_41_to_fp16, y = ts_125_cast_fp16)[name = tensor("op_8052_cast_fp16")]; + tensor rotr_41_cast_fp16 = cos(x = var_8052_cast_fp16)[name = tensor("rotr_41_cast_fp16")]; + tensor roti_41_cast_fp16 = sin(x = var_8052_cast_fp16)[name = tensor("roti_41_cast_fp16")]; + tensor var_8056_cast_fp16 = mul(x = var_8022_cast_fp16, y = rotr_41_cast_fp16)[name = tensor("op_8056_cast_fp16")]; + tensor var_8057_cast_fp16 = mul(x = var_8030_cast_fp16, y = roti_41_cast_fp16)[name = tensor("op_8057_cast_fp16")]; + tensor qor_81_cast_fp16 = sub(x = var_8056_cast_fp16, y = var_8057_cast_fp16)[name = tensor("qor_81_cast_fp16")]; + tensor var_8060_cast_fp16 = mul(x = var_8022_cast_fp16, y = roti_41_cast_fp16)[name = tensor("op_8060_cast_fp16")]; + tensor var_8061_cast_fp16 = mul(x = var_8030_cast_fp16, y = rotr_41_cast_fp16)[name = tensor("op_8061_cast_fp16")]; + tensor qoi_81_cast_fp16 = add(x = var_8060_cast_fp16, y = var_8061_cast_fp16)[name = tensor("qoi_81_cast_fp16")]; + tensor var_8064_cast_fp16 = mul(x = var_8038_cast_fp16, y = rotr_41_cast_fp16)[name = tensor("op_8064_cast_fp16")]; + tensor var_8065_cast_fp16 = mul(x = var_8046_cast_fp16, y = roti_41_cast_fp16)[name = tensor("op_8065_cast_fp16")]; + tensor kor_81_cast_fp16 = sub(x = var_8064_cast_fp16, y = var_8065_cast_fp16)[name = tensor("kor_81_cast_fp16")]; + tensor var_8068_cast_fp16 = mul(x = var_8038_cast_fp16, y = roti_41_cast_fp16)[name = tensor("op_8068_cast_fp16")]; + tensor var_8069_cast_fp16 = mul(x = var_8046_cast_fp16, y = rotr_41_cast_fp16)[name = tensor("op_8069_cast_fp16")]; + tensor koi_81_cast_fp16 = add(x = var_8068_cast_fp16, y = var_8069_cast_fp16)[name = tensor("koi_81_cast_fp16")]; + tensor qo_41_axis_0 = const()[name = tensor("qo_41_axis_0"), val = tensor(-1)]; + tensor qo_41_cast_fp16 = stack(axis = qo_41_axis_0, values = (qor_81_cast_fp16, qoi_81_cast_fp16))[name = tensor("qo_41_cast_fp16")]; + tensor ko_41_axis_0 = const()[name = tensor("ko_41_axis_0"), val = tensor(-1)]; + tensor ko_41_cast_fp16 = stack(axis = ko_41_axis_0, values = (kor_81_cast_fp16, koi_81_cast_fp16))[name = tensor("ko_41_cast_fp16")]; + tensor var_8102 = const()[name = tensor("op_8102"), val = tensor([1, 1, 16, 64])]; + tensor q_123_cast_fp16 = reshape(shape = var_8102, x = qo_41_cast_fp16)[name = tensor("q_123_cast_fp16")]; + tensor var_8108 = const()[name = tensor("op_8108"), val = tensor([1, 1, 16, 64])]; + tensor k_83_cast_fp16 = reshape(shape = var_8108, x = ko_41_cast_fp16)[name = tensor("k_83_cast_fp16")]; + tensor _inversed_8133_y_0_to_fp16 = const()[name = tensor("_inversed_8133_y_0_to_fp16"), val = tensor(0x1p-9)]; + tensor _inversed_8133_cast_fp16 = mul(x = ts_125_cast_fp16, y = _inversed_8133_y_0_to_fp16)[name = tensor("_inversed_8133_cast_fp16")]; + tensor var_8134_cast_fp16 = floor(x = _inversed_8133_cast_fp16)[name = tensor("op_8134_cast_fp16")]; + tensor var_8135_to_fp16 = const()[name = tensor("op_8135_to_fp16"), val = tensor(0x1p+9)]; + tensor var_8136_cast_fp16 = mul(x = var_8134_cast_fp16, y = var_8135_to_fp16)[name = tensor("op_8136_cast_fp16")]; + tensor write_indices_float_83_cast_fp16 = sub(x = ts_125_cast_fp16, y = var_8136_cast_fp16)[name = tensor("write_indices_float_83_cast_fp16")]; + tensor var_8143_dtype_0 = const()[name = tensor("op_8143_dtype_0"), val = tensor("int32")]; + tensor write_indices_41_reps_0 = const()[name = tensor("write_indices_41_reps_0"), val = tensor([1, 1, 16, 64])]; + tensor write_indices_float_83_cast_fp16_to_int32 = cast(dtype = var_8143_dtype_0, x = write_indices_float_83_cast_fp16)[name = tensor("cast_452")]; + tensor write_indices_41 = tile(reps = write_indices_41_reps_0, x = write_indices_float_83_cast_fp16_to_int32)[name = tensor("write_indices_41")]; + tensor var_8153_begin_0 = const()[name = tensor("op_8153_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_8153_end_0 = const()[name = tensor("op_8153_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor var_8153_end_mask_0 = const()[name = tensor("op_8153_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_8153_squeeze_mask_0 = const()[name = tensor("op_8153_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor cache20_to_fp16_dtype_0 = const()[name = tensor("cache20_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor cache20_to_fp16 = cast(dtype = cache20_to_fp16_dtype_0, x = cache20)[name = tensor("cast_451")]; + tensor var_8153_cast_fp16 = slice_by_index(begin = var_8153_begin_0, end = var_8153_end_0, end_mask = var_8153_end_mask_0, squeeze_mask = var_8153_squeeze_mask_0, x = cache20_to_fp16)[name = tensor("op_8153_cast_fp16")]; + tensor var_8155_axis_0 = const()[name = tensor("op_8155_axis_0"), val = tensor(1)]; + tensor var_8155_mode_0 = const()[name = tensor("op_8155_mode_0"), val = tensor("update")]; + tensor var_8155_validate_indices_0 = const()[name = tensor("op_8155_validate_indices_0"), val = tensor(false)]; + tensor var_8155_cast_fp16 = scatter_along_axis(axis = var_8155_axis_0, data = var_8153_cast_fp16, indices = write_indices_41, mode = var_8155_mode_0, updates = k_83_cast_fp16, validate_indices = var_8155_validate_indices_0)[name = tensor("op_8155_cast_fp16")]; + tensor concat_142 = const()[name = tensor("concat_142"), val = tensor([0, 0, 0, 0, 0])]; + tensor concat_143 = const()[name = tensor("concat_143"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_41_internal_tensor_assign_1_stride_0 = const()[name = tensor("new_cache_41_internal_tensor_assign_1_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_41_internal_tensor_assign_1_begin_mask_0 = const()[name = tensor("new_cache_41_internal_tensor_assign_1_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_41_internal_tensor_assign_1_end_mask_0 = const()[name = tensor("new_cache_41_internal_tensor_assign_1_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_41_internal_tensor_assign_1_squeeze_mask_0 = const()[name = tensor("new_cache_41_internal_tensor_assign_1_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_88 = const()[name = tensor("shape_88"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_40 = const()[name = tensor("reduce_prod_40"), val = tensor(1048576)]; + tensor range_1d_40_start_0 = const()[name = tensor("range_1d_40_start_0"), val = tensor(0)]; + tensor range_1d_40_step_0 = const()[name = tensor("range_1d_40_step_0"), val = tensor(1)]; + tensor range_1d_40 = range_1d(end = reduce_prod_40, start = range_1d_40_start_0, step = range_1d_40_step_0)[name = tensor("range_1d_40")]; + tensor reshape_200 = reshape(shape = shape_88, x = range_1d_40)[name = tensor("reshape_200")]; + tensor slice_by_index_40 = slice_by_index(begin = concat_142, begin_mask = new_cache_41_internal_tensor_assign_1_begin_mask_0, end = concat_143, end_mask = new_cache_41_internal_tensor_assign_1_end_mask_0, squeeze_mask = new_cache_41_internal_tensor_assign_1_squeeze_mask_0, stride = new_cache_41_internal_tensor_assign_1_stride_0, x = reshape_200)[name = tensor("slice_by_index_40")]; + tensor reshape_201_shape_0 = const()[name = tensor("reshape_201_shape_0"), val = tensor([-1])]; + tensor reshape_201 = reshape(shape = reshape_201_shape_0, x = slice_by_index_40)[name = tensor("reshape_201")]; + tensor reshape_202_shape_0 = const()[name = tensor("reshape_202_shape_0"), val = tensor([-1])]; + tensor reshape_202_cast_fp16 = reshape(shape = reshape_202_shape_0, x = var_8155_cast_fp16)[name = tensor("reshape_202_cast_fp16")]; + tensor reshape_203_shape_0 = const()[name = tensor("reshape_203_shape_0"), val = tensor([-1])]; + tensor reshape_203_cast_fp16 = reshape(shape = reshape_203_shape_0, x = cache20_to_fp16)[name = tensor("reshape_203_cast_fp16")]; + tensor scatter_40_mode_0 = const()[name = tensor("scatter_40_mode_0"), val = tensor("update")]; + tensor scatter_40_axis_0 = const()[name = tensor("scatter_40_axis_0"), val = tensor(0)]; + tensor scatter_40_validate_indices_0 = const()[name = tensor("scatter_40_validate_indices_0"), val = tensor(false)]; + tensor scatter_40_cast_fp16 = scatter(axis = scatter_40_axis_0, data = reshape_203_cast_fp16, indices = reshape_201, mode = scatter_40_mode_0, updates = reshape_202_cast_fp16, validate_indices = scatter_40_validate_indices_0)[name = tensor("scatter_40_cast_fp16")]; + tensor reshape_204_cast_fp16 = reshape(shape = shape_88, x = scatter_40_cast_fp16)[name = tensor("reshape_204_cast_fp16")]; + tensor var_8163_begin_0 = const()[name = tensor("op_8163_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor var_8163_end_0 = const()[name = tensor("op_8163_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor var_8163_end_mask_0 = const()[name = tensor("op_8163_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_8163_squeeze_mask_0 = const()[name = tensor("op_8163_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor var_8163_cast_fp16 = slice_by_index(begin = var_8163_begin_0, end = var_8163_end_0, end_mask = var_8163_end_mask_0, squeeze_mask = var_8163_squeeze_mask_0, x = reshape_204_cast_fp16)[name = tensor("op_8163_cast_fp16")]; + tensor var_8165_axis_0 = const()[name = tensor("op_8165_axis_0"), val = tensor(1)]; + tensor var_8165_mode_0 = const()[name = tensor("op_8165_mode_0"), val = tensor("update")]; + tensor var_8165_validate_indices_0 = const()[name = tensor("op_8165_validate_indices_0"), val = tensor(false)]; + tensor var_8165_cast_fp16 = scatter_along_axis(axis = var_8165_axis_0, data = var_8163_cast_fp16, indices = write_indices_41, mode = var_8165_mode_0, updates = v_41_cast_fp16, validate_indices = var_8165_validate_indices_0)[name = tensor("op_8165_cast_fp16")]; + tensor concat_144 = const()[name = tensor("concat_144"), val = tensor([1, 0, 0, 0, 0])]; + tensor concat_145 = const()[name = tensor("concat_145"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_41_internal_tensor_assign_2_stride_0 = const()[name = tensor("new_cache_41_internal_tensor_assign_2_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_41_internal_tensor_assign_2_begin_mask_0 = const()[name = tensor("new_cache_41_internal_tensor_assign_2_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_41_internal_tensor_assign_2_end_mask_0 = const()[name = tensor("new_cache_41_internal_tensor_assign_2_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_41_internal_tensor_assign_2_squeeze_mask_0 = const()[name = tensor("new_cache_41_internal_tensor_assign_2_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_89 = const()[name = tensor("shape_89"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_41 = const()[name = tensor("reduce_prod_41"), val = tensor(1048576)]; + tensor range_1d_41_start_0 = const()[name = tensor("range_1d_41_start_0"), val = tensor(0)]; + tensor range_1d_41_step_0 = const()[name = tensor("range_1d_41_step_0"), val = tensor(1)]; + tensor range_1d_41 = range_1d(end = reduce_prod_41, start = range_1d_41_start_0, step = range_1d_41_step_0)[name = tensor("range_1d_41")]; + tensor reshape_205 = reshape(shape = shape_89, x = range_1d_41)[name = tensor("reshape_205")]; + tensor slice_by_index_41 = slice_by_index(begin = concat_144, begin_mask = new_cache_41_internal_tensor_assign_2_begin_mask_0, end = concat_145, end_mask = new_cache_41_internal_tensor_assign_2_end_mask_0, squeeze_mask = new_cache_41_internal_tensor_assign_2_squeeze_mask_0, stride = new_cache_41_internal_tensor_assign_2_stride_0, x = reshape_205)[name = tensor("slice_by_index_41")]; + tensor reshape_206_shape_0 = const()[name = tensor("reshape_206_shape_0"), val = tensor([-1])]; + tensor reshape_206 = reshape(shape = reshape_206_shape_0, x = slice_by_index_41)[name = tensor("reshape_206")]; + tensor reshape_207_shape_0 = const()[name = tensor("reshape_207_shape_0"), val = tensor([-1])]; + tensor reshape_207_cast_fp16 = reshape(shape = reshape_207_shape_0, x = var_8165_cast_fp16)[name = tensor("reshape_207_cast_fp16")]; + tensor reshape_208_shape_0 = const()[name = tensor("reshape_208_shape_0"), val = tensor([-1])]; + tensor reshape_208_cast_fp16 = reshape(shape = reshape_208_shape_0, x = reshape_204_cast_fp16)[name = tensor("reshape_208_cast_fp16")]; + tensor scatter_41_mode_0 = const()[name = tensor("scatter_41_mode_0"), val = tensor("update")]; + tensor scatter_41_axis_0 = const()[name = tensor("scatter_41_axis_0"), val = tensor(0)]; + tensor scatter_41_validate_indices_0 = const()[name = tensor("scatter_41_validate_indices_0"), val = tensor(false)]; + tensor scatter_41_cast_fp16 = scatter(axis = scatter_41_axis_0, data = reshape_208_cast_fp16, indices = reshape_206, mode = scatter_41_mode_0, updates = reshape_207_cast_fp16, validate_indices = scatter_41_validate_indices_0)[name = tensor("scatter_41_cast_fp16")]; + tensor reshape_209_cast_fp16 = reshape(shape = shape_89, x = scatter_41_cast_fp16)[name = tensor("reshape_209_cast_fp16")]; + tensor reshape_209_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("reshape_209_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor keys_121_begin_0 = const()[name = tensor("keys_121_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor keys_121_end_0 = const()[name = tensor("keys_121_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor keys_121_end_mask_0 = const()[name = tensor("keys_121_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor keys_121_squeeze_mask_0 = const()[name = tensor("keys_121_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor keys_121_cast_fp16 = slice_by_index(begin = keys_121_begin_0, end = keys_121_end_0, end_mask = keys_121_end_mask_0, squeeze_mask = keys_121_squeeze_mask_0, x = reshape_209_cast_fp16)[name = tensor("keys_121_cast_fp16")]; + tensor values_121_begin_0 = const()[name = tensor("values_121_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor values_121_end_0 = const()[name = tensor("values_121_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor values_121_end_mask_0 = const()[name = tensor("values_121_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor values_121_squeeze_mask_0 = const()[name = tensor("values_121_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor values_121_cast_fp16 = slice_by_index(begin = values_121_begin_0, end = values_121_end_0, end_mask = values_121_end_mask_0, squeeze_mask = values_121_squeeze_mask_0, x = reshape_209_cast_fp16)[name = tensor("values_121_cast_fp16")]; + tensor var_8177_cast_fp16 = not_equal(x = keys_121_cast_fp16, y = keys_121_cast_fp16)[name = tensor("op_8177_cast_fp16")]; + tensor keys_123_cast_fp16 = select(a = var_463_to_fp16, b = keys_121_cast_fp16, cond = var_8177_cast_fp16)[name = tensor("keys_123_cast_fp16")]; + tensor var_8185_cast_fp16 = not_equal(x = values_121_cast_fp16, y = values_121_cast_fp16)[name = tensor("op_8185_cast_fp16")]; + tensor values_123_cast_fp16 = select(a = var_463_to_fp16, b = values_121_cast_fp16, cond = var_8185_cast_fp16)[name = tensor("values_123_cast_fp16")]; + tensor var_8209 = const()[name = tensor("op_8209"), val = tensor([0, 2, 1, 3])]; + tensor var_8225 = const()[name = tensor("op_8225"), val = tensor([1, 1, 1])]; + tensor var_8226_cast_fp16 = reshape(shape = var_8225, x = position20_to_fp16)[name = tensor("op_8226_cast_fp16")]; + tensor var_8246_to_fp16 = const()[name = tensor("op_8246_to_fp16"), val = tensor(0x1p+0)]; + tensor valid_len_41_cast_fp16 = add(x = var_8226_cast_fp16, y = var_8246_to_fp16)[name = tensor("valid_len_41_cast_fp16")]; + tensor valid_mask_41_cast_fp16 = less(x = k_positions_1_promoted_to_fp16, y = valid_len_41_cast_fp16)[name = tensor("valid_mask_41_cast_fp16")]; + tensor causal_mask_41_cast_fp16 = less_equal(x = k_positions_1_promoted_to_fp16, y = var_8226_cast_fp16)[name = tensor("causal_mask_41_cast_fp16")]; + tensor attn_mask_81 = logical_and(x = valid_mask_41_cast_fp16, y = causal_mask_41_cast_fp16)[name = tensor("attn_mask_81")]; + tensor attn_mask_83_axes_0 = const()[name = tensor("attn_mask_83_axes_0"), val = tensor([1])]; + tensor attn_mask_83 = expand_dims(axes = attn_mask_83_axes_0, x = attn_mask_81)[name = tensor("attn_mask_83")]; + tensor var_8264_transpose_x_0 = const()[name = tensor("op_8264_transpose_x_0"), val = tensor(false)]; + tensor var_8264_transpose_y_0 = const()[name = tensor("op_8264_transpose_y_0"), val = tensor(false)]; + tensor transpose_112_perm_0 = const()[name = tensor("transpose_112_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_113_perm_0 = const()[name = tensor("transpose_113_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_113 = transpose(perm = transpose_113_perm_0, x = keys_123_cast_fp16)[name = tensor("transpose_133")]; + tensor transpose_112 = transpose(perm = transpose_112_perm_0, x = q_123_cast_fp16)[name = tensor("transpose_134")]; + tensor var_8264_cast_fp16 = matmul(transpose_x = var_8264_transpose_x_0, transpose_y = var_8264_transpose_y_0, x = transpose_112, y = transpose_113)[name = tensor("op_8264_cast_fp16")]; + tensor var_8258_to_fp16 = const()[name = tensor("op_8258_to_fp16"), val = tensor([0x1p-3])]; + tensor attn_weights_121_cast_fp16 = mul(x = var_8264_cast_fp16, y = var_8258_to_fp16)[name = tensor("attn_weights_121_cast_fp16")]; + tensor var_8266 = logical_not(x = attn_mask_83)[name = tensor("op_8266")]; + tensor var_8267_to_fp16 = const()[name = tensor("op_8267_to_fp16"), val = tensor(-inf)]; + tensor attn_weights_123_cast_fp16 = select(a = var_8267_to_fp16, b = attn_weights_121_cast_fp16, cond = var_8266)[name = tensor("attn_weights_123_cast_fp16")]; + tensor var_8269 = const()[name = tensor("op_8269"), val = tensor(-1)]; + tensor attn_weights_125_cast_fp16 = softmax(axis = var_8269, x = attn_weights_123_cast_fp16)[name = tensor("attn_weights_125_cast_fp16")]; + tensor attn_output_41_transpose_x_0 = const()[name = tensor("attn_output_41_transpose_x_0"), val = tensor(false)]; + tensor attn_output_41_transpose_y_0 = const()[name = tensor("attn_output_41_transpose_y_0"), val = tensor(false)]; + tensor values_125_cast_fp16 = transpose(perm = var_8209, x = values_123_cast_fp16)[name = tensor("transpose_135")]; + tensor attn_output_41_cast_fp16 = matmul(transpose_x = attn_output_41_transpose_x_0, transpose_y = attn_output_41_transpose_y_0, x = attn_weights_125_cast_fp16, y = values_125_cast_fp16)[name = tensor("attn_output_41_cast_fp16")]; + tensor var_8277 = const()[name = tensor("op_8277"), val = tensor([0, 2, 1, 3])]; + tensor var_8282 = const()[name = tensor("op_8282"), val = tensor([1, 1, 1024])]; + tensor var_8278_cast_fp16 = transpose(perm = var_8277, x = attn_output_41_cast_fp16)[name = tensor("transpose_132")]; + tensor input_247_cast_fp16 = reshape(shape = var_8282, x = var_8278_cast_fp16)[name = tensor("input_247_cast_fp16")]; + tensor attn20_out_proj_weight_to_fp16 = const()[name = tensor("attn20_out_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(510920960)))]; + tensor linear_82_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = attn20_out_proj_weight_to_fp16, x = input_247_cast_fp16)[name = tensor("linear_82_cast_fp16")]; + tensor var_8288_to_fp16 = const()[name = tensor("op_8288_to_fp16"), val = tensor(0x1p+0)]; + tensor var_8289_cast_fp16 = add(x = position20_to_fp16, y = var_8288_to_fp16)[name = tensor("op_8289_cast_fp16")]; + tensor var_8289_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_8289_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor input_249_cast_fp16 = add(x = input_243_cast_fp16, y = linear_82_cast_fp16)[name = tensor("input_249_cast_fp16")]; + tensor input_251_axes_0 = const()[name = tensor("input_251_axes_0"), val = tensor([-1])]; + tensor norm20_2_weight_to_fp16 = const()[name = tensor("norm20_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(513018176)))]; + tensor norm20_2_bias_to_fp16 = const()[name = tensor("norm20_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(513020288)))]; + tensor var_8293_to_fp16 = const()[name = tensor("op_8293_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_251_cast_fp16 = layer_norm(axes = input_251_axes_0, beta = norm20_2_bias_to_fp16, epsilon = var_8293_to_fp16, gamma = norm20_2_weight_to_fp16, x = input_249_cast_fp16)[name = tensor("input_251_cast_fp16")]; + tensor linear20_1_weight_to_fp16 = const()[name = tensor("linear20_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(513022400)))]; + tensor linear_83_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = linear20_1_weight_to_fp16, x = input_251_cast_fp16)[name = tensor("linear_83_cast_fp16")]; + tensor input_253_mode_0 = const()[name = tensor("input_253_mode_0"), val = tensor("EXACT")]; + tensor input_253_cast_fp16 = gelu(mode = input_253_mode_0, x = linear_83_cast_fp16)[name = tensor("input_253_cast_fp16")]; + tensor linear20_2_weight_to_fp16 = const()[name = tensor("linear20_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(521411072)))]; + tensor linear_84_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = linear20_2_weight_to_fp16, x = input_253_cast_fp16)[name = tensor("linear_84_cast_fp16")]; + tensor input_255_cast_fp16 = add(x = input_249_cast_fp16, y = linear_84_cast_fp16)[name = tensor("input_255_cast_fp16")]; + tensor input_257_axes_0 = const()[name = tensor("input_257_axes_0"), val = tensor([-1])]; + tensor norm21_1_weight_to_fp16 = const()[name = tensor("norm21_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(529799744)))]; + tensor norm21_1_bias_to_fp16 = const()[name = tensor("norm21_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(529801856)))]; + tensor var_8310_to_fp16 = const()[name = tensor("op_8310_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_257_cast_fp16 = layer_norm(axes = input_257_axes_0, beta = norm21_1_bias_to_fp16, epsilon = var_8310_to_fp16, gamma = norm21_1_weight_to_fp16, x = input_255_cast_fp16)[name = tensor("input_257_cast_fp16")]; + tensor attn21_in_proj_weight_to_fp16 = const()[name = tensor("attn21_in_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(529803968)))]; + tensor linear_85_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = attn21_in_proj_weight_to_fp16, x = input_257_cast_fp16)[name = tensor("linear_85_cast_fp16")]; + tensor var_8324 = const()[name = tensor("op_8324"), val = tensor([1, 1, 3, 16, 64])]; + tensor qkv_43_cast_fp16 = reshape(shape = var_8324, x = linear_85_cast_fp16)[name = tensor("qkv_43_cast_fp16")]; + tensor q_127_begin_0 = const()[name = tensor("q_127_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor q_127_end_0 = const()[name = tensor("q_127_end_0"), val = tensor([1, 1, 1, 16, 64])]; + tensor q_127_end_mask_0 = const()[name = tensor("q_127_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor q_127_squeeze_mask_0 = const()[name = tensor("q_127_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor q_127_cast_fp16 = slice_by_index(begin = q_127_begin_0, end = q_127_end_0, end_mask = q_127_end_mask_0, squeeze_mask = q_127_squeeze_mask_0, x = qkv_43_cast_fp16)[name = tensor("q_127_cast_fp16")]; + tensor k_85_begin_0 = const()[name = tensor("k_85_begin_0"), val = tensor([0, 0, 1, 0, 0])]; + tensor k_85_end_0 = const()[name = tensor("k_85_end_0"), val = tensor([1, 1, 2, 16, 64])]; + tensor k_85_end_mask_0 = const()[name = tensor("k_85_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor k_85_squeeze_mask_0 = const()[name = tensor("k_85_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor k_85_cast_fp16 = slice_by_index(begin = k_85_begin_0, end = k_85_end_0, end_mask = k_85_end_mask_0, squeeze_mask = k_85_squeeze_mask_0, x = qkv_43_cast_fp16)[name = tensor("k_85_cast_fp16")]; + tensor v_43_begin_0 = const()[name = tensor("v_43_begin_0"), val = tensor([0, 0, 2, 0, 0])]; + tensor v_43_end_0 = const()[name = tensor("v_43_end_0"), val = tensor([1, 1, 3, 16, 64])]; + tensor v_43_end_mask_0 = const()[name = tensor("v_43_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor v_43_squeeze_mask_0 = const()[name = tensor("v_43_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor v_43_cast_fp16 = slice_by_index(begin = v_43_begin_0, end = v_43_end_0, end_mask = v_43_end_mask_0, squeeze_mask = v_43_squeeze_mask_0, x = qkv_43_cast_fp16)[name = tensor("v_43_cast_fp16")]; + tensor var_8390 = const()[name = tensor("op_8390"), val = tensor([1, 1, 1, 1])]; + tensor position21_to_fp16_dtype_0 = const()[name = tensor("position21_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor position21_to_fp16 = cast(dtype = position21_to_fp16_dtype_0, x = position21)[name = tensor("cast_448")]; + tensor ts_131_cast_fp16 = reshape(shape = var_8390, x = position21_to_fp16)[name = tensor("ts_131_cast_fp16")]; + tensor var_8397 = const()[name = tensor("op_8397"), val = tensor([1, 1, 16, 32, 2])]; + tensor q_complex_43_cast_fp16 = reshape(shape = var_8397, x = q_127_cast_fp16)[name = tensor("q_complex_43_cast_fp16")]; + tensor var_8404 = const()[name = tensor("op_8404"), val = tensor([1, 1, 16, 32, 2])]; + tensor k_complex_43_cast_fp16 = reshape(shape = var_8404, x = k_85_cast_fp16)[name = tensor("k_complex_43_cast_fp16")]; + tensor var_8408_begin_0 = const()[name = tensor("op_8408_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_8408_end_0 = const()[name = tensor("op_8408_end_0"), val = tensor([1, 1, 16, 32, 1])]; + tensor var_8408_end_mask_0 = const()[name = tensor("op_8408_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_8408_squeeze_mask_0 = const()[name = tensor("op_8408_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_8408_cast_fp16 = slice_by_index(begin = var_8408_begin_0, end = var_8408_end_0, end_mask = var_8408_end_mask_0, squeeze_mask = var_8408_squeeze_mask_0, x = q_complex_43_cast_fp16)[name = tensor("op_8408_cast_fp16")]; + tensor var_8416_begin_0 = const()[name = tensor("op_8416_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_8416_end_0 = const()[name = tensor("op_8416_end_0"), val = tensor([1, 1, 16, 32, 2])]; + tensor var_8416_end_mask_0 = const()[name = tensor("op_8416_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_8416_squeeze_mask_0 = const()[name = tensor("op_8416_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_8416_cast_fp16 = slice_by_index(begin = var_8416_begin_0, end = var_8416_end_0, end_mask = var_8416_end_mask_0, squeeze_mask = var_8416_squeeze_mask_0, x = q_complex_43_cast_fp16)[name = tensor("op_8416_cast_fp16")]; + tensor var_8424_begin_0 = const()[name = tensor("op_8424_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_8424_end_0 = const()[name = tensor("op_8424_end_0"), val = tensor([1, 1, 16, 32, 1])]; + tensor var_8424_end_mask_0 = const()[name = tensor("op_8424_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_8424_squeeze_mask_0 = const()[name = tensor("op_8424_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_8424_cast_fp16 = slice_by_index(begin = var_8424_begin_0, end = var_8424_end_0, end_mask = var_8424_end_mask_0, squeeze_mask = var_8424_squeeze_mask_0, x = k_complex_43_cast_fp16)[name = tensor("op_8424_cast_fp16")]; + tensor var_8432_begin_0 = const()[name = tensor("op_8432_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_8432_end_0 = const()[name = tensor("op_8432_end_0"), val = tensor([1, 1, 16, 32, 2])]; + tensor var_8432_end_mask_0 = const()[name = tensor("op_8432_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_8432_squeeze_mask_0 = const()[name = tensor("op_8432_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_8432_cast_fp16 = slice_by_index(begin = var_8432_begin_0, end = var_8432_end_0, end_mask = var_8432_end_mask_0, squeeze_mask = var_8432_squeeze_mask_0, x = k_complex_43_cast_fp16)[name = tensor("op_8432_cast_fp16")]; + tensor freqs_43_to_fp16 = const()[name = tensor("freqs_43_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(536095488)))]; + tensor var_8438_cast_fp16 = mul(x = freqs_43_to_fp16, y = ts_131_cast_fp16)[name = tensor("op_8438_cast_fp16")]; + tensor rotr_43_cast_fp16 = cos(x = var_8438_cast_fp16)[name = tensor("rotr_43_cast_fp16")]; + tensor roti_43_cast_fp16 = sin(x = var_8438_cast_fp16)[name = tensor("roti_43_cast_fp16")]; + tensor var_8442_cast_fp16 = mul(x = var_8408_cast_fp16, y = rotr_43_cast_fp16)[name = tensor("op_8442_cast_fp16")]; + tensor var_8443_cast_fp16 = mul(x = var_8416_cast_fp16, y = roti_43_cast_fp16)[name = tensor("op_8443_cast_fp16")]; + tensor qor_85_cast_fp16 = sub(x = var_8442_cast_fp16, y = var_8443_cast_fp16)[name = tensor("qor_85_cast_fp16")]; + tensor var_8446_cast_fp16 = mul(x = var_8408_cast_fp16, y = roti_43_cast_fp16)[name = tensor("op_8446_cast_fp16")]; + tensor var_8447_cast_fp16 = mul(x = var_8416_cast_fp16, y = rotr_43_cast_fp16)[name = tensor("op_8447_cast_fp16")]; + tensor qoi_85_cast_fp16 = add(x = var_8446_cast_fp16, y = var_8447_cast_fp16)[name = tensor("qoi_85_cast_fp16")]; + tensor var_8450_cast_fp16 = mul(x = var_8424_cast_fp16, y = rotr_43_cast_fp16)[name = tensor("op_8450_cast_fp16")]; + tensor var_8451_cast_fp16 = mul(x = var_8432_cast_fp16, y = roti_43_cast_fp16)[name = tensor("op_8451_cast_fp16")]; + tensor kor_85_cast_fp16 = sub(x = var_8450_cast_fp16, y = var_8451_cast_fp16)[name = tensor("kor_85_cast_fp16")]; + tensor var_8454_cast_fp16 = mul(x = var_8424_cast_fp16, y = roti_43_cast_fp16)[name = tensor("op_8454_cast_fp16")]; + tensor var_8455_cast_fp16 = mul(x = var_8432_cast_fp16, y = rotr_43_cast_fp16)[name = tensor("op_8455_cast_fp16")]; + tensor koi_85_cast_fp16 = add(x = var_8454_cast_fp16, y = var_8455_cast_fp16)[name = tensor("koi_85_cast_fp16")]; + tensor qo_43_axis_0 = const()[name = tensor("qo_43_axis_0"), val = tensor(-1)]; + tensor qo_43_cast_fp16 = stack(axis = qo_43_axis_0, values = (qor_85_cast_fp16, qoi_85_cast_fp16))[name = tensor("qo_43_cast_fp16")]; + tensor ko_43_axis_0 = const()[name = tensor("ko_43_axis_0"), val = tensor(-1)]; + tensor ko_43_cast_fp16 = stack(axis = ko_43_axis_0, values = (kor_85_cast_fp16, koi_85_cast_fp16))[name = tensor("ko_43_cast_fp16")]; + tensor var_8488 = const()[name = tensor("op_8488"), val = tensor([1, 1, 16, 64])]; + tensor q_129_cast_fp16 = reshape(shape = var_8488, x = qo_43_cast_fp16)[name = tensor("q_129_cast_fp16")]; + tensor var_8494 = const()[name = tensor("op_8494"), val = tensor([1, 1, 16, 64])]; + tensor k_87_cast_fp16 = reshape(shape = var_8494, x = ko_43_cast_fp16)[name = tensor("k_87_cast_fp16")]; + tensor _inversed_8519_y_0_to_fp16 = const()[name = tensor("_inversed_8519_y_0_to_fp16"), val = tensor(0x1p-9)]; + tensor _inversed_8519_cast_fp16 = mul(x = ts_131_cast_fp16, y = _inversed_8519_y_0_to_fp16)[name = tensor("_inversed_8519_cast_fp16")]; + tensor var_8520_cast_fp16 = floor(x = _inversed_8519_cast_fp16)[name = tensor("op_8520_cast_fp16")]; + tensor var_8521_to_fp16 = const()[name = tensor("op_8521_to_fp16"), val = tensor(0x1p+9)]; + tensor var_8522_cast_fp16 = mul(x = var_8520_cast_fp16, y = var_8521_to_fp16)[name = tensor("op_8522_cast_fp16")]; + tensor write_indices_float_87_cast_fp16 = sub(x = ts_131_cast_fp16, y = var_8522_cast_fp16)[name = tensor("write_indices_float_87_cast_fp16")]; + tensor var_8529_dtype_0 = const()[name = tensor("op_8529_dtype_0"), val = tensor("int32")]; + tensor write_indices_43_reps_0 = const()[name = tensor("write_indices_43_reps_0"), val = tensor([1, 1, 16, 64])]; + tensor write_indices_float_87_cast_fp16_to_int32 = cast(dtype = var_8529_dtype_0, x = write_indices_float_87_cast_fp16)[name = tensor("cast_447")]; + tensor write_indices_43 = tile(reps = write_indices_43_reps_0, x = write_indices_float_87_cast_fp16_to_int32)[name = tensor("write_indices_43")]; + tensor var_8539_begin_0 = const()[name = tensor("op_8539_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_8539_end_0 = const()[name = tensor("op_8539_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor var_8539_end_mask_0 = const()[name = tensor("op_8539_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_8539_squeeze_mask_0 = const()[name = tensor("op_8539_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor cache21_to_fp16_dtype_0 = const()[name = tensor("cache21_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor cache21_to_fp16 = cast(dtype = cache21_to_fp16_dtype_0, x = cache21)[name = tensor("cast_446")]; + tensor var_8539_cast_fp16 = slice_by_index(begin = var_8539_begin_0, end = var_8539_end_0, end_mask = var_8539_end_mask_0, squeeze_mask = var_8539_squeeze_mask_0, x = cache21_to_fp16)[name = tensor("op_8539_cast_fp16")]; + tensor var_8541_axis_0 = const()[name = tensor("op_8541_axis_0"), val = tensor(1)]; + tensor var_8541_mode_0 = const()[name = tensor("op_8541_mode_0"), val = tensor("update")]; + tensor var_8541_validate_indices_0 = const()[name = tensor("op_8541_validate_indices_0"), val = tensor(false)]; + tensor var_8541_cast_fp16 = scatter_along_axis(axis = var_8541_axis_0, data = var_8539_cast_fp16, indices = write_indices_43, mode = var_8541_mode_0, updates = k_87_cast_fp16, validate_indices = var_8541_validate_indices_0)[name = tensor("op_8541_cast_fp16")]; + tensor concat_149 = const()[name = tensor("concat_149"), val = tensor([0, 0, 0, 0, 0])]; + tensor concat_150 = const()[name = tensor("concat_150"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_43_internal_tensor_assign_1_stride_0 = const()[name = tensor("new_cache_43_internal_tensor_assign_1_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_43_internal_tensor_assign_1_begin_mask_0 = const()[name = tensor("new_cache_43_internal_tensor_assign_1_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_43_internal_tensor_assign_1_end_mask_0 = const()[name = tensor("new_cache_43_internal_tensor_assign_1_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_43_internal_tensor_assign_1_squeeze_mask_0 = const()[name = tensor("new_cache_43_internal_tensor_assign_1_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_90 = const()[name = tensor("shape_90"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_42 = const()[name = tensor("reduce_prod_42"), val = tensor(1048576)]; + tensor range_1d_42_start_0 = const()[name = tensor("range_1d_42_start_0"), val = tensor(0)]; + tensor range_1d_42_step_0 = const()[name = tensor("range_1d_42_step_0"), val = tensor(1)]; + tensor range_1d_42 = range_1d(end = reduce_prod_42, start = range_1d_42_start_0, step = range_1d_42_step_0)[name = tensor("range_1d_42")]; + tensor reshape_210 = reshape(shape = shape_90, x = range_1d_42)[name = tensor("reshape_210")]; + tensor slice_by_index_42 = slice_by_index(begin = concat_149, begin_mask = new_cache_43_internal_tensor_assign_1_begin_mask_0, end = concat_150, end_mask = new_cache_43_internal_tensor_assign_1_end_mask_0, squeeze_mask = new_cache_43_internal_tensor_assign_1_squeeze_mask_0, stride = new_cache_43_internal_tensor_assign_1_stride_0, x = reshape_210)[name = tensor("slice_by_index_42")]; + tensor reshape_211_shape_0 = const()[name = tensor("reshape_211_shape_0"), val = tensor([-1])]; + tensor reshape_211 = reshape(shape = reshape_211_shape_0, x = slice_by_index_42)[name = tensor("reshape_211")]; + tensor reshape_212_shape_0 = const()[name = tensor("reshape_212_shape_0"), val = tensor([-1])]; + tensor reshape_212_cast_fp16 = reshape(shape = reshape_212_shape_0, x = var_8541_cast_fp16)[name = tensor("reshape_212_cast_fp16")]; + tensor reshape_213_shape_0 = const()[name = tensor("reshape_213_shape_0"), val = tensor([-1])]; + tensor reshape_213_cast_fp16 = reshape(shape = reshape_213_shape_0, x = cache21_to_fp16)[name = tensor("reshape_213_cast_fp16")]; + tensor scatter_42_mode_0 = const()[name = tensor("scatter_42_mode_0"), val = tensor("update")]; + tensor scatter_42_axis_0 = const()[name = tensor("scatter_42_axis_0"), val = tensor(0)]; + tensor scatter_42_validate_indices_0 = const()[name = tensor("scatter_42_validate_indices_0"), val = tensor(false)]; + tensor scatter_42_cast_fp16 = scatter(axis = scatter_42_axis_0, data = reshape_213_cast_fp16, indices = reshape_211, mode = scatter_42_mode_0, updates = reshape_212_cast_fp16, validate_indices = scatter_42_validate_indices_0)[name = tensor("scatter_42_cast_fp16")]; + tensor reshape_214_cast_fp16 = reshape(shape = shape_90, x = scatter_42_cast_fp16)[name = tensor("reshape_214_cast_fp16")]; + tensor var_8549_begin_0 = const()[name = tensor("op_8549_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor var_8549_end_0 = const()[name = tensor("op_8549_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor var_8549_end_mask_0 = const()[name = tensor("op_8549_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_8549_squeeze_mask_0 = const()[name = tensor("op_8549_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor var_8549_cast_fp16 = slice_by_index(begin = var_8549_begin_0, end = var_8549_end_0, end_mask = var_8549_end_mask_0, squeeze_mask = var_8549_squeeze_mask_0, x = reshape_214_cast_fp16)[name = tensor("op_8549_cast_fp16")]; + tensor var_8551_axis_0 = const()[name = tensor("op_8551_axis_0"), val = tensor(1)]; + tensor var_8551_mode_0 = const()[name = tensor("op_8551_mode_0"), val = tensor("update")]; + tensor var_8551_validate_indices_0 = const()[name = tensor("op_8551_validate_indices_0"), val = tensor(false)]; + tensor var_8551_cast_fp16 = scatter_along_axis(axis = var_8551_axis_0, data = var_8549_cast_fp16, indices = write_indices_43, mode = var_8551_mode_0, updates = v_43_cast_fp16, validate_indices = var_8551_validate_indices_0)[name = tensor("op_8551_cast_fp16")]; + tensor concat_151 = const()[name = tensor("concat_151"), val = tensor([1, 0, 0, 0, 0])]; + tensor concat_152 = const()[name = tensor("concat_152"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_43_internal_tensor_assign_2_stride_0 = const()[name = tensor("new_cache_43_internal_tensor_assign_2_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_43_internal_tensor_assign_2_begin_mask_0 = const()[name = tensor("new_cache_43_internal_tensor_assign_2_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_43_internal_tensor_assign_2_end_mask_0 = const()[name = tensor("new_cache_43_internal_tensor_assign_2_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_43_internal_tensor_assign_2_squeeze_mask_0 = const()[name = tensor("new_cache_43_internal_tensor_assign_2_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_91 = const()[name = tensor("shape_91"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_43 = const()[name = tensor("reduce_prod_43"), val = tensor(1048576)]; + tensor range_1d_43_start_0 = const()[name = tensor("range_1d_43_start_0"), val = tensor(0)]; + tensor range_1d_43_step_0 = const()[name = tensor("range_1d_43_step_0"), val = tensor(1)]; + tensor range_1d_43 = range_1d(end = reduce_prod_43, start = range_1d_43_start_0, step = range_1d_43_step_0)[name = tensor("range_1d_43")]; + tensor reshape_215 = reshape(shape = shape_91, x = range_1d_43)[name = tensor("reshape_215")]; + tensor slice_by_index_43 = slice_by_index(begin = concat_151, begin_mask = new_cache_43_internal_tensor_assign_2_begin_mask_0, end = concat_152, end_mask = new_cache_43_internal_tensor_assign_2_end_mask_0, squeeze_mask = new_cache_43_internal_tensor_assign_2_squeeze_mask_0, stride = new_cache_43_internal_tensor_assign_2_stride_0, x = reshape_215)[name = tensor("slice_by_index_43")]; + tensor reshape_216_shape_0 = const()[name = tensor("reshape_216_shape_0"), val = tensor([-1])]; + tensor reshape_216 = reshape(shape = reshape_216_shape_0, x = slice_by_index_43)[name = tensor("reshape_216")]; + tensor reshape_217_shape_0 = const()[name = tensor("reshape_217_shape_0"), val = tensor([-1])]; + tensor reshape_217_cast_fp16 = reshape(shape = reshape_217_shape_0, x = var_8551_cast_fp16)[name = tensor("reshape_217_cast_fp16")]; + tensor reshape_218_shape_0 = const()[name = tensor("reshape_218_shape_0"), val = tensor([-1])]; + tensor reshape_218_cast_fp16 = reshape(shape = reshape_218_shape_0, x = reshape_214_cast_fp16)[name = tensor("reshape_218_cast_fp16")]; + tensor scatter_43_mode_0 = const()[name = tensor("scatter_43_mode_0"), val = tensor("update")]; + tensor scatter_43_axis_0 = const()[name = tensor("scatter_43_axis_0"), val = tensor(0)]; + tensor scatter_43_validate_indices_0 = const()[name = tensor("scatter_43_validate_indices_0"), val = tensor(false)]; + tensor scatter_43_cast_fp16 = scatter(axis = scatter_43_axis_0, data = reshape_218_cast_fp16, indices = reshape_216, mode = scatter_43_mode_0, updates = reshape_217_cast_fp16, validate_indices = scatter_43_validate_indices_0)[name = tensor("scatter_43_cast_fp16")]; + tensor reshape_219_cast_fp16 = reshape(shape = shape_91, x = scatter_43_cast_fp16)[name = tensor("reshape_219_cast_fp16")]; + tensor reshape_219_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("reshape_219_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor keys_127_begin_0 = const()[name = tensor("keys_127_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor keys_127_end_0 = const()[name = tensor("keys_127_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor keys_127_end_mask_0 = const()[name = tensor("keys_127_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor keys_127_squeeze_mask_0 = const()[name = tensor("keys_127_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor keys_127_cast_fp16 = slice_by_index(begin = keys_127_begin_0, end = keys_127_end_0, end_mask = keys_127_end_mask_0, squeeze_mask = keys_127_squeeze_mask_0, x = reshape_219_cast_fp16)[name = tensor("keys_127_cast_fp16")]; + tensor values_127_begin_0 = const()[name = tensor("values_127_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor values_127_end_0 = const()[name = tensor("values_127_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor values_127_end_mask_0 = const()[name = tensor("values_127_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor values_127_squeeze_mask_0 = const()[name = tensor("values_127_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor values_127_cast_fp16 = slice_by_index(begin = values_127_begin_0, end = values_127_end_0, end_mask = values_127_end_mask_0, squeeze_mask = values_127_squeeze_mask_0, x = reshape_219_cast_fp16)[name = tensor("values_127_cast_fp16")]; + tensor var_8563_cast_fp16 = not_equal(x = keys_127_cast_fp16, y = keys_127_cast_fp16)[name = tensor("op_8563_cast_fp16")]; + tensor keys_129_cast_fp16 = select(a = var_463_to_fp16, b = keys_127_cast_fp16, cond = var_8563_cast_fp16)[name = tensor("keys_129_cast_fp16")]; + tensor var_8571_cast_fp16 = not_equal(x = values_127_cast_fp16, y = values_127_cast_fp16)[name = tensor("op_8571_cast_fp16")]; + tensor values_129_cast_fp16 = select(a = var_463_to_fp16, b = values_127_cast_fp16, cond = var_8571_cast_fp16)[name = tensor("values_129_cast_fp16")]; + tensor var_8595 = const()[name = tensor("op_8595"), val = tensor([0, 2, 1, 3])]; + tensor var_8611 = const()[name = tensor("op_8611"), val = tensor([1, 1, 1])]; + tensor var_8612_cast_fp16 = reshape(shape = var_8611, x = position21_to_fp16)[name = tensor("op_8612_cast_fp16")]; + tensor var_8632_to_fp16 = const()[name = tensor("op_8632_to_fp16"), val = tensor(0x1p+0)]; + tensor valid_len_43_cast_fp16 = add(x = var_8612_cast_fp16, y = var_8632_to_fp16)[name = tensor("valid_len_43_cast_fp16")]; + tensor valid_mask_43_cast_fp16 = less(x = k_positions_1_promoted_to_fp16, y = valid_len_43_cast_fp16)[name = tensor("valid_mask_43_cast_fp16")]; + tensor causal_mask_43_cast_fp16 = less_equal(x = k_positions_1_promoted_to_fp16, y = var_8612_cast_fp16)[name = tensor("causal_mask_43_cast_fp16")]; + tensor attn_mask_85 = logical_and(x = valid_mask_43_cast_fp16, y = causal_mask_43_cast_fp16)[name = tensor("attn_mask_85")]; + tensor attn_mask_87_axes_0 = const()[name = tensor("attn_mask_87_axes_0"), val = tensor([1])]; + tensor attn_mask_87 = expand_dims(axes = attn_mask_87_axes_0, x = attn_mask_85)[name = tensor("attn_mask_87")]; + tensor var_8650_transpose_x_0 = const()[name = tensor("op_8650_transpose_x_0"), val = tensor(false)]; + tensor var_8650_transpose_y_0 = const()[name = tensor("op_8650_transpose_y_0"), val = tensor(false)]; + tensor transpose_114_perm_0 = const()[name = tensor("transpose_114_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_115_perm_0 = const()[name = tensor("transpose_115_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_115 = transpose(perm = transpose_115_perm_0, x = keys_129_cast_fp16)[name = tensor("transpose_129")]; + tensor transpose_114 = transpose(perm = transpose_114_perm_0, x = q_129_cast_fp16)[name = tensor("transpose_130")]; + tensor var_8650_cast_fp16 = matmul(transpose_x = var_8650_transpose_x_0, transpose_y = var_8650_transpose_y_0, x = transpose_114, y = transpose_115)[name = tensor("op_8650_cast_fp16")]; + tensor var_8644_to_fp16 = const()[name = tensor("op_8644_to_fp16"), val = tensor([0x1p-3])]; + tensor attn_weights_127_cast_fp16 = mul(x = var_8650_cast_fp16, y = var_8644_to_fp16)[name = tensor("attn_weights_127_cast_fp16")]; + tensor var_8652 = logical_not(x = attn_mask_87)[name = tensor("op_8652")]; + tensor var_8653_to_fp16 = const()[name = tensor("op_8653_to_fp16"), val = tensor(-inf)]; + tensor attn_weights_129_cast_fp16 = select(a = var_8653_to_fp16, b = attn_weights_127_cast_fp16, cond = var_8652)[name = tensor("attn_weights_129_cast_fp16")]; + tensor var_8655 = const()[name = tensor("op_8655"), val = tensor(-1)]; + tensor attn_weights_131_cast_fp16 = softmax(axis = var_8655, x = attn_weights_129_cast_fp16)[name = tensor("attn_weights_131_cast_fp16")]; + tensor attn_output_43_transpose_x_0 = const()[name = tensor("attn_output_43_transpose_x_0"), val = tensor(false)]; + tensor attn_output_43_transpose_y_0 = const()[name = tensor("attn_output_43_transpose_y_0"), val = tensor(false)]; + tensor values_131_cast_fp16 = transpose(perm = var_8595, x = values_129_cast_fp16)[name = tensor("transpose_131")]; + tensor attn_output_43_cast_fp16 = matmul(transpose_x = attn_output_43_transpose_x_0, transpose_y = attn_output_43_transpose_y_0, x = attn_weights_131_cast_fp16, y = values_131_cast_fp16)[name = tensor("attn_output_43_cast_fp16")]; + tensor var_8663 = const()[name = tensor("op_8663"), val = tensor([0, 2, 1, 3])]; + tensor var_8668 = const()[name = tensor("op_8668"), val = tensor([1, 1, 1024])]; + tensor var_8664_cast_fp16 = transpose(perm = var_8663, x = attn_output_43_cast_fp16)[name = tensor("transpose_128")]; + tensor input_259_cast_fp16 = reshape(shape = var_8668, x = var_8664_cast_fp16)[name = tensor("input_259_cast_fp16")]; + tensor attn21_out_proj_weight_to_fp16 = const()[name = tensor("attn21_out_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(536095616)))]; + tensor linear_86_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = attn21_out_proj_weight_to_fp16, x = input_259_cast_fp16)[name = tensor("linear_86_cast_fp16")]; + tensor var_8674_to_fp16 = const()[name = tensor("op_8674_to_fp16"), val = tensor(0x1p+0)]; + tensor var_8675_cast_fp16 = add(x = position21_to_fp16, y = var_8674_to_fp16)[name = tensor("op_8675_cast_fp16")]; + tensor var_8675_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_8675_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor input_261_cast_fp16 = add(x = input_255_cast_fp16, y = linear_86_cast_fp16)[name = tensor("input_261_cast_fp16")]; + tensor input_263_axes_0 = const()[name = tensor("input_263_axes_0"), val = tensor([-1])]; + tensor norm21_2_weight_to_fp16 = const()[name = tensor("norm21_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(538192832)))]; + tensor norm21_2_bias_to_fp16 = const()[name = tensor("norm21_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(538194944)))]; + tensor var_8679_to_fp16 = const()[name = tensor("op_8679_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_263_cast_fp16 = layer_norm(axes = input_263_axes_0, beta = norm21_2_bias_to_fp16, epsilon = var_8679_to_fp16, gamma = norm21_2_weight_to_fp16, x = input_261_cast_fp16)[name = tensor("input_263_cast_fp16")]; + tensor linear21_1_weight_to_fp16 = const()[name = tensor("linear21_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(538197056)))]; + tensor linear_87_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = linear21_1_weight_to_fp16, x = input_263_cast_fp16)[name = tensor("linear_87_cast_fp16")]; + tensor input_265_mode_0 = const()[name = tensor("input_265_mode_0"), val = tensor("EXACT")]; + tensor input_265_cast_fp16 = gelu(mode = input_265_mode_0, x = linear_87_cast_fp16)[name = tensor("input_265_cast_fp16")]; + tensor linear21_2_weight_to_fp16 = const()[name = tensor("linear21_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(546585728)))]; + tensor linear_88_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = linear21_2_weight_to_fp16, x = input_265_cast_fp16)[name = tensor("linear_88_cast_fp16")]; + tensor input_267_cast_fp16 = add(x = input_261_cast_fp16, y = linear_88_cast_fp16)[name = tensor("input_267_cast_fp16")]; + tensor input_269_axes_0 = const()[name = tensor("input_269_axes_0"), val = tensor([-1])]; + tensor norm22_1_weight_to_fp16 = const()[name = tensor("norm22_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(554974400)))]; + tensor norm22_1_bias_to_fp16 = const()[name = tensor("norm22_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(554976512)))]; + tensor var_8696_to_fp16 = const()[name = tensor("op_8696_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_269_cast_fp16 = layer_norm(axes = input_269_axes_0, beta = norm22_1_bias_to_fp16, epsilon = var_8696_to_fp16, gamma = norm22_1_weight_to_fp16, x = input_267_cast_fp16)[name = tensor("input_269_cast_fp16")]; + tensor attn22_in_proj_weight_to_fp16 = const()[name = tensor("attn22_in_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(554978624)))]; + tensor linear_89_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = attn22_in_proj_weight_to_fp16, x = input_269_cast_fp16)[name = tensor("linear_89_cast_fp16")]; + tensor var_8710 = const()[name = tensor("op_8710"), val = tensor([1, 1, 3, 16, 64])]; + tensor qkv_45_cast_fp16 = reshape(shape = var_8710, x = linear_89_cast_fp16)[name = tensor("qkv_45_cast_fp16")]; + tensor q_133_begin_0 = const()[name = tensor("q_133_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor q_133_end_0 = const()[name = tensor("q_133_end_0"), val = tensor([1, 1, 1, 16, 64])]; + tensor q_133_end_mask_0 = const()[name = tensor("q_133_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor q_133_squeeze_mask_0 = const()[name = tensor("q_133_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor q_133_cast_fp16 = slice_by_index(begin = q_133_begin_0, end = q_133_end_0, end_mask = q_133_end_mask_0, squeeze_mask = q_133_squeeze_mask_0, x = qkv_45_cast_fp16)[name = tensor("q_133_cast_fp16")]; + tensor k_89_begin_0 = const()[name = tensor("k_89_begin_0"), val = tensor([0, 0, 1, 0, 0])]; + tensor k_89_end_0 = const()[name = tensor("k_89_end_0"), val = tensor([1, 1, 2, 16, 64])]; + tensor k_89_end_mask_0 = const()[name = tensor("k_89_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor k_89_squeeze_mask_0 = const()[name = tensor("k_89_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor k_89_cast_fp16 = slice_by_index(begin = k_89_begin_0, end = k_89_end_0, end_mask = k_89_end_mask_0, squeeze_mask = k_89_squeeze_mask_0, x = qkv_45_cast_fp16)[name = tensor("k_89_cast_fp16")]; + tensor v_45_begin_0 = const()[name = tensor("v_45_begin_0"), val = tensor([0, 0, 2, 0, 0])]; + tensor v_45_end_0 = const()[name = tensor("v_45_end_0"), val = tensor([1, 1, 3, 16, 64])]; + tensor v_45_end_mask_0 = const()[name = tensor("v_45_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor v_45_squeeze_mask_0 = const()[name = tensor("v_45_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor v_45_cast_fp16 = slice_by_index(begin = v_45_begin_0, end = v_45_end_0, end_mask = v_45_end_mask_0, squeeze_mask = v_45_squeeze_mask_0, x = qkv_45_cast_fp16)[name = tensor("v_45_cast_fp16")]; + tensor var_8776 = const()[name = tensor("op_8776"), val = tensor([1, 1, 1, 1])]; + tensor position22_to_fp16_dtype_0 = const()[name = tensor("position22_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor position22_to_fp16 = cast(dtype = position22_to_fp16_dtype_0, x = position22)[name = tensor("cast_443")]; + tensor ts_137_cast_fp16 = reshape(shape = var_8776, x = position22_to_fp16)[name = tensor("ts_137_cast_fp16")]; + tensor var_8783 = const()[name = tensor("op_8783"), val = tensor([1, 1, 16, 32, 2])]; + tensor q_complex_45_cast_fp16 = reshape(shape = var_8783, x = q_133_cast_fp16)[name = tensor("q_complex_45_cast_fp16")]; + tensor var_8790 = const()[name = tensor("op_8790"), val = tensor([1, 1, 16, 32, 2])]; + tensor k_complex_45_cast_fp16 = reshape(shape = var_8790, x = k_89_cast_fp16)[name = tensor("k_complex_45_cast_fp16")]; + tensor var_8794_begin_0 = const()[name = tensor("op_8794_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_8794_end_0 = const()[name = tensor("op_8794_end_0"), val = tensor([1, 1, 16, 32, 1])]; + tensor var_8794_end_mask_0 = const()[name = tensor("op_8794_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_8794_squeeze_mask_0 = const()[name = tensor("op_8794_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_8794_cast_fp16 = slice_by_index(begin = var_8794_begin_0, end = var_8794_end_0, end_mask = var_8794_end_mask_0, squeeze_mask = var_8794_squeeze_mask_0, x = q_complex_45_cast_fp16)[name = tensor("op_8794_cast_fp16")]; + tensor var_8802_begin_0 = const()[name = tensor("op_8802_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_8802_end_0 = const()[name = tensor("op_8802_end_0"), val = tensor([1, 1, 16, 32, 2])]; + tensor var_8802_end_mask_0 = const()[name = tensor("op_8802_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_8802_squeeze_mask_0 = const()[name = tensor("op_8802_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_8802_cast_fp16 = slice_by_index(begin = var_8802_begin_0, end = var_8802_end_0, end_mask = var_8802_end_mask_0, squeeze_mask = var_8802_squeeze_mask_0, x = q_complex_45_cast_fp16)[name = tensor("op_8802_cast_fp16")]; + tensor var_8810_begin_0 = const()[name = tensor("op_8810_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_8810_end_0 = const()[name = tensor("op_8810_end_0"), val = tensor([1, 1, 16, 32, 1])]; + tensor var_8810_end_mask_0 = const()[name = tensor("op_8810_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_8810_squeeze_mask_0 = const()[name = tensor("op_8810_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_8810_cast_fp16 = slice_by_index(begin = var_8810_begin_0, end = var_8810_end_0, end_mask = var_8810_end_mask_0, squeeze_mask = var_8810_squeeze_mask_0, x = k_complex_45_cast_fp16)[name = tensor("op_8810_cast_fp16")]; + tensor var_8818_begin_0 = const()[name = tensor("op_8818_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_8818_end_0 = const()[name = tensor("op_8818_end_0"), val = tensor([1, 1, 16, 32, 2])]; + tensor var_8818_end_mask_0 = const()[name = tensor("op_8818_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_8818_squeeze_mask_0 = const()[name = tensor("op_8818_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_8818_cast_fp16 = slice_by_index(begin = var_8818_begin_0, end = var_8818_end_0, end_mask = var_8818_end_mask_0, squeeze_mask = var_8818_squeeze_mask_0, x = k_complex_45_cast_fp16)[name = tensor("op_8818_cast_fp16")]; + tensor freqs_45_to_fp16 = const()[name = tensor("freqs_45_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(561270144)))]; + tensor var_8824_cast_fp16 = mul(x = freqs_45_to_fp16, y = ts_137_cast_fp16)[name = tensor("op_8824_cast_fp16")]; + tensor rotr_45_cast_fp16 = cos(x = var_8824_cast_fp16)[name = tensor("rotr_45_cast_fp16")]; + tensor roti_45_cast_fp16 = sin(x = var_8824_cast_fp16)[name = tensor("roti_45_cast_fp16")]; + tensor var_8828_cast_fp16 = mul(x = var_8794_cast_fp16, y = rotr_45_cast_fp16)[name = tensor("op_8828_cast_fp16")]; + tensor var_8829_cast_fp16 = mul(x = var_8802_cast_fp16, y = roti_45_cast_fp16)[name = tensor("op_8829_cast_fp16")]; + tensor qor_89_cast_fp16 = sub(x = var_8828_cast_fp16, y = var_8829_cast_fp16)[name = tensor("qor_89_cast_fp16")]; + tensor var_8832_cast_fp16 = mul(x = var_8794_cast_fp16, y = roti_45_cast_fp16)[name = tensor("op_8832_cast_fp16")]; + tensor var_8833_cast_fp16 = mul(x = var_8802_cast_fp16, y = rotr_45_cast_fp16)[name = tensor("op_8833_cast_fp16")]; + tensor qoi_89_cast_fp16 = add(x = var_8832_cast_fp16, y = var_8833_cast_fp16)[name = tensor("qoi_89_cast_fp16")]; + tensor var_8836_cast_fp16 = mul(x = var_8810_cast_fp16, y = rotr_45_cast_fp16)[name = tensor("op_8836_cast_fp16")]; + tensor var_8837_cast_fp16 = mul(x = var_8818_cast_fp16, y = roti_45_cast_fp16)[name = tensor("op_8837_cast_fp16")]; + tensor kor_89_cast_fp16 = sub(x = var_8836_cast_fp16, y = var_8837_cast_fp16)[name = tensor("kor_89_cast_fp16")]; + tensor var_8840_cast_fp16 = mul(x = var_8810_cast_fp16, y = roti_45_cast_fp16)[name = tensor("op_8840_cast_fp16")]; + tensor var_8841_cast_fp16 = mul(x = var_8818_cast_fp16, y = rotr_45_cast_fp16)[name = tensor("op_8841_cast_fp16")]; + tensor koi_89_cast_fp16 = add(x = var_8840_cast_fp16, y = var_8841_cast_fp16)[name = tensor("koi_89_cast_fp16")]; + tensor qo_45_axis_0 = const()[name = tensor("qo_45_axis_0"), val = tensor(-1)]; + tensor qo_45_cast_fp16 = stack(axis = qo_45_axis_0, values = (qor_89_cast_fp16, qoi_89_cast_fp16))[name = tensor("qo_45_cast_fp16")]; + tensor ko_45_axis_0 = const()[name = tensor("ko_45_axis_0"), val = tensor(-1)]; + tensor ko_45_cast_fp16 = stack(axis = ko_45_axis_0, values = (kor_89_cast_fp16, koi_89_cast_fp16))[name = tensor("ko_45_cast_fp16")]; + tensor var_8874 = const()[name = tensor("op_8874"), val = tensor([1, 1, 16, 64])]; + tensor q_135_cast_fp16 = reshape(shape = var_8874, x = qo_45_cast_fp16)[name = tensor("q_135_cast_fp16")]; + tensor var_8880 = const()[name = tensor("op_8880"), val = tensor([1, 1, 16, 64])]; + tensor k_91_cast_fp16 = reshape(shape = var_8880, x = ko_45_cast_fp16)[name = tensor("k_91_cast_fp16")]; + tensor _inversed_8905_y_0_to_fp16 = const()[name = tensor("_inversed_8905_y_0_to_fp16"), val = tensor(0x1p-9)]; + tensor _inversed_8905_cast_fp16 = mul(x = ts_137_cast_fp16, y = _inversed_8905_y_0_to_fp16)[name = tensor("_inversed_8905_cast_fp16")]; + tensor var_8906_cast_fp16 = floor(x = _inversed_8905_cast_fp16)[name = tensor("op_8906_cast_fp16")]; + tensor var_8907_to_fp16 = const()[name = tensor("op_8907_to_fp16"), val = tensor(0x1p+9)]; + tensor var_8908_cast_fp16 = mul(x = var_8906_cast_fp16, y = var_8907_to_fp16)[name = tensor("op_8908_cast_fp16")]; + tensor write_indices_float_91_cast_fp16 = sub(x = ts_137_cast_fp16, y = var_8908_cast_fp16)[name = tensor("write_indices_float_91_cast_fp16")]; + tensor var_8915_dtype_0 = const()[name = tensor("op_8915_dtype_0"), val = tensor("int32")]; + tensor write_indices_45_reps_0 = const()[name = tensor("write_indices_45_reps_0"), val = tensor([1, 1, 16, 64])]; + tensor write_indices_float_91_cast_fp16_to_int32 = cast(dtype = var_8915_dtype_0, x = write_indices_float_91_cast_fp16)[name = tensor("cast_442")]; + tensor write_indices_45 = tile(reps = write_indices_45_reps_0, x = write_indices_float_91_cast_fp16_to_int32)[name = tensor("write_indices_45")]; + tensor var_8925_begin_0 = const()[name = tensor("op_8925_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_8925_end_0 = const()[name = tensor("op_8925_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor var_8925_end_mask_0 = const()[name = tensor("op_8925_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_8925_squeeze_mask_0 = const()[name = tensor("op_8925_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor cache22_to_fp16_dtype_0 = const()[name = tensor("cache22_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor cache22_to_fp16 = cast(dtype = cache22_to_fp16_dtype_0, x = cache22)[name = tensor("cast_441")]; + tensor var_8925_cast_fp16 = slice_by_index(begin = var_8925_begin_0, end = var_8925_end_0, end_mask = var_8925_end_mask_0, squeeze_mask = var_8925_squeeze_mask_0, x = cache22_to_fp16)[name = tensor("op_8925_cast_fp16")]; + tensor var_8927_axis_0 = const()[name = tensor("op_8927_axis_0"), val = tensor(1)]; + tensor var_8927_mode_0 = const()[name = tensor("op_8927_mode_0"), val = tensor("update")]; + tensor var_8927_validate_indices_0 = const()[name = tensor("op_8927_validate_indices_0"), val = tensor(false)]; + tensor var_8927_cast_fp16 = scatter_along_axis(axis = var_8927_axis_0, data = var_8925_cast_fp16, indices = write_indices_45, mode = var_8927_mode_0, updates = k_91_cast_fp16, validate_indices = var_8927_validate_indices_0)[name = tensor("op_8927_cast_fp16")]; + tensor concat_156 = const()[name = tensor("concat_156"), val = tensor([0, 0, 0, 0, 0])]; + tensor concat_157 = const()[name = tensor("concat_157"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_45_internal_tensor_assign_1_stride_0 = const()[name = tensor("new_cache_45_internal_tensor_assign_1_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_45_internal_tensor_assign_1_begin_mask_0 = const()[name = tensor("new_cache_45_internal_tensor_assign_1_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_45_internal_tensor_assign_1_end_mask_0 = const()[name = tensor("new_cache_45_internal_tensor_assign_1_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_45_internal_tensor_assign_1_squeeze_mask_0 = const()[name = tensor("new_cache_45_internal_tensor_assign_1_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_92 = const()[name = tensor("shape_92"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_44 = const()[name = tensor("reduce_prod_44"), val = tensor(1048576)]; + tensor range_1d_44_start_0 = const()[name = tensor("range_1d_44_start_0"), val = tensor(0)]; + tensor range_1d_44_step_0 = const()[name = tensor("range_1d_44_step_0"), val = tensor(1)]; + tensor range_1d_44 = range_1d(end = reduce_prod_44, start = range_1d_44_start_0, step = range_1d_44_step_0)[name = tensor("range_1d_44")]; + tensor reshape_220 = reshape(shape = shape_92, x = range_1d_44)[name = tensor("reshape_220")]; + tensor slice_by_index_44 = slice_by_index(begin = concat_156, begin_mask = new_cache_45_internal_tensor_assign_1_begin_mask_0, end = concat_157, end_mask = new_cache_45_internal_tensor_assign_1_end_mask_0, squeeze_mask = new_cache_45_internal_tensor_assign_1_squeeze_mask_0, stride = new_cache_45_internal_tensor_assign_1_stride_0, x = reshape_220)[name = tensor("slice_by_index_44")]; + tensor reshape_221_shape_0 = const()[name = tensor("reshape_221_shape_0"), val = tensor([-1])]; + tensor reshape_221 = reshape(shape = reshape_221_shape_0, x = slice_by_index_44)[name = tensor("reshape_221")]; + tensor reshape_222_shape_0 = const()[name = tensor("reshape_222_shape_0"), val = tensor([-1])]; + tensor reshape_222_cast_fp16 = reshape(shape = reshape_222_shape_0, x = var_8927_cast_fp16)[name = tensor("reshape_222_cast_fp16")]; + tensor reshape_223_shape_0 = const()[name = tensor("reshape_223_shape_0"), val = tensor([-1])]; + tensor reshape_223_cast_fp16 = reshape(shape = reshape_223_shape_0, x = cache22_to_fp16)[name = tensor("reshape_223_cast_fp16")]; + tensor scatter_44_mode_0 = const()[name = tensor("scatter_44_mode_0"), val = tensor("update")]; + tensor scatter_44_axis_0 = const()[name = tensor("scatter_44_axis_0"), val = tensor(0)]; + tensor scatter_44_validate_indices_0 = const()[name = tensor("scatter_44_validate_indices_0"), val = tensor(false)]; + tensor scatter_44_cast_fp16 = scatter(axis = scatter_44_axis_0, data = reshape_223_cast_fp16, indices = reshape_221, mode = scatter_44_mode_0, updates = reshape_222_cast_fp16, validate_indices = scatter_44_validate_indices_0)[name = tensor("scatter_44_cast_fp16")]; + tensor reshape_224_cast_fp16 = reshape(shape = shape_92, x = scatter_44_cast_fp16)[name = tensor("reshape_224_cast_fp16")]; + tensor var_8935_begin_0 = const()[name = tensor("op_8935_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor var_8935_end_0 = const()[name = tensor("op_8935_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor var_8935_end_mask_0 = const()[name = tensor("op_8935_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_8935_squeeze_mask_0 = const()[name = tensor("op_8935_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor var_8935_cast_fp16 = slice_by_index(begin = var_8935_begin_0, end = var_8935_end_0, end_mask = var_8935_end_mask_0, squeeze_mask = var_8935_squeeze_mask_0, x = reshape_224_cast_fp16)[name = tensor("op_8935_cast_fp16")]; + tensor var_8937_axis_0 = const()[name = tensor("op_8937_axis_0"), val = tensor(1)]; + tensor var_8937_mode_0 = const()[name = tensor("op_8937_mode_0"), val = tensor("update")]; + tensor var_8937_validate_indices_0 = const()[name = tensor("op_8937_validate_indices_0"), val = tensor(false)]; + tensor var_8937_cast_fp16 = scatter_along_axis(axis = var_8937_axis_0, data = var_8935_cast_fp16, indices = write_indices_45, mode = var_8937_mode_0, updates = v_45_cast_fp16, validate_indices = var_8937_validate_indices_0)[name = tensor("op_8937_cast_fp16")]; + tensor concat_158 = const()[name = tensor("concat_158"), val = tensor([1, 0, 0, 0, 0])]; + tensor concat_159 = const()[name = tensor("concat_159"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_45_internal_tensor_assign_2_stride_0 = const()[name = tensor("new_cache_45_internal_tensor_assign_2_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_45_internal_tensor_assign_2_begin_mask_0 = const()[name = tensor("new_cache_45_internal_tensor_assign_2_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_45_internal_tensor_assign_2_end_mask_0 = const()[name = tensor("new_cache_45_internal_tensor_assign_2_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_45_internal_tensor_assign_2_squeeze_mask_0 = const()[name = tensor("new_cache_45_internal_tensor_assign_2_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_93 = const()[name = tensor("shape_93"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_45 = const()[name = tensor("reduce_prod_45"), val = tensor(1048576)]; + tensor range_1d_45_start_0 = const()[name = tensor("range_1d_45_start_0"), val = tensor(0)]; + tensor range_1d_45_step_0 = const()[name = tensor("range_1d_45_step_0"), val = tensor(1)]; + tensor range_1d_45 = range_1d(end = reduce_prod_45, start = range_1d_45_start_0, step = range_1d_45_step_0)[name = tensor("range_1d_45")]; + tensor reshape_225 = reshape(shape = shape_93, x = range_1d_45)[name = tensor("reshape_225")]; + tensor slice_by_index_45 = slice_by_index(begin = concat_158, begin_mask = new_cache_45_internal_tensor_assign_2_begin_mask_0, end = concat_159, end_mask = new_cache_45_internal_tensor_assign_2_end_mask_0, squeeze_mask = new_cache_45_internal_tensor_assign_2_squeeze_mask_0, stride = new_cache_45_internal_tensor_assign_2_stride_0, x = reshape_225)[name = tensor("slice_by_index_45")]; + tensor reshape_226_shape_0 = const()[name = tensor("reshape_226_shape_0"), val = tensor([-1])]; + tensor reshape_226 = reshape(shape = reshape_226_shape_0, x = slice_by_index_45)[name = tensor("reshape_226")]; + tensor reshape_227_shape_0 = const()[name = tensor("reshape_227_shape_0"), val = tensor([-1])]; + tensor reshape_227_cast_fp16 = reshape(shape = reshape_227_shape_0, x = var_8937_cast_fp16)[name = tensor("reshape_227_cast_fp16")]; + tensor reshape_228_shape_0 = const()[name = tensor("reshape_228_shape_0"), val = tensor([-1])]; + tensor reshape_228_cast_fp16 = reshape(shape = reshape_228_shape_0, x = reshape_224_cast_fp16)[name = tensor("reshape_228_cast_fp16")]; + tensor scatter_45_mode_0 = const()[name = tensor("scatter_45_mode_0"), val = tensor("update")]; + tensor scatter_45_axis_0 = const()[name = tensor("scatter_45_axis_0"), val = tensor(0)]; + tensor scatter_45_validate_indices_0 = const()[name = tensor("scatter_45_validate_indices_0"), val = tensor(false)]; + tensor scatter_45_cast_fp16 = scatter(axis = scatter_45_axis_0, data = reshape_228_cast_fp16, indices = reshape_226, mode = scatter_45_mode_0, updates = reshape_227_cast_fp16, validate_indices = scatter_45_validate_indices_0)[name = tensor("scatter_45_cast_fp16")]; + tensor reshape_229_cast_fp16 = reshape(shape = shape_93, x = scatter_45_cast_fp16)[name = tensor("reshape_229_cast_fp16")]; + tensor reshape_229_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("reshape_229_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor keys_133_begin_0 = const()[name = tensor("keys_133_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor keys_133_end_0 = const()[name = tensor("keys_133_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor keys_133_end_mask_0 = const()[name = tensor("keys_133_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor keys_133_squeeze_mask_0 = const()[name = tensor("keys_133_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor keys_133_cast_fp16 = slice_by_index(begin = keys_133_begin_0, end = keys_133_end_0, end_mask = keys_133_end_mask_0, squeeze_mask = keys_133_squeeze_mask_0, x = reshape_229_cast_fp16)[name = tensor("keys_133_cast_fp16")]; + tensor values_133_begin_0 = const()[name = tensor("values_133_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor values_133_end_0 = const()[name = tensor("values_133_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor values_133_end_mask_0 = const()[name = tensor("values_133_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor values_133_squeeze_mask_0 = const()[name = tensor("values_133_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor values_133_cast_fp16 = slice_by_index(begin = values_133_begin_0, end = values_133_end_0, end_mask = values_133_end_mask_0, squeeze_mask = values_133_squeeze_mask_0, x = reshape_229_cast_fp16)[name = tensor("values_133_cast_fp16")]; + tensor var_8949_cast_fp16 = not_equal(x = keys_133_cast_fp16, y = keys_133_cast_fp16)[name = tensor("op_8949_cast_fp16")]; + tensor keys_135_cast_fp16 = select(a = var_463_to_fp16, b = keys_133_cast_fp16, cond = var_8949_cast_fp16)[name = tensor("keys_135_cast_fp16")]; + tensor var_8957_cast_fp16 = not_equal(x = values_133_cast_fp16, y = values_133_cast_fp16)[name = tensor("op_8957_cast_fp16")]; + tensor values_135_cast_fp16 = select(a = var_463_to_fp16, b = values_133_cast_fp16, cond = var_8957_cast_fp16)[name = tensor("values_135_cast_fp16")]; + tensor var_8981 = const()[name = tensor("op_8981"), val = tensor([0, 2, 1, 3])]; + tensor var_8997 = const()[name = tensor("op_8997"), val = tensor([1, 1, 1])]; + tensor var_8998_cast_fp16 = reshape(shape = var_8997, x = position22_to_fp16)[name = tensor("op_8998_cast_fp16")]; + tensor var_9018_to_fp16 = const()[name = tensor("op_9018_to_fp16"), val = tensor(0x1p+0)]; + tensor valid_len_45_cast_fp16 = add(x = var_8998_cast_fp16, y = var_9018_to_fp16)[name = tensor("valid_len_45_cast_fp16")]; + tensor valid_mask_45_cast_fp16 = less(x = k_positions_1_promoted_to_fp16, y = valid_len_45_cast_fp16)[name = tensor("valid_mask_45_cast_fp16")]; + tensor causal_mask_45_cast_fp16 = less_equal(x = k_positions_1_promoted_to_fp16, y = var_8998_cast_fp16)[name = tensor("causal_mask_45_cast_fp16")]; + tensor attn_mask_89 = logical_and(x = valid_mask_45_cast_fp16, y = causal_mask_45_cast_fp16)[name = tensor("attn_mask_89")]; + tensor attn_mask_91_axes_0 = const()[name = tensor("attn_mask_91_axes_0"), val = tensor([1])]; + tensor attn_mask_91 = expand_dims(axes = attn_mask_91_axes_0, x = attn_mask_89)[name = tensor("attn_mask_91")]; + tensor var_9036_transpose_x_0 = const()[name = tensor("op_9036_transpose_x_0"), val = tensor(false)]; + tensor var_9036_transpose_y_0 = const()[name = tensor("op_9036_transpose_y_0"), val = tensor(false)]; + tensor transpose_116_perm_0 = const()[name = tensor("transpose_116_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_117_perm_0 = const()[name = tensor("transpose_117_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_117 = transpose(perm = transpose_117_perm_0, x = keys_135_cast_fp16)[name = tensor("transpose_125")]; + tensor transpose_116 = transpose(perm = transpose_116_perm_0, x = q_135_cast_fp16)[name = tensor("transpose_126")]; + tensor var_9036_cast_fp16 = matmul(transpose_x = var_9036_transpose_x_0, transpose_y = var_9036_transpose_y_0, x = transpose_116, y = transpose_117)[name = tensor("op_9036_cast_fp16")]; + tensor var_9030_to_fp16 = const()[name = tensor("op_9030_to_fp16"), val = tensor([0x1p-3])]; + tensor attn_weights_133_cast_fp16 = mul(x = var_9036_cast_fp16, y = var_9030_to_fp16)[name = tensor("attn_weights_133_cast_fp16")]; + tensor var_9038 = logical_not(x = attn_mask_91)[name = tensor("op_9038")]; + tensor var_9039_to_fp16 = const()[name = tensor("op_9039_to_fp16"), val = tensor(-inf)]; + tensor attn_weights_135_cast_fp16 = select(a = var_9039_to_fp16, b = attn_weights_133_cast_fp16, cond = var_9038)[name = tensor("attn_weights_135_cast_fp16")]; + tensor var_9041 = const()[name = tensor("op_9041"), val = tensor(-1)]; + tensor attn_weights_137_cast_fp16 = softmax(axis = var_9041, x = attn_weights_135_cast_fp16)[name = tensor("attn_weights_137_cast_fp16")]; + tensor attn_output_45_transpose_x_0 = const()[name = tensor("attn_output_45_transpose_x_0"), val = tensor(false)]; + tensor attn_output_45_transpose_y_0 = const()[name = tensor("attn_output_45_transpose_y_0"), val = tensor(false)]; + tensor values_137_cast_fp16 = transpose(perm = var_8981, x = values_135_cast_fp16)[name = tensor("transpose_127")]; + tensor attn_output_45_cast_fp16 = matmul(transpose_x = attn_output_45_transpose_x_0, transpose_y = attn_output_45_transpose_y_0, x = attn_weights_137_cast_fp16, y = values_137_cast_fp16)[name = tensor("attn_output_45_cast_fp16")]; + tensor var_9049 = const()[name = tensor("op_9049"), val = tensor([0, 2, 1, 3])]; + tensor var_9054 = const()[name = tensor("op_9054"), val = tensor([1, 1, 1024])]; + tensor var_9050_cast_fp16 = transpose(perm = var_9049, x = attn_output_45_cast_fp16)[name = tensor("transpose_124")]; + tensor input_271_cast_fp16 = reshape(shape = var_9054, x = var_9050_cast_fp16)[name = tensor("input_271_cast_fp16")]; + tensor attn22_out_proj_weight_to_fp16 = const()[name = tensor("attn22_out_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(561270272)))]; + tensor linear_90_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = attn22_out_proj_weight_to_fp16, x = input_271_cast_fp16)[name = tensor("linear_90_cast_fp16")]; + tensor var_9060_to_fp16 = const()[name = tensor("op_9060_to_fp16"), val = tensor(0x1p+0)]; + tensor var_9061_cast_fp16 = add(x = position22_to_fp16, y = var_9060_to_fp16)[name = tensor("op_9061_cast_fp16")]; + tensor var_9061_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_9061_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor input_273_cast_fp16 = add(x = input_267_cast_fp16, y = linear_90_cast_fp16)[name = tensor("input_273_cast_fp16")]; + tensor input_275_axes_0 = const()[name = tensor("input_275_axes_0"), val = tensor([-1])]; + tensor norm22_2_weight_to_fp16 = const()[name = tensor("norm22_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(563367488)))]; + tensor norm22_2_bias_to_fp16 = const()[name = tensor("norm22_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(563369600)))]; + tensor var_9065_to_fp16 = const()[name = tensor("op_9065_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_275_cast_fp16 = layer_norm(axes = input_275_axes_0, beta = norm22_2_bias_to_fp16, epsilon = var_9065_to_fp16, gamma = norm22_2_weight_to_fp16, x = input_273_cast_fp16)[name = tensor("input_275_cast_fp16")]; + tensor linear22_1_weight_to_fp16 = const()[name = tensor("linear22_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(563371712)))]; + tensor linear_91_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = linear22_1_weight_to_fp16, x = input_275_cast_fp16)[name = tensor("linear_91_cast_fp16")]; + tensor input_277_mode_0 = const()[name = tensor("input_277_mode_0"), val = tensor("EXACT")]; + tensor input_277_cast_fp16 = gelu(mode = input_277_mode_0, x = linear_91_cast_fp16)[name = tensor("input_277_cast_fp16")]; + tensor linear22_2_weight_to_fp16 = const()[name = tensor("linear22_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(571760384)))]; + tensor linear_92_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = linear22_2_weight_to_fp16, x = input_277_cast_fp16)[name = tensor("linear_92_cast_fp16")]; + tensor input_279_cast_fp16 = add(x = input_273_cast_fp16, y = linear_92_cast_fp16)[name = tensor("input_279_cast_fp16")]; + tensor input_281_axes_0 = const()[name = tensor("input_281_axes_0"), val = tensor([-1])]; + tensor norm23_1_weight_to_fp16 = const()[name = tensor("norm23_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(580149056)))]; + tensor norm23_1_bias_to_fp16 = const()[name = tensor("norm23_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(580151168)))]; + tensor var_9082_to_fp16 = const()[name = tensor("op_9082_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_281_cast_fp16 = layer_norm(axes = input_281_axes_0, beta = norm23_1_bias_to_fp16, epsilon = var_9082_to_fp16, gamma = norm23_1_weight_to_fp16, x = input_279_cast_fp16)[name = tensor("input_281_cast_fp16")]; + tensor attn23_in_proj_weight_to_fp16 = const()[name = tensor("attn23_in_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(580153280)))]; + tensor linear_93_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = attn23_in_proj_weight_to_fp16, x = input_281_cast_fp16)[name = tensor("linear_93_cast_fp16")]; + tensor var_9096 = const()[name = tensor("op_9096"), val = tensor([1, 1, 3, 16, 64])]; + tensor qkv_cast_fp16 = reshape(shape = var_9096, x = linear_93_cast_fp16)[name = tensor("qkv_cast_fp16")]; + tensor q_139_begin_0 = const()[name = tensor("q_139_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor q_139_end_0 = const()[name = tensor("q_139_end_0"), val = tensor([1, 1, 1, 16, 64])]; + tensor q_139_end_mask_0 = const()[name = tensor("q_139_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor q_139_squeeze_mask_0 = const()[name = tensor("q_139_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor q_139_cast_fp16 = slice_by_index(begin = q_139_begin_0, end = q_139_end_0, end_mask = q_139_end_mask_0, squeeze_mask = q_139_squeeze_mask_0, x = qkv_cast_fp16)[name = tensor("q_139_cast_fp16")]; + tensor k_93_begin_0 = const()[name = tensor("k_93_begin_0"), val = tensor([0, 0, 1, 0, 0])]; + tensor k_93_end_0 = const()[name = tensor("k_93_end_0"), val = tensor([1, 1, 2, 16, 64])]; + tensor k_93_end_mask_0 = const()[name = tensor("k_93_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor k_93_squeeze_mask_0 = const()[name = tensor("k_93_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor k_93_cast_fp16 = slice_by_index(begin = k_93_begin_0, end = k_93_end_0, end_mask = k_93_end_mask_0, squeeze_mask = k_93_squeeze_mask_0, x = qkv_cast_fp16)[name = tensor("k_93_cast_fp16")]; + tensor v_begin_0 = const()[name = tensor("v_begin_0"), val = tensor([0, 0, 2, 0, 0])]; + tensor v_end_0 = const()[name = tensor("v_end_0"), val = tensor([1, 1, 3, 16, 64])]; + tensor v_end_mask_0 = const()[name = tensor("v_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor v_squeeze_mask_0 = const()[name = tensor("v_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor v_cast_fp16 = slice_by_index(begin = v_begin_0, end = v_end_0, end_mask = v_end_mask_0, squeeze_mask = v_squeeze_mask_0, x = qkv_cast_fp16)[name = tensor("v_cast_fp16")]; + tensor var_9162 = const()[name = tensor("op_9162"), val = tensor([1, 1, 1, 1])]; + tensor position23_to_fp16_dtype_0 = const()[name = tensor("position23_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor position23_to_fp16 = cast(dtype = position23_to_fp16_dtype_0, x = position23)[name = tensor("cast_438")]; + tensor ts_cast_fp16 = reshape(shape = var_9162, x = position23_to_fp16)[name = tensor("ts_cast_fp16")]; + tensor var_9169 = const()[name = tensor("op_9169"), val = tensor([1, 1, 16, 32, 2])]; + tensor q_complex_cast_fp16 = reshape(shape = var_9169, x = q_139_cast_fp16)[name = tensor("q_complex_cast_fp16")]; + tensor var_9176 = const()[name = tensor("op_9176"), val = tensor([1, 1, 16, 32, 2])]; + tensor k_complex_cast_fp16 = reshape(shape = var_9176, x = k_93_cast_fp16)[name = tensor("k_complex_cast_fp16")]; + tensor var_9180_begin_0 = const()[name = tensor("op_9180_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_9180_end_0 = const()[name = tensor("op_9180_end_0"), val = tensor([1, 1, 16, 32, 1])]; + tensor var_9180_end_mask_0 = const()[name = tensor("op_9180_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_9180_squeeze_mask_0 = const()[name = tensor("op_9180_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_9180_cast_fp16 = slice_by_index(begin = var_9180_begin_0, end = var_9180_end_0, end_mask = var_9180_end_mask_0, squeeze_mask = var_9180_squeeze_mask_0, x = q_complex_cast_fp16)[name = tensor("op_9180_cast_fp16")]; + tensor var_9188_begin_0 = const()[name = tensor("op_9188_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_9188_end_0 = const()[name = tensor("op_9188_end_0"), val = tensor([1, 1, 16, 32, 2])]; + tensor var_9188_end_mask_0 = const()[name = tensor("op_9188_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_9188_squeeze_mask_0 = const()[name = tensor("op_9188_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_9188_cast_fp16 = slice_by_index(begin = var_9188_begin_0, end = var_9188_end_0, end_mask = var_9188_end_mask_0, squeeze_mask = var_9188_squeeze_mask_0, x = q_complex_cast_fp16)[name = tensor("op_9188_cast_fp16")]; + tensor var_9196_begin_0 = const()[name = tensor("op_9196_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_9196_end_0 = const()[name = tensor("op_9196_end_0"), val = tensor([1, 1, 16, 32, 1])]; + tensor var_9196_end_mask_0 = const()[name = tensor("op_9196_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_9196_squeeze_mask_0 = const()[name = tensor("op_9196_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_9196_cast_fp16 = slice_by_index(begin = var_9196_begin_0, end = var_9196_end_0, end_mask = var_9196_end_mask_0, squeeze_mask = var_9196_squeeze_mask_0, x = k_complex_cast_fp16)[name = tensor("op_9196_cast_fp16")]; + tensor var_9204_begin_0 = const()[name = tensor("op_9204_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_9204_end_0 = const()[name = tensor("op_9204_end_0"), val = tensor([1, 1, 16, 32, 2])]; + tensor var_9204_end_mask_0 = const()[name = tensor("op_9204_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_9204_squeeze_mask_0 = const()[name = tensor("op_9204_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_9204_cast_fp16 = slice_by_index(begin = var_9204_begin_0, end = var_9204_end_0, end_mask = var_9204_end_mask_0, squeeze_mask = var_9204_squeeze_mask_0, x = k_complex_cast_fp16)[name = tensor("op_9204_cast_fp16")]; + tensor freqs_to_fp16 = const()[name = tensor("freqs_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(586444800)))]; + tensor var_9210_cast_fp16 = mul(x = freqs_to_fp16, y = ts_cast_fp16)[name = tensor("op_9210_cast_fp16")]; + tensor rotr_cast_fp16 = cos(x = var_9210_cast_fp16)[name = tensor("rotr_cast_fp16")]; + tensor roti_cast_fp16 = sin(x = var_9210_cast_fp16)[name = tensor("roti_cast_fp16")]; + tensor var_9214_cast_fp16 = mul(x = var_9180_cast_fp16, y = rotr_cast_fp16)[name = tensor("op_9214_cast_fp16")]; + tensor var_9215_cast_fp16 = mul(x = var_9188_cast_fp16, y = roti_cast_fp16)[name = tensor("op_9215_cast_fp16")]; + tensor qor_93_cast_fp16 = sub(x = var_9214_cast_fp16, y = var_9215_cast_fp16)[name = tensor("qor_93_cast_fp16")]; + tensor var_9218_cast_fp16 = mul(x = var_9180_cast_fp16, y = roti_cast_fp16)[name = tensor("op_9218_cast_fp16")]; + tensor var_9219_cast_fp16 = mul(x = var_9188_cast_fp16, y = rotr_cast_fp16)[name = tensor("op_9219_cast_fp16")]; + tensor qoi_93_cast_fp16 = add(x = var_9218_cast_fp16, y = var_9219_cast_fp16)[name = tensor("qoi_93_cast_fp16")]; + tensor var_9222_cast_fp16 = mul(x = var_9196_cast_fp16, y = rotr_cast_fp16)[name = tensor("op_9222_cast_fp16")]; + tensor var_9223_cast_fp16 = mul(x = var_9204_cast_fp16, y = roti_cast_fp16)[name = tensor("op_9223_cast_fp16")]; + tensor kor_93_cast_fp16 = sub(x = var_9222_cast_fp16, y = var_9223_cast_fp16)[name = tensor("kor_93_cast_fp16")]; + tensor var_9226_cast_fp16 = mul(x = var_9196_cast_fp16, y = roti_cast_fp16)[name = tensor("op_9226_cast_fp16")]; + tensor var_9227_cast_fp16 = mul(x = var_9204_cast_fp16, y = rotr_cast_fp16)[name = tensor("op_9227_cast_fp16")]; + tensor koi_93_cast_fp16 = add(x = var_9226_cast_fp16, y = var_9227_cast_fp16)[name = tensor("koi_93_cast_fp16")]; + tensor qo_axis_0 = const()[name = tensor("qo_axis_0"), val = tensor(-1)]; + tensor qo_cast_fp16 = stack(axis = qo_axis_0, values = (qor_93_cast_fp16, qoi_93_cast_fp16))[name = tensor("qo_cast_fp16")]; + tensor ko_axis_0 = const()[name = tensor("ko_axis_0"), val = tensor(-1)]; + tensor ko_cast_fp16 = stack(axis = ko_axis_0, values = (kor_93_cast_fp16, koi_93_cast_fp16))[name = tensor("ko_cast_fp16")]; + tensor var_9260 = const()[name = tensor("op_9260"), val = tensor([1, 1, 16, 64])]; + tensor q_141_cast_fp16 = reshape(shape = var_9260, x = qo_cast_fp16)[name = tensor("q_141_cast_fp16")]; + tensor var_9266 = const()[name = tensor("op_9266"), val = tensor([1, 1, 16, 64])]; + tensor k_cast_fp16 = reshape(shape = var_9266, x = ko_cast_fp16)[name = tensor("k_cast_fp16")]; + tensor _inversed_9291_y_0_to_fp16 = const()[name = tensor("_inversed_9291_y_0_to_fp16"), val = tensor(0x1p-9)]; + tensor _inversed_9291_cast_fp16 = mul(x = ts_cast_fp16, y = _inversed_9291_y_0_to_fp16)[name = tensor("_inversed_9291_cast_fp16")]; + tensor var_9292_cast_fp16 = floor(x = _inversed_9291_cast_fp16)[name = tensor("op_9292_cast_fp16")]; + tensor var_9293_to_fp16 = const()[name = tensor("op_9293_to_fp16"), val = tensor(0x1p+9)]; + tensor var_9294_cast_fp16 = mul(x = var_9292_cast_fp16, y = var_9293_to_fp16)[name = tensor("op_9294_cast_fp16")]; + tensor write_indices_float_cast_fp16 = sub(x = ts_cast_fp16, y = var_9294_cast_fp16)[name = tensor("write_indices_float_cast_fp16")]; + tensor var_9301_dtype_0 = const()[name = tensor("op_9301_dtype_0"), val = tensor("int32")]; + tensor write_indices_reps_0 = const()[name = tensor("write_indices_reps_0"), val = tensor([1, 1, 16, 64])]; + tensor write_indices_float_cast_fp16_to_int32 = cast(dtype = var_9301_dtype_0, x = write_indices_float_cast_fp16)[name = tensor("cast_437")]; + tensor write_indices = tile(reps = write_indices_reps_0, x = write_indices_float_cast_fp16_to_int32)[name = tensor("write_indices")]; + tensor var_9311_begin_0 = const()[name = tensor("op_9311_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_9311_end_0 = const()[name = tensor("op_9311_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor var_9311_end_mask_0 = const()[name = tensor("op_9311_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_9311_squeeze_mask_0 = const()[name = tensor("op_9311_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor cache23_to_fp16_dtype_0 = const()[name = tensor("cache23_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor cache23_to_fp16 = cast(dtype = cache23_to_fp16_dtype_0, x = cache23)[name = tensor("cast_436")]; + tensor var_9311_cast_fp16 = slice_by_index(begin = var_9311_begin_0, end = var_9311_end_0, end_mask = var_9311_end_mask_0, squeeze_mask = var_9311_squeeze_mask_0, x = cache23_to_fp16)[name = tensor("op_9311_cast_fp16")]; + tensor var_9313_axis_0 = const()[name = tensor("op_9313_axis_0"), val = tensor(1)]; + tensor var_9313_mode_0 = const()[name = tensor("op_9313_mode_0"), val = tensor("update")]; + tensor var_9313_validate_indices_0 = const()[name = tensor("op_9313_validate_indices_0"), val = tensor(false)]; + tensor var_9313_cast_fp16 = scatter_along_axis(axis = var_9313_axis_0, data = var_9311_cast_fp16, indices = write_indices, mode = var_9313_mode_0, updates = k_cast_fp16, validate_indices = var_9313_validate_indices_0)[name = tensor("op_9313_cast_fp16")]; + tensor concat_163 = const()[name = tensor("concat_163"), val = tensor([0, 0, 0, 0, 0])]; + tensor concat_164 = const()[name = tensor("concat_164"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_internal_tensor_assign_1_stride_0 = const()[name = tensor("new_cache_internal_tensor_assign_1_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_internal_tensor_assign_1_begin_mask_0 = const()[name = tensor("new_cache_internal_tensor_assign_1_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_internal_tensor_assign_1_end_mask_0 = const()[name = tensor("new_cache_internal_tensor_assign_1_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_internal_tensor_assign_1_squeeze_mask_0 = const()[name = tensor("new_cache_internal_tensor_assign_1_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_94 = const()[name = tensor("shape_94"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_46 = const()[name = tensor("reduce_prod_46"), val = tensor(1048576)]; + tensor range_1d_46_start_0 = const()[name = tensor("range_1d_46_start_0"), val = tensor(0)]; + tensor range_1d_46_step_0 = const()[name = tensor("range_1d_46_step_0"), val = tensor(1)]; + tensor range_1d_46 = range_1d(end = reduce_prod_46, start = range_1d_46_start_0, step = range_1d_46_step_0)[name = tensor("range_1d_46")]; + tensor reshape_230 = reshape(shape = shape_94, x = range_1d_46)[name = tensor("reshape_230")]; + tensor slice_by_index_46 = slice_by_index(begin = concat_163, begin_mask = new_cache_internal_tensor_assign_1_begin_mask_0, end = concat_164, end_mask = new_cache_internal_tensor_assign_1_end_mask_0, squeeze_mask = new_cache_internal_tensor_assign_1_squeeze_mask_0, stride = new_cache_internal_tensor_assign_1_stride_0, x = reshape_230)[name = tensor("slice_by_index_46")]; + tensor reshape_231_shape_0 = const()[name = tensor("reshape_231_shape_0"), val = tensor([-1])]; + tensor reshape_231 = reshape(shape = reshape_231_shape_0, x = slice_by_index_46)[name = tensor("reshape_231")]; + tensor reshape_232_shape_0 = const()[name = tensor("reshape_232_shape_0"), val = tensor([-1])]; + tensor reshape_232_cast_fp16 = reshape(shape = reshape_232_shape_0, x = var_9313_cast_fp16)[name = tensor("reshape_232_cast_fp16")]; + tensor reshape_233_shape_0 = const()[name = tensor("reshape_233_shape_0"), val = tensor([-1])]; + tensor reshape_233_cast_fp16 = reshape(shape = reshape_233_shape_0, x = cache23_to_fp16)[name = tensor("reshape_233_cast_fp16")]; + tensor scatter_46_mode_0 = const()[name = tensor("scatter_46_mode_0"), val = tensor("update")]; + tensor scatter_46_axis_0 = const()[name = tensor("scatter_46_axis_0"), val = tensor(0)]; + tensor scatter_46_validate_indices_0 = const()[name = tensor("scatter_46_validate_indices_0"), val = tensor(false)]; + tensor scatter_46_cast_fp16 = scatter(axis = scatter_46_axis_0, data = reshape_233_cast_fp16, indices = reshape_231, mode = scatter_46_mode_0, updates = reshape_232_cast_fp16, validate_indices = scatter_46_validate_indices_0)[name = tensor("scatter_46_cast_fp16")]; + tensor reshape_234_cast_fp16 = reshape(shape = shape_94, x = scatter_46_cast_fp16)[name = tensor("reshape_234_cast_fp16")]; + tensor var_9321_begin_0 = const()[name = tensor("op_9321_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor var_9321_end_0 = const()[name = tensor("op_9321_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor var_9321_end_mask_0 = const()[name = tensor("op_9321_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_9321_squeeze_mask_0 = const()[name = tensor("op_9321_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor var_9321_cast_fp16 = slice_by_index(begin = var_9321_begin_0, end = var_9321_end_0, end_mask = var_9321_end_mask_0, squeeze_mask = var_9321_squeeze_mask_0, x = reshape_234_cast_fp16)[name = tensor("op_9321_cast_fp16")]; + tensor var_9323_axis_0 = const()[name = tensor("op_9323_axis_0"), val = tensor(1)]; + tensor var_9323_mode_0 = const()[name = tensor("op_9323_mode_0"), val = tensor("update")]; + tensor var_9323_validate_indices_0 = const()[name = tensor("op_9323_validate_indices_0"), val = tensor(false)]; + tensor var_9323_cast_fp16 = scatter_along_axis(axis = var_9323_axis_0, data = var_9321_cast_fp16, indices = write_indices, mode = var_9323_mode_0, updates = v_cast_fp16, validate_indices = var_9323_validate_indices_0)[name = tensor("op_9323_cast_fp16")]; + tensor concat_165 = const()[name = tensor("concat_165"), val = tensor([1, 0, 0, 0, 0])]; + tensor concat_166 = const()[name = tensor("concat_166"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_internal_tensor_assign_2_stride_0 = const()[name = tensor("new_cache_internal_tensor_assign_2_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_internal_tensor_assign_2_begin_mask_0 = const()[name = tensor("new_cache_internal_tensor_assign_2_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_internal_tensor_assign_2_end_mask_0 = const()[name = tensor("new_cache_internal_tensor_assign_2_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_internal_tensor_assign_2_squeeze_mask_0 = const()[name = tensor("new_cache_internal_tensor_assign_2_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_95 = const()[name = tensor("shape_95"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_47 = const()[name = tensor("reduce_prod_47"), val = tensor(1048576)]; + tensor range_1d_47_start_0 = const()[name = tensor("range_1d_47_start_0"), val = tensor(0)]; + tensor range_1d_47_step_0 = const()[name = tensor("range_1d_47_step_0"), val = tensor(1)]; + tensor range_1d_47 = range_1d(end = reduce_prod_47, start = range_1d_47_start_0, step = range_1d_47_step_0)[name = tensor("range_1d_47")]; + tensor reshape_235 = reshape(shape = shape_95, x = range_1d_47)[name = tensor("reshape_235")]; + tensor slice_by_index_47 = slice_by_index(begin = concat_165, begin_mask = new_cache_internal_tensor_assign_2_begin_mask_0, end = concat_166, end_mask = new_cache_internal_tensor_assign_2_end_mask_0, squeeze_mask = new_cache_internal_tensor_assign_2_squeeze_mask_0, stride = new_cache_internal_tensor_assign_2_stride_0, x = reshape_235)[name = tensor("slice_by_index_47")]; + tensor reshape_236_shape_0 = const()[name = tensor("reshape_236_shape_0"), val = tensor([-1])]; + tensor reshape_236 = reshape(shape = reshape_236_shape_0, x = slice_by_index_47)[name = tensor("reshape_236")]; + tensor reshape_237_shape_0 = const()[name = tensor("reshape_237_shape_0"), val = tensor([-1])]; + tensor reshape_237_cast_fp16 = reshape(shape = reshape_237_shape_0, x = var_9323_cast_fp16)[name = tensor("reshape_237_cast_fp16")]; + tensor reshape_238_shape_0 = const()[name = tensor("reshape_238_shape_0"), val = tensor([-1])]; + tensor reshape_238_cast_fp16 = reshape(shape = reshape_238_shape_0, x = reshape_234_cast_fp16)[name = tensor("reshape_238_cast_fp16")]; + tensor scatter_47_mode_0 = const()[name = tensor("scatter_47_mode_0"), val = tensor("update")]; + tensor scatter_47_axis_0 = const()[name = tensor("scatter_47_axis_0"), val = tensor(0)]; + tensor scatter_47_validate_indices_0 = const()[name = tensor("scatter_47_validate_indices_0"), val = tensor(false)]; + tensor scatter_47_cast_fp16 = scatter(axis = scatter_47_axis_0, data = reshape_238_cast_fp16, indices = reshape_236, mode = scatter_47_mode_0, updates = reshape_237_cast_fp16, validate_indices = scatter_47_validate_indices_0)[name = tensor("scatter_47_cast_fp16")]; + tensor reshape_239_cast_fp16 = reshape(shape = shape_95, x = scatter_47_cast_fp16)[name = tensor("reshape_239_cast_fp16")]; + tensor reshape_239_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("reshape_239_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor keys_139_begin_0 = const()[name = tensor("keys_139_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor keys_139_end_0 = const()[name = tensor("keys_139_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor keys_139_end_mask_0 = const()[name = tensor("keys_139_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor keys_139_squeeze_mask_0 = const()[name = tensor("keys_139_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor keys_139_cast_fp16 = slice_by_index(begin = keys_139_begin_0, end = keys_139_end_0, end_mask = keys_139_end_mask_0, squeeze_mask = keys_139_squeeze_mask_0, x = reshape_239_cast_fp16)[name = tensor("keys_139_cast_fp16")]; + tensor values_139_begin_0 = const()[name = tensor("values_139_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor values_139_end_0 = const()[name = tensor("values_139_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor values_139_end_mask_0 = const()[name = tensor("values_139_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor values_139_squeeze_mask_0 = const()[name = tensor("values_139_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor values_139_cast_fp16 = slice_by_index(begin = values_139_begin_0, end = values_139_end_0, end_mask = values_139_end_mask_0, squeeze_mask = values_139_squeeze_mask_0, x = reshape_239_cast_fp16)[name = tensor("values_139_cast_fp16")]; + tensor var_9335_cast_fp16 = not_equal(x = keys_139_cast_fp16, y = keys_139_cast_fp16)[name = tensor("op_9335_cast_fp16")]; + tensor keys_141_cast_fp16 = select(a = var_463_to_fp16, b = keys_139_cast_fp16, cond = var_9335_cast_fp16)[name = tensor("keys_141_cast_fp16")]; + tensor var_9343_cast_fp16 = not_equal(x = values_139_cast_fp16, y = values_139_cast_fp16)[name = tensor("op_9343_cast_fp16")]; + tensor values_141_cast_fp16 = select(a = var_463_to_fp16, b = values_139_cast_fp16, cond = var_9343_cast_fp16)[name = tensor("values_141_cast_fp16")]; + tensor var_9367 = const()[name = tensor("op_9367"), val = tensor([0, 2, 1, 3])]; + tensor var_9383 = const()[name = tensor("op_9383"), val = tensor([1, 1, 1])]; + tensor var_9384_cast_fp16 = reshape(shape = var_9383, x = position23_to_fp16)[name = tensor("op_9384_cast_fp16")]; + tensor var_9404_to_fp16 = const()[name = tensor("op_9404_to_fp16"), val = tensor(0x1p+0)]; + tensor valid_len_cast_fp16 = add(x = var_9384_cast_fp16, y = var_9404_to_fp16)[name = tensor("valid_len_cast_fp16")]; + tensor valid_mask_cast_fp16 = less(x = k_positions_1_promoted_to_fp16, y = valid_len_cast_fp16)[name = tensor("valid_mask_cast_fp16")]; + tensor causal_mask_cast_fp16 = less_equal(x = k_positions_1_promoted_to_fp16, y = var_9384_cast_fp16)[name = tensor("causal_mask_cast_fp16")]; + tensor attn_mask_93 = logical_and(x = valid_mask_cast_fp16, y = causal_mask_cast_fp16)[name = tensor("attn_mask_93")]; + tensor attn_mask_axes_0 = const()[name = tensor("attn_mask_axes_0"), val = tensor([1])]; + tensor attn_mask = expand_dims(axes = attn_mask_axes_0, x = attn_mask_93)[name = tensor("attn_mask")]; + tensor var_9422_transpose_x_0 = const()[name = tensor("op_9422_transpose_x_0"), val = tensor(false)]; + tensor var_9422_transpose_y_0 = const()[name = tensor("op_9422_transpose_y_0"), val = tensor(false)]; + tensor transpose_118_perm_0 = const()[name = tensor("transpose_118_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_119_perm_0 = const()[name = tensor("transpose_119_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_119 = transpose(perm = transpose_119_perm_0, x = keys_141_cast_fp16)[name = tensor("transpose_121")]; + tensor transpose_118 = transpose(perm = transpose_118_perm_0, x = q_141_cast_fp16)[name = tensor("transpose_122")]; + tensor var_9422_cast_fp16 = matmul(transpose_x = var_9422_transpose_x_0, transpose_y = var_9422_transpose_y_0, x = transpose_118, y = transpose_119)[name = tensor("op_9422_cast_fp16")]; + tensor var_9416_to_fp16 = const()[name = tensor("op_9416_to_fp16"), val = tensor([0x1p-3])]; + tensor attn_weights_139_cast_fp16 = mul(x = var_9422_cast_fp16, y = var_9416_to_fp16)[name = tensor("attn_weights_139_cast_fp16")]; + tensor var_9424 = logical_not(x = attn_mask)[name = tensor("op_9424")]; + tensor var_9425_to_fp16 = const()[name = tensor("op_9425_to_fp16"), val = tensor(-inf)]; + tensor attn_weights_141_cast_fp16 = select(a = var_9425_to_fp16, b = attn_weights_139_cast_fp16, cond = var_9424)[name = tensor("attn_weights_141_cast_fp16")]; + tensor var_9427 = const()[name = tensor("op_9427"), val = tensor(-1)]; + tensor attn_weights_cast_fp16 = softmax(axis = var_9427, x = attn_weights_141_cast_fp16)[name = tensor("attn_weights_cast_fp16")]; + tensor attn_output_transpose_x_0 = const()[name = tensor("attn_output_transpose_x_0"), val = tensor(false)]; + tensor attn_output_transpose_y_0 = const()[name = tensor("attn_output_transpose_y_0"), val = tensor(false)]; + tensor values_cast_fp16 = transpose(perm = var_9367, x = values_141_cast_fp16)[name = tensor("transpose_123")]; + tensor attn_output_cast_fp16 = matmul(transpose_x = attn_output_transpose_x_0, transpose_y = attn_output_transpose_y_0, x = attn_weights_cast_fp16, y = values_cast_fp16)[name = tensor("attn_output_cast_fp16")]; + tensor var_9435 = const()[name = tensor("op_9435"), val = tensor([0, 2, 1, 3])]; + tensor var_9440 = const()[name = tensor("op_9440"), val = tensor([1, 1, 1024])]; + tensor var_9436_cast_fp16 = transpose(perm = var_9435, x = attn_output_cast_fp16)[name = tensor("transpose_120")]; + tensor input_283_cast_fp16 = reshape(shape = var_9440, x = var_9436_cast_fp16)[name = tensor("input_283_cast_fp16")]; + tensor attn23_out_proj_weight_to_fp16 = const()[name = tensor("attn23_out_proj_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(586444928)))]; + tensor linear_94_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = attn23_out_proj_weight_to_fp16, x = input_283_cast_fp16)[name = tensor("linear_94_cast_fp16")]; + tensor var_9446_to_fp16 = const()[name = tensor("op_9446_to_fp16"), val = tensor(0x1p+0)]; + tensor var_9447_cast_fp16 = add(x = position23_to_fp16, y = var_9446_to_fp16)[name = tensor("op_9447_cast_fp16")]; + tensor var_9447_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_9447_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor input_285_cast_fp16 = add(x = input_279_cast_fp16, y = linear_94_cast_fp16)[name = tensor("input_285_cast_fp16")]; + tensor input_287_axes_0 = const()[name = tensor("input_287_axes_0"), val = tensor([-1])]; + tensor norm23_2_weight_to_fp16 = const()[name = tensor("norm23_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(588542144)))]; + tensor norm23_2_bias_to_fp16 = const()[name = tensor("norm23_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(588544256)))]; + tensor var_9451_to_fp16 = const()[name = tensor("op_9451_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_287_cast_fp16 = layer_norm(axes = input_287_axes_0, beta = norm23_2_bias_to_fp16, epsilon = var_9451_to_fp16, gamma = norm23_2_weight_to_fp16, x = input_285_cast_fp16)[name = tensor("input_287_cast_fp16")]; + tensor linear23_1_weight_to_fp16 = const()[name = tensor("linear23_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(588546368)))]; + tensor linear_95_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = linear23_1_weight_to_fp16, x = input_287_cast_fp16)[name = tensor("linear_95_cast_fp16")]; + tensor input_289_mode_0 = const()[name = tensor("input_289_mode_0"), val = tensor("EXACT")]; + tensor input_289_cast_fp16 = gelu(mode = input_289_mode_0, x = linear_95_cast_fp16)[name = tensor("input_289_cast_fp16")]; + tensor linear23_2_weight_to_fp16 = const()[name = tensor("linear23_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(596935040)))]; + tensor linear_96_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = linear23_2_weight_to_fp16, x = input_289_cast_fp16)[name = tensor("linear_96_cast_fp16")]; + tensor input_291_cast_fp16 = add(x = input_285_cast_fp16, y = linear_96_cast_fp16)[name = tensor("input_291_cast_fp16")]; + tensor input_axes_0 = const()[name = tensor("input_axes_0"), val = tensor([-1])]; + tensor out_norm_weight_to_fp16 = const()[name = tensor("out_norm_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(605323712)))]; + tensor out_norm_bias_to_fp16 = const()[name = tensor("out_norm_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(605325824)))]; + tensor var_9468_to_fp16 = const()[name = tensor("op_9468_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_cast_fp16 = layer_norm(axes = input_axes_0, beta = out_norm_bias_to_fp16, epsilon = var_9468_to_fp16, gamma = out_norm_weight_to_fp16, x = input_291_cast_fp16)[name = tensor("input_cast_fp16")]; + tensor input_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("input_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor out_eos_weight_to_fp16 = const()[name = tensor("out_eos_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(605327936)))]; + tensor out_eos_bias_to_fp16 = const()[name = tensor("out_eos_bias_to_fp16"), val = tensor([-0x1.5ep-3])]; + tensor linear_97_cast_fp16 = linear(bias = out_eos_bias_to_fp16, weight = out_eos_weight_to_fp16, x = input_cast_fp16)[name = tensor("linear_97_cast_fp16")]; + tensor linear_97_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("linear_97_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor var_9476 = cast(dtype = linear_97_cast_fp16_to_fp32_dtype_0, x = linear_97_cast_fp16)[name = tensor("cast_432")]; + tensor input = cast(dtype = input_cast_fp16_to_fp32_dtype_0, x = input_cast_fp16)[name = tensor("cast_433")]; + tensor var_9447 = cast(dtype = var_9447_cast_fp16_to_fp32_dtype_0, x = var_9447_cast_fp16)[name = tensor("cast_434")]; + tensor new_cache_internal_tensor_assign_2 = cast(dtype = reshape_239_cast_fp16_to_fp32_dtype_0, x = reshape_239_cast_fp16)[name = tensor("cast_435")]; + tensor var_9061 = cast(dtype = var_9061_cast_fp16_to_fp32_dtype_0, x = var_9061_cast_fp16)[name = tensor("cast_439")]; + tensor new_cache_45_internal_tensor_assign_2 = cast(dtype = reshape_229_cast_fp16_to_fp32_dtype_0, x = reshape_229_cast_fp16)[name = tensor("cast_440")]; + tensor var_8675 = cast(dtype = var_8675_cast_fp16_to_fp32_dtype_0, x = var_8675_cast_fp16)[name = tensor("cast_444")]; + tensor new_cache_43_internal_tensor_assign_2 = cast(dtype = reshape_219_cast_fp16_to_fp32_dtype_0, x = reshape_219_cast_fp16)[name = tensor("cast_445")]; + tensor var_8289 = cast(dtype = var_8289_cast_fp16_to_fp32_dtype_0, x = var_8289_cast_fp16)[name = tensor("cast_449")]; + tensor new_cache_41_internal_tensor_assign_2 = cast(dtype = reshape_209_cast_fp16_to_fp32_dtype_0, x = reshape_209_cast_fp16)[name = tensor("cast_450")]; + tensor var_7903 = cast(dtype = var_7903_cast_fp16_to_fp32_dtype_0, x = var_7903_cast_fp16)[name = tensor("cast_454")]; + tensor new_cache_39_internal_tensor_assign_2 = cast(dtype = reshape_199_cast_fp16_to_fp32_dtype_0, x = reshape_199_cast_fp16)[name = tensor("cast_455")]; + tensor var_7517 = cast(dtype = var_7517_cast_fp16_to_fp32_dtype_0, x = var_7517_cast_fp16)[name = tensor("cast_459")]; + tensor new_cache_37_internal_tensor_assign_2 = cast(dtype = reshape_189_cast_fp16_to_fp32_dtype_0, x = reshape_189_cast_fp16)[name = tensor("cast_460")]; + tensor var_7131 = cast(dtype = var_7131_cast_fp16_to_fp32_dtype_0, x = var_7131_cast_fp16)[name = tensor("cast_464")]; + tensor new_cache_35_internal_tensor_assign_2 = cast(dtype = reshape_179_cast_fp16_to_fp32_dtype_0, x = reshape_179_cast_fp16)[name = tensor("cast_465")]; + tensor var_6745 = cast(dtype = var_6745_cast_fp16_to_fp32_dtype_0, x = var_6745_cast_fp16)[name = tensor("cast_469")]; + tensor new_cache_33_internal_tensor_assign_2 = cast(dtype = reshape_169_cast_fp16_to_fp32_dtype_0, x = reshape_169_cast_fp16)[name = tensor("cast_470")]; + tensor var_6359 = cast(dtype = var_6359_cast_fp16_to_fp32_dtype_0, x = var_6359_cast_fp16)[name = tensor("cast_474")]; + tensor new_cache_31_internal_tensor_assign_2 = cast(dtype = reshape_159_cast_fp16_to_fp32_dtype_0, x = reshape_159_cast_fp16)[name = tensor("cast_475")]; + tensor var_5973 = cast(dtype = var_5973_cast_fp16_to_fp32_dtype_0, x = var_5973_cast_fp16)[name = tensor("cast_479")]; + tensor new_cache_29_internal_tensor_assign_2 = cast(dtype = reshape_149_cast_fp16_to_fp32_dtype_0, x = reshape_149_cast_fp16)[name = tensor("cast_480")]; + tensor var_5587 = cast(dtype = var_5587_cast_fp16_to_fp32_dtype_0, x = var_5587_cast_fp16)[name = tensor("cast_484")]; + tensor new_cache_27_internal_tensor_assign_2 = cast(dtype = reshape_139_cast_fp16_to_fp32_dtype_0, x = reshape_139_cast_fp16)[name = tensor("cast_485")]; + tensor var_5201 = cast(dtype = var_5201_cast_fp16_to_fp32_dtype_0, x = var_5201_cast_fp16)[name = tensor("cast_489")]; + tensor new_cache_25_internal_tensor_assign_2 = cast(dtype = reshape_129_cast_fp16_to_fp32_dtype_0, x = reshape_129_cast_fp16)[name = tensor("cast_490")]; + tensor var_4815 = cast(dtype = var_4815_cast_fp16_to_fp32_dtype_0, x = var_4815_cast_fp16)[name = tensor("cast_494")]; + tensor new_cache_23_internal_tensor_assign_2 = cast(dtype = reshape_119_cast_fp16_to_fp32_dtype_0, x = reshape_119_cast_fp16)[name = tensor("cast_495")]; + tensor var_4429 = cast(dtype = var_4429_cast_fp16_to_fp32_dtype_0, x = var_4429_cast_fp16)[name = tensor("cast_499")]; + tensor new_cache_21_internal_tensor_assign_2 = cast(dtype = reshape_109_cast_fp16_to_fp32_dtype_0, x = reshape_109_cast_fp16)[name = tensor("cast_500")]; + tensor var_4043 = cast(dtype = var_4043_cast_fp16_to_fp32_dtype_0, x = var_4043_cast_fp16)[name = tensor("cast_504")]; + tensor new_cache_19_internal_tensor_assign_2 = cast(dtype = reshape_99_cast_fp16_to_fp32_dtype_0, x = reshape_99_cast_fp16)[name = tensor("cast_505")]; + tensor var_3657 = cast(dtype = var_3657_cast_fp16_to_fp32_dtype_0, x = var_3657_cast_fp16)[name = tensor("cast_509")]; + tensor new_cache_17_internal_tensor_assign_2 = cast(dtype = reshape_89_cast_fp16_to_fp32_dtype_0, x = reshape_89_cast_fp16)[name = tensor("cast_510")]; + tensor var_3271 = cast(dtype = var_3271_cast_fp16_to_fp32_dtype_0, x = var_3271_cast_fp16)[name = tensor("cast_514")]; + tensor new_cache_15_internal_tensor_assign_2 = cast(dtype = reshape_79_cast_fp16_to_fp32_dtype_0, x = reshape_79_cast_fp16)[name = tensor("cast_515")]; + tensor var_2885 = cast(dtype = var_2885_cast_fp16_to_fp32_dtype_0, x = var_2885_cast_fp16)[name = tensor("cast_519")]; + tensor new_cache_13_internal_tensor_assign_2 = cast(dtype = reshape_69_cast_fp16_to_fp32_dtype_0, x = reshape_69_cast_fp16)[name = tensor("cast_520")]; + tensor var_2499 = cast(dtype = var_2499_cast_fp16_to_fp32_dtype_0, x = var_2499_cast_fp16)[name = tensor("cast_524")]; + tensor new_cache_11_internal_tensor_assign_2 = cast(dtype = reshape_59_cast_fp16_to_fp32_dtype_0, x = reshape_59_cast_fp16)[name = tensor("cast_525")]; + tensor var_2113 = cast(dtype = var_2113_cast_fp16_to_fp32_dtype_0, x = var_2113_cast_fp16)[name = tensor("cast_529")]; + tensor new_cache_9_internal_tensor_assign_2 = cast(dtype = reshape_49_cast_fp16_to_fp32_dtype_0, x = reshape_49_cast_fp16)[name = tensor("cast_530")]; + tensor var_1727 = cast(dtype = var_1727_cast_fp16_to_fp32_dtype_0, x = var_1727_cast_fp16)[name = tensor("cast_534")]; + tensor new_cache_7_internal_tensor_assign_2 = cast(dtype = reshape_39_cast_fp16_to_fp32_dtype_0, x = reshape_39_cast_fp16)[name = tensor("cast_535")]; + tensor var_1341 = cast(dtype = var_1341_cast_fp16_to_fp32_dtype_0, x = var_1341_cast_fp16)[name = tensor("cast_539")]; + tensor new_cache_5_internal_tensor_assign_2 = cast(dtype = reshape_29_cast_fp16_to_fp32_dtype_0, x = reshape_29_cast_fp16)[name = tensor("cast_540")]; + tensor var_955 = cast(dtype = var_955_cast_fp16_to_fp32_dtype_0, x = var_955_cast_fp16)[name = tensor("cast_544")]; + tensor new_cache_3_internal_tensor_assign_2 = cast(dtype = reshape_19_cast_fp16_to_fp32_dtype_0, x = reshape_19_cast_fp16)[name = tensor("cast_545")]; + tensor var_569 = cast(dtype = var_569_cast_fp16_to_fp32_dtype_0, x = var_569_cast_fp16)[name = tensor("cast_549")]; + tensor new_cache_1_internal_tensor_assign_2 = cast(dtype = reshape_9_cast_fp16_to_fp32_dtype_0, x = reshape_9_cast_fp16)[name = tensor("cast_550")]; + } -> (input, var_9476, new_cache_1_internal_tensor_assign_2, var_569, new_cache_3_internal_tensor_assign_2, var_955, new_cache_5_internal_tensor_assign_2, var_1341, new_cache_7_internal_tensor_assign_2, var_1727, new_cache_9_internal_tensor_assign_2, var_2113, new_cache_11_internal_tensor_assign_2, var_2499, new_cache_13_internal_tensor_assign_2, var_2885, new_cache_15_internal_tensor_assign_2, var_3271, new_cache_17_internal_tensor_assign_2, var_3657, new_cache_19_internal_tensor_assign_2, var_4043, new_cache_21_internal_tensor_assign_2, var_4429, new_cache_23_internal_tensor_assign_2, var_4815, new_cache_25_internal_tensor_assign_2, var_5201, new_cache_27_internal_tensor_assign_2, var_5587, new_cache_29_internal_tensor_assign_2, var_5973, new_cache_31_internal_tensor_assign_2, var_6359, new_cache_33_internal_tensor_assign_2, var_6745, new_cache_35_internal_tensor_assign_2, var_7131, new_cache_37_internal_tensor_assign_2, var_7517, new_cache_39_internal_tensor_assign_2, var_7903, new_cache_41_internal_tensor_assign_2, var_8289, new_cache_43_internal_tensor_assign_2, var_8675, new_cache_45_internal_tensor_assign_2, var_9061, new_cache_internal_tensor_assign_2, var_9447); +} \ No newline at end of file diff --git a/v2.1/spanish_24l/flowlm_step.mlmodelc/weights/weight.bin b/v2.1/spanish_24l/flowlm_step.mlmodelc/weights/weight.bin new file mode 100644 index 0000000000000000000000000000000000000000..93dc53df6ce08ebaed8b165fe0b30ee554ca0be1 --- /dev/null +++ b/v2.1/spanish_24l/flowlm_step.mlmodelc/weights/weight.bin @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:037a47c87454c8926d7c1b38e955eb41080c5bc8685a35f00de558c962e2ab53 +size 605330048 diff --git a/v2.1/spanish_24l/flowlm_step.mlpackage/Data/com.apple.CoreML/model.mlmodel b/v2.1/spanish_24l/flowlm_step.mlpackage/Data/com.apple.CoreML/model.mlmodel new file mode 100644 index 0000000000000000000000000000000000000000..859b8fda91fb2705c751d520ad149ea216c98c78 --- /dev/null +++ b/v2.1/spanish_24l/flowlm_step.mlpackage/Data/com.apple.CoreML/model.mlmodel @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:33036b133ceaa042887760ae001b7cd705f58796187d678b906d788bc02f39c1 +size 825108 diff --git a/v2.1/spanish_24l/flowlm_step.mlpackage/Data/com.apple.CoreML/weights/weight.bin b/v2.1/spanish_24l/flowlm_step.mlpackage/Data/com.apple.CoreML/weights/weight.bin new file mode 100644 index 0000000000000000000000000000000000000000..93dc53df6ce08ebaed8b165fe0b30ee554ca0be1 --- /dev/null +++ b/v2.1/spanish_24l/flowlm_step.mlpackage/Data/com.apple.CoreML/weights/weight.bin @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:037a47c87454c8926d7c1b38e955eb41080c5bc8685a35f00de558c962e2ab53 +size 605330048 diff --git a/v2.1/spanish_24l/flowlm_step.mlpackage/Manifest.json b/v2.1/spanish_24l/flowlm_step.mlpackage/Manifest.json new file mode 100644 index 0000000000000000000000000000000000000000..e50cfc94aafe45321a55dd3b55011d7e07ce200e --- /dev/null +++ b/v2.1/spanish_24l/flowlm_step.mlpackage/Manifest.json @@ -0,0 +1,18 @@ +{ + "fileFormatVersion": "1.0.0", + "itemInfoEntries": { + "2DFC79AE-40F1-4A23-B6AD-EBD0C3746635": { + "author": "com.apple.CoreML", + "description": "CoreML Model Specification", + "name": "model.mlmodel", + "path": "com.apple.CoreML/model.mlmodel" + }, + "4EEC9ECB-B02B-4D28-975C-40995B2FE2CA": { + "author": "com.apple.CoreML", + "description": "CoreML Model Weights", + "name": "weights", + "path": "com.apple.CoreML/weights" + } + }, + "rootModelIdentifier": "2DFC79AE-40F1-4A23-B6AD-EBD0C3746635" +} diff --git a/v2.1/spanish_24l/flowlm_stepv2.mlmodelc/analytics/coremldata.bin b/v2.1/spanish_24l/flowlm_stepv2.mlmodelc/analytics/coremldata.bin new file mode 100644 index 0000000000000000000000000000000000000000..8512933b7772f492b9999bed183b170a15094728 --- /dev/null +++ b/v2.1/spanish_24l/flowlm_stepv2.mlmodelc/analytics/coremldata.bin @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:fdc2fac192f57020f12ab5f9659313bc06839302e60eb69db52601f7346fd08f +size 243 diff --git a/v2.1/spanish_24l/flowlm_stepv2.mlmodelc/coremldata.bin b/v2.1/spanish_24l/flowlm_stepv2.mlmodelc/coremldata.bin new file mode 100644 index 0000000000000000000000000000000000000000..29f2b7048f66ee6e4201e1ec85baef14a779744c --- /dev/null +++ b/v2.1/spanish_24l/flowlm_stepv2.mlmodelc/coremldata.bin @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:0f588a391a6438abaf363bda542c72a0b7370e0a858f00bcbbc0235403d55f23 +size 3561 diff --git a/v2.1/spanish_24l/flowlm_stepv2.mlmodelc/metadata.json b/v2.1/spanish_24l/flowlm_stepv2.mlmodelc/metadata.json new file mode 100644 index 0000000000000000000000000000000000000000..629da9a69819375413a6c11a7789eb2ca4f04e28 --- /dev/null +++ b/v2.1/spanish_24l/flowlm_stepv2.mlmodelc/metadata.json @@ -0,0 +1,1068 @@ +[ + { + "metadataOutputVersion" : "3.0", + "storagePrecision" : "Mixed (Float16, Int8)", + "outputSchema" : [ + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1 × 1 × 1024)", + "shortDescription" : "", + "shape" : "[1, 1, 1024]", + "name" : "input", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1 × 1 × 1)", + "shortDescription" : "", + "shape" : "[1, 1, 1]", + "name" : "var_10268", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 2 × 1 × 512 × 16 × 64)", + "shortDescription" : "", + "shape" : "[2, 1, 512, 16, 64]", + "name" : "new_cache_1_internal_tensor_assign_2", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1)", + "shortDescription" : "", + "shape" : "[1]", + "name" : "var_602", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 2 × 1 × 512 × 16 × 64)", + "shortDescription" : "", + "shape" : "[2, 1, 512, 16, 64]", + "name" : "new_cache_3_internal_tensor_assign_2", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1)", + "shortDescription" : "", + "shape" : "[1]", + "name" : "var_1021", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 2 × 1 × 512 × 16 × 64)", + "shortDescription" : "", + "shape" : "[2, 1, 512, 16, 64]", + "name" : "new_cache_5_internal_tensor_assign_2", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1)", + "shortDescription" : "", + "shape" : "[1]", + "name" : "var_1440", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 2 × 1 × 512 × 16 × 64)", + "shortDescription" : "", + "shape" : "[2, 1, 512, 16, 64]", + "name" : "new_cache_7_internal_tensor_assign_2", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1)", + "shortDescription" : "", + "shape" : "[1]", + "name" : "var_1859", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 2 × 1 × 512 × 16 × 64)", + "shortDescription" : "", + "shape" : "[2, 1, 512, 16, 64]", + "name" : "new_cache_9_internal_tensor_assign_2", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1)", + "shortDescription" : "", + "shape" : "[1]", + "name" : "var_2278", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 2 × 1 × 512 × 16 × 64)", + "shortDescription" : "", + "shape" : "[2, 1, 512, 16, 64]", + "name" : "new_cache_11_internal_tensor_assign_2", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1)", + "shortDescription" : "", + "shape" : "[1]", + "name" : "var_2697", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 2 × 1 × 512 × 16 × 64)", + "shortDescription" : "", + "shape" : "[2, 1, 512, 16, 64]", + "name" : "new_cache_13_internal_tensor_assign_2", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1)", + "shortDescription" : "", + "shape" : "[1]", + "name" : "var_3116", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 2 × 1 × 512 × 16 × 64)", + "shortDescription" : "", + "shape" : "[2, 1, 512, 16, 64]", + "name" : "new_cache_15_internal_tensor_assign_2", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1)", + "shortDescription" : "", + "shape" : "[1]", + "name" : "var_3535", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 2 × 1 × 512 × 16 × 64)", + "shortDescription" : "", + "shape" : "[2, 1, 512, 16, 64]", + "name" : "new_cache_17_internal_tensor_assign_2", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1)", + "shortDescription" : "", + "shape" : "[1]", + "name" : "var_3954", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 2 × 1 × 512 × 16 × 64)", + "shortDescription" : "", + "shape" : "[2, 1, 512, 16, 64]", + "name" : "new_cache_19_internal_tensor_assign_2", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1)", + "shortDescription" : "", + "shape" : "[1]", + "name" : "var_4373", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 2 × 1 × 512 × 16 × 64)", + "shortDescription" : "", + "shape" : "[2, 1, 512, 16, 64]", + "name" : "new_cache_21_internal_tensor_assign_2", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1)", + "shortDescription" : "", + "shape" : "[1]", + "name" : "var_4792", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 2 × 1 × 512 × 16 × 64)", + "shortDescription" : "", + "shape" : "[2, 1, 512, 16, 64]", + "name" : "new_cache_23_internal_tensor_assign_2", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1)", + "shortDescription" : "", + "shape" : "[1]", + "name" : "var_5211", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 2 × 1 × 512 × 16 × 64)", + "shortDescription" : "", + "shape" : "[2, 1, 512, 16, 64]", + "name" : "new_cache_25_internal_tensor_assign_2", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1)", + "shortDescription" : "", + "shape" : "[1]", + "name" : "var_5630", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 2 × 1 × 512 × 16 × 64)", + "shortDescription" : "", + "shape" : "[2, 1, 512, 16, 64]", + "name" : "new_cache_27_internal_tensor_assign_2", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1)", + "shortDescription" : "", + "shape" : "[1]", + "name" : "var_6049", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 2 × 1 × 512 × 16 × 64)", + "shortDescription" : "", + "shape" : "[2, 1, 512, 16, 64]", + "name" : "new_cache_29_internal_tensor_assign_2", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1)", + "shortDescription" : "", + "shape" : "[1]", + "name" : "var_6468", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 2 × 1 × 512 × 16 × 64)", + "shortDescription" : "", + "shape" : "[2, 1, 512, 16, 64]", + "name" : "new_cache_31_internal_tensor_assign_2", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1)", + "shortDescription" : "", + "shape" : "[1]", + "name" : "var_6887", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 2 × 1 × 512 × 16 × 64)", + "shortDescription" : "", + "shape" : "[2, 1, 512, 16, 64]", + "name" : "new_cache_33_internal_tensor_assign_2", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1)", + "shortDescription" : "", + "shape" : "[1]", + "name" : "var_7306", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 2 × 1 × 512 × 16 × 64)", + "shortDescription" : "", + "shape" : "[2, 1, 512, 16, 64]", + "name" : "new_cache_35_internal_tensor_assign_2", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1)", + "shortDescription" : "", + "shape" : "[1]", + "name" : "var_7725", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 2 × 1 × 512 × 16 × 64)", + "shortDescription" : "", + "shape" : "[2, 1, 512, 16, 64]", + "name" : "new_cache_37_internal_tensor_assign_2", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1)", + "shortDescription" : "", + "shape" : "[1]", + "name" : "var_8144", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 2 × 1 × 512 × 16 × 64)", + "shortDescription" : "", + "shape" : "[2, 1, 512, 16, 64]", + "name" : "new_cache_39_internal_tensor_assign_2", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1)", + "shortDescription" : "", + "shape" : "[1]", + "name" : "var_8563", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 2 × 1 × 512 × 16 × 64)", + "shortDescription" : "", + "shape" : "[2, 1, 512, 16, 64]", + "name" : "new_cache_41_internal_tensor_assign_2", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1)", + "shortDescription" : "", + "shape" : "[1]", + "name" : "var_8982", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 2 × 1 × 512 × 16 × 64)", + "shortDescription" : "", + "shape" : "[2, 1, 512, 16, 64]", + "name" : "new_cache_43_internal_tensor_assign_2", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1)", + "shortDescription" : "", + "shape" : "[1]", + "name" : "var_9401", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 2 × 1 × 512 × 16 × 64)", + "shortDescription" : "", + "shape" : "[2, 1, 512, 16, 64]", + "name" : "new_cache_45_internal_tensor_assign_2", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1)", + "shortDescription" : "", + "shape" : "[1]", + "name" : "var_9820", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 2 × 1 × 512 × 16 × 64)", + "shortDescription" : "", + "shape" : "[2, 1, 512, 16, 64]", + "name" : "new_cache_internal_tensor_assign_2", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1)", + "shortDescription" : "", + "shape" : "[1]", + "name" : "var_10239", + "type" : "MultiArray" + } + ], + "modelParameters" : [ + + ], + "specificationVersion" : 8, + "mlProgramOperationTypeHistogram" : { + "Ios17.cos" : 24, + "Range1d" : 48, + "Ios17.notEqual" : 49, + "Ios17.floor" : 24, + "Ios17.reshape" : 432, + "Ios17.logicalAnd" : 24, + "Ios16.softmax" : 24, + "Ios17.scatter" : 48, + "Ios17.matmul" : 48, + "Ios17.sin" : 24, + "Ios17.transpose" : 96, + "Select" : 73, + "Ios17.expandDims" : 25, + "Ios17.add" : 144, + "Tile" : 24, + "Ios17.sliceByIndex" : 312, + "Ios17.lessEqual" : 24, + "Ios17.scatterAlongAxis" : 48, + "Ios17.logicalNot" : 24, + "Ios17.layerNorm" : 49, + "Ios16.constexprAffineDequantize" : 96, + "Ios17.cast" : 124, + "Ios17.sub" : 72, + "Ios17.less" : 24, + "Ios16.gelu" : 24, + "Ios17.linear" : 98, + "Stack" : 48, + "Ios17.mul" : 288 + }, + "computePrecision" : "Mixed (Float16, Float32, Int32)", + "isUpdatable" : "0", + "stateSchema" : [ + + ], + "availability" : { + "macOS" : "14.0", + "tvOS" : "17.0", + "visionOS" : "1.0", + "watchOS" : "10.0", + "iOS" : "17.0", + "macCatalyst" : "17.0" + }, + "modelType" : { + "name" : "MLModelType_mlProgram" + }, + "userDefinedMetadata" : { + "com.github.apple.coremltools.conversion_date" : "2026-04-27", + "com.github.apple.coremltools.source" : "torch==2.9.1", + "com.github.apple.coremltools.version" : "9.0", + "com.github.apple.coremltools.source_dialect" : "TorchScript" + }, + "inputSchema" : [ + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1 × 1 × 32)", + "shortDescription" : "", + "shape" : "[1, 1, 32]", + "name" : "sequence", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 32)", + "shortDescription" : "", + "shape" : "[32]", + "name" : "bos_emb", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 2 × 1 × 512 × 16 × 64)", + "shortDescription" : "", + "shape" : "[2, 1, 512, 16, 64]", + "name" : "cache0", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1)", + "shortDescription" : "", + "shape" : "[1]", + "name" : "position0", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 2 × 1 × 512 × 16 × 64)", + "shortDescription" : "", + "shape" : "[2, 1, 512, 16, 64]", + "name" : "cache1", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1)", + "shortDescription" : "", + "shape" : "[1]", + "name" : "position1", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 2 × 1 × 512 × 16 × 64)", + "shortDescription" : "", + "shape" : "[2, 1, 512, 16, 64]", + "name" : "cache2", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1)", + "shortDescription" : "", + "shape" : "[1]", + "name" : "position2", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 2 × 1 × 512 × 16 × 64)", + "shortDescription" : "", + "shape" : "[2, 1, 512, 16, 64]", + "name" : "cache3", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1)", + "shortDescription" : "", + "shape" : "[1]", + "name" : "position3", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 2 × 1 × 512 × 16 × 64)", + "shortDescription" : "", + "shape" : "[2, 1, 512, 16, 64]", + "name" : "cache4", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1)", + "shortDescription" : "", + "shape" : "[1]", + "name" : "position4", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 2 × 1 × 512 × 16 × 64)", + "shortDescription" : "", + "shape" : "[2, 1, 512, 16, 64]", + "name" : "cache5", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1)", + "shortDescription" : "", + "shape" : "[1]", + "name" : "position5", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 2 × 1 × 512 × 16 × 64)", + "shortDescription" : "", + "shape" : "[2, 1, 512, 16, 64]", + "name" : "cache6", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1)", + "shortDescription" : "", + "shape" : "[1]", + "name" : "position6", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 2 × 1 × 512 × 16 × 64)", + "shortDescription" : "", + "shape" : "[2, 1, 512, 16, 64]", + "name" : "cache7", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1)", + "shortDescription" : "", + "shape" : "[1]", + "name" : "position7", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 2 × 1 × 512 × 16 × 64)", + "shortDescription" : "", + "shape" : "[2, 1, 512, 16, 64]", + "name" : "cache8", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1)", + "shortDescription" : "", + "shape" : "[1]", + "name" : "position8", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 2 × 1 × 512 × 16 × 64)", + "shortDescription" : "", + "shape" : "[2, 1, 512, 16, 64]", + "name" : "cache9", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1)", + "shortDescription" : "", + "shape" : "[1]", + "name" : "position9", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 2 × 1 × 512 × 16 × 64)", + "shortDescription" : "", + "shape" : "[2, 1, 512, 16, 64]", + "name" : "cache10", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1)", + "shortDescription" : "", + "shape" : "[1]", + "name" : "position10", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 2 × 1 × 512 × 16 × 64)", + "shortDescription" : "", + "shape" : "[2, 1, 512, 16, 64]", + "name" : "cache11", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1)", + "shortDescription" : "", + "shape" : "[1]", + "name" : "position11", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 2 × 1 × 512 × 16 × 64)", + "shortDescription" : "", + "shape" : "[2, 1, 512, 16, 64]", + "name" : "cache12", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1)", + "shortDescription" : "", + "shape" : "[1]", + "name" : "position12", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 2 × 1 × 512 × 16 × 64)", + "shortDescription" : "", + "shape" : "[2, 1, 512, 16, 64]", + "name" : "cache13", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1)", + "shortDescription" : "", + "shape" : "[1]", + "name" : "position13", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 2 × 1 × 512 × 16 × 64)", + "shortDescription" : "", + "shape" : "[2, 1, 512, 16, 64]", + "name" : "cache14", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1)", + "shortDescription" : "", + "shape" : "[1]", + "name" : "position14", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 2 × 1 × 512 × 16 × 64)", + "shortDescription" : "", + "shape" : "[2, 1, 512, 16, 64]", + "name" : "cache15", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1)", + "shortDescription" : "", + "shape" : "[1]", + "name" : "position15", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 2 × 1 × 512 × 16 × 64)", + "shortDescription" : "", + "shape" : "[2, 1, 512, 16, 64]", + "name" : "cache16", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1)", + "shortDescription" : "", + "shape" : "[1]", + "name" : "position16", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 2 × 1 × 512 × 16 × 64)", + "shortDescription" : "", + "shape" : "[2, 1, 512, 16, 64]", + "name" : "cache17", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1)", + "shortDescription" : "", + "shape" : "[1]", + "name" : "position17", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 2 × 1 × 512 × 16 × 64)", + "shortDescription" : "", + "shape" : "[2, 1, 512, 16, 64]", + "name" : "cache18", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1)", + "shortDescription" : "", + "shape" : "[1]", + "name" : "position18", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 2 × 1 × 512 × 16 × 64)", + "shortDescription" : "", + "shape" : "[2, 1, 512, 16, 64]", + "name" : "cache19", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1)", + "shortDescription" : "", + "shape" : "[1]", + "name" : "position19", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 2 × 1 × 512 × 16 × 64)", + "shortDescription" : "", + "shape" : "[2, 1, 512, 16, 64]", + "name" : "cache20", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1)", + "shortDescription" : "", + "shape" : "[1]", + "name" : "position20", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 2 × 1 × 512 × 16 × 64)", + "shortDescription" : "", + "shape" : "[2, 1, 512, 16, 64]", + "name" : "cache21", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1)", + "shortDescription" : "", + "shape" : "[1]", + "name" : "position21", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 2 × 1 × 512 × 16 × 64)", + "shortDescription" : "", + "shape" : "[2, 1, 512, 16, 64]", + "name" : "cache22", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1)", + "shortDescription" : "", + "shape" : "[1]", + "name" : "position22", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 2 × 1 × 512 × 16 × 64)", + "shortDescription" : "", + "shape" : "[2, 1, 512, 16, 64]", + "name" : "cache23", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1)", + "shortDescription" : "", + "shape" : "[1]", + "name" : "position23", + "type" : "MultiArray" + } + ], + "generatedClassName" : "flowlm_stepv2", + "method" : "predict" + } +] \ No newline at end of file diff --git a/v2.1/spanish_24l/flowlm_stepv2.mlmodelc/model.mil b/v2.1/spanish_24l/flowlm_stepv2.mlmodelc/model.mil new file mode 100644 index 0000000000000000000000000000000000000000..50bc47c018725c1e0a044bf569de211771195a6f --- /dev/null +++ b/v2.1/spanish_24l/flowlm_stepv2.mlmodelc/model.mil @@ -0,0 +1,5337 @@ +program(1.0) +[buildInfo = dict, tensor>({{"coremlc-component-MIL", "3520.4.1"}, {"coremlc-version", "3520.5.1"}, {"coremltools-component-torch", "2.9.1"}, {"coremltools-source-dialect", "TorchScript"}, {"coremltools-version", "9.0"}})] +{ + func main(tensor bos_emb, tensor cache0, tensor cache1, tensor cache10, tensor cache11, tensor cache12, tensor cache13, tensor cache14, tensor cache15, tensor cache16, tensor cache17, tensor cache18, tensor cache19, tensor cache2, tensor cache20, tensor cache21, tensor cache22, tensor cache23, tensor cache3, tensor cache4, tensor cache5, tensor cache6, tensor cache7, tensor cache8, tensor cache9, tensor position0, tensor position1, tensor position10, tensor position11, tensor position12, tensor position13, tensor position14, tensor position15, tensor position16, tensor position17, tensor position18, tensor position19, tensor position2, tensor position20, tensor position21, tensor position22, tensor position23, tensor position3, tensor position4, tensor position5, tensor position6, tensor position7, tensor position8, tensor position9, tensor sequence) { + tensor attn0_in_proj_weight_cast_fp16 = constexpr_affine_dequantize()[axis = tensor(0), name = tensor("attn0_in_proj_weight_cast_fp16"), quantized_data = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(64))), scale = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(3148992))), zero_point = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(3145856)))]; + tensor attn0_out_proj_weight_cast_fp16 = constexpr_affine_dequantize()[axis = tensor(0), name = tensor("attn0_out_proj_weight_cast_fp16"), quantized_data = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(3155200))), scale = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(4204928))), zero_point = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(4203840)))]; + tensor linear0_1_weight_cast_fp16 = constexpr_affine_dequantize()[axis = tensor(0), name = tensor("linear0_1_weight_cast_fp16"), quantized_data = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(4207040))), scale = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8405568))), zero_point = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8401408)))]; + tensor linear0_2_weight_cast_fp16 = constexpr_affine_dequantize()[axis = tensor(0), name = tensor("linear0_2_weight_cast_fp16"), quantized_data = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8413824))), scale = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(12608192))), zero_point = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(4203840)))]; + tensor attn1_in_proj_weight_cast_fp16 = constexpr_affine_dequantize()[axis = tensor(0), name = tensor("attn1_in_proj_weight_cast_fp16"), quantized_data = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(12610304))), scale = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15756096))), zero_point = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(3145856)))]; + tensor attn1_out_proj_weight_cast_fp16 = constexpr_affine_dequantize()[axis = tensor(0), name = tensor("attn1_out_proj_weight_cast_fp16"), quantized_data = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15762304))), scale = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(16810944))), zero_point = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(4203840)))]; + tensor linear1_1_weight_cast_fp16 = constexpr_affine_dequantize()[axis = tensor(0), name = tensor("linear1_1_weight_cast_fp16"), quantized_data = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(16813056))), scale = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21007424))), zero_point = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8401408)))]; + tensor linear1_2_weight_cast_fp16 = constexpr_affine_dequantize()[axis = tensor(0), name = tensor("linear1_2_weight_cast_fp16"), quantized_data = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21015680))), scale = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(25210048))), zero_point = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(4203840)))]; + tensor attn2_in_proj_weight_cast_fp16 = constexpr_affine_dequantize()[axis = tensor(0), name = tensor("attn2_in_proj_weight_cast_fp16"), quantized_data = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(25212160))), scale = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28357952))), zero_point = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(3145856)))]; + tensor attn2_out_proj_weight_cast_fp16 = constexpr_affine_dequantize()[axis = tensor(0), name = tensor("attn2_out_proj_weight_cast_fp16"), quantized_data = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(28364160))), scale = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(29412800))), zero_point = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(4203840)))]; + tensor linear2_1_weight_cast_fp16 = constexpr_affine_dequantize()[axis = tensor(0), name = tensor("linear2_1_weight_cast_fp16"), quantized_data = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(29414912))), scale = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(33609280))), zero_point = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8401408)))]; + tensor linear2_2_weight_cast_fp16 = constexpr_affine_dequantize()[axis = tensor(0), name = tensor("linear2_2_weight_cast_fp16"), quantized_data = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(33617536))), scale = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(37811904))), zero_point = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(4203840)))]; + tensor attn3_in_proj_weight_cast_fp16 = constexpr_affine_dequantize()[axis = tensor(0), name = tensor("attn3_in_proj_weight_cast_fp16"), quantized_data = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(37814016))), scale = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(40959808))), zero_point = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(3145856)))]; + tensor attn3_out_proj_weight_cast_fp16 = constexpr_affine_dequantize()[axis = tensor(0), name = tensor("attn3_out_proj_weight_cast_fp16"), quantized_data = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(40966016))), scale = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(42014656))), zero_point = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(4203840)))]; + tensor linear3_1_weight_cast_fp16 = constexpr_affine_dequantize()[axis = tensor(0), name = tensor("linear3_1_weight_cast_fp16"), quantized_data = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(42016768))), scale = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(46211136))), zero_point = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8401408)))]; + tensor linear3_2_weight_cast_fp16 = constexpr_affine_dequantize()[axis = tensor(0), name = tensor("linear3_2_weight_cast_fp16"), quantized_data = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(46219392))), scale = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(50413760))), zero_point = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(4203840)))]; + tensor attn4_in_proj_weight_cast_fp16 = constexpr_affine_dequantize()[axis = tensor(0), name = tensor("attn4_in_proj_weight_cast_fp16"), quantized_data = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(50415872))), scale = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(53561664))), zero_point = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(3145856)))]; + tensor attn4_out_proj_weight_cast_fp16 = constexpr_affine_dequantize()[axis = tensor(0), name = tensor("attn4_out_proj_weight_cast_fp16"), quantized_data = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(53567872))), scale = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(54616512))), zero_point = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(4203840)))]; + tensor linear4_1_weight_cast_fp16 = constexpr_affine_dequantize()[axis = tensor(0), name = tensor("linear4_1_weight_cast_fp16"), quantized_data = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(54618624))), scale = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(58812992))), zero_point = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8401408)))]; + tensor linear4_2_weight_cast_fp16 = constexpr_affine_dequantize()[axis = tensor(0), name = tensor("linear4_2_weight_cast_fp16"), quantized_data = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(58821248))), scale = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(63015616))), zero_point = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(4203840)))]; + tensor attn5_in_proj_weight_cast_fp16 = constexpr_affine_dequantize()[axis = tensor(0), name = tensor("attn5_in_proj_weight_cast_fp16"), quantized_data = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(63017728))), scale = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(66163520))), zero_point = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(3145856)))]; + tensor attn5_out_proj_weight_cast_fp16 = constexpr_affine_dequantize()[axis = tensor(0), name = tensor("attn5_out_proj_weight_cast_fp16"), quantized_data = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(66169728))), scale = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(67218368))), zero_point = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(4203840)))]; + tensor linear5_1_weight_cast_fp16 = constexpr_affine_dequantize()[axis = tensor(0), name = tensor("linear5_1_weight_cast_fp16"), quantized_data = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(67220480))), scale = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(71414848))), zero_point = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8401408)))]; + tensor linear5_2_weight_cast_fp16 = constexpr_affine_dequantize()[axis = tensor(0), name = tensor("linear5_2_weight_cast_fp16"), quantized_data = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(71423104))), scale = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(75617472))), zero_point = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(4203840)))]; + tensor attn6_in_proj_weight_cast_fp16 = constexpr_affine_dequantize()[axis = tensor(0), name = tensor("attn6_in_proj_weight_cast_fp16"), quantized_data = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(75619584))), scale = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(78765376))), zero_point = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(3145856)))]; + tensor attn6_out_proj_weight_cast_fp16 = constexpr_affine_dequantize()[axis = tensor(0), name = tensor("attn6_out_proj_weight_cast_fp16"), quantized_data = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(78771584))), scale = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(79820224))), zero_point = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(4203840)))]; + tensor linear6_1_weight_cast_fp16 = constexpr_affine_dequantize()[axis = tensor(0), name = tensor("linear6_1_weight_cast_fp16"), quantized_data = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(79822336))), scale = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(84016704))), zero_point = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8401408)))]; + tensor linear6_2_weight_cast_fp16 = constexpr_affine_dequantize()[axis = tensor(0), name = tensor("linear6_2_weight_cast_fp16"), quantized_data = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(84024960))), scale = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(88219328))), zero_point = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(4203840)))]; + tensor attn7_in_proj_weight_cast_fp16 = constexpr_affine_dequantize()[axis = tensor(0), name = tensor("attn7_in_proj_weight_cast_fp16"), quantized_data = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(88221440))), scale = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(91367232))), zero_point = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(3145856)))]; + tensor attn7_out_proj_weight_cast_fp16 = constexpr_affine_dequantize()[axis = tensor(0), name = tensor("attn7_out_proj_weight_cast_fp16"), quantized_data = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(91373440))), scale = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(92422080))), zero_point = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(4203840)))]; + tensor linear7_1_weight_cast_fp16 = constexpr_affine_dequantize()[axis = tensor(0), name = tensor("linear7_1_weight_cast_fp16"), quantized_data = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(92424192))), scale = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(96618560))), zero_point = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8401408)))]; + tensor linear7_2_weight_cast_fp16 = constexpr_affine_dequantize()[axis = tensor(0), name = tensor("linear7_2_weight_cast_fp16"), quantized_data = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(96626816))), scale = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(100821184))), zero_point = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(4203840)))]; + tensor attn8_in_proj_weight_cast_fp16 = constexpr_affine_dequantize()[axis = tensor(0), name = tensor("attn8_in_proj_weight_cast_fp16"), quantized_data = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(100823296))), scale = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(103969088))), zero_point = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(3145856)))]; + tensor attn8_out_proj_weight_cast_fp16 = constexpr_affine_dequantize()[axis = tensor(0), name = tensor("attn8_out_proj_weight_cast_fp16"), quantized_data = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(103975296))), scale = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(105023936))), zero_point = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(4203840)))]; + tensor linear8_1_weight_cast_fp16 = constexpr_affine_dequantize()[axis = tensor(0), name = tensor("linear8_1_weight_cast_fp16"), quantized_data = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(105026048))), scale = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(109220416))), zero_point = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8401408)))]; + tensor linear8_2_weight_cast_fp16 = constexpr_affine_dequantize()[axis = tensor(0), name = tensor("linear8_2_weight_cast_fp16"), quantized_data = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(109228672))), scale = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(113423040))), zero_point = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(4203840)))]; + tensor attn9_in_proj_weight_cast_fp16 = constexpr_affine_dequantize()[axis = tensor(0), name = tensor("attn9_in_proj_weight_cast_fp16"), quantized_data = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(113425152))), scale = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(116570944))), zero_point = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(3145856)))]; + tensor attn9_out_proj_weight_cast_fp16 = constexpr_affine_dequantize()[axis = tensor(0), name = tensor("attn9_out_proj_weight_cast_fp16"), quantized_data = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(116577152))), scale = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(117625792))), zero_point = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(4203840)))]; + tensor linear9_1_weight_cast_fp16 = constexpr_affine_dequantize()[axis = tensor(0), name = tensor("linear9_1_weight_cast_fp16"), quantized_data = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(117627904))), scale = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(121822272))), zero_point = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8401408)))]; + tensor linear9_2_weight_cast_fp16 = constexpr_affine_dequantize()[axis = tensor(0), name = tensor("linear9_2_weight_cast_fp16"), quantized_data = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(121830528))), scale = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(126024896))), zero_point = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(4203840)))]; + tensor attn10_in_proj_weight_cast_fp16 = constexpr_affine_dequantize()[axis = tensor(0), name = tensor("attn10_in_proj_weight_cast_fp16"), quantized_data = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(126027008))), scale = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(129172800))), zero_point = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(3145856)))]; + tensor attn10_out_proj_weight_cast_fp16 = constexpr_affine_dequantize()[axis = tensor(0), name = tensor("attn10_out_proj_weight_cast_fp16"), quantized_data = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(129179008))), scale = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(130227648))), zero_point = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(4203840)))]; + tensor linear10_1_weight_cast_fp16 = constexpr_affine_dequantize()[axis = tensor(0), name = tensor("linear10_1_weight_cast_fp16"), quantized_data = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(130229760))), scale = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(134424128))), zero_point = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8401408)))]; + tensor linear10_2_weight_cast_fp16 = constexpr_affine_dequantize()[axis = tensor(0), name = tensor("linear10_2_weight_cast_fp16"), quantized_data = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(134432384))), scale = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(138626752))), zero_point = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(4203840)))]; + tensor attn11_in_proj_weight_cast_fp16 = constexpr_affine_dequantize()[axis = tensor(0), name = tensor("attn11_in_proj_weight_cast_fp16"), quantized_data = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(138628864))), scale = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(141774656))), zero_point = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(3145856)))]; + tensor attn11_out_proj_weight_cast_fp16 = constexpr_affine_dequantize()[axis = tensor(0), name = tensor("attn11_out_proj_weight_cast_fp16"), quantized_data = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(141780864))), scale = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(142829504))), zero_point = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(4203840)))]; + tensor linear11_1_weight_cast_fp16 = constexpr_affine_dequantize()[axis = tensor(0), name = tensor("linear11_1_weight_cast_fp16"), quantized_data = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(142831616))), scale = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(147025984))), zero_point = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8401408)))]; + tensor linear11_2_weight_cast_fp16 = constexpr_affine_dequantize()[axis = tensor(0), name = tensor("linear11_2_weight_cast_fp16"), quantized_data = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(147034240))), scale = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(151228608))), zero_point = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(4203840)))]; + tensor attn12_in_proj_weight_cast_fp16 = constexpr_affine_dequantize()[axis = tensor(0), name = tensor("attn12_in_proj_weight_cast_fp16"), quantized_data = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(151230720))), scale = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(154376512))), zero_point = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(3145856)))]; + tensor attn12_out_proj_weight_cast_fp16 = constexpr_affine_dequantize()[axis = tensor(0), name = tensor("attn12_out_proj_weight_cast_fp16"), quantized_data = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(154382720))), scale = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(155431360))), zero_point = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(4203840)))]; + tensor linear12_1_weight_cast_fp16 = constexpr_affine_dequantize()[axis = tensor(0), name = tensor("linear12_1_weight_cast_fp16"), quantized_data = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(155433472))), scale = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(159627840))), zero_point = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8401408)))]; + tensor linear12_2_weight_cast_fp16 = constexpr_affine_dequantize()[axis = tensor(0), name = tensor("linear12_2_weight_cast_fp16"), quantized_data = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(159636096))), scale = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(163830464))), zero_point = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(4203840)))]; + tensor attn13_in_proj_weight_cast_fp16 = constexpr_affine_dequantize()[axis = tensor(0), name = tensor("attn13_in_proj_weight_cast_fp16"), quantized_data = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(163832576))), scale = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(166978368))), zero_point = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(3145856)))]; + tensor attn13_out_proj_weight_cast_fp16 = constexpr_affine_dequantize()[axis = tensor(0), name = tensor("attn13_out_proj_weight_cast_fp16"), quantized_data = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(166984576))), scale = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(168033216))), zero_point = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(4203840)))]; + tensor linear13_1_weight_cast_fp16 = constexpr_affine_dequantize()[axis = tensor(0), name = tensor("linear13_1_weight_cast_fp16"), quantized_data = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(168035328))), scale = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(172229696))), zero_point = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8401408)))]; + tensor linear13_2_weight_cast_fp16 = constexpr_affine_dequantize()[axis = tensor(0), name = tensor("linear13_2_weight_cast_fp16"), quantized_data = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(172237952))), scale = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(176432320))), zero_point = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(4203840)))]; + tensor attn14_in_proj_weight_cast_fp16 = constexpr_affine_dequantize()[axis = tensor(0), name = tensor("attn14_in_proj_weight_cast_fp16"), quantized_data = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(176434432))), scale = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(179580224))), zero_point = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(3145856)))]; + tensor attn14_out_proj_weight_cast_fp16 = constexpr_affine_dequantize()[axis = tensor(0), name = tensor("attn14_out_proj_weight_cast_fp16"), quantized_data = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(179586432))), scale = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(180635072))), zero_point = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(4203840)))]; + tensor linear14_1_weight_cast_fp16 = constexpr_affine_dequantize()[axis = tensor(0), name = tensor("linear14_1_weight_cast_fp16"), quantized_data = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(180637184))), scale = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(184831552))), zero_point = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8401408)))]; + tensor linear14_2_weight_cast_fp16 = constexpr_affine_dequantize()[axis = tensor(0), name = tensor("linear14_2_weight_cast_fp16"), quantized_data = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(184839808))), scale = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(189034176))), zero_point = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(4203840)))]; + tensor attn15_in_proj_weight_cast_fp16 = constexpr_affine_dequantize()[axis = tensor(0), name = tensor("attn15_in_proj_weight_cast_fp16"), quantized_data = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(189036288))), scale = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(192182080))), zero_point = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(3145856)))]; + tensor attn15_out_proj_weight_cast_fp16 = constexpr_affine_dequantize()[axis = tensor(0), name = tensor("attn15_out_proj_weight_cast_fp16"), quantized_data = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(192188288))), scale = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(193236928))), zero_point = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(4203840)))]; + tensor linear15_1_weight_cast_fp16 = constexpr_affine_dequantize()[axis = tensor(0), name = tensor("linear15_1_weight_cast_fp16"), quantized_data = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(193239040))), scale = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(197433408))), zero_point = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8401408)))]; + tensor linear15_2_weight_cast_fp16 = constexpr_affine_dequantize()[axis = tensor(0), name = tensor("linear15_2_weight_cast_fp16"), quantized_data = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(197441664))), scale = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(201636032))), zero_point = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(4203840)))]; + tensor attn16_in_proj_weight_cast_fp16 = constexpr_affine_dequantize()[axis = tensor(0), name = tensor("attn16_in_proj_weight_cast_fp16"), quantized_data = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(201638144))), scale = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(204783936))), zero_point = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(3145856)))]; + tensor attn16_out_proj_weight_cast_fp16 = constexpr_affine_dequantize()[axis = tensor(0), name = tensor("attn16_out_proj_weight_cast_fp16"), quantized_data = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(204790144))), scale = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(205838784))), zero_point = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(4203840)))]; + tensor linear16_1_weight_cast_fp16 = constexpr_affine_dequantize()[axis = tensor(0), name = tensor("linear16_1_weight_cast_fp16"), quantized_data = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(205840896))), scale = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(210035264))), zero_point = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8401408)))]; + tensor linear16_2_weight_cast_fp16 = constexpr_affine_dequantize()[axis = tensor(0), name = tensor("linear16_2_weight_cast_fp16"), quantized_data = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(210043520))), scale = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(214237888))), zero_point = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(4203840)))]; + tensor attn17_in_proj_weight_cast_fp16 = constexpr_affine_dequantize()[axis = tensor(0), name = tensor("attn17_in_proj_weight_cast_fp16"), quantized_data = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(214240000))), scale = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(217385792))), zero_point = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(3145856)))]; + tensor attn17_out_proj_weight_cast_fp16 = constexpr_affine_dequantize()[axis = tensor(0), name = tensor("attn17_out_proj_weight_cast_fp16"), quantized_data = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(217392000))), scale = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(218440640))), zero_point = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(4203840)))]; + tensor linear17_1_weight_cast_fp16 = constexpr_affine_dequantize()[axis = tensor(0), name = tensor("linear17_1_weight_cast_fp16"), quantized_data = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(218442752))), scale = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(222637120))), zero_point = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8401408)))]; + tensor linear17_2_weight_cast_fp16 = constexpr_affine_dequantize()[axis = tensor(0), name = tensor("linear17_2_weight_cast_fp16"), quantized_data = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(222645376))), scale = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(226839744))), zero_point = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(4203840)))]; + tensor attn18_in_proj_weight_cast_fp16 = constexpr_affine_dequantize()[axis = tensor(0), name = tensor("attn18_in_proj_weight_cast_fp16"), quantized_data = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(226841856))), scale = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(229987648))), zero_point = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(3145856)))]; + tensor attn18_out_proj_weight_cast_fp16 = constexpr_affine_dequantize()[axis = tensor(0), name = tensor("attn18_out_proj_weight_cast_fp16"), quantized_data = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(229993856))), scale = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(231042496))), zero_point = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(4203840)))]; + tensor linear18_1_weight_cast_fp16 = constexpr_affine_dequantize()[axis = tensor(0), name = tensor("linear18_1_weight_cast_fp16"), quantized_data = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(231044608))), scale = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(235238976))), zero_point = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8401408)))]; + tensor linear18_2_weight_cast_fp16 = constexpr_affine_dequantize()[axis = tensor(0), name = tensor("linear18_2_weight_cast_fp16"), quantized_data = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(235247232))), scale = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(239441600))), zero_point = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(4203840)))]; + tensor attn19_in_proj_weight_cast_fp16 = constexpr_affine_dequantize()[axis = tensor(0), name = tensor("attn19_in_proj_weight_cast_fp16"), quantized_data = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(239443712))), scale = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(242589504))), zero_point = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(3145856)))]; + tensor attn19_out_proj_weight_cast_fp16 = constexpr_affine_dequantize()[axis = tensor(0), name = tensor("attn19_out_proj_weight_cast_fp16"), quantized_data = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(242595712))), scale = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(243644352))), zero_point = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(4203840)))]; + tensor linear19_1_weight_cast_fp16 = constexpr_affine_dequantize()[axis = tensor(0), name = tensor("linear19_1_weight_cast_fp16"), quantized_data = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(243646464))), scale = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(247840832))), zero_point = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8401408)))]; + tensor linear19_2_weight_cast_fp16 = constexpr_affine_dequantize()[axis = tensor(0), name = tensor("linear19_2_weight_cast_fp16"), quantized_data = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(247849088))), scale = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(252043456))), zero_point = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(4203840)))]; + tensor attn20_in_proj_weight_cast_fp16 = constexpr_affine_dequantize()[axis = tensor(0), name = tensor("attn20_in_proj_weight_cast_fp16"), quantized_data = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(252045568))), scale = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(255191360))), zero_point = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(3145856)))]; + tensor attn20_out_proj_weight_cast_fp16 = constexpr_affine_dequantize()[axis = tensor(0), name = tensor("attn20_out_proj_weight_cast_fp16"), quantized_data = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(255197568))), scale = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(256246208))), zero_point = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(4203840)))]; + tensor linear20_1_weight_cast_fp16 = constexpr_affine_dequantize()[axis = tensor(0), name = tensor("linear20_1_weight_cast_fp16"), quantized_data = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(256248320))), scale = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(260442688))), zero_point = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8401408)))]; + tensor linear20_2_weight_cast_fp16 = constexpr_affine_dequantize()[axis = tensor(0), name = tensor("linear20_2_weight_cast_fp16"), quantized_data = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(260450944))), scale = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(264645312))), zero_point = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(4203840)))]; + tensor attn21_in_proj_weight_cast_fp16 = constexpr_affine_dequantize()[axis = tensor(0), name = tensor("attn21_in_proj_weight_cast_fp16"), quantized_data = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(264647424))), scale = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(267793216))), zero_point = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(3145856)))]; + tensor attn21_out_proj_weight_cast_fp16 = constexpr_affine_dequantize()[axis = tensor(0), name = tensor("attn21_out_proj_weight_cast_fp16"), quantized_data = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(267799424))), scale = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(268848064))), zero_point = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(4203840)))]; + tensor linear21_1_weight_cast_fp16 = constexpr_affine_dequantize()[axis = tensor(0), name = tensor("linear21_1_weight_cast_fp16"), quantized_data = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(268850176))), scale = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(273044544))), zero_point = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8401408)))]; + tensor linear21_2_weight_cast_fp16 = constexpr_affine_dequantize()[axis = tensor(0), name = tensor("linear21_2_weight_cast_fp16"), quantized_data = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(273052800))), scale = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(277247168))), zero_point = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(4203840)))]; + tensor attn22_in_proj_weight_cast_fp16 = constexpr_affine_dequantize()[axis = tensor(0), name = tensor("attn22_in_proj_weight_cast_fp16"), quantized_data = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(277249280))), scale = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(280395072))), zero_point = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(3145856)))]; + tensor attn22_out_proj_weight_cast_fp16 = constexpr_affine_dequantize()[axis = tensor(0), name = tensor("attn22_out_proj_weight_cast_fp16"), quantized_data = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(280401280))), scale = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(281449920))), zero_point = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(4203840)))]; + tensor linear22_1_weight_cast_fp16 = constexpr_affine_dequantize()[axis = tensor(0), name = tensor("linear22_1_weight_cast_fp16"), quantized_data = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(281452032))), scale = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(285646400))), zero_point = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8401408)))]; + tensor linear22_2_weight_cast_fp16 = constexpr_affine_dequantize()[axis = tensor(0), name = tensor("linear22_2_weight_cast_fp16"), quantized_data = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(285654656))), scale = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(289849024))), zero_point = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(4203840)))]; + tensor attn23_in_proj_weight_cast_fp16 = constexpr_affine_dequantize()[axis = tensor(0), name = tensor("attn23_in_proj_weight_cast_fp16"), quantized_data = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(289851136))), scale = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(292996928))), zero_point = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(3145856)))]; + tensor attn23_out_proj_weight_cast_fp16 = constexpr_affine_dequantize()[axis = tensor(0), name = tensor("attn23_out_proj_weight_cast_fp16"), quantized_data = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(293003136))), scale = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(294051776))), zero_point = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(4203840)))]; + tensor linear23_1_weight_cast_fp16 = constexpr_affine_dequantize()[axis = tensor(0), name = tensor("linear23_1_weight_cast_fp16"), quantized_data = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(294053888))), scale = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(298248256))), zero_point = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8401408)))]; + tensor linear23_2_weight_cast_fp16 = constexpr_affine_dequantize()[axis = tensor(0), name = tensor("linear23_2_weight_cast_fp16"), quantized_data = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(298256512))), scale = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(302450880))), zero_point = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(4203840)))]; + tensor sequence_to_fp16_dtype_0 = const()[name = tensor("sequence_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor sequence_to_fp16 = cast(dtype = sequence_to_fp16_dtype_0, x = sequence)[name = tensor("cast_555")]; + tensor var_198_cast_fp16 = not_equal(x = sequence_to_fp16, y = sequence_to_fp16)[name = tensor("op_198_cast_fp16")]; + tensor expand_dims_0_axes_0 = const()[name = tensor("expand_dims_0_axes_0"), val = tensor([0, 1])]; + tensor bos_emb_to_fp16_dtype_0 = const()[name = tensor("bos_emb_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor bos_emb_to_fp16 = cast(dtype = bos_emb_to_fp16_dtype_0, x = bos_emb)[name = tensor("cast_554")]; + tensor expand_dims_0_cast_fp16 = expand_dims(axes = expand_dims_0_axes_0, x = bos_emb_to_fp16)[name = tensor("expand_dims_0_cast_fp16")]; + tensor input_1_cast_fp16 = select(a = expand_dims_0_cast_fp16, b = sequence_to_fp16, cond = var_198_cast_fp16)[name = tensor("input_1_cast_fp16")]; + tensor input_linear_weight_to_fp16 = const()[name = tensor("input_linear_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(302452992)))]; + tensor linear_0_bias_0_to_fp16 = const()[name = tensor("linear_0_bias_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(302518592)))]; + tensor linear_0_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = input_linear_weight_to_fp16, x = input_1_cast_fp16)[name = tensor("linear_0_cast_fp16")]; + tensor x_1_axes_0 = const()[name = tensor("x_1_axes_0"), val = tensor([-1])]; + tensor norm0_1_weight_to_fp16 = const()[name = tensor("norm0_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(302520704)))]; + tensor norm0_1_bias_to_fp16 = const()[name = tensor("norm0_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(302522816)))]; + tensor var_204_to_fp16 = const()[name = tensor("op_204_to_fp16"), val = tensor(0x1.5p-17)]; + tensor x_1_cast_fp16 = layer_norm(axes = x_1_axes_0, beta = norm0_1_bias_to_fp16, epsilon = var_204_to_fp16, gamma = norm0_1_weight_to_fp16, x = linear_0_cast_fp16)[name = tensor("x_1_cast_fp16")]; + tensor linear_1_bias_0_to_fp16 = const()[name = tensor("linear_1_bias_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(302524928)))]; + tensor linear_1_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = attn0_in_proj_weight_cast_fp16, x = x_1_cast_fp16)[name = tensor("linear_1_cast_fp16")]; + tensor var_240 = const()[name = tensor("op_240"), val = tensor([1, 1, 3, 16, 64])]; + tensor qkv_1_cast_fp16 = reshape(shape = var_240, x = linear_1_cast_fp16)[name = tensor("qkv_1_cast_fp16")]; + tensor q_1_begin_0 = const()[name = tensor("q_1_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor q_1_end_0 = const()[name = tensor("q_1_end_0"), val = tensor([1, 1, 1, 16, 64])]; + tensor q_1_end_mask_0 = const()[name = tensor("q_1_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor q_1_squeeze_mask_0 = const()[name = tensor("q_1_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor q_1_cast_fp16 = slice_by_index(begin = q_1_begin_0, end = q_1_end_0, end_mask = q_1_end_mask_0, squeeze_mask = q_1_squeeze_mask_0, x = qkv_1_cast_fp16)[name = tensor("q_1_cast_fp16")]; + tensor k_1_begin_0 = const()[name = tensor("k_1_begin_0"), val = tensor([0, 0, 1, 0, 0])]; + tensor k_1_end_0 = const()[name = tensor("k_1_end_0"), val = tensor([1, 1, 2, 16, 64])]; + tensor k_1_end_mask_0 = const()[name = tensor("k_1_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor k_1_squeeze_mask_0 = const()[name = tensor("k_1_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor k_1_cast_fp16 = slice_by_index(begin = k_1_begin_0, end = k_1_end_0, end_mask = k_1_end_mask_0, squeeze_mask = k_1_squeeze_mask_0, x = qkv_1_cast_fp16)[name = tensor("k_1_cast_fp16")]; + tensor v_1_begin_0 = const()[name = tensor("v_1_begin_0"), val = tensor([0, 0, 2, 0, 0])]; + tensor v_1_end_0 = const()[name = tensor("v_1_end_0"), val = tensor([1, 1, 3, 16, 64])]; + tensor v_1_end_mask_0 = const()[name = tensor("v_1_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor v_1_squeeze_mask_0 = const()[name = tensor("v_1_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor v_1_cast_fp16 = slice_by_index(begin = v_1_begin_0, end = v_1_end_0, end_mask = v_1_end_mask_0, squeeze_mask = v_1_squeeze_mask_0, x = qkv_1_cast_fp16)[name = tensor("v_1_cast_fp16")]; + tensor var_344 = const()[name = tensor("op_344"), val = tensor([1, 1, 1, 1])]; + tensor position0_to_fp16_dtype_0 = const()[name = tensor("position0_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor position0_to_fp16 = cast(dtype = position0_to_fp16_dtype_0, x = position0)[name = tensor("cast_553")]; + tensor ts_5_cast_fp16 = reshape(shape = var_344, x = position0_to_fp16)[name = tensor("ts_5_cast_fp16")]; + tensor var_348 = const()[name = tensor("op_348"), val = tensor([1, 1, 16, 32, 2])]; + tensor q_complex_1_cast_fp16 = reshape(shape = var_348, x = q_1_cast_fp16)[name = tensor("q_complex_1_cast_fp16")]; + tensor var_352 = const()[name = tensor("op_352"), val = tensor([1, 1, 16, 32, 2])]; + tensor k_complex_1_cast_fp16 = reshape(shape = var_352, x = k_1_cast_fp16)[name = tensor("k_complex_1_cast_fp16")]; + tensor var_356_begin_0 = const()[name = tensor("op_356_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_356_end_0 = const()[name = tensor("op_356_end_0"), val = tensor([1, 1, 16, 32, 1])]; + tensor var_356_end_mask_0 = const()[name = tensor("op_356_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_356_squeeze_mask_0 = const()[name = tensor("op_356_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_356_cast_fp16 = slice_by_index(begin = var_356_begin_0, end = var_356_end_0, end_mask = var_356_end_mask_0, squeeze_mask = var_356_squeeze_mask_0, x = q_complex_1_cast_fp16)[name = tensor("op_356_cast_fp16")]; + tensor var_364_begin_0 = const()[name = tensor("op_364_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_364_end_0 = const()[name = tensor("op_364_end_0"), val = tensor([1, 1, 16, 32, 2])]; + tensor var_364_end_mask_0 = const()[name = tensor("op_364_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_364_squeeze_mask_0 = const()[name = tensor("op_364_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_364_cast_fp16 = slice_by_index(begin = var_364_begin_0, end = var_364_end_0, end_mask = var_364_end_mask_0, squeeze_mask = var_364_squeeze_mask_0, x = q_complex_1_cast_fp16)[name = tensor("op_364_cast_fp16")]; + tensor var_372_begin_0 = const()[name = tensor("op_372_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_372_end_0 = const()[name = tensor("op_372_end_0"), val = tensor([1, 1, 16, 32, 1])]; + tensor var_372_end_mask_0 = const()[name = tensor("op_372_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_372_squeeze_mask_0 = const()[name = tensor("op_372_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_372_cast_fp16 = slice_by_index(begin = var_372_begin_0, end = var_372_end_0, end_mask = var_372_end_mask_0, squeeze_mask = var_372_squeeze_mask_0, x = k_complex_1_cast_fp16)[name = tensor("op_372_cast_fp16")]; + tensor var_380_begin_0 = const()[name = tensor("op_380_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_380_end_0 = const()[name = tensor("op_380_end_0"), val = tensor([1, 1, 16, 32, 2])]; + tensor var_380_end_mask_0 = const()[name = tensor("op_380_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_380_squeeze_mask_0 = const()[name = tensor("op_380_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_380_cast_fp16 = slice_by_index(begin = var_380_begin_0, end = var_380_end_0, end_mask = var_380_end_mask_0, squeeze_mask = var_380_squeeze_mask_0, x = k_complex_1_cast_fp16)[name = tensor("op_380_cast_fp16")]; + tensor freqs_1_to_fp16 = const()[name = tensor("freqs_1_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(302531136)))]; + tensor var_386_cast_fp16 = mul(x = freqs_1_to_fp16, y = ts_5_cast_fp16)[name = tensor("op_386_cast_fp16")]; + tensor rotr_1_cast_fp16 = cos(x = var_386_cast_fp16)[name = tensor("rotr_1_cast_fp16")]; + tensor roti_1_cast_fp16 = sin(x = var_386_cast_fp16)[name = tensor("roti_1_cast_fp16")]; + tensor var_390_cast_fp16 = mul(x = var_356_cast_fp16, y = rotr_1_cast_fp16)[name = tensor("op_390_cast_fp16")]; + tensor var_391_cast_fp16 = mul(x = var_364_cast_fp16, y = roti_1_cast_fp16)[name = tensor("op_391_cast_fp16")]; + tensor qor_1_cast_fp16 = sub(x = var_390_cast_fp16, y = var_391_cast_fp16)[name = tensor("qor_1_cast_fp16")]; + tensor var_394_cast_fp16 = mul(x = var_356_cast_fp16, y = roti_1_cast_fp16)[name = tensor("op_394_cast_fp16")]; + tensor var_395_cast_fp16 = mul(x = var_364_cast_fp16, y = rotr_1_cast_fp16)[name = tensor("op_395_cast_fp16")]; + tensor qoi_1_cast_fp16 = add(x = var_394_cast_fp16, y = var_395_cast_fp16)[name = tensor("qoi_1_cast_fp16")]; + tensor var_398_cast_fp16 = mul(x = var_372_cast_fp16, y = rotr_1_cast_fp16)[name = tensor("op_398_cast_fp16")]; + tensor var_399_cast_fp16 = mul(x = var_380_cast_fp16, y = roti_1_cast_fp16)[name = tensor("op_399_cast_fp16")]; + tensor kor_1_cast_fp16 = sub(x = var_398_cast_fp16, y = var_399_cast_fp16)[name = tensor("kor_1_cast_fp16")]; + tensor var_402_cast_fp16 = mul(x = var_372_cast_fp16, y = roti_1_cast_fp16)[name = tensor("op_402_cast_fp16")]; + tensor var_403_cast_fp16 = mul(x = var_380_cast_fp16, y = rotr_1_cast_fp16)[name = tensor("op_403_cast_fp16")]; + tensor koi_1_cast_fp16 = add(x = var_402_cast_fp16, y = var_403_cast_fp16)[name = tensor("koi_1_cast_fp16")]; + tensor qo_1_axis_0 = const()[name = tensor("qo_1_axis_0"), val = tensor(-1)]; + tensor qo_1_cast_fp16 = stack(axis = qo_1_axis_0, values = (qor_1_cast_fp16, qoi_1_cast_fp16))[name = tensor("qo_1_cast_fp16")]; + tensor ko_1_axis_0 = const()[name = tensor("ko_1_axis_0"), val = tensor(-1)]; + tensor ko_1_cast_fp16 = stack(axis = ko_1_axis_0, values = (kor_1_cast_fp16, koi_1_cast_fp16))[name = tensor("ko_1_cast_fp16")]; + tensor var_432 = const()[name = tensor("op_432"), val = tensor([1, 1, 16, 64])]; + tensor q_3_cast_fp16 = reshape(shape = var_432, x = qo_1_cast_fp16)[name = tensor("q_3_cast_fp16")]; + tensor var_434 = const()[name = tensor("op_434"), val = tensor([1, 1, 16, 64])]; + tensor k_3_cast_fp16 = reshape(shape = var_434, x = ko_1_cast_fp16)[name = tensor("k_3_cast_fp16")]; + tensor _inversed_456_y_0_to_fp16 = const()[name = tensor("_inversed_456_y_0_to_fp16"), val = tensor(0x1p-9)]; + tensor _inversed_456_cast_fp16 = mul(x = ts_5_cast_fp16, y = _inversed_456_y_0_to_fp16)[name = tensor("_inversed_456_cast_fp16")]; + tensor var_457_cast_fp16 = floor(x = _inversed_456_cast_fp16)[name = tensor("op_457_cast_fp16")]; + tensor var_458_to_fp16 = const()[name = tensor("op_458_to_fp16"), val = tensor(0x1p+9)]; + tensor var_459_cast_fp16 = mul(x = var_457_cast_fp16, y = var_458_to_fp16)[name = tensor("op_459_cast_fp16")]; + tensor write_indices_float_3_cast_fp16 = sub(x = ts_5_cast_fp16, y = var_459_cast_fp16)[name = tensor("write_indices_float_3_cast_fp16")]; + tensor var_466_dtype_0 = const()[name = tensor("op_466_dtype_0"), val = tensor("int32")]; + tensor write_indices_1_reps_0 = const()[name = tensor("write_indices_1_reps_0"), val = tensor([1, 1, 16, 64])]; + tensor write_indices_float_3_cast_fp16_to_int32 = cast(dtype = var_466_dtype_0, x = write_indices_float_3_cast_fp16)[name = tensor("cast_552")]; + tensor write_indices_1 = tile(reps = write_indices_1_reps_0, x = write_indices_float_3_cast_fp16_to_int32)[name = tensor("write_indices_1")]; + tensor var_474_begin_0 = const()[name = tensor("op_474_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_474_end_0 = const()[name = tensor("op_474_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor var_474_end_mask_0 = const()[name = tensor("op_474_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_474_squeeze_mask_0 = const()[name = tensor("op_474_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor cache0_to_fp16_dtype_0 = const()[name = tensor("cache0_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor cache0_to_fp16 = cast(dtype = cache0_to_fp16_dtype_0, x = cache0)[name = tensor("cast_551")]; + tensor var_474_cast_fp16 = slice_by_index(begin = var_474_begin_0, end = var_474_end_0, end_mask = var_474_end_mask_0, squeeze_mask = var_474_squeeze_mask_0, x = cache0_to_fp16)[name = tensor("op_474_cast_fp16")]; + tensor var_476_axis_0 = const()[name = tensor("op_476_axis_0"), val = tensor(1)]; + tensor var_476_mode_0 = const()[name = tensor("op_476_mode_0"), val = tensor("update")]; + tensor var_476_validate_indices_0 = const()[name = tensor("op_476_validate_indices_0"), val = tensor(false)]; + tensor var_476_cast_fp16 = scatter_along_axis(axis = var_476_axis_0, data = var_474_cast_fp16, indices = write_indices_1, mode = var_476_mode_0, updates = k_3_cast_fp16, validate_indices = var_476_validate_indices_0)[name = tensor("op_476_cast_fp16")]; + tensor concat_2 = const()[name = tensor("concat_2"), val = tensor([0, 0, 0, 0, 0])]; + tensor concat_3 = const()[name = tensor("concat_3"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_1_internal_tensor_assign_1_stride_0 = const()[name = tensor("new_cache_1_internal_tensor_assign_1_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_1_internal_tensor_assign_1_begin_mask_0 = const()[name = tensor("new_cache_1_internal_tensor_assign_1_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_1_internal_tensor_assign_1_end_mask_0 = const()[name = tensor("new_cache_1_internal_tensor_assign_1_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_1_internal_tensor_assign_1_squeeze_mask_0 = const()[name = tensor("new_cache_1_internal_tensor_assign_1_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_48 = const()[name = tensor("shape_48"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_0 = const()[name = tensor("reduce_prod_0"), val = tensor(1048576)]; + tensor range_1d_0_start_0 = const()[name = tensor("range_1d_0_start_0"), val = tensor(0)]; + tensor range_1d_0_step_0 = const()[name = tensor("range_1d_0_step_0"), val = tensor(1)]; + tensor range_1d_0 = range_1d(end = reduce_prod_0, start = range_1d_0_start_0, step = range_1d_0_step_0)[name = tensor("range_1d_0")]; + tensor reshape_0 = reshape(shape = shape_48, x = range_1d_0)[name = tensor("reshape_0")]; + tensor slice_by_index_0 = slice_by_index(begin = concat_2, begin_mask = new_cache_1_internal_tensor_assign_1_begin_mask_0, end = concat_3, end_mask = new_cache_1_internal_tensor_assign_1_end_mask_0, squeeze_mask = new_cache_1_internal_tensor_assign_1_squeeze_mask_0, stride = new_cache_1_internal_tensor_assign_1_stride_0, x = reshape_0)[name = tensor("slice_by_index_0")]; + tensor reshape_1_shape_0 = const()[name = tensor("reshape_1_shape_0"), val = tensor([-1])]; + tensor reshape_1 = reshape(shape = reshape_1_shape_0, x = slice_by_index_0)[name = tensor("reshape_1")]; + tensor reshape_2_shape_0 = const()[name = tensor("reshape_2_shape_0"), val = tensor([-1])]; + tensor reshape_2_cast_fp16 = reshape(shape = reshape_2_shape_0, x = var_476_cast_fp16)[name = tensor("reshape_2_cast_fp16")]; + tensor reshape_3_shape_0 = const()[name = tensor("reshape_3_shape_0"), val = tensor([-1])]; + tensor reshape_3_cast_fp16 = reshape(shape = reshape_3_shape_0, x = cache0_to_fp16)[name = tensor("reshape_3_cast_fp16")]; + tensor scatter_0_mode_0 = const()[name = tensor("scatter_0_mode_0"), val = tensor("update")]; + tensor scatter_0_axis_0 = const()[name = tensor("scatter_0_axis_0"), val = tensor(0)]; + tensor scatter_0_validate_indices_0 = const()[name = tensor("scatter_0_validate_indices_0"), val = tensor(false)]; + tensor scatter_0_cast_fp16 = scatter(axis = scatter_0_axis_0, data = reshape_3_cast_fp16, indices = reshape_1, mode = scatter_0_mode_0, updates = reshape_2_cast_fp16, validate_indices = scatter_0_validate_indices_0)[name = tensor("scatter_0_cast_fp16")]; + tensor reshape_4_cast_fp16 = reshape(shape = shape_48, x = scatter_0_cast_fp16)[name = tensor("reshape_4_cast_fp16")]; + tensor var_484_begin_0 = const()[name = tensor("op_484_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor var_484_end_0 = const()[name = tensor("op_484_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor var_484_end_mask_0 = const()[name = tensor("op_484_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_484_squeeze_mask_0 = const()[name = tensor("op_484_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor var_484_cast_fp16 = slice_by_index(begin = var_484_begin_0, end = var_484_end_0, end_mask = var_484_end_mask_0, squeeze_mask = var_484_squeeze_mask_0, x = reshape_4_cast_fp16)[name = tensor("op_484_cast_fp16")]; + tensor var_486_axis_0 = const()[name = tensor("op_486_axis_0"), val = tensor(1)]; + tensor var_486_mode_0 = const()[name = tensor("op_486_mode_0"), val = tensor("update")]; + tensor var_486_validate_indices_0 = const()[name = tensor("op_486_validate_indices_0"), val = tensor(false)]; + tensor var_486_cast_fp16 = scatter_along_axis(axis = var_486_axis_0, data = var_484_cast_fp16, indices = write_indices_1, mode = var_486_mode_0, updates = v_1_cast_fp16, validate_indices = var_486_validate_indices_0)[name = tensor("op_486_cast_fp16")]; + tensor concat_4 = const()[name = tensor("concat_4"), val = tensor([1, 0, 0, 0, 0])]; + tensor concat_5 = const()[name = tensor("concat_5"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_1_internal_tensor_assign_2_stride_0 = const()[name = tensor("new_cache_1_internal_tensor_assign_2_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_1_internal_tensor_assign_2_begin_mask_0 = const()[name = tensor("new_cache_1_internal_tensor_assign_2_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_1_internal_tensor_assign_2_end_mask_0 = const()[name = tensor("new_cache_1_internal_tensor_assign_2_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_1_internal_tensor_assign_2_squeeze_mask_0 = const()[name = tensor("new_cache_1_internal_tensor_assign_2_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_49 = const()[name = tensor("shape_49"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_1 = const()[name = tensor("reduce_prod_1"), val = tensor(1048576)]; + tensor range_1d_1_start_0 = const()[name = tensor("range_1d_1_start_0"), val = tensor(0)]; + tensor range_1d_1_step_0 = const()[name = tensor("range_1d_1_step_0"), val = tensor(1)]; + tensor range_1d_1 = range_1d(end = reduce_prod_1, start = range_1d_1_start_0, step = range_1d_1_step_0)[name = tensor("range_1d_1")]; + tensor reshape_5 = reshape(shape = shape_49, x = range_1d_1)[name = tensor("reshape_5")]; + tensor slice_by_index_1 = slice_by_index(begin = concat_4, begin_mask = new_cache_1_internal_tensor_assign_2_begin_mask_0, end = concat_5, end_mask = new_cache_1_internal_tensor_assign_2_end_mask_0, squeeze_mask = new_cache_1_internal_tensor_assign_2_squeeze_mask_0, stride = new_cache_1_internal_tensor_assign_2_stride_0, x = reshape_5)[name = tensor("slice_by_index_1")]; + tensor reshape_6_shape_0 = const()[name = tensor("reshape_6_shape_0"), val = tensor([-1])]; + tensor reshape_6 = reshape(shape = reshape_6_shape_0, x = slice_by_index_1)[name = tensor("reshape_6")]; + tensor reshape_7_shape_0 = const()[name = tensor("reshape_7_shape_0"), val = tensor([-1])]; + tensor reshape_7_cast_fp16 = reshape(shape = reshape_7_shape_0, x = var_486_cast_fp16)[name = tensor("reshape_7_cast_fp16")]; + tensor reshape_8_shape_0 = const()[name = tensor("reshape_8_shape_0"), val = tensor([-1])]; + tensor reshape_8_cast_fp16 = reshape(shape = reshape_8_shape_0, x = reshape_4_cast_fp16)[name = tensor("reshape_8_cast_fp16")]; + tensor scatter_1_mode_0 = const()[name = tensor("scatter_1_mode_0"), val = tensor("update")]; + tensor scatter_1_axis_0 = const()[name = tensor("scatter_1_axis_0"), val = tensor(0)]; + tensor scatter_1_validate_indices_0 = const()[name = tensor("scatter_1_validate_indices_0"), val = tensor(false)]; + tensor scatter_1_cast_fp16 = scatter(axis = scatter_1_axis_0, data = reshape_8_cast_fp16, indices = reshape_6, mode = scatter_1_mode_0, updates = reshape_7_cast_fp16, validate_indices = scatter_1_validate_indices_0)[name = tensor("scatter_1_cast_fp16")]; + tensor reshape_9_cast_fp16 = reshape(shape = shape_49, x = scatter_1_cast_fp16)[name = tensor("reshape_9_cast_fp16")]; + tensor reshape_9_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("reshape_9_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor keys_1_begin_0 = const()[name = tensor("keys_1_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor keys_1_end_0 = const()[name = tensor("keys_1_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor keys_1_end_mask_0 = const()[name = tensor("keys_1_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor keys_1_squeeze_mask_0 = const()[name = tensor("keys_1_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor keys_1_cast_fp16 = slice_by_index(begin = keys_1_begin_0, end = keys_1_end_0, end_mask = keys_1_end_mask_0, squeeze_mask = keys_1_squeeze_mask_0, x = reshape_9_cast_fp16)[name = tensor("keys_1_cast_fp16")]; + tensor values_1_begin_0 = const()[name = tensor("values_1_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor values_1_end_0 = const()[name = tensor("values_1_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor values_1_end_mask_0 = const()[name = tensor("values_1_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor values_1_squeeze_mask_0 = const()[name = tensor("values_1_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor values_1_cast_fp16 = slice_by_index(begin = values_1_begin_0, end = values_1_end_0, end_mask = values_1_end_mask_0, squeeze_mask = values_1_squeeze_mask_0, x = reshape_9_cast_fp16)[name = tensor("values_1_cast_fp16")]; + tensor var_498_cast_fp16 = not_equal(x = keys_1_cast_fp16, y = keys_1_cast_fp16)[name = tensor("op_498_cast_fp16")]; + tensor var_504_to_fp16 = const()[name = tensor("op_504_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(302531264)))]; + tensor keys_3_cast_fp16 = select(a = var_504_to_fp16, b = keys_1_cast_fp16, cond = var_498_cast_fp16)[name = tensor("keys_3_cast_fp16")]; + tensor var_506_cast_fp16 = not_equal(x = values_1_cast_fp16, y = values_1_cast_fp16)[name = tensor("op_506_cast_fp16")]; + tensor values_3_cast_fp16 = select(a = var_504_to_fp16, b = values_1_cast_fp16, cond = var_506_cast_fp16)[name = tensor("values_3_cast_fp16")]; + tensor var_530 = const()[name = tensor("op_530"), val = tensor([0, 2, 1, 3])]; + tensor var_543 = const()[name = tensor("op_543"), val = tensor([1, 1, 1])]; + tensor var_544_cast_fp16 = reshape(shape = var_543, x = position0_to_fp16)[name = tensor("op_544_cast_fp16")]; + tensor var_561_to_fp16 = const()[name = tensor("op_561_to_fp16"), val = tensor(0x1p+0)]; + tensor valid_len_1_cast_fp16 = add(x = var_544_cast_fp16, y = var_561_to_fp16)[name = tensor("valid_len_1_cast_fp16")]; + tensor k_positions_1_promoted_to_fp16 = const()[name = tensor("k_positions_1_promoted_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(303579904)))]; + tensor valid_mask_1_cast_fp16 = less(x = k_positions_1_promoted_to_fp16, y = valid_len_1_cast_fp16)[name = tensor("valid_mask_1_cast_fp16")]; + tensor causal_mask_1_cast_fp16 = less_equal(x = k_positions_1_promoted_to_fp16, y = var_544_cast_fp16)[name = tensor("causal_mask_1_cast_fp16")]; + tensor attn_mask_1 = logical_and(x = valid_mask_1_cast_fp16, y = causal_mask_1_cast_fp16)[name = tensor("attn_mask_1")]; + tensor attn_mask_3_axes_0 = const()[name = tensor("attn_mask_3_axes_0"), val = tensor([1])]; + tensor attn_mask_3 = expand_dims(axes = attn_mask_3_axes_0, x = attn_mask_1)[name = tensor("attn_mask_3")]; + tensor var_579_transpose_x_0 = const()[name = tensor("op_579_transpose_x_0"), val = tensor(false)]; + tensor var_579_transpose_y_0 = const()[name = tensor("op_579_transpose_y_0"), val = tensor(false)]; + tensor transpose_72_perm_0 = const()[name = tensor("transpose_72_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_73_perm_0 = const()[name = tensor("transpose_73_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_73 = transpose(perm = transpose_73_perm_0, x = keys_3_cast_fp16)[name = tensor("transpose_213")]; + tensor transpose_72 = transpose(perm = transpose_72_perm_0, x = q_3_cast_fp16)[name = tensor("transpose_214")]; + tensor var_579_cast_fp16 = matmul(transpose_x = var_579_transpose_x_0, transpose_y = var_579_transpose_y_0, x = transpose_72, y = transpose_73)[name = tensor("op_579_cast_fp16")]; + tensor var_573_to_fp16 = const()[name = tensor("op_573_to_fp16"), val = tensor([0x1p-3])]; + tensor attn_weights_1_cast_fp16 = mul(x = var_579_cast_fp16, y = var_573_to_fp16)[name = tensor("attn_weights_1_cast_fp16")]; + tensor var_581 = logical_not(x = attn_mask_3)[name = tensor("op_581")]; + tensor var_582_to_fp16 = const()[name = tensor("op_582_to_fp16"), val = tensor(-inf)]; + tensor attn_weights_3_cast_fp16 = select(a = var_582_to_fp16, b = attn_weights_1_cast_fp16, cond = var_581)[name = tensor("attn_weights_3_cast_fp16")]; + tensor var_584 = const()[name = tensor("op_584"), val = tensor(-1)]; + tensor attn_weights_5_cast_fp16 = softmax(axis = var_584, x = attn_weights_3_cast_fp16)[name = tensor("attn_weights_5_cast_fp16")]; + tensor attn_output_1_transpose_x_0 = const()[name = tensor("attn_output_1_transpose_x_0"), val = tensor(false)]; + tensor attn_output_1_transpose_y_0 = const()[name = tensor("attn_output_1_transpose_y_0"), val = tensor(false)]; + tensor values_5_cast_fp16 = transpose(perm = var_530, x = values_3_cast_fp16)[name = tensor("transpose_215")]; + tensor attn_output_1_cast_fp16 = matmul(transpose_x = attn_output_1_transpose_x_0, transpose_y = attn_output_1_transpose_y_0, x = attn_weights_5_cast_fp16, y = values_5_cast_fp16)[name = tensor("attn_output_1_cast_fp16")]; + tensor var_592 = const()[name = tensor("op_592"), val = tensor([0, 2, 1, 3])]; + tensor var_595 = const()[name = tensor("op_595"), val = tensor([1, 1, 1024])]; + tensor var_593_cast_fp16 = transpose(perm = var_592, x = attn_output_1_cast_fp16)[name = tensor("transpose_212")]; + tensor input_5_cast_fp16 = reshape(shape = var_595, x = var_593_cast_fp16)[name = tensor("input_5_cast_fp16")]; + tensor linear_2_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = attn0_out_proj_weight_cast_fp16, x = input_5_cast_fp16)[name = tensor("linear_2_cast_fp16")]; + tensor var_601_to_fp16 = const()[name = tensor("op_601_to_fp16"), val = tensor(0x1p+0)]; + tensor var_602_cast_fp16 = add(x = position0_to_fp16, y = var_601_to_fp16)[name = tensor("op_602_cast_fp16")]; + tensor var_602_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_602_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor input_7_cast_fp16 = add(x = linear_0_cast_fp16, y = linear_2_cast_fp16)[name = tensor("input_7_cast_fp16")]; + tensor input_9_axes_0 = const()[name = tensor("input_9_axes_0"), val = tensor([-1])]; + tensor norm0_2_weight_to_fp16 = const()[name = tensor("norm0_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(303580992)))]; + tensor norm0_2_bias_to_fp16 = const()[name = tensor("norm0_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(303583104)))]; + tensor var_606_to_fp16 = const()[name = tensor("op_606_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_9_cast_fp16 = layer_norm(axes = input_9_axes_0, beta = norm0_2_bias_to_fp16, epsilon = var_606_to_fp16, gamma = norm0_2_weight_to_fp16, x = input_7_cast_fp16)[name = tensor("input_9_cast_fp16")]; + tensor linear_3_bias_0_to_fp16 = const()[name = tensor("linear_3_bias_0_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(303585216)))]; + tensor linear_3_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = linear0_1_weight_cast_fp16, x = input_9_cast_fp16)[name = tensor("linear_3_cast_fp16")]; + tensor input_11_mode_0 = const()[name = tensor("input_11_mode_0"), val = tensor("EXACT")]; + tensor input_11_cast_fp16 = gelu(mode = input_11_mode_0, x = linear_3_cast_fp16)[name = tensor("input_11_cast_fp16")]; + tensor linear_4_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = linear0_2_weight_cast_fp16, x = input_11_cast_fp16)[name = tensor("linear_4_cast_fp16")]; + tensor input_13_cast_fp16 = add(x = input_7_cast_fp16, y = linear_4_cast_fp16)[name = tensor("input_13_cast_fp16")]; + tensor x_3_axes_0 = const()[name = tensor("x_3_axes_0"), val = tensor([-1])]; + tensor norm1_1_weight_to_fp16 = const()[name = tensor("norm1_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(303593472)))]; + tensor norm1_1_bias_to_fp16 = const()[name = tensor("norm1_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(303595584)))]; + tensor var_623_to_fp16 = const()[name = tensor("op_623_to_fp16"), val = tensor(0x1.5p-17)]; + tensor x_3_cast_fp16 = layer_norm(axes = x_3_axes_0, beta = norm1_1_bias_to_fp16, epsilon = var_623_to_fp16, gamma = norm1_1_weight_to_fp16, x = input_13_cast_fp16)[name = tensor("x_3_cast_fp16")]; + tensor linear_5_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = attn1_in_proj_weight_cast_fp16, x = x_3_cast_fp16)[name = tensor("linear_5_cast_fp16")]; + tensor var_659 = const()[name = tensor("op_659"), val = tensor([1, 1, 3, 16, 64])]; + tensor qkv_3_cast_fp16 = reshape(shape = var_659, x = linear_5_cast_fp16)[name = tensor("qkv_3_cast_fp16")]; + tensor q_7_begin_0 = const()[name = tensor("q_7_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor q_7_end_0 = const()[name = tensor("q_7_end_0"), val = tensor([1, 1, 1, 16, 64])]; + tensor q_7_end_mask_0 = const()[name = tensor("q_7_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor q_7_squeeze_mask_0 = const()[name = tensor("q_7_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor q_7_cast_fp16 = slice_by_index(begin = q_7_begin_0, end = q_7_end_0, end_mask = q_7_end_mask_0, squeeze_mask = q_7_squeeze_mask_0, x = qkv_3_cast_fp16)[name = tensor("q_7_cast_fp16")]; + tensor k_5_begin_0 = const()[name = tensor("k_5_begin_0"), val = tensor([0, 0, 1, 0, 0])]; + tensor k_5_end_0 = const()[name = tensor("k_5_end_0"), val = tensor([1, 1, 2, 16, 64])]; + tensor k_5_end_mask_0 = const()[name = tensor("k_5_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor k_5_squeeze_mask_0 = const()[name = tensor("k_5_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor k_5_cast_fp16 = slice_by_index(begin = k_5_begin_0, end = k_5_end_0, end_mask = k_5_end_mask_0, squeeze_mask = k_5_squeeze_mask_0, x = qkv_3_cast_fp16)[name = tensor("k_5_cast_fp16")]; + tensor v_3_begin_0 = const()[name = tensor("v_3_begin_0"), val = tensor([0, 0, 2, 0, 0])]; + tensor v_3_end_0 = const()[name = tensor("v_3_end_0"), val = tensor([1, 1, 3, 16, 64])]; + tensor v_3_end_mask_0 = const()[name = tensor("v_3_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor v_3_squeeze_mask_0 = const()[name = tensor("v_3_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor v_3_cast_fp16 = slice_by_index(begin = v_3_begin_0, end = v_3_end_0, end_mask = v_3_end_mask_0, squeeze_mask = v_3_squeeze_mask_0, x = qkv_3_cast_fp16)[name = tensor("v_3_cast_fp16")]; + tensor var_763 = const()[name = tensor("op_763"), val = tensor([1, 1, 1, 1])]; + tensor position1_to_fp16_dtype_0 = const()[name = tensor("position1_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor position1_to_fp16 = cast(dtype = position1_to_fp16_dtype_0, x = position1)[name = tensor("cast_548")]; + tensor ts_11_cast_fp16 = reshape(shape = var_763, x = position1_to_fp16)[name = tensor("ts_11_cast_fp16")]; + tensor var_767 = const()[name = tensor("op_767"), val = tensor([1, 1, 16, 32, 2])]; + tensor q_complex_3_cast_fp16 = reshape(shape = var_767, x = q_7_cast_fp16)[name = tensor("q_complex_3_cast_fp16")]; + tensor var_771 = const()[name = tensor("op_771"), val = tensor([1, 1, 16, 32, 2])]; + tensor k_complex_3_cast_fp16 = reshape(shape = var_771, x = k_5_cast_fp16)[name = tensor("k_complex_3_cast_fp16")]; + tensor var_775_begin_0 = const()[name = tensor("op_775_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_775_end_0 = const()[name = tensor("op_775_end_0"), val = tensor([1, 1, 16, 32, 1])]; + tensor var_775_end_mask_0 = const()[name = tensor("op_775_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_775_squeeze_mask_0 = const()[name = tensor("op_775_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_775_cast_fp16 = slice_by_index(begin = var_775_begin_0, end = var_775_end_0, end_mask = var_775_end_mask_0, squeeze_mask = var_775_squeeze_mask_0, x = q_complex_3_cast_fp16)[name = tensor("op_775_cast_fp16")]; + tensor var_783_begin_0 = const()[name = tensor("op_783_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_783_end_0 = const()[name = tensor("op_783_end_0"), val = tensor([1, 1, 16, 32, 2])]; + tensor var_783_end_mask_0 = const()[name = tensor("op_783_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_783_squeeze_mask_0 = const()[name = tensor("op_783_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_783_cast_fp16 = slice_by_index(begin = var_783_begin_0, end = var_783_end_0, end_mask = var_783_end_mask_0, squeeze_mask = var_783_squeeze_mask_0, x = q_complex_3_cast_fp16)[name = tensor("op_783_cast_fp16")]; + tensor var_791_begin_0 = const()[name = tensor("op_791_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_791_end_0 = const()[name = tensor("op_791_end_0"), val = tensor([1, 1, 16, 32, 1])]; + tensor var_791_end_mask_0 = const()[name = tensor("op_791_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_791_squeeze_mask_0 = const()[name = tensor("op_791_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_791_cast_fp16 = slice_by_index(begin = var_791_begin_0, end = var_791_end_0, end_mask = var_791_end_mask_0, squeeze_mask = var_791_squeeze_mask_0, x = k_complex_3_cast_fp16)[name = tensor("op_791_cast_fp16")]; + tensor var_799_begin_0 = const()[name = tensor("op_799_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_799_end_0 = const()[name = tensor("op_799_end_0"), val = tensor([1, 1, 16, 32, 2])]; + tensor var_799_end_mask_0 = const()[name = tensor("op_799_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_799_squeeze_mask_0 = const()[name = tensor("op_799_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_799_cast_fp16 = slice_by_index(begin = var_799_begin_0, end = var_799_end_0, end_mask = var_799_end_mask_0, squeeze_mask = var_799_squeeze_mask_0, x = k_complex_3_cast_fp16)[name = tensor("op_799_cast_fp16")]; + tensor freqs_3_to_fp16 = const()[name = tensor("freqs_3_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(303597696)))]; + tensor var_805_cast_fp16 = mul(x = freqs_3_to_fp16, y = ts_11_cast_fp16)[name = tensor("op_805_cast_fp16")]; + tensor rotr_3_cast_fp16 = cos(x = var_805_cast_fp16)[name = tensor("rotr_3_cast_fp16")]; + tensor roti_3_cast_fp16 = sin(x = var_805_cast_fp16)[name = tensor("roti_3_cast_fp16")]; + tensor var_809_cast_fp16 = mul(x = var_775_cast_fp16, y = rotr_3_cast_fp16)[name = tensor("op_809_cast_fp16")]; + tensor var_810_cast_fp16 = mul(x = var_783_cast_fp16, y = roti_3_cast_fp16)[name = tensor("op_810_cast_fp16")]; + tensor qor_5_cast_fp16 = sub(x = var_809_cast_fp16, y = var_810_cast_fp16)[name = tensor("qor_5_cast_fp16")]; + tensor var_813_cast_fp16 = mul(x = var_775_cast_fp16, y = roti_3_cast_fp16)[name = tensor("op_813_cast_fp16")]; + tensor var_814_cast_fp16 = mul(x = var_783_cast_fp16, y = rotr_3_cast_fp16)[name = tensor("op_814_cast_fp16")]; + tensor qoi_5_cast_fp16 = add(x = var_813_cast_fp16, y = var_814_cast_fp16)[name = tensor("qoi_5_cast_fp16")]; + tensor var_817_cast_fp16 = mul(x = var_791_cast_fp16, y = rotr_3_cast_fp16)[name = tensor("op_817_cast_fp16")]; + tensor var_818_cast_fp16 = mul(x = var_799_cast_fp16, y = roti_3_cast_fp16)[name = tensor("op_818_cast_fp16")]; + tensor kor_5_cast_fp16 = sub(x = var_817_cast_fp16, y = var_818_cast_fp16)[name = tensor("kor_5_cast_fp16")]; + tensor var_821_cast_fp16 = mul(x = var_791_cast_fp16, y = roti_3_cast_fp16)[name = tensor("op_821_cast_fp16")]; + tensor var_822_cast_fp16 = mul(x = var_799_cast_fp16, y = rotr_3_cast_fp16)[name = tensor("op_822_cast_fp16")]; + tensor koi_5_cast_fp16 = add(x = var_821_cast_fp16, y = var_822_cast_fp16)[name = tensor("koi_5_cast_fp16")]; + tensor qo_3_axis_0 = const()[name = tensor("qo_3_axis_0"), val = tensor(-1)]; + tensor qo_3_cast_fp16 = stack(axis = qo_3_axis_0, values = (qor_5_cast_fp16, qoi_5_cast_fp16))[name = tensor("qo_3_cast_fp16")]; + tensor ko_3_axis_0 = const()[name = tensor("ko_3_axis_0"), val = tensor(-1)]; + tensor ko_3_cast_fp16 = stack(axis = ko_3_axis_0, values = (kor_5_cast_fp16, koi_5_cast_fp16))[name = tensor("ko_3_cast_fp16")]; + tensor var_851 = const()[name = tensor("op_851"), val = tensor([1, 1, 16, 64])]; + tensor q_9_cast_fp16 = reshape(shape = var_851, x = qo_3_cast_fp16)[name = tensor("q_9_cast_fp16")]; + tensor var_853 = const()[name = tensor("op_853"), val = tensor([1, 1, 16, 64])]; + tensor k_7_cast_fp16 = reshape(shape = var_853, x = ko_3_cast_fp16)[name = tensor("k_7_cast_fp16")]; + tensor _inversed_875_y_0_to_fp16 = const()[name = tensor("_inversed_875_y_0_to_fp16"), val = tensor(0x1p-9)]; + tensor _inversed_875_cast_fp16 = mul(x = ts_11_cast_fp16, y = _inversed_875_y_0_to_fp16)[name = tensor("_inversed_875_cast_fp16")]; + tensor var_876_cast_fp16 = floor(x = _inversed_875_cast_fp16)[name = tensor("op_876_cast_fp16")]; + tensor var_877_to_fp16 = const()[name = tensor("op_877_to_fp16"), val = tensor(0x1p+9)]; + tensor var_878_cast_fp16 = mul(x = var_876_cast_fp16, y = var_877_to_fp16)[name = tensor("op_878_cast_fp16")]; + tensor write_indices_float_7_cast_fp16 = sub(x = ts_11_cast_fp16, y = var_878_cast_fp16)[name = tensor("write_indices_float_7_cast_fp16")]; + tensor var_885_dtype_0 = const()[name = tensor("op_885_dtype_0"), val = tensor("int32")]; + tensor write_indices_3_reps_0 = const()[name = tensor("write_indices_3_reps_0"), val = tensor([1, 1, 16, 64])]; + tensor write_indices_float_7_cast_fp16_to_int32 = cast(dtype = var_885_dtype_0, x = write_indices_float_7_cast_fp16)[name = tensor("cast_547")]; + tensor write_indices_3 = tile(reps = write_indices_3_reps_0, x = write_indices_float_7_cast_fp16_to_int32)[name = tensor("write_indices_3")]; + tensor var_893_begin_0 = const()[name = tensor("op_893_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_893_end_0 = const()[name = tensor("op_893_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor var_893_end_mask_0 = const()[name = tensor("op_893_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_893_squeeze_mask_0 = const()[name = tensor("op_893_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor cache1_to_fp16_dtype_0 = const()[name = tensor("cache1_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor cache1_to_fp16 = cast(dtype = cache1_to_fp16_dtype_0, x = cache1)[name = tensor("cast_546")]; + tensor var_893_cast_fp16 = slice_by_index(begin = var_893_begin_0, end = var_893_end_0, end_mask = var_893_end_mask_0, squeeze_mask = var_893_squeeze_mask_0, x = cache1_to_fp16)[name = tensor("op_893_cast_fp16")]; + tensor var_895_axis_0 = const()[name = tensor("op_895_axis_0"), val = tensor(1)]; + tensor var_895_mode_0 = const()[name = tensor("op_895_mode_0"), val = tensor("update")]; + tensor var_895_validate_indices_0 = const()[name = tensor("op_895_validate_indices_0"), val = tensor(false)]; + tensor var_895_cast_fp16 = scatter_along_axis(axis = var_895_axis_0, data = var_893_cast_fp16, indices = write_indices_3, mode = var_895_mode_0, updates = k_7_cast_fp16, validate_indices = var_895_validate_indices_0)[name = tensor("op_895_cast_fp16")]; + tensor concat_9 = const()[name = tensor("concat_9"), val = tensor([0, 0, 0, 0, 0])]; + tensor concat_10 = const()[name = tensor("concat_10"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_3_internal_tensor_assign_1_stride_0 = const()[name = tensor("new_cache_3_internal_tensor_assign_1_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_3_internal_tensor_assign_1_begin_mask_0 = const()[name = tensor("new_cache_3_internal_tensor_assign_1_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_3_internal_tensor_assign_1_end_mask_0 = const()[name = tensor("new_cache_3_internal_tensor_assign_1_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_3_internal_tensor_assign_1_squeeze_mask_0 = const()[name = tensor("new_cache_3_internal_tensor_assign_1_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_50 = const()[name = tensor("shape_50"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_2 = const()[name = tensor("reduce_prod_2"), val = tensor(1048576)]; + tensor range_1d_2_start_0 = const()[name = tensor("range_1d_2_start_0"), val = tensor(0)]; + tensor range_1d_2_step_0 = const()[name = tensor("range_1d_2_step_0"), val = tensor(1)]; + tensor range_1d_2 = range_1d(end = reduce_prod_2, start = range_1d_2_start_0, step = range_1d_2_step_0)[name = tensor("range_1d_2")]; + tensor reshape_10 = reshape(shape = shape_50, x = range_1d_2)[name = tensor("reshape_10")]; + tensor slice_by_index_2 = slice_by_index(begin = concat_9, begin_mask = new_cache_3_internal_tensor_assign_1_begin_mask_0, end = concat_10, end_mask = new_cache_3_internal_tensor_assign_1_end_mask_0, squeeze_mask = new_cache_3_internal_tensor_assign_1_squeeze_mask_0, stride = new_cache_3_internal_tensor_assign_1_stride_0, x = reshape_10)[name = tensor("slice_by_index_2")]; + tensor reshape_11_shape_0 = const()[name = tensor("reshape_11_shape_0"), val = tensor([-1])]; + tensor reshape_11 = reshape(shape = reshape_11_shape_0, x = slice_by_index_2)[name = tensor("reshape_11")]; + tensor reshape_12_shape_0 = const()[name = tensor("reshape_12_shape_0"), val = tensor([-1])]; + tensor reshape_12_cast_fp16 = reshape(shape = reshape_12_shape_0, x = var_895_cast_fp16)[name = tensor("reshape_12_cast_fp16")]; + tensor reshape_13_shape_0 = const()[name = tensor("reshape_13_shape_0"), val = tensor([-1])]; + tensor reshape_13_cast_fp16 = reshape(shape = reshape_13_shape_0, x = cache1_to_fp16)[name = tensor("reshape_13_cast_fp16")]; + tensor scatter_2_mode_0 = const()[name = tensor("scatter_2_mode_0"), val = tensor("update")]; + tensor scatter_2_axis_0 = const()[name = tensor("scatter_2_axis_0"), val = tensor(0)]; + tensor scatter_2_validate_indices_0 = const()[name = tensor("scatter_2_validate_indices_0"), val = tensor(false)]; + tensor scatter_2_cast_fp16 = scatter(axis = scatter_2_axis_0, data = reshape_13_cast_fp16, indices = reshape_11, mode = scatter_2_mode_0, updates = reshape_12_cast_fp16, validate_indices = scatter_2_validate_indices_0)[name = tensor("scatter_2_cast_fp16")]; + tensor reshape_14_cast_fp16 = reshape(shape = shape_50, x = scatter_2_cast_fp16)[name = tensor("reshape_14_cast_fp16")]; + tensor var_903_begin_0 = const()[name = tensor("op_903_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor var_903_end_0 = const()[name = tensor("op_903_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor var_903_end_mask_0 = const()[name = tensor("op_903_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_903_squeeze_mask_0 = const()[name = tensor("op_903_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor var_903_cast_fp16 = slice_by_index(begin = var_903_begin_0, end = var_903_end_0, end_mask = var_903_end_mask_0, squeeze_mask = var_903_squeeze_mask_0, x = reshape_14_cast_fp16)[name = tensor("op_903_cast_fp16")]; + tensor var_905_axis_0 = const()[name = tensor("op_905_axis_0"), val = tensor(1)]; + tensor var_905_mode_0 = const()[name = tensor("op_905_mode_0"), val = tensor("update")]; + tensor var_905_validate_indices_0 = const()[name = tensor("op_905_validate_indices_0"), val = tensor(false)]; + tensor var_905_cast_fp16 = scatter_along_axis(axis = var_905_axis_0, data = var_903_cast_fp16, indices = write_indices_3, mode = var_905_mode_0, updates = v_3_cast_fp16, validate_indices = var_905_validate_indices_0)[name = tensor("op_905_cast_fp16")]; + tensor concat_11 = const()[name = tensor("concat_11"), val = tensor([1, 0, 0, 0, 0])]; + tensor concat_12 = const()[name = tensor("concat_12"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_3_internal_tensor_assign_2_stride_0 = const()[name = tensor("new_cache_3_internal_tensor_assign_2_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_3_internal_tensor_assign_2_begin_mask_0 = const()[name = tensor("new_cache_3_internal_tensor_assign_2_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_3_internal_tensor_assign_2_end_mask_0 = const()[name = tensor("new_cache_3_internal_tensor_assign_2_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_3_internal_tensor_assign_2_squeeze_mask_0 = const()[name = tensor("new_cache_3_internal_tensor_assign_2_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_51 = const()[name = tensor("shape_51"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_3 = const()[name = tensor("reduce_prod_3"), val = tensor(1048576)]; + tensor range_1d_3_start_0 = const()[name = tensor("range_1d_3_start_0"), val = tensor(0)]; + tensor range_1d_3_step_0 = const()[name = tensor("range_1d_3_step_0"), val = tensor(1)]; + tensor range_1d_3 = range_1d(end = reduce_prod_3, start = range_1d_3_start_0, step = range_1d_3_step_0)[name = tensor("range_1d_3")]; + tensor reshape_15 = reshape(shape = shape_51, x = range_1d_3)[name = tensor("reshape_15")]; + tensor slice_by_index_3 = slice_by_index(begin = concat_11, begin_mask = new_cache_3_internal_tensor_assign_2_begin_mask_0, end = concat_12, end_mask = new_cache_3_internal_tensor_assign_2_end_mask_0, squeeze_mask = new_cache_3_internal_tensor_assign_2_squeeze_mask_0, stride = new_cache_3_internal_tensor_assign_2_stride_0, x = reshape_15)[name = tensor("slice_by_index_3")]; + tensor reshape_16_shape_0 = const()[name = tensor("reshape_16_shape_0"), val = tensor([-1])]; + tensor reshape_16 = reshape(shape = reshape_16_shape_0, x = slice_by_index_3)[name = tensor("reshape_16")]; + tensor reshape_17_shape_0 = const()[name = tensor("reshape_17_shape_0"), val = tensor([-1])]; + tensor reshape_17_cast_fp16 = reshape(shape = reshape_17_shape_0, x = var_905_cast_fp16)[name = tensor("reshape_17_cast_fp16")]; + tensor reshape_18_shape_0 = const()[name = tensor("reshape_18_shape_0"), val = tensor([-1])]; + tensor reshape_18_cast_fp16 = reshape(shape = reshape_18_shape_0, x = reshape_14_cast_fp16)[name = tensor("reshape_18_cast_fp16")]; + tensor scatter_3_mode_0 = const()[name = tensor("scatter_3_mode_0"), val = tensor("update")]; + tensor scatter_3_axis_0 = const()[name = tensor("scatter_3_axis_0"), val = tensor(0)]; + tensor scatter_3_validate_indices_0 = const()[name = tensor("scatter_3_validate_indices_0"), val = tensor(false)]; + tensor scatter_3_cast_fp16 = scatter(axis = scatter_3_axis_0, data = reshape_18_cast_fp16, indices = reshape_16, mode = scatter_3_mode_0, updates = reshape_17_cast_fp16, validate_indices = scatter_3_validate_indices_0)[name = tensor("scatter_3_cast_fp16")]; + tensor reshape_19_cast_fp16 = reshape(shape = shape_51, x = scatter_3_cast_fp16)[name = tensor("reshape_19_cast_fp16")]; + tensor reshape_19_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("reshape_19_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor keys_7_begin_0 = const()[name = tensor("keys_7_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor keys_7_end_0 = const()[name = tensor("keys_7_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor keys_7_end_mask_0 = const()[name = tensor("keys_7_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor keys_7_squeeze_mask_0 = const()[name = tensor("keys_7_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor keys_7_cast_fp16 = slice_by_index(begin = keys_7_begin_0, end = keys_7_end_0, end_mask = keys_7_end_mask_0, squeeze_mask = keys_7_squeeze_mask_0, x = reshape_19_cast_fp16)[name = tensor("keys_7_cast_fp16")]; + tensor values_7_begin_0 = const()[name = tensor("values_7_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor values_7_end_0 = const()[name = tensor("values_7_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor values_7_end_mask_0 = const()[name = tensor("values_7_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor values_7_squeeze_mask_0 = const()[name = tensor("values_7_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor values_7_cast_fp16 = slice_by_index(begin = values_7_begin_0, end = values_7_end_0, end_mask = values_7_end_mask_0, squeeze_mask = values_7_squeeze_mask_0, x = reshape_19_cast_fp16)[name = tensor("values_7_cast_fp16")]; + tensor var_917_cast_fp16 = not_equal(x = keys_7_cast_fp16, y = keys_7_cast_fp16)[name = tensor("op_917_cast_fp16")]; + tensor keys_9_cast_fp16 = select(a = var_504_to_fp16, b = keys_7_cast_fp16, cond = var_917_cast_fp16)[name = tensor("keys_9_cast_fp16")]; + tensor var_925_cast_fp16 = not_equal(x = values_7_cast_fp16, y = values_7_cast_fp16)[name = tensor("op_925_cast_fp16")]; + tensor values_9_cast_fp16 = select(a = var_504_to_fp16, b = values_7_cast_fp16, cond = var_925_cast_fp16)[name = tensor("values_9_cast_fp16")]; + tensor var_949 = const()[name = tensor("op_949"), val = tensor([0, 2, 1, 3])]; + tensor var_962 = const()[name = tensor("op_962"), val = tensor([1, 1, 1])]; + tensor var_963_cast_fp16 = reshape(shape = var_962, x = position1_to_fp16)[name = tensor("op_963_cast_fp16")]; + tensor var_980_to_fp16 = const()[name = tensor("op_980_to_fp16"), val = tensor(0x1p+0)]; + tensor valid_len_3_cast_fp16 = add(x = var_963_cast_fp16, y = var_980_to_fp16)[name = tensor("valid_len_3_cast_fp16")]; + tensor valid_mask_3_cast_fp16 = less(x = k_positions_1_promoted_to_fp16, y = valid_len_3_cast_fp16)[name = tensor("valid_mask_3_cast_fp16")]; + tensor causal_mask_3_cast_fp16 = less_equal(x = k_positions_1_promoted_to_fp16, y = var_963_cast_fp16)[name = tensor("causal_mask_3_cast_fp16")]; + tensor attn_mask_5 = logical_and(x = valid_mask_3_cast_fp16, y = causal_mask_3_cast_fp16)[name = tensor("attn_mask_5")]; + tensor attn_mask_7_axes_0 = const()[name = tensor("attn_mask_7_axes_0"), val = tensor([1])]; + tensor attn_mask_7 = expand_dims(axes = attn_mask_7_axes_0, x = attn_mask_5)[name = tensor("attn_mask_7")]; + tensor var_998_transpose_x_0 = const()[name = tensor("op_998_transpose_x_0"), val = tensor(false)]; + tensor var_998_transpose_y_0 = const()[name = tensor("op_998_transpose_y_0"), val = tensor(false)]; + tensor transpose_74_perm_0 = const()[name = tensor("transpose_74_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_75_perm_0 = const()[name = tensor("transpose_75_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_75 = transpose(perm = transpose_75_perm_0, x = keys_9_cast_fp16)[name = tensor("transpose_209")]; + tensor transpose_74 = transpose(perm = transpose_74_perm_0, x = q_9_cast_fp16)[name = tensor("transpose_210")]; + tensor var_998_cast_fp16 = matmul(transpose_x = var_998_transpose_x_0, transpose_y = var_998_transpose_y_0, x = transpose_74, y = transpose_75)[name = tensor("op_998_cast_fp16")]; + tensor var_992_to_fp16 = const()[name = tensor("op_992_to_fp16"), val = tensor([0x1p-3])]; + tensor attn_weights_7_cast_fp16 = mul(x = var_998_cast_fp16, y = var_992_to_fp16)[name = tensor("attn_weights_7_cast_fp16")]; + tensor var_1000 = logical_not(x = attn_mask_7)[name = tensor("op_1000")]; + tensor var_1001_to_fp16 = const()[name = tensor("op_1001_to_fp16"), val = tensor(-inf)]; + tensor attn_weights_9_cast_fp16 = select(a = var_1001_to_fp16, b = attn_weights_7_cast_fp16, cond = var_1000)[name = tensor("attn_weights_9_cast_fp16")]; + tensor var_1003 = const()[name = tensor("op_1003"), val = tensor(-1)]; + tensor attn_weights_11_cast_fp16 = softmax(axis = var_1003, x = attn_weights_9_cast_fp16)[name = tensor("attn_weights_11_cast_fp16")]; + tensor attn_output_3_transpose_x_0 = const()[name = tensor("attn_output_3_transpose_x_0"), val = tensor(false)]; + tensor attn_output_3_transpose_y_0 = const()[name = tensor("attn_output_3_transpose_y_0"), val = tensor(false)]; + tensor values_11_cast_fp16 = transpose(perm = var_949, x = values_9_cast_fp16)[name = tensor("transpose_211")]; + tensor attn_output_3_cast_fp16 = matmul(transpose_x = attn_output_3_transpose_x_0, transpose_y = attn_output_3_transpose_y_0, x = attn_weights_11_cast_fp16, y = values_11_cast_fp16)[name = tensor("attn_output_3_cast_fp16")]; + tensor var_1011 = const()[name = tensor("op_1011"), val = tensor([0, 2, 1, 3])]; + tensor var_1014 = const()[name = tensor("op_1014"), val = tensor([1, 1, 1024])]; + tensor var_1012_cast_fp16 = transpose(perm = var_1011, x = attn_output_3_cast_fp16)[name = tensor("transpose_208")]; + tensor input_15_cast_fp16 = reshape(shape = var_1014, x = var_1012_cast_fp16)[name = tensor("input_15_cast_fp16")]; + tensor linear_6_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = attn1_out_proj_weight_cast_fp16, x = input_15_cast_fp16)[name = tensor("linear_6_cast_fp16")]; + tensor var_1020_to_fp16 = const()[name = tensor("op_1020_to_fp16"), val = tensor(0x1p+0)]; + tensor var_1021_cast_fp16 = add(x = position1_to_fp16, y = var_1020_to_fp16)[name = tensor("op_1021_cast_fp16")]; + tensor var_1021_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_1021_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor input_17_cast_fp16 = add(x = input_13_cast_fp16, y = linear_6_cast_fp16)[name = tensor("input_17_cast_fp16")]; + tensor input_19_axes_0 = const()[name = tensor("input_19_axes_0"), val = tensor([-1])]; + tensor norm1_2_weight_to_fp16 = const()[name = tensor("norm1_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(303597824)))]; + tensor norm1_2_bias_to_fp16 = const()[name = tensor("norm1_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(303599936)))]; + tensor var_1025_to_fp16 = const()[name = tensor("op_1025_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_19_cast_fp16 = layer_norm(axes = input_19_axes_0, beta = norm1_2_bias_to_fp16, epsilon = var_1025_to_fp16, gamma = norm1_2_weight_to_fp16, x = input_17_cast_fp16)[name = tensor("input_19_cast_fp16")]; + tensor linear_7_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = linear1_1_weight_cast_fp16, x = input_19_cast_fp16)[name = tensor("linear_7_cast_fp16")]; + tensor input_21_mode_0 = const()[name = tensor("input_21_mode_0"), val = tensor("EXACT")]; + tensor input_21_cast_fp16 = gelu(mode = input_21_mode_0, x = linear_7_cast_fp16)[name = tensor("input_21_cast_fp16")]; + tensor linear_8_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = linear1_2_weight_cast_fp16, x = input_21_cast_fp16)[name = tensor("linear_8_cast_fp16")]; + tensor input_23_cast_fp16 = add(x = input_17_cast_fp16, y = linear_8_cast_fp16)[name = tensor("input_23_cast_fp16")]; + tensor x_5_axes_0 = const()[name = tensor("x_5_axes_0"), val = tensor([-1])]; + tensor norm2_1_weight_to_fp16 = const()[name = tensor("norm2_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(303602048)))]; + tensor norm2_1_bias_to_fp16 = const()[name = tensor("norm2_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(303604160)))]; + tensor var_1042_to_fp16 = const()[name = tensor("op_1042_to_fp16"), val = tensor(0x1.5p-17)]; + tensor x_5_cast_fp16 = layer_norm(axes = x_5_axes_0, beta = norm2_1_bias_to_fp16, epsilon = var_1042_to_fp16, gamma = norm2_1_weight_to_fp16, x = input_23_cast_fp16)[name = tensor("x_5_cast_fp16")]; + tensor linear_9_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = attn2_in_proj_weight_cast_fp16, x = x_5_cast_fp16)[name = tensor("linear_9_cast_fp16")]; + tensor var_1078 = const()[name = tensor("op_1078"), val = tensor([1, 1, 3, 16, 64])]; + tensor qkv_5_cast_fp16 = reshape(shape = var_1078, x = linear_9_cast_fp16)[name = tensor("qkv_5_cast_fp16")]; + tensor q_13_begin_0 = const()[name = tensor("q_13_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor q_13_end_0 = const()[name = tensor("q_13_end_0"), val = tensor([1, 1, 1, 16, 64])]; + tensor q_13_end_mask_0 = const()[name = tensor("q_13_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor q_13_squeeze_mask_0 = const()[name = tensor("q_13_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor q_13_cast_fp16 = slice_by_index(begin = q_13_begin_0, end = q_13_end_0, end_mask = q_13_end_mask_0, squeeze_mask = q_13_squeeze_mask_0, x = qkv_5_cast_fp16)[name = tensor("q_13_cast_fp16")]; + tensor k_9_begin_0 = const()[name = tensor("k_9_begin_0"), val = tensor([0, 0, 1, 0, 0])]; + tensor k_9_end_0 = const()[name = tensor("k_9_end_0"), val = tensor([1, 1, 2, 16, 64])]; + tensor k_9_end_mask_0 = const()[name = tensor("k_9_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor k_9_squeeze_mask_0 = const()[name = tensor("k_9_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor k_9_cast_fp16 = slice_by_index(begin = k_9_begin_0, end = k_9_end_0, end_mask = k_9_end_mask_0, squeeze_mask = k_9_squeeze_mask_0, x = qkv_5_cast_fp16)[name = tensor("k_9_cast_fp16")]; + tensor v_5_begin_0 = const()[name = tensor("v_5_begin_0"), val = tensor([0, 0, 2, 0, 0])]; + tensor v_5_end_0 = const()[name = tensor("v_5_end_0"), val = tensor([1, 1, 3, 16, 64])]; + tensor v_5_end_mask_0 = const()[name = tensor("v_5_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor v_5_squeeze_mask_0 = const()[name = tensor("v_5_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor v_5_cast_fp16 = slice_by_index(begin = v_5_begin_0, end = v_5_end_0, end_mask = v_5_end_mask_0, squeeze_mask = v_5_squeeze_mask_0, x = qkv_5_cast_fp16)[name = tensor("v_5_cast_fp16")]; + tensor var_1182 = const()[name = tensor("op_1182"), val = tensor([1, 1, 1, 1])]; + tensor position2_to_fp16_dtype_0 = const()[name = tensor("position2_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor position2_to_fp16 = cast(dtype = position2_to_fp16_dtype_0, x = position2)[name = tensor("cast_543")]; + tensor ts_17_cast_fp16 = reshape(shape = var_1182, x = position2_to_fp16)[name = tensor("ts_17_cast_fp16")]; + tensor var_1186 = const()[name = tensor("op_1186"), val = tensor([1, 1, 16, 32, 2])]; + tensor q_complex_5_cast_fp16 = reshape(shape = var_1186, x = q_13_cast_fp16)[name = tensor("q_complex_5_cast_fp16")]; + tensor var_1190 = const()[name = tensor("op_1190"), val = tensor([1, 1, 16, 32, 2])]; + tensor k_complex_5_cast_fp16 = reshape(shape = var_1190, x = k_9_cast_fp16)[name = tensor("k_complex_5_cast_fp16")]; + tensor var_1194_begin_0 = const()[name = tensor("op_1194_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_1194_end_0 = const()[name = tensor("op_1194_end_0"), val = tensor([1, 1, 16, 32, 1])]; + tensor var_1194_end_mask_0 = const()[name = tensor("op_1194_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_1194_squeeze_mask_0 = const()[name = tensor("op_1194_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_1194_cast_fp16 = slice_by_index(begin = var_1194_begin_0, end = var_1194_end_0, end_mask = var_1194_end_mask_0, squeeze_mask = var_1194_squeeze_mask_0, x = q_complex_5_cast_fp16)[name = tensor("op_1194_cast_fp16")]; + tensor var_1202_begin_0 = const()[name = tensor("op_1202_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_1202_end_0 = const()[name = tensor("op_1202_end_0"), val = tensor([1, 1, 16, 32, 2])]; + tensor var_1202_end_mask_0 = const()[name = tensor("op_1202_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_1202_squeeze_mask_0 = const()[name = tensor("op_1202_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_1202_cast_fp16 = slice_by_index(begin = var_1202_begin_0, end = var_1202_end_0, end_mask = var_1202_end_mask_0, squeeze_mask = var_1202_squeeze_mask_0, x = q_complex_5_cast_fp16)[name = tensor("op_1202_cast_fp16")]; + tensor var_1210_begin_0 = const()[name = tensor("op_1210_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_1210_end_0 = const()[name = tensor("op_1210_end_0"), val = tensor([1, 1, 16, 32, 1])]; + tensor var_1210_end_mask_0 = const()[name = tensor("op_1210_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_1210_squeeze_mask_0 = const()[name = tensor("op_1210_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_1210_cast_fp16 = slice_by_index(begin = var_1210_begin_0, end = var_1210_end_0, end_mask = var_1210_end_mask_0, squeeze_mask = var_1210_squeeze_mask_0, x = k_complex_5_cast_fp16)[name = tensor("op_1210_cast_fp16")]; + tensor var_1218_begin_0 = const()[name = tensor("op_1218_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_1218_end_0 = const()[name = tensor("op_1218_end_0"), val = tensor([1, 1, 16, 32, 2])]; + tensor var_1218_end_mask_0 = const()[name = tensor("op_1218_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_1218_squeeze_mask_0 = const()[name = tensor("op_1218_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_1218_cast_fp16 = slice_by_index(begin = var_1218_begin_0, end = var_1218_end_0, end_mask = var_1218_end_mask_0, squeeze_mask = var_1218_squeeze_mask_0, x = k_complex_5_cast_fp16)[name = tensor("op_1218_cast_fp16")]; + tensor freqs_5_to_fp16 = const()[name = tensor("freqs_5_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(303606272)))]; + tensor var_1224_cast_fp16 = mul(x = freqs_5_to_fp16, y = ts_17_cast_fp16)[name = tensor("op_1224_cast_fp16")]; + tensor rotr_5_cast_fp16 = cos(x = var_1224_cast_fp16)[name = tensor("rotr_5_cast_fp16")]; + tensor roti_5_cast_fp16 = sin(x = var_1224_cast_fp16)[name = tensor("roti_5_cast_fp16")]; + tensor var_1228_cast_fp16 = mul(x = var_1194_cast_fp16, y = rotr_5_cast_fp16)[name = tensor("op_1228_cast_fp16")]; + tensor var_1229_cast_fp16 = mul(x = var_1202_cast_fp16, y = roti_5_cast_fp16)[name = tensor("op_1229_cast_fp16")]; + tensor qor_9_cast_fp16 = sub(x = var_1228_cast_fp16, y = var_1229_cast_fp16)[name = tensor("qor_9_cast_fp16")]; + tensor var_1232_cast_fp16 = mul(x = var_1194_cast_fp16, y = roti_5_cast_fp16)[name = tensor("op_1232_cast_fp16")]; + tensor var_1233_cast_fp16 = mul(x = var_1202_cast_fp16, y = rotr_5_cast_fp16)[name = tensor("op_1233_cast_fp16")]; + tensor qoi_9_cast_fp16 = add(x = var_1232_cast_fp16, y = var_1233_cast_fp16)[name = tensor("qoi_9_cast_fp16")]; + tensor var_1236_cast_fp16 = mul(x = var_1210_cast_fp16, y = rotr_5_cast_fp16)[name = tensor("op_1236_cast_fp16")]; + tensor var_1237_cast_fp16 = mul(x = var_1218_cast_fp16, y = roti_5_cast_fp16)[name = tensor("op_1237_cast_fp16")]; + tensor kor_9_cast_fp16 = sub(x = var_1236_cast_fp16, y = var_1237_cast_fp16)[name = tensor("kor_9_cast_fp16")]; + tensor var_1240_cast_fp16 = mul(x = var_1210_cast_fp16, y = roti_5_cast_fp16)[name = tensor("op_1240_cast_fp16")]; + tensor var_1241_cast_fp16 = mul(x = var_1218_cast_fp16, y = rotr_5_cast_fp16)[name = tensor("op_1241_cast_fp16")]; + tensor koi_9_cast_fp16 = add(x = var_1240_cast_fp16, y = var_1241_cast_fp16)[name = tensor("koi_9_cast_fp16")]; + tensor qo_5_axis_0 = const()[name = tensor("qo_5_axis_0"), val = tensor(-1)]; + tensor qo_5_cast_fp16 = stack(axis = qo_5_axis_0, values = (qor_9_cast_fp16, qoi_9_cast_fp16))[name = tensor("qo_5_cast_fp16")]; + tensor ko_5_axis_0 = const()[name = tensor("ko_5_axis_0"), val = tensor(-1)]; + tensor ko_5_cast_fp16 = stack(axis = ko_5_axis_0, values = (kor_9_cast_fp16, koi_9_cast_fp16))[name = tensor("ko_5_cast_fp16")]; + tensor var_1270 = const()[name = tensor("op_1270"), val = tensor([1, 1, 16, 64])]; + tensor q_15_cast_fp16 = reshape(shape = var_1270, x = qo_5_cast_fp16)[name = tensor("q_15_cast_fp16")]; + tensor var_1272 = const()[name = tensor("op_1272"), val = tensor([1, 1, 16, 64])]; + tensor k_11_cast_fp16 = reshape(shape = var_1272, x = ko_5_cast_fp16)[name = tensor("k_11_cast_fp16")]; + tensor _inversed_1294_y_0_to_fp16 = const()[name = tensor("_inversed_1294_y_0_to_fp16"), val = tensor(0x1p-9)]; + tensor _inversed_1294_cast_fp16 = mul(x = ts_17_cast_fp16, y = _inversed_1294_y_0_to_fp16)[name = tensor("_inversed_1294_cast_fp16")]; + tensor var_1295_cast_fp16 = floor(x = _inversed_1294_cast_fp16)[name = tensor("op_1295_cast_fp16")]; + tensor var_1296_to_fp16 = const()[name = tensor("op_1296_to_fp16"), val = tensor(0x1p+9)]; + tensor var_1297_cast_fp16 = mul(x = var_1295_cast_fp16, y = var_1296_to_fp16)[name = tensor("op_1297_cast_fp16")]; + tensor write_indices_float_11_cast_fp16 = sub(x = ts_17_cast_fp16, y = var_1297_cast_fp16)[name = tensor("write_indices_float_11_cast_fp16")]; + tensor var_1304_dtype_0 = const()[name = tensor("op_1304_dtype_0"), val = tensor("int32")]; + tensor write_indices_5_reps_0 = const()[name = tensor("write_indices_5_reps_0"), val = tensor([1, 1, 16, 64])]; + tensor write_indices_float_11_cast_fp16_to_int32 = cast(dtype = var_1304_dtype_0, x = write_indices_float_11_cast_fp16)[name = tensor("cast_542")]; + tensor write_indices_5 = tile(reps = write_indices_5_reps_0, x = write_indices_float_11_cast_fp16_to_int32)[name = tensor("write_indices_5")]; + tensor var_1312_begin_0 = const()[name = tensor("op_1312_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_1312_end_0 = const()[name = tensor("op_1312_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor var_1312_end_mask_0 = const()[name = tensor("op_1312_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_1312_squeeze_mask_0 = const()[name = tensor("op_1312_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor cache2_to_fp16_dtype_0 = const()[name = tensor("cache2_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor cache2_to_fp16 = cast(dtype = cache2_to_fp16_dtype_0, x = cache2)[name = tensor("cast_541")]; + tensor var_1312_cast_fp16 = slice_by_index(begin = var_1312_begin_0, end = var_1312_end_0, end_mask = var_1312_end_mask_0, squeeze_mask = var_1312_squeeze_mask_0, x = cache2_to_fp16)[name = tensor("op_1312_cast_fp16")]; + tensor var_1314_axis_0 = const()[name = tensor("op_1314_axis_0"), val = tensor(1)]; + tensor var_1314_mode_0 = const()[name = tensor("op_1314_mode_0"), val = tensor("update")]; + tensor var_1314_validate_indices_0 = const()[name = tensor("op_1314_validate_indices_0"), val = tensor(false)]; + tensor var_1314_cast_fp16 = scatter_along_axis(axis = var_1314_axis_0, data = var_1312_cast_fp16, indices = write_indices_5, mode = var_1314_mode_0, updates = k_11_cast_fp16, validate_indices = var_1314_validate_indices_0)[name = tensor("op_1314_cast_fp16")]; + tensor concat_16 = const()[name = tensor("concat_16"), val = tensor([0, 0, 0, 0, 0])]; + tensor concat_17 = const()[name = tensor("concat_17"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_5_internal_tensor_assign_1_stride_0 = const()[name = tensor("new_cache_5_internal_tensor_assign_1_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_5_internal_tensor_assign_1_begin_mask_0 = const()[name = tensor("new_cache_5_internal_tensor_assign_1_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_5_internal_tensor_assign_1_end_mask_0 = const()[name = tensor("new_cache_5_internal_tensor_assign_1_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_5_internal_tensor_assign_1_squeeze_mask_0 = const()[name = tensor("new_cache_5_internal_tensor_assign_1_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_52 = const()[name = tensor("shape_52"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_4 = const()[name = tensor("reduce_prod_4"), val = tensor(1048576)]; + tensor range_1d_4_start_0 = const()[name = tensor("range_1d_4_start_0"), val = tensor(0)]; + tensor range_1d_4_step_0 = const()[name = tensor("range_1d_4_step_0"), val = tensor(1)]; + tensor range_1d_4 = range_1d(end = reduce_prod_4, start = range_1d_4_start_0, step = range_1d_4_step_0)[name = tensor("range_1d_4")]; + tensor reshape_20 = reshape(shape = shape_52, x = range_1d_4)[name = tensor("reshape_20")]; + tensor slice_by_index_4 = slice_by_index(begin = concat_16, begin_mask = new_cache_5_internal_tensor_assign_1_begin_mask_0, end = concat_17, end_mask = new_cache_5_internal_tensor_assign_1_end_mask_0, squeeze_mask = new_cache_5_internal_tensor_assign_1_squeeze_mask_0, stride = new_cache_5_internal_tensor_assign_1_stride_0, x = reshape_20)[name = tensor("slice_by_index_4")]; + tensor reshape_21_shape_0 = const()[name = tensor("reshape_21_shape_0"), val = tensor([-1])]; + tensor reshape_21 = reshape(shape = reshape_21_shape_0, x = slice_by_index_4)[name = tensor("reshape_21")]; + tensor reshape_22_shape_0 = const()[name = tensor("reshape_22_shape_0"), val = tensor([-1])]; + tensor reshape_22_cast_fp16 = reshape(shape = reshape_22_shape_0, x = var_1314_cast_fp16)[name = tensor("reshape_22_cast_fp16")]; + tensor reshape_23_shape_0 = const()[name = tensor("reshape_23_shape_0"), val = tensor([-1])]; + tensor reshape_23_cast_fp16 = reshape(shape = reshape_23_shape_0, x = cache2_to_fp16)[name = tensor("reshape_23_cast_fp16")]; + tensor scatter_4_mode_0 = const()[name = tensor("scatter_4_mode_0"), val = tensor("update")]; + tensor scatter_4_axis_0 = const()[name = tensor("scatter_4_axis_0"), val = tensor(0)]; + tensor scatter_4_validate_indices_0 = const()[name = tensor("scatter_4_validate_indices_0"), val = tensor(false)]; + tensor scatter_4_cast_fp16 = scatter(axis = scatter_4_axis_0, data = reshape_23_cast_fp16, indices = reshape_21, mode = scatter_4_mode_0, updates = reshape_22_cast_fp16, validate_indices = scatter_4_validate_indices_0)[name = tensor("scatter_4_cast_fp16")]; + tensor reshape_24_cast_fp16 = reshape(shape = shape_52, x = scatter_4_cast_fp16)[name = tensor("reshape_24_cast_fp16")]; + tensor var_1322_begin_0 = const()[name = tensor("op_1322_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor var_1322_end_0 = const()[name = tensor("op_1322_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor var_1322_end_mask_0 = const()[name = tensor("op_1322_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_1322_squeeze_mask_0 = const()[name = tensor("op_1322_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor var_1322_cast_fp16 = slice_by_index(begin = var_1322_begin_0, end = var_1322_end_0, end_mask = var_1322_end_mask_0, squeeze_mask = var_1322_squeeze_mask_0, x = reshape_24_cast_fp16)[name = tensor("op_1322_cast_fp16")]; + tensor var_1324_axis_0 = const()[name = tensor("op_1324_axis_0"), val = tensor(1)]; + tensor var_1324_mode_0 = const()[name = tensor("op_1324_mode_0"), val = tensor("update")]; + tensor var_1324_validate_indices_0 = const()[name = tensor("op_1324_validate_indices_0"), val = tensor(false)]; + tensor var_1324_cast_fp16 = scatter_along_axis(axis = var_1324_axis_0, data = var_1322_cast_fp16, indices = write_indices_5, mode = var_1324_mode_0, updates = v_5_cast_fp16, validate_indices = var_1324_validate_indices_0)[name = tensor("op_1324_cast_fp16")]; + tensor concat_18 = const()[name = tensor("concat_18"), val = tensor([1, 0, 0, 0, 0])]; + tensor concat_19 = const()[name = tensor("concat_19"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_5_internal_tensor_assign_2_stride_0 = const()[name = tensor("new_cache_5_internal_tensor_assign_2_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_5_internal_tensor_assign_2_begin_mask_0 = const()[name = tensor("new_cache_5_internal_tensor_assign_2_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_5_internal_tensor_assign_2_end_mask_0 = const()[name = tensor("new_cache_5_internal_tensor_assign_2_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_5_internal_tensor_assign_2_squeeze_mask_0 = const()[name = tensor("new_cache_5_internal_tensor_assign_2_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_53 = const()[name = tensor("shape_53"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_5 = const()[name = tensor("reduce_prod_5"), val = tensor(1048576)]; + tensor range_1d_5_start_0 = const()[name = tensor("range_1d_5_start_0"), val = tensor(0)]; + tensor range_1d_5_step_0 = const()[name = tensor("range_1d_5_step_0"), val = tensor(1)]; + tensor range_1d_5 = range_1d(end = reduce_prod_5, start = range_1d_5_start_0, step = range_1d_5_step_0)[name = tensor("range_1d_5")]; + tensor reshape_25 = reshape(shape = shape_53, x = range_1d_5)[name = tensor("reshape_25")]; + tensor slice_by_index_5 = slice_by_index(begin = concat_18, begin_mask = new_cache_5_internal_tensor_assign_2_begin_mask_0, end = concat_19, end_mask = new_cache_5_internal_tensor_assign_2_end_mask_0, squeeze_mask = new_cache_5_internal_tensor_assign_2_squeeze_mask_0, stride = new_cache_5_internal_tensor_assign_2_stride_0, x = reshape_25)[name = tensor("slice_by_index_5")]; + tensor reshape_26_shape_0 = const()[name = tensor("reshape_26_shape_0"), val = tensor([-1])]; + tensor reshape_26 = reshape(shape = reshape_26_shape_0, x = slice_by_index_5)[name = tensor("reshape_26")]; + tensor reshape_27_shape_0 = const()[name = tensor("reshape_27_shape_0"), val = tensor([-1])]; + tensor reshape_27_cast_fp16 = reshape(shape = reshape_27_shape_0, x = var_1324_cast_fp16)[name = tensor("reshape_27_cast_fp16")]; + tensor reshape_28_shape_0 = const()[name = tensor("reshape_28_shape_0"), val = tensor([-1])]; + tensor reshape_28_cast_fp16 = reshape(shape = reshape_28_shape_0, x = reshape_24_cast_fp16)[name = tensor("reshape_28_cast_fp16")]; + tensor scatter_5_mode_0 = const()[name = tensor("scatter_5_mode_0"), val = tensor("update")]; + tensor scatter_5_axis_0 = const()[name = tensor("scatter_5_axis_0"), val = tensor(0)]; + tensor scatter_5_validate_indices_0 = const()[name = tensor("scatter_5_validate_indices_0"), val = tensor(false)]; + tensor scatter_5_cast_fp16 = scatter(axis = scatter_5_axis_0, data = reshape_28_cast_fp16, indices = reshape_26, mode = scatter_5_mode_0, updates = reshape_27_cast_fp16, validate_indices = scatter_5_validate_indices_0)[name = tensor("scatter_5_cast_fp16")]; + tensor reshape_29_cast_fp16 = reshape(shape = shape_53, x = scatter_5_cast_fp16)[name = tensor("reshape_29_cast_fp16")]; + tensor reshape_29_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("reshape_29_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor keys_13_begin_0 = const()[name = tensor("keys_13_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor keys_13_end_0 = const()[name = tensor("keys_13_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor keys_13_end_mask_0 = const()[name = tensor("keys_13_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor keys_13_squeeze_mask_0 = const()[name = tensor("keys_13_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor keys_13_cast_fp16 = slice_by_index(begin = keys_13_begin_0, end = keys_13_end_0, end_mask = keys_13_end_mask_0, squeeze_mask = keys_13_squeeze_mask_0, x = reshape_29_cast_fp16)[name = tensor("keys_13_cast_fp16")]; + tensor values_13_begin_0 = const()[name = tensor("values_13_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor values_13_end_0 = const()[name = tensor("values_13_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor values_13_end_mask_0 = const()[name = tensor("values_13_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor values_13_squeeze_mask_0 = const()[name = tensor("values_13_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor values_13_cast_fp16 = slice_by_index(begin = values_13_begin_0, end = values_13_end_0, end_mask = values_13_end_mask_0, squeeze_mask = values_13_squeeze_mask_0, x = reshape_29_cast_fp16)[name = tensor("values_13_cast_fp16")]; + tensor var_1336_cast_fp16 = not_equal(x = keys_13_cast_fp16, y = keys_13_cast_fp16)[name = tensor("op_1336_cast_fp16")]; + tensor keys_15_cast_fp16 = select(a = var_504_to_fp16, b = keys_13_cast_fp16, cond = var_1336_cast_fp16)[name = tensor("keys_15_cast_fp16")]; + tensor var_1344_cast_fp16 = not_equal(x = values_13_cast_fp16, y = values_13_cast_fp16)[name = tensor("op_1344_cast_fp16")]; + tensor values_15_cast_fp16 = select(a = var_504_to_fp16, b = values_13_cast_fp16, cond = var_1344_cast_fp16)[name = tensor("values_15_cast_fp16")]; + tensor var_1368 = const()[name = tensor("op_1368"), val = tensor([0, 2, 1, 3])]; + tensor var_1381 = const()[name = tensor("op_1381"), val = tensor([1, 1, 1])]; + tensor var_1382_cast_fp16 = reshape(shape = var_1381, x = position2_to_fp16)[name = tensor("op_1382_cast_fp16")]; + tensor var_1399_to_fp16 = const()[name = tensor("op_1399_to_fp16"), val = tensor(0x1p+0)]; + tensor valid_len_5_cast_fp16 = add(x = var_1382_cast_fp16, y = var_1399_to_fp16)[name = tensor("valid_len_5_cast_fp16")]; + tensor valid_mask_5_cast_fp16 = less(x = k_positions_1_promoted_to_fp16, y = valid_len_5_cast_fp16)[name = tensor("valid_mask_5_cast_fp16")]; + tensor causal_mask_5_cast_fp16 = less_equal(x = k_positions_1_promoted_to_fp16, y = var_1382_cast_fp16)[name = tensor("causal_mask_5_cast_fp16")]; + tensor attn_mask_9 = logical_and(x = valid_mask_5_cast_fp16, y = causal_mask_5_cast_fp16)[name = tensor("attn_mask_9")]; + tensor attn_mask_11_axes_0 = const()[name = tensor("attn_mask_11_axes_0"), val = tensor([1])]; + tensor attn_mask_11 = expand_dims(axes = attn_mask_11_axes_0, x = attn_mask_9)[name = tensor("attn_mask_11")]; + tensor var_1417_transpose_x_0 = const()[name = tensor("op_1417_transpose_x_0"), val = tensor(false)]; + tensor var_1417_transpose_y_0 = const()[name = tensor("op_1417_transpose_y_0"), val = tensor(false)]; + tensor transpose_76_perm_0 = const()[name = tensor("transpose_76_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_77_perm_0 = const()[name = tensor("transpose_77_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_77 = transpose(perm = transpose_77_perm_0, x = keys_15_cast_fp16)[name = tensor("transpose_205")]; + tensor transpose_76 = transpose(perm = transpose_76_perm_0, x = q_15_cast_fp16)[name = tensor("transpose_206")]; + tensor var_1417_cast_fp16 = matmul(transpose_x = var_1417_transpose_x_0, transpose_y = var_1417_transpose_y_0, x = transpose_76, y = transpose_77)[name = tensor("op_1417_cast_fp16")]; + tensor var_1411_to_fp16 = const()[name = tensor("op_1411_to_fp16"), val = tensor([0x1p-3])]; + tensor attn_weights_13_cast_fp16 = mul(x = var_1417_cast_fp16, y = var_1411_to_fp16)[name = tensor("attn_weights_13_cast_fp16")]; + tensor var_1419 = logical_not(x = attn_mask_11)[name = tensor("op_1419")]; + tensor var_1420_to_fp16 = const()[name = tensor("op_1420_to_fp16"), val = tensor(-inf)]; + tensor attn_weights_15_cast_fp16 = select(a = var_1420_to_fp16, b = attn_weights_13_cast_fp16, cond = var_1419)[name = tensor("attn_weights_15_cast_fp16")]; + tensor var_1422 = const()[name = tensor("op_1422"), val = tensor(-1)]; + tensor attn_weights_17_cast_fp16 = softmax(axis = var_1422, x = attn_weights_15_cast_fp16)[name = tensor("attn_weights_17_cast_fp16")]; + tensor attn_output_5_transpose_x_0 = const()[name = tensor("attn_output_5_transpose_x_0"), val = tensor(false)]; + tensor attn_output_5_transpose_y_0 = const()[name = tensor("attn_output_5_transpose_y_0"), val = tensor(false)]; + tensor values_17_cast_fp16 = transpose(perm = var_1368, x = values_15_cast_fp16)[name = tensor("transpose_207")]; + tensor attn_output_5_cast_fp16 = matmul(transpose_x = attn_output_5_transpose_x_0, transpose_y = attn_output_5_transpose_y_0, x = attn_weights_17_cast_fp16, y = values_17_cast_fp16)[name = tensor("attn_output_5_cast_fp16")]; + tensor var_1430 = const()[name = tensor("op_1430"), val = tensor([0, 2, 1, 3])]; + tensor var_1433 = const()[name = tensor("op_1433"), val = tensor([1, 1, 1024])]; + tensor var_1431_cast_fp16 = transpose(perm = var_1430, x = attn_output_5_cast_fp16)[name = tensor("transpose_204")]; + tensor input_25_cast_fp16 = reshape(shape = var_1433, x = var_1431_cast_fp16)[name = tensor("input_25_cast_fp16")]; + tensor linear_10_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = attn2_out_proj_weight_cast_fp16, x = input_25_cast_fp16)[name = tensor("linear_10_cast_fp16")]; + tensor var_1439_to_fp16 = const()[name = tensor("op_1439_to_fp16"), val = tensor(0x1p+0)]; + tensor var_1440_cast_fp16 = add(x = position2_to_fp16, y = var_1439_to_fp16)[name = tensor("op_1440_cast_fp16")]; + tensor var_1440_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_1440_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor input_27_cast_fp16 = add(x = input_23_cast_fp16, y = linear_10_cast_fp16)[name = tensor("input_27_cast_fp16")]; + tensor input_29_axes_0 = const()[name = tensor("input_29_axes_0"), val = tensor([-1])]; + tensor norm2_2_weight_to_fp16 = const()[name = tensor("norm2_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(303606400)))]; + tensor norm2_2_bias_to_fp16 = const()[name = tensor("norm2_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(303608512)))]; + tensor var_1444_to_fp16 = const()[name = tensor("op_1444_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_29_cast_fp16 = layer_norm(axes = input_29_axes_0, beta = norm2_2_bias_to_fp16, epsilon = var_1444_to_fp16, gamma = norm2_2_weight_to_fp16, x = input_27_cast_fp16)[name = tensor("input_29_cast_fp16")]; + tensor linear_11_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = linear2_1_weight_cast_fp16, x = input_29_cast_fp16)[name = tensor("linear_11_cast_fp16")]; + tensor input_31_mode_0 = const()[name = tensor("input_31_mode_0"), val = tensor("EXACT")]; + tensor input_31_cast_fp16 = gelu(mode = input_31_mode_0, x = linear_11_cast_fp16)[name = tensor("input_31_cast_fp16")]; + tensor linear_12_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = linear2_2_weight_cast_fp16, x = input_31_cast_fp16)[name = tensor("linear_12_cast_fp16")]; + tensor input_33_cast_fp16 = add(x = input_27_cast_fp16, y = linear_12_cast_fp16)[name = tensor("input_33_cast_fp16")]; + tensor x_7_axes_0 = const()[name = tensor("x_7_axes_0"), val = tensor([-1])]; + tensor norm3_1_weight_to_fp16 = const()[name = tensor("norm3_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(303610624)))]; + tensor norm3_1_bias_to_fp16 = const()[name = tensor("norm3_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(303612736)))]; + tensor var_1461_to_fp16 = const()[name = tensor("op_1461_to_fp16"), val = tensor(0x1.5p-17)]; + tensor x_7_cast_fp16 = layer_norm(axes = x_7_axes_0, beta = norm3_1_bias_to_fp16, epsilon = var_1461_to_fp16, gamma = norm3_1_weight_to_fp16, x = input_33_cast_fp16)[name = tensor("x_7_cast_fp16")]; + tensor linear_13_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = attn3_in_proj_weight_cast_fp16, x = x_7_cast_fp16)[name = tensor("linear_13_cast_fp16")]; + tensor var_1497 = const()[name = tensor("op_1497"), val = tensor([1, 1, 3, 16, 64])]; + tensor qkv_7_cast_fp16 = reshape(shape = var_1497, x = linear_13_cast_fp16)[name = tensor("qkv_7_cast_fp16")]; + tensor q_19_begin_0 = const()[name = tensor("q_19_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor q_19_end_0 = const()[name = tensor("q_19_end_0"), val = tensor([1, 1, 1, 16, 64])]; + tensor q_19_end_mask_0 = const()[name = tensor("q_19_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor q_19_squeeze_mask_0 = const()[name = tensor("q_19_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor q_19_cast_fp16 = slice_by_index(begin = q_19_begin_0, end = q_19_end_0, end_mask = q_19_end_mask_0, squeeze_mask = q_19_squeeze_mask_0, x = qkv_7_cast_fp16)[name = tensor("q_19_cast_fp16")]; + tensor k_13_begin_0 = const()[name = tensor("k_13_begin_0"), val = tensor([0, 0, 1, 0, 0])]; + tensor k_13_end_0 = const()[name = tensor("k_13_end_0"), val = tensor([1, 1, 2, 16, 64])]; + tensor k_13_end_mask_0 = const()[name = tensor("k_13_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor k_13_squeeze_mask_0 = const()[name = tensor("k_13_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor k_13_cast_fp16 = slice_by_index(begin = k_13_begin_0, end = k_13_end_0, end_mask = k_13_end_mask_0, squeeze_mask = k_13_squeeze_mask_0, x = qkv_7_cast_fp16)[name = tensor("k_13_cast_fp16")]; + tensor v_7_begin_0 = const()[name = tensor("v_7_begin_0"), val = tensor([0, 0, 2, 0, 0])]; + tensor v_7_end_0 = const()[name = tensor("v_7_end_0"), val = tensor([1, 1, 3, 16, 64])]; + tensor v_7_end_mask_0 = const()[name = tensor("v_7_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor v_7_squeeze_mask_0 = const()[name = tensor("v_7_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor v_7_cast_fp16 = slice_by_index(begin = v_7_begin_0, end = v_7_end_0, end_mask = v_7_end_mask_0, squeeze_mask = v_7_squeeze_mask_0, x = qkv_7_cast_fp16)[name = tensor("v_7_cast_fp16")]; + tensor var_1601 = const()[name = tensor("op_1601"), val = tensor([1, 1, 1, 1])]; + tensor position3_to_fp16_dtype_0 = const()[name = tensor("position3_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor position3_to_fp16 = cast(dtype = position3_to_fp16_dtype_0, x = position3)[name = tensor("cast_538")]; + tensor ts_23_cast_fp16 = reshape(shape = var_1601, x = position3_to_fp16)[name = tensor("ts_23_cast_fp16")]; + tensor var_1605 = const()[name = tensor("op_1605"), val = tensor([1, 1, 16, 32, 2])]; + tensor q_complex_7_cast_fp16 = reshape(shape = var_1605, x = q_19_cast_fp16)[name = tensor("q_complex_7_cast_fp16")]; + tensor var_1609 = const()[name = tensor("op_1609"), val = tensor([1, 1, 16, 32, 2])]; + tensor k_complex_7_cast_fp16 = reshape(shape = var_1609, x = k_13_cast_fp16)[name = tensor("k_complex_7_cast_fp16")]; + tensor var_1613_begin_0 = const()[name = tensor("op_1613_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_1613_end_0 = const()[name = tensor("op_1613_end_0"), val = tensor([1, 1, 16, 32, 1])]; + tensor var_1613_end_mask_0 = const()[name = tensor("op_1613_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_1613_squeeze_mask_0 = const()[name = tensor("op_1613_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_1613_cast_fp16 = slice_by_index(begin = var_1613_begin_0, end = var_1613_end_0, end_mask = var_1613_end_mask_0, squeeze_mask = var_1613_squeeze_mask_0, x = q_complex_7_cast_fp16)[name = tensor("op_1613_cast_fp16")]; + tensor var_1621_begin_0 = const()[name = tensor("op_1621_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_1621_end_0 = const()[name = tensor("op_1621_end_0"), val = tensor([1, 1, 16, 32, 2])]; + tensor var_1621_end_mask_0 = const()[name = tensor("op_1621_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_1621_squeeze_mask_0 = const()[name = tensor("op_1621_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_1621_cast_fp16 = slice_by_index(begin = var_1621_begin_0, end = var_1621_end_0, end_mask = var_1621_end_mask_0, squeeze_mask = var_1621_squeeze_mask_0, x = q_complex_7_cast_fp16)[name = tensor("op_1621_cast_fp16")]; + tensor var_1629_begin_0 = const()[name = tensor("op_1629_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_1629_end_0 = const()[name = tensor("op_1629_end_0"), val = tensor([1, 1, 16, 32, 1])]; + tensor var_1629_end_mask_0 = const()[name = tensor("op_1629_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_1629_squeeze_mask_0 = const()[name = tensor("op_1629_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_1629_cast_fp16 = slice_by_index(begin = var_1629_begin_0, end = var_1629_end_0, end_mask = var_1629_end_mask_0, squeeze_mask = var_1629_squeeze_mask_0, x = k_complex_7_cast_fp16)[name = tensor("op_1629_cast_fp16")]; + tensor var_1637_begin_0 = const()[name = tensor("op_1637_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_1637_end_0 = const()[name = tensor("op_1637_end_0"), val = tensor([1, 1, 16, 32, 2])]; + tensor var_1637_end_mask_0 = const()[name = tensor("op_1637_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_1637_squeeze_mask_0 = const()[name = tensor("op_1637_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_1637_cast_fp16 = slice_by_index(begin = var_1637_begin_0, end = var_1637_end_0, end_mask = var_1637_end_mask_0, squeeze_mask = var_1637_squeeze_mask_0, x = k_complex_7_cast_fp16)[name = tensor("op_1637_cast_fp16")]; + tensor freqs_7_to_fp16 = const()[name = tensor("freqs_7_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(303614848)))]; + tensor var_1643_cast_fp16 = mul(x = freqs_7_to_fp16, y = ts_23_cast_fp16)[name = tensor("op_1643_cast_fp16")]; + tensor rotr_7_cast_fp16 = cos(x = var_1643_cast_fp16)[name = tensor("rotr_7_cast_fp16")]; + tensor roti_7_cast_fp16 = sin(x = var_1643_cast_fp16)[name = tensor("roti_7_cast_fp16")]; + tensor var_1647_cast_fp16 = mul(x = var_1613_cast_fp16, y = rotr_7_cast_fp16)[name = tensor("op_1647_cast_fp16")]; + tensor var_1648_cast_fp16 = mul(x = var_1621_cast_fp16, y = roti_7_cast_fp16)[name = tensor("op_1648_cast_fp16")]; + tensor qor_13_cast_fp16 = sub(x = var_1647_cast_fp16, y = var_1648_cast_fp16)[name = tensor("qor_13_cast_fp16")]; + tensor var_1651_cast_fp16 = mul(x = var_1613_cast_fp16, y = roti_7_cast_fp16)[name = tensor("op_1651_cast_fp16")]; + tensor var_1652_cast_fp16 = mul(x = var_1621_cast_fp16, y = rotr_7_cast_fp16)[name = tensor("op_1652_cast_fp16")]; + tensor qoi_13_cast_fp16 = add(x = var_1651_cast_fp16, y = var_1652_cast_fp16)[name = tensor("qoi_13_cast_fp16")]; + tensor var_1655_cast_fp16 = mul(x = var_1629_cast_fp16, y = rotr_7_cast_fp16)[name = tensor("op_1655_cast_fp16")]; + tensor var_1656_cast_fp16 = mul(x = var_1637_cast_fp16, y = roti_7_cast_fp16)[name = tensor("op_1656_cast_fp16")]; + tensor kor_13_cast_fp16 = sub(x = var_1655_cast_fp16, y = var_1656_cast_fp16)[name = tensor("kor_13_cast_fp16")]; + tensor var_1659_cast_fp16 = mul(x = var_1629_cast_fp16, y = roti_7_cast_fp16)[name = tensor("op_1659_cast_fp16")]; + tensor var_1660_cast_fp16 = mul(x = var_1637_cast_fp16, y = rotr_7_cast_fp16)[name = tensor("op_1660_cast_fp16")]; + tensor koi_13_cast_fp16 = add(x = var_1659_cast_fp16, y = var_1660_cast_fp16)[name = tensor("koi_13_cast_fp16")]; + tensor qo_7_axis_0 = const()[name = tensor("qo_7_axis_0"), val = tensor(-1)]; + tensor qo_7_cast_fp16 = stack(axis = qo_7_axis_0, values = (qor_13_cast_fp16, qoi_13_cast_fp16))[name = tensor("qo_7_cast_fp16")]; + tensor ko_7_axis_0 = const()[name = tensor("ko_7_axis_0"), val = tensor(-1)]; + tensor ko_7_cast_fp16 = stack(axis = ko_7_axis_0, values = (kor_13_cast_fp16, koi_13_cast_fp16))[name = tensor("ko_7_cast_fp16")]; + tensor var_1689 = const()[name = tensor("op_1689"), val = tensor([1, 1, 16, 64])]; + tensor q_21_cast_fp16 = reshape(shape = var_1689, x = qo_7_cast_fp16)[name = tensor("q_21_cast_fp16")]; + tensor var_1691 = const()[name = tensor("op_1691"), val = tensor([1, 1, 16, 64])]; + tensor k_15_cast_fp16 = reshape(shape = var_1691, x = ko_7_cast_fp16)[name = tensor("k_15_cast_fp16")]; + tensor _inversed_1713_y_0_to_fp16 = const()[name = tensor("_inversed_1713_y_0_to_fp16"), val = tensor(0x1p-9)]; + tensor _inversed_1713_cast_fp16 = mul(x = ts_23_cast_fp16, y = _inversed_1713_y_0_to_fp16)[name = tensor("_inversed_1713_cast_fp16")]; + tensor var_1714_cast_fp16 = floor(x = _inversed_1713_cast_fp16)[name = tensor("op_1714_cast_fp16")]; + tensor var_1715_to_fp16 = const()[name = tensor("op_1715_to_fp16"), val = tensor(0x1p+9)]; + tensor var_1716_cast_fp16 = mul(x = var_1714_cast_fp16, y = var_1715_to_fp16)[name = tensor("op_1716_cast_fp16")]; + tensor write_indices_float_15_cast_fp16 = sub(x = ts_23_cast_fp16, y = var_1716_cast_fp16)[name = tensor("write_indices_float_15_cast_fp16")]; + tensor var_1723_dtype_0 = const()[name = tensor("op_1723_dtype_0"), val = tensor("int32")]; + tensor write_indices_7_reps_0 = const()[name = tensor("write_indices_7_reps_0"), val = tensor([1, 1, 16, 64])]; + tensor write_indices_float_15_cast_fp16_to_int32 = cast(dtype = var_1723_dtype_0, x = write_indices_float_15_cast_fp16)[name = tensor("cast_537")]; + tensor write_indices_7 = tile(reps = write_indices_7_reps_0, x = write_indices_float_15_cast_fp16_to_int32)[name = tensor("write_indices_7")]; + tensor var_1731_begin_0 = const()[name = tensor("op_1731_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_1731_end_0 = const()[name = tensor("op_1731_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor var_1731_end_mask_0 = const()[name = tensor("op_1731_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_1731_squeeze_mask_0 = const()[name = tensor("op_1731_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor cache3_to_fp16_dtype_0 = const()[name = tensor("cache3_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor cache3_to_fp16 = cast(dtype = cache3_to_fp16_dtype_0, x = cache3)[name = tensor("cast_536")]; + tensor var_1731_cast_fp16 = slice_by_index(begin = var_1731_begin_0, end = var_1731_end_0, end_mask = var_1731_end_mask_0, squeeze_mask = var_1731_squeeze_mask_0, x = cache3_to_fp16)[name = tensor("op_1731_cast_fp16")]; + tensor var_1733_axis_0 = const()[name = tensor("op_1733_axis_0"), val = tensor(1)]; + tensor var_1733_mode_0 = const()[name = tensor("op_1733_mode_0"), val = tensor("update")]; + tensor var_1733_validate_indices_0 = const()[name = tensor("op_1733_validate_indices_0"), val = tensor(false)]; + tensor var_1733_cast_fp16 = scatter_along_axis(axis = var_1733_axis_0, data = var_1731_cast_fp16, indices = write_indices_7, mode = var_1733_mode_0, updates = k_15_cast_fp16, validate_indices = var_1733_validate_indices_0)[name = tensor("op_1733_cast_fp16")]; + tensor concat_23 = const()[name = tensor("concat_23"), val = tensor([0, 0, 0, 0, 0])]; + tensor concat_24 = const()[name = tensor("concat_24"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_7_internal_tensor_assign_1_stride_0 = const()[name = tensor("new_cache_7_internal_tensor_assign_1_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_7_internal_tensor_assign_1_begin_mask_0 = const()[name = tensor("new_cache_7_internal_tensor_assign_1_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_7_internal_tensor_assign_1_end_mask_0 = const()[name = tensor("new_cache_7_internal_tensor_assign_1_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_7_internal_tensor_assign_1_squeeze_mask_0 = const()[name = tensor("new_cache_7_internal_tensor_assign_1_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_54 = const()[name = tensor("shape_54"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_6 = const()[name = tensor("reduce_prod_6"), val = tensor(1048576)]; + tensor range_1d_6_start_0 = const()[name = tensor("range_1d_6_start_0"), val = tensor(0)]; + tensor range_1d_6_step_0 = const()[name = tensor("range_1d_6_step_0"), val = tensor(1)]; + tensor range_1d_6 = range_1d(end = reduce_prod_6, start = range_1d_6_start_0, step = range_1d_6_step_0)[name = tensor("range_1d_6")]; + tensor reshape_30 = reshape(shape = shape_54, x = range_1d_6)[name = tensor("reshape_30")]; + tensor slice_by_index_6 = slice_by_index(begin = concat_23, begin_mask = new_cache_7_internal_tensor_assign_1_begin_mask_0, end = concat_24, end_mask = new_cache_7_internal_tensor_assign_1_end_mask_0, squeeze_mask = new_cache_7_internal_tensor_assign_1_squeeze_mask_0, stride = new_cache_7_internal_tensor_assign_1_stride_0, x = reshape_30)[name = tensor("slice_by_index_6")]; + tensor reshape_31_shape_0 = const()[name = tensor("reshape_31_shape_0"), val = tensor([-1])]; + tensor reshape_31 = reshape(shape = reshape_31_shape_0, x = slice_by_index_6)[name = tensor("reshape_31")]; + tensor reshape_32_shape_0 = const()[name = tensor("reshape_32_shape_0"), val = tensor([-1])]; + tensor reshape_32_cast_fp16 = reshape(shape = reshape_32_shape_0, x = var_1733_cast_fp16)[name = tensor("reshape_32_cast_fp16")]; + tensor reshape_33_shape_0 = const()[name = tensor("reshape_33_shape_0"), val = tensor([-1])]; + tensor reshape_33_cast_fp16 = reshape(shape = reshape_33_shape_0, x = cache3_to_fp16)[name = tensor("reshape_33_cast_fp16")]; + tensor scatter_6_mode_0 = const()[name = tensor("scatter_6_mode_0"), val = tensor("update")]; + tensor scatter_6_axis_0 = const()[name = tensor("scatter_6_axis_0"), val = tensor(0)]; + tensor scatter_6_validate_indices_0 = const()[name = tensor("scatter_6_validate_indices_0"), val = tensor(false)]; + tensor scatter_6_cast_fp16 = scatter(axis = scatter_6_axis_0, data = reshape_33_cast_fp16, indices = reshape_31, mode = scatter_6_mode_0, updates = reshape_32_cast_fp16, validate_indices = scatter_6_validate_indices_0)[name = tensor("scatter_6_cast_fp16")]; + tensor reshape_34_cast_fp16 = reshape(shape = shape_54, x = scatter_6_cast_fp16)[name = tensor("reshape_34_cast_fp16")]; + tensor var_1741_begin_0 = const()[name = tensor("op_1741_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor var_1741_end_0 = const()[name = tensor("op_1741_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor var_1741_end_mask_0 = const()[name = tensor("op_1741_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_1741_squeeze_mask_0 = const()[name = tensor("op_1741_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor var_1741_cast_fp16 = slice_by_index(begin = var_1741_begin_0, end = var_1741_end_0, end_mask = var_1741_end_mask_0, squeeze_mask = var_1741_squeeze_mask_0, x = reshape_34_cast_fp16)[name = tensor("op_1741_cast_fp16")]; + tensor var_1743_axis_0 = const()[name = tensor("op_1743_axis_0"), val = tensor(1)]; + tensor var_1743_mode_0 = const()[name = tensor("op_1743_mode_0"), val = tensor("update")]; + tensor var_1743_validate_indices_0 = const()[name = tensor("op_1743_validate_indices_0"), val = tensor(false)]; + tensor var_1743_cast_fp16 = scatter_along_axis(axis = var_1743_axis_0, data = var_1741_cast_fp16, indices = write_indices_7, mode = var_1743_mode_0, updates = v_7_cast_fp16, validate_indices = var_1743_validate_indices_0)[name = tensor("op_1743_cast_fp16")]; + tensor concat_25 = const()[name = tensor("concat_25"), val = tensor([1, 0, 0, 0, 0])]; + tensor concat_26 = const()[name = tensor("concat_26"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_7_internal_tensor_assign_2_stride_0 = const()[name = tensor("new_cache_7_internal_tensor_assign_2_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_7_internal_tensor_assign_2_begin_mask_0 = const()[name = tensor("new_cache_7_internal_tensor_assign_2_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_7_internal_tensor_assign_2_end_mask_0 = const()[name = tensor("new_cache_7_internal_tensor_assign_2_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_7_internal_tensor_assign_2_squeeze_mask_0 = const()[name = tensor("new_cache_7_internal_tensor_assign_2_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_55 = const()[name = tensor("shape_55"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_7 = const()[name = tensor("reduce_prod_7"), val = tensor(1048576)]; + tensor range_1d_7_start_0 = const()[name = tensor("range_1d_7_start_0"), val = tensor(0)]; + tensor range_1d_7_step_0 = const()[name = tensor("range_1d_7_step_0"), val = tensor(1)]; + tensor range_1d_7 = range_1d(end = reduce_prod_7, start = range_1d_7_start_0, step = range_1d_7_step_0)[name = tensor("range_1d_7")]; + tensor reshape_35 = reshape(shape = shape_55, x = range_1d_7)[name = tensor("reshape_35")]; + tensor slice_by_index_7 = slice_by_index(begin = concat_25, begin_mask = new_cache_7_internal_tensor_assign_2_begin_mask_0, end = concat_26, end_mask = new_cache_7_internal_tensor_assign_2_end_mask_0, squeeze_mask = new_cache_7_internal_tensor_assign_2_squeeze_mask_0, stride = new_cache_7_internal_tensor_assign_2_stride_0, x = reshape_35)[name = tensor("slice_by_index_7")]; + tensor reshape_36_shape_0 = const()[name = tensor("reshape_36_shape_0"), val = tensor([-1])]; + tensor reshape_36 = reshape(shape = reshape_36_shape_0, x = slice_by_index_7)[name = tensor("reshape_36")]; + tensor reshape_37_shape_0 = const()[name = tensor("reshape_37_shape_0"), val = tensor([-1])]; + tensor reshape_37_cast_fp16 = reshape(shape = reshape_37_shape_0, x = var_1743_cast_fp16)[name = tensor("reshape_37_cast_fp16")]; + tensor reshape_38_shape_0 = const()[name = tensor("reshape_38_shape_0"), val = tensor([-1])]; + tensor reshape_38_cast_fp16 = reshape(shape = reshape_38_shape_0, x = reshape_34_cast_fp16)[name = tensor("reshape_38_cast_fp16")]; + tensor scatter_7_mode_0 = const()[name = tensor("scatter_7_mode_0"), val = tensor("update")]; + tensor scatter_7_axis_0 = const()[name = tensor("scatter_7_axis_0"), val = tensor(0)]; + tensor scatter_7_validate_indices_0 = const()[name = tensor("scatter_7_validate_indices_0"), val = tensor(false)]; + tensor scatter_7_cast_fp16 = scatter(axis = scatter_7_axis_0, data = reshape_38_cast_fp16, indices = reshape_36, mode = scatter_7_mode_0, updates = reshape_37_cast_fp16, validate_indices = scatter_7_validate_indices_0)[name = tensor("scatter_7_cast_fp16")]; + tensor reshape_39_cast_fp16 = reshape(shape = shape_55, x = scatter_7_cast_fp16)[name = tensor("reshape_39_cast_fp16")]; + tensor reshape_39_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("reshape_39_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor keys_19_begin_0 = const()[name = tensor("keys_19_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor keys_19_end_0 = const()[name = tensor("keys_19_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor keys_19_end_mask_0 = const()[name = tensor("keys_19_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor keys_19_squeeze_mask_0 = const()[name = tensor("keys_19_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor keys_19_cast_fp16 = slice_by_index(begin = keys_19_begin_0, end = keys_19_end_0, end_mask = keys_19_end_mask_0, squeeze_mask = keys_19_squeeze_mask_0, x = reshape_39_cast_fp16)[name = tensor("keys_19_cast_fp16")]; + tensor values_19_begin_0 = const()[name = tensor("values_19_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor values_19_end_0 = const()[name = tensor("values_19_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor values_19_end_mask_0 = const()[name = tensor("values_19_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor values_19_squeeze_mask_0 = const()[name = tensor("values_19_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor values_19_cast_fp16 = slice_by_index(begin = values_19_begin_0, end = values_19_end_0, end_mask = values_19_end_mask_0, squeeze_mask = values_19_squeeze_mask_0, x = reshape_39_cast_fp16)[name = tensor("values_19_cast_fp16")]; + tensor var_1755_cast_fp16 = not_equal(x = keys_19_cast_fp16, y = keys_19_cast_fp16)[name = tensor("op_1755_cast_fp16")]; + tensor keys_21_cast_fp16 = select(a = var_504_to_fp16, b = keys_19_cast_fp16, cond = var_1755_cast_fp16)[name = tensor("keys_21_cast_fp16")]; + tensor var_1763_cast_fp16 = not_equal(x = values_19_cast_fp16, y = values_19_cast_fp16)[name = tensor("op_1763_cast_fp16")]; + tensor values_21_cast_fp16 = select(a = var_504_to_fp16, b = values_19_cast_fp16, cond = var_1763_cast_fp16)[name = tensor("values_21_cast_fp16")]; + tensor var_1787 = const()[name = tensor("op_1787"), val = tensor([0, 2, 1, 3])]; + tensor var_1800 = const()[name = tensor("op_1800"), val = tensor([1, 1, 1])]; + tensor var_1801_cast_fp16 = reshape(shape = var_1800, x = position3_to_fp16)[name = tensor("op_1801_cast_fp16")]; + tensor var_1818_to_fp16 = const()[name = tensor("op_1818_to_fp16"), val = tensor(0x1p+0)]; + tensor valid_len_7_cast_fp16 = add(x = var_1801_cast_fp16, y = var_1818_to_fp16)[name = tensor("valid_len_7_cast_fp16")]; + tensor valid_mask_7_cast_fp16 = less(x = k_positions_1_promoted_to_fp16, y = valid_len_7_cast_fp16)[name = tensor("valid_mask_7_cast_fp16")]; + tensor causal_mask_7_cast_fp16 = less_equal(x = k_positions_1_promoted_to_fp16, y = var_1801_cast_fp16)[name = tensor("causal_mask_7_cast_fp16")]; + tensor attn_mask_13 = logical_and(x = valid_mask_7_cast_fp16, y = causal_mask_7_cast_fp16)[name = tensor("attn_mask_13")]; + tensor attn_mask_15_axes_0 = const()[name = tensor("attn_mask_15_axes_0"), val = tensor([1])]; + tensor attn_mask_15 = expand_dims(axes = attn_mask_15_axes_0, x = attn_mask_13)[name = tensor("attn_mask_15")]; + tensor var_1836_transpose_x_0 = const()[name = tensor("op_1836_transpose_x_0"), val = tensor(false)]; + tensor var_1836_transpose_y_0 = const()[name = tensor("op_1836_transpose_y_0"), val = tensor(false)]; + tensor transpose_78_perm_0 = const()[name = tensor("transpose_78_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_79_perm_0 = const()[name = tensor("transpose_79_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_79 = transpose(perm = transpose_79_perm_0, x = keys_21_cast_fp16)[name = tensor("transpose_201")]; + tensor transpose_78 = transpose(perm = transpose_78_perm_0, x = q_21_cast_fp16)[name = tensor("transpose_202")]; + tensor var_1836_cast_fp16 = matmul(transpose_x = var_1836_transpose_x_0, transpose_y = var_1836_transpose_y_0, x = transpose_78, y = transpose_79)[name = tensor("op_1836_cast_fp16")]; + tensor var_1830_to_fp16 = const()[name = tensor("op_1830_to_fp16"), val = tensor([0x1p-3])]; + tensor attn_weights_19_cast_fp16 = mul(x = var_1836_cast_fp16, y = var_1830_to_fp16)[name = tensor("attn_weights_19_cast_fp16")]; + tensor var_1838 = logical_not(x = attn_mask_15)[name = tensor("op_1838")]; + tensor var_1839_to_fp16 = const()[name = tensor("op_1839_to_fp16"), val = tensor(-inf)]; + tensor attn_weights_21_cast_fp16 = select(a = var_1839_to_fp16, b = attn_weights_19_cast_fp16, cond = var_1838)[name = tensor("attn_weights_21_cast_fp16")]; + tensor var_1841 = const()[name = tensor("op_1841"), val = tensor(-1)]; + tensor attn_weights_23_cast_fp16 = softmax(axis = var_1841, x = attn_weights_21_cast_fp16)[name = tensor("attn_weights_23_cast_fp16")]; + tensor attn_output_7_transpose_x_0 = const()[name = tensor("attn_output_7_transpose_x_0"), val = tensor(false)]; + tensor attn_output_7_transpose_y_0 = const()[name = tensor("attn_output_7_transpose_y_0"), val = tensor(false)]; + tensor values_23_cast_fp16 = transpose(perm = var_1787, x = values_21_cast_fp16)[name = tensor("transpose_203")]; + tensor attn_output_7_cast_fp16 = matmul(transpose_x = attn_output_7_transpose_x_0, transpose_y = attn_output_7_transpose_y_0, x = attn_weights_23_cast_fp16, y = values_23_cast_fp16)[name = tensor("attn_output_7_cast_fp16")]; + tensor var_1849 = const()[name = tensor("op_1849"), val = tensor([0, 2, 1, 3])]; + tensor var_1852 = const()[name = tensor("op_1852"), val = tensor([1, 1, 1024])]; + tensor var_1850_cast_fp16 = transpose(perm = var_1849, x = attn_output_7_cast_fp16)[name = tensor("transpose_200")]; + tensor input_35_cast_fp16 = reshape(shape = var_1852, x = var_1850_cast_fp16)[name = tensor("input_35_cast_fp16")]; + tensor linear_14_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = attn3_out_proj_weight_cast_fp16, x = input_35_cast_fp16)[name = tensor("linear_14_cast_fp16")]; + tensor var_1858_to_fp16 = const()[name = tensor("op_1858_to_fp16"), val = tensor(0x1p+0)]; + tensor var_1859_cast_fp16 = add(x = position3_to_fp16, y = var_1858_to_fp16)[name = tensor("op_1859_cast_fp16")]; + tensor var_1859_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_1859_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor input_37_cast_fp16 = add(x = input_33_cast_fp16, y = linear_14_cast_fp16)[name = tensor("input_37_cast_fp16")]; + tensor input_39_axes_0 = const()[name = tensor("input_39_axes_0"), val = tensor([-1])]; + tensor norm3_2_weight_to_fp16 = const()[name = tensor("norm3_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(303614976)))]; + tensor norm3_2_bias_to_fp16 = const()[name = tensor("norm3_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(303617088)))]; + tensor var_1863_to_fp16 = const()[name = tensor("op_1863_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_39_cast_fp16 = layer_norm(axes = input_39_axes_0, beta = norm3_2_bias_to_fp16, epsilon = var_1863_to_fp16, gamma = norm3_2_weight_to_fp16, x = input_37_cast_fp16)[name = tensor("input_39_cast_fp16")]; + tensor linear_15_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = linear3_1_weight_cast_fp16, x = input_39_cast_fp16)[name = tensor("linear_15_cast_fp16")]; + tensor input_41_mode_0 = const()[name = tensor("input_41_mode_0"), val = tensor("EXACT")]; + tensor input_41_cast_fp16 = gelu(mode = input_41_mode_0, x = linear_15_cast_fp16)[name = tensor("input_41_cast_fp16")]; + tensor linear_16_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = linear3_2_weight_cast_fp16, x = input_41_cast_fp16)[name = tensor("linear_16_cast_fp16")]; + tensor input_43_cast_fp16 = add(x = input_37_cast_fp16, y = linear_16_cast_fp16)[name = tensor("input_43_cast_fp16")]; + tensor x_9_axes_0 = const()[name = tensor("x_9_axes_0"), val = tensor([-1])]; + tensor norm4_1_weight_to_fp16 = const()[name = tensor("norm4_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(303619200)))]; + tensor norm4_1_bias_to_fp16 = const()[name = tensor("norm4_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(303621312)))]; + tensor var_1880_to_fp16 = const()[name = tensor("op_1880_to_fp16"), val = tensor(0x1.5p-17)]; + tensor x_9_cast_fp16 = layer_norm(axes = x_9_axes_0, beta = norm4_1_bias_to_fp16, epsilon = var_1880_to_fp16, gamma = norm4_1_weight_to_fp16, x = input_43_cast_fp16)[name = tensor("x_9_cast_fp16")]; + tensor linear_17_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = attn4_in_proj_weight_cast_fp16, x = x_9_cast_fp16)[name = tensor("linear_17_cast_fp16")]; + tensor var_1916 = const()[name = tensor("op_1916"), val = tensor([1, 1, 3, 16, 64])]; + tensor qkv_9_cast_fp16 = reshape(shape = var_1916, x = linear_17_cast_fp16)[name = tensor("qkv_9_cast_fp16")]; + tensor q_25_begin_0 = const()[name = tensor("q_25_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor q_25_end_0 = const()[name = tensor("q_25_end_0"), val = tensor([1, 1, 1, 16, 64])]; + tensor q_25_end_mask_0 = const()[name = tensor("q_25_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor q_25_squeeze_mask_0 = const()[name = tensor("q_25_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor q_25_cast_fp16 = slice_by_index(begin = q_25_begin_0, end = q_25_end_0, end_mask = q_25_end_mask_0, squeeze_mask = q_25_squeeze_mask_0, x = qkv_9_cast_fp16)[name = tensor("q_25_cast_fp16")]; + tensor k_17_begin_0 = const()[name = tensor("k_17_begin_0"), val = tensor([0, 0, 1, 0, 0])]; + tensor k_17_end_0 = const()[name = tensor("k_17_end_0"), val = tensor([1, 1, 2, 16, 64])]; + tensor k_17_end_mask_0 = const()[name = tensor("k_17_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor k_17_squeeze_mask_0 = const()[name = tensor("k_17_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor k_17_cast_fp16 = slice_by_index(begin = k_17_begin_0, end = k_17_end_0, end_mask = k_17_end_mask_0, squeeze_mask = k_17_squeeze_mask_0, x = qkv_9_cast_fp16)[name = tensor("k_17_cast_fp16")]; + tensor v_9_begin_0 = const()[name = tensor("v_9_begin_0"), val = tensor([0, 0, 2, 0, 0])]; + tensor v_9_end_0 = const()[name = tensor("v_9_end_0"), val = tensor([1, 1, 3, 16, 64])]; + tensor v_9_end_mask_0 = const()[name = tensor("v_9_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor v_9_squeeze_mask_0 = const()[name = tensor("v_9_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor v_9_cast_fp16 = slice_by_index(begin = v_9_begin_0, end = v_9_end_0, end_mask = v_9_end_mask_0, squeeze_mask = v_9_squeeze_mask_0, x = qkv_9_cast_fp16)[name = tensor("v_9_cast_fp16")]; + tensor var_2020 = const()[name = tensor("op_2020"), val = tensor([1, 1, 1, 1])]; + tensor position4_to_fp16_dtype_0 = const()[name = tensor("position4_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor position4_to_fp16 = cast(dtype = position4_to_fp16_dtype_0, x = position4)[name = tensor("cast_533")]; + tensor ts_29_cast_fp16 = reshape(shape = var_2020, x = position4_to_fp16)[name = tensor("ts_29_cast_fp16")]; + tensor var_2024 = const()[name = tensor("op_2024"), val = tensor([1, 1, 16, 32, 2])]; + tensor q_complex_9_cast_fp16 = reshape(shape = var_2024, x = q_25_cast_fp16)[name = tensor("q_complex_9_cast_fp16")]; + tensor var_2028 = const()[name = tensor("op_2028"), val = tensor([1, 1, 16, 32, 2])]; + tensor k_complex_9_cast_fp16 = reshape(shape = var_2028, x = k_17_cast_fp16)[name = tensor("k_complex_9_cast_fp16")]; + tensor var_2032_begin_0 = const()[name = tensor("op_2032_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_2032_end_0 = const()[name = tensor("op_2032_end_0"), val = tensor([1, 1, 16, 32, 1])]; + tensor var_2032_end_mask_0 = const()[name = tensor("op_2032_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_2032_squeeze_mask_0 = const()[name = tensor("op_2032_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_2032_cast_fp16 = slice_by_index(begin = var_2032_begin_0, end = var_2032_end_0, end_mask = var_2032_end_mask_0, squeeze_mask = var_2032_squeeze_mask_0, x = q_complex_9_cast_fp16)[name = tensor("op_2032_cast_fp16")]; + tensor var_2040_begin_0 = const()[name = tensor("op_2040_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_2040_end_0 = const()[name = tensor("op_2040_end_0"), val = tensor([1, 1, 16, 32, 2])]; + tensor var_2040_end_mask_0 = const()[name = tensor("op_2040_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_2040_squeeze_mask_0 = const()[name = tensor("op_2040_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_2040_cast_fp16 = slice_by_index(begin = var_2040_begin_0, end = var_2040_end_0, end_mask = var_2040_end_mask_0, squeeze_mask = var_2040_squeeze_mask_0, x = q_complex_9_cast_fp16)[name = tensor("op_2040_cast_fp16")]; + tensor var_2048_begin_0 = const()[name = tensor("op_2048_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_2048_end_0 = const()[name = tensor("op_2048_end_0"), val = tensor([1, 1, 16, 32, 1])]; + tensor var_2048_end_mask_0 = const()[name = tensor("op_2048_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_2048_squeeze_mask_0 = const()[name = tensor("op_2048_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_2048_cast_fp16 = slice_by_index(begin = var_2048_begin_0, end = var_2048_end_0, end_mask = var_2048_end_mask_0, squeeze_mask = var_2048_squeeze_mask_0, x = k_complex_9_cast_fp16)[name = tensor("op_2048_cast_fp16")]; + tensor var_2056_begin_0 = const()[name = tensor("op_2056_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_2056_end_0 = const()[name = tensor("op_2056_end_0"), val = tensor([1, 1, 16, 32, 2])]; + tensor var_2056_end_mask_0 = const()[name = tensor("op_2056_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_2056_squeeze_mask_0 = const()[name = tensor("op_2056_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_2056_cast_fp16 = slice_by_index(begin = var_2056_begin_0, end = var_2056_end_0, end_mask = var_2056_end_mask_0, squeeze_mask = var_2056_squeeze_mask_0, x = k_complex_9_cast_fp16)[name = tensor("op_2056_cast_fp16")]; + tensor freqs_9_to_fp16 = const()[name = tensor("freqs_9_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(303623424)))]; + tensor var_2062_cast_fp16 = mul(x = freqs_9_to_fp16, y = ts_29_cast_fp16)[name = tensor("op_2062_cast_fp16")]; + tensor rotr_9_cast_fp16 = cos(x = var_2062_cast_fp16)[name = tensor("rotr_9_cast_fp16")]; + tensor roti_9_cast_fp16 = sin(x = var_2062_cast_fp16)[name = tensor("roti_9_cast_fp16")]; + tensor var_2066_cast_fp16 = mul(x = var_2032_cast_fp16, y = rotr_9_cast_fp16)[name = tensor("op_2066_cast_fp16")]; + tensor var_2067_cast_fp16 = mul(x = var_2040_cast_fp16, y = roti_9_cast_fp16)[name = tensor("op_2067_cast_fp16")]; + tensor qor_17_cast_fp16 = sub(x = var_2066_cast_fp16, y = var_2067_cast_fp16)[name = tensor("qor_17_cast_fp16")]; + tensor var_2070_cast_fp16 = mul(x = var_2032_cast_fp16, y = roti_9_cast_fp16)[name = tensor("op_2070_cast_fp16")]; + tensor var_2071_cast_fp16 = mul(x = var_2040_cast_fp16, y = rotr_9_cast_fp16)[name = tensor("op_2071_cast_fp16")]; + tensor qoi_17_cast_fp16 = add(x = var_2070_cast_fp16, y = var_2071_cast_fp16)[name = tensor("qoi_17_cast_fp16")]; + tensor var_2074_cast_fp16 = mul(x = var_2048_cast_fp16, y = rotr_9_cast_fp16)[name = tensor("op_2074_cast_fp16")]; + tensor var_2075_cast_fp16 = mul(x = var_2056_cast_fp16, y = roti_9_cast_fp16)[name = tensor("op_2075_cast_fp16")]; + tensor kor_17_cast_fp16 = sub(x = var_2074_cast_fp16, y = var_2075_cast_fp16)[name = tensor("kor_17_cast_fp16")]; + tensor var_2078_cast_fp16 = mul(x = var_2048_cast_fp16, y = roti_9_cast_fp16)[name = tensor("op_2078_cast_fp16")]; + tensor var_2079_cast_fp16 = mul(x = var_2056_cast_fp16, y = rotr_9_cast_fp16)[name = tensor("op_2079_cast_fp16")]; + tensor koi_17_cast_fp16 = add(x = var_2078_cast_fp16, y = var_2079_cast_fp16)[name = tensor("koi_17_cast_fp16")]; + tensor qo_9_axis_0 = const()[name = tensor("qo_9_axis_0"), val = tensor(-1)]; + tensor qo_9_cast_fp16 = stack(axis = qo_9_axis_0, values = (qor_17_cast_fp16, qoi_17_cast_fp16))[name = tensor("qo_9_cast_fp16")]; + tensor ko_9_axis_0 = const()[name = tensor("ko_9_axis_0"), val = tensor(-1)]; + tensor ko_9_cast_fp16 = stack(axis = ko_9_axis_0, values = (kor_17_cast_fp16, koi_17_cast_fp16))[name = tensor("ko_9_cast_fp16")]; + tensor var_2108 = const()[name = tensor("op_2108"), val = tensor([1, 1, 16, 64])]; + tensor q_27_cast_fp16 = reshape(shape = var_2108, x = qo_9_cast_fp16)[name = tensor("q_27_cast_fp16")]; + tensor var_2110 = const()[name = tensor("op_2110"), val = tensor([1, 1, 16, 64])]; + tensor k_19_cast_fp16 = reshape(shape = var_2110, x = ko_9_cast_fp16)[name = tensor("k_19_cast_fp16")]; + tensor _inversed_2132_y_0_to_fp16 = const()[name = tensor("_inversed_2132_y_0_to_fp16"), val = tensor(0x1p-9)]; + tensor _inversed_2132_cast_fp16 = mul(x = ts_29_cast_fp16, y = _inversed_2132_y_0_to_fp16)[name = tensor("_inversed_2132_cast_fp16")]; + tensor var_2133_cast_fp16 = floor(x = _inversed_2132_cast_fp16)[name = tensor("op_2133_cast_fp16")]; + tensor var_2134_to_fp16 = const()[name = tensor("op_2134_to_fp16"), val = tensor(0x1p+9)]; + tensor var_2135_cast_fp16 = mul(x = var_2133_cast_fp16, y = var_2134_to_fp16)[name = tensor("op_2135_cast_fp16")]; + tensor write_indices_float_19_cast_fp16 = sub(x = ts_29_cast_fp16, y = var_2135_cast_fp16)[name = tensor("write_indices_float_19_cast_fp16")]; + tensor var_2142_dtype_0 = const()[name = tensor("op_2142_dtype_0"), val = tensor("int32")]; + tensor write_indices_9_reps_0 = const()[name = tensor("write_indices_9_reps_0"), val = tensor([1, 1, 16, 64])]; + tensor write_indices_float_19_cast_fp16_to_int32 = cast(dtype = var_2142_dtype_0, x = write_indices_float_19_cast_fp16)[name = tensor("cast_532")]; + tensor write_indices_9 = tile(reps = write_indices_9_reps_0, x = write_indices_float_19_cast_fp16_to_int32)[name = tensor("write_indices_9")]; + tensor var_2150_begin_0 = const()[name = tensor("op_2150_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_2150_end_0 = const()[name = tensor("op_2150_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor var_2150_end_mask_0 = const()[name = tensor("op_2150_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_2150_squeeze_mask_0 = const()[name = tensor("op_2150_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor cache4_to_fp16_dtype_0 = const()[name = tensor("cache4_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor cache4_to_fp16 = cast(dtype = cache4_to_fp16_dtype_0, x = cache4)[name = tensor("cast_531")]; + tensor var_2150_cast_fp16 = slice_by_index(begin = var_2150_begin_0, end = var_2150_end_0, end_mask = var_2150_end_mask_0, squeeze_mask = var_2150_squeeze_mask_0, x = cache4_to_fp16)[name = tensor("op_2150_cast_fp16")]; + tensor var_2152_axis_0 = const()[name = tensor("op_2152_axis_0"), val = tensor(1)]; + tensor var_2152_mode_0 = const()[name = tensor("op_2152_mode_0"), val = tensor("update")]; + tensor var_2152_validate_indices_0 = const()[name = tensor("op_2152_validate_indices_0"), val = tensor(false)]; + tensor var_2152_cast_fp16 = scatter_along_axis(axis = var_2152_axis_0, data = var_2150_cast_fp16, indices = write_indices_9, mode = var_2152_mode_0, updates = k_19_cast_fp16, validate_indices = var_2152_validate_indices_0)[name = tensor("op_2152_cast_fp16")]; + tensor concat_30 = const()[name = tensor("concat_30"), val = tensor([0, 0, 0, 0, 0])]; + tensor concat_31 = const()[name = tensor("concat_31"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_9_internal_tensor_assign_1_stride_0 = const()[name = tensor("new_cache_9_internal_tensor_assign_1_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_9_internal_tensor_assign_1_begin_mask_0 = const()[name = tensor("new_cache_9_internal_tensor_assign_1_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_9_internal_tensor_assign_1_end_mask_0 = const()[name = tensor("new_cache_9_internal_tensor_assign_1_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_9_internal_tensor_assign_1_squeeze_mask_0 = const()[name = tensor("new_cache_9_internal_tensor_assign_1_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_56 = const()[name = tensor("shape_56"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_8 = const()[name = tensor("reduce_prod_8"), val = tensor(1048576)]; + tensor range_1d_8_start_0 = const()[name = tensor("range_1d_8_start_0"), val = tensor(0)]; + tensor range_1d_8_step_0 = const()[name = tensor("range_1d_8_step_0"), val = tensor(1)]; + tensor range_1d_8 = range_1d(end = reduce_prod_8, start = range_1d_8_start_0, step = range_1d_8_step_0)[name = tensor("range_1d_8")]; + tensor reshape_40 = reshape(shape = shape_56, x = range_1d_8)[name = tensor("reshape_40")]; + tensor slice_by_index_8 = slice_by_index(begin = concat_30, begin_mask = new_cache_9_internal_tensor_assign_1_begin_mask_0, end = concat_31, end_mask = new_cache_9_internal_tensor_assign_1_end_mask_0, squeeze_mask = new_cache_9_internal_tensor_assign_1_squeeze_mask_0, stride = new_cache_9_internal_tensor_assign_1_stride_0, x = reshape_40)[name = tensor("slice_by_index_8")]; + tensor reshape_41_shape_0 = const()[name = tensor("reshape_41_shape_0"), val = tensor([-1])]; + tensor reshape_41 = reshape(shape = reshape_41_shape_0, x = slice_by_index_8)[name = tensor("reshape_41")]; + tensor reshape_42_shape_0 = const()[name = tensor("reshape_42_shape_0"), val = tensor([-1])]; + tensor reshape_42_cast_fp16 = reshape(shape = reshape_42_shape_0, x = var_2152_cast_fp16)[name = tensor("reshape_42_cast_fp16")]; + tensor reshape_43_shape_0 = const()[name = tensor("reshape_43_shape_0"), val = tensor([-1])]; + tensor reshape_43_cast_fp16 = reshape(shape = reshape_43_shape_0, x = cache4_to_fp16)[name = tensor("reshape_43_cast_fp16")]; + tensor scatter_8_mode_0 = const()[name = tensor("scatter_8_mode_0"), val = tensor("update")]; + tensor scatter_8_axis_0 = const()[name = tensor("scatter_8_axis_0"), val = tensor(0)]; + tensor scatter_8_validate_indices_0 = const()[name = tensor("scatter_8_validate_indices_0"), val = tensor(false)]; + tensor scatter_8_cast_fp16 = scatter(axis = scatter_8_axis_0, data = reshape_43_cast_fp16, indices = reshape_41, mode = scatter_8_mode_0, updates = reshape_42_cast_fp16, validate_indices = scatter_8_validate_indices_0)[name = tensor("scatter_8_cast_fp16")]; + tensor reshape_44_cast_fp16 = reshape(shape = shape_56, x = scatter_8_cast_fp16)[name = tensor("reshape_44_cast_fp16")]; + tensor var_2160_begin_0 = const()[name = tensor("op_2160_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor var_2160_end_0 = const()[name = tensor("op_2160_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor var_2160_end_mask_0 = const()[name = tensor("op_2160_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_2160_squeeze_mask_0 = const()[name = tensor("op_2160_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor var_2160_cast_fp16 = slice_by_index(begin = var_2160_begin_0, end = var_2160_end_0, end_mask = var_2160_end_mask_0, squeeze_mask = var_2160_squeeze_mask_0, x = reshape_44_cast_fp16)[name = tensor("op_2160_cast_fp16")]; + tensor var_2162_axis_0 = const()[name = tensor("op_2162_axis_0"), val = tensor(1)]; + tensor var_2162_mode_0 = const()[name = tensor("op_2162_mode_0"), val = tensor("update")]; + tensor var_2162_validate_indices_0 = const()[name = tensor("op_2162_validate_indices_0"), val = tensor(false)]; + tensor var_2162_cast_fp16 = scatter_along_axis(axis = var_2162_axis_0, data = var_2160_cast_fp16, indices = write_indices_9, mode = var_2162_mode_0, updates = v_9_cast_fp16, validate_indices = var_2162_validate_indices_0)[name = tensor("op_2162_cast_fp16")]; + tensor concat_32 = const()[name = tensor("concat_32"), val = tensor([1, 0, 0, 0, 0])]; + tensor concat_33 = const()[name = tensor("concat_33"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_9_internal_tensor_assign_2_stride_0 = const()[name = tensor("new_cache_9_internal_tensor_assign_2_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_9_internal_tensor_assign_2_begin_mask_0 = const()[name = tensor("new_cache_9_internal_tensor_assign_2_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_9_internal_tensor_assign_2_end_mask_0 = const()[name = tensor("new_cache_9_internal_tensor_assign_2_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_9_internal_tensor_assign_2_squeeze_mask_0 = const()[name = tensor("new_cache_9_internal_tensor_assign_2_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_57 = const()[name = tensor("shape_57"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_9 = const()[name = tensor("reduce_prod_9"), val = tensor(1048576)]; + tensor range_1d_9_start_0 = const()[name = tensor("range_1d_9_start_0"), val = tensor(0)]; + tensor range_1d_9_step_0 = const()[name = tensor("range_1d_9_step_0"), val = tensor(1)]; + tensor range_1d_9 = range_1d(end = reduce_prod_9, start = range_1d_9_start_0, step = range_1d_9_step_0)[name = tensor("range_1d_9")]; + tensor reshape_45 = reshape(shape = shape_57, x = range_1d_9)[name = tensor("reshape_45")]; + tensor slice_by_index_9 = slice_by_index(begin = concat_32, begin_mask = new_cache_9_internal_tensor_assign_2_begin_mask_0, end = concat_33, end_mask = new_cache_9_internal_tensor_assign_2_end_mask_0, squeeze_mask = new_cache_9_internal_tensor_assign_2_squeeze_mask_0, stride = new_cache_9_internal_tensor_assign_2_stride_0, x = reshape_45)[name = tensor("slice_by_index_9")]; + tensor reshape_46_shape_0 = const()[name = tensor("reshape_46_shape_0"), val = tensor([-1])]; + tensor reshape_46 = reshape(shape = reshape_46_shape_0, x = slice_by_index_9)[name = tensor("reshape_46")]; + tensor reshape_47_shape_0 = const()[name = tensor("reshape_47_shape_0"), val = tensor([-1])]; + tensor reshape_47_cast_fp16 = reshape(shape = reshape_47_shape_0, x = var_2162_cast_fp16)[name = tensor("reshape_47_cast_fp16")]; + tensor reshape_48_shape_0 = const()[name = tensor("reshape_48_shape_0"), val = tensor([-1])]; + tensor reshape_48_cast_fp16 = reshape(shape = reshape_48_shape_0, x = reshape_44_cast_fp16)[name = tensor("reshape_48_cast_fp16")]; + tensor scatter_9_mode_0 = const()[name = tensor("scatter_9_mode_0"), val = tensor("update")]; + tensor scatter_9_axis_0 = const()[name = tensor("scatter_9_axis_0"), val = tensor(0)]; + tensor scatter_9_validate_indices_0 = const()[name = tensor("scatter_9_validate_indices_0"), val = tensor(false)]; + tensor scatter_9_cast_fp16 = scatter(axis = scatter_9_axis_0, data = reshape_48_cast_fp16, indices = reshape_46, mode = scatter_9_mode_0, updates = reshape_47_cast_fp16, validate_indices = scatter_9_validate_indices_0)[name = tensor("scatter_9_cast_fp16")]; + tensor reshape_49_cast_fp16 = reshape(shape = shape_57, x = scatter_9_cast_fp16)[name = tensor("reshape_49_cast_fp16")]; + tensor reshape_49_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("reshape_49_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor keys_25_begin_0 = const()[name = tensor("keys_25_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor keys_25_end_0 = const()[name = tensor("keys_25_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor keys_25_end_mask_0 = const()[name = tensor("keys_25_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor keys_25_squeeze_mask_0 = const()[name = tensor("keys_25_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor keys_25_cast_fp16 = slice_by_index(begin = keys_25_begin_0, end = keys_25_end_0, end_mask = keys_25_end_mask_0, squeeze_mask = keys_25_squeeze_mask_0, x = reshape_49_cast_fp16)[name = tensor("keys_25_cast_fp16")]; + tensor values_25_begin_0 = const()[name = tensor("values_25_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor values_25_end_0 = const()[name = tensor("values_25_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor values_25_end_mask_0 = const()[name = tensor("values_25_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor values_25_squeeze_mask_0 = const()[name = tensor("values_25_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor values_25_cast_fp16 = slice_by_index(begin = values_25_begin_0, end = values_25_end_0, end_mask = values_25_end_mask_0, squeeze_mask = values_25_squeeze_mask_0, x = reshape_49_cast_fp16)[name = tensor("values_25_cast_fp16")]; + tensor var_2174_cast_fp16 = not_equal(x = keys_25_cast_fp16, y = keys_25_cast_fp16)[name = tensor("op_2174_cast_fp16")]; + tensor keys_27_cast_fp16 = select(a = var_504_to_fp16, b = keys_25_cast_fp16, cond = var_2174_cast_fp16)[name = tensor("keys_27_cast_fp16")]; + tensor var_2182_cast_fp16 = not_equal(x = values_25_cast_fp16, y = values_25_cast_fp16)[name = tensor("op_2182_cast_fp16")]; + tensor values_27_cast_fp16 = select(a = var_504_to_fp16, b = values_25_cast_fp16, cond = var_2182_cast_fp16)[name = tensor("values_27_cast_fp16")]; + tensor var_2206 = const()[name = tensor("op_2206"), val = tensor([0, 2, 1, 3])]; + tensor var_2219 = const()[name = tensor("op_2219"), val = tensor([1, 1, 1])]; + tensor var_2220_cast_fp16 = reshape(shape = var_2219, x = position4_to_fp16)[name = tensor("op_2220_cast_fp16")]; + tensor var_2237_to_fp16 = const()[name = tensor("op_2237_to_fp16"), val = tensor(0x1p+0)]; + tensor valid_len_9_cast_fp16 = add(x = var_2220_cast_fp16, y = var_2237_to_fp16)[name = tensor("valid_len_9_cast_fp16")]; + tensor valid_mask_9_cast_fp16 = less(x = k_positions_1_promoted_to_fp16, y = valid_len_9_cast_fp16)[name = tensor("valid_mask_9_cast_fp16")]; + tensor causal_mask_9_cast_fp16 = less_equal(x = k_positions_1_promoted_to_fp16, y = var_2220_cast_fp16)[name = tensor("causal_mask_9_cast_fp16")]; + tensor attn_mask_17 = logical_and(x = valid_mask_9_cast_fp16, y = causal_mask_9_cast_fp16)[name = tensor("attn_mask_17")]; + tensor attn_mask_19_axes_0 = const()[name = tensor("attn_mask_19_axes_0"), val = tensor([1])]; + tensor attn_mask_19 = expand_dims(axes = attn_mask_19_axes_0, x = attn_mask_17)[name = tensor("attn_mask_19")]; + tensor var_2255_transpose_x_0 = const()[name = tensor("op_2255_transpose_x_0"), val = tensor(false)]; + tensor var_2255_transpose_y_0 = const()[name = tensor("op_2255_transpose_y_0"), val = tensor(false)]; + tensor transpose_80_perm_0 = const()[name = tensor("transpose_80_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_81_perm_0 = const()[name = tensor("transpose_81_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_81 = transpose(perm = transpose_81_perm_0, x = keys_27_cast_fp16)[name = tensor("transpose_197")]; + tensor transpose_80 = transpose(perm = transpose_80_perm_0, x = q_27_cast_fp16)[name = tensor("transpose_198")]; + tensor var_2255_cast_fp16 = matmul(transpose_x = var_2255_transpose_x_0, transpose_y = var_2255_transpose_y_0, x = transpose_80, y = transpose_81)[name = tensor("op_2255_cast_fp16")]; + tensor var_2249_to_fp16 = const()[name = tensor("op_2249_to_fp16"), val = tensor([0x1p-3])]; + tensor attn_weights_25_cast_fp16 = mul(x = var_2255_cast_fp16, y = var_2249_to_fp16)[name = tensor("attn_weights_25_cast_fp16")]; + tensor var_2257 = logical_not(x = attn_mask_19)[name = tensor("op_2257")]; + tensor var_2258_to_fp16 = const()[name = tensor("op_2258_to_fp16"), val = tensor(-inf)]; + tensor attn_weights_27_cast_fp16 = select(a = var_2258_to_fp16, b = attn_weights_25_cast_fp16, cond = var_2257)[name = tensor("attn_weights_27_cast_fp16")]; + tensor var_2260 = const()[name = tensor("op_2260"), val = tensor(-1)]; + tensor attn_weights_29_cast_fp16 = softmax(axis = var_2260, x = attn_weights_27_cast_fp16)[name = tensor("attn_weights_29_cast_fp16")]; + tensor attn_output_9_transpose_x_0 = const()[name = tensor("attn_output_9_transpose_x_0"), val = tensor(false)]; + tensor attn_output_9_transpose_y_0 = const()[name = tensor("attn_output_9_transpose_y_0"), val = tensor(false)]; + tensor values_29_cast_fp16 = transpose(perm = var_2206, x = values_27_cast_fp16)[name = tensor("transpose_199")]; + tensor attn_output_9_cast_fp16 = matmul(transpose_x = attn_output_9_transpose_x_0, transpose_y = attn_output_9_transpose_y_0, x = attn_weights_29_cast_fp16, y = values_29_cast_fp16)[name = tensor("attn_output_9_cast_fp16")]; + tensor var_2268 = const()[name = tensor("op_2268"), val = tensor([0, 2, 1, 3])]; + tensor var_2271 = const()[name = tensor("op_2271"), val = tensor([1, 1, 1024])]; + tensor var_2269_cast_fp16 = transpose(perm = var_2268, x = attn_output_9_cast_fp16)[name = tensor("transpose_196")]; + tensor input_45_cast_fp16 = reshape(shape = var_2271, x = var_2269_cast_fp16)[name = tensor("input_45_cast_fp16")]; + tensor linear_18_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = attn4_out_proj_weight_cast_fp16, x = input_45_cast_fp16)[name = tensor("linear_18_cast_fp16")]; + tensor var_2277_to_fp16 = const()[name = tensor("op_2277_to_fp16"), val = tensor(0x1p+0)]; + tensor var_2278_cast_fp16 = add(x = position4_to_fp16, y = var_2277_to_fp16)[name = tensor("op_2278_cast_fp16")]; + tensor var_2278_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_2278_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor input_47_cast_fp16 = add(x = input_43_cast_fp16, y = linear_18_cast_fp16)[name = tensor("input_47_cast_fp16")]; + tensor input_49_axes_0 = const()[name = tensor("input_49_axes_0"), val = tensor([-1])]; + tensor norm4_2_weight_to_fp16 = const()[name = tensor("norm4_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(303623552)))]; + tensor norm4_2_bias_to_fp16 = const()[name = tensor("norm4_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(303625664)))]; + tensor var_2282_to_fp16 = const()[name = tensor("op_2282_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_49_cast_fp16 = layer_norm(axes = input_49_axes_0, beta = norm4_2_bias_to_fp16, epsilon = var_2282_to_fp16, gamma = norm4_2_weight_to_fp16, x = input_47_cast_fp16)[name = tensor("input_49_cast_fp16")]; + tensor linear_19_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = linear4_1_weight_cast_fp16, x = input_49_cast_fp16)[name = tensor("linear_19_cast_fp16")]; + tensor input_51_mode_0 = const()[name = tensor("input_51_mode_0"), val = tensor("EXACT")]; + tensor input_51_cast_fp16 = gelu(mode = input_51_mode_0, x = linear_19_cast_fp16)[name = tensor("input_51_cast_fp16")]; + tensor linear_20_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = linear4_2_weight_cast_fp16, x = input_51_cast_fp16)[name = tensor("linear_20_cast_fp16")]; + tensor input_53_cast_fp16 = add(x = input_47_cast_fp16, y = linear_20_cast_fp16)[name = tensor("input_53_cast_fp16")]; + tensor x_11_axes_0 = const()[name = tensor("x_11_axes_0"), val = tensor([-1])]; + tensor norm5_1_weight_to_fp16 = const()[name = tensor("norm5_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(303627776)))]; + tensor norm5_1_bias_to_fp16 = const()[name = tensor("norm5_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(303629888)))]; + tensor var_2299_to_fp16 = const()[name = tensor("op_2299_to_fp16"), val = tensor(0x1.5p-17)]; + tensor x_11_cast_fp16 = layer_norm(axes = x_11_axes_0, beta = norm5_1_bias_to_fp16, epsilon = var_2299_to_fp16, gamma = norm5_1_weight_to_fp16, x = input_53_cast_fp16)[name = tensor("x_11_cast_fp16")]; + tensor linear_21_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = attn5_in_proj_weight_cast_fp16, x = x_11_cast_fp16)[name = tensor("linear_21_cast_fp16")]; + tensor var_2335 = const()[name = tensor("op_2335"), val = tensor([1, 1, 3, 16, 64])]; + tensor qkv_11_cast_fp16 = reshape(shape = var_2335, x = linear_21_cast_fp16)[name = tensor("qkv_11_cast_fp16")]; + tensor q_31_begin_0 = const()[name = tensor("q_31_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor q_31_end_0 = const()[name = tensor("q_31_end_0"), val = tensor([1, 1, 1, 16, 64])]; + tensor q_31_end_mask_0 = const()[name = tensor("q_31_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor q_31_squeeze_mask_0 = const()[name = tensor("q_31_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor q_31_cast_fp16 = slice_by_index(begin = q_31_begin_0, end = q_31_end_0, end_mask = q_31_end_mask_0, squeeze_mask = q_31_squeeze_mask_0, x = qkv_11_cast_fp16)[name = tensor("q_31_cast_fp16")]; + tensor k_21_begin_0 = const()[name = tensor("k_21_begin_0"), val = tensor([0, 0, 1, 0, 0])]; + tensor k_21_end_0 = const()[name = tensor("k_21_end_0"), val = tensor([1, 1, 2, 16, 64])]; + tensor k_21_end_mask_0 = const()[name = tensor("k_21_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor k_21_squeeze_mask_0 = const()[name = tensor("k_21_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor k_21_cast_fp16 = slice_by_index(begin = k_21_begin_0, end = k_21_end_0, end_mask = k_21_end_mask_0, squeeze_mask = k_21_squeeze_mask_0, x = qkv_11_cast_fp16)[name = tensor("k_21_cast_fp16")]; + tensor v_11_begin_0 = const()[name = tensor("v_11_begin_0"), val = tensor([0, 0, 2, 0, 0])]; + tensor v_11_end_0 = const()[name = tensor("v_11_end_0"), val = tensor([1, 1, 3, 16, 64])]; + tensor v_11_end_mask_0 = const()[name = tensor("v_11_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor v_11_squeeze_mask_0 = const()[name = tensor("v_11_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor v_11_cast_fp16 = slice_by_index(begin = v_11_begin_0, end = v_11_end_0, end_mask = v_11_end_mask_0, squeeze_mask = v_11_squeeze_mask_0, x = qkv_11_cast_fp16)[name = tensor("v_11_cast_fp16")]; + tensor var_2439 = const()[name = tensor("op_2439"), val = tensor([1, 1, 1, 1])]; + tensor position5_to_fp16_dtype_0 = const()[name = tensor("position5_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor position5_to_fp16 = cast(dtype = position5_to_fp16_dtype_0, x = position5)[name = tensor("cast_528")]; + tensor ts_35_cast_fp16 = reshape(shape = var_2439, x = position5_to_fp16)[name = tensor("ts_35_cast_fp16")]; + tensor var_2443 = const()[name = tensor("op_2443"), val = tensor([1, 1, 16, 32, 2])]; + tensor q_complex_11_cast_fp16 = reshape(shape = var_2443, x = q_31_cast_fp16)[name = tensor("q_complex_11_cast_fp16")]; + tensor var_2447 = const()[name = tensor("op_2447"), val = tensor([1, 1, 16, 32, 2])]; + tensor k_complex_11_cast_fp16 = reshape(shape = var_2447, x = k_21_cast_fp16)[name = tensor("k_complex_11_cast_fp16")]; + tensor var_2451_begin_0 = const()[name = tensor("op_2451_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_2451_end_0 = const()[name = tensor("op_2451_end_0"), val = tensor([1, 1, 16, 32, 1])]; + tensor var_2451_end_mask_0 = const()[name = tensor("op_2451_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_2451_squeeze_mask_0 = const()[name = tensor("op_2451_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_2451_cast_fp16 = slice_by_index(begin = var_2451_begin_0, end = var_2451_end_0, end_mask = var_2451_end_mask_0, squeeze_mask = var_2451_squeeze_mask_0, x = q_complex_11_cast_fp16)[name = tensor("op_2451_cast_fp16")]; + tensor var_2459_begin_0 = const()[name = tensor("op_2459_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_2459_end_0 = const()[name = tensor("op_2459_end_0"), val = tensor([1, 1, 16, 32, 2])]; + tensor var_2459_end_mask_0 = const()[name = tensor("op_2459_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_2459_squeeze_mask_0 = const()[name = tensor("op_2459_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_2459_cast_fp16 = slice_by_index(begin = var_2459_begin_0, end = var_2459_end_0, end_mask = var_2459_end_mask_0, squeeze_mask = var_2459_squeeze_mask_0, x = q_complex_11_cast_fp16)[name = tensor("op_2459_cast_fp16")]; + tensor var_2467_begin_0 = const()[name = tensor("op_2467_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_2467_end_0 = const()[name = tensor("op_2467_end_0"), val = tensor([1, 1, 16, 32, 1])]; + tensor var_2467_end_mask_0 = const()[name = tensor("op_2467_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_2467_squeeze_mask_0 = const()[name = tensor("op_2467_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_2467_cast_fp16 = slice_by_index(begin = var_2467_begin_0, end = var_2467_end_0, end_mask = var_2467_end_mask_0, squeeze_mask = var_2467_squeeze_mask_0, x = k_complex_11_cast_fp16)[name = tensor("op_2467_cast_fp16")]; + tensor var_2475_begin_0 = const()[name = tensor("op_2475_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_2475_end_0 = const()[name = tensor("op_2475_end_0"), val = tensor([1, 1, 16, 32, 2])]; + tensor var_2475_end_mask_0 = const()[name = tensor("op_2475_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_2475_squeeze_mask_0 = const()[name = tensor("op_2475_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_2475_cast_fp16 = slice_by_index(begin = var_2475_begin_0, end = var_2475_end_0, end_mask = var_2475_end_mask_0, squeeze_mask = var_2475_squeeze_mask_0, x = k_complex_11_cast_fp16)[name = tensor("op_2475_cast_fp16")]; + tensor freqs_11_to_fp16 = const()[name = tensor("freqs_11_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(303632000)))]; + tensor var_2481_cast_fp16 = mul(x = freqs_11_to_fp16, y = ts_35_cast_fp16)[name = tensor("op_2481_cast_fp16")]; + tensor rotr_11_cast_fp16 = cos(x = var_2481_cast_fp16)[name = tensor("rotr_11_cast_fp16")]; + tensor roti_11_cast_fp16 = sin(x = var_2481_cast_fp16)[name = tensor("roti_11_cast_fp16")]; + tensor var_2485_cast_fp16 = mul(x = var_2451_cast_fp16, y = rotr_11_cast_fp16)[name = tensor("op_2485_cast_fp16")]; + tensor var_2486_cast_fp16 = mul(x = var_2459_cast_fp16, y = roti_11_cast_fp16)[name = tensor("op_2486_cast_fp16")]; + tensor qor_21_cast_fp16 = sub(x = var_2485_cast_fp16, y = var_2486_cast_fp16)[name = tensor("qor_21_cast_fp16")]; + tensor var_2489_cast_fp16 = mul(x = var_2451_cast_fp16, y = roti_11_cast_fp16)[name = tensor("op_2489_cast_fp16")]; + tensor var_2490_cast_fp16 = mul(x = var_2459_cast_fp16, y = rotr_11_cast_fp16)[name = tensor("op_2490_cast_fp16")]; + tensor qoi_21_cast_fp16 = add(x = var_2489_cast_fp16, y = var_2490_cast_fp16)[name = tensor("qoi_21_cast_fp16")]; + tensor var_2493_cast_fp16 = mul(x = var_2467_cast_fp16, y = rotr_11_cast_fp16)[name = tensor("op_2493_cast_fp16")]; + tensor var_2494_cast_fp16 = mul(x = var_2475_cast_fp16, y = roti_11_cast_fp16)[name = tensor("op_2494_cast_fp16")]; + tensor kor_21_cast_fp16 = sub(x = var_2493_cast_fp16, y = var_2494_cast_fp16)[name = tensor("kor_21_cast_fp16")]; + tensor var_2497_cast_fp16 = mul(x = var_2467_cast_fp16, y = roti_11_cast_fp16)[name = tensor("op_2497_cast_fp16")]; + tensor var_2498_cast_fp16 = mul(x = var_2475_cast_fp16, y = rotr_11_cast_fp16)[name = tensor("op_2498_cast_fp16")]; + tensor koi_21_cast_fp16 = add(x = var_2497_cast_fp16, y = var_2498_cast_fp16)[name = tensor("koi_21_cast_fp16")]; + tensor qo_11_axis_0 = const()[name = tensor("qo_11_axis_0"), val = tensor(-1)]; + tensor qo_11_cast_fp16 = stack(axis = qo_11_axis_0, values = (qor_21_cast_fp16, qoi_21_cast_fp16))[name = tensor("qo_11_cast_fp16")]; + tensor ko_11_axis_0 = const()[name = tensor("ko_11_axis_0"), val = tensor(-1)]; + tensor ko_11_cast_fp16 = stack(axis = ko_11_axis_0, values = (kor_21_cast_fp16, koi_21_cast_fp16))[name = tensor("ko_11_cast_fp16")]; + tensor var_2527 = const()[name = tensor("op_2527"), val = tensor([1, 1, 16, 64])]; + tensor q_33_cast_fp16 = reshape(shape = var_2527, x = qo_11_cast_fp16)[name = tensor("q_33_cast_fp16")]; + tensor var_2529 = const()[name = tensor("op_2529"), val = tensor([1, 1, 16, 64])]; + tensor k_23_cast_fp16 = reshape(shape = var_2529, x = ko_11_cast_fp16)[name = tensor("k_23_cast_fp16")]; + tensor _inversed_2551_y_0_to_fp16 = const()[name = tensor("_inversed_2551_y_0_to_fp16"), val = tensor(0x1p-9)]; + tensor _inversed_2551_cast_fp16 = mul(x = ts_35_cast_fp16, y = _inversed_2551_y_0_to_fp16)[name = tensor("_inversed_2551_cast_fp16")]; + tensor var_2552_cast_fp16 = floor(x = _inversed_2551_cast_fp16)[name = tensor("op_2552_cast_fp16")]; + tensor var_2553_to_fp16 = const()[name = tensor("op_2553_to_fp16"), val = tensor(0x1p+9)]; + tensor var_2554_cast_fp16 = mul(x = var_2552_cast_fp16, y = var_2553_to_fp16)[name = tensor("op_2554_cast_fp16")]; + tensor write_indices_float_23_cast_fp16 = sub(x = ts_35_cast_fp16, y = var_2554_cast_fp16)[name = tensor("write_indices_float_23_cast_fp16")]; + tensor var_2561_dtype_0 = const()[name = tensor("op_2561_dtype_0"), val = tensor("int32")]; + tensor write_indices_11_reps_0 = const()[name = tensor("write_indices_11_reps_0"), val = tensor([1, 1, 16, 64])]; + tensor write_indices_float_23_cast_fp16_to_int32 = cast(dtype = var_2561_dtype_0, x = write_indices_float_23_cast_fp16)[name = tensor("cast_527")]; + tensor write_indices_11 = tile(reps = write_indices_11_reps_0, x = write_indices_float_23_cast_fp16_to_int32)[name = tensor("write_indices_11")]; + tensor var_2569_begin_0 = const()[name = tensor("op_2569_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_2569_end_0 = const()[name = tensor("op_2569_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor var_2569_end_mask_0 = const()[name = tensor("op_2569_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_2569_squeeze_mask_0 = const()[name = tensor("op_2569_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor cache5_to_fp16_dtype_0 = const()[name = tensor("cache5_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor cache5_to_fp16 = cast(dtype = cache5_to_fp16_dtype_0, x = cache5)[name = tensor("cast_526")]; + tensor var_2569_cast_fp16 = slice_by_index(begin = var_2569_begin_0, end = var_2569_end_0, end_mask = var_2569_end_mask_0, squeeze_mask = var_2569_squeeze_mask_0, x = cache5_to_fp16)[name = tensor("op_2569_cast_fp16")]; + tensor var_2571_axis_0 = const()[name = tensor("op_2571_axis_0"), val = tensor(1)]; + tensor var_2571_mode_0 = const()[name = tensor("op_2571_mode_0"), val = tensor("update")]; + tensor var_2571_validate_indices_0 = const()[name = tensor("op_2571_validate_indices_0"), val = tensor(false)]; + tensor var_2571_cast_fp16 = scatter_along_axis(axis = var_2571_axis_0, data = var_2569_cast_fp16, indices = write_indices_11, mode = var_2571_mode_0, updates = k_23_cast_fp16, validate_indices = var_2571_validate_indices_0)[name = tensor("op_2571_cast_fp16")]; + tensor concat_37 = const()[name = tensor("concat_37"), val = tensor([0, 0, 0, 0, 0])]; + tensor concat_38 = const()[name = tensor("concat_38"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_11_internal_tensor_assign_1_stride_0 = const()[name = tensor("new_cache_11_internal_tensor_assign_1_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_11_internal_tensor_assign_1_begin_mask_0 = const()[name = tensor("new_cache_11_internal_tensor_assign_1_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_11_internal_tensor_assign_1_end_mask_0 = const()[name = tensor("new_cache_11_internal_tensor_assign_1_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_11_internal_tensor_assign_1_squeeze_mask_0 = const()[name = tensor("new_cache_11_internal_tensor_assign_1_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_58 = const()[name = tensor("shape_58"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_10 = const()[name = tensor("reduce_prod_10"), val = tensor(1048576)]; + tensor range_1d_10_start_0 = const()[name = tensor("range_1d_10_start_0"), val = tensor(0)]; + tensor range_1d_10_step_0 = const()[name = tensor("range_1d_10_step_0"), val = tensor(1)]; + tensor range_1d_10 = range_1d(end = reduce_prod_10, start = range_1d_10_start_0, step = range_1d_10_step_0)[name = tensor("range_1d_10")]; + tensor reshape_50 = reshape(shape = shape_58, x = range_1d_10)[name = tensor("reshape_50")]; + tensor slice_by_index_10 = slice_by_index(begin = concat_37, begin_mask = new_cache_11_internal_tensor_assign_1_begin_mask_0, end = concat_38, end_mask = new_cache_11_internal_tensor_assign_1_end_mask_0, squeeze_mask = new_cache_11_internal_tensor_assign_1_squeeze_mask_0, stride = new_cache_11_internal_tensor_assign_1_stride_0, x = reshape_50)[name = tensor("slice_by_index_10")]; + tensor reshape_51_shape_0 = const()[name = tensor("reshape_51_shape_0"), val = tensor([-1])]; + tensor reshape_51 = reshape(shape = reshape_51_shape_0, x = slice_by_index_10)[name = tensor("reshape_51")]; + tensor reshape_52_shape_0 = const()[name = tensor("reshape_52_shape_0"), val = tensor([-1])]; + tensor reshape_52_cast_fp16 = reshape(shape = reshape_52_shape_0, x = var_2571_cast_fp16)[name = tensor("reshape_52_cast_fp16")]; + tensor reshape_53_shape_0 = const()[name = tensor("reshape_53_shape_0"), val = tensor([-1])]; + tensor reshape_53_cast_fp16 = reshape(shape = reshape_53_shape_0, x = cache5_to_fp16)[name = tensor("reshape_53_cast_fp16")]; + tensor scatter_10_mode_0 = const()[name = tensor("scatter_10_mode_0"), val = tensor("update")]; + tensor scatter_10_axis_0 = const()[name = tensor("scatter_10_axis_0"), val = tensor(0)]; + tensor scatter_10_validate_indices_0 = const()[name = tensor("scatter_10_validate_indices_0"), val = tensor(false)]; + tensor scatter_10_cast_fp16 = scatter(axis = scatter_10_axis_0, data = reshape_53_cast_fp16, indices = reshape_51, mode = scatter_10_mode_0, updates = reshape_52_cast_fp16, validate_indices = scatter_10_validate_indices_0)[name = tensor("scatter_10_cast_fp16")]; + tensor reshape_54_cast_fp16 = reshape(shape = shape_58, x = scatter_10_cast_fp16)[name = tensor("reshape_54_cast_fp16")]; + tensor var_2579_begin_0 = const()[name = tensor("op_2579_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor var_2579_end_0 = const()[name = tensor("op_2579_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor var_2579_end_mask_0 = const()[name = tensor("op_2579_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_2579_squeeze_mask_0 = const()[name = tensor("op_2579_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor var_2579_cast_fp16 = slice_by_index(begin = var_2579_begin_0, end = var_2579_end_0, end_mask = var_2579_end_mask_0, squeeze_mask = var_2579_squeeze_mask_0, x = reshape_54_cast_fp16)[name = tensor("op_2579_cast_fp16")]; + tensor var_2581_axis_0 = const()[name = tensor("op_2581_axis_0"), val = tensor(1)]; + tensor var_2581_mode_0 = const()[name = tensor("op_2581_mode_0"), val = tensor("update")]; + tensor var_2581_validate_indices_0 = const()[name = tensor("op_2581_validate_indices_0"), val = tensor(false)]; + tensor var_2581_cast_fp16 = scatter_along_axis(axis = var_2581_axis_0, data = var_2579_cast_fp16, indices = write_indices_11, mode = var_2581_mode_0, updates = v_11_cast_fp16, validate_indices = var_2581_validate_indices_0)[name = tensor("op_2581_cast_fp16")]; + tensor concat_39 = const()[name = tensor("concat_39"), val = tensor([1, 0, 0, 0, 0])]; + tensor concat_40 = const()[name = tensor("concat_40"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_11_internal_tensor_assign_2_stride_0 = const()[name = tensor("new_cache_11_internal_tensor_assign_2_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_11_internal_tensor_assign_2_begin_mask_0 = const()[name = tensor("new_cache_11_internal_tensor_assign_2_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_11_internal_tensor_assign_2_end_mask_0 = const()[name = tensor("new_cache_11_internal_tensor_assign_2_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_11_internal_tensor_assign_2_squeeze_mask_0 = const()[name = tensor("new_cache_11_internal_tensor_assign_2_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_59 = const()[name = tensor("shape_59"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_11 = const()[name = tensor("reduce_prod_11"), val = tensor(1048576)]; + tensor range_1d_11_start_0 = const()[name = tensor("range_1d_11_start_0"), val = tensor(0)]; + tensor range_1d_11_step_0 = const()[name = tensor("range_1d_11_step_0"), val = tensor(1)]; + tensor range_1d_11 = range_1d(end = reduce_prod_11, start = range_1d_11_start_0, step = range_1d_11_step_0)[name = tensor("range_1d_11")]; + tensor reshape_55 = reshape(shape = shape_59, x = range_1d_11)[name = tensor("reshape_55")]; + tensor slice_by_index_11 = slice_by_index(begin = concat_39, begin_mask = new_cache_11_internal_tensor_assign_2_begin_mask_0, end = concat_40, end_mask = new_cache_11_internal_tensor_assign_2_end_mask_0, squeeze_mask = new_cache_11_internal_tensor_assign_2_squeeze_mask_0, stride = new_cache_11_internal_tensor_assign_2_stride_0, x = reshape_55)[name = tensor("slice_by_index_11")]; + tensor reshape_56_shape_0 = const()[name = tensor("reshape_56_shape_0"), val = tensor([-1])]; + tensor reshape_56 = reshape(shape = reshape_56_shape_0, x = slice_by_index_11)[name = tensor("reshape_56")]; + tensor reshape_57_shape_0 = const()[name = tensor("reshape_57_shape_0"), val = tensor([-1])]; + tensor reshape_57_cast_fp16 = reshape(shape = reshape_57_shape_0, x = var_2581_cast_fp16)[name = tensor("reshape_57_cast_fp16")]; + tensor reshape_58_shape_0 = const()[name = tensor("reshape_58_shape_0"), val = tensor([-1])]; + tensor reshape_58_cast_fp16 = reshape(shape = reshape_58_shape_0, x = reshape_54_cast_fp16)[name = tensor("reshape_58_cast_fp16")]; + tensor scatter_11_mode_0 = const()[name = tensor("scatter_11_mode_0"), val = tensor("update")]; + tensor scatter_11_axis_0 = const()[name = tensor("scatter_11_axis_0"), val = tensor(0)]; + tensor scatter_11_validate_indices_0 = const()[name = tensor("scatter_11_validate_indices_0"), val = tensor(false)]; + tensor scatter_11_cast_fp16 = scatter(axis = scatter_11_axis_0, data = reshape_58_cast_fp16, indices = reshape_56, mode = scatter_11_mode_0, updates = reshape_57_cast_fp16, validate_indices = scatter_11_validate_indices_0)[name = tensor("scatter_11_cast_fp16")]; + tensor reshape_59_cast_fp16 = reshape(shape = shape_59, x = scatter_11_cast_fp16)[name = tensor("reshape_59_cast_fp16")]; + tensor reshape_59_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("reshape_59_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor keys_31_begin_0 = const()[name = tensor("keys_31_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor keys_31_end_0 = const()[name = tensor("keys_31_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor keys_31_end_mask_0 = const()[name = tensor("keys_31_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor keys_31_squeeze_mask_0 = const()[name = tensor("keys_31_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor keys_31_cast_fp16 = slice_by_index(begin = keys_31_begin_0, end = keys_31_end_0, end_mask = keys_31_end_mask_0, squeeze_mask = keys_31_squeeze_mask_0, x = reshape_59_cast_fp16)[name = tensor("keys_31_cast_fp16")]; + tensor values_31_begin_0 = const()[name = tensor("values_31_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor values_31_end_0 = const()[name = tensor("values_31_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor values_31_end_mask_0 = const()[name = tensor("values_31_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor values_31_squeeze_mask_0 = const()[name = tensor("values_31_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor values_31_cast_fp16 = slice_by_index(begin = values_31_begin_0, end = values_31_end_0, end_mask = values_31_end_mask_0, squeeze_mask = values_31_squeeze_mask_0, x = reshape_59_cast_fp16)[name = tensor("values_31_cast_fp16")]; + tensor var_2593_cast_fp16 = not_equal(x = keys_31_cast_fp16, y = keys_31_cast_fp16)[name = tensor("op_2593_cast_fp16")]; + tensor keys_33_cast_fp16 = select(a = var_504_to_fp16, b = keys_31_cast_fp16, cond = var_2593_cast_fp16)[name = tensor("keys_33_cast_fp16")]; + tensor var_2601_cast_fp16 = not_equal(x = values_31_cast_fp16, y = values_31_cast_fp16)[name = tensor("op_2601_cast_fp16")]; + tensor values_33_cast_fp16 = select(a = var_504_to_fp16, b = values_31_cast_fp16, cond = var_2601_cast_fp16)[name = tensor("values_33_cast_fp16")]; + tensor var_2625 = const()[name = tensor("op_2625"), val = tensor([0, 2, 1, 3])]; + tensor var_2638 = const()[name = tensor("op_2638"), val = tensor([1, 1, 1])]; + tensor var_2639_cast_fp16 = reshape(shape = var_2638, x = position5_to_fp16)[name = tensor("op_2639_cast_fp16")]; + tensor var_2656_to_fp16 = const()[name = tensor("op_2656_to_fp16"), val = tensor(0x1p+0)]; + tensor valid_len_11_cast_fp16 = add(x = var_2639_cast_fp16, y = var_2656_to_fp16)[name = tensor("valid_len_11_cast_fp16")]; + tensor valid_mask_11_cast_fp16 = less(x = k_positions_1_promoted_to_fp16, y = valid_len_11_cast_fp16)[name = tensor("valid_mask_11_cast_fp16")]; + tensor causal_mask_11_cast_fp16 = less_equal(x = k_positions_1_promoted_to_fp16, y = var_2639_cast_fp16)[name = tensor("causal_mask_11_cast_fp16")]; + tensor attn_mask_21 = logical_and(x = valid_mask_11_cast_fp16, y = causal_mask_11_cast_fp16)[name = tensor("attn_mask_21")]; + tensor attn_mask_23_axes_0 = const()[name = tensor("attn_mask_23_axes_0"), val = tensor([1])]; + tensor attn_mask_23 = expand_dims(axes = attn_mask_23_axes_0, x = attn_mask_21)[name = tensor("attn_mask_23")]; + tensor var_2674_transpose_x_0 = const()[name = tensor("op_2674_transpose_x_0"), val = tensor(false)]; + tensor var_2674_transpose_y_0 = const()[name = tensor("op_2674_transpose_y_0"), val = tensor(false)]; + tensor transpose_82_perm_0 = const()[name = tensor("transpose_82_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_83_perm_0 = const()[name = tensor("transpose_83_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_83 = transpose(perm = transpose_83_perm_0, x = keys_33_cast_fp16)[name = tensor("transpose_193")]; + tensor transpose_82 = transpose(perm = transpose_82_perm_0, x = q_33_cast_fp16)[name = tensor("transpose_194")]; + tensor var_2674_cast_fp16 = matmul(transpose_x = var_2674_transpose_x_0, transpose_y = var_2674_transpose_y_0, x = transpose_82, y = transpose_83)[name = tensor("op_2674_cast_fp16")]; + tensor var_2668_to_fp16 = const()[name = tensor("op_2668_to_fp16"), val = tensor([0x1p-3])]; + tensor attn_weights_31_cast_fp16 = mul(x = var_2674_cast_fp16, y = var_2668_to_fp16)[name = tensor("attn_weights_31_cast_fp16")]; + tensor var_2676 = logical_not(x = attn_mask_23)[name = tensor("op_2676")]; + tensor var_2677_to_fp16 = const()[name = tensor("op_2677_to_fp16"), val = tensor(-inf)]; + tensor attn_weights_33_cast_fp16 = select(a = var_2677_to_fp16, b = attn_weights_31_cast_fp16, cond = var_2676)[name = tensor("attn_weights_33_cast_fp16")]; + tensor var_2679 = const()[name = tensor("op_2679"), val = tensor(-1)]; + tensor attn_weights_35_cast_fp16 = softmax(axis = var_2679, x = attn_weights_33_cast_fp16)[name = tensor("attn_weights_35_cast_fp16")]; + tensor attn_output_11_transpose_x_0 = const()[name = tensor("attn_output_11_transpose_x_0"), val = tensor(false)]; + tensor attn_output_11_transpose_y_0 = const()[name = tensor("attn_output_11_transpose_y_0"), val = tensor(false)]; + tensor values_35_cast_fp16 = transpose(perm = var_2625, x = values_33_cast_fp16)[name = tensor("transpose_195")]; + tensor attn_output_11_cast_fp16 = matmul(transpose_x = attn_output_11_transpose_x_0, transpose_y = attn_output_11_transpose_y_0, x = attn_weights_35_cast_fp16, y = values_35_cast_fp16)[name = tensor("attn_output_11_cast_fp16")]; + tensor var_2687 = const()[name = tensor("op_2687"), val = tensor([0, 2, 1, 3])]; + tensor var_2690 = const()[name = tensor("op_2690"), val = tensor([1, 1, 1024])]; + tensor var_2688_cast_fp16 = transpose(perm = var_2687, x = attn_output_11_cast_fp16)[name = tensor("transpose_192")]; + tensor input_55_cast_fp16 = reshape(shape = var_2690, x = var_2688_cast_fp16)[name = tensor("input_55_cast_fp16")]; + tensor linear_22_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = attn5_out_proj_weight_cast_fp16, x = input_55_cast_fp16)[name = tensor("linear_22_cast_fp16")]; + tensor var_2696_to_fp16 = const()[name = tensor("op_2696_to_fp16"), val = tensor(0x1p+0)]; + tensor var_2697_cast_fp16 = add(x = position5_to_fp16, y = var_2696_to_fp16)[name = tensor("op_2697_cast_fp16")]; + tensor var_2697_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_2697_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor input_57_cast_fp16 = add(x = input_53_cast_fp16, y = linear_22_cast_fp16)[name = tensor("input_57_cast_fp16")]; + tensor input_59_axes_0 = const()[name = tensor("input_59_axes_0"), val = tensor([-1])]; + tensor norm5_2_weight_to_fp16 = const()[name = tensor("norm5_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(303632128)))]; + tensor norm5_2_bias_to_fp16 = const()[name = tensor("norm5_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(303634240)))]; + tensor var_2701_to_fp16 = const()[name = tensor("op_2701_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_59_cast_fp16 = layer_norm(axes = input_59_axes_0, beta = norm5_2_bias_to_fp16, epsilon = var_2701_to_fp16, gamma = norm5_2_weight_to_fp16, x = input_57_cast_fp16)[name = tensor("input_59_cast_fp16")]; + tensor linear_23_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = linear5_1_weight_cast_fp16, x = input_59_cast_fp16)[name = tensor("linear_23_cast_fp16")]; + tensor input_61_mode_0 = const()[name = tensor("input_61_mode_0"), val = tensor("EXACT")]; + tensor input_61_cast_fp16 = gelu(mode = input_61_mode_0, x = linear_23_cast_fp16)[name = tensor("input_61_cast_fp16")]; + tensor linear_24_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = linear5_2_weight_cast_fp16, x = input_61_cast_fp16)[name = tensor("linear_24_cast_fp16")]; + tensor input_63_cast_fp16 = add(x = input_57_cast_fp16, y = linear_24_cast_fp16)[name = tensor("input_63_cast_fp16")]; + tensor x_13_axes_0 = const()[name = tensor("x_13_axes_0"), val = tensor([-1])]; + tensor norm6_1_weight_to_fp16 = const()[name = tensor("norm6_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(303636352)))]; + tensor norm6_1_bias_to_fp16 = const()[name = tensor("norm6_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(303638464)))]; + tensor var_2718_to_fp16 = const()[name = tensor("op_2718_to_fp16"), val = tensor(0x1.5p-17)]; + tensor x_13_cast_fp16 = layer_norm(axes = x_13_axes_0, beta = norm6_1_bias_to_fp16, epsilon = var_2718_to_fp16, gamma = norm6_1_weight_to_fp16, x = input_63_cast_fp16)[name = tensor("x_13_cast_fp16")]; + tensor linear_25_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = attn6_in_proj_weight_cast_fp16, x = x_13_cast_fp16)[name = tensor("linear_25_cast_fp16")]; + tensor var_2754 = const()[name = tensor("op_2754"), val = tensor([1, 1, 3, 16, 64])]; + tensor qkv_13_cast_fp16 = reshape(shape = var_2754, x = linear_25_cast_fp16)[name = tensor("qkv_13_cast_fp16")]; + tensor q_37_begin_0 = const()[name = tensor("q_37_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor q_37_end_0 = const()[name = tensor("q_37_end_0"), val = tensor([1, 1, 1, 16, 64])]; + tensor q_37_end_mask_0 = const()[name = tensor("q_37_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor q_37_squeeze_mask_0 = const()[name = tensor("q_37_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor q_37_cast_fp16 = slice_by_index(begin = q_37_begin_0, end = q_37_end_0, end_mask = q_37_end_mask_0, squeeze_mask = q_37_squeeze_mask_0, x = qkv_13_cast_fp16)[name = tensor("q_37_cast_fp16")]; + tensor k_25_begin_0 = const()[name = tensor("k_25_begin_0"), val = tensor([0, 0, 1, 0, 0])]; + tensor k_25_end_0 = const()[name = tensor("k_25_end_0"), val = tensor([1, 1, 2, 16, 64])]; + tensor k_25_end_mask_0 = const()[name = tensor("k_25_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor k_25_squeeze_mask_0 = const()[name = tensor("k_25_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor k_25_cast_fp16 = slice_by_index(begin = k_25_begin_0, end = k_25_end_0, end_mask = k_25_end_mask_0, squeeze_mask = k_25_squeeze_mask_0, x = qkv_13_cast_fp16)[name = tensor("k_25_cast_fp16")]; + tensor v_13_begin_0 = const()[name = tensor("v_13_begin_0"), val = tensor([0, 0, 2, 0, 0])]; + tensor v_13_end_0 = const()[name = tensor("v_13_end_0"), val = tensor([1, 1, 3, 16, 64])]; + tensor v_13_end_mask_0 = const()[name = tensor("v_13_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor v_13_squeeze_mask_0 = const()[name = tensor("v_13_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor v_13_cast_fp16 = slice_by_index(begin = v_13_begin_0, end = v_13_end_0, end_mask = v_13_end_mask_0, squeeze_mask = v_13_squeeze_mask_0, x = qkv_13_cast_fp16)[name = tensor("v_13_cast_fp16")]; + tensor var_2858 = const()[name = tensor("op_2858"), val = tensor([1, 1, 1, 1])]; + tensor position6_to_fp16_dtype_0 = const()[name = tensor("position6_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor position6_to_fp16 = cast(dtype = position6_to_fp16_dtype_0, x = position6)[name = tensor("cast_523")]; + tensor ts_41_cast_fp16 = reshape(shape = var_2858, x = position6_to_fp16)[name = tensor("ts_41_cast_fp16")]; + tensor var_2862 = const()[name = tensor("op_2862"), val = tensor([1, 1, 16, 32, 2])]; + tensor q_complex_13_cast_fp16 = reshape(shape = var_2862, x = q_37_cast_fp16)[name = tensor("q_complex_13_cast_fp16")]; + tensor var_2866 = const()[name = tensor("op_2866"), val = tensor([1, 1, 16, 32, 2])]; + tensor k_complex_13_cast_fp16 = reshape(shape = var_2866, x = k_25_cast_fp16)[name = tensor("k_complex_13_cast_fp16")]; + tensor var_2870_begin_0 = const()[name = tensor("op_2870_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_2870_end_0 = const()[name = tensor("op_2870_end_0"), val = tensor([1, 1, 16, 32, 1])]; + tensor var_2870_end_mask_0 = const()[name = tensor("op_2870_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_2870_squeeze_mask_0 = const()[name = tensor("op_2870_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_2870_cast_fp16 = slice_by_index(begin = var_2870_begin_0, end = var_2870_end_0, end_mask = var_2870_end_mask_0, squeeze_mask = var_2870_squeeze_mask_0, x = q_complex_13_cast_fp16)[name = tensor("op_2870_cast_fp16")]; + tensor var_2878_begin_0 = const()[name = tensor("op_2878_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_2878_end_0 = const()[name = tensor("op_2878_end_0"), val = tensor([1, 1, 16, 32, 2])]; + tensor var_2878_end_mask_0 = const()[name = tensor("op_2878_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_2878_squeeze_mask_0 = const()[name = tensor("op_2878_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_2878_cast_fp16 = slice_by_index(begin = var_2878_begin_0, end = var_2878_end_0, end_mask = var_2878_end_mask_0, squeeze_mask = var_2878_squeeze_mask_0, x = q_complex_13_cast_fp16)[name = tensor("op_2878_cast_fp16")]; + tensor var_2886_begin_0 = const()[name = tensor("op_2886_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_2886_end_0 = const()[name = tensor("op_2886_end_0"), val = tensor([1, 1, 16, 32, 1])]; + tensor var_2886_end_mask_0 = const()[name = tensor("op_2886_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_2886_squeeze_mask_0 = const()[name = tensor("op_2886_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_2886_cast_fp16 = slice_by_index(begin = var_2886_begin_0, end = var_2886_end_0, end_mask = var_2886_end_mask_0, squeeze_mask = var_2886_squeeze_mask_0, x = k_complex_13_cast_fp16)[name = tensor("op_2886_cast_fp16")]; + tensor var_2894_begin_0 = const()[name = tensor("op_2894_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_2894_end_0 = const()[name = tensor("op_2894_end_0"), val = tensor([1, 1, 16, 32, 2])]; + tensor var_2894_end_mask_0 = const()[name = tensor("op_2894_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_2894_squeeze_mask_0 = const()[name = tensor("op_2894_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_2894_cast_fp16 = slice_by_index(begin = var_2894_begin_0, end = var_2894_end_0, end_mask = var_2894_end_mask_0, squeeze_mask = var_2894_squeeze_mask_0, x = k_complex_13_cast_fp16)[name = tensor("op_2894_cast_fp16")]; + tensor freqs_13_to_fp16 = const()[name = tensor("freqs_13_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(303640576)))]; + tensor var_2900_cast_fp16 = mul(x = freqs_13_to_fp16, y = ts_41_cast_fp16)[name = tensor("op_2900_cast_fp16")]; + tensor rotr_13_cast_fp16 = cos(x = var_2900_cast_fp16)[name = tensor("rotr_13_cast_fp16")]; + tensor roti_13_cast_fp16 = sin(x = var_2900_cast_fp16)[name = tensor("roti_13_cast_fp16")]; + tensor var_2904_cast_fp16 = mul(x = var_2870_cast_fp16, y = rotr_13_cast_fp16)[name = tensor("op_2904_cast_fp16")]; + tensor var_2905_cast_fp16 = mul(x = var_2878_cast_fp16, y = roti_13_cast_fp16)[name = tensor("op_2905_cast_fp16")]; + tensor qor_25_cast_fp16 = sub(x = var_2904_cast_fp16, y = var_2905_cast_fp16)[name = tensor("qor_25_cast_fp16")]; + tensor var_2908_cast_fp16 = mul(x = var_2870_cast_fp16, y = roti_13_cast_fp16)[name = tensor("op_2908_cast_fp16")]; + tensor var_2909_cast_fp16 = mul(x = var_2878_cast_fp16, y = rotr_13_cast_fp16)[name = tensor("op_2909_cast_fp16")]; + tensor qoi_25_cast_fp16 = add(x = var_2908_cast_fp16, y = var_2909_cast_fp16)[name = tensor("qoi_25_cast_fp16")]; + tensor var_2912_cast_fp16 = mul(x = var_2886_cast_fp16, y = rotr_13_cast_fp16)[name = tensor("op_2912_cast_fp16")]; + tensor var_2913_cast_fp16 = mul(x = var_2894_cast_fp16, y = roti_13_cast_fp16)[name = tensor("op_2913_cast_fp16")]; + tensor kor_25_cast_fp16 = sub(x = var_2912_cast_fp16, y = var_2913_cast_fp16)[name = tensor("kor_25_cast_fp16")]; + tensor var_2916_cast_fp16 = mul(x = var_2886_cast_fp16, y = roti_13_cast_fp16)[name = tensor("op_2916_cast_fp16")]; + tensor var_2917_cast_fp16 = mul(x = var_2894_cast_fp16, y = rotr_13_cast_fp16)[name = tensor("op_2917_cast_fp16")]; + tensor koi_25_cast_fp16 = add(x = var_2916_cast_fp16, y = var_2917_cast_fp16)[name = tensor("koi_25_cast_fp16")]; + tensor qo_13_axis_0 = const()[name = tensor("qo_13_axis_0"), val = tensor(-1)]; + tensor qo_13_cast_fp16 = stack(axis = qo_13_axis_0, values = (qor_25_cast_fp16, qoi_25_cast_fp16))[name = tensor("qo_13_cast_fp16")]; + tensor ko_13_axis_0 = const()[name = tensor("ko_13_axis_0"), val = tensor(-1)]; + tensor ko_13_cast_fp16 = stack(axis = ko_13_axis_0, values = (kor_25_cast_fp16, koi_25_cast_fp16))[name = tensor("ko_13_cast_fp16")]; + tensor var_2946 = const()[name = tensor("op_2946"), val = tensor([1, 1, 16, 64])]; + tensor q_39_cast_fp16 = reshape(shape = var_2946, x = qo_13_cast_fp16)[name = tensor("q_39_cast_fp16")]; + tensor var_2948 = const()[name = tensor("op_2948"), val = tensor([1, 1, 16, 64])]; + tensor k_27_cast_fp16 = reshape(shape = var_2948, x = ko_13_cast_fp16)[name = tensor("k_27_cast_fp16")]; + tensor _inversed_2970_y_0_to_fp16 = const()[name = tensor("_inversed_2970_y_0_to_fp16"), val = tensor(0x1p-9)]; + tensor _inversed_2970_cast_fp16 = mul(x = ts_41_cast_fp16, y = _inversed_2970_y_0_to_fp16)[name = tensor("_inversed_2970_cast_fp16")]; + tensor var_2971_cast_fp16 = floor(x = _inversed_2970_cast_fp16)[name = tensor("op_2971_cast_fp16")]; + tensor var_2972_to_fp16 = const()[name = tensor("op_2972_to_fp16"), val = tensor(0x1p+9)]; + tensor var_2973_cast_fp16 = mul(x = var_2971_cast_fp16, y = var_2972_to_fp16)[name = tensor("op_2973_cast_fp16")]; + tensor write_indices_float_27_cast_fp16 = sub(x = ts_41_cast_fp16, y = var_2973_cast_fp16)[name = tensor("write_indices_float_27_cast_fp16")]; + tensor var_2980_dtype_0 = const()[name = tensor("op_2980_dtype_0"), val = tensor("int32")]; + tensor write_indices_13_reps_0 = const()[name = tensor("write_indices_13_reps_0"), val = tensor([1, 1, 16, 64])]; + tensor write_indices_float_27_cast_fp16_to_int32 = cast(dtype = var_2980_dtype_0, x = write_indices_float_27_cast_fp16)[name = tensor("cast_522")]; + tensor write_indices_13 = tile(reps = write_indices_13_reps_0, x = write_indices_float_27_cast_fp16_to_int32)[name = tensor("write_indices_13")]; + tensor var_2988_begin_0 = const()[name = tensor("op_2988_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_2988_end_0 = const()[name = tensor("op_2988_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor var_2988_end_mask_0 = const()[name = tensor("op_2988_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_2988_squeeze_mask_0 = const()[name = tensor("op_2988_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor cache6_to_fp16_dtype_0 = const()[name = tensor("cache6_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor cache6_to_fp16 = cast(dtype = cache6_to_fp16_dtype_0, x = cache6)[name = tensor("cast_521")]; + tensor var_2988_cast_fp16 = slice_by_index(begin = var_2988_begin_0, end = var_2988_end_0, end_mask = var_2988_end_mask_0, squeeze_mask = var_2988_squeeze_mask_0, x = cache6_to_fp16)[name = tensor("op_2988_cast_fp16")]; + tensor var_2990_axis_0 = const()[name = tensor("op_2990_axis_0"), val = tensor(1)]; + tensor var_2990_mode_0 = const()[name = tensor("op_2990_mode_0"), val = tensor("update")]; + tensor var_2990_validate_indices_0 = const()[name = tensor("op_2990_validate_indices_0"), val = tensor(false)]; + tensor var_2990_cast_fp16 = scatter_along_axis(axis = var_2990_axis_0, data = var_2988_cast_fp16, indices = write_indices_13, mode = var_2990_mode_0, updates = k_27_cast_fp16, validate_indices = var_2990_validate_indices_0)[name = tensor("op_2990_cast_fp16")]; + tensor concat_44 = const()[name = tensor("concat_44"), val = tensor([0, 0, 0, 0, 0])]; + tensor concat_45 = const()[name = tensor("concat_45"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_13_internal_tensor_assign_1_stride_0 = const()[name = tensor("new_cache_13_internal_tensor_assign_1_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_13_internal_tensor_assign_1_begin_mask_0 = const()[name = tensor("new_cache_13_internal_tensor_assign_1_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_13_internal_tensor_assign_1_end_mask_0 = const()[name = tensor("new_cache_13_internal_tensor_assign_1_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_13_internal_tensor_assign_1_squeeze_mask_0 = const()[name = tensor("new_cache_13_internal_tensor_assign_1_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_60 = const()[name = tensor("shape_60"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_12 = const()[name = tensor("reduce_prod_12"), val = tensor(1048576)]; + tensor range_1d_12_start_0 = const()[name = tensor("range_1d_12_start_0"), val = tensor(0)]; + tensor range_1d_12_step_0 = const()[name = tensor("range_1d_12_step_0"), val = tensor(1)]; + tensor range_1d_12 = range_1d(end = reduce_prod_12, start = range_1d_12_start_0, step = range_1d_12_step_0)[name = tensor("range_1d_12")]; + tensor reshape_60 = reshape(shape = shape_60, x = range_1d_12)[name = tensor("reshape_60")]; + tensor slice_by_index_12 = slice_by_index(begin = concat_44, begin_mask = new_cache_13_internal_tensor_assign_1_begin_mask_0, end = concat_45, end_mask = new_cache_13_internal_tensor_assign_1_end_mask_0, squeeze_mask = new_cache_13_internal_tensor_assign_1_squeeze_mask_0, stride = new_cache_13_internal_tensor_assign_1_stride_0, x = reshape_60)[name = tensor("slice_by_index_12")]; + tensor reshape_61_shape_0 = const()[name = tensor("reshape_61_shape_0"), val = tensor([-1])]; + tensor reshape_61 = reshape(shape = reshape_61_shape_0, x = slice_by_index_12)[name = tensor("reshape_61")]; + tensor reshape_62_shape_0 = const()[name = tensor("reshape_62_shape_0"), val = tensor([-1])]; + tensor reshape_62_cast_fp16 = reshape(shape = reshape_62_shape_0, x = var_2990_cast_fp16)[name = tensor("reshape_62_cast_fp16")]; + tensor reshape_63_shape_0 = const()[name = tensor("reshape_63_shape_0"), val = tensor([-1])]; + tensor reshape_63_cast_fp16 = reshape(shape = reshape_63_shape_0, x = cache6_to_fp16)[name = tensor("reshape_63_cast_fp16")]; + tensor scatter_12_mode_0 = const()[name = tensor("scatter_12_mode_0"), val = tensor("update")]; + tensor scatter_12_axis_0 = const()[name = tensor("scatter_12_axis_0"), val = tensor(0)]; + tensor scatter_12_validate_indices_0 = const()[name = tensor("scatter_12_validate_indices_0"), val = tensor(false)]; + tensor scatter_12_cast_fp16 = scatter(axis = scatter_12_axis_0, data = reshape_63_cast_fp16, indices = reshape_61, mode = scatter_12_mode_0, updates = reshape_62_cast_fp16, validate_indices = scatter_12_validate_indices_0)[name = tensor("scatter_12_cast_fp16")]; + tensor reshape_64_cast_fp16 = reshape(shape = shape_60, x = scatter_12_cast_fp16)[name = tensor("reshape_64_cast_fp16")]; + tensor var_2998_begin_0 = const()[name = tensor("op_2998_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor var_2998_end_0 = const()[name = tensor("op_2998_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor var_2998_end_mask_0 = const()[name = tensor("op_2998_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_2998_squeeze_mask_0 = const()[name = tensor("op_2998_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor var_2998_cast_fp16 = slice_by_index(begin = var_2998_begin_0, end = var_2998_end_0, end_mask = var_2998_end_mask_0, squeeze_mask = var_2998_squeeze_mask_0, x = reshape_64_cast_fp16)[name = tensor("op_2998_cast_fp16")]; + tensor var_3000_axis_0 = const()[name = tensor("op_3000_axis_0"), val = tensor(1)]; + tensor var_3000_mode_0 = const()[name = tensor("op_3000_mode_0"), val = tensor("update")]; + tensor var_3000_validate_indices_0 = const()[name = tensor("op_3000_validate_indices_0"), val = tensor(false)]; + tensor var_3000_cast_fp16 = scatter_along_axis(axis = var_3000_axis_0, data = var_2998_cast_fp16, indices = write_indices_13, mode = var_3000_mode_0, updates = v_13_cast_fp16, validate_indices = var_3000_validate_indices_0)[name = tensor("op_3000_cast_fp16")]; + tensor concat_46 = const()[name = tensor("concat_46"), val = tensor([1, 0, 0, 0, 0])]; + tensor concat_47 = const()[name = tensor("concat_47"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_13_internal_tensor_assign_2_stride_0 = const()[name = tensor("new_cache_13_internal_tensor_assign_2_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_13_internal_tensor_assign_2_begin_mask_0 = const()[name = tensor("new_cache_13_internal_tensor_assign_2_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_13_internal_tensor_assign_2_end_mask_0 = const()[name = tensor("new_cache_13_internal_tensor_assign_2_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_13_internal_tensor_assign_2_squeeze_mask_0 = const()[name = tensor("new_cache_13_internal_tensor_assign_2_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_61 = const()[name = tensor("shape_61"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_13 = const()[name = tensor("reduce_prod_13"), val = tensor(1048576)]; + tensor range_1d_13_start_0 = const()[name = tensor("range_1d_13_start_0"), val = tensor(0)]; + tensor range_1d_13_step_0 = const()[name = tensor("range_1d_13_step_0"), val = tensor(1)]; + tensor range_1d_13 = range_1d(end = reduce_prod_13, start = range_1d_13_start_0, step = range_1d_13_step_0)[name = tensor("range_1d_13")]; + tensor reshape_65 = reshape(shape = shape_61, x = range_1d_13)[name = tensor("reshape_65")]; + tensor slice_by_index_13 = slice_by_index(begin = concat_46, begin_mask = new_cache_13_internal_tensor_assign_2_begin_mask_0, end = concat_47, end_mask = new_cache_13_internal_tensor_assign_2_end_mask_0, squeeze_mask = new_cache_13_internal_tensor_assign_2_squeeze_mask_0, stride = new_cache_13_internal_tensor_assign_2_stride_0, x = reshape_65)[name = tensor("slice_by_index_13")]; + tensor reshape_66_shape_0 = const()[name = tensor("reshape_66_shape_0"), val = tensor([-1])]; + tensor reshape_66 = reshape(shape = reshape_66_shape_0, x = slice_by_index_13)[name = tensor("reshape_66")]; + tensor reshape_67_shape_0 = const()[name = tensor("reshape_67_shape_0"), val = tensor([-1])]; + tensor reshape_67_cast_fp16 = reshape(shape = reshape_67_shape_0, x = var_3000_cast_fp16)[name = tensor("reshape_67_cast_fp16")]; + tensor reshape_68_shape_0 = const()[name = tensor("reshape_68_shape_0"), val = tensor([-1])]; + tensor reshape_68_cast_fp16 = reshape(shape = reshape_68_shape_0, x = reshape_64_cast_fp16)[name = tensor("reshape_68_cast_fp16")]; + tensor scatter_13_mode_0 = const()[name = tensor("scatter_13_mode_0"), val = tensor("update")]; + tensor scatter_13_axis_0 = const()[name = tensor("scatter_13_axis_0"), val = tensor(0)]; + tensor scatter_13_validate_indices_0 = const()[name = tensor("scatter_13_validate_indices_0"), val = tensor(false)]; + tensor scatter_13_cast_fp16 = scatter(axis = scatter_13_axis_0, data = reshape_68_cast_fp16, indices = reshape_66, mode = scatter_13_mode_0, updates = reshape_67_cast_fp16, validate_indices = scatter_13_validate_indices_0)[name = tensor("scatter_13_cast_fp16")]; + tensor reshape_69_cast_fp16 = reshape(shape = shape_61, x = scatter_13_cast_fp16)[name = tensor("reshape_69_cast_fp16")]; + tensor reshape_69_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("reshape_69_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor keys_37_begin_0 = const()[name = tensor("keys_37_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor keys_37_end_0 = const()[name = tensor("keys_37_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor keys_37_end_mask_0 = const()[name = tensor("keys_37_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor keys_37_squeeze_mask_0 = const()[name = tensor("keys_37_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor keys_37_cast_fp16 = slice_by_index(begin = keys_37_begin_0, end = keys_37_end_0, end_mask = keys_37_end_mask_0, squeeze_mask = keys_37_squeeze_mask_0, x = reshape_69_cast_fp16)[name = tensor("keys_37_cast_fp16")]; + tensor values_37_begin_0 = const()[name = tensor("values_37_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor values_37_end_0 = const()[name = tensor("values_37_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor values_37_end_mask_0 = const()[name = tensor("values_37_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor values_37_squeeze_mask_0 = const()[name = tensor("values_37_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor values_37_cast_fp16 = slice_by_index(begin = values_37_begin_0, end = values_37_end_0, end_mask = values_37_end_mask_0, squeeze_mask = values_37_squeeze_mask_0, x = reshape_69_cast_fp16)[name = tensor("values_37_cast_fp16")]; + tensor var_3012_cast_fp16 = not_equal(x = keys_37_cast_fp16, y = keys_37_cast_fp16)[name = tensor("op_3012_cast_fp16")]; + tensor keys_39_cast_fp16 = select(a = var_504_to_fp16, b = keys_37_cast_fp16, cond = var_3012_cast_fp16)[name = tensor("keys_39_cast_fp16")]; + tensor var_3020_cast_fp16 = not_equal(x = values_37_cast_fp16, y = values_37_cast_fp16)[name = tensor("op_3020_cast_fp16")]; + tensor values_39_cast_fp16 = select(a = var_504_to_fp16, b = values_37_cast_fp16, cond = var_3020_cast_fp16)[name = tensor("values_39_cast_fp16")]; + tensor var_3044 = const()[name = tensor("op_3044"), val = tensor([0, 2, 1, 3])]; + tensor var_3057 = const()[name = tensor("op_3057"), val = tensor([1, 1, 1])]; + tensor var_3058_cast_fp16 = reshape(shape = var_3057, x = position6_to_fp16)[name = tensor("op_3058_cast_fp16")]; + tensor var_3075_to_fp16 = const()[name = tensor("op_3075_to_fp16"), val = tensor(0x1p+0)]; + tensor valid_len_13_cast_fp16 = add(x = var_3058_cast_fp16, y = var_3075_to_fp16)[name = tensor("valid_len_13_cast_fp16")]; + tensor valid_mask_13_cast_fp16 = less(x = k_positions_1_promoted_to_fp16, y = valid_len_13_cast_fp16)[name = tensor("valid_mask_13_cast_fp16")]; + tensor causal_mask_13_cast_fp16 = less_equal(x = k_positions_1_promoted_to_fp16, y = var_3058_cast_fp16)[name = tensor("causal_mask_13_cast_fp16")]; + tensor attn_mask_25 = logical_and(x = valid_mask_13_cast_fp16, y = causal_mask_13_cast_fp16)[name = tensor("attn_mask_25")]; + tensor attn_mask_27_axes_0 = const()[name = tensor("attn_mask_27_axes_0"), val = tensor([1])]; + tensor attn_mask_27 = expand_dims(axes = attn_mask_27_axes_0, x = attn_mask_25)[name = tensor("attn_mask_27")]; + tensor var_3093_transpose_x_0 = const()[name = tensor("op_3093_transpose_x_0"), val = tensor(false)]; + tensor var_3093_transpose_y_0 = const()[name = tensor("op_3093_transpose_y_0"), val = tensor(false)]; + tensor transpose_84_perm_0 = const()[name = tensor("transpose_84_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_85_perm_0 = const()[name = tensor("transpose_85_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_85 = transpose(perm = transpose_85_perm_0, x = keys_39_cast_fp16)[name = tensor("transpose_189")]; + tensor transpose_84 = transpose(perm = transpose_84_perm_0, x = q_39_cast_fp16)[name = tensor("transpose_190")]; + tensor var_3093_cast_fp16 = matmul(transpose_x = var_3093_transpose_x_0, transpose_y = var_3093_transpose_y_0, x = transpose_84, y = transpose_85)[name = tensor("op_3093_cast_fp16")]; + tensor var_3087_to_fp16 = const()[name = tensor("op_3087_to_fp16"), val = tensor([0x1p-3])]; + tensor attn_weights_37_cast_fp16 = mul(x = var_3093_cast_fp16, y = var_3087_to_fp16)[name = tensor("attn_weights_37_cast_fp16")]; + tensor var_3095 = logical_not(x = attn_mask_27)[name = tensor("op_3095")]; + tensor var_3096_to_fp16 = const()[name = tensor("op_3096_to_fp16"), val = tensor(-inf)]; + tensor attn_weights_39_cast_fp16 = select(a = var_3096_to_fp16, b = attn_weights_37_cast_fp16, cond = var_3095)[name = tensor("attn_weights_39_cast_fp16")]; + tensor var_3098 = const()[name = tensor("op_3098"), val = tensor(-1)]; + tensor attn_weights_41_cast_fp16 = softmax(axis = var_3098, x = attn_weights_39_cast_fp16)[name = tensor("attn_weights_41_cast_fp16")]; + tensor attn_output_13_transpose_x_0 = const()[name = tensor("attn_output_13_transpose_x_0"), val = tensor(false)]; + tensor attn_output_13_transpose_y_0 = const()[name = tensor("attn_output_13_transpose_y_0"), val = tensor(false)]; + tensor values_41_cast_fp16 = transpose(perm = var_3044, x = values_39_cast_fp16)[name = tensor("transpose_191")]; + tensor attn_output_13_cast_fp16 = matmul(transpose_x = attn_output_13_transpose_x_0, transpose_y = attn_output_13_transpose_y_0, x = attn_weights_41_cast_fp16, y = values_41_cast_fp16)[name = tensor("attn_output_13_cast_fp16")]; + tensor var_3106 = const()[name = tensor("op_3106"), val = tensor([0, 2, 1, 3])]; + tensor var_3109 = const()[name = tensor("op_3109"), val = tensor([1, 1, 1024])]; + tensor var_3107_cast_fp16 = transpose(perm = var_3106, x = attn_output_13_cast_fp16)[name = tensor("transpose_188")]; + tensor input_65_cast_fp16 = reshape(shape = var_3109, x = var_3107_cast_fp16)[name = tensor("input_65_cast_fp16")]; + tensor linear_26_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = attn6_out_proj_weight_cast_fp16, x = input_65_cast_fp16)[name = tensor("linear_26_cast_fp16")]; + tensor var_3115_to_fp16 = const()[name = tensor("op_3115_to_fp16"), val = tensor(0x1p+0)]; + tensor var_3116_cast_fp16 = add(x = position6_to_fp16, y = var_3115_to_fp16)[name = tensor("op_3116_cast_fp16")]; + tensor var_3116_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_3116_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor input_67_cast_fp16 = add(x = input_63_cast_fp16, y = linear_26_cast_fp16)[name = tensor("input_67_cast_fp16")]; + tensor input_69_axes_0 = const()[name = tensor("input_69_axes_0"), val = tensor([-1])]; + tensor norm6_2_weight_to_fp16 = const()[name = tensor("norm6_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(303640704)))]; + tensor norm6_2_bias_to_fp16 = const()[name = tensor("norm6_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(303642816)))]; + tensor var_3120_to_fp16 = const()[name = tensor("op_3120_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_69_cast_fp16 = layer_norm(axes = input_69_axes_0, beta = norm6_2_bias_to_fp16, epsilon = var_3120_to_fp16, gamma = norm6_2_weight_to_fp16, x = input_67_cast_fp16)[name = tensor("input_69_cast_fp16")]; + tensor linear_27_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = linear6_1_weight_cast_fp16, x = input_69_cast_fp16)[name = tensor("linear_27_cast_fp16")]; + tensor input_71_mode_0 = const()[name = tensor("input_71_mode_0"), val = tensor("EXACT")]; + tensor input_71_cast_fp16 = gelu(mode = input_71_mode_0, x = linear_27_cast_fp16)[name = tensor("input_71_cast_fp16")]; + tensor linear_28_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = linear6_2_weight_cast_fp16, x = input_71_cast_fp16)[name = tensor("linear_28_cast_fp16")]; + tensor input_73_cast_fp16 = add(x = input_67_cast_fp16, y = linear_28_cast_fp16)[name = tensor("input_73_cast_fp16")]; + tensor x_15_axes_0 = const()[name = tensor("x_15_axes_0"), val = tensor([-1])]; + tensor norm7_1_weight_to_fp16 = const()[name = tensor("norm7_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(303644928)))]; + tensor norm7_1_bias_to_fp16 = const()[name = tensor("norm7_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(303647040)))]; + tensor var_3137_to_fp16 = const()[name = tensor("op_3137_to_fp16"), val = tensor(0x1.5p-17)]; + tensor x_15_cast_fp16 = layer_norm(axes = x_15_axes_0, beta = norm7_1_bias_to_fp16, epsilon = var_3137_to_fp16, gamma = norm7_1_weight_to_fp16, x = input_73_cast_fp16)[name = tensor("x_15_cast_fp16")]; + tensor linear_29_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = attn7_in_proj_weight_cast_fp16, x = x_15_cast_fp16)[name = tensor("linear_29_cast_fp16")]; + tensor var_3173 = const()[name = tensor("op_3173"), val = tensor([1, 1, 3, 16, 64])]; + tensor qkv_15_cast_fp16 = reshape(shape = var_3173, x = linear_29_cast_fp16)[name = tensor("qkv_15_cast_fp16")]; + tensor q_43_begin_0 = const()[name = tensor("q_43_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor q_43_end_0 = const()[name = tensor("q_43_end_0"), val = tensor([1, 1, 1, 16, 64])]; + tensor q_43_end_mask_0 = const()[name = tensor("q_43_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor q_43_squeeze_mask_0 = const()[name = tensor("q_43_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor q_43_cast_fp16 = slice_by_index(begin = q_43_begin_0, end = q_43_end_0, end_mask = q_43_end_mask_0, squeeze_mask = q_43_squeeze_mask_0, x = qkv_15_cast_fp16)[name = tensor("q_43_cast_fp16")]; + tensor k_29_begin_0 = const()[name = tensor("k_29_begin_0"), val = tensor([0, 0, 1, 0, 0])]; + tensor k_29_end_0 = const()[name = tensor("k_29_end_0"), val = tensor([1, 1, 2, 16, 64])]; + tensor k_29_end_mask_0 = const()[name = tensor("k_29_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor k_29_squeeze_mask_0 = const()[name = tensor("k_29_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor k_29_cast_fp16 = slice_by_index(begin = k_29_begin_0, end = k_29_end_0, end_mask = k_29_end_mask_0, squeeze_mask = k_29_squeeze_mask_0, x = qkv_15_cast_fp16)[name = tensor("k_29_cast_fp16")]; + tensor v_15_begin_0 = const()[name = tensor("v_15_begin_0"), val = tensor([0, 0, 2, 0, 0])]; + tensor v_15_end_0 = const()[name = tensor("v_15_end_0"), val = tensor([1, 1, 3, 16, 64])]; + tensor v_15_end_mask_0 = const()[name = tensor("v_15_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor v_15_squeeze_mask_0 = const()[name = tensor("v_15_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor v_15_cast_fp16 = slice_by_index(begin = v_15_begin_0, end = v_15_end_0, end_mask = v_15_end_mask_0, squeeze_mask = v_15_squeeze_mask_0, x = qkv_15_cast_fp16)[name = tensor("v_15_cast_fp16")]; + tensor var_3277 = const()[name = tensor("op_3277"), val = tensor([1, 1, 1, 1])]; + tensor position7_to_fp16_dtype_0 = const()[name = tensor("position7_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor position7_to_fp16 = cast(dtype = position7_to_fp16_dtype_0, x = position7)[name = tensor("cast_518")]; + tensor ts_47_cast_fp16 = reshape(shape = var_3277, x = position7_to_fp16)[name = tensor("ts_47_cast_fp16")]; + tensor var_3281 = const()[name = tensor("op_3281"), val = tensor([1, 1, 16, 32, 2])]; + tensor q_complex_15_cast_fp16 = reshape(shape = var_3281, x = q_43_cast_fp16)[name = tensor("q_complex_15_cast_fp16")]; + tensor var_3285 = const()[name = tensor("op_3285"), val = tensor([1, 1, 16, 32, 2])]; + tensor k_complex_15_cast_fp16 = reshape(shape = var_3285, x = k_29_cast_fp16)[name = tensor("k_complex_15_cast_fp16")]; + tensor var_3289_begin_0 = const()[name = tensor("op_3289_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_3289_end_0 = const()[name = tensor("op_3289_end_0"), val = tensor([1, 1, 16, 32, 1])]; + tensor var_3289_end_mask_0 = const()[name = tensor("op_3289_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_3289_squeeze_mask_0 = const()[name = tensor("op_3289_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_3289_cast_fp16 = slice_by_index(begin = var_3289_begin_0, end = var_3289_end_0, end_mask = var_3289_end_mask_0, squeeze_mask = var_3289_squeeze_mask_0, x = q_complex_15_cast_fp16)[name = tensor("op_3289_cast_fp16")]; + tensor var_3297_begin_0 = const()[name = tensor("op_3297_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_3297_end_0 = const()[name = tensor("op_3297_end_0"), val = tensor([1, 1, 16, 32, 2])]; + tensor var_3297_end_mask_0 = const()[name = tensor("op_3297_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_3297_squeeze_mask_0 = const()[name = tensor("op_3297_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_3297_cast_fp16 = slice_by_index(begin = var_3297_begin_0, end = var_3297_end_0, end_mask = var_3297_end_mask_0, squeeze_mask = var_3297_squeeze_mask_0, x = q_complex_15_cast_fp16)[name = tensor("op_3297_cast_fp16")]; + tensor var_3305_begin_0 = const()[name = tensor("op_3305_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_3305_end_0 = const()[name = tensor("op_3305_end_0"), val = tensor([1, 1, 16, 32, 1])]; + tensor var_3305_end_mask_0 = const()[name = tensor("op_3305_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_3305_squeeze_mask_0 = const()[name = tensor("op_3305_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_3305_cast_fp16 = slice_by_index(begin = var_3305_begin_0, end = var_3305_end_0, end_mask = var_3305_end_mask_0, squeeze_mask = var_3305_squeeze_mask_0, x = k_complex_15_cast_fp16)[name = tensor("op_3305_cast_fp16")]; + tensor var_3313_begin_0 = const()[name = tensor("op_3313_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_3313_end_0 = const()[name = tensor("op_3313_end_0"), val = tensor([1, 1, 16, 32, 2])]; + tensor var_3313_end_mask_0 = const()[name = tensor("op_3313_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_3313_squeeze_mask_0 = const()[name = tensor("op_3313_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_3313_cast_fp16 = slice_by_index(begin = var_3313_begin_0, end = var_3313_end_0, end_mask = var_3313_end_mask_0, squeeze_mask = var_3313_squeeze_mask_0, x = k_complex_15_cast_fp16)[name = tensor("op_3313_cast_fp16")]; + tensor freqs_15_to_fp16 = const()[name = tensor("freqs_15_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(303649152)))]; + tensor var_3319_cast_fp16 = mul(x = freqs_15_to_fp16, y = ts_47_cast_fp16)[name = tensor("op_3319_cast_fp16")]; + tensor rotr_15_cast_fp16 = cos(x = var_3319_cast_fp16)[name = tensor("rotr_15_cast_fp16")]; + tensor roti_15_cast_fp16 = sin(x = var_3319_cast_fp16)[name = tensor("roti_15_cast_fp16")]; + tensor var_3323_cast_fp16 = mul(x = var_3289_cast_fp16, y = rotr_15_cast_fp16)[name = tensor("op_3323_cast_fp16")]; + tensor var_3324_cast_fp16 = mul(x = var_3297_cast_fp16, y = roti_15_cast_fp16)[name = tensor("op_3324_cast_fp16")]; + tensor qor_29_cast_fp16 = sub(x = var_3323_cast_fp16, y = var_3324_cast_fp16)[name = tensor("qor_29_cast_fp16")]; + tensor var_3327_cast_fp16 = mul(x = var_3289_cast_fp16, y = roti_15_cast_fp16)[name = tensor("op_3327_cast_fp16")]; + tensor var_3328_cast_fp16 = mul(x = var_3297_cast_fp16, y = rotr_15_cast_fp16)[name = tensor("op_3328_cast_fp16")]; + tensor qoi_29_cast_fp16 = add(x = var_3327_cast_fp16, y = var_3328_cast_fp16)[name = tensor("qoi_29_cast_fp16")]; + tensor var_3331_cast_fp16 = mul(x = var_3305_cast_fp16, y = rotr_15_cast_fp16)[name = tensor("op_3331_cast_fp16")]; + tensor var_3332_cast_fp16 = mul(x = var_3313_cast_fp16, y = roti_15_cast_fp16)[name = tensor("op_3332_cast_fp16")]; + tensor kor_29_cast_fp16 = sub(x = var_3331_cast_fp16, y = var_3332_cast_fp16)[name = tensor("kor_29_cast_fp16")]; + tensor var_3335_cast_fp16 = mul(x = var_3305_cast_fp16, y = roti_15_cast_fp16)[name = tensor("op_3335_cast_fp16")]; + tensor var_3336_cast_fp16 = mul(x = var_3313_cast_fp16, y = rotr_15_cast_fp16)[name = tensor("op_3336_cast_fp16")]; + tensor koi_29_cast_fp16 = add(x = var_3335_cast_fp16, y = var_3336_cast_fp16)[name = tensor("koi_29_cast_fp16")]; + tensor qo_15_axis_0 = const()[name = tensor("qo_15_axis_0"), val = tensor(-1)]; + tensor qo_15_cast_fp16 = stack(axis = qo_15_axis_0, values = (qor_29_cast_fp16, qoi_29_cast_fp16))[name = tensor("qo_15_cast_fp16")]; + tensor ko_15_axis_0 = const()[name = tensor("ko_15_axis_0"), val = tensor(-1)]; + tensor ko_15_cast_fp16 = stack(axis = ko_15_axis_0, values = (kor_29_cast_fp16, koi_29_cast_fp16))[name = tensor("ko_15_cast_fp16")]; + tensor var_3365 = const()[name = tensor("op_3365"), val = tensor([1, 1, 16, 64])]; + tensor q_45_cast_fp16 = reshape(shape = var_3365, x = qo_15_cast_fp16)[name = tensor("q_45_cast_fp16")]; + tensor var_3367 = const()[name = tensor("op_3367"), val = tensor([1, 1, 16, 64])]; + tensor k_31_cast_fp16 = reshape(shape = var_3367, x = ko_15_cast_fp16)[name = tensor("k_31_cast_fp16")]; + tensor _inversed_3389_y_0_to_fp16 = const()[name = tensor("_inversed_3389_y_0_to_fp16"), val = tensor(0x1p-9)]; + tensor _inversed_3389_cast_fp16 = mul(x = ts_47_cast_fp16, y = _inversed_3389_y_0_to_fp16)[name = tensor("_inversed_3389_cast_fp16")]; + tensor var_3390_cast_fp16 = floor(x = _inversed_3389_cast_fp16)[name = tensor("op_3390_cast_fp16")]; + tensor var_3391_to_fp16 = const()[name = tensor("op_3391_to_fp16"), val = tensor(0x1p+9)]; + tensor var_3392_cast_fp16 = mul(x = var_3390_cast_fp16, y = var_3391_to_fp16)[name = tensor("op_3392_cast_fp16")]; + tensor write_indices_float_31_cast_fp16 = sub(x = ts_47_cast_fp16, y = var_3392_cast_fp16)[name = tensor("write_indices_float_31_cast_fp16")]; + tensor var_3399_dtype_0 = const()[name = tensor("op_3399_dtype_0"), val = tensor("int32")]; + tensor write_indices_15_reps_0 = const()[name = tensor("write_indices_15_reps_0"), val = tensor([1, 1, 16, 64])]; + tensor write_indices_float_31_cast_fp16_to_int32 = cast(dtype = var_3399_dtype_0, x = write_indices_float_31_cast_fp16)[name = tensor("cast_517")]; + tensor write_indices_15 = tile(reps = write_indices_15_reps_0, x = write_indices_float_31_cast_fp16_to_int32)[name = tensor("write_indices_15")]; + tensor var_3407_begin_0 = const()[name = tensor("op_3407_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_3407_end_0 = const()[name = tensor("op_3407_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor var_3407_end_mask_0 = const()[name = tensor("op_3407_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_3407_squeeze_mask_0 = const()[name = tensor("op_3407_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor cache7_to_fp16_dtype_0 = const()[name = tensor("cache7_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor cache7_to_fp16 = cast(dtype = cache7_to_fp16_dtype_0, x = cache7)[name = tensor("cast_516")]; + tensor var_3407_cast_fp16 = slice_by_index(begin = var_3407_begin_0, end = var_3407_end_0, end_mask = var_3407_end_mask_0, squeeze_mask = var_3407_squeeze_mask_0, x = cache7_to_fp16)[name = tensor("op_3407_cast_fp16")]; + tensor var_3409_axis_0 = const()[name = tensor("op_3409_axis_0"), val = tensor(1)]; + tensor var_3409_mode_0 = const()[name = tensor("op_3409_mode_0"), val = tensor("update")]; + tensor var_3409_validate_indices_0 = const()[name = tensor("op_3409_validate_indices_0"), val = tensor(false)]; + tensor var_3409_cast_fp16 = scatter_along_axis(axis = var_3409_axis_0, data = var_3407_cast_fp16, indices = write_indices_15, mode = var_3409_mode_0, updates = k_31_cast_fp16, validate_indices = var_3409_validate_indices_0)[name = tensor("op_3409_cast_fp16")]; + tensor concat_51 = const()[name = tensor("concat_51"), val = tensor([0, 0, 0, 0, 0])]; + tensor concat_52 = const()[name = tensor("concat_52"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_15_internal_tensor_assign_1_stride_0 = const()[name = tensor("new_cache_15_internal_tensor_assign_1_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_15_internal_tensor_assign_1_begin_mask_0 = const()[name = tensor("new_cache_15_internal_tensor_assign_1_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_15_internal_tensor_assign_1_end_mask_0 = const()[name = tensor("new_cache_15_internal_tensor_assign_1_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_15_internal_tensor_assign_1_squeeze_mask_0 = const()[name = tensor("new_cache_15_internal_tensor_assign_1_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_62 = const()[name = tensor("shape_62"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_14 = const()[name = tensor("reduce_prod_14"), val = tensor(1048576)]; + tensor range_1d_14_start_0 = const()[name = tensor("range_1d_14_start_0"), val = tensor(0)]; + tensor range_1d_14_step_0 = const()[name = tensor("range_1d_14_step_0"), val = tensor(1)]; + tensor range_1d_14 = range_1d(end = reduce_prod_14, start = range_1d_14_start_0, step = range_1d_14_step_0)[name = tensor("range_1d_14")]; + tensor reshape_70 = reshape(shape = shape_62, x = range_1d_14)[name = tensor("reshape_70")]; + tensor slice_by_index_14 = slice_by_index(begin = concat_51, begin_mask = new_cache_15_internal_tensor_assign_1_begin_mask_0, end = concat_52, end_mask = new_cache_15_internal_tensor_assign_1_end_mask_0, squeeze_mask = new_cache_15_internal_tensor_assign_1_squeeze_mask_0, stride = new_cache_15_internal_tensor_assign_1_stride_0, x = reshape_70)[name = tensor("slice_by_index_14")]; + tensor reshape_71_shape_0 = const()[name = tensor("reshape_71_shape_0"), val = tensor([-1])]; + tensor reshape_71 = reshape(shape = reshape_71_shape_0, x = slice_by_index_14)[name = tensor("reshape_71")]; + tensor reshape_72_shape_0 = const()[name = tensor("reshape_72_shape_0"), val = tensor([-1])]; + tensor reshape_72_cast_fp16 = reshape(shape = reshape_72_shape_0, x = var_3409_cast_fp16)[name = tensor("reshape_72_cast_fp16")]; + tensor reshape_73_shape_0 = const()[name = tensor("reshape_73_shape_0"), val = tensor([-1])]; + tensor reshape_73_cast_fp16 = reshape(shape = reshape_73_shape_0, x = cache7_to_fp16)[name = tensor("reshape_73_cast_fp16")]; + tensor scatter_14_mode_0 = const()[name = tensor("scatter_14_mode_0"), val = tensor("update")]; + tensor scatter_14_axis_0 = const()[name = tensor("scatter_14_axis_0"), val = tensor(0)]; + tensor scatter_14_validate_indices_0 = const()[name = tensor("scatter_14_validate_indices_0"), val = tensor(false)]; + tensor scatter_14_cast_fp16 = scatter(axis = scatter_14_axis_0, data = reshape_73_cast_fp16, indices = reshape_71, mode = scatter_14_mode_0, updates = reshape_72_cast_fp16, validate_indices = scatter_14_validate_indices_0)[name = tensor("scatter_14_cast_fp16")]; + tensor reshape_74_cast_fp16 = reshape(shape = shape_62, x = scatter_14_cast_fp16)[name = tensor("reshape_74_cast_fp16")]; + tensor var_3417_begin_0 = const()[name = tensor("op_3417_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor var_3417_end_0 = const()[name = tensor("op_3417_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor var_3417_end_mask_0 = const()[name = tensor("op_3417_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_3417_squeeze_mask_0 = const()[name = tensor("op_3417_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor var_3417_cast_fp16 = slice_by_index(begin = var_3417_begin_0, end = var_3417_end_0, end_mask = var_3417_end_mask_0, squeeze_mask = var_3417_squeeze_mask_0, x = reshape_74_cast_fp16)[name = tensor("op_3417_cast_fp16")]; + tensor var_3419_axis_0 = const()[name = tensor("op_3419_axis_0"), val = tensor(1)]; + tensor var_3419_mode_0 = const()[name = tensor("op_3419_mode_0"), val = tensor("update")]; + tensor var_3419_validate_indices_0 = const()[name = tensor("op_3419_validate_indices_0"), val = tensor(false)]; + tensor var_3419_cast_fp16 = scatter_along_axis(axis = var_3419_axis_0, data = var_3417_cast_fp16, indices = write_indices_15, mode = var_3419_mode_0, updates = v_15_cast_fp16, validate_indices = var_3419_validate_indices_0)[name = tensor("op_3419_cast_fp16")]; + tensor concat_53 = const()[name = tensor("concat_53"), val = tensor([1, 0, 0, 0, 0])]; + tensor concat_54 = const()[name = tensor("concat_54"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_15_internal_tensor_assign_2_stride_0 = const()[name = tensor("new_cache_15_internal_tensor_assign_2_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_15_internal_tensor_assign_2_begin_mask_0 = const()[name = tensor("new_cache_15_internal_tensor_assign_2_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_15_internal_tensor_assign_2_end_mask_0 = const()[name = tensor("new_cache_15_internal_tensor_assign_2_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_15_internal_tensor_assign_2_squeeze_mask_0 = const()[name = tensor("new_cache_15_internal_tensor_assign_2_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_63 = const()[name = tensor("shape_63"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_15 = const()[name = tensor("reduce_prod_15"), val = tensor(1048576)]; + tensor range_1d_15_start_0 = const()[name = tensor("range_1d_15_start_0"), val = tensor(0)]; + tensor range_1d_15_step_0 = const()[name = tensor("range_1d_15_step_0"), val = tensor(1)]; + tensor range_1d_15 = range_1d(end = reduce_prod_15, start = range_1d_15_start_0, step = range_1d_15_step_0)[name = tensor("range_1d_15")]; + tensor reshape_75 = reshape(shape = shape_63, x = range_1d_15)[name = tensor("reshape_75")]; + tensor slice_by_index_15 = slice_by_index(begin = concat_53, begin_mask = new_cache_15_internal_tensor_assign_2_begin_mask_0, end = concat_54, end_mask = new_cache_15_internal_tensor_assign_2_end_mask_0, squeeze_mask = new_cache_15_internal_tensor_assign_2_squeeze_mask_0, stride = new_cache_15_internal_tensor_assign_2_stride_0, x = reshape_75)[name = tensor("slice_by_index_15")]; + tensor reshape_76_shape_0 = const()[name = tensor("reshape_76_shape_0"), val = tensor([-1])]; + tensor reshape_76 = reshape(shape = reshape_76_shape_0, x = slice_by_index_15)[name = tensor("reshape_76")]; + tensor reshape_77_shape_0 = const()[name = tensor("reshape_77_shape_0"), val = tensor([-1])]; + tensor reshape_77_cast_fp16 = reshape(shape = reshape_77_shape_0, x = var_3419_cast_fp16)[name = tensor("reshape_77_cast_fp16")]; + tensor reshape_78_shape_0 = const()[name = tensor("reshape_78_shape_0"), val = tensor([-1])]; + tensor reshape_78_cast_fp16 = reshape(shape = reshape_78_shape_0, x = reshape_74_cast_fp16)[name = tensor("reshape_78_cast_fp16")]; + tensor scatter_15_mode_0 = const()[name = tensor("scatter_15_mode_0"), val = tensor("update")]; + tensor scatter_15_axis_0 = const()[name = tensor("scatter_15_axis_0"), val = tensor(0)]; + tensor scatter_15_validate_indices_0 = const()[name = tensor("scatter_15_validate_indices_0"), val = tensor(false)]; + tensor scatter_15_cast_fp16 = scatter(axis = scatter_15_axis_0, data = reshape_78_cast_fp16, indices = reshape_76, mode = scatter_15_mode_0, updates = reshape_77_cast_fp16, validate_indices = scatter_15_validate_indices_0)[name = tensor("scatter_15_cast_fp16")]; + tensor reshape_79_cast_fp16 = reshape(shape = shape_63, x = scatter_15_cast_fp16)[name = tensor("reshape_79_cast_fp16")]; + tensor reshape_79_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("reshape_79_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor keys_43_begin_0 = const()[name = tensor("keys_43_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor keys_43_end_0 = const()[name = tensor("keys_43_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor keys_43_end_mask_0 = const()[name = tensor("keys_43_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor keys_43_squeeze_mask_0 = const()[name = tensor("keys_43_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor keys_43_cast_fp16 = slice_by_index(begin = keys_43_begin_0, end = keys_43_end_0, end_mask = keys_43_end_mask_0, squeeze_mask = keys_43_squeeze_mask_0, x = reshape_79_cast_fp16)[name = tensor("keys_43_cast_fp16")]; + tensor values_43_begin_0 = const()[name = tensor("values_43_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor values_43_end_0 = const()[name = tensor("values_43_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor values_43_end_mask_0 = const()[name = tensor("values_43_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor values_43_squeeze_mask_0 = const()[name = tensor("values_43_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor values_43_cast_fp16 = slice_by_index(begin = values_43_begin_0, end = values_43_end_0, end_mask = values_43_end_mask_0, squeeze_mask = values_43_squeeze_mask_0, x = reshape_79_cast_fp16)[name = tensor("values_43_cast_fp16")]; + tensor var_3431_cast_fp16 = not_equal(x = keys_43_cast_fp16, y = keys_43_cast_fp16)[name = tensor("op_3431_cast_fp16")]; + tensor keys_45_cast_fp16 = select(a = var_504_to_fp16, b = keys_43_cast_fp16, cond = var_3431_cast_fp16)[name = tensor("keys_45_cast_fp16")]; + tensor var_3439_cast_fp16 = not_equal(x = values_43_cast_fp16, y = values_43_cast_fp16)[name = tensor("op_3439_cast_fp16")]; + tensor values_45_cast_fp16 = select(a = var_504_to_fp16, b = values_43_cast_fp16, cond = var_3439_cast_fp16)[name = tensor("values_45_cast_fp16")]; + tensor var_3463 = const()[name = tensor("op_3463"), val = tensor([0, 2, 1, 3])]; + tensor var_3476 = const()[name = tensor("op_3476"), val = tensor([1, 1, 1])]; + tensor var_3477_cast_fp16 = reshape(shape = var_3476, x = position7_to_fp16)[name = tensor("op_3477_cast_fp16")]; + tensor var_3494_to_fp16 = const()[name = tensor("op_3494_to_fp16"), val = tensor(0x1p+0)]; + tensor valid_len_15_cast_fp16 = add(x = var_3477_cast_fp16, y = var_3494_to_fp16)[name = tensor("valid_len_15_cast_fp16")]; + tensor valid_mask_15_cast_fp16 = less(x = k_positions_1_promoted_to_fp16, y = valid_len_15_cast_fp16)[name = tensor("valid_mask_15_cast_fp16")]; + tensor causal_mask_15_cast_fp16 = less_equal(x = k_positions_1_promoted_to_fp16, y = var_3477_cast_fp16)[name = tensor("causal_mask_15_cast_fp16")]; + tensor attn_mask_29 = logical_and(x = valid_mask_15_cast_fp16, y = causal_mask_15_cast_fp16)[name = tensor("attn_mask_29")]; + tensor attn_mask_31_axes_0 = const()[name = tensor("attn_mask_31_axes_0"), val = tensor([1])]; + tensor attn_mask_31 = expand_dims(axes = attn_mask_31_axes_0, x = attn_mask_29)[name = tensor("attn_mask_31")]; + tensor var_3512_transpose_x_0 = const()[name = tensor("op_3512_transpose_x_0"), val = tensor(false)]; + tensor var_3512_transpose_y_0 = const()[name = tensor("op_3512_transpose_y_0"), val = tensor(false)]; + tensor transpose_86_perm_0 = const()[name = tensor("transpose_86_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_87_perm_0 = const()[name = tensor("transpose_87_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_87 = transpose(perm = transpose_87_perm_0, x = keys_45_cast_fp16)[name = tensor("transpose_185")]; + tensor transpose_86 = transpose(perm = transpose_86_perm_0, x = q_45_cast_fp16)[name = tensor("transpose_186")]; + tensor var_3512_cast_fp16 = matmul(transpose_x = var_3512_transpose_x_0, transpose_y = var_3512_transpose_y_0, x = transpose_86, y = transpose_87)[name = tensor("op_3512_cast_fp16")]; + tensor var_3506_to_fp16 = const()[name = tensor("op_3506_to_fp16"), val = tensor([0x1p-3])]; + tensor attn_weights_43_cast_fp16 = mul(x = var_3512_cast_fp16, y = var_3506_to_fp16)[name = tensor("attn_weights_43_cast_fp16")]; + tensor var_3514 = logical_not(x = attn_mask_31)[name = tensor("op_3514")]; + tensor var_3515_to_fp16 = const()[name = tensor("op_3515_to_fp16"), val = tensor(-inf)]; + tensor attn_weights_45_cast_fp16 = select(a = var_3515_to_fp16, b = attn_weights_43_cast_fp16, cond = var_3514)[name = tensor("attn_weights_45_cast_fp16")]; + tensor var_3517 = const()[name = tensor("op_3517"), val = tensor(-1)]; + tensor attn_weights_47_cast_fp16 = softmax(axis = var_3517, x = attn_weights_45_cast_fp16)[name = tensor("attn_weights_47_cast_fp16")]; + tensor attn_output_15_transpose_x_0 = const()[name = tensor("attn_output_15_transpose_x_0"), val = tensor(false)]; + tensor attn_output_15_transpose_y_0 = const()[name = tensor("attn_output_15_transpose_y_0"), val = tensor(false)]; + tensor values_47_cast_fp16 = transpose(perm = var_3463, x = values_45_cast_fp16)[name = tensor("transpose_187")]; + tensor attn_output_15_cast_fp16 = matmul(transpose_x = attn_output_15_transpose_x_0, transpose_y = attn_output_15_transpose_y_0, x = attn_weights_47_cast_fp16, y = values_47_cast_fp16)[name = tensor("attn_output_15_cast_fp16")]; + tensor var_3525 = const()[name = tensor("op_3525"), val = tensor([0, 2, 1, 3])]; + tensor var_3528 = const()[name = tensor("op_3528"), val = tensor([1, 1, 1024])]; + tensor var_3526_cast_fp16 = transpose(perm = var_3525, x = attn_output_15_cast_fp16)[name = tensor("transpose_184")]; + tensor input_75_cast_fp16 = reshape(shape = var_3528, x = var_3526_cast_fp16)[name = tensor("input_75_cast_fp16")]; + tensor linear_30_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = attn7_out_proj_weight_cast_fp16, x = input_75_cast_fp16)[name = tensor("linear_30_cast_fp16")]; + tensor var_3534_to_fp16 = const()[name = tensor("op_3534_to_fp16"), val = tensor(0x1p+0)]; + tensor var_3535_cast_fp16 = add(x = position7_to_fp16, y = var_3534_to_fp16)[name = tensor("op_3535_cast_fp16")]; + tensor var_3535_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_3535_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor input_77_cast_fp16 = add(x = input_73_cast_fp16, y = linear_30_cast_fp16)[name = tensor("input_77_cast_fp16")]; + tensor input_79_axes_0 = const()[name = tensor("input_79_axes_0"), val = tensor([-1])]; + tensor norm7_2_weight_to_fp16 = const()[name = tensor("norm7_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(303649280)))]; + tensor norm7_2_bias_to_fp16 = const()[name = tensor("norm7_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(303651392)))]; + tensor var_3539_to_fp16 = const()[name = tensor("op_3539_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_79_cast_fp16 = layer_norm(axes = input_79_axes_0, beta = norm7_2_bias_to_fp16, epsilon = var_3539_to_fp16, gamma = norm7_2_weight_to_fp16, x = input_77_cast_fp16)[name = tensor("input_79_cast_fp16")]; + tensor linear_31_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = linear7_1_weight_cast_fp16, x = input_79_cast_fp16)[name = tensor("linear_31_cast_fp16")]; + tensor input_81_mode_0 = const()[name = tensor("input_81_mode_0"), val = tensor("EXACT")]; + tensor input_81_cast_fp16 = gelu(mode = input_81_mode_0, x = linear_31_cast_fp16)[name = tensor("input_81_cast_fp16")]; + tensor linear_32_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = linear7_2_weight_cast_fp16, x = input_81_cast_fp16)[name = tensor("linear_32_cast_fp16")]; + tensor input_83_cast_fp16 = add(x = input_77_cast_fp16, y = linear_32_cast_fp16)[name = tensor("input_83_cast_fp16")]; + tensor x_17_axes_0 = const()[name = tensor("x_17_axes_0"), val = tensor([-1])]; + tensor norm8_1_weight_to_fp16 = const()[name = tensor("norm8_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(303653504)))]; + tensor norm8_1_bias_to_fp16 = const()[name = tensor("norm8_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(303655616)))]; + tensor var_3556_to_fp16 = const()[name = tensor("op_3556_to_fp16"), val = tensor(0x1.5p-17)]; + tensor x_17_cast_fp16 = layer_norm(axes = x_17_axes_0, beta = norm8_1_bias_to_fp16, epsilon = var_3556_to_fp16, gamma = norm8_1_weight_to_fp16, x = input_83_cast_fp16)[name = tensor("x_17_cast_fp16")]; + tensor linear_33_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = attn8_in_proj_weight_cast_fp16, x = x_17_cast_fp16)[name = tensor("linear_33_cast_fp16")]; + tensor var_3592 = const()[name = tensor("op_3592"), val = tensor([1, 1, 3, 16, 64])]; + tensor qkv_17_cast_fp16 = reshape(shape = var_3592, x = linear_33_cast_fp16)[name = tensor("qkv_17_cast_fp16")]; + tensor q_49_begin_0 = const()[name = tensor("q_49_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor q_49_end_0 = const()[name = tensor("q_49_end_0"), val = tensor([1, 1, 1, 16, 64])]; + tensor q_49_end_mask_0 = const()[name = tensor("q_49_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor q_49_squeeze_mask_0 = const()[name = tensor("q_49_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor q_49_cast_fp16 = slice_by_index(begin = q_49_begin_0, end = q_49_end_0, end_mask = q_49_end_mask_0, squeeze_mask = q_49_squeeze_mask_0, x = qkv_17_cast_fp16)[name = tensor("q_49_cast_fp16")]; + tensor k_33_begin_0 = const()[name = tensor("k_33_begin_0"), val = tensor([0, 0, 1, 0, 0])]; + tensor k_33_end_0 = const()[name = tensor("k_33_end_0"), val = tensor([1, 1, 2, 16, 64])]; + tensor k_33_end_mask_0 = const()[name = tensor("k_33_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor k_33_squeeze_mask_0 = const()[name = tensor("k_33_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor k_33_cast_fp16 = slice_by_index(begin = k_33_begin_0, end = k_33_end_0, end_mask = k_33_end_mask_0, squeeze_mask = k_33_squeeze_mask_0, x = qkv_17_cast_fp16)[name = tensor("k_33_cast_fp16")]; + tensor v_17_begin_0 = const()[name = tensor("v_17_begin_0"), val = tensor([0, 0, 2, 0, 0])]; + tensor v_17_end_0 = const()[name = tensor("v_17_end_0"), val = tensor([1, 1, 3, 16, 64])]; + tensor v_17_end_mask_0 = const()[name = tensor("v_17_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor v_17_squeeze_mask_0 = const()[name = tensor("v_17_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor v_17_cast_fp16 = slice_by_index(begin = v_17_begin_0, end = v_17_end_0, end_mask = v_17_end_mask_0, squeeze_mask = v_17_squeeze_mask_0, x = qkv_17_cast_fp16)[name = tensor("v_17_cast_fp16")]; + tensor var_3696 = const()[name = tensor("op_3696"), val = tensor([1, 1, 1, 1])]; + tensor position8_to_fp16_dtype_0 = const()[name = tensor("position8_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor position8_to_fp16 = cast(dtype = position8_to_fp16_dtype_0, x = position8)[name = tensor("cast_513")]; + tensor ts_53_cast_fp16 = reshape(shape = var_3696, x = position8_to_fp16)[name = tensor("ts_53_cast_fp16")]; + tensor var_3700 = const()[name = tensor("op_3700"), val = tensor([1, 1, 16, 32, 2])]; + tensor q_complex_17_cast_fp16 = reshape(shape = var_3700, x = q_49_cast_fp16)[name = tensor("q_complex_17_cast_fp16")]; + tensor var_3704 = const()[name = tensor("op_3704"), val = tensor([1, 1, 16, 32, 2])]; + tensor k_complex_17_cast_fp16 = reshape(shape = var_3704, x = k_33_cast_fp16)[name = tensor("k_complex_17_cast_fp16")]; + tensor var_3708_begin_0 = const()[name = tensor("op_3708_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_3708_end_0 = const()[name = tensor("op_3708_end_0"), val = tensor([1, 1, 16, 32, 1])]; + tensor var_3708_end_mask_0 = const()[name = tensor("op_3708_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_3708_squeeze_mask_0 = const()[name = tensor("op_3708_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_3708_cast_fp16 = slice_by_index(begin = var_3708_begin_0, end = var_3708_end_0, end_mask = var_3708_end_mask_0, squeeze_mask = var_3708_squeeze_mask_0, x = q_complex_17_cast_fp16)[name = tensor("op_3708_cast_fp16")]; + tensor var_3716_begin_0 = const()[name = tensor("op_3716_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_3716_end_0 = const()[name = tensor("op_3716_end_0"), val = tensor([1, 1, 16, 32, 2])]; + tensor var_3716_end_mask_0 = const()[name = tensor("op_3716_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_3716_squeeze_mask_0 = const()[name = tensor("op_3716_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_3716_cast_fp16 = slice_by_index(begin = var_3716_begin_0, end = var_3716_end_0, end_mask = var_3716_end_mask_0, squeeze_mask = var_3716_squeeze_mask_0, x = q_complex_17_cast_fp16)[name = tensor("op_3716_cast_fp16")]; + tensor var_3724_begin_0 = const()[name = tensor("op_3724_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_3724_end_0 = const()[name = tensor("op_3724_end_0"), val = tensor([1, 1, 16, 32, 1])]; + tensor var_3724_end_mask_0 = const()[name = tensor("op_3724_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_3724_squeeze_mask_0 = const()[name = tensor("op_3724_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_3724_cast_fp16 = slice_by_index(begin = var_3724_begin_0, end = var_3724_end_0, end_mask = var_3724_end_mask_0, squeeze_mask = var_3724_squeeze_mask_0, x = k_complex_17_cast_fp16)[name = tensor("op_3724_cast_fp16")]; + tensor var_3732_begin_0 = const()[name = tensor("op_3732_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_3732_end_0 = const()[name = tensor("op_3732_end_0"), val = tensor([1, 1, 16, 32, 2])]; + tensor var_3732_end_mask_0 = const()[name = tensor("op_3732_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_3732_squeeze_mask_0 = const()[name = tensor("op_3732_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_3732_cast_fp16 = slice_by_index(begin = var_3732_begin_0, end = var_3732_end_0, end_mask = var_3732_end_mask_0, squeeze_mask = var_3732_squeeze_mask_0, x = k_complex_17_cast_fp16)[name = tensor("op_3732_cast_fp16")]; + tensor freqs_17_to_fp16 = const()[name = tensor("freqs_17_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(303657728)))]; + tensor var_3738_cast_fp16 = mul(x = freqs_17_to_fp16, y = ts_53_cast_fp16)[name = tensor("op_3738_cast_fp16")]; + tensor rotr_17_cast_fp16 = cos(x = var_3738_cast_fp16)[name = tensor("rotr_17_cast_fp16")]; + tensor roti_17_cast_fp16 = sin(x = var_3738_cast_fp16)[name = tensor("roti_17_cast_fp16")]; + tensor var_3742_cast_fp16 = mul(x = var_3708_cast_fp16, y = rotr_17_cast_fp16)[name = tensor("op_3742_cast_fp16")]; + tensor var_3743_cast_fp16 = mul(x = var_3716_cast_fp16, y = roti_17_cast_fp16)[name = tensor("op_3743_cast_fp16")]; + tensor qor_33_cast_fp16 = sub(x = var_3742_cast_fp16, y = var_3743_cast_fp16)[name = tensor("qor_33_cast_fp16")]; + tensor var_3746_cast_fp16 = mul(x = var_3708_cast_fp16, y = roti_17_cast_fp16)[name = tensor("op_3746_cast_fp16")]; + tensor var_3747_cast_fp16 = mul(x = var_3716_cast_fp16, y = rotr_17_cast_fp16)[name = tensor("op_3747_cast_fp16")]; + tensor qoi_33_cast_fp16 = add(x = var_3746_cast_fp16, y = var_3747_cast_fp16)[name = tensor("qoi_33_cast_fp16")]; + tensor var_3750_cast_fp16 = mul(x = var_3724_cast_fp16, y = rotr_17_cast_fp16)[name = tensor("op_3750_cast_fp16")]; + tensor var_3751_cast_fp16 = mul(x = var_3732_cast_fp16, y = roti_17_cast_fp16)[name = tensor("op_3751_cast_fp16")]; + tensor kor_33_cast_fp16 = sub(x = var_3750_cast_fp16, y = var_3751_cast_fp16)[name = tensor("kor_33_cast_fp16")]; + tensor var_3754_cast_fp16 = mul(x = var_3724_cast_fp16, y = roti_17_cast_fp16)[name = tensor("op_3754_cast_fp16")]; + tensor var_3755_cast_fp16 = mul(x = var_3732_cast_fp16, y = rotr_17_cast_fp16)[name = tensor("op_3755_cast_fp16")]; + tensor koi_33_cast_fp16 = add(x = var_3754_cast_fp16, y = var_3755_cast_fp16)[name = tensor("koi_33_cast_fp16")]; + tensor qo_17_axis_0 = const()[name = tensor("qo_17_axis_0"), val = tensor(-1)]; + tensor qo_17_cast_fp16 = stack(axis = qo_17_axis_0, values = (qor_33_cast_fp16, qoi_33_cast_fp16))[name = tensor("qo_17_cast_fp16")]; + tensor ko_17_axis_0 = const()[name = tensor("ko_17_axis_0"), val = tensor(-1)]; + tensor ko_17_cast_fp16 = stack(axis = ko_17_axis_0, values = (kor_33_cast_fp16, koi_33_cast_fp16))[name = tensor("ko_17_cast_fp16")]; + tensor var_3784 = const()[name = tensor("op_3784"), val = tensor([1, 1, 16, 64])]; + tensor q_51_cast_fp16 = reshape(shape = var_3784, x = qo_17_cast_fp16)[name = tensor("q_51_cast_fp16")]; + tensor var_3786 = const()[name = tensor("op_3786"), val = tensor([1, 1, 16, 64])]; + tensor k_35_cast_fp16 = reshape(shape = var_3786, x = ko_17_cast_fp16)[name = tensor("k_35_cast_fp16")]; + tensor _inversed_3808_y_0_to_fp16 = const()[name = tensor("_inversed_3808_y_0_to_fp16"), val = tensor(0x1p-9)]; + tensor _inversed_3808_cast_fp16 = mul(x = ts_53_cast_fp16, y = _inversed_3808_y_0_to_fp16)[name = tensor("_inversed_3808_cast_fp16")]; + tensor var_3809_cast_fp16 = floor(x = _inversed_3808_cast_fp16)[name = tensor("op_3809_cast_fp16")]; + tensor var_3810_to_fp16 = const()[name = tensor("op_3810_to_fp16"), val = tensor(0x1p+9)]; + tensor var_3811_cast_fp16 = mul(x = var_3809_cast_fp16, y = var_3810_to_fp16)[name = tensor("op_3811_cast_fp16")]; + tensor write_indices_float_35_cast_fp16 = sub(x = ts_53_cast_fp16, y = var_3811_cast_fp16)[name = tensor("write_indices_float_35_cast_fp16")]; + tensor var_3818_dtype_0 = const()[name = tensor("op_3818_dtype_0"), val = tensor("int32")]; + tensor write_indices_17_reps_0 = const()[name = tensor("write_indices_17_reps_0"), val = tensor([1, 1, 16, 64])]; + tensor write_indices_float_35_cast_fp16_to_int32 = cast(dtype = var_3818_dtype_0, x = write_indices_float_35_cast_fp16)[name = tensor("cast_512")]; + tensor write_indices_17 = tile(reps = write_indices_17_reps_0, x = write_indices_float_35_cast_fp16_to_int32)[name = tensor("write_indices_17")]; + tensor var_3826_begin_0 = const()[name = tensor("op_3826_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_3826_end_0 = const()[name = tensor("op_3826_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor var_3826_end_mask_0 = const()[name = tensor("op_3826_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_3826_squeeze_mask_0 = const()[name = tensor("op_3826_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor cache8_to_fp16_dtype_0 = const()[name = tensor("cache8_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor cache8_to_fp16 = cast(dtype = cache8_to_fp16_dtype_0, x = cache8)[name = tensor("cast_511")]; + tensor var_3826_cast_fp16 = slice_by_index(begin = var_3826_begin_0, end = var_3826_end_0, end_mask = var_3826_end_mask_0, squeeze_mask = var_3826_squeeze_mask_0, x = cache8_to_fp16)[name = tensor("op_3826_cast_fp16")]; + tensor var_3828_axis_0 = const()[name = tensor("op_3828_axis_0"), val = tensor(1)]; + tensor var_3828_mode_0 = const()[name = tensor("op_3828_mode_0"), val = tensor("update")]; + tensor var_3828_validate_indices_0 = const()[name = tensor("op_3828_validate_indices_0"), val = tensor(false)]; + tensor var_3828_cast_fp16 = scatter_along_axis(axis = var_3828_axis_0, data = var_3826_cast_fp16, indices = write_indices_17, mode = var_3828_mode_0, updates = k_35_cast_fp16, validate_indices = var_3828_validate_indices_0)[name = tensor("op_3828_cast_fp16")]; + tensor concat_58 = const()[name = tensor("concat_58"), val = tensor([0, 0, 0, 0, 0])]; + tensor concat_59 = const()[name = tensor("concat_59"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_17_internal_tensor_assign_1_stride_0 = const()[name = tensor("new_cache_17_internal_tensor_assign_1_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_17_internal_tensor_assign_1_begin_mask_0 = const()[name = tensor("new_cache_17_internal_tensor_assign_1_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_17_internal_tensor_assign_1_end_mask_0 = const()[name = tensor("new_cache_17_internal_tensor_assign_1_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_17_internal_tensor_assign_1_squeeze_mask_0 = const()[name = tensor("new_cache_17_internal_tensor_assign_1_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_64 = const()[name = tensor("shape_64"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_16 = const()[name = tensor("reduce_prod_16"), val = tensor(1048576)]; + tensor range_1d_16_start_0 = const()[name = tensor("range_1d_16_start_0"), val = tensor(0)]; + tensor range_1d_16_step_0 = const()[name = tensor("range_1d_16_step_0"), val = tensor(1)]; + tensor range_1d_16 = range_1d(end = reduce_prod_16, start = range_1d_16_start_0, step = range_1d_16_step_0)[name = tensor("range_1d_16")]; + tensor reshape_80 = reshape(shape = shape_64, x = range_1d_16)[name = tensor("reshape_80")]; + tensor slice_by_index_16 = slice_by_index(begin = concat_58, begin_mask = new_cache_17_internal_tensor_assign_1_begin_mask_0, end = concat_59, end_mask = new_cache_17_internal_tensor_assign_1_end_mask_0, squeeze_mask = new_cache_17_internal_tensor_assign_1_squeeze_mask_0, stride = new_cache_17_internal_tensor_assign_1_stride_0, x = reshape_80)[name = tensor("slice_by_index_16")]; + tensor reshape_81_shape_0 = const()[name = tensor("reshape_81_shape_0"), val = tensor([-1])]; + tensor reshape_81 = reshape(shape = reshape_81_shape_0, x = slice_by_index_16)[name = tensor("reshape_81")]; + tensor reshape_82_shape_0 = const()[name = tensor("reshape_82_shape_0"), val = tensor([-1])]; + tensor reshape_82_cast_fp16 = reshape(shape = reshape_82_shape_0, x = var_3828_cast_fp16)[name = tensor("reshape_82_cast_fp16")]; + tensor reshape_83_shape_0 = const()[name = tensor("reshape_83_shape_0"), val = tensor([-1])]; + tensor reshape_83_cast_fp16 = reshape(shape = reshape_83_shape_0, x = cache8_to_fp16)[name = tensor("reshape_83_cast_fp16")]; + tensor scatter_16_mode_0 = const()[name = tensor("scatter_16_mode_0"), val = tensor("update")]; + tensor scatter_16_axis_0 = const()[name = tensor("scatter_16_axis_0"), val = tensor(0)]; + tensor scatter_16_validate_indices_0 = const()[name = tensor("scatter_16_validate_indices_0"), val = tensor(false)]; + tensor scatter_16_cast_fp16 = scatter(axis = scatter_16_axis_0, data = reshape_83_cast_fp16, indices = reshape_81, mode = scatter_16_mode_0, updates = reshape_82_cast_fp16, validate_indices = scatter_16_validate_indices_0)[name = tensor("scatter_16_cast_fp16")]; + tensor reshape_84_cast_fp16 = reshape(shape = shape_64, x = scatter_16_cast_fp16)[name = tensor("reshape_84_cast_fp16")]; + tensor var_3836_begin_0 = const()[name = tensor("op_3836_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor var_3836_end_0 = const()[name = tensor("op_3836_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor var_3836_end_mask_0 = const()[name = tensor("op_3836_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_3836_squeeze_mask_0 = const()[name = tensor("op_3836_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor var_3836_cast_fp16 = slice_by_index(begin = var_3836_begin_0, end = var_3836_end_0, end_mask = var_3836_end_mask_0, squeeze_mask = var_3836_squeeze_mask_0, x = reshape_84_cast_fp16)[name = tensor("op_3836_cast_fp16")]; + tensor var_3838_axis_0 = const()[name = tensor("op_3838_axis_0"), val = tensor(1)]; + tensor var_3838_mode_0 = const()[name = tensor("op_3838_mode_0"), val = tensor("update")]; + tensor var_3838_validate_indices_0 = const()[name = tensor("op_3838_validate_indices_0"), val = tensor(false)]; + tensor var_3838_cast_fp16 = scatter_along_axis(axis = var_3838_axis_0, data = var_3836_cast_fp16, indices = write_indices_17, mode = var_3838_mode_0, updates = v_17_cast_fp16, validate_indices = var_3838_validate_indices_0)[name = tensor("op_3838_cast_fp16")]; + tensor concat_60 = const()[name = tensor("concat_60"), val = tensor([1, 0, 0, 0, 0])]; + tensor concat_61 = const()[name = tensor("concat_61"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_17_internal_tensor_assign_2_stride_0 = const()[name = tensor("new_cache_17_internal_tensor_assign_2_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_17_internal_tensor_assign_2_begin_mask_0 = const()[name = tensor("new_cache_17_internal_tensor_assign_2_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_17_internal_tensor_assign_2_end_mask_0 = const()[name = tensor("new_cache_17_internal_tensor_assign_2_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_17_internal_tensor_assign_2_squeeze_mask_0 = const()[name = tensor("new_cache_17_internal_tensor_assign_2_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_65 = const()[name = tensor("shape_65"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_17 = const()[name = tensor("reduce_prod_17"), val = tensor(1048576)]; + tensor range_1d_17_start_0 = const()[name = tensor("range_1d_17_start_0"), val = tensor(0)]; + tensor range_1d_17_step_0 = const()[name = tensor("range_1d_17_step_0"), val = tensor(1)]; + tensor range_1d_17 = range_1d(end = reduce_prod_17, start = range_1d_17_start_0, step = range_1d_17_step_0)[name = tensor("range_1d_17")]; + tensor reshape_85 = reshape(shape = shape_65, x = range_1d_17)[name = tensor("reshape_85")]; + tensor slice_by_index_17 = slice_by_index(begin = concat_60, begin_mask = new_cache_17_internal_tensor_assign_2_begin_mask_0, end = concat_61, end_mask = new_cache_17_internal_tensor_assign_2_end_mask_0, squeeze_mask = new_cache_17_internal_tensor_assign_2_squeeze_mask_0, stride = new_cache_17_internal_tensor_assign_2_stride_0, x = reshape_85)[name = tensor("slice_by_index_17")]; + tensor reshape_86_shape_0 = const()[name = tensor("reshape_86_shape_0"), val = tensor([-1])]; + tensor reshape_86 = reshape(shape = reshape_86_shape_0, x = slice_by_index_17)[name = tensor("reshape_86")]; + tensor reshape_87_shape_0 = const()[name = tensor("reshape_87_shape_0"), val = tensor([-1])]; + tensor reshape_87_cast_fp16 = reshape(shape = reshape_87_shape_0, x = var_3838_cast_fp16)[name = tensor("reshape_87_cast_fp16")]; + tensor reshape_88_shape_0 = const()[name = tensor("reshape_88_shape_0"), val = tensor([-1])]; + tensor reshape_88_cast_fp16 = reshape(shape = reshape_88_shape_0, x = reshape_84_cast_fp16)[name = tensor("reshape_88_cast_fp16")]; + tensor scatter_17_mode_0 = const()[name = tensor("scatter_17_mode_0"), val = tensor("update")]; + tensor scatter_17_axis_0 = const()[name = tensor("scatter_17_axis_0"), val = tensor(0)]; + tensor scatter_17_validate_indices_0 = const()[name = tensor("scatter_17_validate_indices_0"), val = tensor(false)]; + tensor scatter_17_cast_fp16 = scatter(axis = scatter_17_axis_0, data = reshape_88_cast_fp16, indices = reshape_86, mode = scatter_17_mode_0, updates = reshape_87_cast_fp16, validate_indices = scatter_17_validate_indices_0)[name = tensor("scatter_17_cast_fp16")]; + tensor reshape_89_cast_fp16 = reshape(shape = shape_65, x = scatter_17_cast_fp16)[name = tensor("reshape_89_cast_fp16")]; + tensor reshape_89_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("reshape_89_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor keys_49_begin_0 = const()[name = tensor("keys_49_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor keys_49_end_0 = const()[name = tensor("keys_49_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor keys_49_end_mask_0 = const()[name = tensor("keys_49_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor keys_49_squeeze_mask_0 = const()[name = tensor("keys_49_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor keys_49_cast_fp16 = slice_by_index(begin = keys_49_begin_0, end = keys_49_end_0, end_mask = keys_49_end_mask_0, squeeze_mask = keys_49_squeeze_mask_0, x = reshape_89_cast_fp16)[name = tensor("keys_49_cast_fp16")]; + tensor values_49_begin_0 = const()[name = tensor("values_49_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor values_49_end_0 = const()[name = tensor("values_49_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor values_49_end_mask_0 = const()[name = tensor("values_49_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor values_49_squeeze_mask_0 = const()[name = tensor("values_49_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor values_49_cast_fp16 = slice_by_index(begin = values_49_begin_0, end = values_49_end_0, end_mask = values_49_end_mask_0, squeeze_mask = values_49_squeeze_mask_0, x = reshape_89_cast_fp16)[name = tensor("values_49_cast_fp16")]; + tensor var_3850_cast_fp16 = not_equal(x = keys_49_cast_fp16, y = keys_49_cast_fp16)[name = tensor("op_3850_cast_fp16")]; + tensor keys_51_cast_fp16 = select(a = var_504_to_fp16, b = keys_49_cast_fp16, cond = var_3850_cast_fp16)[name = tensor("keys_51_cast_fp16")]; + tensor var_3858_cast_fp16 = not_equal(x = values_49_cast_fp16, y = values_49_cast_fp16)[name = tensor("op_3858_cast_fp16")]; + tensor values_51_cast_fp16 = select(a = var_504_to_fp16, b = values_49_cast_fp16, cond = var_3858_cast_fp16)[name = tensor("values_51_cast_fp16")]; + tensor var_3882 = const()[name = tensor("op_3882"), val = tensor([0, 2, 1, 3])]; + tensor var_3895 = const()[name = tensor("op_3895"), val = tensor([1, 1, 1])]; + tensor var_3896_cast_fp16 = reshape(shape = var_3895, x = position8_to_fp16)[name = tensor("op_3896_cast_fp16")]; + tensor var_3913_to_fp16 = const()[name = tensor("op_3913_to_fp16"), val = tensor(0x1p+0)]; + tensor valid_len_17_cast_fp16 = add(x = var_3896_cast_fp16, y = var_3913_to_fp16)[name = tensor("valid_len_17_cast_fp16")]; + tensor valid_mask_17_cast_fp16 = less(x = k_positions_1_promoted_to_fp16, y = valid_len_17_cast_fp16)[name = tensor("valid_mask_17_cast_fp16")]; + tensor causal_mask_17_cast_fp16 = less_equal(x = k_positions_1_promoted_to_fp16, y = var_3896_cast_fp16)[name = tensor("causal_mask_17_cast_fp16")]; + tensor attn_mask_33 = logical_and(x = valid_mask_17_cast_fp16, y = causal_mask_17_cast_fp16)[name = tensor("attn_mask_33")]; + tensor attn_mask_35_axes_0 = const()[name = tensor("attn_mask_35_axes_0"), val = tensor([1])]; + tensor attn_mask_35 = expand_dims(axes = attn_mask_35_axes_0, x = attn_mask_33)[name = tensor("attn_mask_35")]; + tensor var_3931_transpose_x_0 = const()[name = tensor("op_3931_transpose_x_0"), val = tensor(false)]; + tensor var_3931_transpose_y_0 = const()[name = tensor("op_3931_transpose_y_0"), val = tensor(false)]; + tensor transpose_88_perm_0 = const()[name = tensor("transpose_88_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_89_perm_0 = const()[name = tensor("transpose_89_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_89 = transpose(perm = transpose_89_perm_0, x = keys_51_cast_fp16)[name = tensor("transpose_181")]; + tensor transpose_88 = transpose(perm = transpose_88_perm_0, x = q_51_cast_fp16)[name = tensor("transpose_182")]; + tensor var_3931_cast_fp16 = matmul(transpose_x = var_3931_transpose_x_0, transpose_y = var_3931_transpose_y_0, x = transpose_88, y = transpose_89)[name = tensor("op_3931_cast_fp16")]; + tensor var_3925_to_fp16 = const()[name = tensor("op_3925_to_fp16"), val = tensor([0x1p-3])]; + tensor attn_weights_49_cast_fp16 = mul(x = var_3931_cast_fp16, y = var_3925_to_fp16)[name = tensor("attn_weights_49_cast_fp16")]; + tensor var_3933 = logical_not(x = attn_mask_35)[name = tensor("op_3933")]; + tensor var_3934_to_fp16 = const()[name = tensor("op_3934_to_fp16"), val = tensor(-inf)]; + tensor attn_weights_51_cast_fp16 = select(a = var_3934_to_fp16, b = attn_weights_49_cast_fp16, cond = var_3933)[name = tensor("attn_weights_51_cast_fp16")]; + tensor var_3936 = const()[name = tensor("op_3936"), val = tensor(-1)]; + tensor attn_weights_53_cast_fp16 = softmax(axis = var_3936, x = attn_weights_51_cast_fp16)[name = tensor("attn_weights_53_cast_fp16")]; + tensor attn_output_17_transpose_x_0 = const()[name = tensor("attn_output_17_transpose_x_0"), val = tensor(false)]; + tensor attn_output_17_transpose_y_0 = const()[name = tensor("attn_output_17_transpose_y_0"), val = tensor(false)]; + tensor values_53_cast_fp16 = transpose(perm = var_3882, x = values_51_cast_fp16)[name = tensor("transpose_183")]; + tensor attn_output_17_cast_fp16 = matmul(transpose_x = attn_output_17_transpose_x_0, transpose_y = attn_output_17_transpose_y_0, x = attn_weights_53_cast_fp16, y = values_53_cast_fp16)[name = tensor("attn_output_17_cast_fp16")]; + tensor var_3944 = const()[name = tensor("op_3944"), val = tensor([0, 2, 1, 3])]; + tensor var_3947 = const()[name = tensor("op_3947"), val = tensor([1, 1, 1024])]; + tensor var_3945_cast_fp16 = transpose(perm = var_3944, x = attn_output_17_cast_fp16)[name = tensor("transpose_180")]; + tensor input_85_cast_fp16 = reshape(shape = var_3947, x = var_3945_cast_fp16)[name = tensor("input_85_cast_fp16")]; + tensor linear_34_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = attn8_out_proj_weight_cast_fp16, x = input_85_cast_fp16)[name = tensor("linear_34_cast_fp16")]; + tensor var_3953_to_fp16 = const()[name = tensor("op_3953_to_fp16"), val = tensor(0x1p+0)]; + tensor var_3954_cast_fp16 = add(x = position8_to_fp16, y = var_3953_to_fp16)[name = tensor("op_3954_cast_fp16")]; + tensor var_3954_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_3954_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor input_87_cast_fp16 = add(x = input_83_cast_fp16, y = linear_34_cast_fp16)[name = tensor("input_87_cast_fp16")]; + tensor input_89_axes_0 = const()[name = tensor("input_89_axes_0"), val = tensor([-1])]; + tensor norm8_2_weight_to_fp16 = const()[name = tensor("norm8_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(303657856)))]; + tensor norm8_2_bias_to_fp16 = const()[name = tensor("norm8_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(303659968)))]; + tensor var_3958_to_fp16 = const()[name = tensor("op_3958_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_89_cast_fp16 = layer_norm(axes = input_89_axes_0, beta = norm8_2_bias_to_fp16, epsilon = var_3958_to_fp16, gamma = norm8_2_weight_to_fp16, x = input_87_cast_fp16)[name = tensor("input_89_cast_fp16")]; + tensor linear_35_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = linear8_1_weight_cast_fp16, x = input_89_cast_fp16)[name = tensor("linear_35_cast_fp16")]; + tensor input_91_mode_0 = const()[name = tensor("input_91_mode_0"), val = tensor("EXACT")]; + tensor input_91_cast_fp16 = gelu(mode = input_91_mode_0, x = linear_35_cast_fp16)[name = tensor("input_91_cast_fp16")]; + tensor linear_36_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = linear8_2_weight_cast_fp16, x = input_91_cast_fp16)[name = tensor("linear_36_cast_fp16")]; + tensor input_93_cast_fp16 = add(x = input_87_cast_fp16, y = linear_36_cast_fp16)[name = tensor("input_93_cast_fp16")]; + tensor x_19_axes_0 = const()[name = tensor("x_19_axes_0"), val = tensor([-1])]; + tensor norm9_1_weight_to_fp16 = const()[name = tensor("norm9_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(303662080)))]; + tensor norm9_1_bias_to_fp16 = const()[name = tensor("norm9_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(303664192)))]; + tensor var_3975_to_fp16 = const()[name = tensor("op_3975_to_fp16"), val = tensor(0x1.5p-17)]; + tensor x_19_cast_fp16 = layer_norm(axes = x_19_axes_0, beta = norm9_1_bias_to_fp16, epsilon = var_3975_to_fp16, gamma = norm9_1_weight_to_fp16, x = input_93_cast_fp16)[name = tensor("x_19_cast_fp16")]; + tensor linear_37_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = attn9_in_proj_weight_cast_fp16, x = x_19_cast_fp16)[name = tensor("linear_37_cast_fp16")]; + tensor var_4011 = const()[name = tensor("op_4011"), val = tensor([1, 1, 3, 16, 64])]; + tensor qkv_19_cast_fp16 = reshape(shape = var_4011, x = linear_37_cast_fp16)[name = tensor("qkv_19_cast_fp16")]; + tensor q_55_begin_0 = const()[name = tensor("q_55_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor q_55_end_0 = const()[name = tensor("q_55_end_0"), val = tensor([1, 1, 1, 16, 64])]; + tensor q_55_end_mask_0 = const()[name = tensor("q_55_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor q_55_squeeze_mask_0 = const()[name = tensor("q_55_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor q_55_cast_fp16 = slice_by_index(begin = q_55_begin_0, end = q_55_end_0, end_mask = q_55_end_mask_0, squeeze_mask = q_55_squeeze_mask_0, x = qkv_19_cast_fp16)[name = tensor("q_55_cast_fp16")]; + tensor k_37_begin_0 = const()[name = tensor("k_37_begin_0"), val = tensor([0, 0, 1, 0, 0])]; + tensor k_37_end_0 = const()[name = tensor("k_37_end_0"), val = tensor([1, 1, 2, 16, 64])]; + tensor k_37_end_mask_0 = const()[name = tensor("k_37_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor k_37_squeeze_mask_0 = const()[name = tensor("k_37_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor k_37_cast_fp16 = slice_by_index(begin = k_37_begin_0, end = k_37_end_0, end_mask = k_37_end_mask_0, squeeze_mask = k_37_squeeze_mask_0, x = qkv_19_cast_fp16)[name = tensor("k_37_cast_fp16")]; + tensor v_19_begin_0 = const()[name = tensor("v_19_begin_0"), val = tensor([0, 0, 2, 0, 0])]; + tensor v_19_end_0 = const()[name = tensor("v_19_end_0"), val = tensor([1, 1, 3, 16, 64])]; + tensor v_19_end_mask_0 = const()[name = tensor("v_19_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor v_19_squeeze_mask_0 = const()[name = tensor("v_19_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor v_19_cast_fp16 = slice_by_index(begin = v_19_begin_0, end = v_19_end_0, end_mask = v_19_end_mask_0, squeeze_mask = v_19_squeeze_mask_0, x = qkv_19_cast_fp16)[name = tensor("v_19_cast_fp16")]; + tensor var_4115 = const()[name = tensor("op_4115"), val = tensor([1, 1, 1, 1])]; + tensor position9_to_fp16_dtype_0 = const()[name = tensor("position9_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor position9_to_fp16 = cast(dtype = position9_to_fp16_dtype_0, x = position9)[name = tensor("cast_508")]; + tensor ts_59_cast_fp16 = reshape(shape = var_4115, x = position9_to_fp16)[name = tensor("ts_59_cast_fp16")]; + tensor var_4119 = const()[name = tensor("op_4119"), val = tensor([1, 1, 16, 32, 2])]; + tensor q_complex_19_cast_fp16 = reshape(shape = var_4119, x = q_55_cast_fp16)[name = tensor("q_complex_19_cast_fp16")]; + tensor var_4123 = const()[name = tensor("op_4123"), val = tensor([1, 1, 16, 32, 2])]; + tensor k_complex_19_cast_fp16 = reshape(shape = var_4123, x = k_37_cast_fp16)[name = tensor("k_complex_19_cast_fp16")]; + tensor var_4127_begin_0 = const()[name = tensor("op_4127_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_4127_end_0 = const()[name = tensor("op_4127_end_0"), val = tensor([1, 1, 16, 32, 1])]; + tensor var_4127_end_mask_0 = const()[name = tensor("op_4127_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_4127_squeeze_mask_0 = const()[name = tensor("op_4127_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_4127_cast_fp16 = slice_by_index(begin = var_4127_begin_0, end = var_4127_end_0, end_mask = var_4127_end_mask_0, squeeze_mask = var_4127_squeeze_mask_0, x = q_complex_19_cast_fp16)[name = tensor("op_4127_cast_fp16")]; + tensor var_4135_begin_0 = const()[name = tensor("op_4135_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_4135_end_0 = const()[name = tensor("op_4135_end_0"), val = tensor([1, 1, 16, 32, 2])]; + tensor var_4135_end_mask_0 = const()[name = tensor("op_4135_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_4135_squeeze_mask_0 = const()[name = tensor("op_4135_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_4135_cast_fp16 = slice_by_index(begin = var_4135_begin_0, end = var_4135_end_0, end_mask = var_4135_end_mask_0, squeeze_mask = var_4135_squeeze_mask_0, x = q_complex_19_cast_fp16)[name = tensor("op_4135_cast_fp16")]; + tensor var_4143_begin_0 = const()[name = tensor("op_4143_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_4143_end_0 = const()[name = tensor("op_4143_end_0"), val = tensor([1, 1, 16, 32, 1])]; + tensor var_4143_end_mask_0 = const()[name = tensor("op_4143_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_4143_squeeze_mask_0 = const()[name = tensor("op_4143_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_4143_cast_fp16 = slice_by_index(begin = var_4143_begin_0, end = var_4143_end_0, end_mask = var_4143_end_mask_0, squeeze_mask = var_4143_squeeze_mask_0, x = k_complex_19_cast_fp16)[name = tensor("op_4143_cast_fp16")]; + tensor var_4151_begin_0 = const()[name = tensor("op_4151_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_4151_end_0 = const()[name = tensor("op_4151_end_0"), val = tensor([1, 1, 16, 32, 2])]; + tensor var_4151_end_mask_0 = const()[name = tensor("op_4151_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_4151_squeeze_mask_0 = const()[name = tensor("op_4151_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_4151_cast_fp16 = slice_by_index(begin = var_4151_begin_0, end = var_4151_end_0, end_mask = var_4151_end_mask_0, squeeze_mask = var_4151_squeeze_mask_0, x = k_complex_19_cast_fp16)[name = tensor("op_4151_cast_fp16")]; + tensor freqs_19_to_fp16 = const()[name = tensor("freqs_19_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(303666304)))]; + tensor var_4157_cast_fp16 = mul(x = freqs_19_to_fp16, y = ts_59_cast_fp16)[name = tensor("op_4157_cast_fp16")]; + tensor rotr_19_cast_fp16 = cos(x = var_4157_cast_fp16)[name = tensor("rotr_19_cast_fp16")]; + tensor roti_19_cast_fp16 = sin(x = var_4157_cast_fp16)[name = tensor("roti_19_cast_fp16")]; + tensor var_4161_cast_fp16 = mul(x = var_4127_cast_fp16, y = rotr_19_cast_fp16)[name = tensor("op_4161_cast_fp16")]; + tensor var_4162_cast_fp16 = mul(x = var_4135_cast_fp16, y = roti_19_cast_fp16)[name = tensor("op_4162_cast_fp16")]; + tensor qor_37_cast_fp16 = sub(x = var_4161_cast_fp16, y = var_4162_cast_fp16)[name = tensor("qor_37_cast_fp16")]; + tensor var_4165_cast_fp16 = mul(x = var_4127_cast_fp16, y = roti_19_cast_fp16)[name = tensor("op_4165_cast_fp16")]; + tensor var_4166_cast_fp16 = mul(x = var_4135_cast_fp16, y = rotr_19_cast_fp16)[name = tensor("op_4166_cast_fp16")]; + tensor qoi_37_cast_fp16 = add(x = var_4165_cast_fp16, y = var_4166_cast_fp16)[name = tensor("qoi_37_cast_fp16")]; + tensor var_4169_cast_fp16 = mul(x = var_4143_cast_fp16, y = rotr_19_cast_fp16)[name = tensor("op_4169_cast_fp16")]; + tensor var_4170_cast_fp16 = mul(x = var_4151_cast_fp16, y = roti_19_cast_fp16)[name = tensor("op_4170_cast_fp16")]; + tensor kor_37_cast_fp16 = sub(x = var_4169_cast_fp16, y = var_4170_cast_fp16)[name = tensor("kor_37_cast_fp16")]; + tensor var_4173_cast_fp16 = mul(x = var_4143_cast_fp16, y = roti_19_cast_fp16)[name = tensor("op_4173_cast_fp16")]; + tensor var_4174_cast_fp16 = mul(x = var_4151_cast_fp16, y = rotr_19_cast_fp16)[name = tensor("op_4174_cast_fp16")]; + tensor koi_37_cast_fp16 = add(x = var_4173_cast_fp16, y = var_4174_cast_fp16)[name = tensor("koi_37_cast_fp16")]; + tensor qo_19_axis_0 = const()[name = tensor("qo_19_axis_0"), val = tensor(-1)]; + tensor qo_19_cast_fp16 = stack(axis = qo_19_axis_0, values = (qor_37_cast_fp16, qoi_37_cast_fp16))[name = tensor("qo_19_cast_fp16")]; + tensor ko_19_axis_0 = const()[name = tensor("ko_19_axis_0"), val = tensor(-1)]; + tensor ko_19_cast_fp16 = stack(axis = ko_19_axis_0, values = (kor_37_cast_fp16, koi_37_cast_fp16))[name = tensor("ko_19_cast_fp16")]; + tensor var_4203 = const()[name = tensor("op_4203"), val = tensor([1, 1, 16, 64])]; + tensor q_57_cast_fp16 = reshape(shape = var_4203, x = qo_19_cast_fp16)[name = tensor("q_57_cast_fp16")]; + tensor var_4205 = const()[name = tensor("op_4205"), val = tensor([1, 1, 16, 64])]; + tensor k_39_cast_fp16 = reshape(shape = var_4205, x = ko_19_cast_fp16)[name = tensor("k_39_cast_fp16")]; + tensor _inversed_4227_y_0_to_fp16 = const()[name = tensor("_inversed_4227_y_0_to_fp16"), val = tensor(0x1p-9)]; + tensor _inversed_4227_cast_fp16 = mul(x = ts_59_cast_fp16, y = _inversed_4227_y_0_to_fp16)[name = tensor("_inversed_4227_cast_fp16")]; + tensor var_4228_cast_fp16 = floor(x = _inversed_4227_cast_fp16)[name = tensor("op_4228_cast_fp16")]; + tensor var_4229_to_fp16 = const()[name = tensor("op_4229_to_fp16"), val = tensor(0x1p+9)]; + tensor var_4230_cast_fp16 = mul(x = var_4228_cast_fp16, y = var_4229_to_fp16)[name = tensor("op_4230_cast_fp16")]; + tensor write_indices_float_39_cast_fp16 = sub(x = ts_59_cast_fp16, y = var_4230_cast_fp16)[name = tensor("write_indices_float_39_cast_fp16")]; + tensor var_4237_dtype_0 = const()[name = tensor("op_4237_dtype_0"), val = tensor("int32")]; + tensor write_indices_19_reps_0 = const()[name = tensor("write_indices_19_reps_0"), val = tensor([1, 1, 16, 64])]; + tensor write_indices_float_39_cast_fp16_to_int32 = cast(dtype = var_4237_dtype_0, x = write_indices_float_39_cast_fp16)[name = tensor("cast_507")]; + tensor write_indices_19 = tile(reps = write_indices_19_reps_0, x = write_indices_float_39_cast_fp16_to_int32)[name = tensor("write_indices_19")]; + tensor var_4245_begin_0 = const()[name = tensor("op_4245_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_4245_end_0 = const()[name = tensor("op_4245_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor var_4245_end_mask_0 = const()[name = tensor("op_4245_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_4245_squeeze_mask_0 = const()[name = tensor("op_4245_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor cache9_to_fp16_dtype_0 = const()[name = tensor("cache9_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor cache9_to_fp16 = cast(dtype = cache9_to_fp16_dtype_0, x = cache9)[name = tensor("cast_506")]; + tensor var_4245_cast_fp16 = slice_by_index(begin = var_4245_begin_0, end = var_4245_end_0, end_mask = var_4245_end_mask_0, squeeze_mask = var_4245_squeeze_mask_0, x = cache9_to_fp16)[name = tensor("op_4245_cast_fp16")]; + tensor var_4247_axis_0 = const()[name = tensor("op_4247_axis_0"), val = tensor(1)]; + tensor var_4247_mode_0 = const()[name = tensor("op_4247_mode_0"), val = tensor("update")]; + tensor var_4247_validate_indices_0 = const()[name = tensor("op_4247_validate_indices_0"), val = tensor(false)]; + tensor var_4247_cast_fp16 = scatter_along_axis(axis = var_4247_axis_0, data = var_4245_cast_fp16, indices = write_indices_19, mode = var_4247_mode_0, updates = k_39_cast_fp16, validate_indices = var_4247_validate_indices_0)[name = tensor("op_4247_cast_fp16")]; + tensor concat_65 = const()[name = tensor("concat_65"), val = tensor([0, 0, 0, 0, 0])]; + tensor concat_66 = const()[name = tensor("concat_66"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_19_internal_tensor_assign_1_stride_0 = const()[name = tensor("new_cache_19_internal_tensor_assign_1_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_19_internal_tensor_assign_1_begin_mask_0 = const()[name = tensor("new_cache_19_internal_tensor_assign_1_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_19_internal_tensor_assign_1_end_mask_0 = const()[name = tensor("new_cache_19_internal_tensor_assign_1_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_19_internal_tensor_assign_1_squeeze_mask_0 = const()[name = tensor("new_cache_19_internal_tensor_assign_1_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_66 = const()[name = tensor("shape_66"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_18 = const()[name = tensor("reduce_prod_18"), val = tensor(1048576)]; + tensor range_1d_18_start_0 = const()[name = tensor("range_1d_18_start_0"), val = tensor(0)]; + tensor range_1d_18_step_0 = const()[name = tensor("range_1d_18_step_0"), val = tensor(1)]; + tensor range_1d_18 = range_1d(end = reduce_prod_18, start = range_1d_18_start_0, step = range_1d_18_step_0)[name = tensor("range_1d_18")]; + tensor reshape_90 = reshape(shape = shape_66, x = range_1d_18)[name = tensor("reshape_90")]; + tensor slice_by_index_18 = slice_by_index(begin = concat_65, begin_mask = new_cache_19_internal_tensor_assign_1_begin_mask_0, end = concat_66, end_mask = new_cache_19_internal_tensor_assign_1_end_mask_0, squeeze_mask = new_cache_19_internal_tensor_assign_1_squeeze_mask_0, stride = new_cache_19_internal_tensor_assign_1_stride_0, x = reshape_90)[name = tensor("slice_by_index_18")]; + tensor reshape_91_shape_0 = const()[name = tensor("reshape_91_shape_0"), val = tensor([-1])]; + tensor reshape_91 = reshape(shape = reshape_91_shape_0, x = slice_by_index_18)[name = tensor("reshape_91")]; + tensor reshape_92_shape_0 = const()[name = tensor("reshape_92_shape_0"), val = tensor([-1])]; + tensor reshape_92_cast_fp16 = reshape(shape = reshape_92_shape_0, x = var_4247_cast_fp16)[name = tensor("reshape_92_cast_fp16")]; + tensor reshape_93_shape_0 = const()[name = tensor("reshape_93_shape_0"), val = tensor([-1])]; + tensor reshape_93_cast_fp16 = reshape(shape = reshape_93_shape_0, x = cache9_to_fp16)[name = tensor("reshape_93_cast_fp16")]; + tensor scatter_18_mode_0 = const()[name = tensor("scatter_18_mode_0"), val = tensor("update")]; + tensor scatter_18_axis_0 = const()[name = tensor("scatter_18_axis_0"), val = tensor(0)]; + tensor scatter_18_validate_indices_0 = const()[name = tensor("scatter_18_validate_indices_0"), val = tensor(false)]; + tensor scatter_18_cast_fp16 = scatter(axis = scatter_18_axis_0, data = reshape_93_cast_fp16, indices = reshape_91, mode = scatter_18_mode_0, updates = reshape_92_cast_fp16, validate_indices = scatter_18_validate_indices_0)[name = tensor("scatter_18_cast_fp16")]; + tensor reshape_94_cast_fp16 = reshape(shape = shape_66, x = scatter_18_cast_fp16)[name = tensor("reshape_94_cast_fp16")]; + tensor var_4255_begin_0 = const()[name = tensor("op_4255_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor var_4255_end_0 = const()[name = tensor("op_4255_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor var_4255_end_mask_0 = const()[name = tensor("op_4255_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_4255_squeeze_mask_0 = const()[name = tensor("op_4255_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor var_4255_cast_fp16 = slice_by_index(begin = var_4255_begin_0, end = var_4255_end_0, end_mask = var_4255_end_mask_0, squeeze_mask = var_4255_squeeze_mask_0, x = reshape_94_cast_fp16)[name = tensor("op_4255_cast_fp16")]; + tensor var_4257_axis_0 = const()[name = tensor("op_4257_axis_0"), val = tensor(1)]; + tensor var_4257_mode_0 = const()[name = tensor("op_4257_mode_0"), val = tensor("update")]; + tensor var_4257_validate_indices_0 = const()[name = tensor("op_4257_validate_indices_0"), val = tensor(false)]; + tensor var_4257_cast_fp16 = scatter_along_axis(axis = var_4257_axis_0, data = var_4255_cast_fp16, indices = write_indices_19, mode = var_4257_mode_0, updates = v_19_cast_fp16, validate_indices = var_4257_validate_indices_0)[name = tensor("op_4257_cast_fp16")]; + tensor concat_67 = const()[name = tensor("concat_67"), val = tensor([1, 0, 0, 0, 0])]; + tensor concat_68 = const()[name = tensor("concat_68"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_19_internal_tensor_assign_2_stride_0 = const()[name = tensor("new_cache_19_internal_tensor_assign_2_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_19_internal_tensor_assign_2_begin_mask_0 = const()[name = tensor("new_cache_19_internal_tensor_assign_2_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_19_internal_tensor_assign_2_end_mask_0 = const()[name = tensor("new_cache_19_internal_tensor_assign_2_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_19_internal_tensor_assign_2_squeeze_mask_0 = const()[name = tensor("new_cache_19_internal_tensor_assign_2_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_67 = const()[name = tensor("shape_67"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_19 = const()[name = tensor("reduce_prod_19"), val = tensor(1048576)]; + tensor range_1d_19_start_0 = const()[name = tensor("range_1d_19_start_0"), val = tensor(0)]; + tensor range_1d_19_step_0 = const()[name = tensor("range_1d_19_step_0"), val = tensor(1)]; + tensor range_1d_19 = range_1d(end = reduce_prod_19, start = range_1d_19_start_0, step = range_1d_19_step_0)[name = tensor("range_1d_19")]; + tensor reshape_95 = reshape(shape = shape_67, x = range_1d_19)[name = tensor("reshape_95")]; + tensor slice_by_index_19 = slice_by_index(begin = concat_67, begin_mask = new_cache_19_internal_tensor_assign_2_begin_mask_0, end = concat_68, end_mask = new_cache_19_internal_tensor_assign_2_end_mask_0, squeeze_mask = new_cache_19_internal_tensor_assign_2_squeeze_mask_0, stride = new_cache_19_internal_tensor_assign_2_stride_0, x = reshape_95)[name = tensor("slice_by_index_19")]; + tensor reshape_96_shape_0 = const()[name = tensor("reshape_96_shape_0"), val = tensor([-1])]; + tensor reshape_96 = reshape(shape = reshape_96_shape_0, x = slice_by_index_19)[name = tensor("reshape_96")]; + tensor reshape_97_shape_0 = const()[name = tensor("reshape_97_shape_0"), val = tensor([-1])]; + tensor reshape_97_cast_fp16 = reshape(shape = reshape_97_shape_0, x = var_4257_cast_fp16)[name = tensor("reshape_97_cast_fp16")]; + tensor reshape_98_shape_0 = const()[name = tensor("reshape_98_shape_0"), val = tensor([-1])]; + tensor reshape_98_cast_fp16 = reshape(shape = reshape_98_shape_0, x = reshape_94_cast_fp16)[name = tensor("reshape_98_cast_fp16")]; + tensor scatter_19_mode_0 = const()[name = tensor("scatter_19_mode_0"), val = tensor("update")]; + tensor scatter_19_axis_0 = const()[name = tensor("scatter_19_axis_0"), val = tensor(0)]; + tensor scatter_19_validate_indices_0 = const()[name = tensor("scatter_19_validate_indices_0"), val = tensor(false)]; + tensor scatter_19_cast_fp16 = scatter(axis = scatter_19_axis_0, data = reshape_98_cast_fp16, indices = reshape_96, mode = scatter_19_mode_0, updates = reshape_97_cast_fp16, validate_indices = scatter_19_validate_indices_0)[name = tensor("scatter_19_cast_fp16")]; + tensor reshape_99_cast_fp16 = reshape(shape = shape_67, x = scatter_19_cast_fp16)[name = tensor("reshape_99_cast_fp16")]; + tensor reshape_99_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("reshape_99_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor keys_55_begin_0 = const()[name = tensor("keys_55_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor keys_55_end_0 = const()[name = tensor("keys_55_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor keys_55_end_mask_0 = const()[name = tensor("keys_55_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor keys_55_squeeze_mask_0 = const()[name = tensor("keys_55_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor keys_55_cast_fp16 = slice_by_index(begin = keys_55_begin_0, end = keys_55_end_0, end_mask = keys_55_end_mask_0, squeeze_mask = keys_55_squeeze_mask_0, x = reshape_99_cast_fp16)[name = tensor("keys_55_cast_fp16")]; + tensor values_55_begin_0 = const()[name = tensor("values_55_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor values_55_end_0 = const()[name = tensor("values_55_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor values_55_end_mask_0 = const()[name = tensor("values_55_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor values_55_squeeze_mask_0 = const()[name = tensor("values_55_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor values_55_cast_fp16 = slice_by_index(begin = values_55_begin_0, end = values_55_end_0, end_mask = values_55_end_mask_0, squeeze_mask = values_55_squeeze_mask_0, x = reshape_99_cast_fp16)[name = tensor("values_55_cast_fp16")]; + tensor var_4269_cast_fp16 = not_equal(x = keys_55_cast_fp16, y = keys_55_cast_fp16)[name = tensor("op_4269_cast_fp16")]; + tensor keys_57_cast_fp16 = select(a = var_504_to_fp16, b = keys_55_cast_fp16, cond = var_4269_cast_fp16)[name = tensor("keys_57_cast_fp16")]; + tensor var_4277_cast_fp16 = not_equal(x = values_55_cast_fp16, y = values_55_cast_fp16)[name = tensor("op_4277_cast_fp16")]; + tensor values_57_cast_fp16 = select(a = var_504_to_fp16, b = values_55_cast_fp16, cond = var_4277_cast_fp16)[name = tensor("values_57_cast_fp16")]; + tensor var_4301 = const()[name = tensor("op_4301"), val = tensor([0, 2, 1, 3])]; + tensor var_4314 = const()[name = tensor("op_4314"), val = tensor([1, 1, 1])]; + tensor var_4315_cast_fp16 = reshape(shape = var_4314, x = position9_to_fp16)[name = tensor("op_4315_cast_fp16")]; + tensor var_4332_to_fp16 = const()[name = tensor("op_4332_to_fp16"), val = tensor(0x1p+0)]; + tensor valid_len_19_cast_fp16 = add(x = var_4315_cast_fp16, y = var_4332_to_fp16)[name = tensor("valid_len_19_cast_fp16")]; + tensor valid_mask_19_cast_fp16 = less(x = k_positions_1_promoted_to_fp16, y = valid_len_19_cast_fp16)[name = tensor("valid_mask_19_cast_fp16")]; + tensor causal_mask_19_cast_fp16 = less_equal(x = k_positions_1_promoted_to_fp16, y = var_4315_cast_fp16)[name = tensor("causal_mask_19_cast_fp16")]; + tensor attn_mask_37 = logical_and(x = valid_mask_19_cast_fp16, y = causal_mask_19_cast_fp16)[name = tensor("attn_mask_37")]; + tensor attn_mask_39_axes_0 = const()[name = tensor("attn_mask_39_axes_0"), val = tensor([1])]; + tensor attn_mask_39 = expand_dims(axes = attn_mask_39_axes_0, x = attn_mask_37)[name = tensor("attn_mask_39")]; + tensor var_4350_transpose_x_0 = const()[name = tensor("op_4350_transpose_x_0"), val = tensor(false)]; + tensor var_4350_transpose_y_0 = const()[name = tensor("op_4350_transpose_y_0"), val = tensor(false)]; + tensor transpose_90_perm_0 = const()[name = tensor("transpose_90_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_91_perm_0 = const()[name = tensor("transpose_91_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_91 = transpose(perm = transpose_91_perm_0, x = keys_57_cast_fp16)[name = tensor("transpose_177")]; + tensor transpose_90 = transpose(perm = transpose_90_perm_0, x = q_57_cast_fp16)[name = tensor("transpose_178")]; + tensor var_4350_cast_fp16 = matmul(transpose_x = var_4350_transpose_x_0, transpose_y = var_4350_transpose_y_0, x = transpose_90, y = transpose_91)[name = tensor("op_4350_cast_fp16")]; + tensor var_4344_to_fp16 = const()[name = tensor("op_4344_to_fp16"), val = tensor([0x1p-3])]; + tensor attn_weights_55_cast_fp16 = mul(x = var_4350_cast_fp16, y = var_4344_to_fp16)[name = tensor("attn_weights_55_cast_fp16")]; + tensor var_4352 = logical_not(x = attn_mask_39)[name = tensor("op_4352")]; + tensor var_4353_to_fp16 = const()[name = tensor("op_4353_to_fp16"), val = tensor(-inf)]; + tensor attn_weights_57_cast_fp16 = select(a = var_4353_to_fp16, b = attn_weights_55_cast_fp16, cond = var_4352)[name = tensor("attn_weights_57_cast_fp16")]; + tensor var_4355 = const()[name = tensor("op_4355"), val = tensor(-1)]; + tensor attn_weights_59_cast_fp16 = softmax(axis = var_4355, x = attn_weights_57_cast_fp16)[name = tensor("attn_weights_59_cast_fp16")]; + tensor attn_output_19_transpose_x_0 = const()[name = tensor("attn_output_19_transpose_x_0"), val = tensor(false)]; + tensor attn_output_19_transpose_y_0 = const()[name = tensor("attn_output_19_transpose_y_0"), val = tensor(false)]; + tensor values_59_cast_fp16 = transpose(perm = var_4301, x = values_57_cast_fp16)[name = tensor("transpose_179")]; + tensor attn_output_19_cast_fp16 = matmul(transpose_x = attn_output_19_transpose_x_0, transpose_y = attn_output_19_transpose_y_0, x = attn_weights_59_cast_fp16, y = values_59_cast_fp16)[name = tensor("attn_output_19_cast_fp16")]; + tensor var_4363 = const()[name = tensor("op_4363"), val = tensor([0, 2, 1, 3])]; + tensor var_4366 = const()[name = tensor("op_4366"), val = tensor([1, 1, 1024])]; + tensor var_4364_cast_fp16 = transpose(perm = var_4363, x = attn_output_19_cast_fp16)[name = tensor("transpose_176")]; + tensor input_95_cast_fp16 = reshape(shape = var_4366, x = var_4364_cast_fp16)[name = tensor("input_95_cast_fp16")]; + tensor linear_38_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = attn9_out_proj_weight_cast_fp16, x = input_95_cast_fp16)[name = tensor("linear_38_cast_fp16")]; + tensor var_4372_to_fp16 = const()[name = tensor("op_4372_to_fp16"), val = tensor(0x1p+0)]; + tensor var_4373_cast_fp16 = add(x = position9_to_fp16, y = var_4372_to_fp16)[name = tensor("op_4373_cast_fp16")]; + tensor var_4373_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_4373_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor input_97_cast_fp16 = add(x = input_93_cast_fp16, y = linear_38_cast_fp16)[name = tensor("input_97_cast_fp16")]; + tensor input_99_axes_0 = const()[name = tensor("input_99_axes_0"), val = tensor([-1])]; + tensor norm9_2_weight_to_fp16 = const()[name = tensor("norm9_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(303666432)))]; + tensor norm9_2_bias_to_fp16 = const()[name = tensor("norm9_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(303668544)))]; + tensor var_4377_to_fp16 = const()[name = tensor("op_4377_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_99_cast_fp16 = layer_norm(axes = input_99_axes_0, beta = norm9_2_bias_to_fp16, epsilon = var_4377_to_fp16, gamma = norm9_2_weight_to_fp16, x = input_97_cast_fp16)[name = tensor("input_99_cast_fp16")]; + tensor linear_39_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = linear9_1_weight_cast_fp16, x = input_99_cast_fp16)[name = tensor("linear_39_cast_fp16")]; + tensor input_101_mode_0 = const()[name = tensor("input_101_mode_0"), val = tensor("EXACT")]; + tensor input_101_cast_fp16 = gelu(mode = input_101_mode_0, x = linear_39_cast_fp16)[name = tensor("input_101_cast_fp16")]; + tensor linear_40_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = linear9_2_weight_cast_fp16, x = input_101_cast_fp16)[name = tensor("linear_40_cast_fp16")]; + tensor input_103_cast_fp16 = add(x = input_97_cast_fp16, y = linear_40_cast_fp16)[name = tensor("input_103_cast_fp16")]; + tensor x_21_axes_0 = const()[name = tensor("x_21_axes_0"), val = tensor([-1])]; + tensor norm10_1_weight_to_fp16 = const()[name = tensor("norm10_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(303670656)))]; + tensor norm10_1_bias_to_fp16 = const()[name = tensor("norm10_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(303672768)))]; + tensor var_4394_to_fp16 = const()[name = tensor("op_4394_to_fp16"), val = tensor(0x1.5p-17)]; + tensor x_21_cast_fp16 = layer_norm(axes = x_21_axes_0, beta = norm10_1_bias_to_fp16, epsilon = var_4394_to_fp16, gamma = norm10_1_weight_to_fp16, x = input_103_cast_fp16)[name = tensor("x_21_cast_fp16")]; + tensor linear_41_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = attn10_in_proj_weight_cast_fp16, x = x_21_cast_fp16)[name = tensor("linear_41_cast_fp16")]; + tensor var_4430 = const()[name = tensor("op_4430"), val = tensor([1, 1, 3, 16, 64])]; + tensor qkv_21_cast_fp16 = reshape(shape = var_4430, x = linear_41_cast_fp16)[name = tensor("qkv_21_cast_fp16")]; + tensor q_61_begin_0 = const()[name = tensor("q_61_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor q_61_end_0 = const()[name = tensor("q_61_end_0"), val = tensor([1, 1, 1, 16, 64])]; + tensor q_61_end_mask_0 = const()[name = tensor("q_61_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor q_61_squeeze_mask_0 = const()[name = tensor("q_61_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor q_61_cast_fp16 = slice_by_index(begin = q_61_begin_0, end = q_61_end_0, end_mask = q_61_end_mask_0, squeeze_mask = q_61_squeeze_mask_0, x = qkv_21_cast_fp16)[name = tensor("q_61_cast_fp16")]; + tensor k_41_begin_0 = const()[name = tensor("k_41_begin_0"), val = tensor([0, 0, 1, 0, 0])]; + tensor k_41_end_0 = const()[name = tensor("k_41_end_0"), val = tensor([1, 1, 2, 16, 64])]; + tensor k_41_end_mask_0 = const()[name = tensor("k_41_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor k_41_squeeze_mask_0 = const()[name = tensor("k_41_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor k_41_cast_fp16 = slice_by_index(begin = k_41_begin_0, end = k_41_end_0, end_mask = k_41_end_mask_0, squeeze_mask = k_41_squeeze_mask_0, x = qkv_21_cast_fp16)[name = tensor("k_41_cast_fp16")]; + tensor v_21_begin_0 = const()[name = tensor("v_21_begin_0"), val = tensor([0, 0, 2, 0, 0])]; + tensor v_21_end_0 = const()[name = tensor("v_21_end_0"), val = tensor([1, 1, 3, 16, 64])]; + tensor v_21_end_mask_0 = const()[name = tensor("v_21_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor v_21_squeeze_mask_0 = const()[name = tensor("v_21_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor v_21_cast_fp16 = slice_by_index(begin = v_21_begin_0, end = v_21_end_0, end_mask = v_21_end_mask_0, squeeze_mask = v_21_squeeze_mask_0, x = qkv_21_cast_fp16)[name = tensor("v_21_cast_fp16")]; + tensor var_4534 = const()[name = tensor("op_4534"), val = tensor([1, 1, 1, 1])]; + tensor position10_to_fp16_dtype_0 = const()[name = tensor("position10_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor position10_to_fp16 = cast(dtype = position10_to_fp16_dtype_0, x = position10)[name = tensor("cast_503")]; + tensor ts_65_cast_fp16 = reshape(shape = var_4534, x = position10_to_fp16)[name = tensor("ts_65_cast_fp16")]; + tensor var_4538 = const()[name = tensor("op_4538"), val = tensor([1, 1, 16, 32, 2])]; + tensor q_complex_21_cast_fp16 = reshape(shape = var_4538, x = q_61_cast_fp16)[name = tensor("q_complex_21_cast_fp16")]; + tensor var_4542 = const()[name = tensor("op_4542"), val = tensor([1, 1, 16, 32, 2])]; + tensor k_complex_21_cast_fp16 = reshape(shape = var_4542, x = k_41_cast_fp16)[name = tensor("k_complex_21_cast_fp16")]; + tensor var_4546_begin_0 = const()[name = tensor("op_4546_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_4546_end_0 = const()[name = tensor("op_4546_end_0"), val = tensor([1, 1, 16, 32, 1])]; + tensor var_4546_end_mask_0 = const()[name = tensor("op_4546_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_4546_squeeze_mask_0 = const()[name = tensor("op_4546_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_4546_cast_fp16 = slice_by_index(begin = var_4546_begin_0, end = var_4546_end_0, end_mask = var_4546_end_mask_0, squeeze_mask = var_4546_squeeze_mask_0, x = q_complex_21_cast_fp16)[name = tensor("op_4546_cast_fp16")]; + tensor var_4554_begin_0 = const()[name = tensor("op_4554_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_4554_end_0 = const()[name = tensor("op_4554_end_0"), val = tensor([1, 1, 16, 32, 2])]; + tensor var_4554_end_mask_0 = const()[name = tensor("op_4554_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_4554_squeeze_mask_0 = const()[name = tensor("op_4554_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_4554_cast_fp16 = slice_by_index(begin = var_4554_begin_0, end = var_4554_end_0, end_mask = var_4554_end_mask_0, squeeze_mask = var_4554_squeeze_mask_0, x = q_complex_21_cast_fp16)[name = tensor("op_4554_cast_fp16")]; + tensor var_4562_begin_0 = const()[name = tensor("op_4562_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_4562_end_0 = const()[name = tensor("op_4562_end_0"), val = tensor([1, 1, 16, 32, 1])]; + tensor var_4562_end_mask_0 = const()[name = tensor("op_4562_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_4562_squeeze_mask_0 = const()[name = tensor("op_4562_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_4562_cast_fp16 = slice_by_index(begin = var_4562_begin_0, end = var_4562_end_0, end_mask = var_4562_end_mask_0, squeeze_mask = var_4562_squeeze_mask_0, x = k_complex_21_cast_fp16)[name = tensor("op_4562_cast_fp16")]; + tensor var_4570_begin_0 = const()[name = tensor("op_4570_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_4570_end_0 = const()[name = tensor("op_4570_end_0"), val = tensor([1, 1, 16, 32, 2])]; + tensor var_4570_end_mask_0 = const()[name = tensor("op_4570_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_4570_squeeze_mask_0 = const()[name = tensor("op_4570_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_4570_cast_fp16 = slice_by_index(begin = var_4570_begin_0, end = var_4570_end_0, end_mask = var_4570_end_mask_0, squeeze_mask = var_4570_squeeze_mask_0, x = k_complex_21_cast_fp16)[name = tensor("op_4570_cast_fp16")]; + tensor freqs_21_to_fp16 = const()[name = tensor("freqs_21_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(303674880)))]; + tensor var_4576_cast_fp16 = mul(x = freqs_21_to_fp16, y = ts_65_cast_fp16)[name = tensor("op_4576_cast_fp16")]; + tensor rotr_21_cast_fp16 = cos(x = var_4576_cast_fp16)[name = tensor("rotr_21_cast_fp16")]; + tensor roti_21_cast_fp16 = sin(x = var_4576_cast_fp16)[name = tensor("roti_21_cast_fp16")]; + tensor var_4580_cast_fp16 = mul(x = var_4546_cast_fp16, y = rotr_21_cast_fp16)[name = tensor("op_4580_cast_fp16")]; + tensor var_4581_cast_fp16 = mul(x = var_4554_cast_fp16, y = roti_21_cast_fp16)[name = tensor("op_4581_cast_fp16")]; + tensor qor_41_cast_fp16 = sub(x = var_4580_cast_fp16, y = var_4581_cast_fp16)[name = tensor("qor_41_cast_fp16")]; + tensor var_4584_cast_fp16 = mul(x = var_4546_cast_fp16, y = roti_21_cast_fp16)[name = tensor("op_4584_cast_fp16")]; + tensor var_4585_cast_fp16 = mul(x = var_4554_cast_fp16, y = rotr_21_cast_fp16)[name = tensor("op_4585_cast_fp16")]; + tensor qoi_41_cast_fp16 = add(x = var_4584_cast_fp16, y = var_4585_cast_fp16)[name = tensor("qoi_41_cast_fp16")]; + tensor var_4588_cast_fp16 = mul(x = var_4562_cast_fp16, y = rotr_21_cast_fp16)[name = tensor("op_4588_cast_fp16")]; + tensor var_4589_cast_fp16 = mul(x = var_4570_cast_fp16, y = roti_21_cast_fp16)[name = tensor("op_4589_cast_fp16")]; + tensor kor_41_cast_fp16 = sub(x = var_4588_cast_fp16, y = var_4589_cast_fp16)[name = tensor("kor_41_cast_fp16")]; + tensor var_4592_cast_fp16 = mul(x = var_4562_cast_fp16, y = roti_21_cast_fp16)[name = tensor("op_4592_cast_fp16")]; + tensor var_4593_cast_fp16 = mul(x = var_4570_cast_fp16, y = rotr_21_cast_fp16)[name = tensor("op_4593_cast_fp16")]; + tensor koi_41_cast_fp16 = add(x = var_4592_cast_fp16, y = var_4593_cast_fp16)[name = tensor("koi_41_cast_fp16")]; + tensor qo_21_axis_0 = const()[name = tensor("qo_21_axis_0"), val = tensor(-1)]; + tensor qo_21_cast_fp16 = stack(axis = qo_21_axis_0, values = (qor_41_cast_fp16, qoi_41_cast_fp16))[name = tensor("qo_21_cast_fp16")]; + tensor ko_21_axis_0 = const()[name = tensor("ko_21_axis_0"), val = tensor(-1)]; + tensor ko_21_cast_fp16 = stack(axis = ko_21_axis_0, values = (kor_41_cast_fp16, koi_41_cast_fp16))[name = tensor("ko_21_cast_fp16")]; + tensor var_4622 = const()[name = tensor("op_4622"), val = tensor([1, 1, 16, 64])]; + tensor q_63_cast_fp16 = reshape(shape = var_4622, x = qo_21_cast_fp16)[name = tensor("q_63_cast_fp16")]; + tensor var_4624 = const()[name = tensor("op_4624"), val = tensor([1, 1, 16, 64])]; + tensor k_43_cast_fp16 = reshape(shape = var_4624, x = ko_21_cast_fp16)[name = tensor("k_43_cast_fp16")]; + tensor _inversed_4646_y_0_to_fp16 = const()[name = tensor("_inversed_4646_y_0_to_fp16"), val = tensor(0x1p-9)]; + tensor _inversed_4646_cast_fp16 = mul(x = ts_65_cast_fp16, y = _inversed_4646_y_0_to_fp16)[name = tensor("_inversed_4646_cast_fp16")]; + tensor var_4647_cast_fp16 = floor(x = _inversed_4646_cast_fp16)[name = tensor("op_4647_cast_fp16")]; + tensor var_4648_to_fp16 = const()[name = tensor("op_4648_to_fp16"), val = tensor(0x1p+9)]; + tensor var_4649_cast_fp16 = mul(x = var_4647_cast_fp16, y = var_4648_to_fp16)[name = tensor("op_4649_cast_fp16")]; + tensor write_indices_float_43_cast_fp16 = sub(x = ts_65_cast_fp16, y = var_4649_cast_fp16)[name = tensor("write_indices_float_43_cast_fp16")]; + tensor var_4656_dtype_0 = const()[name = tensor("op_4656_dtype_0"), val = tensor("int32")]; + tensor write_indices_21_reps_0 = const()[name = tensor("write_indices_21_reps_0"), val = tensor([1, 1, 16, 64])]; + tensor write_indices_float_43_cast_fp16_to_int32 = cast(dtype = var_4656_dtype_0, x = write_indices_float_43_cast_fp16)[name = tensor("cast_502")]; + tensor write_indices_21 = tile(reps = write_indices_21_reps_0, x = write_indices_float_43_cast_fp16_to_int32)[name = tensor("write_indices_21")]; + tensor var_4664_begin_0 = const()[name = tensor("op_4664_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_4664_end_0 = const()[name = tensor("op_4664_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor var_4664_end_mask_0 = const()[name = tensor("op_4664_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_4664_squeeze_mask_0 = const()[name = tensor("op_4664_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor cache10_to_fp16_dtype_0 = const()[name = tensor("cache10_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor cache10_to_fp16 = cast(dtype = cache10_to_fp16_dtype_0, x = cache10)[name = tensor("cast_501")]; + tensor var_4664_cast_fp16 = slice_by_index(begin = var_4664_begin_0, end = var_4664_end_0, end_mask = var_4664_end_mask_0, squeeze_mask = var_4664_squeeze_mask_0, x = cache10_to_fp16)[name = tensor("op_4664_cast_fp16")]; + tensor var_4666_axis_0 = const()[name = tensor("op_4666_axis_0"), val = tensor(1)]; + tensor var_4666_mode_0 = const()[name = tensor("op_4666_mode_0"), val = tensor("update")]; + tensor var_4666_validate_indices_0 = const()[name = tensor("op_4666_validate_indices_0"), val = tensor(false)]; + tensor var_4666_cast_fp16 = scatter_along_axis(axis = var_4666_axis_0, data = var_4664_cast_fp16, indices = write_indices_21, mode = var_4666_mode_0, updates = k_43_cast_fp16, validate_indices = var_4666_validate_indices_0)[name = tensor("op_4666_cast_fp16")]; + tensor concat_72 = const()[name = tensor("concat_72"), val = tensor([0, 0, 0, 0, 0])]; + tensor concat_73 = const()[name = tensor("concat_73"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_21_internal_tensor_assign_1_stride_0 = const()[name = tensor("new_cache_21_internal_tensor_assign_1_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_21_internal_tensor_assign_1_begin_mask_0 = const()[name = tensor("new_cache_21_internal_tensor_assign_1_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_21_internal_tensor_assign_1_end_mask_0 = const()[name = tensor("new_cache_21_internal_tensor_assign_1_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_21_internal_tensor_assign_1_squeeze_mask_0 = const()[name = tensor("new_cache_21_internal_tensor_assign_1_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_68 = const()[name = tensor("shape_68"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_20 = const()[name = tensor("reduce_prod_20"), val = tensor(1048576)]; + tensor range_1d_20_start_0 = const()[name = tensor("range_1d_20_start_0"), val = tensor(0)]; + tensor range_1d_20_step_0 = const()[name = tensor("range_1d_20_step_0"), val = tensor(1)]; + tensor range_1d_20 = range_1d(end = reduce_prod_20, start = range_1d_20_start_0, step = range_1d_20_step_0)[name = tensor("range_1d_20")]; + tensor reshape_100 = reshape(shape = shape_68, x = range_1d_20)[name = tensor("reshape_100")]; + tensor slice_by_index_20 = slice_by_index(begin = concat_72, begin_mask = new_cache_21_internal_tensor_assign_1_begin_mask_0, end = concat_73, end_mask = new_cache_21_internal_tensor_assign_1_end_mask_0, squeeze_mask = new_cache_21_internal_tensor_assign_1_squeeze_mask_0, stride = new_cache_21_internal_tensor_assign_1_stride_0, x = reshape_100)[name = tensor("slice_by_index_20")]; + tensor reshape_101_shape_0 = const()[name = tensor("reshape_101_shape_0"), val = tensor([-1])]; + tensor reshape_101 = reshape(shape = reshape_101_shape_0, x = slice_by_index_20)[name = tensor("reshape_101")]; + tensor reshape_102_shape_0 = const()[name = tensor("reshape_102_shape_0"), val = tensor([-1])]; + tensor reshape_102_cast_fp16 = reshape(shape = reshape_102_shape_0, x = var_4666_cast_fp16)[name = tensor("reshape_102_cast_fp16")]; + tensor reshape_103_shape_0 = const()[name = tensor("reshape_103_shape_0"), val = tensor([-1])]; + tensor reshape_103_cast_fp16 = reshape(shape = reshape_103_shape_0, x = cache10_to_fp16)[name = tensor("reshape_103_cast_fp16")]; + tensor scatter_20_mode_0 = const()[name = tensor("scatter_20_mode_0"), val = tensor("update")]; + tensor scatter_20_axis_0 = const()[name = tensor("scatter_20_axis_0"), val = tensor(0)]; + tensor scatter_20_validate_indices_0 = const()[name = tensor("scatter_20_validate_indices_0"), val = tensor(false)]; + tensor scatter_20_cast_fp16 = scatter(axis = scatter_20_axis_0, data = reshape_103_cast_fp16, indices = reshape_101, mode = scatter_20_mode_0, updates = reshape_102_cast_fp16, validate_indices = scatter_20_validate_indices_0)[name = tensor("scatter_20_cast_fp16")]; + tensor reshape_104_cast_fp16 = reshape(shape = shape_68, x = scatter_20_cast_fp16)[name = tensor("reshape_104_cast_fp16")]; + tensor var_4674_begin_0 = const()[name = tensor("op_4674_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor var_4674_end_0 = const()[name = tensor("op_4674_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor var_4674_end_mask_0 = const()[name = tensor("op_4674_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_4674_squeeze_mask_0 = const()[name = tensor("op_4674_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor var_4674_cast_fp16 = slice_by_index(begin = var_4674_begin_0, end = var_4674_end_0, end_mask = var_4674_end_mask_0, squeeze_mask = var_4674_squeeze_mask_0, x = reshape_104_cast_fp16)[name = tensor("op_4674_cast_fp16")]; + tensor var_4676_axis_0 = const()[name = tensor("op_4676_axis_0"), val = tensor(1)]; + tensor var_4676_mode_0 = const()[name = tensor("op_4676_mode_0"), val = tensor("update")]; + tensor var_4676_validate_indices_0 = const()[name = tensor("op_4676_validate_indices_0"), val = tensor(false)]; + tensor var_4676_cast_fp16 = scatter_along_axis(axis = var_4676_axis_0, data = var_4674_cast_fp16, indices = write_indices_21, mode = var_4676_mode_0, updates = v_21_cast_fp16, validate_indices = var_4676_validate_indices_0)[name = tensor("op_4676_cast_fp16")]; + tensor concat_74 = const()[name = tensor("concat_74"), val = tensor([1, 0, 0, 0, 0])]; + tensor concat_75 = const()[name = tensor("concat_75"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_21_internal_tensor_assign_2_stride_0 = const()[name = tensor("new_cache_21_internal_tensor_assign_2_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_21_internal_tensor_assign_2_begin_mask_0 = const()[name = tensor("new_cache_21_internal_tensor_assign_2_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_21_internal_tensor_assign_2_end_mask_0 = const()[name = tensor("new_cache_21_internal_tensor_assign_2_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_21_internal_tensor_assign_2_squeeze_mask_0 = const()[name = tensor("new_cache_21_internal_tensor_assign_2_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_69 = const()[name = tensor("shape_69"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_21 = const()[name = tensor("reduce_prod_21"), val = tensor(1048576)]; + tensor range_1d_21_start_0 = const()[name = tensor("range_1d_21_start_0"), val = tensor(0)]; + tensor range_1d_21_step_0 = const()[name = tensor("range_1d_21_step_0"), val = tensor(1)]; + tensor range_1d_21 = range_1d(end = reduce_prod_21, start = range_1d_21_start_0, step = range_1d_21_step_0)[name = tensor("range_1d_21")]; + tensor reshape_105 = reshape(shape = shape_69, x = range_1d_21)[name = tensor("reshape_105")]; + tensor slice_by_index_21 = slice_by_index(begin = concat_74, begin_mask = new_cache_21_internal_tensor_assign_2_begin_mask_0, end = concat_75, end_mask = new_cache_21_internal_tensor_assign_2_end_mask_0, squeeze_mask = new_cache_21_internal_tensor_assign_2_squeeze_mask_0, stride = new_cache_21_internal_tensor_assign_2_stride_0, x = reshape_105)[name = tensor("slice_by_index_21")]; + tensor reshape_106_shape_0 = const()[name = tensor("reshape_106_shape_0"), val = tensor([-1])]; + tensor reshape_106 = reshape(shape = reshape_106_shape_0, x = slice_by_index_21)[name = tensor("reshape_106")]; + tensor reshape_107_shape_0 = const()[name = tensor("reshape_107_shape_0"), val = tensor([-1])]; + tensor reshape_107_cast_fp16 = reshape(shape = reshape_107_shape_0, x = var_4676_cast_fp16)[name = tensor("reshape_107_cast_fp16")]; + tensor reshape_108_shape_0 = const()[name = tensor("reshape_108_shape_0"), val = tensor([-1])]; + tensor reshape_108_cast_fp16 = reshape(shape = reshape_108_shape_0, x = reshape_104_cast_fp16)[name = tensor("reshape_108_cast_fp16")]; + tensor scatter_21_mode_0 = const()[name = tensor("scatter_21_mode_0"), val = tensor("update")]; + tensor scatter_21_axis_0 = const()[name = tensor("scatter_21_axis_0"), val = tensor(0)]; + tensor scatter_21_validate_indices_0 = const()[name = tensor("scatter_21_validate_indices_0"), val = tensor(false)]; + tensor scatter_21_cast_fp16 = scatter(axis = scatter_21_axis_0, data = reshape_108_cast_fp16, indices = reshape_106, mode = scatter_21_mode_0, updates = reshape_107_cast_fp16, validate_indices = scatter_21_validate_indices_0)[name = tensor("scatter_21_cast_fp16")]; + tensor reshape_109_cast_fp16 = reshape(shape = shape_69, x = scatter_21_cast_fp16)[name = tensor("reshape_109_cast_fp16")]; + tensor reshape_109_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("reshape_109_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor keys_61_begin_0 = const()[name = tensor("keys_61_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor keys_61_end_0 = const()[name = tensor("keys_61_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor keys_61_end_mask_0 = const()[name = tensor("keys_61_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor keys_61_squeeze_mask_0 = const()[name = tensor("keys_61_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor keys_61_cast_fp16 = slice_by_index(begin = keys_61_begin_0, end = keys_61_end_0, end_mask = keys_61_end_mask_0, squeeze_mask = keys_61_squeeze_mask_0, x = reshape_109_cast_fp16)[name = tensor("keys_61_cast_fp16")]; + tensor values_61_begin_0 = const()[name = tensor("values_61_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor values_61_end_0 = const()[name = tensor("values_61_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor values_61_end_mask_0 = const()[name = tensor("values_61_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor values_61_squeeze_mask_0 = const()[name = tensor("values_61_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor values_61_cast_fp16 = slice_by_index(begin = values_61_begin_0, end = values_61_end_0, end_mask = values_61_end_mask_0, squeeze_mask = values_61_squeeze_mask_0, x = reshape_109_cast_fp16)[name = tensor("values_61_cast_fp16")]; + tensor var_4688_cast_fp16 = not_equal(x = keys_61_cast_fp16, y = keys_61_cast_fp16)[name = tensor("op_4688_cast_fp16")]; + tensor keys_63_cast_fp16 = select(a = var_504_to_fp16, b = keys_61_cast_fp16, cond = var_4688_cast_fp16)[name = tensor("keys_63_cast_fp16")]; + tensor var_4696_cast_fp16 = not_equal(x = values_61_cast_fp16, y = values_61_cast_fp16)[name = tensor("op_4696_cast_fp16")]; + tensor values_63_cast_fp16 = select(a = var_504_to_fp16, b = values_61_cast_fp16, cond = var_4696_cast_fp16)[name = tensor("values_63_cast_fp16")]; + tensor var_4720 = const()[name = tensor("op_4720"), val = tensor([0, 2, 1, 3])]; + tensor var_4733 = const()[name = tensor("op_4733"), val = tensor([1, 1, 1])]; + tensor var_4734_cast_fp16 = reshape(shape = var_4733, x = position10_to_fp16)[name = tensor("op_4734_cast_fp16")]; + tensor var_4751_to_fp16 = const()[name = tensor("op_4751_to_fp16"), val = tensor(0x1p+0)]; + tensor valid_len_21_cast_fp16 = add(x = var_4734_cast_fp16, y = var_4751_to_fp16)[name = tensor("valid_len_21_cast_fp16")]; + tensor valid_mask_21_cast_fp16 = less(x = k_positions_1_promoted_to_fp16, y = valid_len_21_cast_fp16)[name = tensor("valid_mask_21_cast_fp16")]; + tensor causal_mask_21_cast_fp16 = less_equal(x = k_positions_1_promoted_to_fp16, y = var_4734_cast_fp16)[name = tensor("causal_mask_21_cast_fp16")]; + tensor attn_mask_41 = logical_and(x = valid_mask_21_cast_fp16, y = causal_mask_21_cast_fp16)[name = tensor("attn_mask_41")]; + tensor attn_mask_43_axes_0 = const()[name = tensor("attn_mask_43_axes_0"), val = tensor([1])]; + tensor attn_mask_43 = expand_dims(axes = attn_mask_43_axes_0, x = attn_mask_41)[name = tensor("attn_mask_43")]; + tensor var_4769_transpose_x_0 = const()[name = tensor("op_4769_transpose_x_0"), val = tensor(false)]; + tensor var_4769_transpose_y_0 = const()[name = tensor("op_4769_transpose_y_0"), val = tensor(false)]; + tensor transpose_92_perm_0 = const()[name = tensor("transpose_92_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_93_perm_0 = const()[name = tensor("transpose_93_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_93 = transpose(perm = transpose_93_perm_0, x = keys_63_cast_fp16)[name = tensor("transpose_173")]; + tensor transpose_92 = transpose(perm = transpose_92_perm_0, x = q_63_cast_fp16)[name = tensor("transpose_174")]; + tensor var_4769_cast_fp16 = matmul(transpose_x = var_4769_transpose_x_0, transpose_y = var_4769_transpose_y_0, x = transpose_92, y = transpose_93)[name = tensor("op_4769_cast_fp16")]; + tensor var_4763_to_fp16 = const()[name = tensor("op_4763_to_fp16"), val = tensor([0x1p-3])]; + tensor attn_weights_61_cast_fp16 = mul(x = var_4769_cast_fp16, y = var_4763_to_fp16)[name = tensor("attn_weights_61_cast_fp16")]; + tensor var_4771 = logical_not(x = attn_mask_43)[name = tensor("op_4771")]; + tensor var_4772_to_fp16 = const()[name = tensor("op_4772_to_fp16"), val = tensor(-inf)]; + tensor attn_weights_63_cast_fp16 = select(a = var_4772_to_fp16, b = attn_weights_61_cast_fp16, cond = var_4771)[name = tensor("attn_weights_63_cast_fp16")]; + tensor var_4774 = const()[name = tensor("op_4774"), val = tensor(-1)]; + tensor attn_weights_65_cast_fp16 = softmax(axis = var_4774, x = attn_weights_63_cast_fp16)[name = tensor("attn_weights_65_cast_fp16")]; + tensor attn_output_21_transpose_x_0 = const()[name = tensor("attn_output_21_transpose_x_0"), val = tensor(false)]; + tensor attn_output_21_transpose_y_0 = const()[name = tensor("attn_output_21_transpose_y_0"), val = tensor(false)]; + tensor values_65_cast_fp16 = transpose(perm = var_4720, x = values_63_cast_fp16)[name = tensor("transpose_175")]; + tensor attn_output_21_cast_fp16 = matmul(transpose_x = attn_output_21_transpose_x_0, transpose_y = attn_output_21_transpose_y_0, x = attn_weights_65_cast_fp16, y = values_65_cast_fp16)[name = tensor("attn_output_21_cast_fp16")]; + tensor var_4782 = const()[name = tensor("op_4782"), val = tensor([0, 2, 1, 3])]; + tensor var_4785 = const()[name = tensor("op_4785"), val = tensor([1, 1, 1024])]; + tensor var_4783_cast_fp16 = transpose(perm = var_4782, x = attn_output_21_cast_fp16)[name = tensor("transpose_172")]; + tensor input_105_cast_fp16 = reshape(shape = var_4785, x = var_4783_cast_fp16)[name = tensor("input_105_cast_fp16")]; + tensor linear_42_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = attn10_out_proj_weight_cast_fp16, x = input_105_cast_fp16)[name = tensor("linear_42_cast_fp16")]; + tensor var_4791_to_fp16 = const()[name = tensor("op_4791_to_fp16"), val = tensor(0x1p+0)]; + tensor var_4792_cast_fp16 = add(x = position10_to_fp16, y = var_4791_to_fp16)[name = tensor("op_4792_cast_fp16")]; + tensor var_4792_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_4792_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor input_107_cast_fp16 = add(x = input_103_cast_fp16, y = linear_42_cast_fp16)[name = tensor("input_107_cast_fp16")]; + tensor input_109_axes_0 = const()[name = tensor("input_109_axes_0"), val = tensor([-1])]; + tensor norm10_2_weight_to_fp16 = const()[name = tensor("norm10_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(303675008)))]; + tensor norm10_2_bias_to_fp16 = const()[name = tensor("norm10_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(303677120)))]; + tensor var_4796_to_fp16 = const()[name = tensor("op_4796_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_109_cast_fp16 = layer_norm(axes = input_109_axes_0, beta = norm10_2_bias_to_fp16, epsilon = var_4796_to_fp16, gamma = norm10_2_weight_to_fp16, x = input_107_cast_fp16)[name = tensor("input_109_cast_fp16")]; + tensor linear_43_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = linear10_1_weight_cast_fp16, x = input_109_cast_fp16)[name = tensor("linear_43_cast_fp16")]; + tensor input_111_mode_0 = const()[name = tensor("input_111_mode_0"), val = tensor("EXACT")]; + tensor input_111_cast_fp16 = gelu(mode = input_111_mode_0, x = linear_43_cast_fp16)[name = tensor("input_111_cast_fp16")]; + tensor linear_44_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = linear10_2_weight_cast_fp16, x = input_111_cast_fp16)[name = tensor("linear_44_cast_fp16")]; + tensor input_113_cast_fp16 = add(x = input_107_cast_fp16, y = linear_44_cast_fp16)[name = tensor("input_113_cast_fp16")]; + tensor x_23_axes_0 = const()[name = tensor("x_23_axes_0"), val = tensor([-1])]; + tensor norm11_1_weight_to_fp16 = const()[name = tensor("norm11_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(303679232)))]; + tensor norm11_1_bias_to_fp16 = const()[name = tensor("norm11_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(303681344)))]; + tensor var_4813_to_fp16 = const()[name = tensor("op_4813_to_fp16"), val = tensor(0x1.5p-17)]; + tensor x_23_cast_fp16 = layer_norm(axes = x_23_axes_0, beta = norm11_1_bias_to_fp16, epsilon = var_4813_to_fp16, gamma = norm11_1_weight_to_fp16, x = input_113_cast_fp16)[name = tensor("x_23_cast_fp16")]; + tensor linear_45_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = attn11_in_proj_weight_cast_fp16, x = x_23_cast_fp16)[name = tensor("linear_45_cast_fp16")]; + tensor var_4849 = const()[name = tensor("op_4849"), val = tensor([1, 1, 3, 16, 64])]; + tensor qkv_23_cast_fp16 = reshape(shape = var_4849, x = linear_45_cast_fp16)[name = tensor("qkv_23_cast_fp16")]; + tensor q_67_begin_0 = const()[name = tensor("q_67_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor q_67_end_0 = const()[name = tensor("q_67_end_0"), val = tensor([1, 1, 1, 16, 64])]; + tensor q_67_end_mask_0 = const()[name = tensor("q_67_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor q_67_squeeze_mask_0 = const()[name = tensor("q_67_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor q_67_cast_fp16 = slice_by_index(begin = q_67_begin_0, end = q_67_end_0, end_mask = q_67_end_mask_0, squeeze_mask = q_67_squeeze_mask_0, x = qkv_23_cast_fp16)[name = tensor("q_67_cast_fp16")]; + tensor k_45_begin_0 = const()[name = tensor("k_45_begin_0"), val = tensor([0, 0, 1, 0, 0])]; + tensor k_45_end_0 = const()[name = tensor("k_45_end_0"), val = tensor([1, 1, 2, 16, 64])]; + tensor k_45_end_mask_0 = const()[name = tensor("k_45_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor k_45_squeeze_mask_0 = const()[name = tensor("k_45_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor k_45_cast_fp16 = slice_by_index(begin = k_45_begin_0, end = k_45_end_0, end_mask = k_45_end_mask_0, squeeze_mask = k_45_squeeze_mask_0, x = qkv_23_cast_fp16)[name = tensor("k_45_cast_fp16")]; + tensor v_23_begin_0 = const()[name = tensor("v_23_begin_0"), val = tensor([0, 0, 2, 0, 0])]; + tensor v_23_end_0 = const()[name = tensor("v_23_end_0"), val = tensor([1, 1, 3, 16, 64])]; + tensor v_23_end_mask_0 = const()[name = tensor("v_23_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor v_23_squeeze_mask_0 = const()[name = tensor("v_23_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor v_23_cast_fp16 = slice_by_index(begin = v_23_begin_0, end = v_23_end_0, end_mask = v_23_end_mask_0, squeeze_mask = v_23_squeeze_mask_0, x = qkv_23_cast_fp16)[name = tensor("v_23_cast_fp16")]; + tensor var_4953 = const()[name = tensor("op_4953"), val = tensor([1, 1, 1, 1])]; + tensor position11_to_fp16_dtype_0 = const()[name = tensor("position11_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor position11_to_fp16 = cast(dtype = position11_to_fp16_dtype_0, x = position11)[name = tensor("cast_498")]; + tensor ts_71_cast_fp16 = reshape(shape = var_4953, x = position11_to_fp16)[name = tensor("ts_71_cast_fp16")]; + tensor var_4957 = const()[name = tensor("op_4957"), val = tensor([1, 1, 16, 32, 2])]; + tensor q_complex_23_cast_fp16 = reshape(shape = var_4957, x = q_67_cast_fp16)[name = tensor("q_complex_23_cast_fp16")]; + tensor var_4961 = const()[name = tensor("op_4961"), val = tensor([1, 1, 16, 32, 2])]; + tensor k_complex_23_cast_fp16 = reshape(shape = var_4961, x = k_45_cast_fp16)[name = tensor("k_complex_23_cast_fp16")]; + tensor var_4965_begin_0 = const()[name = tensor("op_4965_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_4965_end_0 = const()[name = tensor("op_4965_end_0"), val = tensor([1, 1, 16, 32, 1])]; + tensor var_4965_end_mask_0 = const()[name = tensor("op_4965_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_4965_squeeze_mask_0 = const()[name = tensor("op_4965_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_4965_cast_fp16 = slice_by_index(begin = var_4965_begin_0, end = var_4965_end_0, end_mask = var_4965_end_mask_0, squeeze_mask = var_4965_squeeze_mask_0, x = q_complex_23_cast_fp16)[name = tensor("op_4965_cast_fp16")]; + tensor var_4973_begin_0 = const()[name = tensor("op_4973_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_4973_end_0 = const()[name = tensor("op_4973_end_0"), val = tensor([1, 1, 16, 32, 2])]; + tensor var_4973_end_mask_0 = const()[name = tensor("op_4973_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_4973_squeeze_mask_0 = const()[name = tensor("op_4973_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_4973_cast_fp16 = slice_by_index(begin = var_4973_begin_0, end = var_4973_end_0, end_mask = var_4973_end_mask_0, squeeze_mask = var_4973_squeeze_mask_0, x = q_complex_23_cast_fp16)[name = tensor("op_4973_cast_fp16")]; + tensor var_4981_begin_0 = const()[name = tensor("op_4981_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_4981_end_0 = const()[name = tensor("op_4981_end_0"), val = tensor([1, 1, 16, 32, 1])]; + tensor var_4981_end_mask_0 = const()[name = tensor("op_4981_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_4981_squeeze_mask_0 = const()[name = tensor("op_4981_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_4981_cast_fp16 = slice_by_index(begin = var_4981_begin_0, end = var_4981_end_0, end_mask = var_4981_end_mask_0, squeeze_mask = var_4981_squeeze_mask_0, x = k_complex_23_cast_fp16)[name = tensor("op_4981_cast_fp16")]; + tensor var_4989_begin_0 = const()[name = tensor("op_4989_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_4989_end_0 = const()[name = tensor("op_4989_end_0"), val = tensor([1, 1, 16, 32, 2])]; + tensor var_4989_end_mask_0 = const()[name = tensor("op_4989_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_4989_squeeze_mask_0 = const()[name = tensor("op_4989_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_4989_cast_fp16 = slice_by_index(begin = var_4989_begin_0, end = var_4989_end_0, end_mask = var_4989_end_mask_0, squeeze_mask = var_4989_squeeze_mask_0, x = k_complex_23_cast_fp16)[name = tensor("op_4989_cast_fp16")]; + tensor freqs_23_to_fp16 = const()[name = tensor("freqs_23_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(303683456)))]; + tensor var_4995_cast_fp16 = mul(x = freqs_23_to_fp16, y = ts_71_cast_fp16)[name = tensor("op_4995_cast_fp16")]; + tensor rotr_23_cast_fp16 = cos(x = var_4995_cast_fp16)[name = tensor("rotr_23_cast_fp16")]; + tensor roti_23_cast_fp16 = sin(x = var_4995_cast_fp16)[name = tensor("roti_23_cast_fp16")]; + tensor var_4999_cast_fp16 = mul(x = var_4965_cast_fp16, y = rotr_23_cast_fp16)[name = tensor("op_4999_cast_fp16")]; + tensor var_5000_cast_fp16 = mul(x = var_4973_cast_fp16, y = roti_23_cast_fp16)[name = tensor("op_5000_cast_fp16")]; + tensor qor_45_cast_fp16 = sub(x = var_4999_cast_fp16, y = var_5000_cast_fp16)[name = tensor("qor_45_cast_fp16")]; + tensor var_5003_cast_fp16 = mul(x = var_4965_cast_fp16, y = roti_23_cast_fp16)[name = tensor("op_5003_cast_fp16")]; + tensor var_5004_cast_fp16 = mul(x = var_4973_cast_fp16, y = rotr_23_cast_fp16)[name = tensor("op_5004_cast_fp16")]; + tensor qoi_45_cast_fp16 = add(x = var_5003_cast_fp16, y = var_5004_cast_fp16)[name = tensor("qoi_45_cast_fp16")]; + tensor var_5007_cast_fp16 = mul(x = var_4981_cast_fp16, y = rotr_23_cast_fp16)[name = tensor("op_5007_cast_fp16")]; + tensor var_5008_cast_fp16 = mul(x = var_4989_cast_fp16, y = roti_23_cast_fp16)[name = tensor("op_5008_cast_fp16")]; + tensor kor_45_cast_fp16 = sub(x = var_5007_cast_fp16, y = var_5008_cast_fp16)[name = tensor("kor_45_cast_fp16")]; + tensor var_5011_cast_fp16 = mul(x = var_4981_cast_fp16, y = roti_23_cast_fp16)[name = tensor("op_5011_cast_fp16")]; + tensor var_5012_cast_fp16 = mul(x = var_4989_cast_fp16, y = rotr_23_cast_fp16)[name = tensor("op_5012_cast_fp16")]; + tensor koi_45_cast_fp16 = add(x = var_5011_cast_fp16, y = var_5012_cast_fp16)[name = tensor("koi_45_cast_fp16")]; + tensor qo_23_axis_0 = const()[name = tensor("qo_23_axis_0"), val = tensor(-1)]; + tensor qo_23_cast_fp16 = stack(axis = qo_23_axis_0, values = (qor_45_cast_fp16, qoi_45_cast_fp16))[name = tensor("qo_23_cast_fp16")]; + tensor ko_23_axis_0 = const()[name = tensor("ko_23_axis_0"), val = tensor(-1)]; + tensor ko_23_cast_fp16 = stack(axis = ko_23_axis_0, values = (kor_45_cast_fp16, koi_45_cast_fp16))[name = tensor("ko_23_cast_fp16")]; + tensor var_5041 = const()[name = tensor("op_5041"), val = tensor([1, 1, 16, 64])]; + tensor q_69_cast_fp16 = reshape(shape = var_5041, x = qo_23_cast_fp16)[name = tensor("q_69_cast_fp16")]; + tensor var_5043 = const()[name = tensor("op_5043"), val = tensor([1, 1, 16, 64])]; + tensor k_47_cast_fp16 = reshape(shape = var_5043, x = ko_23_cast_fp16)[name = tensor("k_47_cast_fp16")]; + tensor _inversed_5065_y_0_to_fp16 = const()[name = tensor("_inversed_5065_y_0_to_fp16"), val = tensor(0x1p-9)]; + tensor _inversed_5065_cast_fp16 = mul(x = ts_71_cast_fp16, y = _inversed_5065_y_0_to_fp16)[name = tensor("_inversed_5065_cast_fp16")]; + tensor var_5066_cast_fp16 = floor(x = _inversed_5065_cast_fp16)[name = tensor("op_5066_cast_fp16")]; + tensor var_5067_to_fp16 = const()[name = tensor("op_5067_to_fp16"), val = tensor(0x1p+9)]; + tensor var_5068_cast_fp16 = mul(x = var_5066_cast_fp16, y = var_5067_to_fp16)[name = tensor("op_5068_cast_fp16")]; + tensor write_indices_float_47_cast_fp16 = sub(x = ts_71_cast_fp16, y = var_5068_cast_fp16)[name = tensor("write_indices_float_47_cast_fp16")]; + tensor var_5075_dtype_0 = const()[name = tensor("op_5075_dtype_0"), val = tensor("int32")]; + tensor write_indices_23_reps_0 = const()[name = tensor("write_indices_23_reps_0"), val = tensor([1, 1, 16, 64])]; + tensor write_indices_float_47_cast_fp16_to_int32 = cast(dtype = var_5075_dtype_0, x = write_indices_float_47_cast_fp16)[name = tensor("cast_497")]; + tensor write_indices_23 = tile(reps = write_indices_23_reps_0, x = write_indices_float_47_cast_fp16_to_int32)[name = tensor("write_indices_23")]; + tensor var_5083_begin_0 = const()[name = tensor("op_5083_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_5083_end_0 = const()[name = tensor("op_5083_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor var_5083_end_mask_0 = const()[name = tensor("op_5083_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_5083_squeeze_mask_0 = const()[name = tensor("op_5083_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor cache11_to_fp16_dtype_0 = const()[name = tensor("cache11_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor cache11_to_fp16 = cast(dtype = cache11_to_fp16_dtype_0, x = cache11)[name = tensor("cast_496")]; + tensor var_5083_cast_fp16 = slice_by_index(begin = var_5083_begin_0, end = var_5083_end_0, end_mask = var_5083_end_mask_0, squeeze_mask = var_5083_squeeze_mask_0, x = cache11_to_fp16)[name = tensor("op_5083_cast_fp16")]; + tensor var_5085_axis_0 = const()[name = tensor("op_5085_axis_0"), val = tensor(1)]; + tensor var_5085_mode_0 = const()[name = tensor("op_5085_mode_0"), val = tensor("update")]; + tensor var_5085_validate_indices_0 = const()[name = tensor("op_5085_validate_indices_0"), val = tensor(false)]; + tensor var_5085_cast_fp16 = scatter_along_axis(axis = var_5085_axis_0, data = var_5083_cast_fp16, indices = write_indices_23, mode = var_5085_mode_0, updates = k_47_cast_fp16, validate_indices = var_5085_validate_indices_0)[name = tensor("op_5085_cast_fp16")]; + tensor concat_79 = const()[name = tensor("concat_79"), val = tensor([0, 0, 0, 0, 0])]; + tensor concat_80 = const()[name = tensor("concat_80"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_23_internal_tensor_assign_1_stride_0 = const()[name = tensor("new_cache_23_internal_tensor_assign_1_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_23_internal_tensor_assign_1_begin_mask_0 = const()[name = tensor("new_cache_23_internal_tensor_assign_1_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_23_internal_tensor_assign_1_end_mask_0 = const()[name = tensor("new_cache_23_internal_tensor_assign_1_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_23_internal_tensor_assign_1_squeeze_mask_0 = const()[name = tensor("new_cache_23_internal_tensor_assign_1_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_70 = const()[name = tensor("shape_70"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_22 = const()[name = tensor("reduce_prod_22"), val = tensor(1048576)]; + tensor range_1d_22_start_0 = const()[name = tensor("range_1d_22_start_0"), val = tensor(0)]; + tensor range_1d_22_step_0 = const()[name = tensor("range_1d_22_step_0"), val = tensor(1)]; + tensor range_1d_22 = range_1d(end = reduce_prod_22, start = range_1d_22_start_0, step = range_1d_22_step_0)[name = tensor("range_1d_22")]; + tensor reshape_110 = reshape(shape = shape_70, x = range_1d_22)[name = tensor("reshape_110")]; + tensor slice_by_index_22 = slice_by_index(begin = concat_79, begin_mask = new_cache_23_internal_tensor_assign_1_begin_mask_0, end = concat_80, end_mask = new_cache_23_internal_tensor_assign_1_end_mask_0, squeeze_mask = new_cache_23_internal_tensor_assign_1_squeeze_mask_0, stride = new_cache_23_internal_tensor_assign_1_stride_0, x = reshape_110)[name = tensor("slice_by_index_22")]; + tensor reshape_111_shape_0 = const()[name = tensor("reshape_111_shape_0"), val = tensor([-1])]; + tensor reshape_111 = reshape(shape = reshape_111_shape_0, x = slice_by_index_22)[name = tensor("reshape_111")]; + tensor reshape_112_shape_0 = const()[name = tensor("reshape_112_shape_0"), val = tensor([-1])]; + tensor reshape_112_cast_fp16 = reshape(shape = reshape_112_shape_0, x = var_5085_cast_fp16)[name = tensor("reshape_112_cast_fp16")]; + tensor reshape_113_shape_0 = const()[name = tensor("reshape_113_shape_0"), val = tensor([-1])]; + tensor reshape_113_cast_fp16 = reshape(shape = reshape_113_shape_0, x = cache11_to_fp16)[name = tensor("reshape_113_cast_fp16")]; + tensor scatter_22_mode_0 = const()[name = tensor("scatter_22_mode_0"), val = tensor("update")]; + tensor scatter_22_axis_0 = const()[name = tensor("scatter_22_axis_0"), val = tensor(0)]; + tensor scatter_22_validate_indices_0 = const()[name = tensor("scatter_22_validate_indices_0"), val = tensor(false)]; + tensor scatter_22_cast_fp16 = scatter(axis = scatter_22_axis_0, data = reshape_113_cast_fp16, indices = reshape_111, mode = scatter_22_mode_0, updates = reshape_112_cast_fp16, validate_indices = scatter_22_validate_indices_0)[name = tensor("scatter_22_cast_fp16")]; + tensor reshape_114_cast_fp16 = reshape(shape = shape_70, x = scatter_22_cast_fp16)[name = tensor("reshape_114_cast_fp16")]; + tensor var_5093_begin_0 = const()[name = tensor("op_5093_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor var_5093_end_0 = const()[name = tensor("op_5093_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor var_5093_end_mask_0 = const()[name = tensor("op_5093_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_5093_squeeze_mask_0 = const()[name = tensor("op_5093_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor var_5093_cast_fp16 = slice_by_index(begin = var_5093_begin_0, end = var_5093_end_0, end_mask = var_5093_end_mask_0, squeeze_mask = var_5093_squeeze_mask_0, x = reshape_114_cast_fp16)[name = tensor("op_5093_cast_fp16")]; + tensor var_5095_axis_0 = const()[name = tensor("op_5095_axis_0"), val = tensor(1)]; + tensor var_5095_mode_0 = const()[name = tensor("op_5095_mode_0"), val = tensor("update")]; + tensor var_5095_validate_indices_0 = const()[name = tensor("op_5095_validate_indices_0"), val = tensor(false)]; + tensor var_5095_cast_fp16 = scatter_along_axis(axis = var_5095_axis_0, data = var_5093_cast_fp16, indices = write_indices_23, mode = var_5095_mode_0, updates = v_23_cast_fp16, validate_indices = var_5095_validate_indices_0)[name = tensor("op_5095_cast_fp16")]; + tensor concat_81 = const()[name = tensor("concat_81"), val = tensor([1, 0, 0, 0, 0])]; + tensor concat_82 = const()[name = tensor("concat_82"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_23_internal_tensor_assign_2_stride_0 = const()[name = tensor("new_cache_23_internal_tensor_assign_2_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_23_internal_tensor_assign_2_begin_mask_0 = const()[name = tensor("new_cache_23_internal_tensor_assign_2_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_23_internal_tensor_assign_2_end_mask_0 = const()[name = tensor("new_cache_23_internal_tensor_assign_2_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_23_internal_tensor_assign_2_squeeze_mask_0 = const()[name = tensor("new_cache_23_internal_tensor_assign_2_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_71 = const()[name = tensor("shape_71"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_23 = const()[name = tensor("reduce_prod_23"), val = tensor(1048576)]; + tensor range_1d_23_start_0 = const()[name = tensor("range_1d_23_start_0"), val = tensor(0)]; + tensor range_1d_23_step_0 = const()[name = tensor("range_1d_23_step_0"), val = tensor(1)]; + tensor range_1d_23 = range_1d(end = reduce_prod_23, start = range_1d_23_start_0, step = range_1d_23_step_0)[name = tensor("range_1d_23")]; + tensor reshape_115 = reshape(shape = shape_71, x = range_1d_23)[name = tensor("reshape_115")]; + tensor slice_by_index_23 = slice_by_index(begin = concat_81, begin_mask = new_cache_23_internal_tensor_assign_2_begin_mask_0, end = concat_82, end_mask = new_cache_23_internal_tensor_assign_2_end_mask_0, squeeze_mask = new_cache_23_internal_tensor_assign_2_squeeze_mask_0, stride = new_cache_23_internal_tensor_assign_2_stride_0, x = reshape_115)[name = tensor("slice_by_index_23")]; + tensor reshape_116_shape_0 = const()[name = tensor("reshape_116_shape_0"), val = tensor([-1])]; + tensor reshape_116 = reshape(shape = reshape_116_shape_0, x = slice_by_index_23)[name = tensor("reshape_116")]; + tensor reshape_117_shape_0 = const()[name = tensor("reshape_117_shape_0"), val = tensor([-1])]; + tensor reshape_117_cast_fp16 = reshape(shape = reshape_117_shape_0, x = var_5095_cast_fp16)[name = tensor("reshape_117_cast_fp16")]; + tensor reshape_118_shape_0 = const()[name = tensor("reshape_118_shape_0"), val = tensor([-1])]; + tensor reshape_118_cast_fp16 = reshape(shape = reshape_118_shape_0, x = reshape_114_cast_fp16)[name = tensor("reshape_118_cast_fp16")]; + tensor scatter_23_mode_0 = const()[name = tensor("scatter_23_mode_0"), val = tensor("update")]; + tensor scatter_23_axis_0 = const()[name = tensor("scatter_23_axis_0"), val = tensor(0)]; + tensor scatter_23_validate_indices_0 = const()[name = tensor("scatter_23_validate_indices_0"), val = tensor(false)]; + tensor scatter_23_cast_fp16 = scatter(axis = scatter_23_axis_0, data = reshape_118_cast_fp16, indices = reshape_116, mode = scatter_23_mode_0, updates = reshape_117_cast_fp16, validate_indices = scatter_23_validate_indices_0)[name = tensor("scatter_23_cast_fp16")]; + tensor reshape_119_cast_fp16 = reshape(shape = shape_71, x = scatter_23_cast_fp16)[name = tensor("reshape_119_cast_fp16")]; + tensor reshape_119_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("reshape_119_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor keys_67_begin_0 = const()[name = tensor("keys_67_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor keys_67_end_0 = const()[name = tensor("keys_67_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor keys_67_end_mask_0 = const()[name = tensor("keys_67_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor keys_67_squeeze_mask_0 = const()[name = tensor("keys_67_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor keys_67_cast_fp16 = slice_by_index(begin = keys_67_begin_0, end = keys_67_end_0, end_mask = keys_67_end_mask_0, squeeze_mask = keys_67_squeeze_mask_0, x = reshape_119_cast_fp16)[name = tensor("keys_67_cast_fp16")]; + tensor values_67_begin_0 = const()[name = tensor("values_67_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor values_67_end_0 = const()[name = tensor("values_67_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor values_67_end_mask_0 = const()[name = tensor("values_67_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor values_67_squeeze_mask_0 = const()[name = tensor("values_67_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor values_67_cast_fp16 = slice_by_index(begin = values_67_begin_0, end = values_67_end_0, end_mask = values_67_end_mask_0, squeeze_mask = values_67_squeeze_mask_0, x = reshape_119_cast_fp16)[name = tensor("values_67_cast_fp16")]; + tensor var_5107_cast_fp16 = not_equal(x = keys_67_cast_fp16, y = keys_67_cast_fp16)[name = tensor("op_5107_cast_fp16")]; + tensor keys_69_cast_fp16 = select(a = var_504_to_fp16, b = keys_67_cast_fp16, cond = var_5107_cast_fp16)[name = tensor("keys_69_cast_fp16")]; + tensor var_5115_cast_fp16 = not_equal(x = values_67_cast_fp16, y = values_67_cast_fp16)[name = tensor("op_5115_cast_fp16")]; + tensor values_69_cast_fp16 = select(a = var_504_to_fp16, b = values_67_cast_fp16, cond = var_5115_cast_fp16)[name = tensor("values_69_cast_fp16")]; + tensor var_5139 = const()[name = tensor("op_5139"), val = tensor([0, 2, 1, 3])]; + tensor var_5152 = const()[name = tensor("op_5152"), val = tensor([1, 1, 1])]; + tensor var_5153_cast_fp16 = reshape(shape = var_5152, x = position11_to_fp16)[name = tensor("op_5153_cast_fp16")]; + tensor var_5170_to_fp16 = const()[name = tensor("op_5170_to_fp16"), val = tensor(0x1p+0)]; + tensor valid_len_23_cast_fp16 = add(x = var_5153_cast_fp16, y = var_5170_to_fp16)[name = tensor("valid_len_23_cast_fp16")]; + tensor valid_mask_23_cast_fp16 = less(x = k_positions_1_promoted_to_fp16, y = valid_len_23_cast_fp16)[name = tensor("valid_mask_23_cast_fp16")]; + tensor causal_mask_23_cast_fp16 = less_equal(x = k_positions_1_promoted_to_fp16, y = var_5153_cast_fp16)[name = tensor("causal_mask_23_cast_fp16")]; + tensor attn_mask_45 = logical_and(x = valid_mask_23_cast_fp16, y = causal_mask_23_cast_fp16)[name = tensor("attn_mask_45")]; + tensor attn_mask_47_axes_0 = const()[name = tensor("attn_mask_47_axes_0"), val = tensor([1])]; + tensor attn_mask_47 = expand_dims(axes = attn_mask_47_axes_0, x = attn_mask_45)[name = tensor("attn_mask_47")]; + tensor var_5188_transpose_x_0 = const()[name = tensor("op_5188_transpose_x_0"), val = tensor(false)]; + tensor var_5188_transpose_y_0 = const()[name = tensor("op_5188_transpose_y_0"), val = tensor(false)]; + tensor transpose_94_perm_0 = const()[name = tensor("transpose_94_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_95_perm_0 = const()[name = tensor("transpose_95_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_95 = transpose(perm = transpose_95_perm_0, x = keys_69_cast_fp16)[name = tensor("transpose_169")]; + tensor transpose_94 = transpose(perm = transpose_94_perm_0, x = q_69_cast_fp16)[name = tensor("transpose_170")]; + tensor var_5188_cast_fp16 = matmul(transpose_x = var_5188_transpose_x_0, transpose_y = var_5188_transpose_y_0, x = transpose_94, y = transpose_95)[name = tensor("op_5188_cast_fp16")]; + tensor var_5182_to_fp16 = const()[name = tensor("op_5182_to_fp16"), val = tensor([0x1p-3])]; + tensor attn_weights_67_cast_fp16 = mul(x = var_5188_cast_fp16, y = var_5182_to_fp16)[name = tensor("attn_weights_67_cast_fp16")]; + tensor var_5190 = logical_not(x = attn_mask_47)[name = tensor("op_5190")]; + tensor var_5191_to_fp16 = const()[name = tensor("op_5191_to_fp16"), val = tensor(-inf)]; + tensor attn_weights_69_cast_fp16 = select(a = var_5191_to_fp16, b = attn_weights_67_cast_fp16, cond = var_5190)[name = tensor("attn_weights_69_cast_fp16")]; + tensor var_5193 = const()[name = tensor("op_5193"), val = tensor(-1)]; + tensor attn_weights_71_cast_fp16 = softmax(axis = var_5193, x = attn_weights_69_cast_fp16)[name = tensor("attn_weights_71_cast_fp16")]; + tensor attn_output_23_transpose_x_0 = const()[name = tensor("attn_output_23_transpose_x_0"), val = tensor(false)]; + tensor attn_output_23_transpose_y_0 = const()[name = tensor("attn_output_23_transpose_y_0"), val = tensor(false)]; + tensor values_71_cast_fp16 = transpose(perm = var_5139, x = values_69_cast_fp16)[name = tensor("transpose_171")]; + tensor attn_output_23_cast_fp16 = matmul(transpose_x = attn_output_23_transpose_x_0, transpose_y = attn_output_23_transpose_y_0, x = attn_weights_71_cast_fp16, y = values_71_cast_fp16)[name = tensor("attn_output_23_cast_fp16")]; + tensor var_5201 = const()[name = tensor("op_5201"), val = tensor([0, 2, 1, 3])]; + tensor var_5204 = const()[name = tensor("op_5204"), val = tensor([1, 1, 1024])]; + tensor var_5202_cast_fp16 = transpose(perm = var_5201, x = attn_output_23_cast_fp16)[name = tensor("transpose_168")]; + tensor input_115_cast_fp16 = reshape(shape = var_5204, x = var_5202_cast_fp16)[name = tensor("input_115_cast_fp16")]; + tensor linear_46_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = attn11_out_proj_weight_cast_fp16, x = input_115_cast_fp16)[name = tensor("linear_46_cast_fp16")]; + tensor var_5210_to_fp16 = const()[name = tensor("op_5210_to_fp16"), val = tensor(0x1p+0)]; + tensor var_5211_cast_fp16 = add(x = position11_to_fp16, y = var_5210_to_fp16)[name = tensor("op_5211_cast_fp16")]; + tensor var_5211_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_5211_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor input_117_cast_fp16 = add(x = input_113_cast_fp16, y = linear_46_cast_fp16)[name = tensor("input_117_cast_fp16")]; + tensor input_119_axes_0 = const()[name = tensor("input_119_axes_0"), val = tensor([-1])]; + tensor norm11_2_weight_to_fp16 = const()[name = tensor("norm11_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(303683584)))]; + tensor norm11_2_bias_to_fp16 = const()[name = tensor("norm11_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(303685696)))]; + tensor var_5215_to_fp16 = const()[name = tensor("op_5215_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_119_cast_fp16 = layer_norm(axes = input_119_axes_0, beta = norm11_2_bias_to_fp16, epsilon = var_5215_to_fp16, gamma = norm11_2_weight_to_fp16, x = input_117_cast_fp16)[name = tensor("input_119_cast_fp16")]; + tensor linear_47_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = linear11_1_weight_cast_fp16, x = input_119_cast_fp16)[name = tensor("linear_47_cast_fp16")]; + tensor input_121_mode_0 = const()[name = tensor("input_121_mode_0"), val = tensor("EXACT")]; + tensor input_121_cast_fp16 = gelu(mode = input_121_mode_0, x = linear_47_cast_fp16)[name = tensor("input_121_cast_fp16")]; + tensor linear_48_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = linear11_2_weight_cast_fp16, x = input_121_cast_fp16)[name = tensor("linear_48_cast_fp16")]; + tensor input_123_cast_fp16 = add(x = input_117_cast_fp16, y = linear_48_cast_fp16)[name = tensor("input_123_cast_fp16")]; + tensor x_25_axes_0 = const()[name = tensor("x_25_axes_0"), val = tensor([-1])]; + tensor norm12_1_weight_to_fp16 = const()[name = tensor("norm12_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(303687808)))]; + tensor norm12_1_bias_to_fp16 = const()[name = tensor("norm12_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(303689920)))]; + tensor var_5232_to_fp16 = const()[name = tensor("op_5232_to_fp16"), val = tensor(0x1.5p-17)]; + tensor x_25_cast_fp16 = layer_norm(axes = x_25_axes_0, beta = norm12_1_bias_to_fp16, epsilon = var_5232_to_fp16, gamma = norm12_1_weight_to_fp16, x = input_123_cast_fp16)[name = tensor("x_25_cast_fp16")]; + tensor linear_49_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = attn12_in_proj_weight_cast_fp16, x = x_25_cast_fp16)[name = tensor("linear_49_cast_fp16")]; + tensor var_5268 = const()[name = tensor("op_5268"), val = tensor([1, 1, 3, 16, 64])]; + tensor qkv_25_cast_fp16 = reshape(shape = var_5268, x = linear_49_cast_fp16)[name = tensor("qkv_25_cast_fp16")]; + tensor q_73_begin_0 = const()[name = tensor("q_73_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor q_73_end_0 = const()[name = tensor("q_73_end_0"), val = tensor([1, 1, 1, 16, 64])]; + tensor q_73_end_mask_0 = const()[name = tensor("q_73_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor q_73_squeeze_mask_0 = const()[name = tensor("q_73_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor q_73_cast_fp16 = slice_by_index(begin = q_73_begin_0, end = q_73_end_0, end_mask = q_73_end_mask_0, squeeze_mask = q_73_squeeze_mask_0, x = qkv_25_cast_fp16)[name = tensor("q_73_cast_fp16")]; + tensor k_49_begin_0 = const()[name = tensor("k_49_begin_0"), val = tensor([0, 0, 1, 0, 0])]; + tensor k_49_end_0 = const()[name = tensor("k_49_end_0"), val = tensor([1, 1, 2, 16, 64])]; + tensor k_49_end_mask_0 = const()[name = tensor("k_49_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor k_49_squeeze_mask_0 = const()[name = tensor("k_49_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor k_49_cast_fp16 = slice_by_index(begin = k_49_begin_0, end = k_49_end_0, end_mask = k_49_end_mask_0, squeeze_mask = k_49_squeeze_mask_0, x = qkv_25_cast_fp16)[name = tensor("k_49_cast_fp16")]; + tensor v_25_begin_0 = const()[name = tensor("v_25_begin_0"), val = tensor([0, 0, 2, 0, 0])]; + tensor v_25_end_0 = const()[name = tensor("v_25_end_0"), val = tensor([1, 1, 3, 16, 64])]; + tensor v_25_end_mask_0 = const()[name = tensor("v_25_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor v_25_squeeze_mask_0 = const()[name = tensor("v_25_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor v_25_cast_fp16 = slice_by_index(begin = v_25_begin_0, end = v_25_end_0, end_mask = v_25_end_mask_0, squeeze_mask = v_25_squeeze_mask_0, x = qkv_25_cast_fp16)[name = tensor("v_25_cast_fp16")]; + tensor var_5372 = const()[name = tensor("op_5372"), val = tensor([1, 1, 1, 1])]; + tensor position12_to_fp16_dtype_0 = const()[name = tensor("position12_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor position12_to_fp16 = cast(dtype = position12_to_fp16_dtype_0, x = position12)[name = tensor("cast_493")]; + tensor ts_77_cast_fp16 = reshape(shape = var_5372, x = position12_to_fp16)[name = tensor("ts_77_cast_fp16")]; + tensor var_5376 = const()[name = tensor("op_5376"), val = tensor([1, 1, 16, 32, 2])]; + tensor q_complex_25_cast_fp16 = reshape(shape = var_5376, x = q_73_cast_fp16)[name = tensor("q_complex_25_cast_fp16")]; + tensor var_5380 = const()[name = tensor("op_5380"), val = tensor([1, 1, 16, 32, 2])]; + tensor k_complex_25_cast_fp16 = reshape(shape = var_5380, x = k_49_cast_fp16)[name = tensor("k_complex_25_cast_fp16")]; + tensor var_5384_begin_0 = const()[name = tensor("op_5384_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_5384_end_0 = const()[name = tensor("op_5384_end_0"), val = tensor([1, 1, 16, 32, 1])]; + tensor var_5384_end_mask_0 = const()[name = tensor("op_5384_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_5384_squeeze_mask_0 = const()[name = tensor("op_5384_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_5384_cast_fp16 = slice_by_index(begin = var_5384_begin_0, end = var_5384_end_0, end_mask = var_5384_end_mask_0, squeeze_mask = var_5384_squeeze_mask_0, x = q_complex_25_cast_fp16)[name = tensor("op_5384_cast_fp16")]; + tensor var_5392_begin_0 = const()[name = tensor("op_5392_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_5392_end_0 = const()[name = tensor("op_5392_end_0"), val = tensor([1, 1, 16, 32, 2])]; + tensor var_5392_end_mask_0 = const()[name = tensor("op_5392_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_5392_squeeze_mask_0 = const()[name = tensor("op_5392_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_5392_cast_fp16 = slice_by_index(begin = var_5392_begin_0, end = var_5392_end_0, end_mask = var_5392_end_mask_0, squeeze_mask = var_5392_squeeze_mask_0, x = q_complex_25_cast_fp16)[name = tensor("op_5392_cast_fp16")]; + tensor var_5400_begin_0 = const()[name = tensor("op_5400_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_5400_end_0 = const()[name = tensor("op_5400_end_0"), val = tensor([1, 1, 16, 32, 1])]; + tensor var_5400_end_mask_0 = const()[name = tensor("op_5400_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_5400_squeeze_mask_0 = const()[name = tensor("op_5400_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_5400_cast_fp16 = slice_by_index(begin = var_5400_begin_0, end = var_5400_end_0, end_mask = var_5400_end_mask_0, squeeze_mask = var_5400_squeeze_mask_0, x = k_complex_25_cast_fp16)[name = tensor("op_5400_cast_fp16")]; + tensor var_5408_begin_0 = const()[name = tensor("op_5408_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_5408_end_0 = const()[name = tensor("op_5408_end_0"), val = tensor([1, 1, 16, 32, 2])]; + tensor var_5408_end_mask_0 = const()[name = tensor("op_5408_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_5408_squeeze_mask_0 = const()[name = tensor("op_5408_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_5408_cast_fp16 = slice_by_index(begin = var_5408_begin_0, end = var_5408_end_0, end_mask = var_5408_end_mask_0, squeeze_mask = var_5408_squeeze_mask_0, x = k_complex_25_cast_fp16)[name = tensor("op_5408_cast_fp16")]; + tensor freqs_25_to_fp16 = const()[name = tensor("freqs_25_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(303692032)))]; + tensor var_5414_cast_fp16 = mul(x = freqs_25_to_fp16, y = ts_77_cast_fp16)[name = tensor("op_5414_cast_fp16")]; + tensor rotr_25_cast_fp16 = cos(x = var_5414_cast_fp16)[name = tensor("rotr_25_cast_fp16")]; + tensor roti_25_cast_fp16 = sin(x = var_5414_cast_fp16)[name = tensor("roti_25_cast_fp16")]; + tensor var_5418_cast_fp16 = mul(x = var_5384_cast_fp16, y = rotr_25_cast_fp16)[name = tensor("op_5418_cast_fp16")]; + tensor var_5419_cast_fp16 = mul(x = var_5392_cast_fp16, y = roti_25_cast_fp16)[name = tensor("op_5419_cast_fp16")]; + tensor qor_49_cast_fp16 = sub(x = var_5418_cast_fp16, y = var_5419_cast_fp16)[name = tensor("qor_49_cast_fp16")]; + tensor var_5422_cast_fp16 = mul(x = var_5384_cast_fp16, y = roti_25_cast_fp16)[name = tensor("op_5422_cast_fp16")]; + tensor var_5423_cast_fp16 = mul(x = var_5392_cast_fp16, y = rotr_25_cast_fp16)[name = tensor("op_5423_cast_fp16")]; + tensor qoi_49_cast_fp16 = add(x = var_5422_cast_fp16, y = var_5423_cast_fp16)[name = tensor("qoi_49_cast_fp16")]; + tensor var_5426_cast_fp16 = mul(x = var_5400_cast_fp16, y = rotr_25_cast_fp16)[name = tensor("op_5426_cast_fp16")]; + tensor var_5427_cast_fp16 = mul(x = var_5408_cast_fp16, y = roti_25_cast_fp16)[name = tensor("op_5427_cast_fp16")]; + tensor kor_49_cast_fp16 = sub(x = var_5426_cast_fp16, y = var_5427_cast_fp16)[name = tensor("kor_49_cast_fp16")]; + tensor var_5430_cast_fp16 = mul(x = var_5400_cast_fp16, y = roti_25_cast_fp16)[name = tensor("op_5430_cast_fp16")]; + tensor var_5431_cast_fp16 = mul(x = var_5408_cast_fp16, y = rotr_25_cast_fp16)[name = tensor("op_5431_cast_fp16")]; + tensor koi_49_cast_fp16 = add(x = var_5430_cast_fp16, y = var_5431_cast_fp16)[name = tensor("koi_49_cast_fp16")]; + tensor qo_25_axis_0 = const()[name = tensor("qo_25_axis_0"), val = tensor(-1)]; + tensor qo_25_cast_fp16 = stack(axis = qo_25_axis_0, values = (qor_49_cast_fp16, qoi_49_cast_fp16))[name = tensor("qo_25_cast_fp16")]; + tensor ko_25_axis_0 = const()[name = tensor("ko_25_axis_0"), val = tensor(-1)]; + tensor ko_25_cast_fp16 = stack(axis = ko_25_axis_0, values = (kor_49_cast_fp16, koi_49_cast_fp16))[name = tensor("ko_25_cast_fp16")]; + tensor var_5460 = const()[name = tensor("op_5460"), val = tensor([1, 1, 16, 64])]; + tensor q_75_cast_fp16 = reshape(shape = var_5460, x = qo_25_cast_fp16)[name = tensor("q_75_cast_fp16")]; + tensor var_5462 = const()[name = tensor("op_5462"), val = tensor([1, 1, 16, 64])]; + tensor k_51_cast_fp16 = reshape(shape = var_5462, x = ko_25_cast_fp16)[name = tensor("k_51_cast_fp16")]; + tensor _inversed_5484_y_0_to_fp16 = const()[name = tensor("_inversed_5484_y_0_to_fp16"), val = tensor(0x1p-9)]; + tensor _inversed_5484_cast_fp16 = mul(x = ts_77_cast_fp16, y = _inversed_5484_y_0_to_fp16)[name = tensor("_inversed_5484_cast_fp16")]; + tensor var_5485_cast_fp16 = floor(x = _inversed_5484_cast_fp16)[name = tensor("op_5485_cast_fp16")]; + tensor var_5486_to_fp16 = const()[name = tensor("op_5486_to_fp16"), val = tensor(0x1p+9)]; + tensor var_5487_cast_fp16 = mul(x = var_5485_cast_fp16, y = var_5486_to_fp16)[name = tensor("op_5487_cast_fp16")]; + tensor write_indices_float_51_cast_fp16 = sub(x = ts_77_cast_fp16, y = var_5487_cast_fp16)[name = tensor("write_indices_float_51_cast_fp16")]; + tensor var_5494_dtype_0 = const()[name = tensor("op_5494_dtype_0"), val = tensor("int32")]; + tensor write_indices_25_reps_0 = const()[name = tensor("write_indices_25_reps_0"), val = tensor([1, 1, 16, 64])]; + tensor write_indices_float_51_cast_fp16_to_int32 = cast(dtype = var_5494_dtype_0, x = write_indices_float_51_cast_fp16)[name = tensor("cast_492")]; + tensor write_indices_25 = tile(reps = write_indices_25_reps_0, x = write_indices_float_51_cast_fp16_to_int32)[name = tensor("write_indices_25")]; + tensor var_5502_begin_0 = const()[name = tensor("op_5502_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_5502_end_0 = const()[name = tensor("op_5502_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor var_5502_end_mask_0 = const()[name = tensor("op_5502_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_5502_squeeze_mask_0 = const()[name = tensor("op_5502_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor cache12_to_fp16_dtype_0 = const()[name = tensor("cache12_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor cache12_to_fp16 = cast(dtype = cache12_to_fp16_dtype_0, x = cache12)[name = tensor("cast_491")]; + tensor var_5502_cast_fp16 = slice_by_index(begin = var_5502_begin_0, end = var_5502_end_0, end_mask = var_5502_end_mask_0, squeeze_mask = var_5502_squeeze_mask_0, x = cache12_to_fp16)[name = tensor("op_5502_cast_fp16")]; + tensor var_5504_axis_0 = const()[name = tensor("op_5504_axis_0"), val = tensor(1)]; + tensor var_5504_mode_0 = const()[name = tensor("op_5504_mode_0"), val = tensor("update")]; + tensor var_5504_validate_indices_0 = const()[name = tensor("op_5504_validate_indices_0"), val = tensor(false)]; + tensor var_5504_cast_fp16 = scatter_along_axis(axis = var_5504_axis_0, data = var_5502_cast_fp16, indices = write_indices_25, mode = var_5504_mode_0, updates = k_51_cast_fp16, validate_indices = var_5504_validate_indices_0)[name = tensor("op_5504_cast_fp16")]; + tensor concat_86 = const()[name = tensor("concat_86"), val = tensor([0, 0, 0, 0, 0])]; + tensor concat_87 = const()[name = tensor("concat_87"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_25_internal_tensor_assign_1_stride_0 = const()[name = tensor("new_cache_25_internal_tensor_assign_1_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_25_internal_tensor_assign_1_begin_mask_0 = const()[name = tensor("new_cache_25_internal_tensor_assign_1_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_25_internal_tensor_assign_1_end_mask_0 = const()[name = tensor("new_cache_25_internal_tensor_assign_1_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_25_internal_tensor_assign_1_squeeze_mask_0 = const()[name = tensor("new_cache_25_internal_tensor_assign_1_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_72 = const()[name = tensor("shape_72"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_24 = const()[name = tensor("reduce_prod_24"), val = tensor(1048576)]; + tensor range_1d_24_start_0 = const()[name = tensor("range_1d_24_start_0"), val = tensor(0)]; + tensor range_1d_24_step_0 = const()[name = tensor("range_1d_24_step_0"), val = tensor(1)]; + tensor range_1d_24 = range_1d(end = reduce_prod_24, start = range_1d_24_start_0, step = range_1d_24_step_0)[name = tensor("range_1d_24")]; + tensor reshape_120 = reshape(shape = shape_72, x = range_1d_24)[name = tensor("reshape_120")]; + tensor slice_by_index_24 = slice_by_index(begin = concat_86, begin_mask = new_cache_25_internal_tensor_assign_1_begin_mask_0, end = concat_87, end_mask = new_cache_25_internal_tensor_assign_1_end_mask_0, squeeze_mask = new_cache_25_internal_tensor_assign_1_squeeze_mask_0, stride = new_cache_25_internal_tensor_assign_1_stride_0, x = reshape_120)[name = tensor("slice_by_index_24")]; + tensor reshape_121_shape_0 = const()[name = tensor("reshape_121_shape_0"), val = tensor([-1])]; + tensor reshape_121 = reshape(shape = reshape_121_shape_0, x = slice_by_index_24)[name = tensor("reshape_121")]; + tensor reshape_122_shape_0 = const()[name = tensor("reshape_122_shape_0"), val = tensor([-1])]; + tensor reshape_122_cast_fp16 = reshape(shape = reshape_122_shape_0, x = var_5504_cast_fp16)[name = tensor("reshape_122_cast_fp16")]; + tensor reshape_123_shape_0 = const()[name = tensor("reshape_123_shape_0"), val = tensor([-1])]; + tensor reshape_123_cast_fp16 = reshape(shape = reshape_123_shape_0, x = cache12_to_fp16)[name = tensor("reshape_123_cast_fp16")]; + tensor scatter_24_mode_0 = const()[name = tensor("scatter_24_mode_0"), val = tensor("update")]; + tensor scatter_24_axis_0 = const()[name = tensor("scatter_24_axis_0"), val = tensor(0)]; + tensor scatter_24_validate_indices_0 = const()[name = tensor("scatter_24_validate_indices_0"), val = tensor(false)]; + tensor scatter_24_cast_fp16 = scatter(axis = scatter_24_axis_0, data = reshape_123_cast_fp16, indices = reshape_121, mode = scatter_24_mode_0, updates = reshape_122_cast_fp16, validate_indices = scatter_24_validate_indices_0)[name = tensor("scatter_24_cast_fp16")]; + tensor reshape_124_cast_fp16 = reshape(shape = shape_72, x = scatter_24_cast_fp16)[name = tensor("reshape_124_cast_fp16")]; + tensor var_5512_begin_0 = const()[name = tensor("op_5512_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor var_5512_end_0 = const()[name = tensor("op_5512_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor var_5512_end_mask_0 = const()[name = tensor("op_5512_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_5512_squeeze_mask_0 = const()[name = tensor("op_5512_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor var_5512_cast_fp16 = slice_by_index(begin = var_5512_begin_0, end = var_5512_end_0, end_mask = var_5512_end_mask_0, squeeze_mask = var_5512_squeeze_mask_0, x = reshape_124_cast_fp16)[name = tensor("op_5512_cast_fp16")]; + tensor var_5514_axis_0 = const()[name = tensor("op_5514_axis_0"), val = tensor(1)]; + tensor var_5514_mode_0 = const()[name = tensor("op_5514_mode_0"), val = tensor("update")]; + tensor var_5514_validate_indices_0 = const()[name = tensor("op_5514_validate_indices_0"), val = tensor(false)]; + tensor var_5514_cast_fp16 = scatter_along_axis(axis = var_5514_axis_0, data = var_5512_cast_fp16, indices = write_indices_25, mode = var_5514_mode_0, updates = v_25_cast_fp16, validate_indices = var_5514_validate_indices_0)[name = tensor("op_5514_cast_fp16")]; + tensor concat_88 = const()[name = tensor("concat_88"), val = tensor([1, 0, 0, 0, 0])]; + tensor concat_89 = const()[name = tensor("concat_89"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_25_internal_tensor_assign_2_stride_0 = const()[name = tensor("new_cache_25_internal_tensor_assign_2_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_25_internal_tensor_assign_2_begin_mask_0 = const()[name = tensor("new_cache_25_internal_tensor_assign_2_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_25_internal_tensor_assign_2_end_mask_0 = const()[name = tensor("new_cache_25_internal_tensor_assign_2_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_25_internal_tensor_assign_2_squeeze_mask_0 = const()[name = tensor("new_cache_25_internal_tensor_assign_2_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_73 = const()[name = tensor("shape_73"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_25 = const()[name = tensor("reduce_prod_25"), val = tensor(1048576)]; + tensor range_1d_25_start_0 = const()[name = tensor("range_1d_25_start_0"), val = tensor(0)]; + tensor range_1d_25_step_0 = const()[name = tensor("range_1d_25_step_0"), val = tensor(1)]; + tensor range_1d_25 = range_1d(end = reduce_prod_25, start = range_1d_25_start_0, step = range_1d_25_step_0)[name = tensor("range_1d_25")]; + tensor reshape_125 = reshape(shape = shape_73, x = range_1d_25)[name = tensor("reshape_125")]; + tensor slice_by_index_25 = slice_by_index(begin = concat_88, begin_mask = new_cache_25_internal_tensor_assign_2_begin_mask_0, end = concat_89, end_mask = new_cache_25_internal_tensor_assign_2_end_mask_0, squeeze_mask = new_cache_25_internal_tensor_assign_2_squeeze_mask_0, stride = new_cache_25_internal_tensor_assign_2_stride_0, x = reshape_125)[name = tensor("slice_by_index_25")]; + tensor reshape_126_shape_0 = const()[name = tensor("reshape_126_shape_0"), val = tensor([-1])]; + tensor reshape_126 = reshape(shape = reshape_126_shape_0, x = slice_by_index_25)[name = tensor("reshape_126")]; + tensor reshape_127_shape_0 = const()[name = tensor("reshape_127_shape_0"), val = tensor([-1])]; + tensor reshape_127_cast_fp16 = reshape(shape = reshape_127_shape_0, x = var_5514_cast_fp16)[name = tensor("reshape_127_cast_fp16")]; + tensor reshape_128_shape_0 = const()[name = tensor("reshape_128_shape_0"), val = tensor([-1])]; + tensor reshape_128_cast_fp16 = reshape(shape = reshape_128_shape_0, x = reshape_124_cast_fp16)[name = tensor("reshape_128_cast_fp16")]; + tensor scatter_25_mode_0 = const()[name = tensor("scatter_25_mode_0"), val = tensor("update")]; + tensor scatter_25_axis_0 = const()[name = tensor("scatter_25_axis_0"), val = tensor(0)]; + tensor scatter_25_validate_indices_0 = const()[name = tensor("scatter_25_validate_indices_0"), val = tensor(false)]; + tensor scatter_25_cast_fp16 = scatter(axis = scatter_25_axis_0, data = reshape_128_cast_fp16, indices = reshape_126, mode = scatter_25_mode_0, updates = reshape_127_cast_fp16, validate_indices = scatter_25_validate_indices_0)[name = tensor("scatter_25_cast_fp16")]; + tensor reshape_129_cast_fp16 = reshape(shape = shape_73, x = scatter_25_cast_fp16)[name = tensor("reshape_129_cast_fp16")]; + tensor reshape_129_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("reshape_129_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor keys_73_begin_0 = const()[name = tensor("keys_73_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor keys_73_end_0 = const()[name = tensor("keys_73_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor keys_73_end_mask_0 = const()[name = tensor("keys_73_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor keys_73_squeeze_mask_0 = const()[name = tensor("keys_73_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor keys_73_cast_fp16 = slice_by_index(begin = keys_73_begin_0, end = keys_73_end_0, end_mask = keys_73_end_mask_0, squeeze_mask = keys_73_squeeze_mask_0, x = reshape_129_cast_fp16)[name = tensor("keys_73_cast_fp16")]; + tensor values_73_begin_0 = const()[name = tensor("values_73_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor values_73_end_0 = const()[name = tensor("values_73_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor values_73_end_mask_0 = const()[name = tensor("values_73_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor values_73_squeeze_mask_0 = const()[name = tensor("values_73_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor values_73_cast_fp16 = slice_by_index(begin = values_73_begin_0, end = values_73_end_0, end_mask = values_73_end_mask_0, squeeze_mask = values_73_squeeze_mask_0, x = reshape_129_cast_fp16)[name = tensor("values_73_cast_fp16")]; + tensor var_5526_cast_fp16 = not_equal(x = keys_73_cast_fp16, y = keys_73_cast_fp16)[name = tensor("op_5526_cast_fp16")]; + tensor keys_75_cast_fp16 = select(a = var_504_to_fp16, b = keys_73_cast_fp16, cond = var_5526_cast_fp16)[name = tensor("keys_75_cast_fp16")]; + tensor var_5534_cast_fp16 = not_equal(x = values_73_cast_fp16, y = values_73_cast_fp16)[name = tensor("op_5534_cast_fp16")]; + tensor values_75_cast_fp16 = select(a = var_504_to_fp16, b = values_73_cast_fp16, cond = var_5534_cast_fp16)[name = tensor("values_75_cast_fp16")]; + tensor var_5558 = const()[name = tensor("op_5558"), val = tensor([0, 2, 1, 3])]; + tensor var_5571 = const()[name = tensor("op_5571"), val = tensor([1, 1, 1])]; + tensor var_5572_cast_fp16 = reshape(shape = var_5571, x = position12_to_fp16)[name = tensor("op_5572_cast_fp16")]; + tensor var_5589_to_fp16 = const()[name = tensor("op_5589_to_fp16"), val = tensor(0x1p+0)]; + tensor valid_len_25_cast_fp16 = add(x = var_5572_cast_fp16, y = var_5589_to_fp16)[name = tensor("valid_len_25_cast_fp16")]; + tensor valid_mask_25_cast_fp16 = less(x = k_positions_1_promoted_to_fp16, y = valid_len_25_cast_fp16)[name = tensor("valid_mask_25_cast_fp16")]; + tensor causal_mask_25_cast_fp16 = less_equal(x = k_positions_1_promoted_to_fp16, y = var_5572_cast_fp16)[name = tensor("causal_mask_25_cast_fp16")]; + tensor attn_mask_49 = logical_and(x = valid_mask_25_cast_fp16, y = causal_mask_25_cast_fp16)[name = tensor("attn_mask_49")]; + tensor attn_mask_51_axes_0 = const()[name = tensor("attn_mask_51_axes_0"), val = tensor([1])]; + tensor attn_mask_51 = expand_dims(axes = attn_mask_51_axes_0, x = attn_mask_49)[name = tensor("attn_mask_51")]; + tensor var_5607_transpose_x_0 = const()[name = tensor("op_5607_transpose_x_0"), val = tensor(false)]; + tensor var_5607_transpose_y_0 = const()[name = tensor("op_5607_transpose_y_0"), val = tensor(false)]; + tensor transpose_96_perm_0 = const()[name = tensor("transpose_96_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_97_perm_0 = const()[name = tensor("transpose_97_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_97 = transpose(perm = transpose_97_perm_0, x = keys_75_cast_fp16)[name = tensor("transpose_165")]; + tensor transpose_96 = transpose(perm = transpose_96_perm_0, x = q_75_cast_fp16)[name = tensor("transpose_166")]; + tensor var_5607_cast_fp16 = matmul(transpose_x = var_5607_transpose_x_0, transpose_y = var_5607_transpose_y_0, x = transpose_96, y = transpose_97)[name = tensor("op_5607_cast_fp16")]; + tensor var_5601_to_fp16 = const()[name = tensor("op_5601_to_fp16"), val = tensor([0x1p-3])]; + tensor attn_weights_73_cast_fp16 = mul(x = var_5607_cast_fp16, y = var_5601_to_fp16)[name = tensor("attn_weights_73_cast_fp16")]; + tensor var_5609 = logical_not(x = attn_mask_51)[name = tensor("op_5609")]; + tensor var_5610_to_fp16 = const()[name = tensor("op_5610_to_fp16"), val = tensor(-inf)]; + tensor attn_weights_75_cast_fp16 = select(a = var_5610_to_fp16, b = attn_weights_73_cast_fp16, cond = var_5609)[name = tensor("attn_weights_75_cast_fp16")]; + tensor var_5612 = const()[name = tensor("op_5612"), val = tensor(-1)]; + tensor attn_weights_77_cast_fp16 = softmax(axis = var_5612, x = attn_weights_75_cast_fp16)[name = tensor("attn_weights_77_cast_fp16")]; + tensor attn_output_25_transpose_x_0 = const()[name = tensor("attn_output_25_transpose_x_0"), val = tensor(false)]; + tensor attn_output_25_transpose_y_0 = const()[name = tensor("attn_output_25_transpose_y_0"), val = tensor(false)]; + tensor values_77_cast_fp16 = transpose(perm = var_5558, x = values_75_cast_fp16)[name = tensor("transpose_167")]; + tensor attn_output_25_cast_fp16 = matmul(transpose_x = attn_output_25_transpose_x_0, transpose_y = attn_output_25_transpose_y_0, x = attn_weights_77_cast_fp16, y = values_77_cast_fp16)[name = tensor("attn_output_25_cast_fp16")]; + tensor var_5620 = const()[name = tensor("op_5620"), val = tensor([0, 2, 1, 3])]; + tensor var_5623 = const()[name = tensor("op_5623"), val = tensor([1, 1, 1024])]; + tensor var_5621_cast_fp16 = transpose(perm = var_5620, x = attn_output_25_cast_fp16)[name = tensor("transpose_164")]; + tensor input_125_cast_fp16 = reshape(shape = var_5623, x = var_5621_cast_fp16)[name = tensor("input_125_cast_fp16")]; + tensor linear_50_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = attn12_out_proj_weight_cast_fp16, x = input_125_cast_fp16)[name = tensor("linear_50_cast_fp16")]; + tensor var_5629_to_fp16 = const()[name = tensor("op_5629_to_fp16"), val = tensor(0x1p+0)]; + tensor var_5630_cast_fp16 = add(x = position12_to_fp16, y = var_5629_to_fp16)[name = tensor("op_5630_cast_fp16")]; + tensor var_5630_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_5630_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor input_127_cast_fp16 = add(x = input_123_cast_fp16, y = linear_50_cast_fp16)[name = tensor("input_127_cast_fp16")]; + tensor input_129_axes_0 = const()[name = tensor("input_129_axes_0"), val = tensor([-1])]; + tensor norm12_2_weight_to_fp16 = const()[name = tensor("norm12_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(303692160)))]; + tensor norm12_2_bias_to_fp16 = const()[name = tensor("norm12_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(303694272)))]; + tensor var_5634_to_fp16 = const()[name = tensor("op_5634_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_129_cast_fp16 = layer_norm(axes = input_129_axes_0, beta = norm12_2_bias_to_fp16, epsilon = var_5634_to_fp16, gamma = norm12_2_weight_to_fp16, x = input_127_cast_fp16)[name = tensor("input_129_cast_fp16")]; + tensor linear_51_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = linear12_1_weight_cast_fp16, x = input_129_cast_fp16)[name = tensor("linear_51_cast_fp16")]; + tensor input_131_mode_0 = const()[name = tensor("input_131_mode_0"), val = tensor("EXACT")]; + tensor input_131_cast_fp16 = gelu(mode = input_131_mode_0, x = linear_51_cast_fp16)[name = tensor("input_131_cast_fp16")]; + tensor linear_52_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = linear12_2_weight_cast_fp16, x = input_131_cast_fp16)[name = tensor("linear_52_cast_fp16")]; + tensor input_133_cast_fp16 = add(x = input_127_cast_fp16, y = linear_52_cast_fp16)[name = tensor("input_133_cast_fp16")]; + tensor x_27_axes_0 = const()[name = tensor("x_27_axes_0"), val = tensor([-1])]; + tensor norm13_1_weight_to_fp16 = const()[name = tensor("norm13_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(303696384)))]; + tensor norm13_1_bias_to_fp16 = const()[name = tensor("norm13_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(303698496)))]; + tensor var_5651_to_fp16 = const()[name = tensor("op_5651_to_fp16"), val = tensor(0x1.5p-17)]; + tensor x_27_cast_fp16 = layer_norm(axes = x_27_axes_0, beta = norm13_1_bias_to_fp16, epsilon = var_5651_to_fp16, gamma = norm13_1_weight_to_fp16, x = input_133_cast_fp16)[name = tensor("x_27_cast_fp16")]; + tensor linear_53_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = attn13_in_proj_weight_cast_fp16, x = x_27_cast_fp16)[name = tensor("linear_53_cast_fp16")]; + tensor var_5687 = const()[name = tensor("op_5687"), val = tensor([1, 1, 3, 16, 64])]; + tensor qkv_27_cast_fp16 = reshape(shape = var_5687, x = linear_53_cast_fp16)[name = tensor("qkv_27_cast_fp16")]; + tensor q_79_begin_0 = const()[name = tensor("q_79_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor q_79_end_0 = const()[name = tensor("q_79_end_0"), val = tensor([1, 1, 1, 16, 64])]; + tensor q_79_end_mask_0 = const()[name = tensor("q_79_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor q_79_squeeze_mask_0 = const()[name = tensor("q_79_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor q_79_cast_fp16 = slice_by_index(begin = q_79_begin_0, end = q_79_end_0, end_mask = q_79_end_mask_0, squeeze_mask = q_79_squeeze_mask_0, x = qkv_27_cast_fp16)[name = tensor("q_79_cast_fp16")]; + tensor k_53_begin_0 = const()[name = tensor("k_53_begin_0"), val = tensor([0, 0, 1, 0, 0])]; + tensor k_53_end_0 = const()[name = tensor("k_53_end_0"), val = tensor([1, 1, 2, 16, 64])]; + tensor k_53_end_mask_0 = const()[name = tensor("k_53_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor k_53_squeeze_mask_0 = const()[name = tensor("k_53_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor k_53_cast_fp16 = slice_by_index(begin = k_53_begin_0, end = k_53_end_0, end_mask = k_53_end_mask_0, squeeze_mask = k_53_squeeze_mask_0, x = qkv_27_cast_fp16)[name = tensor("k_53_cast_fp16")]; + tensor v_27_begin_0 = const()[name = tensor("v_27_begin_0"), val = tensor([0, 0, 2, 0, 0])]; + tensor v_27_end_0 = const()[name = tensor("v_27_end_0"), val = tensor([1, 1, 3, 16, 64])]; + tensor v_27_end_mask_0 = const()[name = tensor("v_27_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor v_27_squeeze_mask_0 = const()[name = tensor("v_27_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor v_27_cast_fp16 = slice_by_index(begin = v_27_begin_0, end = v_27_end_0, end_mask = v_27_end_mask_0, squeeze_mask = v_27_squeeze_mask_0, x = qkv_27_cast_fp16)[name = tensor("v_27_cast_fp16")]; + tensor var_5791 = const()[name = tensor("op_5791"), val = tensor([1, 1, 1, 1])]; + tensor position13_to_fp16_dtype_0 = const()[name = tensor("position13_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor position13_to_fp16 = cast(dtype = position13_to_fp16_dtype_0, x = position13)[name = tensor("cast_488")]; + tensor ts_83_cast_fp16 = reshape(shape = var_5791, x = position13_to_fp16)[name = tensor("ts_83_cast_fp16")]; + tensor var_5795 = const()[name = tensor("op_5795"), val = tensor([1, 1, 16, 32, 2])]; + tensor q_complex_27_cast_fp16 = reshape(shape = var_5795, x = q_79_cast_fp16)[name = tensor("q_complex_27_cast_fp16")]; + tensor var_5799 = const()[name = tensor("op_5799"), val = tensor([1, 1, 16, 32, 2])]; + tensor k_complex_27_cast_fp16 = reshape(shape = var_5799, x = k_53_cast_fp16)[name = tensor("k_complex_27_cast_fp16")]; + tensor var_5803_begin_0 = const()[name = tensor("op_5803_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_5803_end_0 = const()[name = tensor("op_5803_end_0"), val = tensor([1, 1, 16, 32, 1])]; + tensor var_5803_end_mask_0 = const()[name = tensor("op_5803_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_5803_squeeze_mask_0 = const()[name = tensor("op_5803_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_5803_cast_fp16 = slice_by_index(begin = var_5803_begin_0, end = var_5803_end_0, end_mask = var_5803_end_mask_0, squeeze_mask = var_5803_squeeze_mask_0, x = q_complex_27_cast_fp16)[name = tensor("op_5803_cast_fp16")]; + tensor var_5811_begin_0 = const()[name = tensor("op_5811_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_5811_end_0 = const()[name = tensor("op_5811_end_0"), val = tensor([1, 1, 16, 32, 2])]; + tensor var_5811_end_mask_0 = const()[name = tensor("op_5811_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_5811_squeeze_mask_0 = const()[name = tensor("op_5811_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_5811_cast_fp16 = slice_by_index(begin = var_5811_begin_0, end = var_5811_end_0, end_mask = var_5811_end_mask_0, squeeze_mask = var_5811_squeeze_mask_0, x = q_complex_27_cast_fp16)[name = tensor("op_5811_cast_fp16")]; + tensor var_5819_begin_0 = const()[name = tensor("op_5819_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_5819_end_0 = const()[name = tensor("op_5819_end_0"), val = tensor([1, 1, 16, 32, 1])]; + tensor var_5819_end_mask_0 = const()[name = tensor("op_5819_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_5819_squeeze_mask_0 = const()[name = tensor("op_5819_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_5819_cast_fp16 = slice_by_index(begin = var_5819_begin_0, end = var_5819_end_0, end_mask = var_5819_end_mask_0, squeeze_mask = var_5819_squeeze_mask_0, x = k_complex_27_cast_fp16)[name = tensor("op_5819_cast_fp16")]; + tensor var_5827_begin_0 = const()[name = tensor("op_5827_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_5827_end_0 = const()[name = tensor("op_5827_end_0"), val = tensor([1, 1, 16, 32, 2])]; + tensor var_5827_end_mask_0 = const()[name = tensor("op_5827_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_5827_squeeze_mask_0 = const()[name = tensor("op_5827_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_5827_cast_fp16 = slice_by_index(begin = var_5827_begin_0, end = var_5827_end_0, end_mask = var_5827_end_mask_0, squeeze_mask = var_5827_squeeze_mask_0, x = k_complex_27_cast_fp16)[name = tensor("op_5827_cast_fp16")]; + tensor freqs_27_to_fp16 = const()[name = tensor("freqs_27_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(303700608)))]; + tensor var_5833_cast_fp16 = mul(x = freqs_27_to_fp16, y = ts_83_cast_fp16)[name = tensor("op_5833_cast_fp16")]; + tensor rotr_27_cast_fp16 = cos(x = var_5833_cast_fp16)[name = tensor("rotr_27_cast_fp16")]; + tensor roti_27_cast_fp16 = sin(x = var_5833_cast_fp16)[name = tensor("roti_27_cast_fp16")]; + tensor var_5837_cast_fp16 = mul(x = var_5803_cast_fp16, y = rotr_27_cast_fp16)[name = tensor("op_5837_cast_fp16")]; + tensor var_5838_cast_fp16 = mul(x = var_5811_cast_fp16, y = roti_27_cast_fp16)[name = tensor("op_5838_cast_fp16")]; + tensor qor_53_cast_fp16 = sub(x = var_5837_cast_fp16, y = var_5838_cast_fp16)[name = tensor("qor_53_cast_fp16")]; + tensor var_5841_cast_fp16 = mul(x = var_5803_cast_fp16, y = roti_27_cast_fp16)[name = tensor("op_5841_cast_fp16")]; + tensor var_5842_cast_fp16 = mul(x = var_5811_cast_fp16, y = rotr_27_cast_fp16)[name = tensor("op_5842_cast_fp16")]; + tensor qoi_53_cast_fp16 = add(x = var_5841_cast_fp16, y = var_5842_cast_fp16)[name = tensor("qoi_53_cast_fp16")]; + tensor var_5845_cast_fp16 = mul(x = var_5819_cast_fp16, y = rotr_27_cast_fp16)[name = tensor("op_5845_cast_fp16")]; + tensor var_5846_cast_fp16 = mul(x = var_5827_cast_fp16, y = roti_27_cast_fp16)[name = tensor("op_5846_cast_fp16")]; + tensor kor_53_cast_fp16 = sub(x = var_5845_cast_fp16, y = var_5846_cast_fp16)[name = tensor("kor_53_cast_fp16")]; + tensor var_5849_cast_fp16 = mul(x = var_5819_cast_fp16, y = roti_27_cast_fp16)[name = tensor("op_5849_cast_fp16")]; + tensor var_5850_cast_fp16 = mul(x = var_5827_cast_fp16, y = rotr_27_cast_fp16)[name = tensor("op_5850_cast_fp16")]; + tensor koi_53_cast_fp16 = add(x = var_5849_cast_fp16, y = var_5850_cast_fp16)[name = tensor("koi_53_cast_fp16")]; + tensor qo_27_axis_0 = const()[name = tensor("qo_27_axis_0"), val = tensor(-1)]; + tensor qo_27_cast_fp16 = stack(axis = qo_27_axis_0, values = (qor_53_cast_fp16, qoi_53_cast_fp16))[name = tensor("qo_27_cast_fp16")]; + tensor ko_27_axis_0 = const()[name = tensor("ko_27_axis_0"), val = tensor(-1)]; + tensor ko_27_cast_fp16 = stack(axis = ko_27_axis_0, values = (kor_53_cast_fp16, koi_53_cast_fp16))[name = tensor("ko_27_cast_fp16")]; + tensor var_5879 = const()[name = tensor("op_5879"), val = tensor([1, 1, 16, 64])]; + tensor q_81_cast_fp16 = reshape(shape = var_5879, x = qo_27_cast_fp16)[name = tensor("q_81_cast_fp16")]; + tensor var_5881 = const()[name = tensor("op_5881"), val = tensor([1, 1, 16, 64])]; + tensor k_55_cast_fp16 = reshape(shape = var_5881, x = ko_27_cast_fp16)[name = tensor("k_55_cast_fp16")]; + tensor _inversed_5903_y_0_to_fp16 = const()[name = tensor("_inversed_5903_y_0_to_fp16"), val = tensor(0x1p-9)]; + tensor _inversed_5903_cast_fp16 = mul(x = ts_83_cast_fp16, y = _inversed_5903_y_0_to_fp16)[name = tensor("_inversed_5903_cast_fp16")]; + tensor var_5904_cast_fp16 = floor(x = _inversed_5903_cast_fp16)[name = tensor("op_5904_cast_fp16")]; + tensor var_5905_to_fp16 = const()[name = tensor("op_5905_to_fp16"), val = tensor(0x1p+9)]; + tensor var_5906_cast_fp16 = mul(x = var_5904_cast_fp16, y = var_5905_to_fp16)[name = tensor("op_5906_cast_fp16")]; + tensor write_indices_float_55_cast_fp16 = sub(x = ts_83_cast_fp16, y = var_5906_cast_fp16)[name = tensor("write_indices_float_55_cast_fp16")]; + tensor var_5913_dtype_0 = const()[name = tensor("op_5913_dtype_0"), val = tensor("int32")]; + tensor write_indices_27_reps_0 = const()[name = tensor("write_indices_27_reps_0"), val = tensor([1, 1, 16, 64])]; + tensor write_indices_float_55_cast_fp16_to_int32 = cast(dtype = var_5913_dtype_0, x = write_indices_float_55_cast_fp16)[name = tensor("cast_487")]; + tensor write_indices_27 = tile(reps = write_indices_27_reps_0, x = write_indices_float_55_cast_fp16_to_int32)[name = tensor("write_indices_27")]; + tensor var_5921_begin_0 = const()[name = tensor("op_5921_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_5921_end_0 = const()[name = tensor("op_5921_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor var_5921_end_mask_0 = const()[name = tensor("op_5921_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_5921_squeeze_mask_0 = const()[name = tensor("op_5921_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor cache13_to_fp16_dtype_0 = const()[name = tensor("cache13_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor cache13_to_fp16 = cast(dtype = cache13_to_fp16_dtype_0, x = cache13)[name = tensor("cast_486")]; + tensor var_5921_cast_fp16 = slice_by_index(begin = var_5921_begin_0, end = var_5921_end_0, end_mask = var_5921_end_mask_0, squeeze_mask = var_5921_squeeze_mask_0, x = cache13_to_fp16)[name = tensor("op_5921_cast_fp16")]; + tensor var_5923_axis_0 = const()[name = tensor("op_5923_axis_0"), val = tensor(1)]; + tensor var_5923_mode_0 = const()[name = tensor("op_5923_mode_0"), val = tensor("update")]; + tensor var_5923_validate_indices_0 = const()[name = tensor("op_5923_validate_indices_0"), val = tensor(false)]; + tensor var_5923_cast_fp16 = scatter_along_axis(axis = var_5923_axis_0, data = var_5921_cast_fp16, indices = write_indices_27, mode = var_5923_mode_0, updates = k_55_cast_fp16, validate_indices = var_5923_validate_indices_0)[name = tensor("op_5923_cast_fp16")]; + tensor concat_93 = const()[name = tensor("concat_93"), val = tensor([0, 0, 0, 0, 0])]; + tensor concat_94 = const()[name = tensor("concat_94"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_27_internal_tensor_assign_1_stride_0 = const()[name = tensor("new_cache_27_internal_tensor_assign_1_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_27_internal_tensor_assign_1_begin_mask_0 = const()[name = tensor("new_cache_27_internal_tensor_assign_1_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_27_internal_tensor_assign_1_end_mask_0 = const()[name = tensor("new_cache_27_internal_tensor_assign_1_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_27_internal_tensor_assign_1_squeeze_mask_0 = const()[name = tensor("new_cache_27_internal_tensor_assign_1_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_74 = const()[name = tensor("shape_74"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_26 = const()[name = tensor("reduce_prod_26"), val = tensor(1048576)]; + tensor range_1d_26_start_0 = const()[name = tensor("range_1d_26_start_0"), val = tensor(0)]; + tensor range_1d_26_step_0 = const()[name = tensor("range_1d_26_step_0"), val = tensor(1)]; + tensor range_1d_26 = range_1d(end = reduce_prod_26, start = range_1d_26_start_0, step = range_1d_26_step_0)[name = tensor("range_1d_26")]; + tensor reshape_130 = reshape(shape = shape_74, x = range_1d_26)[name = tensor("reshape_130")]; + tensor slice_by_index_26 = slice_by_index(begin = concat_93, begin_mask = new_cache_27_internal_tensor_assign_1_begin_mask_0, end = concat_94, end_mask = new_cache_27_internal_tensor_assign_1_end_mask_0, squeeze_mask = new_cache_27_internal_tensor_assign_1_squeeze_mask_0, stride = new_cache_27_internal_tensor_assign_1_stride_0, x = reshape_130)[name = tensor("slice_by_index_26")]; + tensor reshape_131_shape_0 = const()[name = tensor("reshape_131_shape_0"), val = tensor([-1])]; + tensor reshape_131 = reshape(shape = reshape_131_shape_0, x = slice_by_index_26)[name = tensor("reshape_131")]; + tensor reshape_132_shape_0 = const()[name = tensor("reshape_132_shape_0"), val = tensor([-1])]; + tensor reshape_132_cast_fp16 = reshape(shape = reshape_132_shape_0, x = var_5923_cast_fp16)[name = tensor("reshape_132_cast_fp16")]; + tensor reshape_133_shape_0 = const()[name = tensor("reshape_133_shape_0"), val = tensor([-1])]; + tensor reshape_133_cast_fp16 = reshape(shape = reshape_133_shape_0, x = cache13_to_fp16)[name = tensor("reshape_133_cast_fp16")]; + tensor scatter_26_mode_0 = const()[name = tensor("scatter_26_mode_0"), val = tensor("update")]; + tensor scatter_26_axis_0 = const()[name = tensor("scatter_26_axis_0"), val = tensor(0)]; + tensor scatter_26_validate_indices_0 = const()[name = tensor("scatter_26_validate_indices_0"), val = tensor(false)]; + tensor scatter_26_cast_fp16 = scatter(axis = scatter_26_axis_0, data = reshape_133_cast_fp16, indices = reshape_131, mode = scatter_26_mode_0, updates = reshape_132_cast_fp16, validate_indices = scatter_26_validate_indices_0)[name = tensor("scatter_26_cast_fp16")]; + tensor reshape_134_cast_fp16 = reshape(shape = shape_74, x = scatter_26_cast_fp16)[name = tensor("reshape_134_cast_fp16")]; + tensor var_5931_begin_0 = const()[name = tensor("op_5931_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor var_5931_end_0 = const()[name = tensor("op_5931_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor var_5931_end_mask_0 = const()[name = tensor("op_5931_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_5931_squeeze_mask_0 = const()[name = tensor("op_5931_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor var_5931_cast_fp16 = slice_by_index(begin = var_5931_begin_0, end = var_5931_end_0, end_mask = var_5931_end_mask_0, squeeze_mask = var_5931_squeeze_mask_0, x = reshape_134_cast_fp16)[name = tensor("op_5931_cast_fp16")]; + tensor var_5933_axis_0 = const()[name = tensor("op_5933_axis_0"), val = tensor(1)]; + tensor var_5933_mode_0 = const()[name = tensor("op_5933_mode_0"), val = tensor("update")]; + tensor var_5933_validate_indices_0 = const()[name = tensor("op_5933_validate_indices_0"), val = tensor(false)]; + tensor var_5933_cast_fp16 = scatter_along_axis(axis = var_5933_axis_0, data = var_5931_cast_fp16, indices = write_indices_27, mode = var_5933_mode_0, updates = v_27_cast_fp16, validate_indices = var_5933_validate_indices_0)[name = tensor("op_5933_cast_fp16")]; + tensor concat_95 = const()[name = tensor("concat_95"), val = tensor([1, 0, 0, 0, 0])]; + tensor concat_96 = const()[name = tensor("concat_96"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_27_internal_tensor_assign_2_stride_0 = const()[name = tensor("new_cache_27_internal_tensor_assign_2_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_27_internal_tensor_assign_2_begin_mask_0 = const()[name = tensor("new_cache_27_internal_tensor_assign_2_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_27_internal_tensor_assign_2_end_mask_0 = const()[name = tensor("new_cache_27_internal_tensor_assign_2_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_27_internal_tensor_assign_2_squeeze_mask_0 = const()[name = tensor("new_cache_27_internal_tensor_assign_2_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_75 = const()[name = tensor("shape_75"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_27 = const()[name = tensor("reduce_prod_27"), val = tensor(1048576)]; + tensor range_1d_27_start_0 = const()[name = tensor("range_1d_27_start_0"), val = tensor(0)]; + tensor range_1d_27_step_0 = const()[name = tensor("range_1d_27_step_0"), val = tensor(1)]; + tensor range_1d_27 = range_1d(end = reduce_prod_27, start = range_1d_27_start_0, step = range_1d_27_step_0)[name = tensor("range_1d_27")]; + tensor reshape_135 = reshape(shape = shape_75, x = range_1d_27)[name = tensor("reshape_135")]; + tensor slice_by_index_27 = slice_by_index(begin = concat_95, begin_mask = new_cache_27_internal_tensor_assign_2_begin_mask_0, end = concat_96, end_mask = new_cache_27_internal_tensor_assign_2_end_mask_0, squeeze_mask = new_cache_27_internal_tensor_assign_2_squeeze_mask_0, stride = new_cache_27_internal_tensor_assign_2_stride_0, x = reshape_135)[name = tensor("slice_by_index_27")]; + tensor reshape_136_shape_0 = const()[name = tensor("reshape_136_shape_0"), val = tensor([-1])]; + tensor reshape_136 = reshape(shape = reshape_136_shape_0, x = slice_by_index_27)[name = tensor("reshape_136")]; + tensor reshape_137_shape_0 = const()[name = tensor("reshape_137_shape_0"), val = tensor([-1])]; + tensor reshape_137_cast_fp16 = reshape(shape = reshape_137_shape_0, x = var_5933_cast_fp16)[name = tensor("reshape_137_cast_fp16")]; + tensor reshape_138_shape_0 = const()[name = tensor("reshape_138_shape_0"), val = tensor([-1])]; + tensor reshape_138_cast_fp16 = reshape(shape = reshape_138_shape_0, x = reshape_134_cast_fp16)[name = tensor("reshape_138_cast_fp16")]; + tensor scatter_27_mode_0 = const()[name = tensor("scatter_27_mode_0"), val = tensor("update")]; + tensor scatter_27_axis_0 = const()[name = tensor("scatter_27_axis_0"), val = tensor(0)]; + tensor scatter_27_validate_indices_0 = const()[name = tensor("scatter_27_validate_indices_0"), val = tensor(false)]; + tensor scatter_27_cast_fp16 = scatter(axis = scatter_27_axis_0, data = reshape_138_cast_fp16, indices = reshape_136, mode = scatter_27_mode_0, updates = reshape_137_cast_fp16, validate_indices = scatter_27_validate_indices_0)[name = tensor("scatter_27_cast_fp16")]; + tensor reshape_139_cast_fp16 = reshape(shape = shape_75, x = scatter_27_cast_fp16)[name = tensor("reshape_139_cast_fp16")]; + tensor reshape_139_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("reshape_139_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor keys_79_begin_0 = const()[name = tensor("keys_79_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor keys_79_end_0 = const()[name = tensor("keys_79_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor keys_79_end_mask_0 = const()[name = tensor("keys_79_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor keys_79_squeeze_mask_0 = const()[name = tensor("keys_79_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor keys_79_cast_fp16 = slice_by_index(begin = keys_79_begin_0, end = keys_79_end_0, end_mask = keys_79_end_mask_0, squeeze_mask = keys_79_squeeze_mask_0, x = reshape_139_cast_fp16)[name = tensor("keys_79_cast_fp16")]; + tensor values_79_begin_0 = const()[name = tensor("values_79_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor values_79_end_0 = const()[name = tensor("values_79_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor values_79_end_mask_0 = const()[name = tensor("values_79_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor values_79_squeeze_mask_0 = const()[name = tensor("values_79_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor values_79_cast_fp16 = slice_by_index(begin = values_79_begin_0, end = values_79_end_0, end_mask = values_79_end_mask_0, squeeze_mask = values_79_squeeze_mask_0, x = reshape_139_cast_fp16)[name = tensor("values_79_cast_fp16")]; + tensor var_5945_cast_fp16 = not_equal(x = keys_79_cast_fp16, y = keys_79_cast_fp16)[name = tensor("op_5945_cast_fp16")]; + tensor keys_81_cast_fp16 = select(a = var_504_to_fp16, b = keys_79_cast_fp16, cond = var_5945_cast_fp16)[name = tensor("keys_81_cast_fp16")]; + tensor var_5953_cast_fp16 = not_equal(x = values_79_cast_fp16, y = values_79_cast_fp16)[name = tensor("op_5953_cast_fp16")]; + tensor values_81_cast_fp16 = select(a = var_504_to_fp16, b = values_79_cast_fp16, cond = var_5953_cast_fp16)[name = tensor("values_81_cast_fp16")]; + tensor var_5977 = const()[name = tensor("op_5977"), val = tensor([0, 2, 1, 3])]; + tensor var_5990 = const()[name = tensor("op_5990"), val = tensor([1, 1, 1])]; + tensor var_5991_cast_fp16 = reshape(shape = var_5990, x = position13_to_fp16)[name = tensor("op_5991_cast_fp16")]; + tensor var_6008_to_fp16 = const()[name = tensor("op_6008_to_fp16"), val = tensor(0x1p+0)]; + tensor valid_len_27_cast_fp16 = add(x = var_5991_cast_fp16, y = var_6008_to_fp16)[name = tensor("valid_len_27_cast_fp16")]; + tensor valid_mask_27_cast_fp16 = less(x = k_positions_1_promoted_to_fp16, y = valid_len_27_cast_fp16)[name = tensor("valid_mask_27_cast_fp16")]; + tensor causal_mask_27_cast_fp16 = less_equal(x = k_positions_1_promoted_to_fp16, y = var_5991_cast_fp16)[name = tensor("causal_mask_27_cast_fp16")]; + tensor attn_mask_53 = logical_and(x = valid_mask_27_cast_fp16, y = causal_mask_27_cast_fp16)[name = tensor("attn_mask_53")]; + tensor attn_mask_55_axes_0 = const()[name = tensor("attn_mask_55_axes_0"), val = tensor([1])]; + tensor attn_mask_55 = expand_dims(axes = attn_mask_55_axes_0, x = attn_mask_53)[name = tensor("attn_mask_55")]; + tensor var_6026_transpose_x_0 = const()[name = tensor("op_6026_transpose_x_0"), val = tensor(false)]; + tensor var_6026_transpose_y_0 = const()[name = tensor("op_6026_transpose_y_0"), val = tensor(false)]; + tensor transpose_98_perm_0 = const()[name = tensor("transpose_98_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_99_perm_0 = const()[name = tensor("transpose_99_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_99 = transpose(perm = transpose_99_perm_0, x = keys_81_cast_fp16)[name = tensor("transpose_161")]; + tensor transpose_98 = transpose(perm = transpose_98_perm_0, x = q_81_cast_fp16)[name = tensor("transpose_162")]; + tensor var_6026_cast_fp16 = matmul(transpose_x = var_6026_transpose_x_0, transpose_y = var_6026_transpose_y_0, x = transpose_98, y = transpose_99)[name = tensor("op_6026_cast_fp16")]; + tensor var_6020_to_fp16 = const()[name = tensor("op_6020_to_fp16"), val = tensor([0x1p-3])]; + tensor attn_weights_79_cast_fp16 = mul(x = var_6026_cast_fp16, y = var_6020_to_fp16)[name = tensor("attn_weights_79_cast_fp16")]; + tensor var_6028 = logical_not(x = attn_mask_55)[name = tensor("op_6028")]; + tensor var_6029_to_fp16 = const()[name = tensor("op_6029_to_fp16"), val = tensor(-inf)]; + tensor attn_weights_81_cast_fp16 = select(a = var_6029_to_fp16, b = attn_weights_79_cast_fp16, cond = var_6028)[name = tensor("attn_weights_81_cast_fp16")]; + tensor var_6031 = const()[name = tensor("op_6031"), val = tensor(-1)]; + tensor attn_weights_83_cast_fp16 = softmax(axis = var_6031, x = attn_weights_81_cast_fp16)[name = tensor("attn_weights_83_cast_fp16")]; + tensor attn_output_27_transpose_x_0 = const()[name = tensor("attn_output_27_transpose_x_0"), val = tensor(false)]; + tensor attn_output_27_transpose_y_0 = const()[name = tensor("attn_output_27_transpose_y_0"), val = tensor(false)]; + tensor values_83_cast_fp16 = transpose(perm = var_5977, x = values_81_cast_fp16)[name = tensor("transpose_163")]; + tensor attn_output_27_cast_fp16 = matmul(transpose_x = attn_output_27_transpose_x_0, transpose_y = attn_output_27_transpose_y_0, x = attn_weights_83_cast_fp16, y = values_83_cast_fp16)[name = tensor("attn_output_27_cast_fp16")]; + tensor var_6039 = const()[name = tensor("op_6039"), val = tensor([0, 2, 1, 3])]; + tensor var_6042 = const()[name = tensor("op_6042"), val = tensor([1, 1, 1024])]; + tensor var_6040_cast_fp16 = transpose(perm = var_6039, x = attn_output_27_cast_fp16)[name = tensor("transpose_160")]; + tensor input_135_cast_fp16 = reshape(shape = var_6042, x = var_6040_cast_fp16)[name = tensor("input_135_cast_fp16")]; + tensor linear_54_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = attn13_out_proj_weight_cast_fp16, x = input_135_cast_fp16)[name = tensor("linear_54_cast_fp16")]; + tensor var_6048_to_fp16 = const()[name = tensor("op_6048_to_fp16"), val = tensor(0x1p+0)]; + tensor var_6049_cast_fp16 = add(x = position13_to_fp16, y = var_6048_to_fp16)[name = tensor("op_6049_cast_fp16")]; + tensor var_6049_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_6049_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor input_137_cast_fp16 = add(x = input_133_cast_fp16, y = linear_54_cast_fp16)[name = tensor("input_137_cast_fp16")]; + tensor input_139_axes_0 = const()[name = tensor("input_139_axes_0"), val = tensor([-1])]; + tensor norm13_2_weight_to_fp16 = const()[name = tensor("norm13_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(303700736)))]; + tensor norm13_2_bias_to_fp16 = const()[name = tensor("norm13_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(303702848)))]; + tensor var_6053_to_fp16 = const()[name = tensor("op_6053_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_139_cast_fp16 = layer_norm(axes = input_139_axes_0, beta = norm13_2_bias_to_fp16, epsilon = var_6053_to_fp16, gamma = norm13_2_weight_to_fp16, x = input_137_cast_fp16)[name = tensor("input_139_cast_fp16")]; + tensor linear_55_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = linear13_1_weight_cast_fp16, x = input_139_cast_fp16)[name = tensor("linear_55_cast_fp16")]; + tensor input_141_mode_0 = const()[name = tensor("input_141_mode_0"), val = tensor("EXACT")]; + tensor input_141_cast_fp16 = gelu(mode = input_141_mode_0, x = linear_55_cast_fp16)[name = tensor("input_141_cast_fp16")]; + tensor linear_56_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = linear13_2_weight_cast_fp16, x = input_141_cast_fp16)[name = tensor("linear_56_cast_fp16")]; + tensor input_143_cast_fp16 = add(x = input_137_cast_fp16, y = linear_56_cast_fp16)[name = tensor("input_143_cast_fp16")]; + tensor x_29_axes_0 = const()[name = tensor("x_29_axes_0"), val = tensor([-1])]; + tensor norm14_1_weight_to_fp16 = const()[name = tensor("norm14_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(303704960)))]; + tensor norm14_1_bias_to_fp16 = const()[name = tensor("norm14_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(303707072)))]; + tensor var_6070_to_fp16 = const()[name = tensor("op_6070_to_fp16"), val = tensor(0x1.5p-17)]; + tensor x_29_cast_fp16 = layer_norm(axes = x_29_axes_0, beta = norm14_1_bias_to_fp16, epsilon = var_6070_to_fp16, gamma = norm14_1_weight_to_fp16, x = input_143_cast_fp16)[name = tensor("x_29_cast_fp16")]; + tensor linear_57_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = attn14_in_proj_weight_cast_fp16, x = x_29_cast_fp16)[name = tensor("linear_57_cast_fp16")]; + tensor var_6106 = const()[name = tensor("op_6106"), val = tensor([1, 1, 3, 16, 64])]; + tensor qkv_29_cast_fp16 = reshape(shape = var_6106, x = linear_57_cast_fp16)[name = tensor("qkv_29_cast_fp16")]; + tensor q_85_begin_0 = const()[name = tensor("q_85_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor q_85_end_0 = const()[name = tensor("q_85_end_0"), val = tensor([1, 1, 1, 16, 64])]; + tensor q_85_end_mask_0 = const()[name = tensor("q_85_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor q_85_squeeze_mask_0 = const()[name = tensor("q_85_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor q_85_cast_fp16 = slice_by_index(begin = q_85_begin_0, end = q_85_end_0, end_mask = q_85_end_mask_0, squeeze_mask = q_85_squeeze_mask_0, x = qkv_29_cast_fp16)[name = tensor("q_85_cast_fp16")]; + tensor k_57_begin_0 = const()[name = tensor("k_57_begin_0"), val = tensor([0, 0, 1, 0, 0])]; + tensor k_57_end_0 = const()[name = tensor("k_57_end_0"), val = tensor([1, 1, 2, 16, 64])]; + tensor k_57_end_mask_0 = const()[name = tensor("k_57_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor k_57_squeeze_mask_0 = const()[name = tensor("k_57_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor k_57_cast_fp16 = slice_by_index(begin = k_57_begin_0, end = k_57_end_0, end_mask = k_57_end_mask_0, squeeze_mask = k_57_squeeze_mask_0, x = qkv_29_cast_fp16)[name = tensor("k_57_cast_fp16")]; + tensor v_29_begin_0 = const()[name = tensor("v_29_begin_0"), val = tensor([0, 0, 2, 0, 0])]; + tensor v_29_end_0 = const()[name = tensor("v_29_end_0"), val = tensor([1, 1, 3, 16, 64])]; + tensor v_29_end_mask_0 = const()[name = tensor("v_29_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor v_29_squeeze_mask_0 = const()[name = tensor("v_29_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor v_29_cast_fp16 = slice_by_index(begin = v_29_begin_0, end = v_29_end_0, end_mask = v_29_end_mask_0, squeeze_mask = v_29_squeeze_mask_0, x = qkv_29_cast_fp16)[name = tensor("v_29_cast_fp16")]; + tensor var_6210 = const()[name = tensor("op_6210"), val = tensor([1, 1, 1, 1])]; + tensor position14_to_fp16_dtype_0 = const()[name = tensor("position14_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor position14_to_fp16 = cast(dtype = position14_to_fp16_dtype_0, x = position14)[name = tensor("cast_483")]; + tensor ts_89_cast_fp16 = reshape(shape = var_6210, x = position14_to_fp16)[name = tensor("ts_89_cast_fp16")]; + tensor var_6214 = const()[name = tensor("op_6214"), val = tensor([1, 1, 16, 32, 2])]; + tensor q_complex_29_cast_fp16 = reshape(shape = var_6214, x = q_85_cast_fp16)[name = tensor("q_complex_29_cast_fp16")]; + tensor var_6218 = const()[name = tensor("op_6218"), val = tensor([1, 1, 16, 32, 2])]; + tensor k_complex_29_cast_fp16 = reshape(shape = var_6218, x = k_57_cast_fp16)[name = tensor("k_complex_29_cast_fp16")]; + tensor var_6222_begin_0 = const()[name = tensor("op_6222_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_6222_end_0 = const()[name = tensor("op_6222_end_0"), val = tensor([1, 1, 16, 32, 1])]; + tensor var_6222_end_mask_0 = const()[name = tensor("op_6222_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_6222_squeeze_mask_0 = const()[name = tensor("op_6222_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_6222_cast_fp16 = slice_by_index(begin = var_6222_begin_0, end = var_6222_end_0, end_mask = var_6222_end_mask_0, squeeze_mask = var_6222_squeeze_mask_0, x = q_complex_29_cast_fp16)[name = tensor("op_6222_cast_fp16")]; + tensor var_6230_begin_0 = const()[name = tensor("op_6230_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_6230_end_0 = const()[name = tensor("op_6230_end_0"), val = tensor([1, 1, 16, 32, 2])]; + tensor var_6230_end_mask_0 = const()[name = tensor("op_6230_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_6230_squeeze_mask_0 = const()[name = tensor("op_6230_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_6230_cast_fp16 = slice_by_index(begin = var_6230_begin_0, end = var_6230_end_0, end_mask = var_6230_end_mask_0, squeeze_mask = var_6230_squeeze_mask_0, x = q_complex_29_cast_fp16)[name = tensor("op_6230_cast_fp16")]; + tensor var_6238_begin_0 = const()[name = tensor("op_6238_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_6238_end_0 = const()[name = tensor("op_6238_end_0"), val = tensor([1, 1, 16, 32, 1])]; + tensor var_6238_end_mask_0 = const()[name = tensor("op_6238_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_6238_squeeze_mask_0 = const()[name = tensor("op_6238_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_6238_cast_fp16 = slice_by_index(begin = var_6238_begin_0, end = var_6238_end_0, end_mask = var_6238_end_mask_0, squeeze_mask = var_6238_squeeze_mask_0, x = k_complex_29_cast_fp16)[name = tensor("op_6238_cast_fp16")]; + tensor var_6246_begin_0 = const()[name = tensor("op_6246_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_6246_end_0 = const()[name = tensor("op_6246_end_0"), val = tensor([1, 1, 16, 32, 2])]; + tensor var_6246_end_mask_0 = const()[name = tensor("op_6246_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_6246_squeeze_mask_0 = const()[name = tensor("op_6246_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_6246_cast_fp16 = slice_by_index(begin = var_6246_begin_0, end = var_6246_end_0, end_mask = var_6246_end_mask_0, squeeze_mask = var_6246_squeeze_mask_0, x = k_complex_29_cast_fp16)[name = tensor("op_6246_cast_fp16")]; + tensor freqs_29_to_fp16 = const()[name = tensor("freqs_29_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(303709184)))]; + tensor var_6252_cast_fp16 = mul(x = freqs_29_to_fp16, y = ts_89_cast_fp16)[name = tensor("op_6252_cast_fp16")]; + tensor rotr_29_cast_fp16 = cos(x = var_6252_cast_fp16)[name = tensor("rotr_29_cast_fp16")]; + tensor roti_29_cast_fp16 = sin(x = var_6252_cast_fp16)[name = tensor("roti_29_cast_fp16")]; + tensor var_6256_cast_fp16 = mul(x = var_6222_cast_fp16, y = rotr_29_cast_fp16)[name = tensor("op_6256_cast_fp16")]; + tensor var_6257_cast_fp16 = mul(x = var_6230_cast_fp16, y = roti_29_cast_fp16)[name = tensor("op_6257_cast_fp16")]; + tensor qor_57_cast_fp16 = sub(x = var_6256_cast_fp16, y = var_6257_cast_fp16)[name = tensor("qor_57_cast_fp16")]; + tensor var_6260_cast_fp16 = mul(x = var_6222_cast_fp16, y = roti_29_cast_fp16)[name = tensor("op_6260_cast_fp16")]; + tensor var_6261_cast_fp16 = mul(x = var_6230_cast_fp16, y = rotr_29_cast_fp16)[name = tensor("op_6261_cast_fp16")]; + tensor qoi_57_cast_fp16 = add(x = var_6260_cast_fp16, y = var_6261_cast_fp16)[name = tensor("qoi_57_cast_fp16")]; + tensor var_6264_cast_fp16 = mul(x = var_6238_cast_fp16, y = rotr_29_cast_fp16)[name = tensor("op_6264_cast_fp16")]; + tensor var_6265_cast_fp16 = mul(x = var_6246_cast_fp16, y = roti_29_cast_fp16)[name = tensor("op_6265_cast_fp16")]; + tensor kor_57_cast_fp16 = sub(x = var_6264_cast_fp16, y = var_6265_cast_fp16)[name = tensor("kor_57_cast_fp16")]; + tensor var_6268_cast_fp16 = mul(x = var_6238_cast_fp16, y = roti_29_cast_fp16)[name = tensor("op_6268_cast_fp16")]; + tensor var_6269_cast_fp16 = mul(x = var_6246_cast_fp16, y = rotr_29_cast_fp16)[name = tensor("op_6269_cast_fp16")]; + tensor koi_57_cast_fp16 = add(x = var_6268_cast_fp16, y = var_6269_cast_fp16)[name = tensor("koi_57_cast_fp16")]; + tensor qo_29_axis_0 = const()[name = tensor("qo_29_axis_0"), val = tensor(-1)]; + tensor qo_29_cast_fp16 = stack(axis = qo_29_axis_0, values = (qor_57_cast_fp16, qoi_57_cast_fp16))[name = tensor("qo_29_cast_fp16")]; + tensor ko_29_axis_0 = const()[name = tensor("ko_29_axis_0"), val = tensor(-1)]; + tensor ko_29_cast_fp16 = stack(axis = ko_29_axis_0, values = (kor_57_cast_fp16, koi_57_cast_fp16))[name = tensor("ko_29_cast_fp16")]; + tensor var_6298 = const()[name = tensor("op_6298"), val = tensor([1, 1, 16, 64])]; + tensor q_87_cast_fp16 = reshape(shape = var_6298, x = qo_29_cast_fp16)[name = tensor("q_87_cast_fp16")]; + tensor var_6300 = const()[name = tensor("op_6300"), val = tensor([1, 1, 16, 64])]; + tensor k_59_cast_fp16 = reshape(shape = var_6300, x = ko_29_cast_fp16)[name = tensor("k_59_cast_fp16")]; + tensor _inversed_6322_y_0_to_fp16 = const()[name = tensor("_inversed_6322_y_0_to_fp16"), val = tensor(0x1p-9)]; + tensor _inversed_6322_cast_fp16 = mul(x = ts_89_cast_fp16, y = _inversed_6322_y_0_to_fp16)[name = tensor("_inversed_6322_cast_fp16")]; + tensor var_6323_cast_fp16 = floor(x = _inversed_6322_cast_fp16)[name = tensor("op_6323_cast_fp16")]; + tensor var_6324_to_fp16 = const()[name = tensor("op_6324_to_fp16"), val = tensor(0x1p+9)]; + tensor var_6325_cast_fp16 = mul(x = var_6323_cast_fp16, y = var_6324_to_fp16)[name = tensor("op_6325_cast_fp16")]; + tensor write_indices_float_59_cast_fp16 = sub(x = ts_89_cast_fp16, y = var_6325_cast_fp16)[name = tensor("write_indices_float_59_cast_fp16")]; + tensor var_6332_dtype_0 = const()[name = tensor("op_6332_dtype_0"), val = tensor("int32")]; + tensor write_indices_29_reps_0 = const()[name = tensor("write_indices_29_reps_0"), val = tensor([1, 1, 16, 64])]; + tensor write_indices_float_59_cast_fp16_to_int32 = cast(dtype = var_6332_dtype_0, x = write_indices_float_59_cast_fp16)[name = tensor("cast_482")]; + tensor write_indices_29 = tile(reps = write_indices_29_reps_0, x = write_indices_float_59_cast_fp16_to_int32)[name = tensor("write_indices_29")]; + tensor var_6340_begin_0 = const()[name = tensor("op_6340_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_6340_end_0 = const()[name = tensor("op_6340_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor var_6340_end_mask_0 = const()[name = tensor("op_6340_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_6340_squeeze_mask_0 = const()[name = tensor("op_6340_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor cache14_to_fp16_dtype_0 = const()[name = tensor("cache14_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor cache14_to_fp16 = cast(dtype = cache14_to_fp16_dtype_0, x = cache14)[name = tensor("cast_481")]; + tensor var_6340_cast_fp16 = slice_by_index(begin = var_6340_begin_0, end = var_6340_end_0, end_mask = var_6340_end_mask_0, squeeze_mask = var_6340_squeeze_mask_0, x = cache14_to_fp16)[name = tensor("op_6340_cast_fp16")]; + tensor var_6342_axis_0 = const()[name = tensor("op_6342_axis_0"), val = tensor(1)]; + tensor var_6342_mode_0 = const()[name = tensor("op_6342_mode_0"), val = tensor("update")]; + tensor var_6342_validate_indices_0 = const()[name = tensor("op_6342_validate_indices_0"), val = tensor(false)]; + tensor var_6342_cast_fp16 = scatter_along_axis(axis = var_6342_axis_0, data = var_6340_cast_fp16, indices = write_indices_29, mode = var_6342_mode_0, updates = k_59_cast_fp16, validate_indices = var_6342_validate_indices_0)[name = tensor("op_6342_cast_fp16")]; + tensor concat_100 = const()[name = tensor("concat_100"), val = tensor([0, 0, 0, 0, 0])]; + tensor concat_101 = const()[name = tensor("concat_101"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_29_internal_tensor_assign_1_stride_0 = const()[name = tensor("new_cache_29_internal_tensor_assign_1_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_29_internal_tensor_assign_1_begin_mask_0 = const()[name = tensor("new_cache_29_internal_tensor_assign_1_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_29_internal_tensor_assign_1_end_mask_0 = const()[name = tensor("new_cache_29_internal_tensor_assign_1_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_29_internal_tensor_assign_1_squeeze_mask_0 = const()[name = tensor("new_cache_29_internal_tensor_assign_1_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_76 = const()[name = tensor("shape_76"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_28 = const()[name = tensor("reduce_prod_28"), val = tensor(1048576)]; + tensor range_1d_28_start_0 = const()[name = tensor("range_1d_28_start_0"), val = tensor(0)]; + tensor range_1d_28_step_0 = const()[name = tensor("range_1d_28_step_0"), val = tensor(1)]; + tensor range_1d_28 = range_1d(end = reduce_prod_28, start = range_1d_28_start_0, step = range_1d_28_step_0)[name = tensor("range_1d_28")]; + tensor reshape_140 = reshape(shape = shape_76, x = range_1d_28)[name = tensor("reshape_140")]; + tensor slice_by_index_28 = slice_by_index(begin = concat_100, begin_mask = new_cache_29_internal_tensor_assign_1_begin_mask_0, end = concat_101, end_mask = new_cache_29_internal_tensor_assign_1_end_mask_0, squeeze_mask = new_cache_29_internal_tensor_assign_1_squeeze_mask_0, stride = new_cache_29_internal_tensor_assign_1_stride_0, x = reshape_140)[name = tensor("slice_by_index_28")]; + tensor reshape_141_shape_0 = const()[name = tensor("reshape_141_shape_0"), val = tensor([-1])]; + tensor reshape_141 = reshape(shape = reshape_141_shape_0, x = slice_by_index_28)[name = tensor("reshape_141")]; + tensor reshape_142_shape_0 = const()[name = tensor("reshape_142_shape_0"), val = tensor([-1])]; + tensor reshape_142_cast_fp16 = reshape(shape = reshape_142_shape_0, x = var_6342_cast_fp16)[name = tensor("reshape_142_cast_fp16")]; + tensor reshape_143_shape_0 = const()[name = tensor("reshape_143_shape_0"), val = tensor([-1])]; + tensor reshape_143_cast_fp16 = reshape(shape = reshape_143_shape_0, x = cache14_to_fp16)[name = tensor("reshape_143_cast_fp16")]; + tensor scatter_28_mode_0 = const()[name = tensor("scatter_28_mode_0"), val = tensor("update")]; + tensor scatter_28_axis_0 = const()[name = tensor("scatter_28_axis_0"), val = tensor(0)]; + tensor scatter_28_validate_indices_0 = const()[name = tensor("scatter_28_validate_indices_0"), val = tensor(false)]; + tensor scatter_28_cast_fp16 = scatter(axis = scatter_28_axis_0, data = reshape_143_cast_fp16, indices = reshape_141, mode = scatter_28_mode_0, updates = reshape_142_cast_fp16, validate_indices = scatter_28_validate_indices_0)[name = tensor("scatter_28_cast_fp16")]; + tensor reshape_144_cast_fp16 = reshape(shape = shape_76, x = scatter_28_cast_fp16)[name = tensor("reshape_144_cast_fp16")]; + tensor var_6350_begin_0 = const()[name = tensor("op_6350_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor var_6350_end_0 = const()[name = tensor("op_6350_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor var_6350_end_mask_0 = const()[name = tensor("op_6350_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_6350_squeeze_mask_0 = const()[name = tensor("op_6350_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor var_6350_cast_fp16 = slice_by_index(begin = var_6350_begin_0, end = var_6350_end_0, end_mask = var_6350_end_mask_0, squeeze_mask = var_6350_squeeze_mask_0, x = reshape_144_cast_fp16)[name = tensor("op_6350_cast_fp16")]; + tensor var_6352_axis_0 = const()[name = tensor("op_6352_axis_0"), val = tensor(1)]; + tensor var_6352_mode_0 = const()[name = tensor("op_6352_mode_0"), val = tensor("update")]; + tensor var_6352_validate_indices_0 = const()[name = tensor("op_6352_validate_indices_0"), val = tensor(false)]; + tensor var_6352_cast_fp16 = scatter_along_axis(axis = var_6352_axis_0, data = var_6350_cast_fp16, indices = write_indices_29, mode = var_6352_mode_0, updates = v_29_cast_fp16, validate_indices = var_6352_validate_indices_0)[name = tensor("op_6352_cast_fp16")]; + tensor concat_102 = const()[name = tensor("concat_102"), val = tensor([1, 0, 0, 0, 0])]; + tensor concat_103 = const()[name = tensor("concat_103"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_29_internal_tensor_assign_2_stride_0 = const()[name = tensor("new_cache_29_internal_tensor_assign_2_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_29_internal_tensor_assign_2_begin_mask_0 = const()[name = tensor("new_cache_29_internal_tensor_assign_2_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_29_internal_tensor_assign_2_end_mask_0 = const()[name = tensor("new_cache_29_internal_tensor_assign_2_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_29_internal_tensor_assign_2_squeeze_mask_0 = const()[name = tensor("new_cache_29_internal_tensor_assign_2_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_77 = const()[name = tensor("shape_77"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_29 = const()[name = tensor("reduce_prod_29"), val = tensor(1048576)]; + tensor range_1d_29_start_0 = const()[name = tensor("range_1d_29_start_0"), val = tensor(0)]; + tensor range_1d_29_step_0 = const()[name = tensor("range_1d_29_step_0"), val = tensor(1)]; + tensor range_1d_29 = range_1d(end = reduce_prod_29, start = range_1d_29_start_0, step = range_1d_29_step_0)[name = tensor("range_1d_29")]; + tensor reshape_145 = reshape(shape = shape_77, x = range_1d_29)[name = tensor("reshape_145")]; + tensor slice_by_index_29 = slice_by_index(begin = concat_102, begin_mask = new_cache_29_internal_tensor_assign_2_begin_mask_0, end = concat_103, end_mask = new_cache_29_internal_tensor_assign_2_end_mask_0, squeeze_mask = new_cache_29_internal_tensor_assign_2_squeeze_mask_0, stride = new_cache_29_internal_tensor_assign_2_stride_0, x = reshape_145)[name = tensor("slice_by_index_29")]; + tensor reshape_146_shape_0 = const()[name = tensor("reshape_146_shape_0"), val = tensor([-1])]; + tensor reshape_146 = reshape(shape = reshape_146_shape_0, x = slice_by_index_29)[name = tensor("reshape_146")]; + tensor reshape_147_shape_0 = const()[name = tensor("reshape_147_shape_0"), val = tensor([-1])]; + tensor reshape_147_cast_fp16 = reshape(shape = reshape_147_shape_0, x = var_6352_cast_fp16)[name = tensor("reshape_147_cast_fp16")]; + tensor reshape_148_shape_0 = const()[name = tensor("reshape_148_shape_0"), val = tensor([-1])]; + tensor reshape_148_cast_fp16 = reshape(shape = reshape_148_shape_0, x = reshape_144_cast_fp16)[name = tensor("reshape_148_cast_fp16")]; + tensor scatter_29_mode_0 = const()[name = tensor("scatter_29_mode_0"), val = tensor("update")]; + tensor scatter_29_axis_0 = const()[name = tensor("scatter_29_axis_0"), val = tensor(0)]; + tensor scatter_29_validate_indices_0 = const()[name = tensor("scatter_29_validate_indices_0"), val = tensor(false)]; + tensor scatter_29_cast_fp16 = scatter(axis = scatter_29_axis_0, data = reshape_148_cast_fp16, indices = reshape_146, mode = scatter_29_mode_0, updates = reshape_147_cast_fp16, validate_indices = scatter_29_validate_indices_0)[name = tensor("scatter_29_cast_fp16")]; + tensor reshape_149_cast_fp16 = reshape(shape = shape_77, x = scatter_29_cast_fp16)[name = tensor("reshape_149_cast_fp16")]; + tensor reshape_149_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("reshape_149_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor keys_85_begin_0 = const()[name = tensor("keys_85_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor keys_85_end_0 = const()[name = tensor("keys_85_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor keys_85_end_mask_0 = const()[name = tensor("keys_85_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor keys_85_squeeze_mask_0 = const()[name = tensor("keys_85_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor keys_85_cast_fp16 = slice_by_index(begin = keys_85_begin_0, end = keys_85_end_0, end_mask = keys_85_end_mask_0, squeeze_mask = keys_85_squeeze_mask_0, x = reshape_149_cast_fp16)[name = tensor("keys_85_cast_fp16")]; + tensor values_85_begin_0 = const()[name = tensor("values_85_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor values_85_end_0 = const()[name = tensor("values_85_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor values_85_end_mask_0 = const()[name = tensor("values_85_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor values_85_squeeze_mask_0 = const()[name = tensor("values_85_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor values_85_cast_fp16 = slice_by_index(begin = values_85_begin_0, end = values_85_end_0, end_mask = values_85_end_mask_0, squeeze_mask = values_85_squeeze_mask_0, x = reshape_149_cast_fp16)[name = tensor("values_85_cast_fp16")]; + tensor var_6364_cast_fp16 = not_equal(x = keys_85_cast_fp16, y = keys_85_cast_fp16)[name = tensor("op_6364_cast_fp16")]; + tensor keys_87_cast_fp16 = select(a = var_504_to_fp16, b = keys_85_cast_fp16, cond = var_6364_cast_fp16)[name = tensor("keys_87_cast_fp16")]; + tensor var_6372_cast_fp16 = not_equal(x = values_85_cast_fp16, y = values_85_cast_fp16)[name = tensor("op_6372_cast_fp16")]; + tensor values_87_cast_fp16 = select(a = var_504_to_fp16, b = values_85_cast_fp16, cond = var_6372_cast_fp16)[name = tensor("values_87_cast_fp16")]; + tensor var_6396 = const()[name = tensor("op_6396"), val = tensor([0, 2, 1, 3])]; + tensor var_6409 = const()[name = tensor("op_6409"), val = tensor([1, 1, 1])]; + tensor var_6410_cast_fp16 = reshape(shape = var_6409, x = position14_to_fp16)[name = tensor("op_6410_cast_fp16")]; + tensor var_6427_to_fp16 = const()[name = tensor("op_6427_to_fp16"), val = tensor(0x1p+0)]; + tensor valid_len_29_cast_fp16 = add(x = var_6410_cast_fp16, y = var_6427_to_fp16)[name = tensor("valid_len_29_cast_fp16")]; + tensor valid_mask_29_cast_fp16 = less(x = k_positions_1_promoted_to_fp16, y = valid_len_29_cast_fp16)[name = tensor("valid_mask_29_cast_fp16")]; + tensor causal_mask_29_cast_fp16 = less_equal(x = k_positions_1_promoted_to_fp16, y = var_6410_cast_fp16)[name = tensor("causal_mask_29_cast_fp16")]; + tensor attn_mask_57 = logical_and(x = valid_mask_29_cast_fp16, y = causal_mask_29_cast_fp16)[name = tensor("attn_mask_57")]; + tensor attn_mask_59_axes_0 = const()[name = tensor("attn_mask_59_axes_0"), val = tensor([1])]; + tensor attn_mask_59 = expand_dims(axes = attn_mask_59_axes_0, x = attn_mask_57)[name = tensor("attn_mask_59")]; + tensor var_6445_transpose_x_0 = const()[name = tensor("op_6445_transpose_x_0"), val = tensor(false)]; + tensor var_6445_transpose_y_0 = const()[name = tensor("op_6445_transpose_y_0"), val = tensor(false)]; + tensor transpose_100_perm_0 = const()[name = tensor("transpose_100_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_101_perm_0 = const()[name = tensor("transpose_101_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_101 = transpose(perm = transpose_101_perm_0, x = keys_87_cast_fp16)[name = tensor("transpose_157")]; + tensor transpose_100 = transpose(perm = transpose_100_perm_0, x = q_87_cast_fp16)[name = tensor("transpose_158")]; + tensor var_6445_cast_fp16 = matmul(transpose_x = var_6445_transpose_x_0, transpose_y = var_6445_transpose_y_0, x = transpose_100, y = transpose_101)[name = tensor("op_6445_cast_fp16")]; + tensor var_6439_to_fp16 = const()[name = tensor("op_6439_to_fp16"), val = tensor([0x1p-3])]; + tensor attn_weights_85_cast_fp16 = mul(x = var_6445_cast_fp16, y = var_6439_to_fp16)[name = tensor("attn_weights_85_cast_fp16")]; + tensor var_6447 = logical_not(x = attn_mask_59)[name = tensor("op_6447")]; + tensor var_6448_to_fp16 = const()[name = tensor("op_6448_to_fp16"), val = tensor(-inf)]; + tensor attn_weights_87_cast_fp16 = select(a = var_6448_to_fp16, b = attn_weights_85_cast_fp16, cond = var_6447)[name = tensor("attn_weights_87_cast_fp16")]; + tensor var_6450 = const()[name = tensor("op_6450"), val = tensor(-1)]; + tensor attn_weights_89_cast_fp16 = softmax(axis = var_6450, x = attn_weights_87_cast_fp16)[name = tensor("attn_weights_89_cast_fp16")]; + tensor attn_output_29_transpose_x_0 = const()[name = tensor("attn_output_29_transpose_x_0"), val = tensor(false)]; + tensor attn_output_29_transpose_y_0 = const()[name = tensor("attn_output_29_transpose_y_0"), val = tensor(false)]; + tensor values_89_cast_fp16 = transpose(perm = var_6396, x = values_87_cast_fp16)[name = tensor("transpose_159")]; + tensor attn_output_29_cast_fp16 = matmul(transpose_x = attn_output_29_transpose_x_0, transpose_y = attn_output_29_transpose_y_0, x = attn_weights_89_cast_fp16, y = values_89_cast_fp16)[name = tensor("attn_output_29_cast_fp16")]; + tensor var_6458 = const()[name = tensor("op_6458"), val = tensor([0, 2, 1, 3])]; + tensor var_6461 = const()[name = tensor("op_6461"), val = tensor([1, 1, 1024])]; + tensor var_6459_cast_fp16 = transpose(perm = var_6458, x = attn_output_29_cast_fp16)[name = tensor("transpose_156")]; + tensor input_145_cast_fp16 = reshape(shape = var_6461, x = var_6459_cast_fp16)[name = tensor("input_145_cast_fp16")]; + tensor linear_58_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = attn14_out_proj_weight_cast_fp16, x = input_145_cast_fp16)[name = tensor("linear_58_cast_fp16")]; + tensor var_6467_to_fp16 = const()[name = tensor("op_6467_to_fp16"), val = tensor(0x1p+0)]; + tensor var_6468_cast_fp16 = add(x = position14_to_fp16, y = var_6467_to_fp16)[name = tensor("op_6468_cast_fp16")]; + tensor var_6468_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_6468_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor input_147_cast_fp16 = add(x = input_143_cast_fp16, y = linear_58_cast_fp16)[name = tensor("input_147_cast_fp16")]; + tensor input_149_axes_0 = const()[name = tensor("input_149_axes_0"), val = tensor([-1])]; + tensor norm14_2_weight_to_fp16 = const()[name = tensor("norm14_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(303709312)))]; + tensor norm14_2_bias_to_fp16 = const()[name = tensor("norm14_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(303711424)))]; + tensor var_6472_to_fp16 = const()[name = tensor("op_6472_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_149_cast_fp16 = layer_norm(axes = input_149_axes_0, beta = norm14_2_bias_to_fp16, epsilon = var_6472_to_fp16, gamma = norm14_2_weight_to_fp16, x = input_147_cast_fp16)[name = tensor("input_149_cast_fp16")]; + tensor linear_59_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = linear14_1_weight_cast_fp16, x = input_149_cast_fp16)[name = tensor("linear_59_cast_fp16")]; + tensor input_151_mode_0 = const()[name = tensor("input_151_mode_0"), val = tensor("EXACT")]; + tensor input_151_cast_fp16 = gelu(mode = input_151_mode_0, x = linear_59_cast_fp16)[name = tensor("input_151_cast_fp16")]; + tensor linear_60_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = linear14_2_weight_cast_fp16, x = input_151_cast_fp16)[name = tensor("linear_60_cast_fp16")]; + tensor input_153_cast_fp16 = add(x = input_147_cast_fp16, y = linear_60_cast_fp16)[name = tensor("input_153_cast_fp16")]; + tensor x_31_axes_0 = const()[name = tensor("x_31_axes_0"), val = tensor([-1])]; + tensor norm15_1_weight_to_fp16 = const()[name = tensor("norm15_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(303713536)))]; + tensor norm15_1_bias_to_fp16 = const()[name = tensor("norm15_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(303715648)))]; + tensor var_6489_to_fp16 = const()[name = tensor("op_6489_to_fp16"), val = tensor(0x1.5p-17)]; + tensor x_31_cast_fp16 = layer_norm(axes = x_31_axes_0, beta = norm15_1_bias_to_fp16, epsilon = var_6489_to_fp16, gamma = norm15_1_weight_to_fp16, x = input_153_cast_fp16)[name = tensor("x_31_cast_fp16")]; + tensor linear_61_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = attn15_in_proj_weight_cast_fp16, x = x_31_cast_fp16)[name = tensor("linear_61_cast_fp16")]; + tensor var_6525 = const()[name = tensor("op_6525"), val = tensor([1, 1, 3, 16, 64])]; + tensor qkv_31_cast_fp16 = reshape(shape = var_6525, x = linear_61_cast_fp16)[name = tensor("qkv_31_cast_fp16")]; + tensor q_91_begin_0 = const()[name = tensor("q_91_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor q_91_end_0 = const()[name = tensor("q_91_end_0"), val = tensor([1, 1, 1, 16, 64])]; + tensor q_91_end_mask_0 = const()[name = tensor("q_91_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor q_91_squeeze_mask_0 = const()[name = tensor("q_91_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor q_91_cast_fp16 = slice_by_index(begin = q_91_begin_0, end = q_91_end_0, end_mask = q_91_end_mask_0, squeeze_mask = q_91_squeeze_mask_0, x = qkv_31_cast_fp16)[name = tensor("q_91_cast_fp16")]; + tensor k_61_begin_0 = const()[name = tensor("k_61_begin_0"), val = tensor([0, 0, 1, 0, 0])]; + tensor k_61_end_0 = const()[name = tensor("k_61_end_0"), val = tensor([1, 1, 2, 16, 64])]; + tensor k_61_end_mask_0 = const()[name = tensor("k_61_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor k_61_squeeze_mask_0 = const()[name = tensor("k_61_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor k_61_cast_fp16 = slice_by_index(begin = k_61_begin_0, end = k_61_end_0, end_mask = k_61_end_mask_0, squeeze_mask = k_61_squeeze_mask_0, x = qkv_31_cast_fp16)[name = tensor("k_61_cast_fp16")]; + tensor v_31_begin_0 = const()[name = tensor("v_31_begin_0"), val = tensor([0, 0, 2, 0, 0])]; + tensor v_31_end_0 = const()[name = tensor("v_31_end_0"), val = tensor([1, 1, 3, 16, 64])]; + tensor v_31_end_mask_0 = const()[name = tensor("v_31_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor v_31_squeeze_mask_0 = const()[name = tensor("v_31_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor v_31_cast_fp16 = slice_by_index(begin = v_31_begin_0, end = v_31_end_0, end_mask = v_31_end_mask_0, squeeze_mask = v_31_squeeze_mask_0, x = qkv_31_cast_fp16)[name = tensor("v_31_cast_fp16")]; + tensor var_6629 = const()[name = tensor("op_6629"), val = tensor([1, 1, 1, 1])]; + tensor position15_to_fp16_dtype_0 = const()[name = tensor("position15_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor position15_to_fp16 = cast(dtype = position15_to_fp16_dtype_0, x = position15)[name = tensor("cast_478")]; + tensor ts_95_cast_fp16 = reshape(shape = var_6629, x = position15_to_fp16)[name = tensor("ts_95_cast_fp16")]; + tensor var_6633 = const()[name = tensor("op_6633"), val = tensor([1, 1, 16, 32, 2])]; + tensor q_complex_31_cast_fp16 = reshape(shape = var_6633, x = q_91_cast_fp16)[name = tensor("q_complex_31_cast_fp16")]; + tensor var_6637 = const()[name = tensor("op_6637"), val = tensor([1, 1, 16, 32, 2])]; + tensor k_complex_31_cast_fp16 = reshape(shape = var_6637, x = k_61_cast_fp16)[name = tensor("k_complex_31_cast_fp16")]; + tensor var_6641_begin_0 = const()[name = tensor("op_6641_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_6641_end_0 = const()[name = tensor("op_6641_end_0"), val = tensor([1, 1, 16, 32, 1])]; + tensor var_6641_end_mask_0 = const()[name = tensor("op_6641_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_6641_squeeze_mask_0 = const()[name = tensor("op_6641_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_6641_cast_fp16 = slice_by_index(begin = var_6641_begin_0, end = var_6641_end_0, end_mask = var_6641_end_mask_0, squeeze_mask = var_6641_squeeze_mask_0, x = q_complex_31_cast_fp16)[name = tensor("op_6641_cast_fp16")]; + tensor var_6649_begin_0 = const()[name = tensor("op_6649_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_6649_end_0 = const()[name = tensor("op_6649_end_0"), val = tensor([1, 1, 16, 32, 2])]; + tensor var_6649_end_mask_0 = const()[name = tensor("op_6649_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_6649_squeeze_mask_0 = const()[name = tensor("op_6649_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_6649_cast_fp16 = slice_by_index(begin = var_6649_begin_0, end = var_6649_end_0, end_mask = var_6649_end_mask_0, squeeze_mask = var_6649_squeeze_mask_0, x = q_complex_31_cast_fp16)[name = tensor("op_6649_cast_fp16")]; + tensor var_6657_begin_0 = const()[name = tensor("op_6657_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_6657_end_0 = const()[name = tensor("op_6657_end_0"), val = tensor([1, 1, 16, 32, 1])]; + tensor var_6657_end_mask_0 = const()[name = tensor("op_6657_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_6657_squeeze_mask_0 = const()[name = tensor("op_6657_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_6657_cast_fp16 = slice_by_index(begin = var_6657_begin_0, end = var_6657_end_0, end_mask = var_6657_end_mask_0, squeeze_mask = var_6657_squeeze_mask_0, x = k_complex_31_cast_fp16)[name = tensor("op_6657_cast_fp16")]; + tensor var_6665_begin_0 = const()[name = tensor("op_6665_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_6665_end_0 = const()[name = tensor("op_6665_end_0"), val = tensor([1, 1, 16, 32, 2])]; + tensor var_6665_end_mask_0 = const()[name = tensor("op_6665_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_6665_squeeze_mask_0 = const()[name = tensor("op_6665_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_6665_cast_fp16 = slice_by_index(begin = var_6665_begin_0, end = var_6665_end_0, end_mask = var_6665_end_mask_0, squeeze_mask = var_6665_squeeze_mask_0, x = k_complex_31_cast_fp16)[name = tensor("op_6665_cast_fp16")]; + tensor freqs_31_to_fp16 = const()[name = tensor("freqs_31_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(303717760)))]; + tensor var_6671_cast_fp16 = mul(x = freqs_31_to_fp16, y = ts_95_cast_fp16)[name = tensor("op_6671_cast_fp16")]; + tensor rotr_31_cast_fp16 = cos(x = var_6671_cast_fp16)[name = tensor("rotr_31_cast_fp16")]; + tensor roti_31_cast_fp16 = sin(x = var_6671_cast_fp16)[name = tensor("roti_31_cast_fp16")]; + tensor var_6675_cast_fp16 = mul(x = var_6641_cast_fp16, y = rotr_31_cast_fp16)[name = tensor("op_6675_cast_fp16")]; + tensor var_6676_cast_fp16 = mul(x = var_6649_cast_fp16, y = roti_31_cast_fp16)[name = tensor("op_6676_cast_fp16")]; + tensor qor_61_cast_fp16 = sub(x = var_6675_cast_fp16, y = var_6676_cast_fp16)[name = tensor("qor_61_cast_fp16")]; + tensor var_6679_cast_fp16 = mul(x = var_6641_cast_fp16, y = roti_31_cast_fp16)[name = tensor("op_6679_cast_fp16")]; + tensor var_6680_cast_fp16 = mul(x = var_6649_cast_fp16, y = rotr_31_cast_fp16)[name = tensor("op_6680_cast_fp16")]; + tensor qoi_61_cast_fp16 = add(x = var_6679_cast_fp16, y = var_6680_cast_fp16)[name = tensor("qoi_61_cast_fp16")]; + tensor var_6683_cast_fp16 = mul(x = var_6657_cast_fp16, y = rotr_31_cast_fp16)[name = tensor("op_6683_cast_fp16")]; + tensor var_6684_cast_fp16 = mul(x = var_6665_cast_fp16, y = roti_31_cast_fp16)[name = tensor("op_6684_cast_fp16")]; + tensor kor_61_cast_fp16 = sub(x = var_6683_cast_fp16, y = var_6684_cast_fp16)[name = tensor("kor_61_cast_fp16")]; + tensor var_6687_cast_fp16 = mul(x = var_6657_cast_fp16, y = roti_31_cast_fp16)[name = tensor("op_6687_cast_fp16")]; + tensor var_6688_cast_fp16 = mul(x = var_6665_cast_fp16, y = rotr_31_cast_fp16)[name = tensor("op_6688_cast_fp16")]; + tensor koi_61_cast_fp16 = add(x = var_6687_cast_fp16, y = var_6688_cast_fp16)[name = tensor("koi_61_cast_fp16")]; + tensor qo_31_axis_0 = const()[name = tensor("qo_31_axis_0"), val = tensor(-1)]; + tensor qo_31_cast_fp16 = stack(axis = qo_31_axis_0, values = (qor_61_cast_fp16, qoi_61_cast_fp16))[name = tensor("qo_31_cast_fp16")]; + tensor ko_31_axis_0 = const()[name = tensor("ko_31_axis_0"), val = tensor(-1)]; + tensor ko_31_cast_fp16 = stack(axis = ko_31_axis_0, values = (kor_61_cast_fp16, koi_61_cast_fp16))[name = tensor("ko_31_cast_fp16")]; + tensor var_6717 = const()[name = tensor("op_6717"), val = tensor([1, 1, 16, 64])]; + tensor q_93_cast_fp16 = reshape(shape = var_6717, x = qo_31_cast_fp16)[name = tensor("q_93_cast_fp16")]; + tensor var_6719 = const()[name = tensor("op_6719"), val = tensor([1, 1, 16, 64])]; + tensor k_63_cast_fp16 = reshape(shape = var_6719, x = ko_31_cast_fp16)[name = tensor("k_63_cast_fp16")]; + tensor _inversed_6741_y_0_to_fp16 = const()[name = tensor("_inversed_6741_y_0_to_fp16"), val = tensor(0x1p-9)]; + tensor _inversed_6741_cast_fp16 = mul(x = ts_95_cast_fp16, y = _inversed_6741_y_0_to_fp16)[name = tensor("_inversed_6741_cast_fp16")]; + tensor var_6742_cast_fp16 = floor(x = _inversed_6741_cast_fp16)[name = tensor("op_6742_cast_fp16")]; + tensor var_6743_to_fp16 = const()[name = tensor("op_6743_to_fp16"), val = tensor(0x1p+9)]; + tensor var_6744_cast_fp16 = mul(x = var_6742_cast_fp16, y = var_6743_to_fp16)[name = tensor("op_6744_cast_fp16")]; + tensor write_indices_float_63_cast_fp16 = sub(x = ts_95_cast_fp16, y = var_6744_cast_fp16)[name = tensor("write_indices_float_63_cast_fp16")]; + tensor var_6751_dtype_0 = const()[name = tensor("op_6751_dtype_0"), val = tensor("int32")]; + tensor write_indices_31_reps_0 = const()[name = tensor("write_indices_31_reps_0"), val = tensor([1, 1, 16, 64])]; + tensor write_indices_float_63_cast_fp16_to_int32 = cast(dtype = var_6751_dtype_0, x = write_indices_float_63_cast_fp16)[name = tensor("cast_477")]; + tensor write_indices_31 = tile(reps = write_indices_31_reps_0, x = write_indices_float_63_cast_fp16_to_int32)[name = tensor("write_indices_31")]; + tensor var_6759_begin_0 = const()[name = tensor("op_6759_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_6759_end_0 = const()[name = tensor("op_6759_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor var_6759_end_mask_0 = const()[name = tensor("op_6759_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_6759_squeeze_mask_0 = const()[name = tensor("op_6759_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor cache15_to_fp16_dtype_0 = const()[name = tensor("cache15_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor cache15_to_fp16 = cast(dtype = cache15_to_fp16_dtype_0, x = cache15)[name = tensor("cast_476")]; + tensor var_6759_cast_fp16 = slice_by_index(begin = var_6759_begin_0, end = var_6759_end_0, end_mask = var_6759_end_mask_0, squeeze_mask = var_6759_squeeze_mask_0, x = cache15_to_fp16)[name = tensor("op_6759_cast_fp16")]; + tensor var_6761_axis_0 = const()[name = tensor("op_6761_axis_0"), val = tensor(1)]; + tensor var_6761_mode_0 = const()[name = tensor("op_6761_mode_0"), val = tensor("update")]; + tensor var_6761_validate_indices_0 = const()[name = tensor("op_6761_validate_indices_0"), val = tensor(false)]; + tensor var_6761_cast_fp16 = scatter_along_axis(axis = var_6761_axis_0, data = var_6759_cast_fp16, indices = write_indices_31, mode = var_6761_mode_0, updates = k_63_cast_fp16, validate_indices = var_6761_validate_indices_0)[name = tensor("op_6761_cast_fp16")]; + tensor concat_107 = const()[name = tensor("concat_107"), val = tensor([0, 0, 0, 0, 0])]; + tensor concat_108 = const()[name = tensor("concat_108"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_31_internal_tensor_assign_1_stride_0 = const()[name = tensor("new_cache_31_internal_tensor_assign_1_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_31_internal_tensor_assign_1_begin_mask_0 = const()[name = tensor("new_cache_31_internal_tensor_assign_1_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_31_internal_tensor_assign_1_end_mask_0 = const()[name = tensor("new_cache_31_internal_tensor_assign_1_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_31_internal_tensor_assign_1_squeeze_mask_0 = const()[name = tensor("new_cache_31_internal_tensor_assign_1_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_78 = const()[name = tensor("shape_78"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_30 = const()[name = tensor("reduce_prod_30"), val = tensor(1048576)]; + tensor range_1d_30_start_0 = const()[name = tensor("range_1d_30_start_0"), val = tensor(0)]; + tensor range_1d_30_step_0 = const()[name = tensor("range_1d_30_step_0"), val = tensor(1)]; + tensor range_1d_30 = range_1d(end = reduce_prod_30, start = range_1d_30_start_0, step = range_1d_30_step_0)[name = tensor("range_1d_30")]; + tensor reshape_150 = reshape(shape = shape_78, x = range_1d_30)[name = tensor("reshape_150")]; + tensor slice_by_index_30 = slice_by_index(begin = concat_107, begin_mask = new_cache_31_internal_tensor_assign_1_begin_mask_0, end = concat_108, end_mask = new_cache_31_internal_tensor_assign_1_end_mask_0, squeeze_mask = new_cache_31_internal_tensor_assign_1_squeeze_mask_0, stride = new_cache_31_internal_tensor_assign_1_stride_0, x = reshape_150)[name = tensor("slice_by_index_30")]; + tensor reshape_151_shape_0 = const()[name = tensor("reshape_151_shape_0"), val = tensor([-1])]; + tensor reshape_151 = reshape(shape = reshape_151_shape_0, x = slice_by_index_30)[name = tensor("reshape_151")]; + tensor reshape_152_shape_0 = const()[name = tensor("reshape_152_shape_0"), val = tensor([-1])]; + tensor reshape_152_cast_fp16 = reshape(shape = reshape_152_shape_0, x = var_6761_cast_fp16)[name = tensor("reshape_152_cast_fp16")]; + tensor reshape_153_shape_0 = const()[name = tensor("reshape_153_shape_0"), val = tensor([-1])]; + tensor reshape_153_cast_fp16 = reshape(shape = reshape_153_shape_0, x = cache15_to_fp16)[name = tensor("reshape_153_cast_fp16")]; + tensor scatter_30_mode_0 = const()[name = tensor("scatter_30_mode_0"), val = tensor("update")]; + tensor scatter_30_axis_0 = const()[name = tensor("scatter_30_axis_0"), val = tensor(0)]; + tensor scatter_30_validate_indices_0 = const()[name = tensor("scatter_30_validate_indices_0"), val = tensor(false)]; + tensor scatter_30_cast_fp16 = scatter(axis = scatter_30_axis_0, data = reshape_153_cast_fp16, indices = reshape_151, mode = scatter_30_mode_0, updates = reshape_152_cast_fp16, validate_indices = scatter_30_validate_indices_0)[name = tensor("scatter_30_cast_fp16")]; + tensor reshape_154_cast_fp16 = reshape(shape = shape_78, x = scatter_30_cast_fp16)[name = tensor("reshape_154_cast_fp16")]; + tensor var_6769_begin_0 = const()[name = tensor("op_6769_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor var_6769_end_0 = const()[name = tensor("op_6769_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor var_6769_end_mask_0 = const()[name = tensor("op_6769_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_6769_squeeze_mask_0 = const()[name = tensor("op_6769_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor var_6769_cast_fp16 = slice_by_index(begin = var_6769_begin_0, end = var_6769_end_0, end_mask = var_6769_end_mask_0, squeeze_mask = var_6769_squeeze_mask_0, x = reshape_154_cast_fp16)[name = tensor("op_6769_cast_fp16")]; + tensor var_6771_axis_0 = const()[name = tensor("op_6771_axis_0"), val = tensor(1)]; + tensor var_6771_mode_0 = const()[name = tensor("op_6771_mode_0"), val = tensor("update")]; + tensor var_6771_validate_indices_0 = const()[name = tensor("op_6771_validate_indices_0"), val = tensor(false)]; + tensor var_6771_cast_fp16 = scatter_along_axis(axis = var_6771_axis_0, data = var_6769_cast_fp16, indices = write_indices_31, mode = var_6771_mode_0, updates = v_31_cast_fp16, validate_indices = var_6771_validate_indices_0)[name = tensor("op_6771_cast_fp16")]; + tensor concat_109 = const()[name = tensor("concat_109"), val = tensor([1, 0, 0, 0, 0])]; + tensor concat_110 = const()[name = tensor("concat_110"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_31_internal_tensor_assign_2_stride_0 = const()[name = tensor("new_cache_31_internal_tensor_assign_2_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_31_internal_tensor_assign_2_begin_mask_0 = const()[name = tensor("new_cache_31_internal_tensor_assign_2_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_31_internal_tensor_assign_2_end_mask_0 = const()[name = tensor("new_cache_31_internal_tensor_assign_2_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_31_internal_tensor_assign_2_squeeze_mask_0 = const()[name = tensor("new_cache_31_internal_tensor_assign_2_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_79 = const()[name = tensor("shape_79"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_31 = const()[name = tensor("reduce_prod_31"), val = tensor(1048576)]; + tensor range_1d_31_start_0 = const()[name = tensor("range_1d_31_start_0"), val = tensor(0)]; + tensor range_1d_31_step_0 = const()[name = tensor("range_1d_31_step_0"), val = tensor(1)]; + tensor range_1d_31 = range_1d(end = reduce_prod_31, start = range_1d_31_start_0, step = range_1d_31_step_0)[name = tensor("range_1d_31")]; + tensor reshape_155 = reshape(shape = shape_79, x = range_1d_31)[name = tensor("reshape_155")]; + tensor slice_by_index_31 = slice_by_index(begin = concat_109, begin_mask = new_cache_31_internal_tensor_assign_2_begin_mask_0, end = concat_110, end_mask = new_cache_31_internal_tensor_assign_2_end_mask_0, squeeze_mask = new_cache_31_internal_tensor_assign_2_squeeze_mask_0, stride = new_cache_31_internal_tensor_assign_2_stride_0, x = reshape_155)[name = tensor("slice_by_index_31")]; + tensor reshape_156_shape_0 = const()[name = tensor("reshape_156_shape_0"), val = tensor([-1])]; + tensor reshape_156 = reshape(shape = reshape_156_shape_0, x = slice_by_index_31)[name = tensor("reshape_156")]; + tensor reshape_157_shape_0 = const()[name = tensor("reshape_157_shape_0"), val = tensor([-1])]; + tensor reshape_157_cast_fp16 = reshape(shape = reshape_157_shape_0, x = var_6771_cast_fp16)[name = tensor("reshape_157_cast_fp16")]; + tensor reshape_158_shape_0 = const()[name = tensor("reshape_158_shape_0"), val = tensor([-1])]; + tensor reshape_158_cast_fp16 = reshape(shape = reshape_158_shape_0, x = reshape_154_cast_fp16)[name = tensor("reshape_158_cast_fp16")]; + tensor scatter_31_mode_0 = const()[name = tensor("scatter_31_mode_0"), val = tensor("update")]; + tensor scatter_31_axis_0 = const()[name = tensor("scatter_31_axis_0"), val = tensor(0)]; + tensor scatter_31_validate_indices_0 = const()[name = tensor("scatter_31_validate_indices_0"), val = tensor(false)]; + tensor scatter_31_cast_fp16 = scatter(axis = scatter_31_axis_0, data = reshape_158_cast_fp16, indices = reshape_156, mode = scatter_31_mode_0, updates = reshape_157_cast_fp16, validate_indices = scatter_31_validate_indices_0)[name = tensor("scatter_31_cast_fp16")]; + tensor reshape_159_cast_fp16 = reshape(shape = shape_79, x = scatter_31_cast_fp16)[name = tensor("reshape_159_cast_fp16")]; + tensor reshape_159_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("reshape_159_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor keys_91_begin_0 = const()[name = tensor("keys_91_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor keys_91_end_0 = const()[name = tensor("keys_91_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor keys_91_end_mask_0 = const()[name = tensor("keys_91_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor keys_91_squeeze_mask_0 = const()[name = tensor("keys_91_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor keys_91_cast_fp16 = slice_by_index(begin = keys_91_begin_0, end = keys_91_end_0, end_mask = keys_91_end_mask_0, squeeze_mask = keys_91_squeeze_mask_0, x = reshape_159_cast_fp16)[name = tensor("keys_91_cast_fp16")]; + tensor values_91_begin_0 = const()[name = tensor("values_91_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor values_91_end_0 = const()[name = tensor("values_91_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor values_91_end_mask_0 = const()[name = tensor("values_91_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor values_91_squeeze_mask_0 = const()[name = tensor("values_91_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor values_91_cast_fp16 = slice_by_index(begin = values_91_begin_0, end = values_91_end_0, end_mask = values_91_end_mask_0, squeeze_mask = values_91_squeeze_mask_0, x = reshape_159_cast_fp16)[name = tensor("values_91_cast_fp16")]; + tensor var_6783_cast_fp16 = not_equal(x = keys_91_cast_fp16, y = keys_91_cast_fp16)[name = tensor("op_6783_cast_fp16")]; + tensor keys_93_cast_fp16 = select(a = var_504_to_fp16, b = keys_91_cast_fp16, cond = var_6783_cast_fp16)[name = tensor("keys_93_cast_fp16")]; + tensor var_6791_cast_fp16 = not_equal(x = values_91_cast_fp16, y = values_91_cast_fp16)[name = tensor("op_6791_cast_fp16")]; + tensor values_93_cast_fp16 = select(a = var_504_to_fp16, b = values_91_cast_fp16, cond = var_6791_cast_fp16)[name = tensor("values_93_cast_fp16")]; + tensor var_6815 = const()[name = tensor("op_6815"), val = tensor([0, 2, 1, 3])]; + tensor var_6828 = const()[name = tensor("op_6828"), val = tensor([1, 1, 1])]; + tensor var_6829_cast_fp16 = reshape(shape = var_6828, x = position15_to_fp16)[name = tensor("op_6829_cast_fp16")]; + tensor var_6846_to_fp16 = const()[name = tensor("op_6846_to_fp16"), val = tensor(0x1p+0)]; + tensor valid_len_31_cast_fp16 = add(x = var_6829_cast_fp16, y = var_6846_to_fp16)[name = tensor("valid_len_31_cast_fp16")]; + tensor valid_mask_31_cast_fp16 = less(x = k_positions_1_promoted_to_fp16, y = valid_len_31_cast_fp16)[name = tensor("valid_mask_31_cast_fp16")]; + tensor causal_mask_31_cast_fp16 = less_equal(x = k_positions_1_promoted_to_fp16, y = var_6829_cast_fp16)[name = tensor("causal_mask_31_cast_fp16")]; + tensor attn_mask_61 = logical_and(x = valid_mask_31_cast_fp16, y = causal_mask_31_cast_fp16)[name = tensor("attn_mask_61")]; + tensor attn_mask_63_axes_0 = const()[name = tensor("attn_mask_63_axes_0"), val = tensor([1])]; + tensor attn_mask_63 = expand_dims(axes = attn_mask_63_axes_0, x = attn_mask_61)[name = tensor("attn_mask_63")]; + tensor var_6864_transpose_x_0 = const()[name = tensor("op_6864_transpose_x_0"), val = tensor(false)]; + tensor var_6864_transpose_y_0 = const()[name = tensor("op_6864_transpose_y_0"), val = tensor(false)]; + tensor transpose_102_perm_0 = const()[name = tensor("transpose_102_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_103_perm_0 = const()[name = tensor("transpose_103_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_103 = transpose(perm = transpose_103_perm_0, x = keys_93_cast_fp16)[name = tensor("transpose_153")]; + tensor transpose_102 = transpose(perm = transpose_102_perm_0, x = q_93_cast_fp16)[name = tensor("transpose_154")]; + tensor var_6864_cast_fp16 = matmul(transpose_x = var_6864_transpose_x_0, transpose_y = var_6864_transpose_y_0, x = transpose_102, y = transpose_103)[name = tensor("op_6864_cast_fp16")]; + tensor var_6858_to_fp16 = const()[name = tensor("op_6858_to_fp16"), val = tensor([0x1p-3])]; + tensor attn_weights_91_cast_fp16 = mul(x = var_6864_cast_fp16, y = var_6858_to_fp16)[name = tensor("attn_weights_91_cast_fp16")]; + tensor var_6866 = logical_not(x = attn_mask_63)[name = tensor("op_6866")]; + tensor var_6867_to_fp16 = const()[name = tensor("op_6867_to_fp16"), val = tensor(-inf)]; + tensor attn_weights_93_cast_fp16 = select(a = var_6867_to_fp16, b = attn_weights_91_cast_fp16, cond = var_6866)[name = tensor("attn_weights_93_cast_fp16")]; + tensor var_6869 = const()[name = tensor("op_6869"), val = tensor(-1)]; + tensor attn_weights_95_cast_fp16 = softmax(axis = var_6869, x = attn_weights_93_cast_fp16)[name = tensor("attn_weights_95_cast_fp16")]; + tensor attn_output_31_transpose_x_0 = const()[name = tensor("attn_output_31_transpose_x_0"), val = tensor(false)]; + tensor attn_output_31_transpose_y_0 = const()[name = tensor("attn_output_31_transpose_y_0"), val = tensor(false)]; + tensor values_95_cast_fp16 = transpose(perm = var_6815, x = values_93_cast_fp16)[name = tensor("transpose_155")]; + tensor attn_output_31_cast_fp16 = matmul(transpose_x = attn_output_31_transpose_x_0, transpose_y = attn_output_31_transpose_y_0, x = attn_weights_95_cast_fp16, y = values_95_cast_fp16)[name = tensor("attn_output_31_cast_fp16")]; + tensor var_6877 = const()[name = tensor("op_6877"), val = tensor([0, 2, 1, 3])]; + tensor var_6880 = const()[name = tensor("op_6880"), val = tensor([1, 1, 1024])]; + tensor var_6878_cast_fp16 = transpose(perm = var_6877, x = attn_output_31_cast_fp16)[name = tensor("transpose_152")]; + tensor input_155_cast_fp16 = reshape(shape = var_6880, x = var_6878_cast_fp16)[name = tensor("input_155_cast_fp16")]; + tensor linear_62_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = attn15_out_proj_weight_cast_fp16, x = input_155_cast_fp16)[name = tensor("linear_62_cast_fp16")]; + tensor var_6886_to_fp16 = const()[name = tensor("op_6886_to_fp16"), val = tensor(0x1p+0)]; + tensor var_6887_cast_fp16 = add(x = position15_to_fp16, y = var_6886_to_fp16)[name = tensor("op_6887_cast_fp16")]; + tensor var_6887_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_6887_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor input_157_cast_fp16 = add(x = input_153_cast_fp16, y = linear_62_cast_fp16)[name = tensor("input_157_cast_fp16")]; + tensor input_159_axes_0 = const()[name = tensor("input_159_axes_0"), val = tensor([-1])]; + tensor norm15_2_weight_to_fp16 = const()[name = tensor("norm15_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(303717888)))]; + tensor norm15_2_bias_to_fp16 = const()[name = tensor("norm15_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(303720000)))]; + tensor var_6891_to_fp16 = const()[name = tensor("op_6891_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_159_cast_fp16 = layer_norm(axes = input_159_axes_0, beta = norm15_2_bias_to_fp16, epsilon = var_6891_to_fp16, gamma = norm15_2_weight_to_fp16, x = input_157_cast_fp16)[name = tensor("input_159_cast_fp16")]; + tensor linear_63_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = linear15_1_weight_cast_fp16, x = input_159_cast_fp16)[name = tensor("linear_63_cast_fp16")]; + tensor input_161_mode_0 = const()[name = tensor("input_161_mode_0"), val = tensor("EXACT")]; + tensor input_161_cast_fp16 = gelu(mode = input_161_mode_0, x = linear_63_cast_fp16)[name = tensor("input_161_cast_fp16")]; + tensor linear_64_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = linear15_2_weight_cast_fp16, x = input_161_cast_fp16)[name = tensor("linear_64_cast_fp16")]; + tensor input_163_cast_fp16 = add(x = input_157_cast_fp16, y = linear_64_cast_fp16)[name = tensor("input_163_cast_fp16")]; + tensor x_33_axes_0 = const()[name = tensor("x_33_axes_0"), val = tensor([-1])]; + tensor norm16_1_weight_to_fp16 = const()[name = tensor("norm16_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(303722112)))]; + tensor norm16_1_bias_to_fp16 = const()[name = tensor("norm16_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(303724224)))]; + tensor var_6908_to_fp16 = const()[name = tensor("op_6908_to_fp16"), val = tensor(0x1.5p-17)]; + tensor x_33_cast_fp16 = layer_norm(axes = x_33_axes_0, beta = norm16_1_bias_to_fp16, epsilon = var_6908_to_fp16, gamma = norm16_1_weight_to_fp16, x = input_163_cast_fp16)[name = tensor("x_33_cast_fp16")]; + tensor linear_65_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = attn16_in_proj_weight_cast_fp16, x = x_33_cast_fp16)[name = tensor("linear_65_cast_fp16")]; + tensor var_6944 = const()[name = tensor("op_6944"), val = tensor([1, 1, 3, 16, 64])]; + tensor qkv_33_cast_fp16 = reshape(shape = var_6944, x = linear_65_cast_fp16)[name = tensor("qkv_33_cast_fp16")]; + tensor q_97_begin_0 = const()[name = tensor("q_97_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor q_97_end_0 = const()[name = tensor("q_97_end_0"), val = tensor([1, 1, 1, 16, 64])]; + tensor q_97_end_mask_0 = const()[name = tensor("q_97_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor q_97_squeeze_mask_0 = const()[name = tensor("q_97_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor q_97_cast_fp16 = slice_by_index(begin = q_97_begin_0, end = q_97_end_0, end_mask = q_97_end_mask_0, squeeze_mask = q_97_squeeze_mask_0, x = qkv_33_cast_fp16)[name = tensor("q_97_cast_fp16")]; + tensor k_65_begin_0 = const()[name = tensor("k_65_begin_0"), val = tensor([0, 0, 1, 0, 0])]; + tensor k_65_end_0 = const()[name = tensor("k_65_end_0"), val = tensor([1, 1, 2, 16, 64])]; + tensor k_65_end_mask_0 = const()[name = tensor("k_65_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor k_65_squeeze_mask_0 = const()[name = tensor("k_65_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor k_65_cast_fp16 = slice_by_index(begin = k_65_begin_0, end = k_65_end_0, end_mask = k_65_end_mask_0, squeeze_mask = k_65_squeeze_mask_0, x = qkv_33_cast_fp16)[name = tensor("k_65_cast_fp16")]; + tensor v_33_begin_0 = const()[name = tensor("v_33_begin_0"), val = tensor([0, 0, 2, 0, 0])]; + tensor v_33_end_0 = const()[name = tensor("v_33_end_0"), val = tensor([1, 1, 3, 16, 64])]; + tensor v_33_end_mask_0 = const()[name = tensor("v_33_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor v_33_squeeze_mask_0 = const()[name = tensor("v_33_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor v_33_cast_fp16 = slice_by_index(begin = v_33_begin_0, end = v_33_end_0, end_mask = v_33_end_mask_0, squeeze_mask = v_33_squeeze_mask_0, x = qkv_33_cast_fp16)[name = tensor("v_33_cast_fp16")]; + tensor var_7048 = const()[name = tensor("op_7048"), val = tensor([1, 1, 1, 1])]; + tensor position16_to_fp16_dtype_0 = const()[name = tensor("position16_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor position16_to_fp16 = cast(dtype = position16_to_fp16_dtype_0, x = position16)[name = tensor("cast_473")]; + tensor ts_101_cast_fp16 = reshape(shape = var_7048, x = position16_to_fp16)[name = tensor("ts_101_cast_fp16")]; + tensor var_7052 = const()[name = tensor("op_7052"), val = tensor([1, 1, 16, 32, 2])]; + tensor q_complex_33_cast_fp16 = reshape(shape = var_7052, x = q_97_cast_fp16)[name = tensor("q_complex_33_cast_fp16")]; + tensor var_7056 = const()[name = tensor("op_7056"), val = tensor([1, 1, 16, 32, 2])]; + tensor k_complex_33_cast_fp16 = reshape(shape = var_7056, x = k_65_cast_fp16)[name = tensor("k_complex_33_cast_fp16")]; + tensor var_7060_begin_0 = const()[name = tensor("op_7060_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_7060_end_0 = const()[name = tensor("op_7060_end_0"), val = tensor([1, 1, 16, 32, 1])]; + tensor var_7060_end_mask_0 = const()[name = tensor("op_7060_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_7060_squeeze_mask_0 = const()[name = tensor("op_7060_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_7060_cast_fp16 = slice_by_index(begin = var_7060_begin_0, end = var_7060_end_0, end_mask = var_7060_end_mask_0, squeeze_mask = var_7060_squeeze_mask_0, x = q_complex_33_cast_fp16)[name = tensor("op_7060_cast_fp16")]; + tensor var_7068_begin_0 = const()[name = tensor("op_7068_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_7068_end_0 = const()[name = tensor("op_7068_end_0"), val = tensor([1, 1, 16, 32, 2])]; + tensor var_7068_end_mask_0 = const()[name = tensor("op_7068_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_7068_squeeze_mask_0 = const()[name = tensor("op_7068_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_7068_cast_fp16 = slice_by_index(begin = var_7068_begin_0, end = var_7068_end_0, end_mask = var_7068_end_mask_0, squeeze_mask = var_7068_squeeze_mask_0, x = q_complex_33_cast_fp16)[name = tensor("op_7068_cast_fp16")]; + tensor var_7076_begin_0 = const()[name = tensor("op_7076_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_7076_end_0 = const()[name = tensor("op_7076_end_0"), val = tensor([1, 1, 16, 32, 1])]; + tensor var_7076_end_mask_0 = const()[name = tensor("op_7076_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_7076_squeeze_mask_0 = const()[name = tensor("op_7076_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_7076_cast_fp16 = slice_by_index(begin = var_7076_begin_0, end = var_7076_end_0, end_mask = var_7076_end_mask_0, squeeze_mask = var_7076_squeeze_mask_0, x = k_complex_33_cast_fp16)[name = tensor("op_7076_cast_fp16")]; + tensor var_7084_begin_0 = const()[name = tensor("op_7084_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_7084_end_0 = const()[name = tensor("op_7084_end_0"), val = tensor([1, 1, 16, 32, 2])]; + tensor var_7084_end_mask_0 = const()[name = tensor("op_7084_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_7084_squeeze_mask_0 = const()[name = tensor("op_7084_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_7084_cast_fp16 = slice_by_index(begin = var_7084_begin_0, end = var_7084_end_0, end_mask = var_7084_end_mask_0, squeeze_mask = var_7084_squeeze_mask_0, x = k_complex_33_cast_fp16)[name = tensor("op_7084_cast_fp16")]; + tensor freqs_33_to_fp16 = const()[name = tensor("freqs_33_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(303726336)))]; + tensor var_7090_cast_fp16 = mul(x = freqs_33_to_fp16, y = ts_101_cast_fp16)[name = tensor("op_7090_cast_fp16")]; + tensor rotr_33_cast_fp16 = cos(x = var_7090_cast_fp16)[name = tensor("rotr_33_cast_fp16")]; + tensor roti_33_cast_fp16 = sin(x = var_7090_cast_fp16)[name = tensor("roti_33_cast_fp16")]; + tensor var_7094_cast_fp16 = mul(x = var_7060_cast_fp16, y = rotr_33_cast_fp16)[name = tensor("op_7094_cast_fp16")]; + tensor var_7095_cast_fp16 = mul(x = var_7068_cast_fp16, y = roti_33_cast_fp16)[name = tensor("op_7095_cast_fp16")]; + tensor qor_65_cast_fp16 = sub(x = var_7094_cast_fp16, y = var_7095_cast_fp16)[name = tensor("qor_65_cast_fp16")]; + tensor var_7098_cast_fp16 = mul(x = var_7060_cast_fp16, y = roti_33_cast_fp16)[name = tensor("op_7098_cast_fp16")]; + tensor var_7099_cast_fp16 = mul(x = var_7068_cast_fp16, y = rotr_33_cast_fp16)[name = tensor("op_7099_cast_fp16")]; + tensor qoi_65_cast_fp16 = add(x = var_7098_cast_fp16, y = var_7099_cast_fp16)[name = tensor("qoi_65_cast_fp16")]; + tensor var_7102_cast_fp16 = mul(x = var_7076_cast_fp16, y = rotr_33_cast_fp16)[name = tensor("op_7102_cast_fp16")]; + tensor var_7103_cast_fp16 = mul(x = var_7084_cast_fp16, y = roti_33_cast_fp16)[name = tensor("op_7103_cast_fp16")]; + tensor kor_65_cast_fp16 = sub(x = var_7102_cast_fp16, y = var_7103_cast_fp16)[name = tensor("kor_65_cast_fp16")]; + tensor var_7106_cast_fp16 = mul(x = var_7076_cast_fp16, y = roti_33_cast_fp16)[name = tensor("op_7106_cast_fp16")]; + tensor var_7107_cast_fp16 = mul(x = var_7084_cast_fp16, y = rotr_33_cast_fp16)[name = tensor("op_7107_cast_fp16")]; + tensor koi_65_cast_fp16 = add(x = var_7106_cast_fp16, y = var_7107_cast_fp16)[name = tensor("koi_65_cast_fp16")]; + tensor qo_33_axis_0 = const()[name = tensor("qo_33_axis_0"), val = tensor(-1)]; + tensor qo_33_cast_fp16 = stack(axis = qo_33_axis_0, values = (qor_65_cast_fp16, qoi_65_cast_fp16))[name = tensor("qo_33_cast_fp16")]; + tensor ko_33_axis_0 = const()[name = tensor("ko_33_axis_0"), val = tensor(-1)]; + tensor ko_33_cast_fp16 = stack(axis = ko_33_axis_0, values = (kor_65_cast_fp16, koi_65_cast_fp16))[name = tensor("ko_33_cast_fp16")]; + tensor var_7136 = const()[name = tensor("op_7136"), val = tensor([1, 1, 16, 64])]; + tensor q_99_cast_fp16 = reshape(shape = var_7136, x = qo_33_cast_fp16)[name = tensor("q_99_cast_fp16")]; + tensor var_7138 = const()[name = tensor("op_7138"), val = tensor([1, 1, 16, 64])]; + tensor k_67_cast_fp16 = reshape(shape = var_7138, x = ko_33_cast_fp16)[name = tensor("k_67_cast_fp16")]; + tensor _inversed_7160_y_0_to_fp16 = const()[name = tensor("_inversed_7160_y_0_to_fp16"), val = tensor(0x1p-9)]; + tensor _inversed_7160_cast_fp16 = mul(x = ts_101_cast_fp16, y = _inversed_7160_y_0_to_fp16)[name = tensor("_inversed_7160_cast_fp16")]; + tensor var_7161_cast_fp16 = floor(x = _inversed_7160_cast_fp16)[name = tensor("op_7161_cast_fp16")]; + tensor var_7162_to_fp16 = const()[name = tensor("op_7162_to_fp16"), val = tensor(0x1p+9)]; + tensor var_7163_cast_fp16 = mul(x = var_7161_cast_fp16, y = var_7162_to_fp16)[name = tensor("op_7163_cast_fp16")]; + tensor write_indices_float_67_cast_fp16 = sub(x = ts_101_cast_fp16, y = var_7163_cast_fp16)[name = tensor("write_indices_float_67_cast_fp16")]; + tensor var_7170_dtype_0 = const()[name = tensor("op_7170_dtype_0"), val = tensor("int32")]; + tensor write_indices_33_reps_0 = const()[name = tensor("write_indices_33_reps_0"), val = tensor([1, 1, 16, 64])]; + tensor write_indices_float_67_cast_fp16_to_int32 = cast(dtype = var_7170_dtype_0, x = write_indices_float_67_cast_fp16)[name = tensor("cast_472")]; + tensor write_indices_33 = tile(reps = write_indices_33_reps_0, x = write_indices_float_67_cast_fp16_to_int32)[name = tensor("write_indices_33")]; + tensor var_7178_begin_0 = const()[name = tensor("op_7178_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_7178_end_0 = const()[name = tensor("op_7178_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor var_7178_end_mask_0 = const()[name = tensor("op_7178_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_7178_squeeze_mask_0 = const()[name = tensor("op_7178_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor cache16_to_fp16_dtype_0 = const()[name = tensor("cache16_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor cache16_to_fp16 = cast(dtype = cache16_to_fp16_dtype_0, x = cache16)[name = tensor("cast_471")]; + tensor var_7178_cast_fp16 = slice_by_index(begin = var_7178_begin_0, end = var_7178_end_0, end_mask = var_7178_end_mask_0, squeeze_mask = var_7178_squeeze_mask_0, x = cache16_to_fp16)[name = tensor("op_7178_cast_fp16")]; + tensor var_7180_axis_0 = const()[name = tensor("op_7180_axis_0"), val = tensor(1)]; + tensor var_7180_mode_0 = const()[name = tensor("op_7180_mode_0"), val = tensor("update")]; + tensor var_7180_validate_indices_0 = const()[name = tensor("op_7180_validate_indices_0"), val = tensor(false)]; + tensor var_7180_cast_fp16 = scatter_along_axis(axis = var_7180_axis_0, data = var_7178_cast_fp16, indices = write_indices_33, mode = var_7180_mode_0, updates = k_67_cast_fp16, validate_indices = var_7180_validate_indices_0)[name = tensor("op_7180_cast_fp16")]; + tensor concat_114 = const()[name = tensor("concat_114"), val = tensor([0, 0, 0, 0, 0])]; + tensor concat_115 = const()[name = tensor("concat_115"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_33_internal_tensor_assign_1_stride_0 = const()[name = tensor("new_cache_33_internal_tensor_assign_1_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_33_internal_tensor_assign_1_begin_mask_0 = const()[name = tensor("new_cache_33_internal_tensor_assign_1_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_33_internal_tensor_assign_1_end_mask_0 = const()[name = tensor("new_cache_33_internal_tensor_assign_1_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_33_internal_tensor_assign_1_squeeze_mask_0 = const()[name = tensor("new_cache_33_internal_tensor_assign_1_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_80 = const()[name = tensor("shape_80"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_32 = const()[name = tensor("reduce_prod_32"), val = tensor(1048576)]; + tensor range_1d_32_start_0 = const()[name = tensor("range_1d_32_start_0"), val = tensor(0)]; + tensor range_1d_32_step_0 = const()[name = tensor("range_1d_32_step_0"), val = tensor(1)]; + tensor range_1d_32 = range_1d(end = reduce_prod_32, start = range_1d_32_start_0, step = range_1d_32_step_0)[name = tensor("range_1d_32")]; + tensor reshape_160 = reshape(shape = shape_80, x = range_1d_32)[name = tensor("reshape_160")]; + tensor slice_by_index_32 = slice_by_index(begin = concat_114, begin_mask = new_cache_33_internal_tensor_assign_1_begin_mask_0, end = concat_115, end_mask = new_cache_33_internal_tensor_assign_1_end_mask_0, squeeze_mask = new_cache_33_internal_tensor_assign_1_squeeze_mask_0, stride = new_cache_33_internal_tensor_assign_1_stride_0, x = reshape_160)[name = tensor("slice_by_index_32")]; + tensor reshape_161_shape_0 = const()[name = tensor("reshape_161_shape_0"), val = tensor([-1])]; + tensor reshape_161 = reshape(shape = reshape_161_shape_0, x = slice_by_index_32)[name = tensor("reshape_161")]; + tensor reshape_162_shape_0 = const()[name = tensor("reshape_162_shape_0"), val = tensor([-1])]; + tensor reshape_162_cast_fp16 = reshape(shape = reshape_162_shape_0, x = var_7180_cast_fp16)[name = tensor("reshape_162_cast_fp16")]; + tensor reshape_163_shape_0 = const()[name = tensor("reshape_163_shape_0"), val = tensor([-1])]; + tensor reshape_163_cast_fp16 = reshape(shape = reshape_163_shape_0, x = cache16_to_fp16)[name = tensor("reshape_163_cast_fp16")]; + tensor scatter_32_mode_0 = const()[name = tensor("scatter_32_mode_0"), val = tensor("update")]; + tensor scatter_32_axis_0 = const()[name = tensor("scatter_32_axis_0"), val = tensor(0)]; + tensor scatter_32_validate_indices_0 = const()[name = tensor("scatter_32_validate_indices_0"), val = tensor(false)]; + tensor scatter_32_cast_fp16 = scatter(axis = scatter_32_axis_0, data = reshape_163_cast_fp16, indices = reshape_161, mode = scatter_32_mode_0, updates = reshape_162_cast_fp16, validate_indices = scatter_32_validate_indices_0)[name = tensor("scatter_32_cast_fp16")]; + tensor reshape_164_cast_fp16 = reshape(shape = shape_80, x = scatter_32_cast_fp16)[name = tensor("reshape_164_cast_fp16")]; + tensor var_7188_begin_0 = const()[name = tensor("op_7188_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor var_7188_end_0 = const()[name = tensor("op_7188_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor var_7188_end_mask_0 = const()[name = tensor("op_7188_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_7188_squeeze_mask_0 = const()[name = tensor("op_7188_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor var_7188_cast_fp16 = slice_by_index(begin = var_7188_begin_0, end = var_7188_end_0, end_mask = var_7188_end_mask_0, squeeze_mask = var_7188_squeeze_mask_0, x = reshape_164_cast_fp16)[name = tensor("op_7188_cast_fp16")]; + tensor var_7190_axis_0 = const()[name = tensor("op_7190_axis_0"), val = tensor(1)]; + tensor var_7190_mode_0 = const()[name = tensor("op_7190_mode_0"), val = tensor("update")]; + tensor var_7190_validate_indices_0 = const()[name = tensor("op_7190_validate_indices_0"), val = tensor(false)]; + tensor var_7190_cast_fp16 = scatter_along_axis(axis = var_7190_axis_0, data = var_7188_cast_fp16, indices = write_indices_33, mode = var_7190_mode_0, updates = v_33_cast_fp16, validate_indices = var_7190_validate_indices_0)[name = tensor("op_7190_cast_fp16")]; + tensor concat_116 = const()[name = tensor("concat_116"), val = tensor([1, 0, 0, 0, 0])]; + tensor concat_117 = const()[name = tensor("concat_117"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_33_internal_tensor_assign_2_stride_0 = const()[name = tensor("new_cache_33_internal_tensor_assign_2_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_33_internal_tensor_assign_2_begin_mask_0 = const()[name = tensor("new_cache_33_internal_tensor_assign_2_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_33_internal_tensor_assign_2_end_mask_0 = const()[name = tensor("new_cache_33_internal_tensor_assign_2_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_33_internal_tensor_assign_2_squeeze_mask_0 = const()[name = tensor("new_cache_33_internal_tensor_assign_2_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_81 = const()[name = tensor("shape_81"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_33 = const()[name = tensor("reduce_prod_33"), val = tensor(1048576)]; + tensor range_1d_33_start_0 = const()[name = tensor("range_1d_33_start_0"), val = tensor(0)]; + tensor range_1d_33_step_0 = const()[name = tensor("range_1d_33_step_0"), val = tensor(1)]; + tensor range_1d_33 = range_1d(end = reduce_prod_33, start = range_1d_33_start_0, step = range_1d_33_step_0)[name = tensor("range_1d_33")]; + tensor reshape_165 = reshape(shape = shape_81, x = range_1d_33)[name = tensor("reshape_165")]; + tensor slice_by_index_33 = slice_by_index(begin = concat_116, begin_mask = new_cache_33_internal_tensor_assign_2_begin_mask_0, end = concat_117, end_mask = new_cache_33_internal_tensor_assign_2_end_mask_0, squeeze_mask = new_cache_33_internal_tensor_assign_2_squeeze_mask_0, stride = new_cache_33_internal_tensor_assign_2_stride_0, x = reshape_165)[name = tensor("slice_by_index_33")]; + tensor reshape_166_shape_0 = const()[name = tensor("reshape_166_shape_0"), val = tensor([-1])]; + tensor reshape_166 = reshape(shape = reshape_166_shape_0, x = slice_by_index_33)[name = tensor("reshape_166")]; + tensor reshape_167_shape_0 = const()[name = tensor("reshape_167_shape_0"), val = tensor([-1])]; + tensor reshape_167_cast_fp16 = reshape(shape = reshape_167_shape_0, x = var_7190_cast_fp16)[name = tensor("reshape_167_cast_fp16")]; + tensor reshape_168_shape_0 = const()[name = tensor("reshape_168_shape_0"), val = tensor([-1])]; + tensor reshape_168_cast_fp16 = reshape(shape = reshape_168_shape_0, x = reshape_164_cast_fp16)[name = tensor("reshape_168_cast_fp16")]; + tensor scatter_33_mode_0 = const()[name = tensor("scatter_33_mode_0"), val = tensor("update")]; + tensor scatter_33_axis_0 = const()[name = tensor("scatter_33_axis_0"), val = tensor(0)]; + tensor scatter_33_validate_indices_0 = const()[name = tensor("scatter_33_validate_indices_0"), val = tensor(false)]; + tensor scatter_33_cast_fp16 = scatter(axis = scatter_33_axis_0, data = reshape_168_cast_fp16, indices = reshape_166, mode = scatter_33_mode_0, updates = reshape_167_cast_fp16, validate_indices = scatter_33_validate_indices_0)[name = tensor("scatter_33_cast_fp16")]; + tensor reshape_169_cast_fp16 = reshape(shape = shape_81, x = scatter_33_cast_fp16)[name = tensor("reshape_169_cast_fp16")]; + tensor reshape_169_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("reshape_169_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor keys_97_begin_0 = const()[name = tensor("keys_97_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor keys_97_end_0 = const()[name = tensor("keys_97_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor keys_97_end_mask_0 = const()[name = tensor("keys_97_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor keys_97_squeeze_mask_0 = const()[name = tensor("keys_97_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor keys_97_cast_fp16 = slice_by_index(begin = keys_97_begin_0, end = keys_97_end_0, end_mask = keys_97_end_mask_0, squeeze_mask = keys_97_squeeze_mask_0, x = reshape_169_cast_fp16)[name = tensor("keys_97_cast_fp16")]; + tensor values_97_begin_0 = const()[name = tensor("values_97_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor values_97_end_0 = const()[name = tensor("values_97_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor values_97_end_mask_0 = const()[name = tensor("values_97_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor values_97_squeeze_mask_0 = const()[name = tensor("values_97_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor values_97_cast_fp16 = slice_by_index(begin = values_97_begin_0, end = values_97_end_0, end_mask = values_97_end_mask_0, squeeze_mask = values_97_squeeze_mask_0, x = reshape_169_cast_fp16)[name = tensor("values_97_cast_fp16")]; + tensor var_7202_cast_fp16 = not_equal(x = keys_97_cast_fp16, y = keys_97_cast_fp16)[name = tensor("op_7202_cast_fp16")]; + tensor keys_99_cast_fp16 = select(a = var_504_to_fp16, b = keys_97_cast_fp16, cond = var_7202_cast_fp16)[name = tensor("keys_99_cast_fp16")]; + tensor var_7210_cast_fp16 = not_equal(x = values_97_cast_fp16, y = values_97_cast_fp16)[name = tensor("op_7210_cast_fp16")]; + tensor values_99_cast_fp16 = select(a = var_504_to_fp16, b = values_97_cast_fp16, cond = var_7210_cast_fp16)[name = tensor("values_99_cast_fp16")]; + tensor var_7234 = const()[name = tensor("op_7234"), val = tensor([0, 2, 1, 3])]; + tensor var_7247 = const()[name = tensor("op_7247"), val = tensor([1, 1, 1])]; + tensor var_7248_cast_fp16 = reshape(shape = var_7247, x = position16_to_fp16)[name = tensor("op_7248_cast_fp16")]; + tensor var_7265_to_fp16 = const()[name = tensor("op_7265_to_fp16"), val = tensor(0x1p+0)]; + tensor valid_len_33_cast_fp16 = add(x = var_7248_cast_fp16, y = var_7265_to_fp16)[name = tensor("valid_len_33_cast_fp16")]; + tensor valid_mask_33_cast_fp16 = less(x = k_positions_1_promoted_to_fp16, y = valid_len_33_cast_fp16)[name = tensor("valid_mask_33_cast_fp16")]; + tensor causal_mask_33_cast_fp16 = less_equal(x = k_positions_1_promoted_to_fp16, y = var_7248_cast_fp16)[name = tensor("causal_mask_33_cast_fp16")]; + tensor attn_mask_65 = logical_and(x = valid_mask_33_cast_fp16, y = causal_mask_33_cast_fp16)[name = tensor("attn_mask_65")]; + tensor attn_mask_67_axes_0 = const()[name = tensor("attn_mask_67_axes_0"), val = tensor([1])]; + tensor attn_mask_67 = expand_dims(axes = attn_mask_67_axes_0, x = attn_mask_65)[name = tensor("attn_mask_67")]; + tensor var_7283_transpose_x_0 = const()[name = tensor("op_7283_transpose_x_0"), val = tensor(false)]; + tensor var_7283_transpose_y_0 = const()[name = tensor("op_7283_transpose_y_0"), val = tensor(false)]; + tensor transpose_104_perm_0 = const()[name = tensor("transpose_104_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_105_perm_0 = const()[name = tensor("transpose_105_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_105 = transpose(perm = transpose_105_perm_0, x = keys_99_cast_fp16)[name = tensor("transpose_149")]; + tensor transpose_104 = transpose(perm = transpose_104_perm_0, x = q_99_cast_fp16)[name = tensor("transpose_150")]; + tensor var_7283_cast_fp16 = matmul(transpose_x = var_7283_transpose_x_0, transpose_y = var_7283_transpose_y_0, x = transpose_104, y = transpose_105)[name = tensor("op_7283_cast_fp16")]; + tensor var_7277_to_fp16 = const()[name = tensor("op_7277_to_fp16"), val = tensor([0x1p-3])]; + tensor attn_weights_97_cast_fp16 = mul(x = var_7283_cast_fp16, y = var_7277_to_fp16)[name = tensor("attn_weights_97_cast_fp16")]; + tensor var_7285 = logical_not(x = attn_mask_67)[name = tensor("op_7285")]; + tensor var_7286_to_fp16 = const()[name = tensor("op_7286_to_fp16"), val = tensor(-inf)]; + tensor attn_weights_99_cast_fp16 = select(a = var_7286_to_fp16, b = attn_weights_97_cast_fp16, cond = var_7285)[name = tensor("attn_weights_99_cast_fp16")]; + tensor var_7288 = const()[name = tensor("op_7288"), val = tensor(-1)]; + tensor attn_weights_101_cast_fp16 = softmax(axis = var_7288, x = attn_weights_99_cast_fp16)[name = tensor("attn_weights_101_cast_fp16")]; + tensor attn_output_33_transpose_x_0 = const()[name = tensor("attn_output_33_transpose_x_0"), val = tensor(false)]; + tensor attn_output_33_transpose_y_0 = const()[name = tensor("attn_output_33_transpose_y_0"), val = tensor(false)]; + tensor values_101_cast_fp16 = transpose(perm = var_7234, x = values_99_cast_fp16)[name = tensor("transpose_151")]; + tensor attn_output_33_cast_fp16 = matmul(transpose_x = attn_output_33_transpose_x_0, transpose_y = attn_output_33_transpose_y_0, x = attn_weights_101_cast_fp16, y = values_101_cast_fp16)[name = tensor("attn_output_33_cast_fp16")]; + tensor var_7296 = const()[name = tensor("op_7296"), val = tensor([0, 2, 1, 3])]; + tensor var_7299 = const()[name = tensor("op_7299"), val = tensor([1, 1, 1024])]; + tensor var_7297_cast_fp16 = transpose(perm = var_7296, x = attn_output_33_cast_fp16)[name = tensor("transpose_148")]; + tensor input_165_cast_fp16 = reshape(shape = var_7299, x = var_7297_cast_fp16)[name = tensor("input_165_cast_fp16")]; + tensor linear_66_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = attn16_out_proj_weight_cast_fp16, x = input_165_cast_fp16)[name = tensor("linear_66_cast_fp16")]; + tensor var_7305_to_fp16 = const()[name = tensor("op_7305_to_fp16"), val = tensor(0x1p+0)]; + tensor var_7306_cast_fp16 = add(x = position16_to_fp16, y = var_7305_to_fp16)[name = tensor("op_7306_cast_fp16")]; + tensor var_7306_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_7306_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor input_167_cast_fp16 = add(x = input_163_cast_fp16, y = linear_66_cast_fp16)[name = tensor("input_167_cast_fp16")]; + tensor input_169_axes_0 = const()[name = tensor("input_169_axes_0"), val = tensor([-1])]; + tensor norm16_2_weight_to_fp16 = const()[name = tensor("norm16_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(303726464)))]; + tensor norm16_2_bias_to_fp16 = const()[name = tensor("norm16_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(303728576)))]; + tensor var_7310_to_fp16 = const()[name = tensor("op_7310_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_169_cast_fp16 = layer_norm(axes = input_169_axes_0, beta = norm16_2_bias_to_fp16, epsilon = var_7310_to_fp16, gamma = norm16_2_weight_to_fp16, x = input_167_cast_fp16)[name = tensor("input_169_cast_fp16")]; + tensor linear_67_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = linear16_1_weight_cast_fp16, x = input_169_cast_fp16)[name = tensor("linear_67_cast_fp16")]; + tensor input_171_mode_0 = const()[name = tensor("input_171_mode_0"), val = tensor("EXACT")]; + tensor input_171_cast_fp16 = gelu(mode = input_171_mode_0, x = linear_67_cast_fp16)[name = tensor("input_171_cast_fp16")]; + tensor linear_68_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = linear16_2_weight_cast_fp16, x = input_171_cast_fp16)[name = tensor("linear_68_cast_fp16")]; + tensor input_173_cast_fp16 = add(x = input_167_cast_fp16, y = linear_68_cast_fp16)[name = tensor("input_173_cast_fp16")]; + tensor x_35_axes_0 = const()[name = tensor("x_35_axes_0"), val = tensor([-1])]; + tensor norm17_1_weight_to_fp16 = const()[name = tensor("norm17_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(303730688)))]; + tensor norm17_1_bias_to_fp16 = const()[name = tensor("norm17_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(303732800)))]; + tensor var_7327_to_fp16 = const()[name = tensor("op_7327_to_fp16"), val = tensor(0x1.5p-17)]; + tensor x_35_cast_fp16 = layer_norm(axes = x_35_axes_0, beta = norm17_1_bias_to_fp16, epsilon = var_7327_to_fp16, gamma = norm17_1_weight_to_fp16, x = input_173_cast_fp16)[name = tensor("x_35_cast_fp16")]; + tensor linear_69_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = attn17_in_proj_weight_cast_fp16, x = x_35_cast_fp16)[name = tensor("linear_69_cast_fp16")]; + tensor var_7363 = const()[name = tensor("op_7363"), val = tensor([1, 1, 3, 16, 64])]; + tensor qkv_35_cast_fp16 = reshape(shape = var_7363, x = linear_69_cast_fp16)[name = tensor("qkv_35_cast_fp16")]; + tensor q_103_begin_0 = const()[name = tensor("q_103_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor q_103_end_0 = const()[name = tensor("q_103_end_0"), val = tensor([1, 1, 1, 16, 64])]; + tensor q_103_end_mask_0 = const()[name = tensor("q_103_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor q_103_squeeze_mask_0 = const()[name = tensor("q_103_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor q_103_cast_fp16 = slice_by_index(begin = q_103_begin_0, end = q_103_end_0, end_mask = q_103_end_mask_0, squeeze_mask = q_103_squeeze_mask_0, x = qkv_35_cast_fp16)[name = tensor("q_103_cast_fp16")]; + tensor k_69_begin_0 = const()[name = tensor("k_69_begin_0"), val = tensor([0, 0, 1, 0, 0])]; + tensor k_69_end_0 = const()[name = tensor("k_69_end_0"), val = tensor([1, 1, 2, 16, 64])]; + tensor k_69_end_mask_0 = const()[name = tensor("k_69_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor k_69_squeeze_mask_0 = const()[name = tensor("k_69_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor k_69_cast_fp16 = slice_by_index(begin = k_69_begin_0, end = k_69_end_0, end_mask = k_69_end_mask_0, squeeze_mask = k_69_squeeze_mask_0, x = qkv_35_cast_fp16)[name = tensor("k_69_cast_fp16")]; + tensor v_35_begin_0 = const()[name = tensor("v_35_begin_0"), val = tensor([0, 0, 2, 0, 0])]; + tensor v_35_end_0 = const()[name = tensor("v_35_end_0"), val = tensor([1, 1, 3, 16, 64])]; + tensor v_35_end_mask_0 = const()[name = tensor("v_35_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor v_35_squeeze_mask_0 = const()[name = tensor("v_35_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor v_35_cast_fp16 = slice_by_index(begin = v_35_begin_0, end = v_35_end_0, end_mask = v_35_end_mask_0, squeeze_mask = v_35_squeeze_mask_0, x = qkv_35_cast_fp16)[name = tensor("v_35_cast_fp16")]; + tensor var_7467 = const()[name = tensor("op_7467"), val = tensor([1, 1, 1, 1])]; + tensor position17_to_fp16_dtype_0 = const()[name = tensor("position17_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor position17_to_fp16 = cast(dtype = position17_to_fp16_dtype_0, x = position17)[name = tensor("cast_468")]; + tensor ts_107_cast_fp16 = reshape(shape = var_7467, x = position17_to_fp16)[name = tensor("ts_107_cast_fp16")]; + tensor var_7471 = const()[name = tensor("op_7471"), val = tensor([1, 1, 16, 32, 2])]; + tensor q_complex_35_cast_fp16 = reshape(shape = var_7471, x = q_103_cast_fp16)[name = tensor("q_complex_35_cast_fp16")]; + tensor var_7475 = const()[name = tensor("op_7475"), val = tensor([1, 1, 16, 32, 2])]; + tensor k_complex_35_cast_fp16 = reshape(shape = var_7475, x = k_69_cast_fp16)[name = tensor("k_complex_35_cast_fp16")]; + tensor var_7479_begin_0 = const()[name = tensor("op_7479_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_7479_end_0 = const()[name = tensor("op_7479_end_0"), val = tensor([1, 1, 16, 32, 1])]; + tensor var_7479_end_mask_0 = const()[name = tensor("op_7479_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_7479_squeeze_mask_0 = const()[name = tensor("op_7479_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_7479_cast_fp16 = slice_by_index(begin = var_7479_begin_0, end = var_7479_end_0, end_mask = var_7479_end_mask_0, squeeze_mask = var_7479_squeeze_mask_0, x = q_complex_35_cast_fp16)[name = tensor("op_7479_cast_fp16")]; + tensor var_7487_begin_0 = const()[name = tensor("op_7487_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_7487_end_0 = const()[name = tensor("op_7487_end_0"), val = tensor([1, 1, 16, 32, 2])]; + tensor var_7487_end_mask_0 = const()[name = tensor("op_7487_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_7487_squeeze_mask_0 = const()[name = tensor("op_7487_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_7487_cast_fp16 = slice_by_index(begin = var_7487_begin_0, end = var_7487_end_0, end_mask = var_7487_end_mask_0, squeeze_mask = var_7487_squeeze_mask_0, x = q_complex_35_cast_fp16)[name = tensor("op_7487_cast_fp16")]; + tensor var_7495_begin_0 = const()[name = tensor("op_7495_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_7495_end_0 = const()[name = tensor("op_7495_end_0"), val = tensor([1, 1, 16, 32, 1])]; + tensor var_7495_end_mask_0 = const()[name = tensor("op_7495_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_7495_squeeze_mask_0 = const()[name = tensor("op_7495_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_7495_cast_fp16 = slice_by_index(begin = var_7495_begin_0, end = var_7495_end_0, end_mask = var_7495_end_mask_0, squeeze_mask = var_7495_squeeze_mask_0, x = k_complex_35_cast_fp16)[name = tensor("op_7495_cast_fp16")]; + tensor var_7503_begin_0 = const()[name = tensor("op_7503_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_7503_end_0 = const()[name = tensor("op_7503_end_0"), val = tensor([1, 1, 16, 32, 2])]; + tensor var_7503_end_mask_0 = const()[name = tensor("op_7503_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_7503_squeeze_mask_0 = const()[name = tensor("op_7503_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_7503_cast_fp16 = slice_by_index(begin = var_7503_begin_0, end = var_7503_end_0, end_mask = var_7503_end_mask_0, squeeze_mask = var_7503_squeeze_mask_0, x = k_complex_35_cast_fp16)[name = tensor("op_7503_cast_fp16")]; + tensor freqs_35_to_fp16 = const()[name = tensor("freqs_35_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(303734912)))]; + tensor var_7509_cast_fp16 = mul(x = freqs_35_to_fp16, y = ts_107_cast_fp16)[name = tensor("op_7509_cast_fp16")]; + tensor rotr_35_cast_fp16 = cos(x = var_7509_cast_fp16)[name = tensor("rotr_35_cast_fp16")]; + tensor roti_35_cast_fp16 = sin(x = var_7509_cast_fp16)[name = tensor("roti_35_cast_fp16")]; + tensor var_7513_cast_fp16 = mul(x = var_7479_cast_fp16, y = rotr_35_cast_fp16)[name = tensor("op_7513_cast_fp16")]; + tensor var_7514_cast_fp16 = mul(x = var_7487_cast_fp16, y = roti_35_cast_fp16)[name = tensor("op_7514_cast_fp16")]; + tensor qor_69_cast_fp16 = sub(x = var_7513_cast_fp16, y = var_7514_cast_fp16)[name = tensor("qor_69_cast_fp16")]; + tensor var_7517_cast_fp16 = mul(x = var_7479_cast_fp16, y = roti_35_cast_fp16)[name = tensor("op_7517_cast_fp16")]; + tensor var_7518_cast_fp16 = mul(x = var_7487_cast_fp16, y = rotr_35_cast_fp16)[name = tensor("op_7518_cast_fp16")]; + tensor qoi_69_cast_fp16 = add(x = var_7517_cast_fp16, y = var_7518_cast_fp16)[name = tensor("qoi_69_cast_fp16")]; + tensor var_7521_cast_fp16 = mul(x = var_7495_cast_fp16, y = rotr_35_cast_fp16)[name = tensor("op_7521_cast_fp16")]; + tensor var_7522_cast_fp16 = mul(x = var_7503_cast_fp16, y = roti_35_cast_fp16)[name = tensor("op_7522_cast_fp16")]; + tensor kor_69_cast_fp16 = sub(x = var_7521_cast_fp16, y = var_7522_cast_fp16)[name = tensor("kor_69_cast_fp16")]; + tensor var_7525_cast_fp16 = mul(x = var_7495_cast_fp16, y = roti_35_cast_fp16)[name = tensor("op_7525_cast_fp16")]; + tensor var_7526_cast_fp16 = mul(x = var_7503_cast_fp16, y = rotr_35_cast_fp16)[name = tensor("op_7526_cast_fp16")]; + tensor koi_69_cast_fp16 = add(x = var_7525_cast_fp16, y = var_7526_cast_fp16)[name = tensor("koi_69_cast_fp16")]; + tensor qo_35_axis_0 = const()[name = tensor("qo_35_axis_0"), val = tensor(-1)]; + tensor qo_35_cast_fp16 = stack(axis = qo_35_axis_0, values = (qor_69_cast_fp16, qoi_69_cast_fp16))[name = tensor("qo_35_cast_fp16")]; + tensor ko_35_axis_0 = const()[name = tensor("ko_35_axis_0"), val = tensor(-1)]; + tensor ko_35_cast_fp16 = stack(axis = ko_35_axis_0, values = (kor_69_cast_fp16, koi_69_cast_fp16))[name = tensor("ko_35_cast_fp16")]; + tensor var_7555 = const()[name = tensor("op_7555"), val = tensor([1, 1, 16, 64])]; + tensor q_105_cast_fp16 = reshape(shape = var_7555, x = qo_35_cast_fp16)[name = tensor("q_105_cast_fp16")]; + tensor var_7557 = const()[name = tensor("op_7557"), val = tensor([1, 1, 16, 64])]; + tensor k_71_cast_fp16 = reshape(shape = var_7557, x = ko_35_cast_fp16)[name = tensor("k_71_cast_fp16")]; + tensor _inversed_7579_y_0_to_fp16 = const()[name = tensor("_inversed_7579_y_0_to_fp16"), val = tensor(0x1p-9)]; + tensor _inversed_7579_cast_fp16 = mul(x = ts_107_cast_fp16, y = _inversed_7579_y_0_to_fp16)[name = tensor("_inversed_7579_cast_fp16")]; + tensor var_7580_cast_fp16 = floor(x = _inversed_7579_cast_fp16)[name = tensor("op_7580_cast_fp16")]; + tensor var_7581_to_fp16 = const()[name = tensor("op_7581_to_fp16"), val = tensor(0x1p+9)]; + tensor var_7582_cast_fp16 = mul(x = var_7580_cast_fp16, y = var_7581_to_fp16)[name = tensor("op_7582_cast_fp16")]; + tensor write_indices_float_71_cast_fp16 = sub(x = ts_107_cast_fp16, y = var_7582_cast_fp16)[name = tensor("write_indices_float_71_cast_fp16")]; + tensor var_7589_dtype_0 = const()[name = tensor("op_7589_dtype_0"), val = tensor("int32")]; + tensor write_indices_35_reps_0 = const()[name = tensor("write_indices_35_reps_0"), val = tensor([1, 1, 16, 64])]; + tensor write_indices_float_71_cast_fp16_to_int32 = cast(dtype = var_7589_dtype_0, x = write_indices_float_71_cast_fp16)[name = tensor("cast_467")]; + tensor write_indices_35 = tile(reps = write_indices_35_reps_0, x = write_indices_float_71_cast_fp16_to_int32)[name = tensor("write_indices_35")]; + tensor var_7597_begin_0 = const()[name = tensor("op_7597_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_7597_end_0 = const()[name = tensor("op_7597_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor var_7597_end_mask_0 = const()[name = tensor("op_7597_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_7597_squeeze_mask_0 = const()[name = tensor("op_7597_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor cache17_to_fp16_dtype_0 = const()[name = tensor("cache17_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor cache17_to_fp16 = cast(dtype = cache17_to_fp16_dtype_0, x = cache17)[name = tensor("cast_466")]; + tensor var_7597_cast_fp16 = slice_by_index(begin = var_7597_begin_0, end = var_7597_end_0, end_mask = var_7597_end_mask_0, squeeze_mask = var_7597_squeeze_mask_0, x = cache17_to_fp16)[name = tensor("op_7597_cast_fp16")]; + tensor var_7599_axis_0 = const()[name = tensor("op_7599_axis_0"), val = tensor(1)]; + tensor var_7599_mode_0 = const()[name = tensor("op_7599_mode_0"), val = tensor("update")]; + tensor var_7599_validate_indices_0 = const()[name = tensor("op_7599_validate_indices_0"), val = tensor(false)]; + tensor var_7599_cast_fp16 = scatter_along_axis(axis = var_7599_axis_0, data = var_7597_cast_fp16, indices = write_indices_35, mode = var_7599_mode_0, updates = k_71_cast_fp16, validate_indices = var_7599_validate_indices_0)[name = tensor("op_7599_cast_fp16")]; + tensor concat_121 = const()[name = tensor("concat_121"), val = tensor([0, 0, 0, 0, 0])]; + tensor concat_122 = const()[name = tensor("concat_122"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_35_internal_tensor_assign_1_stride_0 = const()[name = tensor("new_cache_35_internal_tensor_assign_1_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_35_internal_tensor_assign_1_begin_mask_0 = const()[name = tensor("new_cache_35_internal_tensor_assign_1_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_35_internal_tensor_assign_1_end_mask_0 = const()[name = tensor("new_cache_35_internal_tensor_assign_1_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_35_internal_tensor_assign_1_squeeze_mask_0 = const()[name = tensor("new_cache_35_internal_tensor_assign_1_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_82 = const()[name = tensor("shape_82"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_34 = const()[name = tensor("reduce_prod_34"), val = tensor(1048576)]; + tensor range_1d_34_start_0 = const()[name = tensor("range_1d_34_start_0"), val = tensor(0)]; + tensor range_1d_34_step_0 = const()[name = tensor("range_1d_34_step_0"), val = tensor(1)]; + tensor range_1d_34 = range_1d(end = reduce_prod_34, start = range_1d_34_start_0, step = range_1d_34_step_0)[name = tensor("range_1d_34")]; + tensor reshape_170 = reshape(shape = shape_82, x = range_1d_34)[name = tensor("reshape_170")]; + tensor slice_by_index_34 = slice_by_index(begin = concat_121, begin_mask = new_cache_35_internal_tensor_assign_1_begin_mask_0, end = concat_122, end_mask = new_cache_35_internal_tensor_assign_1_end_mask_0, squeeze_mask = new_cache_35_internal_tensor_assign_1_squeeze_mask_0, stride = new_cache_35_internal_tensor_assign_1_stride_0, x = reshape_170)[name = tensor("slice_by_index_34")]; + tensor reshape_171_shape_0 = const()[name = tensor("reshape_171_shape_0"), val = tensor([-1])]; + tensor reshape_171 = reshape(shape = reshape_171_shape_0, x = slice_by_index_34)[name = tensor("reshape_171")]; + tensor reshape_172_shape_0 = const()[name = tensor("reshape_172_shape_0"), val = tensor([-1])]; + tensor reshape_172_cast_fp16 = reshape(shape = reshape_172_shape_0, x = var_7599_cast_fp16)[name = tensor("reshape_172_cast_fp16")]; + tensor reshape_173_shape_0 = const()[name = tensor("reshape_173_shape_0"), val = tensor([-1])]; + tensor reshape_173_cast_fp16 = reshape(shape = reshape_173_shape_0, x = cache17_to_fp16)[name = tensor("reshape_173_cast_fp16")]; + tensor scatter_34_mode_0 = const()[name = tensor("scatter_34_mode_0"), val = tensor("update")]; + tensor scatter_34_axis_0 = const()[name = tensor("scatter_34_axis_0"), val = tensor(0)]; + tensor scatter_34_validate_indices_0 = const()[name = tensor("scatter_34_validate_indices_0"), val = tensor(false)]; + tensor scatter_34_cast_fp16 = scatter(axis = scatter_34_axis_0, data = reshape_173_cast_fp16, indices = reshape_171, mode = scatter_34_mode_0, updates = reshape_172_cast_fp16, validate_indices = scatter_34_validate_indices_0)[name = tensor("scatter_34_cast_fp16")]; + tensor reshape_174_cast_fp16 = reshape(shape = shape_82, x = scatter_34_cast_fp16)[name = tensor("reshape_174_cast_fp16")]; + tensor var_7607_begin_0 = const()[name = tensor("op_7607_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor var_7607_end_0 = const()[name = tensor("op_7607_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor var_7607_end_mask_0 = const()[name = tensor("op_7607_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_7607_squeeze_mask_0 = const()[name = tensor("op_7607_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor var_7607_cast_fp16 = slice_by_index(begin = var_7607_begin_0, end = var_7607_end_0, end_mask = var_7607_end_mask_0, squeeze_mask = var_7607_squeeze_mask_0, x = reshape_174_cast_fp16)[name = tensor("op_7607_cast_fp16")]; + tensor var_7609_axis_0 = const()[name = tensor("op_7609_axis_0"), val = tensor(1)]; + tensor var_7609_mode_0 = const()[name = tensor("op_7609_mode_0"), val = tensor("update")]; + tensor var_7609_validate_indices_0 = const()[name = tensor("op_7609_validate_indices_0"), val = tensor(false)]; + tensor var_7609_cast_fp16 = scatter_along_axis(axis = var_7609_axis_0, data = var_7607_cast_fp16, indices = write_indices_35, mode = var_7609_mode_0, updates = v_35_cast_fp16, validate_indices = var_7609_validate_indices_0)[name = tensor("op_7609_cast_fp16")]; + tensor concat_123 = const()[name = tensor("concat_123"), val = tensor([1, 0, 0, 0, 0])]; + tensor concat_124 = const()[name = tensor("concat_124"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_35_internal_tensor_assign_2_stride_0 = const()[name = tensor("new_cache_35_internal_tensor_assign_2_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_35_internal_tensor_assign_2_begin_mask_0 = const()[name = tensor("new_cache_35_internal_tensor_assign_2_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_35_internal_tensor_assign_2_end_mask_0 = const()[name = tensor("new_cache_35_internal_tensor_assign_2_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_35_internal_tensor_assign_2_squeeze_mask_0 = const()[name = tensor("new_cache_35_internal_tensor_assign_2_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_83 = const()[name = tensor("shape_83"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_35 = const()[name = tensor("reduce_prod_35"), val = tensor(1048576)]; + tensor range_1d_35_start_0 = const()[name = tensor("range_1d_35_start_0"), val = tensor(0)]; + tensor range_1d_35_step_0 = const()[name = tensor("range_1d_35_step_0"), val = tensor(1)]; + tensor range_1d_35 = range_1d(end = reduce_prod_35, start = range_1d_35_start_0, step = range_1d_35_step_0)[name = tensor("range_1d_35")]; + tensor reshape_175 = reshape(shape = shape_83, x = range_1d_35)[name = tensor("reshape_175")]; + tensor slice_by_index_35 = slice_by_index(begin = concat_123, begin_mask = new_cache_35_internal_tensor_assign_2_begin_mask_0, end = concat_124, end_mask = new_cache_35_internal_tensor_assign_2_end_mask_0, squeeze_mask = new_cache_35_internal_tensor_assign_2_squeeze_mask_0, stride = new_cache_35_internal_tensor_assign_2_stride_0, x = reshape_175)[name = tensor("slice_by_index_35")]; + tensor reshape_176_shape_0 = const()[name = tensor("reshape_176_shape_0"), val = tensor([-1])]; + tensor reshape_176 = reshape(shape = reshape_176_shape_0, x = slice_by_index_35)[name = tensor("reshape_176")]; + tensor reshape_177_shape_0 = const()[name = tensor("reshape_177_shape_0"), val = tensor([-1])]; + tensor reshape_177_cast_fp16 = reshape(shape = reshape_177_shape_0, x = var_7609_cast_fp16)[name = tensor("reshape_177_cast_fp16")]; + tensor reshape_178_shape_0 = const()[name = tensor("reshape_178_shape_0"), val = tensor([-1])]; + tensor reshape_178_cast_fp16 = reshape(shape = reshape_178_shape_0, x = reshape_174_cast_fp16)[name = tensor("reshape_178_cast_fp16")]; + tensor scatter_35_mode_0 = const()[name = tensor("scatter_35_mode_0"), val = tensor("update")]; + tensor scatter_35_axis_0 = const()[name = tensor("scatter_35_axis_0"), val = tensor(0)]; + tensor scatter_35_validate_indices_0 = const()[name = tensor("scatter_35_validate_indices_0"), val = tensor(false)]; + tensor scatter_35_cast_fp16 = scatter(axis = scatter_35_axis_0, data = reshape_178_cast_fp16, indices = reshape_176, mode = scatter_35_mode_0, updates = reshape_177_cast_fp16, validate_indices = scatter_35_validate_indices_0)[name = tensor("scatter_35_cast_fp16")]; + tensor reshape_179_cast_fp16 = reshape(shape = shape_83, x = scatter_35_cast_fp16)[name = tensor("reshape_179_cast_fp16")]; + tensor reshape_179_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("reshape_179_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor keys_103_begin_0 = const()[name = tensor("keys_103_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor keys_103_end_0 = const()[name = tensor("keys_103_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor keys_103_end_mask_0 = const()[name = tensor("keys_103_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor keys_103_squeeze_mask_0 = const()[name = tensor("keys_103_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor keys_103_cast_fp16 = slice_by_index(begin = keys_103_begin_0, end = keys_103_end_0, end_mask = keys_103_end_mask_0, squeeze_mask = keys_103_squeeze_mask_0, x = reshape_179_cast_fp16)[name = tensor("keys_103_cast_fp16")]; + tensor values_103_begin_0 = const()[name = tensor("values_103_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor values_103_end_0 = const()[name = tensor("values_103_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor values_103_end_mask_0 = const()[name = tensor("values_103_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor values_103_squeeze_mask_0 = const()[name = tensor("values_103_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor values_103_cast_fp16 = slice_by_index(begin = values_103_begin_0, end = values_103_end_0, end_mask = values_103_end_mask_0, squeeze_mask = values_103_squeeze_mask_0, x = reshape_179_cast_fp16)[name = tensor("values_103_cast_fp16")]; + tensor var_7621_cast_fp16 = not_equal(x = keys_103_cast_fp16, y = keys_103_cast_fp16)[name = tensor("op_7621_cast_fp16")]; + tensor keys_105_cast_fp16 = select(a = var_504_to_fp16, b = keys_103_cast_fp16, cond = var_7621_cast_fp16)[name = tensor("keys_105_cast_fp16")]; + tensor var_7629_cast_fp16 = not_equal(x = values_103_cast_fp16, y = values_103_cast_fp16)[name = tensor("op_7629_cast_fp16")]; + tensor values_105_cast_fp16 = select(a = var_504_to_fp16, b = values_103_cast_fp16, cond = var_7629_cast_fp16)[name = tensor("values_105_cast_fp16")]; + tensor var_7653 = const()[name = tensor("op_7653"), val = tensor([0, 2, 1, 3])]; + tensor var_7666 = const()[name = tensor("op_7666"), val = tensor([1, 1, 1])]; + tensor var_7667_cast_fp16 = reshape(shape = var_7666, x = position17_to_fp16)[name = tensor("op_7667_cast_fp16")]; + tensor var_7684_to_fp16 = const()[name = tensor("op_7684_to_fp16"), val = tensor(0x1p+0)]; + tensor valid_len_35_cast_fp16 = add(x = var_7667_cast_fp16, y = var_7684_to_fp16)[name = tensor("valid_len_35_cast_fp16")]; + tensor valid_mask_35_cast_fp16 = less(x = k_positions_1_promoted_to_fp16, y = valid_len_35_cast_fp16)[name = tensor("valid_mask_35_cast_fp16")]; + tensor causal_mask_35_cast_fp16 = less_equal(x = k_positions_1_promoted_to_fp16, y = var_7667_cast_fp16)[name = tensor("causal_mask_35_cast_fp16")]; + tensor attn_mask_69 = logical_and(x = valid_mask_35_cast_fp16, y = causal_mask_35_cast_fp16)[name = tensor("attn_mask_69")]; + tensor attn_mask_71_axes_0 = const()[name = tensor("attn_mask_71_axes_0"), val = tensor([1])]; + tensor attn_mask_71 = expand_dims(axes = attn_mask_71_axes_0, x = attn_mask_69)[name = tensor("attn_mask_71")]; + tensor var_7702_transpose_x_0 = const()[name = tensor("op_7702_transpose_x_0"), val = tensor(false)]; + tensor var_7702_transpose_y_0 = const()[name = tensor("op_7702_transpose_y_0"), val = tensor(false)]; + tensor transpose_106_perm_0 = const()[name = tensor("transpose_106_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_107_perm_0 = const()[name = tensor("transpose_107_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_107 = transpose(perm = transpose_107_perm_0, x = keys_105_cast_fp16)[name = tensor("transpose_145")]; + tensor transpose_106 = transpose(perm = transpose_106_perm_0, x = q_105_cast_fp16)[name = tensor("transpose_146")]; + tensor var_7702_cast_fp16 = matmul(transpose_x = var_7702_transpose_x_0, transpose_y = var_7702_transpose_y_0, x = transpose_106, y = transpose_107)[name = tensor("op_7702_cast_fp16")]; + tensor var_7696_to_fp16 = const()[name = tensor("op_7696_to_fp16"), val = tensor([0x1p-3])]; + tensor attn_weights_103_cast_fp16 = mul(x = var_7702_cast_fp16, y = var_7696_to_fp16)[name = tensor("attn_weights_103_cast_fp16")]; + tensor var_7704 = logical_not(x = attn_mask_71)[name = tensor("op_7704")]; + tensor var_7705_to_fp16 = const()[name = tensor("op_7705_to_fp16"), val = tensor(-inf)]; + tensor attn_weights_105_cast_fp16 = select(a = var_7705_to_fp16, b = attn_weights_103_cast_fp16, cond = var_7704)[name = tensor("attn_weights_105_cast_fp16")]; + tensor var_7707 = const()[name = tensor("op_7707"), val = tensor(-1)]; + tensor attn_weights_107_cast_fp16 = softmax(axis = var_7707, x = attn_weights_105_cast_fp16)[name = tensor("attn_weights_107_cast_fp16")]; + tensor attn_output_35_transpose_x_0 = const()[name = tensor("attn_output_35_transpose_x_0"), val = tensor(false)]; + tensor attn_output_35_transpose_y_0 = const()[name = tensor("attn_output_35_transpose_y_0"), val = tensor(false)]; + tensor values_107_cast_fp16 = transpose(perm = var_7653, x = values_105_cast_fp16)[name = tensor("transpose_147")]; + tensor attn_output_35_cast_fp16 = matmul(transpose_x = attn_output_35_transpose_x_0, transpose_y = attn_output_35_transpose_y_0, x = attn_weights_107_cast_fp16, y = values_107_cast_fp16)[name = tensor("attn_output_35_cast_fp16")]; + tensor var_7715 = const()[name = tensor("op_7715"), val = tensor([0, 2, 1, 3])]; + tensor var_7718 = const()[name = tensor("op_7718"), val = tensor([1, 1, 1024])]; + tensor var_7716_cast_fp16 = transpose(perm = var_7715, x = attn_output_35_cast_fp16)[name = tensor("transpose_144")]; + tensor input_175_cast_fp16 = reshape(shape = var_7718, x = var_7716_cast_fp16)[name = tensor("input_175_cast_fp16")]; + tensor linear_70_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = attn17_out_proj_weight_cast_fp16, x = input_175_cast_fp16)[name = tensor("linear_70_cast_fp16")]; + tensor var_7724_to_fp16 = const()[name = tensor("op_7724_to_fp16"), val = tensor(0x1p+0)]; + tensor var_7725_cast_fp16 = add(x = position17_to_fp16, y = var_7724_to_fp16)[name = tensor("op_7725_cast_fp16")]; + tensor var_7725_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_7725_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor input_177_cast_fp16 = add(x = input_173_cast_fp16, y = linear_70_cast_fp16)[name = tensor("input_177_cast_fp16")]; + tensor input_179_axes_0 = const()[name = tensor("input_179_axes_0"), val = tensor([-1])]; + tensor norm17_2_weight_to_fp16 = const()[name = tensor("norm17_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(303735040)))]; + tensor norm17_2_bias_to_fp16 = const()[name = tensor("norm17_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(303737152)))]; + tensor var_7729_to_fp16 = const()[name = tensor("op_7729_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_179_cast_fp16 = layer_norm(axes = input_179_axes_0, beta = norm17_2_bias_to_fp16, epsilon = var_7729_to_fp16, gamma = norm17_2_weight_to_fp16, x = input_177_cast_fp16)[name = tensor("input_179_cast_fp16")]; + tensor linear_71_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = linear17_1_weight_cast_fp16, x = input_179_cast_fp16)[name = tensor("linear_71_cast_fp16")]; + tensor input_181_mode_0 = const()[name = tensor("input_181_mode_0"), val = tensor("EXACT")]; + tensor input_181_cast_fp16 = gelu(mode = input_181_mode_0, x = linear_71_cast_fp16)[name = tensor("input_181_cast_fp16")]; + tensor linear_72_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = linear17_2_weight_cast_fp16, x = input_181_cast_fp16)[name = tensor("linear_72_cast_fp16")]; + tensor input_183_cast_fp16 = add(x = input_177_cast_fp16, y = linear_72_cast_fp16)[name = tensor("input_183_cast_fp16")]; + tensor x_37_axes_0 = const()[name = tensor("x_37_axes_0"), val = tensor([-1])]; + tensor norm18_1_weight_to_fp16 = const()[name = tensor("norm18_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(303739264)))]; + tensor norm18_1_bias_to_fp16 = const()[name = tensor("norm18_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(303741376)))]; + tensor var_7746_to_fp16 = const()[name = tensor("op_7746_to_fp16"), val = tensor(0x1.5p-17)]; + tensor x_37_cast_fp16 = layer_norm(axes = x_37_axes_0, beta = norm18_1_bias_to_fp16, epsilon = var_7746_to_fp16, gamma = norm18_1_weight_to_fp16, x = input_183_cast_fp16)[name = tensor("x_37_cast_fp16")]; + tensor linear_73_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = attn18_in_proj_weight_cast_fp16, x = x_37_cast_fp16)[name = tensor("linear_73_cast_fp16")]; + tensor var_7782 = const()[name = tensor("op_7782"), val = tensor([1, 1, 3, 16, 64])]; + tensor qkv_37_cast_fp16 = reshape(shape = var_7782, x = linear_73_cast_fp16)[name = tensor("qkv_37_cast_fp16")]; + tensor q_109_begin_0 = const()[name = tensor("q_109_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor q_109_end_0 = const()[name = tensor("q_109_end_0"), val = tensor([1, 1, 1, 16, 64])]; + tensor q_109_end_mask_0 = const()[name = tensor("q_109_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor q_109_squeeze_mask_0 = const()[name = tensor("q_109_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor q_109_cast_fp16 = slice_by_index(begin = q_109_begin_0, end = q_109_end_0, end_mask = q_109_end_mask_0, squeeze_mask = q_109_squeeze_mask_0, x = qkv_37_cast_fp16)[name = tensor("q_109_cast_fp16")]; + tensor k_73_begin_0 = const()[name = tensor("k_73_begin_0"), val = tensor([0, 0, 1, 0, 0])]; + tensor k_73_end_0 = const()[name = tensor("k_73_end_0"), val = tensor([1, 1, 2, 16, 64])]; + tensor k_73_end_mask_0 = const()[name = tensor("k_73_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor k_73_squeeze_mask_0 = const()[name = tensor("k_73_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor k_73_cast_fp16 = slice_by_index(begin = k_73_begin_0, end = k_73_end_0, end_mask = k_73_end_mask_0, squeeze_mask = k_73_squeeze_mask_0, x = qkv_37_cast_fp16)[name = tensor("k_73_cast_fp16")]; + tensor v_37_begin_0 = const()[name = tensor("v_37_begin_0"), val = tensor([0, 0, 2, 0, 0])]; + tensor v_37_end_0 = const()[name = tensor("v_37_end_0"), val = tensor([1, 1, 3, 16, 64])]; + tensor v_37_end_mask_0 = const()[name = tensor("v_37_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor v_37_squeeze_mask_0 = const()[name = tensor("v_37_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor v_37_cast_fp16 = slice_by_index(begin = v_37_begin_0, end = v_37_end_0, end_mask = v_37_end_mask_0, squeeze_mask = v_37_squeeze_mask_0, x = qkv_37_cast_fp16)[name = tensor("v_37_cast_fp16")]; + tensor var_7886 = const()[name = tensor("op_7886"), val = tensor([1, 1, 1, 1])]; + tensor position18_to_fp16_dtype_0 = const()[name = tensor("position18_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor position18_to_fp16 = cast(dtype = position18_to_fp16_dtype_0, x = position18)[name = tensor("cast_463")]; + tensor ts_113_cast_fp16 = reshape(shape = var_7886, x = position18_to_fp16)[name = tensor("ts_113_cast_fp16")]; + tensor var_7890 = const()[name = tensor("op_7890"), val = tensor([1, 1, 16, 32, 2])]; + tensor q_complex_37_cast_fp16 = reshape(shape = var_7890, x = q_109_cast_fp16)[name = tensor("q_complex_37_cast_fp16")]; + tensor var_7894 = const()[name = tensor("op_7894"), val = tensor([1, 1, 16, 32, 2])]; + tensor k_complex_37_cast_fp16 = reshape(shape = var_7894, x = k_73_cast_fp16)[name = tensor("k_complex_37_cast_fp16")]; + tensor var_7898_begin_0 = const()[name = tensor("op_7898_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_7898_end_0 = const()[name = tensor("op_7898_end_0"), val = tensor([1, 1, 16, 32, 1])]; + tensor var_7898_end_mask_0 = const()[name = tensor("op_7898_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_7898_squeeze_mask_0 = const()[name = tensor("op_7898_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_7898_cast_fp16 = slice_by_index(begin = var_7898_begin_0, end = var_7898_end_0, end_mask = var_7898_end_mask_0, squeeze_mask = var_7898_squeeze_mask_0, x = q_complex_37_cast_fp16)[name = tensor("op_7898_cast_fp16")]; + tensor var_7906_begin_0 = const()[name = tensor("op_7906_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_7906_end_0 = const()[name = tensor("op_7906_end_0"), val = tensor([1, 1, 16, 32, 2])]; + tensor var_7906_end_mask_0 = const()[name = tensor("op_7906_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_7906_squeeze_mask_0 = const()[name = tensor("op_7906_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_7906_cast_fp16 = slice_by_index(begin = var_7906_begin_0, end = var_7906_end_0, end_mask = var_7906_end_mask_0, squeeze_mask = var_7906_squeeze_mask_0, x = q_complex_37_cast_fp16)[name = tensor("op_7906_cast_fp16")]; + tensor var_7914_begin_0 = const()[name = tensor("op_7914_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_7914_end_0 = const()[name = tensor("op_7914_end_0"), val = tensor([1, 1, 16, 32, 1])]; + tensor var_7914_end_mask_0 = const()[name = tensor("op_7914_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_7914_squeeze_mask_0 = const()[name = tensor("op_7914_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_7914_cast_fp16 = slice_by_index(begin = var_7914_begin_0, end = var_7914_end_0, end_mask = var_7914_end_mask_0, squeeze_mask = var_7914_squeeze_mask_0, x = k_complex_37_cast_fp16)[name = tensor("op_7914_cast_fp16")]; + tensor var_7922_begin_0 = const()[name = tensor("op_7922_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_7922_end_0 = const()[name = tensor("op_7922_end_0"), val = tensor([1, 1, 16, 32, 2])]; + tensor var_7922_end_mask_0 = const()[name = tensor("op_7922_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_7922_squeeze_mask_0 = const()[name = tensor("op_7922_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_7922_cast_fp16 = slice_by_index(begin = var_7922_begin_0, end = var_7922_end_0, end_mask = var_7922_end_mask_0, squeeze_mask = var_7922_squeeze_mask_0, x = k_complex_37_cast_fp16)[name = tensor("op_7922_cast_fp16")]; + tensor freqs_37_to_fp16 = const()[name = tensor("freqs_37_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(303743488)))]; + tensor var_7928_cast_fp16 = mul(x = freqs_37_to_fp16, y = ts_113_cast_fp16)[name = tensor("op_7928_cast_fp16")]; + tensor rotr_37_cast_fp16 = cos(x = var_7928_cast_fp16)[name = tensor("rotr_37_cast_fp16")]; + tensor roti_37_cast_fp16 = sin(x = var_7928_cast_fp16)[name = tensor("roti_37_cast_fp16")]; + tensor var_7932_cast_fp16 = mul(x = var_7898_cast_fp16, y = rotr_37_cast_fp16)[name = tensor("op_7932_cast_fp16")]; + tensor var_7933_cast_fp16 = mul(x = var_7906_cast_fp16, y = roti_37_cast_fp16)[name = tensor("op_7933_cast_fp16")]; + tensor qor_73_cast_fp16 = sub(x = var_7932_cast_fp16, y = var_7933_cast_fp16)[name = tensor("qor_73_cast_fp16")]; + tensor var_7936_cast_fp16 = mul(x = var_7898_cast_fp16, y = roti_37_cast_fp16)[name = tensor("op_7936_cast_fp16")]; + tensor var_7937_cast_fp16 = mul(x = var_7906_cast_fp16, y = rotr_37_cast_fp16)[name = tensor("op_7937_cast_fp16")]; + tensor qoi_73_cast_fp16 = add(x = var_7936_cast_fp16, y = var_7937_cast_fp16)[name = tensor("qoi_73_cast_fp16")]; + tensor var_7940_cast_fp16 = mul(x = var_7914_cast_fp16, y = rotr_37_cast_fp16)[name = tensor("op_7940_cast_fp16")]; + tensor var_7941_cast_fp16 = mul(x = var_7922_cast_fp16, y = roti_37_cast_fp16)[name = tensor("op_7941_cast_fp16")]; + tensor kor_73_cast_fp16 = sub(x = var_7940_cast_fp16, y = var_7941_cast_fp16)[name = tensor("kor_73_cast_fp16")]; + tensor var_7944_cast_fp16 = mul(x = var_7914_cast_fp16, y = roti_37_cast_fp16)[name = tensor("op_7944_cast_fp16")]; + tensor var_7945_cast_fp16 = mul(x = var_7922_cast_fp16, y = rotr_37_cast_fp16)[name = tensor("op_7945_cast_fp16")]; + tensor koi_73_cast_fp16 = add(x = var_7944_cast_fp16, y = var_7945_cast_fp16)[name = tensor("koi_73_cast_fp16")]; + tensor qo_37_axis_0 = const()[name = tensor("qo_37_axis_0"), val = tensor(-1)]; + tensor qo_37_cast_fp16 = stack(axis = qo_37_axis_0, values = (qor_73_cast_fp16, qoi_73_cast_fp16))[name = tensor("qo_37_cast_fp16")]; + tensor ko_37_axis_0 = const()[name = tensor("ko_37_axis_0"), val = tensor(-1)]; + tensor ko_37_cast_fp16 = stack(axis = ko_37_axis_0, values = (kor_73_cast_fp16, koi_73_cast_fp16))[name = tensor("ko_37_cast_fp16")]; + tensor var_7974 = const()[name = tensor("op_7974"), val = tensor([1, 1, 16, 64])]; + tensor q_111_cast_fp16 = reshape(shape = var_7974, x = qo_37_cast_fp16)[name = tensor("q_111_cast_fp16")]; + tensor var_7976 = const()[name = tensor("op_7976"), val = tensor([1, 1, 16, 64])]; + tensor k_75_cast_fp16 = reshape(shape = var_7976, x = ko_37_cast_fp16)[name = tensor("k_75_cast_fp16")]; + tensor _inversed_7998_y_0_to_fp16 = const()[name = tensor("_inversed_7998_y_0_to_fp16"), val = tensor(0x1p-9)]; + tensor _inversed_7998_cast_fp16 = mul(x = ts_113_cast_fp16, y = _inversed_7998_y_0_to_fp16)[name = tensor("_inversed_7998_cast_fp16")]; + tensor var_7999_cast_fp16 = floor(x = _inversed_7998_cast_fp16)[name = tensor("op_7999_cast_fp16")]; + tensor var_8000_to_fp16 = const()[name = tensor("op_8000_to_fp16"), val = tensor(0x1p+9)]; + tensor var_8001_cast_fp16 = mul(x = var_7999_cast_fp16, y = var_8000_to_fp16)[name = tensor("op_8001_cast_fp16")]; + tensor write_indices_float_75_cast_fp16 = sub(x = ts_113_cast_fp16, y = var_8001_cast_fp16)[name = tensor("write_indices_float_75_cast_fp16")]; + tensor var_8008_dtype_0 = const()[name = tensor("op_8008_dtype_0"), val = tensor("int32")]; + tensor write_indices_37_reps_0 = const()[name = tensor("write_indices_37_reps_0"), val = tensor([1, 1, 16, 64])]; + tensor write_indices_float_75_cast_fp16_to_int32 = cast(dtype = var_8008_dtype_0, x = write_indices_float_75_cast_fp16)[name = tensor("cast_462")]; + tensor write_indices_37 = tile(reps = write_indices_37_reps_0, x = write_indices_float_75_cast_fp16_to_int32)[name = tensor("write_indices_37")]; + tensor var_8016_begin_0 = const()[name = tensor("op_8016_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_8016_end_0 = const()[name = tensor("op_8016_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor var_8016_end_mask_0 = const()[name = tensor("op_8016_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_8016_squeeze_mask_0 = const()[name = tensor("op_8016_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor cache18_to_fp16_dtype_0 = const()[name = tensor("cache18_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor cache18_to_fp16 = cast(dtype = cache18_to_fp16_dtype_0, x = cache18)[name = tensor("cast_461")]; + tensor var_8016_cast_fp16 = slice_by_index(begin = var_8016_begin_0, end = var_8016_end_0, end_mask = var_8016_end_mask_0, squeeze_mask = var_8016_squeeze_mask_0, x = cache18_to_fp16)[name = tensor("op_8016_cast_fp16")]; + tensor var_8018_axis_0 = const()[name = tensor("op_8018_axis_0"), val = tensor(1)]; + tensor var_8018_mode_0 = const()[name = tensor("op_8018_mode_0"), val = tensor("update")]; + tensor var_8018_validate_indices_0 = const()[name = tensor("op_8018_validate_indices_0"), val = tensor(false)]; + tensor var_8018_cast_fp16 = scatter_along_axis(axis = var_8018_axis_0, data = var_8016_cast_fp16, indices = write_indices_37, mode = var_8018_mode_0, updates = k_75_cast_fp16, validate_indices = var_8018_validate_indices_0)[name = tensor("op_8018_cast_fp16")]; + tensor concat_128 = const()[name = tensor("concat_128"), val = tensor([0, 0, 0, 0, 0])]; + tensor concat_129 = const()[name = tensor("concat_129"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_37_internal_tensor_assign_1_stride_0 = const()[name = tensor("new_cache_37_internal_tensor_assign_1_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_37_internal_tensor_assign_1_begin_mask_0 = const()[name = tensor("new_cache_37_internal_tensor_assign_1_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_37_internal_tensor_assign_1_end_mask_0 = const()[name = tensor("new_cache_37_internal_tensor_assign_1_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_37_internal_tensor_assign_1_squeeze_mask_0 = const()[name = tensor("new_cache_37_internal_tensor_assign_1_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_84 = const()[name = tensor("shape_84"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_36 = const()[name = tensor("reduce_prod_36"), val = tensor(1048576)]; + tensor range_1d_36_start_0 = const()[name = tensor("range_1d_36_start_0"), val = tensor(0)]; + tensor range_1d_36_step_0 = const()[name = tensor("range_1d_36_step_0"), val = tensor(1)]; + tensor range_1d_36 = range_1d(end = reduce_prod_36, start = range_1d_36_start_0, step = range_1d_36_step_0)[name = tensor("range_1d_36")]; + tensor reshape_180 = reshape(shape = shape_84, x = range_1d_36)[name = tensor("reshape_180")]; + tensor slice_by_index_36 = slice_by_index(begin = concat_128, begin_mask = new_cache_37_internal_tensor_assign_1_begin_mask_0, end = concat_129, end_mask = new_cache_37_internal_tensor_assign_1_end_mask_0, squeeze_mask = new_cache_37_internal_tensor_assign_1_squeeze_mask_0, stride = new_cache_37_internal_tensor_assign_1_stride_0, x = reshape_180)[name = tensor("slice_by_index_36")]; + tensor reshape_181_shape_0 = const()[name = tensor("reshape_181_shape_0"), val = tensor([-1])]; + tensor reshape_181 = reshape(shape = reshape_181_shape_0, x = slice_by_index_36)[name = tensor("reshape_181")]; + tensor reshape_182_shape_0 = const()[name = tensor("reshape_182_shape_0"), val = tensor([-1])]; + tensor reshape_182_cast_fp16 = reshape(shape = reshape_182_shape_0, x = var_8018_cast_fp16)[name = tensor("reshape_182_cast_fp16")]; + tensor reshape_183_shape_0 = const()[name = tensor("reshape_183_shape_0"), val = tensor([-1])]; + tensor reshape_183_cast_fp16 = reshape(shape = reshape_183_shape_0, x = cache18_to_fp16)[name = tensor("reshape_183_cast_fp16")]; + tensor scatter_36_mode_0 = const()[name = tensor("scatter_36_mode_0"), val = tensor("update")]; + tensor scatter_36_axis_0 = const()[name = tensor("scatter_36_axis_0"), val = tensor(0)]; + tensor scatter_36_validate_indices_0 = const()[name = tensor("scatter_36_validate_indices_0"), val = tensor(false)]; + tensor scatter_36_cast_fp16 = scatter(axis = scatter_36_axis_0, data = reshape_183_cast_fp16, indices = reshape_181, mode = scatter_36_mode_0, updates = reshape_182_cast_fp16, validate_indices = scatter_36_validate_indices_0)[name = tensor("scatter_36_cast_fp16")]; + tensor reshape_184_cast_fp16 = reshape(shape = shape_84, x = scatter_36_cast_fp16)[name = tensor("reshape_184_cast_fp16")]; + tensor var_8026_begin_0 = const()[name = tensor("op_8026_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor var_8026_end_0 = const()[name = tensor("op_8026_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor var_8026_end_mask_0 = const()[name = tensor("op_8026_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_8026_squeeze_mask_0 = const()[name = tensor("op_8026_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor var_8026_cast_fp16 = slice_by_index(begin = var_8026_begin_0, end = var_8026_end_0, end_mask = var_8026_end_mask_0, squeeze_mask = var_8026_squeeze_mask_0, x = reshape_184_cast_fp16)[name = tensor("op_8026_cast_fp16")]; + tensor var_8028_axis_0 = const()[name = tensor("op_8028_axis_0"), val = tensor(1)]; + tensor var_8028_mode_0 = const()[name = tensor("op_8028_mode_0"), val = tensor("update")]; + tensor var_8028_validate_indices_0 = const()[name = tensor("op_8028_validate_indices_0"), val = tensor(false)]; + tensor var_8028_cast_fp16 = scatter_along_axis(axis = var_8028_axis_0, data = var_8026_cast_fp16, indices = write_indices_37, mode = var_8028_mode_0, updates = v_37_cast_fp16, validate_indices = var_8028_validate_indices_0)[name = tensor("op_8028_cast_fp16")]; + tensor concat_130 = const()[name = tensor("concat_130"), val = tensor([1, 0, 0, 0, 0])]; + tensor concat_131 = const()[name = tensor("concat_131"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_37_internal_tensor_assign_2_stride_0 = const()[name = tensor("new_cache_37_internal_tensor_assign_2_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_37_internal_tensor_assign_2_begin_mask_0 = const()[name = tensor("new_cache_37_internal_tensor_assign_2_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_37_internal_tensor_assign_2_end_mask_0 = const()[name = tensor("new_cache_37_internal_tensor_assign_2_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_37_internal_tensor_assign_2_squeeze_mask_0 = const()[name = tensor("new_cache_37_internal_tensor_assign_2_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_85 = const()[name = tensor("shape_85"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_37 = const()[name = tensor("reduce_prod_37"), val = tensor(1048576)]; + tensor range_1d_37_start_0 = const()[name = tensor("range_1d_37_start_0"), val = tensor(0)]; + tensor range_1d_37_step_0 = const()[name = tensor("range_1d_37_step_0"), val = tensor(1)]; + tensor range_1d_37 = range_1d(end = reduce_prod_37, start = range_1d_37_start_0, step = range_1d_37_step_0)[name = tensor("range_1d_37")]; + tensor reshape_185 = reshape(shape = shape_85, x = range_1d_37)[name = tensor("reshape_185")]; + tensor slice_by_index_37 = slice_by_index(begin = concat_130, begin_mask = new_cache_37_internal_tensor_assign_2_begin_mask_0, end = concat_131, end_mask = new_cache_37_internal_tensor_assign_2_end_mask_0, squeeze_mask = new_cache_37_internal_tensor_assign_2_squeeze_mask_0, stride = new_cache_37_internal_tensor_assign_2_stride_0, x = reshape_185)[name = tensor("slice_by_index_37")]; + tensor reshape_186_shape_0 = const()[name = tensor("reshape_186_shape_0"), val = tensor([-1])]; + tensor reshape_186 = reshape(shape = reshape_186_shape_0, x = slice_by_index_37)[name = tensor("reshape_186")]; + tensor reshape_187_shape_0 = const()[name = tensor("reshape_187_shape_0"), val = tensor([-1])]; + tensor reshape_187_cast_fp16 = reshape(shape = reshape_187_shape_0, x = var_8028_cast_fp16)[name = tensor("reshape_187_cast_fp16")]; + tensor reshape_188_shape_0 = const()[name = tensor("reshape_188_shape_0"), val = tensor([-1])]; + tensor reshape_188_cast_fp16 = reshape(shape = reshape_188_shape_0, x = reshape_184_cast_fp16)[name = tensor("reshape_188_cast_fp16")]; + tensor scatter_37_mode_0 = const()[name = tensor("scatter_37_mode_0"), val = tensor("update")]; + tensor scatter_37_axis_0 = const()[name = tensor("scatter_37_axis_0"), val = tensor(0)]; + tensor scatter_37_validate_indices_0 = const()[name = tensor("scatter_37_validate_indices_0"), val = tensor(false)]; + tensor scatter_37_cast_fp16 = scatter(axis = scatter_37_axis_0, data = reshape_188_cast_fp16, indices = reshape_186, mode = scatter_37_mode_0, updates = reshape_187_cast_fp16, validate_indices = scatter_37_validate_indices_0)[name = tensor("scatter_37_cast_fp16")]; + tensor reshape_189_cast_fp16 = reshape(shape = shape_85, x = scatter_37_cast_fp16)[name = tensor("reshape_189_cast_fp16")]; + tensor reshape_189_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("reshape_189_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor keys_109_begin_0 = const()[name = tensor("keys_109_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor keys_109_end_0 = const()[name = tensor("keys_109_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor keys_109_end_mask_0 = const()[name = tensor("keys_109_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor keys_109_squeeze_mask_0 = const()[name = tensor("keys_109_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor keys_109_cast_fp16 = slice_by_index(begin = keys_109_begin_0, end = keys_109_end_0, end_mask = keys_109_end_mask_0, squeeze_mask = keys_109_squeeze_mask_0, x = reshape_189_cast_fp16)[name = tensor("keys_109_cast_fp16")]; + tensor values_109_begin_0 = const()[name = tensor("values_109_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor values_109_end_0 = const()[name = tensor("values_109_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor values_109_end_mask_0 = const()[name = tensor("values_109_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor values_109_squeeze_mask_0 = const()[name = tensor("values_109_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor values_109_cast_fp16 = slice_by_index(begin = values_109_begin_0, end = values_109_end_0, end_mask = values_109_end_mask_0, squeeze_mask = values_109_squeeze_mask_0, x = reshape_189_cast_fp16)[name = tensor("values_109_cast_fp16")]; + tensor var_8040_cast_fp16 = not_equal(x = keys_109_cast_fp16, y = keys_109_cast_fp16)[name = tensor("op_8040_cast_fp16")]; + tensor keys_111_cast_fp16 = select(a = var_504_to_fp16, b = keys_109_cast_fp16, cond = var_8040_cast_fp16)[name = tensor("keys_111_cast_fp16")]; + tensor var_8048_cast_fp16 = not_equal(x = values_109_cast_fp16, y = values_109_cast_fp16)[name = tensor("op_8048_cast_fp16")]; + tensor values_111_cast_fp16 = select(a = var_504_to_fp16, b = values_109_cast_fp16, cond = var_8048_cast_fp16)[name = tensor("values_111_cast_fp16")]; + tensor var_8072 = const()[name = tensor("op_8072"), val = tensor([0, 2, 1, 3])]; + tensor var_8085 = const()[name = tensor("op_8085"), val = tensor([1, 1, 1])]; + tensor var_8086_cast_fp16 = reshape(shape = var_8085, x = position18_to_fp16)[name = tensor("op_8086_cast_fp16")]; + tensor var_8103_to_fp16 = const()[name = tensor("op_8103_to_fp16"), val = tensor(0x1p+0)]; + tensor valid_len_37_cast_fp16 = add(x = var_8086_cast_fp16, y = var_8103_to_fp16)[name = tensor("valid_len_37_cast_fp16")]; + tensor valid_mask_37_cast_fp16 = less(x = k_positions_1_promoted_to_fp16, y = valid_len_37_cast_fp16)[name = tensor("valid_mask_37_cast_fp16")]; + tensor causal_mask_37_cast_fp16 = less_equal(x = k_positions_1_promoted_to_fp16, y = var_8086_cast_fp16)[name = tensor("causal_mask_37_cast_fp16")]; + tensor attn_mask_73 = logical_and(x = valid_mask_37_cast_fp16, y = causal_mask_37_cast_fp16)[name = tensor("attn_mask_73")]; + tensor attn_mask_75_axes_0 = const()[name = tensor("attn_mask_75_axes_0"), val = tensor([1])]; + tensor attn_mask_75 = expand_dims(axes = attn_mask_75_axes_0, x = attn_mask_73)[name = tensor("attn_mask_75")]; + tensor var_8121_transpose_x_0 = const()[name = tensor("op_8121_transpose_x_0"), val = tensor(false)]; + tensor var_8121_transpose_y_0 = const()[name = tensor("op_8121_transpose_y_0"), val = tensor(false)]; + tensor transpose_108_perm_0 = const()[name = tensor("transpose_108_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_109_perm_0 = const()[name = tensor("transpose_109_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_109 = transpose(perm = transpose_109_perm_0, x = keys_111_cast_fp16)[name = tensor("transpose_141")]; + tensor transpose_108 = transpose(perm = transpose_108_perm_0, x = q_111_cast_fp16)[name = tensor("transpose_142")]; + tensor var_8121_cast_fp16 = matmul(transpose_x = var_8121_transpose_x_0, transpose_y = var_8121_transpose_y_0, x = transpose_108, y = transpose_109)[name = tensor("op_8121_cast_fp16")]; + tensor var_8115_to_fp16 = const()[name = tensor("op_8115_to_fp16"), val = tensor([0x1p-3])]; + tensor attn_weights_109_cast_fp16 = mul(x = var_8121_cast_fp16, y = var_8115_to_fp16)[name = tensor("attn_weights_109_cast_fp16")]; + tensor var_8123 = logical_not(x = attn_mask_75)[name = tensor("op_8123")]; + tensor var_8124_to_fp16 = const()[name = tensor("op_8124_to_fp16"), val = tensor(-inf)]; + tensor attn_weights_111_cast_fp16 = select(a = var_8124_to_fp16, b = attn_weights_109_cast_fp16, cond = var_8123)[name = tensor("attn_weights_111_cast_fp16")]; + tensor var_8126 = const()[name = tensor("op_8126"), val = tensor(-1)]; + tensor attn_weights_113_cast_fp16 = softmax(axis = var_8126, x = attn_weights_111_cast_fp16)[name = tensor("attn_weights_113_cast_fp16")]; + tensor attn_output_37_transpose_x_0 = const()[name = tensor("attn_output_37_transpose_x_0"), val = tensor(false)]; + tensor attn_output_37_transpose_y_0 = const()[name = tensor("attn_output_37_transpose_y_0"), val = tensor(false)]; + tensor values_113_cast_fp16 = transpose(perm = var_8072, x = values_111_cast_fp16)[name = tensor("transpose_143")]; + tensor attn_output_37_cast_fp16 = matmul(transpose_x = attn_output_37_transpose_x_0, transpose_y = attn_output_37_transpose_y_0, x = attn_weights_113_cast_fp16, y = values_113_cast_fp16)[name = tensor("attn_output_37_cast_fp16")]; + tensor var_8134 = const()[name = tensor("op_8134"), val = tensor([0, 2, 1, 3])]; + tensor var_8137 = const()[name = tensor("op_8137"), val = tensor([1, 1, 1024])]; + tensor var_8135_cast_fp16 = transpose(perm = var_8134, x = attn_output_37_cast_fp16)[name = tensor("transpose_140")]; + tensor input_185_cast_fp16 = reshape(shape = var_8137, x = var_8135_cast_fp16)[name = tensor("input_185_cast_fp16")]; + tensor linear_74_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = attn18_out_proj_weight_cast_fp16, x = input_185_cast_fp16)[name = tensor("linear_74_cast_fp16")]; + tensor var_8143_to_fp16 = const()[name = tensor("op_8143_to_fp16"), val = tensor(0x1p+0)]; + tensor var_8144_cast_fp16 = add(x = position18_to_fp16, y = var_8143_to_fp16)[name = tensor("op_8144_cast_fp16")]; + tensor var_8144_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_8144_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor input_187_cast_fp16 = add(x = input_183_cast_fp16, y = linear_74_cast_fp16)[name = tensor("input_187_cast_fp16")]; + tensor input_189_axes_0 = const()[name = tensor("input_189_axes_0"), val = tensor([-1])]; + tensor norm18_2_weight_to_fp16 = const()[name = tensor("norm18_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(303743616)))]; + tensor norm18_2_bias_to_fp16 = const()[name = tensor("norm18_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(303745728)))]; + tensor var_8148_to_fp16 = const()[name = tensor("op_8148_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_189_cast_fp16 = layer_norm(axes = input_189_axes_0, beta = norm18_2_bias_to_fp16, epsilon = var_8148_to_fp16, gamma = norm18_2_weight_to_fp16, x = input_187_cast_fp16)[name = tensor("input_189_cast_fp16")]; + tensor linear_75_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = linear18_1_weight_cast_fp16, x = input_189_cast_fp16)[name = tensor("linear_75_cast_fp16")]; + tensor input_191_mode_0 = const()[name = tensor("input_191_mode_0"), val = tensor("EXACT")]; + tensor input_191_cast_fp16 = gelu(mode = input_191_mode_0, x = linear_75_cast_fp16)[name = tensor("input_191_cast_fp16")]; + tensor linear_76_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = linear18_2_weight_cast_fp16, x = input_191_cast_fp16)[name = tensor("linear_76_cast_fp16")]; + tensor input_193_cast_fp16 = add(x = input_187_cast_fp16, y = linear_76_cast_fp16)[name = tensor("input_193_cast_fp16")]; + tensor x_39_axes_0 = const()[name = tensor("x_39_axes_0"), val = tensor([-1])]; + tensor norm19_1_weight_to_fp16 = const()[name = tensor("norm19_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(303747840)))]; + tensor norm19_1_bias_to_fp16 = const()[name = tensor("norm19_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(303749952)))]; + tensor var_8165_to_fp16 = const()[name = tensor("op_8165_to_fp16"), val = tensor(0x1.5p-17)]; + tensor x_39_cast_fp16 = layer_norm(axes = x_39_axes_0, beta = norm19_1_bias_to_fp16, epsilon = var_8165_to_fp16, gamma = norm19_1_weight_to_fp16, x = input_193_cast_fp16)[name = tensor("x_39_cast_fp16")]; + tensor linear_77_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = attn19_in_proj_weight_cast_fp16, x = x_39_cast_fp16)[name = tensor("linear_77_cast_fp16")]; + tensor var_8201 = const()[name = tensor("op_8201"), val = tensor([1, 1, 3, 16, 64])]; + tensor qkv_39_cast_fp16 = reshape(shape = var_8201, x = linear_77_cast_fp16)[name = tensor("qkv_39_cast_fp16")]; + tensor q_115_begin_0 = const()[name = tensor("q_115_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor q_115_end_0 = const()[name = tensor("q_115_end_0"), val = tensor([1, 1, 1, 16, 64])]; + tensor q_115_end_mask_0 = const()[name = tensor("q_115_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor q_115_squeeze_mask_0 = const()[name = tensor("q_115_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor q_115_cast_fp16 = slice_by_index(begin = q_115_begin_0, end = q_115_end_0, end_mask = q_115_end_mask_0, squeeze_mask = q_115_squeeze_mask_0, x = qkv_39_cast_fp16)[name = tensor("q_115_cast_fp16")]; + tensor k_77_begin_0 = const()[name = tensor("k_77_begin_0"), val = tensor([0, 0, 1, 0, 0])]; + tensor k_77_end_0 = const()[name = tensor("k_77_end_0"), val = tensor([1, 1, 2, 16, 64])]; + tensor k_77_end_mask_0 = const()[name = tensor("k_77_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor k_77_squeeze_mask_0 = const()[name = tensor("k_77_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor k_77_cast_fp16 = slice_by_index(begin = k_77_begin_0, end = k_77_end_0, end_mask = k_77_end_mask_0, squeeze_mask = k_77_squeeze_mask_0, x = qkv_39_cast_fp16)[name = tensor("k_77_cast_fp16")]; + tensor v_39_begin_0 = const()[name = tensor("v_39_begin_0"), val = tensor([0, 0, 2, 0, 0])]; + tensor v_39_end_0 = const()[name = tensor("v_39_end_0"), val = tensor([1, 1, 3, 16, 64])]; + tensor v_39_end_mask_0 = const()[name = tensor("v_39_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor v_39_squeeze_mask_0 = const()[name = tensor("v_39_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor v_39_cast_fp16 = slice_by_index(begin = v_39_begin_0, end = v_39_end_0, end_mask = v_39_end_mask_0, squeeze_mask = v_39_squeeze_mask_0, x = qkv_39_cast_fp16)[name = tensor("v_39_cast_fp16")]; + tensor var_8305 = const()[name = tensor("op_8305"), val = tensor([1, 1, 1, 1])]; + tensor position19_to_fp16_dtype_0 = const()[name = tensor("position19_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor position19_to_fp16 = cast(dtype = position19_to_fp16_dtype_0, x = position19)[name = tensor("cast_458")]; + tensor ts_119_cast_fp16 = reshape(shape = var_8305, x = position19_to_fp16)[name = tensor("ts_119_cast_fp16")]; + tensor var_8309 = const()[name = tensor("op_8309"), val = tensor([1, 1, 16, 32, 2])]; + tensor q_complex_39_cast_fp16 = reshape(shape = var_8309, x = q_115_cast_fp16)[name = tensor("q_complex_39_cast_fp16")]; + tensor var_8313 = const()[name = tensor("op_8313"), val = tensor([1, 1, 16, 32, 2])]; + tensor k_complex_39_cast_fp16 = reshape(shape = var_8313, x = k_77_cast_fp16)[name = tensor("k_complex_39_cast_fp16")]; + tensor var_8317_begin_0 = const()[name = tensor("op_8317_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_8317_end_0 = const()[name = tensor("op_8317_end_0"), val = tensor([1, 1, 16, 32, 1])]; + tensor var_8317_end_mask_0 = const()[name = tensor("op_8317_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_8317_squeeze_mask_0 = const()[name = tensor("op_8317_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_8317_cast_fp16 = slice_by_index(begin = var_8317_begin_0, end = var_8317_end_0, end_mask = var_8317_end_mask_0, squeeze_mask = var_8317_squeeze_mask_0, x = q_complex_39_cast_fp16)[name = tensor("op_8317_cast_fp16")]; + tensor var_8325_begin_0 = const()[name = tensor("op_8325_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_8325_end_0 = const()[name = tensor("op_8325_end_0"), val = tensor([1, 1, 16, 32, 2])]; + tensor var_8325_end_mask_0 = const()[name = tensor("op_8325_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_8325_squeeze_mask_0 = const()[name = tensor("op_8325_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_8325_cast_fp16 = slice_by_index(begin = var_8325_begin_0, end = var_8325_end_0, end_mask = var_8325_end_mask_0, squeeze_mask = var_8325_squeeze_mask_0, x = q_complex_39_cast_fp16)[name = tensor("op_8325_cast_fp16")]; + tensor var_8333_begin_0 = const()[name = tensor("op_8333_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_8333_end_0 = const()[name = tensor("op_8333_end_0"), val = tensor([1, 1, 16, 32, 1])]; + tensor var_8333_end_mask_0 = const()[name = tensor("op_8333_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_8333_squeeze_mask_0 = const()[name = tensor("op_8333_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_8333_cast_fp16 = slice_by_index(begin = var_8333_begin_0, end = var_8333_end_0, end_mask = var_8333_end_mask_0, squeeze_mask = var_8333_squeeze_mask_0, x = k_complex_39_cast_fp16)[name = tensor("op_8333_cast_fp16")]; + tensor var_8341_begin_0 = const()[name = tensor("op_8341_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_8341_end_0 = const()[name = tensor("op_8341_end_0"), val = tensor([1, 1, 16, 32, 2])]; + tensor var_8341_end_mask_0 = const()[name = tensor("op_8341_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_8341_squeeze_mask_0 = const()[name = tensor("op_8341_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_8341_cast_fp16 = slice_by_index(begin = var_8341_begin_0, end = var_8341_end_0, end_mask = var_8341_end_mask_0, squeeze_mask = var_8341_squeeze_mask_0, x = k_complex_39_cast_fp16)[name = tensor("op_8341_cast_fp16")]; + tensor freqs_39_to_fp16 = const()[name = tensor("freqs_39_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(303752064)))]; + tensor var_8347_cast_fp16 = mul(x = freqs_39_to_fp16, y = ts_119_cast_fp16)[name = tensor("op_8347_cast_fp16")]; + tensor rotr_39_cast_fp16 = cos(x = var_8347_cast_fp16)[name = tensor("rotr_39_cast_fp16")]; + tensor roti_39_cast_fp16 = sin(x = var_8347_cast_fp16)[name = tensor("roti_39_cast_fp16")]; + tensor var_8351_cast_fp16 = mul(x = var_8317_cast_fp16, y = rotr_39_cast_fp16)[name = tensor("op_8351_cast_fp16")]; + tensor var_8352_cast_fp16 = mul(x = var_8325_cast_fp16, y = roti_39_cast_fp16)[name = tensor("op_8352_cast_fp16")]; + tensor qor_77_cast_fp16 = sub(x = var_8351_cast_fp16, y = var_8352_cast_fp16)[name = tensor("qor_77_cast_fp16")]; + tensor var_8355_cast_fp16 = mul(x = var_8317_cast_fp16, y = roti_39_cast_fp16)[name = tensor("op_8355_cast_fp16")]; + tensor var_8356_cast_fp16 = mul(x = var_8325_cast_fp16, y = rotr_39_cast_fp16)[name = tensor("op_8356_cast_fp16")]; + tensor qoi_77_cast_fp16 = add(x = var_8355_cast_fp16, y = var_8356_cast_fp16)[name = tensor("qoi_77_cast_fp16")]; + tensor var_8359_cast_fp16 = mul(x = var_8333_cast_fp16, y = rotr_39_cast_fp16)[name = tensor("op_8359_cast_fp16")]; + tensor var_8360_cast_fp16 = mul(x = var_8341_cast_fp16, y = roti_39_cast_fp16)[name = tensor("op_8360_cast_fp16")]; + tensor kor_77_cast_fp16 = sub(x = var_8359_cast_fp16, y = var_8360_cast_fp16)[name = tensor("kor_77_cast_fp16")]; + tensor var_8363_cast_fp16 = mul(x = var_8333_cast_fp16, y = roti_39_cast_fp16)[name = tensor("op_8363_cast_fp16")]; + tensor var_8364_cast_fp16 = mul(x = var_8341_cast_fp16, y = rotr_39_cast_fp16)[name = tensor("op_8364_cast_fp16")]; + tensor koi_77_cast_fp16 = add(x = var_8363_cast_fp16, y = var_8364_cast_fp16)[name = tensor("koi_77_cast_fp16")]; + tensor qo_39_axis_0 = const()[name = tensor("qo_39_axis_0"), val = tensor(-1)]; + tensor qo_39_cast_fp16 = stack(axis = qo_39_axis_0, values = (qor_77_cast_fp16, qoi_77_cast_fp16))[name = tensor("qo_39_cast_fp16")]; + tensor ko_39_axis_0 = const()[name = tensor("ko_39_axis_0"), val = tensor(-1)]; + tensor ko_39_cast_fp16 = stack(axis = ko_39_axis_0, values = (kor_77_cast_fp16, koi_77_cast_fp16))[name = tensor("ko_39_cast_fp16")]; + tensor var_8393 = const()[name = tensor("op_8393"), val = tensor([1, 1, 16, 64])]; + tensor q_117_cast_fp16 = reshape(shape = var_8393, x = qo_39_cast_fp16)[name = tensor("q_117_cast_fp16")]; + tensor var_8395 = const()[name = tensor("op_8395"), val = tensor([1, 1, 16, 64])]; + tensor k_79_cast_fp16 = reshape(shape = var_8395, x = ko_39_cast_fp16)[name = tensor("k_79_cast_fp16")]; + tensor _inversed_8417_y_0_to_fp16 = const()[name = tensor("_inversed_8417_y_0_to_fp16"), val = tensor(0x1p-9)]; + tensor _inversed_8417_cast_fp16 = mul(x = ts_119_cast_fp16, y = _inversed_8417_y_0_to_fp16)[name = tensor("_inversed_8417_cast_fp16")]; + tensor var_8418_cast_fp16 = floor(x = _inversed_8417_cast_fp16)[name = tensor("op_8418_cast_fp16")]; + tensor var_8419_to_fp16 = const()[name = tensor("op_8419_to_fp16"), val = tensor(0x1p+9)]; + tensor var_8420_cast_fp16 = mul(x = var_8418_cast_fp16, y = var_8419_to_fp16)[name = tensor("op_8420_cast_fp16")]; + tensor write_indices_float_79_cast_fp16 = sub(x = ts_119_cast_fp16, y = var_8420_cast_fp16)[name = tensor("write_indices_float_79_cast_fp16")]; + tensor var_8427_dtype_0 = const()[name = tensor("op_8427_dtype_0"), val = tensor("int32")]; + tensor write_indices_39_reps_0 = const()[name = tensor("write_indices_39_reps_0"), val = tensor([1, 1, 16, 64])]; + tensor write_indices_float_79_cast_fp16_to_int32 = cast(dtype = var_8427_dtype_0, x = write_indices_float_79_cast_fp16)[name = tensor("cast_457")]; + tensor write_indices_39 = tile(reps = write_indices_39_reps_0, x = write_indices_float_79_cast_fp16_to_int32)[name = tensor("write_indices_39")]; + tensor var_8435_begin_0 = const()[name = tensor("op_8435_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_8435_end_0 = const()[name = tensor("op_8435_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor var_8435_end_mask_0 = const()[name = tensor("op_8435_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_8435_squeeze_mask_0 = const()[name = tensor("op_8435_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor cache19_to_fp16_dtype_0 = const()[name = tensor("cache19_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor cache19_to_fp16 = cast(dtype = cache19_to_fp16_dtype_0, x = cache19)[name = tensor("cast_456")]; + tensor var_8435_cast_fp16 = slice_by_index(begin = var_8435_begin_0, end = var_8435_end_0, end_mask = var_8435_end_mask_0, squeeze_mask = var_8435_squeeze_mask_0, x = cache19_to_fp16)[name = tensor("op_8435_cast_fp16")]; + tensor var_8437_axis_0 = const()[name = tensor("op_8437_axis_0"), val = tensor(1)]; + tensor var_8437_mode_0 = const()[name = tensor("op_8437_mode_0"), val = tensor("update")]; + tensor var_8437_validate_indices_0 = const()[name = tensor("op_8437_validate_indices_0"), val = tensor(false)]; + tensor var_8437_cast_fp16 = scatter_along_axis(axis = var_8437_axis_0, data = var_8435_cast_fp16, indices = write_indices_39, mode = var_8437_mode_0, updates = k_79_cast_fp16, validate_indices = var_8437_validate_indices_0)[name = tensor("op_8437_cast_fp16")]; + tensor concat_135 = const()[name = tensor("concat_135"), val = tensor([0, 0, 0, 0, 0])]; + tensor concat_136 = const()[name = tensor("concat_136"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_39_internal_tensor_assign_1_stride_0 = const()[name = tensor("new_cache_39_internal_tensor_assign_1_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_39_internal_tensor_assign_1_begin_mask_0 = const()[name = tensor("new_cache_39_internal_tensor_assign_1_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_39_internal_tensor_assign_1_end_mask_0 = const()[name = tensor("new_cache_39_internal_tensor_assign_1_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_39_internal_tensor_assign_1_squeeze_mask_0 = const()[name = tensor("new_cache_39_internal_tensor_assign_1_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_86 = const()[name = tensor("shape_86"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_38 = const()[name = tensor("reduce_prod_38"), val = tensor(1048576)]; + tensor range_1d_38_start_0 = const()[name = tensor("range_1d_38_start_0"), val = tensor(0)]; + tensor range_1d_38_step_0 = const()[name = tensor("range_1d_38_step_0"), val = tensor(1)]; + tensor range_1d_38 = range_1d(end = reduce_prod_38, start = range_1d_38_start_0, step = range_1d_38_step_0)[name = tensor("range_1d_38")]; + tensor reshape_190 = reshape(shape = shape_86, x = range_1d_38)[name = tensor("reshape_190")]; + tensor slice_by_index_38 = slice_by_index(begin = concat_135, begin_mask = new_cache_39_internal_tensor_assign_1_begin_mask_0, end = concat_136, end_mask = new_cache_39_internal_tensor_assign_1_end_mask_0, squeeze_mask = new_cache_39_internal_tensor_assign_1_squeeze_mask_0, stride = new_cache_39_internal_tensor_assign_1_stride_0, x = reshape_190)[name = tensor("slice_by_index_38")]; + tensor reshape_191_shape_0 = const()[name = tensor("reshape_191_shape_0"), val = tensor([-1])]; + tensor reshape_191 = reshape(shape = reshape_191_shape_0, x = slice_by_index_38)[name = tensor("reshape_191")]; + tensor reshape_192_shape_0 = const()[name = tensor("reshape_192_shape_0"), val = tensor([-1])]; + tensor reshape_192_cast_fp16 = reshape(shape = reshape_192_shape_0, x = var_8437_cast_fp16)[name = tensor("reshape_192_cast_fp16")]; + tensor reshape_193_shape_0 = const()[name = tensor("reshape_193_shape_0"), val = tensor([-1])]; + tensor reshape_193_cast_fp16 = reshape(shape = reshape_193_shape_0, x = cache19_to_fp16)[name = tensor("reshape_193_cast_fp16")]; + tensor scatter_38_mode_0 = const()[name = tensor("scatter_38_mode_0"), val = tensor("update")]; + tensor scatter_38_axis_0 = const()[name = tensor("scatter_38_axis_0"), val = tensor(0)]; + tensor scatter_38_validate_indices_0 = const()[name = tensor("scatter_38_validate_indices_0"), val = tensor(false)]; + tensor scatter_38_cast_fp16 = scatter(axis = scatter_38_axis_0, data = reshape_193_cast_fp16, indices = reshape_191, mode = scatter_38_mode_0, updates = reshape_192_cast_fp16, validate_indices = scatter_38_validate_indices_0)[name = tensor("scatter_38_cast_fp16")]; + tensor reshape_194_cast_fp16 = reshape(shape = shape_86, x = scatter_38_cast_fp16)[name = tensor("reshape_194_cast_fp16")]; + tensor var_8445_begin_0 = const()[name = tensor("op_8445_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor var_8445_end_0 = const()[name = tensor("op_8445_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor var_8445_end_mask_0 = const()[name = tensor("op_8445_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_8445_squeeze_mask_0 = const()[name = tensor("op_8445_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor var_8445_cast_fp16 = slice_by_index(begin = var_8445_begin_0, end = var_8445_end_0, end_mask = var_8445_end_mask_0, squeeze_mask = var_8445_squeeze_mask_0, x = reshape_194_cast_fp16)[name = tensor("op_8445_cast_fp16")]; + tensor var_8447_axis_0 = const()[name = tensor("op_8447_axis_0"), val = tensor(1)]; + tensor var_8447_mode_0 = const()[name = tensor("op_8447_mode_0"), val = tensor("update")]; + tensor var_8447_validate_indices_0 = const()[name = tensor("op_8447_validate_indices_0"), val = tensor(false)]; + tensor var_8447_cast_fp16 = scatter_along_axis(axis = var_8447_axis_0, data = var_8445_cast_fp16, indices = write_indices_39, mode = var_8447_mode_0, updates = v_39_cast_fp16, validate_indices = var_8447_validate_indices_0)[name = tensor("op_8447_cast_fp16")]; + tensor concat_137 = const()[name = tensor("concat_137"), val = tensor([1, 0, 0, 0, 0])]; + tensor concat_138 = const()[name = tensor("concat_138"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_39_internal_tensor_assign_2_stride_0 = const()[name = tensor("new_cache_39_internal_tensor_assign_2_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_39_internal_tensor_assign_2_begin_mask_0 = const()[name = tensor("new_cache_39_internal_tensor_assign_2_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_39_internal_tensor_assign_2_end_mask_0 = const()[name = tensor("new_cache_39_internal_tensor_assign_2_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_39_internal_tensor_assign_2_squeeze_mask_0 = const()[name = tensor("new_cache_39_internal_tensor_assign_2_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_87 = const()[name = tensor("shape_87"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_39 = const()[name = tensor("reduce_prod_39"), val = tensor(1048576)]; + tensor range_1d_39_start_0 = const()[name = tensor("range_1d_39_start_0"), val = tensor(0)]; + tensor range_1d_39_step_0 = const()[name = tensor("range_1d_39_step_0"), val = tensor(1)]; + tensor range_1d_39 = range_1d(end = reduce_prod_39, start = range_1d_39_start_0, step = range_1d_39_step_0)[name = tensor("range_1d_39")]; + tensor reshape_195 = reshape(shape = shape_87, x = range_1d_39)[name = tensor("reshape_195")]; + tensor slice_by_index_39 = slice_by_index(begin = concat_137, begin_mask = new_cache_39_internal_tensor_assign_2_begin_mask_0, end = concat_138, end_mask = new_cache_39_internal_tensor_assign_2_end_mask_0, squeeze_mask = new_cache_39_internal_tensor_assign_2_squeeze_mask_0, stride = new_cache_39_internal_tensor_assign_2_stride_0, x = reshape_195)[name = tensor("slice_by_index_39")]; + tensor reshape_196_shape_0 = const()[name = tensor("reshape_196_shape_0"), val = tensor([-1])]; + tensor reshape_196 = reshape(shape = reshape_196_shape_0, x = slice_by_index_39)[name = tensor("reshape_196")]; + tensor reshape_197_shape_0 = const()[name = tensor("reshape_197_shape_0"), val = tensor([-1])]; + tensor reshape_197_cast_fp16 = reshape(shape = reshape_197_shape_0, x = var_8447_cast_fp16)[name = tensor("reshape_197_cast_fp16")]; + tensor reshape_198_shape_0 = const()[name = tensor("reshape_198_shape_0"), val = tensor([-1])]; + tensor reshape_198_cast_fp16 = reshape(shape = reshape_198_shape_0, x = reshape_194_cast_fp16)[name = tensor("reshape_198_cast_fp16")]; + tensor scatter_39_mode_0 = const()[name = tensor("scatter_39_mode_0"), val = tensor("update")]; + tensor scatter_39_axis_0 = const()[name = tensor("scatter_39_axis_0"), val = tensor(0)]; + tensor scatter_39_validate_indices_0 = const()[name = tensor("scatter_39_validate_indices_0"), val = tensor(false)]; + tensor scatter_39_cast_fp16 = scatter(axis = scatter_39_axis_0, data = reshape_198_cast_fp16, indices = reshape_196, mode = scatter_39_mode_0, updates = reshape_197_cast_fp16, validate_indices = scatter_39_validate_indices_0)[name = tensor("scatter_39_cast_fp16")]; + tensor reshape_199_cast_fp16 = reshape(shape = shape_87, x = scatter_39_cast_fp16)[name = tensor("reshape_199_cast_fp16")]; + tensor reshape_199_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("reshape_199_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor keys_115_begin_0 = const()[name = tensor("keys_115_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor keys_115_end_0 = const()[name = tensor("keys_115_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor keys_115_end_mask_0 = const()[name = tensor("keys_115_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor keys_115_squeeze_mask_0 = const()[name = tensor("keys_115_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor keys_115_cast_fp16 = slice_by_index(begin = keys_115_begin_0, end = keys_115_end_0, end_mask = keys_115_end_mask_0, squeeze_mask = keys_115_squeeze_mask_0, x = reshape_199_cast_fp16)[name = tensor("keys_115_cast_fp16")]; + tensor values_115_begin_0 = const()[name = tensor("values_115_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor values_115_end_0 = const()[name = tensor("values_115_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor values_115_end_mask_0 = const()[name = tensor("values_115_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor values_115_squeeze_mask_0 = const()[name = tensor("values_115_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor values_115_cast_fp16 = slice_by_index(begin = values_115_begin_0, end = values_115_end_0, end_mask = values_115_end_mask_0, squeeze_mask = values_115_squeeze_mask_0, x = reshape_199_cast_fp16)[name = tensor("values_115_cast_fp16")]; + tensor var_8459_cast_fp16 = not_equal(x = keys_115_cast_fp16, y = keys_115_cast_fp16)[name = tensor("op_8459_cast_fp16")]; + tensor keys_117_cast_fp16 = select(a = var_504_to_fp16, b = keys_115_cast_fp16, cond = var_8459_cast_fp16)[name = tensor("keys_117_cast_fp16")]; + tensor var_8467_cast_fp16 = not_equal(x = values_115_cast_fp16, y = values_115_cast_fp16)[name = tensor("op_8467_cast_fp16")]; + tensor values_117_cast_fp16 = select(a = var_504_to_fp16, b = values_115_cast_fp16, cond = var_8467_cast_fp16)[name = tensor("values_117_cast_fp16")]; + tensor var_8491 = const()[name = tensor("op_8491"), val = tensor([0, 2, 1, 3])]; + tensor var_8504 = const()[name = tensor("op_8504"), val = tensor([1, 1, 1])]; + tensor var_8505_cast_fp16 = reshape(shape = var_8504, x = position19_to_fp16)[name = tensor("op_8505_cast_fp16")]; + tensor var_8522_to_fp16 = const()[name = tensor("op_8522_to_fp16"), val = tensor(0x1p+0)]; + tensor valid_len_39_cast_fp16 = add(x = var_8505_cast_fp16, y = var_8522_to_fp16)[name = tensor("valid_len_39_cast_fp16")]; + tensor valid_mask_39_cast_fp16 = less(x = k_positions_1_promoted_to_fp16, y = valid_len_39_cast_fp16)[name = tensor("valid_mask_39_cast_fp16")]; + tensor causal_mask_39_cast_fp16 = less_equal(x = k_positions_1_promoted_to_fp16, y = var_8505_cast_fp16)[name = tensor("causal_mask_39_cast_fp16")]; + tensor attn_mask_77 = logical_and(x = valid_mask_39_cast_fp16, y = causal_mask_39_cast_fp16)[name = tensor("attn_mask_77")]; + tensor attn_mask_79_axes_0 = const()[name = tensor("attn_mask_79_axes_0"), val = tensor([1])]; + tensor attn_mask_79 = expand_dims(axes = attn_mask_79_axes_0, x = attn_mask_77)[name = tensor("attn_mask_79")]; + tensor var_8540_transpose_x_0 = const()[name = tensor("op_8540_transpose_x_0"), val = tensor(false)]; + tensor var_8540_transpose_y_0 = const()[name = tensor("op_8540_transpose_y_0"), val = tensor(false)]; + tensor transpose_110_perm_0 = const()[name = tensor("transpose_110_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_111_perm_0 = const()[name = tensor("transpose_111_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_111 = transpose(perm = transpose_111_perm_0, x = keys_117_cast_fp16)[name = tensor("transpose_137")]; + tensor transpose_110 = transpose(perm = transpose_110_perm_0, x = q_117_cast_fp16)[name = tensor("transpose_138")]; + tensor var_8540_cast_fp16 = matmul(transpose_x = var_8540_transpose_x_0, transpose_y = var_8540_transpose_y_0, x = transpose_110, y = transpose_111)[name = tensor("op_8540_cast_fp16")]; + tensor var_8534_to_fp16 = const()[name = tensor("op_8534_to_fp16"), val = tensor([0x1p-3])]; + tensor attn_weights_115_cast_fp16 = mul(x = var_8540_cast_fp16, y = var_8534_to_fp16)[name = tensor("attn_weights_115_cast_fp16")]; + tensor var_8542 = logical_not(x = attn_mask_79)[name = tensor("op_8542")]; + tensor var_8543_to_fp16 = const()[name = tensor("op_8543_to_fp16"), val = tensor(-inf)]; + tensor attn_weights_117_cast_fp16 = select(a = var_8543_to_fp16, b = attn_weights_115_cast_fp16, cond = var_8542)[name = tensor("attn_weights_117_cast_fp16")]; + tensor var_8545 = const()[name = tensor("op_8545"), val = tensor(-1)]; + tensor attn_weights_119_cast_fp16 = softmax(axis = var_8545, x = attn_weights_117_cast_fp16)[name = tensor("attn_weights_119_cast_fp16")]; + tensor attn_output_39_transpose_x_0 = const()[name = tensor("attn_output_39_transpose_x_0"), val = tensor(false)]; + tensor attn_output_39_transpose_y_0 = const()[name = tensor("attn_output_39_transpose_y_0"), val = tensor(false)]; + tensor values_119_cast_fp16 = transpose(perm = var_8491, x = values_117_cast_fp16)[name = tensor("transpose_139")]; + tensor attn_output_39_cast_fp16 = matmul(transpose_x = attn_output_39_transpose_x_0, transpose_y = attn_output_39_transpose_y_0, x = attn_weights_119_cast_fp16, y = values_119_cast_fp16)[name = tensor("attn_output_39_cast_fp16")]; + tensor var_8553 = const()[name = tensor("op_8553"), val = tensor([0, 2, 1, 3])]; + tensor var_8556 = const()[name = tensor("op_8556"), val = tensor([1, 1, 1024])]; + tensor var_8554_cast_fp16 = transpose(perm = var_8553, x = attn_output_39_cast_fp16)[name = tensor("transpose_136")]; + tensor input_195_cast_fp16 = reshape(shape = var_8556, x = var_8554_cast_fp16)[name = tensor("input_195_cast_fp16")]; + tensor linear_78_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = attn19_out_proj_weight_cast_fp16, x = input_195_cast_fp16)[name = tensor("linear_78_cast_fp16")]; + tensor var_8562_to_fp16 = const()[name = tensor("op_8562_to_fp16"), val = tensor(0x1p+0)]; + tensor var_8563_cast_fp16 = add(x = position19_to_fp16, y = var_8562_to_fp16)[name = tensor("op_8563_cast_fp16")]; + tensor var_8563_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_8563_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor input_197_cast_fp16 = add(x = input_193_cast_fp16, y = linear_78_cast_fp16)[name = tensor("input_197_cast_fp16")]; + tensor input_199_axes_0 = const()[name = tensor("input_199_axes_0"), val = tensor([-1])]; + tensor norm19_2_weight_to_fp16 = const()[name = tensor("norm19_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(303752192)))]; + tensor norm19_2_bias_to_fp16 = const()[name = tensor("norm19_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(303754304)))]; + tensor var_8567_to_fp16 = const()[name = tensor("op_8567_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_199_cast_fp16 = layer_norm(axes = input_199_axes_0, beta = norm19_2_bias_to_fp16, epsilon = var_8567_to_fp16, gamma = norm19_2_weight_to_fp16, x = input_197_cast_fp16)[name = tensor("input_199_cast_fp16")]; + tensor linear_79_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = linear19_1_weight_cast_fp16, x = input_199_cast_fp16)[name = tensor("linear_79_cast_fp16")]; + tensor input_201_mode_0 = const()[name = tensor("input_201_mode_0"), val = tensor("EXACT")]; + tensor input_201_cast_fp16 = gelu(mode = input_201_mode_0, x = linear_79_cast_fp16)[name = tensor("input_201_cast_fp16")]; + tensor linear_80_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = linear19_2_weight_cast_fp16, x = input_201_cast_fp16)[name = tensor("linear_80_cast_fp16")]; + tensor input_203_cast_fp16 = add(x = input_197_cast_fp16, y = linear_80_cast_fp16)[name = tensor("input_203_cast_fp16")]; + tensor x_41_axes_0 = const()[name = tensor("x_41_axes_0"), val = tensor([-1])]; + tensor norm20_1_weight_to_fp16 = const()[name = tensor("norm20_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(303756416)))]; + tensor norm20_1_bias_to_fp16 = const()[name = tensor("norm20_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(303758528)))]; + tensor var_8584_to_fp16 = const()[name = tensor("op_8584_to_fp16"), val = tensor(0x1.5p-17)]; + tensor x_41_cast_fp16 = layer_norm(axes = x_41_axes_0, beta = norm20_1_bias_to_fp16, epsilon = var_8584_to_fp16, gamma = norm20_1_weight_to_fp16, x = input_203_cast_fp16)[name = tensor("x_41_cast_fp16")]; + tensor linear_81_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = attn20_in_proj_weight_cast_fp16, x = x_41_cast_fp16)[name = tensor("linear_81_cast_fp16")]; + tensor var_8620 = const()[name = tensor("op_8620"), val = tensor([1, 1, 3, 16, 64])]; + tensor qkv_41_cast_fp16 = reshape(shape = var_8620, x = linear_81_cast_fp16)[name = tensor("qkv_41_cast_fp16")]; + tensor q_121_begin_0 = const()[name = tensor("q_121_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor q_121_end_0 = const()[name = tensor("q_121_end_0"), val = tensor([1, 1, 1, 16, 64])]; + tensor q_121_end_mask_0 = const()[name = tensor("q_121_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor q_121_squeeze_mask_0 = const()[name = tensor("q_121_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor q_121_cast_fp16 = slice_by_index(begin = q_121_begin_0, end = q_121_end_0, end_mask = q_121_end_mask_0, squeeze_mask = q_121_squeeze_mask_0, x = qkv_41_cast_fp16)[name = tensor("q_121_cast_fp16")]; + tensor k_81_begin_0 = const()[name = tensor("k_81_begin_0"), val = tensor([0, 0, 1, 0, 0])]; + tensor k_81_end_0 = const()[name = tensor("k_81_end_0"), val = tensor([1, 1, 2, 16, 64])]; + tensor k_81_end_mask_0 = const()[name = tensor("k_81_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor k_81_squeeze_mask_0 = const()[name = tensor("k_81_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor k_81_cast_fp16 = slice_by_index(begin = k_81_begin_0, end = k_81_end_0, end_mask = k_81_end_mask_0, squeeze_mask = k_81_squeeze_mask_0, x = qkv_41_cast_fp16)[name = tensor("k_81_cast_fp16")]; + tensor v_41_begin_0 = const()[name = tensor("v_41_begin_0"), val = tensor([0, 0, 2, 0, 0])]; + tensor v_41_end_0 = const()[name = tensor("v_41_end_0"), val = tensor([1, 1, 3, 16, 64])]; + tensor v_41_end_mask_0 = const()[name = tensor("v_41_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor v_41_squeeze_mask_0 = const()[name = tensor("v_41_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor v_41_cast_fp16 = slice_by_index(begin = v_41_begin_0, end = v_41_end_0, end_mask = v_41_end_mask_0, squeeze_mask = v_41_squeeze_mask_0, x = qkv_41_cast_fp16)[name = tensor("v_41_cast_fp16")]; + tensor var_8724 = const()[name = tensor("op_8724"), val = tensor([1, 1, 1, 1])]; + tensor position20_to_fp16_dtype_0 = const()[name = tensor("position20_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor position20_to_fp16 = cast(dtype = position20_to_fp16_dtype_0, x = position20)[name = tensor("cast_453")]; + tensor ts_125_cast_fp16 = reshape(shape = var_8724, x = position20_to_fp16)[name = tensor("ts_125_cast_fp16")]; + tensor var_8728 = const()[name = tensor("op_8728"), val = tensor([1, 1, 16, 32, 2])]; + tensor q_complex_41_cast_fp16 = reshape(shape = var_8728, x = q_121_cast_fp16)[name = tensor("q_complex_41_cast_fp16")]; + tensor var_8732 = const()[name = tensor("op_8732"), val = tensor([1, 1, 16, 32, 2])]; + tensor k_complex_41_cast_fp16 = reshape(shape = var_8732, x = k_81_cast_fp16)[name = tensor("k_complex_41_cast_fp16")]; + tensor var_8736_begin_0 = const()[name = tensor("op_8736_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_8736_end_0 = const()[name = tensor("op_8736_end_0"), val = tensor([1, 1, 16, 32, 1])]; + tensor var_8736_end_mask_0 = const()[name = tensor("op_8736_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_8736_squeeze_mask_0 = const()[name = tensor("op_8736_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_8736_cast_fp16 = slice_by_index(begin = var_8736_begin_0, end = var_8736_end_0, end_mask = var_8736_end_mask_0, squeeze_mask = var_8736_squeeze_mask_0, x = q_complex_41_cast_fp16)[name = tensor("op_8736_cast_fp16")]; + tensor var_8744_begin_0 = const()[name = tensor("op_8744_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_8744_end_0 = const()[name = tensor("op_8744_end_0"), val = tensor([1, 1, 16, 32, 2])]; + tensor var_8744_end_mask_0 = const()[name = tensor("op_8744_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_8744_squeeze_mask_0 = const()[name = tensor("op_8744_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_8744_cast_fp16 = slice_by_index(begin = var_8744_begin_0, end = var_8744_end_0, end_mask = var_8744_end_mask_0, squeeze_mask = var_8744_squeeze_mask_0, x = q_complex_41_cast_fp16)[name = tensor("op_8744_cast_fp16")]; + tensor var_8752_begin_0 = const()[name = tensor("op_8752_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_8752_end_0 = const()[name = tensor("op_8752_end_0"), val = tensor([1, 1, 16, 32, 1])]; + tensor var_8752_end_mask_0 = const()[name = tensor("op_8752_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_8752_squeeze_mask_0 = const()[name = tensor("op_8752_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_8752_cast_fp16 = slice_by_index(begin = var_8752_begin_0, end = var_8752_end_0, end_mask = var_8752_end_mask_0, squeeze_mask = var_8752_squeeze_mask_0, x = k_complex_41_cast_fp16)[name = tensor("op_8752_cast_fp16")]; + tensor var_8760_begin_0 = const()[name = tensor("op_8760_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_8760_end_0 = const()[name = tensor("op_8760_end_0"), val = tensor([1, 1, 16, 32, 2])]; + tensor var_8760_end_mask_0 = const()[name = tensor("op_8760_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_8760_squeeze_mask_0 = const()[name = tensor("op_8760_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_8760_cast_fp16 = slice_by_index(begin = var_8760_begin_0, end = var_8760_end_0, end_mask = var_8760_end_mask_0, squeeze_mask = var_8760_squeeze_mask_0, x = k_complex_41_cast_fp16)[name = tensor("op_8760_cast_fp16")]; + tensor freqs_41_to_fp16 = const()[name = tensor("freqs_41_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(303760640)))]; + tensor var_8766_cast_fp16 = mul(x = freqs_41_to_fp16, y = ts_125_cast_fp16)[name = tensor("op_8766_cast_fp16")]; + tensor rotr_41_cast_fp16 = cos(x = var_8766_cast_fp16)[name = tensor("rotr_41_cast_fp16")]; + tensor roti_41_cast_fp16 = sin(x = var_8766_cast_fp16)[name = tensor("roti_41_cast_fp16")]; + tensor var_8770_cast_fp16 = mul(x = var_8736_cast_fp16, y = rotr_41_cast_fp16)[name = tensor("op_8770_cast_fp16")]; + tensor var_8771_cast_fp16 = mul(x = var_8744_cast_fp16, y = roti_41_cast_fp16)[name = tensor("op_8771_cast_fp16")]; + tensor qor_81_cast_fp16 = sub(x = var_8770_cast_fp16, y = var_8771_cast_fp16)[name = tensor("qor_81_cast_fp16")]; + tensor var_8774_cast_fp16 = mul(x = var_8736_cast_fp16, y = roti_41_cast_fp16)[name = tensor("op_8774_cast_fp16")]; + tensor var_8775_cast_fp16 = mul(x = var_8744_cast_fp16, y = rotr_41_cast_fp16)[name = tensor("op_8775_cast_fp16")]; + tensor qoi_81_cast_fp16 = add(x = var_8774_cast_fp16, y = var_8775_cast_fp16)[name = tensor("qoi_81_cast_fp16")]; + tensor var_8778_cast_fp16 = mul(x = var_8752_cast_fp16, y = rotr_41_cast_fp16)[name = tensor("op_8778_cast_fp16")]; + tensor var_8779_cast_fp16 = mul(x = var_8760_cast_fp16, y = roti_41_cast_fp16)[name = tensor("op_8779_cast_fp16")]; + tensor kor_81_cast_fp16 = sub(x = var_8778_cast_fp16, y = var_8779_cast_fp16)[name = tensor("kor_81_cast_fp16")]; + tensor var_8782_cast_fp16 = mul(x = var_8752_cast_fp16, y = roti_41_cast_fp16)[name = tensor("op_8782_cast_fp16")]; + tensor var_8783_cast_fp16 = mul(x = var_8760_cast_fp16, y = rotr_41_cast_fp16)[name = tensor("op_8783_cast_fp16")]; + tensor koi_81_cast_fp16 = add(x = var_8782_cast_fp16, y = var_8783_cast_fp16)[name = tensor("koi_81_cast_fp16")]; + tensor qo_41_axis_0 = const()[name = tensor("qo_41_axis_0"), val = tensor(-1)]; + tensor qo_41_cast_fp16 = stack(axis = qo_41_axis_0, values = (qor_81_cast_fp16, qoi_81_cast_fp16))[name = tensor("qo_41_cast_fp16")]; + tensor ko_41_axis_0 = const()[name = tensor("ko_41_axis_0"), val = tensor(-1)]; + tensor ko_41_cast_fp16 = stack(axis = ko_41_axis_0, values = (kor_81_cast_fp16, koi_81_cast_fp16))[name = tensor("ko_41_cast_fp16")]; + tensor var_8812 = const()[name = tensor("op_8812"), val = tensor([1, 1, 16, 64])]; + tensor q_123_cast_fp16 = reshape(shape = var_8812, x = qo_41_cast_fp16)[name = tensor("q_123_cast_fp16")]; + tensor var_8814 = const()[name = tensor("op_8814"), val = tensor([1, 1, 16, 64])]; + tensor k_83_cast_fp16 = reshape(shape = var_8814, x = ko_41_cast_fp16)[name = tensor("k_83_cast_fp16")]; + tensor _inversed_8836_y_0_to_fp16 = const()[name = tensor("_inversed_8836_y_0_to_fp16"), val = tensor(0x1p-9)]; + tensor _inversed_8836_cast_fp16 = mul(x = ts_125_cast_fp16, y = _inversed_8836_y_0_to_fp16)[name = tensor("_inversed_8836_cast_fp16")]; + tensor var_8837_cast_fp16 = floor(x = _inversed_8836_cast_fp16)[name = tensor("op_8837_cast_fp16")]; + tensor var_8838_to_fp16 = const()[name = tensor("op_8838_to_fp16"), val = tensor(0x1p+9)]; + tensor var_8839_cast_fp16 = mul(x = var_8837_cast_fp16, y = var_8838_to_fp16)[name = tensor("op_8839_cast_fp16")]; + tensor write_indices_float_83_cast_fp16 = sub(x = ts_125_cast_fp16, y = var_8839_cast_fp16)[name = tensor("write_indices_float_83_cast_fp16")]; + tensor var_8846_dtype_0 = const()[name = tensor("op_8846_dtype_0"), val = tensor("int32")]; + tensor write_indices_41_reps_0 = const()[name = tensor("write_indices_41_reps_0"), val = tensor([1, 1, 16, 64])]; + tensor write_indices_float_83_cast_fp16_to_int32 = cast(dtype = var_8846_dtype_0, x = write_indices_float_83_cast_fp16)[name = tensor("cast_452")]; + tensor write_indices_41 = tile(reps = write_indices_41_reps_0, x = write_indices_float_83_cast_fp16_to_int32)[name = tensor("write_indices_41")]; + tensor var_8854_begin_0 = const()[name = tensor("op_8854_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_8854_end_0 = const()[name = tensor("op_8854_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor var_8854_end_mask_0 = const()[name = tensor("op_8854_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_8854_squeeze_mask_0 = const()[name = tensor("op_8854_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor cache20_to_fp16_dtype_0 = const()[name = tensor("cache20_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor cache20_to_fp16 = cast(dtype = cache20_to_fp16_dtype_0, x = cache20)[name = tensor("cast_451")]; + tensor var_8854_cast_fp16 = slice_by_index(begin = var_8854_begin_0, end = var_8854_end_0, end_mask = var_8854_end_mask_0, squeeze_mask = var_8854_squeeze_mask_0, x = cache20_to_fp16)[name = tensor("op_8854_cast_fp16")]; + tensor var_8856_axis_0 = const()[name = tensor("op_8856_axis_0"), val = tensor(1)]; + tensor var_8856_mode_0 = const()[name = tensor("op_8856_mode_0"), val = tensor("update")]; + tensor var_8856_validate_indices_0 = const()[name = tensor("op_8856_validate_indices_0"), val = tensor(false)]; + tensor var_8856_cast_fp16 = scatter_along_axis(axis = var_8856_axis_0, data = var_8854_cast_fp16, indices = write_indices_41, mode = var_8856_mode_0, updates = k_83_cast_fp16, validate_indices = var_8856_validate_indices_0)[name = tensor("op_8856_cast_fp16")]; + tensor concat_142 = const()[name = tensor("concat_142"), val = tensor([0, 0, 0, 0, 0])]; + tensor concat_143 = const()[name = tensor("concat_143"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_41_internal_tensor_assign_1_stride_0 = const()[name = tensor("new_cache_41_internal_tensor_assign_1_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_41_internal_tensor_assign_1_begin_mask_0 = const()[name = tensor("new_cache_41_internal_tensor_assign_1_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_41_internal_tensor_assign_1_end_mask_0 = const()[name = tensor("new_cache_41_internal_tensor_assign_1_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_41_internal_tensor_assign_1_squeeze_mask_0 = const()[name = tensor("new_cache_41_internal_tensor_assign_1_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_88 = const()[name = tensor("shape_88"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_40 = const()[name = tensor("reduce_prod_40"), val = tensor(1048576)]; + tensor range_1d_40_start_0 = const()[name = tensor("range_1d_40_start_0"), val = tensor(0)]; + tensor range_1d_40_step_0 = const()[name = tensor("range_1d_40_step_0"), val = tensor(1)]; + tensor range_1d_40 = range_1d(end = reduce_prod_40, start = range_1d_40_start_0, step = range_1d_40_step_0)[name = tensor("range_1d_40")]; + tensor reshape_200 = reshape(shape = shape_88, x = range_1d_40)[name = tensor("reshape_200")]; + tensor slice_by_index_40 = slice_by_index(begin = concat_142, begin_mask = new_cache_41_internal_tensor_assign_1_begin_mask_0, end = concat_143, end_mask = new_cache_41_internal_tensor_assign_1_end_mask_0, squeeze_mask = new_cache_41_internal_tensor_assign_1_squeeze_mask_0, stride = new_cache_41_internal_tensor_assign_1_stride_0, x = reshape_200)[name = tensor("slice_by_index_40")]; + tensor reshape_201_shape_0 = const()[name = tensor("reshape_201_shape_0"), val = tensor([-1])]; + tensor reshape_201 = reshape(shape = reshape_201_shape_0, x = slice_by_index_40)[name = tensor("reshape_201")]; + tensor reshape_202_shape_0 = const()[name = tensor("reshape_202_shape_0"), val = tensor([-1])]; + tensor reshape_202_cast_fp16 = reshape(shape = reshape_202_shape_0, x = var_8856_cast_fp16)[name = tensor("reshape_202_cast_fp16")]; + tensor reshape_203_shape_0 = const()[name = tensor("reshape_203_shape_0"), val = tensor([-1])]; + tensor reshape_203_cast_fp16 = reshape(shape = reshape_203_shape_0, x = cache20_to_fp16)[name = tensor("reshape_203_cast_fp16")]; + tensor scatter_40_mode_0 = const()[name = tensor("scatter_40_mode_0"), val = tensor("update")]; + tensor scatter_40_axis_0 = const()[name = tensor("scatter_40_axis_0"), val = tensor(0)]; + tensor scatter_40_validate_indices_0 = const()[name = tensor("scatter_40_validate_indices_0"), val = tensor(false)]; + tensor scatter_40_cast_fp16 = scatter(axis = scatter_40_axis_0, data = reshape_203_cast_fp16, indices = reshape_201, mode = scatter_40_mode_0, updates = reshape_202_cast_fp16, validate_indices = scatter_40_validate_indices_0)[name = tensor("scatter_40_cast_fp16")]; + tensor reshape_204_cast_fp16 = reshape(shape = shape_88, x = scatter_40_cast_fp16)[name = tensor("reshape_204_cast_fp16")]; + tensor var_8864_begin_0 = const()[name = tensor("op_8864_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor var_8864_end_0 = const()[name = tensor("op_8864_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor var_8864_end_mask_0 = const()[name = tensor("op_8864_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_8864_squeeze_mask_0 = const()[name = tensor("op_8864_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor var_8864_cast_fp16 = slice_by_index(begin = var_8864_begin_0, end = var_8864_end_0, end_mask = var_8864_end_mask_0, squeeze_mask = var_8864_squeeze_mask_0, x = reshape_204_cast_fp16)[name = tensor("op_8864_cast_fp16")]; + tensor var_8866_axis_0 = const()[name = tensor("op_8866_axis_0"), val = tensor(1)]; + tensor var_8866_mode_0 = const()[name = tensor("op_8866_mode_0"), val = tensor("update")]; + tensor var_8866_validate_indices_0 = const()[name = tensor("op_8866_validate_indices_0"), val = tensor(false)]; + tensor var_8866_cast_fp16 = scatter_along_axis(axis = var_8866_axis_0, data = var_8864_cast_fp16, indices = write_indices_41, mode = var_8866_mode_0, updates = v_41_cast_fp16, validate_indices = var_8866_validate_indices_0)[name = tensor("op_8866_cast_fp16")]; + tensor concat_144 = const()[name = tensor("concat_144"), val = tensor([1, 0, 0, 0, 0])]; + tensor concat_145 = const()[name = tensor("concat_145"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_41_internal_tensor_assign_2_stride_0 = const()[name = tensor("new_cache_41_internal_tensor_assign_2_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_41_internal_tensor_assign_2_begin_mask_0 = const()[name = tensor("new_cache_41_internal_tensor_assign_2_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_41_internal_tensor_assign_2_end_mask_0 = const()[name = tensor("new_cache_41_internal_tensor_assign_2_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_41_internal_tensor_assign_2_squeeze_mask_0 = const()[name = tensor("new_cache_41_internal_tensor_assign_2_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_89 = const()[name = tensor("shape_89"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_41 = const()[name = tensor("reduce_prod_41"), val = tensor(1048576)]; + tensor range_1d_41_start_0 = const()[name = tensor("range_1d_41_start_0"), val = tensor(0)]; + tensor range_1d_41_step_0 = const()[name = tensor("range_1d_41_step_0"), val = tensor(1)]; + tensor range_1d_41 = range_1d(end = reduce_prod_41, start = range_1d_41_start_0, step = range_1d_41_step_0)[name = tensor("range_1d_41")]; + tensor reshape_205 = reshape(shape = shape_89, x = range_1d_41)[name = tensor("reshape_205")]; + tensor slice_by_index_41 = slice_by_index(begin = concat_144, begin_mask = new_cache_41_internal_tensor_assign_2_begin_mask_0, end = concat_145, end_mask = new_cache_41_internal_tensor_assign_2_end_mask_0, squeeze_mask = new_cache_41_internal_tensor_assign_2_squeeze_mask_0, stride = new_cache_41_internal_tensor_assign_2_stride_0, x = reshape_205)[name = tensor("slice_by_index_41")]; + tensor reshape_206_shape_0 = const()[name = tensor("reshape_206_shape_0"), val = tensor([-1])]; + tensor reshape_206 = reshape(shape = reshape_206_shape_0, x = slice_by_index_41)[name = tensor("reshape_206")]; + tensor reshape_207_shape_0 = const()[name = tensor("reshape_207_shape_0"), val = tensor([-1])]; + tensor reshape_207_cast_fp16 = reshape(shape = reshape_207_shape_0, x = var_8866_cast_fp16)[name = tensor("reshape_207_cast_fp16")]; + tensor reshape_208_shape_0 = const()[name = tensor("reshape_208_shape_0"), val = tensor([-1])]; + tensor reshape_208_cast_fp16 = reshape(shape = reshape_208_shape_0, x = reshape_204_cast_fp16)[name = tensor("reshape_208_cast_fp16")]; + tensor scatter_41_mode_0 = const()[name = tensor("scatter_41_mode_0"), val = tensor("update")]; + tensor scatter_41_axis_0 = const()[name = tensor("scatter_41_axis_0"), val = tensor(0)]; + tensor scatter_41_validate_indices_0 = const()[name = tensor("scatter_41_validate_indices_0"), val = tensor(false)]; + tensor scatter_41_cast_fp16 = scatter(axis = scatter_41_axis_0, data = reshape_208_cast_fp16, indices = reshape_206, mode = scatter_41_mode_0, updates = reshape_207_cast_fp16, validate_indices = scatter_41_validate_indices_0)[name = tensor("scatter_41_cast_fp16")]; + tensor reshape_209_cast_fp16 = reshape(shape = shape_89, x = scatter_41_cast_fp16)[name = tensor("reshape_209_cast_fp16")]; + tensor reshape_209_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("reshape_209_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor keys_121_begin_0 = const()[name = tensor("keys_121_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor keys_121_end_0 = const()[name = tensor("keys_121_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor keys_121_end_mask_0 = const()[name = tensor("keys_121_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor keys_121_squeeze_mask_0 = const()[name = tensor("keys_121_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor keys_121_cast_fp16 = slice_by_index(begin = keys_121_begin_0, end = keys_121_end_0, end_mask = keys_121_end_mask_0, squeeze_mask = keys_121_squeeze_mask_0, x = reshape_209_cast_fp16)[name = tensor("keys_121_cast_fp16")]; + tensor values_121_begin_0 = const()[name = tensor("values_121_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor values_121_end_0 = const()[name = tensor("values_121_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor values_121_end_mask_0 = const()[name = tensor("values_121_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor values_121_squeeze_mask_0 = const()[name = tensor("values_121_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor values_121_cast_fp16 = slice_by_index(begin = values_121_begin_0, end = values_121_end_0, end_mask = values_121_end_mask_0, squeeze_mask = values_121_squeeze_mask_0, x = reshape_209_cast_fp16)[name = tensor("values_121_cast_fp16")]; + tensor var_8878_cast_fp16 = not_equal(x = keys_121_cast_fp16, y = keys_121_cast_fp16)[name = tensor("op_8878_cast_fp16")]; + tensor keys_123_cast_fp16 = select(a = var_504_to_fp16, b = keys_121_cast_fp16, cond = var_8878_cast_fp16)[name = tensor("keys_123_cast_fp16")]; + tensor var_8886_cast_fp16 = not_equal(x = values_121_cast_fp16, y = values_121_cast_fp16)[name = tensor("op_8886_cast_fp16")]; + tensor values_123_cast_fp16 = select(a = var_504_to_fp16, b = values_121_cast_fp16, cond = var_8886_cast_fp16)[name = tensor("values_123_cast_fp16")]; + tensor var_8910 = const()[name = tensor("op_8910"), val = tensor([0, 2, 1, 3])]; + tensor var_8923 = const()[name = tensor("op_8923"), val = tensor([1, 1, 1])]; + tensor var_8924_cast_fp16 = reshape(shape = var_8923, x = position20_to_fp16)[name = tensor("op_8924_cast_fp16")]; + tensor var_8941_to_fp16 = const()[name = tensor("op_8941_to_fp16"), val = tensor(0x1p+0)]; + tensor valid_len_41_cast_fp16 = add(x = var_8924_cast_fp16, y = var_8941_to_fp16)[name = tensor("valid_len_41_cast_fp16")]; + tensor valid_mask_41_cast_fp16 = less(x = k_positions_1_promoted_to_fp16, y = valid_len_41_cast_fp16)[name = tensor("valid_mask_41_cast_fp16")]; + tensor causal_mask_41_cast_fp16 = less_equal(x = k_positions_1_promoted_to_fp16, y = var_8924_cast_fp16)[name = tensor("causal_mask_41_cast_fp16")]; + tensor attn_mask_81 = logical_and(x = valid_mask_41_cast_fp16, y = causal_mask_41_cast_fp16)[name = tensor("attn_mask_81")]; + tensor attn_mask_83_axes_0 = const()[name = tensor("attn_mask_83_axes_0"), val = tensor([1])]; + tensor attn_mask_83 = expand_dims(axes = attn_mask_83_axes_0, x = attn_mask_81)[name = tensor("attn_mask_83")]; + tensor var_8959_transpose_x_0 = const()[name = tensor("op_8959_transpose_x_0"), val = tensor(false)]; + tensor var_8959_transpose_y_0 = const()[name = tensor("op_8959_transpose_y_0"), val = tensor(false)]; + tensor transpose_112_perm_0 = const()[name = tensor("transpose_112_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_113_perm_0 = const()[name = tensor("transpose_113_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_113 = transpose(perm = transpose_113_perm_0, x = keys_123_cast_fp16)[name = tensor("transpose_133")]; + tensor transpose_112 = transpose(perm = transpose_112_perm_0, x = q_123_cast_fp16)[name = tensor("transpose_134")]; + tensor var_8959_cast_fp16 = matmul(transpose_x = var_8959_transpose_x_0, transpose_y = var_8959_transpose_y_0, x = transpose_112, y = transpose_113)[name = tensor("op_8959_cast_fp16")]; + tensor var_8953_to_fp16 = const()[name = tensor("op_8953_to_fp16"), val = tensor([0x1p-3])]; + tensor attn_weights_121_cast_fp16 = mul(x = var_8959_cast_fp16, y = var_8953_to_fp16)[name = tensor("attn_weights_121_cast_fp16")]; + tensor var_8961 = logical_not(x = attn_mask_83)[name = tensor("op_8961")]; + tensor var_8962_to_fp16 = const()[name = tensor("op_8962_to_fp16"), val = tensor(-inf)]; + tensor attn_weights_123_cast_fp16 = select(a = var_8962_to_fp16, b = attn_weights_121_cast_fp16, cond = var_8961)[name = tensor("attn_weights_123_cast_fp16")]; + tensor var_8964 = const()[name = tensor("op_8964"), val = tensor(-1)]; + tensor attn_weights_125_cast_fp16 = softmax(axis = var_8964, x = attn_weights_123_cast_fp16)[name = tensor("attn_weights_125_cast_fp16")]; + tensor attn_output_41_transpose_x_0 = const()[name = tensor("attn_output_41_transpose_x_0"), val = tensor(false)]; + tensor attn_output_41_transpose_y_0 = const()[name = tensor("attn_output_41_transpose_y_0"), val = tensor(false)]; + tensor values_125_cast_fp16 = transpose(perm = var_8910, x = values_123_cast_fp16)[name = tensor("transpose_135")]; + tensor attn_output_41_cast_fp16 = matmul(transpose_x = attn_output_41_transpose_x_0, transpose_y = attn_output_41_transpose_y_0, x = attn_weights_125_cast_fp16, y = values_125_cast_fp16)[name = tensor("attn_output_41_cast_fp16")]; + tensor var_8972 = const()[name = tensor("op_8972"), val = tensor([0, 2, 1, 3])]; + tensor var_8975 = const()[name = tensor("op_8975"), val = tensor([1, 1, 1024])]; + tensor var_8973_cast_fp16 = transpose(perm = var_8972, x = attn_output_41_cast_fp16)[name = tensor("transpose_132")]; + tensor input_205_cast_fp16 = reshape(shape = var_8975, x = var_8973_cast_fp16)[name = tensor("input_205_cast_fp16")]; + tensor linear_82_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = attn20_out_proj_weight_cast_fp16, x = input_205_cast_fp16)[name = tensor("linear_82_cast_fp16")]; + tensor var_8981_to_fp16 = const()[name = tensor("op_8981_to_fp16"), val = tensor(0x1p+0)]; + tensor var_8982_cast_fp16 = add(x = position20_to_fp16, y = var_8981_to_fp16)[name = tensor("op_8982_cast_fp16")]; + tensor var_8982_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_8982_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor input_207_cast_fp16 = add(x = input_203_cast_fp16, y = linear_82_cast_fp16)[name = tensor("input_207_cast_fp16")]; + tensor input_209_axes_0 = const()[name = tensor("input_209_axes_0"), val = tensor([-1])]; + tensor norm20_2_weight_to_fp16 = const()[name = tensor("norm20_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(303760768)))]; + tensor norm20_2_bias_to_fp16 = const()[name = tensor("norm20_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(303762880)))]; + tensor var_8986_to_fp16 = const()[name = tensor("op_8986_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_209_cast_fp16 = layer_norm(axes = input_209_axes_0, beta = norm20_2_bias_to_fp16, epsilon = var_8986_to_fp16, gamma = norm20_2_weight_to_fp16, x = input_207_cast_fp16)[name = tensor("input_209_cast_fp16")]; + tensor linear_83_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = linear20_1_weight_cast_fp16, x = input_209_cast_fp16)[name = tensor("linear_83_cast_fp16")]; + tensor input_211_mode_0 = const()[name = tensor("input_211_mode_0"), val = tensor("EXACT")]; + tensor input_211_cast_fp16 = gelu(mode = input_211_mode_0, x = linear_83_cast_fp16)[name = tensor("input_211_cast_fp16")]; + tensor linear_84_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = linear20_2_weight_cast_fp16, x = input_211_cast_fp16)[name = tensor("linear_84_cast_fp16")]; + tensor input_213_cast_fp16 = add(x = input_207_cast_fp16, y = linear_84_cast_fp16)[name = tensor("input_213_cast_fp16")]; + tensor x_43_axes_0 = const()[name = tensor("x_43_axes_0"), val = tensor([-1])]; + tensor norm21_1_weight_to_fp16 = const()[name = tensor("norm21_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(303764992)))]; + tensor norm21_1_bias_to_fp16 = const()[name = tensor("norm21_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(303767104)))]; + tensor var_9003_to_fp16 = const()[name = tensor("op_9003_to_fp16"), val = tensor(0x1.5p-17)]; + tensor x_43_cast_fp16 = layer_norm(axes = x_43_axes_0, beta = norm21_1_bias_to_fp16, epsilon = var_9003_to_fp16, gamma = norm21_1_weight_to_fp16, x = input_213_cast_fp16)[name = tensor("x_43_cast_fp16")]; + tensor linear_85_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = attn21_in_proj_weight_cast_fp16, x = x_43_cast_fp16)[name = tensor("linear_85_cast_fp16")]; + tensor var_9039 = const()[name = tensor("op_9039"), val = tensor([1, 1, 3, 16, 64])]; + tensor qkv_43_cast_fp16 = reshape(shape = var_9039, x = linear_85_cast_fp16)[name = tensor("qkv_43_cast_fp16")]; + tensor q_127_begin_0 = const()[name = tensor("q_127_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor q_127_end_0 = const()[name = tensor("q_127_end_0"), val = tensor([1, 1, 1, 16, 64])]; + tensor q_127_end_mask_0 = const()[name = tensor("q_127_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor q_127_squeeze_mask_0 = const()[name = tensor("q_127_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor q_127_cast_fp16 = slice_by_index(begin = q_127_begin_0, end = q_127_end_0, end_mask = q_127_end_mask_0, squeeze_mask = q_127_squeeze_mask_0, x = qkv_43_cast_fp16)[name = tensor("q_127_cast_fp16")]; + tensor k_85_begin_0 = const()[name = tensor("k_85_begin_0"), val = tensor([0, 0, 1, 0, 0])]; + tensor k_85_end_0 = const()[name = tensor("k_85_end_0"), val = tensor([1, 1, 2, 16, 64])]; + tensor k_85_end_mask_0 = const()[name = tensor("k_85_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor k_85_squeeze_mask_0 = const()[name = tensor("k_85_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor k_85_cast_fp16 = slice_by_index(begin = k_85_begin_0, end = k_85_end_0, end_mask = k_85_end_mask_0, squeeze_mask = k_85_squeeze_mask_0, x = qkv_43_cast_fp16)[name = tensor("k_85_cast_fp16")]; + tensor v_43_begin_0 = const()[name = tensor("v_43_begin_0"), val = tensor([0, 0, 2, 0, 0])]; + tensor v_43_end_0 = const()[name = tensor("v_43_end_0"), val = tensor([1, 1, 3, 16, 64])]; + tensor v_43_end_mask_0 = const()[name = tensor("v_43_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor v_43_squeeze_mask_0 = const()[name = tensor("v_43_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor v_43_cast_fp16 = slice_by_index(begin = v_43_begin_0, end = v_43_end_0, end_mask = v_43_end_mask_0, squeeze_mask = v_43_squeeze_mask_0, x = qkv_43_cast_fp16)[name = tensor("v_43_cast_fp16")]; + tensor var_9143 = const()[name = tensor("op_9143"), val = tensor([1, 1, 1, 1])]; + tensor position21_to_fp16_dtype_0 = const()[name = tensor("position21_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor position21_to_fp16 = cast(dtype = position21_to_fp16_dtype_0, x = position21)[name = tensor("cast_448")]; + tensor ts_131_cast_fp16 = reshape(shape = var_9143, x = position21_to_fp16)[name = tensor("ts_131_cast_fp16")]; + tensor var_9147 = const()[name = tensor("op_9147"), val = tensor([1, 1, 16, 32, 2])]; + tensor q_complex_43_cast_fp16 = reshape(shape = var_9147, x = q_127_cast_fp16)[name = tensor("q_complex_43_cast_fp16")]; + tensor var_9151 = const()[name = tensor("op_9151"), val = tensor([1, 1, 16, 32, 2])]; + tensor k_complex_43_cast_fp16 = reshape(shape = var_9151, x = k_85_cast_fp16)[name = tensor("k_complex_43_cast_fp16")]; + tensor var_9155_begin_0 = const()[name = tensor("op_9155_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_9155_end_0 = const()[name = tensor("op_9155_end_0"), val = tensor([1, 1, 16, 32, 1])]; + tensor var_9155_end_mask_0 = const()[name = tensor("op_9155_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_9155_squeeze_mask_0 = const()[name = tensor("op_9155_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_9155_cast_fp16 = slice_by_index(begin = var_9155_begin_0, end = var_9155_end_0, end_mask = var_9155_end_mask_0, squeeze_mask = var_9155_squeeze_mask_0, x = q_complex_43_cast_fp16)[name = tensor("op_9155_cast_fp16")]; + tensor var_9163_begin_0 = const()[name = tensor("op_9163_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_9163_end_0 = const()[name = tensor("op_9163_end_0"), val = tensor([1, 1, 16, 32, 2])]; + tensor var_9163_end_mask_0 = const()[name = tensor("op_9163_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_9163_squeeze_mask_0 = const()[name = tensor("op_9163_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_9163_cast_fp16 = slice_by_index(begin = var_9163_begin_0, end = var_9163_end_0, end_mask = var_9163_end_mask_0, squeeze_mask = var_9163_squeeze_mask_0, x = q_complex_43_cast_fp16)[name = tensor("op_9163_cast_fp16")]; + tensor var_9171_begin_0 = const()[name = tensor("op_9171_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_9171_end_0 = const()[name = tensor("op_9171_end_0"), val = tensor([1, 1, 16, 32, 1])]; + tensor var_9171_end_mask_0 = const()[name = tensor("op_9171_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_9171_squeeze_mask_0 = const()[name = tensor("op_9171_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_9171_cast_fp16 = slice_by_index(begin = var_9171_begin_0, end = var_9171_end_0, end_mask = var_9171_end_mask_0, squeeze_mask = var_9171_squeeze_mask_0, x = k_complex_43_cast_fp16)[name = tensor("op_9171_cast_fp16")]; + tensor var_9179_begin_0 = const()[name = tensor("op_9179_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_9179_end_0 = const()[name = tensor("op_9179_end_0"), val = tensor([1, 1, 16, 32, 2])]; + tensor var_9179_end_mask_0 = const()[name = tensor("op_9179_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_9179_squeeze_mask_0 = const()[name = tensor("op_9179_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_9179_cast_fp16 = slice_by_index(begin = var_9179_begin_0, end = var_9179_end_0, end_mask = var_9179_end_mask_0, squeeze_mask = var_9179_squeeze_mask_0, x = k_complex_43_cast_fp16)[name = tensor("op_9179_cast_fp16")]; + tensor freqs_43_to_fp16 = const()[name = tensor("freqs_43_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(303769216)))]; + tensor var_9185_cast_fp16 = mul(x = freqs_43_to_fp16, y = ts_131_cast_fp16)[name = tensor("op_9185_cast_fp16")]; + tensor rotr_43_cast_fp16 = cos(x = var_9185_cast_fp16)[name = tensor("rotr_43_cast_fp16")]; + tensor roti_43_cast_fp16 = sin(x = var_9185_cast_fp16)[name = tensor("roti_43_cast_fp16")]; + tensor var_9189_cast_fp16 = mul(x = var_9155_cast_fp16, y = rotr_43_cast_fp16)[name = tensor("op_9189_cast_fp16")]; + tensor var_9190_cast_fp16 = mul(x = var_9163_cast_fp16, y = roti_43_cast_fp16)[name = tensor("op_9190_cast_fp16")]; + tensor qor_85_cast_fp16 = sub(x = var_9189_cast_fp16, y = var_9190_cast_fp16)[name = tensor("qor_85_cast_fp16")]; + tensor var_9193_cast_fp16 = mul(x = var_9155_cast_fp16, y = roti_43_cast_fp16)[name = tensor("op_9193_cast_fp16")]; + tensor var_9194_cast_fp16 = mul(x = var_9163_cast_fp16, y = rotr_43_cast_fp16)[name = tensor("op_9194_cast_fp16")]; + tensor qoi_85_cast_fp16 = add(x = var_9193_cast_fp16, y = var_9194_cast_fp16)[name = tensor("qoi_85_cast_fp16")]; + tensor var_9197_cast_fp16 = mul(x = var_9171_cast_fp16, y = rotr_43_cast_fp16)[name = tensor("op_9197_cast_fp16")]; + tensor var_9198_cast_fp16 = mul(x = var_9179_cast_fp16, y = roti_43_cast_fp16)[name = tensor("op_9198_cast_fp16")]; + tensor kor_85_cast_fp16 = sub(x = var_9197_cast_fp16, y = var_9198_cast_fp16)[name = tensor("kor_85_cast_fp16")]; + tensor var_9201_cast_fp16 = mul(x = var_9171_cast_fp16, y = roti_43_cast_fp16)[name = tensor("op_9201_cast_fp16")]; + tensor var_9202_cast_fp16 = mul(x = var_9179_cast_fp16, y = rotr_43_cast_fp16)[name = tensor("op_9202_cast_fp16")]; + tensor koi_85_cast_fp16 = add(x = var_9201_cast_fp16, y = var_9202_cast_fp16)[name = tensor("koi_85_cast_fp16")]; + tensor qo_43_axis_0 = const()[name = tensor("qo_43_axis_0"), val = tensor(-1)]; + tensor qo_43_cast_fp16 = stack(axis = qo_43_axis_0, values = (qor_85_cast_fp16, qoi_85_cast_fp16))[name = tensor("qo_43_cast_fp16")]; + tensor ko_43_axis_0 = const()[name = tensor("ko_43_axis_0"), val = tensor(-1)]; + tensor ko_43_cast_fp16 = stack(axis = ko_43_axis_0, values = (kor_85_cast_fp16, koi_85_cast_fp16))[name = tensor("ko_43_cast_fp16")]; + tensor var_9231 = const()[name = tensor("op_9231"), val = tensor([1, 1, 16, 64])]; + tensor q_129_cast_fp16 = reshape(shape = var_9231, x = qo_43_cast_fp16)[name = tensor("q_129_cast_fp16")]; + tensor var_9233 = const()[name = tensor("op_9233"), val = tensor([1, 1, 16, 64])]; + tensor k_87_cast_fp16 = reshape(shape = var_9233, x = ko_43_cast_fp16)[name = tensor("k_87_cast_fp16")]; + tensor _inversed_9255_y_0_to_fp16 = const()[name = tensor("_inversed_9255_y_0_to_fp16"), val = tensor(0x1p-9)]; + tensor _inversed_9255_cast_fp16 = mul(x = ts_131_cast_fp16, y = _inversed_9255_y_0_to_fp16)[name = tensor("_inversed_9255_cast_fp16")]; + tensor var_9256_cast_fp16 = floor(x = _inversed_9255_cast_fp16)[name = tensor("op_9256_cast_fp16")]; + tensor var_9257_to_fp16 = const()[name = tensor("op_9257_to_fp16"), val = tensor(0x1p+9)]; + tensor var_9258_cast_fp16 = mul(x = var_9256_cast_fp16, y = var_9257_to_fp16)[name = tensor("op_9258_cast_fp16")]; + tensor write_indices_float_87_cast_fp16 = sub(x = ts_131_cast_fp16, y = var_9258_cast_fp16)[name = tensor("write_indices_float_87_cast_fp16")]; + tensor var_9265_dtype_0 = const()[name = tensor("op_9265_dtype_0"), val = tensor("int32")]; + tensor write_indices_43_reps_0 = const()[name = tensor("write_indices_43_reps_0"), val = tensor([1, 1, 16, 64])]; + tensor write_indices_float_87_cast_fp16_to_int32 = cast(dtype = var_9265_dtype_0, x = write_indices_float_87_cast_fp16)[name = tensor("cast_447")]; + tensor write_indices_43 = tile(reps = write_indices_43_reps_0, x = write_indices_float_87_cast_fp16_to_int32)[name = tensor("write_indices_43")]; + tensor var_9273_begin_0 = const()[name = tensor("op_9273_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_9273_end_0 = const()[name = tensor("op_9273_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor var_9273_end_mask_0 = const()[name = tensor("op_9273_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_9273_squeeze_mask_0 = const()[name = tensor("op_9273_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor cache21_to_fp16_dtype_0 = const()[name = tensor("cache21_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor cache21_to_fp16 = cast(dtype = cache21_to_fp16_dtype_0, x = cache21)[name = tensor("cast_446")]; + tensor var_9273_cast_fp16 = slice_by_index(begin = var_9273_begin_0, end = var_9273_end_0, end_mask = var_9273_end_mask_0, squeeze_mask = var_9273_squeeze_mask_0, x = cache21_to_fp16)[name = tensor("op_9273_cast_fp16")]; + tensor var_9275_axis_0 = const()[name = tensor("op_9275_axis_0"), val = tensor(1)]; + tensor var_9275_mode_0 = const()[name = tensor("op_9275_mode_0"), val = tensor("update")]; + tensor var_9275_validate_indices_0 = const()[name = tensor("op_9275_validate_indices_0"), val = tensor(false)]; + tensor var_9275_cast_fp16 = scatter_along_axis(axis = var_9275_axis_0, data = var_9273_cast_fp16, indices = write_indices_43, mode = var_9275_mode_0, updates = k_87_cast_fp16, validate_indices = var_9275_validate_indices_0)[name = tensor("op_9275_cast_fp16")]; + tensor concat_149 = const()[name = tensor("concat_149"), val = tensor([0, 0, 0, 0, 0])]; + tensor concat_150 = const()[name = tensor("concat_150"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_43_internal_tensor_assign_1_stride_0 = const()[name = tensor("new_cache_43_internal_tensor_assign_1_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_43_internal_tensor_assign_1_begin_mask_0 = const()[name = tensor("new_cache_43_internal_tensor_assign_1_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_43_internal_tensor_assign_1_end_mask_0 = const()[name = tensor("new_cache_43_internal_tensor_assign_1_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_43_internal_tensor_assign_1_squeeze_mask_0 = const()[name = tensor("new_cache_43_internal_tensor_assign_1_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_90 = const()[name = tensor("shape_90"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_42 = const()[name = tensor("reduce_prod_42"), val = tensor(1048576)]; + tensor range_1d_42_start_0 = const()[name = tensor("range_1d_42_start_0"), val = tensor(0)]; + tensor range_1d_42_step_0 = const()[name = tensor("range_1d_42_step_0"), val = tensor(1)]; + tensor range_1d_42 = range_1d(end = reduce_prod_42, start = range_1d_42_start_0, step = range_1d_42_step_0)[name = tensor("range_1d_42")]; + tensor reshape_210 = reshape(shape = shape_90, x = range_1d_42)[name = tensor("reshape_210")]; + tensor slice_by_index_42 = slice_by_index(begin = concat_149, begin_mask = new_cache_43_internal_tensor_assign_1_begin_mask_0, end = concat_150, end_mask = new_cache_43_internal_tensor_assign_1_end_mask_0, squeeze_mask = new_cache_43_internal_tensor_assign_1_squeeze_mask_0, stride = new_cache_43_internal_tensor_assign_1_stride_0, x = reshape_210)[name = tensor("slice_by_index_42")]; + tensor reshape_211_shape_0 = const()[name = tensor("reshape_211_shape_0"), val = tensor([-1])]; + tensor reshape_211 = reshape(shape = reshape_211_shape_0, x = slice_by_index_42)[name = tensor("reshape_211")]; + tensor reshape_212_shape_0 = const()[name = tensor("reshape_212_shape_0"), val = tensor([-1])]; + tensor reshape_212_cast_fp16 = reshape(shape = reshape_212_shape_0, x = var_9275_cast_fp16)[name = tensor("reshape_212_cast_fp16")]; + tensor reshape_213_shape_0 = const()[name = tensor("reshape_213_shape_0"), val = tensor([-1])]; + tensor reshape_213_cast_fp16 = reshape(shape = reshape_213_shape_0, x = cache21_to_fp16)[name = tensor("reshape_213_cast_fp16")]; + tensor scatter_42_mode_0 = const()[name = tensor("scatter_42_mode_0"), val = tensor("update")]; + tensor scatter_42_axis_0 = const()[name = tensor("scatter_42_axis_0"), val = tensor(0)]; + tensor scatter_42_validate_indices_0 = const()[name = tensor("scatter_42_validate_indices_0"), val = tensor(false)]; + tensor scatter_42_cast_fp16 = scatter(axis = scatter_42_axis_0, data = reshape_213_cast_fp16, indices = reshape_211, mode = scatter_42_mode_0, updates = reshape_212_cast_fp16, validate_indices = scatter_42_validate_indices_0)[name = tensor("scatter_42_cast_fp16")]; + tensor reshape_214_cast_fp16 = reshape(shape = shape_90, x = scatter_42_cast_fp16)[name = tensor("reshape_214_cast_fp16")]; + tensor var_9283_begin_0 = const()[name = tensor("op_9283_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor var_9283_end_0 = const()[name = tensor("op_9283_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor var_9283_end_mask_0 = const()[name = tensor("op_9283_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_9283_squeeze_mask_0 = const()[name = tensor("op_9283_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor var_9283_cast_fp16 = slice_by_index(begin = var_9283_begin_0, end = var_9283_end_0, end_mask = var_9283_end_mask_0, squeeze_mask = var_9283_squeeze_mask_0, x = reshape_214_cast_fp16)[name = tensor("op_9283_cast_fp16")]; + tensor var_9285_axis_0 = const()[name = tensor("op_9285_axis_0"), val = tensor(1)]; + tensor var_9285_mode_0 = const()[name = tensor("op_9285_mode_0"), val = tensor("update")]; + tensor var_9285_validate_indices_0 = const()[name = tensor("op_9285_validate_indices_0"), val = tensor(false)]; + tensor var_9285_cast_fp16 = scatter_along_axis(axis = var_9285_axis_0, data = var_9283_cast_fp16, indices = write_indices_43, mode = var_9285_mode_0, updates = v_43_cast_fp16, validate_indices = var_9285_validate_indices_0)[name = tensor("op_9285_cast_fp16")]; + tensor concat_151 = const()[name = tensor("concat_151"), val = tensor([1, 0, 0, 0, 0])]; + tensor concat_152 = const()[name = tensor("concat_152"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_43_internal_tensor_assign_2_stride_0 = const()[name = tensor("new_cache_43_internal_tensor_assign_2_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_43_internal_tensor_assign_2_begin_mask_0 = const()[name = tensor("new_cache_43_internal_tensor_assign_2_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_43_internal_tensor_assign_2_end_mask_0 = const()[name = tensor("new_cache_43_internal_tensor_assign_2_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_43_internal_tensor_assign_2_squeeze_mask_0 = const()[name = tensor("new_cache_43_internal_tensor_assign_2_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_91 = const()[name = tensor("shape_91"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_43 = const()[name = tensor("reduce_prod_43"), val = tensor(1048576)]; + tensor range_1d_43_start_0 = const()[name = tensor("range_1d_43_start_0"), val = tensor(0)]; + tensor range_1d_43_step_0 = const()[name = tensor("range_1d_43_step_0"), val = tensor(1)]; + tensor range_1d_43 = range_1d(end = reduce_prod_43, start = range_1d_43_start_0, step = range_1d_43_step_0)[name = tensor("range_1d_43")]; + tensor reshape_215 = reshape(shape = shape_91, x = range_1d_43)[name = tensor("reshape_215")]; + tensor slice_by_index_43 = slice_by_index(begin = concat_151, begin_mask = new_cache_43_internal_tensor_assign_2_begin_mask_0, end = concat_152, end_mask = new_cache_43_internal_tensor_assign_2_end_mask_0, squeeze_mask = new_cache_43_internal_tensor_assign_2_squeeze_mask_0, stride = new_cache_43_internal_tensor_assign_2_stride_0, x = reshape_215)[name = tensor("slice_by_index_43")]; + tensor reshape_216_shape_0 = const()[name = tensor("reshape_216_shape_0"), val = tensor([-1])]; + tensor reshape_216 = reshape(shape = reshape_216_shape_0, x = slice_by_index_43)[name = tensor("reshape_216")]; + tensor reshape_217_shape_0 = const()[name = tensor("reshape_217_shape_0"), val = tensor([-1])]; + tensor reshape_217_cast_fp16 = reshape(shape = reshape_217_shape_0, x = var_9285_cast_fp16)[name = tensor("reshape_217_cast_fp16")]; + tensor reshape_218_shape_0 = const()[name = tensor("reshape_218_shape_0"), val = tensor([-1])]; + tensor reshape_218_cast_fp16 = reshape(shape = reshape_218_shape_0, x = reshape_214_cast_fp16)[name = tensor("reshape_218_cast_fp16")]; + tensor scatter_43_mode_0 = const()[name = tensor("scatter_43_mode_0"), val = tensor("update")]; + tensor scatter_43_axis_0 = const()[name = tensor("scatter_43_axis_0"), val = tensor(0)]; + tensor scatter_43_validate_indices_0 = const()[name = tensor("scatter_43_validate_indices_0"), val = tensor(false)]; + tensor scatter_43_cast_fp16 = scatter(axis = scatter_43_axis_0, data = reshape_218_cast_fp16, indices = reshape_216, mode = scatter_43_mode_0, updates = reshape_217_cast_fp16, validate_indices = scatter_43_validate_indices_0)[name = tensor("scatter_43_cast_fp16")]; + tensor reshape_219_cast_fp16 = reshape(shape = shape_91, x = scatter_43_cast_fp16)[name = tensor("reshape_219_cast_fp16")]; + tensor reshape_219_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("reshape_219_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor keys_127_begin_0 = const()[name = tensor("keys_127_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor keys_127_end_0 = const()[name = tensor("keys_127_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor keys_127_end_mask_0 = const()[name = tensor("keys_127_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor keys_127_squeeze_mask_0 = const()[name = tensor("keys_127_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor keys_127_cast_fp16 = slice_by_index(begin = keys_127_begin_0, end = keys_127_end_0, end_mask = keys_127_end_mask_0, squeeze_mask = keys_127_squeeze_mask_0, x = reshape_219_cast_fp16)[name = tensor("keys_127_cast_fp16")]; + tensor values_127_begin_0 = const()[name = tensor("values_127_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor values_127_end_0 = const()[name = tensor("values_127_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor values_127_end_mask_0 = const()[name = tensor("values_127_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor values_127_squeeze_mask_0 = const()[name = tensor("values_127_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor values_127_cast_fp16 = slice_by_index(begin = values_127_begin_0, end = values_127_end_0, end_mask = values_127_end_mask_0, squeeze_mask = values_127_squeeze_mask_0, x = reshape_219_cast_fp16)[name = tensor("values_127_cast_fp16")]; + tensor var_9297_cast_fp16 = not_equal(x = keys_127_cast_fp16, y = keys_127_cast_fp16)[name = tensor("op_9297_cast_fp16")]; + tensor keys_129_cast_fp16 = select(a = var_504_to_fp16, b = keys_127_cast_fp16, cond = var_9297_cast_fp16)[name = tensor("keys_129_cast_fp16")]; + tensor var_9305_cast_fp16 = not_equal(x = values_127_cast_fp16, y = values_127_cast_fp16)[name = tensor("op_9305_cast_fp16")]; + tensor values_129_cast_fp16 = select(a = var_504_to_fp16, b = values_127_cast_fp16, cond = var_9305_cast_fp16)[name = tensor("values_129_cast_fp16")]; + tensor var_9329 = const()[name = tensor("op_9329"), val = tensor([0, 2, 1, 3])]; + tensor var_9342 = const()[name = tensor("op_9342"), val = tensor([1, 1, 1])]; + tensor var_9343_cast_fp16 = reshape(shape = var_9342, x = position21_to_fp16)[name = tensor("op_9343_cast_fp16")]; + tensor var_9360_to_fp16 = const()[name = tensor("op_9360_to_fp16"), val = tensor(0x1p+0)]; + tensor valid_len_43_cast_fp16 = add(x = var_9343_cast_fp16, y = var_9360_to_fp16)[name = tensor("valid_len_43_cast_fp16")]; + tensor valid_mask_43_cast_fp16 = less(x = k_positions_1_promoted_to_fp16, y = valid_len_43_cast_fp16)[name = tensor("valid_mask_43_cast_fp16")]; + tensor causal_mask_43_cast_fp16 = less_equal(x = k_positions_1_promoted_to_fp16, y = var_9343_cast_fp16)[name = tensor("causal_mask_43_cast_fp16")]; + tensor attn_mask_85 = logical_and(x = valid_mask_43_cast_fp16, y = causal_mask_43_cast_fp16)[name = tensor("attn_mask_85")]; + tensor attn_mask_87_axes_0 = const()[name = tensor("attn_mask_87_axes_0"), val = tensor([1])]; + tensor attn_mask_87 = expand_dims(axes = attn_mask_87_axes_0, x = attn_mask_85)[name = tensor("attn_mask_87")]; + tensor var_9378_transpose_x_0 = const()[name = tensor("op_9378_transpose_x_0"), val = tensor(false)]; + tensor var_9378_transpose_y_0 = const()[name = tensor("op_9378_transpose_y_0"), val = tensor(false)]; + tensor transpose_114_perm_0 = const()[name = tensor("transpose_114_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_115_perm_0 = const()[name = tensor("transpose_115_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_115 = transpose(perm = transpose_115_perm_0, x = keys_129_cast_fp16)[name = tensor("transpose_129")]; + tensor transpose_114 = transpose(perm = transpose_114_perm_0, x = q_129_cast_fp16)[name = tensor("transpose_130")]; + tensor var_9378_cast_fp16 = matmul(transpose_x = var_9378_transpose_x_0, transpose_y = var_9378_transpose_y_0, x = transpose_114, y = transpose_115)[name = tensor("op_9378_cast_fp16")]; + tensor var_9372_to_fp16 = const()[name = tensor("op_9372_to_fp16"), val = tensor([0x1p-3])]; + tensor attn_weights_127_cast_fp16 = mul(x = var_9378_cast_fp16, y = var_9372_to_fp16)[name = tensor("attn_weights_127_cast_fp16")]; + tensor var_9380 = logical_not(x = attn_mask_87)[name = tensor("op_9380")]; + tensor var_9381_to_fp16 = const()[name = tensor("op_9381_to_fp16"), val = tensor(-inf)]; + tensor attn_weights_129_cast_fp16 = select(a = var_9381_to_fp16, b = attn_weights_127_cast_fp16, cond = var_9380)[name = tensor("attn_weights_129_cast_fp16")]; + tensor var_9383 = const()[name = tensor("op_9383"), val = tensor(-1)]; + tensor attn_weights_131_cast_fp16 = softmax(axis = var_9383, x = attn_weights_129_cast_fp16)[name = tensor("attn_weights_131_cast_fp16")]; + tensor attn_output_43_transpose_x_0 = const()[name = tensor("attn_output_43_transpose_x_0"), val = tensor(false)]; + tensor attn_output_43_transpose_y_0 = const()[name = tensor("attn_output_43_transpose_y_0"), val = tensor(false)]; + tensor values_131_cast_fp16 = transpose(perm = var_9329, x = values_129_cast_fp16)[name = tensor("transpose_131")]; + tensor attn_output_43_cast_fp16 = matmul(transpose_x = attn_output_43_transpose_x_0, transpose_y = attn_output_43_transpose_y_0, x = attn_weights_131_cast_fp16, y = values_131_cast_fp16)[name = tensor("attn_output_43_cast_fp16")]; + tensor var_9391 = const()[name = tensor("op_9391"), val = tensor([0, 2, 1, 3])]; + tensor var_9394 = const()[name = tensor("op_9394"), val = tensor([1, 1, 1024])]; + tensor var_9392_cast_fp16 = transpose(perm = var_9391, x = attn_output_43_cast_fp16)[name = tensor("transpose_128")]; + tensor input_215_cast_fp16 = reshape(shape = var_9394, x = var_9392_cast_fp16)[name = tensor("input_215_cast_fp16")]; + tensor linear_86_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = attn21_out_proj_weight_cast_fp16, x = input_215_cast_fp16)[name = tensor("linear_86_cast_fp16")]; + tensor var_9400_to_fp16 = const()[name = tensor("op_9400_to_fp16"), val = tensor(0x1p+0)]; + tensor var_9401_cast_fp16 = add(x = position21_to_fp16, y = var_9400_to_fp16)[name = tensor("op_9401_cast_fp16")]; + tensor var_9401_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_9401_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor input_217_cast_fp16 = add(x = input_213_cast_fp16, y = linear_86_cast_fp16)[name = tensor("input_217_cast_fp16")]; + tensor input_219_axes_0 = const()[name = tensor("input_219_axes_0"), val = tensor([-1])]; + tensor norm21_2_weight_to_fp16 = const()[name = tensor("norm21_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(303769344)))]; + tensor norm21_2_bias_to_fp16 = const()[name = tensor("norm21_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(303771456)))]; + tensor var_9405_to_fp16 = const()[name = tensor("op_9405_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_219_cast_fp16 = layer_norm(axes = input_219_axes_0, beta = norm21_2_bias_to_fp16, epsilon = var_9405_to_fp16, gamma = norm21_2_weight_to_fp16, x = input_217_cast_fp16)[name = tensor("input_219_cast_fp16")]; + tensor linear_87_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = linear21_1_weight_cast_fp16, x = input_219_cast_fp16)[name = tensor("linear_87_cast_fp16")]; + tensor input_221_mode_0 = const()[name = tensor("input_221_mode_0"), val = tensor("EXACT")]; + tensor input_221_cast_fp16 = gelu(mode = input_221_mode_0, x = linear_87_cast_fp16)[name = tensor("input_221_cast_fp16")]; + tensor linear_88_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = linear21_2_weight_cast_fp16, x = input_221_cast_fp16)[name = tensor("linear_88_cast_fp16")]; + tensor input_223_cast_fp16 = add(x = input_217_cast_fp16, y = linear_88_cast_fp16)[name = tensor("input_223_cast_fp16")]; + tensor x_45_axes_0 = const()[name = tensor("x_45_axes_0"), val = tensor([-1])]; + tensor norm22_1_weight_to_fp16 = const()[name = tensor("norm22_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(303773568)))]; + tensor norm22_1_bias_to_fp16 = const()[name = tensor("norm22_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(303775680)))]; + tensor var_9422_to_fp16 = const()[name = tensor("op_9422_to_fp16"), val = tensor(0x1.5p-17)]; + tensor x_45_cast_fp16 = layer_norm(axes = x_45_axes_0, beta = norm22_1_bias_to_fp16, epsilon = var_9422_to_fp16, gamma = norm22_1_weight_to_fp16, x = input_223_cast_fp16)[name = tensor("x_45_cast_fp16")]; + tensor linear_89_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = attn22_in_proj_weight_cast_fp16, x = x_45_cast_fp16)[name = tensor("linear_89_cast_fp16")]; + tensor var_9458 = const()[name = tensor("op_9458"), val = tensor([1, 1, 3, 16, 64])]; + tensor qkv_45_cast_fp16 = reshape(shape = var_9458, x = linear_89_cast_fp16)[name = tensor("qkv_45_cast_fp16")]; + tensor q_133_begin_0 = const()[name = tensor("q_133_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor q_133_end_0 = const()[name = tensor("q_133_end_0"), val = tensor([1, 1, 1, 16, 64])]; + tensor q_133_end_mask_0 = const()[name = tensor("q_133_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor q_133_squeeze_mask_0 = const()[name = tensor("q_133_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor q_133_cast_fp16 = slice_by_index(begin = q_133_begin_0, end = q_133_end_0, end_mask = q_133_end_mask_0, squeeze_mask = q_133_squeeze_mask_0, x = qkv_45_cast_fp16)[name = tensor("q_133_cast_fp16")]; + tensor k_89_begin_0 = const()[name = tensor("k_89_begin_0"), val = tensor([0, 0, 1, 0, 0])]; + tensor k_89_end_0 = const()[name = tensor("k_89_end_0"), val = tensor([1, 1, 2, 16, 64])]; + tensor k_89_end_mask_0 = const()[name = tensor("k_89_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor k_89_squeeze_mask_0 = const()[name = tensor("k_89_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor k_89_cast_fp16 = slice_by_index(begin = k_89_begin_0, end = k_89_end_0, end_mask = k_89_end_mask_0, squeeze_mask = k_89_squeeze_mask_0, x = qkv_45_cast_fp16)[name = tensor("k_89_cast_fp16")]; + tensor v_45_begin_0 = const()[name = tensor("v_45_begin_0"), val = tensor([0, 0, 2, 0, 0])]; + tensor v_45_end_0 = const()[name = tensor("v_45_end_0"), val = tensor([1, 1, 3, 16, 64])]; + tensor v_45_end_mask_0 = const()[name = tensor("v_45_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor v_45_squeeze_mask_0 = const()[name = tensor("v_45_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor v_45_cast_fp16 = slice_by_index(begin = v_45_begin_0, end = v_45_end_0, end_mask = v_45_end_mask_0, squeeze_mask = v_45_squeeze_mask_0, x = qkv_45_cast_fp16)[name = tensor("v_45_cast_fp16")]; + tensor var_9562 = const()[name = tensor("op_9562"), val = tensor([1, 1, 1, 1])]; + tensor position22_to_fp16_dtype_0 = const()[name = tensor("position22_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor position22_to_fp16 = cast(dtype = position22_to_fp16_dtype_0, x = position22)[name = tensor("cast_443")]; + tensor ts_137_cast_fp16 = reshape(shape = var_9562, x = position22_to_fp16)[name = tensor("ts_137_cast_fp16")]; + tensor var_9566 = const()[name = tensor("op_9566"), val = tensor([1, 1, 16, 32, 2])]; + tensor q_complex_45_cast_fp16 = reshape(shape = var_9566, x = q_133_cast_fp16)[name = tensor("q_complex_45_cast_fp16")]; + tensor var_9570 = const()[name = tensor("op_9570"), val = tensor([1, 1, 16, 32, 2])]; + tensor k_complex_45_cast_fp16 = reshape(shape = var_9570, x = k_89_cast_fp16)[name = tensor("k_complex_45_cast_fp16")]; + tensor var_9574_begin_0 = const()[name = tensor("op_9574_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_9574_end_0 = const()[name = tensor("op_9574_end_0"), val = tensor([1, 1, 16, 32, 1])]; + tensor var_9574_end_mask_0 = const()[name = tensor("op_9574_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_9574_squeeze_mask_0 = const()[name = tensor("op_9574_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_9574_cast_fp16 = slice_by_index(begin = var_9574_begin_0, end = var_9574_end_0, end_mask = var_9574_end_mask_0, squeeze_mask = var_9574_squeeze_mask_0, x = q_complex_45_cast_fp16)[name = tensor("op_9574_cast_fp16")]; + tensor var_9582_begin_0 = const()[name = tensor("op_9582_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_9582_end_0 = const()[name = tensor("op_9582_end_0"), val = tensor([1, 1, 16, 32, 2])]; + tensor var_9582_end_mask_0 = const()[name = tensor("op_9582_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_9582_squeeze_mask_0 = const()[name = tensor("op_9582_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_9582_cast_fp16 = slice_by_index(begin = var_9582_begin_0, end = var_9582_end_0, end_mask = var_9582_end_mask_0, squeeze_mask = var_9582_squeeze_mask_0, x = q_complex_45_cast_fp16)[name = tensor("op_9582_cast_fp16")]; + tensor var_9590_begin_0 = const()[name = tensor("op_9590_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_9590_end_0 = const()[name = tensor("op_9590_end_0"), val = tensor([1, 1, 16, 32, 1])]; + tensor var_9590_end_mask_0 = const()[name = tensor("op_9590_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_9590_squeeze_mask_0 = const()[name = tensor("op_9590_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_9590_cast_fp16 = slice_by_index(begin = var_9590_begin_0, end = var_9590_end_0, end_mask = var_9590_end_mask_0, squeeze_mask = var_9590_squeeze_mask_0, x = k_complex_45_cast_fp16)[name = tensor("op_9590_cast_fp16")]; + tensor var_9598_begin_0 = const()[name = tensor("op_9598_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_9598_end_0 = const()[name = tensor("op_9598_end_0"), val = tensor([1, 1, 16, 32, 2])]; + tensor var_9598_end_mask_0 = const()[name = tensor("op_9598_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_9598_squeeze_mask_0 = const()[name = tensor("op_9598_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_9598_cast_fp16 = slice_by_index(begin = var_9598_begin_0, end = var_9598_end_0, end_mask = var_9598_end_mask_0, squeeze_mask = var_9598_squeeze_mask_0, x = k_complex_45_cast_fp16)[name = tensor("op_9598_cast_fp16")]; + tensor freqs_45_to_fp16 = const()[name = tensor("freqs_45_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(303777792)))]; + tensor var_9604_cast_fp16 = mul(x = freqs_45_to_fp16, y = ts_137_cast_fp16)[name = tensor("op_9604_cast_fp16")]; + tensor rotr_45_cast_fp16 = cos(x = var_9604_cast_fp16)[name = tensor("rotr_45_cast_fp16")]; + tensor roti_45_cast_fp16 = sin(x = var_9604_cast_fp16)[name = tensor("roti_45_cast_fp16")]; + tensor var_9608_cast_fp16 = mul(x = var_9574_cast_fp16, y = rotr_45_cast_fp16)[name = tensor("op_9608_cast_fp16")]; + tensor var_9609_cast_fp16 = mul(x = var_9582_cast_fp16, y = roti_45_cast_fp16)[name = tensor("op_9609_cast_fp16")]; + tensor qor_89_cast_fp16 = sub(x = var_9608_cast_fp16, y = var_9609_cast_fp16)[name = tensor("qor_89_cast_fp16")]; + tensor var_9612_cast_fp16 = mul(x = var_9574_cast_fp16, y = roti_45_cast_fp16)[name = tensor("op_9612_cast_fp16")]; + tensor var_9613_cast_fp16 = mul(x = var_9582_cast_fp16, y = rotr_45_cast_fp16)[name = tensor("op_9613_cast_fp16")]; + tensor qoi_89_cast_fp16 = add(x = var_9612_cast_fp16, y = var_9613_cast_fp16)[name = tensor("qoi_89_cast_fp16")]; + tensor var_9616_cast_fp16 = mul(x = var_9590_cast_fp16, y = rotr_45_cast_fp16)[name = tensor("op_9616_cast_fp16")]; + tensor var_9617_cast_fp16 = mul(x = var_9598_cast_fp16, y = roti_45_cast_fp16)[name = tensor("op_9617_cast_fp16")]; + tensor kor_89_cast_fp16 = sub(x = var_9616_cast_fp16, y = var_9617_cast_fp16)[name = tensor("kor_89_cast_fp16")]; + tensor var_9620_cast_fp16 = mul(x = var_9590_cast_fp16, y = roti_45_cast_fp16)[name = tensor("op_9620_cast_fp16")]; + tensor var_9621_cast_fp16 = mul(x = var_9598_cast_fp16, y = rotr_45_cast_fp16)[name = tensor("op_9621_cast_fp16")]; + tensor koi_89_cast_fp16 = add(x = var_9620_cast_fp16, y = var_9621_cast_fp16)[name = tensor("koi_89_cast_fp16")]; + tensor qo_45_axis_0 = const()[name = tensor("qo_45_axis_0"), val = tensor(-1)]; + tensor qo_45_cast_fp16 = stack(axis = qo_45_axis_0, values = (qor_89_cast_fp16, qoi_89_cast_fp16))[name = tensor("qo_45_cast_fp16")]; + tensor ko_45_axis_0 = const()[name = tensor("ko_45_axis_0"), val = tensor(-1)]; + tensor ko_45_cast_fp16 = stack(axis = ko_45_axis_0, values = (kor_89_cast_fp16, koi_89_cast_fp16))[name = tensor("ko_45_cast_fp16")]; + tensor var_9650 = const()[name = tensor("op_9650"), val = tensor([1, 1, 16, 64])]; + tensor q_135_cast_fp16 = reshape(shape = var_9650, x = qo_45_cast_fp16)[name = tensor("q_135_cast_fp16")]; + tensor var_9652 = const()[name = tensor("op_9652"), val = tensor([1, 1, 16, 64])]; + tensor k_91_cast_fp16 = reshape(shape = var_9652, x = ko_45_cast_fp16)[name = tensor("k_91_cast_fp16")]; + tensor _inversed_9674_y_0_to_fp16 = const()[name = tensor("_inversed_9674_y_0_to_fp16"), val = tensor(0x1p-9)]; + tensor _inversed_9674_cast_fp16 = mul(x = ts_137_cast_fp16, y = _inversed_9674_y_0_to_fp16)[name = tensor("_inversed_9674_cast_fp16")]; + tensor var_9675_cast_fp16 = floor(x = _inversed_9674_cast_fp16)[name = tensor("op_9675_cast_fp16")]; + tensor var_9676_to_fp16 = const()[name = tensor("op_9676_to_fp16"), val = tensor(0x1p+9)]; + tensor var_9677_cast_fp16 = mul(x = var_9675_cast_fp16, y = var_9676_to_fp16)[name = tensor("op_9677_cast_fp16")]; + tensor write_indices_float_91_cast_fp16 = sub(x = ts_137_cast_fp16, y = var_9677_cast_fp16)[name = tensor("write_indices_float_91_cast_fp16")]; + tensor var_9684_dtype_0 = const()[name = tensor("op_9684_dtype_0"), val = tensor("int32")]; + tensor write_indices_45_reps_0 = const()[name = tensor("write_indices_45_reps_0"), val = tensor([1, 1, 16, 64])]; + tensor write_indices_float_91_cast_fp16_to_int32 = cast(dtype = var_9684_dtype_0, x = write_indices_float_91_cast_fp16)[name = tensor("cast_442")]; + tensor write_indices_45 = tile(reps = write_indices_45_reps_0, x = write_indices_float_91_cast_fp16_to_int32)[name = tensor("write_indices_45")]; + tensor var_9692_begin_0 = const()[name = tensor("op_9692_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_9692_end_0 = const()[name = tensor("op_9692_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor var_9692_end_mask_0 = const()[name = tensor("op_9692_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_9692_squeeze_mask_0 = const()[name = tensor("op_9692_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor cache22_to_fp16_dtype_0 = const()[name = tensor("cache22_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor cache22_to_fp16 = cast(dtype = cache22_to_fp16_dtype_0, x = cache22)[name = tensor("cast_441")]; + tensor var_9692_cast_fp16 = slice_by_index(begin = var_9692_begin_0, end = var_9692_end_0, end_mask = var_9692_end_mask_0, squeeze_mask = var_9692_squeeze_mask_0, x = cache22_to_fp16)[name = tensor("op_9692_cast_fp16")]; + tensor var_9694_axis_0 = const()[name = tensor("op_9694_axis_0"), val = tensor(1)]; + tensor var_9694_mode_0 = const()[name = tensor("op_9694_mode_0"), val = tensor("update")]; + tensor var_9694_validate_indices_0 = const()[name = tensor("op_9694_validate_indices_0"), val = tensor(false)]; + tensor var_9694_cast_fp16 = scatter_along_axis(axis = var_9694_axis_0, data = var_9692_cast_fp16, indices = write_indices_45, mode = var_9694_mode_0, updates = k_91_cast_fp16, validate_indices = var_9694_validate_indices_0)[name = tensor("op_9694_cast_fp16")]; + tensor concat_156 = const()[name = tensor("concat_156"), val = tensor([0, 0, 0, 0, 0])]; + tensor concat_157 = const()[name = tensor("concat_157"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_45_internal_tensor_assign_1_stride_0 = const()[name = tensor("new_cache_45_internal_tensor_assign_1_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_45_internal_tensor_assign_1_begin_mask_0 = const()[name = tensor("new_cache_45_internal_tensor_assign_1_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_45_internal_tensor_assign_1_end_mask_0 = const()[name = tensor("new_cache_45_internal_tensor_assign_1_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_45_internal_tensor_assign_1_squeeze_mask_0 = const()[name = tensor("new_cache_45_internal_tensor_assign_1_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_92 = const()[name = tensor("shape_92"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_44 = const()[name = tensor("reduce_prod_44"), val = tensor(1048576)]; + tensor range_1d_44_start_0 = const()[name = tensor("range_1d_44_start_0"), val = tensor(0)]; + tensor range_1d_44_step_0 = const()[name = tensor("range_1d_44_step_0"), val = tensor(1)]; + tensor range_1d_44 = range_1d(end = reduce_prod_44, start = range_1d_44_start_0, step = range_1d_44_step_0)[name = tensor("range_1d_44")]; + tensor reshape_220 = reshape(shape = shape_92, x = range_1d_44)[name = tensor("reshape_220")]; + tensor slice_by_index_44 = slice_by_index(begin = concat_156, begin_mask = new_cache_45_internal_tensor_assign_1_begin_mask_0, end = concat_157, end_mask = new_cache_45_internal_tensor_assign_1_end_mask_0, squeeze_mask = new_cache_45_internal_tensor_assign_1_squeeze_mask_0, stride = new_cache_45_internal_tensor_assign_1_stride_0, x = reshape_220)[name = tensor("slice_by_index_44")]; + tensor reshape_221_shape_0 = const()[name = tensor("reshape_221_shape_0"), val = tensor([-1])]; + tensor reshape_221 = reshape(shape = reshape_221_shape_0, x = slice_by_index_44)[name = tensor("reshape_221")]; + tensor reshape_222_shape_0 = const()[name = tensor("reshape_222_shape_0"), val = tensor([-1])]; + tensor reshape_222_cast_fp16 = reshape(shape = reshape_222_shape_0, x = var_9694_cast_fp16)[name = tensor("reshape_222_cast_fp16")]; + tensor reshape_223_shape_0 = const()[name = tensor("reshape_223_shape_0"), val = tensor([-1])]; + tensor reshape_223_cast_fp16 = reshape(shape = reshape_223_shape_0, x = cache22_to_fp16)[name = tensor("reshape_223_cast_fp16")]; + tensor scatter_44_mode_0 = const()[name = tensor("scatter_44_mode_0"), val = tensor("update")]; + tensor scatter_44_axis_0 = const()[name = tensor("scatter_44_axis_0"), val = tensor(0)]; + tensor scatter_44_validate_indices_0 = const()[name = tensor("scatter_44_validate_indices_0"), val = tensor(false)]; + tensor scatter_44_cast_fp16 = scatter(axis = scatter_44_axis_0, data = reshape_223_cast_fp16, indices = reshape_221, mode = scatter_44_mode_0, updates = reshape_222_cast_fp16, validate_indices = scatter_44_validate_indices_0)[name = tensor("scatter_44_cast_fp16")]; + tensor reshape_224_cast_fp16 = reshape(shape = shape_92, x = scatter_44_cast_fp16)[name = tensor("reshape_224_cast_fp16")]; + tensor var_9702_begin_0 = const()[name = tensor("op_9702_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor var_9702_end_0 = const()[name = tensor("op_9702_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor var_9702_end_mask_0 = const()[name = tensor("op_9702_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_9702_squeeze_mask_0 = const()[name = tensor("op_9702_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor var_9702_cast_fp16 = slice_by_index(begin = var_9702_begin_0, end = var_9702_end_0, end_mask = var_9702_end_mask_0, squeeze_mask = var_9702_squeeze_mask_0, x = reshape_224_cast_fp16)[name = tensor("op_9702_cast_fp16")]; + tensor var_9704_axis_0 = const()[name = tensor("op_9704_axis_0"), val = tensor(1)]; + tensor var_9704_mode_0 = const()[name = tensor("op_9704_mode_0"), val = tensor("update")]; + tensor var_9704_validate_indices_0 = const()[name = tensor("op_9704_validate_indices_0"), val = tensor(false)]; + tensor var_9704_cast_fp16 = scatter_along_axis(axis = var_9704_axis_0, data = var_9702_cast_fp16, indices = write_indices_45, mode = var_9704_mode_0, updates = v_45_cast_fp16, validate_indices = var_9704_validate_indices_0)[name = tensor("op_9704_cast_fp16")]; + tensor concat_158 = const()[name = tensor("concat_158"), val = tensor([1, 0, 0, 0, 0])]; + tensor concat_159 = const()[name = tensor("concat_159"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_45_internal_tensor_assign_2_stride_0 = const()[name = tensor("new_cache_45_internal_tensor_assign_2_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_45_internal_tensor_assign_2_begin_mask_0 = const()[name = tensor("new_cache_45_internal_tensor_assign_2_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_45_internal_tensor_assign_2_end_mask_0 = const()[name = tensor("new_cache_45_internal_tensor_assign_2_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_45_internal_tensor_assign_2_squeeze_mask_0 = const()[name = tensor("new_cache_45_internal_tensor_assign_2_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_93 = const()[name = tensor("shape_93"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_45 = const()[name = tensor("reduce_prod_45"), val = tensor(1048576)]; + tensor range_1d_45_start_0 = const()[name = tensor("range_1d_45_start_0"), val = tensor(0)]; + tensor range_1d_45_step_0 = const()[name = tensor("range_1d_45_step_0"), val = tensor(1)]; + tensor range_1d_45 = range_1d(end = reduce_prod_45, start = range_1d_45_start_0, step = range_1d_45_step_0)[name = tensor("range_1d_45")]; + tensor reshape_225 = reshape(shape = shape_93, x = range_1d_45)[name = tensor("reshape_225")]; + tensor slice_by_index_45 = slice_by_index(begin = concat_158, begin_mask = new_cache_45_internal_tensor_assign_2_begin_mask_0, end = concat_159, end_mask = new_cache_45_internal_tensor_assign_2_end_mask_0, squeeze_mask = new_cache_45_internal_tensor_assign_2_squeeze_mask_0, stride = new_cache_45_internal_tensor_assign_2_stride_0, x = reshape_225)[name = tensor("slice_by_index_45")]; + tensor reshape_226_shape_0 = const()[name = tensor("reshape_226_shape_0"), val = tensor([-1])]; + tensor reshape_226 = reshape(shape = reshape_226_shape_0, x = slice_by_index_45)[name = tensor("reshape_226")]; + tensor reshape_227_shape_0 = const()[name = tensor("reshape_227_shape_0"), val = tensor([-1])]; + tensor reshape_227_cast_fp16 = reshape(shape = reshape_227_shape_0, x = var_9704_cast_fp16)[name = tensor("reshape_227_cast_fp16")]; + tensor reshape_228_shape_0 = const()[name = tensor("reshape_228_shape_0"), val = tensor([-1])]; + tensor reshape_228_cast_fp16 = reshape(shape = reshape_228_shape_0, x = reshape_224_cast_fp16)[name = tensor("reshape_228_cast_fp16")]; + tensor scatter_45_mode_0 = const()[name = tensor("scatter_45_mode_0"), val = tensor("update")]; + tensor scatter_45_axis_0 = const()[name = tensor("scatter_45_axis_0"), val = tensor(0)]; + tensor scatter_45_validate_indices_0 = const()[name = tensor("scatter_45_validate_indices_0"), val = tensor(false)]; + tensor scatter_45_cast_fp16 = scatter(axis = scatter_45_axis_0, data = reshape_228_cast_fp16, indices = reshape_226, mode = scatter_45_mode_0, updates = reshape_227_cast_fp16, validate_indices = scatter_45_validate_indices_0)[name = tensor("scatter_45_cast_fp16")]; + tensor reshape_229_cast_fp16 = reshape(shape = shape_93, x = scatter_45_cast_fp16)[name = tensor("reshape_229_cast_fp16")]; + tensor reshape_229_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("reshape_229_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor keys_133_begin_0 = const()[name = tensor("keys_133_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor keys_133_end_0 = const()[name = tensor("keys_133_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor keys_133_end_mask_0 = const()[name = tensor("keys_133_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor keys_133_squeeze_mask_0 = const()[name = tensor("keys_133_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor keys_133_cast_fp16 = slice_by_index(begin = keys_133_begin_0, end = keys_133_end_0, end_mask = keys_133_end_mask_0, squeeze_mask = keys_133_squeeze_mask_0, x = reshape_229_cast_fp16)[name = tensor("keys_133_cast_fp16")]; + tensor values_133_begin_0 = const()[name = tensor("values_133_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor values_133_end_0 = const()[name = tensor("values_133_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor values_133_end_mask_0 = const()[name = tensor("values_133_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor values_133_squeeze_mask_0 = const()[name = tensor("values_133_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor values_133_cast_fp16 = slice_by_index(begin = values_133_begin_0, end = values_133_end_0, end_mask = values_133_end_mask_0, squeeze_mask = values_133_squeeze_mask_0, x = reshape_229_cast_fp16)[name = tensor("values_133_cast_fp16")]; + tensor var_9716_cast_fp16 = not_equal(x = keys_133_cast_fp16, y = keys_133_cast_fp16)[name = tensor("op_9716_cast_fp16")]; + tensor keys_135_cast_fp16 = select(a = var_504_to_fp16, b = keys_133_cast_fp16, cond = var_9716_cast_fp16)[name = tensor("keys_135_cast_fp16")]; + tensor var_9724_cast_fp16 = not_equal(x = values_133_cast_fp16, y = values_133_cast_fp16)[name = tensor("op_9724_cast_fp16")]; + tensor values_135_cast_fp16 = select(a = var_504_to_fp16, b = values_133_cast_fp16, cond = var_9724_cast_fp16)[name = tensor("values_135_cast_fp16")]; + tensor var_9748 = const()[name = tensor("op_9748"), val = tensor([0, 2, 1, 3])]; + tensor var_9761 = const()[name = tensor("op_9761"), val = tensor([1, 1, 1])]; + tensor var_9762_cast_fp16 = reshape(shape = var_9761, x = position22_to_fp16)[name = tensor("op_9762_cast_fp16")]; + tensor var_9779_to_fp16 = const()[name = tensor("op_9779_to_fp16"), val = tensor(0x1p+0)]; + tensor valid_len_45_cast_fp16 = add(x = var_9762_cast_fp16, y = var_9779_to_fp16)[name = tensor("valid_len_45_cast_fp16")]; + tensor valid_mask_45_cast_fp16 = less(x = k_positions_1_promoted_to_fp16, y = valid_len_45_cast_fp16)[name = tensor("valid_mask_45_cast_fp16")]; + tensor causal_mask_45_cast_fp16 = less_equal(x = k_positions_1_promoted_to_fp16, y = var_9762_cast_fp16)[name = tensor("causal_mask_45_cast_fp16")]; + tensor attn_mask_89 = logical_and(x = valid_mask_45_cast_fp16, y = causal_mask_45_cast_fp16)[name = tensor("attn_mask_89")]; + tensor attn_mask_91_axes_0 = const()[name = tensor("attn_mask_91_axes_0"), val = tensor([1])]; + tensor attn_mask_91 = expand_dims(axes = attn_mask_91_axes_0, x = attn_mask_89)[name = tensor("attn_mask_91")]; + tensor var_9797_transpose_x_0 = const()[name = tensor("op_9797_transpose_x_0"), val = tensor(false)]; + tensor var_9797_transpose_y_0 = const()[name = tensor("op_9797_transpose_y_0"), val = tensor(false)]; + tensor transpose_116_perm_0 = const()[name = tensor("transpose_116_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_117_perm_0 = const()[name = tensor("transpose_117_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_117 = transpose(perm = transpose_117_perm_0, x = keys_135_cast_fp16)[name = tensor("transpose_125")]; + tensor transpose_116 = transpose(perm = transpose_116_perm_0, x = q_135_cast_fp16)[name = tensor("transpose_126")]; + tensor var_9797_cast_fp16 = matmul(transpose_x = var_9797_transpose_x_0, transpose_y = var_9797_transpose_y_0, x = transpose_116, y = transpose_117)[name = tensor("op_9797_cast_fp16")]; + tensor var_9791_to_fp16 = const()[name = tensor("op_9791_to_fp16"), val = tensor([0x1p-3])]; + tensor attn_weights_133_cast_fp16 = mul(x = var_9797_cast_fp16, y = var_9791_to_fp16)[name = tensor("attn_weights_133_cast_fp16")]; + tensor var_9799 = logical_not(x = attn_mask_91)[name = tensor("op_9799")]; + tensor var_9800_to_fp16 = const()[name = tensor("op_9800_to_fp16"), val = tensor(-inf)]; + tensor attn_weights_135_cast_fp16 = select(a = var_9800_to_fp16, b = attn_weights_133_cast_fp16, cond = var_9799)[name = tensor("attn_weights_135_cast_fp16")]; + tensor var_9802 = const()[name = tensor("op_9802"), val = tensor(-1)]; + tensor attn_weights_137_cast_fp16 = softmax(axis = var_9802, x = attn_weights_135_cast_fp16)[name = tensor("attn_weights_137_cast_fp16")]; + tensor attn_output_45_transpose_x_0 = const()[name = tensor("attn_output_45_transpose_x_0"), val = tensor(false)]; + tensor attn_output_45_transpose_y_0 = const()[name = tensor("attn_output_45_transpose_y_0"), val = tensor(false)]; + tensor values_137_cast_fp16 = transpose(perm = var_9748, x = values_135_cast_fp16)[name = tensor("transpose_127")]; + tensor attn_output_45_cast_fp16 = matmul(transpose_x = attn_output_45_transpose_x_0, transpose_y = attn_output_45_transpose_y_0, x = attn_weights_137_cast_fp16, y = values_137_cast_fp16)[name = tensor("attn_output_45_cast_fp16")]; + tensor var_9810 = const()[name = tensor("op_9810"), val = tensor([0, 2, 1, 3])]; + tensor var_9813 = const()[name = tensor("op_9813"), val = tensor([1, 1, 1024])]; + tensor var_9811_cast_fp16 = transpose(perm = var_9810, x = attn_output_45_cast_fp16)[name = tensor("transpose_124")]; + tensor input_225_cast_fp16 = reshape(shape = var_9813, x = var_9811_cast_fp16)[name = tensor("input_225_cast_fp16")]; + tensor linear_90_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = attn22_out_proj_weight_cast_fp16, x = input_225_cast_fp16)[name = tensor("linear_90_cast_fp16")]; + tensor var_9819_to_fp16 = const()[name = tensor("op_9819_to_fp16"), val = tensor(0x1p+0)]; + tensor var_9820_cast_fp16 = add(x = position22_to_fp16, y = var_9819_to_fp16)[name = tensor("op_9820_cast_fp16")]; + tensor var_9820_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_9820_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor input_227_cast_fp16 = add(x = input_223_cast_fp16, y = linear_90_cast_fp16)[name = tensor("input_227_cast_fp16")]; + tensor input_229_axes_0 = const()[name = tensor("input_229_axes_0"), val = tensor([-1])]; + tensor norm22_2_weight_to_fp16 = const()[name = tensor("norm22_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(303777920)))]; + tensor norm22_2_bias_to_fp16 = const()[name = tensor("norm22_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(303780032)))]; + tensor var_9824_to_fp16 = const()[name = tensor("op_9824_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_229_cast_fp16 = layer_norm(axes = input_229_axes_0, beta = norm22_2_bias_to_fp16, epsilon = var_9824_to_fp16, gamma = norm22_2_weight_to_fp16, x = input_227_cast_fp16)[name = tensor("input_229_cast_fp16")]; + tensor linear_91_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = linear22_1_weight_cast_fp16, x = input_229_cast_fp16)[name = tensor("linear_91_cast_fp16")]; + tensor input_231_mode_0 = const()[name = tensor("input_231_mode_0"), val = tensor("EXACT")]; + tensor input_231_cast_fp16 = gelu(mode = input_231_mode_0, x = linear_91_cast_fp16)[name = tensor("input_231_cast_fp16")]; + tensor linear_92_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = linear22_2_weight_cast_fp16, x = input_231_cast_fp16)[name = tensor("linear_92_cast_fp16")]; + tensor input_233_cast_fp16 = add(x = input_227_cast_fp16, y = linear_92_cast_fp16)[name = tensor("input_233_cast_fp16")]; + tensor x_axes_0 = const()[name = tensor("x_axes_0"), val = tensor([-1])]; + tensor norm23_1_weight_to_fp16 = const()[name = tensor("norm23_1_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(303782144)))]; + tensor norm23_1_bias_to_fp16 = const()[name = tensor("norm23_1_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(303784256)))]; + tensor var_9841_to_fp16 = const()[name = tensor("op_9841_to_fp16"), val = tensor(0x1.5p-17)]; + tensor x_cast_fp16 = layer_norm(axes = x_axes_0, beta = norm23_1_bias_to_fp16, epsilon = var_9841_to_fp16, gamma = norm23_1_weight_to_fp16, x = input_233_cast_fp16)[name = tensor("x_cast_fp16")]; + tensor linear_93_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = attn23_in_proj_weight_cast_fp16, x = x_cast_fp16)[name = tensor("linear_93_cast_fp16")]; + tensor var_9877 = const()[name = tensor("op_9877"), val = tensor([1, 1, 3, 16, 64])]; + tensor qkv_cast_fp16 = reshape(shape = var_9877, x = linear_93_cast_fp16)[name = tensor("qkv_cast_fp16")]; + tensor q_139_begin_0 = const()[name = tensor("q_139_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor q_139_end_0 = const()[name = tensor("q_139_end_0"), val = tensor([1, 1, 1, 16, 64])]; + tensor q_139_end_mask_0 = const()[name = tensor("q_139_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor q_139_squeeze_mask_0 = const()[name = tensor("q_139_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor q_139_cast_fp16 = slice_by_index(begin = q_139_begin_0, end = q_139_end_0, end_mask = q_139_end_mask_0, squeeze_mask = q_139_squeeze_mask_0, x = qkv_cast_fp16)[name = tensor("q_139_cast_fp16")]; + tensor k_93_begin_0 = const()[name = tensor("k_93_begin_0"), val = tensor([0, 0, 1, 0, 0])]; + tensor k_93_end_0 = const()[name = tensor("k_93_end_0"), val = tensor([1, 1, 2, 16, 64])]; + tensor k_93_end_mask_0 = const()[name = tensor("k_93_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor k_93_squeeze_mask_0 = const()[name = tensor("k_93_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor k_93_cast_fp16 = slice_by_index(begin = k_93_begin_0, end = k_93_end_0, end_mask = k_93_end_mask_0, squeeze_mask = k_93_squeeze_mask_0, x = qkv_cast_fp16)[name = tensor("k_93_cast_fp16")]; + tensor v_begin_0 = const()[name = tensor("v_begin_0"), val = tensor([0, 0, 2, 0, 0])]; + tensor v_end_0 = const()[name = tensor("v_end_0"), val = tensor([1, 1, 3, 16, 64])]; + tensor v_end_mask_0 = const()[name = tensor("v_end_mask_0"), val = tensor([true, true, false, true, true])]; + tensor v_squeeze_mask_0 = const()[name = tensor("v_squeeze_mask_0"), val = tensor([false, false, true, false, false])]; + tensor v_cast_fp16 = slice_by_index(begin = v_begin_0, end = v_end_0, end_mask = v_end_mask_0, squeeze_mask = v_squeeze_mask_0, x = qkv_cast_fp16)[name = tensor("v_cast_fp16")]; + tensor var_9981 = const()[name = tensor("op_9981"), val = tensor([1, 1, 1, 1])]; + tensor position23_to_fp16_dtype_0 = const()[name = tensor("position23_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor position23_to_fp16 = cast(dtype = position23_to_fp16_dtype_0, x = position23)[name = tensor("cast_438")]; + tensor ts_cast_fp16 = reshape(shape = var_9981, x = position23_to_fp16)[name = tensor("ts_cast_fp16")]; + tensor var_9985 = const()[name = tensor("op_9985"), val = tensor([1, 1, 16, 32, 2])]; + tensor q_complex_cast_fp16 = reshape(shape = var_9985, x = q_139_cast_fp16)[name = tensor("q_complex_cast_fp16")]; + tensor var_9989 = const()[name = tensor("op_9989"), val = tensor([1, 1, 16, 32, 2])]; + tensor k_complex_cast_fp16 = reshape(shape = var_9989, x = k_93_cast_fp16)[name = tensor("k_complex_cast_fp16")]; + tensor var_9993_begin_0 = const()[name = tensor("op_9993_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_9993_end_0 = const()[name = tensor("op_9993_end_0"), val = tensor([1, 1, 16, 32, 1])]; + tensor var_9993_end_mask_0 = const()[name = tensor("op_9993_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_9993_squeeze_mask_0 = const()[name = tensor("op_9993_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_9993_cast_fp16 = slice_by_index(begin = var_9993_begin_0, end = var_9993_end_0, end_mask = var_9993_end_mask_0, squeeze_mask = var_9993_squeeze_mask_0, x = q_complex_cast_fp16)[name = tensor("op_9993_cast_fp16")]; + tensor var_10001_begin_0 = const()[name = tensor("op_10001_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_10001_end_0 = const()[name = tensor("op_10001_end_0"), val = tensor([1, 1, 16, 32, 2])]; + tensor var_10001_end_mask_0 = const()[name = tensor("op_10001_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_10001_squeeze_mask_0 = const()[name = tensor("op_10001_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_10001_cast_fp16 = slice_by_index(begin = var_10001_begin_0, end = var_10001_end_0, end_mask = var_10001_end_mask_0, squeeze_mask = var_10001_squeeze_mask_0, x = q_complex_cast_fp16)[name = tensor("op_10001_cast_fp16")]; + tensor var_10009_begin_0 = const()[name = tensor("op_10009_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_10009_end_0 = const()[name = tensor("op_10009_end_0"), val = tensor([1, 1, 16, 32, 1])]; + tensor var_10009_end_mask_0 = const()[name = tensor("op_10009_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_10009_squeeze_mask_0 = const()[name = tensor("op_10009_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_10009_cast_fp16 = slice_by_index(begin = var_10009_begin_0, end = var_10009_end_0, end_mask = var_10009_end_mask_0, squeeze_mask = var_10009_squeeze_mask_0, x = k_complex_cast_fp16)[name = tensor("op_10009_cast_fp16")]; + tensor var_10017_begin_0 = const()[name = tensor("op_10017_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_10017_end_0 = const()[name = tensor("op_10017_end_0"), val = tensor([1, 1, 16, 32, 2])]; + tensor var_10017_end_mask_0 = const()[name = tensor("op_10017_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_10017_squeeze_mask_0 = const()[name = tensor("op_10017_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_10017_cast_fp16 = slice_by_index(begin = var_10017_begin_0, end = var_10017_end_0, end_mask = var_10017_end_mask_0, squeeze_mask = var_10017_squeeze_mask_0, x = k_complex_cast_fp16)[name = tensor("op_10017_cast_fp16")]; + tensor freqs_to_fp16 = const()[name = tensor("freqs_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(303786368)))]; + tensor var_10023_cast_fp16 = mul(x = freqs_to_fp16, y = ts_cast_fp16)[name = tensor("op_10023_cast_fp16")]; + tensor rotr_cast_fp16 = cos(x = var_10023_cast_fp16)[name = tensor("rotr_cast_fp16")]; + tensor roti_cast_fp16 = sin(x = var_10023_cast_fp16)[name = tensor("roti_cast_fp16")]; + tensor var_10027_cast_fp16 = mul(x = var_9993_cast_fp16, y = rotr_cast_fp16)[name = tensor("op_10027_cast_fp16")]; + tensor var_10028_cast_fp16 = mul(x = var_10001_cast_fp16, y = roti_cast_fp16)[name = tensor("op_10028_cast_fp16")]; + tensor qor_93_cast_fp16 = sub(x = var_10027_cast_fp16, y = var_10028_cast_fp16)[name = tensor("qor_93_cast_fp16")]; + tensor var_10031_cast_fp16 = mul(x = var_9993_cast_fp16, y = roti_cast_fp16)[name = tensor("op_10031_cast_fp16")]; + tensor var_10032_cast_fp16 = mul(x = var_10001_cast_fp16, y = rotr_cast_fp16)[name = tensor("op_10032_cast_fp16")]; + tensor qoi_93_cast_fp16 = add(x = var_10031_cast_fp16, y = var_10032_cast_fp16)[name = tensor("qoi_93_cast_fp16")]; + tensor var_10035_cast_fp16 = mul(x = var_10009_cast_fp16, y = rotr_cast_fp16)[name = tensor("op_10035_cast_fp16")]; + tensor var_10036_cast_fp16 = mul(x = var_10017_cast_fp16, y = roti_cast_fp16)[name = tensor("op_10036_cast_fp16")]; + tensor kor_93_cast_fp16 = sub(x = var_10035_cast_fp16, y = var_10036_cast_fp16)[name = tensor("kor_93_cast_fp16")]; + tensor var_10039_cast_fp16 = mul(x = var_10009_cast_fp16, y = roti_cast_fp16)[name = tensor("op_10039_cast_fp16")]; + tensor var_10040_cast_fp16 = mul(x = var_10017_cast_fp16, y = rotr_cast_fp16)[name = tensor("op_10040_cast_fp16")]; + tensor koi_93_cast_fp16 = add(x = var_10039_cast_fp16, y = var_10040_cast_fp16)[name = tensor("koi_93_cast_fp16")]; + tensor qo_axis_0 = const()[name = tensor("qo_axis_0"), val = tensor(-1)]; + tensor qo_cast_fp16 = stack(axis = qo_axis_0, values = (qor_93_cast_fp16, qoi_93_cast_fp16))[name = tensor("qo_cast_fp16")]; + tensor ko_axis_0 = const()[name = tensor("ko_axis_0"), val = tensor(-1)]; + tensor ko_cast_fp16 = stack(axis = ko_axis_0, values = (kor_93_cast_fp16, koi_93_cast_fp16))[name = tensor("ko_cast_fp16")]; + tensor var_10069 = const()[name = tensor("op_10069"), val = tensor([1, 1, 16, 64])]; + tensor q_141_cast_fp16 = reshape(shape = var_10069, x = qo_cast_fp16)[name = tensor("q_141_cast_fp16")]; + tensor var_10071 = const()[name = tensor("op_10071"), val = tensor([1, 1, 16, 64])]; + tensor k_cast_fp16 = reshape(shape = var_10071, x = ko_cast_fp16)[name = tensor("k_cast_fp16")]; + tensor _inversed_10093_y_0_to_fp16 = const()[name = tensor("_inversed_10093_y_0_to_fp16"), val = tensor(0x1p-9)]; + tensor _inversed_10093_cast_fp16 = mul(x = ts_cast_fp16, y = _inversed_10093_y_0_to_fp16)[name = tensor("_inversed_10093_cast_fp16")]; + tensor var_10094_cast_fp16 = floor(x = _inversed_10093_cast_fp16)[name = tensor("op_10094_cast_fp16")]; + tensor var_10095_to_fp16 = const()[name = tensor("op_10095_to_fp16"), val = tensor(0x1p+9)]; + tensor var_10096_cast_fp16 = mul(x = var_10094_cast_fp16, y = var_10095_to_fp16)[name = tensor("op_10096_cast_fp16")]; + tensor write_indices_float_cast_fp16 = sub(x = ts_cast_fp16, y = var_10096_cast_fp16)[name = tensor("write_indices_float_cast_fp16")]; + tensor var_10103_dtype_0 = const()[name = tensor("op_10103_dtype_0"), val = tensor("int32")]; + tensor write_indices_reps_0 = const()[name = tensor("write_indices_reps_0"), val = tensor([1, 1, 16, 64])]; + tensor write_indices_float_cast_fp16_to_int32 = cast(dtype = var_10103_dtype_0, x = write_indices_float_cast_fp16)[name = tensor("cast_437")]; + tensor write_indices = tile(reps = write_indices_reps_0, x = write_indices_float_cast_fp16_to_int32)[name = tensor("write_indices")]; + tensor var_10111_begin_0 = const()[name = tensor("op_10111_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_10111_end_0 = const()[name = tensor("op_10111_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor var_10111_end_mask_0 = const()[name = tensor("op_10111_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_10111_squeeze_mask_0 = const()[name = tensor("op_10111_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor cache23_to_fp16_dtype_0 = const()[name = tensor("cache23_to_fp16_dtype_0"), val = tensor("fp16")]; + tensor cache23_to_fp16 = cast(dtype = cache23_to_fp16_dtype_0, x = cache23)[name = tensor("cast_436")]; + tensor var_10111_cast_fp16 = slice_by_index(begin = var_10111_begin_0, end = var_10111_end_0, end_mask = var_10111_end_mask_0, squeeze_mask = var_10111_squeeze_mask_0, x = cache23_to_fp16)[name = tensor("op_10111_cast_fp16")]; + tensor var_10113_axis_0 = const()[name = tensor("op_10113_axis_0"), val = tensor(1)]; + tensor var_10113_mode_0 = const()[name = tensor("op_10113_mode_0"), val = tensor("update")]; + tensor var_10113_validate_indices_0 = const()[name = tensor("op_10113_validate_indices_0"), val = tensor(false)]; + tensor var_10113_cast_fp16 = scatter_along_axis(axis = var_10113_axis_0, data = var_10111_cast_fp16, indices = write_indices, mode = var_10113_mode_0, updates = k_cast_fp16, validate_indices = var_10113_validate_indices_0)[name = tensor("op_10113_cast_fp16")]; + tensor concat_163 = const()[name = tensor("concat_163"), val = tensor([0, 0, 0, 0, 0])]; + tensor concat_164 = const()[name = tensor("concat_164"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_internal_tensor_assign_1_stride_0 = const()[name = tensor("new_cache_internal_tensor_assign_1_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_internal_tensor_assign_1_begin_mask_0 = const()[name = tensor("new_cache_internal_tensor_assign_1_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_internal_tensor_assign_1_end_mask_0 = const()[name = tensor("new_cache_internal_tensor_assign_1_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_internal_tensor_assign_1_squeeze_mask_0 = const()[name = tensor("new_cache_internal_tensor_assign_1_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_94 = const()[name = tensor("shape_94"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_46 = const()[name = tensor("reduce_prod_46"), val = tensor(1048576)]; + tensor range_1d_46_start_0 = const()[name = tensor("range_1d_46_start_0"), val = tensor(0)]; + tensor range_1d_46_step_0 = const()[name = tensor("range_1d_46_step_0"), val = tensor(1)]; + tensor range_1d_46 = range_1d(end = reduce_prod_46, start = range_1d_46_start_0, step = range_1d_46_step_0)[name = tensor("range_1d_46")]; + tensor reshape_230 = reshape(shape = shape_94, x = range_1d_46)[name = tensor("reshape_230")]; + tensor slice_by_index_46 = slice_by_index(begin = concat_163, begin_mask = new_cache_internal_tensor_assign_1_begin_mask_0, end = concat_164, end_mask = new_cache_internal_tensor_assign_1_end_mask_0, squeeze_mask = new_cache_internal_tensor_assign_1_squeeze_mask_0, stride = new_cache_internal_tensor_assign_1_stride_0, x = reshape_230)[name = tensor("slice_by_index_46")]; + tensor reshape_231_shape_0 = const()[name = tensor("reshape_231_shape_0"), val = tensor([-1])]; + tensor reshape_231 = reshape(shape = reshape_231_shape_0, x = slice_by_index_46)[name = tensor("reshape_231")]; + tensor reshape_232_shape_0 = const()[name = tensor("reshape_232_shape_0"), val = tensor([-1])]; + tensor reshape_232_cast_fp16 = reshape(shape = reshape_232_shape_0, x = var_10113_cast_fp16)[name = tensor("reshape_232_cast_fp16")]; + tensor reshape_233_shape_0 = const()[name = tensor("reshape_233_shape_0"), val = tensor([-1])]; + tensor reshape_233_cast_fp16 = reshape(shape = reshape_233_shape_0, x = cache23_to_fp16)[name = tensor("reshape_233_cast_fp16")]; + tensor scatter_46_mode_0 = const()[name = tensor("scatter_46_mode_0"), val = tensor("update")]; + tensor scatter_46_axis_0 = const()[name = tensor("scatter_46_axis_0"), val = tensor(0)]; + tensor scatter_46_validate_indices_0 = const()[name = tensor("scatter_46_validate_indices_0"), val = tensor(false)]; + tensor scatter_46_cast_fp16 = scatter(axis = scatter_46_axis_0, data = reshape_233_cast_fp16, indices = reshape_231, mode = scatter_46_mode_0, updates = reshape_232_cast_fp16, validate_indices = scatter_46_validate_indices_0)[name = tensor("scatter_46_cast_fp16")]; + tensor reshape_234_cast_fp16 = reshape(shape = shape_94, x = scatter_46_cast_fp16)[name = tensor("reshape_234_cast_fp16")]; + tensor var_10121_begin_0 = const()[name = tensor("op_10121_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor var_10121_end_0 = const()[name = tensor("op_10121_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor var_10121_end_mask_0 = const()[name = tensor("op_10121_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_10121_squeeze_mask_0 = const()[name = tensor("op_10121_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor var_10121_cast_fp16 = slice_by_index(begin = var_10121_begin_0, end = var_10121_end_0, end_mask = var_10121_end_mask_0, squeeze_mask = var_10121_squeeze_mask_0, x = reshape_234_cast_fp16)[name = tensor("op_10121_cast_fp16")]; + tensor var_10123_axis_0 = const()[name = tensor("op_10123_axis_0"), val = tensor(1)]; + tensor var_10123_mode_0 = const()[name = tensor("op_10123_mode_0"), val = tensor("update")]; + tensor var_10123_validate_indices_0 = const()[name = tensor("op_10123_validate_indices_0"), val = tensor(false)]; + tensor var_10123_cast_fp16 = scatter_along_axis(axis = var_10123_axis_0, data = var_10121_cast_fp16, indices = write_indices, mode = var_10123_mode_0, updates = v_cast_fp16, validate_indices = var_10123_validate_indices_0)[name = tensor("op_10123_cast_fp16")]; + tensor concat_165 = const()[name = tensor("concat_165"), val = tensor([1, 0, 0, 0, 0])]; + tensor concat_166 = const()[name = tensor("concat_166"), val = tensor([0, 0, 0, 0, 0])]; + tensor new_cache_internal_tensor_assign_2_stride_0 = const()[name = tensor("new_cache_internal_tensor_assign_2_stride_0"), val = tensor([1, 1, 1, 1, 1])]; + tensor new_cache_internal_tensor_assign_2_begin_mask_0 = const()[name = tensor("new_cache_internal_tensor_assign_2_begin_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_internal_tensor_assign_2_end_mask_0 = const()[name = tensor("new_cache_internal_tensor_assign_2_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor new_cache_internal_tensor_assign_2_squeeze_mask_0 = const()[name = tensor("new_cache_internal_tensor_assign_2_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor shape_95 = const()[name = tensor("shape_95"), val = tensor([2, 1, 512, 16, 64])]; + tensor reduce_prod_47 = const()[name = tensor("reduce_prod_47"), val = tensor(1048576)]; + tensor range_1d_47_start_0 = const()[name = tensor("range_1d_47_start_0"), val = tensor(0)]; + tensor range_1d_47_step_0 = const()[name = tensor("range_1d_47_step_0"), val = tensor(1)]; + tensor range_1d_47 = range_1d(end = reduce_prod_47, start = range_1d_47_start_0, step = range_1d_47_step_0)[name = tensor("range_1d_47")]; + tensor reshape_235 = reshape(shape = shape_95, x = range_1d_47)[name = tensor("reshape_235")]; + tensor slice_by_index_47 = slice_by_index(begin = concat_165, begin_mask = new_cache_internal_tensor_assign_2_begin_mask_0, end = concat_166, end_mask = new_cache_internal_tensor_assign_2_end_mask_0, squeeze_mask = new_cache_internal_tensor_assign_2_squeeze_mask_0, stride = new_cache_internal_tensor_assign_2_stride_0, x = reshape_235)[name = tensor("slice_by_index_47")]; + tensor reshape_236_shape_0 = const()[name = tensor("reshape_236_shape_0"), val = tensor([-1])]; + tensor reshape_236 = reshape(shape = reshape_236_shape_0, x = slice_by_index_47)[name = tensor("reshape_236")]; + tensor reshape_237_shape_0 = const()[name = tensor("reshape_237_shape_0"), val = tensor([-1])]; + tensor reshape_237_cast_fp16 = reshape(shape = reshape_237_shape_0, x = var_10123_cast_fp16)[name = tensor("reshape_237_cast_fp16")]; + tensor reshape_238_shape_0 = const()[name = tensor("reshape_238_shape_0"), val = tensor([-1])]; + tensor reshape_238_cast_fp16 = reshape(shape = reshape_238_shape_0, x = reshape_234_cast_fp16)[name = tensor("reshape_238_cast_fp16")]; + tensor scatter_47_mode_0 = const()[name = tensor("scatter_47_mode_0"), val = tensor("update")]; + tensor scatter_47_axis_0 = const()[name = tensor("scatter_47_axis_0"), val = tensor(0)]; + tensor scatter_47_validate_indices_0 = const()[name = tensor("scatter_47_validate_indices_0"), val = tensor(false)]; + tensor scatter_47_cast_fp16 = scatter(axis = scatter_47_axis_0, data = reshape_238_cast_fp16, indices = reshape_236, mode = scatter_47_mode_0, updates = reshape_237_cast_fp16, validate_indices = scatter_47_validate_indices_0)[name = tensor("scatter_47_cast_fp16")]; + tensor reshape_239_cast_fp16 = reshape(shape = shape_95, x = scatter_47_cast_fp16)[name = tensor("reshape_239_cast_fp16")]; + tensor reshape_239_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("reshape_239_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor keys_139_begin_0 = const()[name = tensor("keys_139_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor keys_139_end_0 = const()[name = tensor("keys_139_end_0"), val = tensor([1, 1, 512, 16, 64])]; + tensor keys_139_end_mask_0 = const()[name = tensor("keys_139_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor keys_139_squeeze_mask_0 = const()[name = tensor("keys_139_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor keys_139_cast_fp16 = slice_by_index(begin = keys_139_begin_0, end = keys_139_end_0, end_mask = keys_139_end_mask_0, squeeze_mask = keys_139_squeeze_mask_0, x = reshape_239_cast_fp16)[name = tensor("keys_139_cast_fp16")]; + tensor values_139_begin_0 = const()[name = tensor("values_139_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor values_139_end_0 = const()[name = tensor("values_139_end_0"), val = tensor([2, 1, 512, 16, 64])]; + tensor values_139_end_mask_0 = const()[name = tensor("values_139_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor values_139_squeeze_mask_0 = const()[name = tensor("values_139_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor values_139_cast_fp16 = slice_by_index(begin = values_139_begin_0, end = values_139_end_0, end_mask = values_139_end_mask_0, squeeze_mask = values_139_squeeze_mask_0, x = reshape_239_cast_fp16)[name = tensor("values_139_cast_fp16")]; + tensor var_10135_cast_fp16 = not_equal(x = keys_139_cast_fp16, y = keys_139_cast_fp16)[name = tensor("op_10135_cast_fp16")]; + tensor keys_141_cast_fp16 = select(a = var_504_to_fp16, b = keys_139_cast_fp16, cond = var_10135_cast_fp16)[name = tensor("keys_141_cast_fp16")]; + tensor var_10143_cast_fp16 = not_equal(x = values_139_cast_fp16, y = values_139_cast_fp16)[name = tensor("op_10143_cast_fp16")]; + tensor values_141_cast_fp16 = select(a = var_504_to_fp16, b = values_139_cast_fp16, cond = var_10143_cast_fp16)[name = tensor("values_141_cast_fp16")]; + tensor var_10167 = const()[name = tensor("op_10167"), val = tensor([0, 2, 1, 3])]; + tensor var_10180 = const()[name = tensor("op_10180"), val = tensor([1, 1, 1])]; + tensor var_10181_cast_fp16 = reshape(shape = var_10180, x = position23_to_fp16)[name = tensor("op_10181_cast_fp16")]; + tensor var_10198_to_fp16 = const()[name = tensor("op_10198_to_fp16"), val = tensor(0x1p+0)]; + tensor valid_len_cast_fp16 = add(x = var_10181_cast_fp16, y = var_10198_to_fp16)[name = tensor("valid_len_cast_fp16")]; + tensor valid_mask_cast_fp16 = less(x = k_positions_1_promoted_to_fp16, y = valid_len_cast_fp16)[name = tensor("valid_mask_cast_fp16")]; + tensor causal_mask_cast_fp16 = less_equal(x = k_positions_1_promoted_to_fp16, y = var_10181_cast_fp16)[name = tensor("causal_mask_cast_fp16")]; + tensor attn_mask_93 = logical_and(x = valid_mask_cast_fp16, y = causal_mask_cast_fp16)[name = tensor("attn_mask_93")]; + tensor attn_mask_axes_0 = const()[name = tensor("attn_mask_axes_0"), val = tensor([1])]; + tensor attn_mask = expand_dims(axes = attn_mask_axes_0, x = attn_mask_93)[name = tensor("attn_mask")]; + tensor var_10216_transpose_x_0 = const()[name = tensor("op_10216_transpose_x_0"), val = tensor(false)]; + tensor var_10216_transpose_y_0 = const()[name = tensor("op_10216_transpose_y_0"), val = tensor(false)]; + tensor transpose_118_perm_0 = const()[name = tensor("transpose_118_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_119_perm_0 = const()[name = tensor("transpose_119_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_119 = transpose(perm = transpose_119_perm_0, x = keys_141_cast_fp16)[name = tensor("transpose_121")]; + tensor transpose_118 = transpose(perm = transpose_118_perm_0, x = q_141_cast_fp16)[name = tensor("transpose_122")]; + tensor var_10216_cast_fp16 = matmul(transpose_x = var_10216_transpose_x_0, transpose_y = var_10216_transpose_y_0, x = transpose_118, y = transpose_119)[name = tensor("op_10216_cast_fp16")]; + tensor var_10210_to_fp16 = const()[name = tensor("op_10210_to_fp16"), val = tensor([0x1p-3])]; + tensor attn_weights_139_cast_fp16 = mul(x = var_10216_cast_fp16, y = var_10210_to_fp16)[name = tensor("attn_weights_139_cast_fp16")]; + tensor var_10218 = logical_not(x = attn_mask)[name = tensor("op_10218")]; + tensor var_10219_to_fp16 = const()[name = tensor("op_10219_to_fp16"), val = tensor(-inf)]; + tensor attn_weights_141_cast_fp16 = select(a = var_10219_to_fp16, b = attn_weights_139_cast_fp16, cond = var_10218)[name = tensor("attn_weights_141_cast_fp16")]; + tensor var_10221 = const()[name = tensor("op_10221"), val = tensor(-1)]; + tensor attn_weights_cast_fp16 = softmax(axis = var_10221, x = attn_weights_141_cast_fp16)[name = tensor("attn_weights_cast_fp16")]; + tensor attn_output_transpose_x_0 = const()[name = tensor("attn_output_transpose_x_0"), val = tensor(false)]; + tensor attn_output_transpose_y_0 = const()[name = tensor("attn_output_transpose_y_0"), val = tensor(false)]; + tensor values_cast_fp16 = transpose(perm = var_10167, x = values_141_cast_fp16)[name = tensor("transpose_123")]; + tensor attn_output_cast_fp16 = matmul(transpose_x = attn_output_transpose_x_0, transpose_y = attn_output_transpose_y_0, x = attn_weights_cast_fp16, y = values_cast_fp16)[name = tensor("attn_output_cast_fp16")]; + tensor var_10229 = const()[name = tensor("op_10229"), val = tensor([0, 2, 1, 3])]; + tensor var_10232 = const()[name = tensor("op_10232"), val = tensor([1, 1, 1024])]; + tensor var_10230_cast_fp16 = transpose(perm = var_10229, x = attn_output_cast_fp16)[name = tensor("transpose_120")]; + tensor input_235_cast_fp16 = reshape(shape = var_10232, x = var_10230_cast_fp16)[name = tensor("input_235_cast_fp16")]; + tensor linear_94_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = attn23_out_proj_weight_cast_fp16, x = input_235_cast_fp16)[name = tensor("linear_94_cast_fp16")]; + tensor var_10238_to_fp16 = const()[name = tensor("op_10238_to_fp16"), val = tensor(0x1p+0)]; + tensor var_10239_cast_fp16 = add(x = position23_to_fp16, y = var_10238_to_fp16)[name = tensor("op_10239_cast_fp16")]; + tensor var_10239_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("op_10239_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor input_237_cast_fp16 = add(x = input_233_cast_fp16, y = linear_94_cast_fp16)[name = tensor("input_237_cast_fp16")]; + tensor input_239_axes_0 = const()[name = tensor("input_239_axes_0"), val = tensor([-1])]; + tensor norm23_2_weight_to_fp16 = const()[name = tensor("norm23_2_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(303786496)))]; + tensor norm23_2_bias_to_fp16 = const()[name = tensor("norm23_2_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(303788608)))]; + tensor var_10243_to_fp16 = const()[name = tensor("op_10243_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_239_cast_fp16 = layer_norm(axes = input_239_axes_0, beta = norm23_2_bias_to_fp16, epsilon = var_10243_to_fp16, gamma = norm23_2_weight_to_fp16, x = input_237_cast_fp16)[name = tensor("input_239_cast_fp16")]; + tensor linear_95_cast_fp16 = linear(bias = linear_3_bias_0_to_fp16, weight = linear23_1_weight_cast_fp16, x = input_239_cast_fp16)[name = tensor("linear_95_cast_fp16")]; + tensor input_241_mode_0 = const()[name = tensor("input_241_mode_0"), val = tensor("EXACT")]; + tensor input_241_cast_fp16 = gelu(mode = input_241_mode_0, x = linear_95_cast_fp16)[name = tensor("input_241_cast_fp16")]; + tensor linear_96_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = linear23_2_weight_cast_fp16, x = input_241_cast_fp16)[name = tensor("linear_96_cast_fp16")]; + tensor input_243_cast_fp16 = add(x = input_237_cast_fp16, y = linear_96_cast_fp16)[name = tensor("input_243_cast_fp16")]; + tensor input_axes_0 = const()[name = tensor("input_axes_0"), val = tensor([-1])]; + tensor out_norm_weight_to_fp16 = const()[name = tensor("out_norm_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(303790720)))]; + tensor out_norm_bias_to_fp16 = const()[name = tensor("out_norm_bias_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(303792832)))]; + tensor var_10260_to_fp16 = const()[name = tensor("op_10260_to_fp16"), val = tensor(0x1.5p-17)]; + tensor input_cast_fp16 = layer_norm(axes = input_axes_0, beta = out_norm_bias_to_fp16, epsilon = var_10260_to_fp16, gamma = out_norm_weight_to_fp16, x = input_243_cast_fp16)[name = tensor("input_cast_fp16")]; + tensor input_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("input_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor out_eos_weight_to_fp16 = const()[name = tensor("out_eos_weight_to_fp16"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(303794944)))]; + tensor out_eos_bias_to_fp16 = const()[name = tensor("out_eos_bias_to_fp16"), val = tensor([-0x1.5ep-3])]; + tensor linear_97_cast_fp16 = linear(bias = out_eos_bias_to_fp16, weight = out_eos_weight_to_fp16, x = input_cast_fp16)[name = tensor("linear_97_cast_fp16")]; + tensor linear_97_cast_fp16_to_fp32_dtype_0 = const()[name = tensor("linear_97_cast_fp16_to_fp32_dtype_0"), val = tensor("fp32")]; + tensor var_10268 = cast(dtype = linear_97_cast_fp16_to_fp32_dtype_0, x = linear_97_cast_fp16)[name = tensor("cast_432")]; + tensor input = cast(dtype = input_cast_fp16_to_fp32_dtype_0, x = input_cast_fp16)[name = tensor("cast_433")]; + tensor var_10239 = cast(dtype = var_10239_cast_fp16_to_fp32_dtype_0, x = var_10239_cast_fp16)[name = tensor("cast_434")]; + tensor new_cache_internal_tensor_assign_2 = cast(dtype = reshape_239_cast_fp16_to_fp32_dtype_0, x = reshape_239_cast_fp16)[name = tensor("cast_435")]; + tensor var_9820 = cast(dtype = var_9820_cast_fp16_to_fp32_dtype_0, x = var_9820_cast_fp16)[name = tensor("cast_439")]; + tensor new_cache_45_internal_tensor_assign_2 = cast(dtype = reshape_229_cast_fp16_to_fp32_dtype_0, x = reshape_229_cast_fp16)[name = tensor("cast_440")]; + tensor var_9401 = cast(dtype = var_9401_cast_fp16_to_fp32_dtype_0, x = var_9401_cast_fp16)[name = tensor("cast_444")]; + tensor new_cache_43_internal_tensor_assign_2 = cast(dtype = reshape_219_cast_fp16_to_fp32_dtype_0, x = reshape_219_cast_fp16)[name = tensor("cast_445")]; + tensor var_8982 = cast(dtype = var_8982_cast_fp16_to_fp32_dtype_0, x = var_8982_cast_fp16)[name = tensor("cast_449")]; + tensor new_cache_41_internal_tensor_assign_2 = cast(dtype = reshape_209_cast_fp16_to_fp32_dtype_0, x = reshape_209_cast_fp16)[name = tensor("cast_450")]; + tensor var_8563 = cast(dtype = var_8563_cast_fp16_to_fp32_dtype_0, x = var_8563_cast_fp16)[name = tensor("cast_454")]; + tensor new_cache_39_internal_tensor_assign_2 = cast(dtype = reshape_199_cast_fp16_to_fp32_dtype_0, x = reshape_199_cast_fp16)[name = tensor("cast_455")]; + tensor var_8144 = cast(dtype = var_8144_cast_fp16_to_fp32_dtype_0, x = var_8144_cast_fp16)[name = tensor("cast_459")]; + tensor new_cache_37_internal_tensor_assign_2 = cast(dtype = reshape_189_cast_fp16_to_fp32_dtype_0, x = reshape_189_cast_fp16)[name = tensor("cast_460")]; + tensor var_7725 = cast(dtype = var_7725_cast_fp16_to_fp32_dtype_0, x = var_7725_cast_fp16)[name = tensor("cast_464")]; + tensor new_cache_35_internal_tensor_assign_2 = cast(dtype = reshape_179_cast_fp16_to_fp32_dtype_0, x = reshape_179_cast_fp16)[name = tensor("cast_465")]; + tensor var_7306 = cast(dtype = var_7306_cast_fp16_to_fp32_dtype_0, x = var_7306_cast_fp16)[name = tensor("cast_469")]; + tensor new_cache_33_internal_tensor_assign_2 = cast(dtype = reshape_169_cast_fp16_to_fp32_dtype_0, x = reshape_169_cast_fp16)[name = tensor("cast_470")]; + tensor var_6887 = cast(dtype = var_6887_cast_fp16_to_fp32_dtype_0, x = var_6887_cast_fp16)[name = tensor("cast_474")]; + tensor new_cache_31_internal_tensor_assign_2 = cast(dtype = reshape_159_cast_fp16_to_fp32_dtype_0, x = reshape_159_cast_fp16)[name = tensor("cast_475")]; + tensor var_6468 = cast(dtype = var_6468_cast_fp16_to_fp32_dtype_0, x = var_6468_cast_fp16)[name = tensor("cast_479")]; + tensor new_cache_29_internal_tensor_assign_2 = cast(dtype = reshape_149_cast_fp16_to_fp32_dtype_0, x = reshape_149_cast_fp16)[name = tensor("cast_480")]; + tensor var_6049 = cast(dtype = var_6049_cast_fp16_to_fp32_dtype_0, x = var_6049_cast_fp16)[name = tensor("cast_484")]; + tensor new_cache_27_internal_tensor_assign_2 = cast(dtype = reshape_139_cast_fp16_to_fp32_dtype_0, x = reshape_139_cast_fp16)[name = tensor("cast_485")]; + tensor var_5630 = cast(dtype = var_5630_cast_fp16_to_fp32_dtype_0, x = var_5630_cast_fp16)[name = tensor("cast_489")]; + tensor new_cache_25_internal_tensor_assign_2 = cast(dtype = reshape_129_cast_fp16_to_fp32_dtype_0, x = reshape_129_cast_fp16)[name = tensor("cast_490")]; + tensor var_5211 = cast(dtype = var_5211_cast_fp16_to_fp32_dtype_0, x = var_5211_cast_fp16)[name = tensor("cast_494")]; + tensor new_cache_23_internal_tensor_assign_2 = cast(dtype = reshape_119_cast_fp16_to_fp32_dtype_0, x = reshape_119_cast_fp16)[name = tensor("cast_495")]; + tensor var_4792 = cast(dtype = var_4792_cast_fp16_to_fp32_dtype_0, x = var_4792_cast_fp16)[name = tensor("cast_499")]; + tensor new_cache_21_internal_tensor_assign_2 = cast(dtype = reshape_109_cast_fp16_to_fp32_dtype_0, x = reshape_109_cast_fp16)[name = tensor("cast_500")]; + tensor var_4373 = cast(dtype = var_4373_cast_fp16_to_fp32_dtype_0, x = var_4373_cast_fp16)[name = tensor("cast_504")]; + tensor new_cache_19_internal_tensor_assign_2 = cast(dtype = reshape_99_cast_fp16_to_fp32_dtype_0, x = reshape_99_cast_fp16)[name = tensor("cast_505")]; + tensor var_3954 = cast(dtype = var_3954_cast_fp16_to_fp32_dtype_0, x = var_3954_cast_fp16)[name = tensor("cast_509")]; + tensor new_cache_17_internal_tensor_assign_2 = cast(dtype = reshape_89_cast_fp16_to_fp32_dtype_0, x = reshape_89_cast_fp16)[name = tensor("cast_510")]; + tensor var_3535 = cast(dtype = var_3535_cast_fp16_to_fp32_dtype_0, x = var_3535_cast_fp16)[name = tensor("cast_514")]; + tensor new_cache_15_internal_tensor_assign_2 = cast(dtype = reshape_79_cast_fp16_to_fp32_dtype_0, x = reshape_79_cast_fp16)[name = tensor("cast_515")]; + tensor var_3116 = cast(dtype = var_3116_cast_fp16_to_fp32_dtype_0, x = var_3116_cast_fp16)[name = tensor("cast_519")]; + tensor new_cache_13_internal_tensor_assign_2 = cast(dtype = reshape_69_cast_fp16_to_fp32_dtype_0, x = reshape_69_cast_fp16)[name = tensor("cast_520")]; + tensor var_2697 = cast(dtype = var_2697_cast_fp16_to_fp32_dtype_0, x = var_2697_cast_fp16)[name = tensor("cast_524")]; + tensor new_cache_11_internal_tensor_assign_2 = cast(dtype = reshape_59_cast_fp16_to_fp32_dtype_0, x = reshape_59_cast_fp16)[name = tensor("cast_525")]; + tensor var_2278 = cast(dtype = var_2278_cast_fp16_to_fp32_dtype_0, x = var_2278_cast_fp16)[name = tensor("cast_529")]; + tensor new_cache_9_internal_tensor_assign_2 = cast(dtype = reshape_49_cast_fp16_to_fp32_dtype_0, x = reshape_49_cast_fp16)[name = tensor("cast_530")]; + tensor var_1859 = cast(dtype = var_1859_cast_fp16_to_fp32_dtype_0, x = var_1859_cast_fp16)[name = tensor("cast_534")]; + tensor new_cache_7_internal_tensor_assign_2 = cast(dtype = reshape_39_cast_fp16_to_fp32_dtype_0, x = reshape_39_cast_fp16)[name = tensor("cast_535")]; + tensor var_1440 = cast(dtype = var_1440_cast_fp16_to_fp32_dtype_0, x = var_1440_cast_fp16)[name = tensor("cast_539")]; + tensor new_cache_5_internal_tensor_assign_2 = cast(dtype = reshape_29_cast_fp16_to_fp32_dtype_0, x = reshape_29_cast_fp16)[name = tensor("cast_540")]; + tensor var_1021 = cast(dtype = var_1021_cast_fp16_to_fp32_dtype_0, x = var_1021_cast_fp16)[name = tensor("cast_544")]; + tensor new_cache_3_internal_tensor_assign_2 = cast(dtype = reshape_19_cast_fp16_to_fp32_dtype_0, x = reshape_19_cast_fp16)[name = tensor("cast_545")]; + tensor var_602 = cast(dtype = var_602_cast_fp16_to_fp32_dtype_0, x = var_602_cast_fp16)[name = tensor("cast_549")]; + tensor new_cache_1_internal_tensor_assign_2 = cast(dtype = reshape_9_cast_fp16_to_fp32_dtype_0, x = reshape_9_cast_fp16)[name = tensor("cast_550")]; + } -> (input, var_10268, new_cache_1_internal_tensor_assign_2, var_602, new_cache_3_internal_tensor_assign_2, var_1021, new_cache_5_internal_tensor_assign_2, var_1440, new_cache_7_internal_tensor_assign_2, var_1859, new_cache_9_internal_tensor_assign_2, var_2278, new_cache_11_internal_tensor_assign_2, var_2697, new_cache_13_internal_tensor_assign_2, var_3116, new_cache_15_internal_tensor_assign_2, var_3535, new_cache_17_internal_tensor_assign_2, var_3954, new_cache_19_internal_tensor_assign_2, var_4373, new_cache_21_internal_tensor_assign_2, var_4792, new_cache_23_internal_tensor_assign_2, var_5211, new_cache_25_internal_tensor_assign_2, var_5630, new_cache_27_internal_tensor_assign_2, var_6049, new_cache_29_internal_tensor_assign_2, var_6468, new_cache_31_internal_tensor_assign_2, var_6887, new_cache_33_internal_tensor_assign_2, var_7306, new_cache_35_internal_tensor_assign_2, var_7725, new_cache_37_internal_tensor_assign_2, var_8144, new_cache_39_internal_tensor_assign_2, var_8563, new_cache_41_internal_tensor_assign_2, var_8982, new_cache_43_internal_tensor_assign_2, var_9401, new_cache_45_internal_tensor_assign_2, var_9820, new_cache_internal_tensor_assign_2, var_10239); +} \ No newline at end of file diff --git a/v2.1/spanish_24l/flowlm_stepv2.mlmodelc/weights/weight.bin b/v2.1/spanish_24l/flowlm_stepv2.mlmodelc/weights/weight.bin new file mode 100644 index 0000000000000000000000000000000000000000..a0dabc0b8fc628d1f30b32a73f179dc52c92f699 --- /dev/null +++ b/v2.1/spanish_24l/flowlm_stepv2.mlmodelc/weights/weight.bin @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:c602a9c0b379c64efa6de2ab44b4f77771d6cd5cc463fcf5b2375b38edeff5bd +size 303797056 diff --git a/v2.1/spanish_24l/flowlm_stepv2.mlpackage/Data/com.apple.CoreML/model.mlmodel b/v2.1/spanish_24l/flowlm_stepv2.mlpackage/Data/com.apple.CoreML/model.mlmodel new file mode 100644 index 0000000000000000000000000000000000000000..9dcf6c9b379124322089a31cd33f2edef2036085 --- /dev/null +++ b/v2.1/spanish_24l/flowlm_stepv2.mlpackage/Data/com.apple.CoreML/model.mlmodel @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:6e876a0e64b10c40e15cd77ade6e35a688fbf8d74774158780c555b8763d77c4 +size 843748 diff --git a/v2.1/spanish_24l/flowlm_stepv2.mlpackage/Data/com.apple.CoreML/weights/weight.bin b/v2.1/spanish_24l/flowlm_stepv2.mlpackage/Data/com.apple.CoreML/weights/weight.bin new file mode 100644 index 0000000000000000000000000000000000000000..a0dabc0b8fc628d1f30b32a73f179dc52c92f699 --- /dev/null +++ b/v2.1/spanish_24l/flowlm_stepv2.mlpackage/Data/com.apple.CoreML/weights/weight.bin @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:c602a9c0b379c64efa6de2ab44b4f77771d6cd5cc463fcf5b2375b38edeff5bd +size 303797056 diff --git a/v2.1/spanish_24l/flowlm_stepv2.mlpackage/Manifest.json b/v2.1/spanish_24l/flowlm_stepv2.mlpackage/Manifest.json new file mode 100644 index 0000000000000000000000000000000000000000..14918ffefdfa764a7a90cd483571468323fb4067 --- /dev/null +++ b/v2.1/spanish_24l/flowlm_stepv2.mlpackage/Manifest.json @@ -0,0 +1,18 @@ +{ + "fileFormatVersion": "1.0.0", + "itemInfoEntries": { + "7843107D-DFD3-4FD1-A3B4-921835E031E1": { + "author": "com.apple.CoreML", + "description": "CoreML Model Weights", + "name": "weights", + "path": "com.apple.CoreML/weights" + }, + "791A9167-6457-4CF1-9DF1-C7B761BE17E6": { + "author": "com.apple.CoreML", + "description": "CoreML Model Specification", + "name": "model.mlmodel", + "path": "com.apple.CoreML/model.mlmodel" + } + }, + "rootModelIdentifier": "791A9167-6457-4CF1-9DF1-C7B761BE17E6" +} diff --git a/v2.1/spanish_24l/manifest.json b/v2.1/spanish_24l/manifest.json new file mode 100644 index 0000000000000000000000000000000000000000..fc7da98920126d1261a9c47517b1303db1599db7 --- /dev/null +++ b/v2.1/spanish_24l/manifest.json @@ -0,0 +1,27 @@ +{ + "version": "2.1", + "base": "v2", + "kind": "optimization-reconvert", + "self_contained": true, + "note": "Same weights as v2. Re-converted for speed: fused flow decoder (100% ANE), one-shot cond prefill, fp16 flowlm. NOT a finetune.", + "new": [ + "flow_decoder_fused", + "cond_prefill", + "flowlm_step" + ], + "copied_from_v2": [ + "flowlm_stepv2.mlpackage", + "flowlm_stepv2.mlmodelc", + "mimi_decoder.mlpackage", + "mimi_decoder.mlmodelc", + "constants", + "constants_bin" + ], + "compute_units": { + "flow_decoder_fused": "all (ANE)", + "cond_prefill": "all (GPU)", + "flowlm_step": "all (GPU, fp16)", + "flowlm_stepv2": "cpuAndGpu (GPU, int8 \u2014 fastest flowlm)", + "mimi_decoder": "cpuOnly" + } +} diff --git a/v2.1/spanish_24l/mimi_decoder.mlmodelc/analytics/coremldata.bin b/v2.1/spanish_24l/mimi_decoder.mlmodelc/analytics/coremldata.bin new file mode 100644 index 0000000000000000000000000000000000000000..ccd13d21e0d395a10c5887caa6375a56f60607b1 --- /dev/null +++ b/v2.1/spanish_24l/mimi_decoder.mlmodelc/analytics/coremldata.bin @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:82f67b93642e157ae69dec33904e792b2f8f7b4d561e91b30b3d9d3b1e257add +size 243 diff --git a/v2.1/spanish_24l/mimi_decoder.mlmodelc/coremldata.bin b/v2.1/spanish_24l/mimi_decoder.mlmodelc/coremldata.bin new file mode 100644 index 0000000000000000000000000000000000000000..178af78caf046ea9fe0fa97b8ea797138f2c527b --- /dev/null +++ b/v2.1/spanish_24l/mimi_decoder.mlmodelc/coremldata.bin @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:691a933cff2ced4478d0c56ca1a0765d650b6001969c3e01a02229de7f9094f4 +size 1777 diff --git a/v2.1/spanish_24l/mimi_decoder.mlmodelc/metadata.json b/v2.1/spanish_24l/mimi_decoder.mlmodelc/metadata.json new file mode 100644 index 0000000000000000000000000000000000000000..f66f9e7ff5be2ffbc44c8300f3501f67a77afaf6 --- /dev/null +++ b/v2.1/spanish_24l/mimi_decoder.mlmodelc/metadata.json @@ -0,0 +1,573 @@ +[ + { + "metadataOutputVersion" : "3.0", + "storagePrecision" : "Float32", + "outputSchema" : [ + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1 × 1 × 1920)", + "shortDescription" : "", + "shape" : "[1, 1, 1920]", + "name" : "var_724", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1 × 512 × 16)", + "shortDescription" : "", + "shape" : "[1, 512, 16]", + "name" : "var_77", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 2 × 1 × 256 × 8 × 64)", + "shortDescription" : "", + "shape" : "[2, 1, 256, 8, 64]", + "name" : "var_210", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1)", + "shortDescription" : "", + "shape" : "[1]", + "name" : "var_741", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 2 × 1 × 256 × 8 × 64)", + "shortDescription" : "", + "shape" : "[2, 1, 256, 8, 64]", + "name" : "var_400", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1)", + "shortDescription" : "", + "shape" : "[1]", + "name" : "var_744", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1 × 512 × 6)", + "shortDescription" : "", + "shape" : "[1, 512, 6]", + "name" : "var_542", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1)", + "shortDescription" : "", + "shape" : "[1]", + "name" : "conv0_first", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1 × 256 × 6)", + "shortDescription" : "", + "shape" : "[1, 256, 6]", + "name" : "var_565", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1 × 256 × 2)", + "shortDescription" : "", + "shape" : "[1, 256, 2]", + "name" : "var_585", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1)", + "shortDescription" : "", + "shape" : "[1]", + "name" : "res0_conv0_first", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1 × 128 × 0)", + "shortDescription" : "", + "shape" : "[1, 128, 0]", + "name" : "res0_conv1_prev", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1)", + "shortDescription" : "", + "shape" : "[1]", + "name" : "res0_conv1_first", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1 × 128 × 5)", + "shortDescription" : "", + "shape" : "[1, 128, 5]", + "name" : "var_621", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1 × 128 × 2)", + "shortDescription" : "", + "shape" : "[1, 128, 2]", + "name" : "var_641", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1)", + "shortDescription" : "", + "shape" : "[1]", + "name" : "res1_conv0_first", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1 × 64 × 0)", + "shortDescription" : "", + "shape" : "[1, 64, 0]", + "name" : "res1_conv1_prev", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1)", + "shortDescription" : "", + "shape" : "[1]", + "name" : "res1_conv1_first", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1 × 64 × 4)", + "shortDescription" : "", + "shape" : "[1, 64, 4]", + "name" : "var_677", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1 × 64 × 2)", + "shortDescription" : "", + "shape" : "[1, 64, 2]", + "name" : "var_697", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1)", + "shortDescription" : "", + "shape" : "[1]", + "name" : "res2_conv0_first", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1 × 32 × 0)", + "shortDescription" : "", + "shape" : "[1, 32, 0]", + "name" : "res2_conv1_prev", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1)", + "shortDescription" : "", + "shape" : "[1]", + "name" : "res2_conv1_first", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1 × 64 × 2)", + "shortDescription" : "", + "shape" : "[1, 64, 2]", + "name" : "var_725", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1)", + "shortDescription" : "", + "shape" : "[1]", + "name" : "conv_final_first", + "type" : "MultiArray" + } + ], + "modelParameters" : [ + + ], + "specificationVersion" : 8, + "mlProgramOperationTypeHistogram" : { + "Ios17.notEqual" : 4, + "Tile" : 2, + "Ios17.lessEqual" : 2, + "Ios17.layerNorm" : 4, + "Stack" : 6, + "Ios17.greaterEqual" : 4, + "Ios17.scatterAlongAxis" : 4, + "Ios17.sub" : 17, + "Identity" : 11, + "Ios17.elu" : 10, + "Ios17.logicalNot" : 2, + "Ios17.sliceByIndex" : 35, + "Ios17.concat" : 9, + "Ios17.convTranspose" : 4, + "Ios17.floorDiv" : 4, + "Ios17.expandDims" : 7, + "Ios17.less" : 2, + "Ios17.mul" : 29, + "Ios17.transpose" : 10, + "Ios17.matmul" : 4, + "Ios16.gelu" : 2, + "Ios17.conv" : 9, + "Ios17.reshape" : 22, + "Split" : 2, + "Ios17.squeeze" : 6, + "Ios17.cast" : 6, + "Ios17.sin" : 2, + "Ios17.linear" : 8, + "Select" : 6, + "Ios16.softmax" : 2, + "Ios17.add" : 28, + "Ios17.cos" : 2, + "Ios17.logicalAnd" : 6 + }, + "computePrecision" : "Mixed (Float32, Int32)", + "isUpdatable" : "0", + "stateSchema" : [ + + ], + "availability" : { + "macOS" : "14.0", + "tvOS" : "17.0", + "visionOS" : "1.0", + "watchOS" : "10.0", + "iOS" : "17.0", + "macCatalyst" : "17.0" + }, + "modelType" : { + "name" : "MLModelType_mlProgram" + }, + "userDefinedMetadata" : { + "com.github.apple.coremltools.conversion_date" : "2026-04-24", + "com.github.apple.coremltools.source" : "torch==2.9.1", + "com.github.apple.coremltools.version" : "9.0", + "com.github.apple.coremltools.source_dialect" : "TorchScript" + }, + "inputSchema" : [ + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1 × 32)", + "shortDescription" : "", + "shape" : "[1, 32]", + "name" : "latent", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1 × 512 × 16)", + "shortDescription" : "", + "shape" : "[1, 512, 16]", + "name" : "upsample_partial", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 2 × 1 × 256 × 8 × 64)", + "shortDescription" : "", + "shape" : "[2, 1, 256, 8, 64]", + "name" : "attn0_cache", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1)", + "shortDescription" : "", + "shape" : "[1]", + "name" : "attn0_offset", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 2 × 1 × 256 × 8 × 64)", + "shortDescription" : "", + "shape" : "[2, 1, 256, 8, 64]", + "name" : "attn1_cache", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1)", + "shortDescription" : "", + "shape" : "[1]", + "name" : "attn1_offset", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1 × 512 × 6)", + "shortDescription" : "", + "shape" : "[1, 512, 6]", + "name" : "conv0_prev", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1)", + "shortDescription" : "", + "shape" : "[1]", + "name" : "conv0_first", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1 × 256 × 6)", + "shortDescription" : "", + "shape" : "[1, 256, 6]", + "name" : "convtr0_partial", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1 × 256 × 2)", + "shortDescription" : "", + "shape" : "[1, 256, 2]", + "name" : "res0_conv0_prev", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1)", + "shortDescription" : "", + "shape" : "[1]", + "name" : "res0_conv0_first", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1 × 128 × 0)", + "shortDescription" : "", + "shape" : "[1, 128, 0]", + "name" : "res0_conv1_prev", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1)", + "shortDescription" : "", + "shape" : "[1]", + "name" : "res0_conv1_first", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1 × 128 × 5)", + "shortDescription" : "", + "shape" : "[1, 128, 5]", + "name" : "convtr1_partial", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1 × 128 × 2)", + "shortDescription" : "", + "shape" : "[1, 128, 2]", + "name" : "res1_conv0_prev", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1)", + "shortDescription" : "", + "shape" : "[1]", + "name" : "res1_conv0_first", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1 × 64 × 0)", + "shortDescription" : "", + "shape" : "[1, 64, 0]", + "name" : "res1_conv1_prev", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1)", + "shortDescription" : "", + "shape" : "[1]", + "name" : "res1_conv1_first", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1 × 64 × 4)", + "shortDescription" : "", + "shape" : "[1, 64, 4]", + "name" : "convtr2_partial", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1 × 64 × 2)", + "shortDescription" : "", + "shape" : "[1, 64, 2]", + "name" : "res2_conv0_prev", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1)", + "shortDescription" : "", + "shape" : "[1]", + "name" : "res2_conv0_first", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1 × 32 × 0)", + "shortDescription" : "", + "shape" : "[1, 32, 0]", + "name" : "res2_conv1_prev", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1)", + "shortDescription" : "", + "shape" : "[1]", + "name" : "res2_conv1_first", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1 × 64 × 2)", + "shortDescription" : "", + "shape" : "[1, 64, 2]", + "name" : "conv_final_prev", + "type" : "MultiArray" + }, + { + "hasShapeFlexibility" : "0", + "isOptional" : "0", + "dataType" : "Float32", + "formattedType" : "MultiArray (Float32 1)", + "shortDescription" : "", + "shape" : "[1]", + "name" : "conv_final_first", + "type" : "MultiArray" + } + ], + "generatedClassName" : "mimi_decoder", + "method" : "predict" + } +] \ No newline at end of file diff --git a/v2.1/spanish_24l/mimi_decoder.mlmodelc/model.mil b/v2.1/spanish_24l/mimi_decoder.mlmodelc/model.mil new file mode 100644 index 0000000000000000000000000000000000000000..312fdf981100e6ea1ead8963b7a3aa07c21d32b8 --- /dev/null +++ b/v2.1/spanish_24l/mimi_decoder.mlmodelc/model.mil @@ -0,0 +1,646 @@ +program(1.0) +[buildInfo = dict, tensor>({{"coremlc-component-MIL", "3520.4.1"}, {"coremlc-version", "3520.5.1"}, {"coremltools-component-torch", "2.9.1"}, {"coremltools-source-dialect", "TorchScript"}, {"coremltools-version", "9.0"}})] +{ + func main(tensor attn0_cache, tensor attn0_offset, tensor attn1_cache, tensor attn1_offset, tensor conv0_first, tensor conv0_prev, tensor conv_final_first, tensor conv_final_prev, tensor convtr0_partial, tensor convtr1_partial, tensor convtr2_partial, tensor latent, tensor res0_conv0_first, tensor res0_conv0_prev, tensor res0_conv1_first, tensor res0_conv1_prev, tensor res1_conv0_first, tensor res1_conv0_prev, tensor res1_conv1_first, tensor res1_conv1_prev, tensor res2_conv0_first, tensor res2_conv0_prev, tensor res2_conv1_first, tensor res2_conv1_prev, tensor upsample_partial) { + tensor emb_mean = const()[name = tensor("emb_mean"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(64)))]; + tensor emb_std = const()[name = tensor("emb_std"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(256)))]; + tensor mimi_quantizer_output_proj_weight = const()[name = tensor("mimi_quantizer_output_proj_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(448)))]; + tensor mimi_upsample_convtr_convtr_weight = const()[name = tensor("mimi_upsample_convtr_convtr_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(66048)))]; + tensor mimi_decoder_transformer_transformer_layers_0_norm1_bias = const()[name = tensor("mimi_decoder_transformer_transformer_layers_0_norm1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(131648)))]; + tensor mimi_decoder_transformer_transformer_layers_0_norm1_weight = const()[name = tensor("mimi_decoder_transformer_transformer_layers_0_norm1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(133760)))]; + tensor mimi_decoder_transformer_transformer_layers_0_self_attn_in_proj_weight = const()[name = tensor("mimi_decoder_transformer_transformer_layers_0_self_attn_in_proj_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(135872)))]; + tensor mimi_decoder_transformer_transformer_layers_0_self_attn_out_proj_weight = const()[name = tensor("mimi_decoder_transformer_transformer_layers_0_self_attn_out_proj_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(3281664)))]; + tensor mimi_decoder_transformer_transformer_layers_0_layer_scale_1_scale = const()[name = tensor("mimi_decoder_transformer_transformer_layers_0_layer_scale_1_scale"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(4330304)))]; + tensor mimi_decoder_transformer_transformer_layers_0_norm2_bias = const()[name = tensor("mimi_decoder_transformer_transformer_layers_0_norm2_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(4332416)))]; + tensor mimi_decoder_transformer_transformer_layers_0_norm2_weight = const()[name = tensor("mimi_decoder_transformer_transformer_layers_0_norm2_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(4334528)))]; + tensor mimi_decoder_transformer_transformer_layers_0_linear1_weight = const()[name = tensor("mimi_decoder_transformer_transformer_layers_0_linear1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(4336640)))]; + tensor mimi_decoder_transformer_transformer_layers_0_linear2_weight = const()[name = tensor("mimi_decoder_transformer_transformer_layers_0_linear2_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(8531008)))]; + tensor mimi_decoder_transformer_transformer_layers_0_layer_scale_2_scale = const()[name = tensor("mimi_decoder_transformer_transformer_layers_0_layer_scale_2_scale"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(12725376)))]; + tensor mimi_decoder_transformer_transformer_layers_1_norm1_bias = const()[name = tensor("mimi_decoder_transformer_transformer_layers_1_norm1_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(12727488)))]; + tensor mimi_decoder_transformer_transformer_layers_1_norm1_weight = const()[name = tensor("mimi_decoder_transformer_transformer_layers_1_norm1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(12729600)))]; + tensor mimi_decoder_transformer_transformer_layers_1_self_attn_in_proj_weight = const()[name = tensor("mimi_decoder_transformer_transformer_layers_1_self_attn_in_proj_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(12731712)))]; + tensor mimi_decoder_transformer_transformer_layers_1_self_attn_out_proj_weight = const()[name = tensor("mimi_decoder_transformer_transformer_layers_1_self_attn_out_proj_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(15877504)))]; + tensor mimi_decoder_transformer_transformer_layers_1_layer_scale_1_scale = const()[name = tensor("mimi_decoder_transformer_transformer_layers_1_layer_scale_1_scale"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(16926144)))]; + tensor mimi_decoder_transformer_transformer_layers_1_norm2_bias = const()[name = tensor("mimi_decoder_transformer_transformer_layers_1_norm2_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(16928256)))]; + tensor mimi_decoder_transformer_transformer_layers_1_norm2_weight = const()[name = tensor("mimi_decoder_transformer_transformer_layers_1_norm2_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(16930368)))]; + tensor mimi_decoder_transformer_transformer_layers_1_linear1_weight = const()[name = tensor("mimi_decoder_transformer_transformer_layers_1_linear1_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(16932480)))]; + tensor mimi_decoder_transformer_transformer_layers_1_linear2_weight = const()[name = tensor("mimi_decoder_transformer_transformer_layers_1_linear2_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(21126848)))]; + tensor mimi_decoder_transformer_transformer_layers_1_layer_scale_2_scale = const()[name = tensor("mimi_decoder_transformer_transformer_layers_1_layer_scale_2_scale"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(25321216)))]; + tensor mimi_decoder_model_0_conv_bias = const()[name = tensor("mimi_decoder_model_0_conv_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(25323328)))]; + tensor mimi_decoder_model_0_conv_weight = const()[name = tensor("mimi_decoder_model_0_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(25325440)))]; + tensor mimi_decoder_model_2_convtr_bias = const()[name = tensor("mimi_decoder_model_2_convtr_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(32665536)))]; + tensor mimi_decoder_model_2_convtr_weight = const()[name = tensor("mimi_decoder_model_2_convtr_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(32666624)))]; + tensor mimi_decoder_model_3_block_1_conv_bias = const()[name = tensor("mimi_decoder_model_3_block_1_conv_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(38958144)))]; + tensor mimi_decoder_model_3_block_1_conv_weight = const()[name = tensor("mimi_decoder_model_3_block_1_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(38958720)))]; + tensor mimi_decoder_model_3_block_3_conv_bias = const()[name = tensor("mimi_decoder_model_3_block_3_conv_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(39352000)))]; + tensor mimi_decoder_model_3_block_3_conv_weight = const()[name = tensor("mimi_decoder_model_3_block_3_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(39353088)))]; + tensor mimi_decoder_model_5_convtr_bias = const()[name = tensor("mimi_decoder_model_5_convtr_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(39484224)))]; + tensor mimi_decoder_model_5_convtr_weight = const()[name = tensor("mimi_decoder_model_5_convtr_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(39484800)))]; + tensor mimi_decoder_model_6_block_1_conv_bias = const()[name = tensor("mimi_decoder_model_6_block_1_conv_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(40795584)))]; + tensor mimi_decoder_model_6_block_1_conv_weight = const()[name = tensor("mimi_decoder_model_6_block_1_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(40795904)))]; + tensor mimi_decoder_model_6_block_3_conv_bias = const()[name = tensor("mimi_decoder_model_6_block_3_conv_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(40894272)))]; + tensor mimi_decoder_model_6_block_3_conv_weight = const()[name = tensor("mimi_decoder_model_6_block_3_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(40894848)))]; + tensor mimi_decoder_model_8_convtr_bias = const()[name = tensor("mimi_decoder_model_8_convtr_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(40927680)))]; + tensor mimi_decoder_model_8_convtr_weight = const()[name = tensor("mimi_decoder_model_8_convtr_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(40928000)))]; + tensor mimi_decoder_model_9_block_1_conv_bias = const()[name = tensor("mimi_decoder_model_9_block_1_conv_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(41190208)))]; + tensor mimi_decoder_model_9_block_1_conv_weight = const()[name = tensor("mimi_decoder_model_9_block_1_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(41190400)))]; + tensor mimi_decoder_model_9_block_3_conv_bias = const()[name = tensor("mimi_decoder_model_9_block_3_conv_bias"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(41215040)))]; + tensor mimi_decoder_model_9_block_3_conv_weight = const()[name = tensor("mimi_decoder_model_9_block_3_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(41215360)))]; + tensor mimi_decoder_model_11_conv_bias = const()[name = tensor("mimi_decoder_model_11_conv_bias"), val = tensor([-0x1.96p-16])]; + tensor mimi_decoder_model_11_conv_weight = const()[name = tensor("mimi_decoder_model_11_conv_weight"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(41223616)))]; + tensor var_38 = mul(x = latent, y = emb_std)[name = tensor("op_38")]; + tensor denorm = add(x = var_38, y = emb_mean)[name = tensor("denorm")]; + tensor input_1_axes_0 = const()[name = tensor("input_1_axes_0"), val = tensor([-1])]; + tensor input_1 = expand_dims(axes = input_1_axes_0, x = denorm)[name = tensor("input_1")]; + tensor x_1_pad_type_0 = const()[name = tensor("x_1_pad_type_0"), val = tensor("valid")]; + tensor x_1_strides_0 = const()[name = tensor("x_1_strides_0"), val = tensor([1])]; + tensor x_1_pad_0 = const()[name = tensor("x_1_pad_0"), val = tensor([0, 0])]; + tensor x_1_dilations_0 = const()[name = tensor("x_1_dilations_0"), val = tensor([1])]; + tensor x_1_groups_0 = const()[name = tensor("x_1_groups_0"), val = tensor(1)]; + tensor x_1 = conv(dilations = x_1_dilations_0, groups = x_1_groups_0, pad = x_1_pad_0, pad_type = x_1_pad_type_0, strides = x_1_strides_0, weight = mimi_quantizer_output_proj_weight, x = input_1)[name = tensor("x_1")]; + tensor var_62 = const()[name = tensor("op_62"), val = tensor(-1)]; + tensor y_1_pad_type_0 = const()[name = tensor("y_1_pad_type_0"), val = tensor("valid")]; + tensor y_1_strides_0 = const()[name = tensor("y_1_strides_0"), val = tensor([16])]; + tensor y_1_groups_0 = const()[name = tensor("y_1_groups_0"), val = tensor(512)]; + tensor y_1_pad_0 = const()[name = tensor("y_1_pad_0"), val = tensor([0, 0])]; + tensor y_1_dilations_0 = const()[name = tensor("y_1_dilations_0"), val = tensor([1])]; + tensor y_1_has_output_shape_output_shape_0 = const()[name = tensor("y_1_has_output_shape_output_shape_0"), val = tensor([1, 512, 32])]; + tensor y_1_has_output_shape = conv_transpose(dilations = y_1_dilations_0, groups = y_1_groups_0, output_shape = y_1_has_output_shape_output_shape_0, pad = y_1_pad_0, pad_type = y_1_pad_type_0, strides = y_1_strides_0, weight = mimi_upsample_convtr_convtr_weight, x = x_1)[name = tensor("y_1_has_output_shape")]; + tensor var_72_begin_0 = const()[name = tensor("op_72_begin_0"), val = tensor([0, 0, 0])]; + tensor var_72_end_0 = const()[name = tensor("op_72_end_0"), val = tensor([1, 512, 16])]; + tensor var_72_end_mask_0 = const()[name = tensor("op_72_end_mask_0"), val = tensor([true, true, false])]; + tensor var_72 = slice_by_index(begin = var_72_begin_0, end = var_72_end_0, end_mask = var_72_end_mask_0, x = y_1_has_output_shape)[name = tensor("op_72")]; + tensor var_73 = add(x = var_72, y = upsample_partial)[name = tensor("op_73")]; + tensor var_74_begin_0 = const()[name = tensor("op_74_begin_0"), val = tensor([0, 0, 16])]; + tensor var_74_end_0 = const()[name = tensor("op_74_end_0"), val = tensor([1, 512, 32])]; + tensor var_74_end_mask_0 = const()[name = tensor("op_74_end_mask_0"), val = tensor([true, true, true])]; + tensor var_74 = slice_by_index(begin = var_74_begin_0, end = var_74_end_0, end_mask = var_74_end_mask_0, x = y_1_has_output_shape)[name = tensor("op_74")]; + tensor y_3_interleave_0 = const()[name = tensor("y_3_interleave_0"), val = tensor(false)]; + tensor y_3 = concat(axis = var_62, interleave = y_3_interleave_0, values = (var_73, var_74))[name = tensor("y_3")]; + tensor var_77_begin_0 = const()[name = tensor("op_77_begin_0"), val = tensor([0, 0, 16])]; + tensor var_77_end_0 = const()[name = tensor("op_77_end_0"), val = tensor([1, 512, 32])]; + tensor var_77_end_mask_0 = const()[name = tensor("op_77_end_mask_0"), val = tensor([true, true, true])]; + tensor var_77 = slice_by_index(begin = var_77_begin_0, end = var_77_end_0, end_mask = var_77_end_mask_0, x = y_3)[name = tensor("op_77")]; + tensor x_3_begin_0 = const()[name = tensor("x_3_begin_0"), val = tensor([0, 0, 0])]; + tensor x_3_end_0 = const()[name = tensor("x_3_end_0"), val = tensor([1, 512, 16])]; + tensor x_3_end_mask_0 = const()[name = tensor("x_3_end_mask_0"), val = tensor([true, true, false])]; + tensor x_3 = slice_by_index(begin = x_3_begin_0, end = x_3_end_0, end_mask = x_3_end_mask_0, x = y_3)[name = tensor("x_3")]; + tensor var_86 = const()[name = tensor("op_86"), val = tensor(0)]; + tensor var_91 = const()[name = tensor("op_91"), val = tensor(-1)]; + tensor var_100 = const()[name = tensor("op_100"), val = tensor(-0x1.ff933cp+127)]; + tensor var_102 = const()[name = tensor("op_102"), val = tensor(0x1.4f8b58p-17)]; + tensor input_3_perm_0 = const()[name = tensor("input_3_perm_0"), val = tensor([0, 2, 1])]; + tensor query_1_axes_0 = const()[name = tensor("query_1_axes_0"), val = tensor([-1])]; + tensor input_3 = transpose(perm = input_3_perm_0, x = x_3)[name = tensor("transpose_19")]; + tensor query_1 = layer_norm(axes = query_1_axes_0, beta = mimi_decoder_transformer_transformer_layers_0_norm1_bias, epsilon = var_102, gamma = mimi_decoder_transformer_transformer_layers_0_norm1_weight, x = input_3)[name = tensor("query_1")]; + tensor linear_0_bias_0 = const()[name = tensor("linear_0_bias_0"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(41224448)))]; + tensor projected_1 = linear(bias = linear_0_bias_0, weight = mimi_decoder_transformer_transformer_layers_0_self_attn_in_proj_weight, x = query_1)[name = tensor("linear_0")]; + tensor var_130 = const()[name = tensor("op_130"), val = tensor([1, 16, 3, 8, 64])]; + tensor packed_1 = reshape(shape = var_130, x = projected_1)[name = tensor("packed_1")]; + tensor var_132_split_sizes_0 = const()[name = tensor("op_132_split_sizes_0"), val = tensor([1, 1, 1])]; + tensor var_132_axis_0 = const()[name = tensor("op_132_axis_0"), val = tensor(2)]; + tensor var_132_0, tensor var_132_1, tensor var_132_2 = split(axis = var_132_axis_0, split_sizes = var_132_split_sizes_0, x = packed_1)[name = tensor("op_132")]; + tensor squeeze_0_axes_0 = const()[name = tensor("squeeze_0_axes_0"), val = tensor([2])]; + tensor squeeze_0 = squeeze(axes = squeeze_0_axes_0, x = var_132_0)[name = tensor("squeeze_0")]; + tensor squeeze_1_axes_0 = const()[name = tensor("squeeze_1_axes_0"), val = tensor([2])]; + tensor squeeze_1 = squeeze(axes = squeeze_1_axes_0, x = var_132_1)[name = tensor("squeeze_1")]; + tensor squeeze_2_axes_0 = const()[name = tensor("squeeze_2_axes_0"), val = tensor([2])]; + tensor squeeze_2 = squeeze(axes = squeeze_2_axes_0, x = var_132_2)[name = tensor("squeeze_2")]; + tensor offset_3_begin_0 = const()[name = tensor("offset_3_begin_0"), val = tensor([0])]; + tensor offset_3_end_0 = const()[name = tensor("offset_3_end_0"), val = tensor([1])]; + tensor offset_3_end_mask_0 = const()[name = tensor("offset_3_end_mask_0"), val = tensor([false])]; + tensor offset_3_squeeze_mask_0 = const()[name = tensor("offset_3_squeeze_mask_0"), val = tensor([true])]; + tensor offset_3 = slice_by_index(begin = offset_3_begin_0, end = offset_3_end_0, end_mask = offset_3_end_mask_0, squeeze_mask = offset_3_squeeze_mask_0, x = attn0_offset)[name = tensor("offset_3")]; + tensor freqs_1 = const()[name = tensor("freqs_1"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(41230656)))]; + tensor ts_1_promoted = const()[name = tensor("ts_1_promoted"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(41230848)))]; + tensor ts_3 = add(x = ts_1_promoted, y = offset_3)[name = tensor("ts_3")]; + tensor var_148 = const()[name = tensor("op_148"), val = tensor([-1, 1, 1])]; + tensor ts_5 = reshape(shape = var_148, x = ts_3)[name = tensor("ts_5")]; + tensor var_150 = const()[name = tensor("op_150"), val = tensor([1, 16, 8, 32, 2])]; + tensor q_3 = reshape(shape = var_150, x = squeeze_0)[name = tensor("q_3")]; + tensor var_152 = const()[name = tensor("op_152"), val = tensor([1, 16, 8, 32, 2])]; + tensor k_3 = reshape(shape = var_152, x = squeeze_1)[name = tensor("k_3")]; + tensor var_154_begin_0 = const()[name = tensor("op_154_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_154_end_0 = const()[name = tensor("op_154_end_0"), val = tensor([1, 16, 8, 32, 1])]; + tensor var_154_end_mask_0 = const()[name = tensor("op_154_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_154_squeeze_mask_0 = const()[name = tensor("op_154_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_154 = slice_by_index(begin = var_154_begin_0, end = var_154_end_0, end_mask = var_154_end_mask_0, squeeze_mask = var_154_squeeze_mask_0, x = q_3)[name = tensor("op_154")]; + tensor var_156_begin_0 = const()[name = tensor("op_156_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_156_end_0 = const()[name = tensor("op_156_end_0"), val = tensor([1, 16, 8, 32, 2])]; + tensor var_156_end_mask_0 = const()[name = tensor("op_156_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_156_squeeze_mask_0 = const()[name = tensor("op_156_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_156 = slice_by_index(begin = var_156_begin_0, end = var_156_end_0, end_mask = var_156_end_mask_0, squeeze_mask = var_156_squeeze_mask_0, x = q_3)[name = tensor("op_156")]; + tensor var_158_begin_0 = const()[name = tensor("op_158_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_158_end_0 = const()[name = tensor("op_158_end_0"), val = tensor([1, 16, 8, 32, 1])]; + tensor var_158_end_mask_0 = const()[name = tensor("op_158_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_158_squeeze_mask_0 = const()[name = tensor("op_158_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_158 = slice_by_index(begin = var_158_begin_0, end = var_158_end_0, end_mask = var_158_end_mask_0, squeeze_mask = var_158_squeeze_mask_0, x = k_3)[name = tensor("op_158")]; + tensor var_160_begin_0 = const()[name = tensor("op_160_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_160_end_0 = const()[name = tensor("op_160_end_0"), val = tensor([1, 16, 8, 32, 2])]; + tensor var_160_end_mask_0 = const()[name = tensor("op_160_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_160_squeeze_mask_0 = const()[name = tensor("op_160_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_160 = slice_by_index(begin = var_160_begin_0, end = var_160_end_0, end_mask = var_160_end_mask_0, squeeze_mask = var_160_squeeze_mask_0, x = k_3)[name = tensor("op_160")]; + tensor var_162 = mul(x = freqs_1, y = ts_5)[name = tensor("op_162")]; + tensor rotr_1 = cos(x = var_162)[name = tensor("rotr_1")]; + tensor roti_1 = sin(x = var_162)[name = tensor("roti_1")]; + tensor var_166 = mul(x = var_154, y = rotr_1)[name = tensor("op_166")]; + tensor var_167 = mul(x = var_156, y = roti_1)[name = tensor("op_167")]; + tensor qor_1 = sub(x = var_166, y = var_167)[name = tensor("qor_1")]; + tensor var_169 = mul(x = var_154, y = roti_1)[name = tensor("op_169")]; + tensor var_170 = mul(x = var_156, y = rotr_1)[name = tensor("op_170")]; + tensor qoi_1 = add(x = var_169, y = var_170)[name = tensor("qoi_1")]; + tensor var_172 = mul(x = var_158, y = rotr_1)[name = tensor("op_172")]; + tensor var_173 = mul(x = var_160, y = roti_1)[name = tensor("op_173")]; + tensor kor_1 = sub(x = var_172, y = var_173)[name = tensor("kor_1")]; + tensor var_175 = mul(x = var_158, y = roti_1)[name = tensor("op_175")]; + tensor var_176 = mul(x = var_160, y = rotr_1)[name = tensor("op_176")]; + tensor koi_1 = add(x = var_175, y = var_176)[name = tensor("koi_1")]; + tensor qo_1_axis_0 = const()[name = tensor("qo_1_axis_0"), val = tensor(-1)]; + tensor qo_1 = stack(axis = qo_1_axis_0, values = (qor_1, qoi_1))[name = tensor("qo_1")]; + tensor ko_1_axis_0 = const()[name = tensor("ko_1_axis_0"), val = tensor(-1)]; + tensor ko_1 = stack(axis = ko_1_axis_0, values = (kor_1, koi_1))[name = tensor("ko_1")]; + tensor var_186 = const()[name = tensor("op_186"), val = tensor([1, 16, 8, 64])]; + tensor q_5 = reshape(shape = var_186, x = qo_1)[name = tensor("q_5")]; + tensor var_188 = const()[name = tensor("op_188"), val = tensor([1, 16, 8, 64])]; + tensor k_5 = reshape(shape = var_188, x = ko_1)[name = tensor("k_5")]; + tensor capacity_1 = const()[name = tensor("capacity_1"), val = tensor([256])]; + tensor var_193_dtype_0 = const()[name = tensor("op_193_dtype_0"), val = tensor("int32")]; + tensor var_194 = const()[name = tensor("op_194"), val = tensor([1, 1])]; + tensor var_193 = cast(dtype = var_193_dtype_0, x = attn0_offset)[name = tensor("cast_49")]; + tensor write_base_1 = reshape(shape = var_194, x = var_193)[name = tensor("write_base_1")]; + tensor write_range_1 = const()[name = tensor("write_range_1"), val = tensor([[0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15]])]; + tensor abs_idx_1 = add(x = write_base_1, y = write_range_1)[name = tensor("abs_idx_1")]; + tensor wrapped_1_div = floor_div(x = abs_idx_1, y = capacity_1)[name = tensor("wrapped_1_div")]; + tensor wrapped_1_div_scaled = mul(x = wrapped_1_div, y = capacity_1)[name = tensor("wrapped_1_div_scaled")]; + tensor wrapped_1 = sub(x = abs_idx_1, y = wrapped_1_div_scaled)[name = tensor("wrapped_1")]; + tensor var_201 = const()[name = tensor("op_201"), val = tensor([1, 16, 1, 1])]; + tensor var_202 = reshape(shape = var_201, x = wrapped_1)[name = tensor("op_202")]; + tensor write_indexes_1_reps_0 = const()[name = tensor("write_indexes_1_reps_0"), val = tensor([1, 1, 8, 64])]; + tensor write_indexes_1 = tile(reps = write_indexes_1_reps_0, x = var_202)[name = tensor("write_indexes_1")]; + tensor var_205_begin_0 = const()[name = tensor("op_205_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_205_end_0 = const()[name = tensor("op_205_end_0"), val = tensor([1, 1, 256, 8, 64])]; + tensor var_205_end_mask_0 = const()[name = tensor("op_205_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_205_squeeze_mask_0 = const()[name = tensor("op_205_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor var_205 = slice_by_index(begin = var_205_begin_0, end = var_205_end_0, end_mask = var_205_end_mask_0, squeeze_mask = var_205_squeeze_mask_0, x = attn0_cache)[name = tensor("op_205")]; + tensor new_k_cache_1_axis_0 = const()[name = tensor("new_k_cache_1_axis_0"), val = tensor(1)]; + tensor new_k_cache_1_mode_0 = const()[name = tensor("new_k_cache_1_mode_0"), val = tensor("update")]; + tensor new_k_cache_1_validate_indices_0 = const()[name = tensor("new_k_cache_1_validate_indices_0"), val = tensor(false)]; + tensor new_k_cache_1 = scatter_along_axis(axis = new_k_cache_1_axis_0, data = var_205, indices = write_indexes_1, mode = new_k_cache_1_mode_0, updates = k_5, validate_indices = new_k_cache_1_validate_indices_0)[name = tensor("new_k_cache_1")]; + tensor var_207_begin_0 = const()[name = tensor("op_207_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor var_207_end_0 = const()[name = tensor("op_207_end_0"), val = tensor([2, 1, 256, 8, 64])]; + tensor var_207_end_mask_0 = const()[name = tensor("op_207_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_207_squeeze_mask_0 = const()[name = tensor("op_207_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor var_207 = slice_by_index(begin = var_207_begin_0, end = var_207_end_0, end_mask = var_207_end_mask_0, squeeze_mask = var_207_squeeze_mask_0, x = attn0_cache)[name = tensor("op_207")]; + tensor new_v_cache_1_axis_0 = const()[name = tensor("new_v_cache_1_axis_0"), val = tensor(1)]; + tensor new_v_cache_1_mode_0 = const()[name = tensor("new_v_cache_1_mode_0"), val = tensor("update")]; + tensor new_v_cache_1_validate_indices_0 = const()[name = tensor("new_v_cache_1_validate_indices_0"), val = tensor(false)]; + tensor new_v_cache_1 = scatter_along_axis(axis = new_v_cache_1_axis_0, data = var_207, indices = write_indexes_1, mode = new_v_cache_1_mode_0, updates = squeeze_2, validate_indices = new_v_cache_1_validate_indices_0)[name = tensor("new_v_cache_1")]; + tensor var_210_axis_0 = const()[name = tensor("op_210_axis_0"), val = tensor(0)]; + tensor var_210 = stack(axis = var_210_axis_0, values = (new_k_cache_1, new_v_cache_1))[name = tensor("op_210")]; + tensor var_211 = not_equal(x = new_k_cache_1, y = new_k_cache_1)[name = tensor("op_211")]; + tensor var_212 = const()[name = tensor("op_212"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(41230976)))]; + tensor new_k_cache_3 = select(a = var_212, b = new_k_cache_1, cond = var_211)[name = tensor("new_k_cache_3")]; + tensor var_214 = not_equal(x = new_v_cache_1, y = new_v_cache_1)[name = tensor("op_214")]; + tensor new_v_cache_3 = select(a = var_212, b = new_v_cache_1, cond = var_214)[name = tensor("new_v_cache_3")]; + tensor var_219 = const()[name = tensor("op_219"), val = tensor([0, 2, 1, 3])]; + tensor var_221 = const()[name = tensor("op_221"), val = tensor([1, 1])]; + tensor var_222 = reshape(shape = var_221, x = attn0_offset)[name = tensor("op_222")]; + tensor var_224_promoted = const()[name = tensor("op_224_promoted"), val = tensor([0x1.ep+3])]; + tensor var_225 = add(x = var_222, y = var_224_promoted)[name = tensor("op_225")]; + tensor last_pos_1_dtype_0 = const()[name = tensor("last_pos_1_dtype_0"), val = tensor("int32")]; + tensor slot_idx_1 = const()[name = tensor("slot_idx_1"), val = tensor([[0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19, 20, 21, 22, 23, 24, 25, 26, 27, 28, 29, 30, 31, 32, 33, 34, 35, 36, 37, 38, 39, 40, 41, 42, 43, 44, 45, 46, 47, 48, 49, 50, 51, 52, 53, 54, 55, 56, 57, 58, 59, 60, 61, 62, 63, 64, 65, 66, 67, 68, 69, 70, 71, 72, 73, 74, 75, 76, 77, 78, 79, 80, 81, 82, 83, 84, 85, 86, 87, 88, 89, 90, 91, 92, 93, 94, 95, 96, 97, 98, 99, 100, 101, 102, 103, 104, 105, 106, 107, 108, 109, 110, 111, 112, 113, 114, 115, 116, 117, 118, 119, 120, 121, 122, 123, 124, 125, 126, 127, 128, 129, 130, 131, 132, 133, 134, 135, 136, 137, 138, 139, 140, 141, 142, 143, 144, 145, 146, 147, 148, 149, 150, 151, 152, 153, 154, 155, 156, 157, 158, 159, 160, 161, 162, 163, 164, 165, 166, 167, 168, 169, 170, 171, 172, 173, 174, 175, 176, 177, 178, 179, 180, 181, 182, 183, 184, 185, 186, 187, 188, 189, 190, 191, 192, 193, 194, 195, 196, 197, 198, 199, 200, 201, 202, 203, 204, 205, 206, 207, 208, 209, 210, 211, 212, 213, 214, 215, 216, 217, 218, 219, 220, 221, 222, 223, 224, 225, 226, 227, 228, 229, 230, 231, 232, 233, 234, 235, 236, 237, 238, 239, 240, 241, 242, 243, 244, 245, 246, 247, 248, 249, 250, 251, 252, 253, 254, 255]])]; + tensor last_pos_1 = cast(dtype = last_pos_1_dtype_0, x = var_225)[name = tensor("cast_48")]; + tensor diff_1 = sub(x = last_pos_1, y = slot_idx_1)[name = tensor("diff_1")]; + tensor var_231_div = floor_div(x = diff_1, y = capacity_1)[name = tensor("op_231_div")]; + tensor var_231_div_scaled = mul(x = var_231_div, y = capacity_1)[name = tensor("op_231_div_scaled")]; + tensor var_231 = sub(x = diff_1, y = var_231_div_scaled)[name = tensor("op_231")]; + tensor pos_k_1 = sub(x = last_pos_1, y = var_231)[name = tensor("pos_k_1")]; + tensor var_237_promoted = const()[name = tensor("op_237_promoted"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(41755328)))]; + tensor pos_q_1 = add(x = var_222, y = var_237_promoted)[name = tensor("pos_q_1")]; + tensor var_241_axes_0 = const()[name = tensor("op_241_axes_0"), val = tensor([2])]; + tensor var_241 = expand_dims(axes = var_241_axes_0, x = pos_q_1)[name = tensor("op_241")]; + tensor var_243_axes_0 = const()[name = tensor("op_243_axes_0"), val = tensor([1])]; + tensor var_243 = expand_dims(axes = var_243_axes_0, x = pos_k_1)[name = tensor("op_243")]; + tensor var_244_promoted_dtype_0 = const()[name = tensor("op_244_promoted_dtype_0"), val = tensor("fp32")]; + tensor var_244_promoted = cast(dtype = var_244_promoted_dtype_0, x = var_243)[name = tensor("cast_47")]; + tensor delta_1 = sub(x = var_241, y = var_244_promoted)[name = tensor("delta_1")]; + tensor valid_1 = greater_equal(x = var_243, y = var_86)[name = tensor("valid_1")]; + tensor var_253 = const()[name = tensor("op_253"), val = tensor([1, 1, 1])]; + tensor var_254 = reshape(shape = var_253, x = attn0_offset)[name = tensor("op_254")]; + tensor var_256_promoted = const()[name = tensor("op_256_promoted"), val = tensor([0x1.ep+3])]; + tensor var_257 = add(x = var_254, y = var_256_promoted)[name = tensor("op_257")]; + tensor var_258 = less_equal(x = var_244_promoted, y = var_257)[name = tensor("op_258")]; + tensor valid_3 = logical_and(x = valid_1, y = var_258)[name = tensor("valid_3")]; + tensor var_86_promoted = const()[name = tensor("op_86_promoted"), val = tensor(0x0p+0)]; + tensor var_260 = greater_equal(x = delta_1, y = var_86_promoted)[name = tensor("op_260")]; + tensor attn_mask_1 = logical_and(x = valid_3, y = var_260)[name = tensor("attn_mask_1")]; + tensor var_98_promoted = const()[name = tensor("op_98_promoted"), val = tensor(0x1.f4p+7)]; + tensor var_262 = less(x = delta_1, y = var_98_promoted)[name = tensor("op_262")]; + tensor attn_mask_3 = logical_and(x = attn_mask_1, y = var_262)[name = tensor("attn_mask_3")]; + tensor attn_mask_5_axes_0 = const()[name = tensor("attn_mask_5_axes_0"), val = tensor([1])]; + tensor attn_mask_5 = expand_dims(axes = attn_mask_5_axes_0, x = attn_mask_3)[name = tensor("attn_mask_5")]; + tensor var_267_transpose_x_0 = const()[name = tensor("op_267_transpose_x_0"), val = tensor(false)]; + tensor var_267_transpose_y_0 = const()[name = tensor("op_267_transpose_y_0"), val = tensor(false)]; + tensor transpose_6_perm_0 = const()[name = tensor("transpose_6_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_7_perm_0 = const()[name = tensor("transpose_7_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_7 = transpose(perm = transpose_7_perm_0, x = new_k_cache_3)[name = tensor("transpose_16")]; + tensor transpose_6 = transpose(perm = transpose_6_perm_0, x = q_5)[name = tensor("transpose_17")]; + tensor var_267 = matmul(transpose_x = var_267_transpose_x_0, transpose_y = var_267_transpose_y_0, x = transpose_6, y = transpose_7)[name = tensor("op_267")]; + tensor var_268 = const()[name = tensor("op_268"), val = tensor(0x1p-3)]; + tensor attn_1 = mul(x = var_267, y = var_268)[name = tensor("attn_1")]; + tensor var_270 = logical_not(x = attn_mask_5)[name = tensor("op_270")]; + tensor attn_3 = select(a = var_100, b = attn_1, cond = var_270)[name = tensor("attn_3")]; + tensor attn_5 = softmax(axis = var_91, x = attn_3)[name = tensor("attn_5")]; + tensor x_5_transpose_x_0 = const()[name = tensor("x_5_transpose_x_0"), val = tensor(false)]; + tensor x_5_transpose_y_0 = const()[name = tensor("x_5_transpose_y_0"), val = tensor(false)]; + tensor v_attn_1 = transpose(perm = var_219, x = new_v_cache_3)[name = tensor("transpose_18")]; + tensor x_5 = matmul(transpose_x = x_5_transpose_x_0, transpose_y = x_5_transpose_y_0, x = attn_5, y = v_attn_1)[name = tensor("x_5")]; + tensor var_274_perm_0 = const()[name = tensor("op_274_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_275 = const()[name = tensor("op_275"), val = tensor([1, 16, 512])]; + tensor var_274 = transpose(perm = var_274_perm_0, x = x_5)[name = tensor("transpose_15")]; + tensor input_5 = reshape(shape = var_275, x = var_274)[name = tensor("input_5")]; + tensor linear_1_bias_0 = const()[name = tensor("linear_1_bias_0"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(41755456)))]; + tensor x_7 = linear(bias = linear_1_bias_0, weight = mimi_decoder_transformer_transformer_layers_0_self_attn_out_proj_weight, x = input_5)[name = tensor("linear_1")]; + tensor var_284 = mul(x = mimi_decoder_transformer_transformer_layers_0_layer_scale_1_scale, y = x_7)[name = tensor("op_284")]; + tensor input_7 = add(x = input_3, y = var_284)[name = tensor("input_7")]; + tensor input_9_axes_0 = const()[name = tensor("input_9_axes_0"), val = tensor([-1])]; + tensor input_9 = layer_norm(axes = input_9_axes_0, beta = mimi_decoder_transformer_transformer_layers_0_norm2_bias, epsilon = var_102, gamma = mimi_decoder_transformer_transformer_layers_0_norm2_weight, x = input_7)[name = tensor("input_9")]; + tensor linear_2_bias_0 = const()[name = tensor("linear_2_bias_0"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(41757568)))]; + tensor var_291 = linear(bias = linear_2_bias_0, weight = mimi_decoder_transformer_transformer_layers_0_linear1_weight, x = input_9)[name = tensor("linear_2")]; + tensor input_11_mode_0 = const()[name = tensor("input_11_mode_0"), val = tensor("EXACT")]; + tensor input_11 = gelu(mode = input_11_mode_0, x = var_291)[name = tensor("input_11")]; + tensor x_9 = linear(bias = linear_1_bias_0, weight = mimi_decoder_transformer_transformer_layers_0_linear2_weight, x = input_11)[name = tensor("linear_3")]; + tensor var_297 = mul(x = mimi_decoder_transformer_transformer_layers_0_layer_scale_2_scale, y = x_9)[name = tensor("op_297")]; + tensor input_13 = add(x = input_7, y = var_297)[name = tensor("input_13")]; + tensor query_axes_0 = const()[name = tensor("query_axes_0"), val = tensor([-1])]; + tensor query = layer_norm(axes = query_axes_0, beta = mimi_decoder_transformer_transformer_layers_1_norm1_bias, epsilon = var_102, gamma = mimi_decoder_transformer_transformer_layers_1_norm1_weight, x = input_13)[name = tensor("query")]; + tensor projected = linear(bias = linear_0_bias_0, weight = mimi_decoder_transformer_transformer_layers_1_self_attn_in_proj_weight, x = query)[name = tensor("linear_4")]; + tensor var_320 = const()[name = tensor("op_320"), val = tensor([1, 16, 3, 8, 64])]; + tensor packed = reshape(shape = var_320, x = projected)[name = tensor("packed")]; + tensor var_322_split_sizes_0 = const()[name = tensor("op_322_split_sizes_0"), val = tensor([1, 1, 1])]; + tensor var_322_axis_0 = const()[name = tensor("op_322_axis_0"), val = tensor(2)]; + tensor var_322_0, tensor var_322_1, tensor var_322_2 = split(axis = var_322_axis_0, split_sizes = var_322_split_sizes_0, x = packed)[name = tensor("op_322")]; + tensor squeeze_3_axes_0 = const()[name = tensor("squeeze_3_axes_0"), val = tensor([2])]; + tensor squeeze_3 = squeeze(axes = squeeze_3_axes_0, x = var_322_0)[name = tensor("squeeze_3")]; + tensor squeeze_4_axes_0 = const()[name = tensor("squeeze_4_axes_0"), val = tensor([2])]; + tensor squeeze_4 = squeeze(axes = squeeze_4_axes_0, x = var_322_1)[name = tensor("squeeze_4")]; + tensor squeeze_5_axes_0 = const()[name = tensor("squeeze_5_axes_0"), val = tensor([2])]; + tensor squeeze_5 = squeeze(axes = squeeze_5_axes_0, x = var_322_2)[name = tensor("squeeze_5")]; + tensor offset_begin_0 = const()[name = tensor("offset_begin_0"), val = tensor([0])]; + tensor offset_end_0 = const()[name = tensor("offset_end_0"), val = tensor([1])]; + tensor offset_end_mask_0 = const()[name = tensor("offset_end_mask_0"), val = tensor([false])]; + tensor offset_squeeze_mask_0 = const()[name = tensor("offset_squeeze_mask_0"), val = tensor([true])]; + tensor offset = slice_by_index(begin = offset_begin_0, end = offset_end_0, end_mask = offset_end_mask_0, squeeze_mask = offset_squeeze_mask_0, x = attn1_offset)[name = tensor("offset")]; + tensor freqs = const()[name = tensor("freqs"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(41765824)))]; + tensor ts_7_promoted = const()[name = tensor("ts_7_promoted"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(41766016)))]; + tensor ts_9 = add(x = ts_7_promoted, y = offset)[name = tensor("ts_9")]; + tensor var_338 = const()[name = tensor("op_338"), val = tensor([-1, 1, 1])]; + tensor ts = reshape(shape = var_338, x = ts_9)[name = tensor("ts")]; + tensor var_340 = const()[name = tensor("op_340"), val = tensor([1, 16, 8, 32, 2])]; + tensor q_9 = reshape(shape = var_340, x = squeeze_3)[name = tensor("q_9")]; + tensor var_342 = const()[name = tensor("op_342"), val = tensor([1, 16, 8, 32, 2])]; + tensor k_9 = reshape(shape = var_342, x = squeeze_4)[name = tensor("k_9")]; + tensor var_344_begin_0 = const()[name = tensor("op_344_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_344_end_0 = const()[name = tensor("op_344_end_0"), val = tensor([1, 16, 8, 32, 1])]; + tensor var_344_end_mask_0 = const()[name = tensor("op_344_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_344_squeeze_mask_0 = const()[name = tensor("op_344_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_344 = slice_by_index(begin = var_344_begin_0, end = var_344_end_0, end_mask = var_344_end_mask_0, squeeze_mask = var_344_squeeze_mask_0, x = q_9)[name = tensor("op_344")]; + tensor var_346_begin_0 = const()[name = tensor("op_346_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_346_end_0 = const()[name = tensor("op_346_end_0"), val = tensor([1, 16, 8, 32, 2])]; + tensor var_346_end_mask_0 = const()[name = tensor("op_346_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_346_squeeze_mask_0 = const()[name = tensor("op_346_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_346 = slice_by_index(begin = var_346_begin_0, end = var_346_end_0, end_mask = var_346_end_mask_0, squeeze_mask = var_346_squeeze_mask_0, x = q_9)[name = tensor("op_346")]; + tensor var_348_begin_0 = const()[name = tensor("op_348_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_348_end_0 = const()[name = tensor("op_348_end_0"), val = tensor([1, 16, 8, 32, 1])]; + tensor var_348_end_mask_0 = const()[name = tensor("op_348_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_348_squeeze_mask_0 = const()[name = tensor("op_348_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_348 = slice_by_index(begin = var_348_begin_0, end = var_348_end_0, end_mask = var_348_end_mask_0, squeeze_mask = var_348_squeeze_mask_0, x = k_9)[name = tensor("op_348")]; + tensor var_350_begin_0 = const()[name = tensor("op_350_begin_0"), val = tensor([0, 0, 0, 0, 1])]; + tensor var_350_end_0 = const()[name = tensor("op_350_end_0"), val = tensor([1, 16, 8, 32, 2])]; + tensor var_350_end_mask_0 = const()[name = tensor("op_350_end_mask_0"), val = tensor([true, true, true, true, false])]; + tensor var_350_squeeze_mask_0 = const()[name = tensor("op_350_squeeze_mask_0"), val = tensor([false, false, false, false, true])]; + tensor var_350 = slice_by_index(begin = var_350_begin_0, end = var_350_end_0, end_mask = var_350_end_mask_0, squeeze_mask = var_350_squeeze_mask_0, x = k_9)[name = tensor("op_350")]; + tensor var_352 = mul(x = freqs, y = ts)[name = tensor("op_352")]; + tensor rotr = cos(x = var_352)[name = tensor("rotr")]; + tensor roti = sin(x = var_352)[name = tensor("roti")]; + tensor var_356 = mul(x = var_344, y = rotr)[name = tensor("op_356")]; + tensor var_357 = mul(x = var_346, y = roti)[name = tensor("op_357")]; + tensor qor_5 = sub(x = var_356, y = var_357)[name = tensor("qor_5")]; + tensor var_359 = mul(x = var_344, y = roti)[name = tensor("op_359")]; + tensor var_360 = mul(x = var_346, y = rotr)[name = tensor("op_360")]; + tensor qoi_5 = add(x = var_359, y = var_360)[name = tensor("qoi_5")]; + tensor var_362 = mul(x = var_348, y = rotr)[name = tensor("op_362")]; + tensor var_363 = mul(x = var_350, y = roti)[name = tensor("op_363")]; + tensor kor_5 = sub(x = var_362, y = var_363)[name = tensor("kor_5")]; + tensor var_365 = mul(x = var_348, y = roti)[name = tensor("op_365")]; + tensor var_366 = mul(x = var_350, y = rotr)[name = tensor("op_366")]; + tensor koi_5 = add(x = var_365, y = var_366)[name = tensor("koi_5")]; + tensor qo_axis_0 = const()[name = tensor("qo_axis_0"), val = tensor(-1)]; + tensor qo = stack(axis = qo_axis_0, values = (qor_5, qoi_5))[name = tensor("qo")]; + tensor ko_axis_0 = const()[name = tensor("ko_axis_0"), val = tensor(-1)]; + tensor ko = stack(axis = ko_axis_0, values = (kor_5, koi_5))[name = tensor("ko")]; + tensor var_376 = const()[name = tensor("op_376"), val = tensor([1, 16, 8, 64])]; + tensor q = reshape(shape = var_376, x = qo)[name = tensor("q")]; + tensor var_378 = const()[name = tensor("op_378"), val = tensor([1, 16, 8, 64])]; + tensor k = reshape(shape = var_378, x = ko)[name = tensor("k")]; + tensor capacity = const()[name = tensor("capacity"), val = tensor([256])]; + tensor var_383_dtype_0 = const()[name = tensor("op_383_dtype_0"), val = tensor("int32")]; + tensor var_384 = const()[name = tensor("op_384"), val = tensor([1, 1])]; + tensor var_383 = cast(dtype = var_383_dtype_0, x = attn1_offset)[name = tensor("cast_46")]; + tensor write_base = reshape(shape = var_384, x = var_383)[name = tensor("write_base")]; + tensor write_range = const()[name = tensor("write_range"), val = tensor([[0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15]])]; + tensor abs_idx = add(x = write_base, y = write_range)[name = tensor("abs_idx")]; + tensor wrapped_div = floor_div(x = abs_idx, y = capacity)[name = tensor("wrapped_div")]; + tensor wrapped_div_scaled = mul(x = wrapped_div, y = capacity)[name = tensor("wrapped_div_scaled")]; + tensor wrapped = sub(x = abs_idx, y = wrapped_div_scaled)[name = tensor("wrapped")]; + tensor var_391 = const()[name = tensor("op_391"), val = tensor([1, 16, 1, 1])]; + tensor var_392 = reshape(shape = var_391, x = wrapped)[name = tensor("op_392")]; + tensor write_indexes_reps_0 = const()[name = tensor("write_indexes_reps_0"), val = tensor([1, 1, 8, 64])]; + tensor write_indexes = tile(reps = write_indexes_reps_0, x = var_392)[name = tensor("write_indexes")]; + tensor var_395_begin_0 = const()[name = tensor("op_395_begin_0"), val = tensor([0, 0, 0, 0, 0])]; + tensor var_395_end_0 = const()[name = tensor("op_395_end_0"), val = tensor([1, 1, 256, 8, 64])]; + tensor var_395_end_mask_0 = const()[name = tensor("op_395_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_395_squeeze_mask_0 = const()[name = tensor("op_395_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor var_395 = slice_by_index(begin = var_395_begin_0, end = var_395_end_0, end_mask = var_395_end_mask_0, squeeze_mask = var_395_squeeze_mask_0, x = attn1_cache)[name = tensor("op_395")]; + tensor new_k_cache_5_axis_0 = const()[name = tensor("new_k_cache_5_axis_0"), val = tensor(1)]; + tensor new_k_cache_5_mode_0 = const()[name = tensor("new_k_cache_5_mode_0"), val = tensor("update")]; + tensor new_k_cache_5_validate_indices_0 = const()[name = tensor("new_k_cache_5_validate_indices_0"), val = tensor(false)]; + tensor new_k_cache_5 = scatter_along_axis(axis = new_k_cache_5_axis_0, data = var_395, indices = write_indexes, mode = new_k_cache_5_mode_0, updates = k, validate_indices = new_k_cache_5_validate_indices_0)[name = tensor("new_k_cache_5")]; + tensor var_397_begin_0 = const()[name = tensor("op_397_begin_0"), val = tensor([1, 0, 0, 0, 0])]; + tensor var_397_end_0 = const()[name = tensor("op_397_end_0"), val = tensor([2, 1, 256, 8, 64])]; + tensor var_397_end_mask_0 = const()[name = tensor("op_397_end_mask_0"), val = tensor([false, true, true, true, true])]; + tensor var_397_squeeze_mask_0 = const()[name = tensor("op_397_squeeze_mask_0"), val = tensor([true, false, false, false, false])]; + tensor var_397 = slice_by_index(begin = var_397_begin_0, end = var_397_end_0, end_mask = var_397_end_mask_0, squeeze_mask = var_397_squeeze_mask_0, x = attn1_cache)[name = tensor("op_397")]; + tensor new_v_cache_5_axis_0 = const()[name = tensor("new_v_cache_5_axis_0"), val = tensor(1)]; + tensor new_v_cache_5_mode_0 = const()[name = tensor("new_v_cache_5_mode_0"), val = tensor("update")]; + tensor new_v_cache_5_validate_indices_0 = const()[name = tensor("new_v_cache_5_validate_indices_0"), val = tensor(false)]; + tensor new_v_cache_5 = scatter_along_axis(axis = new_v_cache_5_axis_0, data = var_397, indices = write_indexes, mode = new_v_cache_5_mode_0, updates = squeeze_5, validate_indices = new_v_cache_5_validate_indices_0)[name = tensor("new_v_cache_5")]; + tensor var_400_axis_0 = const()[name = tensor("op_400_axis_0"), val = tensor(0)]; + tensor var_400 = stack(axis = var_400_axis_0, values = (new_k_cache_5, new_v_cache_5))[name = tensor("op_400")]; + tensor var_401 = not_equal(x = new_k_cache_5, y = new_k_cache_5)[name = tensor("op_401")]; + tensor new_k_cache = select(a = var_212, b = new_k_cache_5, cond = var_401)[name = tensor("new_k_cache")]; + tensor var_404 = not_equal(x = new_v_cache_5, y = new_v_cache_5)[name = tensor("op_404")]; + tensor new_v_cache = select(a = var_212, b = new_v_cache_5, cond = var_404)[name = tensor("new_v_cache")]; + tensor var_409 = const()[name = tensor("op_409"), val = tensor([0, 2, 1, 3])]; + tensor var_411 = const()[name = tensor("op_411"), val = tensor([1, 1])]; + tensor var_412 = reshape(shape = var_411, x = attn1_offset)[name = tensor("op_412")]; + tensor var_414_promoted = const()[name = tensor("op_414_promoted"), val = tensor([0x1.ep+3])]; + tensor var_415 = add(x = var_412, y = var_414_promoted)[name = tensor("op_415")]; + tensor last_pos_dtype_0 = const()[name = tensor("last_pos_dtype_0"), val = tensor("int32")]; + tensor last_pos = cast(dtype = last_pos_dtype_0, x = var_415)[name = tensor("cast_45")]; + tensor diff = sub(x = last_pos, y = slot_idx_1)[name = tensor("diff")]; + tensor var_421_div = floor_div(x = diff, y = capacity)[name = tensor("op_421_div")]; + tensor var_421_div_scaled = mul(x = var_421_div, y = capacity)[name = tensor("op_421_div_scaled")]; + tensor var_421 = sub(x = diff, y = var_421_div_scaled)[name = tensor("op_421")]; + tensor pos_k = sub(x = last_pos, y = var_421)[name = tensor("pos_k")]; + tensor var_427_promoted = const()[name = tensor("op_427_promoted"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(41766144)))]; + tensor pos_q = add(x = var_412, y = var_427_promoted)[name = tensor("pos_q")]; + tensor var_431_axes_0 = const()[name = tensor("op_431_axes_0"), val = tensor([2])]; + tensor var_431 = expand_dims(axes = var_431_axes_0, x = pos_q)[name = tensor("op_431")]; + tensor var_433_axes_0 = const()[name = tensor("op_433_axes_0"), val = tensor([1])]; + tensor var_433 = expand_dims(axes = var_433_axes_0, x = pos_k)[name = tensor("op_433")]; + tensor var_434_promoted_dtype_0 = const()[name = tensor("op_434_promoted_dtype_0"), val = tensor("fp32")]; + tensor var_434_promoted = cast(dtype = var_434_promoted_dtype_0, x = var_433)[name = tensor("cast_44")]; + tensor delta = sub(x = var_431, y = var_434_promoted)[name = tensor("delta")]; + tensor valid_5 = greater_equal(x = var_433, y = var_86)[name = tensor("valid_5")]; + tensor var_443 = const()[name = tensor("op_443"), val = tensor([1, 1, 1])]; + tensor var_444 = reshape(shape = var_443, x = attn1_offset)[name = tensor("op_444")]; + tensor var_446_promoted = const()[name = tensor("op_446_promoted"), val = tensor([0x1.ep+3])]; + tensor var_447 = add(x = var_444, y = var_446_promoted)[name = tensor("op_447")]; + tensor var_448 = less_equal(x = var_434_promoted, y = var_447)[name = tensor("op_448")]; + tensor valid = logical_and(x = valid_5, y = var_448)[name = tensor("valid")]; + tensor var_86_promoted_1 = const()[name = tensor("op_86_promoted_1"), val = tensor(0x0p+0)]; + tensor var_450 = greater_equal(x = delta, y = var_86_promoted_1)[name = tensor("op_450")]; + tensor attn_mask_7 = logical_and(x = valid, y = var_450)[name = tensor("attn_mask_7")]; + tensor var_98_promoted_1 = const()[name = tensor("op_98_promoted_1"), val = tensor(0x1.f4p+7)]; + tensor var_452 = less(x = delta, y = var_98_promoted_1)[name = tensor("op_452")]; + tensor attn_mask_9 = logical_and(x = attn_mask_7, y = var_452)[name = tensor("attn_mask_9")]; + tensor attn_mask_axes_0 = const()[name = tensor("attn_mask_axes_0"), val = tensor([1])]; + tensor attn_mask = expand_dims(axes = attn_mask_axes_0, x = attn_mask_9)[name = tensor("attn_mask")]; + tensor var_457_transpose_x_0 = const()[name = tensor("op_457_transpose_x_0"), val = tensor(false)]; + tensor var_457_transpose_y_0 = const()[name = tensor("op_457_transpose_y_0"), val = tensor(false)]; + tensor transpose_8_perm_0 = const()[name = tensor("transpose_8_perm_0"), val = tensor([0, 2, -3, -1])]; + tensor transpose_9_perm_0 = const()[name = tensor("transpose_9_perm_0"), val = tensor([0, 2, -1, -3])]; + tensor transpose_9 = transpose(perm = transpose_9_perm_0, x = new_k_cache)[name = tensor("transpose_12")]; + tensor transpose_8 = transpose(perm = transpose_8_perm_0, x = q)[name = tensor("transpose_13")]; + tensor var_457 = matmul(transpose_x = var_457_transpose_x_0, transpose_y = var_457_transpose_y_0, x = transpose_8, y = transpose_9)[name = tensor("op_457")]; + tensor var_458 = const()[name = tensor("op_458"), val = tensor(0x1p-3)]; + tensor attn_7 = mul(x = var_457, y = var_458)[name = tensor("attn_7")]; + tensor var_460 = logical_not(x = attn_mask)[name = tensor("op_460")]; + tensor attn_9 = select(a = var_100, b = attn_7, cond = var_460)[name = tensor("attn_9")]; + tensor attn = softmax(axis = var_91, x = attn_9)[name = tensor("attn")]; + tensor x_11_transpose_x_0 = const()[name = tensor("x_11_transpose_x_0"), val = tensor(false)]; + tensor x_11_transpose_y_0 = const()[name = tensor("x_11_transpose_y_0"), val = tensor(false)]; + tensor v_attn = transpose(perm = var_409, x = new_v_cache)[name = tensor("transpose_14")]; + tensor x_11 = matmul(transpose_x = x_11_transpose_x_0, transpose_y = x_11_transpose_y_0, x = attn, y = v_attn)[name = tensor("x_11")]; + tensor var_464_perm_0 = const()[name = tensor("op_464_perm_0"), val = tensor([0, 2, 1, 3])]; + tensor var_465 = const()[name = tensor("op_465"), val = tensor([1, 16, 512])]; + tensor var_464 = transpose(perm = var_464_perm_0, x = x_11)[name = tensor("transpose_11")]; + tensor input_15 = reshape(shape = var_465, x = var_464)[name = tensor("input_15")]; + tensor x_13 = linear(bias = linear_1_bias_0, weight = mimi_decoder_transformer_transformer_layers_1_self_attn_out_proj_weight, x = input_15)[name = tensor("linear_5")]; + tensor var_474 = mul(x = mimi_decoder_transformer_transformer_layers_1_layer_scale_1_scale, y = x_13)[name = tensor("op_474")]; + tensor input_17 = add(x = input_13, y = var_474)[name = tensor("input_17")]; + tensor input_19_axes_0 = const()[name = tensor("input_19_axes_0"), val = tensor([-1])]; + tensor input_19 = layer_norm(axes = input_19_axes_0, beta = mimi_decoder_transformer_transformer_layers_1_norm2_bias, epsilon = var_102, gamma = mimi_decoder_transformer_transformer_layers_1_norm2_weight, x = input_17)[name = tensor("input_19")]; + tensor var_481 = linear(bias = linear_2_bias_0, weight = mimi_decoder_transformer_transformer_layers_1_linear1_weight, x = input_19)[name = tensor("linear_6")]; + tensor input_21_mode_0 = const()[name = tensor("input_21_mode_0"), val = tensor("EXACT")]; + tensor input_21 = gelu(mode = input_21_mode_0, x = var_481)[name = tensor("input_21")]; + tensor x_15 = linear(bias = linear_1_bias_0, weight = mimi_decoder_transformer_transformer_layers_1_linear2_weight, x = input_21)[name = tensor("linear_7")]; + tensor var_487 = mul(x = mimi_decoder_transformer_transformer_layers_1_layer_scale_2_scale, y = x_15)[name = tensor("op_487")]; + tensor z = add(x = input_17, y = var_487)[name = tensor("z")]; + tensor x_17_perm_0 = const()[name = tensor("x_17_perm_0"), val = tensor([0, 2, 1])]; + tensor var_507 = const()[name = tensor("op_507"), val = tensor(0x1p+0)]; + tensor var_508 = const()[name = tensor("op_508"), val = tensor(-1)]; + tensor input_23_interleave_0 = const()[name = tensor("input_23_interleave_0"), val = tensor(false)]; + tensor x_17 = transpose(perm = x_17_perm_0, x = z)[name = tensor("transpose_10")]; + tensor input_23 = concat(axis = var_508, interleave = input_23_interleave_0, values = (conv0_prev, x_17))[name = tensor("input_23")]; + tensor input_25_pad_type_0 = const()[name = tensor("input_25_pad_type_0"), val = tensor("valid")]; + tensor input_25_strides_0 = const()[name = tensor("input_25_strides_0"), val = tensor([1])]; + tensor input_25_pad_0 = const()[name = tensor("input_25_pad_0"), val = tensor([0, 0])]; + tensor input_25_dilations_0 = const()[name = tensor("input_25_dilations_0"), val = tensor([1])]; + tensor input_25_groups_0 = const()[name = tensor("input_25_groups_0"), val = tensor(1)]; + tensor input_25 = conv(bias = mimi_decoder_model_0_conv_bias, dilations = input_25_dilations_0, groups = input_25_groups_0, pad = input_25_pad_0, pad_type = input_25_pad_type_0, strides = input_25_strides_0, weight = mimi_decoder_model_0_conv_weight, x = input_23)[name = tensor("input_25")]; + tensor var_542_begin_0 = const()[name = tensor("op_542_begin_0"), val = tensor([0, 0, 16])]; + tensor var_542_end_0 = const()[name = tensor("op_542_end_0"), val = tensor([1, 512, 22])]; + tensor var_542_end_mask_0 = const()[name = tensor("op_542_end_mask_0"), val = tensor([true, true, true])]; + tensor var_542 = slice_by_index(begin = var_542_begin_0, end = var_542_end_0, end_mask = var_542_end_mask_0, x = input_23)[name = tensor("op_542")]; + tensor input_27 = elu(alpha = var_507, x = input_25)[name = tensor("input_27")]; + tensor y_5_pad_type_0 = const()[name = tensor("y_5_pad_type_0"), val = tensor("valid")]; + tensor y_5_strides_0 = const()[name = tensor("y_5_strides_0"), val = tensor([6])]; + tensor y_5_pad_0 = const()[name = tensor("y_5_pad_0"), val = tensor([0, 0])]; + tensor y_5_dilations_0 = const()[name = tensor("y_5_dilations_0"), val = tensor([1])]; + tensor y_5_groups_0 = const()[name = tensor("y_5_groups_0"), val = tensor(1)]; + tensor y_5_has_output_shape_output_shape_0 = const()[name = tensor("y_5_has_output_shape_output_shape_0"), val = tensor([1, 256, 102])]; + tensor y_5_has_output_shape = conv_transpose(bias = mimi_decoder_model_2_convtr_bias, dilations = y_5_dilations_0, groups = y_5_groups_0, output_shape = y_5_has_output_shape_output_shape_0, pad = y_5_pad_0, pad_type = y_5_pad_type_0, strides = y_5_strides_0, weight = mimi_decoder_model_2_convtr_weight, x = input_27)[name = tensor("y_5_has_output_shape")]; + tensor var_557_begin_0 = const()[name = tensor("op_557_begin_0"), val = tensor([0, 0, 0])]; + tensor var_557_end_0 = const()[name = tensor("op_557_end_0"), val = tensor([1, 256, 6])]; + tensor var_557_end_mask_0 = const()[name = tensor("op_557_end_mask_0"), val = tensor([true, true, false])]; + tensor var_557 = slice_by_index(begin = var_557_begin_0, end = var_557_end_0, end_mask = var_557_end_mask_0, x = y_5_has_output_shape)[name = tensor("op_557")]; + tensor var_558 = add(x = var_557, y = convtr0_partial)[name = tensor("op_558")]; + tensor var_559_begin_0 = const()[name = tensor("op_559_begin_0"), val = tensor([0, 0, 6])]; + tensor var_559_end_0 = const()[name = tensor("op_559_end_0"), val = tensor([1, 256, 102])]; + tensor var_559_end_mask_0 = const()[name = tensor("op_559_end_mask_0"), val = tensor([true, true, true])]; + tensor var_559 = slice_by_index(begin = var_559_begin_0, end = var_559_end_0, end_mask = var_559_end_mask_0, x = y_5_has_output_shape)[name = tensor("op_559")]; + tensor y_7_interleave_0 = const()[name = tensor("y_7_interleave_0"), val = tensor(false)]; + tensor y_7 = concat(axis = var_508, interleave = y_7_interleave_0, values = (var_558, var_559))[name = tensor("y_7")]; + tensor new_partial_1_begin_0 = const()[name = tensor("new_partial_1_begin_0"), val = tensor([0, 0, 96])]; + tensor new_partial_1_end_0 = const()[name = tensor("new_partial_1_end_0"), val = tensor([1, 256, 102])]; + tensor new_partial_1_end_mask_0 = const()[name = tensor("new_partial_1_end_mask_0"), val = tensor([true, true, true])]; + tensor new_partial_1 = slice_by_index(begin = new_partial_1_begin_0, end = new_partial_1_end_0, end_mask = new_partial_1_end_mask_0, x = y_7)[name = tensor("new_partial_1")]; + tensor var_564 = const()[name = tensor("op_564"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(41766272)))]; + tensor var_565 = sub(x = new_partial_1, y = var_564)[name = tensor("op_565")]; + tensor input_29_begin_0 = const()[name = tensor("input_29_begin_0"), val = tensor([0, 0, 0])]; + tensor input_29_end_0 = const()[name = tensor("input_29_end_0"), val = tensor([1, 256, 96])]; + tensor input_29_end_mask_0 = const()[name = tensor("input_29_end_mask_0"), val = tensor([true, true, false])]; + tensor input_29 = slice_by_index(begin = input_29_begin_0, end = input_29_end_0, end_mask = input_29_end_mask_0, x = y_7)[name = tensor("input_29")]; + tensor x_19 = elu(alpha = var_507, x = input_29)[name = tensor("x_19")]; + tensor input_31_interleave_0 = const()[name = tensor("input_31_interleave_0"), val = tensor(false)]; + tensor input_31 = concat(axis = var_508, interleave = input_31_interleave_0, values = (res0_conv0_prev, x_19))[name = tensor("input_31")]; + tensor input_33_pad_type_0 = const()[name = tensor("input_33_pad_type_0"), val = tensor("valid")]; + tensor input_33_strides_0 = const()[name = tensor("input_33_strides_0"), val = tensor([1])]; + tensor input_33_pad_0 = const()[name = tensor("input_33_pad_0"), val = tensor([0, 0])]; + tensor input_33_dilations_0 = const()[name = tensor("input_33_dilations_0"), val = tensor([1])]; + tensor input_33_groups_0 = const()[name = tensor("input_33_groups_0"), val = tensor(1)]; + tensor input_33 = conv(bias = mimi_decoder_model_3_block_1_conv_bias, dilations = input_33_dilations_0, groups = input_33_groups_0, pad = input_33_pad_0, pad_type = input_33_pad_type_0, strides = input_33_strides_0, weight = mimi_decoder_model_3_block_1_conv_weight, x = input_31)[name = tensor("input_33")]; + tensor var_585_begin_0 = const()[name = tensor("op_585_begin_0"), val = tensor([0, 0, 96])]; + tensor var_585_end_0 = const()[name = tensor("op_585_end_0"), val = tensor([1, 256, 98])]; + tensor var_585_end_mask_0 = const()[name = tensor("op_585_end_mask_0"), val = tensor([true, true, true])]; + tensor var_585 = slice_by_index(begin = var_585_begin_0, end = var_585_end_0, end_mask = var_585_end_mask_0, x = input_31)[name = tensor("op_585")]; + tensor x_21 = elu(alpha = var_507, x = input_33)[name = tensor("x_21")]; + tensor v_5_pad_type_0 = const()[name = tensor("v_5_pad_type_0"), val = tensor("valid")]; + tensor v_5_strides_0 = const()[name = tensor("v_5_strides_0"), val = tensor([1])]; + tensor v_5_pad_0 = const()[name = tensor("v_5_pad_0"), val = tensor([0, 0])]; + tensor v_5_dilations_0 = const()[name = tensor("v_5_dilations_0"), val = tensor([1])]; + tensor v_5_groups_0 = const()[name = tensor("v_5_groups_0"), val = tensor(1)]; + tensor v_5 = conv(bias = mimi_decoder_model_3_block_3_conv_bias, dilations = v_5_dilations_0, groups = v_5_groups_0, pad = v_5_pad_0, pad_type = v_5_pad_type_0, strides = v_5_strides_0, weight = mimi_decoder_model_3_block_3_conv_weight, x = x_21)[name = tensor("v_5")]; + tensor input_35 = add(x = input_29, y = v_5)[name = tensor("input_35")]; + tensor input_37 = elu(alpha = var_507, x = input_35)[name = tensor("input_37")]; + tensor y_9_pad_type_0 = const()[name = tensor("y_9_pad_type_0"), val = tensor("valid")]; + tensor y_9_strides_0 = const()[name = tensor("y_9_strides_0"), val = tensor([5])]; + tensor y_9_pad_0 = const()[name = tensor("y_9_pad_0"), val = tensor([0, 0])]; + tensor y_9_dilations_0 = const()[name = tensor("y_9_dilations_0"), val = tensor([1])]; + tensor y_9_groups_0 = const()[name = tensor("y_9_groups_0"), val = tensor(1)]; + tensor y_9_has_output_shape_output_shape_0 = const()[name = tensor("y_9_has_output_shape_output_shape_0"), val = tensor([1, 128, 485])]; + tensor y_9_has_output_shape = conv_transpose(bias = mimi_decoder_model_5_convtr_bias, dilations = y_9_dilations_0, groups = y_9_groups_0, output_shape = y_9_has_output_shape_output_shape_0, pad = y_9_pad_0, pad_type = y_9_pad_type_0, strides = y_9_strides_0, weight = mimi_decoder_model_5_convtr_weight, x = input_37)[name = tensor("y_9_has_output_shape")]; + tensor var_613_begin_0 = const()[name = tensor("op_613_begin_0"), val = tensor([0, 0, 0])]; + tensor var_613_end_0 = const()[name = tensor("op_613_end_0"), val = tensor([1, 128, 5])]; + tensor var_613_end_mask_0 = const()[name = tensor("op_613_end_mask_0"), val = tensor([true, true, false])]; + tensor var_613 = slice_by_index(begin = var_613_begin_0, end = var_613_end_0, end_mask = var_613_end_mask_0, x = y_9_has_output_shape)[name = tensor("op_613")]; + tensor var_614 = add(x = var_613, y = convtr1_partial)[name = tensor("op_614")]; + tensor var_615_begin_0 = const()[name = tensor("op_615_begin_0"), val = tensor([0, 0, 5])]; + tensor var_615_end_0 = const()[name = tensor("op_615_end_0"), val = tensor([1, 128, 485])]; + tensor var_615_end_mask_0 = const()[name = tensor("op_615_end_mask_0"), val = tensor([true, true, true])]; + tensor var_615 = slice_by_index(begin = var_615_begin_0, end = var_615_end_0, end_mask = var_615_end_mask_0, x = y_9_has_output_shape)[name = tensor("op_615")]; + tensor y_11_interleave_0 = const()[name = tensor("y_11_interleave_0"), val = tensor(false)]; + tensor y_11 = concat(axis = var_508, interleave = y_11_interleave_0, values = (var_614, var_615))[name = tensor("y_11")]; + tensor new_partial_3_begin_0 = const()[name = tensor("new_partial_3_begin_0"), val = tensor([0, 0, 480])]; + tensor new_partial_3_end_0 = const()[name = tensor("new_partial_3_end_0"), val = tensor([1, 128, 485])]; + tensor new_partial_3_end_mask_0 = const()[name = tensor("new_partial_3_end_mask_0"), val = tensor([true, true, true])]; + tensor new_partial_3 = slice_by_index(begin = new_partial_3_begin_0, end = new_partial_3_end_0, end_mask = new_partial_3_end_mask_0, x = y_11)[name = tensor("new_partial_3")]; + tensor var_620 = const()[name = tensor("op_620"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(41767360)))]; + tensor var_621 = sub(x = new_partial_3, y = var_620)[name = tensor("op_621")]; + tensor input_39_begin_0 = const()[name = tensor("input_39_begin_0"), val = tensor([0, 0, 0])]; + tensor input_39_end_0 = const()[name = tensor("input_39_end_0"), val = tensor([1, 128, 480])]; + tensor input_39_end_mask_0 = const()[name = tensor("input_39_end_mask_0"), val = tensor([true, true, false])]; + tensor input_39 = slice_by_index(begin = input_39_begin_0, end = input_39_end_0, end_mask = input_39_end_mask_0, x = y_11)[name = tensor("input_39")]; + tensor x_23 = elu(alpha = var_507, x = input_39)[name = tensor("x_23")]; + tensor input_41_interleave_0 = const()[name = tensor("input_41_interleave_0"), val = tensor(false)]; + tensor input_41 = concat(axis = var_508, interleave = input_41_interleave_0, values = (res1_conv0_prev, x_23))[name = tensor("input_41")]; + tensor input_43_pad_type_0 = const()[name = tensor("input_43_pad_type_0"), val = tensor("valid")]; + tensor input_43_strides_0 = const()[name = tensor("input_43_strides_0"), val = tensor([1])]; + tensor input_43_pad_0 = const()[name = tensor("input_43_pad_0"), val = tensor([0, 0])]; + tensor input_43_dilations_0 = const()[name = tensor("input_43_dilations_0"), val = tensor([1])]; + tensor input_43_groups_0 = const()[name = tensor("input_43_groups_0"), val = tensor(1)]; + tensor input_43 = conv(bias = mimi_decoder_model_6_block_1_conv_bias, dilations = input_43_dilations_0, groups = input_43_groups_0, pad = input_43_pad_0, pad_type = input_43_pad_type_0, strides = input_43_strides_0, weight = mimi_decoder_model_6_block_1_conv_weight, x = input_41)[name = tensor("input_43")]; + tensor var_641_begin_0 = const()[name = tensor("op_641_begin_0"), val = tensor([0, 0, 480])]; + tensor var_641_end_0 = const()[name = tensor("op_641_end_0"), val = tensor([1, 128, 482])]; + tensor var_641_end_mask_0 = const()[name = tensor("op_641_end_mask_0"), val = tensor([true, true, true])]; + tensor var_641 = slice_by_index(begin = var_641_begin_0, end = var_641_end_0, end_mask = var_641_end_mask_0, x = input_41)[name = tensor("op_641")]; + tensor x_25 = elu(alpha = var_507, x = input_43)[name = tensor("x_25")]; + tensor v_7_pad_type_0 = const()[name = tensor("v_7_pad_type_0"), val = tensor("valid")]; + tensor v_7_strides_0 = const()[name = tensor("v_7_strides_0"), val = tensor([1])]; + tensor v_7_pad_0 = const()[name = tensor("v_7_pad_0"), val = tensor([0, 0])]; + tensor v_7_dilations_0 = const()[name = tensor("v_7_dilations_0"), val = tensor([1])]; + tensor v_7_groups_0 = const()[name = tensor("v_7_groups_0"), val = tensor(1)]; + tensor v_7 = conv(bias = mimi_decoder_model_6_block_3_conv_bias, dilations = v_7_dilations_0, groups = v_7_groups_0, pad = v_7_pad_0, pad_type = v_7_pad_type_0, strides = v_7_strides_0, weight = mimi_decoder_model_6_block_3_conv_weight, x = x_25)[name = tensor("v_7")]; + tensor input_45 = add(x = input_39, y = v_7)[name = tensor("input_45")]; + tensor input_47 = elu(alpha = var_507, x = input_45)[name = tensor("input_47")]; + tensor y_13_pad_type_0 = const()[name = tensor("y_13_pad_type_0"), val = tensor("valid")]; + tensor y_13_strides_0 = const()[name = tensor("y_13_strides_0"), val = tensor([4])]; + tensor y_13_pad_0 = const()[name = tensor("y_13_pad_0"), val = tensor([0, 0])]; + tensor y_13_dilations_0 = const()[name = tensor("y_13_dilations_0"), val = tensor([1])]; + tensor y_13_groups_0 = const()[name = tensor("y_13_groups_0"), val = tensor(1)]; + tensor y_13_has_output_shape_output_shape_0 = const()[name = tensor("y_13_has_output_shape_output_shape_0"), val = tensor([1, 64, 1924])]; + tensor y_13_has_output_shape = conv_transpose(bias = mimi_decoder_model_8_convtr_bias, dilations = y_13_dilations_0, groups = y_13_groups_0, output_shape = y_13_has_output_shape_output_shape_0, pad = y_13_pad_0, pad_type = y_13_pad_type_0, strides = y_13_strides_0, weight = mimi_decoder_model_8_convtr_weight, x = input_47)[name = tensor("y_13_has_output_shape")]; + tensor var_669_begin_0 = const()[name = tensor("op_669_begin_0"), val = tensor([0, 0, 0])]; + tensor var_669_end_0 = const()[name = tensor("op_669_end_0"), val = tensor([1, 64, 4])]; + tensor var_669_end_mask_0 = const()[name = tensor("op_669_end_mask_0"), val = tensor([true, true, false])]; + tensor var_669 = slice_by_index(begin = var_669_begin_0, end = var_669_end_0, end_mask = var_669_end_mask_0, x = y_13_has_output_shape)[name = tensor("op_669")]; + tensor var_670 = add(x = var_669, y = convtr2_partial)[name = tensor("op_670")]; + tensor var_671_begin_0 = const()[name = tensor("op_671_begin_0"), val = tensor([0, 0, 4])]; + tensor var_671_end_0 = const()[name = tensor("op_671_end_0"), val = tensor([1, 64, 1924])]; + tensor var_671_end_mask_0 = const()[name = tensor("op_671_end_mask_0"), val = tensor([true, true, true])]; + tensor var_671 = slice_by_index(begin = var_671_begin_0, end = var_671_end_0, end_mask = var_671_end_mask_0, x = y_13_has_output_shape)[name = tensor("op_671")]; + tensor y_interleave_0 = const()[name = tensor("y_interleave_0"), val = tensor(false)]; + tensor y = concat(axis = var_508, interleave = y_interleave_0, values = (var_670, var_671))[name = tensor("y")]; + tensor new_partial_begin_0 = const()[name = tensor("new_partial_begin_0"), val = tensor([0, 0, 1920])]; + tensor new_partial_end_0 = const()[name = tensor("new_partial_end_0"), val = tensor([1, 64, 1924])]; + tensor new_partial_end_mask_0 = const()[name = tensor("new_partial_end_mask_0"), val = tensor([true, true, true])]; + tensor new_partial = slice_by_index(begin = new_partial_begin_0, end = new_partial_end_0, end_mask = new_partial_end_mask_0, x = y)[name = tensor("new_partial")]; + tensor var_676 = const()[name = tensor("op_676"), val = tensor(BLOBFILE(path = tensor("@model_path/weights/weight.bin"), offset = tensor(41767936)))]; + tensor var_677 = sub(x = new_partial, y = var_676)[name = tensor("op_677")]; + tensor input_49_begin_0 = const()[name = tensor("input_49_begin_0"), val = tensor([0, 0, 0])]; + tensor input_49_end_0 = const()[name = tensor("input_49_end_0"), val = tensor([1, 64, 1920])]; + tensor input_49_end_mask_0 = const()[name = tensor("input_49_end_mask_0"), val = tensor([true, true, false])]; + tensor input_49 = slice_by_index(begin = input_49_begin_0, end = input_49_end_0, end_mask = input_49_end_mask_0, x = y)[name = tensor("input_49")]; + tensor x_27 = elu(alpha = var_507, x = input_49)[name = tensor("x_27")]; + tensor input_51_interleave_0 = const()[name = tensor("input_51_interleave_0"), val = tensor(false)]; + tensor input_51 = concat(axis = var_508, interleave = input_51_interleave_0, values = (res2_conv0_prev, x_27))[name = tensor("input_51")]; + tensor input_53_pad_type_0 = const()[name = tensor("input_53_pad_type_0"), val = tensor("valid")]; + tensor input_53_strides_0 = const()[name = tensor("input_53_strides_0"), val = tensor([1])]; + tensor input_53_pad_0 = const()[name = tensor("input_53_pad_0"), val = tensor([0, 0])]; + tensor input_53_dilations_0 = const()[name = tensor("input_53_dilations_0"), val = tensor([1])]; + tensor input_53_groups_0 = const()[name = tensor("input_53_groups_0"), val = tensor(1)]; + tensor input_53 = conv(bias = mimi_decoder_model_9_block_1_conv_bias, dilations = input_53_dilations_0, groups = input_53_groups_0, pad = input_53_pad_0, pad_type = input_53_pad_type_0, strides = input_53_strides_0, weight = mimi_decoder_model_9_block_1_conv_weight, x = input_51)[name = tensor("input_53")]; + tensor var_697_begin_0 = const()[name = tensor("op_697_begin_0"), val = tensor([0, 0, 1920])]; + tensor var_697_end_0 = const()[name = tensor("op_697_end_0"), val = tensor([1, 64, 1922])]; + tensor var_697_end_mask_0 = const()[name = tensor("op_697_end_mask_0"), val = tensor([true, true, true])]; + tensor var_697 = slice_by_index(begin = var_697_begin_0, end = var_697_end_0, end_mask = var_697_end_mask_0, x = input_51)[name = tensor("op_697")]; + tensor x_29 = elu(alpha = var_507, x = input_53)[name = tensor("x_29")]; + tensor v_pad_type_0 = const()[name = tensor("v_pad_type_0"), val = tensor("valid")]; + tensor v_strides_0 = const()[name = tensor("v_strides_0"), val = tensor([1])]; + tensor v_pad_0 = const()[name = tensor("v_pad_0"), val = tensor([0, 0])]; + tensor v_dilations_0 = const()[name = tensor("v_dilations_0"), val = tensor([1])]; + tensor v_groups_0 = const()[name = tensor("v_groups_0"), val = tensor(1)]; + tensor v = conv(bias = mimi_decoder_model_9_block_3_conv_bias, dilations = v_dilations_0, groups = v_groups_0, pad = v_pad_0, pad_type = v_pad_type_0, strides = v_strides_0, weight = mimi_decoder_model_9_block_3_conv_weight, x = x_29)[name = tensor("v")]; + tensor input_55 = add(x = input_49, y = v)[name = tensor("input_55")]; + tensor x = elu(alpha = var_507, x = input_55)[name = tensor("x")]; + tensor input_interleave_0 = const()[name = tensor("input_interleave_0"), val = tensor(false)]; + tensor input = concat(axis = var_508, interleave = input_interleave_0, values = (conv_final_prev, x))[name = tensor("input")]; + tensor var_724_pad_type_0 = const()[name = tensor("op_724_pad_type_0"), val = tensor("valid")]; + tensor var_724_strides_0 = const()[name = tensor("op_724_strides_0"), val = tensor([1])]; + tensor var_724_pad_0 = const()[name = tensor("op_724_pad_0"), val = tensor([0, 0])]; + tensor var_724_dilations_0 = const()[name = tensor("op_724_dilations_0"), val = tensor([1])]; + tensor var_724_groups_0 = const()[name = tensor("op_724_groups_0"), val = tensor(1)]; + tensor var_724 = conv(bias = mimi_decoder_model_11_conv_bias, dilations = var_724_dilations_0, groups = var_724_groups_0, pad = var_724_pad_0, pad_type = var_724_pad_type_0, strides = var_724_strides_0, weight = mimi_decoder_model_11_conv_weight, x = input)[name = tensor("op_724")]; + tensor var_725_begin_0 = const()[name = tensor("op_725_begin_0"), val = tensor([0, 0, 1920])]; + tensor var_725_end_0 = const()[name = tensor("op_725_end_0"), val = tensor([1, 64, 1922])]; + tensor var_725_end_mask_0 = const()[name = tensor("op_725_end_mask_0"), val = tensor([true, true, true])]; + tensor var_725 = slice_by_index(begin = var_725_begin_0, end = var_725_end_0, end_mask = var_725_end_mask_0, x = input)[name = tensor("op_725")]; + tensor var_740_promoted = const()[name = tensor("op_740_promoted"), val = tensor(0x1p+4)]; + tensor var_741 = add(x = attn0_offset, y = var_740_promoted)[name = tensor("op_741")]; + tensor var_743_promoted = const()[name = tensor("op_743_promoted"), val = tensor(0x1p+4)]; + tensor var_744 = add(x = attn1_offset, y = var_743_promoted)[name = tensor("op_744")]; + tensor conv0_first_tmp = identity(x = conv0_first)[name = tensor("conv0_first_tmp")]; + tensor res0_conv0_first_tmp = identity(x = res0_conv0_first)[name = tensor("res0_conv0_first_tmp")]; + tensor res0_conv1_prev_tmp = identity(x = res0_conv1_prev)[name = tensor("res0_conv1_prev_tmp")]; + tensor res0_conv1_first_tmp = identity(x = res0_conv1_first)[name = tensor("res0_conv1_first_tmp")]; + tensor res1_conv0_first_tmp = identity(x = res1_conv0_first)[name = tensor("res1_conv0_first_tmp")]; + tensor res1_conv1_prev_tmp = identity(x = res1_conv1_prev)[name = tensor("res1_conv1_prev_tmp")]; + tensor res1_conv1_first_tmp = identity(x = res1_conv1_first)[name = tensor("res1_conv1_first_tmp")]; + tensor res2_conv0_first_tmp = identity(x = res2_conv0_first)[name = tensor("res2_conv0_first_tmp")]; + tensor res2_conv1_prev_tmp = identity(x = res2_conv1_prev)[name = tensor("res2_conv1_prev_tmp")]; + tensor res2_conv1_first_tmp = identity(x = res2_conv1_first)[name = tensor("res2_conv1_first_tmp")]; + tensor conv_final_first_tmp = identity(x = conv_final_first)[name = tensor("conv_final_first_tmp")]; + } -> (var_724, var_77, var_210, var_741, var_400, var_744, var_542, conv0_first, var_565, var_585, res0_conv0_first, res0_conv1_prev, res0_conv1_first, var_621, var_641, res1_conv0_first, res1_conv1_prev, res1_conv1_first, var_677, var_697, res2_conv0_first, res2_conv1_prev, res2_conv1_first, var_725, conv_final_first); +} \ No newline at end of file diff --git a/v2.1/spanish_24l/mimi_decoder.mlmodelc/weights/weight.bin b/v2.1/spanish_24l/mimi_decoder.mlmodelc/weights/weight.bin new file mode 100644 index 0000000000000000000000000000000000000000..bd5e7b9637062a46e05dbd4dc1bda16bdc3045d8 --- /dev/null +++ b/v2.1/spanish_24l/mimi_decoder.mlmodelc/weights/weight.bin @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:bb7420c1509b2cffe63c553fa375975d9600df265bbb60aa8e200626f849b8e3 +size 41768256 diff --git a/v2.1/spanish_24l/mimi_decoder.mlpackage/Data/com.apple.CoreML/model.mlmodel b/v2.1/spanish_24l/mimi_decoder.mlpackage/Data/com.apple.CoreML/model.mlmodel new file mode 100644 index 0000000000000000000000000000000000000000..0a1d0bdb4f67d46ce73ccf8546acdfecc80ca8da --- /dev/null +++ b/v2.1/spanish_24l/mimi_decoder.mlpackage/Data/com.apple.CoreML/model.mlmodel @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:a640fae337ebc437fe039ecfd5e936f005f3e098515d025e67c8776c82f68066 +size 93270 diff --git a/v2.1/spanish_24l/mimi_decoder.mlpackage/Data/com.apple.CoreML/weights/weight.bin b/v2.1/spanish_24l/mimi_decoder.mlpackage/Data/com.apple.CoreML/weights/weight.bin new file mode 100644 index 0000000000000000000000000000000000000000..bd5e7b9637062a46e05dbd4dc1bda16bdc3045d8 --- /dev/null +++ b/v2.1/spanish_24l/mimi_decoder.mlpackage/Data/com.apple.CoreML/weights/weight.bin @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:bb7420c1509b2cffe63c553fa375975d9600df265bbb60aa8e200626f849b8e3 +size 41768256 diff --git a/v2.1/spanish_24l/mimi_decoder.mlpackage/Manifest.json b/v2.1/spanish_24l/mimi_decoder.mlpackage/Manifest.json new file mode 100644 index 0000000000000000000000000000000000000000..ce44e01af849695df42c78bf71350dff18bc8c39 --- /dev/null +++ b/v2.1/spanish_24l/mimi_decoder.mlpackage/Manifest.json @@ -0,0 +1,18 @@ +{ + "fileFormatVersion": "1.0.0", + "itemInfoEntries": { + "6E14DD69-0866-4ED3-9F2A-675A02B9536F": { + "author": "com.apple.CoreML", + "description": "CoreML Model Weights", + "name": "weights", + "path": "com.apple.CoreML/weights" + }, + "77834B4C-E044-4503-ADD6-AE395E120782": { + "author": "com.apple.CoreML", + "description": "CoreML Model Specification", + "name": "model.mlmodel", + "path": "com.apple.CoreML/model.mlmodel" + } + }, + "rootModelIdentifier": "77834B4C-E044-4503-ADD6-AE395E120782" +}